diff --git a/.gitattributes b/.gitattributes index ee64ee236a65156af215e31aecfda874854b50e9..683517c990bb5aca85114ff289a31d74277068c0 100644 --- a/.gitattributes +++ b/.gitattributes @@ -1909,3 +1909,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text 2024/SINDER_[[:space:]]Repairing[[:space:]]the[[:space:]]Singular[[:space:]]Defects[[:space:]]of[[:space:]]DINOv2/a11fb003-56e4-45a0-8b06-a6dadfe8a645_origin.pdf filter=lfs diff=lfs merge=lfs -text 2024/SLAck_[[:space:]]Semantic,[[:space:]]Location,[[:space:]]and[[:space:]]Appearance[[:space:]]Aware[[:space:]]Open-Vocabulary[[:space:]]Tracking/9878a5ca-4a3f-4a61-88bc-bdfef4a54480_origin.pdf filter=lfs diff=lfs merge=lfs -text 2024/SLEDGE_[[:space:]]Synthesizing[[:space:]]Driving[[:space:]]Environments[[:space:]]with[[:space:]]Generative[[:space:]]Models[[:space:]]and[[:space:]]Rule-Based[[:space:]]Traffic/35e94c06-38fe-4e5f-b845-97c01c4d8bfd_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/SLIM_[[:space:]]Spuriousness[[:space:]]Mitigation[[:space:]]with[[:space:]]Minimal[[:space:]]Human[[:space:]]Annotations/460c9d54-c412-4380-844b-3f656829c454_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/SMFANet_[[:space:]]A[[:space:]]Lightweight[[:space:]]Self-Modulation[[:space:]]Feature[[:space:]]Aggregation[[:space:]]Network[[:space:]]for[[:space:]]Efficient[[:space:]]Image[[:space:]]Super-Resolution/b8afe796-2e9d-45e7-96aa-cb21a9d8b896_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/SMILe_[[:space:]]Leveraging[[:space:]]Submodular[[:space:]]Mutual[[:space:]]Information[[:space:]]For[[:space:]]Robust[[:space:]]Few-Shot[[:space:]]Object[[:space:]]Detection/57f9e916-d947-4fa8-bb11-b89fc66615b3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/SMooDi_[[:space:]]Stylized[[:space:]]Motion[[:space:]]Diffusion[[:space:]]Model/10d37622-04ad-4741-9458-84cef09d41ff_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/SNP_[[:space:]]Structured[[:space:]]Neuron-level[[:space:]]Pruning[[:space:]]to[[:space:]]Preserve[[:space:]]Attention[[:space:]]Scores/cb547506-91af-4f34-aebd-6546cde66d30_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/SNeRV_[[:space:]]Spectra-preserving[[:space:]]Neural[[:space:]]Representation[[:space:]]for[[:space:]]Video/5d3bbd00-0b38-4046-ad52-4dc33526757f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/SOS_[[:space:]]Segment[[:space:]]Object[[:space:]]System[[:space:]]for[[:space:]]Open-World[[:space:]]Instance[[:space:]]Segmentation[[:space:]]With[[:space:]]Object[[:space:]]Priors/466dafa4-76e1-41c2-9ff8-766257e5e796_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/SPAMming[[:space:]]Labels_[[:space:]]Efficient[[:space:]]Annotations[[:space:]]for[[:space:]]the[[:space:]]Trackers[[:space:]]of[[:space:]]Tomorrow/0cf5ca3c-0b18-48c1-b497-d85ab9c4bdac_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/SPARO_[[:space:]]Selective[[:space:]]Attention[[:space:]]for[[:space:]]Robust[[:space:]]and[[:space:]]Compositional[[:space:]]Transformer[[:space:]]Encodings[[:space:]]for[[:space:]]Vision/4bccfc91-c07f-416d-bd12-0ad527681e54_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/SPHINX_[[:space:]]A[[:space:]]Mixer[[:space:]]of[[:space:]]Weights,[[:space:]]Visual[[:space:]]Embeddings[[:space:]]and[[:space:]]Image[[:space:]]Scales[[:space:]]for[[:space:]]Multi-modal[[:space:]]Large[[:space:]]Language[[:space:]]Models/cd141534-c355-432b-969a-838688ba26e3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/SPIN_[[:space:]]Hierarchical[[:space:]]Segmentation[[:space:]]with[[:space:]]Subpart[[:space:]]Granularity[[:space:]]in[[:space:]]Natural[[:space:]]Images/88afeb42-c98d-4785-a2a2-08cde2fa67ab_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/SPIRE_[[:space:]]Semantic[[:space:]]Prompt-Driven[[:space:]]Image[[:space:]]Restoration/0f07d6c7-f604-4c26-b13e-8eef2743217e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/SPVLoc_[[:space:]]Semantic[[:space:]]Panoramic[[:space:]]Viewport[[:space:]]Matching[[:space:]]for[[:space:]]6D[[:space:]]Camera[[:space:]]Localization[[:space:]]in[[:space:]]Unseen[[:space:]]Environments/a0f540fa-1e67-4e90-a2ae-34590f8ccef2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/SQ-LLaVA_[[:space:]]Self-Questioning[[:space:]]for[[:space:]]Large[[:space:]]Vision-Language[[:space:]]Assistant/3909e8d0-9729-43b7-abd9-e1c4bba247cf_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/SRPose_[[:space:]]Two-view[[:space:]]Relative[[:space:]]Pose[[:space:]]Estimation[[:space:]]with[[:space:]]Sparse[[:space:]]Keypoints/e429b1d6-1778-4297-b226-264e6bf36452_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/SSL-Cleanse_[[:space:]]Trojan[[:space:]]Detection[[:space:]]and[[:space:]]Mitigation[[:space:]]in[[:space:]]Self-Supervised[[:space:]]Learning/a42be677-fdb6-4ca1-8003-8fcf32b7a42b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/ST-LDM_[[:space:]]A[[:space:]]Universal[[:space:]]Framework[[:space:]]for[[:space:]]Text-Grounded[[:space:]]Object[[:space:]]Generation[[:space:]]in[[:space:]]Real[[:space:]]Images/ea78a6df-2485-4a42-a86e-90a6085c52c4_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/ST-LLM_[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]Are[[:space:]]Effective[[:space:]]Temporal[[:space:]]Learners/f824889a-c655-4483-9dd5-c5cb83e43784_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/STAG4D_[[:space:]]Spatial-Temporal[[:space:]]Anchored[[:space:]]Generative[[:space:]]4D[[:space:]]Gaussians/51eb6cce-864e-4035-8901-6f032e5306da_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/STAMP_[[:space:]]Outlier-Aware[[:space:]]Test-Time[[:space:]]Adaptation[[:space:]]with[[:space:]]Stable[[:space:]]Memory[[:space:]]Replay/20b10b92-8cd4-4e35-ad35-a50ac0ff44cf_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/STSP_[[:space:]]Spatial-Temporal[[:space:]]Subspace[[:space:]]Projection[[:space:]]for[[:space:]]Video[[:space:]]Class-incremental[[:space:]]Learning/9413ef29-ccb6-4b51-bb25-c4b8dee0d763_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/SUMix_[[:space:]]Mixup[[:space:]]with[[:space:]]Semantic[[:space:]]and[[:space:]]Uncertain[[:space:]]Information/f0041215-e301-48ad-9db8-48c026069726_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/SUP-NeRF_[[:space:]]A[[:space:]]Streamlined[[:space:]]Unification[[:space:]]of[[:space:]]Pose[[:space:]]Estimation[[:space:]]and[[:space:]]NeRF[[:space:]]for[[:space:]]Monocular[[:space:]]3D[[:space:]]Object[[:space:]]Reconstruction/c913b291-07c4-425a-807f-281296d9ef23_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/SV3D_[[:space:]]Novel[[:space:]]Multi-view[[:space:]]Synthesis[[:space:]]and[[:space:]]3D[[:space:]]Generation[[:space:]]from[[:space:]]a[[:space:]]Single[[:space:]]Image[[:space:]]using[[:space:]]Latent[[:space:]]Video[[:space:]]Diffusion/55466408-8e29-4541-9064-bae13a2c8fbf_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/SWAG_[[:space:]]Splatting[[:space:]]in[[:space:]]the[[:space:]]Wild[[:space:]]images[[:space:]]with[[:space:]]Appearance-conditioned[[:space:]]Gaussians/3925312b-25b5-4314-8bc0-d6cd9369f8b8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/SWinGS_[[:space:]]Sliding[[:space:]]Windows[[:space:]]for[[:space:]]Dynamic[[:space:]]3D[[:space:]]Gaussian[[:space:]]Splatting/1089e742-f963-4d64-a493-ab6447fb05de_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/S^3D-NeRF_[[:space:]]Single-Shot[[:space:]]Speech-Driven[[:space:]]Neural[[:space:]]Radiance[[:space:]]Field[[:space:]]for[[:space:]]High[[:space:]]Fidelity[[:space:]]Talking[[:space:]]Head[[:space:]]Synthesis/49757949-cf34-4879-95e0-e53ca1bcc2a6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Safe-CLIP_[[:space:]]Removing[[:space:]]NSFW[[:space:]]Concepts[[:space:]]from[[:space:]]Vision-and-Language[[:space:]]Models/48fc840c-c29c-4199-8bf2-a181f591f8ca_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Safe-Sim_[[:space:]]Safety-Critical[[:space:]]Closed-Loop[[:space:]]Traffic[[:space:]]Simulation[[:space:]]with[[:space:]]Diffusion-Controllable[[:space:]]Adversaries/548909b9-64d9-4498-bf91-808d29ce48c5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Safeguard[[:space:]]Text-to-Image[[:space:]]Diffusion[[:space:]]Models[[:space:]]with[[:space:]]Human[[:space:]]Feedback[[:space:]]Inversion/603d7974-42c1-4361-a264-c2589e2c9681_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Salience-Based[[:space:]]Adaptive[[:space:]]Masking_[[:space:]]Revisiting[[:space:]]Token[[:space:]]Dynamics[[:space:]]for[[:space:]]Enhanced[[:space:]]Pre-training/05c12db4-7790-4b9e-84f8-e76fa8191a90_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Sapiens_[[:space:]]Foundation[[:space:]]for[[:space:]]Human[[:space:]]Vision[[:space:]]Models/726c3ced-d623-4195-b91f-26812a59bf2e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Scalable[[:space:]]Group[[:space:]]Choreography[[:space:]]via[[:space:]]Variational[[:space:]]Phase[[:space:]]Manifold[[:space:]]Learning/6dc9e2ab-125f-402e-97a2-dfb708ce273d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Scalar[[:space:]]Function[[:space:]]Topology[[:space:]]Divergence_[[:space:]]Comparing[[:space:]]Topology[[:space:]]of[[:space:]]3D[[:space:]]Objects/228bb063-0fe3-438c-a000-0e252dd17cc3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/ScaleDreamer_[[:space:]]Scalable[[:space:]]Text-to-3D[[:space:]]Synthesis[[:space:]]with[[:space:]]Asynchronous[[:space:]]Score[[:space:]]Distillation/0ae3774a-e55b-4154-9080-499a583a8778_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Scaling[[:space:]]Backwards_[[:space:]]Minimal[[:space:]]Synthetic[[:space:]]Pre-training_/82902d64-0219-4fba-9a2b-20c3a32a53cf_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Scaling[[:space:]]Up[[:space:]]Personalized[[:space:]]Image[[:space:]]Aesthetic[[:space:]]Assessment[[:space:]]via[[:space:]]Task[[:space:]]Vector[[:space:]]Customization/77e3f8ef-24e2-4d25-ba6f-999ef6bc7191_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/ScanReason_[[:space:]]Empowering[[:space:]]3D[[:space:]]Visual[[:space:]]Grounding[[:space:]]with[[:space:]]Reasoning[[:space:]]Capabilities/c0158ad2-ca00-4982-8b2e-6cbbd910c1e7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/ScanTalk_[[:space:]]3D[[:space:]]Talking[[:space:]]Heads[[:space:]]from[[:space:]]Unregistered[[:space:]]Scans/34e5d7c9-005f-4b11-aa72-5d7f3ba885f0_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/ScatterFormer_[[:space:]]Efficient[[:space:]]Voxel[[:space:]]Transformer[[:space:]]with[[:space:]]Scattered[[:space:]]Linear[[:space:]]Attention/1045fb37-bbfd-4adc-8de1-f127bb3078ca_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Scene[[:space:]]Coordinate[[:space:]]Reconstruction_[[:space:]]Posing[[:space:]]of[[:space:]]Image[[:space:]]Collections[[:space:]]via[[:space:]]Incremental[[:space:]]Learning[[:space:]]of[[:space:]]a[[:space:]]Relocalizer/c513fe61-35e0-435f-acdb-dbf8715686b2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Scene-Conditional[[:space:]]3D[[:space:]]Object[[:space:]]Stylization[[:space:]]and[[:space:]]Composition/ed573969-4bd1-4a9b-b748-12ae2bb2b2d3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Scene-Graph[[:space:]]ViT_[[:space:]]End-to-End[[:space:]]Open-Vocabulary[[:space:]]Visual[[:space:]]Relationship[[:space:]]Detection/07085046-a430-402b-be81-31056599f4f5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Scene-aware[[:space:]]Human[[:space:]]Motion[[:space:]]Forecasting[[:space:]]via[[:space:]]Mutual[[:space:]]Distance[[:space:]]Prediction/abecdb1c-2c3f-4eb7-820c-6ae2bc3ee016_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/SceneGraphLoc_[[:space:]]Cross-Modal[[:space:]]Coarse[[:space:]]Visual[[:space:]]Localization[[:space:]]on[[:space:]]3D[[:space:]]Scene[[:space:]]Graphs/ca6ad507-bc20-40c4-b703-2093582a66a7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/SceneScript_[[:space:]]Reconstructing[[:space:]]Scenes[[:space:]]With[[:space:]]An[[:space:]]Autoregressive[[:space:]]Structured[[:space:]]Language[[:space:]]Model/6d586386-d370-4aeb-a27d-9451b501b037_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/SceneTeller_[[:space:]]Language-to-3D[[:space:]]Scene[[:space:]]Generation/dec651d5-2a06-4ad2-af7e-aeaf7a7dadd2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/SceneVerse_[[:space:]]Scaling[[:space:]]3D[[:space:]]Vision-Language[[:space:]]Learning[[:space:]]for[[:space:]]Grounded[[:space:]]Scene[[:space:]]Understanding/72c8327d-dc6a-436c-89b3-89891fd18764_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Scissorhands_[[:space:]]Scrub[[:space:]]Data[[:space:]]Influence[[:space:]]via[[:space:]]Connection[[:space:]]Sensitivity[[:space:]]in[[:space:]]Networks/43a51c48-5759-4f46-b600-93f450455ce7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Score[[:space:]]Distillation[[:space:]]Sampling[[:space:]]with[[:space:]]Learned[[:space:]]Manifold[[:space:]]Corrective/71b1df1d-3510-4a6f-9a35-a2f80ebbbb53_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/ScribblePrompt_[[:space:]]Fast[[:space:]]and[[:space:]]Flexible[[:space:]]Interactive[[:space:]]Segmentation[[:space:]]for[[:space:]]Any[[:space:]]Biomedical[[:space:]]Image/97fd2117-c4b3-47cc-8dc3-f7ab8ca17648_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/SeA_[[:space:]]Semantic[[:space:]]Adversarial[[:space:]]Augmentation[[:space:]]for[[:space:]]Last[[:space:]]Layer[[:space:]]Features[[:space:]]from[[:space:]]Unsupervised[[:space:]]Representation[[:space:]]Learning/ca46bbef-cd36-48f1-b01d-aebb511e39a1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/SeFlow_[[:space:]]A[[:space:]]Self-Supervised[[:space:]]Scene[[:space:]]Flow[[:space:]]Method[[:space:]]in[[:space:]]Autonomous[[:space:]]Driving/e7db8302-f2d3-4bc6-8b61-d5de1eb958ee_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/See[[:space:]]and[[:space:]]Think_[[:space:]]Embodied[[:space:]]Agent[[:space:]]in[[:space:]]Virtual[[:space:]]Environment/0ba9ae38-8422-4b1f-90c0-ee0bf0ad307a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Seeing[[:space:]]Faces[[:space:]]in[[:space:]]Things_[[:space:]]A[[:space:]]Model[[:space:]]and[[:space:]]Dataset[[:space:]]for[[:space:]]Pareidolia/78844de0-7820-498e-be36-c38ded4cc8e3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Seeing[[:space:]]the[[:space:]]Unseen_[[:space:]]A[[:space:]]Frequency[[:space:]]Prompt[[:space:]]Guided[[:space:]]Transformer[[:space:]]for[[:space:]]Image[[:space:]]Restoration/73dcd519-18ab-4cac-87c9-f1d814570e98_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/SegGen_[[:space:]]Supercharging[[:space:]]Segmentation[[:space:]]Models[[:space:]]with[[:space:]]Text2Mask[[:space:]]and[[:space:]]Mask2Img[[:space:]]Synthesis/acac7544-25ea-4a13-aba6-d24e96031cd3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/SegPoint_[[:space:]]Segment[[:space:]]Any[[:space:]]Point[[:space:]]Cloud[[:space:]]via[[:space:]]Large[[:space:]]Language[[:space:]]Model/d283f64a-a017-47de-9be3-9206675b2c4f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/SegVG_[[:space:]]Transferring[[:space:]]Object[[:space:]]Bounding[[:space:]]Box[[:space:]]to[[:space:]]Segmentation[[:space:]]for[[:space:]]Visual[[:space:]]Grounding/4d7de385-7f41-401e-b1cf-a9c2348c4d1c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Segment[[:space:]]and[[:space:]]Recognize[[:space:]]Anything[[:space:]]at[[:space:]]Any[[:space:]]Granularity/ec47fa22-1632-444f-b911-9d6b0be51867_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Segment,[[:space:]]Lift[[:space:]]and[[:space:]]Fit_[[:space:]]Automatic[[:space:]]3D[[:space:]]Shape[[:space:]]Labeling[[:space:]]from[[:space:]]2D[[:space:]]Prompts/478f2a06-4d04-42d7-815f-2c137ffb864d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Segment3D_[[:space:]]Learning[[:space:]]Fine-Grained[[:space:]]Class-Agnostic[[:space:]]3D[[:space:]]Segmentation[[:space:]]without[[:space:]]Manual[[:space:]]Labels/a98903ad-e70d-49e3-988e-6063b3741036_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Segmentation-guided[[:space:]]Layer-wise[[:space:]]Image[[:space:]]Vectorization[[:space:]]with[[:space:]]Gradient[[:space:]]Fills/01f8c0e2-d236-4d33-947c-2747a133c0bb_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/SeiT++_[[:space:]]Masked[[:space:]]Token[[:space:]]Modeling[[:space:]]Improves[[:space:]]Storage-efficient[[:space:]]Training/a6a99e76-474a-47eb-80d4-d9020aeb146a_origin.pdf filter=lfs diff=lfs merge=lfs -text diff --git a/2024/SLIM_ Spuriousness Mitigation with Minimal Human Annotations/460c9d54-c412-4380-844b-3f656829c454_content_list.json b/2024/SLIM_ Spuriousness Mitigation with Minimal Human Annotations/460c9d54-c412-4380-844b-3f656829c454_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ed7616bcd14a39307835f1978a3407ad2c814680 --- /dev/null +++ b/2024/SLIM_ Spuriousness Mitigation with Minimal Human Annotations/460c9d54-c412-4380-844b-3f656829c454_content_list.json @@ -0,0 +1,1544 @@ +[ + { + "type": "text", + "text": "SLIM: Spuriousness Mitigation with Minimal Human Annotations", + "text_level": 1, + "bbox": [ + 303, + 141, + 702, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiwei Xuan $^{1}$ , Ziquan Deng $^{1}$ , Hsuan-Tien Lin $^{2}$ , and Kwan-Liu Ma $^{1}$", + "bbox": [ + 238, + 212, + 764, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 University of California, Davis", + "bbox": [ + 393, + 239, + 611, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ National Taiwan University", + "bbox": [ + 401, + 253, + 601, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{xwxuan, ziqdeng, klma}@ucdavis.edu, htlin@csie.ntu.edu.tw", + "bbox": [ + 276, + 268, + 725, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Recent studies highlight that deep learning models often learn spurious features mistakenly linked to labels, compromising their reliability in real-world scenarios where such correlations do not hold. Despite the increasing research effort, existing solutions often face two main challenges: they either demand substantial annotations of spurious attributes, or they yield less competitive outcomes with expensive training when additional annotations are absent. In this paper, we introduce SLIM, a cost-effective and performance-targeted approach to reducing spurious correlations in deep learning. Our method leverages a human-in-the-loop protocol featuring a novel attention labeling mechanism with a constructed attention representation space. SLIM significantly reduces the need for exhaustive additional labeling, requiring human input for fewer than $3\\%$ of instances. By prioritizing data quality over complicated training strategies, SLIM curates a smaller yet more feature-balanced data subset, fostering the development of spuriousness-robust models. Experimental validations across key benchmarks demonstrate that SLIM competes with or exceeds the performance of leading methods while significantly reducing costs. The SLIM framework thus presents a promising path for developing reliable models more efficiently. Our code is available in https://github.com/xiweix/SLIM.git/.", + "bbox": [ + 261, + 321, + 743, + 599 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Model spuriousness $\\cdot$ Human-in-the-loop $\\cdot$ Data distribution", + "bbox": [ + 261, + 613, + 740, + 627 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 655, + 375, + 671 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Spurious correlations, where models mistakenly rely on irrelevant features to make decisions, pose a significant challenge in machine learning. Consequently, certain groups may experience deceptively inflated performance metrics, while accuracy reduces for others. This prevalence of \"right answers for wrong reasons\" fosters a misleading impression of model proficiency and constrains its utility across diverse contexts. Meanwhile, rectifying the issue of spuriousness is a challenging task, particularly when seeking to enhance the performance of underperforming groups. A notable example is the ISIC skin cancer dataset [5], where a typical spurious correlation occurs when color patches frequently co-occur with benign-labeled data. As shown in Fig. 1, a biased model may focus", + "bbox": [ + 212, + 689, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7b1247d216167746b4d9731ef33e662ebef38e44f8d72ee78a1a2d2a7e33efa9.jpg", + "image_caption": [ + "Fig. 1: (Left) GradCAM visualizations highlighting model attention on ISIC. (a, b) showcase a model biased towards patches, leading to a correct prediction with wrong reasons in (a) but an incorrect prediction despite correct focus in (b). (c, d) depict a spuriousness-robust model consistently focusing on core features. (Right) The table details the training data distribution by class (benign or malignant) and the presence of color patches, illustrating the imbalance and potential for spurious correlations." + ], + "image_footnote": [], + "bbox": [ + 217, + 145, + 785, + 203 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "on patches rather than skin lesions to identify melanoma, and fail when such patches are missing, while a desirable case would be the model captures correct skin regions regardless of spurious features' existence. Enhancing model robustness to address such dangerous misdiagnoses is crucial, driving a surge in research towards more effective solutions.", + "bbox": [ + 212, + 323, + 787, + 398 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A typical approach in this context is to seek for optimal data distribution. GDRO [27] proves the efficacy of group-balanced data in enhancing model robustness. However, along with subsequent studies [14, 22], such an approach requires spuriousness labels at dataset scale, consuming a considerable cost. Our investigations reveal notable inconsistencies in human-annotated spuriousness labels, raising concerns about their relia", + "bbox": [ + 212, + 407, + 509, + 573 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/7f66332c8cd36ad0b8883d48978b554e56a85cd43e874b201e9a2ddddfa93d62.jpg", + "table_caption": [ + "Table 1: Annotation forms and quantities required by different methods on Waterbirds, highlighting the reduced workload of SLIM. (Att. = Attention, T. = Training set, V. = Validation set)" + ], + "table_footnote": [], + "table_body": "
MethodAnno. FormAnno. Amount
RRR [26]Map100%×T.
GradIA [9]Att. Yes-No, Map100%×T.
Energy Loss [24]Bounding box100%×T.
CRAYON [15]Att. Yes-No100%×T.
GDRO [27]Spuriousness label100%×T.
DFR [14]Spuriousness label100%×V.
SLIM (ours)Att. Yes-No2.5%×T. or 3%×V.
", + "bbox": [ + 522, + 497, + 784, + 570 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "bility (refer to Table 6). There are other approaches with more direct human involvement, such as providing yes-no feedback on attention correctness [15,36]; or outlining core features with maps [9,26] or bounding boxes [24]. Despite different formats of queried information, they require extensive effort across the entire dataset, as summarized in Table 1. Considering the empirical applicability, our work aims to streamline this process by reducing human annotation workload and minimizing the ambiguity of collected data. Methods requiring spuriousness annotation tend to excel in performance at higher costs, while minimizing such expenses typically results in performance trade-offs [1,6]. Facing this challenge, techniques such as JTT [17] and CNC [41] design training or data-augmenting phases to reduce spuriousness without additional annotations. However, the extra computational cost raised by multiple training iterations is undesired. We believe, as spuriousness is indeed caused by data biases, improving data quality is more essential than designing complex learning strategies.", + "bbox": [ + 212, + 574, + 787, + 787 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Therefore, the remaining research gap calls for a data quality-oriented approach that balances performance and cost. This is where our framework, SLIM, steps in. $SLIM$ offers a novel approach to mitigating spurious correlations by", + "bbox": [ + 212, + 794, + 787, + 842 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "X. Xuan et al.", + "bbox": [ + 271, + 114, + 369, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "intelligently leveraging a minimal set of human-annotated data and curating a feature-balanced data set to enhance model robustness without any significant computational overhead incurred by many existing methods. As shown in Fig. 2, inferred from a reference model, $SLIM$ first constructs an attention representation space that reflects locally consistent data features and model attention, enabling efficient attention label estimation. — Representative instances can be sampled from this space for querying attention correctness labels and such labels can be expanded to neighbor instances. Subsequently, utilizing a visual explanation method that highlights model attention, like GradCAM [29], $SLIM$ disentangles core and environment features within the latent space. Here, \"core features\" denote features critical for class determination, whereas \"environment features\" represent the surrounding contextual information. We agree with GDRO that balanced data distribution is crucial, but instead of using human-annotated labels to approximate spuriousness distribution, we leverage latent space features directly disentangled from model attention to balance data distribution, which is more aligned with model understanding and involving less human assumptions. Through this, $SLIM$ assembles data groups with diverse feature combinations. Focused on data quality, a consistency-aware sampling mechanism is employed to curate a subset with optimal feature distributions — ensuring balanced core features across various environmental contexts. Finally, we conclude our framework with a straightforward training procedure.", + "bbox": [ + 217, + 146, + 785, + 462 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The experimental results demonstrate SLIM's advanced capability in spuriousness mitigation by performing comparably to, or exceeding state-of-the-art (SOTA) methods. Notably, SLIM accomplishes this with (1) a significantly reduced requirement for human annotations: $0.12\\% - 2.5\\%$ of the training set size and $0.35\\% - 4\\%$ of the validation set; and (2) a lightweight training for a robust model using our constructed data with $5\\% - 30\\%$ of the original set. Our code will be published after the review process is complete. Our contributions are:", + "bbox": [ + 217, + 464, + 785, + 569 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Data-quality-oriented spuriousness mitigation: SLIM introduces an interactive, simple, and effective data construction pipeline to create high-quality data for spuriousness mitigation.", + "- Cost-efficient framework: In contrast to related works with leading performance, SLIM demonstrates a significant reduction in both annotation requirements and computational cost.", + "- Competitive or superior model performance: Without incurring excessive costs, SLIM achieves comparable or superior performance to existing methods, confirming its effectiveness in mitigating spurious correlations." + ], + "bbox": [ + 215, + 577, + 784, + 709 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 217, + 734, + 385, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Spuriousness mitigation with spuriousness labels. When a comprehensive set of spuriousness labels is accessible during training, methods like class balancing [11] and importance weighting [32] can be employed for mitigating spurious correlations, which utilizes data groups defined by class and spuriousness labels. Following the same schema, GDRO [27] refines the training approach", + "bbox": [ + 217, + 763, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "SLIM: Spuriousness Mitigation with Minimal Human Annotations", + "bbox": [ + 287, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "by focusing on the worst-performing groups as determined by ERM loss. Aiming to reduce the dependency on comprehensive spuriousness labeling, several methods [1,6,17,21,30,32,41] have emerged to approximate group information using a reference model. While these approaches are technically sophisticated, they often fail to achieve the same level of performance as approaches informed by spuriousness labels. To balance the tradeoff between performance and cost, a further research direction reduces the requirement for extensive labeling by leveraging spuriousness labels in smaller validation datasets [14,22]. For example, SSA [22] adopts semi-supervised learning to predict group labels in training data according to information from validation, then applies GDRO to reduce spurious correlations. However, even with this reduced scale of validation sets, the labor-intensive nature of spuriousness labeling remains a barrier to scalability. Moreover, groups in these methods are formed based on human-defined criteria, risking a misalignment with the actual categories a model might discern. This drawback can hinder the true representation of spurious correlations and may lead to suboptimal mitigation outcomes.", + "bbox": [ + 212, + 146, + 787, + 387 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Spuriousness mitigation with human-in-the-loop. Integrating human insights has proven beneficial in mitigating model spuriousness, as evidenced by various human-involved strategies. Considering the valuable human effort, [28] provides an efficient solution by querying data samples for issue mitigation with domain experts in the loop, which overlooks potential issues in the broader data scope. Another option for avoiding redundant human annotation is to query global spurious concepts [31, 38] from knowledgeable experts, which can be ambiguous and fail to cover all possibilities. More recent approaches [10, 24, 26] raise concerns about the inconsistency of spuriousness labels, introducing alternatives such as human-generated attention maps or bounding boxes outlining core features. These human-generated references help align the models' attention with the actual features relevant to decision-making. However, producing such annotations is challenging, and these techniques are resource-intensive as they necessitate annotations across the entire dataset. In response, more recent research research [2, 40] simplifies the annotation process to binary feedback on model's attention correctness. While less demanding, its requirement to annotate all instances still poses challenges in real practices.", + "bbox": [ + 212, + 388, + 787, + 645 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Problem Formulation", + "text_level": 1, + "bbox": [ + 215, + 669, + 460, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We introduce the formal definition of the spuriousness mitigation problem [13]. Assume training data $\\mathcal{D} = \\{(x_i, y_i)\\}_{i=1}^n$ with input features $\\boldsymbol{x}_i \\in \\mathbb{R}^d$ and labels $y_i \\in [\\mathcal{L}]$ , where $\\mathcal{L} = \\{L_1, \\ldots, L_l\\}$ are the classes in $\\mathcal{D}$ . Machine learning models for classification are trained to minimize the empirical risk:", + "bbox": [ + 212, + 700, + 787, + 762 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\theta} ^ {*} \\in \\underset {\\boldsymbol {\\theta}} {\\arg \\min } \\mathbb {E} _ {(\\boldsymbol {x} _ {i}, y _ {i}) \\in \\mathcal {D}} [ l (f (\\boldsymbol {\\theta}, \\boldsymbol {x} _ {i}), y _ {i}) ], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 777, + 785, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pmb{\\theta}$ represent model parameters, $f(\\pmb{\\theta},\\pmb{x}_i)$ and $l(*)$ are the model outputs and the loss function, respectively.", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "X. Xuan et al.", + "bbox": [ + 271, + 114, + 367, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Groups. In spuriousness mitigation, we assume there exists a core feature set $\\mathcal{C} = \\{c_1,\\dots ,c_l\\}$ and a spurious feature set $\\mathcal{S} = \\{s_1,\\dots ,s_k\\}$ in $\\mathcal{D}$ . Each class $L_{i}$ is uniquely identified by a core feature $c_{i}$ , while spurious features $s_j$ can be found across different classes. We consider an instance in the dataset as having both a core feature $c_{i}$ and a spurious feature $s_j$ , and each instance belongs to a unique group $g_{c_i,s_j}$ . Core features are reliable class indicators in both training and testing sets, while spurious features with misleading co-occurrence with class in training may miss such correlations in the testing phase. If the model mistakenly links spurious features with classes in training, it may fail in testing when such spurious features are absent. This can lead to lower accuracy rates, especially for less common group variations that do not exhibit the spurious feature the model has learned to rely on.", + "bbox": [ + 212, + 146, + 787, + 327 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Worst-group accuracy. The vital evaluation of spuriousness mitigation focuses on the model's performance in the subgroup where it is least effective – the worst-group accuracy. It is calculated by:", + "bbox": [ + 212, + 328, + 785, + 375 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nA c c _ {w} = \\min _ {g \\in \\mathcal {G}} \\left(\\frac {1}{| g |} \\sum_ {\\left(x _ {i}, y _ {i}\\right) \\in g} \\mathbf {1} [ f \\left(x _ {i}\\right) = y _ {i} ]\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 356, + 386, + 785, + 419 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{G} = \\{\\cup_{c_i,s_j}g_{c_i,s_j}\\}$ is the union of subgroups defined by core feature $c_{i}$ and spurious feature $s_j$ , and $f(x_{i})$ is the label predicted by the model. This metric is crucial because it directly assesses how effectively a model can generalize across different subgroups, especially those that might be mostly underrepresented because of spurious correlations. It ensures that improvements in model performance are not just reflective of gains on the majority or less challenging subsets of the data.", + "bbox": [ + 212, + 429, + 787, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 Method", + "text_level": 1, + "bbox": [ + 215, + 561, + 331, + 578 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we detail the method behind SLIM. In alignment with related works requiring no spuriousness label, our approach also follows a two-stage procedure: inferring groups with an ERM-trained reference model $f_{r}$ , and then leveraging the inferred groups to train a robust model. As an approach oriented by data quality, SLIM framework comprises a three-phase data construction pipeline, as illustrated in Fig. 2. Firstly, we build an attention representation space to group instances with both similar data features and model attention, preparing for the data sampling and label propagation. Following this, we sample typical instances for human to evaluate model's attention correctness with yes-no feedback, then estimate neighboring instances' attention labels accordingly. In the last phase, we filter out data with wrong attention and leverage attention matrix to disentangle core and environment features. With constructed feature sets, we apply consistency-weighted data sampling to curate a feature-balanced subset, with a desired core-environment balanced distribution. Finally, SLIM framework is concluded by a training process with the curated subset, resulting in a model exhibiting robustness against spurious correlations.", + "bbox": [ + 212, + 597, + 787, + 843 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "SLIM: Spuriousness Mitigation with Minimal Human Annotations", + "bbox": [ + 287, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0cdc12ac114e2af3890c359276af5715a31132619d73866b44116d1be9cfcb99.jpg", + "image_caption": [ + "Fig. 2: Overview of SLIM framework, with a data construction pipeline consisting of three phases: (1) Attention Space Construction, creating a space with data features and model attention aligned locally; (2) Attention Annotation and Expansion, where instances sampled from the attention space are labeled by human for attention correctness, and labels are propagated to neighboring instances; and (3) Balanced Data Curation, which filters out instances with incorrect attention, and utilize attention-weighted $(F_A)$ and inverse-attention-weighted $(F_{\\bar{A}})$ feature vectors to create core and environment feature sets, forming subgroups to assemble a feature-balanced subset for training a spuriousness-robust model." + ], + "image_footnote": [], + "bbox": [ + 232, + 142, + 772, + 320 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1 Attention Space Construction", + "text_level": 1, + "bbox": [ + 215, + 460, + 508, + 477 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Considering the benefits of being aware of spuriousness information and drawbacks of the annotation cost, in our framework, we deploy a binary labeling scheme to simplify descriptive labeling tasks into yes-no feedback on model's attention correctness. Moreover, we aim to further facilitate this process by avoiding the necessity of exhaustive labeling. Our empirical observations indicate that a model would have consistent attention mechanisms on specific data subgroups. If such subgroups can be found, it would largely aid the labeling workload by enabling label estimations among neighbors.", + "bbox": [ + 212, + 489, + 787, + 611 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Driven by this, we aim to find a representation space where instances with similar attention outcomes (as indicated by highlighted image features) are closely grouped. However, our analysis reveals that conventional feature representation spaces formed directly from feature vectors do not fulfill this requirement - such spaces only maintain local consistency in data features but fail to capture the nuances of model attention (refer to Sec. 5.3).", + "bbox": [ + 212, + 612, + 787, + 702 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Thus, we have developed a process utilizing model attention in the latent space to support the establishment of such space, as illustrated in Fig. 3. For each image, we obtain its feature vector $F$ and model attribution $A$ in the latent space at first. Note that $A$ represents the model's attention focus and can be linearly interpolated to visualize an attention heatmap. We then use $A$ as a weight matrix for $F$ , producing an attention-weighted feature vector $F_{A}$ . Finally, with the dimensionality reduction method UMAP [20], we project vectors $F_{A}$ into a 2D space, namely the attention space, which ensures local consistency encompasses both data features and the model's attention mechanisms.", + "bbox": [ + 212, + 704, + 787, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "X. Xuan et al.", + "bbox": [ + 271, + 114, + 369, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a052730a8a30e45226e81fecedbdef87655946b17163286dbcb43f04a1db1175.jpg", + "image_caption": [ + "Fig. 3: Construction of the attention space. In the latent space, feature vector $F$ and model's attribution vector $A$ are extracted, representing input features and model's attention, representatively. By weighting $F$ with $A$ , the attention-weighted feature vector $F_{A}$ emphasizes features of model's top interest. All $F_{A}$ are then projected to form the attention space featuring locally consistent features and model attention." + ], + "image_footnote": [], + "bbox": [ + 232, + 143, + 772, + 209 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Note that for a fair comparison, in alignment with baseline methods, experiments in this paper involve the CNN model (ResNet) for classification and GradCAM for obtaining model attention. But our framework is fully applicable to (1) other vision model architectures such as vision transformers (ViT) [8] or Swin transformers [18], and (2) other heatmap-based model visualization techniques, such as RISE [23] or ScoreCAM [33].", + "bbox": [ + 212, + 297, + 784, + 387 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2 Attention Annotation and Expansion", + "text_level": 1, + "bbox": [ + 215, + 412, + 568, + 426 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Given that our attention space ensures locally consistent model attention, it allows for annotating a selectively sampled subset of instances and then estimating neighboring labels accordingly, thereby enabling efficient transfer of human expertise. The criteria for selecting these instances are twofold: (1) diversity, to encompass a broad covering of various data features and model attention patterns; (2) typicality, ensuring they are representative of their neighboring instances. To achieve this, we utilize the k-means clustering algorithm to partition our constructed attention space into $n$ clusters (diversity). We then select $n$ instances that are closest to each cluster's center (typicality). The choice of $n$ is aimed at optimizing the balance between minimizing intra-cluster variance in data features and model attention patterns, and keeping the annotation budget manageable. Our empirical analysis, detailed in Sec. 5.3, reveals that the intra-cluster consistencies increase with $n$ , but this trend levels off upon reaching a specific value of $n$ -the elbow point. Thus, selecting $n$ at this elbow point optimizes consistency without unnecessarily increasing the annotation workload.", + "bbox": [ + 212, + 439, + 784, + 666 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We gather binary yes-no feedback from human experts on the correctness of the model's attention for these $n$ instances, feeding this feedback along with our attention space embedding into the label spreading algorithm. This algorithm leverages data similarity and density to extend attention labels to previously unlabeled data [43]. Ultimately, this process yields a probability score for correct attention for each instance across the dataset.", + "bbox": [ + 212, + 667, + 785, + 755 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3 Balanced Data Curation", + "text_level": 1, + "bbox": [ + 215, + 782, + 464, + 796 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Following Sec. 4.2 that provides attention accuracy scores, we proceed to construct our data subset to support model's unbiased learning, as described below:", + "bbox": [ + 212, + 809, + 784, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "SLIM: Spuriousness Mitigation with Minimal Human Annotations", + "bbox": [ + 287, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Data screening. Inspired by works [3,7,12], our theoretical analysis further supports the premise that data receiving correct attention from a reference model are more likely to enhance the learning of core features in a robust model within a more balanced data distribution, as substantiated in Appx. A.2. Furthermore, our study indicates that even a reference model with inherent biases has the potential to accurately identify core features for a data subset. Such correctly captured features typically exhibit diversity, ensuring good coverage across the spectrum of the core feature set. This understanding informs our approach to curating a data subset that: (1) selects instances based on accurate model attention, and (2) ensures a balanced representation of core features across various environment features. Thus we first apply data screening to filter out instances with wrong attention according to attention accuracy scores.", + "bbox": [ + 212, + 146, + 787, + 327 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Disentangling core and environment features. After screening, the retained data subset mainly comprises instances with correct attention. This is reflected by attention masks that highlight core features, meaning that the corresponding attribution vector $A$ in the latent space also highlights core features of feature vector $F$ . Moreover, existing studies have found that a model's feature vectors are representative of most input features, regardless of whether they are used for decision-making [14, 37]. Therefore, utilizing $A$ as the weight matrix, we disentangle core and environment features in the latent space — weighting $F$ by attention masks $A$ and inverse attention masks $\\bar{A}$ . Subsequently, k-means clustering is applied to partition all attention-weighted feature vectors $F_{A}$ , and inverse-attention-weighted feature vectors $F_{\\bar{A}}$ , respectively, facilitating the formation of representative subgroups for the core feature set $(\\hat{C} = \\hat{c}_1,\\dots ,\\hat{c}_N)$ and the environment feature set $(\\mathcal{E} = e_1,\\ldots ,e_M)$ . The number of clusters, determined through the elbow method, culminates in $M\\times N$ subgroups $g_{\\hat{c}_i,e_j}$ , each characterized by a distinct combination of core and environment features $(\\hat{c}_i,e_j)$ . Data sampling. Given the unequal distribution of core and environment features, the internal consistency across each constructed subgroup $g_{\\hat{c}_i,e_j}$ varies. Our sampling strategy aims to mitigate over-representation of particular features to prevent overfitting, while ensuring broad coverage of diverse core and environment features. We assess cluster consistency as a metric for determining the sample size from each cluster. For a cluster $\\hat{c}_i$ in the core feature space, denote the distance between an instance $x$ and the cluster center as $d_{\\hat{c}_i}(x)$ . The cluster consistency is defined as the reciprocal of the average distance per cluster:", + "bbox": [ + 212, + 329, + 787, + 680 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\rho_ {\\hat {c} _ {i}} = \\frac {1}{\\frac {1}{| \\hat {c} _ {i} |} \\sum_ {x \\in \\hat {c} _ {i}} d _ {\\hat {c} _ {i}} (x)}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 416, + 694, + 785, + 729 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A lower $\\rho_{\\hat{c}_i}$ signals less consistent features, necessitating a higher sample weight for adequately representing them. A cluster's sampling weight is defined as $w_{\\hat{c}_i} = \\frac{1}{\\rho_{\\hat{c}_i}}$ and the number of samples to be drawn from a core cluster $\\hat{c}_i$ is:", + "bbox": [ + 215, + 739, + 787, + 790 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nn _ {\\hat {c} _ {i}} = N \\cdot \\frac {1 / \\rho_ {\\hat {c} _ {i}}}{\\sum_ {j = 1} ^ {| \\hat {C} |} 1 / \\rho_ {\\hat {c} _ {j}}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 419, + 804, + 785, + 844 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "X. Xuan et al.", + "bbox": [ + 271, + 114, + 367, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $N$ is the sampling budget. Moving to the environment feature space, the $n_{\\hat{c}_i}$ instances from a core feature cluster $\\hat{c}_i$ may distribute across multiple environment clusters $\\{e_1,\\dots ,e_M\\}$ and formulate subgroups $\\{g_{\\hat{c}_i,e_1},\\ldots ,g_{\\hat{c}_i,e_M}\\}$ . A similar methodology is employed for a certain group $g_{\\hat{c}_i,e_i}$ . The average distance for $g_{\\hat{c}_i,e_i}$ in the environment representation space is:", + "bbox": [ + 214, + 146, + 787, + 223 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\rho_ {\\hat {c} _ {i}, e _ {i}} = \\frac {1}{\\frac {1}{| \\hat {c} _ {i} , e _ {i} |} \\sum_ {x \\in (\\hat {c} _ {i} , e _ {i})} d _ {e _ {i}} (x)}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 234, + 785, + 271 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Similarly, we have the sampling weight $w_{\\hat{c}_i,e_i} = \\frac{1}{\\rho_{\\hat{c}_i,e_i}}$ and the number of samples to be drawn from a certain group $g_{\\hat{c}_i,e_i}$ is finally defined by:", + "bbox": [ + 214, + 280, + 785, + 316 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nn _ {\\hat {c} _ {i}, e _ {i}} = n _ {\\hat {c} _ {i}} \\cdot \\frac {w _ {\\hat {c} _ {i} , e _ {i}}}{\\sum_ {j = 1} ^ {| \\mathcal {E} (\\hat {c} _ {i}) |} w _ {\\hat {c} _ {j} , e _ {i}}} = N \\cdot \\frac {1 / \\rho_ {\\hat {c} _ {i}}}{\\sum_ {j = 1} ^ {| \\hat {\\mathcal {C}} |} 1 / \\rho_ {\\hat {c} _ {j}}} \\cdot \\frac {1 / \\rho_ {\\hat {c} _ {i} , e _ {i}}}{\\sum_ {j = 1} ^ {| \\mathcal {E} (\\hat {c} _ {i}) |} 1 / \\rho_ {\\hat {c} _ {i} , e _ {j}}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 272, + 327, + 785, + 367 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $|\\hat{\\mathcal{C}}|$ denotes the number of core clusters and $|\\mathcal{E}(\\hat{c}_i)|$ denotes the number of subgroups formulated by instances from cluster $\\hat{c}_i$ . To summary, our sampling strategy takes feature consistencies and sampling budget into consideration, and ensures fair representation of different core-environment feature combinations across the feature spaces.", + "bbox": [ + 214, + 378, + 787, + 455 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.4 Training", + "text_level": 1, + "bbox": [ + 215, + 477, + 333, + 493 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In our pursuit to mitigate spurious correlations, we emphasize the pivotal role of data quality by ensuring a balanced distribution of core features across various environmental contexts within our dataset. Given this focus, we employ a basic training approach for our model to learn from this well-prepared dataset.", + "bbox": [ + 212, + 502, + 785, + 563 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We apply our data curation pipeline separately to the training $(SLIM_{Tr})$ and validation $(SLIM_{Val})$ sets. For the data constructed from the training set, we adopt the conventional ERM. For the dataset derived from the validation set, we employ the training technique presented by DFR [14]. It involves retraining only the final linear layer of a classification model that was initially trained using ERM. Note that our approach diverges from DFR practices by utilizing our curated dataset instead of a metagroup-balanced dataset.", + "bbox": [ + 214, + 563, + 787, + 669 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 691, + 375, + 709 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 215, + 723, + 428, + 739 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Datasets. Our study explores a range of datasets with inherent spurious correlations to challenge image classification models, including: (1) Waterbirds [27]. Contrasting landbirds and waterbirds against congruent/incongruent backgrounds, this dataset tests models' ability to focus on relevant features for bird type classification. (2) CelebA [19]. A comprehensive celebrity image dataset where the focus is on distinguishing hair color (blond vs. non-blond) to examine", + "bbox": [ + 214, + 750, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "SLIM: Spuriousness Mitigation with Minimal Human Annotations", + "bbox": [ + 287, + 114, + 732, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 127 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "biases in associating hair color with gender. (3) ISIC [5]. Aiming to differentiate benign from malignant melanoma in dermoscopic images, we target colorful patches as spurious features, aligning with previous studies [25]. (4) NICO. Derived from $\\mathrm{NICO}++$ [42], this dataset features various object categories in shifted contexts to probe spurious correlations. (5) ImageNet-9 [35]. This dataset is crafted from ImageNet to test models' robustness against background variations. Detailed information is available in Appx. A.4.", + "bbox": [ + 212, + 146, + 787, + 253 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Baselines. We benchmark SLIM against a spectrum of SOTA spuriousness mitigation techniques, categorized as follows: (1) Techniques not reliant on spuriousness labels: EIIL [6], PGI [1], GEORGE [30], LfF [21], CIM [32], JTT [17], and CNC [41]; (2) Techniques requiring additional information such as predefined spuriousness concepts: DISC [34]; (3) Methods utilizing spuriousness labels from validation sets: SSA [22] and DFR [14]; (4) Approaches utilizing spuriousness labels from training sets: GDRO [27] and PDE [7].", + "bbox": [ + 212, + 260, + 787, + 367 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation. Aligning with baselines, our experiments involve CNN models with ResNet50 architecture and GradCAM for visualizing model attention. For the capability of performance enhancement, we evaluate the worst-group and average accuracy. Additionally, we consider the necessity for extra labels and the additional (add.) training cost to evaluate computational efficiency. Note that the add. training cost is aligned with [13, 39], measuring add. data amounts required for second-stage training, where 1 means the amount of data equals the training set size. Furthermore, we examine the attention precision to evaluate whether model attention is corrected. All quantification experiments are based on five independent runs with different random seeds.", + "bbox": [ + 212, + 375, + 787, + 526 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "More experimental results with other architecture (ViT) and datasets (MetaShift [16] and FMoW [4]) are provided in Appx. A.8.", + "bbox": [ + 214, + 534, + 785, + 566 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/0589e3ce026a905e800fcca963926a64b2e33b8f1b8a8e62a3261ad98aede71a.jpg", + "table_caption": [ + "Table 2: Worst-group and average accuracy (%) comparison on binary classification tasks, highlighting best and second-best performances. $\\spadesuit$ and $\\diamond$ indicate methods requiring spuriousness concepts and attention labels, respectively." + ], + "table_footnote": [], + "table_body": "
MethodSpuriousness labelAdd. training costWaterbirdsCelebAISIC
WorstAvgWorstAvgWorstAvg
ERMX0×62.6±0.397.3±1.047.7±2.194.9±0.356.7±2.279.0±1.3
EIILX0×83.5±2.894.2±1.381.7±0.885.7±0.172.3±3.986.4±1.7
PGIX0×79.5±1.995.5±0.885.3±0.387.3±0.167.4±1.985.7±1.1
GEORGEX1×76.2±2.095.7±0.554.9±1.994.6±0.258.8±2.183.2±1.4
LfFX1×77.3±2.391.4±1.776.8±1.284.5±0.772.1±2.185.4±0.5
CIMX1×77.8±1.594.7±1.183.8±1.090.5±0.671.7±1.685.9±1.1
JTTX5×-6×83.1±3.590.6±0.381.5±1.788.1±0.328.1±4.386.2±0.3
CNCX2×-12×88.5±0.390.9±0.188.8±0.989.9±0.568.2±0.284.6±0.2
DISCX◇1×88.7±0.793.8±0.785.7±1.189.9±0.773.4±0.484.7±0.4
SLIMTrX◇0.1×-0.2×89.1±0.691.7±0.589.4±0.390.1±0.279.1±0.588.5±0.4
SLIMValX◇0.05×-0.3×91.0±0.493.8±0.489.4±0.590.7±0.274.0±0.285.4±0.2
SSAvalidation1.5×-5×89.0±0.692.2±0.989.8±1.392.8±0.161.7±0.777.7±1.0
DFRvalidation0.05×-0.3×89.6±0.792.1±0.287.3±1.090.2±0.871.9±0.784.3±0.3
GDROtraining0×89.9±0.692.6±0.188.9±1.393.9±0.166.8±3.580.5±1.9
PDEtraining0×85.9±2.790.5±1.885.0±4.092.3±0.866.4±2.382.2±1.3
", + "bbox": [ + 215, + 637, + 787, + 825 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "X. Xuan et al.", + "bbox": [ + 271, + 114, + 367, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/6d6062decb3ee33841b4b96c986a76fad480734d075a79dcc125dd4b2f06b8dc.jpg", + "table_caption": [ + "Table 3: Worst-group and average accuracy (%) comparison on multi-classification tasks, highlighting best and second-best results. $\\diamond$ indicates methods requiring attention labels. - represents inapplicable methods because such methods require spuriousness labels while the dataset does not provide them." + ], + "table_footnote": [], + "table_body": "
MethodSpuriousness labelAdd. training costNICOImageNet9
WorstAvgWorstAvg
ERMX0×11.9±0.270.9±1.255.3±2.075.6±1.1
CNCX5×-12×72.6±1.789.5±0.658.9±0.674.3±0.2
SLIMTrX◇0.1×-0.2×78.0±0.892.1±0.168.1±0.374.5±0.2
SLIMValX◇0.1×-0.2×73.8±0.490.7±0.264.6±0.474.2±0.3
DFRvalidation0.4×70.9±2.185.8±0.6--
GDROtraining0×73.3±1.090.3±0.5--
", + "bbox": [ + 272, + 210, + 730, + 305 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/1499461a1571aafc5beaafbe54dad50f3a4e5491894b856e4930a0a140f08a21.jpg", + "table_caption": [ + "Table 4: SLIM's primary cost, including the data amount required for attention annotation (Att.), and the constructed data size for model training (Training)." + ], + "table_footnote": [], + "table_body": "
MethodWaterbirdCelebAISICNICOImageNet 9
Att.TrainingAtt.TrainingAtt.TrainingAtt.TrainingAtt.Training
SLIMTr120 (2.5%)0.1×200 (0.12%)0.1×120 (0.7%)0.2×40 (1.8%)0.2×180 (0.4%)0.1×
SLIMVal40 (3%)0.1×70 (0.35%)0.05×40 (2%)0.3×20 (2.7%)0.2×50 (1%)0.1×
", + "bbox": [ + 217, + 345, + 785, + 391 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.2 Spuriousness Mitigation Performance", + "text_level": 1, + "bbox": [ + 214, + 393, + 571, + 407 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In this section, we present the performance evaluation by comparing different methods across multiple binary and multiclass classification datasets mentioned in Sec. 5.1. Key findings are demonstrated below.", + "bbox": [ + 212, + 417, + 784, + 460 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Superior worst-group performance. The metric of worst-group accuracy $(Acc_w)$ is pivotal for spuriousness mitigation methods (refer to Sec. 3), and the results in Tables 2 and 3 showcase SLIM's superior capability in this regard. Across various classification tasks, SLIM consistently outperforms competing methods in achieving the highest $Acc_w$ while maintaining high average accuracy. Specifically, in binary classification scenarios (Table 2), SLIM leads in performance for Waterbirds and ISIC and secures a close second in CelebA. In multi-classification tasks (Table 3), $SLIM_{Tr}$ and $SLIM_{Val}$ rank as the top performers across the board, underscoring the effectiveness of our approach in spuriousness mitigation.", + "bbox": [ + 212, + 462, + 785, + 612 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Cost-efficient spuriousness mitigation. Methods in Tables 2 and 3 are divided into four groups from top to bottom, as per the categorization established in Sec. 5.1 according to the computational cost. The \"Add. training cost\" columns denote the additional training cost for mitigating spuriousness apart from the initial training, such as additional learning phases (CNC, SSA) or data augmentation (JTT). Results in the table substantiate that methods with higher costs (higher Add. training costs or more spuriousness labels) tend to outperform others, aligning with prior research findings. Nevertheless, among the well-performed methods, it is worth mentioning that $SLIM$ is more cost-efficient: $SLIM_{Tr}$ used 0.1-0.2 proportion of training sets and $SLIM_{Val}$ used 0.05-0.3 proportion of validation sets for additional training, which is the same as DFR and much smaller than others. Moreover, $SLIM$ requires labels for attention correctness, which is more cost-efficient to obtain than spuriousness labels (refer to Sec. 5.3 for more details). Table 4 presents $SLIM$ 's primary cost, including the training data amount for mitigating spuriousness, which is the detailed version of \"Add. train", + "bbox": [ + 212, + 613, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "SLIM: Spuriousness Mitigation with Minimal Human Annotations", + "bbox": [ + 287, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/339c57f4cfb0e174ef20e2d2c3c0ebc8ecbdeda076aa133cd9baeb2516005855.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 228, + 186, + 588, + 296 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/e398f742806b00861e83c37042025ed288352d685411b4dc361db3c8cd56c557.jpg", + "table_caption": [ + "Table 5: Attention accuracy evaluation. (Left) GradCAM visualization of model attention. (Right) Worst group and average attention accuracy of selected methods." + ], + "table_footnote": [], + "table_body": "
WaterbirdsAIoU
WorstAvg
ERM.25±.05.47±.03
CNC.48±.05.49±.04
GDRO.49±.03.52±.03
PDE.50±.02.52±.03
SLIMTr.53±.02.61±.02
ImageNet 9AIoU
WorstAvg
ERM.33±.07.55±.03
CNC.44±.03.61±.03
SLIMTr.58±.04.68±.03
", + "bbox": [ + 632, + 184, + 767, + 296 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "ing cost\" in Tables 2 and 3, and the data amount requiring attention labels (Att.). Notably, our approach only requires no more than $3\\%$ of the corresponding data split (train or validation) for attention-label annotation, further proving the cost-efficiency of SLIM.", + "bbox": [ + 215, + 297, + 782, + 358 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Enhanced attention accuracy. Beyond classification accuracy, we further evaluate the performance of spuriousness mitigation methods via attention accuracy. Specifically, for models resulting from each method, we randomly sample data instances to visualize model attention, and measure attention accuracy scores in both worst-group and dataset levels, consistent with our evaluation of classification accuracy. Align with the related work [40], we employ the Adjusted Intersection-over-Union (AIoU) score as our attention accuracy measurement, which is an adjusted IoU score with a threshold, aiming at measuring the overlap between model's highest attributed region and the ground-truth bounding box. Details related to AIoU are given in Appx. A.7. Table 5 presents the evaluation results. The left GradCAM examples showcase SLIM's capability of correcting model's wrong attention, and the right table compares SLIM with SOTA methods, where we can find our method secures the best performance in terms of both worst-group and average attention accuracy.", + "bbox": [ + 212, + 359, + 787, + 570 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.3 Attention Labeling Performance", + "text_level": 1, + "bbox": [ + 215, + 590, + 529, + 606 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To compare the processes of annotating spuriousness versus attention correctness, we employed crowdsourcing tasks using the Waterbirds and NICO datasets, as both of which come with ground-truth spuriousness labels. For this study, we selected a random set of 120 images from each dataset and established two separate tasks: one for spuriousness labeling and one for attention correctness labeling. For each task, we recruited 60 workers independently from the Prolific platform to prevent learning biases from cross-task participation. We meticulously recorded the time taken to label each image and reviewed all submissions to exclude any outliers with unusually long or short annotation times. Ensuring ethical research standards, our study avoided collecting personally identifiable information and excluded any potentially offensive content. Detailed instructions provided to the crowdsourcing participants can be found in Appx. A.6.", + "bbox": [ + 212, + 613, + 787, + 794 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Evaluation Consistency. Table 6 presents metrics for this study. While the annotation time for Waterbirds remains similar across tasks, attention correctness for NICO is annotated more quickly than the other. This difference may stem", + "bbox": [ + 212, + 795, + 787, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "X. Xuan et al.", + "bbox": [ + 271, + 114, + 367, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/6c0101a6dc68b2589b4520a97959d1928b1e9456a3b5f9fafe0a88cf29cb0947.jpg", + "table_caption": [ + "Table 6: Comparison of crowdsourced labeling tasks: spuriousness vs. attention correctness." + ], + "table_footnote": [], + "table_body": "
WaterbirdsAvg. TimeAvg. Consis.
Spuriousness4.9±2.3 sec81.9±14.8%
Att.4.5±1.4 sec94.6±10.5%
NICOAvg. TimeAvg. Consis.
Spuriousness7.9±3.8 sec70.1±19.1%
Att.5.3±1.6 sec95.1±6.1%
", + "bbox": [ + 220, + 205, + 442, + 280 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/0e4c5b3e9219ea3099038f43ce33447737c73f5a34bbff9735ddbce52d86de9b.jpg", + "image_caption": [ + "Fig. 4: Comparison of neighbor consistency over original feature space and our attention space on Waterbirds, w.r.t. model attention (left) and data features (right)." + ], + "image_footnote": [], + "bbox": [ + 454, + 143, + 753, + 220 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "from NICO's broader array of spuriousness choices versus the binary attention annotations. To measure the annotation consistency—whether a single receives divergent labels—we use the percentage of the most frequent label as a consensus measure. A higher percentage suggests greater consistency. Our findings indicate that attention correctness annotations exhibit a notably higher consistency, suggesting that it is less ambiguous to determine correct attention than to identify spurious features.", + "bbox": [ + 212, + 285, + 785, + 388 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Attention Consistency. $SLIM$ requires spreading labels in a space where attention patterns are consistently similar in nearby areas. To evaluate its feasibility, we compared how similar neighbors are in the original feature space versus the attention space created by $SLIM$ . We used k-means clustering with varying numbers of clusters to measure the intra-cluster similarity for both the model's attention matrices and the feature vectors. A higher similarity score means the space has more consistent patterns, which is essential for the accurate spreading of labels. As Fig. 4 shows, our attention space significantly improves consistency in the model's attention, while also maintaining a good level of data feature consistency. Increasing the number of clusters generally increases consistency, but there is a point where adding more clusters does not lead to further gains, as shown by the curve flattening at $k = 120$ in the left plot.", + "bbox": [ + 212, + 390, + 785, + 571 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In summary, we have demonstrated that the annotation module of SLIM is less time-consuming and yields more consistent labeling feedback. Additionally, the attention space engineered by SLIM establishes a solid foundation for propagating attention labels, considering the consistent model attention observed among neighboring instances.", + "bbox": [ + 212, + 571, + 785, + 647 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.4 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 666, + 390, + 681 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We investigate two critical factors of SLIM, which are detailed below. Other factors such as the dimension of the attention space are discussed in Appx. A.8. Influence of data amount for attention annotations. We change the data amount for attention annotations to investigate their effect on performance, keeping other variables constant. Table 7 illustrates that both the worst-group and average accuracy tend to rise with more annotations, while such incremental benefit stabilizes after reaching a certain point, helping us identify an optimal number of annotations that balance performance with annotation cost.", + "bbox": [ + 212, + 688, + 785, + 809 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Influence of training set size. This size refers to the size of the data subset constructed in SLIM's final phase for training a robust model. By only varying", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "SLIM: Spuriousness Mitigation with Minimal Human Annotations", + "bbox": [ + 287, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/15fd323e671d15385071fe992d8ea68532b7a4900d3076f2914670e0447bdeaf.jpg", + "table_caption": [ + "Table 7: Performance variation with different attention annotation amounts in SLIM. Marks in table: best and second best performance; the amounts used in Table 2." + ], + "table_footnote": [], + "table_body": "
MethodWaterbirdsCelebAISIC
AmountWorstAvgAmountWorstAvgAmountWorstAvg
SLIMTr6078.2±0.586.0±0.310086.4±0.290.2±0.16075.3±0.286.6±0.9
9085.9±0.989.3±0.315087.8±0.491.1±0.79074.4±0.487.4±0.7
12089.1±0.691.7±0.520089.4±0.490.1±0.212079.1±0.588.5±0.4
15089.4±0.691.7±0.225089.4±0.290.9±0.115079.9±0.888.8±0.7
18089.3±0.391.0±1.530089.5±0.191.2±0.218079.8±0.388.6±0.3
SLIMVal1078.9±8.786.5±5.14084.7±0.487.4±0.31070.5±0.583.3±1.2
2589.2±0.693.4±1.05587.6±0.390.4±0.62572.7±0.185.2±0.8
4091.0±0.493.8±0.47089.4±0.490.7±0.24074.0±0.285.4±0.2
5591.5±0.393.8±0.18589.7±0.190.5±0.45574.0±0.285.9±0.5
7090.7±0.394.0±0.310089.7±0.290.5±0.27074.1±0.285.8±0.2
", + "bbox": [ + 256, + 184, + 746, + 324 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/82001d851d184b45d5c1f91e7eaf2c3237add32877152dad009ab10af44ce692.jpg", + "table_caption": [ + "Table 8: Performance variation with different training sizes. $SLIM_{Tr}$ and $SLIM_{Val}$ are trained on different proportions (Prop) of training sets and validation sets, respectively." + ], + "table_footnote": [], + "table_body": "
MethodWaterbirdsCelebAISIC
Prop%WorstAvgProp%WorstAvgProp%WorstAvg
SLIMTr582.4±1.083.6±1.3588.9±0.190.0±0.11577.6±2.187.1±1.2
1089.1±0.691.7±0.51089.4±0.390.1±0.22079.1±0.588.5±0.4
1282.3±1.387.8±0.31589.4±0.491.0±0.42577.4±1.588.2±1.2
1583.5±2.789.6±0.52088.9±0.190.8±0.33078.3±1.588.3±0.9
SLIMVal584.2±3.689.2±2.6385.7±1.388.5±1.12568.2±2.181.9±0.8
1091.0±0.493.8±0.4589.4±0.490.7±0.23074.0±0.285.4±0.2
1287.6±0.693.9±1.31088.3±1.490.3±0.53572.5±0.585.3±0.1
1586.8±2.094.6±0.51587.2±0.490.3±0.44070.9±1.285.4±3.5
", + "bbox": [ + 256, + 367, + 746, + 484 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "this parameter (Table 8), we find that the optimal size is inherently linked to the size of our constructed subgroups, $|g_{\\hat{c}_i,e_i}|$ . An optimal size is reached when data from various groups are sufficiently sampled without redundant oversampling. These findings guide the hyperparameter selection in this phase, ensuring an effective cost allocation.", + "bbox": [ + 215, + 489, + 784, + 563 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 217, + 592, + 356, + 607 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we introduce SLIM, a novel framework for spuriousness mitigating in deep learning. With minimal human annotation effort, SLIM provides a data construction pipeline to create a feature-balanced subset for training a robust model. Extensive evaluation demonstrates the superior cost-effectiveness of SLIM, which matches or outperforms state-of-the-art methods with far less costs. The success of SLIM underscores the potential of data quality and efficient human supervision in the development of robust models. We believe that SLIM will inspire further research in this direction, ultimately leading to more cost-effective machine learning systems for reliable AI.", + "bbox": [ + 215, + 627, + 784, + 762 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements. This research is supported in part by Bosch Research and the National Institute of Health with grants P41-EB032840 and R01CA270454. H.-T. Lin is partially supported by the National Taiwan University Center for Data Intelligence via NTU-112L900901 and the Ministry of Science and Technology in Taiwan via MOST 112-2628-E-002-030.", + "bbox": [ + 215, + 763, + 784, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "X. Xuan et al.", + "bbox": [ + 271, + 114, + 367, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Ahmed, F., Bengio, Y., Van Seijen, H., Courville, A.: Systematic generalisation with group invariant predictions. In: International Conference on Learning Representations (2020)", + "2. Bontempelli, A., Teso, S., Tentori, K., Giunchiglia, F., Passerini, A.: Concept-level debugging of part-prototype networks. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=oiwXwPDTyNk", + "3. Chen, Z., Deng, Y., Wu, Y., Gu, Q., Li, Y.: Towards understanding the mixture-of-experts layer in deep learning. Advances in neural information processing systems 35, 23049-23062 (2022)", + "4. Christie, G., Fendley, N., Wilson, J., Mukherjee, R.: Functional map of the world. In: CVPR (2018)", + "5. Codella, N.C.F., Gutman, D., Celebi, M.E., Helba, B., Marchetti, M.A., Dusza, S.W., Kalloo, A., Liopyris, K., Mishra, N., Kittler, H., Halpern, A.: Skin lesion analysis toward melanoma detection: A challenge at the 2017 international symposium on biomedical imaging (isbi), hosted by the international skin imaging collaboration (isic). In: 2018 IEEE 15th International Symposium on Biomedical Imaging (ISBI 2018). pp. 168-172 (2018). https://doi.org/10.1109/ISBI.2018.8363547", + "6. Creager, E., Jacobsen, J.H., Zemel, R.: Environment inference for invariant learning. In: International Conference on Machine Learning. pp. 2189-2200. PMLR (2021)", + "7. Deng, Y., Yang, Y., Mirzasoleiman, B., Gu, Q.: Robust learning with progressive data expansion against spurious correlation. In: Thirty-seventh Conference on Neural Information Processing Systems (2023), https://openreview.net/forum?id=9QEVJ9qm46", + "8. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)", + "9. Gao, Y., Sun, T.S., Zhao, L., Hong, S.R.: Aligning eyes between humans and deep neural network through interactive attention alignment. Proc. ACM Hum-. Comput. Interact. 6(CSCW2) (nov 2022). https://doi.org/10.1145/3555590, https://doi.org/10.1145/3555590", + "0. Gao, Y., Sun, T.S., Zhao, L., Hong, S.R.: Aligning eyes between humans and deep neural network through interactive attention alignment. Proceedings of the ACM on Human-Computer Interaction 6(CSCW2), 1-28 (2022)", + "1. He, H., Garcia, E.A.: Learning from imbalanced data. IEEE Transactions on knowledge and data engineering 21(9), 1263-1284 (2009)", + "2. Jelassi, S., Li, Y.: Towards understanding how momentum improves generalization in deep learning. In: International Conference on Machine Learning. pp. 9965-10040. PMLR (2022)", + "3. Joshi, S., Yang, Y., Xue, Y., Yang, W., Mirzasoleiman, B.: Towards mitigating spurious correlations in the wild: A benchmark & a more realistic dataset. arXiv preprint arXiv:2306.11957 (2023)", + "4. Kirichenko, P., Izmailov, P., Wilson, A.G.: Last layer re-training is sufficient for robustness to spurious correlations. arXiv preprint arXiv:2204.02937 (2022)", + "5. Lee, S., Payani, A., Chau, D.H.P.: Towards mitigating spurious correlations in image classifiers with simple yes-no feedback. In: AI and HCI Workshop at the 40th International Conference on Machine Learning (2023)" + ], + "bbox": [ + 225, + 172, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "SLIM: Spuriousness Mitigation with Minimal Human Annotations", + "bbox": [ + 287, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Liang, W., Zou, J.: Metashift: A dataset of datasets for evaluating contextual distribution shifts and training conflicts. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=MTex8qKavoS", + "17. Liu, E.Z., Haghgoo, B., Chen, A.S., Raghunathan, A., Koh, P.W., Sagawa, S., Liang, P., Finn, C.: Just train twice: Improving group robustness without training group information. In: International Conference on Machine Learning. pp. 6781-6792. PMLR (2021)", + "18. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021)", + "19. Liu, Z., Luo, P., Wang, X., Tang, X.: Deep learning face attributes in the wild. In: Proceedings of the IEEE international conference on computer vision. pp. 3730-3738 (2015)", + "20. McInnes, L., Healy, J., Melville, J.: Umap: Uniform manifold approximation and projection for dimension reduction. arXiv preprint arXiv:1802.03426 (2018)", + "21. Nam, J., Cha, H., Ahn, S., Lee, J., Shin, J.: Learning from failure: De-biasing classifier from biased classifier. Advances in Neural Information Processing Systems 33, 20673-20684 (2020)", + "22. Nam, J., Kim, J., Lee, J., Shin, J.: Spread spurious attribute: Improving worst-group accuracy with spurious attribute estimation. arXiv preprint arXiv:2204.02070 (2022)", + "23. Petsiuk, V., Das, A., Saenko, K.: Rise: Randomized input sampling for explanation of black-box models. arXiv preprint arXiv:1806.07421 (2018)", + "24. Rao, S., Böhle, M., Parchami-Araghi, A., Schiele, B.: Using explanations to guide models. arXiv preprint arXiv:2303.11932 (2023)", + "25. Rieger, L., Singh, C., Murdoch, W.J., Yu, B.: Interpretations are useful: Penalizing explanations to align neural networks with prior knowledge. In: Proceedings of the 37th International Conference on Machine Learning. ICML'20, JMLR.org (2020)", + "26. Ross, A.S., Hughes, M.C., Doshi-Velez, F.: Right for the right reasons: Training differentiable models by constraining their explanations. In: Proceedings of the Twenty-Sixth International Joint Conference on Artificial Intelligence, IJCAI-17. pp. 2662-2670 (2017). https://doi.org/10.24963/ijcai.2017/371, https://doi.org/10.24963/ijcai.2017/371", + "27. Sagawa*, S., Koh*, P.W., Hashimoto, T.B., Liang, P.: Distributionally robust neural networks. In: International Conference on Learning Representations (2020), https://openreview.net/forum?id=ryxGuJrFvS", + "28. Schramowski, P., Stammer, W., Teso, S., Brugger, A., Herbert, F., Shao, X., Luigs, H.G., Mahlein, A.K., Kersting, K.: Making deep neural networks right for the right scientific reasons by interacting with their explanations. Nature Machine Intelligence 2(8), 476-486 (2020)", + "29. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Gradcam: Visual explanations from deep networks via gradient-based localization. In: Proceedings of the IEEE international conference on computer vision. pp. 618-626 (2017)", + "30. Sohoni, N., Dunnmon, J., Angus, G., Gu, A., Ré, C.: No subclass left behind: Fine-grained robustness in coarse-grained classification problems. Advances in Neural Information Processing Systems 33, 19339-19352 (2020)", + "31. Srivastava, M., Hashimoto, T., Liang, P.: Robustness to spurious correlations via human annotations. In: International Conference on Machine Learning. pp. 9109-9119. PMLR (2020)" + ], + "bbox": [ + 215, + 147, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "X. Xuan et al.", + "bbox": [ + 271, + 114, + 367, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "32. Taghanaki, S.A., Choi, K., Khasahmadi, A.H., Goyal, A.: Robust representation learning via perceptual similarity metrics. In: International Conference on Machine Learning. pp. 10043-10053. PMLR (2021)", + "33. Wang, H., Wang, Z., Du, M., Yang, F., Zhang, Z., Ding, S., Mardziel, P., Hu, X.: Score-cam: Score-weighted visual explanations for convolutional neural networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 24-25 (2020)", + "34. Wu, S., Yuksekgonul, M., Zhang, L., Zou, J.: Discover and cure: Concept-aware mitigation of spurious correlation. In: Proceedings of the 40th International Conference on Machine Learning. ICML'23, JMLR.org (2023)", + "35. Xiao, K., Engstrom, L., Ilyas, A., Madry, A.: Noise or signal: The role of image backgrounds in object recognition. ArXiv preprint arXiv:2006.09994 (2020)", + "36. Xuan, X., Ono, J.P., Gou, L., Ma, K.L., Ren, L.: Attributionscanner: A visual analytics system for metadata-free data-slicing based model validation. arXiv preprint arXiv:2401.06462 (2024)", + "37. Xue, Y., Payani, A., Yang, Y., Mirzasoleiman, B.: Eliminating spurious correlations from pre-trained models via data mixing. arXiv preprint arXiv:2305.14521 (2023)", + "38. Yan, S., Yu, Z., Zhang, X., Mahapatra, D., Chandra, S.S., Janda, M., Soyer, P., Ge, Z.: Towards trustable skin cancer diagnosis via rewriting model's decision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11568-11577 (2023)", + "39. Yang, Y., Gan, E., Karolina Dziugaite, G., Mirzasoleiman, B.: Identifying spurious biases early in training through the lens of simplicity bias. In: Proceedings of The 27th International Conference on Artificial Intelligence and Statistics. Proceedings of Machine Learning Research, vol. 238, pp. 2953-2961. PMLR (02-04 May 2024)", + "40. Yang, Y., Nushi, B., Palangi, H., Mirzasoleiman, B.: Mitigating spurious correlations in multi-modal models during fine-tuning. In: Proceedings of the 40th International Conference on Machine Learning. ICML'23, JMLR.org (2023)", + "41. Zhang, M., Sohoni, N.S., Zhang, H.R., Finn, C., Ré, C.: Correct-n-contrast: A contrastive approach for improving robustness to spurious correlations. arXiv preprint arXiv:2203.01517 (2022)", + "42. Zhang, X., He, Y., Xu, R., Yu, H., Shen, Z., Cui, P.: Nico++: Towards better benchmarking for domain generalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16036-16047 (2023)", + "43. Zhou, D., Bousquet, O., Lal, T., Weston, J., Schölkopf, B.: Learning with local and global consistency. Advances in neural information processing systems 16 (2003)" + ], + "bbox": [ + 212, + 146, + 787, + 646 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "SLIM: Spuriousness Mitigation with Minimal Human Annotations", + "bbox": [ + 287, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/SLIM_ Spuriousness Mitigation with Minimal Human Annotations/460c9d54-c412-4380-844b-3f656829c454_model.json b/2024/SLIM_ Spuriousness Mitigation with Minimal Human Annotations/460c9d54-c412-4380-844b-3f656829c454_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b01e4ef6657a23d398e81460202524944569f168 --- /dev/null +++ b/2024/SLIM_ Spuriousness Mitigation with Minimal Human Annotations/460c9d54-c412-4380-844b-3f656829c454_model.json @@ -0,0 +1,2082 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.304, + 0.142, + 0.703, + 0.185 + ], + "angle": 0, + "content": "SLIM: Spuriousness Mitigation with Minimal Human Annotations" + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.213, + 0.765, + 0.229 + ], + "angle": 0, + "content": "Xiwei Xuan\\(^{1}\\), Ziquan Deng\\(^{1}\\), Hsuan-Tien Lin\\(^{2}\\), and Kwan-Liu Ma\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.394, + 0.24, + 0.612, + 0.254 + ], + "angle": 0, + "content": "1 University of California, Davis" + }, + { + "type": "text", + "bbox": [ + 0.402, + 0.254, + 0.602, + 0.269 + ], + "angle": 0, + "content": "\\(^{2}\\) National Taiwan University" + }, + { + "type": "text", + "bbox": [ + 0.277, + 0.27, + 0.726, + 0.283 + ], + "angle": 0, + "content": "{xwxuan, ziqdeng, klma}@ucdavis.edu, htlin@csie.ntu.edu.tw" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.323, + 0.744, + 0.601 + ], + "angle": 0, + "content": "Abstract. Recent studies highlight that deep learning models often learn spurious features mistakenly linked to labels, compromising their reliability in real-world scenarios where such correlations do not hold. Despite the increasing research effort, existing solutions often face two main challenges: they either demand substantial annotations of spurious attributes, or they yield less competitive outcomes with expensive training when additional annotations are absent. In this paper, we introduce SLIM, a cost-effective and performance-targeted approach to reducing spurious correlations in deep learning. Our method leverages a human-in-the-loop protocol featuring a novel attention labeling mechanism with a constructed attention representation space. SLIM significantly reduces the need for exhaustive additional labeling, requiring human input for fewer than \\(3\\%\\) of instances. By prioritizing data quality over complicated training strategies, SLIM curates a smaller yet more feature-balanced data subset, fostering the development of spuriousness-robust models. Experimental validations across key benchmarks demonstrate that SLIM competes with or exceeds the performance of leading methods while significantly reducing costs. The SLIM framework thus presents a promising path for developing reliable models more efficiently. Our code is available in https://github.com/xiweix/SLIM.git/." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.614, + 0.741, + 0.628 + ], + "angle": 0, + "content": "Keywords: Model spuriousness \\(\\cdot\\) Human-in-the-loop \\(\\cdot\\) Data distribution" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.656, + 0.377, + 0.672 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Spurious correlations, where models mistakenly rely on irrelevant features to make decisions, pose a significant challenge in machine learning. Consequently, certain groups may experience deceptively inflated performance metrics, while accuracy reduces for others. This prevalence of \"right answers for wrong reasons\" fosters a misleading impression of model proficiency and constrains its utility across diverse contexts. Meanwhile, rectifying the issue of spuriousness is a challenging task, particularly when seeking to enhance the performance of underperforming groups. A notable example is the ISIC skin cancer dataset [5], where a typical spurious correlation occurs when color patches frequently co-occur with benign-labeled data. As shown in Fig. 1, a biased model may focus" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.37, + 0.128 + ], + "angle": 0, + "content": "X. Xuan et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.146, + 0.787, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.218, + 0.788, + 0.304 + ], + "angle": 0, + "content": "Fig. 1: (Left) GradCAM visualizations highlighting model attention on ISIC. (a, b) showcase a model biased towards patches, leading to a correct prediction with wrong reasons in (a) but an incorrect prediction despite correct focus in (b). (c, d) depict a spuriousness-robust model consistently focusing on core features. (Right) The table details the training data distribution by class (benign or malignant) and the presence of color patches, illustrating the imbalance and potential for spurious correlations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.324, + 0.788, + 0.4 + ], + "angle": 0, + "content": "on patches rather than skin lesions to identify melanoma, and fail when such patches are missing, while a desirable case would be the model captures correct skin regions regardless of spurious features' existence. Enhancing model robustness to address such dangerous misdiagnoses is crucial, driving a surge in research towards more effective solutions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.409, + 0.511, + 0.574 + ], + "angle": 0, + "content": "A typical approach in this context is to seek for optimal data distribution. GDRO [27] proves the efficacy of group-balanced data in enhancing model robustness. However, along with subsequent studies [14, 22], such an approach requires spuriousness labels at dataset scale, consuming a considerable cost. Our investigations reveal notable inconsistencies in human-annotated spuriousness labels, raising concerns about their relia" + }, + { + "type": "table_caption", + "bbox": [ + 0.521, + 0.427, + 0.788, + 0.496 + ], + "angle": 0, + "content": "Table 1: Annotation forms and quantities required by different methods on Waterbirds, highlighting the reduced workload of SLIM. (Att. = Attention, T. = Training set, V. = Validation set)" + }, + { + "type": "table", + "bbox": [ + 0.524, + 0.498, + 0.785, + 0.571 + ], + "angle": 0, + "content": "
MethodAnno. FormAnno. Amount
RRR [26]Map100%×T.
GradIA [9]Att. Yes-No, Map100%×T.
Energy Loss [24]Bounding box100%×T.
CRAYON [15]Att. Yes-No100%×T.
GDRO [27]Spuriousness label100%×T.
DFR [14]Spuriousness label100%×V.
SLIM (ours)Att. Yes-No2.5%×T. or 3%×V.
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.575, + 0.788, + 0.788 + ], + "angle": 0, + "content": "bility (refer to Table 6). There are other approaches with more direct human involvement, such as providing yes-no feedback on attention correctness [15,36]; or outlining core features with maps [9,26] or bounding boxes [24]. Despite different formats of queried information, they require extensive effort across the entire dataset, as summarized in Table 1. Considering the empirical applicability, our work aims to streamline this process by reducing human annotation workload and minimizing the ambiguity of collected data. Methods requiring spuriousness annotation tend to excel in performance at higher costs, while minimizing such expenses typically results in performance trade-offs [1,6]. Facing this challenge, techniques such as JTT [17] and CNC [41] design training or data-augmenting phases to reduce spuriousness without additional annotations. However, the extra computational cost raised by multiple training iterations is undesired. We believe, as spuriousness is indeed caused by data biases, improving data quality is more essential than designing complex learning strategies." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Therefore, the remaining research gap calls for a data quality-oriented approach that balances performance and cost. This is where our framework, SLIM, steps in. \\(SLIM\\) offers a novel approach to mitigating spurious correlations by" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.289, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SLIM: Spuriousness Mitigation with Minimal Human Annotations" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.147, + 0.786, + 0.463 + ], + "angle": 0, + "content": "intelligently leveraging a minimal set of human-annotated data and curating a feature-balanced data set to enhance model robustness without any significant computational overhead incurred by many existing methods. As shown in Fig. 2, inferred from a reference model, \\( SLIM \\) first constructs an attention representation space that reflects locally consistent data features and model attention, enabling efficient attention label estimation. — Representative instances can be sampled from this space for querying attention correctness labels and such labels can be expanded to neighbor instances. Subsequently, utilizing a visual explanation method that highlights model attention, like GradCAM [29], \\( SLIM \\) disentangles core and environment features within the latent space. Here, \"core features\" denote features critical for class determination, whereas \"environment features\" represent the surrounding contextual information. We agree with GDRO that balanced data distribution is crucial, but instead of using human-annotated labels to approximate spuriousness distribution, we leverage latent space features directly disentangled from model attention to balance data distribution, which is more aligned with model understanding and involving less human assumptions. Through this, \\( SLIM \\) assembles data groups with diverse feature combinations. Focused on data quality, a consistency-aware sampling mechanism is employed to curate a subset with optimal feature distributions — ensuring balanced core features across various environmental contexts. Finally, we conclude our framework with a straightforward training procedure." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.465, + 0.786, + 0.57 + ], + "angle": 0, + "content": "The experimental results demonstrate SLIM's advanced capability in spuriousness mitigation by performing comparably to, or exceeding state-of-the-art (SOTA) methods. Notably, SLIM accomplishes this with (1) a significantly reduced requirement for human annotations: \\(0.12\\% - 2.5\\%\\) of the training set size and \\(0.35\\% - 4\\%\\) of the validation set; and (2) a lightweight training for a robust model using our constructed data with \\(5\\% - 30\\%\\) of the original set. Our code will be published after the review process is complete. Our contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.621 + ], + "angle": 0, + "content": "- Data-quality-oriented spuriousness mitigation: SLIM introduces an interactive, simple, and effective data construction pipeline to create high-quality data for spuriousness mitigation." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.623, + 0.785, + 0.666 + ], + "angle": 0, + "content": "- Cost-efficient framework: In contrast to related works with leading performance, SLIM demonstrates a significant reduction in both annotation requirements and computational cost." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.668, + 0.785, + 0.71 + ], + "angle": 0, + "content": "- Competitive or superior model performance: Without incurring excessive costs, SLIM achieves comparable or superior performance to existing methods, confirming its effectiveness in mitigating spurious correlations." + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.71 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.735, + 0.387, + 0.751 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.765, + 0.786, + 0.84 + ], + "angle": 0, + "content": "Spuriousness mitigation with spuriousness labels. When a comprehensive set of spuriousness labels is accessible during training, methods like class balancing [11] and importance weighting [32] can be employed for mitigating spurious correlations, which utilizes data groups defined by class and spuriousness labels. Following the same schema, GDRO [27] refines the training approach" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.128 + ], + "angle": 0, + "content": "X. Xuan et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.388 + ], + "angle": 0, + "content": "by focusing on the worst-performing groups as determined by ERM loss. Aiming to reduce the dependency on comprehensive spuriousness labeling, several methods [1,6,17,21,30,32,41] have emerged to approximate group information using a reference model. While these approaches are technically sophisticated, they often fail to achieve the same level of performance as approaches informed by spuriousness labels. To balance the tradeoff between performance and cost, a further research direction reduces the requirement for extensive labeling by leveraging spuriousness labels in smaller validation datasets [14,22]. For example, SSA [22] adopts semi-supervised learning to predict group labels in training data according to information from validation, then applies GDRO to reduce spurious correlations. However, even with this reduced scale of validation sets, the labor-intensive nature of spuriousness labeling remains a barrier to scalability. Moreover, groups in these methods are formed based on human-defined criteria, risking a misalignment with the actual categories a model might discern. This drawback can hinder the true representation of spurious correlations and may lead to suboptimal mitigation outcomes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.389, + 0.788, + 0.646 + ], + "angle": 0, + "content": "Spuriousness mitigation with human-in-the-loop. Integrating human insights has proven beneficial in mitigating model spuriousness, as evidenced by various human-involved strategies. Considering the valuable human effort, [28] provides an efficient solution by querying data samples for issue mitigation with domain experts in the loop, which overlooks potential issues in the broader data scope. Another option for avoiding redundant human annotation is to query global spurious concepts [31, 38] from knowledgeable experts, which can be ambiguous and fail to cover all possibilities. More recent approaches [10, 24, 26] raise concerns about the inconsistency of spuriousness labels, introducing alternatives such as human-generated attention maps or bounding boxes outlining core features. These human-generated references help align the models' attention with the actual features relevant to decision-making. However, producing such annotations is challenging, and these techniques are resource-intensive as they necessitate annotations across the entire dataset. In response, more recent research research [2, 40] simplifies the annotation process to binary feedback on model's attention correctness. While less demanding, its requirement to annotate all instances still poses challenges in real practices." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.67, + 0.461, + 0.687 + ], + "angle": 0, + "content": "3 Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.702, + 0.788, + 0.763 + ], + "angle": 0, + "content": "We introduce the formal definition of the spuriousness mitigation problem [13]. Assume training data \\(\\mathcal{D} = \\{(x_i, y_i)\\}_{i=1}^n\\) with input features \\(\\boldsymbol{x}_i \\in \\mathbb{R}^d\\) and labels \\(y_i \\in [\\mathcal{L}]\\), where \\(\\mathcal{L} = \\{L_1, \\ldots, L_l\\}\\) are the classes in \\(\\mathcal{D}\\). Machine learning models for classification are trained to minimize the empirical risk:" + }, + { + "type": "equation", + "bbox": [ + 0.373, + 0.779, + 0.786, + 0.803 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\theta} ^ {*} \\in \\underset {\\boldsymbol {\\theta}} {\\arg \\min } \\mathbb {E} _ {(\\boldsymbol {x} _ {i}, y _ {i}) \\in \\mathcal {D}} [ l (f (\\boldsymbol {\\theta}, \\boldsymbol {x} _ {i}), y _ {i}) ], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "where \\(\\pmb{\\theta}\\) represent model parameters, \\(f(\\pmb{\\theta},\\pmb{x}_i)\\) and \\(l(*)\\) are the model outputs and the loss function, respectively." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.289, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "SLIM: Spuriousness Mitigation with Minimal Human Annotations" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.328 + ], + "angle": 0, + "content": "Groups. In spuriousness mitigation, we assume there exists a core feature set \\(\\mathcal{C} = \\{c_1,\\dots ,c_l\\}\\) and a spurious feature set \\(\\mathcal{S} = \\{s_1,\\dots ,s_k\\}\\) in \\(\\mathcal{D}\\). Each class \\(L_{i}\\) is uniquely identified by a core feature \\(c_{i}\\), while spurious features \\(s_j\\) can be found across different classes. We consider an instance in the dataset as having both a core feature \\(c_{i}\\) and a spurious feature \\(s_j\\), and each instance belongs to a unique group \\(g_{c_i,s_j}\\). Core features are reliable class indicators in both training and testing sets, while spurious features with misleading co-occurrence with class in training may miss such correlations in the testing phase. If the model mistakenly links spurious features with classes in training, it may fail in testing when such spurious features are absent. This can lead to lower accuracy rates, especially for less common group variations that do not exhibit the spurious feature the model has learned to rely on." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.329, + 0.787, + 0.375 + ], + "angle": 0, + "content": "Worst-group accuracy. The vital evaluation of spuriousness mitigation focuses on the model's performance in the subgroup where it is least effective – the worst-group accuracy. It is calculated by:" + }, + { + "type": "equation", + "bbox": [ + 0.357, + 0.387, + 0.786, + 0.42 + ], + "angle": 0, + "content": "\\[\nA c c _ {w} = \\min _ {g \\in \\mathcal {G}} \\left(\\frac {1}{| g |} \\sum_ {\\left(x _ {i}, y _ {i}\\right) \\in g} \\mathbf {1} [ f \\left(x _ {i}\\right) = y _ {i} ]\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.788, + 0.538 + ], + "angle": 0, + "content": "where \\(\\mathcal{G} = \\{\\cup_{c_i,s_j}g_{c_i,s_j}\\}\\) is the union of subgroups defined by core feature \\(c_{i}\\) and spurious feature \\(s_j\\), and \\(f(x_{i})\\) is the label predicted by the model. This metric is crucial because it directly assesses how effectively a model can generalize across different subgroups, especially those that might be mostly underrepresented because of spurious correlations. It ensures that improvements in model performance are not just reflective of gains on the majority or less challenging subsets of the data." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.563, + 0.332, + 0.579 + ], + "angle": 0, + "content": "4 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.789, + 0.844 + ], + "angle": 0, + "content": "In this section, we detail the method behind SLIM. In alignment with related works requiring no spuriousness label, our approach also follows a two-stage procedure: inferring groups with an ERM-trained reference model \\( f_{r} \\), and then leveraging the inferred groups to train a robust model. As an approach oriented by data quality, SLIM framework comprises a three-phase data construction pipeline, as illustrated in Fig. 2. Firstly, we build an attention representation space to group instances with both similar data features and model attention, preparing for the data sampling and label propagation. Following this, we sample typical instances for human to evaluate model's attention correctness with yes-no feedback, then estimate neighboring instances' attention labels accordingly. In the last phase, we filter out data with wrong attention and leverage attention matrix to disentangle core and environment features. With constructed feature sets, we apply consistency-weighted data sampling to curate a feature-balanced subset, with a desired core-environment balanced distribution. Finally, SLIM framework is concluded by a training process with the curated subset, resulting in a model exhibiting robustness against spurious correlations." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.37, + 0.128 + ], + "angle": 0, + "content": "X. Xuan et al." + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.143, + 0.773, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.331, + 0.788, + 0.457 + ], + "angle": 0, + "content": "Fig. 2: Overview of SLIM framework, with a data construction pipeline consisting of three phases: (1) Attention Space Construction, creating a space with data features and model attention aligned locally; (2) Attention Annotation and Expansion, where instances sampled from the attention space are labeled by human for attention correctness, and labels are propagated to neighboring instances; and (3) Balanced Data Curation, which filters out instances with incorrect attention, and utilize attention-weighted \\((F_A)\\) and inverse-attention-weighted \\((F_{\\bar{A}})\\) feature vectors to create core and environment feature sets, forming subgroups to assemble a feature-balanced subset for training a spuriousness-robust model." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.462, + 0.509, + 0.478 + ], + "angle": 0, + "content": "4.1 Attention Space Construction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.491, + 0.788, + 0.612 + ], + "angle": 0, + "content": "Considering the benefits of being aware of spuriousness information and drawbacks of the annotation cost, in our framework, we deploy a binary labeling scheme to simplify descriptive labeling tasks into yes-no feedback on model's attention correctness. Moreover, we aim to further facilitate this process by avoiding the necessity of exhaustive labeling. Our empirical observations indicate that a model would have consistent attention mechanisms on specific data subgroups. If such subgroups can be found, it would largely aid the labeling workload by enabling label estimations among neighbors." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.788, + 0.703 + ], + "angle": 0, + "content": "Driven by this, we aim to find a representation space where instances with similar attention outcomes (as indicated by highlighted image features) are closely grouped. However, our analysis reveals that conventional feature representation spaces formed directly from feature vectors do not fulfill this requirement - such spaces only maintain local consistency in data features but fail to capture the nuances of model attention (refer to Sec. 5.3)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.789, + 0.84 + ], + "angle": 0, + "content": "Thus, we have developed a process utilizing model attention in the latent space to support the establishment of such space, as illustrated in Fig. 3. For each image, we obtain its feature vector \\( F \\) and model attribution \\( A \\) in the latent space at first. Note that \\( A \\) represents the model's attention focus and can be linearly interpolated to visualize an attention heatmap. We then use \\( A \\) as a weight matrix for \\( F \\), producing an attention-weighted feature vector \\( F_{A} \\). Finally, with the dimensionality reduction method UMAP [20], we project vectors \\( F_{A} \\) into a 2D space, namely the attention space, which ensures local consistency encompasses both data features and the model's attention mechanisms." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.289, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "SLIM: Spuriousness Mitigation with Minimal Human Annotations" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.145, + 0.773, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.223, + 0.784, + 0.292 + ], + "angle": 0, + "content": "Fig. 3: Construction of the attention space. In the latent space, feature vector \\( F \\) and model's attribution vector \\( A \\) are extracted, representing input features and model's attention, representatively. By weighting \\( F \\) with \\( A \\), the attention-weighted feature vector \\( F_{A} \\) emphasizes features of model's top interest. All \\( F_{A} \\) are then projected to form the attention space featuring locally consistent features and model attention." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.298, + 0.785, + 0.388 + ], + "angle": 0, + "content": "Note that for a fair comparison, in alignment with baseline methods, experiments in this paper involve the CNN model (ResNet) for classification and GradCAM for obtaining model attention. But our framework is fully applicable to (1) other vision model architectures such as vision transformers (ViT) [8] or Swin transformers [18], and (2) other heatmap-based model visualization techniques, such as RISE [23] or ScoreCAM [33]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.413, + 0.569, + 0.428 + ], + "angle": 0, + "content": "4.2 Attention Annotation and Expansion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.44, + 0.785, + 0.667 + ], + "angle": 0, + "content": "Given that our attention space ensures locally consistent model attention, it allows for annotating a selectively sampled subset of instances and then estimating neighboring labels accordingly, thereby enabling efficient transfer of human expertise. The criteria for selecting these instances are twofold: (1) diversity, to encompass a broad covering of various data features and model attention patterns; (2) typicality, ensuring they are representative of their neighboring instances. To achieve this, we utilize the k-means clustering algorithm to partition our constructed attention space into \\( n \\) clusters (diversity). We then select \\( n \\) instances that are closest to each cluster's center (typicality). The choice of \\( n \\) is aimed at optimizing the balance between minimizing intra-cluster variance in data features and model attention patterns, and keeping the annotation budget manageable. Our empirical analysis, detailed in Sec. 5.3, reveals that the intra-cluster consistencies increase with \\( n \\), but this trend levels off upon reaching a specific value of \\( n \\) -the elbow point. Thus, selecting \\( n \\) at this elbow point optimizes consistency without unnecessarily increasing the annotation workload." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.668, + 0.787, + 0.756 + ], + "angle": 0, + "content": "We gather binary yes-no feedback from human experts on the correctness of the model's attention for these \\( n \\) instances, feeding this feedback along with our attention space embedding into the label spreading algorithm. This algorithm leverages data similarity and density to extend attention labels to previously unlabeled data [43]. Ultimately, this process yields a probability score for correct attention for each instance across the dataset." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.783, + 0.465, + 0.797 + ], + "angle": 0, + "content": "4.3 Balanced Data Curation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Following Sec. 4.2 that provides attention accuracy scores, we proceed to construct our data subset to support model's unbiased learning, as described below:" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.128 + ], + "angle": 0, + "content": "X. Xuan et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.328 + ], + "angle": 0, + "content": "Data screening. Inspired by works [3,7,12], our theoretical analysis further supports the premise that data receiving correct attention from a reference model are more likely to enhance the learning of core features in a robust model within a more balanced data distribution, as substantiated in Appx. A.2. Furthermore, our study indicates that even a reference model with inherent biases has the potential to accurately identify core features for a data subset. Such correctly captured features typically exhibit diversity, ensuring good coverage across the spectrum of the core feature set. This understanding informs our approach to curating a data subset that: (1) selects instances based on accurate model attention, and (2) ensures a balanced representation of core features across various environment features. Thus we first apply data screening to filter out instances with wrong attention according to attention accuracy scores." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.33, + 0.789, + 0.681 + ], + "angle": 0, + "content": "Disentangling core and environment features. After screening, the retained data subset mainly comprises instances with correct attention. This is reflected by attention masks that highlight core features, meaning that the corresponding attribution vector \\( A \\) in the latent space also highlights core features of feature vector \\( F \\). Moreover, existing studies have found that a model's feature vectors are representative of most input features, regardless of whether they are used for decision-making [14, 37]. Therefore, utilizing \\( A \\) as the weight matrix, we disentangle core and environment features in the latent space — weighting \\( F \\) by attention masks \\( A \\) and inverse attention masks \\( \\bar{A} \\). Subsequently, k-means clustering is applied to partition all attention-weighted feature vectors \\( F_{A} \\), and inverse-attention-weighted feature vectors \\( F_{\\bar{A}} \\), respectively, facilitating the formation of representative subgroups for the core feature set \\( (\\hat{C} = \\hat{c}_1,\\dots ,\\hat{c}_N) \\) and the environment feature set \\( (\\mathcal{E} = e_1,\\ldots ,e_M) \\). The number of clusters, determined through the elbow method, culminates in \\( M\\times N \\) subgroups \\( g_{\\hat{c}_i,e_j} \\), each characterized by a distinct combination of core and environment features \\( (\\hat{c}_i,e_j) \\). Data sampling. Given the unequal distribution of core and environment features, the internal consistency across each constructed subgroup \\( g_{\\hat{c}_i,e_j} \\) varies. Our sampling strategy aims to mitigate over-representation of particular features to prevent overfitting, while ensuring broad coverage of diverse core and environment features. We assess cluster consistency as a metric for determining the sample size from each cluster. For a cluster \\( \\hat{c}_i \\) in the core feature space, denote the distance between an instance \\( x \\) and the cluster center as \\( d_{\\hat{c}_i}(x) \\). The cluster consistency is defined as the reciprocal of the average distance per cluster:" + }, + { + "type": "equation", + "bbox": [ + 0.417, + 0.695, + 0.787, + 0.731 + ], + "angle": 0, + "content": "\\[\n\\rho_ {\\hat {c} _ {i}} = \\frac {1}{\\frac {1}{| \\hat {c} _ {i} |} \\sum_ {x \\in \\hat {c} _ {i}} d _ {\\hat {c} _ {i}} (x)}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.74, + 0.788, + 0.791 + ], + "angle": 0, + "content": "A lower \\(\\rho_{\\hat{c}_i}\\) signals less consistent features, necessitating a higher sample weight for adequately representing them. A cluster's sampling weight is defined as \\(w_{\\hat{c}_i} = \\frac{1}{\\rho_{\\hat{c}_i}}\\) and the number of samples to be drawn from a core cluster \\(\\hat{c}_i\\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.42, + 0.805, + 0.787, + 0.845 + ], + "angle": 0, + "content": "\\[\nn _ {\\hat {c} _ {i}} = N \\cdot \\frac {1 / \\rho_ {\\hat {c} _ {i}}}{\\sum_ {j = 1} ^ {| \\hat {C} |} 1 / \\rho_ {\\hat {c} _ {j}}}, \\tag {4}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.289, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "SLIM: Spuriousness Mitigation with Minimal Human Annotations" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.128 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "where \\(N\\) is the sampling budget. Moving to the environment feature space, the \\(n_{\\hat{c}_i}\\) instances from a core feature cluster \\(\\hat{c}_i\\) may distribute across multiple environment clusters \\(\\{e_1,\\dots ,e_M\\}\\) and formulate subgroups \\(\\{g_{\\hat{c}_i,e_1},\\ldots ,g_{\\hat{c}_i,e_M}\\}\\). A similar methodology is employed for a certain group \\(g_{\\hat{c}_i,e_i}\\). The average distance for \\(g_{\\hat{c}_i,e_i}\\) in the environment representation space is:" + }, + { + "type": "equation", + "bbox": [ + 0.389, + 0.235, + 0.786, + 0.272 + ], + "angle": 0, + "content": "\\[\n\\rho_ {\\hat {c} _ {i}, e _ {i}} = \\frac {1}{\\frac {1}{| \\hat {c} _ {i} , e _ {i} |} \\sum_ {x \\in (\\hat {c} _ {i} , e _ {i})} d _ {e _ {i}} (x)}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.281, + 0.787, + 0.317 + ], + "angle": 0, + "content": "Similarly, we have the sampling weight \\( w_{\\hat{c}_i,e_i} = \\frac{1}{\\rho_{\\hat{c}_i,e_i}} \\) and the number of samples to be drawn from a certain group \\( g_{\\hat{c}_i,e_i} \\) is finally defined by:" + }, + { + "type": "equation", + "bbox": [ + 0.273, + 0.328, + 0.786, + 0.368 + ], + "angle": 0, + "content": "\\[\nn _ {\\hat {c} _ {i}, e _ {i}} = n _ {\\hat {c} _ {i}} \\cdot \\frac {w _ {\\hat {c} _ {i} , e _ {i}}}{\\sum_ {j = 1} ^ {| \\mathcal {E} (\\hat {c} _ {i}) |} w _ {\\hat {c} _ {j} , e _ {i}}} = N \\cdot \\frac {1 / \\rho_ {\\hat {c} _ {i}}}{\\sum_ {j = 1} ^ {| \\hat {\\mathcal {C}} |} 1 / \\rho_ {\\hat {c} _ {j}}} \\cdot \\frac {1 / \\rho_ {\\hat {c} _ {i} , e _ {i}}}{\\sum_ {j = 1} ^ {| \\mathcal {E} (\\hat {c} _ {i}) |} 1 / \\rho_ {\\hat {c} _ {i} , e _ {j}}}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.379, + 0.788, + 0.456 + ], + "angle": 0, + "content": "where \\( |\\hat{\\mathcal{C}}| \\) denotes the number of core clusters and \\( |\\mathcal{E}(\\hat{c}_i)| \\) denotes the number of subgroups formulated by instances from cluster \\( \\hat{c}_i \\). To summary, our sampling strategy takes feature consistencies and sampling budget into consideration, and ensures fair representation of different core-environment feature combinations across the feature spaces." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.478, + 0.334, + 0.494 + ], + "angle": 0, + "content": "4.4 Training" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.503, + 0.787, + 0.564 + ], + "angle": 0, + "content": "In our pursuit to mitigate spurious correlations, we emphasize the pivotal role of data quality by ensuring a balanced distribution of core features across various environmental contexts within our dataset. Given this focus, we employ a basic training approach for our model to learn from this well-prepared dataset." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.564, + 0.788, + 0.67 + ], + "angle": 0, + "content": "We apply our data curation pipeline separately to the training \\((SLIM_{Tr})\\) and validation \\((SLIM_{Val})\\) sets. For the data constructed from the training set, we adopt the conventional ERM. For the dataset derived from the validation set, we employ the training technique presented by DFR [14]. It involves retraining only the final linear layer of a classification model that was initially trained using ERM. Note that our approach diverges from DFR practices by utilizing our curated dataset instead of a metagroup-balanced dataset." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.692, + 0.377, + 0.71 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.724, + 0.429, + 0.74 + ], + "angle": 0, + "content": "5.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.75, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Datasets. Our study explores a range of datasets with inherent spurious correlations to challenge image classification models, including: (1) Waterbirds [27]. Contrasting landbirds and waterbirds against congruent/incongruent backgrounds, this dataset tests models' ability to focus on relevant features for bird type classification. (2) CelebA [19]. A comprehensive celebrity image dataset where the focus is on distinguishing hair color (blond vs. non-blond) to examine" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.128 + ], + "angle": 0, + "content": "X. Xuan et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.254 + ], + "angle": 0, + "content": "biases in associating hair color with gender. (3) ISIC [5]. Aiming to differentiate benign from malignant melanoma in dermoscopic images, we target colorful patches as spurious features, aligning with previous studies [25]. (4) NICO. Derived from \\(\\mathrm{NICO}++\\) [42], this dataset features various object categories in shifted contexts to probe spurious correlations. (5) ImageNet-9 [35]. This dataset is crafted from ImageNet to test models' robustness against background variations. Detailed information is available in Appx. A.4." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.261, + 0.788, + 0.368 + ], + "angle": 0, + "content": "Baselines. We benchmark SLIM against a spectrum of SOTA spuriousness mitigation techniques, categorized as follows: (1) Techniques not reliant on spuriousness labels: EIIL [6], PGI [1], GEORGE [30], LfF [21], CIM [32], JTT [17], and CNC [41]; (2) Techniques requiring additional information such as predefined spuriousness concepts: DISC [34]; (3) Methods utilizing spuriousness labels from validation sets: SSA [22] and DFR [14]; (4) Approaches utilizing spuriousness labels from training sets: GDRO [27] and PDE [7]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.375, + 0.788, + 0.527 + ], + "angle": 0, + "content": "Evaluation. Aligning with baselines, our experiments involve CNN models with ResNet50 architecture and GradCAM for visualizing model attention. For the capability of performance enhancement, we evaluate the worst-group and average accuracy. Additionally, we consider the necessity for extra labels and the additional (add.) training cost to evaluate computational efficiency. Note that the add. training cost is aligned with [13, 39], measuring add. data amounts required for second-stage training, where 1 means the amount of data equals the training set size. Furthermore, we examine the attention precision to evaluate whether model attention is corrected. All quantification experiments are based on five independent runs with different random seeds." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.535, + 0.787, + 0.567 + ], + "angle": 0, + "content": "More experimental results with other architecture (ViT) and datasets (MetaShift [16] and FMoW [4]) are provided in Appx. A.8." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.584, + 0.788, + 0.627 + ], + "angle": 0, + "content": "Table 2: Worst-group and average accuracy (%) comparison on binary classification tasks, highlighting best and second-best performances. \\(\\spadesuit\\) and \\(\\diamond\\) indicate methods requiring spuriousness concepts and attention labels, respectively." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.638, + 0.788, + 0.827 + ], + "angle": 0, + "content": "
MethodSpuriousness labelAdd. training costWaterbirdsCelebAISIC
WorstAvgWorstAvgWorstAvg
ERMX0×62.6±0.397.3±1.047.7±2.194.9±0.356.7±2.279.0±1.3
EIILX0×83.5±2.894.2±1.381.7±0.885.7±0.172.3±3.986.4±1.7
PGIX0×79.5±1.995.5±0.885.3±0.387.3±0.167.4±1.985.7±1.1
GEORGEX1×76.2±2.095.7±0.554.9±1.994.6±0.258.8±2.183.2±1.4
LfFX1×77.3±2.391.4±1.776.8±1.284.5±0.772.1±2.185.4±0.5
CIMX1×77.8±1.594.7±1.183.8±1.090.5±0.671.7±1.685.9±1.1
JTTX5×-6×83.1±3.590.6±0.381.5±1.788.1±0.328.1±4.386.2±0.3
CNCX2×-12×88.5±0.390.9±0.188.8±0.989.9±0.568.2±0.284.6±0.2
DISCX◇1×88.7±0.793.8±0.785.7±1.189.9±0.773.4±0.484.7±0.4
SLIMTrX◇0.1×-0.2×89.1±0.691.7±0.589.4±0.390.1±0.279.1±0.588.5±0.4
SLIMValX◇0.05×-0.3×91.0±0.493.8±0.489.4±0.590.7±0.274.0±0.285.4±0.2
SSAvalidation1.5×-5×89.0±0.692.2±0.989.8±1.392.8±0.161.7±0.777.7±1.0
DFRvalidation0.05×-0.3×89.6±0.792.1±0.287.3±1.090.2±0.871.9±0.784.3±0.3
GDROtraining0×89.9±0.692.6±0.188.9±1.393.9±0.166.8±3.580.5±1.9
PDEtraining0×85.9±2.790.5±1.885.0±4.092.3±0.866.4±2.382.2±1.3
" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.289, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SLIM: Spuriousness Mitigation with Minimal Human Annotations" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 3: Worst-group and average accuracy (%) comparison on multi-classification tasks, highlighting best and second-best results. \\(\\diamond\\) indicates methods requiring attention labels. - represents inapplicable methods because such methods require spuriousness labels while the dataset does not provide them." + }, + { + "type": "table", + "bbox": [ + 0.274, + 0.212, + 0.731, + 0.306 + ], + "angle": 0, + "content": "
MethodSpuriousness labelAdd. training costNICOImageNet9
WorstAvgWorstAvg
ERMX0×11.9±0.270.9±1.255.3±2.075.6±1.1
CNCX5×-12×72.6±1.789.5±0.658.9±0.674.3±0.2
SLIMTrX◇0.1×-0.2×78.0±0.892.1±0.168.1±0.374.5±0.2
SLIMValX◇0.1×-0.2×73.8±0.490.7±0.264.6±0.474.2±0.3
DFRvalidation0.4×70.9±2.185.8±0.6--
GDROtraining0×73.3±1.090.3±0.5--
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.307, + 0.784, + 0.334 + ], + "angle": 0, + "content": "Table 4: SLIM's primary cost, including the data amount required for attention annotation (Att.), and the constructed data size for model training (Training)." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.347, + 0.787, + 0.392 + ], + "angle": 0, + "content": "
MethodWaterbirdCelebAISICNICOImageNet 9
Att.TrainingAtt.TrainingAtt.TrainingAtt.TrainingAtt.Training
SLIMTr120 (2.5%)0.1×200 (0.12%)0.1×120 (0.7%)0.2×40 (1.8%)0.2×180 (0.4%)0.1×
SLIMVal40 (3%)0.1×70 (0.35%)0.05×40 (2%)0.3×20 (2.7%)0.2×50 (1%)0.1×
" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.394, + 0.573, + 0.409 + ], + "angle": 0, + "content": "5.2 Spuriousness Mitigation Performance" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.418, + 0.785, + 0.461 + ], + "angle": 0, + "content": "In this section, we present the performance evaluation by comparing different methods across multiple binary and multiclass classification datasets mentioned in Sec. 5.1. Key findings are demonstrated below." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.463, + 0.787, + 0.613 + ], + "angle": 0, + "content": "Superior worst-group performance. The metric of worst-group accuracy \\((Acc_w)\\) is pivotal for spuriousness mitigation methods (refer to Sec. 3), and the results in Tables 2 and 3 showcase SLIM's superior capability in this regard. Across various classification tasks, SLIM consistently outperforms competing methods in achieving the highest \\(Acc_w\\) while maintaining high average accuracy. Specifically, in binary classification scenarios (Table 2), SLIM leads in performance for Waterbirds and ISIC and secures a close second in CelebA. In multi-classification tasks (Table 3), \\(SLIM_{Tr}\\) and \\(SLIM_{Val}\\) rank as the top performers across the board, underscoring the effectiveness of our approach in spuriousness mitigation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Cost-efficient spuriousness mitigation. Methods in Tables 2 and 3 are divided into four groups from top to bottom, as per the categorization established in Sec. 5.1 according to the computational cost. The \"Add. training cost\" columns denote the additional training cost for mitigating spuriousness apart from the initial training, such as additional learning phases (CNC, SSA) or data augmentation (JTT). Results in the table substantiate that methods with higher costs (higher Add. training costs or more spuriousness labels) tend to outperform others, aligning with prior research findings. Nevertheless, among the well-performed methods, it is worth mentioning that \\(SLIM\\) is more cost-efficient: \\(SLIM_{Tr}\\) used 0.1-0.2 proportion of training sets and \\(SLIM_{Val}\\) used 0.05-0.3 proportion of validation sets for additional training, which is the same as DFR and much smaller than others. Moreover, \\(SLIM\\) requires labels for attention correctness, which is more cost-efficient to obtain than spuriousness labels (refer to Sec. 5.3 for more details). Table 4 presents \\(SLIM\\)'s primary cost, including the training data amount for mitigating spuriousness, which is the detailed version of \"Add. train" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.127 + ], + "angle": 0, + "content": "X. Xuan et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.784, + 0.174 + ], + "angle": 0, + "content": "Table 5: Attention accuracy evaluation. (Left) GradCAM visualization of model attention. (Right) Worst group and average attention accuracy of selected methods." + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.187, + 0.589, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "table", + "bbox": [ + 0.633, + 0.185, + 0.768, + 0.297 + ], + "angle": 0, + "content": "
WaterbirdsAIoU
WorstAvg
ERM.25±.05.47±.03
CNC.48±.05.49±.04
GDRO.49±.03.52±.03
PDE.50±.02.52±.03
SLIMTr.53±.02.61±.02
ImageNet 9AIoU
WorstAvg
ERM.33±.07.55±.03
CNC.44±.03.61±.03
SLIMTr.58±.04.68±.03
" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.299, + 0.784, + 0.359 + ], + "angle": 0, + "content": "ing cost\" in Tables 2 and 3, and the data amount requiring attention labels (Att.). Notably, our approach only requires no more than \\(3\\%\\) of the corresponding data split (train or validation) for attention-label annotation, further proving the cost-efficiency of SLIM." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.36, + 0.788, + 0.571 + ], + "angle": 0, + "content": "Enhanced attention accuracy. Beyond classification accuracy, we further evaluate the performance of spuriousness mitigation methods via attention accuracy. Specifically, for models resulting from each method, we randomly sample data instances to visualize model attention, and measure attention accuracy scores in both worst-group and dataset levels, consistent with our evaluation of classification accuracy. Align with the related work [40], we employ the Adjusted Intersection-over-Union (AIoU) score as our attention accuracy measurement, which is an adjusted IoU score with a threshold, aiming at measuring the overlap between model's highest attributed region and the ground-truth bounding box. Details related to AIoU are given in Appx. A.7. Table 5 presents the evaluation results. The left GradCAM examples showcase SLIM's capability of correcting model's wrong attention, and the right table compares SLIM with SOTA methods, where we can find our method secures the best performance in terms of both worst-group and average attention accuracy." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.591, + 0.53, + 0.607 + ], + "angle": 0, + "content": "5.3 Attention Labeling Performance" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.795 + ], + "angle": 0, + "content": "To compare the processes of annotating spuriousness versus attention correctness, we employed crowdsourcing tasks using the Waterbirds and NICO datasets, as both of which come with ground-truth spuriousness labels. For this study, we selected a random set of 120 images from each dataset and established two separate tasks: one for spuriousness labeling and one for attention correctness labeling. For each task, we recruited 60 workers independently from the Prolific platform to prevent learning biases from cross-task participation. We meticulously recorded the time taken to label each image and reviewed all submissions to exclude any outliers with unusually long or short annotation times. Ensuring ethical research standards, our study avoided collecting personally identifiable information and excluded any potentially offensive content. Detailed instructions provided to the crowdsourcing participants can be found in Appx. A.6." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Evaluation Consistency. Table 6 presents metrics for this study. While the annotation time for Waterbirds remains similar across tasks, attention correctness for NICO is annotated more quickly than the other. This difference may stem" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.289, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "SLIM: Spuriousness Mitigation with Minimal Human Annotations" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.153, + 0.446, + 0.194 + ], + "angle": 0, + "content": "Table 6: Comparison of crowdsourced labeling tasks: spuriousness vs. attention correctness." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.207, + 0.444, + 0.281 + ], + "angle": 0, + "content": "
WaterbirdsAvg. TimeAvg. Consis.
Spuriousness4.9±2.3 sec81.9±14.8%
Att.4.5±1.4 sec94.6±10.5%
NICOAvg. TimeAvg. Consis.
Spuriousness7.9±3.8 sec70.1±19.1%
Att.5.3±1.6 sec95.1±6.1%
" + }, + { + "type": "image", + "bbox": [ + 0.455, + 0.144, + 0.754, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.447, + 0.231, + 0.758, + 0.286 + ], + "angle": 0, + "content": "Fig. 4: Comparison of neighbor consistency over original feature space and our attention space on Waterbirds, w.r.t. model attention (left) and data features (right)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.286, + 0.787, + 0.39 + ], + "angle": 0, + "content": "from NICO's broader array of spuriousness choices versus the binary attention annotations. To measure the annotation consistency—whether a single receives divergent labels—we use the percentage of the most frequent label as a consensus measure. A higher percentage suggests greater consistency. Our findings indicate that attention correctness annotations exhibit a notably higher consistency, suggesting that it is less ambiguous to determine correct attention than to identify spurious features." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.391, + 0.787, + 0.572 + ], + "angle": 0, + "content": "Attention Consistency. \\(SLIM\\) requires spreading labels in a space where attention patterns are consistently similar in nearby areas. To evaluate its feasibility, we compared how similar neighbors are in the original feature space versus the attention space created by \\(SLIM\\). We used k-means clustering with varying numbers of clusters to measure the intra-cluster similarity for both the model's attention matrices and the feature vectors. A higher similarity score means the space has more consistent patterns, which is essential for the accurate spreading of labels. As Fig. 4 shows, our attention space significantly improves consistency in the model's attention, while also maintaining a good level of data feature consistency. Increasing the number of clusters generally increases consistency, but there is a point where adding more clusters does not lead to further gains, as shown by the curve flattening at \\(k = 120\\) in the left plot." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.573, + 0.787, + 0.648 + ], + "angle": 0, + "content": "In summary, we have demonstrated that the annotation module of SLIM is less time-consuming and yields more consistent labeling feedback. Additionally, the attention space engineered by SLIM establishes a solid foundation for propagating attention labels, considering the consistent model attention observed among neighboring instances." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.667, + 0.391, + 0.683 + ], + "angle": 0, + "content": "5.4 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.787, + 0.81 + ], + "angle": 0, + "content": "We investigate two critical factors of SLIM, which are detailed below. Other factors such as the dimension of the attention space are discussed in Appx. A.8. Influence of data amount for attention annotations. We change the data amount for attention annotations to investigate their effect on performance, keeping other variables constant. Table 7 illustrates that both the worst-group and average accuracy tend to rise with more annotations, while such incremental benefit stabilizes after reaching a certain point, helping us identify an optimal number of annotations that balance performance with annotation cost." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Influence of training set size. This size refers to the size of the data subset constructed in SLIM's final phase for training a robust model. By only varying" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.128 + ], + "angle": 0, + "content": "X. Xuan et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.785, + 0.173 + ], + "angle": 0, + "content": "Table 7: Performance variation with different attention annotation amounts in SLIM. Marks in table: best and second best performance; the amounts used in Table 2." + }, + { + "type": "table", + "bbox": [ + 0.258, + 0.185, + 0.747, + 0.325 + ], + "angle": 0, + "content": "
MethodWaterbirdsCelebAISIC
AmountWorstAvgAmountWorstAvgAmountWorstAvg
SLIMTr6078.2±0.586.0±0.310086.4±0.290.2±0.16075.3±0.286.6±0.9
9085.9±0.989.3±0.315087.8±0.491.1±0.79074.4±0.487.4±0.7
12089.1±0.691.7±0.520089.4±0.490.1±0.212079.1±0.588.5±0.4
15089.4±0.691.7±0.225089.4±0.290.9±0.115079.9±0.888.8±0.7
18089.3±0.391.0±1.530089.5±0.191.2±0.218079.8±0.388.6±0.3
SLIMVal1078.9±8.786.5±5.14084.7±0.487.4±0.31070.5±0.583.3±1.2
2589.2±0.693.4±1.05587.6±0.390.4±0.62572.7±0.185.2±0.8
4091.0±0.493.8±0.47089.4±0.490.7±0.24074.0±0.285.4±0.2
5591.5±0.393.8±0.18589.7±0.190.5±0.45574.0±0.285.9±0.5
7090.7±0.394.0±0.310089.7±0.290.5±0.27074.1±0.285.8±0.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.356 + ], + "angle": 0, + "content": "Table 8: Performance variation with different training sizes. \\(SLIM_{Tr}\\) and \\(SLIM_{Val}\\) are trained on different proportions (Prop) of training sets and validation sets, respectively." + }, + { + "type": "table", + "bbox": [ + 0.258, + 0.368, + 0.747, + 0.485 + ], + "angle": 0, + "content": "
MethodWaterbirdsCelebAISIC
Prop%WorstAvgProp%WorstAvgProp%WorstAvg
SLIMTr582.4±1.083.6±1.3588.9±0.190.0±0.11577.6±2.187.1±1.2
1089.1±0.691.7±0.51089.4±0.390.1±0.22079.1±0.588.5±0.4
1282.3±1.387.8±0.31589.4±0.491.0±0.42577.4±1.588.2±1.2
1583.5±2.789.6±0.52088.9±0.190.8±0.33078.3±1.588.3±0.9
SLIMVal584.2±3.689.2±2.6385.7±1.388.5±1.12568.2±2.181.9±0.8
1091.0±0.493.8±0.4589.4±0.490.7±0.23074.0±0.285.4±0.2
1287.6±0.693.9±1.31088.3±1.490.3±0.53572.5±0.585.3±0.1
1586.8±2.094.6±0.51587.2±0.490.3±0.44070.9±1.285.4±3.5
" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.49, + 0.785, + 0.564 + ], + "angle": 0, + "content": "this parameter (Table 8), we find that the optimal size is inherently linked to the size of our constructed subgroups, \\( |g_{\\hat{c}_i,e_i}| \\). An optimal size is reached when data from various groups are sufficiently sampled without redundant oversampling. These findings guide the hyperparameter selection in this phase, ensuring an effective cost allocation." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.593, + 0.357, + 0.608 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.628, + 0.785, + 0.763 + ], + "angle": 0, + "content": "In this paper, we introduce SLIM, a novel framework for spuriousness mitigating in deep learning. With minimal human annotation effort, SLIM provides a data construction pipeline to create a feature-balanced subset for training a robust model. Extensive evaluation demonstrates the superior cost-effectiveness of SLIM, which matches or outperforms state-of-the-art methods with far less costs. The success of SLIM underscores the potential of data quality and efficient human supervision in the development of robust models. We believe that SLIM will inspire further research in this direction, ultimately leading to more cost-effective machine learning systems for reliable AI." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.765, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Acknowledgements. This research is supported in part by Bosch Research and the National Institute of Health with grants P41-EB032840 and R01CA270454. H.-T. Lin is partially supported by the National Taiwan University Center for Data Intelligence via NTU-112L900901 and the Ministry of Science and Technology in Taiwan via MOST 112-2628-E-002-030." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.289, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SLIM: Spuriousness Mitigation with Minimal Human Annotations" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.173, + 0.785, + 0.214 + ], + "angle": 0, + "content": "1. Ahmed, F., Bengio, Y., Van Seijen, H., Courville, A.: Systematic generalisation with group invariant predictions. In: International Conference on Learning Representations (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.215, + 0.785, + 0.268 + ], + "angle": 0, + "content": "2. Bontempelli, A., Teso, S., Tentori, K., Giunchiglia, F., Passerini, A.: Concept-level debugging of part-prototype networks. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=oiwXwPDTyNk" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.269, + 0.785, + 0.31 + ], + "angle": 0, + "content": "3. Chen, Z., Deng, Y., Wu, Y., Gu, Q., Li, Y.: Towards understanding the mixture-of-experts layer in deep learning. Advances in neural information processing systems 35, 23049-23062 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.31, + 0.785, + 0.337 + ], + "angle": 0, + "content": "4. Christie, G., Fendley, N., Wilson, J., Mukherjee, R.: Functional map of the world. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.337, + 0.785, + 0.419 + ], + "angle": 0, + "content": "5. Codella, N.C.F., Gutman, D., Celebi, M.E., Helba, B., Marchetti, M.A., Dusza, S.W., Kalloo, A., Liopyris, K., Mishra, N., Kittler, H., Halpern, A.: Skin lesion analysis toward melanoma detection: A challenge at the 2017 international symposium on biomedical imaging (isbi), hosted by the international skin imaging collaboration (isic). In: 2018 IEEE 15th International Symposium on Biomedical Imaging (ISBI 2018). pp. 168-172 (2018). https://doi.org/10.1109/ISBI.2018.8363547" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.42, + 0.785, + 0.46 + ], + "angle": 0, + "content": "6. Creager, E., Jacobsen, J.H., Zemel, R.: Environment inference for invariant learning. In: International Conference on Machine Learning. pp. 2189-2200. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.46, + 0.785, + 0.514 + ], + "angle": 0, + "content": "7. Deng, Y., Yang, Y., Mirzasoleiman, B., Gu, Q.: Robust learning with progressive data expansion against spurious correlation. In: Thirty-seventh Conference on Neural Information Processing Systems (2023), https://openreview.net/forum?id=9QEVJ9qm46" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.514, + 0.785, + 0.568 + ], + "angle": 0, + "content": "8. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.568, + 0.785, + 0.624 + ], + "angle": 0, + "content": "9. Gao, Y., Sun, T.S., Zhao, L., Hong, S.R.: Aligning eyes between humans and deep neural network through interactive attention alignment. Proc. ACM Hum-. Comput. Interact. 6(CSCW2) (nov 2022). https://doi.org/10.1145/3555590, https://doi.org/10.1145/3555590" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.624, + 0.785, + 0.664 + ], + "angle": 0, + "content": "0. Gao, Y., Sun, T.S., Zhao, L., Hong, S.R.: Aligning eyes between humans and deep neural network through interactive attention alignment. Proceedings of the ACM on Human-Computer Interaction 6(CSCW2), 1-28 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.664, + 0.785, + 0.692 + ], + "angle": 0, + "content": "1. He, H., Garcia, E.A.: Learning from imbalanced data. IEEE Transactions on knowledge and data engineering 21(9), 1263-1284 (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.692, + 0.785, + 0.732 + ], + "angle": 0, + "content": "2. Jelassi, S., Li, Y.: Towards understanding how momentum improves generalization in deep learning. In: International Conference on Machine Learning. pp. 9965-10040. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.732, + 0.785, + 0.772 + ], + "angle": 0, + "content": "3. Joshi, S., Yang, Y., Xue, Y., Yang, W., Mirzasoleiman, B.: Towards mitigating spurious correlations in the wild: A benchmark & a more realistic dataset. arXiv preprint arXiv:2306.11957 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.785, + 0.8 + ], + "angle": 0, + "content": "4. Kirichenko, P., Izmailov, P., Wilson, A.G.: Last layer re-training is sufficient for robustness to spurious correlations. arXiv preprint arXiv:2204.02937 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "5. Lee, S., Payani, A., Chau, D.H.P.: Towards mitigating spurious correlations in image classifiers with simple yes-no feedback. In: AI and HCI Workshop at the 40th International Conference on Machine Learning (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.173, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.128 + ], + "angle": 0, + "content": "X. Xuan et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "16. Liang, W., Zou, J.: Metashift: A dataset of datasets for evaluating contextual distribution shifts and training conflicts. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=MTex8qKavoS" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.787, + 0.243 + ], + "angle": 0, + "content": "17. Liu, E.Z., Haghgoo, B., Chen, A.S., Raghunathan, A., Koh, P.W., Sagawa, S., Liang, P., Finn, C.: Just train twice: Improving group robustness without training group information. In: International Conference on Machine Learning. pp. 6781-6792. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.243, + 0.785, + 0.298 + ], + "angle": 0, + "content": "18. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.298, + 0.785, + 0.338 + ], + "angle": 0, + "content": "19. Liu, Z., Luo, P., Wang, X., Tang, X.: Deep learning face attributes in the wild. In: Proceedings of the IEEE international conference on computer vision. pp. 3730-3738 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.339, + 0.785, + 0.366 + ], + "angle": 0, + "content": "20. McInnes, L., Healy, J., Melville, J.: Umap: Uniform manifold approximation and projection for dimension reduction. arXiv preprint arXiv:1802.03426 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.366, + 0.785, + 0.405 + ], + "angle": 0, + "content": "21. Nam, J., Cha, H., Ahn, S., Lee, J., Shin, J.: Learning from failure: De-biasing classifier from biased classifier. Advances in Neural Information Processing Systems 33, 20673-20684 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.406, + 0.785, + 0.446 + ], + "angle": 0, + "content": "22. Nam, J., Kim, J., Lee, J., Shin, J.: Spread spurious attribute: Improving worst-group accuracy with spurious attribute estimation. arXiv preprint arXiv:2204.02070 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.446, + 0.785, + 0.473 + ], + "angle": 0, + "content": "23. Petsiuk, V., Das, A., Saenko, K.: Rise: Randomized input sampling for explanation of black-box models. arXiv preprint arXiv:1806.07421 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.785, + 0.5 + ], + "angle": 0, + "content": "24. Rao, S., Böhle, M., Parchami-Araghi, A., Schiele, B.: Using explanations to guide models. arXiv preprint arXiv:2303.11932 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.785, + 0.541 + ], + "angle": 0, + "content": "25. Rieger, L., Singh, C., Murdoch, W.J., Yu, B.: Interpretations are useful: Penalizing explanations to align neural networks with prior knowledge. In: Proceedings of the 37th International Conference on Machine Learning. ICML'20, JMLR.org (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.541, + 0.785, + 0.609 + ], + "angle": 0, + "content": "26. Ross, A.S., Hughes, M.C., Doshi-Velez, F.: Right for the right reasons: Training differentiable models by constraining their explanations. In: Proceedings of the Twenty-Sixth International Joint Conference on Artificial Intelligence, IJCAI-17. pp. 2662-2670 (2017). https://doi.org/10.24963/ijcai.2017/371, https://doi.org/10.24963/ijcai.2017/371" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.609, + 0.785, + 0.65 + ], + "angle": 0, + "content": "27. Sagawa*, S., Koh*, P.W., Hashimoto, T.B., Liang, P.: Distributionally robust neural networks. In: International Conference on Learning Representations (2020), https://openreview.net/forum?id=ryxGuJrFvS" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.65, + 0.785, + 0.704 + ], + "angle": 0, + "content": "28. Schramowski, P., Stammer, W., Teso, S., Brugger, A., Herbert, F., Shao, X., Luigs, H.G., Mahlein, A.K., Kersting, K.: Making deep neural networks right for the right scientific reasons by interacting with their explanations. Nature Machine Intelligence 2(8), 476-486 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.704, + 0.785, + 0.759 + ], + "angle": 0, + "content": "29. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Gradcam: Visual explanations from deep networks via gradient-based localization. In: Proceedings of the IEEE international conference on computer vision. pp. 618-626 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.785, + 0.799 + ], + "angle": 0, + "content": "30. Sohoni, N., Dunnmon, J., Angus, G., Gu, A., Ré, C.: No subclass left behind: Fine-grained robustness in coarse-grained classification problems. Advances in Neural Information Processing Systems 33, 19339-19352 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "31. Srivastava, M., Hashimoto, T., Liang, P.: Robustness to spurious correlations via human annotations. In: International Conference on Machine Learning. pp. 9109-9119. PMLR (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.289, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SLIM: Spuriousness Mitigation with Minimal Human Annotations" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "32. Taghanaki, S.A., Choi, K., Khasahmadi, A.H., Goyal, A.: Robust representation learning via perceptual similarity metrics. In: International Conference on Machine Learning. pp. 10043-10053. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.788, + 0.245 + ], + "angle": 0, + "content": "33. Wang, H., Wang, Z., Du, M., Yang, F., Zhang, Z., Ding, S., Mardziel, P., Hu, X.: Score-cam: Score-weighted visual explanations for convolutional neural networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 24-25 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.245, + 0.788, + 0.286 + ], + "angle": 0, + "content": "34. Wu, S., Yuksekgonul, M., Zhang, L., Zou, J.: Discover and cure: Concept-aware mitigation of spurious correlation. In: Proceedings of the 40th International Conference on Machine Learning. ICML'23, JMLR.org (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.286, + 0.788, + 0.314 + ], + "angle": 0, + "content": "35. Xiao, K., Engstrom, L., Ilyas, A., Madry, A.: Noise or signal: The role of image backgrounds in object recognition. ArXiv preprint arXiv:2006.09994 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.314, + 0.788, + 0.355 + ], + "angle": 0, + "content": "36. Xuan, X., Ono, J.P., Gou, L., Ma, K.L., Ren, L.: Attributionscanner: A visual analytics system for metadata-free data-slicing based model validation. arXiv preprint arXiv:2401.06462 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.355, + 0.788, + 0.383 + ], + "angle": 0, + "content": "37. Xue, Y., Payani, A., Yang, Y., Mirzasoleiman, B.: Eliminating spurious correlations from pre-trained models via data mixing. arXiv preprint arXiv:2305.14521 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.383, + 0.788, + 0.438 + ], + "angle": 0, + "content": "38. Yan, S., Yu, Z., Zhang, X., Mahapatra, D., Chandra, S.S., Janda, M., Soyer, P., Ge, Z.: Towards trustable skin cancer diagnosis via rewriting model's decision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11568-11577 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.438, + 0.788, + 0.494 + ], + "angle": 0, + "content": "39. Yang, Y., Gan, E., Karolina Dziugaite, G., Mirzasoleiman, B.: Identifying spurious biases early in training through the lens of simplicity bias. In: Proceedings of The 27th International Conference on Artificial Intelligence and Statistics. Proceedings of Machine Learning Research, vol. 238, pp. 2953-2961. PMLR (02-04 May 2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.494, + 0.788, + 0.535 + ], + "angle": 0, + "content": "40. Yang, Y., Nushi, B., Palangi, H., Mirzasoleiman, B.: Mitigating spurious correlations in multi-modal models during fine-tuning. In: Proceedings of the 40th International Conference on Machine Learning. ICML'23, JMLR.org (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.535, + 0.788, + 0.577 + ], + "angle": 0, + "content": "41. Zhang, M., Sohoni, N.S., Zhang, H.R., Finn, C., Ré, C.: Correct-n-contrast: A contrastive approach for improving robustness to spurious correlations. arXiv preprint arXiv:2203.01517 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.577, + 0.788, + 0.618 + ], + "angle": 0, + "content": "42. Zhang, X., He, Y., Xu, R., Yu, H., Shen, Z., Cui, P.: Nico++: Towards better benchmarking for domain generalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16036-16047 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.618, + 0.788, + 0.647 + ], + "angle": 0, + "content": "43. Zhou, D., Bousquet, O., Lal, T., Weston, J., Schölkopf, B.: Learning with local and global consistency. Advances in neural information processing systems 16 (2003)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.647 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/SLIM_ Spuriousness Mitigation with Minimal Human Annotations/460c9d54-c412-4380-844b-3f656829c454_origin.pdf b/2024/SLIM_ Spuriousness Mitigation with Minimal Human Annotations/460c9d54-c412-4380-844b-3f656829c454_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a6c062a540a382a8bbf873e94459f456bc2c5048 --- /dev/null +++ b/2024/SLIM_ Spuriousness Mitigation with Minimal Human Annotations/460c9d54-c412-4380-844b-3f656829c454_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:51b3cfb249256026a2ec4fc148a879681dcef9565a564cec9768e9abbf88cbbb +size 10059437 diff --git a/2024/SLIM_ Spuriousness Mitigation with Minimal Human Annotations/full.md b/2024/SLIM_ Spuriousness Mitigation with Minimal Human Annotations/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8b0bfc9bd7fd4ce5493458ebbbf8ad957976a767 --- /dev/null +++ b/2024/SLIM_ Spuriousness Mitigation with Minimal Human Annotations/full.md @@ -0,0 +1,265 @@ +# SLIM: Spuriousness Mitigation with Minimal Human Annotations + +Xiwei Xuan $^{1}$ , Ziquan Deng $^{1}$ , Hsuan-Tien Lin $^{2}$ , and Kwan-Liu Ma $^{1}$ + +1 University of California, Davis + +$^{2}$ National Taiwan University + +{xwxuan, ziqdeng, klma}@ucdavis.edu, htlin@csie.ntu.edu.tw + +Abstract. Recent studies highlight that deep learning models often learn spurious features mistakenly linked to labels, compromising their reliability in real-world scenarios where such correlations do not hold. Despite the increasing research effort, existing solutions often face two main challenges: they either demand substantial annotations of spurious attributes, or they yield less competitive outcomes with expensive training when additional annotations are absent. In this paper, we introduce SLIM, a cost-effective and performance-targeted approach to reducing spurious correlations in deep learning. Our method leverages a human-in-the-loop protocol featuring a novel attention labeling mechanism with a constructed attention representation space. SLIM significantly reduces the need for exhaustive additional labeling, requiring human input for fewer than $3\%$ of instances. By prioritizing data quality over complicated training strategies, SLIM curates a smaller yet more feature-balanced data subset, fostering the development of spuriousness-robust models. Experimental validations across key benchmarks demonstrate that SLIM competes with or exceeds the performance of leading methods while significantly reducing costs. The SLIM framework thus presents a promising path for developing reliable models more efficiently. Our code is available in https://github.com/xiweix/SLIM.git/. + +Keywords: Model spuriousness $\cdot$ Human-in-the-loop $\cdot$ Data distribution + +# 1 Introduction + +Spurious correlations, where models mistakenly rely on irrelevant features to make decisions, pose a significant challenge in machine learning. Consequently, certain groups may experience deceptively inflated performance metrics, while accuracy reduces for others. This prevalence of "right answers for wrong reasons" fosters a misleading impression of model proficiency and constrains its utility across diverse contexts. Meanwhile, rectifying the issue of spuriousness is a challenging task, particularly when seeking to enhance the performance of underperforming groups. A notable example is the ISIC skin cancer dataset [5], where a typical spurious correlation occurs when color patches frequently co-occur with benign-labeled data. As shown in Fig. 1, a biased model may focus + +![](images/7b1247d216167746b4d9731ef33e662ebef38e44f8d72ee78a1a2d2a7e33efa9.jpg) +Fig. 1: (Left) GradCAM visualizations highlighting model attention on ISIC. (a, b) showcase a model biased towards patches, leading to a correct prediction with wrong reasons in (a) but an incorrect prediction despite correct focus in (b). (c, d) depict a spuriousness-robust model consistently focusing on core features. (Right) The table details the training data distribution by class (benign or malignant) and the presence of color patches, illustrating the imbalance and potential for spurious correlations. + +on patches rather than skin lesions to identify melanoma, and fail when such patches are missing, while a desirable case would be the model captures correct skin regions regardless of spurious features' existence. Enhancing model robustness to address such dangerous misdiagnoses is crucial, driving a surge in research towards more effective solutions. + +A typical approach in this context is to seek for optimal data distribution. GDRO [27] proves the efficacy of group-balanced data in enhancing model robustness. However, along with subsequent studies [14, 22], such an approach requires spuriousness labels at dataset scale, consuming a considerable cost. Our investigations reveal notable inconsistencies in human-annotated spuriousness labels, raising concerns about their relia + +Table 1: Annotation forms and quantities required by different methods on Waterbirds, highlighting the reduced workload of SLIM. (Att. = Attention, T. = Training set, V. = Validation set) + +
MethodAnno. FormAnno. Amount
RRR [26]Map100%×T.
GradIA [9]Att. Yes-No, Map100%×T.
Energy Loss [24]Bounding box100%×T.
CRAYON [15]Att. Yes-No100%×T.
GDRO [27]Spuriousness label100%×T.
DFR [14]Spuriousness label100%×V.
SLIM (ours)Att. Yes-No2.5%×T. or 3%×V.
+ +bility (refer to Table 6). There are other approaches with more direct human involvement, such as providing yes-no feedback on attention correctness [15,36]; or outlining core features with maps [9,26] or bounding boxes [24]. Despite different formats of queried information, they require extensive effort across the entire dataset, as summarized in Table 1. Considering the empirical applicability, our work aims to streamline this process by reducing human annotation workload and minimizing the ambiguity of collected data. Methods requiring spuriousness annotation tend to excel in performance at higher costs, while minimizing such expenses typically results in performance trade-offs [1,6]. Facing this challenge, techniques such as JTT [17] and CNC [41] design training or data-augmenting phases to reduce spuriousness without additional annotations. However, the extra computational cost raised by multiple training iterations is undesired. We believe, as spuriousness is indeed caused by data biases, improving data quality is more essential than designing complex learning strategies. + +Therefore, the remaining research gap calls for a data quality-oriented approach that balances performance and cost. This is where our framework, SLIM, steps in. $SLIM$ offers a novel approach to mitigating spurious correlations by + +intelligently leveraging a minimal set of human-annotated data and curating a feature-balanced data set to enhance model robustness without any significant computational overhead incurred by many existing methods. As shown in Fig. 2, inferred from a reference model, $SLIM$ first constructs an attention representation space that reflects locally consistent data features and model attention, enabling efficient attention label estimation. — Representative instances can be sampled from this space for querying attention correctness labels and such labels can be expanded to neighbor instances. Subsequently, utilizing a visual explanation method that highlights model attention, like GradCAM [29], $SLIM$ disentangles core and environment features within the latent space. Here, "core features" denote features critical for class determination, whereas "environment features" represent the surrounding contextual information. We agree with GDRO that balanced data distribution is crucial, but instead of using human-annotated labels to approximate spuriousness distribution, we leverage latent space features directly disentangled from model attention to balance data distribution, which is more aligned with model understanding and involving less human assumptions. Through this, $SLIM$ assembles data groups with diverse feature combinations. Focused on data quality, a consistency-aware sampling mechanism is employed to curate a subset with optimal feature distributions — ensuring balanced core features across various environmental contexts. Finally, we conclude our framework with a straightforward training procedure. + +The experimental results demonstrate SLIM's advanced capability in spuriousness mitigation by performing comparably to, or exceeding state-of-the-art (SOTA) methods. Notably, SLIM accomplishes this with (1) a significantly reduced requirement for human annotations: $0.12\% - 2.5\%$ of the training set size and $0.35\% - 4\%$ of the validation set; and (2) a lightweight training for a robust model using our constructed data with $5\% - 30\%$ of the original set. Our code will be published after the review process is complete. Our contributions are: + +- Data-quality-oriented spuriousness mitigation: SLIM introduces an interactive, simple, and effective data construction pipeline to create high-quality data for spuriousness mitigation. +- Cost-efficient framework: In contrast to related works with leading performance, SLIM demonstrates a significant reduction in both annotation requirements and computational cost. +- Competitive or superior model performance: Without incurring excessive costs, SLIM achieves comparable or superior performance to existing methods, confirming its effectiveness in mitigating spurious correlations. + +# 2 Related Work + +Spuriousness mitigation with spuriousness labels. When a comprehensive set of spuriousness labels is accessible during training, methods like class balancing [11] and importance weighting [32] can be employed for mitigating spurious correlations, which utilizes data groups defined by class and spuriousness labels. Following the same schema, GDRO [27] refines the training approach + +by focusing on the worst-performing groups as determined by ERM loss. Aiming to reduce the dependency on comprehensive spuriousness labeling, several methods [1,6,17,21,30,32,41] have emerged to approximate group information using a reference model. While these approaches are technically sophisticated, they often fail to achieve the same level of performance as approaches informed by spuriousness labels. To balance the tradeoff between performance and cost, a further research direction reduces the requirement for extensive labeling by leveraging spuriousness labels in smaller validation datasets [14,22]. For example, SSA [22] adopts semi-supervised learning to predict group labels in training data according to information from validation, then applies GDRO to reduce spurious correlations. However, even with this reduced scale of validation sets, the labor-intensive nature of spuriousness labeling remains a barrier to scalability. Moreover, groups in these methods are formed based on human-defined criteria, risking a misalignment with the actual categories a model might discern. This drawback can hinder the true representation of spurious correlations and may lead to suboptimal mitigation outcomes. + +Spuriousness mitigation with human-in-the-loop. Integrating human insights has proven beneficial in mitigating model spuriousness, as evidenced by various human-involved strategies. Considering the valuable human effort, [28] provides an efficient solution by querying data samples for issue mitigation with domain experts in the loop, which overlooks potential issues in the broader data scope. Another option for avoiding redundant human annotation is to query global spurious concepts [31, 38] from knowledgeable experts, which can be ambiguous and fail to cover all possibilities. More recent approaches [10, 24, 26] raise concerns about the inconsistency of spuriousness labels, introducing alternatives such as human-generated attention maps or bounding boxes outlining core features. These human-generated references help align the models' attention with the actual features relevant to decision-making. However, producing such annotations is challenging, and these techniques are resource-intensive as they necessitate annotations across the entire dataset. In response, more recent research research [2, 40] simplifies the annotation process to binary feedback on model's attention correctness. While less demanding, its requirement to annotate all instances still poses challenges in real practices. + +# 3 Problem Formulation + +We introduce the formal definition of the spuriousness mitigation problem [13]. Assume training data $\mathcal{D} = \{(x_i, y_i)\}_{i=1}^n$ with input features $\boldsymbol{x}_i \in \mathbb{R}^d$ and labels $y_i \in [\mathcal{L}]$ , where $\mathcal{L} = \{L_1, \ldots, L_l\}$ are the classes in $\mathcal{D}$ . Machine learning models for classification are trained to minimize the empirical risk: + +$$ +\boldsymbol {\theta} ^ {*} \in \underset {\boldsymbol {\theta}} {\arg \min } \mathbb {E} _ {(\boldsymbol {x} _ {i}, y _ {i}) \in \mathcal {D}} [ l (f (\boldsymbol {\theta}, \boldsymbol {x} _ {i}), y _ {i}) ], \tag {1} +$$ + +where $\pmb{\theta}$ represent model parameters, $f(\pmb{\theta},\pmb{x}_i)$ and $l(*)$ are the model outputs and the loss function, respectively. + +Groups. In spuriousness mitigation, we assume there exists a core feature set $\mathcal{C} = \{c_1,\dots ,c_l\}$ and a spurious feature set $\mathcal{S} = \{s_1,\dots ,s_k\}$ in $\mathcal{D}$ . Each class $L_{i}$ is uniquely identified by a core feature $c_{i}$ , while spurious features $s_j$ can be found across different classes. We consider an instance in the dataset as having both a core feature $c_{i}$ and a spurious feature $s_j$ , and each instance belongs to a unique group $g_{c_i,s_j}$ . Core features are reliable class indicators in both training and testing sets, while spurious features with misleading co-occurrence with class in training may miss such correlations in the testing phase. If the model mistakenly links spurious features with classes in training, it may fail in testing when such spurious features are absent. This can lead to lower accuracy rates, especially for less common group variations that do not exhibit the spurious feature the model has learned to rely on. + +Worst-group accuracy. The vital evaluation of spuriousness mitigation focuses on the model's performance in the subgroup where it is least effective – the worst-group accuracy. It is calculated by: + +$$ +A c c _ {w} = \min _ {g \in \mathcal {G}} \left(\frac {1}{| g |} \sum_ {\left(x _ {i}, y _ {i}\right) \in g} \mathbf {1} [ f \left(x _ {i}\right) = y _ {i} ]\right), \tag {2} +$$ + +where $\mathcal{G} = \{\cup_{c_i,s_j}g_{c_i,s_j}\}$ is the union of subgroups defined by core feature $c_{i}$ and spurious feature $s_j$ , and $f(x_{i})$ is the label predicted by the model. This metric is crucial because it directly assesses how effectively a model can generalize across different subgroups, especially those that might be mostly underrepresented because of spurious correlations. It ensures that improvements in model performance are not just reflective of gains on the majority or less challenging subsets of the data. + +# 4 Method + +In this section, we detail the method behind SLIM. In alignment with related works requiring no spuriousness label, our approach also follows a two-stage procedure: inferring groups with an ERM-trained reference model $f_{r}$ , and then leveraging the inferred groups to train a robust model. As an approach oriented by data quality, SLIM framework comprises a three-phase data construction pipeline, as illustrated in Fig. 2. Firstly, we build an attention representation space to group instances with both similar data features and model attention, preparing for the data sampling and label propagation. Following this, we sample typical instances for human to evaluate model's attention correctness with yes-no feedback, then estimate neighboring instances' attention labels accordingly. In the last phase, we filter out data with wrong attention and leverage attention matrix to disentangle core and environment features. With constructed feature sets, we apply consistency-weighted data sampling to curate a feature-balanced subset, with a desired core-environment balanced distribution. Finally, SLIM framework is concluded by a training process with the curated subset, resulting in a model exhibiting robustness against spurious correlations. + +![](images/0cdc12ac114e2af3890c359276af5715a31132619d73866b44116d1be9cfcb99.jpg) +Fig. 2: Overview of SLIM framework, with a data construction pipeline consisting of three phases: (1) Attention Space Construction, creating a space with data features and model attention aligned locally; (2) Attention Annotation and Expansion, where instances sampled from the attention space are labeled by human for attention correctness, and labels are propagated to neighboring instances; and (3) Balanced Data Curation, which filters out instances with incorrect attention, and utilize attention-weighted $(F_A)$ and inverse-attention-weighted $(F_{\bar{A}})$ feature vectors to create core and environment feature sets, forming subgroups to assemble a feature-balanced subset for training a spuriousness-robust model. + +# 4.1 Attention Space Construction + +Considering the benefits of being aware of spuriousness information and drawbacks of the annotation cost, in our framework, we deploy a binary labeling scheme to simplify descriptive labeling tasks into yes-no feedback on model's attention correctness. Moreover, we aim to further facilitate this process by avoiding the necessity of exhaustive labeling. Our empirical observations indicate that a model would have consistent attention mechanisms on specific data subgroups. If such subgroups can be found, it would largely aid the labeling workload by enabling label estimations among neighbors. + +Driven by this, we aim to find a representation space where instances with similar attention outcomes (as indicated by highlighted image features) are closely grouped. However, our analysis reveals that conventional feature representation spaces formed directly from feature vectors do not fulfill this requirement - such spaces only maintain local consistency in data features but fail to capture the nuances of model attention (refer to Sec. 5.3). + +Thus, we have developed a process utilizing model attention in the latent space to support the establishment of such space, as illustrated in Fig. 3. For each image, we obtain its feature vector $F$ and model attribution $A$ in the latent space at first. Note that $A$ represents the model's attention focus and can be linearly interpolated to visualize an attention heatmap. We then use $A$ as a weight matrix for $F$ , producing an attention-weighted feature vector $F_{A}$ . Finally, with the dimensionality reduction method UMAP [20], we project vectors $F_{A}$ into a 2D space, namely the attention space, which ensures local consistency encompasses both data features and the model's attention mechanisms. + +![](images/a052730a8a30e45226e81fecedbdef87655946b17163286dbcb43f04a1db1175.jpg) +Fig. 3: Construction of the attention space. In the latent space, feature vector $F$ and model's attribution vector $A$ are extracted, representing input features and model's attention, representatively. By weighting $F$ with $A$ , the attention-weighted feature vector $F_{A}$ emphasizes features of model's top interest. All $F_{A}$ are then projected to form the attention space featuring locally consistent features and model attention. + +Note that for a fair comparison, in alignment with baseline methods, experiments in this paper involve the CNN model (ResNet) for classification and GradCAM for obtaining model attention. But our framework is fully applicable to (1) other vision model architectures such as vision transformers (ViT) [8] or Swin transformers [18], and (2) other heatmap-based model visualization techniques, such as RISE [23] or ScoreCAM [33]. + +# 4.2 Attention Annotation and Expansion + +Given that our attention space ensures locally consistent model attention, it allows for annotating a selectively sampled subset of instances and then estimating neighboring labels accordingly, thereby enabling efficient transfer of human expertise. The criteria for selecting these instances are twofold: (1) diversity, to encompass a broad covering of various data features and model attention patterns; (2) typicality, ensuring they are representative of their neighboring instances. To achieve this, we utilize the k-means clustering algorithm to partition our constructed attention space into $n$ clusters (diversity). We then select $n$ instances that are closest to each cluster's center (typicality). The choice of $n$ is aimed at optimizing the balance between minimizing intra-cluster variance in data features and model attention patterns, and keeping the annotation budget manageable. Our empirical analysis, detailed in Sec. 5.3, reveals that the intra-cluster consistencies increase with $n$ , but this trend levels off upon reaching a specific value of $n$ -the elbow point. Thus, selecting $n$ at this elbow point optimizes consistency without unnecessarily increasing the annotation workload. + +We gather binary yes-no feedback from human experts on the correctness of the model's attention for these $n$ instances, feeding this feedback along with our attention space embedding into the label spreading algorithm. This algorithm leverages data similarity and density to extend attention labels to previously unlabeled data [43]. Ultimately, this process yields a probability score for correct attention for each instance across the dataset. + +# 4.3 Balanced Data Curation + +Following Sec. 4.2 that provides attention accuracy scores, we proceed to construct our data subset to support model's unbiased learning, as described below: + +Data screening. Inspired by works [3,7,12], our theoretical analysis further supports the premise that data receiving correct attention from a reference model are more likely to enhance the learning of core features in a robust model within a more balanced data distribution, as substantiated in Appx. A.2. Furthermore, our study indicates that even a reference model with inherent biases has the potential to accurately identify core features for a data subset. Such correctly captured features typically exhibit diversity, ensuring good coverage across the spectrum of the core feature set. This understanding informs our approach to curating a data subset that: (1) selects instances based on accurate model attention, and (2) ensures a balanced representation of core features across various environment features. Thus we first apply data screening to filter out instances with wrong attention according to attention accuracy scores. + +Disentangling core and environment features. After screening, the retained data subset mainly comprises instances with correct attention. This is reflected by attention masks that highlight core features, meaning that the corresponding attribution vector $A$ in the latent space also highlights core features of feature vector $F$ . Moreover, existing studies have found that a model's feature vectors are representative of most input features, regardless of whether they are used for decision-making [14, 37]. Therefore, utilizing $A$ as the weight matrix, we disentangle core and environment features in the latent space — weighting $F$ by attention masks $A$ and inverse attention masks $\bar{A}$ . Subsequently, k-means clustering is applied to partition all attention-weighted feature vectors $F_{A}$ , and inverse-attention-weighted feature vectors $F_{\bar{A}}$ , respectively, facilitating the formation of representative subgroups for the core feature set $(\hat{C} = \hat{c}_1,\dots ,\hat{c}_N)$ and the environment feature set $(\mathcal{E} = e_1,\ldots ,e_M)$ . The number of clusters, determined through the elbow method, culminates in $M\times N$ subgroups $g_{\hat{c}_i,e_j}$ , each characterized by a distinct combination of core and environment features $(\hat{c}_i,e_j)$ . Data sampling. Given the unequal distribution of core and environment features, the internal consistency across each constructed subgroup $g_{\hat{c}_i,e_j}$ varies. Our sampling strategy aims to mitigate over-representation of particular features to prevent overfitting, while ensuring broad coverage of diverse core and environment features. We assess cluster consistency as a metric for determining the sample size from each cluster. For a cluster $\hat{c}_i$ in the core feature space, denote the distance between an instance $x$ and the cluster center as $d_{\hat{c}_i}(x)$ . The cluster consistency is defined as the reciprocal of the average distance per cluster: + +$$ +\rho_ {\hat {c} _ {i}} = \frac {1}{\frac {1}{| \hat {c} _ {i} |} \sum_ {x \in \hat {c} _ {i}} d _ {\hat {c} _ {i}} (x)}. \tag {3} +$$ + +A lower $\rho_{\hat{c}_i}$ signals less consistent features, necessitating a higher sample weight for adequately representing them. A cluster's sampling weight is defined as $w_{\hat{c}_i} = \frac{1}{\rho_{\hat{c}_i}}$ and the number of samples to be drawn from a core cluster $\hat{c}_i$ is: + +$$ +n _ {\hat {c} _ {i}} = N \cdot \frac {1 / \rho_ {\hat {c} _ {i}}}{\sum_ {j = 1} ^ {| \hat {C} |} 1 / \rho_ {\hat {c} _ {j}}}, \tag {4} +$$ + +where $N$ is the sampling budget. Moving to the environment feature space, the $n_{\hat{c}_i}$ instances from a core feature cluster $\hat{c}_i$ may distribute across multiple environment clusters $\{e_1,\dots ,e_M\}$ and formulate subgroups $\{g_{\hat{c}_i,e_1},\ldots ,g_{\hat{c}_i,e_M}\}$ . A similar methodology is employed for a certain group $g_{\hat{c}_i,e_i}$ . The average distance for $g_{\hat{c}_i,e_i}$ in the environment representation space is: + +$$ +\rho_ {\hat {c} _ {i}, e _ {i}} = \frac {1}{\frac {1}{| \hat {c} _ {i} , e _ {i} |} \sum_ {x \in (\hat {c} _ {i} , e _ {i})} d _ {e _ {i}} (x)}. \tag {5} +$$ + +Similarly, we have the sampling weight $w_{\hat{c}_i,e_i} = \frac{1}{\rho_{\hat{c}_i,e_i}}$ and the number of samples to be drawn from a certain group $g_{\hat{c}_i,e_i}$ is finally defined by: + +$$ +n _ {\hat {c} _ {i}, e _ {i}} = n _ {\hat {c} _ {i}} \cdot \frac {w _ {\hat {c} _ {i} , e _ {i}}}{\sum_ {j = 1} ^ {| \mathcal {E} (\hat {c} _ {i}) |} w _ {\hat {c} _ {j} , e _ {i}}} = N \cdot \frac {1 / \rho_ {\hat {c} _ {i}}}{\sum_ {j = 1} ^ {| \hat {\mathcal {C}} |} 1 / \rho_ {\hat {c} _ {j}}} \cdot \frac {1 / \rho_ {\hat {c} _ {i} , e _ {i}}}{\sum_ {j = 1} ^ {| \mathcal {E} (\hat {c} _ {i}) |} 1 / \rho_ {\hat {c} _ {i} , e _ {j}}}, \tag {6} +$$ + +where $|\hat{\mathcal{C}}|$ denotes the number of core clusters and $|\mathcal{E}(\hat{c}_i)|$ denotes the number of subgroups formulated by instances from cluster $\hat{c}_i$ . To summary, our sampling strategy takes feature consistencies and sampling budget into consideration, and ensures fair representation of different core-environment feature combinations across the feature spaces. + +# 4.4 Training + +In our pursuit to mitigate spurious correlations, we emphasize the pivotal role of data quality by ensuring a balanced distribution of core features across various environmental contexts within our dataset. Given this focus, we employ a basic training approach for our model to learn from this well-prepared dataset. + +We apply our data curation pipeline separately to the training $(SLIM_{Tr})$ and validation $(SLIM_{Val})$ sets. For the data constructed from the training set, we adopt the conventional ERM. For the dataset derived from the validation set, we employ the training technique presented by DFR [14]. It involves retraining only the final linear layer of a classification model that was initially trained using ERM. Note that our approach diverges from DFR practices by utilizing our curated dataset instead of a metagroup-balanced dataset. + +# 5 Experiments + +# 5.1 Experimental Setup + +Datasets. Our study explores a range of datasets with inherent spurious correlations to challenge image classification models, including: (1) Waterbirds [27]. Contrasting landbirds and waterbirds against congruent/incongruent backgrounds, this dataset tests models' ability to focus on relevant features for bird type classification. (2) CelebA [19]. A comprehensive celebrity image dataset where the focus is on distinguishing hair color (blond vs. non-blond) to examine + +biases in associating hair color with gender. (3) ISIC [5]. Aiming to differentiate benign from malignant melanoma in dermoscopic images, we target colorful patches as spurious features, aligning with previous studies [25]. (4) NICO. Derived from $\mathrm{NICO}++$ [42], this dataset features various object categories in shifted contexts to probe spurious correlations. (5) ImageNet-9 [35]. This dataset is crafted from ImageNet to test models' robustness against background variations. Detailed information is available in Appx. A.4. + +Baselines. We benchmark SLIM against a spectrum of SOTA spuriousness mitigation techniques, categorized as follows: (1) Techniques not reliant on spuriousness labels: EIIL [6], PGI [1], GEORGE [30], LfF [21], CIM [32], JTT [17], and CNC [41]; (2) Techniques requiring additional information such as predefined spuriousness concepts: DISC [34]; (3) Methods utilizing spuriousness labels from validation sets: SSA [22] and DFR [14]; (4) Approaches utilizing spuriousness labels from training sets: GDRO [27] and PDE [7]. + +Evaluation. Aligning with baselines, our experiments involve CNN models with ResNet50 architecture and GradCAM for visualizing model attention. For the capability of performance enhancement, we evaluate the worst-group and average accuracy. Additionally, we consider the necessity for extra labels and the additional (add.) training cost to evaluate computational efficiency. Note that the add. training cost is aligned with [13, 39], measuring add. data amounts required for second-stage training, where 1 means the amount of data equals the training set size. Furthermore, we examine the attention precision to evaluate whether model attention is corrected. All quantification experiments are based on five independent runs with different random seeds. + +More experimental results with other architecture (ViT) and datasets (MetaShift [16] and FMoW [4]) are provided in Appx. A.8. + +Table 2: Worst-group and average accuracy (%) comparison on binary classification tasks, highlighting best and second-best performances. $\spadesuit$ and $\diamond$ indicate methods requiring spuriousness concepts and attention labels, respectively. + +
MethodSpuriousness labelAdd. training costWaterbirdsCelebAISIC
WorstAvgWorstAvgWorstAvg
ERMX0×62.6±0.397.3±1.047.7±2.194.9±0.356.7±2.279.0±1.3
EIILX0×83.5±2.894.2±1.381.7±0.885.7±0.172.3±3.986.4±1.7
PGIX0×79.5±1.995.5±0.885.3±0.387.3±0.167.4±1.985.7±1.1
GEORGEX1×76.2±2.095.7±0.554.9±1.994.6±0.258.8±2.183.2±1.4
LfFX1×77.3±2.391.4±1.776.8±1.284.5±0.772.1±2.185.4±0.5
CIMX1×77.8±1.594.7±1.183.8±1.090.5±0.671.7±1.685.9±1.1
JTTX5×-6×83.1±3.590.6±0.381.5±1.788.1±0.328.1±4.386.2±0.3
CNCX2×-12×88.5±0.390.9±0.188.8±0.989.9±0.568.2±0.284.6±0.2
DISCX◇1×88.7±0.793.8±0.785.7±1.189.9±0.773.4±0.484.7±0.4
SLIMTrX◇0.1×-0.2×89.1±0.691.7±0.589.4±0.390.1±0.279.1±0.588.5±0.4
SLIMValX◇0.05×-0.3×91.0±0.493.8±0.489.4±0.590.7±0.274.0±0.285.4±0.2
SSAvalidation1.5×-5×89.0±0.692.2±0.989.8±1.392.8±0.161.7±0.777.7±1.0
DFRvalidation0.05×-0.3×89.6±0.792.1±0.287.3±1.090.2±0.871.9±0.784.3±0.3
GDROtraining0×89.9±0.692.6±0.188.9±1.393.9±0.166.8±3.580.5±1.9
PDEtraining0×85.9±2.790.5±1.885.0±4.092.3±0.866.4±2.382.2±1.3
+ +Table 3: Worst-group and average accuracy (%) comparison on multi-classification tasks, highlighting best and second-best results. $\diamond$ indicates methods requiring attention labels. - represents inapplicable methods because such methods require spuriousness labels while the dataset does not provide them. + +
MethodSpuriousness labelAdd. training costNICOImageNet9
WorstAvgWorstAvg
ERMX0×11.9±0.270.9±1.255.3±2.075.6±1.1
CNCX5×-12×72.6±1.789.5±0.658.9±0.674.3±0.2
SLIMTrX◇0.1×-0.2×78.0±0.892.1±0.168.1±0.374.5±0.2
SLIMValX◇0.1×-0.2×73.8±0.490.7±0.264.6±0.474.2±0.3
DFRvalidation0.4×70.9±2.185.8±0.6--
GDROtraining0×73.3±1.090.3±0.5--
+ +Table 4: SLIM's primary cost, including the data amount required for attention annotation (Att.), and the constructed data size for model training (Training). + +
MethodWaterbirdCelebAISICNICOImageNet 9
Att.TrainingAtt.TrainingAtt.TrainingAtt.TrainingAtt.Training
SLIMTr120 (2.5%)0.1×200 (0.12%)0.1×120 (0.7%)0.2×40 (1.8%)0.2×180 (0.4%)0.1×
SLIMVal40 (3%)0.1×70 (0.35%)0.05×40 (2%)0.3×20 (2.7%)0.2×50 (1%)0.1×
+ +# 5.2 Spuriousness Mitigation Performance + +In this section, we present the performance evaluation by comparing different methods across multiple binary and multiclass classification datasets mentioned in Sec. 5.1. Key findings are demonstrated below. + +Superior worst-group performance. The metric of worst-group accuracy $(Acc_w)$ is pivotal for spuriousness mitigation methods (refer to Sec. 3), and the results in Tables 2 and 3 showcase SLIM's superior capability in this regard. Across various classification tasks, SLIM consistently outperforms competing methods in achieving the highest $Acc_w$ while maintaining high average accuracy. Specifically, in binary classification scenarios (Table 2), SLIM leads in performance for Waterbirds and ISIC and secures a close second in CelebA. In multi-classification tasks (Table 3), $SLIM_{Tr}$ and $SLIM_{Val}$ rank as the top performers across the board, underscoring the effectiveness of our approach in spuriousness mitigation. + +Cost-efficient spuriousness mitigation. Methods in Tables 2 and 3 are divided into four groups from top to bottom, as per the categorization established in Sec. 5.1 according to the computational cost. The "Add. training cost" columns denote the additional training cost for mitigating spuriousness apart from the initial training, such as additional learning phases (CNC, SSA) or data augmentation (JTT). Results in the table substantiate that methods with higher costs (higher Add. training costs or more spuriousness labels) tend to outperform others, aligning with prior research findings. Nevertheless, among the well-performed methods, it is worth mentioning that $SLIM$ is more cost-efficient: $SLIM_{Tr}$ used 0.1-0.2 proportion of training sets and $SLIM_{Val}$ used 0.05-0.3 proportion of validation sets for additional training, which is the same as DFR and much smaller than others. Moreover, $SLIM$ requires labels for attention correctness, which is more cost-efficient to obtain than spuriousness labels (refer to Sec. 5.3 for more details). Table 4 presents $SLIM$ 's primary cost, including the training data amount for mitigating spuriousness, which is the detailed version of "Add. train + +![](images/339c57f4cfb0e174ef20e2d2c3c0ebc8ecbdeda076aa133cd9baeb2516005855.jpg) + +Table 5: Attention accuracy evaluation. (Left) GradCAM visualization of model attention. (Right) Worst group and average attention accuracy of selected methods. + +
WaterbirdsAIoU
WorstAvg
ERM.25±.05.47±.03
CNC.48±.05.49±.04
GDRO.49±.03.52±.03
PDE.50±.02.52±.03
SLIMTr.53±.02.61±.02
ImageNet 9AIoU
WorstAvg
ERM.33±.07.55±.03
CNC.44±.03.61±.03
SLIMTr.58±.04.68±.03
+ +ing cost" in Tables 2 and 3, and the data amount requiring attention labels (Att.). Notably, our approach only requires no more than $3\%$ of the corresponding data split (train or validation) for attention-label annotation, further proving the cost-efficiency of SLIM. + +Enhanced attention accuracy. Beyond classification accuracy, we further evaluate the performance of spuriousness mitigation methods via attention accuracy. Specifically, for models resulting from each method, we randomly sample data instances to visualize model attention, and measure attention accuracy scores in both worst-group and dataset levels, consistent with our evaluation of classification accuracy. Align with the related work [40], we employ the Adjusted Intersection-over-Union (AIoU) score as our attention accuracy measurement, which is an adjusted IoU score with a threshold, aiming at measuring the overlap between model's highest attributed region and the ground-truth bounding box. Details related to AIoU are given in Appx. A.7. Table 5 presents the evaluation results. The left GradCAM examples showcase SLIM's capability of correcting model's wrong attention, and the right table compares SLIM with SOTA methods, where we can find our method secures the best performance in terms of both worst-group and average attention accuracy. + +# 5.3 Attention Labeling Performance + +To compare the processes of annotating spuriousness versus attention correctness, we employed crowdsourcing tasks using the Waterbirds and NICO datasets, as both of which come with ground-truth spuriousness labels. For this study, we selected a random set of 120 images from each dataset and established two separate tasks: one for spuriousness labeling and one for attention correctness labeling. For each task, we recruited 60 workers independently from the Prolific platform to prevent learning biases from cross-task participation. We meticulously recorded the time taken to label each image and reviewed all submissions to exclude any outliers with unusually long or short annotation times. Ensuring ethical research standards, our study avoided collecting personally identifiable information and excluded any potentially offensive content. Detailed instructions provided to the crowdsourcing participants can be found in Appx. A.6. + +Evaluation Consistency. Table 6 presents metrics for this study. While the annotation time for Waterbirds remains similar across tasks, attention correctness for NICO is annotated more quickly than the other. This difference may stem + +Table 6: Comparison of crowdsourced labeling tasks: spuriousness vs. attention correctness. + +
WaterbirdsAvg. TimeAvg. Consis.
Spuriousness4.9±2.3 sec81.9±14.8%
Att.4.5±1.4 sec94.6±10.5%
NICOAvg. TimeAvg. Consis.
Spuriousness7.9±3.8 sec70.1±19.1%
Att.5.3±1.6 sec95.1±6.1%
+ +![](images/0e4c5b3e9219ea3099038f43ce33447737c73f5a34bbff9735ddbce52d86de9b.jpg) +Fig. 4: Comparison of neighbor consistency over original feature space and our attention space on Waterbirds, w.r.t. model attention (left) and data features (right). + +from NICO's broader array of spuriousness choices versus the binary attention annotations. To measure the annotation consistency—whether a single receives divergent labels—we use the percentage of the most frequent label as a consensus measure. A higher percentage suggests greater consistency. Our findings indicate that attention correctness annotations exhibit a notably higher consistency, suggesting that it is less ambiguous to determine correct attention than to identify spurious features. + +Attention Consistency. $SLIM$ requires spreading labels in a space where attention patterns are consistently similar in nearby areas. To evaluate its feasibility, we compared how similar neighbors are in the original feature space versus the attention space created by $SLIM$ . We used k-means clustering with varying numbers of clusters to measure the intra-cluster similarity for both the model's attention matrices and the feature vectors. A higher similarity score means the space has more consistent patterns, which is essential for the accurate spreading of labels. As Fig. 4 shows, our attention space significantly improves consistency in the model's attention, while also maintaining a good level of data feature consistency. Increasing the number of clusters generally increases consistency, but there is a point where adding more clusters does not lead to further gains, as shown by the curve flattening at $k = 120$ in the left plot. + +In summary, we have demonstrated that the annotation module of SLIM is less time-consuming and yields more consistent labeling feedback. Additionally, the attention space engineered by SLIM establishes a solid foundation for propagating attention labels, considering the consistent model attention observed among neighboring instances. + +# 5.4 Ablation Study + +We investigate two critical factors of SLIM, which are detailed below. Other factors such as the dimension of the attention space are discussed in Appx. A.8. Influence of data amount for attention annotations. We change the data amount for attention annotations to investigate their effect on performance, keeping other variables constant. Table 7 illustrates that both the worst-group and average accuracy tend to rise with more annotations, while such incremental benefit stabilizes after reaching a certain point, helping us identify an optimal number of annotations that balance performance with annotation cost. + +Influence of training set size. This size refers to the size of the data subset constructed in SLIM's final phase for training a robust model. By only varying + +Table 7: Performance variation with different attention annotation amounts in SLIM. Marks in table: best and second best performance; the amounts used in Table 2. + +
MethodWaterbirdsCelebAISIC
AmountWorstAvgAmountWorstAvgAmountWorstAvg
SLIMTr6078.2±0.586.0±0.310086.4±0.290.2±0.16075.3±0.286.6±0.9
9085.9±0.989.3±0.315087.8±0.491.1±0.79074.4±0.487.4±0.7
12089.1±0.691.7±0.520089.4±0.490.1±0.212079.1±0.588.5±0.4
15089.4±0.691.7±0.225089.4±0.290.9±0.115079.9±0.888.8±0.7
18089.3±0.391.0±1.530089.5±0.191.2±0.218079.8±0.388.6±0.3
SLIMVal1078.9±8.786.5±5.14084.7±0.487.4±0.31070.5±0.583.3±1.2
2589.2±0.693.4±1.05587.6±0.390.4±0.62572.7±0.185.2±0.8
4091.0±0.493.8±0.47089.4±0.490.7±0.24074.0±0.285.4±0.2
5591.5±0.393.8±0.18589.7±0.190.5±0.45574.0±0.285.9±0.5
7090.7±0.394.0±0.310089.7±0.290.5±0.27074.1±0.285.8±0.2
+ +Table 8: Performance variation with different training sizes. $SLIM_{Tr}$ and $SLIM_{Val}$ are trained on different proportions (Prop) of training sets and validation sets, respectively. + +
MethodWaterbirdsCelebAISIC
Prop%WorstAvgProp%WorstAvgProp%WorstAvg
SLIMTr582.4±1.083.6±1.3588.9±0.190.0±0.11577.6±2.187.1±1.2
1089.1±0.691.7±0.51089.4±0.390.1±0.22079.1±0.588.5±0.4
1282.3±1.387.8±0.31589.4±0.491.0±0.42577.4±1.588.2±1.2
1583.5±2.789.6±0.52088.9±0.190.8±0.33078.3±1.588.3±0.9
SLIMVal584.2±3.689.2±2.6385.7±1.388.5±1.12568.2±2.181.9±0.8
1091.0±0.493.8±0.4589.4±0.490.7±0.23074.0±0.285.4±0.2
1287.6±0.693.9±1.31088.3±1.490.3±0.53572.5±0.585.3±0.1
1586.8±2.094.6±0.51587.2±0.490.3±0.44070.9±1.285.4±3.5
+ +this parameter (Table 8), we find that the optimal size is inherently linked to the size of our constructed subgroups, $|g_{\hat{c}_i,e_i}|$ . An optimal size is reached when data from various groups are sufficiently sampled without redundant oversampling. These findings guide the hyperparameter selection in this phase, ensuring an effective cost allocation. + +# 6 Conclusion + +In this paper, we introduce SLIM, a novel framework for spuriousness mitigating in deep learning. With minimal human annotation effort, SLIM provides a data construction pipeline to create a feature-balanced subset for training a robust model. Extensive evaluation demonstrates the superior cost-effectiveness of SLIM, which matches or outperforms state-of-the-art methods with far less costs. The success of SLIM underscores the potential of data quality and efficient human supervision in the development of robust models. We believe that SLIM will inspire further research in this direction, ultimately leading to more cost-effective machine learning systems for reliable AI. + +Acknowledgements. This research is supported in part by Bosch Research and the National Institute of Health with grants P41-EB032840 and R01CA270454. H.-T. Lin is partially supported by the National Taiwan University Center for Data Intelligence via NTU-112L900901 and the Ministry of Science and Technology in Taiwan via MOST 112-2628-E-002-030. + +# References + +1. Ahmed, F., Bengio, Y., Van Seijen, H., Courville, A.: Systematic generalisation with group invariant predictions. In: International Conference on Learning Representations (2020) +2. Bontempelli, A., Teso, S., Tentori, K., Giunchiglia, F., Passerini, A.: Concept-level debugging of part-prototype networks. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=oiwXwPDTyNk +3. Chen, Z., Deng, Y., Wu, Y., Gu, Q., Li, Y.: Towards understanding the mixture-of-experts layer in deep learning. Advances in neural information processing systems 35, 23049-23062 (2022) +4. Christie, G., Fendley, N., Wilson, J., Mukherjee, R.: Functional map of the world. In: CVPR (2018) +5. Codella, N.C.F., Gutman, D., Celebi, M.E., Helba, B., Marchetti, M.A., Dusza, S.W., Kalloo, A., Liopyris, K., Mishra, N., Kittler, H., Halpern, A.: Skin lesion analysis toward melanoma detection: A challenge at the 2017 international symposium on biomedical imaging (isbi), hosted by the international skin imaging collaboration (isic). In: 2018 IEEE 15th International Symposium on Biomedical Imaging (ISBI 2018). pp. 168-172 (2018). https://doi.org/10.1109/ISBI.2018.8363547 +6. Creager, E., Jacobsen, J.H., Zemel, R.: Environment inference for invariant learning. In: International Conference on Machine Learning. pp. 2189-2200. PMLR (2021) +7. Deng, Y., Yang, Y., Mirzasoleiman, B., Gu, Q.: Robust learning with progressive data expansion against spurious correlation. In: Thirty-seventh Conference on Neural Information Processing Systems (2023), https://openreview.net/forum?id=9QEVJ9qm46 +8. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) +9. Gao, Y., Sun, T.S., Zhao, L., Hong, S.R.: Aligning eyes between humans and deep neural network through interactive attention alignment. Proc. ACM Hum-. Comput. Interact. 6(CSCW2) (nov 2022). https://doi.org/10.1145/3555590, https://doi.org/10.1145/3555590 +0. Gao, Y., Sun, T.S., Zhao, L., Hong, S.R.: Aligning eyes between humans and deep neural network through interactive attention alignment. Proceedings of the ACM on Human-Computer Interaction 6(CSCW2), 1-28 (2022) +1. He, H., Garcia, E.A.: Learning from imbalanced data. IEEE Transactions on knowledge and data engineering 21(9), 1263-1284 (2009) +2. Jelassi, S., Li, Y.: Towards understanding how momentum improves generalization in deep learning. In: International Conference on Machine Learning. pp. 9965-10040. PMLR (2022) +3. Joshi, S., Yang, Y., Xue, Y., Yang, W., Mirzasoleiman, B.: Towards mitigating spurious correlations in the wild: A benchmark & a more realistic dataset. arXiv preprint arXiv:2306.11957 (2023) +4. Kirichenko, P., Izmailov, P., Wilson, A.G.: Last layer re-training is sufficient for robustness to spurious correlations. arXiv preprint arXiv:2204.02937 (2022) +5. Lee, S., Payani, A., Chau, D.H.P.: Towards mitigating spurious correlations in image classifiers with simple yes-no feedback. In: AI and HCI Workshop at the 40th International Conference on Machine Learning (2023) + +16. Liang, W., Zou, J.: Metashift: A dataset of datasets for evaluating contextual distribution shifts and training conflicts. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=MTex8qKavoS +17. Liu, E.Z., Haghgoo, B., Chen, A.S., Raghunathan, A., Koh, P.W., Sagawa, S., Liang, P., Finn, C.: Just train twice: Improving group robustness without training group information. In: International Conference on Machine Learning. pp. 6781-6792. PMLR (2021) +18. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021) +19. Liu, Z., Luo, P., Wang, X., Tang, X.: Deep learning face attributes in the wild. In: Proceedings of the IEEE international conference on computer vision. pp. 3730-3738 (2015) +20. McInnes, L., Healy, J., Melville, J.: Umap: Uniform manifold approximation and projection for dimension reduction. arXiv preprint arXiv:1802.03426 (2018) +21. Nam, J., Cha, H., Ahn, S., Lee, J., Shin, J.: Learning from failure: De-biasing classifier from biased classifier. Advances in Neural Information Processing Systems 33, 20673-20684 (2020) +22. Nam, J., Kim, J., Lee, J., Shin, J.: Spread spurious attribute: Improving worst-group accuracy with spurious attribute estimation. arXiv preprint arXiv:2204.02070 (2022) +23. Petsiuk, V., Das, A., Saenko, K.: Rise: Randomized input sampling for explanation of black-box models. arXiv preprint arXiv:1806.07421 (2018) +24. Rao, S., Böhle, M., Parchami-Araghi, A., Schiele, B.: Using explanations to guide models. arXiv preprint arXiv:2303.11932 (2023) +25. Rieger, L., Singh, C., Murdoch, W.J., Yu, B.: Interpretations are useful: Penalizing explanations to align neural networks with prior knowledge. In: Proceedings of the 37th International Conference on Machine Learning. ICML'20, JMLR.org (2020) +26. Ross, A.S., Hughes, M.C., Doshi-Velez, F.: Right for the right reasons: Training differentiable models by constraining their explanations. In: Proceedings of the Twenty-Sixth International Joint Conference on Artificial Intelligence, IJCAI-17. pp. 2662-2670 (2017). https://doi.org/10.24963/ijcai.2017/371, https://doi.org/10.24963/ijcai.2017/371 +27. Sagawa*, S., Koh*, P.W., Hashimoto, T.B., Liang, P.: Distributionally robust neural networks. In: International Conference on Learning Representations (2020), https://openreview.net/forum?id=ryxGuJrFvS +28. Schramowski, P., Stammer, W., Teso, S., Brugger, A., Herbert, F., Shao, X., Luigs, H.G., Mahlein, A.K., Kersting, K.: Making deep neural networks right for the right scientific reasons by interacting with their explanations. Nature Machine Intelligence 2(8), 476-486 (2020) +29. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Gradcam: Visual explanations from deep networks via gradient-based localization. In: Proceedings of the IEEE international conference on computer vision. pp. 618-626 (2017) +30. Sohoni, N., Dunnmon, J., Angus, G., Gu, A., Ré, C.: No subclass left behind: Fine-grained robustness in coarse-grained classification problems. Advances in Neural Information Processing Systems 33, 19339-19352 (2020) +31. Srivastava, M., Hashimoto, T., Liang, P.: Robustness to spurious correlations via human annotations. In: International Conference on Machine Learning. pp. 9109-9119. PMLR (2020) + +32. Taghanaki, S.A., Choi, K., Khasahmadi, A.H., Goyal, A.: Robust representation learning via perceptual similarity metrics. In: International Conference on Machine Learning. pp. 10043-10053. PMLR (2021) +33. Wang, H., Wang, Z., Du, M., Yang, F., Zhang, Z., Ding, S., Mardziel, P., Hu, X.: Score-cam: Score-weighted visual explanations for convolutional neural networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 24-25 (2020) +34. Wu, S., Yuksekgonul, M., Zhang, L., Zou, J.: Discover and cure: Concept-aware mitigation of spurious correlation. In: Proceedings of the 40th International Conference on Machine Learning. ICML'23, JMLR.org (2023) +35. Xiao, K., Engstrom, L., Ilyas, A., Madry, A.: Noise or signal: The role of image backgrounds in object recognition. ArXiv preprint arXiv:2006.09994 (2020) +36. Xuan, X., Ono, J.P., Gou, L., Ma, K.L., Ren, L.: Attributionscanner: A visual analytics system for metadata-free data-slicing based model validation. arXiv preprint arXiv:2401.06462 (2024) +37. Xue, Y., Payani, A., Yang, Y., Mirzasoleiman, B.: Eliminating spurious correlations from pre-trained models via data mixing. arXiv preprint arXiv:2305.14521 (2023) +38. Yan, S., Yu, Z., Zhang, X., Mahapatra, D., Chandra, S.S., Janda, M., Soyer, P., Ge, Z.: Towards trustable skin cancer diagnosis via rewriting model's decision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11568-11577 (2023) +39. Yang, Y., Gan, E., Karolina Dziugaite, G., Mirzasoleiman, B.: Identifying spurious biases early in training through the lens of simplicity bias. In: Proceedings of The 27th International Conference on Artificial Intelligence and Statistics. Proceedings of Machine Learning Research, vol. 238, pp. 2953-2961. PMLR (02-04 May 2024) +40. Yang, Y., Nushi, B., Palangi, H., Mirzasoleiman, B.: Mitigating spurious correlations in multi-modal models during fine-tuning. In: Proceedings of the 40th International Conference on Machine Learning. ICML'23, JMLR.org (2023) +41. Zhang, M., Sohoni, N.S., Zhang, H.R., Finn, C., Ré, C.: Correct-n-contrast: A contrastive approach for improving robustness to spurious correlations. arXiv preprint arXiv:2203.01517 (2022) +42. Zhang, X., He, Y., Xu, R., Yu, H., Shen, Z., Cui, P.: Nico++: Towards better benchmarking for domain generalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16036-16047 (2023) +43. Zhou, D., Bousquet, O., Lal, T., Weston, J., Schölkopf, B.: Learning with local and global consistency. Advances in neural information processing systems 16 (2003) \ No newline at end of file diff --git a/2024/SLIM_ Spuriousness Mitigation with Minimal Human Annotations/images.zip b/2024/SLIM_ Spuriousness Mitigation with Minimal Human Annotations/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f9accfa6e3e8f4bfe4ac06b108ac23f51a53eb74 --- /dev/null +++ b/2024/SLIM_ Spuriousness Mitigation with Minimal Human Annotations/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2569e5a9fbade50b36c1650d785ea3255c0ff2602ed28c96ff4517b27e284352 +size 430493 diff --git a/2024/SLIM_ Spuriousness Mitigation with Minimal Human Annotations/layout.json b/2024/SLIM_ Spuriousness Mitigation with Minimal Human Annotations/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b785b7d7219c1b6210d7ff6aa321ad8dbfa17de8 --- /dev/null +++ b/2024/SLIM_ Spuriousness Mitigation with Minimal Human Annotations/layout.json @@ -0,0 +1,8664 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 186, + 112, + 430, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 112, + 430, + 146 + ], + "spans": [ + { + "bbox": [ + 186, + 112, + 430, + 146 + ], + "type": "text", + "content": "SLIM: Spuriousness Mitigation with Minimal Human Annotations" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 168, + 468, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 168, + 468, + 181 + ], + "spans": [ + { + "bbox": [ + 146, + 168, + 468, + 181 + ], + "type": "text", + "content": "Xiwei Xuan" + }, + { + "bbox": [ + 146, + 168, + 468, + 181 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 146, + 168, + 468, + 181 + ], + "type": "text", + "content": ", Ziquan Deng" + }, + { + "bbox": [ + 146, + 168, + 468, + 181 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 146, + 168, + 468, + 181 + ], + "type": "text", + "content": ", Hsuan-Tien Lin" + }, + { + "bbox": [ + 146, + 168, + 468, + 181 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 146, + 168, + 468, + 181 + ], + "type": "text", + "content": ", and Kwan-Liu Ma" + }, + { + "bbox": [ + 146, + 168, + 468, + 181 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 241, + 190, + 374, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 190, + 374, + 201 + ], + "spans": [ + { + "bbox": [ + 241, + 190, + 374, + 201 + ], + "type": "text", + "content": "1 University of California, Davis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 246, + 201, + 368, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 201, + 368, + 213 + ], + "spans": [ + { + "bbox": [ + 246, + 201, + 368, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 246, + 201, + 368, + 213 + ], + "type": "text", + "content": " National Taiwan University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 169, + 213, + 444, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 213, + 444, + 224 + ], + "spans": [ + { + "bbox": [ + 169, + 213, + 444, + 224 + ], + "type": "text", + "content": "{xwxuan, ziqdeng, klma}@ucdavis.edu, htlin@csie.ntu.edu.tw" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 255, + 455, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 255, + 455, + 475 + ], + "spans": [ + { + "bbox": [ + 160, + 255, + 455, + 475 + ], + "type": "text", + "content": "Abstract. Recent studies highlight that deep learning models often learn spurious features mistakenly linked to labels, compromising their reliability in real-world scenarios where such correlations do not hold. Despite the increasing research effort, existing solutions often face two main challenges: they either demand substantial annotations of spurious attributes, or they yield less competitive outcomes with expensive training when additional annotations are absent. In this paper, we introduce SLIM, a cost-effective and performance-targeted approach to reducing spurious correlations in deep learning. Our method leverages a human-in-the-loop protocol featuring a novel attention labeling mechanism with a constructed attention representation space. SLIM significantly reduces the need for exhaustive additional labeling, requiring human input for fewer than " + }, + { + "bbox": [ + 160, + 255, + 455, + 475 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 160, + 255, + 455, + 475 + ], + "type": "text", + "content": " of instances. By prioritizing data quality over complicated training strategies, SLIM curates a smaller yet more feature-balanced data subset, fostering the development of spuriousness-robust models. Experimental validations across key benchmarks demonstrate that SLIM competes with or exceeds the performance of leading methods while significantly reducing costs. The SLIM framework thus presents a promising path for developing reliable models more efficiently. Our code is available in https://github.com/xiweix/SLIM.git/." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 160, + 486, + 453, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 486, + 453, + 497 + ], + "spans": [ + { + "bbox": [ + 160, + 486, + 453, + 497 + ], + "type": "text", + "content": "Keywords: Model spuriousness " + }, + { + "bbox": [ + 160, + 486, + 453, + 497 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 486, + 453, + 497 + ], + "type": "text", + "content": " Human-in-the-loop " + }, + { + "bbox": [ + 160, + 486, + 453, + 497 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 486, + 453, + 497 + ], + "type": "text", + "content": " Data distribution" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 519, + 230, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 519, + 230, + 532 + ], + "spans": [ + { + "bbox": [ + 132, + 519, + 230, + 532 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": "Spurious correlations, where models mistakenly rely on irrelevant features to make decisions, pose a significant challenge in machine learning. Consequently, certain groups may experience deceptively inflated performance metrics, while accuracy reduces for others. This prevalence of \"right answers for wrong reasons\" fosters a misleading impression of model proficiency and constrains its utility across diverse contexts. Meanwhile, rectifying the issue of spuriousness is a challenging task, particularly when seeking to enhance the performance of underperforming groups. A notable example is the ISIC skin cancer dataset [5], where a typical spurious correlation occurs when color patches frequently co-occur with benign-labeled data. As shown in Fig. 1, a biased model may focus" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 115, + 481, + 161 + ], + "blocks": [ + { + "bbox": [ + 133, + 115, + 481, + 161 + ], + "lines": [ + { + "bbox": [ + 133, + 115, + 481, + 161 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 481, + 161 + ], + "type": "image", + "image_path": "7b1247d216167746b4d9731ef33e662ebef38e44f8d72ee78a1a2d2a7e33efa9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 172, + 482, + 240 + ], + "lines": [ + { + "bbox": [ + 130, + 172, + 482, + 240 + ], + "spans": [ + { + "bbox": [ + 130, + 172, + 482, + 240 + ], + "type": "text", + "content": "Fig. 1: (Left) GradCAM visualizations highlighting model attention on ISIC. (a, b) showcase a model biased towards patches, leading to a correct prediction with wrong reasons in (a) but an incorrect prediction despite correct focus in (b). (c, d) depict a spuriousness-robust model consistently focusing on core features. (Right) The table details the training data distribution by class (benign or malignant) and the presence of color patches, illustrating the imbalance and potential for spurious correlations." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 256, + 482, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 256, + 482, + 316 + ], + "spans": [ + { + "bbox": [ + 130, + 256, + 482, + 316 + ], + "type": "text", + "content": "on patches rather than skin lesions to identify melanoma, and fail when such patches are missing, while a desirable case would be the model captures correct skin regions regardless of spurious features' existence. Enhancing model robustness to address such dangerous misdiagnoses is crucial, driving a surge in research towards more effective solutions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 323, + 312, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 323, + 312, + 454 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 312, + 454 + ], + "type": "text", + "content": "A typical approach in this context is to seek for optimal data distribution. GDRO [27] proves the efficacy of group-balanced data in enhancing model robustness. However, along with subsequent studies [14, 22], such an approach requires spuriousness labels at dataset scale, consuming a considerable cost. Our investigations reveal notable inconsistencies in human-annotated spuriousness labels, raising concerns about their relia" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 320, + 394, + 480, + 452 + ], + "blocks": [ + { + "bbox": [ + 318, + 338, + 482, + 392 + ], + "lines": [ + { + "bbox": [ + 318, + 338, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 318, + 338, + 482, + 392 + ], + "type": "text", + "content": "Table 1: Annotation forms and quantities required by different methods on Waterbirds, highlighting the reduced workload of SLIM. (Att. = Attention, T. = Training set, V. = Validation set)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 320, + 394, + 480, + 452 + ], + "lines": [ + { + "bbox": [ + 320, + 394, + 480, + 452 + ], + "spans": [ + { + "bbox": [ + 320, + 394, + 480, + 452 + ], + "type": "table", + "html": "
MethodAnno. FormAnno. Amount
RRR [26]Map100%×T.
GradIA [9]Att. Yes-No, Map100%×T.
Energy Loss [24]Bounding box100%×T.
CRAYON [15]Att. Yes-No100%×T.
GDRO [27]Spuriousness label100%×T.
DFR [14]Spuriousness label100%×V.
SLIM (ours)Att. Yes-No2.5%×T. or 3%×V.
", + "image_path": "7f66332c8cd36ad0b8883d48978b554e56a85cd43e874b201e9a2ddddfa93d62.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 455, + 482, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 455, + 482, + 624 + ], + "spans": [ + { + "bbox": [ + 130, + 455, + 482, + 624 + ], + "type": "text", + "content": "bility (refer to Table 6). There are other approaches with more direct human involvement, such as providing yes-no feedback on attention correctness [15,36]; or outlining core features with maps [9,26] or bounding boxes [24]. Despite different formats of queried information, they require extensive effort across the entire dataset, as summarized in Table 1. Considering the empirical applicability, our work aims to streamline this process by reducing human annotation workload and minimizing the ambiguity of collected data. Methods requiring spuriousness annotation tend to excel in performance at higher costs, while minimizing such expenses typically results in performance trade-offs [1,6]. Facing this challenge, techniques such as JTT [17] and CNC [41] design training or data-augmenting phases to reduce spuriousness without additional annotations. However, the extra computational cost raised by multiple training iterations is undesired. We believe, as spuriousness is indeed caused by data biases, improving data quality is more essential than designing complex learning strategies." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "Therefore, the remaining research gap calls for a data quality-oriented approach that balances performance and cost. This is where our framework, SLIM, steps in. " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "SLIM" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " offers a novel approach to mitigating spurious correlations by" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 226, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 226, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 226, + 101 + ], + "type": "text", + "content": "X. Xuan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 366 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 366 + ], + "type": "text", + "content": "intelligently leveraging a minimal set of human-annotated data and curating a feature-balanced data set to enhance model robustness without any significant computational overhead incurred by many existing methods. As shown in Fig. 2, inferred from a reference model, " + }, + { + "bbox": [ + 133, + 116, + 481, + 366 + ], + "type": "inline_equation", + "content": "SLIM" + }, + { + "bbox": [ + 133, + 116, + 481, + 366 + ], + "type": "text", + "content": " first constructs an attention representation space that reflects locally consistent data features and model attention, enabling efficient attention label estimation. — Representative instances can be sampled from this space for querying attention correctness labels and such labels can be expanded to neighbor instances. Subsequently, utilizing a visual explanation method that highlights model attention, like GradCAM [29], " + }, + { + "bbox": [ + 133, + 116, + 481, + 366 + ], + "type": "inline_equation", + "content": "SLIM" + }, + { + "bbox": [ + 133, + 116, + 481, + 366 + ], + "type": "text", + "content": " disentangles core and environment features within the latent space. Here, \"core features\" denote features critical for class determination, whereas \"environment features\" represent the surrounding contextual information. We agree with GDRO that balanced data distribution is crucial, but instead of using human-annotated labels to approximate spuriousness distribution, we leverage latent space features directly disentangled from model attention to balance data distribution, which is more aligned with model understanding and involving less human assumptions. Through this, " + }, + { + "bbox": [ + 133, + 116, + 481, + 366 + ], + "type": "inline_equation", + "content": "SLIM" + }, + { + "bbox": [ + 133, + 116, + 481, + 366 + ], + "type": "text", + "content": " assembles data groups with diverse feature combinations. Focused on data quality, a consistency-aware sampling mechanism is employed to curate a subset with optimal feature distributions — ensuring balanced core features across various environmental contexts. Finally, we conclude our framework with a straightforward training procedure." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 368, + 481, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 368, + 481, + 451 + ], + "spans": [ + { + "bbox": [ + 133, + 368, + 481, + 451 + ], + "type": "text", + "content": "The experimental results demonstrate SLIM's advanced capability in spuriousness mitigation by performing comparably to, or exceeding state-of-the-art (SOTA) methods. Notably, SLIM accomplishes this with (1) a significantly reduced requirement for human annotations: " + }, + { + "bbox": [ + 133, + 368, + 481, + 451 + ], + "type": "inline_equation", + "content": "0.12\\% - 2.5\\%" + }, + { + "bbox": [ + 133, + 368, + 481, + 451 + ], + "type": "text", + "content": " of the training set size and " + }, + { + "bbox": [ + 133, + 368, + 481, + 451 + ], + "type": "inline_equation", + "content": "0.35\\% - 4\\%" + }, + { + "bbox": [ + 133, + 368, + 481, + 451 + ], + "type": "text", + "content": " of the validation set; and (2) a lightweight training for a robust model using our constructed data with " + }, + { + "bbox": [ + 133, + 368, + 481, + 451 + ], + "type": "inline_equation", + "content": "5\\% - 30\\%" + }, + { + "bbox": [ + 133, + 368, + 481, + 451 + ], + "type": "text", + "content": " of the original set. Our code will be published after the review process is complete. Our contributions are:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 457, + 480, + 562 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 132, + 457, + 480, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 491 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 491 + ], + "type": "text", + "content": "- Data-quality-oriented spuriousness mitigation: SLIM introduces an interactive, simple, and effective data construction pipeline to create high-quality data for spuriousness mitigation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 493, + 480, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 493, + 480, + 527 + ], + "spans": [ + { + "bbox": [ + 132, + 493, + 480, + 527 + ], + "type": "text", + "content": "- Cost-efficient framework: In contrast to related works with leading performance, SLIM demonstrates a significant reduction in both annotation requirements and computational cost." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 529, + 480, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 529, + 480, + 562 + ], + "spans": [ + { + "bbox": [ + 132, + 529, + 480, + 562 + ], + "type": "text", + "content": "- Competitive or superior model performance: Without incurring excessive costs, SLIM achieves comparable or superior performance to existing methods, confirming its effectiveness in mitigating spurious correlations." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 133, + 582, + 236, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 582, + 236, + 594 + ], + "spans": [ + { + "bbox": [ + 133, + 582, + 236, + 594 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 605, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 605, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 605, + 481, + 665 + ], + "type": "text", + "content": "Spuriousness mitigation with spuriousness labels. When a comprehensive set of spuriousness labels is accessible during training, methods like class balancing [11] and importance weighting [32] can be employed for mitigating spurious correlations, which utilizes data groups defined by class and spuriousness labels. Following the same schema, GDRO [27] refines the training approach" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 176, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 176, + 91, + 447, + 102 + ], + "type": "text", + "content": "SLIM: Spuriousness Mitigation with Minimal Human Annotations" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 307 + ], + "type": "text", + "content": "by focusing on the worst-performing groups as determined by ERM loss. Aiming to reduce the dependency on comprehensive spuriousness labeling, several methods [1,6,17,21,30,32,41] have emerged to approximate group information using a reference model. While these approaches are technically sophisticated, they often fail to achieve the same level of performance as approaches informed by spuriousness labels. To balance the tradeoff between performance and cost, a further research direction reduces the requirement for extensive labeling by leveraging spuriousness labels in smaller validation datasets [14,22]. For example, SSA [22] adopts semi-supervised learning to predict group labels in training data according to information from validation, then applies GDRO to reduce spurious correlations. However, even with this reduced scale of validation sets, the labor-intensive nature of spuriousness labeling remains a barrier to scalability. Moreover, groups in these methods are formed based on human-defined criteria, risking a misalignment with the actual categories a model might discern. This drawback can hinder the true representation of spurious correlations and may lead to suboptimal mitigation outcomes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 308, + 482, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 482, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 482, + 511 + ], + "type": "text", + "content": "Spuriousness mitigation with human-in-the-loop. Integrating human insights has proven beneficial in mitigating model spuriousness, as evidenced by various human-involved strategies. Considering the valuable human effort, [28] provides an efficient solution by querying data samples for issue mitigation with domain experts in the loop, which overlooks potential issues in the broader data scope. Another option for avoiding redundant human annotation is to query global spurious concepts [31, 38] from knowledgeable experts, which can be ambiguous and fail to cover all possibilities. More recent approaches [10, 24, 26] raise concerns about the inconsistency of spuriousness labels, introducing alternatives such as human-generated attention maps or bounding boxes outlining core features. These human-generated references help align the models' attention with the actual features relevant to decision-making. However, producing such annotations is challenging, and these techniques are resource-intensive as they necessitate annotations across the entire dataset. In response, more recent research research [2, 40] simplifies the annotation process to binary feedback on model's attention correctness. While less demanding, its requirement to annotate all instances still poses challenges in real practices." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 530, + 282, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 530, + 282, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 530, + 282, + 544 + ], + "type": "text", + "content": "3 Problem Formulation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 555, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 555, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 555, + 482, + 604 + ], + "type": "text", + "content": "We introduce the formal definition of the spuriousness mitigation problem [13]. Assume training data " + }, + { + "bbox": [ + 130, + 555, + 482, + 604 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{(x_i, y_i)\\}_{i=1}^n" + }, + { + "bbox": [ + 130, + 555, + 482, + 604 + ], + "type": "text", + "content": " with input features " + }, + { + "bbox": [ + 130, + 555, + 482, + 604 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_i \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 130, + 555, + 482, + 604 + ], + "type": "text", + "content": " and labels " + }, + { + "bbox": [ + 130, + 555, + 482, + 604 + ], + "type": "inline_equation", + "content": "y_i \\in [\\mathcal{L}]" + }, + { + "bbox": [ + 130, + 555, + 482, + 604 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 555, + 482, + 604 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\{L_1, \\ldots, L_l\\}" + }, + { + "bbox": [ + 130, + 555, + 482, + 604 + ], + "type": "text", + "content": " are the classes in " + }, + { + "bbox": [ + 130, + 555, + 482, + 604 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 555, + 482, + 604 + ], + "type": "text", + "content": ". Machine learning models for classification are trained to minimize the empirical risk:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 228, + 616, + 481, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 616, + 481, + 635 + ], + "spans": [ + { + "bbox": [ + 228, + 616, + 481, + 635 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\theta} ^ {*} \\in \\underset {\\boldsymbol {\\theta}} {\\arg \\min } \\mathbb {E} _ {(\\boldsymbol {x} _ {i}, y _ {i}) \\in \\mathcal {D}} [ l (f (\\boldsymbol {\\theta}, \\boldsymbol {x} _ {i}), y _ {i}) ], \\tag {1}", + "image_path": "656986117ed9794ddfe7f5f09507d4744c161cf58de4954147352dd8c7da91e1.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}" + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": " represent model parameters, " + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "inline_equation", + "content": "f(\\pmb{\\theta},\\pmb{x}_i)" + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "inline_equation", + "content": "l(*)" + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": " are the model outputs and the loss function, respectively." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "text", + "content": "X. Xuan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "content": "Groups. In spuriousness mitigation, we assume there exists a core feature set " + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = \\{c_1,\\dots ,c_l\\}" + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "content": " and a spurious feature set " + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "inline_equation", + "content": "\\mathcal{S} = \\{s_1,\\dots ,s_k\\}" + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "content": ". Each class " + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "inline_equation", + "content": "L_{i}" + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "content": " is uniquely identified by a core feature " + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "content": ", while spurious features " + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "inline_equation", + "content": "s_j" + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "content": " can be found across different classes. We consider an instance in the dataset as having both a core feature " + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "content": " and a spurious feature " + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "inline_equation", + "content": "s_j" + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "content": ", and each instance belongs to a unique group " + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "inline_equation", + "content": "g_{c_i,s_j}" + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "content": ". Core features are reliable class indicators in both training and testing sets, while spurious features with misleading co-occurrence with class in training may miss such correlations in the testing phase. If the model mistakenly links spurious features with classes in training, it may fail in testing when such spurious features are absent. This can lead to lower accuracy rates, especially for less common group variations that do not exhibit the spurious feature the model has learned to rely on." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 260, + 481, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 260, + 481, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 260, + 481, + 297 + ], + "type": "text", + "content": "Worst-group accuracy. The vital evaluation of spuriousness mitigation focuses on the model's performance in the subgroup where it is least effective – the worst-group accuracy. It is calculated by:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 218, + 306, + 481, + 332 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 218, + 306, + 481, + 332 + ], + "spans": [ + { + "bbox": [ + 218, + 306, + 481, + 332 + ], + "type": "interline_equation", + "content": "A c c _ {w} = \\min _ {g \\in \\mathcal {G}} \\left(\\frac {1}{| g |} \\sum_ {\\left(x _ {i}, y _ {i}\\right) \\in g} \\mathbf {1} [ f \\left(x _ {i}\\right) = y _ {i} ]\\right), \\tag {2}", + "image_path": "52d4c5c983bee0b4cc1e950b19b23bf0b935ad1b4fdaad9e416e443709cb6c40.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 340, + 482, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 482, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 482, + 426 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 340, + 482, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{G} = \\{\\cup_{c_i,s_j}g_{c_i,s_j}\\}" + }, + { + "bbox": [ + 130, + 340, + 482, + 426 + ], + "type": "text", + "content": " is the union of subgroups defined by core feature " + }, + { + "bbox": [ + 130, + 340, + 482, + 426 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 130, + 340, + 482, + 426 + ], + "type": "text", + "content": " and spurious feature " + }, + { + "bbox": [ + 130, + 340, + 482, + 426 + ], + "type": "inline_equation", + "content": "s_j" + }, + { + "bbox": [ + 130, + 340, + 482, + 426 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 340, + 482, + 426 + ], + "type": "inline_equation", + "content": "f(x_{i})" + }, + { + "bbox": [ + 130, + 340, + 482, + 426 + ], + "type": "text", + "content": " is the label predicted by the model. This metric is crucial because it directly assesses how effectively a model can generalize across different subgroups, especially those that might be mostly underrepresented because of spurious correlations. It ensures that improvements in model performance are not just reflective of gains on the majority or less challenging subsets of the data." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 445, + 203, + 458 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 445, + 203, + 458 + ], + "spans": [ + { + "bbox": [ + 132, + 445, + 203, + 458 + ], + "type": "text", + "content": "4 Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 473, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 482, + 668 + ], + "type": "text", + "content": "In this section, we detail the method behind SLIM. In alignment with related works requiring no spuriousness label, our approach also follows a two-stage procedure: inferring groups with an ERM-trained reference model " + }, + { + "bbox": [ + 130, + 473, + 482, + 668 + ], + "type": "inline_equation", + "content": "f_{r}" + }, + { + "bbox": [ + 130, + 473, + 482, + 668 + ], + "type": "text", + "content": ", and then leveraging the inferred groups to train a robust model. As an approach oriented by data quality, SLIM framework comprises a three-phase data construction pipeline, as illustrated in Fig. 2. Firstly, we build an attention representation space to group instances with both similar data features and model attention, preparing for the data sampling and label propagation. Following this, we sample typical instances for human to evaluate model's attention correctness with yes-no feedback, then estimate neighboring instances' attention labels accordingly. In the last phase, we filter out data with wrong attention and leverage attention matrix to disentangle core and environment features. With constructed feature sets, we apply consistency-weighted data sampling to curate a feature-balanced subset, with a desired core-environment balanced distribution. Finally, SLIM framework is concluded by a training process with the curated subset, resulting in a model exhibiting robustness against spurious correlations." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 176, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 176, + 91, + 448, + 102 + ], + "type": "text", + "content": "SLIM: Spuriousness Mitigation with Minimal Human Annotations" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 142, + 113, + 473, + 254 + ], + "blocks": [ + { + "bbox": [ + 142, + 113, + 473, + 254 + ], + "lines": [ + { + "bbox": [ + 142, + 113, + 473, + 254 + ], + "spans": [ + { + "bbox": [ + 142, + 113, + 473, + 254 + ], + "type": "image", + "image_path": "0cdc12ac114e2af3890c359276af5715a31132619d73866b44116d1be9cfcb99.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 262, + 482, + 361 + ], + "lines": [ + { + "bbox": [ + 130, + 262, + 482, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 482, + 361 + ], + "type": "text", + "content": "Fig. 2: Overview of SLIM framework, with a data construction pipeline consisting of three phases: (1) Attention Space Construction, creating a space with data features and model attention aligned locally; (2) Attention Annotation and Expansion, where instances sampled from the attention space are labeled by human for attention correctness, and labels are propagated to neighboring instances; and (3) Balanced Data Curation, which filters out instances with incorrect attention, and utilize attention-weighted " + }, + { + "bbox": [ + 130, + 262, + 482, + 361 + ], + "type": "inline_equation", + "content": "(F_A)" + }, + { + "bbox": [ + 130, + 262, + 482, + 361 + ], + "type": "text", + "content": " and inverse-attention-weighted " + }, + { + "bbox": [ + 130, + 262, + 482, + 361 + ], + "type": "inline_equation", + "content": "(F_{\\bar{A}})" + }, + { + "bbox": [ + 130, + 262, + 482, + 361 + ], + "type": "text", + "content": " feature vectors to create core and environment feature sets, forming subgroups to assemble a feature-balanced subset for training a spuriousness-robust model." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 365, + 311, + 378 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 311, + 378 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 311, + 378 + ], + "type": "text", + "content": "4.1 Attention Space Construction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 388, + 482, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 482, + 484 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 482, + 484 + ], + "type": "text", + "content": "Considering the benefits of being aware of spuriousness information and drawbacks of the annotation cost, in our framework, we deploy a binary labeling scheme to simplify descriptive labeling tasks into yes-no feedback on model's attention correctness. Moreover, we aim to further facilitate this process by avoiding the necessity of exhaustive labeling. Our empirical observations indicate that a model would have consistent attention mechanisms on specific data subgroups. If such subgroups can be found, it would largely aid the labeling workload by enabling label estimations among neighbors." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 485, + 482, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 482, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 482, + 556 + ], + "type": "text", + "content": "Driven by this, we aim to find a representation space where instances with similar attention outcomes (as indicated by highlighted image features) are closely grouped. However, our analysis reveals that conventional feature representation spaces formed directly from feature vectors do not fulfill this requirement - such spaces only maintain local consistency in data features but fail to capture the nuances of model attention (refer to Sec. 5.3)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": "Thus, we have developed a process utilizing model attention in the latent space to support the establishment of such space, as illustrated in Fig. 3. For each image, we obtain its feature vector " + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": " and model attribution " + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": " in the latent space at first. Note that " + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": " represents the model's attention focus and can be linearly interpolated to visualize an attention heatmap. We then use " + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": " as a weight matrix for " + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": ", producing an attention-weighted feature vector " + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "inline_equation", + "content": "F_{A}" + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": ". Finally, with the dimensionality reduction method UMAP [20], we project vectors " + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "inline_equation", + "content": "F_{A}" + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": " into a 2D space, namely the attention space, which ensures local consistency encompasses both data features and the model's attention mechanisms." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 226, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 226, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 226, + 101 + ], + "type": "text", + "content": "X. Xuan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 142, + 114, + 473, + 166 + ], + "blocks": [ + { + "bbox": [ + 142, + 114, + 473, + 166 + ], + "lines": [ + { + "bbox": [ + 142, + 114, + 473, + 166 + ], + "spans": [ + { + "bbox": [ + 142, + 114, + 473, + 166 + ], + "type": "image", + "image_path": "a052730a8a30e45226e81fecedbdef87655946b17163286dbcb43f04a1db1175.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 176, + 479, + 231 + ], + "lines": [ + { + "bbox": [ + 130, + 176, + 479, + 231 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 479, + 231 + ], + "type": "text", + "content": "Fig. 3: Construction of the attention space. In the latent space, feature vector " + }, + { + "bbox": [ + 130, + 176, + 479, + 231 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 176, + 479, + 231 + ], + "type": "text", + "content": " and model's attribution vector " + }, + { + "bbox": [ + 130, + 176, + 479, + 231 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 130, + 176, + 479, + 231 + ], + "type": "text", + "content": " are extracted, representing input features and model's attention, representatively. By weighting " + }, + { + "bbox": [ + 130, + 176, + 479, + 231 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 176, + 479, + 231 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 176, + 479, + 231 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 130, + 176, + 479, + 231 + ], + "type": "text", + "content": ", the attention-weighted feature vector " + }, + { + "bbox": [ + 130, + 176, + 479, + 231 + ], + "type": "inline_equation", + "content": "F_{A}" + }, + { + "bbox": [ + 130, + 176, + 479, + 231 + ], + "type": "text", + "content": " emphasizes features of model's top interest. All " + }, + { + "bbox": [ + 130, + 176, + 479, + 231 + ], + "type": "inline_equation", + "content": "F_{A}" + }, + { + "bbox": [ + 130, + 176, + 479, + 231 + ], + "type": "text", + "content": " are then projected to form the attention space featuring locally consistent features and model attention." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 236, + 480, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 480, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 480, + 307 + ], + "type": "text", + "content": "Note that for a fair comparison, in alignment with baseline methods, experiments in this paper involve the CNN model (ResNet) for classification and GradCAM for obtaining model attention. But our framework is fully applicable to (1) other vision model architectures such as vision transformers (ViT) [8] or Swin transformers [18], and (2) other heatmap-based model visualization techniques, such as RISE [23] or ScoreCAM [33]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 327, + 348, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 327, + 348, + 338 + ], + "spans": [ + { + "bbox": [ + 132, + 327, + 348, + 338 + ], + "type": "text", + "content": "4.2 Attention Annotation and Expansion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 348, + 480, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 348, + 480, + 528 + ], + "spans": [ + { + "bbox": [ + 130, + 348, + 480, + 528 + ], + "type": "text", + "content": "Given that our attention space ensures locally consistent model attention, it allows for annotating a selectively sampled subset of instances and then estimating neighboring labels accordingly, thereby enabling efficient transfer of human expertise. The criteria for selecting these instances are twofold: (1) diversity, to encompass a broad covering of various data features and model attention patterns; (2) typicality, ensuring they are representative of their neighboring instances. To achieve this, we utilize the k-means clustering algorithm to partition our constructed attention space into " + }, + { + "bbox": [ + 130, + 348, + 480, + 528 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 348, + 480, + 528 + ], + "type": "text", + "content": " clusters (diversity). We then select " + }, + { + "bbox": [ + 130, + 348, + 480, + 528 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 348, + 480, + 528 + ], + "type": "text", + "content": " instances that are closest to each cluster's center (typicality). The choice of " + }, + { + "bbox": [ + 130, + 348, + 480, + 528 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 348, + 480, + 528 + ], + "type": "text", + "content": " is aimed at optimizing the balance between minimizing intra-cluster variance in data features and model attention patterns, and keeping the annotation budget manageable. Our empirical analysis, detailed in Sec. 5.3, reveals that the intra-cluster consistencies increase with " + }, + { + "bbox": [ + 130, + 348, + 480, + 528 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 348, + 480, + 528 + ], + "type": "text", + "content": ", but this trend levels off upon reaching a specific value of " + }, + { + "bbox": [ + 130, + 348, + 480, + 528 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 348, + 480, + 528 + ], + "type": "text", + "content": " -the elbow point. Thus, selecting " + }, + { + "bbox": [ + 130, + 348, + 480, + 528 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 348, + 480, + 528 + ], + "type": "text", + "content": " at this elbow point optimizes consistency without unnecessarily increasing the annotation workload." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 529, + 481, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 529, + 481, + 598 + ], + "spans": [ + { + "bbox": [ + 130, + 529, + 481, + 598 + ], + "type": "text", + "content": "We gather binary yes-no feedback from human experts on the correctness of the model's attention for these " + }, + { + "bbox": [ + 130, + 529, + 481, + 598 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 529, + 481, + 598 + ], + "type": "text", + "content": " instances, feeding this feedback along with our attention space embedding into the label spreading algorithm. This algorithm leverages data similarity and density to extend attention labels to previously unlabeled data [43]. Ultimately, this process yields a probability score for correct attention for each instance across the dataset." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 620, + 284, + 631 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 620, + 284, + 631 + ], + "spans": [ + { + "bbox": [ + 132, + 620, + 284, + 631 + ], + "type": "text", + "content": "4.3 Balanced Data Curation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "content": "Following Sec. 4.2 that provides attention accuracy scores, we proceed to construct our data subset to support model's unbiased learning, as described below:" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 176, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 176, + 91, + 448, + 102 + ], + "type": "text", + "content": "SLIM: Spuriousness Mitigation with Minimal Human Annotations" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "content": "Data screening. Inspired by works [3,7,12], our theoretical analysis further supports the premise that data receiving correct attention from a reference model are more likely to enhance the learning of core features in a robust model within a more balanced data distribution, as substantiated in Appx. A.2. Furthermore, our study indicates that even a reference model with inherent biases has the potential to accurately identify core features for a data subset. Such correctly captured features typically exhibit diversity, ensuring good coverage across the spectrum of the core feature set. This understanding informs our approach to curating a data subset that: (1) selects instances based on accurate model attention, and (2) ensures a balanced representation of core features across various environment features. Thus we first apply data screening to filter out instances with wrong attention according to attention accuracy scores." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 261, + 482, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 261, + 482, + 539 + ], + "spans": [ + { + "bbox": [ + 130, + 261, + 482, + 539 + ], + "type": "text", + "content": "Disentangling core and environment features. After screening, the retained data subset mainly comprises instances with correct attention. This is reflected by attention masks that highlight core features, meaning that the corresponding attribution vector " + }, + { + "bbox": [ + 130, + 261, + 482, + 539 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 130, + 261, + 482, + 539 + ], + "type": "text", + "content": " in the latent space also highlights core features of feature vector " + }, + { + "bbox": [ + 130, + 261, + 482, + 539 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 261, + 482, + 539 + ], + "type": "text", + "content": ". Moreover, existing studies have found that a model's feature vectors are representative of most input features, regardless of whether they are used for decision-making [14, 37]. Therefore, utilizing " + }, + { + "bbox": [ + 130, + 261, + 482, + 539 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 130, + 261, + 482, + 539 + ], + "type": "text", + "content": " as the weight matrix, we disentangle core and environment features in the latent space — weighting " + }, + { + "bbox": [ + 130, + 261, + 482, + 539 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 261, + 482, + 539 + ], + "type": "text", + "content": " by attention masks " + }, + { + "bbox": [ + 130, + 261, + 482, + 539 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 130, + 261, + 482, + 539 + ], + "type": "text", + "content": " and inverse attention masks " + }, + { + "bbox": [ + 130, + 261, + 482, + 539 + ], + "type": "inline_equation", + "content": "\\bar{A}" + }, + { + "bbox": [ + 130, + 261, + 482, + 539 + ], + "type": "text", + "content": ". Subsequently, k-means clustering is applied to partition all attention-weighted feature vectors " + }, + { + "bbox": [ + 130, + 261, + 482, + 539 + ], + "type": "inline_equation", + "content": "F_{A}" + }, + { + "bbox": [ + 130, + 261, + 482, + 539 + ], + "type": "text", + "content": ", and inverse-attention-weighted feature vectors " + }, + { + "bbox": [ + 130, + 261, + 482, + 539 + ], + "type": "inline_equation", + "content": "F_{\\bar{A}}" + }, + { + "bbox": [ + 130, + 261, + 482, + 539 + ], + "type": "text", + "content": ", respectively, facilitating the formation of representative subgroups for the core feature set " + }, + { + "bbox": [ + 130, + 261, + 482, + 539 + ], + "type": "inline_equation", + "content": "(\\hat{C} = \\hat{c}_1,\\dots ,\\hat{c}_N)" + }, + { + "bbox": [ + 130, + 261, + 482, + 539 + ], + "type": "text", + "content": " and the environment feature set " + }, + { + "bbox": [ + 130, + 261, + 482, + 539 + ], + "type": "inline_equation", + "content": "(\\mathcal{E} = e_1,\\ldots ,e_M)" + }, + { + "bbox": [ + 130, + 261, + 482, + 539 + ], + "type": "text", + "content": ". The number of clusters, determined through the elbow method, culminates in " + }, + { + "bbox": [ + 130, + 261, + 482, + 539 + ], + "type": "inline_equation", + "content": "M\\times N" + }, + { + "bbox": [ + 130, + 261, + 482, + 539 + ], + "type": "text", + "content": " subgroups " + }, + { + "bbox": [ + 130, + 261, + 482, + 539 + ], + "type": "inline_equation", + "content": "g_{\\hat{c}_i,e_j}" + }, + { + "bbox": [ + 130, + 261, + 482, + 539 + ], + "type": "text", + "content": ", each characterized by a distinct combination of core and environment features " + }, + { + "bbox": [ + 130, + 261, + 482, + 539 + ], + "type": "inline_equation", + "content": "(\\hat{c}_i,e_j)" + }, + { + "bbox": [ + 130, + 261, + 482, + 539 + ], + "type": "text", + "content": ". Data sampling. Given the unequal distribution of core and environment features, the internal consistency across each constructed subgroup " + }, + { + "bbox": [ + 130, + 261, + 482, + 539 + ], + "type": "inline_equation", + "content": "g_{\\hat{c}_i,e_j}" + }, + { + "bbox": [ + 130, + 261, + 482, + 539 + ], + "type": "text", + "content": " varies. Our sampling strategy aims to mitigate over-representation of particular features to prevent overfitting, while ensuring broad coverage of diverse core and environment features. We assess cluster consistency as a metric for determining the sample size from each cluster. For a cluster " + }, + { + "bbox": [ + 130, + 261, + 482, + 539 + ], + "type": "inline_equation", + "content": "\\hat{c}_i" + }, + { + "bbox": [ + 130, + 261, + 482, + 539 + ], + "type": "text", + "content": " in the core feature space, denote the distance between an instance " + }, + { + "bbox": [ + 130, + 261, + 482, + 539 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 261, + 482, + 539 + ], + "type": "text", + "content": " and the cluster center as " + }, + { + "bbox": [ + 130, + 261, + 482, + 539 + ], + "type": "inline_equation", + "content": "d_{\\hat{c}_i}(x)" + }, + { + "bbox": [ + 130, + 261, + 482, + 539 + ], + "type": "text", + "content": ". The cluster consistency is defined as the reciprocal of the average distance per cluster:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 255, + 550, + 481, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 550, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 255, + 550, + 481, + 578 + ], + "type": "interline_equation", + "content": "\\rho_ {\\hat {c} _ {i}} = \\frac {1}{\\frac {1}{| \\hat {c} _ {i} |} \\sum_ {x \\in \\hat {c} _ {i}} d _ {\\hat {c} _ {i}} (x)}. \\tag {3}", + "image_path": "0633664509c7a7dfcdff7862f687f5b2d1db67146dcbdd745e042835ecdff6db.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 586, + 482, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 586, + 482, + 626 + ], + "spans": [ + { + "bbox": [ + 132, + 586, + 482, + 626 + ], + "type": "text", + "content": "A lower " + }, + { + "bbox": [ + 132, + 586, + 482, + 626 + ], + "type": "inline_equation", + "content": "\\rho_{\\hat{c}_i}" + }, + { + "bbox": [ + 132, + 586, + 482, + 626 + ], + "type": "text", + "content": " signals less consistent features, necessitating a higher sample weight for adequately representing them. A cluster's sampling weight is defined as " + }, + { + "bbox": [ + 132, + 586, + 482, + 626 + ], + "type": "inline_equation", + "content": "w_{\\hat{c}_i} = \\frac{1}{\\rho_{\\hat{c}_i}}" + }, + { + "bbox": [ + 132, + 586, + 482, + 626 + ], + "type": "text", + "content": " and the number of samples to be drawn from a core cluster " + }, + { + "bbox": [ + 132, + 586, + 482, + 626 + ], + "type": "inline_equation", + "content": "\\hat{c}_i" + }, + { + "bbox": [ + 132, + 586, + 482, + 626 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 257, + 637, + 481, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 637, + 481, + 669 + ], + "spans": [ + { + "bbox": [ + 257, + 637, + 481, + 669 + ], + "type": "interline_equation", + "content": "n _ {\\hat {c} _ {i}} = N \\cdot \\frac {1 / \\rho_ {\\hat {c} _ {i}}}{\\sum_ {j = 1} ^ {| \\hat {C} |} 1 / \\rho_ {\\hat {c} _ {j}}}, \\tag {4}", + "image_path": "37bbe2ce88ccd4a6b4b3eaf7f245f5af60abdde4bf20cad245fd9782a834d375.jpg" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "text", + "content": "X. Xuan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 482, + 177 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 131, + 116, + 482, + 177 + ], + "type": "text", + "content": " is the sampling budget. Moving to the environment feature space, the " + }, + { + "bbox": [ + 131, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "n_{\\hat{c}_i}" + }, + { + "bbox": [ + 131, + 116, + 482, + 177 + ], + "type": "text", + "content": " instances from a core feature cluster " + }, + { + "bbox": [ + 131, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "\\hat{c}_i" + }, + { + "bbox": [ + 131, + 116, + 482, + 177 + ], + "type": "text", + "content": " may distribute across multiple environment clusters " + }, + { + "bbox": [ + 131, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "\\{e_1,\\dots ,e_M\\}" + }, + { + "bbox": [ + 131, + 116, + 482, + 177 + ], + "type": "text", + "content": " and formulate subgroups " + }, + { + "bbox": [ + 131, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "\\{g_{\\hat{c}_i,e_1},\\ldots ,g_{\\hat{c}_i,e_M}\\}" + }, + { + "bbox": [ + 131, + 116, + 482, + 177 + ], + "type": "text", + "content": ". A similar methodology is employed for a certain group " + }, + { + "bbox": [ + 131, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "g_{\\hat{c}_i,e_i}" + }, + { + "bbox": [ + 131, + 116, + 482, + 177 + ], + "type": "text", + "content": ". The average distance for " + }, + { + "bbox": [ + 131, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "g_{\\hat{c}_i,e_i}" + }, + { + "bbox": [ + 131, + 116, + 482, + 177 + ], + "type": "text", + "content": " in the environment representation space is:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 238, + 186, + 481, + 215 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 186, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 238, + 186, + 481, + 215 + ], + "type": "interline_equation", + "content": "\\rho_ {\\hat {c} _ {i}, e _ {i}} = \\frac {1}{\\frac {1}{| \\hat {c} _ {i} , e _ {i} |} \\sum_ {x \\in (\\hat {c} _ {i} , e _ {i})} d _ {e _ {i}} (x)}. \\tag {5}", + "image_path": "f78f02c6a1feb7a190fe71352c00e2cf8a66a7471fc192b1eb7346038be4f627.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 222, + 481, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 222, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 131, + 222, + 481, + 251 + ], + "type": "text", + "content": "Similarly, we have the sampling weight " + }, + { + "bbox": [ + 131, + 222, + 481, + 251 + ], + "type": "inline_equation", + "content": "w_{\\hat{c}_i,e_i} = \\frac{1}{\\rho_{\\hat{c}_i,e_i}}" + }, + { + "bbox": [ + 131, + 222, + 481, + 251 + ], + "type": "text", + "content": " and the number of samples to be drawn from a certain group " + }, + { + "bbox": [ + 131, + 222, + 481, + 251 + ], + "type": "inline_equation", + "content": "g_{\\hat{c}_i,e_i}" + }, + { + "bbox": [ + 131, + 222, + 481, + 251 + ], + "type": "text", + "content": " is finally defined by:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 167, + 259, + 481, + 291 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 259, + 481, + 291 + ], + "spans": [ + { + "bbox": [ + 167, + 259, + 481, + 291 + ], + "type": "interline_equation", + "content": "n _ {\\hat {c} _ {i}, e _ {i}} = n _ {\\hat {c} _ {i}} \\cdot \\frac {w _ {\\hat {c} _ {i} , e _ {i}}}{\\sum_ {j = 1} ^ {| \\mathcal {E} (\\hat {c} _ {i}) |} w _ {\\hat {c} _ {j} , e _ {i}}} = N \\cdot \\frac {1 / \\rho_ {\\hat {c} _ {i}}}{\\sum_ {j = 1} ^ {| \\hat {\\mathcal {C}} |} 1 / \\rho_ {\\hat {c} _ {j}}} \\cdot \\frac {1 / \\rho_ {\\hat {c} _ {i} , e _ {i}}}{\\sum_ {j = 1} ^ {| \\mathcal {E} (\\hat {c} _ {i}) |} 1 / \\rho_ {\\hat {c} _ {i} , e _ {j}}}, \\tag {6}", + "image_path": "58b9757e5e3ed028a36ccc5c919186d692ece02db6d08e94311ebd095d970541.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 300, + 482, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 300, + 482, + 361 + ], + "spans": [ + { + "bbox": [ + 131, + 300, + 482, + 361 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 300, + 482, + 361 + ], + "type": "inline_equation", + "content": "|\\hat{\\mathcal{C}}|" + }, + { + "bbox": [ + 131, + 300, + 482, + 361 + ], + "type": "text", + "content": " denotes the number of core clusters and " + }, + { + "bbox": [ + 131, + 300, + 482, + 361 + ], + "type": "inline_equation", + "content": "|\\mathcal{E}(\\hat{c}_i)|" + }, + { + "bbox": [ + 131, + 300, + 482, + 361 + ], + "type": "text", + "content": " denotes the number of subgroups formulated by instances from cluster " + }, + { + "bbox": [ + 131, + 300, + 482, + 361 + ], + "type": "inline_equation", + "content": "\\hat{c}_i" + }, + { + "bbox": [ + 131, + 300, + 482, + 361 + ], + "type": "text", + "content": ". To summary, our sampling strategy takes feature consistencies and sampling budget into consideration, and ensures fair representation of different core-environment feature combinations across the feature spaces." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 378, + 204, + 391 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 378, + 204, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 378, + 204, + 391 + ], + "type": "text", + "content": "4.4 Training" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 398, + 481, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 398, + 481, + 446 + ], + "spans": [ + { + "bbox": [ + 130, + 398, + 481, + 446 + ], + "type": "text", + "content": "In our pursuit to mitigate spurious correlations, we emphasize the pivotal role of data quality by ensuring a balanced distribution of core features across various environmental contexts within our dataset. Given this focus, we employ a basic training approach for our model to learn from this well-prepared dataset." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 446, + 482, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 446, + 482, + 530 + ], + "spans": [ + { + "bbox": [ + 131, + 446, + 482, + 530 + ], + "type": "text", + "content": "We apply our data curation pipeline separately to the training " + }, + { + "bbox": [ + 131, + 446, + 482, + 530 + ], + "type": "inline_equation", + "content": "(SLIM_{Tr})" + }, + { + "bbox": [ + 131, + 446, + 482, + 530 + ], + "type": "text", + "content": " and validation " + }, + { + "bbox": [ + 131, + 446, + 482, + 530 + ], + "type": "inline_equation", + "content": "(SLIM_{Val})" + }, + { + "bbox": [ + 131, + 446, + 482, + 530 + ], + "type": "text", + "content": " sets. For the data constructed from the training set, we adopt the conventional ERM. For the dataset derived from the validation set, we employ the training technique presented by DFR [14]. It involves retraining only the final linear layer of a classification model that was initially trained using ERM. Note that our approach diverges from DFR practices by utilizing our curated dataset instead of a metagroup-balanced dataset." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 548, + 230, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 548, + 230, + 562 + ], + "spans": [ + { + "bbox": [ + 132, + 548, + 230, + 562 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 573, + 262, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 573, + 262, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 573, + 262, + 586 + ], + "type": "text", + "content": "5.1 Experimental Setup" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 594, + 482, + 666 + ], + "type": "text", + "content": "Datasets. Our study explores a range of datasets with inherent spurious correlations to challenge image classification models, including: (1) Waterbirds [27]. Contrasting landbirds and waterbirds against congruent/incongruent backgrounds, this dataset tests models' ability to focus on relevant features for bird type classification. (2) CelebA [19]. A comprehensive celebrity image dataset where the focus is on distinguishing hair color (blond vs. non-blond) to examine" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 176, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 176, + 91, + 448, + 103 + ], + "type": "text", + "content": "SLIM: Spuriousness Mitigation with Minimal Human Annotations" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": "biases in associating hair color with gender. (3) ISIC [5]. Aiming to differentiate benign from malignant melanoma in dermoscopic images, we target colorful patches as spurious features, aligning with previous studies [25]. (4) NICO. Derived from " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "\\mathrm{NICO}++" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": " [42], this dataset features various object categories in shifted contexts to probe spurious correlations. (5) ImageNet-9 [35]. This dataset is crafted from ImageNet to test models' robustness against background variations. Detailed information is available in Appx. A.4." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 206, + 482, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 206, + 482, + 291 + ], + "spans": [ + { + "bbox": [ + 130, + 206, + 482, + 291 + ], + "type": "text", + "content": "Baselines. We benchmark SLIM against a spectrum of SOTA spuriousness mitigation techniques, categorized as follows: (1) Techniques not reliant on spuriousness labels: EIIL [6], PGI [1], GEORGE [30], LfF [21], CIM [32], JTT [17], and CNC [41]; (2) Techniques requiring additional information such as predefined spuriousness concepts: DISC [34]; (3) Methods utilizing spuriousness labels from validation sets: SSA [22] and DFR [14]; (4) Approaches utilizing spuriousness labels from training sets: GDRO [27] and PDE [7]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 297, + 482, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 482, + 417 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 482, + 417 + ], + "type": "text", + "content": "Evaluation. Aligning with baselines, our experiments involve CNN models with ResNet50 architecture and GradCAM for visualizing model attention. For the capability of performance enhancement, we evaluate the worst-group and average accuracy. Additionally, we consider the necessity for extra labels and the additional (add.) training cost to evaluate computational efficiency. Note that the add. training cost is aligned with [13, 39], measuring add. data amounts required for second-stage training, where 1 means the amount of data equals the training set size. Furthermore, we examine the attention precision to evaluate whether model attention is corrected. All quantification experiments are based on five independent runs with different random seeds." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 423, + 481, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 423, + 481, + 449 + ], + "spans": [ + { + "bbox": [ + 131, + 423, + 481, + 449 + ], + "type": "text", + "content": "More experimental results with other architecture (ViT) and datasets (MetaShift [16] and FMoW [4]) are provided in Appx. A.8." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 132, + 505, + 482, + 654 + ], + "blocks": [ + { + "bbox": [ + 130, + 462, + 482, + 496 + ], + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 496 + ], + "type": "text", + "content": "Table 2: Worst-group and average accuracy (%) comparison on binary classification tasks, highlighting best and second-best performances. " + }, + { + "bbox": [ + 130, + 462, + 482, + 496 + ], + "type": "inline_equation", + "content": "\\spadesuit" + }, + { + "bbox": [ + 130, + 462, + 482, + 496 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 462, + 482, + 496 + ], + "type": "inline_equation", + "content": "\\diamond" + }, + { + "bbox": [ + 130, + 462, + 482, + 496 + ], + "type": "text", + "content": " indicate methods requiring spuriousness concepts and attention labels, respectively." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 505, + 482, + 654 + ], + "lines": [ + { + "bbox": [ + 132, + 505, + 482, + 654 + ], + "spans": [ + { + "bbox": [ + 132, + 505, + 482, + 654 + ], + "type": "table", + "html": "
MethodSpuriousness labelAdd. training costWaterbirdsCelebAISIC
WorstAvgWorstAvgWorstAvg
ERMX0×62.6±0.397.3±1.047.7±2.194.9±0.356.7±2.279.0±1.3
EIILX0×83.5±2.894.2±1.381.7±0.885.7±0.172.3±3.986.4±1.7
PGIX0×79.5±1.995.5±0.885.3±0.387.3±0.167.4±1.985.7±1.1
GEORGEX1×76.2±2.095.7±0.554.9±1.994.6±0.258.8±2.183.2±1.4
LfFX1×77.3±2.391.4±1.776.8±1.284.5±0.772.1±2.185.4±0.5
CIMX1×77.8±1.594.7±1.183.8±1.090.5±0.671.7±1.685.9±1.1
JTTX5×-6×83.1±3.590.6±0.381.5±1.788.1±0.328.1±4.386.2±0.3
CNCX2×-12×88.5±0.390.9±0.188.8±0.989.9±0.568.2±0.284.6±0.2
DISCX◇1×88.7±0.793.8±0.785.7±1.189.9±0.773.4±0.484.7±0.4
SLIMTrX◇0.1×-0.2×89.1±0.691.7±0.589.4±0.390.1±0.279.1±0.588.5±0.4
SLIMValX◇0.05×-0.3×91.0±0.493.8±0.489.4±0.590.7±0.274.0±0.285.4±0.2
SSAvalidation1.5×-5×89.0±0.692.2±0.989.8±1.392.8±0.161.7±0.777.7±1.0
DFRvalidation0.05×-0.3×89.6±0.792.1±0.287.3±1.090.2±0.871.9±0.784.3±0.3
GDROtraining0×89.9±0.692.6±0.188.9±1.393.9±0.166.8±3.580.5±1.9
PDEtraining0×85.9±2.790.5±1.885.0±4.092.3±0.866.4±2.382.2±1.3
", + "image_path": "0589e3ce026a905e800fcca963926a64b2e33b8f1b8a8e62a3261ad98aede71a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "text", + "content": "X. Xuan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 167, + 167, + 447, + 242 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 3: Worst-group and average accuracy (%) comparison on multi-classification tasks, highlighting best and second-best results. " + }, + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "inline_equation", + "content": "\\diamond" + }, + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": " indicates methods requiring attention labels. - represents inapplicable methods because such methods require spuriousness labels while the dataset does not provide them." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 167, + 167, + 447, + 242 + ], + "lines": [ + { + "bbox": [ + 167, + 167, + 447, + 242 + ], + "spans": [ + { + "bbox": [ + 167, + 167, + 447, + 242 + ], + "type": "table", + "html": "
MethodSpuriousness labelAdd. training costNICOImageNet9
WorstAvgWorstAvg
ERMX0×11.9±0.270.9±1.255.3±2.075.6±1.1
CNCX5×-12×72.6±1.789.5±0.658.9±0.674.3±0.2
SLIMTrX◇0.1×-0.2×78.0±0.892.1±0.168.1±0.374.5±0.2
SLIMValX◇0.1×-0.2×73.8±0.490.7±0.264.6±0.474.2±0.3
DFRvalidation0.4×70.9±2.185.8±0.6--
GDROtraining0×73.3±1.090.3±0.5--
", + "image_path": "6d6062decb3ee33841b4b96c986a76fad480734d075a79dcc125dd4b2f06b8dc.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 133, + 274, + 481, + 310 + ], + "blocks": [ + { + "bbox": [ + 130, + 243, + 479, + 264 + ], + "lines": [ + { + "bbox": [ + 130, + 243, + 479, + 264 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 479, + 264 + ], + "type": "text", + "content": "Table 4: SLIM's primary cost, including the data amount required for attention annotation (Att.), and the constructed data size for model training (Training)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 274, + 481, + 310 + ], + "lines": [ + { + "bbox": [ + 133, + 274, + 481, + 310 + ], + "spans": [ + { + "bbox": [ + 133, + 274, + 481, + 310 + ], + "type": "table", + "html": "
MethodWaterbirdCelebAISICNICOImageNet 9
Att.TrainingAtt.TrainingAtt.TrainingAtt.TrainingAtt.Training
SLIMTr120 (2.5%)0.1×200 (0.12%)0.1×120 (0.7%)0.2×40 (1.8%)0.2×180 (0.4%)0.1×
SLIMVal40 (3%)0.1×70 (0.35%)0.05×40 (2%)0.3×20 (2.7%)0.2×50 (1%)0.1×
", + "image_path": "1499461a1571aafc5beaafbe54dad50f3a4e5491894b856e4930a0a140f08a21.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 312, + 350, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 312, + 350, + 323 + ], + "spans": [ + { + "bbox": [ + 131, + 312, + 350, + 323 + ], + "type": "text", + "content": "5.2 Spuriousness Mitigation Performance" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 331, + 480, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 480, + 365 + ], + "type": "text", + "content": "In this section, we present the performance evaluation by comparing different methods across multiple binary and multiclass classification datasets mentioned in Sec. 5.1. Key findings are demonstrated below." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 366, + 481, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 366, + 481, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 366, + 481, + 485 + ], + "type": "text", + "content": "Superior worst-group performance. The metric of worst-group accuracy " + }, + { + "bbox": [ + 130, + 366, + 481, + 485 + ], + "type": "inline_equation", + "content": "(Acc_w)" + }, + { + "bbox": [ + 130, + 366, + 481, + 485 + ], + "type": "text", + "content": " is pivotal for spuriousness mitigation methods (refer to Sec. 3), and the results in Tables 2 and 3 showcase SLIM's superior capability in this regard. Across various classification tasks, SLIM consistently outperforms competing methods in achieving the highest " + }, + { + "bbox": [ + 130, + 366, + 481, + 485 + ], + "type": "inline_equation", + "content": "Acc_w" + }, + { + "bbox": [ + 130, + 366, + 481, + 485 + ], + "type": "text", + "content": " while maintaining high average accuracy. Specifically, in binary classification scenarios (Table 2), SLIM leads in performance for Waterbirds and ISIC and secures a close second in CelebA. In multi-classification tasks (Table 3), " + }, + { + "bbox": [ + 130, + 366, + 481, + 485 + ], + "type": "inline_equation", + "content": "SLIM_{Tr}" + }, + { + "bbox": [ + 130, + 366, + 481, + 485 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 366, + 481, + 485 + ], + "type": "inline_equation", + "content": "SLIM_{Val}" + }, + { + "bbox": [ + 130, + 366, + 481, + 485 + ], + "type": "text", + "content": " rank as the top performers across the board, underscoring the effectiveness of our approach in spuriousness mitigation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 486, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 481, + 665 + ], + "type": "text", + "content": "Cost-efficient spuriousness mitigation. Methods in Tables 2 and 3 are divided into four groups from top to bottom, as per the categorization established in Sec. 5.1 according to the computational cost. The \"Add. training cost\" columns denote the additional training cost for mitigating spuriousness apart from the initial training, such as additional learning phases (CNC, SSA) or data augmentation (JTT). Results in the table substantiate that methods with higher costs (higher Add. training costs or more spuriousness labels) tend to outperform others, aligning with prior research findings. Nevertheless, among the well-performed methods, it is worth mentioning that " + }, + { + "bbox": [ + 130, + 486, + 481, + 665 + ], + "type": "inline_equation", + "content": "SLIM" + }, + { + "bbox": [ + 130, + 486, + 481, + 665 + ], + "type": "text", + "content": " is more cost-efficient: " + }, + { + "bbox": [ + 130, + 486, + 481, + 665 + ], + "type": "inline_equation", + "content": "SLIM_{Tr}" + }, + { + "bbox": [ + 130, + 486, + 481, + 665 + ], + "type": "text", + "content": " used 0.1-0.2 proportion of training sets and " + }, + { + "bbox": [ + 130, + 486, + 481, + 665 + ], + "type": "inline_equation", + "content": "SLIM_{Val}" + }, + { + "bbox": [ + 130, + 486, + 481, + 665 + ], + "type": "text", + "content": " used 0.05-0.3 proportion of validation sets for additional training, which is the same as DFR and much smaller than others. Moreover, " + }, + { + "bbox": [ + 130, + 486, + 481, + 665 + ], + "type": "inline_equation", + "content": "SLIM" + }, + { + "bbox": [ + 130, + 486, + 481, + 665 + ], + "type": "text", + "content": " requires labels for attention correctness, which is more cost-efficient to obtain than spuriousness labels (refer to Sec. 5.3 for more details). Table 4 presents " + }, + { + "bbox": [ + 130, + 486, + 481, + 665 + ], + "type": "inline_equation", + "content": "SLIM" + }, + { + "bbox": [ + 130, + 486, + 481, + 665 + ], + "type": "text", + "content": "'s primary cost, including the training data amount for mitigating spuriousness, which is the detailed version of \"Add. train" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 176, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 176, + 91, + 447, + 102 + ], + "type": "text", + "content": "SLIM: Spuriousness Mitigation with Minimal Human Annotations" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 148, + 360, + 235 + ], + "blocks": [ + { + "bbox": [ + 140, + 148, + 360, + 235 + ], + "lines": [ + { + "bbox": [ + 140, + 148, + 360, + 235 + ], + "spans": [ + { + "bbox": [ + 140, + 148, + 360, + 235 + ], + "type": "image", + "image_path": "339c57f4cfb0e174ef20e2d2c3c0ebc8ecbdeda076aa133cd9baeb2516005855.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 387, + 146, + 470, + 235 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "text", + "content": "Table 5: Attention accuracy evaluation. (Left) GradCAM visualization of model attention. (Right) Worst group and average attention accuracy of selected methods." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 387, + 146, + 470, + 235 + ], + "lines": [ + { + "bbox": [ + 387, + 146, + 470, + 235 + ], + "spans": [ + { + "bbox": [ + 387, + 146, + 470, + 235 + ], + "type": "table", + "html": "
WaterbirdsAIoU
WorstAvg
ERM.25±.05.47±.03
CNC.48±.05.49±.04
GDRO.49±.03.52±.03
PDE.50±.02.52±.03
SLIMTr.53±.02.61±.02
ImageNet 9AIoU
WorstAvg
ERM.33±.07.55±.03
CNC.44±.03.61±.03
SLIMTr.58±.04.68±.03
", + "image_path": "e398f742806b00861e83c37042025ed288352d685411b4dc361db3c8cd56c557.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 236, + 479, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 479, + 284 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 479, + 284 + ], + "type": "text", + "content": "ing cost\" in Tables 2 and 3, and the data amount requiring attention labels (Att.). Notably, our approach only requires no more than " + }, + { + "bbox": [ + 132, + 236, + 479, + 284 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 132, + 236, + 479, + 284 + ], + "type": "text", + "content": " of the corresponding data split (train or validation) for attention-label annotation, further proving the cost-efficiency of SLIM." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 285, + 482, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 482, + 452 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 482, + 452 + ], + "type": "text", + "content": "Enhanced attention accuracy. Beyond classification accuracy, we further evaluate the performance of spuriousness mitigation methods via attention accuracy. Specifically, for models resulting from each method, we randomly sample data instances to visualize model attention, and measure attention accuracy scores in both worst-group and dataset levels, consistent with our evaluation of classification accuracy. Align with the related work [40], we employ the Adjusted Intersection-over-Union (AIoU) score as our attention accuracy measurement, which is an adjusted IoU score with a threshold, aiming at measuring the overlap between model's highest attributed region and the ground-truth bounding box. Details related to AIoU are given in Appx. A.7. Table 5 presents the evaluation results. The left GradCAM examples showcase SLIM's capability of correcting model's wrong attention, and the right table compares SLIM with SOTA methods, where we can find our method secures the best performance in terms of both worst-group and average attention accuracy." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 468, + 324, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 324, + 480 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 324, + 480 + ], + "type": "text", + "content": "5.3 Attention Labeling Performance" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 486, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 629 + ], + "type": "text", + "content": "To compare the processes of annotating spuriousness versus attention correctness, we employed crowdsourcing tasks using the Waterbirds and NICO datasets, as both of which come with ground-truth spuriousness labels. For this study, we selected a random set of 120 images from each dataset and established two separate tasks: one for spuriousness labeling and one for attention correctness labeling. For each task, we recruited 60 workers independently from the Prolific platform to prevent learning biases from cross-task participation. We meticulously recorded the time taken to label each image and reviewed all submissions to exclude any outliers with unusually long or short annotation times. Ensuring ethical research standards, our study avoided collecting personally identifiable information and excluded any potentially offensive content. Detailed instructions provided to the crowdsourcing participants can be found in Appx. A.6." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "content": "Evaluation Consistency. Table 6 presents metrics for this study. While the annotation time for Waterbirds remains similar across tasks, attention correctness for NICO is annotated more quickly than the other. This difference may stem" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "text", + "content": "X. Xuan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 163, + 271, + 222 + ], + "blocks": [ + { + "bbox": [ + 132, + 121, + 272, + 153 + ], + "lines": [ + { + "bbox": [ + 132, + 121, + 272, + 153 + ], + "spans": [ + { + "bbox": [ + 132, + 121, + 272, + 153 + ], + "type": "text", + "content": "Table 6: Comparison of crowdsourced labeling tasks: spuriousness vs. attention correctness." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 163, + 271, + 222 + ], + "lines": [ + { + "bbox": [ + 135, + 163, + 271, + 222 + ], + "spans": [ + { + "bbox": [ + 135, + 163, + 271, + 222 + ], + "type": "table", + "html": "
WaterbirdsAvg. TimeAvg. Consis.
Spuriousness4.9±2.3 sec81.9±14.8%
Att.4.5±1.4 sec94.6±10.5%
NICOAvg. TimeAvg. Consis.
Spuriousness7.9±3.8 sec70.1±19.1%
Att.5.3±1.6 sec95.1±6.1%
", + "image_path": "6c0101a6dc68b2589b4520a97959d1928b1e9456a3b5f9fafe0a88cf29cb0947.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 278, + 114, + 461, + 175 + ], + "blocks": [ + { + "bbox": [ + 278, + 114, + 461, + 175 + ], + "lines": [ + { + "bbox": [ + 278, + 114, + 461, + 175 + ], + "spans": [ + { + "bbox": [ + 278, + 114, + 461, + 175 + ], + "type": "image", + "image_path": "0e4c5b3e9219ea3099038f43ce33447737c73f5a34bbff9735ddbce52d86de9b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 273, + 182, + 463, + 226 + ], + "lines": [ + { + "bbox": [ + 273, + 182, + 463, + 226 + ], + "spans": [ + { + "bbox": [ + 273, + 182, + 463, + 226 + ], + "type": "text", + "content": "Fig. 4: Comparison of neighbor consistency over original feature space and our attention space on Waterbirds, w.r.t. model attention (left) and data features (right)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 226, + 481, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 226, + 481, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 481, + 308 + ], + "type": "text", + "content": "from NICO's broader array of spuriousness choices versus the binary attention annotations. To measure the annotation consistency—whether a single receives divergent labels—we use the percentage of the most frequent label as a consensus measure. A higher percentage suggests greater consistency. Our findings indicate that attention correctness annotations exhibit a notably higher consistency, suggesting that it is less ambiguous to determine correct attention than to identify spurious features." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 309, + 481, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 309, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 309, + 481, + 453 + ], + "type": "text", + "content": "Attention Consistency. " + }, + { + "bbox": [ + 130, + 309, + 481, + 453 + ], + "type": "inline_equation", + "content": "SLIM" + }, + { + "bbox": [ + 130, + 309, + 481, + 453 + ], + "type": "text", + "content": " requires spreading labels in a space where attention patterns are consistently similar in nearby areas. To evaluate its feasibility, we compared how similar neighbors are in the original feature space versus the attention space created by " + }, + { + "bbox": [ + 130, + 309, + 481, + 453 + ], + "type": "inline_equation", + "content": "SLIM" + }, + { + "bbox": [ + 130, + 309, + 481, + 453 + ], + "type": "text", + "content": ". We used k-means clustering with varying numbers of clusters to measure the intra-cluster similarity for both the model's attention matrices and the feature vectors. A higher similarity score means the space has more consistent patterns, which is essential for the accurate spreading of labels. As Fig. 4 shows, our attention space significantly improves consistency in the model's attention, while also maintaining a good level of data feature consistency. Increasing the number of clusters generally increases consistency, but there is a point where adding more clusters does not lead to further gains, as shown by the curve flattening at " + }, + { + "bbox": [ + 130, + 309, + 481, + 453 + ], + "type": "inline_equation", + "content": "k = 120" + }, + { + "bbox": [ + 130, + 309, + 481, + 453 + ], + "type": "text", + "content": " in the left plot." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 453, + 481, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 453, + 481, + 513 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 481, + 513 + ], + "type": "text", + "content": "In summary, we have demonstrated that the annotation module of SLIM is less time-consuming and yields more consistent labeling feedback. Additionally, the attention space engineered by SLIM establishes a solid foundation for propagating attention labels, considering the consistent model attention observed among neighboring instances." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 528, + 239, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 528, + 239, + 540 + ], + "spans": [ + { + "bbox": [ + 132, + 528, + 239, + 540 + ], + "type": "text", + "content": "5.4 Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 545, + 481, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 481, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 481, + 641 + ], + "type": "text", + "content": "We investigate two critical factors of SLIM, which are detailed below. Other factors such as the dimension of the attention space are discussed in Appx. A.8. Influence of data amount for attention annotations. We change the data amount for attention annotations to investigate their effect on performance, keeping other variables constant. Table 7 illustrates that both the worst-group and average accuracy tend to rise with more annotations, while such incremental benefit stabilizes after reaching a certain point, helping us identify an optimal number of annotations that balance performance with annotation cost." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Influence of training set size. This size refers to the size of the data subset constructed in SLIM's final phase for training a robust model. By only varying" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 176, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 176, + 91, + 448, + 102 + ], + "type": "text", + "content": "SLIM: Spuriousness Mitigation with Minimal Human Annotations" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 157, + 146, + 457, + 257 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 7: Performance variation with different attention annotation amounts in SLIM. Marks in table: best and second best performance; the amounts used in Table 2." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 157, + 146, + 457, + 257 + ], + "lines": [ + { + "bbox": [ + 157, + 146, + 457, + 257 + ], + "spans": [ + { + "bbox": [ + 157, + 146, + 457, + 257 + ], + "type": "table", + "html": "
MethodWaterbirdsCelebAISIC
AmountWorstAvgAmountWorstAvgAmountWorstAvg
SLIMTr6078.2±0.586.0±0.310086.4±0.290.2±0.16075.3±0.286.6±0.9
9085.9±0.989.3±0.315087.8±0.491.1±0.79074.4±0.487.4±0.7
12089.1±0.691.7±0.520089.4±0.490.1±0.212079.1±0.588.5±0.4
15089.4±0.691.7±0.225089.4±0.290.9±0.115079.9±0.888.8±0.7
18089.3±0.391.0±1.530089.5±0.191.2±0.218079.8±0.388.6±0.3
SLIMVal1078.9±8.786.5±5.14084.7±0.487.4±0.31070.5±0.583.3±1.2
2589.2±0.693.4±1.05587.6±0.390.4±0.62572.7±0.185.2±0.8
4091.0±0.493.8±0.47089.4±0.490.7±0.24074.0±0.285.4±0.2
5591.5±0.393.8±0.18589.7±0.190.5±0.45574.0±0.285.9±0.5
7090.7±0.394.0±0.310089.7±0.290.5±0.27074.1±0.285.8±0.2
", + "image_path": "15fd323e671d15385071fe992d8ea68532b7a4900d3076f2914670e0447bdeaf.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 157, + 291, + 457, + 384 + ], + "blocks": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "text", + "content": "Table 8: Performance variation with different training sizes. " + }, + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "inline_equation", + "content": "SLIM_{Tr}" + }, + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "inline_equation", + "content": "SLIM_{Val}" + }, + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "text", + "content": " are trained on different proportions (Prop) of training sets and validation sets, respectively." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 157, + 291, + 457, + 384 + ], + "lines": [ + { + "bbox": [ + 157, + 291, + 457, + 384 + ], + "spans": [ + { + "bbox": [ + 157, + 291, + 457, + 384 + ], + "type": "table", + "html": "
MethodWaterbirdsCelebAISIC
Prop%WorstAvgProp%WorstAvgProp%WorstAvg
SLIMTr582.4±1.083.6±1.3588.9±0.190.0±0.11577.6±2.187.1±1.2
1089.1±0.691.7±0.51089.4±0.390.1±0.22079.1±0.588.5±0.4
1282.3±1.387.8±0.31589.4±0.491.0±0.42577.4±1.588.2±1.2
1583.5±2.789.6±0.52088.9±0.190.8±0.33078.3±1.588.3±0.9
SLIMVal584.2±3.689.2±2.6385.7±1.388.5±1.12568.2±2.181.9±0.8
1091.0±0.493.8±0.4589.4±0.490.7±0.23074.0±0.285.4±0.2
1287.6±0.693.9±1.31088.3±1.490.3±0.53572.5±0.585.3±0.1
1586.8±2.094.6±0.51587.2±0.490.3±0.44070.9±1.285.4±3.5
", + "image_path": "82001d851d184b45d5c1f91e7eaf2c3237add32877152dad009ab10af44ce692.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 388, + 480, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 388, + 480, + 446 + ], + "spans": [ + { + "bbox": [ + 132, + 388, + 480, + 446 + ], + "type": "text", + "content": "this parameter (Table 8), we find that the optimal size is inherently linked to the size of our constructed subgroups, " + }, + { + "bbox": [ + 132, + 388, + 480, + 446 + ], + "type": "inline_equation", + "content": "|g_{\\hat{c}_i,e_i}|" + }, + { + "bbox": [ + 132, + 388, + 480, + 446 + ], + "type": "text", + "content": ". An optimal size is reached when data from various groups are sufficiently sampled without redundant oversampling. These findings guide the hyperparameter selection in this phase, ensuring an effective cost allocation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 469, + 218, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 469, + 218, + 481 + ], + "spans": [ + { + "bbox": [ + 133, + 469, + 218, + 481 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 497, + 480, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 497, + 480, + 604 + ], + "spans": [ + { + "bbox": [ + 132, + 497, + 480, + 604 + ], + "type": "text", + "content": "In this paper, we introduce SLIM, a novel framework for spuriousness mitigating in deep learning. With minimal human annotation effort, SLIM provides a data construction pipeline to create a feature-balanced subset for training a robust model. Extensive evaluation demonstrates the superior cost-effectiveness of SLIM, which matches or outperforms state-of-the-art methods with far less costs. The success of SLIM underscores the potential of data quality and efficient human supervision in the development of robust models. We believe that SLIM will inspire further research in this direction, ultimately leading to more cost-effective machine learning systems for reliable AI." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 605, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 605, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 605, + 480, + 665 + ], + "type": "text", + "content": "Acknowledgements. This research is supported in part by Bosch Research and the National Institute of Health with grants P41-EB032840 and R01CA270454. H.-T. Lin is partially supported by the National Taiwan University Center for Data Intelligence via NTU-112L900901 and the Ministry of Science and Technology in Taiwan via MOST 112-2628-E-002-030." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "text", + "content": "X. Xuan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 137, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 137, + 480, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 137, + 480, + 169 + ], + "spans": [ + { + "bbox": [ + 138, + 137, + 480, + 169 + ], + "type": "text", + "content": "1. Ahmed, F., Bengio, Y., Van Seijen, H., Courville, A.: Systematic generalisation with group invariant predictions. In: International Conference on Learning Representations (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 170, + 480, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 170, + 480, + 212 + ], + "spans": [ + { + "bbox": [ + 138, + 170, + 480, + 212 + ], + "type": "text", + "content": "2. Bontempelli, A., Teso, S., Tentori, K., Giunchiglia, F., Passerini, A.: Concept-level debugging of part-prototype networks. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=oiwXwPDTyNk" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 213, + 480, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 213, + 480, + 245 + ], + "spans": [ + { + "bbox": [ + 138, + 213, + 480, + 245 + ], + "type": "text", + "content": "3. Chen, Z., Deng, Y., Wu, Y., Gu, Q., Li, Y.: Towards understanding the mixture-of-experts layer in deep learning. Advances in neural information processing systems 35, 23049-23062 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 245, + 480, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 245, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 138, + 245, + 480, + 266 + ], + "type": "text", + "content": "4. Christie, G., Fendley, N., Wilson, J., Mukherjee, R.: Functional map of the world. In: CVPR (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 266, + 480, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 266, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 138, + 266, + 480, + 331 + ], + "type": "text", + "content": "5. Codella, N.C.F., Gutman, D., Celebi, M.E., Helba, B., Marchetti, M.A., Dusza, S.W., Kalloo, A., Liopyris, K., Mishra, N., Kittler, H., Halpern, A.: Skin lesion analysis toward melanoma detection: A challenge at the 2017 international symposium on biomedical imaging (isbi), hosted by the international skin imaging collaboration (isic). In: 2018 IEEE 15th International Symposium on Biomedical Imaging (ISBI 2018). pp. 168-172 (2018). https://doi.org/10.1109/ISBI.2018.8363547" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 332, + 480, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 332, + 480, + 364 + ], + "spans": [ + { + "bbox": [ + 138, + 332, + 480, + 364 + ], + "type": "text", + "content": "6. Creager, E., Jacobsen, J.H., Zemel, R.: Environment inference for invariant learning. In: International Conference on Machine Learning. pp. 2189-2200. PMLR (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 364, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 364, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 138, + 364, + 480, + 407 + ], + "type": "text", + "content": "7. Deng, Y., Yang, Y., Mirzasoleiman, B., Gu, Q.: Robust learning with progressive data expansion against spurious correlation. In: Thirty-seventh Conference on Neural Information Processing Systems (2023), https://openreview.net/forum?id=9QEVJ9qm46" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 407, + 480, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 407, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 138, + 407, + 480, + 449 + ], + "type": "text", + "content": "8. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 449, + 480, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 449, + 480, + 494 + ], + "spans": [ + { + "bbox": [ + 138, + 449, + 480, + 494 + ], + "type": "text", + "content": "9. Gao, Y., Sun, T.S., Zhao, L., Hong, S.R.: Aligning eyes between humans and deep neural network through interactive attention alignment. Proc. ACM Hum-. Comput. Interact. 6(CSCW2) (nov 2022). https://doi.org/10.1145/3555590, https://doi.org/10.1145/3555590" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 494, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 494, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 138, + 494, + 480, + 525 + ], + "type": "text", + "content": "0. Gao, Y., Sun, T.S., Zhao, L., Hong, S.R.: Aligning eyes between humans and deep neural network through interactive attention alignment. Proceedings of the ACM on Human-Computer Interaction 6(CSCW2), 1-28 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 525, + 480, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 525, + 480, + 548 + ], + "spans": [ + { + "bbox": [ + 138, + 525, + 480, + 548 + ], + "type": "text", + "content": "1. He, H., Garcia, E.A.: Learning from imbalanced data. IEEE Transactions on knowledge and data engineering 21(9), 1263-1284 (2009)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 548, + 480, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 548, + 480, + 579 + ], + "spans": [ + { + "bbox": [ + 138, + 548, + 480, + 579 + ], + "type": "text", + "content": "2. Jelassi, S., Li, Y.: Towards understanding how momentum improves generalization in deep learning. In: International Conference on Machine Learning. pp. 9965-10040. PMLR (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 579, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 579, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 138, + 579, + 480, + 611 + ], + "type": "text", + "content": "3. Joshi, S., Yang, Y., Xue, Y., Yang, W., Mirzasoleiman, B.: Towards mitigating spurious correlations in the wild: A benchmark & a more realistic dataset. arXiv preprint arXiv:2306.11957 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 611, + 480, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 480, + 633 + ], + "type": "text", + "content": "4. Kirichenko, P., Izmailov, P., Wilson, A.G.: Last layer re-training is sufficient for robustness to spurious correlations. arXiv preprint arXiv:2204.02937 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "text", + "content": "5. Lee, S., Payani, A., Chau, D.H.P.: Towards mitigating spurious correlations in image classifiers with simple yes-no feedback. In: AI and HCI Workshop at the 40th International Conference on Machine Learning (2023)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 176, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 176, + 91, + 447, + 102 + ], + "type": "text", + "content": "SLIM: Spuriousness Mitigation with Minimal Human Annotations" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "16. Liang, W., Zou, J.: Metashift: A dataset of datasets for evaluating contextual distribution shifts and training conflicts. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=MTex8qKavoS" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 481, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 481, + 192 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 481, + 192 + ], + "type": "text", + "content": "17. Liu, E.Z., Haghgoo, B., Chen, A.S., Raghunathan, A., Koh, P.W., Sagawa, S., Liang, P., Finn, C.: Just train twice: Improving group robustness without training group information. In: International Conference on Machine Learning. pp. 6781-6792. PMLR (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 192, + 480, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 192, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 133, + 192, + 480, + 236 + ], + "type": "text", + "content": "18. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 236, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 480, + 267 + ], + "type": "text", + "content": "19. Liu, Z., Luo, P., Wang, X., Tang, X.: Deep learning face attributes in the wild. In: Proceedings of the IEEE international conference on computer vision. pp. 3730-3738 (2015)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 268, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 268, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 268, + 480, + 289 + ], + "type": "text", + "content": "20. McInnes, L., Healy, J., Melville, J.: Umap: Uniform manifold approximation and projection for dimension reduction. arXiv preprint arXiv:1802.03426 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 289, + 480, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 480, + 320 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 480, + 320 + ], + "type": "text", + "content": "21. Nam, J., Cha, H., Ahn, S., Lee, J., Shin, J.: Learning from failure: De-biasing classifier from biased classifier. Advances in Neural Information Processing Systems 33, 20673-20684 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 321, + 480, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 321, + 480, + 353 + ], + "spans": [ + { + "bbox": [ + 132, + 321, + 480, + 353 + ], + "type": "text", + "content": "22. Nam, J., Kim, J., Lee, J., Shin, J.: Spread spurious attribute: Improving worst-group accuracy with spurious attribute estimation. arXiv preprint arXiv:2204.02070 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 353, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 353, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 353, + 480, + 374 + ], + "type": "text", + "content": "23. Petsiuk, V., Das, A., Saenko, K.: Rise: Randomized input sampling for explanation of black-box models. arXiv preprint arXiv:1806.07421 (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 374, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 480, + 396 + ], + "type": "text", + "content": "24. Rao, S., Böhle, M., Parchami-Araghi, A., Schiele, B.: Using explanations to guide models. arXiv preprint arXiv:2303.11932 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 396, + 480, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 480, + 428 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 480, + 428 + ], + "type": "text", + "content": "25. Rieger, L., Singh, C., Murdoch, W.J., Yu, B.: Interpretations are useful: Penalizing explanations to align neural networks with prior knowledge. In: Proceedings of the 37th International Conference on Machine Learning. ICML'20, JMLR.org (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 428, + 480, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 428, + 480, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 428, + 480, + 482 + ], + "type": "text", + "content": "26. Ross, A.S., Hughes, M.C., Doshi-Velez, F.: Right for the right reasons: Training differentiable models by constraining their explanations. In: Proceedings of the Twenty-Sixth International Joint Conference on Artificial Intelligence, IJCAI-17. pp. 2662-2670 (2017). https://doi.org/10.24963/ijcai.2017/371, https://doi.org/10.24963/ijcai.2017/371" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 482, + 480, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 482, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 482, + 480, + 514 + ], + "type": "text", + "content": "27. Sagawa*, S., Koh*, P.W., Hashimoto, T.B., Liang, P.: Distributionally robust neural networks. In: International Conference on Learning Representations (2020), https://openreview.net/forum?id=ryxGuJrFvS" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 514, + 480, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 480, + 557 + ], + "type": "text", + "content": "28. Schramowski, P., Stammer, W., Teso, S., Brugger, A., Herbert, F., Shao, X., Luigs, H.G., Mahlein, A.K., Kersting, K.: Making deep neural networks right for the right scientific reasons by interacting with their explanations. Nature Machine Intelligence 2(8), 476-486 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 557, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 480, + 601 + ], + "type": "text", + "content": "29. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Gradcam: Visual explanations from deep networks via gradient-based localization. In: Proceedings of the IEEE international conference on computer vision. pp. 618-626 (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 601, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 480, + 632 + ], + "type": "text", + "content": "30. Sohoni, N., Dunnmon, J., Angus, G., Gu, A., Ré, C.: No subclass left behind: Fine-grained robustness in coarse-grained classification problems. Advances in Neural Information Processing Systems 33, 19339-19352 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "31. Srivastava, M., Hashimoto, T., Liang, P.: Robustness to spurious correlations via human annotations. In: International Conference on Machine Learning. pp. 9109-9119. PMLR (2020)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "text", + "content": "X. Xuan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 512 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "32. Taghanaki, S.A., Choi, K., Khasahmadi, A.H., Goyal, A.: Robust representation learning via perceptual similarity metrics. In: International Conference on Machine Learning. pp. 10043-10053. PMLR (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 482, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 482, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 482, + 194 + ], + "type": "text", + "content": "33. Wang, H., Wang, Z., Du, M., Yang, F., Zhang, Z., Ding, S., Mardziel, P., Hu, X.: Score-cam: Score-weighted visual explanations for convolutional neural networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 24-25 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 194, + 482, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 194, + 482, + 226 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 482, + 226 + ], + "type": "text", + "content": "34. Wu, S., Yuksekgonul, M., Zhang, L., Zou, J.: Discover and cure: Concept-aware mitigation of spurious correlation. In: Proceedings of the 40th International Conference on Machine Learning. ICML'23, JMLR.org (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 226, + 482, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 226, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 482, + 248 + ], + "type": "text", + "content": "35. Xiao, K., Engstrom, L., Ilyas, A., Madry, A.: Noise or signal: The role of image backgrounds in object recognition. ArXiv preprint arXiv:2006.09994 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 248, + 482, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 482, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 482, + 281 + ], + "type": "text", + "content": "36. Xuan, X., Ono, J.P., Gou, L., Ma, K.L., Ren, L.: Attributionscanner: A visual analytics system for metadata-free data-slicing based model validation. arXiv preprint arXiv:2401.06462 (2024)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 281, + 482, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 303 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 303 + ], + "type": "text", + "content": "37. Xue, Y., Payani, A., Yang, Y., Mirzasoleiman, B.: Eliminating spurious correlations from pre-trained models via data mixing. arXiv preprint arXiv:2305.14521 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 303, + 482, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 303, + 482, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 303, + 482, + 346 + ], + "type": "text", + "content": "38. Yan, S., Yu, Z., Zhang, X., Mahapatra, D., Chandra, S.S., Janda, M., Soyer, P., Ge, Z.: Towards trustable skin cancer diagnosis via rewriting model's decision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11568-11577 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 346, + 482, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 482, + 391 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 482, + 391 + ], + "type": "text", + "content": "39. Yang, Y., Gan, E., Karolina Dziugaite, G., Mirzasoleiman, B.: Identifying spurious biases early in training through the lens of simplicity bias. In: Proceedings of The 27th International Conference on Artificial Intelligence and Statistics. Proceedings of Machine Learning Research, vol. 238, pp. 2953-2961. PMLR (02-04 May 2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 391, + 482, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 391, + 482, + 423 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 482, + 423 + ], + "type": "text", + "content": "40. Yang, Y., Nushi, B., Palangi, H., Mirzasoleiman, B.: Mitigating spurious correlations in multi-modal models during fine-tuning. In: Proceedings of the 40th International Conference on Machine Learning. ICML'23, JMLR.org (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 423, + 482, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 482, + 456 + ], + "type": "text", + "content": "41. Zhang, M., Sohoni, N.S., Zhang, H.R., Finn, C., Ré, C.: Correct-n-contrast: A contrastive approach for improving robustness to spurious correlations. arXiv preprint arXiv:2203.01517 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 456, + 482, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 456, + 482, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 456, + 482, + 489 + ], + "type": "text", + "content": "42. Zhang, X., He, Y., Xu, R., Yu, H., Shen, Z., Cui, P.: Nico++: Towards better benchmarking for domain generalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16036-16047 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 489, + 482, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 489, + 482, + 512 + ], + "spans": [ + { + "bbox": [ + 130, + 489, + 482, + 512 + ], + "type": "text", + "content": "43. Zhou, D., Bousquet, O., Lal, T., Weston, J., Schölkopf, B.: Learning with local and global consistency. Advances in neural information processing systems 16 (2003)" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 176, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 176, + 91, + 447, + 102 + ], + "type": "text", + "content": "SLIM: Spuriousness Mitigation with Minimal Human Annotations" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/SMFANet_ A Lightweight Self-Modulation Feature Aggregation Network for Efficient Image Super-Resolution/b8afe796-2e9d-45e7-96aa-cb21a9d8b896_content_list.json b/2024/SMFANet_ A Lightweight Self-Modulation Feature Aggregation Network for Efficient Image Super-Resolution/b8afe796-2e9d-45e7-96aa-cb21a9d8b896_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..449ca3cd89391f8a4b4df8a32a25dcc7964a39de --- /dev/null +++ b/2024/SMFANet_ A Lightweight Self-Modulation Feature Aggregation Network for Efficient Image Super-Resolution/b8afe796-2e9d-45e7-96aa-cb21a9d8b896_content_list.json @@ -0,0 +1,2272 @@ +[ + { + "type": "text", + "text": "SMFANet: A Lightweight Self-Modulation Feature Aggregation Network for Efficient Image Super-Resolution", + "text_level": 1, + "bbox": [ + 259, + 140, + 743, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mingjun Zheng*, Long Sun*, Jiangxin Dong, and Jinshan Pan†", + "bbox": [ + 269, + 234, + 730, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "School of Computer Science and Engineering, Nanjing University of Science and Technology, Nanjing, China {mingjunzheng, cs.longsun, jxdong, jspan}@njust.edu.cn", + "bbox": [ + 290, + 262, + 712, + 305 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Transformer-based restoration methods achieve significant performance as the self-attention (SA) of the Transformer can explore non-local information for better high-resolution image reconstruction. However, the key dot-product SA requires substantial computational resources, which limits its application in low-power devices. Moreover, the low-pass nature of the SA mechanism limits its capacity for capturing local details, consequently leading to smooth reconstruction results. To address these issues, we propose a self-modulation feature aggregation (SMFA) module to collaboratively exploit both local and non-local feature interactions for a more accurate reconstruction. Specifically, the SMFA module employs an efficient approximation of self-attention (EASA) branch to model non-local information and uses a local detail estimation (LDE) branch to capture local details. Additionally, we further introduce a partial convolution-based feed-forward network (PCFN) to refine the representative features derived from the SMFA. Extensive experiments show that the proposed SMFANet family achieve a better trade-off between reconstruction performance and computational efficiency on public benchmark datasets. In particular, compared to the $\\times 4$ SwinIR-light, SMFANet+ achieves 0.14dB higher performance over five public testsets on average, and $\\times 10$ times faster runtime, with only about $43\\%$ of the model complexity (e.g., FLOPs). Our source codes and pretrained models are available at: https://github.com/Zheng-MJ/SMFANet.", + "bbox": [ + 261, + 332, + 745, + 636 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Efficient image SR $\\cdot$ Self-attention $\\cdot$ Feature aggregation", + "bbox": [ + 261, + 648, + 723, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 685, + 375, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Single-image super-resolution (SISR) refers to restoring a high-resolution (HR) image from the given degraded low-resolution (LR) counterpart. With the rapid rise of streaming platforms and ultra-high-definition (UHD) content becoming the industrial standard, SISR has attracted significant attention from academia to industry for its ability to provide a pleasurable watching experience at a lower cost of media transmission. Concurrently, the ill-posed nature of SISR", + "bbox": [ + 212, + 713, + 787, + 805 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution", + "bbox": [ + 232, + 811, + 374, + 824 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 232, + 824, + 395, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/78771184017821c5d09f37149b61361ade882cb0edce7faad22089e66e9b8639.jpg", + "image_caption": [ + "Fig. 1: Overall performance comparison between our proposed method and other state-of-the-art lightweight methods on Manga109 [33] for $\\times 4$ SR. The circle sizes in the left and right figures represent the number of FLOPs of the model and the GPU usage required for inference, respectively. The proposed SMFANet family achieve a better trade-off between computational efficiency and reconstruction performance." + ], + "image_footnote": [], + "bbox": [ + 236, + 143, + 500, + 297 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/10722de5e0c7e0ece6e0ccef1eeef7780e804dc370b6a9a0dc192598a42e660f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 143, + 764, + 297 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "task makes it challenging for conventional hand-crafted prior-based [8,39,44], or interpolation-based approaches to decently solve this problem.", + "bbox": [ + 212, + 414, + 782, + 444 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Over the past decade, deep learning has revolutionized the field of SISR. Various convolutional neural networks (CNNs) [1,9,10,17,24,26,35-37,50] have been developed to solve this problem. As the basic operation of CNNs, the convolution operator is translation-invariant and has limited receptive fields, which limits its ability to model non-local information. For better image restoration, CNN-based methods [26, 50, 51] become deeper and larger to enhance the representation ability. These high-capacity networks consume high computational costs, e.g., RCAN [50] has 15.59M parameters with over 400 layers, which cannot be applied to source-limited devices, e.g., smart phones.", + "bbox": [ + 212, + 445, + 784, + 580 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, the vision Transformer (ViT) [12] achieves impressive success on both high-level [29] and low-level vision tasks [4, 7, 21, 25, 27, 41, 47, 49, 53], as the self-attention (SA) mechanism in the ViT can model non-local information effectively. However, the SA mechanism requires high computational resources with massive memory consumption. To reduce the computational cost, window-based self-attention [7, 25, 29, 49], transposed SA [45], and weight multiplexing strategy [48] have been developed to ease the computational burden. However, these still require a long time to learn feature dependencies for image super-resolution.", + "bbox": [ + 212, + 582, + 785, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Furthermore, recent studies [11, 34] reveal that ViTs favor learning low-frequency components, which leads to smoothed reconstruction results. Thus, these limitations motivate us to think: Is it possible to develop an efficient feature modulation block that explores non-local information in a SA-like manner while modeling local details for efficient image super-resolution?", + "bbox": [ + 212, + 718, + 785, + 794 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we propose a self-modulation feature aggregation (SMFA) module that contains an efficient approximation of self-attention (EASA) branch to exploit non-local information, and a local detail estimation (LDE) branch to", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Zheng et al.", + "bbox": [ + 271, + 114, + 352, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "model local features. On the EASA branch, we acquire the low-frequency contents through a downsampling operation, calculate the global variance of the input features to modulate the processed low-frequency features, and then use the modulated features to aggregate the input features adaptively. As the EASA prioritizes non-local structure information exploration, we use the LDE branch with convolutional layers to capture local features in parallel.", + "bbox": [ + 212, + 146, + 784, + 236 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In addition, we develop an efficient partial convolution-based feed-forward network (PCFN), which further refines the representative features derived from the SMFA in both spatial and channel dimensions. We formulate the proposed SMFA module and PCFN into an end-to-end trainable network, called SMFANet, to solve SISR. Extensive experiments show that the proposed SMFANet achieves a favorable trade-off between computational efficiency and reconstruction performance (see Figure 1).", + "bbox": [ + 212, + 237, + 785, + 340 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The main contributions are summarized as follows:", + "bbox": [ + 240, + 343, + 607, + 357 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We develop an efficient SMFA module to extract representative features, where the EASA branch for exploring non-local information and the LDE branch for capturing local features.", + "- We present a lightweight PCFN to refine the features derived from the SMFA in spatial and channel dimensions further.", + "- We conduct quantitative and qualitative evaluations for our proposed SMFANet on public benchmark datasets, and the results demonstrate that our method achieves a favorable trade-off between model complexity and reconstruction performance." + ], + "bbox": [ + 225, + 366, + 784, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 522, + 387, + 537 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "CNN-based super-resolution. SRCNN [9] first introduces effective end-to-end trainable CNNs to map LR directly to its counterpart HR, with superior performance over previous hand-crafted prior and interpolation methods. Then, FSRCNN [10] and ESPCN [35] adopt a post-upsampling strategy to improve efficiency. VDSR [18] uses global residual learning to solve the training problem of deep networks effectively. Therefore, subsequent CNN-based methods can stack deeper and larger layers to capture more information for better image restoration. EDSR [26] increases the model footprint to 43M with significant performance improvement, and RCAN [50], based on channel attention, builds a model with over 400 layers. However, these large models that require high computational costs make them challenging to deploy in real-world applications.", + "bbox": [ + 212, + 551, + 785, + 718 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To alleviate the computational burden, many CNN-based methods aim to strike a balance between reconstruction performance and model complexity. CARN [1] applies group convolution and cascades an efficient residual network. IMDN [17] adopts the information multi-distillation blocks to split, refine, and aggregate the feature mapping step-by-step, which significantly reduces the model parameters. Furthermore, its improvement work RFDN [28] is the winning solution in the NTIRE 2022 Efficient SR Challenge. BSRN [24] uses blueprint convolutions to reduce computational redundancy significantly. Unlike the above", + "bbox": [ + 212, + 719, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "SMFANet for Efficient Image Super-Resolution", + "bbox": [ + 416, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "approaches that employ progressive feature refinement techniques, recent studies [32, 36, 52] tend to introduce non-local feature modulation to extract representative features for reconstruction. VapSR [52] uses large kernel convolutions in the attention branch to enlarge the receptive field. SAFMN [36] enlarges the receptive field by introducing different down-sampling rates to obtain multi-scale spatial features, and then aggregates these features to generate the modulation map. MDRN [32] employs the parallel implementation of multiple stridden convolutions and polling operations to capture finer global structural information. Although these methods can model non-local feature interactions at a low computational cost and achieve good performance, performing reconstruction using aggregated non-local features only and ignoring local features leads to artifacts in local details (e.g., corners and edges) of the super-resolved image. To avoid this problem, we propose an efficient SMFA module that collaboratively models local and non-local features for more accurate reconstruction.", + "bbox": [ + 212, + 146, + 787, + 356 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ViT-based super-resolution. The vision Transformer (ViT) [12] utilizes the self-attention (SA) mechanism to explore global information for image reconstruction and achieves great success. IPT [4] first introduces a standard vision Transformer to solve the image SR problem. However, the SA mechanism of the ViT consumes huge resources, so various attention variants have been proposed to reduce the computational burden. SwinIR [25] employs window self-attention and outperforms the CNN-based large model [26]. ELAN [49] proposes a group-wise self-attention module and shares the weights to reduce complexity significantly. Moreover, recent studies [11,34] have pointed out that ViTs have the nature of the low-pass filter, resulting in smooth reconstruction results. Therefore, many ViT-based methods [6, 20, 23] utilize convolutional operators to enhance local details for higher-quality image reconstruction. Although these methods integrate the advantages of both convolutional and transformer structures and utilize both local and non-local information to achieve high-quality image reconstruction results, they still rely on window-based attention variants, and feature interactions across windows require much time to perform. Different from these methods, we develop an efficient approximation of self-attention (EASA) branch to model non-local information with much lower model complexity.", + "bbox": [ + 212, + 358, + 787, + 632 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Proposed Method", + "text_level": 1, + "bbox": [ + 215, + 662, + 426, + 680 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our goal is to develop a simple yet effective CNN model to collaboratively explore local and non-local feature information for accurate image super-resolution. To this end, we first develop a self-modulation feature aggregation (SMFA) module to effectively exploit representative features, where an efficient approximate self-attention (EASA) module based on feature modulation is used for non-local feature interactions, while an additional local detail estimation (LDE) branch is used for local information exploration. To refine the features generated by the SMFA layer, we further introduce a partial convolution-based feed-forward network (PCFN). We incorporate these components into a unified unit that", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Zheng et al.", + "bbox": [ + 271, + 114, + 352, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f26eeaf4b8abd0718cd1dc1bd92e899e54a84fc2317a29b21f021c82af3c16ed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 222, + 143, + 671, + 203 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "$⑦$ Adaptive Max Pooling", + "$\\bullet$ Element-wise Product", + "Nearest Up-sampling", + "$\\bigoplus$ Element-wise Addition", + "$\\widehat{\\mathbf{S}}$ Channel Split", + "Variance Calculation", + "GELU Activation", + "C Channel Concatenate" + ], + "bbox": [ + 681, + 148, + 774, + 301 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f5e576f6a8729a2c802f75042e163aaaf05dd20fa8fc1bf78799db84d7b52083.jpg", + "image_caption": [ + "Fig.2: Network architecture of the proposed SMFANet. The proposed SMFANet consists of a shallow feature extraction module, feature modulation blocks (FMBs), and a lightweight image reconstruction module. Feature modulation block contains one self-modulation feature aggregation (SMFA) module and one partial convolution-based feed-forward network (PCFN)." + ], + "image_footnote": [], + "bbox": [ + 218, + 204, + 604, + 309 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8bff7d6ed5fed818f7433ebadce750206c0ab98974bd2da416129fc4e952de27.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 607, + 205, + 666, + 308 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "pays more attention to information-rich regions for better reconstruction. In the following, we explain the main ideas of each component in detail.", + "bbox": [ + 212, + 426, + 784, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Overall architecture", + "text_level": 1, + "bbox": [ + 214, + 479, + 426, + 492 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 2 illustrates the overall architecture of our proposed SMFANet. It takes a low-resolution image as input, and uses a $3 \\times 3$ convolutional layer to extract shallow features. The extracted shallow features are then fed to a sequence of feature modulation blocks (FMBs) to produce deep representative features, where the FMB consists of a self-modulation feature aggregation (SMFA) module and a partial convolution-based feed-forward network (PCFN). After the feature modulation block, the representative features are processed by an image reconstruction module to reconstruct the high-quality output. To make the reconstruction module as lightweight as possible, a $3 \\times 3$ convolutional layer is used to convert the channel dimensions to a specific size that adjusts to the upsampling ratio, and a PixelShuffle layer [35] is used for enlargement. To facilitate high-frequency information learning, we insert a global residual connection before the image reconstruction module.", + "bbox": [ + 212, + 503, + 787, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Self-modulation feature aggregation module", + "text_level": 1, + "bbox": [ + 214, + 723, + 624, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Exploring non-local information is important for image SR reconstruction. Recent ViT-based SR methods [4, 7, 25, 47, 49] utilize various self-attention mechanisms to explore non-local information and achieve impressive reconstruction performance. However, these self-attention variants are computationally expensive and have a limited ability to model local details, as their low-pass filter nature makes them preferentially capture low-frequency information [11, 20, 34].", + "bbox": [ + 212, + 750, + 787, + 842 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "SMFANet for Efficient Image Super-Resolution", + "bbox": [ + 416, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address this problem, we develop a lightweight self-modulation feature aggregation (SMFA) module to collaboratively model local and non-local features for accurate reconstruction. Within the SMFA module, an efficient approximation of self-attention (EASA) branch is implemented for non-local information exploration at a moderate cost, and a local detail estimation (LDE) branch is used for capturing local information.", + "bbox": [ + 212, + 146, + 782, + 234 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given the input feature $F_{in} \\in \\mathbb{R}^{H \\times W \\times C}$ , where $H \\times W$ denotes the spatial size and $C$ is the number of channels, we first apply a $1 \\times 1$ convolution to the normalized $F_{in}$ to expand the channel, and then split the channel into two parts as inputs to the EASA and LDE branches:", + "bbox": [ + 212, + 236, + 784, + 297 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\{X, Y \\} = \\mathcal {S} \\left(\\operatorname {C o n v} _ {1 \\times 1} \\left(\\left| \\left| F _ {i n} \\right| \\right| _ {2}\\right)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 390, + 305, + 784, + 321 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $||\\cdot||_2$ is the $L_{2}$ normalization, $Conv_{1\\times 1}(\\cdot)$ denotes a $1\\times 1$ convolutional layer, $S(\\cdot)$ denotes a channel splitting operation, $X\\in \\mathbb{R}^{H\\times W\\times C}$ and $Y\\in \\mathbb{R}^{H\\times W\\times C}$ . We then process the features $X$ and $Y$ in parallel via the EASA and LED branches, producing the non-local feature $X_{l}$ and local feature $Y_{d}$ , respectively. Finally, we fuse $X_{l}$ and $Y_{d}$ together with element-wise addition and feed them into a $1\\times 1$ convolution to form a representative output of the SMFA module. This process can be formulated as:", + "bbox": [ + 214, + 327, + 782, + 431 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nF _ {\\rho} = \\operatorname {C o n v} _ {1 \\times 1} \\left(X _ {l} + Y _ {d}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 415, + 441, + 784, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $F_{\\rho}\\in \\mathbb{R}^{H\\times W\\times C}$ is the output feature.", + "bbox": [ + 214, + 460, + 529, + 477 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Efficient approximation of self-attention. We obtain the low-frequency components through a downsampling operation and feed them into a $3 \\times 3$ depth-wise convolution to generate non-local structure information $X_{s} \\in \\mathbb{R}^{H / 8 \\times W / 8 \\times C}$ :", + "bbox": [ + 214, + 478, + 782, + 523 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nX _ {s} = D W C o n v _ {3 \\times 3} (\\mathcal {D} (X)), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 405, + 539, + 784, + 553 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{D}(\\cdot)$ denotes the adaptive max pooling with a scaling factor of 8, $DWConv_{3\\times 3}(\\cdot)$ is a $3\\times 3$ depth-wise convolutional layer. To embed global descriptions for modulating non-local representation $X_{s}$ , we introduce the variance of $X$ as the statistical divergence of spatial information, and merge it with the non-local representation $X_{s}$ through a $1\\times 1$ convolution:", + "bbox": [ + 214, + 556, + 782, + 632 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma^ {2} (X) = \\frac {1}{N} \\sum_ {i = 0} ^ {N - 1} \\left(x _ {i} - \\mu\\right) ^ {2}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 641, + 784, + 680 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nX _ {m} = \\operatorname {C o n v} _ {1 \\times 1} \\left(X _ {s} + \\sigma^ {2} (X)\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 684, + 607, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\sigma^2 (X)\\in \\mathbb{R}^{1\\times 1\\times C}$ is the variance of $X$ , $N$ is the total number of pixels, $x_{i}$ denotes the value of each pixel, $\\mu$ is the mean of all pixel values, and $X_{m}\\in \\mathbb{R}^{H\\times W\\times C}$ represents the modulated feature. This variance modulation mechanism facilitates better exploring non-local information, and we study it in detail in Section 5.", + "bbox": [ + 214, + 705, + 782, + 781 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, we use the modulated features to aggregate the input feature $X$ for extracting the representative structure information $X_{l}$ :", + "bbox": [ + 214, + 782, + 782, + 813 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nX _ {l} = X \\odot \\mathcal {U} (\\phi (X _ {m})), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 823, + 782, + 838 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Zheng et al.", + "bbox": [ + 271, + 114, + 352, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\phi(\\cdot)$ refers to the GELU activation function [15], $\\mathcal{U}(\\cdot)$ denotes a nearest upsampling operation, and $\\odot$ represents the element-wise product operation.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Local detail estimation. Local details are important for the pleasing high-frequency reconstruction. As the EASA prioritizes non-local structure information exploration, we develop a simple local detail estimation layer to capture local features simultaneously. In detail, an expanded depth-wise convolution with a kernel size of $3 \\times 3$ is used to encode local information $Y_{h}$ from the input features $Y$ . Then, we use two $1 \\times 1$ convolutions with a hidden GELU activation to generate the enhanced local feature $Y_{d}$ , which is achieved by:", + "bbox": [ + 212, + 176, + 784, + 282 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} Y _ {h} = \\operatorname {C o n v} _ {1 \\times 1} (D W \\operatorname {C o n v} _ {3 \\times 3} (Y)), \\\\ Y _ {v} = \\operatorname {C o n v} _ {1} (\\phi (Y)) \\end{array} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 382, + 290, + 784, + 316 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $Y_{h}\\in \\mathbb{R}^{H\\times W\\times 2C}$ is the encoded local information.", + "bbox": [ + 214, + 330, + 617, + 348 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Partial convolution-based feed-forward network", + "text_level": 1, + "bbox": [ + 214, + 367, + 651, + 381 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The regular feed-forward network (FFN) [12, 40] operates on each pixel location identically, which lacks the exchange of information in the spatial dimension. Inspired by [5, 37], we improve the FFN and propose an efficient partial convolution-based feed-forward network (PCFN) to further refine the representative features derived from the SMFA.", + "bbox": [ + 212, + 388, + 782, + 463 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 2(c) shows that the PCFN uses a $1 \\times 1$ convolution with the GELU activation function to perform cross-channel interaction on the expanded hidden space. Then, it splits the hidden features into two chunks of $\\{F_{\\rho}^{1}, F_{\\rho}^{2}\\}$ and employs a $3 \\times 3$ convolution followed by a GELU activation to process $F_{\\rho}^{1}$ to encode local contextual information. The processed $F_{\\rho}^{1}$ and $F_{\\rho}^{2}$ are then concatenated and fed into a $1 \\times 1$ convolution for further feature mixing and reducing the hidden channel back to the original input dimension. This process is defined as:", + "bbox": [ + 212, + 465, + 784, + 570 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{F _ {\\rho} ^ {1}, F _ {\\rho} ^ {2} \\right\\} = \\mathcal {S} \\left(\\phi \\left(\\operatorname {C o n v} _ {1 \\times 1} \\left(\\left| \\left| F _ {\\rho} \\right| \\right| _ {2}\\right)\\right)\\right), \\\\ \\hat {\\hat {r}} - C = \\left(\\mathcal {A} \\left[ + (C - \\left(F _ {1}\\right)) - F _ {2} ^ {2} \\right]\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 590, + 784, + 617 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {F} _ {\\rho} = C o n v _ {1 \\times 1} (\\mathcal {C} ([ \\phi (C o n v _ {3 \\times 3} (F _ {\\rho} ^ {1})), F _ {\\rho} ^ {2} ])),\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 611, + 648, + 630 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $F_{\\rho}^{1}\\in \\mathbb{R}^{H\\times W\\times C / 2}$ and $F_{\\rho}^{2}\\in \\mathbb{R}^{H\\times W\\times 3C / 2}$ , $\\mathcal{C}(\\cdot)$ represents a concatenation operation, and $\\operatorname{Conv}_{3\\times 3}(\\cdot)$ denotes a $3\\times 3$ convolutional layer.", + "bbox": [ + 212, + 636, + 782, + 669 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Feature modulation block", + "text_level": 1, + "bbox": [ + 214, + 686, + 472, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As mentioned previously, we formulate the proposed SMFA and PCFN into a feature modulation block (FMB) to produce deep representative features. To stabilize the model training and encourage more information flow, the residual connection is employed. The FMB can be written as:", + "bbox": [ + 212, + 709, + 782, + 768 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} F _ {\\rho} = S M F A \\left(F _ {i n}\\right) + F _ {i n}, \\\\ \\hat {\\hat {F}} = \\operatorname {R G E N} (F _ {i}) + F _ {i}. \\end{array} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 411, + 777, + 784, + 803 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {F} _ {\\rho} = P C F N (F _ {\\rho}) + F _ {\\rho},\n$$\n", + "text_format": "latex", + "bbox": [ + 411, + 797, + 578, + 814 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $F_{in}$ , $F_{\\rho}$ and $\\hat{F}_{\\rho}$ remain consistent with the previous definition.", + "bbox": [ + 214, + 823, + 709, + 842 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "SMFANet for Efficient Image Super-Resolution", + "bbox": [ + 416, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Experimental Results", + "text_level": 1, + "bbox": [ + 215, + 143, + 460, + 162 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Dataset and implementation", + "text_level": 1, + "bbox": [ + 215, + 180, + 496, + 196 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Datasets. The high-quality DIV2K [38] and DIV2K + Flickr2K (DF2K) [38] datasets are commonly used for training SR models, and previous methods [14,17,25,26,36] chose one of them as the training dataset. To make a fair comparison with recent state-of-the-art methods, we train our proposed method on the DIV2K and DF2K datasets, respectively. In addition, we evaluate our method on the commonly used test datasets, including Set5 [3], Set14 [46], B100 [2], Urban100 [16] and Manga109 [33]. We transform the image into YCbCr color space and calculate the peak signal-to-noise ratio (PSNR) and structure similarity index (SSIM) on the Y channel of the image to evaluate the quality of the recovered image.", + "bbox": [ + 212, + 208, + 787, + 359 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Implementation details. During the training, we randomly crop 64 patches of size $64 \\times 64$ with random horizontal flips and rotations from LR images as the basic training inputs. The proposed model is trained with the same loss function as SAFMN [36] and optimized by Adam [19] with $\\beta_{1} = 0.9$ and $\\beta_{2} = 0.99$ . We set the initial learning rate to $1 \\times 10^{-3}$ and the minimum one to $1 \\times 10^{-5}$ , which is updated by the Cosine Annealing scheme [30]. The number of iterations for all experiments is set to 1,000,000. All experiments are conducted with the PyTorch framework on an NVIDIA GeForce RTX 3090 GPU. We train two versions of SMFANet with different complexity. The standard SMFANet consists of 8 FMBs with 36 channels while the large version, SMFANet+, has 12 FMBs with 48 channels. Due to the page limit, we include more experimental results in the supplemental material.", + "bbox": [ + 212, + 361, + 787, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 Comparisons with state-of-the-art methods", + "text_level": 1, + "bbox": [ + 215, + 568, + 617, + 583 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Quantitative comparison. To fully evaluate the performance of our method, we first compare our SMFANet with state-of-the-art CNN-based lightweight SR methods, including FSRCNN [10], CARN [1], EDSR-baseline [26], IMDN [17], LAPAR-A [22], SMSR [42], ShuffleMixer [37], SAFMN [36], and further compare the SMFANet+ with recent ViT-based lightweight algorithms, including ESRT [31], SwinIR-light [25], ELAN-light [49], HPINet-S [27], NGswin [7], SPI-N [47] and SRFormer-light [53]. The quantitative comparisons on benchmark datasets for the upscaling factors of $\\times 2$ , $\\times 3$ , and $\\times 4$ are reported in Table 1 and Table 2, where SMFANet (DIV2K) denotes training using the DIV2K dataset, and SMFANet (DF2K) uses DF2K as the training dataset. In addition to PSNR and SSIM metrics, we also list the number of parameters (#Params) and FLOPs (#FLOPs). For a fair comparison, we calculate the model complexity of all evaluated methods with the fvcore library (i.e., fvcore.nn.flop_count_str) under a setting of super resolving an LR image to $1280 \\times 720$ pixels.", + "bbox": [ + 212, + 597, + 787, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Benefiting from the ability to model non-local information of the SMFA module, SMFANet can effectively explore more information compared to previous", + "bbox": [ + 215, + 809, + 785, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Zheng et al.", + "bbox": [ + 271, + 114, + 352, + 128 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b2a3b82cfe2cb70915fca86fff22396e3aa6c95f211a4b940fc42e2fa5cd87e1.jpg", + "table_caption": [ + "Table 1: Comparison with CNN-based lightweight SR methods on public benchmark datasets. All PSNR/SSIM results are calculated on the Y-channel. #FLOPs is measured corresponding to an HR image of the size $1280 \\times 720$ pixels. The best and second-best performances are highlighted in Red and Blue colors." + ], + "table_footnote": [], + "table_body": "
ScaleMethods#Params (K)#FLOPs (G)Set5Set14B100Urban100Manga109
×2FSRCNN [10]12637.00/0.955832.63/0.908831.53/0.892029.88/0.902036.67/0.9694
CARN [1]159222337.76/0.959033.52/0.916632.09/0.897831.92/0.925638.36/0.9765
EDSR-baseline [26]137031637.99/0.960433.57/0.917532.16/0.899431.98/0.927238.54/0.9769
IMDN [17]69416138.00/0.960533.63/0.917732.19/0.899632.17/0.928338.88/0.9774
LAPAR-A [22]54817138.01/0.960533.62/0.918332.19/0.899932.10/0.928338.67/0.9772
SMSR [42]98513238.00/0.960133.64/0.917932.17/0.899932.19/0.928438.76/0.9771
ShuffleMixer [37]3949138.01/0.960633.63/0.918032.17/0.899531.89/0.925738.83/0.9774
SAFMN [36]2285238.00/0.960533.54/0.917732.16/0.899531.84/0.925638.71/0.9771
SMFANet (DIV2K)1864138.04/0.960633.65/0.918632.19/0.899932.22/0.928338.98/0.9776
SMFANet (DF2K)1864138.08/0.960733.65/0.918532.22/0.900232.20/0.928239.11/0.9779
×3FSRCNN [10]12533.16/0.914029.43/0.824228.53/0.791026.43/0.808030.98/0.9212
CARN [1]159211934.29/0.925530.29/0.840729.06/0.803428.06/0.849333.50/0.9440
EDSR-baseline [26]155516034.37/0.927030.28/0.841729.09/0.805228.15/0.852733.45/0.9439
IMDN [17]7037234.36/0.927030.32/0.841729.09/0.804628.17/0.851933.61/0.9445
LAPAR-A [22]59411434.36/0.926730.34/0.841229.11/0.805428.15/0.852333.51/0.9441
SMSR [42]9936834.40/0.927030.33/0.841229.10/0.805028.25/0.853633.68/0.9445
ShuffleMixer [37]4154334.40/0.927230.37/0.842329.12/0.805128.08/0.849833.69/0.9448
SAFMN [36]2332334.34/0.926730.33/0.841829.08/0.804827.95/0.847433.52/0.9437
SMFANet (DIV2K)1911934.46/0.927530.39/0.843229.13/0.805928.25/0.852533.84/0.9454
SMFANet (DF2K)1911934.42/0.927430.41/0.843029.16/0.806528.22/0.852333.96/0.9460
×4FSRCNN [10]12530.71/0.865727.59/0.753526.98/0.715024.62/0.728027.90/0.8517
CARN [1]15929132.13/0.893728.60/0.780627.58/0.734926.07/0.783730.47/0.9084
EDSR-baseline [26]151811432.09/0.893828.58/0.781327.57/0.735726.04/0.784930.35/0.9067
IMDN [17]7154132.21/0.894828.58/0.781127.56/0.735326.04/0.783830.45/0.9075
LAPAR-A [22]6599432.15/0.894428.61/0.781827.61/0.736626.14/0.787130.42/0.9074
SMSR [42]10064232.12/0.893228.55/0.780827.55/0.735126.11/0.786830.54/0.9085
ShuffleMixer [37]4112832.21/0.895328.66/0.782727.61/0.736626.08/0.783530.65/0.9093
SAFMN [36]2401432.18/0.894828.60/0.781327.58/0.735925.97/0.780930.43/0.9063
SMFANet (DIV2K)1971132.25/0.895628.67/0.782527.61/0.737126.19/0.786130.72/0.9097
SMFANet (DF2K)1971132.25/0.895628.71/0.783327.64/0.737726.18/0.786230.82/0.9104
", + "bbox": [ + 225, + 212, + 782, + 502 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "CNN-based methods. Table 1 shows that our SMFANet achieves better performance on almost all benchmark datasets. For instance, on the $\\times 4$ Manga109 dataset, SMFANet (DIV2K) outperforms IMDN [17] by $0.27\\mathrm{dB}$ with only $27\\%$ FLOPs. When training SMFANet on the larger DF2K dataset, the performance gain is further improved by $0.1\\mathrm{dB}$ .", + "bbox": [ + 212, + 534, + 782, + 609 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 2 shows that the proposed SMFANet+ achieves competitive performance with lower model complexity. Note that ViT-based methods are based on the self-attention mechanism to model long-range dependency, which consumes significant computational costs. In contrast, our SMFA module, with only a few low-consumption operators, efficiently explores non-local information for high-quality reconstruction. On the five benchmark datasets, SMFANet $^+$ (DIV2K) shows similar performance to SwinIR-light [25] with only $43\\%$ FLOPs, and SMFANet $^+$ (DF2K) improves the average PSNR gain of the $\\times 4$ SR to 0.14dB. These comparisons suggest that our proposed SMFANet can exploit more information than previous CNN-based and ViT-based approaches.", + "bbox": [ + 212, + 611, + 784, + 762 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Qualitative comparisons. We compare the visual results of our proposed SMFANet and SMFANet+ with CNN-based methods and ViT-based methods on the $\\times 4$ Urban100 dataset, respectively. The evaluated CNN-based methods [17,22,26,36,37] are listed at the top of Figure 3, which suffer from blurring artifacts and distorted lines. In contrast, our method can restore more accurate", + "bbox": [ + 212, + 763, + 785, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "SMFANet for Efficient Image Super-Resolution", + "bbox": [ + 416, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/f030380f25b420aefb339c43f6d901135ab2ef8b520ec103c8792f90c4a18da5.jpg", + "table_caption": [ + "Table 2: Comparison with ViT-based lightweight SR methods on public benchmark datasets. All measured metrics are calculated in the same way as in Table 1. The best and second-best performances are highlighted in Red and Blue colors." + ], + "table_footnote": [], + "table_body": "
ScaleMethods#Params (K)#FLOPs (G)Set5Set14B100Urban100Manga109
×2ESRT [31]678111638.03/0.960033.75/0.918432.25/0.900132.58/0.931839.12/0.9774
SwinIR-light [25]91024438.14/0.961133.86/0.920632.31/0.901232.76/0.934039.12/0.9783
ELAN-light [49]62120338.17/0.961133.94/0.920732.30/0.901232.76/0.934039.11/0.9782
NGswin [7]99814638.05/0.961033.79/0.919932.27/0.900832.53/0.932438.97/0.9777
SPIN [47]49711438.20/0.961533.90/0.921532.31/0.901532.79/0.934039.18/0.9784
SMFANet+ (DIV2K)48010838.18/0.961133.82/0.920232.28/0.901132.64/0.932339.25/0.9777
SMFANet+ (DF2K)48010838.19/0.961133.92/0.920732.32/0.901532.70/0.933139.46/0.9787
×3ESRT [31]77083534.42/0.926830.43/0.843329.15/0.806328.46/0.857433.95/0.9455
SwinIR-light [25]91811434.62/0.928930.54/0.846329.20/0.808228.66/0.862433.98/0.9478
ELAN-light [49]6299034.61/0.928830.55/0.846329.21/0.808128.69/0.862434.00/0.9478
NGswin [7]10076634.52/0.928230.53/0.845629.19/0.807828.52/0.860333.89/0.9470
SPIN [47]5695834.65/0.929330.57/0.846429.23/0.808928.71/0.862734.24/0.9489
SMFANet+ (DIV2K)4874834.63/0.928530.52/0.845629.23/0.808428.59/0.859434.17/0.9478
SMFANet+ (DF2K)4874834.66/0.929230.57/0.846129.25/0.809028.67/0.861134.45/0.9490
×4ESRT [31]75229832.19/0.894728.69/0.783327.69/0.737926.39/0.796230.75/0.9100
SwinIR-light [25]9306532.44/0.897628.77/0.785827.69/0.740626.47/0.798030.92/0.9151
ELAN-light [49]6405432.43/0.897528.78/0.785827.69/0.740626.54/0.798230.92/0.9150
HPINet-S [27]4638832.47/0.898728.80/0.787227.69/0.741626.59/0.801630.92/0.9143
NGswin [7]10194032.33/0.896328.78/0.785927.66/0.739626.45/0.796330.80/0.9128
SPIN [47]5554232.48/0.898328.80/0.786227.70/0.741526.55/0.799830.98/0.9156
SRFormer-light [53]8736332.51/0.898828.82/0.787227.73/0.742226.67/0.803231.17/0.9165
SMFANet+ (DIV2K)4962832.43/0.897928.77/0.784927.70/0.740026.45/0.794331.06/0.9138
SMFANet+ (DF2K)4962832.51/0.898528.87/0.787227.74/0.741226.56/0.797631.29/0.9163
", + "bbox": [ + 220, + 198, + 774, + 419 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/36f84ab653d4e200fbc05b9a8c3bd5e71b1f19bf6154f2f64ac830bbc1d0538d.jpg", + "table_caption": [ + "Table 3: Memory and running time comparisons on $\\times 4$ SR. #GPU Mem. denotes the maximum GPU memory consumption during the inference phase, derived with the Pytorch torch.cuda.max_memory_allocated() function. #Avg.Time is the average running time on 500 LR images of $320 \\times 180$ pixels. The best and second-best performances are highlighted in Red and Blue colors." + ], + "table_footnote": [], + "table_body": "
TypeMethodsQuality MetricsTypeMethodsQuality Metrics
#GPU Mem. [M]#Avg. Time [ms]#GPU Mem. [M]#Avg. Time [ms]
CNNCARN [1]702.0713.54ViTSwinIR-light [25]342.44177.06
EDSR-baseline [26]507.1315.21ELAN-light [49]241.3442.75
IMDN [17]204.279.64HPINet-S [27]445.92140.45
LAPAR-A [22]1811.4624.40NGswin [7]372.85108.25
ShuffleMixer [37]474.7919.21SPIN [47]441.52798.44
SAFMN [36]65.268.44SRFormer-light [53]320.95145.99
SMFANet93.209.26SMFANet+247.5117.39
", + "bbox": [ + 225, + 513, + 777, + 614 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "parallel patterns with sharper boundaries. The bottom of Figure 3 shows that ViT-based methods [7,25,27,31,49] do not restore the structures well, e.g., strips, while our SMFANet+ recovers an image with better structures.", + "bbox": [ + 215, + 643, + 785, + 686 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Memory and running time comparisons. To fully demonstrate the efficiency of our proposed SMFANet, we further compare our method with the CNN-based [1, 17, 22, 26, 36, 37] and ViT-based [7, 25, 27, 47, 49, 53] methods in terms of GPU memory consumption and inference time on $\\times 4$ SR. Table 3 shows the comparisons of GPU memory and inference time. Our SMFANet has faster inference than the listed CNN methods except SAFMN, and the larger SMFANet+ presents a significant efficiency advantage over ViT-based methods. The running time of SMFANet+ is $17.39\\mathrm{ms}$ , which is $\\times 10$ times faster than SwinIR-light [25] and $\\times 6$ times faster than NGswin [7]. Moreover, when compared to ELAN-light [49], which uses a shared group-wise multi-scale self", + "bbox": [ + 215, + 688, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Zheng et al.", + "bbox": [ + 271, + 114, + 352, + 128 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/da02855a3f993c548d98bbc00b8f54bb6cb1e80e646c27f5a2e9610bdeccbe51.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 223, + 145, + 395, + 246 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/3c49c7915166b4ec255f9c3a4c077cb6354f02faa702a56815e458f21e1bf5a8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 397, + 143, + 493, + 246 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/bf69c49b83ff8702696c8ffd1088a1aa2d529d5387357074eab9afeccd05d290.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 143, + 596, + 246 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/950d668936f5318b57c5a7b816a6bd5d56ec76dc22806b58b0e09bf813b71291.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 602, + 143, + 700, + 246 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/04dc92e0af253e375ccced95f91bf572b0a5db0fa871ead81ec94c22cd50aaf6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 705, + 143, + 803, + 246 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/6ef858126139393627546d8b35daca3cc8bb61cccaebb2daac5f899485beddfb.jpg", + "image_caption": [ + "image078 from Urban100 [16]", + "Fig. 3: Visual comparisons for $\\times 4$ SR on the Urban100 dataset. The comparisons between SMFANet and CNN-based methods are shown at the top of the Figure, while the comparisons between SMFANet+ and ViT-based methods are exhibited at the bottom. Our proposed method recovers the image with more accurate structures." + ], + "image_footnote": [], + "bbox": [ + 223, + 246, + 395, + 339 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/94e40c544286f9883005bb3e9d5b47640eee5ddd4bdf1503f32e3b18624e394b.jpg", + "image_caption": [ + "(e) ELAN-light [49]" + ], + "image_footnote": [], + "bbox": [ + 397, + 246, + 493, + 339 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/388a397b683ee5ea39a9ad7844877948b238e6595dae3baa3b78eb7bedbe8600.jpg", + "image_caption": [ + "(f) NGswin [7]" + ], + "image_footnote": [], + "bbox": [ + 500, + 246, + 596, + 339 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/d4ec60e93bd9def1e099d492a5a45ada74fde113a9dbac37526c4ee2dcb85c45.jpg", + "image_caption": [ + "(g) HPINet-S [27]" + ], + "image_footnote": [], + "bbox": [ + 602, + 246, + 700, + 339 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/45cf5cc0d2c73b2a58c747f15cfb22af6f4c100ea6034f686656c25270507f78.jpg", + "image_caption": [ + "(h) SMFANet+" + ], + "image_footnote": [], + "bbox": [ + 705, + 246, + 803, + 339 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "attention to accelerate model efficiency, the SMFANet+ still has a much faster runtime (17.39ms vs. 42.75ms). These quantitative and qualitative comparison results show that our method obtains even better results against advanced SR models but with significantly less complexity.", + "bbox": [ + 212, + 458, + 784, + 518 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "LAM comparisons. The local attribution map (LAM) [13] indicates the significant correlation between the red pixels and the rectangular position patch during the restoration process. We compare the LAM [13] results between our method and the evaluated CNN-based [17, 36, 37] and ViT-based [25, 27, 49] methods in Figure 4, and label the corresponding diffusion index (DI) value below each subgraph, where a larger DI value indicates a wider range of pixels involved. These results demonstrate that the proposed SMFANet can explore more non-local information for accurate image super-resolution.", + "bbox": [ + 212, + 518, + 784, + 640 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5 Analysis and Discussions", + "text_level": 1, + "bbox": [ + 215, + 667, + 493, + 684 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In this section, we conduct extensive ablation studies to analyze and evaluate the effect of each component in the proposed SMFANet. We implement all ablation experiments based on the $\\times 4$ SMFANet model and train them with the DIV2K [38] dataset for fair comparisons. The quantitative ablation results in Table 4 are measured on the Urban100 [16] and Manga109 [33] datasets.", + "bbox": [ + 212, + 703, + 784, + 777 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Effectiveness of the SMFA. The proposed SMFA module utilizes a parallel structure that absorbs local and non-local feature extraction to enhance reconstruction accuracy. To demonstrate its effectiveness, we first disable the SMFA module to compare it with the baseline SMFANet. Table 4 shows that the PSNR", + "bbox": [ + 212, + 779, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "SMFANet for Efficient Image Super-Resolution", + "bbox": [ + 416, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/f5a746bf1626f46cbf91b017d7a43fb313bb2ec6f7493ecb62997fcd28db0854.jpg", + "image_caption": [ + "image010 from Urban100" + ], + "image_footnote": [], + "bbox": [ + 223, + 145, + 395, + 239 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/0f5b1dd82bb967f95f6bd51237ceb4435dc2d26ca8464f81762de9f2d30130b7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 403, + 146, + 495, + 191 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/dcacc3c8257f03960ed1e90eaf8ab19c2000513d218b8dbb952a594ce00aa0ee.jpg", + "image_caption": [ + "(a) ShuffleMixer [37] \nDI: 3.5808" + ], + "image_footnote": [], + "bbox": [ + 403, + 191, + 495, + 238 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/dda0f3e965ed30a9e3c9660d63f1fa81d93e1e61560a77df6895dbebab4f274a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 145, + 594, + 191 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/67808f8cfb400f680e076f1edaa238597826d14799f37b9992bcd1db5bdc98f7.jpg", + "image_caption": [ + "(b) SAFMN [36] \nDI: 8.6313" + ], + "image_footnote": [], + "bbox": [ + 504, + 191, + 594, + 238 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/b566e27607b5623d28d3aec9523b1002af3d74e0970ec0f79ab1f4138d3f6cc0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 599, + 145, + 689, + 191 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/94f9024109556785e7875fa8367c7cc1ef9ed81758c2869b0df50eb835e00b6b.jpg", + "image_caption": [ + "(c) IMDN [17] DI: 10.3650" + ], + "image_footnote": [], + "bbox": [ + 598, + 191, + 687, + 238 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/0a9a975d4a3ce5713532b8a1e825d89d9d8db44e96ac6953d13bb551b66a25ba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 696, + 145, + 785, + 191 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/6766065451f11839c234fc266db8f164cc9809bb85fcd4407f4a12476b219984.jpg", + "image_caption": [ + "(d) SMFANet DI 12.0899" + ], + "image_footnote": [], + "bbox": [ + 696, + 191, + 785, + 238 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/515868b5a042d5407350a0ec7fa4874227234c83e8a093e67cd87a02bcfebc94.jpg", + "image_caption": [ + "img005 from Urban100", + "Fig. 4: Comparison of local attribution maps (LAMs) [13] and diffusion indices (DIs) [13]. Our proposed SMFANet family exploit more feature information and reconstructs a more accurate image structure." + ], + "image_footnote": [], + "bbox": [ + 222, + 261, + 395, + 354 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/2c4bc7c6ac7ced408b455bb46a783b4b7a2643cebce02cbca429e7ea55326af7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 401, + 261, + 495, + 306 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/0675c69552ca2fb49eafa7c97475c1da7de6d907b7739674d7882494344760bf.jpg", + "image_caption": [ + "(a) SwinIR-light [25] \nDI: 6.74785" + ], + "image_footnote": [], + "bbox": [ + 401, + 306, + 495, + 354 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/dd5bd54d14477446067fddcb9616c38c604910aadb487a7e51cb7ffa2ac95f70.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 261, + 594, + 306 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/df0cc27ce4c00485fba494013181bb95ab3b1a6b1df1c7dffc2aa716d2180e0b.jpg", + "image_caption": [ + "b) HPINet-S [27] \nDI: 12.8471" + ], + "image_footnote": [], + "bbox": [ + 504, + 306, + 594, + 354 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/55fb970d289cd7538733bc3b9b64ae24725808ecdec4829532b50186ad272984.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 598, + 261, + 689, + 306 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/e73e425d874198f09c8812aac9044827c29044036b57d102f67cd7cc1fabffae.jpg", + "image_caption": [ + "(c) ELAN-light [49] \nDI: 18.2614" + ], + "image_footnote": [], + "bbox": [ + 598, + 306, + 689, + 354 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/58f50301a00a053071d401e70befeb31e290eb86fe6990ecfb8d2c104fc2a8ba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 696, + 261, + 785, + 306 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/b03a000236565a3a6265e625991add603e6de9b2bde01ea3f747b02944f1ac4d.jpg", + "image_caption": [ + "(d) SMFANet+ \nDI: 19.6512" + ], + "image_footnote": [], + "bbox": [ + 696, + 306, + 785, + 354 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "values drop by 0.65dB and 0.96dB on the Urban100 and Manga109 datasets. These results show the importance of the SMFA.", + "bbox": [ + 212, + 472, + 782, + 502 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In addition, as the proposed SMFA module mainly contains an EASA branch to explore non-local information and a LDE branch to capture local features, we conduct ablation studies with respect to these components to demonstrate their effectiveness on image SR. Without the EASA branch, significant performance drops of $0.25\\mathrm{dB}$ and $0.33\\mathrm{dB}$ are observed on the Urban100 and Manga109 datasets. Figure 5 (c) shows the corresponding visual results with blurred structure. Moreover, replacing EASA with the window-based SA [25] yields almost the same performance, but the corresponding runtime and GPU consumption are three times higher. These results indicate that the EASA branch can efficiently approximate SA to explore non-local information for accurate structural reconstruction. As for the LDE branch, removing it, the model only achieves the performance of $25.92\\mathrm{dB}$ and $30.22\\mathrm{dB}$ on the Urban100 and Manga109. Figure 5 (d) shows the reconstructed images with distorted lines, suggesting that the LDE branch contributes to recovering image details. Figure 6 shows the power spectral density maps, which intuitively illustrate the complementary roles of the EASA and LDE. The PSD of $X_{l}$ derived from the EASA shows that the energy density is more concentrated in the center region, whereas the PSD of $Y_{d}$ obtained from the LDE scatters in the surrounding area compared to $F_{in}$ and $X_{l}$ .", + "bbox": [ + 212, + 505, + 785, + 777 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To further validate the effectiveness of this complementary two-branch design, we replace SMFA with either dual LDE branches or dual EASA branches. With these two alternatives, the PSNR values on the Urban100 dataset drop significantly by 0.17dB and 0.25dB, respectively. We also investigated experiments", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Zheng et al.", + "bbox": [ + 271, + 114, + 352, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/7e30f05a56e3bead604cbf45935f3892a3a768bf12a1b4890a3f8fd230510e35.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 222, + 143, + 331, + 207 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/f9e7b5c314f67120b84faed2e38318eb817b67af3787bdba545da834af8db9b5.jpg", + "image_caption": [ + "(a) HR patch" + ], + "image_footnote": [], + "bbox": [ + 228, + 210, + 330, + 270 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/e612a2cd12dafce68be24143128e012f06053988cf6fa273270085c0d2219336.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 334, + 143, + 442, + 207 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/c0b03589eeb15392652fe753f81fc093dd1e1358b3c251cd82097ce3161a95e5.jpg", + "image_caption": [ + "(b) LR patch" + ], + "image_footnote": [], + "bbox": [ + 334, + 209, + 444, + 271 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/512fe8f9ca834e2e47dbbb432d7c5f8910114adb985de56bcbce925dcb50b06e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 446, + 143, + 555, + 207 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/dc9f0eb35e92e0f035274a9bc793e3a5f76ec97d59250ddfd49b2e2865db03a6.jpg", + "image_caption": [ + "(c) w/o EASA" + ], + "image_footnote": [], + "bbox": [ + 450, + 209, + 553, + 270 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/f690950930d464221fd546f824e0e2f86753942330f04803ca42daccdf3ab627.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 558, + 143, + 666, + 207 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/0340d8020ebf862ab88732c223b24031583c2dec748c19fccf0b2a17a5d37133.jpg", + "image_caption": [ + "(d) w/o LDE" + ], + "image_footnote": [], + "bbox": [ + 563, + 209, + 663, + 270 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/be622c2a3e2bad6b10ec69f7352f7c1616d731b4d2f0e20e26c20ba0796159d1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 143, + 779, + 207 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/da9bdd48e10af41c326fad8f77a23fdbed34e4a3f23191cb1d5523773d052251.jpg", + "image_caption": [ + "(e) SMFANet" + ], + "image_footnote": [], + "bbox": [ + 676, + 209, + 779, + 270 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/5f56e33f72f35ba4128d3b68639f94d6a1b969598451439768bc7a01f0c4d782.jpg", + "image_caption": [ + "Fig. 5: Effectiveness of the SMFA for image super-resolution on the $\\times 4$ Urban100 and Manga109 datasets." + ], + "image_footnote": [], + "bbox": [ + 233, + 356, + 364, + 419 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/cab105c5193fe800703ab74b72f66e74252dd0051b6128f248fd08582e83514d.jpg", + "image_caption": [ + "HR image" + ], + "image_footnote": [], + "bbox": [ + 233, + 421, + 366, + 484 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/b6cda250e48d7fe1d828e067f4aabbadd63de947fe02b64a68d55dffe78001ac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 356, + 500, + 419 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/5c1e44240c5fc96459029bcb7c64204cea7cb866b55ae7b321cd6df0d45f4060.jpg", + "image_caption": [ + "Input feature $F_{in}$" + ], + "image_footnote": [], + "bbox": [ + 369, + 421, + 500, + 484 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/4d6917221a1c6b19dfd7a4900ebddac4a2fdad208f88e6de30c949785748b790.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 356, + 633, + 419 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/1170d6da30fae62866a76bd910c1a6a6a3a0b90a37cfb6e86dbc86f00abaa4b1.jpg", + "image_caption": [ + "$X_{l}$ from EASA" + ], + "image_footnote": [], + "bbox": [ + 501, + 421, + 633, + 484 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/678c0309701644ec5974c3936f3895a50db6b37277883a19d7ce177e42428c56.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 637, + 356, + 769, + 419 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/24721c17aae3917b6c4c9ca66c319cfa2ca111b05d48a4b1ace89e925fb750cf.jpg", + "image_caption": [ + "$Y_{d}$ from LDE", + "Fig. 6: The power spectral density (PSD) visualizations of feature $F_{in}$ , $X_{l}$ , $Y_{d}$ . We perform a periodic shift of the spectrum map such that the low-frequency component is moved to the center. The EASA activates more low-frequency components for feature $X_{l}$ , and the LDE enhances high-frequency representations for feature $Y_{d}$ ." + ], + "image_footnote": [], + "bbox": [ + 637, + 421, + 769, + 484 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "that sequentially process the EASA and LDE, whose corresponding reconstruction performances are clearly decreased as well. These reduced results prove the effectiveness of such a parallel design.", + "bbox": [ + 212, + 613, + 784, + 657 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effectiveness of the variance modulation. The variance modulation is used in the EASA branch to facilitate better non-local information exploration. To demonstrate its effectiveness, we remove this operation from the EASA branch. Table 5 presents that the PSNR values drop by $0.06\\mathrm{dB}$ and $0.08\\mathrm{dB}$ on the Urban100 and Manga109 datasets. For investigating the effect of $\\sigma^2 (X)$ , we further replace $\\sigma^2 (X)$ with the variance of $X_{s}$ and a learnable parameter $\\alpha \\in \\mathbb{R}^{1\\times 1\\times C}$ for feature modulation, and their PSNR values on the Urban100 datasets are reduced to 26.12dB and 26.13dB, respectively. These results suggest that the variance of $X_{s}$ or the learnable parameter $\\alpha$ does not accurately represent the global statistical divergence of the input feature.", + "bbox": [ + 212, + 657, + 785, + 809 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In addition, variance modulation with element-wise product incurs a 0.05dB and 0.11dB PSNR reduction on Urban100 and Manga109, respectively. This", + "bbox": [ + 214, + 809, + 784, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "SMFANet for Efficient Image Super-Resolution", + "bbox": [ + 416, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/5509c7d8ac736e1302e84060fd11c16bd0e8dadb04d7c3caf1562770b29efdc0.jpg", + "table_caption": [ + "Table 4: Ablation experiments for $\\times 4$ SMFANet on Urban100 and Manga109 datasets. All measured metrics are calculated in the same way as in Table 1. “A → B” is to replace A with B, and “A → None” is to remove operator A." + ], + "table_footnote": [], + "table_body": "
AblationVariant#Params(K)#FLOPs(G)#GPU Mem. (M)#Avg. Time(ms)Urban100Manga109
BaselineSMFANet1971193.209.2626.19/0.786130.72/0.9097
SMFASMFA → None82556.903.1225.54/0.764629.76/0.8974
EASA → None1831077.317.2825.94/0.778330.39/0.9061
EASA → Window-based SA23716269.6937.5126.20/0.788230.74/0.9105
7722128776.746.9525.92/0.777230.92/0.9032
SMFA → Dual LDE branches2631485.729.6326.02/0.781030.52/0.9076
SMFA → Dual EASA branches143888.459.2825.94/0.778330.29/0.9076
SMFA → "EASA then LDE"1871177.168.9026.04/0.782430.57/0.9077
SMFA → "LDE then EASA"1871172.988.9526.11/0.784130.64/0.9083
Variance ModulationVariance σ2(X) → None1971193.208.8726.13/0.785230.64/0.9091
Variance σ2(X) → σ2(Xs)1971193.209.3126.12/0.783930.60/0.9086
Variance σ2(X) → Learnable parameter α1971193.209.1826.13/0.785330.65/0.9092
Addition → Element-wise Product1971193.209.1026.14/0.784130.66/0.9083
PCFNPCFN → None132792.946.6425.89/0.777030.38/0.9051
PCFN → FFN174993.118.7126.06/0.782530.56/0.9079
PCFN → ConvFFN2131293.1311.0326.19/0.786130.71/0.9097
PCFN → GDFN20412193.4810.8326.22/0.787230.75/0.9101
", + "bbox": [ + 218, + 198, + 782, + 305 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "performance drop is caused by the element-wise product operation suppressing the feature channel with low variance, resulting in a decreased information flow. Effectiveness of the PCFN. The PCFN is used to facilitate the fusion of the representative features extracted by the SMFA module. To demonstrate the effectiveness of this module, we performed ablation experiments with removal of PCFN and replacement of FFN [40], respectively. Table 5 shows that the exclusion of PCFN and the substitution of FFN each leads to a respective decrease in PSNR values of 0.34dB and 0.16dB on the Manga109 [33] dataset, compared to the baseline.", + "bbox": [ + 212, + 337, + 785, + 474 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Recent FFN variants [36, 43, 45, 53] introduce convolution to boost local encoding capability for better reconstruction performance. However, the expanded hidden features contain redundant information, and processing them with convolution directly increases the computational cost. For example, we use GDFN [45] for channel mixing, resulting in a $\\times 2$ increase in GPU consumption over PCFN, with only a slight improvement in reconstruction performance.", + "bbox": [ + 212, + 476, + 785, + 568 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Thus, our PCFN introduces a partial convolution that processes only $1/4$ of the hidden channels, which reduces redundant computation and memory usage while efficiently encoding local information. Table 4 shows that PCFN achieves comparable SR performance with less computational overhead compared to ConvFFN [53] and GDFN [45].", + "bbox": [ + 212, + 569, + 785, + 645 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 676, + 359, + 694 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We have proposed a simple yet effective SMFANet method to solve image superresolution efficiently. As the key component of SMFANet, our SMFA module contains the EASA for non-local information exploration and the LDE for local detail modeling. Moreover, we develop a PCFN to facilitate the fusion of the non-local and local information extracted by the SMFA module both in the channel and spatial dimensions. Extensive experimental results show that the proposed SMFANet family achieve a favorable trade-off between reconstruction performance and computational efficiency.", + "bbox": [ + 212, + 718, + 785, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Zheng et al.", + "bbox": [ + 271, + 114, + 352, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work has been partly supported by the National Natural Science Foundation of China (Nos. U22B2049, 62272233, 62332010), and the Fundamental Research Funds for the Central Universities (No. 30922010910).", + "bbox": [ + 212, + 178, + 785, + 223 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 247, + 321, + 263 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Ahn, N., Kang, B., Sohn, K.A.: Fast, accurate, and lightweight super-resolution with cascading residual network. In: ECCV (2018)", + "2. Arbeláez, P., Maire, M., Fowlkes, C.C., Malik, J.: Contour detection and hierarchical image segmentation. PAMI 33(5), 898-916 (2011)", + "3. Bevilacqua, M., Roumy, A., Guillemot, C., line Alberi Morel, M.: Low-complexity single-image super-resolution based on nonnegative neighbor embedding. In: BMVC (2012)", + "4. Chen, H., Wang, Y., Guo, T., Xu, C., Deng, Y., Liu, Z., Ma, S., Xu, C., Xu, C., Gao, W.: Pre-trained image processing transformer. In: CVPR (2021)", + "5. Chen, J., hong Kao, S., He, H., Zhuo, W., Wen, S., Lee, C.H., Chan, S.H.G.: Run, don't walk: Chasing higher flops for faster neural networks. In: CVPR (2023)", + "6. Chen, X., Wang, X., Zhou, J., Dong, C.: Activating more pixels in image superresolution transformer. In: CVPR (2023)", + "7. Choi, H., Lee, J., Yang, J.: N-gram in swim transformers for efficient lightweight image super-resolution. In: CVPR (2023)", + "8. Dai, S., Han, M., Xu, W., Wu, Y., Gong, Y.: Soft edge smoothness prior for alpha channel super resolution. In: CVPR (2007)", + "9. Dong, C., Loy, C.C., He, K., Tang, X.: Image super-resolution using deep convolutional networks. PAMI 38(2), 295-307 (2016)", + "0. Dong, C., Loy, C.C., Tang, X.: Accelerating the super-resolution convolutional neural network. In: ECCV (2016)", + "1. Dong, J., Pan, J., Yang, Z., Tang, J.: Multi-scale residual low-pass filter network for image deblurring. In: ICCV (2023)", + "12. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2021)", + "13. Gu, J., Dong, C.: Interpreting super-resolution networks with local attribution maps. In: CVPR (2021)", + "14. Guo, H., Li, J., Dai, T., Ouyang, Z., Ren, X., Xia, S.T.: Mambair: A simple baseline for image restoration with state-space model. arXiv preprint arXiv:2402.15648 (2024)", + "15. Hendrycks, D., Gimpel, K.: Gaussian error linear units. arXiv preprint arXiv:1606.08415 (2016)", + "16. Huang, J.B., Singh, A., Ahuja, N.: Single image super-resolution from transformed self-exemplars. In: CVPR (2015)", + "17. Hui, Z., Gao, X., Yang, Y., Wang, X.: Lightweight image super-resolution with information multi-distillation network. In: ACM MM (2019)", + "18. Kim, J., Lee, J.K., Lee, K.M.: Accurate image super-resolution using very deep convolutional networks. In: CVPR (2016)" + ], + "bbox": [ + 218, + 281, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "SMFANet for Efficient Image Super-Resolution", + "bbox": [ + 416, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "19. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. In: ICLR (2015)", + "20. Li, A., Zhang, L., Liu, Y., Zhu, C.: Feature modulation transformer: Crossrefinement of global representation via high-frequency prior for image superresolution. In: ICCV (2023)", + "21. Li, M., Ma, B., Zhang, Y.: Lightweight image super-resolution with pyramid clustering transformer. TCSVT pp. 1-1 (2023)", + "22. Li, W., Zhou, K., Qi, L., Jiang, N., Lu, J., Jia, J.: LAPAR: Linearly-assembled pixel-adaptive regression network for single image super-resolution and beyond. In: NeurIPS (2020)", + "23. Li, Y., Fan, Y., Xiang, X., Demandolx, D., Ranjan, R., Timofte, R., Gool, L.V.: Efficient and explicit modelling of image hierarchies for image restoration. In: CVPR (2023)", + "24. Li, Z., Liu, Y., Chen, X., Cai, H., Gu, J., Qiao, Y., Dong, C.: Blueprint separable residual network for efficient image super-resolution. In: CVPR Workshops (2022)", + "25. Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: SwinIR: Image restoration using swin transformer. In: ICCV Workshops (2021)", + "26. Lim, B., Son, S., Kim, H., Nah, S., Lee, K.M.: Enhanced deep residual networks for single image super-resolution. In: CVPR Workshops (2017)", + "27. Liu, J., Chen, C., Tang, J., Wu, G.: From coarse to fine: Hierarchical pixel integration for lightweight image super-resolution. In: AAAI (2023)", + "28. Liu, J., Tang, J., Wu, G.: Residual feature distillation network for lightweight image super-resolution. In: ECCV Workshops (2020)", + "29. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin Transformer: Hierarchical vision transformer using shifted windows. In: ICCV (2021)", + "30. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. In: ICLR (2017)", + "31. Lu, Z., Li, J., Liu, H., Huang, C., Zhang, L., Zeng, T.: Transformer for single image super-resolution. In: CVPR Workshops (2022)", + "32. Mao, Y., Zhang, N., Wang, Q., Bai, B., Bai, W., Fang, H., Liu, P., Li, M., Yan, S.: Multi-level dispersion residual network for efficient image super-resolution. In: CVPR Workshops (2023)", + "33. Matsui, Y., Ito, K., Aramaki, Y., Yamasaki, T., Aizawa, K.: Sketch-based manga retrieval using manga109 dataset. arXiv preprint arXiv:1510.04389 (2015)", + "34. Park, N., Kim, S.: How do vision transformers work? In: ICLR (2022)", + "35. Shi, W., Caballero, J., Huszar, F., Totz, J., Aitken, A.P., Bishop, R., Rueckert, D., Wang, Z.: Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In: CVPR (2016)", + "36. Sun, L., Dong, J., Tang, J., Pan, J.: Spatially-adaptive feature modulation for efficient image super-resolution. In: ICCV (2023)", + "37. Sun, L., Pan, J., Tang, J.: ShuffleMixer: An efficient convnet for image superresolution. In: NeurIPS (2022)", + "38. Timofte, R., Agustsson, E., Van Gool, L., Yang, M.H., Zhang, L., et al.: NTIRE 2017 challenge on single image super-resolution: Methods and results. In: CVPR Workshops (2017)", + "39. Timofte, R., DeSmet, V., Van Gool, L.: A+: Adjusted anchored neighborhood regression for fast super-resolution. In: ACCV 2014 (2015)", + "40. Vaswani, Shazeer, Parmar, Uszkoreit, Jones, Gomez, Kaiser, ukasz, L., Polosukhin, Illia: Attention is all you need. In: NeurIPS (2017)", + "41. Wang, H., Chen, X., Ni, B., Liu, Y., Liu, j.: Omni aggregation networks for lightweight image super-resolution. In: CVPR (2023)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Zheng et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "42. Wang, L., Dong, X., Wang, Y., Ying, X., Lin, Z., An, W., Guo, Y.: Exploring sparsity in image super-resolution for efficient inference. In: CVPR (2021)", + "43. Wang, W., Xie, E., Li, X., Fan, D.P., Song, K., Liang, D., Lu, T., Luo, P., Shao, L.: Pvtv2: Improved baselines with pyramid vision transformer. Computational Visual Media 8(3), 1-10 (2022)", + "44. Yang, J., Wright, J., Huang, T.S., Ma, Y.: Image super-resolution via sparse representation. TIP 19(11), 2861-2873 (2010)", + "45. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: CVPR (2022)", + "46. Zeyde, R., Elad, M., Protter, M.: On single image scale-up using sparse-representations. In: Curves and Surfaces (2012)", + "47. Zhang, A., Ren, W., Liu, Y., Cao, X.: Lightweight image super-resolution with superpixel token interaction. In: ICCV (2023)", + "48. Zhang, J., Peng, H., Wu, K., Liu, M., Xiao, B., Fu, J., Yuan, L.: Minivit: Compressing vision transformers with weight multiplexing. In: CVPR (2022)", + "49. Zhang, X., Zeng, H., Guo, S., Zhang, L.: Efficient long-range attention network for image super-resolution. In: ECCV (2022)", + "50. Zhang, Y., Li, K., Li, K., Wang, L., Zhong, B., Fu, Y.: Image super-resolution using very deep residual channel attention networks. In: ECCV (2018)", + "51. Zhang, Y., Tian, Y., Kong, Y., Zhong, B., Fu, Y.: Residual dense network for image super-resolution. In: CVPR (2018)", + "52. Zhou, L., Cai, H., Gu, J., Li, Z., Liu, Y., Chen, X., Qiao, Y., Dong, C.: Efficient image super-resolution using vast-receptive-field attention. In: ECCV Workshops (2022)", + "53. Zhou, Y., Li, Z., Guo, C., Bai, S., Cheng, M., Hou, Q.: Srformer: Permuted self-attention for single image super-resolution. In: ICCV (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 507 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "SMFANet for Efficient Image Super-Resolution", + "bbox": [ + 416, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/SMFANet_ A Lightweight Self-Modulation Feature Aggregation Network for Efficient Image Super-Resolution/b8afe796-2e9d-45e7-96aa-cb21a9d8b896_model.json b/2024/SMFANet_ A Lightweight Self-Modulation Feature Aggregation Network for Efficient Image Super-Resolution/b8afe796-2e9d-45e7-96aa-cb21a9d8b896_model.json new file mode 100644 index 0000000000000000000000000000000000000000..14425178e5775fd1070259776d452570dfca52c7 --- /dev/null +++ b/2024/SMFANet_ A Lightweight Self-Modulation Feature Aggregation Network for Efficient Image Super-Resolution/b8afe796-2e9d-45e7-96aa-cb21a9d8b896_model.json @@ -0,0 +1,3116 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.261, + 0.141, + 0.744, + 0.21 + ], + "angle": 0, + "content": "SMFANet: A Lightweight Self-Modulation Feature Aggregation Network for Efficient Image Super-Resolution" + }, + { + "type": "text", + "bbox": [ + 0.271, + 0.235, + 0.731, + 0.253 + ], + "angle": 0, + "content": "Mingjun Zheng*, Long Sun*, Jiangxin Dong, and Jinshan Pan†" + }, + { + "type": "text", + "bbox": [ + 0.292, + 0.263, + 0.713, + 0.306 + ], + "angle": 0, + "content": "School of Computer Science and Engineering, Nanjing University of Science and Technology, Nanjing, China {mingjunzheng, cs.longsun, jxdong, jspan}@njust.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.333, + 0.746, + 0.637 + ], + "angle": 0, + "content": "Abstract. Transformer-based restoration methods achieve significant performance as the self-attention (SA) of the Transformer can explore non-local information for better high-resolution image reconstruction. However, the key dot-product SA requires substantial computational resources, which limits its application in low-power devices. Moreover, the low-pass nature of the SA mechanism limits its capacity for capturing local details, consequently leading to smooth reconstruction results. To address these issues, we propose a self-modulation feature aggregation (SMFA) module to collaboratively exploit both local and non-local feature interactions for a more accurate reconstruction. Specifically, the SMFA module employs an efficient approximation of self-attention (EASA) branch to model non-local information and uses a local detail estimation (LDE) branch to capture local details. Additionally, we further introduce a partial convolution-based feed-forward network (PCFN) to refine the representative features derived from the SMFA. Extensive experiments show that the proposed SMFANet family achieve a better trade-off between reconstruction performance and computational efficiency on public benchmark datasets. In particular, compared to the \\(\\times 4\\) SwinIR-light, SMFANet+ achieves 0.14dB higher performance over five public testsets on average, and \\(\\times 10\\) times faster runtime, with only about \\(43\\%\\) of the model complexity (e.g., FLOPs). Our source codes and pretrained models are available at: https://github.com/Zheng-MJ/SMFANet." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.65, + 0.725, + 0.664 + ], + "angle": 0, + "content": "Keywords: Efficient image SR \\(\\cdot\\) Self-attention \\(\\cdot\\) Feature aggregation" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.686, + 0.377, + 0.701 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.714, + 0.788, + 0.806 + ], + "angle": 0, + "content": "Single-image super-resolution (SISR) refers to restoring a high-resolution (HR) image from the given degraded low-resolution (LR) counterpart. With the rapid rise of streaming platforms and ultra-high-definition (UHD) content becoming the industrial standard, SISR has attracted significant attention from academia to industry for its ability to provide a pleasurable watching experience at a lower cost of media transmission. Concurrently, the ill-posed nature of SISR" + }, + { + "type": "page_footnote", + "bbox": [ + 0.233, + 0.812, + 0.375, + 0.825 + ], + "angle": 0, + "content": "* Equal contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.233, + 0.825, + 0.396, + 0.84 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.233, + 0.812, + 0.396, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.129 + ], + "angle": 0, + "content": "Zheng et al." + }, + { + "type": "image", + "bbox": [ + 0.237, + 0.144, + 0.501, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.145, + 0.766, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.314, + 0.788, + 0.385 + ], + "angle": 0, + "content": "Fig. 1: Overall performance comparison between our proposed method and other state-of-the-art lightweight methods on Manga109 [33] for \\(\\times 4\\) SR. The circle sizes in the left and right figures represent the number of FLOPs of the model and the GPU usage required for inference, respectively. The proposed SMFANet family achieve a better trade-off between computational efficiency and reconstruction performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.415, + 0.784, + 0.445 + ], + "angle": 0, + "content": "task makes it challenging for conventional hand-crafted prior-based [8,39,44], or interpolation-based approaches to decently solve this problem." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.446, + 0.785, + 0.581 + ], + "angle": 0, + "content": "Over the past decade, deep learning has revolutionized the field of SISR. Various convolutional neural networks (CNNs) [1,9,10,17,24,26,35-37,50] have been developed to solve this problem. As the basic operation of CNNs, the convolution operator is translation-invariant and has limited receptive fields, which limits its ability to model non-local information. For better image restoration, CNN-based methods [26, 50, 51] become deeper and larger to enhance the representation ability. These high-capacity networks consume high computational costs, e.g., RCAN [50] has 15.59M parameters with over 400 layers, which cannot be applied to source-limited devices, e.g., smart phones." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.786, + 0.717 + ], + "angle": 0, + "content": "Recently, the vision Transformer (ViT) [12] achieves impressive success on both high-level [29] and low-level vision tasks [4, 7, 21, 25, 27, 41, 47, 49, 53], as the self-attention (SA) mechanism in the ViT can model non-local information effectively. However, the SA mechanism requires high computational resources with massive memory consumption. To reduce the computational cost, window-based self-attention [7, 25, 29, 49], transposed SA [45], and weight multiplexing strategy [48] have been developed to ease the computational burden. However, these still require a long time to learn feature dependencies for image super-resolution." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.786, + 0.795 + ], + "angle": 0, + "content": "Furthermore, recent studies [11, 34] reveal that ViTs favor learning low-frequency components, which leads to smoothed reconstruction results. Thus, these limitations motivate us to think: Is it possible to develop an efficient feature modulation block that explores non-local information in a SA-like manner while modeling local details for efficient image super-resolution?" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.786, + 0.84 + ], + "angle": 0, + "content": "To this end, we propose a self-modulation feature aggregation (SMFA) module that contains an efficient approximation of self-attention (EASA) branch to exploit non-local information, and a local detail estimation (LDE) branch to" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.418, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SMFANet for Efficient Image Super-Resolution" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.237 + ], + "angle": 0, + "content": "model local features. On the EASA branch, we acquire the low-frequency contents through a downsampling operation, calculate the global variance of the input features to modulate the processed low-frequency features, and then use the modulated features to aggregate the input features adaptively. As the EASA prioritizes non-local structure information exploration, we use the LDE branch with convolutional layers to capture local features in parallel." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.787, + 0.342 + ], + "angle": 0, + "content": "In addition, we develop an efficient partial convolution-based feed-forward network (PCFN), which further refines the representative features derived from the SMFA in both spatial and channel dimensions. We formulate the proposed SMFA module and PCFN into an end-to-end trainable network, called SMFANet, to solve SISR. Extensive experiments show that the proposed SMFANet achieves a favorable trade-off between computational efficiency and reconstruction performance (see Figure 1)." + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.344, + 0.608, + 0.358 + ], + "angle": 0, + "content": "The main contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.367, + 0.784, + 0.41 + ], + "angle": 0, + "content": "- We develop an efficient SMFA module to extract representative features, where the EASA branch for exploring non-local information and the LDE branch for capturing local features." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.412, + 0.784, + 0.44 + ], + "angle": 0, + "content": "- We present a lightweight PCFN to refine the features derived from the SMFA in spatial and channel dimensions further." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.442, + 0.785, + 0.5 + ], + "angle": 0, + "content": "- We conduct quantitative and qualitative evaluations for our proposed SMFANet on public benchmark datasets, and the results demonstrate that our method achieves a favorable trade-off between model complexity and reconstruction performance." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.367, + 0.785, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.523, + 0.388, + 0.539 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.787, + 0.719 + ], + "angle": 0, + "content": "CNN-based super-resolution. SRCNN [9] first introduces effective end-to-end trainable CNNs to map LR directly to its counterpart HR, with superior performance over previous hand-crafted prior and interpolation methods. Then, FSRCNN [10] and ESPCN [35] adopt a post-upsampling strategy to improve efficiency. VDSR [18] uses global residual learning to solve the training problem of deep networks effectively. Therefore, subsequent CNN-based methods can stack deeper and larger layers to capture more information for better image restoration. EDSR [26] increases the model footprint to 43M with significant performance improvement, and RCAN [50], based on channel attention, builds a model with over 400 layers. However, these large models that require high computational costs make them challenging to deploy in real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.787, + 0.84 + ], + "angle": 0, + "content": "To alleviate the computational burden, many CNN-based methods aim to strike a balance between reconstruction performance and model complexity. CARN [1] applies group convolution and cascades an efficient residual network. IMDN [17] adopts the information multi-distillation blocks to split, refine, and aggregate the feature mapping step-by-step, which significantly reduces the model parameters. Furthermore, its improvement work RFDN [28] is the winning solution in the NTIRE 2022 Efficient SR Challenge. BSRN [24] uses blueprint convolutions to reduce computational redundancy significantly. Unlike the above" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.129 + ], + "angle": 0, + "content": "Zheng et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.357 + ], + "angle": 0, + "content": "approaches that employ progressive feature refinement techniques, recent studies [32, 36, 52] tend to introduce non-local feature modulation to extract representative features for reconstruction. VapSR [52] uses large kernel convolutions in the attention branch to enlarge the receptive field. SAFMN [36] enlarges the receptive field by introducing different down-sampling rates to obtain multi-scale spatial features, and then aggregates these features to generate the modulation map. MDRN [32] employs the parallel implementation of multiple stridden convolutions and polling operations to capture finer global structural information. Although these methods can model non-local feature interactions at a low computational cost and achieve good performance, performing reconstruction using aggregated non-local features only and ignoring local features leads to artifacts in local details (e.g., corners and edges) of the super-resolved image. To avoid this problem, we propose an efficient SMFA module that collaboratively models local and non-local features for more accurate reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.359, + 0.788, + 0.633 + ], + "angle": 0, + "content": "ViT-based super-resolution. The vision Transformer (ViT) [12] utilizes the self-attention (SA) mechanism to explore global information for image reconstruction and achieves great success. IPT [4] first introduces a standard vision Transformer to solve the image SR problem. However, the SA mechanism of the ViT consumes huge resources, so various attention variants have been proposed to reduce the computational burden. SwinIR [25] employs window self-attention and outperforms the CNN-based large model [26]. ELAN [49] proposes a group-wise self-attention module and shares the weights to reduce complexity significantly. Moreover, recent studies [11,34] have pointed out that ViTs have the nature of the low-pass filter, resulting in smooth reconstruction results. Therefore, many ViT-based methods [6, 20, 23] utilize convolutional operators to enhance local details for higher-quality image reconstruction. Although these methods integrate the advantages of both convolutional and transformer structures and utilize both local and non-local information to achieve high-quality image reconstruction results, they still rely on window-based attention variants, and feature interactions across windows require much time to perform. Different from these methods, we develop an efficient approximation of self-attention (EASA) branch to model non-local information with much lower model complexity." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.664, + 0.427, + 0.681 + ], + "angle": 0, + "content": "3 Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Our goal is to develop a simple yet effective CNN model to collaboratively explore local and non-local feature information for accurate image super-resolution. To this end, we first develop a self-modulation feature aggregation (SMFA) module to effectively exploit representative features, where an efficient approximate self-attention (EASA) module based on feature modulation is used for non-local feature interactions, while an additional local detail estimation (LDE) branch is used for local information exploration. To refine the features generated by the SMFA layer, we further introduce a partial convolution-based feed-forward network (PCFN). We incorporate these components into a unified unit that" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.418, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "SMFANet for Efficient Image Super-Resolution" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.145, + 0.673, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.683, + 0.149, + 0.776, + 0.16 + ], + "angle": 0, + "content": "\\(⑦\\) Adaptive Max Pooling" + }, + { + "type": "text", + "bbox": [ + 0.684, + 0.171, + 0.773, + 0.181 + ], + "angle": 0, + "content": "\\(\\bullet\\) Element-wise Product" + }, + { + "type": "text", + "bbox": [ + 0.684, + 0.19, + 0.771, + 0.202 + ], + "angle": 0, + "content": "Nearest Up-sampling" + }, + { + "type": "text", + "bbox": [ + 0.684, + 0.211, + 0.776, + 0.221 + ], + "angle": 0, + "content": "\\(\\bigoplus\\) Element-wise Addition" + }, + { + "type": "text", + "bbox": [ + 0.684, + 0.231, + 0.746, + 0.242 + ], + "angle": 0, + "content": "\\(\\widehat{\\mathbf{S}}\\) Channel Split" + }, + { + "type": "text", + "bbox": [ + 0.684, + 0.252, + 0.771, + 0.261 + ], + "angle": 0, + "content": "Variance Calculation" + }, + { + "type": "text", + "bbox": [ + 0.684, + 0.272, + 0.758, + 0.283 + ], + "angle": 0, + "content": "GELU Activation" + }, + { + "type": "text", + "bbox": [ + 0.684, + 0.293, + 0.771, + 0.303 + ], + "angle": 0, + "content": "C Channel Concatenate" + }, + { + "type": "list", + "bbox": [ + 0.683, + 0.149, + 0.776, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.205, + 0.605, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.608, + 0.207, + 0.667, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.326, + 0.785, + 0.397 + ], + "angle": 0, + "content": "Fig.2: Network architecture of the proposed SMFANet. The proposed SMFANet consists of a shallow feature extraction module, feature modulation blocks (FMBs), and a lightweight image reconstruction module. Feature modulation block contains one self-modulation feature aggregation (SMFA) module and one partial convolution-based feed-forward network (PCFN)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.427, + 0.785, + 0.457 + ], + "angle": 0, + "content": "pays more attention to information-rich regions for better reconstruction. In the following, we explain the main ideas of each component in detail." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.48, + 0.427, + 0.493 + ], + "angle": 0, + "content": "3.1 Overall architecture" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.505, + 0.788, + 0.701 + ], + "angle": 0, + "content": "Figure 2 illustrates the overall architecture of our proposed SMFANet. It takes a low-resolution image as input, and uses a \\(3 \\times 3\\) convolutional layer to extract shallow features. The extracted shallow features are then fed to a sequence of feature modulation blocks (FMBs) to produce deep representative features, where the FMB consists of a self-modulation feature aggregation (SMFA) module and a partial convolution-based feed-forward network (PCFN). After the feature modulation block, the representative features are processed by an image reconstruction module to reconstruct the high-quality output. To make the reconstruction module as lightweight as possible, a \\(3 \\times 3\\) convolutional layer is used to convert the channel dimensions to a specific size that adjusts to the upsampling ratio, and a PixelShuffle layer [35] is used for enlargement. To facilitate high-frequency information learning, we insert a global residual connection before the image reconstruction module." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.724, + 0.625, + 0.74 + ], + "angle": 0, + "content": "3.2 Self-modulation feature aggregation module" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Exploring non-local information is important for image SR reconstruction. Recent ViT-based SR methods [4, 7, 25, 47, 49] utilize various self-attention mechanisms to explore non-local information and achieve impressive reconstruction performance. However, these self-attention variants are computationally expensive and have a limited ability to model local details, as their low-pass filter nature makes them preferentially capture low-frequency information [11, 20, 34]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.129 + ], + "angle": 0, + "content": "Zheng et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.236 + ], + "angle": 0, + "content": "To address this problem, we develop a lightweight self-modulation feature aggregation (SMFA) module to collaboratively model local and non-local features for accurate reconstruction. Within the SMFA module, an efficient approximation of self-attention (EASA) branch is implemented for non-local information exploration at a moderate cost, and a local detail estimation (LDE) branch is used for capturing local information." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.237, + 0.785, + 0.298 + ], + "angle": 0, + "content": "Given the input feature \\( F_{in} \\in \\mathbb{R}^{H \\times W \\times C} \\), where \\( H \\times W \\) denotes the spatial size and \\( C \\) is the number of channels, we first apply a \\( 1 \\times 1 \\) convolution to the normalized \\( F_{in} \\) to expand the channel, and then split the channel into two parts as inputs to the EASA and LDE branches:" + }, + { + "type": "equation", + "bbox": [ + 0.391, + 0.306, + 0.785, + 0.322 + ], + "angle": 0, + "content": "\\[\n\\{X, Y \\} = \\mathcal {S} \\left(\\operatorname {C o n v} _ {1 \\times 1} \\left(\\left| \\left| F _ {i n} \\right| \\right| _ {2}\\right)\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.328, + 0.784, + 0.433 + ], + "angle": 0, + "content": "where \\( ||\\cdot||_2 \\) is the \\( L_{2} \\) normalization, \\( Conv_{1\\times 1}(\\cdot) \\) denotes a \\( 1\\times 1 \\) convolutional layer, \\( S(\\cdot) \\) denotes a channel splitting operation, \\( X\\in \\mathbb{R}^{H\\times W\\times C} \\) and \\( Y\\in \\mathbb{R}^{H\\times W\\times C} \\). We then process the features \\( X \\) and \\( Y \\) in parallel via the EASA and LED branches, producing the non-local feature \\( X_{l} \\) and local feature \\( Y_{d} \\), respectively. Finally, we fuse \\( X_{l} \\) and \\( Y_{d} \\) together with element-wise addition and feed them into a \\( 1\\times 1 \\) convolution to form a representative output of the SMFA module. This process can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.416, + 0.442, + 0.785, + 0.457 + ], + "angle": 0, + "content": "\\[\nF _ {\\rho} = \\operatorname {C o n v} _ {1 \\times 1} \\left(X _ {l} + Y _ {d}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.462, + 0.531, + 0.478 + ], + "angle": 0, + "content": "where \\(F_{\\rho}\\in \\mathbb{R}^{H\\times W\\times C}\\) is the output feature." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.479, + 0.784, + 0.524 + ], + "angle": 0, + "content": "Efficient approximation of self-attention. We obtain the low-frequency components through a downsampling operation and feed them into a \\(3 \\times 3\\) depth-wise convolution to generate non-local structure information \\(X_{s} \\in \\mathbb{R}^{H / 8 \\times W / 8 \\times C}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.406, + 0.54, + 0.785, + 0.554 + ], + "angle": 0, + "content": "\\[\nX _ {s} = D W C o n v _ {3 \\times 3} (\\mathcal {D} (X)), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.558, + 0.784, + 0.633 + ], + "angle": 0, + "content": "where \\(\\mathcal{D}(\\cdot)\\) denotes the adaptive max pooling with a scaling factor of 8, \\(DWConv_{3\\times 3}(\\cdot)\\) is a \\(3\\times 3\\) depth-wise convolutional layer. To embed global descriptions for modulating non-local representation \\(X_{s}\\), we introduce the variance of \\(X\\) as the statistical divergence of spatial information, and merge it with the non-local representation \\(X_{s}\\) through a \\(1\\times 1\\) convolution:" + }, + { + "type": "equation", + "bbox": [ + 0.397, + 0.642, + 0.785, + 0.681 + ], + "angle": 0, + "content": "\\[\n\\sigma^ {2} (X) = \\frac {1}{N} \\sum_ {i = 0} ^ {N - 1} \\left(x _ {i} - \\mu\\right) ^ {2}, \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.397, + 0.685, + 0.609, + 0.701 + ], + "angle": 0, + "content": "\\[\nX _ {m} = \\operatorname {C o n v} _ {1 \\times 1} \\left(X _ {s} + \\sigma^ {2} (X)\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.707, + 0.784, + 0.782 + ], + "angle": 0, + "content": "where \\(\\sigma^2 (X)\\in \\mathbb{R}^{1\\times 1\\times C}\\) is the variance of \\(X\\), \\(N\\) is the total number of pixels, \\(x_{i}\\) denotes the value of each pixel, \\(\\mu\\) is the mean of all pixel values, and \\(X_{m}\\in \\mathbb{R}^{H\\times W\\times C}\\) represents the modulated feature. This variance modulation mechanism facilitates better exploring non-local information, and we study it in detail in Section 5." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.784, + 0.784, + 0.814 + ], + "angle": 0, + "content": "Finally, we use the modulated features to aggregate the input feature \\( X \\) for extracting the representative structure information \\( X_{l} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.427, + 0.824, + 0.784, + 0.839 + ], + "angle": 0, + "content": "\\[\nX _ {l} = X \\odot \\mathcal {U} (\\phi (X _ {m})), \\tag {5}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.418, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "SMFANet for Efficient Image Super-Resolution" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "where \\(\\phi(\\cdot)\\) refers to the GELU activation function [15], \\(\\mathcal{U}(\\cdot)\\) denotes a nearest upsampling operation, and \\(\\odot\\) represents the element-wise product operation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.785, + 0.283 + ], + "angle": 0, + "content": "Local detail estimation. Local details are important for the pleasing high-frequency reconstruction. As the EASA prioritizes non-local structure information exploration, we develop a simple local detail estimation layer to capture local features simultaneously. In detail, an expanded depth-wise convolution with a kernel size of \\(3 \\times 3\\) is used to encode local information \\(Y_{h}\\) from the input features \\(Y\\). Then, we use two \\(1 \\times 1\\) convolutions with a hidden GELU activation to generate the enhanced local feature \\(Y_{d}\\), which is achieved by:" + }, + { + "type": "equation", + "bbox": [ + 0.383, + 0.291, + 0.785, + 0.317 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} Y _ {h} = \\operatorname {C o n v} _ {1 \\times 1} (D W \\operatorname {C o n v} _ {3 \\times 3} (Y)), \\\\ Y _ {v} = \\operatorname {C o n v} _ {1} (\\phi (Y)) \\end{array} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.332, + 0.618, + 0.349 + ], + "angle": 0, + "content": "where \\(Y_{h}\\in \\mathbb{R}^{H\\times W\\times 2C}\\) is the encoded local information." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.368, + 0.652, + 0.382 + ], + "angle": 0, + "content": "3.3 Partial convolution-based feed-forward network" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.39, + 0.784, + 0.464 + ], + "angle": 0, + "content": "The regular feed-forward network (FFN) [12, 40] operates on each pixel location identically, which lacks the exchange of information in the spatial dimension. Inspired by [5, 37], we improve the FFN and propose an efficient partial convolution-based feed-forward network (PCFN) to further refine the representative features derived from the SMFA." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.466, + 0.785, + 0.571 + ], + "angle": 0, + "content": "Figure 2(c) shows that the PCFN uses a \\(1 \\times 1\\) convolution with the GELU activation function to perform cross-channel interaction on the expanded hidden space. Then, it splits the hidden features into two chunks of \\(\\{F_{\\rho}^{1}, F_{\\rho}^{2}\\}\\) and employs a \\(3 \\times 3\\) convolution followed by a GELU activation to process \\(F_{\\rho}^{1}\\) to encode local contextual information. The processed \\(F_{\\rho}^{1}\\) and \\(F_{\\rho}^{2}\\) are then concatenated and fed into a \\(1 \\times 1\\) convolution for further feature mixing and reducing the hidden channel back to the original input dimension. This process is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.354, + 0.591, + 0.785, + 0.618 + ], + "angle": 0, + "content": "\\[\n\\left\\{F _ {\\rho} ^ {1}, F _ {\\rho} ^ {2} \\right\\} = \\mathcal {S} \\left(\\phi \\left(\\operatorname {C o n v} _ {1 \\times 1} \\left(\\left| \\left| F _ {\\rho} \\right| \\right| _ {2}\\right)\\right)\\right), \\\\ \\hat {\\hat {r}} - C = \\left(\\mathcal {A} \\left[ + (C - \\left(F _ {1}\\right)) - F _ {2} ^ {2} \\right]\\right) \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.354, + 0.612, + 0.65, + 0.631 + ], + "angle": 0, + "content": "\\[\n\\hat {F} _ {\\rho} = C o n v _ {1 \\times 1} (\\mathcal {C} ([ \\phi (C o n v _ {3 \\times 3} (F _ {\\rho} ^ {1})), F _ {\\rho} ^ {2} ])),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.637, + 0.784, + 0.67 + ], + "angle": 0, + "content": "where \\(F_{\\rho}^{1}\\in \\mathbb{R}^{H\\times W\\times C / 2}\\) and \\(F_{\\rho}^{2}\\in \\mathbb{R}^{H\\times W\\times 3C / 2}\\), \\(\\mathcal{C}(\\cdot)\\) represents a concatenation operation, and \\(\\operatorname{Conv}_{3\\times 3}(\\cdot)\\) denotes a \\(3\\times 3\\) convolutional layer." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.688, + 0.473, + 0.702 + ], + "angle": 0, + "content": "3.4 Feature modulation block" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.71, + 0.784, + 0.77 + ], + "angle": 0, + "content": "As mentioned previously, we formulate the proposed SMFA and PCFN into a feature modulation block (FMB) to produce deep representative features. To stabilize the model training and encourage more information flow, the residual connection is employed. The FMB can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.413, + 0.779, + 0.785, + 0.804 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} F _ {\\rho} = S M F A \\left(F _ {i n}\\right) + F _ {i n}, \\\\ \\hat {\\hat {F}} = \\operatorname {R G E N} (F _ {i}) + F _ {i}. \\end{array} \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.413, + 0.799, + 0.58, + 0.815 + ], + "angle": 0, + "content": "\\[\n\\hat {F} _ {\\rho} = P C F N (F _ {\\rho}) + F _ {\\rho},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.824, + 0.71, + 0.843 + ], + "angle": 0, + "content": "where \\(F_{in}\\), \\(F_{\\rho}\\) and \\(\\hat{F}_{\\rho}\\) remain consistent with the previous definition." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.129 + ], + "angle": 0, + "content": "Zheng et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.461, + 0.163 + ], + "angle": 0, + "content": "4 Experimental Results" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.181, + 0.498, + 0.197 + ], + "angle": 0, + "content": "4.1 Dataset and implementation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.209, + 0.788, + 0.36 + ], + "angle": 0, + "content": "Datasets. The high-quality DIV2K [38] and DIV2K + Flickr2K (DF2K) [38] datasets are commonly used for training SR models, and previous methods [14,17,25,26,36] chose one of them as the training dataset. To make a fair comparison with recent state-of-the-art methods, we train our proposed method on the DIV2K and DF2K datasets, respectively. In addition, we evaluate our method on the commonly used test datasets, including Set5 [3], Set14 [46], B100 [2], Urban100 [16] and Manga109 [33]. We transform the image into YCbCr color space and calculate the peak signal-to-noise ratio (PSNR) and structure similarity index (SSIM) on the Y channel of the image to evaluate the quality of the recovered image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.362, + 0.788, + 0.543 + ], + "angle": 0, + "content": "Implementation details. During the training, we randomly crop 64 patches of size \\(64 \\times 64\\) with random horizontal flips and rotations from LR images as the basic training inputs. The proposed model is trained with the same loss function as SAFMN [36] and optimized by Adam [19] with \\(\\beta_{1} = 0.9\\) and \\(\\beta_{2} = 0.99\\). We set the initial learning rate to \\(1 \\times 10^{-3}\\) and the minimum one to \\(1 \\times 10^{-5}\\), which is updated by the Cosine Annealing scheme [30]. The number of iterations for all experiments is set to 1,000,000. All experiments are conducted with the PyTorch framework on an NVIDIA GeForce RTX 3090 GPU. We train two versions of SMFANet with different complexity. The standard SMFANet consists of 8 FMBs with 36 channels while the large version, SMFANet+, has 12 FMBs with 48 channels. Due to the page limit, we include more experimental results in the supplemental material." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.569, + 0.618, + 0.584 + ], + "angle": 0, + "content": "4.2 Comparisons with state-of-the-art methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.788, + 0.81 + ], + "angle": 0, + "content": "Quantitative comparison. To fully evaluate the performance of our method, we first compare our SMFANet with state-of-the-art CNN-based lightweight SR methods, including FSRCNN [10], CARN [1], EDSR-baseline [26], IMDN [17], LAPAR-A [22], SMSR [42], ShuffleMixer [37], SAFMN [36], and further compare the SMFANet+ with recent ViT-based lightweight algorithms, including ESRT [31], SwinIR-light [25], ELAN-light [49], HPINet-S [27], NGswin [7], SPI-N [47] and SRFormer-light [53]. The quantitative comparisons on benchmark datasets for the upscaling factors of \\(\\times 2\\), \\(\\times 3\\), and \\(\\times 4\\) are reported in Table 1 and Table 2, where SMFANet (DIV2K) denotes training using the DIV2K dataset, and SMFANet (DF2K) uses DF2K as the training dataset. In addition to PSNR and SSIM metrics, we also list the number of parameters (#Params) and FLOPs (#FLOPs). For a fair comparison, we calculate the model complexity of all evaluated methods with the fvcore library (i.e., fvcore.nn.flop_count_str) under a setting of super resolving an LR image to \\(1280 \\times 720\\) pixels." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Benefiting from the ability to model non-local information of the SMFA module, SMFANet can effectively explore more information compared to previous" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.418, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "SMFANet for Efficient Image Super-Resolution" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 1: Comparison with CNN-based lightweight SR methods on public benchmark datasets. All PSNR/SSIM results are calculated on the Y-channel. #FLOPs is measured corresponding to an HR image of the size \\(1280 \\times 720\\) pixels. The best and second-best performances are highlighted in Red and Blue colors." + }, + { + "type": "table", + "bbox": [ + 0.226, + 0.213, + 0.784, + 0.503 + ], + "angle": 0, + "content": "
ScaleMethods#Params (K)#FLOPs (G)Set5Set14B100Urban100Manga109
×2FSRCNN [10]12637.00/0.955832.63/0.908831.53/0.892029.88/0.902036.67/0.9694
CARN [1]159222337.76/0.959033.52/0.916632.09/0.897831.92/0.925638.36/0.9765
EDSR-baseline [26]137031637.99/0.960433.57/0.917532.16/0.899431.98/0.927238.54/0.9769
IMDN [17]69416138.00/0.960533.63/0.917732.19/0.899632.17/0.928338.88/0.9774
LAPAR-A [22]54817138.01/0.960533.62/0.918332.19/0.899932.10/0.928338.67/0.9772
SMSR [42]98513238.00/0.960133.64/0.917932.17/0.899932.19/0.928438.76/0.9771
ShuffleMixer [37]3949138.01/0.960633.63/0.918032.17/0.899531.89/0.925738.83/0.9774
SAFMN [36]2285238.00/0.960533.54/0.917732.16/0.899531.84/0.925638.71/0.9771
SMFANet (DIV2K)1864138.04/0.960633.65/0.918632.19/0.899932.22/0.928338.98/0.9776
SMFANet (DF2K)1864138.08/0.960733.65/0.918532.22/0.900232.20/0.928239.11/0.9779
×3FSRCNN [10]12533.16/0.914029.43/0.824228.53/0.791026.43/0.808030.98/0.9212
CARN [1]159211934.29/0.925530.29/0.840729.06/0.803428.06/0.849333.50/0.9440
EDSR-baseline [26]155516034.37/0.927030.28/0.841729.09/0.805228.15/0.852733.45/0.9439
IMDN [17]7037234.36/0.927030.32/0.841729.09/0.804628.17/0.851933.61/0.9445
LAPAR-A [22]59411434.36/0.926730.34/0.841229.11/0.805428.15/0.852333.51/0.9441
SMSR [42]9936834.40/0.927030.33/0.841229.10/0.805028.25/0.853633.68/0.9445
ShuffleMixer [37]4154334.40/0.927230.37/0.842329.12/0.805128.08/0.849833.69/0.9448
SAFMN [36]2332334.34/0.926730.33/0.841829.08/0.804827.95/0.847433.52/0.9437
SMFANet (DIV2K)1911934.46/0.927530.39/0.843229.13/0.805928.25/0.852533.84/0.9454
SMFANet (DF2K)1911934.42/0.927430.41/0.843029.16/0.806528.22/0.852333.96/0.9460
×4FSRCNN [10]12530.71/0.865727.59/0.753526.98/0.715024.62/0.728027.90/0.8517
CARN [1]15929132.13/0.893728.60/0.780627.58/0.734926.07/0.783730.47/0.9084
EDSR-baseline [26]151811432.09/0.893828.58/0.781327.57/0.735726.04/0.784930.35/0.9067
IMDN [17]7154132.21/0.894828.58/0.781127.56/0.735326.04/0.783830.45/0.9075
LAPAR-A [22]6599432.15/0.894428.61/0.781827.61/0.736626.14/0.787130.42/0.9074
SMSR [42]10064232.12/0.893228.55/0.780827.55/0.735126.11/0.786830.54/0.9085
ShuffleMixer [37]4112832.21/0.895328.66/0.782727.61/0.736626.08/0.783530.65/0.9093
SAFMN [36]2401432.18/0.894828.60/0.781327.58/0.735925.97/0.780930.43/0.9063
SMFANet (DIV2K)1971132.25/0.895628.67/0.782527.61/0.737126.19/0.786130.72/0.9097
SMFANet (DF2K)1971132.25/0.895628.71/0.783327.64/0.737726.18/0.786230.82/0.9104
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.535, + 0.784, + 0.61 + ], + "angle": 0, + "content": "CNN-based methods. Table 1 shows that our SMFANet achieves better performance on almost all benchmark datasets. For instance, on the \\(\\times 4\\) Manga109 dataset, SMFANet (DIV2K) outperforms IMDN [17] by \\(0.27\\mathrm{dB}\\) with only \\(27\\%\\) FLOPs. When training SMFANet on the larger DF2K dataset, the performance gain is further improved by \\(0.1\\mathrm{dB}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.612, + 0.785, + 0.763 + ], + "angle": 0, + "content": "Table 2 shows that the proposed SMFANet+ achieves competitive performance with lower model complexity. Note that ViT-based methods are based on the self-attention mechanism to model long-range dependency, which consumes significant computational costs. In contrast, our SMFA module, with only a few low-consumption operators, efficiently explores non-local information for high-quality reconstruction. On the five benchmark datasets, SMFANet \\(^+\\) (DIV2K) shows similar performance to SwinIR-light [25] with only \\(43\\%\\) FLOPs, and SMFANet \\(^+\\) (DF2K) improves the average PSNR gain of the \\(\\times 4\\) SR to 0.14dB. These comparisons suggest that our proposed SMFANet can exploit more information than previous CNN-based and ViT-based approaches." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Qualitative comparisons. We compare the visual results of our proposed SMFANet and SMFANet+ with CNN-based methods and ViT-based methods on the \\(\\times 4\\) Urban100 dataset, respectively. The evaluated CNN-based methods [17,22,26,36,37] are listed at the top of Figure 3, which suffer from blurring artifacts and distorted lines. In contrast, our method can restore more accurate" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.129 + ], + "angle": 0, + "content": "Zheng et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 2: Comparison with ViT-based lightweight SR methods on public benchmark datasets. All measured metrics are calculated in the same way as in Table 1. The best and second-best performances are highlighted in Red and Blue colors." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.199, + 0.776, + 0.42 + ], + "angle": 0, + "content": "
ScaleMethods#Params (K)#FLOPs (G)Set5Set14B100Urban100Manga109
×2ESRT [31]678111638.03/0.960033.75/0.918432.25/0.900132.58/0.931839.12/0.9774
SwinIR-light [25]91024438.14/0.961133.86/0.920632.31/0.901232.76/0.934039.12/0.9783
ELAN-light [49]62120338.17/0.961133.94/0.920732.30/0.901232.76/0.934039.11/0.9782
NGswin [7]99814638.05/0.961033.79/0.919932.27/0.900832.53/0.932438.97/0.9777
SPIN [47]49711438.20/0.961533.90/0.921532.31/0.901532.79/0.934039.18/0.9784
SMFANet+ (DIV2K)48010838.18/0.961133.82/0.920232.28/0.901132.64/0.932339.25/0.9777
SMFANet+ (DF2K)48010838.19/0.961133.92/0.920732.32/0.901532.70/0.933139.46/0.9787
×3ESRT [31]77083534.42/0.926830.43/0.843329.15/0.806328.46/0.857433.95/0.9455
SwinIR-light [25]91811434.62/0.928930.54/0.846329.20/0.808228.66/0.862433.98/0.9478
ELAN-light [49]6299034.61/0.928830.55/0.846329.21/0.808128.69/0.862434.00/0.9478
NGswin [7]10076634.52/0.928230.53/0.845629.19/0.807828.52/0.860333.89/0.9470
SPIN [47]5695834.65/0.929330.57/0.846429.23/0.808928.71/0.862734.24/0.9489
SMFANet+ (DIV2K)4874834.63/0.928530.52/0.845629.23/0.808428.59/0.859434.17/0.9478
SMFANet+ (DF2K)4874834.66/0.929230.57/0.846129.25/0.809028.67/0.861134.45/0.9490
×4ESRT [31]75229832.19/0.894728.69/0.783327.69/0.737926.39/0.796230.75/0.9100
SwinIR-light [25]9306532.44/0.897628.77/0.785827.69/0.740626.47/0.798030.92/0.9151
ELAN-light [49]6405432.43/0.897528.78/0.785827.69/0.740626.54/0.798230.92/0.9150
HPINet-S [27]4638832.47/0.898728.80/0.787227.69/0.741626.59/0.801630.92/0.9143
NGswin [7]10194032.33/0.896328.78/0.785927.66/0.739626.45/0.796330.80/0.9128
SPIN [47]5554232.48/0.898328.80/0.786227.70/0.741526.55/0.799830.98/0.9156
SRFormer-light [53]8736332.51/0.898828.82/0.787227.73/0.742226.67/0.803231.17/0.9165
SMFANet+ (DIV2K)4962832.43/0.897928.77/0.784927.70/0.740026.45/0.794331.06/0.9138
SMFANet+ (DF2K)4962832.51/0.898528.87/0.787227.74/0.741226.56/0.797631.29/0.9163
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.433, + 0.788, + 0.504 + ], + "angle": 0, + "content": "Table 3: Memory and running time comparisons on \\(\\times 4\\) SR. #GPU Mem. denotes the maximum GPU memory consumption during the inference phase, derived with the Pytorch torch.cuda.max_memory_allocated() function. #Avg.Time is the average running time on 500 LR images of \\(320 \\times 180\\) pixels. The best and second-best performances are highlighted in Red and Blue colors." + }, + { + "type": "table", + "bbox": [ + 0.226, + 0.515, + 0.779, + 0.616 + ], + "angle": 0, + "content": "
TypeMethodsQuality MetricsTypeMethodsQuality Metrics
#GPU Mem. [M]#Avg. Time [ms]#GPU Mem. [M]#Avg. Time [ms]
CNNCARN [1]702.0713.54ViTSwinIR-light [25]342.44177.06
EDSR-baseline [26]507.1315.21ELAN-light [49]241.3442.75
IMDN [17]204.279.64HPINet-S [27]445.92140.45
LAPAR-A [22]1811.4624.40NGswin [7]372.85108.25
ShuffleMixer [37]474.7919.21SPIN [47]441.52798.44
SAFMN [36]65.268.44SRFormer-light [53]320.95145.99
SMFANet93.209.26SMFANet+247.5117.39
" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.644, + 0.787, + 0.688 + ], + "angle": 0, + "content": "parallel patterns with sharper boundaries. The bottom of Figure 3 shows that ViT-based methods [7,25,27,31,49] do not restore the structures well, e.g., strips, while our SMFANet+ recovers an image with better structures." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Memory and running time comparisons. To fully demonstrate the efficiency of our proposed SMFANet, we further compare our method with the CNN-based [1, 17, 22, 26, 36, 37] and ViT-based [7, 25, 27, 47, 49, 53] methods in terms of GPU memory consumption and inference time on \\(\\times 4\\) SR. Table 3 shows the comparisons of GPU memory and inference time. Our SMFANet has faster inference than the listed CNN methods except SAFMN, and the larger SMFANet+ presents a significant efficiency advantage over ViT-based methods. The running time of SMFANet+ is \\(17.39\\mathrm{ms}\\), which is \\(\\times 10\\) times faster than SwinIR-light [25] and \\(\\times 6\\) times faster than NGswin [7]. Moreover, when compared to ELAN-light [49], which uses a shared group-wise multi-scale self" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.418, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SMFANet for Efficient Image Super-Resolution" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.146, + 0.396, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.398, + 0.145, + 0.495, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.145, + 0.598, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.603, + 0.145, + 0.701, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.145, + 0.805, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.247, + 0.396, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.236, + 0.339, + 0.381, + 0.347 + ], + "angle": 0, + "content": "image078 from Urban100 [16]" + }, + { + "type": "image", + "bbox": [ + 0.398, + 0.247, + 0.495, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.4, + 0.339, + 0.495, + 0.347 + ], + "angle": 0, + "content": "(e) ELAN-light [49]" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.247, + 0.598, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.514, + 0.339, + 0.585, + 0.347 + ], + "angle": 0, + "content": "(f) NGswin [7]" + }, + { + "type": "image", + "bbox": [ + 0.604, + 0.247, + 0.701, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.609, + 0.339, + 0.695, + 0.347 + ], + "angle": 0, + "content": "(g) HPINet-S [27]" + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.247, + 0.805, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.719, + 0.339, + 0.793, + 0.347 + ], + "angle": 0, + "content": "(h) SMFANet+" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.371, + 0.788, + 0.427 + ], + "angle": 0, + "content": "Fig. 3: Visual comparisons for \\(\\times 4\\) SR on the Urban100 dataset. The comparisons between SMFANet and CNN-based methods are shown at the top of the Figure, while the comparisons between SMFANet+ and ViT-based methods are exhibited at the bottom. Our proposed method recovers the image with more accurate structures." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.459, + 0.785, + 0.519 + ], + "angle": 0, + "content": "attention to accelerate model efficiency, the SMFANet+ still has a much faster runtime (17.39ms vs. 42.75ms). These quantitative and qualitative comparison results show that our method obtains even better results against advanced SR models but with significantly less complexity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.52, + 0.785, + 0.641 + ], + "angle": 0, + "content": "LAM comparisons. The local attribution map (LAM) [13] indicates the significant correlation between the red pixels and the rectangular position patch during the restoration process. We compare the LAM [13] results between our method and the evaluated CNN-based [17, 36, 37] and ViT-based [25, 27, 49] methods in Figure 4, and label the corresponding diffusion index (DI) value below each subgraph, where a larger DI value indicates a wider range of pixels involved. These results demonstrate that the proposed SMFANet can explore more non-local information for accurate image super-resolution." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.668, + 0.494, + 0.685 + ], + "angle": 0, + "content": "5 Analysis and Discussions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.785, + 0.779 + ], + "angle": 0, + "content": "In this section, we conduct extensive ablation studies to analyze and evaluate the effect of each component in the proposed SMFANet. We implement all ablation experiments based on the \\(\\times 4\\) SMFANet model and train them with the DIV2K [38] dataset for fair comparisons. The quantitative ablation results in Table 4 are measured on the Urban100 [16] and Manga109 [33] datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Effectiveness of the SMFA. The proposed SMFA module utilizes a parallel structure that absorbs local and non-local feature extraction to enhance reconstruction accuracy. To demonstrate its effectiveness, we first disable the SMFA module to compare it with the baseline SMFANet. Table 4 shows that the PSNR" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.129 + ], + "angle": 0, + "content": "Zheng et al." + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.146, + 0.396, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.246, + 0.24, + 0.371, + 0.248 + ], + "angle": 0, + "content": "image010 from Urban100" + }, + { + "type": "image", + "bbox": [ + 0.404, + 0.147, + 0.496, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.404, + 0.193, + 0.496, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.4, + 0.239, + 0.5, + 0.256 + ], + "angle": 0, + "content": "(a) ShuffleMixer [37] \nDI: 3.5808" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.146, + 0.596, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.193, + 0.595, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.239, + 0.59, + 0.256 + ], + "angle": 0, + "content": "(b) SAFMN [36] \nDI: 8.6313" + }, + { + "type": "image", + "bbox": [ + 0.6, + 0.146, + 0.691, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.193, + 0.689, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.609, + 0.239, + 0.68, + 0.256 + ], + "angle": 0, + "content": "(c) IMDN [17] DI: 10.3650" + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.146, + 0.786, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.193, + 0.786, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.708, + 0.239, + 0.774, + 0.256 + ], + "angle": 0, + "content": "(d) SMFANet DI 12.0899" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.262, + 0.396, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.252, + 0.357, + 0.365, + 0.364 + ], + "angle": 0, + "content": "img005 from Urban100" + }, + { + "type": "image", + "bbox": [ + 0.403, + 0.262, + 0.496, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.402, + 0.308, + 0.496, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.4, + 0.356, + 0.499, + 0.372 + ], + "angle": 0, + "content": "(a) SwinIR-light [25] \nDI: 6.74785" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.262, + 0.596, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.308, + 0.595, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.356, + 0.593, + 0.372 + ], + "angle": 0, + "content": "b) HPINet-S [27] \nDI: 12.8471" + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.262, + 0.691, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.308, + 0.691, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.598, + 0.356, + 0.692, + 0.372 + ], + "angle": 0, + "content": "(c) ELAN-light [49] \nDI: 18.2614" + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.262, + 0.786, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.308, + 0.786, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.701, + 0.356, + 0.776, + 0.372 + ], + "angle": 0, + "content": "(d) SMFANet+ \nDI: 19.6512" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.396, + 0.785, + 0.438 + ], + "angle": 0, + "content": "Fig. 4: Comparison of local attribution maps (LAMs) [13] and diffusion indices (DIs) [13]. Our proposed SMFANet family exploit more feature information and reconstructs a more accurate image structure." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.473, + 0.784, + 0.503 + ], + "angle": 0, + "content": "values drop by 0.65dB and 0.96dB on the Urban100 and Manga109 datasets. These results show the importance of the SMFA." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.506, + 0.787, + 0.778 + ], + "angle": 0, + "content": "In addition, as the proposed SMFA module mainly contains an EASA branch to explore non-local information and a LDE branch to capture local features, we conduct ablation studies with respect to these components to demonstrate their effectiveness on image SR. Without the EASA branch, significant performance drops of \\(0.25\\mathrm{dB}\\) and \\(0.33\\mathrm{dB}\\) are observed on the Urban100 and Manga109 datasets. Figure 5 (c) shows the corresponding visual results with blurred structure. Moreover, replacing EASA with the window-based SA [25] yields almost the same performance, but the corresponding runtime and GPU consumption are three times higher. These results indicate that the EASA branch can efficiently approximate SA to explore non-local information for accurate structural reconstruction. As for the LDE branch, removing it, the model only achieves the performance of \\(25.92\\mathrm{dB}\\) and \\(30.22\\mathrm{dB}\\) on the Urban100 and Manga109. Figure 5 (d) shows the reconstructed images with distorted lines, suggesting that the LDE branch contributes to recovering image details. Figure 6 shows the power spectral density maps, which intuitively illustrate the complementary roles of the EASA and LDE. The PSD of \\(X_{l}\\) derived from the EASA shows that the energy density is more concentrated in the center region, whereas the PSD of \\(Y_{d}\\) obtained from the LDE scatters in the surrounding area compared to \\(F_{in}\\) and \\(X_{l}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.786, + 0.841 + ], + "angle": 0, + "content": "To further validate the effectiveness of this complementary two-branch design, we replace SMFA with either dual LDE branches or dual EASA branches. With these two alternatives, the PSNR values on the Urban100 dataset drop significantly by 0.17dB and 0.25dB, respectively. We also investigated experiments" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.418, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SMFANet for Efficient Image Super-Resolution" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.144, + 0.333, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.211, + 0.331, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.234, + 0.275, + 0.324, + 0.29 + ], + "angle": 0, + "content": "(a) HR patch" + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.144, + 0.444, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.21, + 0.445, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.346, + 0.275, + 0.435, + 0.29 + ], + "angle": 0, + "content": "(b) LR patch" + }, + { + "type": "image", + "bbox": [ + 0.447, + 0.144, + 0.556, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.451, + 0.21, + 0.554, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.454, + 0.275, + 0.55, + 0.29 + ], + "angle": 0, + "content": "(c) w/o EASA" + }, + { + "type": "image", + "bbox": [ + 0.559, + 0.144, + 0.668, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.565, + 0.21, + 0.665, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.569, + 0.275, + 0.659, + 0.29 + ], + "angle": 0, + "content": "(d) w/o LDE" + }, + { + "type": "image", + "bbox": [ + 0.67, + 0.144, + 0.781, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.678, + 0.21, + 0.78, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.681, + 0.275, + 0.772, + 0.29 + ], + "angle": 0, + "content": "(e) SMFANet" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.314, + 0.785, + 0.342 + ], + "angle": 0, + "content": "Fig. 5: Effectiveness of the SMFA for image super-resolution on the \\(\\times 4\\) Urban100 and Manga109 datasets." + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.357, + 0.366, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.422, + 0.367, + 0.486 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.266, + 0.489, + 0.334, + 0.502 + ], + "angle": 0, + "content": "HR image" + }, + { + "type": "image", + "bbox": [ + 0.369, + 0.357, + 0.501, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.37, + 0.422, + 0.501, + 0.486 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.376, + 0.489, + 0.492, + 0.502 + ], + "angle": 0, + "content": "Input feature \\(F_{in}\\)" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.357, + 0.635, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.422, + 0.635, + 0.486 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.519, + 0.489, + 0.619, + 0.501 + ], + "angle": 0, + "content": "\\(X_{l}\\) from EASA" + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.357, + 0.77, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.422, + 0.77, + 0.486 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.659, + 0.489, + 0.749, + 0.501 + ], + "angle": 0, + "content": "\\(Y_{d}\\) from LDE" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.527, + 0.785, + 0.582 + ], + "angle": 0, + "content": "Fig. 6: The power spectral density (PSD) visualizations of feature \\( F_{in} \\), \\( X_{l} \\), \\( Y_{d} \\). We perform a periodic shift of the spectrum map such that the low-frequency component is moved to the center. The EASA activates more low-frequency components for feature \\( X_{l} \\), and the LDE enhances high-frequency representations for feature \\( Y_{d} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.785, + 0.658 + ], + "angle": 0, + "content": "that sequentially process the EASA and LDE, whose corresponding reconstruction performances are clearly decreased as well. These reduced results prove the effectiveness of such a parallel design." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.787, + 0.81 + ], + "angle": 0, + "content": "Effectiveness of the variance modulation. The variance modulation is used in the EASA branch to facilitate better non-local information exploration. To demonstrate its effectiveness, we remove this operation from the EASA branch. Table 5 presents that the PSNR values drop by \\(0.06\\mathrm{dB}\\) and \\(0.08\\mathrm{dB}\\) on the Urban100 and Manga109 datasets. For investigating the effect of \\(\\sigma^2 (X)\\), we further replace \\(\\sigma^2 (X)\\) with the variance of \\(X_{s}\\) and a learnable parameter \\(\\alpha \\in \\mathbb{R}^{1\\times 1\\times C}\\) for feature modulation, and their PSNR values on the Urban100 datasets are reduced to 26.12dB and 26.13dB, respectively. These results suggest that the variance of \\(X_{s}\\) or the learnable parameter \\(\\alpha\\) does not accurately represent the global statistical divergence of the input feature." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "In addition, variance modulation with element-wise product incurs a 0.05dB and 0.11dB PSNR reduction on Urban100 and Manga109, respectively. This" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.129 + ], + "angle": 0, + "content": "Zheng et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 4: Ablation experiments for \\(\\times 4\\) SMFANet on Urban100 and Manga109 datasets. All measured metrics are calculated in the same way as in Table 1. “A → B” is to replace A with B, and “A → None” is to remove operator A." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.199, + 0.784, + 0.306 + ], + "angle": 0, + "content": "
AblationVariant#Params(K)#FLOPs(G)#GPU Mem. (M)#Avg. Time(ms)Urban100Manga109
BaselineSMFANet1971193.209.2626.19/0.786130.72/0.9097
SMFASMFA → None82556.903.1225.54/0.764629.76/0.8974
EASA → None1831077.317.2825.94/0.778330.39/0.9061
EASA → Window-based SA23716269.6937.5126.20/0.788230.74/0.9105
7722128776.746.9525.92/0.777230.92/0.9032
SMFA → Dual LDE branches2631485.729.6326.02/0.781030.52/0.9076
SMFA → Dual EASA branches143888.459.2825.94/0.778330.29/0.9076
SMFA → "EASA then LDE"1871177.168.9026.04/0.782430.57/0.9077
SMFA → "LDE then EASA"1871172.988.9526.11/0.784130.64/0.9083
Variance ModulationVariance σ2(X) → None1971193.208.8726.13/0.785230.64/0.9091
Variance σ2(X) → σ2(Xs)1971193.209.3126.12/0.783930.60/0.9086
Variance σ2(X) → Learnable parameter α1971193.209.1826.13/0.785330.65/0.9092
Addition → Element-wise Product1971193.209.1026.14/0.784130.66/0.9083
PCFNPCFN → None132792.946.6425.89/0.777030.38/0.9051
PCFN → FFN174993.118.7126.06/0.782530.56/0.9079
PCFN → ConvFFN2131293.1311.0326.19/0.786130.71/0.9097
PCFN → GDFN20412193.4810.8326.22/0.787230.75/0.9101
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.338, + 0.787, + 0.475 + ], + "angle": 0, + "content": "performance drop is caused by the element-wise product operation suppressing the feature channel with low variance, resulting in a decreased information flow. Effectiveness of the PCFN. The PCFN is used to facilitate the fusion of the representative features extracted by the SMFA module. To demonstrate the effectiveness of this module, we performed ablation experiments with removal of PCFN and replacement of FFN [40], respectively. Table 5 shows that the exclusion of PCFN and the substitution of FFN each leads to a respective decrease in PSNR values of 0.34dB and 0.16dB on the Manga109 [33] dataset, compared to the baseline." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.477, + 0.787, + 0.569 + ], + "angle": 0, + "content": "Recent FFN variants [36, 43, 45, 53] introduce convolution to boost local encoding capability for better reconstruction performance. However, the expanded hidden features contain redundant information, and processing them with convolution directly increases the computational cost. For example, we use GDFN [45] for channel mixing, resulting in a \\(\\times 2\\) increase in GPU consumption over PCFN, with only a slight improvement in reconstruction performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.57, + 0.787, + 0.646 + ], + "angle": 0, + "content": "Thus, our PCFN introduces a partial convolution that processes only \\(1/4\\) of the hidden channels, which reduces redundant computation and memory usage while efficiently encoding local information. Table 4 shows that PCFN achieves comparable SR performance with less computational overhead compared to ConvFFN [53] and GDFN [45]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.678, + 0.36, + 0.695 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.787, + 0.841 + ], + "angle": 0, + "content": "We have proposed a simple yet effective SMFANet method to solve image superresolution efficiently. As the key component of SMFANet, our SMFA module contains the EASA for non-local information exploration and the LDE for local detail modeling. Moreover, we develop a PCFN to facilitate the fusion of the non-local and local information extracted by the SMFA module both in the channel and spatial dimensions. Extensive experimental results show that the proposed SMFANet family achieve a favorable trade-off between reconstruction performance and computational efficiency." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.418, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SMFANet for Efficient Image Super-Resolution" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.787, + 0.224 + ], + "angle": 0, + "content": "This work has been partly supported by the National Natural Science Foundation of China (Nos. U22B2049, 62272233, 62332010), and the Fundamental Research Funds for the Central Universities (No. 30922010910)." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.248, + 0.323, + 0.264 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.282, + 0.785, + 0.309 + ], + "angle": 0, + "content": "1. Ahn, N., Kang, B., Sohn, K.A.: Fast, accurate, and lightweight super-resolution with cascading residual network. In: ECCV (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.311, + 0.785, + 0.338 + ], + "angle": 0, + "content": "2. Arbeláez, P., Maire, M., Fowlkes, C.C., Malik, J.: Contour detection and hierarchical image segmentation. PAMI 33(5), 898-916 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.34, + 0.785, + 0.379 + ], + "angle": 0, + "content": "3. Bevilacqua, M., Roumy, A., Guillemot, C., line Alberi Morel, M.: Low-complexity single-image super-resolution based on nonnegative neighbor embedding. In: BMVC (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.381, + 0.785, + 0.407 + ], + "angle": 0, + "content": "4. Chen, H., Wang, Y., Guo, T., Xu, C., Deng, Y., Liu, Z., Ma, S., Xu, C., Xu, C., Gao, W.: Pre-trained image processing transformer. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.409, + 0.785, + 0.435 + ], + "angle": 0, + "content": "5. Chen, J., hong Kao, S., He, H., Zhuo, W., Wen, S., Lee, C.H., Chan, S.H.G.: Run, don't walk: Chasing higher flops for faster neural networks. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.437, + 0.785, + 0.463 + ], + "angle": 0, + "content": "6. Chen, X., Wang, X., Zhou, J., Dong, C.: Activating more pixels in image superresolution transformer. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.465, + 0.785, + 0.491 + ], + "angle": 0, + "content": "7. Choi, H., Lee, J., Yang, J.: N-gram in swim transformers for efficient lightweight image super-resolution. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.493, + 0.785, + 0.519 + ], + "angle": 0, + "content": "8. Dai, S., Han, M., Xu, W., Wu, Y., Gong, Y.: Soft edge smoothness prior for alpha channel super resolution. In: CVPR (2007)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.521, + 0.785, + 0.547 + ], + "angle": 0, + "content": "9. Dong, C., Loy, C.C., He, K., Tang, X.: Image super-resolution using deep convolutional networks. PAMI 38(2), 295-307 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.549, + 0.785, + 0.575 + ], + "angle": 0, + "content": "0. Dong, C., Loy, C.C., Tang, X.: Accelerating the super-resolution convolutional neural network. In: ECCV (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.577, + 0.785, + 0.603 + ], + "angle": 0, + "content": "1. Dong, J., Pan, J., Yang, Z., Tang, J.: Multi-scale residual low-pass filter network for image deblurring. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.605, + 0.785, + 0.658 + ], + "angle": 0, + "content": "12. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.66, + 0.785, + 0.687 + ], + "angle": 0, + "content": "13. Gu, J., Dong, C.: Interpreting super-resolution networks with local attribution maps. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.689, + 0.785, + 0.728 + ], + "angle": 0, + "content": "14. Guo, H., Li, J., Dai, T., Ouyang, Z., Ren, X., Xia, S.T.: Mambair: A simple baseline for image restoration with state-space model. arXiv preprint arXiv:2402.15648 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.73, + 0.785, + 0.756 + ], + "angle": 0, + "content": "15. Hendrycks, D., Gimpel, K.: Gaussian error linear units. arXiv preprint arXiv:1606.08415 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.758, + 0.785, + 0.784 + ], + "angle": 0, + "content": "16. Huang, J.B., Singh, A., Ahuja, N.: Single image super-resolution from transformed self-exemplars. In: CVPR (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.786, + 0.785, + 0.812 + ], + "angle": 0, + "content": "17. Hui, Z., Gao, X., Yang, Y., Wang, X.: Lightweight image super-resolution with information multi-distillation network. In: ACM MM (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "18. Kim, J., Lee, J.K., Lee, K.M.: Accurate image super-resolution using very deep convolutional networks. In: CVPR (2016)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.282, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "Zheng et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "19. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. In: ICLR (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.785, + 0.217 + ], + "angle": 0, + "content": "20. Li, A., Zhang, L., Liu, Y., Zhu, C.: Feature modulation transformer: Crossrefinement of global representation via high-frequency prior for image superresolution. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.217, + 0.785, + 0.243 + ], + "angle": 0, + "content": "21. Li, M., Ma, B., Zhang, Y.: Lightweight image super-resolution with pyramid clustering transformer. TCSVT pp. 1-1 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.243, + 0.785, + 0.284 + ], + "angle": 0, + "content": "22. Li, W., Zhou, K., Qi, L., Jiang, N., Lu, J., Jia, J.: LAPAR: Linearly-assembled pixel-adaptive regression network for single image super-resolution and beyond. In: NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.284, + 0.785, + 0.325 + ], + "angle": 0, + "content": "23. Li, Y., Fan, Y., Xiang, X., Demandolx, D., Ranjan, R., Timofte, R., Gool, L.V.: Efficient and explicit modelling of image hierarchies for image restoration. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.325, + 0.785, + 0.353 + ], + "angle": 0, + "content": "24. Li, Z., Liu, Y., Chen, X., Cai, H., Gu, J., Qiao, Y., Dong, C.: Blueprint separable residual network for efficient image super-resolution. In: CVPR Workshops (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.353, + 0.785, + 0.379 + ], + "angle": 0, + "content": "25. Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: SwinIR: Image restoration using swin transformer. In: ICCV Workshops (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.785, + 0.407 + ], + "angle": 0, + "content": "26. Lim, B., Son, S., Kim, H., Nah, S., Lee, K.M.: Enhanced deep residual networks for single image super-resolution. In: CVPR Workshops (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.407, + 0.785, + 0.434 + ], + "angle": 0, + "content": "27. Liu, J., Chen, C., Tang, J., Wu, G.: From coarse to fine: Hierarchical pixel integration for lightweight image super-resolution. In: AAAI (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.434, + 0.785, + 0.461 + ], + "angle": 0, + "content": "28. Liu, J., Tang, J., Wu, G.: Residual feature distillation network for lightweight image super-resolution. In: ECCV Workshops (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.461, + 0.785, + 0.487 + ], + "angle": 0, + "content": "29. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin Transformer: Hierarchical vision transformer using shifted windows. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.487, + 0.785, + 0.514 + ], + "angle": 0, + "content": "30. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. In: ICLR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.514, + 0.785, + 0.542 + ], + "angle": 0, + "content": "31. Lu, Z., Li, J., Liu, H., Huang, C., Zhang, L., Zeng, T.: Transformer for single image super-resolution. In: CVPR Workshops (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.542, + 0.785, + 0.582 + ], + "angle": 0, + "content": "32. Mao, Y., Zhang, N., Wang, Q., Bai, B., Bai, W., Fang, H., Liu, P., Li, M., Yan, S.: Multi-level dispersion residual network for efficient image super-resolution. In: CVPR Workshops (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.582, + 0.785, + 0.61 + ], + "angle": 0, + "content": "33. Matsui, Y., Ito, K., Aramaki, Y., Yamasaki, T., Aizawa, K.: Sketch-based manga retrieval using manga109 dataset. arXiv preprint arXiv:1510.04389 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.61, + 0.709, + 0.623 + ], + "angle": 0, + "content": "34. Park, N., Kim, S.: How do vision transformers work? In: ICLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.623, + 0.785, + 0.663 + ], + "angle": 0, + "content": "35. Shi, W., Caballero, J., Huszar, F., Totz, J., Aitken, A.P., Bishop, R., Rueckert, D., Wang, Z.: Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In: CVPR (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.663, + 0.785, + 0.691 + ], + "angle": 0, + "content": "36. Sun, L., Dong, J., Tang, J., Pan, J.: Spatially-adaptive feature modulation for efficient image super-resolution. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.691, + 0.785, + 0.718 + ], + "angle": 0, + "content": "37. Sun, L., Pan, J., Tang, J.: ShuffleMixer: An efficient convnet for image superresolution. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.718, + 0.785, + 0.759 + ], + "angle": 0, + "content": "38. Timofte, R., Agustsson, E., Van Gool, L., Yang, M.H., Zhang, L., et al.: NTIRE 2017 challenge on single image super-resolution: Methods and results. In: CVPR Workshops (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.785, + 0.786 + ], + "angle": 0, + "content": "39. Timofte, R., DeSmet, V., Van Gool, L.: A+: Adjusted anchored neighborhood regression for fast super-resolution. In: ACCV 2014 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.785, + 0.813 + ], + "angle": 0, + "content": "40. Vaswani, Shazeer, Parmar, Uszkoreit, Jones, Gomez, Kaiser, ukasz, L., Polosukhin, Illia: Attention is all you need. In: NeurIPS (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "41. Wang, H., Chen, X., Ni, B., Liu, Y., Liu, j.: Omni aggregation networks for lightweight image super-resolution. In: CVPR (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.418, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SMFANet for Efficient Image Super-Resolution" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.175 + ], + "angle": 0, + "content": "42. Wang, L., Dong, X., Wang, Y., Ying, X., Lin, Z., An, W., Guo, Y.: Exploring sparsity in image super-resolution for efficient inference. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.177, + 0.788, + 0.217 + ], + "angle": 0, + "content": "43. Wang, W., Xie, E., Li, X., Fan, D.P., Song, K., Liang, D., Lu, T., Luo, P., Shao, L.: Pvtv2: Improved baselines with pyramid vision transformer. Computational Visual Media 8(3), 1-10 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.218, + 0.787, + 0.245 + ], + "angle": 0, + "content": "44. Yang, J., Wright, J., Huang, T.S., Ma, Y.: Image super-resolution via sparse representation. TIP 19(11), 2861-2873 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.246, + 0.786, + 0.273 + ], + "angle": 0, + "content": "45. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.273, + 0.787, + 0.3 + ], + "angle": 0, + "content": "46. Zeyde, R., Elad, M., Protter, M.: On single image scale-up using sparse-representations. In: Curves and Surfaces (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.301, + 0.787, + 0.327 + ], + "angle": 0, + "content": "47. Zhang, A., Ren, W., Liu, Y., Cao, X.: Lightweight image super-resolution with superpixel token interaction. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.328, + 0.787, + 0.356 + ], + "angle": 0, + "content": "48. Zhang, J., Peng, H., Wu, K., Liu, M., Xiao, B., Fu, J., Yuan, L.: Minivit: Compressing vision transformers with weight multiplexing. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.357, + 0.787, + 0.383 + ], + "angle": 0, + "content": "49. Zhang, X., Zeng, H., Guo, S., Zhang, L.: Efficient long-range attention network for image super-resolution. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.384, + 0.787, + 0.411 + ], + "angle": 0, + "content": "50. Zhang, Y., Li, K., Li, K., Wang, L., Zhong, B., Fu, Y.: Image super-resolution using very deep residual channel attention networks. In: ECCV (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.412, + 0.787, + 0.438 + ], + "angle": 0, + "content": "51. Zhang, Y., Tian, Y., Kong, Y., Zhong, B., Fu, Y.: Residual dense network for image super-resolution. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.44, + 0.787, + 0.48 + ], + "angle": 0, + "content": "52. Zhou, L., Cai, H., Gu, J., Li, Z., Liu, Y., Chen, X., Qiao, Y., Dong, C.: Efficient image super-resolution using vast-receptive-field attention. In: ECCV Workshops (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.481, + 0.787, + 0.508 + ], + "angle": 0, + "content": "53. Zhou, Y., Li, Z., Guo, C., Bai, S., Cheng, M., Hou, Q.: Srformer: Permuted self-attention for single image super-resolution. In: ICCV (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.508 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/SMFANet_ A Lightweight Self-Modulation Feature Aggregation Network for Efficient Image Super-Resolution/b8afe796-2e9d-45e7-96aa-cb21a9d8b896_origin.pdf b/2024/SMFANet_ A Lightweight Self-Modulation Feature Aggregation Network for Efficient Image Super-Resolution/b8afe796-2e9d-45e7-96aa-cb21a9d8b896_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..583922df0c4ecbd56b45f03942627558d2fdf5be --- /dev/null +++ b/2024/SMFANet_ A Lightweight Self-Modulation Feature Aggregation Network for Efficient Image Super-Resolution/b8afe796-2e9d-45e7-96aa-cb21a9d8b896_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a61bbc790dfe4847c4d22999579f04d7b3b55866592eb910a1e8df31f8fc2aba +size 9725146 diff --git a/2024/SMFANet_ A Lightweight Self-Modulation Feature Aggregation Network for Efficient Image Super-Resolution/full.md b/2024/SMFANet_ A Lightweight Self-Modulation Feature Aggregation Network for Efficient Image Super-Resolution/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9768c0f6987ee24787111de5047d31cc736eb808 --- /dev/null +++ b/2024/SMFANet_ A Lightweight Self-Modulation Feature Aggregation Network for Efficient Image Super-Resolution/full.md @@ -0,0 +1,418 @@ +# SMFANet: A Lightweight Self-Modulation Feature Aggregation Network for Efficient Image Super-Resolution + +Mingjun Zheng*, Long Sun*, Jiangxin Dong, and Jinshan Pan† + +School of Computer Science and Engineering, Nanjing University of Science and Technology, Nanjing, China {mingjunzheng, cs.longsun, jxdong, jspan}@njust.edu.cn + +Abstract. Transformer-based restoration methods achieve significant performance as the self-attention (SA) of the Transformer can explore non-local information for better high-resolution image reconstruction. However, the key dot-product SA requires substantial computational resources, which limits its application in low-power devices. Moreover, the low-pass nature of the SA mechanism limits its capacity for capturing local details, consequently leading to smooth reconstruction results. To address these issues, we propose a self-modulation feature aggregation (SMFA) module to collaboratively exploit both local and non-local feature interactions for a more accurate reconstruction. Specifically, the SMFA module employs an efficient approximation of self-attention (EASA) branch to model non-local information and uses a local detail estimation (LDE) branch to capture local details. Additionally, we further introduce a partial convolution-based feed-forward network (PCFN) to refine the representative features derived from the SMFA. Extensive experiments show that the proposed SMFANet family achieve a better trade-off between reconstruction performance and computational efficiency on public benchmark datasets. In particular, compared to the $\times 4$ SwinIR-light, SMFANet+ achieves 0.14dB higher performance over five public testsets on average, and $\times 10$ times faster runtime, with only about $43\%$ of the model complexity (e.g., FLOPs). Our source codes and pretrained models are available at: https://github.com/Zheng-MJ/SMFANet. + +Keywords: Efficient image SR $\cdot$ Self-attention $\cdot$ Feature aggregation + +# 1 Introduction + +Single-image super-resolution (SISR) refers to restoring a high-resolution (HR) image from the given degraded low-resolution (LR) counterpart. With the rapid rise of streaming platforms and ultra-high-definition (UHD) content becoming the industrial standard, SISR has attracted significant attention from academia to industry for its ability to provide a pleasurable watching experience at a lower cost of media transmission. Concurrently, the ill-posed nature of SISR + +![](images/78771184017821c5d09f37149b61361ade882cb0edce7faad22089e66e9b8639.jpg) +Fig. 1: Overall performance comparison between our proposed method and other state-of-the-art lightweight methods on Manga109 [33] for $\times 4$ SR. The circle sizes in the left and right figures represent the number of FLOPs of the model and the GPU usage required for inference, respectively. The proposed SMFANet family achieve a better trade-off between computational efficiency and reconstruction performance. + +![](images/10722de5e0c7e0ece6e0ccef1eeef7780e804dc370b6a9a0dc192598a42e660f.jpg) + +task makes it challenging for conventional hand-crafted prior-based [8,39,44], or interpolation-based approaches to decently solve this problem. + +Over the past decade, deep learning has revolutionized the field of SISR. Various convolutional neural networks (CNNs) [1,9,10,17,24,26,35-37,50] have been developed to solve this problem. As the basic operation of CNNs, the convolution operator is translation-invariant and has limited receptive fields, which limits its ability to model non-local information. For better image restoration, CNN-based methods [26, 50, 51] become deeper and larger to enhance the representation ability. These high-capacity networks consume high computational costs, e.g., RCAN [50] has 15.59M parameters with over 400 layers, which cannot be applied to source-limited devices, e.g., smart phones. + +Recently, the vision Transformer (ViT) [12] achieves impressive success on both high-level [29] and low-level vision tasks [4, 7, 21, 25, 27, 41, 47, 49, 53], as the self-attention (SA) mechanism in the ViT can model non-local information effectively. However, the SA mechanism requires high computational resources with massive memory consumption. To reduce the computational cost, window-based self-attention [7, 25, 29, 49], transposed SA [45], and weight multiplexing strategy [48] have been developed to ease the computational burden. However, these still require a long time to learn feature dependencies for image super-resolution. + +Furthermore, recent studies [11, 34] reveal that ViTs favor learning low-frequency components, which leads to smoothed reconstruction results. Thus, these limitations motivate us to think: Is it possible to develop an efficient feature modulation block that explores non-local information in a SA-like manner while modeling local details for efficient image super-resolution? + +To this end, we propose a self-modulation feature aggregation (SMFA) module that contains an efficient approximation of self-attention (EASA) branch to exploit non-local information, and a local detail estimation (LDE) branch to + +model local features. On the EASA branch, we acquire the low-frequency contents through a downsampling operation, calculate the global variance of the input features to modulate the processed low-frequency features, and then use the modulated features to aggregate the input features adaptively. As the EASA prioritizes non-local structure information exploration, we use the LDE branch with convolutional layers to capture local features in parallel. + +In addition, we develop an efficient partial convolution-based feed-forward network (PCFN), which further refines the representative features derived from the SMFA in both spatial and channel dimensions. We formulate the proposed SMFA module and PCFN into an end-to-end trainable network, called SMFANet, to solve SISR. Extensive experiments show that the proposed SMFANet achieves a favorable trade-off between computational efficiency and reconstruction performance (see Figure 1). + +The main contributions are summarized as follows: + +- We develop an efficient SMFA module to extract representative features, where the EASA branch for exploring non-local information and the LDE branch for capturing local features. +- We present a lightweight PCFN to refine the features derived from the SMFA in spatial and channel dimensions further. +- We conduct quantitative and qualitative evaluations for our proposed SMFANet on public benchmark datasets, and the results demonstrate that our method achieves a favorable trade-off between model complexity and reconstruction performance. + +# 2 Related Work + +CNN-based super-resolution. SRCNN [9] first introduces effective end-to-end trainable CNNs to map LR directly to its counterpart HR, with superior performance over previous hand-crafted prior and interpolation methods. Then, FSRCNN [10] and ESPCN [35] adopt a post-upsampling strategy to improve efficiency. VDSR [18] uses global residual learning to solve the training problem of deep networks effectively. Therefore, subsequent CNN-based methods can stack deeper and larger layers to capture more information for better image restoration. EDSR [26] increases the model footprint to 43M with significant performance improvement, and RCAN [50], based on channel attention, builds a model with over 400 layers. However, these large models that require high computational costs make them challenging to deploy in real-world applications. + +To alleviate the computational burden, many CNN-based methods aim to strike a balance between reconstruction performance and model complexity. CARN [1] applies group convolution and cascades an efficient residual network. IMDN [17] adopts the information multi-distillation blocks to split, refine, and aggregate the feature mapping step-by-step, which significantly reduces the model parameters. Furthermore, its improvement work RFDN [28] is the winning solution in the NTIRE 2022 Efficient SR Challenge. BSRN [24] uses blueprint convolutions to reduce computational redundancy significantly. Unlike the above + +approaches that employ progressive feature refinement techniques, recent studies [32, 36, 52] tend to introduce non-local feature modulation to extract representative features for reconstruction. VapSR [52] uses large kernel convolutions in the attention branch to enlarge the receptive field. SAFMN [36] enlarges the receptive field by introducing different down-sampling rates to obtain multi-scale spatial features, and then aggregates these features to generate the modulation map. MDRN [32] employs the parallel implementation of multiple stridden convolutions and polling operations to capture finer global structural information. Although these methods can model non-local feature interactions at a low computational cost and achieve good performance, performing reconstruction using aggregated non-local features only and ignoring local features leads to artifacts in local details (e.g., corners and edges) of the super-resolved image. To avoid this problem, we propose an efficient SMFA module that collaboratively models local and non-local features for more accurate reconstruction. + +ViT-based super-resolution. The vision Transformer (ViT) [12] utilizes the self-attention (SA) mechanism to explore global information for image reconstruction and achieves great success. IPT [4] first introduces a standard vision Transformer to solve the image SR problem. However, the SA mechanism of the ViT consumes huge resources, so various attention variants have been proposed to reduce the computational burden. SwinIR [25] employs window self-attention and outperforms the CNN-based large model [26]. ELAN [49] proposes a group-wise self-attention module and shares the weights to reduce complexity significantly. Moreover, recent studies [11,34] have pointed out that ViTs have the nature of the low-pass filter, resulting in smooth reconstruction results. Therefore, many ViT-based methods [6, 20, 23] utilize convolutional operators to enhance local details for higher-quality image reconstruction. Although these methods integrate the advantages of both convolutional and transformer structures and utilize both local and non-local information to achieve high-quality image reconstruction results, they still rely on window-based attention variants, and feature interactions across windows require much time to perform. Different from these methods, we develop an efficient approximation of self-attention (EASA) branch to model non-local information with much lower model complexity. + +# 3 Proposed Method + +Our goal is to develop a simple yet effective CNN model to collaboratively explore local and non-local feature information for accurate image super-resolution. To this end, we first develop a self-modulation feature aggregation (SMFA) module to effectively exploit representative features, where an efficient approximate self-attention (EASA) module based on feature modulation is used for non-local feature interactions, while an additional local detail estimation (LDE) branch is used for local information exploration. To refine the features generated by the SMFA layer, we further introduce a partial convolution-based feed-forward network (PCFN). We incorporate these components into a unified unit that + +![](images/f26eeaf4b8abd0718cd1dc1bd92e899e54a84fc2317a29b21f021c82af3c16ed.jpg) + +$⑦$ Adaptive Max Pooling +$\bullet$ Element-wise Product +Nearest Up-sampling +$\bigoplus$ Element-wise Addition +$\widehat{\mathbf{S}}$ Channel Split +Variance Calculation +GELU Activation +C Channel Concatenate + +![](images/f5e576f6a8729a2c802f75042e163aaaf05dd20fa8fc1bf78799db84d7b52083.jpg) +Fig.2: Network architecture of the proposed SMFANet. The proposed SMFANet consists of a shallow feature extraction module, feature modulation blocks (FMBs), and a lightweight image reconstruction module. Feature modulation block contains one self-modulation feature aggregation (SMFA) module and one partial convolution-based feed-forward network (PCFN). + +![](images/8bff7d6ed5fed818f7433ebadce750206c0ab98974bd2da416129fc4e952de27.jpg) + +pays more attention to information-rich regions for better reconstruction. In the following, we explain the main ideas of each component in detail. + +# 3.1 Overall architecture + +Figure 2 illustrates the overall architecture of our proposed SMFANet. It takes a low-resolution image as input, and uses a $3 \times 3$ convolutional layer to extract shallow features. The extracted shallow features are then fed to a sequence of feature modulation blocks (FMBs) to produce deep representative features, where the FMB consists of a self-modulation feature aggregation (SMFA) module and a partial convolution-based feed-forward network (PCFN). After the feature modulation block, the representative features are processed by an image reconstruction module to reconstruct the high-quality output. To make the reconstruction module as lightweight as possible, a $3 \times 3$ convolutional layer is used to convert the channel dimensions to a specific size that adjusts to the upsampling ratio, and a PixelShuffle layer [35] is used for enlargement. To facilitate high-frequency information learning, we insert a global residual connection before the image reconstruction module. + +# 3.2 Self-modulation feature aggregation module + +Exploring non-local information is important for image SR reconstruction. Recent ViT-based SR methods [4, 7, 25, 47, 49] utilize various self-attention mechanisms to explore non-local information and achieve impressive reconstruction performance. However, these self-attention variants are computationally expensive and have a limited ability to model local details, as their low-pass filter nature makes them preferentially capture low-frequency information [11, 20, 34]. + +To address this problem, we develop a lightweight self-modulation feature aggregation (SMFA) module to collaboratively model local and non-local features for accurate reconstruction. Within the SMFA module, an efficient approximation of self-attention (EASA) branch is implemented for non-local information exploration at a moderate cost, and a local detail estimation (LDE) branch is used for capturing local information. + +Given the input feature $F_{in} \in \mathbb{R}^{H \times W \times C}$ , where $H \times W$ denotes the spatial size and $C$ is the number of channels, we first apply a $1 \times 1$ convolution to the normalized $F_{in}$ to expand the channel, and then split the channel into two parts as inputs to the EASA and LDE branches: + +$$ +\{X, Y \} = \mathcal {S} \left(\operatorname {C o n v} _ {1 \times 1} \left(\left| \left| F _ {i n} \right| \right| _ {2}\right)\right), \tag {1} +$$ + +where $||\cdot||_2$ is the $L_{2}$ normalization, $Conv_{1\times 1}(\cdot)$ denotes a $1\times 1$ convolutional layer, $S(\cdot)$ denotes a channel splitting operation, $X\in \mathbb{R}^{H\times W\times C}$ and $Y\in \mathbb{R}^{H\times W\times C}$ . We then process the features $X$ and $Y$ in parallel via the EASA and LED branches, producing the non-local feature $X_{l}$ and local feature $Y_{d}$ , respectively. Finally, we fuse $X_{l}$ and $Y_{d}$ together with element-wise addition and feed them into a $1\times 1$ convolution to form a representative output of the SMFA module. This process can be formulated as: + +$$ +F _ {\rho} = \operatorname {C o n v} _ {1 \times 1} \left(X _ {l} + Y _ {d}\right), \tag {2} +$$ + +where $F_{\rho}\in \mathbb{R}^{H\times W\times C}$ is the output feature. + +Efficient approximation of self-attention. We obtain the low-frequency components through a downsampling operation and feed them into a $3 \times 3$ depth-wise convolution to generate non-local structure information $X_{s} \in \mathbb{R}^{H / 8 \times W / 8 \times C}$ : + +$$ +X _ {s} = D W C o n v _ {3 \times 3} (\mathcal {D} (X)), \tag {3} +$$ + +where $\mathcal{D}(\cdot)$ denotes the adaptive max pooling with a scaling factor of 8, $DWConv_{3\times 3}(\cdot)$ is a $3\times 3$ depth-wise convolutional layer. To embed global descriptions for modulating non-local representation $X_{s}$ , we introduce the variance of $X$ as the statistical divergence of spatial information, and merge it with the non-local representation $X_{s}$ through a $1\times 1$ convolution: + +$$ +\sigma^ {2} (X) = \frac {1}{N} \sum_ {i = 0} ^ {N - 1} \left(x _ {i} - \mu\right) ^ {2}, \tag {4} +$$ + +$$ +X _ {m} = \operatorname {C o n v} _ {1 \times 1} \left(X _ {s} + \sigma^ {2} (X)\right), +$$ + +where $\sigma^2 (X)\in \mathbb{R}^{1\times 1\times C}$ is the variance of $X$ , $N$ is the total number of pixels, $x_{i}$ denotes the value of each pixel, $\mu$ is the mean of all pixel values, and $X_{m}\in \mathbb{R}^{H\times W\times C}$ represents the modulated feature. This variance modulation mechanism facilitates better exploring non-local information, and we study it in detail in Section 5. + +Finally, we use the modulated features to aggregate the input feature $X$ for extracting the representative structure information $X_{l}$ : + +$$ +X _ {l} = X \odot \mathcal {U} (\phi (X _ {m})), \tag {5} +$$ + +where $\phi(\cdot)$ refers to the GELU activation function [15], $\mathcal{U}(\cdot)$ denotes a nearest upsampling operation, and $\odot$ represents the element-wise product operation. + +Local detail estimation. Local details are important for the pleasing high-frequency reconstruction. As the EASA prioritizes non-local structure information exploration, we develop a simple local detail estimation layer to capture local features simultaneously. In detail, an expanded depth-wise convolution with a kernel size of $3 \times 3$ is used to encode local information $Y_{h}$ from the input features $Y$ . Then, we use two $1 \times 1$ convolutions with a hidden GELU activation to generate the enhanced local feature $Y_{d}$ , which is achieved by: + +$$ +\begin{array}{l} Y _ {h} = \operatorname {C o n v} _ {1 \times 1} (D W \operatorname {C o n v} _ {3 \times 3} (Y)), \\ Y _ {v} = \operatorname {C o n v} _ {1} (\phi (Y)) \end{array} \tag {6} +$$ + +where $Y_{h}\in \mathbb{R}^{H\times W\times 2C}$ is the encoded local information. + +# 3.3 Partial convolution-based feed-forward network + +The regular feed-forward network (FFN) [12, 40] operates on each pixel location identically, which lacks the exchange of information in the spatial dimension. Inspired by [5, 37], we improve the FFN and propose an efficient partial convolution-based feed-forward network (PCFN) to further refine the representative features derived from the SMFA. + +Figure 2(c) shows that the PCFN uses a $1 \times 1$ convolution with the GELU activation function to perform cross-channel interaction on the expanded hidden space. Then, it splits the hidden features into two chunks of $\{F_{\rho}^{1}, F_{\rho}^{2}\}$ and employs a $3 \times 3$ convolution followed by a GELU activation to process $F_{\rho}^{1}$ to encode local contextual information. The processed $F_{\rho}^{1}$ and $F_{\rho}^{2}$ are then concatenated and fed into a $1 \times 1$ convolution for further feature mixing and reducing the hidden channel back to the original input dimension. This process is defined as: + +$$ +\left\{F _ {\rho} ^ {1}, F _ {\rho} ^ {2} \right\} = \mathcal {S} \left(\phi \left(\operatorname {C o n v} _ {1 \times 1} \left(\left| \left| F _ {\rho} \right| \right| _ {2}\right)\right)\right), \\ \hat {\hat {r}} - C = \left(\mathcal {A} \left[ + (C - \left(F _ {1}\right)) - F _ {2} ^ {2} \right]\right) \tag {7} +$$ + +$$ +\hat {F} _ {\rho} = C o n v _ {1 \times 1} (\mathcal {C} ([ \phi (C o n v _ {3 \times 3} (F _ {\rho} ^ {1})), F _ {\rho} ^ {2} ])), +$$ + +where $F_{\rho}^{1}\in \mathbb{R}^{H\times W\times C / 2}$ and $F_{\rho}^{2}\in \mathbb{R}^{H\times W\times 3C / 2}$ , $\mathcal{C}(\cdot)$ represents a concatenation operation, and $\operatorname{Conv}_{3\times 3}(\cdot)$ denotes a $3\times 3$ convolutional layer. + +# 3.4 Feature modulation block + +As mentioned previously, we formulate the proposed SMFA and PCFN into a feature modulation block (FMB) to produce deep representative features. To stabilize the model training and encourage more information flow, the residual connection is employed. The FMB can be written as: + +$$ +\begin{array}{l} F _ {\rho} = S M F A \left(F _ {i n}\right) + F _ {i n}, \\ \hat {\hat {F}} = \operatorname {R G E N} (F _ {i}) + F _ {i}. \end{array} \tag {8} +$$ + +$$ +\hat {F} _ {\rho} = P C F N (F _ {\rho}) + F _ {\rho}, +$$ + +where $F_{in}$ , $F_{\rho}$ and $\hat{F}_{\rho}$ remain consistent with the previous definition. + +# 4 Experimental Results + +# 4.1 Dataset and implementation + +Datasets. The high-quality DIV2K [38] and DIV2K + Flickr2K (DF2K) [38] datasets are commonly used for training SR models, and previous methods [14,17,25,26,36] chose one of them as the training dataset. To make a fair comparison with recent state-of-the-art methods, we train our proposed method on the DIV2K and DF2K datasets, respectively. In addition, we evaluate our method on the commonly used test datasets, including Set5 [3], Set14 [46], B100 [2], Urban100 [16] and Manga109 [33]. We transform the image into YCbCr color space and calculate the peak signal-to-noise ratio (PSNR) and structure similarity index (SSIM) on the Y channel of the image to evaluate the quality of the recovered image. + +Implementation details. During the training, we randomly crop 64 patches of size $64 \times 64$ with random horizontal flips and rotations from LR images as the basic training inputs. The proposed model is trained with the same loss function as SAFMN [36] and optimized by Adam [19] with $\beta_{1} = 0.9$ and $\beta_{2} = 0.99$ . We set the initial learning rate to $1 \times 10^{-3}$ and the minimum one to $1 \times 10^{-5}$ , which is updated by the Cosine Annealing scheme [30]. The number of iterations for all experiments is set to 1,000,000. All experiments are conducted with the PyTorch framework on an NVIDIA GeForce RTX 3090 GPU. We train two versions of SMFANet with different complexity. The standard SMFANet consists of 8 FMBs with 36 channels while the large version, SMFANet+, has 12 FMBs with 48 channels. Due to the page limit, we include more experimental results in the supplemental material. + +# 4.2 Comparisons with state-of-the-art methods + +Quantitative comparison. To fully evaluate the performance of our method, we first compare our SMFANet with state-of-the-art CNN-based lightweight SR methods, including FSRCNN [10], CARN [1], EDSR-baseline [26], IMDN [17], LAPAR-A [22], SMSR [42], ShuffleMixer [37], SAFMN [36], and further compare the SMFANet+ with recent ViT-based lightweight algorithms, including ESRT [31], SwinIR-light [25], ELAN-light [49], HPINet-S [27], NGswin [7], SPI-N [47] and SRFormer-light [53]. The quantitative comparisons on benchmark datasets for the upscaling factors of $\times 2$ , $\times 3$ , and $\times 4$ are reported in Table 1 and Table 2, where SMFANet (DIV2K) denotes training using the DIV2K dataset, and SMFANet (DF2K) uses DF2K as the training dataset. In addition to PSNR and SSIM metrics, we also list the number of parameters (#Params) and FLOPs (#FLOPs). For a fair comparison, we calculate the model complexity of all evaluated methods with the fvcore library (i.e., fvcore.nn.flop_count_str) under a setting of super resolving an LR image to $1280 \times 720$ pixels. + +Benefiting from the ability to model non-local information of the SMFA module, SMFANet can effectively explore more information compared to previous + +Table 1: Comparison with CNN-based lightweight SR methods on public benchmark datasets. All PSNR/SSIM results are calculated on the Y-channel. #FLOPs is measured corresponding to an HR image of the size $1280 \times 720$ pixels. The best and second-best performances are highlighted in Red and Blue colors. + +
ScaleMethods#Params (K)#FLOPs (G)Set5Set14B100Urban100Manga109
×2FSRCNN [10]12637.00/0.955832.63/0.908831.53/0.892029.88/0.902036.67/0.9694
CARN [1]159222337.76/0.959033.52/0.916632.09/0.897831.92/0.925638.36/0.9765
EDSR-baseline [26]137031637.99/0.960433.57/0.917532.16/0.899431.98/0.927238.54/0.9769
IMDN [17]69416138.00/0.960533.63/0.917732.19/0.899632.17/0.928338.88/0.9774
LAPAR-A [22]54817138.01/0.960533.62/0.918332.19/0.899932.10/0.928338.67/0.9772
SMSR [42]98513238.00/0.960133.64/0.917932.17/0.899932.19/0.928438.76/0.9771
ShuffleMixer [37]3949138.01/0.960633.63/0.918032.17/0.899531.89/0.925738.83/0.9774
SAFMN [36]2285238.00/0.960533.54/0.917732.16/0.899531.84/0.925638.71/0.9771
SMFANet (DIV2K)1864138.04/0.960633.65/0.918632.19/0.899932.22/0.928338.98/0.9776
SMFANet (DF2K)1864138.08/0.960733.65/0.918532.22/0.900232.20/0.928239.11/0.9779
×3FSRCNN [10]12533.16/0.914029.43/0.824228.53/0.791026.43/0.808030.98/0.9212
CARN [1]159211934.29/0.925530.29/0.840729.06/0.803428.06/0.849333.50/0.9440
EDSR-baseline [26]155516034.37/0.927030.28/0.841729.09/0.805228.15/0.852733.45/0.9439
IMDN [17]7037234.36/0.927030.32/0.841729.09/0.804628.17/0.851933.61/0.9445
LAPAR-A [22]59411434.36/0.926730.34/0.841229.11/0.805428.15/0.852333.51/0.9441
SMSR [42]9936834.40/0.927030.33/0.841229.10/0.805028.25/0.853633.68/0.9445
ShuffleMixer [37]4154334.40/0.927230.37/0.842329.12/0.805128.08/0.849833.69/0.9448
SAFMN [36]2332334.34/0.926730.33/0.841829.08/0.804827.95/0.847433.52/0.9437
SMFANet (DIV2K)1911934.46/0.927530.39/0.843229.13/0.805928.25/0.852533.84/0.9454
SMFANet (DF2K)1911934.42/0.927430.41/0.843029.16/0.806528.22/0.852333.96/0.9460
×4FSRCNN [10]12530.71/0.865727.59/0.753526.98/0.715024.62/0.728027.90/0.8517
CARN [1]15929132.13/0.893728.60/0.780627.58/0.734926.07/0.783730.47/0.9084
EDSR-baseline [26]151811432.09/0.893828.58/0.781327.57/0.735726.04/0.784930.35/0.9067
IMDN [17]7154132.21/0.894828.58/0.781127.56/0.735326.04/0.783830.45/0.9075
LAPAR-A [22]6599432.15/0.894428.61/0.781827.61/0.736626.14/0.787130.42/0.9074
SMSR [42]10064232.12/0.893228.55/0.780827.55/0.735126.11/0.786830.54/0.9085
ShuffleMixer [37]4112832.21/0.895328.66/0.782727.61/0.736626.08/0.783530.65/0.9093
SAFMN [36]2401432.18/0.894828.60/0.781327.58/0.735925.97/0.780930.43/0.9063
SMFANet (DIV2K)1971132.25/0.895628.67/0.782527.61/0.737126.19/0.786130.72/0.9097
SMFANet (DF2K)1971132.25/0.895628.71/0.783327.64/0.737726.18/0.786230.82/0.9104
+ +CNN-based methods. Table 1 shows that our SMFANet achieves better performance on almost all benchmark datasets. For instance, on the $\times 4$ Manga109 dataset, SMFANet (DIV2K) outperforms IMDN [17] by $0.27\mathrm{dB}$ with only $27\%$ FLOPs. When training SMFANet on the larger DF2K dataset, the performance gain is further improved by $0.1\mathrm{dB}$ . + +Table 2 shows that the proposed SMFANet+ achieves competitive performance with lower model complexity. Note that ViT-based methods are based on the self-attention mechanism to model long-range dependency, which consumes significant computational costs. In contrast, our SMFA module, with only a few low-consumption operators, efficiently explores non-local information for high-quality reconstruction. On the five benchmark datasets, SMFANet $^+$ (DIV2K) shows similar performance to SwinIR-light [25] with only $43\%$ FLOPs, and SMFANet $^+$ (DF2K) improves the average PSNR gain of the $\times 4$ SR to 0.14dB. These comparisons suggest that our proposed SMFANet can exploit more information than previous CNN-based and ViT-based approaches. + +Qualitative comparisons. We compare the visual results of our proposed SMFANet and SMFANet+ with CNN-based methods and ViT-based methods on the $\times 4$ Urban100 dataset, respectively. The evaluated CNN-based methods [17,22,26,36,37] are listed at the top of Figure 3, which suffer from blurring artifacts and distorted lines. In contrast, our method can restore more accurate + +Table 2: Comparison with ViT-based lightweight SR methods on public benchmark datasets. All measured metrics are calculated in the same way as in Table 1. The best and second-best performances are highlighted in Red and Blue colors. + +
ScaleMethods#Params (K)#FLOPs (G)Set5Set14B100Urban100Manga109
×2ESRT [31]678111638.03/0.960033.75/0.918432.25/0.900132.58/0.931839.12/0.9774
SwinIR-light [25]91024438.14/0.961133.86/0.920632.31/0.901232.76/0.934039.12/0.9783
ELAN-light [49]62120338.17/0.961133.94/0.920732.30/0.901232.76/0.934039.11/0.9782
NGswin [7]99814638.05/0.961033.79/0.919932.27/0.900832.53/0.932438.97/0.9777
SPIN [47]49711438.20/0.961533.90/0.921532.31/0.901532.79/0.934039.18/0.9784
SMFANet+ (DIV2K)48010838.18/0.961133.82/0.920232.28/0.901132.64/0.932339.25/0.9777
SMFANet+ (DF2K)48010838.19/0.961133.92/0.920732.32/0.901532.70/0.933139.46/0.9787
×3ESRT [31]77083534.42/0.926830.43/0.843329.15/0.806328.46/0.857433.95/0.9455
SwinIR-light [25]91811434.62/0.928930.54/0.846329.20/0.808228.66/0.862433.98/0.9478
ELAN-light [49]6299034.61/0.928830.55/0.846329.21/0.808128.69/0.862434.00/0.9478
NGswin [7]10076634.52/0.928230.53/0.845629.19/0.807828.52/0.860333.89/0.9470
SPIN [47]5695834.65/0.929330.57/0.846429.23/0.808928.71/0.862734.24/0.9489
SMFANet+ (DIV2K)4874834.63/0.928530.52/0.845629.23/0.808428.59/0.859434.17/0.9478
SMFANet+ (DF2K)4874834.66/0.929230.57/0.846129.25/0.809028.67/0.861134.45/0.9490
×4ESRT [31]75229832.19/0.894728.69/0.783327.69/0.737926.39/0.796230.75/0.9100
SwinIR-light [25]9306532.44/0.897628.77/0.785827.69/0.740626.47/0.798030.92/0.9151
ELAN-light [49]6405432.43/0.897528.78/0.785827.69/0.740626.54/0.798230.92/0.9150
HPINet-S [27]4638832.47/0.898728.80/0.787227.69/0.741626.59/0.801630.92/0.9143
NGswin [7]10194032.33/0.896328.78/0.785927.66/0.739626.45/0.796330.80/0.9128
SPIN [47]5554232.48/0.898328.80/0.786227.70/0.741526.55/0.799830.98/0.9156
SRFormer-light [53]8736332.51/0.898828.82/0.787227.73/0.742226.67/0.803231.17/0.9165
SMFANet+ (DIV2K)4962832.43/0.897928.77/0.784927.70/0.740026.45/0.794331.06/0.9138
SMFANet+ (DF2K)4962832.51/0.898528.87/0.787227.74/0.741226.56/0.797631.29/0.9163
+ +Table 3: Memory and running time comparisons on $\times 4$ SR. #GPU Mem. denotes the maximum GPU memory consumption during the inference phase, derived with the Pytorch torch.cuda.max_memory_allocated() function. #Avg.Time is the average running time on 500 LR images of $320 \times 180$ pixels. The best and second-best performances are highlighted in Red and Blue colors. + +
TypeMethodsQuality MetricsTypeMethodsQuality Metrics
#GPU Mem. [M]#Avg. Time [ms]#GPU Mem. [M]#Avg. Time [ms]
CNNCARN [1]702.0713.54ViTSwinIR-light [25]342.44177.06
EDSR-baseline [26]507.1315.21ELAN-light [49]241.3442.75
IMDN [17]204.279.64HPINet-S [27]445.92140.45
LAPAR-A [22]1811.4624.40NGswin [7]372.85108.25
ShuffleMixer [37]474.7919.21SPIN [47]441.52798.44
SAFMN [36]65.268.44SRFormer-light [53]320.95145.99
SMFANet93.209.26SMFANet+247.5117.39
+ +parallel patterns with sharper boundaries. The bottom of Figure 3 shows that ViT-based methods [7,25,27,31,49] do not restore the structures well, e.g., strips, while our SMFANet+ recovers an image with better structures. + +Memory and running time comparisons. To fully demonstrate the efficiency of our proposed SMFANet, we further compare our method with the CNN-based [1, 17, 22, 26, 36, 37] and ViT-based [7, 25, 27, 47, 49, 53] methods in terms of GPU memory consumption and inference time on $\times 4$ SR. Table 3 shows the comparisons of GPU memory and inference time. Our SMFANet has faster inference than the listed CNN methods except SAFMN, and the larger SMFANet+ presents a significant efficiency advantage over ViT-based methods. The running time of SMFANet+ is $17.39\mathrm{ms}$ , which is $\times 10$ times faster than SwinIR-light [25] and $\times 6$ times faster than NGswin [7]. Moreover, when compared to ELAN-light [49], which uses a shared group-wise multi-scale self + +![](images/da02855a3f993c548d98bbc00b8f54bb6cb1e80e646c27f5a2e9610bdeccbe51.jpg) + +![](images/3c49c7915166b4ec255f9c3a4c077cb6354f02faa702a56815e458f21e1bf5a8.jpg) + +![](images/bf69c49b83ff8702696c8ffd1088a1aa2d529d5387357074eab9afeccd05d290.jpg) + +![](images/950d668936f5318b57c5a7b816a6bd5d56ec76dc22806b58b0e09bf813b71291.jpg) + +![](images/04dc92e0af253e375ccced95f91bf572b0a5db0fa871ead81ec94c22cd50aaf6.jpg) + +![](images/6ef858126139393627546d8b35daca3cc8bb61cccaebb2daac5f899485beddfb.jpg) +image078 from Urban100 [16] +Fig. 3: Visual comparisons for $\times 4$ SR on the Urban100 dataset. The comparisons between SMFANet and CNN-based methods are shown at the top of the Figure, while the comparisons between SMFANet+ and ViT-based methods are exhibited at the bottom. Our proposed method recovers the image with more accurate structures. + +![](images/94e40c544286f9883005bb3e9d5b47640eee5ddd4bdf1503f32e3b18624e394b.jpg) +(e) ELAN-light [49] + +![](images/388a397b683ee5ea39a9ad7844877948b238e6595dae3baa3b78eb7bedbe8600.jpg) +(f) NGswin [7] + +![](images/d4ec60e93bd9def1e099d492a5a45ada74fde113a9dbac37526c4ee2dcb85c45.jpg) +(g) HPINet-S [27] + +![](images/45cf5cc0d2c73b2a58c747f15cfb22af6f4c100ea6034f686656c25270507f78.jpg) +(h) SMFANet+ + +attention to accelerate model efficiency, the SMFANet+ still has a much faster runtime (17.39ms vs. 42.75ms). These quantitative and qualitative comparison results show that our method obtains even better results against advanced SR models but with significantly less complexity. + +LAM comparisons. The local attribution map (LAM) [13] indicates the significant correlation between the red pixels and the rectangular position patch during the restoration process. We compare the LAM [13] results between our method and the evaluated CNN-based [17, 36, 37] and ViT-based [25, 27, 49] methods in Figure 4, and label the corresponding diffusion index (DI) value below each subgraph, where a larger DI value indicates a wider range of pixels involved. These results demonstrate that the proposed SMFANet can explore more non-local information for accurate image super-resolution. + +# 5 Analysis and Discussions + +In this section, we conduct extensive ablation studies to analyze and evaluate the effect of each component in the proposed SMFANet. We implement all ablation experiments based on the $\times 4$ SMFANet model and train them with the DIV2K [38] dataset for fair comparisons. The quantitative ablation results in Table 4 are measured on the Urban100 [16] and Manga109 [33] datasets. + +Effectiveness of the SMFA. The proposed SMFA module utilizes a parallel structure that absorbs local and non-local feature extraction to enhance reconstruction accuracy. To demonstrate its effectiveness, we first disable the SMFA module to compare it with the baseline SMFANet. Table 4 shows that the PSNR + +![](images/f5a746bf1626f46cbf91b017d7a43fb313bb2ec6f7493ecb62997fcd28db0854.jpg) +image010 from Urban100 + +![](images/0f5b1dd82bb967f95f6bd51237ceb4435dc2d26ca8464f81762de9f2d30130b7.jpg) + +![](images/dcacc3c8257f03960ed1e90eaf8ab19c2000513d218b8dbb952a594ce00aa0ee.jpg) +(a) ShuffleMixer [37] +DI: 3.5808 + +![](images/dda0f3e965ed30a9e3c9660d63f1fa81d93e1e61560a77df6895dbebab4f274a.jpg) + +![](images/67808f8cfb400f680e076f1edaa238597826d14799f37b9992bcd1db5bdc98f7.jpg) +(b) SAFMN [36] +DI: 8.6313 + +![](images/b566e27607b5623d28d3aec9523b1002af3d74e0970ec0f79ab1f4138d3f6cc0.jpg) + +![](images/94f9024109556785e7875fa8367c7cc1ef9ed81758c2869b0df50eb835e00b6b.jpg) +(c) IMDN [17] DI: 10.3650 + +![](images/0a9a975d4a3ce5713532b8a1e825d89d9d8db44e96ac6953d13bb551b66a25ba.jpg) + +![](images/6766065451f11839c234fc266db8f164cc9809bb85fcd4407f4a12476b219984.jpg) +(d) SMFANet DI 12.0899 + +![](images/515868b5a042d5407350a0ec7fa4874227234c83e8a093e67cd87a02bcfebc94.jpg) +img005 from Urban100 +Fig. 4: Comparison of local attribution maps (LAMs) [13] and diffusion indices (DIs) [13]. Our proposed SMFANet family exploit more feature information and reconstructs a more accurate image structure. + +![](images/2c4bc7c6ac7ced408b455bb46a783b4b7a2643cebce02cbca429e7ea55326af7.jpg) + +![](images/0675c69552ca2fb49eafa7c97475c1da7de6d907b7739674d7882494344760bf.jpg) +(a) SwinIR-light [25] +DI: 6.74785 + +![](images/dd5bd54d14477446067fddcb9616c38c604910aadb487a7e51cb7ffa2ac95f70.jpg) + +![](images/df0cc27ce4c00485fba494013181bb95ab3b1a6b1df1c7dffc2aa716d2180e0b.jpg) +b) HPINet-S [27] +DI: 12.8471 + +![](images/55fb970d289cd7538733bc3b9b64ae24725808ecdec4829532b50186ad272984.jpg) + +![](images/e73e425d874198f09c8812aac9044827c29044036b57d102f67cd7cc1fabffae.jpg) +(c) ELAN-light [49] +DI: 18.2614 + +![](images/58f50301a00a053071d401e70befeb31e290eb86fe6990ecfb8d2c104fc2a8ba.jpg) + +![](images/b03a000236565a3a6265e625991add603e6de9b2bde01ea3f747b02944f1ac4d.jpg) +(d) SMFANet+ +DI: 19.6512 + +values drop by 0.65dB and 0.96dB on the Urban100 and Manga109 datasets. These results show the importance of the SMFA. + +In addition, as the proposed SMFA module mainly contains an EASA branch to explore non-local information and a LDE branch to capture local features, we conduct ablation studies with respect to these components to demonstrate their effectiveness on image SR. Without the EASA branch, significant performance drops of $0.25\mathrm{dB}$ and $0.33\mathrm{dB}$ are observed on the Urban100 and Manga109 datasets. Figure 5 (c) shows the corresponding visual results with blurred structure. Moreover, replacing EASA with the window-based SA [25] yields almost the same performance, but the corresponding runtime and GPU consumption are three times higher. These results indicate that the EASA branch can efficiently approximate SA to explore non-local information for accurate structural reconstruction. As for the LDE branch, removing it, the model only achieves the performance of $25.92\mathrm{dB}$ and $30.22\mathrm{dB}$ on the Urban100 and Manga109. Figure 5 (d) shows the reconstructed images with distorted lines, suggesting that the LDE branch contributes to recovering image details. Figure 6 shows the power spectral density maps, which intuitively illustrate the complementary roles of the EASA and LDE. The PSD of $X_{l}$ derived from the EASA shows that the energy density is more concentrated in the center region, whereas the PSD of $Y_{d}$ obtained from the LDE scatters in the surrounding area compared to $F_{in}$ and $X_{l}$ . + +To further validate the effectiveness of this complementary two-branch design, we replace SMFA with either dual LDE branches or dual EASA branches. With these two alternatives, the PSNR values on the Urban100 dataset drop significantly by 0.17dB and 0.25dB, respectively. We also investigated experiments + +![](images/7e30f05a56e3bead604cbf45935f3892a3a768bf12a1b4890a3f8fd230510e35.jpg) + +![](images/f9e7b5c314f67120b84faed2e38318eb817b67af3787bdba545da834af8db9b5.jpg) +(a) HR patch + +![](images/e612a2cd12dafce68be24143128e012f06053988cf6fa273270085c0d2219336.jpg) + +![](images/c0b03589eeb15392652fe753f81fc093dd1e1358b3c251cd82097ce3161a95e5.jpg) +(b) LR patch + +![](images/512fe8f9ca834e2e47dbbb432d7c5f8910114adb985de56bcbce925dcb50b06e.jpg) + +![](images/dc9f0eb35e92e0f035274a9bc793e3a5f76ec97d59250ddfd49b2e2865db03a6.jpg) +(c) w/o EASA + +![](images/f690950930d464221fd546f824e0e2f86753942330f04803ca42daccdf3ab627.jpg) + +![](images/0340d8020ebf862ab88732c223b24031583c2dec748c19fccf0b2a17a5d37133.jpg) +(d) w/o LDE + +![](images/be622c2a3e2bad6b10ec69f7352f7c1616d731b4d2f0e20e26c20ba0796159d1.jpg) + +![](images/da9bdd48e10af41c326fad8f77a23fdbed34e4a3f23191cb1d5523773d052251.jpg) +(e) SMFANet + +![](images/5f56e33f72f35ba4128d3b68639f94d6a1b969598451439768bc7a01f0c4d782.jpg) +Fig. 5: Effectiveness of the SMFA for image super-resolution on the $\times 4$ Urban100 and Manga109 datasets. + +![](images/cab105c5193fe800703ab74b72f66e74252dd0051b6128f248fd08582e83514d.jpg) +HR image + +![](images/b6cda250e48d7fe1d828e067f4aabbadd63de947fe02b64a68d55dffe78001ac.jpg) + +![](images/5c1e44240c5fc96459029bcb7c64204cea7cb866b55ae7b321cd6df0d45f4060.jpg) +Input feature $F_{in}$ + +![](images/4d6917221a1c6b19dfd7a4900ebddac4a2fdad208f88e6de30c949785748b790.jpg) + +![](images/1170d6da30fae62866a76bd910c1a6a6a3a0b90a37cfb6e86dbc86f00abaa4b1.jpg) +$X_{l}$ from EASA + +![](images/678c0309701644ec5974c3936f3895a50db6b37277883a19d7ce177e42428c56.jpg) + +![](images/24721c17aae3917b6c4c9ca66c319cfa2ca111b05d48a4b1ace89e925fb750cf.jpg) +$Y_{d}$ from LDE +Fig. 6: The power spectral density (PSD) visualizations of feature $F_{in}$ , $X_{l}$ , $Y_{d}$ . We perform a periodic shift of the spectrum map such that the low-frequency component is moved to the center. The EASA activates more low-frequency components for feature $X_{l}$ , and the LDE enhances high-frequency representations for feature $Y_{d}$ . + +that sequentially process the EASA and LDE, whose corresponding reconstruction performances are clearly decreased as well. These reduced results prove the effectiveness of such a parallel design. + +Effectiveness of the variance modulation. The variance modulation is used in the EASA branch to facilitate better non-local information exploration. To demonstrate its effectiveness, we remove this operation from the EASA branch. Table 5 presents that the PSNR values drop by $0.06\mathrm{dB}$ and $0.08\mathrm{dB}$ on the Urban100 and Manga109 datasets. For investigating the effect of $\sigma^2 (X)$ , we further replace $\sigma^2 (X)$ with the variance of $X_{s}$ and a learnable parameter $\alpha \in \mathbb{R}^{1\times 1\times C}$ for feature modulation, and their PSNR values on the Urban100 datasets are reduced to 26.12dB and 26.13dB, respectively. These results suggest that the variance of $X_{s}$ or the learnable parameter $\alpha$ does not accurately represent the global statistical divergence of the input feature. + +In addition, variance modulation with element-wise product incurs a 0.05dB and 0.11dB PSNR reduction on Urban100 and Manga109, respectively. This + +Table 4: Ablation experiments for $\times 4$ SMFANet on Urban100 and Manga109 datasets. All measured metrics are calculated in the same way as in Table 1. “A → B” is to replace A with B, and “A → None” is to remove operator A. + +
AblationVariant#Params(K)#FLOPs(G)#GPU Mem. (M)#Avg. Time(ms)Urban100Manga109
BaselineSMFANet1971193.209.2626.19/0.786130.72/0.9097
SMFASMFA → None82556.903.1225.54/0.764629.76/0.8974
EASA → None1831077.317.2825.94/0.778330.39/0.9061
EASA → Window-based SA23716269.6937.5126.20/0.788230.74/0.9105
7722128776.746.9525.92/0.777230.92/0.9032
SMFA → Dual LDE branches2631485.729.6326.02/0.781030.52/0.9076
SMFA → Dual EASA branches143888.459.2825.94/0.778330.29/0.9076
SMFA → "EASA then LDE"1871177.168.9026.04/0.782430.57/0.9077
SMFA → "LDE then EASA"1871172.988.9526.11/0.784130.64/0.9083
Variance ModulationVariance σ2(X) → None1971193.208.8726.13/0.785230.64/0.9091
Variance σ2(X) → σ2(Xs)1971193.209.3126.12/0.783930.60/0.9086
Variance σ2(X) → Learnable parameter α1971193.209.1826.13/0.785330.65/0.9092
Addition → Element-wise Product1971193.209.1026.14/0.784130.66/0.9083
PCFNPCFN → None132792.946.6425.89/0.777030.38/0.9051
PCFN → FFN174993.118.7126.06/0.782530.56/0.9079
PCFN → ConvFFN2131293.1311.0326.19/0.786130.71/0.9097
PCFN → GDFN20412193.4810.8326.22/0.787230.75/0.9101
+ +performance drop is caused by the element-wise product operation suppressing the feature channel with low variance, resulting in a decreased information flow. Effectiveness of the PCFN. The PCFN is used to facilitate the fusion of the representative features extracted by the SMFA module. To demonstrate the effectiveness of this module, we performed ablation experiments with removal of PCFN and replacement of FFN [40], respectively. Table 5 shows that the exclusion of PCFN and the substitution of FFN each leads to a respective decrease in PSNR values of 0.34dB and 0.16dB on the Manga109 [33] dataset, compared to the baseline. + +Recent FFN variants [36, 43, 45, 53] introduce convolution to boost local encoding capability for better reconstruction performance. However, the expanded hidden features contain redundant information, and processing them with convolution directly increases the computational cost. For example, we use GDFN [45] for channel mixing, resulting in a $\times 2$ increase in GPU consumption over PCFN, with only a slight improvement in reconstruction performance. + +Thus, our PCFN introduces a partial convolution that processes only $1/4$ of the hidden channels, which reduces redundant computation and memory usage while efficiently encoding local information. Table 4 shows that PCFN achieves comparable SR performance with less computational overhead compared to ConvFFN [53] and GDFN [45]. + +# 6 Conclusion + +We have proposed a simple yet effective SMFANet method to solve image superresolution efficiently. As the key component of SMFANet, our SMFA module contains the EASA for non-local information exploration and the LDE for local detail modeling. Moreover, we develop a PCFN to facilitate the fusion of the non-local and local information extracted by the SMFA module both in the channel and spatial dimensions. Extensive experimental results show that the proposed SMFANet family achieve a favorable trade-off between reconstruction performance and computational efficiency. + +# Acknowledgements + +This work has been partly supported by the National Natural Science Foundation of China (Nos. U22B2049, 62272233, 62332010), and the Fundamental Research Funds for the Central Universities (No. 30922010910). + +# References + +1. Ahn, N., Kang, B., Sohn, K.A.: Fast, accurate, and lightweight super-resolution with cascading residual network. In: ECCV (2018) +2. Arbeláez, P., Maire, M., Fowlkes, C.C., Malik, J.: Contour detection and hierarchical image segmentation. PAMI 33(5), 898-916 (2011) +3. Bevilacqua, M., Roumy, A., Guillemot, C., line Alberi Morel, M.: Low-complexity single-image super-resolution based on nonnegative neighbor embedding. In: BMVC (2012) +4. Chen, H., Wang, Y., Guo, T., Xu, C., Deng, Y., Liu, Z., Ma, S., Xu, C., Xu, C., Gao, W.: Pre-trained image processing transformer. In: CVPR (2021) +5. Chen, J., hong Kao, S., He, H., Zhuo, W., Wen, S., Lee, C.H., Chan, S.H.G.: Run, don't walk: Chasing higher flops for faster neural networks. In: CVPR (2023) +6. Chen, X., Wang, X., Zhou, J., Dong, C.: Activating more pixels in image superresolution transformer. In: CVPR (2023) +7. Choi, H., Lee, J., Yang, J.: N-gram in swim transformers for efficient lightweight image super-resolution. In: CVPR (2023) +8. Dai, S., Han, M., Xu, W., Wu, Y., Gong, Y.: Soft edge smoothness prior for alpha channel super resolution. In: CVPR (2007) +9. Dong, C., Loy, C.C., He, K., Tang, X.: Image super-resolution using deep convolutional networks. PAMI 38(2), 295-307 (2016) +0. Dong, C., Loy, C.C., Tang, X.: Accelerating the super-resolution convolutional neural network. In: ECCV (2016) +1. Dong, J., Pan, J., Yang, Z., Tang, J.: Multi-scale residual low-pass filter network for image deblurring. In: ICCV (2023) +12. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2021) +13. Gu, J., Dong, C.: Interpreting super-resolution networks with local attribution maps. In: CVPR (2021) +14. Guo, H., Li, J., Dai, T., Ouyang, Z., Ren, X., Xia, S.T.: Mambair: A simple baseline for image restoration with state-space model. arXiv preprint arXiv:2402.15648 (2024) +15. Hendrycks, D., Gimpel, K.: Gaussian error linear units. arXiv preprint arXiv:1606.08415 (2016) +16. Huang, J.B., Singh, A., Ahuja, N.: Single image super-resolution from transformed self-exemplars. In: CVPR (2015) +17. Hui, Z., Gao, X., Yang, Y., Wang, X.: Lightweight image super-resolution with information multi-distillation network. In: ACM MM (2019) +18. Kim, J., Lee, J.K., Lee, K.M.: Accurate image super-resolution using very deep convolutional networks. In: CVPR (2016) + +19. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. In: ICLR (2015) +20. Li, A., Zhang, L., Liu, Y., Zhu, C.: Feature modulation transformer: Crossrefinement of global representation via high-frequency prior for image superresolution. In: ICCV (2023) +21. Li, M., Ma, B., Zhang, Y.: Lightweight image super-resolution with pyramid clustering transformer. TCSVT pp. 1-1 (2023) +22. Li, W., Zhou, K., Qi, L., Jiang, N., Lu, J., Jia, J.: LAPAR: Linearly-assembled pixel-adaptive regression network for single image super-resolution and beyond. In: NeurIPS (2020) +23. Li, Y., Fan, Y., Xiang, X., Demandolx, D., Ranjan, R., Timofte, R., Gool, L.V.: Efficient and explicit modelling of image hierarchies for image restoration. In: CVPR (2023) +24. Li, Z., Liu, Y., Chen, X., Cai, H., Gu, J., Qiao, Y., Dong, C.: Blueprint separable residual network for efficient image super-resolution. In: CVPR Workshops (2022) +25. Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: SwinIR: Image restoration using swin transformer. In: ICCV Workshops (2021) +26. Lim, B., Son, S., Kim, H., Nah, S., Lee, K.M.: Enhanced deep residual networks for single image super-resolution. In: CVPR Workshops (2017) +27. Liu, J., Chen, C., Tang, J., Wu, G.: From coarse to fine: Hierarchical pixel integration for lightweight image super-resolution. In: AAAI (2023) +28. Liu, J., Tang, J., Wu, G.: Residual feature distillation network for lightweight image super-resolution. In: ECCV Workshops (2020) +29. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin Transformer: Hierarchical vision transformer using shifted windows. In: ICCV (2021) +30. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. In: ICLR (2017) +31. Lu, Z., Li, J., Liu, H., Huang, C., Zhang, L., Zeng, T.: Transformer for single image super-resolution. In: CVPR Workshops (2022) +32. Mao, Y., Zhang, N., Wang, Q., Bai, B., Bai, W., Fang, H., Liu, P., Li, M., Yan, S.: Multi-level dispersion residual network for efficient image super-resolution. In: CVPR Workshops (2023) +33. Matsui, Y., Ito, K., Aramaki, Y., Yamasaki, T., Aizawa, K.: Sketch-based manga retrieval using manga109 dataset. arXiv preprint arXiv:1510.04389 (2015) +34. Park, N., Kim, S.: How do vision transformers work? In: ICLR (2022) +35. Shi, W., Caballero, J., Huszar, F., Totz, J., Aitken, A.P., Bishop, R., Rueckert, D., Wang, Z.: Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In: CVPR (2016) +36. Sun, L., Dong, J., Tang, J., Pan, J.: Spatially-adaptive feature modulation for efficient image super-resolution. In: ICCV (2023) +37. Sun, L., Pan, J., Tang, J.: ShuffleMixer: An efficient convnet for image superresolution. In: NeurIPS (2022) +38. Timofte, R., Agustsson, E., Van Gool, L., Yang, M.H., Zhang, L., et al.: NTIRE 2017 challenge on single image super-resolution: Methods and results. In: CVPR Workshops (2017) +39. Timofte, R., DeSmet, V., Van Gool, L.: A+: Adjusted anchored neighborhood regression for fast super-resolution. In: ACCV 2014 (2015) +40. Vaswani, Shazeer, Parmar, Uszkoreit, Jones, Gomez, Kaiser, ukasz, L., Polosukhin, Illia: Attention is all you need. In: NeurIPS (2017) +41. Wang, H., Chen, X., Ni, B., Liu, Y., Liu, j.: Omni aggregation networks for lightweight image super-resolution. In: CVPR (2023) + +42. Wang, L., Dong, X., Wang, Y., Ying, X., Lin, Z., An, W., Guo, Y.: Exploring sparsity in image super-resolution for efficient inference. In: CVPR (2021) +43. Wang, W., Xie, E., Li, X., Fan, D.P., Song, K., Liang, D., Lu, T., Luo, P., Shao, L.: Pvtv2: Improved baselines with pyramid vision transformer. Computational Visual Media 8(3), 1-10 (2022) +44. Yang, J., Wright, J., Huang, T.S., Ma, Y.: Image super-resolution via sparse representation. TIP 19(11), 2861-2873 (2010) +45. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: CVPR (2022) +46. Zeyde, R., Elad, M., Protter, M.: On single image scale-up using sparse-representations. In: Curves and Surfaces (2012) +47. Zhang, A., Ren, W., Liu, Y., Cao, X.: Lightweight image super-resolution with superpixel token interaction. In: ICCV (2023) +48. Zhang, J., Peng, H., Wu, K., Liu, M., Xiao, B., Fu, J., Yuan, L.: Minivit: Compressing vision transformers with weight multiplexing. In: CVPR (2022) +49. Zhang, X., Zeng, H., Guo, S., Zhang, L.: Efficient long-range attention network for image super-resolution. In: ECCV (2022) +50. Zhang, Y., Li, K., Li, K., Wang, L., Zhong, B., Fu, Y.: Image super-resolution using very deep residual channel attention networks. In: ECCV (2018) +51. Zhang, Y., Tian, Y., Kong, Y., Zhong, B., Fu, Y.: Residual dense network for image super-resolution. In: CVPR (2018) +52. Zhou, L., Cai, H., Gu, J., Li, Z., Liu, Y., Chen, X., Qiao, Y., Dong, C.: Efficient image super-resolution using vast-receptive-field attention. In: ECCV Workshops (2022) +53. Zhou, Y., Li, Z., Guo, C., Bai, S., Cheng, M., Hou, Q.: Srformer: Permuted self-attention for single image super-resolution. In: ICCV (2023) \ No newline at end of file diff --git a/2024/SMFANet_ A Lightweight Self-Modulation Feature Aggregation Network for Efficient Image Super-Resolution/images.zip b/2024/SMFANet_ A Lightweight Self-Modulation Feature Aggregation Network for Efficient Image Super-Resolution/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9d7ee080db9ad6b1d41fd18751a6457ab9ef6081 --- /dev/null +++ b/2024/SMFANet_ A Lightweight Self-Modulation Feature Aggregation Network for Efficient Image Super-Resolution/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f79f9fef3f09b5dd26869cac4ce9f8be21acdcfa1db7b79eebe36fb25eae8bfd +size 785422 diff --git a/2024/SMFANet_ A Lightweight Self-Modulation Feature Aggregation Network for Efficient Image Super-Resolution/layout.json b/2024/SMFANet_ A Lightweight Self-Modulation Feature Aggregation Network for Efficient Image Super-Resolution/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..99aaffd37215e26d5696c66510be8c7e7ab612d3 --- /dev/null +++ b/2024/SMFANet_ A Lightweight Self-Modulation Feature Aggregation Network for Efficient Image Super-Resolution/layout.json @@ -0,0 +1,12470 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 159, + 111, + 455, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 111, + 455, + 166 + ], + "spans": [ + { + "bbox": [ + 159, + 111, + 455, + 166 + ], + "type": "text", + "content": "SMFANet: A Lightweight Self-Modulation Feature Aggregation Network for Efficient Image Super-Resolution" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 186, + 447, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 186, + 447, + 200 + ], + "spans": [ + { + "bbox": [ + 165, + 186, + 447, + 200 + ], + "type": "text", + "content": "Mingjun Zheng*, Long Sun*, Jiangxin Dong, and Jinshan Pan†" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 178, + 208, + 436, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 208, + 436, + 242 + ], + "spans": [ + { + "bbox": [ + 178, + 208, + 436, + 242 + ], + "type": "text", + "content": "School of Computer Science and Engineering, Nanjing University of Science and Technology, Nanjing, China {mingjunzheng, cs.longsun, jxdong, jspan}@njust.edu.cn" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 263, + 456, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 263, + 456, + 504 + ], + "spans": [ + { + "bbox": [ + 160, + 263, + 456, + 504 + ], + "type": "text", + "content": "Abstract. Transformer-based restoration methods achieve significant performance as the self-attention (SA) of the Transformer can explore non-local information for better high-resolution image reconstruction. However, the key dot-product SA requires substantial computational resources, which limits its application in low-power devices. Moreover, the low-pass nature of the SA mechanism limits its capacity for capturing local details, consequently leading to smooth reconstruction results. To address these issues, we propose a self-modulation feature aggregation (SMFA) module to collaboratively exploit both local and non-local feature interactions for a more accurate reconstruction. Specifically, the SMFA module employs an efficient approximation of self-attention (EASA) branch to model non-local information and uses a local detail estimation (LDE) branch to capture local details. Additionally, we further introduce a partial convolution-based feed-forward network (PCFN) to refine the representative features derived from the SMFA. Extensive experiments show that the proposed SMFANet family achieve a better trade-off between reconstruction performance and computational efficiency on public benchmark datasets. In particular, compared to the " + }, + { + "bbox": [ + 160, + 263, + 456, + 504 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 160, + 263, + 456, + 504 + ], + "type": "text", + "content": " SwinIR-light, SMFANet+ achieves 0.14dB higher performance over five public testsets on average, and " + }, + { + "bbox": [ + 160, + 263, + 456, + 504 + ], + "type": "inline_equation", + "content": "\\times 10" + }, + { + "bbox": [ + 160, + 263, + 456, + 504 + ], + "type": "text", + "content": " times faster runtime, with only about " + }, + { + "bbox": [ + 160, + 263, + 456, + 504 + ], + "type": "inline_equation", + "content": "43\\%" + }, + { + "bbox": [ + 160, + 263, + 456, + 504 + ], + "type": "text", + "content": " of the model complexity (e.g., FLOPs). Our source codes and pretrained models are available at: https://github.com/Zheng-MJ/SMFANet." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 514, + 443, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 514, + 443, + 525 + ], + "spans": [ + { + "bbox": [ + 160, + 514, + 443, + 525 + ], + "type": "text", + "content": "Keywords: Efficient image SR " + }, + { + "bbox": [ + 160, + 514, + 443, + 525 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 514, + 443, + 525 + ], + "type": "text", + "content": " Self-attention " + }, + { + "bbox": [ + 160, + 514, + 443, + 525 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 514, + 443, + 525 + ], + "type": "text", + "content": " Feature aggregation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 543, + 230, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 543, + 230, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 543, + 230, + 555 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 565, + 482, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 565, + 482, + 638 + ], + "spans": [ + { + "bbox": [ + 130, + 565, + 482, + 638 + ], + "type": "text", + "content": "Single-image super-resolution (SISR) refers to restoring a high-resolution (HR) image from the given degraded low-resolution (LR) counterpart. With the rapid rise of streaming platforms and ultra-high-definition (UHD) content becoming the industrial standard, SISR has attracted significant attention from academia to industry for its ability to provide a pleasurable watching experience at a lower cost of media transmission. Concurrently, the ill-posed nature of SISR" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 142, + 643, + 229, + 653 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 643, + 229, + 653 + ], + "spans": [ + { + "bbox": [ + 142, + 643, + 229, + 653 + ], + "type": "text", + "content": "* Equal contribution" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 142, + 653, + 242, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 653, + 242, + 665 + ], + "spans": [ + { + "bbox": [ + 142, + 653, + 242, + 665 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 145, + 114, + 306, + 236 + ], + "blocks": [ + { + "bbox": [ + 145, + 114, + 306, + 236 + ], + "lines": [ + { + "bbox": [ + 145, + 114, + 306, + 236 + ], + "spans": [ + { + "bbox": [ + 145, + 114, + 306, + 236 + ], + "type": "image", + "image_path": "78771184017821c5d09f37149b61361ade882cb0edce7faad22089e66e9b8639.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 248, + 482, + 304 + ], + "lines": [ + { + "bbox": [ + 130, + 248, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 482, + 304 + ], + "type": "text", + "content": "Fig. 1: Overall performance comparison between our proposed method and other state-of-the-art lightweight methods on Manga109 [33] for " + }, + { + "bbox": [ + 130, + 248, + 482, + 304 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 130, + 248, + 482, + 304 + ], + "type": "text", + "content": " SR. The circle sizes in the left and right figures represent the number of FLOPs of the model and the GPU usage required for inference, respectively. The proposed SMFANet family achieve a better trade-off between computational efficiency and reconstruction performance." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 307, + 114, + 468, + 236 + ], + "blocks": [ + { + "bbox": [ + 307, + 114, + 468, + 236 + ], + "lines": [ + { + "bbox": [ + 307, + 114, + 468, + 236 + ], + "spans": [ + { + "bbox": [ + 307, + 114, + 468, + 236 + ], + "type": "image", + "image_path": "10722de5e0c7e0ece6e0ccef1eeef7780e804dc370b6a9a0dc192598a42e660f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 328, + 479, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 328, + 479, + 352 + ], + "spans": [ + { + "bbox": [ + 130, + 328, + 479, + 352 + ], + "type": "text", + "content": "task makes it challenging for conventional hand-crafted prior-based [8,39,44], or interpolation-based approaches to decently solve this problem." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 353, + 480, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 353, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 353, + 480, + 460 + ], + "type": "text", + "content": "Over the past decade, deep learning has revolutionized the field of SISR. Various convolutional neural networks (CNNs) [1,9,10,17,24,26,35-37,50] have been developed to solve this problem. As the basic operation of CNNs, the convolution operator is translation-invariant and has limited receptive fields, which limits its ability to model non-local information. For better image restoration, CNN-based methods [26, 50, 51] become deeper and larger to enhance the representation ability. These high-capacity networks consume high computational costs, e.g., RCAN [50] has 15.59M parameters with over 400 layers, which cannot be applied to source-limited devices, e.g., smart phones." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 461, + 481, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 481, + 567 + ], + "type": "text", + "content": "Recently, the vision Transformer (ViT) [12] achieves impressive success on both high-level [29] and low-level vision tasks [4, 7, 21, 25, 27, 41, 47, 49, 53], as the self-attention (SA) mechanism in the ViT can model non-local information effectively. However, the SA mechanism requires high computational resources with massive memory consumption. To reduce the computational cost, window-based self-attention [7, 25, 29, 49], transposed SA [45], and weight multiplexing strategy [48] have been developed to ease the computational burden. However, these still require a long time to learn feature dependencies for image super-resolution." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 569, + 481, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 481, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 481, + 629 + ], + "type": "text", + "content": "Furthermore, recent studies [11, 34] reveal that ViTs favor learning low-frequency components, which leads to smoothed reconstruction results. Thus, these limitations motivate us to think: Is it possible to develop an efficient feature modulation block that explores non-local information in a SA-like manner while modeling local details for efficient image super-resolution?" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "To this end, we propose a self-modulation feature aggregation (SMFA) module that contains an efficient approximation of self-attention (EASA) branch to exploit non-local information, and a local detail estimation (LDE) branch to" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "type": "text", + "content": "Zheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 187 + ], + "type": "text", + "content": "model local features. On the EASA branch, we acquire the low-frequency contents through a downsampling operation, calculate the global variance of the input features to modulate the processed low-frequency features, and then use the modulated features to aggregate the input features adaptively. As the EASA prioritizes non-local structure information exploration, we use the LDE branch with convolutional layers to capture local features in parallel." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 481, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 481, + 270 + ], + "type": "text", + "content": "In addition, we develop an efficient partial convolution-based feed-forward network (PCFN), which further refines the representative features derived from the SMFA in both spatial and channel dimensions. We formulate the proposed SMFA module and PCFN into an end-to-end trainable network, called SMFANet, to solve SISR. Extensive experiments show that the proposed SMFANet achieves a favorable trade-off between computational efficiency and reconstruction performance (see Figure 1)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 147, + 272, + 372, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 272, + 372, + 283 + ], + "spans": [ + { + "bbox": [ + 147, + 272, + 372, + 283 + ], + "type": "text", + "content": "The main contributions are summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 290, + 480, + 396 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 290, + 479, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 290, + 479, + 324 + ], + "spans": [ + { + "bbox": [ + 138, + 290, + 479, + 324 + ], + "type": "text", + "content": "- We develop an efficient SMFA module to extract representative features, where the EASA branch for exploring non-local information and the LDE branch for capturing local features." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 326, + 479, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 326, + 479, + 348 + ], + "spans": [ + { + "bbox": [ + 138, + 326, + 479, + 348 + ], + "type": "text", + "content": "- We present a lightweight PCFN to refine the features derived from the SMFA in spatial and channel dimensions further." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 350, + 480, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 350, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 138, + 350, + 480, + 396 + ], + "type": "text", + "content": "- We conduct quantitative and qualitative evaluations for our proposed SMFANet on public benchmark datasets, and the results demonstrate that our method achieves a favorable trade-off between model complexity and reconstruction performance." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 414, + 237, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 237, + 426 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 237, + 426 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 437, + 481, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 481, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 481, + 569 + ], + "type": "text", + "content": "CNN-based super-resolution. SRCNN [9] first introduces effective end-to-end trainable CNNs to map LR directly to its counterpart HR, with superior performance over previous hand-crafted prior and interpolation methods. Then, FSRCNN [10] and ESPCN [35] adopt a post-upsampling strategy to improve efficiency. VDSR [18] uses global residual learning to solve the training problem of deep networks effectively. Therefore, subsequent CNN-based methods can stack deeper and larger layers to capture more information for better image restoration. EDSR [26] increases the model footprint to 43M with significant performance improvement, and RCAN [50], based on channel attention, builds a model with over 400 layers. However, these large models that require high computational costs make them challenging to deploy in real-world applications." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "content": "To alleviate the computational burden, many CNN-based methods aim to strike a balance between reconstruction performance and model complexity. CARN [1] applies group convolution and cascades an efficient residual network. IMDN [17] adopts the information multi-distillation blocks to split, refine, and aggregate the feature mapping step-by-step, which significantly reduces the model parameters. Furthermore, its improvement work RFDN [28] is the winning solution in the NTIRE 2022 Efficient SR Challenge. BSRN [24] uses blueprint convolutions to reduce computational redundancy significantly. Unlike the above" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 255, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 255, + 91, + 447, + 102 + ], + "type": "text", + "content": "SMFANet for Efficient Image Super-Resolution" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "text", + "content": "approaches that employ progressive feature refinement techniques, recent studies [32, 36, 52] tend to introduce non-local feature modulation to extract representative features for reconstruction. VapSR [52] uses large kernel convolutions in the attention branch to enlarge the receptive field. SAFMN [36] enlarges the receptive field by introducing different down-sampling rates to obtain multi-scale spatial features, and then aggregates these features to generate the modulation map. MDRN [32] employs the parallel implementation of multiple stridden convolutions and polling operations to capture finer global structural information. Although these methods can model non-local feature interactions at a low computational cost and achieve good performance, performing reconstruction using aggregated non-local features only and ignoring local features leads to artifacts in local details (e.g., corners and edges) of the super-resolved image. To avoid this problem, we propose an efficient SMFA module that collaboratively models local and non-local features for more accurate reconstruction." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 284, + 482, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 284, + 482, + 501 + ], + "spans": [ + { + "bbox": [ + 130, + 284, + 482, + 501 + ], + "type": "text", + "content": "ViT-based super-resolution. The vision Transformer (ViT) [12] utilizes the self-attention (SA) mechanism to explore global information for image reconstruction and achieves great success. IPT [4] first introduces a standard vision Transformer to solve the image SR problem. However, the SA mechanism of the ViT consumes huge resources, so various attention variants have been proposed to reduce the computational burden. SwinIR [25] employs window self-attention and outperforms the CNN-based large model [26]. ELAN [49] proposes a group-wise self-attention module and shares the weights to reduce complexity significantly. Moreover, recent studies [11,34] have pointed out that ViTs have the nature of the low-pass filter, resulting in smooth reconstruction results. Therefore, many ViT-based methods [6, 20, 23] utilize convolutional operators to enhance local details for higher-quality image reconstruction. Although these methods integrate the advantages of both convolutional and transformer structures and utilize both local and non-local information to achieve high-quality image reconstruction results, they still rely on window-based attention variants, and feature interactions across windows require much time to perform. Different from these methods, we develop an efficient approximation of self-attention (EASA) branch to model non-local information with much lower model complexity." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 525, + 261, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 261, + 539 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 261, + 539 + ], + "type": "text", + "content": "3 Proposed Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "Our goal is to develop a simple yet effective CNN model to collaboratively explore local and non-local feature information for accurate image super-resolution. To this end, we first develop a self-modulation feature aggregation (SMFA) module to effectively exploit representative features, where an efficient approximate self-attention (EASA) module based on feature modulation is used for non-local feature interactions, while an additional local detail estimation (LDE) branch is used for local information exploration. To refine the features generated by the SMFA layer, we further introduce a partial convolution-based feed-forward network (PCFN). We incorporate these components into a unified unit that" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "type": "text", + "content": "Zheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 411, + 161 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 411, + 161 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 411, + 161 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 411, + 161 + ], + "type": "image", + "image_path": "f26eeaf4b8abd0718cd1dc1bd92e899e54a84fc2317a29b21f021c82af3c16ed.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 417, + 118, + 474, + 239 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 417, + 118, + 474, + 126 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 417, + 118, + 474, + 126 + ], + "spans": [ + { + "bbox": [ + 417, + 118, + 474, + 126 + ], + "type": "inline_equation", + "content": "⑦" + }, + { + "bbox": [ + 417, + 118, + 474, + 126 + ], + "type": "text", + "content": " Adaptive Max Pooling" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 418, + 135, + 473, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 418, + 135, + 473, + 143 + ], + "spans": [ + { + "bbox": [ + 418, + 135, + 473, + 143 + ], + "type": "inline_equation", + "content": "\\bullet" + }, + { + "bbox": [ + 418, + 135, + 473, + 143 + ], + "type": "text", + "content": " Element-wise Product" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 418, + 150, + 471, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 418, + 150, + 471, + 159 + ], + "spans": [ + { + "bbox": [ + 418, + 150, + 471, + 159 + ], + "type": "text", + "content": "Nearest Up-sampling" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 418, + 167, + 474, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 418, + 167, + 474, + 175 + ], + "spans": [ + { + "bbox": [ + 418, + 167, + 474, + 175 + ], + "type": "inline_equation", + "content": "\\bigoplus" + }, + { + "bbox": [ + 418, + 167, + 474, + 175 + ], + "type": "text", + "content": " Element-wise Addition" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 418, + 182, + 456, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 418, + 182, + 456, + 191 + ], + "spans": [ + { + "bbox": [ + 418, + 182, + 456, + 191 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{S}}" + }, + { + "bbox": [ + 418, + 182, + 456, + 191 + ], + "type": "text", + "content": " Channel Split" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 418, + 199, + 471, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 418, + 199, + 471, + 206 + ], + "spans": [ + { + "bbox": [ + 418, + 199, + 471, + 206 + ], + "type": "text", + "content": "Variance Calculation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 418, + 215, + 463, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 418, + 215, + 463, + 224 + ], + "spans": [ + { + "bbox": [ + 418, + 215, + 463, + 224 + ], + "type": "text", + "content": "GELU Activation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 418, + 232, + 471, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 418, + 232, + 471, + 239 + ], + "spans": [ + { + "bbox": [ + 418, + 232, + 471, + 239 + ], + "type": "text", + "content": "C Channel Concatenate" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 134, + 162, + 370, + 245 + ], + "blocks": [ + { + "bbox": [ + 134, + 162, + 370, + 245 + ], + "lines": [ + { + "bbox": [ + 134, + 162, + 370, + 245 + ], + "spans": [ + { + "bbox": [ + 134, + 162, + 370, + 245 + ], + "type": "image", + "image_path": "f5e576f6a8729a2c802f75042e163aaaf05dd20fa8fc1bf78799db84d7b52083.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 258, + 480, + 314 + ], + "lines": [ + { + "bbox": [ + 130, + 258, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 480, + 314 + ], + "type": "text", + "content": "Fig.2: Network architecture of the proposed SMFANet. The proposed SMFANet consists of a shallow feature extraction module, feature modulation blocks (FMBs), and a lightweight image reconstruction module. Feature modulation block contains one self-modulation feature aggregation (SMFA) module and one partial convolution-based feed-forward network (PCFN)." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 372, + 163, + 408, + 244 + ], + "blocks": [ + { + "bbox": [ + 372, + 163, + 408, + 244 + ], + "lines": [ + { + "bbox": [ + 372, + 163, + 408, + 244 + ], + "spans": [ + { + "bbox": [ + 372, + 163, + 408, + 244 + ], + "type": "image", + "image_path": "8bff7d6ed5fed818f7433ebadce750206c0ab98974bd2da416129fc4e952de27.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 338, + 480, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 338, + 480, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 480, + 361 + ], + "type": "text", + "content": "pays more attention to information-rich regions for better reconstruction. In the following, we explain the main ideas of each component in detail." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 131, + 380, + 261, + 390 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 380, + 261, + 390 + ], + "spans": [ + { + "bbox": [ + 131, + 380, + 261, + 390 + ], + "type": "text", + "content": "3.1 Overall architecture" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 399, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 482, + 555 + ], + "type": "text", + "content": "Figure 2 illustrates the overall architecture of our proposed SMFANet. It takes a low-resolution image as input, and uses a " + }, + { + "bbox": [ + 130, + 399, + 482, + 555 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 399, + 482, + 555 + ], + "type": "text", + "content": " convolutional layer to extract shallow features. The extracted shallow features are then fed to a sequence of feature modulation blocks (FMBs) to produce deep representative features, where the FMB consists of a self-modulation feature aggregation (SMFA) module and a partial convolution-based feed-forward network (PCFN). After the feature modulation block, the representative features are processed by an image reconstruction module to reconstruct the high-quality output. To make the reconstruction module as lightweight as possible, a " + }, + { + "bbox": [ + 130, + 399, + 482, + 555 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 399, + 482, + 555 + ], + "type": "text", + "content": " convolutional layer is used to convert the channel dimensions to a specific size that adjusts to the upsampling ratio, and a PixelShuffle layer [35] is used for enlargement. To facilitate high-frequency information learning, we insert a global residual connection before the image reconstruction module." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 131, + 573, + 382, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 573, + 382, + 586 + ], + "spans": [ + { + "bbox": [ + 131, + 573, + 382, + 586 + ], + "type": "text", + "content": "3.2 Self-modulation feature aggregation module" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": "Exploring non-local information is important for image SR reconstruction. Recent ViT-based SR methods [4, 7, 25, 47, 49] utilize various self-attention mechanisms to explore non-local information and achieve impressive reconstruction performance. However, these self-attention variants are computationally expensive and have a limited ability to model local details, as their low-pass filter nature makes them preferentially capture low-frequency information [11, 20, 34]." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 255, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 255, + 91, + 447, + 102 + ], + "type": "text", + "content": "SMFANet for Efficient Image Super-Resolution" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "content": "To address this problem, we develop a lightweight self-modulation feature aggregation (SMFA) module to collaboratively model local and non-local features for accurate reconstruction. Within the SMFA module, an efficient approximation of self-attention (EASA) branch is implemented for non-local information exploration at a moderate cost, and a local detail estimation (LDE) branch is used for capturing local information." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 187, + 480, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 187, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 187, + 480, + 236 + ], + "type": "text", + "content": "Given the input feature " + }, + { + "bbox": [ + 130, + 187, + 480, + 236 + ], + "type": "inline_equation", + "content": "F_{in} \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 130, + 187, + 480, + 236 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 187, + 480, + 236 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 130, + 187, + 480, + 236 + ], + "type": "text", + "content": " denotes the spatial size and " + }, + { + "bbox": [ + 130, + 187, + 480, + 236 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 187, + 480, + 236 + ], + "type": "text", + "content": " is the number of channels, we first apply a " + }, + { + "bbox": [ + 130, + 187, + 480, + 236 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 130, + 187, + 480, + 236 + ], + "type": "text", + "content": " convolution to the normalized " + }, + { + "bbox": [ + 130, + 187, + 480, + 236 + ], + "type": "inline_equation", + "content": "F_{in}" + }, + { + "bbox": [ + 130, + 187, + 480, + 236 + ], + "type": "text", + "content": " to expand the channel, and then split the channel into two parts as inputs to the EASA and LDE branches:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 239, + 242, + 480, + 255 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 242, + 480, + 255 + ], + "spans": [ + { + "bbox": [ + 239, + 242, + 480, + 255 + ], + "type": "interline_equation", + "content": "\\{X, Y \\} = \\mathcal {S} \\left(\\operatorname {C o n v} _ {1 \\times 1} \\left(\\left| \\left| F _ {i n} \\right| \\right| _ {2}\\right)\\right), \\tag {1}", + "image_path": "746556bde492439599a7e8190926135191868d9a5442afbb506e402537dda92a.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 259, + 479, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 259, + 479, + 342 + ], + "spans": [ + { + "bbox": [ + 131, + 259, + 479, + 342 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 259, + 479, + 342 + ], + "type": "inline_equation", + "content": "||\\cdot||_2" + }, + { + "bbox": [ + 131, + 259, + 479, + 342 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 131, + 259, + 479, + 342 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 131, + 259, + 479, + 342 + ], + "type": "text", + "content": " normalization, " + }, + { + "bbox": [ + 131, + 259, + 479, + 342 + ], + "type": "inline_equation", + "content": "Conv_{1\\times 1}(\\cdot)" + }, + { + "bbox": [ + 131, + 259, + 479, + 342 + ], + "type": "text", + "content": " denotes a " + }, + { + "bbox": [ + 131, + 259, + 479, + 342 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 131, + 259, + 479, + 342 + ], + "type": "text", + "content": " convolutional layer, " + }, + { + "bbox": [ + 131, + 259, + 479, + 342 + ], + "type": "inline_equation", + "content": "S(\\cdot)" + }, + { + "bbox": [ + 131, + 259, + 479, + 342 + ], + "type": "text", + "content": " denotes a channel splitting operation, " + }, + { + "bbox": [ + 131, + 259, + 479, + 342 + ], + "type": "inline_equation", + "content": "X\\in \\mathbb{R}^{H\\times W\\times C}" + }, + { + "bbox": [ + 131, + 259, + 479, + 342 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 259, + 479, + 342 + ], + "type": "inline_equation", + "content": "Y\\in \\mathbb{R}^{H\\times W\\times C}" + }, + { + "bbox": [ + 131, + 259, + 479, + 342 + ], + "type": "text", + "content": ". We then process the features " + }, + { + "bbox": [ + 131, + 259, + 479, + 342 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 131, + 259, + 479, + 342 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 259, + 479, + 342 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 131, + 259, + 479, + 342 + ], + "type": "text", + "content": " in parallel via the EASA and LED branches, producing the non-local feature " + }, + { + "bbox": [ + 131, + 259, + 479, + 342 + ], + "type": "inline_equation", + "content": "X_{l}" + }, + { + "bbox": [ + 131, + 259, + 479, + 342 + ], + "type": "text", + "content": " and local feature " + }, + { + "bbox": [ + 131, + 259, + 479, + 342 + ], + "type": "inline_equation", + "content": "Y_{d}" + }, + { + "bbox": [ + 131, + 259, + 479, + 342 + ], + "type": "text", + "content": ", respectively. Finally, we fuse " + }, + { + "bbox": [ + 131, + 259, + 479, + 342 + ], + "type": "inline_equation", + "content": "X_{l}" + }, + { + "bbox": [ + 131, + 259, + 479, + 342 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 259, + 479, + 342 + ], + "type": "inline_equation", + "content": "Y_{d}" + }, + { + "bbox": [ + 131, + 259, + 479, + 342 + ], + "type": "text", + "content": " together with element-wise addition and feed them into a " + }, + { + "bbox": [ + 131, + 259, + 479, + 342 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 131, + 259, + 479, + 342 + ], + "type": "text", + "content": " convolution to form a representative output of the SMFA module. This process can be formulated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 254, + 350, + 480, + 361 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 254, + 350, + 480, + 361 + ], + "spans": [ + { + "bbox": [ + 254, + 350, + 480, + 361 + ], + "type": "interline_equation", + "content": "F _ {\\rho} = \\operatorname {C o n v} _ {1 \\times 1} \\left(X _ {l} + Y _ {d}\\right), \\tag {2}", + "image_path": "496a38505f30b1c164b3fe3accc0b679e555a9407065a622403ada78ed1ee863.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 365, + 324, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 365, + 324, + 378 + ], + "spans": [ + { + "bbox": [ + 131, + 365, + 324, + 378 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 365, + 324, + 378 + ], + "type": "inline_equation", + "content": "F_{\\rho}\\in \\mathbb{R}^{H\\times W\\times C}" + }, + { + "bbox": [ + 131, + 365, + 324, + 378 + ], + "type": "text", + "content": " is the output feature." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 379, + 479, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 379, + 479, + 415 + ], + "spans": [ + { + "bbox": [ + 131, + 379, + 479, + 415 + ], + "type": "text", + "content": "Efficient approximation of self-attention. We obtain the low-frequency components through a downsampling operation and feed them into a " + }, + { + "bbox": [ + 131, + 379, + 479, + 415 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 131, + 379, + 479, + 415 + ], + "type": "text", + "content": " depth-wise convolution to generate non-local structure information " + }, + { + "bbox": [ + 131, + 379, + 479, + 415 + ], + "type": "inline_equation", + "content": "X_{s} \\in \\mathbb{R}^{H / 8 \\times W / 8 \\times C}" + }, + { + "bbox": [ + 131, + 379, + 479, + 415 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 248, + 427, + 480, + 438 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 248, + 427, + 480, + 438 + ], + "spans": [ + { + "bbox": [ + 248, + 427, + 480, + 438 + ], + "type": "interline_equation", + "content": "X _ {s} = D W C o n v _ {3 \\times 3} (\\mathcal {D} (X)), \\tag {3}", + "image_path": "e60d8b72b343e8b49b4b32c11f80208f234591f34aecbfe8f3cf843a06401e91.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 441, + 479, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 441, + 479, + 501 + ], + "spans": [ + { + "bbox": [ + 131, + 441, + 479, + 501 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 441, + 479, + 501 + ], + "type": "inline_equation", + "content": "\\mathcal{D}(\\cdot)" + }, + { + "bbox": [ + 131, + 441, + 479, + 501 + ], + "type": "text", + "content": " denotes the adaptive max pooling with a scaling factor of 8, " + }, + { + "bbox": [ + 131, + 441, + 479, + 501 + ], + "type": "inline_equation", + "content": "DWConv_{3\\times 3}(\\cdot)" + }, + { + "bbox": [ + 131, + 441, + 479, + 501 + ], + "type": "text", + "content": " is a " + }, + { + "bbox": [ + 131, + 441, + 479, + 501 + ], + "type": "inline_equation", + "content": "3\\times 3" + }, + { + "bbox": [ + 131, + 441, + 479, + 501 + ], + "type": "text", + "content": " depth-wise convolutional layer. To embed global descriptions for modulating non-local representation " + }, + { + "bbox": [ + 131, + 441, + 479, + 501 + ], + "type": "inline_equation", + "content": "X_{s}" + }, + { + "bbox": [ + 131, + 441, + 479, + 501 + ], + "type": "text", + "content": ", we introduce the variance of " + }, + { + "bbox": [ + 131, + 441, + 479, + 501 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 131, + 441, + 479, + 501 + ], + "type": "text", + "content": " as the statistical divergence of spatial information, and merge it with the non-local representation " + }, + { + "bbox": [ + 131, + 441, + 479, + 501 + ], + "type": "inline_equation", + "content": "X_{s}" + }, + { + "bbox": [ + 131, + 441, + 479, + 501 + ], + "type": "text", + "content": " through a " + }, + { + "bbox": [ + 131, + 441, + 479, + 501 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 131, + 441, + 479, + 501 + ], + "type": "text", + "content": " convolution:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 242, + 508, + 480, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 508, + 480, + 539 + ], + "spans": [ + { + "bbox": [ + 242, + 508, + 480, + 539 + ], + "type": "interline_equation", + "content": "\\sigma^ {2} (X) = \\frac {1}{N} \\sum_ {i = 0} ^ {N - 1} \\left(x _ {i} - \\mu\\right) ^ {2}, \\tag {4}", + "image_path": "9a9006fe056008a6683d601a98503cbfe1d9a4b701356acaa4ce1b97e602f2d4.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 242, + 542, + 372, + 555 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 542, + 372, + 555 + ], + "spans": [ + { + "bbox": [ + 242, + 542, + 372, + 555 + ], + "type": "interline_equation", + "content": "X _ {m} = \\operatorname {C o n v} _ {1 \\times 1} \\left(X _ {s} + \\sigma^ {2} (X)\\right),", + "image_path": "a4f30a1f1d6093918490db722379e9a2d3a59fc8ba9c19c9df1b11bc4649330b.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 559, + 479, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 559, + 479, + 619 + ], + "spans": [ + { + "bbox": [ + 131, + 559, + 479, + 619 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 559, + 479, + 619 + ], + "type": "inline_equation", + "content": "\\sigma^2 (X)\\in \\mathbb{R}^{1\\times 1\\times C}" + }, + { + "bbox": [ + 131, + 559, + 479, + 619 + ], + "type": "text", + "content": " is the variance of " + }, + { + "bbox": [ + 131, + 559, + 479, + 619 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 131, + 559, + 479, + 619 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 559, + 479, + 619 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 131, + 559, + 479, + 619 + ], + "type": "text", + "content": " is the total number of pixels, " + }, + { + "bbox": [ + 131, + 559, + 479, + 619 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 131, + 559, + 479, + 619 + ], + "type": "text", + "content": " denotes the value of each pixel, " + }, + { + "bbox": [ + 131, + 559, + 479, + 619 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 131, + 559, + 479, + 619 + ], + "type": "text", + "content": " is the mean of all pixel values, and " + }, + { + "bbox": [ + 131, + 559, + 479, + 619 + ], + "type": "inline_equation", + "content": "X_{m}\\in \\mathbb{R}^{H\\times W\\times C}" + }, + { + "bbox": [ + 131, + 559, + 479, + 619 + ], + "type": "text", + "content": " represents the modulated feature. This variance modulation mechanism facilitates better exploring non-local information, and we study it in detail in Section 5." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 620, + 479, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 620, + 479, + 644 + ], + "spans": [ + { + "bbox": [ + 131, + 620, + 479, + 644 + ], + "type": "text", + "content": "Finally, we use the modulated features to aggregate the input feature " + }, + { + "bbox": [ + 131, + 620, + 479, + 644 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 131, + 620, + 479, + 644 + ], + "type": "text", + "content": " for extracting the representative structure information " + }, + { + "bbox": [ + 131, + 620, + 479, + 644 + ], + "type": "inline_equation", + "content": "X_{l}" + }, + { + "bbox": [ + 131, + 620, + 479, + 644 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 261, + 652, + 479, + 664 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 652, + 479, + 664 + ], + "spans": [ + { + "bbox": [ + 261, + 652, + 479, + 664 + ], + "type": "interline_equation", + "content": "X _ {l} = X \\odot \\mathcal {U} (\\phi (X _ {m})), \\tag {5}", + "image_path": "e5637f4bd4455c18a5a06e2e423b4d0d254cea2444f5679a9eb7f08f111b1e25.jpg" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "type": "text", + "content": "Zheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\phi(\\cdot)" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " refers to the GELU activation function [15], " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\mathcal{U}(\\cdot)" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " denotes a nearest upsampling operation, and " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " represents the element-wise product operation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 480, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 480, + 224 + ], + "type": "text", + "content": "Local detail estimation. Local details are important for the pleasing high-frequency reconstruction. As the EASA prioritizes non-local structure information exploration, we develop a simple local detail estimation layer to capture local features simultaneously. In detail, an expanded depth-wise convolution with a kernel size of " + }, + { + "bbox": [ + 130, + 140, + 480, + 224 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 140, + 480, + 224 + ], + "type": "text", + "content": " is used to encode local information " + }, + { + "bbox": [ + 130, + 140, + 480, + 224 + ], + "type": "inline_equation", + "content": "Y_{h}" + }, + { + "bbox": [ + 130, + 140, + 480, + 224 + ], + "type": "text", + "content": " from the input features " + }, + { + "bbox": [ + 130, + 140, + 480, + 224 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 130, + 140, + 480, + 224 + ], + "type": "text", + "content": ". Then, we use two " + }, + { + "bbox": [ + 130, + 140, + 480, + 224 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 130, + 140, + 480, + 224 + ], + "type": "text", + "content": " convolutions with a hidden GELU activation to generate the enhanced local feature " + }, + { + "bbox": [ + 130, + 140, + 480, + 224 + ], + "type": "inline_equation", + "content": "Y_{d}" + }, + { + "bbox": [ + 130, + 140, + 480, + 224 + ], + "type": "text", + "content": ", which is achieved by:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 234, + 230, + 480, + 251 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 234, + 230, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 234, + 230, + 480, + 251 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} Y _ {h} = \\operatorname {C o n v} _ {1 \\times 1} (D W \\operatorname {C o n v} _ {3 \\times 3} (Y)), \\\\ Y _ {v} = \\operatorname {C o n v} _ {1} (\\phi (Y)) \\end{array} \\tag {6}", + "image_path": "024ec65f16f4ad12851536e6d465189d6067f6120877302620453207b66d4563.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 262, + 378, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 262, + 378, + 276 + ], + "spans": [ + { + "bbox": [ + 131, + 262, + 378, + 276 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 262, + 378, + 276 + ], + "type": "inline_equation", + "content": "Y_{h}\\in \\mathbb{R}^{H\\times W\\times 2C}" + }, + { + "bbox": [ + 131, + 262, + 378, + 276 + ], + "type": "text", + "content": " is the encoded local information." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 291, + 399, + 302 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 291, + 399, + 302 + ], + "spans": [ + { + "bbox": [ + 131, + 291, + 399, + 302 + ], + "type": "text", + "content": "3.3 Partial convolution-based feed-forward network" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 308, + 479, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 479, + 367 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 479, + 367 + ], + "type": "text", + "content": "The regular feed-forward network (FFN) [12, 40] operates on each pixel location identically, which lacks the exchange of information in the spatial dimension. Inspired by [5, 37], we improve the FFN and propose an efficient partial convolution-based feed-forward network (PCFN) to further refine the representative features derived from the SMFA." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 369, + 480, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 480, + 452 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 480, + 452 + ], + "type": "text", + "content": "Figure 2(c) shows that the PCFN uses a " + }, + { + "bbox": [ + 130, + 369, + 480, + 452 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 130, + 369, + 480, + 452 + ], + "type": "text", + "content": " convolution with the GELU activation function to perform cross-channel interaction on the expanded hidden space. Then, it splits the hidden features into two chunks of " + }, + { + "bbox": [ + 130, + 369, + 480, + 452 + ], + "type": "inline_equation", + "content": "\\{F_{\\rho}^{1}, F_{\\rho}^{2}\\}" + }, + { + "bbox": [ + 130, + 369, + 480, + 452 + ], + "type": "text", + "content": " and employs a " + }, + { + "bbox": [ + 130, + 369, + 480, + 452 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 369, + 480, + 452 + ], + "type": "text", + "content": " convolution followed by a GELU activation to process " + }, + { + "bbox": [ + 130, + 369, + 480, + 452 + ], + "type": "inline_equation", + "content": "F_{\\rho}^{1}" + }, + { + "bbox": [ + 130, + 369, + 480, + 452 + ], + "type": "text", + "content": " to encode local contextual information. The processed " + }, + { + "bbox": [ + 130, + 369, + 480, + 452 + ], + "type": "inline_equation", + "content": "F_{\\rho}^{1}" + }, + { + "bbox": [ + 130, + 369, + 480, + 452 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 369, + 480, + 452 + ], + "type": "inline_equation", + "content": "F_{\\rho}^{2}" + }, + { + "bbox": [ + 130, + 369, + 480, + 452 + ], + "type": "text", + "content": " are then concatenated and fed into a " + }, + { + "bbox": [ + 130, + 369, + 480, + 452 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 130, + 369, + 480, + 452 + ], + "type": "text", + "content": " convolution for further feature mixing and reducing the hidden channel back to the original input dimension. This process is defined as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 216, + 468, + 480, + 489 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 468, + 480, + 489 + ], + "spans": [ + { + "bbox": [ + 216, + 468, + 480, + 489 + ], + "type": "interline_equation", + "content": "\\left\\{F _ {\\rho} ^ {1}, F _ {\\rho} ^ {2} \\right\\} = \\mathcal {S} \\left(\\phi \\left(\\operatorname {C o n v} _ {1 \\times 1} \\left(\\left| \\left| F _ {\\rho} \\right| \\right| _ {2}\\right)\\right)\\right), \\\\ \\hat {\\hat {r}} - C = \\left(\\mathcal {A} \\left[ + (C - \\left(F _ {1}\\right)) - F _ {2} ^ {2} \\right]\\right) \\tag {7}", + "image_path": "47b450120f55dceeef0e0bf49fe459074b513baff7a078a01f35c5697ab38630.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 216, + 484, + 397, + 499 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 484, + 397, + 499 + ], + "spans": [ + { + "bbox": [ + 216, + 484, + 397, + 499 + ], + "type": "interline_equation", + "content": "\\hat {F} _ {\\rho} = C o n v _ {1 \\times 1} (\\mathcal {C} ([ \\phi (C o n v _ {3 \\times 3} (F _ {\\rho} ^ {1})), F _ {\\rho} ^ {2} ])),", + "image_path": "26493df7514d3620ac90f0e8bb92835429af96d1606cc8c11781e11b71ef3e08.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 504, + 479, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 504, + 479, + 530 + ], + "spans": [ + { + "bbox": [ + 130, + 504, + 479, + 530 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 504, + 479, + 530 + ], + "type": "inline_equation", + "content": "F_{\\rho}^{1}\\in \\mathbb{R}^{H\\times W\\times C / 2}" + }, + { + "bbox": [ + 130, + 504, + 479, + 530 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 504, + 479, + 530 + ], + "type": "inline_equation", + "content": "F_{\\rho}^{2}\\in \\mathbb{R}^{H\\times W\\times 3C / 2}" + }, + { + "bbox": [ + 130, + 504, + 479, + 530 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 504, + 479, + 530 + ], + "type": "inline_equation", + "content": "\\mathcal{C}(\\cdot)" + }, + { + "bbox": [ + 130, + 504, + 479, + 530 + ], + "type": "text", + "content": " represents a concatenation operation, and " + }, + { + "bbox": [ + 130, + 504, + 479, + 530 + ], + "type": "inline_equation", + "content": "\\operatorname{Conv}_{3\\times 3}(\\cdot)" + }, + { + "bbox": [ + 130, + 504, + 479, + 530 + ], + "type": "text", + "content": " denotes a " + }, + { + "bbox": [ + 130, + 504, + 479, + 530 + ], + "type": "inline_equation", + "content": "3\\times 3" + }, + { + "bbox": [ + 130, + 504, + 479, + 530 + ], + "type": "text", + "content": " convolutional layer." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 544, + 289, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 544, + 289, + 555 + ], + "spans": [ + { + "bbox": [ + 131, + 544, + 289, + 555 + ], + "type": "text", + "content": "3.4 Feature modulation block" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 562, + 479, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 562, + 479, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 562, + 479, + 609 + ], + "type": "text", + "content": "As mentioned previously, we formulate the proposed SMFA and PCFN into a feature modulation block (FMB) to produce deep representative features. To stabilize the model training and encourage more information flow, the residual connection is employed. The FMB can be written as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 252, + 616, + 480, + 636 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 616, + 480, + 636 + ], + "spans": [ + { + "bbox": [ + 252, + 616, + 480, + 636 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} F _ {\\rho} = S M F A \\left(F _ {i n}\\right) + F _ {i n}, \\\\ \\hat {\\hat {F}} = \\operatorname {R G E N} (F _ {i}) + F _ {i}. \\end{array} \\tag {8}", + "image_path": "aecc5c90d0e6505b4306a21158d4c1e6399e67fe6eaae9cc4cc383484b71b51a.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 252, + 632, + 354, + 645 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 632, + 354, + 645 + ], + "spans": [ + { + "bbox": [ + 252, + 632, + 354, + 645 + ], + "type": "interline_equation", + "content": "\\hat {F} _ {\\rho} = P C F N (F _ {\\rho}) + F _ {\\rho},", + "image_path": "88604a3690426e771edf7966b25be95c7b30c552fe1d72203b9c231867e93b01.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 131, + 652, + 434, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 652, + 434, + 667 + ], + "spans": [ + { + "bbox": [ + 131, + 652, + 434, + 667 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 652, + 434, + 667 + ], + "type": "inline_equation", + "content": "F_{in}" + }, + { + "bbox": [ + 131, + 652, + 434, + 667 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 652, + 434, + 667 + ], + "type": "inline_equation", + "content": "F_{\\rho}" + }, + { + "bbox": [ + 131, + 652, + 434, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 652, + 434, + 667 + ], + "type": "inline_equation", + "content": "\\hat{F}_{\\rho}" + }, + { + "bbox": [ + 131, + 652, + 434, + 667 + ], + "type": "text", + "content": " remain consistent with the previous definition." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 255, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 255, + 91, + 447, + 102 + ], + "type": "text", + "content": "SMFANet for Efficient Image Super-Resolution" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 282, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 282, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 282, + 129 + ], + "type": "text", + "content": "4 Experimental Results" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 143, + 304, + 156 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 143, + 304, + 156 + ], + "spans": [ + { + "bbox": [ + 132, + 143, + 304, + 156 + ], + "type": "text", + "content": "4.1 Dataset and implementation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 165, + 482, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 165, + 482, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 165, + 482, + 285 + ], + "type": "text", + "content": "Datasets. The high-quality DIV2K [38] and DIV2K + Flickr2K (DF2K) [38] datasets are commonly used for training SR models, and previous methods [14,17,25,26,36] chose one of them as the training dataset. To make a fair comparison with recent state-of-the-art methods, we train our proposed method on the DIV2K and DF2K datasets, respectively. In addition, we evaluate our method on the commonly used test datasets, including Set5 [3], Set14 [46], B100 [2], Urban100 [16] and Manga109 [33]. We transform the image into YCbCr color space and calculate the peak signal-to-noise ratio (PSNR) and structure similarity index (SSIM) on the Y channel of the image to evaluate the quality of the recovered image." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 286, + 482, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 286, + 482, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 286, + 482, + 430 + ], + "type": "text", + "content": "Implementation details. During the training, we randomly crop 64 patches of size " + }, + { + "bbox": [ + 130, + 286, + 482, + 430 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 130, + 286, + 482, + 430 + ], + "type": "text", + "content": " with random horizontal flips and rotations from LR images as the basic training inputs. The proposed model is trained with the same loss function as SAFMN [36] and optimized by Adam [19] with " + }, + { + "bbox": [ + 130, + 286, + 482, + 430 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 130, + 286, + 482, + 430 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 286, + 482, + 430 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.99" + }, + { + "bbox": [ + 130, + 286, + 482, + 430 + ], + "type": "text", + "content": ". We set the initial learning rate to " + }, + { + "bbox": [ + 130, + 286, + 482, + 430 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-3}" + }, + { + "bbox": [ + 130, + 286, + 482, + 430 + ], + "type": "text", + "content": " and the minimum one to " + }, + { + "bbox": [ + 130, + 286, + 482, + 430 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-5}" + }, + { + "bbox": [ + 130, + 286, + 482, + 430 + ], + "type": "text", + "content": ", which is updated by the Cosine Annealing scheme [30]. The number of iterations for all experiments is set to 1,000,000. All experiments are conducted with the PyTorch framework on an NVIDIA GeForce RTX 3090 GPU. We train two versions of SMFANet with different complexity. The standard SMFANet consists of 8 FMBs with 36 channels while the large version, SMFANet+, has 12 FMBs with 48 channels. Due to the page limit, we include more experimental results in the supplemental material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 450, + 378, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 378, + 462 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 378, + 462 + ], + "type": "text", + "content": "4.2 Comparisons with state-of-the-art methods" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 473, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 482, + 641 + ], + "type": "text", + "content": "Quantitative comparison. To fully evaluate the performance of our method, we first compare our SMFANet with state-of-the-art CNN-based lightweight SR methods, including FSRCNN [10], CARN [1], EDSR-baseline [26], IMDN [17], LAPAR-A [22], SMSR [42], ShuffleMixer [37], SAFMN [36], and further compare the SMFANet+ with recent ViT-based lightweight algorithms, including ESRT [31], SwinIR-light [25], ELAN-light [49], HPINet-S [27], NGswin [7], SPI-N [47] and SRFormer-light [53]. The quantitative comparisons on benchmark datasets for the upscaling factors of " + }, + { + "bbox": [ + 130, + 473, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\times 2" + }, + { + "bbox": [ + 130, + 473, + 482, + 641 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 473, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\times 3" + }, + { + "bbox": [ + 130, + 473, + 482, + 641 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 473, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 130, + 473, + 482, + 641 + ], + "type": "text", + "content": " are reported in Table 1 and Table 2, where SMFANet (DIV2K) denotes training using the DIV2K dataset, and SMFANet (DF2K) uses DF2K as the training dataset. In addition to PSNR and SSIM metrics, we also list the number of parameters (#Params) and FLOPs (#FLOPs). For a fair comparison, we calculate the model complexity of all evaluated methods with the fvcore library (i.e., fvcore.nn.flop_count_str) under a setting of super resolving an LR image to " + }, + { + "bbox": [ + 130, + 473, + 482, + 641 + ], + "type": "inline_equation", + "content": "1280 \\times 720" + }, + { + "bbox": [ + 130, + 473, + 482, + 641 + ], + "type": "text", + "content": " pixels." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "type": "text", + "content": "Benefiting from the ability to model non-local information of the SMFA module, SMFANet can effectively explore more information compared to previous" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "type": "text", + "content": "Zheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 138, + 168, + 479, + 398 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 1: Comparison with CNN-based lightweight SR methods on public benchmark datasets. All PSNR/SSIM results are calculated on the Y-channel. #FLOPs is measured corresponding to an HR image of the size " + }, + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "inline_equation", + "content": "1280 \\times 720" + }, + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "text", + "content": " pixels. The best and second-best performances are highlighted in Red and Blue colors." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 168, + 479, + 398 + ], + "lines": [ + { + "bbox": [ + 138, + 168, + 479, + 398 + ], + "spans": [ + { + "bbox": [ + 138, + 168, + 479, + 398 + ], + "type": "table", + "html": "
ScaleMethods#Params (K)#FLOPs (G)Set5Set14B100Urban100Manga109
×2FSRCNN [10]12637.00/0.955832.63/0.908831.53/0.892029.88/0.902036.67/0.9694
CARN [1]159222337.76/0.959033.52/0.916632.09/0.897831.92/0.925638.36/0.9765
EDSR-baseline [26]137031637.99/0.960433.57/0.917532.16/0.899431.98/0.927238.54/0.9769
IMDN [17]69416138.00/0.960533.63/0.917732.19/0.899632.17/0.928338.88/0.9774
LAPAR-A [22]54817138.01/0.960533.62/0.918332.19/0.899932.10/0.928338.67/0.9772
SMSR [42]98513238.00/0.960133.64/0.917932.17/0.899932.19/0.928438.76/0.9771
ShuffleMixer [37]3949138.01/0.960633.63/0.918032.17/0.899531.89/0.925738.83/0.9774
SAFMN [36]2285238.00/0.960533.54/0.917732.16/0.899531.84/0.925638.71/0.9771
SMFANet (DIV2K)1864138.04/0.960633.65/0.918632.19/0.899932.22/0.928338.98/0.9776
SMFANet (DF2K)1864138.08/0.960733.65/0.918532.22/0.900232.20/0.928239.11/0.9779
×3FSRCNN [10]12533.16/0.914029.43/0.824228.53/0.791026.43/0.808030.98/0.9212
CARN [1]159211934.29/0.925530.29/0.840729.06/0.803428.06/0.849333.50/0.9440
EDSR-baseline [26]155516034.37/0.927030.28/0.841729.09/0.805228.15/0.852733.45/0.9439
IMDN [17]7037234.36/0.927030.32/0.841729.09/0.804628.17/0.851933.61/0.9445
LAPAR-A [22]59411434.36/0.926730.34/0.841229.11/0.805428.15/0.852333.51/0.9441
SMSR [42]9936834.40/0.927030.33/0.841229.10/0.805028.25/0.853633.68/0.9445
ShuffleMixer [37]4154334.40/0.927230.37/0.842329.12/0.805128.08/0.849833.69/0.9448
SAFMN [36]2332334.34/0.926730.33/0.841829.08/0.804827.95/0.847433.52/0.9437
SMFANet (DIV2K)1911934.46/0.927530.39/0.843229.13/0.805928.25/0.852533.84/0.9454
SMFANet (DF2K)1911934.42/0.927430.41/0.843029.16/0.806528.22/0.852333.96/0.9460
×4FSRCNN [10]12530.71/0.865727.59/0.753526.98/0.715024.62/0.728027.90/0.8517
CARN [1]15929132.13/0.893728.60/0.780627.58/0.734926.07/0.783730.47/0.9084
EDSR-baseline [26]151811432.09/0.893828.58/0.781327.57/0.735726.04/0.784930.35/0.9067
IMDN [17]7154132.21/0.894828.58/0.781127.56/0.735326.04/0.783830.45/0.9075
LAPAR-A [22]6599432.15/0.894428.61/0.781827.61/0.736626.14/0.787130.42/0.9074
SMSR [42]10064232.12/0.893228.55/0.780827.55/0.735126.11/0.786830.54/0.9085
ShuffleMixer [37]4112832.21/0.895328.66/0.782727.61/0.736626.08/0.783530.65/0.9093
SAFMN [36]2401432.18/0.894828.60/0.781327.58/0.735925.97/0.780930.43/0.9063
SMFANet (DIV2K)1971132.25/0.895628.67/0.782527.61/0.737126.19/0.786130.72/0.9097
SMFANet (DF2K)1971132.25/0.895628.71/0.783327.64/0.737726.18/0.786230.82/0.9104
", + "image_path": "b2a3b82cfe2cb70915fca86fff22396e3aa6c95f211a4b940fc42e2fa5cd87e1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 423, + 479, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 479, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 479, + 483 + ], + "type": "text", + "content": "CNN-based methods. Table 1 shows that our SMFANet achieves better performance on almost all benchmark datasets. For instance, on the " + }, + { + "bbox": [ + 130, + 423, + 479, + 483 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 130, + 423, + 479, + 483 + ], + "type": "text", + "content": " Manga109 dataset, SMFANet (DIV2K) outperforms IMDN [17] by " + }, + { + "bbox": [ + 130, + 423, + 479, + 483 + ], + "type": "inline_equation", + "content": "0.27\\mathrm{dB}" + }, + { + "bbox": [ + 130, + 423, + 479, + 483 + ], + "type": "text", + "content": " with only " + }, + { + "bbox": [ + 130, + 423, + 479, + 483 + ], + "type": "inline_equation", + "content": "27\\%" + }, + { + "bbox": [ + 130, + 423, + 479, + 483 + ], + "type": "text", + "content": " FLOPs. When training SMFANet on the larger DF2K dataset, the performance gain is further improved by " + }, + { + "bbox": [ + 130, + 423, + 479, + 483 + ], + "type": "inline_equation", + "content": "0.1\\mathrm{dB}" + }, + { + "bbox": [ + 130, + 423, + 479, + 483 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 484, + 480, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 484, + 480, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 484, + 480, + 604 + ], + "type": "text", + "content": "Table 2 shows that the proposed SMFANet+ achieves competitive performance with lower model complexity. Note that ViT-based methods are based on the self-attention mechanism to model long-range dependency, which consumes significant computational costs. In contrast, our SMFA module, with only a few low-consumption operators, efficiently explores non-local information for high-quality reconstruction. On the five benchmark datasets, SMFANet " + }, + { + "bbox": [ + 130, + 484, + 480, + 604 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 130, + 484, + 480, + 604 + ], + "type": "text", + "content": " (DIV2K) shows similar performance to SwinIR-light [25] with only " + }, + { + "bbox": [ + 130, + 484, + 480, + 604 + ], + "type": "inline_equation", + "content": "43\\%" + }, + { + "bbox": [ + 130, + 484, + 480, + 604 + ], + "type": "text", + "content": " FLOPs, and SMFANet " + }, + { + "bbox": [ + 130, + 484, + 480, + 604 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 130, + 484, + 480, + 604 + ], + "type": "text", + "content": " (DF2K) improves the average PSNR gain of the " + }, + { + "bbox": [ + 130, + 484, + 480, + 604 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 130, + 484, + 480, + 604 + ], + "type": "text", + "content": " SR to 0.14dB. These comparisons suggest that our proposed SMFANet can exploit more information than previous CNN-based and ViT-based approaches." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": "Qualitative comparisons. We compare the visual results of our proposed SMFANet and SMFANet+ with CNN-based methods and ViT-based methods on the " + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": " Urban100 dataset, respectively. The evaluated CNN-based methods [17,22,26,36,37] are listed at the top of Figure 3, which suffer from blurring artifacts and distorted lines. In contrast, our method can restore more accurate" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 255, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 255, + 91, + 447, + 102 + ], + "type": "text", + "content": "SMFANet for Efficient Image Super-Resolution" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 157, + 474, + 332 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 2: Comparison with ViT-based lightweight SR methods on public benchmark datasets. All measured metrics are calculated in the same way as in Table 1. The best and second-best performances are highlighted in Red and Blue colors." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 157, + 474, + 332 + ], + "lines": [ + { + "bbox": [ + 135, + 157, + 474, + 332 + ], + "spans": [ + { + "bbox": [ + 135, + 157, + 474, + 332 + ], + "type": "table", + "html": "
ScaleMethods#Params (K)#FLOPs (G)Set5Set14B100Urban100Manga109
×2ESRT [31]678111638.03/0.960033.75/0.918432.25/0.900132.58/0.931839.12/0.9774
SwinIR-light [25]91024438.14/0.961133.86/0.920632.31/0.901232.76/0.934039.12/0.9783
ELAN-light [49]62120338.17/0.961133.94/0.920732.30/0.901232.76/0.934039.11/0.9782
NGswin [7]99814638.05/0.961033.79/0.919932.27/0.900832.53/0.932438.97/0.9777
SPIN [47]49711438.20/0.961533.90/0.921532.31/0.901532.79/0.934039.18/0.9784
SMFANet+ (DIV2K)48010838.18/0.961133.82/0.920232.28/0.901132.64/0.932339.25/0.9777
SMFANet+ (DF2K)48010838.19/0.961133.92/0.920732.32/0.901532.70/0.933139.46/0.9787
×3ESRT [31]77083534.42/0.926830.43/0.843329.15/0.806328.46/0.857433.95/0.9455
SwinIR-light [25]91811434.62/0.928930.54/0.846329.20/0.808228.66/0.862433.98/0.9478
ELAN-light [49]6299034.61/0.928830.55/0.846329.21/0.808128.69/0.862434.00/0.9478
NGswin [7]10076634.52/0.928230.53/0.845629.19/0.807828.52/0.860333.89/0.9470
SPIN [47]5695834.65/0.929330.57/0.846429.23/0.808928.71/0.862734.24/0.9489
SMFANet+ (DIV2K)4874834.63/0.928530.52/0.845629.23/0.808428.59/0.859434.17/0.9478
SMFANet+ (DF2K)4874834.66/0.929230.57/0.846129.25/0.809028.67/0.861134.45/0.9490
×4ESRT [31]75229832.19/0.894728.69/0.783327.69/0.737926.39/0.796230.75/0.9100
SwinIR-light [25]9306532.44/0.897628.77/0.785827.69/0.740626.47/0.798030.92/0.9151
ELAN-light [49]6405432.43/0.897528.78/0.785827.69/0.740626.54/0.798230.92/0.9150
HPINet-S [27]4638832.47/0.898728.80/0.787227.69/0.741626.59/0.801630.92/0.9143
NGswin [7]10194032.33/0.896328.78/0.785927.66/0.739626.45/0.796330.80/0.9128
SPIN [47]5554232.48/0.898328.80/0.786227.70/0.741526.55/0.799830.98/0.9156
SRFormer-light [53]8736332.51/0.898828.82/0.787227.73/0.742226.67/0.803231.17/0.9165
SMFANet+ (DIV2K)4962832.43/0.897928.77/0.784927.70/0.740026.45/0.794331.06/0.9138
SMFANet+ (DF2K)4962832.51/0.898528.87/0.787227.74/0.741226.56/0.797631.29/0.9163
", + "image_path": "f030380f25b420aefb339c43f6d901135ab2ef8b520ec103c8792f90c4a18da5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 138, + 407, + 476, + 487 + ], + "blocks": [ + { + "bbox": [ + 132, + 342, + 482, + 399 + ], + "lines": [ + { + "bbox": [ + 132, + 342, + 482, + 399 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 482, + 399 + ], + "type": "text", + "content": "Table 3: Memory and running time comparisons on " + }, + { + "bbox": [ + 132, + 342, + 482, + 399 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 132, + 342, + 482, + 399 + ], + "type": "text", + "content": " SR. #GPU Mem. denotes the maximum GPU memory consumption during the inference phase, derived with the Pytorch torch.cuda.max_memory_allocated() function. #Avg.Time is the average running time on 500 LR images of " + }, + { + "bbox": [ + 132, + 342, + 482, + 399 + ], + "type": "inline_equation", + "content": "320 \\times 180" + }, + { + "bbox": [ + 132, + 342, + 482, + 399 + ], + "type": "text", + "content": " pixels. The best and second-best performances are highlighted in Red and Blue colors." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 407, + 476, + 487 + ], + "lines": [ + { + "bbox": [ + 138, + 407, + 476, + 487 + ], + "spans": [ + { + "bbox": [ + 138, + 407, + 476, + 487 + ], + "type": "table", + "html": "
TypeMethodsQuality MetricsTypeMethodsQuality Metrics
#GPU Mem. [M]#Avg. Time [ms]#GPU Mem. [M]#Avg. Time [ms]
CNNCARN [1]702.0713.54ViTSwinIR-light [25]342.44177.06
EDSR-baseline [26]507.1315.21ELAN-light [49]241.3442.75
IMDN [17]204.279.64HPINet-S [27]445.92140.45
LAPAR-A [22]1811.4624.40NGswin [7]372.85108.25
ShuffleMixer [37]474.7919.21SPIN [47]441.52798.44
SAFMN [36]65.268.44SRFormer-light [53]320.95145.99
SMFANet93.209.26SMFANet+247.5117.39
", + "image_path": "36f84ab653d4e200fbc05b9a8c3bd5e71b1f19bf6154f2f64ac830bbc1d0538d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 510, + 481, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 510, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 510, + 481, + 544 + ], + "type": "text", + "content": "parallel patterns with sharper boundaries. The bottom of Figure 3 shows that ViT-based methods [7,25,27,31,49] do not restore the structures well, e.g., strips, while our SMFANet+ recovers an image with better structures." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 482, + 666 + ], + "type": "text", + "content": "Memory and running time comparisons. To fully demonstrate the efficiency of our proposed SMFANet, we further compare our method with the CNN-based [1, 17, 22, 26, 36, 37] and ViT-based [7, 25, 27, 47, 49, 53] methods in terms of GPU memory consumption and inference time on " + }, + { + "bbox": [ + 132, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 132, + 545, + 482, + 666 + ], + "type": "text", + "content": " SR. Table 3 shows the comparisons of GPU memory and inference time. Our SMFANet has faster inference than the listed CNN methods except SAFMN, and the larger SMFANet+ presents a significant efficiency advantage over ViT-based methods. The running time of SMFANet+ is " + }, + { + "bbox": [ + 132, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "17.39\\mathrm{ms}" + }, + { + "bbox": [ + 132, + 545, + 482, + 666 + ], + "type": "text", + "content": ", which is " + }, + { + "bbox": [ + 132, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\times 10" + }, + { + "bbox": [ + 132, + 545, + 482, + 666 + ], + "type": "text", + "content": " times faster than SwinIR-light [25] and " + }, + { + "bbox": [ + 132, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\times 6" + }, + { + "bbox": [ + 132, + 545, + 482, + 666 + ], + "type": "text", + "content": " times faster than NGswin [7]. Moreover, when compared to ELAN-light [49], which uses a shared group-wise multi-scale self" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "type": "text", + "content": "Zheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 115, + 242, + 195 + ], + "blocks": [ + { + "bbox": [ + 137, + 115, + 242, + 195 + ], + "lines": [ + { + "bbox": [ + 137, + 115, + 242, + 195 + ], + "spans": [ + { + "bbox": [ + 137, + 115, + 242, + 195 + ], + "type": "image", + "image_path": "da02855a3f993c548d98bbc00b8f54bb6cb1e80e646c27f5a2e9610bdeccbe51.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 243, + 114, + 302, + 195 + ], + "blocks": [ + { + "bbox": [ + 243, + 114, + 302, + 195 + ], + "lines": [ + { + "bbox": [ + 243, + 114, + 302, + 195 + ], + "spans": [ + { + "bbox": [ + 243, + 114, + 302, + 195 + ], + "type": "image", + "image_path": "3c49c7915166b4ec255f9c3a4c077cb6354f02faa702a56815e458f21e1bf5a8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 306, + 114, + 365, + 195 + ], + "blocks": [ + { + "bbox": [ + 306, + 114, + 365, + 195 + ], + "lines": [ + { + "bbox": [ + 306, + 114, + 365, + 195 + ], + "spans": [ + { + "bbox": [ + 306, + 114, + 365, + 195 + ], + "type": "image", + "image_path": "bf69c49b83ff8702696c8ffd1088a1aa2d529d5387357074eab9afeccd05d290.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 369, + 114, + 429, + 195 + ], + "blocks": [ + { + "bbox": [ + 369, + 114, + 429, + 195 + ], + "lines": [ + { + "bbox": [ + 369, + 114, + 429, + 195 + ], + "spans": [ + { + "bbox": [ + 369, + 114, + 429, + 195 + ], + "type": "image", + "image_path": "950d668936f5318b57c5a7b816a6bd5d56ec76dc22806b58b0e09bf813b71291.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 432, + 114, + 492, + 195 + ], + "blocks": [ + { + "bbox": [ + 432, + 114, + 492, + 195 + ], + "lines": [ + { + "bbox": [ + 432, + 114, + 492, + 195 + ], + "spans": [ + { + "bbox": [ + 432, + 114, + 492, + 195 + ], + "type": "image", + "image_path": "04dc92e0af253e375ccced95f91bf572b0a5db0fa871ead81ec94c22cd50aaf6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 137, + 195, + 242, + 269 + ], + "blocks": [ + { + "bbox": [ + 137, + 195, + 242, + 269 + ], + "lines": [ + { + "bbox": [ + 137, + 195, + 242, + 269 + ], + "spans": [ + { + "bbox": [ + 137, + 195, + 242, + 269 + ], + "type": "image", + "image_path": "6ef858126139393627546d8b35daca3cc8bb61cccaebb2daac5f899485beddfb.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 144, + 268, + 233, + 274 + ], + "lines": [ + { + "bbox": [ + 144, + 268, + 233, + 274 + ], + "spans": [ + { + "bbox": [ + 144, + 268, + 233, + 274 + ], + "type": "text", + "content": "image078 from Urban100 [16]" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 293, + 482, + 338 + ], + "lines": [ + { + "bbox": [ + 131, + 293, + 482, + 338 + ], + "spans": [ + { + "bbox": [ + 131, + 293, + 482, + 338 + ], + "type": "text", + "content": "Fig. 3: Visual comparisons for " + }, + { + "bbox": [ + 131, + 293, + 482, + 338 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 131, + 293, + 482, + 338 + ], + "type": "text", + "content": " SR on the Urban100 dataset. The comparisons between SMFANet and CNN-based methods are shown at the top of the Figure, while the comparisons between SMFANet+ and ViT-based methods are exhibited at the bottom. Our proposed method recovers the image with more accurate structures." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 243, + 195, + 302, + 269 + ], + "blocks": [ + { + "bbox": [ + 243, + 195, + 302, + 269 + ], + "lines": [ + { + "bbox": [ + 243, + 195, + 302, + 269 + ], + "spans": [ + { + "bbox": [ + 243, + 195, + 302, + 269 + ], + "type": "image", + "image_path": "94e40c544286f9883005bb3e9d5b47640eee5ddd4bdf1503f32e3b18624e394b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 244, + 268, + 302, + 274 + ], + "lines": [ + { + "bbox": [ + 244, + 268, + 302, + 274 + ], + "spans": [ + { + "bbox": [ + 244, + 268, + 302, + 274 + ], + "type": "text", + "content": "(e) ELAN-light [49]" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 306, + 195, + 365, + 269 + ], + "blocks": [ + { + "bbox": [ + 306, + 195, + 365, + 269 + ], + "lines": [ + { + "bbox": [ + 306, + 195, + 365, + 269 + ], + "spans": [ + { + "bbox": [ + 306, + 195, + 365, + 269 + ], + "type": "image", + "image_path": "388a397b683ee5ea39a9ad7844877948b238e6595dae3baa3b78eb7bedbe8600.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 314, + 268, + 358, + 274 + ], + "lines": [ + { + "bbox": [ + 314, + 268, + 358, + 274 + ], + "spans": [ + { + "bbox": [ + 314, + 268, + 358, + 274 + ], + "type": "text", + "content": "(f) NGswin [7]" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 369, + 195, + 429, + 269 + ], + "blocks": [ + { + "bbox": [ + 369, + 195, + 429, + 269 + ], + "lines": [ + { + "bbox": [ + 369, + 195, + 429, + 269 + ], + "spans": [ + { + "bbox": [ + 369, + 195, + 429, + 269 + ], + "type": "image", + "image_path": "d4ec60e93bd9def1e099d492a5a45ada74fde113a9dbac37526c4ee2dcb85c45.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 372, + 268, + 425, + 274 + ], + "lines": [ + { + "bbox": [ + 372, + 268, + 425, + 274 + ], + "spans": [ + { + "bbox": [ + 372, + 268, + 425, + 274 + ], + "type": "text", + "content": "(g) HPINet-S [27]" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 432, + 195, + 492, + 269 + ], + "blocks": [ + { + "bbox": [ + 432, + 195, + 492, + 269 + ], + "lines": [ + { + "bbox": [ + 432, + 195, + 492, + 269 + ], + "spans": [ + { + "bbox": [ + 432, + 195, + 492, + 269 + ], + "type": "image", + "image_path": "45cf5cc0d2c73b2a58c747f15cfb22af6f4c100ea6034f686656c25270507f78.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 440, + 268, + 485, + 274 + ], + "lines": [ + { + "bbox": [ + 440, + 268, + 485, + 274 + ], + "spans": [ + { + "bbox": [ + 440, + 268, + 485, + 274 + ], + "type": "text", + "content": "(h) SMFANet+" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 363, + 480, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 363, + 480, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 363, + 480, + 411 + ], + "type": "text", + "content": "attention to accelerate model efficiency, the SMFANet+ still has a much faster runtime (17.39ms vs. 42.75ms). These quantitative and qualitative comparison results show that our method obtains even better results against advanced SR models but with significantly less complexity." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 411, + 480, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 480, + 507 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 480, + 507 + ], + "type": "text", + "content": "LAM comparisons. The local attribution map (LAM) [13] indicates the significant correlation between the red pixels and the rectangular position patch during the restoration process. We compare the LAM [13] results between our method and the evaluated CNN-based [17, 36, 37] and ViT-based [25, 27, 49] methods in Figure 4, and label the corresponding diffusion index (DI) value below each subgraph, where a larger DI value indicates a wider range of pixels involved. These results demonstrate that the proposed SMFANet can explore more non-local information for accurate image super-resolution." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 529, + 302, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 529, + 302, + 542 + ], + "spans": [ + { + "bbox": [ + 132, + 529, + 302, + 542 + ], + "type": "text", + "content": "5 Analysis and Discussions" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 557, + 480, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 480, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 480, + 616 + ], + "type": "text", + "content": "In this section, we conduct extensive ablation studies to analyze and evaluate the effect of each component in the proposed SMFANet. We implement all ablation experiments based on the " + }, + { + "bbox": [ + 130, + 557, + 480, + 616 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 130, + 557, + 480, + 616 + ], + "type": "text", + "content": " SMFANet model and train them with the DIV2K [38] dataset for fair comparisons. The quantitative ablation results in Table 4 are measured on the Urban100 [16] and Manga109 [33] datasets." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "content": "Effectiveness of the SMFA. The proposed SMFA module utilizes a parallel structure that absorbs local and non-local feature extraction to enhance reconstruction accuracy. To demonstrate its effectiveness, we first disable the SMFA module to compare it with the baseline SMFANet. Table 4 shows that the PSNR" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 255, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 255, + 91, + 447, + 102 + ], + "type": "text", + "content": "SMFANet for Efficient Image Super-Resolution" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 115, + 242, + 190 + ], + "blocks": [ + { + "bbox": [ + 137, + 115, + 242, + 190 + ], + "lines": [ + { + "bbox": [ + 137, + 115, + 242, + 190 + ], + "spans": [ + { + "bbox": [ + 137, + 115, + 242, + 190 + ], + "type": "image", + "image_path": "f5a746bf1626f46cbf91b017d7a43fb313bb2ec6f7493ecb62997fcd28db0854.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 150, + 190, + 227, + 196 + ], + "lines": [ + { + "bbox": [ + 150, + 190, + 227, + 196 + ], + "spans": [ + { + "bbox": [ + 150, + 190, + 227, + 196 + ], + "type": "text", + "content": "image010 from Urban100" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 247, + 116, + 303, + 152 + ], + "blocks": [ + { + "bbox": [ + 247, + 116, + 303, + 152 + ], + "lines": [ + { + "bbox": [ + 247, + 116, + 303, + 152 + ], + "spans": [ + { + "bbox": [ + 247, + 116, + 303, + 152 + ], + "type": "image", + "image_path": "0f5b1dd82bb967f95f6bd51237ceb4435dc2d26ca8464f81762de9f2d30130b7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 247, + 152, + 303, + 189 + ], + "blocks": [ + { + "bbox": [ + 247, + 152, + 303, + 189 + ], + "lines": [ + { + "bbox": [ + 247, + 152, + 303, + 189 + ], + "spans": [ + { + "bbox": [ + 247, + 152, + 303, + 189 + ], + "type": "image", + "image_path": "dcacc3c8257f03960ed1e90eaf8ab19c2000513d218b8dbb952a594ce00aa0ee.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 244, + 189, + 306, + 202 + ], + "lines": [ + { + "bbox": [ + 244, + 189, + 306, + 202 + ], + "spans": [ + { + "bbox": [ + 244, + 189, + 306, + 202 + ], + "type": "text", + "content": "(a) ShuffleMixer [37] \nDI: 3.5808" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 309, + 115, + 364, + 152 + ], + "blocks": [ + { + "bbox": [ + 309, + 115, + 364, + 152 + ], + "lines": [ + { + "bbox": [ + 309, + 115, + 364, + 152 + ], + "spans": [ + { + "bbox": [ + 309, + 115, + 364, + 152 + ], + "type": "image", + "image_path": "dda0f3e965ed30a9e3c9660d63f1fa81d93e1e61560a77df6895dbebab4f274a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 309, + 152, + 364, + 189 + ], + "blocks": [ + { + "bbox": [ + 309, + 152, + 364, + 189 + ], + "lines": [ + { + "bbox": [ + 309, + 152, + 364, + 189 + ], + "spans": [ + { + "bbox": [ + 309, + 152, + 364, + 189 + ], + "type": "image", + "image_path": "67808f8cfb400f680e076f1edaa238597826d14799f37b9992bcd1db5bdc98f7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 311, + 189, + 361, + 202 + ], + "lines": [ + { + "bbox": [ + 311, + 189, + 361, + 202 + ], + "spans": [ + { + "bbox": [ + 311, + 189, + 361, + 202 + ], + "type": "text", + "content": "(b) SAFMN [36] \nDI: 8.6313" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 367, + 115, + 422, + 152 + ], + "blocks": [ + { + "bbox": [ + 367, + 115, + 422, + 152 + ], + "lines": [ + { + "bbox": [ + 367, + 115, + 422, + 152 + ], + "spans": [ + { + "bbox": [ + 367, + 115, + 422, + 152 + ], + "type": "image", + "image_path": "b566e27607b5623d28d3aec9523b1002af3d74e0970ec0f79ab1f4138d3f6cc0.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 366, + 152, + 421, + 189 + ], + "blocks": [ + { + "bbox": [ + 366, + 152, + 421, + 189 + ], + "lines": [ + { + "bbox": [ + 366, + 152, + 421, + 189 + ], + "spans": [ + { + "bbox": [ + 366, + 152, + 421, + 189 + ], + "type": "image", + "image_path": "94f9024109556785e7875fa8367c7cc1ef9ed81758c2869b0df50eb835e00b6b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 372, + 189, + 416, + 202 + ], + "lines": [ + { + "bbox": [ + 372, + 189, + 416, + 202 + ], + "spans": [ + { + "bbox": [ + 372, + 189, + 416, + 202 + ], + "type": "text", + "content": "(c) IMDN [17] DI: 10.3650" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 426, + 115, + 481, + 152 + ], + "blocks": [ + { + "bbox": [ + 426, + 115, + 481, + 152 + ], + "lines": [ + { + "bbox": [ + 426, + 115, + 481, + 152 + ], + "spans": [ + { + "bbox": [ + 426, + 115, + 481, + 152 + ], + "type": "image", + "image_path": "0a9a975d4a3ce5713532b8a1e825d89d9d8db44e96ac6953d13bb551b66a25ba.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 426, + 152, + 481, + 189 + ], + "blocks": [ + { + "bbox": [ + 426, + 152, + 481, + 189 + ], + "lines": [ + { + "bbox": [ + 426, + 152, + 481, + 189 + ], + "spans": [ + { + "bbox": [ + 426, + 152, + 481, + 189 + ], + "type": "image", + "image_path": "6766065451f11839c234fc266db8f164cc9809bb85fcd4407f4a12476b219984.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 433, + 189, + 473, + 202 + ], + "lines": [ + { + "bbox": [ + 433, + 189, + 473, + 202 + ], + "spans": [ + { + "bbox": [ + 433, + 189, + 473, + 202 + ], + "type": "text", + "content": "(d) SMFANet DI 12.0899" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 136, + 207, + 242, + 281 + ], + "blocks": [ + { + "bbox": [ + 136, + 207, + 242, + 281 + ], + "lines": [ + { + "bbox": [ + 136, + 207, + 242, + 281 + ], + "spans": [ + { + "bbox": [ + 136, + 207, + 242, + 281 + ], + "type": "image", + "image_path": "515868b5a042d5407350a0ec7fa4874227234c83e8a093e67cd87a02bcfebc94.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 154, + 282, + 223, + 288 + ], + "lines": [ + { + "bbox": [ + 154, + 282, + 223, + 288 + ], + "spans": [ + { + "bbox": [ + 154, + 282, + 223, + 288 + ], + "type": "text", + "content": "img005 from Urban100" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 313, + 480, + 346 + ], + "lines": [ + { + "bbox": [ + 130, + 313, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 313, + 480, + 346 + ], + "type": "text", + "content": "Fig. 4: Comparison of local attribution maps (LAMs) [13] and diffusion indices (DIs) [13]. Our proposed SMFANet family exploit more feature information and reconstructs a more accurate image structure." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 246, + 207, + 303, + 243 + ], + "blocks": [ + { + "bbox": [ + 246, + 207, + 303, + 243 + ], + "lines": [ + { + "bbox": [ + 246, + 207, + 303, + 243 + ], + "spans": [ + { + "bbox": [ + 246, + 207, + 303, + 243 + ], + "type": "image", + "image_path": "2c4bc7c6ac7ced408b455bb46a783b4b7a2643cebce02cbca429e7ea55326af7.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 246, + 243, + 303, + 281 + ], + "blocks": [ + { + "bbox": [ + 246, + 243, + 303, + 281 + ], + "lines": [ + { + "bbox": [ + 246, + 243, + 303, + 281 + ], + "spans": [ + { + "bbox": [ + 246, + 243, + 303, + 281 + ], + "type": "image", + "image_path": "0675c69552ca2fb49eafa7c97475c1da7de6d907b7739674d7882494344760bf.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 244, + 281, + 305, + 294 + ], + "lines": [ + { + "bbox": [ + 244, + 281, + 305, + 294 + ], + "spans": [ + { + "bbox": [ + 244, + 281, + 305, + 294 + ], + "type": "text", + "content": "(a) SwinIR-light [25] \nDI: 6.74785" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 309, + 207, + 364, + 243 + ], + "blocks": [ + { + "bbox": [ + 309, + 207, + 364, + 243 + ], + "lines": [ + { + "bbox": [ + 309, + 207, + 364, + 243 + ], + "spans": [ + { + "bbox": [ + 309, + 207, + 364, + 243 + ], + "type": "image", + "image_path": "dd5bd54d14477446067fddcb9616c38c604910aadb487a7e51cb7ffa2ac95f70.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 309, + 243, + 364, + 281 + ], + "blocks": [ + { + "bbox": [ + 309, + 243, + 364, + 281 + ], + "lines": [ + { + "bbox": [ + 309, + 243, + 364, + 281 + ], + "spans": [ + { + "bbox": [ + 309, + 243, + 364, + 281 + ], + "type": "image", + "image_path": "df0cc27ce4c00485fba494013181bb95ab3b1a6b1df1c7dffc2aa716d2180e0b.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 311, + 281, + 362, + 294 + ], + "lines": [ + { + "bbox": [ + 311, + 281, + 362, + 294 + ], + "spans": [ + { + "bbox": [ + 311, + 281, + 362, + 294 + ], + "type": "text", + "content": "b) HPINet-S [27] \nDI: 12.8471" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 366, + 207, + 422, + 243 + ], + "blocks": [ + { + "bbox": [ + 366, + 207, + 422, + 243 + ], + "lines": [ + { + "bbox": [ + 366, + 207, + 422, + 243 + ], + "spans": [ + { + "bbox": [ + 366, + 207, + 422, + 243 + ], + "type": "image", + "image_path": "55fb970d289cd7538733bc3b9b64ae24725808ecdec4829532b50186ad272984.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 366, + 243, + 422, + 281 + ], + "blocks": [ + { + "bbox": [ + 366, + 243, + 422, + 281 + ], + "lines": [ + { + "bbox": [ + 366, + 243, + 422, + 281 + ], + "spans": [ + { + "bbox": [ + 366, + 243, + 422, + 281 + ], + "type": "image", + "image_path": "e73e425d874198f09c8812aac9044827c29044036b57d102f67cd7cc1fabffae.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 365, + 281, + 423, + 294 + ], + "lines": [ + { + "bbox": [ + 365, + 281, + 423, + 294 + ], + "spans": [ + { + "bbox": [ + 365, + 281, + 423, + 294 + ], + "type": "text", + "content": "(c) ELAN-light [49] \nDI: 18.2614" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 426, + 207, + 481, + 243 + ], + "blocks": [ + { + "bbox": [ + 426, + 207, + 481, + 243 + ], + "lines": [ + { + "bbox": [ + 426, + 207, + 481, + 243 + ], + "spans": [ + { + "bbox": [ + 426, + 207, + 481, + 243 + ], + "type": "image", + "image_path": "58f50301a00a053071d401e70befeb31e290eb86fe6990ecfb8d2c104fc2a8ba.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 426, + 243, + 481, + 281 + ], + "blocks": [ + { + "bbox": [ + 426, + 243, + 481, + 281 + ], + "lines": [ + { + "bbox": [ + 426, + 243, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 426, + 243, + 481, + 281 + ], + "type": "image", + "image_path": "b03a000236565a3a6265e625991add603e6de9b2bde01ea3f747b02944f1ac4d.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 429, + 281, + 474, + 294 + ], + "lines": [ + { + "bbox": [ + 429, + 281, + 474, + 294 + ], + "spans": [ + { + "bbox": [ + 429, + 281, + 474, + 294 + ], + "type": "text", + "content": "(d) SMFANet+ \nDI: 19.6512" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "bbox": [ + 130, + 374, + 479, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 374, + 479, + 398 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 479, + 398 + ], + "type": "text", + "content": "values drop by 0.65dB and 0.96dB on the Urban100 and Manga109 datasets. These results show the importance of the SMFA." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 130, + 400, + 481, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 400, + 481, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 400, + 481, + 616 + ], + "type": "text", + "content": "In addition, as the proposed SMFA module mainly contains an EASA branch to explore non-local information and a LDE branch to capture local features, we conduct ablation studies with respect to these components to demonstrate their effectiveness on image SR. Without the EASA branch, significant performance drops of " + }, + { + "bbox": [ + 130, + 400, + 481, + 616 + ], + "type": "inline_equation", + "content": "0.25\\mathrm{dB}" + }, + { + "bbox": [ + 130, + 400, + 481, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 400, + 481, + 616 + ], + "type": "inline_equation", + "content": "0.33\\mathrm{dB}" + }, + { + "bbox": [ + 130, + 400, + 481, + 616 + ], + "type": "text", + "content": " are observed on the Urban100 and Manga109 datasets. Figure 5 (c) shows the corresponding visual results with blurred structure. Moreover, replacing EASA with the window-based SA [25] yields almost the same performance, but the corresponding runtime and GPU consumption are three times higher. These results indicate that the EASA branch can efficiently approximate SA to explore non-local information for accurate structural reconstruction. As for the LDE branch, removing it, the model only achieves the performance of " + }, + { + "bbox": [ + 130, + 400, + 481, + 616 + ], + "type": "inline_equation", + "content": "25.92\\mathrm{dB}" + }, + { + "bbox": [ + 130, + 400, + 481, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 400, + 481, + 616 + ], + "type": "inline_equation", + "content": "30.22\\mathrm{dB}" + }, + { + "bbox": [ + 130, + 400, + 481, + 616 + ], + "type": "text", + "content": " on the Urban100 and Manga109. Figure 5 (d) shows the reconstructed images with distorted lines, suggesting that the LDE branch contributes to recovering image details. Figure 6 shows the power spectral density maps, which intuitively illustrate the complementary roles of the EASA and LDE. The PSD of " + }, + { + "bbox": [ + 130, + 400, + 481, + 616 + ], + "type": "inline_equation", + "content": "X_{l}" + }, + { + "bbox": [ + 130, + 400, + 481, + 616 + ], + "type": "text", + "content": " derived from the EASA shows that the energy density is more concentrated in the center region, whereas the PSD of " + }, + { + "bbox": [ + 130, + 400, + 481, + 616 + ], + "type": "inline_equation", + "content": "Y_{d}" + }, + { + "bbox": [ + 130, + 400, + 481, + 616 + ], + "type": "text", + "content": " obtained from the LDE scatters in the surrounding area compared to " + }, + { + "bbox": [ + 130, + 400, + 481, + 616 + ], + "type": "inline_equation", + "content": "F_{in}" + }, + { + "bbox": [ + 130, + 400, + 481, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 400, + 481, + 616 + ], + "type": "inline_equation", + "content": "X_{l}" + }, + { + "bbox": [ + 130, + 400, + 481, + 616 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "To further validate the effectiveness of this complementary two-branch design, we replace SMFA with either dual LDE branches or dual EASA branches. With these two alternatives, the PSNR values on the Urban100 dataset drop significantly by 0.17dB and 0.25dB, respectively. We also investigated experiments" + } + ] + } + ], + "index": 33 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "type": "text", + "content": "Zheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 203, + 164 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 203, + 164 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 203, + 164 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 203, + 164 + ], + "type": "image", + "image_path": "7e30f05a56e3bead604cbf45935f3892a3a768bf12a1b4890a3f8fd230510e35.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 140, + 167, + 202, + 214 + ], + "blocks": [ + { + "bbox": [ + 140, + 167, + 202, + 214 + ], + "lines": [ + { + "bbox": [ + 140, + 167, + 202, + 214 + ], + "spans": [ + { + "bbox": [ + 140, + 167, + 202, + 214 + ], + "type": "image", + "image_path": "f9e7b5c314f67120b84faed2e38318eb817b67af3787bdba545da834af8db9b5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 143, + 217, + 198, + 229 + ], + "lines": [ + { + "bbox": [ + 143, + 217, + 198, + 229 + ], + "spans": [ + { + "bbox": [ + 143, + 217, + 198, + 229 + ], + "type": "text", + "content": "(a) HR patch" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 205, + 114, + 271, + 164 + ], + "blocks": [ + { + "bbox": [ + 205, + 114, + 271, + 164 + ], + "lines": [ + { + "bbox": [ + 205, + 114, + 271, + 164 + ], + "spans": [ + { + "bbox": [ + 205, + 114, + 271, + 164 + ], + "type": "image", + "image_path": "e612a2cd12dafce68be24143128e012f06053988cf6fa273270085c0d2219336.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 205, + 166, + 272, + 215 + ], + "blocks": [ + { + "bbox": [ + 205, + 166, + 272, + 215 + ], + "lines": [ + { + "bbox": [ + 205, + 166, + 272, + 215 + ], + "spans": [ + { + "bbox": [ + 205, + 166, + 272, + 215 + ], + "type": "image", + "image_path": "c0b03589eeb15392652fe753f81fc093dd1e1358b3c251cd82097ce3161a95e5.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 211, + 217, + 266, + 229 + ], + "lines": [ + { + "bbox": [ + 211, + 217, + 266, + 229 + ], + "spans": [ + { + "bbox": [ + 211, + 217, + 266, + 229 + ], + "type": "text", + "content": "(b) LR patch" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 273, + 114, + 340, + 164 + ], + "blocks": [ + { + "bbox": [ + 273, + 114, + 340, + 164 + ], + "lines": [ + { + "bbox": [ + 273, + 114, + 340, + 164 + ], + "spans": [ + { + "bbox": [ + 273, + 114, + 340, + 164 + ], + "type": "image", + "image_path": "512fe8f9ca834e2e47dbbb432d7c5f8910114adb985de56bcbce925dcb50b06e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 276, + 166, + 339, + 214 + ], + "blocks": [ + { + "bbox": [ + 276, + 166, + 339, + 214 + ], + "lines": [ + { + "bbox": [ + 276, + 166, + 339, + 214 + ], + "spans": [ + { + "bbox": [ + 276, + 166, + 339, + 214 + ], + "type": "image", + "image_path": "dc9f0eb35e92e0f035274a9bc793e3a5f76ec97d59250ddfd49b2e2865db03a6.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 277, + 217, + 336, + 229 + ], + "lines": [ + { + "bbox": [ + 277, + 217, + 336, + 229 + ], + "spans": [ + { + "bbox": [ + 277, + 217, + 336, + 229 + ], + "type": "text", + "content": "(c) w/o EASA" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 342, + 114, + 408, + 164 + ], + "blocks": [ + { + "bbox": [ + 342, + 114, + 408, + 164 + ], + "lines": [ + { + "bbox": [ + 342, + 114, + 408, + 164 + ], + "spans": [ + { + "bbox": [ + 342, + 114, + 408, + 164 + ], + "type": "image", + "image_path": "f690950930d464221fd546f824e0e2f86753942330f04803ca42daccdf3ab627.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 345, + 166, + 406, + 214 + ], + "blocks": [ + { + "bbox": [ + 345, + 166, + 406, + 214 + ], + "lines": [ + { + "bbox": [ + 345, + 166, + 406, + 214 + ], + "spans": [ + { + "bbox": [ + 345, + 166, + 406, + 214 + ], + "type": "image", + "image_path": "0340d8020ebf862ab88732c223b24031583c2dec748c19fccf0b2a17a5d37133.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 348, + 217, + 403, + 229 + ], + "lines": [ + { + "bbox": [ + 348, + 217, + 403, + 229 + ], + "spans": [ + { + "bbox": [ + 348, + 217, + 403, + 229 + ], + "type": "text", + "content": "(d) w/o LDE" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 410, + 114, + 477, + 164 + ], + "blocks": [ + { + "bbox": [ + 410, + 114, + 477, + 164 + ], + "lines": [ + { + "bbox": [ + 410, + 114, + 477, + 164 + ], + "spans": [ + { + "bbox": [ + 410, + 114, + 477, + 164 + ], + "type": "image", + "image_path": "be622c2a3e2bad6b10ec69f7352f7c1616d731b4d2f0e20e26c20ba0796159d1.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 414, + 166, + 477, + 214 + ], + "blocks": [ + { + "bbox": [ + 414, + 166, + 477, + 214 + ], + "lines": [ + { + "bbox": [ + 414, + 166, + 477, + 214 + ], + "spans": [ + { + "bbox": [ + 414, + 166, + 477, + 214 + ], + "type": "image", + "image_path": "da9bdd48e10af41c326fad8f77a23fdbed34e4a3f23191cb1d5523773d052251.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 416, + 217, + 472, + 229 + ], + "lines": [ + { + "bbox": [ + 416, + 217, + 472, + 229 + ], + "spans": [ + { + "bbox": [ + 416, + 217, + 472, + 229 + ], + "type": "text", + "content": "(e) SMFANet" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 143, + 282, + 223, + 332 + ], + "blocks": [ + { + "bbox": [ + 131, + 248, + 480, + 270 + ], + "lines": [ + { + "bbox": [ + 131, + 248, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 131, + 248, + 480, + 270 + ], + "type": "text", + "content": "Fig. 5: Effectiveness of the SMFA for image super-resolution on the " + }, + { + "bbox": [ + 131, + 248, + 480, + 270 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 131, + 248, + 480, + 270 + ], + "type": "text", + "content": " Urban100 and Manga109 datasets." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 143, + 282, + 223, + 332 + ], + "lines": [ + { + "bbox": [ + 143, + 282, + 223, + 332 + ], + "spans": [ + { + "bbox": [ + 143, + 282, + 223, + 332 + ], + "type": "image", + "image_path": "5f56e33f72f35ba4128d3b68639f94d6a1b969598451439768bc7a01f0c4d782.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 143, + 334, + 224, + 384 + ], + "blocks": [ + { + "bbox": [ + 143, + 334, + 224, + 384 + ], + "lines": [ + { + "bbox": [ + 143, + 334, + 224, + 384 + ], + "spans": [ + { + "bbox": [ + 143, + 334, + 224, + 384 + ], + "type": "image", + "image_path": "cab105c5193fe800703ab74b72f66e74252dd0051b6128f248fd08582e83514d.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 162, + 387, + 204, + 397 + ], + "lines": [ + { + "bbox": [ + 162, + 387, + 204, + 397 + ], + "spans": [ + { + "bbox": [ + 162, + 387, + 204, + 397 + ], + "type": "text", + "content": "HR image" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 225, + 282, + 306, + 332 + ], + "blocks": [ + { + "bbox": [ + 225, + 282, + 306, + 332 + ], + "lines": [ + { + "bbox": [ + 225, + 282, + 306, + 332 + ], + "spans": [ + { + "bbox": [ + 225, + 282, + 306, + 332 + ], + "type": "image", + "image_path": "b6cda250e48d7fe1d828e067f4aabbadd63de947fe02b64a68d55dffe78001ac.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 226, + 334, + 306, + 384 + ], + "blocks": [ + { + "bbox": [ + 226, + 334, + 306, + 384 + ], + "lines": [ + { + "bbox": [ + 226, + 334, + 306, + 384 + ], + "spans": [ + { + "bbox": [ + 226, + 334, + 306, + 384 + ], + "type": "image", + "image_path": "5c1e44240c5fc96459029bcb7c64204cea7cb866b55ae7b321cd6df0d45f4060.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 230, + 387, + 301, + 397 + ], + "lines": [ + { + "bbox": [ + 230, + 387, + 301, + 397 + ], + "spans": [ + { + "bbox": [ + 230, + 387, + 301, + 397 + ], + "type": "text", + "content": "Input feature " + }, + { + "bbox": [ + 230, + 387, + 301, + 397 + ], + "type": "inline_equation", + "content": "F_{in}" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 307, + 282, + 388, + 332 + ], + "blocks": [ + { + "bbox": [ + 307, + 282, + 388, + 332 + ], + "lines": [ + { + "bbox": [ + 307, + 282, + 388, + 332 + ], + "spans": [ + { + "bbox": [ + 307, + 282, + 388, + 332 + ], + "type": "image", + "image_path": "4d6917221a1c6b19dfd7a4900ebddac4a2fdad208f88e6de30c949785748b790.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 307, + 334, + 388, + 384 + ], + "blocks": [ + { + "bbox": [ + 307, + 334, + 388, + 384 + ], + "lines": [ + { + "bbox": [ + 307, + 334, + 388, + 384 + ], + "spans": [ + { + "bbox": [ + 307, + 334, + 388, + 384 + ], + "type": "image", + "image_path": "1170d6da30fae62866a76bd910c1a6a6a3a0b90a37cfb6e86dbc86f00abaa4b1.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 387, + 378, + 396 + ], + "lines": [ + { + "bbox": [ + 317, + 387, + 378, + 396 + ], + "spans": [ + { + "bbox": [ + 317, + 387, + 378, + 396 + ], + "type": "inline_equation", + "content": "X_{l}" + }, + { + "bbox": [ + 317, + 387, + 378, + 396 + ], + "type": "text", + "content": " from EASA" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 390, + 282, + 471, + 332 + ], + "blocks": [ + { + "bbox": [ + 390, + 282, + 471, + 332 + ], + "lines": [ + { + "bbox": [ + 390, + 282, + 471, + 332 + ], + "spans": [ + { + "bbox": [ + 390, + 282, + 471, + 332 + ], + "type": "image", + "image_path": "678c0309701644ec5974c3936f3895a50db6b37277883a19d7ce177e42428c56.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 390, + 334, + 471, + 384 + ], + "blocks": [ + { + "bbox": [ + 390, + 334, + 471, + 384 + ], + "lines": [ + { + "bbox": [ + 390, + 334, + 471, + 384 + ], + "spans": [ + { + "bbox": [ + 390, + 334, + 471, + 384 + ], + "type": "image", + "image_path": "24721c17aae3917b6c4c9ca66c319cfa2ca111b05d48a4b1ace89e925fb750cf.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 387, + 458, + 396 + ], + "lines": [ + { + "bbox": [ + 403, + 387, + 458, + 396 + ], + "spans": [ + { + "bbox": [ + 403, + 387, + 458, + 396 + ], + "type": "inline_equation", + "content": "Y_{d}" + }, + { + "bbox": [ + 403, + 387, + 458, + 396 + ], + "type": "text", + "content": " from LDE" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 417, + 480, + 460 + ], + "lines": [ + { + "bbox": [ + 131, + 417, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 131, + 417, + 480, + 460 + ], + "type": "text", + "content": "Fig. 6: The power spectral density (PSD) visualizations of feature " + }, + { + "bbox": [ + 131, + 417, + 480, + 460 + ], + "type": "inline_equation", + "content": "F_{in}" + }, + { + "bbox": [ + 131, + 417, + 480, + 460 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 417, + 480, + 460 + ], + "type": "inline_equation", + "content": "X_{l}" + }, + { + "bbox": [ + 131, + 417, + 480, + 460 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 417, + 480, + 460 + ], + "type": "inline_equation", + "content": "Y_{d}" + }, + { + "bbox": [ + 131, + 417, + 480, + 460 + ], + "type": "text", + "content": ". We perform a periodic shift of the spectrum map such that the low-frequency component is moved to the center. The EASA activates more low-frequency components for feature " + }, + { + "bbox": [ + 131, + 417, + 480, + 460 + ], + "type": "inline_equation", + "content": "X_{l}" + }, + { + "bbox": [ + 131, + 417, + 480, + 460 + ], + "type": "text", + "content": ", and the LDE enhances high-frequency representations for feature " + }, + { + "bbox": [ + 131, + 417, + 480, + 460 + ], + "type": "inline_equation", + "content": "Y_{d}" + }, + { + "bbox": [ + 131, + 417, + 480, + 460 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "bbox": [ + 130, + 486, + 480, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 480, + 521 + ], + "type": "text", + "content": "that sequentially process the EASA and LDE, whose corresponding reconstruction performances are clearly decreased as well. These reduced results prove the effectiveness of such a parallel design." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 130, + 521, + 481, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 481, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 481, + 641 + ], + "type": "text", + "content": "Effectiveness of the variance modulation. The variance modulation is used in the EASA branch to facilitate better non-local information exploration. To demonstrate its effectiveness, we remove this operation from the EASA branch. Table 5 presents that the PSNR values drop by " + }, + { + "bbox": [ + 130, + 521, + 481, + 641 + ], + "type": "inline_equation", + "content": "0.06\\mathrm{dB}" + }, + { + "bbox": [ + 130, + 521, + 481, + 641 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 521, + 481, + 641 + ], + "type": "inline_equation", + "content": "0.08\\mathrm{dB}" + }, + { + "bbox": [ + 130, + 521, + 481, + 641 + ], + "type": "text", + "content": " on the Urban100 and Manga109 datasets. For investigating the effect of " + }, + { + "bbox": [ + 130, + 521, + 481, + 641 + ], + "type": "inline_equation", + "content": "\\sigma^2 (X)" + }, + { + "bbox": [ + 130, + 521, + 481, + 641 + ], + "type": "text", + "content": ", we further replace " + }, + { + "bbox": [ + 130, + 521, + 481, + 641 + ], + "type": "inline_equation", + "content": "\\sigma^2 (X)" + }, + { + "bbox": [ + 130, + 521, + 481, + 641 + ], + "type": "text", + "content": " with the variance of " + }, + { + "bbox": [ + 130, + 521, + 481, + 641 + ], + "type": "inline_equation", + "content": "X_{s}" + }, + { + "bbox": [ + 130, + 521, + 481, + 641 + ], + "type": "text", + "content": " and a learnable parameter " + }, + { + "bbox": [ + 130, + 521, + 481, + 641 + ], + "type": "inline_equation", + "content": "\\alpha \\in \\mathbb{R}^{1\\times 1\\times C}" + }, + { + "bbox": [ + 130, + 521, + 481, + 641 + ], + "type": "text", + "content": " for feature modulation, and their PSNR values on the Urban100 datasets are reduced to 26.12dB and 26.13dB, respectively. These results suggest that the variance of " + }, + { + "bbox": [ + 130, + 521, + 481, + 641 + ], + "type": "inline_equation", + "content": "X_{s}" + }, + { + "bbox": [ + 130, + 521, + 481, + 641 + ], + "type": "text", + "content": " or the learnable parameter " + }, + { + "bbox": [ + 130, + 521, + 481, + 641 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 521, + 481, + 641 + ], + "type": "text", + "content": " does not accurately represent the global statistical divergence of the input feature." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 131, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 480, + 665 + ], + "type": "text", + "content": "In addition, variance modulation with element-wise product incurs a 0.05dB and 0.11dB PSNR reduction on Urban100 and Manga109, respectively. This" + } + ] + } + ], + "index": 33 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 255, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 255, + 91, + 447, + 102 + ], + "type": "text", + "content": "SMFANet for Efficient Image Super-Resolution" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 157, + 479, + 242 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 4: Ablation experiments for " + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": " SMFANet on Urban100 and Manga109 datasets. All measured metrics are calculated in the same way as in Table 1. “A → B” is to replace A with B, and “A → None” is to remove operator A." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 157, + 479, + 242 + ], + "lines": [ + { + "bbox": [ + 134, + 157, + 479, + 242 + ], + "spans": [ + { + "bbox": [ + 134, + 157, + 479, + 242 + ], + "type": "table", + "html": "
AblationVariant#Params(K)#FLOPs(G)#GPU Mem. (M)#Avg. Time(ms)Urban100Manga109
BaselineSMFANet1971193.209.2626.19/0.786130.72/0.9097
SMFASMFA → None82556.903.1225.54/0.764629.76/0.8974
EASA → None1831077.317.2825.94/0.778330.39/0.9061
EASA → Window-based SA23716269.6937.5126.20/0.788230.74/0.9105
7722128776.746.9525.92/0.777230.92/0.9032
SMFA → Dual LDE branches2631485.729.6326.02/0.781030.52/0.9076
SMFA → Dual EASA branches143888.459.2825.94/0.778330.29/0.9076
SMFA → "EASA then LDE"1871177.168.9026.04/0.782430.57/0.9077
SMFA → "LDE then EASA"1871172.988.9526.11/0.784130.64/0.9083
Variance ModulationVariance σ2(X) → None1971193.208.8726.13/0.785230.64/0.9091
Variance σ2(X) → σ2(Xs)1971193.209.3126.12/0.783930.60/0.9086
Variance σ2(X) → Learnable parameter α1971193.209.1826.13/0.785330.65/0.9092
Addition → Element-wise Product1971193.209.1026.14/0.784130.66/0.9083
PCFNPCFN → None132792.946.6425.89/0.777030.38/0.9051
PCFN → FFN174993.118.7126.06/0.782530.56/0.9079
PCFN → ConvFFN2131293.1311.0326.19/0.786130.71/0.9097
PCFN → GDFN20412193.4810.8326.22/0.787230.75/0.9101
", + "image_path": "5509c7d8ac736e1302e84060fd11c16bd0e8dadb04d7c3caf1562770b29efdc0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 267, + 481, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 267, + 481, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 267, + 481, + 376 + ], + "type": "text", + "content": "performance drop is caused by the element-wise product operation suppressing the feature channel with low variance, resulting in a decreased information flow. Effectiveness of the PCFN. The PCFN is used to facilitate the fusion of the representative features extracted by the SMFA module. To demonstrate the effectiveness of this module, we performed ablation experiments with removal of PCFN and replacement of FFN [40], respectively. Table 5 shows that the exclusion of PCFN and the substitution of FFN each leads to a respective decrease in PSNR values of 0.34dB and 0.16dB on the Manga109 [33] dataset, compared to the baseline." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 377, + 481, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 377, + 481, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 377, + 481, + 450 + ], + "type": "text", + "content": "Recent FFN variants [36, 43, 45, 53] introduce convolution to boost local encoding capability for better reconstruction performance. However, the expanded hidden features contain redundant information, and processing them with convolution directly increases the computational cost. For example, we use GDFN [45] for channel mixing, resulting in a " + }, + { + "bbox": [ + 130, + 377, + 481, + 450 + ], + "type": "inline_equation", + "content": "\\times 2" + }, + { + "bbox": [ + 130, + 377, + 481, + 450 + ], + "type": "text", + "content": " increase in GPU consumption over PCFN, with only a slight improvement in reconstruction performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 451, + 481, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 451, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 451, + 481, + 511 + ], + "type": "text", + "content": "Thus, our PCFN introduces a partial convolution that processes only " + }, + { + "bbox": [ + 130, + 451, + 481, + 511 + ], + "type": "inline_equation", + "content": "1/4" + }, + { + "bbox": [ + 130, + 451, + 481, + 511 + ], + "type": "text", + "content": " of the hidden channels, which reduces redundant computation and memory usage while efficiently encoding local information. Table 4 shows that PCFN achieves comparable SR performance with less computational overhead compared to ConvFFN [53] and GDFN [45]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 536, + 220, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 220, + 550 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 220, + 550 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 569, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 481, + 666 + ], + "type": "text", + "content": "We have proposed a simple yet effective SMFANet method to solve image superresolution efficiently. As the key component of SMFANet, our SMFA module contains the EASA for non-local information exploration and the LDE for local detail modeling. Moreover, we develop a PCFN to facilitate the fusion of the non-local and local information extracted by the SMFA module both in the channel and spatial dimensions. Extensive experimental results show that the proposed SMFANet family achieve a favorable trade-off between reconstruction performance and computational efficiency." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "type": "text", + "content": "Zheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 481, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 481, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 481, + 177 + ], + "type": "text", + "content": "This work has been partly supported by the National Natural Science Foundation of China (Nos. U22B2049, 62272233, 62332010), and the Fundamental Research Funds for the Central Universities (No. 30922010910)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 196, + 197, + 209 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 196, + 197, + 209 + ], + "spans": [ + { + "bbox": [ + 133, + 196, + 197, + 209 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 223, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 138, + 223, + 480, + 244 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 223, + 480, + 244 + ], + "spans": [ + { + "bbox": [ + 138, + 223, + 480, + 244 + ], + "type": "text", + "content": "1. Ahn, N., Kang, B., Sohn, K.A.: Fast, accurate, and lightweight super-resolution with cascading residual network. In: ECCV (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 246, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 246, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 138, + 246, + 480, + 267 + ], + "type": "text", + "content": "2. Arbeláez, P., Maire, M., Fowlkes, C.C., Malik, J.: Contour detection and hierarchical image segmentation. PAMI 33(5), 898-916 (2011)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 269, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 269, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 138, + 269, + 480, + 300 + ], + "type": "text", + "content": "3. Bevilacqua, M., Roumy, A., Guillemot, C., line Alberi Morel, M.: Low-complexity single-image super-resolution based on nonnegative neighbor embedding. In: BMVC (2012)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 301, + 480, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 301, + 480, + 322 + ], + "spans": [ + { + "bbox": [ + 138, + 301, + 480, + 322 + ], + "type": "text", + "content": "4. Chen, H., Wang, Y., Guo, T., Xu, C., Deng, Y., Liu, Z., Ma, S., Xu, C., Xu, C., Gao, W.: Pre-trained image processing transformer. In: CVPR (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 323, + 480, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 323, + 480, + 344 + ], + "spans": [ + { + "bbox": [ + 138, + 323, + 480, + 344 + ], + "type": "text", + "content": "5. Chen, J., hong Kao, S., He, H., Zhuo, W., Wen, S., Lee, C.H., Chan, S.H.G.: Run, don't walk: Chasing higher flops for faster neural networks. In: CVPR (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 346, + 480, + 366 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 346, + 480, + 366 + ], + "spans": [ + { + "bbox": [ + 138, + 346, + 480, + 366 + ], + "type": "text", + "content": "6. Chen, X., Wang, X., Zhou, J., Dong, C.: Activating more pixels in image superresolution transformer. In: CVPR (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 368, + 480, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 368, + 480, + 388 + ], + "spans": [ + { + "bbox": [ + 138, + 368, + 480, + 388 + ], + "type": "text", + "content": "7. Choi, H., Lee, J., Yang, J.: N-gram in swim transformers for efficient lightweight image super-resolution. In: CVPR (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 390, + 480, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 390, + 480, + 411 + ], + "spans": [ + { + "bbox": [ + 138, + 390, + 480, + 411 + ], + "type": "text", + "content": "8. Dai, S., Han, M., Xu, W., Wu, Y., Gong, Y.: Soft edge smoothness prior for alpha channel super resolution. In: CVPR (2007)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 412, + 480, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 412, + 480, + 433 + ], + "spans": [ + { + "bbox": [ + 138, + 412, + 480, + 433 + ], + "type": "text", + "content": "9. Dong, C., Loy, C.C., He, K., Tang, X.: Image super-resolution using deep convolutional networks. PAMI 38(2), 295-307 (2016)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 434, + 480, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 434, + 480, + 455 + ], + "spans": [ + { + "bbox": [ + 138, + 434, + 480, + 455 + ], + "type": "text", + "content": "0. Dong, C., Loy, C.C., Tang, X.: Accelerating the super-resolution convolutional neural network. In: ECCV (2016)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 456, + 480, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 456, + 480, + 477 + ], + "spans": [ + { + "bbox": [ + 138, + 456, + 480, + 477 + ], + "type": "text", + "content": "1. Dong, J., Pan, J., Yang, Z., Tang, J.: Multi-scale residual low-pass filter network for image deblurring. In: ICCV (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 479, + 480, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 479, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 134, + 479, + 480, + 521 + ], + "type": "text", + "content": "12. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 522, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 522, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 134, + 522, + 480, + 544 + ], + "type": "text", + "content": "13. Gu, J., Dong, C.: Interpreting super-resolution networks with local attribution maps. In: CVPR (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 545, + 480, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 545, + 480, + 576 + ], + "spans": [ + { + "bbox": [ + 134, + 545, + 480, + 576 + ], + "type": "text", + "content": "14. Guo, H., Li, J., Dai, T., Ouyang, Z., Ren, X., Xia, S.T.: Mambair: A simple baseline for image restoration with state-space model. arXiv preprint arXiv:2402.15648 (2024)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 134, + 578, + 480, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 578, + 480, + 598 + ], + "spans": [ + { + "bbox": [ + 134, + 578, + 480, + 598 + ], + "type": "text", + "content": "15. Hendrycks, D., Gimpel, K.: Gaussian error linear units. arXiv preprint arXiv:1606.08415 (2016)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 134, + 600, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 600, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 134, + 600, + 480, + 620 + ], + "type": "text", + "content": "16. Huang, J.B., Singh, A., Ahuja, N.: Single image super-resolution from transformed self-exemplars. In: CVPR (2015)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 134, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 134, + 622, + 480, + 643 + ], + "type": "text", + "content": "17. Hui, Z., Gao, X., Yang, Y., Wang, X.: Lightweight image super-resolution with information multi-distillation network. In: ACM MM (2019)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 134, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 644, + 480, + 665 + ], + "type": "text", + "content": "18. Kim, J., Lee, J.K., Lee, K.M.: Accurate image super-resolution using very deep convolutional networks. In: CVPR (2016)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 255, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 255, + 91, + 447, + 102 + ], + "type": "text", + "content": "SMFANet for Efficient Image Super-Resolution" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "19. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. In: ICLR (2015)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 480, + 171 + ], + "type": "text", + "content": "20. Li, A., Zhang, L., Liu, Y., Zhu, C.: Feature modulation transformer: Crossrefinement of global representation via high-frequency prior for image superresolution. In: ICCV (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 171, + 480, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 171, + 480, + 192 + ], + "spans": [ + { + "bbox": [ + 132, + 171, + 480, + 192 + ], + "type": "text", + "content": "21. Li, M., Ma, B., Zhang, Y.: Lightweight image super-resolution with pyramid clustering transformer. TCSVT pp. 1-1 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 192, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 192, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 192, + 480, + 224 + ], + "type": "text", + "content": "22. Li, W., Zhou, K., Qi, L., Jiang, N., Lu, J., Jia, J.: LAPAR: Linearly-assembled pixel-adaptive regression network for single image super-resolution and beyond. In: NeurIPS (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 224, + 480, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 480, + 257 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 480, + 257 + ], + "type": "text", + "content": "23. Li, Y., Fan, Y., Xiang, X., Demandolx, D., Ranjan, R., Timofte, R., Gool, L.V.: Efficient and explicit modelling of image hierarchies for image restoration. In: CVPR (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 257, + 480, + 279 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 257, + 480, + 279 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 480, + 279 + ], + "type": "text", + "content": "24. Li, Z., Liu, Y., Chen, X., Cai, H., Gu, J., Qiao, Y., Dong, C.: Blueprint separable residual network for efficient image super-resolution. In: CVPR Workshops (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 279, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 279, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 279, + 480, + 300 + ], + "type": "text", + "content": "25. Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: SwinIR: Image restoration using swin transformer. In: ICCV Workshops (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 300, + 480, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 480, + 322 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 480, + 322 + ], + "type": "text", + "content": "26. Lim, B., Son, S., Kim, H., Nah, S., Lee, K.M.: Enhanced deep residual networks for single image super-resolution. In: CVPR Workshops (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 322, + 480, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 322, + 480, + 343 + ], + "spans": [ + { + "bbox": [ + 132, + 322, + 480, + 343 + ], + "type": "text", + "content": "27. Liu, J., Chen, C., Tang, J., Wu, G.: From coarse to fine: Hierarchical pixel integration for lightweight image super-resolution. In: AAAI (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 343, + 480, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 480, + 365 + ], + "type": "text", + "content": "28. Liu, J., Tang, J., Wu, G.: Residual feature distillation network for lightweight image super-resolution. In: ECCV Workshops (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 365, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 480, + 385 + ], + "type": "text", + "content": "29. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin Transformer: Hierarchical vision transformer using shifted windows. In: ICCV (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 385, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 385, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 385, + 480, + 407 + ], + "type": "text", + "content": "30. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. In: ICLR (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 407, + 480, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 480, + 429 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 480, + 429 + ], + "type": "text", + "content": "31. Lu, Z., Li, J., Liu, H., Huang, C., Zhang, L., Zeng, T.: Transformer for single image super-resolution. In: CVPR Workshops (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 429, + 480, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 429, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 480, + 460 + ], + "type": "text", + "content": "32. Mao, Y., Zhang, N., Wang, Q., Bai, B., Bai, W., Fang, H., Liu, P., Li, M., Yan, S.: Multi-level dispersion residual network for efficient image super-resolution. In: CVPR Workshops (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 460, + 480, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 460, + 480, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 480, + 483 + ], + "type": "text", + "content": "33. Matsui, Y., Ito, K., Aramaki, Y., Yamasaki, T., Aizawa, K.: Sketch-based manga retrieval using manga109 dataset. arXiv preprint arXiv:1510.04389 (2015)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 483, + 433, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 433, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 433, + 493 + ], + "type": "text", + "content": "34. Park, N., Kim, S.: How do vision transformers work? In: ICLR (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 493, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 493, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 493, + 480, + 525 + ], + "type": "text", + "content": "35. Shi, W., Caballero, J., Huszar, F., Totz, J., Aitken, A.P., Bishop, R., Rueckert, D., Wang, Z.: Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In: CVPR (2016)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 525, + 480, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 480, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 480, + 547 + ], + "type": "text", + "content": "36. Sun, L., Dong, J., Tang, J., Pan, J.: Spatially-adaptive feature modulation for efficient image super-resolution. In: ICCV (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 547, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 547, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 547, + 480, + 568 + ], + "type": "text", + "content": "37. Sun, L., Pan, J., Tang, J.: ShuffleMixer: An efficient convnet for image superresolution. In: NeurIPS (2022)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 568, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 480, + 601 + ], + "type": "text", + "content": "38. Timofte, R., Agustsson, E., Van Gool, L., Yang, M.H., Zhang, L., et al.: NTIRE 2017 challenge on single image super-resolution: Methods and results. In: CVPR Workshops (2017)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 601, + 480, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 480, + 622 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 480, + 622 + ], + "type": "text", + "content": "39. Timofte, R., DeSmet, V., Van Gool, L.: A+: Adjusted anchored neighborhood regression for fast super-resolution. In: ACCV 2014 (2015)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "text", + "content": "40. Vaswani, Shazeer, Parmar, Uszkoreit, Jones, Gomez, Kaiser, ukasz, L., Polosukhin, Illia: Attention is all you need. In: NeurIPS (2017)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "41. Wang, H., Chen, X., Ni, B., Liu, Y., Liu, j.: Omni aggregation networks for lightweight image super-resolution. In: CVPR (2023)" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "Zheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 402 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "text", + "content": "42. Wang, L., Dong, X., Wang, Y., Ying, X., Lin, Z., An, W., Guo, Y.: Exploring sparsity in image super-resolution for efficient inference. In: CVPR (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 171 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 171 + ], + "type": "text", + "content": "43. Wang, W., Xie, E., Li, X., Fan, D.P., Song, K., Liang, D., Lu, T., Luo, P., Shao, L.: Pvtv2: Improved baselines with pyramid vision transformer. Computational Visual Media 8(3), 1-10 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 194 + ], + "type": "text", + "content": "44. Yang, J., Wright, J., Huang, T.S., Ma, Y.: Image super-resolution via sparse representation. TIP 19(11), 2861-2873 (2010)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "text", + "content": "45. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: CVPR (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "type": "text", + "content": "46. Zeyde, R., Elad, M., Protter, M.: On single image scale-up using sparse-representations. In: Curves and Surfaces (2012)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 238, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 481, + 258 + ], + "type": "text", + "content": "47. Zhang, A., Ren, W., Liu, Y., Cao, X.: Lightweight image super-resolution with superpixel token interaction. In: ICCV (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "type": "text", + "content": "48. Zhang, J., Peng, H., Wu, K., Liu, M., Xiao, B., Fu, J., Yuan, L.: Minivit: Compressing vision transformers with weight multiplexing. In: CVPR (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 282, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 303 + ], + "type": "text", + "content": "49. Zhang, X., Zeng, H., Guo, S., Zhang, L.: Efficient long-range attention network for image super-resolution. In: ECCV (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "type": "text", + "content": "50. Zhang, Y., Li, K., Li, K., Wang, L., Zhong, B., Fu, Y.: Image super-resolution using very deep residual channel attention networks. In: ECCV (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 326, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 346 + ], + "type": "text", + "content": "51. Zhang, Y., Tian, Y., Kong, Y., Zhong, B., Fu, Y.: Residual dense network for image super-resolution. In: CVPR (2018)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 348, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 348, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 348, + 481, + 380 + ], + "type": "text", + "content": "52. Zhou, L., Cai, H., Gu, J., Li, Z., Liu, Y., Chen, X., Qiao, Y., Dong, C.: Efficient image super-resolution using vast-receptive-field attention. In: ECCV Workshops (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 380, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 402 + ], + "type": "text", + "content": "53. Zhou, Y., Li, Z., Guo, C., Bai, S., Cheng, M., Hou, Q.: Srformer: Permuted self-attention for single image super-resolution. In: ICCV (2023)" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 255, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 255, + 91, + 447, + 102 + ], + "type": "text", + "content": "SMFANet for Efficient Image Super-Resolution" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/SMILe_ Leveraging Submodular Mutual Information For Robust Few-Shot Object Detection/57f9e916-d947-4fa8-bb11-b89fc66615b3_content_list.json b/2024/SMILe_ Leveraging Submodular Mutual Information For Robust Few-Shot Object Detection/57f9e916-d947-4fa8-bb11-b89fc66615b3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..132d245c04aec7d2e60ea1cf41b42a3061f0bf01 --- /dev/null +++ b/2024/SMILe_ Leveraging Submodular Mutual Information For Robust Few-Shot Object Detection/57f9e916-d947-4fa8-bb11-b89fc66615b3_content_list.json @@ -0,0 +1,1484 @@ +[ + { + "type": "text", + "text": "SMILE: Leveraging Submodular Mutual Information For Robust Few-Shot Object Detection", + "text_level": 1, + "bbox": [ + 266, + 140, + 736, + 205 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Anay Majee1, Ryan Sharp2, and Rishabh Iyer1", + "bbox": [ + 310, + 234, + 692, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 The University of Texas at Dallas, TX, USA", + "bbox": [ + 344, + 261, + 656, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{anay.majee, rishabh.iley}@utdallas.edu", + "bbox": [ + 344, + 276, + 651, + 290 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 IGS Energy, OH, USA", + "bbox": [ + 418, + 290, + 583, + 304 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "rysharp@igsenergy.com", + "bbox": [ + 418, + 305, + 583, + 318 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Confusion and forgetting of object classes have been challenges of prime interest in Few-Shot Object Detection (FSOD). To overcome these pitfalls in metric learning based FSOD techniques, we introduce a novel Submodular Mutual Information Learning (SMILE3) framework for loss functions which adopts combinatorial mutual information functions as learning objectives to enforce learning of well-separated feature clusters between the base and novel classes. Additionally, the joint objective in SMILE minimizes the total submodular information contained in a class leading to discriminative feature clusters. The combined effect of this joint objective demonstrates significant improvements in class confusion and forgetting in FSOD. Further we show that SMILE generalizes to several existing approaches in FSOD, improving their performance, agnostic of the backbone architecture. Experiments on popular FSOD benchmarks, PASCAL-VOC and MS-COCO show that our approach generalizes to State-of-the-Art (SoTA) approaches improving their novel class performance by up to $5.7\\%$ (3.3 mAP points) and $5.4\\%$ (2.6 mAP points) on the 10-shot setting of VOC (split 3) and 30-shot setting of COCO datasets respectively. Our experiments also demonstrate better retention of base class performance and up to $2\\times$ faster convergence over existing approaches agnostic of the underlying architecture.", + "bbox": [ + 259, + 357, + 743, + 636 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 662, + 375, + 678 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances in Deep Neural networks (DNNs) have enabled models to learn discriminative feature representations from large-scale image benchmarks. Unfortunately, these architectures fail to adapt to few-shot settings tasked to recognize novel objects over existing ones with few examples, closely resembling human-like perception. Although recent research has shown significant promise in few-shot image recognition [10, 12, 33, 35, 36], Few-Shot Object Detection (FSOD) remains a challenge with recent works [27, 28, 37, 38] highlighting two", + "bbox": [ + 212, + 694, + 787, + 801 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Work done as a Graduate student at The University of Texas at Dallas.", + "bbox": [ + 217, + 810, + 712, + 824 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "3 Project page: https://anaymajee.me/ assets/project_pages/smile.html.", + "bbox": [ + 217, + 824, + 732, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1ea79b0189d765270c8b55465a89be4be6ce67ec6f448f069b3c494f4fecaab2.jpg", + "image_caption": [ + "Fig. 1: Functionality of components in $L_{comb}$ proposed in the SMILE (ours) framework, (a) $L_{comb}^{inter}$ promotes separation between $C_b$ and $C_n$ while (c) $L_{comb}^{intra}$ promotes intra-class compactness." + ], + "image_footnote": [], + "bbox": [ + 250, + 148, + 751, + 308 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "major challenges - Class Confusion and Catastrophic Forgetting. Class confusion, as highlighted in [29] manifests itself through mis-prediction of instances belonging to a newly learnt (novel) class, as one or more instances of the already learnt (base) classes. Authors in [1, 38] attribute this to the sharing of visual information between classes resulting in increased inter-class bias due to overlapping feature clusters as shown in Fig. 1(a). Catastrophic forgetting refers to the gradual degradation in the performance of already learnt classes in the quest to learn the novel ones, as shown in Fig. 2(a), seldom overfitting to rare classes [29, 38]. Further, large feature diversity (intra-class variance) among base classes lead to formation of non-discriminative feature clusters as shown in Fig. 1(b), aggravating the existing inter-class bias in the feature space. Unlike existing approaches (refer Sec. 2) which target either confusion or forgetting, our paper presents a unified approach to tackle both these challenges in FSOD. Although, recent approaches [1, 28, 34] attempt to tackle these challenges through contrastive learning strategies, such approaches have been limited by their capability to overcome either inter-class bias or intra-class variance [30, 38] and poor generalization to longtail settings [30] (FSOD being a extreme case).", + "bbox": [ + 212, + 385, + 787, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we introduce a combinatorial viewpoint in FSOD considering each object class $i \\in [1, C]$ in the dataset $\\mathcal{T}$ as a set $A_i$ of samples, where $\\mathcal{T} = \\{A_1, \\dots, A_C\\}$ , facilitating the application of combinatorial functions as learning objectives. We aim to overcome the aforementioned challenges through representation learning in the low-data regime by adopting this formulation through the SMILE: Submodular Mutual Information Learning framework, wherein we introduce novel, set-based combinatorial objective functions for FSOD as shown in Fig. 3. SMILE introduces a joint objective formulation $L_{comb}$ (Eq. (3)) based on two popular flavors of submodular information functions - Submodular Mutual Information [22] (SMI) and Total Submodular Information [11] targeting the root causes of confusion and forgetting in FSOD. At first, SMILE is the first to introduce pairwise SMI functions $I_f$ in representation learning which model the common (overlapping) information between two", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Majee et al.", + "bbox": [ + 271, + 114, + 352, + 128 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/de43313bc2880b20646758366c8511fda729561914c5d681b588d8c3844dd8d4.jpg", + "image_caption": [ + "Fig. 2: Resilience to Catastrophic forgetting and faster convergence in SMILE over SoTA approaches. (a) shows that combinatorial losses in SMILE are robust to catastrophic forgetting, while (b) shows that objectives in SMILE results in faster convergence over SoTA FSOD methods (AGCM and DiGeo)." + ], + "image_footnote": [], + "bbox": [ + 246, + 143, + 493, + 281 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a9150b517d2da2d94c869bc75c30efe377be18107d2e72d210e55782362a4c13.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 143, + 754, + 281 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sets. Minimizing $I_{f}$ through the joint objective $L_{comb}$ reduces feature overlap between base and novel classes alleviating inter-class bias in the model towards abundantly sampled classes as shown in Fig. 1(b). We extend this property of SMI functions to the novel classes minimizing the inter-cluster overlap between few-shot classes, promoting learning of discriminative features from just few samples. Secondly, SMILE preserves the diversity within each class by minimizing the total submodular information contained within each set as shown in Fig. 1(c), minimizing the impact of forgetting. This formulation closely follows the observation in [30] which models cooperation [15] between instances in a set by minimizing a submodular function over a set, to preserve representative features. The unified objective $L_{comb}$ introduced in SMILE models both these necessary properties through a weighted sum of two distinct objectives $L_{comb}^{inter}$ and $L_{comb}^{intra}$ as shown in Fig. 3 balancing the tradeoff between inter-cluster separation and intra-cluster compactness respectively. This allows us to introduce a family of loss functions which inherently eliminates confusion and forgetting as shown in Tab. 5. We conduct our experiments on two popular FSOD benchmarks, PASCAL-VOC [6] and MS-COCO [26] for several few-shot settings and demonstrate the following contributions of SMILE:", + "bbox": [ + 212, + 366, + 787, + 638 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- SMILE introduces a novel set-based combinatorial viewpoint in FSOD by applying combinatorial Mutual Information based objective to discriminate between base and novel classes, in conjunction with submodular total information to minimize intra-class variance as the objective function.", + "- SMILE generalizes to existing approaches in FSOD, irrespective of the underlying architecture demonstrating up to $5.7\\%$ improvement in novel class performance (Tab. 2) over the baseline FSOD approach.", + "SMILE demonstrates up to $2 \\times$ faster convergence (Fig. 2(b)) over existing SoTA approaches resulting in faster generalization to unknown object classes.", + "- Finally, SMILE demonstrates up to $11\\%$ and $3.5\\%$ reduction in class confusion and catastrophic forgetting while achieving SoTA performance on popular FSOD benchmarks like PASCAL-VOC (by $5.7\\%$ on split 2, 10-shot setting) and MS-COCO ( $5.4\\%$ on 30-shot setting)." + ], + "bbox": [ + 225, + 645, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "SMILE: Submodular Mutual Information Learner", + "bbox": [ + 401, + 114, + 732, + 127 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 143, + 388, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Few-Shot Object Detection (FSOD): Classical FSOD approaches utilize finetuning [3] or distance metric learning [17] to adapt features to novel classes. Recent methods employ meta-learning techniques [16,40,41] with episodic training to learn class-specific features. Meta-Reweight [16] and Meta-RCNN [41] use additional feature extractors, while Add-Info [40] leverages feature differences between support and query images. Techniques like [43] enhance class-specific features through information sharing, and CME [24] aims to reduce class confusion. Attention mechanisms [7,44] are used to identify discriminative features. However, meta-learning approaches are resource-intensive and may fail to generalize to significantly different novel classes. Metric learning strategies like FsDet [37], FSCE [34], and SRR-FSD [45] offer better generalization without additional overheads. PNPDet [42] partially addresses catastrophic forgetting and class confusion. GFSD [8] proposes a Bias-Balanced RPN to mitigate overfitting in metric learners.", + "bbox": [ + 212, + 174, + 787, + 385 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recent approaches like [18,32] adopt weak supervision from unlabelled data or low confidence predictions in RoI pooling layers to generalize to novel classes. These methods often use abundant samples from base classes [28] to prevent catastrophic forgetting, adding computational overhead in low-shot settings. Vision transformers [4] have been adopted in FSOD through methods like imTED [27] and PDC [23], with reduced computational overhead by using pre-trained attention heads. Alternatively, DiGeo [28] and PDC [23] learn the geometry or difference in distributions of RoI proposals [13] between object classes to overcome forgetting and confusion. However, these approaches rely on contrastive learning objectives [20] that struggle to learn discriminative feature embeddings due to adoption of pairwise similarity metrics. Our work, SMILE, aims to improve the feature learning capacity of existing SoTA approaches, irrespective of their underlying architectures.", + "bbox": [ + 212, + 386, + 787, + 582 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Submodular Functions and Combinatorial Objectives: Submodular functions are recognized as set functions with an inherent diminishing returns characteristic. Defined as a set function $f: 2^{\\mathcal{V}} \\to \\mathbb{R}$ operating on a ground-set $\\mathcal{V}$ , a function is termed submodular if it adheres to the condition $f(X) + f(Y) \\geq f(X \\cup Y) + f(X \\cap Y), \\forall X, Y \\subseteq \\mathcal{V}$ [11]. These functions have garnered considerable attention in research, particularly in fields like data subset selection [22], active learning [21], and video summarization [19,22] through their ability in modeling concepts such as diversity, relevance, set-cover and representation. A subclass of submodular functions, namely Submodular Mutual Information (SMI) functions introduced in [22] model the similarity and diversity between pairs of object classes establishing itself as a powerful tool to model inter-class bias. Recently, Majee et al. [30] introduces these set-based combinatorial functions as objectives in representation learning and demonstrates their capability in overcoming inter-class bias (by minimizing the similarity between nonidentical object classes) and intra-class variance (maximizing the similarity between instances of the same object class). However, these functions are yet to be studied in the context of few-shot learning. We introduce novel instances of SMI based objectives in", + "bbox": [ + 212, + 583, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 230, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Majee et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/41d46461f410f09a8a7a20db24290b5e43a772732f76334dd723fb4d8a6a0cd8.jpg", + "image_caption": [ + "Fig. 3: Overview of our SMILE framework highlighting the application of Mutual Information function based objectives in SMILE for the fine-tuning stage of Few-Shot Object Detection." + ], + "image_footnote": [], + "bbox": [ + 212, + 143, + 787, + 313 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SMILEs to minimize inter-class bias between base and novel classes. To the best of our knowledge we are the first to introduce novel SMI based combinatorial objectives in conjunction with total information based combinatorial functions through SMILEs in a quest to minimize confusion and forgetting in few-shot object detection.", + "bbox": [ + 212, + 386, + 787, + 463 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 484, + 328, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Problem Definition : Few-Shot Object Detection", + "text_level": 1, + "bbox": [ + 214, + 513, + 661, + 529 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We define a few-shot learner $h(x, \\theta)$ as shown in Fig. 3 that receives input data $x$ from base classes $C_b \\in [1, |C_b|]$ and novel classes $C_n \\in [1, |C_n|]$ such that $C = \\{C_b \\cup C_n\\}$ and $\\{C_b \\cap C_n\\} = \\emptyset$ . Here, $\\theta$ denotes the learnable parameters. The training data can be divided into two distinct parts, base $D_{base}$ and novel $D_{novel}$ such that, $\\mathcal{T} = \\{D_{base} \\cup D_{novel}\\}$ and $\\{D_{base} \\cap D_{novel}\\} = \\emptyset$ . SMILe introduces a paradigm shift in FSOD by imbibing a combinatorial viewpoint, where the base dataset, $D_{base} = [A_1^b, A_2^b, \\dots, A_{|C_b|}^b]$ , containing abundant training examples from $C_b$ base classes and the novel dataset, $D_{novel} = [A_1^n, A_2^n, \\dots, A_{|C_n|}^n]$ containing only $K$ -shot ( $|A_i^n| = K$ for $i \\in [1, C_n]$ ) training examples from $C_n$ novel classes. The objective of the few-shot learner $h(x, \\theta)$ is to learn discriminative representation from classes in $D_{novel}$ without degradation in performance on classes in $D_{base}$ . Following FSCE [34] we adopt a two-stage training strategy. In the base training stage we train $h(x, \\theta)$ on abundant samples in $D_{base}$ , allowing the model to generalize on the domain of $D_{base}$ . The few-shot adaptation stage adapts $h(x, \\theta)$ to previously unseen $K$ -shot data by fine-tuning on data samples from $D_{base} \\cup D_{novel}$ where $|A_k| = K$ for $k \\in \\{C_b \\cup C_n\\}$ . The goal of SMILe is to overcome class confusion and forgetting in FSOD resulting from elevated inter-class bias and intra-class variance as observed in [1, 29, 38]. The final model $h(x, \\theta)$ obtained after two training stages is evaluated on $D_{test}$ containing unseen data samples from both $C_b \\cup C_n$ .", + "bbox": [ + 212, + 537, + 787, + 843 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "SMILE: Submodular Mutual Information Learner", + "bbox": [ + 401, + 114, + 730, + 127 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 The SMILE Framework", + "text_level": 1, + "bbox": [ + 215, + 146, + 454, + 160 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Adopting a combinatorial viewpoint as disclosed earlier allows us to employ submodular combinatorial functions as learning objectives to tackle confusion and forgetting in FSOD. As discussed in Sec. 2, minimizing a Submodular functions naturally models cooperation [15] while maximizing it models diversity [25] due to their inherent diminishing marginal returns property. SMILE adopts the aforementioned properties of submodular functions to define a novel family of combinatorial objective (loss) functions $L_{comb}(\\theta)$ which enforces orthogonality in the feature space when applied on Region-of-Interest (RoI) features in FSOD models. The loss function $L_{comb}(\\theta)$ can be decomposed into two major components - $L_{comb}^{inter}$ minimizes inter-class bias between base and novel classes and $L_{comb}^{intra}$ maximizes intra-class compactness within abundant classes.", + "bbox": [ + 212, + 169, + 787, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For $L_{comb}^{inter}$ , SMILE explores a sub-category of combinatorial functions, namely Submodular Mutual Information (SMI) which can be defined as $I_{f}(A_{i}, A_{j}) = f(A_{i}) + f(A_{j}) - f(A_{i} \\cup A_{j})$ [11, 22], and models the common information between two sets $A_{i}$ and $A_{j}$ , $\\forall i, j \\in \\mathcal{T}$ . Results in [11, 22] portray $I_{f}(A_{i}, A_{j}; \\theta)$ as a measure of the degree of similarity between object classes $A_{i}$ and $A_{j}$ . Adopting this definition of SMI, $L_{comb}^{inter}$ minimizes the SMI between the base $C_{b}$ and the novel $C_{n}$ classes, ensuring sufficient inter-cluster separation (by minimizing inter-class bias) as shown in Eq. (1). $L_{comb}^{inter}$ further minimizes the mutual information between classes in $C_{n}$ , minimizing inter-cluster overlaps between the novel classes. This is visually depicted in Fig. 1(b) and has been shown to be effective in mitigating class confusion in FSOD through our experiments in Sec. 4.", + "bbox": [ + 212, + 335, + 787, + 518 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {c o m b}} ^ {\\text {i n t e r}} (\\theta) = \\sum_ {\\substack {b \\in C _ {b} \\\\ n \\in C _ {n}}} I _ {f} (A _ {b}, A _ {n}; \\theta) + \\sum_ {\\substack {i, j \\in C _ {n} \\\\ i \\neq j}} I _ {f} (A _ {i}, A _ {j}; \\theta) = \\sum_ {\\substack {i \\in (C _ {b} \\cup C _ {n}) \\\\ j \\in C _ {n}: i \\neq j}} I _ {f} (A _ {i}, A _ {j}; \\theta) \\tag{1}\n$$\n", + "text_format": "latex", + "bbox": [ + 233, + 526, + 785, + 585 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addition to confusion which stems from inter-class bias, SMILEs aims at mitigating catastrophic forgetting [29] in FSOD which has been attributed to large intra-class variance among abundant object classes in [1,38]. In coherence to the combinatorial formulation in SMILEs we achieve this through $L_{comb}^{intra}$ which minimizes the Total Submodular Information, defined as $S_f(A_1, \\dots, A_{|C|}) = \\sum_{k=1}^{|C|} f(A_k; \\theta)$ , over sets $A_k \\in \\mathcal{T}$ , given a submodular function $f(A_k; \\theta)$ . As discussed earlier, minimizing the submodular information models cooperation which asserts that minimizing $L_{comb}^{inter}$ promotes learning of discriminative feature clusters, penalizing abundant classes to have large feature variance in the embedding space as shown in Fig. 1(c). Although submodular functions have been studied in the field of representation learning to minimize intra-class variance in [30], but primarily differs from SMILEs in modeling a longtail recognition task by minimizing the total submodular correlation, which models gain in information when new features are added to a set. The formulation of $L_{comb}^{intra}$ has been shown in Eq. (2) where we minimize the total submodular information within samples in each class in $C_b \\cup C_n$ and our experiments in Tab. 5", + "bbox": [ + 212, + 594, + 787, + 843 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Majee et al.", + "bbox": [ + 271, + 114, + 352, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "show the effectiveness of $L_{comb}^{intra}$ in boosting base class performance asserting the mitigation of catastrophic forgetting.", + "bbox": [ + 214, + 146, + 787, + 176 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nL _ {c o m b} ^ {i n t r a} (\\theta) = \\sum_ {b \\in C _ {b}} f (A _ {b}, \\theta) + \\sum_ {n \\in C _ {n}} f (A _ {n}, \\theta) = \\sum_ {k \\in \\left(C _ {b} \\cup C _ {n}\\right)} f (A _ {k}, \\theta) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 285, + 189, + 785, + 222 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablating on the choice of the submodular function $f$ and SMI functions $I_{f}$ we introduce several instances of SMILE objectives as discussed in Tab. 1.", + "bbox": [ + 214, + 233, + 785, + 262 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Encapsulating the aforementioned formulations of $L_{comb}^{inter}$ and $L_{comb}^{intra}$ in SMILE we define a joint objective $L_{comb}(\\theta)$ which tackles both the challenges of confusion and forgetting. We thus define $L_{comb}(\\theta)$ in Eq. (3) which is the weighted algebraic sum of $L_{comb}^{inter}$ and $L_{comb}^{intra}$ with the weighting factor $\\eta$ .", + "bbox": [ + 214, + 263, + 785, + 325 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {c o m b} (\\theta) = (1 - \\eta) L _ {c o m b} ^ {i n t r a} (\\theta) + \\eta L _ {c o m b} ^ {i n t e r} (\\theta) \\\\ = \\sum_ {i \\in C _ {b} \\cup C _ {n}} \\left[ (1 - \\eta) f \\left(A _ {i}, \\theta\\right) + \\eta \\sum_ {\\substack {j \\in C _ {n} \\\\ i \\neq j}} I _ {f} \\left(A _ {i}, A _ {j}; \\theta\\right) \\right] \\tag{3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 299, + 338, + 785, + 407 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Note, that the combinatorial objective $L_{comb}(\\theta)$ is applied on output features from the RoI Pooling layers in proposal-based [31,38] architectures. To promote adoption of SMILE agnostic of the backbone architecture we introduce a combinatorial head $Z_{comb} = Comb(h,\\theta)$ which projects the RoI features to 128-dimensional feature vectors [20], $Z_{comb}$ on which $L_{comb}(\\theta)$ is applied during the few-shot adaptation stage.", + "bbox": [ + 214, + 419, + 785, + 508 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Finally, we summarize the total classification loss in SMILEs as depicted in Eq. (4) as the sum over all three objectives: the classification head $L_{Clf}$ , the box regression head $L_{bbox}$ and the combinatorial head $L_{comb}(\\theta)$ . Note that the objectives proposed in SMILEs apply only to $Comb(h,\\theta)$ while the RoI classification and regression heads are unchanged. This follows the observations in [28,38] which warrants the boost in performance originating from learning robust feature representations for each RoI predicted by the model.", + "bbox": [ + 214, + 510, + 785, + 616 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nL _ {c l s} (\\theta) = L _ {C l f} (\\theta) + L _ {b b o x} (\\theta) + L _ {c o m b} (\\theta) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 359, + 628, + 785, + 646 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Instantiations of $L_{comb}^{inter}$ and $L_{comb}^{intra}$ in the SMILE Framework", + "text_level": 1, + "bbox": [ + 214, + 667, + 756, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Given a submodular function $f(A)$ and a Submodular Mutual Information (SMI) function $I_{f}(A,Q)$ over sets $A$ and $Q$ , we derive two instances $L_{comb}^{inter}$ and $L_{comb}^{intra}$ objectives in SMILEs. Depending on the choice of $f(A)$ we define two instances: Facility-Location Mutual Information (SMILE-FLMI) and Graph-Cut Mutual Information (SMILE-GCMI). Inherently, both objectives adopt the cosine similarity metric $S_{ij}(\\theta)$ as used in SupCon [20] which can be defined as $S_{ij}(\\theta) = \\frac{Z_{comb_i}^{\\mathrm{T}} \\cdot Z_{comb_j}}{||Z_{comb_i}|| \\cdot ||Z_{comb_j}||}$ to compute similarity between sets in the learning objective. Although the similarity kernel used in SMILEs is computed in a pairwise fashion, objectives defined under $L_{comb}$ use it to only compute feature interactions", + "bbox": [ + 214, + 693, + 787, + 842 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "SMILE: Submodular Mutual Information Learner", + "bbox": [ + 401, + 114, + 732, + 127 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8ac960cc7f5c4206e4dd60abece09b20c0dd526c24b83514bfeaaeaeb1e25f7a.jpg", + "table_caption": [ + "Table 1: Summary of various instantiations of SMILE highlighting the components of the combinatorial objective, $L_{comb}^{inter}$ and $L_{comb}^{intra}$ ." + ], + "table_footnote": [], + "table_body": "
ObjectiveInstances of Lintercomb(θ)Instances of Intracomb(θ)
SMILE-GCMI (ours)∑k∈(Ck∪Cn)2λ∑iAksjA1Sij(θ)∑k∈Ck∪Cn∑iAksjTsetAksij(θ) - λ∑i,jAksij(θ)
∑k∈(Ck∪Cn)2λ∑iAksjA1maxSij(θ) + λ∑iAksjA1maxSij(θ)∑k∈Ck∪Cn∑iTsetAksij(θ)
SMILE-FLMI (ours)
", + "bbox": [ + 215, + 184, + 785, + 252 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "between samples, differing from existing approaches in aggregation of pairwise similarities to compute total information and mutual information over classes in $\\mathcal{T}$ .", + "bbox": [ + 212, + 271, + 787, + 316 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "SMILE-FLMI based objective is derived from the Facility-Location Mutual Information (FLMI) [22] function, expressed as $I_{f}(Q,A) = \\sum_{i\\in Q}\\max_{j\\in A}S_{ij}(\\theta) + \\lambda \\sum_{i\\in A}\\max_{j\\in Q}S_{ij}(\\theta)$ and minimizes the maximum similarity (most similar) between sets $Q$ and $A$ . Given the facility-location (FL) submodular function $f(A,\\theta) = \\sum_{i\\in \\mathcal{T}}\\max_{j\\in A}S_{ij}(\\theta)$ over the set $A$ , we can derive $L_{comb}^{inter}(\\theta)$ and $L_{comb}^{intra}(\\theta)$ shown in Eq. (5) as the SMILE-FLMI objective. Note that $L_{comb}^{inter}(\\theta)$ is applied between object classes in $C_b\\cup C_n$ and $C_n$ while $L_{comb}^{intra}(\\theta)$ is applied over all classes in $C_b\\cup C_n$ .", + "bbox": [ + 212, + 340, + 787, + 481 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {c o m b} ^ {i n t e r} (\\theta) = \\sum_ {k \\in \\left(C _ {b} \\cup C _ {n}\\right)} \\sum_ {i \\in A _ {k}} \\max _ {j \\in A _ {l}} S _ {i j} (\\theta) + \\lambda \\sum_ {i \\in A _ {l}} \\max _ {j \\in A _ {k}} S _ {i j} (\\theta), \\\\ l \\in C _ {n}: k \\neq l \\tag {5} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 300, + 491, + 784, + 539 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nL _ {c o m b} ^ {i n t r a} (\\theta) = \\sum_ {k \\in C _ {b} \\cup C _ {n}} \\sum_ {i \\in \\mathcal {T} \\backslash A _ {k}} \\max _ {j \\in A _ {k}} S _ {i j} (\\theta)\n$$\n", + "text_format": "latex", + "bbox": [ + 300, + 539, + 571, + 570 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Minimizing the $L_{comb}^{inter}$ objective function ensures that the sets $A_{l} \\in C_{n}$ and $A_{k} \\in C_{b} \\cup C_{n}$ are disjoint by minimizing the similarity between features in $A_{k}$ and the hardest negative $(\\sum_{i \\in A_{k}} \\max_{j \\in A_{l}} S_{ij}(\\theta)$ for $k \\in C_{b} \\cup C_{n}$ and $l \\in C_{n})$ feature vectors in $A_{l}$ . Further, $L_{comb}^{inter}$ enforces sufficient separation between the novel classes themselves to promote learning of disjoint feature clusters even with few-shot data overcoming confusion. Additionally, $L_{comb}^{intra}$ minimizes the total information contained in each set $A_{k} \\in (C_{b} \\cup C_{n})$ . This objective retains discriminative feature information from each class in $\\mathcal{T}$ reducing the impact of forgetting.", + "bbox": [ + 212, + 584, + 787, + 708 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "SMILE-GCMI based objective described in Eq. (6) minimizes the pairwise similarity of feature vectors between a positive set $A_{k} \\in C_{b} \\cup C_{n}$ and the sets in $A_{l} \\in C_{n}$ while maximizing the similarity between features in each set $A_{k} \\in C_{b} \\cup C_{n}$ . Given two sets $Q$ and $A$ , [22] defines the Graph-Cut SMI to be $I_{f}(Q, A) = 2\\lambda \\sum_{i \\in Q} \\sum_{j \\in A} S_{ij}(\\theta)$ , where the Graph-Cut function over a set $A$ is given by $f(A, \\theta) = \\sum_{i \\in A} \\sum_{j \\in T \\setminus A_{k}} S_{ij}(\\theta) - \\lambda \\sum_{i,j \\in A} S_{ij}(\\theta)$ . Given the Graph-Cut and the Graph-Cut SMI functions, we derive $L_{comb}^{inter}(\\theta)$ and $L_{comb}^{intra}(\\theta)$", + "bbox": [ + 212, + 731, + 787, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Majee et al.", + "bbox": [ + 271, + 114, + 352, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "shown in Eq. (6) as the SMILE-GCMI objective. Similar to SMILE-FLMI, the $L_{comb}^{inter}(\\theta)$ is applied between object classes in $C_b \\cup C_n$ and $C_n$ while $L_{comb}^{intra}(\\theta)$ is applied over all classes in $C_b \\cup C_n$ .", + "bbox": [ + 215, + 146, + 787, + 191 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nL _ {c o m b} ^ {i n t e r} (\\theta) = \\sum_ {\\substack {k \\in (C _ {b} \\cup C _ {n}) \\\\ l \\in C _ {n}: k \\neq l}} 2 \\lambda \\sum_ {i \\in A _ {k}} \\sum_ {j \\in A _ {l}} S _ {i j} (\\theta), \\tag{6}\n$$\n", + "text_format": "latex", + "bbox": [ + 307, + 200, + 784, + 247 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nL _ {c o m b} ^ {i n t r a} (\\theta) = \\sum_ {k \\in C _ {b} \\cup C _ {n}} \\sum_ {i \\in A _ {k}} \\sum_ {j \\in \\mathcal {T} \\backslash A _ {k}} S _ {i j} (\\theta) - \\lambda \\sum_ {i, j \\in A _ {k}} S _ {i j} (\\theta)\n$$\n", + "text_format": "latex", + "bbox": [ + 307, + 246, + 696, + 279 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Although objectives in SMILE-FLMI and SMILE-GCMI are tasked with similar functions, the $L_{comb}^{inter}$ in SMILE-GCMI minimizes the pairwise similarity between sets in $C_b \\cup C_n$ and $C_n$ rather than the most similar set in SMILE-FLMI. Further, the $L_{comb}^{intra}$ in SMILE-GCMI scales linearly with size of $A_k$ as described in [30]. This does not allow the model to substantially improve performance on learning discriminative feature representations for classes in both $C_b$ and $C_n$ as the $|A_k| = K$ (number of shots) thus failing to outperform the model trained using SMILE-FLMI.", + "bbox": [ + 212, + 286, + 782, + 405 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The detailed derivations of the aforementioned instances are included in the Supplementary material. Our experiments in Sec. 4.4 elucidates the fact that SMILE-FLMI is a better choice to overcome forgetting and confusion in FSOD.", + "bbox": [ + 212, + 407, + 782, + 452 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 473, + 375, + 491 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We evaluate models in SMILE by adopting standard evaluation criterion in FSOD [16, 37] and report the Mean Average Precision (mAP) at $50\\%$ Intersection Over Union (IoU) for all our experiments.", + "bbox": [ + 212, + 502, + 782, + 547 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 215, + 566, + 426, + 583 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Datasets We evaluate our proposed SMILE approach on two few-shot object detection datasets - and PASCAL-VOC [5] and MS-COCO [26] datasets.", + "bbox": [ + 212, + 589, + 782, + 619 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "PASCAL-VOC [5] dataset consists of 20 classes, out of which 15 are considered as base and 5 as novel classes. The novel classes are chosen at random giving rise to three data splits namely, split-1 (bird, bus, cow, motorbike, sofa), split-2 (aeroplane, bottle, cow, horse, sofa) and split-3 (boat, cat, motorbike, sheep, sofa). Following previous works [16], we use the combined VOC $07 + 12$ datasets for training and evaluate our models on the complete validation set of VOC 2007 for 1, 5, and 10 shot settings.", + "bbox": [ + 212, + 631, + 784, + 736 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "MS-COCO [26] dataset consists of 80 classes, out of which 60 are considered as base and 20 as novel classes. Following existing approaches in FSOD [41] we randomly select 5k samples from $(D_{base} \\cup D_{novel})$ to use as the validation set while the remaining samples are used to generate random 10 and 30-shot splits for training of the MS-COCO 2014 dataset. The key difference between VOC and COCO is the large intra-class variance and class-imbalance in COCO.", + "bbox": [ + 212, + 750, + 782, + 838 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "SMILE: Submodular Mutual Information Learner", + "bbox": [ + 401, + 114, + 732, + 127 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/4ce735bb899e250f1d59da34769396814a4bb9dee644fad0b9c1daf73722aa9a.jpg", + "table_caption": [ + "Table 2: Quantitative analysis on PASCAL-VOC dataset: Few-shot object detection performance ( $mAP_{novel}$ ) on novel class splits of PASCAL-VOC dataset. We tabulate results for $K = 1, 5, 10$ shots from various SoTA techniques in FSOD. * indicates that the results are averaged over 10 random seeds. † indicates a meta-learning strategy (N-way, K-shot training)." + ], + "table_footnote": [], + "table_body": "
MethodLearner TypeBackboneSplit 1Split 2Split 3
K=151015101510
†Meta-RCNN [41]MetaFRCN-R10119.945.751.510.434.845.414.341.248.1
†Meta-Reweight [16]MetaYOLO V214.833.947.215.730.140.521.342.845.9
†MetaDet [38]MetaFRCN-R10118.936.849.621.831.743.020.643.944.1
†Add-Info [40]MetaFRCN-R10124.249.157.421.637.045.721.243.849.6
†CME [24]MetaYOLO V217.844.847.512.733.740.015.744.948.8
PNPDet [42]MetricDLA-3418.2-41.016.6-36.418.9-36.2
FsDet w/ FC [37]MetricFRCN-R10136.855.757.018.235.539.027.748.750.2
FsDet w/ cos [37]MetricFRCN-R10139.855.756.023.535.139.130.849.549.8
Retentive-RCNN [9]MetricFRCN-R10140.153.756.121.737.040.330.249.750.1
FSCE [34]MetricFRCN-R10141.057.457.827.344.449.840.153.257.7
\\( FSCE + SMILE \\) (ours)Comb.FRCN-R10141.257.961.129.244.650.541.355.659.0
AGCM [1]MetricFRCN-R10140.358.559.927.549.350.642.154.258.2
\\( AGCM + SMILE \\) (ours)Comb.FRCN-R10140.959.762.031.949.552.342.656.461.4
DiGeo [28]MetricFRCN-R10136.054.160.920.742.847.127.547.352.9
DiGeo + SMILE(ours)Comb.FRCN-R10136.156.662.326.544.147.333.151.956.4
imTED [27]MetricViT-B31.971.977.022.752.257.712.669.672.8
imTED + PDC [23]MetricViT-B36.673.177.115.551.856.018.967.972.8
PDC + SMILE (ours)Comb.ViT-B36.675.277.927.152.758.315.170.074.7
", + "bbox": [ + 218, + 226, + 789, + 467 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Implementation Details The SMILE framework adopts a architecture agnostic approach and adopt several backbones including Faster-RCNN [31] and ViT [27]. For VOC, the input batch size to the network is set to 16 and 2 in the base training and few-shot adaptation stages for Faster-RCNN and ViT based approaches. The input resolution is set to $764 \\times 1333$ pixels for data splits in COCO, while it is set to $800 \\times 600$ pixels for PASCAL-VOC. The hyper-parameters used in the formulation of SMILE, namely $\\eta$ and similarity kernel $S$ , are chosen through ablation experiments described in Sec. 4.4. Results from existing methods are a reproduction of the algorithm from publicly available codebases. All our experiments are performed on 4 NVIDIA GTX 1080 Ti GPUs with additional details in the supplementary material and code released at https://github.com/amajee11us/SMILE-FSOD.git.", + "bbox": [ + 212, + 493, + 787, + 676 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Results on Few-Shot PASCAL VOC Dataset", + "text_level": 1, + "bbox": [ + 215, + 695, + 630, + 709 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 2 records the results obtained from our SMILE framework on novel splits of the PASCAL-VOC dataset and contrasts it against SoTA FSOD techniques. We adopt four SoTA approaches FSCE [34], AGCM [1], DiGeo [28] and imTED [27], covering several backbone architectures Faster-RCNN + FPN (FSCE, AGCM and DiGeo) alongside ViT (imTED, PDC) and introducing SMILE ( $M + SMILE$ ) approach into existing architectures $M$ . For Faster-RCNN based architectures (FSCE) we show a maximum of $5.7\\%$ (3.3 map points) improvement while for FPN based architectures (AGCM and DiGeo) we show a $3.5\\%$ (2.1 map points", + "bbox": [ + 212, + 718, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Majee et al.", + "bbox": [ + 271, + 114, + 352, + 128 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "for AGCM+SMILE) improvement. It is interesting to note that unlike FSCE and AGCM, DiGeo uses abundant samples from $C_b$ alongside few-shot samples in $C_n$ (with upsampling) during finetuning introducing a large inter-class bias. SMILE outperforms DiGeo by up to 2.3 $mAP$ points (split 2, 10-shot) showing the resilience of SMILE towards imbalance, thus overcoming confusion in FSOD. Additionally, for recently introduced transformer based architectures (imTED + PDC [23]) SMILE outperforms the existing SoTA with a maximum improvement of 4.9 $mAP$ points (split 2, 5-shot) thus establishing SMILE as the SoTA on few-shot splits of VOC. Note, that the choice of objective functions $L_{comb}^{inter}$ and $L_{comb}^{intra}$ for this experiment has been determined to be SMILE+FLMI through an ablation on the different instances in SMILEs as described in Sec. 4.4. Finally, Fig. 2(b) shows that introduction of SMILE framework to existing SoTA approaches leads to rapid convergence on the novel classes up to 2x over existing SoTA. This is significant for mission critical tasks like autonomous driving where the model is required to rapidly learn novel objects to reduce turn-around time.", + "bbox": [ + 212, + 145, + 787, + 373 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Results on Few-Shot MS-COCO Dataset", + "text_level": 1, + "bbox": [ + 215, + 398, + 596, + 414 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Similar to the results in PASCAL VOC we demonstrate the results of our SMILE framework on MS-COCO dataset. In contrast to VOC, COCO presents an extremely imbalanced setting with a long-tail distribution within $D_{base}$ itself making it really hard for FSOD approaches to achieve SoTA through primitive objective functions. Following the ablation experiments in Sec. 4.4 we adopt the SMILE+FLMI objective (best performing) to conduct the experiments on 20 few-shot classes of MS-COCO dataset. We show that SMILE generalizes existing SoTA approach (imTED + PDC) for COCO dataset by $5.4\\%$ (2.6 mAP points, 30-shot setting). This further establishes the generalizability of our approach over varying data distributions (VOC and COCO) while achieving SoTA in FSOD tasks.", + "bbox": [ + 212, + 426, + 787, + 592 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.4 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 619, + 388, + 635 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We conduct ablation experiments on the 10-shot split of VOC (split 1) with hyper-parameters $\\eta = 0.5$ , cosine similarity metric and $\\lambda = 1.0$ . Ablations for hyper-parameters are detailed in the supplementary material.", + "bbox": [ + 212, + 647, + 785, + 694 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Components of SMILE Instantiations in SMILEs consists of two main components - $L_{comb}^{inter}$ and $L_{comb}^{intra}$ . We consider three baselines FSCE, AGCM and DiGeo which follow the Faster-RCNN/FPN backbone for this experiment. First, we introduce $L_{comb}^{intra}$ by adopting the FL based objective as determined through ablation experiments below. This objective models intra-class variance and ensures reduction in intra-class variance characterized by boost in base class performance. Secondly, following the SMILE-FLMI formulation in Eq. (5) we introduce $L_{comb}^{inter}$ during the few-shot adaptation stage. Applying this objective minimizes", + "bbox": [ + 212, + 719, + 787, + 842 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "SMILE: Submodular Mutual Information Learner", + "bbox": [ + 401, + 114, + 730, + 127 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/ce88f033d4a6c31997410fe43e549c3acb88a4ec1950a9c6a24b6b486d6875fa.jpg", + "table_caption": [ + "Table 3: Performance of SMILE on MS COCO dataset : Our SMILE objectives demonstrate better generalizability while outperforming SoTA FSOD approaches on novel class performance $mAP_{50}$ (novel)." + ], + "table_footnote": [], + "table_body": "
Method\\( {mAP} \\)\\( {mAP}_{50} \\)\\( {mAP}_{75} \\)\\( {mAP} \\)\\( {mAP}_{50} \\)\\( {mAP}_{75} \\)
10-shot30-shot
Meta-Reweight [16]5.612.34.69.119.07.6
Meta-RCNN [41]8.719.16.612.425.310.8
TFA w/cos [37]10.0-9.313.7-13.4
Add-Info [40]12.527.39.814.730.612.2
MPSR [39]9.817.99.714.125.414.2
FSCE [34]11.9-10.516.4-16.2
FADI [2]12.222.711.916.129.115.8
CME [24]15.124.616.416.2--
FCT [14]17.130.217.021.435.522.1
imTED-B [27]22.536.623.730.247.432.5
imTED-B+PDC [23]23.438.124.530.847.333.5
PDC + SMILE (ours)25.840.126.131.049.933.6
", + "bbox": [ + 292, + 198, + 709, + 380 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/ed3a0f7f188f7fb22316323b1e1b8345208d3174b963fbb4add03f20bfbb8be9.jpg", + "table_caption": [ + "Table 4: Ablation on various components of the proposed SMILE approach." + ], + "table_footnote": [], + "table_body": "
MethodBaselinef(Ai)(Lintra)If(Ai,Aj)(Lintercomb)mAPbasemAPnovel
FsDet w/ cos---23.639.8
FSCE✓86.157.8
✓✓89.660.1
✓✓88.361.0
✓✓✓89.861.1
AGCM✓87.658.0
✓✓88.661.3
✓✓88.961.8
✓✓✓89.361.8
DiGeo✓90.560.9
✓✓92.361.7
✓✓91.462.0
✓✓✓92.662.3
", + "bbox": [ + 310, + 421, + 687, + 617 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "the inter-class bias between $C_b \\cup C_n$ and $C_n$ , thus improving novel class performance significantly. Nevertheless, we see a slight drop in base class performance due to forgetting prevalent in FSOD tasks. Finally, we combine both instantiations in SMILE into one single objective as in Eq. (3) with $\\eta = 0.5$ and show that SMILE improves both base and novel class performance emerging as the best choice for FSOD. To demonstrate generalization, we perform this experiment on several SoTA approaches as baseline and show that the results discussed in aforementioned section hold. We summarize all the results in Tab. 4.", + "bbox": [ + 212, + 647, + 787, + 768 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Choice of Combinatorial functions in $L_{comb}$ SMILE introduces several instances of $L_{comb}^{intra}$ and $L_{comb}^{inter}$ . To clearly understand the contributions of each of these instances we conduct experiments tabulated in Tab. 5 and determine the", + "bbox": [ + 212, + 794, + 787, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Majee et al.", + "bbox": [ + 271, + 114, + 352, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/14912a88ec6bf36a2df0f19fbfc3f9aee23db7ba163ba927c86509d5a96a87e9.jpg", + "table_caption": [ + "Table 5: Ablation on the choice of Submodular Information function $I_{f}$ and Submodular Information function $f$ for $L_{comb}$ in SMILE." + ], + "table_footnote": [], + "table_body": "
Modelf(A,θ)If(Ai,Aj)mAPbasemAPnovel
DiGeo [28]--87.957.4
GCGCMI92.660.9
FLFLMI93.162.3
FSCE [34]--86.157.8
GCGCMI87.460.3
FLFLMI89.861.1
", + "bbox": [ + 357, + 184, + 643, + 268 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "best performing formulation which generalizes to existing FSOD architectures. Unlike other ablation experiments in Sec. 4.4, we conduct our experiments on DiGeo which introduces an extremely imbalanced scenario by using abundant samples in $D_{base}$ . We conclude that SMILE-FLMI which considers the Facility-Location based objective as $L_{comb}^{intra}$ and Facility-Location Mutual Information based objective as $L_{comb}^{inter}$ as the best performing instantiation. This result follows the formulation in Eq. (5) where FL naturally models intra-class compactness in class-imbalanced settings [30] while FLMI penalizes the classes in $C_b$ to learn overlapping feature representations with the classes in $C_n$ . We use SMILE-FLMI for all benchmark experiments in Sec. 4.2 and Sec. 4.3.", + "bbox": [ + 212, + 297, + 787, + 449 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Robustness to Catastrophic Forgetting One of the most significant challenges in FSOD is the elimination of catastrophic forgetting which manifests as the degradation in the performance of classes in $C_b$ while learning classes in $C_n$ . This primarily occurs due to the lack of discriminative feature representations from instances in $D_{base}$ during the few-shot adaptation (stage 2) stage. We plot the change in base class performance as the training progresses in existing SoTA methods AGCM and DiGeo against number of training iterations in Fig. 2(a). At first, we contrast the change in base class performance $mAP_{base}$ between AGCM and AGCM+SMILE and observe that AGCM overfits on the few-shot samples in $D_{base}$ reducing the performance on $C_{base}$ as the training progresses. AGCM + SMILE on the other hand better retains the performance on base classes with $\\sim 3.5\\%$ better retention in base class performance. Interestingly, DiGeo is able to retain most of the base class performance with a very small degradation over the roofline (a model trained with only the base classes until convergence). Our DiGeo+SMILE approach outperforms Digo by demonstrating base class performance even higher than the roofline establishing the supremacy of $L_{comb}$ in overcoming inter-class bias and intra-class variance resulting in robustness against catastrophic forgetting.", + "bbox": [ + 212, + 470, + 787, + 743 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Overcoming Class Confusion Figure 4 highlights the supremacy of the proposed SMILE framework in mitigating class confusion through confusion matrix plots. We compare the confusion between classes in $C_b \\cup C_n$ of SoTA approaches AGCM and DiGeo before and after introduction of combinatorial objectives in SMILE. Although both approaches use K-shot examples for classes in $C_n$ , Di", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "SMILE: Submodular Mutual Information Learner", + "bbox": [ + 401, + 114, + 730, + 127 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/21f3a143182809eff3cca7577c7548522161ae5987ee05923df949cf9e9624f3.jpg", + "image_caption": [ + "Fig. 4: Ablation on Overcoming Class Confusion in SMILE. (a,b) SMILE demonstrates $11\\%$ lower confusion over AGCM and (c,d) $4\\%$ lower confusion over DiGeo. Only significant numbers are highlighted. Best viewed in $200\\%$ zoom." + ], + "image_footnote": [], + "bbox": [ + 215, + 146, + 362, + 265 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/e66a23625cc084bf94517d28a182a2cf8e48f318423e47579ad4849ec8db6bab.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 146, + 495, + 265 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/c3c2890c81276f9bb2dd8825bc0aa0e27b5a08a24c0d73ae3e10c69de2be3dcf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 495, + 148, + 627, + 265 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/b5352161d55bd070afbe86f70f9bf6e94be824d53405d25622a51c09e25b6497.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 627, + 147, + 779, + 266 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Geo differs from AGCM by adopting an upsampling strategy which allows the utilization of abundant examples in $C_b$ while upsampling the instances in $C_n$ . This injects different degrees of inter-class biases for models trained by adopting AGCM and DiGeo which has been demonstrated as the primary reason for confusion in previous work [29]. At first, we observe from Fig. 4 that by adopting the upsampling based strategy, DiGeo achieves very low confusion between already learnt base classes, leading to significantly lower confusion (5% among $C_b$ and $C_n$ ). Further, confusion matrix plots in Fig. 4 show that AGCM+SMILE demonstrates 11% lower confusion than AGCM and DiGeo+SMILE shows 4% lower confusion and DiGeo. This proves the efficacy of combinatorial objectives ( $L_{comb}^{inter}$ ) in mitigating inter-class bias, thereby reducing confusion between classes.", + "bbox": [ + 212, + 335, + 787, + 517 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 561, + 359, + 577 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we have presented a novel approach to Few-Shot Object Detection (FSOD) by introducing a combinatorial viewpoint through the SMILE framework. By leveraging the properties of set-based combinatorial functions, SMILE aims to address the challenges of class confusion and catastrophic forgetting, which are prevalent in FSOD tasks. Our approach incorporates Submodular Mutual Information (SMI) and Submodular Information Measures (SIM) to penalize overlapping features between base and novel classes and to ensure the formation of compact feature clusters, respectively. The experimental results on PASCAL-VOC and MS-COCO benchmarks demonstrate the effectiveness of SMILE, showing significant improvements in novel class performance, faster convergence, and a reduction in class confusion and catastrophic forgetting. Overall, SMILE offers a promising direction for advancing the state-of-the-art in FSOD by providing a generalized framework that is adaptable to various underlying architectures and capable of handling the complexities associated with few-shot learning in object detection.", + "bbox": [ + 212, + 613, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Majee et al.", + "bbox": [ + 271, + 114, + 352, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "We gratefully thank anonymous reviewers for their valuable comments. This work is supported by the National Science Foundation under Grant Numbers IIS-2106937, a gift from Google Research, an Amazon Research Award, and the Adobe Data Science Research award. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the National Science Foundation, Google or Adobe.", + "bbox": [ + 212, + 176, + 787, + 282 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 306, + 321, + 321 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Agarwal, A., Majee, A., Subramanian, A., Arora, C.: Attention guided cosine margin to overcome class-imbalance in few-shot road object detection. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) Workshops. pp. 221-230 (2022)", + "2. Cao, Y., Wang, J., Jin, Y., Wu, T., Chen, K., Liu, Z., Lin, D.: Few-shot object detection via association and discrimination. In: Thirty-Fifth Conference on Neural Information Processing Systems (2021)", + "3. Chen, H., Wang, Y., Wang, G., Qiao, Y.: LSTD: A Low-Shot Transfer Detector For Object Detection. In: AAAI. pp. 2836-2843 (2018)", + "4. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2021)", + "5. Everingham, M., Van Gool, L., Williams, C.K.I., Winn, J., Zisserman, A.: The Pascal Visual Object Classes (VOC) Challenge. IJCV pp. 303-338 (2010)", + "6. Everingham, M., Van Gool, L., Williams, C., Winn, J., Zisserman, A.: The Pascal visual object classes (voc) challenge. International Journal of Computer Vision 88, 303-338 (06 2010)", + "7. Fan, Q., Zhuo, W., Tang, C.K., Tai, Y.W.: Few-Shot Object Detection With Attention-RPN And Multi-Relation Detector. In: CVPR (2020)", + "8. Fan, Z., Ma, Y., Li, Z., Sun, J.: Generalized Few-Shot Object Detection Without Forgetting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4527-4536 (June 2021)", + "9. Fan, Z., Ma, Y., Li, Z., Sun, J.: Generalized few-shot object detection without forgetting (2021)", + "0. Finn, C., Abbeel, P., Levine, S.: Model-Agnostic Meta-Learning For Fast Adaptation Of Deep Networks. In: ICML (2017)", + "1. Fujishige, S.: Submodular Functions and Optimization, vol. 58. Elsevier (2005)", + "2. Gidaris, S., Komodakis, N.: Dynamic Few-Shot Visual Learning Without Forgetting. In: CVPR (2018)", + "3. Girshick, R.B.: Fast R-CNN. ICCV (2015)", + "4. Han, G., Ma, J., Huang, S., Chen, L., Chang, S.F.: Few-shot object detection with fully cross-transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5321-5330 (2022)", + "5. Jegelka, S., Bilmes, J.: Submodularity beyond submodular energies: Coupling edges in graph cuts. In: CVPR 2011 (2011)" + ], + "bbox": [ + 225, + 339, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "SMILE: Submodular Mutual Information Learner", + "bbox": [ + 401, + 114, + 730, + 127 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Kang, B., Liu, Z., Wang, X., Yu, F., Feng, J., Darrell, T.: Few-shot Object Detection Via Feature Reweighting. In: ICCV (2019)", + "17. Karlinsky, L., Shtok, J., Harary, S., Schwartz, E., Aides, A., Feris, R., Giryes, R., Bronstein, A.M.: RepMet: Representative-Based Metric Learning For Classification And Few-Shot Object Detection. In: CVPR (2019)", + "18. Kaul, P., Xie, W., Zisserman, A.: Label, Verify, Correct: A Simple Few-Shot Object Detection Method. In: IEEE Conference on Computer Vision and Pattern Recognition (2022)", + "19. Kaushal, V., Iyer, R., Doctor, K., Sahoo, A., Dubal, P., Kothawade, S., Mahadev, R., Dargan, K., Ramakrishnan, G.: Demystifying multi-faceted video summarization: Tradeoff between diversity, representation, coverage and importance. In: 2019 IEEE Winter Conference on Applications of Computer Vision (WACV). pp. 452-461 (2019)", + "20. Khosla, P., Teterwak, P., Wang, C., Sarna, A., Tian, Y., Isola, P., Maschinot, A., Liu, C., Krishnan, D.: Supervised contrastive learning. In: Advances in Neural Information Processing Systems (2020)", + "21. Kothawade, S., Ghosh, S., Shekhar, S., Xiang, Y., Iyer, R.K.: Talisman: Targeted active learning for object detection with rare classes and slices using submodular mutual information. In: Computer Vision - ECCV 2022 - 17th European Conference (2022)", + "22. Kothawade, S., Kaushal, V., Ramakrishnan, G., Bilmes, J.A., Iyer, R.K.: PRISM: A rich class of parameterized submodular information measures for guided data subset selection. In: Thirty-Sixth AAAI Conference on Artificial Intelligence, AAAI. pp. 10238-10246 (2022)", + "23. Li, B., Liu, C., Shi, M., Chen, X., Ji, X., Ye, Q.: Proposal distribution calibration for few-shot object detection. IEEE transactions on neural networks and learning systems (2022)", + "24. Li, B., Yang, B., Liu, C., Liu, F., Ji, R., Ye, Q.: Beyond Max-Margin: Class Margin Equilibrium For Few-Shot Object Detection. In: CVPR (June 2021)", + "25. Lin, H., Bilmes, J.: A class of submodular functions for document summarization. In: Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies (2011)", + "26. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision - ECCV 2014. pp. 740-755. Springer International Publishing, Cham (2014)", + "27. Liu, F., Zhang, X., Peng, Z., Guo, Z., Wan, F., Ji, X., Ye, Q.: Integrally Migrating Pre-trained Transformer Encoder-decoders for Visual Object Detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6825-6834 (2023)", + "28. Ma, J., Niu, Y., Xu, J., Huang, S., Han, G., Chang, S.F.: Digeo: Discriminative geometry-aware learning for generalized few-shot object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2023)", + "29. Majee, A., Agrawal, K., Subramanian, A.: Few-Shot Learning For Road Object Detection. In: AAAI Workshop on Meta-Learning and MetaDL Challenge. vol. 140, pp. 115-126 (2021)", + "30. Majee, A., Kothawade, S.N., Killamsetty, K., Iyer, R.K.: SCoRe: Submodular Combinatorial Representation Learning. In: Forty-first International Conference on Machine Learning (ICML) (2024)" + ], + "bbox": [ + 215, + 146, + 784, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Majee et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "31. Ren, S., He, K., Girshick, R.B., Sun, J.: Faster R-CNN: Towards Real-Time Object Detection With Region Proposal Networks. IEEE Trans. on Pattern Analysis and Machine Intelligence (2015)", + "32. Shangguan, Z., Rostami, M.: Identification of novel classes for improving few-shot object detection (2023)", + "33. Snell, J., Swersky, K., Zemel, R.: Prototypical Networks For Few-shot Learning. In: NeurIPS. pp. 4077-4087 (2017)", + "34. Sun, B., Li, B., Cai, S., Yuan, Y., Zhang, C.: FSCE: Few-Shot Object Detection Via Contrastive Proposal Encoding. In: CVPR (June 2021)", + "35. Sung, F., Yang, Y., Zhang, L., Xiang, T., Torr, P.H., Hospedales, T.M.: Learning To Compare: Relation Network For Few-Shot Learning. In: CVPR (June 2018)", + "36. Vinyals, O., Blundell, C., Lillicrap, T., kavukcuoglu, k., Wierstra, D.: Matching Networks For One Shot Learning. In: NeurIPS (2016)", + "37. Wang, X., Huang, T.E., Darrell, T., Gonzalez, J.E., Yu, F.: Frustratingly Simple Few-Shot Object Detection. In: ICML (2020)", + "38. Wang, Y.X., Ramanan, D., Hebert, M.: Meta-Learning To Detect Rare Objects. In: ICCV (2019)", + "39. Wu, J., Liu, S., Huang, D., Wang, Y.: Multi-scale positive sample refinement for few-shot object detection. In: European Conference on Computer Vision (2020)", + "40. Xiao, Y., Marlet, R.: Few-Shot Object Detection And Viewpoint Estimation For Objects In The Wild. In: ECCV (2020)", + "41. Yan, X., Chen, Z., Xu, A., Wang, X., Liang, X., Lin, L.: Meta R-CNN: Towards General Solver For Instance-Level Low-Shot Learning. In: CVPR. pp. 9577-9586 (2019)", + "42. Zhang, G., Cui, K., Wu, R., Lu, S., Tian, Y.: PNPDet: Efficient Few-Shot Detection Without Forgetting Via Plug-And-Play Sub-Networks. In: WACV. pp. 3823-3832 (2021)", + "43. Zhang, L., Zhou, S., Guan, J., Zhang, J.: Accurate Few-Shot Object Detection With Support-Query Mutual Guidance And Hybrid Loss. In: CVPR (June 2021)", + "44. Zhang, S., Luo, D., Wang, L., Koniusz, P.: Few-Shot Object Detection By Second-order Pooling. In: Proceedings of the Asian Conference on Computer Vision (ACCV) (November 2020)", + "45. Zhu, C., Chen, F., Ahmed, U., Shen, Z., Savvides, M.: Semantic Relation Reasoning For Shot-Stable Few-Shot Object Detection. In: CVPR (June 2021)" + ], + "bbox": [ + 212, + 146, + 787, + 618 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "SMILE: Submodular Mutual Information Learner", + "bbox": [ + 401, + 114, + 730, + 127 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/SMILe_ Leveraging Submodular Mutual Information For Robust Few-Shot Object Detection/57f9e916-d947-4fa8-bb11-b89fc66615b3_model.json b/2024/SMILe_ Leveraging Submodular Mutual Information For Robust Few-Shot Object Detection/57f9e916-d947-4fa8-bb11-b89fc66615b3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6a6e79f6c9ce5c156a0f15e967e64270ee4f73f2 --- /dev/null +++ b/2024/SMILe_ Leveraging Submodular Mutual Information For Robust Few-Shot Object Detection/57f9e916-d947-4fa8-bb11-b89fc66615b3_model.json @@ -0,0 +1,2038 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.267, + 0.141, + 0.737, + 0.207 + ], + "angle": 0, + "content": "SMILE: Leveraging Submodular Mutual Information For Robust Few-Shot Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.311, + 0.235, + 0.694, + 0.253 + ], + "angle": 0, + "content": "Anay Majee1, Ryan Sharp2, and Rishabh Iyer1" + }, + { + "type": "text", + "bbox": [ + 0.345, + 0.262, + 0.658, + 0.277 + ], + "angle": 0, + "content": "1 The University of Texas at Dallas, TX, USA" + }, + { + "type": "text", + "bbox": [ + 0.345, + 0.277, + 0.653, + 0.291 + ], + "angle": 0, + "content": "{anay.majee, rishabh.iley}@utdallas.edu" + }, + { + "type": "text", + "bbox": [ + 0.419, + 0.291, + 0.584, + 0.305 + ], + "angle": 0, + "content": "2 IGS Energy, OH, USA" + }, + { + "type": "text", + "bbox": [ + 0.419, + 0.306, + 0.584, + 0.319 + ], + "angle": 0, + "content": "rysharp@igsenergy.com" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.358, + 0.744, + 0.637 + ], + "angle": 0, + "content": "Abstract. Confusion and forgetting of object classes have been challenges of prime interest in Few-Shot Object Detection (FSOD). To overcome these pitfalls in metric learning based FSOD techniques, we introduce a novel Submodular Mutual Information Learning (SMILE3) framework for loss functions which adopts combinatorial mutual information functions as learning objectives to enforce learning of well-separated feature clusters between the base and novel classes. Additionally, the joint objective in SMILE minimizes the total submodular information contained in a class leading to discriminative feature clusters. The combined effect of this joint objective demonstrates significant improvements in class confusion and forgetting in FSOD. Further we show that SMILE generalizes to several existing approaches in FSOD, improving their performance, agnostic of the backbone architecture. Experiments on popular FSOD benchmarks, PASCAL-VOC and MS-COCO show that our approach generalizes to State-of-the-Art (SoTA) approaches improving their novel class performance by up to \\(5.7\\%\\) (3.3 mAP points) and \\(5.4\\%\\) (2.6 mAP points) on the 10-shot setting of VOC (split 3) and 30-shot setting of COCO datasets respectively. Our experiments also demonstrate better retention of base class performance and up to \\(2\\times\\) faster convergence over existing approaches agnostic of the underlying architecture." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.663, + 0.376, + 0.679 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.695, + 0.788, + 0.803 + ], + "angle": 0, + "content": "Recent advances in Deep Neural networks (DNNs) have enabled models to learn discriminative feature representations from large-scale image benchmarks. Unfortunately, these architectures fail to adapt to few-shot settings tasked to recognize novel objects over existing ones with few examples, closely resembling human-like perception. Although recent research has shown significant promise in few-shot image recognition [10, 12, 33, 35, 36], Few-Shot Object Detection (FSOD) remains a challenge with recent works [27, 28, 37, 38] highlighting two" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.713, + 0.825 + ], + "angle": 0, + "content": "* Work done as a Graduate student at The University of Texas at Dallas." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.733, + 0.84 + ], + "angle": 0, + "content": "3 Project page: https://anaymajee.me/ assets/project_pages/smile.html." + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.811, + 0.733, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.129 + ], + "angle": 0, + "content": "Majee et al." + }, + { + "type": "image", + "bbox": [ + 0.251, + 0.149, + 0.753, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.321, + 0.788, + 0.364 + ], + "angle": 0, + "content": "Fig. 1: Functionality of components in \\( L_{comb} \\) proposed in the SMILE (ours) framework, (a) \\( L_{comb}^{inter} \\) promotes separation between \\( C_b \\) and \\( C_n \\) while (c) \\( L_{comb}^{intra} \\) promotes intra-class compactness." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.386, + 0.788, + 0.644 + ], + "angle": 0, + "content": "major challenges - Class Confusion and Catastrophic Forgetting. Class confusion, as highlighted in [29] manifests itself through mis-prediction of instances belonging to a newly learnt (novel) class, as one or more instances of the already learnt (base) classes. Authors in [1, 38] attribute this to the sharing of visual information between classes resulting in increased inter-class bias due to overlapping feature clusters as shown in Fig. 1(a). Catastrophic forgetting refers to the gradual degradation in the performance of already learnt classes in the quest to learn the novel ones, as shown in Fig. 2(a), seldom overfitting to rare classes [29, 38]. Further, large feature diversity (intra-class variance) among base classes lead to formation of non-discriminative feature clusters as shown in Fig. 1(b), aggravating the existing inter-class bias in the feature space. Unlike existing approaches (refer Sec. 2) which target either confusion or forgetting, our paper presents a unified approach to tackle both these challenges in FSOD. Although, recent approaches [1, 28, 34] attempt to tackle these challenges through contrastive learning strategies, such approaches have been limited by their capability to overcome either inter-class bias or intra-class variance [30, 38] and poor generalization to longtail settings [30] (FSOD being a extreme case)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this paper, we introduce a combinatorial viewpoint in FSOD considering each object class \\( i \\in [1, C] \\) in the dataset \\( \\mathcal{T} \\) as a set \\( A_i \\) of samples, where \\( \\mathcal{T} = \\{A_1, \\dots, A_C\\} \\), facilitating the application of combinatorial functions as learning objectives. We aim to overcome the aforementioned challenges through representation learning in the low-data regime by adopting this formulation through the SMILE: Submodular Mutual Information Learning framework, wherein we introduce novel, set-based combinatorial objective functions for FSOD as shown in Fig. 3. SMILE introduces a joint objective formulation \\( L_{comb} \\) (Eq. (3)) based on two popular flavors of submodular information functions - Submodular Mutual Information [22] (SMI) and Total Submodular Information [11] targeting the root causes of confusion and forgetting in FSOD. At first, SMILE is the first to introduce pairwise SMI functions \\( I_f \\) in representation learning which model the common (overlapping) information between two" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.403, + 0.115, + 0.733, + 0.128 + ], + "angle": 0, + "content": "SMILE: Submodular Mutual Information Learner" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.145, + 0.495, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.145, + 0.756, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.294, + 0.788, + 0.351 + ], + "angle": 0, + "content": "Fig. 2: Resilience to Catastrophic forgetting and faster convergence in SMILE over SoTA approaches. (a) shows that combinatorial losses in SMILE are robust to catastrophic forgetting, while (b) shows that objectives in SMILE results in faster convergence over SoTA FSOD methods (AGCM and DiGeo)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.367, + 0.789, + 0.639 + ], + "angle": 0, + "content": "sets. Minimizing \\( I_{f} \\) through the joint objective \\( L_{comb} \\) reduces feature overlap between base and novel classes alleviating inter-class bias in the model towards abundantly sampled classes as shown in Fig. 1(b). We extend this property of SMI functions to the novel classes minimizing the inter-cluster overlap between few-shot classes, promoting learning of discriminative features from just few samples. Secondly, SMILE preserves the diversity within each class by minimizing the total submodular information contained within each set as shown in Fig. 1(c), minimizing the impact of forgetting. This formulation closely follows the observation in [30] which models cooperation [15] between instances in a set by minimizing a submodular function over a set, to preserve representative features. The unified objective \\( L_{comb} \\) introduced in SMILE models both these necessary properties through a weighted sum of two distinct objectives \\( L_{comb}^{inter} \\) and \\( L_{comb}^{intra} \\) as shown in Fig. 3 balancing the tradeoff between inter-cluster separation and intra-cluster compactness respectively. This allows us to introduce a family of loss functions which inherently eliminates confusion and forgetting as shown in Tab. 5. We conduct our experiments on two popular FSOD benchmarks, PASCAL-VOC [6] and MS-COCO [26] for several few-shot settings and demonstrate the following contributions of SMILE:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.646, + 0.787, + 0.706 + ], + "angle": 0, + "content": "- SMILE introduces a novel set-based combinatorial viewpoint in FSOD by applying combinatorial Mutual Information based objective to discriminate between base and novel classes, in conjunction with submodular total information to minimize intra-class variance as the objective function." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.707, + 0.787, + 0.751 + ], + "angle": 0, + "content": "- SMILE generalizes to existing approaches in FSOD, irrespective of the underlying architecture demonstrating up to \\(5.7\\%\\) improvement in novel class performance (Tab. 2) over the baseline FSOD approach." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.751, + 0.786, + 0.781 + ], + "angle": 0, + "content": "SMILE demonstrates up to \\(2 \\times\\) faster convergence (Fig. 2(b)) over existing SoTA approaches resulting in faster generalization to unknown object classes." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.781, + 0.787, + 0.84 + ], + "angle": 0, + "content": "- Finally, SMILE demonstrates up to \\(11\\%\\) and \\(3.5\\%\\) reduction in class confusion and catastrophic forgetting while achieving SoTA performance on popular FSOD benchmarks like PASCAL-VOC (by \\(5.7\\%\\) on split 2, 10-shot setting) and MS-COCO (\\(5.4\\%\\) on 30-shot setting)." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.646, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.232, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Majee et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.39, + 0.161 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.386 + ], + "angle": 0, + "content": "Few-Shot Object Detection (FSOD): Classical FSOD approaches utilize finetuning [3] or distance metric learning [17] to adapt features to novel classes. Recent methods employ meta-learning techniques [16,40,41] with episodic training to learn class-specific features. Meta-Reweight [16] and Meta-RCNN [41] use additional feature extractors, while Add-Info [40] leverages feature differences between support and query images. Techniques like [43] enhance class-specific features through information sharing, and CME [24] aims to reduce class confusion. Attention mechanisms [7,44] are used to identify discriminative features. However, meta-learning approaches are resource-intensive and may fail to generalize to significantly different novel classes. Metric learning strategies like FsDet [37], FSCE [34], and SRR-FSD [45] offer better generalization without additional overheads. PNPDet [42] partially addresses catastrophic forgetting and class confusion. GFSD [8] proposes a Bias-Balanced RPN to mitigate overfitting in metric learners." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.387, + 0.788, + 0.583 + ], + "angle": 0, + "content": "Recent approaches like [18,32] adopt weak supervision from unlabelled data or low confidence predictions in RoI pooling layers to generalize to novel classes. These methods often use abundant samples from base classes [28] to prevent catastrophic forgetting, adding computational overhead in low-shot settings. Vision transformers [4] have been adopted in FSOD through methods like imTED [27] and PDC [23], with reduced computational overhead by using pre-trained attention heads. Alternatively, DiGeo [28] and PDC [23] learn the geometry or difference in distributions of RoI proposals [13] between object classes to overcome forgetting and confusion. However, these approaches rely on contrastive learning objectives [20] that struggle to learn discriminative feature embeddings due to adoption of pairwise similarity metrics. Our work, SMILE, aims to improve the feature learning capacity of existing SoTA approaches, irrespective of their underlying architectures." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Submodular Functions and Combinatorial Objectives: Submodular functions are recognized as set functions with an inherent diminishing returns characteristic. Defined as a set function \\( f: 2^{\\mathcal{V}} \\to \\mathbb{R} \\) operating on a ground-set \\( \\mathcal{V} \\), a function is termed submodular if it adheres to the condition \\( f(X) + f(Y) \\geq f(X \\cup Y) + f(X \\cap Y), \\forall X, Y \\subseteq \\mathcal{V} \\) [11]. These functions have garnered considerable attention in research, particularly in fields like data subset selection [22], active learning [21], and video summarization [19,22] through their ability in modeling concepts such as diversity, relevance, set-cover and representation. A subclass of submodular functions, namely Submodular Mutual Information (SMI) functions introduced in [22] model the similarity and diversity between pairs of object classes establishing itself as a powerful tool to model inter-class bias. Recently, Majee et al. [30] introduces these set-based combinatorial functions as objectives in representation learning and demonstrates their capability in overcoming inter-class bias (by minimizing the similarity between nonidentical object classes) and intra-class variance (maximizing the similarity between instances of the same object class). However, these functions are yet to be studied in the context of few-shot learning. We introduce novel instances of SMI based objectives in" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.403, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "SMILE: Submodular Mutual Information Learner" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.144, + 0.788, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.327, + 0.788, + 0.369 + ], + "angle": 0, + "content": "Fig. 3: Overview of our SMILE framework highlighting the application of Mutual Information function based objectives in SMILE for the fine-tuning stage of Few-Shot Object Detection." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.387, + 0.788, + 0.464 + ], + "angle": 0, + "content": "SMILEs to minimize inter-class bias between base and novel classes. To the best of our knowledge we are the first to introduce novel SMI based combinatorial objectives in conjunction with total information based combinatorial functions through SMILEs in a quest to minimize confusion and forgetting in few-shot object detection." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.485, + 0.33, + 0.5 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.515, + 0.662, + 0.53 + ], + "angle": 0, + "content": "3.1 Problem Definition : Few-Shot Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.789, + 0.844 + ], + "angle": 0, + "content": "We define a few-shot learner \\( h(x, \\theta) \\) as shown in Fig. 3 that receives input data \\( x \\) from base classes \\( C_b \\in [1, |C_b|] \\) and novel classes \\( C_n \\in [1, |C_n|] \\) such that \\( C = \\{C_b \\cup C_n\\} \\) and \\( \\{C_b \\cap C_n\\} = \\emptyset \\). Here, \\( \\theta \\) denotes the learnable parameters. The training data can be divided into two distinct parts, base \\( D_{base} \\) and novel \\( D_{novel} \\) such that, \\( \\mathcal{T} = \\{D_{base} \\cup D_{novel}\\} \\) and \\( \\{D_{base} \\cap D_{novel}\\} = \\emptyset \\). SMILe introduces a paradigm shift in FSOD by imbibing a combinatorial viewpoint, where the base dataset, \\( D_{base} = [A_1^b, A_2^b, \\dots, A_{|C_b|}^b] \\), containing abundant training examples from \\( C_b \\) base classes and the novel dataset, \\( D_{novel} = [A_1^n, A_2^n, \\dots, A_{|C_n|}^n] \\) containing only \\( K \\)-shot ( \\( |A_i^n| = K \\) for \\( i \\in [1, C_n] \\)) training examples from \\( C_n \\) novel classes. The objective of the few-shot learner \\( h(x, \\theta) \\) is to learn discriminative representation from classes in \\( D_{novel} \\) without degradation in performance on classes in \\( D_{base} \\). Following FSCE [34] we adopt a two-stage training strategy. In the base training stage we train \\( h(x, \\theta) \\) on abundant samples in \\( D_{base} \\), allowing the model to generalize on the domain of \\( D_{base} \\). The few-shot adaptation stage adapts \\( h(x, \\theta) \\) to previously unseen \\( K \\)-shot data by fine-tuning on data samples from \\( D_{base} \\cup D_{novel} \\) where \\( |A_k| = K \\) for \\( k \\in \\{C_b \\cup C_n\\} \\). The goal of SMILe is to overcome class confusion and forgetting in FSOD resulting from elevated inter-class bias and intra-class variance as observed in [1, 29, 38]. The final model \\( h(x, \\theta) \\) obtained after two training stages is evaluated on \\( D_{test} \\) containing unseen data samples from both \\( C_b \\cup C_n \\)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.129 + ], + "angle": 0, + "content": "Majee et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.455, + 0.161 + ], + "angle": 0, + "content": "3.2 The SMILE Framework" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.17, + 0.788, + 0.336 + ], + "angle": 0, + "content": "Adopting a combinatorial viewpoint as disclosed earlier allows us to employ submodular combinatorial functions as learning objectives to tackle confusion and forgetting in FSOD. As discussed in Sec. 2, minimizing a Submodular functions naturally models cooperation [15] while maximizing it models diversity [25] due to their inherent diminishing marginal returns property. SMILE adopts the aforementioned properties of submodular functions to define a novel family of combinatorial objective (loss) functions \\( L_{comb}(\\theta) \\) which enforces orthogonality in the feature space when applied on Region-of-Interest (RoI) features in FSOD models. The loss function \\( L_{comb}(\\theta) \\) can be decomposed into two major components - \\( L_{comb}^{inter} \\) minimizes inter-class bias between base and novel classes and \\( L_{comb}^{intra} \\) maximizes intra-class compactness within abundant classes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.337, + 0.788, + 0.519 + ], + "angle": 0, + "content": "For \\( L_{comb}^{inter} \\), SMILE explores a sub-category of combinatorial functions, namely Submodular Mutual Information (SMI) which can be defined as \\( I_{f}(A_{i}, A_{j}) = f(A_{i}) + f(A_{j}) - f(A_{i} \\cup A_{j}) \\) [11, 22], and models the common information between two sets \\( A_{i} \\) and \\( A_{j} \\), \\( \\forall i, j \\in \\mathcal{T} \\). Results in [11, 22] portray \\( I_{f}(A_{i}, A_{j}; \\theta) \\) as a measure of the degree of similarity between object classes \\( A_{i} \\) and \\( A_{j} \\). Adopting this definition of SMI, \\( L_{comb}^{inter} \\) minimizes the SMI between the base \\( C_{b} \\) and the novel \\( C_{n} \\) classes, ensuring sufficient inter-cluster separation (by minimizing inter-class bias) as shown in Eq. (1). \\( L_{comb}^{inter} \\) further minimizes the mutual information between classes in \\( C_{n} \\), minimizing inter-cluster overlaps between the novel classes. This is visually depicted in Fig. 1(b) and has been shown to be effective in mitigating class confusion in FSOD through our experiments in Sec. 4." + }, + { + "type": "equation", + "bbox": [ + 0.235, + 0.527, + 0.786, + 0.587 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {c o m b}} ^ {\\text {i n t e r}} (\\theta) = \\sum_ {\\substack {b \\in C _ {b} \\\\ n \\in C _ {n}}} I _ {f} (A _ {b}, A _ {n}; \\theta) + \\sum_ {\\substack {i, j \\in C _ {n} \\\\ i \\neq j}} I _ {f} (A _ {i}, A _ {j}; \\theta) = \\sum_ {\\substack {i \\in (C _ {b} \\cup C _ {n}) \\\\ j \\in C _ {n}: i \\neq j}} I _ {f} (A _ {i}, A _ {j}; \\theta) \\tag{1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.595, + 0.789, + 0.844 + ], + "angle": 0, + "content": "In addition to confusion which stems from inter-class bias, SMILEs aims at mitigating catastrophic forgetting [29] in FSOD which has been attributed to large intra-class variance among abundant object classes in [1,38]. In coherence to the combinatorial formulation in SMILEs we achieve this through \\( L_{comb}^{intra} \\) which minimizes the Total Submodular Information, defined as \\( S_f(A_1, \\dots, A_{|C|}) = \\sum_{k=1}^{|C|} f(A_k; \\theta) \\), over sets \\( A_k \\in \\mathcal{T} \\), given a submodular function \\( f(A_k; \\theta) \\). As discussed earlier, minimizing the submodular information models cooperation which asserts that minimizing \\( L_{comb}^{inter} \\) promotes learning of discriminative feature clusters, penalizing abundant classes to have large feature variance in the embedding space as shown in Fig. 1(c). Although submodular functions have been studied in the field of representation learning to minimize intra-class variance in [30], but primarily differs from SMILEs in modeling a longtail recognition task by minimizing the total submodular correlation, which models gain in information when new features are added to a set. The formulation of \\( L_{comb}^{intra} \\) has been shown in Eq. (2) where we minimize the total submodular information within samples in each class in \\( C_b \\cup C_n \\) and our experiments in Tab. 5" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.403, + 0.115, + 0.733, + 0.128 + ], + "angle": 0, + "content": "SMILE: Submodular Mutual Information Learner" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.788, + 0.178 + ], + "angle": 0, + "content": "show the effectiveness of \\( L_{comb}^{intra} \\) in boosting base class performance asserting the mitigation of catastrophic forgetting." + }, + { + "type": "equation", + "bbox": [ + 0.287, + 0.19, + 0.786, + 0.223 + ], + "angle": 0, + "content": "\\[\nL _ {c o m b} ^ {i n t r a} (\\theta) = \\sum_ {b \\in C _ {b}} f (A _ {b}, \\theta) + \\sum_ {n \\in C _ {n}} f (A _ {n}, \\theta) = \\sum_ {k \\in \\left(C _ {b} \\cup C _ {n}\\right)} f (A _ {k}, \\theta) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.234, + 0.786, + 0.263 + ], + "angle": 0, + "content": "Ablating on the choice of the submodular function \\( f \\) and SMI functions \\( I_{f} \\) we introduce several instances of SMILE objectives as discussed in Tab. 1." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.264, + 0.787, + 0.326 + ], + "angle": 0, + "content": "Encapsulating the aforementioned formulations of \\( L_{comb}^{inter} \\) and \\( L_{comb}^{intra} \\) in SMILE we define a joint objective \\( L_{comb}(\\theta) \\) which tackles both the challenges of confusion and forgetting. We thus define \\( L_{comb}(\\theta) \\) in Eq. (3) which is the weighted algebraic sum of \\( L_{comb}^{inter} \\) and \\( L_{comb}^{intra} \\) with the weighting factor \\( \\eta \\)." + }, + { + "type": "equation", + "bbox": [ + 0.3, + 0.339, + 0.786, + 0.409 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {c o m b} (\\theta) = (1 - \\eta) L _ {c o m b} ^ {i n t r a} (\\theta) + \\eta L _ {c o m b} ^ {i n t e r} (\\theta) \\\\ = \\sum_ {i \\in C _ {b} \\cup C _ {n}} \\left[ (1 - \\eta) f \\left(A _ {i}, \\theta\\right) + \\eta \\sum_ {\\substack {j \\in C _ {n} \\\\ i \\neq j}} I _ {f} \\left(A _ {i}, A _ {j}; \\theta\\right) \\right] \\tag{3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.42, + 0.787, + 0.51 + ], + "angle": 0, + "content": "Note, that the combinatorial objective \\( L_{comb}(\\theta) \\) is applied on output features from the RoI Pooling layers in proposal-based [31,38] architectures. To promote adoption of SMILE agnostic of the backbone architecture we introduce a combinatorial head \\( Z_{comb} = Comb(h,\\theta) \\) which projects the RoI features to 128-dimensional feature vectors [20], \\( Z_{comb} \\) on which \\( L_{comb}(\\theta) \\) is applied during the few-shot adaptation stage." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.511, + 0.787, + 0.617 + ], + "angle": 0, + "content": "Finally, we summarize the total classification loss in SMILEs as depicted in Eq. (4) as the sum over all three objectives: the classification head \\( L_{Clf} \\), the box regression head \\( L_{bbox} \\) and the combinatorial head \\( L_{comb}(\\theta) \\). Note that the objectives proposed in SMILEs apply only to \\( Comb(h,\\theta) \\) while the RoI classification and regression heads are unchanged. This follows the observations in [28,38] which warrants the boost in performance originating from learning robust feature representations for each RoI predicted by the model." + }, + { + "type": "equation", + "bbox": [ + 0.36, + 0.63, + 0.786, + 0.647 + ], + "angle": 0, + "content": "\\[\nL _ {c l s} (\\theta) = L _ {C l f} (\\theta) + L _ {b b o x} (\\theta) + L _ {c o m b} (\\theta) \\tag {4}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.668, + 0.757, + 0.687 + ], + "angle": 0, + "content": "3.3 Instantiations of \\(L_{comb}^{inter}\\) and \\(L_{comb}^{intra}\\) in the SMILE Framework" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.694, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Given a submodular function \\( f(A) \\) and a Submodular Mutual Information (SMI) function \\( I_{f}(A,Q) \\) over sets \\( A \\) and \\( Q \\), we derive two instances \\( L_{comb}^{inter} \\) and \\( L_{comb}^{intra} \\) objectives in SMILEs. Depending on the choice of \\( f(A) \\) we define two instances: Facility-Location Mutual Information (SMILE-FLMI) and Graph-Cut Mutual Information (SMILE-GCMI). Inherently, both objectives adopt the cosine similarity metric \\( S_{ij}(\\theta) \\) as used in SupCon [20] which can be defined as \\( S_{ij}(\\theta) = \\frac{Z_{comb_i}^{\\mathrm{T}} \\cdot Z_{comb_j}}{||Z_{comb_i}|| \\cdot ||Z_{comb_j}||} \\) to compute similarity between sets in the learning objective. Although the similarity kernel used in SMILEs is computed in a pairwise fashion, objectives defined under \\( L_{comb} \\) use it to only compute feature interactions" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.129 + ], + "angle": 0, + "content": "Majee et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 1: Summary of various instantiations of SMILE highlighting the components of the combinatorial objective, \\( L_{comb}^{inter} \\) and \\( L_{comb}^{intra} \\)." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.185, + 0.787, + 0.253 + ], + "angle": 0, + "content": "
ObjectiveInstances of Lintercomb(θ)Instances of Intracomb(θ)
SMILE-GCMI (ours)∑k∈(Ck∪Cn)2λ∑iAksjA1Sij(θ)∑k∈Ck∪Cn∑iAksjTsetAksij(θ) - λ∑i,jAksij(θ)
∑k∈(Ck∪Cn)2λ∑iAksjA1maxSij(θ) + λ∑iAksjA1maxSij(θ)∑k∈Ck∪Cn∑iTsetAksij(θ)
SMILE-FLMI (ours)
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.272, + 0.789, + 0.318 + ], + "angle": 0, + "content": "between samples, differing from existing approaches in aggregation of pairwise similarities to compute total information and mutual information over classes in \\(\\mathcal{T}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.341, + 0.789, + 0.482 + ], + "angle": 0, + "content": "SMILE-FLMI based objective is derived from the Facility-Location Mutual Information (FLMI) [22] function, expressed as \\( I_{f}(Q,A) = \\sum_{i\\in Q}\\max_{j\\in A}S_{ij}(\\theta) + \\lambda \\sum_{i\\in A}\\max_{j\\in Q}S_{ij}(\\theta) \\) and minimizes the maximum similarity (most similar) between sets \\( Q \\) and \\( A \\). Given the facility-location (FL) submodular function \\( f(A,\\theta) = \\sum_{i\\in \\mathcal{T}}\\max_{j\\in A}S_{ij}(\\theta) \\) over the set \\( A \\), we can derive \\( L_{comb}^{inter}(\\theta) \\) and \\( L_{comb}^{intra}(\\theta) \\) shown in Eq. (5) as the SMILE-FLMI objective. Note that \\( L_{comb}^{inter}(\\theta) \\) is applied between object classes in \\( C_b\\cup C_n \\) and \\( C_n \\) while \\( L_{comb}^{intra}(\\theta) \\) is applied over all classes in \\( C_b\\cup C_n \\)." + }, + { + "type": "equation", + "bbox": [ + 0.301, + 0.492, + 0.785, + 0.54 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {c o m b} ^ {i n t e r} (\\theta) = \\sum_ {k \\in \\left(C _ {b} \\cup C _ {n}\\right)} \\sum_ {i \\in A _ {k}} \\max _ {j \\in A _ {l}} S _ {i j} (\\theta) + \\lambda \\sum_ {i \\in A _ {l}} \\max _ {j \\in A _ {k}} S _ {i j} (\\theta), \\\\ l \\in C _ {n}: k \\neq l \\tag {5} \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.301, + 0.54, + 0.572, + 0.571 + ], + "angle": 0, + "content": "\\[\nL _ {c o m b} ^ {i n t r a} (\\theta) = \\sum_ {k \\in C _ {b} \\cup C _ {n}} \\sum_ {i \\in \\mathcal {T} \\backslash A _ {k}} \\max _ {j \\in A _ {k}} S _ {i j} (\\theta)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.585, + 0.789, + 0.709 + ], + "angle": 0, + "content": "Minimizing the \\(L_{comb}^{inter}\\) objective function ensures that the sets \\(A_{l} \\in C_{n}\\) and \\(A_{k} \\in C_{b} \\cup C_{n}\\) are disjoint by minimizing the similarity between features in \\(A_{k}\\) and the hardest negative \\((\\sum_{i \\in A_{k}} \\max_{j \\in A_{l}} S_{ij}(\\theta)\\) for \\(k \\in C_{b} \\cup C_{n}\\) and \\(l \\in C_{n})\\) feature vectors in \\(A_{l}\\). Further, \\(L_{comb}^{inter}\\) enforces sufficient separation between the novel classes themselves to promote learning of disjoint feature clusters even with few-shot data overcoming confusion. Additionally, \\(L_{comb}^{intra}\\) minimizes the total information contained in each set \\(A_{k} \\in (C_{b} \\cup C_{n})\\). This objective retains discriminative feature information from each class in \\(\\mathcal{T}\\) reducing the impact of forgetting." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.732, + 0.789, + 0.843 + ], + "angle": 0, + "content": "SMILE-GCMI based objective described in Eq. (6) minimizes the pairwise similarity of feature vectors between a positive set \\( A_{k} \\in C_{b} \\cup C_{n} \\) and the sets in \\( A_{l} \\in C_{n} \\) while maximizing the similarity between features in each set \\( A_{k} \\in C_{b} \\cup C_{n} \\). Given two sets \\( Q \\) and \\( A \\), [22] defines the Graph-Cut SMI to be \\( I_{f}(Q, A) = 2\\lambda \\sum_{i \\in Q} \\sum_{j \\in A} S_{ij}(\\theta) \\), where the Graph-Cut function over a set \\( A \\) is given by \\( f(A, \\theta) = \\sum_{i \\in A} \\sum_{j \\in T \\setminus A_{k}} S_{ij}(\\theta) - \\lambda \\sum_{i,j \\in A} S_{ij}(\\theta) \\). Given the Graph-Cut and the Graph-Cut SMI functions, we derive \\( L_{comb}^{inter}(\\theta) \\) and \\( L_{comb}^{intra}(\\theta) \\)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.403, + 0.115, + 0.733, + 0.128 + ], + "angle": 0, + "content": "SMILE: Submodular Mutual Information Learner" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.788, + 0.193 + ], + "angle": 0, + "content": "shown in Eq. (6) as the SMILE-GCMI objective. Similar to SMILE-FLMI, the \\( L_{comb}^{inter}(\\theta) \\) is applied between object classes in \\( C_b \\cup C_n \\) and \\( C_n \\) while \\( L_{comb}^{intra}(\\theta) \\) is applied over all classes in \\( C_b \\cup C_n \\)." + }, + { + "type": "equation", + "bbox": [ + 0.308, + 0.201, + 0.785, + 0.248 + ], + "angle": 0, + "content": "\\[\nL _ {c o m b} ^ {i n t e r} (\\theta) = \\sum_ {\\substack {k \\in (C _ {b} \\cup C _ {n}) \\\\ l \\in C _ {n}: k \\neq l}} 2 \\lambda \\sum_ {i \\in A _ {k}} \\sum_ {j \\in A _ {l}} S _ {i j} (\\theta), \\tag{6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.308, + 0.247, + 0.697, + 0.28 + ], + "angle": 0, + "content": "\\[\nL _ {c o m b} ^ {i n t r a} (\\theta) = \\sum_ {k \\in C _ {b} \\cup C _ {n}} \\sum_ {i \\in A _ {k}} \\sum_ {j \\in \\mathcal {T} \\backslash A _ {k}} S _ {i j} (\\theta) - \\lambda \\sum_ {i, j \\in A _ {k}} S _ {i j} (\\theta)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.287, + 0.784, + 0.406 + ], + "angle": 0, + "content": "Although objectives in SMILE-FLMI and SMILE-GCMI are tasked with similar functions, the \\( L_{comb}^{inter} \\) in SMILE-GCMI minimizes the pairwise similarity between sets in \\( C_b \\cup C_n \\) and \\( C_n \\) rather than the most similar set in SMILE-FLMI. Further, the \\( L_{comb}^{intra} \\) in SMILE-GCMI scales linearly with size of \\( A_k \\) as described in [30]. This does not allow the model to substantially improve performance on learning discriminative feature representations for classes in both \\( C_b \\) and \\( C_n \\) as the \\( |A_k| = K \\) (number of shots) thus failing to outperform the model trained using SMILE-FLMI." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.409, + 0.784, + 0.453 + ], + "angle": 0, + "content": "The detailed derivations of the aforementioned instances are included in the Supplementary material. Our experiments in Sec. 4.4 elucidates the fact that SMILE-FLMI is a better choice to overcome forgetting and confusion in FSOD." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.474, + 0.376, + 0.492 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.503, + 0.784, + 0.548 + ], + "angle": 0, + "content": "We evaluate models in SMILE by adopting standard evaluation criterion in FSOD [16, 37] and report the Mean Average Precision (mAP) at \\(50\\%\\) Intersection Over Union (IoU) for all our experiments." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.568, + 0.427, + 0.584 + ], + "angle": 0, + "content": "4.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.59, + 0.784, + 0.621 + ], + "angle": 0, + "content": "Datasets We evaluate our proposed SMILE approach on two few-shot object detection datasets - and PASCAL-VOC [5] and MS-COCO [26] datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.632, + 0.785, + 0.737 + ], + "angle": 0, + "content": "PASCAL-VOC [5] dataset consists of 20 classes, out of which 15 are considered as base and 5 as novel classes. The novel classes are chosen at random giving rise to three data splits namely, split-1 (bird, bus, cow, motorbike, sofa), split-2 (aeroplane, bottle, cow, horse, sofa) and split-3 (boat, cat, motorbike, sheep, sofa). Following previous works [16], we use the combined VOC \\(07 + 12\\) datasets for training and evaluate our models on the complete validation set of VOC 2007 for 1, 5, and 10 shot settings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.784, + 0.839 + ], + "angle": 0, + "content": "MS-COCO [26] dataset consists of 80 classes, out of which 60 are considered as base and 20 as novel classes. Following existing approaches in FSOD [41] we randomly select 5k samples from \\((D_{base} \\cup D_{novel})\\) to use as the validation set while the remaining samples are used to generate random 10 and 30-shot splits for training of the MS-COCO 2014 dataset. The key difference between VOC and COCO is the large intra-class variance and class-imbalance in COCO." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.129 + ], + "angle": 0, + "content": "Majee et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 2: Quantitative analysis on PASCAL-VOC dataset: Few-shot object detection performance (\\( mAP_{novel} \\)) on novel class splits of PASCAL-VOC dataset. We tabulate results for \\( K = 1, 5, 10 \\) shots from various SoTA techniques in FSOD. * indicates that the results are averaged over 10 random seeds. † indicates a meta-learning strategy (N-way, K-shot training)." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.227, + 0.79, + 0.468 + ], + "angle": 0, + "content": "
MethodLearner TypeBackboneSplit 1Split 2Split 3
K=151015101510
†Meta-RCNN [41]MetaFRCN-R10119.945.751.510.434.845.414.341.248.1
†Meta-Reweight [16]MetaYOLO V214.833.947.215.730.140.521.342.845.9
†MetaDet [38]MetaFRCN-R10118.936.849.621.831.743.020.643.944.1
†Add-Info [40]MetaFRCN-R10124.249.157.421.637.045.721.243.849.6
†CME [24]MetaYOLO V217.844.847.512.733.740.015.744.948.8
PNPDet [42]MetricDLA-3418.2-41.016.6-36.418.9-36.2
FsDet w/ FC [37]MetricFRCN-R10136.855.757.018.235.539.027.748.750.2
FsDet w/ cos [37]MetricFRCN-R10139.855.756.023.535.139.130.849.549.8
Retentive-RCNN [9]MetricFRCN-R10140.153.756.121.737.040.330.249.750.1
FSCE [34]MetricFRCN-R10141.057.457.827.344.449.840.153.257.7
\\( FSCE + SMILE \\) (ours)Comb.FRCN-R10141.257.961.129.244.650.541.355.659.0
AGCM [1]MetricFRCN-R10140.358.559.927.549.350.642.154.258.2
\\( AGCM + SMILE \\) (ours)Comb.FRCN-R10140.959.762.031.949.552.342.656.461.4
DiGeo [28]MetricFRCN-R10136.054.160.920.742.847.127.547.352.9
DiGeo + SMILE(ours)Comb.FRCN-R10136.156.662.326.544.147.333.151.956.4
imTED [27]MetricViT-B31.971.977.022.752.257.712.669.672.8
imTED + PDC [23]MetricViT-B36.673.177.115.551.856.018.967.972.8
PDC + SMILE (ours)Comb.ViT-B36.675.277.927.152.758.315.170.074.7
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.494, + 0.789, + 0.678 + ], + "angle": 0, + "content": "Implementation Details The SMILE framework adopts a architecture agnostic approach and adopt several backbones including Faster-RCNN [31] and ViT [27]. For VOC, the input batch size to the network is set to 16 and 2 in the base training and few-shot adaptation stages for Faster-RCNN and ViT based approaches. The input resolution is set to \\(764 \\times 1333\\) pixels for data splits in COCO, while it is set to \\(800 \\times 600\\) pixels for PASCAL-VOC. The hyper-parameters used in the formulation of SMILE, namely \\(\\eta\\) and similarity kernel \\(S\\), are chosen through ablation experiments described in Sec. 4.4. Results from existing methods are a reproduction of the algorithm from publicly available codebases. All our experiments are performed on 4 NVIDIA GTX 1080 Ti GPUs with additional details in the supplementary material and code released at https://github.com/amajee11us/SMILE-FSOD.git." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.696, + 0.631, + 0.71 + ], + "angle": 0, + "content": "4.2 Results on Few-Shot PASCAL VOC Dataset" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Table 2 records the results obtained from our SMILE framework on novel splits of the PASCAL-VOC dataset and contrasts it against SoTA FSOD techniques. We adopt four SoTA approaches FSCE [34], AGCM [1], DiGeo [28] and imTED [27], covering several backbone architectures Faster-RCNN + FPN (FSCE, AGCM and DiGeo) alongside ViT (imTED, PDC) and introducing SMILE (\\(M + SMILE\\)) approach into existing architectures \\(M\\). For Faster-RCNN based architectures (FSCE) we show a maximum of \\(5.7\\%\\) (3.3 map points) improvement while for FPN based architectures (AGCM and DiGeo) we show a \\(3.5\\%\\) (2.1 map points" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.403, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "SMILE: Submodular Mutual Information Learner" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.374 + ], + "angle": 0, + "content": "for AGCM+SMILE) improvement. It is interesting to note that unlike FSCE and AGCM, DiGeo uses abundant samples from \\( C_b \\) alongside few-shot samples in \\( C_n \\) (with upsampling) during finetuning introducing a large inter-class bias. SMILE outperforms DiGeo by up to 2.3 \\( mAP \\) points (split 2, 10-shot) showing the resilience of SMILE towards imbalance, thus overcoming confusion in FSOD. Additionally, for recently introduced transformer based architectures (imTED + PDC [23]) SMILE outperforms the existing SoTA with a maximum improvement of 4.9 \\( mAP \\) points (split 2, 5-shot) thus establishing SMILE as the SoTA on few-shot splits of VOC. Note, that the choice of objective functions \\( L_{comb}^{inter} \\) and \\( L_{comb}^{intra} \\) for this experiment has been determined to be SMILE+FLMI through an ablation on the different instances in SMILEs as described in Sec. 4.4. Finally, Fig. 2(b) shows that introduction of SMILE framework to existing SoTA approaches leads to rapid convergence on the novel classes up to 2x over existing SoTA. This is significant for mission critical tasks like autonomous driving where the model is required to rapidly learn novel objects to reduce turn-around time." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.399, + 0.597, + 0.415 + ], + "angle": 0, + "content": "4.3 Results on Few-Shot MS-COCO Dataset" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.428, + 0.789, + 0.593 + ], + "angle": 0, + "content": "Similar to the results in PASCAL VOC we demonstrate the results of our SMILE framework on MS-COCO dataset. In contrast to VOC, COCO presents an extremely imbalanced setting with a long-tail distribution within \\( D_{base} \\) itself making it really hard for FSOD approaches to achieve SoTA through primitive objective functions. Following the ablation experiments in Sec. 4.4 we adopt the SMILE+FLMI objective (best performing) to conduct the experiments on 20 few-shot classes of MS-COCO dataset. We show that SMILE generalizes existing SoTA approach (imTED + PDC) for COCO dataset by \\( 5.4\\% \\) (2.6 mAP points, 30-shot setting). This further establishes the generalizability of our approach over varying data distributions (VOC and COCO) while achieving SoTA in FSOD tasks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.62, + 0.39, + 0.636 + ], + "angle": 0, + "content": "4.4 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.648, + 0.787, + 0.695 + ], + "angle": 0, + "content": "We conduct ablation experiments on the 10-shot split of VOC (split 1) with hyper-parameters \\(\\eta = 0.5\\), cosine similarity metric and \\(\\lambda = 1.0\\). Ablations for hyper-parameters are detailed in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Components of SMILE Instantiations in SMILEs consists of two main components - \\( L_{comb}^{inter} \\) and \\( L_{comb}^{intra} \\). We consider three baselines FSCE, AGCM and DiGeo which follow the Faster-RCNN/FPN backbone for this experiment. First, we introduce \\( L_{comb}^{intra} \\) by adopting the FL based objective as determined through ablation experiments below. This objective models intra-class variance and ensures reduction in intra-class variance characterized by boost in base class performance. Secondly, following the SMILE-FLMI formulation in Eq. (5) we introduce \\( L_{comb}^{inter} \\) during the few-shot adaptation stage. Applying this objective minimizes" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.129 + ], + "angle": 0, + "content": "Majee et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 3: Performance of SMILE on MS COCO dataset : Our SMILE objectives demonstrate better generalizability while outperforming SoTA FSOD approaches on novel class performance \\( mAP_{50} \\) (novel)." + }, + { + "type": "table", + "bbox": [ + 0.293, + 0.199, + 0.71, + 0.381 + ], + "angle": 0, + "content": "
Method\\( {mAP} \\)\\( {mAP}_{50} \\)\\( {mAP}_{75} \\)\\( {mAP} \\)\\( {mAP}_{50} \\)\\( {mAP}_{75} \\)
10-shot30-shot
Meta-Reweight [16]5.612.34.69.119.07.6
Meta-RCNN [41]8.719.16.612.425.310.8
TFA w/cos [37]10.0-9.313.7-13.4
Add-Info [40]12.527.39.814.730.612.2
MPSR [39]9.817.99.714.125.414.2
FSCE [34]11.9-10.516.4-16.2
FADI [2]12.222.711.916.129.115.8
CME [24]15.124.616.416.2--
FCT [14]17.130.217.021.435.522.1
imTED-B [27]22.536.623.730.247.432.5
imTED-B+PDC [23]23.438.124.530.847.333.5
PDC + SMILE (ours)25.840.126.131.049.933.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.247, + 0.396, + 0.754, + 0.411 + ], + "angle": 0, + "content": "Table 4: Ablation on various components of the proposed SMILE approach." + }, + { + "type": "table", + "bbox": [ + 0.312, + 0.422, + 0.688, + 0.618 + ], + "angle": 0, + "content": "
MethodBaselinef(Ai)(Lintra)If(Ai,Aj)(Lintercomb)mAPbasemAPnovel
FsDet w/ cos---23.639.8
FSCE✓86.157.8
✓✓89.660.1
✓✓88.361.0
✓✓✓89.861.1
AGCM✓87.658.0
✓✓88.661.3
✓✓88.961.8
✓✓✓89.361.8
DiGeo✓90.560.9
✓✓92.361.7
✓✓91.462.0
✓✓✓92.662.3
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.648, + 0.788, + 0.769 + ], + "angle": 0, + "content": "the inter-class bias between \\( C_b \\cup C_n \\) and \\( C_n \\), thus improving novel class performance significantly. Nevertheless, we see a slight drop in base class performance due to forgetting prevalent in FSOD tasks. Finally, we combine both instantiations in SMILE into one single objective as in Eq. (3) with \\( \\eta = 0.5 \\) and show that SMILE improves both base and novel class performance emerging as the best choice for FSOD. To demonstrate generalization, we perform this experiment on several SoTA approaches as baseline and show that the results discussed in aforementioned section hold. We summarize all the results in Tab. 4." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Choice of Combinatorial functions in \\( L_{comb} \\) SMILE introduces several instances of \\( L_{comb}^{intra} \\) and \\( L_{comb}^{inter} \\). To clearly understand the contributions of each of these instances we conduct experiments tabulated in Tab. 5 and determine the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.403, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "SMILE: Submodular Mutual Information Learner" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.785, + 0.173 + ], + "angle": 0, + "content": "Table 5: Ablation on the choice of Submodular Information function \\( I_{f} \\) and Submodular Information function \\( f \\) for \\( L_{comb} \\) in SMILE." + }, + { + "type": "table", + "bbox": [ + 0.359, + 0.185, + 0.645, + 0.27 + ], + "angle": 0, + "content": "
Modelf(A,θ)If(Ai,Aj)mAPbasemAPnovel
DiGeo [28]--87.957.4
GCGCMI92.660.9
FLFLMI93.162.3
FSCE [34]--86.157.8
GCGCMI87.460.3
FLFLMI89.861.1
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.298, + 0.788, + 0.45 + ], + "angle": 0, + "content": "best performing formulation which generalizes to existing FSOD architectures. Unlike other ablation experiments in Sec. 4.4, we conduct our experiments on DiGeo which introduces an extremely imbalanced scenario by using abundant samples in \\( D_{base} \\). We conclude that SMILE-FLMI which considers the Facility-Location based objective as \\( L_{comb}^{intra} \\) and Facility-Location Mutual Information based objective as \\( L_{comb}^{inter} \\) as the best performing instantiation. This result follows the formulation in Eq. (5) where FL naturally models intra-class compactness in class-imbalanced settings [30] while FLMI penalizes the classes in \\( C_b \\) to learn overlapping feature representations with the classes in \\( C_n \\). We use SMILE-FLMI for all benchmark experiments in Sec. 4.2 and Sec. 4.3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.471, + 0.789, + 0.744 + ], + "angle": 0, + "content": "Robustness to Catastrophic Forgetting One of the most significant challenges in FSOD is the elimination of catastrophic forgetting which manifests as the degradation in the performance of classes in \\( C_b \\) while learning classes in \\( C_n \\). This primarily occurs due to the lack of discriminative feature representations from instances in \\( D_{base} \\) during the few-shot adaptation (stage 2) stage. We plot the change in base class performance as the training progresses in existing SoTA methods AGCM and DiGeo against number of training iterations in Fig. 2(a). At first, we contrast the change in base class performance \\( mAP_{base} \\) between AGCM and AGCM+SMILE and observe that AGCM overfits on the few-shot samples in \\( D_{base} \\) reducing the performance on \\( C_{base} \\) as the training progresses. AGCM + SMILE on the other hand better retains the performance on base classes with \\( \\sim 3.5\\% \\) better retention in base class performance. Interestingly, DiGeo is able to retain most of the base class performance with a very small degradation over the roofline (a model trained with only the base classes until convergence). Our DiGeo+SMILE approach outperforms Digo by demonstrating base class performance even higher than the roofline establishing the supremacy of \\( L_{comb} \\) in overcoming inter-class bias and intra-class variance resulting in robustness against catastrophic forgetting." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Overcoming Class Confusion Figure 4 highlights the supremacy of the proposed SMILE framework in mitigating class confusion through confusion matrix plots. We compare the confusion between classes in \\( C_b \\cup C_n \\) of SoTA approaches AGCM and DiGeo before and after introduction of combinatorial objectives in SMILE. Although both approaches use K-shot examples for classes in \\( C_n \\), Di" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.129 + ], + "angle": 0, + "content": "Majee et al." + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.147, + 0.363, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.147, + 0.496, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.496, + 0.15, + 0.629, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.629, + 0.148, + 0.781, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.277, + 0.788, + 0.32 + ], + "angle": 0, + "content": "Fig. 4: Ablation on Overcoming Class Confusion in SMILE. (a,b) SMILE demonstrates \\(11\\%\\) lower confusion over AGCM and (c,d) \\(4\\%\\) lower confusion over DiGeo. Only significant numbers are highlighted. Best viewed in \\(200\\%\\) zoom." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.337, + 0.789, + 0.518 + ], + "angle": 0, + "content": "Geo differs from AGCM by adopting an upsampling strategy which allows the utilization of abundant examples in \\( C_b \\) while upsampling the instances in \\( C_n \\). This injects different degrees of inter-class biases for models trained by adopting AGCM and DiGeo which has been demonstrated as the primary reason for confusion in previous work [29]. At first, we observe from Fig. 4 that by adopting the upsampling based strategy, DiGeo achieves very low confusion between already learnt base classes, leading to significantly lower confusion (5% among \\( C_b \\) and \\( C_n \\)). Further, confusion matrix plots in Fig. 4 show that AGCM+SMILE demonstrates 11% lower confusion than AGCM and DiGeo+SMILE shows 4% lower confusion and DiGeo. This proves the efficacy of combinatorial objectives (\\( L_{comb}^{inter} \\)) in mitigating inter-class bias, thereby reducing confusion between classes." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.562, + 0.36, + 0.578 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.789, + 0.841 + ], + "angle": 0, + "content": "In this work, we have presented a novel approach to Few-Shot Object Detection (FSOD) by introducing a combinatorial viewpoint through the SMILE framework. By leveraging the properties of set-based combinatorial functions, SMILE aims to address the challenges of class confusion and catastrophic forgetting, which are prevalent in FSOD tasks. Our approach incorporates Submodular Mutual Information (SMI) and Submodular Information Measures (SIM) to penalize overlapping features between base and novel classes and to ensure the formation of compact feature clusters, respectively. The experimental results on PASCAL-VOC and MS-COCO benchmarks demonstrate the effectiveness of SMILE, showing significant improvements in novel class performance, faster convergence, and a reduction in class confusion and catastrophic forgetting. Overall, SMILE offers a promising direction for advancing the state-of-the-art in FSOD by providing a generalized framework that is adaptable to various underlying architectures and capable of handling the complexities associated with few-shot learning in object detection." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.403, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "SMILE: Submodular Mutual Information Learner" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.283 + ], + "angle": 0, + "content": "We gratefully thank anonymous reviewers for their valuable comments. This work is supported by the National Science Foundation under Grant Numbers IIS-2106937, a gift from Google Research, an Amazon Research Award, and the Adobe Data Science Research award. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the National Science Foundation, Google or Adobe." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.308, + 0.323, + 0.323 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.34, + 0.785, + 0.395 + ], + "angle": 0, + "content": "1. Agarwal, A., Majee, A., Subramanian, A., Arora, C.: Attention guided cosine margin to overcome class-imbalance in few-shot road object detection. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) Workshops. pp. 221-230 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.396, + 0.785, + 0.437 + ], + "angle": 0, + "content": "2. Cao, Y., Wang, J., Jin, Y., Wu, T., Chen, K., Liu, Z., Lin, D.: Few-shot object detection via association and discrimination. In: Thirty-Fifth Conference on Neural Information Processing Systems (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.438, + 0.785, + 0.465 + ], + "angle": 0, + "content": "3. Chen, H., Wang, Y., Wang, G., Qiao, Y.: LSTD: A Low-Shot Transfer Detector For Object Detection. In: AAAI. pp. 2836-2843 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.466, + 0.785, + 0.52 + ], + "angle": 0, + "content": "4. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.521, + 0.785, + 0.548 + ], + "angle": 0, + "content": "5. Everingham, M., Van Gool, L., Williams, C.K.I., Winn, J., Zisserman, A.: The Pascal Visual Object Classes (VOC) Challenge. IJCV pp. 303-338 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.549, + 0.785, + 0.589 + ], + "angle": 0, + "content": "6. Everingham, M., Van Gool, L., Williams, C., Winn, J., Zisserman, A.: The Pascal visual object classes (voc) challenge. International Journal of Computer Vision 88, 303-338 (06 2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.59, + 0.785, + 0.618 + ], + "angle": 0, + "content": "7. Fan, Q., Zhuo, W., Tang, C.K., Tai, Y.W.: Few-Shot Object Detection With Attention-RPN And Multi-Relation Detector. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.619, + 0.785, + 0.659 + ], + "angle": 0, + "content": "8. Fan, Z., Ma, Y., Li, Z., Sun, J.: Generalized Few-Shot Object Detection Without Forgetting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4527-4536 (June 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.66, + 0.785, + 0.687 + ], + "angle": 0, + "content": "9. Fan, Z., Ma, Y., Li, Z., Sun, J.: Generalized few-shot object detection without forgetting (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.688, + 0.785, + 0.715 + ], + "angle": 0, + "content": "0. Finn, C., Abbeel, P., Levine, S.: Model-Agnostic Meta-Learning For Fast Adaptation Of Deep Networks. In: ICML (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.716, + 0.767, + 0.729 + ], + "angle": 0, + "content": "1. Fujishige, S.: Submodular Functions and Optimization, vol. 58. Elsevier (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.73, + 0.785, + 0.756 + ], + "angle": 0, + "content": "2. Gidaris, S., Komodakis, N.: Dynamic Few-Shot Visual Learning Without Forgetting. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.757, + 0.526, + 0.771 + ], + "angle": 0, + "content": "3. Girshick, R.B.: Fast R-CNN. ICCV (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.785, + 0.812 + ], + "angle": 0, + "content": "4. Han, G., Ma, J., Huang, S., Chen, L., Chang, S.F.: Few-shot object detection with fully cross-transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5321-5330 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "5. Jegelka, S., Bilmes, J.: Submodularity beyond submodular energies: Coupling edges in graph cuts. In: CVPR 2011 (2011)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.34, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "Majee et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "16. Kang, B., Liu, Z., Wang, X., Yu, F., Feng, J., Darrell, T.: Few-shot Object Detection Via Feature Reweighting. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.218 + ], + "angle": 0, + "content": "17. Karlinsky, L., Shtok, J., Harary, S., Schwartz, E., Aides, A., Feris, R., Giryes, R., Bronstein, A.M.: RepMet: Representative-Based Metric Learning For Classification And Few-Shot Object Detection. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.219, + 0.785, + 0.261 + ], + "angle": 0, + "content": "18. Kaul, P., Xie, W., Zisserman, A.: Label, Verify, Correct: A Simple Few-Shot Object Detection Method. In: IEEE Conference on Computer Vision and Pattern Recognition (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.262, + 0.785, + 0.33 + ], + "angle": 0, + "content": "19. Kaushal, V., Iyer, R., Doctor, K., Sahoo, A., Dubal, P., Kothawade, S., Mahadev, R., Dargan, K., Ramakrishnan, G.: Demystifying multi-faceted video summarization: Tradeoff between diversity, representation, coverage and importance. In: 2019 IEEE Winter Conference on Applications of Computer Vision (WACV). pp. 452-461 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.332, + 0.785, + 0.374 + ], + "angle": 0, + "content": "20. Khosla, P., Teterwak, P., Wang, C., Sarna, A., Tian, Y., Isola, P., Maschinot, A., Liu, C., Krishnan, D.: Supervised contrastive learning. In: Advances in Neural Information Processing Systems (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.375, + 0.785, + 0.43 + ], + "angle": 0, + "content": "21. Kothawade, S., Ghosh, S., Shekhar, S., Xiang, Y., Iyer, R.K.: Talisman: Targeted active learning for object detection with rare classes and slices using submodular mutual information. In: Computer Vision - ECCV 2022 - 17th European Conference (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.431, + 0.785, + 0.487 + ], + "angle": 0, + "content": "22. Kothawade, S., Kaushal, V., Ramakrishnan, G., Bilmes, J.A., Iyer, R.K.: PRISM: A rich class of parameterized submodular information measures for guided data subset selection. In: Thirty-Sixth AAAI Conference on Artificial Intelligence, AAAI. pp. 10238-10246 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.785, + 0.529 + ], + "angle": 0, + "content": "23. Li, B., Liu, C., Shi, M., Chen, X., Ji, X., Ye, Q.: Proposal distribution calibration for few-shot object detection. IEEE transactions on neural networks and learning systems (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.53, + 0.785, + 0.557 + ], + "angle": 0, + "content": "24. Li, B., Yang, B., Liu, C., Liu, F., Ji, R., Ye, Q.: Beyond Max-Margin: Class Margin Equilibrium For Few-Shot Object Detection. In: CVPR (June 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.559, + 0.785, + 0.6 + ], + "angle": 0, + "content": "25. Lin, H., Bilmes, J.: A class of submodular functions for document summarization. In: Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.601, + 0.785, + 0.642 + ], + "angle": 0, + "content": "26. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision - ECCV 2014. pp. 740-755. Springer International Publishing, Cham (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.643, + 0.785, + 0.698 + ], + "angle": 0, + "content": "27. Liu, F., Zhang, X., Peng, Z., Guo, Z., Wan, F., Ji, X., Ye, Q.: Integrally Migrating Pre-trained Transformer Encoder-decoders for Visual Object Detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6825-6834 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.7, + 0.785, + 0.755 + ], + "angle": 0, + "content": "28. Ma, J., Niu, Y., Xu, J., Huang, S., Han, G., Chang, S.F.: Digeo: Discriminative geometry-aware learning for generalized few-shot object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.757, + 0.785, + 0.799 + ], + "angle": 0, + "content": "29. Majee, A., Agrawal, K., Subramanian, A.: Few-Shot Learning For Road Object Detection. In: AAAI Workshop on Meta-Learning and MetaDL Challenge. vol. 140, pp. 115-126 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.841 + ], + "angle": 0, + "content": "30. Majee, A., Kothawade, S.N., Killamsetty, K., Iyer, R.K.: SCoRe: Submodular Combinatorial Representation Learning. In: Forty-first International Conference on Machine Learning (ICML) (2024)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.403, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "SMILE: Submodular Mutual Information Learner" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "31. Ren, S., He, K., Girshick, R.B., Sun, J.: Faster R-CNN: Towards Real-Time Object Detection With Region Proposal Networks. IEEE Trans. on Pattern Analysis and Machine Intelligence (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.787, + 0.217 + ], + "angle": 0, + "content": "32. Shangguan, Z., Rostami, M.: Identification of novel classes for improving few-shot object detection (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.218, + 0.786, + 0.244 + ], + "angle": 0, + "content": "33. Snell, J., Swersky, K., Zemel, R.: Prototypical Networks For Few-shot Learning. In: NeurIPS. pp. 4077-4087 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.245, + 0.787, + 0.272 + ], + "angle": 0, + "content": "34. Sun, B., Li, B., Cai, S., Yuan, Y., Zhang, C.: FSCE: Few-Shot Object Detection Via Contrastive Proposal Encoding. In: CVPR (June 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.273, + 0.787, + 0.3 + ], + "angle": 0, + "content": "35. Sung, F., Yang, Y., Zhang, L., Xiang, T., Torr, P.H., Hospedales, T.M.: Learning To Compare: Relation Network For Few-Shot Learning. In: CVPR (June 2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.301, + 0.787, + 0.327 + ], + "angle": 0, + "content": "36. Vinyals, O., Blundell, C., Lillicrap, T., kavukcuoglu, k., Wierstra, D.: Matching Networks For One Shot Learning. In: NeurIPS (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.328, + 0.787, + 0.355 + ], + "angle": 0, + "content": "37. Wang, X., Huang, T.E., Darrell, T., Gonzalez, J.E., Yu, F.: Frustratingly Simple Few-Shot Object Detection. In: ICML (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.356, + 0.787, + 0.383 + ], + "angle": 0, + "content": "38. Wang, Y.X., Ramanan, D., Hebert, M.: Meta-Learning To Detect Rare Objects. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.384, + 0.787, + 0.411 + ], + "angle": 0, + "content": "39. Wu, J., Liu, S., Huang, D., Wang, Y.: Multi-scale positive sample refinement for few-shot object detection. In: European Conference on Computer Vision (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.412, + 0.787, + 0.438 + ], + "angle": 0, + "content": "40. Xiao, Y., Marlet, R.: Few-Shot Object Detection And Viewpoint Estimation For Objects In The Wild. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.44, + 0.787, + 0.48 + ], + "angle": 0, + "content": "41. Yan, X., Chen, Z., Xu, A., Wang, X., Liang, X., Lin, L.: Meta R-CNN: Towards General Solver For Instance-Level Low-Shot Learning. In: CVPR. pp. 9577-9586 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.481, + 0.787, + 0.522 + ], + "angle": 0, + "content": "42. Zhang, G., Cui, K., Wu, R., Lu, S., Tian, Y.: PNPDet: Efficient Few-Shot Detection Without Forgetting Via Plug-And-Play Sub-Networks. In: WACV. pp. 3823-3832 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.523, + 0.787, + 0.55 + ], + "angle": 0, + "content": "43. Zhang, L., Zhou, S., Guan, J., Zhang, J.: Accurate Few-Shot Object Detection With Support-Query Mutual Guidance And Hybrid Loss. In: CVPR (June 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.551, + 0.787, + 0.591 + ], + "angle": 0, + "content": "44. Zhang, S., Luo, D., Wang, L., Koniusz, P.: Few-Shot Object Detection By Second-order Pooling. In: Proceedings of the Asian Conference on Computer Vision (ACCV) (November 2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.592, + 0.787, + 0.619 + ], + "angle": 0, + "content": "45. Zhu, C., Chen, F., Ahmed, U., Shen, Z., Savvides, M.: Semantic Relation Reasoning For Shot-Stable Few-Shot Object Detection. In: CVPR (June 2021)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.619 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/SMILe_ Leveraging Submodular Mutual Information For Robust Few-Shot Object Detection/57f9e916-d947-4fa8-bb11-b89fc66615b3_origin.pdf b/2024/SMILe_ Leveraging Submodular Mutual Information For Robust Few-Shot Object Detection/57f9e916-d947-4fa8-bb11-b89fc66615b3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2eec13c7fbd4db9384583f7a94efcfe9fc33a5d9 --- /dev/null +++ b/2024/SMILe_ Leveraging Submodular Mutual Information For Robust Few-Shot Object Detection/57f9e916-d947-4fa8-bb11-b89fc66615b3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0d277937cf5e92581d3b280e799a86b1c80c4672de23f5fe9df26573911053c2 +size 5558133 diff --git a/2024/SMILe_ Leveraging Submodular Mutual Information For Robust Few-Shot Object Detection/full.md b/2024/SMILe_ Leveraging Submodular Mutual Information For Robust Few-Shot Object Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..db7586ee68b1867c70d0a922b808948b31de8a35 --- /dev/null +++ b/2024/SMILe_ Leveraging Submodular Mutual Information For Robust Few-Shot Object Detection/full.md @@ -0,0 +1,252 @@ +# SMILE: Leveraging Submodular Mutual Information For Robust Few-Shot Object Detection + +Anay Majee1, Ryan Sharp2, and Rishabh Iyer1 + +1 The University of Texas at Dallas, TX, USA + +{anay.majee, rishabh.iley}@utdallas.edu + +2 IGS Energy, OH, USA + +rysharp@igsenergy.com + +Abstract. Confusion and forgetting of object classes have been challenges of prime interest in Few-Shot Object Detection (FSOD). To overcome these pitfalls in metric learning based FSOD techniques, we introduce a novel Submodular Mutual Information Learning (SMILE3) framework for loss functions which adopts combinatorial mutual information functions as learning objectives to enforce learning of well-separated feature clusters between the base and novel classes. Additionally, the joint objective in SMILE minimizes the total submodular information contained in a class leading to discriminative feature clusters. The combined effect of this joint objective demonstrates significant improvements in class confusion and forgetting in FSOD. Further we show that SMILE generalizes to several existing approaches in FSOD, improving their performance, agnostic of the backbone architecture. Experiments on popular FSOD benchmarks, PASCAL-VOC and MS-COCO show that our approach generalizes to State-of-the-Art (SoTA) approaches improving their novel class performance by up to $5.7\%$ (3.3 mAP points) and $5.4\%$ (2.6 mAP points) on the 10-shot setting of VOC (split 3) and 30-shot setting of COCO datasets respectively. Our experiments also demonstrate better retention of base class performance and up to $2\times$ faster convergence over existing approaches agnostic of the underlying architecture. + +# 1 Introduction + +Recent advances in Deep Neural networks (DNNs) have enabled models to learn discriminative feature representations from large-scale image benchmarks. Unfortunately, these architectures fail to adapt to few-shot settings tasked to recognize novel objects over existing ones with few examples, closely resembling human-like perception. Although recent research has shown significant promise in few-shot image recognition [10, 12, 33, 35, 36], Few-Shot Object Detection (FSOD) remains a challenge with recent works [27, 28, 37, 38] highlighting two + +![](images/1ea79b0189d765270c8b55465a89be4be6ce67ec6f448f069b3c494f4fecaab2.jpg) +Fig. 1: Functionality of components in $L_{comb}$ proposed in the SMILE (ours) framework, (a) $L_{comb}^{inter}$ promotes separation between $C_b$ and $C_n$ while (c) $L_{comb}^{intra}$ promotes intra-class compactness. + +major challenges - Class Confusion and Catastrophic Forgetting. Class confusion, as highlighted in [29] manifests itself through mis-prediction of instances belonging to a newly learnt (novel) class, as one or more instances of the already learnt (base) classes. Authors in [1, 38] attribute this to the sharing of visual information between classes resulting in increased inter-class bias due to overlapping feature clusters as shown in Fig. 1(a). Catastrophic forgetting refers to the gradual degradation in the performance of already learnt classes in the quest to learn the novel ones, as shown in Fig. 2(a), seldom overfitting to rare classes [29, 38]. Further, large feature diversity (intra-class variance) among base classes lead to formation of non-discriminative feature clusters as shown in Fig. 1(b), aggravating the existing inter-class bias in the feature space. Unlike existing approaches (refer Sec. 2) which target either confusion or forgetting, our paper presents a unified approach to tackle both these challenges in FSOD. Although, recent approaches [1, 28, 34] attempt to tackle these challenges through contrastive learning strategies, such approaches have been limited by their capability to overcome either inter-class bias or intra-class variance [30, 38] and poor generalization to longtail settings [30] (FSOD being a extreme case). + +In this paper, we introduce a combinatorial viewpoint in FSOD considering each object class $i \in [1, C]$ in the dataset $\mathcal{T}$ as a set $A_i$ of samples, where $\mathcal{T} = \{A_1, \dots, A_C\}$ , facilitating the application of combinatorial functions as learning objectives. We aim to overcome the aforementioned challenges through representation learning in the low-data regime by adopting this formulation through the SMILE: Submodular Mutual Information Learning framework, wherein we introduce novel, set-based combinatorial objective functions for FSOD as shown in Fig. 3. SMILE introduces a joint objective formulation $L_{comb}$ (Eq. (3)) based on two popular flavors of submodular information functions - Submodular Mutual Information [22] (SMI) and Total Submodular Information [11] targeting the root causes of confusion and forgetting in FSOD. At first, SMILE is the first to introduce pairwise SMI functions $I_f$ in representation learning which model the common (overlapping) information between two + +![](images/de43313bc2880b20646758366c8511fda729561914c5d681b588d8c3844dd8d4.jpg) +Fig. 2: Resilience to Catastrophic forgetting and faster convergence in SMILE over SoTA approaches. (a) shows that combinatorial losses in SMILE are robust to catastrophic forgetting, while (b) shows that objectives in SMILE results in faster convergence over SoTA FSOD methods (AGCM and DiGeo). + +![](images/a9150b517d2da2d94c869bc75c30efe377be18107d2e72d210e55782362a4c13.jpg) + +sets. Minimizing $I_{f}$ through the joint objective $L_{comb}$ reduces feature overlap between base and novel classes alleviating inter-class bias in the model towards abundantly sampled classes as shown in Fig. 1(b). We extend this property of SMI functions to the novel classes minimizing the inter-cluster overlap between few-shot classes, promoting learning of discriminative features from just few samples. Secondly, SMILE preserves the diversity within each class by minimizing the total submodular information contained within each set as shown in Fig. 1(c), minimizing the impact of forgetting. This formulation closely follows the observation in [30] which models cooperation [15] between instances in a set by minimizing a submodular function over a set, to preserve representative features. The unified objective $L_{comb}$ introduced in SMILE models both these necessary properties through a weighted sum of two distinct objectives $L_{comb}^{inter}$ and $L_{comb}^{intra}$ as shown in Fig. 3 balancing the tradeoff between inter-cluster separation and intra-cluster compactness respectively. This allows us to introduce a family of loss functions which inherently eliminates confusion and forgetting as shown in Tab. 5. We conduct our experiments on two popular FSOD benchmarks, PASCAL-VOC [6] and MS-COCO [26] for several few-shot settings and demonstrate the following contributions of SMILE: + +- SMILE introduces a novel set-based combinatorial viewpoint in FSOD by applying combinatorial Mutual Information based objective to discriminate between base and novel classes, in conjunction with submodular total information to minimize intra-class variance as the objective function. +- SMILE generalizes to existing approaches in FSOD, irrespective of the underlying architecture demonstrating up to $5.7\%$ improvement in novel class performance (Tab. 2) over the baseline FSOD approach. +SMILE demonstrates up to $2 \times$ faster convergence (Fig. 2(b)) over existing SoTA approaches resulting in faster generalization to unknown object classes. +- Finally, SMILE demonstrates up to $11\%$ and $3.5\%$ reduction in class confusion and catastrophic forgetting while achieving SoTA performance on popular FSOD benchmarks like PASCAL-VOC (by $5.7\%$ on split 2, 10-shot setting) and MS-COCO ( $5.4\%$ on 30-shot setting). + +# 2 Related Work + +Few-Shot Object Detection (FSOD): Classical FSOD approaches utilize finetuning [3] or distance metric learning [17] to adapt features to novel classes. Recent methods employ meta-learning techniques [16,40,41] with episodic training to learn class-specific features. Meta-Reweight [16] and Meta-RCNN [41] use additional feature extractors, while Add-Info [40] leverages feature differences between support and query images. Techniques like [43] enhance class-specific features through information sharing, and CME [24] aims to reduce class confusion. Attention mechanisms [7,44] are used to identify discriminative features. However, meta-learning approaches are resource-intensive and may fail to generalize to significantly different novel classes. Metric learning strategies like FsDet [37], FSCE [34], and SRR-FSD [45] offer better generalization without additional overheads. PNPDet [42] partially addresses catastrophic forgetting and class confusion. GFSD [8] proposes a Bias-Balanced RPN to mitigate overfitting in metric learners. + +Recent approaches like [18,32] adopt weak supervision from unlabelled data or low confidence predictions in RoI pooling layers to generalize to novel classes. These methods often use abundant samples from base classes [28] to prevent catastrophic forgetting, adding computational overhead in low-shot settings. Vision transformers [4] have been adopted in FSOD through methods like imTED [27] and PDC [23], with reduced computational overhead by using pre-trained attention heads. Alternatively, DiGeo [28] and PDC [23] learn the geometry or difference in distributions of RoI proposals [13] between object classes to overcome forgetting and confusion. However, these approaches rely on contrastive learning objectives [20] that struggle to learn discriminative feature embeddings due to adoption of pairwise similarity metrics. Our work, SMILE, aims to improve the feature learning capacity of existing SoTA approaches, irrespective of their underlying architectures. + +Submodular Functions and Combinatorial Objectives: Submodular functions are recognized as set functions with an inherent diminishing returns characteristic. Defined as a set function $f: 2^{\mathcal{V}} \to \mathbb{R}$ operating on a ground-set $\mathcal{V}$ , a function is termed submodular if it adheres to the condition $f(X) + f(Y) \geq f(X \cup Y) + f(X \cap Y), \forall X, Y \subseteq \mathcal{V}$ [11]. These functions have garnered considerable attention in research, particularly in fields like data subset selection [22], active learning [21], and video summarization [19,22] through their ability in modeling concepts such as diversity, relevance, set-cover and representation. A subclass of submodular functions, namely Submodular Mutual Information (SMI) functions introduced in [22] model the similarity and diversity between pairs of object classes establishing itself as a powerful tool to model inter-class bias. Recently, Majee et al. [30] introduces these set-based combinatorial functions as objectives in representation learning and demonstrates their capability in overcoming inter-class bias (by minimizing the similarity between nonidentical object classes) and intra-class variance (maximizing the similarity between instances of the same object class). However, these functions are yet to be studied in the context of few-shot learning. We introduce novel instances of SMI based objectives in + +![](images/41d46461f410f09a8a7a20db24290b5e43a772732f76334dd723fb4d8a6a0cd8.jpg) +Fig. 3: Overview of our SMILE framework highlighting the application of Mutual Information function based objectives in SMILE for the fine-tuning stage of Few-Shot Object Detection. + +SMILEs to minimize inter-class bias between base and novel classes. To the best of our knowledge we are the first to introduce novel SMI based combinatorial objectives in conjunction with total information based combinatorial functions through SMILEs in a quest to minimize confusion and forgetting in few-shot object detection. + +# 3 Method + +# 3.1 Problem Definition : Few-Shot Object Detection + +We define a few-shot learner $h(x, \theta)$ as shown in Fig. 3 that receives input data $x$ from base classes $C_b \in [1, |C_b|]$ and novel classes $C_n \in [1, |C_n|]$ such that $C = \{C_b \cup C_n\}$ and $\{C_b \cap C_n\} = \emptyset$ . Here, $\theta$ denotes the learnable parameters. The training data can be divided into two distinct parts, base $D_{base}$ and novel $D_{novel}$ such that, $\mathcal{T} = \{D_{base} \cup D_{novel}\}$ and $\{D_{base} \cap D_{novel}\} = \emptyset$ . SMILe introduces a paradigm shift in FSOD by imbibing a combinatorial viewpoint, where the base dataset, $D_{base} = [A_1^b, A_2^b, \dots, A_{|C_b|}^b]$ , containing abundant training examples from $C_b$ base classes and the novel dataset, $D_{novel} = [A_1^n, A_2^n, \dots, A_{|C_n|}^n]$ containing only $K$ -shot ( $|A_i^n| = K$ for $i \in [1, C_n]$ ) training examples from $C_n$ novel classes. The objective of the few-shot learner $h(x, \theta)$ is to learn discriminative representation from classes in $D_{novel}$ without degradation in performance on classes in $D_{base}$ . Following FSCE [34] we adopt a two-stage training strategy. In the base training stage we train $h(x, \theta)$ on abundant samples in $D_{base}$ , allowing the model to generalize on the domain of $D_{base}$ . The few-shot adaptation stage adapts $h(x, \theta)$ to previously unseen $K$ -shot data by fine-tuning on data samples from $D_{base} \cup D_{novel}$ where $|A_k| = K$ for $k \in \{C_b \cup C_n\}$ . The goal of SMILe is to overcome class confusion and forgetting in FSOD resulting from elevated inter-class bias and intra-class variance as observed in [1, 29, 38]. The final model $h(x, \theta)$ obtained after two training stages is evaluated on $D_{test}$ containing unseen data samples from both $C_b \cup C_n$ . + +# 3.2 The SMILE Framework + +Adopting a combinatorial viewpoint as disclosed earlier allows us to employ submodular combinatorial functions as learning objectives to tackle confusion and forgetting in FSOD. As discussed in Sec. 2, minimizing a Submodular functions naturally models cooperation [15] while maximizing it models diversity [25] due to their inherent diminishing marginal returns property. SMILE adopts the aforementioned properties of submodular functions to define a novel family of combinatorial objective (loss) functions $L_{comb}(\theta)$ which enforces orthogonality in the feature space when applied on Region-of-Interest (RoI) features in FSOD models. The loss function $L_{comb}(\theta)$ can be decomposed into two major components - $L_{comb}^{inter}$ minimizes inter-class bias between base and novel classes and $L_{comb}^{intra}$ maximizes intra-class compactness within abundant classes. + +For $L_{comb}^{inter}$ , SMILE explores a sub-category of combinatorial functions, namely Submodular Mutual Information (SMI) which can be defined as $I_{f}(A_{i}, A_{j}) = f(A_{i}) + f(A_{j}) - f(A_{i} \cup A_{j})$ [11, 22], and models the common information between two sets $A_{i}$ and $A_{j}$ , $\forall i, j \in \mathcal{T}$ . Results in [11, 22] portray $I_{f}(A_{i}, A_{j}; \theta)$ as a measure of the degree of similarity between object classes $A_{i}$ and $A_{j}$ . Adopting this definition of SMI, $L_{comb}^{inter}$ minimizes the SMI between the base $C_{b}$ and the novel $C_{n}$ classes, ensuring sufficient inter-cluster separation (by minimizing inter-class bias) as shown in Eq. (1). $L_{comb}^{inter}$ further minimizes the mutual information between classes in $C_{n}$ , minimizing inter-cluster overlaps between the novel classes. This is visually depicted in Fig. 1(b) and has been shown to be effective in mitigating class confusion in FSOD through our experiments in Sec. 4. + +$$ +L _ {\text {c o m b}} ^ {\text {i n t e r}} (\theta) = \sum_ {\substack {b \in C _ {b} \\ n \in C _ {n}}} I _ {f} (A _ {b}, A _ {n}; \theta) + \sum_ {\substack {i, j \in C _ {n} \\ i \neq j}} I _ {f} (A _ {i}, A _ {j}; \theta) = \sum_ {\substack {i \in (C _ {b} \cup C _ {n}) \\ j \in C _ {n}: i \neq j}} I _ {f} (A _ {i}, A _ {j}; \theta) \tag{1} +$$ + +In addition to confusion which stems from inter-class bias, SMILEs aims at mitigating catastrophic forgetting [29] in FSOD which has been attributed to large intra-class variance among abundant object classes in [1,38]. In coherence to the combinatorial formulation in SMILEs we achieve this through $L_{comb}^{intra}$ which minimizes the Total Submodular Information, defined as $S_f(A_1, \dots, A_{|C|}) = \sum_{k=1}^{|C|} f(A_k; \theta)$ , over sets $A_k \in \mathcal{T}$ , given a submodular function $f(A_k; \theta)$ . As discussed earlier, minimizing the submodular information models cooperation which asserts that minimizing $L_{comb}^{inter}$ promotes learning of discriminative feature clusters, penalizing abundant classes to have large feature variance in the embedding space as shown in Fig. 1(c). Although submodular functions have been studied in the field of representation learning to minimize intra-class variance in [30], but primarily differs from SMILEs in modeling a longtail recognition task by minimizing the total submodular correlation, which models gain in information when new features are added to a set. The formulation of $L_{comb}^{intra}$ has been shown in Eq. (2) where we minimize the total submodular information within samples in each class in $C_b \cup C_n$ and our experiments in Tab. 5 + +show the effectiveness of $L_{comb}^{intra}$ in boosting base class performance asserting the mitigation of catastrophic forgetting. + +$$ +L _ {c o m b} ^ {i n t r a} (\theta) = \sum_ {b \in C _ {b}} f (A _ {b}, \theta) + \sum_ {n \in C _ {n}} f (A _ {n}, \theta) = \sum_ {k \in \left(C _ {b} \cup C _ {n}\right)} f (A _ {k}, \theta) \tag {2} +$$ + +Ablating on the choice of the submodular function $f$ and SMI functions $I_{f}$ we introduce several instances of SMILE objectives as discussed in Tab. 1. + +Encapsulating the aforementioned formulations of $L_{comb}^{inter}$ and $L_{comb}^{intra}$ in SMILE we define a joint objective $L_{comb}(\theta)$ which tackles both the challenges of confusion and forgetting. We thus define $L_{comb}(\theta)$ in Eq. (3) which is the weighted algebraic sum of $L_{comb}^{inter}$ and $L_{comb}^{intra}$ with the weighting factor $\eta$ . + +$$ +\begin{array}{l} L _ {c o m b} (\theta) = (1 - \eta) L _ {c o m b} ^ {i n t r a} (\theta) + \eta L _ {c o m b} ^ {i n t e r} (\theta) \\ = \sum_ {i \in C _ {b} \cup C _ {n}} \left[ (1 - \eta) f \left(A _ {i}, \theta\right) + \eta \sum_ {\substack {j \in C _ {n} \\ i \neq j}} I _ {f} \left(A _ {i}, A _ {j}; \theta\right) \right] \tag{3} \\ \end{array} +$$ + +Note, that the combinatorial objective $L_{comb}(\theta)$ is applied on output features from the RoI Pooling layers in proposal-based [31,38] architectures. To promote adoption of SMILE agnostic of the backbone architecture we introduce a combinatorial head $Z_{comb} = Comb(h,\theta)$ which projects the RoI features to 128-dimensional feature vectors [20], $Z_{comb}$ on which $L_{comb}(\theta)$ is applied during the few-shot adaptation stage. + +Finally, we summarize the total classification loss in SMILEs as depicted in Eq. (4) as the sum over all three objectives: the classification head $L_{Clf}$ , the box regression head $L_{bbox}$ and the combinatorial head $L_{comb}(\theta)$ . Note that the objectives proposed in SMILEs apply only to $Comb(h,\theta)$ while the RoI classification and regression heads are unchanged. This follows the observations in [28,38] which warrants the boost in performance originating from learning robust feature representations for each RoI predicted by the model. + +$$ +L _ {c l s} (\theta) = L _ {C l f} (\theta) + L _ {b b o x} (\theta) + L _ {c o m b} (\theta) \tag {4} +$$ + +# 3.3 Instantiations of $L_{comb}^{inter}$ and $L_{comb}^{intra}$ in the SMILE Framework + +Given a submodular function $f(A)$ and a Submodular Mutual Information (SMI) function $I_{f}(A,Q)$ over sets $A$ and $Q$ , we derive two instances $L_{comb}^{inter}$ and $L_{comb}^{intra}$ objectives in SMILEs. Depending on the choice of $f(A)$ we define two instances: Facility-Location Mutual Information (SMILE-FLMI) and Graph-Cut Mutual Information (SMILE-GCMI). Inherently, both objectives adopt the cosine similarity metric $S_{ij}(\theta)$ as used in SupCon [20] which can be defined as $S_{ij}(\theta) = \frac{Z_{comb_i}^{\mathrm{T}} \cdot Z_{comb_j}}{||Z_{comb_i}|| \cdot ||Z_{comb_j}||}$ to compute similarity between sets in the learning objective. Although the similarity kernel used in SMILEs is computed in a pairwise fashion, objectives defined under $L_{comb}$ use it to only compute feature interactions + +Table 1: Summary of various instantiations of SMILE highlighting the components of the combinatorial objective, $L_{comb}^{inter}$ and $L_{comb}^{intra}$ . + +
ObjectiveInstances of Lintercomb(θ)Instances of Intracomb(θ)
SMILE-GCMI (ours)∑k∈(Ck∪Cn)2λ∑iAksjA1Sij(θ)∑k∈Ck∪Cn∑iAksjTsetAksij(θ) - λ∑i,jAksij(θ)
∑k∈(Ck∪Cn)2λ∑iAksjA1maxSij(θ) + λ∑iAksjA1maxSij(θ)∑k∈Ck∪Cn∑iTsetAksij(θ)
SMILE-FLMI (ours)
+ +between samples, differing from existing approaches in aggregation of pairwise similarities to compute total information and mutual information over classes in $\mathcal{T}$ . + +SMILE-FLMI based objective is derived from the Facility-Location Mutual Information (FLMI) [22] function, expressed as $I_{f}(Q,A) = \sum_{i\in Q}\max_{j\in A}S_{ij}(\theta) + \lambda \sum_{i\in A}\max_{j\in Q}S_{ij}(\theta)$ and minimizes the maximum similarity (most similar) between sets $Q$ and $A$ . Given the facility-location (FL) submodular function $f(A,\theta) = \sum_{i\in \mathcal{T}}\max_{j\in A}S_{ij}(\theta)$ over the set $A$ , we can derive $L_{comb}^{inter}(\theta)$ and $L_{comb}^{intra}(\theta)$ shown in Eq. (5) as the SMILE-FLMI objective. Note that $L_{comb}^{inter}(\theta)$ is applied between object classes in $C_b\cup C_n$ and $C_n$ while $L_{comb}^{intra}(\theta)$ is applied over all classes in $C_b\cup C_n$ . + +$$ +\begin{array}{l} L _ {c o m b} ^ {i n t e r} (\theta) = \sum_ {k \in \left(C _ {b} \cup C _ {n}\right)} \sum_ {i \in A _ {k}} \max _ {j \in A _ {l}} S _ {i j} (\theta) + \lambda \sum_ {i \in A _ {l}} \max _ {j \in A _ {k}} S _ {i j} (\theta), \\ l \in C _ {n}: k \neq l \tag {5} \\ \end{array} +$$ + +$$ +L _ {c o m b} ^ {i n t r a} (\theta) = \sum_ {k \in C _ {b} \cup C _ {n}} \sum_ {i \in \mathcal {T} \backslash A _ {k}} \max _ {j \in A _ {k}} S _ {i j} (\theta) +$$ + +Minimizing the $L_{comb}^{inter}$ objective function ensures that the sets $A_{l} \in C_{n}$ and $A_{k} \in C_{b} \cup C_{n}$ are disjoint by minimizing the similarity between features in $A_{k}$ and the hardest negative $(\sum_{i \in A_{k}} \max_{j \in A_{l}} S_{ij}(\theta)$ for $k \in C_{b} \cup C_{n}$ and $l \in C_{n})$ feature vectors in $A_{l}$ . Further, $L_{comb}^{inter}$ enforces sufficient separation between the novel classes themselves to promote learning of disjoint feature clusters even with few-shot data overcoming confusion. Additionally, $L_{comb}^{intra}$ minimizes the total information contained in each set $A_{k} \in (C_{b} \cup C_{n})$ . This objective retains discriminative feature information from each class in $\mathcal{T}$ reducing the impact of forgetting. + +SMILE-GCMI based objective described in Eq. (6) minimizes the pairwise similarity of feature vectors between a positive set $A_{k} \in C_{b} \cup C_{n}$ and the sets in $A_{l} \in C_{n}$ while maximizing the similarity between features in each set $A_{k} \in C_{b} \cup C_{n}$ . Given two sets $Q$ and $A$ , [22] defines the Graph-Cut SMI to be $I_{f}(Q, A) = 2\lambda \sum_{i \in Q} \sum_{j \in A} S_{ij}(\theta)$ , where the Graph-Cut function over a set $A$ is given by $f(A, \theta) = \sum_{i \in A} \sum_{j \in T \setminus A_{k}} S_{ij}(\theta) - \lambda \sum_{i,j \in A} S_{ij}(\theta)$ . Given the Graph-Cut and the Graph-Cut SMI functions, we derive $L_{comb}^{inter}(\theta)$ and $L_{comb}^{intra}(\theta)$ + +shown in Eq. (6) as the SMILE-GCMI objective. Similar to SMILE-FLMI, the $L_{comb}^{inter}(\theta)$ is applied between object classes in $C_b \cup C_n$ and $C_n$ while $L_{comb}^{intra}(\theta)$ is applied over all classes in $C_b \cup C_n$ . + +$$ +L _ {c o m b} ^ {i n t e r} (\theta) = \sum_ {\substack {k \in (C _ {b} \cup C _ {n}) \\ l \in C _ {n}: k \neq l}} 2 \lambda \sum_ {i \in A _ {k}} \sum_ {j \in A _ {l}} S _ {i j} (\theta), \tag{6} +$$ + +$$ +L _ {c o m b} ^ {i n t r a} (\theta) = \sum_ {k \in C _ {b} \cup C _ {n}} \sum_ {i \in A _ {k}} \sum_ {j \in \mathcal {T} \backslash A _ {k}} S _ {i j} (\theta) - \lambda \sum_ {i, j \in A _ {k}} S _ {i j} (\theta) +$$ + +Although objectives in SMILE-FLMI and SMILE-GCMI are tasked with similar functions, the $L_{comb}^{inter}$ in SMILE-GCMI minimizes the pairwise similarity between sets in $C_b \cup C_n$ and $C_n$ rather than the most similar set in SMILE-FLMI. Further, the $L_{comb}^{intra}$ in SMILE-GCMI scales linearly with size of $A_k$ as described in [30]. This does not allow the model to substantially improve performance on learning discriminative feature representations for classes in both $C_b$ and $C_n$ as the $|A_k| = K$ (number of shots) thus failing to outperform the model trained using SMILE-FLMI. + +The detailed derivations of the aforementioned instances are included in the Supplementary material. Our experiments in Sec. 4.4 elucidates the fact that SMILE-FLMI is a better choice to overcome forgetting and confusion in FSOD. + +# 4 Experiments + +We evaluate models in SMILE by adopting standard evaluation criterion in FSOD [16, 37] and report the Mean Average Precision (mAP) at $50\%$ Intersection Over Union (IoU) for all our experiments. + +# 4.1 Experimental Setup + +Datasets We evaluate our proposed SMILE approach on two few-shot object detection datasets - and PASCAL-VOC [5] and MS-COCO [26] datasets. + +PASCAL-VOC [5] dataset consists of 20 classes, out of which 15 are considered as base and 5 as novel classes. The novel classes are chosen at random giving rise to three data splits namely, split-1 (bird, bus, cow, motorbike, sofa), split-2 (aeroplane, bottle, cow, horse, sofa) and split-3 (boat, cat, motorbike, sheep, sofa). Following previous works [16], we use the combined VOC $07 + 12$ datasets for training and evaluate our models on the complete validation set of VOC 2007 for 1, 5, and 10 shot settings. + +MS-COCO [26] dataset consists of 80 classes, out of which 60 are considered as base and 20 as novel classes. Following existing approaches in FSOD [41] we randomly select 5k samples from $(D_{base} \cup D_{novel})$ to use as the validation set while the remaining samples are used to generate random 10 and 30-shot splits for training of the MS-COCO 2014 dataset. The key difference between VOC and COCO is the large intra-class variance and class-imbalance in COCO. + +Table 2: Quantitative analysis on PASCAL-VOC dataset: Few-shot object detection performance ( $mAP_{novel}$ ) on novel class splits of PASCAL-VOC dataset. We tabulate results for $K = 1, 5, 10$ shots from various SoTA techniques in FSOD. * indicates that the results are averaged over 10 random seeds. † indicates a meta-learning strategy (N-way, K-shot training). + +
MethodLearner TypeBackboneSplit 1Split 2Split 3
K=151015101510
†Meta-RCNN [41]MetaFRCN-R10119.945.751.510.434.845.414.341.248.1
†Meta-Reweight [16]MetaYOLO V214.833.947.215.730.140.521.342.845.9
†MetaDet [38]MetaFRCN-R10118.936.849.621.831.743.020.643.944.1
†Add-Info [40]MetaFRCN-R10124.249.157.421.637.045.721.243.849.6
†CME [24]MetaYOLO V217.844.847.512.733.740.015.744.948.8
PNPDet [42]MetricDLA-3418.2-41.016.6-36.418.9-36.2
FsDet w/ FC [37]MetricFRCN-R10136.855.757.018.235.539.027.748.750.2
FsDet w/ cos [37]MetricFRCN-R10139.855.756.023.535.139.130.849.549.8
Retentive-RCNN [9]MetricFRCN-R10140.153.756.121.737.040.330.249.750.1
FSCE [34]MetricFRCN-R10141.057.457.827.344.449.840.153.257.7
\( FSCE + SMILE \) (ours)Comb.FRCN-R10141.257.961.129.244.650.541.355.659.0
AGCM [1]MetricFRCN-R10140.358.559.927.549.350.642.154.258.2
\( AGCM + SMILE \) (ours)Comb.FRCN-R10140.959.762.031.949.552.342.656.461.4
DiGeo [28]MetricFRCN-R10136.054.160.920.742.847.127.547.352.9
DiGeo + SMILE(ours)Comb.FRCN-R10136.156.662.326.544.147.333.151.956.4
imTED [27]MetricViT-B31.971.977.022.752.257.712.669.672.8
imTED + PDC [23]MetricViT-B36.673.177.115.551.856.018.967.972.8
PDC + SMILE (ours)Comb.ViT-B36.675.277.927.152.758.315.170.074.7
+ +Implementation Details The SMILE framework adopts a architecture agnostic approach and adopt several backbones including Faster-RCNN [31] and ViT [27]. For VOC, the input batch size to the network is set to 16 and 2 in the base training and few-shot adaptation stages for Faster-RCNN and ViT based approaches. The input resolution is set to $764 \times 1333$ pixels for data splits in COCO, while it is set to $800 \times 600$ pixels for PASCAL-VOC. The hyper-parameters used in the formulation of SMILE, namely $\eta$ and similarity kernel $S$ , are chosen through ablation experiments described in Sec. 4.4. Results from existing methods are a reproduction of the algorithm from publicly available codebases. All our experiments are performed on 4 NVIDIA GTX 1080 Ti GPUs with additional details in the supplementary material and code released at https://github.com/amajee11us/SMILE-FSOD.git. + +# 4.2 Results on Few-Shot PASCAL VOC Dataset + +Table 2 records the results obtained from our SMILE framework on novel splits of the PASCAL-VOC dataset and contrasts it against SoTA FSOD techniques. We adopt four SoTA approaches FSCE [34], AGCM [1], DiGeo [28] and imTED [27], covering several backbone architectures Faster-RCNN + FPN (FSCE, AGCM and DiGeo) alongside ViT (imTED, PDC) and introducing SMILE ( $M + SMILE$ ) approach into existing architectures $M$ . For Faster-RCNN based architectures (FSCE) we show a maximum of $5.7\%$ (3.3 map points) improvement while for FPN based architectures (AGCM and DiGeo) we show a $3.5\%$ (2.1 map points + +for AGCM+SMILE) improvement. It is interesting to note that unlike FSCE and AGCM, DiGeo uses abundant samples from $C_b$ alongside few-shot samples in $C_n$ (with upsampling) during finetuning introducing a large inter-class bias. SMILE outperforms DiGeo by up to 2.3 $mAP$ points (split 2, 10-shot) showing the resilience of SMILE towards imbalance, thus overcoming confusion in FSOD. Additionally, for recently introduced transformer based architectures (imTED + PDC [23]) SMILE outperforms the existing SoTA with a maximum improvement of 4.9 $mAP$ points (split 2, 5-shot) thus establishing SMILE as the SoTA on few-shot splits of VOC. Note, that the choice of objective functions $L_{comb}^{inter}$ and $L_{comb}^{intra}$ for this experiment has been determined to be SMILE+FLMI through an ablation on the different instances in SMILEs as described in Sec. 4.4. Finally, Fig. 2(b) shows that introduction of SMILE framework to existing SoTA approaches leads to rapid convergence on the novel classes up to 2x over existing SoTA. This is significant for mission critical tasks like autonomous driving where the model is required to rapidly learn novel objects to reduce turn-around time. + +# 4.3 Results on Few-Shot MS-COCO Dataset + +Similar to the results in PASCAL VOC we demonstrate the results of our SMILE framework on MS-COCO dataset. In contrast to VOC, COCO presents an extremely imbalanced setting with a long-tail distribution within $D_{base}$ itself making it really hard for FSOD approaches to achieve SoTA through primitive objective functions. Following the ablation experiments in Sec. 4.4 we adopt the SMILE+FLMI objective (best performing) to conduct the experiments on 20 few-shot classes of MS-COCO dataset. We show that SMILE generalizes existing SoTA approach (imTED + PDC) for COCO dataset by $5.4\%$ (2.6 mAP points, 30-shot setting). This further establishes the generalizability of our approach over varying data distributions (VOC and COCO) while achieving SoTA in FSOD tasks. + +# 4.4 Ablation Study + +We conduct ablation experiments on the 10-shot split of VOC (split 1) with hyper-parameters $\eta = 0.5$ , cosine similarity metric and $\lambda = 1.0$ . Ablations for hyper-parameters are detailed in the supplementary material. + +Components of SMILE Instantiations in SMILEs consists of two main components - $L_{comb}^{inter}$ and $L_{comb}^{intra}$ . We consider three baselines FSCE, AGCM and DiGeo which follow the Faster-RCNN/FPN backbone for this experiment. First, we introduce $L_{comb}^{intra}$ by adopting the FL based objective as determined through ablation experiments below. This objective models intra-class variance and ensures reduction in intra-class variance characterized by boost in base class performance. Secondly, following the SMILE-FLMI formulation in Eq. (5) we introduce $L_{comb}^{inter}$ during the few-shot adaptation stage. Applying this objective minimizes + +Table 3: Performance of SMILE on MS COCO dataset : Our SMILE objectives demonstrate better generalizability while outperforming SoTA FSOD approaches on novel class performance $mAP_{50}$ (novel). + +
Method\( {mAP} \)\( {mAP}_{50} \)\( {mAP}_{75} \)\( {mAP} \)\( {mAP}_{50} \)\( {mAP}_{75} \)
10-shot30-shot
Meta-Reweight [16]5.612.34.69.119.07.6
Meta-RCNN [41]8.719.16.612.425.310.8
TFA w/cos [37]10.0-9.313.7-13.4
Add-Info [40]12.527.39.814.730.612.2
MPSR [39]9.817.99.714.125.414.2
FSCE [34]11.9-10.516.4-16.2
FADI [2]12.222.711.916.129.115.8
CME [24]15.124.616.416.2--
FCT [14]17.130.217.021.435.522.1
imTED-B [27]22.536.623.730.247.432.5
imTED-B+PDC [23]23.438.124.530.847.333.5
PDC + SMILE (ours)25.840.126.131.049.933.6
+ +Table 4: Ablation on various components of the proposed SMILE approach. + +
MethodBaselinef(Ai)(Lintra)If(Ai,Aj)(Lintercomb)mAPbasemAPnovel
FsDet w/ cos---23.639.8
FSCE✓86.157.8
✓✓89.660.1
✓✓88.361.0
✓✓✓89.861.1
AGCM✓87.658.0
✓✓88.661.3
✓✓88.961.8
✓✓✓89.361.8
DiGeo✓90.560.9
✓✓92.361.7
✓✓91.462.0
✓✓✓92.662.3
+ +the inter-class bias between $C_b \cup C_n$ and $C_n$ , thus improving novel class performance significantly. Nevertheless, we see a slight drop in base class performance due to forgetting prevalent in FSOD tasks. Finally, we combine both instantiations in SMILE into one single objective as in Eq. (3) with $\eta = 0.5$ and show that SMILE improves both base and novel class performance emerging as the best choice for FSOD. To demonstrate generalization, we perform this experiment on several SoTA approaches as baseline and show that the results discussed in aforementioned section hold. We summarize all the results in Tab. 4. + +Choice of Combinatorial functions in $L_{comb}$ SMILE introduces several instances of $L_{comb}^{intra}$ and $L_{comb}^{inter}$ . To clearly understand the contributions of each of these instances we conduct experiments tabulated in Tab. 5 and determine the + +Table 5: Ablation on the choice of Submodular Information function $I_{f}$ and Submodular Information function $f$ for $L_{comb}$ in SMILE. + +
Modelf(A,θ)If(Ai,Aj)mAPbasemAPnovel
DiGeo [28]--87.957.4
GCGCMI92.660.9
FLFLMI93.162.3
FSCE [34]--86.157.8
GCGCMI87.460.3
FLFLMI89.861.1
+ +best performing formulation which generalizes to existing FSOD architectures. Unlike other ablation experiments in Sec. 4.4, we conduct our experiments on DiGeo which introduces an extremely imbalanced scenario by using abundant samples in $D_{base}$ . We conclude that SMILE-FLMI which considers the Facility-Location based objective as $L_{comb}^{intra}$ and Facility-Location Mutual Information based objective as $L_{comb}^{inter}$ as the best performing instantiation. This result follows the formulation in Eq. (5) where FL naturally models intra-class compactness in class-imbalanced settings [30] while FLMI penalizes the classes in $C_b$ to learn overlapping feature representations with the classes in $C_n$ . We use SMILE-FLMI for all benchmark experiments in Sec. 4.2 and Sec. 4.3. + +Robustness to Catastrophic Forgetting One of the most significant challenges in FSOD is the elimination of catastrophic forgetting which manifests as the degradation in the performance of classes in $C_b$ while learning classes in $C_n$ . This primarily occurs due to the lack of discriminative feature representations from instances in $D_{base}$ during the few-shot adaptation (stage 2) stage. We plot the change in base class performance as the training progresses in existing SoTA methods AGCM and DiGeo against number of training iterations in Fig. 2(a). At first, we contrast the change in base class performance $mAP_{base}$ between AGCM and AGCM+SMILE and observe that AGCM overfits on the few-shot samples in $D_{base}$ reducing the performance on $C_{base}$ as the training progresses. AGCM + SMILE on the other hand better retains the performance on base classes with $\sim 3.5\%$ better retention in base class performance. Interestingly, DiGeo is able to retain most of the base class performance with a very small degradation over the roofline (a model trained with only the base classes until convergence). Our DiGeo+SMILE approach outperforms Digo by demonstrating base class performance even higher than the roofline establishing the supremacy of $L_{comb}$ in overcoming inter-class bias and intra-class variance resulting in robustness against catastrophic forgetting. + +Overcoming Class Confusion Figure 4 highlights the supremacy of the proposed SMILE framework in mitigating class confusion through confusion matrix plots. We compare the confusion between classes in $C_b \cup C_n$ of SoTA approaches AGCM and DiGeo before and after introduction of combinatorial objectives in SMILE. Although both approaches use K-shot examples for classes in $C_n$ , Di + +![](images/21f3a143182809eff3cca7577c7548522161ae5987ee05923df949cf9e9624f3.jpg) +Fig. 4: Ablation on Overcoming Class Confusion in SMILE. (a,b) SMILE demonstrates $11\%$ lower confusion over AGCM and (c,d) $4\%$ lower confusion over DiGeo. Only significant numbers are highlighted. Best viewed in $200\%$ zoom. + +![](images/e66a23625cc084bf94517d28a182a2cf8e48f318423e47579ad4849ec8db6bab.jpg) + +![](images/c3c2890c81276f9bb2dd8825bc0aa0e27b5a08a24c0d73ae3e10c69de2be3dcf.jpg) + +![](images/b5352161d55bd070afbe86f70f9bf6e94be824d53405d25622a51c09e25b6497.jpg) + +Geo differs from AGCM by adopting an upsampling strategy which allows the utilization of abundant examples in $C_b$ while upsampling the instances in $C_n$ . This injects different degrees of inter-class biases for models trained by adopting AGCM and DiGeo which has been demonstrated as the primary reason for confusion in previous work [29]. At first, we observe from Fig. 4 that by adopting the upsampling based strategy, DiGeo achieves very low confusion between already learnt base classes, leading to significantly lower confusion (5% among $C_b$ and $C_n$ ). Further, confusion matrix plots in Fig. 4 show that AGCM+SMILE demonstrates 11% lower confusion than AGCM and DiGeo+SMILE shows 4% lower confusion and DiGeo. This proves the efficacy of combinatorial objectives ( $L_{comb}^{inter}$ ) in mitigating inter-class bias, thereby reducing confusion between classes. + +# 5 Conclusion + +In this work, we have presented a novel approach to Few-Shot Object Detection (FSOD) by introducing a combinatorial viewpoint through the SMILE framework. By leveraging the properties of set-based combinatorial functions, SMILE aims to address the challenges of class confusion and catastrophic forgetting, which are prevalent in FSOD tasks. Our approach incorporates Submodular Mutual Information (SMI) and Submodular Information Measures (SIM) to penalize overlapping features between base and novel classes and to ensure the formation of compact feature clusters, respectively. The experimental results on PASCAL-VOC and MS-COCO benchmarks demonstrate the effectiveness of SMILE, showing significant improvements in novel class performance, faster convergence, and a reduction in class confusion and catastrophic forgetting. Overall, SMILE offers a promising direction for advancing the state-of-the-art in FSOD by providing a generalized framework that is adaptable to various underlying architectures and capable of handling the complexities associated with few-shot learning in object detection. + +# Acknowledgements + +We gratefully thank anonymous reviewers for their valuable comments. This work is supported by the National Science Foundation under Grant Numbers IIS-2106937, a gift from Google Research, an Amazon Research Award, and the Adobe Data Science Research award. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the National Science Foundation, Google or Adobe. + +# References + +1. Agarwal, A., Majee, A., Subramanian, A., Arora, C.: Attention guided cosine margin to overcome class-imbalance in few-shot road object detection. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) Workshops. pp. 221-230 (2022) +2. Cao, Y., Wang, J., Jin, Y., Wu, T., Chen, K., Liu, Z., Lin, D.: Few-shot object detection via association and discrimination. In: Thirty-Fifth Conference on Neural Information Processing Systems (2021) +3. Chen, H., Wang, Y., Wang, G., Qiao, Y.: LSTD: A Low-Shot Transfer Detector For Object Detection. In: AAAI. pp. 2836-2843 (2018) +4. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2021) +5. Everingham, M., Van Gool, L., Williams, C.K.I., Winn, J., Zisserman, A.: The Pascal Visual Object Classes (VOC) Challenge. IJCV pp. 303-338 (2010) +6. Everingham, M., Van Gool, L., Williams, C., Winn, J., Zisserman, A.: The Pascal visual object classes (voc) challenge. International Journal of Computer Vision 88, 303-338 (06 2010) +7. Fan, Q., Zhuo, W., Tang, C.K., Tai, Y.W.: Few-Shot Object Detection With Attention-RPN And Multi-Relation Detector. In: CVPR (2020) +8. Fan, Z., Ma, Y., Li, Z., Sun, J.: Generalized Few-Shot Object Detection Without Forgetting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4527-4536 (June 2021) +9. Fan, Z., Ma, Y., Li, Z., Sun, J.: Generalized few-shot object detection without forgetting (2021) +0. Finn, C., Abbeel, P., Levine, S.: Model-Agnostic Meta-Learning For Fast Adaptation Of Deep Networks. In: ICML (2017) +1. Fujishige, S.: Submodular Functions and Optimization, vol. 58. Elsevier (2005) +2. Gidaris, S., Komodakis, N.: Dynamic Few-Shot Visual Learning Without Forgetting. In: CVPR (2018) +3. Girshick, R.B.: Fast R-CNN. ICCV (2015) +4. Han, G., Ma, J., Huang, S., Chen, L., Chang, S.F.: Few-shot object detection with fully cross-transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5321-5330 (2022) +5. Jegelka, S., Bilmes, J.: Submodularity beyond submodular energies: Coupling edges in graph cuts. In: CVPR 2011 (2011) + +16. Kang, B., Liu, Z., Wang, X., Yu, F., Feng, J., Darrell, T.: Few-shot Object Detection Via Feature Reweighting. In: ICCV (2019) +17. Karlinsky, L., Shtok, J., Harary, S., Schwartz, E., Aides, A., Feris, R., Giryes, R., Bronstein, A.M.: RepMet: Representative-Based Metric Learning For Classification And Few-Shot Object Detection. In: CVPR (2019) +18. Kaul, P., Xie, W., Zisserman, A.: Label, Verify, Correct: A Simple Few-Shot Object Detection Method. In: IEEE Conference on Computer Vision and Pattern Recognition (2022) +19. Kaushal, V., Iyer, R., Doctor, K., Sahoo, A., Dubal, P., Kothawade, S., Mahadev, R., Dargan, K., Ramakrishnan, G.: Demystifying multi-faceted video summarization: Tradeoff between diversity, representation, coverage and importance. In: 2019 IEEE Winter Conference on Applications of Computer Vision (WACV). pp. 452-461 (2019) +20. Khosla, P., Teterwak, P., Wang, C., Sarna, A., Tian, Y., Isola, P., Maschinot, A., Liu, C., Krishnan, D.: Supervised contrastive learning. In: Advances in Neural Information Processing Systems (2020) +21. Kothawade, S., Ghosh, S., Shekhar, S., Xiang, Y., Iyer, R.K.: Talisman: Targeted active learning for object detection with rare classes and slices using submodular mutual information. In: Computer Vision - ECCV 2022 - 17th European Conference (2022) +22. Kothawade, S., Kaushal, V., Ramakrishnan, G., Bilmes, J.A., Iyer, R.K.: PRISM: A rich class of parameterized submodular information measures for guided data subset selection. In: Thirty-Sixth AAAI Conference on Artificial Intelligence, AAAI. pp. 10238-10246 (2022) +23. Li, B., Liu, C., Shi, M., Chen, X., Ji, X., Ye, Q.: Proposal distribution calibration for few-shot object detection. IEEE transactions on neural networks and learning systems (2022) +24. Li, B., Yang, B., Liu, C., Liu, F., Ji, R., Ye, Q.: Beyond Max-Margin: Class Margin Equilibrium For Few-Shot Object Detection. In: CVPR (June 2021) +25. Lin, H., Bilmes, J.: A class of submodular functions for document summarization. In: Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies (2011) +26. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision - ECCV 2014. pp. 740-755. Springer International Publishing, Cham (2014) +27. Liu, F., Zhang, X., Peng, Z., Guo, Z., Wan, F., Ji, X., Ye, Q.: Integrally Migrating Pre-trained Transformer Encoder-decoders for Visual Object Detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6825-6834 (2023) +28. Ma, J., Niu, Y., Xu, J., Huang, S., Han, G., Chang, S.F.: Digeo: Discriminative geometry-aware learning for generalized few-shot object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2023) +29. Majee, A., Agrawal, K., Subramanian, A.: Few-Shot Learning For Road Object Detection. In: AAAI Workshop on Meta-Learning and MetaDL Challenge. vol. 140, pp. 115-126 (2021) +30. Majee, A., Kothawade, S.N., Killamsetty, K., Iyer, R.K.: SCoRe: Submodular Combinatorial Representation Learning. In: Forty-first International Conference on Machine Learning (ICML) (2024) + +31. Ren, S., He, K., Girshick, R.B., Sun, J.: Faster R-CNN: Towards Real-Time Object Detection With Region Proposal Networks. IEEE Trans. on Pattern Analysis and Machine Intelligence (2015) +32. Shangguan, Z., Rostami, M.: Identification of novel classes for improving few-shot object detection (2023) +33. Snell, J., Swersky, K., Zemel, R.: Prototypical Networks For Few-shot Learning. In: NeurIPS. pp. 4077-4087 (2017) +34. Sun, B., Li, B., Cai, S., Yuan, Y., Zhang, C.: FSCE: Few-Shot Object Detection Via Contrastive Proposal Encoding. In: CVPR (June 2021) +35. Sung, F., Yang, Y., Zhang, L., Xiang, T., Torr, P.H., Hospedales, T.M.: Learning To Compare: Relation Network For Few-Shot Learning. In: CVPR (June 2018) +36. Vinyals, O., Blundell, C., Lillicrap, T., kavukcuoglu, k., Wierstra, D.: Matching Networks For One Shot Learning. In: NeurIPS (2016) +37. Wang, X., Huang, T.E., Darrell, T., Gonzalez, J.E., Yu, F.: Frustratingly Simple Few-Shot Object Detection. In: ICML (2020) +38. Wang, Y.X., Ramanan, D., Hebert, M.: Meta-Learning To Detect Rare Objects. In: ICCV (2019) +39. Wu, J., Liu, S., Huang, D., Wang, Y.: Multi-scale positive sample refinement for few-shot object detection. In: European Conference on Computer Vision (2020) +40. Xiao, Y., Marlet, R.: Few-Shot Object Detection And Viewpoint Estimation For Objects In The Wild. In: ECCV (2020) +41. Yan, X., Chen, Z., Xu, A., Wang, X., Liang, X., Lin, L.: Meta R-CNN: Towards General Solver For Instance-Level Low-Shot Learning. In: CVPR. pp. 9577-9586 (2019) +42. Zhang, G., Cui, K., Wu, R., Lu, S., Tian, Y.: PNPDet: Efficient Few-Shot Detection Without Forgetting Via Plug-And-Play Sub-Networks. In: WACV. pp. 3823-3832 (2021) +43. Zhang, L., Zhou, S., Guan, J., Zhang, J.: Accurate Few-Shot Object Detection With Support-Query Mutual Guidance And Hybrid Loss. In: CVPR (June 2021) +44. Zhang, S., Luo, D., Wang, L., Koniusz, P.: Few-Shot Object Detection By Second-order Pooling. In: Proceedings of the Asian Conference on Computer Vision (ACCV) (November 2020) +45. Zhu, C., Chen, F., Ahmed, U., Shen, Z., Savvides, M.: Semantic Relation Reasoning For Shot-Stable Few-Shot Object Detection. In: CVPR (June 2021) \ No newline at end of file diff --git a/2024/SMILe_ Leveraging Submodular Mutual Information For Robust Few-Shot Object Detection/images.zip b/2024/SMILe_ Leveraging Submodular Mutual Information For Robust Few-Shot Object Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8e85f5b6aea1a4dd8c190a4fdebfeb5d67207ff7 --- /dev/null +++ b/2024/SMILe_ Leveraging Submodular Mutual Information For Robust Few-Shot Object Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aabb9b38bcfbb1d46671b7083dfde316ad4a1c7f4e9f2842eebfd858a1719050 +size 508473 diff --git a/2024/SMILe_ Leveraging Submodular Mutual Information For Robust Few-Shot Object Detection/layout.json b/2024/SMILe_ Leveraging Submodular Mutual Information For Robust Few-Shot Object Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0d8d80f6d60afb5e65ec59facc13d0221688c287 --- /dev/null +++ b/2024/SMILe_ Leveraging Submodular Mutual Information For Robust Few-Shot Object Detection/layout.json @@ -0,0 +1,10729 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 163, + 111, + 451, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 163, + 111, + 451, + 163 + ], + "spans": [ + { + "bbox": [ + 163, + 111, + 451, + 163 + ], + "type": "text", + "content": "SMILE: Leveraging Submodular Mutual Information For Robust Few-Shot Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 190, + 186, + 424, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 186, + 424, + 200 + ], + "spans": [ + { + "bbox": [ + 190, + 186, + 424, + 200 + ], + "type": "text", + "content": "Anay Majee1, Ryan Sharp2, and Rishabh Iyer1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 211, + 207, + 402, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 207, + 402, + 219 + ], + "spans": [ + { + "bbox": [ + 211, + 207, + 402, + 219 + ], + "type": "text", + "content": "1 The University of Texas at Dallas, TX, USA" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 211, + 219, + 399, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 219, + 399, + 230 + ], + "spans": [ + { + "bbox": [ + 211, + 219, + 399, + 230 + ], + "type": "text", + "content": "{anay.majee, rishabh.iley}@utdallas.edu" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 256, + 230, + 357, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 230, + 357, + 241 + ], + "spans": [ + { + "bbox": [ + 256, + 230, + 357, + 241 + ], + "type": "text", + "content": "2 IGS Energy, OH, USA" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 256, + 242, + 357, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 242, + 357, + 252 + ], + "spans": [ + { + "bbox": [ + 256, + 242, + 357, + 252 + ], + "type": "text", + "content": "rysharp@igsenergy.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 283, + 455, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 283, + 455, + 504 + ], + "spans": [ + { + "bbox": [ + 159, + 283, + 455, + 504 + ], + "type": "text", + "content": "Abstract. Confusion and forgetting of object classes have been challenges of prime interest in Few-Shot Object Detection (FSOD). To overcome these pitfalls in metric learning based FSOD techniques, we introduce a novel Submodular Mutual Information Learning (SMILE3) framework for loss functions which adopts combinatorial mutual information functions as learning objectives to enforce learning of well-separated feature clusters between the base and novel classes. Additionally, the joint objective in SMILE minimizes the total submodular information contained in a class leading to discriminative feature clusters. The combined effect of this joint objective demonstrates significant improvements in class confusion and forgetting in FSOD. Further we show that SMILE generalizes to several existing approaches in FSOD, improving their performance, agnostic of the backbone architecture. Experiments on popular FSOD benchmarks, PASCAL-VOC and MS-COCO show that our approach generalizes to State-of-the-Art (SoTA) approaches improving their novel class performance by up to " + }, + { + "bbox": [ + 159, + 283, + 455, + 504 + ], + "type": "inline_equation", + "content": "5.7\\%" + }, + { + "bbox": [ + 159, + 283, + 455, + 504 + ], + "type": "text", + "content": " (3.3 mAP points) and " + }, + { + "bbox": [ + 159, + 283, + 455, + 504 + ], + "type": "inline_equation", + "content": "5.4\\%" + }, + { + "bbox": [ + 159, + 283, + 455, + 504 + ], + "type": "text", + "content": " (2.6 mAP points) on the 10-shot setting of VOC (split 3) and 30-shot setting of COCO datasets respectively. Our experiments also demonstrate better retention of base class performance and up to " + }, + { + "bbox": [ + 159, + 283, + 455, + 504 + ], + "type": "inline_equation", + "content": "2\\times" + }, + { + "bbox": [ + 159, + 283, + 455, + 504 + ], + "type": "text", + "content": " faster convergence over existing approaches agnostic of the underlying architecture." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 525, + 230, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 230, + 537 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 230, + 537 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 550, + 482, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 550, + 482, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 550, + 482, + 635 + ], + "type": "text", + "content": "Recent advances in Deep Neural networks (DNNs) have enabled models to learn discriminative feature representations from large-scale image benchmarks. Unfortunately, these architectures fail to adapt to few-shot settings tasked to recognize novel objects over existing ones with few examples, closely resembling human-like perception. Although recent research has shown significant promise in few-shot image recognition [10, 12, 33, 35, 36], Few-Shot Object Detection (FSOD) remains a challenge with recent works [27, 28, 37, 38] highlighting two" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 642, + 436, + 653 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 436, + 653 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 436, + 653 + ], + "type": "text", + "content": "* Work done as a Graduate student at The University of Texas at Dallas." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 653, + 448, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 448, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 448, + 665 + ], + "type": "text", + "content": "3 Project page: https://anaymajee.me/ assets/project_pages/smile.html." + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 153, + 118, + 460, + 244 + ], + "blocks": [ + { + "bbox": [ + 153, + 118, + 460, + 244 + ], + "lines": [ + { + "bbox": [ + 153, + 118, + 460, + 244 + ], + "spans": [ + { + "bbox": [ + 153, + 118, + 460, + 244 + ], + "type": "image", + "image_path": "1ea79b0189d765270c8b55465a89be4be6ce67ec6f448f069b3c494f4fecaab2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 254, + 482, + 288 + ], + "lines": [ + { + "bbox": [ + 130, + 254, + 482, + 288 + ], + "spans": [ + { + "bbox": [ + 130, + 254, + 482, + 288 + ], + "type": "text", + "content": "Fig. 1: Functionality of components in " + }, + { + "bbox": [ + 130, + 254, + 482, + 288 + ], + "type": "inline_equation", + "content": "L_{comb}" + }, + { + "bbox": [ + 130, + 254, + 482, + 288 + ], + "type": "text", + "content": " proposed in the SMILE (ours) framework, (a) " + }, + { + "bbox": [ + 130, + 254, + 482, + 288 + ], + "type": "inline_equation", + "content": "L_{comb}^{inter}" + }, + { + "bbox": [ + 130, + 254, + 482, + 288 + ], + "type": "text", + "content": " promotes separation between " + }, + { + "bbox": [ + 130, + 254, + 482, + 288 + ], + "type": "inline_equation", + "content": "C_b" + }, + { + "bbox": [ + 130, + 254, + 482, + 288 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 254, + 482, + 288 + ], + "type": "inline_equation", + "content": "C_n" + }, + { + "bbox": [ + 130, + 254, + 482, + 288 + ], + "type": "text", + "content": " while (c) " + }, + { + "bbox": [ + 130, + 254, + 482, + 288 + ], + "type": "inline_equation", + "content": "L_{comb}^{intra}" + }, + { + "bbox": [ + 130, + 254, + 482, + 288 + ], + "type": "text", + "content": " promotes intra-class compactness." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 305, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 305, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 305, + 482, + 510 + ], + "type": "text", + "content": "major challenges - Class Confusion and Catastrophic Forgetting. Class confusion, as highlighted in [29] manifests itself through mis-prediction of instances belonging to a newly learnt (novel) class, as one or more instances of the already learnt (base) classes. Authors in [1, 38] attribute this to the sharing of visual information between classes resulting in increased inter-class bias due to overlapping feature clusters as shown in Fig. 1(a). Catastrophic forgetting refers to the gradual degradation in the performance of already learnt classes in the quest to learn the novel ones, as shown in Fig. 2(a), seldom overfitting to rare classes [29, 38]. Further, large feature diversity (intra-class variance) among base classes lead to formation of non-discriminative feature clusters as shown in Fig. 1(b), aggravating the existing inter-class bias in the feature space. Unlike existing approaches (refer Sec. 2) which target either confusion or forgetting, our paper presents a unified approach to tackle both these challenges in FSOD. Although, recent approaches [1, 28, 34] attempt to tackle these challenges through contrastive learning strategies, such approaches have been limited by their capability to overcome either inter-class bias or intra-class variance [30, 38] and poor generalization to longtail settings [30] (FSOD being a extreme case)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we introduce a combinatorial viewpoint in FSOD considering each object class " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "i \\in [1, C]" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " in the dataset " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " as a set " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "A_i" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " of samples, where " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{T} = \\{A_1, \\dots, A_C\\}" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": ", facilitating the application of combinatorial functions as learning objectives. We aim to overcome the aforementioned challenges through representation learning in the low-data regime by adopting this formulation through the SMILE: Submodular Mutual Information Learning framework, wherein we introduce novel, set-based combinatorial objective functions for FSOD as shown in Fig. 3. SMILE introduces a joint objective formulation " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "L_{comb}" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " (Eq. (3)) based on two popular flavors of submodular information functions - Submodular Mutual Information [22] (SMI) and Total Submodular Information [11] targeting the root causes of confusion and forgetting in FSOD. At first, SMILE is the first to introduce pairwise SMI functions " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "I_f" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " in representation learning which model the common (overlapping) information between two" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "type": "text", + "content": "Majee et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 151, + 114, + 302, + 223 + ], + "blocks": [ + { + "bbox": [ + 151, + 114, + 302, + 223 + ], + "lines": [ + { + "bbox": [ + 151, + 114, + 302, + 223 + ], + "spans": [ + { + "bbox": [ + 151, + 114, + 302, + 223 + ], + "type": "image", + "image_path": "de43313bc2880b20646758366c8511fda729561914c5d681b588d8c3844dd8d4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 232, + 482, + 277 + ], + "lines": [ + { + "bbox": [ + 130, + 232, + 482, + 277 + ], + "spans": [ + { + "bbox": [ + 130, + 232, + 482, + 277 + ], + "type": "text", + "content": "Fig. 2: Resilience to Catastrophic forgetting and faster convergence in SMILE over SoTA approaches. (a) shows that combinatorial losses in SMILE are robust to catastrophic forgetting, while (b) shows that objectives in SMILE results in faster convergence over SoTA FSOD methods (AGCM and DiGeo)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 312, + 114, + 462, + 223 + ], + "blocks": [ + { + "bbox": [ + 312, + 114, + 462, + 223 + ], + "lines": [ + { + "bbox": [ + 312, + 114, + 462, + 223 + ], + "spans": [ + { + "bbox": [ + 312, + 114, + 462, + 223 + ], + "type": "image", + "image_path": "a9150b517d2da2d94c869bc75c30efe377be18107d2e72d210e55782362a4c13.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 290, + 482, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 290, + 482, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 290, + 482, + 506 + ], + "type": "text", + "content": "sets. Minimizing " + }, + { + "bbox": [ + 130, + 290, + 482, + 506 + ], + "type": "inline_equation", + "content": "I_{f}" + }, + { + "bbox": [ + 130, + 290, + 482, + 506 + ], + "type": "text", + "content": " through the joint objective " + }, + { + "bbox": [ + 130, + 290, + 482, + 506 + ], + "type": "inline_equation", + "content": "L_{comb}" + }, + { + "bbox": [ + 130, + 290, + 482, + 506 + ], + "type": "text", + "content": " reduces feature overlap between base and novel classes alleviating inter-class bias in the model towards abundantly sampled classes as shown in Fig. 1(b). We extend this property of SMI functions to the novel classes minimizing the inter-cluster overlap between few-shot classes, promoting learning of discriminative features from just few samples. Secondly, SMILE preserves the diversity within each class by minimizing the total submodular information contained within each set as shown in Fig. 1(c), minimizing the impact of forgetting. This formulation closely follows the observation in [30] which models cooperation [15] between instances in a set by minimizing a submodular function over a set, to preserve representative features. The unified objective " + }, + { + "bbox": [ + 130, + 290, + 482, + 506 + ], + "type": "inline_equation", + "content": "L_{comb}" + }, + { + "bbox": [ + 130, + 290, + 482, + 506 + ], + "type": "text", + "content": " introduced in SMILE models both these necessary properties through a weighted sum of two distinct objectives " + }, + { + "bbox": [ + 130, + 290, + 482, + 506 + ], + "type": "inline_equation", + "content": "L_{comb}^{inter}" + }, + { + "bbox": [ + 130, + 290, + 482, + 506 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 290, + 482, + 506 + ], + "type": "inline_equation", + "content": "L_{comb}^{intra}" + }, + { + "bbox": [ + 130, + 290, + 482, + 506 + ], + "type": "text", + "content": " as shown in Fig. 3 balancing the tradeoff between inter-cluster separation and intra-cluster compactness respectively. This allows us to introduce a family of loss functions which inherently eliminates confusion and forgetting as shown in Tab. 5. We conduct our experiments on two popular FSOD benchmarks, PASCAL-VOC [6] and MS-COCO [26] for several few-shot settings and demonstrate the following contributions of SMILE:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 511, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 138, + 511, + 481, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 511, + 481, + 559 + ], + "spans": [ + { + "bbox": [ + 138, + 511, + 481, + 559 + ], + "type": "text", + "content": "- SMILE introduces a novel set-based combinatorial viewpoint in FSOD by applying combinatorial Mutual Information based objective to discriminate between base and novel classes, in conjunction with submodular total information to minimize intra-class variance as the objective function." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 559, + 481, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 559, + 481, + 594 + ], + "spans": [ + { + "bbox": [ + 138, + 559, + 481, + 594 + ], + "type": "text", + "content": "- SMILE generalizes to existing approaches in FSOD, irrespective of the underlying architecture demonstrating up to " + }, + { + "bbox": [ + 138, + 559, + 481, + 594 + ], + "type": "inline_equation", + "content": "5.7\\%" + }, + { + "bbox": [ + 138, + 559, + 481, + 594 + ], + "type": "text", + "content": " improvement in novel class performance (Tab. 2) over the baseline FSOD approach." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 594, + 481, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 594, + 481, + 618 + ], + "spans": [ + { + "bbox": [ + 138, + 594, + 481, + 618 + ], + "type": "text", + "content": "SMILE demonstrates up to " + }, + { + "bbox": [ + 138, + 594, + 481, + 618 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 138, + 594, + 481, + 618 + ], + "type": "text", + "content": " faster convergence (Fig. 2(b)) over existing SoTA approaches resulting in faster generalization to unknown object classes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 618, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 618, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 618, + 481, + 665 + ], + "type": "text", + "content": "- Finally, SMILE demonstrates up to " + }, + { + "bbox": [ + 138, + 618, + 481, + 665 + ], + "type": "inline_equation", + "content": "11\\%" + }, + { + "bbox": [ + 138, + 618, + 481, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 138, + 618, + 481, + 665 + ], + "type": "inline_equation", + "content": "3.5\\%" + }, + { + "bbox": [ + 138, + 618, + 481, + 665 + ], + "type": "text", + "content": " reduction in class confusion and catastrophic forgetting while achieving SoTA performance on popular FSOD benchmarks like PASCAL-VOC (by " + }, + { + "bbox": [ + 138, + 618, + 481, + 665 + ], + "type": "inline_equation", + "content": "5.7\\%" + }, + { + "bbox": [ + 138, + 618, + 481, + 665 + ], + "type": "text", + "content": " on split 2, 10-shot setting) and MS-COCO (" + }, + { + "bbox": [ + 138, + 618, + 481, + 665 + ], + "type": "inline_equation", + "content": "5.4\\%" + }, + { + "bbox": [ + 138, + 618, + 481, + 665 + ], + "type": "text", + "content": " on 30-shot setting)." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 246, + 91, + 448, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 91, + 448, + 101 + ], + "spans": [ + { + "bbox": [ + 246, + 91, + 448, + 101 + ], + "type": "text", + "content": "SMILE: Submodular Mutual Information Learner" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 305 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 305 + ], + "type": "text", + "content": "Few-Shot Object Detection (FSOD): Classical FSOD approaches utilize finetuning [3] or distance metric learning [17] to adapt features to novel classes. Recent methods employ meta-learning techniques [16,40,41] with episodic training to learn class-specific features. Meta-Reweight [16] and Meta-RCNN [41] use additional feature extractors, while Add-Info [40] leverages feature differences between support and query images. Techniques like [43] enhance class-specific features through information sharing, and CME [24] aims to reduce class confusion. Attention mechanisms [7,44] are used to identify discriminative features. However, meta-learning approaches are resource-intensive and may fail to generalize to significantly different novel classes. Metric learning strategies like FsDet [37], FSCE [34], and SRR-FSD [45] offer better generalization without additional overheads. PNPDet [42] partially addresses catastrophic forgetting and class confusion. GFSD [8] proposes a Bias-Balanced RPN to mitigate overfitting in metric learners." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 306, + 482, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 306, + 482, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 306, + 482, + 461 + ], + "type": "text", + "content": "Recent approaches like [18,32] adopt weak supervision from unlabelled data or low confidence predictions in RoI pooling layers to generalize to novel classes. These methods often use abundant samples from base classes [28] to prevent catastrophic forgetting, adding computational overhead in low-shot settings. Vision transformers [4] have been adopted in FSOD through methods like imTED [27] and PDC [23], with reduced computational overhead by using pre-trained attention heads. Alternatively, DiGeo [28] and PDC [23] learn the geometry or difference in distributions of RoI proposals [13] between object classes to overcome forgetting and confusion. However, these approaches rely on contrastive learning objectives [20] that struggle to learn discriminative feature embeddings due to adoption of pairwise similarity metrics. Our work, SMILE, aims to improve the feature learning capacity of existing SoTA approaches, irrespective of their underlying architectures." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "content": "Submodular Functions and Combinatorial Objectives: Submodular functions are recognized as set functions with an inherent diminishing returns characteristic. Defined as a set function " + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "inline_equation", + "content": "f: 2^{\\mathcal{V}} \\to \\mathbb{R}" + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "content": " operating on a ground-set " + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "content": ", a function is termed submodular if it adheres to the condition " + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "inline_equation", + "content": "f(X) + f(Y) \\geq f(X \\cup Y) + f(X \\cap Y), \\forall X, Y \\subseteq \\mathcal{V}" + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "content": " [11]. These functions have garnered considerable attention in research, particularly in fields like data subset selection [22], active learning [21], and video summarization [19,22] through their ability in modeling concepts such as diversity, relevance, set-cover and representation. A subclass of submodular functions, namely Submodular Mutual Information (SMI) functions introduced in [22] model the similarity and diversity between pairs of object classes establishing itself as a powerful tool to model inter-class bias. Recently, Majee et al. [30] introduces these set-based combinatorial functions as objectives in representation learning and demonstrates their capability in overcoming inter-class bias (by minimizing the similarity between nonidentical object classes) and intra-class variance (maximizing the similarity between instances of the same object class). However, these functions are yet to be studied in the context of few-shot learning. We introduce novel instances of SMI based objectives in" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Majee et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 130, + 114, + 482, + 248 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 248 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 248 + ], + "type": "image", + "image_path": "41d46461f410f09a8a7a20db24290b5e43a772732f76334dd723fb4d8a6a0cd8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 258, + 482, + 292 + ], + "lines": [ + { + "bbox": [ + 130, + 258, + 482, + 292 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 482, + 292 + ], + "type": "text", + "content": "Fig. 3: Overview of our SMILE framework highlighting the application of Mutual Information function based objectives in SMILE for the fine-tuning stage of Few-Shot Object Detection." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 306, + 482, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 306, + 482, + 367 + ], + "spans": [ + { + "bbox": [ + 130, + 306, + 482, + 367 + ], + "type": "text", + "content": "SMILEs to minimize inter-class bias between base and novel classes. To the best of our knowledge we are the first to introduce novel SMI based combinatorial objectives in conjunction with total information based combinatorial functions through SMILEs in a quest to minimize confusion and forgetting in few-shot object detection." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 384, + 201, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 384, + 201, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 384, + 201, + 396 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 407, + 405, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 407, + 405, + 419 + ], + "spans": [ + { + "bbox": [ + 131, + 407, + 405, + 419 + ], + "type": "text", + "content": "3.1 Problem Definition : Few-Shot Object Detection" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": "We define a few-shot learner " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "h(x, \\theta)" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": " as shown in Fig. 3 that receives input data " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": " from base classes " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "C_b \\in [1, |C_b|]" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": " and novel classes " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "C_n \\in [1, |C_n|]" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "C = \\{C_b \\cup C_n\\}" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\{C_b \\cap C_n\\} = \\emptyset" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": " denotes the learnable parameters. The training data can be divided into two distinct parts, base " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "D_{base}" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": " and novel " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "D_{novel}" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": " such that, " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{T} = \\{D_{base} \\cup D_{novel}\\}" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\{D_{base} \\cap D_{novel}\\} = \\emptyset" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": ". SMILe introduces a paradigm shift in FSOD by imbibing a combinatorial viewpoint, where the base dataset, " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "D_{base} = [A_1^b, A_2^b, \\dots, A_{|C_b|}^b]" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": ", containing abundant training examples from " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "C_b" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": " base classes and the novel dataset, " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "D_{novel} = [A_1^n, A_2^n, \\dots, A_{|C_n|}^n]" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": " containing only " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": "-shot ( " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "|A_i^n| = K" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "i \\in [1, C_n]" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": ") training examples from " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "C_n" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": " novel classes. The objective of the few-shot learner " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "h(x, \\theta)" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": " is to learn discriminative representation from classes in " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "D_{novel}" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": " without degradation in performance on classes in " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "D_{base}" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": ". Following FSCE [34] we adopt a two-stage training strategy. In the base training stage we train " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "h(x, \\theta)" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": " on abundant samples in " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "D_{base}" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": ", allowing the model to generalize on the domain of " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "D_{base}" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": ". The few-shot adaptation stage adapts " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "h(x, \\theta)" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": " to previously unseen " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": "-shot data by fine-tuning on data samples from " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "D_{base} \\cup D_{novel}" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "|A_k| = K" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "k \\in \\{C_b \\cup C_n\\}" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": ". The goal of SMILe is to overcome class confusion and forgetting in FSOD resulting from elevated inter-class bias and intra-class variance as observed in [1, 29, 38]. The final model " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "h(x, \\theta)" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": " obtained after two training stages is evaluated on " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "D_{test}" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": " containing unseen data samples from both " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "C_b \\cup C_n" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 246, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 246, + 91, + 447, + 101 + ], + "type": "text", + "content": "SMILE: Submodular Mutual Information Learner" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 278, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 278, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 278, + 127 + ], + "type": "text", + "content": "3.2 The SMILE Framework" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 134, + 482, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 134, + 482, + 266 + ], + "spans": [ + { + "bbox": [ + 130, + 134, + 482, + 266 + ], + "type": "text", + "content": "Adopting a combinatorial viewpoint as disclosed earlier allows us to employ submodular combinatorial functions as learning objectives to tackle confusion and forgetting in FSOD. As discussed in Sec. 2, minimizing a Submodular functions naturally models cooperation [15] while maximizing it models diversity [25] due to their inherent diminishing marginal returns property. SMILE adopts the aforementioned properties of submodular functions to define a novel family of combinatorial objective (loss) functions " + }, + { + "bbox": [ + 130, + 134, + 482, + 266 + ], + "type": "inline_equation", + "content": "L_{comb}(\\theta)" + }, + { + "bbox": [ + 130, + 134, + 482, + 266 + ], + "type": "text", + "content": " which enforces orthogonality in the feature space when applied on Region-of-Interest (RoI) features in FSOD models. The loss function " + }, + { + "bbox": [ + 130, + 134, + 482, + 266 + ], + "type": "inline_equation", + "content": "L_{comb}(\\theta)" + }, + { + "bbox": [ + 130, + 134, + 482, + 266 + ], + "type": "text", + "content": " can be decomposed into two major components - " + }, + { + "bbox": [ + 130, + 134, + 482, + 266 + ], + "type": "inline_equation", + "content": "L_{comb}^{inter}" + }, + { + "bbox": [ + 130, + 134, + 482, + 266 + ], + "type": "text", + "content": " minimizes inter-class bias between base and novel classes and " + }, + { + "bbox": [ + 130, + 134, + 482, + 266 + ], + "type": "inline_equation", + "content": "L_{comb}^{intra}" + }, + { + "bbox": [ + 130, + 134, + 482, + 266 + ], + "type": "text", + "content": " maximizes intra-class compactness within abundant classes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 266, + 482, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 266, + 482, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 482, + 411 + ], + "type": "text", + "content": "For " + }, + { + "bbox": [ + 130, + 266, + 482, + 411 + ], + "type": "inline_equation", + "content": "L_{comb}^{inter}" + }, + { + "bbox": [ + 130, + 266, + 482, + 411 + ], + "type": "text", + "content": ", SMILE explores a sub-category of combinatorial functions, namely Submodular Mutual Information (SMI) which can be defined as " + }, + { + "bbox": [ + 130, + 266, + 482, + 411 + ], + "type": "inline_equation", + "content": "I_{f}(A_{i}, A_{j}) = f(A_{i}) + f(A_{j}) - f(A_{i} \\cup A_{j})" + }, + { + "bbox": [ + 130, + 266, + 482, + 411 + ], + "type": "text", + "content": " [11, 22], and models the common information between two sets " + }, + { + "bbox": [ + 130, + 266, + 482, + 411 + ], + "type": "inline_equation", + "content": "A_{i}" + }, + { + "bbox": [ + 130, + 266, + 482, + 411 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 266, + 482, + 411 + ], + "type": "inline_equation", + "content": "A_{j}" + }, + { + "bbox": [ + 130, + 266, + 482, + 411 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 266, + 482, + 411 + ], + "type": "inline_equation", + "content": "\\forall i, j \\in \\mathcal{T}" + }, + { + "bbox": [ + 130, + 266, + 482, + 411 + ], + "type": "text", + "content": ". Results in [11, 22] portray " + }, + { + "bbox": [ + 130, + 266, + 482, + 411 + ], + "type": "inline_equation", + "content": "I_{f}(A_{i}, A_{j}; \\theta)" + }, + { + "bbox": [ + 130, + 266, + 482, + 411 + ], + "type": "text", + "content": " as a measure of the degree of similarity between object classes " + }, + { + "bbox": [ + 130, + 266, + 482, + 411 + ], + "type": "inline_equation", + "content": "A_{i}" + }, + { + "bbox": [ + 130, + 266, + 482, + 411 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 266, + 482, + 411 + ], + "type": "inline_equation", + "content": "A_{j}" + }, + { + "bbox": [ + 130, + 266, + 482, + 411 + ], + "type": "text", + "content": ". Adopting this definition of SMI, " + }, + { + "bbox": [ + 130, + 266, + 482, + 411 + ], + "type": "inline_equation", + "content": "L_{comb}^{inter}" + }, + { + "bbox": [ + 130, + 266, + 482, + 411 + ], + "type": "text", + "content": " minimizes the SMI between the base " + }, + { + "bbox": [ + 130, + 266, + 482, + 411 + ], + "type": "inline_equation", + "content": "C_{b}" + }, + { + "bbox": [ + 130, + 266, + 482, + 411 + ], + "type": "text", + "content": " and the novel " + }, + { + "bbox": [ + 130, + 266, + 482, + 411 + ], + "type": "inline_equation", + "content": "C_{n}" + }, + { + "bbox": [ + 130, + 266, + 482, + 411 + ], + "type": "text", + "content": " classes, ensuring sufficient inter-cluster separation (by minimizing inter-class bias) as shown in Eq. (1). " + }, + { + "bbox": [ + 130, + 266, + 482, + 411 + ], + "type": "inline_equation", + "content": "L_{comb}^{inter}" + }, + { + "bbox": [ + 130, + 266, + 482, + 411 + ], + "type": "text", + "content": " further minimizes the mutual information between classes in " + }, + { + "bbox": [ + 130, + 266, + 482, + 411 + ], + "type": "inline_equation", + "content": "C_{n}" + }, + { + "bbox": [ + 130, + 266, + 482, + 411 + ], + "type": "text", + "content": ", minimizing inter-cluster overlaps between the novel classes. This is visually depicted in Fig. 1(b) and has been shown to be effective in mitigating class confusion in FSOD through our experiments in Sec. 4." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 417, + 481, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 417, + 481, + 464 + ], + "spans": [ + { + "bbox": [ + 143, + 417, + 481, + 464 + ], + "type": "interline_equation", + "content": "L _ {\\text {c o m b}} ^ {\\text {i n t e r}} (\\theta) = \\sum_ {\\substack {b \\in C _ {b} \\\\ n \\in C _ {n}}} I _ {f} (A _ {b}, A _ {n}; \\theta) + \\sum_ {\\substack {i, j \\in C _ {n} \\\\ i \\neq j}} I _ {f} (A _ {i}, A _ {j}; \\theta) = \\sum_ {\\substack {i \\in (C _ {b} \\cup C _ {n}) \\\\ j \\in C _ {n}: i \\neq j}} I _ {f} (A _ {i}, A _ {j}; \\theta) \\tag{1}", + "image_path": "d23734feb84a705dcdad54e52264d379392e9d21898dd3b2772b37ab88091955.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 471, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 471, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 471, + 482, + 668 + ], + "type": "text", + "content": "In addition to confusion which stems from inter-class bias, SMILEs aims at mitigating catastrophic forgetting [29] in FSOD which has been attributed to large intra-class variance among abundant object classes in [1,38]. In coherence to the combinatorial formulation in SMILEs we achieve this through " + }, + { + "bbox": [ + 130, + 471, + 482, + 668 + ], + "type": "inline_equation", + "content": "L_{comb}^{intra}" + }, + { + "bbox": [ + 130, + 471, + 482, + 668 + ], + "type": "text", + "content": " which minimizes the Total Submodular Information, defined as " + }, + { + "bbox": [ + 130, + 471, + 482, + 668 + ], + "type": "inline_equation", + "content": "S_f(A_1, \\dots, A_{|C|}) = \\sum_{k=1}^{|C|} f(A_k; \\theta)" + }, + { + "bbox": [ + 130, + 471, + 482, + 668 + ], + "type": "text", + "content": ", over sets " + }, + { + "bbox": [ + 130, + 471, + 482, + 668 + ], + "type": "inline_equation", + "content": "A_k \\in \\mathcal{T}" + }, + { + "bbox": [ + 130, + 471, + 482, + 668 + ], + "type": "text", + "content": ", given a submodular function " + }, + { + "bbox": [ + 130, + 471, + 482, + 668 + ], + "type": "inline_equation", + "content": "f(A_k; \\theta)" + }, + { + "bbox": [ + 130, + 471, + 482, + 668 + ], + "type": "text", + "content": ". As discussed earlier, minimizing the submodular information models cooperation which asserts that minimizing " + }, + { + "bbox": [ + 130, + 471, + 482, + 668 + ], + "type": "inline_equation", + "content": "L_{comb}^{inter}" + }, + { + "bbox": [ + 130, + 471, + 482, + 668 + ], + "type": "text", + "content": " promotes learning of discriminative feature clusters, penalizing abundant classes to have large feature variance in the embedding space as shown in Fig. 1(c). Although submodular functions have been studied in the field of representation learning to minimize intra-class variance in [30], but primarily differs from SMILEs in modeling a longtail recognition task by minimizing the total submodular correlation, which models gain in information when new features are added to a set. The formulation of " + }, + { + "bbox": [ + 130, + 471, + 482, + 668 + ], + "type": "inline_equation", + "content": "L_{comb}^{intra}" + }, + { + "bbox": [ + 130, + 471, + 482, + 668 + ], + "type": "text", + "content": " has been shown in Eq. (2) where we minimize the total submodular information within samples in each class in " + }, + { + "bbox": [ + 130, + 471, + 482, + 668 + ], + "type": "inline_equation", + "content": "C_b \\cup C_n" + }, + { + "bbox": [ + 130, + 471, + 482, + 668 + ], + "type": "text", + "content": " and our experiments in Tab. 5" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "type": "text", + "content": "Majee et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 482, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 482, + 140 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 482, + 140 + ], + "type": "text", + "content": "show the effectiveness of " + }, + { + "bbox": [ + 131, + 116, + 482, + 140 + ], + "type": "inline_equation", + "content": "L_{comb}^{intra}" + }, + { + "bbox": [ + 131, + 116, + 482, + 140 + ], + "type": "text", + "content": " in boosting base class performance asserting the mitigation of catastrophic forgetting." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 175, + 150, + 481, + 176 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 150, + 481, + 176 + ], + "spans": [ + { + "bbox": [ + 175, + 150, + 481, + 176 + ], + "type": "interline_equation", + "content": "L _ {c o m b} ^ {i n t r a} (\\theta) = \\sum_ {b \\in C _ {b}} f (A _ {b}, \\theta) + \\sum_ {n \\in C _ {n}} f (A _ {n}, \\theta) = \\sum_ {k \\in \\left(C _ {b} \\cup C _ {n}\\right)} f (A _ {k}, \\theta) \\tag {2}", + "image_path": "c4e38faf1a4b2dd01e4ebd2ea5292876a18eb0e1dfc9f310dbfa1448fb82b5e3.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 185, + 481, + 208 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 185, + 481, + 208 + ], + "spans": [ + { + "bbox": [ + 131, + 185, + 481, + 208 + ], + "type": "text", + "content": "Ablating on the choice of the submodular function " + }, + { + "bbox": [ + 131, + 185, + 481, + 208 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 131, + 185, + 481, + 208 + ], + "type": "text", + "content": " and SMI functions " + }, + { + "bbox": [ + 131, + 185, + 481, + 208 + ], + "type": "inline_equation", + "content": "I_{f}" + }, + { + "bbox": [ + 131, + 185, + 481, + 208 + ], + "type": "text", + "content": " we introduce several instances of SMILE objectives as discussed in Tab. 1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 209, + 481, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 209, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 131, + 209, + 481, + 258 + ], + "type": "text", + "content": "Encapsulating the aforementioned formulations of " + }, + { + "bbox": [ + 131, + 209, + 481, + 258 + ], + "type": "inline_equation", + "content": "L_{comb}^{inter}" + }, + { + "bbox": [ + 131, + 209, + 481, + 258 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 209, + 481, + 258 + ], + "type": "inline_equation", + "content": "L_{comb}^{intra}" + }, + { + "bbox": [ + 131, + 209, + 481, + 258 + ], + "type": "text", + "content": " in SMILE we define a joint objective " + }, + { + "bbox": [ + 131, + 209, + 481, + 258 + ], + "type": "inline_equation", + "content": "L_{comb}(\\theta)" + }, + { + "bbox": [ + 131, + 209, + 481, + 258 + ], + "type": "text", + "content": " which tackles both the challenges of confusion and forgetting. We thus define " + }, + { + "bbox": [ + 131, + 209, + 481, + 258 + ], + "type": "inline_equation", + "content": "L_{comb}(\\theta)" + }, + { + "bbox": [ + 131, + 209, + 481, + 258 + ], + "type": "text", + "content": " in Eq. (3) which is the weighted algebraic sum of " + }, + { + "bbox": [ + 131, + 209, + 481, + 258 + ], + "type": "inline_equation", + "content": "L_{comb}^{inter}" + }, + { + "bbox": [ + 131, + 209, + 481, + 258 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 209, + 481, + 258 + ], + "type": "inline_equation", + "content": "L_{comb}^{intra}" + }, + { + "bbox": [ + 131, + 209, + 481, + 258 + ], + "type": "text", + "content": " with the weighting factor " + }, + { + "bbox": [ + 131, + 209, + 481, + 258 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 131, + 209, + 481, + 258 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 183, + 268, + 481, + 323 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 183, + 268, + 481, + 323 + ], + "spans": [ + { + "bbox": [ + 183, + 268, + 481, + 323 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {c o m b} (\\theta) = (1 - \\eta) L _ {c o m b} ^ {i n t r a} (\\theta) + \\eta L _ {c o m b} ^ {i n t e r} (\\theta) \\\\ = \\sum_ {i \\in C _ {b} \\cup C _ {n}} \\left[ (1 - \\eta) f \\left(A _ {i}, \\theta\\right) + \\eta \\sum_ {\\substack {j \\in C _ {n} \\\\ i \\neq j}} I _ {f} \\left(A _ {i}, A _ {j}; \\theta\\right) \\right] \\tag{3} \\\\ \\end{array}", + "image_path": "57ed2fba962fc419784896083f72c1aab959ec80d832dcea2df1d43d9779480d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 332, + 481, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 332, + 481, + 403 + ], + "spans": [ + { + "bbox": [ + 131, + 332, + 481, + 403 + ], + "type": "text", + "content": "Note, that the combinatorial objective " + }, + { + "bbox": [ + 131, + 332, + 481, + 403 + ], + "type": "inline_equation", + "content": "L_{comb}(\\theta)" + }, + { + "bbox": [ + 131, + 332, + 481, + 403 + ], + "type": "text", + "content": " is applied on output features from the RoI Pooling layers in proposal-based [31,38] architectures. To promote adoption of SMILE agnostic of the backbone architecture we introduce a combinatorial head " + }, + { + "bbox": [ + 131, + 332, + 481, + 403 + ], + "type": "inline_equation", + "content": "Z_{comb} = Comb(h,\\theta)" + }, + { + "bbox": [ + 131, + 332, + 481, + 403 + ], + "type": "text", + "content": " which projects the RoI features to 128-dimensional feature vectors [20], " + }, + { + "bbox": [ + 131, + 332, + 481, + 403 + ], + "type": "inline_equation", + "content": "Z_{comb}" + }, + { + "bbox": [ + 131, + 332, + 481, + 403 + ], + "type": "text", + "content": " on which " + }, + { + "bbox": [ + 131, + 332, + 481, + 403 + ], + "type": "inline_equation", + "content": "L_{comb}(\\theta)" + }, + { + "bbox": [ + 131, + 332, + 481, + 403 + ], + "type": "text", + "content": " is applied during the few-shot adaptation stage." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 404, + 481, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 404, + 481, + 488 + ], + "spans": [ + { + "bbox": [ + 131, + 404, + 481, + 488 + ], + "type": "text", + "content": "Finally, we summarize the total classification loss in SMILEs as depicted in Eq. (4) as the sum over all three objectives: the classification head " + }, + { + "bbox": [ + 131, + 404, + 481, + 488 + ], + "type": "inline_equation", + "content": "L_{Clf}" + }, + { + "bbox": [ + 131, + 404, + 481, + 488 + ], + "type": "text", + "content": ", the box regression head " + }, + { + "bbox": [ + 131, + 404, + 481, + 488 + ], + "type": "inline_equation", + "content": "L_{bbox}" + }, + { + "bbox": [ + 131, + 404, + 481, + 488 + ], + "type": "text", + "content": " and the combinatorial head " + }, + { + "bbox": [ + 131, + 404, + 481, + 488 + ], + "type": "inline_equation", + "content": "L_{comb}(\\theta)" + }, + { + "bbox": [ + 131, + 404, + 481, + 488 + ], + "type": "text", + "content": ". Note that the objectives proposed in SMILEs apply only to " + }, + { + "bbox": [ + 131, + 404, + 481, + 488 + ], + "type": "inline_equation", + "content": "Comb(h,\\theta)" + }, + { + "bbox": [ + 131, + 404, + 481, + 488 + ], + "type": "text", + "content": " while the RoI classification and regression heads are unchanged. This follows the observations in [28,38] which warrants the boost in performance originating from learning robust feature representations for each RoI predicted by the model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 220, + 498, + 481, + 512 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 498, + 481, + 512 + ], + "spans": [ + { + "bbox": [ + 220, + 498, + 481, + 512 + ], + "type": "interline_equation", + "content": "L _ {c l s} (\\theta) = L _ {C l f} (\\theta) + L _ {b b o x} (\\theta) + L _ {c o m b} (\\theta) \\tag {4}", + "image_path": "07c1963c8c38862f5bd84babc0861405b658ef678681e09262b4d9bafe8f67ab.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 529, + 463, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 529, + 463, + 544 + ], + "spans": [ + { + "bbox": [ + 131, + 529, + 463, + 544 + ], + "type": "text", + "content": "3.3 Instantiations of " + }, + { + "bbox": [ + 131, + 529, + 463, + 544 + ], + "type": "inline_equation", + "content": "L_{comb}^{inter}" + }, + { + "bbox": [ + 131, + 529, + 463, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 529, + 463, + 544 + ], + "type": "inline_equation", + "content": "L_{comb}^{intra}" + }, + { + "bbox": [ + 131, + 529, + 463, + 544 + ], + "type": "text", + "content": " in the SMILE Framework" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 549, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 549, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 131, + 549, + 482, + 667 + ], + "type": "text", + "content": "Given a submodular function " + }, + { + "bbox": [ + 131, + 549, + 482, + 667 + ], + "type": "inline_equation", + "content": "f(A)" + }, + { + "bbox": [ + 131, + 549, + 482, + 667 + ], + "type": "text", + "content": " and a Submodular Mutual Information (SMI) function " + }, + { + "bbox": [ + 131, + 549, + 482, + 667 + ], + "type": "inline_equation", + "content": "I_{f}(A,Q)" + }, + { + "bbox": [ + 131, + 549, + 482, + 667 + ], + "type": "text", + "content": " over sets " + }, + { + "bbox": [ + 131, + 549, + 482, + 667 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 131, + 549, + 482, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 549, + 482, + 667 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 131, + 549, + 482, + 667 + ], + "type": "text", + "content": ", we derive two instances " + }, + { + "bbox": [ + 131, + 549, + 482, + 667 + ], + "type": "inline_equation", + "content": "L_{comb}^{inter}" + }, + { + "bbox": [ + 131, + 549, + 482, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 549, + 482, + 667 + ], + "type": "inline_equation", + "content": "L_{comb}^{intra}" + }, + { + "bbox": [ + 131, + 549, + 482, + 667 + ], + "type": "text", + "content": " objectives in SMILEs. Depending on the choice of " + }, + { + "bbox": [ + 131, + 549, + 482, + 667 + ], + "type": "inline_equation", + "content": "f(A)" + }, + { + "bbox": [ + 131, + 549, + 482, + 667 + ], + "type": "text", + "content": " we define two instances: Facility-Location Mutual Information (SMILE-FLMI) and Graph-Cut Mutual Information (SMILE-GCMI). Inherently, both objectives adopt the cosine similarity metric " + }, + { + "bbox": [ + 131, + 549, + 482, + 667 + ], + "type": "inline_equation", + "content": "S_{ij}(\\theta)" + }, + { + "bbox": [ + 131, + 549, + 482, + 667 + ], + "type": "text", + "content": " as used in SupCon [20] which can be defined as " + }, + { + "bbox": [ + 131, + 549, + 482, + 667 + ], + "type": "inline_equation", + "content": "S_{ij}(\\theta) = \\frac{Z_{comb_i}^{\\mathrm{T}} \\cdot Z_{comb_j}}{||Z_{comb_i}|| \\cdot ||Z_{comb_j}||}" + }, + { + "bbox": [ + 131, + 549, + 482, + 667 + ], + "type": "text", + "content": " to compute similarity between sets in the learning objective. Although the similarity kernel used in SMILEs is computed in a pairwise fashion, objectives defined under " + }, + { + "bbox": [ + 131, + 549, + 482, + 667 + ], + "type": "inline_equation", + "content": "L_{comb}" + }, + { + "bbox": [ + 131, + 549, + 482, + 667 + ], + "type": "text", + "content": " use it to only compute feature interactions" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 246, + 91, + 448, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 91, + 448, + 101 + ], + "spans": [ + { + "bbox": [ + 246, + 91, + 448, + 101 + ], + "type": "text", + "content": "SMILE: Submodular Mutual Information Learner" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 146, + 481, + 200 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 1: Summary of various instantiations of SMILE highlighting the components of the combinatorial objective, " + }, + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "inline_equation", + "content": "L_{comb}^{inter}" + }, + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "inline_equation", + "content": "L_{comb}^{intra}" + }, + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 146, + 481, + 200 + ], + "lines": [ + { + "bbox": [ + 132, + 146, + 481, + 200 + ], + "spans": [ + { + "bbox": [ + 132, + 146, + 481, + 200 + ], + "type": "table", + "html": "
ObjectiveInstances of Lintercomb(θ)Instances of Intracomb(θ)
SMILE-GCMI (ours)∑k∈(Ck∪Cn)2λ∑iAksjA1Sij(θ)∑k∈Ck∪Cn∑iAksjTsetAksij(θ) - λ∑i,jAksij(θ)
∑k∈(Ck∪Cn)2λ∑iAksjA1maxSij(θ) + λ∑iAksjA1maxSij(θ)∑k∈Ck∪Cn∑iTsetAksij(θ)
SMILE-FLMI (ours)
", + "image_path": "8ac960cc7f5c4206e4dd60abece09b20c0dd526c24b83514bfeaaeaeb1e25f7a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 215, + 482, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 215, + 482, + 251 + ], + "spans": [ + { + "bbox": [ + 130, + 215, + 482, + 251 + ], + "type": "text", + "content": "between samples, differing from existing approaches in aggregation of pairwise similarities to compute total information and mutual information over classes in " + }, + { + "bbox": [ + 130, + 215, + 482, + 251 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 130, + 215, + 482, + 251 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 270, + 482, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 381 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 381 + ], + "type": "text", + "content": "SMILE-FLMI based objective is derived from the Facility-Location Mutual Information (FLMI) [22] function, expressed as " + }, + { + "bbox": [ + 130, + 270, + 482, + 381 + ], + "type": "inline_equation", + "content": "I_{f}(Q,A) = \\sum_{i\\in Q}\\max_{j\\in A}S_{ij}(\\theta) + \\lambda \\sum_{i\\in A}\\max_{j\\in Q}S_{ij}(\\theta)" + }, + { + "bbox": [ + 130, + 270, + 482, + 381 + ], + "type": "text", + "content": " and minimizes the maximum similarity (most similar) between sets " + }, + { + "bbox": [ + 130, + 270, + 482, + 381 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 130, + 270, + 482, + 381 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 270, + 482, + 381 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 130, + 270, + 482, + 381 + ], + "type": "text", + "content": ". Given the facility-location (FL) submodular function " + }, + { + "bbox": [ + 130, + 270, + 482, + 381 + ], + "type": "inline_equation", + "content": "f(A,\\theta) = \\sum_{i\\in \\mathcal{T}}\\max_{j\\in A}S_{ij}(\\theta)" + }, + { + "bbox": [ + 130, + 270, + 482, + 381 + ], + "type": "text", + "content": " over the set " + }, + { + "bbox": [ + 130, + 270, + 482, + 381 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 130, + 270, + 482, + 381 + ], + "type": "text", + "content": ", we can derive " + }, + { + "bbox": [ + 130, + 270, + 482, + 381 + ], + "type": "inline_equation", + "content": "L_{comb}^{inter}(\\theta)" + }, + { + "bbox": [ + 130, + 270, + 482, + 381 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 270, + 482, + 381 + ], + "type": "inline_equation", + "content": "L_{comb}^{intra}(\\theta)" + }, + { + "bbox": [ + 130, + 270, + 482, + 381 + ], + "type": "text", + "content": " shown in Eq. (5) as the SMILE-FLMI objective. Note that " + }, + { + "bbox": [ + 130, + 270, + 482, + 381 + ], + "type": "inline_equation", + "content": "L_{comb}^{inter}(\\theta)" + }, + { + "bbox": [ + 130, + 270, + 482, + 381 + ], + "type": "text", + "content": " is applied between object classes in " + }, + { + "bbox": [ + 130, + 270, + 482, + 381 + ], + "type": "inline_equation", + "content": "C_b\\cup C_n" + }, + { + "bbox": [ + 130, + 270, + 482, + 381 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 270, + 482, + 381 + ], + "type": "inline_equation", + "content": "C_n" + }, + { + "bbox": [ + 130, + 270, + 482, + 381 + ], + "type": "text", + "content": " while " + }, + { + "bbox": [ + 130, + 270, + 482, + 381 + ], + "type": "inline_equation", + "content": "L_{comb}^{intra}(\\theta)" + }, + { + "bbox": [ + 130, + 270, + 482, + 381 + ], + "type": "text", + "content": " is applied over all classes in " + }, + { + "bbox": [ + 130, + 270, + 482, + 381 + ], + "type": "inline_equation", + "content": "C_b\\cup C_n" + }, + { + "bbox": [ + 130, + 270, + 482, + 381 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 184, + 389, + 480, + 427 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 389, + 480, + 427 + ], + "spans": [ + { + "bbox": [ + 184, + 389, + 480, + 427 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {c o m b} ^ {i n t e r} (\\theta) = \\sum_ {k \\in \\left(C _ {b} \\cup C _ {n}\\right)} \\sum_ {i \\in A _ {k}} \\max _ {j \\in A _ {l}} S _ {i j} (\\theta) + \\lambda \\sum_ {i \\in A _ {l}} \\max _ {j \\in A _ {k}} S _ {i j} (\\theta), \\\\ l \\in C _ {n}: k \\neq l \\tag {5} \\\\ \\end{array}", + "image_path": "1f10e661dab67158335718f34d6df175c9cfc4d9991d1ee0b7ab689ac5582f1f.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 184, + 427, + 350, + 452 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 427, + 350, + 452 + ], + "spans": [ + { + "bbox": [ + 184, + 427, + 350, + 452 + ], + "type": "interline_equation", + "content": "L _ {c o m b} ^ {i n t r a} (\\theta) = \\sum_ {k \\in C _ {b} \\cup C _ {n}} \\sum_ {i \\in \\mathcal {T} \\backslash A _ {k}} \\max _ {j \\in A _ {k}} S _ {i j} (\\theta)", + "image_path": "183f3f73810c1aaa00c98275b8435091edeff96738a1c490fa73bda8fb0f2f55.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 463, + 482, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 463, + 482, + 561 + ], + "spans": [ + { + "bbox": [ + 130, + 463, + 482, + 561 + ], + "type": "text", + "content": "Minimizing the " + }, + { + "bbox": [ + 130, + 463, + 482, + 561 + ], + "type": "inline_equation", + "content": "L_{comb}^{inter}" + }, + { + "bbox": [ + 130, + 463, + 482, + 561 + ], + "type": "text", + "content": " objective function ensures that the sets " + }, + { + "bbox": [ + 130, + 463, + 482, + 561 + ], + "type": "inline_equation", + "content": "A_{l} \\in C_{n}" + }, + { + "bbox": [ + 130, + 463, + 482, + 561 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 463, + 482, + 561 + ], + "type": "inline_equation", + "content": "A_{k} \\in C_{b} \\cup C_{n}" + }, + { + "bbox": [ + 130, + 463, + 482, + 561 + ], + "type": "text", + "content": " are disjoint by minimizing the similarity between features in " + }, + { + "bbox": [ + 130, + 463, + 482, + 561 + ], + "type": "inline_equation", + "content": "A_{k}" + }, + { + "bbox": [ + 130, + 463, + 482, + 561 + ], + "type": "text", + "content": " and the hardest negative " + }, + { + "bbox": [ + 130, + 463, + 482, + 561 + ], + "type": "inline_equation", + "content": "(\\sum_{i \\in A_{k}} \\max_{j \\in A_{l}} S_{ij}(\\theta)" + }, + { + "bbox": [ + 130, + 463, + 482, + 561 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 130, + 463, + 482, + 561 + ], + "type": "inline_equation", + "content": "k \\in C_{b} \\cup C_{n}" + }, + { + "bbox": [ + 130, + 463, + 482, + 561 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 463, + 482, + 561 + ], + "type": "inline_equation", + "content": "l \\in C_{n})" + }, + { + "bbox": [ + 130, + 463, + 482, + 561 + ], + "type": "text", + "content": " feature vectors in " + }, + { + "bbox": [ + 130, + 463, + 482, + 561 + ], + "type": "inline_equation", + "content": "A_{l}" + }, + { + "bbox": [ + 130, + 463, + 482, + 561 + ], + "type": "text", + "content": ". Further, " + }, + { + "bbox": [ + 130, + 463, + 482, + 561 + ], + "type": "inline_equation", + "content": "L_{comb}^{inter}" + }, + { + "bbox": [ + 130, + 463, + 482, + 561 + ], + "type": "text", + "content": " enforces sufficient separation between the novel classes themselves to promote learning of disjoint feature clusters even with few-shot data overcoming confusion. Additionally, " + }, + { + "bbox": [ + 130, + 463, + 482, + 561 + ], + "type": "inline_equation", + "content": "L_{comb}^{intra}" + }, + { + "bbox": [ + 130, + 463, + 482, + 561 + ], + "type": "text", + "content": " minimizes the total information contained in each set " + }, + { + "bbox": [ + 130, + 463, + 482, + 561 + ], + "type": "inline_equation", + "content": "A_{k} \\in (C_{b} \\cup C_{n})" + }, + { + "bbox": [ + 130, + 463, + 482, + 561 + ], + "type": "text", + "content": ". This objective retains discriminative feature information from each class in " + }, + { + "bbox": [ + 130, + 463, + 482, + 561 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 130, + 463, + 482, + 561 + ], + "type": "text", + "content": " reducing the impact of forgetting." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 579, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 579, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 579, + 482, + 667 + ], + "type": "text", + "content": "SMILE-GCMI based objective described in Eq. (6) minimizes the pairwise similarity of feature vectors between a positive set " + }, + { + "bbox": [ + 130, + 579, + 482, + 667 + ], + "type": "inline_equation", + "content": "A_{k} \\in C_{b} \\cup C_{n}" + }, + { + "bbox": [ + 130, + 579, + 482, + 667 + ], + "type": "text", + "content": " and the sets in " + }, + { + "bbox": [ + 130, + 579, + 482, + 667 + ], + "type": "inline_equation", + "content": "A_{l} \\in C_{n}" + }, + { + "bbox": [ + 130, + 579, + 482, + 667 + ], + "type": "text", + "content": " while maximizing the similarity between features in each set " + }, + { + "bbox": [ + 130, + 579, + 482, + 667 + ], + "type": "inline_equation", + "content": "A_{k} \\in C_{b} \\cup C_{n}" + }, + { + "bbox": [ + 130, + 579, + 482, + 667 + ], + "type": "text", + "content": ". Given two sets " + }, + { + "bbox": [ + 130, + 579, + 482, + 667 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 130, + 579, + 482, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 579, + 482, + 667 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 130, + 579, + 482, + 667 + ], + "type": "text", + "content": ", [22] defines the Graph-Cut SMI to be " + }, + { + "bbox": [ + 130, + 579, + 482, + 667 + ], + "type": "inline_equation", + "content": "I_{f}(Q, A) = 2\\lambda \\sum_{i \\in Q} \\sum_{j \\in A} S_{ij}(\\theta)" + }, + { + "bbox": [ + 130, + 579, + 482, + 667 + ], + "type": "text", + "content": ", where the Graph-Cut function over a set " + }, + { + "bbox": [ + 130, + 579, + 482, + 667 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 130, + 579, + 482, + 667 + ], + "type": "text", + "content": " is given by " + }, + { + "bbox": [ + 130, + 579, + 482, + 667 + ], + "type": "inline_equation", + "content": "f(A, \\theta) = \\sum_{i \\in A} \\sum_{j \\in T \\setminus A_{k}} S_{ij}(\\theta) - \\lambda \\sum_{i,j \\in A} S_{ij}(\\theta)" + }, + { + "bbox": [ + 130, + 579, + 482, + 667 + ], + "type": "text", + "content": ". Given the Graph-Cut and the Graph-Cut SMI functions, we derive " + }, + { + "bbox": [ + 130, + 579, + 482, + 667 + ], + "type": "inline_equation", + "content": "L_{comb}^{inter}(\\theta)" + }, + { + "bbox": [ + 130, + 579, + 482, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 579, + 482, + 667 + ], + "type": "inline_equation", + "content": "L_{comb}^{intra}(\\theta)" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "type": "text", + "content": "Majee et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "type": "text", + "content": "shown in Eq. (6) as the SMILE-GCMI objective. Similar to SMILE-FLMI, the " + }, + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "L_{comb}^{inter}(\\theta)" + }, + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "type": "text", + "content": " is applied between object classes in " + }, + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "C_b \\cup C_n" + }, + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "C_n" + }, + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "type": "text", + "content": " while " + }, + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "L_{comb}^{intra}(\\theta)" + }, + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "type": "text", + "content": " is applied over all classes in " + }, + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "C_b \\cup C_n" + }, + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 188, + 159, + 480, + 196 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 159, + 480, + 196 + ], + "spans": [ + { + "bbox": [ + 188, + 159, + 480, + 196 + ], + "type": "interline_equation", + "content": "L _ {c o m b} ^ {i n t e r} (\\theta) = \\sum_ {\\substack {k \\in (C _ {b} \\cup C _ {n}) \\\\ l \\in C _ {n}: k \\neq l}} 2 \\lambda \\sum_ {i \\in A _ {k}} \\sum_ {j \\in A _ {l}} S _ {i j} (\\theta), \\tag{6}", + "image_path": "724eb4db5b0032852a360ec3def4ebd0b3fe19b20e09cf7f91f199f5674bce5c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 188, + 195, + 426, + 221 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 195, + 426, + 221 + ], + "spans": [ + { + "bbox": [ + 188, + 195, + 426, + 221 + ], + "type": "interline_equation", + "content": "L _ {c o m b} ^ {i n t r a} (\\theta) = \\sum_ {k \\in C _ {b} \\cup C _ {n}} \\sum_ {i \\in A _ {k}} \\sum_ {j \\in \\mathcal {T} \\backslash A _ {k}} S _ {i j} (\\theta) - \\lambda \\sum_ {i, j \\in A _ {k}} S _ {i j} (\\theta)", + "image_path": "905d910c366a0b6ac47628d0b3ae2f56c1356663c7b404e00c0c1991089832c7.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 227, + 479, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 227, + 479, + 321 + ], + "spans": [ + { + "bbox": [ + 130, + 227, + 479, + 321 + ], + "type": "text", + "content": "Although objectives in SMILE-FLMI and SMILE-GCMI are tasked with similar functions, the " + }, + { + "bbox": [ + 130, + 227, + 479, + 321 + ], + "type": "inline_equation", + "content": "L_{comb}^{inter}" + }, + { + "bbox": [ + 130, + 227, + 479, + 321 + ], + "type": "text", + "content": " in SMILE-GCMI minimizes the pairwise similarity between sets in " + }, + { + "bbox": [ + 130, + 227, + 479, + 321 + ], + "type": "inline_equation", + "content": "C_b \\cup C_n" + }, + { + "bbox": [ + 130, + 227, + 479, + 321 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 227, + 479, + 321 + ], + "type": "inline_equation", + "content": "C_n" + }, + { + "bbox": [ + 130, + 227, + 479, + 321 + ], + "type": "text", + "content": " rather than the most similar set in SMILE-FLMI. Further, the " + }, + { + "bbox": [ + 130, + 227, + 479, + 321 + ], + "type": "inline_equation", + "content": "L_{comb}^{intra}" + }, + { + "bbox": [ + 130, + 227, + 479, + 321 + ], + "type": "text", + "content": " in SMILE-GCMI scales linearly with size of " + }, + { + "bbox": [ + 130, + 227, + 479, + 321 + ], + "type": "inline_equation", + "content": "A_k" + }, + { + "bbox": [ + 130, + 227, + 479, + 321 + ], + "type": "text", + "content": " as described in [30]. This does not allow the model to substantially improve performance on learning discriminative feature representations for classes in both " + }, + { + "bbox": [ + 130, + 227, + 479, + 321 + ], + "type": "inline_equation", + "content": "C_b" + }, + { + "bbox": [ + 130, + 227, + 479, + 321 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 227, + 479, + 321 + ], + "type": "inline_equation", + "content": "C_n" + }, + { + "bbox": [ + 130, + 227, + 479, + 321 + ], + "type": "text", + "content": " as the " + }, + { + "bbox": [ + 130, + 227, + 479, + 321 + ], + "type": "inline_equation", + "content": "|A_k| = K" + }, + { + "bbox": [ + 130, + 227, + 479, + 321 + ], + "type": "text", + "content": " (number of shots) thus failing to outperform the model trained using SMILE-FLMI." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 323, + 479, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 323, + 479, + 358 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 479, + 358 + ], + "type": "text", + "content": "The detailed derivations of the aforementioned instances are included in the Supplementary material. Our experiments in Sec. 4.4 elucidates the fact that SMILE-FLMI is a better choice to overcome forgetting and confusion in FSOD." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 375, + 230, + 389 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 375, + 230, + 389 + ], + "spans": [ + { + "bbox": [ + 132, + 375, + 230, + 389 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 398, + 479, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 398, + 479, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 398, + 479, + 434 + ], + "type": "text", + "content": "We evaluate models in SMILE by adopting standard evaluation criterion in FSOD [16, 37] and report the Mean Average Precision (mAP) at " + }, + { + "bbox": [ + 130, + 398, + 479, + 434 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 398, + 479, + 434 + ], + "type": "text", + "content": " Intersection Over Union (IoU) for all our experiments." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 449, + 261, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 449, + 261, + 462 + ], + "spans": [ + { + "bbox": [ + 132, + 449, + 261, + 462 + ], + "type": "text", + "content": "4.1 Experimental Setup" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 467, + 479, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 467, + 479, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 467, + 479, + 491 + ], + "type": "text", + "content": "Datasets We evaluate our proposed SMILE approach on two few-shot object detection datasets - and PASCAL-VOC [5] and MS-COCO [26] datasets." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 500, + 480, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 500, + 480, + 583 + ], + "spans": [ + { + "bbox": [ + 130, + 500, + 480, + 583 + ], + "type": "text", + "content": "PASCAL-VOC [5] dataset consists of 20 classes, out of which 15 are considered as base and 5 as novel classes. The novel classes are chosen at random giving rise to three data splits namely, split-1 (bird, bus, cow, motorbike, sofa), split-2 (aeroplane, bottle, cow, horse, sofa) and split-3 (boat, cat, motorbike, sheep, sofa). Following previous works [16], we use the combined VOC " + }, + { + "bbox": [ + 130, + 500, + 480, + 583 + ], + "type": "inline_equation", + "content": "07 + 12" + }, + { + "bbox": [ + 130, + 500, + 480, + 583 + ], + "type": "text", + "content": " datasets for training and evaluate our models on the complete validation set of VOC 2007 for 1, 5, and 10 shot settings." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 594, + 479, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 479, + 664 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 479, + 664 + ], + "type": "text", + "content": "MS-COCO [26] dataset consists of 80 classes, out of which 60 are considered as base and 20 as novel classes. Following existing approaches in FSOD [41] we randomly select 5k samples from " + }, + { + "bbox": [ + 130, + 594, + 479, + 664 + ], + "type": "inline_equation", + "content": "(D_{base} \\cup D_{novel})" + }, + { + "bbox": [ + 130, + 594, + 479, + 664 + ], + "type": "text", + "content": " to use as the validation set while the remaining samples are used to generate random 10 and 30-shot splits for training of the MS-COCO 2014 dataset. The key difference between VOC and COCO is the large intra-class variance and class-imbalance in COCO." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 246, + 91, + 448, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 91, + 448, + 101 + ], + "spans": [ + { + "bbox": [ + 246, + 91, + 448, + 101 + ], + "type": "text", + "content": "SMILE: Submodular Mutual Information Learner" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 179, + 483, + 370 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 2: Quantitative analysis on PASCAL-VOC dataset: Few-shot object detection performance (" + }, + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "mAP_{novel}" + }, + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "type": "text", + "content": ") on novel class splits of PASCAL-VOC dataset. We tabulate results for " + }, + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "K = 1, 5, 10" + }, + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "type": "text", + "content": " shots from various SoTA techniques in FSOD. * indicates that the results are averaged over 10 random seeds. † indicates a meta-learning strategy (N-way, K-shot training)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 179, + 483, + 370 + ], + "lines": [ + { + "bbox": [ + 134, + 179, + 483, + 370 + ], + "spans": [ + { + "bbox": [ + 134, + 179, + 483, + 370 + ], + "type": "table", + "html": "
MethodLearner TypeBackboneSplit 1Split 2Split 3
K=151015101510
†Meta-RCNN [41]MetaFRCN-R10119.945.751.510.434.845.414.341.248.1
†Meta-Reweight [16]MetaYOLO V214.833.947.215.730.140.521.342.845.9
†MetaDet [38]MetaFRCN-R10118.936.849.621.831.743.020.643.944.1
†Add-Info [40]MetaFRCN-R10124.249.157.421.637.045.721.243.849.6
†CME [24]MetaYOLO V217.844.847.512.733.740.015.744.948.8
PNPDet [42]MetricDLA-3418.2-41.016.6-36.418.9-36.2
FsDet w/ FC [37]MetricFRCN-R10136.855.757.018.235.539.027.748.750.2
FsDet w/ cos [37]MetricFRCN-R10139.855.756.023.535.139.130.849.549.8
Retentive-RCNN [9]MetricFRCN-R10140.153.756.121.737.040.330.249.750.1
FSCE [34]MetricFRCN-R10141.057.457.827.344.449.840.153.257.7
\\( FSCE + SMILE \\) (ours)Comb.FRCN-R10141.257.961.129.244.650.541.355.659.0
AGCM [1]MetricFRCN-R10140.358.559.927.549.350.642.154.258.2
\\( AGCM + SMILE \\) (ours)Comb.FRCN-R10140.959.762.031.949.552.342.656.461.4
DiGeo [28]MetricFRCN-R10136.054.160.920.742.847.127.547.352.9
DiGeo + SMILE(ours)Comb.FRCN-R10136.156.662.326.544.147.333.151.956.4
imTED [27]MetricViT-B31.971.977.022.752.257.712.669.672.8
imTED + PDC [23]MetricViT-B36.673.177.115.551.856.018.967.972.8
PDC + SMILE (ours)Comb.ViT-B36.675.277.927.152.758.315.170.074.7
", + "image_path": "4ce735bb899e250f1d59da34769396814a4bb9dee644fad0b9c1daf73722aa9a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 391, + 482, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 391, + 482, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 482, + 536 + ], + "type": "text", + "content": "Implementation Details The SMILE framework adopts a architecture agnostic approach and adopt several backbones including Faster-RCNN [31] and ViT [27]. For VOC, the input batch size to the network is set to 16 and 2 in the base training and few-shot adaptation stages for Faster-RCNN and ViT based approaches. The input resolution is set to " + }, + { + "bbox": [ + 130, + 391, + 482, + 536 + ], + "type": "inline_equation", + "content": "764 \\times 1333" + }, + { + "bbox": [ + 130, + 391, + 482, + 536 + ], + "type": "text", + "content": " pixels for data splits in COCO, while it is set to " + }, + { + "bbox": [ + 130, + 391, + 482, + 536 + ], + "type": "inline_equation", + "content": "800 \\times 600" + }, + { + "bbox": [ + 130, + 391, + 482, + 536 + ], + "type": "text", + "content": " pixels for PASCAL-VOC. The hyper-parameters used in the formulation of SMILE, namely " + }, + { + "bbox": [ + 130, + 391, + 482, + 536 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 130, + 391, + 482, + 536 + ], + "type": "text", + "content": " and similarity kernel " + }, + { + "bbox": [ + 130, + 391, + 482, + 536 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 391, + 482, + 536 + ], + "type": "text", + "content": ", are chosen through ablation experiments described in Sec. 4.4. Results from existing methods are a reproduction of the algorithm from publicly available codebases. All our experiments are performed on 4 NVIDIA GTX 1080 Ti GPUs with additional details in the supplementary material and code released at https://github.com/amajee11us/SMILE-FSOD.git." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 551, + 386, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 551, + 386, + 562 + ], + "spans": [ + { + "bbox": [ + 132, + 551, + 386, + 562 + ], + "type": "text", + "content": "4.2 Results on Few-Shot PASCAL VOC Dataset" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": "Table 2 records the results obtained from our SMILE framework on novel splits of the PASCAL-VOC dataset and contrasts it against SoTA FSOD techniques. We adopt four SoTA approaches FSCE [34], AGCM [1], DiGeo [28] and imTED [27], covering several backbone architectures Faster-RCNN + FPN (FSCE, AGCM and DiGeo) alongside ViT (imTED, PDC) and introducing SMILE (" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "M + SMILE" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": ") approach into existing architectures " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": ". For Faster-RCNN based architectures (FSCE) we show a maximum of " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "5.7\\%" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": " (3.3 map points) improvement while for FPN based architectures (AGCM and DiGeo) we show a " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "3.5\\%" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": " (2.1 map points" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "type": "text", + "content": "Majee et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 296 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 296 + ], + "type": "text", + "content": "for AGCM+SMILE) improvement. It is interesting to note that unlike FSCE and AGCM, DiGeo uses abundant samples from " + }, + { + "bbox": [ + 130, + 115, + 482, + 296 + ], + "type": "inline_equation", + "content": "C_b" + }, + { + "bbox": [ + 130, + 115, + 482, + 296 + ], + "type": "text", + "content": " alongside few-shot samples in " + }, + { + "bbox": [ + 130, + 115, + 482, + 296 + ], + "type": "inline_equation", + "content": "C_n" + }, + { + "bbox": [ + 130, + 115, + 482, + 296 + ], + "type": "text", + "content": " (with upsampling) during finetuning introducing a large inter-class bias. SMILE outperforms DiGeo by up to 2.3 " + }, + { + "bbox": [ + 130, + 115, + 482, + 296 + ], + "type": "inline_equation", + "content": "mAP" + }, + { + "bbox": [ + 130, + 115, + 482, + 296 + ], + "type": "text", + "content": " points (split 2, 10-shot) showing the resilience of SMILE towards imbalance, thus overcoming confusion in FSOD. Additionally, for recently introduced transformer based architectures (imTED + PDC [23]) SMILE outperforms the existing SoTA with a maximum improvement of 4.9 " + }, + { + "bbox": [ + 130, + 115, + 482, + 296 + ], + "type": "inline_equation", + "content": "mAP" + }, + { + "bbox": [ + 130, + 115, + 482, + 296 + ], + "type": "text", + "content": " points (split 2, 5-shot) thus establishing SMILE as the SoTA on few-shot splits of VOC. Note, that the choice of objective functions " + }, + { + "bbox": [ + 130, + 115, + 482, + 296 + ], + "type": "inline_equation", + "content": "L_{comb}^{inter}" + }, + { + "bbox": [ + 130, + 115, + 482, + 296 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 115, + 482, + 296 + ], + "type": "inline_equation", + "content": "L_{comb}^{intra}" + }, + { + "bbox": [ + 130, + 115, + 482, + 296 + ], + "type": "text", + "content": " for this experiment has been determined to be SMILE+FLMI through an ablation on the different instances in SMILEs as described in Sec. 4.4. Finally, Fig. 2(b) shows that introduction of SMILE framework to existing SoTA approaches leads to rapid convergence on the novel classes up to 2x over existing SoTA. This is significant for mission critical tasks like autonomous driving where the model is required to rapidly learn novel objects to reduce turn-around time." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 316, + 365, + 328 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 316, + 365, + 328 + ], + "spans": [ + { + "bbox": [ + 132, + 316, + 365, + 328 + ], + "type": "text", + "content": "4.3 Results on Few-Shot MS-COCO Dataset" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 338, + 482, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 338, + 482, + 469 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 482, + 469 + ], + "type": "text", + "content": "Similar to the results in PASCAL VOC we demonstrate the results of our SMILE framework on MS-COCO dataset. In contrast to VOC, COCO presents an extremely imbalanced setting with a long-tail distribution within " + }, + { + "bbox": [ + 130, + 338, + 482, + 469 + ], + "type": "inline_equation", + "content": "D_{base}" + }, + { + "bbox": [ + 130, + 338, + 482, + 469 + ], + "type": "text", + "content": " itself making it really hard for FSOD approaches to achieve SoTA through primitive objective functions. Following the ablation experiments in Sec. 4.4 we adopt the SMILE+FLMI objective (best performing) to conduct the experiments on 20 few-shot classes of MS-COCO dataset. We show that SMILE generalizes existing SoTA approach (imTED + PDC) for COCO dataset by " + }, + { + "bbox": [ + 130, + 338, + 482, + 469 + ], + "type": "inline_equation", + "content": "5.4\\%" + }, + { + "bbox": [ + 130, + 338, + 482, + 469 + ], + "type": "text", + "content": " (2.6 mAP points, 30-shot setting). This further establishes the generalizability of our approach over varying data distributions (VOC and COCO) while achieving SoTA in FSOD tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 491, + 238, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 238, + 503 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 238, + 503 + ], + "type": "text", + "content": "4.4 Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 513, + 481, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 513, + 481, + 550 + ], + "spans": [ + { + "bbox": [ + 130, + 513, + 481, + 550 + ], + "type": "text", + "content": "We conduct ablation experiments on the 10-shot split of VOC (split 1) with hyper-parameters " + }, + { + "bbox": [ + 130, + 513, + 481, + 550 + ], + "type": "inline_equation", + "content": "\\eta = 0.5" + }, + { + "bbox": [ + 130, + 513, + 481, + 550 + ], + "type": "text", + "content": ", cosine similarity metric and " + }, + { + "bbox": [ + 130, + 513, + 481, + 550 + ], + "type": "inline_equation", + "content": "\\lambda = 1.0" + }, + { + "bbox": [ + 130, + 513, + 481, + 550 + ], + "type": "text", + "content": ". Ablations for hyper-parameters are detailed in the supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": "Components of SMILE Instantiations in SMILEs consists of two main components - " + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "L_{comb}^{inter}" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "L_{comb}^{intra}" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": ". We consider three baselines FSCE, AGCM and DiGeo which follow the Faster-RCNN/FPN backbone for this experiment. First, we introduce " + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "L_{comb}^{intra}" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": " by adopting the FL based objective as determined through ablation experiments below. This objective models intra-class variance and ensures reduction in intra-class variance characterized by boost in base class performance. Secondly, following the SMILE-FLMI formulation in Eq. (5) we introduce " + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "L_{comb}^{inter}" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": " during the few-shot adaptation stage. Applying this objective minimizes" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 246, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 246, + 91, + 447, + 101 + ], + "type": "text", + "content": "SMILE: Submodular Mutual Information Learner" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 179, + 157, + 434, + 301 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 3: Performance of SMILE on MS COCO dataset : Our SMILE objectives demonstrate better generalizability while outperforming SoTA FSOD approaches on novel class performance " + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "mAP_{50}" + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": " (novel)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 179, + 157, + 434, + 301 + ], + "lines": [ + { + "bbox": [ + 179, + 157, + 434, + 301 + ], + "spans": [ + { + "bbox": [ + 179, + 157, + 434, + 301 + ], + "type": "table", + "html": "
Method\\( {mAP} \\)\\( {mAP}_{50} \\)\\( {mAP}_{75} \\)\\( {mAP} \\)\\( {mAP}_{50} \\)\\( {mAP}_{75} \\)
10-shot30-shot
Meta-Reweight [16]5.612.34.69.119.07.6
Meta-RCNN [41]8.719.16.612.425.310.8
TFA w/cos [37]10.0-9.313.7-13.4
Add-Info [40]12.527.39.814.730.612.2
MPSR [39]9.817.99.714.125.414.2
FSCE [34]11.9-10.516.4-16.2
FADI [2]12.222.711.916.129.115.8
CME [24]15.124.616.416.2--
FCT [14]17.130.217.021.435.522.1
imTED-B [27]22.536.623.730.247.432.5
imTED-B+PDC [23]23.438.124.530.847.333.5
PDC + SMILE (ours)25.840.126.131.049.933.6
", + "image_path": "ce88f033d4a6c31997410fe43e549c3acb88a4ec1950a9c6a24b6b486d6875fa.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 190, + 334, + 421, + 489 + ], + "blocks": [ + { + "bbox": [ + 151, + 313, + 461, + 325 + ], + "lines": [ + { + "bbox": [ + 151, + 313, + 461, + 325 + ], + "spans": [ + { + "bbox": [ + 151, + 313, + 461, + 325 + ], + "type": "text", + "content": "Table 4: Ablation on various components of the proposed SMILE approach." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 190, + 334, + 421, + 489 + ], + "lines": [ + { + "bbox": [ + 190, + 334, + 421, + 489 + ], + "spans": [ + { + "bbox": [ + 190, + 334, + 421, + 489 + ], + "type": "table", + "html": "
MethodBaselinef(Ai)(Lintra)If(Ai,Aj)(Lintercomb)mAPbasemAPnovel
FsDet w/ cos---23.639.8
FSCE✓86.157.8
✓✓89.660.1
✓✓88.361.0
✓✓✓89.861.1
AGCM✓87.658.0
✓✓88.661.3
✓✓88.961.8
✓✓✓89.361.8
DiGeo✓90.560.9
✓✓92.361.7
✓✓91.462.0
✓✓✓92.662.3
", + "image_path": "ed3a0f7f188f7fb22316323b1e1b8345208d3174b963fbb4add03f20bfbb8be9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 513, + 482, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 513, + 482, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 513, + 482, + 609 + ], + "type": "text", + "content": "the inter-class bias between " + }, + { + "bbox": [ + 130, + 513, + 482, + 609 + ], + "type": "inline_equation", + "content": "C_b \\cup C_n" + }, + { + "bbox": [ + 130, + 513, + 482, + 609 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 513, + 482, + 609 + ], + "type": "inline_equation", + "content": "C_n" + }, + { + "bbox": [ + 130, + 513, + 482, + 609 + ], + "type": "text", + "content": ", thus improving novel class performance significantly. Nevertheless, we see a slight drop in base class performance due to forgetting prevalent in FSOD tasks. Finally, we combine both instantiations in SMILE into one single objective as in Eq. (3) with " + }, + { + "bbox": [ + 130, + 513, + 482, + 609 + ], + "type": "inline_equation", + "content": "\\eta = 0.5" + }, + { + "bbox": [ + 130, + 513, + 482, + 609 + ], + "type": "text", + "content": " and show that SMILE improves both base and novel class performance emerging as the best choice for FSOD. To demonstrate generalization, we perform this experiment on several SoTA approaches as baseline and show that the results discussed in aforementioned section hold. We summarize all the results in Tab. 4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": "Choice of Combinatorial functions in " + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "inline_equation", + "content": "L_{comb}" + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": " SMILE introduces several instances of " + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "inline_equation", + "content": "L_{comb}^{intra}" + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "inline_equation", + "content": "L_{comb}^{inter}" + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": ". To clearly understand the contributions of each of these instances we conduct experiments tabulated in Tab. 5 and determine the" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "type": "text", + "content": "Majee et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 219, + 146, + 394, + 213 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 5: Ablation on the choice of Submodular Information function " + }, + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "inline_equation", + "content": "I_{f}" + }, + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": " and Submodular Information function " + }, + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "inline_equation", + "content": "L_{comb}" + }, + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": " in SMILE." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 219, + 146, + 394, + 213 + ], + "lines": [ + { + "bbox": [ + 219, + 146, + 394, + 213 + ], + "spans": [ + { + "bbox": [ + 219, + 146, + 394, + 213 + ], + "type": "table", + "html": "
Modelf(A,θ)If(Ai,Aj)mAPbasemAPnovel
DiGeo [28]--87.957.4
GCGCMI92.660.9
FLFLMI93.162.3
FSCE [34]--86.157.8
GCGCMI87.460.3
FLFLMI89.861.1
", + "image_path": "14912a88ec6bf36a2df0f19fbfc3f9aee23db7ba163ba927c86509d5a96a87e9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 236, + 482, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 482, + 356 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 482, + 356 + ], + "type": "text", + "content": "best performing formulation which generalizes to existing FSOD architectures. Unlike other ablation experiments in Sec. 4.4, we conduct our experiments on DiGeo which introduces an extremely imbalanced scenario by using abundant samples in " + }, + { + "bbox": [ + 130, + 236, + 482, + 356 + ], + "type": "inline_equation", + "content": "D_{base}" + }, + { + "bbox": [ + 130, + 236, + 482, + 356 + ], + "type": "text", + "content": ". We conclude that SMILE-FLMI which considers the Facility-Location based objective as " + }, + { + "bbox": [ + 130, + 236, + 482, + 356 + ], + "type": "inline_equation", + "content": "L_{comb}^{intra}" + }, + { + "bbox": [ + 130, + 236, + 482, + 356 + ], + "type": "text", + "content": " and Facility-Location Mutual Information based objective as " + }, + { + "bbox": [ + 130, + 236, + 482, + 356 + ], + "type": "inline_equation", + "content": "L_{comb}^{inter}" + }, + { + "bbox": [ + 130, + 236, + 482, + 356 + ], + "type": "text", + "content": " as the best performing instantiation. This result follows the formulation in Eq. (5) where FL naturally models intra-class compactness in class-imbalanced settings [30] while FLMI penalizes the classes in " + }, + { + "bbox": [ + 130, + 236, + 482, + 356 + ], + "type": "inline_equation", + "content": "C_b" + }, + { + "bbox": [ + 130, + 236, + 482, + 356 + ], + "type": "text", + "content": " to learn overlapping feature representations with the classes in " + }, + { + "bbox": [ + 130, + 236, + 482, + 356 + ], + "type": "inline_equation", + "content": "C_n" + }, + { + "bbox": [ + 130, + 236, + 482, + 356 + ], + "type": "text", + "content": ". We use SMILE-FLMI for all benchmark experiments in Sec. 4.2 and Sec. 4.3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 373, + 482, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 373, + 482, + 589 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 482, + 589 + ], + "type": "text", + "content": "Robustness to Catastrophic Forgetting One of the most significant challenges in FSOD is the elimination of catastrophic forgetting which manifests as the degradation in the performance of classes in " + }, + { + "bbox": [ + 130, + 373, + 482, + 589 + ], + "type": "inline_equation", + "content": "C_b" + }, + { + "bbox": [ + 130, + 373, + 482, + 589 + ], + "type": "text", + "content": " while learning classes in " + }, + { + "bbox": [ + 130, + 373, + 482, + 589 + ], + "type": "inline_equation", + "content": "C_n" + }, + { + "bbox": [ + 130, + 373, + 482, + 589 + ], + "type": "text", + "content": ". This primarily occurs due to the lack of discriminative feature representations from instances in " + }, + { + "bbox": [ + 130, + 373, + 482, + 589 + ], + "type": "inline_equation", + "content": "D_{base}" + }, + { + "bbox": [ + 130, + 373, + 482, + 589 + ], + "type": "text", + "content": " during the few-shot adaptation (stage 2) stage. We plot the change in base class performance as the training progresses in existing SoTA methods AGCM and DiGeo against number of training iterations in Fig. 2(a). At first, we contrast the change in base class performance " + }, + { + "bbox": [ + 130, + 373, + 482, + 589 + ], + "type": "inline_equation", + "content": "mAP_{base}" + }, + { + "bbox": [ + 130, + 373, + 482, + 589 + ], + "type": "text", + "content": " between AGCM and AGCM+SMILE and observe that AGCM overfits on the few-shot samples in " + }, + { + "bbox": [ + 130, + 373, + 482, + 589 + ], + "type": "inline_equation", + "content": "D_{base}" + }, + { + "bbox": [ + 130, + 373, + 482, + 589 + ], + "type": "text", + "content": " reducing the performance on " + }, + { + "bbox": [ + 130, + 373, + 482, + 589 + ], + "type": "inline_equation", + "content": "C_{base}" + }, + { + "bbox": [ + 130, + 373, + 482, + 589 + ], + "type": "text", + "content": " as the training progresses. AGCM + SMILE on the other hand better retains the performance on base classes with " + }, + { + "bbox": [ + 130, + 373, + 482, + 589 + ], + "type": "inline_equation", + "content": "\\sim 3.5\\%" + }, + { + "bbox": [ + 130, + 373, + 482, + 589 + ], + "type": "text", + "content": " better retention in base class performance. Interestingly, DiGeo is able to retain most of the base class performance with a very small degradation over the roofline (a model trained with only the base classes until convergence). Our DiGeo+SMILE approach outperforms Digo by demonstrating base class performance even higher than the roofline establishing the supremacy of " + }, + { + "bbox": [ + 130, + 373, + 482, + 589 + ], + "type": "inline_equation", + "content": "L_{comb}" + }, + { + "bbox": [ + 130, + 373, + 482, + 589 + ], + "type": "text", + "content": " in overcoming inter-class bias and intra-class variance resulting in robustness against catastrophic forgetting." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Overcoming Class Confusion Figure 4 highlights the supremacy of the proposed SMILE framework in mitigating class confusion through confusion matrix plots. We compare the confusion between classes in " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "C_b \\cup C_n" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " of SoTA approaches AGCM and DiGeo before and after introduction of combinatorial objectives in SMILE. Although both approaches use K-shot examples for classes in " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "C_n" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": ", Di" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 246, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 246, + 91, + 447, + 101 + ], + "type": "text", + "content": "SMILE: Submodular Mutual Information Learner" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 116, + 222, + 210 + ], + "blocks": [ + { + "bbox": [ + 132, + 116, + 222, + 210 + ], + "lines": [ + { + "bbox": [ + 132, + 116, + 222, + 210 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 222, + 210 + ], + "type": "image", + "image_path": "21f3a143182809eff3cca7577c7548522161ae5987ee05923df949cf9e9624f3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 219, + 482, + 253 + ], + "lines": [ + { + "bbox": [ + 130, + 219, + 482, + 253 + ], + "spans": [ + { + "bbox": [ + 130, + 219, + 482, + 253 + ], + "type": "text", + "content": "Fig. 4: Ablation on Overcoming Class Confusion in SMILE. (a,b) SMILE demonstrates " + }, + { + "bbox": [ + 130, + 219, + 482, + 253 + ], + "type": "inline_equation", + "content": "11\\%" + }, + { + "bbox": [ + 130, + 219, + 482, + 253 + ], + "type": "text", + "content": " lower confusion over AGCM and (c,d) " + }, + { + "bbox": [ + 130, + 219, + 482, + 253 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 130, + 219, + 482, + 253 + ], + "type": "text", + "content": " lower confusion over DiGeo. Only significant numbers are highlighted. Best viewed in " + }, + { + "bbox": [ + 130, + 219, + 482, + 253 + ], + "type": "inline_equation", + "content": "200\\%" + }, + { + "bbox": [ + 130, + 219, + 482, + 253 + ], + "type": "text", + "content": " zoom." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 223, + 116, + 303, + 210 + ], + "blocks": [ + { + "bbox": [ + 223, + 116, + 303, + 210 + ], + "lines": [ + { + "bbox": [ + 223, + 116, + 303, + 210 + ], + "spans": [ + { + "bbox": [ + 223, + 116, + 303, + 210 + ], + "type": "image", + "image_path": "e66a23625cc084bf94517d28a182a2cf8e48f318423e47579ad4849ec8db6bab.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 303, + 118, + 384, + 210 + ], + "blocks": [ + { + "bbox": [ + 303, + 118, + 384, + 210 + ], + "lines": [ + { + "bbox": [ + 303, + 118, + 384, + 210 + ], + "spans": [ + { + "bbox": [ + 303, + 118, + 384, + 210 + ], + "type": "image", + "image_path": "c3c2890c81276f9bb2dd8825bc0aa0e27b5a08a24c0d73ae3e10c69de2be3dcf.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 384, + 117, + 477, + 211 + ], + "blocks": [ + { + "bbox": [ + 384, + 117, + 477, + 211 + ], + "lines": [ + { + "bbox": [ + 384, + 117, + 477, + 211 + ], + "spans": [ + { + "bbox": [ + 384, + 117, + 477, + 211 + ], + "type": "image", + "image_path": "b5352161d55bd070afbe86f70f9bf6e94be824d53405d25622a51c09e25b6497.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 266, + 482, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 266, + 482, + 410 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 482, + 410 + ], + "type": "text", + "content": "Geo differs from AGCM by adopting an upsampling strategy which allows the utilization of abundant examples in " + }, + { + "bbox": [ + 130, + 266, + 482, + 410 + ], + "type": "inline_equation", + "content": "C_b" + }, + { + "bbox": [ + 130, + 266, + 482, + 410 + ], + "type": "text", + "content": " while upsampling the instances in " + }, + { + "bbox": [ + 130, + 266, + 482, + 410 + ], + "type": "inline_equation", + "content": "C_n" + }, + { + "bbox": [ + 130, + 266, + 482, + 410 + ], + "type": "text", + "content": ". This injects different degrees of inter-class biases for models trained by adopting AGCM and DiGeo which has been demonstrated as the primary reason for confusion in previous work [29]. At first, we observe from Fig. 4 that by adopting the upsampling based strategy, DiGeo achieves very low confusion between already learnt base classes, leading to significantly lower confusion (5% among " + }, + { + "bbox": [ + 130, + 266, + 482, + 410 + ], + "type": "inline_equation", + "content": "C_b" + }, + { + "bbox": [ + 130, + 266, + 482, + 410 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 266, + 482, + 410 + ], + "type": "inline_equation", + "content": "C_n" + }, + { + "bbox": [ + 130, + 266, + 482, + 410 + ], + "type": "text", + "content": "). Further, confusion matrix plots in Fig. 4 show that AGCM+SMILE demonstrates 11% lower confusion than AGCM and DiGeo+SMILE shows 4% lower confusion and DiGeo. This proves the efficacy of combinatorial objectives (" + }, + { + "bbox": [ + 130, + 266, + 482, + 410 + ], + "type": "inline_equation", + "content": "L_{comb}^{inter}" + }, + { + "bbox": [ + 130, + 266, + 482, + 410 + ], + "type": "text", + "content": ") in mitigating inter-class bias, thereby reducing confusion between classes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 445, + 220, + 457 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 445, + 220, + 457 + ], + "spans": [ + { + "bbox": [ + 132, + 445, + 220, + 457 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": "In this work, we have presented a novel approach to Few-Shot Object Detection (FSOD) by introducing a combinatorial viewpoint through the SMILE framework. By leveraging the properties of set-based combinatorial functions, SMILE aims to address the challenges of class confusion and catastrophic forgetting, which are prevalent in FSOD tasks. Our approach incorporates Submodular Mutual Information (SMI) and Submodular Information Measures (SIM) to penalize overlapping features between base and novel classes and to ensure the formation of compact feature clusters, respectively. The experimental results on PASCAL-VOC and MS-COCO benchmarks demonstrate the effectiveness of SMILE, showing significant improvements in novel class performance, faster convergence, and a reduction in class confusion and catastrophic forgetting. Overall, SMILE offers a promising direction for advancing the state-of-the-art in FSOD by providing a generalized framework that is adaptable to various underlying architectures and capable of handling the complexities associated with few-shot learning in object detection." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "type": "text", + "content": "Majee et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "content": "We gratefully thank anonymous reviewers for their valuable comments. This work is supported by the National Science Foundation under Grant Numbers IIS-2106937, a gift from Google Research, an Amazon Research Award, and the Adobe Data Science Research award. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the National Science Foundation, Google or Adobe." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 243, + 197, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 243, + 197, + 255 + ], + "spans": [ + { + "bbox": [ + 133, + 243, + 197, + 255 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 269, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 269, + 480, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 269, + 480, + 312 + ], + "spans": [ + { + "bbox": [ + 138, + 269, + 480, + 312 + ], + "type": "text", + "content": "1. Agarwal, A., Majee, A., Subramanian, A., Arora, C.: Attention guided cosine margin to overcome class-imbalance in few-shot road object detection. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) Workshops. pp. 221-230 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 313, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 313, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 138, + 313, + 480, + 346 + ], + "type": "text", + "content": "2. Cao, Y., Wang, J., Jin, Y., Wu, T., Chen, K., Liu, Z., Lin, D.: Few-shot object detection via association and discrimination. In: Thirty-Fifth Conference on Neural Information Processing Systems (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 346, + 480, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 346, + 480, + 368 + ], + "spans": [ + { + "bbox": [ + 138, + 346, + 480, + 368 + ], + "type": "text", + "content": "3. Chen, H., Wang, Y., Wang, G., Qiao, Y.: LSTD: A Low-Shot Transfer Detector For Object Detection. In: AAAI. pp. 2836-2843 (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 369, + 480, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 369, + 480, + 411 + ], + "spans": [ + { + "bbox": [ + 138, + 369, + 480, + 411 + ], + "type": "text", + "content": "4. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 412, + 480, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 412, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 138, + 412, + 480, + 434 + ], + "type": "text", + "content": "5. Everingham, M., Van Gool, L., Williams, C.K.I., Winn, J., Zisserman, A.: The Pascal Visual Object Classes (VOC) Challenge. IJCV pp. 303-338 (2010)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 434, + 480, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 434, + 480, + 466 + ], + "spans": [ + { + "bbox": [ + 138, + 434, + 480, + 466 + ], + "type": "text", + "content": "6. Everingham, M., Van Gool, L., Williams, C., Winn, J., Zisserman, A.: The Pascal visual object classes (voc) challenge. International Journal of Computer Vision 88, 303-338 (06 2010)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 467, + 480, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 467, + 480, + 489 + ], + "spans": [ + { + "bbox": [ + 138, + 467, + 480, + 489 + ], + "type": "text", + "content": "7. Fan, Q., Zhuo, W., Tang, C.K., Tai, Y.W.: Few-Shot Object Detection With Attention-RPN And Multi-Relation Detector. In: CVPR (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 490, + 480, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 490, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 138, + 490, + 480, + 521 + ], + "type": "text", + "content": "8. Fan, Z., Ma, Y., Li, Z., Sun, J.: Generalized Few-Shot Object Detection Without Forgetting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4527-4536 (June 2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 522, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 522, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 522, + 480, + 544 + ], + "type": "text", + "content": "9. Fan, Z., Ma, Y., Li, Z., Sun, J.: Generalized few-shot object detection without forgetting (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 544, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 544, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 138, + 544, + 480, + 566 + ], + "type": "text", + "content": "0. Finn, C., Abbeel, P., Levine, S.: Model-Agnostic Meta-Learning For Fast Adaptation Of Deep Networks. In: ICML (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 567, + 469, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 469, + 577 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 469, + 577 + ], + "type": "text", + "content": "1. Fujishige, S.: Submodular Functions and Optimization, vol. 58. Elsevier (2005)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 578, + 480, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 480, + 598 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 480, + 598 + ], + "type": "text", + "content": "2. Gidaris, S., Komodakis, N.: Dynamic Few-Shot Visual Learning Without Forgetting. In: CVPR (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 599, + 321, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 599, + 321, + 610 + ], + "spans": [ + { + "bbox": [ + 138, + 599, + 321, + 610 + ], + "type": "text", + "content": "3. Girshick, R.B.: Fast R-CNN. ICCV (2015)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 611, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 480, + 643 + ], + "type": "text", + "content": "4. Han, G., Ma, J., Huang, S., Chen, L., Chang, S.F.: Few-shot object detection with fully cross-transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5321-5330 (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "5. Jegelka, S., Bilmes, J.: Submodularity beyond submodular energies: Coupling edges in graph cuts. In: CVPR 2011 (2011)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 246, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 246, + 91, + 447, + 101 + ], + "type": "text", + "content": "SMILE: Submodular Mutual Information Learner" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "16. Kang, B., Liu, Z., Wang, X., Yu, F., Feng, J., Darrell, T.: Few-shot Object Detection Via Feature Reweighting. In: ICCV (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 172 + ], + "type": "text", + "content": "17. Karlinsky, L., Shtok, J., Harary, S., Schwartz, E., Aides, A., Feris, R., Giryes, R., Bronstein, A.M.: RepMet: Representative-Based Metric Learning For Classification And Few-Shot Object Detection. In: CVPR (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 173, + 480, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 173, + 480, + 206 + ], + "spans": [ + { + "bbox": [ + 132, + 173, + 480, + 206 + ], + "type": "text", + "content": "18. Kaul, P., Xie, W., Zisserman, A.: Label, Verify, Correct: A Simple Few-Shot Object Detection Method. In: IEEE Conference on Computer Vision and Pattern Recognition (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 207, + 480, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 207, + 480, + 261 + ], + "spans": [ + { + "bbox": [ + 132, + 207, + 480, + 261 + ], + "type": "text", + "content": "19. Kaushal, V., Iyer, R., Doctor, K., Sahoo, A., Dubal, P., Kothawade, S., Mahadev, R., Dargan, K., Ramakrishnan, G.: Demystifying multi-faceted video summarization: Tradeoff between diversity, representation, coverage and importance. In: 2019 IEEE Winter Conference on Applications of Computer Vision (WACV). pp. 452-461 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 262, + 480, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 262, + 480, + 296 + ], + "spans": [ + { + "bbox": [ + 132, + 262, + 480, + 296 + ], + "type": "text", + "content": "20. Khosla, P., Teterwak, P., Wang, C., Sarna, A., Tian, Y., Isola, P., Maschinot, A., Liu, C., Krishnan, D.: Supervised contrastive learning. In: Advances in Neural Information Processing Systems (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 297, + 480, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 297, + 480, + 340 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 480, + 340 + ], + "type": "text", + "content": "21. Kothawade, S., Ghosh, S., Shekhar, S., Xiang, Y., Iyer, R.K.: Talisman: Targeted active learning for object detection with rare classes and slices using submodular mutual information. In: Computer Vision - ECCV 2022 - 17th European Conference (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 341, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 341, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 341, + 480, + 385 + ], + "type": "text", + "content": "22. Kothawade, S., Kaushal, V., Ramakrishnan, G., Bilmes, J.A., Iyer, R.K.: PRISM: A rich class of parameterized submodular information measures for guided data subset selection. In: Thirty-Sixth AAAI Conference on Artificial Intelligence, AAAI. pp. 10238-10246 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "type": "text", + "content": "23. Li, B., Liu, C., Shi, M., Chen, X., Ji, X., Ye, Q.: Proposal distribution calibration for few-shot object detection. IEEE transactions on neural networks and learning systems (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 419, + 480, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 419, + 480, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 419, + 480, + 441 + ], + "type": "text", + "content": "24. Li, B., Yang, B., Liu, C., Liu, F., Ji, R., Ye, Q.: Beyond Max-Margin: Class Margin Equilibrium For Few-Shot Object Detection. In: CVPR (June 2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 442, + 480, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 442, + 480, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 442, + 480, + 475 + ], + "type": "text", + "content": "25. Lin, H., Bilmes, J.: A class of submodular functions for document summarization. In: Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies (2011)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 475, + 480, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 480, + 508 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 480, + 508 + ], + "type": "text", + "content": "26. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision - ECCV 2014. pp. 740-755. Springer International Publishing, Cham (2014)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 509, + 480, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 509, + 480, + 552 + ], + "spans": [ + { + "bbox": [ + 132, + 509, + 480, + 552 + ], + "type": "text", + "content": "27. Liu, F., Zhang, X., Peng, Z., Guo, Z., Wan, F., Ji, X., Ye, Q.: Integrally Migrating Pre-trained Transformer Encoder-decoders for Visual Object Detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6825-6834 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 554, + 480, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 554, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 554, + 480, + 597 + ], + "type": "text", + "content": "28. Ma, J., Niu, Y., Xu, J., Huang, S., Han, G., Chang, S.F.: Digeo: Discriminative geometry-aware learning for generalized few-shot object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 599, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 480, + 632 + ], + "type": "text", + "content": "29. Majee, A., Agrawal, K., Subramanian, A.: Few-Shot Learning For Road Object Detection. In: AAAI Workshop on Meta-Learning and MetaDL Challenge. vol. 140, pp. 115-126 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "text", + "content": "30. Majee, A., Kothawade, S.N., Killamsetty, K., Iyer, R.K.: SCoRe: Submodular Combinatorial Representation Learning. In: Forty-first International Conference on Machine Learning (ICML) (2024)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "Majee et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 490 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "31. Ren, S., He, K., Girshick, R.B., Sun, J.: Faster R-CNN: Towards Real-Time Object Detection With Region Proposal Networks. IEEE Trans. on Pattern Analysis and Machine Intelligence (2015)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 481, + 171 + ], + "type": "text", + "content": "32. Shangguan, Z., Rostami, M.: Identification of novel classes for improving few-shot object detection (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 193 + ], + "type": "text", + "content": "33. Snell, J., Swersky, K., Zemel, R.: Prototypical Networks For Few-shot Learning. In: NeurIPS. pp. 4077-4087 (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 215 + ], + "type": "text", + "content": "34. Sun, B., Li, B., Cai, S., Yuan, Y., Zhang, C.: FSCE: Few-Shot Object Detection Via Contrastive Proposal Encoding. In: CVPR (June 2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "type": "text", + "content": "35. Sung, F., Yang, Y., Zhang, L., Xiang, T., Torr, P.H., Hospedales, T.M.: Learning To Compare: Relation Network For Few-Shot Learning. In: CVPR (June 2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 238, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 481, + 258 + ], + "type": "text", + "content": "36. Vinyals, O., Blundell, C., Lillicrap, T., kavukcuoglu, k., Wierstra, D.: Matching Networks For One Shot Learning. In: NeurIPS (2016)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "type": "text", + "content": "37. Wang, X., Huang, T.E., Darrell, T., Gonzalez, J.E., Yu, F.: Frustratingly Simple Few-Shot Object Detection. In: ICML (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 281, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 481, + 303 + ], + "type": "text", + "content": "38. Wang, Y.X., Ramanan, D., Hebert, M.: Meta-Learning To Detect Rare Objects. In: ICCV (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "type": "text", + "content": "39. Wu, J., Liu, S., Huang, D., Wang, Y.: Multi-scale positive sample refinement for few-shot object detection. In: European Conference on Computer Vision (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 326, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 346 + ], + "type": "text", + "content": "40. Xiao, Y., Marlet, R.: Few-Shot Object Detection And Viewpoint Estimation For Objects In The Wild. In: ECCV (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 348, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 348, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 348, + 481, + 380 + ], + "type": "text", + "content": "41. Yan, X., Chen, Z., Xu, A., Wang, X., Liang, X., Lin, L.: Meta R-CNN: Towards General Solver For Instance-Level Low-Shot Learning. In: CVPR. pp. 9577-9586 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "text", + "content": "42. Zhang, G., Cui, K., Wu, R., Lu, S., Tian, Y.: PNPDet: Efficient Few-Shot Detection Without Forgetting Via Plug-And-Play Sub-Networks. In: WACV. pp. 3823-3832 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 414, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 481, + 435 + ], + "type": "text", + "content": "43. Zhang, L., Zhou, S., Guan, J., Zhang, J.: Accurate Few-Shot Object Detection With Support-Query Mutual Guidance And Hybrid Loss. In: CVPR (June 2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 436, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 436, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 436, + 481, + 468 + ], + "type": "text", + "content": "44. Zhang, S., Luo, D., Wang, L., Koniusz, P.: Few-Shot Object Detection By Second-order Pooling. In: Proceedings of the Asian Conference on Computer Vision (ACCV) (November 2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 468, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 490 + ], + "type": "text", + "content": "45. Zhu, C., Chen, F., Ahmed, U., Shen, Z., Savvides, M.: Semantic Relation Reasoning For Shot-Stable Few-Shot Object Detection. In: CVPR (June 2021)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 246, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 246, + 91, + 447, + 101 + ], + "type": "text", + "content": "SMILE: Submodular Mutual Information Learner" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/SMooDi_ Stylized Motion Diffusion Model/10d37622-04ad-4741-9458-84cef09d41ff_content_list.json b/2024/SMooDi_ Stylized Motion Diffusion Model/10d37622-04ad-4741-9458-84cef09d41ff_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..12dde8ab72706507c9318bf70c8b02f81c67ef0a --- /dev/null +++ b/2024/SMooDi_ Stylized Motion Diffusion Model/10d37622-04ad-4741-9458-84cef09d41ff_content_list.json @@ -0,0 +1,1764 @@ +[ + { + "type": "text", + "text": "SMooDi: Stylized Motion Diffusion Model", + "text_level": 1, + "bbox": [ + 259, + 140, + 741, + 162 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lei Zhong $^{1}$ , Yiming Xie $^{1}$ , Varun Jampani $^{2}$ , Deqing Sun $^{3}$ , and Huaizu Jiang $^{1}$", + "bbox": [ + 227, + 188, + 774, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Northeastern University", + "bbox": [ + 411, + 231, + 589, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Stability AI", + "bbox": [ + 454, + 246, + 549, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Google Research", + "bbox": [ + 437, + 260, + 563, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{le.zhong,xie.yim,h.jiang}@northeastern.edu", + "bbox": [ + 336, + 275, + 666, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "varunjampani@gmail.com", + "bbox": [ + 416, + 290, + 586, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "deqingsun@google.com", + "bbox": [ + 423, + 303, + 578, + 315 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. We introduce a novel Stylized Motion Diffusion model, dubbed SMooDi, to generate stylized motion driven by content texts and style motion sequences. Unlike existing methods that either generate motion of various content or transfer style from one sequence to another, SMooDi can rapidly generate motion across a broad range of content and diverse styles. To this end, we tailor a pre-trained text-to-motion model for stylization. Specifically, we propose style guidance to ensure that the generated motion closely matches the reference style, alongside a lightweight style adaptor that directs the motion towards the desired style while ensuring realism. Experiments across various applications demonstrate that our proposed framework outperforms existing methods in stylized motion generation. Project Page: https://neu-vi.github.io/SMooDi/", + "bbox": [ + 259, + 349, + 746, + 515 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Motion synthesis $\\cdot$ Diffusion model $\\cdot$ Stylized motion", + "bbox": [ + 261, + 527, + 699, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 566, + 375, + 583 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We address the problem of generating stylized motion from a content text and a style motion sequence, as shown in Fig. 1. Human motion can typically be characterized by two components: content and style. Motion content represents the nature of a movement, such as walking and waving, and motion style reflects individual characteristics, such as personality traits (e.g., old, childlike) and emotions (e.g., happy, angry). Traditional pipelines create stylized motions via motion capture from actors, and are both labor-intensive and time-consuming. Therefore, decades of research have focused on developing automatic methods to assist stylized motion creation [1,11, 19].", + "bbox": [ + 212, + 597, + 785, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Motion style transfer [1, 45] is a practical and popular approach for the creation of stylized motion. It transfers the style from an existing style motion sequence to another existing content motion sequence. However, when a broad array of motion needs to be stylized, the pipeline may be inefficient - it would first require the collection of a large number of content motion sequences and then apply a motion style transfer method to process each sequence independently. Moreover, motion sequences are not always readily available, especially", + "bbox": [ + 212, + 734, + 785, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1a78d6a219a0266697792277f5abdf7c44a6ed82ffac53eaa9e197a415310631.jpg", + "image_caption": [ + "Fig. 1: SMooDi can generate realistic, stylized human motions given a content text a style motion sequence. It also accepts a motion sequence as content input. Darker color indicates later frames in the sequence. To better showcase the stylized motion generation, we place the style label for the each of the style motion sequence. Note that such style labels are not used as model input and shown here for visualization purpose only. (Best viewed in color.)" + ], + "image_footnote": [], + "bbox": [ + 236, + 148, + 759, + 366 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "for some customized content, such as running along a specific trajectory. They may still need to be created first by actors or animators for stylization.", + "bbox": [ + 212, + 500, + 785, + 530 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent advances of human motion generation with diffusion models [10, 17] have shown impressive results of creating diverse and realistic human motions. But most efforts have concentrated on efficiently and accurately translating textual prompts into human motions, focusing on the content only [6, 31, 46]. Integrating the style condition to generate stylized motions remains under-explored.", + "bbox": [ + 212, + 532, + 787, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Combining these two lines of research is a straightforward approach to tackle stylized motion generation, where a motion style transfer method [1, 45] can be applied to each motion sequence generated by a text-driven motion diffusion model [6, 31, 46]. However, in addition to the aforementioned inefficiency issue, this approach has two more limitations. First, error may accumulate across the pipeline. As, motion style transfer methods are usually trained with high-quality real-world motion sequences, we empirically observe that their performance may significantly degrade for imperfect motions produced by text-to-motion techniques. Second, existing motion style transfer methods rely on specialized style datasets [1, 27, 52] with limited motion content, which restricts their applications to motion diffusion models.", + "bbox": [ + 212, + 611, + 787, + 776 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we present a novel stylized motion diffusion model, dubbed SMooDi, that customizes a pre-trained text-to-motion model for stylization. Built upon the pre-trained motion latent diffusion model (MLD) [6], SMooDi inherits MLD's ability to generate diverse motion content. At the same time,", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Zhong et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SMooDi can generate motions in a variety of styles according to different style reference conditions, as shown in Fig. 1. Our main novelty is the style modulation module, which consists of a style adaptor and a style guidance module. First of all, drawing inspiration from controllable image generation [61], the style adaptor is designed to predict residual features conditioned on style reference motion sequence within each attention layer of MLD. It is useful for incorporating the style condition while ensuring the realism of the generated motion. Second, we design both classifier-free and classifier-based style guidance to more precisely control the stylized motion generation. Specifically, the classifier-free style and content guidance are linearly combined, where we can easily strike a balance between preserving content and reflecting style within the generated motion. At the same time, we design an additional classifier-based style guidance mechanism. It is an analytic function quantifying the disparity between the generated motion and the style reference motion in a style-centric embedding space, whose gradients are subsequently employed to guide the generated motion closer to the intended style. Our style adaptor and guidance module are designed to be complementary, which lead to high-quality stylized motion generation. The two modules are jointly optimized in a feature space instead of sequence-wise separate stitching, thereby avoiding the error accumulation issue.", + "bbox": [ + 217, + 146, + 787, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Although our approach is primarily designed for stylized motion generation driven by content text, we can utilize DDIM-Inversion [43] to identify the noisy latent corresponding to the content motion sequence. Following the same procedure as for text-driven content, SMooDi is capable of facilitating stylized motion generation based on content motion sequences. In other words, motion style transfer is a downstream application of our approach should it be desired in practice, e.g., to stylize the already created motion sequences.", + "bbox": [ + 212, + 458, + 787, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Experiments on the HumanML3D [16] and 100STYLE [27] datasets demonstrate that SMooDi surpasses other baseline models in generating stylized motion driven by content text, excelling in both content preservation and style reflection. More importantly, unlike previous methods that require individual fine-tuning for each style [12, 30, 54], SMooDi successfully integrates diverse content from the HumanML3D dataset and various styles from the 100STYLE dataset into a single model without requiring additional tuning during inference.", + "bbox": [ + 212, + 588, + 787, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To summarize, our contributions are: (1) To our knowledge, SMooDi is the first approach that adapts a pre-trained text-to-motion model to generate diverse stylized motion. (2) We introduce a novel style modulation module that utilizes a stylized adaptor and a style classifier guidance to enable stylized motion generation while ensuring style reflection, content preservation, and realism. (3) Experiments demonstrate that SMooDi not only sets a new state of the art in stylized motion generation driven by content text but also achieves performance comparable to state-of-the-art methods in motion style transfer.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "SMooDi: Stylized Motion Diffusion Model", + "bbox": [ + 450, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 143, + 387, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.1 Human Motion Generation", + "text_level": 1, + "bbox": [ + 215, + 172, + 485, + 186 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Human motion generation has attracted great attention [4, 5, 7, 9, 14, 16, 20, 32, 34-36, 47, 50, 51, 57, 58, 63]. Inspired by the impressive performance of diffusion models in image generation, a lot of works [6, 8, 13, 18, 22, 23, 25, 29, 32, 33, 39, 41, 46, 48, 53, 56, 60, 62, 64] utilize diffusion models to generate human motion. MDM [46] facilitates high-quality generation and versatile conditioning, providing a solid baseline for novel motion generation tasks. MLD [6] minimizes computational overhead during both training and inference by establishing the diffusion process within the latent space. Driven by the efficacy of diffusion models for control and conditioning, several studies have leveraged pre-trained motion diffusion models to generate long-sequence motions [41], enable human-object interactions [29], and control the joint trajectory of generated motions [23, 53]. However, there is no work exploring how to leverage pre-trained motion diffusion models to generate diverse stylized motion. While some studies [2, 3, 37] have enabled stylized motion generation in their diffusion pipeline, their methods are trained from scratch, and the supported styles are restricted by their motion content dataset. It is challenging for them to simultaneously support diverse motion content and style. In this work, we build upon a pre-trained motion diffusion model, MLD, and explore how to fine-tune it on a larger motion style dataset, 100STYLE, to learn diverse motion styles while retaining the ability to support motion generation across a wide range of content.", + "bbox": [ + 217, + 193, + 787, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Motion Style Transfer", + "text_level": 1, + "bbox": [ + 215, + 513, + 442, + 529 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recently, motion style transfer has seen quality enhancements through the adoption of various advanced neural architectures and generative models, such as graph neural networks [28], time-series models [27, 45], normalizing flows [49], and diffusion models [3,37]. Specifically, Aberman et al. [1] designed a two-branch generative adversarial network to disentangle motion style from content and facilitate their re-composition. Their approach effectively breaks the constraint of requiring a paired motion dataset. Motion Puzzle [19] realizes a framework that can control the style of individual body parts. Above methods extract both content and style features from the motion sequence. Moreover, Guo et al. [15] leverage the latent space of pre-trained motion models to enhance the extraction and infusion of motion content and style. However, a major limitation of these models is their reliance on specialized style datasets [1, 52] with limited motion content, which restricts their applications. In this work, we customize a pre-trained text-to-motion model for stylization, thus inheriting its ability to generate diverse motion content.", + "bbox": [ + 217, + 536, + 787, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Stylized Motion Diffusion Model", + "text_level": 1, + "bbox": [ + 215, + 781, + 568, + 799 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we introduce our proposed SMooDi for incorporating style conditions from a style motion sequence into a content-oriented pre-trained motion", + "bbox": [ + 215, + 809, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Zhong et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a065bb04d06a62b99007a08d915c79fc33dbe7da04b3af8b8c815c11e28fe1ec.jpg", + "image_caption": [ + "Fig. 2: Overview of SMooDi. Our model generates stylized human motions from content text and a style motion sequence. At the denoising step $t$ , our model takes the content text $\\mathbf{c}$ , style motion $\\mathbf{s}$ , and noisy latent $\\mathbf{z}_t$ as input and predicts $\\epsilon_t$ , which is then transferred to $\\mathbf{z}_{t-1}$ . This denoising step is repeated $T$ times to obtain the noise-free motion latent $\\mathbf{z}_0$ , which is fed into a motion decoder $D$ to produce the stylized motion." + ], + "image_footnote": [], + "bbox": [ + 254, + 150, + 754, + 276 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "diffusion model (MLD [6]). Fig. 2 presents an overview of SMooDi. Following the setting in MLD [6], we place the diffusion process in the motion latent space. Let $\\epsilon_{\\theta}$ denote the latent denoiser (a UNet parameterized by $\\theta$ ), and $\\{z_t\\}_{t=0}^T$ denote the sequence of noisy latents, where $z_T$ is a Gaussian noise. Given a content prompt $\\mathbf{c}$ and a style prompt $\\mathbf{s}$ , we define $\\epsilon_t = \\epsilon_{\\theta}(z_t, t, \\mathbf{c}, \\mathbf{s})$ for the denoising at step $t$ ( $0 < t \\leq T$ ). A cleaner noisy latent $z_{t-1}$ can be obtained by subtracting $\\epsilon_t$ from $z_t$ . The denoising step is repeated $T$ iterations until a clean latent $z_0$ is obtained. It can then be decoded by a motion decoder $\\mathbf{D}$ into a realistic motion sequence $\\mathbf{x} \\in \\mathbb{R}^{N \\times H}$ that accurately reflects both the content and style conditions. Here, $N$ represents the length of the motion sequence, and $H$ is the dimension of human motion representations. We employ the same motion representations as in HumanML3D [16], where $H = 263$ .", + "bbox": [ + 212, + 402, + 787, + 584 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in the Fig. 2, the content prompt is a text description, and the style prompt is provided by a reference style motion sequence $\\mathbf{s} \\in \\mathbb{R}^{N \\times H}$ . In this section, we focus on using a text description as the content prompt $\\mathbf{c}$ to explain our proposed stylized motion diffusion model. By employing the DDIM-Inversion [43] to identify the noisy latent corresponding to a motion content sequence, we can effectively use motion sequences as content prompts to generate stylized motions. In other words, motion style transfer is a downstream application of our proposed approach.", + "bbox": [ + 212, + 585, + 787, + 707 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our proposed stylization module consists of a style adaptor and a style guidance module. We will explain them separately in the rest of this section.", + "bbox": [ + 212, + 708, + 785, + 738 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Style Adaptor", + "text_level": 1, + "bbox": [ + 214, + 763, + 380, + 780 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although LoRA has been successfully used to incorporate \"style\" into the models in image domain [21, 42], they typically require training a separate LoRA for each style. In contrast, we focus on fine-tuning the model just once to adapt to", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "SMooDi: Stylized Motion Diffusion Model", + "bbox": [ + 450, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/176bb973d7f55c23ac7ec4bcd6f259025e607ffd350f5cebff31c4ab290e720c.jpg", + "image_caption": [ + "Fig. 3: Detailed illustration of our proposed style adaptor. The style adaptor is connected to the motion diffusion model via zero linear layer. The output of the style adaptor from each Transformer encoder is added to the motion diffusion model to steer the predicted noise towards the target style." + ], + "image_footnote": [], + "bbox": [ + 223, + 143, + 777, + 256 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "various motion styles, where adapting ControlNet [61] is more suitable. Therefore, we design a content-aware style adaptor based on ControlNet. This adaptor incorporates the motion style condition into the pre-trained MLD [6].", + "bbox": [ + 212, + 353, + 782, + 398 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Instead of learning to disentangle motion style from content from scratch on a large motion dataset, we redirect our focus towards capturing the motion style while ensuring the preservation of diverse motion content within the pre-trained MLD framework. Specifically, it consists of a trainable copy of the Transformer Encoder from the latent diffusion model in MLD. The architecture of the style adaptor is illustrated in Fig. 3. An independent style encoder is utilized to extract the style embedding from the style motion sequence $\\mathbf{s}$ . The style adaptor takes the same content prompt $\\mathbf{c}$ , the noised latent $\\mathbf{z}_t$ and timestep $t$ as in MLD, and the extracted style embedding. Each Transformer layer in the original latent diffusion model and the style adaptor is connected via a linear layer, with both weight and bias initially set to zeros. As training progresses, the style adaptor learns the style constraints and gradually applies the learned feature corrections to the corresponding layers in the latent diffusion model, thereby implicitly steering the output towards the desired style.", + "bbox": [ + 212, + 398, + 787, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Style Guidance", + "text_level": 1, + "bbox": [ + 215, + 635, + 387, + 648 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The style adaptor alone may not be sufficient to successfully incorporate the style condition. We further leverage both the classifier-free and classifier guidance to further enhance the stylization of a motion diffusion model. The combination of two types of guidance effectively ensures that generated motion meets multiple constraints while maintaining realism, complementing the style adaptor.", + "bbox": [ + 212, + 662, + 787, + 738 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Classifier-free Style Guidance. With the introduction of an extra style condition, we can divide the conditioned classifier-free guidance into two parts.", + "bbox": [ + 212, + 739, + 785, + 768 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon_ {\\theta} \\left(\\boldsymbol {z} _ {t}, t, \\mathbf {c}, \\mathbf {s}\\right) = \\epsilon_ {\\theta} \\left(\\boldsymbol {z} _ {t}, t, \\emptyset , \\emptyset\\right) +\n$$\n", + "text_format": "latex", + "bbox": [ + 225, + 782, + 434, + 797 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\underbrace {w _ {c} (\\epsilon_ {\\theta} (\\boldsymbol {z} _ {t} , t , \\mathbf {c} , \\emptyset) - \\epsilon_ {\\theta} (\\boldsymbol {z} _ {t} , t , \\emptyset , \\emptyset))} + \\underbrace {w _ {s} (\\epsilon_ {\\theta} (\\boldsymbol {z} _ {t} , t , \\mathbf {c} , \\mathbf {s}) - \\epsilon_ {\\theta} (\\boldsymbol {z} _ {t} , t , \\mathbf {c} , \\emptyset))} _ {\\text {(1)}}\n$$\n", + "text_format": "latex", + "bbox": [ + 282, + 801, + 785, + 825 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Classifier-free Content Guidance Classifier-free Style Guidance", + "bbox": [ + 303, + 825, + 720, + 835 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Zhong et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $w_{c}$ and $w_{s}$ represent the strengths of the classifier-free guidance for the condition $\\mathbf{c}$ and $\\mathbf{s}$ , respectively. We slightly abuse the notations here by using $\\emptyset$ to denote a condition is not used. The classifier-free content guidance is the same as in MLD, which can facilitate the text-to-motion generation process in combination with the first term's unconditioned guidance. Our proposed classifier-free style guidance works in a similar way. Note that $\\epsilon_{\\theta}(z_t,t,c,s)$ is MLD model with the style adaptor incorporated introduced in the previous section, which takes both a textual prompt and style motion sequence as input. By contrasting the text-driven denoising output with and without the style condition, it can highlight the effectiveness of the style input $s$ and facilitate the generation of stylized motion driven by content text. Our insight here is that by dividing the conditioned guidance into content and style components separately, we can easily strike a balance between preserving content and reflecting style within the generated motion.", + "bbox": [ + 212, + 146, + 787, + 356 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To better understand the classifier-free content and style guidance, we visualize each of them through decoding denoised latent $\\mathbf{z}_0$ into the motion space. As illustrated in Fig.4(a), the content guidance ensures the motion generation is faithful to the textual prompt, while the style guidance, as shown in Fig.4(b), emphasizes style-related characteristics in the output. Combining both forms of guidance results in a stylized motion that adheres to both content and style conditions, as illustrated in Fig. 4(c).", + "bbox": [ + 212, + 358, + 787, + 463 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Classifier-based Style Guidance. To further improve the stylization of a motion diffusion model, we adopt the classifier guidance [10,59] to provide stronger guidance to the generated motion towards the desired style. The core of our classifier-based style guidance is a novel analytic function $G(\\mathbf{z}_t,t,\\mathbf{s})$ , which calculates the $L_{1}$ distance between the style embedding of the generated clean motion $\\hat{\\pmb{x}}_0$ at denoising step $t$ and the reference style motions $\\mathbf{s}$ . The function's gradient is utilized to steer the generated motion towards the desired style.", + "bbox": [ + 212, + 464, + 787, + 571 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon_ {\\boldsymbol {\\theta}} (\\boldsymbol {z} _ {t}, t, \\mathbf {c}, \\mathbf {s}) = \\epsilon_ {\\boldsymbol {\\theta}} (\\boldsymbol {z} _ {t}, t, \\mathbf {c}, \\mathbf {s}) + \\tau \\nabla_ {\\boldsymbol {z} _ {t}} G (\\boldsymbol {z} _ {t}, t, \\mathbf {s}),\n$$\n", + "text_format": "latex", + "bbox": [ + 341, + 585, + 663, + 599 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nG \\left(\\boldsymbol {z} _ {t}, t, \\mathbf {s}\\right) = | f \\left(\\hat {\\boldsymbol {x}} _ {0}\\right) - f (\\mathbf {s}) |, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 602, + 784, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\tau$ adjusts the strength of reference-style guidance and $f$ denotes the style feature extractor. The generated motion $\\hat{\\pmb{x}}_0$ is obtained by first converting the denoising output latent $\\pmb{z}_t$ into the predicted clean latent as shown below:", + "bbox": [ + 214, + 631, + 787, + 678 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {z}} _ {0} = \\frac {\\boldsymbol {z} _ {t} - \\sqrt {1 - \\alpha_ {t}} \\varepsilon_ {\\theta} (\\boldsymbol {z} _ {t} , t , \\boldsymbol {c} , \\boldsymbol {s})}{\\sqrt {\\alpha_ {t}}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 689, + 785, + 722 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\alpha_{t}$ denotes the pre-defined noise scale in the forward process of the diffusion model. The predicted clean latent $\\hat{z}_0$ is then input into the motion decoder $\\mathbf{D}$ to obtain the generated motion.", + "bbox": [ + 212, + 733, + 785, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We obtain the style feature extractor by training a style classifier on the 100STYLE dataset [27] and removing its final layer. We refer the readers to more details in the supplementary materials. The training of the style feature extractor with ground-truth style labels for supervision enables it to effectively", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "SMooDi: Stylized Motion Diffusion Model", + "bbox": [ + 450, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7b64e7f402554b1ab66ab627b226f6f54cfa96271300186a8592e72263d86444.jpg", + "image_caption": [ + "Text: A person walks forward and then sits down.", + "Fig. 4: Visual illustrations of the classifier-free and classifier-based style guidance. (a) and (b) respectively show the classifier-free content and style guidance; (c) displays the initial stylized motion resulting from the combination of (a) and (b); (d) illustrates the refined stylized motion modified by the classifier-based style guidance." + ], + "image_footnote": [], + "bbox": [ + 243, + 164, + 754, + 232 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "capture style-related features. Therefore, style classifier guidance can provide more guidance to the stylized motion generation.", + "bbox": [ + 212, + 330, + 782, + 359 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Combination of the Two Style Guidance. The classifier-free and classifier-based style guidance are designed to complement each other, each playing a vital role in accurately reflecting the target style in the generated motions. As illustrated in Fig. 4, the desired style motion is \"arms open wide to the sides like an airplane\". The classifier-free style guidance (Fig. 4(b)) can capture style-related characteristic in a reasonably accurate manner. When combined with the classifier-free content guidance, it depicts the desired style (Fig.4(c)). Refined further by the classifier-based style guidance, the stylization is more authentic, where the person's harms are more open (Fig.4(d)). In addition to such visual results, quantitative ablation studies also verify the effectiveness of our proposed both classifier-free and classifier-based style guidance.", + "bbox": [ + 212, + 359, + 784, + 525 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "At the same time, although classifier-based style guidance offers precise style control, its effectiveness may be compromised when the content text significantly diverges from locomotion-related movements. This is because the style feature function is trained solely on the 100STYLE dataset, which contains only such movements. Over-reliance on style classifier guidance risks producing motions that fail to execute the desired actions, leading to unrealistic and physically implausible movements. Therefore, we leverage a content-aware style adaptor that establishes the fundamental style direction, while style classifier guidance refines this base for a more precise outcome. The effectiveness of this design is verified in our ablation studies.", + "bbox": [ + 212, + 526, + 787, + 676 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 Learning Scheme", + "text_level": 1, + "bbox": [ + 214, + 696, + 405, + 713 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Following [61], a straightforward approach to train SMooDi is to freeze the parameters of MLD and solely train the style adaptor on the 100STYLE dataset using the following loss function:", + "bbox": [ + 212, + 720, + 784, + 767 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s t d} = \\mathbb {E} _ {\\epsilon , \\mathbf {z}} \\left[ \\| \\epsilon_ {\\theta} (\\mathbf {z} _ {t}, t, \\mathbf {c}, \\mathbf {s}) - \\epsilon \\| _ {2} ^ {2} \\right], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 775, + 785, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\epsilon \\sim \\mathcal{N}(0, \\mathbf{I})$ represents the ground-truth noise added to $z_0$ . In our experiments, however, we found that this loss function alone leads to an issue of", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 116, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Zhong et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "\"content-forgetting\", where the model progressively looses the MLD's ability to generate motions with diverse contents. To address this issue, we design a content prior preservation loss $L_{pr}$ . Specifically, we randomly sample motions from the HumanML3D dataset to compute a prior preservation loss when fine-tuning SMooDi on the 100STYLE dataset.", + "bbox": [ + 212, + 146, + 787, + 220 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {p r} = \\mathbb {E} _ {\\epsilon^ {\\prime}, \\boldsymbol {z} ^ {\\prime}} \\left[ \\| \\epsilon_ {\\theta} (\\boldsymbol {z} _ {t} ^ {\\prime}, t, \\mathbf {c} ^ {\\prime}, \\mathbf {s} ^ {\\prime}) - \\epsilon^ {\\prime} \\| _ {2} ^ {2} \\right], \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 229, + 785, + 255 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\mathbf{z}_t^\\prime$ , $\\mathbf{c}^\\prime$ and $\\mathbf{s}^\\prime$ represents the motion latent, content prompt and style motion sequence derived from the HumanML3D dataset. $\\epsilon^\\prime$ is the noise map added to $\\mathbf{z}_0^\\prime$ . A similar solution is used in DreamBooth [40] to solve the \"language drift\" problem, where images generated from the frozen pretrained image generation model are utilized to enforce a class-prior preservation loss during model finetuning. Our content preservation loss can effectively mitigate content forgetting while learning diverse motion styles from the 100STYLE dataset.", + "bbox": [ + 212, + 263, + 787, + 369 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To further encourage the style adaptor to focus on motion style, while also ensuring that the latent diffusion model in MLD handles motion content well, we introduce an additional cycle prior-preservation loss, inspired by [19,55]. Specifically, we start this process by randomly sampling content text and motion style sequences from both the 100STYLE and HumanML3D datasets simultaneously. Then, we intermix the content text and motion style from these sequences with each other. Finally, we repeat this process to reconstruct the original motion sequences. The formula is expressed as follows:", + "bbox": [ + 212, + 369, + 787, + 489 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c y c} = \\mathbb {E} _ {\\boldsymbol {z}, \\boldsymbol {z} ^ {\\prime}, \\epsilon , \\epsilon^ {\\prime}} \\left[ \\left\\| \\epsilon_ {\\theta} (\\boldsymbol {z} _ {t} ^ {s h}, t, \\mathbf {c}, \\boldsymbol {s} ^ {h s}) + \\epsilon_ {\\theta} (\\boldsymbol {z} _ {t} ^ {h s}, t, \\mathbf {c} ^ {\\prime}, \\boldsymbol {s} ^ {s h}) - \\epsilon - \\epsilon^ {\\prime} \\right\\| _ {2} ^ {2} \\right], \\qquad (6)\n$$\n", + "text_format": "latex", + "bbox": [ + 269, + 501, + 785, + 527 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $s^{hs}$ denotes the motion sequence created by merging content from the HumanML3D dataset with style from the 100STYLE dataset. Similarly, $s^{sh}$ represents the sequence where content is sourced from the 100STYLE dataset and style from the HumanML3D dataset. The noised latent codes $z_{t}^{sh}$ and $z_{t}^{hs}$ correspond to $s^{sh}$ and $s^{hs}$ , respectively. Essentially, the cycle prior-preservation loss exchanges diverse content and style between two datasets, encouraging the content text to remain invariant in the generated motion under forward and backward translation. Overall, the training loss function of our framework is defined as follows:", + "bbox": [ + 212, + 537, + 787, + 671 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a l l}} = \\mathcal {L} _ {\\text {s t d}} + \\lambda_ {p r} \\mathcal {L} _ {p r} + \\lambda_ {\\text {c y c}} \\mathcal {L} _ {\\text {c y c}} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 674, + 784, + 690 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\lambda_{pr}$ and $\\lambda_{cyc}$ are hyperparameters. We refer readers to the pseudocode and illustration for training in the supplementary material for more details.", + "bbox": [ + 212, + 695, + 787, + 726 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 748, + 375, + 765 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We conduct experiments on both stylized text2motion and motion style transfer to demonstrate the effectiveness of our framework. Both tasks use a motion sequence as a style prompt, with the primary difference being their content prompt input: the former utilizes text, while the latter relies on motion sequences.", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "SMooDi: Stylized Motion Diffusion Model", + "bbox": [ + 450, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/8bc8b194ccbd0264c168f138b088101389be86adb4f9771a80ea88a9e764ddb3.jpg", + "table_caption": [ + "Table 1: Comparison with baseline methods on stylized motion generation driven by content text, using a combination of the 100STYLE (providing style) and HumanML3D datasets (providing content)." + ], + "table_footnote": [], + "table_body": "
MethodFID ↓ Foot skatingMM Dist ↓ R-precision ↑ Diversity → SRA ↑ ratio ↓
Ours1.6090.1244.4770.5719.23572.418
MLD+Motion Puzzle [19]6.1270.1856.4670.2906.476263.769
MLD+Aberman et al. [1]3.3090.3475.9830.4068.81654.367
ChatGPT+MLD0.6140.1314.3130.6058.8364.819
", + "bbox": [ + 261, + 198, + 741, + 282 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/8a9bb57b0d85d341d5bb052f1244e30c7e415cee0faba7aa1771752ae6c75e3b.jpg", + "image_caption": [ + "A person kicks." + ], + "image_footnote": [], + "bbox": [ + 250, + 303, + 359, + 368 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/5690f5675f7815d3acc4f2ec27d669e871e5bb5f76569ffeba5ad839e218ba7a.jpg", + "image_caption": [ + "(a) $\\mathrm{MLD + }$ Motion Puzzle (b) $\\mathrm{MLD + }$ Aberman et al." + ], + "image_footnote": [], + "bbox": [ + 460, + 306, + 531, + 359 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/d20b7f278bef388d97796ed81374f419ea0c4a86bbdb00444dfa53071d2365c8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 550, + 308, + 640, + 361 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/ba94ee848bc02ff1ac8a9219c441a35904447cfdf8276e3c99011f2b9ceb7ddd.jpg", + "image_caption": [ + "(c) Ours" + ], + "image_footnote": [], + "bbox": [ + 653, + 308, + 728, + 359 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/64b79eba7852a8af79637162ebaf2c7c698a869850f7b9c5b3ab9835c71c04b7.jpg", + "image_caption": [ + "Fig. 5: Qualitative comparisons of our approach and baseline methods on two stylized motion generation task." + ], + "image_footnote": [], + "bbox": [ + 250, + 371, + 354, + 433 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/0c83bd5658cd74abe2de00e7783896b233f1194aef387151aa646e2baf18b223.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 371, + 444, + 433 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/635b950c9548f5ef96be20ee1baaac4f7ad758840f926e21b03dc39487b7f76b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 460, + 371, + 550, + 434 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/4a540e666fc0cbc5de51165940de3a0905c63a35a52da71294eb2de9ab55ef05.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 563, + 371, + 643, + 434 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/58c685bcf66f0b58dad0b423be5d3619937b4c86c1e0669d46f06a3acab567b9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 650, + 371, + 736, + 433 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Datasets. We utilize the HumanML3D dataset [16] as our motion content dataset and the 100STYLE dataset [27] as our motion style dataset. The HumanML3D dataset is the largest motion capture dataset, featuring text annotations and comprising 14,646 motions and 44,970 motion annotations. Following the processing approach outlined in [16], we preprocess the HumanML3D dataset to obtain consistent motion representations. On the other hand, the 100STYLE dataset [27], being the largest motion style dataset, comprises up to 1,125 minutes of motion sequences, showcasing a wide array of 100 diverse locomotion styles. Due to differences in skeletons between the 100STYLE dataset and HumanML3D, we retarget the motions from 100STYLE to match the HumanML3D (SMPL-H) skeleton. Following this alignment, we apply the same processing steps as used for the HumanML3D dataset to preprocess the 100STYLE dataset. Moreover, as the 100STYLE dataset lacks text descriptions, we leverage MotionGPT [20] to generate pseudo text descriptions for the motion sequences in the 100STYLE dataset.", + "bbox": [ + 212, + 507, + 787, + 733 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation metrics are designed to assess three dimensions: Content Preservation, Style Reflection, and Realism. For content preservation and style reflection assessment, we employ metrics consistent with those used in [6]: motion-retrieval precision (R precision), Multi-modal Distance (MM Dist), Diversity, and Frechet Inception Distance (FID). Additionally, recognizing the common foot skating issues in kinetics-based motion generation methods, we incorporate the foot skating ratio metric proposed by [23] into our motion quality evaluation. For style", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Zhong et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "reflection, we employ Style Recognition Accuracy (SRA) [19]. During evaluation, we randomly select a content text from the HumanML3D dataset and a motion style sequence from the 100STYLE dataset to generate the stylized motion. We then use a pre-trained style classifier to compute the SRA for the generated motion. It's noteworthy that some motion style labels in the 100STYLE dataset, like 'kick' and 'jump,' inherently convey motion content, which may conflict with the content text in HumanML3D dataset. To address this, we categorize the motion style labels into server groups following the approach by Kim et al. [24], Specifically excluding the 'ACT' group ensures that only motion style labels not conflicting with motion content are considered when computing the SRA metric. Further details about the evaluation metrics are provided in the supplementary material.", + "bbox": [ + 212, + 146, + 787, + 327 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Baselines. For motion style transfer task, we compare our methods with two state-of-the-art methods, namely Motion Puzzle [19] and Aberman et al. [1]. To ensure a fair comparison, we train the compared methods under the same settings as ours, using a combined dataset comprising HumanML3D and 100STYLE. Due to the constraints of the multi-class discriminator in Aberman et al., which requires style labels, we adopt the training method outlined in Motion Puzzle to eliminate the need for style labels. For stylized text2motion task, we compare our method against baselines capable of generating stylized motion from content text and style motion sequences. The straightforward baselines involve applying motion style transfer methods to the motion sequences generated by the text2motion model. To align with our approach that uses a pre-trained motion diffusion model, the text2motion models in the baselines for stylized motion generation select MLD, and the motion style transfer methods are consistent with those used in the motion style transfer task. For the stylized text2motion task, we compare our method against two kinds of baselines capable of generating stylized motion from content text and style motion sequences. The first kind of baseline involves applying motion style transfer methods to the motion sequences generated by the text2motion model. To align with our approach that uses a pretrained motion diffusion model, the text2motion models select MLD [6], and the motion style transfer methods are consistent with those used in the motion style transfer task. The second kind of baseline involves using ChatGPT to merge style labels from 100STYLE with text from HumanML3D into a sentence. For example, given the content text 'a person walks.' and the style label 'old,' we obtain 'an elderly person walks.' This merged sentence is then fed to MLD.", + "bbox": [ + 212, + 329, + 787, + 690 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.1 Comparison to Baseline Methods", + "text_level": 1, + "bbox": [ + 214, + 710, + 535, + 724 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Quantitative and Qualitative For the task of stylized text2motion, Table 1 reports the comparisons of our method with the three baseline methods.", + "bbox": [ + 212, + 734, + 782, + 762 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "As shown in the $3^{rd}$ row of Table 1, ChatGPT+MLD only achieves around $5.29\\%$ in terms of SRA, indicating that MLD cannot enable stylized generation from text alone, even though it contains style descriptions. Notably, our method outperforms the two baselines that combine MLD with motion style transfer methods in all metrics.", + "bbox": [ + 212, + 763, + 785, + 838 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "SMooDi: Stylized Motion Diffusion Model", + "bbox": [ + 450, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "(a) Evaluation on HumanML3D dataset", + "bbox": [ + 243, + 183, + 477, + 196 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/1866ec59a6e555e9a0d5f77a9fac17ce646ad284c17370c40d7e018957c94724.jpg", + "table_caption": [ + "Table 2: Comparison with baseline methods on motion style transfer." + ], + "table_footnote": [], + "table_body": "
MethodFoot skating ratio ↓
Ours0.0951.58265.147
Motion Puzzle [19]0.1976.87167.233
(Aberman et al [1])0.3383.89261.006
", + "bbox": [ + 238, + 205, + 480, + 271 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/dcf9f82e01b7034c50ee9b03d78220f3c51b5f91ae71003d289fa8f05e69e858.jpg", + "table_caption": [ + "(b) Evaluation on Xia dataset" + ], + "table_footnote": [], + "table_body": "
MethodFoot skating ratio↓FID↓SRA↑(%)CRA↑(%)
Ours0.03174.66361.11145.555
Motion Puzzle [19]0.03165.36067.77825.556
(Aberman et al [1])0.02605.68156.66734.444
", + "bbox": [ + 501, + 205, + 787, + 270 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Specifically, our method performs better than MLD+Motion Puzzle in the SRA metric by $13.56\\%$ and significantly outperforms MLD+Aberman et al. in the FID metric by $51.38\\%$ and $0.64\\%$ in the R-precision metric. The first row of Fig. 5 validates our observation, where the motion generated by our method performs better in adhering to both content and style constraints than baseline methods. In contrast, MLP+Aberman et al. can successfully perform the action but fail to reflect the motion style in Fig.5(b), while Motion Puzzle can accurately reflect the motion style but struggles to effectively perform the action in Fig.5(a).", + "bbox": [ + 212, + 305, + 784, + 426 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "For the task of motion style transfer, since it does not take content text as input, text-motion related metrics such as MM Dist, R-precision, and Diversity are not applicable and thus are not reported. Part (a) of Table 2 presents a comparison between our method and the two baseline methods, using the HumanML3D dataset as the motion content source and drawing motion styles from the 100STYLE dataset. Our method delivers competitive results in the SRA metric and excels in the FID and foot skating ratio metrics. Specifically, we see a substantial $59.35\\%$ improvement in the FID metric over Aberman et al. [1]. To more effectively compare the generalizability of different methods, we conduct experiments on the Xia dataset [52], a small, specific motion style dataset that was unseen by our and the baseline models during training. Because motion content labels are present in the Xia dataset, we report the Content Recognition Accuracy (CRA). Part (b) of Table 2 showcases the results. Our method maintains competitive performance in the SRA metric, with only a marginal $9.83\\%$ decrease in SRA compared to Motion Puzzle. On the contrary, our method exhibits a significant $32.26\\%$ increase in the CAR metric relative to Aberman et al., and a notable $78.26\\%$ enhancement over Motion Puzzle. Our method achieves a better balance between style reflection and content preservation. The second row of Fig. 5 validates this observation. It is worth noting that, unlike other motion style transfer methods, our method does not incorporate objectives for enabling stylized motion generation using motion content sequences. Through simple DDIM-Inversion and without any additional optimization or regularization, our method achieves performance comparable to existing motion style transfer methods.", + "bbox": [ + 212, + 429, + 787, + 790 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "User Study. Due to the highly subjective nature of stylized motion, we conduct User studies using pairwise comparisons to further evaluate our proposed method in the tasks of stylized motion generation and motion style transfer. We", + "bbox": [ + 212, + 794, + 787, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Zhong et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/78c81354b646e1c0ff69593bd136bcabf93540288b16b965e256231df98fe389.jpg", + "table_caption": [ + "Table 3: Ablation Studies on HumanML3D Content and 100STYLE Styles." + ], + "table_footnote": [], + "table_body": "
MethodFID↓ ratio↓MM Dist↓R-precision↑ (Top-3)Diversity→ SRA(%)↑
Ours (on all)1.6090.1244.4770.5719.23572.418
w/o Lcyc2.0460.1364.4650.5698.86964.866
w/o Lpr + Lcyc5.9960.1666.0980.3357.45681.841
w/o classifier-based1.0500.1114.0850.6309.44520.245
w/o adaptor2.9840.1234.5260.5508.37269.952
", + "bbox": [ + 289, + 170, + 715, + 266 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Text: A person walks forward and then sits down.", + "bbox": [ + 336, + 281, + 617, + 294 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/9cc5618563a75e1af0ddadb6dcabde399b405b079118742d11e117b59ea4d369.jpg", + "image_caption": [ + "Fig. 6: Visual comparisons of the ablation designs and our full model." + ], + "image_footnote": [], + "bbox": [ + 243, + 294, + 746, + 359 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "recruited 22 human subjects to participate in the study. In each test, participants are presented with two 4-second video clips synthesized by our method and one comparison method. They are then required to select their preferred clip while considering Realism, Style Reflection, and Content Preservation dimensions, respectively. As shown in Fig. 7, our method receives more user appreciation compared to two baselines across three dimensions in two tasks. Further user study details are provided in supplementary.", + "bbox": [ + 212, + 415, + 787, + 521 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.2 Ablation studies", + "text_level": 1, + "bbox": [ + 215, + 542, + 398, + 556 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "To validate the effectiveness of our framework's design choices, we have conducted several ablation studies: the first assesses the impact of each loss function term, while the second evaluates the influence of the style adaptor and style guidance during sampling.", + "bbox": [ + 212, + 568, + 785, + 628 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Loss Components. Firstly, we exclude the cycle-prior term in the loss function, denoted as $w / o$ $L_{cyc}$ . Comparing the results in the $1^{\\mathrm{st}}$ and $2^{\\mathrm{nd}}$ rows in Table 3, we observe that our full model outperforms in all content preservation and style reflection metrics. The motion generated by our approach can still perform the content adhering to the text description but performs worse in accurately reflecting the motion style, as reflected by the arms not being fully extended horizontally.", + "bbox": [ + 212, + 628, + 787, + 733 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Since the cycle prior-preservation term is built upon the prior-preservation term, it is meaningless to exclude $L_{pr}$ while retaining $L_{cyc}$ . Therefore, we further exclude both the prior-preservation and cycle-prior term, denoted as $w / oL_{pr} + L_{cyc}$ in the $3^{\\mathrm{rd}}$ row. By comparing the results in the $2^{\\mathrm{nd}}$ and $3^{\\mathrm{rd}}$ rows of Table 3, we notice that while the number of SRAs is higher in the third row, other metrics show a significant decline. Specifically, in terms of the FID metric, performance deteriorates by more than $229\\%$ . Indeed, without $L_{pr}$ , the", + "bbox": [ + 212, + 734, + 787, + 842 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "SMooDi: Stylized Motion Diffusion Model", + "bbox": [ + 450, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/c4a1190969a76af004907c207d9a7b1c4fa11a0ca8aec274d55bec4405b3e88c.jpg", + "image_caption": [ + "Fig. 7: User Study on two stylized motion generation tasks." + ], + "image_footnote": [], + "bbox": [ + 272, + 146, + 495, + 270 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/da2f3a7d738cef23e56906e87ff523501e30364fbf6d2e37a13f1ebe0745bcc2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 146, + 723, + 268 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "model tends to lose the ability to translate content text into corresponding motion, a phenomenon named 'content-forgetting' as described in Sec. 3.3. Fig. 6(a) validates our observation, showing that the content in the generated motion significantly deviates from the text descriptions and closely resembles the style motion sequence.", + "bbox": [ + 212, + 325, + 784, + 400 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Style Adaptor and Style Guidance. Initially, we compare our model to a variant without the classifier-based guidance, $w / o$ classifier-based, to demonstrate its effectiveness. The $5^{\\text{th}}$ row of Table 3 presents the results. Consistent with the findings in [19], a reasonable trade-off between content preservation and style reflection is observed. Although classifier-based style guidance may slightly affect the content preservation metrics, it significantly boosts the model's performance in the SRA metric, yielding an impressive $208\\%$ improvement.Fig. 6(d) demonstrates that, without classifier-based style guidance, the generated motions can reflect the motion style, yet they still fall short of fully achieving the target style. Classifier-based style guidance can effectively bridge this gap.", + "bbox": [ + 212, + 412, + 784, + 561 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Subsequently, as shown in Table 3, we evaluate a variant without the style adaptor, denoted as $w / o$ adaptor (the last row). In cases where the SRA values are close, the style adaptor improved the FID metric by about $80.46\\%$ . Fig. 6(c) shows that the generated motion can greatly perform the 'walk' action while successfully reflecting the style, but fails to perform the 'sit' action. This indicates that the effectiveness of classifier-based style guidance diminishes when the content text deviates from locomotion-related movements. Relying solely on it may even adversely affect action performance.", + "bbox": [ + 212, + 564, + 785, + 684 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 705, + 359, + 720 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we introduce the Stylized Motion Diffusion Model, a novel approach that leverages a pre-trained motion diffusion model to facilitate stylized motion generation driven by content text. By integrating a style adaptor and style classifier guidance, our method is capable of producing realistic human motions that accurately reflect both the content text descriptions and the desired motion style from motion sequences. Through detailed ablation studies, we have demonstrated the effectiveness of each component in our framework.", + "bbox": [ + 212, + 734, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Zhong et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Aberman, K., Weng, Y., Lischinski, D., Cohen-Or, D., Chen, B.: Unpaired motion style transfer from video to animation. TOG (2020)", + "2. Alexanderson, S., Nagy, R., Beskow, J., Henter, G.E.: Listen, denoise, action! audio-driven motion synthesis with diffusion models. TOG (2023)", + "3. Ao, T., Zhang, Z., Liu, L.: Gesturediffuclip: Gesture diffusion model with clip latents. TOG (2023)", + "4. Cen, Z., Pi, H., Peng, S., Shen, Z., Yang, M., Zhu, S., Bao, H., Zhou, X.: Generating human motion in 3d scenes from text descriptions. In: CVPR (2024)", + "5. Chen, L.H., Lu, S., Zeng, A., Zhang, H., Wang, B., Zhang, R., Zhang, L.: Motionllm: Understanding human behaviors from human motions and videos. ArXiv (2024)", + "6. Chen, X., Jiang, B., Liu, W., Huang, Z., Fu, B., Chen, T., Yu, G.: Executing your commands via motion diffusion in latent space. In: CVPR (2023)", + "7. Cohan, S., Tevet, G., Reda, D., Peng, X.B., van de Panne, M.: Flexible motion in-between with diffusion models. ArXiv (2024)", + "8. Dabral, R., Mughal, M.H., Golyanik, V., Theobalt, C.: Mofusion: A framework for denoising-diffusion-based motion synthesis. In: CVPR (2023)", + "9. Dai, W., Chen, L.H., Wang, J., Liu, J., Dai, B., Tang, Y.: Motionlcm: Real-time controllable motion generation via latent consistency model. ArXiv (2024)", + "0. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. NeurIPS (2021)", + "1. Du, H., Herrmann, E., Sprenger, J., Fischer, K., Slusallek, P.: Stylistic locomotion modeling and synthesis using variational generative models. In: Proceedings of the 12th ACM SIGGRAPH Conference on Motion, Interaction and Games (2019)", + "2. Everaert, M.N., Bocchio, M., Arpa, S., Susstrunk, S., Achanta, R.: Diffusion in style. In: ICCV (2023)", + "3. Ghosh, A., Dabral, R., Golyanik, V., Theobalt, C., Slusallek, P.: Remos: Reactive 3d motion synthesis for two-person interactions. In: ArXiv (2023)", + "4. Guo, C., Mu, Y., Javed, M.G., Wang, S., Cheng, L.: Momask: Generative masked modeling of 3d human motions. In: CVPR (2024)", + "5. Guo, C., Mu, Y., Zuo, X., Dai, P., Yan, Y., Lu, J., Cheng, L.: Generative human motion stylization in latent space. ArXiv (2024)", + "6. Guo, C., Zou, S., Zuo, X., Wang, S., Ji, W., Li, X., Cheng, L.: Generating diverse and natural 3d human motions from text. In: CVPR (2022)", + "7. Ho, J., Salimans, T.: Classifier-free diffusion guidance. ArXiv (2022)", + "8. Huang, S., Wang, Z., Li, P., Jia, B., Liu, T., Zhu, Y., Liang, W., Zhu, S.C.: Diffusion-based generation, optimization, and planning in 3d scenes. In: CVPR (2023)", + "9. Jang, D.K., Park, S., Lee, S.H.: Motion puzzle: Arbitrary motion style transfer by body part. TOG (2022)", + "20. Jiang, B., Chen, X., Liu, W., Yu, J., Yu, G., Chen, T.: Motiongpt: Human motion as a foreign language. ArXiv (2023)", + "21. Jones, M., Wang, S.Y., Kumari, N., Bau, D., Zhu, J.Y.: Customizing text-to-image models with a single image pair. ArXiv (2024)", + "22. Karunratanakul, K., Preechakul, K., Aksan, E., Beeler, T., Suwajanakorn, S., Tang, S.: Optimizing diffusion noise can serve as universal motion priors. In: Arxiv (2023)" + ], + "bbox": [ + 225, + 178, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "SMooDi: Stylized Motion Diffusion Model", + "bbox": [ + 450, + 114, + 730, + 127 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "23. Karunratanakul, K., Preechakul, K., Suwajanakorn, S., Tang, S.: Gmd: Controllable human motion synthesis via guided diffusion models. In: ICCV (2023)", + "24. Kim, H.J., Lee, S.H.: Perceptual characteristics by motion style category. In: Eurographics (Short Papers) (2019)", + "25. Kulkarni, N., Rempe, D., Genova, K., Kundu, A., Johnson, J., Fouhey, D., Guibas, L.: Nifty: Neural object interaction fields for guided human motion synthesis. ArXiv (2023)", + "26. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. ArXiv (2017)", + "27. Mason, I., Starke, S., Komura, T.: Real-time style modelling of human locomotion via feature-wise transformations and local motion phases. Proceedings of the ACM on Computer Graphics and Interactive Techniques (2022)", + "28. Park, S., Jang, D.K., Lee, S.H.: Diverse motion stylization for multiple style domains via spatial-temporal graph-based generative model. Proceedings of the ACM on Computer Graphics and Interactive Techniques (2021)", + "29. Peng, X., Xie, Y., Wu, Z., Jampani, V., Sun, D., Jiang, H.: Hoi-diff: Text-driven synthesis of 3d human-object interactions using diffusion models. ArXiv (2023)", + "30. Peng, X.B., Ma, Z., Abbeel, P., Levine, S., Kanazawa, A.: Amp: Adversarial motion priors for stylized physics-based character control. TOG (2021)", + "31. Petrovich, M., Black, M.J., Varol, G.: Temos: Generating diverse human motions from textual descriptions. In: ECCV (2022)", + "32. Petrovich, M., Litany, O., Iqbal, U., Black, M.J., Varol, G., Bin Peng, X., Rempe, D.: Multi-track timeline control for text-driven 3d human motion generation. In: CVPR (2024)", + "33. Pi, H., Peng, S., Yang, M., Zhou, X., Bao, H.: Hierarchical generation of human-object interactions with diffusion probabilistic models. In: ICCV (2023)", + "34. Pinyoanuntapong, E., Saleem, M.U., Wang, P., Lee, M., Das, S., Chen, C.: Bamm: Bidirectional autoregressive motion model. ArXiv (2024)", + "35. Pinyoanuntapong, E., Wang, P., Lee, M., Chen, C.: Mmm: Generative masked motion model. In: CVPR (2024)", + "36. Raab, S., Gat, I., Sala, N., Tevet, G., Shalev-Arkushin, R., Fried, O., Bermano, A.H., Cohen-Or, D.: Monkey see, monkey do: Harnessing self-attention in motion diffusion for zero-shot motion transfer. ArXiv (2024)", + "37. Raab, S., Leibovitch, I., Tevet, G., Arar, M., Bermano, A.H., Cohen-Or, D.: Single motion diffusion. ArXiv (2023)", + "38. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021)", + "39. Rempe, D., Luo, Z., Peng, X.B., Yuan, Y., Kitani, K., Kreis, K., Fidler, S., Litany, O.: Trace and pace: Controllable pedestrian animation via guided trajectory diffusion. In: CVPR (2023)", + "40. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: CVPR (2023)", + "41. Shafir, Y., Tevet, G., Kapon, R., Bermano, A.H.: Human motion diffusion as a generative prior. ArXiv (2023)", + "42. Shah, V., Ruiz, N., Cole, F., Lu, E., Lazebnik, S., Li, Y., Jampani, V.:Ziplora: Any subject in any style by effectively merging loras. ArXiv (2023)", + "43. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. ArXiv (2020)", + "44. Song, Y., Dhariwal, P., Chen, M., Sutskever, I.: Consistency models. ArXiv (2023)", + "45. Tao, T., Zhan, X., Chen, Z., van de Panne, M.: Style-erd: Responsive and coherent online motion style transfer. Arxiv (2022)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Zhong et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "46. Tevet, G., Raab, S., Gordon, B., Shafir, Y., Cohen-or, D., Bermano, A.H.: Human motion diffusion model. In: ICLR (2023)", + "47. Wan, W., Dou, Z., Komura, T., Wang, W., Jayaraman, D., Liu, L.: Tlcontrol: Trajectory and language control for human motion synthesis. ArXiv (2023)", + "48. Wang, Z., Wang, J., Lin, D., Dai, B.: Intercontrol: Generate human motion interactions by controlling every joint. ArXiv (2023)", + "49. Wen, Y.H., Yang, Z., Fu, H., Gao, L., Sun, Y., Liu, Y.J.: Autoregressive stylized motion synthesis with generative flow. In: CVPR (2021)", + "50. Wu, Q., Zhao, Y., Wang, Y., Tai, Y.W., Tang, C.K.: Motionllm: Multimodal motion-language learning with large language models. ArXiv (2024)", + "51. Wu, Q., Shi, Y., Huang, X., Yu, J., Xu, L., Wang, J.: Thor: Text to human-object interaction diffusion via relation intervention. ArXiv (2024)", + "52. Xia, S., Wang, C., Chai, J., Hodgins, J.: Realtime style transfer for unlabeled heterogeneous human motion. TOG (2015)", + "53. Xie, Y., Jampani, V., Zhong, L., Sun, D., Jiang, H.: Omnicontrol: Control any joint at any time for human motion generation. In: ICLR (2024)", + "54. Xu, P., Xie, K., Andrews, S., Kry, P.G., Neff, M., McGuire, M., Karamouzas, I., Zordan, V.: Adaptnet: Policy adaptation for physics-based character control. TOG (2023)", + "55. Xu, S., Ma, Z., Huang, Y., Lee, H., Chai, J.: Rethinking cycle consistency in text-guided diffusion for image manipulation. NeurIPS (2024)", + "56. Xu, S., Li, Z., Wang, Y.X., Gui, L.Y.: Interdiff: Generating 3d human-object interactions with physics-informed diffusion. In: ICCV (2023)", + "57. Xu, S., Wang, Z., Wang, Y.X., Gui, L.Y.: Interdreamer: Zero-shot text to 3d dynamic human-object interaction. ArXiv (2024)", + "58. Yi, H., Thies, J., Black, M.J., Peng, X.B., Rempe, D.: Generating human interaction motions in scenes with text control. ArXiv (2024)", + "59. Yu, J., Wang, Y., Zhao, C., Ghanem, B., Zhang, J.: Freedom: Training-free energy-guided conditional diffusion model. ArXiv (2023)", + "60. Yuan, Y., Song, J., Iqbal, U., Vahdat, A., Kautz, J.: Physdiff: Physics-guided human motion diffusion model. In: ICCV (2023)", + "61. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: ICCV (2023)", + "62. Zhang, M., Cai, Z., Pan, L., Hong, F., Guo, X., Yang, L., Liu, Z.: Motiondiffuse: Text-driven human motion generation with diffusion model. PAMI (2024)", + "63. Zhang, Y., Huang, D., Liu, B., Tang, S., Lu, Y., Chen, L., Bai, L., Chu, Q., Yu, N., Ouyang, W.: Motiongpt: Finetuned llms are general-purpose motion generators. ArXiv (2023)", + "64. Zhou, W., Dou, Z., Cao, Z., Liao, Z., Wang, J., Wang, W., Liu, Y., Komura, T., Wang, W., Liu, L.: Emdm: Efficient motion diffusion model for fast, high-quality motion generation. Arxiv (2023)", + "65. Zhou, Y., Barnes, C., Lu, J., Yang, J., Li, H.: On the continuity of rotation representations in neural networks. In: CVPR (2019)" + ], + "bbox": [ + 212, + 146, + 787, + 742 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "SMooDi: Stylized Motion Diffusion Model", + "bbox": [ + 450, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/SMooDi_ Stylized Motion Diffusion Model/10d37622-04ad-4741-9458-84cef09d41ff_model.json b/2024/SMooDi_ Stylized Motion Diffusion Model/10d37622-04ad-4741-9458-84cef09d41ff_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a036274512095c2e6359f03e2935d4c82b493f7c --- /dev/null +++ b/2024/SMooDi_ Stylized Motion Diffusion Model/10d37622-04ad-4741-9458-84cef09d41ff_model.json @@ -0,0 +1,2511 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.26, + 0.141, + 0.743, + 0.163 + ], + "angle": 0, + "content": "SMooDi: Stylized Motion Diffusion Model" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.189, + 0.776, + 0.221 + ], + "angle": 0, + "content": "Lei Zhong\\(^{1}\\), Yiming Xie\\(^{1}\\), Varun Jampani\\(^{2}\\), Deqing Sun\\(^{3}\\), and Huaizu Jiang\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.413, + 0.232, + 0.59, + 0.246 + ], + "angle": 0, + "content": "1 Northeastern University" + }, + { + "type": "text", + "bbox": [ + 0.455, + 0.247, + 0.55, + 0.26 + ], + "angle": 0, + "content": "2 Stability AI" + }, + { + "type": "text", + "bbox": [ + 0.439, + 0.261, + 0.564, + 0.274 + ], + "angle": 0, + "content": "3 Google Research" + }, + { + "type": "text", + "bbox": [ + 0.337, + 0.276, + 0.668, + 0.289 + ], + "angle": 0, + "content": "{le.zhong,xie.yim,h.jiang}@northeastern.edu" + }, + { + "type": "text", + "bbox": [ + 0.418, + 0.291, + 0.588, + 0.303 + ], + "angle": 0, + "content": "varunjampani@gmail.com" + }, + { + "type": "text", + "bbox": [ + 0.424, + 0.304, + 0.58, + 0.316 + ], + "angle": 0, + "content": "deqingsun@google.com" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.35, + 0.747, + 0.516 + ], + "angle": 0, + "content": "Abstract. We introduce a novel Stylized Motion Diffusion model, dubbed SMooDi, to generate stylized motion driven by content texts and style motion sequences. Unlike existing methods that either generate motion of various content or transfer style from one sequence to another, SMooDi can rapidly generate motion across a broad range of content and diverse styles. To this end, we tailor a pre-trained text-to-motion model for stylization. Specifically, we propose style guidance to ensure that the generated motion closely matches the reference style, alongside a lightweight style adaptor that directs the motion towards the desired style while ensuring realism. Experiments across various applications demonstrate that our proposed framework outperforms existing methods in stylized motion generation. Project Page: https://neu-vi.github.io/SMooDi/" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.529, + 0.7, + 0.543 + ], + "angle": 0, + "content": "Keywords: Motion synthesis \\(\\cdot\\) Diffusion model \\(\\cdot\\) Stylized motion" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.568, + 0.377, + 0.584 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.787, + 0.734 + ], + "angle": 0, + "content": "We address the problem of generating stylized motion from a content text and a style motion sequence, as shown in Fig. 1. Human motion can typically be characterized by two components: content and style. Motion content represents the nature of a movement, such as walking and waving, and motion style reflects individual characteristics, such as personality traits (e.g., old, childlike) and emotions (e.g., happy, angry). Traditional pipelines create stylized motions via motion capture from actors, and are both labor-intensive and time-consuming. Therefore, decades of research have focused on developing automatic methods to assist stylized motion creation [1,11, 19]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Motion style transfer [1, 45] is a practical and popular approach for the creation of stylized motion. It transfers the style from an existing style motion sequence to another existing content motion sequence. However, when a broad array of motion needs to be stylized, the pipeline may be inefficient - it would first require the collection of a large number of content motion sequences and then apply a motion style transfer method to process each sequence independently. Moreover, motion sequences are not always readily available, especially" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Zhong et al." + }, + { + "type": "image", + "bbox": [ + 0.237, + 0.149, + 0.761, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.381, + 0.789, + 0.466 + ], + "angle": 0, + "content": "Fig. 1: SMooDi can generate realistic, stylized human motions given a content text a style motion sequence. It also accepts a motion sequence as content input. Darker color indicates later frames in the sequence. To better showcase the stylized motion generation, we place the style label for the each of the style motion sequence. Note that such style labels are not used as model input and shown here for visualization purpose only. (Best viewed in color.)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.501, + 0.787, + 0.531 + ], + "angle": 0, + "content": "for some customized content, such as running along a specific trajectory. They may still need to be created first by actors or animators for stylization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.533, + 0.788, + 0.61 + ], + "angle": 0, + "content": "Recent advances of human motion generation with diffusion models [10, 17] have shown impressive results of creating diverse and realistic human motions. But most efforts have concentrated on efficiently and accurately translating textual prompts into human motions, focusing on the content only [6, 31, 46]. Integrating the style condition to generate stylized motions remains under-explored." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.612, + 0.788, + 0.777 + ], + "angle": 0, + "content": "Combining these two lines of research is a straightforward approach to tackle stylized motion generation, where a motion style transfer method [1, 45] can be applied to each motion sequence generated by a text-driven motion diffusion model [6, 31, 46]. However, in addition to the aforementioned inefficiency issue, this approach has two more limitations. First, error may accumulate across the pipeline. As, motion style transfer methods are usually trained with high-quality real-world motion sequences, we empirically observe that their performance may significantly degrade for imperfect motions produced by text-to-motion techniques. Second, existing motion style transfer methods rely on specialized style datasets [1, 27, 52] with limited motion content, which restricts their applications to motion diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this paper, we present a novel stylized motion diffusion model, dubbed SMooDi, that customizes a pre-trained text-to-motion model for stylization. Built upon the pre-trained motion latent diffusion model (MLD) [6], SMooDi inherits MLD's ability to generate diverse motion content. At the same time," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.451, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SMooDi: Stylized Motion Diffusion Model" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.147, + 0.788, + 0.433 + ], + "angle": 0, + "content": "SMooDi can generate motions in a variety of styles according to different style reference conditions, as shown in Fig. 1. Our main novelty is the style modulation module, which consists of a style adaptor and a style guidance module. First of all, drawing inspiration from controllable image generation [61], the style adaptor is designed to predict residual features conditioned on style reference motion sequence within each attention layer of MLD. It is useful for incorporating the style condition while ensuring the realism of the generated motion. Second, we design both classifier-free and classifier-based style guidance to more precisely control the stylized motion generation. Specifically, the classifier-free style and content guidance are linearly combined, where we can easily strike a balance between preserving content and reflecting style within the generated motion. At the same time, we design an additional classifier-based style guidance mechanism. It is an analytic function quantifying the disparity between the generated motion and the style reference motion in a style-centric embedding space, whose gradients are subsequently employed to guide the generated motion closer to the intended style. Our style adaptor and guidance module are designed to be complementary, which lead to high-quality stylized motion generation. The two modules are jointly optimized in a feature space instead of sequence-wise separate stitching, thereby avoiding the error accumulation issue." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.459, + 0.788, + 0.567 + ], + "angle": 0, + "content": "Although our approach is primarily designed for stylized motion generation driven by content text, we can utilize DDIM-Inversion [43] to identify the noisy latent corresponding to the content motion sequence. Following the same procedure as for text-driven content, SMooDi is capable of facilitating stylized motion generation based on content motion sequences. In other words, motion style transfer is a downstream application of our approach should it be desired in practice, e.g., to stylize the already created motion sequences." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.589, + 0.788, + 0.697 + ], + "angle": 0, + "content": "Experiments on the HumanML3D [16] and 100STYLE [27] datasets demonstrate that SMooDi surpasses other baseline models in generating stylized motion driven by content text, excelling in both content preservation and style reflection. More importantly, unlike previous methods that require individual fine-tuning for each style [12, 30, 54], SMooDi successfully integrates diverse content from the HumanML3D dataset and various styles from the 100STYLE dataset into a single model without requiring additional tuning during inference." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To summarize, our contributions are: (1) To our knowledge, SMooDi is the first approach that adapts a pre-trained text-to-motion model to generate diverse stylized motion. (2) We introduce a novel style modulation module that utilizes a stylized adaptor and a style classifier guidance to enable stylized motion generation while ensuring style reflection, content preservation, and realism. (3) Experiments demonstrate that SMooDi not only sets a new state of the art in stylized motion generation driven by content text but also achieves performance comparable to state-of-the-art methods in motion style transfer." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Zhong et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.388, + 0.161 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.173, + 0.486, + 0.187 + ], + "angle": 0, + "content": "2.1 Human Motion Generation" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.194, + 0.788, + 0.496 + ], + "angle": 0, + "content": "Human motion generation has attracted great attention [4, 5, 7, 9, 14, 16, 20, 32, 34-36, 47, 50, 51, 57, 58, 63]. Inspired by the impressive performance of diffusion models in image generation, a lot of works [6, 8, 13, 18, 22, 23, 25, 29, 32, 33, 39, 41, 46, 48, 53, 56, 60, 62, 64] utilize diffusion models to generate human motion. MDM [46] facilitates high-quality generation and versatile conditioning, providing a solid baseline for novel motion generation tasks. MLD [6] minimizes computational overhead during both training and inference by establishing the diffusion process within the latent space. Driven by the efficacy of diffusion models for control and conditioning, several studies have leveraged pre-trained motion diffusion models to generate long-sequence motions [41], enable human-object interactions [29], and control the joint trajectory of generated motions [23, 53]. However, there is no work exploring how to leverage pre-trained motion diffusion models to generate diverse stylized motion. While some studies [2, 3, 37] have enabled stylized motion generation in their diffusion pipeline, their methods are trained from scratch, and the supported styles are restricted by their motion content dataset. It is challenging for them to simultaneously support diverse motion content and style. In this work, we build upon a pre-trained motion diffusion model, MLD, and explore how to fine-tune it on a larger motion style dataset, 100STYLE, to learn diverse motion styles while retaining the ability to support motion generation across a wide range of content." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.515, + 0.444, + 0.53 + ], + "angle": 0, + "content": "2.2 Motion Style Transfer" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.537, + 0.788, + 0.763 + ], + "angle": 0, + "content": "Recently, motion style transfer has seen quality enhancements through the adoption of various advanced neural architectures and generative models, such as graph neural networks [28], time-series models [27, 45], normalizing flows [49], and diffusion models [3,37]. Specifically, Aberman et al. [1] designed a two-branch generative adversarial network to disentangle motion style from content and facilitate their re-composition. Their approach effectively breaks the constraint of requiring a paired motion dataset. Motion Puzzle [19] realizes a framework that can control the style of individual body parts. Above methods extract both content and style features from the motion sequence. Moreover, Guo et al. [15] leverage the latent space of pre-trained motion models to enhance the extraction and infusion of motion content and style. However, a major limitation of these models is their reliance on specialized style datasets [1, 52] with limited motion content, which restricts their applications. In this work, we customize a pre-trained text-to-motion model for stylization, thus inheriting its ability to generate diverse motion content." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.782, + 0.569, + 0.8 + ], + "angle": 0, + "content": "3 Stylized Motion Diffusion Model" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In this section, we introduce our proposed SMooDi for incorporating style conditions from a style motion sequence into a content-oriented pre-trained motion" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.451, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SMooDi: Stylized Motion Diffusion Model" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.255, + 0.151, + 0.756, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.289, + 0.788, + 0.371 + ], + "angle": 0, + "content": "Fig. 2: Overview of SMooDi. Our model generates stylized human motions from content text and a style motion sequence. At the denoising step \\( t \\), our model takes the content text \\( \\mathbf{c} \\), style motion \\( \\mathbf{s} \\), and noisy latent \\( \\mathbf{z}_t \\) as input and predicts \\( \\epsilon_t \\), which is then transferred to \\( \\mathbf{z}_{t-1} \\). This denoising step is repeated \\( T \\) times to obtain the noise-free motion latent \\( \\mathbf{z}_0 \\), which is fed into a motion decoder \\( D \\) to produce the stylized motion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.404, + 0.789, + 0.585 + ], + "angle": 0, + "content": "diffusion model (MLD [6]). Fig. 2 presents an overview of SMooDi. Following the setting in MLD [6], we place the diffusion process in the motion latent space. Let \\(\\epsilon_{\\theta}\\) denote the latent denoiser (a UNet parameterized by \\(\\theta\\)), and \\(\\{z_t\\}_{t=0}^T\\) denote the sequence of noisy latents, where \\(z_T\\) is a Gaussian noise. Given a content prompt \\(\\mathbf{c}\\) and a style prompt \\(\\mathbf{s}\\), we define \\(\\epsilon_t = \\epsilon_{\\theta}(z_t, t, \\mathbf{c}, \\mathbf{s})\\) for the denoising at step \\(t\\) (\\(0 < t \\leq T\\)). A cleaner noisy latent \\(z_{t-1}\\) can be obtained by subtracting \\(\\epsilon_t\\) from \\(z_t\\). The denoising step is repeated \\(T\\) iterations until a clean latent \\(z_0\\) is obtained. It can then be decoded by a motion decoder \\(\\mathbf{D}\\) into a realistic motion sequence \\(\\mathbf{x} \\in \\mathbb{R}^{N \\times H}\\) that accurately reflects both the content and style conditions. Here, \\(N\\) represents the length of the motion sequence, and \\(H\\) is the dimension of human motion representations. We employ the same motion representations as in HumanML3D [16], where \\(H = 263\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.587, + 0.789, + 0.708 + ], + "angle": 0, + "content": "As shown in the Fig. 2, the content prompt is a text description, and the style prompt is provided by a reference style motion sequence \\(\\mathbf{s} \\in \\mathbb{R}^{N \\times H}\\). In this section, we focus on using a text description as the content prompt \\(\\mathbf{c}\\) to explain our proposed stylized motion diffusion model. By employing the DDIM-Inversion [43] to identify the noisy latent corresponding to a motion content sequence, we can effectively use motion sequences as content prompts to generate stylized motions. In other words, motion style transfer is a downstream application of our proposed approach." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.709, + 0.787, + 0.739 + ], + "angle": 0, + "content": "Our proposed stylization module consists of a style adaptor and a style guidance module. We will explain them separately in the rest of this section." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.765, + 0.381, + 0.781 + ], + "angle": 0, + "content": "3.1 Style Adaptor" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Although LoRA has been successfully used to incorporate \"style\" into the models in image domain [21, 42], they typically require training a separate LoRA for each style. In contrast, we focus on fine-tuning the model just once to adapt to" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Zhong et al." + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.144, + 0.778, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.266, + 0.789, + 0.325 + ], + "angle": 0, + "content": "Fig. 3: Detailed illustration of our proposed style adaptor. The style adaptor is connected to the motion diffusion model via zero linear layer. The output of the style adaptor from each Transformer encoder is added to the motion diffusion model to steer the predicted noise towards the target style." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.354, + 0.784, + 0.399 + ], + "angle": 0, + "content": "various motion styles, where adapting ControlNet [61] is more suitable. Therefore, we design a content-aware style adaptor based on ControlNet. This adaptor incorporates the motion style condition into the pre-trained MLD [6]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.4, + 0.789, + 0.611 + ], + "angle": 0, + "content": "Instead of learning to disentangle motion style from content from scratch on a large motion dataset, we redirect our focus towards capturing the motion style while ensuring the preservation of diverse motion content within the pre-trained MLD framework. Specifically, it consists of a trainable copy of the Transformer Encoder from the latent diffusion model in MLD. The architecture of the style adaptor is illustrated in Fig. 3. An independent style encoder is utilized to extract the style embedding from the style motion sequence \\(\\mathbf{s}\\). The style adaptor takes the same content prompt \\(\\mathbf{c}\\), the noised latent \\(\\mathbf{z}_t\\) and timestep \\(t\\) as in MLD, and the extracted style embedding. Each Transformer layer in the original latent diffusion model and the style adaptor is connected via a linear layer, with both weight and bias initially set to zeros. As training progresses, the style adaptor learns the style constraints and gradually applies the learned feature corrections to the corresponding layers in the latent diffusion model, thereby implicitly steering the output towards the desired style." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.636, + 0.388, + 0.65 + ], + "angle": 0, + "content": "3.2 Style Guidance" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.663, + 0.788, + 0.739 + ], + "angle": 0, + "content": "The style adaptor alone may not be sufficient to successfully incorporate the style condition. We further leverage both the classifier-free and classifier guidance to further enhance the stylization of a motion diffusion model. The combination of two types of guidance effectively ensures that generated motion meets multiple constraints while maintaining realism, complementing the style adaptor." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.74, + 0.787, + 0.77 + ], + "angle": 0, + "content": "Classifier-free Style Guidance. With the introduction of an extra style condition, we can divide the conditioned classifier-free guidance into two parts." + }, + { + "type": "equation", + "bbox": [ + 0.226, + 0.783, + 0.435, + 0.799 + ], + "angle": 0, + "content": "\\[\n\\epsilon_ {\\theta} \\left(\\boldsymbol {z} _ {t}, t, \\mathbf {c}, \\mathbf {s}\\right) = \\epsilon_ {\\theta} \\left(\\boldsymbol {z} _ {t}, t, \\emptyset , \\emptyset\\right) +\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.284, + 0.802, + 0.786, + 0.826 + ], + "angle": 0, + "content": "\\[\n\\underbrace {w _ {c} (\\epsilon_ {\\theta} (\\boldsymbol {z} _ {t} , t , \\mathbf {c} , \\emptyset) - \\epsilon_ {\\theta} (\\boldsymbol {z} _ {t} , t , \\emptyset , \\emptyset))} + \\underbrace {w _ {s} (\\epsilon_ {\\theta} (\\boldsymbol {z} _ {t} , t , \\mathbf {c} , \\mathbf {s}) - \\epsilon_ {\\theta} (\\boldsymbol {z} _ {t} , t , \\mathbf {c} , \\emptyset))} _ {\\text {(1)}}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.304, + 0.826, + 0.722, + 0.836 + ], + "angle": 0, + "content": "Classifier-free Content Guidance Classifier-free Style Guidance" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.451, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SMooDi: Stylized Motion Diffusion Model" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.357 + ], + "angle": 0, + "content": "where \\( w_{c} \\) and \\( w_{s} \\) represent the strengths of the classifier-free guidance for the condition \\( \\mathbf{c} \\) and \\( \\mathbf{s} \\), respectively. We slightly abuse the notations here by using \\( \\emptyset \\) to denote a condition is not used. The classifier-free content guidance is the same as in MLD, which can facilitate the text-to-motion generation process in combination with the first term's unconditioned guidance. Our proposed classifier-free style guidance works in a similar way. Note that \\( \\epsilon_{\\theta}(z_t,t,c,s) \\) is MLD model with the style adaptor incorporated introduced in the previous section, which takes both a textual prompt and style motion sequence as input. By contrasting the text-driven denoising output with and without the style condition, it can highlight the effectiveness of the style input \\( s \\) and facilitate the generation of stylized motion driven by content text. Our insight here is that by dividing the conditioned guidance into content and style components separately, we can easily strike a balance between preserving content and reflecting style within the generated motion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.359, + 0.788, + 0.464 + ], + "angle": 0, + "content": "To better understand the classifier-free content and style guidance, we visualize each of them through decoding denoised latent \\( \\mathbf{z}_0 \\) into the motion space. As illustrated in Fig.4(a), the content guidance ensures the motion generation is faithful to the textual prompt, while the style guidance, as shown in Fig.4(b), emphasizes style-related characteristics in the output. Combining both forms of guidance results in a stylized motion that adheres to both content and style conditions, as illustrated in Fig. 4(c)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.465, + 0.788, + 0.572 + ], + "angle": 0, + "content": "Classifier-based Style Guidance. To further improve the stylization of a motion diffusion model, we adopt the classifier guidance [10,59] to provide stronger guidance to the generated motion towards the desired style. The core of our classifier-based style guidance is a novel analytic function \\( G(\\mathbf{z}_t,t,\\mathbf{s}) \\), which calculates the \\( L_{1} \\) distance between the style embedding of the generated clean motion \\( \\hat{\\pmb{x}}_0 \\) at denoising step \\( t \\) and the reference style motions \\( \\mathbf{s} \\). The function's gradient is utilized to steer the generated motion towards the desired style." + }, + { + "type": "equation", + "bbox": [ + 0.342, + 0.586, + 0.664, + 0.6 + ], + "angle": 0, + "content": "\\[\n\\epsilon_ {\\boldsymbol {\\theta}} (\\boldsymbol {z} _ {t}, t, \\mathbf {c}, \\mathbf {s}) = \\epsilon_ {\\boldsymbol {\\theta}} (\\boldsymbol {z} _ {t}, t, \\mathbf {c}, \\mathbf {s}) + \\tau \\nabla_ {\\boldsymbol {z} _ {t}} G (\\boldsymbol {z} _ {t}, t, \\mathbf {s}),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.358, + 0.603, + 0.785, + 0.62 + ], + "angle": 0, + "content": "\\[\nG \\left(\\boldsymbol {z} _ {t}, t, \\mathbf {s}\\right) = | f \\left(\\hat {\\boldsymbol {x}} _ {0}\\right) - f (\\mathbf {s}) |, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.632, + 0.788, + 0.679 + ], + "angle": 0, + "content": "where \\(\\tau\\) adjusts the strength of reference-style guidance and \\(f\\) denotes the style feature extractor. The generated motion \\(\\hat{\\pmb{x}}_0\\) is obtained by first converting the denoising output latent \\(\\pmb{z}_t\\) into the predicted clean latent as shown below:" + }, + { + "type": "equation", + "bbox": [ + 0.389, + 0.69, + 0.786, + 0.723 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {z}} _ {0} = \\frac {\\boldsymbol {z} _ {t} - \\sqrt {1 - \\alpha_ {t}} \\varepsilon_ {\\theta} (\\boldsymbol {z} _ {t} , t , \\boldsymbol {c} , \\boldsymbol {s})}{\\sqrt {\\alpha_ {t}}}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.787, + 0.779 + ], + "angle": 0, + "content": "where \\(\\alpha_{t}\\) denotes the pre-defined noise scale in the forward process of the diffusion model. The predicted clean latent \\(\\hat{z}_0\\) is then input into the motion decoder \\(\\mathbf{D}\\) to obtain the generated motion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We obtain the style feature extractor by training a style classifier on the 100STYLE dataset [27] and removing its final layer. We refer the readers to more details in the supplementary materials. The training of the style feature extractor with ground-truth style labels for supervision enables it to effectively" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.117, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Zhong et al." + }, + { + "type": "image_caption", + "bbox": [ + 0.331, + 0.15, + 0.631, + 0.164 + ], + "angle": 0, + "content": "Text: A person walks forward and then sits down." + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.165, + 0.756, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.246, + 0.788, + 0.304 + ], + "angle": 0, + "content": "Fig. 4: Visual illustrations of the classifier-free and classifier-based style guidance. (a) and (b) respectively show the classifier-free content and style guidance; (c) displays the initial stylized motion resulting from the combination of (a) and (b); (d) illustrates the refined stylized motion modified by the classifier-based style guidance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.331, + 0.784, + 0.36 + ], + "angle": 0, + "content": "capture style-related features. Therefore, style classifier guidance can provide more guidance to the stylized motion generation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.361, + 0.785, + 0.526 + ], + "angle": 0, + "content": "Combination of the Two Style Guidance. The classifier-free and classifier-based style guidance are designed to complement each other, each playing a vital role in accurately reflecting the target style in the generated motions. As illustrated in Fig. 4, the desired style motion is \"arms open wide to the sides like an airplane\". The classifier-free style guidance (Fig. 4(b)) can capture style-related characteristic in a reasonably accurate manner. When combined with the classifier-free content guidance, it depicts the desired style (Fig.4(c)). Refined further by the classifier-based style guidance, the stylization is more authentic, where the person's harms are more open (Fig.4(d)). In addition to such visual results, quantitative ablation studies also verify the effectiveness of our proposed both classifier-free and classifier-based style guidance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.527, + 0.789, + 0.678 + ], + "angle": 0, + "content": "At the same time, although classifier-based style guidance offers precise style control, its effectiveness may be compromised when the content text significantly diverges from locomotion-related movements. This is because the style feature function is trained solely on the 100STYLE dataset, which contains only such movements. Over-reliance on style classifier guidance risks producing motions that fail to execute the desired actions, leading to unrealistic and physically implausible movements. Therefore, we leverage a content-aware style adaptor that establishes the fundamental style direction, while style classifier guidance refines this base for a more precise outcome. The effectiveness of this design is verified in our ablation studies." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.698, + 0.406, + 0.714 + ], + "angle": 0, + "content": "3.3 Learning Scheme" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.722, + 0.785, + 0.768 + ], + "angle": 0, + "content": "Following [61], a straightforward approach to train SMooDi is to freeze the parameters of MLD and solely train the style adaptor on the 100STYLE dataset using the following loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.378, + 0.776, + 0.787, + 0.803 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s t d} = \\mathbb {E} _ {\\epsilon , \\mathbf {z}} \\left[ \\| \\epsilon_ {\\theta} (\\mathbf {z} _ {t}, t, \\mathbf {c}, \\mathbf {s}) - \\epsilon \\| _ {2} ^ {2} \\right], \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.789, + 0.842 + ], + "angle": 0, + "content": "where \\(\\epsilon \\sim \\mathcal{N}(0, \\mathbf{I})\\) represents the ground-truth noise added to \\(z_0\\). In our experiments, however, we found that this loss function alone leads to an issue of" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.451, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SMooDi: Stylized Motion Diffusion Model" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.222 + ], + "angle": 0, + "content": "\"content-forgetting\", where the model progressively looses the MLD's ability to generate motions with diverse contents. To address this issue, we design a content prior preservation loss \\( L_{pr} \\). Specifically, we randomly sample motions from the HumanML3D dataset to compute a prior preservation loss when fine-tuning SMooDi on the 100STYLE dataset." + }, + { + "type": "equation", + "bbox": [ + 0.373, + 0.231, + 0.786, + 0.256 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {p r} = \\mathbb {E} _ {\\epsilon^ {\\prime}, \\boldsymbol {z} ^ {\\prime}} \\left[ \\| \\epsilon_ {\\theta} (\\boldsymbol {z} _ {t} ^ {\\prime}, t, \\mathbf {c} ^ {\\prime}, \\mathbf {s} ^ {\\prime}) - \\epsilon^ {\\prime} \\| _ {2} ^ {2} \\right], \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.265, + 0.789, + 0.37 + ], + "angle": 0, + "content": "where \\(\\mathbf{z}_t^\\prime\\), \\(\\mathbf{c}^\\prime\\) and \\(\\mathbf{s}^\\prime\\) represents the motion latent, content prompt and style motion sequence derived from the HumanML3D dataset. \\(\\epsilon^\\prime\\) is the noise map added to \\(\\mathbf{z}_0^\\prime\\). A similar solution is used in DreamBooth [40] to solve the \"language drift\" problem, where images generated from the frozen pretrained image generation model are utilized to enforce a class-prior preservation loss during model finetuning. Our content preservation loss can effectively mitigate content forgetting while learning diverse motion styles from the 100STYLE dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.371, + 0.789, + 0.491 + ], + "angle": 0, + "content": "To further encourage the style adaptor to focus on motion style, while also ensuring that the latent diffusion model in MLD handles motion content well, we introduce an additional cycle prior-preservation loss, inspired by [19,55]. Specifically, we start this process by randomly sampling content text and motion style sequences from both the 100STYLE and HumanML3D datasets simultaneously. Then, we intermix the content text and motion style from these sequences with each other. Finally, we repeat this process to reconstruct the original motion sequences. The formula is expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.271, + 0.502, + 0.786, + 0.528 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c y c} = \\mathbb {E} _ {\\boldsymbol {z}, \\boldsymbol {z} ^ {\\prime}, \\epsilon , \\epsilon^ {\\prime}} \\left[ \\left\\| \\epsilon_ {\\theta} (\\boldsymbol {z} _ {t} ^ {s h}, t, \\mathbf {c}, \\boldsymbol {s} ^ {h s}) + \\epsilon_ {\\theta} (\\boldsymbol {z} _ {t} ^ {h s}, t, \\mathbf {c} ^ {\\prime}, \\boldsymbol {s} ^ {s h}) - \\epsilon - \\epsilon^ {\\prime} \\right\\| _ {2} ^ {2} \\right], \\qquad (6)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.788, + 0.672 + ], + "angle": 0, + "content": "where \\( s^{hs} \\) denotes the motion sequence created by merging content from the HumanML3D dataset with style from the 100STYLE dataset. Similarly, \\( s^{sh} \\) represents the sequence where content is sourced from the 100STYLE dataset and style from the HumanML3D dataset. The noised latent codes \\( z_{t}^{sh} \\) and \\( z_{t}^{hs} \\) correspond to \\( s^{sh} \\) and \\( s^{hs} \\), respectively. Essentially, the cycle prior-preservation loss exchanges diverse content and style between two datasets, encouraging the content text to remain invariant in the generated motion under forward and backward translation. Overall, the training loss function of our framework is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.389, + 0.675, + 0.785, + 0.691 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a l l}} = \\mathcal {L} _ {\\text {s t d}} + \\lambda_ {p r} \\mathcal {L} _ {p r} + \\lambda_ {\\text {c y c}} \\mathcal {L} _ {\\text {c y c}} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.696, + 0.788, + 0.727 + ], + "angle": 0, + "content": "where \\(\\lambda_{pr}\\) and \\(\\lambda_{cyc}\\) are hyperparameters. We refer readers to the pseudocode and illustration for training in the supplementary material for more details." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.749, + 0.377, + 0.766 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We conduct experiments on both stylized text2motion and motion style transfer to demonstrate the effectiveness of our framework. Both tasks use a motion sequence as a style prompt, with the primary difference being their content prompt input: the former utilizes text, while the latter relies on motion sequences." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Zhong et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.188 + ], + "angle": 0, + "content": "Table 1: Comparison with baseline methods on stylized motion generation driven by content text, using a combination of the 100STYLE (providing style) and HumanML3D datasets (providing content)." + }, + { + "type": "table", + "bbox": [ + 0.263, + 0.199, + 0.742, + 0.284 + ], + "angle": 0, + "content": "
MethodFID ↓ Foot skatingMM Dist ↓ R-precision ↑ Diversity → SRA ↑ ratio ↓
Ours1.6090.1244.4770.5719.23572.418
MLD+Motion Puzzle [19]6.1270.1856.4670.2906.476263.769
MLD+Aberman et al. [1]3.3090.3475.9830.4068.81654.367
ChatGPT+MLD0.6140.1314.3130.6058.8364.819
" + }, + { + "type": "image", + "bbox": [ + 0.25, + 0.304, + 0.361, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.362, + 0.329, + 0.444, + 0.34 + ], + "angle": 0, + "content": "A person kicks." + }, + { + "type": "image", + "bbox": [ + 0.461, + 0.308, + 0.532, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.551, + 0.309, + 0.642, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.655, + 0.309, + 0.73, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.46, + 0.361, + 0.641, + 0.369 + ], + "angle": 0, + "content": "(a) \\(\\mathrm{MLD + }\\) Motion Puzzle (b) \\(\\mathrm{MLD + }\\) Aberman et al." + }, + { + "type": "image_caption", + "bbox": [ + 0.669, + 0.361, + 0.702, + 0.369 + ], + "angle": 0, + "content": "(c) Ours" + }, + { + "type": "image", + "bbox": [ + 0.25, + 0.372, + 0.355, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.372, + 0.445, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.462, + 0.372, + 0.551, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.564, + 0.372, + 0.644, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.651, + 0.372, + 0.738, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.45, + 0.789, + 0.478 + ], + "angle": 0, + "content": "Fig. 5: Qualitative comparisons of our approach and baseline methods on two stylized motion generation task." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.789, + 0.734 + ], + "angle": 0, + "content": "Datasets. We utilize the HumanML3D dataset [16] as our motion content dataset and the 100STYLE dataset [27] as our motion style dataset. The HumanML3D dataset is the largest motion capture dataset, featuring text annotations and comprising 14,646 motions and 44,970 motion annotations. Following the processing approach outlined in [16], we preprocess the HumanML3D dataset to obtain consistent motion representations. On the other hand, the 100STYLE dataset [27], being the largest motion style dataset, comprises up to 1,125 minutes of motion sequences, showcasing a wide array of 100 diverse locomotion styles. Due to differences in skeletons between the 100STYLE dataset and HumanML3D, we retarget the motions from 100STYLE to match the HumanML3D (SMPL-H) skeleton. Following this alignment, we apply the same processing steps as used for the HumanML3D dataset to preprocess the 100STYLE dataset. Moreover, as the 100STYLE dataset lacks text descriptions, we leverage MotionGPT [20] to generate pseudo text descriptions for the motion sequences in the 100STYLE dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Evaluation metrics are designed to assess three dimensions: Content Preservation, Style Reflection, and Realism. For content preservation and style reflection assessment, we employ metrics consistent with those used in [6]: motion-retrieval precision (R precision), Multi-modal Distance (MM Dist), Diversity, and Frechet Inception Distance (FID). Additionally, recognizing the common foot skating issues in kinetics-based motion generation methods, we incorporate the foot skating ratio metric proposed by [23] into our motion quality evaluation. For style" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.451, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SMooDi: Stylized Motion Diffusion Model" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.328 + ], + "angle": 0, + "content": "reflection, we employ Style Recognition Accuracy (SRA) [19]. During evaluation, we randomly select a content text from the HumanML3D dataset and a motion style sequence from the 100STYLE dataset to generate the stylized motion. We then use a pre-trained style classifier to compute the SRA for the generated motion. It's noteworthy that some motion style labels in the 100STYLE dataset, like 'kick' and 'jump,' inherently convey motion content, which may conflict with the content text in HumanML3D dataset. To address this, we categorize the motion style labels into server groups following the approach by Kim et al. [24], Specifically excluding the 'ACT' group ensures that only motion style labels not conflicting with motion content are considered when computing the SRA metric. Further details about the evaluation metrics are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.33, + 0.789, + 0.691 + ], + "angle": 0, + "content": "Baselines. For motion style transfer task, we compare our methods with two state-of-the-art methods, namely Motion Puzzle [19] and Aberman et al. [1]. To ensure a fair comparison, we train the compared methods under the same settings as ours, using a combined dataset comprising HumanML3D and 100STYLE. Due to the constraints of the multi-class discriminator in Aberman et al., which requires style labels, we adopt the training method outlined in Motion Puzzle to eliminate the need for style labels. For stylized text2motion task, we compare our method against baselines capable of generating stylized motion from content text and style motion sequences. The straightforward baselines involve applying motion style transfer methods to the motion sequences generated by the text2motion model. To align with our approach that uses a pre-trained motion diffusion model, the text2motion models in the baselines for stylized motion generation select MLD, and the motion style transfer methods are consistent with those used in the motion style transfer task. For the stylized text2motion task, we compare our method against two kinds of baselines capable of generating stylized motion from content text and style motion sequences. The first kind of baseline involves applying motion style transfer methods to the motion sequences generated by the text2motion model. To align with our approach that uses a pretrained motion diffusion model, the text2motion models select MLD [6], and the motion style transfer methods are consistent with those used in the motion style transfer task. The second kind of baseline involves using ChatGPT to merge style labels from 100STYLE with text from HumanML3D into a sentence. For example, given the content text 'a person walks.' and the style label 'old,' we obtain 'an elderly person walks.' This merged sentence is then fed to MLD." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.711, + 0.536, + 0.726 + ], + "angle": 0, + "content": "4.1 Comparison to Baseline Methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.784, + 0.763 + ], + "angle": 0, + "content": "Quantitative and Qualitative For the task of stylized text2motion, Table 1 reports the comparisons of our method with the three baseline methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.839 + ], + "angle": 0, + "content": "As shown in the \\(3^{rd}\\) row of Table 1, ChatGPT+MLD only achieves around \\(5.29\\%\\) in terms of SRA, indicating that MLD cannot enable stylized generation from text alone, even though it contains style descriptions. Notably, our method outperforms the two baselines that combine MLD with motion style transfer methods in all metrics." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Zhong et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.267, + 0.145, + 0.735, + 0.16 + ], + "angle": 0, + "content": "Table 2: Comparison with baseline methods on motion style transfer." + }, + { + "type": "table_caption", + "bbox": [ + 0.244, + 0.184, + 0.478, + 0.197 + ], + "angle": 0, + "content": "(a) Evaluation on HumanML3D dataset" + }, + { + "type": "table", + "bbox": [ + 0.239, + 0.207, + 0.482, + 0.272 + ], + "angle": 0, + "content": "
MethodFoot skating ratio ↓
Ours0.0951.58265.147
Motion Puzzle [19]0.1976.87167.233
(Aberman et al [1])0.3383.89261.006
" + }, + { + "type": "table_caption", + "bbox": [ + 0.556, + 0.184, + 0.731, + 0.196 + ], + "angle": 0, + "content": "(b) Evaluation on Xia dataset" + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.207, + 0.789, + 0.271 + ], + "angle": 0, + "content": "
MethodFoot skating ratio↓FID↓SRA↑(%)CRA↑(%)
Ours0.03174.66361.11145.555
Motion Puzzle [19]0.03165.36067.77825.556
(Aberman et al [1])0.02605.68156.66734.444
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.306, + 0.785, + 0.427 + ], + "angle": 0, + "content": "Specifically, our method performs better than MLD+Motion Puzzle in the SRA metric by \\(13.56\\%\\) and significantly outperforms MLD+Aberman et al. in the FID metric by \\(51.38\\%\\) and \\(0.64\\%\\) in the R-precision metric. The first row of Fig. 5 validates our observation, where the motion generated by our method performs better in adhering to both content and style constraints than baseline methods. In contrast, MLP+Aberman et al. can successfully perform the action but fail to reflect the motion style in Fig.5(b), while Motion Puzzle can accurately reflect the motion style but struggles to effectively perform the action in Fig.5(a)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.788, + 0.791 + ], + "angle": 0, + "content": "For the task of motion style transfer, since it does not take content text as input, text-motion related metrics such as MM Dist, R-precision, and Diversity are not applicable and thus are not reported. Part (a) of Table 2 presents a comparison between our method and the two baseline methods, using the HumanML3D dataset as the motion content source and drawing motion styles from the 100STYLE dataset. Our method delivers competitive results in the SRA metric and excels in the FID and foot skating ratio metrics. Specifically, we see a substantial \\(59.35\\%\\) improvement in the FID metric over Aberman et al. [1]. To more effectively compare the generalizability of different methods, we conduct experiments on the Xia dataset [52], a small, specific motion style dataset that was unseen by our and the baseline models during training. Because motion content labels are present in the Xia dataset, we report the Content Recognition Accuracy (CRA). Part (b) of Table 2 showcases the results. Our method maintains competitive performance in the SRA metric, with only a marginal \\(9.83\\%\\) decrease in SRA compared to Motion Puzzle. On the contrary, our method exhibits a significant \\(32.26\\%\\) increase in the CAR metric relative to Aberman et al., and a notable \\(78.26\\%\\) enhancement over Motion Puzzle. Our method achieves a better balance between style reflection and content preservation. The second row of Fig. 5 validates this observation. It is worth noting that, unlike other motion style transfer methods, our method does not incorporate objectives for enabling stylized motion generation using motion content sequences. Through simple DDIM-Inversion and without any additional optimization or regularization, our method achieves performance comparable to existing motion style transfer methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.84 + ], + "angle": 0, + "content": "User Study. Due to the highly subjective nature of stylized motion, we conduct User studies using pairwise comparisons to further evaluate our proposed method in the tasks of stylized motion generation and motion style transfer. We" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.451, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SMooDi: Stylized Motion Diffusion Model" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.246, + 0.145, + 0.756, + 0.16 + ], + "angle": 0, + "content": "Table 3: Ablation Studies on HumanML3D Content and 100STYLE Styles." + }, + { + "type": "table", + "bbox": [ + 0.29, + 0.171, + 0.716, + 0.267 + ], + "angle": 0, + "content": "
MethodFID↓ ratio↓MM Dist↓R-precision↑ (Top-3)Diversity→ SRA(%)↑
Ours (on all)1.6090.1244.4770.5719.23572.418
w/o Lcyc2.0460.1364.4650.5698.86964.866
w/o Lpr + Lcyc5.9960.1666.0980.3357.45681.841
w/o classifier-based1.0500.1114.0850.6309.44520.245
w/o adaptor2.9840.1234.5260.5508.37269.952
" + }, + { + "type": "table_caption", + "bbox": [ + 0.338, + 0.282, + 0.619, + 0.295 + ], + "angle": 0, + "content": "Text: A person walks forward and then sits down." + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.295, + 0.747, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.257, + 0.372, + 0.746, + 0.387 + ], + "angle": 0, + "content": "Fig. 6: Visual comparisons of the ablation designs and our full model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.416, + 0.788, + 0.522 + ], + "angle": 0, + "content": "recruited 22 human subjects to participate in the study. In each test, participants are presented with two 4-second video clips synthesized by our method and one comparison method. They are then required to select their preferred clip while considering Realism, Style Reflection, and Content Preservation dimensions, respectively. As shown in Fig. 7, our method receives more user appreciation compared to two baselines across three dimensions in two tasks. Further user study details are provided in supplementary." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.544, + 0.399, + 0.558 + ], + "angle": 0, + "content": "4.2 Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.787, + 0.629 + ], + "angle": 0, + "content": "To validate the effectiveness of our framework's design choices, we have conducted several ablation studies: the first assesses the impact of each loss function term, while the second evaluates the influence of the style adaptor and style guidance during sampling." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.788, + 0.734 + ], + "angle": 0, + "content": "Loss Components. Firstly, we exclude the cycle-prior term in the loss function, denoted as \\( w / o \\) \\( L_{cyc} \\). Comparing the results in the \\( 1^{\\mathrm{st}} \\) and \\( 2^{\\mathrm{nd}} \\) rows in Table 3, we observe that our full model outperforms in all content preservation and style reflection metrics. The motion generated by our approach can still perform the content adhering to the text description but performs worse in accurately reflecting the motion style, as reflected by the arms not being fully extended horizontally." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.789, + 0.843 + ], + "angle": 0, + "content": "Since the cycle prior-preservation term is built upon the prior-preservation term, it is meaningless to exclude \\( L_{pr} \\) while retaining \\( L_{cyc} \\). Therefore, we further exclude both the prior-preservation and cycle-prior term, denoted as \\( w / oL_{pr} + L_{cyc} \\) in the \\( 3^{\\mathrm{rd}} \\) row. By comparing the results in the \\( 2^{\\mathrm{nd}} \\) and \\( 3^{\\mathrm{rd}} \\) rows of Table 3, we notice that while the number of SRAs is higher in the third row, other metrics show a significant decline. Specifically, in terms of the FID metric, performance deteriorates by more than \\( 229\\% \\). Indeed, without \\( L_{pr} \\), the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Zhong et al." + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.147, + 0.496, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.147, + 0.725, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.293, + 0.284, + 0.706, + 0.298 + ], + "angle": 0, + "content": "Fig. 7: User Study on two stylized motion generation tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.326, + 0.785, + 0.401 + ], + "angle": 0, + "content": "model tends to lose the ability to translate content text into corresponding motion, a phenomenon named 'content-forgetting' as described in Sec. 3.3. Fig. 6(a) validates our observation, showing that the content in the generated motion significantly deviates from the text descriptions and closely resembles the style motion sequence." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.413, + 0.785, + 0.563 + ], + "angle": 0, + "content": "Style Adaptor and Style Guidance. Initially, we compare our model to a variant without the classifier-based guidance, \\( w / o \\) classifier-based, to demonstrate its effectiveness. The \\( 5^{\\text{th}} \\) row of Table 3 presents the results. Consistent with the findings in [19], a reasonable trade-off between content preservation and style reflection is observed. Although classifier-based style guidance may slightly affect the content preservation metrics, it significantly boosts the model's performance in the SRA metric, yielding an impressive \\( 208\\% \\) improvement.Fig. 6(d) demonstrates that, without classifier-based style guidance, the generated motions can reflect the motion style, yet they still fall short of fully achieving the target style. Classifier-based style guidance can effectively bridge this gap." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.565, + 0.787, + 0.685 + ], + "angle": 0, + "content": "Subsequently, as shown in Table 3, we evaluate a variant without the style adaptor, denoted as \\( w / o \\) adaptor (the last row). In cases where the SRA values are close, the style adaptor improved the FID metric by about \\( 80.46\\% \\). Fig. 6(c) shows that the generated motion can greatly perform the 'walk' action while successfully reflecting the style, but fails to perform the 'sit' action. This indicates that the effectiveness of classifier-based style guidance diminishes when the content text deviates from locomotion-related movements. Relying solely on it may even adversely affect action performance." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.706, + 0.36, + 0.721 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In this work, we introduce the Stylized Motion Diffusion Model, a novel approach that leverages a pre-trained motion diffusion model to facilitate stylized motion generation driven by content text. By integrating a style adaptor and style classifier guidance, our method is capable of producing realistic human motions that accurately reflect both the content text descriptions and the desired motion style from motion sequences. Through detailed ablation studies, we have demonstrated the effectiveness of each component in our framework." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.451, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "SMooDi: Stylized Motion Diffusion Model" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.179, + 0.785, + 0.207 + ], + "angle": 0, + "content": "1. Aberman, K., Weng, Y., Lischinski, D., Cohen-Or, D., Chen, B.: Unpaired motion style transfer from video to animation. TOG (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.208, + 0.785, + 0.235 + ], + "angle": 0, + "content": "2. Alexanderson, S., Nagy, R., Beskow, J., Henter, G.E.: Listen, denoise, action! audio-driven motion synthesis with diffusion models. TOG (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.236, + 0.785, + 0.263 + ], + "angle": 0, + "content": "3. Ao, T., Zhang, Z., Liu, L.: Gesturediffuclip: Gesture diffusion model with clip latents. TOG (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.264, + 0.785, + 0.292 + ], + "angle": 0, + "content": "4. Cen, Z., Pi, H., Peng, S., Shen, Z., Yang, M., Zhu, S., Bao, H., Zhou, X.: Generating human motion in 3d scenes from text descriptions. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.293, + 0.785, + 0.333 + ], + "angle": 0, + "content": "5. Chen, L.H., Lu, S., Zeng, A., Zhang, H., Wang, B., Zhang, R., Zhang, L.: Motionllm: Understanding human behaviors from human motions and videos. ArXiv (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.334, + 0.785, + 0.362 + ], + "angle": 0, + "content": "6. Chen, X., Jiang, B., Liu, W., Huang, Z., Fu, B., Chen, T., Yu, G.: Executing your commands via motion diffusion in latent space. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.363, + 0.785, + 0.39 + ], + "angle": 0, + "content": "7. Cohan, S., Tevet, G., Reda, D., Peng, X.B., van de Panne, M.: Flexible motion in-between with diffusion models. ArXiv (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.391, + 0.785, + 0.418 + ], + "angle": 0, + "content": "8. Dabral, R., Mughal, M.H., Golyanik, V., Theobalt, C.: Mofusion: A framework for denoising-diffusion-based motion synthesis. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.419, + 0.785, + 0.447 + ], + "angle": 0, + "content": "9. Dai, W., Chen, L.H., Wang, J., Liu, J., Dai, B., Tang, Y.: Motionlcm: Real-time controllable motion generation via latent consistency model. ArXiv (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.448, + 0.785, + 0.474 + ], + "angle": 0, + "content": "0. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. NeurIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.476, + 0.785, + 0.516 + ], + "angle": 0, + "content": "1. Du, H., Herrmann, E., Sprenger, J., Fischer, K., Slusallek, P.: Stylistic locomotion modeling and synthesis using variational generative models. In: Proceedings of the 12th ACM SIGGRAPH Conference on Motion, Interaction and Games (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.517, + 0.785, + 0.544 + ], + "angle": 0, + "content": "2. Everaert, M.N., Bocchio, M., Arpa, S., Susstrunk, S., Achanta, R.: Diffusion in style. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.545, + 0.785, + 0.573 + ], + "angle": 0, + "content": "3. Ghosh, A., Dabral, R., Golyanik, V., Theobalt, C., Slusallek, P.: Remos: Reactive 3d motion synthesis for two-person interactions. In: ArXiv (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.574, + 0.785, + 0.601 + ], + "angle": 0, + "content": "4. Guo, C., Mu, Y., Javed, M.G., Wang, S., Cheng, L.: Momask: Generative masked modeling of 3d human motions. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.602, + 0.785, + 0.629 + ], + "angle": 0, + "content": "5. Guo, C., Mu, Y., Zuo, X., Dai, P., Yan, Y., Lu, J., Cheng, L.: Generative human motion stylization in latent space. ArXiv (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.63, + 0.785, + 0.657 + ], + "angle": 0, + "content": "6. Guo, C., Zou, S., Zuo, X., Wang, S., Ji, W., Li, X., Cheng, L.: Generating diverse and natural 3d human motions from text. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.659, + 0.697, + 0.672 + ], + "angle": 0, + "content": "7. Ho, J., Salimans, T.: Classifier-free diffusion guidance. ArXiv (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.673, + 0.785, + 0.713 + ], + "angle": 0, + "content": "8. Huang, S., Wang, Z., Li, P., Jia, B., Liu, T., Zhu, Y., Liang, W., Zhu, S.C.: Diffusion-based generation, optimization, and planning in 3d scenes. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.715, + 0.785, + 0.742 + ], + "angle": 0, + "content": "9. Jang, D.K., Park, S., Lee, S.H.: Motion puzzle: Arbitrary motion style transfer by body part. TOG (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.743, + 0.785, + 0.77 + ], + "angle": 0, + "content": "20. Jiang, B., Chen, X., Liu, W., Yu, J., Yu, G., Chen, T.: Motiongpt: Human motion as a foreign language. ArXiv (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.771, + 0.785, + 0.798 + ], + "angle": 0, + "content": "21. Jones, M., Wang, S.Y., Kumari, N., Bau, D., Zhu, J.Y.: Customizing text-to-image models with a single image pair. ArXiv (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "22. Karunratanakul, K., Preechakul, K., Aksan, E., Beeler, T., Suwajanakorn, S., Tang, S.: Optimizing diffusion noise can serve as universal motion priors. In: Arxiv (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.179, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Zhong et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "23. Karunratanakul, K., Preechakul, K., Suwajanakorn, S., Tang, S.: Gmd: Controllable human motion synthesis via guided diffusion models. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.785, + 0.202 + ], + "angle": 0, + "content": "24. Kim, H.J., Lee, S.H.: Perceptual characteristics by motion style category. In: Eurographics (Short Papers) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.203, + 0.785, + 0.242 + ], + "angle": 0, + "content": "25. Kulkarni, N., Rempe, D., Genova, K., Kundu, A., Johnson, J., Fouhey, D., Guibas, L.: Nifty: Neural object interaction fields for guided human motion synthesis. ArXiv (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.243, + 0.771, + 0.257 + ], + "angle": 0, + "content": "26. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. ArXiv (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.258, + 0.785, + 0.297 + ], + "angle": 0, + "content": "27. Mason, I., Starke, S., Komura, T.: Real-time style modelling of human locomotion via feature-wise transformations and local motion phases. Proceedings of the ACM on Computer Graphics and Interactive Techniques (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.298, + 0.785, + 0.338 + ], + "angle": 0, + "content": "28. Park, S., Jang, D.K., Lee, S.H.: Diverse motion stylization for multiple style domains via spatial-temporal graph-based generative model. Proceedings of the ACM on Computer Graphics and Interactive Techniques (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.339, + 0.785, + 0.366 + ], + "angle": 0, + "content": "29. Peng, X., Xie, Y., Wu, Z., Jampani, V., Sun, D., Jiang, H.: Hoi-diff: Text-driven synthesis of 3d human-object interactions using diffusion models. ArXiv (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.367, + 0.785, + 0.392 + ], + "angle": 0, + "content": "30. Peng, X.B., Ma, Z., Abbeel, P., Levine, S., Kanazawa, A.: Amp: Adversarial motion priors for stylized physics-based character control. TOG (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.393, + 0.785, + 0.419 + ], + "angle": 0, + "content": "31. Petrovich, M., Black, M.J., Varol, G.: Temos: Generating diverse human motions from textual descriptions. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.42, + 0.785, + 0.46 + ], + "angle": 0, + "content": "32. Petrovich, M., Litany, O., Iqbal, U., Black, M.J., Varol, G., Bin Peng, X., Rempe, D.: Multi-track timeline control for text-driven 3d human motion generation. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.461, + 0.785, + 0.487 + ], + "angle": 0, + "content": "33. Pi, H., Peng, S., Yang, M., Zhou, X., Bao, H.: Hierarchical generation of human-object interactions with diffusion probabilistic models. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.785, + 0.514 + ], + "angle": 0, + "content": "34. Pinyoanuntapong, E., Saleem, M.U., Wang, P., Lee, M., Das, S., Chen, C.: Bamm: Bidirectional autoregressive motion model. ArXiv (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.515, + 0.785, + 0.541 + ], + "angle": 0, + "content": "35. Pinyoanuntapong, E., Wang, P., Lee, M., Chen, C.: Mmm: Generative masked motion model. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.542, + 0.785, + 0.582 + ], + "angle": 0, + "content": "36. Raab, S., Gat, I., Sala, N., Tevet, G., Shalev-Arkushin, R., Fried, O., Bermano, A.H., Cohen-Or, D.: Monkey see, monkey do: Harnessing self-attention in motion diffusion for zero-shot motion transfer. ArXiv (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.583, + 0.785, + 0.609 + ], + "angle": 0, + "content": "37. Raab, S., Leibovitch, I., Tevet, G., Arar, M., Bermano, A.H., Cohen-Or, D.: Single motion diffusion. ArXiv (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.61, + 0.785, + 0.651 + ], + "angle": 0, + "content": "38. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.652, + 0.785, + 0.691 + ], + "angle": 0, + "content": "39. Rempe, D., Luo, Z., Peng, X.B., Yuan, Y., Kitani, K., Kreis, K., Fidler, S., Litany, O.: Trace and pace: Controllable pedestrian animation via guided trajectory diffusion. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.692, + 0.785, + 0.732 + ], + "angle": 0, + "content": "40. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.733, + 0.785, + 0.759 + ], + "angle": 0, + "content": "41. Shafir, Y., Tevet, G., Kapon, R., Bermano, A.H.: Human motion diffusion as a generative prior. ArXiv (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.76, + 0.785, + 0.786 + ], + "angle": 0, + "content": "42. Shah, V., Ruiz, N., Cole, F., Lu, E., Lazebnik, S., Li, Y., Jampani, V.:Ziplora: Any subject in any style by effectively merging loras. ArXiv (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.787, + 0.785, + 0.8 + ], + "angle": 0, + "content": "43. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. ArXiv (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.801, + 0.785, + 0.813 + ], + "angle": 0, + "content": "44. Song, Y., Dhariwal, P., Chen, M., Sutskever, I.: Consistency models. ArXiv (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "45. Tao, T., Zhan, X., Chen, Z., van de Panne, M.: Style-erd: Responsive and coherent online motion style transfer. Arxiv (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.451, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SMooDi: Stylized Motion Diffusion Model" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.175 + ], + "angle": 0, + "content": "46. Tevet, G., Raab, S., Gordon, B., Shafir, Y., Cohen-or, D., Bermano, A.H.: Human motion diffusion model. In: ICLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.175, + 0.787, + 0.203 + ], + "angle": 0, + "content": "47. Wan, W., Dou, Z., Komura, T., Wang, W., Jayaraman, D., Liu, L.: Tlcontrol: Trajectory and language control for human motion synthesis. ArXiv (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.203, + 0.787, + 0.231 + ], + "angle": 0, + "content": "48. Wang, Z., Wang, J., Lin, D., Dai, B.: Intercontrol: Generate human motion interactions by controlling every joint. ArXiv (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.231, + 0.787, + 0.259 + ], + "angle": 0, + "content": "49. Wen, Y.H., Yang, Z., Fu, H., Gao, L., Sun, Y., Liu, Y.J.: Autoregressive stylized motion synthesis with generative flow. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.259, + 0.787, + 0.286 + ], + "angle": 0, + "content": "50. Wu, Q., Zhao, Y., Wang, Y., Tai, Y.W., Tang, C.K.: Motionllm: Multimodal motion-language learning with large language models. ArXiv (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.286, + 0.787, + 0.314 + ], + "angle": 0, + "content": "51. Wu, Q., Shi, Y., Huang, X., Yu, J., Xu, L., Wang, J.: Thor: Text to human-object interaction diffusion via relation intervention. ArXiv (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.314, + 0.787, + 0.342 + ], + "angle": 0, + "content": "52. Xia, S., Wang, C., Chai, J., Hodgins, J.: Realtime style transfer for unlabeled heterogeneous human motion. TOG (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.342, + 0.787, + 0.37 + ], + "angle": 0, + "content": "53. Xie, Y., Jampani, V., Zhong, L., Sun, D., Jiang, H.: Omnicontrol: Control any joint at any time for human motion generation. In: ICLR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.37, + 0.787, + 0.411 + ], + "angle": 0, + "content": "54. Xu, P., Xie, K., Andrews, S., Kry, P.G., Neff, M., McGuire, M., Karamouzas, I., Zordan, V.: Adaptnet: Policy adaptation for physics-based character control. TOG (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.411, + 0.787, + 0.438 + ], + "angle": 0, + "content": "55. Xu, S., Ma, Z., Huang, Y., Lee, H., Chai, J.: Rethinking cycle consistency in text-guided diffusion for image manipulation. NeurIPS (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.438, + 0.787, + 0.466 + ], + "angle": 0, + "content": "56. Xu, S., Li, Z., Wang, Y.X., Gui, L.Y.: Interdiff: Generating 3d human-object interactions with physics-informed diffusion. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.466, + 0.787, + 0.494 + ], + "angle": 0, + "content": "57. Xu, S., Wang, Z., Wang, Y.X., Gui, L.Y.: Interdreamer: Zero-shot text to 3d dynamic human-object interaction. ArXiv (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.494, + 0.787, + 0.522 + ], + "angle": 0, + "content": "58. Yi, H., Thies, J., Black, M.J., Peng, X.B., Rempe, D.: Generating human interaction motions in scenes with text control. ArXiv (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.522, + 0.787, + 0.55 + ], + "angle": 0, + "content": "59. Yu, J., Wang, Y., Zhao, C., Ghanem, B., Zhang, J.: Freedom: Training-free energy-guided conditional diffusion model. ArXiv (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.55, + 0.787, + 0.577 + ], + "angle": 0, + "content": "60. Yuan, Y., Song, J., Iqbal, U., Vahdat, A., Kautz, J.: Physdiff: Physics-guided human motion diffusion model. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.577, + 0.787, + 0.605 + ], + "angle": 0, + "content": "61. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.605, + 0.787, + 0.633 + ], + "angle": 0, + "content": "62. Zhang, M., Cai, Z., Pan, L., Hong, F., Guo, X., Yang, L., Liu, Z.: Motiondiffuse: Text-driven human motion generation with diffusion model. PAMI (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.633, + 0.787, + 0.673 + ], + "angle": 0, + "content": "63. Zhang, Y., Huang, D., Liu, B., Tang, S., Lu, Y., Chen, L., Bai, L., Chu, Q., Yu, N., Ouyang, W.: Motiongpt: Finetuned llms are general-purpose motion generators. ArXiv (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.673, + 0.787, + 0.715 + ], + "angle": 0, + "content": "64. Zhou, W., Dou, Z., Cao, Z., Liao, Z., Wang, J., Wang, W., Liu, Y., Komura, T., Wang, W., Liu, L.: Emdm: Efficient motion diffusion model for fast, high-quality motion generation. Arxiv (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.715, + 0.787, + 0.743 + ], + "angle": 0, + "content": "65. Zhou, Y., Barnes, C., Lu, J., Yang, J., Li, H.: On the continuity of rotation representations in neural networks. In: CVPR (2019)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.743 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/SMooDi_ Stylized Motion Diffusion Model/10d37622-04ad-4741-9458-84cef09d41ff_origin.pdf b/2024/SMooDi_ Stylized Motion Diffusion Model/10d37622-04ad-4741-9458-84cef09d41ff_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..379da316bb77d40a3375420b86935e6a9fad1c99 --- /dev/null +++ b/2024/SMooDi_ Stylized Motion Diffusion Model/10d37622-04ad-4741-9458-84cef09d41ff_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:03161eb5a8faf2015c9be4ed5154ab62561ffb36cfd7ee8e92d0712b0acd057e +size 12480204 diff --git a/2024/SMooDi_ Stylized Motion Diffusion Model/full.md b/2024/SMooDi_ Stylized Motion Diffusion Model/full.md new file mode 100644 index 0000000000000000000000000000000000000000..546083fbe52433b010138dea455bce3e9f707c1f --- /dev/null +++ b/2024/SMooDi_ Stylized Motion Diffusion Model/full.md @@ -0,0 +1,324 @@ +# SMooDi: Stylized Motion Diffusion Model + +Lei Zhong $^{1}$ , Yiming Xie $^{1}$ , Varun Jampani $^{2}$ , Deqing Sun $^{3}$ , and Huaizu Jiang $^{1}$ + +1 Northeastern University + +2 Stability AI + +3 Google Research + +{le.zhong,xie.yim,h.jiang}@northeastern.edu + +varunjampani@gmail.com + +deqingsun@google.com + +Abstract. We introduce a novel Stylized Motion Diffusion model, dubbed SMooDi, to generate stylized motion driven by content texts and style motion sequences. Unlike existing methods that either generate motion of various content or transfer style from one sequence to another, SMooDi can rapidly generate motion across a broad range of content and diverse styles. To this end, we tailor a pre-trained text-to-motion model for stylization. Specifically, we propose style guidance to ensure that the generated motion closely matches the reference style, alongside a lightweight style adaptor that directs the motion towards the desired style while ensuring realism. Experiments across various applications demonstrate that our proposed framework outperforms existing methods in stylized motion generation. Project Page: https://neu-vi.github.io/SMooDi/ + +Keywords: Motion synthesis $\cdot$ Diffusion model $\cdot$ Stylized motion + +# 1 Introduction + +We address the problem of generating stylized motion from a content text and a style motion sequence, as shown in Fig. 1. Human motion can typically be characterized by two components: content and style. Motion content represents the nature of a movement, such as walking and waving, and motion style reflects individual characteristics, such as personality traits (e.g., old, childlike) and emotions (e.g., happy, angry). Traditional pipelines create stylized motions via motion capture from actors, and are both labor-intensive and time-consuming. Therefore, decades of research have focused on developing automatic methods to assist stylized motion creation [1,11, 19]. + +Motion style transfer [1, 45] is a practical and popular approach for the creation of stylized motion. It transfers the style from an existing style motion sequence to another existing content motion sequence. However, when a broad array of motion needs to be stylized, the pipeline may be inefficient - it would first require the collection of a large number of content motion sequences and then apply a motion style transfer method to process each sequence independently. Moreover, motion sequences are not always readily available, especially + +![](images/1a78d6a219a0266697792277f5abdf7c44a6ed82ffac53eaa9e197a415310631.jpg) +Fig. 1: SMooDi can generate realistic, stylized human motions given a content text a style motion sequence. It also accepts a motion sequence as content input. Darker color indicates later frames in the sequence. To better showcase the stylized motion generation, we place the style label for the each of the style motion sequence. Note that such style labels are not used as model input and shown here for visualization purpose only. (Best viewed in color.) + +for some customized content, such as running along a specific trajectory. They may still need to be created first by actors or animators for stylization. + +Recent advances of human motion generation with diffusion models [10, 17] have shown impressive results of creating diverse and realistic human motions. But most efforts have concentrated on efficiently and accurately translating textual prompts into human motions, focusing on the content only [6, 31, 46]. Integrating the style condition to generate stylized motions remains under-explored. + +Combining these two lines of research is a straightforward approach to tackle stylized motion generation, where a motion style transfer method [1, 45] can be applied to each motion sequence generated by a text-driven motion diffusion model [6, 31, 46]. However, in addition to the aforementioned inefficiency issue, this approach has two more limitations. First, error may accumulate across the pipeline. As, motion style transfer methods are usually trained with high-quality real-world motion sequences, we empirically observe that their performance may significantly degrade for imperfect motions produced by text-to-motion techniques. Second, existing motion style transfer methods rely on specialized style datasets [1, 27, 52] with limited motion content, which restricts their applications to motion diffusion models. + +In this paper, we present a novel stylized motion diffusion model, dubbed SMooDi, that customizes a pre-trained text-to-motion model for stylization. Built upon the pre-trained motion latent diffusion model (MLD) [6], SMooDi inherits MLD's ability to generate diverse motion content. At the same time, + +SMooDi can generate motions in a variety of styles according to different style reference conditions, as shown in Fig. 1. Our main novelty is the style modulation module, which consists of a style adaptor and a style guidance module. First of all, drawing inspiration from controllable image generation [61], the style adaptor is designed to predict residual features conditioned on style reference motion sequence within each attention layer of MLD. It is useful for incorporating the style condition while ensuring the realism of the generated motion. Second, we design both classifier-free and classifier-based style guidance to more precisely control the stylized motion generation. Specifically, the classifier-free style and content guidance are linearly combined, where we can easily strike a balance between preserving content and reflecting style within the generated motion. At the same time, we design an additional classifier-based style guidance mechanism. It is an analytic function quantifying the disparity between the generated motion and the style reference motion in a style-centric embedding space, whose gradients are subsequently employed to guide the generated motion closer to the intended style. Our style adaptor and guidance module are designed to be complementary, which lead to high-quality stylized motion generation. The two modules are jointly optimized in a feature space instead of sequence-wise separate stitching, thereby avoiding the error accumulation issue. + +Although our approach is primarily designed for stylized motion generation driven by content text, we can utilize DDIM-Inversion [43] to identify the noisy latent corresponding to the content motion sequence. Following the same procedure as for text-driven content, SMooDi is capable of facilitating stylized motion generation based on content motion sequences. In other words, motion style transfer is a downstream application of our approach should it be desired in practice, e.g., to stylize the already created motion sequences. + +Experiments on the HumanML3D [16] and 100STYLE [27] datasets demonstrate that SMooDi surpasses other baseline models in generating stylized motion driven by content text, excelling in both content preservation and style reflection. More importantly, unlike previous methods that require individual fine-tuning for each style [12, 30, 54], SMooDi successfully integrates diverse content from the HumanML3D dataset and various styles from the 100STYLE dataset into a single model without requiring additional tuning during inference. + +To summarize, our contributions are: (1) To our knowledge, SMooDi is the first approach that adapts a pre-trained text-to-motion model to generate diverse stylized motion. (2) We introduce a novel style modulation module that utilizes a stylized adaptor and a style classifier guidance to enable stylized motion generation while ensuring style reflection, content preservation, and realism. (3) Experiments demonstrate that SMooDi not only sets a new state of the art in stylized motion generation driven by content text but also achieves performance comparable to state-of-the-art methods in motion style transfer. + +# 2 Related Work + +# 2.1 Human Motion Generation + +Human motion generation has attracted great attention [4, 5, 7, 9, 14, 16, 20, 32, 34-36, 47, 50, 51, 57, 58, 63]. Inspired by the impressive performance of diffusion models in image generation, a lot of works [6, 8, 13, 18, 22, 23, 25, 29, 32, 33, 39, 41, 46, 48, 53, 56, 60, 62, 64] utilize diffusion models to generate human motion. MDM [46] facilitates high-quality generation and versatile conditioning, providing a solid baseline for novel motion generation tasks. MLD [6] minimizes computational overhead during both training and inference by establishing the diffusion process within the latent space. Driven by the efficacy of diffusion models for control and conditioning, several studies have leveraged pre-trained motion diffusion models to generate long-sequence motions [41], enable human-object interactions [29], and control the joint trajectory of generated motions [23, 53]. However, there is no work exploring how to leverage pre-trained motion diffusion models to generate diverse stylized motion. While some studies [2, 3, 37] have enabled stylized motion generation in their diffusion pipeline, their methods are trained from scratch, and the supported styles are restricted by their motion content dataset. It is challenging for them to simultaneously support diverse motion content and style. In this work, we build upon a pre-trained motion diffusion model, MLD, and explore how to fine-tune it on a larger motion style dataset, 100STYLE, to learn diverse motion styles while retaining the ability to support motion generation across a wide range of content. + +# 2.2 Motion Style Transfer + +Recently, motion style transfer has seen quality enhancements through the adoption of various advanced neural architectures and generative models, such as graph neural networks [28], time-series models [27, 45], normalizing flows [49], and diffusion models [3,37]. Specifically, Aberman et al. [1] designed a two-branch generative adversarial network to disentangle motion style from content and facilitate their re-composition. Their approach effectively breaks the constraint of requiring a paired motion dataset. Motion Puzzle [19] realizes a framework that can control the style of individual body parts. Above methods extract both content and style features from the motion sequence. Moreover, Guo et al. [15] leverage the latent space of pre-trained motion models to enhance the extraction and infusion of motion content and style. However, a major limitation of these models is their reliance on specialized style datasets [1, 52] with limited motion content, which restricts their applications. In this work, we customize a pre-trained text-to-motion model for stylization, thus inheriting its ability to generate diverse motion content. + +# 3 Stylized Motion Diffusion Model + +In this section, we introduce our proposed SMooDi for incorporating style conditions from a style motion sequence into a content-oriented pre-trained motion + +![](images/a065bb04d06a62b99007a08d915c79fc33dbe7da04b3af8b8c815c11e28fe1ec.jpg) +Fig. 2: Overview of SMooDi. Our model generates stylized human motions from content text and a style motion sequence. At the denoising step $t$ , our model takes the content text $\mathbf{c}$ , style motion $\mathbf{s}$ , and noisy latent $\mathbf{z}_t$ as input and predicts $\epsilon_t$ , which is then transferred to $\mathbf{z}_{t-1}$ . This denoising step is repeated $T$ times to obtain the noise-free motion latent $\mathbf{z}_0$ , which is fed into a motion decoder $D$ to produce the stylized motion. + +diffusion model (MLD [6]). Fig. 2 presents an overview of SMooDi. Following the setting in MLD [6], we place the diffusion process in the motion latent space. Let $\epsilon_{\theta}$ denote the latent denoiser (a UNet parameterized by $\theta$ ), and $\{z_t\}_{t=0}^T$ denote the sequence of noisy latents, where $z_T$ is a Gaussian noise. Given a content prompt $\mathbf{c}$ and a style prompt $\mathbf{s}$ , we define $\epsilon_t = \epsilon_{\theta}(z_t, t, \mathbf{c}, \mathbf{s})$ for the denoising at step $t$ ( $0 < t \leq T$ ). A cleaner noisy latent $z_{t-1}$ can be obtained by subtracting $\epsilon_t$ from $z_t$ . The denoising step is repeated $T$ iterations until a clean latent $z_0$ is obtained. It can then be decoded by a motion decoder $\mathbf{D}$ into a realistic motion sequence $\mathbf{x} \in \mathbb{R}^{N \times H}$ that accurately reflects both the content and style conditions. Here, $N$ represents the length of the motion sequence, and $H$ is the dimension of human motion representations. We employ the same motion representations as in HumanML3D [16], where $H = 263$ . + +As shown in the Fig. 2, the content prompt is a text description, and the style prompt is provided by a reference style motion sequence $\mathbf{s} \in \mathbb{R}^{N \times H}$ . In this section, we focus on using a text description as the content prompt $\mathbf{c}$ to explain our proposed stylized motion diffusion model. By employing the DDIM-Inversion [43] to identify the noisy latent corresponding to a motion content sequence, we can effectively use motion sequences as content prompts to generate stylized motions. In other words, motion style transfer is a downstream application of our proposed approach. + +Our proposed stylization module consists of a style adaptor and a style guidance module. We will explain them separately in the rest of this section. + +# 3.1 Style Adaptor + +Although LoRA has been successfully used to incorporate "style" into the models in image domain [21, 42], they typically require training a separate LoRA for each style. In contrast, we focus on fine-tuning the model just once to adapt to + +![](images/176bb973d7f55c23ac7ec4bcd6f259025e607ffd350f5cebff31c4ab290e720c.jpg) +Fig. 3: Detailed illustration of our proposed style adaptor. The style adaptor is connected to the motion diffusion model via zero linear layer. The output of the style adaptor from each Transformer encoder is added to the motion diffusion model to steer the predicted noise towards the target style. + +various motion styles, where adapting ControlNet [61] is more suitable. Therefore, we design a content-aware style adaptor based on ControlNet. This adaptor incorporates the motion style condition into the pre-trained MLD [6]. + +Instead of learning to disentangle motion style from content from scratch on a large motion dataset, we redirect our focus towards capturing the motion style while ensuring the preservation of diverse motion content within the pre-trained MLD framework. Specifically, it consists of a trainable copy of the Transformer Encoder from the latent diffusion model in MLD. The architecture of the style adaptor is illustrated in Fig. 3. An independent style encoder is utilized to extract the style embedding from the style motion sequence $\mathbf{s}$ . The style adaptor takes the same content prompt $\mathbf{c}$ , the noised latent $\mathbf{z}_t$ and timestep $t$ as in MLD, and the extracted style embedding. Each Transformer layer in the original latent diffusion model and the style adaptor is connected via a linear layer, with both weight and bias initially set to zeros. As training progresses, the style adaptor learns the style constraints and gradually applies the learned feature corrections to the corresponding layers in the latent diffusion model, thereby implicitly steering the output towards the desired style. + +# 3.2 Style Guidance + +The style adaptor alone may not be sufficient to successfully incorporate the style condition. We further leverage both the classifier-free and classifier guidance to further enhance the stylization of a motion diffusion model. The combination of two types of guidance effectively ensures that generated motion meets multiple constraints while maintaining realism, complementing the style adaptor. + +Classifier-free Style Guidance. With the introduction of an extra style condition, we can divide the conditioned classifier-free guidance into two parts. + +$$ +\epsilon_ {\theta} \left(\boldsymbol {z} _ {t}, t, \mathbf {c}, \mathbf {s}\right) = \epsilon_ {\theta} \left(\boldsymbol {z} _ {t}, t, \emptyset , \emptyset\right) + +$$ + +$$ +\underbrace {w _ {c} (\epsilon_ {\theta} (\boldsymbol {z} _ {t} , t , \mathbf {c} , \emptyset) - \epsilon_ {\theta} (\boldsymbol {z} _ {t} , t , \emptyset , \emptyset))} + \underbrace {w _ {s} (\epsilon_ {\theta} (\boldsymbol {z} _ {t} , t , \mathbf {c} , \mathbf {s}) - \epsilon_ {\theta} (\boldsymbol {z} _ {t} , t , \mathbf {c} , \emptyset))} _ {\text {(1)}} +$$ + +Classifier-free Content Guidance Classifier-free Style Guidance + +where $w_{c}$ and $w_{s}$ represent the strengths of the classifier-free guidance for the condition $\mathbf{c}$ and $\mathbf{s}$ , respectively. We slightly abuse the notations here by using $\emptyset$ to denote a condition is not used. The classifier-free content guidance is the same as in MLD, which can facilitate the text-to-motion generation process in combination with the first term's unconditioned guidance. Our proposed classifier-free style guidance works in a similar way. Note that $\epsilon_{\theta}(z_t,t,c,s)$ is MLD model with the style adaptor incorporated introduced in the previous section, which takes both a textual prompt and style motion sequence as input. By contrasting the text-driven denoising output with and without the style condition, it can highlight the effectiveness of the style input $s$ and facilitate the generation of stylized motion driven by content text. Our insight here is that by dividing the conditioned guidance into content and style components separately, we can easily strike a balance between preserving content and reflecting style within the generated motion. + +To better understand the classifier-free content and style guidance, we visualize each of them through decoding denoised latent $\mathbf{z}_0$ into the motion space. As illustrated in Fig.4(a), the content guidance ensures the motion generation is faithful to the textual prompt, while the style guidance, as shown in Fig.4(b), emphasizes style-related characteristics in the output. Combining both forms of guidance results in a stylized motion that adheres to both content and style conditions, as illustrated in Fig. 4(c). + +Classifier-based Style Guidance. To further improve the stylization of a motion diffusion model, we adopt the classifier guidance [10,59] to provide stronger guidance to the generated motion towards the desired style. The core of our classifier-based style guidance is a novel analytic function $G(\mathbf{z}_t,t,\mathbf{s})$ , which calculates the $L_{1}$ distance between the style embedding of the generated clean motion $\hat{\pmb{x}}_0$ at denoising step $t$ and the reference style motions $\mathbf{s}$ . The function's gradient is utilized to steer the generated motion towards the desired style. + +$$ +\epsilon_ {\boldsymbol {\theta}} (\boldsymbol {z} _ {t}, t, \mathbf {c}, \mathbf {s}) = \epsilon_ {\boldsymbol {\theta}} (\boldsymbol {z} _ {t}, t, \mathbf {c}, \mathbf {s}) + \tau \nabla_ {\boldsymbol {z} _ {t}} G (\boldsymbol {z} _ {t}, t, \mathbf {s}), +$$ + +$$ +G \left(\boldsymbol {z} _ {t}, t, \mathbf {s}\right) = | f \left(\hat {\boldsymbol {x}} _ {0}\right) - f (\mathbf {s}) |, \tag {2} +$$ + +where $\tau$ adjusts the strength of reference-style guidance and $f$ denotes the style feature extractor. The generated motion $\hat{\pmb{x}}_0$ is obtained by first converting the denoising output latent $\pmb{z}_t$ into the predicted clean latent as shown below: + +$$ +\hat {\boldsymbol {z}} _ {0} = \frac {\boldsymbol {z} _ {t} - \sqrt {1 - \alpha_ {t}} \varepsilon_ {\theta} (\boldsymbol {z} _ {t} , t , \boldsymbol {c} , \boldsymbol {s})}{\sqrt {\alpha_ {t}}}, \tag {3} +$$ + +where $\alpha_{t}$ denotes the pre-defined noise scale in the forward process of the diffusion model. The predicted clean latent $\hat{z}_0$ is then input into the motion decoder $\mathbf{D}$ to obtain the generated motion. + +We obtain the style feature extractor by training a style classifier on the 100STYLE dataset [27] and removing its final layer. We refer the readers to more details in the supplementary materials. The training of the style feature extractor with ground-truth style labels for supervision enables it to effectively + +![](images/7b64e7f402554b1ab66ab627b226f6f54cfa96271300186a8592e72263d86444.jpg) +Text: A person walks forward and then sits down. +Fig. 4: Visual illustrations of the classifier-free and classifier-based style guidance. (a) and (b) respectively show the classifier-free content and style guidance; (c) displays the initial stylized motion resulting from the combination of (a) and (b); (d) illustrates the refined stylized motion modified by the classifier-based style guidance. + +capture style-related features. Therefore, style classifier guidance can provide more guidance to the stylized motion generation. + +Combination of the Two Style Guidance. The classifier-free and classifier-based style guidance are designed to complement each other, each playing a vital role in accurately reflecting the target style in the generated motions. As illustrated in Fig. 4, the desired style motion is "arms open wide to the sides like an airplane". The classifier-free style guidance (Fig. 4(b)) can capture style-related characteristic in a reasonably accurate manner. When combined with the classifier-free content guidance, it depicts the desired style (Fig.4(c)). Refined further by the classifier-based style guidance, the stylization is more authentic, where the person's harms are more open (Fig.4(d)). In addition to such visual results, quantitative ablation studies also verify the effectiveness of our proposed both classifier-free and classifier-based style guidance. + +At the same time, although classifier-based style guidance offers precise style control, its effectiveness may be compromised when the content text significantly diverges from locomotion-related movements. This is because the style feature function is trained solely on the 100STYLE dataset, which contains only such movements. Over-reliance on style classifier guidance risks producing motions that fail to execute the desired actions, leading to unrealistic and physically implausible movements. Therefore, we leverage a content-aware style adaptor that establishes the fundamental style direction, while style classifier guidance refines this base for a more precise outcome. The effectiveness of this design is verified in our ablation studies. + +# 3.3 Learning Scheme + +Following [61], a straightforward approach to train SMooDi is to freeze the parameters of MLD and solely train the style adaptor on the 100STYLE dataset using the following loss function: + +$$ +\mathcal {L} _ {s t d} = \mathbb {E} _ {\epsilon , \mathbf {z}} \left[ \| \epsilon_ {\theta} (\mathbf {z} _ {t}, t, \mathbf {c}, \mathbf {s}) - \epsilon \| _ {2} ^ {2} \right], \tag {4} +$$ + +where $\epsilon \sim \mathcal{N}(0, \mathbf{I})$ represents the ground-truth noise added to $z_0$ . In our experiments, however, we found that this loss function alone leads to an issue of + +"content-forgetting", where the model progressively looses the MLD's ability to generate motions with diverse contents. To address this issue, we design a content prior preservation loss $L_{pr}$ . Specifically, we randomly sample motions from the HumanML3D dataset to compute a prior preservation loss when fine-tuning SMooDi on the 100STYLE dataset. + +$$ +\mathcal {L} _ {p r} = \mathbb {E} _ {\epsilon^ {\prime}, \boldsymbol {z} ^ {\prime}} \left[ \| \epsilon_ {\theta} (\boldsymbol {z} _ {t} ^ {\prime}, t, \mathbf {c} ^ {\prime}, \mathbf {s} ^ {\prime}) - \epsilon^ {\prime} \| _ {2} ^ {2} \right], \tag {5} +$$ + +where $\mathbf{z}_t^\prime$ , $\mathbf{c}^\prime$ and $\mathbf{s}^\prime$ represents the motion latent, content prompt and style motion sequence derived from the HumanML3D dataset. $\epsilon^\prime$ is the noise map added to $\mathbf{z}_0^\prime$ . A similar solution is used in DreamBooth [40] to solve the "language drift" problem, where images generated from the frozen pretrained image generation model are utilized to enforce a class-prior preservation loss during model finetuning. Our content preservation loss can effectively mitigate content forgetting while learning diverse motion styles from the 100STYLE dataset. + +To further encourage the style adaptor to focus on motion style, while also ensuring that the latent diffusion model in MLD handles motion content well, we introduce an additional cycle prior-preservation loss, inspired by [19,55]. Specifically, we start this process by randomly sampling content text and motion style sequences from both the 100STYLE and HumanML3D datasets simultaneously. Then, we intermix the content text and motion style from these sequences with each other. Finally, we repeat this process to reconstruct the original motion sequences. The formula is expressed as follows: + +$$ +\mathcal {L} _ {c y c} = \mathbb {E} _ {\boldsymbol {z}, \boldsymbol {z} ^ {\prime}, \epsilon , \epsilon^ {\prime}} \left[ \left\| \epsilon_ {\theta} (\boldsymbol {z} _ {t} ^ {s h}, t, \mathbf {c}, \boldsymbol {s} ^ {h s}) + \epsilon_ {\theta} (\boldsymbol {z} _ {t} ^ {h s}, t, \mathbf {c} ^ {\prime}, \boldsymbol {s} ^ {s h}) - \epsilon - \epsilon^ {\prime} \right\| _ {2} ^ {2} \right], \qquad (6) +$$ + +where $s^{hs}$ denotes the motion sequence created by merging content from the HumanML3D dataset with style from the 100STYLE dataset. Similarly, $s^{sh}$ represents the sequence where content is sourced from the 100STYLE dataset and style from the HumanML3D dataset. The noised latent codes $z_{t}^{sh}$ and $z_{t}^{hs}$ correspond to $s^{sh}$ and $s^{hs}$ , respectively. Essentially, the cycle prior-preservation loss exchanges diverse content and style between two datasets, encouraging the content text to remain invariant in the generated motion under forward and backward translation. Overall, the training loss function of our framework is defined as follows: + +$$ +\mathcal {L} _ {\text {a l l}} = \mathcal {L} _ {\text {s t d}} + \lambda_ {p r} \mathcal {L} _ {p r} + \lambda_ {\text {c y c}} \mathcal {L} _ {\text {c y c}} \tag {7} +$$ + +where $\lambda_{pr}$ and $\lambda_{cyc}$ are hyperparameters. We refer readers to the pseudocode and illustration for training in the supplementary material for more details. + +# 4 Experiments + +We conduct experiments on both stylized text2motion and motion style transfer to demonstrate the effectiveness of our framework. Both tasks use a motion sequence as a style prompt, with the primary difference being their content prompt input: the former utilizes text, while the latter relies on motion sequences. + +Table 1: Comparison with baseline methods on stylized motion generation driven by content text, using a combination of the 100STYLE (providing style) and HumanML3D datasets (providing content). + +
MethodFID ↓ Foot skatingMM Dist ↓ R-precision ↑ Diversity → SRA ↑ ratio ↓
Ours1.6090.1244.4770.5719.23572.418
MLD+Motion Puzzle [19]6.1270.1856.4670.2906.476263.769
MLD+Aberman et al. [1]3.3090.3475.9830.4068.81654.367
ChatGPT+MLD0.6140.1314.3130.6058.8364.819
+ +![](images/8a9bb57b0d85d341d5bb052f1244e30c7e415cee0faba7aa1771752ae6c75e3b.jpg) +A person kicks. + +![](images/5690f5675f7815d3acc4f2ec27d669e871e5bb5f76569ffeba5ad839e218ba7a.jpg) +(a) $\mathrm{MLD + }$ Motion Puzzle (b) $\mathrm{MLD + }$ Aberman et al. + +![](images/d20b7f278bef388d97796ed81374f419ea0c4a86bbdb00444dfa53071d2365c8.jpg) + +![](images/ba94ee848bc02ff1ac8a9219c441a35904447cfdf8276e3c99011f2b9ceb7ddd.jpg) +(c) Ours + +![](images/64b79eba7852a8af79637162ebaf2c7c698a869850f7b9c5b3ab9835c71c04b7.jpg) +Fig. 5: Qualitative comparisons of our approach and baseline methods on two stylized motion generation task. + +![](images/0c83bd5658cd74abe2de00e7783896b233f1194aef387151aa646e2baf18b223.jpg) + +![](images/635b950c9548f5ef96be20ee1baaac4f7ad758840f926e21b03dc39487b7f76b.jpg) + +![](images/4a540e666fc0cbc5de51165940de3a0905c63a35a52da71294eb2de9ab55ef05.jpg) + +![](images/58c685bcf66f0b58dad0b423be5d3619937b4c86c1e0669d46f06a3acab567b9.jpg) + +Datasets. We utilize the HumanML3D dataset [16] as our motion content dataset and the 100STYLE dataset [27] as our motion style dataset. The HumanML3D dataset is the largest motion capture dataset, featuring text annotations and comprising 14,646 motions and 44,970 motion annotations. Following the processing approach outlined in [16], we preprocess the HumanML3D dataset to obtain consistent motion representations. On the other hand, the 100STYLE dataset [27], being the largest motion style dataset, comprises up to 1,125 minutes of motion sequences, showcasing a wide array of 100 diverse locomotion styles. Due to differences in skeletons between the 100STYLE dataset and HumanML3D, we retarget the motions from 100STYLE to match the HumanML3D (SMPL-H) skeleton. Following this alignment, we apply the same processing steps as used for the HumanML3D dataset to preprocess the 100STYLE dataset. Moreover, as the 100STYLE dataset lacks text descriptions, we leverage MotionGPT [20] to generate pseudo text descriptions for the motion sequences in the 100STYLE dataset. + +Evaluation metrics are designed to assess three dimensions: Content Preservation, Style Reflection, and Realism. For content preservation and style reflection assessment, we employ metrics consistent with those used in [6]: motion-retrieval precision (R precision), Multi-modal Distance (MM Dist), Diversity, and Frechet Inception Distance (FID). Additionally, recognizing the common foot skating issues in kinetics-based motion generation methods, we incorporate the foot skating ratio metric proposed by [23] into our motion quality evaluation. For style + +reflection, we employ Style Recognition Accuracy (SRA) [19]. During evaluation, we randomly select a content text from the HumanML3D dataset and a motion style sequence from the 100STYLE dataset to generate the stylized motion. We then use a pre-trained style classifier to compute the SRA for the generated motion. It's noteworthy that some motion style labels in the 100STYLE dataset, like 'kick' and 'jump,' inherently convey motion content, which may conflict with the content text in HumanML3D dataset. To address this, we categorize the motion style labels into server groups following the approach by Kim et al. [24], Specifically excluding the 'ACT' group ensures that only motion style labels not conflicting with motion content are considered when computing the SRA metric. Further details about the evaluation metrics are provided in the supplementary material. + +Baselines. For motion style transfer task, we compare our methods with two state-of-the-art methods, namely Motion Puzzle [19] and Aberman et al. [1]. To ensure a fair comparison, we train the compared methods under the same settings as ours, using a combined dataset comprising HumanML3D and 100STYLE. Due to the constraints of the multi-class discriminator in Aberman et al., which requires style labels, we adopt the training method outlined in Motion Puzzle to eliminate the need for style labels. For stylized text2motion task, we compare our method against baselines capable of generating stylized motion from content text and style motion sequences. The straightforward baselines involve applying motion style transfer methods to the motion sequences generated by the text2motion model. To align with our approach that uses a pre-trained motion diffusion model, the text2motion models in the baselines for stylized motion generation select MLD, and the motion style transfer methods are consistent with those used in the motion style transfer task. For the stylized text2motion task, we compare our method against two kinds of baselines capable of generating stylized motion from content text and style motion sequences. The first kind of baseline involves applying motion style transfer methods to the motion sequences generated by the text2motion model. To align with our approach that uses a pretrained motion diffusion model, the text2motion models select MLD [6], and the motion style transfer methods are consistent with those used in the motion style transfer task. The second kind of baseline involves using ChatGPT to merge style labels from 100STYLE with text from HumanML3D into a sentence. For example, given the content text 'a person walks.' and the style label 'old,' we obtain 'an elderly person walks.' This merged sentence is then fed to MLD. + +# 4.1 Comparison to Baseline Methods + +Quantitative and Qualitative For the task of stylized text2motion, Table 1 reports the comparisons of our method with the three baseline methods. + +As shown in the $3^{rd}$ row of Table 1, ChatGPT+MLD only achieves around $5.29\%$ in terms of SRA, indicating that MLD cannot enable stylized generation from text alone, even though it contains style descriptions. Notably, our method outperforms the two baselines that combine MLD with motion style transfer methods in all metrics. + +(a) Evaluation on HumanML3D dataset + +Table 2: Comparison with baseline methods on motion style transfer. + +
MethodFoot skating ratio ↓
Ours0.0951.58265.147
Motion Puzzle [19]0.1976.87167.233
(Aberman et al [1])0.3383.89261.006
+ +(b) Evaluation on Xia dataset + +
MethodFoot skating ratio↓FID↓SRA↑(%)CRA↑(%)
Ours0.03174.66361.11145.555
Motion Puzzle [19]0.03165.36067.77825.556
(Aberman et al [1])0.02605.68156.66734.444
+ +Specifically, our method performs better than MLD+Motion Puzzle in the SRA metric by $13.56\%$ and significantly outperforms MLD+Aberman et al. in the FID metric by $51.38\%$ and $0.64\%$ in the R-precision metric. The first row of Fig. 5 validates our observation, where the motion generated by our method performs better in adhering to both content and style constraints than baseline methods. In contrast, MLP+Aberman et al. can successfully perform the action but fail to reflect the motion style in Fig.5(b), while Motion Puzzle can accurately reflect the motion style but struggles to effectively perform the action in Fig.5(a). + +For the task of motion style transfer, since it does not take content text as input, text-motion related metrics such as MM Dist, R-precision, and Diversity are not applicable and thus are not reported. Part (a) of Table 2 presents a comparison between our method and the two baseline methods, using the HumanML3D dataset as the motion content source and drawing motion styles from the 100STYLE dataset. Our method delivers competitive results in the SRA metric and excels in the FID and foot skating ratio metrics. Specifically, we see a substantial $59.35\%$ improvement in the FID metric over Aberman et al. [1]. To more effectively compare the generalizability of different methods, we conduct experiments on the Xia dataset [52], a small, specific motion style dataset that was unseen by our and the baseline models during training. Because motion content labels are present in the Xia dataset, we report the Content Recognition Accuracy (CRA). Part (b) of Table 2 showcases the results. Our method maintains competitive performance in the SRA metric, with only a marginal $9.83\%$ decrease in SRA compared to Motion Puzzle. On the contrary, our method exhibits a significant $32.26\%$ increase in the CAR metric relative to Aberman et al., and a notable $78.26\%$ enhancement over Motion Puzzle. Our method achieves a better balance between style reflection and content preservation. The second row of Fig. 5 validates this observation. It is worth noting that, unlike other motion style transfer methods, our method does not incorporate objectives for enabling stylized motion generation using motion content sequences. Through simple DDIM-Inversion and without any additional optimization or regularization, our method achieves performance comparable to existing motion style transfer methods. + +User Study. Due to the highly subjective nature of stylized motion, we conduct User studies using pairwise comparisons to further evaluate our proposed method in the tasks of stylized motion generation and motion style transfer. We + +Table 3: Ablation Studies on HumanML3D Content and 100STYLE Styles. + +
MethodFID↓ ratio↓MM Dist↓R-precision↑ (Top-3)Diversity→ SRA(%)↑
Ours (on all)1.6090.1244.4770.5719.23572.418
w/o Lcyc2.0460.1364.4650.5698.86964.866
w/o Lpr + Lcyc5.9960.1666.0980.3357.45681.841
w/o classifier-based1.0500.1114.0850.6309.44520.245
w/o adaptor2.9840.1234.5260.5508.37269.952
+ +Text: A person walks forward and then sits down. + +![](images/9cc5618563a75e1af0ddadb6dcabde399b405b079118742d11e117b59ea4d369.jpg) +Fig. 6: Visual comparisons of the ablation designs and our full model. + +recruited 22 human subjects to participate in the study. In each test, participants are presented with two 4-second video clips synthesized by our method and one comparison method. They are then required to select their preferred clip while considering Realism, Style Reflection, and Content Preservation dimensions, respectively. As shown in Fig. 7, our method receives more user appreciation compared to two baselines across three dimensions in two tasks. Further user study details are provided in supplementary. + +# 4.2 Ablation studies + +To validate the effectiveness of our framework's design choices, we have conducted several ablation studies: the first assesses the impact of each loss function term, while the second evaluates the influence of the style adaptor and style guidance during sampling. + +Loss Components. Firstly, we exclude the cycle-prior term in the loss function, denoted as $w / o$ $L_{cyc}$ . Comparing the results in the $1^{\mathrm{st}}$ and $2^{\mathrm{nd}}$ rows in Table 3, we observe that our full model outperforms in all content preservation and style reflection metrics. The motion generated by our approach can still perform the content adhering to the text description but performs worse in accurately reflecting the motion style, as reflected by the arms not being fully extended horizontally. + +Since the cycle prior-preservation term is built upon the prior-preservation term, it is meaningless to exclude $L_{pr}$ while retaining $L_{cyc}$ . Therefore, we further exclude both the prior-preservation and cycle-prior term, denoted as $w / oL_{pr} + L_{cyc}$ in the $3^{\mathrm{rd}}$ row. By comparing the results in the $2^{\mathrm{nd}}$ and $3^{\mathrm{rd}}$ rows of Table 3, we notice that while the number of SRAs is higher in the third row, other metrics show a significant decline. Specifically, in terms of the FID metric, performance deteriorates by more than $229\%$ . Indeed, without $L_{pr}$ , the + +![](images/c4a1190969a76af004907c207d9a7b1c4fa11a0ca8aec274d55bec4405b3e88c.jpg) +Fig. 7: User Study on two stylized motion generation tasks. + +![](images/da2f3a7d738cef23e56906e87ff523501e30364fbf6d2e37a13f1ebe0745bcc2.jpg) + +model tends to lose the ability to translate content text into corresponding motion, a phenomenon named 'content-forgetting' as described in Sec. 3.3. Fig. 6(a) validates our observation, showing that the content in the generated motion significantly deviates from the text descriptions and closely resembles the style motion sequence. + +Style Adaptor and Style Guidance. Initially, we compare our model to a variant without the classifier-based guidance, $w / o$ classifier-based, to demonstrate its effectiveness. The $5^{\text{th}}$ row of Table 3 presents the results. Consistent with the findings in [19], a reasonable trade-off between content preservation and style reflection is observed. Although classifier-based style guidance may slightly affect the content preservation metrics, it significantly boosts the model's performance in the SRA metric, yielding an impressive $208\%$ improvement.Fig. 6(d) demonstrates that, without classifier-based style guidance, the generated motions can reflect the motion style, yet they still fall short of fully achieving the target style. Classifier-based style guidance can effectively bridge this gap. + +Subsequently, as shown in Table 3, we evaluate a variant without the style adaptor, denoted as $w / o$ adaptor (the last row). In cases where the SRA values are close, the style adaptor improved the FID metric by about $80.46\%$ . Fig. 6(c) shows that the generated motion can greatly perform the 'walk' action while successfully reflecting the style, but fails to perform the 'sit' action. This indicates that the effectiveness of classifier-based style guidance diminishes when the content text deviates from locomotion-related movements. Relying solely on it may even adversely affect action performance. + +# 5 Conclusion + +In this work, we introduce the Stylized Motion Diffusion Model, a novel approach that leverages a pre-trained motion diffusion model to facilitate stylized motion generation driven by content text. By integrating a style adaptor and style classifier guidance, our method is capable of producing realistic human motions that accurately reflect both the content text descriptions and the desired motion style from motion sequences. Through detailed ablation studies, we have demonstrated the effectiveness of each component in our framework. + +# References + +1. Aberman, K., Weng, Y., Lischinski, D., Cohen-Or, D., Chen, B.: Unpaired motion style transfer from video to animation. TOG (2020) +2. Alexanderson, S., Nagy, R., Beskow, J., Henter, G.E.: Listen, denoise, action! audio-driven motion synthesis with diffusion models. TOG (2023) +3. Ao, T., Zhang, Z., Liu, L.: Gesturediffuclip: Gesture diffusion model with clip latents. TOG (2023) +4. Cen, Z., Pi, H., Peng, S., Shen, Z., Yang, M., Zhu, S., Bao, H., Zhou, X.: Generating human motion in 3d scenes from text descriptions. In: CVPR (2024) +5. Chen, L.H., Lu, S., Zeng, A., Zhang, H., Wang, B., Zhang, R., Zhang, L.: Motionllm: Understanding human behaviors from human motions and videos. ArXiv (2024) +6. Chen, X., Jiang, B., Liu, W., Huang, Z., Fu, B., Chen, T., Yu, G.: Executing your commands via motion diffusion in latent space. In: CVPR (2023) +7. Cohan, S., Tevet, G., Reda, D., Peng, X.B., van de Panne, M.: Flexible motion in-between with diffusion models. ArXiv (2024) +8. Dabral, R., Mughal, M.H., Golyanik, V., Theobalt, C.: Mofusion: A framework for denoising-diffusion-based motion synthesis. In: CVPR (2023) +9. Dai, W., Chen, L.H., Wang, J., Liu, J., Dai, B., Tang, Y.: Motionlcm: Real-time controllable motion generation via latent consistency model. ArXiv (2024) +0. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. NeurIPS (2021) +1. Du, H., Herrmann, E., Sprenger, J., Fischer, K., Slusallek, P.: Stylistic locomotion modeling and synthesis using variational generative models. In: Proceedings of the 12th ACM SIGGRAPH Conference on Motion, Interaction and Games (2019) +2. Everaert, M.N., Bocchio, M., Arpa, S., Susstrunk, S., Achanta, R.: Diffusion in style. In: ICCV (2023) +3. Ghosh, A., Dabral, R., Golyanik, V., Theobalt, C., Slusallek, P.: Remos: Reactive 3d motion synthesis for two-person interactions. In: ArXiv (2023) +4. Guo, C., Mu, Y., Javed, M.G., Wang, S., Cheng, L.: Momask: Generative masked modeling of 3d human motions. In: CVPR (2024) +5. Guo, C., Mu, Y., Zuo, X., Dai, P., Yan, Y., Lu, J., Cheng, L.: Generative human motion stylization in latent space. ArXiv (2024) +6. Guo, C., Zou, S., Zuo, X., Wang, S., Ji, W., Li, X., Cheng, L.: Generating diverse and natural 3d human motions from text. In: CVPR (2022) +7. Ho, J., Salimans, T.: Classifier-free diffusion guidance. ArXiv (2022) +8. Huang, S., Wang, Z., Li, P., Jia, B., Liu, T., Zhu, Y., Liang, W., Zhu, S.C.: Diffusion-based generation, optimization, and planning in 3d scenes. In: CVPR (2023) +9. Jang, D.K., Park, S., Lee, S.H.: Motion puzzle: Arbitrary motion style transfer by body part. TOG (2022) +20. Jiang, B., Chen, X., Liu, W., Yu, J., Yu, G., Chen, T.: Motiongpt: Human motion as a foreign language. ArXiv (2023) +21. Jones, M., Wang, S.Y., Kumari, N., Bau, D., Zhu, J.Y.: Customizing text-to-image models with a single image pair. ArXiv (2024) +22. Karunratanakul, K., Preechakul, K., Aksan, E., Beeler, T., Suwajanakorn, S., Tang, S.: Optimizing diffusion noise can serve as universal motion priors. In: Arxiv (2023) + +23. Karunratanakul, K., Preechakul, K., Suwajanakorn, S., Tang, S.: Gmd: Controllable human motion synthesis via guided diffusion models. In: ICCV (2023) +24. Kim, H.J., Lee, S.H.: Perceptual characteristics by motion style category. In: Eurographics (Short Papers) (2019) +25. Kulkarni, N., Rempe, D., Genova, K., Kundu, A., Johnson, J., Fouhey, D., Guibas, L.: Nifty: Neural object interaction fields for guided human motion synthesis. ArXiv (2023) +26. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. ArXiv (2017) +27. Mason, I., Starke, S., Komura, T.: Real-time style modelling of human locomotion via feature-wise transformations and local motion phases. Proceedings of the ACM on Computer Graphics and Interactive Techniques (2022) +28. Park, S., Jang, D.K., Lee, S.H.: Diverse motion stylization for multiple style domains via spatial-temporal graph-based generative model. Proceedings of the ACM on Computer Graphics and Interactive Techniques (2021) +29. Peng, X., Xie, Y., Wu, Z., Jampani, V., Sun, D., Jiang, H.: Hoi-diff: Text-driven synthesis of 3d human-object interactions using diffusion models. ArXiv (2023) +30. Peng, X.B., Ma, Z., Abbeel, P., Levine, S., Kanazawa, A.: Amp: Adversarial motion priors for stylized physics-based character control. TOG (2021) +31. Petrovich, M., Black, M.J., Varol, G.: Temos: Generating diverse human motions from textual descriptions. In: ECCV (2022) +32. Petrovich, M., Litany, O., Iqbal, U., Black, M.J., Varol, G., Bin Peng, X., Rempe, D.: Multi-track timeline control for text-driven 3d human motion generation. In: CVPR (2024) +33. Pi, H., Peng, S., Yang, M., Zhou, X., Bao, H.: Hierarchical generation of human-object interactions with diffusion probabilistic models. In: ICCV (2023) +34. Pinyoanuntapong, E., Saleem, M.U., Wang, P., Lee, M., Das, S., Chen, C.: Bamm: Bidirectional autoregressive motion model. ArXiv (2024) +35. Pinyoanuntapong, E., Wang, P., Lee, M., Chen, C.: Mmm: Generative masked motion model. In: CVPR (2024) +36. Raab, S., Gat, I., Sala, N., Tevet, G., Shalev-Arkushin, R., Fried, O., Bermano, A.H., Cohen-Or, D.: Monkey see, monkey do: Harnessing self-attention in motion diffusion for zero-shot motion transfer. ArXiv (2024) +37. Raab, S., Leibovitch, I., Tevet, G., Arar, M., Bermano, A.H., Cohen-Or, D.: Single motion diffusion. ArXiv (2023) +38. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021) +39. Rempe, D., Luo, Z., Peng, X.B., Yuan, Y., Kitani, K., Kreis, K., Fidler, S., Litany, O.: Trace and pace: Controllable pedestrian animation via guided trajectory diffusion. In: CVPR (2023) +40. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: CVPR (2023) +41. Shafir, Y., Tevet, G., Kapon, R., Bermano, A.H.: Human motion diffusion as a generative prior. ArXiv (2023) +42. Shah, V., Ruiz, N., Cole, F., Lu, E., Lazebnik, S., Li, Y., Jampani, V.:Ziplora: Any subject in any style by effectively merging loras. ArXiv (2023) +43. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. ArXiv (2020) +44. Song, Y., Dhariwal, P., Chen, M., Sutskever, I.: Consistency models. ArXiv (2023) +45. Tao, T., Zhan, X., Chen, Z., van de Panne, M.: Style-erd: Responsive and coherent online motion style transfer. Arxiv (2022) + +46. Tevet, G., Raab, S., Gordon, B., Shafir, Y., Cohen-or, D., Bermano, A.H.: Human motion diffusion model. In: ICLR (2023) +47. Wan, W., Dou, Z., Komura, T., Wang, W., Jayaraman, D., Liu, L.: Tlcontrol: Trajectory and language control for human motion synthesis. ArXiv (2023) +48. Wang, Z., Wang, J., Lin, D., Dai, B.: Intercontrol: Generate human motion interactions by controlling every joint. ArXiv (2023) +49. Wen, Y.H., Yang, Z., Fu, H., Gao, L., Sun, Y., Liu, Y.J.: Autoregressive stylized motion synthesis with generative flow. In: CVPR (2021) +50. Wu, Q., Zhao, Y., Wang, Y., Tai, Y.W., Tang, C.K.: Motionllm: Multimodal motion-language learning with large language models. ArXiv (2024) +51. Wu, Q., Shi, Y., Huang, X., Yu, J., Xu, L., Wang, J.: Thor: Text to human-object interaction diffusion via relation intervention. ArXiv (2024) +52. Xia, S., Wang, C., Chai, J., Hodgins, J.: Realtime style transfer for unlabeled heterogeneous human motion. TOG (2015) +53. Xie, Y., Jampani, V., Zhong, L., Sun, D., Jiang, H.: Omnicontrol: Control any joint at any time for human motion generation. In: ICLR (2024) +54. Xu, P., Xie, K., Andrews, S., Kry, P.G., Neff, M., McGuire, M., Karamouzas, I., Zordan, V.: Adaptnet: Policy adaptation for physics-based character control. TOG (2023) +55. Xu, S., Ma, Z., Huang, Y., Lee, H., Chai, J.: Rethinking cycle consistency in text-guided diffusion for image manipulation. NeurIPS (2024) +56. Xu, S., Li, Z., Wang, Y.X., Gui, L.Y.: Interdiff: Generating 3d human-object interactions with physics-informed diffusion. In: ICCV (2023) +57. Xu, S., Wang, Z., Wang, Y.X., Gui, L.Y.: Interdreamer: Zero-shot text to 3d dynamic human-object interaction. ArXiv (2024) +58. Yi, H., Thies, J., Black, M.J., Peng, X.B., Rempe, D.: Generating human interaction motions in scenes with text control. ArXiv (2024) +59. Yu, J., Wang, Y., Zhao, C., Ghanem, B., Zhang, J.: Freedom: Training-free energy-guided conditional diffusion model. ArXiv (2023) +60. Yuan, Y., Song, J., Iqbal, U., Vahdat, A., Kautz, J.: Physdiff: Physics-guided human motion diffusion model. In: ICCV (2023) +61. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: ICCV (2023) +62. Zhang, M., Cai, Z., Pan, L., Hong, F., Guo, X., Yang, L., Liu, Z.: Motiondiffuse: Text-driven human motion generation with diffusion model. PAMI (2024) +63. Zhang, Y., Huang, D., Liu, B., Tang, S., Lu, Y., Chen, L., Bai, L., Chu, Q., Yu, N., Ouyang, W.: Motiongpt: Finetuned llms are general-purpose motion generators. ArXiv (2023) +64. Zhou, W., Dou, Z., Cao, Z., Liao, Z., Wang, J., Wang, W., Liu, Y., Komura, T., Wang, W., Liu, L.: Emdm: Efficient motion diffusion model for fast, high-quality motion generation. Arxiv (2023) +65. Zhou, Y., Barnes, C., Lu, J., Yang, J., Li, H.: On the continuity of rotation representations in neural networks. In: CVPR (2019) \ No newline at end of file diff --git a/2024/SMooDi_ Stylized Motion Diffusion Model/images.zip b/2024/SMooDi_ Stylized Motion Diffusion Model/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d4f3ead732c697726887067cf1be8d561134b6f9 --- /dev/null +++ b/2024/SMooDi_ Stylized Motion Diffusion Model/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:59da127f190b2291d5e8c806cc74b3394fc6f7eb6f5d21e90e8f87730fcc5682 +size 353093 diff --git a/2024/SMooDi_ Stylized Motion Diffusion Model/layout.json b/2024/SMooDi_ Stylized Motion Diffusion Model/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..629f12c7eb48850876b245b5245c12e8dff028d0 --- /dev/null +++ b/2024/SMooDi_ Stylized Motion Diffusion Model/layout.json @@ -0,0 +1,9893 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 159, + 111, + 454, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 111, + 454, + 129 + ], + "spans": [ + { + "bbox": [ + 159, + 111, + 454, + 129 + ], + "type": "text", + "content": "SMooDi: Stylized Motion Diffusion Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 139, + 149, + 474, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 149, + 474, + 175 + ], + "spans": [ + { + "bbox": [ + 139, + 149, + 474, + 175 + ], + "type": "text", + "content": "Lei Zhong" + }, + { + "bbox": [ + 139, + 149, + 474, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 139, + 149, + 474, + 175 + ], + "type": "text", + "content": ", Yiming Xie" + }, + { + "bbox": [ + 139, + 149, + 474, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 139, + 149, + 474, + 175 + ], + "type": "text", + "content": ", Varun Jampani" + }, + { + "bbox": [ + 139, + 149, + 474, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 139, + 149, + 474, + 175 + ], + "type": "text", + "content": ", Deqing Sun" + }, + { + "bbox": [ + 139, + 149, + 474, + 175 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 139, + 149, + 474, + 175 + ], + "type": "text", + "content": ", and Huaizu Jiang" + }, + { + "bbox": [ + 139, + 149, + 474, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 252, + 183, + 361, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 183, + 361, + 194 + ], + "spans": [ + { + "bbox": [ + 252, + 183, + 361, + 194 + ], + "type": "text", + "content": "1 Northeastern University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 278, + 195, + 336, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 278, + 195, + 336, + 205 + ], + "spans": [ + { + "bbox": [ + 278, + 195, + 336, + 205 + ], + "type": "text", + "content": "2 Stability AI" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 268, + 206, + 345, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 268, + 206, + 345, + 217 + ], + "spans": [ + { + "bbox": [ + 268, + 206, + 345, + 217 + ], + "type": "text", + "content": "3 Google Research" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 206, + 218, + 408, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 218, + 408, + 228 + ], + "spans": [ + { + "bbox": [ + 206, + 218, + 408, + 228 + ], + "type": "text", + "content": "{le.zhong,xie.yim,h.jiang}@northeastern.edu" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 255, + 230, + 359, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 230, + 359, + 239 + ], + "spans": [ + { + "bbox": [ + 255, + 230, + 359, + 239 + ], + "type": "text", + "content": "varunjampani@gmail.com" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 259, + 240, + 354, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 259, + 240, + 354, + 250 + ], + "spans": [ + { + "bbox": [ + 259, + 240, + 354, + 250 + ], + "type": "text", + "content": "deqingsun@google.com" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 159, + 277, + 457, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 277, + 457, + 408 + ], + "spans": [ + { + "bbox": [ + 159, + 277, + 457, + 408 + ], + "type": "text", + "content": "Abstract. We introduce a novel Stylized Motion Diffusion model, dubbed SMooDi, to generate stylized motion driven by content texts and style motion sequences. Unlike existing methods that either generate motion of various content or transfer style from one sequence to another, SMooDi can rapidly generate motion across a broad range of content and diverse styles. To this end, we tailor a pre-trained text-to-motion model for stylization. Specifically, we propose style guidance to ensure that the generated motion closely matches the reference style, alongside a lightweight style adaptor that directs the motion towards the desired style while ensuring realism. Experiments across various applications demonstrate that our proposed framework outperforms existing methods in stylized motion generation. Project Page: https://neu-vi.github.io/SMooDi/" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 160, + 418, + 428, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 418, + 428, + 430 + ], + "spans": [ + { + "bbox": [ + 160, + 418, + 428, + 430 + ], + "type": "text", + "content": "Keywords: Motion synthesis " + }, + { + "bbox": [ + 160, + 418, + 428, + 430 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 418, + 428, + 430 + ], + "type": "text", + "content": " Diffusion model " + }, + { + "bbox": [ + 160, + 418, + 428, + 430 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 418, + 428, + 430 + ], + "type": "text", + "content": " Stylized motion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 449, + 230, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 449, + 230, + 462 + ], + "spans": [ + { + "bbox": [ + 132, + 449, + 230, + 462 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 473, + 481, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 481, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 481, + 581 + ], + "type": "text", + "content": "We address the problem of generating stylized motion from a content text and a style motion sequence, as shown in Fig. 1. Human motion can typically be characterized by two components: content and style. Motion content represents the nature of a movement, such as walking and waving, and motion style reflects individual characteristics, such as personality traits (e.g., old, childlike) and emotions (e.g., happy, angry). Traditional pipelines create stylized motions via motion capture from actors, and are both labor-intensive and time-consuming. Therefore, decades of research have focused on developing automatic methods to assist stylized motion creation [1,11, 19]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 582, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 481, + 666 + ], + "type": "text", + "content": "Motion style transfer [1, 45] is a practical and popular approach for the creation of stylized motion. It transfers the style from an existing style motion sequence to another existing content motion sequence. However, when a broad array of motion needs to be stylized, the pipeline may be inefficient - it would first require the collection of a large number of content motion sequences and then apply a motion style transfer method to process each sequence independently. Moreover, motion sequences are not always readily available, especially" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 145, + 118, + 465, + 290 + ], + "blocks": [ + { + "bbox": [ + 145, + 118, + 465, + 290 + ], + "lines": [ + { + "bbox": [ + 145, + 118, + 465, + 290 + ], + "spans": [ + { + "bbox": [ + 145, + 118, + 465, + 290 + ], + "type": "image", + "image_path": "1a78d6a219a0266697792277f5abdf7c44a6ed82ffac53eaa9e197a415310631.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 301, + 482, + 369 + ], + "lines": [ + { + "bbox": [ + 130, + 301, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 301, + 482, + 369 + ], + "type": "text", + "content": "Fig. 1: SMooDi can generate realistic, stylized human motions given a content text a style motion sequence. It also accepts a motion sequence as content input. Darker color indicates later frames in the sequence. To better showcase the stylized motion generation, we place the style label for the each of the style motion sequence. Note that such style labels are not used as model input and shown here for visualization purpose only. (Best viewed in color.)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 396, + 481, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 396, + 481, + 420 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 481, + 420 + ], + "type": "text", + "content": "for some customized content, such as running along a specific trajectory. They may still need to be created first by actors or animators for stylization." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 422, + 482, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 482, + 483 + ], + "type": "text", + "content": "Recent advances of human motion generation with diffusion models [10, 17] have shown impressive results of creating diverse and realistic human motions. But most efforts have concentrated on efficiently and accurately translating textual prompts into human motions, focusing on the content only [6, 31, 46]. Integrating the style condition to generate stylized motions remains under-explored." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 484, + 482, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 484, + 482, + 615 + ], + "spans": [ + { + "bbox": [ + 130, + 484, + 482, + 615 + ], + "type": "text", + "content": "Combining these two lines of research is a straightforward approach to tackle stylized motion generation, where a motion style transfer method [1, 45] can be applied to each motion sequence generated by a text-driven motion diffusion model [6, 31, 46]. However, in addition to the aforementioned inefficiency issue, this approach has two more limitations. First, error may accumulate across the pipeline. As, motion style transfer methods are usually trained with high-quality real-world motion sequences, we empirically observe that their performance may significantly degrade for imperfect motions produced by text-to-motion techniques. Second, existing motion style transfer methods rely on specialized style datasets [1, 27, 52] with limited motion content, which restricts their applications to motion diffusion models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we present a novel stylized motion diffusion model, dubbed SMooDi, that customizes a pre-trained text-to-motion model for stylization. Built upon the pre-trained motion latent diffusion model (MLD) [6], SMooDi inherits MLD's ability to generate diverse motion content. At the same time," + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Zhong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 482, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 482, + 342 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 482, + 342 + ], + "type": "text", + "content": "SMooDi can generate motions in a variety of styles according to different style reference conditions, as shown in Fig. 1. Our main novelty is the style modulation module, which consists of a style adaptor and a style guidance module. First of all, drawing inspiration from controllable image generation [61], the style adaptor is designed to predict residual features conditioned on style reference motion sequence within each attention layer of MLD. It is useful for incorporating the style condition while ensuring the realism of the generated motion. Second, we design both classifier-free and classifier-based style guidance to more precisely control the stylized motion generation. Specifically, the classifier-free style and content guidance are linearly combined, where we can easily strike a balance between preserving content and reflecting style within the generated motion. At the same time, we design an additional classifier-based style guidance mechanism. It is an analytic function quantifying the disparity between the generated motion and the style reference motion in a style-centric embedding space, whose gradients are subsequently employed to guide the generated motion closer to the intended style. Our style adaptor and guidance module are designed to be complementary, which lead to high-quality stylized motion generation. The two modules are jointly optimized in a feature space instead of sequence-wise separate stitching, thereby avoiding the error accumulation issue." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 363, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 363, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 363, + 482, + 449 + ], + "type": "text", + "content": "Although our approach is primarily designed for stylized motion generation driven by content text, we can utilize DDIM-Inversion [43] to identify the noisy latent corresponding to the content motion sequence. Following the same procedure as for text-driven content, SMooDi is capable of facilitating stylized motion generation based on content motion sequences. In other words, motion style transfer is a downstream application of our approach should it be desired in practice, e.g., to stylize the already created motion sequences." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 466, + 482, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 466, + 482, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 466, + 482, + 552 + ], + "type": "text", + "content": "Experiments on the HumanML3D [16] and 100STYLE [27] datasets demonstrate that SMooDi surpasses other baseline models in generating stylized motion driven by content text, excelling in both content preservation and style reflection. More importantly, unlike previous methods that require individual fine-tuning for each style [12, 30, 54], SMooDi successfully integrates diverse content from the HumanML3D dataset and various styles from the 100STYLE dataset into a single model without requiring additional tuning during inference." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "To summarize, our contributions are: (1) To our knowledge, SMooDi is the first approach that adapts a pre-trained text-to-motion model to generate diverse stylized motion. (2) We introduce a novel style modulation module that utilizes a stylized adaptor and a style classifier guidance to enable stylized motion generation while ensuring style reflection, content preservation, and realism. (3) Experiments demonstrate that SMooDi not only sets a new state of the art in stylized motion generation driven by content text but also achieves performance comparable to state-of-the-art methods in motion style transfer." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 276, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 276, + 91, + 447, + 102 + ], + "type": "text", + "content": "SMooDi: Stylized Motion Diffusion Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 137, + 297, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 137, + 297, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 137, + 297, + 148 + ], + "type": "text", + "content": "2.1 Human Motion Generation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 153, + 482, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 153, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 133, + 153, + 482, + 392 + ], + "type": "text", + "content": "Human motion generation has attracted great attention [4, 5, 7, 9, 14, 16, 20, 32, 34-36, 47, 50, 51, 57, 58, 63]. Inspired by the impressive performance of diffusion models in image generation, a lot of works [6, 8, 13, 18, 22, 23, 25, 29, 32, 33, 39, 41, 46, 48, 53, 56, 60, 62, 64] utilize diffusion models to generate human motion. MDM [46] facilitates high-quality generation and versatile conditioning, providing a solid baseline for novel motion generation tasks. MLD [6] minimizes computational overhead during both training and inference by establishing the diffusion process within the latent space. Driven by the efficacy of diffusion models for control and conditioning, several studies have leveraged pre-trained motion diffusion models to generate long-sequence motions [41], enable human-object interactions [29], and control the joint trajectory of generated motions [23, 53]. However, there is no work exploring how to leverage pre-trained motion diffusion models to generate diverse stylized motion. While some studies [2, 3, 37] have enabled stylized motion generation in their diffusion pipeline, their methods are trained from scratch, and the supported styles are restricted by their motion content dataset. It is challenging for them to simultaneously support diverse motion content and style. In this work, we build upon a pre-trained motion diffusion model, MLD, and explore how to fine-tune it on a larger motion style dataset, 100STYLE, to learn diverse motion styles while retaining the ability to support motion generation across a wide range of content." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 407, + 271, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 271, + 419 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 271, + 419 + ], + "type": "text", + "content": "2.2 Motion Style Transfer" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 425, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 425, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 133, + 425, + 482, + 604 + ], + "type": "text", + "content": "Recently, motion style transfer has seen quality enhancements through the adoption of various advanced neural architectures and generative models, such as graph neural networks [28], time-series models [27, 45], normalizing flows [49], and diffusion models [3,37]. Specifically, Aberman et al. [1] designed a two-branch generative adversarial network to disentangle motion style from content and facilitate their re-composition. Their approach effectively breaks the constraint of requiring a paired motion dataset. Motion Puzzle [19] realizes a framework that can control the style of individual body parts. Above methods extract both content and style features from the motion sequence. Moreover, Guo et al. [15] leverage the latent space of pre-trained motion models to enhance the extraction and infusion of motion content and style. However, a major limitation of these models is their reliance on specialized style datasets [1, 52] with limited motion content, which restricts their applications. In this work, we customize a pre-trained text-to-motion model for stylization, thus inheriting its ability to generate diverse motion content." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 619, + 348, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 619, + 348, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 619, + 348, + 633 + ], + "type": "text", + "content": "3 Stylized Motion Diffusion Model" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "type": "text", + "content": "In this section, we introduce our proposed SMooDi for incorporating style conditions from a style motion sequence into a content-oriented pre-trained motion" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Zhong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 156, + 119, + 462, + 219 + ], + "blocks": [ + { + "bbox": [ + 156, + 119, + 462, + 219 + ], + "lines": [ + { + "bbox": [ + 156, + 119, + 462, + 219 + ], + "spans": [ + { + "bbox": [ + 156, + 119, + 462, + 219 + ], + "type": "image", + "image_path": "a065bb04d06a62b99007a08d915c79fc33dbe7da04b3af8b8c815c11e28fe1ec.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 228, + 482, + 293 + ], + "lines": [ + { + "bbox": [ + 130, + 228, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 228, + 482, + 293 + ], + "type": "text", + "content": "Fig. 2: Overview of SMooDi. Our model generates stylized human motions from content text and a style motion sequence. At the denoising step " + }, + { + "bbox": [ + 130, + 228, + 482, + 293 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 228, + 482, + 293 + ], + "type": "text", + "content": ", our model takes the content text " + }, + { + "bbox": [ + 130, + 228, + 482, + 293 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 130, + 228, + 482, + 293 + ], + "type": "text", + "content": ", style motion " + }, + { + "bbox": [ + 130, + 228, + 482, + 293 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 130, + 228, + 482, + 293 + ], + "type": "text", + "content": ", and noisy latent " + }, + { + "bbox": [ + 130, + 228, + 482, + 293 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 130, + 228, + 482, + 293 + ], + "type": "text", + "content": " as input and predicts " + }, + { + "bbox": [ + 130, + 228, + 482, + 293 + ], + "type": "inline_equation", + "content": "\\epsilon_t" + }, + { + "bbox": [ + 130, + 228, + 482, + 293 + ], + "type": "text", + "content": ", which is then transferred to " + }, + { + "bbox": [ + 130, + 228, + 482, + 293 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{t-1}" + }, + { + "bbox": [ + 130, + 228, + 482, + 293 + ], + "type": "text", + "content": ". This denoising step is repeated " + }, + { + "bbox": [ + 130, + 228, + 482, + 293 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 228, + 482, + 293 + ], + "type": "text", + "content": " times to obtain the noise-free motion latent " + }, + { + "bbox": [ + 130, + 228, + 482, + 293 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0" + }, + { + "bbox": [ + 130, + 228, + 482, + 293 + ], + "type": "text", + "content": ", which is fed into a motion decoder " + }, + { + "bbox": [ + 130, + 228, + 482, + 293 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 228, + 482, + 293 + ], + "type": "text", + "content": " to produce the stylized motion." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "text", + "content": "diffusion model (MLD [6]). Fig. 2 presents an overview of SMooDi. Following the setting in MLD [6], we place the diffusion process in the motion latent space. Let " + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "text", + "content": " denote the latent denoiser (a UNet parameterized by " + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "text", + "content": "), and " + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "inline_equation", + "content": "\\{z_t\\}_{t=0}^T" + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "text", + "content": " denote the sequence of noisy latents, where " + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "inline_equation", + "content": "z_T" + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "text", + "content": " is a Gaussian noise. Given a content prompt " + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "text", + "content": " and a style prompt " + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "text", + "content": ", we define " + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "inline_equation", + "content": "\\epsilon_t = \\epsilon_{\\theta}(z_t, t, \\mathbf{c}, \\mathbf{s})" + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "text", + "content": " for the denoising at step " + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "inline_equation", + "content": "0 < t \\leq T" + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "text", + "content": "). A cleaner noisy latent " + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "inline_equation", + "content": "z_{t-1}" + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "text", + "content": " can be obtained by subtracting " + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "inline_equation", + "content": "\\epsilon_t" + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "inline_equation", + "content": "z_t" + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "text", + "content": ". The denoising step is repeated " + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "text", + "content": " iterations until a clean latent " + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "text", + "content": " is obtained. It can then be decoded by a motion decoder " + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "text", + "content": " into a realistic motion sequence " + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^{N \\times H}" + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "text", + "content": " that accurately reflects both the content and style conditions. Here, " + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "text", + "content": " represents the length of the motion sequence, and " + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "text", + "content": " is the dimension of human motion representations. We employ the same motion representations as in HumanML3D [16], where " + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "inline_equation", + "content": "H = 263" + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 464, + 482, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 464, + 482, + 560 + ], + "spans": [ + { + "bbox": [ + 130, + 464, + 482, + 560 + ], + "type": "text", + "content": "As shown in the Fig. 2, the content prompt is a text description, and the style prompt is provided by a reference style motion sequence " + }, + { + "bbox": [ + 130, + 464, + 482, + 560 + ], + "type": "inline_equation", + "content": "\\mathbf{s} \\in \\mathbb{R}^{N \\times H}" + }, + { + "bbox": [ + 130, + 464, + 482, + 560 + ], + "type": "text", + "content": ". In this section, we focus on using a text description as the content prompt " + }, + { + "bbox": [ + 130, + 464, + 482, + 560 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 130, + 464, + 482, + 560 + ], + "type": "text", + "content": " to explain our proposed stylized motion diffusion model. By employing the DDIM-Inversion [43] to identify the noisy latent corresponding to a motion content sequence, we can effectively use motion sequences as content prompts to generate stylized motions. In other words, motion style transfer is a downstream application of our proposed approach." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 561, + 481, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 561, + 481, + 585 + ], + "spans": [ + { + "bbox": [ + 130, + 561, + 481, + 585 + ], + "type": "text", + "content": "Our proposed stylization module consists of a style adaptor and a style guidance module. We will explain them separately in the rest of this section." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 605, + 233, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 605, + 233, + 618 + ], + "spans": [ + { + "bbox": [ + 131, + 605, + 233, + 618 + ], + "type": "text", + "content": "3.1 Style Adaptor" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "Although LoRA has been successfully used to incorporate \"style\" into the models in image domain [21, 42], they typically require training a separate LoRA for each style. In contrast, we focus on fine-tuning the model just once to adapt to" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 276, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 276, + 91, + 447, + 102 + ], + "type": "text", + "content": "SMooDi: Stylized Motion Diffusion Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 114, + 476, + 203 + ], + "blocks": [ + { + "bbox": [ + 137, + 114, + 476, + 203 + ], + "lines": [ + { + "bbox": [ + 137, + 114, + 476, + 203 + ], + "spans": [ + { + "bbox": [ + 137, + 114, + 476, + 203 + ], + "type": "image", + "image_path": "176bb973d7f55c23ac7ec4bcd6f259025e607ffd350f5cebff31c4ab290e720c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 210, + 482, + 257 + ], + "lines": [ + { + "bbox": [ + 130, + 210, + 482, + 257 + ], + "spans": [ + { + "bbox": [ + 130, + 210, + 482, + 257 + ], + "type": "text", + "content": "Fig. 3: Detailed illustration of our proposed style adaptor. The style adaptor is connected to the motion diffusion model via zero linear layer. The output of the style adaptor from each Transformer encoder is added to the motion diffusion model to steer the predicted noise towards the target style." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 280, + 479, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 280, + 479, + 316 + ], + "spans": [ + { + "bbox": [ + 130, + 280, + 479, + 316 + ], + "type": "text", + "content": "various motion styles, where adapting ControlNet [61] is more suitable. Therefore, we design a content-aware style adaptor based on ControlNet. This adaptor incorporates the motion style condition into the pre-trained MLD [6]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 316, + 482, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 482, + 483 + ], + "type": "text", + "content": "Instead of learning to disentangle motion style from content from scratch on a large motion dataset, we redirect our focus towards capturing the motion style while ensuring the preservation of diverse motion content within the pre-trained MLD framework. Specifically, it consists of a trainable copy of the Transformer Encoder from the latent diffusion model in MLD. The architecture of the style adaptor is illustrated in Fig. 3. An independent style encoder is utilized to extract the style embedding from the style motion sequence " + }, + { + "bbox": [ + 130, + 316, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 130, + 316, + 482, + 483 + ], + "type": "text", + "content": ". The style adaptor takes the same content prompt " + }, + { + "bbox": [ + 130, + 316, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 130, + 316, + 482, + 483 + ], + "type": "text", + "content": ", the noised latent " + }, + { + "bbox": [ + 130, + 316, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 130, + 316, + 482, + 483 + ], + "type": "text", + "content": " and timestep " + }, + { + "bbox": [ + 130, + 316, + 482, + 483 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 316, + 482, + 483 + ], + "type": "text", + "content": " as in MLD, and the extracted style embedding. Each Transformer layer in the original latent diffusion model and the style adaptor is connected via a linear layer, with both weight and bias initially set to zeros. As training progresses, the style adaptor learns the style constraints and gradually applies the learned feature corrections to the corresponding layers in the latent diffusion model, thereby implicitly steering the output towards the desired style." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 503, + 237, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 503, + 237, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 503, + 237, + 514 + ], + "type": "text", + "content": "3.2 Style Guidance" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 525, + 482, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 525, + 482, + 585 + ], + "spans": [ + { + "bbox": [ + 130, + 525, + 482, + 585 + ], + "type": "text", + "content": "The style adaptor alone may not be sufficient to successfully incorporate the style condition. We further leverage both the classifier-free and classifier guidance to further enhance the stylization of a motion diffusion model. The combination of two types of guidance effectively ensures that generated motion meets multiple constraints while maintaining realism, complementing the style adaptor." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 586, + 481, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 586, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 586, + 481, + 609 + ], + "type": "text", + "content": "Classifier-free Style Guidance. With the introduction of an extra style condition, we can divide the conditioned classifier-free guidance into two parts." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 620, + 266, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 620, + 266, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 620, + 266, + 632 + ], + "type": "interline_equation", + "content": "\\epsilon_ {\\theta} \\left(\\boldsymbol {z} _ {t}, t, \\mathbf {c}, \\mathbf {s}\\right) = \\epsilon_ {\\theta} \\left(\\boldsymbol {z} _ {t}, t, \\emptyset , \\emptyset\\right) +", + "image_path": "3abd0ec791a69c4a49119bac43b8efe9e8019356ec9d3b7a16ef0ee8d8c11bce.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 173, + 635, + 481, + 654 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 635, + 481, + 654 + ], + "spans": [ + { + "bbox": [ + 173, + 635, + 481, + 654 + ], + "type": "interline_equation", + "content": "\\underbrace {w _ {c} (\\epsilon_ {\\theta} (\\boldsymbol {z} _ {t} , t , \\mathbf {c} , \\emptyset) - \\epsilon_ {\\theta} (\\boldsymbol {z} _ {t} , t , \\emptyset , \\emptyset))} + \\underbrace {w _ {s} (\\epsilon_ {\\theta} (\\boldsymbol {z} _ {t} , t , \\mathbf {c} , \\mathbf {s}) - \\epsilon_ {\\theta} (\\boldsymbol {z} _ {t} , t , \\mathbf {c} , \\emptyset))} _ {\\text {(1)}}", + "image_path": "41416dc102dc9a885a841e2b375074cc6a7d0aeff83999d7363f6ea3041c22e2.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 186, + 654, + 441, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 654, + 441, + 662 + ], + "spans": [ + { + "bbox": [ + 186, + 654, + 441, + 662 + ], + "type": "text", + "content": "Classifier-free Content Guidance Classifier-free Style Guidance" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Zhong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "inline_equation", + "content": "w_{c}" + }, + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "inline_equation", + "content": "w_{s}" + }, + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "text", + "content": " represent the strengths of the classifier-free guidance for the condition " + }, + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "text", + "content": ", respectively. We slightly abuse the notations here by using " + }, + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "inline_equation", + "content": "\\emptyset" + }, + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "text", + "content": " to denote a condition is not used. The classifier-free content guidance is the same as in MLD, which can facilitate the text-to-motion generation process in combination with the first term's unconditioned guidance. Our proposed classifier-free style guidance works in a similar way. Note that " + }, + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(z_t,t,c,s)" + }, + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "text", + "content": " is MLD model with the style adaptor incorporated introduced in the previous section, which takes both a textual prompt and style motion sequence as input. By contrasting the text-driven denoising output with and without the style condition, it can highlight the effectiveness of the style input " + }, + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "text", + "content": " and facilitate the generation of stylized motion driven by content text. Our insight here is that by dividing the conditioned guidance into content and style components separately, we can easily strike a balance between preserving content and reflecting style within the generated motion." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 284, + 482, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 284, + 482, + 367 + ], + "spans": [ + { + "bbox": [ + 130, + 284, + 482, + 367 + ], + "type": "text", + "content": "To better understand the classifier-free content and style guidance, we visualize each of them through decoding denoised latent " + }, + { + "bbox": [ + 130, + 284, + 482, + 367 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0" + }, + { + "bbox": [ + 130, + 284, + 482, + 367 + ], + "type": "text", + "content": " into the motion space. As illustrated in Fig.4(a), the content guidance ensures the motion generation is faithful to the textual prompt, while the style guidance, as shown in Fig.4(b), emphasizes style-related characteristics in the output. Combining both forms of guidance results in a stylized motion that adheres to both content and style conditions, as illustrated in Fig. 4(c)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 368, + 482, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 368, + 482, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 368, + 482, + 453 + ], + "type": "text", + "content": "Classifier-based Style Guidance. To further improve the stylization of a motion diffusion model, we adopt the classifier guidance [10,59] to provide stronger guidance to the generated motion towards the desired style. The core of our classifier-based style guidance is a novel analytic function " + }, + { + "bbox": [ + 130, + 368, + 482, + 453 + ], + "type": "inline_equation", + "content": "G(\\mathbf{z}_t,t,\\mathbf{s})" + }, + { + "bbox": [ + 130, + 368, + 482, + 453 + ], + "type": "text", + "content": ", which calculates the " + }, + { + "bbox": [ + 130, + 368, + 482, + 453 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 130, + 368, + 482, + 453 + ], + "type": "text", + "content": " distance between the style embedding of the generated clean motion " + }, + { + "bbox": [ + 130, + 368, + 482, + 453 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}}_0" + }, + { + "bbox": [ + 130, + 368, + 482, + 453 + ], + "type": "text", + "content": " at denoising step " + }, + { + "bbox": [ + 130, + 368, + 482, + 453 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 368, + 482, + 453 + ], + "type": "text", + "content": " and the reference style motions " + }, + { + "bbox": [ + 130, + 368, + 482, + 453 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 130, + 368, + 482, + 453 + ], + "type": "text", + "content": ". The function's gradient is utilized to steer the generated motion towards the desired style." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 209, + 464, + 406, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 464, + 406, + 475 + ], + "spans": [ + { + "bbox": [ + 209, + 464, + 406, + 475 + ], + "type": "interline_equation", + "content": "\\epsilon_ {\\boldsymbol {\\theta}} (\\boldsymbol {z} _ {t}, t, \\mathbf {c}, \\mathbf {s}) = \\epsilon_ {\\boldsymbol {\\theta}} (\\boldsymbol {z} _ {t}, t, \\mathbf {c}, \\mathbf {s}) + \\tau \\nabla_ {\\boldsymbol {z} _ {t}} G (\\boldsymbol {z} _ {t}, t, \\mathbf {s}),", + "image_path": "93426a680695985775cd707bb55e1a704101407460d62e780de2d8d509f137d7.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 219, + 477, + 480, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 477, + 480, + 491 + ], + "spans": [ + { + "bbox": [ + 219, + 477, + 480, + 491 + ], + "type": "interline_equation", + "content": "G \\left(\\boldsymbol {z} _ {t}, t, \\mathbf {s}\\right) = | f \\left(\\hat {\\boldsymbol {x}} _ {0}\\right) - f (\\mathbf {s}) |, \\tag {2}", + "image_path": "b0624eda0caad9293e5e805df3e94b5ae3ab3672ebcbaa28ec92fe0cfdc06e42.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 500, + 482, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 500, + 482, + 537 + ], + "spans": [ + { + "bbox": [ + 131, + 500, + 482, + 537 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 500, + 482, + 537 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 131, + 500, + 482, + 537 + ], + "type": "text", + "content": " adjusts the strength of reference-style guidance and " + }, + { + "bbox": [ + 131, + 500, + 482, + 537 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 131, + 500, + 482, + 537 + ], + "type": "text", + "content": " denotes the style feature extractor. The generated motion " + }, + { + "bbox": [ + 131, + 500, + 482, + 537 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}}_0" + }, + { + "bbox": [ + 131, + 500, + 482, + 537 + ], + "type": "text", + "content": " is obtained by first converting the denoising output latent " + }, + { + "bbox": [ + 131, + 500, + 482, + 537 + ], + "type": "inline_equation", + "content": "\\pmb{z}_t" + }, + { + "bbox": [ + 131, + 500, + 482, + 537 + ], + "type": "text", + "content": " into the predicted clean latent as shown below:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 238, + 546, + 481, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 546, + 481, + 572 + ], + "spans": [ + { + "bbox": [ + 238, + 546, + 481, + 572 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {z}} _ {0} = \\frac {\\boldsymbol {z} _ {t} - \\sqrt {1 - \\alpha_ {t}} \\varepsilon_ {\\theta} (\\boldsymbol {z} _ {t} , t , \\boldsymbol {c} , \\boldsymbol {s})}{\\sqrt {\\alpha_ {t}}}, \\tag {3}", + "image_path": "8d18e931fbf9c84d095079381d78ed13b382e136ed652b8ea008a3940d60a3cb.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 581, + 481, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 481, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 481, + 616 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 581, + 481, + 616 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 130, + 581, + 481, + 616 + ], + "type": "text", + "content": " denotes the pre-defined noise scale in the forward process of the diffusion model. The predicted clean latent " + }, + { + "bbox": [ + 130, + 581, + 481, + 616 + ], + "type": "inline_equation", + "content": "\\hat{z}_0" + }, + { + "bbox": [ + 130, + 581, + 481, + 616 + ], + "type": "text", + "content": " is then input into the motion decoder " + }, + { + "bbox": [ + 130, + 581, + 481, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 130, + 581, + 481, + 616 + ], + "type": "text", + "content": " to obtain the generated motion." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "We obtain the style feature extractor by training a style classifier on the 100STYLE dataset [27] and removing its final layer. We refer the readers to more details in the supplementary materials. The training of the style feature extractor with ground-truth style labels for supervision enables it to effectively" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 276, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 276, + 91, + 447, + 102 + ], + "type": "text", + "content": "SMooDi: Stylized Motion Diffusion Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 130, + 462, + 184 + ], + "blocks": [ + { + "bbox": [ + 202, + 118, + 386, + 129 + ], + "lines": [ + { + "bbox": [ + 202, + 118, + 386, + 129 + ], + "spans": [ + { + "bbox": [ + 202, + 118, + 386, + 129 + ], + "type": "text", + "content": "Text: A person walks forward and then sits down." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 149, + 130, + 462, + 184 + ], + "lines": [ + { + "bbox": [ + 149, + 130, + 462, + 184 + ], + "spans": [ + { + "bbox": [ + 149, + 130, + 462, + 184 + ], + "type": "image", + "image_path": "7b64e7f402554b1ab66ab627b226f6f54cfa96271300186a8592e72263d86444.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 194, + 482, + 240 + ], + "lines": [ + { + "bbox": [ + 130, + 194, + 482, + 240 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 482, + 240 + ], + "type": "text", + "content": "Fig. 4: Visual illustrations of the classifier-free and classifier-based style guidance. (a) and (b) respectively show the classifier-free content and style guidance; (c) displays the initial stylized motion resulting from the combination of (a) and (b); (d) illustrates the refined stylized motion modified by the classifier-based style guidance." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 262, + 479, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 262, + 479, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 479, + 285 + ], + "type": "text", + "content": "capture style-related features. Therefore, style classifier guidance can provide more guidance to the stylized motion generation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 285, + 480, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 480, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 480, + 416 + ], + "type": "text", + "content": "Combination of the Two Style Guidance. The classifier-free and classifier-based style guidance are designed to complement each other, each playing a vital role in accurately reflecting the target style in the generated motions. As illustrated in Fig. 4, the desired style motion is \"arms open wide to the sides like an airplane\". The classifier-free style guidance (Fig. 4(b)) can capture style-related characteristic in a reasonably accurate manner. When combined with the classifier-free content guidance, it depicts the desired style (Fig.4(c)). Refined further by the classifier-based style guidance, the stylization is more authentic, where the person's harms are more open (Fig.4(d)). In addition to such visual results, quantitative ablation studies also verify the effectiveness of our proposed both classifier-free and classifier-based style guidance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 417, + 482, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 417, + 482, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 417, + 482, + 536 + ], + "type": "text", + "content": "At the same time, although classifier-based style guidance offers precise style control, its effectiveness may be compromised when the content text significantly diverges from locomotion-related movements. This is because the style feature function is trained solely on the 100STYLE dataset, which contains only such movements. Over-reliance on style classifier guidance risks producing motions that fail to execute the desired actions, leading to unrealistic and physically implausible movements. Therefore, we leverage a content-aware style adaptor that establishes the fundamental style direction, while style classifier guidance refines this base for a more precise outcome. The effectiveness of this design is verified in our ablation studies." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 552, + 248, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 552, + 248, + 565 + ], + "spans": [ + { + "bbox": [ + 131, + 552, + 248, + 565 + ], + "type": "text", + "content": "3.3 Learning Scheme" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 571, + 480, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 571, + 480, + 608 + ], + "spans": [ + { + "bbox": [ + 130, + 571, + 480, + 608 + ], + "type": "text", + "content": "Following [61], a straightforward approach to train SMooDi is to freeze the parameters of MLD and solely train the style adaptor on the 100STYLE dataset using the following loss function:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 231, + 614, + 481, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 614, + 481, + 635 + ], + "spans": [ + { + "bbox": [ + 231, + 614, + 481, + 635 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s t d} = \\mathbb {E} _ {\\epsilon , \\mathbf {z}} \\left[ \\| \\epsilon_ {\\theta} (\\mathbf {z} _ {t}, t, \\mathbf {c}, \\mathbf {s}) - \\epsilon \\| _ {2} ^ {2} \\right], \\tag {4}", + "image_path": "8c8aca2eba7fc73083e8e4856ec18b1d98b0d2d7b7d0a8bddf82af54261fbc09.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0, \\mathbf{I})" + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": " represents the ground-truth noise added to " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": ". In our experiments, however, we found that this loss function alone leads to an issue of" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Zhong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": "\"content-forgetting\", where the model progressively looses the MLD's ability to generate motions with diverse contents. To address this issue, we design a content prior preservation loss " + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "inline_equation", + "content": "L_{pr}" + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": ". Specifically, we randomly sample motions from the HumanML3D dataset to compute a prior preservation loss when fine-tuning SMooDi on the 100STYLE dataset." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 228, + 182, + 481, + 202 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 182, + 481, + 202 + ], + "spans": [ + { + "bbox": [ + 228, + 182, + 481, + 202 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {p r} = \\mathbb {E} _ {\\epsilon^ {\\prime}, \\boldsymbol {z} ^ {\\prime}} \\left[ \\| \\epsilon_ {\\theta} (\\boldsymbol {z} _ {t} ^ {\\prime}, t, \\mathbf {c} ^ {\\prime}, \\mathbf {s} ^ {\\prime}) - \\epsilon^ {\\prime} \\| _ {2} ^ {2} \\right], \\tag {5}", + "image_path": "85ec5b3d5b4daa8507842f628f0ccea23d15389615e112c10e4525a940c798c7.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 209, + 482, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 209, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 209, + 482, + 293 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 209, + 482, + 293 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t^\\prime" + }, + { + "bbox": [ + 130, + 209, + 482, + 293 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 209, + 482, + 293 + ], + "type": "inline_equation", + "content": "\\mathbf{c}^\\prime" + }, + { + "bbox": [ + 130, + 209, + 482, + 293 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 209, + 482, + 293 + ], + "type": "inline_equation", + "content": "\\mathbf{s}^\\prime" + }, + { + "bbox": [ + 130, + 209, + 482, + 293 + ], + "type": "text", + "content": " represents the motion latent, content prompt and style motion sequence derived from the HumanML3D dataset. " + }, + { + "bbox": [ + 130, + 209, + 482, + 293 + ], + "type": "inline_equation", + "content": "\\epsilon^\\prime" + }, + { + "bbox": [ + 130, + 209, + 482, + 293 + ], + "type": "text", + "content": " is the noise map added to " + }, + { + "bbox": [ + 130, + 209, + 482, + 293 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0^\\prime" + }, + { + "bbox": [ + 130, + 209, + 482, + 293 + ], + "type": "text", + "content": ". A similar solution is used in DreamBooth [40] to solve the \"language drift\" problem, where images generated from the frozen pretrained image generation model are utilized to enforce a class-prior preservation loss during model finetuning. Our content preservation loss can effectively mitigate content forgetting while learning diverse motion styles from the 100STYLE dataset." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 293, + 482, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 388 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 388 + ], + "type": "text", + "content": "To further encourage the style adaptor to focus on motion style, while also ensuring that the latent diffusion model in MLD handles motion content well, we introduce an additional cycle prior-preservation loss, inspired by [19,55]. Specifically, we start this process by randomly sampling content text and motion style sequences from both the 100STYLE and HumanML3D datasets simultaneously. Then, we intermix the content text and motion style from these sequences with each other. Finally, we repeat this process to reconstruct the original motion sequences. The formula is expressed as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 165, + 397, + 481, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 397, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 165, + 397, + 481, + 418 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c y c} = \\mathbb {E} _ {\\boldsymbol {z}, \\boldsymbol {z} ^ {\\prime}, \\epsilon , \\epsilon^ {\\prime}} \\left[ \\left\\| \\epsilon_ {\\theta} (\\boldsymbol {z} _ {t} ^ {s h}, t, \\mathbf {c}, \\boldsymbol {s} ^ {h s}) + \\epsilon_ {\\theta} (\\boldsymbol {z} _ {t} ^ {h s}, t, \\mathbf {c} ^ {\\prime}, \\boldsymbol {s} ^ {s h}) - \\epsilon - \\epsilon^ {\\prime} \\right\\| _ {2} ^ {2} \\right], \\qquad (6)", + "image_path": "2de8ac84f5a8750baa10c90dca583d12d6999fdf367bf3a7a288beb5f4579d02.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 426, + 482, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 532 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 532 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 426, + 482, + 532 + ], + "type": "inline_equation", + "content": "s^{hs}" + }, + { + "bbox": [ + 130, + 426, + 482, + 532 + ], + "type": "text", + "content": " denotes the motion sequence created by merging content from the HumanML3D dataset with style from the 100STYLE dataset. Similarly, " + }, + { + "bbox": [ + 130, + 426, + 482, + 532 + ], + "type": "inline_equation", + "content": "s^{sh}" + }, + { + "bbox": [ + 130, + 426, + 482, + 532 + ], + "type": "text", + "content": " represents the sequence where content is sourced from the 100STYLE dataset and style from the HumanML3D dataset. The noised latent codes " + }, + { + "bbox": [ + 130, + 426, + 482, + 532 + ], + "type": "inline_equation", + "content": "z_{t}^{sh}" + }, + { + "bbox": [ + 130, + 426, + 482, + 532 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 426, + 482, + 532 + ], + "type": "inline_equation", + "content": "z_{t}^{hs}" + }, + { + "bbox": [ + 130, + 426, + 482, + 532 + ], + "type": "text", + "content": " correspond to " + }, + { + "bbox": [ + 130, + 426, + 482, + 532 + ], + "type": "inline_equation", + "content": "s^{sh}" + }, + { + "bbox": [ + 130, + 426, + 482, + 532 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 426, + 482, + 532 + ], + "type": "inline_equation", + "content": "s^{hs}" + }, + { + "bbox": [ + 130, + 426, + 482, + 532 + ], + "type": "text", + "content": ", respectively. Essentially, the cycle prior-preservation loss exchanges diverse content and style between two datasets, encouraging the content text to remain invariant in the generated motion under forward and backward translation. Overall, the training loss function of our framework is defined as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 238, + 534, + 480, + 547 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 534, + 480, + 547 + ], + "spans": [ + { + "bbox": [ + 238, + 534, + 480, + 547 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a l l}} = \\mathcal {L} _ {\\text {s t d}} + \\lambda_ {p r} \\mathcal {L} _ {p r} + \\lambda_ {\\text {c y c}} \\mathcal {L} _ {\\text {c y c}} \\tag {7}", + "image_path": "94f6ab9a152649324b914072f79169c293ddd21f7bfadaaa3c2c4fb3ecf1817c.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 551, + 482, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 551, + 482, + 575 + ], + "spans": [ + { + "bbox": [ + 130, + 551, + 482, + 575 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 551, + 482, + 575 + ], + "type": "inline_equation", + "content": "\\lambda_{pr}" + }, + { + "bbox": [ + 130, + 551, + 482, + 575 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 551, + 482, + 575 + ], + "type": "inline_equation", + "content": "\\lambda_{cyc}" + }, + { + "bbox": [ + 130, + 551, + 482, + 575 + ], + "type": "text", + "content": " are hyperparameters. We refer readers to the pseudocode and illustration for training in the supplementary material for more details." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 593, + 230, + 606 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 593, + 230, + 606 + ], + "spans": [ + { + "bbox": [ + 132, + 593, + 230, + 606 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "We conduct experiments on both stylized text2motion and motion style transfer to demonstrate the effectiveness of our framework. Both tasks use a motion sequence as a style prompt, with the primary difference being their content prompt input: the former utilizes text, while the latter relies on motion sequences." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 276, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 276, + 91, + 447, + 102 + ], + "type": "text", + "content": "SMooDi: Stylized Motion Diffusion Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 160, + 157, + 454, + 224 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 1: Comparison with baseline methods on stylized motion generation driven by content text, using a combination of the 100STYLE (providing style) and HumanML3D datasets (providing content)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 160, + 157, + 454, + 224 + ], + "lines": [ + { + "bbox": [ + 160, + 157, + 454, + 224 + ], + "spans": [ + { + "bbox": [ + 160, + 157, + 454, + 224 + ], + "type": "table", + "html": "
MethodFID ↓ Foot skatingMM Dist ↓ R-precision ↑ Diversity → SRA ↑ ratio ↓
Ours1.6090.1244.4770.5719.23572.418
MLD+Motion Puzzle [19]6.1270.1856.4670.2906.476263.769
MLD+Aberman et al. [1]3.3090.3475.9830.4068.81654.367
ChatGPT+MLD0.6140.1314.3130.6058.8364.819
", + "image_path": "8bc8b194ccbd0264c168f138b088101389be86adb4f9771a80ea88a9e764ddb3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 153, + 240, + 220, + 292 + ], + "blocks": [ + { + "bbox": [ + 153, + 240, + 220, + 292 + ], + "lines": [ + { + "bbox": [ + 153, + 240, + 220, + 292 + ], + "spans": [ + { + "bbox": [ + 153, + 240, + 220, + 292 + ], + "type": "image", + "image_path": "8a9bb57b0d85d341d5bb052f1244e30c7e415cee0faba7aa1771752ae6c75e3b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 221, + 260, + 271, + 269 + ], + "lines": [ + { + "bbox": [ + 221, + 260, + 271, + 269 + ], + "spans": [ + { + "bbox": [ + 221, + 260, + 271, + 269 + ], + "type": "text", + "content": "A person kicks." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 282, + 243, + 325, + 285 + ], + "blocks": [ + { + "bbox": [ + 282, + 243, + 325, + 285 + ], + "lines": [ + { + "bbox": [ + 282, + 243, + 325, + 285 + ], + "spans": [ + { + "bbox": [ + 282, + 243, + 325, + 285 + ], + "type": "image", + "image_path": "5690f5675f7815d3acc4f2ec27d669e871e5bb5f76569ffeba5ad839e218ba7a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 281, + 285, + 392, + 292 + ], + "lines": [ + { + "bbox": [ + 281, + 285, + 392, + 292 + ], + "spans": [ + { + "bbox": [ + 281, + 285, + 392, + 292 + ], + "type": "text", + "content": "(a) " + }, + { + "bbox": [ + 281, + 285, + 392, + 292 + ], + "type": "inline_equation", + "content": "\\mathrm{MLD + }" + }, + { + "bbox": [ + 281, + 285, + 392, + 292 + ], + "type": "text", + "content": " Motion Puzzle (b) " + }, + { + "bbox": [ + 281, + 285, + 392, + 292 + ], + "type": "inline_equation", + "content": "\\mathrm{MLD + }" + }, + { + "bbox": [ + 281, + 285, + 392, + 292 + ], + "type": "text", + "content": " Aberman et al." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 337, + 244, + 392, + 286 + ], + "blocks": [ + { + "bbox": [ + 337, + 244, + 392, + 286 + ], + "lines": [ + { + "bbox": [ + 337, + 244, + 392, + 286 + ], + "spans": [ + { + "bbox": [ + 337, + 244, + 392, + 286 + ], + "type": "image", + "image_path": "d20b7f278bef388d97796ed81374f419ea0c4a86bbdb00444dfa53071d2365c8.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 400, + 244, + 446, + 285 + ], + "blocks": [ + { + "bbox": [ + 400, + 244, + 446, + 285 + ], + "lines": [ + { + "bbox": [ + 400, + 244, + 446, + 285 + ], + "spans": [ + { + "bbox": [ + 400, + 244, + 446, + 285 + ], + "type": "image", + "image_path": "ba94ee848bc02ff1ac8a9219c441a35904447cfdf8276e3c99011f2b9ceb7ddd.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 409, + 285, + 429, + 292 + ], + "lines": [ + { + "bbox": [ + 409, + 285, + 429, + 292 + ], + "spans": [ + { + "bbox": [ + 409, + 285, + 429, + 292 + ], + "type": "text", + "content": "(c) Ours" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 153, + 294, + 217, + 343 + ], + "blocks": [ + { + "bbox": [ + 153, + 294, + 217, + 343 + ], + "lines": [ + { + "bbox": [ + 153, + 294, + 217, + 343 + ], + "spans": [ + { + "bbox": [ + 153, + 294, + 217, + 343 + ], + "type": "image", + "image_path": "64b79eba7852a8af79637162ebaf2c7c698a869850f7b9c5b3ab9835c71c04b7.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 356, + 482, + 378 + ], + "lines": [ + { + "bbox": [ + 131, + 356, + 482, + 378 + ], + "spans": [ + { + "bbox": [ + 131, + 356, + 482, + 378 + ], + "type": "text", + "content": "Fig. 5: Qualitative comparisons of our approach and baseline methods on two stylized motion generation task." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 223, + 294, + 272, + 343 + ], + "blocks": [ + { + "bbox": [ + 223, + 294, + 272, + 343 + ], + "lines": [ + { + "bbox": [ + 223, + 294, + 272, + 343 + ], + "spans": [ + { + "bbox": [ + 223, + 294, + 272, + 343 + ], + "type": "image", + "image_path": "0c83bd5658cd74abe2de00e7783896b233f1194aef387151aa646e2baf18b223.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 282, + 294, + 337, + 344 + ], + "blocks": [ + { + "bbox": [ + 282, + 294, + 337, + 344 + ], + "lines": [ + { + "bbox": [ + 282, + 294, + 337, + 344 + ], + "spans": [ + { + "bbox": [ + 282, + 294, + 337, + 344 + ], + "type": "image", + "image_path": "635b950c9548f5ef96be20ee1baaac4f7ad758840f926e21b03dc39487b7f76b.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 345, + 294, + 394, + 344 + ], + "blocks": [ + { + "bbox": [ + 345, + 294, + 394, + 344 + ], + "lines": [ + { + "bbox": [ + 345, + 294, + 394, + 344 + ], + "spans": [ + { + "bbox": [ + 345, + 294, + 394, + 344 + ], + "type": "image", + "image_path": "4a540e666fc0cbc5de51165940de3a0905c63a35a52da71294eb2de9ab55ef05.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 398, + 294, + 451, + 343 + ], + "blocks": [ + { + "bbox": [ + 398, + 294, + 451, + 343 + ], + "lines": [ + { + "bbox": [ + 398, + 294, + 451, + 343 + ], + "spans": [ + { + "bbox": [ + 398, + 294, + 451, + 343 + ], + "type": "image", + "image_path": "58c685bcf66f0b58dad0b423be5d3619937b4c86c1e0669d46f06a3acab567b9.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 402, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 581 + ], + "type": "text", + "content": "Datasets. We utilize the HumanML3D dataset [16] as our motion content dataset and the 100STYLE dataset [27] as our motion style dataset. The HumanML3D dataset is the largest motion capture dataset, featuring text annotations and comprising 14,646 motions and 44,970 motion annotations. Following the processing approach outlined in [16], we preprocess the HumanML3D dataset to obtain consistent motion representations. On the other hand, the 100STYLE dataset [27], being the largest motion style dataset, comprises up to 1,125 minutes of motion sequences, showcasing a wide array of 100 diverse locomotion styles. Due to differences in skeletons between the 100STYLE dataset and HumanML3D, we retarget the motions from 100STYLE to match the HumanML3D (SMPL-H) skeleton. Following this alignment, we apply the same processing steps as used for the HumanML3D dataset to preprocess the 100STYLE dataset. Moreover, as the 100STYLE dataset lacks text descriptions, we leverage MotionGPT [20] to generate pseudo text descriptions for the motion sequences in the 100STYLE dataset." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Evaluation metrics are designed to assess three dimensions: Content Preservation, Style Reflection, and Realism. For content preservation and style reflection assessment, we employ metrics consistent with those used in [6]: motion-retrieval precision (R precision), Multi-modal Distance (MM Dist), Diversity, and Frechet Inception Distance (FID). Additionally, recognizing the common foot skating issues in kinetics-based motion generation methods, we incorporate the foot skating ratio metric proposed by [23] into our motion quality evaluation. For style" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Zhong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "content": "reflection, we employ Style Recognition Accuracy (SRA) [19]. During evaluation, we randomly select a content text from the HumanML3D dataset and a motion style sequence from the 100STYLE dataset to generate the stylized motion. We then use a pre-trained style classifier to compute the SRA for the generated motion. It's noteworthy that some motion style labels in the 100STYLE dataset, like 'kick' and 'jump,' inherently convey motion content, which may conflict with the content text in HumanML3D dataset. To address this, we categorize the motion style labels into server groups following the approach by Kim et al. [24], Specifically excluding the 'ACT' group ensures that only motion style labels not conflicting with motion content are considered when computing the SRA metric. Further details about the evaluation metrics are provided in the supplementary material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 261, + 482, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 261, + 482, + 547 + ], + "spans": [ + { + "bbox": [ + 130, + 261, + 482, + 547 + ], + "type": "text", + "content": "Baselines. For motion style transfer task, we compare our methods with two state-of-the-art methods, namely Motion Puzzle [19] and Aberman et al. [1]. To ensure a fair comparison, we train the compared methods under the same settings as ours, using a combined dataset comprising HumanML3D and 100STYLE. Due to the constraints of the multi-class discriminator in Aberman et al., which requires style labels, we adopt the training method outlined in Motion Puzzle to eliminate the need for style labels. For stylized text2motion task, we compare our method against baselines capable of generating stylized motion from content text and style motion sequences. The straightforward baselines involve applying motion style transfer methods to the motion sequences generated by the text2motion model. To align with our approach that uses a pre-trained motion diffusion model, the text2motion models in the baselines for stylized motion generation select MLD, and the motion style transfer methods are consistent with those used in the motion style transfer task. For the stylized text2motion task, we compare our method against two kinds of baselines capable of generating stylized motion from content text and style motion sequences. The first kind of baseline involves applying motion style transfer methods to the motion sequences generated by the text2motion model. To align with our approach that uses a pretrained motion diffusion model, the text2motion models select MLD [6], and the motion style transfer methods are consistent with those used in the motion style transfer task. The second kind of baseline involves using ChatGPT to merge style labels from 100STYLE with text from HumanML3D into a sentence. For example, given the content text 'a person walks.' and the style label 'old,' we obtain 'an elderly person walks.' This merged sentence is then fed to MLD." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 563, + 328, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 563, + 328, + 574 + ], + "spans": [ + { + "bbox": [ + 131, + 563, + 328, + 574 + ], + "type": "text", + "content": "4.1 Comparison to Baseline Methods" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 582, + 479, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 479, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 479, + 604 + ], + "type": "text", + "content": "Quantitative and Qualitative For the task of stylized text2motion, Table 1 reports the comparisons of our method with the three baseline methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 605, + 481, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 664 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 664 + ], + "type": "text", + "content": "As shown in the " + }, + { + "bbox": [ + 130, + 605, + 481, + 664 + ], + "type": "inline_equation", + "content": "3^{rd}" + }, + { + "bbox": [ + 130, + 605, + 481, + 664 + ], + "type": "text", + "content": " row of Table 1, ChatGPT+MLD only achieves around " + }, + { + "bbox": [ + 130, + 605, + 481, + 664 + ], + "type": "inline_equation", + "content": "5.29\\%" + }, + { + "bbox": [ + 130, + 605, + 481, + 664 + ], + "type": "text", + "content": " in terms of SRA, indicating that MLD cannot enable stylized generation from text alone, even though it contains style descriptions. Notably, our method outperforms the two baselines that combine MLD with motion style transfer methods in all metrics." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 276, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 276, + 91, + 447, + 102 + ], + "type": "text", + "content": "SMooDi: Stylized Motion Diffusion Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 149, + 145, + 292, + 156 + ], + "lines": [ + { + "bbox": [ + 149, + 145, + 292, + 156 + ], + "spans": [ + { + "bbox": [ + 149, + 145, + 292, + 156 + ], + "type": "text", + "content": "(a) Evaluation on HumanML3D dataset" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "table", + "bbox": [ + 146, + 163, + 294, + 215 + ], + "blocks": [ + { + "bbox": [ + 163, + 114, + 449, + 126 + ], + "lines": [ + { + "bbox": [ + 163, + 114, + 449, + 126 + ], + "spans": [ + { + "bbox": [ + 163, + 114, + 449, + 126 + ], + "type": "text", + "content": "Table 2: Comparison with baseline methods on motion style transfer." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 146, + 163, + 294, + 215 + ], + "lines": [ + { + "bbox": [ + 146, + 163, + 294, + 215 + ], + "spans": [ + { + "bbox": [ + 146, + 163, + 294, + 215 + ], + "type": "table", + "html": "
MethodFoot skating ratio ↓
Ours0.0951.58265.147
Motion Puzzle [19]0.1976.87167.233
(Aberman et al [1])0.3383.89261.006
", + "image_path": "1866ec59a6e555e9a0d5f77a9fac17ce646ad284c17370c40d7e018957c94724.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 307, + 163, + 482, + 214 + ], + "blocks": [ + { + "bbox": [ + 340, + 145, + 447, + 155 + ], + "lines": [ + { + "bbox": [ + 340, + 145, + 447, + 155 + ], + "spans": [ + { + "bbox": [ + 340, + 145, + 447, + 155 + ], + "type": "text", + "content": "(b) Evaluation on Xia dataset" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 163, + 482, + 214 + ], + "lines": [ + { + "bbox": [ + 307, + 163, + 482, + 214 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 482, + 214 + ], + "type": "table", + "html": "
MethodFoot skating ratio↓FID↓SRA↑(%)CRA↑(%)
Ours0.03174.66361.11145.555
Motion Puzzle [19]0.03165.36067.77825.556
(Aberman et al [1])0.02605.68156.66734.444
", + "image_path": "dcf9f82e01b7034c50ee9b03d78220f3c51b5f91ae71003d289fa8f05e69e858.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 242, + 480, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 242, + 480, + 338 + ], + "spans": [ + { + "bbox": [ + 130, + 242, + 480, + 338 + ], + "type": "text", + "content": "Specifically, our method performs better than MLD+Motion Puzzle in the SRA metric by " + }, + { + "bbox": [ + 130, + 242, + 480, + 338 + ], + "type": "inline_equation", + "content": "13.56\\%" + }, + { + "bbox": [ + 130, + 242, + 480, + 338 + ], + "type": "text", + "content": " and significantly outperforms MLD+Aberman et al. in the FID metric by " + }, + { + "bbox": [ + 130, + 242, + 480, + 338 + ], + "type": "inline_equation", + "content": "51.38\\%" + }, + { + "bbox": [ + 130, + 242, + 480, + 338 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 242, + 480, + 338 + ], + "type": "inline_equation", + "content": "0.64\\%" + }, + { + "bbox": [ + 130, + 242, + 480, + 338 + ], + "type": "text", + "content": " in the R-precision metric. The first row of Fig. 5 validates our observation, where the motion generated by our method performs better in adhering to both content and style constraints than baseline methods. In contrast, MLP+Aberman et al. can successfully perform the action but fail to reflect the motion style in Fig.5(b), while Motion Puzzle can accurately reflect the motion style but struggles to effectively perform the action in Fig.5(a)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 340, + 482, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 482, + 626 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 482, + 626 + ], + "type": "text", + "content": "For the task of motion style transfer, since it does not take content text as input, text-motion related metrics such as MM Dist, R-precision, and Diversity are not applicable and thus are not reported. Part (a) of Table 2 presents a comparison between our method and the two baseline methods, using the HumanML3D dataset as the motion content source and drawing motion styles from the 100STYLE dataset. Our method delivers competitive results in the SRA metric and excels in the FID and foot skating ratio metrics. Specifically, we see a substantial " + }, + { + "bbox": [ + 130, + 340, + 482, + 626 + ], + "type": "inline_equation", + "content": "59.35\\%" + }, + { + "bbox": [ + 130, + 340, + 482, + 626 + ], + "type": "text", + "content": " improvement in the FID metric over Aberman et al. [1]. To more effectively compare the generalizability of different methods, we conduct experiments on the Xia dataset [52], a small, specific motion style dataset that was unseen by our and the baseline models during training. Because motion content labels are present in the Xia dataset, we report the Content Recognition Accuracy (CRA). Part (b) of Table 2 showcases the results. Our method maintains competitive performance in the SRA metric, with only a marginal " + }, + { + "bbox": [ + 130, + 340, + 482, + 626 + ], + "type": "inline_equation", + "content": "9.83\\%" + }, + { + "bbox": [ + 130, + 340, + 482, + 626 + ], + "type": "text", + "content": " decrease in SRA compared to Motion Puzzle. On the contrary, our method exhibits a significant " + }, + { + "bbox": [ + 130, + 340, + 482, + 626 + ], + "type": "inline_equation", + "content": "32.26\\%" + }, + { + "bbox": [ + 130, + 340, + 482, + 626 + ], + "type": "text", + "content": " increase in the CAR metric relative to Aberman et al., and a notable " + }, + { + "bbox": [ + 130, + 340, + 482, + 626 + ], + "type": "inline_equation", + "content": "78.26\\%" + }, + { + "bbox": [ + 130, + 340, + 482, + 626 + ], + "type": "text", + "content": " enhancement over Motion Puzzle. Our method achieves a better balance between style reflection and content preservation. The second row of Fig. 5 validates this observation. It is worth noting that, unlike other motion style transfer methods, our method does not incorporate objectives for enabling stylized motion generation using motion content sequences. Through simple DDIM-Inversion and without any additional optimization or regularization, our method achieves performance comparable to existing motion style transfer methods." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": "User Study. Due to the highly subjective nature of stylized motion, we conduct User studies using pairwise comparisons to further evaluate our proposed method in the tasks of stylized motion generation and motion style transfer. We" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Zhong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 177, + 135, + 438, + 211 + ], + "blocks": [ + { + "bbox": [ + 150, + 114, + 462, + 126 + ], + "lines": [ + { + "bbox": [ + 150, + 114, + 462, + 126 + ], + "spans": [ + { + "bbox": [ + 150, + 114, + 462, + 126 + ], + "type": "text", + "content": "Table 3: Ablation Studies on HumanML3D Content and 100STYLE Styles." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 177, + 135, + 438, + 211 + ], + "lines": [ + { + "bbox": [ + 177, + 135, + 438, + 211 + ], + "spans": [ + { + "bbox": [ + 177, + 135, + 438, + 211 + ], + "type": "table", + "html": "
MethodFID↓ ratio↓MM Dist↓R-precision↑ (Top-3)Diversity→ SRA(%)↑
Ours (on all)1.6090.1244.4770.5719.23572.418
w/o Lcyc2.0460.1364.4650.5698.86964.866
w/o Lpr + Lcyc5.9960.1666.0980.3357.45681.841
w/o classifier-based1.0500.1114.0850.6309.44520.245
w/o adaptor2.9840.1234.5260.5508.37269.952
", + "image_path": "78c81354b646e1c0ff69593bd136bcabf93540288b16b965e256231df98fe389.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 206, + 223, + 378, + 233 + ], + "lines": [ + { + "bbox": [ + 206, + 223, + 378, + 233 + ], + "spans": [ + { + "bbox": [ + 206, + 223, + 378, + 233 + ], + "type": "text", + "content": "Text: A person walks forward and then sits down." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 149, + 233, + 457, + 285 + ], + "blocks": [ + { + "bbox": [ + 149, + 233, + 457, + 285 + ], + "lines": [ + { + "bbox": [ + 149, + 233, + 457, + 285 + ], + "spans": [ + { + "bbox": [ + 149, + 233, + 457, + 285 + ], + "type": "image", + "image_path": "9cc5618563a75e1af0ddadb6dcabde399b405b079118742d11e117b59ea4d369.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 157, + 294, + 456, + 306 + ], + "lines": [ + { + "bbox": [ + 157, + 294, + 456, + 306 + ], + "spans": [ + { + "bbox": [ + 157, + 294, + 456, + 306 + ], + "type": "text", + "content": "Fig. 6: Visual comparisons of the ablation designs and our full model." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 329, + 482, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 329, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 329, + 482, + 413 + ], + "type": "text", + "content": "recruited 22 human subjects to participate in the study. In each test, participants are presented with two 4-second video clips synthesized by our method and one comparison method. They are then required to select their preferred clip while considering Realism, Style Reflection, and Content Preservation dimensions, respectively. As shown in Fig. 7, our method receives more user appreciation compared to two baselines across three dimensions in two tasks. Further user study details are provided in supplementary." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 430, + 244, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 244, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 244, + 441 + ], + "type": "text", + "content": "4.2 Ablation studies" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 450, + 481, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 481, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 481, + 498 + ], + "type": "text", + "content": "To validate the effectiveness of our framework's design choices, we have conducted several ablation studies: the first assesses the impact of each loss function term, while the second evaluates the influence of the style adaptor and style guidance during sampling." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 498, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 581 + ], + "type": "text", + "content": "Loss Components. Firstly, we exclude the cycle-prior term in the loss function, denoted as " + }, + { + "bbox": [ + 130, + 498, + 482, + 581 + ], + "type": "inline_equation", + "content": "w / o" + }, + { + "bbox": [ + 130, + 498, + 482, + 581 + ], + "type": "inline_equation", + "content": "L_{cyc}" + }, + { + "bbox": [ + 130, + 498, + 482, + 581 + ], + "type": "text", + "content": ". Comparing the results in the " + }, + { + "bbox": [ + 130, + 498, + 482, + 581 + ], + "type": "inline_equation", + "content": "1^{\\mathrm{st}}" + }, + { + "bbox": [ + 130, + 498, + 482, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 498, + 482, + 581 + ], + "type": "inline_equation", + "content": "2^{\\mathrm{nd}}" + }, + { + "bbox": [ + 130, + 498, + 482, + 581 + ], + "type": "text", + "content": " rows in Table 3, we observe that our full model outperforms in all content preservation and style reflection metrics. The motion generated by our approach can still perform the content adhering to the text description but performs worse in accurately reflecting the motion style, as reflected by the arms not being fully extended horizontally." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "text", + "content": "Since the cycle prior-preservation term is built upon the prior-preservation term, it is meaningless to exclude " + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "inline_equation", + "content": "L_{pr}" + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "text", + "content": " while retaining " + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "inline_equation", + "content": "L_{cyc}" + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "text", + "content": ". Therefore, we further exclude both the prior-preservation and cycle-prior term, denoted as " + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "inline_equation", + "content": "w / oL_{pr} + L_{cyc}" + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "inline_equation", + "content": "3^{\\mathrm{rd}}" + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "text", + "content": " row. By comparing the results in the " + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "inline_equation", + "content": "2^{\\mathrm{nd}}" + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "inline_equation", + "content": "3^{\\mathrm{rd}}" + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "text", + "content": " rows of Table 3, we notice that while the number of SRAs is higher in the third row, other metrics show a significant decline. Specifically, in terms of the FID metric, performance deteriorates by more than " + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "inline_equation", + "content": "229\\%" + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "text", + "content": ". Indeed, without " + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "inline_equation", + "content": "L_{pr}" + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "text", + "content": ", the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 276, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 276, + 91, + 447, + 102 + ], + "type": "text", + "content": "SMooDi: Stylized Motion Diffusion Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 167, + 116, + 303, + 214 + ], + "blocks": [ + { + "bbox": [ + 167, + 116, + 303, + 214 + ], + "lines": [ + { + "bbox": [ + 167, + 116, + 303, + 214 + ], + "spans": [ + { + "bbox": [ + 167, + 116, + 303, + 214 + ], + "type": "image", + "image_path": "c4a1190969a76af004907c207d9a7b1c4fa11a0ca8aec274d55bec4405b3e88c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 179, + 224, + 432, + 236 + ], + "lines": [ + { + "bbox": [ + 179, + 224, + 432, + 236 + ], + "spans": [ + { + "bbox": [ + 179, + 224, + 432, + 236 + ], + "type": "text", + "content": "Fig. 7: User Study on two stylized motion generation tasks." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 309, + 116, + 443, + 213 + ], + "blocks": [ + { + "bbox": [ + 309, + 116, + 443, + 213 + ], + "lines": [ + { + "bbox": [ + 309, + 116, + 443, + 213 + ], + "spans": [ + { + "bbox": [ + 309, + 116, + 443, + 213 + ], + "type": "image", + "image_path": "da2f3a7d738cef23e56906e87ff523501e30364fbf6d2e37a13f1ebe0745bcc2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 258, + 480, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 480, + 317 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 480, + 317 + ], + "type": "text", + "content": "model tends to lose the ability to translate content text into corresponding motion, a phenomenon named 'content-forgetting' as described in Sec. 3.3. Fig. 6(a) validates our observation, showing that the content in the generated motion significantly deviates from the text descriptions and closely resembles the style motion sequence." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 327, + 480, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 480, + 445 + ], + "type": "text", + "content": "Style Adaptor and Style Guidance. Initially, we compare our model to a variant without the classifier-based guidance, " + }, + { + "bbox": [ + 130, + 327, + 480, + 445 + ], + "type": "inline_equation", + "content": "w / o" + }, + { + "bbox": [ + 130, + 327, + 480, + 445 + ], + "type": "text", + "content": " classifier-based, to demonstrate its effectiveness. The " + }, + { + "bbox": [ + 130, + 327, + 480, + 445 + ], + "type": "inline_equation", + "content": "5^{\\text{th}}" + }, + { + "bbox": [ + 130, + 327, + 480, + 445 + ], + "type": "text", + "content": " row of Table 3 presents the results. Consistent with the findings in [19], a reasonable trade-off between content preservation and style reflection is observed. Although classifier-based style guidance may slightly affect the content preservation metrics, it significantly boosts the model's performance in the SRA metric, yielding an impressive " + }, + { + "bbox": [ + 130, + 327, + 480, + 445 + ], + "type": "inline_equation", + "content": "208\\%" + }, + { + "bbox": [ + 130, + 327, + 480, + 445 + ], + "type": "text", + "content": " improvement.Fig. 6(d) demonstrates that, without classifier-based style guidance, the generated motions can reflect the motion style, yet they still fall short of fully achieving the target style. Classifier-based style guidance can effectively bridge this gap." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 447, + 481, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 447, + 481, + 542 + ], + "spans": [ + { + "bbox": [ + 130, + 447, + 481, + 542 + ], + "type": "text", + "content": "Subsequently, as shown in Table 3, we evaluate a variant without the style adaptor, denoted as " + }, + { + "bbox": [ + 130, + 447, + 481, + 542 + ], + "type": "inline_equation", + "content": "w / o" + }, + { + "bbox": [ + 130, + 447, + 481, + 542 + ], + "type": "text", + "content": " adaptor (the last row). In cases where the SRA values are close, the style adaptor improved the FID metric by about " + }, + { + "bbox": [ + 130, + 447, + 481, + 542 + ], + "type": "inline_equation", + "content": "80.46\\%" + }, + { + "bbox": [ + 130, + 447, + 481, + 542 + ], + "type": "text", + "content": ". Fig. 6(c) shows that the generated motion can greatly perform the 'walk' action while successfully reflecting the style, but fails to perform the 'sit' action. This indicates that the effectiveness of classifier-based style guidance diminishes when the content text deviates from locomotion-related movements. Relying solely on it may even adversely affect action performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 559, + 220, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 559, + 220, + 571 + ], + "spans": [ + { + "bbox": [ + 132, + 559, + 220, + 571 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": "In this work, we introduce the Stylized Motion Diffusion Model, a novel approach that leverages a pre-trained motion diffusion model to facilitate stylized motion generation driven by content text. By integrating a style adaptor and style classifier guidance, our method is capable of producing realistic human motions that accurately reflect both the content text descriptions and the desired motion style from motion sequences. Through detailed ablation studies, we have demonstrated the effectiveness of each component in our framework." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Zhong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 141, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 138, + 141, + 480, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 141, + 480, + 163 + ], + "spans": [ + { + "bbox": [ + 138, + 141, + 480, + 163 + ], + "type": "text", + "content": "1. Aberman, K., Weng, Y., Lischinski, D., Cohen-Or, D., Chen, B.: Unpaired motion style transfer from video to animation. TOG (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 164, + 480, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 164, + 480, + 186 + ], + "spans": [ + { + "bbox": [ + 138, + 164, + 480, + 186 + ], + "type": "text", + "content": "2. Alexanderson, S., Nagy, R., Beskow, J., Henter, G.E.: Listen, denoise, action! audio-driven motion synthesis with diffusion models. TOG (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 186, + 480, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 186, + 480, + 208 + ], + "spans": [ + { + "bbox": [ + 138, + 186, + 480, + 208 + ], + "type": "text", + "content": "3. Ao, T., Zhang, Z., Liu, L.: Gesturediffuclip: Gesture diffusion model with clip latents. TOG (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 209, + 480, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 209, + 480, + 231 + ], + "spans": [ + { + "bbox": [ + 138, + 209, + 480, + 231 + ], + "type": "text", + "content": "4. Cen, Z., Pi, H., Peng, S., Shen, Z., Yang, M., Zhu, S., Bao, H., Zhou, X.: Generating human motion in 3d scenes from text descriptions. In: CVPR (2024)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 232, + 480, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 232, + 480, + 263 + ], + "spans": [ + { + "bbox": [ + 138, + 232, + 480, + 263 + ], + "type": "text", + "content": "5. Chen, L.H., Lu, S., Zeng, A., Zhang, H., Wang, B., Zhang, R., Zhang, L.: Motionllm: Understanding human behaviors from human motions and videos. ArXiv (2024)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 264, + 480, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 264, + 480, + 286 + ], + "spans": [ + { + "bbox": [ + 138, + 264, + 480, + 286 + ], + "type": "text", + "content": "6. Chen, X., Jiang, B., Liu, W., Huang, Z., Fu, B., Chen, T., Yu, G.: Executing your commands via motion diffusion in latent space. In: CVPR (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 287, + 480, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 287, + 480, + 308 + ], + "spans": [ + { + "bbox": [ + 138, + 287, + 480, + 308 + ], + "type": "text", + "content": "7. Cohan, S., Tevet, G., Reda, D., Peng, X.B., van de Panne, M.: Flexible motion in-between with diffusion models. ArXiv (2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 309, + 480, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 309, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 138, + 309, + 480, + 331 + ], + "type": "text", + "content": "8. Dabral, R., Mughal, M.H., Golyanik, V., Theobalt, C.: Mofusion: A framework for denoising-diffusion-based motion synthesis. In: CVPR (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 331, + 480, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 331, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 138, + 331, + 480, + 354 + ], + "type": "text", + "content": "9. Dai, W., Chen, L.H., Wang, J., Liu, J., Dai, B., Tang, Y.: Motionlcm: Real-time controllable motion generation via latent consistency model. ArXiv (2024)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 354, + 480, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 354, + 480, + 375 + ], + "spans": [ + { + "bbox": [ + 138, + 354, + 480, + 375 + ], + "type": "text", + "content": "0. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. NeurIPS (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 376, + 480, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 376, + 480, + 408 + ], + "spans": [ + { + "bbox": [ + 138, + 376, + 480, + 408 + ], + "type": "text", + "content": "1. Du, H., Herrmann, E., Sprenger, J., Fischer, K., Slusallek, P.: Stylistic locomotion modeling and synthesis using variational generative models. In: Proceedings of the 12th ACM SIGGRAPH Conference on Motion, Interaction and Games (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 409, + 480, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 409, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 138, + 409, + 480, + 430 + ], + "type": "text", + "content": "2. Everaert, M.N., Bocchio, M., Arpa, S., Susstrunk, S., Achanta, R.: Diffusion in style. In: ICCV (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 431, + 480, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 431, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 138, + 431, + 480, + 453 + ], + "type": "text", + "content": "3. Ghosh, A., Dabral, R., Golyanik, V., Theobalt, C., Slusallek, P.: Remos: Reactive 3d motion synthesis for two-person interactions. In: ArXiv (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 454, + 480, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 454, + 480, + 475 + ], + "spans": [ + { + "bbox": [ + 138, + 454, + 480, + 475 + ], + "type": "text", + "content": "4. Guo, C., Mu, Y., Javed, M.G., Wang, S., Cheng, L.: Momask: Generative masked modeling of 3d human motions. In: CVPR (2024)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 476, + 480, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 476, + 480, + 498 + ], + "spans": [ + { + "bbox": [ + 138, + 476, + 480, + 498 + ], + "type": "text", + "content": "5. Guo, C., Mu, Y., Zuo, X., Dai, P., Yan, Y., Lu, J., Cheng, L.: Generative human motion stylization in latent space. ArXiv (2024)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 498, + 480, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 498, + 480, + 520 + ], + "spans": [ + { + "bbox": [ + 138, + 498, + 480, + 520 + ], + "type": "text", + "content": "6. Guo, C., Zou, S., Zuo, X., Wang, S., Ji, W., Li, X., Cheng, L.: Generating diverse and natural 3d human motions from text. In: CVPR (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 521, + 426, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 521, + 426, + 532 + ], + "spans": [ + { + "bbox": [ + 138, + 521, + 426, + 532 + ], + "type": "text", + "content": "7. Ho, J., Salimans, T.: Classifier-free diffusion guidance. ArXiv (2022)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 533, + 480, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 533, + 480, + 564 + ], + "spans": [ + { + "bbox": [ + 138, + 533, + 480, + 564 + ], + "type": "text", + "content": "8. Huang, S., Wang, Z., Li, P., Jia, B., Liu, T., Zhu, Y., Liang, W., Zhu, S.C.: Diffusion-based generation, optimization, and planning in 3d scenes. In: CVPR (2023)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 566, + 480, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 566, + 480, + 587 + ], + "spans": [ + { + "bbox": [ + 138, + 566, + 480, + 587 + ], + "type": "text", + "content": "9. Jang, D.K., Park, S., Lee, S.H.: Motion puzzle: Arbitrary motion style transfer by body part. TOG (2022)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 138, + 588, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 588, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 138, + 588, + 480, + 609 + ], + "type": "text", + "content": "20. Jiang, B., Chen, X., Liu, W., Yu, J., Yu, G., Chen, T.: Motiongpt: Human motion as a foreign language. ArXiv (2023)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 138, + 610, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 610, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 610, + 480, + 632 + ], + "type": "text", + "content": "21. Jones, M., Wang, S.Y., Kumari, N., Bau, D., Zhu, J.Y.: Customizing text-to-image models with a single image pair. ArXiv (2024)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "text", + "content": "22. Karunratanakul, K., Preechakul, K., Aksan, E., Beeler, T., Suwajanakorn, S., Tang, S.: Optimizing diffusion noise can serve as universal motion priors. In: Arxiv (2023)" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 276, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 276, + 91, + 447, + 101 + ], + "type": "text", + "content": "SMooDi: Stylized Motion Diffusion Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "23. Karunratanakul, K., Preechakul, K., Suwajanakorn, S., Tang, S.: Gmd: Controllable human motion synthesis via guided diffusion models. In: ICCV (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 480, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 480, + 159 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 480, + 159 + ], + "type": "text", + "content": "24. Kim, H.J., Lee, S.H.: Perceptual characteristics by motion style category. In: Eurographics (Short Papers) (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 160, + 480, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 480, + 191 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 480, + 191 + ], + "type": "text", + "content": "25. Kulkarni, N., Rempe, D., Genova, K., Kundu, A., Johnson, J., Fouhey, D., Guibas, L.: Nifty: Neural object interaction fields for guided human motion synthesis. ArXiv (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 192, + 471, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 192, + 471, + 203 + ], + "spans": [ + { + "bbox": [ + 132, + 192, + 471, + 203 + ], + "type": "text", + "content": "26. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. ArXiv (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 204, + 480, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 204, + 480, + 235 + ], + "spans": [ + { + "bbox": [ + 132, + 204, + 480, + 235 + ], + "type": "text", + "content": "27. Mason, I., Starke, S., Komura, T.: Real-time style modelling of human locomotion via feature-wise transformations and local motion phases. Proceedings of the ACM on Computer Graphics and Interactive Techniques (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 236, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 480, + 267 + ], + "type": "text", + "content": "28. Park, S., Jang, D.K., Lee, S.H.: Diverse motion stylization for multiple style domains via spatial-temporal graph-based generative model. Proceedings of the ACM on Computer Graphics and Interactive Techniques (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 268, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 268, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 268, + 480, + 289 + ], + "type": "text", + "content": "29. Peng, X., Xie, Y., Wu, Z., Jampani, V., Sun, D., Jiang, H.: Hoi-diff: Text-driven synthesis of 3d human-object interactions using diffusion models. ArXiv (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 290, + 480, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 290, + 480, + 310 + ], + "spans": [ + { + "bbox": [ + 132, + 290, + 480, + 310 + ], + "type": "text", + "content": "30. Peng, X.B., Ma, Z., Abbeel, P., Levine, S., Kanazawa, A.: Amp: Adversarial motion priors for stylized physics-based character control. TOG (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 311, + 480, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 311, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 132, + 311, + 480, + 331 + ], + "type": "text", + "content": "31. Petrovich, M., Black, M.J., Varol, G.: Temos: Generating diverse human motions from textual descriptions. In: ECCV (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 332, + 480, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 480, + 364 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 480, + 364 + ], + "type": "text", + "content": "32. Petrovich, M., Litany, O., Iqbal, U., Black, M.J., Varol, G., Bin Peng, X., Rempe, D.: Multi-track timeline control for text-driven 3d human motion generation. In: CVPR (2024)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 365, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 480, + 385 + ], + "type": "text", + "content": "33. Pi, H., Peng, S., Yang, M., Zhou, X., Bao, H.: Hierarchical generation of human-object interactions with diffusion probabilistic models. In: ICCV (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 386, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 480, + 407 + ], + "type": "text", + "content": "34. Pinyoanuntapong, E., Saleem, M.U., Wang, P., Lee, M., Das, S., Chen, C.: Bamm: Bidirectional autoregressive motion model. ArXiv (2024)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 407, + 480, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 480, + 428 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 480, + 428 + ], + "type": "text", + "content": "35. Pinyoanuntapong, E., Wang, P., Lee, M., Chen, C.: Mmm: Generative masked motion model. In: CVPR (2024)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 429, + 480, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 429, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 480, + 460 + ], + "type": "text", + "content": "36. Raab, S., Gat, I., Sala, N., Tevet, G., Shalev-Arkushin, R., Fried, O., Bermano, A.H., Cohen-Or, D.: Monkey see, monkey do: Harnessing self-attention in motion diffusion for zero-shot motion transfer. ArXiv (2024)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 461, + 480, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 461, + 480, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 461, + 480, + 482 + ], + "type": "text", + "content": "37. Raab, S., Leibovitch, I., Tevet, G., Arar, M., Bermano, A.H., Cohen-Or, D.: Single motion diffusion. ArXiv (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 483, + 480, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 480, + 515 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 480, + 515 + ], + "type": "text", + "content": "38. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 516, + 480, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 516, + 480, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 516, + 480, + 547 + ], + "type": "text", + "content": "39. Rempe, D., Luo, Z., Peng, X.B., Yuan, Y., Kitani, K., Kreis, K., Fidler, S., Litany, O.: Trace and pace: Controllable pedestrian animation via guided trajectory diffusion. In: CVPR (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 548, + 480, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 548, + 480, + 579 + ], + "spans": [ + { + "bbox": [ + 132, + 548, + 480, + 579 + ], + "type": "text", + "content": "40. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: CVPR (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 580, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 580, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 580, + 480, + 601 + ], + "type": "text", + "content": "41. Shafir, Y., Tevet, G., Kapon, R., Bermano, A.H.: Human motion diffusion as a generative prior. ArXiv (2023)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 601, + 480, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 480, + 622 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 480, + 622 + ], + "type": "text", + "content": "42. Shah, V., Ruiz, N., Cole, F., Lu, E., Lazebnik, S., Li, Y., Jampani, V.:Ziplora: Any subject in any style by effectively merging loras. ArXiv (2023)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 623, + 480, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 623, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 623, + 480, + 633 + ], + "type": "text", + "content": "43. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. ArXiv (2020)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 634, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 634, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 634, + 480, + 643 + ], + "type": "text", + "content": "44. Song, Y., Dhariwal, P., Chen, M., Sutskever, I.: Consistency models. ArXiv (2023)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "type": "text", + "content": "45. Tao, T., Zhan, X., Chen, Z., van de Panne, M.: Style-erd: Responsive and coherent online motion style transfer. Arxiv (2022)" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Zhong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 588 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "text", + "content": "46. Tevet, G., Raab, S., Gordon, B., Shafir, Y., Cohen-or, D., Bermano, A.H.: Human motion diffusion model. In: ICLR (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 481, + 160 + ], + "type": "text", + "content": "47. Wan, W., Dou, Z., Komura, T., Wang, W., Jayaraman, D., Liu, L.: Tlcontrol: Trajectory and language control for human motion synthesis. ArXiv (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 160, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 160, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 160, + 481, + 182 + ], + "type": "text", + "content": "48. Wang, Z., Wang, J., Lin, D., Dai, B.: Intercontrol: Generate human motion interactions by controlling every joint. ArXiv (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 182, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 481, + 205 + ], + "type": "text", + "content": "49. Wen, Y.H., Yang, Z., Fu, H., Gao, L., Sun, Y., Liu, Y.J.: Autoregressive stylized motion synthesis with generative flow. In: CVPR (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 205, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 481, + 226 + ], + "type": "text", + "content": "50. Wu, Q., Zhao, Y., Wang, Y., Tai, Y.W., Tang, C.K.: Motionllm: Multimodal motion-language learning with large language models. ArXiv (2024)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 226, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 226, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 481, + 248 + ], + "type": "text", + "content": "51. Wu, Q., Shi, Y., Huang, X., Yu, J., Xu, L., Wang, J.: Thor: Text to human-object interaction diffusion via relation intervention. ArXiv (2024)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 248, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 481, + 270 + ], + "type": "text", + "content": "52. Xia, S., Wang, C., Chai, J., Hodgins, J.: Realtime style transfer for unlabeled heterogeneous human motion. TOG (2015)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 270, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 481, + 293 + ], + "type": "text", + "content": "53. Xie, Y., Jampani, V., Zhong, L., Sun, D., Jiang, H.: Omnicontrol: Control any joint at any time for human motion generation. In: ICLR (2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 293, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 481, + 325 + ], + "type": "text", + "content": "54. Xu, P., Xie, K., Andrews, S., Kry, P.G., Neff, M., McGuire, M., Karamouzas, I., Zordan, V.: Adaptnet: Policy adaptation for physics-based character control. TOG (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 325, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 325, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 325, + 481, + 346 + ], + "type": "text", + "content": "55. Xu, S., Ma, Z., Huang, Y., Lee, H., Chai, J.: Rethinking cycle consistency in text-guided diffusion for image manipulation. NeurIPS (2024)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 346, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 481, + 369 + ], + "type": "text", + "content": "56. Xu, S., Li, Z., Wang, Y.X., Gui, L.Y.: Interdiff: Generating 3d human-object interactions with physics-informed diffusion. In: ICCV (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 369, + 481, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 481, + 391 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 481, + 391 + ], + "type": "text", + "content": "57. Xu, S., Wang, Z., Wang, Y.X., Gui, L.Y.: Interdreamer: Zero-shot text to 3d dynamic human-object interaction. ArXiv (2024)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 391, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 391, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 481, + 413 + ], + "type": "text", + "content": "58. Yi, H., Thies, J., Black, M.J., Peng, X.B., Rempe, D.: Generating human interaction motions in scenes with text control. ArXiv (2024)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 413, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 481, + 435 + ], + "type": "text", + "content": "59. Yu, J., Wang, Y., Zhao, C., Ghanem, B., Zhang, J.: Freedom: Training-free energy-guided conditional diffusion model. ArXiv (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 435, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 435, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 435, + 481, + 456 + ], + "type": "text", + "content": "60. Yuan, Y., Song, J., Iqbal, U., Vahdat, A., Kautz, J.: Physdiff: Physics-guided human motion diffusion model. In: ICCV (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 456, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 456, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 456, + 481, + 479 + ], + "type": "text", + "content": "61. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: ICCV (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 479, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 481, + 501 + ], + "type": "text", + "content": "62. Zhang, M., Cai, Z., Pan, L., Hong, F., Guo, X., Yang, L., Liu, Z.: Motiondiffuse: Text-driven human motion generation with diffusion model. PAMI (2024)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 501, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 501, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 501, + 481, + 533 + ], + "type": "text", + "content": "63. Zhang, Y., Huang, D., Liu, B., Tang, S., Lu, Y., Chen, L., Bai, L., Chu, Q., Yu, N., Ouyang, W.: Motiongpt: Finetuned llms are general-purpose motion generators. ArXiv (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 533, + 481, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 481, + 566 + ], + "type": "text", + "content": "64. Zhou, W., Dou, Z., Cao, Z., Liao, Z., Wang, J., Wang, W., Liu, Y., Komura, T., Wang, W., Liu, L.: Emdm: Efficient motion diffusion model for fast, high-quality motion generation. Arxiv (2023)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 566, + 481, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 566, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 130, + 566, + 481, + 588 + ], + "type": "text", + "content": "65. Zhou, Y., Barnes, C., Lu, J., Yang, J., Li, H.: On the continuity of rotation representations in neural networks. In: CVPR (2019)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 276, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 276, + 91, + 447, + 102 + ], + "type": "text", + "content": "SMooDi: Stylized Motion Diffusion Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/SNP_ Structured Neuron-level Pruning to Preserve Attention Scores/cb547506-91af-4f34-aebd-6546cde66d30_content_list.json b/2024/SNP_ Structured Neuron-level Pruning to Preserve Attention Scores/cb547506-91af-4f34-aebd-6546cde66d30_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..cfde8de3b0c176e4efc30a86710cd31e3164010b --- /dev/null +++ b/2024/SNP_ Structured Neuron-level Pruning to Preserve Attention Scores/cb547506-91af-4f34-aebd-6546cde66d30_content_list.json @@ -0,0 +1,2669 @@ +[ + { + "type": "text", + "text": "SNP: Structured Neuron-level Pruning to Preserve Attention Scores", + "text_level": 1, + "bbox": [ + 263, + 140, + 743, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kyunghwan Shim1, Jaewoong Yun1, and Shinkook Choi1*", + "bbox": [ + 274, + 210, + 730, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Nota Inc., Korea $^{2}$ {kyunghwan.shim, jwyun, shinkook. Choi}@nota.ai", + "bbox": [ + 321, + 238, + 679, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Multi-head self-attention (MSA) is a key component of Vision Transformers (ViTs), which have achieved great success in various vision tasks. However, their high computational cost and memory footprint hinder their deployment on resource-constrained devices. Conventional pruning approaches can only compress and accelerate the MSA module using head pruning, although the head is not an atomic unit. To address this issue, we propose a novel graph-aware neuron-level pruning method, Structured Neuron-level Pruning (SNP). SNP prunes neurons with less informative attention scores and eliminates redundancy among heads. Specifically, it prunes graphically connected query and key layers having the least informative attention scores while preserving the overall attention scores. Value layers, which can be pruned independently, are pruned to eliminate inter-head redundancy. Our proposed method effectively compresses and accelerates Transformer-based models for both edge devices and server processors. For instance, the DeiT-Small with SNP runs 3.1 times faster than the original model and achieves performance that is $21.94\\%$ faster and $1.12\\%$ higher than the DeiT-Tiny. Additionally, SNP accelerates the efficiently designed Transformer model, EfficientFormer, by 1.74 times on the Jetson Nano with acceptable performance degradation. Source code is at https://github.com/Nota-NetsPresso/SNP", + "bbox": [ + 261, + 303, + 743, + 582 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Network pruning $\\cdot$ Network compression $\\cdot$ Compact models", + "bbox": [ + 261, + 594, + 740, + 609 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 633, + 375, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision Transformers (ViTs) [7,17,24] have outperformed or matched the performance of state-of-the-art convolutional neural networks (CNNs) [10,22,23,28] on various computer vision tasks. The success of ViTs is attributed to the Multihead Self-Attention (MSA) module, which captures intricate relationships in data. However, the Transformer architecture entails substantial computational resources, posing challenges for practical applications on edge devices with constrained storage and computational capabilities. To address this issue, we leverage the graphical components of the MSA module to reduce the dimension of interconnected layers, aimed at effectively reducing their computing budgets while achieving hardware-agnostic speedups.", + "bbox": [ + 212, + 664, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author", + "bbox": [ + 218, + 824, + 382, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c6172df040d4b81d7cc5f6a8ea3566eddbe83324bf39d6bcdb5a235143dc073f.jpg", + "image_caption": [ + "Fig. 1: Comparison of model size, speed, and performance. ImageNet-1K classification results. Latency is profiled by Raspberry Pi 4B. The connected lines represent the compressed models paired with the original model. The size of each circle indicates the number of parameters in respective model. The number adjacent to each compressed model indicates its compressed FLOPs." + ], + "image_footnote": [], + "bbox": [ + 300, + 148, + 697, + 401 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "From the perspective of structured pruning, which aims to reduce the number of dimensions in convolutional or linear layers, the MSA module contains two prunable objectives: heads and neurons. Head pruning, which removes the number of heads, is relatively intuitive to implement due to the reduced complexity of graphical elements compared to neuron-level pruning. In contrast, neuron-level pruning reduces the dimension of individual layers in each head of the MSA module, necessitating a comprehensive understanding of the graphical connectivity of the MSA module. A recent work [31] implements neuron-level pruning by zeroing out individual filters without considering the model's graphical connectivity. This indiscriminate zeroing can negatively affect both accuracy and throughput performance [26].", + "bbox": [ + 212, + 517, + 787, + 683 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we introduce a novel graph-aware neuron-level pruning method called Structured Neuron-level Pruning (SNP) to accelerate and compress ViTs effectively. We propose two pruning criteria based on the function of each layer within the MSA module. SNP prunes filter pairs of query and key layers containing fewer contributions to attention scores. Moreover, SNP aims to reduce redundancy across heads by eliminating the redundant filter from the value layer.", + "bbox": [ + 212, + 685, + 787, + 776 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Furthermore, by removing identical filter indices for all the graphically connected layers, SNP could accelerate various Transformer models on various devices without additional libraries, as shown in Fig. 1.", + "bbox": [ + 212, + 777, + 787, + 823 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, the major contributions of this paper are:", + "bbox": [ + 238, + 824, + 638, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Shim. et al.", + "bbox": [ + 271, + 114, + 349, + 126 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel graph-aware neuron-level pruning method, SNP, for Transformer models. SNP is the first method to use the graphical characteristics of the MSA module to measure the importance score of neurons.", + "- To the best of our knowledge, this is the first work to accelerate Transformer models on any device using neuron-level pruning only.", + "- SNP achieves significant acceleration while maintaining the original performance on several models. Compressed DeiT-Small outperforms DeiT-Tiny by $1.12\\%$ in accuracy, with similar FLOPs, and reduces inference time on various edge devices. Furthermore, the proposed method accelerates the efficiently designed Transformer model, EfficientFormer [16], more than two times with acceptable performance degradation." + ], + "bbox": [ + 223, + 146, + 784, + 309 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 329, + 387, + 345 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Compressing vision Transformers", + "text_level": 1, + "bbox": [ + 215, + 357, + 535, + 372 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The ViTs [7, 17, 24] achieve high performance in numerous vision tasks without specialized image processing modules such as convolutions. The key concept of ViTs is to segment images into patch sequences, convert these patches to token embeddings, and then process them through Transformer encoders [25]. ViTs consist only of Transformer blocks, making them likely to be over-parameterized. For this reason, recent works have aimed to reduce computational cost [3, 24] and be memory efficient [18, 26]. DeiT [24] proposes lightweight ViT architectures through knowledge distillation [13]. ToMe [3] proposes accelerating ViTs by directly combining similar tokens, without the need for training. Liu et al. [18] propose post-training quantization method using nuclear norm based mixed-precision scheme that does not require fine-tuning for the Vision Transformer.", + "bbox": [ + 212, + 378, + 784, + 544 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 Pruning vision Transformers", + "text_level": 1, + "bbox": [ + 215, + 561, + 498, + 577 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unstructured and structured pruning Pruning methods can be broadly categorized into two types, unstructured and structured pruning. Unstructured pruning sets individual weights or parameters to zero, resulting in irregular sparse matrices [9, 14]. Compressed models using unstructured pruning tend to maintain relatively high performance for a given pruning ratio. However, they necessitate additional libraries, such as cuSPARSE [5], Automatic SParsity [20], or SparseDNN [27] to accelerate sparse matrix computations.", + "bbox": [ + 212, + 583, + 784, + 688 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Structured pruning, on the other hand, involves the removal of entire groups of units, such as filters or attention heads. This can be implemented using \"masking\" (zeroing out) [11, 12, 32, 33], or by \"pruning\" [8, 15]. Structured pruning by masking [11, 12, 32, 33] simply sets the group of units to zero, which requires additional libraries to accelerate the model, as unstructured pruning. \"Pruning\" [8, 15], on the other hand, requires a comprehensive understanding of the network's graphical connectivity, including element-wise operations that enforce the same input shape. By considering the graphical connectivity and pruning identical filter indices for inter-connected layers, structured pruning can achieve acceleration on any devices.", + "bbox": [ + 212, + 689, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "SNP: Structured Neuron-level Pruning to Preserve Attention Scores", + "bbox": [ + 277, + 114, + 732, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e5fc9c5a11723ce6c34fb414641fd391df21360b5ad2c1f35e852f92d78d0aca.jpg", + "image_caption": [ + "Fig. 2: Proposed SNP methods, on each prunable component of the Transformer block. (a) SNP for query and key layers to preserve attention scores. (b) prunable components of the Transformer block. (c) SNP for value and other layers, including FFN and patch embedding. (d) conventional zeroing out in the matrix multiplication operator. (e) conventional zeroing out and graph-aware pruning in the residual connection." + ], + "image_footnote": [], + "bbox": [ + 218, + 145, + 784, + 362 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5d3c6fee3f2a567b24a6f3b0f3b7472b81693d0894ed3fc18078d60c048306ed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 720, + 362, + 738, + 373 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Zeroing out", + "bbox": [ + 738, + 363, + 769, + 371 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Head and neuron-level pruning Structured pruning for the MSA module has two pruning objectives: head and neuron. Head pruning [29,30] reduces the number of heads, while neuron-level pruning [31] reduces the dimension of each query, key, and value layer in each head. Recent studies for pruning ViTs have focused on head pruning. X-Pruner [30] proposes a novel head pruning method for ViTs that introduces explainability-aware masks and measure the head importance, resulting in superior model compression. WDPruning [29] propose a method to control the number of attention heads and blocks via threshold on learnable parameters.", + "bbox": [ + 212, + 496, + 787, + 632 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "UVC [31] utilizes knowledge distillation alongside several pruning techniques, such as head pruning, block pruning, and neuron-level pruning. However, neuron-level pruning of UVC is carried out in a masking (zeroing out) manner, converting the weight matrix into a sparse matrix. For this reason, UVC necessitates additional libraries or hardware for accelerating sparse matrices, otherwise, the compressed model with UVC cannot achieve latency gain from the neuron-level pruning.", + "bbox": [ + 212, + 633, + 787, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 215, + 758, + 380, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 787, + 372, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "MSA module takes a single input $X \\in \\mathbb{R}^{N \\times d}$ , where $N$ denotes the input vector length, and $d$ represents the hidden size. It comprises $H$ heads, each consisting", + "bbox": [ + 212, + 808, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Shim. et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "of three linear layers: query, key, and value. Each layer denoted as $W_{\\{q,k,v\\}}^h \\in \\mathbb{R}^{d \\times d_{\\{q,k,v\\}}}$ , where $h$ represents $h$ -th head and $d_{\\{q,k,v\\}}$ indicates the hidden size of each query, key, and value layer. $\\{Q,K,V\\}^h$ signifies the output of each query, key, and value layer in $h$ -th head, with a shape of $\\mathbb{R}^{N \\times d_{\\{q,k,v\\}}}$ .", + "bbox": [ + 212, + 145, + 782, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The self-attention operation for $h$ -th head can be expressed as follows:", + "bbox": [ + 238, + 213, + 746, + 227 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A s} ^ {h} (X) = Q ^ {h} \\cdot \\left(K ^ {h}\\right) ^ {T} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 424, + 234, + 785, + 252 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A t t} ^ {h} (X) = \\operatorname {S o f t m a x} \\left(\\frac {\\operatorname {A s} ^ {h} (X)}{\\sqrt {d _ {q}}}\\right) \\cdot V ^ {h} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 271, + 784, + 306 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {M S A} (X) = \\operatorname {C o n c a t} \\left(\\operatorname {A t t} ^ {1} (X), \\dots , \\operatorname {A t t} ^ {H} (X)\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 349, + 315, + 782, + 332 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "here, As, Att, and MSA represent the functions to calculate attention scores, attention module, and MSA module respectively.", + "bbox": [ + 212, + 335, + 782, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Matrix multiplication, unlike a residual connection, yields zero when either of the inputs is masked, regardless of the value of the other input. Therefore, applying a conventional zeroing-out approach to neuron-level pruning in the MSA module can lead to unintended results, as shown in Fig. 2 (d). To address this issue, we introduce two graph-aware neuron-level pruning criteria to compress and expedite the MSA module.", + "bbox": [ + 212, + 366, + 784, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Preserving attention scores", + "text_level": 1, + "bbox": [ + 215, + 476, + 485, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Attention scores, Eq. (1), of the MSA module learn long-range dependencies between image features by focusing on different parts of the image when processing different features. These attention scores can be recognized as a series of outer product of $Q_{i}^{h}$ and $(K_{i}^{h})^{T}$ as follows:", + "bbox": [ + 212, + 497, + 782, + 559 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\operatorname {A s} ^ {h} (X) = Q ^ {h} \\cdot \\left(K ^ {h}\\right) ^ {T} \\\\ = \\sum_ {i = 1} ^ {d _ {q}} Q _ {i} ^ {h} \\cdot \\left(K _ {i} ^ {h}\\right) ^ {T} \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 411, + 566, + 784, + 630 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "from this perspective, neuron-level pruning, reducing the dimension of query $d_{q}$ and key $d_{k}$ , inevitably distorts the attention scores. For this reason, preserving attention scores is essential to maintain the high performance of the original model.", + "bbox": [ + 212, + 635, + 782, + 694 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To alleviate the distortion, we maintain the graphically connected query-key filter pair $(Q_{i}^{h}$ and $K_{i}^{h})$ , constituting a filter-by-attention score $(Q_{i}^{h}\\cdot (K_{i}^{h})^{T}\\in$ $\\mathbb{R}^{N\\times N})$ , that retains the most significant aspects of the overall attention scores. To identify the most informative filter pair, we initially employ singular value decomposition (SVD) to decompose the original model's attention scores.", + "bbox": [ + 212, + 696, + 782, + 771 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\operatorname {A s} ^ {h} (X) = \\tilde {U} \\cdot \\tilde {S} \\cdot \\tilde {V} ^ {T} \\\\ = \\sum_ {j = 1} ^ {N} \\tilde {\\sigma} _ {j} \\cdot \\tilde {u} _ {j} \\cdot \\tilde {v} _ {j} ^ {T} \\tag {5} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 411, + 779, + 784, + 842 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "SNP: Structured Neuron-level Pruning to Preserve Attention Scores", + "bbox": [ + 277, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tilde{U}$ and $\\tilde{V}$ are the left and right singular vector matrices, respectively, and $\\tilde{S}$ is the diagonal matrix of singular values, with $\\tilde{\\sigma}_1 \\geq \\tilde{\\sigma}_2 \\geq \\ldots \\geq \\tilde{\\sigma}_N$ .", + "bbox": [ + 215, + 143, + 784, + 176 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SVD is a technique for extracting the most informative components of a matrix, those with large singular values, while discarding the less informative components, those with small singular values. To retain the spatial relationships captured by the attention mechanism, we prune the filter pair $Q_{i}^{h}$ and $K_{i}^{h}$ with least correlation with the most informative components of the attention scores.", + "bbox": [ + 212, + 178, + 782, + 252 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To measure the correlation, we adopt the cosine similarity between the attention scores of the $i$ -th filter and the $j$ -th rank matrix, as normalizing the attention scores reduces the impact of magnitude-based approaches. Consequently, the importance score $\\omega_{as,i}^{h}$ is defined as follows:", + "bbox": [ + 212, + 253, + 782, + 316 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\omega_ {a s, i} ^ {h} = \\sum_ {j = 1} ^ {r} | \\cos ((Q _ {i} ^ {h} \\cdot (K _ {i} ^ {h}) ^ {T}), (\\tilde {\\sigma} _ {j} \\cdot \\tilde {u} _ {j} \\cdot \\tilde {v} _ {j} ^ {T})) | \\\\ = \\sum_ {j = 1} ^ {r} | \\frac {\\left(Q _ {i} ^ {h} \\cdot \\left(K _ {i} ^ {h}\\right) ^ {T}\\right) \\cdot \\left(\\tilde {\\sigma} _ {j} \\cdot \\tilde {u} _ {j} \\cdot \\tilde {v} _ {j} ^ {T}\\right)}{| | \\left(Q _ {i} ^ {h} \\cdot \\left(K _ {i} ^ {h}\\right) ^ {T}\\right) | | \\cdot | | \\left(\\tilde {\\sigma} _ {j} \\cdot \\tilde {u} _ {j} \\cdot \\tilde {v} _ {j} ^ {T}\\right) | |} | \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 341, + 332, + 784, + 412 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $r$ is a hyperparameter dictating the quantity of ranks, within the range of $1 \\leq r \\leq N$ , to be compared with the $i$ -th attention scores, while the remaining $N - r$ singular values $(\\tilde{\\sigma}_{r + 1},\\dots,\\tilde{\\sigma}_N)$ are discarded.", + "bbox": [ + 215, + 428, + 782, + 472 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Optimizing $r$ for each attention module can contribute to preserving informative filters and sustaining higher performance. For this reason, we adopt Variational Bayesian Matrix Factorization (VBMF) [21] to determine the optimal rank $r$ for the attention scores. VBMF inherently addresses the challenge of determining the number of latent factors (ranks) in matrix factorization using its probabilistic approach.", + "bbox": [ + 212, + 474, + 782, + 565 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Importance score $\\omega_{as,i}^{h}$ , defined in Eq. (6), represents the importance of the $i$ -th filter for both query and key layers. A larger $\\omega_{as,i}^{h}$ indicates that the filter has a greater impact on the main component of the attention scores, while a lower $\\omega_{as,i}^{h}$ indicates that the filter is less important and can be removed without significantly affecting the attention scores.", + "bbox": [ + 212, + 566, + 782, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Inter-head redundancy removal", + "text_level": 1, + "bbox": [ + 215, + 671, + 519, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the preceding section, we outlined the approach to preserving attention scores even with reduced embedding dimensions in the query and key layers. Here, we introduce a pruning method for the value and other layers, such as FFN or patch embedding layer.", + "bbox": [ + 212, + 702, + 782, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Previous works [2,19] have revealed that a significant proportion of attention heads can be removed without causing significant performance deterioration. To remove this inter-head redundancy through neuron-level pruning, we propose to measure the distance between all the value layers of MSA module, irrespective of the heads. The importance score $\\omega_{v,i}^{h}$ of the $i$ -th filter in the value layer of the", + "bbox": [ + 212, + 763, + 782, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Shim. et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$h$ -th head is as follows:", + "bbox": [ + 215, + 146, + 383, + 160 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\omega_ {v, i} ^ {h} = \\sum_ {j = 1} ^ {H} \\sum_ {l = 1} ^ {d _ {v}} \\left(1 - \\left| \\cos \\left(W _ {i} ^ {h}, W _ {l} ^ {j}\\right) \\right|\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 167, + 785, + 209 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The importance score of the value layer, denoted as $\\omega_{v,i}^{h}$ , indicates the redundancy of filter $i$ in the value layer compared to all value layers of the heads within the corresponding MSA layer. Consequently, filters with the lowest importance scores will be pruned in the initial stages of the pruning process.", + "bbox": [ + 212, + 217, + 784, + 279 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Accelerating Transformers", + "text_level": 1, + "bbox": [ + 215, + 297, + 478, + 313 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As described in Sec. 3.2, SNP removes the least correlated filter pairs $(Q_{i}^{h}, K_{i}^{h})$ to preserve attention scores and enhance the efficiency of the MSA module across various devices.", + "bbox": [ + 212, + 319, + 782, + 364 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Most MSA modules, including MSA modules in DeiTs, are designed to compute all heads in parallel. For achieving this parallel computation, SNP ensures that the number of filter dimensions are consistent across heads, even though the filter indices for each head are independently selected.", + "bbox": [ + 212, + 366, + 784, + 425 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The green highlighted boxes in Fig. 2 (b) represent layers connected by a single residual connection at the last add layer of the Transformer block. Unlike the matrix multiplication operator, the output of the residual connection becomes zero when all the interconnected layers return zero for specific filter indices. For this reason, the residual connection with masking always exceeds the performance of actual pruning, which restricts the set of possible pruning patterns [26].", + "bbox": [ + 212, + 426, + 784, + 531 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To accelerate the residual connection layers, all interconnected layers should be pruned identically. To achieve this, we sum up all the calculated importance scores for interconnected layers based on their filter indices, as shown in Fig. 2 (e). Subsequently, we prune the least important filter indices of all the connected layers.", + "bbox": [ + 212, + 532, + 784, + 606 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Experimental Results", + "text_level": 1, + "bbox": [ + 215, + 627, + 460, + 645 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To ensure a fair comparison with existing methods, we apply SNP to prune the DeiT [24] architectures trained on the ImageNet-1K [6] dataset. Additionally, we conduct experiments on the efficient Transformer model, EfficientFormer-L1, to confirm the robustness of SNP. Furthermore, a series of ablation studies are conducted to gain a comprehensive understanding of our methodology.", + "bbox": [ + 212, + 656, + 782, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1 Implementation details", + "text_level": 1, + "bbox": [ + 215, + 751, + 454, + 767 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The overall pruning and fine-tuning process are executed on the pre-trained $\\mathrm{DeiT}^3$ and EfficientFormer-L1 $^4$ released from the official implementation on", + "bbox": [ + 212, + 773, + 784, + 805 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "SNP: Structured Neuron-level Pruning to Preserve Attention Scores", + "bbox": [ + 277, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "3 https://github.com/facebookresearch/deit", + "bbox": [ + 217, + 809, + 514, + 825 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "4 https://github.com/snap-research/EfficientFormer", + "bbox": [ + 222, + 825, + 568, + 839 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/15471ad10ad38383a868c88a27cfe4a750512f45653f5f0a72ededd9f36ea634.jpg", + "table_caption": [ + "Table 1: Performance comparison of various pruning methods on ImageNet-1K. The \"Pruning\" column represents the pruning methods that corresponding methods use to compress the MSA module. Each of the methods contains one of follows: head pruning (HP), block pruning (BP), neuron-level pruning (NP), neuron-level sparsity (NS)." + ], + "table_footnote": [], + "table_body": "
ModelMethodPruningTop-1 (%)Top-5 (%)GFLOPsParams (M)
DeiT-TOriginal [24]-72.2091.101.35.7
SSViTE [4]HP70.12-0.94.2
WDPruning [29]HP+BP70.3489.820.73.5
X-Pruner [30]HP71.1090.110.6-
SNPNP70.2990.010.63.0
UVC [31]HP+NS+BP70.60-0.5-
DeiT-SOriginal [24]-79.8595.004.622.1
SSViTE [4]HP79.22-3.114.6
WDPruning [29]HP+BP78.3894.052.613.3
X-Pruner [30]HP78.9394.242.4-
UVC [31]HP+NS+BP78.82-2.3-
SNPNP78.5294.372.010.0
SNPNP73.3291.661.36.4
DeiT-BOriginal [24]-81.8095.5917.686.6
SSViTE [4]HP82.22-11.856.8
WDPruning [29]HP+BP80.7695.369.955.3
X-Pruner [30]HP81.0295.388.5-
UVC [31]HP+NS+BP80.57-8.0-
SNPNP79.6394.376.431.6
", + "bbox": [ + 215, + 227, + 785, + 512 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ImageNet-1K. Throughout the fine-tuning phase of the pruned model, we maintain consistent settings across all models, except for the batch size and learning rate. The batch size is set to 256, and to prevent weight explosion, we adjust the learning rate of the compressed model to 1/10 or 1/100 of the original model.", + "bbox": [ + 212, + 545, + 784, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To evaluate the reduced latency using SNP, we have configured four testing scenarios: one on a CPU and another on GPU for both edge devices and server processors. We employ a standard PyTorch model for profiling on the server processors (Intel Xeon Silver 4210R and NVIDIA GeForce RTX 3090). Profiling on the Raspberry Pi 4B and Jetson Nano is conducted using the ONNX and TensorRT formats, respectively. All latencies are measured using a single image as an input, except for the GPU of the server processor (RTX 3090), where it is set to 64 images.", + "bbox": [ + 212, + 606, + 785, + 728 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 Quantitative results", + "text_level": 1, + "bbox": [ + 215, + 752, + 426, + 767 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison with other methods Despite the constraints outlined in Sec. 3.4, SNP not only maintains accuracy comparable to existing methods but also significantly reduces inference time across diverse hardware and data types on the ImageNet-1K dataset, as shown in Tab. 1 and Tab. 2.", + "bbox": [ + 212, + 779, + 785, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Shim. et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9c87df815b8f0d7394ce3702bd49457dda95631b75d88c1445651327c8b089a2.jpg", + "table_caption": [ + "Table 2: Inference speed and Top-1 accuracy of the compressed model across different devices. Performance evaluation involves accuracy on ImageNet-1K and inference time for the compressed DeiTs and EfficientFormer-L1. Latency is benchmarked with 200 warm-up runs and averaged over 1000 runs. In latency measurement, a single image is used as an input, except for the RTX 3090, where 64 images are employed in a single batch." + ], + "table_footnote": [], + "table_body": "
ModelTop-1 (%)GFLOPsEdge devices (ms)Server processors (ms)
Raspberry Pi 4B (.onnx)Jetson Nano (.trt)Xeon Silver 4210R (.pt)RTX 3090 (.pt)
DeiT-Tiny72.201.3139.1341.0334.7418.65
+ SNP (Ours)70.290.681.63 (1.70×)26.67 (1.54×)25.25 (1.38×)17.82 (1.05×)
DeiT-Small79.804.6401.2799.3253.3746.13
+ SNP (Ours)78.522.0199.15 (2.01×)45.51 (2.18×)38.57 (1.38×)32.91 (1.40×)
+ SNP (Ours)73.321.3136.68 (2.94×)32.03 (3.10×)33.46 (1.60×)26.98 (1.71×)
DeiT-Base81.8017.61377.71293.29122.03151.35
+ SNP (Ours)79.636.4565.68 (2.44×)132.55 (2.21×)64.65 (1.89×)72.96 (2.07×)
EfficientFormer-L179.201.3169.1330.9543.7526.19
+ SNP (Ours)75.530.695.12 (1.78×)19.78 (1.56×)38.25 (1.14×)17.24 (1.52×)
+ SNP (Ours)74.510.582.60 (2.05×)17.76 (1.74×)35.15 (1.24×)16.01 (1.64×)
", + "bbox": [ + 220, + 239, + 787, + 366 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In a recent study [26], unconstrained masking generally outperforms posttraining accuracy of pruned models by an average of $2.1\\%$ on ImageNet-1K. Compared to unconstrained head masking approaches like SSVITE [4] and XPruner [30], SNP achieves significantly higher compression rates for all DeiTs FLOPs (30.64% and 10.64%) with minimal performance degradation (0.83% and 0.87%), much less than the $2.1\\%$ average mentioned.", + "bbox": [ + 212, + 397, + 787, + 488 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Furthermore, compared to other pruning approaches like WDPruning [29] and UVC [31], which use a combination of pruning techniques to compress DeiTs, SNP achieves comparable accuracy solely through neuron-level pruning. Notably, DeiT-Small with SNP outperforms WDPruning by $0.14\\%$ , with the removal of 3.3 million parameters and a reduction of 0.6 GFLOPs. Compared to UVC, SNP exhibits negligible performance degradation, averaging $0.51\\%$ across all DeiTs while using $7.65\\%$ fewer FLOPs.", + "bbox": [ + 212, + 488, + 787, + 594 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Large compressed vs. Small hand-crafted DeiT-Small with SNP, a large pruned model, outperforms the smaller, hand-crafted DeiT-Tiny in both accuracy and latency, achieving a notable $1.12\\%$ improvement in top-1 accuracy while maintaining similar FLOPs. Additionally, DeiT-Small with SNP exhibits enhanced speed compared to the original DeiT-Tiny across edge devices and CPU-based server processors. Notably, its speed increases up to $21.94\\%$ compared to the original DeiT-Tiny running on Jetson Nano, a GPU-based edge device. This substantial performance gap underscores the superiority of the compressed model (DeiT-Small with SNP) over the smaller hand-crafted designed model (DeiT-Tiny) in both overall performance and speed.", + "bbox": [ + 212, + 618, + 787, + 770 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Accelerating Transformer-based models As depicted in Tab. 2, SNP achieves impressive acceleration of DeiTs by a factor of $1.44 \\times$ to $2.44 \\times$ on edge devices and $1.05 \\times$ to $2.07 \\times$ on server processors. This acceleration is notable,", + "bbox": [ + 212, + 794, + 787, + 842 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "SNP: Structured Neuron-level Pruning to Preserve Attention Scores", + "bbox": [ + 279, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/6d63344579af9058c87aa70dad1df97d56f7b38419d208c37bd1ecb3cc71554f.jpg", + "image_caption": [ + "Fig.3: Attention maps with varying pruning criteria and compression ratios. All query and key layers are locally pruned based on the specified pruning ratio without fine-tuning. The importance scores of $l2$ -norm and GM on query and key layers are combined by filter index and pruned simultaneously. \"Reverse\" represents the reverse order of SNP." + ], + "image_footnote": [], + "bbox": [ + 331, + 143, + 674, + 440 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "with negligible average performance degradation of $1.79\\%$ , specifically $1.91\\%$ , $1.28\\%$ , and $2.17\\%$ for DeiT-Tiny, DeiT-Small, and DeiT-Base, respectively.", + "bbox": [ + 212, + 556, + 784, + 588 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Compared to WDPruning, which employs both head and block pruning, SNP surpasses in terms of latency for DeiT-Small and DeiT-Base on the RTX 3090 GPU. SNP accelerates the original DeiTs by $1.38 \\times$ and $2.07 \\times$ , respectively, while WDPruning achieves a comparatively modest acceleration of $1.18 \\times$ for both models.", + "bbox": [ + 212, + 590, + 787, + 665 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The superiority of SNP becomes more evident when the target of neuron-level pruning, linear or convolutional layers, contribute a larger proportion of the original model's computation time. we find that SNP further accelerates original model $1.05 \\times$ to $2.07 \\times$ as the DeiT model's size grows on the RTX 3090 GPU. In contrast, WDPruning, involving the removal of entire layers and associated operators in both head and block, consistently shows acceleration rates of $1.18 \\times$ .", + "bbox": [ + 212, + 670, + 787, + 761 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To ascertain the robustness of SNP across diverse Transformer models, especially on the efficiently designed model, we conduct additional experiments on EfficientFormer-L1. SNP accelerates the model by $1.78 \\times$ and $2.05 \\times$ faster on Raspberry Pi 4B and $1.56 \\times$ and $1.74 \\times$ faster on Jetson Nano. Notably, the compressed EfficientFormer-L1 achieves acceptable accuracy of $75.53\\%$ and $74.51\\%$ .", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Shim. et al.", + "bbox": [ + 271, + 114, + 349, + 126 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/fe1f4d12f872427ee2e4c2882de994577a236ac3db60578a73e73447f1e97c39.jpg", + "image_caption": [ + "Original" + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 310, + 213 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/6323a610de3b6a87de2951e2534265b0f3ea1293eae2b48888c961379a775071.jpg", + "image_caption": [ + "Compressed" + ], + "image_footnote": [], + "bbox": [ + 313, + 143, + 401, + 214 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/f495f899f667f74c1ebea9d84560b323a1faed3a302230ccb3ef7f71e9b587c1.jpg", + "image_caption": [ + "Compressed + Fine-tuned" + ], + "image_footnote": [], + "bbox": [ + 405, + 143, + 495, + 214 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/f30d324ea6e826972f11170d584acf8e3fe343abc0f6c1e17e2631f0f4df2e9e.jpg", + "image_caption": [ + "Original" + ], + "image_footnote": [], + "bbox": [ + 506, + 143, + 596, + 214 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/c088f449756ed2a658ee4db4eea0f3dfaa2917b693a573088d8ade161617983d.jpg", + "image_caption": [ + "Compressed" + ], + "image_footnote": [], + "bbox": [ + 598, + 143, + 689, + 214 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/0ea0c7fbda0f1faafa893bf05469fc6f2f2992d7809c48d50a19a03231c65570.jpg", + "image_caption": [ + "Compressed + Fine-tuned" + ], + "image_footnote": [], + "bbox": [ + 692, + 143, + 782, + 214 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/a1e97d41a8cea68255e7a9b05107ab15f937d340b1c469d8eeff847609e8b360.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 222, + 310, + 229 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/e4cf7f34b9d24d2d89564d5c2c695ea84ad328de030288edc83aaa2d40aaafe9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 229, + 310, + 237 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/92ca6126fbf405bbe02594c29cd89eafeb1345f69377f89fd39db6cacdac27df.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 237, + 310, + 244 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/043764bb3f9586de9783dd19fd78bfe6789c008dc0dfe86a14e182e87f85b1e0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 244, + 310, + 253 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/3f28d2f837980e4e95e62d2210f97120d05337bfb2cef20679ab14e68cee5230.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 253, + 310, + 261 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/1e147a49ae4d164c1753880b556a9c73d16b28742f70514b921c7a109872d675.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 261, + 310, + 268 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/e46717d66c317f1268b3e24b641d9d8f10dbae092dba3851ec623446d258ad24.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 268, + 310, + 276 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/02649542649345d8d47fac7611e4659a93655d187db203c620b577879fb96272.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 276, + 310, + 284 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/b9fa9dee01639061707c36607e68657d83d0d4250357ab057048d9a50f5fd60f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 284, + 310, + 291 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/80c3f9e675e8736b403f13dec05ab4f9f08cbffc7d0abf6bbcd7fffe6aedd9c5.jpg", + "image_caption": [ + "Fig. 4: Attention maps from the original, compressed, and fine-tuned DeiT-Tiny with SNP. The attention maps in the first row are visualized using the attention rollout [1]. Each red box contains three attention maps from each head of the MSA module, ordered accordingly." + ], + "image_footnote": [], + "bbox": [ + 218, + 291, + 310, + 297 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/a6a84efa7ff27e3d0f5799df9789e0bf10cce09c7e4dc607d8571f21e8d31059.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 297, + 310, + 305 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/8733c633fd40e21ad93c447045e6e70d37d8c82e6a5128398182a49c7a071c7b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 305, + 310, + 311 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/c4657e25aa77e07d79fd2c1d8f3225347964d110ab3ee705b21897aa451ae4cb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 313, + 222, + 401, + 229 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/733fb88eff8af16a930947a91c86c970e83bcdab4ab20295a91ff3a61cfe05e3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 313, + 229, + 401, + 237 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/adbf66146248093a28a285b7928c9f3c461198191b1a51b8de3b2581ec023cd1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 313, + 237, + 401, + 244 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/0dc1699ad20f9728167c256362b43505ffff8a63f3117f2981d60b51a9aa0c7b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 313, + 244, + 401, + 252 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/5b3d85906cd7f77e264948e19e58181b3d6928c4b8c55f36eb727dd53b8c9134.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 313, + 252, + 401, + 261 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/05e40153d154fde971d94c85587107f7e94c4c1d400362961dfd9b96af661771.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 313, + 261, + 401, + 268 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/2c163d34778577e23b21f38d9f54d65670f591c1a9602c9bd095bbc91b24cc7b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 313, + 268, + 401, + 276 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/122ea387d2a42ed8cd836a60e795faa22985d68c431685d36c6e0b2d78f2ff2c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 313, + 276, + 401, + 284 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/216d524b367e0b7159db2a0df90b7796c13acf549b5b02adcd0a2a9fffab04df.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 313, + 284, + 401, + 291 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/48f57ef9d9e276d804a646c3424e4fb611d9016e3d3305ed1e68662b1a22652e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 313, + 291, + 401, + 297 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/a0071d2447b2285f1beeac11a0a9d1049a6a772332819a5eb5215886eba57ce6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 313, + 297, + 401, + 305 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/7f97ff62864dc438be9a9b156403674afc785197bb8015297585adc5dd2902b5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 313, + 305, + 401, + 311 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/7bfe8e9879866ad1a6c551c96df7557d9964c16717a30e1f6983899afca7a108.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 405, + 222, + 495, + 229 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/2d45862860754f7be5655ecc809419140763d0d2b1a60f89a7e73eb04cdfdefb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 405, + 229, + 495, + 237 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/318dd579dc52f9b742f1d86ef9aa91021a9c238e975c5a4c99728a01a8de4ec3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 405, + 237, + 495, + 244 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/8600cc0e3899fc7aac7bcffef47550512e8fb77623099f2a1b515c27003a720f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 405, + 244, + 495, + 252 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/f98793e63d20f9a9d33927b64ba63087ce6af79b800bb39f5bcea8668e8665b6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 405, + 252, + 495, + 261 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/f556b8f22da9495a87ed7c26bf33cce21f6866d0dd03cf7c29419878a24b0548.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 405, + 261, + 495, + 268 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/dc00dd4deef70eecd465a32d4d5c909f7920a45768ad490b3908478f32e5164c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 405, + 268, + 495, + 276 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/1265d570ac5366ba1c740910c160458a20d26cdcdc9f1ab78331d5268ee51c4b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 405, + 276, + 495, + 282 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/9c41c3b2f1c42b754af61aa24e3c3daed53a7152b46f9e316471384d9c83d11f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 405, + 282, + 495, + 291 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/ab34a27b16544d65dd65588fc3234fc124a3d24e49c664e3638c186460afcb7f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 405, + 291, + 495, + 297 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/68e1a02c5e9e3957959c1d53b488fed6a86730c7a59cc678ed5e984096a344d0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 405, + 297, + 495, + 305 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/f0c77594a2a0657a100a720728263c58e613ad73e9648f8710c6549c9a699d37.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 405, + 305, + 495, + 311 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/618191c91df49a5db67c4f792861f9e4e662552b65fb950866d2bfba05a88c4a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 222, + 596, + 229 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/d506e8f8ab137f12aa0747c2b7be22465379f366ad7775f4b45daf5d8ba4432b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 229, + 596, + 237 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/b94ba33d9bf2a0a0aac9d21e4a294cbe26fa5f485f5219eaaaf53e204c0b5c45.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 237, + 596, + 244 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/0b3d966682b866cb80a0e9ec3586d4e95bf0ae3fe281a6b3623bfbcc89614577.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 244, + 596, + 252 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/b69fc8c7e5b8abad7c52d50cf70066816270dec0650c575792c421a2d00390bd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 252, + 596, + 261 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/354be57d221d5f468ac85f1c2acc6aaea3aabf4269d236655696504b97443df1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 261, + 596, + 268 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/2a5938bd8b133b7d82cfefa4972bc6a980461af895027a54c53243ba0a70de4a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 268, + 596, + 276 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/8c5882ec658ac9467160bd0710b675aae6e37c14569f6882d1d9d6623f4d3f8a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 276, + 596, + 284 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/4ece03b7e0f278b79f3ad24fa430ee3ac8862134854f0aa48d8807b7b6ed1e04.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 284, + 596, + 291 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/fecd71612f8074ebef73698044fc46a9af96362fa4ef774437bfe46e6674c032.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 291, + 596, + 299 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/0d2163883253187f7708895be4ae99120d136bb583e820d409e6acf67c8a9644.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 299, + 596, + 306 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/e056c869856b4f7047484e8018fc91945cd8339e04976e05e83a980f284f77e6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 306, + 596, + 311 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/4c8f13fae9e6650345f57b4d567043ad8c8a10acf3f9a2143b79d749ea85bea1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 311, + 596, + 318 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/1d25f4d1aa674a20408152d2157006593e6d85974ac47f722744149394d933d5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 599, + 222, + 689, + 229 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/76e771d3611843403d36d3da83fde4ccc37213ed51e8d5f5fb21c5db355b7fb1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 599, + 229, + 689, + 237 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/35f8c87b894ea7e4d1368f6b7ab46468ace45a6efdb1b0e9b56d1f4d47fad4e2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 599, + 237, + 689, + 244 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/31c426fa47f0b63bb46e402e6bc6ffaf686ef738eb8439cae1cc7d125d481535.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 599, + 244, + 689, + 252 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/2a2c918e52fa5710e109cb0c128c6c7c22ba9e2c384ef01e318e9a0e1f7e56b7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 599, + 252, + 689, + 261 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/610a3ad04eb3118d53ef7ea46059c44300eda5c6f6032cccbc1cd28dae25ba9f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 599, + 261, + 689, + 268 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/373e4a86d8c2a753fc16a0f8587d8bd1bed4d021efdf9f93a639b8e8b60dd773.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 599, + 268, + 689, + 276 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/966bae9ff81f39d2fc182054cb32e20f63373e819cd950b539564ced0ad021d2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 599, + 276, + 689, + 284 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/f4add919c6a72372542223bba125bcbddca915eb557c7212e4945484844a5cdd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 599, + 284, + 689, + 291 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/607df98602bf3860ee5e4041956aacc419c68b5c6e61b2e8d43be2e564daf380.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 599, + 291, + 689, + 299 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/13276cb8d9e4f91f0dc8e713b4e7b890603f1090284dace91b825fadd2dc7cfb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 599, + 299, + 689, + 306 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/979d3abae462302381b2126b7ae161cfb56fe82f11cc8f00a01130ba70571e48.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 599, + 306, + 689, + 311 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/0c5666a18bfdf3429739304cb15cc376a03fa0b521b4d10cf11e1ede8ff1bbd5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 599, + 311, + 689, + 318 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/dcda949fcbe86e57760bf7fb2ca869b4d853ecfc47a5640786cb24d1e218b9e9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 692, + 222, + 779, + 228 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/d185cbdb87880d909ac1d5b4e423ae5683659da251038a3b582146bc7d7d9ada.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 692, + 228, + 782, + 234 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/c3b74fbd38c0eced21ed711a499a711d3e5933c22bcd5e58a6b1b9ac5f364c2d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 692, + 234, + 782, + 243 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/7c9a5c1ba7551fe44e8eb64d1ff9ff564ec584b523b1a31d16e38f53f8f51dc7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 692, + 243, + 782, + 251 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/42af389f2d089c9cd68bbf2a8671f19c1d6dc20930d8b5690707e89317a99305.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 692, + 251, + 782, + 258 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/ccaf4e741482c658eac0d87b6b322f7e7c40a1e41743d334bfd1e13011e1ffa3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 692, + 258, + 782, + 267 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/79e1fdd3460bfb2ffb97ca83563353083e6491f399ce931481c57f34ffaec664.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 692, + 267, + 782, + 275 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/c58674a657e7f9bc5ba2c64f813bf9ea653692c8c2f5d8e2587e5c8a2b988e7a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 692, + 275, + 782, + 282 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/1e8e03716568d8677f5f7ed812813b2bc65ee25d3aa3f07f2737d5ae99b65b6d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 692, + 282, + 782, + 291 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/23a32049cdd4e27ed7b0e3dce9156011db1140a1bf608b08e547c6895b2c97ec.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 692, + 291, + 782, + 299 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/0c1d342ed7ee47d058557fd7271830d8aacb9f2d342f6dc24a6cc6da9d49db2c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 692, + 299, + 782, + 306 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/c75fa64d8f8053ecb7141f850a5bfa7815e2971e30e7f09e9b9881f3840922f0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 692, + 306, + 782, + 311 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Qualitative results", + "text_level": 1, + "bbox": [ + 215, + 416, + 413, + 430 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In the subsequent sections, attention rollout [1] is employed to randomly selected images from the test datasets to visualize the attention maps of the DeiTs and assess the efficacy of SNP. This evaluation unfolds across two key facets:", + "bbox": [ + 214, + 441, + 785, + 488 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Preserving the attention scores: Visualize attention maps to verify the effectiveness of the proposed importance score in preserving attention scores.", + "- Restoring the attention scores after SNP: Visualize attention maps to find the effectiveness of SNP in restoring attention scores." + ], + "bbox": [ + 223, + 498, + 784, + 559 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Preserving the attention scores To visualize SNP effectiveness, we apply four pruning criteria to compress the query and key filter pairs in all MSA modules of DeiT-Tiny: SNP, $l2$ -norm, geometric median (GM) [11], and \"Reverse\".", + "bbox": [ + 214, + 582, + 785, + 628 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The \"Reverse\" criterion prioritizes pruning the most important filter pairs first, keeping the least important pairs until the end. However, both $l2$ -norm and GM pruning criteria ignore the MSA module's graphical connectivity, evaluating importance scores independently for query and key layers. To handle identical filter indices during pruning, we aggregate scores based on filter indices, removing the least important indices from both layers.", + "bbox": [ + 214, + 628, + 785, + 718 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In Fig. 3, attention maps for the original DeiT-Tiny and locally pruned models are presented across various pruning ratios (10% to 90% with 10% intervals). Our proposed method effectively maintains the original attention map even after pruning over 80% of the filters, whereas other methods show fragmented attention maps at much lower pruning ratios (typically 30% or less). These results highlight the potential of our neuron-level pruning criteria, utilizing SVD to preserve attention scores, in reducing the size and speeding up the execution of MSA modules without compromising accuracy.", + "bbox": [ + 212, + 719, + 785, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "SNP: Structured Neuron-level Pruning to Preserve Attention Scores", + "bbox": [ + 279, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/4f0dcfe461c149defea78ee450ff1aab445d0961931299b153886e598eac9d79.jpg", + "table_caption": [ + "Table 3: Performance of SNP without fine-tuning across different data quantities at various pruning ratios. To determine the proper number of images for calculating the importance score, we conduct local pruning on the query and key layers at pruning ratios of $10\\%$ , $30\\%$ , $50\\%$ , $70\\%$ , and $90\\%$ . All performance metrics are assessed without fine-tuning. The latency is measured on Raspberry Pi 4B." + ], + "table_footnote": [], + "table_body": "
Number of imagesPerformance by pruning ratio
Original10%30%50%70%90%
172.2071.1567.9657.6038.6811.17
472.2070.8267.8657.6538.0511.86
1672.2070.7267.4258.6039.469.11
6472.2070.8367.5559.5042.1312.83
25672.2070.7568.1459.5842.7014.48
Latency (ms)139.13133.60129.29119.41117.48112.32
Params (M)5.75.595.415.235.064.88
", + "bbox": [ + 241, + 226, + 756, + 368 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Restoring the attention scores after SNP Fig. 4 illustrates the overall and per-head attention maps of the original, compressed, and fine-tuned DeiT-Tiny, respectively. The first row shows the overall attention maps of the respective models. Notably, the compressed model maintains a well-preserved overall attention map, despite pruning all layers, including values and FFN, resulting in a $53\\%$ reduction in FLOPs and a $46\\%$ reduction in parameters. Especially, we can observe that attention map is well-restored after the fine-tuning process.", + "bbox": [ + 212, + 397, + 785, + 502 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The twelve red boxes below the overall attention map depict per-head attention maps for twelve layers of each original, compressed, and fine-tuned models respectively. As shown in Fig. 4, it is evident that the attention maps for each head are effectively preserved and restored in each of the compressed and fine-tuned models.", + "bbox": [ + 212, + 503, + 785, + 578 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Ablation studies", + "text_level": 1, + "bbox": [ + 214, + 602, + 398, + 616 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Importance scores by the data quantity Since the attention scores depend on the input, as indicated in Eq. (1), the proposed importance scores for query and key filter pairs (Eq. (6)) may exhibit sensitivity to the distribution of the input image $X$ . To validate the method's robustness, we compute importance scores using various image quantities, pruning query and key layers at different ratios, without fine-tuning process.", + "bbox": [ + 212, + 628, + 785, + 718 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "As depicted in Tab. 3, SNP demonstrates a slight advantage in preserving performance with an increasing number of images, outperforming models compressed with fewer images as the compression ratio rises. However, this improvement comes at the cost of increased computation time for SNP calculations. Considering these factors into account, we opt to use 64 images, which yield the second-best performance among the given number of images. This decision strikes a balance between achieving satisfactory performance and maintaining computational efficiency.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Shim. et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/d1bbf3c5b9cc371fd90ad09fdca666753623755d4200fee638014e2ffaa89677.jpg", + "image_caption": [ + "Fig. 5: Top-1 accuracy of compressed DeiT-Tiny on ImageNet using several pruning criteria without fine-tuning. Query and key layers are locally pruned using various pruning criteria : SNP, GM, l2-norm, reverse order of SNP (\"Reverse\"), and original DeiT-Tiny (\"Baseline\")." + ], + "image_footnote": [], + "bbox": [ + 320, + 146, + 676, + 380 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Performance comparison across pruning ratios To assess the robustness of SNP, we examine the performance of DeiT-Tiny under various pruning criteria and different pruning ratios, as described in Sec. 4.3.", + "bbox": [ + 212, + 479, + 785, + 525 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "As illustrated in Fig. 5, SNP consistently outperforms other pruning criteria across all pruning ratios. In contrast, models compressed with \"Reverse\" criteria exhibit the lowest performance at all pruning ratios, underscoring the robustness of the proposed approach.", + "bbox": [ + 212, + 526, + 785, + 585 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Both Fig. 3 and Fig. 5 confirm that SNP successfully preserves attention scores in both quantitative and qualitative aspects. Even with an $80\\%$ pruning ratio and without fine-tuning, SNP maintains original attention scores and outperforms other pruning criteria.", + "bbox": [ + 212, + 585, + 785, + 647 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 671, + 359, + 686 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In this paper, we propose a novel graph-aware neuron-level pruning method, SNP, designed to compress and accelerate Transformer-based models. SNP proposes two pruning criteria for preserving attention scores and eliminating interhead redundancy. Using SNP, a large compressed model outperforms small, hand-crafted designed models in both performance and latency on edge devices. Moreover, the compressed models exhibit astonishing results in latency on various devices, with negligible performance degradation.", + "bbox": [ + 212, + 703, + 785, + 809 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "As this work is a first attempt to accelerate MSA modules using neuron-level pruning alone, many challenges remain. One is incorporating other pruning", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "SNP: Structured Neuron-level Pruning to Preserve Attention Scores", + "bbox": [ + 279, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "methods such as head or block pruning for a more efficient Transformer model. Another challenge is applying SNP to other vision tasks, including image generation, which requires high computational costs on both training and inference.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We believe that these works encourage the adoption of model pruning as a tool, for both improving the applicability of ViTs in resource-constrained environments and reducing the training costs of large models by integrating with training process.", + "bbox": [ + 212, + 191, + 782, + 252 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 215, + 273, + 401, + 291 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This work was supported by Artificial intelligence industrial convergence cluster development project funded by the Ministry of Science and ICT(MSIT, Korea)&Gwangju Metropolitan City.", + "bbox": [ + 212, + 303, + 782, + 349 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 371, + 321, + 386 + ], + "page_idx": 13 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Abnar, S., Zuidema, W.: Quantifying attention flow in transformers. arXiv preprint arXiv:2005.00928 (2020)", + "2. Bian, Y., Huang, J., Cai, X., Yuan, J., Church, K.: On attention redundancy: A comprehensive study. In: Proceedings of the 2021 conference of the north american chapter of the association for computational linguistics: human language technologies. pp. 930-945 (2021)", + "3. Bolya, D., Fu, C.Y., Dai, X., Zhang, P., Feichtenhofer, C., Hoffman, J.: Token merging: Your vit but faster. arXiv preprint arXiv:2210.09461 (2022)", + "4. Chen, T., Cheng, Y., Gan, Z., Yuan, L., Zhang, L., Wang, Z.: Chasing sparsity in vision transformers: An end-to-end exploration. Advances in Neural Information Processing Systems 34, 19974-19988 (2021)", + "5. Demouth, J.: Sparse matrix-matrix multiplication on thegpu. Tech. rep., NVIDIA (2012)", + "6. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)", + "7. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)", + "8. Fang, G., Ma, X., Song, M., Mi, M.B., Wang, X.: Depgraph: Towards any structural pruning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16091-16101 (2023)", + "9. Han, S., Mao, H., Dally, W.J.: Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149 (2015)", + "10. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)", + "11. He, Y., Liu, P., Wang, Z., Hu, Z., Yang, Y.: Filter pruning via geometric median for deep convolutional neural networks acceleration. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4340-4349 (2019)" + ], + "bbox": [ + 218, + 400, + 784, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Shim. et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 13 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "12. He, Y., Zhang, X., Sun, J.: Channel pruning for accelerating very deep neural networks. In: Proceedings of the IEEE international conference on computer vision. pp. 1389-1397 (2017)", + "13. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015)", + "14. Lee, J., Park, S., Mo, S., Ahn, S., Shin, J.: Layer-adaptive sparsity for the magnitude-based pruning. arXiv preprint arXiv:2010.07611 (2020)", + "15. Li, H., Kadav, A., Durdanovic, I., Samet, H., Graf, H.P.: Pruning filters for efficient convnets. arXiv preprint arXiv:1608.08710 (2016)", + "16. Li, Y., Yuan, G., Wen, Y., Hu, J., Evangelidis, G., Tulyakov, S., Wang, Y., Ren, J.: Efficientformer: Vision transformers at mobilenet speed. Advances in Neural Information Processing Systems 35, 12934-12949 (2022)", + "17. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021)", + "18. Liu, Z., Wang, Y., Han, K., Zhang, W., Ma, S., Gao, W.: Post-training quantization for vision transformer. Advances in Neural Information Processing Systems 34, 28092-28103 (2021)", + "19. Michel, P., Levy, O., Neubig, G.: Are sixteen heads really better than one? Advances in neural information processing systems 32 (2019)", + "20. Mishra, A., Latorre, J.A., Pool, J., Stosic, D., Stosic, D., Venkatesh, G., Yu, C., Micikevicius, P.: Accelerating sparse deep neural networks. arXiv preprint arXiv:2104.08378 (2021)", + "21. Nakajima, S., Sugiyama, M., Babacan, S.D., Tomioka, R.: Global analytic solution of fully-observed variational bayesian matrix factorization. The Journal of Machine Learning Research 14(1), 1-37 (2013)", + "22. Radosavovic, I., Kosaraju, R.P., Girshick, R., He, K., Dólár, P.: Designing network design spaces. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10428-10436 (2020)", + "23. Tan, M., Le, Q.: Efficientnet: Rethinking model scaling for convolutional neural networks. In: International conference on machine learning. pp. 6105-6114. PMLR (2019)", + "24. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jégou, H.: Training data-efficient image transformers & distillation through attention. In: International conference on machine learning. pp. 10347-10357. PMLR (2021)", + "25. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)", + "26. Wan, A., Hao, H., Patnaik, K., Xu, Y., Hadad, O., Güera, D., Ren, Z., Shan, Q.: Upscale: unconstrained channel pruning. In: International Conference on Machine Learning. pp. 35384-35412. PMLR (2023)", + "27. Wang, Z.: Sparsednn: Fast sparse deep learning inference on cpus. arXiv preprint arXiv:2101.07948 (2021)", + "28. Xie, Q., Luong, M.T., Hovy, E., Le, Q.V.: Self-training with noisy student improves imagenet classification. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10687-10698 (2020)", + "29. Yu, F., Huang, K., Wang, M., Cheng, Y., Chu, W., Cui, L.: Width & depth pruning for vision transformers. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 3143-3151 (2022)" + ], + "bbox": [ + 215, + 146, + 784, + 840 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "SNP: Structured Neuron-level Pruning to Preserve Attention Scores", + "bbox": [ + 279, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "30. Yu, L., Xiang, W.: X-pruner: explainable pruning for vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24355-24363 (2023)", + "31. Yu, S., Chen, T., Shen, J., Yuan, H., Tan, J., Yang, S., Liu, J., Wang, Z.: Unified visual transformer compression. arXiv preprint arXiv:2203.08243 (2022)", + "32. Yu, S., Mazaheri, A., Jannesari, A.: Topology-aware network pruning using multi-stage graph embedding and reinforcement learning. In: International conference on machine learning. pp. 25656-25667. PMLR (2022)", + "33. Yu, X., Serra, T., Ramalingam, S., Zhe, S.: The combinatorial brain surgeon: pruning weights that cancel one another in neural networks. In: International Conference on Machine Learning. pp. 25668-25683. PMLR (2022)" + ], + "bbox": [ + 215, + 146, + 784, + 300 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Shim. et al.", + "bbox": [ + 271, + 114, + 349, + 126 + ], + "page_idx": 15 + } +] \ No newline at end of file diff --git a/2024/SNP_ Structured Neuron-level Pruning to Preserve Attention Scores/cb547506-91af-4f34-aebd-6546cde66d30_model.json b/2024/SNP_ Structured Neuron-level Pruning to Preserve Attention Scores/cb547506-91af-4f34-aebd-6546cde66d30_model.json new file mode 100644 index 0000000000000000000000000000000000000000..abdc5be1de520d4ace73f4041e799819fbc9cce0 --- /dev/null +++ b/2024/SNP_ Structured Neuron-level Pruning to Preserve Attention Scores/cb547506-91af-4f34-aebd-6546cde66d30_model.json @@ -0,0 +1,3026 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.264, + 0.141, + 0.744, + 0.187 + ], + "angle": 0, + "content": "SNP: Structured Neuron-level Pruning to Preserve Attention Scores" + }, + { + "type": "text", + "bbox": [ + 0.275, + 0.212, + 0.731, + 0.229 + ], + "angle": 0, + "content": "Kyunghwan Shim1, Jaewoong Yun1, and Shinkook Choi1*" + }, + { + "type": "text", + "bbox": [ + 0.322, + 0.239, + 0.681, + 0.27 + ], + "angle": 0, + "content": "\\(^{1}\\) Nota Inc., Korea \\(^{2}\\) {kyunghwan.shim, jwyun, shinkook. Choi}@nota.ai" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.304, + 0.744, + 0.583 + ], + "angle": 0, + "content": "Abstract. Multi-head self-attention (MSA) is a key component of Vision Transformers (ViTs), which have achieved great success in various vision tasks. However, their high computational cost and memory footprint hinder their deployment on resource-constrained devices. Conventional pruning approaches can only compress and accelerate the MSA module using head pruning, although the head is not an atomic unit. To address this issue, we propose a novel graph-aware neuron-level pruning method, Structured Neuron-level Pruning (SNP). SNP prunes neurons with less informative attention scores and eliminates redundancy among heads. Specifically, it prunes graphically connected query and key layers having the least informative attention scores while preserving the overall attention scores. Value layers, which can be pruned independently, are pruned to eliminate inter-head redundancy. Our proposed method effectively compresses and accelerates Transformer-based models for both edge devices and server processors. For instance, the DeiT-Small with SNP runs 3.1 times faster than the original model and achieves performance that is \\(21.94\\%\\) faster and \\(1.12\\%\\) higher than the DeiT-Tiny. Additionally, SNP accelerates the efficiently designed Transformer model, EfficientFormer, by 1.74 times on the Jetson Nano with acceptable performance degradation. Source code is at https://github.com/Nota-NetsPresso/SNP" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.595, + 0.741, + 0.61 + ], + "angle": 0, + "content": "Keywords: Network pruning \\(\\cdot\\) Network compression \\(\\cdot\\) Compact models" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.635, + 0.377, + 0.65 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.665, + 0.788, + 0.818 + ], + "angle": 0, + "content": "Vision Transformers (ViTs) [7,17,24] have outperformed or matched the performance of state-of-the-art convolutional neural networks (CNNs) [10,22,23,28] on various computer vision tasks. The success of ViTs is attributed to the Multihead Self-Attention (MSA) module, which captures intricate relationships in data. However, the Transformer architecture entails substantial computational resources, posing challenges for practical applications on edge devices with constrained storage and computational capabilities. To address this issue, we leverage the graphical components of the MSA module to reduce the dimension of interconnected layers, aimed at effectively reducing their computing budgets while achieving hardware-agnostic speedups." + }, + { + "type": "page_footnote", + "bbox": [ + 0.22, + 0.825, + 0.383, + 0.84 + ], + "angle": 0, + "content": "* Corresponding author" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.127 + ], + "angle": 0, + "content": "Shim. et al." + }, + { + "type": "image", + "bbox": [ + 0.302, + 0.149, + 0.699, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.414, + 0.788, + 0.484 + ], + "angle": 0, + "content": "Fig. 1: Comparison of model size, speed, and performance. ImageNet-1K classification results. Latency is profiled by Raspberry Pi 4B. The connected lines represent the compressed models paired with the original model. The size of each circle indicates the number of parameters in respective model. The number adjacent to each compressed model indicates its compressed FLOPs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.518, + 0.788, + 0.684 + ], + "angle": 0, + "content": "From the perspective of structured pruning, which aims to reduce the number of dimensions in convolutional or linear layers, the MSA module contains two prunable objectives: heads and neurons. Head pruning, which removes the number of heads, is relatively intuitive to implement due to the reduced complexity of graphical elements compared to neuron-level pruning. In contrast, neuron-level pruning reduces the dimension of individual layers in each head of the MSA module, necessitating a comprehensive understanding of the graphical connectivity of the MSA module. A recent work [31] implements neuron-level pruning by zeroing out individual filters without considering the model's graphical connectivity. This indiscriminate zeroing can negatively affect both accuracy and throughput performance [26]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.686, + 0.788, + 0.777 + ], + "angle": 0, + "content": "In this paper, we introduce a novel graph-aware neuron-level pruning method called Structured Neuron-level Pruning (SNP) to accelerate and compress ViTs effectively. We propose two pruning criteria based on the function of each layer within the MSA module. SNP prunes filter pairs of query and key layers containing fewer contributions to attention scores. Moreover, SNP aims to reduce redundancy across heads by eliminating the redundant filter from the value layer." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.778, + 0.788, + 0.824 + ], + "angle": 0, + "content": "Furthermore, by removing identical filter indices for all the graphically connected layers, SNP could accelerate various Transformer models on various devices without additional libraries, as shown in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.825, + 0.639, + 0.841 + ], + "angle": 0, + "content": "In summary, the major contributions of this paper are:" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.279, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "SNP: Structured Neuron-level Pruning to Preserve Attention Scores" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.147, + 0.784, + 0.191 + ], + "angle": 0, + "content": "- We propose a novel graph-aware neuron-level pruning method, SNP, for Transformer models. SNP is the first method to use the graphical characteristics of the MSA module to measure the importance score of neurons." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.192, + 0.785, + 0.221 + ], + "angle": 0, + "content": "- To the best of our knowledge, this is the first work to accelerate Transformer models on any device using neuron-level pruning only." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.221, + 0.785, + 0.31 + ], + "angle": 0, + "content": "- SNP achieves significant acceleration while maintaining the original performance on several models. Compressed DeiT-Small outperforms DeiT-Tiny by \\(1.12\\%\\) in accuracy, with similar FLOPs, and reduces inference time on various edge devices. Furthermore, the proposed method accelerates the efficiently designed Transformer model, EfficientFormer [16], more than two times with acceptable performance degradation." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.147, + 0.785, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.33, + 0.388, + 0.346 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.358, + 0.536, + 0.373 + ], + "angle": 0, + "content": "2.1 Compressing vision Transformers" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.379, + 0.785, + 0.545 + ], + "angle": 0, + "content": "The ViTs [7, 17, 24] achieve high performance in numerous vision tasks without specialized image processing modules such as convolutions. The key concept of ViTs is to segment images into patch sequences, convert these patches to token embeddings, and then process them through Transformer encoders [25]. ViTs consist only of Transformer blocks, making them likely to be over-parameterized. For this reason, recent works have aimed to reduce computational cost [3, 24] and be memory efficient [18, 26]. DeiT [24] proposes lightweight ViT architectures through knowledge distillation [13]. ToMe [3] proposes accelerating ViTs by directly combining similar tokens, without the need for training. Liu et al. [18] propose post-training quantization method using nuclear norm based mixed-precision scheme that does not require fine-tuning for the Vision Transformer." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.563, + 0.499, + 0.578 + ], + "angle": 0, + "content": "2.2 Pruning vision Transformers" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.785, + 0.689 + ], + "angle": 0, + "content": "Unstructured and structured pruning Pruning methods can be broadly categorized into two types, unstructured and structured pruning. Unstructured pruning sets individual weights or parameters to zero, resulting in irregular sparse matrices [9, 14]. Compressed models using unstructured pruning tend to maintain relatively high performance for a given pruning ratio. However, they necessitate additional libraries, such as cuSPARSE [5], Automatic SParsity [20], or SparseDNN [27] to accelerate sparse matrix computations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Structured pruning, on the other hand, involves the removal of entire groups of units, such as filters or attention heads. This can be implemented using \"masking\" (zeroing out) [11, 12, 32, 33], or by \"pruning\" [8, 15]. Structured pruning by masking [11, 12, 32, 33] simply sets the group of units to zero, which requires additional libraries to accelerate the model, as unstructured pruning. \"Pruning\" [8, 15], on the other hand, requires a comprehensive understanding of the network's graphical connectivity, including element-wise operations that enforce the same input shape. By considering the graphical connectivity and pruning identical filter indices for inter-connected layers, structured pruning can achieve acceleration on any devices." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.128 + ], + "angle": 0, + "content": "Shim. et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.146, + 0.785, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.363, + 0.739, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.74, + 0.364, + 0.77, + 0.372 + ], + "angle": 0, + "content": "Zeroing out" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.387, + 0.788, + 0.47 + ], + "angle": 0, + "content": "Fig. 2: Proposed SNP methods, on each prunable component of the Transformer block. (a) SNP for query and key layers to preserve attention scores. (b) prunable components of the Transformer block. (c) SNP for value and other layers, including FFN and patch embedding. (d) conventional zeroing out in the matrix multiplication operator. (e) conventional zeroing out and graph-aware pruning in the residual connection." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.497, + 0.788, + 0.633 + ], + "angle": 0, + "content": "Head and neuron-level pruning Structured pruning for the MSA module has two pruning objectives: head and neuron. Head pruning [29,30] reduces the number of heads, while neuron-level pruning [31] reduces the dimension of each query, key, and value layer in each head. Recent studies for pruning ViTs have focused on head pruning. X-Pruner [30] proposes a novel head pruning method for ViTs that introduces explainability-aware masks and measure the head importance, resulting in superior model compression. WDPruning [29] propose a method to control the number of attention heads and blocks via threshold on learnable parameters." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.634, + 0.788, + 0.74 + ], + "angle": 0, + "content": "UVC [31] utilizes knowledge distillation alongside several pruning techniques, such as head pruning, block pruning, and neuron-level pruning. However, neuron-level pruning of UVC is carried out in a masking (zeroing out) manner, converting the weight matrix into a sparse matrix. For this reason, UVC necessitates additional libraries or hardware for accelerating sparse matrices, otherwise, the compressed model with UVC cannot achieve latency gain from the neuron-level pruning." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.76, + 0.381, + 0.779 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.788, + 0.373, + 0.803 + ], + "angle": 0, + "content": "3.1 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.809, + 0.787, + 0.842 + ], + "angle": 0, + "content": "MSA module takes a single input \\( X \\in \\mathbb{R}^{N \\times d} \\), where \\( N \\) denotes the input vector length, and \\( d \\) represents the hidden size. It comprises \\( H \\) heads, each consisting" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.279, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "SNP: Structured Neuron-level Pruning to Preserve Attention Scores" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.213 + ], + "angle": 0, + "content": "of three linear layers: query, key, and value. Each layer denoted as \\( W_{\\{q,k,v\\}}^h \\in \\mathbb{R}^{d \\times d_{\\{q,k,v\\}}} \\), where \\( h \\) represents \\( h \\)-th head and \\( d_{\\{q,k,v\\}} \\) indicates the hidden size of each query, key, and value layer. \\( \\{Q,K,V\\}^h \\) signifies the output of each query, key, and value layer in \\( h \\)-th head, with a shape of \\( \\mathbb{R}^{N \\times d_{\\{q,k,v\\}}} \\)." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.214, + 0.747, + 0.228 + ], + "angle": 0, + "content": "The self-attention operation for \\( h \\)-th head can be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.425, + 0.236, + 0.786, + 0.253 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A s} ^ {h} (X) = Q ^ {h} \\cdot \\left(K ^ {h}\\right) ^ {T} \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.379, + 0.272, + 0.785, + 0.308 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A t t} ^ {h} (X) = \\operatorname {S o f t m a x} \\left(\\frac {\\operatorname {A s} ^ {h} (X)}{\\sqrt {d _ {q}}}\\right) \\cdot V ^ {h} \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.351, + 0.316, + 0.784, + 0.333 + ], + "angle": 0, + "content": "\\[\n\\operatorname {M S A} (X) = \\operatorname {C o n c a t} \\left(\\operatorname {A t t} ^ {1} (X), \\dots , \\operatorname {A t t} ^ {H} (X)\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.337, + 0.784, + 0.366 + ], + "angle": 0, + "content": "here, As, Att, and MSA represent the functions to calculate attention scores, attention module, and MSA module respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.367, + 0.785, + 0.457 + ], + "angle": 0, + "content": "Matrix multiplication, unlike a residual connection, yields zero when either of the inputs is masked, regardless of the value of the other input. Therefore, applying a conventional zeroing-out approach to neuron-level pruning in the MSA module can lead to unintended results, as shown in Fig. 2 (d). To address this issue, we introduce two graph-aware neuron-level pruning criteria to compress and expedite the MSA module." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.477, + 0.486, + 0.492 + ], + "angle": 0, + "content": "3.2 Preserving attention scores" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.498, + 0.784, + 0.56 + ], + "angle": 0, + "content": "Attention scores, Eq. (1), of the MSA module learn long-range dependencies between image features by focusing on different parts of the image when processing different features. These attention scores can be recognized as a series of outer product of \\( Q_{i}^{h} \\) and \\( (K_{i}^{h})^{T} \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.412, + 0.567, + 0.785, + 0.631 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\operatorname {A s} ^ {h} (X) = Q ^ {h} \\cdot \\left(K ^ {h}\\right) ^ {T} \\\\ = \\sum_ {i = 1} ^ {d _ {q}} Q _ {i} ^ {h} \\cdot \\left(K _ {i} ^ {h}\\right) ^ {T} \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.636, + 0.784, + 0.695 + ], + "angle": 0, + "content": "from this perspective, neuron-level pruning, reducing the dimension of query \\( d_{q} \\) and key \\( d_{k} \\), inevitably distorts the attention scores. For this reason, preserving attention scores is essential to maintain the high performance of the original model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.697, + 0.784, + 0.772 + ], + "angle": 0, + "content": "To alleviate the distortion, we maintain the graphically connected query-key filter pair \\((Q_{i}^{h}\\) and \\(K_{i}^{h})\\), constituting a filter-by-attention score \\((Q_{i}^{h}\\cdot (K_{i}^{h})^{T}\\in\\) \\(\\mathbb{R}^{N\\times N})\\), that retains the most significant aspects of the overall attention scores. To identify the most informative filter pair, we initially employ singular value decomposition (SVD) to decompose the original model's attention scores." + }, + { + "type": "equation", + "bbox": [ + 0.413, + 0.78, + 0.785, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\operatorname {A s} ^ {h} (X) = \\tilde {U} \\cdot \\tilde {S} \\cdot \\tilde {V} ^ {T} \\\\ = \\sum_ {j = 1} ^ {N} \\tilde {\\sigma} _ {j} \\cdot \\tilde {u} _ {j} \\cdot \\tilde {v} _ {j} ^ {T} \\tag {5} \\\\ \\end{array}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.128 + ], + "angle": 0, + "content": "Shim. et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.145, + 0.785, + 0.177 + ], + "angle": 0, + "content": "where \\(\\tilde{U}\\) and \\(\\tilde{V}\\) are the left and right singular vector matrices, respectively, and \\(\\tilde{S}\\) is the diagonal matrix of singular values, with \\(\\tilde{\\sigma}_1 \\geq \\tilde{\\sigma}_2 \\geq \\ldots \\geq \\tilde{\\sigma}_N\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.784, + 0.253 + ], + "angle": 0, + "content": "SVD is a technique for extracting the most informative components of a matrix, those with large singular values, while discarding the less informative components, those with small singular values. To retain the spatial relationships captured by the attention mechanism, we prune the filter pair \\( Q_{i}^{h} \\) and \\( K_{i}^{h} \\) with least correlation with the most informative components of the attention scores." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.255, + 0.784, + 0.317 + ], + "angle": 0, + "content": "To measure the correlation, we adopt the cosine similarity between the attention scores of the \\(i\\)-th filter and the \\(j\\)-th rank matrix, as normalizing the attention scores reduces the impact of magnitude-based approaches. Consequently, the importance score \\(\\omega_{as,i}^{h}\\) is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.342, + 0.333, + 0.785, + 0.414 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\omega_ {a s, i} ^ {h} = \\sum_ {j = 1} ^ {r} | \\cos ((Q _ {i} ^ {h} \\cdot (K _ {i} ^ {h}) ^ {T}), (\\tilde {\\sigma} _ {j} \\cdot \\tilde {u} _ {j} \\cdot \\tilde {v} _ {j} ^ {T})) | \\\\ = \\sum_ {j = 1} ^ {r} | \\frac {\\left(Q _ {i} ^ {h} \\cdot \\left(K _ {i} ^ {h}\\right) ^ {T}\\right) \\cdot \\left(\\tilde {\\sigma} _ {j} \\cdot \\tilde {u} _ {j} \\cdot \\tilde {v} _ {j} ^ {T}\\right)}{| | \\left(Q _ {i} ^ {h} \\cdot \\left(K _ {i} ^ {h}\\right) ^ {T}\\right) | | \\cdot | | \\left(\\tilde {\\sigma} _ {j} \\cdot \\tilde {u} _ {j} \\cdot \\tilde {v} _ {j} ^ {T}\\right) | |} | \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.429, + 0.784, + 0.473 + ], + "angle": 0, + "content": "where \\( r \\) is a hyperparameter dictating the quantity of ranks, within the range of \\( 1 \\leq r \\leq N \\), to be compared with the \\( i \\)-th attention scores, while the remaining \\( N - r \\) singular values \\( (\\tilde{\\sigma}_{r + 1},\\dots,\\tilde{\\sigma}_N) \\) are discarded." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.476, + 0.784, + 0.566 + ], + "angle": 0, + "content": "Optimizing \\( r \\) for each attention module can contribute to preserving informative filters and sustaining higher performance. For this reason, we adopt Variational Bayesian Matrix Factorization (VBMF) [21] to determine the optimal rank \\( r \\) for the attention scores. VBMF inherently addresses the challenge of determining the number of latent factors (ranks) in matrix factorization using its probabilistic approach." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.567, + 0.784, + 0.645 + ], + "angle": 0, + "content": "Importance score \\(\\omega_{as,i}^{h}\\), defined in Eq. (6), represents the importance of the \\(i\\)-th filter for both query and key layers. A larger \\(\\omega_{as,i}^{h}\\) indicates that the filter has a greater impact on the main component of the attention scores, while a lower \\(\\omega_{as,i}^{h}\\) indicates that the filter is less important and can be removed without significantly affecting the attention scores." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.672, + 0.521, + 0.687 + ], + "angle": 0, + "content": "3.3 Inter-head redundancy removal" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.703, + 0.784, + 0.763 + ], + "angle": 0, + "content": "In the preceding section, we outlined the approach to preserving attention scores even with reduced embedding dimensions in the query and key layers. Here, we introduce a pruning method for the value and other layers, such as FFN or patch embedding layer." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.784, + 0.843 + ], + "angle": 0, + "content": "Previous works [2,19] have revealed that a significant proportion of attention heads can be removed without causing significant performance deterioration. To remove this inter-head redundancy through neuron-level pruning, we propose to measure the distance between all the value layers of MSA module, irrespective of the heads. The importance score \\(\\omega_{v,i}^{h}\\) of the \\(i\\)-th filter in the value layer of the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.279, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "SNP: Structured Neuron-level Pruning to Preserve Attention Scores" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.384, + 0.161 + ], + "angle": 0, + "content": "\\(h\\)-th head is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.379, + 0.168, + 0.786, + 0.21 + ], + "angle": 0, + "content": "\\[\n\\omega_ {v, i} ^ {h} = \\sum_ {j = 1} ^ {H} \\sum_ {l = 1} ^ {d _ {v}} \\left(1 - \\left| \\cos \\left(W _ {i} ^ {h}, W _ {l} ^ {j}\\right) \\right|\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.218, + 0.785, + 0.28 + ], + "angle": 0, + "content": "The importance score of the value layer, denoted as \\(\\omega_{v,i}^{h}\\), indicates the redundancy of filter \\(i\\) in the value layer compared to all value layers of the heads within the corresponding MSA layer. Consequently, filters with the lowest importance scores will be pruned in the initial stages of the pruning process." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.299, + 0.48, + 0.314 + ], + "angle": 0, + "content": "3.4 Accelerating Transformers" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.32, + 0.784, + 0.365 + ], + "angle": 0, + "content": "As described in Sec. 3.2, SNP removes the least correlated filter pairs \\((Q_{i}^{h}, K_{i}^{h})\\) to preserve attention scores and enhance the efficiency of the MSA module across various devices." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.367, + 0.785, + 0.426 + ], + "angle": 0, + "content": "Most MSA modules, including MSA modules in DeiTs, are designed to compute all heads in parallel. For achieving this parallel computation, SNP ensures that the number of filter dimensions are consistent across heads, even though the filter indices for each head are independently selected." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.427, + 0.785, + 0.532 + ], + "angle": 0, + "content": "The green highlighted boxes in Fig. 2 (b) represent layers connected by a single residual connection at the last add layer of the Transformer block. Unlike the matrix multiplication operator, the output of the residual connection becomes zero when all the interconnected layers return zero for specific filter indices. For this reason, the residual connection with masking always exceeds the performance of actual pruning, which restricts the set of possible pruning patterns [26]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.533, + 0.785, + 0.607 + ], + "angle": 0, + "content": "To accelerate the residual connection layers, all interconnected layers should be pruned identically. To achieve this, we sum up all the calculated importance scores for interconnected layers based on their filter indices, as shown in Fig. 2 (e). Subsequently, we prune the least important filter indices of all the connected layers." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.628, + 0.462, + 0.646 + ], + "angle": 0, + "content": "4 Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.657, + 0.784, + 0.734 + ], + "angle": 0, + "content": "To ensure a fair comparison with existing methods, we apply SNP to prune the DeiT [24] architectures trained on the ImageNet-1K [6] dataset. Additionally, we conduct experiments on the efficient Transformer model, EfficientFormer-L1, to confirm the robustness of SNP. Furthermore, a series of ablation studies are conducted to gain a comprehensive understanding of our methodology." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.752, + 0.455, + 0.768 + ], + "angle": 0, + "content": "4.1 Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.775, + 0.785, + 0.806 + ], + "angle": 0, + "content": "The overall pruning and fine-tuning process are executed on the pre-trained \\(\\mathrm{DeiT}^3\\) and EfficientFormer-L1\\(^4\\) released from the official implementation on" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.81, + 0.515, + 0.826 + ], + "angle": 0, + "content": "3 https://github.com/facebookresearch/deit" + }, + { + "type": "page_footnote", + "bbox": [ + 0.223, + 0.826, + 0.569, + 0.84 + ], + "angle": 0, + "content": "4 https://github.com/snap-research/EfficientFormer" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.81, + 0.569, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.128 + ], + "angle": 0, + "content": "Shim. et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.214 + ], + "angle": 0, + "content": "Table 1: Performance comparison of various pruning methods on ImageNet-1K. The \"Pruning\" column represents the pruning methods that corresponding methods use to compress the MSA module. Each of the methods contains one of follows: head pruning (HP), block pruning (BP), neuron-level pruning (NP), neuron-level sparsity (NS)." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.228, + 0.787, + 0.513 + ], + "angle": 0, + "content": "
ModelMethodPruningTop-1 (%)Top-5 (%)GFLOPsParams (M)
DeiT-TOriginal [24]-72.2091.101.35.7
SSViTE [4]HP70.12-0.94.2
WDPruning [29]HP+BP70.3489.820.73.5
X-Pruner [30]HP71.1090.110.6-
SNPNP70.2990.010.63.0
UVC [31]HP+NS+BP70.60-0.5-
DeiT-SOriginal [24]-79.8595.004.622.1
SSViTE [4]HP79.22-3.114.6
WDPruning [29]HP+BP78.3894.052.613.3
X-Pruner [30]HP78.9394.242.4-
UVC [31]HP+NS+BP78.82-2.3-
SNPNP78.5294.372.010.0
SNPNP73.3291.661.36.4
DeiT-BOriginal [24]-81.8095.5917.686.6
SSViTE [4]HP82.22-11.856.8
WDPruning [29]HP+BP80.7695.369.955.3
X-Pruner [30]HP81.0295.388.5-
UVC [31]HP+NS+BP80.57-8.0-
SNPNP79.6394.376.431.6
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.546, + 0.785, + 0.606 + ], + "angle": 0, + "content": "ImageNet-1K. Throughout the fine-tuning phase of the pruned model, we maintain consistent settings across all models, except for the batch size and learning rate. The batch size is set to 256, and to prevent weight explosion, we adjust the learning rate of the compressed model to 1/10 or 1/100 of the original model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.607, + 0.787, + 0.729 + ], + "angle": 0, + "content": "To evaluate the reduced latency using SNP, we have configured four testing scenarios: one on a CPU and another on GPU for both edge devices and server processors. We employ a standard PyTorch model for profiling on the server processors (Intel Xeon Silver 4210R and NVIDIA GeForce RTX 3090). Profiling on the Raspberry Pi 4B and Jetson Nano is conducted using the ONNX and TensorRT formats, respectively. All latencies are measured using a single image as an input, except for the GPU of the server processor (RTX 3090), where it is set to 64 images." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.753, + 0.427, + 0.768 + ], + "angle": 0, + "content": "4.2 Quantitative results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Comparison with other methods Despite the constraints outlined in Sec. 3.4, SNP not only maintains accuracy comparable to existing methods but also significantly reduces inference time across diverse hardware and data types on the ImageNet-1K dataset, as shown in Tab. 1 and Tab. 2." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.28, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SNP: Structured Neuron-level Pruning to Preserve Attention Scores" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.228 + ], + "angle": 0, + "content": "Table 2: Inference speed and Top-1 accuracy of the compressed model across different devices. Performance evaluation involves accuracy on ImageNet-1K and inference time for the compressed DeiTs and EfficientFormer-L1. Latency is benchmarked with 200 warm-up runs and averaged over 1000 runs. In latency measurement, a single image is used as an input, except for the RTX 3090, where 64 images are employed in a single batch." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.241, + 0.788, + 0.367 + ], + "angle": 0, + "content": "
ModelTop-1 (%)GFLOPsEdge devices (ms)Server processors (ms)
Raspberry Pi 4B (.onnx)Jetson Nano (.trt)Xeon Silver 4210R (.pt)RTX 3090 (.pt)
DeiT-Tiny72.201.3139.1341.0334.7418.65
+ SNP (Ours)70.290.681.63 (1.70×)26.67 (1.54×)25.25 (1.38×)17.82 (1.05×)
DeiT-Small79.804.6401.2799.3253.3746.13
+ SNP (Ours)78.522.0199.15 (2.01×)45.51 (2.18×)38.57 (1.38×)32.91 (1.40×)
+ SNP (Ours)73.321.3136.68 (2.94×)32.03 (3.10×)33.46 (1.60×)26.98 (1.71×)
DeiT-Base81.8017.61377.71293.29122.03151.35
+ SNP (Ours)79.636.4565.68 (2.44×)132.55 (2.21×)64.65 (1.89×)72.96 (2.07×)
EfficientFormer-L179.201.3169.1330.9543.7526.19
+ SNP (Ours)75.530.695.12 (1.78×)19.78 (1.56×)38.25 (1.14×)17.24 (1.52×)
+ SNP (Ours)74.510.582.60 (2.05×)17.76 (1.74×)35.15 (1.24×)16.01 (1.64×)
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.398, + 0.788, + 0.489 + ], + "angle": 0, + "content": "In a recent study [26], unconstrained masking generally outperforms posttraining accuracy of pruned models by an average of \\(2.1\\%\\) on ImageNet-1K. Compared to unconstrained head masking approaches like SSVITE [4] and XPruner [30], SNP achieves significantly higher compression rates for all DeiTs FLOPs (30.64% and 10.64%) with minimal performance degradation (0.83% and 0.87%), much less than the \\(2.1\\%\\) average mentioned." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.489, + 0.788, + 0.595 + ], + "angle": 0, + "content": "Furthermore, compared to other pruning approaches like WDPruning [29] and UVC [31], which use a combination of pruning techniques to compress DeiTs, SNP achieves comparable accuracy solely through neuron-level pruning. Notably, DeiT-Small with SNP outperforms WDPruning by \\(0.14\\%\\), with the removal of 3.3 million parameters and a reduction of 0.6 GFLOPs. Compared to UVC, SNP exhibits negligible performance degradation, averaging \\(0.51\\%\\) across all DeiTs while using \\(7.65\\%\\) fewer FLOPs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.619, + 0.788, + 0.771 + ], + "angle": 0, + "content": "Large compressed vs. Small hand-crafted DeiT-Small with SNP, a large pruned model, outperforms the smaller, hand-crafted DeiT-Tiny in both accuracy and latency, achieving a notable \\(1.12\\%\\) improvement in top-1 accuracy while maintaining similar FLOPs. Additionally, DeiT-Small with SNP exhibits enhanced speed compared to the original DeiT-Tiny across edge devices and CPU-based server processors. Notably, its speed increases up to \\(21.94\\%\\) compared to the original DeiT-Tiny running on Jetson Nano, a GPU-based edge device. This substantial performance gap underscores the superiority of the compressed model (DeiT-Small with SNP) over the smaller hand-crafted designed model (DeiT-Tiny) in both overall performance and speed." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Accelerating Transformer-based models As depicted in Tab. 2, SNP achieves impressive acceleration of DeiTs by a factor of \\(1.44 \\times\\) to \\(2.44 \\times\\) on edge devices and \\(1.05 \\times\\) to \\(2.07 \\times\\) on server processors. This acceleration is notable," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.127 + ], + "angle": 0, + "content": "Shim. et al." + }, + { + "type": "image", + "bbox": [ + 0.333, + 0.145, + 0.675, + 0.441 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.451, + 0.788, + 0.52 + ], + "angle": 0, + "content": "Fig.3: Attention maps with varying pruning criteria and compression ratios. All query and key layers are locally pruned based on the specified pruning ratio without fine-tuning. The importance scores of \\( l2 \\)-norm and GM on query and key layers are combined by filter index and pruned simultaneously. \"Reverse\" represents the reverse order of SNP." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.557, + 0.785, + 0.589 + ], + "angle": 0, + "content": "with negligible average performance degradation of \\(1.79\\%\\), specifically \\(1.91\\%\\), \\(1.28\\%\\), and \\(2.17\\%\\) for DeiT-Tiny, DeiT-Small, and DeiT-Base, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.591, + 0.788, + 0.666 + ], + "angle": 0, + "content": "Compared to WDPruning, which employs both head and block pruning, SNP surpasses in terms of latency for DeiT-Small and DeiT-Base on the RTX 3090 GPU. SNP accelerates the original DeiTs by \\(1.38 \\times\\) and \\(2.07 \\times\\), respectively, while WDPruning achieves a comparatively modest acceleration of \\(1.18 \\times\\) for both models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.671, + 0.788, + 0.762 + ], + "angle": 0, + "content": "The superiority of SNP becomes more evident when the target of neuron-level pruning, linear or convolutional layers, contribute a larger proportion of the original model's computation time. we find that SNP further accelerates original model \\(1.05 \\times\\) to \\(2.07 \\times\\) as the DeiT model's size grows on the RTX 3090 GPU. In contrast, WDPruning, involving the removal of entire layers and associated operators in both head and block, consistently shows acceleration rates of \\(1.18 \\times\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "To ascertain the robustness of SNP across diverse Transformer models, especially on the efficiently designed model, we conduct additional experiments on EfficientFormer-L1. SNP accelerates the model by \\(1.78 \\times\\) and \\(2.05 \\times\\) faster on Raspberry Pi 4B and \\(1.56 \\times\\) and \\(1.74 \\times\\) faster on Jetson Nano. Notably, the compressed EfficientFormer-L1 achieves acceptable accuracy of \\(75.53\\%\\) and \\(74.51\\%\\)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.28, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "SNP: Structured Neuron-level Pruning to Preserve Attention Scores" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.145, + 0.311, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.253, + 0.215, + 0.28, + 0.222 + ], + "angle": 0, + "content": "Original" + }, + { + "type": "image", + "bbox": [ + 0.314, + 0.145, + 0.403, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.339, + 0.216, + 0.379, + 0.222 + ], + "angle": 0, + "content": "Compressed" + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.145, + 0.496, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.411, + 0.216, + 0.493, + 0.222 + ], + "angle": 0, + "content": "Compressed + Fine-tuned" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.145, + 0.597, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.539, + 0.216, + 0.565, + 0.222 + ], + "angle": 0, + "content": "Original" + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.145, + 0.69, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.626, + 0.216, + 0.665, + 0.222 + ], + "angle": 0, + "content": "Compressed" + }, + { + "type": "image", + "bbox": [ + 0.693, + 0.145, + 0.784, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.697, + 0.216, + 0.778, + 0.222 + ], + "angle": 0, + "content": "Compressed + Fine-tuned" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.223, + 0.311, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.231, + 0.312, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.238, + 0.312, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.246, + 0.312, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.254, + 0.312, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.262, + 0.312, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.269, + 0.312, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.277, + 0.312, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.285, + 0.312, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.292, + 0.312, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.299, + 0.312, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.306, + 0.312, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.314, + 0.223, + 0.403, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.314, + 0.231, + 0.403, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.314, + 0.238, + 0.403, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.314, + 0.245, + 0.403, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.314, + 0.253, + 0.403, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.314, + 0.262, + 0.403, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.314, + 0.269, + 0.403, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.314, + 0.277, + 0.403, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.314, + 0.285, + 0.403, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.314, + 0.292, + 0.403, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.314, + 0.299, + 0.403, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.314, + 0.306, + 0.403, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.223, + 0.496, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.231, + 0.496, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.238, + 0.496, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.245, + 0.496, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.253, + 0.496, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.262, + 0.496, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.269, + 0.496, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.277, + 0.496, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.284, + 0.496, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.292, + 0.496, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.299, + 0.496, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.306, + 0.496, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.223, + 0.597, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.231, + 0.597, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.238, + 0.597, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.245, + 0.597, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.253, + 0.597, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.262, + 0.597, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.269, + 0.597, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.277, + 0.597, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.285, + 0.597, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.292, + 0.597, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.3, + 0.597, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.307, + 0.597, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.313, + 0.597, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "list", + "bbox": [ + 0.507, + 0.223, + 0.597, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.6, + 0.223, + 0.69, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.6, + 0.231, + 0.69, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.6, + 0.238, + 0.69, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.6, + 0.245, + 0.69, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.6, + 0.253, + 0.69, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.6, + 0.262, + 0.69, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.6, + 0.269, + 0.69, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.6, + 0.277, + 0.69, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.6, + 0.285, + 0.69, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.6, + 0.292, + 0.69, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.6, + 0.3, + 0.69, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.6, + 0.307, + 0.69, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.6, + 0.313, + 0.69, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "list", + "bbox": [ + 0.6, + 0.223, + 0.69, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.693, + 0.223, + 0.78, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.693, + 0.229, + 0.783, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.693, + 0.236, + 0.783, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.693, + 0.244, + 0.783, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.693, + 0.252, + 0.783, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.693, + 0.26, + 0.783, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.693, + 0.268, + 0.783, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.693, + 0.276, + 0.783, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.693, + 0.284, + 0.783, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.693, + 0.292, + 0.783, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.693, + 0.3, + 0.783, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.693, + 0.307, + 0.783, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.33, + 0.787, + 0.387 + ], + "angle": 0, + "content": "Fig. 4: Attention maps from the original, compressed, and fine-tuned DeiT-Tiny with SNP. The attention maps in the first row are visualized using the attention rollout [1]. Each red box contains three attention maps from each head of the MSA module, ordered accordingly." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.417, + 0.415, + 0.431 + ], + "angle": 0, + "content": "4.3 Qualitative results" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.443, + 0.787, + 0.489 + ], + "angle": 0, + "content": "In the subsequent sections, attention rollout [1] is employed to randomly selected images from the test datasets to visualize the attention maps of the DeiTs and assess the efficacy of SNP. This evaluation unfolds across two key facets:" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.499, + 0.785, + 0.529 + ], + "angle": 0, + "content": "- Preserving the attention scores: Visualize attention maps to verify the effectiveness of the proposed importance score in preserving attention scores." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.53, + 0.785, + 0.56 + ], + "angle": 0, + "content": "- Restoring the attention scores after SNP: Visualize attention maps to find the effectiveness of SNP in restoring attention scores." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.499, + 0.785, + 0.56 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.583, + 0.787, + 0.629 + ], + "angle": 0, + "content": "Preserving the attention scores To visualize SNP effectiveness, we apply four pruning criteria to compress the query and key filter pairs in all MSA modules of DeiT-Tiny: SNP, \\( l2 \\)-norm, geometric median (GM) [11], and \"Reverse\"." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.63, + 0.787, + 0.719 + ], + "angle": 0, + "content": "The \"Reverse\" criterion prioritizes pruning the most important filter pairs first, keeping the least important pairs until the end. However, both \\( l2 \\)-norm and GM pruning criteria ignore the MSA module's graphical connectivity, evaluating importance scores independently for query and key layers. To handle identical filter indices during pruning, we aggregate scores based on filter indices, removing the least important indices from both layers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.787, + 0.841 + ], + "angle": 0, + "content": "In Fig. 3, attention maps for the original DeiT-Tiny and locally pruned models are presented across various pruning ratios (10% to 90% with 10% intervals). Our proposed method effectively maintains the original attention map even after pruning over 80% of the filters, whereas other methods show fragmented attention maps at much lower pruning ratios (typically 30% or less). These results highlight the potential of our neuron-level pruning criteria, utilizing SVD to preserve attention scores, in reducing the size and speeding up the execution of MSA modules without compromising accuracy." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.128 + ], + "angle": 0, + "content": "Shim. et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 3: Performance of SNP without fine-tuning across different data quantities at various pruning ratios. To determine the proper number of images for calculating the importance score, we conduct local pruning on the query and key layers at pruning ratios of \\(10\\%\\), \\(30\\%\\), \\(50\\%\\), \\(70\\%\\), and \\(90\\%\\). All performance metrics are assessed without fine-tuning. The latency is measured on Raspberry Pi 4B." + }, + { + "type": "table", + "bbox": [ + 0.242, + 0.227, + 0.758, + 0.369 + ], + "angle": 0, + "content": "
Number of imagesPerformance by pruning ratio
Original10%30%50%70%90%
172.2071.1567.9657.6038.6811.17
472.2070.8267.8657.6538.0511.86
1672.2070.7267.4258.6039.469.11
6472.2070.8367.5559.5042.1312.83
25672.2070.7568.1459.5842.7014.48
Latency (ms)139.13133.60129.29119.41117.48112.32
Params (M)5.75.595.415.235.064.88
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.398, + 0.787, + 0.503 + ], + "angle": 0, + "content": "Restoring the attention scores after SNP Fig. 4 illustrates the overall and per-head attention maps of the original, compressed, and fine-tuned DeiT-Tiny, respectively. The first row shows the overall attention maps of the respective models. Notably, the compressed model maintains a well-preserved overall attention map, despite pruning all layers, including values and FFN, resulting in a \\(53\\%\\) reduction in FLOPs and a \\(46\\%\\) reduction in parameters. Especially, we can observe that attention map is well-restored after the fine-tuning process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.504, + 0.787, + 0.579 + ], + "angle": 0, + "content": "The twelve red boxes below the overall attention map depict per-head attention maps for twelve layers of each original, compressed, and fine-tuned models respectively. As shown in Fig. 4, it is evident that the attention maps for each head are effectively preserved and restored in each of the compressed and fine-tuned models." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.603, + 0.4, + 0.617 + ], + "angle": 0, + "content": "4.4 Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.787, + 0.719 + ], + "angle": 0, + "content": "Importance scores by the data quantity Since the attention scores depend on the input, as indicated in Eq. (1), the proposed importance scores for query and key filter pairs (Eq. (6)) may exhibit sensitivity to the distribution of the input image \\( X \\). To validate the method's robustness, we compute importance scores using various image quantities, pruning query and key layers at different ratios, without fine-tuning process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "As depicted in Tab. 3, SNP demonstrates a slight advantage in preserving performance with an increasing number of images, outperforming models compressed with fewer images as the compression ratio rises. However, this improvement comes at the cost of increased computation time for SNP calculations. Considering these factors into account, we opt to use 64 images, which yield the second-best performance among the given number of images. This decision strikes a balance between achieving satisfactory performance and maintaining computational efficiency." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.28, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SNP: Structured Neuron-level Pruning to Preserve Attention Scores" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.321, + 0.147, + 0.677, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.395, + 0.788, + 0.452 + ], + "angle": 0, + "content": "Fig. 5: Top-1 accuracy of compressed DeiT-Tiny on ImageNet using several pruning criteria without fine-tuning. Query and key layers are locally pruned using various pruning criteria : SNP, GM, l2-norm, reverse order of SNP (\"Reverse\"), and original DeiT-Tiny (\"Baseline\")." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.481, + 0.787, + 0.526 + ], + "angle": 0, + "content": "Performance comparison across pruning ratios To assess the robustness of SNP, we examine the performance of DeiT-Tiny under various pruning criteria and different pruning ratios, as described in Sec. 4.3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.527, + 0.787, + 0.587 + ], + "angle": 0, + "content": "As illustrated in Fig. 5, SNP consistently outperforms other pruning criteria across all pruning ratios. In contrast, models compressed with \"Reverse\" criteria exhibit the lowest performance at all pruning ratios, underscoring the robustness of the proposed approach." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.587, + 0.787, + 0.648 + ], + "angle": 0, + "content": "Both Fig. 3 and Fig. 5 confirm that SNP successfully preserves attention scores in both quantitative and qualitative aspects. Even with an \\(80\\%\\) pruning ratio and without fine-tuning, SNP maintains original attention scores and outperforms other pruning criteria." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.672, + 0.36, + 0.687 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.787, + 0.81 + ], + "angle": 0, + "content": "In this paper, we propose a novel graph-aware neuron-level pruning method, SNP, designed to compress and accelerate Transformer-based models. SNP proposes two pruning criteria for preserving attention scores and eliminating interhead redundancy. Using SNP, a large compressed model outperforms small, hand-crafted designed models in both performance and latency on edge devices. Moreover, the compressed models exhibit astonishing results in latency on various devices, with negligible performance degradation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.841 + ], + "angle": 0, + "content": "As this work is a first attempt to accelerate MSA modules using neuron-level pruning alone, many challenges remain. One is incorporating other pruning" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.128 + ], + "angle": 0, + "content": "Shim. et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.783, + 0.192 + ], + "angle": 0, + "content": "methods such as head or block pruning for a more efficient Transformer model. Another challenge is applying SNP to other vision tasks, including image generation, which requires high computational costs on both training and inference." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.784, + 0.253 + ], + "angle": 0, + "content": "We believe that these works encourage the adoption of model pruning as a tool, for both improving the applicability of ViTs in resource-constrained environments and reducing the training costs of large models by integrating with training process." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.275, + 0.403, + 0.292 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.304, + 0.784, + 0.351 + ], + "angle": 0, + "content": "This work was supported by Artificial intelligence industrial convergence cluster development project funded by the Ministry of Science and ICT(MSIT, Korea)&Gwangju Metropolitan City." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.372, + 0.323, + 0.387 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.401, + 0.785, + 0.429 + ], + "angle": 0, + "content": "1. Abnar, S., Zuidema, W.: Quantifying attention flow in transformers. arXiv preprint arXiv:2005.00928 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.43, + 0.785, + 0.484 + ], + "angle": 0, + "content": "2. Bian, Y., Huang, J., Cai, X., Yuan, J., Church, K.: On attention redundancy: A comprehensive study. In: Proceedings of the 2021 conference of the north american chapter of the association for computational linguistics: human language technologies. pp. 930-945 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.485, + 0.785, + 0.511 + ], + "angle": 0, + "content": "3. Bolya, D., Fu, C.Y., Dai, X., Zhang, P., Feichtenhofer, C., Hoffman, J.: Token merging: Your vit but faster. arXiv preprint arXiv:2210.09461 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.512, + 0.785, + 0.552 + ], + "angle": 0, + "content": "4. Chen, T., Cheng, Y., Gan, Z., Yuan, L., Zhang, L., Wang, Z.: Chasing sparsity in vision transformers: An end-to-end exploration. Advances in Neural Information Processing Systems 34, 19974-19988 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.553, + 0.785, + 0.579 + ], + "angle": 0, + "content": "5. Demouth, J.: Sparse matrix-matrix multiplication on thegpu. Tech. rep., NVIDIA (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.58, + 0.785, + 0.621 + ], + "angle": 0, + "content": "6. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.621, + 0.785, + 0.675 + ], + "angle": 0, + "content": "7. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.676, + 0.785, + 0.717 + ], + "angle": 0, + "content": "8. Fang, G., Ma, X., Song, M., Mi, M.B., Wang, X.: Depgraph: Towards any structural pruning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16091-16101 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.718, + 0.785, + 0.758 + ], + "angle": 0, + "content": "9. Han, S., Mao, H., Dally, W.J.: Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.759, + 0.785, + 0.799 + ], + "angle": 0, + "content": "10. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "11. He, Y., Liu, P., Wang, Z., Hu, Z., Yang, Y.: Filter pruning via geometric median for deep convolutional neural networks acceleration. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4340-4349 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.219, + 0.401, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.28, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SNP: Structured Neuron-level Pruning to Preserve Attention Scores" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "12. He, Y., Zhang, X., Sun, J.: Channel pruning for accelerating very deep neural networks. In: Proceedings of the IEEE international conference on computer vision. pp. 1389-1397 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.218 + ], + "angle": 0, + "content": "13. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.245 + ], + "angle": 0, + "content": "14. Lee, J., Park, S., Mo, S., Ahn, S., Shin, J.: Layer-adaptive sparsity for the magnitude-based pruning. arXiv preprint arXiv:2010.07611 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.785, + 0.273 + ], + "angle": 0, + "content": "15. Li, H., Kadav, A., Durdanovic, I., Samet, H., Graf, H.P.: Pruning filters for efficient convnets. arXiv preprint arXiv:1608.08710 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.314 + ], + "angle": 0, + "content": "16. Li, Y., Yuan, G., Wen, Y., Hu, J., Evangelidis, G., Tulyakov, S., Wang, Y., Ren, J.: Efficientformer: Vision transformers at mobilenet speed. Advances in Neural Information Processing Systems 35, 12934-12949 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.314, + 0.785, + 0.369 + ], + "angle": 0, + "content": "17. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.37, + 0.785, + 0.411 + ], + "angle": 0, + "content": "18. Liu, Z., Wang, Y., Han, K., Zhang, W., Ma, S., Gao, W.: Post-training quantization for vision transformer. Advances in Neural Information Processing Systems 34, 28092-28103 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.785, + 0.439 + ], + "angle": 0, + "content": "19. Michel, P., Levy, O., Neubig, G.: Are sixteen heads really better than one? Advances in neural information processing systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.481 + ], + "angle": 0, + "content": "20. Mishra, A., Latorre, J.A., Pool, J., Stosic, D., Stosic, D., Venkatesh, G., Yu, C., Micikevicius, P.: Accelerating sparse deep neural networks. arXiv preprint arXiv:2104.08378 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.522 + ], + "angle": 0, + "content": "21. Nakajima, S., Sugiyama, M., Babacan, S.D., Tomioka, R.: Global analytic solution of fully-observed variational bayesian matrix factorization. The Journal of Machine Learning Research 14(1), 1-37 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.785, + 0.564 + ], + "angle": 0, + "content": "22. Radosavovic, I., Kosaraju, R.P., Girshick, R., He, K., Dólár, P.: Designing network design spaces. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10428-10436 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.785, + 0.605 + ], + "angle": 0, + "content": "23. Tan, M., Le, Q.: Efficientnet: Rethinking model scaling for convolutional neural networks. In: International conference on machine learning. pp. 6105-6114. PMLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.605, + 0.785, + 0.647 + ], + "angle": 0, + "content": "24. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jégou, H.: Training data-efficient image transformers & distillation through attention. In: International conference on machine learning. pp. 10347-10357. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.688 + ], + "angle": 0, + "content": "25. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.688, + 0.785, + 0.73 + ], + "angle": 0, + "content": "26. Wan, A., Hao, H., Patnaik, K., Xu, Y., Hadad, O., Güera, D., Ren, Z., Shan, Q.: Upscale: unconstrained channel pruning. In: International Conference on Machine Learning. pp. 35384-35412. PMLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.785, + 0.757 + ], + "angle": 0, + "content": "27. Wang, Z.: Sparsednn: Fast sparse deep learning inference on cpus. arXiv preprint arXiv:2101.07948 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "28. Xie, Q., Luong, M.T., Hovy, E., Le, Q.V.: Self-training with noisy student improves imagenet classification. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10687-10698 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.841 + ], + "angle": 0, + "content": "29. Yu, F., Huang, K., Wang, M., Cheng, Y., Chu, W., Cui, L.: Width & depth pruning for vision transformers. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 3143-3151 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.127 + ], + "angle": 0, + "content": "Shim. et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "30. Yu, L., Xiang, W.: X-pruner: explainable pruning for vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24355-24363 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.218 + ], + "angle": 0, + "content": "31. Yu, S., Chen, T., Shen, J., Yuan, H., Tan, J., Yang, S., Liu, J., Wang, Z.: Unified visual transformer compression. arXiv preprint arXiv:2203.08243 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.259 + ], + "angle": 0, + "content": "32. Yu, S., Mazaheri, A., Jannesari, A.: Topology-aware network pruning using multi-stage graph embedding and reinforcement learning. In: International conference on machine learning. pp. 25656-25667. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.785, + 0.301 + ], + "angle": 0, + "content": "33. Yu, X., Serra, T., Ramalingam, S., Zhe, S.: The combinatorial brain surgeon: pruning weights that cancel one another in neural networks. In: International Conference on Machine Learning. pp. 25668-25683. PMLR (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.301 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/SNP_ Structured Neuron-level Pruning to Preserve Attention Scores/cb547506-91af-4f34-aebd-6546cde66d30_origin.pdf b/2024/SNP_ Structured Neuron-level Pruning to Preserve Attention Scores/cb547506-91af-4f34-aebd-6546cde66d30_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..64858a6010f66c23e7091843eaa139088be1cbe9 --- /dev/null +++ b/2024/SNP_ Structured Neuron-level Pruning to Preserve Attention Scores/cb547506-91af-4f34-aebd-6546cde66d30_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f52bd5b6838e30b06fb8f1aaf2ce86324fcd83442fc7aaf48fd282a32e2f74f5 +size 2416237 diff --git a/2024/SNP_ Structured Neuron-level Pruning to Preserve Attention Scores/full.md b/2024/SNP_ Structured Neuron-level Pruning to Preserve Attention Scores/full.md new file mode 100644 index 0000000000000000000000000000000000000000..45785bedb2baeef0acfd0b3c1f91a896854662ca --- /dev/null +++ b/2024/SNP_ Structured Neuron-level Pruning to Preserve Attention Scores/full.md @@ -0,0 +1,431 @@ +# SNP: Structured Neuron-level Pruning to Preserve Attention Scores + +Kyunghwan Shim1, Jaewoong Yun1, and Shinkook Choi1* + +$^{1}$ Nota Inc., Korea $^{2}$ {kyunghwan.shim, jwyun, shinkook. Choi}@nota.ai + +Abstract. Multi-head self-attention (MSA) is a key component of Vision Transformers (ViTs), which have achieved great success in various vision tasks. However, their high computational cost and memory footprint hinder their deployment on resource-constrained devices. Conventional pruning approaches can only compress and accelerate the MSA module using head pruning, although the head is not an atomic unit. To address this issue, we propose a novel graph-aware neuron-level pruning method, Structured Neuron-level Pruning (SNP). SNP prunes neurons with less informative attention scores and eliminates redundancy among heads. Specifically, it prunes graphically connected query and key layers having the least informative attention scores while preserving the overall attention scores. Value layers, which can be pruned independently, are pruned to eliminate inter-head redundancy. Our proposed method effectively compresses and accelerates Transformer-based models for both edge devices and server processors. For instance, the DeiT-Small with SNP runs 3.1 times faster than the original model and achieves performance that is $21.94\%$ faster and $1.12\%$ higher than the DeiT-Tiny. Additionally, SNP accelerates the efficiently designed Transformer model, EfficientFormer, by 1.74 times on the Jetson Nano with acceptable performance degradation. Source code is at https://github.com/Nota-NetsPresso/SNP + +Keywords: Network pruning $\cdot$ Network compression $\cdot$ Compact models + +# 1 Introduction + +Vision Transformers (ViTs) [7,17,24] have outperformed or matched the performance of state-of-the-art convolutional neural networks (CNNs) [10,22,23,28] on various computer vision tasks. The success of ViTs is attributed to the Multihead Self-Attention (MSA) module, which captures intricate relationships in data. However, the Transformer architecture entails substantial computational resources, posing challenges for practical applications on edge devices with constrained storage and computational capabilities. To address this issue, we leverage the graphical components of the MSA module to reduce the dimension of interconnected layers, aimed at effectively reducing their computing budgets while achieving hardware-agnostic speedups. + +![](images/c6172df040d4b81d7cc5f6a8ea3566eddbe83324bf39d6bcdb5a235143dc073f.jpg) +Fig. 1: Comparison of model size, speed, and performance. ImageNet-1K classification results. Latency is profiled by Raspberry Pi 4B. The connected lines represent the compressed models paired with the original model. The size of each circle indicates the number of parameters in respective model. The number adjacent to each compressed model indicates its compressed FLOPs. + +From the perspective of structured pruning, which aims to reduce the number of dimensions in convolutional or linear layers, the MSA module contains two prunable objectives: heads and neurons. Head pruning, which removes the number of heads, is relatively intuitive to implement due to the reduced complexity of graphical elements compared to neuron-level pruning. In contrast, neuron-level pruning reduces the dimension of individual layers in each head of the MSA module, necessitating a comprehensive understanding of the graphical connectivity of the MSA module. A recent work [31] implements neuron-level pruning by zeroing out individual filters without considering the model's graphical connectivity. This indiscriminate zeroing can negatively affect both accuracy and throughput performance [26]. + +In this paper, we introduce a novel graph-aware neuron-level pruning method called Structured Neuron-level Pruning (SNP) to accelerate and compress ViTs effectively. We propose two pruning criteria based on the function of each layer within the MSA module. SNP prunes filter pairs of query and key layers containing fewer contributions to attention scores. Moreover, SNP aims to reduce redundancy across heads by eliminating the redundant filter from the value layer. + +Furthermore, by removing identical filter indices for all the graphically connected layers, SNP could accelerate various Transformer models on various devices without additional libraries, as shown in Fig. 1. + +In summary, the major contributions of this paper are: + +- We propose a novel graph-aware neuron-level pruning method, SNP, for Transformer models. SNP is the first method to use the graphical characteristics of the MSA module to measure the importance score of neurons. +- To the best of our knowledge, this is the first work to accelerate Transformer models on any device using neuron-level pruning only. +- SNP achieves significant acceleration while maintaining the original performance on several models. Compressed DeiT-Small outperforms DeiT-Tiny by $1.12\%$ in accuracy, with similar FLOPs, and reduces inference time on various edge devices. Furthermore, the proposed method accelerates the efficiently designed Transformer model, EfficientFormer [16], more than two times with acceptable performance degradation. + +# 2 Related Work + +# 2.1 Compressing vision Transformers + +The ViTs [7, 17, 24] achieve high performance in numerous vision tasks without specialized image processing modules such as convolutions. The key concept of ViTs is to segment images into patch sequences, convert these patches to token embeddings, and then process them through Transformer encoders [25]. ViTs consist only of Transformer blocks, making them likely to be over-parameterized. For this reason, recent works have aimed to reduce computational cost [3, 24] and be memory efficient [18, 26]. DeiT [24] proposes lightweight ViT architectures through knowledge distillation [13]. ToMe [3] proposes accelerating ViTs by directly combining similar tokens, without the need for training. Liu et al. [18] propose post-training quantization method using nuclear norm based mixed-precision scheme that does not require fine-tuning for the Vision Transformer. + +# 2.2 Pruning vision Transformers + +Unstructured and structured pruning Pruning methods can be broadly categorized into two types, unstructured and structured pruning. Unstructured pruning sets individual weights or parameters to zero, resulting in irregular sparse matrices [9, 14]. Compressed models using unstructured pruning tend to maintain relatively high performance for a given pruning ratio. However, they necessitate additional libraries, such as cuSPARSE [5], Automatic SParsity [20], or SparseDNN [27] to accelerate sparse matrix computations. + +Structured pruning, on the other hand, involves the removal of entire groups of units, such as filters or attention heads. This can be implemented using "masking" (zeroing out) [11, 12, 32, 33], or by "pruning" [8, 15]. Structured pruning by masking [11, 12, 32, 33] simply sets the group of units to zero, which requires additional libraries to accelerate the model, as unstructured pruning. "Pruning" [8, 15], on the other hand, requires a comprehensive understanding of the network's graphical connectivity, including element-wise operations that enforce the same input shape. By considering the graphical connectivity and pruning identical filter indices for inter-connected layers, structured pruning can achieve acceleration on any devices. + +![](images/e5fc9c5a11723ce6c34fb414641fd391df21360b5ad2c1f35e852f92d78d0aca.jpg) +Fig. 2: Proposed SNP methods, on each prunable component of the Transformer block. (a) SNP for query and key layers to preserve attention scores. (b) prunable components of the Transformer block. (c) SNP for value and other layers, including FFN and patch embedding. (d) conventional zeroing out in the matrix multiplication operator. (e) conventional zeroing out and graph-aware pruning in the residual connection. + +![](images/5d3c6fee3f2a567b24a6f3b0f3b7472b81693d0894ed3fc18078d60c048306ed.jpg) + +Zeroing out + +Head and neuron-level pruning Structured pruning for the MSA module has two pruning objectives: head and neuron. Head pruning [29,30] reduces the number of heads, while neuron-level pruning [31] reduces the dimension of each query, key, and value layer in each head. Recent studies for pruning ViTs have focused on head pruning. X-Pruner [30] proposes a novel head pruning method for ViTs that introduces explainability-aware masks and measure the head importance, resulting in superior model compression. WDPruning [29] propose a method to control the number of attention heads and blocks via threshold on learnable parameters. + +UVC [31] utilizes knowledge distillation alongside several pruning techniques, such as head pruning, block pruning, and neuron-level pruning. However, neuron-level pruning of UVC is carried out in a masking (zeroing out) manner, converting the weight matrix into a sparse matrix. For this reason, UVC necessitates additional libraries or hardware for accelerating sparse matrices, otherwise, the compressed model with UVC cannot achieve latency gain from the neuron-level pruning. + +# 3 Methodology + +# 3.1 Preliminaries + +MSA module takes a single input $X \in \mathbb{R}^{N \times d}$ , where $N$ denotes the input vector length, and $d$ represents the hidden size. It comprises $H$ heads, each consisting + +of three linear layers: query, key, and value. Each layer denoted as $W_{\{q,k,v\}}^h \in \mathbb{R}^{d \times d_{\{q,k,v\}}}$ , where $h$ represents $h$ -th head and $d_{\{q,k,v\}}$ indicates the hidden size of each query, key, and value layer. $\{Q,K,V\}^h$ signifies the output of each query, key, and value layer in $h$ -th head, with a shape of $\mathbb{R}^{N \times d_{\{q,k,v\}}}$ . + +The self-attention operation for $h$ -th head can be expressed as follows: + +$$ +\operatorname {A s} ^ {h} (X) = Q ^ {h} \cdot \left(K ^ {h}\right) ^ {T} \tag {1} +$$ + +$$ +\operatorname {A t t} ^ {h} (X) = \operatorname {S o f t m a x} \left(\frac {\operatorname {A s} ^ {h} (X)}{\sqrt {d _ {q}}}\right) \cdot V ^ {h} \tag {2} +$$ + +$$ +\operatorname {M S A} (X) = \operatorname {C o n c a t} \left(\operatorname {A t t} ^ {1} (X), \dots , \operatorname {A t t} ^ {H} (X)\right) \tag {3} +$$ + +here, As, Att, and MSA represent the functions to calculate attention scores, attention module, and MSA module respectively. + +Matrix multiplication, unlike a residual connection, yields zero when either of the inputs is masked, regardless of the value of the other input. Therefore, applying a conventional zeroing-out approach to neuron-level pruning in the MSA module can lead to unintended results, as shown in Fig. 2 (d). To address this issue, we introduce two graph-aware neuron-level pruning criteria to compress and expedite the MSA module. + +# 3.2 Preserving attention scores + +Attention scores, Eq. (1), of the MSA module learn long-range dependencies between image features by focusing on different parts of the image when processing different features. These attention scores can be recognized as a series of outer product of $Q_{i}^{h}$ and $(K_{i}^{h})^{T}$ as follows: + +$$ +\begin{array}{l} \operatorname {A s} ^ {h} (X) = Q ^ {h} \cdot \left(K ^ {h}\right) ^ {T} \\ = \sum_ {i = 1} ^ {d _ {q}} Q _ {i} ^ {h} \cdot \left(K _ {i} ^ {h}\right) ^ {T} \tag {4} \\ \end{array} +$$ + +from this perspective, neuron-level pruning, reducing the dimension of query $d_{q}$ and key $d_{k}$ , inevitably distorts the attention scores. For this reason, preserving attention scores is essential to maintain the high performance of the original model. + +To alleviate the distortion, we maintain the graphically connected query-key filter pair $(Q_{i}^{h}$ and $K_{i}^{h})$ , constituting a filter-by-attention score $(Q_{i}^{h}\cdot (K_{i}^{h})^{T}\in$ $\mathbb{R}^{N\times N})$ , that retains the most significant aspects of the overall attention scores. To identify the most informative filter pair, we initially employ singular value decomposition (SVD) to decompose the original model's attention scores. + +$$ +\begin{array}{l} \operatorname {A s} ^ {h} (X) = \tilde {U} \cdot \tilde {S} \cdot \tilde {V} ^ {T} \\ = \sum_ {j = 1} ^ {N} \tilde {\sigma} _ {j} \cdot \tilde {u} _ {j} \cdot \tilde {v} _ {j} ^ {T} \tag {5} \\ \end{array} +$$ + +where $\tilde{U}$ and $\tilde{V}$ are the left and right singular vector matrices, respectively, and $\tilde{S}$ is the diagonal matrix of singular values, with $\tilde{\sigma}_1 \geq \tilde{\sigma}_2 \geq \ldots \geq \tilde{\sigma}_N$ . + +SVD is a technique for extracting the most informative components of a matrix, those with large singular values, while discarding the less informative components, those with small singular values. To retain the spatial relationships captured by the attention mechanism, we prune the filter pair $Q_{i}^{h}$ and $K_{i}^{h}$ with least correlation with the most informative components of the attention scores. + +To measure the correlation, we adopt the cosine similarity between the attention scores of the $i$ -th filter and the $j$ -th rank matrix, as normalizing the attention scores reduces the impact of magnitude-based approaches. Consequently, the importance score $\omega_{as,i}^{h}$ is defined as follows: + +$$ +\begin{array}{l} \omega_ {a s, i} ^ {h} = \sum_ {j = 1} ^ {r} | \cos ((Q _ {i} ^ {h} \cdot (K _ {i} ^ {h}) ^ {T}), (\tilde {\sigma} _ {j} \cdot \tilde {u} _ {j} \cdot \tilde {v} _ {j} ^ {T})) | \\ = \sum_ {j = 1} ^ {r} | \frac {\left(Q _ {i} ^ {h} \cdot \left(K _ {i} ^ {h}\right) ^ {T}\right) \cdot \left(\tilde {\sigma} _ {j} \cdot \tilde {u} _ {j} \cdot \tilde {v} _ {j} ^ {T}\right)}{| | \left(Q _ {i} ^ {h} \cdot \left(K _ {i} ^ {h}\right) ^ {T}\right) | | \cdot | | \left(\tilde {\sigma} _ {j} \cdot \tilde {u} _ {j} \cdot \tilde {v} _ {j} ^ {T}\right) | |} | \tag {6} \\ \end{array} +$$ + +where $r$ is a hyperparameter dictating the quantity of ranks, within the range of $1 \leq r \leq N$ , to be compared with the $i$ -th attention scores, while the remaining $N - r$ singular values $(\tilde{\sigma}_{r + 1},\dots,\tilde{\sigma}_N)$ are discarded. + +Optimizing $r$ for each attention module can contribute to preserving informative filters and sustaining higher performance. For this reason, we adopt Variational Bayesian Matrix Factorization (VBMF) [21] to determine the optimal rank $r$ for the attention scores. VBMF inherently addresses the challenge of determining the number of latent factors (ranks) in matrix factorization using its probabilistic approach. + +Importance score $\omega_{as,i}^{h}$ , defined in Eq. (6), represents the importance of the $i$ -th filter for both query and key layers. A larger $\omega_{as,i}^{h}$ indicates that the filter has a greater impact on the main component of the attention scores, while a lower $\omega_{as,i}^{h}$ indicates that the filter is less important and can be removed without significantly affecting the attention scores. + +# 3.3 Inter-head redundancy removal + +In the preceding section, we outlined the approach to preserving attention scores even with reduced embedding dimensions in the query and key layers. Here, we introduce a pruning method for the value and other layers, such as FFN or patch embedding layer. + +Previous works [2,19] have revealed that a significant proportion of attention heads can be removed without causing significant performance deterioration. To remove this inter-head redundancy through neuron-level pruning, we propose to measure the distance between all the value layers of MSA module, irrespective of the heads. The importance score $\omega_{v,i}^{h}$ of the $i$ -th filter in the value layer of the + +$h$ -th head is as follows: + +$$ +\omega_ {v, i} ^ {h} = \sum_ {j = 1} ^ {H} \sum_ {l = 1} ^ {d _ {v}} \left(1 - \left| \cos \left(W _ {i} ^ {h}, W _ {l} ^ {j}\right) \right|\right) \tag {7} +$$ + +The importance score of the value layer, denoted as $\omega_{v,i}^{h}$ , indicates the redundancy of filter $i$ in the value layer compared to all value layers of the heads within the corresponding MSA layer. Consequently, filters with the lowest importance scores will be pruned in the initial stages of the pruning process. + +# 3.4 Accelerating Transformers + +As described in Sec. 3.2, SNP removes the least correlated filter pairs $(Q_{i}^{h}, K_{i}^{h})$ to preserve attention scores and enhance the efficiency of the MSA module across various devices. + +Most MSA modules, including MSA modules in DeiTs, are designed to compute all heads in parallel. For achieving this parallel computation, SNP ensures that the number of filter dimensions are consistent across heads, even though the filter indices for each head are independently selected. + +The green highlighted boxes in Fig. 2 (b) represent layers connected by a single residual connection at the last add layer of the Transformer block. Unlike the matrix multiplication operator, the output of the residual connection becomes zero when all the interconnected layers return zero for specific filter indices. For this reason, the residual connection with masking always exceeds the performance of actual pruning, which restricts the set of possible pruning patterns [26]. + +To accelerate the residual connection layers, all interconnected layers should be pruned identically. To achieve this, we sum up all the calculated importance scores for interconnected layers based on their filter indices, as shown in Fig. 2 (e). Subsequently, we prune the least important filter indices of all the connected layers. + +# 4 Experimental Results + +To ensure a fair comparison with existing methods, we apply SNP to prune the DeiT [24] architectures trained on the ImageNet-1K [6] dataset. Additionally, we conduct experiments on the efficient Transformer model, EfficientFormer-L1, to confirm the robustness of SNP. Furthermore, a series of ablation studies are conducted to gain a comprehensive understanding of our methodology. + +# 4.1 Implementation details + +The overall pruning and fine-tuning process are executed on the pre-trained $\mathrm{DeiT}^3$ and EfficientFormer-L1 $^4$ released from the official implementation on + +Table 1: Performance comparison of various pruning methods on ImageNet-1K. The "Pruning" column represents the pruning methods that corresponding methods use to compress the MSA module. Each of the methods contains one of follows: head pruning (HP), block pruning (BP), neuron-level pruning (NP), neuron-level sparsity (NS). + +
ModelMethodPruningTop-1 (%)Top-5 (%)GFLOPsParams (M)
DeiT-TOriginal [24]-72.2091.101.35.7
SSViTE [4]HP70.12-0.94.2
WDPruning [29]HP+BP70.3489.820.73.5
X-Pruner [30]HP71.1090.110.6-
SNPNP70.2990.010.63.0
UVC [31]HP+NS+BP70.60-0.5-
DeiT-SOriginal [24]-79.8595.004.622.1
SSViTE [4]HP79.22-3.114.6
WDPruning [29]HP+BP78.3894.052.613.3
X-Pruner [30]HP78.9394.242.4-
UVC [31]HP+NS+BP78.82-2.3-
SNPNP78.5294.372.010.0
SNPNP73.3291.661.36.4
DeiT-BOriginal [24]-81.8095.5917.686.6
SSViTE [4]HP82.22-11.856.8
WDPruning [29]HP+BP80.7695.369.955.3
X-Pruner [30]HP81.0295.388.5-
UVC [31]HP+NS+BP80.57-8.0-
SNPNP79.6394.376.431.6
+ +ImageNet-1K. Throughout the fine-tuning phase of the pruned model, we maintain consistent settings across all models, except for the batch size and learning rate. The batch size is set to 256, and to prevent weight explosion, we adjust the learning rate of the compressed model to 1/10 or 1/100 of the original model. + +To evaluate the reduced latency using SNP, we have configured four testing scenarios: one on a CPU and another on GPU for both edge devices and server processors. We employ a standard PyTorch model for profiling on the server processors (Intel Xeon Silver 4210R and NVIDIA GeForce RTX 3090). Profiling on the Raspberry Pi 4B and Jetson Nano is conducted using the ONNX and TensorRT formats, respectively. All latencies are measured using a single image as an input, except for the GPU of the server processor (RTX 3090), where it is set to 64 images. + +# 4.2 Quantitative results + +Comparison with other methods Despite the constraints outlined in Sec. 3.4, SNP not only maintains accuracy comparable to existing methods but also significantly reduces inference time across diverse hardware and data types on the ImageNet-1K dataset, as shown in Tab. 1 and Tab. 2. + +Table 2: Inference speed and Top-1 accuracy of the compressed model across different devices. Performance evaluation involves accuracy on ImageNet-1K and inference time for the compressed DeiTs and EfficientFormer-L1. Latency is benchmarked with 200 warm-up runs and averaged over 1000 runs. In latency measurement, a single image is used as an input, except for the RTX 3090, where 64 images are employed in a single batch. + +
ModelTop-1 (%)GFLOPsEdge devices (ms)Server processors (ms)
Raspberry Pi 4B (.onnx)Jetson Nano (.trt)Xeon Silver 4210R (.pt)RTX 3090 (.pt)
DeiT-Tiny72.201.3139.1341.0334.7418.65
+ SNP (Ours)70.290.681.63 (1.70×)26.67 (1.54×)25.25 (1.38×)17.82 (1.05×)
DeiT-Small79.804.6401.2799.3253.3746.13
+ SNP (Ours)78.522.0199.15 (2.01×)45.51 (2.18×)38.57 (1.38×)32.91 (1.40×)
+ SNP (Ours)73.321.3136.68 (2.94×)32.03 (3.10×)33.46 (1.60×)26.98 (1.71×)
DeiT-Base81.8017.61377.71293.29122.03151.35
+ SNP (Ours)79.636.4565.68 (2.44×)132.55 (2.21×)64.65 (1.89×)72.96 (2.07×)
EfficientFormer-L179.201.3169.1330.9543.7526.19
+ SNP (Ours)75.530.695.12 (1.78×)19.78 (1.56×)38.25 (1.14×)17.24 (1.52×)
+ SNP (Ours)74.510.582.60 (2.05×)17.76 (1.74×)35.15 (1.24×)16.01 (1.64×)
+ +In a recent study [26], unconstrained masking generally outperforms posttraining accuracy of pruned models by an average of $2.1\%$ on ImageNet-1K. Compared to unconstrained head masking approaches like SSVITE [4] and XPruner [30], SNP achieves significantly higher compression rates for all DeiTs FLOPs (30.64% and 10.64%) with minimal performance degradation (0.83% and 0.87%), much less than the $2.1\%$ average mentioned. + +Furthermore, compared to other pruning approaches like WDPruning [29] and UVC [31], which use a combination of pruning techniques to compress DeiTs, SNP achieves comparable accuracy solely through neuron-level pruning. Notably, DeiT-Small with SNP outperforms WDPruning by $0.14\%$ , with the removal of 3.3 million parameters and a reduction of 0.6 GFLOPs. Compared to UVC, SNP exhibits negligible performance degradation, averaging $0.51\%$ across all DeiTs while using $7.65\%$ fewer FLOPs. + +Large compressed vs. Small hand-crafted DeiT-Small with SNP, a large pruned model, outperforms the smaller, hand-crafted DeiT-Tiny in both accuracy and latency, achieving a notable $1.12\%$ improvement in top-1 accuracy while maintaining similar FLOPs. Additionally, DeiT-Small with SNP exhibits enhanced speed compared to the original DeiT-Tiny across edge devices and CPU-based server processors. Notably, its speed increases up to $21.94\%$ compared to the original DeiT-Tiny running on Jetson Nano, a GPU-based edge device. This substantial performance gap underscores the superiority of the compressed model (DeiT-Small with SNP) over the smaller hand-crafted designed model (DeiT-Tiny) in both overall performance and speed. + +Accelerating Transformer-based models As depicted in Tab. 2, SNP achieves impressive acceleration of DeiTs by a factor of $1.44 \times$ to $2.44 \times$ on edge devices and $1.05 \times$ to $2.07 \times$ on server processors. This acceleration is notable, + +![](images/6d63344579af9058c87aa70dad1df97d56f7b38419d208c37bd1ecb3cc71554f.jpg) +Fig.3: Attention maps with varying pruning criteria and compression ratios. All query and key layers are locally pruned based on the specified pruning ratio without fine-tuning. The importance scores of $l2$ -norm and GM on query and key layers are combined by filter index and pruned simultaneously. "Reverse" represents the reverse order of SNP. + +with negligible average performance degradation of $1.79\%$ , specifically $1.91\%$ , $1.28\%$ , and $2.17\%$ for DeiT-Tiny, DeiT-Small, and DeiT-Base, respectively. + +Compared to WDPruning, which employs both head and block pruning, SNP surpasses in terms of latency for DeiT-Small and DeiT-Base on the RTX 3090 GPU. SNP accelerates the original DeiTs by $1.38 \times$ and $2.07 \times$ , respectively, while WDPruning achieves a comparatively modest acceleration of $1.18 \times$ for both models. + +The superiority of SNP becomes more evident when the target of neuron-level pruning, linear or convolutional layers, contribute a larger proportion of the original model's computation time. we find that SNP further accelerates original model $1.05 \times$ to $2.07 \times$ as the DeiT model's size grows on the RTX 3090 GPU. In contrast, WDPruning, involving the removal of entire layers and associated operators in both head and block, consistently shows acceleration rates of $1.18 \times$ . + +To ascertain the robustness of SNP across diverse Transformer models, especially on the efficiently designed model, we conduct additional experiments on EfficientFormer-L1. SNP accelerates the model by $1.78 \times$ and $2.05 \times$ faster on Raspberry Pi 4B and $1.56 \times$ and $1.74 \times$ faster on Jetson Nano. Notably, the compressed EfficientFormer-L1 achieves acceptable accuracy of $75.53\%$ and $74.51\%$ . + +![](images/fe1f4d12f872427ee2e4c2882de994577a236ac3db60578a73e73447f1e97c39.jpg) +Original + +![](images/6323a610de3b6a87de2951e2534265b0f3ea1293eae2b48888c961379a775071.jpg) +Compressed + +![](images/f495f899f667f74c1ebea9d84560b323a1faed3a302230ccb3ef7f71e9b587c1.jpg) +Compressed + Fine-tuned + +![](images/f30d324ea6e826972f11170d584acf8e3fe343abc0f6c1e17e2631f0f4df2e9e.jpg) +Original + +![](images/c088f449756ed2a658ee4db4eea0f3dfaa2917b693a573088d8ade161617983d.jpg) +Compressed + +![](images/0ea0c7fbda0f1faafa893bf05469fc6f2f2992d7809c48d50a19a03231c65570.jpg) +Compressed + Fine-tuned + +![](images/a1e97d41a8cea68255e7a9b05107ab15f937d340b1c469d8eeff847609e8b360.jpg) + +![](images/e4cf7f34b9d24d2d89564d5c2c695ea84ad328de030288edc83aaa2d40aaafe9.jpg) + +![](images/92ca6126fbf405bbe02594c29cd89eafeb1345f69377f89fd39db6cacdac27df.jpg) + +![](images/043764bb3f9586de9783dd19fd78bfe6789c008dc0dfe86a14e182e87f85b1e0.jpg) + +![](images/3f28d2f837980e4e95e62d2210f97120d05337bfb2cef20679ab14e68cee5230.jpg) + +![](images/1e147a49ae4d164c1753880b556a9c73d16b28742f70514b921c7a109872d675.jpg) + +![](images/e46717d66c317f1268b3e24b641d9d8f10dbae092dba3851ec623446d258ad24.jpg) + +![](images/02649542649345d8d47fac7611e4659a93655d187db203c620b577879fb96272.jpg) + +![](images/b9fa9dee01639061707c36607e68657d83d0d4250357ab057048d9a50f5fd60f.jpg) + +![](images/80c3f9e675e8736b403f13dec05ab4f9f08cbffc7d0abf6bbcd7fffe6aedd9c5.jpg) +Fig. 4: Attention maps from the original, compressed, and fine-tuned DeiT-Tiny with SNP. The attention maps in the first row are visualized using the attention rollout [1]. Each red box contains three attention maps from each head of the MSA module, ordered accordingly. + +![](images/a6a84efa7ff27e3d0f5799df9789e0bf10cce09c7e4dc607d8571f21e8d31059.jpg) + +![](images/8733c633fd40e21ad93c447045e6e70d37d8c82e6a5128398182a49c7a071c7b.jpg) + +![](images/c4657e25aa77e07d79fd2c1d8f3225347964d110ab3ee705b21897aa451ae4cb.jpg) + +![](images/733fb88eff8af16a930947a91c86c970e83bcdab4ab20295a91ff3a61cfe05e3.jpg) + +![](images/adbf66146248093a28a285b7928c9f3c461198191b1a51b8de3b2581ec023cd1.jpg) + +![](images/0dc1699ad20f9728167c256362b43505ffff8a63f3117f2981d60b51a9aa0c7b.jpg) + +![](images/5b3d85906cd7f77e264948e19e58181b3d6928c4b8c55f36eb727dd53b8c9134.jpg) + +![](images/05e40153d154fde971d94c85587107f7e94c4c1d400362961dfd9b96af661771.jpg) + +![](images/2c163d34778577e23b21f38d9f54d65670f591c1a9602c9bd095bbc91b24cc7b.jpg) + +![](images/122ea387d2a42ed8cd836a60e795faa22985d68c431685d36c6e0b2d78f2ff2c.jpg) + +![](images/216d524b367e0b7159db2a0df90b7796c13acf549b5b02adcd0a2a9fffab04df.jpg) + +![](images/48f57ef9d9e276d804a646c3424e4fb611d9016e3d3305ed1e68662b1a22652e.jpg) + +![](images/a0071d2447b2285f1beeac11a0a9d1049a6a772332819a5eb5215886eba57ce6.jpg) + +![](images/7f97ff62864dc438be9a9b156403674afc785197bb8015297585adc5dd2902b5.jpg) + +![](images/7bfe8e9879866ad1a6c551c96df7557d9964c16717a30e1f6983899afca7a108.jpg) + +![](images/2d45862860754f7be5655ecc809419140763d0d2b1a60f89a7e73eb04cdfdefb.jpg) + +![](images/318dd579dc52f9b742f1d86ef9aa91021a9c238e975c5a4c99728a01a8de4ec3.jpg) + +![](images/8600cc0e3899fc7aac7bcffef47550512e8fb77623099f2a1b515c27003a720f.jpg) + +![](images/f98793e63d20f9a9d33927b64ba63087ce6af79b800bb39f5bcea8668e8665b6.jpg) + +![](images/f556b8f22da9495a87ed7c26bf33cce21f6866d0dd03cf7c29419878a24b0548.jpg) + +![](images/dc00dd4deef70eecd465a32d4d5c909f7920a45768ad490b3908478f32e5164c.jpg) + +![](images/1265d570ac5366ba1c740910c160458a20d26cdcdc9f1ab78331d5268ee51c4b.jpg) + +![](images/9c41c3b2f1c42b754af61aa24e3c3daed53a7152b46f9e316471384d9c83d11f.jpg) + +![](images/ab34a27b16544d65dd65588fc3234fc124a3d24e49c664e3638c186460afcb7f.jpg) + +![](images/68e1a02c5e9e3957959c1d53b488fed6a86730c7a59cc678ed5e984096a344d0.jpg) + +![](images/f0c77594a2a0657a100a720728263c58e613ad73e9648f8710c6549c9a699d37.jpg) + +![](images/618191c91df49a5db67c4f792861f9e4e662552b65fb950866d2bfba05a88c4a.jpg) + +![](images/d506e8f8ab137f12aa0747c2b7be22465379f366ad7775f4b45daf5d8ba4432b.jpg) + +![](images/b94ba33d9bf2a0a0aac9d21e4a294cbe26fa5f485f5219eaaaf53e204c0b5c45.jpg) + +![](images/0b3d966682b866cb80a0e9ec3586d4e95bf0ae3fe281a6b3623bfbcc89614577.jpg) + +![](images/b69fc8c7e5b8abad7c52d50cf70066816270dec0650c575792c421a2d00390bd.jpg) + +![](images/354be57d221d5f468ac85f1c2acc6aaea3aabf4269d236655696504b97443df1.jpg) + +![](images/2a5938bd8b133b7d82cfefa4972bc6a980461af895027a54c53243ba0a70de4a.jpg) + +![](images/8c5882ec658ac9467160bd0710b675aae6e37c14569f6882d1d9d6623f4d3f8a.jpg) + +![](images/4ece03b7e0f278b79f3ad24fa430ee3ac8862134854f0aa48d8807b7b6ed1e04.jpg) + +![](images/fecd71612f8074ebef73698044fc46a9af96362fa4ef774437bfe46e6674c032.jpg) + +![](images/0d2163883253187f7708895be4ae99120d136bb583e820d409e6acf67c8a9644.jpg) + +![](images/e056c869856b4f7047484e8018fc91945cd8339e04976e05e83a980f284f77e6.jpg) + +![](images/4c8f13fae9e6650345f57b4d567043ad8c8a10acf3f9a2143b79d749ea85bea1.jpg) + +![](images/1d25f4d1aa674a20408152d2157006593e6d85974ac47f722744149394d933d5.jpg) + +![](images/76e771d3611843403d36d3da83fde4ccc37213ed51e8d5f5fb21c5db355b7fb1.jpg) + +![](images/35f8c87b894ea7e4d1368f6b7ab46468ace45a6efdb1b0e9b56d1f4d47fad4e2.jpg) + +![](images/31c426fa47f0b63bb46e402e6bc6ffaf686ef738eb8439cae1cc7d125d481535.jpg) + +![](images/2a2c918e52fa5710e109cb0c128c6c7c22ba9e2c384ef01e318e9a0e1f7e56b7.jpg) + +![](images/610a3ad04eb3118d53ef7ea46059c44300eda5c6f6032cccbc1cd28dae25ba9f.jpg) + +![](images/373e4a86d8c2a753fc16a0f8587d8bd1bed4d021efdf9f93a639b8e8b60dd773.jpg) + +![](images/966bae9ff81f39d2fc182054cb32e20f63373e819cd950b539564ced0ad021d2.jpg) + +![](images/f4add919c6a72372542223bba125bcbddca915eb557c7212e4945484844a5cdd.jpg) + +![](images/607df98602bf3860ee5e4041956aacc419c68b5c6e61b2e8d43be2e564daf380.jpg) + +![](images/13276cb8d9e4f91f0dc8e713b4e7b890603f1090284dace91b825fadd2dc7cfb.jpg) + +![](images/979d3abae462302381b2126b7ae161cfb56fe82f11cc8f00a01130ba70571e48.jpg) + +![](images/0c5666a18bfdf3429739304cb15cc376a03fa0b521b4d10cf11e1ede8ff1bbd5.jpg) + +![](images/dcda949fcbe86e57760bf7fb2ca869b4d853ecfc47a5640786cb24d1e218b9e9.jpg) + +![](images/d185cbdb87880d909ac1d5b4e423ae5683659da251038a3b582146bc7d7d9ada.jpg) + +![](images/c3b74fbd38c0eced21ed711a499a711d3e5933c22bcd5e58a6b1b9ac5f364c2d.jpg) + +![](images/7c9a5c1ba7551fe44e8eb64d1ff9ff564ec584b523b1a31d16e38f53f8f51dc7.jpg) + +![](images/42af389f2d089c9cd68bbf2a8671f19c1d6dc20930d8b5690707e89317a99305.jpg) + +![](images/ccaf4e741482c658eac0d87b6b322f7e7c40a1e41743d334bfd1e13011e1ffa3.jpg) + +![](images/79e1fdd3460bfb2ffb97ca83563353083e6491f399ce931481c57f34ffaec664.jpg) + +![](images/c58674a657e7f9bc5ba2c64f813bf9ea653692c8c2f5d8e2587e5c8a2b988e7a.jpg) + +![](images/1e8e03716568d8677f5f7ed812813b2bc65ee25d3aa3f07f2737d5ae99b65b6d.jpg) + +![](images/23a32049cdd4e27ed7b0e3dce9156011db1140a1bf608b08e547c6895b2c97ec.jpg) + +![](images/0c1d342ed7ee47d058557fd7271830d8aacb9f2d342f6dc24a6cc6da9d49db2c.jpg) + +![](images/c75fa64d8f8053ecb7141f850a5bfa7815e2971e30e7f09e9b9881f3840922f0.jpg) + +# 4.3 Qualitative results + +In the subsequent sections, attention rollout [1] is employed to randomly selected images from the test datasets to visualize the attention maps of the DeiTs and assess the efficacy of SNP. This evaluation unfolds across two key facets: + +- Preserving the attention scores: Visualize attention maps to verify the effectiveness of the proposed importance score in preserving attention scores. +- Restoring the attention scores after SNP: Visualize attention maps to find the effectiveness of SNP in restoring attention scores. + +Preserving the attention scores To visualize SNP effectiveness, we apply four pruning criteria to compress the query and key filter pairs in all MSA modules of DeiT-Tiny: SNP, $l2$ -norm, geometric median (GM) [11], and "Reverse". + +The "Reverse" criterion prioritizes pruning the most important filter pairs first, keeping the least important pairs until the end. However, both $l2$ -norm and GM pruning criteria ignore the MSA module's graphical connectivity, evaluating importance scores independently for query and key layers. To handle identical filter indices during pruning, we aggregate scores based on filter indices, removing the least important indices from both layers. + +In Fig. 3, attention maps for the original DeiT-Tiny and locally pruned models are presented across various pruning ratios (10% to 90% with 10% intervals). Our proposed method effectively maintains the original attention map even after pruning over 80% of the filters, whereas other methods show fragmented attention maps at much lower pruning ratios (typically 30% or less). These results highlight the potential of our neuron-level pruning criteria, utilizing SVD to preserve attention scores, in reducing the size and speeding up the execution of MSA modules without compromising accuracy. + +Table 3: Performance of SNP without fine-tuning across different data quantities at various pruning ratios. To determine the proper number of images for calculating the importance score, we conduct local pruning on the query and key layers at pruning ratios of $10\%$ , $30\%$ , $50\%$ , $70\%$ , and $90\%$ . All performance metrics are assessed without fine-tuning. The latency is measured on Raspberry Pi 4B. + +
Number of imagesPerformance by pruning ratio
Original10%30%50%70%90%
172.2071.1567.9657.6038.6811.17
472.2070.8267.8657.6538.0511.86
1672.2070.7267.4258.6039.469.11
6472.2070.8367.5559.5042.1312.83
25672.2070.7568.1459.5842.7014.48
Latency (ms)139.13133.60129.29119.41117.48112.32
Params (M)5.75.595.415.235.064.88
+ +Restoring the attention scores after SNP Fig. 4 illustrates the overall and per-head attention maps of the original, compressed, and fine-tuned DeiT-Tiny, respectively. The first row shows the overall attention maps of the respective models. Notably, the compressed model maintains a well-preserved overall attention map, despite pruning all layers, including values and FFN, resulting in a $53\%$ reduction in FLOPs and a $46\%$ reduction in parameters. Especially, we can observe that attention map is well-restored after the fine-tuning process. + +The twelve red boxes below the overall attention map depict per-head attention maps for twelve layers of each original, compressed, and fine-tuned models respectively. As shown in Fig. 4, it is evident that the attention maps for each head are effectively preserved and restored in each of the compressed and fine-tuned models. + +# 4.4 Ablation studies + +Importance scores by the data quantity Since the attention scores depend on the input, as indicated in Eq. (1), the proposed importance scores for query and key filter pairs (Eq. (6)) may exhibit sensitivity to the distribution of the input image $X$ . To validate the method's robustness, we compute importance scores using various image quantities, pruning query and key layers at different ratios, without fine-tuning process. + +As depicted in Tab. 3, SNP demonstrates a slight advantage in preserving performance with an increasing number of images, outperforming models compressed with fewer images as the compression ratio rises. However, this improvement comes at the cost of increased computation time for SNP calculations. Considering these factors into account, we opt to use 64 images, which yield the second-best performance among the given number of images. This decision strikes a balance between achieving satisfactory performance and maintaining computational efficiency. + +![](images/d1bbf3c5b9cc371fd90ad09fdca666753623755d4200fee638014e2ffaa89677.jpg) +Fig. 5: Top-1 accuracy of compressed DeiT-Tiny on ImageNet using several pruning criteria without fine-tuning. Query and key layers are locally pruned using various pruning criteria : SNP, GM, l2-norm, reverse order of SNP ("Reverse"), and original DeiT-Tiny ("Baseline"). + +Performance comparison across pruning ratios To assess the robustness of SNP, we examine the performance of DeiT-Tiny under various pruning criteria and different pruning ratios, as described in Sec. 4.3. + +As illustrated in Fig. 5, SNP consistently outperforms other pruning criteria across all pruning ratios. In contrast, models compressed with "Reverse" criteria exhibit the lowest performance at all pruning ratios, underscoring the robustness of the proposed approach. + +Both Fig. 3 and Fig. 5 confirm that SNP successfully preserves attention scores in both quantitative and qualitative aspects. Even with an $80\%$ pruning ratio and without fine-tuning, SNP maintains original attention scores and outperforms other pruning criteria. + +# 5 Conclusion + +In this paper, we propose a novel graph-aware neuron-level pruning method, SNP, designed to compress and accelerate Transformer-based models. SNP proposes two pruning criteria for preserving attention scores and eliminating interhead redundancy. Using SNP, a large compressed model outperforms small, hand-crafted designed models in both performance and latency on edge devices. Moreover, the compressed models exhibit astonishing results in latency on various devices, with negligible performance degradation. + +As this work is a first attempt to accelerate MSA modules using neuron-level pruning alone, many challenges remain. One is incorporating other pruning + +methods such as head or block pruning for a more efficient Transformer model. Another challenge is applying SNP to other vision tasks, including image generation, which requires high computational costs on both training and inference. + +We believe that these works encourage the adoption of model pruning as a tool, for both improving the applicability of ViTs in resource-constrained environments and reducing the training costs of large models by integrating with training process. + +# Acknowledgements + +This work was supported by Artificial intelligence industrial convergence cluster development project funded by the Ministry of Science and ICT(MSIT, Korea)&Gwangju Metropolitan City. + +# References + +1. Abnar, S., Zuidema, W.: Quantifying attention flow in transformers. arXiv preprint arXiv:2005.00928 (2020) +2. Bian, Y., Huang, J., Cai, X., Yuan, J., Church, K.: On attention redundancy: A comprehensive study. In: Proceedings of the 2021 conference of the north american chapter of the association for computational linguistics: human language technologies. pp. 930-945 (2021) +3. Bolya, D., Fu, C.Y., Dai, X., Zhang, P., Feichtenhofer, C., Hoffman, J.: Token merging: Your vit but faster. arXiv preprint arXiv:2210.09461 (2022) +4. Chen, T., Cheng, Y., Gan, Z., Yuan, L., Zhang, L., Wang, Z.: Chasing sparsity in vision transformers: An end-to-end exploration. Advances in Neural Information Processing Systems 34, 19974-19988 (2021) +5. Demouth, J.: Sparse matrix-matrix multiplication on thegpu. Tech. rep., NVIDIA (2012) +6. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009) +7. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) +8. Fang, G., Ma, X., Song, M., Mi, M.B., Wang, X.: Depgraph: Towards any structural pruning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16091-16101 (2023) +9. Han, S., Mao, H., Dally, W.J.: Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149 (2015) +10. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) +11. He, Y., Liu, P., Wang, Z., Hu, Z., Yang, Y.: Filter pruning via geometric median for deep convolutional neural networks acceleration. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4340-4349 (2019) + +12. He, Y., Zhang, X., Sun, J.: Channel pruning for accelerating very deep neural networks. In: Proceedings of the IEEE international conference on computer vision. pp. 1389-1397 (2017) +13. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015) +14. Lee, J., Park, S., Mo, S., Ahn, S., Shin, J.: Layer-adaptive sparsity for the magnitude-based pruning. arXiv preprint arXiv:2010.07611 (2020) +15. Li, H., Kadav, A., Durdanovic, I., Samet, H., Graf, H.P.: Pruning filters for efficient convnets. arXiv preprint arXiv:1608.08710 (2016) +16. Li, Y., Yuan, G., Wen, Y., Hu, J., Evangelidis, G., Tulyakov, S., Wang, Y., Ren, J.: Efficientformer: Vision transformers at mobilenet speed. Advances in Neural Information Processing Systems 35, 12934-12949 (2022) +17. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021) +18. Liu, Z., Wang, Y., Han, K., Zhang, W., Ma, S., Gao, W.: Post-training quantization for vision transformer. Advances in Neural Information Processing Systems 34, 28092-28103 (2021) +19. Michel, P., Levy, O., Neubig, G.: Are sixteen heads really better than one? Advances in neural information processing systems 32 (2019) +20. Mishra, A., Latorre, J.A., Pool, J., Stosic, D., Stosic, D., Venkatesh, G., Yu, C., Micikevicius, P.: Accelerating sparse deep neural networks. arXiv preprint arXiv:2104.08378 (2021) +21. Nakajima, S., Sugiyama, M., Babacan, S.D., Tomioka, R.: Global analytic solution of fully-observed variational bayesian matrix factorization. The Journal of Machine Learning Research 14(1), 1-37 (2013) +22. Radosavovic, I., Kosaraju, R.P., Girshick, R., He, K., Dólár, P.: Designing network design spaces. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10428-10436 (2020) +23. Tan, M., Le, Q.: Efficientnet: Rethinking model scaling for convolutional neural networks. In: International conference on machine learning. pp. 6105-6114. PMLR (2019) +24. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jégou, H.: Training data-efficient image transformers & distillation through attention. In: International conference on machine learning. pp. 10347-10357. PMLR (2021) +25. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017) +26. Wan, A., Hao, H., Patnaik, K., Xu, Y., Hadad, O., Güera, D., Ren, Z., Shan, Q.: Upscale: unconstrained channel pruning. In: International Conference on Machine Learning. pp. 35384-35412. PMLR (2023) +27. Wang, Z.: Sparsednn: Fast sparse deep learning inference on cpus. arXiv preprint arXiv:2101.07948 (2021) +28. Xie, Q., Luong, M.T., Hovy, E., Le, Q.V.: Self-training with noisy student improves imagenet classification. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10687-10698 (2020) +29. Yu, F., Huang, K., Wang, M., Cheng, Y., Chu, W., Cui, L.: Width & depth pruning for vision transformers. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 3143-3151 (2022) + +30. Yu, L., Xiang, W.: X-pruner: explainable pruning for vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24355-24363 (2023) +31. Yu, S., Chen, T., Shen, J., Yuan, H., Tan, J., Yang, S., Liu, J., Wang, Z.: Unified visual transformer compression. arXiv preprint arXiv:2203.08243 (2022) +32. Yu, S., Mazaheri, A., Jannesari, A.: Topology-aware network pruning using multi-stage graph embedding and reinforcement learning. In: International conference on machine learning. pp. 25656-25667. PMLR (2022) +33. Yu, X., Serra, T., Ramalingam, S., Zhe, S.: The combinatorial brain surgeon: pruning weights that cancel one another in neural networks. In: International Conference on Machine Learning. pp. 25668-25683. PMLR (2022) \ No newline at end of file diff --git a/2024/SNP_ Structured Neuron-level Pruning to Preserve Attention Scores/images.zip b/2024/SNP_ Structured Neuron-level Pruning to Preserve Attention Scores/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c0f14638dc8b9e9ca13150ea2e349950fad15853 --- /dev/null +++ b/2024/SNP_ Structured Neuron-level Pruning to Preserve Attention Scores/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:368493d3e5e867efe814f59a2f3e3cd1bcb24094af8ab84f19085ef2b6e2023b +size 566270 diff --git a/2024/SNP_ Structured Neuron-level Pruning to Preserve Attention Scores/layout.json b/2024/SNP_ Structured Neuron-level Pruning to Preserve Attention Scores/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d184e1fbab7816fc456421b195dde42f174974a9 --- /dev/null +++ b/2024/SNP_ Structured Neuron-level Pruning to Preserve Attention Scores/layout.json @@ -0,0 +1,11880 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 161, + 111, + 455, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 111, + 455, + 148 + ], + "spans": [ + { + "bbox": [ + 161, + 111, + 455, + 148 + ], + "type": "text", + "content": "SNP: Structured Neuron-level Pruning to Preserve Attention Scores" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 168, + 167, + 447, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 167, + 447, + 181 + ], + "spans": [ + { + "bbox": [ + 168, + 167, + 447, + 181 + ], + "type": "text", + "content": "Kyunghwan Shim1, Jaewoong Yun1, and Shinkook Choi1*" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 197, + 189, + 416, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 189, + 416, + 213 + ], + "spans": [ + { + "bbox": [ + 197, + 189, + 416, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 197, + 189, + 416, + 213 + ], + "type": "text", + "content": " Nota Inc., Korea " + }, + { + "bbox": [ + 197, + 189, + 416, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 197, + 189, + 416, + 213 + ], + "type": "text", + "content": " {kyunghwan.shim, jwyun, shinkook. Choi}@nota.ai" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 240, + 455, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 240, + 455, + 461 + ], + "spans": [ + { + "bbox": [ + 160, + 240, + 455, + 461 + ], + "type": "text", + "content": "Abstract. Multi-head self-attention (MSA) is a key component of Vision Transformers (ViTs), which have achieved great success in various vision tasks. However, their high computational cost and memory footprint hinder their deployment on resource-constrained devices. Conventional pruning approaches can only compress and accelerate the MSA module using head pruning, although the head is not an atomic unit. To address this issue, we propose a novel graph-aware neuron-level pruning method, Structured Neuron-level Pruning (SNP). SNP prunes neurons with less informative attention scores and eliminates redundancy among heads. Specifically, it prunes graphically connected query and key layers having the least informative attention scores while preserving the overall attention scores. Value layers, which can be pruned independently, are pruned to eliminate inter-head redundancy. Our proposed method effectively compresses and accelerates Transformer-based models for both edge devices and server processors. For instance, the DeiT-Small with SNP runs 3.1 times faster than the original model and achieves performance that is " + }, + { + "bbox": [ + 160, + 240, + 455, + 461 + ], + "type": "inline_equation", + "content": "21.94\\%" + }, + { + "bbox": [ + 160, + 240, + 455, + 461 + ], + "type": "text", + "content": " faster and " + }, + { + "bbox": [ + 160, + 240, + 455, + 461 + ], + "type": "inline_equation", + "content": "1.12\\%" + }, + { + "bbox": [ + 160, + 240, + 455, + 461 + ], + "type": "text", + "content": " higher than the DeiT-Tiny. Additionally, SNP accelerates the efficiently designed Transformer model, EfficientFormer, by 1.74 times on the Jetson Nano with acceptable performance degradation. Source code is at https://github.com/Nota-NetsPresso/SNP" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 471, + 453, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 471, + 453, + 483 + ], + "spans": [ + { + "bbox": [ + 160, + 471, + 453, + 483 + ], + "type": "text", + "content": "Keywords: Network pruning " + }, + { + "bbox": [ + 160, + 471, + 453, + 483 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 471, + 453, + 483 + ], + "type": "text", + "content": " Network compression " + }, + { + "bbox": [ + 160, + 471, + 453, + 483 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 471, + 453, + 483 + ], + "type": "text", + "content": " Compact models" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 502, + 230, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 230, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 230, + 514 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 526, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 526, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 526, + 482, + 647 + ], + "type": "text", + "content": "Vision Transformers (ViTs) [7,17,24] have outperformed or matched the performance of state-of-the-art convolutional neural networks (CNNs) [10,22,23,28] on various computer vision tasks. The success of ViTs is attributed to the Multihead Self-Attention (MSA) module, which captures intricate relationships in data. However, the Transformer architecture entails substantial computational resources, posing challenges for practical applications on edge devices with constrained storage and computational capabilities. To address this issue, we leverage the graphical components of the MSA module to reduce the dimension of interconnected layers, aimed at effectively reducing their computing budgets while achieving hardware-agnostic speedups." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 653, + 234, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 653, + 234, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 653, + 234, + 665 + ], + "type": "text", + "content": "* Corresponding author" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 184, + 118, + 427, + 318 + ], + "blocks": [ + { + "bbox": [ + 184, + 118, + 427, + 318 + ], + "lines": [ + { + "bbox": [ + 184, + 118, + 427, + 318 + ], + "spans": [ + { + "bbox": [ + 184, + 118, + 427, + 318 + ], + "type": "image", + "image_path": "c6172df040d4b81d7cc5f6a8ea3566eddbe83324bf39d6bcdb5a235143dc073f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 327, + 482, + 383 + ], + "lines": [ + { + "bbox": [ + 130, + 327, + 482, + 383 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 482, + 383 + ], + "type": "text", + "content": "Fig. 1: Comparison of model size, speed, and performance. ImageNet-1K classification results. Latency is profiled by Raspberry Pi 4B. The connected lines represent the compressed models paired with the original model. The size of each circle indicates the number of parameters in respective model. The number adjacent to each compressed model indicates its compressed FLOPs." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 410, + 482, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 410, + 482, + 541 + ], + "spans": [ + { + "bbox": [ + 130, + 410, + 482, + 541 + ], + "type": "text", + "content": "From the perspective of structured pruning, which aims to reduce the number of dimensions in convolutional or linear layers, the MSA module contains two prunable objectives: heads and neurons. Head pruning, which removes the number of heads, is relatively intuitive to implement due to the reduced complexity of graphical elements compared to neuron-level pruning. In contrast, neuron-level pruning reduces the dimension of individual layers in each head of the MSA module, necessitating a comprehensive understanding of the graphical connectivity of the MSA module. A recent work [31] implements neuron-level pruning by zeroing out individual filters without considering the model's graphical connectivity. This indiscriminate zeroing can negatively affect both accuracy and throughput performance [26]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 543, + 482, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 543, + 482, + 615 + ], + "spans": [ + { + "bbox": [ + 130, + 543, + 482, + 615 + ], + "type": "text", + "content": "In this paper, we introduce a novel graph-aware neuron-level pruning method called Structured Neuron-level Pruning (SNP) to accelerate and compress ViTs effectively. We propose two pruning criteria based on the function of each layer within the MSA module. SNP prunes filter pairs of query and key layers containing fewer contributions to attention scores. Moreover, SNP aims to reduce redundancy across heads by eliminating the redundant filter from the value layer." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 616, + 482, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 616, + 482, + 652 + ], + "spans": [ + { + "bbox": [ + 130, + 616, + 482, + 652 + ], + "type": "text", + "content": "Furthermore, by removing identical filter indices for all the graphically connected layers, SNP could accelerate various Transformer models on various devices without additional libraries, as shown in Fig. 1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 146, + 653, + 391, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 653, + 391, + 666 + ], + "spans": [ + { + "bbox": [ + 146, + 653, + 391, + 666 + ], + "type": "text", + "content": "In summary, the major contributions of this paper are:" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "text", + "content": "Shim. et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 137, + 116, + 480, + 245 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 137, + 116, + 479, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 116, + 479, + 151 + ], + "spans": [ + { + "bbox": [ + 137, + 116, + 479, + 151 + ], + "type": "text", + "content": "- We propose a novel graph-aware neuron-level pruning method, SNP, for Transformer models. SNP is the first method to use the graphical characteristics of the MSA module to measure the importance score of neurons." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 152, + 480, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 152, + 480, + 175 + ], + "spans": [ + { + "bbox": [ + 138, + 152, + 480, + 175 + ], + "type": "text", + "content": "- To the best of our knowledge, this is the first work to accelerate Transformer models on any device using neuron-level pruning only." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 175, + 480, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 175, + 480, + 245 + ], + "spans": [ + { + "bbox": [ + 138, + 175, + 480, + 245 + ], + "type": "text", + "content": "- SNP achieves significant acceleration while maintaining the original performance on several models. Compressed DeiT-Small outperforms DeiT-Tiny by " + }, + { + "bbox": [ + 138, + 175, + 480, + 245 + ], + "type": "inline_equation", + "content": "1.12\\%" + }, + { + "bbox": [ + 138, + 175, + 480, + 245 + ], + "type": "text", + "content": " in accuracy, with similar FLOPs, and reduces inference time on various edge devices. Furthermore, the proposed method accelerates the efficiently designed Transformer model, EfficientFormer [16], more than two times with acceptable performance degradation." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 261, + 237, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 261, + 237, + 274 + ], + "spans": [ + { + "bbox": [ + 132, + 261, + 237, + 274 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 283, + 328, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 283, + 328, + 295 + ], + "spans": [ + { + "bbox": [ + 132, + 283, + 328, + 295 + ], + "type": "text", + "content": "2.1 Compressing vision Transformers" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 300, + 480, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 480, + 431 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 480, + 431 + ], + "type": "text", + "content": "The ViTs [7, 17, 24] achieve high performance in numerous vision tasks without specialized image processing modules such as convolutions. The key concept of ViTs is to segment images into patch sequences, convert these patches to token embeddings, and then process them through Transformer encoders [25]. ViTs consist only of Transformer blocks, making them likely to be over-parameterized. For this reason, recent works have aimed to reduce computational cost [3, 24] and be memory efficient [18, 26]. DeiT [24] proposes lightweight ViT architectures through knowledge distillation [13]. ToMe [3] proposes accelerating ViTs by directly combining similar tokens, without the need for training. Liu et al. [18] propose post-training quantization method using nuclear norm based mixed-precision scheme that does not require fine-tuning for the Vision Transformer." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 445, + 305, + 457 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 445, + 305, + 457 + ], + "spans": [ + { + "bbox": [ + 132, + 445, + 305, + 457 + ], + "type": "text", + "content": "2.2 Pruning vision Transformers" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 462, + 480, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 480, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 480, + 545 + ], + "type": "text", + "content": "Unstructured and structured pruning Pruning methods can be broadly categorized into two types, unstructured and structured pruning. Unstructured pruning sets individual weights or parameters to zero, resulting in irregular sparse matrices [9, 14]. Compressed models using unstructured pruning tend to maintain relatively high performance for a given pruning ratio. However, they necessitate additional libraries, such as cuSPARSE [5], Automatic SParsity [20], or SparseDNN [27] to accelerate sparse matrix computations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 546, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 481, + 665 + ], + "type": "text", + "content": "Structured pruning, on the other hand, involves the removal of entire groups of units, such as filters or attention heads. This can be implemented using \"masking\" (zeroing out) [11, 12, 32, 33], or by \"pruning\" [8, 15]. Structured pruning by masking [11, 12, 32, 33] simply sets the group of units to zero, which requires additional libraries to accelerate the model, as unstructured pruning. \"Pruning\" [8, 15], on the other hand, requires a comprehensive understanding of the network's graphical connectivity, including element-wise operations that enforce the same input shape. By considering the graphical connectivity and pruning identical filter indices for inter-connected layers, structured pruning can achieve acceleration on any devices." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 170, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 170, + 91, + 448, + 102 + ], + "type": "text", + "content": "SNP: Structured Neuron-level Pruning to Preserve Attention Scores" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 115, + 480, + 287 + ], + "blocks": [ + { + "bbox": [ + 134, + 115, + 480, + 287 + ], + "lines": [ + { + "bbox": [ + 134, + 115, + 480, + 287 + ], + "spans": [ + { + "bbox": [ + 134, + 115, + 480, + 287 + ], + "type": "image", + "image_path": "e5fc9c5a11723ce6c34fb414641fd391df21360b5ad2c1f35e852f92d78d0aca.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 306, + 482, + 372 + ], + "lines": [ + { + "bbox": [ + 130, + 306, + 482, + 372 + ], + "spans": [ + { + "bbox": [ + 130, + 306, + 482, + 372 + ], + "type": "text", + "content": "Fig. 2: Proposed SNP methods, on each prunable component of the Transformer block. (a) SNP for query and key layers to preserve attention scores. (b) prunable components of the Transformer block. (c) SNP for value and other layers, including FFN and patch embedding. (d) conventional zeroing out in the matrix multiplication operator. (e) conventional zeroing out and graph-aware pruning in the residual connection." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 441, + 287, + 452, + 296 + ], + "blocks": [ + { + "bbox": [ + 441, + 287, + 452, + 296 + ], + "lines": [ + { + "bbox": [ + 441, + 287, + 452, + 296 + ], + "spans": [ + { + "bbox": [ + 441, + 287, + 452, + 296 + ], + "type": "image", + "image_path": "5d3c6fee3f2a567b24a6f3b0f3b7472b81693d0894ed3fc18078d60c048306ed.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 452, + 288, + 471, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 452, + 288, + 471, + 294 + ], + "spans": [ + { + "bbox": [ + 452, + 288, + 471, + 294 + ], + "type": "text", + "content": "Zeroing out" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 393, + 482, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 393, + 482, + 501 + ], + "spans": [ + { + "bbox": [ + 130, + 393, + 482, + 501 + ], + "type": "text", + "content": "Head and neuron-level pruning Structured pruning for the MSA module has two pruning objectives: head and neuron. Head pruning [29,30] reduces the number of heads, while neuron-level pruning [31] reduces the dimension of each query, key, and value layer in each head. Recent studies for pruning ViTs have focused on head pruning. X-Pruner [30] proposes a novel head pruning method for ViTs that introduces explainability-aware masks and measure the head importance, resulting in superior model compression. WDPruning [29] propose a method to control the number of attention heads and blocks via threshold on learnable parameters." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 502, + 482, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 502, + 482, + 586 + ], + "spans": [ + { + "bbox": [ + 130, + 502, + 482, + 586 + ], + "type": "text", + "content": "UVC [31] utilizes knowledge distillation alongside several pruning techniques, such as head pruning, block pruning, and neuron-level pruning. However, neuron-level pruning of UVC is carried out in a masking (zeroing out) manner, converting the weight matrix into a sparse matrix. For this reason, UVC necessitates additional libraries or hardware for accelerating sparse matrices, otherwise, the compressed model with UVC cannot achieve latency gain from the neuron-level pruning." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 601, + 233, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 233, + 616 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 233, + 616 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 624, + 228, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 624, + 228, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 624, + 228, + 635 + ], + "type": "text", + "content": "3.1 Preliminaries" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 640, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 640, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 640, + 481, + 666 + ], + "type": "text", + "content": "MSA module takes a single input " + }, + { + "bbox": [ + 130, + 640, + 481, + 666 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^{N \\times d}" + }, + { + "bbox": [ + 130, + 640, + 481, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 640, + 481, + 666 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 640, + 481, + 666 + ], + "type": "text", + "content": " denotes the input vector length, and " + }, + { + "bbox": [ + 130, + 640, + 481, + 666 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 640, + 481, + 666 + ], + "type": "text", + "content": " represents the hidden size. It comprises " + }, + { + "bbox": [ + 130, + 640, + 481, + 666 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 130, + 640, + 481, + 666 + ], + "type": "text", + "content": " heads, each consisting" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "Shim. et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 168 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 168 + ], + "type": "text", + "content": "of three linear layers: query, key, and value. Each layer denoted as " + }, + { + "bbox": [ + 130, + 115, + 479, + 168 + ], + "type": "inline_equation", + "content": "W_{\\{q,k,v\\}}^h \\in \\mathbb{R}^{d \\times d_{\\{q,k,v\\}}}" + }, + { + "bbox": [ + 130, + 115, + 479, + 168 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 115, + 479, + 168 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 130, + 115, + 479, + 168 + ], + "type": "text", + "content": " represents " + }, + { + "bbox": [ + 130, + 115, + 479, + 168 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 130, + 115, + 479, + 168 + ], + "type": "text", + "content": "-th head and " + }, + { + "bbox": [ + 130, + 115, + 479, + 168 + ], + "type": "inline_equation", + "content": "d_{\\{q,k,v\\}}" + }, + { + "bbox": [ + 130, + 115, + 479, + 168 + ], + "type": "text", + "content": " indicates the hidden size of each query, key, and value layer. " + }, + { + "bbox": [ + 130, + 115, + 479, + 168 + ], + "type": "inline_equation", + "content": "\\{Q,K,V\\}^h" + }, + { + "bbox": [ + 130, + 115, + 479, + 168 + ], + "type": "text", + "content": " signifies the output of each query, key, and value layer in " + }, + { + "bbox": [ + 130, + 115, + 479, + 168 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 130, + 115, + 479, + 168 + ], + "type": "text", + "content": "-th head, with a shape of " + }, + { + "bbox": [ + 130, + 115, + 479, + 168 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{N \\times d_{\\{q,k,v\\}}}" + }, + { + "bbox": [ + 130, + 115, + 479, + 168 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 146, + 169, + 457, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 169, + 457, + 180 + ], + "spans": [ + { + "bbox": [ + 146, + 169, + 457, + 180 + ], + "type": "text", + "content": "The self-attention operation for " + }, + { + "bbox": [ + 146, + 169, + 457, + 180 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 146, + 169, + 457, + 180 + ], + "type": "text", + "content": "-th head can be expressed as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 260, + 186, + 481, + 200 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 186, + 481, + 200 + ], + "spans": [ + { + "bbox": [ + 260, + 186, + 481, + 200 + ], + "type": "interline_equation", + "content": "\\operatorname {A s} ^ {h} (X) = Q ^ {h} \\cdot \\left(K ^ {h}\\right) ^ {T} \\tag {1}", + "image_path": "01252501c7448330275fffab59993e441885a8ba44fa7f9e5a276d455f940bc1.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 231, + 215, + 480, + 243 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 215, + 480, + 243 + ], + "spans": [ + { + "bbox": [ + 231, + 215, + 480, + 243 + ], + "type": "interline_equation", + "content": "\\operatorname {A t t} ^ {h} (X) = \\operatorname {S o f t m a x} \\left(\\frac {\\operatorname {A s} ^ {h} (X)}{\\sqrt {d _ {q}}}\\right) \\cdot V ^ {h} \\tag {2}", + "image_path": "3b8c3a056b7c2e9cc38e9c0fc633bd73068111937615dfe505a4529882650cc6.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 214, + 250, + 479, + 263 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 250, + 479, + 263 + ], + "spans": [ + { + "bbox": [ + 214, + 250, + 479, + 263 + ], + "type": "interline_equation", + "content": "\\operatorname {M S A} (X) = \\operatorname {C o n c a t} \\left(\\operatorname {A t t} ^ {1} (X), \\dots , \\operatorname {A t t} ^ {H} (X)\\right) \\tag {3}", + "image_path": "69adc0d5ca0211f99fd716b046da3a43b74c502240d2b891c775aa8f65c1df7a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 266, + 479, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 266, + 479, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 479, + 289 + ], + "type": "text", + "content": "here, As, Att, and MSA represent the functions to calculate attention scores, attention module, and MSA module respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 290, + 480, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 290, + 480, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 290, + 480, + 361 + ], + "type": "text", + "content": "Matrix multiplication, unlike a residual connection, yields zero when either of the inputs is masked, regardless of the value of the other input. Therefore, applying a conventional zeroing-out approach to neuron-level pruning in the MSA module can lead to unintended results, as shown in Fig. 2 (d). To address this issue, we introduce two graph-aware neuron-level pruning criteria to compress and expedite the MSA module." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 377, + 297, + 389 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 377, + 297, + 389 + ], + "spans": [ + { + "bbox": [ + 132, + 377, + 297, + 389 + ], + "type": "text", + "content": "3.2 Preserving attention scores" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 394, + 479, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 394, + 479, + 443 + ], + "spans": [ + { + "bbox": [ + 130, + 394, + 479, + 443 + ], + "type": "text", + "content": "Attention scores, Eq. (1), of the MSA module learn long-range dependencies between image features by focusing on different parts of the image when processing different features. These attention scores can be recognized as a series of outer product of " + }, + { + "bbox": [ + 130, + 394, + 479, + 443 + ], + "type": "inline_equation", + "content": "Q_{i}^{h}" + }, + { + "bbox": [ + 130, + 394, + 479, + 443 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 394, + 479, + 443 + ], + "type": "inline_equation", + "content": "(K_{i}^{h})^{T}" + }, + { + "bbox": [ + 130, + 394, + 479, + 443 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 252, + 449, + 480, + 499 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 449, + 480, + 499 + ], + "spans": [ + { + "bbox": [ + 252, + 449, + 480, + 499 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\operatorname {A s} ^ {h} (X) = Q ^ {h} \\cdot \\left(K ^ {h}\\right) ^ {T} \\\\ = \\sum_ {i = 1} ^ {d _ {q}} Q _ {i} ^ {h} \\cdot \\left(K _ {i} ^ {h}\\right) ^ {T} \\tag {4} \\\\ \\end{array}", + "image_path": "8f2fc44789641a98bff50ca195f3fe64875a7dd2c71d6c45cc61c0499ccde16c.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 503, + 479, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 503, + 479, + 550 + ], + "spans": [ + { + "bbox": [ + 130, + 503, + 479, + 550 + ], + "type": "text", + "content": "from this perspective, neuron-level pruning, reducing the dimension of query " + }, + { + "bbox": [ + 130, + 503, + 479, + 550 + ], + "type": "inline_equation", + "content": "d_{q}" + }, + { + "bbox": [ + 130, + 503, + 479, + 550 + ], + "type": "text", + "content": " and key " + }, + { + "bbox": [ + 130, + 503, + 479, + 550 + ], + "type": "inline_equation", + "content": "d_{k}" + }, + { + "bbox": [ + 130, + 503, + 479, + 550 + ], + "type": "text", + "content": ", inevitably distorts the attention scores. For this reason, preserving attention scores is essential to maintain the high performance of the original model." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 552, + 479, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 552, + 479, + 611 + ], + "spans": [ + { + "bbox": [ + 130, + 552, + 479, + 611 + ], + "type": "text", + "content": "To alleviate the distortion, we maintain the graphically connected query-key filter pair " + }, + { + "bbox": [ + 130, + 552, + 479, + 611 + ], + "type": "inline_equation", + "content": "(Q_{i}^{h}" + }, + { + "bbox": [ + 130, + 552, + 479, + 611 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 552, + 479, + 611 + ], + "type": "inline_equation", + "content": "K_{i}^{h})" + }, + { + "bbox": [ + 130, + 552, + 479, + 611 + ], + "type": "text", + "content": ", constituting a filter-by-attention score " + }, + { + "bbox": [ + 130, + 552, + 479, + 611 + ], + "type": "inline_equation", + "content": "(Q_{i}^{h}\\cdot (K_{i}^{h})^{T}\\in" + }, + { + "bbox": [ + 130, + 552, + 479, + 611 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{N\\times N})" + }, + { + "bbox": [ + 130, + 552, + 479, + 611 + ], + "type": "text", + "content": ", that retains the most significant aspects of the overall attention scores. To identify the most informative filter pair, we initially employ singular value decomposition (SVD) to decompose the original model's attention scores." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 252, + 617, + 480, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 617, + 480, + 667 + ], + "spans": [ + { + "bbox": [ + 252, + 617, + 480, + 667 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\operatorname {A s} ^ {h} (X) = \\tilde {U} \\cdot \\tilde {S} \\cdot \\tilde {V} ^ {T} \\\\ = \\sum_ {j = 1} ^ {N} \\tilde {\\sigma} _ {j} \\cdot \\tilde {u} _ {j} \\cdot \\tilde {v} _ {j} ^ {T} \\tag {5} \\\\ \\end{array}", + "image_path": "7ed7463cf6f6d8c72f8d5a7d86a253d7963c84fda9fa9ca128435e8780bf2fc8.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 170, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 170, + 91, + 448, + 102 + ], + "type": "text", + "content": "SNP: Structured Neuron-level Pruning to Preserve Attention Scores" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 140 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 114, + 480, + 140 + ], + "type": "inline_equation", + "content": "\\tilde{U}" + }, + { + "bbox": [ + 132, + 114, + 480, + 140 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 114, + 480, + 140 + ], + "type": "inline_equation", + "content": "\\tilde{V}" + }, + { + "bbox": [ + 132, + 114, + 480, + 140 + ], + "type": "text", + "content": " are the left and right singular vector matrices, respectively, and " + }, + { + "bbox": [ + 132, + 114, + 480, + 140 + ], + "type": "inline_equation", + "content": "\\tilde{S}" + }, + { + "bbox": [ + 132, + 114, + 480, + 140 + ], + "type": "text", + "content": " is the diagonal matrix of singular values, with " + }, + { + "bbox": [ + 132, + 114, + 480, + 140 + ], + "type": "inline_equation", + "content": "\\tilde{\\sigma}_1 \\geq \\tilde{\\sigma}_2 \\geq \\ldots \\geq \\tilde{\\sigma}_N" + }, + { + "bbox": [ + 132, + 114, + 480, + 140 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 479, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 479, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 479, + 200 + ], + "type": "text", + "content": "SVD is a technique for extracting the most informative components of a matrix, those with large singular values, while discarding the less informative components, those with small singular values. To retain the spatial relationships captured by the attention mechanism, we prune the filter pair " + }, + { + "bbox": [ + 130, + 141, + 479, + 200 + ], + "type": "inline_equation", + "content": "Q_{i}^{h}" + }, + { + "bbox": [ + 130, + 141, + 479, + 200 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 141, + 479, + 200 + ], + "type": "inline_equation", + "content": "K_{i}^{h}" + }, + { + "bbox": [ + 130, + 141, + 479, + 200 + ], + "type": "text", + "content": " with least correlation with the most informative components of the attention scores." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 201, + 479, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 201, + 479, + 251 + ], + "spans": [ + { + "bbox": [ + 130, + 201, + 479, + 251 + ], + "type": "text", + "content": "To measure the correlation, we adopt the cosine similarity between the attention scores of the " + }, + { + "bbox": [ + 130, + 201, + 479, + 251 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 201, + 479, + 251 + ], + "type": "text", + "content": "-th filter and the " + }, + { + "bbox": [ + 130, + 201, + 479, + 251 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 201, + 479, + 251 + ], + "type": "text", + "content": "-th rank matrix, as normalizing the attention scores reduces the impact of magnitude-based approaches. Consequently, the importance score " + }, + { + "bbox": [ + 130, + 201, + 479, + 251 + ], + "type": "inline_equation", + "content": "\\omega_{as,i}^{h}" + }, + { + "bbox": [ + 130, + 201, + 479, + 251 + ], + "type": "text", + "content": " is defined as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 209, + 263, + 480, + 327 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 263, + 480, + 327 + ], + "spans": [ + { + "bbox": [ + 209, + 263, + 480, + 327 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\omega_ {a s, i} ^ {h} = \\sum_ {j = 1} ^ {r} | \\cos ((Q _ {i} ^ {h} \\cdot (K _ {i} ^ {h}) ^ {T}), (\\tilde {\\sigma} _ {j} \\cdot \\tilde {u} _ {j} \\cdot \\tilde {v} _ {j} ^ {T})) | \\\\ = \\sum_ {j = 1} ^ {r} | \\frac {\\left(Q _ {i} ^ {h} \\cdot \\left(K _ {i} ^ {h}\\right) ^ {T}\\right) \\cdot \\left(\\tilde {\\sigma} _ {j} \\cdot \\tilde {u} _ {j} \\cdot \\tilde {v} _ {j} ^ {T}\\right)}{| | \\left(Q _ {i} ^ {h} \\cdot \\left(K _ {i} ^ {h}\\right) ^ {T}\\right) | | \\cdot | | \\left(\\tilde {\\sigma} _ {j} \\cdot \\tilde {u} _ {j} \\cdot \\tilde {v} _ {j} ^ {T}\\right) | |} | \\tag {6} \\\\ \\end{array}", + "image_path": "9a4ca0d9584b280671435d0d2fc00f8471595d3341c84490cf32ca1015dd6c52.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 339, + 479, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 339, + 479, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 339, + 479, + 374 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 339, + 479, + 374 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 132, + 339, + 479, + 374 + ], + "type": "text", + "content": " is a hyperparameter dictating the quantity of ranks, within the range of " + }, + { + "bbox": [ + 132, + 339, + 479, + 374 + ], + "type": "inline_equation", + "content": "1 \\leq r \\leq N" + }, + { + "bbox": [ + 132, + 339, + 479, + 374 + ], + "type": "text", + "content": ", to be compared with the " + }, + { + "bbox": [ + 132, + 339, + 479, + 374 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 132, + 339, + 479, + 374 + ], + "type": "text", + "content": "-th attention scores, while the remaining " + }, + { + "bbox": [ + 132, + 339, + 479, + 374 + ], + "type": "inline_equation", + "content": "N - r" + }, + { + "bbox": [ + 132, + 339, + 479, + 374 + ], + "type": "text", + "content": " singular values " + }, + { + "bbox": [ + 132, + 339, + 479, + 374 + ], + "type": "inline_equation", + "content": "(\\tilde{\\sigma}_{r + 1},\\dots,\\tilde{\\sigma}_N)" + }, + { + "bbox": [ + 132, + 339, + 479, + 374 + ], + "type": "text", + "content": " are discarded." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 376, + 479, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 479, + 448 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 479, + 448 + ], + "type": "text", + "content": "Optimizing " + }, + { + "bbox": [ + 130, + 376, + 479, + 448 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 376, + 479, + 448 + ], + "type": "text", + "content": " for each attention module can contribute to preserving informative filters and sustaining higher performance. For this reason, we adopt Variational Bayesian Matrix Factorization (VBMF) [21] to determine the optimal rank " + }, + { + "bbox": [ + 130, + 376, + 479, + 448 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 376, + 479, + 448 + ], + "type": "text", + "content": " for the attention scores. VBMF inherently addresses the challenge of determining the number of latent factors (ranks) in matrix factorization using its probabilistic approach." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 449, + 479, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 479, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 479, + 510 + ], + "type": "text", + "content": "Importance score " + }, + { + "bbox": [ + 130, + 449, + 479, + 510 + ], + "type": "inline_equation", + "content": "\\omega_{as,i}^{h}" + }, + { + "bbox": [ + 130, + 449, + 479, + 510 + ], + "type": "text", + "content": ", defined in Eq. (6), represents the importance of the " + }, + { + "bbox": [ + 130, + 449, + 479, + 510 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 449, + 479, + 510 + ], + "type": "text", + "content": "-th filter for both query and key layers. A larger " + }, + { + "bbox": [ + 130, + 449, + 479, + 510 + ], + "type": "inline_equation", + "content": "\\omega_{as,i}^{h}" + }, + { + "bbox": [ + 130, + 449, + 479, + 510 + ], + "type": "text", + "content": " indicates that the filter has a greater impact on the main component of the attention scores, while a lower " + }, + { + "bbox": [ + 130, + 449, + 479, + 510 + ], + "type": "inline_equation", + "content": "\\omega_{as,i}^{h}" + }, + { + "bbox": [ + 130, + 449, + 479, + 510 + ], + "type": "text", + "content": " indicates that the filter is less important and can be removed without significantly affecting the attention scores." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 532, + 318, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 532, + 318, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 532, + 318, + 544 + ], + "type": "text", + "content": "3.3 Inter-head redundancy removal" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 556, + 479, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 556, + 479, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 556, + 479, + 604 + ], + "type": "text", + "content": "In the preceding section, we outlined the approach to preserving attention scores even with reduced embedding dimensions in the query and key layers. Here, we introduce a pruning method for the value and other layers, such as FFN or patch embedding layer." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 605, + 479, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 479, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 479, + 667 + ], + "type": "text", + "content": "Previous works [2,19] have revealed that a significant proportion of attention heads can be removed without causing significant performance deterioration. To remove this inter-head redundancy through neuron-level pruning, we propose to measure the distance between all the value layers of MSA module, irrespective of the heads. The importance score " + }, + { + "bbox": [ + 130, + 605, + 479, + 667 + ], + "type": "inline_equation", + "content": "\\omega_{v,i}^{h}" + }, + { + "bbox": [ + 130, + 605, + 479, + 667 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 130, + 605, + 479, + 667 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 605, + 479, + 667 + ], + "type": "text", + "content": "-th filter in the value layer of the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "Shim. et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 235, + 127 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 235, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 235, + 127 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 132, + 116, + 235, + 127 + ], + "type": "text", + "content": "-th head is as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 231, + 133, + 481, + 166 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 133, + 481, + 166 + ], + "spans": [ + { + "bbox": [ + 231, + 133, + 481, + 166 + ], + "type": "interline_equation", + "content": "\\omega_ {v, i} ^ {h} = \\sum_ {j = 1} ^ {H} \\sum_ {l = 1} ^ {d _ {v}} \\left(1 - \\left| \\cos \\left(W _ {i} ^ {h}, W _ {l} ^ {j}\\right) \\right|\\right) \\tag {7}", + "image_path": "263451cb62b5daa1284d12fdf66b67853e3d220c33256d78803f1b2f555c6aaa.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 172, + 480, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 172, + 480, + 221 + ], + "spans": [ + { + "bbox": [ + 130, + 172, + 480, + 221 + ], + "type": "text", + "content": "The importance score of the value layer, denoted as " + }, + { + "bbox": [ + 130, + 172, + 480, + 221 + ], + "type": "inline_equation", + "content": "\\omega_{v,i}^{h}" + }, + { + "bbox": [ + 130, + 172, + 480, + 221 + ], + "type": "text", + "content": ", indicates the redundancy of filter " + }, + { + "bbox": [ + 130, + 172, + 480, + 221 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 172, + 480, + 221 + ], + "type": "text", + "content": " in the value layer compared to all value layers of the heads within the corresponding MSA layer. Consequently, filters with the lowest importance scores will be pruned in the initial stages of the pruning process." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 236, + 293, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 293, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 293, + 248 + ], + "type": "text", + "content": "3.4 Accelerating Transformers" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 253, + 479, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 253, + 479, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 253, + 479, + 289 + ], + "type": "text", + "content": "As described in Sec. 3.2, SNP removes the least correlated filter pairs " + }, + { + "bbox": [ + 130, + 253, + 479, + 289 + ], + "type": "inline_equation", + "content": "(Q_{i}^{h}, K_{i}^{h})" + }, + { + "bbox": [ + 130, + 253, + 479, + 289 + ], + "type": "text", + "content": " to preserve attention scores and enhance the efficiency of the MSA module across various devices." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 290, + 480, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 290, + 480, + 337 + ], + "spans": [ + { + "bbox": [ + 130, + 290, + 480, + 337 + ], + "type": "text", + "content": "Most MSA modules, including MSA modules in DeiTs, are designed to compute all heads in parallel. For achieving this parallel computation, SNP ensures that the number of filter dimensions are consistent across heads, even though the filter indices for each head are independently selected." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 338, + 480, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 338, + 480, + 421 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 480, + 421 + ], + "type": "text", + "content": "The green highlighted boxes in Fig. 2 (b) represent layers connected by a single residual connection at the last add layer of the Transformer block. Unlike the matrix multiplication operator, the output of the residual connection becomes zero when all the interconnected layers return zero for specific filter indices. For this reason, the residual connection with masking always exceeds the performance of actual pruning, which restricts the set of possible pruning patterns [26]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 422, + 480, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 480, + 480 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 480, + 480 + ], + "type": "text", + "content": "To accelerate the residual connection layers, all interconnected layers should be pruned identically. To achieve this, we sum up all the calculated importance scores for interconnected layers based on their filter indices, as shown in Fig. 2 (e). Subsequently, we prune the least important filter indices of all the connected layers." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 497, + 282, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 497, + 282, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 497, + 282, + 511 + ], + "type": "text", + "content": "4 Experimental Results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 520, + 479, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 520, + 479, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 520, + 479, + 581 + ], + "type": "text", + "content": "To ensure a fair comparison with existing methods, we apply SNP to prune the DeiT [24] architectures trained on the ImageNet-1K [6] dataset. Additionally, we conduct experiments on the efficient Transformer model, EfficientFormer-L1, to confirm the robustness of SNP. Furthermore, a series of ablation studies are conducted to gain a comprehensive understanding of our methodology." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 595, + 278, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 595, + 278, + 608 + ], + "spans": [ + { + "bbox": [ + 132, + 595, + 278, + 608 + ], + "type": "text", + "content": "4.1 Implementation details" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 613, + 480, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 613, + 480, + 638 + ], + "spans": [ + { + "bbox": [ + 130, + 613, + 480, + 638 + ], + "type": "text", + "content": "The overall pruning and fine-tuning process are executed on the pre-trained " + }, + { + "bbox": [ + 130, + 613, + 480, + 638 + ], + "type": "inline_equation", + "content": "\\mathrm{DeiT}^3" + }, + { + "bbox": [ + 130, + 613, + 480, + 638 + ], + "type": "text", + "content": " and EfficientFormer-L1" + }, + { + "bbox": [ + 130, + 613, + 480, + 638 + ], + "type": "inline_equation", + "content": "^4" + }, + { + "bbox": [ + 130, + 613, + 480, + 638 + ], + "type": "text", + "content": " released from the official implementation on" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 170, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 170, + 91, + 448, + 102 + ], + "type": "text", + "content": "SNP: Structured Neuron-level Pruning to Preserve Attention Scores" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 641, + 315, + 654 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 641, + 315, + 654 + ], + "spans": [ + { + "bbox": [ + 133, + 641, + 315, + 654 + ], + "type": "text", + "content": "3 https://github.com/facebookresearch/deit" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 136, + 654, + 348, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 654, + 348, + 665 + ], + "spans": [ + { + "bbox": [ + 136, + 654, + 348, + 665 + ], + "type": "text", + "content": "4 https://github.com/snap-research/EfficientFormer" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 180, + 481, + 406 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 169 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 169 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 169 + ], + "type": "text", + "content": "Table 1: Performance comparison of various pruning methods on ImageNet-1K. The \"Pruning\" column represents the pruning methods that corresponding methods use to compress the MSA module. Each of the methods contains one of follows: head pruning (HP), block pruning (BP), neuron-level pruning (NP), neuron-level sparsity (NS)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 180, + 481, + 406 + ], + "lines": [ + { + "bbox": [ + 132, + 180, + 481, + 406 + ], + "spans": [ + { + "bbox": [ + 132, + 180, + 481, + 406 + ], + "type": "table", + "html": "
ModelMethodPruningTop-1 (%)Top-5 (%)GFLOPsParams (M)
DeiT-TOriginal [24]-72.2091.101.35.7
SSViTE [4]HP70.12-0.94.2
WDPruning [29]HP+BP70.3489.820.73.5
X-Pruner [30]HP71.1090.110.6-
SNPNP70.2990.010.63.0
UVC [31]HP+NS+BP70.60-0.5-
DeiT-SOriginal [24]-79.8595.004.622.1
SSViTE [4]HP79.22-3.114.6
WDPruning [29]HP+BP78.3894.052.613.3
X-Pruner [30]HP78.9394.242.4-
UVC [31]HP+NS+BP78.82-2.3-
SNPNP78.5294.372.010.0
SNPNP73.3291.661.36.4
DeiT-BOriginal [24]-81.8095.5917.686.6
SSViTE [4]HP82.22-11.856.8
WDPruning [29]HP+BP80.7695.369.955.3
X-Pruner [30]HP81.0295.388.5-
UVC [31]HP+NS+BP80.57-8.0-
SNPNP79.6394.376.431.6
", + "image_path": "15471ad10ad38383a868c88a27cfe4a750512f45653f5f0a72ededd9f36ea634.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 432, + 480, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 432, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 432, + 480, + 479 + ], + "type": "text", + "content": "ImageNet-1K. Throughout the fine-tuning phase of the pruned model, we maintain consistent settings across all models, except for the batch size and learning rate. The batch size is set to 256, and to prevent weight explosion, we adjust the learning rate of the compressed model to 1/10 or 1/100 of the original model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 480, + 481, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 480, + 481, + 577 + ], + "spans": [ + { + "bbox": [ + 130, + 480, + 481, + 577 + ], + "type": "text", + "content": "To evaluate the reduced latency using SNP, we have configured four testing scenarios: one on a CPU and another on GPU for both edge devices and server processors. We employ a standard PyTorch model for profiling on the server processors (Intel Xeon Silver 4210R and NVIDIA GeForce RTX 3090). Profiling on the Raspberry Pi 4B and Jetson Nano is conducted using the ONNX and TensorRT formats, respectively. All latencies are measured using a single image as an input, except for the GPU of the server processor (RTX 3090), where it is set to 64 images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 596, + 261, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 596, + 261, + 608 + ], + "spans": [ + { + "bbox": [ + 132, + 596, + 261, + 608 + ], + "type": "text", + "content": "4.2 Quantitative results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "content": "Comparison with other methods Despite the constraints outlined in Sec. 3.4, SNP not only maintains accuracy comparable to existing methods but also significantly reduces inference time across diverse hardware and data types on the ImageNet-1K dataset, as shown in Tab. 1 and Tab. 2." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "Shim. et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 190, + 482, + 290 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "text", + "content": "Table 2: Inference speed and Top-1 accuracy of the compressed model across different devices. Performance evaluation involves accuracy on ImageNet-1K and inference time for the compressed DeiTs and EfficientFormer-L1. Latency is benchmarked with 200 warm-up runs and averaged over 1000 runs. In latency measurement, a single image is used as an input, except for the RTX 3090, where 64 images are employed in a single batch." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 190, + 482, + 290 + ], + "lines": [ + { + "bbox": [ + 135, + 190, + 482, + 290 + ], + "spans": [ + { + "bbox": [ + 135, + 190, + 482, + 290 + ], + "type": "table", + "html": "
ModelTop-1 (%)GFLOPsEdge devices (ms)Server processors (ms)
Raspberry Pi 4B (.onnx)Jetson Nano (.trt)Xeon Silver 4210R (.pt)RTX 3090 (.pt)
DeiT-Tiny72.201.3139.1341.0334.7418.65
+ SNP (Ours)70.290.681.63 (1.70×)26.67 (1.54×)25.25 (1.38×)17.82 (1.05×)
DeiT-Small79.804.6401.2799.3253.3746.13
+ SNP (Ours)78.522.0199.15 (2.01×)45.51 (2.18×)38.57 (1.38×)32.91 (1.40×)
+ SNP (Ours)73.321.3136.68 (2.94×)32.03 (3.10×)33.46 (1.60×)26.98 (1.71×)
DeiT-Base81.8017.61377.71293.29122.03151.35
+ SNP (Ours)79.636.4565.68 (2.44×)132.55 (2.21×)64.65 (1.89×)72.96 (2.07×)
EfficientFormer-L179.201.3169.1330.9543.7526.19
+ SNP (Ours)75.530.695.12 (1.78×)19.78 (1.56×)38.25 (1.14×)17.24 (1.52×)
+ SNP (Ours)74.510.582.60 (2.05×)17.76 (1.74×)35.15 (1.24×)16.01 (1.64×)
", + "image_path": "9c87df815b8f0d7394ce3702bd49457dda95631b75d88c1445651327c8b089a2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 315, + 482, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 315, + 482, + 387 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 482, + 387 + ], + "type": "text", + "content": "In a recent study [26], unconstrained masking generally outperforms posttraining accuracy of pruned models by an average of " + }, + { + "bbox": [ + 130, + 315, + 482, + 387 + ], + "type": "inline_equation", + "content": "2.1\\%" + }, + { + "bbox": [ + 130, + 315, + 482, + 387 + ], + "type": "text", + "content": " on ImageNet-1K. Compared to unconstrained head masking approaches like SSVITE [4] and XPruner [30], SNP achieves significantly higher compression rates for all DeiTs FLOPs (30.64% and 10.64%) with minimal performance degradation (0.83% and 0.87%), much less than the " + }, + { + "bbox": [ + 130, + 315, + 482, + 387 + ], + "type": "inline_equation", + "content": "2.1\\%" + }, + { + "bbox": [ + 130, + 315, + 482, + 387 + ], + "type": "text", + "content": " average mentioned." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 387, + 482, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 387, + 482, + 471 + ], + "spans": [ + { + "bbox": [ + 130, + 387, + 482, + 471 + ], + "type": "text", + "content": "Furthermore, compared to other pruning approaches like WDPruning [29] and UVC [31], which use a combination of pruning techniques to compress DeiTs, SNP achieves comparable accuracy solely through neuron-level pruning. Notably, DeiT-Small with SNP outperforms WDPruning by " + }, + { + "bbox": [ + 130, + 387, + 482, + 471 + ], + "type": "inline_equation", + "content": "0.14\\%" + }, + { + "bbox": [ + 130, + 387, + 482, + 471 + ], + "type": "text", + "content": ", with the removal of 3.3 million parameters and a reduction of 0.6 GFLOPs. Compared to UVC, SNP exhibits negligible performance degradation, averaging " + }, + { + "bbox": [ + 130, + 387, + 482, + 471 + ], + "type": "inline_equation", + "content": "0.51\\%" + }, + { + "bbox": [ + 130, + 387, + 482, + 471 + ], + "type": "text", + "content": " across all DeiTs while using " + }, + { + "bbox": [ + 130, + 387, + 482, + 471 + ], + "type": "inline_equation", + "content": "7.65\\%" + }, + { + "bbox": [ + 130, + 387, + 482, + 471 + ], + "type": "text", + "content": " fewer FLOPs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 490, + 482, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 490, + 482, + 610 + ], + "spans": [ + { + "bbox": [ + 130, + 490, + 482, + 610 + ], + "type": "text", + "content": "Large compressed vs. Small hand-crafted DeiT-Small with SNP, a large pruned model, outperforms the smaller, hand-crafted DeiT-Tiny in both accuracy and latency, achieving a notable " + }, + { + "bbox": [ + 130, + 490, + 482, + 610 + ], + "type": "inline_equation", + "content": "1.12\\%" + }, + { + "bbox": [ + 130, + 490, + 482, + 610 + ], + "type": "text", + "content": " improvement in top-1 accuracy while maintaining similar FLOPs. Additionally, DeiT-Small with SNP exhibits enhanced speed compared to the original DeiT-Tiny across edge devices and CPU-based server processors. Notably, its speed increases up to " + }, + { + "bbox": [ + 130, + 490, + 482, + 610 + ], + "type": "inline_equation", + "content": "21.94\\%" + }, + { + "bbox": [ + 130, + 490, + 482, + 610 + ], + "type": "text", + "content": " compared to the original DeiT-Tiny running on Jetson Nano, a GPU-based edge device. This substantial performance gap underscores the superiority of the compressed model (DeiT-Small with SNP) over the smaller hand-crafted designed model (DeiT-Tiny) in both overall performance and speed." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "Accelerating Transformer-based models As depicted in Tab. 2, SNP achieves impressive acceleration of DeiTs by a factor of " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "1.44 \\times" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "2.44 \\times" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " on edge devices and " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "1.05 \\times" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "2.07 \\times" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " on server processors. This acceleration is notable," + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 171, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 171, + 91, + 447, + 102 + ], + "type": "text", + "content": "SNP: Structured Neuron-level Pruning to Preserve Attention Scores" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 203, + 114, + 413, + 349 + ], + "blocks": [ + { + "bbox": [ + 203, + 114, + 413, + 349 + ], + "lines": [ + { + "bbox": [ + 203, + 114, + 413, + 349 + ], + "spans": [ + { + "bbox": [ + 203, + 114, + 413, + 349 + ], + "type": "image", + "image_path": "6d63344579af9058c87aa70dad1df97d56f7b38419d208c37bd1ecb3cc71554f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 357, + 482, + 411 + ], + "lines": [ + { + "bbox": [ + 130, + 357, + 482, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 357, + 482, + 411 + ], + "type": "text", + "content": "Fig.3: Attention maps with varying pruning criteria and compression ratios. All query and key layers are locally pruned based on the specified pruning ratio without fine-tuning. The importance scores of " + }, + { + "bbox": [ + 130, + 357, + 482, + 411 + ], + "type": "inline_equation", + "content": "l2" + }, + { + "bbox": [ + 130, + 357, + 482, + 411 + ], + "type": "text", + "content": "-norm and GM on query and key layers are combined by filter index and pruned simultaneously. \"Reverse\" represents the reverse order of SNP." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 441, + 480, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 441, + 480, + 466 + ], + "spans": [ + { + "bbox": [ + 130, + 441, + 480, + 466 + ], + "type": "text", + "content": "with negligible average performance degradation of " + }, + { + "bbox": [ + 130, + 441, + 480, + 466 + ], + "type": "inline_equation", + "content": "1.79\\%" + }, + { + "bbox": [ + 130, + 441, + 480, + 466 + ], + "type": "text", + "content": ", specifically " + }, + { + "bbox": [ + 130, + 441, + 480, + 466 + ], + "type": "inline_equation", + "content": "1.91\\%" + }, + { + "bbox": [ + 130, + 441, + 480, + 466 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 441, + 480, + 466 + ], + "type": "inline_equation", + "content": "1.28\\%" + }, + { + "bbox": [ + 130, + 441, + 480, + 466 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 441, + 480, + 466 + ], + "type": "inline_equation", + "content": "2.17\\%" + }, + { + "bbox": [ + 130, + 441, + 480, + 466 + ], + "type": "text", + "content": " for DeiT-Tiny, DeiT-Small, and DeiT-Base, respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 468, + 482, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 482, + 527 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 482, + 527 + ], + "type": "text", + "content": "Compared to WDPruning, which employs both head and block pruning, SNP surpasses in terms of latency for DeiT-Small and DeiT-Base on the RTX 3090 GPU. SNP accelerates the original DeiTs by " + }, + { + "bbox": [ + 130, + 468, + 482, + 527 + ], + "type": "inline_equation", + "content": "1.38 \\times" + }, + { + "bbox": [ + 130, + 468, + 482, + 527 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 468, + 482, + 527 + ], + "type": "inline_equation", + "content": "2.07 \\times" + }, + { + "bbox": [ + 130, + 468, + 482, + 527 + ], + "type": "text", + "content": ", respectively, while WDPruning achieves a comparatively modest acceleration of " + }, + { + "bbox": [ + 130, + 468, + 482, + 527 + ], + "type": "inline_equation", + "content": "1.18 \\times" + }, + { + "bbox": [ + 130, + 468, + 482, + 527 + ], + "type": "text", + "content": " for both models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 531, + 482, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 531, + 482, + 603 + ], + "spans": [ + { + "bbox": [ + 130, + 531, + 482, + 603 + ], + "type": "text", + "content": "The superiority of SNP becomes more evident when the target of neuron-level pruning, linear or convolutional layers, contribute a larger proportion of the original model's computation time. we find that SNP further accelerates original model " + }, + { + "bbox": [ + 130, + 531, + 482, + 603 + ], + "type": "inline_equation", + "content": "1.05 \\times" + }, + { + "bbox": [ + 130, + 531, + 482, + 603 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 531, + 482, + 603 + ], + "type": "inline_equation", + "content": "2.07 \\times" + }, + { + "bbox": [ + 130, + 531, + 482, + 603 + ], + "type": "text", + "content": " as the DeiT model's size grows on the RTX 3090 GPU. In contrast, WDPruning, involving the removal of entire layers and associated operators in both head and block, consistently shows acceleration rates of " + }, + { + "bbox": [ + 130, + 531, + 482, + 603 + ], + "type": "inline_equation", + "content": "1.18 \\times" + }, + { + "bbox": [ + 130, + 531, + 482, + 603 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "To ascertain the robustness of SNP across diverse Transformer models, especially on the efficiently designed model, we conduct additional experiments on EfficientFormer-L1. SNP accelerates the model by " + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "1.78 \\times" + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "2.05 \\times" + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": " faster on Raspberry Pi 4B and " + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "1.56 \\times" + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "1.74 \\times" + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": " faster on Jetson Nano. Notably, the compressed EfficientFormer-L1 achieves acceptable accuracy of " + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "75.53\\%" + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "74.51\\%" + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "text", + "content": "Shim. et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 190, + 169 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 190, + 169 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 190, + 169 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 190, + 169 + ], + "type": "image", + "image_path": "fe1f4d12f872427ee2e4c2882de994577a236ac3db60578a73e73447f1e97c39.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 154, + 170, + 171, + 175 + ], + "lines": [ + { + "bbox": [ + 154, + 170, + 171, + 175 + ], + "spans": [ + { + "bbox": [ + 154, + 170, + 171, + 175 + ], + "type": "text", + "content": "Original" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 192, + 114, + 246, + 170 + ], + "blocks": [ + { + "bbox": [ + 192, + 114, + 246, + 170 + ], + "lines": [ + { + "bbox": [ + 192, + 114, + 246, + 170 + ], + "spans": [ + { + "bbox": [ + 192, + 114, + 246, + 170 + ], + "type": "image", + "image_path": "6323a610de3b6a87de2951e2534265b0f3ea1293eae2b48888c961379a775071.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 207, + 171, + 231, + 175 + ], + "lines": [ + { + "bbox": [ + 207, + 171, + 231, + 175 + ], + "spans": [ + { + "bbox": [ + 207, + 171, + 231, + 175 + ], + "type": "text", + "content": "Compressed" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 248, + 114, + 303, + 170 + ], + "blocks": [ + { + "bbox": [ + 248, + 114, + 303, + 170 + ], + "lines": [ + { + "bbox": [ + 248, + 114, + 303, + 170 + ], + "spans": [ + { + "bbox": [ + 248, + 114, + 303, + 170 + ], + "type": "image", + "image_path": "f495f899f667f74c1ebea9d84560b323a1faed3a302230ccb3ef7f71e9b587c1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 251, + 171, + 301, + 175 + ], + "lines": [ + { + "bbox": [ + 251, + 171, + 301, + 175 + ], + "spans": [ + { + "bbox": [ + 251, + 171, + 301, + 175 + ], + "type": "text", + "content": "Compressed + Fine-tuned" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 310, + 114, + 365, + 170 + ], + "blocks": [ + { + "bbox": [ + 310, + 114, + 365, + 170 + ], + "lines": [ + { + "bbox": [ + 310, + 114, + 365, + 170 + ], + "spans": [ + { + "bbox": [ + 310, + 114, + 365, + 170 + ], + "type": "image", + "image_path": "f30d324ea6e826972f11170d584acf8e3fe343abc0f6c1e17e2631f0f4df2e9e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 171, + 345, + 175 + ], + "lines": [ + { + "bbox": [ + 329, + 171, + 345, + 175 + ], + "spans": [ + { + "bbox": [ + 329, + 171, + 345, + 175 + ], + "type": "text", + "content": "Original" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 366, + 114, + 422, + 170 + ], + "blocks": [ + { + "bbox": [ + 366, + 114, + 422, + 170 + ], + "lines": [ + { + "bbox": [ + 366, + 114, + 422, + 170 + ], + "spans": [ + { + "bbox": [ + 366, + 114, + 422, + 170 + ], + "type": "image", + "image_path": "c088f449756ed2a658ee4db4eea0f3dfaa2917b693a573088d8ade161617983d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 383, + 171, + 406, + 175 + ], + "lines": [ + { + "bbox": [ + 383, + 171, + 406, + 175 + ], + "spans": [ + { + "bbox": [ + 383, + 171, + 406, + 175 + ], + "type": "text", + "content": "Compressed" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 424, + 114, + 479, + 170 + ], + "blocks": [ + { + "bbox": [ + 424, + 114, + 479, + 170 + ], + "lines": [ + { + "bbox": [ + 424, + 114, + 479, + 170 + ], + "spans": [ + { + "bbox": [ + 424, + 114, + 479, + 170 + ], + "type": "image", + "image_path": "0ea0c7fbda0f1faafa893bf05469fc6f2f2992d7809c48d50a19a03231c65570.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 426, + 171, + 476, + 175 + ], + "lines": [ + { + "bbox": [ + 426, + 171, + 476, + 175 + ], + "spans": [ + { + "bbox": [ + 426, + 171, + 476, + 175 + ], + "type": "text", + "content": "Compressed + Fine-tuned" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 134, + 176, + 190, + 182 + ], + "blocks": [ + { + "bbox": [ + 134, + 176, + 190, + 182 + ], + "lines": [ + { + "bbox": [ + 134, + 176, + 190, + 182 + ], + "spans": [ + { + "bbox": [ + 134, + 176, + 190, + 182 + ], + "type": "image", + "image_path": "a1e97d41a8cea68255e7a9b05107ab15f937d340b1c469d8eeff847609e8b360.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 134, + 182, + 190, + 188 + ], + "blocks": [ + { + "bbox": [ + 134, + 182, + 190, + 188 + ], + "lines": [ + { + "bbox": [ + 134, + 182, + 190, + 188 + ], + "spans": [ + { + "bbox": [ + 134, + 182, + 190, + 188 + ], + "type": "image", + "image_path": "e4cf7f34b9d24d2d89564d5c2c695ea84ad328de030288edc83aaa2d40aaafe9.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 134, + 188, + 190, + 194 + ], + "blocks": [ + { + "bbox": [ + 134, + 188, + 190, + 194 + ], + "lines": [ + { + "bbox": [ + 134, + 188, + 190, + 194 + ], + "spans": [ + { + "bbox": [ + 134, + 188, + 190, + 194 + ], + "type": "image", + "image_path": "92ca6126fbf405bbe02594c29cd89eafeb1345f69377f89fd39db6cacdac27df.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 134, + 194, + 190, + 201 + ], + "blocks": [ + { + "bbox": [ + 134, + 194, + 190, + 201 + ], + "lines": [ + { + "bbox": [ + 134, + 194, + 190, + 201 + ], + "spans": [ + { + "bbox": [ + 134, + 194, + 190, + 201 + ], + "type": "image", + "image_path": "043764bb3f9586de9783dd19fd78bfe6789c008dc0dfe86a14e182e87f85b1e0.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 134, + 201, + 190, + 207 + ], + "blocks": [ + { + "bbox": [ + 134, + 201, + 190, + 207 + ], + "lines": [ + { + "bbox": [ + 134, + 201, + 190, + 207 + ], + "spans": [ + { + "bbox": [ + 134, + 201, + 190, + 207 + ], + "type": "image", + "image_path": "3f28d2f837980e4e95e62d2210f97120d05337bfb2cef20679ab14e68cee5230.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 134, + 207, + 190, + 213 + ], + "blocks": [ + { + "bbox": [ + 134, + 207, + 190, + 213 + ], + "lines": [ + { + "bbox": [ + 134, + 207, + 190, + 213 + ], + "spans": [ + { + "bbox": [ + 134, + 207, + 190, + 213 + ], + "type": "image", + "image_path": "1e147a49ae4d164c1753880b556a9c73d16b28742f70514b921c7a109872d675.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 134, + 213, + 190, + 219 + ], + "blocks": [ + { + "bbox": [ + 134, + 213, + 190, + 219 + ], + "lines": [ + { + "bbox": [ + 134, + 213, + 190, + 219 + ], + "spans": [ + { + "bbox": [ + 134, + 213, + 190, + 219 + ], + "type": "image", + "image_path": "e46717d66c317f1268b3e24b641d9d8f10dbae092dba3851ec623446d258ad24.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 134, + 219, + 190, + 225 + ], + "blocks": [ + { + "bbox": [ + 134, + 219, + 190, + 225 + ], + "lines": [ + { + "bbox": [ + 134, + 219, + 190, + 225 + ], + "spans": [ + { + "bbox": [ + 134, + 219, + 190, + 225 + ], + "type": "image", + "image_path": "02649542649345d8d47fac7611e4659a93655d187db203c620b577879fb96272.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 134, + 225, + 190, + 231 + ], + "blocks": [ + { + "bbox": [ + 134, + 225, + 190, + 231 + ], + "lines": [ + { + "bbox": [ + 134, + 225, + 190, + 231 + ], + "spans": [ + { + "bbox": [ + 134, + 225, + 190, + 231 + ], + "type": "image", + "image_path": "b9fa9dee01639061707c36607e68657d83d0d4250357ab057048d9a50f5fd60f.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 134, + 231, + 190, + 236 + ], + "blocks": [ + { + "bbox": [ + 134, + 231, + 190, + 236 + ], + "lines": [ + { + "bbox": [ + 134, + 231, + 190, + 236 + ], + "spans": [ + { + "bbox": [ + 134, + 231, + 190, + 236 + ], + "type": "image", + "image_path": "80c3f9e675e8736b403f13dec05ab4f9f08cbffc7d0abf6bbcd7fffe6aedd9c5.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 261, + 481, + 306 + ], + "lines": [ + { + "bbox": [ + 132, + 261, + 481, + 306 + ], + "spans": [ + { + "bbox": [ + 132, + 261, + 481, + 306 + ], + "type": "text", + "content": "Fig. 4: Attention maps from the original, compressed, and fine-tuned DeiT-Tiny with SNP. The attention maps in the first row are visualized using the attention rollout [1]. Each red box contains three attention maps from each head of the MSA module, ordered accordingly." + } + ] + } + ], + "index": 90, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 134, + 236, + 190, + 242 + ], + "blocks": [ + { + "bbox": [ + 134, + 236, + 190, + 242 + ], + "lines": [ + { + "bbox": [ + 134, + 236, + 190, + 242 + ], + "spans": [ + { + "bbox": [ + 134, + 236, + 190, + 242 + ], + "type": "image", + "image_path": "a6a84efa7ff27e3d0f5799df9789e0bf10cce09c7e4dc607d8571f21e8d31059.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 134, + 242, + 190, + 247 + ], + "blocks": [ + { + "bbox": [ + 134, + 242, + 190, + 247 + ], + "lines": [ + { + "bbox": [ + 134, + 242, + 190, + 247 + ], + "spans": [ + { + "bbox": [ + 134, + 242, + 190, + 247 + ], + "type": "image", + "image_path": "8733c633fd40e21ad93c447045e6e70d37d8c82e6a5128398182a49c7a071c7b.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 192, + 176, + 246, + 182 + ], + "blocks": [ + { + "bbox": [ + 192, + 176, + 246, + 182 + ], + "lines": [ + { + "bbox": [ + 192, + 176, + 246, + 182 + ], + "spans": [ + { + "bbox": [ + 192, + 176, + 246, + 182 + ], + "type": "image", + "image_path": "c4657e25aa77e07d79fd2c1d8f3225347964d110ab3ee705b21897aa451ae4cb.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 192, + 182, + 246, + 188 + ], + "blocks": [ + { + "bbox": [ + 192, + 182, + 246, + 188 + ], + "lines": [ + { + "bbox": [ + 192, + 182, + 246, + 188 + ], + "spans": [ + { + "bbox": [ + 192, + 182, + 246, + 188 + ], + "type": "image", + "image_path": "733fb88eff8af16a930947a91c86c970e83bcdab4ab20295a91ff3a61cfe05e3.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 192, + 188, + 246, + 194 + ], + "blocks": [ + { + "bbox": [ + 192, + 188, + 246, + 194 + ], + "lines": [ + { + "bbox": [ + 192, + 188, + 246, + 194 + ], + "spans": [ + { + "bbox": [ + 192, + 188, + 246, + 194 + ], + "type": "image", + "image_path": "adbf66146248093a28a285b7928c9f3c461198191b1a51b8de3b2581ec023cd1.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 192, + 194, + 246, + 200 + ], + "blocks": [ + { + "bbox": [ + 192, + 194, + 246, + 200 + ], + "lines": [ + { + "bbox": [ + 192, + 194, + 246, + 200 + ], + "spans": [ + { + "bbox": [ + 192, + 194, + 246, + 200 + ], + "type": "image", + "image_path": "0dc1699ad20f9728167c256362b43505ffff8a63f3117f2981d60b51a9aa0c7b.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 192, + 200, + 246, + 207 + ], + "blocks": [ + { + "bbox": [ + 192, + 200, + 246, + 207 + ], + "lines": [ + { + "bbox": [ + 192, + 200, + 246, + 207 + ], + "spans": [ + { + "bbox": [ + 192, + 200, + 246, + 207 + ], + "type": "image", + "image_path": "5b3d85906cd7f77e264948e19e58181b3d6928c4b8c55f36eb727dd53b8c9134.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 192, + 207, + 246, + 213 + ], + "blocks": [ + { + "bbox": [ + 192, + 207, + 246, + 213 + ], + "lines": [ + { + "bbox": [ + 192, + 207, + 246, + 213 + ], + "spans": [ + { + "bbox": [ + 192, + 207, + 246, + 213 + ], + "type": "image", + "image_path": "05e40153d154fde971d94c85587107f7e94c4c1d400362961dfd9b96af661771.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 192, + 213, + 246, + 219 + ], + "blocks": [ + { + "bbox": [ + 192, + 213, + 246, + 219 + ], + "lines": [ + { + "bbox": [ + 192, + 213, + 246, + 219 + ], + "spans": [ + { + "bbox": [ + 192, + 213, + 246, + 219 + ], + "type": "image", + "image_path": "2c163d34778577e23b21f38d9f54d65670f591c1a9602c9bd095bbc91b24cc7b.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 192, + 219, + 246, + 225 + ], + "blocks": [ + { + "bbox": [ + 192, + 219, + 246, + 225 + ], + "lines": [ + { + "bbox": [ + 192, + 219, + 246, + 225 + ], + "spans": [ + { + "bbox": [ + 192, + 219, + 246, + 225 + ], + "type": "image", + "image_path": "122ea387d2a42ed8cd836a60e795faa22985d68c431685d36c6e0b2d78f2ff2c.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 192, + 225, + 246, + 231 + ], + "blocks": [ + { + "bbox": [ + 192, + 225, + 246, + 231 + ], + "lines": [ + { + "bbox": [ + 192, + 225, + 246, + 231 + ], + "spans": [ + { + "bbox": [ + 192, + 225, + 246, + 231 + ], + "type": "image", + "image_path": "216d524b367e0b7159db2a0df90b7796c13acf549b5b02adcd0a2a9fffab04df.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 192, + 231, + 246, + 236 + ], + "blocks": [ + { + "bbox": [ + 192, + 231, + 246, + 236 + ], + "lines": [ + { + "bbox": [ + 192, + 231, + 246, + 236 + ], + "spans": [ + { + "bbox": [ + 192, + 231, + 246, + 236 + ], + "type": "image", + "image_path": "48f57ef9d9e276d804a646c3424e4fb611d9016e3d3305ed1e68662b1a22652e.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 192, + 236, + 246, + 242 + ], + "blocks": [ + { + "bbox": [ + 192, + 236, + 246, + 242 + ], + "lines": [ + { + "bbox": [ + 192, + 236, + 246, + 242 + ], + "spans": [ + { + "bbox": [ + 192, + 236, + 246, + 242 + ], + "type": "image", + "image_path": "a0071d2447b2285f1beeac11a0a9d1049a6a772332819a5eb5215886eba57ce6.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 192, + 242, + 246, + 247 + ], + "blocks": [ + { + "bbox": [ + 192, + 242, + 246, + 247 + ], + "lines": [ + { + "bbox": [ + 192, + 242, + 246, + 247 + ], + "spans": [ + { + "bbox": [ + 192, + 242, + 246, + 247 + ], + "type": "image", + "image_path": "7f97ff62864dc438be9a9b156403674afc785197bb8015297585adc5dd2902b5.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 248, + 176, + 303, + 182 + ], + "blocks": [ + { + "bbox": [ + 248, + 176, + 303, + 182 + ], + "lines": [ + { + "bbox": [ + 248, + 176, + 303, + 182 + ], + "spans": [ + { + "bbox": [ + 248, + 176, + 303, + 182 + ], + "type": "image", + "image_path": "7bfe8e9879866ad1a6c551c96df7557d9964c16717a30e1f6983899afca7a108.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 248, + 182, + 303, + 188 + ], + "blocks": [ + { + "bbox": [ + 248, + 182, + 303, + 188 + ], + "lines": [ + { + "bbox": [ + 248, + 182, + 303, + 188 + ], + "spans": [ + { + "bbox": [ + 248, + 182, + 303, + 188 + ], + "type": "image", + "image_path": "2d45862860754f7be5655ecc809419140763d0d2b1a60f89a7e73eb04cdfdefb.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 248, + 188, + 303, + 194 + ], + "blocks": [ + { + "bbox": [ + 248, + 188, + 303, + 194 + ], + "lines": [ + { + "bbox": [ + 248, + 188, + 303, + 194 + ], + "spans": [ + { + "bbox": [ + 248, + 188, + 303, + 194 + ], + "type": "image", + "image_path": "318dd579dc52f9b742f1d86ef9aa91021a9c238e975c5a4c99728a01a8de4ec3.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 248, + 194, + 303, + 200 + ], + "blocks": [ + { + "bbox": [ + 248, + 194, + 303, + 200 + ], + "lines": [ + { + "bbox": [ + 248, + 194, + 303, + 200 + ], + "spans": [ + { + "bbox": [ + 248, + 194, + 303, + 200 + ], + "type": "image", + "image_path": "8600cc0e3899fc7aac7bcffef47550512e8fb77623099f2a1b515c27003a720f.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 248, + 200, + 303, + 207 + ], + "blocks": [ + { + "bbox": [ + 248, + 200, + 303, + 207 + ], + "lines": [ + { + "bbox": [ + 248, + 200, + 303, + 207 + ], + "spans": [ + { + "bbox": [ + 248, + 200, + 303, + 207 + ], + "type": "image", + "image_path": "f98793e63d20f9a9d33927b64ba63087ce6af79b800bb39f5bcea8668e8665b6.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 248, + 207, + 303, + 213 + ], + "blocks": [ + { + "bbox": [ + 248, + 207, + 303, + 213 + ], + "lines": [ + { + "bbox": [ + 248, + 207, + 303, + 213 + ], + "spans": [ + { + "bbox": [ + 248, + 207, + 303, + 213 + ], + "type": "image", + "image_path": "f556b8f22da9495a87ed7c26bf33cce21f6866d0dd03cf7c29419878a24b0548.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 248, + 213, + 303, + 219 + ], + "blocks": [ + { + "bbox": [ + 248, + 213, + 303, + 219 + ], + "lines": [ + { + "bbox": [ + 248, + 213, + 303, + 219 + ], + "spans": [ + { + "bbox": [ + 248, + 213, + 303, + 219 + ], + "type": "image", + "image_path": "dc00dd4deef70eecd465a32d4d5c909f7920a45768ad490b3908478f32e5164c.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 248, + 219, + 303, + 224 + ], + "blocks": [ + { + "bbox": [ + 248, + 219, + 303, + 224 + ], + "lines": [ + { + "bbox": [ + 248, + 219, + 303, + 224 + ], + "spans": [ + { + "bbox": [ + 248, + 219, + 303, + 224 + ], + "type": "image", + "image_path": "1265d570ac5366ba1c740910c160458a20d26cdcdc9f1ab78331d5268ee51c4b.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 248, + 224, + 303, + 231 + ], + "blocks": [ + { + "bbox": [ + 248, + 224, + 303, + 231 + ], + "lines": [ + { + "bbox": [ + 248, + 224, + 303, + 231 + ], + "spans": [ + { + "bbox": [ + 248, + 224, + 303, + 231 + ], + "type": "image", + "image_path": "9c41c3b2f1c42b754af61aa24e3c3daed53a7152b46f9e316471384d9c83d11f.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 248, + 231, + 303, + 236 + ], + "blocks": [ + { + "bbox": [ + 248, + 231, + 303, + 236 + ], + "lines": [ + { + "bbox": [ + 248, + 231, + 303, + 236 + ], + "spans": [ + { + "bbox": [ + 248, + 231, + 303, + 236 + ], + "type": "image", + "image_path": "ab34a27b16544d65dd65588fc3234fc124a3d24e49c664e3638c186460afcb7f.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + } + ], + "index": 47 + }, + { + "type": "image", + "bbox": [ + 248, + 236, + 303, + 242 + ], + "blocks": [ + { + "bbox": [ + 248, + 236, + 303, + 242 + ], + "lines": [ + { + "bbox": [ + 248, + 236, + 303, + 242 + ], + "spans": [ + { + "bbox": [ + 248, + 236, + 303, + 242 + ], + "type": "image", + "image_path": "68e1a02c5e9e3957959c1d53b488fed6a86730c7a59cc678ed5e984096a344d0.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + } + ], + "index": 48 + }, + { + "type": "image", + "bbox": [ + 248, + 242, + 303, + 247 + ], + "blocks": [ + { + "bbox": [ + 248, + 242, + 303, + 247 + ], + "lines": [ + { + "bbox": [ + 248, + 242, + 303, + 247 + ], + "spans": [ + { + "bbox": [ + 248, + 242, + 303, + 247 + ], + "type": "image", + "image_path": "f0c77594a2a0657a100a720728263c58e613ad73e9648f8710c6549c9a699d37.jpg" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_body" + } + ], + "index": 49 + }, + { + "type": "image", + "bbox": [ + 310, + 176, + 365, + 182 + ], + "blocks": [ + { + "bbox": [ + 310, + 176, + 365, + 182 + ], + "lines": [ + { + "bbox": [ + 310, + 176, + 365, + 182 + ], + "spans": [ + { + "bbox": [ + 310, + 176, + 365, + 182 + ], + "type": "image", + "image_path": "618191c91df49a5db67c4f792861f9e4e662552b65fb950866d2bfba05a88c4a.jpg" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_body" + } + ], + "index": 50 + }, + { + "type": "image", + "bbox": [ + 310, + 182, + 365, + 188 + ], + "blocks": [ + { + "bbox": [ + 310, + 182, + 365, + 188 + ], + "lines": [ + { + "bbox": [ + 310, + 182, + 365, + 188 + ], + "spans": [ + { + "bbox": [ + 310, + 182, + 365, + 188 + ], + "type": "image", + "image_path": "d506e8f8ab137f12aa0747c2b7be22465379f366ad7775f4b45daf5d8ba4432b.jpg" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_body" + } + ], + "index": 51 + }, + { + "type": "image", + "bbox": [ + 310, + 188, + 365, + 194 + ], + "blocks": [ + { + "bbox": [ + 310, + 188, + 365, + 194 + ], + "lines": [ + { + "bbox": [ + 310, + 188, + 365, + 194 + ], + "spans": [ + { + "bbox": [ + 310, + 188, + 365, + 194 + ], + "type": "image", + "image_path": "b94ba33d9bf2a0a0aac9d21e4a294cbe26fa5f485f5219eaaaf53e204c0b5c45.jpg" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_body" + } + ], + "index": 52 + }, + { + "type": "image", + "bbox": [ + 310, + 194, + 365, + 200 + ], + "blocks": [ + { + "bbox": [ + 310, + 194, + 365, + 200 + ], + "lines": [ + { + "bbox": [ + 310, + 194, + 365, + 200 + ], + "spans": [ + { + "bbox": [ + 310, + 194, + 365, + 200 + ], + "type": "image", + "image_path": "0b3d966682b866cb80a0e9ec3586d4e95bf0ae3fe281a6b3623bfbcc89614577.jpg" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_body" + } + ], + "index": 53 + }, + { + "type": "image", + "bbox": [ + 310, + 200, + 365, + 207 + ], + "blocks": [ + { + "bbox": [ + 310, + 200, + 365, + 207 + ], + "lines": [ + { + "bbox": [ + 310, + 200, + 365, + 207 + ], + "spans": [ + { + "bbox": [ + 310, + 200, + 365, + 207 + ], + "type": "image", + "image_path": "b69fc8c7e5b8abad7c52d50cf70066816270dec0650c575792c421a2d00390bd.jpg" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_body" + } + ], + "index": 54 + }, + { + "type": "image", + "bbox": [ + 310, + 207, + 365, + 213 + ], + "blocks": [ + { + "bbox": [ + 310, + 207, + 365, + 213 + ], + "lines": [ + { + "bbox": [ + 310, + 207, + 365, + 213 + ], + "spans": [ + { + "bbox": [ + 310, + 207, + 365, + 213 + ], + "type": "image", + "image_path": "354be57d221d5f468ac85f1c2acc6aaea3aabf4269d236655696504b97443df1.jpg" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_body" + } + ], + "index": 55 + }, + { + "type": "image", + "bbox": [ + 310, + 213, + 365, + 219 + ], + "blocks": [ + { + "bbox": [ + 310, + 213, + 365, + 219 + ], + "lines": [ + { + "bbox": [ + 310, + 213, + 365, + 219 + ], + "spans": [ + { + "bbox": [ + 310, + 213, + 365, + 219 + ], + "type": "image", + "image_path": "2a5938bd8b133b7d82cfefa4972bc6a980461af895027a54c53243ba0a70de4a.jpg" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_body" + } + ], + "index": 56 + }, + { + "type": "image", + "bbox": [ + 310, + 219, + 365, + 225 + ], + "blocks": [ + { + "bbox": [ + 310, + 219, + 365, + 225 + ], + "lines": [ + { + "bbox": [ + 310, + 219, + 365, + 225 + ], + "spans": [ + { + "bbox": [ + 310, + 219, + 365, + 225 + ], + "type": "image", + "image_path": "8c5882ec658ac9467160bd0710b675aae6e37c14569f6882d1d9d6623f4d3f8a.jpg" + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_body" + } + ], + "index": 57 + }, + { + "type": "image", + "bbox": [ + 310, + 225, + 365, + 231 + ], + "blocks": [ + { + "bbox": [ + 310, + 225, + 365, + 231 + ], + "lines": [ + { + "bbox": [ + 310, + 225, + 365, + 231 + ], + "spans": [ + { + "bbox": [ + 310, + 225, + 365, + 231 + ], + "type": "image", + "image_path": "4ece03b7e0f278b79f3ad24fa430ee3ac8862134854f0aa48d8807b7b6ed1e04.jpg" + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "image_body" + } + ], + "index": 58 + }, + { + "type": "image", + "bbox": [ + 310, + 231, + 365, + 237 + ], + "blocks": [ + { + "bbox": [ + 310, + 231, + 365, + 237 + ], + "lines": [ + { + "bbox": [ + 310, + 231, + 365, + 237 + ], + "spans": [ + { + "bbox": [ + 310, + 231, + 365, + 237 + ], + "type": "image", + "image_path": "fecd71612f8074ebef73698044fc46a9af96362fa4ef774437bfe46e6674c032.jpg" + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_body" + } + ], + "index": 59 + }, + { + "type": "image", + "bbox": [ + 310, + 237, + 365, + 243 + ], + "blocks": [ + { + "bbox": [ + 310, + 237, + 365, + 243 + ], + "lines": [ + { + "bbox": [ + 310, + 237, + 365, + 243 + ], + "spans": [ + { + "bbox": [ + 310, + 237, + 365, + 243 + ], + "type": "image", + "image_path": "0d2163883253187f7708895be4ae99120d136bb583e820d409e6acf67c8a9644.jpg" + } + ] + } + ], + "index": 60, + "angle": 0, + "type": "image_body" + } + ], + "index": 60 + }, + { + "type": "image", + "bbox": [ + 310, + 243, + 365, + 247 + ], + "blocks": [ + { + "bbox": [ + 310, + 243, + 365, + 247 + ], + "lines": [ + { + "bbox": [ + 310, + 243, + 365, + 247 + ], + "spans": [ + { + "bbox": [ + 310, + 243, + 365, + 247 + ], + "type": "image", + "image_path": "e056c869856b4f7047484e8018fc91945cd8339e04976e05e83a980f284f77e6.jpg" + } + ] + } + ], + "index": 61, + "angle": 0, + "type": "image_body" + } + ], + "index": 61 + }, + { + "type": "image", + "bbox": [ + 310, + 247, + 365, + 252 + ], + "blocks": [ + { + "bbox": [ + 310, + 247, + 365, + 252 + ], + "lines": [ + { + "bbox": [ + 310, + 247, + 365, + 252 + ], + "spans": [ + { + "bbox": [ + 310, + 247, + 365, + 252 + ], + "type": "image", + "image_path": "4c8f13fae9e6650345f57b4d567043ad8c8a10acf3f9a2143b79d749ea85bea1.jpg" + } + ] + } + ], + "index": 62, + "angle": 0, + "type": "image_body" + } + ], + "index": 62 + }, + { + "type": "image", + "bbox": [ + 367, + 176, + 422, + 182 + ], + "blocks": [ + { + "bbox": [ + 367, + 176, + 422, + 182 + ], + "lines": [ + { + "bbox": [ + 367, + 176, + 422, + 182 + ], + "spans": [ + { + "bbox": [ + 367, + 176, + 422, + 182 + ], + "type": "image", + "image_path": "1d25f4d1aa674a20408152d2157006593e6d85974ac47f722744149394d933d5.jpg" + } + ] + } + ], + "index": 64, + "angle": 0, + "type": "image_body" + } + ], + "index": 64 + }, + { + "type": "image", + "bbox": [ + 367, + 182, + 422, + 188 + ], + "blocks": [ + { + "bbox": [ + 367, + 182, + 422, + 188 + ], + "lines": [ + { + "bbox": [ + 367, + 182, + 422, + 188 + ], + "spans": [ + { + "bbox": [ + 367, + 182, + 422, + 188 + ], + "type": "image", + "image_path": "76e771d3611843403d36d3da83fde4ccc37213ed51e8d5f5fb21c5db355b7fb1.jpg" + } + ] + } + ], + "index": 65, + "angle": 0, + "type": "image_body" + } + ], + "index": 65 + }, + { + "type": "image", + "bbox": [ + 367, + 188, + 422, + 194 + ], + "blocks": [ + { + "bbox": [ + 367, + 188, + 422, + 194 + ], + "lines": [ + { + "bbox": [ + 367, + 188, + 422, + 194 + ], + "spans": [ + { + "bbox": [ + 367, + 188, + 422, + 194 + ], + "type": "image", + "image_path": "35f8c87b894ea7e4d1368f6b7ab46468ace45a6efdb1b0e9b56d1f4d47fad4e2.jpg" + } + ] + } + ], + "index": 66, + "angle": 0, + "type": "image_body" + } + ], + "index": 66 + }, + { + "type": "image", + "bbox": [ + 367, + 194, + 422, + 200 + ], + "blocks": [ + { + "bbox": [ + 367, + 194, + 422, + 200 + ], + "lines": [ + { + "bbox": [ + 367, + 194, + 422, + 200 + ], + "spans": [ + { + "bbox": [ + 367, + 194, + 422, + 200 + ], + "type": "image", + "image_path": "31c426fa47f0b63bb46e402e6bc6ffaf686ef738eb8439cae1cc7d125d481535.jpg" + } + ] + } + ], + "index": 67, + "angle": 0, + "type": "image_body" + } + ], + "index": 67 + }, + { + "type": "image", + "bbox": [ + 367, + 200, + 422, + 207 + ], + "blocks": [ + { + "bbox": [ + 367, + 200, + 422, + 207 + ], + "lines": [ + { + "bbox": [ + 367, + 200, + 422, + 207 + ], + "spans": [ + { + "bbox": [ + 367, + 200, + 422, + 207 + ], + "type": "image", + "image_path": "2a2c918e52fa5710e109cb0c128c6c7c22ba9e2c384ef01e318e9a0e1f7e56b7.jpg" + } + ] + } + ], + "index": 68, + "angle": 0, + "type": "image_body" + } + ], + "index": 68 + }, + { + "type": "image", + "bbox": [ + 367, + 207, + 422, + 213 + ], + "blocks": [ + { + "bbox": [ + 367, + 207, + 422, + 213 + ], + "lines": [ + { + "bbox": [ + 367, + 207, + 422, + 213 + ], + "spans": [ + { + "bbox": [ + 367, + 207, + 422, + 213 + ], + "type": "image", + "image_path": "610a3ad04eb3118d53ef7ea46059c44300eda5c6f6032cccbc1cd28dae25ba9f.jpg" + } + ] + } + ], + "index": 69, + "angle": 0, + "type": "image_body" + } + ], + "index": 69 + }, + { + "type": "image", + "bbox": [ + 367, + 213, + 422, + 219 + ], + "blocks": [ + { + "bbox": [ + 367, + 213, + 422, + 219 + ], + "lines": [ + { + "bbox": [ + 367, + 213, + 422, + 219 + ], + "spans": [ + { + "bbox": [ + 367, + 213, + 422, + 219 + ], + "type": "image", + "image_path": "373e4a86d8c2a753fc16a0f8587d8bd1bed4d021efdf9f93a639b8e8b60dd773.jpg" + } + ] + } + ], + "index": 70, + "angle": 0, + "type": "image_body" + } + ], + "index": 70 + }, + { + "type": "image", + "bbox": [ + 367, + 219, + 422, + 225 + ], + "blocks": [ + { + "bbox": [ + 367, + 219, + 422, + 225 + ], + "lines": [ + { + "bbox": [ + 367, + 219, + 422, + 225 + ], + "spans": [ + { + "bbox": [ + 367, + 219, + 422, + 225 + ], + "type": "image", + "image_path": "966bae9ff81f39d2fc182054cb32e20f63373e819cd950b539564ced0ad021d2.jpg" + } + ] + } + ], + "index": 71, + "angle": 0, + "type": "image_body" + } + ], + "index": 71 + }, + { + "type": "image", + "bbox": [ + 367, + 225, + 422, + 231 + ], + "blocks": [ + { + "bbox": [ + 367, + 225, + 422, + 231 + ], + "lines": [ + { + "bbox": [ + 367, + 225, + 422, + 231 + ], + "spans": [ + { + "bbox": [ + 367, + 225, + 422, + 231 + ], + "type": "image", + "image_path": "f4add919c6a72372542223bba125bcbddca915eb557c7212e4945484844a5cdd.jpg" + } + ] + } + ], + "index": 72, + "angle": 0, + "type": "image_body" + } + ], + "index": 72 + }, + { + "type": "image", + "bbox": [ + 367, + 231, + 422, + 237 + ], + "blocks": [ + { + "bbox": [ + 367, + 231, + 422, + 237 + ], + "lines": [ + { + "bbox": [ + 367, + 231, + 422, + 237 + ], + "spans": [ + { + "bbox": [ + 367, + 231, + 422, + 237 + ], + "type": "image", + "image_path": "607df98602bf3860ee5e4041956aacc419c68b5c6e61b2e8d43be2e564daf380.jpg" + } + ] + } + ], + "index": 73, + "angle": 0, + "type": "image_body" + } + ], + "index": 73 + }, + { + "type": "image", + "bbox": [ + 367, + 237, + 422, + 243 + ], + "blocks": [ + { + "bbox": [ + 367, + 237, + 422, + 243 + ], + "lines": [ + { + "bbox": [ + 367, + 237, + 422, + 243 + ], + "spans": [ + { + "bbox": [ + 367, + 237, + 422, + 243 + ], + "type": "image", + "image_path": "13276cb8d9e4f91f0dc8e713b4e7b890603f1090284dace91b825fadd2dc7cfb.jpg" + } + ] + } + ], + "index": 74, + "angle": 0, + "type": "image_body" + } + ], + "index": 74 + }, + { + "type": "image", + "bbox": [ + 367, + 243, + 422, + 247 + ], + "blocks": [ + { + "bbox": [ + 367, + 243, + 422, + 247 + ], + "lines": [ + { + "bbox": [ + 367, + 243, + 422, + 247 + ], + "spans": [ + { + "bbox": [ + 367, + 243, + 422, + 247 + ], + "type": "image", + "image_path": "979d3abae462302381b2126b7ae161cfb56fe82f11cc8f00a01130ba70571e48.jpg" + } + ] + } + ], + "index": 75, + "angle": 0, + "type": "image_body" + } + ], + "index": 75 + }, + { + "type": "image", + "bbox": [ + 367, + 247, + 422, + 252 + ], + "blocks": [ + { + "bbox": [ + 367, + 247, + 422, + 252 + ], + "lines": [ + { + "bbox": [ + 367, + 247, + 422, + 252 + ], + "spans": [ + { + "bbox": [ + 367, + 247, + 422, + 252 + ], + "type": "image", + "image_path": "0c5666a18bfdf3429739304cb15cc376a03fa0b521b4d10cf11e1ede8ff1bbd5.jpg" + } + ] + } + ], + "index": 76, + "angle": 0, + "type": "image_body" + } + ], + "index": 76 + }, + { + "type": "image", + "bbox": [ + 424, + 176, + 477, + 181 + ], + "blocks": [ + { + "bbox": [ + 424, + 176, + 477, + 181 + ], + "lines": [ + { + "bbox": [ + 424, + 176, + 477, + 181 + ], + "spans": [ + { + "bbox": [ + 424, + 176, + 477, + 181 + ], + "type": "image", + "image_path": "dcda949fcbe86e57760bf7fb2ca869b4d853ecfc47a5640786cb24d1e218b9e9.jpg" + } + ] + } + ], + "index": 78, + "angle": 0, + "type": "image_body" + } + ], + "index": 78 + }, + { + "type": "image", + "bbox": [ + 424, + 181, + 479, + 186 + ], + "blocks": [ + { + "bbox": [ + 424, + 181, + 479, + 186 + ], + "lines": [ + { + "bbox": [ + 424, + 181, + 479, + 186 + ], + "spans": [ + { + "bbox": [ + 424, + 181, + 479, + 186 + ], + "type": "image", + "image_path": "d185cbdb87880d909ac1d5b4e423ae5683659da251038a3b582146bc7d7d9ada.jpg" + } + ] + } + ], + "index": 79, + "angle": 0, + "type": "image_body" + } + ], + "index": 79 + }, + { + "type": "image", + "bbox": [ + 424, + 186, + 479, + 193 + ], + "blocks": [ + { + "bbox": [ + 424, + 186, + 479, + 193 + ], + "lines": [ + { + "bbox": [ + 424, + 186, + 479, + 193 + ], + "spans": [ + { + "bbox": [ + 424, + 186, + 479, + 193 + ], + "type": "image", + "image_path": "c3b74fbd38c0eced21ed711a499a711d3e5933c22bcd5e58a6b1b9ac5f364c2d.jpg" + } + ] + } + ], + "index": 80, + "angle": 0, + "type": "image_body" + } + ], + "index": 80 + }, + { + "type": "image", + "bbox": [ + 424, + 193, + 479, + 199 + ], + "blocks": [ + { + "bbox": [ + 424, + 193, + 479, + 199 + ], + "lines": [ + { + "bbox": [ + 424, + 193, + 479, + 199 + ], + "spans": [ + { + "bbox": [ + 424, + 193, + 479, + 199 + ], + "type": "image", + "image_path": "7c9a5c1ba7551fe44e8eb64d1ff9ff564ec584b523b1a31d16e38f53f8f51dc7.jpg" + } + ] + } + ], + "index": 81, + "angle": 0, + "type": "image_body" + } + ], + "index": 81 + }, + { + "type": "image", + "bbox": [ + 424, + 199, + 479, + 205 + ], + "blocks": [ + { + "bbox": [ + 424, + 199, + 479, + 205 + ], + "lines": [ + { + "bbox": [ + 424, + 199, + 479, + 205 + ], + "spans": [ + { + "bbox": [ + 424, + 199, + 479, + 205 + ], + "type": "image", + "image_path": "42af389f2d089c9cd68bbf2a8671f19c1d6dc20930d8b5690707e89317a99305.jpg" + } + ] + } + ], + "index": 82, + "angle": 0, + "type": "image_body" + } + ], + "index": 82 + }, + { + "type": "image", + "bbox": [ + 424, + 205, + 479, + 212 + ], + "blocks": [ + { + "bbox": [ + 424, + 205, + 479, + 212 + ], + "lines": [ + { + "bbox": [ + 424, + 205, + 479, + 212 + ], + "spans": [ + { + "bbox": [ + 424, + 205, + 479, + 212 + ], + "type": "image", + "image_path": "ccaf4e741482c658eac0d87b6b322f7e7c40a1e41743d334bfd1e13011e1ffa3.jpg" + } + ] + } + ], + "index": 83, + "angle": 0, + "type": "image_body" + } + ], + "index": 83 + }, + { + "type": "image", + "bbox": [ + 424, + 212, + 479, + 218 + ], + "blocks": [ + { + "bbox": [ + 424, + 212, + 479, + 218 + ], + "lines": [ + { + "bbox": [ + 424, + 212, + 479, + 218 + ], + "spans": [ + { + "bbox": [ + 424, + 212, + 479, + 218 + ], + "type": "image", + "image_path": "79e1fdd3460bfb2ffb97ca83563353083e6491f399ce931481c57f34ffaec664.jpg" + } + ] + } + ], + "index": 84, + "angle": 0, + "type": "image_body" + } + ], + "index": 84 + }, + { + "type": "image", + "bbox": [ + 424, + 218, + 479, + 224 + ], + "blocks": [ + { + "bbox": [ + 424, + 218, + 479, + 224 + ], + "lines": [ + { + "bbox": [ + 424, + 218, + 479, + 224 + ], + "spans": [ + { + "bbox": [ + 424, + 218, + 479, + 224 + ], + "type": "image", + "image_path": "c58674a657e7f9bc5ba2c64f813bf9ea653692c8c2f5d8e2587e5c8a2b988e7a.jpg" + } + ] + } + ], + "index": 85, + "angle": 0, + "type": "image_body" + } + ], + "index": 85 + }, + { + "type": "image", + "bbox": [ + 424, + 224, + 479, + 231 + ], + "blocks": [ + { + "bbox": [ + 424, + 224, + 479, + 231 + ], + "lines": [ + { + "bbox": [ + 424, + 224, + 479, + 231 + ], + "spans": [ + { + "bbox": [ + 424, + 224, + 479, + 231 + ], + "type": "image", + "image_path": "1e8e03716568d8677f5f7ed812813b2bc65ee25d3aa3f07f2737d5ae99b65b6d.jpg" + } + ] + } + ], + "index": 86, + "angle": 0, + "type": "image_body" + } + ], + "index": 86 + }, + { + "type": "image", + "bbox": [ + 424, + 231, + 479, + 237 + ], + "blocks": [ + { + "bbox": [ + 424, + 231, + 479, + 237 + ], + "lines": [ + { + "bbox": [ + 424, + 231, + 479, + 237 + ], + "spans": [ + { + "bbox": [ + 424, + 231, + 479, + 237 + ], + "type": "image", + "image_path": "23a32049cdd4e27ed7b0e3dce9156011db1140a1bf608b08e547c6895b2c97ec.jpg" + } + ] + } + ], + "index": 87, + "angle": 0, + "type": "image_body" + } + ], + "index": 87 + }, + { + "type": "image", + "bbox": [ + 424, + 237, + 479, + 243 + ], + "blocks": [ + { + "bbox": [ + 424, + 237, + 479, + 243 + ], + "lines": [ + { + "bbox": [ + 424, + 237, + 479, + 243 + ], + "spans": [ + { + "bbox": [ + 424, + 237, + 479, + 243 + ], + "type": "image", + "image_path": "0c1d342ed7ee47d058557fd7271830d8aacb9f2d342f6dc24a6cc6da9d49db2c.jpg" + } + ] + } + ], + "index": 88, + "angle": 0, + "type": "image_body" + } + ], + "index": 88 + }, + { + "type": "image", + "bbox": [ + 424, + 243, + 479, + 247 + ], + "blocks": [ + { + "bbox": [ + 424, + 243, + 479, + 247 + ], + "lines": [ + { + "bbox": [ + 424, + 243, + 479, + 247 + ], + "spans": [ + { + "bbox": [ + 424, + 243, + 479, + 247 + ], + "type": "image", + "image_path": "c75fa64d8f8053ecb7141f850a5bfa7815e2971e30e7f09e9b9881f3840922f0.jpg" + } + ] + } + ], + "index": 89, + "angle": 0, + "type": "image_body" + } + ], + "index": 89 + }, + { + "bbox": [ + 132, + 330, + 253, + 341 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 330, + 253, + 341 + ], + "spans": [ + { + "bbox": [ + 132, + 330, + 253, + 341 + ], + "type": "text", + "content": "4.3 Qualitative results" + } + ] + } + ], + "index": 91 + }, + { + "bbox": [ + 131, + 350, + 481, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 350, + 481, + 387 + ], + "spans": [ + { + "bbox": [ + 131, + 350, + 481, + 387 + ], + "type": "text", + "content": "In the subsequent sections, attention rollout [1] is employed to randomly selected images from the test datasets to visualize the attention maps of the DeiTs and assess the efficacy of SNP. This evaluation unfolds across two key facets:" + } + ] + } + ], + "index": 92 + }, + { + "bbox": [ + 137, + 395, + 480, + 443 + ], + "type": "list", + "angle": 0, + "index": 95, + "blocks": [ + { + "bbox": [ + 137, + 395, + 480, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 395, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 137, + 395, + 480, + 418 + ], + "type": "text", + "content": "- Preserving the attention scores: Visualize attention maps to verify the effectiveness of the proposed importance score in preserving attention scores." + } + ] + } + ], + "index": 93 + }, + { + "bbox": [ + 138, + 419, + 480, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 419, + 480, + 443 + ], + "spans": [ + { + "bbox": [ + 138, + 419, + 480, + 443 + ], + "type": "text", + "content": "- Restoring the attention scores after SNP: Visualize attention maps to find the effectiveness of SNP in restoring attention scores." + } + ] + } + ], + "index": 94 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 131, + 461, + 481, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 461, + 481, + 498 + ], + "spans": [ + { + "bbox": [ + 131, + 461, + 481, + 498 + ], + "type": "text", + "content": "Preserving the attention scores To visualize SNP effectiveness, we apply four pruning criteria to compress the query and key filter pairs in all MSA modules of DeiT-Tiny: SNP, " + }, + { + "bbox": [ + 131, + 461, + 481, + 498 + ], + "type": "inline_equation", + "content": "l2" + }, + { + "bbox": [ + 131, + 461, + 481, + 498 + ], + "type": "text", + "content": "-norm, geometric median (GM) [11], and \"Reverse\"." + } + ] + } + ], + "index": 96 + }, + { + "bbox": [ + 131, + 498, + 481, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 498, + 481, + 569 + ], + "spans": [ + { + "bbox": [ + 131, + 498, + 481, + 569 + ], + "type": "text", + "content": "The \"Reverse\" criterion prioritizes pruning the most important filter pairs first, keeping the least important pairs until the end. However, both " + }, + { + "bbox": [ + 131, + 498, + 481, + 569 + ], + "type": "inline_equation", + "content": "l2" + }, + { + "bbox": [ + 131, + 498, + 481, + 569 + ], + "type": "text", + "content": "-norm and GM pruning criteria ignore the MSA module's graphical connectivity, evaluating importance scores independently for query and key layers. To handle identical filter indices during pruning, we aggregate scores based on filter indices, removing the least important indices from both layers." + } + ] + } + ], + "index": 97 + }, + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "text", + "content": "In Fig. 3, attention maps for the original DeiT-Tiny and locally pruned models are presented across various pruning ratios (10% to 90% with 10% intervals). Our proposed method effectively maintains the original attention map even after pruning over 80% of the filters, whereas other methods show fragmented attention maps at much lower pruning ratios (typically 30% or less). These results highlight the potential of our neuron-level pruning criteria, utilizing SVD to preserve attention scores, in reducing the size and speeding up the execution of MSA modules without compromising accuracy." + } + ] + } + ], + "index": 98 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 171, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 171, + 91, + 448, + 102 + ], + "type": "text", + "content": "SNP: Structured Neuron-level Pruning to Preserve Attention Scores" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 148, + 179, + 463, + 292 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 3: Performance of SNP without fine-tuning across different data quantities at various pruning ratios. To determine the proper number of images for calculating the importance score, we conduct local pruning on the query and key layers at pruning ratios of " + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": ". All performance metrics are assessed without fine-tuning. The latency is measured on Raspberry Pi 4B." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 148, + 179, + 463, + 292 + ], + "lines": [ + { + "bbox": [ + 148, + 179, + 463, + 292 + ], + "spans": [ + { + "bbox": [ + 148, + 179, + 463, + 292 + ], + "type": "table", + "html": "
Number of imagesPerformance by pruning ratio
Original10%30%50%70%90%
172.2071.1567.9657.6038.6811.17
472.2070.8267.8657.6538.0511.86
1672.2070.7267.4258.6039.469.11
6472.2070.8367.5559.5042.1312.83
25672.2070.7568.1459.5842.7014.48
Latency (ms)139.13133.60129.29119.41117.48112.32
Params (M)5.75.595.415.235.064.88
", + "image_path": "4f0dcfe461c149defea78ee450ff1aab445d0961931299b153886e598eac9d79.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 315, + 481, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 315, + 481, + 398 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 481, + 398 + ], + "type": "text", + "content": "Restoring the attention scores after SNP Fig. 4 illustrates the overall and per-head attention maps of the original, compressed, and fine-tuned DeiT-Tiny, respectively. The first row shows the overall attention maps of the respective models. Notably, the compressed model maintains a well-preserved overall attention map, despite pruning all layers, including values and FFN, resulting in a " + }, + { + "bbox": [ + 130, + 315, + 481, + 398 + ], + "type": "inline_equation", + "content": "53\\%" + }, + { + "bbox": [ + 130, + 315, + 481, + 398 + ], + "type": "text", + "content": " reduction in FLOPs and a " + }, + { + "bbox": [ + 130, + 315, + 481, + 398 + ], + "type": "inline_equation", + "content": "46\\%" + }, + { + "bbox": [ + 130, + 315, + 481, + 398 + ], + "type": "text", + "content": " reduction in parameters. Especially, we can observe that attention map is well-restored after the fine-tuning process." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 399, + 481, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 481, + 458 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 481, + 458 + ], + "type": "text", + "content": "The twelve red boxes below the overall attention map depict per-head attention maps for twelve layers of each original, compressed, and fine-tuned models respectively. As shown in Fig. 4, it is evident that the attention maps for each head are effectively preserved and restored in each of the compressed and fine-tuned models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 477, + 244, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 477, + 244, + 488 + ], + "spans": [ + { + "bbox": [ + 131, + 477, + 244, + 488 + ], + "type": "text", + "content": "4.4 Ablation studies" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 498, + 481, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 481, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 481, + 569 + ], + "type": "text", + "content": "Importance scores by the data quantity Since the attention scores depend on the input, as indicated in Eq. (1), the proposed importance scores for query and key filter pairs (Eq. (6)) may exhibit sensitivity to the distribution of the input image " + }, + { + "bbox": [ + 130, + 498, + 481, + 569 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 498, + 481, + 569 + ], + "type": "text", + "content": ". To validate the method's robustness, we compute importance scores using various image quantities, pruning query and key layers at different ratios, without fine-tuning process." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "As depicted in Tab. 3, SNP demonstrates a slight advantage in preserving performance with an increasing number of images, outperforming models compressed with fewer images as the compression ratio rises. However, this improvement comes at the cost of increased computation time for SNP calculations. Considering these factors into account, we opt to use 64 images, which yield the second-best performance among the given number of images. This decision strikes a balance between achieving satisfactory performance and maintaining computational efficiency." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "Shim. et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 196, + 116, + 414, + 301 + ], + "blocks": [ + { + "bbox": [ + 196, + 116, + 414, + 301 + ], + "lines": [ + { + "bbox": [ + 196, + 116, + 414, + 301 + ], + "spans": [ + { + "bbox": [ + 196, + 116, + 414, + 301 + ], + "type": "image", + "image_path": "d1bbf3c5b9cc371fd90ad09fdca666753623755d4200fee638014e2ffaa89677.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 312, + 482, + 357 + ], + "lines": [ + { + "bbox": [ + 130, + 312, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 482, + 357 + ], + "type": "text", + "content": "Fig. 5: Top-1 accuracy of compressed DeiT-Tiny on ImageNet using several pruning criteria without fine-tuning. Query and key layers are locally pruned using various pruning criteria : SNP, GM, l2-norm, reverse order of SNP (\"Reverse\"), and original DeiT-Tiny (\"Baseline\")." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 380, + 481, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 481, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 481, + 416 + ], + "type": "text", + "content": "Performance comparison across pruning ratios To assess the robustness of SNP, we examine the performance of DeiT-Tiny under various pruning criteria and different pruning ratios, as described in Sec. 4.3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 417, + 481, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 417, + 481, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 417, + 481, + 464 + ], + "type": "text", + "content": "As illustrated in Fig. 5, SNP consistently outperforms other pruning criteria across all pruning ratios. In contrast, models compressed with \"Reverse\" criteria exhibit the lowest performance at all pruning ratios, underscoring the robustness of the proposed approach." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 464, + 481, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 464, + 481, + 513 + ], + "spans": [ + { + "bbox": [ + 130, + 464, + 481, + 513 + ], + "type": "text", + "content": "Both Fig. 3 and Fig. 5 confirm that SNP successfully preserves attention scores in both quantitative and qualitative aspects. Even with an " + }, + { + "bbox": [ + 130, + 464, + 481, + 513 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 130, + 464, + 481, + 513 + ], + "type": "text", + "content": " pruning ratio and without fine-tuning, SNP maintains original attention scores and outperforms other pruning criteria." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 532, + 220, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 532, + 220, + 544 + ], + "spans": [ + { + "bbox": [ + 131, + 532, + 220, + 544 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 557, + 481, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 481, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 481, + 641 + ], + "type": "text", + "content": "In this paper, we propose a novel graph-aware neuron-level pruning method, SNP, designed to compress and accelerate Transformer-based models. SNP proposes two pruning criteria for preserving attention scores and eliminating interhead redundancy. Using SNP, a large compressed model outperforms small, hand-crafted designed models in both performance and latency on edge devices. Moreover, the compressed models exhibit astonishing results in latency on various devices, with negligible performance degradation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "As this work is a first attempt to accelerate MSA modules using neuron-level pruning alone, many challenges remain. One is incorporating other pruning" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 171, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 171, + 91, + 447, + 102 + ], + "type": "text", + "content": "SNP: Structured Neuron-level Pruning to Preserve Attention Scores" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "methods such as head or block pruning for a more efficient Transformer model. Another challenge is applying SNP to other vision tasks, including image generation, which requires high computational costs on both training and inference." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 479, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 479, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 479, + 200 + ], + "type": "text", + "content": "We believe that these works encourage the adoption of model pruning as a tool, for both improving the applicability of ViTs in resource-constrained environments and reducing the training costs of large models by integrating with training process." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 217, + 246, + 231 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 217, + 246, + 231 + ], + "spans": [ + { + "bbox": [ + 132, + 217, + 246, + 231 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 240, + 479, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 240, + 479, + 277 + ], + "spans": [ + { + "bbox": [ + 130, + 240, + 479, + 277 + ], + "type": "text", + "content": "This work was supported by Artificial intelligence industrial convergence cluster development project funded by the Ministry of Science and ICT(MSIT, Korea)&Gwangju Metropolitan City." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 294, + 197, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 294, + 197, + 306 + ], + "spans": [ + { + "bbox": [ + 132, + 294, + 197, + 306 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 134, + 317, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 136, + 317, + 480, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 317, + 480, + 339 + ], + "spans": [ + { + "bbox": [ + 136, + 317, + 480, + 339 + ], + "type": "text", + "content": "1. Abnar, S., Zuidema, W.: Quantifying attention flow in transformers. arXiv preprint arXiv:2005.00928 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 136, + 340, + 480, + 383 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 340, + 480, + 383 + ], + "spans": [ + { + "bbox": [ + 136, + 340, + 480, + 383 + ], + "type": "text", + "content": "2. Bian, Y., Huang, J., Cai, X., Yuan, J., Church, K.: On attention redundancy: A comprehensive study. In: Proceedings of the 2021 conference of the north american chapter of the association for computational linguistics: human language technologies. pp. 930-945 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 137, + 384, + 480, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 384, + 480, + 404 + ], + "spans": [ + { + "bbox": [ + 137, + 384, + 480, + 404 + ], + "type": "text", + "content": "3. Bolya, D., Fu, C.Y., Dai, X., Zhang, P., Feichtenhofer, C., Hoffman, J.: Token merging: Your vit but faster. arXiv preprint arXiv:2210.09461 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 405, + 480, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 405, + 480, + 437 + ], + "spans": [ + { + "bbox": [ + 138, + 405, + 480, + 437 + ], + "type": "text", + "content": "4. Chen, T., Cheng, Y., Gan, Z., Yuan, L., Zhang, L., Wang, Z.: Chasing sparsity in vision transformers: An end-to-end exploration. Advances in Neural Information Processing Systems 34, 19974-19988 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 437, + 480, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 437, + 480, + 458 + ], + "spans": [ + { + "bbox": [ + 138, + 437, + 480, + 458 + ], + "type": "text", + "content": "5. Demouth, J.: Sparse matrix-matrix multiplication on thegpu. Tech. rep., NVIDIA (2012)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 459, + 480, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 459, + 480, + 491 + ], + "spans": [ + { + "bbox": [ + 138, + 459, + 480, + 491 + ], + "type": "text", + "content": "6. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 491, + 480, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 491, + 480, + 534 + ], + "spans": [ + { + "bbox": [ + 138, + 491, + 480, + 534 + ], + "type": "text", + "content": "7. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 535, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 535, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 535, + 480, + 567 + ], + "type": "text", + "content": "8. Fang, G., Ma, X., Song, M., Mi, M.B., Wang, X.: Depgraph: Towards any structural pruning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16091-16101 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 568, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 568, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 138, + 568, + 480, + 600 + ], + "type": "text", + "content": "9. Han, S., Mao, H., Dally, W.J.: Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149 (2015)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 601, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 601, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 601, + 480, + 632 + ], + "type": "text", + "content": "10. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "text", + "content": "11. He, Y., Liu, P., Wang, Z., Hu, Z., Yang, Y.: Filter pruning via geometric median for deep convolutional neural networks acceleration. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4340-4349 (2019)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "Shim. et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "12. He, Y., Zhang, X., Sun, J.: Channel pruning for accelerating very deep neural networks. In: Proceedings of the IEEE international conference on computer vision. pp. 1389-1397 (2017)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 172 + ], + "type": "text", + "content": "13. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 194 + ], + "type": "text", + "content": "14. Lee, J., Park, S., Mo, S., Ahn, S., Shin, J.: Layer-adaptive sparsity for the magnitude-based pruning. arXiv preprint arXiv:2010.07611 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 216 + ], + "type": "text", + "content": "15. Li, H., Kadav, A., Durdanovic, I., Samet, H., Graf, H.P.: Pruning filters for efficient convnets. arXiv preprint arXiv:1608.08710 (2016)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "type": "text", + "content": "16. Li, Y., Yuan, G., Wen, Y., Hu, J., Evangelidis, G., Tulyakov, S., Wang, Y., Ren, J.: Efficientformer: Vision transformers at mobilenet speed. Advances in Neural Information Processing Systems 35, 12934-12949 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 248, + 480, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 248, + 480, + 292 + ], + "spans": [ + { + "bbox": [ + 132, + 248, + 480, + 292 + ], + "type": "text", + "content": "17. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "text", + "content": "18. Liu, Z., Wang, Y., Han, K., Zhang, W., Ma, S., Gao, W.: Post-training quantization for vision transformer. Advances in Neural Information Processing Systems 34, 28092-28103 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 325, + 480, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 480, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 480, + 347 + ], + "type": "text", + "content": "19. Michel, P., Levy, O., Neubig, G.: Are sixteen heads really better than one? Advances in neural information processing systems 32 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "text", + "content": "20. Mishra, A., Latorre, J.A., Pool, J., Stosic, D., Stosic, D., Venkatesh, G., Yu, C., Micikevicius, P.: Accelerating sparse deep neural networks. arXiv preprint arXiv:2104.08378 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "text", + "content": "21. Nakajima, S., Sugiyama, M., Babacan, S.D., Tomioka, R.: Global analytic solution of fully-observed variational bayesian matrix factorization. The Journal of Machine Learning Research 14(1), 1-37 (2013)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 413, + 480, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 480, + 446 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 480, + 446 + ], + "type": "text", + "content": "22. Radosavovic, I., Kosaraju, R.P., Girshick, R., He, K., Dólár, P.: Designing network design spaces. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10428-10436 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "type": "text", + "content": "23. Tan, M., Le, Q.: Efficientnet: Rethinking model scaling for convolutional neural networks. In: International conference on machine learning. pp. 6105-6114. PMLR (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 479, + 480, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 512 + ], + "type": "text", + "content": "24. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jégou, H.: Training data-efficient image transformers & distillation through attention. In: International conference on machine learning. pp. 10347-10357. PMLR (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "text", + "content": "25. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 544, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 480, + 578 + ], + "type": "text", + "content": "26. Wan, A., Hao, H., Patnaik, K., Xu, Y., Hadad, O., Güera, D., Ren, Z., Shan, Q.: Upscale: unconstrained channel pruning. In: International Conference on Machine Learning. pp. 35384-35412. PMLR (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 578, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 480, + 599 + ], + "type": "text", + "content": "27. Wang, Z.: Sparsednn: Fast sparse deep learning inference on cpus. arXiv preprint arXiv:2101.07948 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "text", + "content": "28. Xie, Q., Luong, M.T., Hovy, E., Le, Q.V.: Self-training with noisy student improves imagenet classification. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10687-10698 (2020)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "text", + "content": "29. Yu, F., Huang, K., Wang, M., Cheng, Y., Chu, W., Cui, L.: Width & depth pruning for vision transformers. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 3143-3151 (2022)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 171, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 171, + 91, + 447, + 102 + ], + "type": "text", + "content": "SNP: Structured Neuron-level Pruning to Preserve Attention Scores" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 238 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "30. Yu, L., Xiang, W.: X-pruner: explainable pruning for vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24355-24363 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 172 + ], + "type": "text", + "content": "31. Yu, S., Chen, T., Shen, J., Yuan, H., Tan, J., Yang, S., Liu, J., Wang, Z.: Unified visual transformer compression. arXiv preprint arXiv:2203.08243 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "type": "text", + "content": "32. Yu, S., Mazaheri, A., Jannesari, A.: Topology-aware network pruning using multi-stage graph embedding and reinforcement learning. In: International conference on machine learning. pp. 25656-25667. PMLR (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 238 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 238 + ], + "type": "text", + "content": "33. Yu, X., Serra, T., Ramalingam, S., Zhe, S.: The combinatorial brain surgeon: pruning weights that cancel one another in neural networks. In: International Conference on Machine Learning. pp. 25668-25683. PMLR (2022)" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "text", + "content": "Shim. et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/SNeRV_ Spectra-preserving Neural Representation for Video/5d3bbd00-0b38-4046-ad52-4dc33526757f_content_list.json b/2024/SNeRV_ Spectra-preserving Neural Representation for Video/5d3bbd00-0b38-4046-ad52-4dc33526757f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f0e18704b0c9d845c527d64f6b8e44fffc8dcb0a --- /dev/null +++ b/2024/SNeRV_ Spectra-preserving Neural Representation for Video/5d3bbd00-0b38-4046-ad52-4dc33526757f_content_list.json @@ -0,0 +1,1758 @@ +[ + { + "type": "text", + "text": "SNeRV: Spectra-preserving Neural Representation for Video", + "text_level": 1, + "bbox": [ + 300, + 140, + 702, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jina Kim\\* JiHoo Lee\\* and Je-Won Kang\\*\\*", + "bbox": [ + 323, + 212, + 678, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Department of Electronic and Electrical Engineering and Graduate Program in Smart Factory, Ewha Womans University, Seoul, South Korea qapo0106@ewhain.net, jihoo.lee@lge.com, jewonk@ewha.ac.kr", + "bbox": [ + 236, + 239, + 764, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Neural representation for video (NeRV), which employs a neural network to parameterize video signals, introduces a novel methodology in video representations. However, existing NeRV-based methods have difficulty in capturing fine spatial details and motion patterns due to spectral bias, in which a neural network learns high-frequency (HF) components at a slower rate than low-frequency (LF) components. In this paper, we propose spectra-preserving NeRV (SNeRV) as a novel approach to enhance implicit video representations by efficiently handling various frequency components. SNeRV uses 2D discrete wavelet transform (DWT) to decompose video into LF and HF features, preserving spatial structures and directly addressing the spectral bias issue. To balance the compactness, we encode only the LF components, while HF components that include fine textures are generated by a decoder. Specialized modules, including a multi-resolution fusion unit (MFU) and a high-frequency restorer (HFR), are integrated into a backbone to facilitate the representation. Furthermore, we extend SNeRV to effectively capture temporal correlations between adjacent video frames, by casting the extension as additional frequency decomposition to a temporal domain. This approach allows us to embed spatio-temporal LF features into the network, using temporally extended up-sampling blocks (TUBs). Experimental results demonstrate that SNeRV outperforms existing NeRV models in capturing fine details and achieves enhanced reconstruction, making it a promising approach in the field of implicit video representations. The codes are available at https://github.com/qwertja/SNeRV.", + "bbox": [ + 261, + 321, + 743, + 654 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Implicit neural representation $\\cdot$ Neural representation for video $\\cdot$ Wavelet transform", + "bbox": [ + 259, + 667, + 743, + 696 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 723, + 377, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Implicit neural representation for video, also referred to as neural representation for video (NeRV), has attracted considerable attention as a promising method", + "bbox": [ + 212, + 756, + 787, + 787 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "\\*Authors contribute equally.", + "bbox": [ + 215, + 797, + 415, + 811 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "\\*\\* Corresponding author.", + "bbox": [ + 215, + 811, + 383, + 824 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Jihoo Lee is currently with SoC R&D Center, LG Electronics.", + "bbox": [ + 233, + 825, + 643, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9d233f82030302960d1de259d0c259dfc565af22d4ec7691c2c6d635325a24f6.jpg", + "image_caption": [ + "Fig. 1: Visual comparisons of the reconstructed HF coefficients of previous NeRV methods [7,48] and the proposed method in \"Jockey\" sequence. Our model is designed to efficiently encode fine details, by implicitly circumventing the spectral bias problem." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 234 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "for video representations [7,8,14,20,21,24,27,48]. NeRV is used to parameterize a video signal with a neural network, in which a space-time coordinate is used as a query, and the network outputs the corresponding RGB value. As a result, a NeRV model itself can serve as a proxy for the video. This approach has proven effective in various tasks, including video compression [20,21], video interpolation [18,34], and video super-resolution [9,10,26].", + "bbox": [ + 212, + 335, + 782, + 426 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Previous studies use pixel-wise [11, 31, 39] or frame-wise mapping models [8] with various positional embedding of space-time coordinates. However, the quality of the reconstructed frame is severely degraded, when a learned decoder fails to capture the spatial details of diverse video contents. HNeRV [7] utilizes a content-adaptive embedding generated by an encoder in addition to a learned decoder. Temporal correlation among adjacent video frames is exploited in the NeRV pipelines using a dynamical system [48] and a decomposition of visual content and motion information [14]. Nevertheless, capturing video representations remains challenging due to the intricate spatio-temporal dynamics in video scenes, while trying to maintain the size of a model.", + "bbox": [ + 212, + 428, + 787, + 578 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The performance of a NeRV model is significantly affected by the characteristics of an input video [7,24,44,48]. When a NeRV approximates an input video, the optimization is carried out to best represent the input within a budget of a model size. However, inherently, the current methods struggle to accurately capture fine spatial details and temporal patterns [7,14,24,48]. According to recent studies on the neural tangent kernel (NTK) [6,36], the learning rate of a neural network is determined by the magnitude of the eigenvalues of the target function. Because natural videos are mostly composed of low-frequency (LF) components, neural networks tend to learn the LF components more quickly, while the learning of high-frequency (HF) components proceeds at a slower rate. Fig. 1 indicates that the existing NeRVs would face challenges in representing HF components due to spectral bias within the vast redundancies. While these behaviors are evident in the learning of implicit neural representations [47], current NeRV methods have not addressed this problem effectively.", + "bbox": [ + 212, + 580, + 787, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a spectra-preserving NeRV (SNeRV) that enhances the learning of implicit video representations, by efficiently processing different frequency components. First, we present a SNeRV backbone model to achieve", + "bbox": [ + 212, + 794, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "J. Kim et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the goal. We employ a discrete wavelet transform (DWT) to decompose a video into LF and HF features, thereby preserving their spatial structures and directly addressing the issue in the frequency domain. The model capacity is determined with the content-adaptive embeddings and the decoder sizes as [7]. To represent a target in more detail, a decoder requires a larger size of a model or an embedding vector, but it compromises the model compactness. Thus, our encoder embeds only the LF component, exploiting its more suitable characteristics to learn the implicit representations. Meanwhile, the HF components including fine textures are generated by a decoder to circumvent the bias problem. This strategy enables a model to reduce redundancies and achieve a more compact representation with several specialized modules such as a multi-resolution fusion unit (MFU) and a high-frequency restorer (HFR) in a backbone model.", + "bbox": [ + 212, + 146, + 787, + 327 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "On top of the backbone, we extend our model to efficiently represent temporal correlations between adjacent video frames, by casting the temporal modeling as spectral decomposition of the video frames. This is accomplished by performing an additional 1D DWT to a temporal dimension, where the motion is considered as HF components. While the backbone lacks the time analysis by maximising the capability of spectral analysis, the temporal extension considers both time and frequency domains in the analysis. By extending the time analysis, the backbone can capture temporal correlation and efficiently perform temporal modeling, well-suited for related tasks such as video interpolation. In experimental results, the backbone and its temporal extension each exhibit a performance trade-off in video regression and interpolation, while maintaining the same model capacity.", + "bbox": [ + 212, + 329, + 787, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our paper has major contributions as follows:", + "bbox": [ + 238, + 510, + 571, + 523 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose the spectra-preserving NeRV that enhances implicit video representations by efficiently processing different frequency components through specialized modules such as an MFU and a HFR.", + "- We extend the backbone to capture temporal correlations between adjacent video frames, by embedding spatio-temporal LF features into the model.", + "- We evaluate the performance of the proposed methods on various video processing tasks and datasets, demonstrating their effectiveness. Our model outperforms previous approaches, emphasizing the advantages of mitigating the spectral bias problem in the INR for videos." + ], + "bbox": [ + 225, + 537, + 782, + 670 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 699, + 387, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Implicit Neural Representation (INR) for Video. INR for video exploited multi-layer perceptron (MLP) transforming the continuous spaces of high-dimensional data into density or RGB values [11,39]. However, the computational complexity increased substantially, as they calculated a neural function for each coordinate. To address this issue, NeRV [8] proposed a method to use a temporal index with positional encoding (PE) and restore up-scaled spatial information using a NeRV block with a PixelShuffle method [38]. Based on this, E-NeRV [24] sepa", + "bbox": [ + 212, + 733, + 820, + 842 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "SNeRV: Spectra-preserving Neural Representation for Video", + "bbox": [ + 328, + 114, + 730, + 130 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "rated spatial and temporal contexts to improve the representation. PS-NeRV [2] further decomposed a video into a set of patches.", + "bbox": [ + 215, + 145, + 782, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pixel-wise [11,31,39] and frame-wise [2,8] NeRV models exploited no spatial prior of a target frame, which compromised the performance. HNeRV [7] proposed a hybrid model using a content-adaptive frame embedding. The embedding was created in the encoder and jointly optimized with the trained decoder parameters. By conveying the content-adaptive spatial features, it enhanced the quality of reconstructed frames. However, as the parameters were fitted to the current frame, the interpolation performance was simultaneously degraded due to the lack of temporal modeling. For this, the motion information in adjacent frames was exploited [14,21,27,48] to achieve better results.", + "bbox": [ + 215, + 186, + 785, + 324 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Neural Networks in Frequency Domain. Discrete wavelet transform (DWT) decomposes a local region of an image into different frequency levels [28] and proven to be effective in various computer vision tasks [16, 17, 46]. Multi-level space-time wavelet decomposition was employed to extract motion information using a video prediction network, while preserving spatial details [17]. This technique was applied to image super resolution [46], enhanced with wavelet attention embeddings for video super resolution [12]. The DWT typically yields sparse yet significant non-zero coefficients when representing the original signal, a property that was actively used in various directions, such as monocular depth estimation [33] and video interpolation [19].", + "bbox": [ + 215, + 334, + 787, + 486 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Neural Tangent Kernel. Modeling the learning process of a neural network has been actively studied using kernel regression with NTK analysis [1,4,6,36], in which the training rate is decided by the eigenvalues of the NTK matrix. The eigenvalue spectrum of an NTK matrix explains the spectral bias of a neural network towards training speed of HF functions [4]. As over-parameterized neural networks would be poorly suited for learning HF functions [1], mitigating the bias is essential for improving the performance of neural networks [39, 41]. Several studies have attempted to alleviate the bias using various input mapping of a coordinate. Sinusoidal functions were modulated into an activation function [39]. Gabor-wavelets were introduced for flexible representations in space and frequency domains [37]. However, such efforts have not been explored for NeRVs to circumvent the spectral bias. Motivated by the NTK, we aim to develop compact implicit representations of diverse frequency components with a reduced number of parameters. We use the inherent sparsity of the HF wavelet transformed components in videos.", + "bbox": [ + 215, + 496, + 787, + 723 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Video Compression. In recent years, deep learning-based video coding techniques were actively studied as alternatives to traditional CODECs [5,40,45]. This approach involved substituting various coding modules, such as frame restoration [30], motion prediction [13,22], and transform and entropy coding [3], with neural networks. DVC [25] proposed an end-to-end video coding model. The NeRV models [8,20,21,27] converted the compression tasks into model compression, using weight pruning, quantization, and entropy coding.", + "bbox": [ + 215, + 733, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "J. Kim et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9ce82f8f63d7d435fd71dd1ce7075cfb8fc8f26c714795185785758566d35a82.jpg", + "image_caption": [ + "Fig. 2: SNeRV backbone encoder and decoder architectures. The encoder applies 2D DWT to extract LF and HF features and embeds only the LF feature to save parameters. The decoder uses MFU and HFR to efficiently process the LF and HF features. CT and RB refer to transposed convolution and residual blocks, respectively." + ], + "image_footnote": [], + "bbox": [ + 241, + 142, + 754, + 280 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Proposed Method", + "text_level": 1, + "bbox": [ + 215, + 377, + 426, + 395 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Overview", + "text_level": 1, + "bbox": [ + 215, + 407, + 339, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our network is designed to enhance the learning of implicit representations of video by mitigating spectral bias. We use a DWT to decompose video into LF and HF components to use an analysis-and-synthesis approach. The LF features with little variation across frames are embedded through the network, because this is inherently suitable for learning a NeRV model. The HF features are restored by a decoder. Specifically, the decoder sequentially forms the HF features with the LF ones as prior rather than directly learning the HF details. This implicit mapping enables the reallocation of the parameters to concentrate on representing LF features. Both features are optimized to consider the analysis-and-synthesis scheme, thus allowing for a more efficient representation of the HF textures and motions, based upon the stable base provided by the LF contents.", + "bbox": [ + 212, + 431, + 787, + 598 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "These schemes are implemented as our backbone model of the SNeRV. It employs the HNeRV framework [7], in which the implicit representations are composed of a set of embedding vectors and learned decoder parameters. Based on this, we explain the novel design principles of the encoder and decoder, as shown in Fig. 2. On top of that, we extend the backbone to effectively represent temporal contexts of video. We explain the extended modules to accommodate these schemes step-by-step in Fig. 3.", + "bbox": [ + 212, + 598, + 787, + 704 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The differences between the backbone and its temporal extension not only clearly demonstrate their distinct architectural characteristics but also assist the performance analysis in various video processing tasks, while the backbone exhibits its superiority to the conventional methods.", + "bbox": [ + 212, + 704, + 787, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 SNeRV Backbone Architecture", + "text_level": 1, + "bbox": [ + 215, + 785, + 517, + 800 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Encoder Design. Given a video sequence $\\mathbf{V} = \\{I_0, I_1, \\dots, I_{T-1}\\}$ , in the backbone, an encoder conducts a 2D DWT along the $x - y$ spatial axes to decompose", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "SNeRV: Spectra-preserving Neural Representation for Video", + "bbox": [ + 328, + 114, + 732, + 130 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "each video frame $I_{t}\\in \\mathbb{R}^{H\\times W\\times 3}$ into multiple wavelet sub-bands. It produces four sets of wavelet coefficients in $\\mathbb{R}^{H / 2\\times W / 2\\times 3}$ , including $C_{LL}$ , $C_{LH}$ , $C_{HL}$ , and $C_{HH}$ , at the current time-step. $C_{LL}$ contains the LF component, which often appears as a blurred and smoothed version of the original frame. The HF components, represented by $C_{LH}$ , $C_{HL}$ , and $C_{HH}$ , encode the horizontal, vertical, and diagonal details of the image. For simplicity, we utilize the Haar WT [28], where the low pass filter is $\\phi_L = [1 / \\sqrt{2},1 / \\sqrt{2} ]$ , and the high pass filter is $\\phi_H = [1 / \\sqrt{2}, - 1 / \\sqrt{2} ]$", + "bbox": [ + 212, + 145, + 787, + 252 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We input only the $C_{LL}$ into a down-sampling block (DB) to generate a content-adaptive and time-specific embedding $e_t$ whose dimension is $16 \\times 2 \\times 4$ . We use the same DB blocks as [7] and keep the same dimension of $e_t$ for fair comparisons. Due to the 2D DWT, the reduced original input size alleviates the burden of parameter sizes on the decoder side. $C_{LH}$ , $C_{HL}$ , and $C_{HH}$ are not explicitly embedded but synthesized in the decoder.", + "bbox": [ + 212, + 252, + 787, + 343 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Decoder Design. The decoder involves the reconstruction process using $\\mathbf{e} = \\{e_t\\}$ and learned decoder parameters. It consists of three components, including up-sampling blocks (UBs), MFU, and HFR. Firstly, UB is used to progressively increase the resolution of the embedding and reconstruct the spectral information of a target time step. We employ five NeRV blocks [8,38] for the UBs. Then, the MFU fuses features from different resolutions to enrich the LF features. The HFR restores $C_{LH}$ , $C_{HL}$ , and $C_{HH}$ from the reconstructed $C_{LL}$ . The target frame is reconstructed from an inverse DWT (IDWT), by combining all the frequency components. Fig. 2 presents the decoder architecture. We explain the MFU and HFR modules in detail as follows.", + "bbox": [ + 212, + 343, + 787, + 492 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Multi-resolution Fusion Unit (MFU). UBs are used to sequentially reconstruct the LF features and recover the original dimension. However, using only UBs has limits in improving the output quality of $C_{LL}$ . To address this, we develop the MFU with a coarse-to-fine structure, in which the blurred outcomes are progressively refined with multi-scaled features during up-sampling. The MFU comprises two MF blocks (MFBs), each consisting of one transposed convolution and six residual blocks.", + "bbox": [ + 212, + 494, + 787, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In each MFB, the current feature $m_{i}$ is up-sampled using the transposed convolution, and then its concatenation with the outcomes of the $i$ -th UB, denoted by $u_{i}$ , goes through the residual blocks, which is mathematically expressed as,", + "bbox": [ + 212, + 601, + 787, + 646 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nm _ {i + 1} = \\mathcal {F} _ {R B} \\left(\\mathcal {F} _ {C T} \\left(m _ {i}\\right) \\circ u _ {i + 3}\\right), \\quad i = 0, 1, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 676, + 784, + 691 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{F}_{CT}$ and $\\mathcal{F}_{RB}$ are the operations of the transposed convolution and residual blocks, respectively, $\\circ$ is a concatenation, and $m_0 = u_2$ . We use $u_i$ from the last three UBs and two MFBs, empirically considering the trade-off between learning ability and the size of parameters.", + "bbox": [ + 212, + 703, + 787, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "High-Frequency Restorer (HFR). Sparse representations of HF wavelet coefficients have been exploited in wavelet video compression and various computer vision researches [33,35,42,43]. The HF coefficients in the flat regions of an image are nearly zero, whereas significant coefficients are found in the edges, which are essential for achieving a high quality image. Motivated by this, the HFR is", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "J. Kim et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ebfeaa92fc242be4e6215a8340337109603337e6051a925a858ffbacf4ae1c93.jpg", + "image_caption": [ + "Fig. 3: Temporal extension from the backbone: the encoder uses additional 1D DWT to generate spatio-temporal embeddings. The decoder uses TUBs to address the features." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 266 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "designed to restore the high spatial-frequency details from the composite of LF features generated by the MFU. We employ three HFR blocks, with each comprising two convolution layers. These HFR blocks are responsible for generating $C_{LH}$ , $C_{HL}$ , and $C_{HH}$ from the results of $\\mathcal{F}_{MFU}$ . This process is shown in Fig. 2. Because the synthesis process requires less parameters than the direct learning of the HF coefficients, the decoder can save the parameters and allow for more parameters to represent the LF ones.", + "bbox": [ + 212, + 349, + 787, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 SNeRV Temporal Extension", + "text_level": 1, + "bbox": [ + 214, + 479, + 495, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Inspired by [48], we extend the backbone to analyse the effect of considering temporal redundancies. We apply 3D $(2 + 1\\mathrm{D})$ DWT to the image sequence $\\mathbf{V}$ to decompose the LF and HF components along temporal axis in addition to the existing 2D DWT. Specifically, we use two sets of adjacent frames, including $\\{I_{t - 1},I_t\\}$ and $\\{I_t,I_{t + 1}\\}$ as a backward and a forward set, respectively. The 3D DWT is applied to the two image sets, in which 2D DWTs are first applied, and then 1D DWT is applied to the temporal axis. Accordingly, it produces two spatio-temporal LF components corresponding to the sets. Then, we acquire temporal embeddings $e_{\\Delta t_b}$ and $e_{\\Delta t_f}$ through DBs to represent the LF components of the backward and forward set, respectively. In the decoder, $e_{\\Delta t}$ s are jointly up-sampled with $e_t$ . For this, we replace the last three UBs with temporally extended UBs (TUBs), since they are used as inputs of the MFU as in the backbone. The overall architecture with temporal extension is shown in Fig. 3. Temporally Extended Up-sampling Block (TUB). We propose the TUBs to process $e_{\\Delta t_b}$ and $e_{\\Delta t_f}$ in addition to $e_t$ using 3D convolution operations. The three inputs are up-sampled using a transposed convolution and then passed through two 3D convolution layers. The computational complexity of the 3D convolution can be significantly reduced using 2D operations. More details of our implementation can be found in the supplementary. During the consecutive 3D convolutions, the channel dimension increases by a factor of 2 and then reduces back to expand the implicit capacity of the network. In the last TUB, all three components reflecting the different time steps are merged to output the", + "bbox": [ + 212, + 507, + 787, + 843 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "SNeRV: Spectra-preserving Neural Representation for Video", + "bbox": [ + 328, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "LF feature for a target at $t$ . This operation allows to exploit temporal information in adjacent frames, resulting in an enhanced expressive capacity of the network over time. The outputs $C_{LL}, C_{LH}, C_{HL}$ , and $C_{HH}$ for the target at $t$ are obtained by passing the results of the last TUB through the MFU and HFR, as described in Sec. 3.2, obtaining the final output by applying 2D IDWT.", + "bbox": [ + 212, + 146, + 787, + 223 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Implementation", + "text_level": 1, + "bbox": [ + 215, + 244, + 392, + 258 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The total size of the decoder parameters and embedding vectors determine the model capacity for representation. Because the performance would vary with the sizes and the extra embeddings are generated at the expense of the decoder size, we carefully explain the modified parameters to keep the total size. We set the size of $e_t$ to $16 \\times 2 \\times 4$ and use five UBs as in [7] for fair comparisons. We use $C_{LL}$ as input that is recovered to the original dimension using IDWT. Because the IDWT replaces a UB in the last stage, we save some parameters by setting a size of a stride to the half of [7, 48]. This parsimony allows the parameters for MFU and HFR. In the temporal extension, we use a dimension of $3 \\times 20 \\times 40$ for $e_{\\Delta t}$ . Because the size is slightly larger than that of $e_t$ , in our implementation, the channel width of a decoder is reduced to maintain the overall size. We analyze the trade-offs with various temporal embedding sizes in Sec. 4.4. More details such as stride and channel sizes can be found in the supplementary materials.", + "bbox": [ + 212, + 270, + 787, + 467 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.5 Loss Function", + "text_level": 1, + "bbox": [ + 215, + 489, + 375, + 503 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To generate visually pleasing frames with the restored HF features, we combine both L1 and SSIM loss functions over frames, given as", + "bbox": [ + 212, + 513, + 785, + 546 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nL \\left(I _ {t}, \\hat {I} _ {t}\\right) = \\frac {1}{T} \\sum_ {t = 0} ^ {T - 1} \\alpha \\left\\| I _ {t} - \\hat {I} _ {t} \\right\\| _ {1} + (1 - \\alpha) \\left(1 - \\operatorname {S S I M} \\left(I _ {t}, \\hat {I} _ {t}\\right)\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 297, + 556, + 785, + 597 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\hat{I}_t$ is the reconstructed frame, and $I_{t}$ is the ground truth at time step $t$ . SSIM $(\\cdot)$ calculates the SSIM score. $\\alpha$ is set to 0.7 during training.", + "bbox": [ + 212, + 609, + 784, + 640 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Moreover, we compute the loss on a set of the four wavelet coefficients $\\mathbf{C}$ for better frequency reconstruction. Our final loss objective is formulated as", + "bbox": [ + 212, + 641, + 785, + 671 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nL _ {t o t a l} = L (I _ {t}, \\hat {I} _ {t}) + L (\\mathbf {C} _ {t}, \\hat {\\mathbf {C}} _ {t}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 683, + 785, + 702 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{C}}_t$ and $\\mathbf{C}_t$ are the output coefficients and the ground truth, respectively.", + "bbox": [ + 212, + 710, + 784, + 729 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 752, + 375, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Experimental Settings", + "text_level": 1, + "bbox": [ + 215, + 784, + 446, + 799 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct experiments on Bunny, UVG [29], and DAVIS [32] datasets. Bunny has 132 frames of $720 \\times 1280$ resolution. UVG is composed of 7 video sequences", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "J. Kim et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2bec3361464c6ad1d1997492b1d24e7e9025db32467406c18000be467ea04105.jpg", + "table_caption": [ + "Table 1: Performance comparisons (dB) in video regression in UVG and DAVIS datasets with different sizes of resolutions." + ], + "table_footnote": [], + "table_body": "
DatasetUVGDAVIS
sizemethodBeautyBosphBeeJockeyReadyShakeYachtAvg.BikeSwanDanceCamelCarCowsTwirlAvg.
960NeRV [8]33.5633.2837.9331.3126.9532.6728.7432.0625.7727.2225.5923.8228.0921.7525.5125.39
E-NeRV [24]34.1232.9635.4531.9230.1432.9129.2232.3927.1228.7027.8725.6828.1824.2426.8126.94
HNeRV [7]34.1236.3838.9633.5829.9734.1531.4034.0830.7531.7728.8426.5933.5023.8528.5529.12
DNeRV [48]34.0235.3938.7134.4530.4033.5730.8533.9130.0432.6129.5226.9732.9723.6428.8929.23
Ours(B)34.4337.7438.8035.6933.1634.5233.4335.4033.2933.8331.4028.6835.7925.1430.4131.22
Ours(T)34.1437.6439.1136.1733.2134.4132.7135.3431.9033.0831.2630.1434.3327.3829.9331.15
480NeRV [8]34.5935.4238.9034.4330.1631.8233.5034.1228.9132.7728.9530.2632.0824.7727.5729.33
E-NeRV [24]35.0836.6339.2834.6231.0632.4333.9834.7329.0333.3529.2130.9834.9925.7626.9230.04
HNeRV [7]35.6838.7339.7237.1234.3834.2437.6336.7936.0739.4034.6235.2038.6129.8933.3435.30
DNeRV [48]35.3038.4039.5037.5634.1833.8337.1436.5635.6639.2534.2235.1438.3529.0833.2434.99
Ours(B)36.0140.0939.5238.3636.9035.1538.9737.8637.3240.5736.0236.6839.0831.7835.1636.66
Ours(T)35.7840.2439.6138.2136.7734.7038.3037.6636.4038.9435.5936.0438.1732.1134.1235.91
", + "bbox": [ + 215, + 184, + 785, + 343 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "with 300 or 600 frames of $1080 \\times 1920$ . We utilize 20 subsets of DAVIS datasets of $1080 \\times 1920$ sized frames. We used the same center-crop as in [7,48]. For training, we utilized an Adam optimizer and a learning rate of $1 \\times 10^{-3}$ with a cosine learning rate scheduling. The batch size is set to 1.", + "bbox": [ + 212, + 372, + 784, + 431 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We use a peak signal-to-noise ratio (PSNR) and a structural similarity index measure (SSIM) for evaluation and bits-per-pixel (bpp) for compression. If none of the conditions are specified, we use 3M-sized models trained for 300 epochs in the experiments. We compare our backbone, denoted as \"Ours(B)\", and the temporal extension, denoted as \"Ours(T)\", with the state-of-the-art methods such as NeRV [8], E-NeRV [24], HNeRV [7], and DNeRV [48]. We train all the models, utilizing the original codes provided by the authors, and report the testing results under the same conditions. To ensure a fair comparison, we maintained the total capacity of the experimented models. To reduce the original model size from 3.8M to 3M in [48], the channel width was decreased from 92 to 70 which varies depending on the embedding sizes, resulting in a slight drop in performance compared to the original results.", + "bbox": [ + 212, + 431, + 785, + 613 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Performance Evaluation and Analysis", + "text_level": 1, + "bbox": [ + 215, + 635, + 570, + 648 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Video Regression. Tab. 1 presents the performance of the tested methods in the tasks of video regression in UVG and DAVIS datasets with various resolutions. In the comparisons, our method presents the highest performance on average among the tested methods. In UVG dataset of $960 \\times 1920$ resolutions, Ours(B) improves the performance approximately by $1.32\\mathrm{dB} \\sim 1.46\\mathrm{dB}$ over HNeRV and DNeRV, respectively. In DAVIS datasets, our methods outperform the previous methods by $2.17\\mathrm{dB} \\sim 2.35\\mathrm{dB}$ . While we show the results from 7 video samples of DAVIS subsets, more results are reported in the supplementary. We observe similar results in $480 \\times 960$ resolutions. These results imply that our methods significantly improve the performance in video regression.", + "bbox": [ + 212, + 657, + 785, + 808 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Tab. 2 exhibits the performance analysis with different model sizes and training epochs in Bunny dataset. We present the results in Tab. 2a for model sizes", + "bbox": [ + 212, + 810, + 785, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "SNeRV: Spectra-preserving Neural Representation for Video", + "bbox": [ + 328, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "(a) Performance evaluation in PSNR (dB) with different model sizes.", + "bbox": [ + 215, + 189, + 473, + 212 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/976ac843a1c8b25650abe443c4fd87580a30bebee27fa5581c46f7e4482ec0ed.jpg", + "table_caption": [ + "Table 2: Performance comparisons (dB) in video regression on Bunny dataset with different model sizes and training epochs." + ], + "table_footnote": [], + "table_body": "
Model size0.35M0.75M1.5M3.0M
NeRV [8]28.1629.8331.8833.01
E-NeRV [24]27.7829.1231.6536.89
HNeRV [7]30.6732.8135.5738.06
DNeRV [48]29.3631.9735.1837.55
Ours(B)30.8833.2536.7639.64
Ours(T)30.4033.1936.6839.09
", + "bbox": [ + 241, + 224, + 450, + 316 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "(b) Performance evaluation with training epochs and model complexity.", + "bbox": [ + 493, + 189, + 779, + 212 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/fff0aa7e98be4963273f8dcab193ad949c1e04ef4928ba293657aa3d9d8777e9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Epochs300600120018002400MACs
NeRV [8]33.0133.5433.9634.1634.21100.9G
E-NeRV [24]36.8938.6939.7139.9039.96103.3G
HNeRV [7]38.0639.2739.8740.1040.1760.89G
DNeRV [48]37.5538.1738.5538.6838.7448.39G
Ours(B)39.6440.4040.8240.9440.9990.49G
Ours(T)39.0939.9540.4140.6340.68181.8G
", + "bbox": [ + 498, + 224, + 771, + 314 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/7fc2e3a2a1f4d06df82d1c6a8450d19336aa461ae59f5cdf6bb17acda6d61819.jpg", + "table_caption": [ + "Table 3: Video interpolation results on UVG and DAVIS datasets." + ], + "table_footnote": [], + "table_body": "
DatasetUVGDAVIS
MethodBeautyBosphBeeJockeyReadyShakeYachtAvg.BikeSwanDanceCamelCarCowsTwirlAvg.
NeRV [8]27.7429.0534.5018.6217.5428.5623.5425.6516.7019.8818.0317.3817.5517.7216.2817.65
E-NeRV [24]26.4928.3135.5518.0717.2128.2223.1725.2916.7620.0218.2018.2516.7617.2516.8017.72
HNeRV [7]30.6435.5738.6021.0821.8230.7827.3129.4019.2620.7720.3319.0218.5118.8717.5619.19
DNeRV [48]30.7934.0738.1628.6824.7630.2425.9730.3420.3321.7121.5119.3021.0518.1118.9620.14
Ours(B)30.6036.4438.3720.8121.7630.7027.0529.3719.4820.8020.2319.1018.3918.8217.6319.21
Ours(T)31.9236.6338.6329.3727.5231.3528.3231.9620.4722.1921.8722.3721.3121.6419.1221.28
", + "bbox": [ + 215, + 359, + 785, + 457 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "of 0.35M, 0.75M, 1.5M, and 3M, where the different sizes correspond to different bit budgets. The results demonstrate the superiority of our backbone model in various model capacities. Ours(T) also secures the second-best results in most scenarios. Tab. 2b exhibits the performance changes, when the epochs vary from 300 to 2400 with 3M-models, implying different encoding times. Ours(B) and Ours(T) exhibit the best and second-best performance, respectively. We also show the encoding complexity of $640 \\times 1280$ video resolutions compared to other NeRV-based methods in Tab. 2b. Despite the higher complexity involved in processing multi-scale features, Ours(B) excels in video regression, delivering superior results over other methods across all sizes and training epochs.", + "bbox": [ + 212, + 489, + 784, + 640 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Video Interpolation. Tab. 3 displays the comparative studies of video interpolation in UVG and DAVIS datasets. We train the model with odd frames and test results with even frames. We observe Ours(T) and DNeRV that embed the frame differences yield better performance than the tested methods without time embeddings because the time embedding is essential to continuously represent videos over time. Ours(T) achieves the highest performance among the tested methods, which is approximately $1.14\\mathrm{dB} \\sim 1.62\\mathrm{dB}$ better than DNeRV in UVG and DAVIS datasets, respectively. The results indicate that Ours(T) provides efficient video representations over time.", + "bbox": [ + 212, + 642, + 785, + 777 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Video Compression. We compare compression results with commercial video CODECs such as H.264 [45] and H.265 [40] and learning-based CODECs such as DVC [25], FVC [15], and DCVC [23]. Fig. 4 presents the rate-distortion curves. Our method improves coding efficiency over H.264 and all the other NeRV-based", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "J. Kim et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/3eb6da531aa5f2090a67f52e8bd4ffdb93f5a66b2179e96c8208cf5fc5efbcd7.jpg", + "image_caption": [ + "Fig. 4: Coding performance comparisons in UVG datasets." + ], + "image_footnote": [], + "bbox": [ + 282, + 157, + 483, + 277 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/dda63efc83aeb83aa9eae93ee9522eb9089a19ca1ba74aa654a8514867928f4e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 157, + 709, + 277 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/bd3ea3d70f16ff47ab159e7bfb5f44e8edffb4ef07c5bdb55c21427f887db598.jpg", + "table_caption": [ + "Table 4: Video in-painting results on DAVIS dataset." + ], + "table_footnote": [], + "table_body": "
MethodIn-painting1In-painting2Avg.
HNeRV [7]29.7629.4129.59
DNeRV [48]30.0129.8229.92
Ours(B)30.7029.9030.30
", + "bbox": [ + 227, + 372, + 470, + 435 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/fffbd7573413caf65d2d0310c41c3339596e6ef09efb495af3c96a4ea265af35.jpg", + "image_caption": [ + "Fig.5: Visualization of In-painting1: masked input (left) and reconstructed output (right)." + ], + "image_footnote": [], + "bbox": [ + 516, + 330, + 650, + 382 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/ef92037fcb378b18b2d1a3b41690845f120cca190cd9a07d41be25e9b6b4e1b3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 330, + 785, + 382 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "methods and achieves comparable coding gains with H.265, DVC, and FVC at low bit-rates with SSIM results. In the comparisons, we use the same coding pipeline and configurations as in [48]. We adopt $10\\%$ model pruning followed by 6-bit quantization for embeddings and 8-bit quantization for decoder parameters for all tested methods. Our methods do not apply additional techniques such as quantization-aware training (QAT) and predictive coding to improve coding efficiency for fair comparisons. HiNeRV [20] provides improved performance with specialized coding tools for NeRVs at the expense of substantial computational complexity and limited performance in video interpolation. Our method provides more fundamental changes as a backbone, which is orthogonal to the direction. Accordingly, we integrate our model into HiNeRV's coding framework and show the results denoted by Ours(B) + HiNeRV in Fig. 4. It improves the BD-PSNR by 0.3dB compared to HiNeRV. Our supplementary results also demonstrate the generalization ability of our study. In terms of model complexity, Ours(B) has 168G multiply-accumulates (MACs) in UVG dataset. HiNeRV and DNeRV have 111.71G MACs and 69.86G MACs in the same dataset, respectively. These values are significantly smaller than those of neural network based video compression (NVC) methods such as DCVC, which has 2,268G MACs as reported in [23]. Ours(B) has 31.3 decoding fps, which is relatively slower than 48.4 fps of HiNeRV but is significantly faster than 1.75 fps of DCVC, measured in NVIDIA GeForce RTX 3090 GPU. When comparing the results, Ours(B) provides a practically reasonable trade-off between coding performance and computational complexity.", + "bbox": [ + 212, + 458, + 787, + 792 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Video In-painting. We evaluate the performance of a video in-painting task. We use $50 \\times 50$ masks with five fixed points and ten random points in In-painting1 and In-painting2 [48], respectively. Tab. 4 presents that our model provides im", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "SNeRV: Spectra-preserving Neural Representation for Video", + "bbox": [ + 328, + 114, + 730, + 130 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/3546ef29a6e13f62093cb4c8e176fd4a4cda25cd12efe6375f6a6af2c7a568ed.jpg", + "image_caption": [ + "(a) Visual comparisons of tested methods in the tasks of video regression. From the top, Bike, Bosph, and Yacht." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 279 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/068a919300fab2b011d571d3e9608b116b749585fc2c1686c45554aafc4cca51.jpg", + "image_caption": [ + "(b) Visual comparisons of tested methods in the tasks of video interpolation. From the top, Jockey and Ready.", + "Fig. 6: Visual comparisons in video regression and interpolation tasks. From the left, ground-truth (GT), HNeRV, DNeRV, and Ours." + ], + "image_footnote": [], + "bbox": [ + 217, + 303, + 785, + 393 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "proved results approximately $0.38\\mathrm{dB} \\sim 0.71\\mathrm{dB}$ over the existing methods. Fig. 5 shows the visualization results of In-painting1.", + "bbox": [ + 212, + 489, + 784, + 520 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Qualitative Performance Analysis. We provide the qualitative results of the tested methods in Fig. 6. The results illustrate that our methods provide visually pleasing results in video regression and interpolation.", + "bbox": [ + 212, + 521, + 784, + 566 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Discussion", + "text_level": 1, + "bbox": [ + 215, + 595, + 346, + 609 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Trade-off between Video Regression and Time Interpolation. Tab. 1 and Tab. 3 display the video regression and interpolation performances, respectively. According to the results, Ours(B) outperforms Ours(T) in regression, but underperforms in interpolation, indicating a trade-off relationship. We explain this trade-off between time duration $(\\Delta t)$ and spectral band $(\\Delta \\omega)$ in terms of learning aspects. Ours(B) sets $\\Delta t = 0$ , lacking the time analysis but maximizing the spectral analysis, while Ours(T) compromises $\\Delta t$ and $\\Delta \\omega$ . Given the limited model parameters, learning the spatial spectral representation of a current frame could fit a reconstruction task but might compromise the generalization ability across time. In contrast, learning temporal relations can make it challenging to fit the parameters to the current time. Similar characteristics are present not only in Ours(B) and Ours(T) but also in other findings. DNeRV [48] offers superior performance in video interpolation compared to HNeRV [7], but not in video reconstruction. This strategy in our approach signifies a new direction in", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "J. Kim et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/48dec2addea558ab2788c4054e35414187caae35d2b315a8db0f58ac765bbe8b.jpg", + "image_caption": [ + "(a) Low-frequency sults of Jockey." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 359, + 200 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/ae021564a6f47e982f462600072c47998d218b5e337d4abb6f00e61c4e1a5511.jpg", + "image_caption": [ + "(b) High-frequency results of Jockey." + ], + "image_footnote": [], + "bbox": [ + 361, + 143, + 496, + 200 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/b050efaad87d189e4014a0025b577ac65ac91de063d2b880a1cda3a1dcb555e5.jpg", + "image_caption": [ + "Fig. 7: Training rates for different frequency components." + ], + "image_footnote": [], + "bbox": [ + 501, + 143, + 640, + 200 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/541841b39c7bda20f47be1c1d39b061dadb0bb59a2b3dc939d75eed7c7bb29ef.jpg", + "image_caption": [ + "(c) Low-frequency results of Ready.", + "(d) High-frequency results of Ready." + ], + "image_footnote": [], + "bbox": [ + 640, + 143, + 781, + 200 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "analyzing and optimizing NeRV for various video processing tasks from a time and spectra perspective, which has not been conducted before.", + "bbox": [ + 212, + 280, + 784, + 309 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Learning Characteristics of LF and HF Components. Fig. 7 displays the learning characteristics of LF and HF components separately for analysis, while they are trained simultaneously. Jockey and Ready sequences contain fine textures and motions, yielding some HF components that are challenging to represent with an implicit neural network. Fig. 7a and Fig. 7c show the results of the LF at training steps. The LFs are steadily trained to reach the convergence points producing similar PSNR values to the tested methods. Besides, Ours(T) and DNeRV exhibit better performance in the early stage, because they use larger sizes of embeddings. However, Fig. 7b and Fig. 7d show significantly different characteristics in learning HF components. While the previous methods fail to efficiently learn the HFs, our methods achieve significantly improved performance with increasing epochs. These results verify that our model could mitigate the spectral bias and successfully preserve HF components.", + "bbox": [ + 212, + 310, + 785, + 507 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 527, + 387, + 542 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "MFU and HFR Modules. Tab. 5a presents the effectiveness of the MFU and HFR. We conduct ablation tests on the Bunny dataset mostly including LF components and the Yacht sequence in UVG including comparatively larger HF components. The best performance is achieved when both the MFU and HFR are utilized. MFU achieves significantly improved performance approximately $1.20\\mathrm{dB} \\sim 1.88\\mathrm{dB}$ , demonstrating the effectiveness of the module. Using HFR, while Bunny increases the performance by $0.21\\mathrm{dB}$ , Yacht observes $1.62\\mathrm{dB}$ . This confirms that HFR is capable of restoring HF components. Further, the HFR contributes approximately $0.44\\mathrm{dB} \\sim 1.50\\mathrm{dB}$ on top of the MFU. This refinement is particularly advantageous in generating HFs that were hardly trained in previous NeRV models.", + "bbox": [ + 212, + 551, + 787, + 718 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Temporal Up-sampling Block. We evaluate the performance of the TUBs, by replacing them in Ours(T) with the conventional NeRV blocks comprising PixelShuffle and 2D convolution layers. Tab. 5b shows that our TUB is suitable for processing time-extended features than the 2D operations. TUB improves the PSNR around 0.80dB.", + "bbox": [ + 212, + 719, + 787, + 792 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Loss Terms. We conduct ablation tests on the loss function in Eq. (2) using a Yacht sequence. Tab. 5b shows the results, when the terms are turned on or off. Both the terms contribute to the performance improvements.", + "bbox": [ + 212, + 794, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "SNeRV: Spectra-preserving Neural Representation for Video", + "bbox": [ + 328, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/20e4e2bf06afd515c48a6a850a43a25dc95c2c7d71941937c14c76d8e4b1049b.jpg", + "table_caption": [ + "Table 5: Ablation studies of the proposed modules and loss terms.", + "(a) MFU and HFR." + ], + "table_footnote": [], + "table_body": "
DatasetBunnyYacht
MFUHFRPSNRSSIMPSNRSSIM
XX38.000.991130.050.9323
✓X39.200.992431.930.9458
X✓38.210.991131.670.9516
✓✓39.640.992833.430.9647
", + "bbox": [ + 251, + 207, + 486, + 300 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/1b17c48bcf9fe06484b8ae3dc7bf5ea891bc43c7697644c26cb1146a6de6289c.jpg", + "table_caption": [ + "(b) Loss terms and TUB." + ], + "table_footnote": [], + "table_body": "
L(It,ˆt) L(Ct,ˆCt)PSNRSSIM
✓X33.290.9600
✓✓33.430.9647
Temporal BlockPSNRSSIM
NeRV Block31.910.9532
TUB32.710.9584
", + "bbox": [ + 532, + 207, + 722, + 306 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/3221893205be72ddee724241e375b20f85a5f3954d4d21c4279666fd52344df7.jpg", + "image_caption": [ + "(a) Varying temporal embedding size with fixed target embedding." + ], + "image_footnote": [], + "bbox": [ + 243, + 323, + 457, + 397 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/6783c7c233d46a02b71fa1736bc1c64afe5e40f68dabc43531ce97cbbb178022.jpg", + "image_caption": [ + "(b) Varying target embedding size with fixed temporal embedding.", + "Fig. 8: Regression and interpolation results with different embedding sizes." + ], + "image_footnote": [], + "bbox": [ + 540, + 323, + 756, + 396 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Effect of Size of Model Components. We analyze the performance with different sizes of target $e_t$ and temporal embeddings $e_{\\Delta t}$ s, with the size of the decoder varying according to these sizes. Fig. 8 presents the performance variations, exploring the impacts of the ratios. The regression performance is reliable in all the cases, when the ratios are different. In interpolation, the size of the temporal embedding greatly affects the performance, because it hints the temporal relations among video frames.", + "bbox": [ + 212, + 478, + 787, + 585 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 606, + 359, + 622 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we proposed the spectra-preserving neural representation for video as an efficient baseline framework. SNeRV successfully learned LF components, while reconstructing fine HF details within a limited capacity of parameters, and surpassed the existing NeRV-based methods in various video-related tasks. More compact and efficient architecture to process the high-frequency components considering time-frequency trade-off will be done as a future work.", + "bbox": [ + 212, + 636, + 787, + 728 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 215, + 750, + 393, + 767 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This work was partly supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No.RS-2021-II212068 and RS-2022-00167169) and was partly supported by the NRF grant funded by MSIT (No.NRF-2022R1A2C4002052).", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "J. Kim et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 323, + 160 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Arora, S., Du, S., Hu, W., Li, Z., Wang, R.: Fine-grained analysis of optimization and generalization for overparameterized two-layer neural networks. International Conf. on Machine Learning (2019)", + "2. Bai, Y., Dong, C., Wang, C., Yuan, C.: Ps-nerv: Patch-wise stylized neural representations for videos. In: Proceedings of the IEEE conference on Image Processing (2023)", + "3. Balle, J., Minnen, D., Singh, S., Hwang, S.J., Johnston, N.: Variational image compression with a scale hyperprior. arXiv preprint arXiv:1802.01436 (2018)", + "4. Basri, R., Galun, M., Geifman, A., Jacobs, D., Kasten, Y., Kritchman, S.: Frequency bias in neural networks for input of non-uniform density. arXiv preprint arXiv 2003.04560 (2020)", + "5. Bross, B., Wang, Y.K., Ye, Y., Liu, S., Chen, J., Sullivan, G.J., Ohm, J.R.: Overview of the versatile video coding (vvc) standard and its applications. IEEE Transactions on Circuits and Systems for Video Technology 31(10), 3736-3764 (2021)", + "6. Cao, Y., Fang, Z., Wu, Y., Zhou, D.X., Gu, Q.: Towards understanding the spectral bias of deep learning. arXiv preprint arXiv:1912.01198 (2019)", + "7. Chen, H., Gwilliam, M., Lim, S., Shrivastava, A.: Hnerv: A hybrid neural representation for videos. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 10270-10279 (2023)", + "8. Chen, H., He, B., Wang, H., Ren, Y., Lim, S.N., Shrivastava, A.: Nerv: Neural representations for videos. Advances in Neural Information Processing Systems 34, 21557-21568 (2021)", + "9. Chen, Y.H., Chen, S.C., Lin, Y.Y., Peng, W.H.: Motif: Learning motion trajectories with local implicit neural functions for continuous space-time video super-resolution. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 23131-23141 (2023)", + "0. Chen, Z., Chen, Y., Liu, J., Xu, X., Goel, V., Wang, Z., Shi, H., Wang, X.: Videoirr: Learning video implicit neural representation for continuous space-time superresolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2047-2057 (2022)", + "1. Chen, Z., Zhang, H.: Learning implicit fields for generative shape modeling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5939-5948 (2019)", + "2. Choi, Y.J., Lee, Y.W., Kim, B.G.: Wavelet attention embedding networks for video super-resolution. In: 2020 25th International conference on pattern recognition (ICPR). pp. 7314-7320. IEEE (2021)", + "3. Djelouah, A., Campos, J., Schaub-Meyer, S., Schroers, C.: Neural inter-frame compression for video coding. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6421-6429 (2019)", + "4. He, B., Yang, X., Wang, H., Wu, Z., Chen, H., Huang, S., Ren, Y., Lim, S.N., Shrivastava, A.: Towards scalable neural representation for diverse videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6132-6142 (2023)", + "5. Hu, Z., Lu, G., Xu, D.: Fvc: A new framework towards deep video compression in feature space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1502-1511 (2021)" + ], + "bbox": [ + 225, + 178, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "SNeRV: Spectra-preserving Neural Representation for Video", + "bbox": [ + 328, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Huang, H., He, R., Sun, Z., Tan, T.: Wavelet domain generative adversarial network for multi-scale face hallucination. International Journal of Computer Vision 127(6-7), 763-784 (2019)", + "17. Jin, B., Hu, Y., Tang, Q., Niu, J., Shi, Z., Han, Y., Li, X.: Exploring spatial-temporal multi-frequency analysis for high-fidelity and temporal-consistency video prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4554-4563 (2020)", + "18. Jung, H., Hui, Z., Luo, L., Yang, H., Liu, F., Yoo, S., Ranjan, R., Demandolx, D.: Anyflow: Arbitrary scale optical flow with implicit neural representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5455-5465 (2023)", + "19. Kong, L., Jiang, B., Luo, D., Chu, W., Tai, Y., Wang, C., Yang, J.: Dynamic frame interpolation in wavelet domain. IEEE Transactions on Image Processing (2023)", + "20. Kwan, H.M., Gao, G., Zhang, F., Gower, A., Bull, D.: Hinerv: Video compression with hierarchical encoding-based neural representation. Advances in Neural Information Processing Systems 36 (2024)", + "21. Lee, J.C., Rho, D., Ko, J.H., Park, E.: Fferv: Flow-guided frame-wise neural representations for videos. arXiv preprint arXiv:2212.12294 (2022)", + "22. Lee, J.K., Kim, N., Cho, S., Kang, J.W.: Deep video prediction network-based inter-frame coding in hevc. IEEE ACCESS 8, 95906-95917 (2020)", + "23. Li, J., Li, B., Lu, Y.: Deep contextual video compression. Advances in Neural Information Processing Systems 34, 18114-18125 (2021)", + "24. Li, Z., Wang, M., Pi, H., Xu, K., Mei, J., Liu, Y.: E-nerv: Expedite neural video representation with disentangled spatial-temporal context. In: European Conference on Computer Vision. pp. 267-284. Springer (2022)", + "25. Lu, G., Ouyang, W., Xu, D., Zhang, X., Cai, C., Gao, Z.: Dvc: An end-to-end deep video compression framework. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11006-11015 (2019)", + "26. Lu, Y., Wang, Z., Liu, M., Wang, H., Wang, L.: Learning spatial-temporal implicit neural representations for event-guided video super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1557-1567 (2023)", + "27. Maiya, S.R., Girish, S., Ehrlich, M., Wang, H., Lee, K.S., Poirson, P., Wu, P., Wang, C., Shrivastava, A.: Nirvana: Neural implicit representations of videos with adaptive networks and autoregressive patch-wise modeling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14378-14387 (2023)", + "28. Mallat, S.: A wavelet tour of signal processing. Elsevier (1999)", + "29. Mercat, A., Viitanen, M., Vanne, J.: Uvg dataset: 50/120fps 4k sequences for video codec analysis and development. In: Proceedings of the 11th ACM Multimedia Systems Conference. pp. 297-302 (2020)", + "30. Pan, Z., Yi, X., Zhang, Y., Jeon, B., Kwong, S.: Efficient in-loop filtering based on enhanced deep convolutional neural networks for hevc. IEEE Transactions on Image Processing 29, 5352-5366 (2020)", + "31. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 165-174 (2019)", + "32. Perazzi, F., Pont-Tuset, J., McWilliams, B., Gool, L.V., Gross, M., Sorkine-Hornung, A.: A benchmark dataset and evaluation methodology for video object" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "J. Kim et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 724-732 (2016)", + "33. Ramamonjisoa, M., Firman, M., Watson, J., Lepetit, V., Turmukhambetov, D.: Single image depth prediction with wavelet decomposition. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11089-11098 (2021)", + "34. Rho, D., Cho, J., Ko, J.H., Park, E.: Neural residual flow fields for efficient video representations. In: Proceedings of the Asian Conference on Computer Vision. pp. 3447-3463 (2022)", + "35. Rho, D., Lee, B., Nam, S., Lee, J.C., Ko, J.H., Park, E.: Masked wavelet representation for compact neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20680-20690 (2023)", + "36. Ronen, B., Jacobs, D., Kasten, Y., Kritchman, S.: The convergence rate of neural networks for learned functions of different frequencies. Advances in Neural Information Processing Systems 32 (2019)", + "37. Saragadam, V., LeJeune, D., Tan, J., Balakrishnan, G., Veeraraghavan, A., Baraniuk, R.G.: Wire: Wavelet implicit neural representations. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 18507-18516 (2023)", + "38. Shi, W., Caballero, J., Huszar, F., Totz, J., Aitken, A.P., Bishop, R., Wang, Z.: Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1874-1883 (2016)", + "39. Sitzmann, V., Martel, J., Bergman, A., Lindell, D., Wetzstein, G.: Implicit neural representations with periodic activation functions. Advances in neural information processing systems 33, 7462-7473 (2020)", + "40. Sullivan, G.J., Ohm, J.R., Han, W.J., Wiegand, T.: Overview of the high efficiency video coding (hevc) standard. IEEE Transactions on circuits and systems for video technology 22(12), 1649-1668 (2012)", + "41. Tancik, M., Srinivasan, P., Mildenhall, B., Fridovich-Keil, S., Raghavan, N., Singhal, U., Ramamoorthi, R., Barron, B., Ng, R.: Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems 33 (2020)", + "42. Taubman, D.S., Marcellin, M.W., Rabbani, M.: Jpeg2000: Image compression fundamentals, standards and practice. Journal of Electronic Imaging 11(2), 286-287 (2002)", + "43. Unser, M., Blu, T.: Mathematical properties of the jpeg2000 wavelet filters. IEEE transactions on image processing 12(9), 1080-1090 (2003)", + "44. Wang, P., Fan, Z., Chen, T., Wang, Z.: Neural implicit dictionary learning via mixture-of-expert training. In: International Conference on Machine Learning. pp. 22613-22624. PMLR (2022)", + "45. Wiegand, T., Sullivan, G.J., Bjontegaard, G., Luthra, A.: Overview of the h. 264/avc video coding standard. IEEE Transactions on circuits and systems for video technology 13(7), 560-576 (2003)", + "46. Xin, J., Li, J., Jiang, X., Wang, N., Huang, H., Gao, X.: Wavelet-based dual recursive network for image super-resolution. IEEE Transactions on Neural Networks and Learning Systems 33(2), 707-720 (2020)", + "47. Yuce, G., Ortiz-Jimenez, G., Besbinar, B., Frossard, P.: A structured dictionary perspective on implicit neural representations. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 19228-19238 (2022)" + ], + "bbox": [ + 215, + 147, + 785, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "SNeRV: Spectra-preserving Neural Representation for Video", + "bbox": [ + 328, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "ref_text", + "text": "48. Zhao, Q., Asif, M.S., Ma, Z.: Dnerv: Modeling inherent dynamics via difference neural representation for videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2031-2040 (2023)", + "bbox": [ + 217, + 146, + 787, + 189 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "J. Kim et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/SNeRV_ Spectra-preserving Neural Representation for Video/5d3bbd00-0b38-4046-ad52-4dc33526757f_model.json b/2024/SNeRV_ Spectra-preserving Neural Representation for Video/5d3bbd00-0b38-4046-ad52-4dc33526757f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7c98f1b10765f946e8fc7703333163287172c5bd --- /dev/null +++ b/2024/SNeRV_ Spectra-preserving Neural Representation for Video/5d3bbd00-0b38-4046-ad52-4dc33526757f_model.json @@ -0,0 +1,2436 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.302, + 0.141, + 0.703, + 0.187 + ], + "angle": 0, + "content": "SNeRV: Spectra-preserving Neural Representation for Video" + }, + { + "type": "text", + "bbox": [ + 0.325, + 0.213, + 0.679, + 0.229 + ], + "angle": 0, + "content": "Jina Kim\\* JiHoo Lee\\* and Je-Won Kang\\*\\*" + }, + { + "type": "text", + "bbox": [ + 0.238, + 0.24, + 0.766, + 0.283 + ], + "angle": 0, + "content": "Department of Electronic and Electrical Engineering and Graduate Program in Smart Factory, Ewha Womans University, Seoul, South Korea qapo0106@ewhain.net, jihoo.lee@lge.com, jewonk@ewha.ac.kr" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.322, + 0.744, + 0.655 + ], + "angle": 0, + "content": "Abstract. Neural representation for video (NeRV), which employs a neural network to parameterize video signals, introduces a novel methodology in video representations. However, existing NeRV-based methods have difficulty in capturing fine spatial details and motion patterns due to spectral bias, in which a neural network learns high-frequency (HF) components at a slower rate than low-frequency (LF) components. In this paper, we propose spectra-preserving NeRV (SNeRV) as a novel approach to enhance implicit video representations by efficiently handling various frequency components. SNeRV uses 2D discrete wavelet transform (DWT) to decompose video into LF and HF features, preserving spatial structures and directly addressing the spectral bias issue. To balance the compactness, we encode only the LF components, while HF components that include fine textures are generated by a decoder. Specialized modules, including a multi-resolution fusion unit (MFU) and a high-frequency restorer (HFR), are integrated into a backbone to facilitate the representation. Furthermore, we extend SNeRV to effectively capture temporal correlations between adjacent video frames, by casting the extension as additional frequency decomposition to a temporal domain. This approach allows us to embed spatio-temporal LF features into the network, using temporally extended up-sampling blocks (TUBs). Experimental results demonstrate that SNeRV outperforms existing NeRV models in capturing fine details and achieves enhanced reconstruction, making it a promising approach in the field of implicit video representations. The codes are available at https://github.com/qwertja/SNeRV." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.668, + 0.744, + 0.697 + ], + "angle": 0, + "content": "Keywords: Implicit neural representation \\(\\cdot\\) Neural representation for video \\(\\cdot\\) Wavelet transform" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.724, + 0.378, + 0.74 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.757, + 0.788, + 0.789 + ], + "angle": 0, + "content": "Implicit neural representation for video, also referred to as neural representation for video (NeRV), has attracted considerable attention as a promising method" + }, + { + "type": "page_footnote", + "bbox": [ + 0.216, + 0.798, + 0.416, + 0.812 + ], + "angle": 0, + "content": "\\*Authors contribute equally." + }, + { + "type": "page_footnote", + "bbox": [ + 0.217, + 0.813, + 0.385, + 0.825 + ], + "angle": 0, + "content": "\\*\\* Corresponding author." + }, + { + "type": "page_footnote", + "bbox": [ + 0.234, + 0.827, + 0.645, + 0.84 + ], + "angle": 0, + "content": "Jihoo Lee is currently with SoC R&D Center, LG Electronics." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "J. Kim et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.787, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.26, + 0.788, + 0.304 + ], + "angle": 0, + "content": "Fig. 1: Visual comparisons of the reconstructed HF coefficients of previous NeRV methods [7,48] and the proposed method in \"Jockey\" sequence. Our model is designed to efficiently encode fine details, by implicitly circumventing the spectral bias problem." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.336, + 0.784, + 0.427 + ], + "angle": 0, + "content": "for video representations [7,8,14,20,21,24,27,48]. NeRV is used to parameterize a video signal with a neural network, in which a space-time coordinate is used as a query, and the network outputs the corresponding RGB value. As a result, a NeRV model itself can serve as a proxy for the video. This approach has proven effective in various tasks, including video compression [20,21], video interpolation [18,34], and video super-resolution [9,10,26]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.429, + 0.789, + 0.579 + ], + "angle": 0, + "content": "Previous studies use pixel-wise [11, 31, 39] or frame-wise mapping models [8] with various positional embedding of space-time coordinates. However, the quality of the reconstructed frame is severely degraded, when a learned decoder fails to capture the spatial details of diverse video contents. HNeRV [7] utilizes a content-adaptive embedding generated by an encoder in addition to a learned decoder. Temporal correlation among adjacent video frames is exploited in the NeRV pipelines using a dynamical system [48] and a decomposition of visual content and motion information [14]. Nevertheless, capturing video representations remains challenging due to the intricate spatio-temporal dynamics in video scenes, while trying to maintain the size of a model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.582, + 0.789, + 0.793 + ], + "angle": 0, + "content": "The performance of a NeRV model is significantly affected by the characteristics of an input video [7,24,44,48]. When a NeRV approximates an input video, the optimization is carried out to best represent the input within a budget of a model size. However, inherently, the current methods struggle to accurately capture fine spatial details and temporal patterns [7,14,24,48]. According to recent studies on the neural tangent kernel (NTK) [6,36], the learning rate of a neural network is determined by the magnitude of the eigenvalues of the target function. Because natural videos are mostly composed of low-frequency (LF) components, neural networks tend to learn the LF components more quickly, while the learning of high-frequency (HF) components proceeds at a slower rate. Fig. 1 indicates that the existing NeRVs would face challenges in representing HF components due to spectral bias within the vast redundancies. While these behaviors are evident in the learning of implicit neural representations [47], current NeRV methods have not addressed this problem effectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.789, + 0.84 + ], + "angle": 0, + "content": "In this paper, we propose a spectra-preserving NeRV (SNeRV) that enhances the learning of implicit video representations, by efficiently processing different frequency components. First, we present a SNeRV backbone model to achieve" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.33, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "SNeRV: Spectra-preserving Neural Representation for Video" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.328 + ], + "angle": 0, + "content": "the goal. We employ a discrete wavelet transform (DWT) to decompose a video into LF and HF features, thereby preserving their spatial structures and directly addressing the issue in the frequency domain. The model capacity is determined with the content-adaptive embeddings and the decoder sizes as [7]. To represent a target in more detail, a decoder requires a larger size of a model or an embedding vector, but it compromises the model compactness. Thus, our encoder embeds only the LF component, exploiting its more suitable characteristics to learn the implicit representations. Meanwhile, the HF components including fine textures are generated by a decoder to circumvent the bias problem. This strategy enables a model to reduce redundancies and achieve a more compact representation with several specialized modules such as a multi-resolution fusion unit (MFU) and a high-frequency restorer (HFR) in a backbone model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.33, + 0.788, + 0.51 + ], + "angle": 0, + "content": "On top of the backbone, we extend our model to efficiently represent temporal correlations between adjacent video frames, by casting the temporal modeling as spectral decomposition of the video frames. This is accomplished by performing an additional 1D DWT to a temporal dimension, where the motion is considered as HF components. While the backbone lacks the time analysis by maximising the capability of spectral analysis, the temporal extension considers both time and frequency domains in the analysis. By extending the time analysis, the backbone can capture temporal correlation and efficiently perform temporal modeling, well-suited for related tasks such as video interpolation. In experimental results, the backbone and its temporal extension each exhibit a performance trade-off in video regression and interpolation, while maintaining the same model capacity." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.511, + 0.573, + 0.525 + ], + "angle": 0, + "content": "Our paper has major contributions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.538, + 0.784, + 0.58 + ], + "angle": 0, + "content": "- We propose the spectra-preserving NeRV that enhances implicit video representations by efficiently processing different frequency components through specialized modules such as an MFU and a HFR." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.584, + 0.784, + 0.61 + ], + "angle": 0, + "content": "- We extend the backbone to capture temporal correlations between adjacent video frames, by embedding spatio-temporal LF features into the model." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.615, + 0.784, + 0.671 + ], + "angle": 0, + "content": "- We evaluate the performance of the proposed methods on various video processing tasks and datasets, demonstrating their effectiveness. Our model outperforms previous approaches, emphasizing the advantages of mitigating the spectral bias problem in the INR for videos." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.538, + 0.784, + 0.671 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.7, + 0.388, + 0.716 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.821, + 0.843 + ], + "angle": 0, + "content": "Implicit Neural Representation (INR) for Video. INR for video exploited multi-layer perceptron (MLP) transforming the continuous spaces of high-dimensional data into density or RGB values [11,39]. However, the computational complexity increased substantially, as they calculated a neural function for each coordinate. To address this issue, NeRV [8] proposed a method to use a temporal index with positional encoding (PE) and restore up-scaled spatial information using a NeRV block with a PixelShuffle method [38]. Based on this, E-NeRV [24] sepa" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "J. Kim et al." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.146, + 0.784, + 0.178 + ], + "angle": 0, + "content": "rated spatial and temporal contexts to improve the representation. PS-NeRV [2] further decomposed a video into a set of patches." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.188, + 0.787, + 0.325 + ], + "angle": 0, + "content": "Pixel-wise [11,31,39] and frame-wise [2,8] NeRV models exploited no spatial prior of a target frame, which compromised the performance. HNeRV [7] proposed a hybrid model using a content-adaptive frame embedding. The embedding was created in the encoder and jointly optimized with the trained decoder parameters. By conveying the content-adaptive spatial features, it enhanced the quality of reconstructed frames. However, as the parameters were fitted to the current frame, the interpolation performance was simultaneously degraded due to the lack of temporal modeling. For this, the motion information in adjacent frames was exploited [14,21,27,48] to achieve better results." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.335, + 0.788, + 0.487 + ], + "angle": 0, + "content": "Neural Networks in Frequency Domain. Discrete wavelet transform (DWT) decomposes a local region of an image into different frequency levels [28] and proven to be effective in various computer vision tasks [16, 17, 46]. Multi-level space-time wavelet decomposition was employed to extract motion information using a video prediction network, while preserving spatial details [17]. This technique was applied to image super resolution [46], enhanced with wavelet attention embeddings for video super resolution [12]. The DWT typically yields sparse yet significant non-zero coefficients when representing the original signal, a property that was actively used in various directions, such as monocular depth estimation [33] and video interpolation [19]." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.497, + 0.788, + 0.724 + ], + "angle": 0, + "content": "Neural Tangent Kernel. Modeling the learning process of a neural network has been actively studied using kernel regression with NTK analysis [1,4,6,36], in which the training rate is decided by the eigenvalues of the NTK matrix. The eigenvalue spectrum of an NTK matrix explains the spectral bias of a neural network towards training speed of HF functions [4]. As over-parameterized neural networks would be poorly suited for learning HF functions [1], mitigating the bias is essential for improving the performance of neural networks [39, 41]. Several studies have attempted to alleviate the bias using various input mapping of a coordinate. Sinusoidal functions were modulated into an activation function [39]. Gabor-wavelets were introduced for flexible representations in space and frequency domains [37]. However, such efforts have not been explored for NeRVs to circumvent the spectral bias. Motivated by the NTK, we aim to develop compact implicit representations of diverse frequency components with a reduced number of parameters. We use the inherent sparsity of the HF wavelet transformed components in videos." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Video Compression. In recent years, deep learning-based video coding techniques were actively studied as alternatives to traditional CODECs [5,40,45]. This approach involved substituting various coding modules, such as frame restoration [30], motion prediction [13,22], and transform and entropy coding [3], with neural networks. DVC [25] proposed an end-to-end video coding model. The NeRV models [8,20,21,27] converted the compression tasks into model compression, using weight pruning, quantization, and entropy coding." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.33, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "SNeRV: Spectra-preserving Neural Representation for Video" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.243, + 0.143, + 0.756, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.295, + 0.788, + 0.352 + ], + "angle": 0, + "content": "Fig. 2: SNeRV backbone encoder and decoder architectures. The encoder applies 2D DWT to extract LF and HF features and embeds only the LF feature to save parameters. The decoder uses MFU and HFR to efficiently process the LF and HF features. CT and RB refer to transposed convolution and residual blocks, respectively." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.378, + 0.427, + 0.396 + ], + "angle": 0, + "content": "3 Proposed Method" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.409, + 0.341, + 0.423 + ], + "angle": 0, + "content": "3.1 Overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.433, + 0.788, + 0.599 + ], + "angle": 0, + "content": "Our network is designed to enhance the learning of implicit representations of video by mitigating spectral bias. We use a DWT to decompose video into LF and HF components to use an analysis-and-synthesis approach. The LF features with little variation across frames are embedded through the network, because this is inherently suitable for learning a NeRV model. The HF features are restored by a decoder. Specifically, the decoder sequentially forms the HF features with the LF ones as prior rather than directly learning the HF details. This implicit mapping enables the reallocation of the parameters to concentrate on representing LF features. Both features are optimized to consider the analysis-and-synthesis scheme, thus allowing for a more efficient representation of the HF textures and motions, based upon the stable base provided by the LF contents." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.788, + 0.705 + ], + "angle": 0, + "content": "These schemes are implemented as our backbone model of the SNeRV. It employs the HNeRV framework [7], in which the implicit representations are composed of a set of embedding vectors and learned decoder parameters. Based on this, we explain the novel design principles of the encoder and decoder, as shown in Fig. 2. On top of that, we extend the backbone to effectively represent temporal contexts of video. We explain the extended modules to accommodate these schemes step-by-step in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.765 + ], + "angle": 0, + "content": "The differences between the backbone and its temporal extension not only clearly demonstrate their distinct architectural characteristics but also assist the performance analysis in various video processing tasks, while the backbone exhibits its superiority to the conventional methods." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.786, + 0.519, + 0.801 + ], + "angle": 0, + "content": "3.2 SNeRV Backbone Architecture" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Encoder Design. Given a video sequence \\(\\mathbf{V} = \\{I_0, I_1, \\dots, I_{T-1}\\}\\), in the backbone, an encoder conducts a 2D DWT along the \\(x - y\\) spatial axes to decompose" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "J. Kim et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.253 + ], + "angle": 0, + "content": "each video frame \\(I_{t}\\in \\mathbb{R}^{H\\times W\\times 3}\\) into multiple wavelet sub-bands. It produces four sets of wavelet coefficients in \\(\\mathbb{R}^{H / 2\\times W / 2\\times 3}\\), including \\(C_{LL}\\), \\(C_{LH}\\), \\(C_{HL}\\), and \\(C_{HH}\\), at the current time-step. \\(C_{LL}\\) contains the LF component, which often appears as a blurred and smoothed version of the original frame. The HF components, represented by \\(C_{LH}\\), \\(C_{HL}\\), and \\(C_{HH}\\), encode the horizontal, vertical, and diagonal details of the image. For simplicity, we utilize the Haar WT [28], where the low pass filter is \\(\\phi_L = [1 / \\sqrt{2},1 / \\sqrt{2} ]\\), and the high pass filter is \\(\\phi_H = [1 / \\sqrt{2}, - 1 / \\sqrt{2} ]\\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.788, + 0.344 + ], + "angle": 0, + "content": "We input only the \\( C_{LL} \\) into a down-sampling block (DB) to generate a content-adaptive and time-specific embedding \\( e_t \\) whose dimension is \\( 16 \\times 2 \\times 4 \\). We use the same DB blocks as [7] and keep the same dimension of \\( e_t \\) for fair comparisons. Due to the 2D DWT, the reduced original input size alleviates the burden of parameter sizes on the decoder side. \\( C_{LH} \\), \\( C_{HL} \\), and \\( C_{HH} \\) are not explicitly embedded but synthesized in the decoder." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.344, + 0.788, + 0.493 + ], + "angle": 0, + "content": "Decoder Design. The decoder involves the reconstruction process using \\(\\mathbf{e} = \\{e_t\\}\\) and learned decoder parameters. It consists of three components, including up-sampling blocks (UBs), MFU, and HFR. Firstly, UB is used to progressively increase the resolution of the embedding and reconstruct the spectral information of a target time step. We employ five NeRV blocks [8,38] for the UBs. Then, the MFU fuses features from different resolutions to enrich the LF features. The HFR restores \\(C_{LH}\\), \\(C_{HL}\\), and \\(C_{HH}\\) from the reconstructed \\(C_{LL}\\). The target frame is reconstructed from an inverse DWT (IDWT), by combining all the frequency components. Fig. 2 presents the decoder architecture. We explain the MFU and HFR modules in detail as follows." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.495, + 0.788, + 0.6 + ], + "angle": 0, + "content": "Multi-resolution Fusion Unit (MFU). UBs are used to sequentially reconstruct the LF features and recover the original dimension. However, using only UBs has limits in improving the output quality of \\( C_{LL} \\). To address this, we develop the MFU with a coarse-to-fine structure, in which the blurred outcomes are progressively refined with multi-scaled features during up-sampling. The MFU comprises two MF blocks (MFBs), each consisting of one transposed convolution and six residual blocks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.602, + 0.788, + 0.647 + ], + "angle": 0, + "content": "In each MFB, the current feature \\( m_{i} \\) is up-sampled using the transposed convolution, and then its concatenation with the outcomes of the \\( i \\)-th UB, denoted by \\( u_{i} \\), goes through the residual blocks, which is mathematically expressed as," + }, + { + "type": "equation", + "bbox": [ + 0.354, + 0.677, + 0.785, + 0.692 + ], + "angle": 0, + "content": "\\[\nm _ {i + 1} = \\mathcal {F} _ {R B} \\left(\\mathcal {F} _ {C T} \\left(m _ {i}\\right) \\circ u _ {i + 3}\\right), \\quad i = 0, 1, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.763 + ], + "angle": 0, + "content": "where \\(\\mathcal{F}_{CT}\\) and \\(\\mathcal{F}_{RB}\\) are the operations of the transposed convolution and residual blocks, respectively, \\(\\circ\\) is a concatenation, and \\(m_0 = u_2\\). We use \\(u_i\\) from the last three UBs and two MFBs, empirically considering the trade-off between learning ability and the size of parameters." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "High-Frequency Restorer (HFR). Sparse representations of HF wavelet coefficients have been exploited in wavelet video compression and various computer vision researches [33,35,42,43]. The HF coefficients in the flat regions of an image are nearly zero, whereas significant coefficients are found in the edges, which are essential for achieving a high quality image. Motivated by this, the HFR is" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.33, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "SNeRV: Spectra-preserving Neural Representation for Video" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.787, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.291, + 0.788, + 0.32 + ], + "angle": 0, + "content": "Fig. 3: Temporal extension from the backbone: the encoder uses additional 1D DWT to generate spatio-temporal embeddings. The decoder uses TUBs to address the features." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.35, + 0.788, + 0.457 + ], + "angle": 0, + "content": "designed to restore the high spatial-frequency details from the composite of LF features generated by the MFU. We employ three HFR blocks, with each comprising two convolution layers. These HFR blocks are responsible for generating \\( C_{LH} \\), \\( C_{HL} \\), and \\( C_{HH} \\) from the results of \\( \\mathcal{F}_{MFU} \\). This process is shown in Fig. 2. Because the synthesis process requires less parameters than the direct learning of the HF coefficients, the decoder can save the parameters and allow for more parameters to represent the LF ones." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.481, + 0.496, + 0.496 + ], + "angle": 0, + "content": "3.3 SNeRV Temporal Extension" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.789, + 0.844 + ], + "angle": 0, + "content": "Inspired by [48], we extend the backbone to analyse the effect of considering temporal redundancies. We apply 3D \\((2 + 1\\mathrm{D})\\) DWT to the image sequence \\(\\mathbf{V}\\) to decompose the LF and HF components along temporal axis in addition to the existing 2D DWT. Specifically, we use two sets of adjacent frames, including \\(\\{I_{t - 1},I_t\\}\\) and \\(\\{I_t,I_{t + 1}\\}\\) as a backward and a forward set, respectively. The 3D DWT is applied to the two image sets, in which 2D DWTs are first applied, and then 1D DWT is applied to the temporal axis. Accordingly, it produces two spatio-temporal LF components corresponding to the sets. Then, we acquire temporal embeddings \\(e_{\\Delta t_b}\\) and \\(e_{\\Delta t_f}\\) through DBs to represent the LF components of the backward and forward set, respectively. In the decoder, \\(e_{\\Delta t}\\)s are jointly up-sampled with \\(e_t\\). For this, we replace the last three UBs with temporally extended UBs (TUBs), since they are used as inputs of the MFU as in the backbone. The overall architecture with temporal extension is shown in Fig. 3. Temporally Extended Up-sampling Block (TUB). We propose the TUBs to process \\(e_{\\Delta t_b}\\) and \\(e_{\\Delta t_f}\\) in addition to \\(e_t\\) using 3D convolution operations. The three inputs are up-sampled using a transposed convolution and then passed through two 3D convolution layers. The computational complexity of the 3D convolution can be significantly reduced using 2D operations. More details of our implementation can be found in the supplementary. During the consecutive 3D convolutions, the channel dimension increases by a factor of 2 and then reduces back to expand the implicit capacity of the network. In the last TUB, all three components reflecting the different time steps are merged to output the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "J. Kim et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "LF feature for a target at \\( t \\). This operation allows to exploit temporal information in adjacent frames, resulting in an enhanced expressive capacity of the network over time. The outputs \\( C_{LL}, C_{LH}, C_{HL} \\), and \\( C_{HH} \\) for the target at \\( t \\) are obtained by passing the results of the last TUB through the MFU and HFR, as described in Sec. 3.2, obtaining the final output by applying 2D IDWT." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.246, + 0.393, + 0.26 + ], + "angle": 0, + "content": "3.4 Implementation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.271, + 0.789, + 0.468 + ], + "angle": 0, + "content": "The total size of the decoder parameters and embedding vectors determine the model capacity for representation. Because the performance would vary with the sizes and the extra embeddings are generated at the expense of the decoder size, we carefully explain the modified parameters to keep the total size. We set the size of \\( e_t \\) to \\( 16 \\times 2 \\times 4 \\) and use five UBs as in [7] for fair comparisons. We use \\( C_{LL} \\) as input that is recovered to the original dimension using IDWT. Because the IDWT replaces a UB in the last stage, we save some parameters by setting a size of a stride to the half of [7, 48]. This parsimony allows the parameters for MFU and HFR. In the temporal extension, we use a dimension of \\( 3 \\times 20 \\times 40 \\) for \\( e_{\\Delta t} \\). Because the size is slightly larger than that of \\( e_t \\), in our implementation, the channel width of a decoder is reduced to maintain the overall size. We analyze the trade-offs with various temporal embedding sizes in Sec. 4.4. More details such as stride and channel sizes can be found in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.49, + 0.377, + 0.504 + ], + "angle": 0, + "content": "3.5 Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.515, + 0.787, + 0.547 + ], + "angle": 0, + "content": "To generate visually pleasing frames with the restored HF features, we combine both L1 and SSIM loss functions over frames, given as" + }, + { + "type": "equation", + "bbox": [ + 0.299, + 0.558, + 0.786, + 0.598 + ], + "angle": 0, + "content": "\\[\nL \\left(I _ {t}, \\hat {I} _ {t}\\right) = \\frac {1}{T} \\sum_ {t = 0} ^ {T - 1} \\alpha \\left\\| I _ {t} - \\hat {I} _ {t} \\right\\| _ {1} + (1 - \\alpha) \\left(1 - \\operatorname {S S I M} \\left(I _ {t}, \\hat {I} _ {t}\\right)\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.611, + 0.785, + 0.641 + ], + "angle": 0, + "content": "where \\(\\hat{I}_t\\) is the reconstructed frame, and \\(I_{t}\\) is the ground truth at time step \\(t\\). SSIM \\((\\cdot)\\) calculates the SSIM score. \\(\\alpha\\) is set to 0.7 during training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.642, + 0.787, + 0.672 + ], + "angle": 0, + "content": "Moreover, we compute the loss on a set of the four wavelet coefficients \\(\\mathbf{C}\\) for better frequency reconstruction. Our final loss objective is formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.397, + 0.684, + 0.786, + 0.703 + ], + "angle": 0, + "content": "\\[\nL _ {t o t a l} = L (I _ {t}, \\hat {I} _ {t}) + L (\\mathbf {C} _ {t}, \\hat {\\mathbf {C}} _ {t}), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.712, + 0.785, + 0.73 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{C}}_t\\) and \\(\\mathbf{C}_t\\) are the output coefficients and the ground truth, respectively." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.753, + 0.377, + 0.77 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.785, + 0.447, + 0.8 + ], + "angle": 0, + "content": "4.1 Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.841 + ], + "angle": 0, + "content": "We conduct experiments on Bunny, UVG [29], and DAVIS [32] datasets. Bunny has 132 frames of \\(720 \\times 1280\\) resolution. UVG is composed of 7 video sequences" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.33, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "SNeRV: Spectra-preserving Neural Representation for Video" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.785, + 0.173 + ], + "angle": 0, + "content": "Table 1: Performance comparisons (dB) in video regression in UVG and DAVIS datasets with different sizes of resolutions." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.185, + 0.787, + 0.344 + ], + "angle": 0, + "content": "
DatasetUVGDAVIS
sizemethodBeautyBosphBeeJockeyReadyShakeYachtAvg.BikeSwanDanceCamelCarCowsTwirlAvg.
960NeRV [8]33.5633.2837.9331.3126.9532.6728.7432.0625.7727.2225.5923.8228.0921.7525.5125.39
E-NeRV [24]34.1232.9635.4531.9230.1432.9129.2232.3927.1228.7027.8725.6828.1824.2426.8126.94
HNeRV [7]34.1236.3838.9633.5829.9734.1531.4034.0830.7531.7728.8426.5933.5023.8528.5529.12
DNeRV [48]34.0235.3938.7134.4530.4033.5730.8533.9130.0432.6129.5226.9732.9723.6428.8929.23
Ours(B)34.4337.7438.8035.6933.1634.5233.4335.4033.2933.8331.4028.6835.7925.1430.4131.22
Ours(T)34.1437.6439.1136.1733.2134.4132.7135.3431.9033.0831.2630.1434.3327.3829.9331.15
480NeRV [8]34.5935.4238.9034.4330.1631.8233.5034.1228.9132.7728.9530.2632.0824.7727.5729.33
E-NeRV [24]35.0836.6339.2834.6231.0632.4333.9834.7329.0333.3529.2130.9834.9925.7626.9230.04
HNeRV [7]35.6838.7339.7237.1234.3834.2437.6336.7936.0739.4034.6235.2038.6129.8933.3435.30
DNeRV [48]35.3038.4039.5037.5634.1833.8337.1436.5635.6639.2534.2235.1438.3529.0833.2434.99
Ours(B)36.0140.0939.5238.3636.9035.1538.9737.8637.3240.5736.0236.6839.0831.7835.1636.66
Ours(T)35.7840.2439.6138.2136.7734.7038.3037.6636.4038.9435.5936.0438.1732.1134.1235.91
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.373, + 0.785, + 0.432 + ], + "angle": 0, + "content": "with 300 or 600 frames of \\(1080 \\times 1920\\). We utilize 20 subsets of DAVIS datasets of \\(1080 \\times 1920\\) sized frames. We used the same center-crop as in [7,48]. For training, we utilized an Adam optimizer and a learning rate of \\(1 \\times 10^{-3}\\) with a cosine learning rate scheduling. The batch size is set to 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.433, + 0.787, + 0.614 + ], + "angle": 0, + "content": "We use a peak signal-to-noise ratio (PSNR) and a structural similarity index measure (SSIM) for evaluation and bits-per-pixel (bpp) for compression. If none of the conditions are specified, we use 3M-sized models trained for 300 epochs in the experiments. We compare our backbone, denoted as \"Ours(B)\", and the temporal extension, denoted as \"Ours(T)\", with the state-of-the-art methods such as NeRV [8], E-NeRV [24], HNeRV [7], and DNeRV [48]. We train all the models, utilizing the original codes provided by the authors, and report the testing results under the same conditions. To ensure a fair comparison, we maintained the total capacity of the experimented models. To reduce the original model size from 3.8M to 3M in [48], the channel width was decreased from 92 to 70 which varies depending on the embedding sizes, resulting in a slight drop in performance compared to the original results." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.636, + 0.571, + 0.65 + ], + "angle": 0, + "content": "4.2 Performance Evaluation and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.787, + 0.809 + ], + "angle": 0, + "content": "Video Regression. Tab. 1 presents the performance of the tested methods in the tasks of video regression in UVG and DAVIS datasets with various resolutions. In the comparisons, our method presents the highest performance on average among the tested methods. In UVG dataset of \\(960 \\times 1920\\) resolutions, Ours(B) improves the performance approximately by \\(1.32\\mathrm{dB} \\sim 1.46\\mathrm{dB}\\) over HNeRV and DNeRV, respectively. In DAVIS datasets, our methods outperform the previous methods by \\(2.17\\mathrm{dB} \\sim 2.35\\mathrm{dB}\\). While we show the results from 7 video samples of DAVIS subsets, more results are reported in the supplementary. We observe similar results in \\(480 \\times 960\\) resolutions. These results imply that our methods significantly improve the performance in video regression." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.811, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Tab. 2 exhibits the performance analysis with different model sizes and training epochs in Bunny dataset. We present the results in Tab. 2a for model sizes" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "J. Kim et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.784, + 0.174 + ], + "angle": 0, + "content": "Table 2: Performance comparisons (dB) in video regression on Bunny dataset with different model sizes and training epochs." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.19, + 0.475, + 0.213 + ], + "angle": 0, + "content": "(a) Performance evaluation in PSNR (dB) with different model sizes." + }, + { + "type": "table", + "bbox": [ + 0.242, + 0.225, + 0.451, + 0.317 + ], + "angle": 0, + "content": "
Model size0.35M0.75M1.5M3.0M
NeRV [8]28.1629.8331.8833.01
E-NeRV [24]27.7829.1231.6536.89
HNeRV [7]30.6732.8135.5738.06
DNeRV [48]29.3631.9735.1837.55
Ours(B)30.8833.2536.7639.64
Ours(T)30.4033.1936.6839.09
" + }, + { + "type": "text", + "bbox": [ + 0.495, + 0.19, + 0.781, + 0.213 + ], + "angle": 0, + "content": "(b) Performance evaluation with training epochs and model complexity." + }, + { + "type": "table", + "bbox": [ + 0.499, + 0.225, + 0.772, + 0.315 + ], + "angle": 0, + "content": "
Epochs300600120018002400MACs
NeRV [8]33.0133.5433.9634.1634.21100.9G
E-NeRV [24]36.8938.6939.7139.9039.96103.3G
HNeRV [7]38.0639.2739.8740.1040.1760.89G
DNeRV [48]37.5538.1738.5538.6838.7448.39G
Ours(B)39.6440.4040.8240.9440.9990.49G
Ours(T)39.0939.9540.4140.6340.68181.8G
" + }, + { + "type": "table_caption", + "bbox": [ + 0.276, + 0.333, + 0.725, + 0.347 + ], + "angle": 0, + "content": "Table 3: Video interpolation results on UVG and DAVIS datasets." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.36, + 0.787, + 0.458 + ], + "angle": 0, + "content": "
DatasetUVGDAVIS
MethodBeautyBosphBeeJockeyReadyShakeYachtAvg.BikeSwanDanceCamelCarCowsTwirlAvg.
NeRV [8]27.7429.0534.5018.6217.5428.5623.5425.6516.7019.8818.0317.3817.5517.7216.2817.65
E-NeRV [24]26.4928.3135.5518.0717.2128.2223.1725.2916.7620.0218.2018.2516.7617.2516.8017.72
HNeRV [7]30.6435.5738.6021.0821.8230.7827.3129.4019.2620.7720.3319.0218.5118.8717.5619.19
DNeRV [48]30.7934.0738.1628.6824.7630.2425.9730.3420.3321.7121.5119.3021.0518.1118.9620.14
Ours(B)30.6036.4438.3720.8121.7630.7027.0529.3719.4820.8020.2319.1018.3918.8217.6319.21
Ours(T)31.9236.6338.6329.3727.5231.3528.3231.9620.4722.1921.8722.3721.3121.6419.1221.28
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.49, + 0.785, + 0.641 + ], + "angle": 0, + "content": "of 0.35M, 0.75M, 1.5M, and 3M, where the different sizes correspond to different bit budgets. The results demonstrate the superiority of our backbone model in various model capacities. Ours(T) also secures the second-best results in most scenarios. Tab. 2b exhibits the performance changes, when the epochs vary from 300 to 2400 with 3M-models, implying different encoding times. Ours(B) and Ours(T) exhibit the best and second-best performance, respectively. We also show the encoding complexity of \\(640 \\times 1280\\) video resolutions compared to other NeRV-based methods in Tab. 2b. Despite the higher complexity involved in processing multi-scale features, Ours(B) excels in video regression, delivering superior results over other methods across all sizes and training epochs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.787, + 0.779 + ], + "angle": 0, + "content": "Video Interpolation. Tab. 3 displays the comparative studies of video interpolation in UVG and DAVIS datasets. We train the model with odd frames and test results with even frames. We observe Ours(T) and DNeRV that embed the frame differences yield better performance than the tested methods without time embeddings because the time embedding is essential to continuously represent videos over time. Ours(T) achieves the highest performance among the tested methods, which is approximately \\(1.14\\mathrm{dB} \\sim 1.62\\mathrm{dB}\\) better than DNeRV in UVG and DAVIS datasets, respectively. The results indicate that Ours(T) provides efficient video representations over time." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Video Compression. We compare compression results with commercial video CODECs such as H.264 [45] and H.265 [40] and learning-based CODECs such as DVC [25], FVC [15], and DCVC [23]. Fig. 4 presents the rate-distortion curves. Our method improves coding efficiency over H.264 and all the other NeRV-based" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.33, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "SNeRV: Spectra-preserving Neural Representation for Video" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.283, + 0.159, + 0.485, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.159, + 0.71, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.304, + 0.289, + 0.698, + 0.305 + ], + "angle": 0, + "content": "Fig. 4: Coding performance comparisons in UVG datasets." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.337, + 0.486, + 0.364 + ], + "angle": 0, + "content": "Table 4: Video in-painting results on DAVIS dataset." + }, + { + "type": "table", + "bbox": [ + 0.228, + 0.373, + 0.472, + 0.436 + ], + "angle": 0, + "content": "
MethodIn-painting1In-painting2Avg.
HNeRV [7]29.7629.4129.59
DNeRV [48]30.0129.8229.92
Ours(B)30.7029.9030.30
" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.331, + 0.651, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.331, + 0.786, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.515, + 0.394, + 0.788, + 0.438 + ], + "angle": 0, + "content": "Fig.5: Visualization of In-painting1: masked input (left) and reconstructed output (right)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.459, + 0.789, + 0.793 + ], + "angle": 0, + "content": "methods and achieves comparable coding gains with H.265, DVC, and FVC at low bit-rates with SSIM results. In the comparisons, we use the same coding pipeline and configurations as in [48]. We adopt \\(10\\%\\) model pruning followed by 6-bit quantization for embeddings and 8-bit quantization for decoder parameters for all tested methods. Our methods do not apply additional techniques such as quantization-aware training (QAT) and predictive coding to improve coding efficiency for fair comparisons. HiNeRV [20] provides improved performance with specialized coding tools for NeRVs at the expense of substantial computational complexity and limited performance in video interpolation. Our method provides more fundamental changes as a backbone, which is orthogonal to the direction. Accordingly, we integrate our model into HiNeRV's coding framework and show the results denoted by Ours(B) + HiNeRV in Fig. 4. It improves the BD-PSNR by 0.3dB compared to HiNeRV. Our supplementary results also demonstrate the generalization ability of our study. In terms of model complexity, Ours(B) has 168G multiply-accumulates (MACs) in UVG dataset. HiNeRV and DNeRV have 111.71G MACs and 69.86G MACs in the same dataset, respectively. These values are significantly smaller than those of neural network based video compression (NVC) methods such as DCVC, which has 2,268G MACs as reported in [23]. Ours(B) has 31.3 decoding fps, which is relatively slower than 48.4 fps of HiNeRV but is significantly faster than 1.75 fps of DCVC, measured in NVIDIA GeForce RTX 3090 GPU. When comparing the results, Ours(B) provides a practically reasonable trade-off between coding performance and computational complexity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Video In-painting. We evaluate the performance of a video in-painting task. We use \\(50 \\times 50\\) masks with five fixed points and ten random points in In-painting1 and In-painting2 [48], respectively. Tab. 4 presents that our model provides im" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "J. Kim et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.787, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.217, + 0.281, + 0.785, + 0.302 + ], + "angle": 0, + "content": "(a) Visual comparisons of tested methods in the tasks of video regression. From the top, Bike, Bosph, and Yacht." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.304, + 0.787, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.217, + 0.396, + 0.785, + 0.418 + ], + "angle": 0, + "content": "(b) Visual comparisons of tested methods in the tasks of video interpolation. From the top, Jockey and Ready." + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.43, + 0.785, + 0.459 + ], + "angle": 0, + "content": "Fig. 6: Visual comparisons in video regression and interpolation tasks. From the left, ground-truth (GT), HNeRV, DNeRV, and Ours." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.49, + 0.785, + 0.521 + ], + "angle": 0, + "content": "proved results approximately \\(0.38\\mathrm{dB} \\sim 0.71\\mathrm{dB}\\) over the existing methods. Fig. 5 shows the visualization results of In-painting1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.522, + 0.785, + 0.568 + ], + "angle": 0, + "content": "Qualitative Performance Analysis. We provide the qualitative results of the tested methods in Fig. 6. The results illustrate that our methods provide visually pleasing results in video regression and interpolation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.597, + 0.348, + 0.61 + ], + "angle": 0, + "content": "4.3 Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Trade-off between Video Regression and Time Interpolation. Tab. 1 and Tab. 3 display the video regression and interpolation performances, respectively. According to the results, Ours(B) outperforms Ours(T) in regression, but underperforms in interpolation, indicating a trade-off relationship. We explain this trade-off between time duration \\((\\Delta t)\\) and spectral band \\((\\Delta \\omega)\\) in terms of learning aspects. Ours(B) sets \\(\\Delta t = 0\\), lacking the time analysis but maximizing the spectral analysis, while Ours(T) compromises \\(\\Delta t\\) and \\(\\Delta \\omega\\). Given the limited model parameters, learning the spatial spectral representation of a current frame could fit a reconstruction task but might compromise the generalization ability across time. In contrast, learning temporal relations can make it challenging to fit the parameters to the current time. Similar characteristics are present not only in Ours(B) and Ours(T) but also in other findings. DNeRV [48] offers superior performance in video interpolation compared to HNeRV [7], but not in video reconstruction. This strategy in our approach signifies a new direction in" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.33, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "SNeRV: Spectra-preserving Neural Representation for Video" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.36, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.221, + 0.203, + 0.337, + 0.225 + ], + "angle": 0, + "content": "(a) Low-frequency sults of Jockey." + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.145, + 0.498, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.362, + 0.203, + 0.5, + 0.225 + ], + "angle": 0, + "content": "(b) High-frequency results of Jockey." + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.145, + 0.641, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.503, + 0.203, + 0.639, + 0.225 + ], + "angle": 0, + "content": "(c) Low-frequency results of Ready." + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.145, + 0.782, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.643, + 0.203, + 0.782, + 0.225 + ], + "angle": 0, + "content": "(d) High-frequency results of Ready." + }, + { + "type": "image_caption", + "bbox": [ + 0.308, + 0.237, + 0.694, + 0.251 + ], + "angle": 0, + "content": "Fig. 7: Training rates for different frequency components." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.281, + 0.785, + 0.31 + ], + "angle": 0, + "content": "analyzing and optimizing NeRV for various video processing tasks from a time and spectra perspective, which has not been conducted before." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.311, + 0.787, + 0.508 + ], + "angle": 0, + "content": "Learning Characteristics of LF and HF Components. Fig. 7 displays the learning characteristics of LF and HF components separately for analysis, while they are trained simultaneously. Jockey and Ready sequences contain fine textures and motions, yielding some HF components that are challenging to represent with an implicit neural network. Fig. 7a and Fig. 7c show the results of the LF at training steps. The LFs are steadily trained to reach the convergence points producing similar PSNR values to the tested methods. Besides, Ours(T) and DNeRV exhibit better performance in the early stage, because they use larger sizes of embeddings. However, Fig. 7b and Fig. 7d show significantly different characteristics in learning HF components. While the previous methods fail to efficiently learn the HFs, our methods achieve significantly improved performance with increasing epochs. These results verify that our model could mitigate the spectral bias and successfully preserve HF components." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.529, + 0.388, + 0.544 + ], + "angle": 0, + "content": "4.4 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.788, + 0.719 + ], + "angle": 0, + "content": "MFU and HFR Modules. Tab. 5a presents the effectiveness of the MFU and HFR. We conduct ablation tests on the Bunny dataset mostly including LF components and the Yacht sequence in UVG including comparatively larger HF components. The best performance is achieved when both the MFU and HFR are utilized. MFU achieves significantly improved performance approximately \\(1.20\\mathrm{dB} \\sim 1.88\\mathrm{dB}\\), demonstrating the effectiveness of the module. Using HFR, while Bunny increases the performance by \\(0.21\\mathrm{dB}\\), Yacht observes \\(1.62\\mathrm{dB}\\). This confirms that HFR is capable of restoring HF components. Further, the HFR contributes approximately \\(0.44\\mathrm{dB} \\sim 1.50\\mathrm{dB}\\) on top of the MFU. This refinement is particularly advantageous in generating HFs that were hardly trained in previous NeRV models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.794 + ], + "angle": 0, + "content": "Temporal Up-sampling Block. We evaluate the performance of the TUBs, by replacing them in Ours(T) with the conventional NeRV blocks comprising PixelShuffle and 2D convolution layers. Tab. 5b shows that our TUB is suitable for processing time-extended features than the 2D operations. TUB improves the PSNR around 0.80dB." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Loss Terms. We conduct ablation tests on the loss function in Eq. (2) using a Yacht sequence. Tab. 5b shows the results, when the terms are turned on or off. Both the terms contribute to the performance improvements." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "J. Kim et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.276, + 0.145, + 0.725, + 0.16 + ], + "angle": 0, + "content": "Table 5: Ablation studies of the proposed modules and loss terms." + }, + { + "type": "table_caption", + "bbox": [ + 0.315, + 0.184, + 0.431, + 0.196 + ], + "angle": 0, + "content": "(a) MFU and HFR." + }, + { + "type": "table", + "bbox": [ + 0.252, + 0.208, + 0.487, + 0.301 + ], + "angle": 0, + "content": "
DatasetBunnyYacht
MFUHFRPSNRSSIMPSNRSSIM
XX38.000.991130.050.9323
✓X39.200.992431.930.9458
X✓38.210.991131.670.9516
✓✓39.640.992833.430.9647
" + }, + { + "type": "table_caption", + "bbox": [ + 0.557, + 0.184, + 0.706, + 0.196 + ], + "angle": 0, + "content": "(b) Loss terms and TUB." + }, + { + "type": "table", + "bbox": [ + 0.534, + 0.208, + 0.723, + 0.307 + ], + "angle": 0, + "content": "
L(It,ˆt) L(Ct,ˆCt)PSNRSSIM
✓X33.290.9600
✓✓33.430.9647
Temporal BlockPSNRSSIM
NeRV Block31.910.9532
TUB32.710.9584
" + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.324, + 0.459, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.403, + 0.486, + 0.424 + ], + "angle": 0, + "content": "(a) Varying temporal embedding size with fixed target embedding." + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.324, + 0.758, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.402, + 0.785, + 0.424 + ], + "angle": 0, + "content": "(b) Varying target embedding size with fixed temporal embedding." + }, + { + "type": "image_caption", + "bbox": [ + 0.25, + 0.436, + 0.751, + 0.451 + ], + "angle": 0, + "content": "Fig. 8: Regression and interpolation results with different embedding sizes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.479, + 0.788, + 0.587 + ], + "angle": 0, + "content": "Effect of Size of Model Components. We analyze the performance with different sizes of target \\( e_t \\) and temporal embeddings \\( e_{\\Delta t} \\)s, with the size of the decoder varying according to these sizes. Fig. 8 presents the performance variations, exploring the impacts of the ratios. The regression performance is reliable in all the cases, when the ratios are different. In interpolation, the size of the temporal embedding greatly affects the performance, because it hints the temporal relations among video frames." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.607, + 0.36, + 0.623 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.637, + 0.788, + 0.729 + ], + "angle": 0, + "content": "In this paper, we proposed the spectra-preserving neural representation for video as an efficient baseline framework. SNeRV successfully learned LF components, while reconstructing fine HF details within a limited capacity of parameters, and surpassed the existing NeRV-based methods in various video-related tasks. More compact and efficient architecture to process the high-frequency components considering time-frequency trade-off will be done as a future work." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.75, + 0.394, + 0.768 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.842 + ], + "angle": 0, + "content": "This work was partly supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No.RS-2021-II212068 and RS-2022-00167169) and was partly supported by the NRF grant funded by MSIT (No.NRF-2022R1A2C4002052)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.33, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SNeRV: Spectra-preserving Neural Representation for Video" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.324, + 0.161 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.179, + 0.785, + 0.223 + ], + "angle": 0, + "content": "1. Arora, S., Du, S., Hu, W., Li, Z., Wang, R.: Fine-grained analysis of optimization and generalization for overparameterized two-layer neural networks. International Conf. on Machine Learning (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.223, + 0.785, + 0.264 + ], + "angle": 0, + "content": "2. Bai, Y., Dong, C., Wang, C., Yuan, C.: Ps-nerv: Patch-wise stylized neural representations for videos. In: Proceedings of the IEEE conference on Image Processing (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.265, + 0.785, + 0.293 + ], + "angle": 0, + "content": "3. Balle, J., Minnen, D., Singh, S., Hwang, S.J., Johnston, N.: Variational image compression with a scale hyperprior. arXiv preprint arXiv:1802.01436 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.293, + 0.785, + 0.334 + ], + "angle": 0, + "content": "4. Basri, R., Galun, M., Geifman, A., Jacobs, D., Kasten, Y., Kritchman, S.: Frequency bias in neural networks for input of non-uniform density. arXiv preprint arXiv 2003.04560 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.334, + 0.785, + 0.39 + ], + "angle": 0, + "content": "5. Bross, B., Wang, Y.K., Ye, Y., Liu, S., Chen, J., Sullivan, G.J., Ohm, J.R.: Overview of the versatile video coding (vvc) standard and its applications. IEEE Transactions on Circuits and Systems for Video Technology 31(10), 3736-3764 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.391, + 0.785, + 0.419 + ], + "angle": 0, + "content": "6. Cao, Y., Fang, Z., Wu, Y., Zhou, D.X., Gu, Q.: Towards understanding the spectral bias of deep learning. arXiv preprint arXiv:1912.01198 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.42, + 0.785, + 0.461 + ], + "angle": 0, + "content": "7. Chen, H., Gwilliam, M., Lim, S., Shrivastava, A.: Hnerv: A hybrid neural representation for videos. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 10270-10279 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.462, + 0.785, + 0.503 + ], + "angle": 0, + "content": "8. Chen, H., He, B., Wang, H., Ren, Y., Lim, S.N., Shrivastava, A.: Nerv: Neural representations for videos. Advances in Neural Information Processing Systems 34, 21557-21568 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.504, + 0.785, + 0.56 + ], + "angle": 0, + "content": "9. Chen, Y.H., Chen, S.C., Lin, Y.Y., Peng, W.H.: Motif: Learning motion trajectories with local implicit neural functions for continuous space-time video super-resolution. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 23131-23141 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.56, + 0.785, + 0.615 + ], + "angle": 0, + "content": "0. Chen, Z., Chen, Y., Liu, J., Xu, X., Goel, V., Wang, Z., Shi, H., Wang, X.: Videoirr: Learning video implicit neural representation for continuous space-time superresolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2047-2057 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.616, + 0.785, + 0.657 + ], + "angle": 0, + "content": "1. Chen, Z., Zhang, H.: Learning implicit fields for generative shape modeling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5939-5948 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.659, + 0.785, + 0.7 + ], + "angle": 0, + "content": "2. Choi, Y.J., Lee, Y.W., Kim, B.G.: Wavelet attention embedding networks for video super-resolution. In: 2020 25th International conference on pattern recognition (ICPR). pp. 7314-7320. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.7, + 0.785, + 0.742 + ], + "angle": 0, + "content": "3. Djelouah, A., Campos, J., Schaub-Meyer, S., Schroers, C.: Neural inter-frame compression for video coding. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6421-6429 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.742, + 0.785, + 0.798 + ], + "angle": 0, + "content": "4. He, B., Yang, X., Wang, H., Wu, Z., Chen, H., Huang, S., Ren, Y., Lim, S.N., Shrivastava, A.: Towards scalable neural representation for diverse videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6132-6142 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "5. Hu, Z., Lu, G., Xu, D.: Fvc: A new framework towards deep video compression in feature space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1502-1511 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.179, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "J. Kim et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "16. Huang, H., He, R., Sun, Z., Tan, T.: Wavelet domain generative adversarial network for multi-scale face hallucination. International Journal of Computer Vision 127(6-7), 763-784 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.245 + ], + "angle": 0, + "content": "17. Jin, B., Hu, Y., Tang, Q., Niu, J., Shi, Z., Han, Y., Li, X.: Exploring spatial-temporal multi-frequency analysis for high-fidelity and temporal-consistency video prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4554-4563 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.246, + 0.785, + 0.3 + ], + "angle": 0, + "content": "18. Jung, H., Hui, Z., Luo, L., Yang, H., Liu, F., Yoo, S., Ranjan, R., Demandolx, D.: Anyflow: Arbitrary scale optical flow with implicit neural representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5455-5465 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.301, + 0.785, + 0.328 + ], + "angle": 0, + "content": "19. Kong, L., Jiang, B., Luo, D., Chu, W., Tai, Y., Wang, C., Yang, J.: Dynamic frame interpolation in wavelet domain. IEEE Transactions on Image Processing (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.328, + 0.785, + 0.37 + ], + "angle": 0, + "content": "20. Kwan, H.M., Gao, G., Zhang, F., Gower, A., Bull, D.: Hinerv: Video compression with hierarchical encoding-based neural representation. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.398 + ], + "angle": 0, + "content": "21. Lee, J.C., Rho, D., Ko, J.H., Park, E.: Fferv: Flow-guided frame-wise neural representations for videos. arXiv preprint arXiv:2212.12294 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.399, + 0.785, + 0.425 + ], + "angle": 0, + "content": "22. Lee, J.K., Kim, N., Cho, S., Kang, J.W.: Deep video prediction network-based inter-frame coding in hevc. IEEE ACCESS 8, 95906-95917 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.426, + 0.785, + 0.452 + ], + "angle": 0, + "content": "23. Li, J., Li, B., Lu, Y.: Deep contextual video compression. Advances in Neural Information Processing Systems 34, 18114-18125 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.495 + ], + "angle": 0, + "content": "24. Li, Z., Wang, M., Pi, H., Xu, K., Mei, J., Liu, Y.: E-nerv: Expedite neural video representation with disentangled spatial-temporal context. In: European Conference on Computer Vision. pp. 267-284. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.785, + 0.536 + ], + "angle": 0, + "content": "25. Lu, G., Ouyang, W., Xu, D., Zhang, X., Cai, C., Gao, Z.: Dvc: An end-to-end deep video compression framework. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11006-11015 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.537, + 0.785, + 0.591 + ], + "angle": 0, + "content": "26. Lu, Y., Wang, Z., Liu, M., Wang, H., Wang, L.: Learning spatial-temporal implicit neural representations for event-guided video super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1557-1567 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.785, + 0.66 + ], + "angle": 0, + "content": "27. Maiya, S.R., Girish, S., Ehrlich, M., Wang, H., Lee, K.S., Poirson, P., Wu, P., Wang, C., Shrivastava, A.: Nirvana: Neural implicit representations of videos with adaptive networks and autoregressive patch-wise modeling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14378-14387 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.659, + 0.674 + ], + "angle": 0, + "content": "28. Mallat, S.: A wavelet tour of signal processing. Elsevier (1999)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.785, + 0.715 + ], + "angle": 0, + "content": "29. Mercat, A., Viitanen, M., Vanne, J.: Uvg dataset: 50/120fps 4k sequences for video codec analysis and development. In: Proceedings of the 11th ACM Multimedia Systems Conference. pp. 297-302 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.785, + 0.757 + ], + "angle": 0, + "content": "30. Pan, Z., Yi, X., Zhang, Y., Jeon, B., Kwong, S.: Efficient in-loop filtering based on enhanced deep convolutional neural networks for hevc. IEEE Transactions on Image Processing 29, 5352-5366 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.813 + ], + "angle": 0, + "content": "31. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 165-174 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "32. Perazzi, F., Pont-Tuset, J., McWilliams, B., Gool, L.V., Gross, M., Sorkine-Hornung, A.: A benchmark dataset and evaluation methodology for video object" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.33, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SNeRV: Spectra-preserving Neural Representation for Video" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.241, + 0.148, + 0.786, + 0.175 + ], + "angle": 0, + "content": "segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 724-732 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.787, + 0.232 + ], + "angle": 0, + "content": "33. Ramamonjisoa, M., Firman, M., Watson, J., Lepetit, V., Turmukhambetov, D.: Single image depth prediction with wavelet decomposition. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11089-11098 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.233, + 0.786, + 0.272 + ], + "angle": 0, + "content": "34. Rho, D., Cho, J., Ko, J.H., Park, E.: Neural residual flow fields for efficient video representations. In: Proceedings of the Asian Conference on Computer Vision. pp. 3447-3463 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.787, + 0.314 + ], + "angle": 0, + "content": "35. Rho, D., Lee, B., Nam, S., Lee, J.C., Ko, J.H., Park, E.: Masked wavelet representation for compact neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20680-20690 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.787, + 0.355 + ], + "angle": 0, + "content": "36. Ronen, B., Jacobs, D., Kasten, Y., Kritchman, S.: The convergence rate of neural networks for learned functions of different frequencies. Advances in Neural Information Processing Systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.787, + 0.411 + ], + "angle": 0, + "content": "37. Saragadam, V., LeJeune, D., Tan, J., Balakrishnan, G., Veeraraghavan, A., Baraniuk, R.G.: Wire: Wavelet implicit neural representations. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 18507-18516 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.787, + 0.467 + ], + "angle": 0, + "content": "38. Shi, W., Caballero, J., Huszar, F., Totz, J., Aitken, A.P., Bishop, R., Wang, Z.: Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1874-1883 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.468, + 0.787, + 0.508 + ], + "angle": 0, + "content": "39. Sitzmann, V., Martel, J., Bergman, A., Lindell, D., Wetzstein, G.: Implicit neural representations with periodic activation functions. Advances in neural information processing systems 33, 7462-7473 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.787, + 0.55 + ], + "angle": 0, + "content": "40. Sullivan, G.J., Ohm, J.R., Han, W.J., Wiegand, T.: Overview of the high efficiency video coding (hevc) standard. IEEE Transactions on circuits and systems for video technology 22(12), 1649-1668 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.551, + 0.787, + 0.605 + ], + "angle": 0, + "content": "41. Tancik, M., Srinivasan, P., Mildenhall, B., Fridovich-Keil, S., Raghavan, N., Singhal, U., Ramamoorthi, R., Barron, B., Ng, R.: Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems 33 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.787, + 0.646 + ], + "angle": 0, + "content": "42. Taubman, D.S., Marcellin, M.W., Rabbani, M.: Jpeg2000: Image compression fundamentals, standards and practice. Journal of Electronic Imaging 11(2), 286-287 (2002)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.787, + 0.674 + ], + "angle": 0, + "content": "43. Unser, M., Blu, T.: Mathematical properties of the jpeg2000 wavelet filters. IEEE transactions on image processing 12(9), 1080-1090 (2003)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.787, + 0.715 + ], + "angle": 0, + "content": "44. Wang, P., Fan, Z., Chen, T., Wang, Z.: Neural implicit dictionary learning via mixture-of-expert training. In: International Conference on Machine Learning. pp. 22613-22624. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.787, + 0.757 + ], + "angle": 0, + "content": "45. Wiegand, T., Sullivan, G.J., Bjontegaard, G., Luthra, A.: Overview of the h. 264/avc video coding standard. IEEE Transactions on circuits and systems for video technology 13(7), 560-576 (2003)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.787, + 0.799 + ], + "angle": 0, + "content": "46. Xin, J., Li, J., Jiang, X., Wang, N., Huang, H., Gao, X.: Wavelet-based dual recursive network for image super-resolution. IEEE Transactions on Neural Networks and Learning Systems 33(2), 707-720 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.787, + 0.84 + ], + "angle": 0, + "content": "47. Yuce, G., Ortiz-Jimenez, G., Besbinar, B., Frossard, P.: A structured dictionary perspective on implicit neural representations. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 19228-19238 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.359, + 0.127 + ], + "angle": 0, + "content": "J. Kim et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.788, + 0.19 + ], + "angle": 0, + "content": "48. Zhao, Q., Asif, M.S., Ma, Z.: Dnerv: Modeling inherent dynamics via difference neural representation for videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2031-2040 (2023)" + } + ] +] \ No newline at end of file diff --git a/2024/SNeRV_ Spectra-preserving Neural Representation for Video/5d3bbd00-0b38-4046-ad52-4dc33526757f_origin.pdf b/2024/SNeRV_ Spectra-preserving Neural Representation for Video/5d3bbd00-0b38-4046-ad52-4dc33526757f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ec502e1192bed6640f6778770fdc5a067037000a --- /dev/null +++ b/2024/SNeRV_ Spectra-preserving Neural Representation for Video/5d3bbd00-0b38-4046-ad52-4dc33526757f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:730183a5ea6cdf549a3986673eb32232f112864749453815d15bbd64c684f80c +size 6274472 diff --git a/2024/SNeRV_ Spectra-preserving Neural Representation for Video/full.md b/2024/SNeRV_ Spectra-preserving Neural Representation for Video/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6132aaaeb46bc82941b16e37cc0c4fc77b74c053 --- /dev/null +++ b/2024/SNeRV_ Spectra-preserving Neural Representation for Video/full.md @@ -0,0 +1,295 @@ +# SNeRV: Spectra-preserving Neural Representation for Video + +Jina Kim\* JiHoo Lee\* and Je-Won Kang\*\* + +Department of Electronic and Electrical Engineering and Graduate Program in Smart Factory, Ewha Womans University, Seoul, South Korea qapo0106@ewhain.net, jihoo.lee@lge.com, jewonk@ewha.ac.kr + +Abstract. Neural representation for video (NeRV), which employs a neural network to parameterize video signals, introduces a novel methodology in video representations. However, existing NeRV-based methods have difficulty in capturing fine spatial details and motion patterns due to spectral bias, in which a neural network learns high-frequency (HF) components at a slower rate than low-frequency (LF) components. In this paper, we propose spectra-preserving NeRV (SNeRV) as a novel approach to enhance implicit video representations by efficiently handling various frequency components. SNeRV uses 2D discrete wavelet transform (DWT) to decompose video into LF and HF features, preserving spatial structures and directly addressing the spectral bias issue. To balance the compactness, we encode only the LF components, while HF components that include fine textures are generated by a decoder. Specialized modules, including a multi-resolution fusion unit (MFU) and a high-frequency restorer (HFR), are integrated into a backbone to facilitate the representation. Furthermore, we extend SNeRV to effectively capture temporal correlations between adjacent video frames, by casting the extension as additional frequency decomposition to a temporal domain. This approach allows us to embed spatio-temporal LF features into the network, using temporally extended up-sampling blocks (TUBs). Experimental results demonstrate that SNeRV outperforms existing NeRV models in capturing fine details and achieves enhanced reconstruction, making it a promising approach in the field of implicit video representations. The codes are available at https://github.com/qwertja/SNeRV. + +Keywords: Implicit neural representation $\cdot$ Neural representation for video $\cdot$ Wavelet transform + +# 1 Introduction + +Implicit neural representation for video, also referred to as neural representation for video (NeRV), has attracted considerable attention as a promising method + +![](images/9d233f82030302960d1de259d0c259dfc565af22d4ec7691c2c6d635325a24f6.jpg) +Fig. 1: Visual comparisons of the reconstructed HF coefficients of previous NeRV methods [7,48] and the proposed method in "Jockey" sequence. Our model is designed to efficiently encode fine details, by implicitly circumventing the spectral bias problem. + +for video representations [7,8,14,20,21,24,27,48]. NeRV is used to parameterize a video signal with a neural network, in which a space-time coordinate is used as a query, and the network outputs the corresponding RGB value. As a result, a NeRV model itself can serve as a proxy for the video. This approach has proven effective in various tasks, including video compression [20,21], video interpolation [18,34], and video super-resolution [9,10,26]. + +Previous studies use pixel-wise [11, 31, 39] or frame-wise mapping models [8] with various positional embedding of space-time coordinates. However, the quality of the reconstructed frame is severely degraded, when a learned decoder fails to capture the spatial details of diverse video contents. HNeRV [7] utilizes a content-adaptive embedding generated by an encoder in addition to a learned decoder. Temporal correlation among adjacent video frames is exploited in the NeRV pipelines using a dynamical system [48] and a decomposition of visual content and motion information [14]. Nevertheless, capturing video representations remains challenging due to the intricate spatio-temporal dynamics in video scenes, while trying to maintain the size of a model. + +The performance of a NeRV model is significantly affected by the characteristics of an input video [7,24,44,48]. When a NeRV approximates an input video, the optimization is carried out to best represent the input within a budget of a model size. However, inherently, the current methods struggle to accurately capture fine spatial details and temporal patterns [7,14,24,48]. According to recent studies on the neural tangent kernel (NTK) [6,36], the learning rate of a neural network is determined by the magnitude of the eigenvalues of the target function. Because natural videos are mostly composed of low-frequency (LF) components, neural networks tend to learn the LF components more quickly, while the learning of high-frequency (HF) components proceeds at a slower rate. Fig. 1 indicates that the existing NeRVs would face challenges in representing HF components due to spectral bias within the vast redundancies. While these behaviors are evident in the learning of implicit neural representations [47], current NeRV methods have not addressed this problem effectively. + +In this paper, we propose a spectra-preserving NeRV (SNeRV) that enhances the learning of implicit video representations, by efficiently processing different frequency components. First, we present a SNeRV backbone model to achieve + +the goal. We employ a discrete wavelet transform (DWT) to decompose a video into LF and HF features, thereby preserving their spatial structures and directly addressing the issue in the frequency domain. The model capacity is determined with the content-adaptive embeddings and the decoder sizes as [7]. To represent a target in more detail, a decoder requires a larger size of a model or an embedding vector, but it compromises the model compactness. Thus, our encoder embeds only the LF component, exploiting its more suitable characteristics to learn the implicit representations. Meanwhile, the HF components including fine textures are generated by a decoder to circumvent the bias problem. This strategy enables a model to reduce redundancies and achieve a more compact representation with several specialized modules such as a multi-resolution fusion unit (MFU) and a high-frequency restorer (HFR) in a backbone model. + +On top of the backbone, we extend our model to efficiently represent temporal correlations between adjacent video frames, by casting the temporal modeling as spectral decomposition of the video frames. This is accomplished by performing an additional 1D DWT to a temporal dimension, where the motion is considered as HF components. While the backbone lacks the time analysis by maximising the capability of spectral analysis, the temporal extension considers both time and frequency domains in the analysis. By extending the time analysis, the backbone can capture temporal correlation and efficiently perform temporal modeling, well-suited for related tasks such as video interpolation. In experimental results, the backbone and its temporal extension each exhibit a performance trade-off in video regression and interpolation, while maintaining the same model capacity. + +Our paper has major contributions as follows: + +- We propose the spectra-preserving NeRV that enhances implicit video representations by efficiently processing different frequency components through specialized modules such as an MFU and a HFR. +- We extend the backbone to capture temporal correlations between adjacent video frames, by embedding spatio-temporal LF features into the model. +- We evaluate the performance of the proposed methods on various video processing tasks and datasets, demonstrating their effectiveness. Our model outperforms previous approaches, emphasizing the advantages of mitigating the spectral bias problem in the INR for videos. + +# 2 Related Work + +Implicit Neural Representation (INR) for Video. INR for video exploited multi-layer perceptron (MLP) transforming the continuous spaces of high-dimensional data into density or RGB values [11,39]. However, the computational complexity increased substantially, as they calculated a neural function for each coordinate. To address this issue, NeRV [8] proposed a method to use a temporal index with positional encoding (PE) and restore up-scaled spatial information using a NeRV block with a PixelShuffle method [38]. Based on this, E-NeRV [24] sepa + +rated spatial and temporal contexts to improve the representation. PS-NeRV [2] further decomposed a video into a set of patches. + +Pixel-wise [11,31,39] and frame-wise [2,8] NeRV models exploited no spatial prior of a target frame, which compromised the performance. HNeRV [7] proposed a hybrid model using a content-adaptive frame embedding. The embedding was created in the encoder and jointly optimized with the trained decoder parameters. By conveying the content-adaptive spatial features, it enhanced the quality of reconstructed frames. However, as the parameters were fitted to the current frame, the interpolation performance was simultaneously degraded due to the lack of temporal modeling. For this, the motion information in adjacent frames was exploited [14,21,27,48] to achieve better results. + +Neural Networks in Frequency Domain. Discrete wavelet transform (DWT) decomposes a local region of an image into different frequency levels [28] and proven to be effective in various computer vision tasks [16, 17, 46]. Multi-level space-time wavelet decomposition was employed to extract motion information using a video prediction network, while preserving spatial details [17]. This technique was applied to image super resolution [46], enhanced with wavelet attention embeddings for video super resolution [12]. The DWT typically yields sparse yet significant non-zero coefficients when representing the original signal, a property that was actively used in various directions, such as monocular depth estimation [33] and video interpolation [19]. + +Neural Tangent Kernel. Modeling the learning process of a neural network has been actively studied using kernel regression with NTK analysis [1,4,6,36], in which the training rate is decided by the eigenvalues of the NTK matrix. The eigenvalue spectrum of an NTK matrix explains the spectral bias of a neural network towards training speed of HF functions [4]. As over-parameterized neural networks would be poorly suited for learning HF functions [1], mitigating the bias is essential for improving the performance of neural networks [39, 41]. Several studies have attempted to alleviate the bias using various input mapping of a coordinate. Sinusoidal functions were modulated into an activation function [39]. Gabor-wavelets were introduced for flexible representations in space and frequency domains [37]. However, such efforts have not been explored for NeRVs to circumvent the spectral bias. Motivated by the NTK, we aim to develop compact implicit representations of diverse frequency components with a reduced number of parameters. We use the inherent sparsity of the HF wavelet transformed components in videos. + +Video Compression. In recent years, deep learning-based video coding techniques were actively studied as alternatives to traditional CODECs [5,40,45]. This approach involved substituting various coding modules, such as frame restoration [30], motion prediction [13,22], and transform and entropy coding [3], with neural networks. DVC [25] proposed an end-to-end video coding model. The NeRV models [8,20,21,27] converted the compression tasks into model compression, using weight pruning, quantization, and entropy coding. + +![](images/9ce82f8f63d7d435fd71dd1ce7075cfb8fc8f26c714795185785758566d35a82.jpg) +Fig. 2: SNeRV backbone encoder and decoder architectures. The encoder applies 2D DWT to extract LF and HF features and embeds only the LF feature to save parameters. The decoder uses MFU and HFR to efficiently process the LF and HF features. CT and RB refer to transposed convolution and residual blocks, respectively. + +# 3 Proposed Method + +# 3.1 Overview + +Our network is designed to enhance the learning of implicit representations of video by mitigating spectral bias. We use a DWT to decompose video into LF and HF components to use an analysis-and-synthesis approach. The LF features with little variation across frames are embedded through the network, because this is inherently suitable for learning a NeRV model. The HF features are restored by a decoder. Specifically, the decoder sequentially forms the HF features with the LF ones as prior rather than directly learning the HF details. This implicit mapping enables the reallocation of the parameters to concentrate on representing LF features. Both features are optimized to consider the analysis-and-synthesis scheme, thus allowing for a more efficient representation of the HF textures and motions, based upon the stable base provided by the LF contents. + +These schemes are implemented as our backbone model of the SNeRV. It employs the HNeRV framework [7], in which the implicit representations are composed of a set of embedding vectors and learned decoder parameters. Based on this, we explain the novel design principles of the encoder and decoder, as shown in Fig. 2. On top of that, we extend the backbone to effectively represent temporal contexts of video. We explain the extended modules to accommodate these schemes step-by-step in Fig. 3. + +The differences between the backbone and its temporal extension not only clearly demonstrate their distinct architectural characteristics but also assist the performance analysis in various video processing tasks, while the backbone exhibits its superiority to the conventional methods. + +# 3.2 SNeRV Backbone Architecture + +Encoder Design. Given a video sequence $\mathbf{V} = \{I_0, I_1, \dots, I_{T-1}\}$ , in the backbone, an encoder conducts a 2D DWT along the $x - y$ spatial axes to decompose + +each video frame $I_{t}\in \mathbb{R}^{H\times W\times 3}$ into multiple wavelet sub-bands. It produces four sets of wavelet coefficients in $\mathbb{R}^{H / 2\times W / 2\times 3}$ , including $C_{LL}$ , $C_{LH}$ , $C_{HL}$ , and $C_{HH}$ , at the current time-step. $C_{LL}$ contains the LF component, which often appears as a blurred and smoothed version of the original frame. The HF components, represented by $C_{LH}$ , $C_{HL}$ , and $C_{HH}$ , encode the horizontal, vertical, and diagonal details of the image. For simplicity, we utilize the Haar WT [28], where the low pass filter is $\phi_L = [1 / \sqrt{2},1 / \sqrt{2} ]$ , and the high pass filter is $\phi_H = [1 / \sqrt{2}, - 1 / \sqrt{2} ]$ + +We input only the $C_{LL}$ into a down-sampling block (DB) to generate a content-adaptive and time-specific embedding $e_t$ whose dimension is $16 \times 2 \times 4$ . We use the same DB blocks as [7] and keep the same dimension of $e_t$ for fair comparisons. Due to the 2D DWT, the reduced original input size alleviates the burden of parameter sizes on the decoder side. $C_{LH}$ , $C_{HL}$ , and $C_{HH}$ are not explicitly embedded but synthesized in the decoder. + +Decoder Design. The decoder involves the reconstruction process using $\mathbf{e} = \{e_t\}$ and learned decoder parameters. It consists of three components, including up-sampling blocks (UBs), MFU, and HFR. Firstly, UB is used to progressively increase the resolution of the embedding and reconstruct the spectral information of a target time step. We employ five NeRV blocks [8,38] for the UBs. Then, the MFU fuses features from different resolutions to enrich the LF features. The HFR restores $C_{LH}$ , $C_{HL}$ , and $C_{HH}$ from the reconstructed $C_{LL}$ . The target frame is reconstructed from an inverse DWT (IDWT), by combining all the frequency components. Fig. 2 presents the decoder architecture. We explain the MFU and HFR modules in detail as follows. + +Multi-resolution Fusion Unit (MFU). UBs are used to sequentially reconstruct the LF features and recover the original dimension. However, using only UBs has limits in improving the output quality of $C_{LL}$ . To address this, we develop the MFU with a coarse-to-fine structure, in which the blurred outcomes are progressively refined with multi-scaled features during up-sampling. The MFU comprises two MF blocks (MFBs), each consisting of one transposed convolution and six residual blocks. + +In each MFB, the current feature $m_{i}$ is up-sampled using the transposed convolution, and then its concatenation with the outcomes of the $i$ -th UB, denoted by $u_{i}$ , goes through the residual blocks, which is mathematically expressed as, + +$$ +m _ {i + 1} = \mathcal {F} _ {R B} \left(\mathcal {F} _ {C T} \left(m _ {i}\right) \circ u _ {i + 3}\right), \quad i = 0, 1, \tag {1} +$$ + +where $\mathcal{F}_{CT}$ and $\mathcal{F}_{RB}$ are the operations of the transposed convolution and residual blocks, respectively, $\circ$ is a concatenation, and $m_0 = u_2$ . We use $u_i$ from the last three UBs and two MFBs, empirically considering the trade-off between learning ability and the size of parameters. + +High-Frequency Restorer (HFR). Sparse representations of HF wavelet coefficients have been exploited in wavelet video compression and various computer vision researches [33,35,42,43]. The HF coefficients in the flat regions of an image are nearly zero, whereas significant coefficients are found in the edges, which are essential for achieving a high quality image. Motivated by this, the HFR is + +![](images/ebfeaa92fc242be4e6215a8340337109603337e6051a925a858ffbacf4ae1c93.jpg) +Fig. 3: Temporal extension from the backbone: the encoder uses additional 1D DWT to generate spatio-temporal embeddings. The decoder uses TUBs to address the features. + +designed to restore the high spatial-frequency details from the composite of LF features generated by the MFU. We employ three HFR blocks, with each comprising two convolution layers. These HFR blocks are responsible for generating $C_{LH}$ , $C_{HL}$ , and $C_{HH}$ from the results of $\mathcal{F}_{MFU}$ . This process is shown in Fig. 2. Because the synthesis process requires less parameters than the direct learning of the HF coefficients, the decoder can save the parameters and allow for more parameters to represent the LF ones. + +# 3.3 SNeRV Temporal Extension + +Inspired by [48], we extend the backbone to analyse the effect of considering temporal redundancies. We apply 3D $(2 + 1\mathrm{D})$ DWT to the image sequence $\mathbf{V}$ to decompose the LF and HF components along temporal axis in addition to the existing 2D DWT. Specifically, we use two sets of adjacent frames, including $\{I_{t - 1},I_t\}$ and $\{I_t,I_{t + 1}\}$ as a backward and a forward set, respectively. The 3D DWT is applied to the two image sets, in which 2D DWTs are first applied, and then 1D DWT is applied to the temporal axis. Accordingly, it produces two spatio-temporal LF components corresponding to the sets. Then, we acquire temporal embeddings $e_{\Delta t_b}$ and $e_{\Delta t_f}$ through DBs to represent the LF components of the backward and forward set, respectively. In the decoder, $e_{\Delta t}$ s are jointly up-sampled with $e_t$ . For this, we replace the last three UBs with temporally extended UBs (TUBs), since they are used as inputs of the MFU as in the backbone. The overall architecture with temporal extension is shown in Fig. 3. Temporally Extended Up-sampling Block (TUB). We propose the TUBs to process $e_{\Delta t_b}$ and $e_{\Delta t_f}$ in addition to $e_t$ using 3D convolution operations. The three inputs are up-sampled using a transposed convolution and then passed through two 3D convolution layers. The computational complexity of the 3D convolution can be significantly reduced using 2D operations. More details of our implementation can be found in the supplementary. During the consecutive 3D convolutions, the channel dimension increases by a factor of 2 and then reduces back to expand the implicit capacity of the network. In the last TUB, all three components reflecting the different time steps are merged to output the + +LF feature for a target at $t$ . This operation allows to exploit temporal information in adjacent frames, resulting in an enhanced expressive capacity of the network over time. The outputs $C_{LL}, C_{LH}, C_{HL}$ , and $C_{HH}$ for the target at $t$ are obtained by passing the results of the last TUB through the MFU and HFR, as described in Sec. 3.2, obtaining the final output by applying 2D IDWT. + +# 3.4 Implementation + +The total size of the decoder parameters and embedding vectors determine the model capacity for representation. Because the performance would vary with the sizes and the extra embeddings are generated at the expense of the decoder size, we carefully explain the modified parameters to keep the total size. We set the size of $e_t$ to $16 \times 2 \times 4$ and use five UBs as in [7] for fair comparisons. We use $C_{LL}$ as input that is recovered to the original dimension using IDWT. Because the IDWT replaces a UB in the last stage, we save some parameters by setting a size of a stride to the half of [7, 48]. This parsimony allows the parameters for MFU and HFR. In the temporal extension, we use a dimension of $3 \times 20 \times 40$ for $e_{\Delta t}$ . Because the size is slightly larger than that of $e_t$ , in our implementation, the channel width of a decoder is reduced to maintain the overall size. We analyze the trade-offs with various temporal embedding sizes in Sec. 4.4. More details such as stride and channel sizes can be found in the supplementary materials. + +# 3.5 Loss Function + +To generate visually pleasing frames with the restored HF features, we combine both L1 and SSIM loss functions over frames, given as + +$$ +L \left(I _ {t}, \hat {I} _ {t}\right) = \frac {1}{T} \sum_ {t = 0} ^ {T - 1} \alpha \left\| I _ {t} - \hat {I} _ {t} \right\| _ {1} + (1 - \alpha) \left(1 - \operatorname {S S I M} \left(I _ {t}, \hat {I} _ {t}\right)\right), \tag {2} +$$ + +where $\hat{I}_t$ is the reconstructed frame, and $I_{t}$ is the ground truth at time step $t$ . SSIM $(\cdot)$ calculates the SSIM score. $\alpha$ is set to 0.7 during training. + +Moreover, we compute the loss on a set of the four wavelet coefficients $\mathbf{C}$ for better frequency reconstruction. Our final loss objective is formulated as + +$$ +L _ {t o t a l} = L (I _ {t}, \hat {I} _ {t}) + L (\mathbf {C} _ {t}, \hat {\mathbf {C}} _ {t}), \tag {3} +$$ + +where $\hat{\mathbf{C}}_t$ and $\mathbf{C}_t$ are the output coefficients and the ground truth, respectively. + +# 4 Experiments + +# 4.1 Experimental Settings + +We conduct experiments on Bunny, UVG [29], and DAVIS [32] datasets. Bunny has 132 frames of $720 \times 1280$ resolution. UVG is composed of 7 video sequences + +Table 1: Performance comparisons (dB) in video regression in UVG and DAVIS datasets with different sizes of resolutions. + +
DatasetUVGDAVIS
sizemethodBeautyBosphBeeJockeyReadyShakeYachtAvg.BikeSwanDanceCamelCarCowsTwirlAvg.
960NeRV [8]33.5633.2837.9331.3126.9532.6728.7432.0625.7727.2225.5923.8228.0921.7525.5125.39
E-NeRV [24]34.1232.9635.4531.9230.1432.9129.2232.3927.1228.7027.8725.6828.1824.2426.8126.94
HNeRV [7]34.1236.3838.9633.5829.9734.1531.4034.0830.7531.7728.8426.5933.5023.8528.5529.12
DNeRV [48]34.0235.3938.7134.4530.4033.5730.8533.9130.0432.6129.5226.9732.9723.6428.8929.23
Ours(B)34.4337.7438.8035.6933.1634.5233.4335.4033.2933.8331.4028.6835.7925.1430.4131.22
Ours(T)34.1437.6439.1136.1733.2134.4132.7135.3431.9033.0831.2630.1434.3327.3829.9331.15
480NeRV [8]34.5935.4238.9034.4330.1631.8233.5034.1228.9132.7728.9530.2632.0824.7727.5729.33
E-NeRV [24]35.0836.6339.2834.6231.0632.4333.9834.7329.0333.3529.2130.9834.9925.7626.9230.04
HNeRV [7]35.6838.7339.7237.1234.3834.2437.6336.7936.0739.4034.6235.2038.6129.8933.3435.30
DNeRV [48]35.3038.4039.5037.5634.1833.8337.1436.5635.6639.2534.2235.1438.3529.0833.2434.99
Ours(B)36.0140.0939.5238.3636.9035.1538.9737.8637.3240.5736.0236.6839.0831.7835.1636.66
Ours(T)35.7840.2439.6138.2136.7734.7038.3037.6636.4038.9435.5936.0438.1732.1134.1235.91
+ +with 300 or 600 frames of $1080 \times 1920$ . We utilize 20 subsets of DAVIS datasets of $1080 \times 1920$ sized frames. We used the same center-crop as in [7,48]. For training, we utilized an Adam optimizer and a learning rate of $1 \times 10^{-3}$ with a cosine learning rate scheduling. The batch size is set to 1. + +We use a peak signal-to-noise ratio (PSNR) and a structural similarity index measure (SSIM) for evaluation and bits-per-pixel (bpp) for compression. If none of the conditions are specified, we use 3M-sized models trained for 300 epochs in the experiments. We compare our backbone, denoted as "Ours(B)", and the temporal extension, denoted as "Ours(T)", with the state-of-the-art methods such as NeRV [8], E-NeRV [24], HNeRV [7], and DNeRV [48]. We train all the models, utilizing the original codes provided by the authors, and report the testing results under the same conditions. To ensure a fair comparison, we maintained the total capacity of the experimented models. To reduce the original model size from 3.8M to 3M in [48], the channel width was decreased from 92 to 70 which varies depending on the embedding sizes, resulting in a slight drop in performance compared to the original results. + +# 4.2 Performance Evaluation and Analysis + +Video Regression. Tab. 1 presents the performance of the tested methods in the tasks of video regression in UVG and DAVIS datasets with various resolutions. In the comparisons, our method presents the highest performance on average among the tested methods. In UVG dataset of $960 \times 1920$ resolutions, Ours(B) improves the performance approximately by $1.32\mathrm{dB} \sim 1.46\mathrm{dB}$ over HNeRV and DNeRV, respectively. In DAVIS datasets, our methods outperform the previous methods by $2.17\mathrm{dB} \sim 2.35\mathrm{dB}$ . While we show the results from 7 video samples of DAVIS subsets, more results are reported in the supplementary. We observe similar results in $480 \times 960$ resolutions. These results imply that our methods significantly improve the performance in video regression. + +Tab. 2 exhibits the performance analysis with different model sizes and training epochs in Bunny dataset. We present the results in Tab. 2a for model sizes + +(a) Performance evaluation in PSNR (dB) with different model sizes. + +Table 2: Performance comparisons (dB) in video regression on Bunny dataset with different model sizes and training epochs. + +
Model size0.35M0.75M1.5M3.0M
NeRV [8]28.1629.8331.8833.01
E-NeRV [24]27.7829.1231.6536.89
HNeRV [7]30.6732.8135.5738.06
DNeRV [48]29.3631.9735.1837.55
Ours(B)30.8833.2536.7639.64
Ours(T)30.4033.1936.6839.09
+ +(b) Performance evaluation with training epochs and model complexity. + +
Epochs300600120018002400MACs
NeRV [8]33.0133.5433.9634.1634.21100.9G
E-NeRV [24]36.8938.6939.7139.9039.96103.3G
HNeRV [7]38.0639.2739.8740.1040.1760.89G
DNeRV [48]37.5538.1738.5538.6838.7448.39G
Ours(B)39.6440.4040.8240.9440.9990.49G
Ours(T)39.0939.9540.4140.6340.68181.8G
+ +Table 3: Video interpolation results on UVG and DAVIS datasets. + +
DatasetUVGDAVIS
MethodBeautyBosphBeeJockeyReadyShakeYachtAvg.BikeSwanDanceCamelCarCowsTwirlAvg.
NeRV [8]27.7429.0534.5018.6217.5428.5623.5425.6516.7019.8818.0317.3817.5517.7216.2817.65
E-NeRV [24]26.4928.3135.5518.0717.2128.2223.1725.2916.7620.0218.2018.2516.7617.2516.8017.72
HNeRV [7]30.6435.5738.6021.0821.8230.7827.3129.4019.2620.7720.3319.0218.5118.8717.5619.19
DNeRV [48]30.7934.0738.1628.6824.7630.2425.9730.3420.3321.7121.5119.3021.0518.1118.9620.14
Ours(B)30.6036.4438.3720.8121.7630.7027.0529.3719.4820.8020.2319.1018.3918.8217.6319.21
Ours(T)31.9236.6338.6329.3727.5231.3528.3231.9620.4722.1921.8722.3721.3121.6419.1221.28
+ +of 0.35M, 0.75M, 1.5M, and 3M, where the different sizes correspond to different bit budgets. The results demonstrate the superiority of our backbone model in various model capacities. Ours(T) also secures the second-best results in most scenarios. Tab. 2b exhibits the performance changes, when the epochs vary from 300 to 2400 with 3M-models, implying different encoding times. Ours(B) and Ours(T) exhibit the best and second-best performance, respectively. We also show the encoding complexity of $640 \times 1280$ video resolutions compared to other NeRV-based methods in Tab. 2b. Despite the higher complexity involved in processing multi-scale features, Ours(B) excels in video regression, delivering superior results over other methods across all sizes and training epochs. + +Video Interpolation. Tab. 3 displays the comparative studies of video interpolation in UVG and DAVIS datasets. We train the model with odd frames and test results with even frames. We observe Ours(T) and DNeRV that embed the frame differences yield better performance than the tested methods without time embeddings because the time embedding is essential to continuously represent videos over time. Ours(T) achieves the highest performance among the tested methods, which is approximately $1.14\mathrm{dB} \sim 1.62\mathrm{dB}$ better than DNeRV in UVG and DAVIS datasets, respectively. The results indicate that Ours(T) provides efficient video representations over time. + +Video Compression. We compare compression results with commercial video CODECs such as H.264 [45] and H.265 [40] and learning-based CODECs such as DVC [25], FVC [15], and DCVC [23]. Fig. 4 presents the rate-distortion curves. Our method improves coding efficiency over H.264 and all the other NeRV-based + +![](images/3eb6da531aa5f2090a67f52e8bd4ffdb93f5a66b2179e96c8208cf5fc5efbcd7.jpg) +Fig. 4: Coding performance comparisons in UVG datasets. + +![](images/dda63efc83aeb83aa9eae93ee9522eb9089a19ca1ba74aa654a8514867928f4e.jpg) + +Table 4: Video in-painting results on DAVIS dataset. + +
MethodIn-painting1In-painting2Avg.
HNeRV [7]29.7629.4129.59
DNeRV [48]30.0129.8229.92
Ours(B)30.7029.9030.30
+ +![](images/fffbd7573413caf65d2d0310c41c3339596e6ef09efb495af3c96a4ea265af35.jpg) +Fig.5: Visualization of In-painting1: masked input (left) and reconstructed output (right). + +![](images/ef92037fcb378b18b2d1a3b41690845f120cca190cd9a07d41be25e9b6b4e1b3.jpg) + +methods and achieves comparable coding gains with H.265, DVC, and FVC at low bit-rates with SSIM results. In the comparisons, we use the same coding pipeline and configurations as in [48]. We adopt $10\%$ model pruning followed by 6-bit quantization for embeddings and 8-bit quantization for decoder parameters for all tested methods. Our methods do not apply additional techniques such as quantization-aware training (QAT) and predictive coding to improve coding efficiency for fair comparisons. HiNeRV [20] provides improved performance with specialized coding tools for NeRVs at the expense of substantial computational complexity and limited performance in video interpolation. Our method provides more fundamental changes as a backbone, which is orthogonal to the direction. Accordingly, we integrate our model into HiNeRV's coding framework and show the results denoted by Ours(B) + HiNeRV in Fig. 4. It improves the BD-PSNR by 0.3dB compared to HiNeRV. Our supplementary results also demonstrate the generalization ability of our study. In terms of model complexity, Ours(B) has 168G multiply-accumulates (MACs) in UVG dataset. HiNeRV and DNeRV have 111.71G MACs and 69.86G MACs in the same dataset, respectively. These values are significantly smaller than those of neural network based video compression (NVC) methods such as DCVC, which has 2,268G MACs as reported in [23]. Ours(B) has 31.3 decoding fps, which is relatively slower than 48.4 fps of HiNeRV but is significantly faster than 1.75 fps of DCVC, measured in NVIDIA GeForce RTX 3090 GPU. When comparing the results, Ours(B) provides a practically reasonable trade-off between coding performance and computational complexity. + +Video In-painting. We evaluate the performance of a video in-painting task. We use $50 \times 50$ masks with five fixed points and ten random points in In-painting1 and In-painting2 [48], respectively. Tab. 4 presents that our model provides im + +![](images/3546ef29a6e13f62093cb4c8e176fd4a4cda25cd12efe6375f6a6af2c7a568ed.jpg) +(a) Visual comparisons of tested methods in the tasks of video regression. From the top, Bike, Bosph, and Yacht. + +![](images/068a919300fab2b011d571d3e9608b116b749585fc2c1686c45554aafc4cca51.jpg) +(b) Visual comparisons of tested methods in the tasks of video interpolation. From the top, Jockey and Ready. +Fig. 6: Visual comparisons in video regression and interpolation tasks. From the left, ground-truth (GT), HNeRV, DNeRV, and Ours. + +proved results approximately $0.38\mathrm{dB} \sim 0.71\mathrm{dB}$ over the existing methods. Fig. 5 shows the visualization results of In-painting1. + +Qualitative Performance Analysis. We provide the qualitative results of the tested methods in Fig. 6. The results illustrate that our methods provide visually pleasing results in video regression and interpolation. + +# 4.3 Discussion + +Trade-off between Video Regression and Time Interpolation. Tab. 1 and Tab. 3 display the video regression and interpolation performances, respectively. According to the results, Ours(B) outperforms Ours(T) in regression, but underperforms in interpolation, indicating a trade-off relationship. We explain this trade-off between time duration $(\Delta t)$ and spectral band $(\Delta \omega)$ in terms of learning aspects. Ours(B) sets $\Delta t = 0$ , lacking the time analysis but maximizing the spectral analysis, while Ours(T) compromises $\Delta t$ and $\Delta \omega$ . Given the limited model parameters, learning the spatial spectral representation of a current frame could fit a reconstruction task but might compromise the generalization ability across time. In contrast, learning temporal relations can make it challenging to fit the parameters to the current time. Similar characteristics are present not only in Ours(B) and Ours(T) but also in other findings. DNeRV [48] offers superior performance in video interpolation compared to HNeRV [7], but not in video reconstruction. This strategy in our approach signifies a new direction in + +![](images/48dec2addea558ab2788c4054e35414187caae35d2b315a8db0f58ac765bbe8b.jpg) +(a) Low-frequency sults of Jockey. + +![](images/ae021564a6f47e982f462600072c47998d218b5e337d4abb6f00e61c4e1a5511.jpg) +(b) High-frequency results of Jockey. + +![](images/b050efaad87d189e4014a0025b577ac65ac91de063d2b880a1cda3a1dcb555e5.jpg) +Fig. 7: Training rates for different frequency components. + +![](images/541841b39c7bda20f47be1c1d39b061dadb0bb59a2b3dc939d75eed7c7bb29ef.jpg) +(c) Low-frequency results of Ready. +(d) High-frequency results of Ready. + +analyzing and optimizing NeRV for various video processing tasks from a time and spectra perspective, which has not been conducted before. + +Learning Characteristics of LF and HF Components. Fig. 7 displays the learning characteristics of LF and HF components separately for analysis, while they are trained simultaneously. Jockey and Ready sequences contain fine textures and motions, yielding some HF components that are challenging to represent with an implicit neural network. Fig. 7a and Fig. 7c show the results of the LF at training steps. The LFs are steadily trained to reach the convergence points producing similar PSNR values to the tested methods. Besides, Ours(T) and DNeRV exhibit better performance in the early stage, because they use larger sizes of embeddings. However, Fig. 7b and Fig. 7d show significantly different characteristics in learning HF components. While the previous methods fail to efficiently learn the HFs, our methods achieve significantly improved performance with increasing epochs. These results verify that our model could mitigate the spectral bias and successfully preserve HF components. + +# 4.4 Ablation Study + +MFU and HFR Modules. Tab. 5a presents the effectiveness of the MFU and HFR. We conduct ablation tests on the Bunny dataset mostly including LF components and the Yacht sequence in UVG including comparatively larger HF components. The best performance is achieved when both the MFU and HFR are utilized. MFU achieves significantly improved performance approximately $1.20\mathrm{dB} \sim 1.88\mathrm{dB}$ , demonstrating the effectiveness of the module. Using HFR, while Bunny increases the performance by $0.21\mathrm{dB}$ , Yacht observes $1.62\mathrm{dB}$ . This confirms that HFR is capable of restoring HF components. Further, the HFR contributes approximately $0.44\mathrm{dB} \sim 1.50\mathrm{dB}$ on top of the MFU. This refinement is particularly advantageous in generating HFs that were hardly trained in previous NeRV models. + +Temporal Up-sampling Block. We evaluate the performance of the TUBs, by replacing them in Ours(T) with the conventional NeRV blocks comprising PixelShuffle and 2D convolution layers. Tab. 5b shows that our TUB is suitable for processing time-extended features than the 2D operations. TUB improves the PSNR around 0.80dB. + +Loss Terms. We conduct ablation tests on the loss function in Eq. (2) using a Yacht sequence. Tab. 5b shows the results, when the terms are turned on or off. Both the terms contribute to the performance improvements. + +Table 5: Ablation studies of the proposed modules and loss terms. +(a) MFU and HFR. + +
DatasetBunnyYacht
MFUHFRPSNRSSIMPSNRSSIM
XX38.000.991130.050.9323
✓X39.200.992431.930.9458
X✓38.210.991131.670.9516
✓✓39.640.992833.430.9647
+ +(b) Loss terms and TUB. + +
L(It,ˆt) L(Ct,ˆCt)PSNRSSIM
✓X33.290.9600
✓✓33.430.9647
Temporal BlockPSNRSSIM
NeRV Block31.910.9532
TUB32.710.9584
+ +![](images/3221893205be72ddee724241e375b20f85a5f3954d4d21c4279666fd52344df7.jpg) +(a) Varying temporal embedding size with fixed target embedding. + +![](images/6783c7c233d46a02b71fa1736bc1c64afe5e40f68dabc43531ce97cbbb178022.jpg) +(b) Varying target embedding size with fixed temporal embedding. +Fig. 8: Regression and interpolation results with different embedding sizes. + +Effect of Size of Model Components. We analyze the performance with different sizes of target $e_t$ and temporal embeddings $e_{\Delta t}$ s, with the size of the decoder varying according to these sizes. Fig. 8 presents the performance variations, exploring the impacts of the ratios. The regression performance is reliable in all the cases, when the ratios are different. In interpolation, the size of the temporal embedding greatly affects the performance, because it hints the temporal relations among video frames. + +# 5 Conclusion + +In this paper, we proposed the spectra-preserving neural representation for video as an efficient baseline framework. SNeRV successfully learned LF components, while reconstructing fine HF details within a limited capacity of parameters, and surpassed the existing NeRV-based methods in various video-related tasks. More compact and efficient architecture to process the high-frequency components considering time-frequency trade-off will be done as a future work. + +# Acknowledgments + +This work was partly supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No.RS-2021-II212068 and RS-2022-00167169) and was partly supported by the NRF grant funded by MSIT (No.NRF-2022R1A2C4002052). + +# References + +1. Arora, S., Du, S., Hu, W., Li, Z., Wang, R.: Fine-grained analysis of optimization and generalization for overparameterized two-layer neural networks. International Conf. on Machine Learning (2019) +2. Bai, Y., Dong, C., Wang, C., Yuan, C.: Ps-nerv: Patch-wise stylized neural representations for videos. In: Proceedings of the IEEE conference on Image Processing (2023) +3. Balle, J., Minnen, D., Singh, S., Hwang, S.J., Johnston, N.: Variational image compression with a scale hyperprior. arXiv preprint arXiv:1802.01436 (2018) +4. Basri, R., Galun, M., Geifman, A., Jacobs, D., Kasten, Y., Kritchman, S.: Frequency bias in neural networks for input of non-uniform density. arXiv preprint arXiv 2003.04560 (2020) +5. Bross, B., Wang, Y.K., Ye, Y., Liu, S., Chen, J., Sullivan, G.J., Ohm, J.R.: Overview of the versatile video coding (vvc) standard and its applications. IEEE Transactions on Circuits and Systems for Video Technology 31(10), 3736-3764 (2021) +6. Cao, Y., Fang, Z., Wu, Y., Zhou, D.X., Gu, Q.: Towards understanding the spectral bias of deep learning. arXiv preprint arXiv:1912.01198 (2019) +7. Chen, H., Gwilliam, M., Lim, S., Shrivastava, A.: Hnerv: A hybrid neural representation for videos. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 10270-10279 (2023) +8. Chen, H., He, B., Wang, H., Ren, Y., Lim, S.N., Shrivastava, A.: Nerv: Neural representations for videos. Advances in Neural Information Processing Systems 34, 21557-21568 (2021) +9. Chen, Y.H., Chen, S.C., Lin, Y.Y., Peng, W.H.: Motif: Learning motion trajectories with local implicit neural functions for continuous space-time video super-resolution. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 23131-23141 (2023) +0. Chen, Z., Chen, Y., Liu, J., Xu, X., Goel, V., Wang, Z., Shi, H., Wang, X.: Videoirr: Learning video implicit neural representation for continuous space-time superresolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2047-2057 (2022) +1. Chen, Z., Zhang, H.: Learning implicit fields for generative shape modeling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5939-5948 (2019) +2. Choi, Y.J., Lee, Y.W., Kim, B.G.: Wavelet attention embedding networks for video super-resolution. In: 2020 25th International conference on pattern recognition (ICPR). pp. 7314-7320. IEEE (2021) +3. Djelouah, A., Campos, J., Schaub-Meyer, S., Schroers, C.: Neural inter-frame compression for video coding. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6421-6429 (2019) +4. He, B., Yang, X., Wang, H., Wu, Z., Chen, H., Huang, S., Ren, Y., Lim, S.N., Shrivastava, A.: Towards scalable neural representation for diverse videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6132-6142 (2023) +5. Hu, Z., Lu, G., Xu, D.: Fvc: A new framework towards deep video compression in feature space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1502-1511 (2021) + +16. Huang, H., He, R., Sun, Z., Tan, T.: Wavelet domain generative adversarial network for multi-scale face hallucination. International Journal of Computer Vision 127(6-7), 763-784 (2019) +17. Jin, B., Hu, Y., Tang, Q., Niu, J., Shi, Z., Han, Y., Li, X.: Exploring spatial-temporal multi-frequency analysis for high-fidelity and temporal-consistency video prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4554-4563 (2020) +18. Jung, H., Hui, Z., Luo, L., Yang, H., Liu, F., Yoo, S., Ranjan, R., Demandolx, D.: Anyflow: Arbitrary scale optical flow with implicit neural representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5455-5465 (2023) +19. Kong, L., Jiang, B., Luo, D., Chu, W., Tai, Y., Wang, C., Yang, J.: Dynamic frame interpolation in wavelet domain. IEEE Transactions on Image Processing (2023) +20. Kwan, H.M., Gao, G., Zhang, F., Gower, A., Bull, D.: Hinerv: Video compression with hierarchical encoding-based neural representation. Advances in Neural Information Processing Systems 36 (2024) +21. Lee, J.C., Rho, D., Ko, J.H., Park, E.: Fferv: Flow-guided frame-wise neural representations for videos. arXiv preprint arXiv:2212.12294 (2022) +22. Lee, J.K., Kim, N., Cho, S., Kang, J.W.: Deep video prediction network-based inter-frame coding in hevc. IEEE ACCESS 8, 95906-95917 (2020) +23. Li, J., Li, B., Lu, Y.: Deep contextual video compression. Advances in Neural Information Processing Systems 34, 18114-18125 (2021) +24. Li, Z., Wang, M., Pi, H., Xu, K., Mei, J., Liu, Y.: E-nerv: Expedite neural video representation with disentangled spatial-temporal context. In: European Conference on Computer Vision. pp. 267-284. Springer (2022) +25. Lu, G., Ouyang, W., Xu, D., Zhang, X., Cai, C., Gao, Z.: Dvc: An end-to-end deep video compression framework. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11006-11015 (2019) +26. Lu, Y., Wang, Z., Liu, M., Wang, H., Wang, L.: Learning spatial-temporal implicit neural representations for event-guided video super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1557-1567 (2023) +27. Maiya, S.R., Girish, S., Ehrlich, M., Wang, H., Lee, K.S., Poirson, P., Wu, P., Wang, C., Shrivastava, A.: Nirvana: Neural implicit representations of videos with adaptive networks and autoregressive patch-wise modeling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14378-14387 (2023) +28. Mallat, S.: A wavelet tour of signal processing. Elsevier (1999) +29. Mercat, A., Viitanen, M., Vanne, J.: Uvg dataset: 50/120fps 4k sequences for video codec analysis and development. In: Proceedings of the 11th ACM Multimedia Systems Conference. pp. 297-302 (2020) +30. Pan, Z., Yi, X., Zhang, Y., Jeon, B., Kwong, S.: Efficient in-loop filtering based on enhanced deep convolutional neural networks for hevc. IEEE Transactions on Image Processing 29, 5352-5366 (2020) +31. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 165-174 (2019) +32. Perazzi, F., Pont-Tuset, J., McWilliams, B., Gool, L.V., Gross, M., Sorkine-Hornung, A.: A benchmark dataset and evaluation methodology for video object + +segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 724-732 (2016) +33. Ramamonjisoa, M., Firman, M., Watson, J., Lepetit, V., Turmukhambetov, D.: Single image depth prediction with wavelet decomposition. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11089-11098 (2021) +34. Rho, D., Cho, J., Ko, J.H., Park, E.: Neural residual flow fields for efficient video representations. In: Proceedings of the Asian Conference on Computer Vision. pp. 3447-3463 (2022) +35. Rho, D., Lee, B., Nam, S., Lee, J.C., Ko, J.H., Park, E.: Masked wavelet representation for compact neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20680-20690 (2023) +36. Ronen, B., Jacobs, D., Kasten, Y., Kritchman, S.: The convergence rate of neural networks for learned functions of different frequencies. Advances in Neural Information Processing Systems 32 (2019) +37. Saragadam, V., LeJeune, D., Tan, J., Balakrishnan, G., Veeraraghavan, A., Baraniuk, R.G.: Wire: Wavelet implicit neural representations. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 18507-18516 (2023) +38. Shi, W., Caballero, J., Huszar, F., Totz, J., Aitken, A.P., Bishop, R., Wang, Z.: Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1874-1883 (2016) +39. Sitzmann, V., Martel, J., Bergman, A., Lindell, D., Wetzstein, G.: Implicit neural representations with periodic activation functions. Advances in neural information processing systems 33, 7462-7473 (2020) +40. Sullivan, G.J., Ohm, J.R., Han, W.J., Wiegand, T.: Overview of the high efficiency video coding (hevc) standard. IEEE Transactions on circuits and systems for video technology 22(12), 1649-1668 (2012) +41. Tancik, M., Srinivasan, P., Mildenhall, B., Fridovich-Keil, S., Raghavan, N., Singhal, U., Ramamoorthi, R., Barron, B., Ng, R.: Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems 33 (2020) +42. Taubman, D.S., Marcellin, M.W., Rabbani, M.: Jpeg2000: Image compression fundamentals, standards and practice. Journal of Electronic Imaging 11(2), 286-287 (2002) +43. Unser, M., Blu, T.: Mathematical properties of the jpeg2000 wavelet filters. IEEE transactions on image processing 12(9), 1080-1090 (2003) +44. Wang, P., Fan, Z., Chen, T., Wang, Z.: Neural implicit dictionary learning via mixture-of-expert training. In: International Conference on Machine Learning. pp. 22613-22624. PMLR (2022) +45. Wiegand, T., Sullivan, G.J., Bjontegaard, G., Luthra, A.: Overview of the h. 264/avc video coding standard. IEEE Transactions on circuits and systems for video technology 13(7), 560-576 (2003) +46. Xin, J., Li, J., Jiang, X., Wang, N., Huang, H., Gao, X.: Wavelet-based dual recursive network for image super-resolution. IEEE Transactions on Neural Networks and Learning Systems 33(2), 707-720 (2020) +47. Yuce, G., Ortiz-Jimenez, G., Besbinar, B., Frossard, P.: A structured dictionary perspective on implicit neural representations. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 19228-19238 (2022) + +48. Zhao, Q., Asif, M.S., Ma, Z.: Dnerv: Modeling inherent dynamics via difference neural representation for videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2031-2040 (2023) \ No newline at end of file diff --git a/2024/SNeRV_ Spectra-preserving Neural Representation for Video/images.zip b/2024/SNeRV_ Spectra-preserving Neural Representation for Video/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..15919a65da483ea04279b08400f0420f0d920d6f --- /dev/null +++ b/2024/SNeRV_ Spectra-preserving Neural Representation for Video/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bc4e855f9d81ab9204187593ae46b9643dbf9539101e6ea8aea553ff0a1e9212 +size 576674 diff --git a/2024/SNeRV_ Spectra-preserving Neural Representation for Video/layout.json b/2024/SNeRV_ Spectra-preserving Neural Representation for Video/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9badfa063bbffeece6d9fd5e699e908c82c1b60a --- /dev/null +++ b/2024/SNeRV_ Spectra-preserving Neural Representation for Video/layout.json @@ -0,0 +1,9466 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 184, + 111, + 430, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 111, + 430, + 148 + ], + "spans": [ + { + "bbox": [ + 184, + 111, + 430, + 148 + ], + "type": "text", + "content": "SNeRV: Spectra-preserving Neural Representation for Video" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 198, + 168, + 415, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 168, + 415, + 181 + ], + "spans": [ + { + "bbox": [ + 198, + 168, + 415, + 181 + ], + "type": "text", + "content": "Jina Kim\\* JiHoo Lee\\* and Je-Won Kang\\*\\*" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 145, + 190, + 468, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 190, + 468, + 224 + ], + "spans": [ + { + "bbox": [ + 145, + 190, + 468, + 224 + ], + "type": "text", + "content": "Department of Electronic and Electrical Engineering and Graduate Program in Smart Factory, Ewha Womans University, Seoul, South Korea qapo0106@ewhain.net, jihoo.lee@lge.com, jewonk@ewha.ac.kr" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 255, + 455, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 255, + 455, + 518 + ], + "spans": [ + { + "bbox": [ + 160, + 255, + 455, + 518 + ], + "type": "text", + "content": "Abstract. Neural representation for video (NeRV), which employs a neural network to parameterize video signals, introduces a novel methodology in video representations. However, existing NeRV-based methods have difficulty in capturing fine spatial details and motion patterns due to spectral bias, in which a neural network learns high-frequency (HF) components at a slower rate than low-frequency (LF) components. In this paper, we propose spectra-preserving NeRV (SNeRV) as a novel approach to enhance implicit video representations by efficiently handling various frequency components. SNeRV uses 2D discrete wavelet transform (DWT) to decompose video into LF and HF features, preserving spatial structures and directly addressing the spectral bias issue. To balance the compactness, we encode only the LF components, while HF components that include fine textures are generated by a decoder. Specialized modules, including a multi-resolution fusion unit (MFU) and a high-frequency restorer (HFR), are integrated into a backbone to facilitate the representation. Furthermore, we extend SNeRV to effectively capture temporal correlations between adjacent video frames, by casting the extension as additional frequency decomposition to a temporal domain. This approach allows us to embed spatio-temporal LF features into the network, using temporally extended up-sampling blocks (TUBs). Experimental results demonstrate that SNeRV outperforms existing NeRV models in capturing fine details and achieves enhanced reconstruction, making it a promising approach in the field of implicit video representations. The codes are available at https://github.com/qwertja/SNeRV." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 529, + 455, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 529, + 455, + 552 + ], + "spans": [ + { + "bbox": [ + 159, + 529, + 455, + 552 + ], + "type": "text", + "content": "Keywords: Implicit neural representation " + }, + { + "bbox": [ + 159, + 529, + 455, + 552 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 529, + 455, + 552 + ], + "type": "text", + "content": " Neural representation for video " + }, + { + "bbox": [ + 159, + 529, + 455, + 552 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 529, + 455, + 552 + ], + "type": "text", + "content": " Wavelet transform" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 573, + 231, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 573, + 231, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 573, + 231, + 586 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 599, + 482, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 599, + 482, + 624 + ], + "spans": [ + { + "bbox": [ + 130, + 599, + 482, + 624 + ], + "type": "text", + "content": "Implicit neural representation for video, also referred to as neural representation for video (NeRV), has attracted considerable attention as a promising method" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 132, + 632, + 254, + 643 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 254, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 254, + 643 + ], + "type": "text", + "content": "\\*Authors contribute equally." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 643, + 235, + 653 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 235, + 653 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 235, + 653 + ], + "type": "text", + "content": "\\*\\* Corresponding author." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 654, + 394, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 654, + 394, + 665 + ], + "spans": [ + { + "bbox": [ + 143, + 654, + 394, + 665 + ], + "type": "text", + "content": "Jihoo Lee is currently with SoC R&D Center, LG Electronics." + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 186 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 186 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 186 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 186 + ], + "type": "image", + "image_path": "9d233f82030302960d1de259d0c259dfc565af22d4ec7691c2c6d635325a24f6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 205, + 482, + 240 + ], + "lines": [ + { + "bbox": [ + 130, + 205, + 482, + 240 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 482, + 240 + ], + "type": "text", + "content": "Fig. 1: Visual comparisons of the reconstructed HF coefficients of previous NeRV methods [7,48] and the proposed method in \"Jockey\" sequence. Our model is designed to efficiently encode fine details, by implicitly circumventing the spectral bias problem." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 266, + 479, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 266, + 479, + 338 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 479, + 338 + ], + "type": "text", + "content": "for video representations [7,8,14,20,21,24,27,48]. NeRV is used to parameterize a video signal with a neural network, in which a space-time coordinate is used as a query, and the network outputs the corresponding RGB value. As a result, a NeRV model itself can serve as a proxy for the video. This approach has proven effective in various tasks, including video compression [20,21], video interpolation [18,34], and video super-resolution [9,10,26]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 339, + 482, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 339, + 482, + 458 + ], + "spans": [ + { + "bbox": [ + 130, + 339, + 482, + 458 + ], + "type": "text", + "content": "Previous studies use pixel-wise [11, 31, 39] or frame-wise mapping models [8] with various positional embedding of space-time coordinates. However, the quality of the reconstructed frame is severely degraded, when a learned decoder fails to capture the spatial details of diverse video contents. HNeRV [7] utilizes a content-adaptive embedding generated by an encoder in addition to a learned decoder. Temporal correlation among adjacent video frames is exploited in the NeRV pipelines using a dynamical system [48] and a decomposition of visual content and motion information [14]. Nevertheless, capturing video representations remains challenging due to the intricate spatio-temporal dynamics in video scenes, while trying to maintain the size of a model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 460, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 482, + 628 + ], + "type": "text", + "content": "The performance of a NeRV model is significantly affected by the characteristics of an input video [7,24,44,48]. When a NeRV approximates an input video, the optimization is carried out to best represent the input within a budget of a model size. However, inherently, the current methods struggle to accurately capture fine spatial details and temporal patterns [7,14,24,48]. According to recent studies on the neural tangent kernel (NTK) [6,36], the learning rate of a neural network is determined by the magnitude of the eigenvalues of the target function. Because natural videos are mostly composed of low-frequency (LF) components, neural networks tend to learn the LF components more quickly, while the learning of high-frequency (HF) components proceeds at a slower rate. Fig. 1 indicates that the existing NeRVs would face challenges in representing HF components due to spectral bias within the vast redundancies. While these behaviors are evident in the learning of implicit neural representations [47], current NeRV methods have not addressed this problem effectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": "In this paper, we propose a spectra-preserving NeRV (SNeRV) that enhances the learning of implicit video representations, by efficiently processing different frequency components. First, we present a SNeRV backbone model to achieve" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "J. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "content": "the goal. We employ a discrete wavelet transform (DWT) to decompose a video into LF and HF features, thereby preserving their spatial structures and directly addressing the issue in the frequency domain. The model capacity is determined with the content-adaptive embeddings and the decoder sizes as [7]. To represent a target in more detail, a decoder requires a larger size of a model or an embedding vector, but it compromises the model compactness. Thus, our encoder embeds only the LF component, exploiting its more suitable characteristics to learn the implicit representations. Meanwhile, the HF components including fine textures are generated by a decoder to circumvent the bias problem. This strategy enables a model to reduce redundancies and achieve a more compact representation with several specialized modules such as a multi-resolution fusion unit (MFU) and a high-frequency restorer (HFR) in a backbone model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 261, + 482, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 261, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 261, + 482, + 403 + ], + "type": "text", + "content": "On top of the backbone, we extend our model to efficiently represent temporal correlations between adjacent video frames, by casting the temporal modeling as spectral decomposition of the video frames. This is accomplished by performing an additional 1D DWT to a temporal dimension, where the motion is considered as HF components. While the backbone lacks the time analysis by maximising the capability of spectral analysis, the temporal extension considers both time and frequency domains in the analysis. By extending the time analysis, the backbone can capture temporal correlation and efficiently perform temporal modeling, well-suited for related tasks such as video interpolation. In experimental results, the backbone and its temporal extension each exhibit a performance trade-off in video regression and interpolation, while maintaining the same model capacity." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 404, + 350, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 404, + 350, + 415 + ], + "spans": [ + { + "bbox": [ + 146, + 404, + 350, + 415 + ], + "type": "text", + "content": "Our paper has major contributions as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 426, + 479, + 531 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 426, + 479, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 426, + 479, + 459 + ], + "spans": [ + { + "bbox": [ + 138, + 426, + 479, + 459 + ], + "type": "text", + "content": "- We propose the spectra-preserving NeRV that enhances implicit video representations by efficiently processing different frequency components through specialized modules such as an MFU and a HFR." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 462, + 479, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 462, + 479, + 483 + ], + "spans": [ + { + "bbox": [ + 138, + 462, + 479, + 483 + ], + "type": "text", + "content": "- We extend the backbone to capture temporal correlations between adjacent video frames, by embedding spatio-temporal LF features into the model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 487, + 479, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 487, + 479, + 531 + ], + "spans": [ + { + "bbox": [ + 138, + 487, + 479, + 531 + ], + "type": "text", + "content": "- We evaluate the performance of the proposed methods on various video processing tasks and datasets, demonstrating their effectiveness. Our model outperforms previous approaches, emphasizing the advantages of mitigating the spectral bias problem in the INR for videos." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 554, + 237, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 554, + 237, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 554, + 237, + 567 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 581, + 502, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 502, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 502, + 667 + ], + "type": "text", + "content": "Implicit Neural Representation (INR) for Video. INR for video exploited multi-layer perceptron (MLP) transforming the continuous spaces of high-dimensional data into density or RGB values [11,39]. However, the computational complexity increased substantially, as they calculated a neural function for each coordinate. To address this issue, NeRV [8] proposed a method to use a temporal index with positional encoding (PE) and restore up-scaled spatial information using a NeRV block with a PixelShuffle method [38]. Based on this, E-NeRV [24] sepa" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 201, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 201, + 91, + 447, + 103 + ], + "type": "text", + "content": "SNeRV: Spectra-preserving Neural Representation for Video" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 115, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 115, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 479, + 140 + ], + "type": "text", + "content": "rated spatial and temporal contexts to improve the representation. PS-NeRV [2] further decomposed a video into a set of patches." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 148, + 481, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 148, + 481, + 257 + ], + "spans": [ + { + "bbox": [ + 132, + 148, + 481, + 257 + ], + "type": "text", + "content": "Pixel-wise [11,31,39] and frame-wise [2,8] NeRV models exploited no spatial prior of a target frame, which compromised the performance. HNeRV [7] proposed a hybrid model using a content-adaptive frame embedding. The embedding was created in the encoder and jointly optimized with the trained decoder parameters. By conveying the content-adaptive spatial features, it enhanced the quality of reconstructed frames. However, as the parameters were fitted to the current frame, the interpolation performance was simultaneously degraded due to the lack of temporal modeling. For this, the motion information in adjacent frames was exploited [14,21,27,48] to achieve better results." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 265, + 482, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 265, + 482, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 265, + 482, + 385 + ], + "type": "text", + "content": "Neural Networks in Frequency Domain. Discrete wavelet transform (DWT) decomposes a local region of an image into different frequency levels [28] and proven to be effective in various computer vision tasks [16, 17, 46]. Multi-level space-time wavelet decomposition was employed to extract motion information using a video prediction network, while preserving spatial details [17]. This technique was applied to image super resolution [46], enhanced with wavelet attention embeddings for video super resolution [12]. The DWT typically yields sparse yet significant non-zero coefficients when representing the original signal, a property that was actively used in various directions, such as monocular depth estimation [33] and video interpolation [19]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 393, + 482, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 393, + 482, + 573 + ], + "spans": [ + { + "bbox": [ + 132, + 393, + 482, + 573 + ], + "type": "text", + "content": "Neural Tangent Kernel. Modeling the learning process of a neural network has been actively studied using kernel regression with NTK analysis [1,4,6,36], in which the training rate is decided by the eigenvalues of the NTK matrix. The eigenvalue spectrum of an NTK matrix explains the spectral bias of a neural network towards training speed of HF functions [4]. As over-parameterized neural networks would be poorly suited for learning HF functions [1], mitigating the bias is essential for improving the performance of neural networks [39, 41]. Several studies have attempted to alleviate the bias using various input mapping of a coordinate. Sinusoidal functions were modulated into an activation function [39]. Gabor-wavelets were introduced for flexible representations in space and frequency domains [37]. However, such efforts have not been explored for NeRVs to circumvent the spectral bias. Motivated by the NTK, we aim to develop compact implicit representations of diverse frequency components with a reduced number of parameters. We use the inherent sparsity of the HF wavelet transformed components in videos." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 581, + 482, + 666 + ], + "type": "text", + "content": "Video Compression. In recent years, deep learning-based video coding techniques were actively studied as alternatives to traditional CODECs [5,40,45]. This approach involved substituting various coding modules, such as frame restoration [30], motion prediction [13,22], and transform and entropy coding [3], with neural networks. DVC [25] proposed an end-to-end video coding model. The NeRV models [8,20,21,27] converted the compression tasks into model compression, using weight pruning, quantization, and entropy coding." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "J. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 148, + 113, + 462, + 222 + ], + "blocks": [ + { + "bbox": [ + 148, + 113, + 462, + 222 + ], + "lines": [ + { + "bbox": [ + 148, + 113, + 462, + 222 + ], + "spans": [ + { + "bbox": [ + 148, + 113, + 462, + 222 + ], + "type": "image", + "image_path": "9ce82f8f63d7d435fd71dd1ce7075cfb8fc8f26c714795185785758566d35a82.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 233, + 482, + 278 + ], + "lines": [ + { + "bbox": [ + 130, + 233, + 482, + 278 + ], + "spans": [ + { + "bbox": [ + 130, + 233, + 482, + 278 + ], + "type": "text", + "content": "Fig. 2: SNeRV backbone encoder and decoder architectures. The encoder applies 2D DWT to extract LF and HF features and embeds only the LF feature to save parameters. The decoder uses MFU and HFR to efficiently process the LF and HF features. CT and RB refer to transposed convolution and residual blocks, respectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 299, + 261, + 313 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 299, + 261, + 313 + ], + "spans": [ + { + "bbox": [ + 132, + 299, + 261, + 313 + ], + "type": "text", + "content": "3 Proposed Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 323, + 208, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 323, + 208, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 323, + 208, + 335 + ], + "type": "text", + "content": "3.1 Overview" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 342, + 482, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 474 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 474 + ], + "type": "text", + "content": "Our network is designed to enhance the learning of implicit representations of video by mitigating spectral bias. We use a DWT to decompose video into LF and HF components to use an analysis-and-synthesis approach. The LF features with little variation across frames are embedded through the network, because this is inherently suitable for learning a NeRV model. The HF features are restored by a decoder. Specifically, the decoder sequentially forms the HF features with the LF ones as prior rather than directly learning the HF details. This implicit mapping enables the reallocation of the parameters to concentrate on representing LF features. Both features are optimized to consider the analysis-and-synthesis scheme, thus allowing for a more efficient representation of the HF textures and motions, based upon the stable base provided by the LF contents." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 474, + 482, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 558 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 558 + ], + "type": "text", + "content": "These schemes are implemented as our backbone model of the SNeRV. It employs the HNeRV framework [7], in which the implicit representations are composed of a set of embedding vectors and learned decoder parameters. Based on this, we explain the novel design principles of the encoder and decoder, as shown in Fig. 2. On top of that, we extend the backbone to effectively represent temporal contexts of video. We explain the extended modules to accommodate these schemes step-by-step in Fig. 3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 558, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 605 + ], + "type": "text", + "content": "The differences between the backbone and its temporal extension not only clearly demonstrate their distinct architectural characteristics but also assist the performance analysis in various video processing tasks, while the backbone exhibits its superiority to the conventional methods." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 622, + 317, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 317, + 634 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 317, + 634 + ], + "type": "text", + "content": "3.2 SNeRV Backbone Architecture" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Encoder Design. Given a video sequence " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{V} = \\{I_0, I_1, \\dots, I_{T-1}\\}" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": ", in the backbone, an encoder conducts a 2D DWT along the " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "x - y" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": " spatial axes to decompose" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 201, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 201, + 91, + 448, + 103 + ], + "type": "text", + "content": "SNeRV: Spectra-preserving Neural Representation for Video" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 200 + ], + "type": "text", + "content": "each video frame " + }, + { + "bbox": [ + 130, + 115, + 482, + 200 + ], + "type": "inline_equation", + "content": "I_{t}\\in \\mathbb{R}^{H\\times W\\times 3}" + }, + { + "bbox": [ + 130, + 115, + 482, + 200 + ], + "type": "text", + "content": " into multiple wavelet sub-bands. It produces four sets of wavelet coefficients in " + }, + { + "bbox": [ + 130, + 115, + 482, + 200 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{H / 2\\times W / 2\\times 3}" + }, + { + "bbox": [ + 130, + 115, + 482, + 200 + ], + "type": "text", + "content": ", including " + }, + { + "bbox": [ + 130, + 115, + 482, + 200 + ], + "type": "inline_equation", + "content": "C_{LL}" + }, + { + "bbox": [ + 130, + 115, + 482, + 200 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 115, + 482, + 200 + ], + "type": "inline_equation", + "content": "C_{LH}" + }, + { + "bbox": [ + 130, + 115, + 482, + 200 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 115, + 482, + 200 + ], + "type": "inline_equation", + "content": "C_{HL}" + }, + { + "bbox": [ + 130, + 115, + 482, + 200 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 115, + 482, + 200 + ], + "type": "inline_equation", + "content": "C_{HH}" + }, + { + "bbox": [ + 130, + 115, + 482, + 200 + ], + "type": "text", + "content": ", at the current time-step. " + }, + { + "bbox": [ + 130, + 115, + 482, + 200 + ], + "type": "inline_equation", + "content": "C_{LL}" + }, + { + "bbox": [ + 130, + 115, + 482, + 200 + ], + "type": "text", + "content": " contains the LF component, which often appears as a blurred and smoothed version of the original frame. The HF components, represented by " + }, + { + "bbox": [ + 130, + 115, + 482, + 200 + ], + "type": "inline_equation", + "content": "C_{LH}" + }, + { + "bbox": [ + 130, + 115, + 482, + 200 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 115, + 482, + 200 + ], + "type": "inline_equation", + "content": "C_{HL}" + }, + { + "bbox": [ + 130, + 115, + 482, + 200 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 115, + 482, + 200 + ], + "type": "inline_equation", + "content": "C_{HH}" + }, + { + "bbox": [ + 130, + 115, + 482, + 200 + ], + "type": "text", + "content": ", encode the horizontal, vertical, and diagonal details of the image. For simplicity, we utilize the Haar WT [28], where the low pass filter is " + }, + { + "bbox": [ + 130, + 115, + 482, + 200 + ], + "type": "inline_equation", + "content": "\\phi_L = [1 / \\sqrt{2},1 / \\sqrt{2} ]" + }, + { + "bbox": [ + 130, + 115, + 482, + 200 + ], + "type": "text", + "content": ", and the high pass filter is " + }, + { + "bbox": [ + 130, + 115, + 482, + 200 + ], + "type": "inline_equation", + "content": "\\phi_H = [1 / \\sqrt{2}, - 1 / \\sqrt{2} ]" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 482, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 482, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 482, + 272 + ], + "type": "text", + "content": "We input only the " + }, + { + "bbox": [ + 130, + 200, + 482, + 272 + ], + "type": "inline_equation", + "content": "C_{LL}" + }, + { + "bbox": [ + 130, + 200, + 482, + 272 + ], + "type": "text", + "content": " into a down-sampling block (DB) to generate a content-adaptive and time-specific embedding " + }, + { + "bbox": [ + 130, + 200, + 482, + 272 + ], + "type": "inline_equation", + "content": "e_t" + }, + { + "bbox": [ + 130, + 200, + 482, + 272 + ], + "type": "text", + "content": " whose dimension is " + }, + { + "bbox": [ + 130, + 200, + 482, + 272 + ], + "type": "inline_equation", + "content": "16 \\times 2 \\times 4" + }, + { + "bbox": [ + 130, + 200, + 482, + 272 + ], + "type": "text", + "content": ". We use the same DB blocks as [7] and keep the same dimension of " + }, + { + "bbox": [ + 130, + 200, + 482, + 272 + ], + "type": "inline_equation", + "content": "e_t" + }, + { + "bbox": [ + 130, + 200, + 482, + 272 + ], + "type": "text", + "content": " for fair comparisons. Due to the 2D DWT, the reduced original input size alleviates the burden of parameter sizes on the decoder side. " + }, + { + "bbox": [ + 130, + 200, + 482, + 272 + ], + "type": "inline_equation", + "content": "C_{LH}" + }, + { + "bbox": [ + 130, + 200, + 482, + 272 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 200, + 482, + 272 + ], + "type": "inline_equation", + "content": "C_{HL}" + }, + { + "bbox": [ + 130, + 200, + 482, + 272 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 200, + 482, + 272 + ], + "type": "inline_equation", + "content": "C_{HH}" + }, + { + "bbox": [ + 130, + 200, + 482, + 272 + ], + "type": "text", + "content": " are not explicitly embedded but synthesized in the decoder." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 272, + 482, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 272, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 482, + 390 + ], + "type": "text", + "content": "Decoder Design. The decoder involves the reconstruction process using " + }, + { + "bbox": [ + 130, + 272, + 482, + 390 + ], + "type": "inline_equation", + "content": "\\mathbf{e} = \\{e_t\\}" + }, + { + "bbox": [ + 130, + 272, + 482, + 390 + ], + "type": "text", + "content": " and learned decoder parameters. It consists of three components, including up-sampling blocks (UBs), MFU, and HFR. Firstly, UB is used to progressively increase the resolution of the embedding and reconstruct the spectral information of a target time step. We employ five NeRV blocks [8,38] for the UBs. Then, the MFU fuses features from different resolutions to enrich the LF features. The HFR restores " + }, + { + "bbox": [ + 130, + 272, + 482, + 390 + ], + "type": "inline_equation", + "content": "C_{LH}" + }, + { + "bbox": [ + 130, + 272, + 482, + 390 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 272, + 482, + 390 + ], + "type": "inline_equation", + "content": "C_{HL}" + }, + { + "bbox": [ + 130, + 272, + 482, + 390 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 272, + 482, + 390 + ], + "type": "inline_equation", + "content": "C_{HH}" + }, + { + "bbox": [ + 130, + 272, + 482, + 390 + ], + "type": "text", + "content": " from the reconstructed " + }, + { + "bbox": [ + 130, + 272, + 482, + 390 + ], + "type": "inline_equation", + "content": "C_{LL}" + }, + { + "bbox": [ + 130, + 272, + 482, + 390 + ], + "type": "text", + "content": ". The target frame is reconstructed from an inverse DWT (IDWT), by combining all the frequency components. Fig. 2 presents the decoder architecture. We explain the MFU and HFR modules in detail as follows." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 392, + 482, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 482, + 475 + ], + "type": "text", + "content": "Multi-resolution Fusion Unit (MFU). UBs are used to sequentially reconstruct the LF features and recover the original dimension. However, using only UBs has limits in improving the output quality of " + }, + { + "bbox": [ + 130, + 392, + 482, + 475 + ], + "type": "inline_equation", + "content": "C_{LL}" + }, + { + "bbox": [ + 130, + 392, + 482, + 475 + ], + "type": "text", + "content": ". To address this, we develop the MFU with a coarse-to-fine structure, in which the blurred outcomes are progressively refined with multi-scaled features during up-sampling. The MFU comprises two MF blocks (MFBs), each consisting of one transposed convolution and six residual blocks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 476, + 482, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 476, + 482, + 512 + ], + "spans": [ + { + "bbox": [ + 130, + 476, + 482, + 512 + ], + "type": "text", + "content": "In each MFB, the current feature " + }, + { + "bbox": [ + 130, + 476, + 482, + 512 + ], + "type": "inline_equation", + "content": "m_{i}" + }, + { + "bbox": [ + 130, + 476, + 482, + 512 + ], + "type": "text", + "content": " is up-sampled using the transposed convolution, and then its concatenation with the outcomes of the " + }, + { + "bbox": [ + 130, + 476, + 482, + 512 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 476, + 482, + 512 + ], + "type": "text", + "content": "-th UB, denoted by " + }, + { + "bbox": [ + 130, + 476, + 482, + 512 + ], + "type": "inline_equation", + "content": "u_{i}" + }, + { + "bbox": [ + 130, + 476, + 482, + 512 + ], + "type": "text", + "content": ", goes through the residual blocks, which is mathematically expressed as," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 216, + 536, + 480, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 536, + 480, + 548 + ], + "spans": [ + { + "bbox": [ + 216, + 536, + 480, + 548 + ], + "type": "interline_equation", + "content": "m _ {i + 1} = \\mathcal {F} _ {R B} \\left(\\mathcal {F} _ {C T} \\left(m _ {i}\\right) \\circ u _ {i + 3}\\right), \\quad i = 0, 1, \\tag {1}", + "image_path": "23e3fac66e1efdf5c7d160013a1e37c143b82b63b524d727fc6c05819282ba23.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 557, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 604 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 557, + 482, + 604 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{CT}" + }, + { + "bbox": [ + 130, + 557, + 482, + 604 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 557, + 482, + 604 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{RB}" + }, + { + "bbox": [ + 130, + 557, + 482, + 604 + ], + "type": "text", + "content": " are the operations of the transposed convolution and residual blocks, respectively, " + }, + { + "bbox": [ + 130, + 557, + 482, + 604 + ], + "type": "inline_equation", + "content": "\\circ" + }, + { + "bbox": [ + 130, + 557, + 482, + 604 + ], + "type": "text", + "content": " is a concatenation, and " + }, + { + "bbox": [ + 130, + 557, + 482, + 604 + ], + "type": "inline_equation", + "content": "m_0 = u_2" + }, + { + "bbox": [ + 130, + 557, + 482, + 604 + ], + "type": "text", + "content": ". We use " + }, + { + "bbox": [ + 130, + 557, + 482, + 604 + ], + "type": "inline_equation", + "content": "u_i" + }, + { + "bbox": [ + 130, + 557, + 482, + 604 + ], + "type": "text", + "content": " from the last three UBs and two MFBs, empirically considering the trade-off between learning ability and the size of parameters." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "High-Frequency Restorer (HFR). Sparse representations of HF wavelet coefficients have been exploited in wavelet video compression and various computer vision researches [33,35,42,43]. The HF coefficients in the flat regions of an image are nearly zero, whereas significant coefficients are found in the edges, which are essential for achieving a high quality image. Motivated by this, the HFR is" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "J. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 211 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 211 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 211 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 211 + ], + "type": "image", + "image_path": "ebfeaa92fc242be4e6215a8340337109603337e6051a925a858ffbacf4ae1c93.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 230, + 482, + 253 + ], + "lines": [ + { + "bbox": [ + 131, + 230, + 482, + 253 + ], + "spans": [ + { + "bbox": [ + 131, + 230, + 482, + 253 + ], + "type": "text", + "content": "Fig. 3: Temporal extension from the backbone: the encoder uses additional 1D DWT to generate spatio-temporal embeddings. The decoder uses TUBs to address the features." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 277, + 482, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 482, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 482, + 361 + ], + "type": "text", + "content": "designed to restore the high spatial-frequency details from the composite of LF features generated by the MFU. We employ three HFR blocks, with each comprising two convolution layers. These HFR blocks are responsible for generating " + }, + { + "bbox": [ + 130, + 277, + 482, + 361 + ], + "type": "inline_equation", + "content": "C_{LH}" + }, + { + "bbox": [ + 130, + 277, + 482, + 361 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 277, + 482, + 361 + ], + "type": "inline_equation", + "content": "C_{HL}" + }, + { + "bbox": [ + 130, + 277, + 482, + 361 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 277, + 482, + 361 + ], + "type": "inline_equation", + "content": "C_{HH}" + }, + { + "bbox": [ + 130, + 277, + 482, + 361 + ], + "type": "text", + "content": " from the results of " + }, + { + "bbox": [ + 130, + 277, + 482, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{MFU}" + }, + { + "bbox": [ + 130, + 277, + 482, + 361 + ], + "type": "text", + "content": ". This process is shown in Fig. 2. Because the synthesis process requires less parameters than the direct learning of the HF coefficients, the decoder can save the parameters and allow for more parameters to represent the LF ones." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 380, + 303, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 380, + 303, + 392 + ], + "spans": [ + { + "bbox": [ + 131, + 380, + 303, + 392 + ], + "type": "text", + "content": "3.3 SNeRV Temporal Extension" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 402, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 668 + ], + "type": "text", + "content": "Inspired by [48], we extend the backbone to analyse the effect of considering temporal redundancies. We apply 3D " + }, + { + "bbox": [ + 130, + 402, + 482, + 668 + ], + "type": "inline_equation", + "content": "(2 + 1\\mathrm{D})" + }, + { + "bbox": [ + 130, + 402, + 482, + 668 + ], + "type": "text", + "content": " DWT to the image sequence " + }, + { + "bbox": [ + 130, + 402, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 130, + 402, + 482, + 668 + ], + "type": "text", + "content": " to decompose the LF and HF components along temporal axis in addition to the existing 2D DWT. Specifically, we use two sets of adjacent frames, including " + }, + { + "bbox": [ + 130, + 402, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\{I_{t - 1},I_t\\}" + }, + { + "bbox": [ + 130, + 402, + 482, + 668 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 402, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\{I_t,I_{t + 1}\\}" + }, + { + "bbox": [ + 130, + 402, + 482, + 668 + ], + "type": "text", + "content": " as a backward and a forward set, respectively. The 3D DWT is applied to the two image sets, in which 2D DWTs are first applied, and then 1D DWT is applied to the temporal axis. Accordingly, it produces two spatio-temporal LF components corresponding to the sets. Then, we acquire temporal embeddings " + }, + { + "bbox": [ + 130, + 402, + 482, + 668 + ], + "type": "inline_equation", + "content": "e_{\\Delta t_b}" + }, + { + "bbox": [ + 130, + 402, + 482, + 668 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 402, + 482, + 668 + ], + "type": "inline_equation", + "content": "e_{\\Delta t_f}" + }, + { + "bbox": [ + 130, + 402, + 482, + 668 + ], + "type": "text", + "content": " through DBs to represent the LF components of the backward and forward set, respectively. In the decoder, " + }, + { + "bbox": [ + 130, + 402, + 482, + 668 + ], + "type": "inline_equation", + "content": "e_{\\Delta t}" + }, + { + "bbox": [ + 130, + 402, + 482, + 668 + ], + "type": "text", + "content": "s are jointly up-sampled with " + }, + { + "bbox": [ + 130, + 402, + 482, + 668 + ], + "type": "inline_equation", + "content": "e_t" + }, + { + "bbox": [ + 130, + 402, + 482, + 668 + ], + "type": "text", + "content": ". For this, we replace the last three UBs with temporally extended UBs (TUBs), since they are used as inputs of the MFU as in the backbone. The overall architecture with temporal extension is shown in Fig. 3. Temporally Extended Up-sampling Block (TUB). We propose the TUBs to process " + }, + { + "bbox": [ + 130, + 402, + 482, + 668 + ], + "type": "inline_equation", + "content": "e_{\\Delta t_b}" + }, + { + "bbox": [ + 130, + 402, + 482, + 668 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 402, + 482, + 668 + ], + "type": "inline_equation", + "content": "e_{\\Delta t_f}" + }, + { + "bbox": [ + 130, + 402, + 482, + 668 + ], + "type": "text", + "content": " in addition to " + }, + { + "bbox": [ + 130, + 402, + 482, + 668 + ], + "type": "inline_equation", + "content": "e_t" + }, + { + "bbox": [ + 130, + 402, + 482, + 668 + ], + "type": "text", + "content": " using 3D convolution operations. The three inputs are up-sampled using a transposed convolution and then passed through two 3D convolution layers. The computational complexity of the 3D convolution can be significantly reduced using 2D operations. More details of our implementation can be found in the supplementary. During the consecutive 3D convolutions, the channel dimension increases by a factor of 2 and then reduces back to expand the implicit capacity of the network. In the last TUB, all three components reflecting the different time steps are merged to output the" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 201, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 201, + 91, + 448, + 102 + ], + "type": "text", + "content": "SNeRV: Spectra-preserving Neural Representation for Video" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": "LF feature for a target at " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": ". This operation allows to exploit temporal information in adjacent frames, resulting in an enhanced expressive capacity of the network over time. The outputs " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "C_{LL}, C_{LH}, C_{HL}" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "C_{HH}" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": " for the target at " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": " are obtained by passing the results of the last TUB through the MFU and HFR, as described in Sec. 3.2, obtaining the final output by applying 2D IDWT." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 194, + 240, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 240, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 240, + 205 + ], + "type": "text", + "content": "3.4 Implementation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 214, + 482, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 214, + 482, + 370 + ], + "spans": [ + { + "bbox": [ + 130, + 214, + 482, + 370 + ], + "type": "text", + "content": "The total size of the decoder parameters and embedding vectors determine the model capacity for representation. Because the performance would vary with the sizes and the extra embeddings are generated at the expense of the decoder size, we carefully explain the modified parameters to keep the total size. We set the size of " + }, + { + "bbox": [ + 130, + 214, + 482, + 370 + ], + "type": "inline_equation", + "content": "e_t" + }, + { + "bbox": [ + 130, + 214, + 482, + 370 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 214, + 482, + 370 + ], + "type": "inline_equation", + "content": "16 \\times 2 \\times 4" + }, + { + "bbox": [ + 130, + 214, + 482, + 370 + ], + "type": "text", + "content": " and use five UBs as in [7] for fair comparisons. We use " + }, + { + "bbox": [ + 130, + 214, + 482, + 370 + ], + "type": "inline_equation", + "content": "C_{LL}" + }, + { + "bbox": [ + 130, + 214, + 482, + 370 + ], + "type": "text", + "content": " as input that is recovered to the original dimension using IDWT. Because the IDWT replaces a UB in the last stage, we save some parameters by setting a size of a stride to the half of [7, 48]. This parsimony allows the parameters for MFU and HFR. In the temporal extension, we use a dimension of " + }, + { + "bbox": [ + 130, + 214, + 482, + 370 + ], + "type": "inline_equation", + "content": "3 \\times 20 \\times 40" + }, + { + "bbox": [ + 130, + 214, + 482, + 370 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 130, + 214, + 482, + 370 + ], + "type": "inline_equation", + "content": "e_{\\Delta t}" + }, + { + "bbox": [ + 130, + 214, + 482, + 370 + ], + "type": "text", + "content": ". Because the size is slightly larger than that of " + }, + { + "bbox": [ + 130, + 214, + 482, + 370 + ], + "type": "inline_equation", + "content": "e_t" + }, + { + "bbox": [ + 130, + 214, + 482, + 370 + ], + "type": "text", + "content": ", in our implementation, the channel width of a decoder is reduced to maintain the overall size. We analyze the trade-offs with various temporal embedding sizes in Sec. 4.4. More details such as stride and channel sizes can be found in the supplementary materials." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 388, + 230, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 388, + 230, + 399 + ], + "spans": [ + { + "bbox": [ + 132, + 388, + 230, + 399 + ], + "type": "text", + "content": "3.5 Loss Function" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 407, + 481, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 481, + 433 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 481, + 433 + ], + "type": "text", + "content": "To generate visually pleasing frames with the restored HF features, we combine both L1 and SSIM loss functions over frames, given as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 182, + 441, + 481, + 473 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 441, + 481, + 473 + ], + "spans": [ + { + "bbox": [ + 182, + 441, + 481, + 473 + ], + "type": "interline_equation", + "content": "L \\left(I _ {t}, \\hat {I} _ {t}\\right) = \\frac {1}{T} \\sum_ {t = 0} ^ {T - 1} \\alpha \\left\\| I _ {t} - \\hat {I} _ {t} \\right\\| _ {1} + (1 - \\alpha) \\left(1 - \\operatorname {S S I M} \\left(I _ {t}, \\hat {I} _ {t}\\right)\\right), \\tag {2}", + "image_path": "4783dc6571a87734d2880b7bde25b4b414c37b03d5ad6d1b08d00c0dfba78d3a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 483, + 480, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 483, + 480, + 507 + ], + "spans": [ + { + "bbox": [ + 130, + 483, + 480, + 507 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 483, + 480, + 507 + ], + "type": "inline_equation", + "content": "\\hat{I}_t" + }, + { + "bbox": [ + 130, + 483, + 480, + 507 + ], + "type": "text", + "content": " is the reconstructed frame, and " + }, + { + "bbox": [ + 130, + 483, + 480, + 507 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 130, + 483, + 480, + 507 + ], + "type": "text", + "content": " is the ground truth at time step " + }, + { + "bbox": [ + 130, + 483, + 480, + 507 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 483, + 480, + 507 + ], + "type": "text", + "content": ". SSIM " + }, + { + "bbox": [ + 130, + 483, + 480, + 507 + ], + "type": "inline_equation", + "content": "(\\cdot)" + }, + { + "bbox": [ + 130, + 483, + 480, + 507 + ], + "type": "text", + "content": " calculates the SSIM score. " + }, + { + "bbox": [ + 130, + 483, + 480, + 507 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 483, + 480, + 507 + ], + "type": "text", + "content": " is set to 0.7 during training." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 508, + 481, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 508, + 481, + 532 + ], + "spans": [ + { + "bbox": [ + 130, + 508, + 481, + 532 + ], + "type": "text", + "content": "Moreover, we compute the loss on a set of the four wavelet coefficients " + }, + { + "bbox": [ + 130, + 508, + 481, + 532 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 130, + 508, + 481, + 532 + ], + "type": "text", + "content": " for better frequency reconstruction. Our final loss objective is formulated as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 242, + 541, + 481, + 556 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 541, + 481, + 556 + ], + "spans": [ + { + "bbox": [ + 242, + 541, + 481, + 556 + ], + "type": "interline_equation", + "content": "L _ {t o t a l} = L (I _ {t}, \\hat {I} _ {t}) + L (\\mathbf {C} _ {t}, \\hat {\\mathbf {C}} _ {t}), \\tag {3}", + "image_path": "295f8c2ae56d53747c1ae8edc056d12b458526fc1b3072fa2574a5107ee3247c.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 563, + 480, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 563, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 563, + 480, + 578 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 563, + 480, + 578 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{C}}_t" + }, + { + "bbox": [ + 130, + 563, + 480, + 578 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 563, + 480, + 578 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_t" + }, + { + "bbox": [ + 130, + 563, + 480, + 578 + ], + "type": "text", + "content": " are the output coefficients and the ground truth, respectively." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 596, + 230, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 596, + 230, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 596, + 230, + 609 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 621, + 273, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 273, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 273, + 633 + ], + "type": "text", + "content": "4.1 Experimental Settings" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "We conduct experiments on Bunny, UVG [29], and DAVIS [32] datasets. Bunny has 132 frames of " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "720 \\times 1280" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": " resolution. UVG is composed of 7 video sequences" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "J. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 146, + 481, + 272 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 1: Performance comparisons (dB) in video regression in UVG and DAVIS datasets with different sizes of resolutions." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 146, + 481, + 272 + ], + "lines": [ + { + "bbox": [ + 132, + 146, + 481, + 272 + ], + "spans": [ + { + "bbox": [ + 132, + 146, + 481, + 272 + ], + "type": "table", + "html": "
DatasetUVGDAVIS
sizemethodBeautyBosphBeeJockeyReadyShakeYachtAvg.BikeSwanDanceCamelCarCowsTwirlAvg.
960NeRV [8]33.5633.2837.9331.3126.9532.6728.7432.0625.7727.2225.5923.8228.0921.7525.5125.39
E-NeRV [24]34.1232.9635.4531.9230.1432.9129.2232.3927.1228.7027.8725.6828.1824.2426.8126.94
HNeRV [7]34.1236.3838.9633.5829.9734.1531.4034.0830.7531.7728.8426.5933.5023.8528.5529.12
DNeRV [48]34.0235.3938.7134.4530.4033.5730.8533.9130.0432.6129.5226.9732.9723.6428.8929.23
Ours(B)34.4337.7438.8035.6933.1634.5233.4335.4033.2933.8331.4028.6835.7925.1430.4131.22
Ours(T)34.1437.6439.1136.1733.2134.4132.7135.3431.9033.0831.2630.1434.3327.3829.9331.15
480NeRV [8]34.5935.4238.9034.4330.1631.8233.5034.1228.9132.7728.9530.2632.0824.7727.5729.33
E-NeRV [24]35.0836.6339.2834.6231.0632.4333.9834.7329.0333.3529.2130.9834.9925.7626.9230.04
HNeRV [7]35.6838.7339.7237.1234.3834.2437.6336.7936.0739.4034.6235.2038.6129.8933.3435.30
DNeRV [48]35.3038.4039.5037.5634.1833.8337.1436.5635.6639.2534.2235.1438.3529.0833.2434.99
Ours(B)36.0140.0939.5238.3636.9035.1538.9737.8637.3240.5736.0236.6839.0831.7835.1636.66
Ours(T)35.7840.2439.6138.2136.7734.7038.3037.6636.4038.9435.5936.0438.1732.1134.1235.91
", + "image_path": "2bec3361464c6ad1d1997492b1d24e7e9025db32467406c18000be467ea04105.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 295, + 480, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 295, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 295, + 480, + 342 + ], + "type": "text", + "content": "with 300 or 600 frames of " + }, + { + "bbox": [ + 130, + 295, + 480, + 342 + ], + "type": "inline_equation", + "content": "1080 \\times 1920" + }, + { + "bbox": [ + 130, + 295, + 480, + 342 + ], + "type": "text", + "content": ". We utilize 20 subsets of DAVIS datasets of " + }, + { + "bbox": [ + 130, + 295, + 480, + 342 + ], + "type": "inline_equation", + "content": "1080 \\times 1920" + }, + { + "bbox": [ + 130, + 295, + 480, + 342 + ], + "type": "text", + "content": " sized frames. We used the same center-crop as in [7,48]. For training, we utilized an Adam optimizer and a learning rate of " + }, + { + "bbox": [ + 130, + 295, + 480, + 342 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-3}" + }, + { + "bbox": [ + 130, + 295, + 480, + 342 + ], + "type": "text", + "content": " with a cosine learning rate scheduling. The batch size is set to 1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 342, + 481, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 481, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 481, + 486 + ], + "type": "text", + "content": "We use a peak signal-to-noise ratio (PSNR) and a structural similarity index measure (SSIM) for evaluation and bits-per-pixel (bpp) for compression. If none of the conditions are specified, we use 3M-sized models trained for 300 epochs in the experiments. We compare our backbone, denoted as \"Ours(B)\", and the temporal extension, denoted as \"Ours(T)\", with the state-of-the-art methods such as NeRV [8], E-NeRV [24], HNeRV [7], and DNeRV [48]. We train all the models, utilizing the original codes provided by the authors, and report the testing results under the same conditions. To ensure a fair comparison, we maintained the total capacity of the experimented models. To reduce the original model size from 3.8M to 3M in [48], the channel width was decreased from 92 to 70 which varies depending on the embedding sizes, resulting in a slight drop in performance compared to the original results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 503, + 349, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 503, + 349, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 503, + 349, + 514 + ], + "type": "text", + "content": "4.2 Performance Evaluation and Analysis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 521, + 481, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 481, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 481, + 640 + ], + "type": "text", + "content": "Video Regression. Tab. 1 presents the performance of the tested methods in the tasks of video regression in UVG and DAVIS datasets with various resolutions. In the comparisons, our method presents the highest performance on average among the tested methods. In UVG dataset of " + }, + { + "bbox": [ + 130, + 521, + 481, + 640 + ], + "type": "inline_equation", + "content": "960 \\times 1920" + }, + { + "bbox": [ + 130, + 521, + 481, + 640 + ], + "type": "text", + "content": " resolutions, Ours(B) improves the performance approximately by " + }, + { + "bbox": [ + 130, + 521, + 481, + 640 + ], + "type": "inline_equation", + "content": "1.32\\mathrm{dB} \\sim 1.46\\mathrm{dB}" + }, + { + "bbox": [ + 130, + 521, + 481, + 640 + ], + "type": "text", + "content": " over HNeRV and DNeRV, respectively. In DAVIS datasets, our methods outperform the previous methods by " + }, + { + "bbox": [ + 130, + 521, + 481, + 640 + ], + "type": "inline_equation", + "content": "2.17\\mathrm{dB} \\sim 2.35\\mathrm{dB}" + }, + { + "bbox": [ + 130, + 521, + 481, + 640 + ], + "type": "text", + "content": ". While we show the results from 7 video samples of DAVIS subsets, more results are reported in the supplementary. We observe similar results in " + }, + { + "bbox": [ + 130, + 521, + 481, + 640 + ], + "type": "inline_equation", + "content": "480 \\times 960" + }, + { + "bbox": [ + 130, + 521, + 481, + 640 + ], + "type": "text", + "content": " resolutions. These results imply that our methods significantly improve the performance in video regression." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 642, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 642, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 642, + 481, + 665 + ], + "type": "text", + "content": "Tab. 2 exhibits the performance analysis with different model sizes and training epochs in Bunny dataset. We present the results in Tab. 2a for model sizes" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "text", + "content": "SNeRV: Spectra-preserving Neural Representation for Video" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 150, + 290, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 290, + 168 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 290, + 168 + ], + "type": "text", + "content": "(a) Performance evaluation in PSNR (dB) with different model sizes." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 148, + 178, + 276, + 251 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "text", + "content": "Table 2: Performance comparisons (dB) in video regression on Bunny dataset with different model sizes and training epochs." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 148, + 178, + 276, + 251 + ], + "lines": [ + { + "bbox": [ + 148, + 178, + 276, + 251 + ], + "spans": [ + { + "bbox": [ + 148, + 178, + 276, + 251 + ], + "type": "table", + "html": "
Model size0.35M0.75M1.5M3.0M
NeRV [8]28.1629.8331.8833.01
E-NeRV [24]27.7829.1231.6536.89
HNeRV [7]30.6732.8135.5738.06
DNeRV [48]29.3631.9735.1837.55
Ours(B)30.8833.2536.7639.64
Ours(T)30.4033.1936.6839.09
", + "image_path": "976ac843a1c8b25650abe443c4fd87580a30bebee27fa5581c46f7e4482ec0ed.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 302, + 150, + 477, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 150, + 477, + 168 + ], + "spans": [ + { + "bbox": [ + 302, + 150, + 477, + 168 + ], + "type": "text", + "content": "(b) Performance evaluation with training epochs and model complexity." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 305, + 178, + 472, + 249 + ], + "blocks": [ + { + "bbox": [ + 305, + 178, + 472, + 249 + ], + "lines": [ + { + "bbox": [ + 305, + 178, + 472, + 249 + ], + "spans": [ + { + "bbox": [ + 305, + 178, + 472, + 249 + ], + "type": "table", + "html": "
Epochs300600120018002400MACs
NeRV [8]33.0133.5433.9634.1634.21100.9G
E-NeRV [24]36.8938.6939.7139.9039.96103.3G
HNeRV [7]38.0639.2739.8740.1040.1760.89G
DNeRV [48]37.5538.1738.5538.6838.7448.39G
Ours(B)39.6440.4040.8240.9440.9990.49G
Ours(T)39.0939.9540.4140.6340.68181.8G
", + "image_path": "fff0aa7e98be4963273f8dcab193ad949c1e04ef4928ba293657aa3d9d8777e9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 132, + 285, + 481, + 362 + ], + "blocks": [ + { + "bbox": [ + 168, + 263, + 443, + 274 + ], + "lines": [ + { + "bbox": [ + 168, + 263, + 443, + 274 + ], + "spans": [ + { + "bbox": [ + 168, + 263, + 443, + 274 + ], + "type": "text", + "content": "Table 3: Video interpolation results on UVG and DAVIS datasets." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 285, + 481, + 362 + ], + "lines": [ + { + "bbox": [ + 132, + 285, + 481, + 362 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 481, + 362 + ], + "type": "table", + "html": "
DatasetUVGDAVIS
MethodBeautyBosphBeeJockeyReadyShakeYachtAvg.BikeSwanDanceCamelCarCowsTwirlAvg.
NeRV [8]27.7429.0534.5018.6217.5428.5623.5425.6516.7019.8818.0317.3817.5517.7216.2817.65
E-NeRV [24]26.4928.3135.5518.0717.2128.2223.1725.2916.7620.0218.2018.2516.7617.2516.8017.72
HNeRV [7]30.6435.5738.6021.0821.8230.7827.3129.4019.2620.7720.3319.0218.5118.8717.5619.19
DNeRV [48]30.7934.0738.1628.6824.7630.2425.9730.3420.3321.7121.5119.3021.0518.1118.9620.14
Ours(B)30.6036.4438.3720.8121.7630.7027.0529.3719.4820.8020.2319.1018.3918.8217.6319.21
Ours(T)31.9236.6338.6329.3727.5231.3528.3231.9620.4722.1921.8722.3721.3121.6419.1221.28
", + "image_path": "7fc2e3a2a1f4d06df82d1c6a8450d19336aa461ae59f5cdf6bb17acda6d61819.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 388, + 480, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 480, + 507 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 480, + 507 + ], + "type": "text", + "content": "of 0.35M, 0.75M, 1.5M, and 3M, where the different sizes correspond to different bit budgets. The results demonstrate the superiority of our backbone model in various model capacities. Ours(T) also secures the second-best results in most scenarios. Tab. 2b exhibits the performance changes, when the epochs vary from 300 to 2400 with 3M-models, implying different encoding times. Ours(B) and Ours(T) exhibit the best and second-best performance, respectively. We also show the encoding complexity of " + }, + { + "bbox": [ + 130, + 388, + 480, + 507 + ], + "type": "inline_equation", + "content": "640 \\times 1280" + }, + { + "bbox": [ + 130, + 388, + 480, + 507 + ], + "type": "text", + "content": " video resolutions compared to other NeRV-based methods in Tab. 2b. Despite the higher complexity involved in processing multi-scale features, Ours(B) excels in video regression, delivering superior results over other methods across all sizes and training epochs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 509, + 481, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 481, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 481, + 616 + ], + "type": "text", + "content": "Video Interpolation. Tab. 3 displays the comparative studies of video interpolation in UVG and DAVIS datasets. We train the model with odd frames and test results with even frames. We observe Ours(T) and DNeRV that embed the frame differences yield better performance than the tested methods without time embeddings because the time embedding is essential to continuously represent videos over time. Ours(T) achieves the highest performance among the tested methods, which is approximately " + }, + { + "bbox": [ + 130, + 509, + 481, + 616 + ], + "type": "inline_equation", + "content": "1.14\\mathrm{dB} \\sim 1.62\\mathrm{dB}" + }, + { + "bbox": [ + 130, + 509, + 481, + 616 + ], + "type": "text", + "content": " better than DNeRV in UVG and DAVIS datasets, respectively. The results indicate that Ours(T) provides efficient video representations over time." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "Video Compression. We compare compression results with commercial video CODECs such as H.264 [45] and H.265 [40] and learning-based CODECs such as DVC [25], FVC [15], and DCVC [23]. Fig. 4 presents the rate-distortion curves. Our method improves coding efficiency over H.264 and all the other NeRV-based" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "J. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 173, + 125, + 296, + 220 + ], + "blocks": [ + { + "bbox": [ + 173, + 125, + 296, + 220 + ], + "lines": [ + { + "bbox": [ + 173, + 125, + 296, + 220 + ], + "spans": [ + { + "bbox": [ + 173, + 125, + 296, + 220 + ], + "type": "image", + "image_path": "3eb6da531aa5f2090a67f52e8bd4ffdb93f5a66b2179e96c8208cf5fc5efbcd7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 186, + 228, + 427, + 241 + ], + "lines": [ + { + "bbox": [ + 186, + 228, + 427, + 241 + ], + "spans": [ + { + "bbox": [ + 186, + 228, + 427, + 241 + ], + "type": "text", + "content": "Fig. 4: Coding performance comparisons in UVG datasets." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 307, + 125, + 434, + 220 + ], + "blocks": [ + { + "bbox": [ + 307, + 125, + 434, + 220 + ], + "lines": [ + { + "bbox": [ + 307, + 125, + 434, + 220 + ], + "spans": [ + { + "bbox": [ + 307, + 125, + 434, + 220 + ], + "type": "image", + "image_path": "dda63efc83aeb83aa9eae93ee9522eb9089a19ca1ba74aa654a8514867928f4e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 139, + 295, + 288, + 345 + ], + "blocks": [ + { + "bbox": [ + 132, + 266, + 297, + 288 + ], + "lines": [ + { + "bbox": [ + 132, + 266, + 297, + 288 + ], + "spans": [ + { + "bbox": [ + 132, + 266, + 297, + 288 + ], + "type": "text", + "content": "Table 4: Video in-painting results on DAVIS dataset." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 139, + 295, + 288, + 345 + ], + "lines": [ + { + "bbox": [ + 139, + 295, + 288, + 345 + ], + "spans": [ + { + "bbox": [ + 139, + 295, + 288, + 345 + ], + "type": "table", + "html": "
MethodIn-painting1In-painting2Avg.
HNeRV [7]29.7629.4129.59
DNeRV [48]30.0129.8229.92
Ours(B)30.7029.9030.30
", + "image_path": "bd3ea3d70f16ff47ab159e7bfb5f44e8edffb4ef07c5bdb55c21427f887db598.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 316, + 262, + 398, + 303 + ], + "blocks": [ + { + "bbox": [ + 316, + 262, + 398, + 303 + ], + "lines": [ + { + "bbox": [ + 316, + 262, + 398, + 303 + ], + "spans": [ + { + "bbox": [ + 316, + 262, + 398, + 303 + ], + "type": "image", + "image_path": "fffbd7573413caf65d2d0310c41c3339596e6ef09efb495af3c96a4ea265af35.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 312, + 482, + 346 + ], + "lines": [ + { + "bbox": [ + 315, + 312, + 482, + 346 + ], + "spans": [ + { + "bbox": [ + 315, + 312, + 482, + 346 + ], + "type": "text", + "content": "Fig.5: Visualization of In-painting1: masked input (left) and reconstructed output (right)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 399, + 262, + 481, + 303 + ], + "blocks": [ + { + "bbox": [ + 399, + 262, + 481, + 303 + ], + "lines": [ + { + "bbox": [ + 399, + 262, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 399, + 262, + 481, + 303 + ], + "type": "image", + "image_path": "ef92037fcb378b18b2d1a3b41690845f120cca190cd9a07d41be25e9b6b4e1b3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 363, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 363, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 363, + 482, + 628 + ], + "type": "text", + "content": "methods and achieves comparable coding gains with H.265, DVC, and FVC at low bit-rates with SSIM results. In the comparisons, we use the same coding pipeline and configurations as in [48]. We adopt " + }, + { + "bbox": [ + 130, + 363, + 482, + 628 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 363, + 482, + 628 + ], + "type": "text", + "content": " model pruning followed by 6-bit quantization for embeddings and 8-bit quantization for decoder parameters for all tested methods. Our methods do not apply additional techniques such as quantization-aware training (QAT) and predictive coding to improve coding efficiency for fair comparisons. HiNeRV [20] provides improved performance with specialized coding tools for NeRVs at the expense of substantial computational complexity and limited performance in video interpolation. Our method provides more fundamental changes as a backbone, which is orthogonal to the direction. Accordingly, we integrate our model into HiNeRV's coding framework and show the results denoted by Ours(B) + HiNeRV in Fig. 4. It improves the BD-PSNR by 0.3dB compared to HiNeRV. Our supplementary results also demonstrate the generalization ability of our study. In terms of model complexity, Ours(B) has 168G multiply-accumulates (MACs) in UVG dataset. HiNeRV and DNeRV have 111.71G MACs and 69.86G MACs in the same dataset, respectively. These values are significantly smaller than those of neural network based video compression (NVC) methods such as DCVC, which has 2,268G MACs as reported in [23]. Ours(B) has 31.3 decoding fps, which is relatively slower than 48.4 fps of HiNeRV but is significantly faster than 1.75 fps of DCVC, measured in NVIDIA GeForce RTX 3090 GPU. When comparing the results, Ours(B) provides a practically reasonable trade-off between coding performance and computational complexity." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Video In-painting. We evaluate the performance of a video in-painting task. We use " + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "50 \\times 50" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": " masks with five fixed points and ten random points in In-painting1 and In-painting2 [48], respectively. Tab. 4 presents that our model provides im" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 201, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 201, + 91, + 447, + 103 + ], + "type": "text", + "content": "SNeRV: Spectra-preserving Neural Representation for Video" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 221 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 221 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 221 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 221 + ], + "type": "image", + "image_path": "3546ef29a6e13f62093cb4c8e176fd4a4cda25cd12efe6375f6a6af2c7a568ed.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 222, + 480, + 239 + ], + "lines": [ + { + "bbox": [ + 132, + 222, + 480, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 222, + 480, + 239 + ], + "type": "text", + "content": "(a) Visual comparisons of tested methods in the tasks of video regression. From the top, Bike, Bosph, and Yacht." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 240, + 481, + 312 + ], + "blocks": [ + { + "bbox": [ + 133, + 240, + 481, + 312 + ], + "lines": [ + { + "bbox": [ + 133, + 240, + 481, + 312 + ], + "spans": [ + { + "bbox": [ + 133, + 240, + 481, + 312 + ], + "type": "image", + "image_path": "068a919300fab2b011d571d3e9608b116b749585fc2c1686c45554aafc4cca51.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 313, + 480, + 331 + ], + "lines": [ + { + "bbox": [ + 132, + 313, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 132, + 313, + 480, + 331 + ], + "type": "text", + "content": "(b) Visual comparisons of tested methods in the tasks of video interpolation. From the top, Jockey and Ready." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 340, + 480, + 363 + ], + "lines": [ + { + "bbox": [ + 132, + 340, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 340, + 480, + 363 + ], + "type": "text", + "content": "Fig. 6: Visual comparisons in video regression and interpolation tasks. From the left, ground-truth (GT), HNeRV, DNeRV, and Ours." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 388, + 480, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 480, + 412 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 480, + 412 + ], + "type": "text", + "content": "proved results approximately " + }, + { + "bbox": [ + 130, + 388, + 480, + 412 + ], + "type": "inline_equation", + "content": "0.38\\mathrm{dB} \\sim 0.71\\mathrm{dB}" + }, + { + "bbox": [ + 130, + 388, + 480, + 412 + ], + "type": "text", + "content": " over the existing methods. Fig. 5 shows the visualization results of In-painting1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 413, + 480, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 480, + 449 + ], + "type": "text", + "content": "Qualitative Performance Analysis. We provide the qualitative results of the tested methods in Fig. 6. The results illustrate that our methods provide visually pleasing results in video regression and interpolation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 472, + 212, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 212, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 212, + 483 + ], + "type": "text", + "content": "4.3 Discussion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "Trade-off between Video Regression and Time Interpolation. Tab. 1 and Tab. 3 display the video regression and interpolation performances, respectively. According to the results, Ours(B) outperforms Ours(T) in regression, but underperforms in interpolation, indicating a trade-off relationship. We explain this trade-off between time duration " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "(\\Delta t)" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " and spectral band " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "(\\Delta \\omega)" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " in terms of learning aspects. Ours(B) sets " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\Delta t = 0" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": ", lacking the time analysis but maximizing the spectral analysis, while Ours(T) compromises " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\Delta \\omega" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": ". Given the limited model parameters, learning the spatial spectral representation of a current frame could fit a reconstruction task but might compromise the generalization ability across time. In contrast, learning temporal relations can make it challenging to fit the parameters to the current time. Similar characteristics are present not only in Ours(B) and Ours(T) but also in other findings. DNeRV [48] offers superior performance in video interpolation compared to HNeRV [7], but not in video reconstruction. This strategy in our approach signifies a new direction in" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "J. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 220, + 159 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 220, + 159 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 220, + 159 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 220, + 159 + ], + "type": "image", + "image_path": "48dec2addea558ab2788c4054e35414187caae35d2b315a8db0f58ac765bbe8b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 135, + 160, + 206, + 178 + ], + "lines": [ + { + "bbox": [ + 135, + 160, + 206, + 178 + ], + "spans": [ + { + "bbox": [ + 135, + 160, + 206, + 178 + ], + "type": "text", + "content": "(a) Low-frequency sults of Jockey." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 221, + 114, + 304, + 159 + ], + "blocks": [ + { + "bbox": [ + 221, + 114, + 304, + 159 + ], + "lines": [ + { + "bbox": [ + 221, + 114, + 304, + 159 + ], + "spans": [ + { + "bbox": [ + 221, + 114, + 304, + 159 + ], + "type": "image", + "image_path": "ae021564a6f47e982f462600072c47998d218b5e337d4abb6f00e61c4e1a5511.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 221, + 160, + 306, + 178 + ], + "lines": [ + { + "bbox": [ + 221, + 160, + 306, + 178 + ], + "spans": [ + { + "bbox": [ + 221, + 160, + 306, + 178 + ], + "type": "text", + "content": "(b) High-frequency results of Jockey." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 114, + 392, + 159 + ], + "blocks": [ + { + "bbox": [ + 307, + 114, + 392, + 159 + ], + "lines": [ + { + "bbox": [ + 307, + 114, + 392, + 159 + ], + "spans": [ + { + "bbox": [ + 307, + 114, + 392, + 159 + ], + "type": "image", + "image_path": "b050efaad87d189e4014a0025b577ac65ac91de063d2b880a1cda3a1dcb555e5.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 188, + 187, + 424, + 198 + ], + "lines": [ + { + "bbox": [ + 188, + 187, + 424, + 198 + ], + "spans": [ + { + "bbox": [ + 188, + 187, + 424, + 198 + ], + "type": "text", + "content": "Fig. 7: Training rates for different frequency components." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 392, + 114, + 478, + 159 + ], + "blocks": [ + { + "bbox": [ + 307, + 160, + 391, + 178 + ], + "lines": [ + { + "bbox": [ + 307, + 160, + 391, + 178 + ], + "spans": [ + { + "bbox": [ + 307, + 160, + 391, + 178 + ], + "type": "text", + "content": "(c) Low-frequency results of Ready." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 392, + 114, + 478, + 159 + ], + "lines": [ + { + "bbox": [ + 392, + 114, + 478, + 159 + ], + "spans": [ + { + "bbox": [ + 392, + 114, + 478, + 159 + ], + "type": "image", + "image_path": "541841b39c7bda20f47be1c1d39b061dadb0bb59a2b3dc939d75eed7c7bb29ef.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 393, + 160, + 478, + 178 + ], + "lines": [ + { + "bbox": [ + 393, + 160, + 478, + 178 + ], + "spans": [ + { + "bbox": [ + 393, + 160, + 478, + 178 + ], + "type": "text", + "content": "(d) High-frequency results of Ready." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 222, + 480, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 222, + 480, + 245 + ], + "spans": [ + { + "bbox": [ + 130, + 222, + 480, + 245 + ], + "type": "text", + "content": "analyzing and optimizing NeRV for various video processing tasks from a time and spectra perspective, which has not been conducted before." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 246, + 481, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 246, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 246, + 481, + 402 + ], + "type": "text", + "content": "Learning Characteristics of LF and HF Components. Fig. 7 displays the learning characteristics of LF and HF components separately for analysis, while they are trained simultaneously. Jockey and Ready sequences contain fine textures and motions, yielding some HF components that are challenging to represent with an implicit neural network. Fig. 7a and Fig. 7c show the results of the LF at training steps. The LFs are steadily trained to reach the convergence points producing similar PSNR values to the tested methods. Besides, Ours(T) and DNeRV exhibit better performance in the early stage, because they use larger sizes of embeddings. However, Fig. 7b and Fig. 7d show significantly different characteristics in learning HF components. While the previous methods fail to efficiently learn the HFs, our methods achieve significantly improved performance with increasing epochs. These results verify that our model could mitigate the spectral bias and successfully preserve HF components." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 418, + 237, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 237, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 237, + 430 + ], + "type": "text", + "content": "4.4 Ablation Study" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 437, + 482, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 569 + ], + "type": "text", + "content": "MFU and HFR Modules. Tab. 5a presents the effectiveness of the MFU and HFR. We conduct ablation tests on the Bunny dataset mostly including LF components and the Yacht sequence in UVG including comparatively larger HF components. The best performance is achieved when both the MFU and HFR are utilized. MFU achieves significantly improved performance approximately " + }, + { + "bbox": [ + 130, + 437, + 482, + 569 + ], + "type": "inline_equation", + "content": "1.20\\mathrm{dB} \\sim 1.88\\mathrm{dB}" + }, + { + "bbox": [ + 130, + 437, + 482, + 569 + ], + "type": "text", + "content": ", demonstrating the effectiveness of the module. Using HFR, while Bunny increases the performance by " + }, + { + "bbox": [ + 130, + 437, + 482, + 569 + ], + "type": "inline_equation", + "content": "0.21\\mathrm{dB}" + }, + { + "bbox": [ + 130, + 437, + 482, + 569 + ], + "type": "text", + "content": ", Yacht observes " + }, + { + "bbox": [ + 130, + 437, + 482, + 569 + ], + "type": "inline_equation", + "content": "1.62\\mathrm{dB}" + }, + { + "bbox": [ + 130, + 437, + 482, + 569 + ], + "type": "text", + "content": ". This confirms that HFR is capable of restoring HF components. Further, the HFR contributes approximately " + }, + { + "bbox": [ + 130, + 437, + 482, + 569 + ], + "type": "inline_equation", + "content": "0.44\\mathrm{dB} \\sim 1.50\\mathrm{dB}" + }, + { + "bbox": [ + 130, + 437, + 482, + 569 + ], + "type": "text", + "content": " on top of the MFU. This refinement is particularly advantageous in generating HFs that were hardly trained in previous NeRV models." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 570, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 628 + ], + "type": "text", + "content": "Temporal Up-sampling Block. We evaluate the performance of the TUBs, by replacing them in Ours(T) with the conventional NeRV blocks comprising PixelShuffle and 2D convolution layers. Tab. 5b shows that our TUB is suitable for processing time-extended features than the 2D operations. TUB improves the PSNR around 0.80dB." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": "Loss Terms. We conduct ablation tests on the loss function in Eq. (2) using a Yacht sequence. Tab. 5b shows the results, when the terms are turned on or off. Both the terms contribute to the performance improvements." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 201, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 201, + 91, + 448, + 102 + ], + "type": "text", + "content": "SNeRV: Spectra-preserving Neural Representation for Video" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 154, + 164, + 298, + 238 + ], + "blocks": [ + { + "bbox": [ + 168, + 114, + 443, + 126 + ], + "lines": [ + { + "bbox": [ + 168, + 114, + 443, + 126 + ], + "spans": [ + { + "bbox": [ + 168, + 114, + 443, + 126 + ], + "type": "text", + "content": "Table 5: Ablation studies of the proposed modules and loss terms." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 192, + 145, + 263, + 155 + ], + "lines": [ + { + "bbox": [ + 192, + 145, + 263, + 155 + ], + "spans": [ + { + "bbox": [ + 192, + 145, + 263, + 155 + ], + "type": "text", + "content": "(a) MFU and HFR." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 154, + 164, + 298, + 238 + ], + "lines": [ + { + "bbox": [ + 154, + 164, + 298, + 238 + ], + "spans": [ + { + "bbox": [ + 154, + 164, + 298, + 238 + ], + "type": "table", + "html": "
DatasetBunnyYacht
MFUHFRPSNRSSIMPSNRSSIM
XX38.000.991130.050.9323
✓X39.200.992431.930.9458
X✓38.210.991131.670.9516
✓✓39.640.992833.430.9647
", + "image_path": "20e4e2bf06afd515c48a6a850a43a25dc95c2c7d71941937c14c76d8e4b1049b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 326, + 164, + 442, + 243 + ], + "blocks": [ + { + "bbox": [ + 340, + 145, + 432, + 155 + ], + "lines": [ + { + "bbox": [ + 340, + 145, + 432, + 155 + ], + "spans": [ + { + "bbox": [ + 340, + 145, + 432, + 155 + ], + "type": "text", + "content": "(b) Loss terms and TUB." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 326, + 164, + 442, + 243 + ], + "lines": [ + { + "bbox": [ + 326, + 164, + 442, + 243 + ], + "spans": [ + { + "bbox": [ + 326, + 164, + 442, + 243 + ], + "type": "table", + "html": "
L(It,ˆt) L(Ct,ˆCt)PSNRSSIM
✓X33.290.9600
✓✓33.430.9647
Temporal BlockPSNRSSIM
NeRV Block31.910.9532
TUB32.710.9584
", + "image_path": "1b17c48bcf9fe06484b8ae3dc7bf5ea891bc43c7697644c26cb1146a6de6289c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 149, + 256, + 280, + 315 + ], + "blocks": [ + { + "bbox": [ + 149, + 256, + 280, + 315 + ], + "lines": [ + { + "bbox": [ + 149, + 256, + 280, + 315 + ], + "spans": [ + { + "bbox": [ + 149, + 256, + 280, + 315 + ], + "type": "image", + "image_path": "3221893205be72ddee724241e375b20f85a5f3954d4d21c4279666fd52344df7.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 133, + 319, + 297, + 335 + ], + "lines": [ + { + "bbox": [ + 133, + 319, + 297, + 335 + ], + "spans": [ + { + "bbox": [ + 133, + 319, + 297, + 335 + ], + "type": "text", + "content": "(a) Varying temporal embedding size with fixed target embedding." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 331, + 256, + 463, + 314 + ], + "blocks": [ + { + "bbox": [ + 331, + 256, + 463, + 314 + ], + "lines": [ + { + "bbox": [ + 331, + 256, + 463, + 314 + ], + "spans": [ + { + "bbox": [ + 331, + 256, + 463, + 314 + ], + "type": "image", + "image_path": "6783c7c233d46a02b71fa1736bc1c64afe5e40f68dabc43531ce97cbbb178022.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 316, + 318, + 480, + 335 + ], + "lines": [ + { + "bbox": [ + 316, + 318, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 316, + 318, + 480, + 335 + ], + "type": "text", + "content": "(b) Varying target embedding size with fixed temporal embedding." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 153, + 345, + 459, + 357 + ], + "lines": [ + { + "bbox": [ + 153, + 345, + 459, + 357 + ], + "spans": [ + { + "bbox": [ + 153, + 345, + 459, + 357 + ], + "type": "text", + "content": "Fig. 8: Regression and interpolation results with different embedding sizes." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 379, + 482, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 379, + 482, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 379, + 482, + 464 + ], + "type": "text", + "content": "Effect of Size of Model Components. We analyze the performance with different sizes of target " + }, + { + "bbox": [ + 130, + 379, + 482, + 464 + ], + "type": "inline_equation", + "content": "e_t" + }, + { + "bbox": [ + 130, + 379, + 482, + 464 + ], + "type": "text", + "content": " and temporal embeddings " + }, + { + "bbox": [ + 130, + 379, + 482, + 464 + ], + "type": "inline_equation", + "content": "e_{\\Delta t}" + }, + { + "bbox": [ + 130, + 379, + 482, + 464 + ], + "type": "text", + "content": "s, with the size of the decoder varying according to these sizes. Fig. 8 presents the performance variations, exploring the impacts of the ratios. The regression performance is reliable in all the cases, when the ratios are different. In interpolation, the size of the temporal embedding greatly affects the performance, because it hints the temporal relations among video frames." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 480, + 220, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 480, + 220, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 480, + 220, + 493 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 504, + 482, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 504, + 482, + 577 + ], + "spans": [ + { + "bbox": [ + 130, + 504, + 482, + 577 + ], + "type": "text", + "content": "In this paper, we proposed the spectra-preserving neural representation for video as an efficient baseline framework. SNeRV successfully learned LF components, while reconstructing fine HF details within a limited capacity of parameters, and surpassed the existing NeRV-based methods in various video-related tasks. More compact and efficient architecture to process the high-frequency components considering time-frequency trade-off will be done as a future work." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 594, + 241, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 594, + 241, + 608 + ], + "spans": [ + { + "bbox": [ + 132, + 594, + 241, + 608 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "This work was partly supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No.RS-2021-II212068 and RS-2022-00167169) and was partly supported by the NRF grant funded by MSIT (No.NRF-2022R1A2C4002052)." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "J. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 141, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 141, + 480, + 176 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 141, + 480, + 176 + ], + "spans": [ + { + "bbox": [ + 138, + 141, + 480, + 176 + ], + "type": "text", + "content": "1. Arora, S., Du, S., Hu, W., Li, Z., Wang, R.: Fine-grained analysis of optimization and generalization for overparameterized two-layer neural networks. International Conf. on Machine Learning (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 176, + 480, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 176, + 480, + 209 + ], + "spans": [ + { + "bbox": [ + 138, + 176, + 480, + 209 + ], + "type": "text", + "content": "2. Bai, Y., Dong, C., Wang, C., Yuan, C.: Ps-nerv: Patch-wise stylized neural representations for videos. In: Proceedings of the IEEE conference on Image Processing (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 209, + 480, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 209, + 480, + 232 + ], + "spans": [ + { + "bbox": [ + 138, + 209, + 480, + 232 + ], + "type": "text", + "content": "3. Balle, J., Minnen, D., Singh, S., Hwang, S.J., Johnston, N.: Variational image compression with a scale hyperprior. arXiv preprint arXiv:1802.01436 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 232, + 480, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 232, + 480, + 264 + ], + "spans": [ + { + "bbox": [ + 138, + 232, + 480, + 264 + ], + "type": "text", + "content": "4. Basri, R., Galun, M., Geifman, A., Jacobs, D., Kasten, Y., Kritchman, S.: Frequency bias in neural networks for input of non-uniform density. arXiv preprint arXiv 2003.04560 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 264, + 480, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 264, + 480, + 308 + ], + "spans": [ + { + "bbox": [ + 138, + 264, + 480, + 308 + ], + "type": "text", + "content": "5. Bross, B., Wang, Y.K., Ye, Y., Liu, S., Chen, J., Sullivan, G.J., Ohm, J.R.: Overview of the versatile video coding (vvc) standard and its applications. IEEE Transactions on Circuits and Systems for Video Technology 31(10), 3736-3764 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 309, + 480, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 309, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 138, + 309, + 480, + 331 + ], + "type": "text", + "content": "6. Cao, Y., Fang, Z., Wu, Y., Zhou, D.X., Gu, Q.: Towards understanding the spectral bias of deep learning. arXiv preprint arXiv:1912.01198 (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 332, + 480, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 332, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 138, + 332, + 480, + 365 + ], + "type": "text", + "content": "7. Chen, H., Gwilliam, M., Lim, S., Shrivastava, A.: Hnerv: A hybrid neural representation for videos. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 10270-10279 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 365, + 480, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 365, + 480, + 398 + ], + "spans": [ + { + "bbox": [ + 138, + 365, + 480, + 398 + ], + "type": "text", + "content": "8. Chen, H., He, B., Wang, H., Ren, Y., Lim, S.N., Shrivastava, A.: Nerv: Neural representations for videos. Advances in Neural Information Processing Systems 34, 21557-21568 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 399, + 480, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 399, + 480, + 443 + ], + "spans": [ + { + "bbox": [ + 138, + 399, + 480, + 443 + ], + "type": "text", + "content": "9. Chen, Y.H., Chen, S.C., Lin, Y.Y., Peng, W.H.: Motif: Learning motion trajectories with local implicit neural functions for continuous space-time video super-resolution. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 23131-23141 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 443, + 480, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 443, + 480, + 487 + ], + "spans": [ + { + "bbox": [ + 138, + 443, + 480, + 487 + ], + "type": "text", + "content": "0. Chen, Z., Chen, Y., Liu, J., Xu, X., Goel, V., Wang, Z., Shi, H., Wang, X.: Videoirr: Learning video implicit neural representation for continuous space-time superresolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2047-2057 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 487, + 480, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 487, + 480, + 520 + ], + "spans": [ + { + "bbox": [ + 138, + 487, + 480, + 520 + ], + "type": "text", + "content": "1. Chen, Z., Zhang, H.: Learning implicit fields for generative shape modeling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5939-5948 (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 521, + 480, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 521, + 480, + 554 + ], + "spans": [ + { + "bbox": [ + 138, + 521, + 480, + 554 + ], + "type": "text", + "content": "2. Choi, Y.J., Lee, Y.W., Kim, B.G.: Wavelet attention embedding networks for video super-resolution. In: 2020 25th International conference on pattern recognition (ICPR). pp. 7314-7320. IEEE (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 554, + 480, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 554, + 480, + 587 + ], + "spans": [ + { + "bbox": [ + 138, + 554, + 480, + 587 + ], + "type": "text", + "content": "3. Djelouah, A., Campos, J., Schaub-Meyer, S., Schroers, C.: Neural inter-frame compression for video coding. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6421-6429 (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 587, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 587, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 587, + 480, + 632 + ], + "type": "text", + "content": "4. He, B., Yang, X., Wang, H., Wu, Z., Chen, H., Huang, S., Ren, Y., Lim, S.N., Shrivastava, A.: Towards scalable neural representation for diverse videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6132-6142 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "text", + "content": "5. Hu, Z., Lu, G., Xu, D.: Fvc: A new framework towards deep video compression in feature space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1502-1511 (2021)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "text", + "content": "SNeRV: Spectra-preserving Neural Representation for Video" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "16. Huang, H., He, R., Sun, Z., Tan, T.: Wavelet domain generative adversarial network for multi-scale face hallucination. International Journal of Computer Vision 127(6-7), 763-784 (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 194 + ], + "type": "text", + "content": "17. Jin, B., Hu, Y., Tang, Q., Niu, J., Shi, Z., Han, Y., Li, X.: Exploring spatial-temporal multi-frequency analysis for high-fidelity and temporal-consistency video prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4554-4563 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 194, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 194, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 133, + 194, + 480, + 237 + ], + "type": "text", + "content": "18. Jung, H., Hui, Z., Luo, L., Yang, H., Liu, F., Yoo, S., Ranjan, R., Demandolx, D.: Anyflow: Arbitrary scale optical flow with implicit neural representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5455-5465 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 238, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 238, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 133, + 238, + 480, + 259 + ], + "type": "text", + "content": "19. Kong, L., Jiang, B., Luo, D., Chu, W., Tai, Y., Wang, C., Yang, J.: Dynamic frame interpolation in wavelet domain. IEEE Transactions on Image Processing (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 259, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 259, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 133, + 259, + 480, + 293 + ], + "type": "text", + "content": "20. Kwan, H.M., Gao, G., Zhang, F., Gower, A., Bull, D.: Hinerv: Video compression with hierarchical encoding-based neural representation. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 480, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 315 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 315 + ], + "type": "text", + "content": "21. Lee, J.C., Rho, D., Ko, J.H., Park, E.: Fferv: Flow-guided frame-wise neural representations for videos. arXiv preprint arXiv:2212.12294 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 316, + 480, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 316, + 480, + 336 + ], + "spans": [ + { + "bbox": [ + 132, + 316, + 480, + 336 + ], + "type": "text", + "content": "22. Lee, J.K., Kim, N., Cho, S., Kang, J.W.: Deep video prediction network-based inter-frame coding in hevc. IEEE ACCESS 8, 95906-95917 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 337, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 337, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 337, + 480, + 357 + ], + "type": "text", + "content": "23. Li, J., Li, B., Lu, Y.: Deep contextual video compression. Advances in Neural Information Processing Systems 34, 18114-18125 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 358, + 480, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 392 + ], + "type": "text", + "content": "24. Li, Z., Wang, M., Pi, H., Xu, K., Mei, J., Liu, Y.: E-nerv: Expedite neural video representation with disentangled spatial-temporal context. In: European Conference on Computer Vision. pp. 267-284. Springer (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 392, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 424 + ], + "type": "text", + "content": "25. Lu, G., Ouyang, W., Xu, D., Zhang, X., Cai, C., Gao, Z.: Dvc: An end-to-end deep video compression framework. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11006-11015 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 425, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 425, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 425, + 480, + 468 + ], + "type": "text", + "content": "26. Lu, Y., Wang, Z., Liu, M., Wang, H., Wang, L.: Learning spatial-temporal implicit neural representations for event-guided video super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1557-1567 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 468, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 522 + ], + "type": "text", + "content": "27. Maiya, S.R., Girish, S., Ehrlich, M., Wang, H., Lee, K.S., Poirson, P., Wu, P., Wang, C., Shrivastava, A.: Nirvana: Neural implicit representations of videos with adaptive networks and autoregressive patch-wise modeling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14378-14387 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 523, + 403, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 403, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 403, + 533 + ], + "type": "text", + "content": "28. Mallat, S.: A wavelet tour of signal processing. Elsevier (1999)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "type": "text", + "content": "29. Mercat, A., Viitanen, M., Vanne, J.: Uvg dataset: 50/120fps 4k sequences for video codec analysis and development. In: Proceedings of the 11th ACM Multimedia Systems Conference. pp. 297-302 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "type": "text", + "content": "30. Pan, Z., Yi, X., Zhang, Y., Jeon, B., Kwong, S.: Efficient in-loop filtering based on enhanced deep convolutional neural networks for hevc. IEEE Transactions on Image Processing 29, 5352-5366 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 600, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 643 + ], + "type": "text", + "content": "31. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 165-174 (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "type": "text", + "content": "32. Perazzi, F., Pont-Tuset, J., McWilliams, B., Gool, L.V., Gross, M., Sorkine-Hornung, A.: A benchmark dataset and evaluation methodology for video object" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "J. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 147, + 117, + 481, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 117, + 481, + 138 + ], + "spans": [ + { + "bbox": [ + 147, + 117, + 481, + 138 + ], + "type": "text", + "content": "segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 724-732 (2016)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 481, + 183 + ], + "type": "text", + "content": "33. Ramamonjisoa, M., Firman, M., Watson, J., Lepetit, V., Turmukhambetov, D.: Single image depth prediction with wavelet decomposition. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11089-11098 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 184, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 184, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 184, + 481, + 215 + ], + "type": "text", + "content": "34. Rho, D., Cho, J., Ko, J.H., Park, E.: Neural residual flow fields for efficient video representations. In: Proceedings of the Asian Conference on Computer Vision. pp. 3447-3463 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "text", + "content": "35. Rho, D., Lee, B., Nam, S., Lee, J.C., Ko, J.H., Park, E.: Masked wavelet representation for compact neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20680-20690 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "text", + "content": "36. Ronen, B., Jacobs, D., Kasten, Y., Kritchman, S.: The convergence rate of neural networks for learned functions of different frequencies. Advances in Neural Information Processing Systems 32 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 282, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 325 + ], + "type": "text", + "content": "37. Saragadam, V., LeJeune, D., Tan, J., Balakrishnan, G., Veeraraghavan, A., Baraniuk, R.G.: Wire: Wavelet implicit neural representations. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 18507-18516 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 326, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 369 + ], + "type": "text", + "content": "38. Shi, W., Caballero, J., Huszar, F., Totz, J., Aitken, A.P., Bishop, R., Wang, Z.: Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1874-1883 (2016)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 370, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 370, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 370, + 481, + 402 + ], + "type": "text", + "content": "39. Sitzmann, V., Martel, J., Bergman, A., Lindell, D., Wetzstein, G.: Implicit neural representations with periodic activation functions. Advances in neural information processing systems 33, 7462-7473 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "type": "text", + "content": "40. Sullivan, G.J., Ohm, J.R., Han, W.J., Wiegand, T.: Overview of the high efficiency video coding (hevc) standard. IEEE Transactions on circuits and systems for video technology 22(12), 1649-1668 (2012)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 436, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 436, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 436, + 481, + 479 + ], + "type": "text", + "content": "41. Tancik, M., Srinivasan, P., Mildenhall, B., Fridovich-Keil, S., Raghavan, N., Singhal, U., Ramamoorthi, R., Barron, B., Ng, R.: Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems 33 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "text", + "content": "42. Taubman, D.S., Marcellin, M.W., Rabbani, M.: Jpeg2000: Image compression fundamentals, standards and practice. Journal of Electronic Imaging 11(2), 286-287 (2002)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "type": "text", + "content": "43. Unser, M., Blu, T.: Mathematical properties of the jpeg2000 wavelet filters. IEEE transactions on image processing 12(9), 1080-1090 (2003)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "type": "text", + "content": "44. Wang, P., Fan, Z., Chen, T., Wang, Z.: Neural implicit dictionary learning via mixture-of-expert training. In: International Conference on Machine Learning. pp. 22613-22624. PMLR (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "type": "text", + "content": "45. Wiegand, T., Sullivan, G.J., Bjontegaard, G., Luthra, A.: Overview of the h. 264/avc video coding standard. IEEE Transactions on circuits and systems for video technology 13(7), 560-576 (2003)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "type": "text", + "content": "46. Xin, J., Li, J., Jiang, X., Wang, N., Huang, H., Gao, X.: Wavelet-based dual recursive network for image super-resolution. IEEE Transactions on Neural Networks and Learning Systems 33(2), 707-720 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "type": "text", + "content": "47. Yuce, G., Ortiz-Jimenez, G., Besbinar, B., Frossard, P.: A structured dictionary perspective on implicit neural representations. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 19228-19238 (2022)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "text", + "content": "SNeRV: Spectra-preserving Neural Representation for Video" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 482, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 482, + 150 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 482, + 150 + ], + "type": "text", + "content": "48. Zhao, Q., Asif, M.S., Ma, Z.: Dnerv: Modeling inherent dynamics via difference neural representation for videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2031-2040 (2023)" + } + ] + } + ], + "index": 2 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "J. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/SOS_ Segment Object System for Open-World Instance Segmentation With Object Priors/466dafa4-76e1-41c2-9ff8-766257e5e796_content_list.json b/2024/SOS_ Segment Object System for Open-World Instance Segmentation With Object Priors/466dafa4-76e1-41c2-9ff8-766257e5e796_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bd7f1a106609a80a1a0d424e1bc3a25caf9adf60 --- /dev/null +++ b/2024/SOS_ Segment Object System for Open-World Instance Segmentation With Object Priors/466dafa4-76e1-41c2-9ff8-766257e5e796_content_list.json @@ -0,0 +1,1773 @@ +[ + { + "type": "text", + "text": "SOS: Segment Object System for Open-World Instance Segmentation With Object Priors", + "text_level": 1, + "bbox": [ + 238, + 140, + 763, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Christian Wilms1, Tim Rolff1,2, Maris Hillemann1, Robert Johanson1, and Simone Frinthrop1", + "bbox": [ + 220, + 210, + 781, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Computer Vision Group, University of Hamburg, Germany \n2 Human-Computer Interaction Group, University of Hamburg, Germany {firstname_lastname}@uni-hamburg.de", + "bbox": [ + 254, + 253, + 746, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. We propose an approach for Open-World Instance Segmentation (OWIS), a task that aims to segment arbitrary unknown objects in images by generalizing from a limited set of annotated object classes during training. Our Segment Object System (SOS) explicitly addresses the generalization ability and the low precision of state-of-the-art systems, which often generate background detections. To this end, we generate high-quality pseudo annotations based on the foundation model SAM [27]. We thoroughly study various object priors to generate prompts for SAM, explicitly focusing the foundation model on objects. The strongest object priors were obtained by self-attention maps from self-supervised Vision Transformers, which we utilize for prompting SAM. Finally, the post-processed segments from SAM are used as pseudo annotations to train a standard instance segmentation system. Our approach shows strong generalization capabilities on COCO, LVIS, and ADE20k datasets and improves on the precision by up to $81.6\\%$ compared to the state-of-the-art. Source code is available at: https://github.com/chwilms/SOS", + "bbox": [ + 259, + 329, + 743, + 564 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Open-world Instance Segmentation $\\cdot$ Object Localization $\\cdot$ Prompting", + "bbox": [ + 259, + 577, + 740, + 606 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 628, + 375, + 645 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Open-World Instance Segmentation (OWIS) is the task of segmenting all object instances in an image by learning from a limited set of known object classes [26, 49, 50]. In contrast to instance segmentation, OWIS is not limited by the closed-world assumption, which assumes that all object classes are known in advance. Since OWIS methods aim to detect not only learned but also unknown object classes, they return class-agnostic detections. This is of particular interest in real-world scenarios with previously unknown object classes (e.g., [55]) and challenges the systems' generalization capabilities. An example of this is shown in the first row of Fig. 1, where annotations for classes such as surfboard or tennis racket do not exist during training, but the objects should be detected during testing.", + "bbox": [ + 212, + 657, + 787, + 809 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Previous OWIS methods mostly generate pseudo annotations for unannotated objects [24, 46, 49, 56] or replace the foreground-background classification", + "bbox": [ + 212, + 810, + 785, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b90e68b72143383adf80f1e1ba9fcc4f24d0c46409d00badc90355a6e7ca539d.jpg", + "image_caption": [ + "Fig. 1: Comparison of (pseudo) annotations (left) used by Mask R-CNN [19], GGN [49], and our SOS, when only original annotations of VOC object classes are given. While Mask R-CNN only uses original annotations without object classes such as tennis racket or surfboard (red arrows), SOS generates pseudo annotations covering those classes (green arrows). GGN generates noisy pseudo annotations including background areas. As a result, only SOS constantly detects these objects not annotated in training (green vs. red arrows on the right). Filled masks denote annotations (left) or detected objects (right), while red frames indicate missed objects." + ], + "image_footnote": [], + "bbox": [ + 215, + 141, + 787, + 309 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "of possible objects in instance segmentation systems with a learning target focusing on the localization quality (e.g., intersection over union) [26,48,54]. While OWIS methods generalize better to unseen object classes than standard instance segmentation systems, they still exhibit low precision [24,26,46,49,54]. One reason for this is the use of noisy pseudo annotations covering background areas as visible in the left part of the second row in Fig. 1.", + "bbox": [ + 212, + 469, + 787, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, foundation models [4] emerged as a promising technique [5,9,23,29, 42]. They are trained on large datasets with a surrogate task and applied to other tasks in a zero- or few-shot manner using prompting. The first foundation model for image segmentation, the Segment Anything Model (SAM), was proposed by [27]. Trained on a large automatically annotated dataset, SAM generates segments for object or stuff regions based on point, box, mask, or text prompts without further training. Despite not only segmenting objects but also stuff regions, given well-designed prompts, SAM is able to generate high-quality object segments. Note that vanilla SAM does not address object segmentation tasks like OWIS itself due to the aforementioned lacking focus on objects.", + "bbox": [ + 212, + 564, + 787, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose the Segment Object System (SOS), a novel OWIS method utilizing high-quality pseudo annotations to improve the generalization capability of a standard instance segmentation system (see lower row in Fig. 1). For generating pseudo annotations, we apply the foundation model SAM with prompts derived from an object prior to roughly localize objects of arbitrary classes and to limit segmentations of stuff regions by SAM. To ensure the high quality of our pseudo annotations based on SAM, improving the overall precision, we thoroughly study various hand-crafted and learned object priors for an", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "C. Wilms et al.", + "bbox": [ + 271, + 114, + 374, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "object-focused application of SAM. This is relevant beyond the scope of OWIS. Using our study findings, SOS utilizes self-attention maps from self-supervised Vision Transformers (ViTs) [11] as object prior for prompting SAM. Finally, SOS filters low-quality pseudo annotations, combines the pseudo annotations with the original annotations of the known classes, and trains a standard instance segmentation system with these mixed annotations. Our extensive evaluation shows the strong generalization abilities of SOS, considerably outperforming all previous state-of-the-art systems across COCO [31], LVIS [17], and ADE20k [61] datasets. Most notably, the results improve by up to $81.6\\%$ in terms of precision over the state-of-the-art due to the high-quality pseudo annotations. We also show that SOS is better suited for OWIS than directly applying SAM.", + "bbox": [ + 212, + 146, + 782, + 311 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overall, our contributions are threefold:", + "bbox": [ + 238, + 313, + 529, + 325 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose SOS, a novel OWIS method based on a learned object prior, prompt-based pseudo annotations, and an arbitrary instance segmentation system.", + "- We thoroughly study various object priors for focusing SAM on objects, leading to high-quality object segments for pseudo annotations and beyond.", + "- Our extensive evaluation shows that the high-quality pseudo annotations in SOS lead to high precision and strong overall results, clearly outperforming other methods." + ], + "bbox": [ + 223, + 340, + 784, + 460 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 488, + 387, + 503 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Open-world Instance Segmentation", + "text_level": 1, + "bbox": [ + 214, + 523, + 550, + 539 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address OWIS, literature mainly offers two streams. First, systems replace the hard classification in instance segmentation systems with a localization score. This avoids classifying unseen objects as background and improves generalization ability [26]. In OLN [26], centerness and Intersection over Union (IoU) are learned as localization scores in a Mask R-CNN system [19]. SWORD [54] and OpenInst [48] follow the same idea for query-based systems.", + "bbox": [ + 212, + 551, + 782, + 642 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The second line of works [24, 46, 49, 56] addresses OWIS by augmenting the annotations from known classes with pseudo annotations aiming to cover objects of unknown classes. To generate pseudo annotations, GGN [49] learns pixel affinities from known classes to generate segments, which are grouped. Similarly, UDOS [24] groups object proposals of object parts to create pseudo annotations, while LDET [46] applies copy-paste-augmentation. Recently, a new stream emerged [58, 62] explicitly utilizing class labels from known classes.", + "bbox": [ + 212, + 643, + 784, + 748 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We follow the second stream and create pseudo annotations within SOS, since it offers the most flexibility w.r.t. the base instance segmentation system. Different from existing approaches that generate noisy pseudo annotations (see GGN in Fig. 1), we investigate and develop object priors to effectively use recent foundation models to create high-quality object-focused pseudo annotations, leading to high recall and precision.", + "bbox": [ + 212, + 750, + 784, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "SOS: Segment Object System", + "bbox": [ + 531, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 Class-agnostic Object Localization", + "text_level": 1, + "bbox": [ + 215, + 146, + 545, + 162 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Localizing objects in images is related to several computer vision tasks. For instance, in traditional object proposal generation [2], cues like saliency [2, 37] or contour information [2, 10, 36, 43] were used to localize class-agnostic object candidates for object detection. Since the advent of deep learning, the models are learned from data [16, 22, 40, 41, 53].", + "bbox": [ + 212, + 172, + 785, + 247 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Focusing on arbitrary salient objects, known as salient object detection, several approaches use the contrast of hand-crafted features such as color [8,28,34, 39] or edge information [28], as well as CNN-based approaches [21, 33, 59]. In a different task, several authors investigated the information stored in learned classifiers or self-supervised feature extractors based on CNNs or ViTs to locate objects. Notable avenues include Class Activation Maps (CAMs) [60] that highlight discriminative object or image parts in CNN-based classifiers and the self-attention maps in the self-supervised DINO feature extractor [6] based on ViTs, which contain information on the scene layout indicating object locations.", + "bbox": [ + 212, + 250, + 785, + 383 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our study on object priors to focus SAM on objects, we investigate several of these object localization cues. In contrast to these approaches, we follow a different goal, focusing prompt-based segmentations on objects for OWIS.", + "bbox": [ + 212, + 383, + 785, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Applications of SAM", + "text_level": 1, + "bbox": [ + 215, + 453, + 434, + 468 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SAM [27] is used in several segmentation tasks [3,38,47]. Similar to us, [7] and [57] use SAM to generate pseudo annotation masks for weakly-supervised semantic segmentation based on image-level supervision. Similarly, [18] generate pseudo annotation masks for concealed object segmentation based on scribble annotations. Finally, [51] generate pseudo annotation masks for weakly-supervised instance segmentation in a multiple instance learning framework.", + "bbox": [ + 212, + 479, + 785, + 570 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Different from the weakly-supervised approaches utilizing SAM, we do not rely on supervision. Hence, a key novelty of our work lies in investigating object priors to focus SAM on segmenting arbitrary objects in the absence of supervision for all object classes. Moreover, we are the first to apply SAM-based pseudo annotations in OWIS.", + "bbox": [ + 212, + 570, + 785, + 646 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Segment Object System for OWIS", + "text_level": 1, + "bbox": [ + 215, + 670, + 584, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address OWIS, we propose our Segment Object System (SOS) that generates high-quality pseudo annotations to train a standard instance segmentation system. As visible from Fig. 2, SOS consists of three main blocks. The first block, the Object Localization Module (OLM), described in Sec. 3.2 (yellow area in Fig. 2), aims to roughly localize objects by sampling locations from an object prior that reflects the object probability per image coordinate. The output of OLM is a set of likely object locations, that is used in the Pseudo Annotation Creator (PAC), described in Sec. 3.3 (green area in Fig. 2), to prompt the pre-trained SAM [27], which we briefly revisit in Sec. 3.1. From the segments", + "bbox": [ + 212, + 703, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "C. Wilms et al.", + "bbox": [ + 271, + 114, + 375, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6f7a464a696853056538da30d03ea8fd1653e04bd52c82b1979dab2ef6b1bbd2.jpg", + "image_caption": [ + "Fig. 2: Overview of our Segment Object System (SOS) for OWIS consisting of three blocks. First, the input image is processed in our Object Localization Module (OLM, yellow area) to create object-focused point prompts roughly localizing objects. Second, our Pseudo Annotations Creator (PAC, green area) generates segments based on the previously generated prompts using SAM [27], and further processes them, leading to a final set of merged original and pseudo annotations. Finally, the merged annotations are used to train an instance segmentation system (blue area)." + ], + "image_footnote": [], + "bbox": [ + 215, + 142, + 787, + 429 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "generated from likely object locations, the PAC generates pseudo annotations by filtering low-quality segments and removing near-duplicates. Finally, the pseudo annotations are merged with the original annotations, and a standard instance segmentation system is trained with the merged annotations (see Sec. 3.4, blue area in Fig. 2). During testing, only this instance segmentation system is used.", + "bbox": [ + 212, + 566, + 787, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Revisit Segment Anything Model", + "text_level": 1, + "bbox": [ + 214, + 662, + 539, + 680 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since SAM [27] is integral to SOS, we will briefly review the model. The idea of SAM is to utilize prompts such as point coordinates to generate segments of image parts like objects, object parts, or stuff regions localized by these prompts. Hence, SAM does not exclusively segment objects unless explicitly instructed by object-focused prompts, as proposed in this paper. SAM is based on a ViT [11] image encoder creating an image embedding, while the embeddings for point prompts are generated based on positional encodings. Both embeddings are processed in a mask decoder utilizing a transformer decoder block and a mask prediction head to create the final segment including a confidence score. Note that SAM is class-agnostic and does not classify the segments.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "SOS: Segment Object System", + "bbox": [ + 532, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For training SAM, [27] propose the new large-scale SA-1B dataset that is annotated with a data engine using SAM itself. In the pre-stage, SAM is trained with public segmentation datasets to aid the first stage of the data engine, where annotators use SAM as an interactive segmentation tool to create initial segmentations. Subsequently, SAM is retrained with the newly annotated data, fully avoiding data leakage w.r.t. public segmentation datasets. In subsequent stages, manual annotations were provided for segments missed by SAM. At last, annotations were created fully automatically by SAM. See [27] for more details.", + "bbox": [ + 212, + 146, + 787, + 268 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Object Localization Module", + "text_level": 1, + "bbox": [ + 214, + 287, + 493, + 304 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The first block of SOS (yellow area in Fig. 2) is the Object Localization Module (OLM). OLM roughly localizes all objects in the input image, and creates object-focused point prompts. As the first step, OLM generates an object prior from the input image. The object prior is a probability mass function constrained to the image plane, although other formulations are possible (see Sec. 4), and indicates the probability of an image coordinate being part of any object. Hence, the object prior is class-agnostic and not limited to known object classes from the training set in OWIS. In our study in Sec. 4, we analyze various object priors. For final SOS, we utilize the six self-attention maps from the self-attention heads of a ViT's final layer, pre-trained in the self-supervised DINO framework [6] on ImageNet [45]. See Sec. 4 for justifications. OLM aggregates the six self-attention maps by elementwise max yielding one map highlighting all relevant scene elements. Subsequently, OLM converts the map to an object prior, by rescaling such that the minimum equals 0 and the sum over the object prior is 1.", + "bbox": [ + 212, + 311, + 787, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Next, OLM randomly samples $S$ (here: $S = 50$ ) image coordinates from the object prior, given the per-coordinate values of the object prior as probabilities. To diversify the samples across multiple objects, we prune parts of the object prior around the sampled coordinate. Specifically, given a coordinate $(x_s, y_s)$ , we set the object prior to zero for all coordinates $\\{(x, y) \\mid |x - x_s| \\leq N \\land |y - y_s| \\leq N\\}$ , with $N = 20$ . To transform the object prior back to a probability mass function, we again apply the rescaling as described above. Subsequently, OLM iteratively applies this sampling until a set of $S$ coordinates are extracted, representing object-focused point prompts. This set is the output of the OLM and will be used for prompting in the Pseudo Annotation Creator.", + "bbox": [ + 212, + 523, + 789, + 676 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Pseudo Annotations Creator", + "text_level": 1, + "bbox": [ + 214, + 695, + 501, + 709 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The Pseudo Annotations Creator (PAC, green area in Fig. 2) is the second block of SOS. It generates pseudo annotations utilizing SAM, given the object-focused point prompts provided by OLM. These object-focused point prompts lifts SAM from a system that segments anything to a system that segments objects. As the first step, PAC prompts a pre-trained SAM with the $S$ point prompts from OLM. To handle ambiguous point prompts that may indicate multiple objects or object parts, we follow [27] and allow SAM to generate three segments per prompt. Each generated segment is a potential pseudo annotation. The resulting set of", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "C. Wilms et al.", + "bbox": [ + 271, + 114, + 375, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "class-agnostic segments is noisy, with segments covering objects and background, or covering the same image area. Therefore, PAC first utilizes SAM's confidence score per segment and removes segments with a confidence below $\\tau_{\\mathrm{conf}}$ . Second, PAC removes near-duplicates by non-maximum suppression with an IoU threshold $\\tau_{\\mathrm{NMS}}$ . We set $\\tau_{\\mathrm{conf}} = 0.9$ and $\\tau_{\\mathrm{NMS}} = 0.95$ for SOS.", + "bbox": [ + 212, + 146, + 787, + 222 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Given this set of pseudo annotations, PAC merges the set with the original annotations representing known classes of the training dataset to combine the knowledge of the human annotators and the knowledge of our object-focused SAM. Since we only want to keep pseudo annotations for objects not covered by the original annotations, we suppress pseudo annotations that have a high IoU ( $\\tau_{\\mathrm{NMS}}$ ) with at least one original annotation. Moreover, we limit the number of pseudo annotations to $P$ in order to balance original and pseudo annotations. In our final SOS, we set $P = 10$ . However, across the COCO training dataset, only 7.8 pseudo annotations are added on average. The other pseudo annotations were suppressed in previous steps. Overall, the output of PAC is a mixed set of annotations, containing all original annotations augmented with high-quality pseudo annotations covering objects of unknown classes.", + "bbox": [ + 212, + 222, + 789, + 402 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Instance Segmentation", + "text_level": 1, + "bbox": [ + 214, + 422, + 450, + 436 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the last block of SOS, we take the merged annotations generated by PAC and train a standard instance segmentation system. While this system can be arbitrary and is generally replaceable, we choose Mask R-CNN [19] with a ResNet-50+FPN backbone [30] trained in a class-agnostic setting, following recent OWIS methods [24, 26, 46, 49]. Overall, this leads to an OWIS method based on high-quality pseudo annotations.", + "bbox": [ + 212, + 445, + 787, + 536 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Object Priors for SOS", + "text_level": 1, + "bbox": [ + 214, + 556, + 468, + 574 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As described in Sec. 3.2, SOS uses an object prior for rough object localization and subsequent point prompt generation for PAC. This section thoroughly investigates the performance of various object priors derived from previous class-agnostic object localization works in SOS. For details on the object priors beyond the subsequent descriptions, we refer to our supplementary.", + "bbox": [ + 212, + 585, + 787, + 662 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1 Object Priors", + "text_level": 1, + "bbox": [ + 214, + 681, + 375, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines: We introduce three baseline object priors that ignore the image content. First, Grid uses SAM with a regular $64 \\times 64$ grid of points, following [27]. This directly leads to prompts without the sampling described in Sec. 3.2. Note that the Grid baseline will lead to segments for both object and stuff regions. Second, we use the spatial distribution of object centroids from the known classes across the training dataset as object prior (Dist). We assume that this generalizes well since it ignores class-specific object features. Finally, we utilize the centroids of the training set's objects from unknown classes as optimal prompts ( $GT$ ), representing an upper bound.", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "SOS: Segment Object System", + "bbox": [ + 531, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Superpixels: Inspired by [15, 52], we utilize the centroids of superpixels generated with an adaptive superpixel segmentation method as object prior ( $Spx$ ). The intuition is that large uniform areas not containing objects should be covered by few superpixels and vice versa. Hence, the density of superpixels is a surrogate for the density of objects. We use the well-known FH superpixels [13] since they adapt their density to the image content following our assumption. Note that every superpixel centroid is a point prompt.", + "bbox": [ + 212, + 146, + 787, + 253 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Contour Density: Since objects are defined by their outer contours [63], we use contour density as an object prior (Contour). Similar to [52], we assume that in areas of high contour density, several objects are located and vice versa. As a surrogate for the contour density, we use edge density based on [10].", + "bbox": [ + 212, + 263, + 789, + 325 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Saliency: Saliency is used in object proposal generation to localize objects [2, 37] as they stick out from their surrounding. Here, we evaluate two saliency methods as object priors. First, we apply the traditional approach VOCUS2 [14] based on color contrast and used by [37] (VOCUS2). Second, we use pre-trained DeepGaze IIE [32] saliency maps learned from eye fixation data (DeepGaze).", + "bbox": [ + 212, + 335, + 787, + 414 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Class Activation Maps: CAMs [60] indicate discriminative image regions for CNN-based classifiers. Since they only need image-level supervision, they are frequently applied in weakly-supervised tasks to locate objects [1, 25, 35]. We use the CAM of the predicted class in a ResNet-50 classifier [20] pre-trained on ImageNet [45] as object prior (CAM).", + "bbox": [ + 212, + 425, + 787, + 502 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Self-attention: Recently, [6] have shown that self-attention maps learned inside self-supervised ViTs encode the scene layout, including object locations. Therefore, we explore the self-attention maps from the final layer of a ViT-S backbone trained in the self-supervised DINO framework [6] as an object prior (DINO).", + "bbox": [ + 212, + 512, + 787, + 574 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Learned Object Locations: Finally, we learn object locations based on the known classes of a dataset using a U-Net [44] in a binary segmentation task (object vs. no object) as object prior ( $U$ -Net). This assumes that, on the level of individual pixels, the model generalizes from known to unknown classes [49,56].", + "bbox": [ + 212, + 585, + 787, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 Study Setup for Object Priors", + "text_level": 1, + "bbox": [ + 212, + 666, + 511, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To assess each object prior, we train SOS with the respective prior on the COCO training set using only annotations from the 20 PASCAL VOC [12] classes. We evaluate on COCO's validation set with annotations from the remaining 60 classes in COCO, following standard COCO (VOC) $\\rightarrow$ COCO (non-VOC) cross-category evaluation in OWIS [26, 46, 49]. Note that we use the default class-agnostic Mask R-CNN in SOS but reduce the training schedule to only a quarter of steps for faster training. We evaluate the results of SOS given each object priors and report Average Recall (AR) for 100 detections and Average Precision (AP) as previous OWIS research [46, 49]. We also report $F_{1}$ score, the harmonic mean between AR and AP, yielding a single number for comparison.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "C. Wilms et al.", + "bbox": [ + 271, + 114, + 375, + 128 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/98c534e1f46cf8af01435b9b48398d2a912be5cb6c56f5d701d91c242fcf430b.jpg", + "image_caption": [ + "Input image" + ], + "image_footnote": [], + "bbox": [ + 233, + 146, + 442, + 256 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/003002287514d3fb661c3340c460d61ece1d66380508a497953a7041dc9c0ebe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 444, + 143, + 552, + 200 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/45e5bc61bd5b28ddc0bdbad119f694a1ab734f9401f8e33570f704fc9f26f2b4.jpg", + "image_caption": [ + "VOCUS2", + "Fig. 3: Object priors VOCUS2, U-Net, and DINO with resulting pseudo annotations." + ], + "image_footnote": [], + "bbox": [ + 444, + 200, + 552, + 257 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/867b8ce8c01c3d142b7ac46172c54b92a51ac566435015d9e7a2a0e8386ed0fe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 553, + 143, + 661, + 200 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/e18673af3e9f66db67f8d6874d2cbc02569e3bc22acb0b15c3d3538437de1770.jpg", + "image_caption": [ + "U-Net" + ], + "image_footnote": [], + "bbox": [ + 553, + 200, + 661, + 257 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/f8091942473796a09656b9da90cd195fc51d5c9f4bf39e4ad0622fae5d28d8ff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 663, + 143, + 769, + 200 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/deb4b2e0d49cc2d6854a38a525247c347665344734ea87235a23b97376720973.jpg", + "image_caption": [ + "DINO" + ], + "image_footnote": [], + "bbox": [ + 663, + 200, + 769, + 257 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3 Results of Object Priors in SOS", + "text_level": 1, + "bbox": [ + 214, + 330, + 526, + 345 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 1 shows the results of our object prior study. First, all priors outperform the baselines Grid and Dist in terms of $\\mathbf{F}_1$ . Both baselines exhibit a lower AP compared to all other object priors, reflecting the missing focus on objects. Hence, simply applying SAM (Grid) is not suitable to segment only objects, as several stuff regions are segmented as well. Second, most priors exhibit a similar performance in terms of all measures. For instance, learning-based CAM and DeepGaze do not outperform simple priors (Spx, Contour, VOCUS2). Only the priors U-Net and DINO perform substantially better, with DINO producing the best result. Comparing DINO to GT reveals that DINO is able to recall almost the same amount of objects, but exhibits a lower precision, as expected.", + "bbox": [ + 212, + 358, + 787, + 508 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To illustrate the results, Fig. 3 depicts object priors for top-performing VOCUS2, $U$ -Net, and DINO, as heatmaps (upper row) with the resulting pseudo annotations generated in SOS (lower row) for an example image. The heatmaps show that DINO is very focused on the discriminative object parts, here the polar bears' faces, while $U$ -Net highlights, as learned, the entire objects. Moreover, $U$ -Net also highlights several background areas. Similarly, VOCUS2 highlights background areas and the polar bears since it is generally focused on high-contrast areas. Overall, only DINO exclusively generates pseudo annotations for the polar bears and some sub-parts, while $U$ -Net and VOCUS2 produce noisy annotations including background patches, similar to previous approaches [49].", + "bbox": [ + 212, + 508, + 787, + 660 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The results show that a well-designed object prior like DINO substantially outperforms the baseline Grid $(+7.5$ in $\\mathrm{F}_1)$ in SOS. Therefore, we use DINO as the object prior in SOS. More generally, the study also reveals important insights on how to prompt SAM for class-agnostic, object-focused results for arbitrary tasks. For more study results, we refer to our supplementary.", + "bbox": [ + 212, + 660, + 787, + 736 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 Evaluation", + "text_level": 1, + "bbox": [ + 214, + 760, + 356, + 776 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In our evaluation, we compare SOS to recent OWIS methods OLN [26], LDET [46], GGN [49], SWORD [54], SOIS [56], OpenInst [48], UDOS [24], and against Mask R-CNN [19] and SAM [27] baselines, depending on the availability of public", + "bbox": [ + 212, + 792, + 792, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "SOS: Segment Object System", + "bbox": [ + 531, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/66088e09959caa645f7fca47fc3764592d6668c76ae9b69873a847a33ac854ca.jpg", + "table_caption": [ + "Table 1: Results of our SOS with various object priors in the cross-category COCO (VOC) $\\rightarrow$ COCO (non-VOC) setting. *: Uses ground truth of unknown classes." + ], + "table_footnote": [], + "table_body": "
Object PriorAPAR100F1Object PriorAPAR100F1
SOS + Grid3.836.56.9SOS + DeepGaze5.435.99.4
SOS + Dist3.427.46.0SOS + CAM5.436.79.4
SOS + Spx5.634.89.6SOS + DINO8.938.114.4
SOS + Contour5.636.69.7SOS + U-Net7.337.312.2
SOS + VOCUS25.137.710.5SOS + GT18.1*42.5*25.4*
", + "bbox": [ + 263, + 184, + 738, + 285 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "source code or reported results. To assess the quality of the approaches, we use four cross-category and cross-dataset settings, commonly used in OWIS. Specifically, we choose COCO (VOC) $\\rightarrow$ COCO (non-VOC) on the COCO dataset [31], already discussed in Sec. 4.2. For cross-dataset evaluation, we use COCO $\\rightarrow$ LVIS, COCO $\\rightarrow$ ADE20k, and COCO $\\rightarrow$ UVO that train on the entire COCO training set and evaluate on the validation sets of LVIS [17], ADE20k [61], and UVO [50] with exhaustive annotations. The LVIS dataset covers the same images as COCO but extends the annotated object classes from 80 to 1203. ADE20k includes 3169 classes of objects, object parts, and stuff regions. Finally, UVO features exhaustive annotations of all objects and, therefore, no system is penalized for detecting objects outside a datasets' taxonomy.", + "bbox": [ + 212, + 316, + 787, + 484 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "As in Sec. 4.2 and following common practice in OWIS [26,46,49], we report AR for up to 100 detections and AP. We again also report the $\\mathrm{F}_1$ score between AR and AP as a single, combined quantity, described in Sec. 4.2. Note that we only evaluate based on masks, not boxes.", + "bbox": [ + 212, + 484, + 787, + 546 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.1 Implementation Details and Data Usage", + "text_level": 1, + "bbox": [ + 214, + 577, + 594, + 593 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "As discussed in Sec. 4.3, SOS uses the object prior DINO for best results. DINO [6] with a ViT-S backbone is used as an ImageNet [45] pre-trained model without fine-tuning. Similarly, SOS uses SAM [27] with the ViT-H backbone as a pre-trained model based on SA-1B [27] without fine-tuning. As an instance segmentation model, we apply Mask R-CNN [19] with an ImageNet pre-trained ResNet-50+FPN backbone [30], and the default configurations except for the class-agnostic training.", + "bbox": [ + 212, + 611, + 787, + 717 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Overall, SOS is based on models pre-trained on standard ImageNet and SA-1B. Pre-training on SA-1B or similar large-scale datasets becomes common with the emergence of foundation models and is widely adopted by systems using SAM in similar contexts [7,51,57]. It is comparable to using models pre-trained on ImageNet that even have class information available during training and also allow a rough localization of objects [6,60]. Moreover, the final, noisy annotations of SA-1B representing objects, object parts, and stuff region were automatically generated by SAM [27], and no data leakage w.r.t. other datasets exists.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "C. Wilms et al.", + "bbox": [ + 271, + 114, + 375, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/e1ecb68f6776a93a15f6a3710ff52f5b88420b7ae1eccc62ac3c834c66237b00.jpg", + "table_caption": [ + "Table 2: Results of two baselines and various OWIS methods in the COCO (VOC) $\\rightarrow$ COCO (non-VOC) setting. $\\dagger$ : uses automatically annotated SA-1B dataset." + ], + "table_footnote": [], + "table_body": "
SystemAPAR100F1
Mask R-CNN [19]1.08.21.8
SAM†[27]3.648.16.7
OLN [26]4.228.47.3
LDET [46]4.324.87.3
GGN [49]4.928.38.4
SWORD [54]4.830.28.3
UDOS [24]2.934.35.3
SOS†(ours)8.939.314.5
", + "bbox": [ + 236, + 224, + 470, + 381 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/623c27c08bdf2b2822d6a7eb01985604d642233c5ec3a310235b526d35de59dd.jpg", + "table_caption": [ + "Table 3: Results of two baselines and various OWIS methods in the COCO $\\rightarrow$ LVIS setting. $^\\dagger$ : uses automatically annotated SA-1B dataset." + ], + "table_footnote": [], + "table_body": "
SystemAPAR100F1
Mask R-CNN [19]7.523.611.4
SAM†[27]6.845.111.8
LDET [46]6.724.810.5
GGN [49]6.527.010.5
SOIS [56]-25.2-
OpenInst [48]-29.3-
UDOS [24]3.924.96.7
SOS†(ours)8.133.313.3
", + "bbox": [ + 531, + 212, + 764, + 367 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.2 Cross-category COCO (VOC) $\\rightarrow$ COCO (non-VOC) Results", + "text_level": 1, + "bbox": [ + 215, + 410, + 756, + 426 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 2 presents the results of various OWIS methods on the cross-category setup COCO (VOC) $\\rightarrow$ COCO (non-VOC). Our SOS clearly outperforms all OWIS methods in $\\mathrm{AR}_{100}$ , AP, and $\\mathbf{F}_1$ . Specifically, SOS outperforms the second-best system in terms of $\\mathrm{F}_1$ , GGN [49], by 6.1. While some of the improvement comes from a better recall ( $+38.9\\%$ compared to GGN), the results are driven by a much-improved precision ( $+81.6\\%$ compared to GGN). This reflects the high quality of the pseudo annotations in SOS on this cross-category setting. Moreover, there is a clear improvement of SOS over Mask R-CNN ( $+12.7$ in $\\mathrm{F}_1$ ). Compared to original SAM, the DINO-based SOS substantially improves the precision by focusing on objects, leading to an improved $\\mathrm{F}_1$ score ( $+7.8$ ).", + "bbox": [ + 212, + 438, + 787, + 590 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The qualitative results of various OWIS methods in Fig. 4 support the quantitative results. In the first example, only GGN [49] and SOS detect both giraffes, however, SOS generates more accurate segmentations. In the second example, only SOS detects all five small surfboards, while other systems detect at most three. Note that class-agnostic Mask R-CNN misses all non-VOC objects in both examples. For more qualitative results, see our supplementary.", + "bbox": [ + 212, + 590, + 787, + 681 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.3 Cross-dataset Results", + "text_level": 1, + "bbox": [ + 215, + 705, + 441, + 720 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "$COCO \\rightarrow LVIS$ : We evaluate the cross-dataset generalization capabilities of OWIS methods, starting with $COCO \\rightarrow LVIS$ . The results in Tab. 3 show a trend similar to the cross-category results. SOS outperforms all OWIS methods across all measures with an improvement of 2.8 in $F_1$ compared to second-best methods GGN [49] and LDET [46]. Similar to the previous results, the improvement is based on gains in both recall and precision, however, the relative improvements in AP and $AR_{100}$ are more similar to each other. Overall, our high-quality pseudo", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "SOS: Segment Object System", + "bbox": [ + 531, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/235737833757eee12f46f0948406e067ab08097a1e13ad6052fe5c94d75f0513.jpg", + "image_caption": [ + "Fig. 4: Qualitative results of OWIS methods and baseline Mask R-CNN in the cross-category COCO (VOC) $\\rightarrow$ COCO (non-VOC) setting. Filled masks denote detected objects, while red frames indicate missed objects." + ], + "image_footnote": [], + "bbox": [ + 218, + 142, + 789, + 290 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/6bc6676155d9cdb73b115bb18cb55aaf8354b3eaaf2183c16ba577deaab74bb4.jpg", + "table_caption": [ + "Table 4: Results of Mask R-CNN baseline and various OWIS methods in the COCO $\\rightarrow$ ADE20k setting. $^\\dagger$ : uses automatically annotated SA-1B dataset." + ], + "table_footnote": [], + "table_body": "
SystemAPAR100F1
Mask R-CNN [19]6.911.98.7
OLN [26]-20.4-
LDET [46]9.518.512.6
GGN [49]9.721.013.3
UDOS [24]7.622.911.4
SOS† (ours)12.526.517.0
", + "bbox": [ + 233, + 431, + 475, + 563 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/e46675cfa54e38ab343a8703b83b50ff2882ecde760656958f3553f891ac9520.jpg", + "table_caption": [ + "Table 5: Results of two baselines and various OWIS methods in the COCO $\\rightarrow$ UVO setting. $^\\dagger$ : uses automatically annotated SA-1B dataset." + ], + "table_footnote": [], + "table_body": "
SystemAPAR100F1
Mask R-CNN [19]20.736.726.5
SAM†[27]11.350.118.4
OLN [26]-41.4-
LDET [46]22.040.428.5
GGN [49]20.343.427.7
UDOS [24]10.643.117.0
SOS†(ours)20.942.328.0
", + "bbox": [ + 526, + 426, + 771, + 568 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "annotations in SOS lead to new state-of-the-art results, which indicate strong generalization to unknown object classes outside COCO.", + "bbox": [ + 212, + 595, + 787, + 628 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "$COCO \\rightarrow ADE20k$ : The results for $COCO \\rightarrow ADE20k$ in Tab. 4, which include labeled object parts of $ADE20k$ , show that SOS again outperforms all other OWIS methods across all measures. The gain over the second-best method in $F_1$ , GGN [49], is 3.7. Similar to $COCO \\rightarrow LVIS$ , the relative gains in $AR_{100}$ and AP are equally distributed. These results indicate that object parts do not degrade the results of SOS. This is in line with SOS's pseudo annotations in Fig. 3, covering entire objects and selected object parts.", + "bbox": [ + 212, + 642, + 787, + 750 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "$COCO \\rightarrow UVO$ : Finally, Tab. 5 presents results for COCO $\\rightarrow$ UVO. SOS still outperforms Mask R-CNN, SAM and recent UDOS [24] in $F_1$ . Outperforming SAM implies that SOS does not exploit the limited taxonomy of annotations in COCO (only 60 object classes in test), but also improves the segmentation of objects outside the COCO object classes on the exhaustively labeled UVO", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "C. Wilms et al.", + "bbox": [ + 271, + 114, + 375, + 127 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/fd46bf28b7c4d2b49bc604b828d7560109f4cc7142fb950d0ed1ae9b51186311.jpg", + "table_caption": [ + "Table 6: Quality of pseudo annotation generated in GGN [49] and our SOS evaluated on the non-VOC annotations of the COCO training dataset. $\\dagger$ : uses automatically annotated SA-1B dataset." + ], + "table_footnote": [], + "table_body": "
SystemPrec.Rec.F1
GGN3 [49]7.312.19.1
SOS3† (ours)19.026.422.1
SOS10† (ours)15.541.722.6
", + "bbox": [ + 254, + 222, + 454, + 301 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/394330a6f99d5e5240e6cc209dbeaa13e7415f348e39041847517301f8eeddb5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 535, + 143, + 647, + 202 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/fe5e1a842cec4d04dd0b27b3f9f5b51cd31e6e815eb15bddde46e336af14f87f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 650, + 143, + 759, + 202 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/9304cbb3b90f3cfe5e5a7ff9f135b3e7c850345a8da0b789f1e5f3bcac0973b1.jpg", + "image_caption": [ + "GGN" + ], + "image_footnote": [], + "bbox": [ + 535, + 203, + 647, + 262 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/0a44d544eb21135caae4783ce969e642135dac6ffc058018478b0abd9141cd59.jpg", + "image_caption": [ + "SOS", + "Fig. 5: Pseudo annotations generated by GGN and our SOS." + ], + "image_footnote": [], + "bbox": [ + 650, + 203, + 759, + 262 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "dataset. However, the results of SOS are below LDET [46] in AP and $\\mathrm{F_1}$ . We attribute this to several classes in UVO not available in ImageNet, which was used to learn DINO object prior. Hence, DINO will miss such objects in the COCO dataset during training, and SOS might miss them during testing. This could be mitigated using more diverse unlabeled images for training DINO.", + "bbox": [ + 212, + 349, + 787, + 425 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.4 Quality of Pseudo Annotations", + "text_level": 1, + "bbox": [ + 214, + 448, + 517, + 463 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We also investigate the quality of the pseudo annotations generated in SOS. To this end, we evaluate the pseudo annotations on the COCO training set against the annotations of non-VOC object classes, reflecting the COCO (VOC) $\\rightarrow$ COCO non-VOC) setting. The results in terms of precision and recall for IoU $= 0.5$ as well as the $\\mathrm{F_1}$ score are presented in Tab. 6. We use up to three and 10 pseudo annotations generated per image in SOS (SOS $_3$ and $\\mathrm{SOS}_{10}$ ), and $\\mathrm{GGN}_3$ [49] that generates three pseudo annotations. The results show that with only three SOS pseudo annotations, more than $25\\%$ of the non-VOC objects are covered. This is substantially more than $\\mathrm{GGN}_3$ ( $12.1\\%$ ). With up to 10 annotations, SOS's pseudo annotations cover more than $40\\%$ of the non-VOC objects. In our supplementary, we further provide class-specific results.", + "bbox": [ + 212, + 473, + 787, + 638 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "To complement these results, we visualize the annotations of $\\mathrm{GGN}_3$ and $\\mathrm{SOS}_3$ on two sample images in Fig. 5. It is visible that $\\mathrm{GGN}_3$ 's annotations cover background regions, leading to low precision in the overall results. Moreover, the pseudo annotations in the upper example do not adhere well to the airplane's boundaries. More qualitative examples are given in Fig. 1 and our supplementary.", + "bbox": [ + 212, + 638, + 787, + 715 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.5 Ablation Studies", + "text_level": 1, + "bbox": [ + 214, + 738, + 401, + 752 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "All ablation studies follow the evaluation setup described in Sec. 4.2. More studies evaluating design choices in SOS are presented in our supplementary.", + "bbox": [ + 212, + 763, + 785, + 794 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Influence of SOS's Components: We investigate the importance of each component in SOS and present the results in Tab. 7. The baseline for this", + "bbox": [ + 214, + 809, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "SOS: Segment Object System", + "bbox": [ + 531, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/48fc97244774b6490dd63b756e42f697266b4851a75b809f661c5b8294941d90.jpg", + "table_caption": [ + "Table 7: Influence of added components in our SOS evaluated in the COCO (VOC) $\\rightarrow$ COCO (non-VOC) setting." + ], + "table_footnote": [], + "table_body": "
ComponentsAPAR100F1
Mask R-CNN1.210.82.2
+ Grid, w/o pp.3.435.26.2
+ DINO, w/o pp.8.937.114.3
+ DINO, w/ pp.8.938.114.4
", + "bbox": [ + 238, + 198, + 470, + 285 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/aa823a562d9ae7e99fbc5f1e6aa9d3073931de7b3c46402c2f0774326ca1b945.jpg", + "table_caption": [ + "Table 8: SOS results with various numbers of pseudo annotations in the COCO (VOC) $\\rightarrow$ COCO (non-VOC) setting." + ], + "table_footnote": [], + "table_body": "
#Pseudo Anns.APAR100F1
38.334.513.4
58.836.214.2
108.938.114.4
208.838.114.3
", + "bbox": [ + 540, + 198, + 756, + 285 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "study is a class-agnostic Mask R-CNN trained without pseudo annotations (first row in Tab. 7). Subsequently, we add pseudo annotations based on SAM with the baseline Grid object prior as in [27] and without our post-processing in PAC (confidence-based thresholding and NMS, see Sec. 3.3). Next, we replace Grid with DINO (third row in Tab. 7), and finally add the post-processing leading to the complete SOS. The results in Tab. 7 show each step improves the $\\mathrm{F}_1$ results. Mainly, introducing SAM for generating pseudo annotations (first row vs. second row) substantially improves $\\mathrm{AR}_{100}$ , while adding the DINO object prior removes background annotations leading to considerably improved precision. Hence, the object prior selection in SOS is crucial for strong results.", + "bbox": [ + 212, + 310, + 787, + 464 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Number of Pseudo Annotations: In the second ablation study, we investigate the influence of the number of pseudo annotations per image on the results of SOS. To this end, we evaluate SOS with up to 3, 5, 10, and 20 pseudo annotations per image. The results in Tab. 8 show that 10 pseudo annotations are preferable. Hence, the ideal number of pseudo annotations in SOS is larger than in GGN [49], which only uses three. This indicates again the high quality of our pseudo annotations, since more than three annotations per image can cover relevant objects without adding too many noisy annotations.", + "bbox": [ + 212, + 473, + 789, + 595 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 614, + 361, + 630 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we addressed the challenging OWIS task and aimed to improve the low precision of previous methods in this open-world task. To improve the precision and the overall detection results, we generate high-quality pseudo annotations based on a prompt-guided foundation model, SAM. To focus the pseudo annotations on objects, we thoroughly investigated various object priors for prompt creation, which revealed important insights on how to prompt SAM for class-agnostic object-focused results in arbitrary tasks. As a result, our novel OWIS method SOS, which uses a self-attention-based object prior for generating pseudo annotations, outperforms recent state-of-the-art OWIS methods across challenging open-world setups in COCO, LVIS, and ADE20k datasets. SOS especially improves the precision based on the high-quality pseudo annotations. Moreover, SOS with its focus on objects also outperforms vanilla SAM that segments both object and stuff regions.", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "C. Wilms et al.", + "bbox": [ + 271, + 114, + 375, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Ahn, J., Cho, S., Kwak, S.: Weakly supervised learning of instance segmentation with inter-pixel relations. In: Conference on Computer Vision and Pattern Recognition (2019)", + "2. Alexe, B., Deselaers, T., Ferrari, V.: What is an object? In: Conference on Computer Vision and Pattern Recognition (2010)", + "3. Baugh, M., Batten, J., Müller, J.P., Kainz, B.: Zero-shot anomaly detection with pre-trained segmentation models. arXiv preprint arXiv:2306.09269 (2023)", + "4. Bommasani, R., Hudson, D.A., Adeli, E., Altman, R., Arora, S., von Arx, S., Bernstein, M.S., Bohg, J., Bosselut, A., Brunskill, E., et al.: On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258 (2021)", + "5. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in Neural Information Processing Systems 33 (2020)", + "6. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: International Conference on Computer Vision (2021)", + "7. Chen, T., Mai, Z., Li, R., Chao, W.l.: Segment anything model (SAM) enhanced pseudo labels for weakly supervised semantic segmentation. arXiv preprint arXiv:2305.05803 (2023)", + "8. Cheng, M.M., Mitra, N.J., Huang, X., Torr, P.H., Hu, S.M.: Global contrast based salient region detection. IEEE Transactions on Pattern Analysis and Machine Intelligence 37(3) (2014)", + "9. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: BERT: Pre-training of deep bidirectional transformers for language understanding. In: Human Language Technology Conference of the North American Chapter of the Association for Computational Linguistics (2019)", + "0. Dollar, P., Zitnick, C.L.: Structured forests for fast edge detection. In: International Conference on Computer Vision (2013)", + "1. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2020)", + "2. Everingham, M., Eslami, S.A., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The PASCAL visual object classes challenge: A retrospective. International Journal of Computer Vision 111 (2015)", + "3. Felzenszwalb, P.F., Huttenlocher, D.P.: Efficient graph-based image segmentation. International Journal of Computer Vision 59 (2004)", + "4. Frintrop, S., Werner, T., Martin Garcia, G.: Traditional saliency reloaded: A good old model in new shape. In: Conference on Computer Vision and Pattern Recognition (2015)", + "5. Gao, G., Lauri, M., Zhang, J., Frintrop, S.: Saliency-guided adaptive seeding for supervoxel segmentation. In: International Conference on Intelligent Robots and Systems (2017)", + "6. Girshick, R.: Fast R-CNN. In: International Conference on Computer Vision (2015)", + "7. Gupta, A., Dolkar, P., Girshick, R.: LVIS: A dataset for large vocabulary instance segmentation. In: Conference on Computer Vision and Pattern Recognition (2019)", + "8. He, C., Li, K., Zhang, Y., Xu, G., Tang, L., Zhang, Y., Guo, Z., Li, X.: Weakly-supervised concealed object segmentation with SAM-based pseudo labeling and multi-scale feature grouping. arXiv preprint arXiv:2305.11003 (2023)" + ], + "bbox": [ + 225, + 172, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "SOS: Segment Object System", + "bbox": [ + 531, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "19. He, K., Gkioxari, G., Dollar, P., Girshick, R.: Mask R-CNN. In: International Conference on Computer Vision (2017)", + "20. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Conference on Computer Vision and Pattern Recognition (2016)", + "21. Hou, Q., Cheng, M.M., Hu, X., Borji, A., Tu, Z., Torr, P.H.: Deeply supervised salient object detection with short connections. In: Conference on Computer Vision and Pattern Recognition (2017)", + "22. Hu, H., Lan, S., Jiang, Y., Cao, Z., Sha, F.: FastMask: Segment multi-scale object candidates in one shot. In: Conference on Computer Vision and Pattern Recognition (2017)", + "23. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International Conference on Machine Learning (2021)", + "24. Kalluri, T., Wang, W., Wang, H., Chandraker, M., Torresani, L., Tran, D.: Openworld instance segmentation: Top-down learning with bottom-up supervision. arXiv preprint arXiv:2303.05503 (2023)", + "25. Kim, B., Yoo, Y., Rhee, C.E., Kim, J.: Beyond semantic to instance segmentation: Weakly-supervised instance segmentation via semantic knowledge transfer and self-refinement. In: Conference on Computer Vision and Pattern Recognition (2022)", + "26. Kim, D., Lin, T.Y., Angelova, A., Kweon, I.S., Kuo, W.: Learning open-world object proposals without learning to classify. IEEE Robotics and Automation Letters 7(2) (2022)", + "27. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: International Conference on Computer Vision (2023)", + "28. Klein, D.A., Frintrop, S.: Center-surround divergence of feature statistics for salient object detection. In: International Conference on Computer Vision (2011)", + "29. Li, J., Li, D., Savarese, S., Hoi, S.: BLIP-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023)", + "30. Lin, T.Y., Dollar, P., Girshick, R., He, K., Hariharan, B., Belongie, S.: Feature pyramid networks for object detection. In: Conference on Computer Vision and Pattern Recognition (2017)", + "31. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft COCO: Common objects in context. In: European Conference on Computer Vision (2014)", + "32. Linardos, A., Kümmerer, M., Press, O., Bethge, M.: DeepGaze IIE: Calibrated prediction in and out-of-domain for state-of-the-art saliency modeling. In: International Conference on Computer Vision (2021)", + "33. Liu, N., Han, J., Yang, M.H.: PiCANet: Learning pixel-wise contextual attention for saliency detection. In: Conference on Computer Vision and Pattern Recognition (2018)", + "34. Liu, T., Yuan, Z., Sun, J., Wang, J., Zheng, N., Tang, X., Shum, H.Y.: Learning to detect a salient object. IEEE Transactions on Pattern analysis and machine intelligence 33(2) (2010)", + "35. Liu, Y., Wu, Y.H., Wen, P., Shi, Y., Qiu, Y., Cheng, M.M.: Leveraging instance-, image-and dataset-level information for weakly supervised instance segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(3) (2020)", + "36. Ma, J., Ming, A., Huang, Z., Wang, X., Zhou, Y.: Object-level proposals. In: International Conference on Computer Vision (2017)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "C. Wilms et al.", + "bbox": [ + 271, + 114, + 374, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "37. Martin García, G., Potapova, E., Werner, T., Zillich, M., Vincze, M., Frintrop, S.: Saliency-based object discovery on RGB-D data with a late-fusion approach. In: International Conference on Robotics and Automation (2015)", + "38. Mazurowski, M.A., Dong, H., Gu, H., Yang, J., Konz, N., Zhang, Y.: Segment anything model for medical image analysis: an experimental study. Medical Image Analysis 89 (2023)", + "39. Perazzi, F., Krahenbuhl, P., Pritch, Y., Hornung, A.: Saliency filters: Contrast based filtering for salient region detection. In: Conference on Computer Vision and Pattern Recognition (2012)", + "40. Pinheiro, P.O., Collobert, R., Dollár, P.: Learning to segment object candidates. Advances in Neural Information Processing Systems 28 (2015)", + "41. Pinheiro, P.O., Lin, T.Y., Collobert, R., Dollár, P.: Learning to refine object segments. In: European Conference on Computer Vision (2016)", + "42. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning (2021)", + "43. Rahtu, E., Kannala, J., Blaschko, M.: Learning a category independent object detection cascade. In: International Conference on Computer Vision (2011)", + "44. Ronneberger, O., Fischer, P., Brox, T.: U-Net: Convolutional networks for biomedical image segmentation. In: International Conference on Medical Image Computing and Computer-Assisted Intervention (2015)", + "45. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al.: Imagenet large scale visual recognition challenge. International Journal of Computer Vision 115 (2015)", + "46. Saito, K., Hu, P., Darrell, T., Saenko, K.: Learning to detect every thing in an open world. In: European Conference on Computer Vision (2022)", + "47. Shi, Z., Sun, Y., Zhang, M.: Training-free object counting with prompts. In: Winter Conference on Applications of Computer Vision (2024)", + "48. Wang, C., Wang, G., Zhang, Q., Guo, P., Liu, W., Wang, X.: OpenInst: A simple query-based method for open-world instance segmentation. arXiv preprint arXiv:2303.15859 (2023)", + "49. Wang, W., Feiszli, M., Wang, H., Malik, J., Tran, D.: Open-world instance segmentation: Exploiting pseudo ground truth from learned pairwise affinity. In: Conference on Computer Vision and Pattern Recognition (2022)", + "50. Wang, W., Feiszli, M., Wang, H., Tran, D.: Unidentified video objects: A benchmark for dense, open-world segmentation. In: International Conference on Computer Vision (2021)", + "51. Wei, Z., Chen, P., Yu, X., Li, G., Jiao, J., Han, Z.: Semantic-aware SAM for point-prompted instance segmentation. arXiv preprint arXiv:2312.15895 (2023)", + "52. Wilms, C., Frintrop, S.: Edge adaptive seeding for superpixel segmentation. In: German Conference on Pattern Recognition (2017)", + "53. Wilms, C., Frintrop, S.: AttentionMask: Attentive, efficient object proposal generation focusing on small objects. In: Asian Conference on Computer Vision (2019)", + "54. Wu, J., Jiang, Y., Yan, B., Lu, H., Yuan, Z., Luo, P.: Exploring transformers for open-world instance segmentation. In: International Conference on Computer Vision (2023)", + "55. Xie, C., Xiang, Y., Mousavian, A., Fox, D.: Unseen object instance segmentation for robotic environments. IEEE Transactions on Robotics 37(5) (2021)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "SOS: Segment Object System", + "bbox": [ + 531, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "56. Xue, X., Yu, D., Liu, L., Liu, Y., Li, Y., Yuan, Z., Song, P., Shou, M.Z.: Single-stage open-world instance segmentation with cross-task consistency regularization. arXiv preprint arXiv:2208.09023 (2022)", + "57. Yang, X., Gong, X.: Foundation model assisted weakly supervised semantic segmentation. In: Winter Conference on Applications of Computer Vision (2024)", + "58. Yang, Y., Zhou, Z., Wu, J., Wang, Y., Xiong, R.: Class semantics modulation for open-set instance segmentation. IEEE Robotics and Automation Letters (2024)", + "59. Zhao, J.X., Liu, J.J., Fan, D.P., Cao, Y., Yang, J., Cheng, M.M.: EGNet: Edge guidance network for salient object detection. In: International Conference on Computer Vision (2019)", + "60. Zhou, B., Khosla, A., Lapedriza, A., Oliva, A., Torralba, A.: Learning deep features for discriminative localization. In: Conference on Computer Vision and Pattern Recognition (2016)", + "61. Zhou, B., Zhao, H., Puig, X., Xiao, T., Fidler, S., Barriuso, A., Torralba, A.: Semantic understanding of scenes through the ADE20K dataset. International Journal of Computer Vision 127 (2019)", + "62. Zhu, M., Li, H., Chen, H., Fan, C., Mao, W., Jing, C., Liu, Y., Shen, C.: SegPrompt: Boosting open-world segmentation via category-level prompt learning. In: International Conference on Computer Vision (2023)", + "63. Zitnick, C.L., Dollár, P.: Edge boxes: Locating object proposals from edges. In: European Conference on Computer Vision (2014)" + ], + "bbox": [ + 215, + 146, + 784, + 438 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "C. Wilms et al.", + "bbox": [ + 271, + 114, + 375, + 127 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/SOS_ Segment Object System for Open-World Instance Segmentation With Object Priors/466dafa4-76e1-41c2-9ff8-766257e5e796_model.json b/2024/SOS_ Segment Object System for Open-World Instance Segmentation With Object Priors/466dafa4-76e1-41c2-9ff8-766257e5e796_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bf3bedd85e3923a781267d3faa3c982e28ae5dc4 --- /dev/null +++ b/2024/SOS_ Segment Object System for Open-World Instance Segmentation With Object Priors/466dafa4-76e1-41c2-9ff8-766257e5e796_model.json @@ -0,0 +1,2557 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.24, + 0.141, + 0.764, + 0.187 + ], + "angle": 0, + "content": "SOS: Segment Object System for Open-World Instance Segmentation With Object Priors" + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.212, + 0.782, + 0.245 + ], + "angle": 0, + "content": "Christian Wilms1, Tim Rolff1,2, Maris Hillemann1, Robert Johanson1, and Simone Frinthrop1" + }, + { + "type": "text", + "bbox": [ + 0.255, + 0.255, + 0.748, + 0.298 + ], + "angle": 0, + "content": "1 Computer Vision Group, University of Hamburg, Germany \n2 Human-Computer Interaction Group, University of Hamburg, Germany {firstname_lastname}@uni-hamburg.de" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.33, + 0.744, + 0.565 + ], + "angle": 0, + "content": "Abstract. We propose an approach for Open-World Instance Segmentation (OWIS), a task that aims to segment arbitrary unknown objects in images by generalizing from a limited set of annotated object classes during training. Our Segment Object System (SOS) explicitly addresses the generalization ability and the low precision of state-of-the-art systems, which often generate background detections. To this end, we generate high-quality pseudo annotations based on the foundation model SAM [27]. We thoroughly study various object priors to generate prompts for SAM, explicitly focusing the foundation model on objects. The strongest object priors were obtained by self-attention maps from self-supervised Vision Transformers, which we utilize for prompting SAM. Finally, the post-processed segments from SAM are used as pseudo annotations to train a standard instance segmentation system. Our approach shows strong generalization capabilities on COCO, LVIS, and ADE20k datasets and improves on the precision by up to \\(81.6\\%\\) compared to the state-of-the-art. Source code is available at: https://github.com/chwilms/SOS" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.578, + 0.741, + 0.607 + ], + "angle": 0, + "content": "Keywords: Open-world Instance Segmentation \\(\\cdot\\) Object Localization \\(\\cdot\\) Prompting" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.629, + 0.376, + 0.646 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.81 + ], + "angle": 0, + "content": "Open-World Instance Segmentation (OWIS) is the task of segmenting all object instances in an image by learning from a limited set of known object classes [26, 49, 50]. In contrast to instance segmentation, OWIS is not limited by the closed-world assumption, which assumes that all object classes are known in advance. Since OWIS methods aim to detect not only learned but also unknown object classes, they return class-agnostic detections. This is of particular interest in real-world scenarios with previously unknown object classes (e.g., [55]) and challenges the systems' generalization capabilities. An example of this is shown in the first row of Fig. 1, where annotations for classes such as surfboard or tennis racket do not exist during training, but the objects should be detected during testing." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.811, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Previous OWIS methods mostly generate pseudo annotations for unannotated objects [24, 46, 49, 56] or replace the foreground-background classification" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.127 + ], + "angle": 0, + "content": "C. Wilms et al." + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.142, + 0.788, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.321, + 0.788, + 0.435 + ], + "angle": 0, + "content": "Fig. 1: Comparison of (pseudo) annotations (left) used by Mask R-CNN [19], GGN [49], and our SOS, when only original annotations of VOC object classes are given. While Mask R-CNN only uses original annotations without object classes such as tennis racket or surfboard (red arrows), SOS generates pseudo annotations covering those classes (green arrows). GGN generates noisy pseudo annotations including background areas. As a result, only SOS constantly detects these objects not annotated in training (green vs. red arrows on the right). Filled masks denote annotations (left) or detected objects (right), while red frames indicate missed objects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.47, + 0.788, + 0.562 + ], + "angle": 0, + "content": "of possible objects in instance segmentation systems with a learning target focusing on the localization quality (e.g., intersection over union) [26,48,54]. While OWIS methods generalize better to unseen object classes than standard instance segmentation systems, they still exhibit low precision [24,26,46,49,54]. One reason for this is the use of noisy pseudo annotations covering background areas as visible in the left part of the second row in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.565, + 0.788, + 0.716 + ], + "angle": 0, + "content": "Recently, foundation models [4] emerged as a promising technique [5,9,23,29, 42]. They are trained on large datasets with a surrogate task and applied to other tasks in a zero- or few-shot manner using prompting. The first foundation model for image segmentation, the Segment Anything Model (SAM), was proposed by [27]. Trained on a large automatically annotated dataset, SAM generates segments for object or stuff regions based on point, box, mask, or text prompts without further training. Despite not only segmenting objects but also stuff regions, given well-designed prompts, SAM is able to generate high-quality object segments. Note that vanilla SAM does not address object segmentation tasks like OWIS itself due to the aforementioned lacking focus on objects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.842 + ], + "angle": 0, + "content": "In this paper, we propose the Segment Object System (SOS), a novel OWIS method utilizing high-quality pseudo annotations to improve the generalization capability of a standard instance segmentation system (see lower row in Fig. 1). For generating pseudo annotations, we apply the foundation model SAM with prompts derived from an object prior to roughly localize objects of arbitrary classes and to limit segmentations of stuff regions by SAM. To ensure the high quality of our pseudo annotations based on SAM, improving the overall precision, we thoroughly study various hand-crafted and learned object priors for an" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.532, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "SOS: Segment Object System" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.312 + ], + "angle": 0, + "content": "object-focused application of SAM. This is relevant beyond the scope of OWIS. Using our study findings, SOS utilizes self-attention maps from self-supervised Vision Transformers (ViTs) [11] as object prior for prompting SAM. Finally, SOS filters low-quality pseudo annotations, combines the pseudo annotations with the original annotations of the known classes, and trains a standard instance segmentation system with these mixed annotations. Our extensive evaluation shows the strong generalization abilities of SOS, considerably outperforming all previous state-of-the-art systems across COCO [31], LVIS [17], and ADE20k [61] datasets. Most notably, the results improve by up to \\(81.6\\%\\) in terms of precision over the state-of-the-art due to the high-quality pseudo annotations. We also show that SOS is better suited for OWIS than directly applying SAM." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.314, + 0.53, + 0.327 + ], + "angle": 0, + "content": "Overall, our contributions are threefold:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.342, + 0.784, + 0.385 + ], + "angle": 0, + "content": "- We propose SOS, a novel OWIS method based on a learned object prior, prompt-based pseudo annotations, and an arbitrary instance segmentation system." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.387, + 0.785, + 0.417 + ], + "angle": 0, + "content": "- We thoroughly study various object priors for focusing SAM on objects, leading to high-quality object segments for pseudo annotations and beyond." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.419, + 0.785, + 0.461 + ], + "angle": 0, + "content": "- Our extensive evaluation shows that the high-quality pseudo annotations in SOS lead to high precision and strong overall results, clearly outperforming other methods." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.342, + 0.785, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.489, + 0.388, + 0.505 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.524, + 0.551, + 0.54 + ], + "angle": 0, + "content": "2.1 Open-world Instance Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.552, + 0.784, + 0.643 + ], + "angle": 0, + "content": "To address OWIS, literature mainly offers two streams. First, systems replace the hard classification in instance segmentation systems with a localization score. This avoids classifying unseen objects as background and improves generalization ability [26]. In OLN [26], centerness and Intersection over Union (IoU) are learned as localization scores in a Mask R-CNN system [19]. SWORD [54] and OpenInst [48] follow the same idea for query-based systems." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.785, + 0.749 + ], + "angle": 0, + "content": "The second line of works [24, 46, 49, 56] addresses OWIS by augmenting the annotations from known classes with pseudo annotations aiming to cover objects of unknown classes. To generate pseudo annotations, GGN [49] learns pixel affinities from known classes to generate segments, which are grouped. Similarly, UDOS [24] groups object proposals of object parts to create pseudo annotations, while LDET [46] applies copy-paste-augmentation. Recently, a new stream emerged [58, 62] explicitly utilizing class labels from known classes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.785, + 0.84 + ], + "angle": 0, + "content": "We follow the second stream and create pseudo annotations within SOS, since it offers the most flexibility w.r.t. the base instance segmentation system. Different from existing approaches that generate noisy pseudo annotations (see GGN in Fig. 1), we investigate and develop object priors to effectively use recent foundation models to create high-quality object-focused pseudo annotations, leading to high recall and precision." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.377, + 0.129 + ], + "angle": 0, + "content": "C. Wilms et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.546, + 0.163 + ], + "angle": 0, + "content": "2.2 Class-agnostic Object Localization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.173, + 0.787, + 0.248 + ], + "angle": 0, + "content": "Localizing objects in images is related to several computer vision tasks. For instance, in traditional object proposal generation [2], cues like saliency [2, 37] or contour information [2, 10, 36, 43] were used to localize class-agnostic object candidates for object detection. Since the advent of deep learning, the models are learned from data [16, 22, 40, 41, 53]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.25, + 0.787, + 0.385 + ], + "angle": 0, + "content": "Focusing on arbitrary salient objects, known as salient object detection, several approaches use the contrast of hand-crafted features such as color [8,28,34, 39] or edge information [28], as well as CNN-based approaches [21, 33, 59]. In a different task, several authors investigated the information stored in learned classifiers or self-supervised feature extractors based on CNNs or ViTs to locate objects. Notable avenues include Class Activation Maps (CAMs) [60] that highlight discriminative object or image parts in CNN-based classifiers and the self-attention maps in the self-supervised DINO feature extractor [6] based on ViTs, which contain information on the scene layout indicating object locations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.385, + 0.787, + 0.432 + ], + "angle": 0, + "content": "In our study on object priors to focus SAM on objects, we investigate several of these object localization cues. In contrast to these approaches, we follow a different goal, focusing prompt-based segmentations on objects for OWIS." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.454, + 0.436, + 0.469 + ], + "angle": 0, + "content": "2.3 Applications of SAM" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.48, + 0.787, + 0.571 + ], + "angle": 0, + "content": "SAM [27] is used in several segmentation tasks [3,38,47]. Similar to us, [7] and [57] use SAM to generate pseudo annotation masks for weakly-supervised semantic segmentation based on image-level supervision. Similarly, [18] generate pseudo annotation masks for concealed object segmentation based on scribble annotations. Finally, [51] generate pseudo annotation masks for weakly-supervised instance segmentation in a multiple instance learning framework." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.571, + 0.787, + 0.647 + ], + "angle": 0, + "content": "Different from the weakly-supervised approaches utilizing SAM, we do not rely on supervision. Hence, a key novelty of our work lies in investigating object priors to focus SAM on segmenting arbitrary objects in the absence of supervision for all object classes. Moreover, we are the first to apply SAM-based pseudo annotations in OWIS." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.671, + 0.586, + 0.689 + ], + "angle": 0, + "content": "3 Segment Object System for OWIS" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.787, + 0.842 + ], + "angle": 0, + "content": "To address OWIS, we propose our Segment Object System (SOS) that generates high-quality pseudo annotations to train a standard instance segmentation system. As visible from Fig. 2, SOS consists of three main blocks. The first block, the Object Localization Module (OLM), described in Sec. 3.2 (yellow area in Fig. 2), aims to roughly localize objects by sampling locations from an object prior that reflects the object probability per image coordinate. The output of OLM is a set of likely object locations, that is used in the Pseudo Annotation Creator (PAC), described in Sec. 3.3 (green area in Fig. 2), to prompt the pre-trained SAM [27], which we briefly revisit in Sec. 3.1. From the segments" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.533, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SOS: Segment Object System" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.143, + 0.789, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.44, + 0.788, + 0.538 + ], + "angle": 0, + "content": "Fig. 2: Overview of our Segment Object System (SOS) for OWIS consisting of three blocks. First, the input image is processed in our Object Localization Module (OLM, yellow area) to create object-focused point prompts roughly localizing objects. Second, our Pseudo Annotations Creator (PAC, green area) generates segments based on the previously generated prompts using SAM [27], and further processes them, leading to a final set of merged original and pseudo annotations. Finally, the merged annotations are used to train an instance segmentation system (blue area)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.567, + 0.788, + 0.645 + ], + "angle": 0, + "content": "generated from likely object locations, the PAC generates pseudo annotations by filtering low-quality segments and removing near-duplicates. Finally, the pseudo annotations are merged with the original annotations, and a standard instance segmentation system is trained with the merged annotations (see Sec. 3.4, blue area in Fig. 2). During testing, only this instance segmentation system is used." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.664, + 0.54, + 0.681 + ], + "angle": 0, + "content": "3.1 Revisit Segment Anything Model" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Since SAM [27] is integral to SOS, we will briefly review the model. The idea of SAM is to utilize prompts such as point coordinates to generate segments of image parts like objects, object parts, or stuff regions localized by these prompts. Hence, SAM does not exclusively segment objects unless explicitly instructed by object-focused prompts, as proposed in this paper. SAM is based on a ViT [11] image encoder creating an image embedding, while the embeddings for point prompts are generated based on positional encodings. Both embeddings are processed in a mask decoder utilizing a transformer decoder block and a mask prediction head to create the final segment including a confidence score. Note that SAM is class-agnostic and does not classify the segments." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.377, + 0.129 + ], + "angle": 0, + "content": "C. Wilms et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.27 + ], + "angle": 0, + "content": "For training SAM, [27] propose the new large-scale SA-1B dataset that is annotated with a data engine using SAM itself. In the pre-stage, SAM is trained with public segmentation datasets to aid the first stage of the data engine, where annotators use SAM as an interactive segmentation tool to create initial segmentations. Subsequently, SAM is retrained with the newly annotated data, fully avoiding data leakage w.r.t. public segmentation datasets. In subsequent stages, manual annotations were provided for segments missed by SAM. At last, annotations were created fully automatically by SAM. See [27] for more details." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.289, + 0.494, + 0.305 + ], + "angle": 0, + "content": "3.2 Object Localization Module" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.312, + 0.789, + 0.524 + ], + "angle": 0, + "content": "The first block of SOS (yellow area in Fig. 2) is the Object Localization Module (OLM). OLM roughly localizes all objects in the input image, and creates object-focused point prompts. As the first step, OLM generates an object prior from the input image. The object prior is a probability mass function constrained to the image plane, although other formulations are possible (see Sec. 4), and indicates the probability of an image coordinate being part of any object. Hence, the object prior is class-agnostic and not limited to known object classes from the training set in OWIS. In our study in Sec. 4, we analyze various object priors. For final SOS, we utilize the six self-attention maps from the self-attention heads of a ViT's final layer, pre-trained in the self-supervised DINO framework [6] on ImageNet [45]. See Sec. 4 for justifications. OLM aggregates the six self-attention maps by elementwise max yielding one map highlighting all relevant scene elements. Subsequently, OLM converts the map to an object prior, by rescaling such that the minimum equals 0 and the sum over the object prior is 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.524, + 0.79, + 0.678 + ], + "angle": 0, + "content": "Next, OLM randomly samples \\( S \\) (here: \\( S = 50 \\)) image coordinates from the object prior, given the per-coordinate values of the object prior as probabilities. To diversify the samples across multiple objects, we prune parts of the object prior around the sampled coordinate. Specifically, given a coordinate \\( (x_s, y_s) \\), we set the object prior to zero for all coordinates \\( \\{(x, y) \\mid |x - x_s| \\leq N \\land |y - y_s| \\leq N\\} \\), with \\( N = 20 \\). To transform the object prior back to a probability mass function, we again apply the rescaling as described above. Subsequently, OLM iteratively applies this sampling until a set of \\( S \\) coordinates are extracted, representing object-focused point prompts. This set is the output of the OLM and will be used for prompting in the Pseudo Annotation Creator." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.696, + 0.502, + 0.71 + ], + "angle": 0, + "content": "3.3 Pseudo Annotations Creator" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.841 + ], + "angle": 0, + "content": "The Pseudo Annotations Creator (PAC, green area in Fig. 2) is the second block of SOS. It generates pseudo annotations utilizing SAM, given the object-focused point prompts provided by OLM. These object-focused point prompts lifts SAM from a system that segments anything to a system that segments objects. As the first step, PAC prompts a pre-trained SAM with the \\(S\\) point prompts from OLM. To handle ambiguous point prompts that may indicate multiple objects or object parts, we follow [27] and allow SAM to generate three segments per prompt. Each generated segment is a potential pseudo annotation. The resulting set of" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.532, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "SOS: Segment Object System" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.223 + ], + "angle": 0, + "content": "class-agnostic segments is noisy, with segments covering objects and background, or covering the same image area. Therefore, PAC first utilizes SAM's confidence score per segment and removes segments with a confidence below \\(\\tau_{\\mathrm{conf}}\\). Second, PAC removes near-duplicates by non-maximum suppression with an IoU threshold \\(\\tau_{\\mathrm{NMS}}\\). We set \\(\\tau_{\\mathrm{conf}} = 0.9\\) and \\(\\tau_{\\mathrm{NMS}} = 0.95\\) for SOS." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.79, + 0.404 + ], + "angle": 0, + "content": "Given this set of pseudo annotations, PAC merges the set with the original annotations representing known classes of the training dataset to combine the knowledge of the human annotators and the knowledge of our object-focused SAM. Since we only want to keep pseudo annotations for objects not covered by the original annotations, we suppress pseudo annotations that have a high IoU (\\(\\tau_{\\mathrm{NMS}}\\)) with at least one original annotation. Moreover, we limit the number of pseudo annotations to \\(P\\) in order to balance original and pseudo annotations. In our final SOS, we set \\(P = 10\\). However, across the COCO training dataset, only 7.8 pseudo annotations are added on average. The other pseudo annotations were suppressed in previous steps. Overall, the output of PAC is a mixed set of annotations, containing all original annotations augmented with high-quality pseudo annotations covering objects of unknown classes." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.424, + 0.452, + 0.438 + ], + "angle": 0, + "content": "3.4 Instance Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.446, + 0.788, + 0.537 + ], + "angle": 0, + "content": "In the last block of SOS, we take the merged annotations generated by PAC and train a standard instance segmentation system. While this system can be arbitrary and is generally replaceable, we choose Mask R-CNN [19] with a ResNet-50+FPN backbone [30] trained in a class-agnostic setting, following recent OWIS methods [24, 26, 46, 49]. Overall, this leads to an OWIS method based on high-quality pseudo annotations." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.557, + 0.469, + 0.575 + ], + "angle": 0, + "content": "4 Object Priors for SOS" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.587, + 0.788, + 0.664 + ], + "angle": 0, + "content": "As described in Sec. 3.2, SOS uses an object prior for rough object localization and subsequent point prompt generation for PAC. This section thoroughly investigates the performance of various object priors derived from previous class-agnostic object localization works in SOS. For details on the object priors beyond the subsequent descriptions, we refer to our supplementary." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.682, + 0.377, + 0.697 + ], + "angle": 0, + "content": "4.1 Object Priors" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Baselines: We introduce three baseline object priors that ignore the image content. First, Grid uses SAM with a regular \\(64 \\times 64\\) grid of points, following [27]. This directly leads to prompts without the sampling described in Sec. 3.2. Note that the Grid baseline will lead to segments for both object and stuff regions. Second, we use the spatial distribution of object centroids from the known classes across the training dataset as object prior (Dist). We assume that this generalizes well since it ignores class-specific object features. Finally, we utilize the centroids of the training set's objects from unknown classes as optimal prompts (\\(GT\\)), representing an upper bound." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.377, + 0.129 + ], + "angle": 0, + "content": "C. Wilms et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.254 + ], + "angle": 0, + "content": "Superpixels: Inspired by [15, 52], we utilize the centroids of superpixels generated with an adaptive superpixel segmentation method as object prior (\\( Spx \\)). The intuition is that large uniform areas not containing objects should be covered by few superpixels and vice versa. Hence, the density of superpixels is a surrogate for the density of objects. We use the well-known FH superpixels [13] since they adapt their density to the image content following our assumption. Note that every superpixel centroid is a point prompt." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.265, + 0.79, + 0.327 + ], + "angle": 0, + "content": "Contour Density: Since objects are defined by their outer contours [63], we use contour density as an object prior (Contour). Similar to [52], we assume that in areas of high contour density, several objects are located and vice versa. As a surrogate for the contour density, we use edge density based on [10]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.337, + 0.788, + 0.415 + ], + "angle": 0, + "content": "Saliency: Saliency is used in object proposal generation to localize objects [2, 37] as they stick out from their surrounding. Here, we evaluate two saliency methods as object priors. First, we apply the traditional approach VOCUS2 [14] based on color contrast and used by [37] (VOCUS2). Second, we use pre-trained DeepGaze IIE [32] saliency maps learned from eye fixation data (DeepGaze)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.426, + 0.788, + 0.503 + ], + "angle": 0, + "content": "Class Activation Maps: CAMs [60] indicate discriminative image regions for CNN-based classifiers. Since they only need image-level supervision, they are frequently applied in weakly-supervised tasks to locate objects [1, 25, 35]. We use the CAM of the predicted class in a ResNet-50 classifier [20] pre-trained on ImageNet [45] as object prior (CAM)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.513, + 0.788, + 0.575 + ], + "angle": 0, + "content": "Self-attention: Recently, [6] have shown that self-attention maps learned inside self-supervised ViTs encode the scene layout, including object locations. Therefore, we explore the self-attention maps from the final layer of a ViT-S backbone trained in the self-supervised DINO framework [6] as an object prior (DINO)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.587, + 0.788, + 0.649 + ], + "angle": 0, + "content": "Learned Object Locations: Finally, we learn object locations based on the known classes of a dataset using a U-Net [44] in a binary segmentation task (object vs. no object) as object prior (\\(U\\)-Net). This assumes that, on the level of individual pixels, the model generalizes from known to unknown classes [49,56]." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.667, + 0.513, + 0.683 + ], + "angle": 0, + "content": "4.2 Study Setup for Object Priors" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.842 + ], + "angle": 0, + "content": "To assess each object prior, we train SOS with the respective prior on the COCO training set using only annotations from the 20 PASCAL VOC [12] classes. We evaluate on COCO's validation set with annotations from the remaining 60 classes in COCO, following standard COCO (VOC) \\(\\rightarrow\\) COCO (non-VOC) cross-category evaluation in OWIS [26, 46, 49]. Note that we use the default class-agnostic Mask R-CNN in SOS but reduce the training schedule to only a quarter of steps for faster training. We evaluate the results of SOS given each object priors and report Average Recall (AR) for 100 detections and Average Precision (AP) as previous OWIS research [46, 49]. We also report \\(F_{1}\\) score, the harmonic mean between AR and AP, yielding a single number for comparison." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.532, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SOS: Segment Object System" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.147, + 0.443, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.297, + 0.261, + 0.381, + 0.275 + ], + "angle": 0, + "content": "Input image" + }, + { + "type": "image", + "bbox": [ + 0.446, + 0.144, + 0.553, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.446, + 0.202, + 0.553, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.469, + 0.261, + 0.532, + 0.274 + ], + "angle": 0, + "content": "VOCUS2" + }, + { + "type": "image", + "bbox": [ + 0.555, + 0.144, + 0.663, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.555, + 0.202, + 0.663, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.589, + 0.261, + 0.631, + 0.274 + ], + "angle": 0, + "content": "U-Net" + }, + { + "type": "image", + "bbox": [ + 0.665, + 0.144, + 0.771, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.664, + 0.202, + 0.771, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.696, + 0.261, + 0.738, + 0.274 + ], + "angle": 0, + "content": "DINO" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.287, + 0.785, + 0.302 + ], + "angle": 0, + "content": "Fig. 3: Object priors VOCUS2, U-Net, and DINO with resulting pseudo annotations." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.332, + 0.527, + 0.347 + ], + "angle": 0, + "content": "4.3 Results of Object Priors in SOS" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.359, + 0.788, + 0.51 + ], + "angle": 0, + "content": "Table 1 shows the results of our object prior study. First, all priors outperform the baselines Grid and Dist in terms of \\(\\mathbf{F}_1\\). Both baselines exhibit a lower AP compared to all other object priors, reflecting the missing focus on objects. Hence, simply applying SAM (Grid) is not suitable to segment only objects, as several stuff regions are segmented as well. Second, most priors exhibit a similar performance in terms of all measures. For instance, learning-based CAM and DeepGaze do not outperform simple priors (Spx, Contour, VOCUS2). Only the priors U-Net and DINO perform substantially better, with DINO producing the best result. Comparing DINO to GT reveals that DINO is able to recall almost the same amount of objects, but exhibits a lower precision, as expected." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.51, + 0.788, + 0.661 + ], + "angle": 0, + "content": "To illustrate the results, Fig. 3 depicts object priors for top-performing VOCUS2, \\( U \\)-Net, and DINO, as heatmaps (upper row) with the resulting pseudo annotations generated in SOS (lower row) for an example image. The heatmaps show that DINO is very focused on the discriminative object parts, here the polar bears' faces, while \\( U \\)-Net highlights, as learned, the entire objects. Moreover, \\( U \\)-Net also highlights several background areas. Similarly, VOCUS2 highlights background areas and the polar bears since it is generally focused on high-contrast areas. Overall, only DINO exclusively generates pseudo annotations for the polar bears and some sub-parts, while \\( U \\)-Net and VOCUS2 produce noisy annotations including background patches, similar to previous approaches [49]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.661, + 0.788, + 0.737 + ], + "angle": 0, + "content": "The results show that a well-designed object prior like DINO substantially outperforms the baseline Grid \\((+7.5\\) in \\(\\mathrm{F}_1)\\) in SOS. Therefore, we use DINO as the object prior in SOS. More generally, the study also reveals important insights on how to prompt SAM for class-agnostic, object-focused results for arbitrary tasks. For more study results, we refer to our supplementary." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.761, + 0.357, + 0.777 + ], + "angle": 0, + "content": "5 Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.794, + 0.793, + 0.842 + ], + "angle": 0, + "content": "In our evaluation, we compare SOS to recent OWIS methods OLN [26], LDET [46], GGN [49], SWORD [54], SOIS [56], OpenInst [48], UDOS [24], and against Mask R-CNN [19] and SAM [27] baselines, depending on the availability of public" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.377, + 0.129 + ], + "angle": 0, + "content": "C. Wilms et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.175 + ], + "angle": 0, + "content": "Table 1: Results of our SOS with various object priors in the cross-category COCO (VOC) \\(\\rightarrow\\) COCO (non-VOC) setting. *: Uses ground truth of unknown classes." + }, + { + "type": "table", + "bbox": [ + 0.264, + 0.185, + 0.74, + 0.286 + ], + "angle": 0, + "content": "
Object PriorAPAR100F1Object PriorAPAR100F1
SOS + Grid3.836.56.9SOS + DeepGaze5.435.99.4
SOS + Dist3.427.46.0SOS + CAM5.436.79.4
SOS + Spx5.634.89.6SOS + DINO8.938.114.4
SOS + Contour5.636.69.7SOS + U-Net7.337.312.2
SOS + VOCUS25.137.710.5SOS + GT18.1*42.5*25.4*
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.318, + 0.788, + 0.485 + ], + "angle": 0, + "content": "source code or reported results. To assess the quality of the approaches, we use four cross-category and cross-dataset settings, commonly used in OWIS. Specifically, we choose COCO (VOC) \\(\\rightarrow\\) COCO (non-VOC) on the COCO dataset [31], already discussed in Sec. 4.2. For cross-dataset evaluation, we use COCO \\(\\rightarrow\\) LVIS, COCO \\(\\rightarrow\\) ADE20k, and COCO \\(\\rightarrow\\) UVO that train on the entire COCO training set and evaluate on the validation sets of LVIS [17], ADE20k [61], and UVO [50] with exhaustive annotations. The LVIS dataset covers the same images as COCO but extends the annotated object classes from 80 to 1203. ADE20k includes 3169 classes of objects, object parts, and stuff regions. Finally, UVO features exhaustive annotations of all objects and, therefore, no system is penalized for detecting objects outside a datasets' taxonomy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.486, + 0.788, + 0.547 + ], + "angle": 0, + "content": "As in Sec. 4.2 and following common practice in OWIS [26,46,49], we report AR for up to 100 detections and AP. We again also report the \\( \\mathrm{F}_1 \\) score between AR and AP as a single, combined quantity, described in Sec. 4.2. Note that we only evaluate based on masks, not boxes." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.578, + 0.595, + 0.594 + ], + "angle": 0, + "content": "5.1 Implementation Details and Data Usage" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.612, + 0.788, + 0.718 + ], + "angle": 0, + "content": "As discussed in Sec. 4.3, SOS uses the object prior DINO for best results. DINO [6] with a ViT-S backbone is used as an ImageNet [45] pre-trained model without fine-tuning. Similarly, SOS uses SAM [27] with the ViT-H backbone as a pre-trained model based on SA-1B [27] without fine-tuning. As an instance segmentation model, we apply Mask R-CNN [19] with an ImageNet pre-trained ResNet-50+FPN backbone [30], and the default configurations except for the class-agnostic training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Overall, SOS is based on models pre-trained on standard ImageNet and SA-1B. Pre-training on SA-1B or similar large-scale datasets becomes common with the emergence of foundation models and is widely adopted by systems using SAM in similar contexts [7,51,57]. It is comparable to using models pre-trained on ImageNet that even have class information available during training and also allow a rough localization of objects [6,60]. Moreover, the final, noisy annotations of SA-1B representing objects, object parts, and stuff region were automatically generated by SAM [27], and no data leakage w.r.t. other datasets exists." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.532, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "SOS: Segment Object System" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.493, + 0.214 + ], + "angle": 0, + "content": "Table 2: Results of two baselines and various OWIS methods in the COCO (VOC) \\(\\rightarrow\\) COCO (non-VOC) setting. \\(\\dagger\\): uses automatically annotated SA-1B dataset." + }, + { + "type": "table", + "bbox": [ + 0.238, + 0.226, + 0.472, + 0.382 + ], + "angle": 0, + "content": "
SystemAPAR100F1
Mask R-CNN [19]1.08.21.8
SAM†[27]3.648.16.7
OLN [26]4.228.47.3
LDET [46]4.324.87.3
GGN [49]4.928.38.4
SWORD [54]4.830.28.3
UDOS [24]2.934.35.3
SOS†(ours)8.939.314.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.509, + 0.145, + 0.788, + 0.2 + ], + "angle": 0, + "content": "Table 3: Results of two baselines and various OWIS methods in the COCO \\(\\rightarrow\\) LVIS setting. \\(^\\dagger\\): uses automatically annotated SA-1B dataset." + }, + { + "type": "table", + "bbox": [ + 0.532, + 0.213, + 0.766, + 0.368 + ], + "angle": 0, + "content": "
SystemAPAR100F1
Mask R-CNN [19]7.523.611.4
SAM†[27]6.845.111.8
LDET [46]6.724.810.5
GGN [49]6.527.010.5
SOIS [56]-25.2-
OpenInst [48]-29.3-
UDOS [24]3.924.96.7
SOS†(ours)8.133.313.3
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.411, + 0.757, + 0.427 + ], + "angle": 0, + "content": "5.2 Cross-category COCO (VOC) \\(\\rightarrow\\) COCO (non-VOC) Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.439, + 0.788, + 0.591 + ], + "angle": 0, + "content": "Table 2 presents the results of various OWIS methods on the cross-category setup COCO (VOC) \\(\\rightarrow\\) COCO (non-VOC). Our SOS clearly outperforms all OWIS methods in \\(\\mathrm{AR}_{100}\\), AP, and \\(\\mathbf{F}_1\\). Specifically, SOS outperforms the second-best system in terms of \\(\\mathrm{F}_1\\), GGN [49], by 6.1. While some of the improvement comes from a better recall (\\(+38.9\\%\\) compared to GGN), the results are driven by a much-improved precision (\\(+81.6\\%\\) compared to GGN). This reflects the high quality of the pseudo annotations in SOS on this cross-category setting. Moreover, there is a clear improvement of SOS over Mask R-CNN (\\(+12.7\\) in \\(\\mathrm{F}_1\\)). Compared to original SAM, the DINO-based SOS substantially improves the precision by focusing on objects, leading to an improved \\(\\mathrm{F}_1\\) score (\\(+7.8\\))." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.591, + 0.788, + 0.682 + ], + "angle": 0, + "content": "The qualitative results of various OWIS methods in Fig. 4 support the quantitative results. In the first example, only GGN [49] and SOS detect both giraffes, however, SOS generates more accurate segmentations. In the second example, only SOS detects all five small surfboards, while other systems detect at most three. Note that class-agnostic Mask R-CNN misses all non-VOC objects in both examples. For more qualitative results, see our supplementary." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.707, + 0.442, + 0.721 + ], + "angle": 0, + "content": "5.3 Cross-dataset Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "\\( COCO \\rightarrow LVIS \\): We evaluate the cross-dataset generalization capabilities of OWIS methods, starting with \\( COCO \\rightarrow LVIS \\). The results in Tab. 3 show a trend similar to the cross-category results. SOS outperforms all OWIS methods across all measures with an improvement of 2.8 in \\( F_1 \\) compared to second-best methods GGN [49] and LDET [46]. Similar to the previous results, the improvement is based on gains in both recall and precision, however, the relative improvements in AP and \\( AR_{100} \\) are more similar to each other. Overall, our high-quality pseudo" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.377, + 0.128 + ], + "angle": 0, + "content": "C. Wilms et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.143, + 0.79, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.301, + 0.788, + 0.345 + ], + "angle": 0, + "content": "Fig. 4: Qualitative results of OWIS methods and baseline Mask R-CNN in the cross-category COCO (VOC) \\(\\rightarrow\\) COCO (non-VOC) setting. Filled masks denote detected objects, while red frames indicate missed objects." + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.365, + 0.495, + 0.422 + ], + "angle": 0, + "content": "Table 4: Results of Mask R-CNN baseline and various OWIS methods in the COCO \\(\\rightarrow\\) ADE20k setting. \\(^\\dagger\\): uses automatically annotated SA-1B dataset." + }, + { + "type": "table", + "bbox": [ + 0.234, + 0.433, + 0.477, + 0.564 + ], + "angle": 0, + "content": "
SystemAPAR100F1
Mask R-CNN [19]6.911.98.7
OLN [26]-20.4-
LDET [46]9.518.512.6
GGN [49]9.721.013.3
UDOS [24]7.622.911.4
SOS† (ours)12.526.517.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.509, + 0.359, + 0.788, + 0.414 + ], + "angle": 0, + "content": "Table 5: Results of two baselines and various OWIS methods in the COCO \\(\\rightarrow\\) UVO setting. \\(^\\dagger\\): uses automatically annotated SA-1B dataset." + }, + { + "type": "table", + "bbox": [ + 0.527, + 0.427, + 0.772, + 0.569 + ], + "angle": 0, + "content": "
SystemAPAR100F1
Mask R-CNN [19]20.736.726.5
SAM†[27]11.350.118.4
OLN [26]-41.4-
LDET [46]22.040.428.5
GGN [49]20.343.427.7
UDOS [24]10.643.117.0
SOS†(ours)20.942.328.0
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.597, + 0.788, + 0.629 + ], + "angle": 0, + "content": "annotations in SOS lead to new state-of-the-art results, which indicate strong generalization to unknown object classes outside COCO." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.788, + 0.751 + ], + "angle": 0, + "content": "\\(COCO \\rightarrow ADE20k\\): The results for \\(COCO \\rightarrow ADE20k\\) in Tab. 4, which include labeled object parts of \\(ADE20k\\), show that SOS again outperforms all other OWIS methods across all measures. The gain over the second-best method in \\(F_1\\), GGN [49], is 3.7. Similar to \\(COCO \\rightarrow LVIS\\), the relative gains in \\(AR_{100}\\) and AP are equally distributed. These results indicate that object parts do not degrade the results of SOS. This is in line with SOS's pseudo annotations in Fig. 3, covering entire objects and selected object parts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.842 + ], + "angle": 0, + "content": "\\( COCO \\rightarrow UVO \\): Finally, Tab. 5 presents results for COCO \\( \\rightarrow \\) UVO. SOS still outperforms Mask R-CNN, SAM and recent UDOS [24] in \\( F_1 \\). Outperforming SAM implies that SOS does not exploit the limited taxonomy of annotations in COCO (only 60 object classes in test), but also improves the segmentation of objects outside the COCO object classes on the exhaustively labeled UVO" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.532, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SOS: Segment Object System" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.14, + 0.493, + 0.209 + ], + "angle": 0, + "content": "Table 6: Quality of pseudo annotation generated in GGN [49] and our SOS evaluated on the non-VOC annotations of the COCO training dataset. \\(\\dagger\\): uses automatically annotated SA-1B dataset." + }, + { + "type": "table", + "bbox": [ + 0.255, + 0.223, + 0.455, + 0.303 + ], + "angle": 0, + "content": "
SystemPrec.Rec.F1
GGN3 [49]7.312.19.1
SOS3† (ours)19.026.422.1
SOS10† (ours)15.541.722.6
" + }, + { + "type": "image", + "bbox": [ + 0.537, + 0.144, + 0.648, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.651, + 0.144, + 0.761, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.537, + 0.204, + 0.648, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.576, + 0.267, + 0.608, + 0.277 + ], + "angle": 0, + "content": "GGN" + }, + { + "type": "image", + "bbox": [ + 0.651, + 0.204, + 0.761, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.692, + 0.267, + 0.719, + 0.277 + ], + "angle": 0, + "content": "SOS" + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.279, + 0.786, + 0.306 + ], + "angle": 0, + "content": "Fig. 5: Pseudo annotations generated by GGN and our SOS." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.35, + 0.788, + 0.426 + ], + "angle": 0, + "content": "dataset. However, the results of SOS are below LDET [46] in AP and \\(\\mathrm{F_1}\\). We attribute this to several classes in UVO not available in ImageNet, which was used to learn DINO object prior. Hence, DINO will miss such objects in the COCO dataset during training, and SOS might miss them during testing. This could be mitigated using more diverse unlabeled images for training DINO." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.449, + 0.518, + 0.464 + ], + "angle": 0, + "content": "5.4 Quality of Pseudo Annotations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.474, + 0.788, + 0.64 + ], + "angle": 0, + "content": "We also investigate the quality of the pseudo annotations generated in SOS. To this end, we evaluate the pseudo annotations on the COCO training set against the annotations of non-VOC object classes, reflecting the COCO (VOC) \\(\\rightarrow\\) COCO non-VOC) setting. The results in terms of precision and recall for IoU \\(= 0.5\\) as well as the \\(\\mathrm{F_1}\\) score are presented in Tab. 6. We use up to three and 10 pseudo annotations generated per image in SOS (SOS\\(_3\\) and \\(\\mathrm{SOS}_{10}\\)), and \\(\\mathrm{GGN}_3\\) [49] that generates three pseudo annotations. The results show that with only three SOS pseudo annotations, more than \\(25\\%\\) of the non-VOC objects are covered. This is substantially more than \\(\\mathrm{GGN}_3\\) (\\(12.1\\%\\)). With up to 10 annotations, SOS's pseudo annotations cover more than \\(40\\%\\) of the non-VOC objects. In our supplementary, we further provide class-specific results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.64, + 0.788, + 0.716 + ], + "angle": 0, + "content": "To complement these results, we visualize the annotations of \\(\\mathrm{GGN}_3\\) and \\(\\mathrm{SOS}_3\\) on two sample images in Fig. 5. It is visible that \\(\\mathrm{GGN}_3\\)'s annotations cover background regions, leading to low precision in the overall results. Moreover, the pseudo annotations in the upper example do not adhere well to the airplane's boundaries. More qualitative examples are given in Fig. 1 and our supplementary." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.739, + 0.403, + 0.753 + ], + "angle": 0, + "content": "5.5 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.786, + 0.795 + ], + "angle": 0, + "content": "All ablation studies follow the evaluation setup described in Sec. 4.2. More studies evaluating design choices in SOS are presented in our supplementary." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Influence of SOS's Components: We investigate the importance of each component in SOS and present the results in Tab. 7. The baseline for this" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.377, + 0.128 + ], + "angle": 0, + "content": "C. Wilms et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.493, + 0.188 + ], + "angle": 0, + "content": "Table 7: Influence of added components in our SOS evaluated in the COCO (VOC) \\(\\rightarrow\\) COCO (non-VOC) setting." + }, + { + "type": "table", + "bbox": [ + 0.239, + 0.199, + 0.472, + 0.286 + ], + "angle": 0, + "content": "
ComponentsAPAR100F1
Mask R-CNN1.210.82.2
+ Grid, w/o pp.3.435.26.2
+ DINO, w/o pp.8.937.114.3
+ DINO, w/ pp.8.938.114.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.51, + 0.145, + 0.788, + 0.188 + ], + "angle": 0, + "content": "Table 8: SOS results with various numbers of pseudo annotations in the COCO (VOC) \\(\\rightarrow\\) COCO (non-VOC) setting." + }, + { + "type": "table", + "bbox": [ + 0.541, + 0.199, + 0.758, + 0.286 + ], + "angle": 0, + "content": "
#Pseudo Anns.APAR100F1
38.334.513.4
58.836.214.2
108.938.114.4
208.838.114.3
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.311, + 0.789, + 0.465 + ], + "angle": 0, + "content": "study is a class-agnostic Mask R-CNN trained without pseudo annotations (first row in Tab. 7). Subsequently, we add pseudo annotations based on SAM with the baseline Grid object prior as in [27] and without our post-processing in PAC (confidence-based thresholding and NMS, see Sec. 3.3). Next, we replace Grid with DINO (third row in Tab. 7), and finally add the post-processing leading to the complete SOS. The results in Tab. 7 show each step improves the \\( \\mathrm{F}_1 \\) results. Mainly, introducing SAM for generating pseudo annotations (first row vs. second row) substantially improves \\( \\mathrm{AR}_{100} \\), while adding the DINO object prior removes background annotations leading to considerably improved precision. Hence, the object prior selection in SOS is crucial for strong results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.474, + 0.79, + 0.596 + ], + "angle": 0, + "content": "Number of Pseudo Annotations: In the second ablation study, we investigate the influence of the number of pseudo annotations per image on the results of SOS. To this end, we evaluate SOS with up to 3, 5, 10, and 20 pseudo annotations per image. The results in Tab. 8 show that 10 pseudo annotations are preferable. Hence, the ideal number of pseudo annotations in SOS is larger than in GGN [49], which only uses three. This indicates again the high quality of our pseudo annotations, since more than three annotations per image can cover relevant objects without adding too many noisy annotations." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.615, + 0.362, + 0.631 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.789, + 0.842 + ], + "angle": 0, + "content": "In this paper, we addressed the challenging OWIS task and aimed to improve the low precision of previous methods in this open-world task. To improve the precision and the overall detection results, we generate high-quality pseudo annotations based on a prompt-guided foundation model, SAM. To focus the pseudo annotations on objects, we thoroughly investigated various object priors for prompt creation, which revealed important insights on how to prompt SAM for class-agnostic object-focused results in arbitrary tasks. As a result, our novel OWIS method SOS, which uses a self-attention-based object prior for generating pseudo annotations, outperforms recent state-of-the-art OWIS methods across challenging open-world setups in COCO, LVIS, and ADE20k datasets. SOS especially improves the precision based on the high-quality pseudo annotations. Moreover, SOS with its focus on objects also outperforms vanilla SAM that segments both object and stuff regions." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.532, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SOS: Segment Object System" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.173, + 0.785, + 0.214 + ], + "angle": 0, + "content": "1. Ahn, J., Cho, S., Kwak, S.: Weakly supervised learning of instance segmentation with inter-pixel relations. In: Conference on Computer Vision and Pattern Recognition (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.215, + 0.785, + 0.242 + ], + "angle": 0, + "content": "2. Alexe, B., Deselaers, T., Ferrari, V.: What is an object? In: Conference on Computer Vision and Pattern Recognition (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.243, + 0.785, + 0.269 + ], + "angle": 0, + "content": "3. Baugh, M., Batten, J., Müller, J.P., Kainz, B.: Zero-shot anomaly detection with pre-trained segmentation models. arXiv preprint arXiv:2306.09269 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.27, + 0.785, + 0.31 + ], + "angle": 0, + "content": "4. Bommasani, R., Hudson, D.A., Adeli, E., Altman, R., Arora, S., von Arx, S., Bernstein, M.S., Bohg, J., Bosselut, A., Brunskill, E., et al.: On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.311, + 0.785, + 0.351 + ], + "angle": 0, + "content": "5. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in Neural Information Processing Systems 33 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.352, + 0.785, + 0.392 + ], + "angle": 0, + "content": "6. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: International Conference on Computer Vision (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.393, + 0.785, + 0.433 + ], + "angle": 0, + "content": "7. Chen, T., Mai, Z., Li, R., Chao, W.l.: Segment anything model (SAM) enhanced pseudo labels for weakly supervised semantic segmentation. arXiv preprint arXiv:2305.05803 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.434, + 0.785, + 0.473 + ], + "angle": 0, + "content": "8. Cheng, M.M., Mitra, N.J., Huang, X., Torr, P.H., Hu, S.M.: Global contrast based salient region detection. IEEE Transactions on Pattern Analysis and Machine Intelligence 37(3) (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.474, + 0.785, + 0.528 + ], + "angle": 0, + "content": "9. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: BERT: Pre-training of deep bidirectional transformers for language understanding. In: Human Language Technology Conference of the North American Chapter of the Association for Computational Linguistics (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.529, + 0.785, + 0.555 + ], + "angle": 0, + "content": "0. Dollar, P., Zitnick, C.L.: Structured forests for fast edge detection. In: International Conference on Computer Vision (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.556, + 0.785, + 0.609 + ], + "angle": 0, + "content": "1. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.61, + 0.785, + 0.651 + ], + "angle": 0, + "content": "2. Everingham, M., Eslami, S.A., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The PASCAL visual object classes challenge: A retrospective. International Journal of Computer Vision 111 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.652, + 0.785, + 0.677 + ], + "angle": 0, + "content": "3. Felzenszwalb, P.F., Huttenlocher, D.P.: Efficient graph-based image segmentation. International Journal of Computer Vision 59 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.678, + 0.785, + 0.718 + ], + "angle": 0, + "content": "4. Frintrop, S., Werner, T., Martin Garcia, G.: Traditional saliency reloaded: A good old model in new shape. In: Conference on Computer Vision and Pattern Recognition (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.719, + 0.785, + 0.759 + ], + "angle": 0, + "content": "5. Gao, G., Lauri, M., Zhang, J., Frintrop, S.: Saliency-guided adaptive seeding for supervoxel segmentation. In: International Conference on Intelligent Robots and Systems (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.76, + 0.785, + 0.772 + ], + "angle": 0, + "content": "6. Girshick, R.: Fast R-CNN. In: International Conference on Computer Vision (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.773, + 0.785, + 0.8 + ], + "angle": 0, + "content": "7. Gupta, A., Dolkar, P., Girshick, R.: LVIS: A dataset for large vocabulary instance segmentation. In: Conference on Computer Vision and Pattern Recognition (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.801, + 0.785, + 0.84 + ], + "angle": 0, + "content": "8. He, C., Li, K., Zhang, Y., Xu, G., Tang, L., Zhang, Y., Guo, Z., Li, X.: Weakly-supervised concealed object segmentation with SAM-based pseudo labeling and multi-scale feature grouping. arXiv preprint arXiv:2305.11003 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.173, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.128 + ], + "angle": 0, + "content": "C. Wilms et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "19. He, K., Gkioxari, G., Dollar, P., Girshick, R.: Mask R-CNN. In: International Conference on Computer Vision (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.203 + ], + "angle": 0, + "content": "20. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Conference on Computer Vision and Pattern Recognition (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.785, + 0.245 + ], + "angle": 0, + "content": "21. Hou, Q., Cheng, M.M., Hu, X., Borji, A., Tu, Z., Torr, P.H.: Deeply supervised salient object detection with short connections. In: Conference on Computer Vision and Pattern Recognition (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.785, + 0.285 + ], + "angle": 0, + "content": "22. Hu, H., Lan, S., Jiang, Y., Cao, Z., Sha, F.: FastMask: Segment multi-scale object candidates in one shot. In: Conference on Computer Vision and Pattern Recognition (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.327 + ], + "angle": 0, + "content": "23. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International Conference on Machine Learning (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.369 + ], + "angle": 0, + "content": "24. Kalluri, T., Wang, W., Wang, H., Chandraker, M., Torresani, L., Tran, D.: Openworld instance segmentation: Top-down learning with bottom-up supervision. arXiv preprint arXiv:2303.05503 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.411 + ], + "angle": 0, + "content": "25. Kim, B., Yoo, Y., Rhee, C.E., Kim, J.: Beyond semantic to instance segmentation: Weakly-supervised instance segmentation via semantic knowledge transfer and self-refinement. In: Conference on Computer Vision and Pattern Recognition (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.452 + ], + "angle": 0, + "content": "26. Kim, D., Lin, T.Y., Angelova, A., Kweon, I.S., Kuo, W.: Learning open-world object proposals without learning to classify. IEEE Robotics and Automation Letters 7(2) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.494 + ], + "angle": 0, + "content": "27. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: International Conference on Computer Vision (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.785, + 0.522 + ], + "angle": 0, + "content": "28. Klein, D.A., Frintrop, S.: Center-surround divergence of feature statistics for salient object detection. In: International Conference on Computer Vision (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.785, + 0.563 + ], + "angle": 0, + "content": "29. Li, J., Li, D., Savarese, S., Hoi, S.: BLIP-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.785, + 0.605 + ], + "angle": 0, + "content": "30. Lin, T.Y., Dollar, P., Girshick, R., He, K., Hariharan, B., Belongie, S.: Feature pyramid networks for object detection. In: Conference on Computer Vision and Pattern Recognition (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.785, + 0.646 + ], + "angle": 0, + "content": "31. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft COCO: Common objects in context. In: European Conference on Computer Vision (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.688 + ], + "angle": 0, + "content": "32. Linardos, A., Kümmerer, M., Press, O., Bethge, M.: DeepGaze IIE: Calibrated prediction in and out-of-domain for state-of-the-art saliency modeling. In: International Conference on Computer Vision (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.785, + 0.729 + ], + "angle": 0, + "content": "33. Liu, N., Han, J., Yang, M.H.: PiCANet: Learning pixel-wise contextual attention for saliency detection. In: Conference on Computer Vision and Pattern Recognition (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.785, + 0.771 + ], + "angle": 0, + "content": "34. Liu, T., Yuan, Z., Sun, J., Wang, J., Zheng, N., Tang, X., Shum, H.Y.: Learning to detect a salient object. IEEE Transactions on Pattern analysis and machine intelligence 33(2) (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.812 + ], + "angle": 0, + "content": "35. Liu, Y., Wu, Y.H., Wen, P., Shi, Y., Qiu, Y., Cheng, M.M.: Leveraging instance-, image-and dataset-level information for weakly supervised instance segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(3) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "36. Ma, J., Ming, A., Huang, Z., Wang, X., Zhou, Y.: Object-level proposals. In: International Conference on Computer Vision (2017)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.532, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SOS: Segment Object System" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "37. Martin García, G., Potapova, E., Werner, T., Zillich, M., Vincze, M., Frintrop, S.: Saliency-based object discovery on RGB-D data with a late-fusion approach. In: International Conference on Robotics and Automation (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "38. Mazurowski, M.A., Dong, H., Gu, H., Yang, J., Konz, N., Zhang, Y.: Segment anything model for medical image analysis: an experimental study. Medical Image Analysis 89 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.233, + 0.786, + 0.274 + ], + "angle": 0, + "content": "39. Perazzi, F., Krahenbuhl, P., Pritch, Y., Hornung, A.: Saliency filters: Contrast based filtering for salient region detection. In: Conference on Computer Vision and Pattern Recognition (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.276, + 0.786, + 0.303 + ], + "angle": 0, + "content": "40. Pinheiro, P.O., Collobert, R., Dollár, P.: Learning to segment object candidates. Advances in Neural Information Processing Systems 28 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.304, + 0.786, + 0.332 + ], + "angle": 0, + "content": "41. Pinheiro, P.O., Lin, T.Y., Collobert, R., Dollár, P.: Learning to refine object segments. In: European Conference on Computer Vision (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.333, + 0.787, + 0.387 + ], + "angle": 0, + "content": "42. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.388, + 0.787, + 0.416 + ], + "angle": 0, + "content": "43. Rahtu, E., Kannala, J., Blaschko, M.: Learning a category independent object detection cascade. In: International Conference on Computer Vision (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.417, + 0.787, + 0.459 + ], + "angle": 0, + "content": "44. Ronneberger, O., Fischer, P., Brox, T.: U-Net: Convolutional networks for biomedical image segmentation. In: International Conference on Medical Image Computing and Computer-Assisted Intervention (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.459, + 0.787, + 0.5 + ], + "angle": 0, + "content": "45. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al.: Imagenet large scale visual recognition challenge. International Journal of Computer Vision 115 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.501, + 0.787, + 0.529 + ], + "angle": 0, + "content": "46. Saito, K., Hu, P., Darrell, T., Saenko, K.: Learning to detect every thing in an open world. In: European Conference on Computer Vision (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.53, + 0.787, + 0.557 + ], + "angle": 0, + "content": "47. Shi, Z., Sun, Y., Zhang, M.: Training-free object counting with prompts. In: Winter Conference on Applications of Computer Vision (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.558, + 0.787, + 0.599 + ], + "angle": 0, + "content": "48. Wang, C., Wang, G., Zhang, Q., Guo, P., Liu, W., Wang, X.: OpenInst: A simple query-based method for open-world instance segmentation. arXiv preprint arXiv:2303.15859 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.6, + 0.787, + 0.642 + ], + "angle": 0, + "content": "49. Wang, W., Feiszli, M., Wang, H., Malik, J., Tran, D.: Open-world instance segmentation: Exploiting pseudo ground truth from learned pairwise affinity. In: Conference on Computer Vision and Pattern Recognition (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.643, + 0.787, + 0.684 + ], + "angle": 0, + "content": "50. Wang, W., Feiszli, M., Wang, H., Tran, D.: Unidentified video objects: A benchmark for dense, open-world segmentation. In: International Conference on Computer Vision (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.685, + 0.787, + 0.713 + ], + "angle": 0, + "content": "51. Wei, Z., Chen, P., Yu, X., Li, G., Jiao, J., Han, Z.: Semantic-aware SAM for point-prompted instance segmentation. arXiv preprint arXiv:2312.15895 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.714, + 0.787, + 0.741 + ], + "angle": 0, + "content": "52. Wilms, C., Frintrop, S.: Edge adaptive seeding for superpixel segmentation. In: German Conference on Pattern Recognition (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.742, + 0.787, + 0.77 + ], + "angle": 0, + "content": "53. Wilms, C., Frintrop, S.: AttentionMask: Attentive, efficient object proposal generation focusing on small objects. In: Asian Conference on Computer Vision (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.771, + 0.787, + 0.812 + ], + "angle": 0, + "content": "54. Wu, J., Jiang, Y., Yan, B., Lu, H., Yuan, Z., Luo, P.: Exploring transformers for open-world instance segmentation. In: International Conference on Computer Vision (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "55. Xie, C., Xiang, Y., Mousavian, A., Fox, D.: Unseen object instance segmentation for robotic environments. IEEE Transactions on Robotics 37(5) (2021)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.377, + 0.128 + ], + "angle": 0, + "content": "C. Wilms et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "56. Xue, X., Yu, D., Liu, L., Liu, Y., Li, Y., Yuan, Z., Song, P., Shou, M.Z.: Single-stage open-world instance segmentation with cross-task consistency regularization. arXiv preprint arXiv:2208.09023 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.217 + ], + "angle": 0, + "content": "57. Yang, X., Gong, X.: Foundation model assisted weakly supervised semantic segmentation. In: Winter Conference on Applications of Computer Vision (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.244 + ], + "angle": 0, + "content": "58. Yang, Y., Zhou, Z., Wu, J., Wang, Y., Xiong, R.: Class semantics modulation for open-set instance segmentation. IEEE Robotics and Automation Letters (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.785, + 0.285 + ], + "angle": 0, + "content": "59. Zhao, J.X., Liu, J.J., Fan, D.P., Cao, Y., Yang, J., Cheng, M.M.: EGNet: Edge guidance network for salient object detection. In: International Conference on Computer Vision (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.327 + ], + "angle": 0, + "content": "60. Zhou, B., Khosla, A., Lapedriza, A., Oliva, A., Torralba, A.: Learning deep features for discriminative localization. In: Conference on Computer Vision and Pattern Recognition (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.369 + ], + "angle": 0, + "content": "61. Zhou, B., Zhao, H., Puig, X., Xiao, T., Fidler, S., Barriuso, A., Torralba, A.: Semantic understanding of scenes through the ADE20K dataset. International Journal of Computer Vision 127 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.37, + 0.785, + 0.411 + ], + "angle": 0, + "content": "62. Zhu, M., Li, H., Chen, H., Fan, C., Mao, W., Jing, C., Liu, Y., Shen, C.: SegPrompt: Boosting open-world segmentation via category-level prompt learning. In: International Conference on Computer Vision (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.785, + 0.439 + ], + "angle": 0, + "content": "63. Zitnick, C.L., Dollár, P.: Edge boxes: Locating object proposals from edges. In: European Conference on Computer Vision (2014)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.439 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/SOS_ Segment Object System for Open-World Instance Segmentation With Object Priors/466dafa4-76e1-41c2-9ff8-766257e5e796_origin.pdf b/2024/SOS_ Segment Object System for Open-World Instance Segmentation With Object Priors/466dafa4-76e1-41c2-9ff8-766257e5e796_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..dc8888e00a695e34f06259b0da583efcbd17d7bd --- /dev/null +++ b/2024/SOS_ Segment Object System for Open-World Instance Segmentation With Object Priors/466dafa4-76e1-41c2-9ff8-766257e5e796_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:75e3e69358f39c9e76225fb5ddf6967c5d4cd77f101be2b4ceebebbee29c66e1 +size 24991148 diff --git a/2024/SOS_ Segment Object System for Open-World Instance Segmentation With Object Priors/full.md b/2024/SOS_ Segment Object System for Open-World Instance Segmentation With Object Priors/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d92e6184be9040ddacf30e3df7da247d4bab5096 --- /dev/null +++ b/2024/SOS_ Segment Object System for Open-World Instance Segmentation With Object Priors/full.md @@ -0,0 +1,308 @@ +# SOS: Segment Object System for Open-World Instance Segmentation With Object Priors + +Christian Wilms1, Tim Rolff1,2, Maris Hillemann1, Robert Johanson1, and Simone Frinthrop1 + +1 Computer Vision Group, University of Hamburg, Germany +2 Human-Computer Interaction Group, University of Hamburg, Germany {firstname_lastname}@uni-hamburg.de + +Abstract. We propose an approach for Open-World Instance Segmentation (OWIS), a task that aims to segment arbitrary unknown objects in images by generalizing from a limited set of annotated object classes during training. Our Segment Object System (SOS) explicitly addresses the generalization ability and the low precision of state-of-the-art systems, which often generate background detections. To this end, we generate high-quality pseudo annotations based on the foundation model SAM [27]. We thoroughly study various object priors to generate prompts for SAM, explicitly focusing the foundation model on objects. The strongest object priors were obtained by self-attention maps from self-supervised Vision Transformers, which we utilize for prompting SAM. Finally, the post-processed segments from SAM are used as pseudo annotations to train a standard instance segmentation system. Our approach shows strong generalization capabilities on COCO, LVIS, and ADE20k datasets and improves on the precision by up to $81.6\%$ compared to the state-of-the-art. Source code is available at: https://github.com/chwilms/SOS + +Keywords: Open-world Instance Segmentation $\cdot$ Object Localization $\cdot$ Prompting + +# 1 Introduction + +Open-World Instance Segmentation (OWIS) is the task of segmenting all object instances in an image by learning from a limited set of known object classes [26, 49, 50]. In contrast to instance segmentation, OWIS is not limited by the closed-world assumption, which assumes that all object classes are known in advance. Since OWIS methods aim to detect not only learned but also unknown object classes, they return class-agnostic detections. This is of particular interest in real-world scenarios with previously unknown object classes (e.g., [55]) and challenges the systems' generalization capabilities. An example of this is shown in the first row of Fig. 1, where annotations for classes such as surfboard or tennis racket do not exist during training, but the objects should be detected during testing. + +Previous OWIS methods mostly generate pseudo annotations for unannotated objects [24, 46, 49, 56] or replace the foreground-background classification + +![](images/b90e68b72143383adf80f1e1ba9fcc4f24d0c46409d00badc90355a6e7ca539d.jpg) +Fig. 1: Comparison of (pseudo) annotations (left) used by Mask R-CNN [19], GGN [49], and our SOS, when only original annotations of VOC object classes are given. While Mask R-CNN only uses original annotations without object classes such as tennis racket or surfboard (red arrows), SOS generates pseudo annotations covering those classes (green arrows). GGN generates noisy pseudo annotations including background areas. As a result, only SOS constantly detects these objects not annotated in training (green vs. red arrows on the right). Filled masks denote annotations (left) or detected objects (right), while red frames indicate missed objects. + +of possible objects in instance segmentation systems with a learning target focusing on the localization quality (e.g., intersection over union) [26,48,54]. While OWIS methods generalize better to unseen object classes than standard instance segmentation systems, they still exhibit low precision [24,26,46,49,54]. One reason for this is the use of noisy pseudo annotations covering background areas as visible in the left part of the second row in Fig. 1. + +Recently, foundation models [4] emerged as a promising technique [5,9,23,29, 42]. They are trained on large datasets with a surrogate task and applied to other tasks in a zero- or few-shot manner using prompting. The first foundation model for image segmentation, the Segment Anything Model (SAM), was proposed by [27]. Trained on a large automatically annotated dataset, SAM generates segments for object or stuff regions based on point, box, mask, or text prompts without further training. Despite not only segmenting objects but also stuff regions, given well-designed prompts, SAM is able to generate high-quality object segments. Note that vanilla SAM does not address object segmentation tasks like OWIS itself due to the aforementioned lacking focus on objects. + +In this paper, we propose the Segment Object System (SOS), a novel OWIS method utilizing high-quality pseudo annotations to improve the generalization capability of a standard instance segmentation system (see lower row in Fig. 1). For generating pseudo annotations, we apply the foundation model SAM with prompts derived from an object prior to roughly localize objects of arbitrary classes and to limit segmentations of stuff regions by SAM. To ensure the high quality of our pseudo annotations based on SAM, improving the overall precision, we thoroughly study various hand-crafted and learned object priors for an + +object-focused application of SAM. This is relevant beyond the scope of OWIS. Using our study findings, SOS utilizes self-attention maps from self-supervised Vision Transformers (ViTs) [11] as object prior for prompting SAM. Finally, SOS filters low-quality pseudo annotations, combines the pseudo annotations with the original annotations of the known classes, and trains a standard instance segmentation system with these mixed annotations. Our extensive evaluation shows the strong generalization abilities of SOS, considerably outperforming all previous state-of-the-art systems across COCO [31], LVIS [17], and ADE20k [61] datasets. Most notably, the results improve by up to $81.6\%$ in terms of precision over the state-of-the-art due to the high-quality pseudo annotations. We also show that SOS is better suited for OWIS than directly applying SAM. + +Overall, our contributions are threefold: + +- We propose SOS, a novel OWIS method based on a learned object prior, prompt-based pseudo annotations, and an arbitrary instance segmentation system. +- We thoroughly study various object priors for focusing SAM on objects, leading to high-quality object segments for pseudo annotations and beyond. +- Our extensive evaluation shows that the high-quality pseudo annotations in SOS lead to high precision and strong overall results, clearly outperforming other methods. + +# 2 Related Work + +# 2.1 Open-world Instance Segmentation + +To address OWIS, literature mainly offers two streams. First, systems replace the hard classification in instance segmentation systems with a localization score. This avoids classifying unseen objects as background and improves generalization ability [26]. In OLN [26], centerness and Intersection over Union (IoU) are learned as localization scores in a Mask R-CNN system [19]. SWORD [54] and OpenInst [48] follow the same idea for query-based systems. + +The second line of works [24, 46, 49, 56] addresses OWIS by augmenting the annotations from known classes with pseudo annotations aiming to cover objects of unknown classes. To generate pseudo annotations, GGN [49] learns pixel affinities from known classes to generate segments, which are grouped. Similarly, UDOS [24] groups object proposals of object parts to create pseudo annotations, while LDET [46] applies copy-paste-augmentation. Recently, a new stream emerged [58, 62] explicitly utilizing class labels from known classes. + +We follow the second stream and create pseudo annotations within SOS, since it offers the most flexibility w.r.t. the base instance segmentation system. Different from existing approaches that generate noisy pseudo annotations (see GGN in Fig. 1), we investigate and develop object priors to effectively use recent foundation models to create high-quality object-focused pseudo annotations, leading to high recall and precision. + +# 2.2 Class-agnostic Object Localization + +Localizing objects in images is related to several computer vision tasks. For instance, in traditional object proposal generation [2], cues like saliency [2, 37] or contour information [2, 10, 36, 43] were used to localize class-agnostic object candidates for object detection. Since the advent of deep learning, the models are learned from data [16, 22, 40, 41, 53]. + +Focusing on arbitrary salient objects, known as salient object detection, several approaches use the contrast of hand-crafted features such as color [8,28,34, 39] or edge information [28], as well as CNN-based approaches [21, 33, 59]. In a different task, several authors investigated the information stored in learned classifiers or self-supervised feature extractors based on CNNs or ViTs to locate objects. Notable avenues include Class Activation Maps (CAMs) [60] that highlight discriminative object or image parts in CNN-based classifiers and the self-attention maps in the self-supervised DINO feature extractor [6] based on ViTs, which contain information on the scene layout indicating object locations. + +In our study on object priors to focus SAM on objects, we investigate several of these object localization cues. In contrast to these approaches, we follow a different goal, focusing prompt-based segmentations on objects for OWIS. + +# 2.3 Applications of SAM + +SAM [27] is used in several segmentation tasks [3,38,47]. Similar to us, [7] and [57] use SAM to generate pseudo annotation masks for weakly-supervised semantic segmentation based on image-level supervision. Similarly, [18] generate pseudo annotation masks for concealed object segmentation based on scribble annotations. Finally, [51] generate pseudo annotation masks for weakly-supervised instance segmentation in a multiple instance learning framework. + +Different from the weakly-supervised approaches utilizing SAM, we do not rely on supervision. Hence, a key novelty of our work lies in investigating object priors to focus SAM on segmenting arbitrary objects in the absence of supervision for all object classes. Moreover, we are the first to apply SAM-based pseudo annotations in OWIS. + +# 3 Segment Object System for OWIS + +To address OWIS, we propose our Segment Object System (SOS) that generates high-quality pseudo annotations to train a standard instance segmentation system. As visible from Fig. 2, SOS consists of three main blocks. The first block, the Object Localization Module (OLM), described in Sec. 3.2 (yellow area in Fig. 2), aims to roughly localize objects by sampling locations from an object prior that reflects the object probability per image coordinate. The output of OLM is a set of likely object locations, that is used in the Pseudo Annotation Creator (PAC), described in Sec. 3.3 (green area in Fig. 2), to prompt the pre-trained SAM [27], which we briefly revisit in Sec. 3.1. From the segments + +![](images/6f7a464a696853056538da30d03ea8fd1653e04bd52c82b1979dab2ef6b1bbd2.jpg) +Fig. 2: Overview of our Segment Object System (SOS) for OWIS consisting of three blocks. First, the input image is processed in our Object Localization Module (OLM, yellow area) to create object-focused point prompts roughly localizing objects. Second, our Pseudo Annotations Creator (PAC, green area) generates segments based on the previously generated prompts using SAM [27], and further processes them, leading to a final set of merged original and pseudo annotations. Finally, the merged annotations are used to train an instance segmentation system (blue area). + +generated from likely object locations, the PAC generates pseudo annotations by filtering low-quality segments and removing near-duplicates. Finally, the pseudo annotations are merged with the original annotations, and a standard instance segmentation system is trained with the merged annotations (see Sec. 3.4, blue area in Fig. 2). During testing, only this instance segmentation system is used. + +# 3.1 Revisit Segment Anything Model + +Since SAM [27] is integral to SOS, we will briefly review the model. The idea of SAM is to utilize prompts such as point coordinates to generate segments of image parts like objects, object parts, or stuff regions localized by these prompts. Hence, SAM does not exclusively segment objects unless explicitly instructed by object-focused prompts, as proposed in this paper. SAM is based on a ViT [11] image encoder creating an image embedding, while the embeddings for point prompts are generated based on positional encodings. Both embeddings are processed in a mask decoder utilizing a transformer decoder block and a mask prediction head to create the final segment including a confidence score. Note that SAM is class-agnostic and does not classify the segments. + +For training SAM, [27] propose the new large-scale SA-1B dataset that is annotated with a data engine using SAM itself. In the pre-stage, SAM is trained with public segmentation datasets to aid the first stage of the data engine, where annotators use SAM as an interactive segmentation tool to create initial segmentations. Subsequently, SAM is retrained with the newly annotated data, fully avoiding data leakage w.r.t. public segmentation datasets. In subsequent stages, manual annotations were provided for segments missed by SAM. At last, annotations were created fully automatically by SAM. See [27] for more details. + +# 3.2 Object Localization Module + +The first block of SOS (yellow area in Fig. 2) is the Object Localization Module (OLM). OLM roughly localizes all objects in the input image, and creates object-focused point prompts. As the first step, OLM generates an object prior from the input image. The object prior is a probability mass function constrained to the image plane, although other formulations are possible (see Sec. 4), and indicates the probability of an image coordinate being part of any object. Hence, the object prior is class-agnostic and not limited to known object classes from the training set in OWIS. In our study in Sec. 4, we analyze various object priors. For final SOS, we utilize the six self-attention maps from the self-attention heads of a ViT's final layer, pre-trained in the self-supervised DINO framework [6] on ImageNet [45]. See Sec. 4 for justifications. OLM aggregates the six self-attention maps by elementwise max yielding one map highlighting all relevant scene elements. Subsequently, OLM converts the map to an object prior, by rescaling such that the minimum equals 0 and the sum over the object prior is 1. + +Next, OLM randomly samples $S$ (here: $S = 50$ ) image coordinates from the object prior, given the per-coordinate values of the object prior as probabilities. To diversify the samples across multiple objects, we prune parts of the object prior around the sampled coordinate. Specifically, given a coordinate $(x_s, y_s)$ , we set the object prior to zero for all coordinates $\{(x, y) \mid |x - x_s| \leq N \land |y - y_s| \leq N\}$ , with $N = 20$ . To transform the object prior back to a probability mass function, we again apply the rescaling as described above. Subsequently, OLM iteratively applies this sampling until a set of $S$ coordinates are extracted, representing object-focused point prompts. This set is the output of the OLM and will be used for prompting in the Pseudo Annotation Creator. + +# 3.3 Pseudo Annotations Creator + +The Pseudo Annotations Creator (PAC, green area in Fig. 2) is the second block of SOS. It generates pseudo annotations utilizing SAM, given the object-focused point prompts provided by OLM. These object-focused point prompts lifts SAM from a system that segments anything to a system that segments objects. As the first step, PAC prompts a pre-trained SAM with the $S$ point prompts from OLM. To handle ambiguous point prompts that may indicate multiple objects or object parts, we follow [27] and allow SAM to generate three segments per prompt. Each generated segment is a potential pseudo annotation. The resulting set of + +class-agnostic segments is noisy, with segments covering objects and background, or covering the same image area. Therefore, PAC first utilizes SAM's confidence score per segment and removes segments with a confidence below $\tau_{\mathrm{conf}}$ . Second, PAC removes near-duplicates by non-maximum suppression with an IoU threshold $\tau_{\mathrm{NMS}}$ . We set $\tau_{\mathrm{conf}} = 0.9$ and $\tau_{\mathrm{NMS}} = 0.95$ for SOS. + +Given this set of pseudo annotations, PAC merges the set with the original annotations representing known classes of the training dataset to combine the knowledge of the human annotators and the knowledge of our object-focused SAM. Since we only want to keep pseudo annotations for objects not covered by the original annotations, we suppress pseudo annotations that have a high IoU ( $\tau_{\mathrm{NMS}}$ ) with at least one original annotation. Moreover, we limit the number of pseudo annotations to $P$ in order to balance original and pseudo annotations. In our final SOS, we set $P = 10$ . However, across the COCO training dataset, only 7.8 pseudo annotations are added on average. The other pseudo annotations were suppressed in previous steps. Overall, the output of PAC is a mixed set of annotations, containing all original annotations augmented with high-quality pseudo annotations covering objects of unknown classes. + +# 3.4 Instance Segmentation + +In the last block of SOS, we take the merged annotations generated by PAC and train a standard instance segmentation system. While this system can be arbitrary and is generally replaceable, we choose Mask R-CNN [19] with a ResNet-50+FPN backbone [30] trained in a class-agnostic setting, following recent OWIS methods [24, 26, 46, 49]. Overall, this leads to an OWIS method based on high-quality pseudo annotations. + +# 4 Object Priors for SOS + +As described in Sec. 3.2, SOS uses an object prior for rough object localization and subsequent point prompt generation for PAC. This section thoroughly investigates the performance of various object priors derived from previous class-agnostic object localization works in SOS. For details on the object priors beyond the subsequent descriptions, we refer to our supplementary. + +# 4.1 Object Priors + +Baselines: We introduce three baseline object priors that ignore the image content. First, Grid uses SAM with a regular $64 \times 64$ grid of points, following [27]. This directly leads to prompts without the sampling described in Sec. 3.2. Note that the Grid baseline will lead to segments for both object and stuff regions. Second, we use the spatial distribution of object centroids from the known classes across the training dataset as object prior (Dist). We assume that this generalizes well since it ignores class-specific object features. Finally, we utilize the centroids of the training set's objects from unknown classes as optimal prompts ( $GT$ ), representing an upper bound. + +Superpixels: Inspired by [15, 52], we utilize the centroids of superpixels generated with an adaptive superpixel segmentation method as object prior ( $Spx$ ). The intuition is that large uniform areas not containing objects should be covered by few superpixels and vice versa. Hence, the density of superpixels is a surrogate for the density of objects. We use the well-known FH superpixels [13] since they adapt their density to the image content following our assumption. Note that every superpixel centroid is a point prompt. + +Contour Density: Since objects are defined by their outer contours [63], we use contour density as an object prior (Contour). Similar to [52], we assume that in areas of high contour density, several objects are located and vice versa. As a surrogate for the contour density, we use edge density based on [10]. + +Saliency: Saliency is used in object proposal generation to localize objects [2, 37] as they stick out from their surrounding. Here, we evaluate two saliency methods as object priors. First, we apply the traditional approach VOCUS2 [14] based on color contrast and used by [37] (VOCUS2). Second, we use pre-trained DeepGaze IIE [32] saliency maps learned from eye fixation data (DeepGaze). + +Class Activation Maps: CAMs [60] indicate discriminative image regions for CNN-based classifiers. Since they only need image-level supervision, they are frequently applied in weakly-supervised tasks to locate objects [1, 25, 35]. We use the CAM of the predicted class in a ResNet-50 classifier [20] pre-trained on ImageNet [45] as object prior (CAM). + +Self-attention: Recently, [6] have shown that self-attention maps learned inside self-supervised ViTs encode the scene layout, including object locations. Therefore, we explore the self-attention maps from the final layer of a ViT-S backbone trained in the self-supervised DINO framework [6] as an object prior (DINO). + +Learned Object Locations: Finally, we learn object locations based on the known classes of a dataset using a U-Net [44] in a binary segmentation task (object vs. no object) as object prior ( $U$ -Net). This assumes that, on the level of individual pixels, the model generalizes from known to unknown classes [49,56]. + +# 4.2 Study Setup for Object Priors + +To assess each object prior, we train SOS with the respective prior on the COCO training set using only annotations from the 20 PASCAL VOC [12] classes. We evaluate on COCO's validation set with annotations from the remaining 60 classes in COCO, following standard COCO (VOC) $\rightarrow$ COCO (non-VOC) cross-category evaluation in OWIS [26, 46, 49]. Note that we use the default class-agnostic Mask R-CNN in SOS but reduce the training schedule to only a quarter of steps for faster training. We evaluate the results of SOS given each object priors and report Average Recall (AR) for 100 detections and Average Precision (AP) as previous OWIS research [46, 49]. We also report $F_{1}$ score, the harmonic mean between AR and AP, yielding a single number for comparison. + +![](images/98c534e1f46cf8af01435b9b48398d2a912be5cb6c56f5d701d91c242fcf430b.jpg) +Input image + +![](images/003002287514d3fb661c3340c460d61ece1d66380508a497953a7041dc9c0ebe.jpg) + +![](images/45e5bc61bd5b28ddc0bdbad119f694a1ab734f9401f8e33570f704fc9f26f2b4.jpg) +VOCUS2 +Fig. 3: Object priors VOCUS2, U-Net, and DINO with resulting pseudo annotations. + +![](images/867b8ce8c01c3d142b7ac46172c54b92a51ac566435015d9e7a2a0e8386ed0fe.jpg) + +![](images/e18673af3e9f66db67f8d6874d2cbc02569e3bc22acb0b15c3d3538437de1770.jpg) +U-Net + +![](images/f8091942473796a09656b9da90cd195fc51d5c9f4bf39e4ad0622fae5d28d8ff.jpg) + +![](images/deb4b2e0d49cc2d6854a38a525247c347665344734ea87235a23b97376720973.jpg) +DINO + +# 4.3 Results of Object Priors in SOS + +Table 1 shows the results of our object prior study. First, all priors outperform the baselines Grid and Dist in terms of $\mathbf{F}_1$ . Both baselines exhibit a lower AP compared to all other object priors, reflecting the missing focus on objects. Hence, simply applying SAM (Grid) is not suitable to segment only objects, as several stuff regions are segmented as well. Second, most priors exhibit a similar performance in terms of all measures. For instance, learning-based CAM and DeepGaze do not outperform simple priors (Spx, Contour, VOCUS2). Only the priors U-Net and DINO perform substantially better, with DINO producing the best result. Comparing DINO to GT reveals that DINO is able to recall almost the same amount of objects, but exhibits a lower precision, as expected. + +To illustrate the results, Fig. 3 depicts object priors for top-performing VOCUS2, $U$ -Net, and DINO, as heatmaps (upper row) with the resulting pseudo annotations generated in SOS (lower row) for an example image. The heatmaps show that DINO is very focused on the discriminative object parts, here the polar bears' faces, while $U$ -Net highlights, as learned, the entire objects. Moreover, $U$ -Net also highlights several background areas. Similarly, VOCUS2 highlights background areas and the polar bears since it is generally focused on high-contrast areas. Overall, only DINO exclusively generates pseudo annotations for the polar bears and some sub-parts, while $U$ -Net and VOCUS2 produce noisy annotations including background patches, similar to previous approaches [49]. + +The results show that a well-designed object prior like DINO substantially outperforms the baseline Grid $(+7.5$ in $\mathrm{F}_1)$ in SOS. Therefore, we use DINO as the object prior in SOS. More generally, the study also reveals important insights on how to prompt SAM for class-agnostic, object-focused results for arbitrary tasks. For more study results, we refer to our supplementary. + +# 5 Evaluation + +In our evaluation, we compare SOS to recent OWIS methods OLN [26], LDET [46], GGN [49], SWORD [54], SOIS [56], OpenInst [48], UDOS [24], and against Mask R-CNN [19] and SAM [27] baselines, depending on the availability of public + +Table 1: Results of our SOS with various object priors in the cross-category COCO (VOC) $\rightarrow$ COCO (non-VOC) setting. *: Uses ground truth of unknown classes. + +
Object PriorAPAR100F1Object PriorAPAR100F1
SOS + Grid3.836.56.9SOS + DeepGaze5.435.99.4
SOS + Dist3.427.46.0SOS + CAM5.436.79.4
SOS + Spx5.634.89.6SOS + DINO8.938.114.4
SOS + Contour5.636.69.7SOS + U-Net7.337.312.2
SOS + VOCUS25.137.710.5SOS + GT18.1*42.5*25.4*
+ +source code or reported results. To assess the quality of the approaches, we use four cross-category and cross-dataset settings, commonly used in OWIS. Specifically, we choose COCO (VOC) $\rightarrow$ COCO (non-VOC) on the COCO dataset [31], already discussed in Sec. 4.2. For cross-dataset evaluation, we use COCO $\rightarrow$ LVIS, COCO $\rightarrow$ ADE20k, and COCO $\rightarrow$ UVO that train on the entire COCO training set and evaluate on the validation sets of LVIS [17], ADE20k [61], and UVO [50] with exhaustive annotations. The LVIS dataset covers the same images as COCO but extends the annotated object classes from 80 to 1203. ADE20k includes 3169 classes of objects, object parts, and stuff regions. Finally, UVO features exhaustive annotations of all objects and, therefore, no system is penalized for detecting objects outside a datasets' taxonomy. + +As in Sec. 4.2 and following common practice in OWIS [26,46,49], we report AR for up to 100 detections and AP. We again also report the $\mathrm{F}_1$ score between AR and AP as a single, combined quantity, described in Sec. 4.2. Note that we only evaluate based on masks, not boxes. + +# 5.1 Implementation Details and Data Usage + +As discussed in Sec. 4.3, SOS uses the object prior DINO for best results. DINO [6] with a ViT-S backbone is used as an ImageNet [45] pre-trained model without fine-tuning. Similarly, SOS uses SAM [27] with the ViT-H backbone as a pre-trained model based on SA-1B [27] without fine-tuning. As an instance segmentation model, we apply Mask R-CNN [19] with an ImageNet pre-trained ResNet-50+FPN backbone [30], and the default configurations except for the class-agnostic training. + +Overall, SOS is based on models pre-trained on standard ImageNet and SA-1B. Pre-training on SA-1B or similar large-scale datasets becomes common with the emergence of foundation models and is widely adopted by systems using SAM in similar contexts [7,51,57]. It is comparable to using models pre-trained on ImageNet that even have class information available during training and also allow a rough localization of objects [6,60]. Moreover, the final, noisy annotations of SA-1B representing objects, object parts, and stuff region were automatically generated by SAM [27], and no data leakage w.r.t. other datasets exists. + +Table 2: Results of two baselines and various OWIS methods in the COCO (VOC) $\rightarrow$ COCO (non-VOC) setting. $\dagger$ : uses automatically annotated SA-1B dataset. + +
SystemAPAR100F1
Mask R-CNN [19]1.08.21.8
SAM†[27]3.648.16.7
OLN [26]4.228.47.3
LDET [46]4.324.87.3
GGN [49]4.928.38.4
SWORD [54]4.830.28.3
UDOS [24]2.934.35.3
SOS†(ours)8.939.314.5
+ +Table 3: Results of two baselines and various OWIS methods in the COCO $\rightarrow$ LVIS setting. $^\dagger$ : uses automatically annotated SA-1B dataset. + +
SystemAPAR100F1
Mask R-CNN [19]7.523.611.4
SAM†[27]6.845.111.8
LDET [46]6.724.810.5
GGN [49]6.527.010.5
SOIS [56]-25.2-
OpenInst [48]-29.3-
UDOS [24]3.924.96.7
SOS†(ours)8.133.313.3
+ +# 5.2 Cross-category COCO (VOC) $\rightarrow$ COCO (non-VOC) Results + +Table 2 presents the results of various OWIS methods on the cross-category setup COCO (VOC) $\rightarrow$ COCO (non-VOC). Our SOS clearly outperforms all OWIS methods in $\mathrm{AR}_{100}$ , AP, and $\mathbf{F}_1$ . Specifically, SOS outperforms the second-best system in terms of $\mathrm{F}_1$ , GGN [49], by 6.1. While some of the improvement comes from a better recall ( $+38.9\%$ compared to GGN), the results are driven by a much-improved precision ( $+81.6\%$ compared to GGN). This reflects the high quality of the pseudo annotations in SOS on this cross-category setting. Moreover, there is a clear improvement of SOS over Mask R-CNN ( $+12.7$ in $\mathrm{F}_1$ ). Compared to original SAM, the DINO-based SOS substantially improves the precision by focusing on objects, leading to an improved $\mathrm{F}_1$ score ( $+7.8$ ). + +The qualitative results of various OWIS methods in Fig. 4 support the quantitative results. In the first example, only GGN [49] and SOS detect both giraffes, however, SOS generates more accurate segmentations. In the second example, only SOS detects all five small surfboards, while other systems detect at most three. Note that class-agnostic Mask R-CNN misses all non-VOC objects in both examples. For more qualitative results, see our supplementary. + +# 5.3 Cross-dataset Results + +$COCO \rightarrow LVIS$ : We evaluate the cross-dataset generalization capabilities of OWIS methods, starting with $COCO \rightarrow LVIS$ . The results in Tab. 3 show a trend similar to the cross-category results. SOS outperforms all OWIS methods across all measures with an improvement of 2.8 in $F_1$ compared to second-best methods GGN [49] and LDET [46]. Similar to the previous results, the improvement is based on gains in both recall and precision, however, the relative improvements in AP and $AR_{100}$ are more similar to each other. Overall, our high-quality pseudo + +![](images/235737833757eee12f46f0948406e067ab08097a1e13ad6052fe5c94d75f0513.jpg) +Fig. 4: Qualitative results of OWIS methods and baseline Mask R-CNN in the cross-category COCO (VOC) $\rightarrow$ COCO (non-VOC) setting. Filled masks denote detected objects, while red frames indicate missed objects. + +Table 4: Results of Mask R-CNN baseline and various OWIS methods in the COCO $\rightarrow$ ADE20k setting. $^\dagger$ : uses automatically annotated SA-1B dataset. + +
SystemAPAR100F1
Mask R-CNN [19]6.911.98.7
OLN [26]-20.4-
LDET [46]9.518.512.6
GGN [49]9.721.013.3
UDOS [24]7.622.911.4
SOS† (ours)12.526.517.0
+ +Table 5: Results of two baselines and various OWIS methods in the COCO $\rightarrow$ UVO setting. $^\dagger$ : uses automatically annotated SA-1B dataset. + +
SystemAPAR100F1
Mask R-CNN [19]20.736.726.5
SAM†[27]11.350.118.4
OLN [26]-41.4-
LDET [46]22.040.428.5
GGN [49]20.343.427.7
UDOS [24]10.643.117.0
SOS†(ours)20.942.328.0
+ +annotations in SOS lead to new state-of-the-art results, which indicate strong generalization to unknown object classes outside COCO. + +$COCO \rightarrow ADE20k$ : The results for $COCO \rightarrow ADE20k$ in Tab. 4, which include labeled object parts of $ADE20k$ , show that SOS again outperforms all other OWIS methods across all measures. The gain over the second-best method in $F_1$ , GGN [49], is 3.7. Similar to $COCO \rightarrow LVIS$ , the relative gains in $AR_{100}$ and AP are equally distributed. These results indicate that object parts do not degrade the results of SOS. This is in line with SOS's pseudo annotations in Fig. 3, covering entire objects and selected object parts. + +$COCO \rightarrow UVO$ : Finally, Tab. 5 presents results for COCO $\rightarrow$ UVO. SOS still outperforms Mask R-CNN, SAM and recent UDOS [24] in $F_1$ . Outperforming SAM implies that SOS does not exploit the limited taxonomy of annotations in COCO (only 60 object classes in test), but also improves the segmentation of objects outside the COCO object classes on the exhaustively labeled UVO + +Table 6: Quality of pseudo annotation generated in GGN [49] and our SOS evaluated on the non-VOC annotations of the COCO training dataset. $\dagger$ : uses automatically annotated SA-1B dataset. + +
SystemPrec.Rec.F1
GGN3 [49]7.312.19.1
SOS3† (ours)19.026.422.1
SOS10† (ours)15.541.722.6
+ +![](images/394330a6f99d5e5240e6cc209dbeaa13e7415f348e39041847517301f8eeddb5.jpg) + +![](images/fe5e1a842cec4d04dd0b27b3f9f5b51cd31e6e815eb15bddde46e336af14f87f.jpg) + +![](images/9304cbb3b90f3cfe5e5a7ff9f135b3e7c850345a8da0b789f1e5f3bcac0973b1.jpg) +GGN + +![](images/0a44d544eb21135caae4783ce969e642135dac6ffc058018478b0abd9141cd59.jpg) +SOS +Fig. 5: Pseudo annotations generated by GGN and our SOS. + +dataset. However, the results of SOS are below LDET [46] in AP and $\mathrm{F_1}$ . We attribute this to several classes in UVO not available in ImageNet, which was used to learn DINO object prior. Hence, DINO will miss such objects in the COCO dataset during training, and SOS might miss them during testing. This could be mitigated using more diverse unlabeled images for training DINO. + +# 5.4 Quality of Pseudo Annotations + +We also investigate the quality of the pseudo annotations generated in SOS. To this end, we evaluate the pseudo annotations on the COCO training set against the annotations of non-VOC object classes, reflecting the COCO (VOC) $\rightarrow$ COCO non-VOC) setting. The results in terms of precision and recall for IoU $= 0.5$ as well as the $\mathrm{F_1}$ score are presented in Tab. 6. We use up to three and 10 pseudo annotations generated per image in SOS (SOS $_3$ and $\mathrm{SOS}_{10}$ ), and $\mathrm{GGN}_3$ [49] that generates three pseudo annotations. The results show that with only three SOS pseudo annotations, more than $25\%$ of the non-VOC objects are covered. This is substantially more than $\mathrm{GGN}_3$ ( $12.1\%$ ). With up to 10 annotations, SOS's pseudo annotations cover more than $40\%$ of the non-VOC objects. In our supplementary, we further provide class-specific results. + +To complement these results, we visualize the annotations of $\mathrm{GGN}_3$ and $\mathrm{SOS}_3$ on two sample images in Fig. 5. It is visible that $\mathrm{GGN}_3$ 's annotations cover background regions, leading to low precision in the overall results. Moreover, the pseudo annotations in the upper example do not adhere well to the airplane's boundaries. More qualitative examples are given in Fig. 1 and our supplementary. + +# 5.5 Ablation Studies + +All ablation studies follow the evaluation setup described in Sec. 4.2. More studies evaluating design choices in SOS are presented in our supplementary. + +Influence of SOS's Components: We investigate the importance of each component in SOS and present the results in Tab. 7. The baseline for this + +Table 7: Influence of added components in our SOS evaluated in the COCO (VOC) $\rightarrow$ COCO (non-VOC) setting. + +
ComponentsAPAR100F1
Mask R-CNN1.210.82.2
+ Grid, w/o pp.3.435.26.2
+ DINO, w/o pp.8.937.114.3
+ DINO, w/ pp.8.938.114.4
+ +Table 8: SOS results with various numbers of pseudo annotations in the COCO (VOC) $\rightarrow$ COCO (non-VOC) setting. + +
#Pseudo Anns.APAR100F1
38.334.513.4
58.836.214.2
108.938.114.4
208.838.114.3
+ +study is a class-agnostic Mask R-CNN trained without pseudo annotations (first row in Tab. 7). Subsequently, we add pseudo annotations based on SAM with the baseline Grid object prior as in [27] and without our post-processing in PAC (confidence-based thresholding and NMS, see Sec. 3.3). Next, we replace Grid with DINO (third row in Tab. 7), and finally add the post-processing leading to the complete SOS. The results in Tab. 7 show each step improves the $\mathrm{F}_1$ results. Mainly, introducing SAM for generating pseudo annotations (first row vs. second row) substantially improves $\mathrm{AR}_{100}$ , while adding the DINO object prior removes background annotations leading to considerably improved precision. Hence, the object prior selection in SOS is crucial for strong results. + +Number of Pseudo Annotations: In the second ablation study, we investigate the influence of the number of pseudo annotations per image on the results of SOS. To this end, we evaluate SOS with up to 3, 5, 10, and 20 pseudo annotations per image. The results in Tab. 8 show that 10 pseudo annotations are preferable. Hence, the ideal number of pseudo annotations in SOS is larger than in GGN [49], which only uses three. This indicates again the high quality of our pseudo annotations, since more than three annotations per image can cover relevant objects without adding too many noisy annotations. + +# 6 Conclusion + +In this paper, we addressed the challenging OWIS task and aimed to improve the low precision of previous methods in this open-world task. To improve the precision and the overall detection results, we generate high-quality pseudo annotations based on a prompt-guided foundation model, SAM. To focus the pseudo annotations on objects, we thoroughly investigated various object priors for prompt creation, which revealed important insights on how to prompt SAM for class-agnostic object-focused results in arbitrary tasks. As a result, our novel OWIS method SOS, which uses a self-attention-based object prior for generating pseudo annotations, outperforms recent state-of-the-art OWIS methods across challenging open-world setups in COCO, LVIS, and ADE20k datasets. SOS especially improves the precision based on the high-quality pseudo annotations. Moreover, SOS with its focus on objects also outperforms vanilla SAM that segments both object and stuff regions. + +# References + +1. Ahn, J., Cho, S., Kwak, S.: Weakly supervised learning of instance segmentation with inter-pixel relations. In: Conference on Computer Vision and Pattern Recognition (2019) +2. Alexe, B., Deselaers, T., Ferrari, V.: What is an object? In: Conference on Computer Vision and Pattern Recognition (2010) +3. Baugh, M., Batten, J., Müller, J.P., Kainz, B.: Zero-shot anomaly detection with pre-trained segmentation models. arXiv preprint arXiv:2306.09269 (2023) +4. Bommasani, R., Hudson, D.A., Adeli, E., Altman, R., Arora, S., von Arx, S., Bernstein, M.S., Bohg, J., Bosselut, A., Brunskill, E., et al.: On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258 (2021) +5. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in Neural Information Processing Systems 33 (2020) +6. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: International Conference on Computer Vision (2021) +7. Chen, T., Mai, Z., Li, R., Chao, W.l.: Segment anything model (SAM) enhanced pseudo labels for weakly supervised semantic segmentation. arXiv preprint arXiv:2305.05803 (2023) +8. Cheng, M.M., Mitra, N.J., Huang, X., Torr, P.H., Hu, S.M.: Global contrast based salient region detection. IEEE Transactions on Pattern Analysis and Machine Intelligence 37(3) (2014) +9. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: BERT: Pre-training of deep bidirectional transformers for language understanding. In: Human Language Technology Conference of the North American Chapter of the Association for Computational Linguistics (2019) +0. Dollar, P., Zitnick, C.L.: Structured forests for fast edge detection. In: International Conference on Computer Vision (2013) +1. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2020) +2. Everingham, M., Eslami, S.A., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The PASCAL visual object classes challenge: A retrospective. International Journal of Computer Vision 111 (2015) +3. Felzenszwalb, P.F., Huttenlocher, D.P.: Efficient graph-based image segmentation. International Journal of Computer Vision 59 (2004) +4. Frintrop, S., Werner, T., Martin Garcia, G.: Traditional saliency reloaded: A good old model in new shape. In: Conference on Computer Vision and Pattern Recognition (2015) +5. Gao, G., Lauri, M., Zhang, J., Frintrop, S.: Saliency-guided adaptive seeding for supervoxel segmentation. In: International Conference on Intelligent Robots and Systems (2017) +6. Girshick, R.: Fast R-CNN. In: International Conference on Computer Vision (2015) +7. Gupta, A., Dolkar, P., Girshick, R.: LVIS: A dataset for large vocabulary instance segmentation. In: Conference on Computer Vision and Pattern Recognition (2019) +8. He, C., Li, K., Zhang, Y., Xu, G., Tang, L., Zhang, Y., Guo, Z., Li, X.: Weakly-supervised concealed object segmentation with SAM-based pseudo labeling and multi-scale feature grouping. arXiv preprint arXiv:2305.11003 (2023) + +19. He, K., Gkioxari, G., Dollar, P., Girshick, R.: Mask R-CNN. In: International Conference on Computer Vision (2017) +20. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Conference on Computer Vision and Pattern Recognition (2016) +21. Hou, Q., Cheng, M.M., Hu, X., Borji, A., Tu, Z., Torr, P.H.: Deeply supervised salient object detection with short connections. In: Conference on Computer Vision and Pattern Recognition (2017) +22. Hu, H., Lan, S., Jiang, Y., Cao, Z., Sha, F.: FastMask: Segment multi-scale object candidates in one shot. In: Conference on Computer Vision and Pattern Recognition (2017) +23. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International Conference on Machine Learning (2021) +24. Kalluri, T., Wang, W., Wang, H., Chandraker, M., Torresani, L., Tran, D.: Openworld instance segmentation: Top-down learning with bottom-up supervision. arXiv preprint arXiv:2303.05503 (2023) +25. Kim, B., Yoo, Y., Rhee, C.E., Kim, J.: Beyond semantic to instance segmentation: Weakly-supervised instance segmentation via semantic knowledge transfer and self-refinement. In: Conference on Computer Vision and Pattern Recognition (2022) +26. Kim, D., Lin, T.Y., Angelova, A., Kweon, I.S., Kuo, W.: Learning open-world object proposals without learning to classify. IEEE Robotics and Automation Letters 7(2) (2022) +27. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: International Conference on Computer Vision (2023) +28. Klein, D.A., Frintrop, S.: Center-surround divergence of feature statistics for salient object detection. In: International Conference on Computer Vision (2011) +29. Li, J., Li, D., Savarese, S., Hoi, S.: BLIP-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023) +30. Lin, T.Y., Dollar, P., Girshick, R., He, K., Hariharan, B., Belongie, S.: Feature pyramid networks for object detection. In: Conference on Computer Vision and Pattern Recognition (2017) +31. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft COCO: Common objects in context. In: European Conference on Computer Vision (2014) +32. Linardos, A., Kümmerer, M., Press, O., Bethge, M.: DeepGaze IIE: Calibrated prediction in and out-of-domain for state-of-the-art saliency modeling. In: International Conference on Computer Vision (2021) +33. Liu, N., Han, J., Yang, M.H.: PiCANet: Learning pixel-wise contextual attention for saliency detection. In: Conference on Computer Vision and Pattern Recognition (2018) +34. Liu, T., Yuan, Z., Sun, J., Wang, J., Zheng, N., Tang, X., Shum, H.Y.: Learning to detect a salient object. IEEE Transactions on Pattern analysis and machine intelligence 33(2) (2010) +35. Liu, Y., Wu, Y.H., Wen, P., Shi, Y., Qiu, Y., Cheng, M.M.: Leveraging instance-, image-and dataset-level information for weakly supervised instance segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(3) (2020) +36. Ma, J., Ming, A., Huang, Z., Wang, X., Zhou, Y.: Object-level proposals. In: International Conference on Computer Vision (2017) + +37. Martin García, G., Potapova, E., Werner, T., Zillich, M., Vincze, M., Frintrop, S.: Saliency-based object discovery on RGB-D data with a late-fusion approach. In: International Conference on Robotics and Automation (2015) +38. Mazurowski, M.A., Dong, H., Gu, H., Yang, J., Konz, N., Zhang, Y.: Segment anything model for medical image analysis: an experimental study. Medical Image Analysis 89 (2023) +39. Perazzi, F., Krahenbuhl, P., Pritch, Y., Hornung, A.: Saliency filters: Contrast based filtering for salient region detection. In: Conference on Computer Vision and Pattern Recognition (2012) +40. Pinheiro, P.O., Collobert, R., Dollár, P.: Learning to segment object candidates. Advances in Neural Information Processing Systems 28 (2015) +41. Pinheiro, P.O., Lin, T.Y., Collobert, R., Dollár, P.: Learning to refine object segments. In: European Conference on Computer Vision (2016) +42. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning (2021) +43. Rahtu, E., Kannala, J., Blaschko, M.: Learning a category independent object detection cascade. In: International Conference on Computer Vision (2011) +44. Ronneberger, O., Fischer, P., Brox, T.: U-Net: Convolutional networks for biomedical image segmentation. In: International Conference on Medical Image Computing and Computer-Assisted Intervention (2015) +45. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al.: Imagenet large scale visual recognition challenge. International Journal of Computer Vision 115 (2015) +46. Saito, K., Hu, P., Darrell, T., Saenko, K.: Learning to detect every thing in an open world. In: European Conference on Computer Vision (2022) +47. Shi, Z., Sun, Y., Zhang, M.: Training-free object counting with prompts. In: Winter Conference on Applications of Computer Vision (2024) +48. Wang, C., Wang, G., Zhang, Q., Guo, P., Liu, W., Wang, X.: OpenInst: A simple query-based method for open-world instance segmentation. arXiv preprint arXiv:2303.15859 (2023) +49. Wang, W., Feiszli, M., Wang, H., Malik, J., Tran, D.: Open-world instance segmentation: Exploiting pseudo ground truth from learned pairwise affinity. In: Conference on Computer Vision and Pattern Recognition (2022) +50. Wang, W., Feiszli, M., Wang, H., Tran, D.: Unidentified video objects: A benchmark for dense, open-world segmentation. In: International Conference on Computer Vision (2021) +51. Wei, Z., Chen, P., Yu, X., Li, G., Jiao, J., Han, Z.: Semantic-aware SAM for point-prompted instance segmentation. arXiv preprint arXiv:2312.15895 (2023) +52. Wilms, C., Frintrop, S.: Edge adaptive seeding for superpixel segmentation. In: German Conference on Pattern Recognition (2017) +53. Wilms, C., Frintrop, S.: AttentionMask: Attentive, efficient object proposal generation focusing on small objects. In: Asian Conference on Computer Vision (2019) +54. Wu, J., Jiang, Y., Yan, B., Lu, H., Yuan, Z., Luo, P.: Exploring transformers for open-world instance segmentation. In: International Conference on Computer Vision (2023) +55. Xie, C., Xiang, Y., Mousavian, A., Fox, D.: Unseen object instance segmentation for robotic environments. IEEE Transactions on Robotics 37(5) (2021) + +56. Xue, X., Yu, D., Liu, L., Liu, Y., Li, Y., Yuan, Z., Song, P., Shou, M.Z.: Single-stage open-world instance segmentation with cross-task consistency regularization. arXiv preprint arXiv:2208.09023 (2022) +57. Yang, X., Gong, X.: Foundation model assisted weakly supervised semantic segmentation. In: Winter Conference on Applications of Computer Vision (2024) +58. Yang, Y., Zhou, Z., Wu, J., Wang, Y., Xiong, R.: Class semantics modulation for open-set instance segmentation. IEEE Robotics and Automation Letters (2024) +59. Zhao, J.X., Liu, J.J., Fan, D.P., Cao, Y., Yang, J., Cheng, M.M.: EGNet: Edge guidance network for salient object detection. In: International Conference on Computer Vision (2019) +60. Zhou, B., Khosla, A., Lapedriza, A., Oliva, A., Torralba, A.: Learning deep features for discriminative localization. In: Conference on Computer Vision and Pattern Recognition (2016) +61. Zhou, B., Zhao, H., Puig, X., Xiao, T., Fidler, S., Barriuso, A., Torralba, A.: Semantic understanding of scenes through the ADE20K dataset. International Journal of Computer Vision 127 (2019) +62. Zhu, M., Li, H., Chen, H., Fan, C., Mao, W., Jing, C., Liu, Y., Shen, C.: SegPrompt: Boosting open-world segmentation via category-level prompt learning. In: International Conference on Computer Vision (2023) +63. Zitnick, C.L., Dollár, P.: Edge boxes: Locating object proposals from edges. In: European Conference on Computer Vision (2014) \ No newline at end of file diff --git a/2024/SOS_ Segment Object System for Open-World Instance Segmentation With Object Priors/images.zip b/2024/SOS_ Segment Object System for Open-World Instance Segmentation With Object Priors/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f0cc525bb0542711b0c68ae0ebdb6a04bba0830b --- /dev/null +++ b/2024/SOS_ Segment Object System for Open-World Instance Segmentation With Object Priors/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e4a24e44c21624dce8b65240d4be56e67cbae60eec45b3b1eebea9f594820c06 +size 489745 diff --git a/2024/SOS_ Segment Object System for Open-World Instance Segmentation With Object Priors/layout.json b/2024/SOS_ Segment Object System for Open-World Instance Segmentation With Object Priors/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a8e56445ddc44e6ab0e31722c031bef9a0786083 --- /dev/null +++ b/2024/SOS_ Segment Object System for Open-World Instance Segmentation With Object Priors/layout.json @@ -0,0 +1,9687 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 146, + 111, + 467, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 111, + 467, + 148 + ], + "spans": [ + { + "bbox": [ + 146, + 111, + 467, + 148 + ], + "type": "text", + "content": "SOS: Segment Object System for Open-World Instance Segmentation With Object Priors" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 135, + 167, + 478, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 167, + 478, + 194 + ], + "spans": [ + { + "bbox": [ + 135, + 167, + 478, + 194 + ], + "type": "text", + "content": "Christian Wilms1, Tim Rolff1,2, Maris Hillemann1, Robert Johanson1, and Simone Frinthrop1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 156, + 201, + 457, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 201, + 457, + 236 + ], + "spans": [ + { + "bbox": [ + 156, + 201, + 457, + 236 + ], + "type": "text", + "content": "1 Computer Vision Group, University of Hamburg, Germany \n2 Human-Computer Interaction Group, University of Hamburg, Germany {firstname_lastname}@uni-hamburg.de" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 159, + 261, + 455, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 261, + 455, + 447 + ], + "spans": [ + { + "bbox": [ + 159, + 261, + 455, + 447 + ], + "type": "text", + "content": "Abstract. We propose an approach for Open-World Instance Segmentation (OWIS), a task that aims to segment arbitrary unknown objects in images by generalizing from a limited set of annotated object classes during training. Our Segment Object System (SOS) explicitly addresses the generalization ability and the low precision of state-of-the-art systems, which often generate background detections. To this end, we generate high-quality pseudo annotations based on the foundation model SAM [27]. We thoroughly study various object priors to generate prompts for SAM, explicitly focusing the foundation model on objects. The strongest object priors were obtained by self-attention maps from self-supervised Vision Transformers, which we utilize for prompting SAM. Finally, the post-processed segments from SAM are used as pseudo annotations to train a standard instance segmentation system. Our approach shows strong generalization capabilities on COCO, LVIS, and ADE20k datasets and improves on the precision by up to " + }, + { + "bbox": [ + 159, + 261, + 455, + 447 + ], + "type": "inline_equation", + "content": "81.6\\%" + }, + { + "bbox": [ + 159, + 261, + 455, + 447 + ], + "type": "text", + "content": " compared to the state-of-the-art. Source code is available at: https://github.com/chwilms/SOS" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 457, + 453, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 457, + 453, + 480 + ], + "spans": [ + { + "bbox": [ + 159, + 457, + 453, + 480 + ], + "type": "text", + "content": "Keywords: Open-world Instance Segmentation " + }, + { + "bbox": [ + 159, + 457, + 453, + 480 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 457, + 453, + 480 + ], + "type": "text", + "content": " Object Localization " + }, + { + "bbox": [ + 159, + 457, + 453, + 480 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 457, + 453, + 480 + ], + "type": "text", + "content": " Prompting" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 498, + 230, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 230, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 230, + 511 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 521, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 641 + ], + "type": "text", + "content": "Open-World Instance Segmentation (OWIS) is the task of segmenting all object instances in an image by learning from a limited set of known object classes [26, 49, 50]. In contrast to instance segmentation, OWIS is not limited by the closed-world assumption, which assumes that all object classes are known in advance. Since OWIS methods aim to detect not only learned but also unknown object classes, they return class-agnostic detections. This is of particular interest in real-world scenarios with previously unknown object classes (e.g., [55]) and challenges the systems' generalization capabilities. An example of this is shown in the first row of Fig. 1, where annotations for classes such as surfboard or tennis racket do not exist during training, but the objects should be detected during testing." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 642, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 642, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 642, + 481, + 666 + ], + "type": "text", + "content": "Previous OWIS methods mostly generate pseudo annotations for unannotated objects [24, 46, 49, 56] or replace the foreground-background classification" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 112, + 482, + 245 + ], + "blocks": [ + { + "bbox": [ + 132, + 112, + 482, + 245 + ], + "lines": [ + { + "bbox": [ + 132, + 112, + 482, + 245 + ], + "spans": [ + { + "bbox": [ + 132, + 112, + 482, + 245 + ], + "type": "image", + "image_path": "b90e68b72143383adf80f1e1ba9fcc4f24d0c46409d00badc90355a6e7ca539d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 254, + 482, + 344 + ], + "lines": [ + { + "bbox": [ + 130, + 254, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 254, + 482, + 344 + ], + "type": "text", + "content": "Fig. 1: Comparison of (pseudo) annotations (left) used by Mask R-CNN [19], GGN [49], and our SOS, when only original annotations of VOC object classes are given. While Mask R-CNN only uses original annotations without object classes such as tennis racket or surfboard (red arrows), SOS generates pseudo annotations covering those classes (green arrows). GGN generates noisy pseudo annotations including background areas. As a result, only SOS constantly detects these objects not annotated in training (green vs. red arrows on the right). Filled masks denote annotations (left) or detected objects (right), while red frames indicate missed objects." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 372, + 482, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 372, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 372, + 482, + 445 + ], + "type": "text", + "content": "of possible objects in instance segmentation systems with a learning target focusing on the localization quality (e.g., intersection over union) [26,48,54]. While OWIS methods generalize better to unseen object classes than standard instance segmentation systems, they still exhibit low precision [24,26,46,49,54]. One reason for this is the use of noisy pseudo annotations covering background areas as visible in the left part of the second row in Fig. 1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 447, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 447, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 447, + 482, + 567 + ], + "type": "text", + "content": "Recently, foundation models [4] emerged as a promising technique [5,9,23,29, 42]. They are trained on large datasets with a surrogate task and applied to other tasks in a zero- or few-shot manner using prompting. The first foundation model for image segmentation, the Segment Anything Model (SAM), was proposed by [27]. Trained on a large automatically annotated dataset, SAM generates segments for object or stuff regions based on point, box, mask, or text prompts without further training. Despite not only segmenting objects but also stuff regions, given well-designed prompts, SAM is able to generate high-quality object segments. Note that vanilla SAM does not address object segmentation tasks like OWIS itself due to the aforementioned lacking focus on objects." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we propose the Segment Object System (SOS), a novel OWIS method utilizing high-quality pseudo annotations to improve the generalization capability of a standard instance segmentation system (see lower row in Fig. 1). For generating pseudo annotations, we apply the foundation model SAM with prompts derived from an object prior to roughly localize objects of arbitrary classes and to limit segmentations of stuff regions by SAM. To ensure the high quality of our pseudo annotations based on SAM, improving the overall precision, we thoroughly study various hand-crafted and learned object priors for an" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 100 + ], + "type": "text", + "content": "C. Wilms et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 247 + ], + "type": "text", + "content": "object-focused application of SAM. This is relevant beyond the scope of OWIS. Using our study findings, SOS utilizes self-attention maps from self-supervised Vision Transformers (ViTs) [11] as object prior for prompting SAM. Finally, SOS filters low-quality pseudo annotations, combines the pseudo annotations with the original annotations of the known classes, and trains a standard instance segmentation system with these mixed annotations. Our extensive evaluation shows the strong generalization abilities of SOS, considerably outperforming all previous state-of-the-art systems across COCO [31], LVIS [17], and ADE20k [61] datasets. Most notably, the results improve by up to " + }, + { + "bbox": [ + 130, + 116, + 479, + 247 + ], + "type": "inline_equation", + "content": "81.6\\%" + }, + { + "bbox": [ + 130, + 116, + 479, + 247 + ], + "type": "text", + "content": " in terms of precision over the state-of-the-art due to the high-quality pseudo annotations. We also show that SOS is better suited for OWIS than directly applying SAM." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 146, + 248, + 324, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 248, + 324, + 258 + ], + "spans": [ + { + "bbox": [ + 146, + 248, + 324, + 258 + ], + "type": "text", + "content": "Overall, our contributions are threefold:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 137, + 270, + 480, + 365 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 138, + 270, + 479, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 270, + 479, + 304 + ], + "spans": [ + { + "bbox": [ + 138, + 270, + 479, + 304 + ], + "type": "text", + "content": "- We propose SOS, a novel OWIS method based on a learned object prior, prompt-based pseudo annotations, and an arbitrary instance segmentation system." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 306, + 480, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 306, + 480, + 330 + ], + "spans": [ + { + "bbox": [ + 137, + 306, + 480, + 330 + ], + "type": "text", + "content": "- We thoroughly study various object priors for focusing SAM on objects, leading to high-quality object segments for pseudo annotations and beyond." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 331, + 480, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 331, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 137, + 331, + 480, + 365 + ], + "type": "text", + "content": "- Our extensive evaluation shows that the high-quality pseudo annotations in SOS lead to high precision and strong overall results, clearly outperforming other methods." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 387, + 237, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 387, + 237, + 399 + ], + "spans": [ + { + "bbox": [ + 132, + 387, + 237, + 399 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 415, + 337, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 415, + 337, + 427 + ], + "spans": [ + { + "bbox": [ + 131, + 415, + 337, + 427 + ], + "type": "text", + "content": "2.1 Open-world Instance Segmentation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 437, + 479, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 479, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 479, + 509 + ], + "type": "text", + "content": "To address OWIS, literature mainly offers two streams. First, systems replace the hard classification in instance segmentation systems with a localization score. This avoids classifying unseen objects as background and improves generalization ability [26]. In OLN [26], centerness and Intersection over Union (IoU) are learned as localization scores in a Mask R-CNN system [19]. SWORD [54] and OpenInst [48] follow the same idea for query-based systems." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 510, + 480, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 480, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 480, + 593 + ], + "type": "text", + "content": "The second line of works [24, 46, 49, 56] addresses OWIS by augmenting the annotations from known classes with pseudo annotations aiming to cover objects of unknown classes. To generate pseudo annotations, GGN [49] learns pixel affinities from known classes to generate segments, which are grouped. Similarly, UDOS [24] groups object proposals of object parts to create pseudo annotations, while LDET [46] applies copy-paste-augmentation. Recently, a new stream emerged [58, 62] explicitly utilizing class labels from known classes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 594, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 480, + 665 + ], + "type": "text", + "content": "We follow the second stream and create pseudo annotations within SOS, since it offers the most flexibility w.r.t. the base instance segmentation system. Different from existing approaches that generate noisy pseudo annotations (see GGN in Fig. 1), we investigate and develop object priors to effectively use recent foundation models to create high-quality object-focused pseudo annotations, leading to high recall and precision." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 325, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 325, + 91, + 447, + 102 + ], + "type": "text", + "content": "SOS: Segment Object System" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 334, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 334, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 334, + 129 + ], + "type": "text", + "content": "2.2 Class-agnostic Object Localization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 481, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 481, + 196 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 481, + 196 + ], + "type": "text", + "content": "Localizing objects in images is related to several computer vision tasks. For instance, in traditional object proposal generation [2], cues like saliency [2, 37] or contour information [2, 10, 36, 43] were used to localize class-agnostic object candidates for object detection. Since the advent of deep learning, the models are learned from data [16, 22, 40, 41, 53]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 198, + 481, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 198, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 198, + 481, + 304 + ], + "type": "text", + "content": "Focusing on arbitrary salient objects, known as salient object detection, several approaches use the contrast of hand-crafted features such as color [8,28,34, 39] or edge information [28], as well as CNN-based approaches [21, 33, 59]. In a different task, several authors investigated the information stored in learned classifiers or self-supervised feature extractors based on CNNs or ViTs to locate objects. Notable avenues include Class Activation Maps (CAMs) [60] that highlight discriminative object or image parts in CNN-based classifiers and the self-attention maps in the self-supervised DINO feature extractor [6] based on ViTs, which contain information on the scene layout indicating object locations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 304, + 481, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 481, + 342 + ], + "type": "text", + "content": "In our study on object priors to focus SAM on objects, we investigate several of these object localization cues. In contrast to these approaches, we follow a different goal, focusing prompt-based segmentations on objects for OWIS." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 359, + 266, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 359, + 266, + 371 + ], + "spans": [ + { + "bbox": [ + 132, + 359, + 266, + 371 + ], + "type": "text", + "content": "2.3 Applications of SAM" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 380, + 481, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 481, + 452 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 481, + 452 + ], + "type": "text", + "content": "SAM [27] is used in several segmentation tasks [3,38,47]. Similar to us, [7] and [57] use SAM to generate pseudo annotation masks for weakly-supervised semantic segmentation based on image-level supervision. Similarly, [18] generate pseudo annotation masks for concealed object segmentation based on scribble annotations. Finally, [51] generate pseudo annotation masks for weakly-supervised instance segmentation in a multiple instance learning framework." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 452, + 481, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 452, + 481, + 512 + ], + "spans": [ + { + "bbox": [ + 130, + 452, + 481, + 512 + ], + "type": "text", + "content": "Different from the weakly-supervised approaches utilizing SAM, we do not rely on supervision. Hence, a key novelty of our work lies in investigating object priors to focus SAM on segmenting arbitrary objects in the absence of supervision for all object classes. Moreover, we are the first to apply SAM-based pseudo annotations in OWIS." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 531, + 358, + 545 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 531, + 358, + 545 + ], + "spans": [ + { + "bbox": [ + 132, + 531, + 358, + 545 + ], + "type": "text", + "content": "3 Segment Object System for OWIS" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 557, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 481, + 666 + ], + "type": "text", + "content": "To address OWIS, we propose our Segment Object System (SOS) that generates high-quality pseudo annotations to train a standard instance segmentation system. As visible from Fig. 2, SOS consists of three main blocks. The first block, the Object Localization Module (OLM), described in Sec. 3.2 (yellow area in Fig. 2), aims to roughly localize objects by sampling locations from an object prior that reflects the object probability per image coordinate. The output of OLM is a set of likely object locations, that is used in the Pseudo Annotation Creator (PAC), described in Sec. 3.3 (green area in Fig. 2), to prompt the pre-trained SAM [27], which we briefly revisit in Sec. 3.1. From the segments" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "text", + "content": "C. Wilms et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 113, + 482, + 340 + ], + "blocks": [ + { + "bbox": [ + 132, + 113, + 482, + 340 + ], + "lines": [ + { + "bbox": [ + 132, + 113, + 482, + 340 + ], + "spans": [ + { + "bbox": [ + 132, + 113, + 482, + 340 + ], + "type": "image", + "image_path": "6f7a464a696853056538da30d03ea8fd1653e04bd52c82b1979dab2ef6b1bbd2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 348, + 482, + 426 + ], + "lines": [ + { + "bbox": [ + 130, + 348, + 482, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 348, + 482, + 426 + ], + "type": "text", + "content": "Fig. 2: Overview of our Segment Object System (SOS) for OWIS consisting of three blocks. First, the input image is processed in our Object Localization Module (OLM, yellow area) to create object-focused point prompts roughly localizing objects. Second, our Pseudo Annotations Creator (PAC, green area) generates segments based on the previously generated prompts using SAM [27], and further processes them, leading to a final set of merged original and pseudo annotations. Finally, the merged annotations are used to train an instance segmentation system (blue area)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 449, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 510 + ], + "type": "text", + "content": "generated from likely object locations, the PAC generates pseudo annotations by filtering low-quality segments and removing near-duplicates. Finally, the pseudo annotations are merged with the original annotations, and a standard instance segmentation system is trained with the merged annotations (see Sec. 3.4, blue area in Fig. 2). During testing, only this instance segmentation system is used." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 525, + 330, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 525, + 330, + 539 + ], + "spans": [ + { + "bbox": [ + 131, + 525, + 330, + 539 + ], + "type": "text", + "content": "3.1 Revisit Segment Anything Model" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Since SAM [27] is integral to SOS, we will briefly review the model. The idea of SAM is to utilize prompts such as point coordinates to generate segments of image parts like objects, object parts, or stuff regions localized by these prompts. Hence, SAM does not exclusively segment objects unless explicitly instructed by object-focused prompts, as proposed in this paper. SAM is based on a ViT [11] image encoder creating an image embedding, while the embeddings for point prompts are generated based on positional encodings. Both embeddings are processed in a mask decoder utilizing a transformer decoder block and a mask prediction head to create the final segment including a confidence score. Note that SAM is class-agnostic and does not classify the segments." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 326, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 326, + 91, + 447, + 102 + ], + "type": "text", + "content": "SOS: Segment Object System" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "For training SAM, [27] propose the new large-scale SA-1B dataset that is annotated with a data engine using SAM itself. In the pre-stage, SAM is trained with public segmentation datasets to aid the first stage of the data engine, where annotators use SAM as an interactive segmentation tool to create initial segmentations. Subsequently, SAM is retrained with the newly annotated data, fully avoiding data leakage w.r.t. public segmentation datasets. In subsequent stages, manual annotations were provided for segments missed by SAM. At last, annotations were created fully automatically by SAM. See [27] for more details." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 228, + 302, + 241 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 228, + 302, + 241 + ], + "spans": [ + { + "bbox": [ + 131, + 228, + 302, + 241 + ], + "type": "text", + "content": "3.2 Object Localization Module" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 247, + 482, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 415 + ], + "type": "text", + "content": "The first block of SOS (yellow area in Fig. 2) is the Object Localization Module (OLM). OLM roughly localizes all objects in the input image, and creates object-focused point prompts. As the first step, OLM generates an object prior from the input image. The object prior is a probability mass function constrained to the image plane, although other formulations are possible (see Sec. 4), and indicates the probability of an image coordinate being part of any object. Hence, the object prior is class-agnostic and not limited to known object classes from the training set in OWIS. In our study in Sec. 4, we analyze various object priors. For final SOS, we utilize the six self-attention maps from the self-attention heads of a ViT's final layer, pre-trained in the self-supervised DINO framework [6] on ImageNet [45]. See Sec. 4 for justifications. OLM aggregates the six self-attention maps by elementwise max yielding one map highlighting all relevant scene elements. Subsequently, OLM converts the map to an object prior, by rescaling such that the minimum equals 0 and the sum over the object prior is 1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 415, + 483, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 483, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 483, + 536 + ], + "type": "text", + "content": "Next, OLM randomly samples " + }, + { + "bbox": [ + 130, + 415, + 483, + 536 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 415, + 483, + 536 + ], + "type": "text", + "content": " (here: " + }, + { + "bbox": [ + 130, + 415, + 483, + 536 + ], + "type": "inline_equation", + "content": "S = 50" + }, + { + "bbox": [ + 130, + 415, + 483, + 536 + ], + "type": "text", + "content": ") image coordinates from the object prior, given the per-coordinate values of the object prior as probabilities. To diversify the samples across multiple objects, we prune parts of the object prior around the sampled coordinate. Specifically, given a coordinate " + }, + { + "bbox": [ + 130, + 415, + 483, + 536 + ], + "type": "inline_equation", + "content": "(x_s, y_s)" + }, + { + "bbox": [ + 130, + 415, + 483, + 536 + ], + "type": "text", + "content": ", we set the object prior to zero for all coordinates " + }, + { + "bbox": [ + 130, + 415, + 483, + 536 + ], + "type": "inline_equation", + "content": "\\{(x, y) \\mid |x - x_s| \\leq N \\land |y - y_s| \\leq N\\}" + }, + { + "bbox": [ + 130, + 415, + 483, + 536 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 130, + 415, + 483, + 536 + ], + "type": "inline_equation", + "content": "N = 20" + }, + { + "bbox": [ + 130, + 415, + 483, + 536 + ], + "type": "text", + "content": ". To transform the object prior back to a probability mass function, we again apply the rescaling as described above. Subsequently, OLM iteratively applies this sampling until a set of " + }, + { + "bbox": [ + 130, + 415, + 483, + 536 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 415, + 483, + 536 + ], + "type": "text", + "content": " coordinates are extracted, representing object-focused point prompts. This set is the output of the OLM and will be used for prompting in the Pseudo Annotation Creator." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 551, + 307, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 551, + 307, + 562 + ], + "spans": [ + { + "bbox": [ + 131, + 551, + 307, + 562 + ], + "type": "text", + "content": "3.3 Pseudo Annotations Creator" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "The Pseudo Annotations Creator (PAC, green area in Fig. 2) is the second block of SOS. It generates pseudo annotations utilizing SAM, given the object-focused point prompts provided by OLM. These object-focused point prompts lifts SAM from a system that segments anything to a system that segments objects. As the first step, PAC prompts a pre-trained SAM with the " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " point prompts from OLM. To handle ambiguous point prompts that may indicate multiple objects or object parts, we follow [27] and allow SAM to generate three segments per prompt. Each generated segment is a potential pseudo annotation. The resulting set of" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "text", + "content": "C. Wilms et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "content": "class-agnostic segments is noisy, with segments covering objects and background, or covering the same image area. Therefore, PAC first utilizes SAM's confidence score per segment and removes segments with a confidence below " + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{conf}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "content": ". Second, PAC removes near-duplicates by non-maximum suppression with an IoU threshold " + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{NMS}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "content": ". We set " + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{conf}} = 0.9" + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{NMS}} = 0.95" + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "content": " for SOS." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 483, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 483, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 483, + 319 + ], + "type": "text", + "content": "Given this set of pseudo annotations, PAC merges the set with the original annotations representing known classes of the training dataset to combine the knowledge of the human annotators and the knowledge of our object-focused SAM. Since we only want to keep pseudo annotations for objects not covered by the original annotations, we suppress pseudo annotations that have a high IoU (" + }, + { + "bbox": [ + 130, + 176, + 483, + 319 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{NMS}}" + }, + { + "bbox": [ + 130, + 176, + 483, + 319 + ], + "type": "text", + "content": ") with at least one original annotation. Moreover, we limit the number of pseudo annotations to " + }, + { + "bbox": [ + 130, + 176, + 483, + 319 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 130, + 176, + 483, + 319 + ], + "type": "text", + "content": " in order to balance original and pseudo annotations. In our final SOS, we set " + }, + { + "bbox": [ + 130, + 176, + 483, + 319 + ], + "type": "inline_equation", + "content": "P = 10" + }, + { + "bbox": [ + 130, + 176, + 483, + 319 + ], + "type": "text", + "content": ". However, across the COCO training dataset, only 7.8 pseudo annotations are added on average. The other pseudo annotations were suppressed in previous steps. Overall, the output of PAC is a mixed set of annotations, containing all original annotations augmented with high-quality pseudo annotations covering objects of unknown classes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 335, + 276, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 335, + 276, + 346 + ], + "spans": [ + { + "bbox": [ + 131, + 335, + 276, + 346 + ], + "type": "text", + "content": "3.4 Instance Segmentation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 353, + 482, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 353, + 482, + 425 + ], + "spans": [ + { + "bbox": [ + 130, + 353, + 482, + 425 + ], + "type": "text", + "content": "In the last block of SOS, we take the merged annotations generated by PAC and train a standard instance segmentation system. While this system can be arbitrary and is generally replaceable, we choose Mask R-CNN [19] with a ResNet-50+FPN backbone [30] trained in a class-agnostic setting, following recent OWIS methods [24, 26, 46, 49]. Overall, this leads to an OWIS method based on high-quality pseudo annotations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 441, + 287, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 441, + 287, + 455 + ], + "spans": [ + { + "bbox": [ + 131, + 441, + 287, + 455 + ], + "type": "text", + "content": "4 Object Priors for SOS" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 464, + 482, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 464, + 482, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 464, + 482, + 525 + ], + "type": "text", + "content": "As described in Sec. 3.2, SOS uses an object prior for rough object localization and subsequent point prompt generation for PAC. This section thoroughly investigates the performance of various object priors derived from previous class-agnostic object localization works in SOS. For details on the object priors beyond the subsequent descriptions, we refer to our supplementary." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 540, + 230, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 540, + 230, + 552 + ], + "spans": [ + { + "bbox": [ + 131, + 540, + 230, + 552 + ], + "type": "text", + "content": "4.1 Object Priors" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "Baselines: We introduce three baseline object priors that ignore the image content. First, Grid uses SAM with a regular " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " grid of points, following [27]. This directly leads to prompts without the sampling described in Sec. 3.2. Note that the Grid baseline will lead to segments for both object and stuff regions. Second, we use the spatial distribution of object centroids from the known classes across the training dataset as object prior (Dist). We assume that this generalizes well since it ignores class-specific object features. Finally, we utilize the centroids of the training set's objects from unknown classes as optimal prompts (" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "GT" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "), representing an upper bound." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 325, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 325, + 91, + 447, + 102 + ], + "type": "text", + "content": "SOS: Segment Object System" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": "Superpixels: Inspired by [15, 52], we utilize the centroids of superpixels generated with an adaptive superpixel segmentation method as object prior (" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "Spx" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": "). The intuition is that large uniform areas not containing objects should be covered by few superpixels and vice versa. Hence, the density of superpixels is a surrogate for the density of objects. We use the well-known FH superpixels [13] since they adapt their density to the image content following our assumption. Note that every superpixel centroid is a point prompt." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 209, + 483, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 209, + 483, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 209, + 483, + 258 + ], + "type": "text", + "content": "Contour Density: Since objects are defined by their outer contours [63], we use contour density as an object prior (Contour). Similar to [52], we assume that in areas of high contour density, several objects are located and vice versa. As a surrogate for the contour density, we use edge density based on [10]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 266, + 482, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 266, + 482, + 328 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 482, + 328 + ], + "type": "text", + "content": "Saliency: Saliency is used in object proposal generation to localize objects [2, 37] as they stick out from their surrounding. Here, we evaluate two saliency methods as object priors. First, we apply the traditional approach VOCUS2 [14] based on color contrast and used by [37] (VOCUS2). Second, we use pre-trained DeepGaze IIE [32] saliency maps learned from eye fixation data (DeepGaze)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 337, + 482, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 337, + 482, + 398 + ], + "spans": [ + { + "bbox": [ + 130, + 337, + 482, + 398 + ], + "type": "text", + "content": "Class Activation Maps: CAMs [60] indicate discriminative image regions for CNN-based classifiers. Since they only need image-level supervision, they are frequently applied in weakly-supervised tasks to locate objects [1, 25, 35]. We use the CAM of the predicted class in a ResNet-50 classifier [20] pre-trained on ImageNet [45] as object prior (CAM)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 406, + 482, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 406, + 482, + 455 + ], + "spans": [ + { + "bbox": [ + 130, + 406, + 482, + 455 + ], + "type": "text", + "content": "Self-attention: Recently, [6] have shown that self-attention maps learned inside self-supervised ViTs encode the scene layout, including object locations. Therefore, we explore the self-attention maps from the final layer of a ViT-S backbone trained in the self-supervised DINO framework [6] as an object prior (DINO)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 464, + 482, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 464, + 482, + 514 + ], + "spans": [ + { + "bbox": [ + 130, + 464, + 482, + 514 + ], + "type": "text", + "content": "Learned Object Locations: Finally, we learn object locations based on the known classes of a dataset using a U-Net [44] in a binary segmentation task (object vs. no object) as object prior (" + }, + { + "bbox": [ + 130, + 464, + 482, + 514 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 130, + 464, + 482, + 514 + ], + "type": "text", + "content": "-Net). This assumes that, on the level of individual pixels, the model generalizes from known to unknown classes [49,56]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 528, + 313, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 528, + 313, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 528, + 313, + 540 + ], + "type": "text", + "content": "4.2 Study Setup for Object Priors" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "To assess each object prior, we train SOS with the respective prior on the COCO training set using only annotations from the 20 PASCAL VOC [12] classes. We evaluate on COCO's validation set with annotations from the remaining 60 classes in COCO, following standard COCO (VOC) " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " COCO (non-VOC) cross-category evaluation in OWIS [26, 46, 49]. Note that we use the default class-agnostic Mask R-CNN in SOS but reduce the training schedule to only a quarter of steps for faster training. We evaluate the results of SOS given each object priors and report Average Recall (AR) for 100 detections and Average Precision (AP) as previous OWIS research [46, 49]. We also report " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "F_{1}" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " score, the harmonic mean between AR and AP, yielding a single number for comparison." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "text", + "content": "C. Wilms et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 116, + 271, + 203 + ], + "blocks": [ + { + "bbox": [ + 143, + 116, + 271, + 203 + ], + "lines": [ + { + "bbox": [ + 143, + 116, + 271, + 203 + ], + "spans": [ + { + "bbox": [ + 143, + 116, + 271, + 203 + ], + "type": "image", + "image_path": "98c534e1f46cf8af01435b9b48398d2a912be5cb6c56f5d701d91c242fcf430b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 181, + 206, + 233, + 217 + ], + "lines": [ + { + "bbox": [ + 181, + 206, + 233, + 217 + ], + "spans": [ + { + "bbox": [ + 181, + 206, + 233, + 217 + ], + "type": "text", + "content": "Input image" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 272, + 114, + 338, + 159 + ], + "blocks": [ + { + "bbox": [ + 272, + 114, + 338, + 159 + ], + "lines": [ + { + "bbox": [ + 272, + 114, + 338, + 159 + ], + "spans": [ + { + "bbox": [ + 272, + 114, + 338, + 159 + ], + "type": "image", + "image_path": "003002287514d3fb661c3340c460d61ece1d66380508a497953a7041dc9c0ebe.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 272, + 159, + 338, + 204 + ], + "blocks": [ + { + "bbox": [ + 272, + 159, + 338, + 204 + ], + "lines": [ + { + "bbox": [ + 272, + 159, + 338, + 204 + ], + "spans": [ + { + "bbox": [ + 272, + 159, + 338, + 204 + ], + "type": "image", + "image_path": "45e5bc61bd5b28ddc0bdbad119f694a1ab734f9401f8e33570f704fc9f26f2b4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 287, + 206, + 325, + 217 + ], + "lines": [ + { + "bbox": [ + 287, + 206, + 325, + 217 + ], + "spans": [ + { + "bbox": [ + 287, + 206, + 325, + 217 + ], + "type": "text", + "content": "VOCUS2" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 227, + 480, + 239 + ], + "lines": [ + { + "bbox": [ + 131, + 227, + 480, + 239 + ], + "spans": [ + { + "bbox": [ + 131, + 227, + 480, + 239 + ], + "type": "text", + "content": "Fig. 3: Object priors VOCUS2, U-Net, and DINO with resulting pseudo annotations." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 339, + 114, + 405, + 159 + ], + "blocks": [ + { + "bbox": [ + 339, + 114, + 405, + 159 + ], + "lines": [ + { + "bbox": [ + 339, + 114, + 405, + 159 + ], + "spans": [ + { + "bbox": [ + 339, + 114, + 405, + 159 + ], + "type": "image", + "image_path": "867b8ce8c01c3d142b7ac46172c54b92a51ac566435015d9e7a2a0e8386ed0fe.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 339, + 159, + 405, + 204 + ], + "blocks": [ + { + "bbox": [ + 339, + 159, + 405, + 204 + ], + "lines": [ + { + "bbox": [ + 339, + 159, + 405, + 204 + ], + "spans": [ + { + "bbox": [ + 339, + 159, + 405, + 204 + ], + "type": "image", + "image_path": "e18673af3e9f66db67f8d6874d2cbc02569e3bc22acb0b15c3d3538437de1770.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 360, + 206, + 386, + 217 + ], + "lines": [ + { + "bbox": [ + 360, + 206, + 386, + 217 + ], + "spans": [ + { + "bbox": [ + 360, + 206, + 386, + 217 + ], + "type": "text", + "content": "U-Net" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 406, + 114, + 471, + 159 + ], + "blocks": [ + { + "bbox": [ + 406, + 114, + 471, + 159 + ], + "lines": [ + { + "bbox": [ + 406, + 114, + 471, + 159 + ], + "spans": [ + { + "bbox": [ + 406, + 114, + 471, + 159 + ], + "type": "image", + "image_path": "f8091942473796a09656b9da90cd195fc51d5c9f4bf39e4ad0622fae5d28d8ff.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 406, + 159, + 471, + 204 + ], + "blocks": [ + { + "bbox": [ + 406, + 159, + 471, + 204 + ], + "lines": [ + { + "bbox": [ + 406, + 159, + 471, + 204 + ], + "spans": [ + { + "bbox": [ + 406, + 159, + 471, + 204 + ], + "type": "image", + "image_path": "deb4b2e0d49cc2d6854a38a525247c347665344734ea87235a23b97376720973.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 425, + 206, + 451, + 217 + ], + "lines": [ + { + "bbox": [ + 425, + 206, + 451, + 217 + ], + "spans": [ + { + "bbox": [ + 425, + 206, + 451, + 217 + ], + "type": "text", + "content": "DINO" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 262, + 322, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 262, + 322, + 274 + ], + "spans": [ + { + "bbox": [ + 131, + 262, + 322, + 274 + ], + "type": "text", + "content": "4.3 Results of Object Priors in SOS" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 284, + 482, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 284, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 284, + 482, + 403 + ], + "type": "text", + "content": "Table 1 shows the results of our object prior study. First, all priors outperform the baselines Grid and Dist in terms of " + }, + { + "bbox": [ + 130, + 284, + 482, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_1" + }, + { + "bbox": [ + 130, + 284, + 482, + 403 + ], + "type": "text", + "content": ". Both baselines exhibit a lower AP compared to all other object priors, reflecting the missing focus on objects. Hence, simply applying SAM (Grid) is not suitable to segment only objects, as several stuff regions are segmented as well. Second, most priors exhibit a similar performance in terms of all measures. For instance, learning-based CAM and DeepGaze do not outperform simple priors (Spx, Contour, VOCUS2). Only the priors U-Net and DINO perform substantially better, with DINO producing the best result. Comparing DINO to GT reveals that DINO is able to recall almost the same amount of objects, but exhibits a lower precision, as expected." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 403, + 482, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 523 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 523 + ], + "type": "text", + "content": "To illustrate the results, Fig. 3 depicts object priors for top-performing VOCUS2, " + }, + { + "bbox": [ + 130, + 403, + 482, + 523 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 130, + 403, + 482, + 523 + ], + "type": "text", + "content": "-Net, and DINO, as heatmaps (upper row) with the resulting pseudo annotations generated in SOS (lower row) for an example image. The heatmaps show that DINO is very focused on the discriminative object parts, here the polar bears' faces, while " + }, + { + "bbox": [ + 130, + 403, + 482, + 523 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 130, + 403, + 482, + 523 + ], + "type": "text", + "content": "-Net highlights, as learned, the entire objects. Moreover, " + }, + { + "bbox": [ + 130, + 403, + 482, + 523 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 130, + 403, + 482, + 523 + ], + "type": "text", + "content": "-Net also highlights several background areas. Similarly, VOCUS2 highlights background areas and the polar bears since it is generally focused on high-contrast areas. Overall, only DINO exclusively generates pseudo annotations for the polar bears and some sub-parts, while " + }, + { + "bbox": [ + 130, + 403, + 482, + 523 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 130, + 403, + 482, + 523 + ], + "type": "text", + "content": "-Net and VOCUS2 produce noisy annotations including background patches, similar to previous approaches [49]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 523, + 482, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 523, + 482, + 583 + ], + "spans": [ + { + "bbox": [ + 130, + 523, + 482, + 583 + ], + "type": "text", + "content": "The results show that a well-designed object prior like DINO substantially outperforms the baseline Grid " + }, + { + "bbox": [ + 130, + 523, + 482, + 583 + ], + "type": "inline_equation", + "content": "(+7.5" + }, + { + "bbox": [ + 130, + 523, + 482, + 583 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 130, + 523, + 482, + 583 + ], + "type": "inline_equation", + "content": "\\mathrm{F}_1)" + }, + { + "bbox": [ + 130, + 523, + 482, + 583 + ], + "type": "text", + "content": " in SOS. Therefore, we use DINO as the object prior in SOS. More generally, the study also reveals important insights on how to prompt SAM for class-agnostic, object-focused results for arbitrary tasks. For more study results, we refer to our supplementary." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 131, + 602, + 218, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 602, + 218, + 615 + ], + "spans": [ + { + "bbox": [ + 131, + 602, + 218, + 615 + ], + "type": "text", + "content": "5 Evaluation" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 628, + 485, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 628, + 485, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 628, + 485, + 666 + ], + "type": "text", + "content": "In our evaluation, we compare SOS to recent OWIS methods OLN [26], LDET [46], GGN [49], SWORD [54], SOIS [56], OpenInst [48], UDOS [24], and against Mask R-CNN [19] and SAM [27] baselines, depending on the availability of public" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 325, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 325, + 91, + 447, + 102 + ], + "type": "text", + "content": "SOS: Segment Object System" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 161, + 146, + 452, + 226 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 138 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 138 + ], + "type": "text", + "content": "Table 1: Results of our SOS with various object priors in the cross-category COCO (VOC) " + }, + { + "bbox": [ + 132, + 114, + 482, + 138 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 132, + 114, + 482, + 138 + ], + "type": "text", + "content": " COCO (non-VOC) setting. *: Uses ground truth of unknown classes." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 161, + 146, + 452, + 226 + ], + "lines": [ + { + "bbox": [ + 161, + 146, + 452, + 226 + ], + "spans": [ + { + "bbox": [ + 161, + 146, + 452, + 226 + ], + "type": "table", + "html": "
Object PriorAPAR100F1Object PriorAPAR100F1
SOS + Grid3.836.56.9SOS + DeepGaze5.435.99.4
SOS + Dist3.427.46.0SOS + CAM5.436.79.4
SOS + Spx5.634.89.6SOS + DINO8.938.114.4
SOS + Contour5.636.69.7SOS + U-Net7.337.312.2
SOS + VOCUS25.137.710.5SOS + GT18.1*42.5*25.4*
", + "image_path": "66088e09959caa645f7fca47fc3764592d6668c76ae9b69873a847a33ac854ca.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 251, + 482, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 251, + 482, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 482, + 384 + ], + "type": "text", + "content": "source code or reported results. To assess the quality of the approaches, we use four cross-category and cross-dataset settings, commonly used in OWIS. Specifically, we choose COCO (VOC) " + }, + { + "bbox": [ + 130, + 251, + 482, + 384 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 251, + 482, + 384 + ], + "type": "text", + "content": " COCO (non-VOC) on the COCO dataset [31], already discussed in Sec. 4.2. For cross-dataset evaluation, we use COCO " + }, + { + "bbox": [ + 130, + 251, + 482, + 384 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 251, + 482, + 384 + ], + "type": "text", + "content": " LVIS, COCO " + }, + { + "bbox": [ + 130, + 251, + 482, + 384 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 251, + 482, + 384 + ], + "type": "text", + "content": " ADE20k, and COCO " + }, + { + "bbox": [ + 130, + 251, + 482, + 384 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 251, + 482, + 384 + ], + "type": "text", + "content": " UVO that train on the entire COCO training set and evaluate on the validation sets of LVIS [17], ADE20k [61], and UVO [50] with exhaustive annotations. The LVIS dataset covers the same images as COCO but extends the annotated object classes from 80 to 1203. ADE20k includes 3169 classes of objects, object parts, and stuff regions. Finally, UVO features exhaustive annotations of all objects and, therefore, no system is penalized for detecting objects outside a datasets' taxonomy." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 384, + 482, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 482, + 433 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 482, + 433 + ], + "type": "text", + "content": "As in Sec. 4.2 and following common practice in OWIS [26,46,49], we report AR for up to 100 detections and AP. We again also report the " + }, + { + "bbox": [ + 130, + 384, + 482, + 433 + ], + "type": "inline_equation", + "content": "\\mathrm{F}_1" + }, + { + "bbox": [ + 130, + 384, + 482, + 433 + ], + "type": "text", + "content": " score between AR and AP as a single, combined quantity, described in Sec. 4.2. Note that we only evaluate based on masks, not boxes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 457, + 364, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 457, + 364, + 470 + ], + "spans": [ + { + "bbox": [ + 131, + 457, + 364, + 470 + ], + "type": "text", + "content": "5.1 Implementation Details and Data Usage" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 484, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 484, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 484, + 482, + 568 + ], + "type": "text", + "content": "As discussed in Sec. 4.3, SOS uses the object prior DINO for best results. DINO [6] with a ViT-S backbone is used as an ImageNet [45] pre-trained model without fine-tuning. Similarly, SOS uses SAM [27] with the ViT-H backbone as a pre-trained model based on SA-1B [27] without fine-tuning. As an instance segmentation model, we apply Mask R-CNN [19] with an ImageNet pre-trained ResNet-50+FPN backbone [30], and the default configurations except for the class-agnostic training." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Overall, SOS is based on models pre-trained on standard ImageNet and SA-1B. Pre-training on SA-1B or similar large-scale datasets becomes common with the emergence of foundation models and is widely adopted by systems using SAM in similar contexts [7,51,57]. It is comparable to using models pre-trained on ImageNet that even have class information available during training and also allow a rough localization of objects [6,60]. Moreover, the final, noisy annotations of SA-1B representing objects, object parts, and stuff region were automatically generated by SAM [27], and no data leakage w.r.t. other datasets exists." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "text", + "content": "C. Wilms et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 145, + 178, + 288, + 302 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 301, + 169 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 301, + 169 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 301, + 169 + ], + "type": "text", + "content": "Table 2: Results of two baselines and various OWIS methods in the COCO (VOC) " + }, + { + "bbox": [ + 132, + 114, + 301, + 169 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 132, + 114, + 301, + 169 + ], + "type": "text", + "content": " COCO (non-VOC) setting. " + }, + { + "bbox": [ + 132, + 114, + 301, + 169 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 132, + 114, + 301, + 169 + ], + "type": "text", + "content": ": uses automatically annotated SA-1B dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 145, + 178, + 288, + 302 + ], + "lines": [ + { + "bbox": [ + 145, + 178, + 288, + 302 + ], + "spans": [ + { + "bbox": [ + 145, + 178, + 288, + 302 + ], + "type": "table", + "html": "
SystemAPAR100F1
Mask R-CNN [19]1.08.21.8
SAM†[27]3.648.16.7
OLN [26]4.228.47.3
LDET [46]4.324.87.3
GGN [49]4.928.38.4
SWORD [54]4.830.28.3
UDOS [24]2.934.35.3
SOS†(ours)8.939.314.5
", + "image_path": "e1ecb68f6776a93a15f6a3710ff52f5b88420b7ae1eccc62ac3c834c66237b00.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 325, + 168, + 468, + 291 + ], + "blocks": [ + { + "bbox": [ + 311, + 114, + 482, + 158 + ], + "lines": [ + { + "bbox": [ + 311, + 114, + 482, + 158 + ], + "spans": [ + { + "bbox": [ + 311, + 114, + 482, + 158 + ], + "type": "text", + "content": "Table 3: Results of two baselines and various OWIS methods in the COCO " + }, + { + "bbox": [ + 311, + 114, + 482, + 158 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 311, + 114, + 482, + 158 + ], + "type": "text", + "content": " LVIS setting. " + }, + { + "bbox": [ + 311, + 114, + 482, + 158 + ], + "type": "inline_equation", + "content": "^\\dagger" + }, + { + "bbox": [ + 311, + 114, + 482, + 158 + ], + "type": "text", + "content": ": uses automatically annotated SA-1B dataset." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 325, + 168, + 468, + 291 + ], + "lines": [ + { + "bbox": [ + 325, + 168, + 468, + 291 + ], + "spans": [ + { + "bbox": [ + 325, + 168, + 468, + 291 + ], + "type": "table", + "html": "
SystemAPAR100F1
Mask R-CNN [19]7.523.611.4
SAM†[27]6.845.111.8
LDET [46]6.724.810.5
GGN [49]6.527.010.5
SOIS [56]-25.2-
OpenInst [48]-29.3-
UDOS [24]3.924.96.7
SOS†(ours)8.133.313.3
", + "image_path": "623c27c08bdf2b2822d6a7eb01985604d642233c5ec3a310235b526d35de59dd.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 325, + 463, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 463, + 338 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 463, + 338 + ], + "type": "text", + "content": "5.2 Cross-category COCO (VOC) " + }, + { + "bbox": [ + 132, + 325, + 463, + 338 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 132, + 325, + 463, + 338 + ], + "type": "text", + "content": " COCO (non-VOC) Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 347, + 482, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 347, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 482, + 468 + ], + "type": "text", + "content": "Table 2 presents the results of various OWIS methods on the cross-category setup COCO (VOC) " + }, + { + "bbox": [ + 130, + 347, + 482, + 468 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 347, + 482, + 468 + ], + "type": "text", + "content": " COCO (non-VOC). Our SOS clearly outperforms all OWIS methods in " + }, + { + "bbox": [ + 130, + 347, + 482, + 468 + ], + "type": "inline_equation", + "content": "\\mathrm{AR}_{100}" + }, + { + "bbox": [ + 130, + 347, + 482, + 468 + ], + "type": "text", + "content": ", AP, and " + }, + { + "bbox": [ + 130, + 347, + 482, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_1" + }, + { + "bbox": [ + 130, + 347, + 482, + 468 + ], + "type": "text", + "content": ". Specifically, SOS outperforms the second-best system in terms of " + }, + { + "bbox": [ + 130, + 347, + 482, + 468 + ], + "type": "inline_equation", + "content": "\\mathrm{F}_1" + }, + { + "bbox": [ + 130, + 347, + 482, + 468 + ], + "type": "text", + "content": ", GGN [49], by 6.1. While some of the improvement comes from a better recall (" + }, + { + "bbox": [ + 130, + 347, + 482, + 468 + ], + "type": "inline_equation", + "content": "+38.9\\%" + }, + { + "bbox": [ + 130, + 347, + 482, + 468 + ], + "type": "text", + "content": " compared to GGN), the results are driven by a much-improved precision (" + }, + { + "bbox": [ + 130, + 347, + 482, + 468 + ], + "type": "inline_equation", + "content": "+81.6\\%" + }, + { + "bbox": [ + 130, + 347, + 482, + 468 + ], + "type": "text", + "content": " compared to GGN). This reflects the high quality of the pseudo annotations in SOS on this cross-category setting. Moreover, there is a clear improvement of SOS over Mask R-CNN (" + }, + { + "bbox": [ + 130, + 347, + 482, + 468 + ], + "type": "inline_equation", + "content": "+12.7" + }, + { + "bbox": [ + 130, + 347, + 482, + 468 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 130, + 347, + 482, + 468 + ], + "type": "inline_equation", + "content": "\\mathrm{F}_1" + }, + { + "bbox": [ + 130, + 347, + 482, + 468 + ], + "type": "text", + "content": "). Compared to original SAM, the DINO-based SOS substantially improves the precision by focusing on objects, leading to an improved " + }, + { + "bbox": [ + 130, + 347, + 482, + 468 + ], + "type": "inline_equation", + "content": "\\mathrm{F}_1" + }, + { + "bbox": [ + 130, + 347, + 482, + 468 + ], + "type": "text", + "content": " score (" + }, + { + "bbox": [ + 130, + 347, + 482, + 468 + ], + "type": "inline_equation", + "content": "+7.8" + }, + { + "bbox": [ + 130, + 347, + 482, + 468 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 468, + 482, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 482, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 482, + 540 + ], + "type": "text", + "content": "The qualitative results of various OWIS methods in Fig. 4 support the quantitative results. In the first example, only GGN [49] and SOS detect both giraffes, however, SOS generates more accurate segmentations. In the second example, only SOS detects all five small surfboards, while other systems detect at most three. Note that class-agnostic Mask R-CNN misses all non-VOC objects in both examples. For more qualitative results, see our supplementary." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 559, + 270, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 559, + 270, + 571 + ], + "spans": [ + { + "bbox": [ + 132, + 559, + 270, + 571 + ], + "type": "text", + "content": "5.3 Cross-dataset Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "COCO \\rightarrow LVIS" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": ": We evaluate the cross-dataset generalization capabilities of OWIS methods, starting with " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "COCO \\rightarrow LVIS" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": ". The results in Tab. 3 show a trend similar to the cross-category results. SOS outperforms all OWIS methods across all measures with an improvement of 2.8 in " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "F_1" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " compared to second-best methods GGN [49] and LDET [46]. Similar to the previous results, the improvement is based on gains in both recall and precision, however, the relative improvements in AP and " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "AR_{100}" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " are more similar to each other. Overall, our high-quality pseudo" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 325, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 325, + 91, + 447, + 102 + ], + "type": "text", + "content": "SOS: Segment Object System" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 113, + 483, + 230 + ], + "blocks": [ + { + "bbox": [ + 134, + 113, + 483, + 230 + ], + "lines": [ + { + "bbox": [ + 134, + 113, + 483, + 230 + ], + "spans": [ + { + "bbox": [ + 134, + 113, + 483, + 230 + ], + "type": "image", + "image_path": "235737833757eee12f46f0948406e067ab08097a1e13ad6052fe5c94d75f0513.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 238, + 482, + 273 + ], + "lines": [ + { + "bbox": [ + 130, + 238, + 482, + 273 + ], + "spans": [ + { + "bbox": [ + 130, + 238, + 482, + 273 + ], + "type": "text", + "content": "Fig. 4: Qualitative results of OWIS methods and baseline Mask R-CNN in the cross-category COCO (VOC) " + }, + { + "bbox": [ + 130, + 238, + 482, + 273 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 238, + 482, + 273 + ], + "type": "text", + "content": " COCO (non-VOC) setting. Filled masks denote detected objects, while red frames indicate missed objects." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 143, + 342, + 291, + 446 + ], + "blocks": [ + { + "bbox": [ + 131, + 289, + 302, + 334 + ], + "lines": [ + { + "bbox": [ + 131, + 289, + 302, + 334 + ], + "spans": [ + { + "bbox": [ + 131, + 289, + 302, + 334 + ], + "type": "text", + "content": "Table 4: Results of Mask R-CNN baseline and various OWIS methods in the COCO " + }, + { + "bbox": [ + 131, + 289, + 302, + 334 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 131, + 289, + 302, + 334 + ], + "type": "text", + "content": " ADE20k setting. " + }, + { + "bbox": [ + 131, + 289, + 302, + 334 + ], + "type": "inline_equation", + "content": "^\\dagger" + }, + { + "bbox": [ + 131, + 289, + 302, + 334 + ], + "type": "text", + "content": ": uses automatically annotated SA-1B dataset." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 143, + 342, + 291, + 446 + ], + "lines": [ + { + "bbox": [ + 143, + 342, + 291, + 446 + ], + "spans": [ + { + "bbox": [ + 143, + 342, + 291, + 446 + ], + "type": "table", + "html": "
SystemAPAR100F1
Mask R-CNN [19]6.911.98.7
OLN [26]-20.4-
LDET [46]9.518.512.6
GGN [49]9.721.013.3
UDOS [24]7.622.911.4
SOS† (ours)12.526.517.0
", + "image_path": "6bc6676155d9cdb73b115bb18cb55aaf8354b3eaaf2183c16ba577deaab74bb4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 322, + 338, + 472, + 450 + ], + "blocks": [ + { + "bbox": [ + 311, + 284, + 482, + 327 + ], + "lines": [ + { + "bbox": [ + 311, + 284, + 482, + 327 + ], + "spans": [ + { + "bbox": [ + 311, + 284, + 482, + 327 + ], + "type": "text", + "content": "Table 5: Results of two baselines and various OWIS methods in the COCO " + }, + { + "bbox": [ + 311, + 284, + 482, + 327 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 311, + 284, + 482, + 327 + ], + "type": "text", + "content": " UVO setting. " + }, + { + "bbox": [ + 311, + 284, + 482, + 327 + ], + "type": "inline_equation", + "content": "^\\dagger" + }, + { + "bbox": [ + 311, + 284, + 482, + 327 + ], + "type": "text", + "content": ": uses automatically annotated SA-1B dataset." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 322, + 338, + 472, + 450 + ], + "lines": [ + { + "bbox": [ + 322, + 338, + 472, + 450 + ], + "spans": [ + { + "bbox": [ + 322, + 338, + 472, + 450 + ], + "type": "table", + "html": "
SystemAPAR100F1
Mask R-CNN [19]20.736.726.5
SAM†[27]11.350.118.4
OLN [26]-41.4-
LDET [46]22.040.428.5
GGN [49]20.343.427.7
UDOS [24]10.643.117.0
SOS†(ours)20.942.328.0
", + "image_path": "e46675cfa54e38ab343a8703b83b50ff2882ecde760656958f3553f891ac9520.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 472, + 482, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 498 + ], + "type": "text", + "content": "annotations in SOS lead to new state-of-the-art results, which indicate strong generalization to unknown object classes outside COCO." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 509, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 482, + 594 + ], + "type": "inline_equation", + "content": "COCO \\rightarrow ADE20k" + }, + { + "bbox": [ + 130, + 509, + 482, + 594 + ], + "type": "text", + "content": ": The results for " + }, + { + "bbox": [ + 130, + 509, + 482, + 594 + ], + "type": "inline_equation", + "content": "COCO \\rightarrow ADE20k" + }, + { + "bbox": [ + 130, + 509, + 482, + 594 + ], + "type": "text", + "content": " in Tab. 4, which include labeled object parts of " + }, + { + "bbox": [ + 130, + 509, + 482, + 594 + ], + "type": "inline_equation", + "content": "ADE20k" + }, + { + "bbox": [ + 130, + 509, + 482, + 594 + ], + "type": "text", + "content": ", show that SOS again outperforms all other OWIS methods across all measures. The gain over the second-best method in " + }, + { + "bbox": [ + 130, + 509, + 482, + 594 + ], + "type": "inline_equation", + "content": "F_1" + }, + { + "bbox": [ + 130, + 509, + 482, + 594 + ], + "type": "text", + "content": ", GGN [49], is 3.7. Similar to " + }, + { + "bbox": [ + 130, + 509, + 482, + 594 + ], + "type": "inline_equation", + "content": "COCO \\rightarrow LVIS" + }, + { + "bbox": [ + 130, + 509, + 482, + 594 + ], + "type": "text", + "content": ", the relative gains in " + }, + { + "bbox": [ + 130, + 509, + 482, + 594 + ], + "type": "inline_equation", + "content": "AR_{100}" + }, + { + "bbox": [ + 130, + 509, + 482, + 594 + ], + "type": "text", + "content": " and AP are equally distributed. These results indicate that object parts do not degrade the results of SOS. This is in line with SOS's pseudo annotations in Fig. 3, covering entire objects and selected object parts." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "COCO \\rightarrow UVO" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": ": Finally, Tab. 5 presents results for COCO " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " UVO. SOS still outperforms Mask R-CNN, SAM and recent UDOS [24] in " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "F_1" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": ". Outperforming SAM implies that SOS does not exploit the limited taxonomy of annotations in COCO (only 60 object classes in test), but also improves the segmentation of objects outside the COCO object classes on the exhaustively labeled UVO" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "type": "text", + "content": "C. Wilms et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 156, + 176, + 278, + 239 + ], + "blocks": [ + { + "bbox": [ + 132, + 110, + 301, + 165 + ], + "lines": [ + { + "bbox": [ + 132, + 110, + 301, + 165 + ], + "spans": [ + { + "bbox": [ + 132, + 110, + 301, + 165 + ], + "type": "text", + "content": "Table 6: Quality of pseudo annotation generated in GGN [49] and our SOS evaluated on the non-VOC annotations of the COCO training dataset. " + }, + { + "bbox": [ + 132, + 110, + 301, + 165 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 132, + 110, + 301, + 165 + ], + "type": "text", + "content": ": uses automatically annotated SA-1B dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 156, + 176, + 278, + 239 + ], + "lines": [ + { + "bbox": [ + 156, + 176, + 278, + 239 + ], + "spans": [ + { + "bbox": [ + 156, + 176, + 278, + 239 + ], + "type": "table", + "html": "
SystemPrec.Rec.F1
GGN3 [49]7.312.19.1
SOS3† (ours)19.026.422.1
SOS10† (ours)15.541.722.6
", + "image_path": "fd46bf28b7c4d2b49bc604b828d7560109f4cc7142fb950d0ed1ae9b51186311.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 328, + 114, + 396, + 160 + ], + "blocks": [ + { + "bbox": [ + 328, + 114, + 396, + 160 + ], + "lines": [ + { + "bbox": [ + 328, + 114, + 396, + 160 + ], + "spans": [ + { + "bbox": [ + 328, + 114, + 396, + 160 + ], + "type": "image", + "image_path": "394330a6f99d5e5240e6cc209dbeaa13e7415f348e39041847517301f8eeddb5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 398, + 114, + 465, + 160 + ], + "blocks": [ + { + "bbox": [ + 398, + 114, + 465, + 160 + ], + "lines": [ + { + "bbox": [ + 398, + 114, + 465, + 160 + ], + "spans": [ + { + "bbox": [ + 398, + 114, + 465, + 160 + ], + "type": "image", + "image_path": "fe5e1a842cec4d04dd0b27b3f9f5b51cd31e6e815eb15bddde46e336af14f87f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 328, + 161, + 396, + 208 + ], + "blocks": [ + { + "bbox": [ + 328, + 161, + 396, + 208 + ], + "lines": [ + { + "bbox": [ + 328, + 161, + 396, + 208 + ], + "spans": [ + { + "bbox": [ + 328, + 161, + 396, + 208 + ], + "type": "image", + "image_path": "9304cbb3b90f3cfe5e5a7ff9f135b3e7c850345a8da0b789f1e5f3bcac0973b1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 352, + 211, + 372, + 219 + ], + "lines": [ + { + "bbox": [ + 352, + 211, + 372, + 219 + ], + "spans": [ + { + "bbox": [ + 352, + 211, + 372, + 219 + ], + "type": "text", + "content": "GGN" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 398, + 161, + 465, + 208 + ], + "blocks": [ + { + "bbox": [ + 398, + 161, + 465, + 208 + ], + "lines": [ + { + "bbox": [ + 398, + 161, + 465, + 208 + ], + "spans": [ + { + "bbox": [ + 398, + 161, + 465, + 208 + ], + "type": "image", + "image_path": "0a44d544eb21135caae4783ce969e642135dac6ffc058018478b0abd9141cd59.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 423, + 211, + 440, + 219 + ], + "lines": [ + { + "bbox": [ + 423, + 211, + 440, + 219 + ], + "spans": [ + { + "bbox": [ + 423, + 211, + 440, + 219 + ], + "type": "text", + "content": "SOS" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 311, + 220, + 481, + 242 + ], + "lines": [ + { + "bbox": [ + 311, + 220, + 481, + 242 + ], + "spans": [ + { + "bbox": [ + 311, + 220, + 481, + 242 + ], + "type": "text", + "content": "Fig. 5: Pseudo annotations generated by GGN and our SOS." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 277, + 482, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 482, + 337 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 482, + 337 + ], + "type": "text", + "content": "dataset. However, the results of SOS are below LDET [46] in AP and " + }, + { + "bbox": [ + 130, + 277, + 482, + 337 + ], + "type": "inline_equation", + "content": "\\mathrm{F_1}" + }, + { + "bbox": [ + 130, + 277, + 482, + 337 + ], + "type": "text", + "content": ". We attribute this to several classes in UVO not available in ImageNet, which was used to learn DINO object prior. Hence, DINO will miss such objects in the COCO dataset during training, and SOS might miss them during testing. This could be mitigated using more diverse unlabeled images for training DINO." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 355, + 317, + 367 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 355, + 317, + 367 + ], + "spans": [ + { + "bbox": [ + 131, + 355, + 317, + 367 + ], + "type": "text", + "content": "5.4 Quality of Pseudo Annotations" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 375, + 482, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 375, + 482, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 375, + 482, + 506 + ], + "type": "text", + "content": "We also investigate the quality of the pseudo annotations generated in SOS. To this end, we evaluate the pseudo annotations on the COCO training set against the annotations of non-VOC object classes, reflecting the COCO (VOC) " + }, + { + "bbox": [ + 130, + 375, + 482, + 506 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 375, + 482, + 506 + ], + "type": "text", + "content": " COCO non-VOC) setting. The results in terms of precision and recall for IoU " + }, + { + "bbox": [ + 130, + 375, + 482, + 506 + ], + "type": "inline_equation", + "content": "= 0.5" + }, + { + "bbox": [ + 130, + 375, + 482, + 506 + ], + "type": "text", + "content": " as well as the " + }, + { + "bbox": [ + 130, + 375, + 482, + 506 + ], + "type": "inline_equation", + "content": "\\mathrm{F_1}" + }, + { + "bbox": [ + 130, + 375, + 482, + 506 + ], + "type": "text", + "content": " score are presented in Tab. 6. We use up to three and 10 pseudo annotations generated per image in SOS (SOS" + }, + { + "bbox": [ + 130, + 375, + 482, + 506 + ], + "type": "inline_equation", + "content": "_3" + }, + { + "bbox": [ + 130, + 375, + 482, + 506 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 375, + 482, + 506 + ], + "type": "inline_equation", + "content": "\\mathrm{SOS}_{10}" + }, + { + "bbox": [ + 130, + 375, + 482, + 506 + ], + "type": "text", + "content": "), and " + }, + { + "bbox": [ + 130, + 375, + 482, + 506 + ], + "type": "inline_equation", + "content": "\\mathrm{GGN}_3" + }, + { + "bbox": [ + 130, + 375, + 482, + 506 + ], + "type": "text", + "content": " [49] that generates three pseudo annotations. The results show that with only three SOS pseudo annotations, more than " + }, + { + "bbox": [ + 130, + 375, + 482, + 506 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 130, + 375, + 482, + 506 + ], + "type": "text", + "content": " of the non-VOC objects are covered. This is substantially more than " + }, + { + "bbox": [ + 130, + 375, + 482, + 506 + ], + "type": "inline_equation", + "content": "\\mathrm{GGN}_3" + }, + { + "bbox": [ + 130, + 375, + 482, + 506 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 375, + 482, + 506 + ], + "type": "inline_equation", + "content": "12.1\\%" + }, + { + "bbox": [ + 130, + 375, + 482, + 506 + ], + "type": "text", + "content": "). With up to 10 annotations, SOS's pseudo annotations cover more than " + }, + { + "bbox": [ + 130, + 375, + 482, + 506 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 130, + 375, + 482, + 506 + ], + "type": "text", + "content": " of the non-VOC objects. In our supplementary, we further provide class-specific results." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 506, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 506, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 506, + 482, + 567 + ], + "type": "text", + "content": "To complement these results, we visualize the annotations of " + }, + { + "bbox": [ + 130, + 506, + 482, + 567 + ], + "type": "inline_equation", + "content": "\\mathrm{GGN}_3" + }, + { + "bbox": [ + 130, + 506, + 482, + 567 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 506, + 482, + 567 + ], + "type": "inline_equation", + "content": "\\mathrm{SOS}_3" + }, + { + "bbox": [ + 130, + 506, + 482, + 567 + ], + "type": "text", + "content": " on two sample images in Fig. 5. It is visible that " + }, + { + "bbox": [ + 130, + 506, + 482, + 567 + ], + "type": "inline_equation", + "content": "\\mathrm{GGN}_3" + }, + { + "bbox": [ + 130, + 506, + 482, + 567 + ], + "type": "text", + "content": "'s annotations cover background regions, leading to low precision in the overall results. Moreover, the pseudo annotations in the upper example do not adhere well to the airplane's boundaries. More qualitative examples are given in Fig. 1 and our supplementary." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 585, + 246, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 585, + 246, + 596 + ], + "spans": [ + { + "bbox": [ + 131, + 585, + 246, + 596 + ], + "type": "text", + "content": "5.5 Ablation Studies" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 605, + 481, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 629 + ], + "type": "text", + "content": "All ablation studies follow the evaluation setup described in Sec. 4.2. More studies evaluating design choices in SOS are presented in our supplementary." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": "Influence of SOS's Components: We investigate the importance of each component in SOS and present the results in Tab. 7. The baseline for this" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 325, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 325, + 91, + 447, + 102 + ], + "type": "text", + "content": "SOS: Segment Object System" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 146, + 157, + 288, + 226 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 301, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 301, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 301, + 148 + ], + "type": "text", + "content": "Table 7: Influence of added components in our SOS evaluated in the COCO (VOC) " + }, + { + "bbox": [ + 132, + 114, + 301, + 148 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 132, + 114, + 301, + 148 + ], + "type": "text", + "content": " COCO (non-VOC) setting." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 146, + 157, + 288, + 226 + ], + "lines": [ + { + "bbox": [ + 146, + 157, + 288, + 226 + ], + "spans": [ + { + "bbox": [ + 146, + 157, + 288, + 226 + ], + "type": "table", + "html": "
ComponentsAPAR100F1
Mask R-CNN1.210.82.2
+ Grid, w/o pp.3.435.26.2
+ DINO, w/o pp.8.937.114.3
+ DINO, w/ pp.8.938.114.4
", + "image_path": "48fc97244774b6490dd63b756e42f697266b4851a75b809f661c5b8294941d90.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 331, + 157, + 463, + 226 + ], + "blocks": [ + { + "bbox": [ + 312, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 312, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 312, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 8: SOS results with various numbers of pseudo annotations in the COCO (VOC) " + }, + { + "bbox": [ + 312, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 312, + 114, + 482, + 148 + ], + "type": "text", + "content": " COCO (non-VOC) setting." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 331, + 157, + 463, + 226 + ], + "lines": [ + { + "bbox": [ + 331, + 157, + 463, + 226 + ], + "spans": [ + { + "bbox": [ + 331, + 157, + 463, + 226 + ], + "type": "table", + "html": "
#Pseudo Anns.APAR100F1
38.334.513.4
58.836.214.2
108.938.114.4
208.838.114.3
", + "image_path": "aa823a562d9ae7e99fbc5f1e6aa9d3073931de7b3c46402c2f0774326ca1b945.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 246, + 482, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 246, + 482, + 368 + ], + "spans": [ + { + "bbox": [ + 130, + 246, + 482, + 368 + ], + "type": "text", + "content": "study is a class-agnostic Mask R-CNN trained without pseudo annotations (first row in Tab. 7). Subsequently, we add pseudo annotations based on SAM with the baseline Grid object prior as in [27] and without our post-processing in PAC (confidence-based thresholding and NMS, see Sec. 3.3). Next, we replace Grid with DINO (third row in Tab. 7), and finally add the post-processing leading to the complete SOS. The results in Tab. 7 show each step improves the " + }, + { + "bbox": [ + 130, + 246, + 482, + 368 + ], + "type": "inline_equation", + "content": "\\mathrm{F}_1" + }, + { + "bbox": [ + 130, + 246, + 482, + 368 + ], + "type": "text", + "content": " results. Mainly, introducing SAM for generating pseudo annotations (first row vs. second row) substantially improves " + }, + { + "bbox": [ + 130, + 246, + 482, + 368 + ], + "type": "inline_equation", + "content": "\\mathrm{AR}_{100}" + }, + { + "bbox": [ + 130, + 246, + 482, + 368 + ], + "type": "text", + "content": ", while adding the DINO object prior removes background annotations leading to considerably improved precision. Hence, the object prior selection in SOS is crucial for strong results." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 375, + 483, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 375, + 483, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 375, + 483, + 472 + ], + "type": "text", + "content": "Number of Pseudo Annotations: In the second ablation study, we investigate the influence of the number of pseudo annotations per image on the results of SOS. To this end, we evaluate SOS with up to 3, 5, 10, and 20 pseudo annotations per image. The results in Tab. 8 show that 10 pseudo annotations are preferable. Hence, the ideal number of pseudo annotations in SOS is larger than in GGN [49], which only uses three. This indicates again the high quality of our pseudo annotations, since more than three annotations per image can cover relevant objects without adding too many noisy annotations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 487, + 221, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 221, + 499 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 221, + 499 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we addressed the challenging OWIS task and aimed to improve the low precision of previous methods in this open-world task. To improve the precision and the overall detection results, we generate high-quality pseudo annotations based on a prompt-guided foundation model, SAM. To focus the pseudo annotations on objects, we thoroughly investigated various object priors for prompt creation, which revealed important insights on how to prompt SAM for class-agnostic object-focused results in arbitrary tasks. As a result, our novel OWIS method SOS, which uses a self-attention-based object prior for generating pseudo annotations, outperforms recent state-of-the-art OWIS methods across challenging open-world setups in COCO, LVIS, and ADE20k datasets. SOS especially improves the precision based on the high-quality pseudo annotations. Moreover, SOS with its focus on objects also outperforms vanilla SAM that segments both object and stuff regions." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "type": "text", + "content": "C. Wilms et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 137, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 138, + 137, + 480, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 137, + 480, + 169 + ], + "spans": [ + { + "bbox": [ + 138, + 137, + 480, + 169 + ], + "type": "text", + "content": "1. Ahn, J., Cho, S., Kwak, S.: Weakly supervised learning of instance segmentation with inter-pixel relations. In: Conference on Computer Vision and Pattern Recognition (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 170, + 480, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 170, + 480, + 191 + ], + "spans": [ + { + "bbox": [ + 138, + 170, + 480, + 191 + ], + "type": "text", + "content": "2. Alexe, B., Deselaers, T., Ferrari, V.: What is an object? In: Conference on Computer Vision and Pattern Recognition (2010)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 192, + 480, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 192, + 480, + 213 + ], + "spans": [ + { + "bbox": [ + 138, + 192, + 480, + 213 + ], + "type": "text", + "content": "3. Baugh, M., Batten, J., Müller, J.P., Kainz, B.: Zero-shot anomaly detection with pre-trained segmentation models. arXiv preprint arXiv:2306.09269 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 213, + 480, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 213, + 480, + 245 + ], + "spans": [ + { + "bbox": [ + 138, + 213, + 480, + 245 + ], + "type": "text", + "content": "4. Bommasani, R., Hudson, D.A., Adeli, E., Altman, R., Arora, S., von Arx, S., Bernstein, M.S., Bohg, J., Bosselut, A., Brunskill, E., et al.: On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 246, + 480, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 246, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 138, + 246, + 480, + 277 + ], + "type": "text", + "content": "5. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in Neural Information Processing Systems 33 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 278, + 480, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 278, + 480, + 310 + ], + "spans": [ + { + "bbox": [ + 138, + 278, + 480, + 310 + ], + "type": "text", + "content": "6. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: International Conference on Computer Vision (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 311, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 311, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 311, + 480, + 342 + ], + "type": "text", + "content": "7. Chen, T., Mai, Z., Li, R., Chao, W.l.: Segment anything model (SAM) enhanced pseudo labels for weakly supervised semantic segmentation. arXiv preprint arXiv:2305.05803 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 343, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 343, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 138, + 343, + 480, + 374 + ], + "type": "text", + "content": "8. Cheng, M.M., Mitra, N.J., Huang, X., Torr, P.H., Hu, S.M.: Global contrast based salient region detection. IEEE Transactions on Pattern Analysis and Machine Intelligence 37(3) (2014)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 375, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 375, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 138, + 375, + 480, + 418 + ], + "type": "text", + "content": "9. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: BERT: Pre-training of deep bidirectional transformers for language understanding. In: Human Language Technology Conference of the North American Chapter of the Association for Computational Linguistics (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 418, + 480, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 418, + 480, + 439 + ], + "spans": [ + { + "bbox": [ + 138, + 418, + 480, + 439 + ], + "type": "text", + "content": "0. Dollar, P., Zitnick, C.L.: Structured forests for fast edge detection. In: International Conference on Computer Vision (2013)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 440, + 480, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 440, + 480, + 482 + ], + "spans": [ + { + "bbox": [ + 138, + 440, + 480, + 482 + ], + "type": "text", + "content": "1. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 483, + 480, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 483, + 480, + 515 + ], + "spans": [ + { + "bbox": [ + 138, + 483, + 480, + 515 + ], + "type": "text", + "content": "2. Everingham, M., Eslami, S.A., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The PASCAL visual object classes challenge: A retrospective. International Journal of Computer Vision 111 (2015)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 516, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 516, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 138, + 516, + 480, + 536 + ], + "type": "text", + "content": "3. Felzenszwalb, P.F., Huttenlocher, D.P.: Efficient graph-based image segmentation. International Journal of Computer Vision 59 (2004)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 536, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 536, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 138, + 536, + 480, + 568 + ], + "type": "text", + "content": "4. Frintrop, S., Werner, T., Martin Garcia, G.: Traditional saliency reloaded: A good old model in new shape. In: Conference on Computer Vision and Pattern Recognition (2015)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 569, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 569, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 138, + 569, + 480, + 601 + ], + "type": "text", + "content": "5. Gao, G., Lauri, M., Zhang, J., Frintrop, S.: Saliency-guided adaptive seeding for supervoxel segmentation. In: International Conference on Intelligent Robots and Systems (2017)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 601, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 480, + 611 + ], + "type": "text", + "content": "6. Girshick, R.: Fast R-CNN. In: International Conference on Computer Vision (2015)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 612, + 480, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 612, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 138, + 612, + 480, + 633 + ], + "type": "text", + "content": "7. Gupta, A., Dolkar, P., Girshick, R.: LVIS: A dataset for large vocabulary instance segmentation. In: Conference on Computer Vision and Pattern Recognition (2019)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 634, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 634, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 634, + 480, + 665 + ], + "type": "text", + "content": "8. He, C., Li, K., Zhang, Y., Xu, G., Tang, L., Zhang, Y., Guo, Z., Li, X.: Weakly-supervised concealed object segmentation with SAM-based pseudo labeling and multi-scale feature grouping. arXiv preprint arXiv:2305.11003 (2023)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 325, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 325, + 91, + 447, + 102 + ], + "type": "text", + "content": "SOS: Segment Object System" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "19. He, K., Gkioxari, G., Dollar, P., Girshick, R.: Mask R-CNN. In: International Conference on Computer Vision (2017)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "type": "text", + "content": "20. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Conference on Computer Vision and Pattern Recognition (2016)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 161, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 480, + 194 + ], + "type": "text", + "content": "21. Hou, Q., Cheng, M.M., Hu, X., Borji, A., Tu, Z., Torr, P.H.: Deeply supervised salient object detection with short connections. In: Conference on Computer Vision and Pattern Recognition (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 225 + ], + "type": "text", + "content": "22. Hu, H., Lan, S., Jiang, Y., Cao, Z., Sha, F.: FastMask: Segment multi-scale object candidates in one shot. In: Conference on Computer Vision and Pattern Recognition (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "text", + "content": "23. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International Conference on Machine Learning (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 480, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 292 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 292 + ], + "type": "text", + "content": "24. Kalluri, T., Wang, W., Wang, H., Chandraker, M., Torresani, L., Tran, D.: Openworld instance segmentation: Top-down learning with bottom-up supervision. arXiv preprint arXiv:2303.05503 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "text", + "content": "25. Kim, B., Yoo, Y., Rhee, C.E., Kim, J.: Beyond semantic to instance segmentation: Weakly-supervised instance segmentation via semantic knowledge transfer and self-refinement. In: Conference on Computer Vision and Pattern Recognition (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "type": "text", + "content": "26. Kim, D., Lin, T.Y., Angelova, A., Kweon, I.S., Kuo, W.: Learning open-world object proposals without learning to classify. IEEE Robotics and Automation Letters 7(2) (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 358, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 391 + ], + "type": "text", + "content": "27. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: International Conference on Computer Vision (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "type": "text", + "content": "28. Klein, D.A., Frintrop, S.: Center-surround divergence of feature statistics for salient object detection. In: International Conference on Computer Vision (2011)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 414, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 480, + 445 + ], + "type": "text", + "content": "29. Li, J., Li, D., Savarese, S., Hoi, S.: BLIP-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 447, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 480, + 479 + ], + "type": "text", + "content": "30. Lin, T.Y., Dollar, P., Girshick, R., He, K., Hariharan, B., Belongie, S.: Feature pyramid networks for object detection. In: Conference on Computer Vision and Pattern Recognition (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "type": "text", + "content": "31. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft COCO: Common objects in context. In: European Conference on Computer Vision (2014)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "text", + "content": "32. Linardos, A., Kümmerer, M., Press, O., Bethge, M.: DeepGaze IIE: Calibrated prediction in and out-of-domain for state-of-the-art saliency modeling. In: International Conference on Computer Vision (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 545, + 480, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 480, + 577 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 480, + 577 + ], + "type": "text", + "content": "33. Liu, N., Han, J., Yang, M.H.: PiCANet: Learning pixel-wise contextual attention for saliency detection. In: Conference on Computer Vision and Pattern Recognition (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 578, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 480, + 610 + ], + "type": "text", + "content": "34. Liu, T., Yuan, Z., Sun, J., Wang, J., Zheng, N., Tang, X., Shum, H.Y.: Learning to detect a salient object. IEEE Transactions on Pattern analysis and machine intelligence 33(2) (2010)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "text", + "content": "35. Liu, Y., Wu, Y.H., Wen, P., Shi, Y., Qiu, Y., Cheng, M.M.: Leveraging instance-, image-and dataset-level information for weakly supervised instance segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(3) (2020)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "36. Ma, J., Ming, A., Huang, Z., Wang, X., Zhou, Y.: Object-level proposals. In: International Conference on Computer Vision (2017)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 101 + ], + "type": "text", + "content": "C. Wilms et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "37. Martin García, G., Potapova, E., Werner, T., Zillich, M., Vincze, M., Frintrop, S.: Saliency-based object discovery on RGB-D data with a late-fusion approach. In: International Conference on Robotics and Automation (2015)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 481, + 183 + ], + "type": "text", + "content": "38. Mazurowski, M.A., Dong, H., Gu, H., Yang, J., Konz, N., Zhang, Y.: Segment anything model for medical image analysis: an experimental study. Medical Image Analysis 89 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 184, + 481, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 184, + 481, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 184, + 481, + 217 + ], + "type": "text", + "content": "39. Perazzi, F., Krahenbuhl, P., Pritch, Y., Hornung, A.: Saliency filters: Contrast based filtering for salient region detection. In: Conference on Computer Vision and Pattern Recognition (2012)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 218, + 481, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 218, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 218, + 481, + 239 + ], + "type": "text", + "content": "40. Pinheiro, P.O., Collobert, R., Dollár, P.: Learning to segment object candidates. Advances in Neural Information Processing Systems 28 (2015)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 240, + 481, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 240, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 240, + 481, + 262 + ], + "type": "text", + "content": "41. Pinheiro, P.O., Lin, T.Y., Collobert, R., Dollár, P.: Learning to refine object segments. In: European Conference on Computer Vision (2016)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 263, + 481, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 263, + 481, + 306 + ], + "spans": [ + { + "bbox": [ + 132, + 263, + 481, + 306 + ], + "type": "text", + "content": "42. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 307, + 481, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 307, + 481, + 329 + ], + "spans": [ + { + "bbox": [ + 132, + 307, + 481, + 329 + ], + "type": "text", + "content": "43. Rahtu, E., Kannala, J., Blaschko, M.: Learning a category independent object detection cascade. In: International Conference on Computer Vision (2011)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 330, + 481, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 330, + 481, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 330, + 481, + 363 + ], + "type": "text", + "content": "44. Ronneberger, O., Fischer, P., Brox, T.: U-Net: Convolutional networks for biomedical image segmentation. In: International Conference on Medical Image Computing and Computer-Assisted Intervention (2015)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 363, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 363, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 363, + 481, + 396 + ], + "type": "text", + "content": "45. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al.: Imagenet large scale visual recognition challenge. International Journal of Computer Vision 115 (2015)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 396, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 481, + 418 + ], + "type": "text", + "content": "46. Saito, K., Hu, P., Darrell, T., Saenko, K.: Learning to detect every thing in an open world. In: European Conference on Computer Vision (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 419, + 481, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 419, + 481, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 419, + 481, + 441 + ], + "type": "text", + "content": "47. Shi, Z., Sun, Y., Zhang, M.: Training-free object counting with prompts. In: Winter Conference on Applications of Computer Vision (2024)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 441, + 481, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 441, + 481, + 474 + ], + "spans": [ + { + "bbox": [ + 132, + 441, + 481, + 474 + ], + "type": "text", + "content": "48. Wang, C., Wang, G., Zhang, Q., Guo, P., Liu, W., Wang, X.: OpenInst: A simple query-based method for open-world instance segmentation. arXiv preprint arXiv:2303.15859 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 475, + 481, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 481, + 508 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 481, + 508 + ], + "type": "text", + "content": "49. Wang, W., Feiszli, M., Wang, H., Malik, J., Tran, D.: Open-world instance segmentation: Exploiting pseudo ground truth from learned pairwise affinity. In: Conference on Computer Vision and Pattern Recognition (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 509, + 481, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 509, + 481, + 541 + ], + "spans": [ + { + "bbox": [ + 132, + 509, + 481, + 541 + ], + "type": "text", + "content": "50. Wang, W., Feiszli, M., Wang, H., Tran, D.: Unidentified video objects: A benchmark for dense, open-world segmentation. In: International Conference on Computer Vision (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 542, + 481, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 542, + 481, + 564 + ], + "spans": [ + { + "bbox": [ + 132, + 542, + 481, + 564 + ], + "type": "text", + "content": "51. Wei, Z., Chen, P., Yu, X., Li, G., Jiao, J., Han, Z.: Semantic-aware SAM for point-prompted instance segmentation. arXiv preprint arXiv:2312.15895 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 565, + 481, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 565, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 565, + 481, + 586 + ], + "type": "text", + "content": "52. Wilms, C., Frintrop, S.: Edge adaptive seeding for superpixel segmentation. In: German Conference on Pattern Recognition (2017)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 587, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 587, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 587, + 481, + 609 + ], + "type": "text", + "content": "53. Wilms, C., Frintrop, S.: AttentionMask: Attentive, efficient object proposal generation focusing on small objects. In: Asian Conference on Computer Vision (2019)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "type": "text", + "content": "54. Wu, J., Jiang, Y., Yan, B., Lu, H., Yuan, Z., Luo, P.: Exploring transformers for open-world instance segmentation. In: International Conference on Computer Vision (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "text", + "content": "55. Xie, C., Xiang, Y., Mousavian, A., Fox, D.: Unseen object instance segmentation for robotic environments. IEEE Transactions on Robotics 37(5) (2021)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 325, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 325, + 91, + 447, + 102 + ], + "type": "text", + "content": "SOS: Segment Object System" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 347 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "56. Xue, X., Yu, D., Liu, L., Liu, Y., Li, Y., Yuan, Z., Song, P., Shou, M.Z.: Single-stage open-world instance segmentation with cross-task consistency regularization. arXiv preprint arXiv:2208.09023 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 171 + ], + "type": "text", + "content": "57. Yang, X., Gong, X.: Foundation model assisted weakly supervised semantic segmentation. In: Winter Conference on Applications of Computer Vision (2024)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "type": "text", + "content": "58. Yang, Y., Zhou, Z., Wu, J., Wang, Y., Xiong, R.: Class semantics modulation for open-set instance segmentation. IEEE Robotics and Automation Letters (2024)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 225 + ], + "type": "text", + "content": "59. Zhao, J.X., Liu, J.J., Fan, D.P., Cao, Y., Yang, J., Cheng, M.M.: EGNet: Edge guidance network for salient object detection. In: International Conference on Computer Vision (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "text", + "content": "60. Zhou, B., Khosla, A., Lapedriza, A., Oliva, A., Torralba, A.: Learning deep features for discriminative localization. In: Conference on Computer Vision and Pattern Recognition (2016)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 480, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 292 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 292 + ], + "type": "text", + "content": "61. Zhou, B., Zhao, H., Puig, X., Xiao, T., Fidler, S., Barriuso, A., Torralba, A.: Semantic understanding of scenes through the ADE20K dataset. International Journal of Computer Vision 127 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "text", + "content": "62. Zhu, M., Li, H., Chen, H., Fan, C., Mao, W., Jing, C., Liu, Y., Shen, C.: SegPrompt: Boosting open-world segmentation via category-level prompt learning. In: International Conference on Computer Vision (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 325, + 480, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 480, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 480, + 347 + ], + "type": "text", + "content": "63. Zitnick, C.L., Dollár, P.: Edge boxes: Locating object proposals from edges. In: European Conference on Computer Vision (2014)" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "type": "text", + "content": "C. Wilms et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/SPAMming Labels_ Efficient Annotations for the Trackers of Tomorrow/0cf5ca3c-0b18-48c1-b497-d85ab9c4bdac_content_list.json b/2024/SPAMming Labels_ Efficient Annotations for the Trackers of Tomorrow/0cf5ca3c-0b18-48c1-b497-d85ab9c4bdac_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b6b738ea165ffec7a0a4773a6d46ed1ef6dfd8a2 --- /dev/null +++ b/2024/SPAMming Labels_ Efficient Annotations for the Trackers of Tomorrow/0cf5ca3c-0b18-48c1-b497-d85ab9c4bdac_content_list.json @@ -0,0 +1,1582 @@ +[ + { + "type": "text", + "text": "SPAMming Labels: Efficient Annotations for the Trackers of Tomorrow", + "text_level": 1, + "bbox": [ + 266, + 140, + 736, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Orcun Cetintas $^{1,2\\dagger}$ , Tim Meinhardt $^{2}$ , Guillem Brasó $^{1,2\\dagger}$ , and Laura Leal-Taixe $^{2}$", + "bbox": [ + 215, + 210, + 785, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Technical University of Munich \n $^{2}$ NVIDIA", + "bbox": [ + 388, + 238, + 612, + 265 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Increasing the annotation efficiency of trajectory annotations from videos has the potential to enable the next generation of data-hungry tracking algorithms to thrive on large-scale datasets. Despite the importance of this task, there are currently very few works exploring how to efficiently label tracking datasets comprehensively. In this work, we introduce SPAM, a video label engine that provides high-quality labels with minimal human intervention. SPAM is built around two key insights: i) most tracking scenarios can be easily resolved. To take advantage of this, we utilize a pre-trained model to generate high-quality pseudo-labels, reserving human involvement for a smaller subset of more difficult instances; ii) handling the spatiotemporal dependencies of track annotations across time can be elegantly and efficiently formulated through graphs. Therefore, we use a unified graph formulation to address the annotation of both detections and identity association for tracks across time. Based on these insights, SPAM produces high-quality annotations with a fraction of ground truth labeling cost. We demonstrate that trackers trained on SPAM labels achieve comparable performance to those trained on human annotations while requiring only $3 - 20\\%$ of the human labeling effort. Hence, SPAM paves the way towards highly efficient labeling of large-scale tracking datasets. We release all models and code.", + "bbox": [ + 261, + 301, + 743, + 592 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Multi-object tracking $\\cdot$ Data labeling $\\cdot$ Active learning", + "bbox": [ + 261, + 604, + 712, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 665, + 375, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Detecting and tracking objects in time is at the core of many challenging vision applications. To address these problems, modern multi-object tracking (MOT) approaches [44, 66, 72, 76, 78] rely on increasingly larger amounts of annotated data. While annotating image datasets is already costly, temporal component of videos further increases the annotation difficulty and data scale requirements. This is due to the fact that redundancies between frames cause the information density to scale poorly with the amount of data, making the overall annotation task more challenging and resource-intensive.", + "bbox": [ + 212, + 696, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Work done during an internship at NVIDIA.", + "bbox": [ + 236, + 825, + 547, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To mitigate these issues, researchers are spending more effort in operating with limited human annotations. Self- and semi-supervised learning (SSL) and pseudo-labeling (PL) approaches are often used to leverage unlabeled data whenever possible [34,35,40], while active learning (AL) strategies aim at selecting the most informative samples to label. All these strategies are commonly applied in various image understanding tasks, such as classification [11,14] or object detection [17,67]. Even if video tasks do not have any lower requirement for data, there are only a handful of works that target efficient labeling of tracking data in the video domain [39,59-61]. Existing works fail to provide comprehensive solutions to labeling by either ignoring dense temporal component of videos [39,61] or working under a limited single-object setup [71]. An effective video labeling engine should operate across multiple objects and frames, considering the intricate dependencies between tracks inherent in the labeling problem.", + "bbox": [ + 212, + 146, + 787, + 343 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Videos have unique characteristics that can be leveraged to unleash the full potential of a modern label engine. Firstly, frame-wise similarities in videos yield data redundancies. In the context of MOT, this implies that some associations between detections are relatively easy to solve [7]. This setup provides an ideal opportunity to leverage state-of-the-art models pre-trained to pseudo-label new data at zero cost. Secondly, the dependencies of objects across time and space cause annotations in one frame of the video to have cascading effects in other frames. For example, solving an association for one track can resolve associations for neighboring tracks. In order to maximize the efficiency of the annotations, we frame the labeling task within a model that takes into account such dependencies. Thus, we adopt a track-centric labeling approach rather than a frame-centric one [39].", + "bbox": [ + 212, + 347, + 787, + 529 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We present the SPAM video label engine, which chains Synthetic pretraining, Pseudo-labeling, and Active learning with a graph-based Model for optimal tracking performance with minimal human effort. In the image domain, there are plenty of large-scale annotated datasets that a model-driven approach can leverage for pre-training [34]. However, this is not the case for multi-object tracking datasets, which are typically much smaller in scale; therefore, we advocate for a Synthetic pre-training. We further analyze which components of a tracking method are more affected by the synthetic-to-real gap. This first experiment gives us an important insight: labeling should be focused on detections and associations, and we can directly use the synthetically trained reID model. With our synthetic pre-trained model, we then propose a model-driven approach to produce Pseudo-labels on real data and retrain the model with its own pseudolabels. We show that with this recipe we can even reach a performance comparable to ground truth without ever training our labeler on manually annotated data.", + "bbox": [ + 212, + 532, + 787, + 760 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Synthetic pre-training and pseudo-labeling allow us to handle easy decisions and defer only uncertain or complex decisions to an annotator. We propose an Active learning strategy where the labeler makes decisions at the track-level instead of frame-level, thereby utilizing the annotation budget more efficiently. Towards this goal, we solve the labeling problem with a Model, based on graph", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "O. Cetintas et al.", + "bbox": [ + 271, + 114, + 388, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "hierarchies [12]. This model captures long-range spatio-temporal dependencies, which means our annotations will have an impact across frames, and not only on the frame where the annotation actually happens. We extend the model with a detection filtering layer so that we can naturally label detection and association within the same framework.", + "bbox": [ + 212, + 146, + 782, + 219 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, we propose SPAM, a video label engine that significantly reduces annotation cost by dealing with easy scenarios with a strong MOT model pre-trained on (i) Synthetic data, followed by (ii) Pseudo-labeling real data for its own re-training. Harder detection and association decisions are carefully selected by an (iii) Active learning scheme that works on a (iv) Model based on graph hierarchies. This allows us to efficiently use the annotator's budget and account for spatio-temporal dependencies in the labels. The resulting model generates labels comparable to ground truth performance, requiring only $3 - 20\\%$ of manual annotation effort across three diverse datasets.", + "bbox": [ + 212, + 222, + 784, + 357 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 382, + 387, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Pre-training for tracking. Over the years, the multi-object tracking (MOT) community has annotated many challenging benchmark datasets [18, 19, 57]. However, due to the significant costs associated with manually labeling temporal data, most benchmarks only provide few annotated video sequences. To obtain top performance, modern data-hungry tracking methods are usually pretrained on additional video and tracking datasets [20, 22, 68, 74, 77]. Many works also rely on relevant image recognition [56] datasets to boost their per-frame object detection performance. An alternative pre-training direction avoids human annotations entirely by relying on simulated synthetic data [23, 24, 26, 29, 58].", + "bbox": [ + 212, + 415, + 782, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our label engine also pre-trains the labeling model synthetically, minimizing its total annotation footprint. To this end, we demonstrate the effectiveness of synthetic pre-training [23] for generating pseudo-labels for self-supervision.", + "bbox": [ + 212, + 551, + 784, + 597 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Self-supervision in the temporal domain. To benefit not only from already annotated datasets, many works explore the utilization of available unlabeled data sources, i.e., how to perform self-supervised learning for videos. Notable examples leverage cycle-consistency-based correspondences [30, 64], color propagation [63] or augmentation-based consistencies [4, 41]. Furthermore, [45] uses contrastive learning over re-identification (reID) embeddings to perform unsupervised association learning. More recently, several works [33-35, 40] in the image domain demonstrate the effectiveness of training on pseudo-labels generated by strong underlying labeling models. In the video domain, this type of self-supervision is mostly unexplored.", + "bbox": [ + 212, + 598, + 784, + 748 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We demonstrate the effectiveness of pseudo-labels for the temporal domain by improving the label quality of our SPAM labeler via self-supervision on its own output. The necessary signal-to-noise ratio for this training loop is obtained by pre-training a strong graph-based model on synthetic data.", + "bbox": [ + 212, + 750, + 782, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Efficient labeling of video datasets. While significant effort has been devoted towards increasing the labeling efficiency of image datasets [1,34], similar", + "bbox": [ + 214, + 809, + 782, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "SPAMming Labels: Efficient Annotations for the Trackers of Tomorrow", + "bbox": [ + 256, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "approaches for the video and tracking domain are still relatively limited. Earlier works reduce the number of frames to annotate within individual videos but still rely heavily on human annotators [62]. Other solutions are limited to only a single object per sequence [16, 60] or still rely on human bounding box annotations [59] to label segmentation masks across time. The authors of [39] select key frames based on their expected impact on tracking performance. While effective within its scope, this approach only yields sparse frame annotations, which are insufficient for training all modern trackers. To reduce the human annotation effort for trajectories, [43] proposes an efficient video interface that still requires labeling all trajectories.", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In contrast, our label engine performs active learning for multiple objects across all video frames and is the first of its kind to tackle long-term tracking problems via a hierarchical labeling approach. The uncertainty measures provided by our graph-based model allow us to focus expensive annotation efforts on individual tracks instead of labeling entire frames in a brute-force manner.", + "bbox": [ + 212, + 297, + 787, + 372 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multi-object tracking models. Tracking objects across time in a video is an established problem with a long research history. The most dominant tracking-by-detection (TbD) paradigm splits the task into two steps: (i) detecting objects in every frame and (ii) associating them to object trajectories, also referred to as data association. Modern trackers have explored performing (i) and (ii) jointly [7, 25, 44, 78], or providing solutions for (ii) relying on motion models [2, 3, 38, 47, 49, 52, 53, 69, 75], appearance models [13, 15, 36, 46, 51, 54, 70, 76] or more complex techniques involving transformers [10, 72]. An alternative direction of work relies on graph formulations [6, 31, 37, 48, 73] to utilize the TbD paradigm. In this formulation, nodes model object detections, and edges, association hypotheses. This long-studied formulation has recently been revived with remarkable success with the help of learnable graph neural network (GNN)architectures [8, 9, 12, 65].", + "bbox": [ + 212, + 375, + 787, + 555 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For our label engine to generate long-term pseudo-labels and uncertainty estimations, we rely on a strong graph-based tracker [12]. To incorporate both detection and association labeling within a unified graph framework, we introduce a novel hierarchy level dedicated to detection filtering.", + "bbox": [ + 212, + 555, + 787, + 614 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 SPAM", + "text_level": 1, + "bbox": [ + 215, + 641, + 316, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Overview", + "text_level": 1, + "bbox": [ + 215, + 675, + 339, + 690 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our main goal is to label multi-object tracking datasets with minimal annotation effort while enabling optimal tracking performance. Our key observation is that most association cases can be solved easily by models pre-trained on synthetic data. Therefore, no human effort is needed, and we can instead use our model's predictions as pseudo-labels. Decisions that are more complex or uncertain are selected by an active learning scheme to be annotated manually.", + "bbox": [ + 212, + 703, + 787, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In order to perform pseudo-labeling and active learning, we require a model that can capture long-term spatio-temporal dependencies between tracks. We leverage a hierarchical graph neural network formulation and cast the annotation", + "bbox": [ + 212, + 794, + 787, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "O. Cetintas et al.", + "bbox": [ + 271, + 114, + 388, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e885ba6615bc01047af4fb79ade4fdf2af243db6d58bab58dd2e78978747dc2f.jpg", + "image_caption": [ + "Fig. 1: Overview of the SPAM model. We first generate a set of detection candidates with our detector. Hierarchical GNNs then classify these candidates into valid and invalid objects via node classification, and assign identities through edge classification." + ], + "image_footnote": [], + "bbox": [ + 222, + 148, + 787, + 247 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "task over a graph. Our architecture (Fig. 1) draws inspiration from SUSHI [12], which demonstrated remarkable association performance. Our model's scope, however, goes beyond association. Given the constraints of the labeling task, we need to equip our model with the ability to reason about detections before performing association. To address this broader goal, our GNN hierarchy, also includes the capability to accept or reject detections by classifying nodes into correct and incorrect object hypotheses.", + "bbox": [ + 212, + 330, + 782, + 435 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our overall model is pre-trained on synthetic data [23] and further fine-tuned on its own predicted pseudo-labels over the target dataset that we aim to annotate. This recipe allows for strong labelling performance without any human annotations, and provides the basis for a scalable video label engine. After this training loop, we use our updated model's predictions as pseudo-labels to annotate a large portion of the target dataset. Decisions that are more complex or uncertain are selected by an active learning scheme to be annotated manually. The resulting set of labels achieves near-GT performance with significantly fewer manual annotations, thereby vastly enhancing the efficiency of the overall process. We provide an overview of our training and annotating recipe in Fig. 2.", + "bbox": [ + 212, + 436, + 787, + 587 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Graph Hierarchies as a Model for Labeling", + "text_level": 1, + "bbox": [ + 214, + 606, + 617, + 622 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Problem formulation. Our model builds upon the standard graph formulation [73] where, given a set of object candidates $\\mathcal{O}$ , our goal is to obtain a set of valid objects $\\mathcal{O}_v \\subset \\mathcal{O}$ and their corresponding trajectories $\\mathcal{T}$ . Each trajectory $T_k \\in \\mathcal{T}$ consists of a set of objects sharing the same identity or, equivalently, the set of edges connecting those nodes. Specifically, we model each object candidate $o_i \\in \\mathcal{O}$ as a node and each association hypothesis among candidates as an edge in an undirected graph $G = (V, E)$ . Within this formulation, nodes $u \\in V$ could be classified into valid objects if $u \\in \\mathcal{O}_v$ . Analogously, edges can be classified into correct or incorrect hypotheses depending on whether they belong to some trajectory or not. This formulation can be generalized to a hierarchical setting by splitting video sequences into non-overlapping subsequences, i.e., disjoint subgraphs, and progressively merging them [12].", + "bbox": [ + 212, + 628, + 787, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Background on graph hierarchies for MOT. Our model consists of a hierarchy of GNNs operating over the graph formulation described above. Intuitively,", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "SPAMming Labels: Efficient Annotations for the Trackers of Tomorrow", + "bbox": [ + 256, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/beb0774fcc317056f4f2f0a806b15ef73c5ecce5e7f19eff5c52537c02fa6a01.jpg", + "image_caption": [ + "Fig. 2: Overview of the SPAM training and annotation pipeline. (a) Initial model training on synthetic data. (b) Application of SPAM to generate pseudo-labels without incurring manual annotation costs on a real dataset, followed by self-training on pseudolabels. (c) Real dataset labeling using pseudo-labels and an uncertainty-based active learning approach." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 781, + 234 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "starting from detection candidates, at each hierarchy level, GNN hierarchy subsequently merge tracklets from the previous level into longer ones. Nodes and edges are represented by initial embeddings and message passing is used to share information within the graph in order to update node and edge embeddings with richer information. The GNNs are trained to classify edges into active and inactive association hypotheses and they share the same features, network architecture and learnable weights across hierarchy levels. In addition to the edge features reported in [12], for each object candidate $o_i$ we concatenate its times-tamp $t_i$ , normalized bounding box coordinates and dimensions $(x_i, y_i, h_i, w_i)$ and corresponding confidence from the detector $c_i$ and embed it as node features.", + "bbox": [ + 212, + 357, + 787, + 508 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Classifying nodes. We aim to minimize the annotation effort; therefore, we ask the annotator to operate on a fixed graph of detection candidates, i.e. annotator does not label new detections after the initial set given by detector. This is why we aim to create an initial graph that is over-complete. In order to do so, we use a low confidence threshold for our detector and obtain a large number of detection candidates. While this ensures a high detection recall, it also introduces a larger number of false positive detections that our model needs to address. As we show in experiments, the original SUSHI architecture [12] cannot handle such false positives, which is why we propose a new layer in the hierarchy, $GNN_{node}$ . Intuitively, $GNN_{node}$ leverages temporal information to validate object candidates. We use its output node embeddings and feed them to a lightweight classifier to enable false positive detection filtering. By doing so, we can consider a large number of input proposals achieving high detection recall by leveraging the spatio-temporal reasoning ability of our GNN for detection candidates.", + "bbox": [ + 212, + 515, + 787, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pre-training and annotation-free fine-tuning. We pre-train all of our model components on the large-scale synthetic dataset MOTSynth [23]. To boost our model's performance and bridge the domain gap with our target dataset for annotation, we conduct further fine-tuning on the target dataset after pre-training. We achieve this by utilizing the raw predictions of our model on the target dataset and employing them as pseudo-labels for self-training. As we show in", + "bbox": [ + 212, + 748, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "O. Cetintas et al.", + "bbox": [ + 271, + 114, + 388, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9bf0b4315066d528c3b5d6ba89673473ac6f5be91aa56f407adc960c1d179393.jpg", + "image_caption": [ + "Fig. 3: Our graph-based labeling pipeline begins with the selection of nodes for annotation. For each node to be annotated, the annotator could be asked to validate the detection, improve the localization by refining the box or perform association." + ], + "image_footnote": [], + "bbox": [ + 267, + 143, + 743, + 276 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "our experimental section, this strategy yields a significant performance improvement without incurring any human annotation cost.", + "bbox": [ + 212, + 372, + 782, + 402 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 How to Label on Graph Hierarchies?", + "text_level": 1, + "bbox": [ + 214, + 430, + 565, + 445 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Manual labeling. Labeling video data at the box level involves several steps: identifying objects of interest, localizing their boundaries with bounding boxes, and assigning them unique identities. Typically, annotators achieve this by examining each frame, detecting and localizing objects individually, and then assigning identities based on temporal context from neighboring frames. In terms of annotation effort in clicks, detecting and localizing objects typically require two clicks (one for the top-left corner and one for the bottom-right corner of the bounding box), while associating objects across time requires one click (identifying the next occurrence of the object in subsequent frames), as reported in [18, 19]. This manual process is labor-intensive, costly, and severely limits the availability of annotated video data.", + "bbox": [ + 212, + 460, + 787, + 626 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Graph-based labeling. Utilizing a graph formulation for detection and tracking enables explicit modeling of objects in the scene and their interactions. In this framework, each node of the graph represents an object, allowing a manual annotation of an object to be framed as selecting a node on this graph. Given a set of nodes to be annotated, possible annotator decisions are: (i) accepting/rejecting a detection as we are working on a large set of detection candidates (one click), (ii) refining the bounding-box to improve localization (two clicks, one for the top-left corner and one for the bottom-right corner of the bounding box) and (iii) associating a node across time (one click) as shown in Fig. 3.", + "bbox": [ + 212, + 628, + 787, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Annotation acquisition. The core idea of our approach is to use our model decisions to guide the annotation process. For the vast majority of scenarios, we directly use our model's predictions as pseudo-labels. For a small subset of more challenging decisions where our model exhibits uncertainty, we validate our decisions with an annotator as described above. Within our graphs,", + "bbox": [ + 212, + 765, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "SPAMming Labels: Efficient Annotations for the Trackers of Tomorrow", + "bbox": [ + 256, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c1ef6f6f66be5c046820990ba3d4e32bd29615bb445768664948ba57b75b8fbf.jpg", + "image_caption": [ + "Fig. 4: Analysis of performance gap between training a model on synthetic and real data for the three most common tracking components: detection, association, re-identification." + ], + "image_footnote": [], + "bbox": [ + 307, + 151, + 697, + 333 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "we adopt an uncertainty-based approach to identify which nodes require annotations. Specifically, for each node $v \\in V$ , we compute its uncertainty as $\\mathrm{uncert}(v) \\coloneqq \\max_{u \\in \\mathbb{N}_v} H(\\hat{y}_{(v,u)})$ , where $\\mathbb{N}_v$ denotes the set of neighboring nodes of $v$ and $\\hat{y}_{(v,u)}$ is the edge prediction for edge $(v,u)$ :", + "bbox": [ + 212, + 414, + 787, + 477 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nH (\\hat {y} _ {(v, u)}) := - (\\hat {y} _ {(v, u)} \\log \\hat {y} _ {(v, u)} + (1 - \\hat {y} _ {(v, u)}) \\log (1 - \\hat {y} _ {(v, u)})) \\qquad (1)\n$$\n", + "text_format": "latex", + "bbox": [ + 289, + 493, + 785, + 512 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "At the $GNN_{node}$ level, where we have node predictions instead of edge predictions, we simply compute $\\text{uncert}(v)$ as $H(\\hat{y}_{(v)})$ instead. Subsequently, we refer the nodes with the highest uncertainty scores to the annotator for manual annotation and use the remaining predictions as pseudo-labels for the rest of the nodes.", + "bbox": [ + 212, + 518, + 787, + 593 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Hierarchical label generation. To fully leverage our hierarchical graph formulation, we apply our label acquisition approach described above in a hierarchical manner. Intuitively, we distribute our annotation budget $B$ across $L$ hierarchical levels as $B_{1},\\ldots ,B_{L}$ , with $B_{1} + \\dots +B_{L} = B$ . In deeper hierarchy levels, nodes represent tracklets instead of single detections. This allows us to propagate annotator decisions for the entire cluster instead of individual detections, resulting in solving multiple uncertainties with a single annotation. As a result, we utilize the annotation budget more effectively while also mitigating error propagation within our pseudo-labels.", + "bbox": [ + 212, + 595, + 787, + 731 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 758, + 377, + 776 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Datasets and Metrics", + "text_level": 1, + "bbox": [ + 214, + 795, + 439, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Datasets. We conduct experiments on three public benchmarks:", + "bbox": [ + 214, + 824, + 684, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "O. Cetintas et al.", + "bbox": [ + 271, + 114, + 388, + 127 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- MOT17 [18] is an established benchmark containing 14 challenging video sequences, exhibiting diverse characteristics, including variations in camera motion, viewpoint, and pedestrian densities.", + "- MOT20 [19] contains a total of 8 video sequences, all featuring extremely crowded scenes. Due to their high pedestrian density, the emphasis is on addressing highly challenging scenarios characterized by frequent and prolonged occlusions.", + "- DanceTrack [57] contains 100 video sequences. These sequences feature dancing individuals with similar appearances, facilitating the evaluation of tracker robustness against highly challenging motion patterns where appearance information has a relatively minor influence. This non-standard setup often leads to a performance drop in trackers [57]." + ], + "bbox": [ + 223, + 145, + 785, + 328 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Metrics. We assess the performance of trackers using three established metrics:", + "bbox": [ + 215, + 339, + 782, + 354 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- MOTA [32] assesses trajectory coverage, by combining object detection measurements with identity switches, with the latter playing a relatively minor role.", + "- IDF1 [50] focuses on identity preservation over the sequences and provides a measure of association performance.", + "- HOTA [42] is a recently introduced metric designed to strike a balance between association and object detection performance within a single score." + ], + "bbox": [ + 223, + 367, + 785, + 472 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 494, + 455, + 508 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Architecture and training. Our model consists of three main components: (i) a detector network, based onYOLOX [27], (ii) a ResNet-based ReID network [28], and (iii) a GNN hierarchy, all pretrained on the synthetic sequences of MOTSynth [23]. After pretraining, we use the resulting model to obtain pseudolabels for our target MOT dataset, and then fine-tune our detector and GNN on them. Note that our Model is never trained on human annotations, it is only trained on synthetic annotations and its own pseudo-labels.", + "bbox": [ + 215, + 520, + 785, + 625 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Experimental setup. Online trackers are real-time tools suitable for deployment on robots or autonomous vehicles, whereas offline trackers are better suited for labeling tasks. Therefore, we structure our experiments to utilize SPAM for labeling and use these labels with online trackers. To evaluate the quality of SPAM labels, we train two state-of-the-art online trackers with these labels: ByteTrack [75] and GHOST [54]. For validation, we follow the standard half-split setting [54, 75, 76] over MOT17 sequences. Annotation budgets are computed based on number of clicks as described in Sec. 3.3.", + "bbox": [ + 215, + 626, + 785, + 746 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3 Synthetic Pretaining", + "text_level": 1, + "bbox": [ + 215, + 768, + 433, + 784 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Several works have shown that synthetic pre-training is a strong starting point for multi-object tracking methods [23, 75]. In this section, we aim to analyze the gap between training a model on synthetic data and training it with real", + "bbox": [ + 215, + 794, + 785, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "SPAMming Labels: Efficient Annotations for the Trackers of Tomorrow", + "bbox": [ + 256, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/3da6dd36b8f91b5cfe7483fb121abd6586fc083974d91b53ae2c76394d104123.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelHigh Conf. BoxLow Conf. Box\\( GNN_{node} \\)MOTA ↑IDF1 ↑HOTA ↑DetA ↑
A✓✘✘64.474.759.955.9
B✓✓✘60.671.458.554.9
C✓✓✓65.475.160.456.6
", + "bbox": [ + 220, + 143, + 787, + 214 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 1: Ablation study investigating the impact of incorporating low confidence bounding boxes with and without the $GNN_{node}$ .", + "bbox": [ + 215, + 217, + 784, + 244 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "data for the three most common tracking components: detection, association, re-identification. For each setting, only the specific component that we are analysing is trained on synthetic data while other components are all trained on real data. We perform this experiment on three datasets and plot the relative HOTA difference (a positive score indicates superior performance with real data training) when training on real data vs. synthetic data in Fig. 4. Our first observation, is that the detector is the component that suffers the domain gap the most, with a performance drop of 9.9 HOTA points. Re-identification, on the other hand, has a similar performance when trained on synthetic or real data. Lastly, association performance is 2.1 HOTA higher with real data training. This first experiment gives us an important insight: labeling and retraining should be focused on detection and association components, and we can directly use the synthetically trained reID model.", + "bbox": [ + 212, + 263, + 787, + 459 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.4 Towards a Model for Labeling", + "text_level": 1, + "bbox": [ + 215, + 482, + 511, + 498 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Graph-based validation of detection candidates. In our framework, annotators have the capability to eliminate detections, but we refrain from requesting them to annotate new detections in order to maintain minimal annotation effort. Instead we start from an initial graph formed by an over-complete set of detections. Towards this end, we use a low confidence threshold for our detector, thereby reducing the number of false negatives in our initial graph and ensuring a high recall. As a result, this leads to an increase in the number of false positives incorporated into the graph, a tendency that is further accentuated by the fact that our detector is exclusively trained on synthetic data. We propose to leverage the graph structure itself to validate the detections based on temporal consistency. Specifically, we add an additional GNN-layer in our hierarchy, $GNN_{node}$ , which intuitively examines the spatial and temporal consistency of the bounding box candidates and filters out implausible ones. In Tab. 1, we assess the effect of this design choice. Model A uses only high confidence bounding boxes, model B creates a graph with also low confidence bounding boxes but without $GNN_{node}$ , while C includes $GNN_{node}$ . As we can see, simply adding low confidence boxes results in a drastic drop in performance on all metrics. Only by adding our $GNN_{node}$ we are able to take full advantage of the low confidence boxes without penalizing our performance with too many errors.", + "bbox": [ + 212, + 507, + 787, + 792 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "SPAM is a SOTA tracker. Before moving to labeling with our model, we want to verify its overall performance and generalization capabilities. Only by using a state-of-the-art model we will be able to reduce annotation effort as much as", + "bbox": [ + 212, + 795, + 787, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "O. Cetintas et al.", + "bbox": [ + 271, + 114, + 388, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/bfbe9dbf7a7db592aa3db48bf834917b5d1922ebd5821a2d8690a0174b4c986a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodHOTA ↑MOTA ↑IDF1 ↑HOTA ↑MOTA ↑IDF1 ↑HOTA ↑MOTA ↑IDF1 ↑
MOT17MOT20DanceTrack
MeMOT [10]56.972.569.054.166.163.7---
TrackFormer [44]57.374.168.054.765.768.6---
MOTR [72]57.868.673.4---54.251.579.7
FairMOT [76]59.373.772.354.661.867.339.782.240.8
QDTrack [46]63.578.777.560.074.773.854.287.750.4
ByteTrack† [75]62.878.977.160.474.274.547.789.653.9
GHOST [54]62.878.777.161.273.775.256.791.357.7
UTM [70]64.081.878.762.578.276.9---
MotionTrack [49]65.181.180.162.878.076.5---
SUSHI [12]66.581.183.164.374.379.863.388.763.4
SPAM (Ours)67.580.784.665.876.581.964.089.263.4
", + "bbox": [ + 220, + 143, + 787, + 314 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 2: Test set results on MOT17, MOT20 and DanceTrack when using SPAM solely as a tracker.", + "bbox": [ + 215, + 315, + 782, + 342 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "possible. We therefore pre-train our model on the synthetic dataset, and finetune it on the specific real dataset using groundtruth. For a fair comparison, we obtain object detections from a YOLOX detector [27] trained following [12, 54, 75]. Our model outperforms state-of-the-art by 1.0 HOTA point on MOT17, 1.5 HOTA points on MOT20, and 0.7 HOTA points on DanceTrack as shown in Tab. 2. Furthermore, on MOT17 and MOT20, there is also a significant improvement in IDF1, showing the identity-preserving capabilities of the model, which is especially important for our video annotation setting. Even if being state-of-the-art is not the main goal of this paper, having a strong performing model allows us to have an excellent starting point for producing pseudo-labels. We would like to emphasize that, we use ground truth annotations only for this experiment just to compare our model fairly with other trackers. From this point onward, we will not use ground truth labels. We will use our synthetic pretraining weights and focus on the pseudo-labeling and active learning aspects of our method to show how to annotate multi-object tracking datasets.", + "bbox": [ + 212, + 356, + 787, + 583 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.5 Enhancing Annotations with Pseudo-labels and Active Learning", + "text_level": 1, + "bbox": [ + 215, + 604, + 785, + 619 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Retraining our Model on Pseudo-labels. In this experiment, we compare two models. The first one, SPAM, is only trained on synthetic data. The second one, SPAM, is trained in three stages: (i) pre-training on synthetic data, (ii) pseudo-labeling the real dataset with our model, and (iii) re-training our model on pseudo-labels. In Sec. 4.5, we observe a significant improvement when using pseudo-labels to re-train our model: an increase of 4-6 HOTA, 4-6 MOTA, and 4-7 IDF1 points on MOT17, MOT20, and DanceTrack, respectively. Note that these improvements come at zero human annotation effort as this process requires no manual annotations.", + "bbox": [ + 212, + 628, + 787, + 763 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Active learning on graph hierarchies. We now test the effectiveness of our proposed active learning labeling on graph hierarchies, and show a comprehensive comparison with several baselines in Fig. 5. We directly measure the accuracy of the output labels given an increasing percentage of manual annotations selected", + "bbox": [ + 212, + 779, + 787, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "SPAMming Labels: Efficient Annotations for the Trackers of Tomorrow", + "bbox": [ + 256, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/968593e5c39ca9296768bd4e3599935b963741a9d00a241671d03e4427ccbfae.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelPseudo-labelHOTA ↑MOTA ↑IDF1 ↑
MOT17
SPAM✘60.065.375.1
SPAM✓63.869.279.7
MOT20
SPAM✘52.265.470.3
SPAM✓58.770.277.1
DanceTrack
SPAM✘41.865.243.4
SPAM✓48.171.050.2
", + "bbox": [ + 217, + 143, + 467, + 281 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 3: Performance boost obtained by our model when retraining with its own pseudo-labels incurring no manual annotation cost.", + "bbox": [ + 215, + 292, + 468, + 348 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/986c70849e4c91b4e5ba3eb90f0241671072a3458beb8f7a0ad06cb95ea25985.jpg", + "image_caption": [ + "Fig. 5: Impact of various active learning strategies on the quality of labeling." + ], + "image_footnote": [], + "bbox": [ + 500, + 143, + 802, + 303 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "by different acquisition functions on MOT17. The vertical axis shows the increase in HOTA over the baseline performance, i.e., the performance without active learning. Oracle is our upper bound and shows the performance if all possible labels are set to their correct value by the annotator. Every method has two different labeling schemes: one for detections (nodes), and one for associations (edges) as explained in Sec. 3.3. In addition to our method, we explore active learning strategies commonly used in object detection, including Coreset [55], Ensemble [5], and Consistency over augmented inputs [21]. The naming is based on the active learning strategy for detection in the first part and the strategy for association in the second part. For instance, EntImg/Rand. indicates image level entropy-based selection for detection and random selection for association.", + "bbox": [ + 212, + 375, + 787, + 541 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The methods shown in red shades and circle markers compare three detection labeling methods while keeping association label selection random. Box-level labeling is far superior to commonly-used image-level labeling, or random labeling. This highlights the potential of performing active learning at the object level rather than the frame level. The methods in blue shades and triangle markers all use our proposed active learning selection on graph hierarchies for the association, and compare several detection selection schemes. The Coreset [55], which prioritizes diversity of the data over uncertainty, emerges as the poorest performing approach. This observation underlines that, when accurate predictions are available, an uncertainty-based approach proves to be a more effective complement to pseudo-labels. Box-level entropy, ensembles, and consistency all perform similarly. Finally, our method, with both detection and association label selection on graph hierarchies is superior to all compared baselines, staying at only 1.2 points from Oracle performance. Notably, we also unify the AL strategy for detection and association with our approach.", + "bbox": [ + 212, + 542, + 787, + 768 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.6 Putting Everything Together: SPAM Labels in Action", + "text_level": 1, + "bbox": [ + 215, + 787, + 709, + 803 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Training trackers with our labels. In this experiment, we ablate the effects of our SPAM pipeline on the actual usecase of training a model with our pro", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "O. Cetintas et al.", + "bbox": [ + 271, + 114, + 388, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/2ab2fc260a6ec865261ca25548f8a26cce2549b0fcbdd0a8ff0d037a5249f2af.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelLabelsPseudo-labelsActive Learn.Ann. Effort ↓HOTA ↑MOTA ↑IDF1 ↑
ByteTrack [75]Ground Truth-100%52.660.465.7
ByteTrack [75]SPAMXXX49.657.863.1
ByteTrack [75]SPAM✓XX50.559.564.5
ByteTrack [75]SPAM✓✓3.3%52.561.866.2
", + "bbox": [ + 218, + 143, + 787, + 223 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 4: Effect of (i) Synthetic training, (ii) Retraining our model with our Pseudolabels, and (iii) Active learning-based acquisition and annotation of a small percentage of real data, on our label quality on the MOT17 validation set.", + "bbox": [ + 215, + 224, + 784, + 265 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "vided labels. We ablate label quality under the following settings: (i) Synthetic training, (ii) Retraining our models on our Pseudo-labels, (iii) Active learning based acquisition and annotation of a small percentage of real data. We train state-of-the-art online tracker ByteTrack [75] from scratch with the aforementioned sets of labels and report its performance on the MOT17 validation set in Tab. 4. The goal performance we want to reach is the one achieved by the model trained on groundtruth annotations (row 1). Row 2 shows our model trained only on synthetic data, which is 3 HOTA points below our goal. Using pseudo-label based training of our model increases performance by almost a HOTA point, and the ground truth level performance is achieved by also using our active learning method to select which samples to label. By annotating only $3.3\\%$ of the data, we can achieve the same results as the model trained on the ground truth data with $100\\%$ annotation effort. This shows the potential of our SPAM in reducing annotation effort for multi-object tracking datasets.", + "bbox": [ + 212, + 303, + 787, + 517 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "How much data do we really need to label? In this final experiment, our goal is to assess the impact of our entire pipeline by answering the question: How much data do we actually need to label in order to achieve the same performance as a model trained on the entire groundtruth training dataset? For that, we choose two well-established trackers, ByteTrack [75] and GHOST [54], and train them from scratch using our SPAM labels. We report their performance on MOT17, MOT20, and DanceTrack test sets in Tab. 5. We first confirm that our results on the validation set also generalize to the test set, and that we can reach GT performance with $3.3\\%$ of annotation effort on MOT17. Note that for MOT17 our annotations yield a slight improvement over GT, which can be attributed to observed annotation noise in the original GT, particularly in moving sequences (further details can be found in the supplementary material). MOT20 and DanceTrack are known to be more challenging datasets [57], which is confirmed by the fact that the trackers still miss a few points when trained on only $3.3\\%$ of manually annotated data. Given that we cannot do a full analysis of different percentages because we are working on the test sets, we choose two percentages of manually annotated data on which to train our trackers: $10\\%$ for MOT20, and $20\\%$ for DanceTrack. In both cases, we achieve on-par performance with the trackers trained on full groundtruth data. This clearly shows the potential in properly leveraging synthetic data, pseudo-labels from a strong model, and active learning on graph hierarchies in order to label multi-object tracking", + "bbox": [ + 212, + 522, + 787, + 843 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "SPAMming Labels: Efficient Annotations for the Trackers of Tomorrow", + "bbox": [ + 256, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/7b3383136a890fee3de50ab6da6cfa55ff292469678d34288ed228ee630464fb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodLabelsAnn. Effort ↓HOTA ↑MOTA ↑IDF1 ↑
MOT17 Test
ByteTrack [75]GT100%50.660.461.1
ByteTrack [75]SPAM3.3%51.664.063.0
GHOST [54]GT100%49.558.059.0
GHOST [54]SPAM3.3%51.361.962.1
MOT20 Test
ByteTrack [75]GT100%48.760.560.0
ByteTrack [75]SPAM3.3%47.957.661.4
ByteTrack [75]SPAM10%48.461.062.8
GHOST [54]GT100%48.359.760.2
GHOST [54]SPAM3.3%47.058.260.7
GHOST [54]SPAM10%46.959.060.8
DanceTrack Test
ByteTrack [75]GT100%41.079.146.0
ByteTrack [75]SPAM3.3%39.576.445.0
ByteTrack [75]SPAM20%41.380.948.1
GHOST [54]GT100%45.880.646.3
GHOST [54]SPAM3.3%41.076.344.8
GHOST [54]SPAM20%43.981.347.7
", + "bbox": [ + 277, + 143, + 730, + 489 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Table 5: Evaluation of SPAM labels on MOT17, MOT20, and DanceTrack test sets with varying annotation budgets by training ByteTrack [75] and GHOST [54] using annotations generated by our method.", + "bbox": [ + 212, + 493, + 787, + 535 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "data, where we can save $96.7\\%$ of the annotation effort on MOT17, $90\\%$ of the annotation effort on MOT20 and $80\\%$ of the annotation effort on DanceTrack.", + "bbox": [ + 212, + 559, + 787, + 590 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 613, + 359, + 628 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We have introduced SPAM, a labeling engine to efficiently obtain trajectory annotations from videos. We have shown that SPAM is able to produce annotations that yield state-of-the-art tracking performance with a significantly reduced human annotation effort. Through our extensive experiments we have shown that (i) our synthetic pretraining $+$ pseudo-label finetuning recipe yields a strong, annotation-free, labeler; (ii) our combination of active learning $+$ pseudolabeling for annotation generation yields high quality and efficient annotations; and (iii) that graphs are an excellent tool to provide comprehensive modelling for trajectory annotation. We believe that SPAM paves the way to make large-scale cheap annotation of tracking datasets a reality, which will in turn power the next generation of data-hungry tracking algorithms.", + "bbox": [ + 212, + 645, + 787, + 811 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "O. Cetintas et al.", + "bbox": [ + 271, + 114, + 388, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Acuna, D., Ling, H., Kar, A., Fidler, S.: Efficient interactive annotation of segmentation datasets with polygon-rnn++. In: CVPR (2018) 3", + "2. Alahi, A., Goel, K., Ramanathan, V., Robicquet, A., Fei-Fei, L., Savarese, S.: Social LSTM: Human trajectory prediction in crowded spaces. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2016) 4", + "3. Andriyenko, A., Schindler, K.: Multi-target tracking by continuous energy minimization. In: CVPR. pp. 1265-1272 (2011). https://doi.org/10.1109/CVPR.2011.5995311 4", + "4. Bastani, F., He, S., Madden, S.: Self-supervised multi-object tracking with cross-input consistency. Advances in Neural Information Processing Systems 34, 13695-13706 (2021) 3", + "5. Beluch, W.H., Genewein, T., Nurnberger, A., Kohler, J.M.: The power of ensembles for active learning in image classification. In: CVPR. pp. 9368-9377 (2018) 12", + "6. Berclaz, J., Fleuret, F., Turetken, E., Fua, P.: Multiple object tracking using k-shortest paths optimization. IEEE TPAMI 33(9), 1806-1819 (2011) 4", + "7. Bergmann, P., Meinhardt, T., Leal-Taixe, L.: Tracking without bells and whistles. In: ICCV. pp. 941-951 (2019) 2, 4", + "8. Brasó, G., Cetintas, O., Leal-Taixe, L.: Multi-object tracking and segmentation via neural message passing. IJCV 130(12), 3035-3053 (2022) 4", + "9. Braso, G., Leal-Taixe, L.: Learning a neural solver for multiple object tracking. In: CVPR (2020) 4", + "0. Cai, J., Xu, M., Li, W., Xiong, Y., Xia, W., Tu, Z., Soatto, S.: Memot: Multi-object tracking with memory. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8090-8100 (2022) 4, 11", + "1. Caron, M., Misra, I., Mairal, J., Goyal, P., Bojanowski, P., Joulin, A.: Unsupervised learning of visual features by contrasting cluster assignments. NeurIPS (2020) 2", + "2. Cetintas, O., Brasó, G., Leal-Taixe, L.: Unifying short and long-term tracking with graph hierarchies. In: CVPR. pp. 22877-22887 (June 2023) 3, 4, 5, 6, 11", + "3. Chen, L., Ai, H., Zhuang, Z., Shang, C.: Real-time multiple people tracking with deeply learned candidate selection and person re-identification. In: 2018 IEEE international conference on multimedia and expo (ICME). pp. 1-6. IEEE (2018) 4", + "4. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: IEEE Int. Conf. Mach. Learn. pp. 1597-1607. PMLR (2020) 2", + "5. Chu, P., Ling, H.: Famnet: Joint learning of feature, affinity and multi-dimensional assignment for online multiple object tracking. In: ICCV (October 2019) 4", + "6. Dai, K., Zhao, J., Wang, L., Wang, D., Li, J., Lu, H., Qian, X., Yang, X.: Video annotation for visual tracking via selection and refinement. In: ICCV (2021) 4", + "7. Dai, Z., Cai, B., Lin, Y., Chen, J.: Up-detr: Unsupervised pre-training for object detection with transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1601-1610 (June 2021) 2", + "8. Dendorfer, P., Osep, A., Milan, A., Schindler, K., Cremers, D., Reid, I., Roth, S., Leal-Taixe, L.: Motchallenge: A benchmark for single-camera multiple target tracking. International Journal of Computer Vision 129(4), 845-881 (2021) 3, 7, 9", + "9. Dendorfer, P., Rezatofighi, H., Milan, A., Shi, J.Q., Cremers, D., Reid, I.D., Roth, S., Schindler, K., Leal-Taix'e, L.: Mot20: A benchmark for multi object tracking in crowded scenes. ArXiv abs/2003.09003 (2020) 3, 7, 9" + ], + "bbox": [ + 225, + 176, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "SPAMming Labels: Efficient Annotations for the Trackers of Tomorrow", + "bbox": [ + 256, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "20. Dólar, P., Wojek, C., Schiele, B., Perona, P.: Pedestrian detection: A benchmark. In: CVPR. pp. 304-311. IEEE (2009) 3", + "21. Elezi, I., Yu, Z., Anandkumar, A., Leal-Taixe, L., Alvarez, J.M.: Not all labels are equal: Rationalizing the labeling costs for training object detection. In: CVPR. pp. 14492-14501 (2022) 12", + "22. Ess, A., Leibe, B., Schindler, K., Van Gool, L.: A mobile vision system for robust multi-person tracking. In: CVPR. pp. 1-8 (2008). https://doi.org/10.1109/CVPR.2008.4587581 3", + "23. Fabbri, M., Brasó, G., Maugeri, G., Cetintas, O., Gasparini, R., Osep, A., Calderara, S., Leal-Taixe, L., Cucchiara, R.: Motsynth: How can synthetic data help pedestrian detection and tracking? In: ICCV. pp. 10849-10859 (October 2021) 3, 5, 6, 9", + "24. Fabbri, M., Lanzi, F., Calderara, S., Palazzi, A., Vezzani, R., Cucchiara, R.: Learning to detect and track visible and occluded body joints in a virtual world. In: ECCV (2018) 3", + "25. Feichtenhofer, C., Pinz, A., Zisserman, A.: Detect to track and track to detect. In: ICCV (Oct 2017) 4", + "26. Gaidon, A., Wang, Q., Cabon, Y., Vig, E.: Virtual worlds as proxy for multi-object tracking analysis. In: CVPR. pp. 4340-4349 (2016) 3", + "27. Ge, Z., Liu, S., Wang, F., Li, Z., Sun, J.: Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430 (2021) 9, 11", + "28. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) 9", + "29. Hu, H.N., Cai, Q.Z., Wang, D., Lin, J., Sun, M., Krahenbihl, P., Darrell, T., Yu, F.: Joint monocular 3d vehicle detection and tracking. In: ICCV (2019) 3", + "30. Jabri, A., Owens, A., Efros, A.: Space-time correspondence as a contrastive random walk. Advances in neural information processing systems 33, 19545-19560 (2020) 3", + "31. Jiang, H., Fels, S., Little, J.J.: A linear programming approach for multiple object tracking. In: CVPR. pp. 1-8 (2007). https://doi.org/10.1109/CVPR.2007.3831804", + "32. Kasturi, R., Goldgof, D., Soundararajan, P., Manohar, V., Garofolo, J., Boonstra, M., Korzhova, V., Zhang, J.: Framework for performance evaluation for face, text and vehicle detection and tracking in video: data, metrics, and protocol. IEEE TPAMI (2009) 9", + "33. Kaul, P., Xie, W., Zisserman, A.: Label, verify, correct: A simple few-shot object detection method. In: IEEE Conference on Computer Vision and Pattern Recognition (2022) 3", + "34. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: ICCV (2023) 2, 3", + "35. Lan, S., Yang, X., Yu, Z., Wu, Z., Alvarez, J.M., Anandkumar, A.: Vision transformers are good mask auto-labelers. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2023) 2, 3", + "36. Leal-Taixe, L., Canton-Ferrer, C., Schindler, K.: Learning by tracking: Siamese cnn for robust target association. In: CVPRW (June 2016) 4", + "37. Leal-Taixe, L., Fenzi, M., Kuznetsova, A., Rosenhahn, B., Savarese, S.: Learning an image-based motion context for multiple people tracking. In: CVPR (June 2014) 4" + ], + "bbox": [ + 215, + 146, + 784, + 837 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "O. Cetintas et al.", + "bbox": [ + 271, + 114, + 387, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "38. Leal-Taixe, L., Pons-Moll, G., Rosenhahn, B.: Everybody needs somebody: Modeling social and grouping behavior on a linear programming multiple people tracker. In: Int. Conf. Comput. Vis. Worksh. pp. 120-127 (2011). https://doi.org/10.1109/ICCVW.2011.6130233 4", + "39. Li, R., Zhang, B., Liu, J., Liu, W., Zhao, J., Teng, Z.: Heterogeneous diversity driven active learning for multi-object tracking. In: ICCV (2023) 2, 4", + "40. Litrico, M., Del Bue, A., Morerio, P.: Guiding pseudo-labels with uncertainty estimation for source-free unsupervised domain adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023) 2, 3", + "41. Liu, K., Jin, S., Fu, Z., Chen, Z., Jiang, R., Ye, J.: Uncertainty-aware unsupervised multi-object tracking. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9996-10005 (2023) 3", + "42. Luiten, J., Osep, A., Dendorfer, P., Torr, P., Geiger, A., Leal-Taixe, L., Leibe, B.: Hota: A higher order metric for evaluating multi-object tracking. IJCV 129(2), 548-578 (2021) 9", + "43. Manen, S., Gygli, M., Dai, D., Van Gool, L.: Pathtrack: Fast trajectory annotation with path supervision. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 290-299 (2017) 4", + "44. Meinhardt, T., Kirillov, A., Leal-Taixe, L., Feichtenhofer, C.: Trackformer: Multi-object tracking with transformers. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022) 1, 4, 11", + "45. Meng, S., Shao, D., Guo, J., Gao, S.: Tracking without label: Unsupervised multiple object tracking via contrastive similarity learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16264-16273 (2023) 3", + "46. Pang, J., Qiu, L., Li, X., Chen, H., Li, Q., Darrell, T., Yu, F.: Quasi-dense similarity learning for multiple object tracking. In: CVPR. pp. 164-173 (2021) 4, 11", + "47. Pellegrini, S., Ess, A., Schindler, K., van Gool, L.: You'll never walk alone: Modeling social behavior for multi-target tracking. In: ICCV. pp. 261-268 (2009). https://doi.org/10.1109/ICCV.2009.5459260 4", + "48. Pirsiavash, H., Ramanan, D., Fowlkes, C.C.: Globally-optimal greedy algorithms for tracking a variable number of objects. In: CVPR. pp. 1201-1208 (2011). https://doi.org/10.1109/CVPR.2011.5995604 4", + "49. Qin, Z., Zhou, S., Wang, L., Duan, J., Hua, G., Tang, W.: Motiontrack: Learning robust short-term and long-term motions for multi-object tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17939-17948 (2023) 4, 11", + "50. Ristani, E., Solera, F., Zou, R., Cucchiara, R., Tomasi, C.: Performance measures and a data set for multi-target, multi-camera tracking. In: Eur. Conf. Comput. Vis. Worksh. pp. 17-35. Springer (2016) 9", + "51. Ristani, E., Tomasi, C.: Features for multi-target multi-camera tracking and re-identification. In: CVPR (June 2018) 4", + "52. Robicquet, A., Sadeghian, A., Alahi, A., Savarese, S.: Learning social etiquette: Human trajectory understanding in crowded scenes. In: ECCV (2016) 4", + "53. Scovanner, P., Tappen, M.F.: Learning pedestrian dynamics from the real world. In: 2009 IEEE 12th International Conference on Computer Vision. pp. 381-388. IEEE (2009) 4", + "54. Seidenschwarz, J., Brasó, G., Serrano, V.C., Elezi, I., Leal-Taixe, L.: Simple cues lead to a strong multi-object tracker. In: CVPR. pp. 13813-13823 (2023) 4, 9, 11, 13, 14" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "SPAMming Labels: Efficient Annotations for the Trackers of Tomorrow", + "bbox": [ + 256, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "55. Sener, O., Savarese, S.: Active learning for convolutional neural networks: A core-set approach. In: ICLR (2018), https://openreview.net/forum?id=H1aIuk-RW 12", + "56. Shao, S., Zhao, Z., Li, B., Xiao, T., Yu, G., Zhang, X., Sun, J.: Crowdhuman: A benchmark for detecting human in a crowd. arXiv preprint arXiv:1805.00123 (2018) 3", + "57. Sun, P., Cao, J., Jiang, Y., Yuan, Z., Bai, S., Kitani, K., Luo, P.: Dancetrack: Multiobject tracking in uniform appearance and diverse motion. In: CVPR (2022) 3, 9, 13", + "58. Sun, S., Akhtar, N., Song, X., Song, H., Mian, A., Shah, M.: Simultaneous detection and tracking with motion modelling for multiple object tracking. In: ECCV. pp. 626-643 (2020) 3", + "59. Voigtlaender, P., Krause, M., Osep, A., Luiten, J., Sekar, B.B.G., Geiger, A., Leibe, B.: Mots: Multi-object tracking and segmentation. In: Proceedings of the iee/cvf conference on computer vision and pattern recognition. pp. 7942-7951 (2019) 2, 4", + "60. Voigtlaender, P., Luo, L., Yuan, C., Jiang, Y., Leibe, B.: Reducing the annotation effort for video object segmentation datasets. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3060-3069 (2021) 2, 4", + "61. Vondrick, C., Ramanan, D.: Video annotation and tracking with active learning. NeurIPS 24 (2011) 2", + "62. Vondrick, C., Ramanan, D., Patterson, D.: Efficiently scaling up video annotation with crowdsourced marketplaces. In: Computer Vision-ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part IV 11. pp. 610-623. Springer (2010) 4", + "63. Vondrick, C., Shrivastava, A., Fathi, A., Guadarrama, S., Murphy, K.: Tracking emerges by colorizing videos. In: Proceedings of the European conference on computer vision (ECCV). pp. 391-408 (2018) 3", + "64. Wang, X., Jabri, A., Efros, A.A.: Learning correspondence from the cycle-consistency of time. In: CVPR (2019) 3", + "65. Wang, Y., Kitani, K., Weng, X.: Joint object detection and multi-object tracking with graph neural networks. In: IEEE Int. Conf. Robotics and Autom. pp. 13708-13715 (2021). https://doi.org/10.1109/ICRA48506.2021.9561110 4", + "66. Wang, Z., Zheng, L., Liu, Y., Wang, S.: Towards real-time multi-object tracking. The European Conference on Computer Vision (ECCV) (2020) 1", + "67. Wei, F., Gao, Y., Wu, Z., Hu, H., Lin, S.: Aligning pretraining for detection via object-level contrastive learning. In: Advances in Neural Information Processing Systems (2021) 2", + "68. Xiao, T., Li, S., Wang, B., Lin, L., Wang, X.: Joint detection and identification feature learning for person search. In: CVPR. pp. 3415-3424 (2017) 3", + "69. Yamaguchi, K., Berg, A.C., Ortiz, L.E., Berg, T.L.: Who are you with and where are you going? In: CVPR. pp. 1345-1352 (2011). https://doi.org/10.1109/CVPR.2011.5995468 4", + "70. You, S., Yao, H., Bao, B.K., Xu, C.: Utm: A unified multiple object tracking model with identity-aware feature enhancement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21876-21886 (2023) 4, 11", + "71. Yuan, D., Chang, X., Liu, Q., Yang, Y., Wang, D., Shu, M., He, Z., Shi, G.: Active learning for deep visual tracking. IEEE Transactions on Neural Networks and Learning Systems (2023) 2" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "O. Cetintas et al.", + "bbox": [ + 271, + 114, + 387, + 126 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "72. Zeng, F., Dong, B., Zhang, Y., Wang, T., Zhang, X., Wei, Y.: Motr: End-to-end multiple-object tracking with transformer. In: European Conference on Computer Vision. pp. 659-675. Springer (2022) 1, 4, 11", + "73. Zhang, L., Li, Y., Nevatia, R.: Global data association for multi-object tracking using network flows. In: CVPR (2008) 4, 5", + "74. Zhang, S., Benenson, R., Schiele, B.: Citypersons: A diverse dataset for pedestrian detection. In: CVPR. pp. 3213-3221 (2017) 3", + "75. Zhang, Y., Sun, P., Jiang, Y., Yu, D., Weng, F., Yuan, Z., Luo, P., Liu, W., Wang, X.: Bytetrack: Multi-object tracking by associating every detection box. In: ECCV. pp. 1-21. Springer (2022) 4, 9, 11, 13, 14", + "76. Zhang, Y., Wang, C., Wang, X., Zeng, W., Liu, W.: Fairmot: On the fairness of detection and re-identification in multiple object tracking. IJCV 129(11), 3069-3087 (2021) 1, 4, 9, 11", + "77. Zheng, L., Zhang, H., Sun, S., Chandraker, M., Yang, Y., Tian, Q.: Person re-identification in the wild. In: CVPR. pp. 1367-1376 (2017) 3", + "78. Zhou, X., Koltun, V., Krahenbuhl, P.: Tracking objects as points. In: ECCV. pp. 474-490. Springer (2020) 1, 4" + ], + "bbox": [ + 215, + 146, + 787, + 383 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "SPAMming Labels: Efficient Annotations for the Trackers of Tomorrow", + "bbox": [ + 256, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/SPAMming Labels_ Efficient Annotations for the Trackers of Tomorrow/0cf5ca3c-0b18-48c1-b497-d85ab9c4bdac_model.json b/2024/SPAMming Labels_ Efficient Annotations for the Trackers of Tomorrow/0cf5ca3c-0b18-48c1-b497-d85ab9c4bdac_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6a02da0f03ab94d3a11b074ee9ae852d739139fa --- /dev/null +++ b/2024/SPAMming Labels_ Efficient Annotations for the Trackers of Tomorrow/0cf5ca3c-0b18-48c1-b497-d85ab9c4bdac_model.json @@ -0,0 +1,2449 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.267, + 0.141, + 0.737, + 0.185 + ], + "angle": 0, + "content": "SPAMming Labels: Efficient Annotations for the Trackers of Tomorrow" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.211, + 0.786, + 0.228 + ], + "angle": 0, + "content": "Orcun Cetintas\\(^{1,2\\dagger}\\), Tim Meinhardt\\(^{2}\\), Guillem Brasó\\(^{1,2\\dagger}\\), and Laura Leal-Taixe\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.389, + 0.239, + 0.613, + 0.266 + ], + "angle": 0, + "content": "\\(^{1}\\) Technical University of Munich \n\\(^{2}\\) NVIDIA" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.302, + 0.744, + 0.593 + ], + "angle": 0, + "content": "Abstract. Increasing the annotation efficiency of trajectory annotations from videos has the potential to enable the next generation of data-hungry tracking algorithms to thrive on large-scale datasets. Despite the importance of this task, there are currently very few works exploring how to efficiently label tracking datasets comprehensively. In this work, we introduce SPAM, a video label engine that provides high-quality labels with minimal human intervention. SPAM is built around two key insights: i) most tracking scenarios can be easily resolved. To take advantage of this, we utilize a pre-trained model to generate high-quality pseudo-labels, reserving human involvement for a smaller subset of more difficult instances; ii) handling the spatiotemporal dependencies of track annotations across time can be elegantly and efficiently formulated through graphs. Therefore, we use a unified graph formulation to address the annotation of both detections and identity association for tracks across time. Based on these insights, SPAM produces high-quality annotations with a fraction of ground truth labeling cost. We demonstrate that trackers trained on SPAM labels achieve comparable performance to those trained on human annotations while requiring only \\(3 - 20\\%\\) of the human labeling effort. Hence, SPAM paves the way towards highly efficient labeling of large-scale tracking datasets. We release all models and code." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.606, + 0.714, + 0.621 + ], + "angle": 0, + "content": "Keywords: Multi-object tracking \\(\\cdot\\) Data labeling \\(\\cdot\\) Active learning" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.666, + 0.377, + 0.683 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.697, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Detecting and tracking objects in time is at the core of many challenging vision applications. To address these problems, modern multi-object tracking (MOT) approaches [44, 66, 72, 76, 78] rely on increasingly larger amounts of annotated data. While annotating image datasets is already costly, temporal component of videos further increases the annotation difficulty and data scale requirements. This is due to the fact that redundancies between frames cause the information density to scale poorly with the amount of data, making the overall annotation task more challenging and resource-intensive." + }, + { + "type": "page_footnote", + "bbox": [ + 0.238, + 0.826, + 0.549, + 0.84 + ], + "angle": 0, + "content": "† Work done during an internship at NVIDIA." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.389, + 0.128 + ], + "angle": 0, + "content": "O. Cetintas et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.344 + ], + "angle": 0, + "content": "To mitigate these issues, researchers are spending more effort in operating with limited human annotations. Self- and semi-supervised learning (SSL) and pseudo-labeling (PL) approaches are often used to leverage unlabeled data whenever possible [34,35,40], while active learning (AL) strategies aim at selecting the most informative samples to label. All these strategies are commonly applied in various image understanding tasks, such as classification [11,14] or object detection [17,67]. Even if video tasks do not have any lower requirement for data, there are only a handful of works that target efficient labeling of tracking data in the video domain [39,59-61]. Existing works fail to provide comprehensive solutions to labeling by either ignoring dense temporal component of videos [39,61] or working under a limited single-object setup [71]. An effective video labeling engine should operate across multiple objects and frames, considering the intricate dependencies between tracks inherent in the labeling problem." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.348, + 0.788, + 0.53 + ], + "angle": 0, + "content": "Videos have unique characteristics that can be leveraged to unleash the full potential of a modern label engine. Firstly, frame-wise similarities in videos yield data redundancies. In the context of MOT, this implies that some associations between detections are relatively easy to solve [7]. This setup provides an ideal opportunity to leverage state-of-the-art models pre-trained to pseudo-label new data at zero cost. Secondly, the dependencies of objects across time and space cause annotations in one frame of the video to have cascading effects in other frames. For example, solving an association for one track can resolve associations for neighboring tracks. In order to maximize the efficiency of the annotations, we frame the labeling task within a model that takes into account such dependencies. Thus, we adopt a track-centric labeling approach rather than a frame-centric one [39]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.534, + 0.788, + 0.761 + ], + "angle": 0, + "content": "We present the SPAM video label engine, which chains Synthetic pretraining, Pseudo-labeling, and Active learning with a graph-based Model for optimal tracking performance with minimal human effort. In the image domain, there are plenty of large-scale annotated datasets that a model-driven approach can leverage for pre-training [34]. However, this is not the case for multi-object tracking datasets, which are typically much smaller in scale; therefore, we advocate for a Synthetic pre-training. We further analyze which components of a tracking method are more affected by the synthetic-to-real gap. This first experiment gives us an important insight: labeling should be focused on detections and associations, and we can directly use the synthetically trained reID model. With our synthetic pre-trained model, we then propose a model-driven approach to produce Pseudo-labels on real data and retrain the model with its own pseudolabels. We show that with this recipe we can even reach a performance comparable to ground truth without ever training our labeler on manually annotated data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Synthetic pre-training and pseudo-labeling allow us to handle easy decisions and defer only uncertain or complex decisions to an annotator. We propose an Active learning strategy where the labeler makes decisions at the track-level instead of frame-level, thereby utilizing the annotation budget more efficiently. Towards this goal, we solve the labeling problem with a Model, based on graph" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.258, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SPAMming Labels: Efficient Annotations for the Trackers of Tomorrow" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.783, + 0.22 + ], + "angle": 0, + "content": "hierarchies [12]. This model captures long-range spatio-temporal dependencies, which means our annotations will have an impact across frames, and not only on the frame where the annotation actually happens. We extend the model with a detection filtering layer so that we can naturally label detection and association within the same framework." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.785, + 0.358 + ], + "angle": 0, + "content": "In summary, we propose SPAM, a video label engine that significantly reduces annotation cost by dealing with easy scenarios with a strong MOT model pre-trained on (i) Synthetic data, followed by (ii) Pseudo-labeling real data for its own re-training. Harder detection and association decisions are carefully selected by an (iii) Active learning scheme that works on a (iv) Model based on graph hierarchies. This allows us to efficiently use the annotator's budget and account for spatio-temporal dependencies in the labels. The resulting model generates labels comparable to ground truth performance, requiring only \\(3 - 20\\%\\) of manual annotation effort across three diverse datasets." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.383, + 0.388, + 0.399 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.416, + 0.783, + 0.552 + ], + "angle": 0, + "content": "Pre-training for tracking. Over the years, the multi-object tracking (MOT) community has annotated many challenging benchmark datasets [18, 19, 57]. However, due to the significant costs associated with manually labeling temporal data, most benchmarks only provide few annotated video sequences. To obtain top performance, modern data-hungry tracking methods are usually pretrained on additional video and tracking datasets [20, 22, 68, 74, 77]. Many works also rely on relevant image recognition [56] datasets to boost their per-frame object detection performance. An alternative pre-training direction avoids human annotations entirely by relying on simulated synthetic data [23, 24, 26, 29, 58]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.785, + 0.598 + ], + "angle": 0, + "content": "Our label engine also pre-trains the labeling model synthetically, minimizing its total annotation footprint. To this end, we demonstrate the effectiveness of synthetic pre-training [23] for generating pseudo-labels for self-supervision." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.785, + 0.749 + ], + "angle": 0, + "content": "Self-supervision in the temporal domain. To benefit not only from already annotated datasets, many works explore the utilization of available unlabeled data sources, i.e., how to perform self-supervised learning for videos. Notable examples leverage cycle-consistency-based correspondences [30, 64], color propagation [63] or augmentation-based consistencies [4, 41]. Furthermore, [45] uses contrastive learning over re-identification (reID) embeddings to perform unsupervised association learning. More recently, several works [33-35, 40] in the image domain demonstrate the effectiveness of training on pseudo-labels generated by strong underlying labeling models. In the video domain, this type of self-supervision is mostly unexplored." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.783, + 0.809 + ], + "angle": 0, + "content": "We demonstrate the effectiveness of pseudo-labels for the temporal domain by improving the label quality of our SPAM labeler via self-supervision on its own output. The necessary signal-to-noise ratio for this training loop is obtained by pre-training a strong graph-based model on synthetic data." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.783, + 0.84 + ], + "angle": 0, + "content": "Efficient labeling of video datasets. While significant effort has been devoted towards increasing the labeling efficiency of image datasets [1,34], similar" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.389, + 0.128 + ], + "angle": 0, + "content": "O. Cetintas et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.298 + ], + "angle": 0, + "content": "approaches for the video and tracking domain are still relatively limited. Earlier works reduce the number of frames to annotate within individual videos but still rely heavily on human annotators [62]. Other solutions are limited to only a single object per sequence [16, 60] or still rely on human bounding box annotations [59] to label segmentation masks across time. The authors of [39] select key frames based on their expected impact on tracking performance. While effective within its scope, this approach only yields sparse frame annotations, which are insufficient for training all modern trackers. To reduce the human annotation effort for trajectories, [43] proposes an efficient video interface that still requires labeling all trajectories." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.299, + 0.788, + 0.373 + ], + "angle": 0, + "content": "In contrast, our label engine performs active learning for multiple objects across all video frames and is the first of its kind to tackle long-term tracking problems via a hierarchical labeling approach. The uncertainty measures provided by our graph-based model allow us to focus expensive annotation efforts on individual tracks instead of labeling entire frames in a brute-force manner." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.375, + 0.788, + 0.556 + ], + "angle": 0, + "content": "Multi-object tracking models. Tracking objects across time in a video is an established problem with a long research history. The most dominant tracking-by-detection (TbD) paradigm splits the task into two steps: (i) detecting objects in every frame and (ii) associating them to object trajectories, also referred to as data association. Modern trackers have explored performing (i) and (ii) jointly [7, 25, 44, 78], or providing solutions for (ii) relying on motion models [2, 3, 38, 47, 49, 52, 53, 69, 75], appearance models [13, 15, 36, 46, 51, 54, 70, 76] or more complex techniques involving transformers [10, 72]. An alternative direction of work relies on graph formulations [6, 31, 37, 48, 73] to utilize the TbD paradigm. In this formulation, nodes model object detections, and edges, association hypotheses. This long-studied formulation has recently been revived with remarkable success with the help of learnable graph neural network (GNN)architectures [8, 9, 12, 65]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.556, + 0.788, + 0.616 + ], + "angle": 0, + "content": "For our label engine to generate long-term pseudo-labels and uncertainty estimations, we rely on a strong graph-based tracker [12]. To incorporate both detection and association labeling within a unified graph framework, we introduce a novel hierarchy level dedicated to detection filtering." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.642, + 0.318, + 0.658 + ], + "angle": 0, + "content": "3 SPAM" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.676, + 0.341, + 0.691 + ], + "angle": 0, + "content": "3.1 Overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.794 + ], + "angle": 0, + "content": "Our main goal is to label multi-object tracking datasets with minimal annotation effort while enabling optimal tracking performance. Our key observation is that most association cases can be solved easily by models pre-trained on synthetic data. Therefore, no human effort is needed, and we can instead use our model's predictions as pseudo-labels. Decisions that are more complex or uncertain are selected by an active learning scheme to be annotated manually." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.84 + ], + "angle": 0, + "content": "In order to perform pseudo-labeling and active learning, we require a model that can capture long-term spatio-temporal dependencies between tracks. We leverage a hierarchical graph neural network formulation and cast the annotation" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.257, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "SPAMming Labels: Efficient Annotations for the Trackers of Tomorrow" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.149, + 0.788, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.262, + 0.788, + 0.306 + ], + "angle": 0, + "content": "Fig. 1: Overview of the SPAM model. We first generate a set of detection candidates with our detector. Hierarchical GNNs then classify these candidates into valid and invalid objects via node classification, and assign identities through edge classification." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.331, + 0.784, + 0.436 + ], + "angle": 0, + "content": "task over a graph. Our architecture (Fig. 1) draws inspiration from SUSHI [12], which demonstrated remarkable association performance. Our model's scope, however, goes beyond association. Given the constraints of the labeling task, we need to equip our model with the ability to reason about detections before performing association. To address this broader goal, our GNN hierarchy, also includes the capability to accept or reject detections by classifying nodes into correct and incorrect object hypotheses." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.437, + 0.788, + 0.588 + ], + "angle": 0, + "content": "Our overall model is pre-trained on synthetic data [23] and further fine-tuned on its own predicted pseudo-labels over the target dataset that we aim to annotate. This recipe allows for strong labelling performance without any human annotations, and provides the basis for a scalable video label engine. After this training loop, we use our updated model's predictions as pseudo-labels to annotate a large portion of the target dataset. Decisions that are more complex or uncertain are selected by an active learning scheme to be annotated manually. The resulting set of labels achieves near-GT performance with significantly fewer manual annotations, thereby vastly enhancing the efficiency of the overall process. We provide an overview of our training and annotating recipe in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.607, + 0.618, + 0.623 + ], + "angle": 0, + "content": "3.2 Graph Hierarchies as a Model for Labeling" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.81 + ], + "angle": 0, + "content": "Problem formulation. Our model builds upon the standard graph formulation [73] where, given a set of object candidates \\(\\mathcal{O}\\), our goal is to obtain a set of valid objects \\(\\mathcal{O}_v \\subset \\mathcal{O}\\) and their corresponding trajectories \\(\\mathcal{T}\\). Each trajectory \\(T_k \\in \\mathcal{T}\\) consists of a set of objects sharing the same identity or, equivalently, the set of edges connecting those nodes. Specifically, we model each object candidate \\(o_i \\in \\mathcal{O}\\) as a node and each association hypothesis among candidates as an edge in an undirected graph \\(G = (V, E)\\). Within this formulation, nodes \\(u \\in V\\) could be classified into valid objects if \\(u \\in \\mathcal{O}_v\\). Analogously, edges can be classified into correct or incorrect hypotheses depending on whether they belong to some trajectory or not. This formulation can be generalized to a hierarchical setting by splitting video sequences into non-overlapping subsequences, i.e., disjoint subgraphs, and progressively merging them [12]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Background on graph hierarchies for MOT. Our model consists of a hierarchy of GNNs operating over the graph formulation described above. Intuitively," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.39, + 0.128 + ], + "angle": 0, + "content": "O. Cetintas et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.782, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.244, + 0.788, + 0.314 + ], + "angle": 0, + "content": "Fig. 2: Overview of the SPAM training and annotation pipeline. (a) Initial model training on synthetic data. (b) Application of SPAM to generate pseudo-labels without incurring manual annotation costs on a real dataset, followed by self-training on pseudolabels. (c) Real dataset labeling using pseudo-labels and an uncertainty-based active learning approach." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.358, + 0.788, + 0.51 + ], + "angle": 0, + "content": "starting from detection candidates, at each hierarchy level, GNN hierarchy subsequently merge tracklets from the previous level into longer ones. Nodes and edges are represented by initial embeddings and message passing is used to share information within the graph in order to update node and edge embeddings with richer information. The GNNs are trained to classify edges into active and inactive association hypotheses and they share the same features, network architecture and learnable weights across hierarchy levels. In addition to the edge features reported in [12], for each object candidate \\( o_i \\) we concatenate its times-tamp \\( t_i \\), normalized bounding box coordinates and dimensions \\( (x_i, y_i, h_i, w_i) \\) and corresponding confidence from the detector \\( c_i \\) and embed it as node features." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.516, + 0.788, + 0.743 + ], + "angle": 0, + "content": "Classifying nodes. We aim to minimize the annotation effort; therefore, we ask the annotator to operate on a fixed graph of detection candidates, i.e. annotator does not label new detections after the initial set given by detector. This is why we aim to create an initial graph that is over-complete. In order to do so, we use a low confidence threshold for our detector and obtain a large number of detection candidates. While this ensures a high detection recall, it also introduces a larger number of false positive detections that our model needs to address. As we show in experiments, the original SUSHI architecture [12] cannot handle such false positives, which is why we propose a new layer in the hierarchy, \\( GNN_{node} \\). Intuitively, \\( GNN_{node} \\) leverages temporal information to validate object candidates. We use its output node embeddings and feed them to a lightweight classifier to enable false positive detection filtering. By doing so, we can consider a large number of input proposals achieving high detection recall by leveraging the spatio-temporal reasoning ability of our GNN for detection candidates." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.749, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Pre-training and annotation-free fine-tuning. We pre-train all of our model components on the large-scale synthetic dataset MOTSynth [23]. To boost our model's performance and bridge the domain gap with our target dataset for annotation, we conduct further fine-tuning on the target dataset after pre-training. We achieve this by utilizing the raw predictions of our model on the target dataset and employing them as pseudo-labels for self-training. As we show in" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.258, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "SPAMming Labels: Efficient Annotations for the Trackers of Tomorrow" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.269, + 0.145, + 0.744, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.299, + 0.785, + 0.341 + ], + "angle": 0, + "content": "Fig. 3: Our graph-based labeling pipeline begins with the selection of nodes for annotation. For each node to be annotated, the annotator could be asked to validate the detection, improve the localization by refining the box or perform association." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.373, + 0.784, + 0.403 + ], + "angle": 0, + "content": "our experimental section, this strategy yields a significant performance improvement without incurring any human annotation cost." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.431, + 0.566, + 0.446 + ], + "angle": 0, + "content": "3.3 How to Label on Graph Hierarchies?" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.461, + 0.788, + 0.627 + ], + "angle": 0, + "content": "Manual labeling. Labeling video data at the box level involves several steps: identifying objects of interest, localizing their boundaries with bounding boxes, and assigning them unique identities. Typically, annotators achieve this by examining each frame, detecting and localizing objects individually, and then assigning identities based on temporal context from neighboring frames. In terms of annotation effort in clicks, detecting and localizing objects typically require two clicks (one for the top-left corner and one for the bottom-right corner of the bounding box), while associating objects across time requires one click (identifying the next occurrence of the object in subsequent frames), as reported in [18, 19]. This manual process is labor-intensive, costly, and severely limits the availability of annotated video data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.765 + ], + "angle": 0, + "content": "Graph-based labeling. Utilizing a graph formulation for detection and tracking enables explicit modeling of objects in the scene and their interactions. In this framework, each node of the graph represents an object, allowing a manual annotation of an object to be framed as selecting a node on this graph. Given a set of nodes to be annotated, possible annotator decisions are: (i) accepting/rejecting a detection as we are working on a large set of detection candidates (one click), (ii) refining the bounding-box to improve localization (two clicks, one for the top-left corner and one for the bottom-right corner of the bounding box) and (iii) associating a node across time (one click) as shown in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Annotation acquisition. The core idea of our approach is to use our model decisions to guide the annotation process. For the vast majority of scenarios, we directly use our model's predictions as pseudo-labels. For a small subset of more challenging decisions where our model exhibits uncertainty, we validate our decisions with an annotator as described above. Within our graphs," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.39, + 0.128 + ], + "angle": 0, + "content": "O. Cetintas et al." + }, + { + "type": "image", + "bbox": [ + 0.308, + 0.152, + 0.699, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.352, + 0.788, + 0.394 + ], + "angle": 0, + "content": "Fig. 4: Analysis of performance gap between training a model on synthetic and real data for the three most common tracking components: detection, association, re-identification." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.415, + 0.789, + 0.478 + ], + "angle": 0, + "content": "we adopt an uncertainty-based approach to identify which nodes require annotations. Specifically, for each node \\( v \\in V \\), we compute its uncertainty as \\( \\mathrm{uncert}(v) \\coloneqq \\max_{u \\in \\mathbb{N}_v} H(\\hat{y}_{(v,u)}) \\), where \\( \\mathbb{N}_v \\) denotes the set of neighboring nodes of \\( v \\) and \\( \\hat{y}_{(v,u)} \\) is the edge prediction for edge \\( (v,u) \\):" + }, + { + "type": "equation", + "bbox": [ + 0.29, + 0.494, + 0.787, + 0.513 + ], + "angle": 0, + "content": "\\[\nH (\\hat {y} _ {(v, u)}) := - (\\hat {y} _ {(v, u)} \\log \\hat {y} _ {(v, u)} + (1 - \\hat {y} _ {(v, u)}) \\log (1 - \\hat {y} _ {(v, u)})) \\qquad (1)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.52, + 0.788, + 0.594 + ], + "angle": 0, + "content": "At the \\( GNN_{node} \\) level, where we have node predictions instead of edge predictions, we simply compute \\( \\text{uncert}(v) \\) as \\( H(\\hat{y}_{(v)}) \\) instead. Subsequently, we refer the nodes with the highest uncertainty scores to the annotator for manual annotation and use the remaining predictions as pseudo-labels for the rest of the nodes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.596, + 0.789, + 0.732 + ], + "angle": 0, + "content": "Hierarchical label generation. To fully leverage our hierarchical graph formulation, we apply our label acquisition approach described above in a hierarchical manner. Intuitively, we distribute our annotation budget \\( B \\) across \\( L \\) hierarchical levels as \\( B_{1},\\ldots ,B_{L} \\), with \\( B_{1} + \\dots +B_{L} = B \\). In deeper hierarchy levels, nodes represent tracklets instead of single detections. This allows us to propagate annotator decisions for the entire cluster instead of individual detections, resulting in solving multiple uncertainties with a single annotation. As a result, we utilize the annotation budget more effectively while also mitigating error propagation within our pseudo-labels." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.76, + 0.378, + 0.777 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.796, + 0.441, + 0.81 + ], + "angle": 0, + "content": "4.1 Datasets and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.825, + 0.686, + 0.841 + ], + "angle": 0, + "content": "Datasets. We conduct experiments on three public benchmarks:" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.257, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SPAMming Labels: Efficient Annotations for the Trackers of Tomorrow" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.146, + 0.786, + 0.191 + ], + "angle": 0, + "content": "- MOT17 [18] is an established benchmark containing 14 challenging video sequences, exhibiting diverse characteristics, including variations in camera motion, viewpoint, and pedestrian densities." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.192, + 0.787, + 0.252 + ], + "angle": 0, + "content": "- MOT20 [19] contains a total of 8 video sequences, all featuring extremely crowded scenes. Due to their high pedestrian density, the emphasis is on addressing highly challenging scenarios characterized by frequent and prolonged occlusions." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.253, + 0.787, + 0.329 + ], + "angle": 0, + "content": "- DanceTrack [57] contains 100 video sequences. These sequences feature dancing individuals with similar appearances, facilitating the evaluation of tracker robustness against highly challenging motion patterns where appearance information has a relatively minor influence. This non-standard setup often leads to a performance drop in trackers [57]." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.146, + 0.787, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.34, + 0.784, + 0.355 + ], + "angle": 0, + "content": "Metrics. We assess the performance of trackers using three established metrics:" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.368, + 0.786, + 0.412 + ], + "angle": 0, + "content": "- MOTA [32] assesses trajectory coverage, by combining object detection measurements with identity switches, with the latter playing a relatively minor role." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.413, + 0.786, + 0.442 + ], + "angle": 0, + "content": "- IDF1 [50] focuses on identity preservation over the sequences and provides a measure of association performance." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.443, + 0.786, + 0.473 + ], + "angle": 0, + "content": "- HOTA [42] is a recently introduced metric designed to strike a balance between association and object detection performance within a single score." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.368, + 0.786, + 0.473 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.496, + 0.456, + 0.51 + ], + "angle": 0, + "content": "4.2 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.521, + 0.787, + 0.626 + ], + "angle": 0, + "content": "Architecture and training. Our model consists of three main components: (i) a detector network, based onYOLOX [27], (ii) a ResNet-based ReID network [28], and (iii) a GNN hierarchy, all pretrained on the synthetic sequences of MOTSynth [23]. After pretraining, we use the resulting model to obtain pseudolabels for our target MOT dataset, and then fine-tune our detector and GNN on them. Note that our Model is never trained on human annotations, it is only trained on synthetic annotations and its own pseudo-labels." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.627, + 0.787, + 0.747 + ], + "angle": 0, + "content": "Experimental setup. Online trackers are real-time tools suitable for deployment on robots or autonomous vehicles, whereas offline trackers are better suited for labeling tasks. Therefore, we structure our experiments to utilize SPAM for labeling and use these labels with online trackers. To evaluate the quality of SPAM labels, we train two state-of-the-art online trackers with these labels: ByteTrack [75] and GHOST [54]. For validation, we follow the standard half-split setting [54, 75, 76] over MOT17 sequences. Annotation budgets are computed based on number of clicks as described in Sec. 3.3." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.77, + 0.434, + 0.785 + ], + "angle": 0, + "content": "4.3 Synthetic Pretaining" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Several works have shown that synthetic pre-training is a strong starting point for multi-object tracking methods [23, 75]. In this section, we aim to analyze the gap between training a model on synthetic data and training it with real" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.389, + 0.128 + ], + "angle": 0, + "content": "O. Cetintas et al." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.145, + 0.788, + 0.215 + ], + "angle": 0, + "content": "
ModelHigh Conf. BoxLow Conf. Box\\( GNN_{node} \\)MOTA ↑IDF1 ↑HOTA ↑DetA ↑
A✓✘✘64.474.759.955.9
B✓✓✘60.671.458.554.9
C✓✓✓65.475.160.456.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.218, + 0.785, + 0.246 + ], + "angle": 0, + "content": "Table 1: Ablation study investigating the impact of incorporating low confidence bounding boxes with and without the \\( GNN_{node} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.264, + 0.788, + 0.46 + ], + "angle": 0, + "content": "data for the three most common tracking components: detection, association, re-identification. For each setting, only the specific component that we are analysing is trained on synthetic data while other components are all trained on real data. We perform this experiment on three datasets and plot the relative HOTA difference (a positive score indicates superior performance with real data training) when training on real data vs. synthetic data in Fig. 4. Our first observation, is that the detector is the component that suffers the domain gap the most, with a performance drop of 9.9 HOTA points. Re-identification, on the other hand, has a similar performance when trained on synthetic or real data. Lastly, association performance is 2.1 HOTA higher with real data training. This first experiment gives us an important insight: labeling and retraining should be focused on detection and association components, and we can directly use the synthetically trained reID model." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.483, + 0.513, + 0.499 + ], + "angle": 0, + "content": "4.4 Towards a Model for Labeling" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.789, + 0.794 + ], + "angle": 0, + "content": "Graph-based validation of detection candidates. In our framework, annotators have the capability to eliminate detections, but we refrain from requesting them to annotate new detections in order to maintain minimal annotation effort. Instead we start from an initial graph formed by an over-complete set of detections. Towards this end, we use a low confidence threshold for our detector, thereby reducing the number of false negatives in our initial graph and ensuring a high recall. As a result, this leads to an increase in the number of false positives incorporated into the graph, a tendency that is further accentuated by the fact that our detector is exclusively trained on synthetic data. We propose to leverage the graph structure itself to validate the detections based on temporal consistency. Specifically, we add an additional GNN-layer in our hierarchy, \\( GNN_{node} \\), which intuitively examines the spatial and temporal consistency of the bounding box candidates and filters out implausible ones. In Tab. 1, we assess the effect of this design choice. Model A uses only high confidence bounding boxes, model B creates a graph with also low confidence bounding boxes but without \\( GNN_{node} \\), while C includes \\( GNN_{node} \\). As we can see, simply adding low confidence boxes results in a drastic drop in performance on all metrics. Only by adding our \\( GNN_{node} \\) we are able to take full advantage of the low confidence boxes without penalizing our performance with too many errors." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.84 + ], + "angle": 0, + "content": "SPAM is a SOTA tracker. Before moving to labeling with our model, we want to verify its overall performance and generalization capabilities. Only by using a state-of-the-art model we will be able to reduce annotation effort as much as" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.258, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SPAMming Labels: Efficient Annotations for the Trackers of Tomorrow" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.145, + 0.788, + 0.315 + ], + "angle": 0, + "content": "
MethodHOTA ↑MOTA ↑IDF1 ↑HOTA ↑MOTA ↑IDF1 ↑HOTA ↑MOTA ↑IDF1 ↑
MOT17MOT20DanceTrack
MeMOT [10]56.972.569.054.166.163.7---
TrackFormer [44]57.374.168.054.765.768.6---
MOTR [72]57.868.673.4---54.251.579.7
FairMOT [76]59.373.772.354.661.867.339.782.240.8
QDTrack [46]63.578.777.560.074.773.854.287.750.4
ByteTrack† [75]62.878.977.160.474.274.547.789.653.9
GHOST [54]62.878.777.161.273.775.256.791.357.7
UTM [70]64.081.878.762.578.276.9---
MotionTrack [49]65.181.180.162.878.076.5---
SUSHI [12]66.581.183.164.374.379.863.388.763.4
SPAM (Ours)67.580.784.665.876.581.964.089.263.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.316, + 0.784, + 0.343 + ], + "angle": 0, + "content": "Table 2: Test set results on MOT17, MOT20 and DanceTrack when using SPAM solely as a tracker." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.357, + 0.788, + 0.584 + ], + "angle": 0, + "content": "possible. We therefore pre-train our model on the synthetic dataset, and finetune it on the specific real dataset using groundtruth. For a fair comparison, we obtain object detections from a YOLOX detector [27] trained following [12, 54, 75]. Our model outperforms state-of-the-art by 1.0 HOTA point on MOT17, 1.5 HOTA points on MOT20, and 0.7 HOTA points on DanceTrack as shown in Tab. 2. Furthermore, on MOT17 and MOT20, there is also a significant improvement in IDF1, showing the identity-preserving capabilities of the model, which is especially important for our video annotation setting. Even if being state-of-the-art is not the main goal of this paper, having a strong performing model allows us to have an excellent starting point for producing pseudo-labels. We would like to emphasize that, we use ground truth annotations only for this experiment just to compare our model fairly with other trackers. From this point onward, we will not use ground truth labels. We will use our synthetic pretraining weights and focus on the pseudo-labeling and active learning aspects of our method to show how to annotate multi-object tracking datasets." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.605, + 0.787, + 0.621 + ], + "angle": 0, + "content": "4.5 Enhancing Annotations with Pseudo-labels and Active Learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.765 + ], + "angle": 0, + "content": "Retraining our Model on Pseudo-labels. In this experiment, we compare two models. The first one, SPAM, is only trained on synthetic data. The second one, SPAM, is trained in three stages: (i) pre-training on synthetic data, (ii) pseudo-labeling the real dataset with our model, and (iii) re-training our model on pseudo-labels. In Sec. 4.5, we observe a significant improvement when using pseudo-labels to re-train our model: an increase of 4-6 HOTA, 4-6 MOTA, and 4-7 IDF1 points on MOT17, MOT20, and DanceTrack, respectively. Note that these improvements come at zero human annotation effort as this process requires no manual annotations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Active learning on graph hierarchies. We now test the effectiveness of our proposed active learning labeling on graph hierarchies, and show a comprehensive comparison with several baselines in Fig. 5. We directly measure the accuracy of the output labels given an increasing percentage of manual annotations selected" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.389, + 0.127 + ], + "angle": 0, + "content": "O. Cetintas et al." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.144, + 0.468, + 0.282 + ], + "angle": 0, + "content": "
ModelPseudo-labelHOTA ↑MOTA ↑IDF1 ↑
MOT17
SPAM✘60.065.375.1
SPAM✓63.869.279.7
MOT20
SPAM✘52.265.470.3
SPAM✓58.770.277.1
DanceTrack
SPAM✘41.865.243.4
SPAM✓48.171.050.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.293, + 0.47, + 0.349 + ], + "angle": 0, + "content": "Table 3: Performance boost obtained by our model when retraining with its own pseudo-labels incurring no manual annotation cost." + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.144, + 0.803, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.487, + 0.304, + 0.788, + 0.333 + ], + "angle": 0, + "content": "Fig. 5: Impact of various active learning strategies on the quality of labeling." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.376, + 0.788, + 0.542 + ], + "angle": 0, + "content": "by different acquisition functions on MOT17. The vertical axis shows the increase in HOTA over the baseline performance, i.e., the performance without active learning. Oracle is our upper bound and shows the performance if all possible labels are set to their correct value by the annotator. Every method has two different labeling schemes: one for detections (nodes), and one for associations (edges) as explained in Sec. 3.3. In addition to our method, we explore active learning strategies commonly used in object detection, including Coreset [55], Ensemble [5], and Consistency over augmented inputs [21]. The naming is based on the active learning strategy for detection in the first part and the strategy for association in the second part. For instance, EntImg/Rand. indicates image level entropy-based selection for detection and random selection for association." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.543, + 0.788, + 0.77 + ], + "angle": 0, + "content": "The methods shown in red shades and circle markers compare three detection labeling methods while keeping association label selection random. Box-level labeling is far superior to commonly-used image-level labeling, or random labeling. This highlights the potential of performing active learning at the object level rather than the frame level. The methods in blue shades and triangle markers all use our proposed active learning selection on graph hierarchies for the association, and compare several detection selection schemes. The Coreset [55], which prioritizes diversity of the data over uncertainty, emerges as the poorest performing approach. This observation underlines that, when accurate predictions are available, an uncertainty-based approach proves to be a more effective complement to pseudo-labels. Box-level entropy, ensembles, and consistency all perform similarly. Finally, our method, with both detection and association label selection on graph hierarchies is superior to all compared baselines, staying at only 1.2 points from Oracle performance. Notably, we also unify the AL strategy for detection and association with our approach." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.788, + 0.71, + 0.804 + ], + "angle": 0, + "content": "4.6 Putting Everything Together: SPAM Labels in Action" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Training trackers with our labels. In this experiment, we ablate the effects of our SPAM pipeline on the actual usecase of training a model with our pro" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.258, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SPAMming Labels: Efficient Annotations for the Trackers of Tomorrow" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.145, + 0.788, + 0.224 + ], + "angle": 0, + "content": "
ModelLabelsPseudo-labelsActive Learn.Ann. Effort ↓HOTA ↑MOTA ↑IDF1 ↑
ByteTrack [75]Ground Truth-100%52.660.465.7
ByteTrack [75]SPAMXXX49.657.863.1
ByteTrack [75]SPAM✓XX50.559.564.5
ByteTrack [75]SPAM✓✓3.3%52.561.866.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.225, + 0.785, + 0.266 + ], + "angle": 0, + "content": "Table 4: Effect of (i) Synthetic training, (ii) Retraining our model with our Pseudolabels, and (iii) Active learning-based acquisition and annotation of a small percentage of real data, on our label quality on the MOT17 validation set." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.304, + 0.789, + 0.518 + ], + "angle": 0, + "content": "vided labels. We ablate label quality under the following settings: (i) Synthetic training, (ii) Retraining our models on our Pseudo-labels, (iii) Active learning based acquisition and annotation of a small percentage of real data. We train state-of-the-art online tracker ByteTrack [75] from scratch with the aforementioned sets of labels and report its performance on the MOT17 validation set in Tab. 4. The goal performance we want to reach is the one achieved by the model trained on groundtruth annotations (row 1). Row 2 shows our model trained only on synthetic data, which is 3 HOTA points below our goal. Using pseudo-label based training of our model increases performance by almost a HOTA point, and the ground truth level performance is achieved by also using our active learning method to select which samples to label. By annotating only \\(3.3\\%\\) of the data, we can achieve the same results as the model trained on the ground truth data with \\(100\\%\\) annotation effort. This shows the potential of our SPAM in reducing annotation effort for multi-object tracking datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.789, + 0.844 + ], + "angle": 0, + "content": "How much data do we really need to label? In this final experiment, our goal is to assess the impact of our entire pipeline by answering the question: How much data do we actually need to label in order to achieve the same performance as a model trained on the entire groundtruth training dataset? For that, we choose two well-established trackers, ByteTrack [75] and GHOST [54], and train them from scratch using our SPAM labels. We report their performance on MOT17, MOT20, and DanceTrack test sets in Tab. 5. We first confirm that our results on the validation set also generalize to the test set, and that we can reach GT performance with \\(3.3\\%\\) of annotation effort on MOT17. Note that for MOT17 our annotations yield a slight improvement over GT, which can be attributed to observed annotation noise in the original GT, particularly in moving sequences (further details can be found in the supplementary material). MOT20 and DanceTrack are known to be more challenging datasets [57], which is confirmed by the fact that the trackers still miss a few points when trained on only \\(3.3\\%\\) of manually annotated data. Given that we cannot do a full analysis of different percentages because we are working on the test sets, we choose two percentages of manually annotated data on which to train our trackers: \\(10\\%\\) for MOT20, and \\(20\\%\\) for DanceTrack. In both cases, we achieve on-par performance with the trackers trained on full groundtruth data. This clearly shows the potential in properly leveraging synthetic data, pseudo-labels from a strong model, and active learning on graph hierarchies in order to label multi-object tracking" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.389, + 0.127 + ], + "angle": 0, + "content": "O. Cetintas et al." + }, + { + "type": "table", + "bbox": [ + 0.278, + 0.144, + 0.731, + 0.491 + ], + "angle": 0, + "content": "
MethodLabelsAnn. Effort ↓HOTA ↑MOTA ↑IDF1 ↑
MOT17 Test
ByteTrack [75]GT100%50.660.461.1
ByteTrack [75]SPAM3.3%51.664.063.0
GHOST [54]GT100%49.558.059.0
GHOST [54]SPAM3.3%51.361.962.1
MOT20 Test
ByteTrack [75]GT100%48.760.560.0
ByteTrack [75]SPAM3.3%47.957.661.4
ByteTrack [75]SPAM10%48.461.062.8
GHOST [54]GT100%48.359.760.2
GHOST [54]SPAM3.3%47.058.260.7
GHOST [54]SPAM10%46.959.060.8
DanceTrack Test
ByteTrack [75]GT100%41.079.146.0
ByteTrack [75]SPAM3.3%39.576.445.0
ByteTrack [75]SPAM20%41.380.948.1
GHOST [54]GT100%45.880.646.3
GHOST [54]SPAM3.3%41.076.344.8
GHOST [54]SPAM20%43.981.347.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.494, + 0.788, + 0.536 + ], + "angle": 0, + "content": "Table 5: Evaluation of SPAM labels on MOT17, MOT20, and DanceTrack test sets with varying annotation budgets by training ByteTrack [75] and GHOST [54] using annotations generated by our method." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.56, + 0.788, + 0.591 + ], + "angle": 0, + "content": "data, where we can save \\(96.7\\%\\) of the annotation effort on MOT17, \\(90\\%\\) of the annotation effort on MOT20 and \\(80\\%\\) of the annotation effort on DanceTrack." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.614, + 0.36, + 0.63 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.646, + 0.788, + 0.813 + ], + "angle": 0, + "content": "We have introduced SPAM, a labeling engine to efficiently obtain trajectory annotations from videos. We have shown that SPAM is able to produce annotations that yield state-of-the-art tracking performance with a significantly reduced human annotation effort. Through our extensive experiments we have shown that (i) our synthetic pretraining \\(+\\) pseudo-label finetuning recipe yields a strong, annotation-free, labeler; (ii) our combination of active learning \\(+\\) pseudolabeling for annotation generation yields high quality and efficient annotations; and (iii) that graphs are an excellent tool to provide comprehensive modelling for trajectory annotation. We believe that SPAM paves the way to make large-scale cheap annotation of tracking datasets a reality, which will in turn power the next generation of data-hungry tracking algorithms." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.258, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SPAMming Labels: Efficient Annotations for the Trackers of Tomorrow" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.204 + ], + "angle": 0, + "content": "1. Acuna, D., Ling, H., Kar, A., Fidler, S.: Efficient interactive annotation of segmentation datasets with polygon-rnn++. In: CVPR (2018) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.205, + 0.785, + 0.245 + ], + "angle": 0, + "content": "2. Alahi, A., Goel, K., Ramanathan, V., Robicquet, A., Fei-Fei, L., Savarese, S.: Social LSTM: Human trajectory prediction in crowded spaces. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2016) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.247, + 0.785, + 0.285 + ], + "angle": 0, + "content": "3. Andriyenko, A., Schindler, K.: Multi-target tracking by continuous energy minimization. In: CVPR. pp. 1265-1272 (2011). https://doi.org/10.1109/CVPR.2011.5995311 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.288, + 0.785, + 0.327 + ], + "angle": 0, + "content": "4. Bastani, F., He, S., Madden, S.: Self-supervised multi-object tracking with cross-input consistency. Advances in Neural Information Processing Systems 34, 13695-13706 (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.329, + 0.785, + 0.355 + ], + "angle": 0, + "content": "5. Beluch, W.H., Genewein, T., Nurnberger, A., Kohler, J.M.: The power of ensembles for active learning in image classification. In: CVPR. pp. 9368-9377 (2018) 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.357, + 0.785, + 0.384 + ], + "angle": 0, + "content": "6. Berclaz, J., Fleuret, F., Turetken, E., Fua, P.: Multiple object tracking using k-shortest paths optimization. IEEE TPAMI 33(9), 1806-1819 (2011) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.385, + 0.785, + 0.412 + ], + "angle": 0, + "content": "7. Bergmann, P., Meinhardt, T., Leal-Taixe, L.: Tracking without bells and whistles. In: ICCV. pp. 941-951 (2019) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.413, + 0.785, + 0.439 + ], + "angle": 0, + "content": "8. Brasó, G., Cetintas, O., Leal-Taixe, L.: Multi-object tracking and segmentation via neural message passing. IJCV 130(12), 3035-3053 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.44, + 0.785, + 0.466 + ], + "angle": 0, + "content": "9. Braso, G., Leal-Taixe, L.: Learning a neural solver for multiple object tracking. In: CVPR (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.468, + 0.785, + 0.508 + ], + "angle": 0, + "content": "0. Cai, J., Xu, M., Li, W., Xiong, Y., Xia, W., Tu, Z., Soatto, S.: Memot: Multi-object tracking with memory. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8090-8100 (2022) 4, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.509, + 0.785, + 0.536 + ], + "angle": 0, + "content": "1. Caron, M., Misra, I., Mairal, J., Goyal, P., Bojanowski, P., Joulin, A.: Unsupervised learning of visual features by contrasting cluster assignments. NeurIPS (2020) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.537, + 0.785, + 0.563 + ], + "angle": 0, + "content": "2. Cetintas, O., Brasó, G., Leal-Taixe, L.: Unifying short and long-term tracking with graph hierarchies. In: CVPR. pp. 22877-22887 (June 2023) 3, 4, 5, 6, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.565, + 0.785, + 0.616 + ], + "angle": 0, + "content": "3. Chen, L., Ai, H., Zhuang, Z., Shang, C.: Real-time multiple people tracking with deeply learned candidate selection and person re-identification. In: 2018 IEEE international conference on multimedia and expo (ICME). pp. 1-6. IEEE (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.619, + 0.785, + 0.66 + ], + "angle": 0, + "content": "4. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: IEEE Int. Conf. Mach. Learn. pp. 1597-1607. PMLR (2020) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.661, + 0.785, + 0.688 + ], + "angle": 0, + "content": "5. Chu, P., Ling, H.: Famnet: Joint learning of feature, affinity and multi-dimensional assignment for online multiple object tracking. In: ICCV (October 2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.689, + 0.785, + 0.716 + ], + "angle": 0, + "content": "6. Dai, K., Zhao, J., Wang, L., Wang, D., Li, J., Lu, H., Qian, X., Yang, X.: Video annotation for visual tracking via selection and refinement. In: ICCV (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.717, + 0.785, + 0.757 + ], + "angle": 0, + "content": "7. Dai, Z., Cai, B., Lin, Y., Chen, J.: Up-detr: Unsupervised pre-training for object detection with transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1601-1610 (June 2021) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.758, + 0.785, + 0.798 + ], + "angle": 0, + "content": "8. Dendorfer, P., Osep, A., Milan, A., Schindler, K., Cremers, D., Reid, I., Roth, S., Leal-Taixe, L.: Motchallenge: A benchmark for single-camera multiple target tracking. International Journal of Computer Vision 129(4), 845-881 (2021) 3, 7, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "9. Dendorfer, P., Rezatofighi, H., Milan, A., Shi, J.Q., Cremers, D., Reid, I.D., Roth, S., Schindler, K., Leal-Taix'e, L.: Mot20: A benchmark for multi object tracking in crowded scenes. ArXiv abs/2003.09003 (2020) 3, 7, 9" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.388, + 0.127 + ], + "angle": 0, + "content": "O. Cetintas et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "20. Dólar, P., Wojek, C., Schiele, B., Perona, P.: Pedestrian detection: A benchmark. In: CVPR. pp. 304-311. IEEE (2009) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.785, + 0.217 + ], + "angle": 0, + "content": "21. Elezi, I., Yu, Z., Anandkumar, A., Leal-Taixe, L., Alvarez, J.M.: Not all labels are equal: Rationalizing the labeling costs for training object detection. In: CVPR. pp. 14492-14501 (2022) 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.258 + ], + "angle": 0, + "content": "22. Ess, A., Leibe, B., Schindler, K., Van Gool, L.: A mobile vision system for robust multi-person tracking. In: CVPR. pp. 1-8 (2008). https://doi.org/10.1109/CVPR.2008.4587581 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.785, + 0.313 + ], + "angle": 0, + "content": "23. Fabbri, M., Brasó, G., Maugeri, G., Cetintas, O., Gasparini, R., Osep, A., Calderara, S., Leal-Taixe, L., Cucchiara, R.: Motsynth: How can synthetic data help pedestrian detection and tracking? In: ICCV. pp. 10849-10859 (October 2021) 3, 5, 6, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.355 + ], + "angle": 0, + "content": "24. Fabbri, M., Lanzi, F., Calderara, S., Palazzi, A., Vezzani, R., Cucchiara, R.: Learning to detect and track visible and occluded body joints in a virtual world. In: ECCV (2018) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.785, + 0.383 + ], + "angle": 0, + "content": "25. Feichtenhofer, C., Pinz, A., Zisserman, A.: Detect to track and track to detect. In: ICCV (Oct 2017) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.411 + ], + "angle": 0, + "content": "26. Gaidon, A., Wang, Q., Cabon, Y., Vig, E.: Virtual worlds as proxy for multi-object tracking analysis. In: CVPR. pp. 4340-4349 (2016) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.438 + ], + "angle": 0, + "content": "27. Ge, Z., Liu, S., Wang, F., Li, Z., Sun, J.: Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430 (2021) 9, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.481 + ], + "angle": 0, + "content": "28. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.508 + ], + "angle": 0, + "content": "29. Hu, H.N., Cai, Q.Z., Wang, D., Lin, J., Sun, M., Krahenbihl, P., Darrell, T., Yu, F.: Joint monocular 3d vehicle detection and tracking. In: ICCV (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.785, + 0.548 + ], + "angle": 0, + "content": "30. Jabri, A., Owens, A., Efros, A.: Space-time correspondence as a contrastive random walk. Advances in neural information processing systems 33, 19545-19560 (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.549, + 0.785, + 0.589 + ], + "angle": 0, + "content": "31. Jiang, H., Fels, S., Little, J.J.: A linear programming approach for multiple object tracking. In: CVPR. pp. 1-8 (2007). https://doi.org/10.1109/CVPR.2007.3831804" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.591, + 0.785, + 0.646 + ], + "angle": 0, + "content": "32. Kasturi, R., Goldgof, D., Soundararajan, P., Manohar, V., Garofolo, J., Boonstra, M., Korzhova, V., Zhang, J.: Framework for performance evaluation for face, text and vehicle detection and tracking in video: data, metrics, and protocol. IEEE TPAMI (2009) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.688 + ], + "angle": 0, + "content": "33. Kaul, P., Xie, W., Zisserman, A.: Label, verify, correct: A simple few-shot object detection method. In: IEEE Conference on Computer Vision and Pattern Recognition (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.785, + 0.728 + ], + "angle": 0, + "content": "34. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: ICCV (2023) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.729, + 0.785, + 0.771 + ], + "angle": 0, + "content": "35. Lan, S., Yang, X., Yu, Z., Wu, Z., Alvarez, J.M., Anandkumar, A.: Vision transformers are good mask auto-labelers. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2023) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.799 + ], + "angle": 0, + "content": "36. Leal-Taixe, L., Canton-Ferrer, C., Schindler, K.: Learning by tracking: Siamese cnn for robust target association. In: CVPRW (June 2016) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.838 + ], + "angle": 0, + "content": "37. Leal-Taixe, L., Fenzi, M., Kuznetsova, A., Rosenhahn, B., Savarese, S.: Learning an image-based motion context for multiple people tracking. In: CVPR (June 2014) 4" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.838 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.258, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SPAMming Labels: Efficient Annotations for the Trackers of Tomorrow" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.202 + ], + "angle": 0, + "content": "38. Leal-Taixe, L., Pons-Moll, G., Rosenhahn, B.: Everybody needs somebody: Modeling social and grouping behavior on a linear programming multiple people tracker. In: Int. Conf. Comput. Vis. Worksh. pp. 120-127 (2011). https://doi.org/10.1109/ICCVW.2011.6130233 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.203, + 0.785, + 0.23 + ], + "angle": 0, + "content": "39. Li, R., Zhang, B., Liu, J., Liu, W., Zhao, J., Teng, Z.: Heterogeneous diversity driven active learning for multi-object tracking. In: ICCV (2023) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.23, + 0.785, + 0.285 + ], + "angle": 0, + "content": "40. Litrico, M., Del Bue, A., Morerio, P.: Guiding pseudo-labels with uncertainty estimation for source-free unsupervised domain adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.285, + 0.785, + 0.325 + ], + "angle": 0, + "content": "41. Liu, K., Jin, S., Fu, Z., Chen, Z., Jiang, R., Ye, J.: Uncertainty-aware unsupervised multi-object tracking. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9996-10005 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.325, + 0.785, + 0.366 + ], + "angle": 0, + "content": "42. Luiten, J., Osep, A., Dendorfer, P., Torr, P., Geiger, A., Leal-Taixe, L., Leibe, B.: Hota: A higher order metric for evaluating multi-object tracking. IJCV 129(2), 548-578 (2021) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.366, + 0.785, + 0.407 + ], + "angle": 0, + "content": "43. Manen, S., Gygli, M., Dai, D., Van Gool, L.: Pathtrack: Fast trajectory annotation with path supervision. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 290-299 (2017) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.407, + 0.785, + 0.447 + ], + "angle": 0, + "content": "44. Meinhardt, T., Kirillov, A., Leal-Taixe, L., Feichtenhofer, C.: Trackformer: Multi-object tracking with transformers. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022) 1, 4, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.447, + 0.785, + 0.5 + ], + "angle": 0, + "content": "45. Meng, S., Shao, D., Guo, J., Gao, S.: Tracking without label: Unsupervised multiple object tracking via contrastive similarity learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16264-16273 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.785, + 0.529 + ], + "angle": 0, + "content": "46. Pang, J., Qiu, L., Li, X., Chen, H., Li, Q., Darrell, T., Yu, F.: Quasi-dense similarity learning for multiple object tracking. In: CVPR. pp. 164-173 (2021) 4, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.529, + 0.785, + 0.569 + ], + "angle": 0, + "content": "47. Pellegrini, S., Ess, A., Schindler, K., van Gool, L.: You'll never walk alone: Modeling social behavior for multi-target tracking. In: ICCV. pp. 261-268 (2009). https://doi.org/10.1109/ICCV.2009.5459260 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.569, + 0.785, + 0.609 + ], + "angle": 0, + "content": "48. Pirsiavash, H., Ramanan, D., Fowlkes, C.C.: Globally-optimal greedy algorithms for tracking a variable number of objects. In: CVPR. pp. 1201-1208 (2011). https://doi.org/10.1109/CVPR.2011.5995604 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.609, + 0.785, + 0.664 + ], + "angle": 0, + "content": "49. Qin, Z., Zhou, S., Wang, L., Duan, J., Hua, G., Tang, W.: Motiontrack: Learning robust short-term and long-term motions for multi-object tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17939-17948 (2023) 4, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.664, + 0.785, + 0.705 + ], + "angle": 0, + "content": "50. Ristani, E., Solera, F., Zou, R., Cucchiara, R., Tomasi, C.: Performance measures and a data set for multi-target, multi-camera tracking. In: Eur. Conf. Comput. Vis. Worksh. pp. 17-35. Springer (2016) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.705, + 0.785, + 0.732 + ], + "angle": 0, + "content": "51. Ristani, E., Tomasi, C.: Features for multi-target multi-camera tracking and re-identification. In: CVPR (June 2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.732, + 0.785, + 0.759 + ], + "angle": 0, + "content": "52. Robicquet, A., Sadeghian, A., Alahi, A., Savarese, S.: Learning social etiquette: Human trajectory understanding in crowded scenes. In: ECCV (2016) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.785, + 0.799 + ], + "angle": 0, + "content": "53. Scovanner, P., Tappen, M.F.: Learning pedestrian dynamics from the real world. In: 2009 IEEE 12th International Conference on Computer Vision. pp. 381-388. IEEE (2009) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "54. Seidenschwarz, J., Brasó, G., Serrano, V.C., Elezi, I., Leal-Taixe, L.: Simple cues lead to a strong multi-object tracker. In: CVPR. pp. 13813-13823 (2023) 4, 9, 11, 13, 14" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.388, + 0.127 + ], + "angle": 0, + "content": "O. Cetintas et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.188 + ], + "angle": 0, + "content": "55. Sener, O., Savarese, S.: Active learning for convolutional neural networks: A core-set approach. In: ICLR (2018), https://openreview.net/forum?id=H1aIuk-RW 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.189, + 0.785, + 0.231 + ], + "angle": 0, + "content": "56. Shao, S., Zhao, Z., Li, B., Xiao, T., Yu, G., Zhang, X., Sun, J.: Crowdhuman: A benchmark for detecting human in a crowd. arXiv preprint arXiv:1805.00123 (2018) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.271 + ], + "angle": 0, + "content": "57. Sun, P., Cao, J., Jiang, Y., Yuan, Z., Bai, S., Kitani, K., Luo, P.: Dancetrack: Multiobject tracking in uniform appearance and diverse motion. In: CVPR (2022) 3, 9, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.313 + ], + "angle": 0, + "content": "58. Sun, S., Akhtar, N., Song, X., Song, H., Mian, A., Shah, M.: Simultaneous detection and tracking with motion modelling for multiple object tracking. In: ECCV. pp. 626-643 (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.357 + ], + "angle": 0, + "content": "59. Voigtlaender, P., Krause, M., Osep, A., Luiten, J., Sekar, B.B.G., Geiger, A., Leibe, B.: Mots: Multi-object tracking and segmentation. In: Proceedings of the iee/cvf conference on computer vision and pattern recognition. pp. 7942-7951 (2019) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.785, + 0.409 + ], + "angle": 0, + "content": "60. Voigtlaender, P., Luo, L., Yuan, C., Jiang, Y., Leibe, B.: Reducing the annotation effort for video object segmentation datasets. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3060-3069 (2021) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.785, + 0.438 + ], + "angle": 0, + "content": "61. Vondrick, C., Ramanan, D.: Video annotation and tracking with active learning. NeurIPS 24 (2011) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.44, + 0.785, + 0.493 + ], + "angle": 0, + "content": "62. Vondrick, C., Ramanan, D., Patterson, D.: Efficiently scaling up video annotation with crowdsourced marketplaces. In: Computer Vision-ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part IV 11. pp. 610-623. Springer (2010) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.785, + 0.536 + ], + "angle": 0, + "content": "63. Vondrick, C., Shrivastava, A., Fathi, A., Guadarrama, S., Murphy, K.: Tracking emerges by colorizing videos. In: Proceedings of the European conference on computer vision (ECCV). pp. 391-408 (2018) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.537, + 0.785, + 0.563 + ], + "angle": 0, + "content": "64. Wang, X., Jabri, A., Efros, A.A.: Learning correspondence from the cycle-consistency of time. In: CVPR (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.785, + 0.605 + ], + "angle": 0, + "content": "65. Wang, Y., Kitani, K., Weng, X.: Joint object detection and multi-object tracking with graph neural networks. In: IEEE Int. Conf. Robotics and Autom. pp. 13708-13715 (2021). https://doi.org/10.1109/ICRA48506.2021.9561110 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.785, + 0.633 + ], + "angle": 0, + "content": "66. Wang, Z., Zheng, L., Liu, Y., Wang, S.: Towards real-time multi-object tracking. The European Conference on Computer Vision (ECCV) (2020) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.634, + 0.785, + 0.673 + ], + "angle": 0, + "content": "67. Wei, F., Gao, Y., Wu, Z., Hu, H., Lin, S.: Aligning pretraining for detection via object-level contrastive learning. In: Advances in Neural Information Processing Systems (2021) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.785, + 0.702 + ], + "angle": 0, + "content": "68. Xiao, T., Li, S., Wang, B., Lin, L., Wang, X.: Joint detection and identification feature learning for person search. In: CVPR. pp. 3415-3424 (2017) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.703, + 0.785, + 0.741 + ], + "angle": 0, + "content": "69. Yamaguchi, K., Berg, A.C., Ortiz, L.E., Berg, T.L.: Who are you with and where are you going? In: CVPR. pp. 1345-1352 (2011). https://doi.org/10.1109/CVPR.2011.5995468 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.743, + 0.785, + 0.797 + ], + "angle": 0, + "content": "70. You, S., Yao, H., Bao, B.K., Xu, C.: Utm: A unified multiple object tracking model with identity-aware feature enhancement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21876-21886 (2023) 4, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "71. Yuan, D., Chang, X., Liu, Q., Yang, Y., Wang, D., Shu, M., He, Z., Shi, G.: Active learning for deep visual tracking. IEEE Transactions on Neural Networks and Learning Systems (2023) 2" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.258, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SPAMming Labels: Efficient Annotations for the Trackers of Tomorrow" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "72. Zeng, F., Dong, B., Zhang, Y., Wang, T., Zhang, X., Wei, Y.: Motr: End-to-end multiple-object tracking with transformer. In: European Conference on Computer Vision. pp. 659-675. Springer (2022) 1, 4, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.217 + ], + "angle": 0, + "content": "73. Zhang, L., Li, Y., Nevatia, R.: Global data association for multi-object tracking using network flows. In: CVPR (2008) 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.787, + 0.244 + ], + "angle": 0, + "content": "74. Zhang, S., Benenson, R., Schiele, B.: Citypersons: A diverse dataset for pedestrian detection. In: CVPR. pp. 3213-3221 (2017) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.787, + 0.286 + ], + "angle": 0, + "content": "75. Zhang, Y., Sun, P., Jiang, Y., Yu, D., Weng, F., Yuan, Z., Luo, P., Liu, W., Wang, X.: Bytetrack: Multi-object tracking by associating every detection box. In: ECCV. pp. 1-21. Springer (2022) 4, 9, 11, 13, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.787, + 0.327 + ], + "angle": 0, + "content": "76. Zhang, Y., Wang, C., Wang, X., Zeng, W., Liu, W.: Fairmot: On the fairness of detection and re-identification in multiple object tracking. IJCV 129(11), 3069-3087 (2021) 1, 4, 9, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.787, + 0.356 + ], + "angle": 0, + "content": "77. Zheng, L., Zhang, H., Sun, S., Chandraker, M., Yang, Y., Tian, Q.: Person re-identification in the wild. In: CVPR. pp. 1367-1376 (2017) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.787, + 0.384 + ], + "angle": 0, + "content": "78. Zhou, X., Koltun, V., Krahenbuhl, P.: Tracking objects as points. In: ECCV. pp. 474-490. Springer (2020) 1, 4" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.384 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/SPAMming Labels_ Efficient Annotations for the Trackers of Tomorrow/0cf5ca3c-0b18-48c1-b497-d85ab9c4bdac_origin.pdf b/2024/SPAMming Labels_ Efficient Annotations for the Trackers of Tomorrow/0cf5ca3c-0b18-48c1-b497-d85ab9c4bdac_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..63150f342b15758e89e111a5498ae076bdf2da38 --- /dev/null +++ b/2024/SPAMming Labels_ Efficient Annotations for the Trackers of Tomorrow/0cf5ca3c-0b18-48c1-b497-d85ab9c4bdac_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9b38c76e008a149b98d1732f1d8bfc027a857f611f96fd49d55f0c3945b97260 +size 721338 diff --git a/2024/SPAMming Labels_ Efficient Annotations for the Trackers of Tomorrow/full.md b/2024/SPAMming Labels_ Efficient Annotations for the Trackers of Tomorrow/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9bbfd46b6bf3bd2a3a1915df8fc85ed5ac0e9ba8 --- /dev/null +++ b/2024/SPAMming Labels_ Efficient Annotations for the Trackers of Tomorrow/full.md @@ -0,0 +1,270 @@ +# SPAMming Labels: Efficient Annotations for the Trackers of Tomorrow + +Orcun Cetintas $^{1,2\dagger}$ , Tim Meinhardt $^{2}$ , Guillem Brasó $^{1,2\dagger}$ , and Laura Leal-Taixe $^{2}$ + +$^{1}$ Technical University of Munich + $^{2}$ NVIDIA + +Abstract. Increasing the annotation efficiency of trajectory annotations from videos has the potential to enable the next generation of data-hungry tracking algorithms to thrive on large-scale datasets. Despite the importance of this task, there are currently very few works exploring how to efficiently label tracking datasets comprehensively. In this work, we introduce SPAM, a video label engine that provides high-quality labels with minimal human intervention. SPAM is built around two key insights: i) most tracking scenarios can be easily resolved. To take advantage of this, we utilize a pre-trained model to generate high-quality pseudo-labels, reserving human involvement for a smaller subset of more difficult instances; ii) handling the spatiotemporal dependencies of track annotations across time can be elegantly and efficiently formulated through graphs. Therefore, we use a unified graph formulation to address the annotation of both detections and identity association for tracks across time. Based on these insights, SPAM produces high-quality annotations with a fraction of ground truth labeling cost. We demonstrate that trackers trained on SPAM labels achieve comparable performance to those trained on human annotations while requiring only $3 - 20\%$ of the human labeling effort. Hence, SPAM paves the way towards highly efficient labeling of large-scale tracking datasets. We release all models and code. + +Keywords: Multi-object tracking $\cdot$ Data labeling $\cdot$ Active learning + +# 1 Introduction + +Detecting and tracking objects in time is at the core of many challenging vision applications. To address these problems, modern multi-object tracking (MOT) approaches [44, 66, 72, 76, 78] rely on increasingly larger amounts of annotated data. While annotating image datasets is already costly, temporal component of videos further increases the annotation difficulty and data scale requirements. This is due to the fact that redundancies between frames cause the information density to scale poorly with the amount of data, making the overall annotation task more challenging and resource-intensive. + +To mitigate these issues, researchers are spending more effort in operating with limited human annotations. Self- and semi-supervised learning (SSL) and pseudo-labeling (PL) approaches are often used to leverage unlabeled data whenever possible [34,35,40], while active learning (AL) strategies aim at selecting the most informative samples to label. All these strategies are commonly applied in various image understanding tasks, such as classification [11,14] or object detection [17,67]. Even if video tasks do not have any lower requirement for data, there are only a handful of works that target efficient labeling of tracking data in the video domain [39,59-61]. Existing works fail to provide comprehensive solutions to labeling by either ignoring dense temporal component of videos [39,61] or working under a limited single-object setup [71]. An effective video labeling engine should operate across multiple objects and frames, considering the intricate dependencies between tracks inherent in the labeling problem. + +Videos have unique characteristics that can be leveraged to unleash the full potential of a modern label engine. Firstly, frame-wise similarities in videos yield data redundancies. In the context of MOT, this implies that some associations between detections are relatively easy to solve [7]. This setup provides an ideal opportunity to leverage state-of-the-art models pre-trained to pseudo-label new data at zero cost. Secondly, the dependencies of objects across time and space cause annotations in one frame of the video to have cascading effects in other frames. For example, solving an association for one track can resolve associations for neighboring tracks. In order to maximize the efficiency of the annotations, we frame the labeling task within a model that takes into account such dependencies. Thus, we adopt a track-centric labeling approach rather than a frame-centric one [39]. + +We present the SPAM video label engine, which chains Synthetic pretraining, Pseudo-labeling, and Active learning with a graph-based Model for optimal tracking performance with minimal human effort. In the image domain, there are plenty of large-scale annotated datasets that a model-driven approach can leverage for pre-training [34]. However, this is not the case for multi-object tracking datasets, which are typically much smaller in scale; therefore, we advocate for a Synthetic pre-training. We further analyze which components of a tracking method are more affected by the synthetic-to-real gap. This first experiment gives us an important insight: labeling should be focused on detections and associations, and we can directly use the synthetically trained reID model. With our synthetic pre-trained model, we then propose a model-driven approach to produce Pseudo-labels on real data and retrain the model with its own pseudolabels. We show that with this recipe we can even reach a performance comparable to ground truth without ever training our labeler on manually annotated data. + +Synthetic pre-training and pseudo-labeling allow us to handle easy decisions and defer only uncertain or complex decisions to an annotator. We propose an Active learning strategy where the labeler makes decisions at the track-level instead of frame-level, thereby utilizing the annotation budget more efficiently. Towards this goal, we solve the labeling problem with a Model, based on graph + +hierarchies [12]. This model captures long-range spatio-temporal dependencies, which means our annotations will have an impact across frames, and not only on the frame where the annotation actually happens. We extend the model with a detection filtering layer so that we can naturally label detection and association within the same framework. + +In summary, we propose SPAM, a video label engine that significantly reduces annotation cost by dealing with easy scenarios with a strong MOT model pre-trained on (i) Synthetic data, followed by (ii) Pseudo-labeling real data for its own re-training. Harder detection and association decisions are carefully selected by an (iii) Active learning scheme that works on a (iv) Model based on graph hierarchies. This allows us to efficiently use the annotator's budget and account for spatio-temporal dependencies in the labels. The resulting model generates labels comparable to ground truth performance, requiring only $3 - 20\%$ of manual annotation effort across three diverse datasets. + +# 2 Related Work + +Pre-training for tracking. Over the years, the multi-object tracking (MOT) community has annotated many challenging benchmark datasets [18, 19, 57]. However, due to the significant costs associated with manually labeling temporal data, most benchmarks only provide few annotated video sequences. To obtain top performance, modern data-hungry tracking methods are usually pretrained on additional video and tracking datasets [20, 22, 68, 74, 77]. Many works also rely on relevant image recognition [56] datasets to boost their per-frame object detection performance. An alternative pre-training direction avoids human annotations entirely by relying on simulated synthetic data [23, 24, 26, 29, 58]. + +Our label engine also pre-trains the labeling model synthetically, minimizing its total annotation footprint. To this end, we demonstrate the effectiveness of synthetic pre-training [23] for generating pseudo-labels for self-supervision. + +Self-supervision in the temporal domain. To benefit not only from already annotated datasets, many works explore the utilization of available unlabeled data sources, i.e., how to perform self-supervised learning for videos. Notable examples leverage cycle-consistency-based correspondences [30, 64], color propagation [63] or augmentation-based consistencies [4, 41]. Furthermore, [45] uses contrastive learning over re-identification (reID) embeddings to perform unsupervised association learning. More recently, several works [33-35, 40] in the image domain demonstrate the effectiveness of training on pseudo-labels generated by strong underlying labeling models. In the video domain, this type of self-supervision is mostly unexplored. + +We demonstrate the effectiveness of pseudo-labels for the temporal domain by improving the label quality of our SPAM labeler via self-supervision on its own output. The necessary signal-to-noise ratio for this training loop is obtained by pre-training a strong graph-based model on synthetic data. + +Efficient labeling of video datasets. While significant effort has been devoted towards increasing the labeling efficiency of image datasets [1,34], similar + +approaches for the video and tracking domain are still relatively limited. Earlier works reduce the number of frames to annotate within individual videos but still rely heavily on human annotators [62]. Other solutions are limited to only a single object per sequence [16, 60] or still rely on human bounding box annotations [59] to label segmentation masks across time. The authors of [39] select key frames based on their expected impact on tracking performance. While effective within its scope, this approach only yields sparse frame annotations, which are insufficient for training all modern trackers. To reduce the human annotation effort for trajectories, [43] proposes an efficient video interface that still requires labeling all trajectories. + +In contrast, our label engine performs active learning for multiple objects across all video frames and is the first of its kind to tackle long-term tracking problems via a hierarchical labeling approach. The uncertainty measures provided by our graph-based model allow us to focus expensive annotation efforts on individual tracks instead of labeling entire frames in a brute-force manner. + +Multi-object tracking models. Tracking objects across time in a video is an established problem with a long research history. The most dominant tracking-by-detection (TbD) paradigm splits the task into two steps: (i) detecting objects in every frame and (ii) associating them to object trajectories, also referred to as data association. Modern trackers have explored performing (i) and (ii) jointly [7, 25, 44, 78], or providing solutions for (ii) relying on motion models [2, 3, 38, 47, 49, 52, 53, 69, 75], appearance models [13, 15, 36, 46, 51, 54, 70, 76] or more complex techniques involving transformers [10, 72]. An alternative direction of work relies on graph formulations [6, 31, 37, 48, 73] to utilize the TbD paradigm. In this formulation, nodes model object detections, and edges, association hypotheses. This long-studied formulation has recently been revived with remarkable success with the help of learnable graph neural network (GNN)architectures [8, 9, 12, 65]. + +For our label engine to generate long-term pseudo-labels and uncertainty estimations, we rely on a strong graph-based tracker [12]. To incorporate both detection and association labeling within a unified graph framework, we introduce a novel hierarchy level dedicated to detection filtering. + +# 3 SPAM + +# 3.1 Overview + +Our main goal is to label multi-object tracking datasets with minimal annotation effort while enabling optimal tracking performance. Our key observation is that most association cases can be solved easily by models pre-trained on synthetic data. Therefore, no human effort is needed, and we can instead use our model's predictions as pseudo-labels. Decisions that are more complex or uncertain are selected by an active learning scheme to be annotated manually. + +In order to perform pseudo-labeling and active learning, we require a model that can capture long-term spatio-temporal dependencies between tracks. We leverage a hierarchical graph neural network formulation and cast the annotation + +![](images/e885ba6615bc01047af4fb79ade4fdf2af243db6d58bab58dd2e78978747dc2f.jpg) +Fig. 1: Overview of the SPAM model. We first generate a set of detection candidates with our detector. Hierarchical GNNs then classify these candidates into valid and invalid objects via node classification, and assign identities through edge classification. + +task over a graph. Our architecture (Fig. 1) draws inspiration from SUSHI [12], which demonstrated remarkable association performance. Our model's scope, however, goes beyond association. Given the constraints of the labeling task, we need to equip our model with the ability to reason about detections before performing association. To address this broader goal, our GNN hierarchy, also includes the capability to accept or reject detections by classifying nodes into correct and incorrect object hypotheses. + +Our overall model is pre-trained on synthetic data [23] and further fine-tuned on its own predicted pseudo-labels over the target dataset that we aim to annotate. This recipe allows for strong labelling performance without any human annotations, and provides the basis for a scalable video label engine. After this training loop, we use our updated model's predictions as pseudo-labels to annotate a large portion of the target dataset. Decisions that are more complex or uncertain are selected by an active learning scheme to be annotated manually. The resulting set of labels achieves near-GT performance with significantly fewer manual annotations, thereby vastly enhancing the efficiency of the overall process. We provide an overview of our training and annotating recipe in Fig. 2. + +# 3.2 Graph Hierarchies as a Model for Labeling + +Problem formulation. Our model builds upon the standard graph formulation [73] where, given a set of object candidates $\mathcal{O}$ , our goal is to obtain a set of valid objects $\mathcal{O}_v \subset \mathcal{O}$ and their corresponding trajectories $\mathcal{T}$ . Each trajectory $T_k \in \mathcal{T}$ consists of a set of objects sharing the same identity or, equivalently, the set of edges connecting those nodes. Specifically, we model each object candidate $o_i \in \mathcal{O}$ as a node and each association hypothesis among candidates as an edge in an undirected graph $G = (V, E)$ . Within this formulation, nodes $u \in V$ could be classified into valid objects if $u \in \mathcal{O}_v$ . Analogously, edges can be classified into correct or incorrect hypotheses depending on whether they belong to some trajectory or not. This formulation can be generalized to a hierarchical setting by splitting video sequences into non-overlapping subsequences, i.e., disjoint subgraphs, and progressively merging them [12]. + +Background on graph hierarchies for MOT. Our model consists of a hierarchy of GNNs operating over the graph formulation described above. Intuitively, + +![](images/beb0774fcc317056f4f2f0a806b15ef73c5ecce5e7f19eff5c52537c02fa6a01.jpg) +Fig. 2: Overview of the SPAM training and annotation pipeline. (a) Initial model training on synthetic data. (b) Application of SPAM to generate pseudo-labels without incurring manual annotation costs on a real dataset, followed by self-training on pseudolabels. (c) Real dataset labeling using pseudo-labels and an uncertainty-based active learning approach. + +starting from detection candidates, at each hierarchy level, GNN hierarchy subsequently merge tracklets from the previous level into longer ones. Nodes and edges are represented by initial embeddings and message passing is used to share information within the graph in order to update node and edge embeddings with richer information. The GNNs are trained to classify edges into active and inactive association hypotheses and they share the same features, network architecture and learnable weights across hierarchy levels. In addition to the edge features reported in [12], for each object candidate $o_i$ we concatenate its times-tamp $t_i$ , normalized bounding box coordinates and dimensions $(x_i, y_i, h_i, w_i)$ and corresponding confidence from the detector $c_i$ and embed it as node features. + +Classifying nodes. We aim to minimize the annotation effort; therefore, we ask the annotator to operate on a fixed graph of detection candidates, i.e. annotator does not label new detections after the initial set given by detector. This is why we aim to create an initial graph that is over-complete. In order to do so, we use a low confidence threshold for our detector and obtain a large number of detection candidates. While this ensures a high detection recall, it also introduces a larger number of false positive detections that our model needs to address. As we show in experiments, the original SUSHI architecture [12] cannot handle such false positives, which is why we propose a new layer in the hierarchy, $GNN_{node}$ . Intuitively, $GNN_{node}$ leverages temporal information to validate object candidates. We use its output node embeddings and feed them to a lightweight classifier to enable false positive detection filtering. By doing so, we can consider a large number of input proposals achieving high detection recall by leveraging the spatio-temporal reasoning ability of our GNN for detection candidates. + +Pre-training and annotation-free fine-tuning. We pre-train all of our model components on the large-scale synthetic dataset MOTSynth [23]. To boost our model's performance and bridge the domain gap with our target dataset for annotation, we conduct further fine-tuning on the target dataset after pre-training. We achieve this by utilizing the raw predictions of our model on the target dataset and employing them as pseudo-labels for self-training. As we show in + +![](images/9bf0b4315066d528c3b5d6ba89673473ac6f5be91aa56f407adc960c1d179393.jpg) +Fig. 3: Our graph-based labeling pipeline begins with the selection of nodes for annotation. For each node to be annotated, the annotator could be asked to validate the detection, improve the localization by refining the box or perform association. + +our experimental section, this strategy yields a significant performance improvement without incurring any human annotation cost. + +# 3.3 How to Label on Graph Hierarchies? + +Manual labeling. Labeling video data at the box level involves several steps: identifying objects of interest, localizing their boundaries with bounding boxes, and assigning them unique identities. Typically, annotators achieve this by examining each frame, detecting and localizing objects individually, and then assigning identities based on temporal context from neighboring frames. In terms of annotation effort in clicks, detecting and localizing objects typically require two clicks (one for the top-left corner and one for the bottom-right corner of the bounding box), while associating objects across time requires one click (identifying the next occurrence of the object in subsequent frames), as reported in [18, 19]. This manual process is labor-intensive, costly, and severely limits the availability of annotated video data. + +Graph-based labeling. Utilizing a graph formulation for detection and tracking enables explicit modeling of objects in the scene and their interactions. In this framework, each node of the graph represents an object, allowing a manual annotation of an object to be framed as selecting a node on this graph. Given a set of nodes to be annotated, possible annotator decisions are: (i) accepting/rejecting a detection as we are working on a large set of detection candidates (one click), (ii) refining the bounding-box to improve localization (two clicks, one for the top-left corner and one for the bottom-right corner of the bounding box) and (iii) associating a node across time (one click) as shown in Fig. 3. + +Annotation acquisition. The core idea of our approach is to use our model decisions to guide the annotation process. For the vast majority of scenarios, we directly use our model's predictions as pseudo-labels. For a small subset of more challenging decisions where our model exhibits uncertainty, we validate our decisions with an annotator as described above. Within our graphs, + +![](images/c1ef6f6f66be5c046820990ba3d4e32bd29615bb445768664948ba57b75b8fbf.jpg) +Fig. 4: Analysis of performance gap between training a model on synthetic and real data for the three most common tracking components: detection, association, re-identification. + +we adopt an uncertainty-based approach to identify which nodes require annotations. Specifically, for each node $v \in V$ , we compute its uncertainty as $\mathrm{uncert}(v) \coloneqq \max_{u \in \mathbb{N}_v} H(\hat{y}_{(v,u)})$ , where $\mathbb{N}_v$ denotes the set of neighboring nodes of $v$ and $\hat{y}_{(v,u)}$ is the edge prediction for edge $(v,u)$ : + +$$ +H (\hat {y} _ {(v, u)}) := - (\hat {y} _ {(v, u)} \log \hat {y} _ {(v, u)} + (1 - \hat {y} _ {(v, u)}) \log (1 - \hat {y} _ {(v, u)})) \qquad (1) +$$ + +At the $GNN_{node}$ level, where we have node predictions instead of edge predictions, we simply compute $\text{uncert}(v)$ as $H(\hat{y}_{(v)})$ instead. Subsequently, we refer the nodes with the highest uncertainty scores to the annotator for manual annotation and use the remaining predictions as pseudo-labels for the rest of the nodes. + +Hierarchical label generation. To fully leverage our hierarchical graph formulation, we apply our label acquisition approach described above in a hierarchical manner. Intuitively, we distribute our annotation budget $B$ across $L$ hierarchical levels as $B_{1},\ldots ,B_{L}$ , with $B_{1} + \dots +B_{L} = B$ . In deeper hierarchy levels, nodes represent tracklets instead of single detections. This allows us to propagate annotator decisions for the entire cluster instead of individual detections, resulting in solving multiple uncertainties with a single annotation. As a result, we utilize the annotation budget more effectively while also mitigating error propagation within our pseudo-labels. + +# 4 Experiments + +# 4.1 Datasets and Metrics + +Datasets. We conduct experiments on three public benchmarks: + +- MOT17 [18] is an established benchmark containing 14 challenging video sequences, exhibiting diverse characteristics, including variations in camera motion, viewpoint, and pedestrian densities. +- MOT20 [19] contains a total of 8 video sequences, all featuring extremely crowded scenes. Due to their high pedestrian density, the emphasis is on addressing highly challenging scenarios characterized by frequent and prolonged occlusions. +- DanceTrack [57] contains 100 video sequences. These sequences feature dancing individuals with similar appearances, facilitating the evaluation of tracker robustness against highly challenging motion patterns where appearance information has a relatively minor influence. This non-standard setup often leads to a performance drop in trackers [57]. + +Metrics. We assess the performance of trackers using three established metrics: + +- MOTA [32] assesses trajectory coverage, by combining object detection measurements with identity switches, with the latter playing a relatively minor role. +- IDF1 [50] focuses on identity preservation over the sequences and provides a measure of association performance. +- HOTA [42] is a recently introduced metric designed to strike a balance between association and object detection performance within a single score. + +# 4.2 Implementation Details + +Architecture and training. Our model consists of three main components: (i) a detector network, based onYOLOX [27], (ii) a ResNet-based ReID network [28], and (iii) a GNN hierarchy, all pretrained on the synthetic sequences of MOTSynth [23]. After pretraining, we use the resulting model to obtain pseudolabels for our target MOT dataset, and then fine-tune our detector and GNN on them. Note that our Model is never trained on human annotations, it is only trained on synthetic annotations and its own pseudo-labels. + +Experimental setup. Online trackers are real-time tools suitable for deployment on robots or autonomous vehicles, whereas offline trackers are better suited for labeling tasks. Therefore, we structure our experiments to utilize SPAM for labeling and use these labels with online trackers. To evaluate the quality of SPAM labels, we train two state-of-the-art online trackers with these labels: ByteTrack [75] and GHOST [54]. For validation, we follow the standard half-split setting [54, 75, 76] over MOT17 sequences. Annotation budgets are computed based on number of clicks as described in Sec. 3.3. + +# 4.3 Synthetic Pretaining + +Several works have shown that synthetic pre-training is a strong starting point for multi-object tracking methods [23, 75]. In this section, we aim to analyze the gap between training a model on synthetic data and training it with real + +
ModelHigh Conf. BoxLow Conf. Box\( GNN_{node} \)MOTA ↑IDF1 ↑HOTA ↑DetA ↑
A✓✘✘64.474.759.955.9
B✓✓✘60.671.458.554.9
C✓✓✓65.475.160.456.6
+ +Table 1: Ablation study investigating the impact of incorporating low confidence bounding boxes with and without the $GNN_{node}$ . + +data for the three most common tracking components: detection, association, re-identification. For each setting, only the specific component that we are analysing is trained on synthetic data while other components are all trained on real data. We perform this experiment on three datasets and plot the relative HOTA difference (a positive score indicates superior performance with real data training) when training on real data vs. synthetic data in Fig. 4. Our first observation, is that the detector is the component that suffers the domain gap the most, with a performance drop of 9.9 HOTA points. Re-identification, on the other hand, has a similar performance when trained on synthetic or real data. Lastly, association performance is 2.1 HOTA higher with real data training. This first experiment gives us an important insight: labeling and retraining should be focused on detection and association components, and we can directly use the synthetically trained reID model. + +# 4.4 Towards a Model for Labeling + +Graph-based validation of detection candidates. In our framework, annotators have the capability to eliminate detections, but we refrain from requesting them to annotate new detections in order to maintain minimal annotation effort. Instead we start from an initial graph formed by an over-complete set of detections. Towards this end, we use a low confidence threshold for our detector, thereby reducing the number of false negatives in our initial graph and ensuring a high recall. As a result, this leads to an increase in the number of false positives incorporated into the graph, a tendency that is further accentuated by the fact that our detector is exclusively trained on synthetic data. We propose to leverage the graph structure itself to validate the detections based on temporal consistency. Specifically, we add an additional GNN-layer in our hierarchy, $GNN_{node}$ , which intuitively examines the spatial and temporal consistency of the bounding box candidates and filters out implausible ones. In Tab. 1, we assess the effect of this design choice. Model A uses only high confidence bounding boxes, model B creates a graph with also low confidence bounding boxes but without $GNN_{node}$ , while C includes $GNN_{node}$ . As we can see, simply adding low confidence boxes results in a drastic drop in performance on all metrics. Only by adding our $GNN_{node}$ we are able to take full advantage of the low confidence boxes without penalizing our performance with too many errors. + +SPAM is a SOTA tracker. Before moving to labeling with our model, we want to verify its overall performance and generalization capabilities. Only by using a state-of-the-art model we will be able to reduce annotation effort as much as + +
MethodHOTA ↑MOTA ↑IDF1 ↑HOTA ↑MOTA ↑IDF1 ↑HOTA ↑MOTA ↑IDF1 ↑
MOT17MOT20DanceTrack
MeMOT [10]56.972.569.054.166.163.7---
TrackFormer [44]57.374.168.054.765.768.6---
MOTR [72]57.868.673.4---54.251.579.7
FairMOT [76]59.373.772.354.661.867.339.782.240.8
QDTrack [46]63.578.777.560.074.773.854.287.750.4
ByteTrack† [75]62.878.977.160.474.274.547.789.653.9
GHOST [54]62.878.777.161.273.775.256.791.357.7
UTM [70]64.081.878.762.578.276.9---
MotionTrack [49]65.181.180.162.878.076.5---
SUSHI [12]66.581.183.164.374.379.863.388.763.4
SPAM (Ours)67.580.784.665.876.581.964.089.263.4
+ +Table 2: Test set results on MOT17, MOT20 and DanceTrack when using SPAM solely as a tracker. + +possible. We therefore pre-train our model on the synthetic dataset, and finetune it on the specific real dataset using groundtruth. For a fair comparison, we obtain object detections from a YOLOX detector [27] trained following [12, 54, 75]. Our model outperforms state-of-the-art by 1.0 HOTA point on MOT17, 1.5 HOTA points on MOT20, and 0.7 HOTA points on DanceTrack as shown in Tab. 2. Furthermore, on MOT17 and MOT20, there is also a significant improvement in IDF1, showing the identity-preserving capabilities of the model, which is especially important for our video annotation setting. Even if being state-of-the-art is not the main goal of this paper, having a strong performing model allows us to have an excellent starting point for producing pseudo-labels. We would like to emphasize that, we use ground truth annotations only for this experiment just to compare our model fairly with other trackers. From this point onward, we will not use ground truth labels. We will use our synthetic pretraining weights and focus on the pseudo-labeling and active learning aspects of our method to show how to annotate multi-object tracking datasets. + +# 4.5 Enhancing Annotations with Pseudo-labels and Active Learning + +Retraining our Model on Pseudo-labels. In this experiment, we compare two models. The first one, SPAM, is only trained on synthetic data. The second one, SPAM, is trained in three stages: (i) pre-training on synthetic data, (ii) pseudo-labeling the real dataset with our model, and (iii) re-training our model on pseudo-labels. In Sec. 4.5, we observe a significant improvement when using pseudo-labels to re-train our model: an increase of 4-6 HOTA, 4-6 MOTA, and 4-7 IDF1 points on MOT17, MOT20, and DanceTrack, respectively. Note that these improvements come at zero human annotation effort as this process requires no manual annotations. + +Active learning on graph hierarchies. We now test the effectiveness of our proposed active learning labeling on graph hierarchies, and show a comprehensive comparison with several baselines in Fig. 5. We directly measure the accuracy of the output labels given an increasing percentage of manual annotations selected + +
ModelPseudo-labelHOTA ↑MOTA ↑IDF1 ↑
MOT17
SPAM✘60.065.375.1
SPAM✓63.869.279.7
MOT20
SPAM✘52.265.470.3
SPAM✓58.770.277.1
DanceTrack
SPAM✘41.865.243.4
SPAM✓48.171.050.2
+ +Table 3: Performance boost obtained by our model when retraining with its own pseudo-labels incurring no manual annotation cost. + +![](images/986c70849e4c91b4e5ba3eb90f0241671072a3458beb8f7a0ad06cb95ea25985.jpg) +Fig. 5: Impact of various active learning strategies on the quality of labeling. + +by different acquisition functions on MOT17. The vertical axis shows the increase in HOTA over the baseline performance, i.e., the performance without active learning. Oracle is our upper bound and shows the performance if all possible labels are set to their correct value by the annotator. Every method has two different labeling schemes: one for detections (nodes), and one for associations (edges) as explained in Sec. 3.3. In addition to our method, we explore active learning strategies commonly used in object detection, including Coreset [55], Ensemble [5], and Consistency over augmented inputs [21]. The naming is based on the active learning strategy for detection in the first part and the strategy for association in the second part. For instance, EntImg/Rand. indicates image level entropy-based selection for detection and random selection for association. + +The methods shown in red shades and circle markers compare three detection labeling methods while keeping association label selection random. Box-level labeling is far superior to commonly-used image-level labeling, or random labeling. This highlights the potential of performing active learning at the object level rather than the frame level. The methods in blue shades and triangle markers all use our proposed active learning selection on graph hierarchies for the association, and compare several detection selection schemes. The Coreset [55], which prioritizes diversity of the data over uncertainty, emerges as the poorest performing approach. This observation underlines that, when accurate predictions are available, an uncertainty-based approach proves to be a more effective complement to pseudo-labels. Box-level entropy, ensembles, and consistency all perform similarly. Finally, our method, with both detection and association label selection on graph hierarchies is superior to all compared baselines, staying at only 1.2 points from Oracle performance. Notably, we also unify the AL strategy for detection and association with our approach. + +# 4.6 Putting Everything Together: SPAM Labels in Action + +Training trackers with our labels. In this experiment, we ablate the effects of our SPAM pipeline on the actual usecase of training a model with our pro + +
ModelLabelsPseudo-labelsActive Learn.Ann. Effort ↓HOTA ↑MOTA ↑IDF1 ↑
ByteTrack [75]Ground Truth-100%52.660.465.7
ByteTrack [75]SPAMXXX49.657.863.1
ByteTrack [75]SPAM✓XX50.559.564.5
ByteTrack [75]SPAM✓✓3.3%52.561.866.2
+ +Table 4: Effect of (i) Synthetic training, (ii) Retraining our model with our Pseudolabels, and (iii) Active learning-based acquisition and annotation of a small percentage of real data, on our label quality on the MOT17 validation set. + +vided labels. We ablate label quality under the following settings: (i) Synthetic training, (ii) Retraining our models on our Pseudo-labels, (iii) Active learning based acquisition and annotation of a small percentage of real data. We train state-of-the-art online tracker ByteTrack [75] from scratch with the aforementioned sets of labels and report its performance on the MOT17 validation set in Tab. 4. The goal performance we want to reach is the one achieved by the model trained on groundtruth annotations (row 1). Row 2 shows our model trained only on synthetic data, which is 3 HOTA points below our goal. Using pseudo-label based training of our model increases performance by almost a HOTA point, and the ground truth level performance is achieved by also using our active learning method to select which samples to label. By annotating only $3.3\%$ of the data, we can achieve the same results as the model trained on the ground truth data with $100\%$ annotation effort. This shows the potential of our SPAM in reducing annotation effort for multi-object tracking datasets. + +How much data do we really need to label? In this final experiment, our goal is to assess the impact of our entire pipeline by answering the question: How much data do we actually need to label in order to achieve the same performance as a model trained on the entire groundtruth training dataset? For that, we choose two well-established trackers, ByteTrack [75] and GHOST [54], and train them from scratch using our SPAM labels. We report their performance on MOT17, MOT20, and DanceTrack test sets in Tab. 5. We first confirm that our results on the validation set also generalize to the test set, and that we can reach GT performance with $3.3\%$ of annotation effort on MOT17. Note that for MOT17 our annotations yield a slight improvement over GT, which can be attributed to observed annotation noise in the original GT, particularly in moving sequences (further details can be found in the supplementary material). MOT20 and DanceTrack are known to be more challenging datasets [57], which is confirmed by the fact that the trackers still miss a few points when trained on only $3.3\%$ of manually annotated data. Given that we cannot do a full analysis of different percentages because we are working on the test sets, we choose two percentages of manually annotated data on which to train our trackers: $10\%$ for MOT20, and $20\%$ for DanceTrack. In both cases, we achieve on-par performance with the trackers trained on full groundtruth data. This clearly shows the potential in properly leveraging synthetic data, pseudo-labels from a strong model, and active learning on graph hierarchies in order to label multi-object tracking + +
MethodLabelsAnn. Effort ↓HOTA ↑MOTA ↑IDF1 ↑
MOT17 Test
ByteTrack [75]GT100%50.660.461.1
ByteTrack [75]SPAM3.3%51.664.063.0
GHOST [54]GT100%49.558.059.0
GHOST [54]SPAM3.3%51.361.962.1
MOT20 Test
ByteTrack [75]GT100%48.760.560.0
ByteTrack [75]SPAM3.3%47.957.661.4
ByteTrack [75]SPAM10%48.461.062.8
GHOST [54]GT100%48.359.760.2
GHOST [54]SPAM3.3%47.058.260.7
GHOST [54]SPAM10%46.959.060.8
DanceTrack Test
ByteTrack [75]GT100%41.079.146.0
ByteTrack [75]SPAM3.3%39.576.445.0
ByteTrack [75]SPAM20%41.380.948.1
GHOST [54]GT100%45.880.646.3
GHOST [54]SPAM3.3%41.076.344.8
GHOST [54]SPAM20%43.981.347.7
+ +Table 5: Evaluation of SPAM labels on MOT17, MOT20, and DanceTrack test sets with varying annotation budgets by training ByteTrack [75] and GHOST [54] using annotations generated by our method. + +data, where we can save $96.7\%$ of the annotation effort on MOT17, $90\%$ of the annotation effort on MOT20 and $80\%$ of the annotation effort on DanceTrack. + +# 5 Conclusion + +We have introduced SPAM, a labeling engine to efficiently obtain trajectory annotations from videos. We have shown that SPAM is able to produce annotations that yield state-of-the-art tracking performance with a significantly reduced human annotation effort. Through our extensive experiments we have shown that (i) our synthetic pretraining $+$ pseudo-label finetuning recipe yields a strong, annotation-free, labeler; (ii) our combination of active learning $+$ pseudolabeling for annotation generation yields high quality and efficient annotations; and (iii) that graphs are an excellent tool to provide comprehensive modelling for trajectory annotation. We believe that SPAM paves the way to make large-scale cheap annotation of tracking datasets a reality, which will in turn power the next generation of data-hungry tracking algorithms. + +# References + +1. Acuna, D., Ling, H., Kar, A., Fidler, S.: Efficient interactive annotation of segmentation datasets with polygon-rnn++. In: CVPR (2018) 3 +2. Alahi, A., Goel, K., Ramanathan, V., Robicquet, A., Fei-Fei, L., Savarese, S.: Social LSTM: Human trajectory prediction in crowded spaces. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2016) 4 +3. Andriyenko, A., Schindler, K.: Multi-target tracking by continuous energy minimization. In: CVPR. pp. 1265-1272 (2011). https://doi.org/10.1109/CVPR.2011.5995311 4 +4. Bastani, F., He, S., Madden, S.: Self-supervised multi-object tracking with cross-input consistency. Advances in Neural Information Processing Systems 34, 13695-13706 (2021) 3 +5. Beluch, W.H., Genewein, T., Nurnberger, A., Kohler, J.M.: The power of ensembles for active learning in image classification. In: CVPR. pp. 9368-9377 (2018) 12 +6. Berclaz, J., Fleuret, F., Turetken, E., Fua, P.: Multiple object tracking using k-shortest paths optimization. IEEE TPAMI 33(9), 1806-1819 (2011) 4 +7. Bergmann, P., Meinhardt, T., Leal-Taixe, L.: Tracking without bells and whistles. In: ICCV. pp. 941-951 (2019) 2, 4 +8. Brasó, G., Cetintas, O., Leal-Taixe, L.: Multi-object tracking and segmentation via neural message passing. IJCV 130(12), 3035-3053 (2022) 4 +9. Braso, G., Leal-Taixe, L.: Learning a neural solver for multiple object tracking. In: CVPR (2020) 4 +0. Cai, J., Xu, M., Li, W., Xiong, Y., Xia, W., Tu, Z., Soatto, S.: Memot: Multi-object tracking with memory. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8090-8100 (2022) 4, 11 +1. Caron, M., Misra, I., Mairal, J., Goyal, P., Bojanowski, P., Joulin, A.: Unsupervised learning of visual features by contrasting cluster assignments. NeurIPS (2020) 2 +2. Cetintas, O., Brasó, G., Leal-Taixe, L.: Unifying short and long-term tracking with graph hierarchies. In: CVPR. pp. 22877-22887 (June 2023) 3, 4, 5, 6, 11 +3. Chen, L., Ai, H., Zhuang, Z., Shang, C.: Real-time multiple people tracking with deeply learned candidate selection and person re-identification. In: 2018 IEEE international conference on multimedia and expo (ICME). pp. 1-6. IEEE (2018) 4 +4. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: IEEE Int. Conf. Mach. Learn. pp. 1597-1607. PMLR (2020) 2 +5. Chu, P., Ling, H.: Famnet: Joint learning of feature, affinity and multi-dimensional assignment for online multiple object tracking. In: ICCV (October 2019) 4 +6. Dai, K., Zhao, J., Wang, L., Wang, D., Li, J., Lu, H., Qian, X., Yang, X.: Video annotation for visual tracking via selection and refinement. In: ICCV (2021) 4 +7. Dai, Z., Cai, B., Lin, Y., Chen, J.: Up-detr: Unsupervised pre-training for object detection with transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1601-1610 (June 2021) 2 +8. Dendorfer, P., Osep, A., Milan, A., Schindler, K., Cremers, D., Reid, I., Roth, S., Leal-Taixe, L.: Motchallenge: A benchmark for single-camera multiple target tracking. International Journal of Computer Vision 129(4), 845-881 (2021) 3, 7, 9 +9. Dendorfer, P., Rezatofighi, H., Milan, A., Shi, J.Q., Cremers, D., Reid, I.D., Roth, S., Schindler, K., Leal-Taix'e, L.: Mot20: A benchmark for multi object tracking in crowded scenes. ArXiv abs/2003.09003 (2020) 3, 7, 9 + +20. Dólar, P., Wojek, C., Schiele, B., Perona, P.: Pedestrian detection: A benchmark. In: CVPR. pp. 304-311. IEEE (2009) 3 +21. Elezi, I., Yu, Z., Anandkumar, A., Leal-Taixe, L., Alvarez, J.M.: Not all labels are equal: Rationalizing the labeling costs for training object detection. In: CVPR. pp. 14492-14501 (2022) 12 +22. Ess, A., Leibe, B., Schindler, K., Van Gool, L.: A mobile vision system for robust multi-person tracking. In: CVPR. pp. 1-8 (2008). https://doi.org/10.1109/CVPR.2008.4587581 3 +23. Fabbri, M., Brasó, G., Maugeri, G., Cetintas, O., Gasparini, R., Osep, A., Calderara, S., Leal-Taixe, L., Cucchiara, R.: Motsynth: How can synthetic data help pedestrian detection and tracking? In: ICCV. pp. 10849-10859 (October 2021) 3, 5, 6, 9 +24. Fabbri, M., Lanzi, F., Calderara, S., Palazzi, A., Vezzani, R., Cucchiara, R.: Learning to detect and track visible and occluded body joints in a virtual world. In: ECCV (2018) 3 +25. Feichtenhofer, C., Pinz, A., Zisserman, A.: Detect to track and track to detect. In: ICCV (Oct 2017) 4 +26. Gaidon, A., Wang, Q., Cabon, Y., Vig, E.: Virtual worlds as proxy for multi-object tracking analysis. In: CVPR. pp. 4340-4349 (2016) 3 +27. Ge, Z., Liu, S., Wang, F., Li, Z., Sun, J.: Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430 (2021) 9, 11 +28. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) 9 +29. Hu, H.N., Cai, Q.Z., Wang, D., Lin, J., Sun, M., Krahenbihl, P., Darrell, T., Yu, F.: Joint monocular 3d vehicle detection and tracking. In: ICCV (2019) 3 +30. Jabri, A., Owens, A., Efros, A.: Space-time correspondence as a contrastive random walk. Advances in neural information processing systems 33, 19545-19560 (2020) 3 +31. Jiang, H., Fels, S., Little, J.J.: A linear programming approach for multiple object tracking. In: CVPR. pp. 1-8 (2007). https://doi.org/10.1109/CVPR.2007.3831804 +32. Kasturi, R., Goldgof, D., Soundararajan, P., Manohar, V., Garofolo, J., Boonstra, M., Korzhova, V., Zhang, J.: Framework for performance evaluation for face, text and vehicle detection and tracking in video: data, metrics, and protocol. IEEE TPAMI (2009) 9 +33. Kaul, P., Xie, W., Zisserman, A.: Label, verify, correct: A simple few-shot object detection method. In: IEEE Conference on Computer Vision and Pattern Recognition (2022) 3 +34. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: ICCV (2023) 2, 3 +35. Lan, S., Yang, X., Yu, Z., Wu, Z., Alvarez, J.M., Anandkumar, A.: Vision transformers are good mask auto-labelers. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2023) 2, 3 +36. Leal-Taixe, L., Canton-Ferrer, C., Schindler, K.: Learning by tracking: Siamese cnn for robust target association. In: CVPRW (June 2016) 4 +37. Leal-Taixe, L., Fenzi, M., Kuznetsova, A., Rosenhahn, B., Savarese, S.: Learning an image-based motion context for multiple people tracking. In: CVPR (June 2014) 4 + +38. Leal-Taixe, L., Pons-Moll, G., Rosenhahn, B.: Everybody needs somebody: Modeling social and grouping behavior on a linear programming multiple people tracker. In: Int. Conf. Comput. Vis. Worksh. pp. 120-127 (2011). https://doi.org/10.1109/ICCVW.2011.6130233 4 +39. Li, R., Zhang, B., Liu, J., Liu, W., Zhao, J., Teng, Z.: Heterogeneous diversity driven active learning for multi-object tracking. In: ICCV (2023) 2, 4 +40. Litrico, M., Del Bue, A., Morerio, P.: Guiding pseudo-labels with uncertainty estimation for source-free unsupervised domain adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023) 2, 3 +41. Liu, K., Jin, S., Fu, Z., Chen, Z., Jiang, R., Ye, J.: Uncertainty-aware unsupervised multi-object tracking. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9996-10005 (2023) 3 +42. Luiten, J., Osep, A., Dendorfer, P., Torr, P., Geiger, A., Leal-Taixe, L., Leibe, B.: Hota: A higher order metric for evaluating multi-object tracking. IJCV 129(2), 548-578 (2021) 9 +43. Manen, S., Gygli, M., Dai, D., Van Gool, L.: Pathtrack: Fast trajectory annotation with path supervision. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 290-299 (2017) 4 +44. Meinhardt, T., Kirillov, A., Leal-Taixe, L., Feichtenhofer, C.: Trackformer: Multi-object tracking with transformers. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022) 1, 4, 11 +45. Meng, S., Shao, D., Guo, J., Gao, S.: Tracking without label: Unsupervised multiple object tracking via contrastive similarity learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16264-16273 (2023) 3 +46. Pang, J., Qiu, L., Li, X., Chen, H., Li, Q., Darrell, T., Yu, F.: Quasi-dense similarity learning for multiple object tracking. In: CVPR. pp. 164-173 (2021) 4, 11 +47. Pellegrini, S., Ess, A., Schindler, K., van Gool, L.: You'll never walk alone: Modeling social behavior for multi-target tracking. In: ICCV. pp. 261-268 (2009). https://doi.org/10.1109/ICCV.2009.5459260 4 +48. Pirsiavash, H., Ramanan, D., Fowlkes, C.C.: Globally-optimal greedy algorithms for tracking a variable number of objects. In: CVPR. pp. 1201-1208 (2011). https://doi.org/10.1109/CVPR.2011.5995604 4 +49. Qin, Z., Zhou, S., Wang, L., Duan, J., Hua, G., Tang, W.: Motiontrack: Learning robust short-term and long-term motions for multi-object tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17939-17948 (2023) 4, 11 +50. Ristani, E., Solera, F., Zou, R., Cucchiara, R., Tomasi, C.: Performance measures and a data set for multi-target, multi-camera tracking. In: Eur. Conf. Comput. Vis. Worksh. pp. 17-35. Springer (2016) 9 +51. Ristani, E., Tomasi, C.: Features for multi-target multi-camera tracking and re-identification. In: CVPR (June 2018) 4 +52. Robicquet, A., Sadeghian, A., Alahi, A., Savarese, S.: Learning social etiquette: Human trajectory understanding in crowded scenes. In: ECCV (2016) 4 +53. Scovanner, P., Tappen, M.F.: Learning pedestrian dynamics from the real world. In: 2009 IEEE 12th International Conference on Computer Vision. pp. 381-388. IEEE (2009) 4 +54. Seidenschwarz, J., Brasó, G., Serrano, V.C., Elezi, I., Leal-Taixe, L.: Simple cues lead to a strong multi-object tracker. In: CVPR. pp. 13813-13823 (2023) 4, 9, 11, 13, 14 + +55. Sener, O., Savarese, S.: Active learning for convolutional neural networks: A core-set approach. In: ICLR (2018), https://openreview.net/forum?id=H1aIuk-RW 12 +56. Shao, S., Zhao, Z., Li, B., Xiao, T., Yu, G., Zhang, X., Sun, J.: Crowdhuman: A benchmark for detecting human in a crowd. arXiv preprint arXiv:1805.00123 (2018) 3 +57. Sun, P., Cao, J., Jiang, Y., Yuan, Z., Bai, S., Kitani, K., Luo, P.: Dancetrack: Multiobject tracking in uniform appearance and diverse motion. In: CVPR (2022) 3, 9, 13 +58. Sun, S., Akhtar, N., Song, X., Song, H., Mian, A., Shah, M.: Simultaneous detection and tracking with motion modelling for multiple object tracking. In: ECCV. pp. 626-643 (2020) 3 +59. Voigtlaender, P., Krause, M., Osep, A., Luiten, J., Sekar, B.B.G., Geiger, A., Leibe, B.: Mots: Multi-object tracking and segmentation. In: Proceedings of the iee/cvf conference on computer vision and pattern recognition. pp. 7942-7951 (2019) 2, 4 +60. Voigtlaender, P., Luo, L., Yuan, C., Jiang, Y., Leibe, B.: Reducing the annotation effort for video object segmentation datasets. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3060-3069 (2021) 2, 4 +61. Vondrick, C., Ramanan, D.: Video annotation and tracking with active learning. NeurIPS 24 (2011) 2 +62. Vondrick, C., Ramanan, D., Patterson, D.: Efficiently scaling up video annotation with crowdsourced marketplaces. In: Computer Vision-ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part IV 11. pp. 610-623. Springer (2010) 4 +63. Vondrick, C., Shrivastava, A., Fathi, A., Guadarrama, S., Murphy, K.: Tracking emerges by colorizing videos. In: Proceedings of the European conference on computer vision (ECCV). pp. 391-408 (2018) 3 +64. Wang, X., Jabri, A., Efros, A.A.: Learning correspondence from the cycle-consistency of time. In: CVPR (2019) 3 +65. Wang, Y., Kitani, K., Weng, X.: Joint object detection and multi-object tracking with graph neural networks. In: IEEE Int. Conf. Robotics and Autom. pp. 13708-13715 (2021). https://doi.org/10.1109/ICRA48506.2021.9561110 4 +66. Wang, Z., Zheng, L., Liu, Y., Wang, S.: Towards real-time multi-object tracking. The European Conference on Computer Vision (ECCV) (2020) 1 +67. Wei, F., Gao, Y., Wu, Z., Hu, H., Lin, S.: Aligning pretraining for detection via object-level contrastive learning. In: Advances in Neural Information Processing Systems (2021) 2 +68. Xiao, T., Li, S., Wang, B., Lin, L., Wang, X.: Joint detection and identification feature learning for person search. In: CVPR. pp. 3415-3424 (2017) 3 +69. Yamaguchi, K., Berg, A.C., Ortiz, L.E., Berg, T.L.: Who are you with and where are you going? In: CVPR. pp. 1345-1352 (2011). https://doi.org/10.1109/CVPR.2011.5995468 4 +70. You, S., Yao, H., Bao, B.K., Xu, C.: Utm: A unified multiple object tracking model with identity-aware feature enhancement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21876-21886 (2023) 4, 11 +71. Yuan, D., Chang, X., Liu, Q., Yang, Y., Wang, D., Shu, M., He, Z., Shi, G.: Active learning for deep visual tracking. IEEE Transactions on Neural Networks and Learning Systems (2023) 2 + +72. Zeng, F., Dong, B., Zhang, Y., Wang, T., Zhang, X., Wei, Y.: Motr: End-to-end multiple-object tracking with transformer. In: European Conference on Computer Vision. pp. 659-675. Springer (2022) 1, 4, 11 +73. Zhang, L., Li, Y., Nevatia, R.: Global data association for multi-object tracking using network flows. In: CVPR (2008) 4, 5 +74. Zhang, S., Benenson, R., Schiele, B.: Citypersons: A diverse dataset for pedestrian detection. In: CVPR. pp. 3213-3221 (2017) 3 +75. Zhang, Y., Sun, P., Jiang, Y., Yu, D., Weng, F., Yuan, Z., Luo, P., Liu, W., Wang, X.: Bytetrack: Multi-object tracking by associating every detection box. In: ECCV. pp. 1-21. Springer (2022) 4, 9, 11, 13, 14 +76. Zhang, Y., Wang, C., Wang, X., Zeng, W., Liu, W.: Fairmot: On the fairness of detection and re-identification in multiple object tracking. IJCV 129(11), 3069-3087 (2021) 1, 4, 9, 11 +77. Zheng, L., Zhang, H., Sun, S., Chandraker, M., Yang, Y., Tian, Q.: Person re-identification in the wild. In: CVPR. pp. 1367-1376 (2017) 3 +78. Zhou, X., Koltun, V., Krahenbuhl, P.: Tracking objects as points. In: ECCV. pp. 474-490. Springer (2020) 1, 4 \ No newline at end of file diff --git a/2024/SPAMming Labels_ Efficient Annotations for the Trackers of Tomorrow/images.zip b/2024/SPAMming Labels_ Efficient Annotations for the Trackers of Tomorrow/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7461532f3a7b02b8f59c5e562e1ca87e6437638e --- /dev/null +++ b/2024/SPAMming Labels_ Efficient Annotations for the Trackers of Tomorrow/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:75ce650021689f1844abe968acff2e139c984a64232e017dd8e2d8c293554b50 +size 321611 diff --git a/2024/SPAMming Labels_ Efficient Annotations for the Trackers of Tomorrow/layout.json b/2024/SPAMming Labels_ Efficient Annotations for the Trackers of Tomorrow/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..79af7e8750a0b9f69d7f0ab9d49b411b5bc855c5 --- /dev/null +++ b/2024/SPAMming Labels_ Efficient Annotations for the Trackers of Tomorrow/layout.json @@ -0,0 +1,8475 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 163, + 111, + 451, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 163, + 111, + 451, + 146 + ], + "spans": [ + { + "bbox": [ + 163, + 111, + 451, + 146 + ], + "type": "text", + "content": "SPAMming Labels: Efficient Annotations for the Trackers of Tomorrow" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 132, + 167, + 481, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 167, + 481, + 180 + ], + "spans": [ + { + "bbox": [ + 132, + 167, + 481, + 180 + ], + "type": "text", + "content": "Orcun Cetintas" + }, + { + "bbox": [ + 132, + 167, + 481, + 180 + ], + "type": "inline_equation", + "content": "^{1,2\\dagger}" + }, + { + "bbox": [ + 132, + 167, + 481, + 180 + ], + "type": "text", + "content": ", Tim Meinhardt" + }, + { + "bbox": [ + 132, + 167, + 481, + 180 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 132, + 167, + 481, + 180 + ], + "type": "text", + "content": ", Guillem Brasó" + }, + { + "bbox": [ + 132, + 167, + 481, + 180 + ], + "type": "inline_equation", + "content": "^{1,2\\dagger}" + }, + { + "bbox": [ + 132, + 167, + 481, + 180 + ], + "type": "text", + "content": ", and Laura Leal-Taixe" + }, + { + "bbox": [ + 132, + 167, + 481, + 180 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 238, + 189, + 375, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 189, + 375, + 210 + ], + "spans": [ + { + "bbox": [ + 238, + 189, + 375, + 210 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 238, + 189, + 375, + 210 + ], + "type": "text", + "content": " Technical University of Munich \n" + }, + { + "bbox": [ + 238, + 189, + 375, + 210 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 238, + 189, + 375, + 210 + ], + "type": "text", + "content": " NVIDIA" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 239, + 455, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 239, + 455, + 469 + ], + "spans": [ + { + "bbox": [ + 160, + 239, + 455, + 469 + ], + "type": "text", + "content": "Abstract. Increasing the annotation efficiency of trajectory annotations from videos has the potential to enable the next generation of data-hungry tracking algorithms to thrive on large-scale datasets. Despite the importance of this task, there are currently very few works exploring how to efficiently label tracking datasets comprehensively. In this work, we introduce SPAM, a video label engine that provides high-quality labels with minimal human intervention. SPAM is built around two key insights: i) most tracking scenarios can be easily resolved. To take advantage of this, we utilize a pre-trained model to generate high-quality pseudo-labels, reserving human involvement for a smaller subset of more difficult instances; ii) handling the spatiotemporal dependencies of track annotations across time can be elegantly and efficiently formulated through graphs. Therefore, we use a unified graph formulation to address the annotation of both detections and identity association for tracks across time. Based on these insights, SPAM produces high-quality annotations with a fraction of ground truth labeling cost. We demonstrate that trackers trained on SPAM labels achieve comparable performance to those trained on human annotations while requiring only " + }, + { + "bbox": [ + 160, + 239, + 455, + 469 + ], + "type": "inline_equation", + "content": "3 - 20\\%" + }, + { + "bbox": [ + 160, + 239, + 455, + 469 + ], + "type": "text", + "content": " of the human labeling effort. Hence, SPAM paves the way towards highly efficient labeling of large-scale tracking datasets. We release all models and code." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 479, + 436, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 479, + 436, + 491 + ], + "spans": [ + { + "bbox": [ + 160, + 479, + 436, + 491 + ], + "type": "text", + "content": "Keywords: Multi-object tracking " + }, + { + "bbox": [ + 160, + 479, + 436, + 491 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 479, + 436, + 491 + ], + "type": "text", + "content": " Data labeling " + }, + { + "bbox": [ + 160, + 479, + 436, + 491 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 479, + 436, + 491 + ], + "type": "text", + "content": " Active learning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 527, + 230, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 527, + 230, + 540 + ], + "spans": [ + { + "bbox": [ + 132, + 527, + 230, + 540 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 552, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 552, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 552, + 482, + 647 + ], + "type": "text", + "content": "Detecting and tracking objects in time is at the core of many challenging vision applications. To address these problems, modern multi-object tracking (MOT) approaches [44, 66, 72, 76, 78] rely on increasingly larger amounts of annotated data. While annotating image datasets is already costly, temporal component of videos further increases the annotation difficulty and data scale requirements. This is due to the fact that redundancies between frames cause the information density to scale poorly with the amount of data, making the overall annotation task more challenging and resource-intensive." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 145, + 654, + 335, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 654, + 335, + 665 + ], + "spans": [ + { + "bbox": [ + 145, + 654, + 335, + 665 + ], + "type": "text", + "content": "† Work done during an internship at NVIDIA." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": "To mitigate these issues, researchers are spending more effort in operating with limited human annotations. Self- and semi-supervised learning (SSL) and pseudo-labeling (PL) approaches are often used to leverage unlabeled data whenever possible [34,35,40], while active learning (AL) strategies aim at selecting the most informative samples to label. All these strategies are commonly applied in various image understanding tasks, such as classification [11,14] or object detection [17,67]. Even if video tasks do not have any lower requirement for data, there are only a handful of works that target efficient labeling of tracking data in the video domain [39,59-61]. Existing works fail to provide comprehensive solutions to labeling by either ignoring dense temporal component of videos [39,61] or working under a limited single-object setup [71]. An effective video labeling engine should operate across multiple objects and frames, considering the intricate dependencies between tracks inherent in the labeling problem." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 275, + 482, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 275, + 482, + 419 + ], + "spans": [ + { + "bbox": [ + 130, + 275, + 482, + 419 + ], + "type": "text", + "content": "Videos have unique characteristics that can be leveraged to unleash the full potential of a modern label engine. Firstly, frame-wise similarities in videos yield data redundancies. In the context of MOT, this implies that some associations between detections are relatively easy to solve [7]. This setup provides an ideal opportunity to leverage state-of-the-art models pre-trained to pseudo-label new data at zero cost. Secondly, the dependencies of objects across time and space cause annotations in one frame of the video to have cascading effects in other frames. For example, solving an association for one track can resolve associations for neighboring tracks. In order to maximize the efficiency of the annotations, we frame the labeling task within a model that takes into account such dependencies. Thus, we adopt a track-centric labeling approach rather than a frame-centric one [39]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 422, + 482, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 482, + 602 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 482, + 602 + ], + "type": "text", + "content": "We present the SPAM video label engine, which chains Synthetic pretraining, Pseudo-labeling, and Active learning with a graph-based Model for optimal tracking performance with minimal human effort. In the image domain, there are plenty of large-scale annotated datasets that a model-driven approach can leverage for pre-training [34]. However, this is not the case for multi-object tracking datasets, which are typically much smaller in scale; therefore, we advocate for a Synthetic pre-training. We further analyze which components of a tracking method are more affected by the synthetic-to-real gap. This first experiment gives us an important insight: labeling should be focused on detections and associations, and we can directly use the synthetically trained reID model. With our synthetic pre-trained model, we then propose a model-driven approach to produce Pseudo-labels on real data and retrain the model with its own pseudolabels. We show that with this recipe we can even reach a performance comparable to ground truth without ever training our labeler on manually annotated data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Synthetic pre-training and pseudo-labeling allow us to handle easy decisions and defer only uncertain or complex decisions to an annotator. We propose an Active learning strategy where the labeler makes decisions at the track-level instead of frame-level, thereby utilizing the annotation budget more efficiently. Towards this goal, we solve the labeling problem with a Model, based on graph" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 238, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 238, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 238, + 101 + ], + "type": "text", + "content": "O. Cetintas et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 174 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 174 + ], + "type": "text", + "content": "hierarchies [12]. This model captures long-range spatio-temporal dependencies, which means our annotations will have an impact across frames, and not only on the frame where the annotation actually happens. We extend the model with a detection filtering layer so that we can naturally label detection and association within the same framework." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 480, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 480, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 480, + 283 + ], + "type": "text", + "content": "In summary, we propose SPAM, a video label engine that significantly reduces annotation cost by dealing with easy scenarios with a strong MOT model pre-trained on (i) Synthetic data, followed by (ii) Pseudo-labeling real data for its own re-training. Harder detection and association decisions are carefully selected by an (iii) Active learning scheme that works on a (iv) Model based on graph hierarchies. This allows us to efficiently use the annotator's budget and account for spatio-temporal dependencies in the labels. The resulting model generates labels comparable to ground truth performance, requiring only " + }, + { + "bbox": [ + 130, + 176, + 480, + 283 + ], + "type": "inline_equation", + "content": "3 - 20\\%" + }, + { + "bbox": [ + 130, + 176, + 480, + 283 + ], + "type": "text", + "content": " of manual annotation effort across three diverse datasets." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 303, + 237, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 303, + 237, + 316 + ], + "spans": [ + { + "bbox": [ + 132, + 303, + 237, + 316 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 329, + 479, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 329, + 479, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 329, + 479, + 437 + ], + "type": "text", + "content": "Pre-training for tracking. Over the years, the multi-object tracking (MOT) community has annotated many challenging benchmark datasets [18, 19, 57]. However, due to the significant costs associated with manually labeling temporal data, most benchmarks only provide few annotated video sequences. To obtain top performance, modern data-hungry tracking methods are usually pretrained on additional video and tracking datasets [20, 22, 68, 74, 77]. Many works also rely on relevant image recognition [56] datasets to boost their per-frame object detection performance. An alternative pre-training direction avoids human annotations entirely by relying on simulated synthetic data [23, 24, 26, 29, 58]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 437, + 480, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 480, + 473 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 480, + 473 + ], + "type": "text", + "content": "Our label engine also pre-trains the labeling model synthetically, minimizing its total annotation footprint. To this end, we demonstrate the effectiveness of synthetic pre-training [23] for generating pseudo-labels for self-supervision." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 474, + 480, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 480, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 480, + 593 + ], + "type": "text", + "content": "Self-supervision in the temporal domain. To benefit not only from already annotated datasets, many works explore the utilization of available unlabeled data sources, i.e., how to perform self-supervised learning for videos. Notable examples leverage cycle-consistency-based correspondences [30, 64], color propagation [63] or augmentation-based consistencies [4, 41]. Furthermore, [45] uses contrastive learning over re-identification (reID) embeddings to perform unsupervised association learning. More recently, several works [33-35, 40] in the image domain demonstrate the effectiveness of training on pseudo-labels generated by strong underlying labeling models. In the video domain, this type of self-supervision is mostly unexplored." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 479, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 479, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 479, + 640 + ], + "type": "text", + "content": "We demonstrate the effectiveness of pseudo-labels for the temporal domain by improving the label quality of our SPAM labeler via self-supervision on its own output. The necessary signal-to-noise ratio for this training loop is obtained by pre-training a strong graph-based model on synthetic data." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "content": "Efficient labeling of video datasets. While significant effort has been devoted towards increasing the labeling efficiency of image datasets [1,34], similar" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "text", + "content": "SPAMming Labels: Efficient Annotations for the Trackers of Tomorrow" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "approaches for the video and tracking domain are still relatively limited. Earlier works reduce the number of frames to annotate within individual videos but still rely heavily on human annotators [62]. Other solutions are limited to only a single object per sequence [16, 60] or still rely on human bounding box annotations [59] to label segmentation masks across time. The authors of [39] select key frames based on their expected impact on tracking performance. While effective within its scope, this approach only yields sparse frame annotations, which are insufficient for training all modern trackers. To reduce the human annotation effort for trajectories, [43] proposes an efficient video interface that still requires labeling all trajectories." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 236, + 482, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 482, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 482, + 295 + ], + "type": "text", + "content": "In contrast, our label engine performs active learning for multiple objects across all video frames and is the first of its kind to tackle long-term tracking problems via a hierarchical labeling approach. The uncertainty measures provided by our graph-based model allow us to focus expensive annotation efforts on individual tracks instead of labeling entire frames in a brute-force manner." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 297, + 482, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 482, + 440 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 482, + 440 + ], + "type": "text", + "content": "Multi-object tracking models. Tracking objects across time in a video is an established problem with a long research history. The most dominant tracking-by-detection (TbD) paradigm splits the task into two steps: (i) detecting objects in every frame and (ii) associating them to object trajectories, also referred to as data association. Modern trackers have explored performing (i) and (ii) jointly [7, 25, 44, 78], or providing solutions for (ii) relying on motion models [2, 3, 38, 47, 49, 52, 53, 69, 75], appearance models [13, 15, 36, 46, 51, 54, 70, 76] or more complex techniques involving transformers [10, 72]. An alternative direction of work relies on graph formulations [6, 31, 37, 48, 73] to utilize the TbD paradigm. In this formulation, nodes model object detections, and edges, association hypotheses. This long-studied formulation has recently been revived with remarkable success with the help of learnable graph neural network (GNN)architectures [8, 9, 12, 65]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 440, + 482, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 440, + 482, + 487 + ], + "spans": [ + { + "bbox": [ + 130, + 440, + 482, + 487 + ], + "type": "text", + "content": "For our label engine to generate long-term pseudo-labels and uncertainty estimations, we rely on a strong graph-based tracker [12]. To incorporate both detection and association labeling within a unified graph framework, we introduce a novel hierarchy level dedicated to detection filtering." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 508, + 194, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 508, + 194, + 521 + ], + "spans": [ + { + "bbox": [ + 132, + 508, + 194, + 521 + ], + "type": "text", + "content": "3 SPAM" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 535, + 208, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 535, + 208, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 535, + 208, + 547 + ], + "type": "text", + "content": "3.1 Overview" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 557, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 628 + ], + "type": "text", + "content": "Our main goal is to label multi-object tracking datasets with minimal annotation effort while enabling optimal tracking performance. Our key observation is that most association cases can be solved easily by models pre-trained on synthetic data. Therefore, no human effort is needed, and we can instead use our model's predictions as pseudo-labels. Decisions that are more complex or uncertain are selected by an active learning scheme to be annotated manually." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": "In order to perform pseudo-labeling and active learning, we require a model that can capture long-term spatio-temporal dependencies between tracks. We leverage a hierarchical graph neural network formulation and cast the annotation" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 238, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 238, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 238, + 101 + ], + "type": "text", + "content": "O. Cetintas et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 118, + 482, + 196 + ], + "blocks": [ + { + "bbox": [ + 136, + 118, + 482, + 196 + ], + "lines": [ + { + "bbox": [ + 136, + 118, + 482, + 196 + ], + "spans": [ + { + "bbox": [ + 136, + 118, + 482, + 196 + ], + "type": "image", + "image_path": "e885ba6615bc01047af4fb79ade4fdf2af243db6d58bab58dd2e78978747dc2f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 207, + 482, + 242 + ], + "lines": [ + { + "bbox": [ + 130, + 207, + 482, + 242 + ], + "spans": [ + { + "bbox": [ + 130, + 207, + 482, + 242 + ], + "type": "text", + "content": "Fig. 1: Overview of the SPAM model. We first generate a set of detection candidates with our detector. Hierarchical GNNs then classify these candidates into valid and invalid objects via node classification, and assign identities through edge classification." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 262, + 479, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 262, + 479, + 345 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 479, + 345 + ], + "type": "text", + "content": "task over a graph. Our architecture (Fig. 1) draws inspiration from SUSHI [12], which demonstrated remarkable association performance. Our model's scope, however, goes beyond association. Given the constraints of the labeling task, we need to equip our model with the ability to reason about detections before performing association. To address this broader goal, our GNN hierarchy, also includes the capability to accept or reject detections by classifying nodes into correct and incorrect object hypotheses." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 346, + 482, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 482, + 465 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 482, + 465 + ], + "type": "text", + "content": "Our overall model is pre-trained on synthetic data [23] and further fine-tuned on its own predicted pseudo-labels over the target dataset that we aim to annotate. This recipe allows for strong labelling performance without any human annotations, and provides the basis for a scalable video label engine. After this training loop, we use our updated model's predictions as pseudo-labels to annotate a large portion of the target dataset. Decisions that are more complex or uncertain are selected by an active learning scheme to be annotated manually. The resulting set of labels achieves near-GT performance with significantly fewer manual annotations, thereby vastly enhancing the efficiency of the overall process. We provide an overview of our training and annotating recipe in Fig. 2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 480, + 378, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 480, + 378, + 493 + ], + "spans": [ + { + "bbox": [ + 131, + 480, + 378, + 493 + ], + "type": "text", + "content": "3.2 Graph Hierarchies as a Model for Labeling" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 498, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 641 + ], + "type": "text", + "content": "Problem formulation. Our model builds upon the standard graph formulation [73] where, given a set of object candidates " + }, + { + "bbox": [ + 130, + 498, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 130, + 498, + 482, + 641 + ], + "type": "text", + "content": ", our goal is to obtain a set of valid objects " + }, + { + "bbox": [ + 130, + 498, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_v \\subset \\mathcal{O}" + }, + { + "bbox": [ + 130, + 498, + 482, + 641 + ], + "type": "text", + "content": " and their corresponding trajectories " + }, + { + "bbox": [ + 130, + 498, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 130, + 498, + 482, + 641 + ], + "type": "text", + "content": ". Each trajectory " + }, + { + "bbox": [ + 130, + 498, + 482, + 641 + ], + "type": "inline_equation", + "content": "T_k \\in \\mathcal{T}" + }, + { + "bbox": [ + 130, + 498, + 482, + 641 + ], + "type": "text", + "content": " consists of a set of objects sharing the same identity or, equivalently, the set of edges connecting those nodes. Specifically, we model each object candidate " + }, + { + "bbox": [ + 130, + 498, + 482, + 641 + ], + "type": "inline_equation", + "content": "o_i \\in \\mathcal{O}" + }, + { + "bbox": [ + 130, + 498, + 482, + 641 + ], + "type": "text", + "content": " as a node and each association hypothesis among candidates as an edge in an undirected graph " + }, + { + "bbox": [ + 130, + 498, + 482, + 641 + ], + "type": "inline_equation", + "content": "G = (V, E)" + }, + { + "bbox": [ + 130, + 498, + 482, + 641 + ], + "type": "text", + "content": ". Within this formulation, nodes " + }, + { + "bbox": [ + 130, + 498, + 482, + 641 + ], + "type": "inline_equation", + "content": "u \\in V" + }, + { + "bbox": [ + 130, + 498, + 482, + 641 + ], + "type": "text", + "content": " could be classified into valid objects if " + }, + { + "bbox": [ + 130, + 498, + 482, + 641 + ], + "type": "inline_equation", + "content": "u \\in \\mathcal{O}_v" + }, + { + "bbox": [ + 130, + 498, + 482, + 641 + ], + "type": "text", + "content": ". Analogously, edges can be classified into correct or incorrect hypotheses depending on whether they belong to some trajectory or not. This formulation can be generalized to a hierarchical setting by splitting video sequences into non-overlapping subsequences, i.e., disjoint subgraphs, and progressively merging them [12]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Background on graph hierarchies for MOT. Our model consists of a hierarchy of GNNs operating over the graph formulation described above. Intuitively," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 157, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 157, + 91, + 448, + 102 + ], + "type": "text", + "content": "SPAMming Labels: Efficient Annotations for the Trackers of Tomorrow" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 478, + 186 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 478, + 186 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 478, + 186 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 478, + 186 + ], + "type": "image", + "image_path": "beb0774fcc317056f4f2f0a806b15ef73c5ecce5e7f19eff5c52537c02fa6a01.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 193, + 482, + 248 + ], + "lines": [ + { + "bbox": [ + 130, + 193, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 193, + 482, + 248 + ], + "type": "text", + "content": "Fig. 2: Overview of the SPAM training and annotation pipeline. (a) Initial model training on synthetic data. (b) Application of SPAM to generate pseudo-labels without incurring manual annotation costs on a real dataset, followed by self-training on pseudolabels. (c) Real dataset labeling using pseudo-labels and an uncertainty-based active learning approach." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 283, + 482, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 283, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 482, + 403 + ], + "type": "text", + "content": "starting from detection candidates, at each hierarchy level, GNN hierarchy subsequently merge tracklets from the previous level into longer ones. Nodes and edges are represented by initial embeddings and message passing is used to share information within the graph in order to update node and edge embeddings with richer information. The GNNs are trained to classify edges into active and inactive association hypotheses and they share the same features, network architecture and learnable weights across hierarchy levels. In addition to the edge features reported in [12], for each object candidate " + }, + { + "bbox": [ + 130, + 283, + 482, + 403 + ], + "type": "inline_equation", + "content": "o_i" + }, + { + "bbox": [ + 130, + 283, + 482, + 403 + ], + "type": "text", + "content": " we concatenate its times-tamp " + }, + { + "bbox": [ + 130, + 283, + 482, + 403 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 130, + 283, + 482, + 403 + ], + "type": "text", + "content": ", normalized bounding box coordinates and dimensions " + }, + { + "bbox": [ + 130, + 283, + 482, + 403 + ], + "type": "inline_equation", + "content": "(x_i, y_i, h_i, w_i)" + }, + { + "bbox": [ + 130, + 283, + 482, + 403 + ], + "type": "text", + "content": " and corresponding confidence from the detector " + }, + { + "bbox": [ + 130, + 283, + 482, + 403 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 130, + 283, + 482, + 403 + ], + "type": "text", + "content": " and embed it as node features." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 408, + 482, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 408, + 482, + 588 + ], + "spans": [ + { + "bbox": [ + 130, + 408, + 482, + 588 + ], + "type": "text", + "content": "Classifying nodes. We aim to minimize the annotation effort; therefore, we ask the annotator to operate on a fixed graph of detection candidates, i.e. annotator does not label new detections after the initial set given by detector. This is why we aim to create an initial graph that is over-complete. In order to do so, we use a low confidence threshold for our detector and obtain a large number of detection candidates. While this ensures a high detection recall, it also introduces a larger number of false positive detections that our model needs to address. As we show in experiments, the original SUSHI architecture [12] cannot handle such false positives, which is why we propose a new layer in the hierarchy, " + }, + { + "bbox": [ + 130, + 408, + 482, + 588 + ], + "type": "inline_equation", + "content": "GNN_{node}" + }, + { + "bbox": [ + 130, + 408, + 482, + 588 + ], + "type": "text", + "content": ". Intuitively, " + }, + { + "bbox": [ + 130, + 408, + 482, + 588 + ], + "type": "inline_equation", + "content": "GNN_{node}" + }, + { + "bbox": [ + 130, + 408, + 482, + 588 + ], + "type": "text", + "content": " leverages temporal information to validate object candidates. We use its output node embeddings and feed them to a lightweight classifier to enable false positive detection filtering. By doing so, we can consider a large number of input proposals achieving high detection recall by leveraging the spatio-temporal reasoning ability of our GNN for detection candidates." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "content": "Pre-training and annotation-free fine-tuning. We pre-train all of our model components on the large-scale synthetic dataset MOTSynth [23]. To boost our model's performance and bridge the domain gap with our target dataset for annotation, we conduct further fine-tuning on the target dataset after pre-training. We achieve this by utilizing the raw predictions of our model on the target dataset and employing them as pseudo-labels for self-training. As we show in" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 238, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 238, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 238, + 101 + ], + "type": "text", + "content": "O. Cetintas et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 164, + 114, + 455, + 219 + ], + "blocks": [ + { + "bbox": [ + 164, + 114, + 455, + 219 + ], + "lines": [ + { + "bbox": [ + 164, + 114, + 455, + 219 + ], + "spans": [ + { + "bbox": [ + 164, + 114, + 455, + 219 + ], + "type": "image", + "image_path": "9bf0b4315066d528c3b5d6ba89673473ac6f5be91aa56f407adc960c1d179393.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 236, + 480, + 270 + ], + "lines": [ + { + "bbox": [ + 131, + 236, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 131, + 236, + 480, + 270 + ], + "type": "text", + "content": "Fig. 3: Our graph-based labeling pipeline begins with the selection of nodes for annotation. For each node to be annotated, the annotator could be asked to validate the detection, improve the localization by refining the box or perform association." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 295, + 479, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 295, + 479, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 295, + 479, + 319 + ], + "type": "text", + "content": "our experimental section, this strategy yields a significant performance improvement without incurring any human annotation cost." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 341, + 346, + 353 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 341, + 346, + 353 + ], + "spans": [ + { + "bbox": [ + 131, + 341, + 346, + 353 + ], + "type": "text", + "content": "3.3 How to Label on Graph Hierarchies?" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 365, + 482, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 496 + ], + "type": "text", + "content": "Manual labeling. Labeling video data at the box level involves several steps: identifying objects of interest, localizing their boundaries with bounding boxes, and assigning them unique identities. Typically, annotators achieve this by examining each frame, detecting and localizing objects individually, and then assigning identities based on temporal context from neighboring frames. In terms of annotation effort in clicks, detecting and localizing objects typically require two clicks (one for the top-left corner and one for the bottom-right corner of the bounding box), while associating objects across time requires one click (identifying the next occurrence of the object in subsequent frames), as reported in [18, 19]. This manual process is labor-intensive, costly, and severely limits the availability of annotated video data." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "type": "text", + "content": "Graph-based labeling. Utilizing a graph formulation for detection and tracking enables explicit modeling of objects in the scene and their interactions. In this framework, each node of the graph represents an object, allowing a manual annotation of an object to be framed as selecting a node on this graph. Given a set of nodes to be annotated, possible annotator decisions are: (i) accepting/rejecting a detection as we are working on a large set of detection candidates (one click), (ii) refining the bounding-box to improve localization (two clicks, one for the top-left corner and one for the bottom-right corner of the bounding box) and (iii) associating a node across time (one click) as shown in Fig. 3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "content": "Annotation acquisition. The core idea of our approach is to use our model decisions to guide the annotation process. For the vast majority of scenarios, we directly use our model's predictions as pseudo-labels. For a small subset of more challenging decisions where our model exhibits uncertainty, we validate our decisions with an annotator as described above. Within our graphs," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "text", + "content": "SPAMming Labels: Efficient Annotations for the Trackers of Tomorrow" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 188, + 120, + 427, + 264 + ], + "blocks": [ + { + "bbox": [ + 188, + 120, + 427, + 264 + ], + "lines": [ + { + "bbox": [ + 188, + 120, + 427, + 264 + ], + "spans": [ + { + "bbox": [ + 188, + 120, + 427, + 264 + ], + "type": "image", + "image_path": "c1ef6f6f66be5c046820990ba3d4e32bd29615bb445768664948ba57b75b8fbf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 278, + 482, + 312 + ], + "lines": [ + { + "bbox": [ + 131, + 278, + 482, + 312 + ], + "spans": [ + { + "bbox": [ + 131, + 278, + 482, + 312 + ], + "type": "text", + "content": "Fig. 4: Analysis of performance gap between training a model on synthetic and real data for the three most common tracking components: detection, association, re-identification." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 328, + 482, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 328, + 482, + 378 + ], + "spans": [ + { + "bbox": [ + 130, + 328, + 482, + 378 + ], + "type": "text", + "content": "we adopt an uncertainty-based approach to identify which nodes require annotations. Specifically, for each node " + }, + { + "bbox": [ + 130, + 328, + 482, + 378 + ], + "type": "inline_equation", + "content": "v \\in V" + }, + { + "bbox": [ + 130, + 328, + 482, + 378 + ], + "type": "text", + "content": ", we compute its uncertainty as " + }, + { + "bbox": [ + 130, + 328, + 482, + 378 + ], + "type": "inline_equation", + "content": "\\mathrm{uncert}(v) \\coloneqq \\max_{u \\in \\mathbb{N}_v} H(\\hat{y}_{(v,u)})" + }, + { + "bbox": [ + 130, + 328, + 482, + 378 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 328, + 482, + 378 + ], + "type": "inline_equation", + "content": "\\mathbb{N}_v" + }, + { + "bbox": [ + 130, + 328, + 482, + 378 + ], + "type": "text", + "content": " denotes the set of neighboring nodes of " + }, + { + "bbox": [ + 130, + 328, + 482, + 378 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 130, + 328, + 482, + 378 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 328, + 482, + 378 + ], + "type": "inline_equation", + "content": "\\hat{y}_{(v,u)}" + }, + { + "bbox": [ + 130, + 328, + 482, + 378 + ], + "type": "text", + "content": " is the edge prediction for edge " + }, + { + "bbox": [ + 130, + 328, + 482, + 378 + ], + "type": "inline_equation", + "content": "(v,u)" + }, + { + "bbox": [ + 130, + 328, + 482, + 378 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 177, + 391, + 481, + 406 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 177, + 391, + 481, + 406 + ], + "spans": [ + { + "bbox": [ + 177, + 391, + 481, + 406 + ], + "type": "interline_equation", + "content": "H (\\hat {y} _ {(v, u)}) := - (\\hat {y} _ {(v, u)} \\log \\hat {y} _ {(v, u)} + (1 - \\hat {y} _ {(v, u)}) \\log (1 - \\hat {y} _ {(v, u)})) \\qquad (1)", + "image_path": "b975c77644a69b50e8c74a89446360d8f850fba2de1d3955335d7d539b2a98c6.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 411, + 482, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 482, + 470 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 482, + 470 + ], + "type": "text", + "content": "At the " + }, + { + "bbox": [ + 130, + 411, + 482, + 470 + ], + "type": "inline_equation", + "content": "GNN_{node}" + }, + { + "bbox": [ + 130, + 411, + 482, + 470 + ], + "type": "text", + "content": " level, where we have node predictions instead of edge predictions, we simply compute " + }, + { + "bbox": [ + 130, + 411, + 482, + 470 + ], + "type": "inline_equation", + "content": "\\text{uncert}(v)" + }, + { + "bbox": [ + 130, + 411, + 482, + 470 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 411, + 482, + 470 + ], + "type": "inline_equation", + "content": "H(\\hat{y}_{(v)})" + }, + { + "bbox": [ + 130, + 411, + 482, + 470 + ], + "type": "text", + "content": " instead. Subsequently, we refer the nodes with the highest uncertainty scores to the annotator for manual annotation and use the remaining predictions as pseudo-labels for the rest of the nodes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 472, + 482, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 579 + ], + "type": "text", + "content": "Hierarchical label generation. To fully leverage our hierarchical graph formulation, we apply our label acquisition approach described above in a hierarchical manner. Intuitively, we distribute our annotation budget " + }, + { + "bbox": [ + 130, + 472, + 482, + 579 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 472, + 482, + 579 + ], + "type": "text", + "content": " across " + }, + { + "bbox": [ + 130, + 472, + 482, + 579 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 472, + 482, + 579 + ], + "type": "text", + "content": " hierarchical levels as " + }, + { + "bbox": [ + 130, + 472, + 482, + 579 + ], + "type": "inline_equation", + "content": "B_{1},\\ldots ,B_{L}" + }, + { + "bbox": [ + 130, + 472, + 482, + 579 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 130, + 472, + 482, + 579 + ], + "type": "inline_equation", + "content": "B_{1} + \\dots +B_{L} = B" + }, + { + "bbox": [ + 130, + 472, + 482, + 579 + ], + "type": "text", + "content": ". In deeper hierarchy levels, nodes represent tracklets instead of single detections. This allows us to propagate annotator decisions for the entire cluster instead of individual detections, resulting in solving multiple uncertainties with a single annotation. As a result, we utilize the annotation budget more effectively while also mitigating error propagation within our pseudo-labels." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 601, + 231, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 601, + 231, + 615 + ], + "spans": [ + { + "bbox": [ + 131, + 601, + 231, + 615 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 630, + 269, + 641 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 630, + 269, + 641 + ], + "spans": [ + { + "bbox": [ + 131, + 630, + 269, + 641 + ], + "type": "text", + "content": "4.1 Datasets and Metrics" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 653, + 419, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 653, + 419, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 653, + 419, + 666 + ], + "type": "text", + "content": "Datasets. We conduct experiments on three public benchmarks:" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 238, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 238, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 238, + 101 + ], + "type": "text", + "content": "O. Cetintas et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 137, + 115, + 481, + 260 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 137, + 115, + 481, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 115, + 481, + 151 + ], + "spans": [ + { + "bbox": [ + 137, + 115, + 481, + 151 + ], + "type": "text", + "content": "- MOT17 [18] is an established benchmark containing 14 challenging video sequences, exhibiting diverse characteristics, including variations in camera motion, viewpoint, and pedestrian densities." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 152, + 481, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 152, + 481, + 199 + ], + "spans": [ + { + "bbox": [ + 138, + 152, + 481, + 199 + ], + "type": "text", + "content": "- MOT20 [19] contains a total of 8 video sequences, all featuring extremely crowded scenes. Due to their high pedestrian density, the emphasis is on addressing highly challenging scenarios characterized by frequent and prolonged occlusions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 200, + 481, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 200, + 481, + 260 + ], + "spans": [ + { + "bbox": [ + 138, + 200, + 481, + 260 + ], + "type": "text", + "content": "- DanceTrack [57] contains 100 video sequences. These sequences feature dancing individuals with similar appearances, facilitating the evaluation of tracker robustness against highly challenging motion patterns where appearance information has a relatively minor influence. This non-standard setup often leads to a performance drop in trackers [57]." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 269, + 479, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 269, + 479, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 269, + 479, + 281 + ], + "type": "text", + "content": "Metrics. We assess the performance of trackers using three established metrics:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 291, + 481, + 374 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 137, + 291, + 481, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 291, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 137, + 291, + 481, + 326 + ], + "type": "text", + "content": "- MOTA [32] assesses trajectory coverage, by combining object detection measurements with identity switches, with the latter playing a relatively minor role." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 137, + 327, + 481, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 327, + 481, + 350 + ], + "spans": [ + { + "bbox": [ + 137, + 327, + 481, + 350 + ], + "type": "text", + "content": "- IDF1 [50] focuses on identity preservation over the sequences and provides a measure of association performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 137, + 350, + 481, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 350, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 137, + 350, + 481, + 374 + ], + "type": "text", + "content": "- HOTA [42] is a recently introduced metric designed to strike a balance between association and object detection performance within a single score." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 392, + 279, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 279, + 403 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 279, + 403 + ], + "type": "text", + "content": "4.2 Implementation Details" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 412, + 481, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 412, + 481, + 495 + ], + "spans": [ + { + "bbox": [ + 132, + 412, + 481, + 495 + ], + "type": "text", + "content": "Architecture and training. Our model consists of three main components: (i) a detector network, based onYOLOX [27], (ii) a ResNet-based ReID network [28], and (iii) a GNN hierarchy, all pretrained on the synthetic sequences of MOTSynth [23]. After pretraining, we use the resulting model to obtain pseudolabels for our target MOT dataset, and then fine-tune our detector and GNN on them. Note that our Model is never trained on human annotations, it is only trained on synthetic annotations and its own pseudo-labels." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 496, + 481, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 496, + 481, + 591 + ], + "spans": [ + { + "bbox": [ + 132, + 496, + 481, + 591 + ], + "type": "text", + "content": "Experimental setup. Online trackers are real-time tools suitable for deployment on robots or autonomous vehicles, whereas offline trackers are better suited for labeling tasks. Therefore, we structure our experiments to utilize SPAM for labeling and use these labels with online trackers. To evaluate the quality of SPAM labels, we train two state-of-the-art online trackers with these labels: ByteTrack [75] and GHOST [54]. For validation, we follow the standard half-split setting [54, 75, 76] over MOT17 sequences. Annotation budgets are computed based on number of clicks as described in Sec. 3.3." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 609, + 265, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 265, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 265, + 621 + ], + "type": "text", + "content": "4.3 Synthetic Pretaining" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 629, + 481, + 665 + ], + "type": "text", + "content": "Several works have shown that synthetic pre-training is a strong starting point for multi-object tracking methods [23, 75]. In this section, we aim to analyze the gap between training a model on synthetic data and training it with real" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "text", + "content": "SPAMming Labels: Efficient Annotations for the Trackers of Tomorrow" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 114, + 482, + 170 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 482, + 170 + ], + "type": "table", + "html": "
ModelHigh Conf. BoxLow Conf. Box\\( GNN_{node} \\)MOTA ↑IDF1 ↑HOTA ↑DetA ↑
A✓✘✘64.474.759.955.9
B✓✓✘60.671.458.554.9
C✓✓✓65.475.160.456.6
", + "image_path": "3da6dd36b8f91b5cfe7483fb121abd6586fc083974d91b53ae2c76394d104123.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 172, + 480, + 194 + ], + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 194 + ], + "type": "text", + "content": "Table 1: Ablation study investigating the impact of incorporating low confidence bounding boxes with and without the " + }, + { + "bbox": [ + 132, + 172, + 480, + 194 + ], + "type": "inline_equation", + "content": "GNN_{node}" + }, + { + "bbox": [ + 132, + 172, + 480, + 194 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 209, + 482, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 209, + 482, + 364 + ], + "spans": [ + { + "bbox": [ + 130, + 209, + 482, + 364 + ], + "type": "text", + "content": "data for the three most common tracking components: detection, association, re-identification. For each setting, only the specific component that we are analysing is trained on synthetic data while other components are all trained on real data. We perform this experiment on three datasets and plot the relative HOTA difference (a positive score indicates superior performance with real data training) when training on real data vs. synthetic data in Fig. 4. Our first observation, is that the detector is the component that suffers the domain gap the most, with a performance drop of 9.9 HOTA points. Re-identification, on the other hand, has a similar performance when trained on synthetic or real data. Lastly, association performance is 2.1 HOTA higher with real data training. This first experiment gives us an important insight: labeling and retraining should be focused on detection and association components, and we can directly use the synthetically trained reID model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 382, + 313, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 382, + 313, + 395 + ], + "spans": [ + { + "bbox": [ + 132, + 382, + 313, + 395 + ], + "type": "text", + "content": "4.4 Towards a Model for Labeling" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 402, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 628 + ], + "type": "text", + "content": "Graph-based validation of detection candidates. In our framework, annotators have the capability to eliminate detections, but we refrain from requesting them to annotate new detections in order to maintain minimal annotation effort. Instead we start from an initial graph formed by an over-complete set of detections. Towards this end, we use a low confidence threshold for our detector, thereby reducing the number of false negatives in our initial graph and ensuring a high recall. As a result, this leads to an increase in the number of false positives incorporated into the graph, a tendency that is further accentuated by the fact that our detector is exclusively trained on synthetic data. We propose to leverage the graph structure itself to validate the detections based on temporal consistency. Specifically, we add an additional GNN-layer in our hierarchy, " + }, + { + "bbox": [ + 130, + 402, + 482, + 628 + ], + "type": "inline_equation", + "content": "GNN_{node}" + }, + { + "bbox": [ + 130, + 402, + 482, + 628 + ], + "type": "text", + "content": ", which intuitively examines the spatial and temporal consistency of the bounding box candidates and filters out implausible ones. In Tab. 1, we assess the effect of this design choice. Model A uses only high confidence bounding boxes, model B creates a graph with also low confidence bounding boxes but without " + }, + { + "bbox": [ + 130, + 402, + 482, + 628 + ], + "type": "inline_equation", + "content": "GNN_{node}" + }, + { + "bbox": [ + 130, + 402, + 482, + 628 + ], + "type": "text", + "content": ", while C includes " + }, + { + "bbox": [ + 130, + 402, + 482, + 628 + ], + "type": "inline_equation", + "content": "GNN_{node}" + }, + { + "bbox": [ + 130, + 402, + 482, + 628 + ], + "type": "text", + "content": ". As we can see, simply adding low confidence boxes results in a drastic drop in performance on all metrics. Only by adding our " + }, + { + "bbox": [ + 130, + 402, + 482, + 628 + ], + "type": "inline_equation", + "content": "GNN_{node}" + }, + { + "bbox": [ + 130, + 402, + 482, + 628 + ], + "type": "text", + "content": " we are able to take full advantage of the low confidence boxes without penalizing our performance with too many errors." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "content": "SPAM is a SOTA tracker. Before moving to labeling with our model, we want to verify its overall performance and generalization capabilities. Only by using a state-of-the-art model we will be able to reduce annotation effort as much as" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 238, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 238, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 238, + 101 + ], + "type": "text", + "content": "O. Cetintas et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 114, + 482, + 249 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 482, + 249 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 482, + 249 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 482, + 249 + ], + "type": "table", + "html": "
MethodHOTA ↑MOTA ↑IDF1 ↑HOTA ↑MOTA ↑IDF1 ↑HOTA ↑MOTA ↑IDF1 ↑
MOT17MOT20DanceTrack
MeMOT [10]56.972.569.054.166.163.7---
TrackFormer [44]57.374.168.054.765.768.6---
MOTR [72]57.868.673.4---54.251.579.7
FairMOT [76]59.373.772.354.661.867.339.782.240.8
QDTrack [46]63.578.777.560.074.773.854.287.750.4
ByteTrack† [75]62.878.977.160.474.274.547.789.653.9
GHOST [54]62.878.777.161.273.775.256.791.357.7
UTM [70]64.081.878.762.578.276.9---
MotionTrack [49]65.181.180.162.878.076.5---
SUSHI [12]66.581.183.164.374.379.863.388.763.4
SPAM (Ours)67.580.784.665.876.581.964.089.263.4
", + "image_path": "bfbe9dbf7a7db592aa3db48bf834917b5d1922ebd5821a2d8690a0174b4c986a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 250, + 479, + 271 + ], + "lines": [ + { + "bbox": [ + 132, + 250, + 479, + 271 + ], + "spans": [ + { + "bbox": [ + 132, + 250, + 479, + 271 + ], + "type": "text", + "content": "Table 2: Test set results on MOT17, MOT20 and DanceTrack when using SPAM solely as a tracker." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 282, + 482, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 282, + 482, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 282, + 482, + 462 + ], + "type": "text", + "content": "possible. We therefore pre-train our model on the synthetic dataset, and finetune it on the specific real dataset using groundtruth. For a fair comparison, we obtain object detections from a YOLOX detector [27] trained following [12, 54, 75]. Our model outperforms state-of-the-art by 1.0 HOTA point on MOT17, 1.5 HOTA points on MOT20, and 0.7 HOTA points on DanceTrack as shown in Tab. 2. Furthermore, on MOT17 and MOT20, there is also a significant improvement in IDF1, showing the identity-preserving capabilities of the model, which is especially important for our video annotation setting. Even if being state-of-the-art is not the main goal of this paper, having a strong performing model allows us to have an excellent starting point for producing pseudo-labels. We would like to emphasize that, we use ground truth annotations only for this experiment just to compare our model fairly with other trackers. From this point onward, we will not use ground truth labels. We will use our synthetic pretraining weights and focus on the pseudo-labeling and active learning aspects of our method to show how to annotate multi-object tracking datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 479, + 481, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 491 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 491 + ], + "type": "text", + "content": "4.5 Enhancing Annotations with Pseudo-labels and Active Learning" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "type": "text", + "content": "Retraining our Model on Pseudo-labels. In this experiment, we compare two models. The first one, SPAM, is only trained on synthetic data. The second one, SPAM, is trained in three stages: (i) pre-training on synthetic data, (ii) pseudo-labeling the real dataset with our model, and (iii) re-training our model on pseudo-labels. In Sec. 4.5, we observe a significant improvement when using pseudo-labels to re-train our model: an increase of 4-6 HOTA, 4-6 MOTA, and 4-7 IDF1 points on MOT17, MOT20, and DanceTrack, respectively. Note that these improvements come at zero human annotation effort as this process requires no manual annotations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "content": "Active learning on graph hierarchies. We now test the effectiveness of our proposed active learning labeling on graph hierarchies, and show a comprehensive comparison with several baselines in Fig. 5. We directly measure the accuracy of the output labels given an increasing percentage of manual annotations selected" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "text", + "content": "SPAMming Labels: Efficient Annotations for the Trackers of Tomorrow" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 114, + 286, + 223 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 286, + 223 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 286, + 223 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 286, + 223 + ], + "type": "table", + "html": "
ModelPseudo-labelHOTA ↑MOTA ↑IDF1 ↑
MOT17
SPAM✘60.065.375.1
SPAM✓63.869.279.7
MOT20
SPAM✘52.265.470.3
SPAM✓58.770.277.1
DanceTrack
SPAM✘41.865.243.4
SPAM✓48.171.050.2
", + "image_path": "968593e5c39ca9296768bd4e3599935b963741a9d00a241671d03e4427ccbfae.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 232, + 287, + 276 + ], + "lines": [ + { + "bbox": [ + 132, + 232, + 287, + 276 + ], + "spans": [ + { + "bbox": [ + 132, + 232, + 287, + 276 + ], + "type": "text", + "content": "Table 3: Performance boost obtained by our model when retraining with its own pseudo-labels incurring no manual annotation cost." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 306, + 114, + 491, + 240 + ], + "blocks": [ + { + "bbox": [ + 306, + 114, + 491, + 240 + ], + "lines": [ + { + "bbox": [ + 306, + 114, + 491, + 240 + ], + "spans": [ + { + "bbox": [ + 306, + 114, + 491, + 240 + ], + "type": "image", + "image_path": "986c70849e4c91b4e5ba3eb90f0241671072a3458beb8f7a0ad06cb95ea25985.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 298, + 240, + 482, + 263 + ], + "lines": [ + { + "bbox": [ + 298, + 240, + 482, + 263 + ], + "spans": [ + { + "bbox": [ + 298, + 240, + 482, + 263 + ], + "type": "text", + "content": "Fig. 5: Impact of various active learning strategies on the quality of labeling." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 297, + 482, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 482, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 482, + 429 + ], + "type": "text", + "content": "by different acquisition functions on MOT17. The vertical axis shows the increase in HOTA over the baseline performance, i.e., the performance without active learning. Oracle is our upper bound and shows the performance if all possible labels are set to their correct value by the annotator. Every method has two different labeling schemes: one for detections (nodes), and one for associations (edges) as explained in Sec. 3.3. In addition to our method, we explore active learning strategies commonly used in object detection, including Coreset [55], Ensemble [5], and Consistency over augmented inputs [21]. The naming is based on the active learning strategy for detection in the first part and the strategy for association in the second part. For instance, EntImg/Rand. indicates image level entropy-based selection for detection and random selection for association." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 430, + 482, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 482, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 482, + 609 + ], + "type": "text", + "content": "The methods shown in red shades and circle markers compare three detection labeling methods while keeping association label selection random. Box-level labeling is far superior to commonly-used image-level labeling, or random labeling. This highlights the potential of performing active learning at the object level rather than the frame level. The methods in blue shades and triangle markers all use our proposed active learning selection on graph hierarchies for the association, and compare several detection selection schemes. The Coreset [55], which prioritizes diversity of the data over uncertainty, emerges as the poorest performing approach. This observation underlines that, when accurate predictions are available, an uncertainty-based approach proves to be a more effective complement to pseudo-labels. Box-level entropy, ensembles, and consistency all perform similarly. Finally, our method, with both detection and association label selection on graph hierarchies is superior to all compared baselines, staying at only 1.2 points from Oracle performance. Notably, we also unify the AL strategy for detection and association with our approach." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 624, + 434, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 624, + 434, + 636 + ], + "spans": [ + { + "bbox": [ + 132, + 624, + 434, + 636 + ], + "type": "text", + "content": "4.6 Putting Everything Together: SPAM Labels in Action" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Training trackers with our labels. In this experiment, we ablate the effects of our SPAM pipeline on the actual usecase of training a model with our pro" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 238, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 238, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 238, + 100 + ], + "type": "text", + "content": "O. Cetintas et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 114, + 482, + 177 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 482, + 177 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 482, + 177 + ], + "type": "table", + "html": "
ModelLabelsPseudo-labelsActive Learn.Ann. Effort ↓HOTA ↑MOTA ↑IDF1 ↑
ByteTrack [75]Ground Truth-100%52.660.465.7
ByteTrack [75]SPAMXXX49.657.863.1
ByteTrack [75]SPAM✓XX50.559.564.5
ByteTrack [75]SPAM✓✓3.3%52.561.866.2
", + "image_path": "2ab2fc260a6ec865261ca25548f8a26cce2549b0fcbdd0a8ff0d037a5249f2af.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 178, + 480, + 210 + ], + "lines": [ + { + "bbox": [ + 132, + 178, + 480, + 210 + ], + "spans": [ + { + "bbox": [ + 132, + 178, + 480, + 210 + ], + "type": "text", + "content": "Table 4: Effect of (i) Synthetic training, (ii) Retraining our model with our Pseudolabels, and (iii) Active learning-based acquisition and annotation of a small percentage of real data, on our label quality on the MOT17 validation set." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 240, + 482, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 240, + 482, + 410 + ], + "spans": [ + { + "bbox": [ + 130, + 240, + 482, + 410 + ], + "type": "text", + "content": "vided labels. We ablate label quality under the following settings: (i) Synthetic training, (ii) Retraining our models on our Pseudo-labels, (iii) Active learning based acquisition and annotation of a small percentage of real data. We train state-of-the-art online tracker ByteTrack [75] from scratch with the aforementioned sets of labels and report its performance on the MOT17 validation set in Tab. 4. The goal performance we want to reach is the one achieved by the model trained on groundtruth annotations (row 1). Row 2 shows our model trained only on synthetic data, which is 3 HOTA points below our goal. Using pseudo-label based training of our model increases performance by almost a HOTA point, and the ground truth level performance is achieved by also using our active learning method to select which samples to label. By annotating only " + }, + { + "bbox": [ + 130, + 240, + 482, + 410 + ], + "type": "inline_equation", + "content": "3.3\\%" + }, + { + "bbox": [ + 130, + 240, + 482, + 410 + ], + "type": "text", + "content": " of the data, we can achieve the same results as the model trained on the ground truth data with " + }, + { + "bbox": [ + 130, + 240, + 482, + 410 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 130, + 240, + 482, + 410 + ], + "type": "text", + "content": " annotation effort. This shows the potential of our SPAM in reducing annotation effort for multi-object tracking datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 414, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 482, + 668 + ], + "type": "text", + "content": "How much data do we really need to label? In this final experiment, our goal is to assess the impact of our entire pipeline by answering the question: How much data do we actually need to label in order to achieve the same performance as a model trained on the entire groundtruth training dataset? For that, we choose two well-established trackers, ByteTrack [75] and GHOST [54], and train them from scratch using our SPAM labels. We report their performance on MOT17, MOT20, and DanceTrack test sets in Tab. 5. We first confirm that our results on the validation set also generalize to the test set, and that we can reach GT performance with " + }, + { + "bbox": [ + 130, + 414, + 482, + 668 + ], + "type": "inline_equation", + "content": "3.3\\%" + }, + { + "bbox": [ + 130, + 414, + 482, + 668 + ], + "type": "text", + "content": " of annotation effort on MOT17. Note that for MOT17 our annotations yield a slight improvement over GT, which can be attributed to observed annotation noise in the original GT, particularly in moving sequences (further details can be found in the supplementary material). MOT20 and DanceTrack are known to be more challenging datasets [57], which is confirmed by the fact that the trackers still miss a few points when trained on only " + }, + { + "bbox": [ + 130, + 414, + 482, + 668 + ], + "type": "inline_equation", + "content": "3.3\\%" + }, + { + "bbox": [ + 130, + 414, + 482, + 668 + ], + "type": "text", + "content": " of manually annotated data. Given that we cannot do a full analysis of different percentages because we are working on the test sets, we choose two percentages of manually annotated data on which to train our trackers: " + }, + { + "bbox": [ + 130, + 414, + 482, + 668 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 414, + 482, + 668 + ], + "type": "text", + "content": " for MOT20, and " + }, + { + "bbox": [ + 130, + 414, + 482, + 668 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 414, + 482, + 668 + ], + "type": "text", + "content": " for DanceTrack. In both cases, we achieve on-par performance with the trackers trained on full groundtruth data. This clearly shows the potential in properly leveraging synthetic data, pseudo-labels from a strong model, and active learning on graph hierarchies in order to label multi-object tracking" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "text", + "content": "SPAMming Labels: Efficient Annotations for the Trackers of Tomorrow" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 170, + 114, + 447, + 388 + ], + "blocks": [ + { + "bbox": [ + 170, + 114, + 447, + 388 + ], + "lines": [ + { + "bbox": [ + 170, + 114, + 447, + 388 + ], + "spans": [ + { + "bbox": [ + 170, + 114, + 447, + 388 + ], + "type": "table", + "html": "
MethodLabelsAnn. Effort ↓HOTA ↑MOTA ↑IDF1 ↑
MOT17 Test
ByteTrack [75]GT100%50.660.461.1
ByteTrack [75]SPAM3.3%51.664.063.0
GHOST [54]GT100%49.558.059.0
GHOST [54]SPAM3.3%51.361.962.1
MOT20 Test
ByteTrack [75]GT100%48.760.560.0
ByteTrack [75]SPAM3.3%47.957.661.4
ByteTrack [75]SPAM10%48.461.062.8
GHOST [54]GT100%48.359.760.2
GHOST [54]SPAM3.3%47.058.260.7
GHOST [54]SPAM10%46.959.060.8
DanceTrack Test
ByteTrack [75]GT100%41.079.146.0
ByteTrack [75]SPAM3.3%39.576.445.0
ByteTrack [75]SPAM20%41.380.948.1
GHOST [54]GT100%45.880.646.3
GHOST [54]SPAM3.3%41.076.344.8
GHOST [54]SPAM20%43.981.347.7
", + "image_path": "7b3383136a890fee3de50ab6da6cfa55ff292469678d34288ed228ee630464fb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 391, + 482, + 424 + ], + "lines": [ + { + "bbox": [ + 130, + 391, + 482, + 424 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 482, + 424 + ], + "type": "text", + "content": "Table 5: Evaluation of SPAM labels on MOT17, MOT20, and DanceTrack test sets with varying annotation budgets by training ByteTrack [75] and GHOST [54] using annotations generated by our method." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 443, + 482, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 443, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 443, + 482, + 468 + ], + "type": "text", + "content": "data, where we can save " + }, + { + "bbox": [ + 130, + 443, + 482, + 468 + ], + "type": "inline_equation", + "content": "96.7\\%" + }, + { + "bbox": [ + 130, + 443, + 482, + 468 + ], + "type": "text", + "content": " of the annotation effort on MOT17, " + }, + { + "bbox": [ + 130, + 443, + 482, + 468 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 130, + 443, + 482, + 468 + ], + "type": "text", + "content": " of the annotation effort on MOT20 and " + }, + { + "bbox": [ + 130, + 443, + 482, + 468 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 130, + 443, + 482, + 468 + ], + "type": "text", + "content": " of the annotation effort on DanceTrack." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 486, + 220, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 486, + 220, + 498 + ], + "spans": [ + { + "bbox": [ + 132, + 486, + 220, + 498 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 511, + 482, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 511, + 482, + 643 + ], + "spans": [ + { + "bbox": [ + 130, + 511, + 482, + 643 + ], + "type": "text", + "content": "We have introduced SPAM, a labeling engine to efficiently obtain trajectory annotations from videos. We have shown that SPAM is able to produce annotations that yield state-of-the-art tracking performance with a significantly reduced human annotation effort. Through our extensive experiments we have shown that (i) our synthetic pretraining " + }, + { + "bbox": [ + 130, + 511, + 482, + 643 + ], + "type": "inline_equation", + "content": "+" + }, + { + "bbox": [ + 130, + 511, + 482, + 643 + ], + "type": "text", + "content": " pseudo-label finetuning recipe yields a strong, annotation-free, labeler; (ii) our combination of active learning " + }, + { + "bbox": [ + 130, + 511, + 482, + 643 + ], + "type": "inline_equation", + "content": "+" + }, + { + "bbox": [ + 130, + 511, + 482, + 643 + ], + "type": "text", + "content": " pseudolabeling for annotation generation yields high quality and efficient annotations; and (iii) that graphs are an excellent tool to provide comprehensive modelling for trajectory annotation. We believe that SPAM paves the way to make large-scale cheap annotation of tracking datasets a reality, which will in turn power the next generation of data-hungry tracking algorithms." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 238, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 238, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 238, + 100 + ], + "type": "text", + "content": "O. Cetintas et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "type": "text", + "content": "1. Acuna, D., Ling, H., Kar, A., Fidler, S.: Efficient interactive annotation of segmentation datasets with polygon-rnn++. In: CVPR (2018) 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 162, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 162, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 138, + 162, + 480, + 194 + ], + "type": "text", + "content": "2. Alahi, A., Goel, K., Ramanathan, V., Robicquet, A., Fei-Fei, L., Savarese, S.: Social LSTM: Human trajectory prediction in crowded spaces. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2016) 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 195, + 480, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 195, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 138, + 195, + 480, + 225 + ], + "type": "text", + "content": "3. Andriyenko, A., Schindler, K.: Multi-target tracking by continuous energy minimization. In: CVPR. pp. 1265-1272 (2011). https://doi.org/10.1109/CVPR.2011.5995311 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 228, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 480, + 258 + ], + "type": "text", + "content": "4. Bastani, F., He, S., Madden, S.: Self-supervised multi-object tracking with cross-input consistency. Advances in Neural Information Processing Systems 34, 13695-13706 (2021) 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 260, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 260, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 138, + 260, + 480, + 281 + ], + "type": "text", + "content": "5. Beluch, W.H., Genewein, T., Nurnberger, A., Kohler, J.M.: The power of ensembles for active learning in image classification. In: CVPR. pp. 9368-9377 (2018) 12" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 282, + 480, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 282, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 138, + 282, + 480, + 304 + ], + "type": "text", + "content": "6. Berclaz, J., Fleuret, F., Turetken, E., Fua, P.: Multiple object tracking using k-shortest paths optimization. IEEE TPAMI 33(9), 1806-1819 (2011) 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 304, + 480, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 304, + 480, + 326 + ], + "spans": [ + { + "bbox": [ + 138, + 304, + 480, + 326 + ], + "type": "text", + "content": "7. Bergmann, P., Meinhardt, T., Leal-Taixe, L.: Tracking without bells and whistles. In: ICCV. pp. 941-951 (2019) 2, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 327, + 480, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 327, + 480, + 347 + ], + "spans": [ + { + "bbox": [ + 138, + 327, + 480, + 347 + ], + "type": "text", + "content": "8. Brasó, G., Cetintas, O., Leal-Taixe, L.: Multi-object tracking and segmentation via neural message passing. IJCV 130(12), 3035-3053 (2022) 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 348, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 348, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 138, + 348, + 480, + 369 + ], + "type": "text", + "content": "9. Braso, G., Leal-Taixe, L.: Learning a neural solver for multiple object tracking. In: CVPR (2020) 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 370, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 370, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 138, + 370, + 480, + 402 + ], + "type": "text", + "content": "0. Cai, J., Xu, M., Li, W., Xiong, Y., Xia, W., Tu, Z., Soatto, S.: Memot: Multi-object tracking with memory. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8090-8100 (2022) 4, 11" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 403, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 403, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 138, + 403, + 480, + 424 + ], + "type": "text", + "content": "1. Caron, M., Misra, I., Mairal, J., Goyal, P., Bojanowski, P., Joulin, A.: Unsupervised learning of visual features by contrasting cluster assignments. NeurIPS (2020) 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 425, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 425, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 138, + 425, + 480, + 445 + ], + "type": "text", + "content": "2. Cetintas, O., Brasó, G., Leal-Taixe, L.: Unifying short and long-term tracking with graph hierarchies. In: CVPR. pp. 22877-22887 (June 2023) 3, 4, 5, 6, 11" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 447, + 480, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 447, + 480, + 487 + ], + "spans": [ + { + "bbox": [ + 138, + 447, + 480, + 487 + ], + "type": "text", + "content": "3. Chen, L., Ai, H., Zhuang, Z., Shang, C.: Real-time multiple people tracking with deeply learned candidate selection and person re-identification. In: 2018 IEEE international conference on multimedia and expo (ICME). pp. 1-6. IEEE (2018) 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 490, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 490, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 138, + 490, + 480, + 522 + ], + "type": "text", + "content": "4. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: IEEE Int. Conf. Mach. Learn. pp. 1597-1607. PMLR (2020) 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 523, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 523, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 523, + 480, + 544 + ], + "type": "text", + "content": "5. Chu, P., Ling, H.: Famnet: Joint learning of feature, affinity and multi-dimensional assignment for online multiple object tracking. In: ICCV (October 2019) 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 545, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 545, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 545, + 480, + 567 + ], + "type": "text", + "content": "6. Dai, K., Zhao, J., Wang, L., Wang, D., Li, J., Lu, H., Qian, X., Yang, X.: Video annotation for visual tracking via selection and refinement. In: ICCV (2021) 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 567, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 480, + 599 + ], + "type": "text", + "content": "7. Dai, Z., Cai, B., Lin, Y., Chen, J.: Up-detr: Unsupervised pre-training for object detection with transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1601-1610 (June 2021) 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "type": "text", + "content": "8. Dendorfer, P., Osep, A., Milan, A., Schindler, K., Cremers, D., Reid, I., Roth, S., Leal-Taixe, L.: Motchallenge: A benchmark for single-camera multiple target tracking. International Journal of Computer Vision 129(4), 845-881 (2021) 3, 7, 9" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "text", + "content": "9. Dendorfer, P., Rezatofighi, H., Milan, A., Shi, J.Q., Cremers, D., Reid, I.D., Roth, S., Schindler, K., Leal-Taix'e, L.: Mot20: A benchmark for multi object tracking in crowded scenes. ArXiv abs/2003.09003 (2020) 3, 7, 9" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "text", + "content": "SPAMming Labels: Efficient Annotations for the Trackers of Tomorrow" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 663 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "20. Dólar, P., Wojek, C., Schiele, B., Perona, P.: Pedestrian detection: A benchmark. In: CVPR. pp. 304-311. IEEE (2009) 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 480, + 171 + ], + "type": "text", + "content": "21. Elezi, I., Yu, Z., Anandkumar, A., Leal-Taixe, L., Alvarez, J.M.: Not all labels are equal: Rationalizing the labeling costs for training object detection. In: CVPR. pp. 14492-14501 (2022) 12" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 204 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 204 + ], + "type": "text", + "content": "22. Ess, A., Leibe, B., Schindler, K., Van Gool, L.: A mobile vision system for robust multi-person tracking. In: CVPR. pp. 1-8 (2008). https://doi.org/10.1109/CVPR.2008.4587581 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 247 + ], + "type": "text", + "content": "23. Fabbri, M., Brasó, G., Maugeri, G., Cetintas, O., Gasparini, R., Osep, A., Calderara, S., Leal-Taixe, L., Cucchiara, R.: Motsynth: How can synthetic data help pedestrian detection and tracking? In: ICCV. pp. 10849-10859 (October 2021) 3, 5, 6, 9" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "type": "text", + "content": "24. Fabbri, M., Lanzi, F., Calderara, S., Palazzi, A., Vezzani, R., Cucchiara, R.: Learning to detect and track visible and occluded body joints in a virtual world. In: ECCV (2018) 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 282, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 480, + 303 + ], + "type": "text", + "content": "25. Feichtenhofer, C., Pinz, A., Zisserman, A.: Detect to track and track to detect. In: ICCV (Oct 2017) 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "text", + "content": "26. Gaidon, A., Wang, Q., Cabon, Y., Vig, E.: Virtual worlds as proxy for multi-object tracking analysis. In: CVPR. pp. 4340-4349 (2016) 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 326, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 346 + ], + "type": "text", + "content": "27. Ge, Z., Liu, S., Wang, F., Li, Z., Sun, J.: Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430 (2021) 9, 11" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "text", + "content": "28. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) 9" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "text", + "content": "29. Hu, H.N., Cai, Q.Z., Wang, D., Lin, J., Sun, M., Krahenbihl, P., Darrell, T., Yu, F.: Joint monocular 3d vehicle detection and tracking. In: ICCV (2019) 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 403, + 480, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 434 + ], + "type": "text", + "content": "30. Jabri, A., Owens, A., Efros, A.: Space-time correspondence as a contrastive random walk. Advances in neural information processing systems 33, 19545-19560 (2020) 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 434, + 480, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 434, + 480, + 466 + ], + "spans": [ + { + "bbox": [ + 132, + 434, + 480, + 466 + ], + "type": "text", + "content": "31. Jiang, H., Fels, S., Little, J.J.: A linear programming approach for multiple object tracking. In: CVPR. pp. 1-8 (2007). https://doi.org/10.1109/CVPR.2007.3831804" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 468, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 511 + ], + "type": "text", + "content": "32. Kasturi, R., Goldgof, D., Soundararajan, P., Manohar, V., Garofolo, J., Boonstra, M., Korzhova, V., Zhang, J.: Framework for performance evaluation for face, text and vehicle detection and tracking in video: data, metrics, and protocol. IEEE TPAMI (2009) 9" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "text", + "content": "33. Kaul, P., Xie, W., Zisserman, A.: Label, verify, correct: A simple few-shot object detection method. In: IEEE Conference on Computer Vision and Pattern Recognition (2022) 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 545, + 480, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 480, + 576 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 480, + 576 + ], + "type": "text", + "content": "34. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: ICCV (2023) 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 577, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 577, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 577, + 480, + 610 + ], + "type": "text", + "content": "35. Lan, S., Yang, X., Yu, Z., Wu, Z., Alvarez, J.M., Anandkumar, A.: Vision transformers are good mask auto-labelers. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2023) 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "text", + "content": "36. Leal-Taixe, L., Canton-Ferrer, C., Schindler, K.: Learning by tracking: Siamese cnn for robust target association. In: CVPRW (June 2016) 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 632, + 480, + 663 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 663 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 663 + ], + "type": "text", + "content": "37. Leal-Taixe, L., Fenzi, M., Kuznetsova, A., Rosenhahn, B., Savarese, S.: Learning an image-based motion context for multiple people tracking. In: CVPR (June 2014) 4" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 237, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 237, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 237, + 100 + ], + "type": "text", + "content": "O. Cetintas et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 159 + ], + "type": "text", + "content": "38. Leal-Taixe, L., Pons-Moll, G., Rosenhahn, B.: Everybody needs somebody: Modeling social and grouping behavior on a linear programming multiple people tracker. In: Int. Conf. Comput. Vis. Worksh. pp. 120-127 (2011). https://doi.org/10.1109/ICCVW.2011.6130233 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 160, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 480, + 182 + ], + "type": "text", + "content": "39. Li, R., Zhang, B., Liu, J., Liu, W., Zhao, J., Teng, Z.: Heterogeneous diversity driven active learning for multi-object tracking. In: ICCV (2023) 2, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 480, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 480, + 225 + ], + "type": "text", + "content": "40. Litrico, M., Del Bue, A., Morerio, P.: Guiding pseudo-labels with uncertainty estimation for source-free unsupervised domain adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023) 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 225, + 480, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 225, + 480, + 257 + ], + "spans": [ + { + "bbox": [ + 132, + 225, + 480, + 257 + ], + "type": "text", + "content": "41. Liu, K., Jin, S., Fu, Z., Chen, Z., Jiang, R., Ye, J.: Uncertainty-aware unsupervised multi-object tracking. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9996-10005 (2023) 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 257, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 257, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 480, + 289 + ], + "type": "text", + "content": "42. Luiten, J., Osep, A., Dendorfer, P., Torr, P., Geiger, A., Leal-Taixe, L., Leibe, B.: Hota: A higher order metric for evaluating multi-object tracking. IJCV 129(2), 548-578 (2021) 9" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 289, + 480, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 480, + 322 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 480, + 322 + ], + "type": "text", + "content": "43. Manen, S., Gygli, M., Dai, D., Van Gool, L.: Pathtrack: Fast trajectory annotation with path supervision. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 290-299 (2017) 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 322, + 480, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 322, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 132, + 322, + 480, + 354 + ], + "type": "text", + "content": "44. Meinhardt, T., Kirillov, A., Leal-Taixe, L., Feichtenhofer, C.: Trackformer: Multi-object tracking with transformers. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022) 1, 4, 11" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 354, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 480, + 396 + ], + "type": "text", + "content": "45. Meng, S., Shao, D., Guo, J., Gao, S.: Tracking without label: Unsupervised multiple object tracking via contrastive similarity learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16264-16273 (2023) 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 396, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 480, + 418 + ], + "type": "text", + "content": "46. Pang, J., Qiu, L., Li, X., Chen, H., Li, Q., Darrell, T., Yu, F.: Quasi-dense similarity learning for multiple object tracking. In: CVPR. pp. 164-173 (2021) 4, 11" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 418, + 480, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 480, + 450 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 480, + 450 + ], + "type": "text", + "content": "47. Pellegrini, S., Ess, A., Schindler, K., van Gool, L.: You'll never walk alone: Modeling social behavior for multi-target tracking. In: ICCV. pp. 261-268 (2009). https://doi.org/10.1109/ICCV.2009.5459260 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 450, + 480, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 480, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 480, + 482 + ], + "type": "text", + "content": "48. Pirsiavash, H., Ramanan, D., Fowlkes, C.C.: Globally-optimal greedy algorithms for tracking a variable number of objects. In: CVPR. pp. 1201-1208 (2011). https://doi.org/10.1109/CVPR.2011.5995604 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 482, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 482, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 482, + 480, + 525 + ], + "type": "text", + "content": "49. Qin, Z., Zhou, S., Wang, L., Duan, J., Hua, G., Tang, W.: Motiontrack: Learning robust short-term and long-term motions for multi-object tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17939-17948 (2023) 4, 11" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 525, + 480, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 480, + 558 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 480, + 558 + ], + "type": "text", + "content": "50. Ristani, E., Solera, F., Zou, R., Cucchiara, R., Tomasi, C.: Performance measures and a data set for multi-target, multi-camera tracking. In: Eur. Conf. Comput. Vis. Worksh. pp. 17-35. Springer (2016) 9" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 558, + 480, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 558, + 480, + 579 + ], + "spans": [ + { + "bbox": [ + 132, + 558, + 480, + 579 + ], + "type": "text", + "content": "51. Ristani, E., Tomasi, C.: Features for multi-target multi-camera tracking and re-identification. In: CVPR (June 2018) 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 579, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 579, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 579, + 480, + 601 + ], + "type": "text", + "content": "52. Robicquet, A., Sadeghian, A., Alahi, A., Savarese, S.: Learning social etiquette: Human trajectory understanding in crowded scenes. In: ECCV (2016) 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 601, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 480, + 632 + ], + "type": "text", + "content": "53. Scovanner, P., Tappen, M.F.: Learning pedestrian dynamics from the real world. In: 2009 IEEE 12th International Conference on Computer Vision. pp. 381-388. IEEE (2009) 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "54. Seidenschwarz, J., Brasó, G., Serrano, V.C., Elezi, I., Leal-Taixe, L.: Simple cues lead to a strong multi-object tracker. In: CVPR. pp. 13813-13823 (2023) 4, 9, 11, 13, 14" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "text", + "content": "SPAMming Labels: Efficient Annotations for the Trackers of Tomorrow" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 148 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 148 + ], + "type": "text", + "content": "55. Sener, O., Savarese, S.: Active learning for convolutional neural networks: A core-set approach. In: ICLR (2018), https://openreview.net/forum?id=H1aIuk-RW 12" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 149, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 480, + 182 + ], + "type": "text", + "content": "56. Shao, S., Zhao, Z., Li, B., Xiao, T., Yu, G., Zhang, X., Sun, J.: Crowdhuman: A benchmark for detecting human in a crowd. arXiv preprint arXiv:1805.00123 (2018) 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 480, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 214 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 214 + ], + "type": "text", + "content": "57. Sun, P., Cao, J., Jiang, Y., Yuan, Z., Bai, S., Kitani, K., Luo, P.: Dancetrack: Multiobject tracking in uniform appearance and diverse motion. In: CVPR (2022) 3, 9, 13" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 480, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 247 + ], + "type": "text", + "content": "58. Sun, S., Akhtar, N., Song, X., Song, H., Mian, A., Shah, M.: Simultaneous detection and tracking with motion modelling for multiple object tracking. In: ECCV. pp. 626-643 (2020) 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 480, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 282 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 282 + ], + "type": "text", + "content": "59. Voigtlaender, P., Krause, M., Osep, A., Luiten, J., Sekar, B.B.G., Geiger, A., Leibe, B.: Mots: Multi-object tracking and segmentation. In: Proceedings of the iee/cvf conference on computer vision and pattern recognition. pp. 7942-7951 (2019) 2, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 282, + 480, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 480, + 323 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 480, + 323 + ], + "type": "text", + "content": "60. Voigtlaender, P., Luo, L., Yuan, C., Jiang, Y., Leibe, B.: Reducing the annotation effort for video object segmentation datasets. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3060-3069 (2021) 2, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 325, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 480, + 346 + ], + "type": "text", + "content": "61. Vondrick, C., Ramanan, D.: Video annotation and tracking with active learning. NeurIPS 24 (2011) 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 348, + 480, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 348, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 348, + 480, + 390 + ], + "type": "text", + "content": "62. Vondrick, C., Ramanan, D., Patterson, D.: Efficiently scaling up video annotation with crowdsourced marketplaces. In: Computer Vision-ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part IV 11. pp. 610-623. Springer (2010) 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 392, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 424 + ], + "type": "text", + "content": "63. Vondrick, C., Shrivastava, A., Fathi, A., Guadarrama, S., Murphy, K.: Tracking emerges by colorizing videos. In: Proceedings of the European conference on computer vision (ECCV). pp. 391-408 (2018) 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 425, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 425, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 425, + 480, + 445 + ], + "type": "text", + "content": "64. Wang, X., Jabri, A., Efros, A.A.: Learning correspondence from the cycle-consistency of time. In: CVPR (2019) 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 447, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 480, + 479 + ], + "type": "text", + "content": "65. Wang, Y., Kitani, K., Weng, X.: Joint object detection and multi-object tracking with graph neural networks. In: IEEE Int. Conf. Robotics and Autom. pp. 13708-13715 (2021). https://doi.org/10.1109/ICRA48506.2021.9561110 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "type": "text", + "content": "66. Wang, Z., Zheng, L., Liu, Y., Wang, S.: Towards real-time multi-object tracking. The European Conference on Computer Vision (ECCV) (2020) 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 502, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 480, + 533 + ], + "type": "text", + "content": "67. Wei, F., Gao, Y., Wu, Z., Hu, H., Lin, S.: Aligning pretraining for detection via object-level contrastive learning. In: Advances in Neural Information Processing Systems (2021) 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "type": "text", + "content": "68. Xiao, T., Li, S., Wang, B., Lin, L., Wang, X.: Joint detection and identification feature learning for person search. In: CVPR. pp. 3415-3424 (2017) 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 556, + 480, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 556, + 480, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 556, + 480, + 586 + ], + "type": "text", + "content": "69. Yamaguchi, K., Berg, A.C., Ortiz, L.E., Berg, T.L.: Who are you with and where are you going? In: CVPR. pp. 1345-1352 (2011). https://doi.org/10.1109/CVPR.2011.5995468 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 588, + 480, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 588, + 480, + 631 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 480, + 631 + ], + "type": "text", + "content": "70. You, S., Yao, H., Bao, B.K., Xu, C.: Utm: A unified multiple object tracking model with identity-aware feature enhancement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21876-21886 (2023) 4, 11" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "71. Yuan, D., Chang, X., Liu, Q., Yang, Y., Wang, D., Shu, M., He, Z., Shi, G.: Active learning for deep visual tracking. IEEE Transactions on Neural Networks and Learning Systems (2023) 2" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 237, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 237, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 237, + 100 + ], + "type": "text", + "content": "O. Cetintas et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 304 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "text", + "content": "72. Zeng, F., Dong, B., Zhang, Y., Wang, T., Zhang, X., Wei, Y.: Motr: End-to-end multiple-object tracking with transformer. In: European Conference on Computer Vision. pp. 659-675. Springer (2022) 1, 4, 11" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 171 + ], + "type": "text", + "content": "73. Zhang, L., Li, Y., Nevatia, R.: Global data association for multi-object tracking using network flows. In: CVPR (2008) 4, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 193 + ], + "type": "text", + "content": "74. Zhang, S., Benenson, R., Schiele, B.: Citypersons: A diverse dataset for pedestrian detection. In: CVPR. pp. 3213-3221 (2017) 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "text", + "content": "75. Zhang, Y., Sun, P., Jiang, Y., Yu, D., Weng, F., Yuan, Z., Luo, P., Liu, W., Wang, X.: Bytetrack: Multi-object tracking by associating every detection box. In: ECCV. pp. 1-21. Springer (2022) 4, 9, 11, 13, 14" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "type": "text", + "content": "76. Zhang, Y., Wang, C., Wang, X., Zeng, W., Liu, W.: Fairmot: On the fairness of detection and re-identification in multiple object tracking. IJCV 129(11), 3069-3087 (2021) 1, 4, 9, 11" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "type": "text", + "content": "77. Zheng, L., Zhang, H., Sun, S., Chandraker, M., Yang, Y., Tian, Q.: Person re-identification in the wild. In: CVPR. pp. 1367-1376 (2017) 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 282, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 304 + ], + "type": "text", + "content": "78. Zhou, X., Koltun, V., Krahenbuhl, P.: Tracking objects as points. In: ECCV. pp. 474-490. Springer (2020) 1, 4" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "text", + "content": "SPAMming Labels: Efficient Annotations for the Trackers of Tomorrow" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/SPARO_ Selective Attention for Robust and Compositional Transformer Encodings for Vision/4bccfc91-c07f-416d-bd12-0ad527681e54_content_list.json b/2024/SPARO_ Selective Attention for Robust and Compositional Transformer Encodings for Vision/4bccfc91-c07f-416d-bd12-0ad527681e54_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0c7537d102786eef26603995dea6ddd0211e53b9 --- /dev/null +++ b/2024/SPARO_ Selective Attention for Robust and Compositional Transformer Encodings for Vision/4bccfc91-c07f-416d-bd12-0ad527681e54_content_list.json @@ -0,0 +1,1797 @@ +[ + { + "type": "image", + "img_path": "images/fcf0f2e780151c3244b95b9fe666092896fd1f14ea02089ec12accca332f2455.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 225, + 143, + 269, + 170 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "SPARO: Selective Attention for Robust and", + "text_level": 1, + "bbox": [ + 272, + 151, + 776, + 171 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Compositional Transformer Encodings for Vision", + "text_level": 1, + "bbox": [ + 223, + 175, + 779, + 195 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ankit Vani1*, Bac Nguyen2, Samuel Lavoie1, Ranjay Krishna3, and Aaron Courville1,4", + "bbox": [ + 235, + 220, + 767, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Mila, Université de Montréal", + "bbox": [ + 395, + 263, + 604, + 277 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Sony AI", + "bbox": [ + 464, + 279, + 535, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 University of Washington, Allen Institute for Artificial Intelligence", + "bbox": [ + 272, + 292, + 728, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 CIFAR AI Chair", + "bbox": [ + 437, + 306, + 565, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Selective attention helps us focus on task-relevant aspects in the constant flood of our sensory input. This constraint in our perception allows us to robustly generalize under distractions and to new compositions of perceivable concepts. Transformers employ a similar notion of attention in their architecture, but representation learning models with transformer backbones like CLIP and DINO often fail to demonstrate robustness and compositionality. We highlight a missing architectural prior: unlike human perception, transformer encodings do not separately attend over individual concepts. In response, we propose SPARO, a readout mechanism that partitions encodings into separately-attended slots, each produced by a single attention head. Using SPARO with CLIP imparts an inductive bias that the vision and text modalities are different views of a shared compositional world with the same corresponding concepts. Using SPARO, we demonstrate improvements on downstream recognition, robustness, retrieval, and compositionality benchmarks with CLIP (up to $+14\\%$ for ImageNet, $+4\\%$ for SugarCrepe), and on nearest neighbors and linear probe for ImageNet with DINO ( $+3\\%$ each). We also showcase a powerful ability to intervene and select individual SPARO concepts to further improve downstream task performance (up from $+4\\%$ to $+9\\%$ for SugarCrepe) and use this ability to study the robustness of SPARO's representation structure. Finally, we provide insights through ablation experiments and visualization of learned concepts.", + "bbox": [ + 259, + 361, + 743, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Selective attention $\\cdot$ Slot representations $\\cdot$ Transformers", + "bbox": [ + 261, + 679, + 723, + 694 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 720, + 377, + 736 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Selective attention is an intrinsic property of human perception [15,44,47], enabling people to focus on task-relevant aspects of their surroundings while tuning out the rest [55]. For instance, it allows a driver to focus on traffic signals, road signs, and other vehicles, or an individual searching for their friend", + "bbox": [ + 212, + 753, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Correspondence at: ankit.vani@umontreal.ca.", + "bbox": [ + 217, + 824, + 550, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "in a crowd to attend to details like people's heights and clothing. More broadly, it empowers us to comprehend the vast complexity of our world with limited mental resources [6] by constraining our perception to one concept at a time. When a task requires information for multiple concepts (e.g., looking at traffic signals and pedestrians), this constraint calls for sequentially attending to separate salient aspects of our stimuli [59]. The cognitive bottleneck of selective attention, with the ability to separately represent concepts from our environment, enables human perception to remain insensitive to irrelevant distractions and generalize to new compositions of perceivable concepts.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In machine learning, the notion of selective attention is party emulated by the attention mechanism 3, which is a core component of transformers 60 such as the vision transformer (ViT) 18. However, transformer encodings learnt using approaches like CLIP 50 and DINO 9 still struggle with robustness 30, 31, 62, 70 and compositional generalization 33, 42, 52, 58, 67, 72. For instance, despite CLIP emerging as the de facto backbone for many vision tasks, it can perform close to chance when evaluated for compositionality 42. While the attention mechanism equips these models with the ability to emphasize relevant aspects of their intermediate representations, it does not endow them with the means to produce encodings for separately-attended concepts. Consequently, transformer-based representation learning models lack an important prior from human cognition: that perception of a complex input can be broken down into perception of its salient concepts.", + "bbox": [ + 212, + 282, + 787, + 479 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This limitation deprives downstream tasks easy access to task-relevant aspects of the data's underlying structure. In a task like ImageNet classification, where most images contain one primary object of interest (e.g., \"dog\"), information like relationships between objects and attributes of the background are irrelevant and can lead to overfitting if considered. However, these details can become crucial to attain high performance in other tasks like image retrieval (e.g., \"dog chasing a frisbee in a park\"). Similar to the generalization benefits of selective attention in human perception, encodings that allow easy separation of attended concepts in the data can generalize to a diverse set of downstream tasks. Without additional constraints, however, the emergence of such compositional structure in encodings is unlikely, especially in the common setting of training on noisy internet data.", + "bbox": [ + 212, + 481, + 787, + 660 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose SPARO (Separate-head attention read-out) as an improved read-out mechanism for transformer encoders in vision inspired by selective attention. SPARO replaces the last transformer block to provide a mechanism for partitioning encodings into slots of separately-attended concepts. We design each slot encoding to be a low-dimensional result of a single-head attention operation, producing a bottleneck that encourages each slot to \"selectively attend.\" In this mechanism, producing multiple slots can be interpreted as selective attention over different concepts occurring in parallel. When training CLIP, SPARO imparts an inductive bias that both modalities are views onto a shared compositional world with the corresponding slots of both encoders representing its concepts.", + "bbox": [ + 212, + 661, + 787, + 813 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "A. Vani et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "5 Source code: https://github.com/ankitkv/sparo-clip", + "bbox": [ + 217, + 823, + 612, + 842 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In contrast, standard CLIP merely embeds the encodings of both modalities in a shared vector space without imposing any additional structure.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Training SPARO using CLIP and DINO exhibits improved generalization, robustness, and compositionality while retaining the same model size. Our experiments show that using CLIP with SPARO improves generalization for zero-shot recognition, robustness, retrieval, and compositionality (e.g., +14% for ImageNet, +4% for SugarCrepe, when trained with 15M examples; +3% each when trained with 400M examples). We also demonstrate improved linear probe performance of CLIP (+10% trained with 15M examples; +2% with 400M examples), as well as both linear probe and nearest neighbors performance of DINO (+3% each). We showcase how SPARO's representational structure enables the ability to manually intervene and select relevant concepts for downstream tasks, leading to improved compositionality and generalization (up from +4% to +9% for SugarCrepe trained with 15M examples; +3% to +6% with 400M examples). We then study the robustness of this structure, validate the choice of separate-head attention through ablations, and provide insights into the effect of the number and dimensionality of concepts. Finally, we provide visualizations for some of the concepts SPARO learns to attend to.", + "bbox": [ + 212, + 176, + 787, + 419 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related work", + "text_level": 1, + "bbox": [ + 215, + 443, + 382, + 458 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We situate our work amongst other transformer read-out mechanisms and slot representation learning methods.", + "bbox": [ + 212, + 476, + 784, + 506 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Transformer read-outs. For transformer [60] decoders with causal masking, the last hidden state, corresponding to the end of sequence (EOS) input token, is used as the output representation. BERT [17] proposed inserting a special CLS token before input sequences for transformer encoders where information from all positions is gathered to produce the encoding. Using the CLS token output is also popular for Vision transformer (ViT) [9, 12, 18] backbones, along with alternatives such as global average pooling [12, 64]. In the context of language, Multi-CLS BERT [10] argues that an input text sequence can have multiple facets, and that samples can be similar along one facet and dissimilar along another. The authors aim to represent these facets by adding multiple CLS tokens to the input sequence. Using separate attention heads for each SPARO slot can also encourage attending to different facets of the inputs, but without requiring expensive accumulations through each layer of the backbone. By keeping SPARO independent of the backbone, we also allow for its easy potential placement on top of frozen pre-trained models without needing to train for added CLS tokens.", + "bbox": [ + 212, + 522, + 787, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SPARO is most related to the attentional pooler [39,66] architecture, which uses multi-head attention with embedded learned queries to produce output encodings. However, SPARO employs a critical structural constraint of not mixing the information from separate attention heads, a design choice we justify empirically in Sec. 5.3. Slot attention [41] also produces slot encodings through the attention mechanism, but rather than having the input positions compete", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "SPARO: Selective Attention in Transformer Encodings for Vision", + "bbox": [ + 303, + 114, + 732, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "for each slot, the slots compete for each input position. Slot attention also proposes an expensive iterative read-out mechanism where the role of each slot is determined by sampling from a per-sample Gaussian distribution, which can be limiting for large real-world datasets.", + "bbox": [ + 212, + 146, + 787, + 207 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Slot representations. Object-centric slot representations have been studied extensively in the literature [7,8,20,21,24-27,41,43,71]. Often, these representations are expected to represent predefined objects and attributes, and evaluated in small-scale or synthetic settings. Recently, the use of slot attention on features produced by a learned encoder [1,49,56], and utilizing unsupervised saliency masks for object discovery [69] has enabled unsupervised learning of object-centric representations in larger scale settings. Related to these ideas, SPARO attends to features produced by a transformer using a single attention head per slot, and each head is capable of learning saliency maps without supervision [9]. However, we do not constrain our slots to correspond to predefined types of objects in the data, nor do we impose any independence conditions between the slots. Self-supervised simplicial embeddings (SEM) [37], with only a soft-discrete slot structure constraint, enable improved generalization on downstream tasks. SPARO replaces the separation of softmax application between SEM slots with a separation of underlying attention operations.", + "bbox": [ + 212, + 223, + 789, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 476, + 330, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Having motivated our approach, we now detail the design of SPARO. We start by introducing the relevant notation, then discuss the SPARO module. Finally, we discuss the addition of SPARO in CLIP and DINO, and the applicability of SPARO beyond these settings.", + "bbox": [ + 212, + 508, + 787, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Notation", + "text_level": 1, + "bbox": [ + 215, + 594, + 336, + 608 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We envision SPARO as a special attention-based layer that modifies the top block of a transformer encoder [51,60] backbone (including the ViT [18]). Transformers take as input a sequence $\\pmb{x} = \\{\\pmb{x}_1,\\dots,\\pmb{x}_n\\}$ of length $n$ , and produce an output state per position $\\{h_1,\\ldots ,h_n\\}$ , where $\\pmb{x}_i\\in \\mathbb{R}^{\\mathrm{inputdim}}$ , $\\pmb{h}_i\\in \\mathbb{R}^d$ , for any $i\\in [n]$ , with $d$ the model width. Finally, a pooling operation reduces the sequence of output states to output encodings with a fixed size $\\pmb{y}\\in \\mathbb{R}^{M}$ . Examples of pooling operations include attentional pooling [39], global average pooling, and extracting the CLS or EOS token representations for images or text respectively.", + "bbox": [ + 212, + 621, + 787, + 743 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2 Separate-head attention read-out (SPARO)", + "text_level": 1, + "bbox": [ + 215, + 767, + 609, + 782 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We replace the pooling operation of transformer encoders with a concatenation of outputs of $L$ single-head attention mechanisms. SPARO acts on the transformer outputs $\\pmb{H} = \\{\\pmb{h}_1, \\dots, \\pmb{h}_n\\} \\in \\mathbb{R}^{n \\times d}$ to produce the encoder output", + "bbox": [ + 212, + 794, + 787, + 842 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "A. Vani et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c0ab253b9f9f3a5a1694cc4ff2e7722630d79db6eaf7a0d69bb5d559a6a8165b.jpg", + "image_caption": [ + "Fig. 1: Illustration of SPARO, a read-out mechanism that structures representations as collections of separately-attended concepts. Take a standard $N$ -block transformer encoder (ViT here as an example), producing an encoding $y$ through extraction of its CLS token output. We can replace the $N$ th transformer block with the SPARO module (typically with equal or fewer parameters) to produce a SPARO encoding $y$ , which is a concatenation of $L$ SPARO slots. Each SPARO slot $y_l$ is produced through single-head attention over the backbone outputs using an embedded query $q_l$ . The value projection is a composition of slot-specific key projection parameterized by $K_l$ and a slot-wise projection shared between all SPARO concepts parameterized by $W$ ." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 787, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\pmb{y} = \\mathrm{concat}(\\pmb{y}_1, \\dots, \\pmb{y}_L) \\in \\mathbb{R}^{LV}$ , where SPARO slot $\\pmb{y}_l \\in \\mathbb{R}^V$ , $l \\in [L]$ . We illustrate the architecture of SPARO in Fig. 1. Each SPARO slot is produced as:", + "bbox": [ + 212, + 484, + 787, + 517 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {y} _ {l} = \\boldsymbol {W} \\boldsymbol {K} _ {l} \\boldsymbol {H} ^ {\\intercal} \\text {s o f t m a x} \\left(\\frac {\\boldsymbol {H} \\boldsymbol {K} _ {l} ^ {\\intercal} \\boldsymbol {q} _ {l}}{\\sqrt {D}}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 369, + 537, + 785, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $\\pmb{K}_l \\in \\mathbb{R}^{D \\times d}$ is a learned key projection weight, and $\\pmb{q}_l \\in \\mathbb{R}^D$ is a learned query embedding for attention. We parameterize the value projection as $\\pmb{W}\\pmb{K}_l$ where $\\pmb{W} \\in \\mathbb{R}^{V \\times D}$ is learned parameter shared between all $L$ slots. Decomposing the value projection in this manner helps reduce the number of model parameters, allowing a larger choice of $L$ with fixed resources.", + "bbox": [ + 212, + 582, + 787, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Generally, we pick $V$ and $D$ to be significantly smaller than the transformer width. In practice, we use the typical value of $D = 64$ used in standard multihead attention modules, and set $V$ to the same value in our experiments. Therefore, each SPARO slot has limited expressivity and representational capacity. The bottleneck underpinning each SPARO slot is a result of these constraints and the mechanistic bottleneck of a single attention mechanism instantiating competition between its input features.", + "bbox": [ + 212, + 659, + 787, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SPARO can be thought of as multi-head attention [60] where the queries are learned embeddings, key and value projection weights are shared, and the output projection weight is block-diagonal with each block containing the same parameters. Assuming $L = V = D = \\sqrt{d}$ , SPARO requires a total of $d^2 + 2d$ parameters, compared to multi-head attention's $4d^2$ .", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "SPARO: Selective Attention in Transformer Encodings for Vision", + "bbox": [ + 303, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3 CLIP with SPARO", + "text_level": 1, + "bbox": [ + 215, + 146, + 410, + 159 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We add SPARO to both the image and text encoders of CLIP [50], setting the same values of $L$ and $V$ in both encoders. Each SPARO slot is $\\ell_2$ -normalized separately, and we divide their concatenation by $\\sqrt{L}$ to produce the global $\\ell_2$ -normalized encoding. With this scheme, the cosine similarity between the image encoding $\\mathbf{y}^{\\mathrm{i}}$ and the text encoding $\\mathbf{y}^{\\mathrm{t}}$ becomes the expected cosine similarity between corresponding slot encodings:", + "bbox": [ + 212, + 167, + 787, + 260 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {y} ^ {\\mathrm {i}} ^ {\\top} \\boldsymbol {y} ^ {\\mathrm {t}} = \\sum_ {l = 1} ^ {L} \\| \\boldsymbol {y} _ {l} ^ {\\mathrm {i}} \\| \\| \\boldsymbol {y} _ {l} ^ {\\mathrm {t}} \\| \\frac {\\boldsymbol {y} _ {l} ^ {\\mathrm {i}} {} ^ {\\top} \\boldsymbol {y} _ {l} ^ {\\mathrm {t}}}{\\| \\boldsymbol {y} _ {l} ^ {\\mathrm {i}} \\| \\| \\boldsymbol {y} _ {l} ^ {\\mathrm {t}} \\|} = \\sum_ {l = 1} ^ {L} \\frac {1}{L} \\operatorname {s i m} (\\boldsymbol {y} _ {l} ^ {\\mathrm {i}}, \\boldsymbol {y} _ {l} ^ {\\mathrm {t}}) = \\underset {l \\sim [ L ]} {\\mathbb {E}} \\left[ \\operatorname {s i m} (\\boldsymbol {y} _ {l} ^ {\\mathrm {i}}, \\boldsymbol {y} _ {l} ^ {\\mathrm {t}}) \\right]. \\quad (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 235, + 266, + 785, + 308 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We train the model using the standard CLIP loss which maximizes $\\mathrm{sim}(\\pmb{y}^{\\mathrm{i}},\\pmb{y}^{\\mathrm{t}})$ of encodings $\\pmb{y}^{\\mathrm{i}}$ and $\\pmb{y}^{\\mathrm{t}}$ for aligned image-text pairs and minimizes it across all other pairings. With SPARO's separate-head attention bottleneck, maximizing $\\mathbb{E}_{l\\sim [L]}\\left[\\mathrm{sim}(\\pmb{y}_l^{\\mathrm{i}},\\pmb{y}_l^{\\mathrm{t}})\\right]$ for aligned pairs across the training dataset encourages both modalities to learn similar attention semantics per slot $l\\in [L]$ . Consequently, SPARO imparts a prior that both modalities encode a shared world with $L$ selectively attendable concepts, and similarities between inputs can be expressed as their average similarity for these concepts.", + "bbox": [ + 212, + 314, + 787, + 436 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4 DINO with SPARO", + "text_level": 1, + "bbox": [ + 215, + 455, + 415, + 469 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The online and momentum encoders of DINO [9] are each comprised of a ViT backbone which produces the encoding, and a DINO head that transforms it into a categorical distribution for distillation. Unlike CLIP, these encoders have an inherent bias for learning similar encoding functions due to one being an exponentially moving average of the other. However, in standard DINO, there are no constraints for selective attention in the encoding structure. We add SPARO to DINO to impart the prior for representing separately-attendable concepts in its encodings. Concretely, we replace the final transformer block of the ViT backbone with the SPARO, but do not modify the DINO head.", + "bbox": [ + 212, + 476, + 787, + 613 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 Results", + "text_level": 1, + "bbox": [ + 215, + 633, + 323, + 648 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Secs. 4 and 5, we evaluate our method on a variety of downstream tasks, analyze SPARO's representation structure with the ability to intervene on selected slots, perform ablations to support our design, and visualize the learned concepts.", + "bbox": [ + 212, + 661, + 785, + 707 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we validate that partitioning the representation structure as a collection of separately-attended concepts leads to improved generalization for downstream recognition, robustness, compositionality, and retrieval, using SPARO with CLIP [50] in Secs. [4.1] to [4.3] and DINO [9] in Sec. [4.3].", + "bbox": [ + 212, + 708, + 785, + 771 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We train our CLIP models on one of Conceptual Captions 3M (CC3M) [57], Conceptual Captions 12M (CC12M) [11], a combination of CC3M and CC12M (CC15M), or LAION-400M (L400M) [54]. Appendix C provides the statistics for our training datasets.", + "bbox": [ + 214, + 779, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "A. Vani et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4f02047679b37ee6918f6e9a857e471f3ebbdc8d930d92c0a1aa34032f8c6178.jpg", + "table_caption": [ + "Table 1: Zero-shot recognition, robustness, and SugarCrepe compositionality for CLIP models trained on Conceptual Captions and LAION-400M." + ], + "table_footnote": [], + "table_body": "
TrainModelImageNet-Object NetSugar Crepe
V1V2SketchAR
CC3MCLIP16 (C)0.1410.1220.0680.0330.1770.0800.611
C+GAP0.1560.1340.0690.0330.1870.0810.616
C+SPARO0.1700.1400.0880.0350.2210.0980.625
CC12MCLIP16 (C)0.3610.3110.2490.0910.4670.2180.697
C+GAP0.3820.3300.2620.1010.5010.2410.695
C+SPARO0.4060.3500.2980.1130.5590.2680.723
CC15MCLIP16 (C)0.3840.3370.2680.1050.5030.2380.699
C+GAP0.3990.3430.2870.1140.5310.2520.701
C+SPARO0.4370.3780.3170.1450.5790.2790.730
L400MCLIP32 (C)0.6170.5310.4820.2020.7190.4230.748
C+GAP0.6230.5370.4920.2120.7250.4400.732
C+SPARO0.6350.5520.5070.2310.7470.4590.770
", + "bbox": [ + 276, + 184, + 725, + 424 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Models. For all experiments in this section, we train CLIP models using the open-source OpenCLIP [34] project. We consider two model sizes: those with a ViT-B/16 visual backbone $(\\mathrm{CLIP}^{16})$ and those with a ViT-B/32 backbone $(\\mathrm{CLIP}^{32})$ . For models trained with SPARO, we replace the last transformer block of the base transformer backbones. For our chosen settings, this ensures that the resulting model size is comparable to the original model size. Further, since only the CLS token output from the image transformer and the EOS token output from the text transformer are used for the standard CLIP encodings, removing one transformer block ensures that we do not gain an unfair advantage by attending to positions from a layer that standard CLIP discards. Note that standard CLIP also retains the MLP as a part of its final transformer block, which is absent in the SPARO module. We also compare our encodings with those produced by CLIP with global average pooling (CLIP+GAP), which enjoys the advantage of being able to use all of the final layer outputs. When using SPARO, we use $L = V = 64$ when training on Conceptual Captions, and $L = 128$ , $V = 64$ when training on LAION-400M. We also provide results for CLIP with residual network (ResNet) [28] encoders in Appendix B.3.", + "bbox": [ + 212, + 453, + 787, + 712 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1 Zero-shot recognition, robustness, and compositionality", + "text_level": 1, + "bbox": [ + 215, + 736, + 718, + 752 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate the zero-shot classification accuracy of trained CLIP models on ImageNet [16] and a set of robustness benchmarks including ImageNet-V2 [53], ImageNet-Sketch [62], ImageNet-A [31], ImageNet-R [30], and ObjectNet [4]. Additionally, to evaluate the compositionality of our learned encodings in terms of objects, attributes, and relations, we test our trained models on SugarCrepe", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "SPARO: Selective Attention in Transformer Encodings for Vision", + "bbox": [ + 303, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a4e4ed91bc6eccbd8d1a29e4a177c6082607043557c7f549f34082801cd4e39f.jpg", + "image_caption": [ + "(a) $\\mathrm{CLIP^{16}}$ models trained on CC3M." + ], + "image_footnote": [], + "bbox": [ + 230, + 146, + 485, + 258 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b2ee92f9fafb6e58631375e2d655d3047e4e2676e91d796712b2e8a95510b0ea.jpg", + "image_caption": [ + "(b) $\\mathrm{CLIP^{16}}$ models trained on CC12M." + ], + "image_footnote": [], + "bbox": [ + 513, + 147, + 769, + 260 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/888dfbd4e6e928bde8154e4532868dd2a102f34b33c0dae68dfcbb083b2f6ff9.jpg", + "image_caption": [ + "(c) $\\mathrm{CLIP^{16}}$ models trained on CC15M." + ], + "image_footnote": [], + "bbox": [ + 228, + 277, + 483, + 388 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/68e09b5ef1d38e54acfe64c539467f560e083602c12c930b69d77cb815fe427c.jpg", + "image_caption": [ + "(d) $\\mathrm{CLIP^{32}}$ models trained on L400M.", + "Fig. 2: Relative differences of $\\mathrm{CLIP + SPARO}$ zero-shot accuracies when compared to $\\mathrm{CLIP + GAP}$ on the VTAB benchmark." + ], + "image_footnote": [], + "bbox": [ + 508, + 277, + 764, + 390 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "33. We report our results in Tab. 1, and provide more fine-grained SugarCrepe numbers in Appendix B.4. We find that using SPARO encodings outperforms CLIP and $\\mathrm{CLIP + GAP}$ in all settings considered. The improved results on the SugarCrepe benchmark suggests that SPARO encodings exhibit compositionality at a level that enable better binding of object, attribute, and relation properties than baseline encodings on average.", + "bbox": [ + 212, + 478, + 787, + 569 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, we also use the VTAB benchmark [70] to evaluate SPARO on a diverse set of datasets [5,13,14,19,22,23,29,35,36,38,45,46,48,61,63,68] that vary more significantly from the training data. We present the relative zero-shot classification accuracy improvements of CLIP + SPARO over CLIP + GAP in Fig. 2 and the corresponding absolute values in Appendix B.6. We see that SPARO outperforms standard CLIP encodings for a majority of VTAB tasks in all four settings.", + "bbox": [ + 212, + 571, + 789, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 Zero-shot image and text retrieval", + "text_level": 1, + "bbox": [ + 214, + 696, + 545, + 712 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To separately evaluate the quality of the learned image and text encodings, we consider zero-shot retrieval based on image and text on MS COCO [40], Flickr8k [32], and Flickr30k [65]. Our results are presented in Tab. 2. To compute Recall@5 for 'Text' we find the nearest 5 texts for an image, calculate the fraction of those texts that match the image, and average this metric over all the images (similar for Recall@5 for 'Image'). We find that SPARO encodings improve retrieval across all three datasets for both modalities.", + "bbox": [ + 212, + 733, + 787, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "A. Vani et al.", + "bbox": [ + 271, + 114, + 362, + 126 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/153a5b1fd04ad83517fdc9b006f242b45ba32e26b887c8ce8929545d3f20333f.jpg", + "table_caption": [ + "Table 2: Image and text zero-shot Recall@5 retrieval results on MS COCO, Flickr8k, and Flickr30k for CLIP models trained on Conceptual Captions and LAION-400M." + ], + "table_footnote": [], + "table_body": "
TrainModelImage R@5Text R@5
COCOF8kF30kCOCOF8kF30k
CC3MCLIP16 (C)0.2370.4000.3530.2940.4890.469
C+GAP0.2600.4300.4000.3120.5340.497
C+SPARO0.2890.4900.4610.3630.5870.581
CC12MCLIP16 (C)0.4740.6910.7150.6140.8370.817
C+GAP0.4990.7140.7250.6280.8190.847
C+SPARO0.5260.7370.7590.6700.8780.878
CC15MCLIP16 (C)0.5120.7380.7570.6360.8440.865
C+GAP0.5350.7610.7720.6590.8490.881
C+SPARO0.5570.7780.7930.6960.8980.905
L400MCLIP32 (C)0.5990.8220.8400.7650.9230.935
C+GAP0.6100.8220.8450.7680.9170.946
C+SPARO0.6160.8360.8540.7740.9350.950
", + "bbox": [ + 292, + 184, + 709, + 422 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/97ac6651a897e6ef7191874042cf4e7931c81498ba89b6c5105e28c892c71b65.jpg", + "table_caption": [ + "Table 3: Left: ImageNet linear probe accuracy for CLIP models trained on Conceptual Captions and LAION-400M. Right: ImageNet 20-nearest neighbors and linear probe accuracy for encodings trained with DINO." + ], + "table_footnote": [], + "table_body": "
ModelImageNet linear probe
Train: CC3MCC12MCC15ML400M
CLIP (C)0.4690.6300.6460.743
C+GAP0.5040.6490.6640.747
C+SPARO0.5610.7000.7110.755
", + "bbox": [ + 225, + 492, + 527, + 585 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/84060768415400d8141c958fc9a1c31c570697bbfc4e89e0703e1b55dceded76.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelImageNet eval
20-NN Linear probe
DINO0.6850.735
DINO+SPARO0.7060.757
", + "bbox": [ + 534, + 500, + 772, + 579 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3 Linear probe and DINO nearest neighbors classification", + "text_level": 1, + "bbox": [ + 214, + 614, + 720, + 630 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We evaluate the classification accuracy of linear probes trained on the encodings of our trained models. For CLIP models, we follow [50] where we take a subset of the training data to use as a validation set, and sweep for the best weight decay hyperparameter on it for each model setting. Additionally, we evaluate SPARO for self-supervised learning in vision using DINO [9]. We train DINO with a ViT-S/16 backbone on ImageNet [16] without labels with 8 GPUs for 100 epochs without changing any other hyperparameters. Here, we use SPARO with $L = V = 64$ , replacing the last transformer block of the ViT encoder. The trained DINO models are evaluated on $k$ -nearest neighbours with $k = 20$ and linear classification on ImageNet with labels. We report our results for CLIP and DINO in Tab. Our results for CLIP align with our previous results, with SPARO outperforming CLIP and CLIP+GAP in all linear probe settings. Furthermore, we see that a straightforward incorporation of SPARO in the DINO encoder", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "SPARO: Selective Attention in Transformer Encodings for Vision", + "bbox": [ + 303, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/8f1095f8c2b31f144cec3356f2ebe6a8126a1c033ebd93652c2a49eec72fef37.jpg", + "table_caption": [ + "Table 4: Effect of intervening to selecting only the top 32 slots based on zero-shot ImageNet validation accuracy on SugarCrepe compositionality for CC15M-trained $\\mathrm{CLIP}^{16} + \\mathrm{SPARO}$ and LAION-400M-trained $\\mathrm{CLIP}^{32} + \\mathrm{SPARO}$ ." + ], + "table_footnote": [], + "table_body": "
SugarCrepeCC15ML400M
Slots:100%50%100%
Object0.8960.9120.9270.929
ReplaceAttribute0.7980.7970.8220.854
Relation0.6660.7130.6780.705
SwapObject0.5930.6590.5850.646
Attribute0.7000.7030.7190.740
AddObject0.7790.8090.8660.857
Attribute0.6790.7370.7940.822
Average0.7300.7610.7700.793
", + "bbox": [ + 346, + 198, + 656, + 378 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "outperforms standard DINO for both nearest neighbors and linear classification. We provide additional experimental details in Appendices C.1 and C.2", + "bbox": [ + 212, + 407, + 782, + 440 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 Analysis", + "text_level": 1, + "bbox": [ + 215, + 464, + 334, + 482 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In this section, we provide insights into encodings learned by SPARO. We demonstrate the ability of using SPARO's slot structure to perform post-hoc selection of slots for improving downstream performance. We then show that without this structure, such selection is more prone to overfitting. Next, we validate the design choice of using single-head attention to model the notion of selective attention through controlled ablations. Finally, we visualize example concepts that CLIP models trained with SPARO learn to attend to.", + "bbox": [ + 212, + 498, + 787, + 604 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.1 Post-hoc concept selection", + "text_level": 1, + "bbox": [ + 215, + 630, + 480, + 645 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "SPARO enables downstream tasks access to a collection of concepts. However, not all concepts may be important for a specific downstream task. For example, attending to watermarks or caption style can be important when trying to determine the source of a sample, but is unnecessary and often harmful for most downstream tasks that focus on the elements in the scenes. The separation of concepts in the structure of a SPARO encoding makes it possible to manually intervene to select desirable slots.", + "bbox": [ + 212, + 657, + 787, + 763 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We demonstrate this ability with a simple heuristic: we perform concept selection by computing the zero-shot ImageNet validation accuracy of each slot separately, and pick only the top-performing 32 slots. In Tab.4, we show that such an intervention exposes improved compositionality when evaluating CLIP+SPARO models trained on CC15M and LAION-400M on SugarCrepe.", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "A. Vani et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/17068aede2c1618842827ddb0319f4389c44acae9e0a0b21c1c000fc9b5351f5.jpg", + "table_caption": [ + "Table 5: Effect of overfitting a partitioned mask for image encodings $\\mathbf{y}$ of frozen CC15M-trained ${\\mathrm{{CLIP}}}^{32} + \\mathrm{{GAP}}$ and ${\\mathrm{{CLIP}}}^{32} + \\mathrm{{SPARO}}$ models by training on SugarCrepe. We normalize the relative change on the evaluation benchmarks by the attained relative change on SugarCrepe. The highlighted rows indicate settings where the structure assumed for the mask aligns with SPARO's representation structure." + ], + "table_footnote": [], + "table_body": "
(# parts, Model Shape of y)TrainingEvaluation
SugarCrepeVTAB AvgImageNet 0-shot Acc
InitialMaskedInitialMaskedRelative changeInitialMaskedRelative change
(64, 64)+GAP(4096)0.6920.7320.2400.217-1.6890.3430.292-2.598
+SPARO(64, 64)0.7100.7640.2620.272+0.4970.3700.364-0.190
(64, 8)+GAP(512)0.6780.7120.2540.219-2.7940.3450.279-3.924
+SPARO(64, 8)0.6960.7410.2490.260+0.6880.3550.347-0.347
(512, 1)+GAP(512)0.6780.7710.2540.215-1.1190.3450.250-2.003
+SPARO(64, 8)0.6960.8090.2490.206-1.0590.3550.224-2.277
(4096, 1)+GAP(4096)0.6920.8150.2400.215-0.5960.3430.286-0.938
+SPARO(64, 64)0.7100.8810.2620.242-0.3160.3700.328-0.465
", + "bbox": [ + 215, + 224, + 787, + 426 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.2 Robustness of SPARO's slot structure", + "text_level": 1, + "bbox": [ + 215, + 455, + 568, + 470 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We further utilize the notion of post-hoc concept selection to study the robustness of SPARO's representational structure, and compare it with that of standard encodings. We set up the experiment as an inverse of Sec. 5.1 — we perform concept selection using the smaller SugarCrepe benchmark and evaluate the resulting encodings on real-world benchmarks. SugarCrepe contains 7,512 example pairs of positive and negative captions, differing only in compositional interpretation in 7 ways, over a set of 1,561 images. Compositionality is a shared high-level notion present in most real-world data, yet there is much information that exists in data beyond what is necessary to solve the SugarCrepe tasks. The small size of the SugarCrepe as a training dataset, together with the ubiquitous nature of compositionality in machine learning tasks, makes it an ideal candidate to overfit concept selection to for stress-testing the representational structure of SPARO by evaluating on other tasks.", + "bbox": [ + 212, + 489, + 787, + 686 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We train a global mask $\\pmb{m} \\in [0,1]^M$ of the image encodings $\\pmb{y} \\in \\mathbb{R}^M$ to maximize SugarCrepe performance, and evaluate the masked encodings $\\pmb{m} \\odot \\pmb{y}$ on VTAB tasks and ImageNet for zero-shot classification. We train masks for encodings of frozen $\\mathrm{CLIP}^{32} + \\mathrm{GAP}$ and $\\mathrm{CLIP}^{32} + \\mathrm{SPARO}$ models trained on CC15M with $M = 512$ and $M = 4096$ each. For the SPARO models, we use $L = 64$ and $V = M / L$ . For each of these settings, we consider two types of masking: per-dimension and per-slot. To compare the $\\mathrm{CLIP}^{32} + \\mathrm{GAP}$ encodings on per-slot masking, we consider their $L$ contiguous equal-sized partitions as the slot encodings. We provide full implementation details of our mask training setup in Appendix C.1.", + "bbox": [ + 212, + 688, + 789, + 842 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "SPARO: Selective Attention in Transformer Encodings for Vision", + "bbox": [ + 303, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/7e04754e3a4962c88a1f44c9746d6f10cf57fcfd55e846fccebbe823d3b7f3e5.jpg", + "table_caption": [ + "Table 6: Ablation experiments showing benefits of SPARO's attention bottleneck (Cross-attention is 'Separate-head') and cross-modal attention alignment (Slot-wise LayerNorm and projection are 'Yes' or 'N/A') with CC15M-trained CLIP, CLIP+GAP, CLIP+SPARO, and ablated variants of CLIP+SPARO that include multi-head cross-attention and the attentional pooler (AttPool) [39][66]." + ], + "table_footnote": [], + "table_body": "
ModelReplace last blockCross- attentionSlot-wise LayerNormSlot-wise projectionParamsFLOPS (×109)ImageNet zero-shot
CLIP32 (C)N/AN/AN/AN/A151M7.40.329
C+GAP151M7.40.345
C+AttPoolNoNo440M8.90.312
Multi-headYes306M8.80.344
YesNo440M8.90.338
Ablated C+SPARONoNoNo306M8.80.359
NoYes306M8.80.344
Separate-headNo172M8.70.358
YesNo306M8.80.344
Yes172M8.70.374
C+SPAROYesN/AN/A161M8.00.372
151M7.40.370
", + "bbox": [ + 217, + 224, + 784, + 479 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We expect the compositionality of the data to be encoded in concepts that are useful for real-world downstream benchmarks. Therefore, the aggressive selection of only the concepts that perform well on a compositionality benchmark should not significantly impact downstream performance. However, since SugarCrepe does not represent all possible compositional manipulations, we expect to see a drop in performance on some tasks due to overfitting. We present our results in Tab. 5. We find that when the mask structure is aligned with the slot structure of SPARO, overfitting concept selection improves the average performance on VTAB tasks, and incurs only a minimal drop in performance on ImageNet. Furthermore, even in settings where the masks are not trained to align with SPARO's slot structure, we see that SPARO remains more resilient to overfitting than standard encodings in a majority of settings.", + "bbox": [ + 212, + 508, + 787, + 690 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.3 Ablating the bottleneck and cross-modal alignment of SPARO", + "text_level": 1, + "bbox": [ + 214, + 710, + 769, + 726 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We evaluate the generalization benefits of the separate-head attention bottleneck and of aligning the SPARO attention heads between the modalities in CLIP. We set up SPARO on one end of the ablation spectrum and the cross-attention readout of attentional pooler (AttPool) [39,66] to the other. AttPool uses multi-head attention to attend to the backbone encoder's outputs using embedded learned queries whereas SPARO uses separate single-head attention mechanisms. Additionally, AttnPool performs layer normalization [2] followed by a linear projec", + "bbox": [ + 212, + 733, + 787, + 843 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "A. Vani et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "tion on the output of attention to produce the output encoding, whereas SPARO directly uses the output of separate-head attention as the encoding.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We train $\\mathrm{CLIP}^{32}$ , $\\mathrm{CLIP}^{32} + \\mathrm{GAP}$ , and variants of $\\mathrm{CLIP}^{32} + \\mathrm{SPARO}$ on CC15M with $L = 128$ , $V = 64$ and evaluate for ImageNet zero-shot accuracy and model size. We start by considering the choice of either separate-head or multi-head cross-attention over the backbone outputs. For each cross-attention setting, we apply layer normalization and linear projection operations to the attention module's outputs, similar to AttnPool. However, we additionally also consider slot-wise variants of these operations. Here, a slot-wise operation $f_{\\phi}$ is one that acts separately on each input slot $\\mathbf{y}_l$ using the same parameters $\\phi$ , and produces the slot-structured output $\\operatorname{concat}(f_{\\phi}(\\mathbf{y}_1), \\ldots, f_{\\phi}(\\mathbf{y}_L))$ . When applied to an input that is not slot-structured, we assume a slot structure by partitioning the input into $L$ contiguous equal-sized partitions. With SPARO, using both slot-wise layer normalization and slot-wise projection enables each read-out attention head from one modality to align with the corresponding head from the other without being influenced by the other heads.", + "bbox": [ + 212, + 176, + 787, + 388 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We present our results in Tab. 6, showing a clear advantage of SPARO over AttnPool. Additionally, both the choices of separate-head attention over multi-head and slot-wise operations over the alternatives result in improved generalization, reduced model size, and fewer floating-point operations per second (FLOPS). While the former supports the effectiveness of the separate-head attention bottleneck, the latter validates the benefits of the prior for shared separately-attendable concepts between the modalities enabled by SPARO.", + "bbox": [ + 212, + 388, + 787, + 494 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.4 Visualizations", + "text_level": 1, + "bbox": [ + 214, + 522, + 377, + 536 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "To conclude our analysis, we qualitatively visualize the concepts represented by SPARO using the $\\mathrm{CLIP}^{16} + \\mathrm{SPARO}$ model trained on CC15M from Sec. 4.1. A limitation of visualizing attended positions is that it only provides partial insight into the represented concept — we can tell where the information is taken from, but not how the information is used. Different SPARO slots can have similar attention maps for the same sample, but reveal different semantics when comparing patterns across other images. For instance, a slot that attends to animals (e.g. dogs, cats, elephants) and another that attends to transportation (e.g. trains, boats, cars) can both have the same attention mask over patches and words for a horse being used for transportation. Furthermore, not all SPARO concepts align between the two modalities for every sample due to information being present in one that is not possible to infer from the other.", + "bbox": [ + 212, + 551, + 787, + 733 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "To ease our interpretation of attended concepts, we choose concepts to visualize that have sharp attention in the text modality, attend to non-overlapping text tokens, and have cosine similarity more than 0.75 with the corresponding image slots. We present examples of these filtered concepts in Fig. 3. In this example, we see that SPARO is capable of attending to different concepts such as subject of the scene, the activity represented, and the surrounding location, for both vision and text. We provide additional visualizations in Appendix B.7", + "bbox": [ + 212, + 734, + 787, + 842 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "SPARO: Selective Attention in Transformer Encodings for Vision", + "bbox": [ + 303, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/6d5b27494b94184ef7d1aa9cae2e65816ac98d50b68948dfbb394ea0a967ead7.jpg", + "image_caption": [ + "A herd of cattle walking down a road being followed by a cowboy" + ], + "image_footnote": [], + "bbox": [ + 236, + 152, + 349, + 215 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/8cb54f7ee842e079c7bdff05d550db44d625d160abc073d1e0411e8a8db43392.jpg", + "image_caption": [ + "everal surfboards standing in a rowon the beach" + ], + "image_footnote": [], + "bbox": [ + 383, + 152, + 493, + 215 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/21d9354066660f9d4080f3b5ad4079f49cb8e16edbdf804d4714ddf10a136e76.jpg", + "image_caption": [ + "Two people riding horses on a rock path" + ], + "image_footnote": [], + "bbox": [ + 521, + 152, + 645, + 215 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/9906ef39be979a259828797000bd8b106e07656c9f3ec60edda9fa1854d03000.jpg", + "image_caption": [ + "A man sitting alone on a park bench in a park" + ], + "image_footnote": [], + "bbox": [ + 678, + 152, + 761, + 215 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/a27332458870ffcd532900de9e89d83dd88adb3042e7bc50ad652b5ade9bfcb4.jpg", + "image_caption": [ + "A herd of cattle walking down a road being followed by a cowboy" + ], + "image_footnote": [], + "bbox": [ + 236, + 229, + 349, + 292 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/89e821ecd4948ab4e7c058dd0c00e29ec8947994e091e4110b486c16a34cc90e.jpg", + "image_caption": [ + "Several surfboards standing in a row on the beach" + ], + "image_footnote": [], + "bbox": [ + 383, + 229, + 493, + 292 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/f8ed22736767397e70be98419c26e2ca1b18f33ba4a0d39e4776c6859bbbfce0.jpg", + "image_caption": [ + "Two people riding horses on a rock path" + ], + "image_footnote": [], + "bbox": [ + 521, + 229, + 643, + 292 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/e0545f5a8b26e7a38db18fe56e1d02f499267c9ba8948fe8a1b6ebc63c93affe.jpg", + "image_caption": [ + "A man sitting alone on a park bench in a park" + ], + "image_footnote": [], + "bbox": [ + 679, + 229, + 761, + 292 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/76ffc4830bb9abc250a8810893259e3ff5bdee2625a556360fac8f07105caf6b.jpg", + "image_caption": [ + "A herd of cattle walking down a road being followed by a cowboy", + "Fig. 3: Visualizing of the attended image and text positions for three SPARO slots (one per row) across four examples (one per column) from MS COCO. We surmise that the SPARO concepts from top to bottom represent the subject, activity, and location." + ], + "image_footnote": [], + "bbox": [ + 236, + 308, + 349, + 369 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/6a6f91537768480583e46fbc5fa75e1437235fbbfcd3d69027d0069e56974497.jpg", + "image_caption": [ + "Several surfboards standing in a row on the beach" + ], + "image_footnote": [], + "bbox": [ + 383, + 308, + 493, + 369 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/f63e2bc8b62bc21dc527adf5d5b2e2907e48aeaac7ced2e0fd7301541bc48c03.jpg", + "image_caption": [ + "Two people riding horses on a rock path" + ], + "image_footnote": [], + "bbox": [ + 521, + 308, + 643, + 369 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/b51dd07a94b93dc37c46b0ad11824a65df7d64e959c852ac62df0a0872e0e203.jpg", + "image_caption": [ + "A man sitting alone on a park bench in a park" + ], + "image_footnote": [], + "bbox": [ + 679, + 308, + 761, + 369 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Discussion", + "text_level": 1, + "bbox": [ + 215, + 467, + 354, + 482 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Humans utilize a prior for compositionality to coherently represent different subsets of salient aspects of the world. We can enrich our comprehension by selectively attending to new aspects or specialize to a task by filtering out distracting ones. We introduce SPARO with a goal of imparting a similar prior to transformers in representation learning frameworks by partitioning encodings into separately-attended concepts. Although we see evidence of disentanglement through positive results of post-hoc concept selection and visualization of attention maps, we do not impose any explicit independence or disentanglement constraints. In our representation learning frameworks, the pressure for learning distinct concepts arises from the need to explain the factors of variation through the training objective with only the features that can be accumulated by a single head of attention with embedded queries and limited dimensionality. However, auxiliary training objectives to impose stronger distributional conditions on the learned concepts can be explored as promising future directions. Beyond pretraining, we highlight that more sophisticated post-hoc concept selection approaches than explored in our work, like using human interaction or set cover algorithms, can further improve the downstream utility of SPARO.", + "bbox": [ + 212, + 497, + 787, + 755 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 215, + 777, + 393, + 795 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This research was funded by Sony and enabled in part by compute resources provided by the Digital Research Alliance of Canada, Mila, and Sony.", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "A. Vani et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 323, + 160 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Aydemir, G., Xie, W., Güney, F.: Self-supervised object-centric learning for videos. arXiv preprint arXiv:2310.06907 (2023)", + "2. Ba, J.L., Kiros, J.R., Hinton, G.E.: Layer normalization. arXiv preprint arXiv:1607.06450 (2016)", + "3. Bahdanau, D., Cho, K., Bengio, Y.: Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473 (2014)", + "4. Barbu, A., Mayo, D., Alverio, J., Luo, W., Wang, C., Gutfreund, D., Tenenbaum, J., Katz, B.: Objectnet: A large-scale bias-controlled dataset for pushing the limits of object recognition models. Advances in neural information processing systems 32 (2019)", + "5. Beattie, C., Leibo, J.Z., Teplyashin, D., Ward, T., Wainwright, M., Kuttler, H., Lefrancq, A., Green, S., Valdés, V., Sadik, A., et al.: Deepmind lab. arXiv preprint arXiv:1612.03801 (2016)", + "6. Boff, K.R., Kaufman, L., Thomas, J.P.: Handbook of perception and human performance, vol. 1. Wiley New York (1986)", + "7. Brady, J., Zimmermann, R.S., Sharma, Y., Schölkopf, B., von Kügelgen, J., Brendel, W.: Provably learning object-centric representations. arXiv preprint arXiv:2305.14229 (2023)", + "8. Burgess, C.P., Matthew, L., Watters, N., Kabra, R., Higgins, I., Botvinick, M., Lerchner, A.: Monet: Unsupervised scene decomposition and representation. arXiv preprint arXiv:1901.11390 (2019)", + "9. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021)", + "0. Chang, H.S., Sun, R.Y., Ricci, K., McCallum, A.: Multi-cls bert: An efficient alternative to traditional ensembling. arXiv preprint arXiv:2210.05043 (2022)", + "1. Changpinyo, S., Sharma, P., Ding, N., Soricut, R.: Conceptual 12M: Pushing web-scale image-text pre-training to recognize long-tail visual concepts. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3558-3568 (2021)", + "2. Chen, X., Xie, S., He, K.: An empirical study of training self-supervised vision transformers. arXiv preprint arXiv:2104.02057 (2021)", + "3. Cheng, G., Han, J., Lu, X.: Remote sensing image scene classification: Benchmark and state of the art. Proceedings of the IEEE 105(10), 1865-1883 (2017)", + "4. Cimpoi, M., Maji, S., Kokkinos, I., Mohamed, S., Vedaldi, A.: Describing textures in the wild. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3606-3613 (2014)", + "5. Colby, C.L., Goldberg, M.E.: Space and attention in parietal cortex. Annual review of neuroscience 22(1), 319-349 (1999)", + "6. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)", + "7. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: BERT: Pre-training of deep bidirectional transformers for language understanding. In: North American Chapter of the Association for Computational Linguistics (2019)", + "8. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + ], + "bbox": [ + 225, + 172, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "SPARO: Selective Attention in Transformer Encodings for Vision", + "bbox": [ + 303, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "19. Dugas, E., Jared, Jorge, Cukierski, W.: Diabetic retinopathy detection (2015), https://kaggle.com/competitions/diabetic-retinopathy-detection", + "20. Engelcke, M., Kosiorek, A.R., Jones, O.P., Posner, I.: Genesis: Generative scene inference and sampling with object-centric latent representations. arXiv preprint arXiv:1907.13052 (2019)", + "21. Eslami, S., Heess, N., Weber, T., Tassa, Y., Szepesvari, D., Hinton, G.E., et al.: Attend, infer, repeat: Fast scene understanding with generative models. Advances in neural information processing systems 29 (2016)", + "22. Fei-Fei, L., Fergus, R., Perona, P.: One-shot learning of object categories. IEEE transactions on pattern analysis and machine intelligence 28(4), 594-611 (2006)", + "23. Geiger, A., Lenz, P., Stiller, C., Urtasun, R.: Vision meets robotics: The kitti dataset. The International Journal of Robotics Research 32(11), 1231-1237 (2013)", + "24. Goyal, A., Didolkar, A., Ke, N.R., Blundell, C., Beaudoin, P., Heess, N., Mozer, M.C., Bengio, Y.: Neural production systems. Advances in Neural Information Processing Systems 34, 25673-25687 (2021)", + "25. Goyal, A., Lamb, A., Gampa, P., Beaudoin, P., Blundell, C., Levine, S., Bengio, Y., Mozer, M.C.: Factorizing declarative and procedural knowledge in structured, dynamical environments. In: International Conference on Learning Representations (2020)", + "26. Goyal, A., Lamb, A., Hoffmann, J., Sodhani, S., Levine, S., Bengio, Y., Scholkopf, B.: Recurrent independent mechanisms. arXiv preprint arXiv:1909.10893 (2019)", + "27. Greff, K., Kaufman, R.L., Kabra, R., Watters, N., Burgess, C., Zoran, D., Matthew, L., Botvinick, M., Lerchner, A.: Multi-object representation learning with iterative variational inference. In: International Conference on Machine Learning. pp. 2424-2433. PMLR (2019)", + "28. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)", + "29. Helber, P., Bischke, B., Dengel, A., Borth, D.: Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 12(7), 2217-2226 (2019)", + "30. Hendrycks, D., Basart, S., Mu, N., Kadavath, S., Wang, F., Dorundo, E., Desai, R., Zhu, T., Parajuli, S., Guo, M., et al.: The many faces of robustness: A critical analysis of out-of-distribution generalization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8340-8349 (2021)", + "31. Hendrycks, D., Zhao, K., Basart, S., Steinhardt, J., Song, D.: Natural adversarial examples. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15262-15271 (2021)", + "32. Hodosh, M., Young, P., Hockenmaier, J.: Framing image description as a ranking task: Data, models and evaluation metrics. Journal of Artificial Intelligence Research 47, 853-899 (2013)", + "33. Hsieh, C.Y., Zhang, J., Ma, Z., Kembhavi, A., Krishna, R.: Sugarcrepe: Fixing hackable benchmarks for vision-language compositionality. Advances in neural information processing systems (2023)", + "34. Ilharco, G., Wortsman, M., Wightman, R., Gordon, C., Carlini, N., Taori, R., Dave, A., Shankar, V., Namkoong, H., Miller, J., Hajishirzi, H., Farhadi, A., Schmidt, L.: OpenCLIP. https://github.com/mlfoundations/open Clip (Jul 2021). https://doi.org/10.5281/zenodo.5143773" + ], + "bbox": [ + 215, + 147, + 784, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "A. Vani et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "35. Johnson, J., Hariharan, B., van der Maaten, L., Fei-Fei, L., Lawrence Zitnick, C., Girshick, R.: Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2901-2910 (2017)", + "36. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images. Technical report (2009)", + "37. Lavoie, S., Tsirigotis, C., Schwarzer, M., Vani, A., Noukhovitch, M., Kawaguchi, K., Courville, A.: Simplicial embeddings in self-supervised learning and downstream classification. In: International Conference on Learning Representations (2023), https://openreview.net/forum?id=RWtGreRpovS", + "38. LeCun, Y., Huang, F.J., Bottou, L.: Learning methods for generic object recognition with invariance to pose and lighting. In: Proceedings of the 2004 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2004. CVPR 2004. vol. 2, pp. II-104. IEEE (2004)", + "39. Lee, J., Lee, Y., Kim, J., Kosiorek, A., Choi, S., Teh, Y.W.: Set transformer: A framework for attention-based permutation-invariant neural networks. In: International conference on machine learning. pp. 3744-3753. PMLR (2019)", + "40. Lin, T., Maire, M., Belongie, S.J., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, C.L.: Microsoft COCO: common objects in context. In: Fleet, D.J., Pajdla, T., Schiele, B., Tuytelaars, T. (eds.) Computer Vision - ECCV 2014 - 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V. Lecture Notes in Computer Science, vol. 8693, pp. 740-755. Springer (2014). https://doi.org/10.1007/978-3-319-10602-1_48, https://doi.org/10.1007/978-3-319-10602-1%5F48", + "41. Locatello, F., Weissenborn, D., Unterthiner, T., Mahendran, A., Heigold, G., Uszkoreit, J., Dosovitskiy, A., Kipf, T.: Object-centric learning with slot attention. Advances in Neural Information Processing Systems 33, 11525-11538 (2020)", + "42. Ma, Z., Hong, J., Gul, M.O., Gandhi, M., Gao, I., Krishna, R.: Crepe: Can vision-language foundation models reason compositionally? In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10910-10921 (2023)", + "43. Mansouri, A., Hartford, J., Zhang, Y., Bengio, Y.: Object centric architectures enable efficient causal representation learning. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=r9FsiXZxZt", + "44. Martinez, A.: Putting spatial attention on the map: timing and localization of stimulus selection processes in striate and extrastriate visual areas. Vision Res 41, 1437-1457 (2001)", + "45. Matthew, L., Higgins, I., Hassabis, D., Lerchner, A.: dsprites: Disentanglement testing sprites dataset. https://github.com/deepmind/dsprites-dataset/ (2017)", + "46. Nilsback, M.E., Zisserman, A.: Automated flower classification over a large number of classes. In: 2008 Sixth Indian conference on computer vision, graphics & image processing. pp. 722-729. IEEE (2008)", + "47. O'Connor, D.H., Fukui, M.M., Pinsk, M.A., Kastner, S.: Attention modulates responses in the human lateral geniculate nucleus. Nature neuroscience 5(11), 1203-1209 (2002)", + "48. Parkhi, O.M., Vedaldi, A., Zisserman, A., Jawahar, C.: Cats and dogs. In: 2012 IEEE conference on computer vision and pattern recognition. pp. 3498-3505. IEEE (2012)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "SPARO: Selective Attention in Transformer Encodings for Vision", + "bbox": [ + 303, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "49. Qian, R., Ding, S., Liu, X., Lin, D.: Semantics meets temporal correspondence: Self-supervised object-centric learning in videos. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16675-16687 (2023)", + "50. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)", + "51. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners. OpenAI blog 1(8), 9 (2019)", + "52. Ray, A., Radenovic, F., Dubey, A., Plummer, B.A., Krishna, R., Saenko, K.: Cola: A benchmark for compositional text-to-image retrieval. In: Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track (2023)", + "53. Recht, B., Roelofs, R., Schmidt, L., Shankar, V.: Do imagenet classifiers generalize to imagenet? In: International conference on machine learning. pp. 5389-5400. PMLR (2019)", + "54. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114 (2021)", + "55. Scott, W.A.: Cognitive complexity and cognitive flexibility. Sociometry pp. 405-414 (1962)", + "56. Seitzer, M., Horn, M., Zadaianchuk, A., Zietlow, D., Xiao, T., Simon-Gabriel, C.J., He, T., Zhang, Z., Scholkopf, B., Brox, T., et al.: Bridging the gap to real-world object-centric learning. arXiv preprint arXiv:2209.14860 (2022)", + "57. Sharma, P., Ding, N., Goodman, S., Soricut, R.: Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In: Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics, ACL 2018, Melbourne, Australia, July 15-20, 2018, Volume 1: Long Papers. pp. 2556-2565. Association for Computational Linguistics (2018)", + "58. Thrush, T., Jiang, R., Bartolo, M., Singh, A., Williams, A., Kiela, D., Ross, C.: Winoground: Probing vision and language models for visio-linguistic compositionality. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5238-5248 (2022)", + "59. Treisman, A.M., Gelade, G.: A feature-integration theory of attention. Cognitive psychology 12(1), 97-136 (1980)", + "60. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)", + "61. Veeling, B.S., Linmans, J., Winkens, J., Cohen, T., Welling, M.: Rotation equivariant cnns for digital pathology. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2018: 21st International Conference, Granada, Spain, September 16-20, 2018, Proceedings, Part II 11. pp. 210-218. Springer (2018)", + "62. Wang, H., Ge, S., Lipton, Z., Xing, E.P.: Learning robust global representations by penalizing local predictive power. In: Advances in Neural Information Processing Systems. pp. 10506-10518 (2019)", + "63. Xiao, J., Hays, J., Ehinger, K.A., Oliva, A., Torralba, A.: Sun database: Large-scale scene recognition from abbey to zoo. In: 2010 IEEE computer society conference on computer vision and pattern recognition. pp. 3485-3492. IEEE (2010)", + "64. Xu, L., Ouyang, W., Bennamoun, M., Boussaid, F., Xu, D.: Multi-class token transformer for weakly supervised semantic segmentation. In: Proceedings of the" + ], + "bbox": [ + 215, + 147, + 785, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "A. Vani et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4310-4319 (2022)", + "65. Young, P., Lai, A., Hodosh, M., Hockenmaier, J.: From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics 2, 67-78 (2014)", + "66. Yu, J., Wang, Z., Vasudevan, V., Yeung, L., Seyedhosseini, M., Wu, Y.: Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917 (2022)", + "67. Yuksekgonul, M., Bianchi, F., Kalluri, P., Jurafsky, D., Zou, J.: When and why vision-language models behave like bags-of-words, and what to do about it? In: The Eleventh International Conference on Learning Representations (2023)", + "68. Yuval, N.: Reading digits in natural images with unsupervised feature learning. In: Proceedings of the NIPS Workshop on Deep Learning and Unsupervised Feature Learning (2011)", + "69. Zadaianchuk, A., Kleindessner, M., Zhu, Y., Locatello, F., Brox, T.: Unsupervised semantic segmentation with self-supervised object-centric representations. arXiv preprint arXiv:2207.05027 (2022)", + "70. Zhai, X., Puigcerver, J., Kolesnikov, A., Ruyssen, P., Riquelme, C., Lucic, M., Djolonga, J., Pinto, A.S., Neumann, M., Dosovitskiy, A., et al.: A large-scale study of representation learning with the visual task adaptation benchmark. arXiv preprint arXiv:1910.04867 (2019)", + "71. Zhang, Y., Hare, J., Prugel-Bennett, A.: Deep set prediction networks. Advances in Neural Information Processing Systems 32 (2019)", + "72. Zhao, T., Zhang, T., Zhu, M., Shen, H., Lee, K., Lu, X., Yin, J.: Vl-checklist: Evaluating pre-trained vision-language models with objects, attributes and relations. arXiv preprint arXiv:2207.00221 (2022)" + ], + "bbox": [ + 215, + 146, + 785, + 507 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "SPARO: Selective Attention in Transformer Encodings for Vision", + "bbox": [ + 303, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/SPARO_ Selective Attention for Robust and Compositional Transformer Encodings for Vision/4bccfc91-c07f-416d-bd12-0ad527681e54_model.json b/2024/SPARO_ Selective Attention for Robust and Compositional Transformer Encodings for Vision/4bccfc91-c07f-416d-bd12-0ad527681e54_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7d66f2fcc3315370132275cdec43d654ea4eb999 --- /dev/null +++ b/2024/SPARO_ Selective Attention for Robust and Compositional Transformer Encodings for Vision/4bccfc91-c07f-416d-bd12-0ad527681e54_model.json @@ -0,0 +1,2702 @@ +[ + [ + { + "type": "image", + "bbox": [ + 0.227, + 0.144, + 0.271, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.274, + 0.152, + 0.777, + 0.172 + ], + "angle": 0, + "content": "SPARO: Selective Attention for Robust and" + }, + { + "type": "title", + "bbox": [ + 0.225, + 0.176, + 0.78, + 0.196 + ], + "angle": 0, + "content": "Compositional Transformer Encodings for Vision" + }, + { + "type": "text", + "bbox": [ + 0.236, + 0.221, + 0.768, + 0.255 + ], + "angle": 0, + "content": "Ankit Vani1*, Bac Nguyen2, Samuel Lavoie1, Ranjay Krishna3, and Aaron Courville1,4" + }, + { + "type": "text", + "bbox": [ + 0.397, + 0.265, + 0.606, + 0.279 + ], + "angle": 0, + "content": "1 Mila, Université de Montréal" + }, + { + "type": "text", + "bbox": [ + 0.465, + 0.28, + 0.537, + 0.293 + ], + "angle": 0, + "content": "2 Sony AI" + }, + { + "type": "text", + "bbox": [ + 0.273, + 0.293, + 0.73, + 0.307 + ], + "angle": 0, + "content": "3 University of Washington, Allen Institute for Artificial Intelligence" + }, + { + "type": "text", + "bbox": [ + 0.438, + 0.307, + 0.566, + 0.32 + ], + "angle": 0, + "content": "4 CIFAR AI Chair" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.362, + 0.744, + 0.668 + ], + "angle": 0, + "content": "Abstract. Selective attention helps us focus on task-relevant aspects in the constant flood of our sensory input. This constraint in our perception allows us to robustly generalize under distractions and to new compositions of perceivable concepts. Transformers employ a similar notion of attention in their architecture, but representation learning models with transformer backbones like CLIP and DINO often fail to demonstrate robustness and compositionality. We highlight a missing architectural prior: unlike human perception, transformer encodings do not separately attend over individual concepts. In response, we propose SPARO, a readout mechanism that partitions encodings into separately-attended slots, each produced by a single attention head. Using SPARO with CLIP imparts an inductive bias that the vision and text modalities are different views of a shared compositional world with the same corresponding concepts. Using SPARO, we demonstrate improvements on downstream recognition, robustness, retrieval, and compositionality benchmarks with CLIP (up to \\(+14\\%\\) for ImageNet, \\(+4\\%\\) for SugarCrepe), and on nearest neighbors and linear probe for ImageNet with DINO (\\(+3\\%\\) each). We also showcase a powerful ability to intervene and select individual SPARO concepts to further improve downstream task performance (up from \\(+4\\%\\) to \\(+9\\%\\) for SugarCrepe) and use this ability to study the robustness of SPARO's representation structure. Finally, we provide insights through ablation experiments and visualization of learned concepts." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.68, + 0.724, + 0.695 + ], + "angle": 0, + "content": "Keywords: Selective attention \\(\\cdot\\) Slot representations \\(\\cdot\\) Transformers" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.722, + 0.378, + 0.737 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.754, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Selective attention is an intrinsic property of human perception [15,44,47], enabling people to focus on task-relevant aspects of their surroundings while tuning out the rest [55]. For instance, it allows a driver to focus on traffic signals, road signs, and other vehicles, or an individual searching for their friend" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.551, + 0.84 + ], + "angle": 0, + "content": "* Correspondence at: ankit.vani@umontreal.ca." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "A. Vani et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.283 + ], + "angle": 0, + "content": "in a crowd to attend to details like people's heights and clothing. More broadly, it empowers us to comprehend the vast complexity of our world with limited mental resources [6] by constraining our perception to one concept at a time. When a task requires information for multiple concepts (e.g., looking at traffic signals and pedestrians), this constraint calls for sequentially attending to separate salient aspects of our stimuli [59]. The cognitive bottleneck of selective attention, with the ability to separately represent concepts from our environment, enables human perception to remain insensitive to irrelevant distractions and generalize to new compositions of perceivable concepts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.284, + 0.788, + 0.48 + ], + "angle": 0, + "content": "In machine learning, the notion of selective attention is party emulated by the attention mechanism 3, which is a core component of transformers 60 such as the vision transformer (ViT) 18. However, transformer encodings learnt using approaches like CLIP 50 and DINO 9 still struggle with robustness 30, 31, 62, 70 and compositional generalization 33, 42, 52, 58, 67, 72. For instance, despite CLIP emerging as the de facto backbone for many vision tasks, it can perform close to chance when evaluated for compositionality 42. While the attention mechanism equips these models with the ability to emphasize relevant aspects of their intermediate representations, it does not endow them with the means to produce encodings for separately-attended concepts. Consequently, transformer-based representation learning models lack an important prior from human cognition: that perception of a complex input can be broken down into perception of its salient concepts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.482, + 0.788, + 0.661 + ], + "angle": 0, + "content": "This limitation deprives downstream tasks easy access to task-relevant aspects of the data's underlying structure. In a task like ImageNet classification, where most images contain one primary object of interest (e.g., \"dog\"), information like relationships between objects and attributes of the background are irrelevant and can lead to overfitting if considered. However, these details can become crucial to attain high performance in other tasks like image retrieval (e.g., \"dog chasing a frisbee in a park\"). Similar to the generalization benefits of selective attention in human perception, encodings that allow easy separation of attended concepts in the data can generalize to a diverse set of downstream tasks. Without additional constraints, however, the emergence of such compositional structure in encodings is unlikely, especially in the common setting of training on noisy internet data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.662, + 0.788, + 0.814 + ], + "angle": 0, + "content": "We propose SPARO (Separate-head attention read-out) as an improved read-out mechanism for transformer encoders in vision inspired by selective attention. SPARO replaces the last transformer block to provide a mechanism for partitioning encodings into slots of separately-attended concepts. We design each slot encoding to be a low-dimensional result of a single-head attention operation, producing a bottleneck that encourages each slot to \"selectively attend.\" In this mechanism, producing multiple slots can be interpreted as selective attention over different concepts occurring in parallel. When training CLIP, SPARO imparts an inductive bias that both modalities are views onto a shared compositional world with the corresponding slots of both encoders representing its concepts." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.824, + 0.613, + 0.843 + ], + "angle": 0, + "content": "5 Source code: https://github.com/ankitkv/sparo-clip" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.305, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "SPARO: Selective Attention in Transformer Encodings for Vision" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "In contrast, standard CLIP merely embeds the encodings of both modalities in a shared vector space without imposing any additional structure." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.42 + ], + "angle": 0, + "content": "Training SPARO using CLIP and DINO exhibits improved generalization, robustness, and compositionality while retaining the same model size. Our experiments show that using CLIP with SPARO improves generalization for zero-shot recognition, robustness, retrieval, and compositionality (e.g., +14% for ImageNet, +4% for SugarCrepe, when trained with 15M examples; +3% each when trained with 400M examples). We also demonstrate improved linear probe performance of CLIP (+10% trained with 15M examples; +2% with 400M examples), as well as both linear probe and nearest neighbors performance of DINO (+3% each). We showcase how SPARO's representational structure enables the ability to manually intervene and select relevant concepts for downstream tasks, leading to improved compositionality and generalization (up from +4% to +9% for SugarCrepe trained with 15M examples; +3% to +6% with 400M examples). We then study the robustness of this structure, validate the choice of separate-head attention through ablations, and provide insights into the effect of the number and dimensionality of concepts. Finally, we provide visualizations for some of the concepts SPARO learns to attend to." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.444, + 0.383, + 0.459 + ], + "angle": 0, + "content": "2 Related work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.477, + 0.785, + 0.507 + ], + "angle": 0, + "content": "We situate our work amongst other transformer read-out mechanisms and slot representation learning methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.788, + 0.75 + ], + "angle": 0, + "content": "Transformer read-outs. For transformer [60] decoders with causal masking, the last hidden state, corresponding to the end of sequence (EOS) input token, is used as the output representation. BERT [17] proposed inserting a special CLS token before input sequences for transformer encoders where information from all positions is gathered to produce the encoding. Using the CLS token output is also popular for Vision transformer (ViT) [9, 12, 18] backbones, along with alternatives such as global average pooling [12, 64]. In the context of language, Multi-CLS BERT [10] argues that an input text sequence can have multiple facets, and that samples can be similar along one facet and dissimilar along another. The authors aim to represent these facets by adding multiple CLS tokens to the input sequence. Using separate attention heads for each SPARO slot can also encourage attending to different facets of the inputs, but without requiring expensive accumulations through each layer of the backbone. By keeping SPARO independent of the backbone, we also allow for its easy potential placement on top of frozen pre-trained models without needing to train for added CLS tokens." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "SPARO is most related to the attentional pooler [39,66] architecture, which uses multi-head attention with embedded learned queries to produce output encodings. However, SPARO employs a critical structural constraint of not mixing the information from separate attention heads, a design choice we justify empirically in Sec. 5.3. Slot attention [41] also produces slot encodings through the attention mechanism, but rather than having the input positions compete" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "A. Vani et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.208 + ], + "angle": 0, + "content": "for each slot, the slots compete for each input position. Slot attention also proposes an expensive iterative read-out mechanism where the role of each slot is determined by sampling from a per-sample Gaussian distribution, which can be limiting for large real-world datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.224, + 0.79, + 0.452 + ], + "angle": 0, + "content": "Slot representations. Object-centric slot representations have been studied extensively in the literature [7,8,20,21,24-27,41,43,71]. Often, these representations are expected to represent predefined objects and attributes, and evaluated in small-scale or synthetic settings. Recently, the use of slot attention on features produced by a learned encoder [1,49,56], and utilizing unsupervised saliency masks for object discovery [69] has enabled unsupervised learning of object-centric representations in larger scale settings. Related to these ideas, SPARO attends to features produced by a transformer using a single attention head per slot, and each head is capable of learning saliency maps without supervision [9]. However, we do not constrain our slots to correspond to predefined types of objects in the data, nor do we impose any independence conditions between the slots. Self-supervised simplicial embeddings (SEM) [37], with only a soft-discrete slot structure constraint, enable improved generalization on downstream tasks. SPARO replaces the separation of softmax application between SEM slots with a separation of underlying attention operations." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.477, + 0.331, + 0.492 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.51, + 0.789, + 0.572 + ], + "angle": 0, + "content": "Having motivated our approach, we now detail the design of SPARO. We start by introducing the relevant notation, then discuss the SPARO module. Finally, we discuss the addition of SPARO in CLIP and DINO, and the applicability of SPARO beyond these settings." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.595, + 0.337, + 0.609 + ], + "angle": 0, + "content": "3.1 Notation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.622, + 0.789, + 0.744 + ], + "angle": 0, + "content": "We envision SPARO as a special attention-based layer that modifies the top block of a transformer encoder [51,60] backbone (including the ViT [18]). Transformers take as input a sequence \\(\\pmb{x} = \\{\\pmb{x}_1,\\dots,\\pmb{x}_n\\}\\) of length \\(n\\), and produce an output state per position \\(\\{h_1,\\ldots ,h_n\\}\\), where \\(\\pmb{x}_i\\in \\mathbb{R}^{\\mathrm{inputdim}}\\), \\(\\pmb{h}_i\\in \\mathbb{R}^d\\), for any \\(i\\in [n]\\), with \\(d\\) the model width. Finally, a pooling operation reduces the sequence of output states to output encodings with a fixed size \\(\\pmb{y}\\in \\mathbb{R}^{M}\\). Examples of pooling operations include attentional pooling [39], global average pooling, and extracting the CLS or EOS token representations for images or text respectively." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.768, + 0.611, + 0.784 + ], + "angle": 0, + "content": "3.2 Separate-head attention read-out (SPARO)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.789, + 0.843 + ], + "angle": 0, + "content": "We replace the pooling operation of transformer encoders with a concatenation of outputs of \\( L \\) single-head attention mechanisms. SPARO acts on the transformer outputs \\( \\pmb{H} = \\{\\pmb{h}_1, \\dots, \\pmb{h}_n\\} \\in \\mathbb{R}^{n \\times d} \\) to produce the encoder output" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.305, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "SPARO: Selective Attention in Transformer Encodings for Vision" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.145, + 0.789, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.331, + 0.788, + 0.457 + ], + "angle": 0, + "content": "Fig. 1: Illustration of SPARO, a read-out mechanism that structures representations as collections of separately-attended concepts. Take a standard \\( N \\)-block transformer encoder (ViT here as an example), producing an encoding \\( y \\) through extraction of its CLS token output. We can replace the \\( N \\)th transformer block with the SPARO module (typically with equal or fewer parameters) to produce a SPARO encoding \\( y \\), which is a concatenation of \\( L \\) SPARO slots. Each SPARO slot \\( y_l \\) is produced through single-head attention over the backbone outputs using an embedded query \\( q_l \\). The value projection is a composition of slot-specific key projection parameterized by \\( K_l \\) and a slot-wise projection shared between all SPARO concepts parameterized by \\( W \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.485, + 0.788, + 0.518 + ], + "angle": 0, + "content": "\\(\\pmb{y} = \\mathrm{concat}(\\pmb{y}_1, \\dots, \\pmb{y}_L) \\in \\mathbb{R}^{LV}\\), where SPARO slot \\(\\pmb{y}_l \\in \\mathbb{R}^V\\), \\(l \\in [L]\\). We illustrate the architecture of SPARO in Fig. 1. Each SPARO slot is produced as:" + }, + { + "type": "equation", + "bbox": [ + 0.37, + 0.538, + 0.787, + 0.573 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {y} _ {l} = \\boldsymbol {W} \\boldsymbol {K} _ {l} \\boldsymbol {H} ^ {\\intercal} \\text {s o f t m a x} \\left(\\frac {\\boldsymbol {H} \\boldsymbol {K} _ {l} ^ {\\intercal} \\boldsymbol {q} _ {l}}{\\sqrt {D}}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.788, + 0.659 + ], + "angle": 0, + "content": "Here, \\( \\pmb{K}_l \\in \\mathbb{R}^{D \\times d} \\) is a learned key projection weight, and \\( \\pmb{q}_l \\in \\mathbb{R}^D \\) is a learned query embedding for attention. We parameterize the value projection as \\( \\pmb{W}\\pmb{K}_l \\) where \\( \\pmb{W} \\in \\mathbb{R}^{V \\times D} \\) is learned parameter shared between all \\( L \\) slots. Decomposing the value projection in this manner helps reduce the number of model parameters, allowing a larger choice of \\( L \\) with fixed resources." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.66, + 0.788, + 0.764 + ], + "angle": 0, + "content": "Generally, we pick \\( V \\) and \\( D \\) to be significantly smaller than the transformer width. In practice, we use the typical value of \\( D = 64 \\) used in standard multihead attention modules, and set \\( V \\) to the same value in our experiments. Therefore, each SPARO slot has limited expressivity and representational capacity. The bottleneck underpinning each SPARO slot is a result of these constraints and the mechanistic bottleneck of a single attention mechanism instantiating competition between its input features." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "SPARO can be thought of as multi-head attention [60] where the queries are learned embeddings, key and value projection weights are shared, and the output projection weight is block-diagonal with each block containing the same parameters. Assuming \\( L = V = D = \\sqrt{d} \\), SPARO requires a total of \\( d^2 + 2d \\) parameters, compared to multi-head attention's \\( 4d^2 \\)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "A. Vani et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.411, + 0.16 + ], + "angle": 0, + "content": "3.3 CLIP with SPARO" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.168, + 0.788, + 0.261 + ], + "angle": 0, + "content": "We add SPARO to both the image and text encoders of CLIP [50], setting the same values of \\( L \\) and \\( V \\) in both encoders. Each SPARO slot is \\( \\ell_2 \\)-normalized separately, and we divide their concatenation by \\( \\sqrt{L} \\) to produce the global \\( \\ell_2 \\)-normalized encoding. With this scheme, the cosine similarity between the image encoding \\( \\mathbf{y}^{\\mathrm{i}} \\) and the text encoding \\( \\mathbf{y}^{\\mathrm{t}} \\) becomes the expected cosine similarity between corresponding slot encodings:" + }, + { + "type": "equation", + "bbox": [ + 0.236, + 0.267, + 0.786, + 0.309 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {y} ^ {\\mathrm {i}} ^ {\\top} \\boldsymbol {y} ^ {\\mathrm {t}} = \\sum_ {l = 1} ^ {L} \\| \\boldsymbol {y} _ {l} ^ {\\mathrm {i}} \\| \\| \\boldsymbol {y} _ {l} ^ {\\mathrm {t}} \\| \\frac {\\boldsymbol {y} _ {l} ^ {\\mathrm {i}} {} ^ {\\top} \\boldsymbol {y} _ {l} ^ {\\mathrm {t}}}{\\| \\boldsymbol {y} _ {l} ^ {\\mathrm {i}} \\| \\| \\boldsymbol {y} _ {l} ^ {\\mathrm {t}} \\|} = \\sum_ {l = 1} ^ {L} \\frac {1}{L} \\operatorname {s i m} (\\boldsymbol {y} _ {l} ^ {\\mathrm {i}}, \\boldsymbol {y} _ {l} ^ {\\mathrm {t}}) = \\underset {l \\sim [ L ]} {\\mathbb {E}} \\left[ \\operatorname {s i m} (\\boldsymbol {y} _ {l} ^ {\\mathrm {i}}, \\boldsymbol {y} _ {l} ^ {\\mathrm {t}}) \\right]. \\quad (2)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.315, + 0.788, + 0.438 + ], + "angle": 0, + "content": "We train the model using the standard CLIP loss which maximizes \\(\\mathrm{sim}(\\pmb{y}^{\\mathrm{i}},\\pmb{y}^{\\mathrm{t}})\\) of encodings \\(\\pmb{y}^{\\mathrm{i}}\\) and \\(\\pmb{y}^{\\mathrm{t}}\\) for aligned image-text pairs and minimizes it across all other pairings. With SPARO's separate-head attention bottleneck, maximizing \\(\\mathbb{E}_{l\\sim [L]}\\left[\\mathrm{sim}(\\pmb{y}_l^{\\mathrm{i}},\\pmb{y}_l^{\\mathrm{t}})\\right]\\) for aligned pairs across the training dataset encourages both modalities to learn similar attention semantics per slot \\(l\\in [L]\\). Consequently, SPARO imparts a prior that both modalities encode a shared world with \\(L\\) selectively attendable concepts, and similarities between inputs can be expressed as their average similarity for these concepts." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.456, + 0.416, + 0.47 + ], + "angle": 0, + "content": "3.4 DINO with SPARO" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.477, + 0.788, + 0.614 + ], + "angle": 0, + "content": "The online and momentum encoders of DINO [9] are each comprised of a ViT backbone which produces the encoding, and a DINO head that transforms it into a categorical distribution for distillation. Unlike CLIP, these encoders have an inherent bias for learning similar encoding functions due to one being an exponentially moving average of the other. However, in standard DINO, there are no constraints for selective attention in the encoding structure. We add SPARO to DINO to impart the prior for representing separately-attendable concepts in its encodings. Concretely, we replace the final transformer block of the ViT backbone with the SPARO, but do not modify the DINO head." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.634, + 0.325, + 0.649 + ], + "angle": 0, + "content": "4 Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.662, + 0.787, + 0.708 + ], + "angle": 0, + "content": "In Secs. 4 and 5, we evaluate our method on a variety of downstream tasks, analyze SPARO's representation structure with the ability to intervene on selected slots, perform ablations to support our design, and visualize the learned concepts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.709, + 0.787, + 0.772 + ], + "angle": 0, + "content": "In this section, we validate that partitioning the representation structure as a collection of separately-attended concepts leads to improved generalization for downstream recognition, robustness, compositionality, and retrieval, using SPARO with CLIP [50] in Secs. [4.1] to [4.3] and DINO [9] in Sec. [4.3]." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Datasets. We train our CLIP models on one of Conceptual Captions 3M (CC3M) [57], Conceptual Captions 12M (CC12M) [11], a combination of CC3M and CC12M (CC15M), or LAION-400M (L400M) [54]. Appendix C provides the statistics for our training datasets." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.305, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "SPARO: Selective Attention in Transformer Encodings for Vision" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.785, + 0.174 + ], + "angle": 0, + "content": "Table 1: Zero-shot recognition, robustness, and SugarCrepe compositionality for CLIP models trained on Conceptual Captions and LAION-400M." + }, + { + "type": "table", + "bbox": [ + 0.277, + 0.185, + 0.727, + 0.425 + ], + "angle": 0, + "content": "
TrainModelImageNet-Object NetSugar Crepe
V1V2SketchAR
CC3MCLIP16 (C)0.1410.1220.0680.0330.1770.0800.611
C+GAP0.1560.1340.0690.0330.1870.0810.616
C+SPARO0.1700.1400.0880.0350.2210.0980.625
CC12MCLIP16 (C)0.3610.3110.2490.0910.4670.2180.697
C+GAP0.3820.3300.2620.1010.5010.2410.695
C+SPARO0.4060.3500.2980.1130.5590.2680.723
CC15MCLIP16 (C)0.3840.3370.2680.1050.5030.2380.699
C+GAP0.3990.3430.2870.1140.5310.2520.701
C+SPARO0.4370.3780.3170.1450.5790.2790.730
L400MCLIP32 (C)0.6170.5310.4820.2020.7190.4230.748
C+GAP0.6230.5370.4920.2120.7250.4400.732
C+SPARO0.6350.5520.5070.2310.7470.4590.770
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.454, + 0.788, + 0.713 + ], + "angle": 0, + "content": "Models. For all experiments in this section, we train CLIP models using the open-source OpenCLIP [34] project. We consider two model sizes: those with a ViT-B/16 visual backbone \\((\\mathrm{CLIP}^{16})\\) and those with a ViT-B/32 backbone \\((\\mathrm{CLIP}^{32})\\). For models trained with SPARO, we replace the last transformer block of the base transformer backbones. For our chosen settings, this ensures that the resulting model size is comparable to the original model size. Further, since only the CLS token output from the image transformer and the EOS token output from the text transformer are used for the standard CLIP encodings, removing one transformer block ensures that we do not gain an unfair advantage by attending to positions from a layer that standard CLIP discards. Note that standard CLIP also retains the MLP as a part of its final transformer block, which is absent in the SPARO module. We also compare our encodings with those produced by CLIP with global average pooling (CLIP+GAP), which enjoys the advantage of being able to use all of the final layer outputs. When using SPARO, we use \\(L = V = 64\\) when training on Conceptual Captions, and \\(L = 128\\), \\(V = 64\\) when training on LAION-400M. We also provide results for CLIP with residual network (ResNet) [28] encoders in Appendix B.3." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.737, + 0.719, + 0.753 + ], + "angle": 0, + "content": "4.1 Zero-shot recognition, robustness, and compositionality" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We evaluate the zero-shot classification accuracy of trained CLIP models on ImageNet [16] and a set of robustness benchmarks including ImageNet-V2 [53], ImageNet-Sketch [62], ImageNet-A [31], ImageNet-R [30], and ObjectNet [4]. Additionally, to evaluate the compositionality of our learned encodings in terms of objects, attributes, and relations, we test our trained models on SugarCrepe" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.127 + ], + "angle": 0, + "content": "A. Vani et al." + }, + { + "type": "image", + "bbox": [ + 0.232, + 0.147, + 0.486, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.253, + 0.262, + 0.472, + 0.274 + ], + "angle": 0, + "content": "(a) \\(\\mathrm{CLIP^{16}}\\) models trained on CC3M." + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.148, + 0.771, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.532, + 0.262, + 0.758, + 0.275 + ], + "angle": 0, + "content": "(b) \\(\\mathrm{CLIP^{16}}\\) models trained on CC12M." + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.278, + 0.484, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.246, + 0.392, + 0.472, + 0.404 + ], + "angle": 0, + "content": "(c) \\(\\mathrm{CLIP^{16}}\\) models trained on CC15M." + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.278, + 0.765, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.392, + 0.751, + 0.405 + ], + "angle": 0, + "content": "(d) \\(\\mathrm{CLIP^{32}}\\) models trained on L400M." + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.416, + 0.788, + 0.445 + ], + "angle": 0, + "content": "Fig. 2: Relative differences of \\(\\mathrm{CLIP + SPARO}\\) zero-shot accuracies when compared to \\(\\mathrm{CLIP + GAP}\\) on the VTAB benchmark." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.479, + 0.789, + 0.57 + ], + "angle": 0, + "content": "33. We report our results in Tab. 1, and provide more fine-grained SugarCrepe numbers in Appendix B.4. We find that using SPARO encodings outperforms CLIP and \\(\\mathrm{CLIP + GAP}\\) in all settings considered. The improved results on the SugarCrepe benchmark suggests that SPARO encodings exhibit compositionality at a level that enable better binding of object, attribute, and relation properties than baseline encodings on average." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.572, + 0.79, + 0.664 + ], + "angle": 0, + "content": "Finally, we also use the VTAB benchmark [70] to evaluate SPARO on a diverse set of datasets [5,13,14,19,22,23,29,35,36,38,45,46,48,61,63,68] that vary more significantly from the training data. We present the relative zero-shot classification accuracy improvements of CLIP + SPARO over CLIP + GAP in Fig. 2 and the corresponding absolute values in Appendix B.6. We see that SPARO outperforms standard CLIP encodings for a majority of VTAB tasks in all four settings." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.698, + 0.547, + 0.713 + ], + "angle": 0, + "content": "4.2 Zero-shot image and text retrieval" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.789, + 0.84 + ], + "angle": 0, + "content": "To separately evaluate the quality of the learned image and text encodings, we consider zero-shot retrieval based on image and text on MS COCO [40], Flickr8k [32], and Flickr30k [65]. Our results are presented in Tab. 2. To compute Recall@5 for 'Text' we find the nearest 5 texts for an image, calculate the fraction of those texts that match the image, and average this metric over all the images (similar for Recall@5 for 'Image'). We find that SPARO encodings improve retrieval across all three datasets for both modalities." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.305, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "SPARO: Selective Attention in Transformer Encodings for Vision" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.784, + 0.174 + ], + "angle": 0, + "content": "Table 2: Image and text zero-shot Recall@5 retrieval results on MS COCO, Flickr8k, and Flickr30k for CLIP models trained on Conceptual Captions and LAION-400M." + }, + { + "type": "table", + "bbox": [ + 0.294, + 0.185, + 0.71, + 0.424 + ], + "angle": 0, + "content": "
TrainModelImage R@5Text R@5
COCOF8kF30kCOCOF8kF30k
CC3MCLIP16 (C)0.2370.4000.3530.2940.4890.469
C+GAP0.2600.4300.4000.3120.5340.497
C+SPARO0.2890.4900.4610.3630.5870.581
CC12MCLIP16 (C)0.4740.6910.7150.6140.8370.817
C+GAP0.4990.7140.7250.6280.8190.847
C+SPARO0.5260.7370.7590.6700.8780.878
CC15MCLIP16 (C)0.5120.7380.7570.6360.8440.865
C+GAP0.5350.7610.7720.6590.8490.881
C+SPARO0.5570.7780.7930.6960.8980.905
L400MCLIP32 (C)0.5990.8220.8400.7650.9230.935
C+GAP0.6100.8220.8450.7680.9170.946
C+SPARO0.6160.8360.8540.7740.9350.950
" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.439, + 0.787, + 0.482 + ], + "angle": 0, + "content": "Table 3: Left: ImageNet linear probe accuracy for CLIP models trained on Conceptual Captions and LAION-400M. Right: ImageNet 20-nearest neighbors and linear probe accuracy for encodings trained with DINO." + }, + { + "type": "table", + "bbox": [ + 0.226, + 0.493, + 0.528, + 0.587 + ], + "angle": 0, + "content": "
ModelImageNet linear probe
Train: CC3MCC12MCC15ML400M
CLIP (C)0.4690.6300.6460.743
C+GAP0.5040.6490.6640.747
C+SPARO0.5610.7000.7110.755
" + }, + { + "type": "table", + "bbox": [ + 0.535, + 0.5, + 0.773, + 0.58 + ], + "angle": 0, + "content": "
ModelImageNet eval
20-NN Linear probe
DINO0.6850.735
DINO+SPARO0.7060.757
" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.616, + 0.722, + 0.631 + ], + "angle": 0, + "content": "4.3 Linear probe and DINO nearest neighbors classification" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We evaluate the classification accuracy of linear probes trained on the encodings of our trained models. For CLIP models, we follow [50] where we take a subset of the training data to use as a validation set, and sweep for the best weight decay hyperparameter on it for each model setting. Additionally, we evaluate SPARO for self-supervised learning in vision using DINO [9]. We train DINO with a ViT-S/16 backbone on ImageNet [16] without labels with 8 GPUs for 100 epochs without changing any other hyperparameters. Here, we use SPARO with \\( L = V = 64 \\), replacing the last transformer block of the ViT encoder. The trained DINO models are evaluated on \\( k \\)-nearest neighbours with \\( k = 20 \\) and linear classification on ImageNet with labels. We report our results for CLIP and DINO in Tab. Our results for CLIP align with our previous results, with SPARO outperforming CLIP and CLIP+GAP in all linear probe settings. Furthermore, we see that a straightforward incorporation of SPARO in the DINO encoder" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "A. Vani et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 4: Effect of intervening to selecting only the top 32 slots based on zero-shot ImageNet validation accuracy on SugarCrepe compositionality for CC15M-trained \\(\\mathrm{CLIP}^{16} + \\mathrm{SPARO}\\) and LAION-400M-trained \\(\\mathrm{CLIP}^{32} + \\mathrm{SPARO}\\)." + }, + { + "type": "table", + "bbox": [ + 0.348, + 0.199, + 0.658, + 0.38 + ], + "angle": 0, + "content": "
SugarCrepeCC15ML400M
Slots:100%50%100%
Object0.8960.9120.9270.929
ReplaceAttribute0.7980.7970.8220.854
Relation0.6660.7130.6780.705
SwapObject0.5930.6590.5850.646
Attribute0.7000.7030.7190.740
AddObject0.7790.8090.8660.857
Attribute0.6790.7370.7940.822
Average0.7300.7610.7700.793
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.409, + 0.784, + 0.441 + ], + "angle": 0, + "content": "outperforms standard DINO for both nearest neighbors and linear classification. We provide additional experimental details in Appendices C.1 and C.2" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.465, + 0.335, + 0.483 + ], + "angle": 0, + "content": "5 Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.499, + 0.788, + 0.606 + ], + "angle": 0, + "content": "In this section, we provide insights into encodings learned by SPARO. We demonstrate the ability of using SPARO's slot structure to perform post-hoc selection of slots for improving downstream performance. We then show that without this structure, such selection is more prone to overfitting. Next, we validate the design choice of using single-head attention to model the notion of selective attention through controlled ablations. Finally, we visualize example concepts that CLIP models trained with SPARO learn to attend to." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.631, + 0.482, + 0.646 + ], + "angle": 0, + "content": "5.1 Post-hoc concept selection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.764 + ], + "angle": 0, + "content": "SPARO enables downstream tasks access to a collection of concepts. However, not all concepts may be important for a specific downstream task. For example, attending to watermarks or caption style can be important when trying to determine the source of a sample, but is unnecessary and often harmful for most downstream tasks that focus on the elements in the scenes. The separation of concepts in the structure of a SPARO encoding makes it possible to manually intervene to select desirable slots." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We demonstrate this ability with a simple heuristic: we perform concept selection by computing the zero-shot ImageNet validation accuracy of each slot separately, and pick only the top-performing 32 slots. In Tab.4, we show that such an intervention exposes improved compositionality when evaluating CLIP+SPARO models trained on CC15M and LAION-400M on SugarCrepe." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.305, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "SPARO: Selective Attention in Transformer Encodings for Vision" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 5: Effect of overfitting a partitioned mask for image encodings \\( \\mathbf{y} \\) of frozen CC15M-trained \\( {\\mathrm{{CLIP}}}^{32} + \\mathrm{{GAP}} \\) and \\( {\\mathrm{{CLIP}}}^{32} + \\mathrm{{SPARO}} \\) models by training on SugarCrepe. We normalize the relative change on the evaluation benchmarks by the attained relative change on SugarCrepe. The highlighted rows indicate settings where the structure assumed for the mask aligns with SPARO's representation structure." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.226, + 0.789, + 0.427 + ], + "angle": 0, + "content": "
(# parts, Model Shape of y)TrainingEvaluation
SugarCrepeVTAB AvgImageNet 0-shot Acc
InitialMaskedInitialMaskedRelative changeInitialMaskedRelative change
(64, 64)+GAP(4096)0.6920.7320.2400.217-1.6890.3430.292-2.598
+SPARO(64, 64)0.7100.7640.2620.272+0.4970.3700.364-0.190
(64, 8)+GAP(512)0.6780.7120.2540.219-2.7940.3450.279-3.924
+SPARO(64, 8)0.6960.7410.2490.260+0.6880.3550.347-0.347
(512, 1)+GAP(512)0.6780.7710.2540.215-1.1190.3450.250-2.003
+SPARO(64, 8)0.6960.8090.2490.206-1.0590.3550.224-2.277
(4096, 1)+GAP(4096)0.6920.8150.2400.215-0.5960.3430.286-0.938
+SPARO(64, 64)0.7100.8810.2620.242-0.3160.3700.328-0.465
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.457, + 0.57, + 0.472 + ], + "angle": 0, + "content": "5.2 Robustness of SPARO's slot structure" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.491, + 0.789, + 0.687 + ], + "angle": 0, + "content": "We further utilize the notion of post-hoc concept selection to study the robustness of SPARO's representational structure, and compare it with that of standard encodings. We set up the experiment as an inverse of Sec. 5.1 — we perform concept selection using the smaller SugarCrepe benchmark and evaluate the resulting encodings on real-world benchmarks. SugarCrepe contains 7,512 example pairs of positive and negative captions, differing only in compositional interpretation in 7 ways, over a set of 1,561 images. Compositionality is a shared high-level notion present in most real-world data, yet there is much information that exists in data beyond what is necessary to solve the SugarCrepe tasks. The small size of the SugarCrepe as a training dataset, together with the ubiquitous nature of compositionality in machine learning tasks, makes it an ideal candidate to overfit concept selection to for stress-testing the representational structure of SPARO by evaluating on other tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.79, + 0.843 + ], + "angle": 0, + "content": "We train a global mask \\( \\pmb{m} \\in [0,1]^M \\) of the image encodings \\( \\pmb{y} \\in \\mathbb{R}^M \\) to maximize SugarCrepe performance, and evaluate the masked encodings \\( \\pmb{m} \\odot \\pmb{y} \\) on VTAB tasks and ImageNet for zero-shot classification. We train masks for encodings of frozen \\( \\mathrm{CLIP}^{32} + \\mathrm{GAP} \\) and \\( \\mathrm{CLIP}^{32} + \\mathrm{SPARO} \\) models trained on CC15M with \\( M = 512 \\) and \\( M = 4096 \\) each. For the SPARO models, we use \\( L = 64 \\) and \\( V = M / L \\). For each of these settings, we consider two types of masking: per-dimension and per-slot. To compare the \\( \\mathrm{CLIP}^{32} + \\mathrm{GAP} \\) encodings on per-slot masking, we consider their \\( L \\) contiguous equal-sized partitions as the slot encodings. We provide full implementation details of our mask training setup in Appendix C.1." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "A. Vani et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.219 + ], + "angle": 0, + "content": "Table 6: Ablation experiments showing benefits of SPARO's attention bottleneck (Cross-attention is 'Separate-head') and cross-modal attention alignment (Slot-wise LayerNorm and projection are 'Yes' or 'N/A') with CC15M-trained CLIP, CLIP+GAP, CLIP+SPARO, and ablated variants of CLIP+SPARO that include multi-head cross-attention and the attentional pooler (AttPool) [39][66]." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.226, + 0.785, + 0.481 + ], + "angle": 0, + "content": "
ModelReplace last blockCross- attentionSlot-wise LayerNormSlot-wise projectionParamsFLOPS (×109)ImageNet zero-shot
CLIP32 (C)N/AN/AN/AN/A151M7.40.329
C+GAP151M7.40.345
C+AttPoolNoNo440M8.90.312
Multi-headYes306M8.80.344
YesNo440M8.90.338
Ablated C+SPARONoNoNo306M8.80.359
NoYes306M8.80.344
Separate-headNo172M8.70.358
YesNo306M8.80.344
Yes172M8.70.374
C+SPAROYesN/AN/A161M8.00.372
151M7.40.370
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.509, + 0.789, + 0.691 + ], + "angle": 0, + "content": "We expect the compositionality of the data to be encoded in concepts that are useful for real-world downstream benchmarks. Therefore, the aggressive selection of only the concepts that perform well on a compositionality benchmark should not significantly impact downstream performance. However, since SugarCrepe does not represent all possible compositional manipulations, we expect to see a drop in performance on some tasks due to overfitting. We present our results in Tab. 5. We find that when the mask structure is aligned with the slot structure of SPARO, overfitting concept selection improves the average performance on VTAB tasks, and incurs only a minimal drop in performance on ImageNet. Furthermore, even in settings where the masks are not trained to align with SPARO's slot structure, we see that SPARO remains more resilient to overfitting than standard encodings in a majority of settings." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.711, + 0.77, + 0.727 + ], + "angle": 0, + "content": "5.3 Ablating the bottleneck and cross-modal alignment of SPARO" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.844 + ], + "angle": 0, + "content": "We evaluate the generalization benefits of the separate-head attention bottleneck and of aligning the SPARO attention heads between the modalities in CLIP. We set up SPARO on one end of the ablation spectrum and the cross-attention readout of attentional pooler (AttPool) [39,66] to the other. AttPool uses multi-head attention to attend to the backbone encoder's outputs using embedded learned queries whereas SPARO uses separate single-head attention mechanisms. Additionally, AttnPool performs layer normalization [2] followed by a linear projec" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.305, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "SPARO: Selective Attention in Transformer Encodings for Vision" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "tion on the output of attention to produce the output encoding, whereas SPARO directly uses the output of separate-head attention as the encoding." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.389 + ], + "angle": 0, + "content": "We train \\(\\mathrm{CLIP}^{32}\\), \\(\\mathrm{CLIP}^{32} + \\mathrm{GAP}\\), and variants of \\(\\mathrm{CLIP}^{32} + \\mathrm{SPARO}\\) on CC15M with \\(L = 128\\), \\(V = 64\\) and evaluate for ImageNet zero-shot accuracy and model size. We start by considering the choice of either separate-head or multi-head cross-attention over the backbone outputs. For each cross-attention setting, we apply layer normalization and linear projection operations to the attention module's outputs, similar to AttnPool. However, we additionally also consider slot-wise variants of these operations. Here, a slot-wise operation \\(f_{\\phi}\\) is one that acts separately on each input slot \\(\\mathbf{y}_l\\) using the same parameters \\(\\phi\\), and produces the slot-structured output \\(\\operatorname{concat}(f_{\\phi}(\\mathbf{y}_1), \\ldots, f_{\\phi}(\\mathbf{y}_L))\\). When applied to an input that is not slot-structured, we assume a slot structure by partitioning the input into \\(L\\) contiguous equal-sized partitions. With SPARO, using both slot-wise layer normalization and slot-wise projection enables each read-out attention head from one modality to align with the corresponding head from the other without being influenced by the other heads." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.39, + 0.788, + 0.496 + ], + "angle": 0, + "content": "We present our results in Tab. 6, showing a clear advantage of SPARO over AttnPool. Additionally, both the choices of separate-head attention over multi-head and slot-wise operations over the alternatives result in improved generalization, reduced model size, and fewer floating-point operations per second (FLOPS). While the former supports the effectiveness of the separate-head attention bottleneck, the latter validates the benefits of the prior for shared separately-attendable concepts between the modalities enabled by SPARO." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.523, + 0.378, + 0.537 + ], + "angle": 0, + "content": "5.4 Visualizations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.788, + 0.734 + ], + "angle": 0, + "content": "To conclude our analysis, we qualitatively visualize the concepts represented by SPARO using the \\(\\mathrm{CLIP}^{16} + \\mathrm{SPARO}\\) model trained on CC15M from Sec. 4.1. A limitation of visualizing attended positions is that it only provides partial insight into the represented concept — we can tell where the information is taken from, but not how the information is used. Different SPARO slots can have similar attention maps for the same sample, but reveal different semantics when comparing patterns across other images. For instance, a slot that attends to animals (e.g. dogs, cats, elephants) and another that attends to transportation (e.g. trains, boats, cars) can both have the same attention mask over patches and words for a horse being used for transportation. Furthermore, not all SPARO concepts align between the two modalities for every sample due to information being present in one that is not possible to infer from the other." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.843 + ], + "angle": 0, + "content": "To ease our interpretation of attended concepts, we choose concepts to visualize that have sharp attention in the text modality, attend to non-overlapping text tokens, and have cosine similarity more than 0.75 with the corresponding image slots. We present examples of these filtered concepts in Fig. 3. In this example, we see that SPARO is capable of attending to different concepts such as subject of the scene, the activity represented, and the surrounding location, for both vision and text. We provide additional visualizations in Appendix B.7" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "A. Vani et al." + }, + { + "type": "image", + "bbox": [ + 0.238, + 0.153, + 0.351, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.225, + 0.216, + 0.364, + 0.232 + ], + "angle": 0, + "content": "A herd of cattle walking down a road being followed by a cowboy" + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.153, + 0.495, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.379, + 0.216, + 0.508, + 0.231 + ], + "angle": 0, + "content": "everal surfboards standing in a rowon the beach" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.153, + 0.646, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.216, + 0.648, + 0.231 + ], + "angle": 0, + "content": "Two people riding horses on a rock path" + }, + { + "type": "image", + "bbox": [ + 0.679, + 0.153, + 0.762, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.665, + 0.216, + 0.777, + 0.231 + ], + "angle": 0, + "content": "A man sitting alone on a park bench in a park" + }, + { + "type": "image", + "bbox": [ + 0.238, + 0.231, + 0.35, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.224, + 0.293, + 0.363, + 0.309 + ], + "angle": 0, + "content": "A herd of cattle walking down a road being followed by a cowboy" + }, + { + "type": "image", + "bbox": [ + 0.385, + 0.231, + 0.495, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.372, + 0.293, + 0.508, + 0.309 + ], + "angle": 0, + "content": "Several surfboards standing in a row on the beach" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.231, + 0.645, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.52, + 0.293, + 0.648, + 0.308 + ], + "angle": 0, + "content": "Two people riding horses on a rock path" + }, + { + "type": "image", + "bbox": [ + 0.68, + 0.231, + 0.762, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.665, + 0.293, + 0.777, + 0.308 + ], + "angle": 0, + "content": "A man sitting alone on a park bench in a park" + }, + { + "type": "image", + "bbox": [ + 0.238, + 0.309, + 0.35, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.225, + 0.371, + 0.362, + 0.387 + ], + "angle": 0, + "content": "A herd of cattle walking down a road being followed by a cowboy" + }, + { + "type": "image", + "bbox": [ + 0.385, + 0.309, + 0.495, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.373, + 0.371, + 0.507, + 0.386 + ], + "angle": 0, + "content": "Several surfboards standing in a row on the beach" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.309, + 0.645, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.52, + 0.371, + 0.648, + 0.386 + ], + "angle": 0, + "content": "Two people riding horses on a rock path" + }, + { + "type": "image", + "bbox": [ + 0.68, + 0.309, + 0.762, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.666, + 0.371, + 0.777, + 0.386 + ], + "angle": 0, + "content": "A man sitting alone on a park bench in a park" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.397, + 0.789, + 0.44 + ], + "angle": 0, + "content": "Fig. 3: Visualizing of the attended image and text positions for three SPARO slots (one per row) across four examples (one per column) from MS COCO. We surmise that the SPARO concepts from top to bottom represent the subject, activity, and location." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.468, + 0.355, + 0.483 + ], + "angle": 0, + "content": "6 Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.498, + 0.789, + 0.756 + ], + "angle": 0, + "content": "Humans utilize a prior for compositionality to coherently represent different subsets of salient aspects of the world. We can enrich our comprehension by selectively attending to new aspects or specialize to a task by filtering out distracting ones. We introduce SPARO with a goal of imparting a similar prior to transformers in representation learning frameworks by partitioning encodings into separately-attended concepts. Although we see evidence of disentanglement through positive results of post-hoc concept selection and visualization of attention maps, we do not impose any explicit independence or disentanglement constraints. In our representation learning frameworks, the pressure for learning distinct concepts arises from the need to explain the factors of variation through the training objective with only the features that can be accumulated by a single head of attention with embedded queries and limited dimensionality. However, auxiliary training objectives to impose stronger distributional conditions on the learned concepts can be explored as promising future directions. Beyond pretraining, we highlight that more sophisticated post-hoc concept selection approaches than explored in our work, like using human interaction or set cover algorithms, can further improve the downstream utility of SPARO." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.779, + 0.394, + 0.796 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "This research was funded by Sony and enabled in part by compute resources provided by the Digital Research Alliance of Canada, Mila, and Sony." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.305, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SPARO: Selective Attention in Transformer Encodings for Vision" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.324, + 0.161 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.173, + 0.785, + 0.2 + ], + "angle": 0, + "content": "1. Aydemir, G., Xie, W., Güney, F.: Self-supervised object-centric learning for videos. arXiv preprint arXiv:2310.06907 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.201, + 0.785, + 0.227 + ], + "angle": 0, + "content": "2. Ba, J.L., Kiros, J.R., Hinton, G.E.: Layer normalization. arXiv preprint arXiv:1607.06450 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.228, + 0.785, + 0.255 + ], + "angle": 0, + "content": "3. Bahdanau, D., Cho, K., Bengio, Y.: Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.255, + 0.785, + 0.309 + ], + "angle": 0, + "content": "4. Barbu, A., Mayo, D., Alverio, J., Luo, W., Wang, C., Gutfreund, D., Tenenbaum, J., Katz, B.: Objectnet: A large-scale bias-controlled dataset for pushing the limits of object recognition models. Advances in neural information processing systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.31, + 0.785, + 0.35 + ], + "angle": 0, + "content": "5. Beattie, C., Leibo, J.Z., Teplyashin, D., Ward, T., Wainwright, M., Kuttler, H., Lefrancq, A., Green, S., Valdés, V., Sadik, A., et al.: Deepmind lab. arXiv preprint arXiv:1612.03801 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.351, + 0.785, + 0.377 + ], + "angle": 0, + "content": "6. Boff, K.R., Kaufman, L., Thomas, J.P.: Handbook of perception and human performance, vol. 1. Wiley New York (1986)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.378, + 0.785, + 0.418 + ], + "angle": 0, + "content": "7. Brady, J., Zimmermann, R.S., Sharma, Y., Schölkopf, B., von Kügelgen, J., Brendel, W.: Provably learning object-centric representations. arXiv preprint arXiv:2305.14229 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.419, + 0.785, + 0.46 + ], + "angle": 0, + "content": "8. Burgess, C.P., Matthew, L., Watters, N., Kabra, R., Higgins, I., Botvinick, M., Lerchner, A.: Monet: Unsupervised scene decomposition and representation. arXiv preprint arXiv:1901.11390 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.46, + 0.785, + 0.5 + ], + "angle": 0, + "content": "9. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.501, + 0.785, + 0.527 + ], + "angle": 0, + "content": "0. Chang, H.S., Sun, R.Y., Ricci, K., McCallum, A.: Multi-cls bert: An efficient alternative to traditional ensembling. arXiv preprint arXiv:2210.05043 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.528, + 0.785, + 0.581 + ], + "angle": 0, + "content": "1. Changpinyo, S., Sharma, P., Ding, N., Soricut, R.: Conceptual 12M: Pushing web-scale image-text pre-training to recognize long-tail visual concepts. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3558-3568 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.582, + 0.785, + 0.609 + ], + "angle": 0, + "content": "2. Chen, X., Xie, S., He, K.: An empirical study of training self-supervised vision transformers. arXiv preprint arXiv:2104.02057 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.61, + 0.785, + 0.636 + ], + "angle": 0, + "content": "3. Cheng, G., Han, J., Lu, X.: Remote sensing image scene classification: Benchmark and state of the art. Proceedings of the IEEE 105(10), 1865-1883 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.636, + 0.785, + 0.677 + ], + "angle": 0, + "content": "4. Cimpoi, M., Maji, S., Kokkinos, I., Mohamed, S., Vedaldi, A.: Describing textures in the wild. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3606-3613 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.677, + 0.785, + 0.703 + ], + "angle": 0, + "content": "5. Colby, C.L., Goldberg, M.E.: Space and attention in parietal cortex. Annual review of neuroscience 22(1), 319-349 (1999)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.704, + 0.785, + 0.744 + ], + "angle": 0, + "content": "6. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.745, + 0.785, + 0.785 + ], + "angle": 0, + "content": "7. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: BERT: Pre-training of deep bidirectional transformers for language understanding. In: North American Chapter of the Association for Computational Linguistics (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.786, + 0.785, + 0.84 + ], + "angle": 0, + "content": "8. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.173, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "A. Vani et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "19. Dugas, E., Jared, Jorge, Cukierski, W.: Diabetic retinopathy detection (2015), https://kaggle.com/competitions/diabetic-retinopathy-detection" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.178, + 0.785, + 0.218 + ], + "angle": 0, + "content": "20. Engelcke, M., Kosiorek, A.R., Jones, O.P., Posner, I.: Genesis: Generative scene inference and sampling with object-centric latent representations. arXiv preprint arXiv:1907.13052 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.22, + 0.785, + 0.261 + ], + "angle": 0, + "content": "21. Eslami, S., Heess, N., Weber, T., Tassa, Y., Szepesvari, D., Hinton, G.E., et al.: Attend, infer, repeat: Fast scene understanding with generative models. Advances in neural information processing systems 29 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.262, + 0.785, + 0.289 + ], + "angle": 0, + "content": "22. Fei-Fei, L., Fergus, R., Perona, P.: One-shot learning of object categories. IEEE transactions on pattern analysis and machine intelligence 28(4), 594-611 (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.291, + 0.785, + 0.318 + ], + "angle": 0, + "content": "23. Geiger, A., Lenz, P., Stiller, C., Urtasun, R.: Vision meets robotics: The kitti dataset. The International Journal of Robotics Research 32(11), 1231-1237 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.32, + 0.785, + 0.36 + ], + "angle": 0, + "content": "24. Goyal, A., Didolkar, A., Ke, N.R., Blundell, C., Beaudoin, P., Heess, N., Mozer, M.C., Bengio, Y.: Neural production systems. Advances in Neural Information Processing Systems 34, 25673-25687 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.362, + 0.785, + 0.416 + ], + "angle": 0, + "content": "25. Goyal, A., Lamb, A., Gampa, P., Beaudoin, P., Blundell, C., Levine, S., Bengio, Y., Mozer, M.C.: Factorizing declarative and procedural knowledge in structured, dynamical environments. In: International Conference on Learning Representations (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.418, + 0.785, + 0.445 + ], + "angle": 0, + "content": "26. Goyal, A., Lamb, A., Hoffmann, J., Sodhani, S., Levine, S., Bengio, Y., Scholkopf, B.: Recurrent independent mechanisms. arXiv preprint arXiv:1909.10893 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.447, + 0.785, + 0.5 + ], + "angle": 0, + "content": "27. Greff, K., Kaufman, R.L., Kabra, R., Watters, N., Burgess, C., Zoran, D., Matthew, L., Botvinick, M., Lerchner, A.: Multi-object representation learning with iterative variational inference. In: International Conference on Machine Learning. pp. 2424-2433. PMLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.503, + 0.785, + 0.544 + ], + "angle": 0, + "content": "28. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.546, + 0.785, + 0.6 + ], + "angle": 0, + "content": "29. Helber, P., Bischke, B., Dengel, A., Borth, D.: Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 12(7), 2217-2226 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.602, + 0.785, + 0.657 + ], + "angle": 0, + "content": "30. Hendrycks, D., Basart, S., Mu, N., Kadavath, S., Wang, F., Dorundo, E., Desai, R., Zhu, T., Parajuli, S., Guo, M., et al.: The many faces of robustness: A critical analysis of out-of-distribution generalization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8340-8349 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.659, + 0.785, + 0.699 + ], + "angle": 0, + "content": "31. Hendrycks, D., Zhao, K., Basart, S., Steinhardt, J., Song, D.: Natural adversarial examples. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15262-15271 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.701, + 0.785, + 0.741 + ], + "angle": 0, + "content": "32. Hodosh, M., Young, P., Hockenmaier, J.: Framing image description as a ranking task: Data, models and evaluation metrics. Journal of Artificial Intelligence Research 47, 853-899 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.743, + 0.785, + 0.784 + ], + "angle": 0, + "content": "33. Hsieh, C.Y., Zhang, J., Ma, Z., Kembhavi, A., Krishna, R.: Sugarcrepe: Fixing hackable benchmarks for vision-language compositionality. Advances in neural information processing systems (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.785, + 0.841 + ], + "angle": 0, + "content": "34. Ilharco, G., Wortsman, M., Wightman, R., Gordon, C., Carlini, N., Taori, R., Dave, A., Shankar, V., Namkoong, H., Miller, J., Hajishirzi, H., Farhadi, A., Schmidt, L.: OpenCLIP. https://github.com/mlfoundations/open Clip (Jul 2021). https://doi.org/10.5281/zenodo.5143773" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.305, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SPARO: Selective Attention in Transformer Encodings for Vision" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.205 + ], + "angle": 0, + "content": "35. Johnson, J., Hariharan, B., van der Maaten, L., Fei-Fei, L., Lawrence Zitnick, C., Girshick, R.: Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2901-2910 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.206, + 0.787, + 0.233 + ], + "angle": 0, + "content": "36. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images. Technical report (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.234, + 0.787, + 0.29 + ], + "angle": 0, + "content": "37. Lavoie, S., Tsirigotis, C., Schwarzer, M., Vani, A., Noukhovitch, M., Kawaguchi, K., Courville, A.: Simplicial embeddings in self-supervised learning and downstream classification. In: International Conference on Learning Representations (2023), https://openreview.net/forum?id=RWtGreRpovS" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.291, + 0.787, + 0.346 + ], + "angle": 0, + "content": "38. LeCun, Y., Huang, F.J., Bottou, L.: Learning methods for generic object recognition with invariance to pose and lighting. In: Proceedings of the 2004 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2004. CVPR 2004. vol. 2, pp. II-104. IEEE (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.346, + 0.787, + 0.389 + ], + "angle": 0, + "content": "39. Lee, J., Lee, Y., Kim, J., Kosiorek, A., Choi, S., Teh, Y.W.: Set transformer: A framework for attention-based permutation-invariant neural networks. In: International conference on machine learning. pp. 3744-3753. PMLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.39, + 0.787, + 0.485 + ], + "angle": 0, + "content": "40. Lin, T., Maire, M., Belongie, S.J., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, C.L.: Microsoft COCO: common objects in context. In: Fleet, D.J., Pajdla, T., Schiele, B., Tuytelaars, T. (eds.) Computer Vision - ECCV 2014 - 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V. Lecture Notes in Computer Science, vol. 8693, pp. 740-755. Springer (2014). https://doi.org/10.1007/978-3-319-10602-1_48, https://doi.org/10.1007/978-3-319-10602-1%5F48" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.487, + 0.787, + 0.529 + ], + "angle": 0, + "content": "41. Locatello, F., Weissenborn, D., Unterthiner, T., Mahendran, A., Heigold, G., Uszkoreit, J., Dosovitskiy, A., Kipf, T.: Object-centric learning with slot attention. Advances in Neural Information Processing Systems 33, 11525-11538 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.53, + 0.787, + 0.585 + ], + "angle": 0, + "content": "42. Ma, Z., Hong, J., Gul, M.O., Gandhi, M., Gao, I., Krishna, R.: Crepe: Can vision-language foundation models reason compositionally? In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10910-10921 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.586, + 0.787, + 0.64 + ], + "angle": 0, + "content": "43. Mansouri, A., Hartford, J., Zhang, Y., Bengio, Y.: Object centric architectures enable efficient causal representation learning. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=r9FsiXZxZt" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.642, + 0.787, + 0.684 + ], + "angle": 0, + "content": "44. Martinez, A.: Putting spatial attention on the map: timing and localization of stimulus selection processes in striate and extrastriate visual areas. Vision Res 41, 1437-1457 (2001)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.685, + 0.787, + 0.713 + ], + "angle": 0, + "content": "45. Matthew, L., Higgins, I., Hassabis, D., Lerchner, A.: dsprites: Disentanglement testing sprites dataset. https://github.com/deepmind/dsprites-dataset/ (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.714, + 0.787, + 0.755 + ], + "angle": 0, + "content": "46. Nilsback, M.E., Zisserman, A.: Automated flower classification over a large number of classes. In: 2008 Sixth Indian conference on computer vision, graphics & image processing. pp. 722-729. IEEE (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.757, + 0.787, + 0.798 + ], + "angle": 0, + "content": "47. O'Connor, D.H., Fukui, M.M., Pinsk, M.A., Kastner, S.: Attention modulates responses in the human lateral geniculate nucleus. Nature neuroscience 5(11), 1203-1209 (2002)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "48. Parkhi, O.M., Vedaldi, A., Zisserman, A., Jawahar, C.: Cats and dogs. In: 2012 IEEE conference on computer vision and pattern recognition. pp. 3498-3505. IEEE (2012)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "A. Vani et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "49. Qian, R., Ding, S., Liu, X., Lin, D.: Semantics meets temporal correspondence: Self-supervised object-centric learning in videos. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16675-16687 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.245 + ], + "angle": 0, + "content": "50. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.786, + 0.273 + ], + "angle": 0, + "content": "51. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners. OpenAI blog 1(8), 9 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.787, + 0.327 + ], + "angle": 0, + "content": "52. Ray, A., Radenovic, F., Dubey, A., Plummer, B.A., Krishna, R., Saenko, K.: Cola: A benchmark for compositional text-to-image retrieval. In: Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.787, + 0.37 + ], + "angle": 0, + "content": "53. Recht, B., Roelofs, R., Schmidt, L., Shankar, V.: Do imagenet classifiers generalize to imagenet? In: International conference on machine learning. pp. 5389-5400. PMLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.787, + 0.412 + ], + "angle": 0, + "content": "54. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.787, + 0.438 + ], + "angle": 0, + "content": "55. Scott, W.A.: Cognitive complexity and cognitive flexibility. Sociometry pp. 405-414 (1962)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.787, + 0.481 + ], + "angle": 0, + "content": "56. Seitzer, M., Horn, M., Zadaianchuk, A., Zietlow, D., Xiao, T., Simon-Gabriel, C.J., He, T., Zhang, Z., Scholkopf, B., Brox, T., et al.: Bridging the gap to real-world object-centric learning. arXiv preprint arXiv:2209.14860 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.787, + 0.55 + ], + "angle": 0, + "content": "57. Sharma, P., Ding, N., Goodman, S., Soricut, R.: Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In: Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics, ACL 2018, Melbourne, Australia, July 15-20, 2018, Volume 1: Long Papers. pp. 2556-2565. Association for Computational Linguistics (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.787, + 0.605 + ], + "angle": 0, + "content": "58. Thrush, T., Jiang, R., Bartolo, M., Singh, A., Williams, A., Kiela, D., Ross, C.: Winoground: Probing vision and language models for visio-linguistic compositionality. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5238-5248 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.787, + 0.633 + ], + "angle": 0, + "content": "59. Treisman, A.M., Gelade, G.: A feature-integration theory of attention. Cognitive psychology 12(1), 97-136 (1980)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.787, + 0.675 + ], + "angle": 0, + "content": "60. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.787, + 0.731 + ], + "angle": 0, + "content": "61. Veeling, B.S., Linmans, J., Winkens, J., Cohen, T., Welling, M.: Rotation equivariant cnns for digital pathology. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2018: 21st International Conference, Granada, Spain, September 16-20, 2018, Proceedings, Part II 11. pp. 210-218. Springer (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.731, + 0.787, + 0.772 + ], + "angle": 0, + "content": "62. Wang, H., Ge, S., Lipton, Z., Xing, E.P.: Learning robust global representations by penalizing local predictive power. In: Advances in Neural Information Processing Systems. pp. 10506-10518 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.787, + 0.813 + ], + "angle": 0, + "content": "63. Xiao, J., Hays, J., Ehinger, K.A., Oliva, A., Torralba, A.: Sun database: Large-scale scene recognition from abbey to zoo. In: 2010 IEEE computer society conference on computer vision and pattern recognition. pp. 3485-3492. IEEE (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "64. Xu, L., Ouyang, W., Bennamoun, M., Boussaid, F., Xu, D.: Multi-class token transformer for weakly supervised semantic segmentation. In: Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.305, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SPARO: Selective Attention in Transformer Encodings for Vision" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.245, + 0.147, + 0.786, + 0.175 + ], + "angle": 0, + "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4310-4319 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.787, + 0.217 + ], + "angle": 0, + "content": "65. Young, P., Lai, A., Hodosh, M., Hockenmaier, J.: From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics 2, 67-78 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.787, + 0.258 + ], + "angle": 0, + "content": "66. Yu, J., Wang, Z., Vasudevan, V., Yeung, L., Seyedhosseini, M., Wu, Y.: Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.787, + 0.3 + ], + "angle": 0, + "content": "67. Yuksekgonul, M., Bianchi, F., Kalluri, P., Jurafsky, D., Zou, J.: When and why vision-language models behave like bags-of-words, and what to do about it? In: The Eleventh International Conference on Learning Representations (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.787, + 0.342 + ], + "angle": 0, + "content": "68. Yuval, N.: Reading digits in natural images with unsupervised feature learning. In: Proceedings of the NIPS Workshop on Deep Learning and Unsupervised Feature Learning (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.787, + 0.383 + ], + "angle": 0, + "content": "69. Zadaianchuk, A., Kleindessner, M., Zhu, Y., Locatello, F., Brox, T.: Unsupervised semantic segmentation with self-supervised object-centric representations. arXiv preprint arXiv:2207.05027 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.383, + 0.787, + 0.438 + ], + "angle": 0, + "content": "70. Zhai, X., Puigcerver, J., Kolesnikov, A., Ruyssen, P., Riquelme, C., Lucic, M., Djolonga, J., Pinto, A.S., Neumann, M., Dosovitskiy, A., et al.: A large-scale study of representation learning with the visual task adaptation benchmark. arXiv preprint arXiv:1910.04867 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.787, + 0.466 + ], + "angle": 0, + "content": "71. Zhang, Y., Hare, J., Prugel-Bennett, A.: Deep set prediction networks. Advances in Neural Information Processing Systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.787, + 0.508 + ], + "angle": 0, + "content": "72. Zhao, T., Zhang, T., Zhu, M., Shen, H., Lee, K., Lu, X., Yin, J.: Vl-checklist: Evaluating pre-trained vision-language models with objects, attributes and relations. arXiv preprint arXiv:2207.00221 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.508 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/SPARO_ Selective Attention for Robust and Compositional Transformer Encodings for Vision/4bccfc91-c07f-416d-bd12-0ad527681e54_origin.pdf b/2024/SPARO_ Selective Attention for Robust and Compositional Transformer Encodings for Vision/4bccfc91-c07f-416d-bd12-0ad527681e54_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d6a11e3d669032031a700d25386e2ae039377a56 --- /dev/null +++ b/2024/SPARO_ Selective Attention for Robust and Compositional Transformer Encodings for Vision/4bccfc91-c07f-416d-bd12-0ad527681e54_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:85b886f32605022520969f81a9a6222cd56c849e0ceddbb584d9879d22ffcc8d +size 9839839 diff --git a/2024/SPARO_ Selective Attention for Robust and Compositional Transformer Encodings for Vision/full.md b/2024/SPARO_ Selective Attention for Robust and Compositional Transformer Encodings for Vision/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c124e35a2f61be2436aa288e3ccac17980d2d79e --- /dev/null +++ b/2024/SPARO_ Selective Attention for Robust and Compositional Transformer Encodings for Vision/full.md @@ -0,0 +1,314 @@ +![](images/fcf0f2e780151c3244b95b9fe666092896fd1f14ea02089ec12accca332f2455.jpg) + +# SPARO: Selective Attention for Robust and + +# Compositional Transformer Encodings for Vision + +Ankit Vani1*, Bac Nguyen2, Samuel Lavoie1, Ranjay Krishna3, and Aaron Courville1,4 + +1 Mila, Université de Montréal + +2 Sony AI + +3 University of Washington, Allen Institute for Artificial Intelligence + +4 CIFAR AI Chair + +Abstract. Selective attention helps us focus on task-relevant aspects in the constant flood of our sensory input. This constraint in our perception allows us to robustly generalize under distractions and to new compositions of perceivable concepts. Transformers employ a similar notion of attention in their architecture, but representation learning models with transformer backbones like CLIP and DINO often fail to demonstrate robustness and compositionality. We highlight a missing architectural prior: unlike human perception, transformer encodings do not separately attend over individual concepts. In response, we propose SPARO, a readout mechanism that partitions encodings into separately-attended slots, each produced by a single attention head. Using SPARO with CLIP imparts an inductive bias that the vision and text modalities are different views of a shared compositional world with the same corresponding concepts. Using SPARO, we demonstrate improvements on downstream recognition, robustness, retrieval, and compositionality benchmarks with CLIP (up to $+14\%$ for ImageNet, $+4\%$ for SugarCrepe), and on nearest neighbors and linear probe for ImageNet with DINO ( $+3\%$ each). We also showcase a powerful ability to intervene and select individual SPARO concepts to further improve downstream task performance (up from $+4\%$ to $+9\%$ for SugarCrepe) and use this ability to study the robustness of SPARO's representation structure. Finally, we provide insights through ablation experiments and visualization of learned concepts. + +Keywords: Selective attention $\cdot$ Slot representations $\cdot$ Transformers + +# 1 Introduction + +Selective attention is an intrinsic property of human perception [15,44,47], enabling people to focus on task-relevant aspects of their surroundings while tuning out the rest [55]. For instance, it allows a driver to focus on traffic signals, road signs, and other vehicles, or an individual searching for their friend + +in a crowd to attend to details like people's heights and clothing. More broadly, it empowers us to comprehend the vast complexity of our world with limited mental resources [6] by constraining our perception to one concept at a time. When a task requires information for multiple concepts (e.g., looking at traffic signals and pedestrians), this constraint calls for sequentially attending to separate salient aspects of our stimuli [59]. The cognitive bottleneck of selective attention, with the ability to separately represent concepts from our environment, enables human perception to remain insensitive to irrelevant distractions and generalize to new compositions of perceivable concepts. + +In machine learning, the notion of selective attention is party emulated by the attention mechanism 3, which is a core component of transformers 60 such as the vision transformer (ViT) 18. However, transformer encodings learnt using approaches like CLIP 50 and DINO 9 still struggle with robustness 30, 31, 62, 70 and compositional generalization 33, 42, 52, 58, 67, 72. For instance, despite CLIP emerging as the de facto backbone for many vision tasks, it can perform close to chance when evaluated for compositionality 42. While the attention mechanism equips these models with the ability to emphasize relevant aspects of their intermediate representations, it does not endow them with the means to produce encodings for separately-attended concepts. Consequently, transformer-based representation learning models lack an important prior from human cognition: that perception of a complex input can be broken down into perception of its salient concepts. + +This limitation deprives downstream tasks easy access to task-relevant aspects of the data's underlying structure. In a task like ImageNet classification, where most images contain one primary object of interest (e.g., "dog"), information like relationships between objects and attributes of the background are irrelevant and can lead to overfitting if considered. However, these details can become crucial to attain high performance in other tasks like image retrieval (e.g., "dog chasing a frisbee in a park"). Similar to the generalization benefits of selective attention in human perception, encodings that allow easy separation of attended concepts in the data can generalize to a diverse set of downstream tasks. Without additional constraints, however, the emergence of such compositional structure in encodings is unlikely, especially in the common setting of training on noisy internet data. + +We propose SPARO (Separate-head attention read-out) as an improved read-out mechanism for transformer encoders in vision inspired by selective attention. SPARO replaces the last transformer block to provide a mechanism for partitioning encodings into slots of separately-attended concepts. We design each slot encoding to be a low-dimensional result of a single-head attention operation, producing a bottleneck that encourages each slot to "selectively attend." In this mechanism, producing multiple slots can be interpreted as selective attention over different concepts occurring in parallel. When training CLIP, SPARO imparts an inductive bias that both modalities are views onto a shared compositional world with the corresponding slots of both encoders representing its concepts. + +In contrast, standard CLIP merely embeds the encodings of both modalities in a shared vector space without imposing any additional structure. + +Training SPARO using CLIP and DINO exhibits improved generalization, robustness, and compositionality while retaining the same model size. Our experiments show that using CLIP with SPARO improves generalization for zero-shot recognition, robustness, retrieval, and compositionality (e.g., +14% for ImageNet, +4% for SugarCrepe, when trained with 15M examples; +3% each when trained with 400M examples). We also demonstrate improved linear probe performance of CLIP (+10% trained with 15M examples; +2% with 400M examples), as well as both linear probe and nearest neighbors performance of DINO (+3% each). We showcase how SPARO's representational structure enables the ability to manually intervene and select relevant concepts for downstream tasks, leading to improved compositionality and generalization (up from +4% to +9% for SugarCrepe trained with 15M examples; +3% to +6% with 400M examples). We then study the robustness of this structure, validate the choice of separate-head attention through ablations, and provide insights into the effect of the number and dimensionality of concepts. Finally, we provide visualizations for some of the concepts SPARO learns to attend to. + +# 2 Related work + +We situate our work amongst other transformer read-out mechanisms and slot representation learning methods. + +Transformer read-outs. For transformer [60] decoders with causal masking, the last hidden state, corresponding to the end of sequence (EOS) input token, is used as the output representation. BERT [17] proposed inserting a special CLS token before input sequences for transformer encoders where information from all positions is gathered to produce the encoding. Using the CLS token output is also popular for Vision transformer (ViT) [9, 12, 18] backbones, along with alternatives such as global average pooling [12, 64]. In the context of language, Multi-CLS BERT [10] argues that an input text sequence can have multiple facets, and that samples can be similar along one facet and dissimilar along another. The authors aim to represent these facets by adding multiple CLS tokens to the input sequence. Using separate attention heads for each SPARO slot can also encourage attending to different facets of the inputs, but without requiring expensive accumulations through each layer of the backbone. By keeping SPARO independent of the backbone, we also allow for its easy potential placement on top of frozen pre-trained models without needing to train for added CLS tokens. + +SPARO is most related to the attentional pooler [39,66] architecture, which uses multi-head attention with embedded learned queries to produce output encodings. However, SPARO employs a critical structural constraint of not mixing the information from separate attention heads, a design choice we justify empirically in Sec. 5.3. Slot attention [41] also produces slot encodings through the attention mechanism, but rather than having the input positions compete + +for each slot, the slots compete for each input position. Slot attention also proposes an expensive iterative read-out mechanism where the role of each slot is determined by sampling from a per-sample Gaussian distribution, which can be limiting for large real-world datasets. + +Slot representations. Object-centric slot representations have been studied extensively in the literature [7,8,20,21,24-27,41,43,71]. Often, these representations are expected to represent predefined objects and attributes, and evaluated in small-scale or synthetic settings. Recently, the use of slot attention on features produced by a learned encoder [1,49,56], and utilizing unsupervised saliency masks for object discovery [69] has enabled unsupervised learning of object-centric representations in larger scale settings. Related to these ideas, SPARO attends to features produced by a transformer using a single attention head per slot, and each head is capable of learning saliency maps without supervision [9]. However, we do not constrain our slots to correspond to predefined types of objects in the data, nor do we impose any independence conditions between the slots. Self-supervised simplicial embeddings (SEM) [37], with only a soft-discrete slot structure constraint, enable improved generalization on downstream tasks. SPARO replaces the separation of softmax application between SEM slots with a separation of underlying attention operations. + +# 3 Method + +Having motivated our approach, we now detail the design of SPARO. We start by introducing the relevant notation, then discuss the SPARO module. Finally, we discuss the addition of SPARO in CLIP and DINO, and the applicability of SPARO beyond these settings. + +# 3.1 Notation + +We envision SPARO as a special attention-based layer that modifies the top block of a transformer encoder [51,60] backbone (including the ViT [18]). Transformers take as input a sequence $\pmb{x} = \{\pmb{x}_1,\dots,\pmb{x}_n\}$ of length $n$ , and produce an output state per position $\{h_1,\ldots ,h_n\}$ , where $\pmb{x}_i\in \mathbb{R}^{\mathrm{inputdim}}$ , $\pmb{h}_i\in \mathbb{R}^d$ , for any $i\in [n]$ , with $d$ the model width. Finally, a pooling operation reduces the sequence of output states to output encodings with a fixed size $\pmb{y}\in \mathbb{R}^{M}$ . Examples of pooling operations include attentional pooling [39], global average pooling, and extracting the CLS or EOS token representations for images or text respectively. + +# 3.2 Separate-head attention read-out (SPARO) + +We replace the pooling operation of transformer encoders with a concatenation of outputs of $L$ single-head attention mechanisms. SPARO acts on the transformer outputs $\pmb{H} = \{\pmb{h}_1, \dots, \pmb{h}_n\} \in \mathbb{R}^{n \times d}$ to produce the encoder output + +![](images/c0ab253b9f9f3a5a1694cc4ff2e7722630d79db6eaf7a0d69bb5d559a6a8165b.jpg) +Fig. 1: Illustration of SPARO, a read-out mechanism that structures representations as collections of separately-attended concepts. Take a standard $N$ -block transformer encoder (ViT here as an example), producing an encoding $y$ through extraction of its CLS token output. We can replace the $N$ th transformer block with the SPARO module (typically with equal or fewer parameters) to produce a SPARO encoding $y$ , which is a concatenation of $L$ SPARO slots. Each SPARO slot $y_l$ is produced through single-head attention over the backbone outputs using an embedded query $q_l$ . The value projection is a composition of slot-specific key projection parameterized by $K_l$ and a slot-wise projection shared between all SPARO concepts parameterized by $W$ . + +$\pmb{y} = \mathrm{concat}(\pmb{y}_1, \dots, \pmb{y}_L) \in \mathbb{R}^{LV}$ , where SPARO slot $\pmb{y}_l \in \mathbb{R}^V$ , $l \in [L]$ . We illustrate the architecture of SPARO in Fig. 1. Each SPARO slot is produced as: + +$$ +\boldsymbol {y} _ {l} = \boldsymbol {W} \boldsymbol {K} _ {l} \boldsymbol {H} ^ {\intercal} \text {s o f t m a x} \left(\frac {\boldsymbol {H} \boldsymbol {K} _ {l} ^ {\intercal} \boldsymbol {q} _ {l}}{\sqrt {D}}\right). \tag {1} +$$ + +Here, $\pmb{K}_l \in \mathbb{R}^{D \times d}$ is a learned key projection weight, and $\pmb{q}_l \in \mathbb{R}^D$ is a learned query embedding for attention. We parameterize the value projection as $\pmb{W}\pmb{K}_l$ where $\pmb{W} \in \mathbb{R}^{V \times D}$ is learned parameter shared between all $L$ slots. Decomposing the value projection in this manner helps reduce the number of model parameters, allowing a larger choice of $L$ with fixed resources. + +Generally, we pick $V$ and $D$ to be significantly smaller than the transformer width. In practice, we use the typical value of $D = 64$ used in standard multihead attention modules, and set $V$ to the same value in our experiments. Therefore, each SPARO slot has limited expressivity and representational capacity. The bottleneck underpinning each SPARO slot is a result of these constraints and the mechanistic bottleneck of a single attention mechanism instantiating competition between its input features. + +SPARO can be thought of as multi-head attention [60] where the queries are learned embeddings, key and value projection weights are shared, and the output projection weight is block-diagonal with each block containing the same parameters. Assuming $L = V = D = \sqrt{d}$ , SPARO requires a total of $d^2 + 2d$ parameters, compared to multi-head attention's $4d^2$ . + +# 3.3 CLIP with SPARO + +We add SPARO to both the image and text encoders of CLIP [50], setting the same values of $L$ and $V$ in both encoders. Each SPARO slot is $\ell_2$ -normalized separately, and we divide their concatenation by $\sqrt{L}$ to produce the global $\ell_2$ -normalized encoding. With this scheme, the cosine similarity between the image encoding $\mathbf{y}^{\mathrm{i}}$ and the text encoding $\mathbf{y}^{\mathrm{t}}$ becomes the expected cosine similarity between corresponding slot encodings: + +$$ +\boldsymbol {y} ^ {\mathrm {i}} ^ {\top} \boldsymbol {y} ^ {\mathrm {t}} = \sum_ {l = 1} ^ {L} \| \boldsymbol {y} _ {l} ^ {\mathrm {i}} \| \| \boldsymbol {y} _ {l} ^ {\mathrm {t}} \| \frac {\boldsymbol {y} _ {l} ^ {\mathrm {i}} {} ^ {\top} \boldsymbol {y} _ {l} ^ {\mathrm {t}}}{\| \boldsymbol {y} _ {l} ^ {\mathrm {i}} \| \| \boldsymbol {y} _ {l} ^ {\mathrm {t}} \|} = \sum_ {l = 1} ^ {L} \frac {1}{L} \operatorname {s i m} (\boldsymbol {y} _ {l} ^ {\mathrm {i}}, \boldsymbol {y} _ {l} ^ {\mathrm {t}}) = \underset {l \sim [ L ]} {\mathbb {E}} \left[ \operatorname {s i m} (\boldsymbol {y} _ {l} ^ {\mathrm {i}}, \boldsymbol {y} _ {l} ^ {\mathrm {t}}) \right]. \quad (2) +$$ + +We train the model using the standard CLIP loss which maximizes $\mathrm{sim}(\pmb{y}^{\mathrm{i}},\pmb{y}^{\mathrm{t}})$ of encodings $\pmb{y}^{\mathrm{i}}$ and $\pmb{y}^{\mathrm{t}}$ for aligned image-text pairs and minimizes it across all other pairings. With SPARO's separate-head attention bottleneck, maximizing $\mathbb{E}_{l\sim [L]}\left[\mathrm{sim}(\pmb{y}_l^{\mathrm{i}},\pmb{y}_l^{\mathrm{t}})\right]$ for aligned pairs across the training dataset encourages both modalities to learn similar attention semantics per slot $l\in [L]$ . Consequently, SPARO imparts a prior that both modalities encode a shared world with $L$ selectively attendable concepts, and similarities between inputs can be expressed as their average similarity for these concepts. + +# 3.4 DINO with SPARO + +The online and momentum encoders of DINO [9] are each comprised of a ViT backbone which produces the encoding, and a DINO head that transforms it into a categorical distribution for distillation. Unlike CLIP, these encoders have an inherent bias for learning similar encoding functions due to one being an exponentially moving average of the other. However, in standard DINO, there are no constraints for selective attention in the encoding structure. We add SPARO to DINO to impart the prior for representing separately-attendable concepts in its encodings. Concretely, we replace the final transformer block of the ViT backbone with the SPARO, but do not modify the DINO head. + +# 4 Results + +In Secs. 4 and 5, we evaluate our method on a variety of downstream tasks, analyze SPARO's representation structure with the ability to intervene on selected slots, perform ablations to support our design, and visualize the learned concepts. + +In this section, we validate that partitioning the representation structure as a collection of separately-attended concepts leads to improved generalization for downstream recognition, robustness, compositionality, and retrieval, using SPARO with CLIP [50] in Secs. [4.1] to [4.3] and DINO [9] in Sec. [4.3]. + +Datasets. We train our CLIP models on one of Conceptual Captions 3M (CC3M) [57], Conceptual Captions 12M (CC12M) [11], a combination of CC3M and CC12M (CC15M), or LAION-400M (L400M) [54]. Appendix C provides the statistics for our training datasets. + +Table 1: Zero-shot recognition, robustness, and SugarCrepe compositionality for CLIP models trained on Conceptual Captions and LAION-400M. + +
TrainModelImageNet-Object NetSugar Crepe
V1V2SketchAR
CC3MCLIP16 (C)0.1410.1220.0680.0330.1770.0800.611
C+GAP0.1560.1340.0690.0330.1870.0810.616
C+SPARO0.1700.1400.0880.0350.2210.0980.625
CC12MCLIP16 (C)0.3610.3110.2490.0910.4670.2180.697
C+GAP0.3820.3300.2620.1010.5010.2410.695
C+SPARO0.4060.3500.2980.1130.5590.2680.723
CC15MCLIP16 (C)0.3840.3370.2680.1050.5030.2380.699
C+GAP0.3990.3430.2870.1140.5310.2520.701
C+SPARO0.4370.3780.3170.1450.5790.2790.730
L400MCLIP32 (C)0.6170.5310.4820.2020.7190.4230.748
C+GAP0.6230.5370.4920.2120.7250.4400.732
C+SPARO0.6350.5520.5070.2310.7470.4590.770
+ +Models. For all experiments in this section, we train CLIP models using the open-source OpenCLIP [34] project. We consider two model sizes: those with a ViT-B/16 visual backbone $(\mathrm{CLIP}^{16})$ and those with a ViT-B/32 backbone $(\mathrm{CLIP}^{32})$ . For models trained with SPARO, we replace the last transformer block of the base transformer backbones. For our chosen settings, this ensures that the resulting model size is comparable to the original model size. Further, since only the CLS token output from the image transformer and the EOS token output from the text transformer are used for the standard CLIP encodings, removing one transformer block ensures that we do not gain an unfair advantage by attending to positions from a layer that standard CLIP discards. Note that standard CLIP also retains the MLP as a part of its final transformer block, which is absent in the SPARO module. We also compare our encodings with those produced by CLIP with global average pooling (CLIP+GAP), which enjoys the advantage of being able to use all of the final layer outputs. When using SPARO, we use $L = V = 64$ when training on Conceptual Captions, and $L = 128$ , $V = 64$ when training on LAION-400M. We also provide results for CLIP with residual network (ResNet) [28] encoders in Appendix B.3. + +# 4.1 Zero-shot recognition, robustness, and compositionality + +We evaluate the zero-shot classification accuracy of trained CLIP models on ImageNet [16] and a set of robustness benchmarks including ImageNet-V2 [53], ImageNet-Sketch [62], ImageNet-A [31], ImageNet-R [30], and ObjectNet [4]. Additionally, to evaluate the compositionality of our learned encodings in terms of objects, attributes, and relations, we test our trained models on SugarCrepe + +![](images/a4e4ed91bc6eccbd8d1a29e4a177c6082607043557c7f549f34082801cd4e39f.jpg) +(a) $\mathrm{CLIP^{16}}$ models trained on CC3M. + +![](images/b2ee92f9fafb6e58631375e2d655d3047e4e2676e91d796712b2e8a95510b0ea.jpg) +(b) $\mathrm{CLIP^{16}}$ models trained on CC12M. + +![](images/888dfbd4e6e928bde8154e4532868dd2a102f34b33c0dae68dfcbb083b2f6ff9.jpg) +(c) $\mathrm{CLIP^{16}}$ models trained on CC15M. + +![](images/68e09b5ef1d38e54acfe64c539467f560e083602c12c930b69d77cb815fe427c.jpg) +(d) $\mathrm{CLIP^{32}}$ models trained on L400M. +Fig. 2: Relative differences of $\mathrm{CLIP + SPARO}$ zero-shot accuracies when compared to $\mathrm{CLIP + GAP}$ on the VTAB benchmark. + +33. We report our results in Tab. 1, and provide more fine-grained SugarCrepe numbers in Appendix B.4. We find that using SPARO encodings outperforms CLIP and $\mathrm{CLIP + GAP}$ in all settings considered. The improved results on the SugarCrepe benchmark suggests that SPARO encodings exhibit compositionality at a level that enable better binding of object, attribute, and relation properties than baseline encodings on average. + +Finally, we also use the VTAB benchmark [70] to evaluate SPARO on a diverse set of datasets [5,13,14,19,22,23,29,35,36,38,45,46,48,61,63,68] that vary more significantly from the training data. We present the relative zero-shot classification accuracy improvements of CLIP + SPARO over CLIP + GAP in Fig. 2 and the corresponding absolute values in Appendix B.6. We see that SPARO outperforms standard CLIP encodings for a majority of VTAB tasks in all four settings. + +# 4.2 Zero-shot image and text retrieval + +To separately evaluate the quality of the learned image and text encodings, we consider zero-shot retrieval based on image and text on MS COCO [40], Flickr8k [32], and Flickr30k [65]. Our results are presented in Tab. 2. To compute Recall@5 for 'Text' we find the nearest 5 texts for an image, calculate the fraction of those texts that match the image, and average this metric over all the images (similar for Recall@5 for 'Image'). We find that SPARO encodings improve retrieval across all three datasets for both modalities. + +Table 2: Image and text zero-shot Recall@5 retrieval results on MS COCO, Flickr8k, and Flickr30k for CLIP models trained on Conceptual Captions and LAION-400M. + +
TrainModelImage R@5Text R@5
COCOF8kF30kCOCOF8kF30k
CC3MCLIP16 (C)0.2370.4000.3530.2940.4890.469
C+GAP0.2600.4300.4000.3120.5340.497
C+SPARO0.2890.4900.4610.3630.5870.581
CC12MCLIP16 (C)0.4740.6910.7150.6140.8370.817
C+GAP0.4990.7140.7250.6280.8190.847
C+SPARO0.5260.7370.7590.6700.8780.878
CC15MCLIP16 (C)0.5120.7380.7570.6360.8440.865
C+GAP0.5350.7610.7720.6590.8490.881
C+SPARO0.5570.7780.7930.6960.8980.905
L400MCLIP32 (C)0.5990.8220.8400.7650.9230.935
C+GAP0.6100.8220.8450.7680.9170.946
C+SPARO0.6160.8360.8540.7740.9350.950
+ +Table 3: Left: ImageNet linear probe accuracy for CLIP models trained on Conceptual Captions and LAION-400M. Right: ImageNet 20-nearest neighbors and linear probe accuracy for encodings trained with DINO. + +
ModelImageNet linear probe
Train: CC3MCC12MCC15ML400M
CLIP (C)0.4690.6300.6460.743
C+GAP0.5040.6490.6640.747
C+SPARO0.5610.7000.7110.755
+ +
ModelImageNet eval
20-NN Linear probe
DINO0.6850.735
DINO+SPARO0.7060.757
+ +# 4.3 Linear probe and DINO nearest neighbors classification + +We evaluate the classification accuracy of linear probes trained on the encodings of our trained models. For CLIP models, we follow [50] where we take a subset of the training data to use as a validation set, and sweep for the best weight decay hyperparameter on it for each model setting. Additionally, we evaluate SPARO for self-supervised learning in vision using DINO [9]. We train DINO with a ViT-S/16 backbone on ImageNet [16] without labels with 8 GPUs for 100 epochs without changing any other hyperparameters. Here, we use SPARO with $L = V = 64$ , replacing the last transformer block of the ViT encoder. The trained DINO models are evaluated on $k$ -nearest neighbours with $k = 20$ and linear classification on ImageNet with labels. We report our results for CLIP and DINO in Tab. Our results for CLIP align with our previous results, with SPARO outperforming CLIP and CLIP+GAP in all linear probe settings. Furthermore, we see that a straightforward incorporation of SPARO in the DINO encoder + +Table 4: Effect of intervening to selecting only the top 32 slots based on zero-shot ImageNet validation accuracy on SugarCrepe compositionality for CC15M-trained $\mathrm{CLIP}^{16} + \mathrm{SPARO}$ and LAION-400M-trained $\mathrm{CLIP}^{32} + \mathrm{SPARO}$ . + +
SugarCrepeCC15ML400M
Slots:100%50%100%
Object0.8960.9120.9270.929
ReplaceAttribute0.7980.7970.8220.854
Relation0.6660.7130.6780.705
SwapObject0.5930.6590.5850.646
Attribute0.7000.7030.7190.740
AddObject0.7790.8090.8660.857
Attribute0.6790.7370.7940.822
Average0.7300.7610.7700.793
+ +outperforms standard DINO for both nearest neighbors and linear classification. We provide additional experimental details in Appendices C.1 and C.2 + +# 5 Analysis + +In this section, we provide insights into encodings learned by SPARO. We demonstrate the ability of using SPARO's slot structure to perform post-hoc selection of slots for improving downstream performance. We then show that without this structure, such selection is more prone to overfitting. Next, we validate the design choice of using single-head attention to model the notion of selective attention through controlled ablations. Finally, we visualize example concepts that CLIP models trained with SPARO learn to attend to. + +# 5.1 Post-hoc concept selection + +SPARO enables downstream tasks access to a collection of concepts. However, not all concepts may be important for a specific downstream task. For example, attending to watermarks or caption style can be important when trying to determine the source of a sample, but is unnecessary and often harmful for most downstream tasks that focus on the elements in the scenes. The separation of concepts in the structure of a SPARO encoding makes it possible to manually intervene to select desirable slots. + +We demonstrate this ability with a simple heuristic: we perform concept selection by computing the zero-shot ImageNet validation accuracy of each slot separately, and pick only the top-performing 32 slots. In Tab.4, we show that such an intervention exposes improved compositionality when evaluating CLIP+SPARO models trained on CC15M and LAION-400M on SugarCrepe. + +Table 5: Effect of overfitting a partitioned mask for image encodings $\mathbf{y}$ of frozen CC15M-trained ${\mathrm{{CLIP}}}^{32} + \mathrm{{GAP}}$ and ${\mathrm{{CLIP}}}^{32} + \mathrm{{SPARO}}$ models by training on SugarCrepe. We normalize the relative change on the evaluation benchmarks by the attained relative change on SugarCrepe. The highlighted rows indicate settings where the structure assumed for the mask aligns with SPARO's representation structure. + +
(# parts, Model Shape of y)TrainingEvaluation
SugarCrepeVTAB AvgImageNet 0-shot Acc
InitialMaskedInitialMaskedRelative changeInitialMaskedRelative change
(64, 64)+GAP(4096)0.6920.7320.2400.217-1.6890.3430.292-2.598
+SPARO(64, 64)0.7100.7640.2620.272+0.4970.3700.364-0.190
(64, 8)+GAP(512)0.6780.7120.2540.219-2.7940.3450.279-3.924
+SPARO(64, 8)0.6960.7410.2490.260+0.6880.3550.347-0.347
(512, 1)+GAP(512)0.6780.7710.2540.215-1.1190.3450.250-2.003
+SPARO(64, 8)0.6960.8090.2490.206-1.0590.3550.224-2.277
(4096, 1)+GAP(4096)0.6920.8150.2400.215-0.5960.3430.286-0.938
+SPARO(64, 64)0.7100.8810.2620.242-0.3160.3700.328-0.465
+ +# 5.2 Robustness of SPARO's slot structure + +We further utilize the notion of post-hoc concept selection to study the robustness of SPARO's representational structure, and compare it with that of standard encodings. We set up the experiment as an inverse of Sec. 5.1 — we perform concept selection using the smaller SugarCrepe benchmark and evaluate the resulting encodings on real-world benchmarks. SugarCrepe contains 7,512 example pairs of positive and negative captions, differing only in compositional interpretation in 7 ways, over a set of 1,561 images. Compositionality is a shared high-level notion present in most real-world data, yet there is much information that exists in data beyond what is necessary to solve the SugarCrepe tasks. The small size of the SugarCrepe as a training dataset, together with the ubiquitous nature of compositionality in machine learning tasks, makes it an ideal candidate to overfit concept selection to for stress-testing the representational structure of SPARO by evaluating on other tasks. + +We train a global mask $\pmb{m} \in [0,1]^M$ of the image encodings $\pmb{y} \in \mathbb{R}^M$ to maximize SugarCrepe performance, and evaluate the masked encodings $\pmb{m} \odot \pmb{y}$ on VTAB tasks and ImageNet for zero-shot classification. We train masks for encodings of frozen $\mathrm{CLIP}^{32} + \mathrm{GAP}$ and $\mathrm{CLIP}^{32} + \mathrm{SPARO}$ models trained on CC15M with $M = 512$ and $M = 4096$ each. For the SPARO models, we use $L = 64$ and $V = M / L$ . For each of these settings, we consider two types of masking: per-dimension and per-slot. To compare the $\mathrm{CLIP}^{32} + \mathrm{GAP}$ encodings on per-slot masking, we consider their $L$ contiguous equal-sized partitions as the slot encodings. We provide full implementation details of our mask training setup in Appendix C.1. + +Table 6: Ablation experiments showing benefits of SPARO's attention bottleneck (Cross-attention is 'Separate-head') and cross-modal attention alignment (Slot-wise LayerNorm and projection are 'Yes' or 'N/A') with CC15M-trained CLIP, CLIP+GAP, CLIP+SPARO, and ablated variants of CLIP+SPARO that include multi-head cross-attention and the attentional pooler (AttPool) [39][66]. + +
ModelReplace last blockCross- attentionSlot-wise LayerNormSlot-wise projectionParamsFLOPS (×109)ImageNet zero-shot
CLIP32 (C)N/AN/AN/AN/A151M7.40.329
C+GAP151M7.40.345
C+AttPoolNoNo440M8.90.312
Multi-headYes306M8.80.344
YesNo440M8.90.338
Ablated C+SPARONoNoNo306M8.80.359
NoYes306M8.80.344
Separate-headNo172M8.70.358
YesNo306M8.80.344
Yes172M8.70.374
C+SPAROYesN/AN/A161M8.00.372
151M7.40.370
+ +We expect the compositionality of the data to be encoded in concepts that are useful for real-world downstream benchmarks. Therefore, the aggressive selection of only the concepts that perform well on a compositionality benchmark should not significantly impact downstream performance. However, since SugarCrepe does not represent all possible compositional manipulations, we expect to see a drop in performance on some tasks due to overfitting. We present our results in Tab. 5. We find that when the mask structure is aligned with the slot structure of SPARO, overfitting concept selection improves the average performance on VTAB tasks, and incurs only a minimal drop in performance on ImageNet. Furthermore, even in settings where the masks are not trained to align with SPARO's slot structure, we see that SPARO remains more resilient to overfitting than standard encodings in a majority of settings. + +# 5.3 Ablating the bottleneck and cross-modal alignment of SPARO + +We evaluate the generalization benefits of the separate-head attention bottleneck and of aligning the SPARO attention heads between the modalities in CLIP. We set up SPARO on one end of the ablation spectrum and the cross-attention readout of attentional pooler (AttPool) [39,66] to the other. AttPool uses multi-head attention to attend to the backbone encoder's outputs using embedded learned queries whereas SPARO uses separate single-head attention mechanisms. Additionally, AttnPool performs layer normalization [2] followed by a linear projec + +tion on the output of attention to produce the output encoding, whereas SPARO directly uses the output of separate-head attention as the encoding. + +We train $\mathrm{CLIP}^{32}$ , $\mathrm{CLIP}^{32} + \mathrm{GAP}$ , and variants of $\mathrm{CLIP}^{32} + \mathrm{SPARO}$ on CC15M with $L = 128$ , $V = 64$ and evaluate for ImageNet zero-shot accuracy and model size. We start by considering the choice of either separate-head or multi-head cross-attention over the backbone outputs. For each cross-attention setting, we apply layer normalization and linear projection operations to the attention module's outputs, similar to AttnPool. However, we additionally also consider slot-wise variants of these operations. Here, a slot-wise operation $f_{\phi}$ is one that acts separately on each input slot $\mathbf{y}_l$ using the same parameters $\phi$ , and produces the slot-structured output $\operatorname{concat}(f_{\phi}(\mathbf{y}_1), \ldots, f_{\phi}(\mathbf{y}_L))$ . When applied to an input that is not slot-structured, we assume a slot structure by partitioning the input into $L$ contiguous equal-sized partitions. With SPARO, using both slot-wise layer normalization and slot-wise projection enables each read-out attention head from one modality to align with the corresponding head from the other without being influenced by the other heads. + +We present our results in Tab. 6, showing a clear advantage of SPARO over AttnPool. Additionally, both the choices of separate-head attention over multi-head and slot-wise operations over the alternatives result in improved generalization, reduced model size, and fewer floating-point operations per second (FLOPS). While the former supports the effectiveness of the separate-head attention bottleneck, the latter validates the benefits of the prior for shared separately-attendable concepts between the modalities enabled by SPARO. + +# 5.4 Visualizations + +To conclude our analysis, we qualitatively visualize the concepts represented by SPARO using the $\mathrm{CLIP}^{16} + \mathrm{SPARO}$ model trained on CC15M from Sec. 4.1. A limitation of visualizing attended positions is that it only provides partial insight into the represented concept — we can tell where the information is taken from, but not how the information is used. Different SPARO slots can have similar attention maps for the same sample, but reveal different semantics when comparing patterns across other images. For instance, a slot that attends to animals (e.g. dogs, cats, elephants) and another that attends to transportation (e.g. trains, boats, cars) can both have the same attention mask over patches and words for a horse being used for transportation. Furthermore, not all SPARO concepts align between the two modalities for every sample due to information being present in one that is not possible to infer from the other. + +To ease our interpretation of attended concepts, we choose concepts to visualize that have sharp attention in the text modality, attend to non-overlapping text tokens, and have cosine similarity more than 0.75 with the corresponding image slots. We present examples of these filtered concepts in Fig. 3. In this example, we see that SPARO is capable of attending to different concepts such as subject of the scene, the activity represented, and the surrounding location, for both vision and text. We provide additional visualizations in Appendix B.7 + +![](images/6d5b27494b94184ef7d1aa9cae2e65816ac98d50b68948dfbb394ea0a967ead7.jpg) +A herd of cattle walking down a road being followed by a cowboy + +![](images/8cb54f7ee842e079c7bdff05d550db44d625d160abc073d1e0411e8a8db43392.jpg) +everal surfboards standing in a rowon the beach + +![](images/21d9354066660f9d4080f3b5ad4079f49cb8e16edbdf804d4714ddf10a136e76.jpg) +Two people riding horses on a rock path + +![](images/9906ef39be979a259828797000bd8b106e07656c9f3ec60edda9fa1854d03000.jpg) +A man sitting alone on a park bench in a park + +![](images/a27332458870ffcd532900de9e89d83dd88adb3042e7bc50ad652b5ade9bfcb4.jpg) +A herd of cattle walking down a road being followed by a cowboy + +![](images/89e821ecd4948ab4e7c058dd0c00e29ec8947994e091e4110b486c16a34cc90e.jpg) +Several surfboards standing in a row on the beach + +![](images/f8ed22736767397e70be98419c26e2ca1b18f33ba4a0d39e4776c6859bbbfce0.jpg) +Two people riding horses on a rock path + +![](images/e0545f5a8b26e7a38db18fe56e1d02f499267c9ba8948fe8a1b6ebc63c93affe.jpg) +A man sitting alone on a park bench in a park + +![](images/76ffc4830bb9abc250a8810893259e3ff5bdee2625a556360fac8f07105caf6b.jpg) +A herd of cattle walking down a road being followed by a cowboy +Fig. 3: Visualizing of the attended image and text positions for three SPARO slots (one per row) across four examples (one per column) from MS COCO. We surmise that the SPARO concepts from top to bottom represent the subject, activity, and location. + +![](images/6a6f91537768480583e46fbc5fa75e1437235fbbfcd3d69027d0069e56974497.jpg) +Several surfboards standing in a row on the beach + +![](images/f63e2bc8b62bc21dc527adf5d5b2e2907e48aeaac7ced2e0fd7301541bc48c03.jpg) +Two people riding horses on a rock path + +![](images/b51dd07a94b93dc37c46b0ad11824a65df7d64e959c852ac62df0a0872e0e203.jpg) +A man sitting alone on a park bench in a park + +# 6 Discussion + +Humans utilize a prior for compositionality to coherently represent different subsets of salient aspects of the world. We can enrich our comprehension by selectively attending to new aspects or specialize to a task by filtering out distracting ones. We introduce SPARO with a goal of imparting a similar prior to transformers in representation learning frameworks by partitioning encodings into separately-attended concepts. Although we see evidence of disentanglement through positive results of post-hoc concept selection and visualization of attention maps, we do not impose any explicit independence or disentanglement constraints. In our representation learning frameworks, the pressure for learning distinct concepts arises from the need to explain the factors of variation through the training objective with only the features that can be accumulated by a single head of attention with embedded queries and limited dimensionality. However, auxiliary training objectives to impose stronger distributional conditions on the learned concepts can be explored as promising future directions. Beyond pretraining, we highlight that more sophisticated post-hoc concept selection approaches than explored in our work, like using human interaction or set cover algorithms, can further improve the downstream utility of SPARO. + +# Acknowledgments + +This research was funded by Sony and enabled in part by compute resources provided by the Digital Research Alliance of Canada, Mila, and Sony. + +# References + +1. Aydemir, G., Xie, W., Güney, F.: Self-supervised object-centric learning for videos. arXiv preprint arXiv:2310.06907 (2023) +2. Ba, J.L., Kiros, J.R., Hinton, G.E.: Layer normalization. arXiv preprint arXiv:1607.06450 (2016) +3. Bahdanau, D., Cho, K., Bengio, Y.: Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473 (2014) +4. Barbu, A., Mayo, D., Alverio, J., Luo, W., Wang, C., Gutfreund, D., Tenenbaum, J., Katz, B.: Objectnet: A large-scale bias-controlled dataset for pushing the limits of object recognition models. Advances in neural information processing systems 32 (2019) +5. Beattie, C., Leibo, J.Z., Teplyashin, D., Ward, T., Wainwright, M., Kuttler, H., Lefrancq, A., Green, S., Valdés, V., Sadik, A., et al.: Deepmind lab. arXiv preprint arXiv:1612.03801 (2016) +6. Boff, K.R., Kaufman, L., Thomas, J.P.: Handbook of perception and human performance, vol. 1. Wiley New York (1986) +7. Brady, J., Zimmermann, R.S., Sharma, Y., Schölkopf, B., von Kügelgen, J., Brendel, W.: Provably learning object-centric representations. arXiv preprint arXiv:2305.14229 (2023) +8. Burgess, C.P., Matthew, L., Watters, N., Kabra, R., Higgins, I., Botvinick, M., Lerchner, A.: Monet: Unsupervised scene decomposition and representation. arXiv preprint arXiv:1901.11390 (2019) +9. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021) +0. Chang, H.S., Sun, R.Y., Ricci, K., McCallum, A.: Multi-cls bert: An efficient alternative to traditional ensembling. arXiv preprint arXiv:2210.05043 (2022) +1. Changpinyo, S., Sharma, P., Ding, N., Soricut, R.: Conceptual 12M: Pushing web-scale image-text pre-training to recognize long-tail visual concepts. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3558-3568 (2021) +2. Chen, X., Xie, S., He, K.: An empirical study of training self-supervised vision transformers. arXiv preprint arXiv:2104.02057 (2021) +3. Cheng, G., Han, J., Lu, X.: Remote sensing image scene classification: Benchmark and state of the art. Proceedings of the IEEE 105(10), 1865-1883 (2017) +4. Cimpoi, M., Maji, S., Kokkinos, I., Mohamed, S., Vedaldi, A.: Describing textures in the wild. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3606-3613 (2014) +5. Colby, C.L., Goldberg, M.E.: Space and attention in parietal cortex. Annual review of neuroscience 22(1), 319-349 (1999) +6. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009) +7. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: BERT: Pre-training of deep bidirectional transformers for language understanding. In: North American Chapter of the Association for Computational Linguistics (2019) +8. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) + +19. Dugas, E., Jared, Jorge, Cukierski, W.: Diabetic retinopathy detection (2015), https://kaggle.com/competitions/diabetic-retinopathy-detection +20. Engelcke, M., Kosiorek, A.R., Jones, O.P., Posner, I.: Genesis: Generative scene inference and sampling with object-centric latent representations. arXiv preprint arXiv:1907.13052 (2019) +21. Eslami, S., Heess, N., Weber, T., Tassa, Y., Szepesvari, D., Hinton, G.E., et al.: Attend, infer, repeat: Fast scene understanding with generative models. Advances in neural information processing systems 29 (2016) +22. Fei-Fei, L., Fergus, R., Perona, P.: One-shot learning of object categories. IEEE transactions on pattern analysis and machine intelligence 28(4), 594-611 (2006) +23. Geiger, A., Lenz, P., Stiller, C., Urtasun, R.: Vision meets robotics: The kitti dataset. The International Journal of Robotics Research 32(11), 1231-1237 (2013) +24. Goyal, A., Didolkar, A., Ke, N.R., Blundell, C., Beaudoin, P., Heess, N., Mozer, M.C., Bengio, Y.: Neural production systems. Advances in Neural Information Processing Systems 34, 25673-25687 (2021) +25. Goyal, A., Lamb, A., Gampa, P., Beaudoin, P., Blundell, C., Levine, S., Bengio, Y., Mozer, M.C.: Factorizing declarative and procedural knowledge in structured, dynamical environments. In: International Conference on Learning Representations (2020) +26. Goyal, A., Lamb, A., Hoffmann, J., Sodhani, S., Levine, S., Bengio, Y., Scholkopf, B.: Recurrent independent mechanisms. arXiv preprint arXiv:1909.10893 (2019) +27. Greff, K., Kaufman, R.L., Kabra, R., Watters, N., Burgess, C., Zoran, D., Matthew, L., Botvinick, M., Lerchner, A.: Multi-object representation learning with iterative variational inference. In: International Conference on Machine Learning. pp. 2424-2433. PMLR (2019) +28. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) +29. Helber, P., Bischke, B., Dengel, A., Borth, D.: Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 12(7), 2217-2226 (2019) +30. Hendrycks, D., Basart, S., Mu, N., Kadavath, S., Wang, F., Dorundo, E., Desai, R., Zhu, T., Parajuli, S., Guo, M., et al.: The many faces of robustness: A critical analysis of out-of-distribution generalization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8340-8349 (2021) +31. Hendrycks, D., Zhao, K., Basart, S., Steinhardt, J., Song, D.: Natural adversarial examples. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15262-15271 (2021) +32. Hodosh, M., Young, P., Hockenmaier, J.: Framing image description as a ranking task: Data, models and evaluation metrics. Journal of Artificial Intelligence Research 47, 853-899 (2013) +33. Hsieh, C.Y., Zhang, J., Ma, Z., Kembhavi, A., Krishna, R.: Sugarcrepe: Fixing hackable benchmarks for vision-language compositionality. Advances in neural information processing systems (2023) +34. Ilharco, G., Wortsman, M., Wightman, R., Gordon, C., Carlini, N., Taori, R., Dave, A., Shankar, V., Namkoong, H., Miller, J., Hajishirzi, H., Farhadi, A., Schmidt, L.: OpenCLIP. https://github.com/mlfoundations/open Clip (Jul 2021). https://doi.org/10.5281/zenodo.5143773 + +35. Johnson, J., Hariharan, B., van der Maaten, L., Fei-Fei, L., Lawrence Zitnick, C., Girshick, R.: Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2901-2910 (2017) +36. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images. Technical report (2009) +37. Lavoie, S., Tsirigotis, C., Schwarzer, M., Vani, A., Noukhovitch, M., Kawaguchi, K., Courville, A.: Simplicial embeddings in self-supervised learning and downstream classification. In: International Conference on Learning Representations (2023), https://openreview.net/forum?id=RWtGreRpovS +38. LeCun, Y., Huang, F.J., Bottou, L.: Learning methods for generic object recognition with invariance to pose and lighting. In: Proceedings of the 2004 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2004. CVPR 2004. vol. 2, pp. II-104. IEEE (2004) +39. Lee, J., Lee, Y., Kim, J., Kosiorek, A., Choi, S., Teh, Y.W.: Set transformer: A framework for attention-based permutation-invariant neural networks. In: International conference on machine learning. pp. 3744-3753. PMLR (2019) +40. Lin, T., Maire, M., Belongie, S.J., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, C.L.: Microsoft COCO: common objects in context. In: Fleet, D.J., Pajdla, T., Schiele, B., Tuytelaars, T. (eds.) Computer Vision - ECCV 2014 - 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V. Lecture Notes in Computer Science, vol. 8693, pp. 740-755. Springer (2014). https://doi.org/10.1007/978-3-319-10602-1_48, https://doi.org/10.1007/978-3-319-10602-1%5F48 +41. Locatello, F., Weissenborn, D., Unterthiner, T., Mahendran, A., Heigold, G., Uszkoreit, J., Dosovitskiy, A., Kipf, T.: Object-centric learning with slot attention. Advances in Neural Information Processing Systems 33, 11525-11538 (2020) +42. Ma, Z., Hong, J., Gul, M.O., Gandhi, M., Gao, I., Krishna, R.: Crepe: Can vision-language foundation models reason compositionally? In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10910-10921 (2023) +43. Mansouri, A., Hartford, J., Zhang, Y., Bengio, Y.: Object centric architectures enable efficient causal representation learning. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=r9FsiXZxZt +44. Martinez, A.: Putting spatial attention on the map: timing and localization of stimulus selection processes in striate and extrastriate visual areas. Vision Res 41, 1437-1457 (2001) +45. Matthew, L., Higgins, I., Hassabis, D., Lerchner, A.: dsprites: Disentanglement testing sprites dataset. https://github.com/deepmind/dsprites-dataset/ (2017) +46. Nilsback, M.E., Zisserman, A.: Automated flower classification over a large number of classes. In: 2008 Sixth Indian conference on computer vision, graphics & image processing. pp. 722-729. IEEE (2008) +47. O'Connor, D.H., Fukui, M.M., Pinsk, M.A., Kastner, S.: Attention modulates responses in the human lateral geniculate nucleus. Nature neuroscience 5(11), 1203-1209 (2002) +48. Parkhi, O.M., Vedaldi, A., Zisserman, A., Jawahar, C.: Cats and dogs. In: 2012 IEEE conference on computer vision and pattern recognition. pp. 3498-3505. IEEE (2012) + +49. Qian, R., Ding, S., Liu, X., Lin, D.: Semantics meets temporal correspondence: Self-supervised object-centric learning in videos. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16675-16687 (2023) +50. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) +51. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners. OpenAI blog 1(8), 9 (2019) +52. Ray, A., Radenovic, F., Dubey, A., Plummer, B.A., Krishna, R., Saenko, K.: Cola: A benchmark for compositional text-to-image retrieval. In: Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track (2023) +53. Recht, B., Roelofs, R., Schmidt, L., Shankar, V.: Do imagenet classifiers generalize to imagenet? In: International conference on machine learning. pp. 5389-5400. PMLR (2019) +54. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114 (2021) +55. Scott, W.A.: Cognitive complexity and cognitive flexibility. Sociometry pp. 405-414 (1962) +56. Seitzer, M., Horn, M., Zadaianchuk, A., Zietlow, D., Xiao, T., Simon-Gabriel, C.J., He, T., Zhang, Z., Scholkopf, B., Brox, T., et al.: Bridging the gap to real-world object-centric learning. arXiv preprint arXiv:2209.14860 (2022) +57. Sharma, P., Ding, N., Goodman, S., Soricut, R.: Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In: Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics, ACL 2018, Melbourne, Australia, July 15-20, 2018, Volume 1: Long Papers. pp. 2556-2565. Association for Computational Linguistics (2018) +58. Thrush, T., Jiang, R., Bartolo, M., Singh, A., Williams, A., Kiela, D., Ross, C.: Winoground: Probing vision and language models for visio-linguistic compositionality. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5238-5248 (2022) +59. Treisman, A.M., Gelade, G.: A feature-integration theory of attention. Cognitive psychology 12(1), 97-136 (1980) +60. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017) +61. Veeling, B.S., Linmans, J., Winkens, J., Cohen, T., Welling, M.: Rotation equivariant cnns for digital pathology. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2018: 21st International Conference, Granada, Spain, September 16-20, 2018, Proceedings, Part II 11. pp. 210-218. Springer (2018) +62. Wang, H., Ge, S., Lipton, Z., Xing, E.P.: Learning robust global representations by penalizing local predictive power. In: Advances in Neural Information Processing Systems. pp. 10506-10518 (2019) +63. Xiao, J., Hays, J., Ehinger, K.A., Oliva, A., Torralba, A.: Sun database: Large-scale scene recognition from abbey to zoo. In: 2010 IEEE computer society conference on computer vision and pattern recognition. pp. 3485-3492. IEEE (2010) +64. Xu, L., Ouyang, W., Bennamoun, M., Boussaid, F., Xu, D.: Multi-class token transformer for weakly supervised semantic segmentation. In: Proceedings of the + +IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4310-4319 (2022) +65. Young, P., Lai, A., Hodosh, M., Hockenmaier, J.: From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics 2, 67-78 (2014) +66. Yu, J., Wang, Z., Vasudevan, V., Yeung, L., Seyedhosseini, M., Wu, Y.: Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917 (2022) +67. Yuksekgonul, M., Bianchi, F., Kalluri, P., Jurafsky, D., Zou, J.: When and why vision-language models behave like bags-of-words, and what to do about it? In: The Eleventh International Conference on Learning Representations (2023) +68. Yuval, N.: Reading digits in natural images with unsupervised feature learning. In: Proceedings of the NIPS Workshop on Deep Learning and Unsupervised Feature Learning (2011) +69. Zadaianchuk, A., Kleindessner, M., Zhu, Y., Locatello, F., Brox, T.: Unsupervised semantic segmentation with self-supervised object-centric representations. arXiv preprint arXiv:2207.05027 (2022) +70. Zhai, X., Puigcerver, J., Kolesnikov, A., Ruyssen, P., Riquelme, C., Lucic, M., Djolonga, J., Pinto, A.S., Neumann, M., Dosovitskiy, A., et al.: A large-scale study of representation learning with the visual task adaptation benchmark. arXiv preprint arXiv:1910.04867 (2019) +71. Zhang, Y., Hare, J., Prugel-Bennett, A.: Deep set prediction networks. Advances in Neural Information Processing Systems 32 (2019) +72. Zhao, T., Zhang, T., Zhu, M., Shen, H., Lee, K., Lu, X., Yin, J.: Vl-checklist: Evaluating pre-trained vision-language models with objects, attributes and relations. arXiv preprint arXiv:2207.00221 (2022) \ No newline at end of file diff --git a/2024/SPARO_ Selective Attention for Robust and Compositional Transformer Encodings for Vision/images.zip b/2024/SPARO_ Selective Attention for Robust and Compositional Transformer Encodings for Vision/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..493227c8dfd127c8890326fe0b50e917cd1fd932 --- /dev/null +++ b/2024/SPARO_ Selective Attention for Robust and Compositional Transformer Encodings for Vision/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1c225cf76fd2c0b47afbbfadf6b370d64ebe42ef97c1de51e582d0d2d9d35f18 +size 605921 diff --git a/2024/SPARO_ Selective Attention for Robust and Compositional Transformer Encodings for Vision/layout.json b/2024/SPARO_ Selective Attention for Robust and Compositional Transformer Encodings for Vision/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..cdc2f6ebf3773bd12e8c82e5d2f8b86d8e4dd160 --- /dev/null +++ b/2024/SPARO_ Selective Attention for Robust and Compositional Transformer Encodings for Vision/layout.json @@ -0,0 +1,10345 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 114, + 165, + 135 + ], + "blocks": [ + { + "bbox": [ + 138, + 114, + 165, + 135 + ], + "lines": [ + { + "bbox": [ + 138, + 114, + 165, + 135 + ], + "spans": [ + { + "bbox": [ + 138, + 114, + 165, + 135 + ], + "type": "image", + "image_path": "fcf0f2e780151c3244b95b9fe666092896fd1f14ea02089ec12accca332f2455.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 167, + 120, + 475, + 136 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 120, + 475, + 136 + ], + "spans": [ + { + "bbox": [ + 167, + 120, + 475, + 136 + ], + "type": "text", + "content": "SPARO: Selective Attention for Robust and" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 137, + 139, + 477, + 155 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 139, + 477, + 155 + ], + "spans": [ + { + "bbox": [ + 137, + 139, + 477, + 155 + ], + "type": "text", + "content": "Compositional Transformer Encodings for Vision" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 144, + 175, + 470, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 175, + 470, + 201 + ], + "spans": [ + { + "bbox": [ + 144, + 175, + 470, + 201 + ], + "type": "text", + "content": "Ankit Vani1*, Bac Nguyen2, Samuel Lavoie1, Ranjay Krishna3, and Aaron Courville1,4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 242, + 209, + 370, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 209, + 370, + 220 + ], + "spans": [ + { + "bbox": [ + 242, + 209, + 370, + 220 + ], + "type": "text", + "content": "1 Mila, Université de Montréal" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 284, + 221, + 328, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 221, + 328, + 232 + ], + "spans": [ + { + "bbox": [ + 284, + 221, + 328, + 232 + ], + "type": "text", + "content": "2 Sony AI" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 167, + 232, + 446, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 232, + 446, + 243 + ], + "spans": [ + { + "bbox": [ + 167, + 232, + 446, + 243 + ], + "type": "text", + "content": "3 University of Washington, Allen Institute for Artificial Intelligence" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 268, + 243, + 346, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 268, + 243, + 346, + 253 + ], + "spans": [ + { + "bbox": [ + 268, + 243, + 346, + 253 + ], + "type": "text", + "content": "4 CIFAR AI Chair" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 159, + 286, + 455, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 286, + 455, + 529 + ], + "spans": [ + { + "bbox": [ + 159, + 286, + 455, + 529 + ], + "type": "text", + "content": "Abstract. Selective attention helps us focus on task-relevant aspects in the constant flood of our sensory input. This constraint in our perception allows us to robustly generalize under distractions and to new compositions of perceivable concepts. Transformers employ a similar notion of attention in their architecture, but representation learning models with transformer backbones like CLIP and DINO often fail to demonstrate robustness and compositionality. We highlight a missing architectural prior: unlike human perception, transformer encodings do not separately attend over individual concepts. In response, we propose SPARO, a readout mechanism that partitions encodings into separately-attended slots, each produced by a single attention head. Using SPARO with CLIP imparts an inductive bias that the vision and text modalities are different views of a shared compositional world with the same corresponding concepts. Using SPARO, we demonstrate improvements on downstream recognition, robustness, retrieval, and compositionality benchmarks with CLIP (up to " + }, + { + "bbox": [ + 159, + 286, + 455, + 529 + ], + "type": "inline_equation", + "content": "+14\\%" + }, + { + "bbox": [ + 159, + 286, + 455, + 529 + ], + "type": "text", + "content": " for ImageNet, " + }, + { + "bbox": [ + 159, + 286, + 455, + 529 + ], + "type": "inline_equation", + "content": "+4\\%" + }, + { + "bbox": [ + 159, + 286, + 455, + 529 + ], + "type": "text", + "content": " for SugarCrepe), and on nearest neighbors and linear probe for ImageNet with DINO (" + }, + { + "bbox": [ + 159, + 286, + 455, + 529 + ], + "type": "inline_equation", + "content": "+3\\%" + }, + { + "bbox": [ + 159, + 286, + 455, + 529 + ], + "type": "text", + "content": " each). We also showcase a powerful ability to intervene and select individual SPARO concepts to further improve downstream task performance (up from " + }, + { + "bbox": [ + 159, + 286, + 455, + 529 + ], + "type": "inline_equation", + "content": "+4\\%" + }, + { + "bbox": [ + 159, + 286, + 455, + 529 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 159, + 286, + 455, + 529 + ], + "type": "inline_equation", + "content": "+9\\%" + }, + { + "bbox": [ + 159, + 286, + 455, + 529 + ], + "type": "text", + "content": " for SugarCrepe) and use this ability to study the robustness of SPARO's representation structure. Finally, we provide insights through ablation experiments and visualization of learned concepts." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 160, + 538, + 443, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 538, + 443, + 550 + ], + "spans": [ + { + "bbox": [ + 160, + 538, + 443, + 550 + ], + "type": "text", + "content": "Keywords: Selective attention " + }, + { + "bbox": [ + 160, + 538, + 443, + 550 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 538, + 443, + 550 + ], + "type": "text", + "content": " Slot representations " + }, + { + "bbox": [ + 160, + 538, + 443, + 550 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 538, + 443, + 550 + ], + "type": "text", + "content": " Transformers" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 571, + 231, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 571, + 231, + 583 + ], + "spans": [ + { + "bbox": [ + 132, + 571, + 231, + 583 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 597, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 597, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 597, + 482, + 647 + ], + "type": "text", + "content": "Selective attention is an intrinsic property of human perception [15,44,47], enabling people to focus on task-relevant aspects of their surroundings while tuning out the rest [55]. For instance, it allows a driver to focus on traffic signals, road signs, and other vehicles, or an individual searching for their friend" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 337, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 337, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 337, + 665 + ], + "type": "text", + "content": "* Correspondence at: ankit.vani@umontreal.ca." + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "in a crowd to attend to details like people's heights and clothing. More broadly, it empowers us to comprehend the vast complexity of our world with limited mental resources [6] by constraining our perception to one concept at a time. When a task requires information for multiple concepts (e.g., looking at traffic signals and pedestrians), this constraint calls for sequentially attending to separate salient aspects of our stimuli [59]. The cognitive bottleneck of selective attention, with the ability to separately represent concepts from our environment, enables human perception to remain insensitive to irrelevant distractions and generalize to new compositions of perceivable concepts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 224, + 482, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 380 + ], + "type": "text", + "content": "In machine learning, the notion of selective attention is party emulated by the attention mechanism 3, which is a core component of transformers 60 such as the vision transformer (ViT) 18. However, transformer encodings learnt using approaches like CLIP 50 and DINO 9 still struggle with robustness 30, 31, 62, 70 and compositional generalization 33, 42, 52, 58, 67, 72. For instance, despite CLIP emerging as the de facto backbone for many vision tasks, it can perform close to chance when evaluated for compositionality 42. While the attention mechanism equips these models with the ability to emphasize relevant aspects of their intermediate representations, it does not endow them with the means to produce encodings for separately-attended concepts. Consequently, transformer-based representation learning models lack an important prior from human cognition: that perception of a complex input can be broken down into perception of its salient concepts." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 381, + 482, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 381, + 482, + 523 + ], + "spans": [ + { + "bbox": [ + 130, + 381, + 482, + 523 + ], + "type": "text", + "content": "This limitation deprives downstream tasks easy access to task-relevant aspects of the data's underlying structure. In a task like ImageNet classification, where most images contain one primary object of interest (e.g., \"dog\"), information like relationships between objects and attributes of the background are irrelevant and can lead to overfitting if considered. However, these details can become crucial to attain high performance in other tasks like image retrieval (e.g., \"dog chasing a frisbee in a park\"). Similar to the generalization benefits of selective attention in human perception, encodings that allow easy separation of attended concepts in the data can generalize to a diverse set of downstream tasks. Without additional constraints, however, the emergence of such compositional structure in encodings is unlikely, especially in the common setting of training on noisy internet data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 524, + 482, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 524, + 482, + 644 + ], + "spans": [ + { + "bbox": [ + 130, + 524, + 482, + 644 + ], + "type": "text", + "content": "We propose SPARO (Separate-head attention read-out) as an improved read-out mechanism for transformer encoders in vision inspired by selective attention. SPARO replaces the last transformer block to provide a mechanism for partitioning encodings into slots of separately-attended concepts. We design each slot encoding to be a low-dimensional result of a single-head attention operation, producing a bottleneck that encourages each slot to \"selectively attend.\" In this mechanism, producing multiple slots can be interpreted as selective attention over different concepts occurring in parallel. When training CLIP, SPARO imparts an inductive bias that both modalities are views onto a shared compositional world with the corresponding slots of both encoders representing its concepts." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "A. Vani et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 652, + 375, + 667 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 652, + 375, + 667 + ], + "spans": [ + { + "bbox": [ + 133, + 652, + 375, + 667 + ], + "type": "text", + "content": "5 Source code: https://github.com/ankitkv/sparo-clip" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "In contrast, standard CLIP merely embeds the encodings of both modalities in a shared vector space without imposing any additional structure." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 332 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 332 + ], + "type": "text", + "content": "Training SPARO using CLIP and DINO exhibits improved generalization, robustness, and compositionality while retaining the same model size. Our experiments show that using CLIP with SPARO improves generalization for zero-shot recognition, robustness, retrieval, and compositionality (e.g., +14% for ImageNet, +4% for SugarCrepe, when trained with 15M examples; +3% each when trained with 400M examples). We also demonstrate improved linear probe performance of CLIP (+10% trained with 15M examples; +2% with 400M examples), as well as both linear probe and nearest neighbors performance of DINO (+3% each). We showcase how SPARO's representational structure enables the ability to manually intervene and select relevant concepts for downstream tasks, leading to improved compositionality and generalization (up from +4% to +9% for SugarCrepe trained with 15M examples; +3% to +6% with 400M examples). We then study the robustness of this structure, validate the choice of separate-head attention through ablations, and provide insights into the effect of the number and dimensionality of concepts. Finally, we provide visualizations for some of the concepts SPARO learns to attend to." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 351, + 234, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 351, + 234, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 351, + 234, + 363 + ], + "type": "text", + "content": "2 Related work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 377, + 480, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 377, + 480, + 401 + ], + "spans": [ + { + "bbox": [ + 130, + 377, + 480, + 401 + ], + "type": "text", + "content": "We situate our work amongst other transformer read-out mechanisms and slot representation learning methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 414, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 482, + 594 + ], + "type": "text", + "content": "Transformer read-outs. For transformer [60] decoders with causal masking, the last hidden state, corresponding to the end of sequence (EOS) input token, is used as the output representation. BERT [17] proposed inserting a special CLS token before input sequences for transformer encoders where information from all positions is gathered to produce the encoding. Using the CLS token output is also popular for Vision transformer (ViT) [9, 12, 18] backbones, along with alternatives such as global average pooling [12, 64]. In the context of language, Multi-CLS BERT [10] argues that an input text sequence can have multiple facets, and that samples can be similar along one facet and dissimilar along another. The authors aim to represent these facets by adding multiple CLS tokens to the input sequence. Using separate attention heads for each SPARO slot can also encourage attending to different facets of the inputs, but without requiring expensive accumulations through each layer of the backbone. By keeping SPARO independent of the backbone, we also allow for its easy potential placement on top of frozen pre-trained models without needing to train for added CLS tokens." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "SPARO is most related to the attentional pooler [39,66] architecture, which uses multi-head attention with embedded learned queries to produce output encodings. However, SPARO employs a critical structural constraint of not mixing the information from separate attention heads, a design choice we justify empirically in Sec. 5.3. Slot attention [41] also produces slot encodings through the attention mechanism, but rather than having the input positions compete" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 186, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 186, + 91, + 448, + 102 + ], + "type": "text", + "content": "SPARO: Selective Attention in Transformer Encodings for Vision" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": "for each slot, the slots compete for each input position. Slot attention also proposes an expensive iterative read-out mechanism where the role of each slot is determined by sampling from a per-sample Gaussian distribution, which can be limiting for large real-world datasets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 177, + 483, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 177, + 483, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 177, + 483, + 357 + ], + "type": "text", + "content": "Slot representations. Object-centric slot representations have been studied extensively in the literature [7,8,20,21,24-27,41,43,71]. Often, these representations are expected to represent predefined objects and attributes, and evaluated in small-scale or synthetic settings. Recently, the use of slot attention on features produced by a learned encoder [1,49,56], and utilizing unsupervised saliency masks for object discovery [69] has enabled unsupervised learning of object-centric representations in larger scale settings. Related to these ideas, SPARO attends to features produced by a transformer using a single attention head per slot, and each head is capable of learning saliency maps without supervision [9]. However, we do not constrain our slots to correspond to predefined types of objects in the data, nor do we impose any independence conditions between the slots. Self-supervised simplicial embeddings (SEM) [37], with only a soft-discrete slot structure constraint, enable improved generalization on downstream tasks. SPARO replaces the separation of softmax application between SEM slots with a separation of underlying attention operations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 377, + 202, + 389 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 377, + 202, + 389 + ], + "spans": [ + { + "bbox": [ + 132, + 377, + 202, + 389 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 403, + 482, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 453 + ], + "type": "text", + "content": "Having motivated our approach, we now detail the design of SPARO. We start by introducing the relevant notation, then discuss the SPARO module. Finally, we discuss the addition of SPARO in CLIP and DINO, and the applicability of SPARO beyond these settings." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 471, + 206, + 482 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 471, + 206, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 471, + 206, + 482 + ], + "type": "text", + "content": "3.1 Notation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 492, + 482, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 492, + 482, + 589 + ], + "spans": [ + { + "bbox": [ + 130, + 492, + 482, + 589 + ], + "type": "text", + "content": "We envision SPARO as a special attention-based layer that modifies the top block of a transformer encoder [51,60] backbone (including the ViT [18]). Transformers take as input a sequence " + }, + { + "bbox": [ + 130, + 492, + 482, + 589 + ], + "type": "inline_equation", + "content": "\\pmb{x} = \\{\\pmb{x}_1,\\dots,\\pmb{x}_n\\}" + }, + { + "bbox": [ + 130, + 492, + 482, + 589 + ], + "type": "text", + "content": " of length " + }, + { + "bbox": [ + 130, + 492, + 482, + 589 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 492, + 482, + 589 + ], + "type": "text", + "content": ", and produce an output state per position " + }, + { + "bbox": [ + 130, + 492, + 482, + 589 + ], + "type": "inline_equation", + "content": "\\{h_1,\\ldots ,h_n\\}" + }, + { + "bbox": [ + 130, + 492, + 482, + 589 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 492, + 482, + 589 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i\\in \\mathbb{R}^{\\mathrm{inputdim}}" + }, + { + "bbox": [ + 130, + 492, + 482, + 589 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 492, + 482, + 589 + ], + "type": "inline_equation", + "content": "\\pmb{h}_i\\in \\mathbb{R}^d" + }, + { + "bbox": [ + 130, + 492, + 482, + 589 + ], + "type": "text", + "content": ", for any " + }, + { + "bbox": [ + 130, + 492, + 482, + 589 + ], + "type": "inline_equation", + "content": "i\\in [n]" + }, + { + "bbox": [ + 130, + 492, + 482, + 589 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 130, + 492, + 482, + 589 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 492, + 482, + 589 + ], + "type": "text", + "content": " the model width. Finally, a pooling operation reduces the sequence of output states to output encodings with a fixed size " + }, + { + "bbox": [ + 130, + 492, + 482, + 589 + ], + "type": "inline_equation", + "content": "\\pmb{y}\\in \\mathbb{R}^{M}" + }, + { + "bbox": [ + 130, + 492, + 482, + 589 + ], + "type": "text", + "content": ". Examples of pooling operations include attentional pooling [39], global average pooling, and extracting the CLS or EOS token representations for images or text respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 608, + 373, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 608, + 373, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 608, + 373, + 620 + ], + "type": "text", + "content": "3.2 Separate-head attention read-out (SPARO)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "We replace the pooling operation of transformer encoders with a concatenation of outputs of " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " single-head attention mechanisms. SPARO acts on the transformer outputs " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\pmb{H} = \\{\\pmb{h}_1, \\dots, \\pmb{h}_n\\} \\in \\mathbb{R}^{n \\times d}" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " to produce the encoder output" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "A. Vani et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 482, + 255 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 255 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 255 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 255 + ], + "type": "image", + "image_path": "c0ab253b9f9f3a5a1694cc4ff2e7722630d79db6eaf7a0d69bb5d559a6a8165b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 262, + 482, + 361 + ], + "lines": [ + { + "bbox": [ + 130, + 262, + 482, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 482, + 361 + ], + "type": "text", + "content": "Fig. 1: Illustration of SPARO, a read-out mechanism that structures representations as collections of separately-attended concepts. Take a standard " + }, + { + "bbox": [ + 130, + 262, + 482, + 361 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 262, + 482, + 361 + ], + "type": "text", + "content": "-block transformer encoder (ViT here as an example), producing an encoding " + }, + { + "bbox": [ + 130, + 262, + 482, + 361 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 262, + 482, + 361 + ], + "type": "text", + "content": " through extraction of its CLS token output. We can replace the " + }, + { + "bbox": [ + 130, + 262, + 482, + 361 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 262, + 482, + 361 + ], + "type": "text", + "content": "th transformer block with the SPARO module (typically with equal or fewer parameters) to produce a SPARO encoding " + }, + { + "bbox": [ + 130, + 262, + 482, + 361 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 262, + 482, + 361 + ], + "type": "text", + "content": ", which is a concatenation of " + }, + { + "bbox": [ + 130, + 262, + 482, + 361 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 262, + 482, + 361 + ], + "type": "text", + "content": " SPARO slots. Each SPARO slot " + }, + { + "bbox": [ + 130, + 262, + 482, + 361 + ], + "type": "inline_equation", + "content": "y_l" + }, + { + "bbox": [ + 130, + 262, + 482, + 361 + ], + "type": "text", + "content": " is produced through single-head attention over the backbone outputs using an embedded query " + }, + { + "bbox": [ + 130, + 262, + 482, + 361 + ], + "type": "inline_equation", + "content": "q_l" + }, + { + "bbox": [ + 130, + 262, + 482, + 361 + ], + "type": "text", + "content": ". The value projection is a composition of slot-specific key projection parameterized by " + }, + { + "bbox": [ + 130, + 262, + 482, + 361 + ], + "type": "inline_equation", + "content": "K_l" + }, + { + "bbox": [ + 130, + 262, + 482, + 361 + ], + "type": "text", + "content": " and a slot-wise projection shared between all SPARO concepts parameterized by " + }, + { + "bbox": [ + 130, + 262, + 482, + 361 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 262, + 482, + 361 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 384, + 482, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 482, + 410 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 482, + 410 + ], + "type": "inline_equation", + "content": "\\pmb{y} = \\mathrm{concat}(\\pmb{y}_1, \\dots, \\pmb{y}_L) \\in \\mathbb{R}^{LV}" + }, + { + "bbox": [ + 130, + 384, + 482, + 410 + ], + "type": "text", + "content": ", where SPARO slot " + }, + { + "bbox": [ + 130, + 384, + 482, + 410 + ], + "type": "inline_equation", + "content": "\\pmb{y}_l \\in \\mathbb{R}^V" + }, + { + "bbox": [ + 130, + 384, + 482, + 410 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 384, + 482, + 410 + ], + "type": "inline_equation", + "content": "l \\in [L]" + }, + { + "bbox": [ + 130, + 384, + 482, + 410 + ], + "type": "text", + "content": ". We illustrate the architecture of SPARO in Fig. 1. Each SPARO slot is produced as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 226, + 426, + 481, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 226, + 426, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 226, + 426, + 481, + 453 + ], + "type": "interline_equation", + "content": "\\boldsymbol {y} _ {l} = \\boldsymbol {W} \\boldsymbol {K} _ {l} \\boldsymbol {H} ^ {\\intercal} \\text {s o f t m a x} \\left(\\frac {\\boldsymbol {H} \\boldsymbol {K} _ {l} ^ {\\intercal} \\boldsymbol {q} _ {l}}{\\sqrt {D}}\\right). \\tag {1}", + "image_path": "3f65ef3bdaaa4c0ad2ed8105df18ee2077ca4de7ff9d12d4b6863ffa66df00a7.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 461, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 521 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 130, + 461, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\pmb{K}_l \\in \\mathbb{R}^{D \\times d}" + }, + { + "bbox": [ + 130, + 461, + 482, + 521 + ], + "type": "text", + "content": " is a learned key projection weight, and " + }, + { + "bbox": [ + 130, + 461, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\pmb{q}_l \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 130, + 461, + 482, + 521 + ], + "type": "text", + "content": " is a learned query embedding for attention. We parameterize the value projection as " + }, + { + "bbox": [ + 130, + 461, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\pmb{W}\\pmb{K}_l" + }, + { + "bbox": [ + 130, + 461, + 482, + 521 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 461, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\pmb{W} \\in \\mathbb{R}^{V \\times D}" + }, + { + "bbox": [ + 130, + 461, + 482, + 521 + ], + "type": "text", + "content": " is learned parameter shared between all " + }, + { + "bbox": [ + 130, + 461, + 482, + 521 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 461, + 482, + 521 + ], + "type": "text", + "content": " slots. Decomposing the value projection in this manner helps reduce the number of model parameters, allowing a larger choice of " + }, + { + "bbox": [ + 130, + 461, + 482, + 521 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 461, + 482, + 521 + ], + "type": "text", + "content": " with fixed resources." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 522, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 482, + 605 + ], + "type": "text", + "content": "Generally, we pick " + }, + { + "bbox": [ + 130, + 522, + 482, + 605 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 522, + 482, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 522, + 482, + 605 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 522, + 482, + 605 + ], + "type": "text", + "content": " to be significantly smaller than the transformer width. In practice, we use the typical value of " + }, + { + "bbox": [ + 130, + 522, + 482, + 605 + ], + "type": "inline_equation", + "content": "D = 64" + }, + { + "bbox": [ + 130, + 522, + 482, + 605 + ], + "type": "text", + "content": " used in standard multihead attention modules, and set " + }, + { + "bbox": [ + 130, + 522, + 482, + 605 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 522, + 482, + 605 + ], + "type": "text", + "content": " to the same value in our experiments. Therefore, each SPARO slot has limited expressivity and representational capacity. The bottleneck underpinning each SPARO slot is a result of these constraints and the mechanistic bottleneck of a single attention mechanism instantiating competition between its input features." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "SPARO can be thought of as multi-head attention [60] where the queries are learned embeddings, key and value projection weights are shared, and the output projection weight is block-diagonal with each block containing the same parameters. Assuming " + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "L = V = D = \\sqrt{d}" + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": ", SPARO requires a total of " + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "d^2 + 2d" + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": " parameters, compared to multi-head attention's " + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "4d^2" + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 186, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 186, + 91, + 448, + 102 + ], + "type": "text", + "content": "SPARO: Selective Attention in Transformer Encodings for Vision" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 251, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 251, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 251, + 126 + ], + "type": "text", + "content": "3.3 CLIP with SPARO" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 133, + 482, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 133, + 482, + 206 + ], + "spans": [ + { + "bbox": [ + 130, + 133, + 482, + 206 + ], + "type": "text", + "content": "We add SPARO to both the image and text encoders of CLIP [50], setting the same values of " + }, + { + "bbox": [ + 130, + 133, + 482, + 206 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 133, + 482, + 206 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 133, + 482, + 206 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 133, + 482, + 206 + ], + "type": "text", + "content": " in both encoders. Each SPARO slot is " + }, + { + "bbox": [ + 130, + 133, + 482, + 206 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 130, + 133, + 482, + 206 + ], + "type": "text", + "content": "-normalized separately, and we divide their concatenation by " + }, + { + "bbox": [ + 130, + 133, + 482, + 206 + ], + "type": "inline_equation", + "content": "\\sqrt{L}" + }, + { + "bbox": [ + 130, + 133, + 482, + 206 + ], + "type": "text", + "content": " to produce the global " + }, + { + "bbox": [ + 130, + 133, + 482, + 206 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 130, + 133, + 482, + 206 + ], + "type": "text", + "content": "-normalized encoding. With this scheme, the cosine similarity between the image encoding " + }, + { + "bbox": [ + 130, + 133, + 482, + 206 + ], + "type": "inline_equation", + "content": "\\mathbf{y}^{\\mathrm{i}}" + }, + { + "bbox": [ + 130, + 133, + 482, + 206 + ], + "type": "text", + "content": " and the text encoding " + }, + { + "bbox": [ + 130, + 133, + 482, + 206 + ], + "type": "inline_equation", + "content": "\\mathbf{y}^{\\mathrm{t}}" + }, + { + "bbox": [ + 130, + 133, + 482, + 206 + ], + "type": "text", + "content": " becomes the expected cosine similarity between corresponding slot encodings:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 144, + 211, + 481, + 244 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 211, + 481, + 244 + ], + "spans": [ + { + "bbox": [ + 144, + 211, + 481, + 244 + ], + "type": "interline_equation", + "content": "\\boldsymbol {y} ^ {\\mathrm {i}} ^ {\\top} \\boldsymbol {y} ^ {\\mathrm {t}} = \\sum_ {l = 1} ^ {L} \\| \\boldsymbol {y} _ {l} ^ {\\mathrm {i}} \\| \\| \\boldsymbol {y} _ {l} ^ {\\mathrm {t}} \\| \\frac {\\boldsymbol {y} _ {l} ^ {\\mathrm {i}} {} ^ {\\top} \\boldsymbol {y} _ {l} ^ {\\mathrm {t}}}{\\| \\boldsymbol {y} _ {l} ^ {\\mathrm {i}} \\| \\| \\boldsymbol {y} _ {l} ^ {\\mathrm {t}} \\|} = \\sum_ {l = 1} ^ {L} \\frac {1}{L} \\operatorname {s i m} (\\boldsymbol {y} _ {l} ^ {\\mathrm {i}}, \\boldsymbol {y} _ {l} ^ {\\mathrm {t}}) = \\underset {l \\sim [ L ]} {\\mathbb {E}} \\left[ \\operatorname {s i m} (\\boldsymbol {y} _ {l} ^ {\\mathrm {i}}, \\boldsymbol {y} _ {l} ^ {\\mathrm {t}}) \\right]. \\quad (2)", + "image_path": "ccb5b8cd1049199387fec626f239e6576943e6269fed5394bea7507006a1a396.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 249, + 482, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 249, + 482, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 249, + 482, + 346 + ], + "type": "text", + "content": "We train the model using the standard CLIP loss which maximizes " + }, + { + "bbox": [ + 130, + 249, + 482, + 346 + ], + "type": "inline_equation", + "content": "\\mathrm{sim}(\\pmb{y}^{\\mathrm{i}},\\pmb{y}^{\\mathrm{t}})" + }, + { + "bbox": [ + 130, + 249, + 482, + 346 + ], + "type": "text", + "content": " of encodings " + }, + { + "bbox": [ + 130, + 249, + 482, + 346 + ], + "type": "inline_equation", + "content": "\\pmb{y}^{\\mathrm{i}}" + }, + { + "bbox": [ + 130, + 249, + 482, + 346 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 249, + 482, + 346 + ], + "type": "inline_equation", + "content": "\\pmb{y}^{\\mathrm{t}}" + }, + { + "bbox": [ + 130, + 249, + 482, + 346 + ], + "type": "text", + "content": " for aligned image-text pairs and minimizes it across all other pairings. With SPARO's separate-head attention bottleneck, maximizing " + }, + { + "bbox": [ + 130, + 249, + 482, + 346 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{l\\sim [L]}\\left[\\mathrm{sim}(\\pmb{y}_l^{\\mathrm{i}},\\pmb{y}_l^{\\mathrm{t}})\\right]" + }, + { + "bbox": [ + 130, + 249, + 482, + 346 + ], + "type": "text", + "content": " for aligned pairs across the training dataset encourages both modalities to learn similar attention semantics per slot " + }, + { + "bbox": [ + 130, + 249, + 482, + 346 + ], + "type": "inline_equation", + "content": "l\\in [L]" + }, + { + "bbox": [ + 130, + 249, + 482, + 346 + ], + "type": "text", + "content": ". Consequently, SPARO imparts a prior that both modalities encode a shared world with " + }, + { + "bbox": [ + 130, + 249, + 482, + 346 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 249, + 482, + 346 + ], + "type": "text", + "content": " selectively attendable concepts, and similarities between inputs can be expressed as their average similarity for these concepts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 361, + 254, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 361, + 254, + 372 + ], + "spans": [ + { + "bbox": [ + 132, + 361, + 254, + 372 + ], + "type": "text", + "content": "3.4 DINO with SPARO" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 377, + 482, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 377, + 482, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 377, + 482, + 486 + ], + "type": "text", + "content": "The online and momentum encoders of DINO [9] are each comprised of a ViT backbone which produces the encoding, and a DINO head that transforms it into a categorical distribution for distillation. Unlike CLIP, these encoders have an inherent bias for learning similar encoding functions due to one being an exponentially moving average of the other. However, in standard DINO, there are no constraints for selective attention in the encoding structure. We add SPARO to DINO to impart the prior for representing separately-attendable concepts in its encodings. Concretely, we replace the final transformer block of the ViT backbone with the SPARO, but do not modify the DINO head." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 502, + 198, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 198, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 198, + 514 + ], + "type": "text", + "content": "4 Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 524, + 481, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 524, + 481, + 560 + ], + "spans": [ + { + "bbox": [ + 130, + 524, + 481, + 560 + ], + "type": "text", + "content": "In Secs. 4 and 5, we evaluate our method on a variety of downstream tasks, analyze SPARO's representation structure with the ability to intervene on selected slots, perform ablations to support our design, and visualize the learned concepts." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 561, + 481, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 561, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 130, + 561, + 481, + 611 + ], + "type": "text", + "content": "In this section, we validate that partitioning the representation structure as a collection of separately-attended concepts leads to improved generalization for downstream recognition, robustness, compositionality, and retrieval, using SPARO with CLIP [50] in Secs. [4.1] to [4.3] and DINO [9] in Sec. [4.3]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 617, + 482, + 666 + ], + "type": "text", + "content": "Datasets. We train our CLIP models on one of Conceptual Captions 3M (CC3M) [57], Conceptual Captions 12M (CC12M) [11], a combination of CC3M and CC12M (CC15M), or LAION-400M (L400M) [54]. Appendix C provides the statistics for our training datasets." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "A. Vani et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 169, + 146, + 444, + 336 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 1: Zero-shot recognition, robustness, and SugarCrepe compositionality for CLIP models trained on Conceptual Captions and LAION-400M." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 169, + 146, + 444, + 336 + ], + "lines": [ + { + "bbox": [ + 169, + 146, + 444, + 336 + ], + "spans": [ + { + "bbox": [ + 169, + 146, + 444, + 336 + ], + "type": "table", + "html": "
TrainModelImageNet-Object NetSugar Crepe
V1V2SketchAR
CC3MCLIP16 (C)0.1410.1220.0680.0330.1770.0800.611
C+GAP0.1560.1340.0690.0330.1870.0810.616
C+SPARO0.1700.1400.0880.0350.2210.0980.625
CC12MCLIP16 (C)0.3610.3110.2490.0910.4670.2180.697
C+GAP0.3820.3300.2620.1010.5010.2410.695
C+SPARO0.4060.3500.2980.1130.5590.2680.723
CC15MCLIP16 (C)0.3840.3370.2680.1050.5030.2380.699
C+GAP0.3990.3430.2870.1140.5310.2520.701
C+SPARO0.4370.3780.3170.1450.5790.2790.730
L400MCLIP32 (C)0.6170.5310.4820.2020.7190.4230.748
C+GAP0.6230.5370.4920.2120.7250.4400.732
C+SPARO0.6350.5520.5070.2310.7470.4590.770
", + "image_path": "4f02047679b37ee6918f6e9a857e471f3ebbdc8d930d92c0a1aa34032f8c6178.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 359, + 482, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 359, + 482, + 564 + ], + "spans": [ + { + "bbox": [ + 130, + 359, + 482, + 564 + ], + "type": "text", + "content": "Models. For all experiments in this section, we train CLIP models using the open-source OpenCLIP [34] project. We consider two model sizes: those with a ViT-B/16 visual backbone " + }, + { + "bbox": [ + 130, + 359, + 482, + 564 + ], + "type": "inline_equation", + "content": "(\\mathrm{CLIP}^{16})" + }, + { + "bbox": [ + 130, + 359, + 482, + 564 + ], + "type": "text", + "content": " and those with a ViT-B/32 backbone " + }, + { + "bbox": [ + 130, + 359, + 482, + 564 + ], + "type": "inline_equation", + "content": "(\\mathrm{CLIP}^{32})" + }, + { + "bbox": [ + 130, + 359, + 482, + 564 + ], + "type": "text", + "content": ". For models trained with SPARO, we replace the last transformer block of the base transformer backbones. For our chosen settings, this ensures that the resulting model size is comparable to the original model size. Further, since only the CLS token output from the image transformer and the EOS token output from the text transformer are used for the standard CLIP encodings, removing one transformer block ensures that we do not gain an unfair advantage by attending to positions from a layer that standard CLIP discards. Note that standard CLIP also retains the MLP as a part of its final transformer block, which is absent in the SPARO module. We also compare our encodings with those produced by CLIP with global average pooling (CLIP+GAP), which enjoys the advantage of being able to use all of the final layer outputs. When using SPARO, we use " + }, + { + "bbox": [ + 130, + 359, + 482, + 564 + ], + "type": "inline_equation", + "content": "L = V = 64" + }, + { + "bbox": [ + 130, + 359, + 482, + 564 + ], + "type": "text", + "content": " when training on Conceptual Captions, and " + }, + { + "bbox": [ + 130, + 359, + 482, + 564 + ], + "type": "inline_equation", + "content": "L = 128" + }, + { + "bbox": [ + 130, + 359, + 482, + 564 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 359, + 482, + 564 + ], + "type": "inline_equation", + "content": "V = 64" + }, + { + "bbox": [ + 130, + 359, + 482, + 564 + ], + "type": "text", + "content": " when training on LAION-400M. We also provide results for CLIP with residual network (ResNet) [28] encoders in Appendix B.3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 583, + 440, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 583, + 440, + 596 + ], + "spans": [ + { + "bbox": [ + 132, + 583, + 440, + 596 + ], + "type": "text", + "content": "4.1 Zero-shot recognition, robustness, and compositionality" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "We evaluate the zero-shot classification accuracy of trained CLIP models on ImageNet [16] and a set of robustness benchmarks including ImageNet-V2 [53], ImageNet-Sketch [62], ImageNet-A [31], ImageNet-R [30], and ObjectNet [4]. Additionally, to evaluate the compositionality of our learned encodings in terms of objects, attributes, and relations, we test our trained models on SugarCrepe" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 186, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 186, + 91, + 448, + 102 + ], + "type": "text", + "content": "SPARO: Selective Attention in Transformer Encodings for Vision" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 141, + 116, + 297, + 205 + ], + "blocks": [ + { + "bbox": [ + 141, + 116, + 297, + 205 + ], + "lines": [ + { + "bbox": [ + 141, + 116, + 297, + 205 + ], + "spans": [ + { + "bbox": [ + 141, + 116, + 297, + 205 + ], + "type": "image", + "image_path": "a4e4ed91bc6eccbd8d1a29e4a177c6082607043557c7f549f34082801cd4e39f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 154, + 207, + 288, + 217 + ], + "lines": [ + { + "bbox": [ + 154, + 207, + 288, + 217 + ], + "spans": [ + { + "bbox": [ + 154, + 207, + 288, + 217 + ], + "type": "text", + "content": "(a) " + }, + { + "bbox": [ + 154, + 207, + 288, + 217 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP^{16}}" + }, + { + "bbox": [ + 154, + 207, + 288, + 217 + ], + "type": "text", + "content": " models trained on CC3M." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 314, + 117, + 471, + 206 + ], + "blocks": [ + { + "bbox": [ + 314, + 117, + 471, + 206 + ], + "lines": [ + { + "bbox": [ + 314, + 117, + 471, + 206 + ], + "spans": [ + { + "bbox": [ + 314, + 117, + 471, + 206 + ], + "type": "image", + "image_path": "b2ee92f9fafb6e58631375e2d655d3047e4e2676e91d796712b2e8a95510b0ea.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 325, + 207, + 463, + 217 + ], + "lines": [ + { + "bbox": [ + 325, + 207, + 463, + 217 + ], + "spans": [ + { + "bbox": [ + 325, + 207, + 463, + 217 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 325, + 207, + 463, + 217 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP^{16}}" + }, + { + "bbox": [ + 325, + 207, + 463, + 217 + ], + "type": "text", + "content": " models trained on CC12M." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 140, + 220, + 296, + 308 + ], + "blocks": [ + { + "bbox": [ + 140, + 220, + 296, + 308 + ], + "lines": [ + { + "bbox": [ + 140, + 220, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 140, + 220, + 296, + 308 + ], + "type": "image", + "image_path": "888dfbd4e6e928bde8154e4532868dd2a102f34b33c0dae68dfcbb083b2f6ff9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 150, + 310, + 288, + 319 + ], + "lines": [ + { + "bbox": [ + 150, + 310, + 288, + 319 + ], + "spans": [ + { + "bbox": [ + 150, + 310, + 288, + 319 + ], + "type": "text", + "content": "(c) " + }, + { + "bbox": [ + 150, + 310, + 288, + 319 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP^{16}}" + }, + { + "bbox": [ + 150, + 310, + 288, + 319 + ], + "type": "text", + "content": " models trained on CC15M." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 311, + 220, + 468, + 309 + ], + "blocks": [ + { + "bbox": [ + 311, + 220, + 468, + 309 + ], + "lines": [ + { + "bbox": [ + 311, + 220, + 468, + 309 + ], + "spans": [ + { + "bbox": [ + 311, + 220, + 468, + 309 + ], + "type": "image", + "image_path": "68e09b5ef1d38e54acfe64c539467f560e083602c12c930b69d77cb815fe427c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 322, + 310, + 459, + 320 + ], + "lines": [ + { + "bbox": [ + 322, + 310, + 459, + 320 + ], + "spans": [ + { + "bbox": [ + 322, + 310, + 459, + 320 + ], + "type": "text", + "content": "(d) " + }, + { + "bbox": [ + 322, + 310, + 459, + 320 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP^{32}}" + }, + { + "bbox": [ + 322, + 310, + 459, + 320 + ], + "type": "text", + "content": " models trained on L400M." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 329, + 482, + 352 + ], + "lines": [ + { + "bbox": [ + 131, + 329, + 482, + 352 + ], + "spans": [ + { + "bbox": [ + 131, + 329, + 482, + 352 + ], + "type": "text", + "content": "Fig. 2: Relative differences of " + }, + { + "bbox": [ + 131, + 329, + 482, + 352 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP + SPARO}" + }, + { + "bbox": [ + 131, + 329, + 482, + 352 + ], + "type": "text", + "content": " zero-shot accuracies when compared to " + }, + { + "bbox": [ + 131, + 329, + 482, + 352 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP + GAP}" + }, + { + "bbox": [ + 131, + 329, + 482, + 352 + ], + "type": "text", + "content": " on the VTAB benchmark." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 379, + 482, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 379, + 482, + 451 + ], + "spans": [ + { + "bbox": [ + 130, + 379, + 482, + 451 + ], + "type": "text", + "content": "33. We report our results in Tab. 1, and provide more fine-grained SugarCrepe numbers in Appendix B.4. We find that using SPARO encodings outperforms CLIP and " + }, + { + "bbox": [ + 130, + 379, + 482, + 451 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP + GAP}" + }, + { + "bbox": [ + 130, + 379, + 482, + 451 + ], + "type": "text", + "content": " in all settings considered. The improved results on the SugarCrepe benchmark suggests that SPARO encodings exhibit compositionality at a level that enable better binding of object, attribute, and relation properties than baseline encodings on average." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 453, + 483, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 453, + 483, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 483, + 525 + ], + "type": "text", + "content": "Finally, we also use the VTAB benchmark [70] to evaluate SPARO on a diverse set of datasets [5,13,14,19,22,23,29,35,36,38,45,46,48,61,63,68] that vary more significantly from the training data. We present the relative zero-shot classification accuracy improvements of CLIP + SPARO over CLIP + GAP in Fig. 2 and the corresponding absolute values in Appendix B.6. We see that SPARO outperforms standard CLIP encodings for a majority of VTAB tasks in all four settings." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 552, + 334, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 552, + 334, + 564 + ], + "spans": [ + { + "bbox": [ + 131, + 552, + 334, + 564 + ], + "type": "text", + "content": "4.2 Zero-shot image and text retrieval" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "type": "text", + "content": "To separately evaluate the quality of the learned image and text encodings, we consider zero-shot retrieval based on image and text on MS COCO [40], Flickr8k [32], and Flickr30k [65]. Our results are presented in Tab. 2. To compute Recall@5 for 'Text' we find the nearest 5 texts for an image, calculate the fraction of those texts that match the image, and average this metric over all the images (similar for Recall@5 for 'Image'). We find that SPARO encodings improve retrieval across all three datasets for both modalities." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "text", + "content": "A. Vani et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 179, + 146, + 434, + 335 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 479, + 137 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 479, + 137 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 479, + 137 + ], + "type": "text", + "content": "Table 2: Image and text zero-shot Recall@5 retrieval results on MS COCO, Flickr8k, and Flickr30k for CLIP models trained on Conceptual Captions and LAION-400M." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 179, + 146, + 434, + 335 + ], + "lines": [ + { + "bbox": [ + 179, + 146, + 434, + 335 + ], + "spans": [ + { + "bbox": [ + 179, + 146, + 434, + 335 + ], + "type": "table", + "html": "
TrainModelImage R@5Text R@5
COCOF8kF30kCOCOF8kF30k
CC3MCLIP16 (C)0.2370.4000.3530.2940.4890.469
C+GAP0.2600.4300.4000.3120.5340.497
C+SPARO0.2890.4900.4610.3630.5870.581
CC12MCLIP16 (C)0.4740.6910.7150.6140.8370.817
C+GAP0.4990.7140.7250.6280.8190.847
C+SPARO0.5260.7370.7590.6700.8780.878
CC15MCLIP16 (C)0.5120.7380.7570.6360.8440.865
C+GAP0.5350.7610.7720.6590.8490.881
C+SPARO0.5570.7780.7930.6960.8980.905
L400MCLIP32 (C)0.5990.8220.8400.7650.9230.935
C+GAP0.6100.8220.8450.7680.9170.946
C+SPARO0.6160.8360.8540.7740.9350.950
", + "image_path": "153a5b1fd04ad83517fdc9b006f242b45ba32e26b887c8ce8929545d3f20333f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 138, + 390, + 323, + 464 + ], + "blocks": [ + { + "bbox": [ + 131, + 347, + 481, + 381 + ], + "lines": [ + { + "bbox": [ + 131, + 347, + 481, + 381 + ], + "spans": [ + { + "bbox": [ + 131, + 347, + 481, + 381 + ], + "type": "text", + "content": "Table 3: Left: ImageNet linear probe accuracy for CLIP models trained on Conceptual Captions and LAION-400M. Right: ImageNet 20-nearest neighbors and linear probe accuracy for encodings trained with DINO." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 390, + 323, + 464 + ], + "lines": [ + { + "bbox": [ + 138, + 390, + 323, + 464 + ], + "spans": [ + { + "bbox": [ + 138, + 390, + 323, + 464 + ], + "type": "table", + "html": "
ModelImageNet linear probe
Train: CC3MCC12MCC15ML400M
CLIP (C)0.4690.6300.6460.743
C+GAP0.5040.6490.6640.747
C+SPARO0.5610.7000.7110.755
", + "image_path": "97ac6651a897e6ef7191874042cf4e7931c81498ba89b6c5105e28c892c71b65.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 327, + 396, + 473, + 459 + ], + "blocks": [ + { + "bbox": [ + 327, + 396, + 473, + 459 + ], + "lines": [ + { + "bbox": [ + 327, + 396, + 473, + 459 + ], + "spans": [ + { + "bbox": [ + 327, + 396, + 473, + 459 + ], + "type": "table", + "html": "
ModelImageNet eval
20-NN Linear probe
DINO0.6850.735
DINO+SPARO0.7060.757
", + "image_path": "84060768415400d8141c958fc9a1c31c570697bbfc4e89e0703e1b55dceded76.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 487, + 441, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 487, + 441, + 499 + ], + "spans": [ + { + "bbox": [ + 131, + 487, + 441, + 499 + ], + "type": "text", + "content": "4.3 Linear probe and DINO nearest neighbors classification" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "We evaluate the classification accuracy of linear probes trained on the encodings of our trained models. For CLIP models, we follow [50] where we take a subset of the training data to use as a validation set, and sweep for the best weight decay hyperparameter on it for each model setting. Additionally, we evaluate SPARO for self-supervised learning in vision using DINO [9]. We train DINO with a ViT-S/16 backbone on ImageNet [16] without labels with 8 GPUs for 100 epochs without changing any other hyperparameters. Here, we use SPARO with " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "L = V = 64" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": ", replacing the last transformer block of the ViT encoder. The trained DINO models are evaluated on " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "-nearest neighbours with " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "k = 20" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " and linear classification on ImageNet with labels. We report our results for CLIP and DINO in Tab. Our results for CLIP align with our previous results, with SPARO outperforming CLIP and CLIP+GAP in all linear probe settings. Furthermore, we see that a straightforward incorporation of SPARO in the DINO encoder" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 186, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 186, + 91, + 448, + 102 + ], + "type": "text", + "content": "SPARO: Selective Attention in Transformer Encodings for Vision" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 212, + 157, + 402, + 300 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 4: Effect of intervening to selecting only the top 32 slots based on zero-shot ImageNet validation accuracy on SugarCrepe compositionality for CC15M-trained " + }, + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}^{16} + \\mathrm{SPARO}" + }, + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "text", + "content": " and LAION-400M-trained " + }, + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}^{32} + \\mathrm{SPARO}" + }, + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 212, + 157, + 402, + 300 + ], + "lines": [ + { + "bbox": [ + 212, + 157, + 402, + 300 + ], + "spans": [ + { + "bbox": [ + 212, + 157, + 402, + 300 + ], + "type": "table", + "html": "
SugarCrepeCC15ML400M
Slots:100%50%100%
Object0.8960.9120.9270.929
ReplaceAttribute0.7980.7970.8220.854
Relation0.6660.7130.6780.705
SwapObject0.5930.6590.5850.646
Attribute0.7000.7030.7190.740
AddObject0.7790.8090.8660.857
Attribute0.6790.7370.7940.822
Average0.7300.7610.7700.793
", + "image_path": "8f1095f8c2b31f144cec3356f2ebe6a8126a1c033ebd93652c2a49eec72fef37.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 323, + 479, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 323, + 479, + 349 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 479, + 349 + ], + "type": "text", + "content": "outperforms standard DINO for both nearest neighbors and linear classification. We provide additional experimental details in Appendices C.1 and C.2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 368, + 205, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 368, + 205, + 382 + ], + "spans": [ + { + "bbox": [ + 132, + 368, + 205, + 382 + ], + "type": "text", + "content": "5 Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 395, + 482, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 395, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 395, + 482, + 479 + ], + "type": "text", + "content": "In this section, we provide insights into encodings learned by SPARO. We demonstrate the ability of using SPARO's slot structure to perform post-hoc selection of slots for improving downstream performance. We then show that without this structure, such selection is more prone to overfitting. Next, we validate the design choice of using single-head attention to model the notion of selective attention through controlled ablations. Finally, we visualize example concepts that CLIP models trained with SPARO learn to attend to." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 499, + 294, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 499, + 294, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 499, + 294, + 511 + ], + "type": "text", + "content": "5.1 Post-hoc concept selection" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "content": "SPARO enables downstream tasks access to a collection of concepts. However, not all concepts may be important for a specific downstream task. For example, attending to watermarks or caption style can be important when trying to determine the source of a sample, but is unnecessary and often harmful for most downstream tasks that focus on the elements in the scenes. The separation of concepts in the structure of a SPARO encoding makes it possible to manually intervene to select desirable slots." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "We demonstrate this ability with a simple heuristic: we perform concept selection by computing the zero-shot ImageNet validation accuracy of each slot separately, and pick only the top-performing 32 slots. In Tab.4, we show that such an intervention exposes improved compositionality when evaluating CLIP+SPARO models trained on CC15M and LAION-400M on SugarCrepe." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "A. Vani et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 178, + 482, + 338 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 5: Effect of overfitting a partitioned mask for image encodings " + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": " of frozen CC15M-trained " + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "{\\mathrm{{CLIP}}}^{32} + \\mathrm{{GAP}}" + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "{\\mathrm{{CLIP}}}^{32} + \\mathrm{{SPARO}}" + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": " models by training on SugarCrepe. We normalize the relative change on the evaluation benchmarks by the attained relative change on SugarCrepe. The highlighted rows indicate settings where the structure assumed for the mask aligns with SPARO's representation structure." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 178, + 482, + 338 + ], + "lines": [ + { + "bbox": [ + 132, + 178, + 482, + 338 + ], + "spans": [ + { + "bbox": [ + 132, + 178, + 482, + 338 + ], + "type": "table", + "html": "
(# parts, Model Shape of y)TrainingEvaluation
SugarCrepeVTAB AvgImageNet 0-shot Acc
InitialMaskedInitialMaskedRelative changeInitialMaskedRelative change
(64, 64)+GAP(4096)0.6920.7320.2400.217-1.6890.3430.292-2.598
+SPARO(64, 64)0.7100.7640.2620.272+0.4970.3700.364-0.190
(64, 8)+GAP(512)0.6780.7120.2540.219-2.7940.3450.279-3.924
+SPARO(64, 8)0.6960.7410.2490.260+0.6880.3550.347-0.347
(512, 1)+GAP(512)0.6780.7710.2540.215-1.1190.3450.250-2.003
+SPARO(64, 8)0.6960.8090.2490.206-1.0590.3550.224-2.277
(4096, 1)+GAP(4096)0.6920.8150.2400.215-0.5960.3430.286-0.938
+SPARO(64, 64)0.7100.8810.2620.242-0.3160.3700.328-0.465
", + "image_path": "17068aede2c1618842827ddb0319f4389c44acae9e0a0b21c1c000fc9b5351f5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 361, + 348, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 361, + 348, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 361, + 348, + 373 + ], + "type": "text", + "content": "5.2 Robustness of SPARO's slot structure" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 388, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 482, + 544 + ], + "type": "text", + "content": "We further utilize the notion of post-hoc concept selection to study the robustness of SPARO's representational structure, and compare it with that of standard encodings. We set up the experiment as an inverse of Sec. 5.1 — we perform concept selection using the smaller SugarCrepe benchmark and evaluate the resulting encodings on real-world benchmarks. SugarCrepe contains 7,512 example pairs of positive and negative captions, differing only in compositional interpretation in 7 ways, over a set of 1,561 images. Compositionality is a shared high-level notion present in most real-world data, yet there is much information that exists in data beyond what is necessary to solve the SugarCrepe tasks. The small size of the SugarCrepe as a training dataset, together with the ubiquitous nature of compositionality in machine learning tasks, makes it an ideal candidate to overfit concept selection to for stress-testing the representational structure of SPARO by evaluating on other tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 545, + 483, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 483, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 483, + 667 + ], + "type": "text", + "content": "We train a global mask " + }, + { + "bbox": [ + 130, + 545, + 483, + 667 + ], + "type": "inline_equation", + "content": "\\pmb{m} \\in [0,1]^M" + }, + { + "bbox": [ + 130, + 545, + 483, + 667 + ], + "type": "text", + "content": " of the image encodings " + }, + { + "bbox": [ + 130, + 545, + 483, + 667 + ], + "type": "inline_equation", + "content": "\\pmb{y} \\in \\mathbb{R}^M" + }, + { + "bbox": [ + 130, + 545, + 483, + 667 + ], + "type": "text", + "content": " to maximize SugarCrepe performance, and evaluate the masked encodings " + }, + { + "bbox": [ + 130, + 545, + 483, + 667 + ], + "type": "inline_equation", + "content": "\\pmb{m} \\odot \\pmb{y}" + }, + { + "bbox": [ + 130, + 545, + 483, + 667 + ], + "type": "text", + "content": " on VTAB tasks and ImageNet for zero-shot classification. We train masks for encodings of frozen " + }, + { + "bbox": [ + 130, + 545, + 483, + 667 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}^{32} + \\mathrm{GAP}" + }, + { + "bbox": [ + 130, + 545, + 483, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 545, + 483, + 667 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}^{32} + \\mathrm{SPARO}" + }, + { + "bbox": [ + 130, + 545, + 483, + 667 + ], + "type": "text", + "content": " models trained on CC15M with " + }, + { + "bbox": [ + 130, + 545, + 483, + 667 + ], + "type": "inline_equation", + "content": "M = 512" + }, + { + "bbox": [ + 130, + 545, + 483, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 545, + 483, + 667 + ], + "type": "inline_equation", + "content": "M = 4096" + }, + { + "bbox": [ + 130, + 545, + 483, + 667 + ], + "type": "text", + "content": " each. For the SPARO models, we use " + }, + { + "bbox": [ + 130, + 545, + 483, + 667 + ], + "type": "inline_equation", + "content": "L = 64" + }, + { + "bbox": [ + 130, + 545, + 483, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 545, + 483, + 667 + ], + "type": "inline_equation", + "content": "V = M / L" + }, + { + "bbox": [ + 130, + 545, + 483, + 667 + ], + "type": "text", + "content": ". For each of these settings, we consider two types of masking: per-dimension and per-slot. To compare the " + }, + { + "bbox": [ + 130, + 545, + 483, + 667 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}^{32} + \\mathrm{GAP}" + }, + { + "bbox": [ + 130, + 545, + 483, + 667 + ], + "type": "text", + "content": " encodings on per-slot masking, we consider their " + }, + { + "bbox": [ + 130, + 545, + 483, + 667 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 545, + 483, + 667 + ], + "type": "text", + "content": " contiguous equal-sized partitions as the slot encodings. We provide full implementation details of our mask training setup in Appendix C.1." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 186, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 186, + 91, + 448, + 102 + ], + "type": "text", + "content": "SPARO: Selective Attention in Transformer Encodings for Vision" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 178, + 480, + 380 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 173 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 173 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 173 + ], + "type": "text", + "content": "Table 6: Ablation experiments showing benefits of SPARO's attention bottleneck (Cross-attention is 'Separate-head') and cross-modal attention alignment (Slot-wise LayerNorm and projection are 'Yes' or 'N/A') with CC15M-trained CLIP, CLIP+GAP, CLIP+SPARO, and ablated variants of CLIP+SPARO that include multi-head cross-attention and the attentional pooler (AttPool) [39][66]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 178, + 480, + 380 + ], + "lines": [ + { + "bbox": [ + 133, + 178, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 133, + 178, + 480, + 380 + ], + "type": "table", + "html": "
ModelReplace last blockCross- attentionSlot-wise LayerNormSlot-wise projectionParamsFLOPS (×109)ImageNet zero-shot
CLIP32 (C)N/AN/AN/AN/A151M7.40.329
C+GAP151M7.40.345
C+AttPoolNoNo440M8.90.312
Multi-headYes306M8.80.344
YesNo440M8.90.338
Ablated C+SPARONoNoNo306M8.80.359
NoYes306M8.80.344
Separate-headNo172M8.70.358
YesNo306M8.80.344
Yes172M8.70.374
C+SPAROYesN/AN/A161M8.00.372
151M7.40.370
", + "image_path": "7e04754e3a4962c88a1f44c9746d6f10cf57fcfd55e846fccebbe823d3b7f3e5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 403, + 482, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 547 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 547 + ], + "type": "text", + "content": "We expect the compositionality of the data to be encoded in concepts that are useful for real-world downstream benchmarks. Therefore, the aggressive selection of only the concepts that perform well on a compositionality benchmark should not significantly impact downstream performance. However, since SugarCrepe does not represent all possible compositional manipulations, we expect to see a drop in performance on some tasks due to overfitting. We present our results in Tab. 5. We find that when the mask structure is aligned with the slot structure of SPARO, overfitting concept selection improves the average performance on VTAB tasks, and incurs only a minimal drop in performance on ImageNet. Furthermore, even in settings where the masks are not trained to align with SPARO's slot structure, we see that SPARO remains more resilient to overfitting than standard encodings in a majority of settings." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 563, + 471, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 563, + 471, + 575 + ], + "spans": [ + { + "bbox": [ + 131, + 563, + 471, + 575 + ], + "type": "text", + "content": "5.3 Ablating the bottleneck and cross-modal alignment of SPARO" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 581, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 668 + ], + "type": "text", + "content": "We evaluate the generalization benefits of the separate-head attention bottleneck and of aligning the SPARO attention heads between the modalities in CLIP. We set up SPARO on one end of the ablation spectrum and the cross-attention readout of attentional pooler (AttPool) [39,66] to the other. AttPool uses multi-head attention to attend to the backbone encoder's outputs using embedded learned queries whereas SPARO uses separate single-head attention mechanisms. Additionally, AttnPool performs layer normalization [2] followed by a linear projec" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "A. Vani et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "tion on the output of attention to produce the output encoding, whereas SPARO directly uses the output of separate-head attention as the encoding." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 308 + ], + "type": "text", + "content": "We train " + }, + { + "bbox": [ + 130, + 140, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}^{32}" + }, + { + "bbox": [ + 130, + 140, + 482, + 308 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 140, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}^{32} + \\mathrm{GAP}" + }, + { + "bbox": [ + 130, + 140, + 482, + 308 + ], + "type": "text", + "content": ", and variants of " + }, + { + "bbox": [ + 130, + 140, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}^{32} + \\mathrm{SPARO}" + }, + { + "bbox": [ + 130, + 140, + 482, + 308 + ], + "type": "text", + "content": " on CC15M with " + }, + { + "bbox": [ + 130, + 140, + 482, + 308 + ], + "type": "inline_equation", + "content": "L = 128" + }, + { + "bbox": [ + 130, + 140, + 482, + 308 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 140, + 482, + 308 + ], + "type": "inline_equation", + "content": "V = 64" + }, + { + "bbox": [ + 130, + 140, + 482, + 308 + ], + "type": "text", + "content": " and evaluate for ImageNet zero-shot accuracy and model size. We start by considering the choice of either separate-head or multi-head cross-attention over the backbone outputs. For each cross-attention setting, we apply layer normalization and linear projection operations to the attention module's outputs, similar to AttnPool. However, we additionally also consider slot-wise variants of these operations. Here, a slot-wise operation " + }, + { + "bbox": [ + 130, + 140, + 482, + 308 + ], + "type": "inline_equation", + "content": "f_{\\phi}" + }, + { + "bbox": [ + 130, + 140, + 482, + 308 + ], + "type": "text", + "content": " is one that acts separately on each input slot " + }, + { + "bbox": [ + 130, + 140, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_l" + }, + { + "bbox": [ + 130, + 140, + 482, + 308 + ], + "type": "text", + "content": " using the same parameters " + }, + { + "bbox": [ + 130, + 140, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 140, + 482, + 308 + ], + "type": "text", + "content": ", and produces the slot-structured output " + }, + { + "bbox": [ + 130, + 140, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\operatorname{concat}(f_{\\phi}(\\mathbf{y}_1), \\ldots, f_{\\phi}(\\mathbf{y}_L))" + }, + { + "bbox": [ + 130, + 140, + 482, + 308 + ], + "type": "text", + "content": ". When applied to an input that is not slot-structured, we assume a slot structure by partitioning the input into " + }, + { + "bbox": [ + 130, + 140, + 482, + 308 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 140, + 482, + 308 + ], + "type": "text", + "content": " contiguous equal-sized partitions. With SPARO, using both slot-wise layer normalization and slot-wise projection enables each read-out attention head from one modality to align with the corresponding head from the other without being influenced by the other heads." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 308, + 482, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 482, + 392 + ], + "type": "text", + "content": "We present our results in Tab. 6, showing a clear advantage of SPARO over AttnPool. Additionally, both the choices of separate-head attention over multi-head and slot-wise operations over the alternatives result in improved generalization, reduced model size, and fewer floating-point operations per second (FLOPS). While the former supports the effectiveness of the separate-head attention bottleneck, the latter validates the benefits of the prior for shared separately-attendable concepts between the modalities enabled by SPARO." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 414, + 231, + 425 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 414, + 231, + 425 + ], + "spans": [ + { + "bbox": [ + 131, + 414, + 231, + 425 + ], + "type": "text", + "content": "5.4 Visualizations" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 437, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 581 + ], + "type": "text", + "content": "To conclude our analysis, we qualitatively visualize the concepts represented by SPARO using the " + }, + { + "bbox": [ + 130, + 437, + 482, + 581 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}^{16} + \\mathrm{SPARO}" + }, + { + "bbox": [ + 130, + 437, + 482, + 581 + ], + "type": "text", + "content": " model trained on CC15M from Sec. 4.1. A limitation of visualizing attended positions is that it only provides partial insight into the represented concept — we can tell where the information is taken from, but not how the information is used. Different SPARO slots can have similar attention maps for the same sample, but reveal different semantics when comparing patterns across other images. For instance, a slot that attends to animals (e.g. dogs, cats, elephants) and another that attends to transportation (e.g. trains, boats, cars) can both have the same attention mask over patches and words for a horse being used for transportation. Furthermore, not all SPARO concepts align between the two modalities for every sample due to information being present in one that is not possible to infer from the other." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "text", + "content": "To ease our interpretation of attended concepts, we choose concepts to visualize that have sharp attention in the text modality, attend to non-overlapping text tokens, and have cosine similarity more than 0.75 with the corresponding image slots. We present examples of these filtered concepts in Fig. 3. In this example, we see that SPARO is capable of attending to different concepts such as subject of the scene, the activity represented, and the surrounding location, for both vision and text. We provide additional visualizations in Appendix B.7" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 186, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 186, + 91, + 448, + 102 + ], + "type": "text", + "content": "SPARO: Selective Attention in Transformer Encodings for Vision" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 145, + 121, + 214, + 171 + ], + "blocks": [ + { + "bbox": [ + 145, + 121, + 214, + 171 + ], + "lines": [ + { + "bbox": [ + 145, + 121, + 214, + 171 + ], + "spans": [ + { + "bbox": [ + 145, + 121, + 214, + 171 + ], + "type": "image", + "image_path": "6d5b27494b94184ef7d1aa9cae2e65816ac98d50b68948dfbb394ea0a967ead7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 137, + 171, + 222, + 183 + ], + "lines": [ + { + "bbox": [ + 137, + 171, + 222, + 183 + ], + "spans": [ + { + "bbox": [ + 137, + 171, + 222, + 183 + ], + "type": "text", + "content": "A herd of cattle walking down a road being followed by a cowboy" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 235, + 121, + 302, + 171 + ], + "blocks": [ + { + "bbox": [ + 235, + 121, + 302, + 171 + ], + "lines": [ + { + "bbox": [ + 235, + 121, + 302, + 171 + ], + "spans": [ + { + "bbox": [ + 235, + 121, + 302, + 171 + ], + "type": "image", + "image_path": "8cb54f7ee842e079c7bdff05d550db44d625d160abc073d1e0411e8a8db43392.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 231, + 171, + 310, + 182 + ], + "lines": [ + { + "bbox": [ + 231, + 171, + 310, + 182 + ], + "spans": [ + { + "bbox": [ + 231, + 171, + 310, + 182 + ], + "type": "text", + "content": "everal surfboards standing in a rowon the beach" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 319, + 121, + 395, + 171 + ], + "blocks": [ + { + "bbox": [ + 319, + 121, + 395, + 171 + ], + "lines": [ + { + "bbox": [ + 319, + 121, + 395, + 171 + ], + "spans": [ + { + "bbox": [ + 319, + 121, + 395, + 171 + ], + "type": "image", + "image_path": "21d9354066660f9d4080f3b5ad4079f49cb8e16edbdf804d4714ddf10a136e76.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 171, + 396, + 182 + ], + "lines": [ + { + "bbox": [ + 317, + 171, + 396, + 182 + ], + "spans": [ + { + "bbox": [ + 317, + 171, + 396, + 182 + ], + "type": "text", + "content": "Two people riding horses on a rock path" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 415, + 121, + 466, + 171 + ], + "blocks": [ + { + "bbox": [ + 415, + 121, + 466, + 171 + ], + "lines": [ + { + "bbox": [ + 415, + 121, + 466, + 171 + ], + "spans": [ + { + "bbox": [ + 415, + 121, + 466, + 171 + ], + "type": "image", + "image_path": "9906ef39be979a259828797000bd8b106e07656c9f3ec60edda9fa1854d03000.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 171, + 475, + 182 + ], + "lines": [ + { + "bbox": [ + 406, + 171, + 475, + 182 + ], + "spans": [ + { + "bbox": [ + 406, + 171, + 475, + 182 + ], + "type": "text", + "content": "A man sitting alone on a park bench in a park" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 145, + 182, + 214, + 232 + ], + "blocks": [ + { + "bbox": [ + 145, + 182, + 214, + 232 + ], + "lines": [ + { + "bbox": [ + 145, + 182, + 214, + 232 + ], + "spans": [ + { + "bbox": [ + 145, + 182, + 214, + 232 + ], + "type": "image", + "image_path": "a27332458870ffcd532900de9e89d83dd88adb3042e7bc50ad652b5ade9bfcb4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 137, + 232, + 222, + 244 + ], + "lines": [ + { + "bbox": [ + 137, + 232, + 222, + 244 + ], + "spans": [ + { + "bbox": [ + 137, + 232, + 222, + 244 + ], + "type": "text", + "content": "A herd of cattle walking down a road being followed by a cowboy" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 235, + 182, + 302, + 232 + ], + "blocks": [ + { + "bbox": [ + 235, + 182, + 302, + 232 + ], + "lines": [ + { + "bbox": [ + 235, + 182, + 302, + 232 + ], + "spans": [ + { + "bbox": [ + 235, + 182, + 302, + 232 + ], + "type": "image", + "image_path": "89e821ecd4948ab4e7c058dd0c00e29ec8947994e091e4110b486c16a34cc90e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 227, + 232, + 310, + 244 + ], + "lines": [ + { + "bbox": [ + 227, + 232, + 310, + 244 + ], + "spans": [ + { + "bbox": [ + 227, + 232, + 310, + 244 + ], + "type": "text", + "content": "Several surfboards standing in a row on the beach" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 319, + 182, + 394, + 232 + ], + "blocks": [ + { + "bbox": [ + 319, + 182, + 394, + 232 + ], + "lines": [ + { + "bbox": [ + 319, + 182, + 394, + 232 + ], + "spans": [ + { + "bbox": [ + 319, + 182, + 394, + 232 + ], + "type": "image", + "image_path": "f8ed22736767397e70be98419c26e2ca1b18f33ba4a0d39e4776c6859bbbfce0.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 232, + 396, + 243 + ], + "lines": [ + { + "bbox": [ + 318, + 232, + 396, + 243 + ], + "spans": [ + { + "bbox": [ + 318, + 232, + 396, + 243 + ], + "type": "text", + "content": "Two people riding horses on a rock path" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 416, + 182, + 466, + 232 + ], + "blocks": [ + { + "bbox": [ + 416, + 182, + 466, + 232 + ], + "lines": [ + { + "bbox": [ + 416, + 182, + 466, + 232 + ], + "spans": [ + { + "bbox": [ + 416, + 182, + 466, + 232 + ], + "type": "image", + "image_path": "e0545f5a8b26e7a38db18fe56e1d02f499267c9ba8948fe8a1b6ebc63c93affe.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 232, + 475, + 243 + ], + "lines": [ + { + "bbox": [ + 406, + 232, + 475, + 243 + ], + "spans": [ + { + "bbox": [ + 406, + 232, + 475, + 243 + ], + "type": "text", + "content": "A man sitting alone on a park bench in a park" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 145, + 244, + 214, + 293 + ], + "blocks": [ + { + "bbox": [ + 145, + 244, + 214, + 293 + ], + "lines": [ + { + "bbox": [ + 145, + 244, + 214, + 293 + ], + "spans": [ + { + "bbox": [ + 145, + 244, + 214, + 293 + ], + "type": "image", + "image_path": "76ffc4830bb9abc250a8810893259e3ff5bdee2625a556360fac8f07105caf6b.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 137, + 293, + 221, + 306 + ], + "lines": [ + { + "bbox": [ + 137, + 293, + 221, + 306 + ], + "spans": [ + { + "bbox": [ + 137, + 293, + 221, + 306 + ], + "type": "text", + "content": "A herd of cattle walking down a road being followed by a cowboy" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 314, + 482, + 348 + ], + "lines": [ + { + "bbox": [ + 131, + 314, + 482, + 348 + ], + "spans": [ + { + "bbox": [ + 131, + 314, + 482, + 348 + ], + "type": "text", + "content": "Fig. 3: Visualizing of the attended image and text positions for three SPARO slots (one per row) across four examples (one per column) from MS COCO. We surmise that the SPARO concepts from top to bottom represent the subject, activity, and location." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 235, + 244, + 302, + 293 + ], + "blocks": [ + { + "bbox": [ + 235, + 244, + 302, + 293 + ], + "lines": [ + { + "bbox": [ + 235, + 244, + 302, + 293 + ], + "spans": [ + { + "bbox": [ + 235, + 244, + 302, + 293 + ], + "type": "image", + "image_path": "6a6f91537768480583e46fbc5fa75e1437235fbbfcd3d69027d0069e56974497.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 228, + 293, + 310, + 305 + ], + "lines": [ + { + "bbox": [ + 228, + 293, + 310, + 305 + ], + "spans": [ + { + "bbox": [ + 228, + 293, + 310, + 305 + ], + "type": "text", + "content": "Several surfboards standing in a row on the beach" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 319, + 244, + 394, + 293 + ], + "blocks": [ + { + "bbox": [ + 319, + 244, + 394, + 293 + ], + "lines": [ + { + "bbox": [ + 319, + 244, + 394, + 293 + ], + "spans": [ + { + "bbox": [ + 319, + 244, + 394, + 293 + ], + "type": "image", + "image_path": "f63e2bc8b62bc21dc527adf5d5b2e2907e48aeaac7ced2e0fd7301541bc48c03.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 293, + 396, + 305 + ], + "lines": [ + { + "bbox": [ + 318, + 293, + 396, + 305 + ], + "spans": [ + { + "bbox": [ + 318, + 293, + 396, + 305 + ], + "type": "text", + "content": "Two people riding horses on a rock path" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 416, + 244, + 466, + 293 + ], + "blocks": [ + { + "bbox": [ + 416, + 244, + 466, + 293 + ], + "lines": [ + { + "bbox": [ + 416, + 244, + 466, + 293 + ], + "spans": [ + { + "bbox": [ + 416, + 244, + 466, + 293 + ], + "type": "image", + "image_path": "b51dd07a94b93dc37c46b0ad11824a65df7d64e959c852ac62df0a0872e0e203.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 407, + 293, + 475, + 305 + ], + "lines": [ + { + "bbox": [ + 407, + 293, + 475, + 305 + ], + "spans": [ + { + "bbox": [ + 407, + 293, + 475, + 305 + ], + "type": "text", + "content": "A man sitting alone on a park bench in a park" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "bbox": [ + 132, + 370, + 217, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 370, + 217, + 382 + ], + "spans": [ + { + "bbox": [ + 132, + 370, + 217, + 382 + ], + "type": "text", + "content": "6 Discussion" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 130, + 394, + 482, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 394, + 482, + 598 + ], + "spans": [ + { + "bbox": [ + 130, + 394, + 482, + 598 + ], + "type": "text", + "content": "Humans utilize a prior for compositionality to coherently represent different subsets of salient aspects of the world. We can enrich our comprehension by selectively attending to new aspects or specialize to a task by filtering out distracting ones. We introduce SPARO with a goal of imparting a similar prior to transformers in representation learning frameworks by partitioning encodings into separately-attended concepts. Although we see evidence of disentanglement through positive results of post-hoc concept selection and visualization of attention maps, we do not impose any explicit independence or disentanglement constraints. In our representation learning frameworks, the pressure for learning distinct concepts arises from the need to explain the factors of variation through the training objective with only the features that can be accumulated by a single head of attention with embedded queries and limited dimensionality. However, auxiliary training objectives to impose stronger distributional conditions on the learned concepts can be explored as promising future directions. Beyond pretraining, we highlight that more sophisticated post-hoc concept selection approaches than explored in our work, like using human interaction or set cover algorithms, can further improve the downstream utility of SPARO." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 132, + 616, + 241, + 630 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 616, + 241, + 630 + ], + "spans": [ + { + "bbox": [ + 132, + 616, + 241, + 630 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "This research was funded by Sony and enabled in part by compute resources provided by the Digital Research Alliance of Canada, Mila, and Sony." + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "A. Vani et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 137, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 138, + 137, + 480, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 137, + 480, + 158 + ], + "spans": [ + { + "bbox": [ + 138, + 137, + 480, + 158 + ], + "type": "text", + "content": "1. Aydemir, G., Xie, W., Güney, F.: Self-supervised object-centric learning for videos. arXiv preprint arXiv:2310.06907 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 159, + 480, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 159, + 480, + 179 + ], + "spans": [ + { + "bbox": [ + 138, + 159, + 480, + 179 + ], + "type": "text", + "content": "2. Ba, J.L., Kiros, J.R., Hinton, G.E.: Layer normalization. arXiv preprint arXiv:1607.06450 (2016)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 180, + 480, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 180, + 480, + 201 + ], + "spans": [ + { + "bbox": [ + 138, + 180, + 480, + 201 + ], + "type": "text", + "content": "3. Bahdanau, D., Cho, K., Bengio, Y.: Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473 (2014)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 201, + 480, + 244 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 201, + 480, + 244 + ], + "spans": [ + { + "bbox": [ + 138, + 201, + 480, + 244 + ], + "type": "text", + "content": "4. Barbu, A., Mayo, D., Alverio, J., Luo, W., Wang, C., Gutfreund, D., Tenenbaum, J., Katz, B.: Objectnet: A large-scale bias-controlled dataset for pushing the limits of object recognition models. Advances in neural information processing systems 32 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 245, + 480, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 245, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 138, + 245, + 480, + 277 + ], + "type": "text", + "content": "5. Beattie, C., Leibo, J.Z., Teplyashin, D., Ward, T., Wainwright, M., Kuttler, H., Lefrancq, A., Green, S., Valdés, V., Sadik, A., et al.: Deepmind lab. arXiv preprint arXiv:1612.03801 (2016)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 277, + 480, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 277, + 480, + 298 + ], + "spans": [ + { + "bbox": [ + 138, + 277, + 480, + 298 + ], + "type": "text", + "content": "6. Boff, K.R., Kaufman, L., Thomas, J.P.: Handbook of perception and human performance, vol. 1. Wiley New York (1986)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 299, + 480, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 299, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 138, + 299, + 480, + 331 + ], + "type": "text", + "content": "7. Brady, J., Zimmermann, R.S., Sharma, Y., Schölkopf, B., von Kügelgen, J., Brendel, W.: Provably learning object-centric representations. arXiv preprint arXiv:2305.14229 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 331, + 480, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 331, + 480, + 364 + ], + "spans": [ + { + "bbox": [ + 138, + 331, + 480, + 364 + ], + "type": "text", + "content": "8. Burgess, C.P., Matthew, L., Watters, N., Kabra, R., Higgins, I., Botvinick, M., Lerchner, A.: Monet: Unsupervised scene decomposition and representation. arXiv preprint arXiv:1901.11390 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 364, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 364, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 138, + 364, + 480, + 396 + ], + "type": "text", + "content": "9. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 396, + 480, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 396, + 480, + 417 + ], + "spans": [ + { + "bbox": [ + 138, + 396, + 480, + 417 + ], + "type": "text", + "content": "0. Chang, H.S., Sun, R.Y., Ricci, K., McCallum, A.: Multi-cls bert: An efficient alternative to traditional ensembling. arXiv preprint arXiv:2210.05043 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 418, + 480, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 418, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 138, + 418, + 480, + 460 + ], + "type": "text", + "content": "1. Changpinyo, S., Sharma, P., Ding, N., Soricut, R.: Conceptual 12M: Pushing web-scale image-text pre-training to recognize long-tail visual concepts. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3558-3568 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 460, + 480, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 460, + 480, + 482 + ], + "spans": [ + { + "bbox": [ + 138, + 460, + 480, + 482 + ], + "type": "text", + "content": "2. Chen, X., Xie, S., He, K.: An empirical study of training self-supervised vision transformers. arXiv preprint arXiv:2104.02057 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 483, + 480, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 483, + 480, + 503 + ], + "spans": [ + { + "bbox": [ + 138, + 483, + 480, + 503 + ], + "type": "text", + "content": "3. Cheng, G., Han, J., Lu, X.: Remote sensing image scene classification: Benchmark and state of the art. Proceedings of the IEEE 105(10), 1865-1883 (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 503, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 503, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 138, + 503, + 480, + 536 + ], + "type": "text", + "content": "4. Cimpoi, M., Maji, S., Kokkinos, I., Mohamed, S., Vedaldi, A.: Describing textures in the wild. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3606-3613 (2014)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 536, + 480, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 536, + 480, + 556 + ], + "spans": [ + { + "bbox": [ + 138, + 536, + 480, + 556 + ], + "type": "text", + "content": "5. Colby, C.L., Goldberg, M.E.: Space and attention in parietal cortex. Annual review of neuroscience 22(1), 319-349 (1999)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 557, + 480, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 557, + 480, + 589 + ], + "spans": [ + { + "bbox": [ + 138, + 557, + 480, + 589 + ], + "type": "text", + "content": "6. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 590, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 590, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 138, + 590, + 480, + 621 + ], + "type": "text", + "content": "7. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: BERT: Pre-training of deep bidirectional transformers for language understanding. In: North American Chapter of the Association for Computational Linguistics (2019)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 622, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 622, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 622, + 480, + 665 + ], + "type": "text", + "content": "8. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 186, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 186, + 91, + 447, + 102 + ], + "type": "text", + "content": "SPARO: Selective Attention in Transformer Encodings for Vision" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "19. Dugas, E., Jared, Jorge, Cukierski, W.: Diabetic retinopathy detection (2015), https://kaggle.com/competitions/diabetic-retinopathy-detection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 172 + ], + "type": "text", + "content": "20. Engelcke, M., Kosiorek, A.R., Jones, O.P., Posner, I.: Genesis: Generative scene inference and sampling with object-centric latent representations. arXiv preprint arXiv:1907.13052 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 174, + 480, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 174, + 480, + 206 + ], + "spans": [ + { + "bbox": [ + 132, + 174, + 480, + 206 + ], + "type": "text", + "content": "21. Eslami, S., Heess, N., Weber, T., Tassa, Y., Szepesvari, D., Hinton, G.E., et al.: Attend, infer, repeat: Fast scene understanding with generative models. Advances in neural information processing systems 29 (2016)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 207, + 480, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 207, + 480, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 207, + 480, + 228 + ], + "type": "text", + "content": "22. Fei-Fei, L., Fergus, R., Perona, P.: One-shot learning of object categories. IEEE transactions on pattern analysis and machine intelligence 28(4), 594-611 (2006)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 230, + 480, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 230, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 230, + 480, + 251 + ], + "type": "text", + "content": "23. Geiger, A., Lenz, P., Stiller, C., Urtasun, R.: Vision meets robotics: The kitti dataset. The International Journal of Robotics Research 32(11), 1231-1237 (2013)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 253, + 480, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 253, + 480, + 285 + ], + "spans": [ + { + "bbox": [ + 132, + 253, + 480, + 285 + ], + "type": "text", + "content": "24. Goyal, A., Didolkar, A., Ke, N.R., Blundell, C., Beaudoin, P., Heess, N., Mozer, M.C., Bengio, Y.: Neural production systems. Advances in Neural Information Processing Systems 34, 25673-25687 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 286, + 480, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 286, + 480, + 329 + ], + "spans": [ + { + "bbox": [ + 132, + 286, + 480, + 329 + ], + "type": "text", + "content": "25. Goyal, A., Lamb, A., Gampa, P., Beaudoin, P., Blundell, C., Levine, S., Bengio, Y., Mozer, M.C.: Factorizing declarative and procedural knowledge in structured, dynamical environments. In: International Conference on Learning Representations (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 331, + 480, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 331, + 480, + 352 + ], + "spans": [ + { + "bbox": [ + 132, + 331, + 480, + 352 + ], + "type": "text", + "content": "26. Goyal, A., Lamb, A., Hoffmann, J., Sodhani, S., Levine, S., Bengio, Y., Scholkopf, B.: Recurrent independent mechanisms. arXiv preprint arXiv:1909.10893 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 354, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 480, + 396 + ], + "type": "text", + "content": "27. Greff, K., Kaufman, R.L., Kabra, R., Watters, N., Burgess, C., Zoran, D., Matthew, L., Botvinick, M., Lerchner, A.: Multi-object representation learning with iterative variational inference. In: International Conference on Machine Learning. pp. 2424-2433. PMLR (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 398, + 480, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 398, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 398, + 480, + 430 + ], + "type": "text", + "content": "28. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 432, + 480, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 432, + 480, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 432, + 480, + 475 + ], + "type": "text", + "content": "29. Helber, P., Bischke, B., Dengel, A., Borth, D.: Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 12(7), 2217-2226 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 476, + 480, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 476, + 480, + 520 + ], + "spans": [ + { + "bbox": [ + 132, + 476, + 480, + 520 + ], + "type": "text", + "content": "30. Hendrycks, D., Basart, S., Mu, N., Kadavath, S., Wang, F., Dorundo, E., Desai, R., Zhu, T., Parajuli, S., Guo, M., et al.: The many faces of robustness: A critical analysis of out-of-distribution generalization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8340-8349 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 521, + 480, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 521, + 480, + 553 + ], + "spans": [ + { + "bbox": [ + 132, + 521, + 480, + 553 + ], + "type": "text", + "content": "31. Hendrycks, D., Zhao, K., Basart, S., Steinhardt, J., Song, D.: Natural adversarial examples. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15262-15271 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 555, + 480, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 480, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 480, + 586 + ], + "type": "text", + "content": "32. Hodosh, M., Young, P., Hockenmaier, J.: Framing image description as a ranking task: Data, models and evaluation metrics. Journal of Artificial Intelligence Research 47, 853-899 (2013)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 588, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 588, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 480, + 620 + ], + "type": "text", + "content": "33. Hsieh, C.Y., Zhang, J., Ma, Z., Kembhavi, A., Krishna, R.: Sugarcrepe: Fixing hackable benchmarks for vision-language compositionality. Advances in neural information processing systems (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 622, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 480, + 666 + ], + "type": "text", + "content": "34. Ilharco, G., Wortsman, M., Wightman, R., Gordon, C., Carlini, N., Taori, R., Dave, A., Shankar, V., Namkoong, H., Miller, J., Hajishirzi, H., Farhadi, A., Schmidt, L.: OpenCLIP. https://github.com/mlfoundations/open Clip (Jul 2021). https://doi.org/10.5281/zenodo.5143773" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "A. Vani et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 162 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 162 + ], + "type": "text", + "content": "35. Johnson, J., Hariharan, B., van der Maaten, L., Fei-Fei, L., Lawrence Zitnick, C., Girshick, R.: Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2901-2910 (2017)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 163, + 481, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 163, + 481, + 184 + ], + "spans": [ + { + "bbox": [ + 130, + 163, + 481, + 184 + ], + "type": "text", + "content": "36. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images. Technical report (2009)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 185, + 481, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 185, + 481, + 229 + ], + "spans": [ + { + "bbox": [ + 130, + 185, + 481, + 229 + ], + "type": "text", + "content": "37. Lavoie, S., Tsirigotis, C., Schwarzer, M., Vani, A., Noukhovitch, M., Kawaguchi, K., Courville, A.: Simplicial embeddings in self-supervised learning and downstream classification. In: International Conference on Learning Representations (2023), https://openreview.net/forum?id=RWtGreRpovS" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 230, + 481, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 230, + 481, + 274 + ], + "spans": [ + { + "bbox": [ + 130, + 230, + 481, + 274 + ], + "type": "text", + "content": "38. LeCun, Y., Huang, F.J., Bottou, L.: Learning methods for generic object recognition with invariance to pose and lighting. In: Proceedings of the 2004 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2004. CVPR 2004. vol. 2, pp. II-104. IEEE (2004)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 274, + 481, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 274, + 481, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 481, + 308 + ], + "type": "text", + "content": "39. Lee, J., Lee, Y., Kim, J., Kosiorek, A., Choi, S., Teh, Y.W.: Set transformer: A framework for attention-based permutation-invariant neural networks. In: International conference on machine learning. pp. 3744-3753. PMLR (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 308, + 481, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 481, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 481, + 384 + ], + "type": "text", + "content": "40. Lin, T., Maire, M., Belongie, S.J., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, C.L.: Microsoft COCO: common objects in context. In: Fleet, D.J., Pajdla, T., Schiele, B., Tuytelaars, T. (eds.) Computer Vision - ECCV 2014 - 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V. Lecture Notes in Computer Science, vol. 8693, pp. 740-755. Springer (2014). https://doi.org/10.1007/978-3-319-10602-1_48, https://doi.org/10.1007/978-3-319-10602-1%5F48" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 385, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 385, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 385, + 481, + 418 + ], + "type": "text", + "content": "41. Locatello, F., Weissenborn, D., Unterthiner, T., Mahendran, A., Heigold, G., Uszkoreit, J., Dosovitskiy, A., Kipf, T.: Object-centric learning with slot attention. Advances in Neural Information Processing Systems 33, 11525-11538 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 419, + 481, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 419, + 481, + 463 + ], + "spans": [ + { + "bbox": [ + 130, + 419, + 481, + 463 + ], + "type": "text", + "content": "42. Ma, Z., Hong, J., Gul, M.O., Gandhi, M., Gao, I., Krishna, R.: Crepe: Can vision-language foundation models reason compositionally? In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10910-10921 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 464, + 481, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 464, + 481, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 464, + 481, + 506 + ], + "type": "text", + "content": "43. Mansouri, A., Hartford, J., Zhang, Y., Bengio, Y.: Object centric architectures enable efficient causal representation learning. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=r9FsiXZxZt" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 508, + 481, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 508, + 481, + 541 + ], + "spans": [ + { + "bbox": [ + 130, + 508, + 481, + 541 + ], + "type": "text", + "content": "44. Martinez, A.: Putting spatial attention on the map: timing and localization of stimulus selection processes in striate and extrastriate visual areas. Vision Res 41, 1437-1457 (2001)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 542, + 481, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 542, + 481, + 564 + ], + "spans": [ + { + "bbox": [ + 130, + 542, + 481, + 564 + ], + "type": "text", + "content": "45. Matthew, L., Higgins, I., Hassabis, D., Lerchner, A.: dsprites: Disentanglement testing sprites dataset. https://github.com/deepmind/dsprites-dataset/ (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 565, + 481, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 565, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 130, + 565, + 481, + 597 + ], + "type": "text", + "content": "46. Nilsback, M.E., Zisserman, A.: Automated flower classification over a large number of classes. In: 2008 Sixth Indian conference on computer vision, graphics & image processing. pp. 722-729. IEEE (2008)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 599, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 599, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 599, + 481, + 632 + ], + "type": "text", + "content": "47. O'Connor, D.H., Fukui, M.M., Pinsk, M.A., Kastner, S.: Attention modulates responses in the human lateral geniculate nucleus. Nature neuroscience 5(11), 1203-1209 (2002)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 632, + 481, + 666 + ], + "type": "text", + "content": "48. Parkhi, O.M., Vedaldi, A., Zisserman, A., Jawahar, C.: Cats and dogs. In: 2012 IEEE conference on computer vision and pattern recognition. pp. 3498-3505. IEEE (2012)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 186, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 186, + 91, + 447, + 102 + ], + "type": "text", + "content": "SPARO: Selective Attention in Transformer Encodings for Vision" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "49. Qian, R., Ding, S., Liu, X., Lin, D.: Semantics meets temporal correspondence: Self-supervised object-centric learning in videos. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16675-16687 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "type": "text", + "content": "50. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "text", + "content": "51. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners. OpenAI blog 1(8), 9 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 258 + ], + "type": "text", + "content": "52. Ray, A., Radenovic, F., Dubey, A., Plummer, B.A., Krishna, R., Saenko, K.: Cola: A benchmark for compositional text-to-image retrieval. In: Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "text", + "content": "53. Recht, B., Roelofs, R., Schmidt, L., Shankar, V.: Do imagenet classifiers generalize to imagenet? In: International conference on machine learning. pp. 5389-5400. PMLR (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 481, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 326 + ], + "type": "text", + "content": "54. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 326, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 346 + ], + "type": "text", + "content": "55. Scott, W.A.: Cognitive complexity and cognitive flexibility. Sociometry pp. 405-414 (1962)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "text", + "content": "56. Seitzer, M., Horn, M., Zadaianchuk, A., Zietlow, D., Xiao, T., Simon-Gabriel, C.J., He, T., Zhang, Z., Scholkopf, B., Brox, T., et al.: Bridging the gap to real-world object-centric learning. arXiv preprint arXiv:2209.14860 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 380, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 435 + ], + "type": "text", + "content": "57. Sharma, P., Ding, N., Goodman, S., Soricut, R.: Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In: Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics, ACL 2018, Melbourne, Australia, July 15-20, 2018, Volume 1: Long Papers. pp. 2556-2565. Association for Computational Linguistics (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 435, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 479 + ], + "type": "text", + "content": "58. Thrush, T., Jiang, R., Bartolo, M., Singh, A., Williams, A., Kiela, D., Ross, C.: Winoground: Probing vision and language models for visio-linguistic compositionality. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5238-5248 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 479, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 501 + ], + "type": "text", + "content": "59. Treisman, A.M., Gelade, G.: A feature-integration theory of attention. Cognitive psychology 12(1), 97-136 (1980)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 501, + 481, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 481, + 534 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 481, + 534 + ], + "type": "text", + "content": "60. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 534, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 578 + ], + "type": "text", + "content": "61. Veeling, B.S., Linmans, J., Winkens, J., Cohen, T., Welling, M.: Rotation equivariant cnns for digital pathology. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2018: 21st International Conference, Granada, Spain, September 16-20, 2018, Proceedings, Part II 11. pp. 210-218. Springer (2018)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 578, + 481, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 481, + 611 + ], + "type": "text", + "content": "62. Wang, H., Ge, S., Lipton, Z., Xing, E.P.: Learning robust global representations by penalizing local predictive power. In: Advances in Neural Information Processing Systems. pp. 10506-10518 (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "text", + "content": "63. Xiao, J., Hays, J., Ehinger, K.A., Oliva, A., Torralba, A.: Sun database: Large-scale scene recognition from abbey to zoo. In: 2010 IEEE computer society conference on computer vision and pattern recognition. pp. 3485-3492. IEEE (2010)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "text", + "content": "64. Xu, L., Ouyang, W., Bennamoun, M., Boussaid, F., Xu, D.: Multi-class token transformer for weakly supervised semantic segmentation. In: Proceedings of the" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "A. Vani et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 402 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 149, + 116, + 481, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 116, + 481, + 138 + ], + "spans": [ + { + "bbox": [ + 149, + 116, + 481, + 138 + ], + "type": "text", + "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4310-4319 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 481, + 171 + ], + "type": "text", + "content": "65. Young, P., Lai, A., Hodosh, M., Hockenmaier, J.: From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics 2, 67-78 (2014)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 204 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 204 + ], + "type": "text", + "content": "66. Yu, J., Wang, Z., Vasudevan, V., Yeung, L., Seyedhosseini, M., Wu, Y.: Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "text", + "content": "67. Yuksekgonul, M., Bianchi, F., Kalluri, P., Jurafsky, D., Zou, J.: When and why vision-language models behave like bags-of-words, and what to do about it? In: The Eleventh International Conference on Learning Representations (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "type": "text", + "content": "68. Yuval, N.: Reading digits in natural images with unsupervised feature learning. In: Proceedings of the NIPS Workshop on Deep Learning and Unsupervised Feature Learning (2011)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 270, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 481, + 303 + ], + "type": "text", + "content": "69. Zadaianchuk, A., Kleindessner, M., Zhu, Y., Locatello, F., Brox, T.: Unsupervised semantic segmentation with self-supervised object-centric representations. arXiv preprint arXiv:2207.05027 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 303, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 303, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 303, + 481, + 346 + ], + "type": "text", + "content": "70. Zhai, X., Puigcerver, J., Kolesnikov, A., Ruyssen, P., Riquelme, C., Lucic, M., Djolonga, J., Pinto, A.S., Neumann, M., Dosovitskiy, A., et al.: A large-scale study of representation learning with the visual task adaptation benchmark. arXiv preprint arXiv:1910.04867 (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 347, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 369 + ], + "type": "text", + "content": "71. Zhang, Y., Hare, J., Prugel-Bennett, A.: Deep set prediction networks. Advances in Neural Information Processing Systems 32 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "type": "text", + "content": "72. Zhao, T., Zhang, T., Zhu, M., Shen, H., Lee, K., Lu, X., Yin, J.: Vl-checklist: Evaluating pre-trained vision-language models with objects, attributes and relations. arXiv preprint arXiv:2207.00221 (2022)" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 186, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 186, + 91, + 447, + 102 + ], + "type": "text", + "content": "SPARO: Selective Attention in Transformer Encodings for Vision" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/SPHINX_ A Mixer of Weights, Visual Embeddings and Image Scales for Multi-modal Large Language Models/cd141534-c355-432b-969a-838688ba26e3_content_list.json b/2024/SPHINX_ A Mixer of Weights, Visual Embeddings and Image Scales for Multi-modal Large Language Models/cd141534-c355-432b-969a-838688ba26e3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a20048d8ee346a1c668e717c39717e6b3aa9cdcf --- /dev/null +++ b/2024/SPHINX_ A Mixer of Weights, Visual Embeddings and Image Scales for Multi-modal Large Language Models/cd141534-c355-432b-969a-838688ba26e3_content_list.json @@ -0,0 +1,2213 @@ +[ + { + "type": "text", + "text": "SPHINX: A Mixer of Weights, Visual Embeddings and Image Scales for Multi-modal Large Language Models", + "text_level": 1, + "bbox": [ + 233, + 140, + 769, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ziyi Lin $^{12*}$ , Dongyang Liu $^{2*}$ , Renrui Zhang $^{12*}$ , Peng Gao $^{2*\\dagger\\dagger}$ , Longtian Qiu $^{23*}$ , Han Xiao $^{2}$ , Han Qiu $^{2}$ , Wenqi Shao $^{2}$ , Keqin Chen $^{2}$ , Jiaming Han $^{12}$ , Siyuan Huang $^{2}$ , Yichi Zhang $^{2}$ , Xuming He $^{3}$ , Yu Qiao $^{2\\dagger}$ , and Hongsheng Li $^{124\\dagger}$", + "bbox": [ + 215, + 233, + 785, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Multimedia Laboratory, The Chinese University of Hong Kong", + "bbox": [ + 285, + 291, + 720, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Shanghai AI Laboratory 3 ShanghaiTech University", + "bbox": [ + 313, + 306, + 689, + 320 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 Centre for Perceptual and Interactive Intelligence Limited", + "bbox": [ + 302, + 320, + 699, + 334 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "* Equal contribution † Equal advisory ‡ Project leader", + "bbox": [ + 287, + 334, + 712, + 348 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zylin@link.cuhk.edu.hk, gaopeng@pjlab.org.cn, hsli@ee.cuhk.edu.hk", + "bbox": [ + 250, + 349, + 751, + 362 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. We present SPHINX, a versatile multi-modal large language model (MLLM) with a joint mixing of model weights, visual embeddings and image scales. First, for stronger vision-language alignment, we unfreeze the large language model (LLM) during pre-training, and introduce a weight mix strategy between LLMs trained by real-world and synthetic data. By directly integrating the weights from two domains, the mixed LLM can efficiently incorporate diverse semantics with favorable robustness. Then, we propose to extract comprehensive visual embeddings from various network architectures, pre-training paradigms, and information granularity, providing language models with more robust image representations. We further propose an efficient strategy aiming to better capture fine-grained appearances of high-resolution images. With a mixing of different scales and high-resolution sub-images, SPHINX attains exceptional visual parsing and reasoning performance on existing evaluation benchmarks. Based on our proposed joint mixing, SPHINX exhibits superior multi-modal understanding capabilities on a wide range of applications, with highlighted fine-grained visual recognition abilities such as region-level understanding, caption grounding, document layout detection, and human pose estimation. We hope our work may cast a light on the exploration of joint mixing in future MLLM research. Code is released at https://github.com/Alpha-VLLM/LLaMA2-Accessory.", + "bbox": [ + 259, + 396, + 743, + 690 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 733, + 375, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Since the era of big data, large language models (LLMs) have attained tremendous strides (7; 61; 62; 87; 108), showcasing unprecedented application scenarios and generalization capabilities. Despite the phenomenal success, plain LLMs are still only reasoning models, and the lack of perception capability has limited their further evolution to autonomous agents with general intelligence. The mission", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Emergent Referring Pose Estimation", + "bbox": [ + 228, + 136, + 346, + 146 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/dbffd25f30333c077f55064be515e1510236d599e46b160aa63ac9f463cc721b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 228, + 146, + 243, + 154 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Else: \n - Select the keypoints of the person wearing a green suit.", + "bbox": [ + 243, + 147, + 369, + 164 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/bff5837be91f9bf40bc591fea0853f6d5418f5359ce5201e9813799030586359.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 228, + 165, + 243, + 180 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SPINDX: The detected key points are left ear[0.76,0.22], left shoulder[0.78,0.32], right shoulder[0.88,0.30], left elbow[0.74,0.49], right elbow[0.90,0.47], wrist[0.68,0.52], right wrist[0.81,0.46], left hip[0.80,0.65], right hip[0.90,0.64], knee[0.71,0.88], and right knee[0.84,0.88].", + "bbox": [ + 245, + 169, + 361, + 204 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/08c1336de65bd73b330d66b057276387e2bc846d3ae38023285ebd48b2597e58.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 138, + 483, + 200 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-level Dense Captioning", + "bbox": [ + 496, + 138, + 589, + 147 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "User: Detect all objectives in this image.", + "bbox": [ + 493, + 148, + 591, + 155 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SHINE: There are 10 objects shown in the image, including a person [0.21, 0.39, 37.0, 76], a helmet[0.30,0.39,0.36,0.43], two sneakers [0.67,0.75,0.85,0.95,0.76], bicycle[0.03,0.53,0.43,0.76], a cell phone[0.22,0.40,0.33,0.42], and four polo shirt[0.50,0.59,0.59,0.59,0.48,0.44,0.48,0.43,0.43,0.16,0.59].", + "bbox": [ + 493, + 156, + 619, + 196 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2a86f2ad90d5b69658e3d97b7061e2153230323078e6175baeb82fdd2e8e4d5d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 624, + 138, + 772, + 214 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Emergent Anomaly Detection", + "bbox": [ + 230, + 217, + 349, + 224 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "User: Is there defect on the image? If yes detect it and plot the box on the image", + "bbox": [ + 230, + 227, + 326, + 247 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SPHINX: Yes, there is a defect on the image. The defect is a scratch on the metal object. The box on the image is $[0.450.46, 0.500.57]$ , which is the scratch on the metal object.", + "bbox": [ + 328, + 227, + 478, + 247 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e18427647bc9055a328fd86d13f8dd6acbd5e67a5a509d6c82b8bebaf215ccab.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 230, + 251, + 313, + 301 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c2114a7b6ea442c886589f514ccb79c8d087bb187efbb3a1470bb3e487d1078c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 316, + 251, + 397, + 301 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b4df21ab2b24e733942d412936c1525ebe66843dfe4b9df8a389eca7072e2b88.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 400, + 251, + 480, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image Editing (w/ Stable Diffusion)", + "bbox": [ + 230, + 310, + 348, + 320 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SPEHIN: Detect road sign in this picture", + "bbox": [ + 236, + 324, + 310, + 335 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/762006744f2e2d5a1a7b63a214c46295da302e9b16ad402b961b9b354d2c45de.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 236, + 335, + 312, + 373 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SPHINX: Detect dog in this", + "bbox": [ + 316, + 324, + 392, + 332 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7be2eb41567a3321e3ccd14803ffa537885a1fbd4373ca7a0b43150e6229b5e1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 318, + 332, + 395, + 373 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SPHINX: Detect monocite in this picture.", + "bbox": [ + 401, + 324, + 475, + 332 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/826bcd4a3347ca2b110f5f2e35362f10a3e0c83d4ea85f57723a2a6a48cd7211.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 401, + 332, + 477, + 373 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Stable Diffusion: A light in the middle of the picture", + "bbox": [ + 236, + 377, + 303, + 386 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/47720c29c7c4bc1c5a495a8f03e6a91cf657d56db06abfa9215db01d0e3003a2.jpg", + "image_caption": [ + "Fig. 1: Examples of multi-task visual reasoning on selected tasks by our proposed SPHINX models, which excels in diverse visual understanding and perception tasks, such as object detection, caption grounding, and region-level description, displaying some emergent capabilities." + ], + "image_footnote": [], + "bbox": [ + 236, + 386, + 312, + 425 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Stable Diffusion: A teddy bear sitting in a basket", + "bbox": [ + 316, + 377, + 395, + 385 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9bdf9cef7c90c7c2cb7054de5c4c616ab725176b36ce1eaaea76cc8b97957610.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 318, + 385, + 395, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Stable Diffusion: A man riding a bicycle", + "bbox": [ + 401, + 377, + 478, + 385 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f25ee9bc17315cf3fb9aa75284d825782cde17bbbb09645737e6686c0e51f79c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 401, + 385, + 477, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "User: Please provide a short description for this document.", + "bbox": [ + 496, + 220, + 563, + 236 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SPHINX: A blue bicycle", + "bbox": [ + 496, + 237, + 558, + 244 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "User: Provide a one-sentence caption for the provided image", + "bbox": [ + 496, + 247, + 581, + 258 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SPHINK: A man riding a bike in a grassy field.", + "bbox": [ + 496, + 261, + 581, + 273 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "User:Generates a detailed description about the image.", + "bbox": [ + 496, + 276, + 581, + 289 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SHANK: The image shows a man standing next to a blue mountain bike in a grassy area. He is standing on the right leg of the bike, with his right leg on the pedal. The man appears to be in the process of mounting the bike, as he is standing on the pedal and holding his foot. There are several yellow flowers scattered throughout the grassy area, adding a touch of color to the scene. This is wearing shorts, which are visible as the stands next to the bike.", + "bbox": [ + 496, + 290, + 583, + 362 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "User: Please provide a short description for this region:", + "bbox": [ + 593, + 220, + 686, + 234 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SPHINK: A man in a blue shirt", + "bbox": [ + 593, + 236, + 671, + 244 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "User: Provide a one-Sentence caption for the provided image", + "bbox": [ + 593, + 247, + 678, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SHINX: A man riding a bike while holding a cell phone.", + "bbox": [ + 593, + 261, + 679, + 273 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "User: Generate a detailed description about the image.", + "bbox": [ + 593, + 275, + 679, + 284 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SHINK: The image shows a man sitting on a bicycle in a grassy area, and the grassy area is filled with people to be taking a selfie or recording a video while sitting on the bike. The grassy area is filled with people who he is prepared for a safe ride. There are several other bicycles in the scene, with one located to the left of the bike, and another located to the right, and a third one further back. The grassy area is scattered with people standing around, riding a couch on color to the scene.", + "bbox": [ + 593, + 287, + 679, + 354 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "User: Please provide a short description of the region [0.37,4.40,97.80]", + "bbox": [ + 684, + 220, + 772, + 236 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SHINK: A cow licking a man's face", + "bbox": [ + 684, + 237, + 769, + 250 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "User: Provide a one-sentence caphor for the provided image", + "bbox": [ + 684, + 252, + 771, + 263 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SPEEDON: A cow is sniffing a bicycle wheel.", + "bbox": [ + 684, + 265, + 758, + 276 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Uc:Generale a deblie du about the image.", + "bbox": [ + 684, + 277, + 751, + 290 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SPINDL: The image features a large brown and white cow standing in a grassy field. The cow is standing on the ground, looking at the edge of the field, and it appears to be sniffing a bicycle tire. Another cow is standing slightly to the right of the main cow. The scene captures the cow's perspective as it interacts with the bicycle tire.", + "bbox": [ + 684, + 292, + 769, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "to integrate reasoning and perception has motivated the research on multi-modal large language models (MLLMs) (41; 50; 106; 109; 112), which make responses to not only human language, but also natural signals, among which visual signals are predominantly studied.", + "bbox": [ + 212, + 520, + 784, + 579 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Early works on MLLMs (22; 50; 102) focus on coarse-grained tasks, such as image captioning and question answering. While this enables visual instruction following or visual dialogue to some extent, fine-grained visual recognition capabilities, such as visual grounding, key-point detection and referring generation, might bring new outlook to the development of MLLMs: They make possible new applications such as precise actions in embodied AI, as well as potentially improved diversity and reduced hallucination on traditional tasks with better access to detailed visual clues.", + "bbox": [ + 212, + 580, + 785, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Besides fine-grained recognition, it is equally important for a foundational MLLM to maintain the performance on a wider domain. Despite the success of previous works such as Shikra (10) which tackles fine-grained tasks alone, it is still challenging to explore the scaling-up methods to utilize larger datasets, higher parameter counts, higher resolutions, etc. in order to equip the models with broader or even emergent abilities.", + "bbox": [ + 212, + 703, + 785, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We thus present the building of SPHINX, our proposed foundational MLLM with versatility and an emphasis on fine-grained visual recognition capabilities. Compared to previous works with similar goals, such as Qwen-VL (4)", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Z. Lin et al.", + "bbox": [ + 271, + 114, + 352, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and MiniGPT-v2 (9), our pipeline can be highlighted with the following unique scaling-up techniques conveniently named as mixing strategies:", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- Mixed model weights. Vision-language data from particular domains might contain special semantics, e.g., synthetic captions (75) compared to real-world ones (76). Considering that directly mixing such data might confuse the MLLM, we introduce a weight-mixing strategy to efficiently combine such domain-specific knowledge. Based on the MLLM pre-trained on real-world data, we fine-tune it on the synthetic data, and then linearly combine the finetuned LLM's weights with the real-world ones. In this way, the two types of models would not be affected by contradictory data and our final SPHINX can effectively integrate knowledge from both synthetic and real-world domains, enabling effective use of datasets from diverse sources.", + "bbox": [ + 225, + 191, + 785, + 340 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- Mixed visual embeddings. To take the advantage of different encoders, we propose to mix the visual embeddings from various vision backbones (41; 64; 70) with different network architectures (CNN vs. ViT), pre-training paradigms (supervised vs. self-supervised), and information granularity (global vs. local). By mixing the different image tokens channel-wisely and sequence-wisely, SPHINX obtains stronger visual representations and leads to better vision-language alignment efficacy.", + "bbox": [ + 225, + 342, + 795, + 446 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- Mixed scales with high-resolution sub-images. To process high-resolution images as required for fine-grained visual recognition, we first spatially divide the input high-resolution image into multiple sub-images, and also downsample it into a low-resolution one. Then, we feed all the images concurrently into the mixed visual encoders, and concatenate the extracted multiple token groups to represent the entire high-resolution visual features. By mixing visual embeddings of different scales and sub-images, our SPHINX can adaptively explore more fine-grained visual semantics from the high resolution and multi-scale image representations, while maintaining encoding efficiency.", + "bbox": [ + 225, + 449, + 800, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "By the proposed three-fold mixing strategy, SPHINX fine-tunes LLMs, e.g., LLaMA-2 (88), to be a powerful MLLM with superior visual instruction-following capacity. Our model excels in a variety of vision tasks, e.g., detecting different objects with remarkable precision and parsing their relations, or accurately interpreting the content within complicated figures. Importantly, SPHINX can achieve impressive fine-grained visual perception for high-resolution images, which exhibits state-of-the-art performance on extensive evaluation benchmarks, e.g., MMBench (52), MME (20), and POPE (45).", + "bbox": [ + 212, + 613, + 792, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related works", + "text_level": 1, + "bbox": [ + 214, + 760, + 390, + 776 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Large language models (LLMs). The field of Natural Language Processing (NLP) has witnessed significant progress over the years, particularly with the advent of LLMs. With Transformer (89) as the fundamental architecture,", + "bbox": [ + 212, + 794, + 782, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "A Mixer of Weights, Visual Embeddings and Image Scales for MLLM", + "bbox": [ + 269, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "LLMs (61; 62; 72) have demonstrated unprecedented performance in modeling intricate language patterns over extensive contexts. Therein, BERT (15) showcases the benefits of pre-training on vast text corpora and fine-tuning on specific tasks, setting new standards on various benchmarks. OpenAI's GPT series (61; 62; 71; 72), especially GPT-3 (7), harness the power of massive model scaling, with billions and even trillions of parameters. To obtain better instruction following ability, InstructGPT (65) and ChatGPT (61) are presented to exhibit exceptional fluency and versatility in open-domain conversation tasks, ranging from text generation to question answering. Recently, the instruction tuning based on LLaMA (87) and LLaMA-2 (88) has gained great popularity as open-source LLMs in the community. Therein, Alpaca (86) and LLaMA-Adapter (106) respectively adopt full and parameter-efficient fine-tuning to acquire favorable instruction-following LLMs. Vicuna (12) and GPT-4-LLM (67) further showcase the improvement brought by higher-quality instruction datasets. Other efforts also extend LLMs for match problem solving (90; 111), visual model system (96; 101), and open-world recognition (107; 113). In this paper, we develop our SPHINX based on the superior language understanding of LLaMA-2 (88) and instruction tuning experience of LLaMA-Adapter series (22; 106), which introduce a three-fold mixer to extend the capability ceiling of instruction-following LLMs for multi-modal input.", + "bbox": [ + 212, + 145, + 787, + 448 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multi-modal large language models (MLLMs). In addition to language instruction following, many efforts have been made to inject multi-modal conditions into LLMs for wider application scenarios. As prior attempts, VisualGPT (8) and BLIP series (14; 41; 42) indicate the potential of aligning LLMs with visual input for image captioning and question answering. Flamingo (3) and Kosmos-1 (32) further exhibit promising multi-modal understanding performance for image-text interleaved contexts. With large-scale pre-training and model sizes, GPT-4 (62) and Bard (24) both showcase remarkable proficiency in vision-language understanding and reasoning over diverse multi-modal tasks. In parallel, a bunch of works have been proposed to align LLaMA with vision modality for advanced visual instruction-following capabilities. LLaVA (50) and MiniGPT-4 (112) utilize a simple projection layer to connect vision encoders (41; 70) with LLMs. LLaMA-Adapter V2 (22) introduces zero-initialized attention mechanisms for efficient visual instruction tuning, and mPLUG-Owl (102) adopts delicately designed intermediate networks for cross-modal alignment. For more modality input, ImageBind-LLM (28) and PandaGPT (83) further incorporate audio and video conditions guided by ImageBind (23). Besides, recent MLLMs are also extended to region-level parsing (10; 68), in-context learning (37; 38), arbitrary image resolutions (6), text-to-image generation (16; 93), and 3D question answering (25; 30; 98). Different from previous works, our SPHINX aims for image-conditioned MLLM, and proposes a three-fold mixer, i.e., model weights, tuning tasks, and visual embeddings, attaining superior generalization capacity for multi-modal learning.", + "bbox": [ + 212, + 455, + 787, + 803 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Weight interpolation. Despite being commonly known to be highly nonlinear (40), neural networks have, under some conditions, exhibited surprisingly", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Z. Lin et al.", + "bbox": [ + 271, + 114, + 352, + 126 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e50bfc0f53a05f22b325d62f9dba5b5fda3ebf8df5e46606a68345a80fa8ef69.jpg", + "image_caption": [ + "Fig. 2: The joint mixing paradigm of SPHINX. with mixed visual embeddings (a) and mixed model weights (b)." + ], + "image_footnote": [], + "bbox": [ + 248, + 148, + 743, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "good results with linearly interpolated weights in several cases, including higher-performing single-task models (11; 95), task interpolation (31; 97; 100), federated or low-communication distributed training (19; 43; 59). A few recent works (2; 80; 84) also investigated weight mixing under the MLLM context. On top of these works, we further demonstrate the effects of inter-domain weight mixing for image-c captioning pre-trained models, as well as how these benefits persist after further task-specific fine-tuning.", + "bbox": [ + 212, + 380, + 784, + 487 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "High-resolution visual embeddings. Fine-grained visual recognition tasks, such as detection and segmentation, traditionally utilize higher-resolution image features to capture detailed information. Convolutional neural networks (CNNs) naturally scale to high-resolution inputs (53) as the convolution operation intrinsically supports varied input sizes, while for the recent transformer-based vision backbones, scaling input resolution is commonly achieved by interpolating the positional embeddings (18). There are also network components specifically designed to obtain high-resolution feature maps, e.g., (44; 46; 103; 110). Unlike those model-surgery-based methods, our mixed-scale solution instead explores the use of fine-tuned large language models to fuse the visual embeddings from multiple parts of the image while leaving the visual encoders themselves untouched. With such a zooming-in mechanism, our framework captures fine-grained visual clues while preserves capabilities of the visual encoders in an architecture-agnostic manner. Interestingly, the recently released GPT-4V API (63) shows a similar correspondence between the image resolution and the number of visual tokens.", + "bbox": [ + 212, + 487, + 785, + 728 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 SPHINX", + "text_level": 1, + "bbox": [ + 215, + 750, + 339, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we introduce a versatile MLLM, SPHINX, with an model overview in Section 3.1, the joint mixing of model weights and visual embeddings in Section 3.2 and the mixing of high-resolution sub-image tokens Section 3.3. Finally, in Section 3.4, we introduce several extended applications of SPHINX.", + "bbox": [ + 212, + 779, + 784, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "A Mixer of Weights, Visual Embeddings and Image Scales for MLLM", + "bbox": [ + 269, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Model overview", + "text_level": 1, + "bbox": [ + 215, + 146, + 395, + 161 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The SPHINX model is composed of an LLM, e.g., LLaMA-2 (88), a mixture of vision encoders and linear projections connecting the vision encoder outputs to the language model inputs. For sequences with image input, visual tokens are extracted from the vision encoders, projected to the LM input space, and then concatenated to the beginning of the sequence. Sequences without image inputs are processed identically as in the language-only LLM.", + "bbox": [ + 212, + 172, + 787, + 263 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We adopt a two-stage training paradigm: the first pre-training stage for vision-language alignment, and the second fine-tuning stage for visual instruction-following learning. In both stages, the LM parameters are fully-tuned, together with the linear layers for visual-language feature alignment, but the visual encoders themselves are kept frozen.", + "bbox": [ + 212, + 263, + 789, + 339 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Pre-training: The joint mixing of visual embeddings and model weights", + "text_level": 1, + "bbox": [ + 214, + 363, + 782, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The pre-training stage (or stage 1) involves two types of mixing: weight mixing and visual embedding mixing, as shown in Figure 2. Since the pre-training datasets are weakly annotated image-caption pairs where the textual information does not contain much detail, we do not use high-resolution images for this stage, thus bypassing the scale mixing.", + "bbox": [ + 212, + 405, + 787, + 479 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Mixed model weights of different domains. Vision-language data from different domains may exhibit distinct characteristics: For example, web-crawled captions like those in LAION-400M (76) tend to be diverse but noisy, but synthetic captions like those in LAION-COCO (75) tend to be coherent but of a relatively limited vocabulary. We propose a weight mixing strategy of domain-specifically tuned LLMs to integrate respective knowledge from real-world and synthetic data. We first utilize the most common domain data (LAION-400M (76)) for pre-training, which endows the MLLM with fundamental visual understanding capabilities. Then, we regard such a pre-trained model as the initial checkpoint to further fine-tune the LLM on synthetic domains, e.g., LAION-COCO (75). Finally, to take advantage of the best data domains, we directly conduct a weighted mixing of two LLMs' weights for semantic aggregation. In detail, we denote the parameters of the fundamental LLM as $\\theta_{real}$ , and the fine-tuned parameters by synthetic data as $\\theta_{syn}$ . The mixing process is formulated as", + "bbox": [ + 212, + 481, + 795, + 693 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {m i x} = \\beta \\cdot \\theta_ {r e a l} + (1 - \\beta) \\cdot \\theta_ {s y n}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 705, + 785, + 722 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\beta$ denotes the mixing coefficient, and $\\theta_{\\text{mix}}$ represents the mixed LLM weights with aggregated semantics. Compared to fusing different domain data for joint pre-training, our weight mix strategy can encourage every MLLM to better learn domain-unique knowledge, and exhibit flexible scalability for any new data domains.", + "bbox": [ + 212, + 733, + 787, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Mixed embeddings for visual encoding. To capture robust visual representations from different aspects, we propose to ensemble a variety of vision", + "bbox": [ + 212, + 809, + 787, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Z. Lin et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/00591b1905a8833dc4bbfc8b522dd169744614c00a7788e3377bd119572b6e9a.jpg", + "image_caption": [ + "Fig. 3: Pipeline of SPHINX for high-resolution images. We propose to further mix different scales and sub-images to better capture fine-grained semantics on high-resolution images." + ], + "image_footnote": [], + "bbox": [ + 279, + 142, + 733, + 359 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "backbones for image encoding. The visual backbones with different characteristics are chosen as follows. 1) Different network architectures. As CNN (29) and ViT (18) mainly aggregate different types of visual appearances, i.e., neighboring dependencies and long-range interactions, we adopt CLIP (70) models respectively with ConvNeXt (94) and ViT image encoders. 2) Different pre-training paradigms. Supervised training can impose explicit semantic information from textual captions or category labels, while self-supervised learning enforces the model to explore implicit pretext task signals. Thus, we further employ the ViT self-supervised by DINOv2 (64) as well as the text-supervised vision encoders, CLIP. 3) Different information granularity. The aforementioned visual encoders all produce visual tokens in the patch level. To better capture global features, we also adopt Q-Former (41) to summarize visual embeddings via querying from the global context. After all the aforementioned encoding, we first channel-wisely concatenate the patch level visual tokens. The channel-wise concatenation allows for better alignment among spatially corresponding features, as well as keeps the number of tokens constant to the number of visual encoders. Then, by using two projection layers for dimension alignment, we spatial-wisely concatenate the representations between those of Q-Former and the other patch-level features. The obtained image tokens are directly placed in front of language instructions, which provide visual context for the language instructions.", + "bbox": [ + 212, + 445, + 787, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Fine-tuning: Mixing scales with high-resolution sub-images", + "text_level": 1, + "bbox": [ + 212, + 777, + 748, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the fine-tuning stage, or stage 2 training, the model from stage 1 will adapt for instruct following as well as high-resolution image recognition. Stage 2 introduces", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "A Mixer of Weights, Visual Embeddings and Image Scales for MLLM", + "bbox": [ + 269, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "scale mixing to process high-resolution input images, which are crucial for fine-grained visual tasks and will be elaborated in this section.", + "bbox": [ + 215, + 146, + 784, + 176 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Mixed scales and high-resolution sub-images. State-of-the-art open-source MLLMs (9; 10; 22; 41; 50; 68) works adopt frozen image encoders during all training stages, in order to preserve the pre-trained visual semantics. Therefore, the image resolution of MLLMs is usually set as $224 \\times 224$ , severely hindering their efficacy for fine-grained visual perception, especially region-level grounding and description. In our SPHINX, we extend the mixing of visual embeddings to more scales and high-resolution sub-images, allowing for efficient high-resolution image encoding. For an input high-resolution image, e.g., $448 \\times 448$ , we construct five corresponding images of $224 \\times 224$ , and feed them as independent images into our mixed vision encoders. Specifically, we first downsample the input image to $224 \\times 224$ as an abstract representation, and also downsample the input image to $448 \\times 448$ and crop four sub-images of $224 \\times 224$ from the four corners of the $448 \\times 448$ image, which preserve the detailed visual information. In this way, we enable MLLMs to not only capture fine-grained visual appearances with $224 \\times 224$ positional encodings, but also achieve favorable computation efficiency. Afterwards, the five groups of image tokens are encoded and concatenated as a long sequence for feeding into LLM, where the first one group encodes global semantics, and the other four record fine-grained local features. Importantly, as the image tokens of different patches do not have interaction through the vision encoders, they are forced to interact within the LLM to obtain complete visual information. Such a strategy, in turn, motivates LLMs to parse the relations within visual conditions for better cross-modal learning. From this perspective, our SPHINX can be regarded as a new paradigm for similar to ViT (18), where the mixed vision encoders serve as a patch embedding layer, and the LLM plays the role for patch interaction as a vision decoder. On visual understanding tasks requiring higher resolutions, SPHINX achieves significant improvement with the mixed visual representations of scales and high-resolution sub-images.", + "bbox": [ + 215, + 181, + 787, + 589 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "On comparison with scaling vision encoder input. The previous standard method to enlarge MLLM input resolutions is to enlarge the input of vision encoders themselves (e.g., by positional embedding interpolation for ViT) and make the vision encoder trainable (4; 9). Compared to vision encoder fine-tuning, we argue that our LLM-based patch fusion is the scalable choice when the vision encoder is large or consists of multiple models as is in our case, because (i) vision encoders can be slow to tune, as the sequence of the same length back-propagating through the vision encoders can be equally costly as backpropagating through the LM, and (ii) vision encoders can be unstable to tune, as different visual backbones may require different tuning configurations (17) given the vast difference in how they are built in the first place. As a result, we end up finding it prohibitively expensive to work out a training recipe for trainable interpolated vision encoders to achieve the same performance as our fixed-vision-encoder method.", + "bbox": [ + 215, + 593, + 787, + 804 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Efficiency. Mixing scales only incurs a $5\\%$ overhead (SPHINX-1k v. SPHINX) during training because $(i)$ we limit mixed scale training to stage 2, which itself", + "bbox": [ + 215, + 809, + 802, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Z. Lin et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "constitutes a smaller portion of the total training time than stage 1, and $(ii)$ high resolution training images are usually paired with longer text, further dragging down the portion of the overhead in the total training cost. We also measured the inference cost on actual hardware: On a A100 GPU with a batch size of 1 and flash decoding enabled, generating a 200-token paragraph without mixed scales takes 29GB of memory and 5.2s, and generating with mixed scales takes 33GB and 5.5s. The memory and time overhead is moderate because $(i)$ the memory is dominated by LLM parameters which are invariant to resolution and $(ii)$ processing image tokens is a prefetching process which utilizes GPU computing cores better than the decoding generation process. As visual encoders are frozen and much smaller (3B parameters combined) compared to the language model (13B parameters), we find mixing visual encoders brings negligible overhead expect for inference-time memory usage (up to 6GB, which is manageable on most modern GPUs).", + "bbox": [ + 212, + 146, + 787, + 357 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.4 Extensions of SPHINX", + "text_level": 1, + "bbox": [ + 215, + 381, + 454, + 396 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "With the strong and general multi-modal understanding capability, SPHINX models can be integrated with other vision models to tackle even more challenging tasks, including language-referred segmentation with SAM (35), image inpainting (104) with LaMa (85) and image editing with Stable Diffusion (73). We have also explored SPHINX as supervised learners, e.g., full-data finetuning on ImageNet-1k (74). The details can be found in the supplementary materials.", + "bbox": [ + 212, + 407, + 787, + 512 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 537, + 375, + 555 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Implementation details", + "text_level": 1, + "bbox": [ + 215, + 571, + 454, + 585 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We provide the important training and evaluation details in this section. More information can be found in the supplementary materials.", + "bbox": [ + 212, + 597, + 785, + 627 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Pre-training datasets. We use two image captioning datasets LAION-400M (76) and LAION-COCO (75) for multi-modal alignment. As we full-fine-tune the language model backbone for long steps, we also jointly train with a text-only dataset RefinedWeb (66) to avoid harming its text reasoning capability due to catastrophic forgetting.", + "bbox": [ + 212, + 628, + 795, + 702 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Pre-training configuration. We fine-tune the weight of the large language model and the randomly-initialized visual projections in the pre-training stage, The visual encoders are kept frozen. We only use the low-resolution view in the pre-training stage, as the web-crawled captions in the LAION datasets are relatively coarse.", + "bbox": [ + 212, + 703, + 785, + 777 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Fine-tuning datasets. In the multi-task fine-tuning phase, our objective is to equip the MLLM with the versatile needs of downstream tasks. Building upon insights from prior research (10; 14; 49; 50; 112), we include instruction following data such as LLaVA (50) and ShareGPT (79), exposing the model to tasks", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "A Mixer of Weights, Visual Embeddings and Image Scales for MLLM", + "bbox": [ + 269, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "requiring explicit directives. For general Vision Question Answering (VQA), we leverage datasets like VQAv2 (1) and GQA (33). Expanding the scope to out-of-domain knowledge, we integrate datasets like OKVQA (58) and A-OKVQA (77), providing the model with information beyond the training data. Optical Character Recognition (OCR) datasets, such as OCRVQA (60) and TextCaps (81) are utilized to increase the text understanding ability of SPHINX. We introduce abundant general object detection and pose estimation datasets, such as COCO (47) and LVIS (26) to inspire the model's capabilities of localization, classification, and human pose estimation. To address grounding tasks, we incorporate RefCOCO (34) and VG (36) datasets, training the model to handle referring object localization. Additionally, Grounding Caption datasets, such as those from Flickr30k (69), further refine the understanding of descriptions in the context of image regions. We streamline the training by converting all datasets into a multi-turn conversation format.", + "bbox": [ + 212, + 146, + 787, + 356 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Fine-tuning configuration. The trained and frozen network components are identical as the pre-training stage. Training data are sampled from the mixture of datasets following their natural frequencies, i.e., the chance of a dataset being sampled from is proportional to its original size.", + "bbox": [ + 212, + 357, + 787, + 417 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Model variant definition. We mainly use 3 model variants for experiments: SPHINX, SPHINX-1k and SPHINX-2k, which take increasingly higher resolution images as inputs. The relevant details are listed in the supplementary materials. All models use LLaMA-2-13B (88) as the language model backbone unless otherwise specified, resulting in a total parameter count of around 16B with the 4 visual encoders.", + "bbox": [ + 212, + 417, + 787, + 507 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation prompt design. In our model evaluation, we prioritize aligning with each benchmark's desired output format. To achieve this, we employ distinct prompts tailored to benchmarks that necessitate long answers, short answers, and multiple-choice responses. The detailed information is provided in the supplementary materials. This approach ensures that our model is capable of handling diverse scenarios.", + "bbox": [ + 212, + 508, + 787, + 599 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Model evaluation", + "text_level": 1, + "bbox": [ + 214, + 619, + 405, + 633 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We provide a comprehensive evaluation of SPHINX and showcase results across multiple benchmarks. Our evaluation encompasses both quantitative metrics (in this section) and qualitative assessments (see the supplementary materials, providing a holistic understanding of our VLM model's performance.", + "bbox": [ + 212, + 643, + 787, + 703 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Benchmarks on multi-modal large language models. We test our model on recently proposed MLLM benchmarks to comprehensively evaluation of the model's characteristic. As shown in Table 1, SPHINX obtains new state-of-the-art performance on 7 out of 11 MLLM benchmarks. Furthermore, the gap between SPHINX and SPHINX-1k on POPE suggests that the introduction of high-resolution sub-images significantly alleviates the visual hallucination problems.", + "bbox": [ + 212, + 704, + 787, + 808 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Visual question answering. We also evaluate our models on various classical visual question answering (VQA) benchmarks in Table 2. SPHINX achieves", + "bbox": [ + 214, + 809, + 787, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Z. Lin et al.", + "bbox": [ + 271, + 114, + 352, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/cd5c70083ad1cbb97cc3f63da6156179b38d91ccd20fb1d13b1d3483e7316da0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsPOPE (45)\\(MME^P\\) (21)\\(MMEC^C\\) (21)MMB (52)SEED (39)\\(LLawa^W\\) (50)MM-Vet (105)CCbench (13)MathVista (54)TinyLVLM (78)Touchstone (5)
BLIP-2 (41)85.31293.8--46.438.122.4--284.7-
InstructBLIP-7B (14)---36.053.460.926.212.125.3300.6552.4
InstructBLIP-13B (14)78.91212.8---58.225.6----
Shikra (10)---58.8-------
LLaMA-AdapterV2 (22)-1328.4356.4------229.2590.1
Qwen-VL-7B (4)---38.256.3--5.5---
Qwen-VL-7B-Chat (4)-1487.6360.760.658.2--39.3-316.8645.2
LLaVA1.5-7B (49)85.91510.7-64.358.663.430.516.4---
LLaVA1.5-13B (49)85.91531.3295.467.761.670.735.426.5---
SPHINX80.71476.1322.266.969.173.536.025.627.0-632.4
SPHINX-1k90.81560.2310.067.171.674.336.627.927.5288.9645.0
SPHINX-2k87.21470.6326.865.971.676.940.227.427.8-659.5
", + "bbox": [ + 218, + 143, + 782, + 287 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/213e9e0329e3906acbe654d23be34a860c038de2c64cdeebd6a964443bac6ad1.jpg", + "table_caption": [ + "Table 1: Comparison with state-of-the-art methods on MLLM benchmarks." + ], + "table_footnote": [], + "table_body": "
MethodsGeneral VQAText-Oriented VQA
OK-VQA (58)VQAv2 (1)VizWiz (27)GQA (33)VSR (48)ScienceQA (55)IconVQA (56)TextVQA (82)OCR-VQA (60)
BLIP-2 (41)45.9-19.641.050.9-40.6-40.6
InstructBLIP (14)--33.449.552.1-44.8-44.8
LLaMA-Adapter-v2 (22)49.670.739.845.1---37.4-
Shikra (10)47.277.4-------
Fuyu-8B (6)60.674.2-------
MiniGPT-v2 (9)57.8-53.660.162.9-51.5--
Qwen-VL-7B (4)58.679.535.259.363.867.1-63.875.7
Qwen-VL-7B-Chat (4)56.678.238.957.561.568.2-61.570.5
LLaVA1.5-7B (49)-78.550.062.0-66.8-58.2-
LLaVA1.5-13B (49)-80.053.663.3-71.6-61.3-
SPHINX62.178.139.962.658.569.350.451.666.0
SPHINX-1k62.280.246.862.965.469.152.758.870.0
SPHINX-2k62.680.744.963.157.170.650.561.267.8
", + "bbox": [ + 218, + 333, + 782, + 520 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 2: Comparison with state-of-the-art methods on classical VQA benchmarks.", + "bbox": [ + 215, + 521, + 782, + 546 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "comparative results across all benchmarks. Notably, SPHINX-1k and SPHINX-2k significantly outperform SPHINX on VQAv2 and text-oriented VQA which demand fine-grained visual information, showcasing the effectiveness of our scale mix-up approach for achieving high resolution without relying on a visual encoder trained specifically on high-resolution images. Although the performances of SPHINX on text-oriented VQA surpass strong baselines such as BLIP-2 and InstructBLIP, it falls behind Qwen-VL-7B, which we conjecture is due to the lack of text-related pre-training data. Introducing more such datasets for training may resolve the issue in the future.", + "bbox": [ + 212, + 595, + 797, + 731 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Visual grounding. Table 3 evaluates SPHINX on REferring Comprehension (REC) benchmarks with RefCOCO (34), RefCOCO+ (57), and RefCOCOg (57) datasets. SPHINX outperforms most state-of-the-art models, including specialist model G-DINO-L (51) and other visual-language generalist models. Compared to a recent strong baseline Qwen-VL-7B (4), which also leverages the large language model for visual understanding, our model still achieves better results across all splits by a large margin. Moreover, SPHINX-1k and SPHINX-2k", + "bbox": [ + 212, + 734, + 787, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "A Mixer of Weights, Visual Embeddings and Image Scales for MLLM", + "bbox": [ + 269, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/885c4c1687d4e89ba7d06026b4a57337d420bf87901ff91e1fe87d206413f55e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsRefCOCO+RefCOCORefCOCOg
valtest-Atest-Bvaltest-Atest-Bval-utest-u
Specialist models
UNINEXT (99)85.2489.6379.7992.6494.3391.4688.7389.37
G-DINO-L (51)82.7588.9575.9290.5693.1988.2486.1387.02
Generalist models
VisionLLM-H (92)----86.70---
OFA-L (91)68.2976.0061.7579.9683.6776.3967.5767.58
Shikra 7B (10)81.6087.3672.1287.0190.6180.2482.2782.19
Shikra 13B (10)82.8987.7974.4187.8391.1181.8182.6483.16
MiniGPT-v2 7B (9)79.9785.1274.4588.6991.6585.3384.4484.66
MiniGPT-v2-7B-Chat (9)79.5885.5273.3288.0691.2984.3084.1984.31
Qwen-VL-7B (4)83.1288.2577.2189.3692.2685.3485.5885.48
Qwen-VL-7B-Chat (4)82.8288.5976.7988.5592.2784.5185.9686.32
SPHINX82.7787.2976.8589.1591.3785.1384.8783.65
SPHINX-1k86.6491.0880.3591.0592.6586.5688.1988.35
SPHINX-2k85.5190.6280.4591.1092.8887.0788.0788.65
", + "bbox": [ + 246, + 143, + 754, + 378 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/e524ddd56c8f032e3352abab11ad43d06ab8cc4facb0ca4da72aab21a8ca3fd2.jpg", + "table_caption": [ + "Table 3: Performance comparisons (Top-1 Accuracy@0.5) on the referring expression comprehension task." + ], + "table_footnote": [], + "table_body": "
patch feat.global feat.MMEGeneral VQAText-oriented VQARefCOCO+RefCOCORefCOCOg
Conv NeXtCLIP LargeDINO v2Q-FormerPCOK VQAViz WizGQAVSRText VQAOCR VQAvaltest-Atest-Bvaltest-Atest-Bvaltest
✓✓✓1392.9299.260.540.359.853.050.461.780.5286.2075.4187.8590.7383.8282.7783.10
✓✓1428.4287.160.347.959.151.649.561.674.7982.3067.4782.1787.8576.7076.4876.20
✓✓✓1246.7267.560.946.160.055.251.365.982.1486.5575.3788.7390.6983.7084.1283.22
✓✓✓1371.7298.959.740.259.751.950.667.079.6185.8173.2386.0289.8080.6581.0081.58
✓✓✓1379.2274.660.547.759.952.251.666.781.0886.5175.0087.7990.9482.5182.1182.14
✓1364.0263.960.740.158.853.450.464.837.2242.4532.0746.5250.2442.6839.8840.79
✓✓✓✓1476.1322.262.139.962.658.551.666.082.7787.2876.8589.1491.3785.1284.8683.64
", + "bbox": [ + 222, + 436, + 782, + 540 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 4: SPHINX visual embedding ablation study. The best result in each column is marked bold and the significantly worse results are marked red.", + "bbox": [ + 215, + 542, + 785, + 569 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "enable the use of high-resolution input images, leading to consecutive improvement over SPHINX and narrowing down the gap to the strong specialist model UNINEXT, which adopts a larger input image size. These results demonstrate the competitive capability of SPHINX for visual grounding.", + "bbox": [ + 212, + 614, + 785, + 676 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Ablation study", + "text_level": 1, + "bbox": [ + 215, + 703, + 385, + 718 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "On the influence of image-text joint pre-training. We show in Figure 4 the effects of jointly training with a text-only dataset (i.e., RefinedWeb) in addition to the image-captioning dataset (i.e., LAION-400M). For comparison, we run an experiment where only LAION-400M is used for training, and observe the text generation loss and image captioning loss as a measurement of its text and image modeling capability, respectively. The text generation loss gradually grows if the model is not trained with RefinedWeb, showing that our joint-training scheme", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Z. Lin et al.", + "bbox": [ + 271, + 114, + 352, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/ede0b422c28622221e07117d519dc8672c092f629d6136092c21b0b8736a5f14.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
viewsMMEGeneral VQAText-oriented VQARefCOCO+RefCOCORefCOCOg
low. sub-res. imagesPCOK VQAViz WizGQAVSRText VQAOCR VQAvaltest-Atest-Bvaltest-Atest-Bvaltest
✓1476.1322.262.139.962.658.551.666.082.7787.2876.8589.1491.3785.1284.8683.64
✓1461.9255.361.846.162.054.057.569.169.8677.4562.7871.7277.0964.7469.1570.89
✓✓1560.2310.062.246.862.965.458.870.086.6491.0880.3591.0592.6586.5688.1988.35
", + "bbox": [ + 222, + 143, + 782, + 224 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 5: Ablation study on mixed scales. For comparison, we remove either the low-resolution view or the sub-image views from SPHINX-1k (note that SPHINX-1k without sub-images is equivalent to SPHINX). The best results are marked bold and the second-best results are marked with underlines.", + "bbox": [ + 215, + 224, + 785, + 280 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "is effective in preserving the text reasoning capability while adapting to cross-modal understanding.", + "bbox": [ + 212, + 324, + 782, + 354 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "On the influence of image-text joint pre-training. We show in Figure 4 the effects of jointly training with a text-only dataset (i.e., RefinedWeb) in addition to the image captioning dataset (i.e., LAION-400M). For comparison, we run an experiment where the model is trained only with LAION-400M and without RefinedWeb, and observe the text generation loss and image captioning loss as a measurement of its text and image modeling capability, respectively. The text generation loss gradually grows if the model is not trained with RefinedWeb, showing that our joint-training scheme is effective in preserving the text reasoning capability while adapting to cross-modal understanding.", + "bbox": [ + 212, + 354, + 784, + 491 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "On the influence of mixed visual embeddings. To validate the effectiveness of visual embedding mix-up, we conducted experiments on SPHINX employing different combinations of visual encoders. The results are presented in Table 4. Interestingly, mixing the encoders exhibits different behaviors on different types of tasks: The mixture of all image encoders do not steadily outperforms its subsets on the VQA benchmarks, which we conjecture is because some datasets are either saturated or out-of-domain for some encoders, while we have noticed a huge gain on visual grounding tasks with the mixture of encoders compared to using only the global features (Q-Former in our case), certifying the importance of spatially-structured features for such fine-grained tasks. We also observed that DINOv2, which is a self-supervised encoder and complementary to the remaining image-text-based encoders, tend to give better results on the visual grounding tsks. As adding visual encoders brings negligible overhead, as analyzed in Sec. 3.3, we choose to use all the encoders by default as they obtain the best performance on most tests.", + "bbox": [ + 212, + 491, + 785, + 717 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "On the influence of mixed scales. To give some insights into our mixed-scale strategy, we perform multi-task fine-tuning with SPHINX-1k, but remove either the low-resolution view or the sub-images to see their respective effects on various benchmarks. The results in Table 5 demonstrate the importance of both types of views: The combination of the two obtains the best results on most benchmarks. Our observation also highlights the critical role of the low-resolution view: Using only the sub-images, the model under-performs the SPHINX baseline on MME and referring comprehension despite the much longer input se", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "A Mixer of Weights, Visual Embeddings and Image Scales for MLLM", + "bbox": [ + 269, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "quence. This emphasizes the importance of the low-resolution view as a global representation which helps glue together the detailed information captured in the sub-images.", + "bbox": [ + 212, + 146, + 787, + 193 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/ff3eb064e3f1f88ecc3a2d26faca562b386a172f29729c3af84f83dadc405c6d.jpg", + "image_caption": [ + "Fig. 4: Loss curve in the pre-training stage with and without optimizing on RefinedWeb. As we perform 1-epoch training on massive datasets, the curve can serve as both training and validation losses, as no sample is seen twice in the process." + ], + "image_footnote": [], + "bbox": [ + 334, + 220, + 669, + 362 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 473, + 359, + 489 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we propose SPHINX, a versatile multi-modal large language model (MLLM) with multi-purpose visual instruction-following capabilities. In our MLLM, we introduce a joint mixing of three different aspects: model weights of pre-trained LLMs by real-world and synthetic data, visual embeddings from different types of vision backbones, and multi-scale views of images with subimage embeddings. Via our proposed three-fold mixing strategy, SPHINX achieves impressive performance over a wide range of multi-modality evaluation benchmarks, and can serve as a strong vision generalist to tackle object detection, region-level captioning, and human pose estimation, etc. Our MLLM can also be integrated with other visual foundation models for wider functionalities, e.g., SAM (35) for language-referred segmentation and Stable Diffusion (73) for image editing.", + "bbox": [ + 212, + 507, + 785, + 686 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations. Limited by the context length of LLaMA, it is currently difficult for our model to process multiple images, interleaved images, videos or very large input images (e.g., beyond $1000 \\times 1000$ ). However, we believe this can be resolved in the near future with long-context LLMs under active research.", + "bbox": [ + 212, + 688, + 785, + 748 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Potential societal impacts. Our method does not pose new negative societal impacts other than those existing in the pre-train models and training datasets themselves to our knowledge. With all base models and datasets used in the paper being publicly available, necessary censorship can be made based on specific ethnicity, regulations, etc. Our method has been shown to reduce hallucination, which could be generally considered a positive societal impacts.", + "bbox": [ + 212, + 750, + 785, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Z. Lin et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements. This project is funded in part by National Key R&D Program of China Project 2022ZD0161100 and 2022ZD0160102, by National Natural Science Foundation of China (Grant No. 62206272), by the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK, by Smart Traffic Fund PSRI/76/2311/PR and by RGC General Research Fund Project 14204021. Hongsheng Li is a PI of CPII under the InnoHK.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Bibliography", + "text_level": 1, + "bbox": [ + 426, + 310, + 576, + 333 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Agrawal, A., Lu, J., Antol, S., Mitchell, M., Zitnick, C.L., Parikh, D., Ba-tra, D.: Vqa: Visual question answering. International Journal of Computer Vision 123, 4 - 31 (2015)", + "[2] Aiello, E., Yu, L., Nie, Y., Aghajanyan, A., Oguz, B.: Jointly training large autoregressive multimodal models. arXiv preprint arXiv:2309.15564 (2023)", + "[3] Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems 35, 23716-23736 (2022)", + "[4] Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-vl: A frontier large vision-language model with versatile abilities. ArXiv abs/2308.12966 (2023)", + "[5] Bai, S., Yang, S., Bai, J., Wang, P., Zhang, X., Lin, J., Wang, X., Zhou, C., Zhou, J.: Touchstone: Evaluating vision-language models by language models (2023)", + "[6] Bavishi, R., Elsen, E., Hawthorne, C., Nye, M., Odena, A., Somani, A., Taşirlar, S.: Introducing our multimodal models (2023), https://www. adept.ai/blog/fuyu-8b", + "[7] Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)", + "[8] Chen, J., Guo, H., Yi, K., Li, B., Elhoseiny, M.: Visualgpt: Data-efficient adaptation of pretrained language models for image captioning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18030-18040 (2022)", + "[9] Chen, J., Li, D.Z.X.S.X., Zhang, Z.L.P., Xiong, R.K.V.C.Y., Elhoseiny, M.: Minigpt-v2: Large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478 (2023)" + ], + "bbox": [ + 233, + 380, + 784, + 840 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[10] Chen, K., Zhang, Z., Zeng, W., Zhang, R., Zhu, F., Zhao, R.: Shikra: Unleashing multimodal lvm's referential dialogue magic. arXiv preprint arXiv:2306.15195 (2023)", + "[11] Chen, X., Wang, X., Changpinyo, S., Piergiovanni, A., Padlewski, P., Salz, D., Goodman, S., Grycner, A., Mustafa, B., Beyer, L., et al.: Pali: A jointly-scaled multilingual language-image model. arXiv preprint arXiv:2209.06794 (2022)", + "[12] Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., Stoica, I., Xing, E.P.: Vicuna: An open-source chatbot impressing gpt-4 with $90\\%$ chatgpt quality. https://lmsys.org/blog/2023-03-30-vicuna/ (March 2023)", + "[13] Contributors, O.: Opencompass: A universal evaluation platform for foundation models. https://github.com/open-compass/opencompass (2023)", + "[14] Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B.A., Fung, P., Hoi, S.C.H.: Instructclip: Towards general-purpose vision-language models with instruction tuning. ArXiv abs/2305.06500 (2023)", + "[15] Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)", + "[16] Dong, R., Han, C., Peng, Y., Qi, Z., Ge, Z., Yang, J., Zhao, L., Sun, J., Zhou, H., Wei, H., et al.: Dreamllm: Synergistic multimodal comprehension and creation. arXiv preprint arXiv:2309.11499 (2023)", + "[17] Dong, X., Bao, J., Zhang, T., Chen, D., Gu, S., Zhang, W., Yuan, L., Chen, D., Wen, F., Yu, N.: Clip itself is a strong fine-tuner: Achieving 85.7% and 88.0% top-1 accuracy with vit-b and vit-l onImagenet. arXiv preprint arXiv:2212.06138 (2022)", + "[18] Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)", + "[19] Douillard, A., Feng, Q., Rusu, A.A., Chhaparia, R., Donchev, Y., Kuncoro, A., Ranzato, M., Szlam, A., Shen, J.: Diloco: Distributed low-communication training of language models. arXiv preprint arXiv:2311.08105 (2023)", + "[20] Fu, C., Chen, P., Shen, Y., Qin, Y., Zhang, M., Lin, X., Qiu, Z., Lin, W., Yang, J., Zheng, X., et al.: Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394 (2023)", + "[21] Fu, C., Chen, P., Shen, Y., Qin, Y., Zhang, M., Lin, X., Qiu, Z., Lin, W., Yang, J., Zheng, X., et al.: Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394 (2023)", + "[22] Gao, P., Han, J., Zhang, R., Lin, Z., Geng, S., Zhou, A., Zhang, W., Lu, P., He, C., Yue, X., et al.: Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010 (2023)" + ], + "bbox": [ + 225, + 145, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Z. Lin et al.", + "bbox": [ + 271, + 114, + 352, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[23] Girdhar, R., El-Nouby, A., Liu, Z., Singh, M., Alwala, K.V., Joulin, A., Misra, I.: Imagebind one embedding space to bind them all. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 15180-15190 (2023)", + "[24] Google: Bard. https://bard.google.com/ (2023)", + "[25] Guo, Z., Zhang, R., Zhu, X., Tang, Y., Ma, X., Han, J., Chen, K., Gao, P., Li, X., Li, H., Heng, P.A.: Point-bind & point-llm: Aligning point cloud with multi-modality for 3d understanding, generation, and instruction following. ArXiv abs/2309.00615 (2023)", + "[26] Gupta, A., Dollar, P., Girshick, R.: Lvis: A dataset for large vocabulary instance segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5356-5364 (2019)", + "[27] Gurari, D., Li, Q., Stangl, A., Guo, A., Lin, C., Grauman, K., Luo, J., Bigham, J.P.: Vizwiz grand challenge: Answering visual questions from blind people. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition pp. 3608-3617 (2018)", + "[28] Han, J., Zhang, R., Shao, W., Gao, P., Xu, P., Xiao, H., Zhang, K., Liu, C., Wen, S., Guo, Z., et al.: Imagebind-llm: Multi-modality instruction tuning. arXiv preprint arXiv:2309.03905 (2023)", + "[29] He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)", + "[30] Hong, Y., Zhen, H., Chen, P., Zheng, S., Du, Y., Chen, Z., Gan, C.: 3d-dllm: Injecting the 3d world into large language models. arXiv preprint arXiv:2307.12981 (2023)", + "[31] Huang, C., Liu, Q., Lin, B.Y., Pang, T., Du, C., Lin, M.: Lorahub: Efficient cross-task generalization via dynamic lora composition. arXiv preprint arXiv:2307.13269 (2023)", + "[32] Huang, S., Dong, L., Wang, W., Hao, Y., Singhal, S., Ma, S., Lv, T., Cui, L., Mohammed, O.K., Liu, Q., et al.: Language is not all you need: Aligning perception with language models. arXiv preprint arXiv:2302.14045 (2023)", + "[33] Hudson, D.A., Manning, C.D.: Gqa: A new dataset for real-world visual reasoning and compositional question answering. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 6693-6702 (2019)", + "[34] Kazemzadeh, S., Ordonez, V., andre Matten, M., Berg, T.L.: Referitgame: Referring to objects in photographs of natural scenes. In: Conference on Empirical Methods in Natural Language Processing (2014)", + "[35] Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)", + "[36] Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., et al.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision 123, 32-73 (2017)" + ], + "bbox": [ + 225, + 145, + 785, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "A Mixer of Weights, Visual Embeddings and Image Scales for MLLM", + "bbox": [ + 269, + 113, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Li, B., Zhang, Y., Chen, L., Wang, J., Pu, F., Yang, J., Li, C., Liu, Z.: Mimic-it: Multi-modal in-context instruction tuning. ArXiv abs/2306.05425 (2023)", + "[38] Li, B., Zhang, Y., Chen, L., Wang, J., Yang, J., Liu, Z.: Otter: A multimodal model with in-context instruction tuning. ArXiv abs/2305.03726 (2023)", + "[39] Li, B., Wang, R., Wang, G., Ge, Y., Ge, Y., Shan, Y.: Seed-bench: Benchmarking multimodal llms with generative comprehension. ArXiv abs/2307.16125 (2023)", + "[40] Li, H., Xu, Z., Taylor, G., Studer, C., Goldstein, T.: Visualizing the loss landscape of neural nets. Advances in neural information processing systems 31 (2018)", + "[41] Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023)", + "[42] Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022)", + "[43] Li, M., Gururangan, S., Dettmers, T., Lewis, M., Althoff, T., Smith, N.A., Zettlemoyer, L.: Branch-train-merge: Embarrassingly parallel training of expert language models. arXiv preprint arXiv:2208.03306 (2022)", + "[44] Li, Y., Mao, H., Girshick, R., He, K.: Exploring plain vision transformer backbones for object detection. In: European Conference on Computer Vision. pp. 280-296. Springer (2022)", + "[45] Li, Y., Du, Y., Zhou, K., Wang, J., Zhao, W.X., Wen, J.R.: Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355 (2023)", + "[46] Lin, T.Y., Dollar, P., Girshick, R., He, K., Hariharan, B., Belongie, S.: Feature pyramid networks for object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2117-2125 (2017)", + "[47] Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dolar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)", + "[48] Liu, F., Emerson, G.E.T., Collier, N.: Visual spatial reasoning. Transactions of the Association for Computational Linguistics (2023)", + "[49] Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. ArXiv abs/2310.03744 (2023)", + "[50] Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv preprint arXiv:2304.08485 (2023)", + "[51] Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., yue Li, C., Yang, J., Su, H., Zhu, J.J., Zhang, L.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. ArXiv abs/2303.05499 (2023)" + ], + "bbox": [ + 225, + 145, + 784, + 840 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Z. Lin et al.", + "bbox": [ + 271, + 114, + 352, + 126 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Liu, Y., Duan, H., Zhang, Y., Li, B., Zhang, S., Zhao, W., Yuan, Y., Wang, J., He, C., Liu, Z., et al.: Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281 (2023)", + "[53] Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3431-3440 (2015)", + "[54] Lu, P., Bansal, H., Xia, T., Liu, J., yue Li, C., Hajishirzi, H., Cheng, H., Chang, K.W., Galley, M., Gao, J.: Mathvista: Evaluating math reasoning in visual contexts with gpt-4v, bard, and other large multimodal models. ArXiv abs/2310.02255 (2023)", + "[55] Lu, P., Mishra, S., Xia, T., Qiu, L., Chang, K.W., Zhu, S.C., Tafjord, O., Clark, P., Kalyan, A.: Learn to explain: Multimodal reasoning via thought chains for science question answering. In: The 36th Conference on Neural Information Processing Systems (NeurIPS) (2022)", + "[56] Lu, P., Qiu, L., Chen, J., Xia, T., Zhao, Y., Zhang, W., Yu, Z., Liang, X., Zhu, S.C.: Iconqa: A new benchmark for abstract diagram understanding and visual language reasoning. In: The 35th Conference on Neural Information Processing Systems (NeurIPS) Track on Datasets and Benchmarks (2021)", + "[57] Mao, J., Huang, J., Toshev, A., Camburu, O.M., Yuille, A.L., Murphy, K.P.: Generation and comprehension of unambiguous object descriptions. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) pp. 11-20 (2015)", + "[58] Marino, K., Rastegari, M., Farhadi, A., Mottaghi, R.: Ok-vqa: A visual question answering benchmark requiring external knowledge. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 3190-3199 (2019)", + "[59] McMahan, B., Moore, E., Ramage, D., Hampson, S., y Arcas, B.A.: Communication-efficient learning of deep networks from decentralized data. In: Artificial intelligence and statistics. pp. 1273-1282. PMLR (2017)", + "[60] Mishra, A., Shekhar, S., Singh, A.K., Chakraborty, A.: Ocr-vqa: Visual question answering by reading text in images. 2019 International Conference on Document Analysis and Recognition (ICDAR) pp. 947-952 (2019)", + "[61] OpenAI: Chatgpt. https://chat.openai.com (2023)", + "[62] OpenAI: Gpt-4 technical report. ArXiv abs/2303.08774 (2023)", + "[63] OpenAI: Vision - openai api. https://platform.openai.com/docs/guides/vision (2023)", + "[64] Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: DiNov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023)", + "[65] Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems 35, 27730-27744 (2022)" + ], + "bbox": [ + 225, + 145, + 785, + 840 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "A Mixer of Weights, Visual Embeddings and Image Scales for MLLM", + "bbox": [ + 269, + 113, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[66] Penedo, G., Malartic, Q., Hesslow, D., Cojocaru, R., Cappelli, A., Alobeidli, H., Pannier, B., Almazrouei, E., Launay, J.: The refined web dataset for falcon llm: outperforming curated corpora with web data, and web data only. arXiv preprint arXiv:2306.01116 (2023)", + "[67] Peng, B., Li, C., He, P., Galley, M., Gao, J.: Instruction tuning with gpt-4. arXiv preprint arXiv:2304.03277 (2023)", + "[68] Peng, Z., Wang, W., Dong, L., Hao, Y., Huang, S., Ma, S., Wei, F.: Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824 (2023)", + "[69] Plummer, B.A., Wang, L., Cervantes, C.M., Caicedo, J.C., Hockenmaier, J., Lazebnik, S.: Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. International Journal of Computer Vision 123, 74 - 93 (2015)", + "[70] Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)", + "[71] Radford, A., Narasimhan, K.: Improving language understanding by generative pre-training (2018)", + "[72] Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners. OpenAI blog 1(8), 9 (2019)", + "[73] Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models (2021)", + "[74] Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al.: Imagenet large scale visual recognition challenge. International journal of computer vision 115, 211-252 (2015)", + "[75] Schuhmann, C., Köpf, A., Vencu, R., Coombes, T., Beaumont, R.: Laion-coco. https://laion.ai/blog/laion-coco/ (2022)", + "[76] Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114 (2021)", + "[77] Schwenk, D., Khandelwal, A., Clark, C., Marino, K., Mottaghi, R.: Akvqa: A benchmark for visual question answering using world knowledge. In: European Conference on Computer Vision (2022)", + "[78] Shao, W., Hu, Y., Gao, P., Lei, M., Zhang, K., Meng, F., Xu, P., Huang, S., Li, H., Qiao, Y., et al.: Tiny lvlm-ehub: Early multimodal experiments with bard. arXiv preprint arXiv:2308.03729 (2023)", + "[79] ShareGPT: Sharegpt. https://sharegpt.com/ (2023)", + "[80] Shukor, M., Dancette, C., Rame, A., Cord, M.: Univalent: Unified model for image, video, audio and language tasks. Transactions on Machine Learning Research Journal (2023)", + "[81] Sidorov, O., Hu, R., Rohrbach, M., Singh, A.: Textcaps: a dataset for image captioning with reading comprehension. ArXiv abs/2003.12462 (2020)" + ], + "bbox": [ + 225, + 145, + 785, + 839 + ], + "page_idx": 19 + }, + { + "type": "page_number", + "text": "20", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 19 + }, + { + "type": "header", + "text": "Z. Lin et al.", + "bbox": [ + 271, + 114, + 352, + 126 + ], + "page_idx": 19 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[82] Singh, A., Natarajan, V., Shah, M., Jiang, Y., Chen, X., Batra, D., Parikh, D., Rohrbach, M.: Towards vqa models that can read. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 8309-8318 (2019)", + "[83] Su, Y., Lan, T., Li, H., Xu, J., Wang, Y., Cai, D.: Pandagpt: One model to instruction-follow them all. ArXiv abs/2305.16355 (2023)", + "[84] Sung, Y.L., Li, L., Lin, K., Gan, Z., Bansal, M., Wang, L.: An empirical study of multimodal model merging. arXiv preprint arXiv:2304.14933 (2023)", + "[85] Suvorov, R., Logacheva, E., Mashikhin, A., Remizova, A., Ashukha, A., Silvestrov, A., Kong, N., Goka, H., Park, K., Lempitsky, V.: Resolution-robust large mask inpainting with fourier convolutions. arXiv preprint arXiv:2109.07161 (2021)", + "[86] Taori, R., Gulrajani, I., Zhang, T., Dubois, Y., Li, X., Guestrin, C., Liang, P., Hashimoto, T.B.: Stanford alpaca: An instruction-following llama model. https://github.com/tatsu-lab/stanford_alpaca (2023)", + "[87] Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., Lample, G.: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)", + "[88] Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023)", + "[89] Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)", + "[90] Wang, K., Ren, H., Zhou, A., Lu, Z., Luo, S., Shi, W., Zhang, R., Song, L., Zhan, M., Li, H.: Mathcoder: Seamless code integration in llms for enhanced mathematical reasoning. arXiv preprint arXiv:2310.03731 (2023)", + "[91] Wang, P., Yang, A., Men, R., Lin, J., Bai, S., Li, Z., Ma, J., Zhou, C., Zhou, J., Yang, H.: Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In: International Conference on Machine Learning. pp. 23318-23340. PMLR (2022)", + "[92] Wang, W., Chen, Z., Chen, X., Wu, J., Zhu, X., Zeng, G., Luo, P., Lu, T., Zhou, J., Qiao, Y., et al.: Visionllm: Large language model is also an open-ended decoder for vision-centric tasks. arXiv preprint arXiv:2305.11175 (2023)", + "[93] Wen, S., Fang, G., Zhang, R., Gao, P., Dong, H., Metaxas, D.: Improving compositional text-to-image generation with large vision-language models. arXiv preprint arXiv:2310.06311 (2023)", + "[94] Woo, S., Debnath, S., Hu, R., Chen, X., Liu, Z., Kweon, I.S., Xie, S.: Convnext v2: Co-designing and scaling convnets with masked autoencoders. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16133-16142 (2023)" + ], + "bbox": [ + 225, + 145, + 784, + 839 + ], + "page_idx": 20 + }, + { + "type": "header", + "text": "A Mixer of Weights, Visual Embeddings and Image Scales for MLLM", + "bbox": [ + 269, + 113, + 730, + 128 + ], + "page_idx": 20 + }, + { + "type": "page_number", + "text": "21", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 20 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[95] Wortsman, M., Ilharco, G., Gadre, S.Y., Roelofs, R., Gontijo-Lopes, R., Morcos, A.S., Namkoong, H., Farhadi, A., Carmon, Y., Kornblith, S., et al.: Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In: International Conference on Machine Learning. pp. 23965-23998. PMLR (2022)", + "[96] Wu, C., Yin, S., Qi, W., Wang, X., Tang, Z., Duan, N.: Visual chatgpt: Talking, drawing and editing with visual foundation models. arXiv preprint arXiv:2303.04671 (2023)", + "[97] Wu, C., Wang, T., Ge, Y., Lu, Z., Zhou, R., Shan, Y., Luo, P.: $\\pi$ -tuning: Transferring multimodal foundation models with optimal multi-task interpolation. In: International Conference on Machine Learning. pp. 37713-37727. PMLR (2023)", + "[98] Xu, R., Wang, X., Wang, T., Chen, Y., Pang, J., Lin, D.: Pointllm: Empowering large language models to understand point clouds. ArXiv abs/2308.16911 (2023)", + "[99] Yan, B., Jiang, Y., Wu, J., Wang, D., Luo, P., Yuan, Z., Lu, H.: Universal instance perception as object discovery and retrieval. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 15325-15336 (2023)", + "[100] Yang, E., Wang, Z., Shen, L., Liu, S., Guo, G., Wang, X., Tao, D.: Adamerging: Adaptive model merging for multi-task learning. arXiv preprint arXiv:2310.02575 (2023)", + "[101] Yang, Z., Li, L., Wang, J., Lin, K., Azarnasab, E., Ahmed, F., Liu, Z., Liu, C., Zeng, M., Wang, L.: Mm-react: Prompting chatgpt for multimodal reasoning and action. arXiv preprint arXiv:2303.11381 (2023)", + "[102] Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178 (2023)", + "[103] Yu, F., Koltun, V.: Multi-scale context aggregation by dilated convolutions. arXiv preprint arXiv:1511.07122 (2015)", + "[104] Yu, T., Feng, R., Feng, R., Liu, J., Jin, X., Zeng, W., Chen, Z.: Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790 (2023)", + "[105] Yu, W., Yang, Z., Li, L., Wang, J., Lin, K., Liu, Z., Wang, X., Wang, L.: Mm-vet: Evaluating large multimodal models for integrated capabilities. ArXiv abs/2308.02490 (2023)", + "[106] Zhang, R., Han, J., Zhou, A., Hu, X., Yan, S., Lu, P., Li, H., Gao, P., Qiao, Y.: Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199 (2023)", + "[107] Zhang, R., Hu, X., Li, B., Huang, S., Deng, H., Li, H., Qiao, Y., Gao, P.: Prompt, generate, then cache: Cascade of foundation models makes strong few-shot learners. CVPR 2023 (2023)", + "[108] Zhang, S., Roller, S., Goyal, N., Artetxe, M., Chen, M., Chen, S., Dewan, C., Diab, M., Li, X., Lin, X.V., et al.: Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068 (2022)" + ], + "bbox": [ + 217, + 145, + 785, + 840 + ], + "page_idx": 21 + }, + { + "type": "page_number", + "text": "22", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 21 + }, + { + "type": "header", + "text": "Z. Lin et al.", + "bbox": [ + 271, + 114, + 352, + 126 + ], + "page_idx": 21 + } +] \ No newline at end of file diff --git a/2024/SPHINX_ A Mixer of Weights, Visual Embeddings and Image Scales for Multi-modal Large Language Models/cd141534-c355-432b-969a-838688ba26e3_model.json b/2024/SPHINX_ A Mixer of Weights, Visual Embeddings and Image Scales for Multi-modal Large Language Models/cd141534-c355-432b-969a-838688ba26e3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..27f011189ceb721fac11fd019b09b577e212c2fc --- /dev/null +++ b/2024/SPHINX_ A Mixer of Weights, Visual Embeddings and Image Scales for Multi-modal Large Language Models/cd141534-c355-432b-969a-838688ba26e3_model.json @@ -0,0 +1,3313 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.234, + 0.141, + 0.77, + 0.21 + ], + "angle": 0, + "content": "SPHINX: A Mixer of Weights, Visual Embeddings and Image Scales for Multi-modal Large Language Models" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.234, + 0.787, + 0.283 + ], + "angle": 0, + "content": "Ziyi Lin\\(^{12*}\\), Dongyang Liu\\(^{2*}\\), Renrui Zhang\\(^{12*}\\), Peng Gao\\(^{2*\\dagger\\dagger}\\), Longtian Qiu\\(^{23*}\\), Han Xiao\\(^{2}\\), Han Qiu\\(^{2}\\), Wenqi Shao\\(^{2}\\), Keqin Chen\\(^{2}\\), Jiaming Han\\(^{12}\\), Siyuan Huang\\(^{2}\\), Yichi Zhang\\(^{2}\\), Xuming He\\(^{3}\\), Yu Qiao\\(^{2\\dagger}\\), and Hongsheng Li\\(^{124\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.286, + 0.292, + 0.721, + 0.308 + ], + "angle": 0, + "content": "1 Multimedia Laboratory, The Chinese University of Hong Kong" + }, + { + "type": "text", + "bbox": [ + 0.315, + 0.307, + 0.69, + 0.321 + ], + "angle": 0, + "content": "2 Shanghai AI Laboratory 3 ShanghaiTech University" + }, + { + "type": "text", + "bbox": [ + 0.303, + 0.321, + 0.7, + 0.335 + ], + "angle": 0, + "content": "4 Centre for Perceptual and Interactive Intelligence Limited" + }, + { + "type": "text", + "bbox": [ + 0.289, + 0.335, + 0.714, + 0.349 + ], + "angle": 0, + "content": "* Equal contribution † Equal advisory ‡ Project leader" + }, + { + "type": "text", + "bbox": [ + 0.251, + 0.35, + 0.753, + 0.363 + ], + "angle": 0, + "content": "zylin@link.cuhk.edu.hk, gaopeng@pjlab.org.cn, hsli@ee.cuhk.edu.hk" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.397, + 0.744, + 0.691 + ], + "angle": 0, + "content": "Abstract. We present SPHINX, a versatile multi-modal large language model (MLLM) with a joint mixing of model weights, visual embeddings and image scales. First, for stronger vision-language alignment, we unfreeze the large language model (LLM) during pre-training, and introduce a weight mix strategy between LLMs trained by real-world and synthetic data. By directly integrating the weights from two domains, the mixed LLM can efficiently incorporate diverse semantics with favorable robustness. Then, we propose to extract comprehensive visual embeddings from various network architectures, pre-training paradigms, and information granularity, providing language models with more robust image representations. We further propose an efficient strategy aiming to better capture fine-grained appearances of high-resolution images. With a mixing of different scales and high-resolution sub-images, SPHINX attains exceptional visual parsing and reasoning performance on existing evaluation benchmarks. Based on our proposed joint mixing, SPHINX exhibits superior multi-modal understanding capabilities on a wide range of applications, with highlighted fine-grained visual recognition abilities such as region-level understanding, caption grounding, document layout detection, and human pose estimation. We hope our work may cast a light on the exploration of joint mixing in future MLLM research. Code is released at https://github.com/Alpha-VLLM/LLaMA2-Accessory." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.734, + 0.376, + 0.751 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Since the era of big data, large language models (LLMs) have attained tremendous strides (7; 61; 62; 87; 108), showcasing unprecedented application scenarios and generalization capabilities. Despite the phenomenal success, plain LLMs are still only reasoning models, and the lack of perception capability has limited their further evolution to autonomous agents with general intelligence. The mission" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.127 + ], + "angle": 0, + "content": "Z. Lin et al." + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.137, + 0.348, + 0.147 + ], + "angle": 0, + "content": "Emergent Referring Pose Estimation" + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.147, + 0.244, + 0.155 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.148, + 0.37, + 0.165 + ], + "angle": 0, + "content": "Else: \n - Select the keypoints of the person wearing a green suit." + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.166, + 0.244, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.246, + 0.17, + 0.362, + 0.205 + ], + "angle": 0, + "content": "SPINDX: The detected key points are left ear[0.76,0.22], left shoulder[0.78,0.32], right shoulder[0.88,0.30], left elbow[0.74,0.49], right elbow[0.90,0.47], wrist[0.68,0.52], right wrist[0.81,0.46], left hip[0.80,0.65], right hip[0.90,0.64], knee[0.71,0.88], and right knee[0.84,0.88]." + }, + { + "type": "image", + "bbox": [ + 0.374, + 0.14, + 0.484, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.139, + 0.591, + 0.148 + ], + "angle": 0, + "content": "Multi-level Dense Captioning" + }, + { + "type": "text", + "bbox": [ + 0.495, + 0.149, + 0.592, + 0.156 + ], + "angle": 0, + "content": "User: Detect all objectives in this image." + }, + { + "type": "text", + "bbox": [ + 0.495, + 0.157, + 0.62, + 0.198 + ], + "angle": 0, + "content": "SHINE: There are 10 objects shown in the image, including a person [0.21, 0.39, 37.0, 76], a helmet[0.30,0.39,0.36,0.43], two sneakers [0.67,0.75,0.85,0.95,0.76], bicycle[0.03,0.53,0.43,0.76], a cell phone[0.22,0.40,0.33,0.42], and four polo shirt[0.50,0.59,0.59,0.59,0.48,0.44,0.48,0.43,0.43,0.16,0.59]." + }, + { + "type": "image", + "bbox": [ + 0.625, + 0.14, + 0.773, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.231, + 0.218, + 0.351, + 0.226 + ], + "angle": 0, + "content": "Emergent Anomaly Detection" + }, + { + "type": "text", + "bbox": [ + 0.231, + 0.228, + 0.327, + 0.248 + ], + "angle": 0, + "content": "User: Is there defect on the image? If yes detect it and plot the box on the image" + }, + { + "type": "text", + "bbox": [ + 0.33, + 0.228, + 0.48, + 0.248 + ], + "angle": 0, + "content": "SPHINX: Yes, there is a defect on the image. The defect is a scratch on the metal object. The box on the image is \\([0.450.46, 0.500.57]\\), which is the scratch on the metal object." + }, + { + "type": "image", + "bbox": [ + 0.232, + 0.252, + 0.315, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.317, + 0.252, + 0.398, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.401, + 0.252, + 0.481, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.311, + 0.349, + 0.321 + ], + "angle": 0, + "content": "Image Editing (w/ Stable Diffusion)" + }, + { + "type": "text", + "bbox": [ + 0.237, + 0.325, + 0.312, + 0.336 + ], + "angle": 0, + "content": "SPEHIN: Detect road sign in this picture" + }, + { + "type": "image", + "bbox": [ + 0.237, + 0.336, + 0.313, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.318, + 0.325, + 0.393, + 0.333 + ], + "angle": 0, + "content": "SPHINX: Detect dog in this" + }, + { + "type": "image", + "bbox": [ + 0.32, + 0.333, + 0.397, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.402, + 0.325, + 0.477, + 0.333 + ], + "angle": 0, + "content": "SPHINX: Detect monocite in this picture." + }, + { + "type": "image", + "bbox": [ + 0.402, + 0.333, + 0.478, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.237, + 0.378, + 0.304, + 0.387 + ], + "angle": 0, + "content": "Stable Diffusion: A light in the middle of the picture" + }, + { + "type": "image", + "bbox": [ + 0.237, + 0.387, + 0.313, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.318, + 0.378, + 0.396, + 0.386 + ], + "angle": 0, + "content": "Stable Diffusion: A teddy bear sitting in a basket" + }, + { + "type": "image", + "bbox": [ + 0.32, + 0.386, + 0.397, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.402, + 0.378, + 0.479, + 0.386 + ], + "angle": 0, + "content": "Stable Diffusion: A man riding a bicycle" + }, + { + "type": "image", + "bbox": [ + 0.402, + 0.386, + 0.478, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.221, + 0.564, + 0.237 + ], + "angle": 0, + "content": "User: Please provide a short description for this document." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.238, + 0.559, + 0.245 + ], + "angle": 0, + "content": "SPHINX: A blue bicycle" + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.248, + 0.582, + 0.26 + ], + "angle": 0, + "content": "User: Provide a one-sentence caption for the provided image" + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.262, + 0.582, + 0.274 + ], + "angle": 0, + "content": "SPHINK: A man riding a bike in a grassy field." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.277, + 0.582, + 0.29 + ], + "angle": 0, + "content": "User:Generates a detailed description about the image." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.291, + 0.584, + 0.363 + ], + "angle": 0, + "content": "SHANK: The image shows a man standing next to a blue mountain bike in a grassy area. He is standing on the right leg of the bike, with his right leg on the pedal. The man appears to be in the process of mounting the bike, as he is standing on the pedal and holding his foot. There are several yellow flowers scattered throughout the grassy area, adding a touch of color to the scene. This is wearing shorts, which are visible as the stands next to the bike." + }, + { + "type": "text", + "bbox": [ + 0.594, + 0.221, + 0.687, + 0.236 + ], + "angle": 0, + "content": "User: Please provide a short description for this region:" + }, + { + "type": "text", + "bbox": [ + 0.594, + 0.237, + 0.672, + 0.245 + ], + "angle": 0, + "content": "SPHINK: A man in a blue shirt" + }, + { + "type": "text", + "bbox": [ + 0.594, + 0.248, + 0.679, + 0.258 + ], + "angle": 0, + "content": "User: Provide a one-Sentence caption for the provided image" + }, + { + "type": "text", + "bbox": [ + 0.594, + 0.262, + 0.681, + 0.274 + ], + "angle": 0, + "content": "SHINX: A man riding a bike while holding a cell phone." + }, + { + "type": "text", + "bbox": [ + 0.594, + 0.276, + 0.681, + 0.285 + ], + "angle": 0, + "content": "User: Generate a detailed description about the image." + }, + { + "type": "text", + "bbox": [ + 0.594, + 0.288, + 0.68, + 0.355 + ], + "angle": 0, + "content": "SHINK: The image shows a man sitting on a bicycle in a grassy area, and the grassy area is filled with people to be taking a selfie or recording a video while sitting on the bike. The grassy area is filled with people who he is prepared for a safe ride. There are several other bicycles in the scene, with one located to the left of the bike, and another located to the right, and a third one further back. The grassy area is scattered with people standing around, riding a couch on color to the scene." + }, + { + "type": "text", + "bbox": [ + 0.686, + 0.221, + 0.773, + 0.237 + ], + "angle": 0, + "content": "User: Please provide a short description of the region [0.37,4.40,97.80]" + }, + { + "type": "text", + "bbox": [ + 0.686, + 0.238, + 0.771, + 0.251 + ], + "angle": 0, + "content": "SHINK: A cow licking a man's face" + }, + { + "type": "text", + "bbox": [ + 0.686, + 0.253, + 0.772, + 0.264 + ], + "angle": 0, + "content": "User: Provide a one-sentence caphor for the provided image" + }, + { + "type": "text", + "bbox": [ + 0.686, + 0.266, + 0.759, + 0.277 + ], + "angle": 0, + "content": "SPEEDON: A cow is sniffing a bicycle wheel." + }, + { + "type": "text", + "bbox": [ + 0.686, + 0.279, + 0.753, + 0.291 + ], + "angle": 0, + "content": "Uc:Generale a deblie du about the image." + }, + { + "type": "text", + "bbox": [ + 0.686, + 0.294, + 0.771, + 0.35 + ], + "angle": 0, + "content": "SPINDL: The image features a large brown and white cow standing in a grassy field. The cow is standing on the ground, looking at the edge of the field, and it appears to be sniffing a bicycle tire. Another cow is standing slightly to the right of the main cow. The scene captures the cow's perspective as it interacts with the bicycle tire." + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.451, + 0.788, + 0.507 + ], + "angle": 0, + "content": "Fig. 1: Examples of multi-task visual reasoning on selected tasks by our proposed SPHINX models, which excels in diverse visual understanding and perception tasks, such as object detection, caption grounding, and region-level description, displaying some emergent capabilities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.521, + 0.785, + 0.58 + ], + "angle": 0, + "content": "to integrate reasoning and perception has motivated the research on multi-modal large language models (MLLMs) (41; 50; 106; 109; 112), which make responses to not only human language, but also natural signals, among which visual signals are predominantly studied." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.582, + 0.787, + 0.702 + ], + "angle": 0, + "content": "Early works on MLLMs (22; 50; 102) focus on coarse-grained tasks, such as image captioning and question answering. While this enables visual instruction following or visual dialogue to some extent, fine-grained visual recognition capabilities, such as visual grounding, key-point detection and referring generation, might bring new outlook to the development of MLLMs: They make possible new applications such as precise actions in embodied AI, as well as potentially improved diversity and reduced hallucination on traditional tasks with better access to detailed visual clues." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.787, + 0.794 + ], + "angle": 0, + "content": "Besides fine-grained recognition, it is equally important for a foundational MLLM to maintain the performance on a wider domain. Despite the success of previous works such as Shikra (10) which tackles fine-grained tasks alone, it is still challenging to explore the scaling-up methods to utilize larger datasets, higher parameter counts, higher resolutions, etc. in order to equip the models with broader or even emergent abilities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "We thus present the building of SPHINX, our proposed foundational MLLM with versatility and an emphasis on fine-grained visual recognition capabilities. Compared to previous works with similar goals, such as Qwen-VL (4)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.27, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "A Mixer of Weights, Visual Embeddings and Image Scales for MLLM" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "and MiniGPT-v2 (9), our pipeline can be highlighted with the following unique scaling-up techniques conveniently named as mixing strategies:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.192, + 0.786, + 0.341 + ], + "angle": 0, + "content": "- Mixed model weights. Vision-language data from particular domains might contain special semantics, e.g., synthetic captions (75) compared to real-world ones (76). Considering that directly mixing such data might confuse the MLLM, we introduce a weight-mixing strategy to efficiently combine such domain-specific knowledge. Based on the MLLM pre-trained on real-world data, we fine-tune it on the synthetic data, and then linearly combine the finetuned LLM's weights with the real-world ones. In this way, the two types of models would not be affected by contradictory data and our final SPHINX can effectively integrate knowledge from both synthetic and real-world domains, enabling effective use of datasets from diverse sources." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.343, + 0.796, + 0.448 + ], + "angle": 0, + "content": "- Mixed visual embeddings. To take the advantage of different encoders, we propose to mix the visual embeddings from various vision backbones (41; 64; 70) with different network architectures (CNN vs. ViT), pre-training paradigms (supervised vs. self-supervised), and information granularity (global vs. local). By mixing the different image tokens channel-wisely and sequence-wisely, SPHINX obtains stronger visual representations and leads to better vision-language alignment efficacy." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.45, + 0.802, + 0.601 + ], + "angle": 0, + "content": "- Mixed scales with high-resolution sub-images. To process high-resolution images as required for fine-grained visual recognition, we first spatially divide the input high-resolution image into multiple sub-images, and also downsample it into a low-resolution one. Then, we feed all the images concurrently into the mixed visual encoders, and concatenate the extracted multiple token groups to represent the entire high-resolution visual features. By mixing visual embeddings of different scales and sub-images, our SPHINX can adaptively explore more fine-grained visual semantics from the high resolution and multi-scale image representations, while maintaining encoding efficiency." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.794, + 0.736 + ], + "angle": 0, + "content": "By the proposed three-fold mixing strategy, SPHINX fine-tunes LLMs, e.g., LLaMA-2 (88), to be a powerful MLLM with superior visual instruction-following capacity. Our model excels in a variety of vision tasks, e.g., detecting different objects with remarkable precision and parsing their relations, or accurately interpreting the content within complicated figures. Importantly, SPHINX can achieve impressive fine-grained visual perception for high-resolution images, which exhibits state-of-the-art performance on extensive evaluation benchmarks, e.g., MMBench (52), MME (20), and POPE (45)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.761, + 0.391, + 0.777 + ], + "angle": 0, + "content": "2 Related works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Large language models (LLMs). The field of Natural Language Processing (NLP) has witnessed significant progress over the years, particularly with the advent of LLMs. With Transformer (89) as the fundamental architecture," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.127 + ], + "angle": 0, + "content": "Z. Lin et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.449 + ], + "angle": 0, + "content": "LLMs (61; 62; 72) have demonstrated unprecedented performance in modeling intricate language patterns over extensive contexts. Therein, BERT (15) showcases the benefits of pre-training on vast text corpora and fine-tuning on specific tasks, setting new standards on various benchmarks. OpenAI's GPT series (61; 62; 71; 72), especially GPT-3 (7), harness the power of massive model scaling, with billions and even trillions of parameters. To obtain better instruction following ability, InstructGPT (65) and ChatGPT (61) are presented to exhibit exceptional fluency and versatility in open-domain conversation tasks, ranging from text generation to question answering. Recently, the instruction tuning based on LLaMA (87) and LLaMA-2 (88) has gained great popularity as open-source LLMs in the community. Therein, Alpaca (86) and LLaMA-Adapter (106) respectively adopt full and parameter-efficient fine-tuning to acquire favorable instruction-following LLMs. Vicuna (12) and GPT-4-LLM (67) further showcase the improvement brought by higher-quality instruction datasets. Other efforts also extend LLMs for match problem solving (90; 111), visual model system (96; 101), and open-world recognition (107; 113). In this paper, we develop our SPHINX based on the superior language understanding of LLaMA-2 (88) and instruction tuning experience of LLaMA-Adapter series (22; 106), which introduce a three-fold mixer to extend the capability ceiling of instruction-following LLMs for multi-modal input." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.456, + 0.788, + 0.804 + ], + "angle": 0, + "content": "Multi-modal large language models (MLLMs). In addition to language instruction following, many efforts have been made to inject multi-modal conditions into LLMs for wider application scenarios. As prior attempts, VisualGPT (8) and BLIP series (14; 41; 42) indicate the potential of aligning LLMs with visual input for image captioning and question answering. Flamingo (3) and Kosmos-1 (32) further exhibit promising multi-modal understanding performance for image-text interleaved contexts. With large-scale pre-training and model sizes, GPT-4 (62) and Bard (24) both showcase remarkable proficiency in vision-language understanding and reasoning over diverse multi-modal tasks. In parallel, a bunch of works have been proposed to align LLaMA with vision modality for advanced visual instruction-following capabilities. LLaVA (50) and MiniGPT-4 (112) utilize a simple projection layer to connect vision encoders (41; 70) with LLMs. LLaMA-Adapter V2 (22) introduces zero-initialized attention mechanisms for efficient visual instruction tuning, and mPLUG-Owl (102) adopts delicately designed intermediate networks for cross-modal alignment. For more modality input, ImageBind-LLM (28) and PandaGPT (83) further incorporate audio and video conditions guided by ImageBind (23). Besides, recent MLLMs are also extended to region-level parsing (10; 68), in-context learning (37; 38), arbitrary image resolutions (6), text-to-image generation (16; 93), and 3D question answering (25; 30; 98). Different from previous works, our SPHINX aims for image-conditioned MLLM, and proposes a three-fold mixer, i.e., model weights, tuning tasks, and visual embeddings, attaining superior generalization capacity for multi-modal learning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Weight interpolation. Despite being commonly known to be highly nonlinear (40), neural networks have, under some conditions, exhibited surprisingly" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.27, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "A Mixer of Weights, Visual Embeddings and Image Scales for MLLM" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.249, + 0.149, + 0.744, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.325, + 0.785, + 0.354 + ], + "angle": 0, + "content": "Fig. 2: The joint mixing paradigm of SPHINX. with mixed visual embeddings (a) and mixed model weights (b)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.381, + 0.785, + 0.488 + ], + "angle": 0, + "content": "good results with linearly interpolated weights in several cases, including higher-performing single-task models (11; 95), task interpolation (31; 97; 100), federated or low-communication distributed training (19; 43; 59). A few recent works (2; 80; 84) also investigated weight mixing under the MLLM context. On top of these works, we further demonstrate the effects of inter-domain weight mixing for image-c captioning pre-trained models, as well as how these benefits persist after further task-specific fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.488, + 0.787, + 0.729 + ], + "angle": 0, + "content": "High-resolution visual embeddings. Fine-grained visual recognition tasks, such as detection and segmentation, traditionally utilize higher-resolution image features to capture detailed information. Convolutional neural networks (CNNs) naturally scale to high-resolution inputs (53) as the convolution operation intrinsically supports varied input sizes, while for the recent transformer-based vision backbones, scaling input resolution is commonly achieved by interpolating the positional embeddings (18). There are also network components specifically designed to obtain high-resolution feature maps, e.g., (44; 46; 103; 110). Unlike those model-surgery-based methods, our mixed-scale solution instead explores the use of fine-tuned large language models to fuse the visual embeddings from multiple parts of the image while leaving the visual encoders themselves untouched. With such a zooming-in mechanism, our framework captures fine-grained visual clues while preserves capabilities of the visual encoders in an architecture-agnostic manner. Interestingly, the recently released GPT-4V API (63) shows a similar correspondence between the image resolution and the number of visual tokens." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.75, + 0.341, + 0.765 + ], + "angle": 0, + "content": "3 SPHINX" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.785, + 0.84 + ], + "angle": 0, + "content": "In this section, we introduce a versatile MLLM, SPHINX, with an model overview in Section 3.1, the joint mixing of model weights and visual embeddings in Section 3.2 and the mixing of high-resolution sub-image tokens Section 3.3. Finally, in Section 3.4, we introduce several extended applications of SPHINX." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.128 + ], + "angle": 0, + "content": "Z. Lin et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.396, + 0.162 + ], + "angle": 0, + "content": "3.1 Model overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.173, + 0.788, + 0.264 + ], + "angle": 0, + "content": "The SPHINX model is composed of an LLM, e.g., LLaMA-2 (88), a mixture of vision encoders and linear projections connecting the vision encoder outputs to the language model inputs. For sequences with image input, visual tokens are extracted from the vision encoders, projected to the LM input space, and then concatenated to the beginning of the sequence. Sequences without image inputs are processed identically as in the language-only LLM." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.265, + 0.79, + 0.34 + ], + "angle": 0, + "content": "We adopt a two-stage training paradigm: the first pre-training stage for vision-language alignment, and the second fine-tuning stage for visual instruction-following learning. In both stages, the LM parameters are fully-tuned, together with the linear layers for visual-language feature alignment, but the visual encoders themselves are kept frozen." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.364, + 0.784, + 0.395 + ], + "angle": 0, + "content": "3.2 Pre-training: The joint mixing of visual embeddings and model weights" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.406, + 0.789, + 0.481 + ], + "angle": 0, + "content": "The pre-training stage (or stage 1) involves two types of mixing: weight mixing and visual embedding mixing, as shown in Figure 2. Since the pre-training datasets are weakly annotated image-caption pairs where the textual information does not contain much detail, we do not use high-resolution images for this stage, thus bypassing the scale mixing." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.482, + 0.796, + 0.694 + ], + "angle": 0, + "content": "Mixed model weights of different domains. Vision-language data from different domains may exhibit distinct characteristics: For example, web-crawled captions like those in LAION-400M (76) tend to be diverse but noisy, but synthetic captions like those in LAION-COCO (75) tend to be coherent but of a relatively limited vocabulary. We propose a weight mixing strategy of domain-specifically tuned LLMs to integrate respective knowledge from real-world and synthetic data. We first utilize the most common domain data (LAION-400M (76)) for pre-training, which endows the MLLM with fundamental visual understanding capabilities. Then, we regard such a pre-trained model as the initial checkpoint to further fine-tune the LLM on synthetic domains, e.g., LAION-COCO (75). Finally, to take advantage of the best data domains, we directly conduct a weighted mixing of two LLMs' weights for semantic aggregation. In detail, we denote the parameters of the fundamental LLM as \\(\\theta_{real}\\), and the fine-tuned parameters by synthetic data as \\(\\theta_{syn}\\). The mixing process is formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.389, + 0.707, + 0.787, + 0.723 + ], + "angle": 0, + "content": "\\[\n\\theta_ {m i x} = \\beta \\cdot \\theta_ {r e a l} + (1 - \\beta) \\cdot \\theta_ {s y n}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.809 + ], + "angle": 0, + "content": "where \\(\\beta\\) denotes the mixing coefficient, and \\(\\theta_{\\text{mix}}\\) represents the mixed LLM weights with aggregated semantics. Compared to fusing different domain data for joint pre-training, our weight mix strategy can encourage every MLLM to better learn domain-unique knowledge, and exhibit flexible scalability for any new data domains." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Mixed embeddings for visual encoding. To capture robust visual representations from different aspects, we propose to ensemble a variety of vision" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.27, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "A Mixer of Weights, Visual Embeddings and Image Scales for MLLM" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.143, + 0.734, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.371, + 0.788, + 0.414 + ], + "angle": 0, + "content": "Fig. 3: Pipeline of SPHINX for high-resolution images. We propose to further mix different scales and sub-images to better capture fine-grained semantics on high-resolution images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.446, + 0.789, + 0.751 + ], + "angle": 0, + "content": "backbones for image encoding. The visual backbones with different characteristics are chosen as follows. 1) Different network architectures. As CNN (29) and ViT (18) mainly aggregate different types of visual appearances, i.e., neighboring dependencies and long-range interactions, we adopt CLIP (70) models respectively with ConvNeXt (94) and ViT image encoders. 2) Different pre-training paradigms. Supervised training can impose explicit semantic information from textual captions or category labels, while self-supervised learning enforces the model to explore implicit pretext task signals. Thus, we further employ the ViT self-supervised by DINOv2 (64) as well as the text-supervised vision encoders, CLIP. 3) Different information granularity. The aforementioned visual encoders all produce visual tokens in the patch level. To better capture global features, we also adopt Q-Former (41) to summarize visual embeddings via querying from the global context. After all the aforementioned encoding, we first channel-wisely concatenate the patch level visual tokens. The channel-wise concatenation allows for better alignment among spatially corresponding features, as well as keeps the number of tokens constant to the number of visual encoders. Then, by using two projection layers for dimension alignment, we spatial-wisely concatenate the representations between those of Q-Former and the other patch-level features. The obtained image tokens are directly placed in front of language instructions, which provide visual context for the language instructions." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.778, + 0.749, + 0.794 + ], + "angle": 0, + "content": "3.3 Fine-tuning: Mixing scales with high-resolution sub-images" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "In the fine-tuning stage, or stage 2 training, the model from stage 1 will adapt for instruct following as well as high-resolution image recognition. Stage 2 introduces" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Z. Lin et al." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.178 + ], + "angle": 0, + "content": "scale mixing to process high-resolution input images, which are crucial for fine-grained visual tasks and will be elaborated in this section." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.182, + 0.788, + 0.59 + ], + "angle": 0, + "content": "Mixed scales and high-resolution sub-images. State-of-the-art open-source MLLMs (9; 10; 22; 41; 50; 68) works adopt frozen image encoders during all training stages, in order to preserve the pre-trained visual semantics. Therefore, the image resolution of MLLMs is usually set as \\(224 \\times 224\\), severely hindering their efficacy for fine-grained visual perception, especially region-level grounding and description. In our SPHINX, we extend the mixing of visual embeddings to more scales and high-resolution sub-images, allowing for efficient high-resolution image encoding. For an input high-resolution image, e.g., \\(448 \\times 448\\), we construct five corresponding images of \\(224 \\times 224\\), and feed them as independent images into our mixed vision encoders. Specifically, we first downsample the input image to \\(224 \\times 224\\) as an abstract representation, and also downsample the input image to \\(448 \\times 448\\) and crop four sub-images of \\(224 \\times 224\\) from the four corners of the \\(448 \\times 448\\) image, which preserve the detailed visual information. In this way, we enable MLLMs to not only capture fine-grained visual appearances with \\(224 \\times 224\\) positional encodings, but also achieve favorable computation efficiency. Afterwards, the five groups of image tokens are encoded and concatenated as a long sequence for feeding into LLM, where the first one group encodes global semantics, and the other four record fine-grained local features. Importantly, as the image tokens of different patches do not have interaction through the vision encoders, they are forced to interact within the LLM to obtain complete visual information. Such a strategy, in turn, motivates LLMs to parse the relations within visual conditions for better cross-modal learning. From this perspective, our SPHINX can be regarded as a new paradigm for similar to ViT (18), where the mixed vision encoders serve as a patch embedding layer, and the LLM plays the role for patch interaction as a vision decoder. On visual understanding tasks requiring higher resolutions, SPHINX achieves significant improvement with the mixed visual representations of scales and high-resolution sub-images." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.594, + 0.788, + 0.805 + ], + "angle": 0, + "content": "On comparison with scaling vision encoder input. The previous standard method to enlarge MLLM input resolutions is to enlarge the input of vision encoders themselves (e.g., by positional embedding interpolation for ViT) and make the vision encoder trainable (4; 9). Compared to vision encoder fine-tuning, we argue that our LLM-based patch fusion is the scalable choice when the vision encoder is large or consists of multiple models as is in our case, because (i) vision encoders can be slow to tune, as the sequence of the same length back-propagating through the vision encoders can be equally costly as backpropagating through the LM, and (ii) vision encoders can be unstable to tune, as different visual backbones may require different tuning configurations (17) given the vast difference in how they are built in the first place. As a result, we end up finding it prohibitively expensive to work out a training recipe for trainable interpolated vision encoders to achieve the same performance as our fixed-vision-encoder method." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.81, + 0.803, + 0.842 + ], + "angle": 0, + "content": "Efficiency. Mixing scales only incurs a \\(5\\%\\) overhead (SPHINX-1k v. SPHINX) during training because \\((i)\\) we limit mixed scale training to stage 2, which itself" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.27, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "A Mixer of Weights, Visual Embeddings and Image Scales for MLLM" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.358 + ], + "angle": 0, + "content": "constitutes a smaller portion of the total training time than stage 1, and \\((ii)\\) high resolution training images are usually paired with longer text, further dragging down the portion of the overhead in the total training cost. We also measured the inference cost on actual hardware: On a A100 GPU with a batch size of 1 and flash decoding enabled, generating a 200-token paragraph without mixed scales takes 29GB of memory and 5.2s, and generating with mixed scales takes 33GB and 5.5s. The memory and time overhead is moderate because \\((i)\\) the memory is dominated by LLM parameters which are invariant to resolution and \\((ii)\\) processing image tokens is a prefetching process which utilizes GPU computing cores better than the decoding generation process. As visual encoders are frozen and much smaller (3B parameters combined) compared to the language model (13B parameters), we find mixing visual encoders brings negligible overhead expect for inference-time memory usage (up to 6GB, which is manageable on most modern GPUs)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.382, + 0.455, + 0.397 + ], + "angle": 0, + "content": "3.4 Extensions of SPHINX" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.409, + 0.788, + 0.513 + ], + "angle": 0, + "content": "With the strong and general multi-modal understanding capability, SPHINX models can be integrated with other vision models to tackle even more challenging tasks, including language-referred segmentation with SAM (35), image inpainting (104) with LaMa (85) and image editing with Stable Diffusion (73). We have also explored SPHINX as supervised learners, e.g., full-data finetuning on ImageNet-1k (74). The details can be found in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.539, + 0.377, + 0.556 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.572, + 0.455, + 0.587 + ], + "angle": 0, + "content": "4.1 Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.787, + 0.628 + ], + "angle": 0, + "content": "We provide the important training and evaluation details in this section. More information can be found in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.796, + 0.703 + ], + "angle": 0, + "content": "Pre-training datasets. We use two image captioning datasets LAION-400M (76) and LAION-COCO (75) for multi-modal alignment. As we full-fine-tune the language model backbone for long steps, we also jointly train with a text-only dataset RefinedWeb (66) to avoid harming its text reasoning capability due to catastrophic forgetting." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.787, + 0.779 + ], + "angle": 0, + "content": "Pre-training configuration. We fine-tune the weight of the large language model and the randomly-initialized visual projections in the pre-training stage, The visual encoders are kept frozen. We only use the low-resolution view in the pre-training stage, as the web-crawled captions in the LAION datasets are relatively coarse." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Fine-tuning datasets. In the multi-task fine-tuning phase, our objective is to equip the MLLM with the versatile needs of downstream tasks. Building upon insights from prior research (10; 14; 49; 50; 112), we include instruction following data such as LLaVA (50) and ShareGPT (79), exposing the model to tasks" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.127 + ], + "angle": 0, + "content": "Z. Lin et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.357 + ], + "angle": 0, + "content": "requiring explicit directives. For general Vision Question Answering (VQA), we leverage datasets like VQAv2 (1) and GQA (33). Expanding the scope to out-of-domain knowledge, we integrate datasets like OKVQA (58) and A-OKVQA (77), providing the model with information beyond the training data. Optical Character Recognition (OCR) datasets, such as OCRVQA (60) and TextCaps (81) are utilized to increase the text understanding ability of SPHINX. We introduce abundant general object detection and pose estimation datasets, such as COCO (47) and LVIS (26) to inspire the model's capabilities of localization, classification, and human pose estimation. To address grounding tasks, we incorporate RefCOCO (34) and VG (36) datasets, training the model to handle referring object localization. Additionally, Grounding Caption datasets, such as those from Flickr30k (69), further refine the understanding of descriptions in the context of image regions. We streamline the training by converting all datasets into a multi-turn conversation format." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.358, + 0.788, + 0.418 + ], + "angle": 0, + "content": "Fine-tuning configuration. The trained and frozen network components are identical as the pre-training stage. Training data are sampled from the mixture of datasets following their natural frequencies, i.e., the chance of a dataset being sampled from is proportional to its original size." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.419, + 0.788, + 0.508 + ], + "angle": 0, + "content": "Model variant definition. We mainly use 3 model variants for experiments: SPHINX, SPHINX-1k and SPHINX-2k, which take increasingly higher resolution images as inputs. The relevant details are listed in the supplementary materials. All models use LLaMA-2-13B (88) as the language model backbone unless otherwise specified, resulting in a total parameter count of around 16B with the 4 visual encoders." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.509, + 0.788, + 0.6 + ], + "angle": 0, + "content": "Evaluation prompt design. In our model evaluation, we prioritize aligning with each benchmark's desired output format. To achieve this, we employ distinct prompts tailored to benchmarks that necessitate long answers, short answers, and multiple-choice responses. The detailed information is provided in the supplementary materials. This approach ensures that our model is capable of handling diverse scenarios." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.621, + 0.406, + 0.635 + ], + "angle": 0, + "content": "4.2 Model evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.704 + ], + "angle": 0, + "content": "We provide a comprehensive evaluation of SPHINX and showcase results across multiple benchmarks. Our evaluation encompasses both quantitative metrics (in this section) and qualitative assessments (see the supplementary materials, providing a holistic understanding of our VLM model's performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.809 + ], + "angle": 0, + "content": "Benchmarks on multi-modal large language models. We test our model on recently proposed MLLM benchmarks to comprehensively evaluation of the model's characteristic. As shown in Table 1, SPHINX obtains new state-of-the-art performance on 7 out of 11 MLLM benchmarks. Furthermore, the gap between SPHINX and SPHINX-1k on POPE suggests that the introduction of high-resolution sub-images significantly alleviates the visual hallucination problems." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Visual question answering. We also evaluate our models on various classical visual question answering (VQA) benchmarks in Table 2. SPHINX achieves" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.27, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "A Mixer of Weights, Visual Embeddings and Image Scales for MLLM" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table", + "bbox": [ + 0.219, + 0.144, + 0.784, + 0.288 + ], + "angle": 0, + "content": "
MethodsPOPE (45)\\(MME^P\\) (21)\\(MMEC^C\\) (21)MMB (52)SEED (39)\\(LLawa^W\\) (50)MM-Vet (105)CCbench (13)MathVista (54)TinyLVLM (78)Touchstone (5)
BLIP-2 (41)85.31293.8--46.438.122.4--284.7-
InstructBLIP-7B (14)---36.053.460.926.212.125.3300.6552.4
InstructBLIP-13B (14)78.91212.8---58.225.6----
Shikra (10)---58.8-------
LLaMA-AdapterV2 (22)-1328.4356.4------229.2590.1
Qwen-VL-7B (4)---38.256.3--5.5---
Qwen-VL-7B-Chat (4)-1487.6360.760.658.2--39.3-316.8645.2
LLaVA1.5-7B (49)85.91510.7-64.358.663.430.516.4---
LLaVA1.5-13B (49)85.91531.3295.467.761.670.735.426.5---
SPHINX80.71476.1322.266.969.173.536.025.627.0-632.4
SPHINX-1k90.81560.2310.067.171.674.336.627.927.5288.9645.0
SPHINX-2k87.21470.6326.865.971.676.940.227.427.8-659.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.218, + 0.289, + 0.784, + 0.302 + ], + "angle": 0, + "content": "Table 1: Comparison with state-of-the-art methods on MLLM benchmarks." + }, + { + "type": "table", + "bbox": [ + 0.219, + 0.334, + 0.784, + 0.521 + ], + "angle": 0, + "content": "
MethodsGeneral VQAText-Oriented VQA
OK-VQA (58)VQAv2 (1)VizWiz (27)GQA (33)VSR (48)ScienceQA (55)IconVQA (56)TextVQA (82)OCR-VQA (60)
BLIP-2 (41)45.9-19.641.050.9-40.6-40.6
InstructBLIP (14)--33.449.552.1-44.8-44.8
LLaMA-Adapter-v2 (22)49.670.739.845.1---37.4-
Shikra (10)47.277.4-------
Fuyu-8B (6)60.674.2-------
MiniGPT-v2 (9)57.8-53.660.162.9-51.5--
Qwen-VL-7B (4)58.679.535.259.363.867.1-63.875.7
Qwen-VL-7B-Chat (4)56.678.238.957.561.568.2-61.570.5
LLaVA1.5-7B (49)-78.550.062.0-66.8-58.2-
LLaVA1.5-13B (49)-80.053.663.3-71.6-61.3-
SPHINX62.178.139.962.658.569.350.451.666.0
SPHINX-1k62.280.246.862.965.469.152.758.870.0
SPHINX-2k62.680.744.963.157.170.650.561.267.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.522, + 0.784, + 0.547 + ], + "angle": 0, + "content": "Table 2: Comparison with state-of-the-art methods on classical VQA benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.596, + 0.799, + 0.732 + ], + "angle": 0, + "content": "comparative results across all benchmarks. Notably, SPHINX-1k and SPHINX-2k significantly outperform SPHINX on VQAv2 and text-oriented VQA which demand fine-grained visual information, showcasing the effectiveness of our scale mix-up approach for achieving high resolution without relying on a visual encoder trained specifically on high-resolution images. Although the performances of SPHINX on text-oriented VQA surpass strong baselines such as BLIP-2 and InstructBLIP, it falls behind Qwen-VL-7B, which we conjecture is due to the lack of text-related pre-training data. Introducing more such datasets for training may resolve the issue in the future." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Visual grounding. Table 3 evaluates SPHINX on REferring Comprehension (REC) benchmarks with RefCOCO (34), RefCOCO+ (57), and RefCOCOg (57) datasets. SPHINX outperforms most state-of-the-art models, including specialist model G-DINO-L (51) and other visual-language generalist models. Compared to a recent strong baseline Qwen-VL-7B (4), which also leverages the large language model for visual understanding, our model still achieves better results across all splits by a large margin. Moreover, SPHINX-1k and SPHINX-2k" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.127 + ], + "angle": 0, + "content": "Z. Lin et al." + }, + { + "type": "table", + "bbox": [ + 0.248, + 0.145, + 0.756, + 0.38 + ], + "angle": 0, + "content": "
MethodsRefCOCO+RefCOCORefCOCOg
valtest-Atest-Bvaltest-Atest-Bval-utest-u
Specialist models
UNINEXT (99)85.2489.6379.7992.6494.3391.4688.7389.37
G-DINO-L (51)82.7588.9575.9290.5693.1988.2486.1387.02
Generalist models
VisionLLM-H (92)----86.70---
OFA-L (91)68.2976.0061.7579.9683.6776.3967.5767.58
Shikra 7B (10)81.6087.3672.1287.0190.6180.2482.2782.19
Shikra 13B (10)82.8987.7974.4187.8391.1181.8182.6483.16
MiniGPT-v2 7B (9)79.9785.1274.4588.6991.6585.3384.4484.66
MiniGPT-v2-7B-Chat (9)79.5885.5273.3288.0691.2984.3084.1984.31
Qwen-VL-7B (4)83.1288.2577.2189.3692.2685.3485.5885.48
Qwen-VL-7B-Chat (4)82.8288.5976.7988.5592.2784.5185.9686.32
SPHINX82.7787.2976.8589.1591.3785.1384.8783.65
SPHINX-1k86.6491.0880.3591.0592.6586.5688.1988.35
SPHINX-2k85.5190.6280.4591.1092.8887.0788.0788.65
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.382, + 0.786, + 0.41 + ], + "angle": 0, + "content": "Table 3: Performance comparisons (Top-1 Accuracy@0.5) on the referring expression comprehension task." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.438, + 0.784, + 0.541 + ], + "angle": 0, + "content": "
patch feat.global feat.MMEGeneral VQAText-oriented VQARefCOCO+RefCOCORefCOCOg
Conv NeXtCLIP LargeDINO v2Q-FormerPCOK VQAViz WizGQAVSRText VQAOCR VQAvaltest-Atest-Bvaltest-Atest-Bvaltest
✓✓✓1392.9299.260.540.359.853.050.461.780.5286.2075.4187.8590.7383.8282.7783.10
✓✓1428.4287.160.347.959.151.649.561.674.7982.3067.4782.1787.8576.7076.4876.20
✓✓✓1246.7267.560.946.160.055.251.365.982.1486.5575.3788.7390.6983.7084.1283.22
✓✓✓1371.7298.959.740.259.751.950.667.079.6185.8173.2386.0289.8080.6581.0081.58
✓✓✓1379.2274.660.547.759.952.251.666.781.0886.5175.0087.7990.9482.5182.1182.14
✓1364.0263.960.740.158.853.450.464.837.2242.4532.0746.5250.2442.6839.8840.79
✓✓✓✓1476.1322.262.139.962.658.551.666.082.7787.2876.8589.1491.3785.1284.8683.64
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.543, + 0.786, + 0.57 + ], + "angle": 0, + "content": "Table 4: SPHINX visual embedding ablation study. The best result in each column is marked bold and the significantly worse results are marked red." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.615, + 0.787, + 0.677 + ], + "angle": 0, + "content": "enable the use of high-resolution input images, leading to consecutive improvement over SPHINX and narrowing down the gap to the strong specialist model UNINEXT, which adopts a larger input image size. These results demonstrate the competitive capability of SPHINX for visual grounding." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.704, + 0.387, + 0.719 + ], + "angle": 0, + "content": "4.3 Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "On the influence of image-text joint pre-training. We show in Figure 4 the effects of jointly training with a text-only dataset (i.e., RefinedWeb) in addition to the image-captioning dataset (i.e., LAION-400M). For comparison, we run an experiment where only LAION-400M is used for training, and observe the text generation loss and image captioning loss as a measurement of its text and image modeling capability, respectively. The text generation loss gradually grows if the model is not trained with RefinedWeb, showing that our joint-training scheme" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.27, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "A Mixer of Weights, Visual Embeddings and Image Scales for MLLM" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.144, + 0.783, + 0.226 + ], + "angle": 0, + "content": "
viewsMMEGeneral VQAText-oriented VQARefCOCO+RefCOCORefCOCOg
low. sub-res. imagesPCOK VQAViz WizGQAVSRText VQAOCR VQAvaltest-Atest-Bvaltest-Atest-Bvaltest
✓1476.1322.262.139.962.658.551.666.082.7787.2876.8589.1491.3785.1284.8683.64
✓1461.9255.361.846.162.054.057.569.169.8677.4562.7871.7277.0964.7469.1570.89
✓✓1560.2310.062.246.862.965.458.870.086.6491.0880.3591.0592.6586.5688.1988.35
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.226, + 0.787, + 0.281 + ], + "angle": 0, + "content": "Table 5: Ablation study on mixed scales. For comparison, we remove either the low-resolution view or the sub-image views from SPHINX-1k (note that SPHINX-1k without sub-images is equivalent to SPHINX). The best results are marked bold and the second-best results are marked with underlines." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.325, + 0.784, + 0.355 + ], + "angle": 0, + "content": "is effective in preserving the text reasoning capability while adapting to cross-modal understanding." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.356, + 0.785, + 0.492 + ], + "angle": 0, + "content": "On the influence of image-text joint pre-training. We show in Figure 4 the effects of jointly training with a text-only dataset (i.e., RefinedWeb) in addition to the image captioning dataset (i.e., LAION-400M). For comparison, we run an experiment where the model is trained only with LAION-400M and without RefinedWeb, and observe the text generation loss and image captioning loss as a measurement of its text and image modeling capability, respectively. The text generation loss gradually grows if the model is not trained with RefinedWeb, showing that our joint-training scheme is effective in preserving the text reasoning capability while adapting to cross-modal understanding." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.492, + 0.787, + 0.718 + ], + "angle": 0, + "content": "On the influence of mixed visual embeddings. To validate the effectiveness of visual embedding mix-up, we conducted experiments on SPHINX employing different combinations of visual encoders. The results are presented in Table 4. Interestingly, mixing the encoders exhibits different behaviors on different types of tasks: The mixture of all image encoders do not steadily outperforms its subsets on the VQA benchmarks, which we conjecture is because some datasets are either saturated or out-of-domain for some encoders, while we have noticed a huge gain on visual grounding tasks with the mixture of encoders compared to using only the global features (Q-Former in our case), certifying the importance of spatially-structured features for such fine-grained tasks. We also observed that DINOv2, which is a self-supervised encoder and complementary to the remaining image-text-based encoders, tend to give better results on the visual grounding tsks. As adding visual encoders brings negligible overhead, as analyzed in Sec. 3.3, we choose to use all the encoders by default as they obtain the best performance on most tests." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "On the influence of mixed scales. To give some insights into our mixed-scale strategy, we perform multi-task fine-tuning with SPHINX-1k, but remove either the low-resolution view or the sub-images to see their respective effects on various benchmarks. The results in Table 5 demonstrate the importance of both types of views: The combination of the two obtains the best results on most benchmarks. Our observation also highlights the critical role of the low-resolution view: Using only the sub-images, the model under-performs the SPHINX baseline on MME and referring comprehension despite the much longer input se" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.128 + ], + "angle": 0, + "content": "Z. Lin et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.194 + ], + "angle": 0, + "content": "quence. This emphasizes the importance of the low-resolution view as a global representation which helps glue together the detailed information captured in the sub-images." + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.221, + 0.67, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.377, + 0.788, + 0.421 + ], + "angle": 0, + "content": "Fig. 4: Loss curve in the pre-training stage with and without optimizing on RefinedWeb. As we perform 1-epoch training on massive datasets, the curve can serve as both training and validation losses, as no sample is seen twice in the process." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.474, + 0.36, + 0.491 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.787, + 0.688 + ], + "angle": 0, + "content": "In this paper, we propose SPHINX, a versatile multi-modal large language model (MLLM) with multi-purpose visual instruction-following capabilities. In our MLLM, we introduce a joint mixing of three different aspects: model weights of pre-trained LLMs by real-world and synthetic data, visual embeddings from different types of vision backbones, and multi-scale views of images with subimage embeddings. Via our proposed three-fold mixing strategy, SPHINX achieves impressive performance over a wide range of multi-modality evaluation benchmarks, and can serve as a strong vision generalist to tackle object detection, region-level captioning, and human pose estimation, etc. Our MLLM can also be integrated with other visual foundation models for wider functionalities, e.g., SAM (35) for language-referred segmentation and Stable Diffusion (73) for image editing." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.787, + 0.749 + ], + "angle": 0, + "content": "Limitations. Limited by the context length of LLaMA, it is currently difficult for our model to process multiple images, interleaved images, videos or very large input images (e.g., beyond \\(1000 \\times 1000\\)). However, we believe this can be resolved in the near future with long-context LLMs under active research." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Potential societal impacts. Our method does not pose new negative societal impacts other than those existing in the pre-train models and training datasets themselves to our knowledge. With all base models and datasets used in the paper being publicly available, necessary censorship can be made based on specific ethnicity, regulations, etc. Our method has been shown to reduce hallucination, which could be generally considered a positive societal impacts." + } + ], + [ + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "Acknowledgements. This project is funded in part by National Key R&D Program of China Project 2022ZD0161100 and 2022ZD0160102, by National Natural Science Foundation of China (Grant No. 62206272), by the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK, by Smart Traffic Fund PSRI/76/2311/PR and by RGC General Research Fund Project 14204021. Hongsheng Li is a PI of CPII under the InnoHK." + }, + { + "type": "title", + "bbox": [ + 0.428, + 0.311, + 0.577, + 0.334 + ], + "angle": 0, + "content": "Bibliography" + }, + { + "type": "ref_text", + "bbox": [ + 0.234, + 0.381, + 0.785, + 0.428 + ], + "angle": 0, + "content": "[1] Agrawal, A., Lu, J., Antol, S., Mitchell, M., Zitnick, C.L., Parikh, D., Ba-tra, D.: Vqa: Visual question answering. International Journal of Computer Vision 123, 4 - 31 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.235, + 0.43, + 0.785, + 0.461 + ], + "angle": 0, + "content": "[2] Aiello, E., Yu, L., Nie, Y., Aghajanyan, A., Oguz, B.: Jointly training large autoregressive multimodal models. arXiv preprint arXiv:2309.15564 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.235, + 0.463, + 0.785, + 0.523 + ], + "angle": 0, + "content": "[3] Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems 35, 23716-23736 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.235, + 0.525, + 0.785, + 0.572 + ], + "angle": 0, + "content": "[4] Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-vl: A frontier large vision-language model with versatile abilities. ArXiv abs/2308.12966 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.235, + 0.573, + 0.785, + 0.619 + ], + "angle": 0, + "content": "[5] Bai, S., Yang, S., Bai, J., Wang, P., Zhang, X., Lin, J., Wang, X., Zhou, C., Zhou, J.: Touchstone: Evaluating vision-language models by language models (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.235, + 0.621, + 0.785, + 0.667 + ], + "angle": 0, + "content": "[6] Bavishi, R., Elsen, E., Hawthorne, C., Nye, M., Odena, A., Somani, A., Taşirlar, S.: Introducing our multimodal models (2023), https://www. adept.ai/blog/fuyu-8b" + }, + { + "type": "ref_text", + "bbox": [ + 0.235, + 0.669, + 0.785, + 0.731 + ], + "angle": 0, + "content": "[7] Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.235, + 0.732, + 0.785, + 0.793 + ], + "angle": 0, + "content": "[8] Chen, J., Guo, H., Yi, K., Li, B., Elhoseiny, M.: Visualgpt: Data-efficient adaptation of pretrained language models for image captioning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18030-18040 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.235, + 0.794, + 0.785, + 0.842 + ], + "angle": 0, + "content": "[9] Chen, J., Li, D.Z.X.S.X., Zhang, Z.L.P., Xiong, R.K.V.C.Y., Elhoseiny, M.: Minigpt-v2: Large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.234, + 0.381, + 0.785, + 0.842 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.127 + ], + "angle": 0, + "content": "Z. Lin et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.146, + 0.785, + 0.191 + ], + "angle": 0, + "content": "[10] Chen, K., Zhang, Z., Zeng, W., Zhang, R., Zhu, F., Zhao, R.: Shikra: Unleashing multimodal lvm's referential dialogue magic. arXiv preprint arXiv:2306.15195 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.193, + 0.787, + 0.253 + ], + "angle": 0, + "content": "[11] Chen, X., Wang, X., Changpinyo, S., Piergiovanni, A., Padlewski, P., Salz, D., Goodman, S., Grycner, A., Mustafa, B., Beyer, L., et al.: Pali: A jointly-scaled multilingual language-image model. arXiv preprint arXiv:2209.06794 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.255, + 0.785, + 0.316 + ], + "angle": 0, + "content": "[12] Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., Stoica, I., Xing, E.P.: Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\) chatgpt quality. https://lmsys.org/blog/2023-03-30-vicuna/ (March 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.317, + 0.785, + 0.348 + ], + "angle": 0, + "content": "[13] Contributors, O.: Opencompass: A universal evaluation platform for foundation models. https://github.com/open-compass/opencompass (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.349, + 0.785, + 0.393 + ], + "angle": 0, + "content": "[14] Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B.A., Fung, P., Hoi, S.C.H.: Instructclip: Towards general-purpose vision-language models with instruction tuning. ArXiv abs/2305.06500 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.395, + 0.785, + 0.44 + ], + "angle": 0, + "content": "[15] Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.441, + 0.785, + 0.486 + ], + "angle": 0, + "content": "[16] Dong, R., Han, C., Peng, Y., Qi, Z., Ge, Z., Yang, J., Zhao, L., Sun, J., Zhou, H., Wei, H., et al.: Dreamllm: Synergistic multimodal comprehension and creation. arXiv preprint arXiv:2309.11499 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.488, + 0.785, + 0.548 + ], + "angle": 0, + "content": "[17] Dong, X., Bao, J., Zhang, T., Chen, D., Gu, S., Zhang, W., Yuan, L., Chen, D., Wen, F., Yu, N.: Clip itself is a strong fine-tuner: Achieving 85.7% and 88.0% top-1 accuracy with vit-b and vit-l onImagenet. arXiv preprint arXiv:2212.06138 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.549, + 0.785, + 0.609 + ], + "angle": 0, + "content": "[18] Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.611, + 0.785, + 0.671 + ], + "angle": 0, + "content": "[19] Douillard, A., Feng, Q., Rusu, A.A., Chhaparia, R., Donchev, Y., Kuncoro, A., Ranzato, M., Szlam, A., Shen, J.: Diloco: Distributed low-communication training of language models. arXiv preprint arXiv:2311.08105 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.672, + 0.785, + 0.733 + ], + "angle": 0, + "content": "[20] Fu, C., Chen, P., Shen, Y., Qin, Y., Zhang, M., Lin, X., Qiu, Z., Lin, W., Yang, J., Zheng, X., et al.: Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.734, + 0.785, + 0.794 + ], + "angle": 0, + "content": "[21] Fu, C., Chen, P., Shen, Y., Qin, Y., Zhang, M., Lin, X., Qiu, Z., Lin, W., Yang, J., Zheng, X., et al.: Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.795, + 0.785, + 0.841 + ], + "angle": 0, + "content": "[22] Gao, P., Han, J., Zhang, R., Lin, Z., Geng, S., Zhou, A., Zhang, W., Lu, P., He, C., Yue, X., et al.: Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.146, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.27, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "A Mixer of Weights, Visual Embeddings and Image Scales for MLLM" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.146, + 0.786, + 0.208 + ], + "angle": 0, + "content": "[23] Girdhar, R., El-Nouby, A., Liu, Z., Singh, M., Alwala, K.V., Joulin, A., Misra, I.: Imagebind one embedding space to bind them all. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 15180-15190 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.209, + 0.623, + 0.224 + ], + "angle": 0, + "content": "[24] Google: Bard. https://bard.google.com/ (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.228, + 0.225, + 0.787, + 0.285 + ], + "angle": 0, + "content": "[25] Guo, Z., Zhang, R., Zhu, X., Tang, Y., Ma, X., Han, J., Chen, K., Gao, P., Li, X., Li, H., Heng, P.A.: Point-bind & point-llm: Aligning point cloud with multi-modality for 3d understanding, generation, and instruction following. ArXiv abs/2309.00615 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.228, + 0.287, + 0.786, + 0.331 + ], + "angle": 0, + "content": "[26] Gupta, A., Dollar, P., Girshick, R.: Lvis: A dataset for large vocabulary instance segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5356-5364 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.228, + 0.333, + 0.786, + 0.393 + ], + "angle": 0, + "content": "[27] Gurari, D., Li, Q., Stangl, A., Guo, A., Lin, C., Grauman, K., Luo, J., Bigham, J.P.: Vizwiz grand challenge: Answering visual questions from blind people. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition pp. 3608-3617 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.228, + 0.395, + 0.786, + 0.44 + ], + "angle": 0, + "content": "[28] Han, J., Zhang, R., Shao, W., Gao, P., Xu, P., Xiao, H., Zhang, K., Liu, C., Wen, S., Guo, Z., et al.: Imagebind-llm: Multi-modality instruction tuning. arXiv preprint arXiv:2309.03905 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.228, + 0.441, + 0.786, + 0.486 + ], + "angle": 0, + "content": "[29] He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.228, + 0.488, + 0.786, + 0.532 + ], + "angle": 0, + "content": "[30] Hong, Y., Zhen, H., Chen, P., Zheng, S., Du, Y., Chen, Z., Gan, C.: 3d-dllm: Injecting the 3d world into large language models. arXiv preprint arXiv:2307.12981 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.228, + 0.534, + 0.786, + 0.579 + ], + "angle": 0, + "content": "[31] Huang, C., Liu, Q., Lin, B.Y., Pang, T., Du, C., Lin, M.: Lorahub: Efficient cross-task generalization via dynamic lora composition. arXiv preprint arXiv:2307.13269 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.228, + 0.58, + 0.786, + 0.625 + ], + "angle": 0, + "content": "[32] Huang, S., Dong, L., Wang, W., Hao, Y., Singhal, S., Ma, S., Lv, T., Cui, L., Mohammed, O.K., Liu, Q., et al.: Language is not all you need: Aligning perception with language models. arXiv preprint arXiv:2302.14045 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.228, + 0.627, + 0.786, + 0.686 + ], + "angle": 0, + "content": "[33] Hudson, D.A., Manning, C.D.: Gqa: A new dataset for real-world visual reasoning and compositional question answering. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 6693-6702 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.228, + 0.688, + 0.786, + 0.734 + ], + "angle": 0, + "content": "[34] Kazemzadeh, S., Ordonez, V., andre Matten, M., Berg, T.L.: Referitgame: Referring to objects in photographs of natural scenes. In: Conference on Empirical Methods in Natural Language Processing (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.228, + 0.735, + 0.786, + 0.779 + ], + "angle": 0, + "content": "[35] Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.228, + 0.781, + 0.786, + 0.841 + ], + "angle": 0, + "content": "[36] Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., et al.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision 123, 32-73 (2017)" + }, + { + "type": "list", + "bbox": [ + 0.227, + 0.146, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.127 + ], + "angle": 0, + "content": "Z. Lin et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.146, + 0.785, + 0.192 + ], + "angle": 0, + "content": "[37] Li, B., Zhang, Y., Chen, L., Wang, J., Pu, F., Yang, J., Li, C., Liu, Z.: Mimic-it: Multi-modal in-context instruction tuning. ArXiv abs/2306.05425 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.192, + 0.785, + 0.237 + ], + "angle": 0, + "content": "[38] Li, B., Zhang, Y., Chen, L., Wang, J., Yang, J., Liu, Z.: Otter: A multimodal model with in-context instruction tuning. ArXiv abs/2305.03726 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.237, + 0.785, + 0.283 + ], + "angle": 0, + "content": "[39] Li, B., Wang, R., Wang, G., Ge, Y., Ge, Y., Shan, Y.: Seed-bench: Benchmarking multimodal llms with generative comprehension. ArXiv abs/2307.16125 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.283, + 0.785, + 0.327 + ], + "angle": 0, + "content": "[40] Li, H., Xu, Z., Taylor, G., Studer, C., Goldstein, T.: Visualizing the loss landscape of neural nets. Advances in neural information processing systems 31 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.328, + 0.785, + 0.373 + ], + "angle": 0, + "content": "[41] Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.373, + 0.785, + 0.433 + ], + "angle": 0, + "content": "[42] Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.433, + 0.785, + 0.478 + ], + "angle": 0, + "content": "[43] Li, M., Gururangan, S., Dettmers, T., Lewis, M., Althoff, T., Smith, N.A., Zettlemoyer, L.: Branch-train-merge: Embarrassingly parallel training of expert language models. arXiv preprint arXiv:2208.03306 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.479, + 0.785, + 0.524 + ], + "angle": 0, + "content": "[44] Li, Y., Mao, H., Girshick, R., He, K.: Exploring plain vision transformer backbones for object detection. In: European Conference on Computer Vision. pp. 280-296. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.228, + 0.524, + 0.785, + 0.569 + ], + "angle": 0, + "content": "[45] Li, Y., Du, Y., Zhou, K., Wang, J., Zhao, W.X., Wen, J.R.: Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.228, + 0.569, + 0.785, + 0.629 + ], + "angle": 0, + "content": "[46] Lin, T.Y., Dollar, P., Girshick, R., He, K., Hariharan, B., Belongie, S.: Feature pyramid networks for object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2117-2125 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.228, + 0.629, + 0.785, + 0.704 + ], + "angle": 0, + "content": "[47] Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dolar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.228, + 0.705, + 0.785, + 0.734 + ], + "angle": 0, + "content": "[48] Liu, F., Emerson, G.E.T., Collier, N.: Visual spatial reasoning. Transactions of the Association for Computational Linguistics (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.228, + 0.735, + 0.785, + 0.765 + ], + "angle": 0, + "content": "[49] Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. ArXiv abs/2310.03744 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.228, + 0.765, + 0.785, + 0.794 + ], + "angle": 0, + "content": "[50] Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv preprint arXiv:2304.08485 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.228, + 0.795, + 0.785, + 0.841 + ], + "angle": 0, + "content": "[51] Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., yue Li, C., Yang, J., Su, H., Zhu, J.J., Zhang, L.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. ArXiv abs/2303.05499 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.227, + 0.146, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.27, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "A Mixer of Weights, Visual Embeddings and Image Scales for MLLM" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.146, + 0.786, + 0.193 + ], + "angle": 0, + "content": "[52] Liu, Y., Duan, H., Zhang, Y., Li, B., Zhang, S., Zhao, W., Yuan, Y., Wang, J., He, C., Liu, Z., et al.: Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.194, + 0.787, + 0.239 + ], + "angle": 0, + "content": "[53] Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3431-3440 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.24, + 0.787, + 0.3 + ], + "angle": 0, + "content": "[54] Lu, P., Bansal, H., Xia, T., Liu, J., yue Li, C., Hajishirzi, H., Cheng, H., Chang, K.W., Galley, M., Gao, J.: Mathvista: Evaluating math reasoning in visual contexts with gpt-4v, bard, and other large multimodal models. ArXiv abs/2310.02255 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.301, + 0.787, + 0.362 + ], + "angle": 0, + "content": "[55] Lu, P., Mishra, S., Xia, T., Qiu, L., Chang, K.W., Zhu, S.C., Tafjord, O., Clark, P., Kalyan, A.: Learn to explain: Multimodal reasoning via thought chains for science question answering. In: The 36th Conference on Neural Information Processing Systems (NeurIPS) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.363, + 0.787, + 0.438 + ], + "angle": 0, + "content": "[56] Lu, P., Qiu, L., Chen, J., Xia, T., Zhao, Y., Zhang, W., Yu, Z., Liang, X., Zhu, S.C.: Iconqa: A new benchmark for abstract diagram understanding and visual language reasoning. In: The 35th Conference on Neural Information Processing Systems (NeurIPS) Track on Datasets and Benchmarks (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.439, + 0.787, + 0.5 + ], + "angle": 0, + "content": "[57] Mao, J., Huang, J., Toshev, A., Camburu, O.M., Yuille, A.L., Murphy, K.P.: Generation and comprehension of unambiguous object descriptions. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) pp. 11-20 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.501, + 0.787, + 0.562 + ], + "angle": 0, + "content": "[58] Marino, K., Rastegari, M., Farhadi, A., Mottaghi, R.: Ok-vqa: A visual question answering benchmark requiring external knowledge. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 3190-3199 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.228, + 0.563, + 0.787, + 0.608 + ], + "angle": 0, + "content": "[59] McMahan, B., Moore, E., Ramage, D., Hampson, S., y Arcas, B.A.: Communication-efficient learning of deep networks from decentralized data. In: Artificial intelligence and statistics. pp. 1273-1282. PMLR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.228, + 0.609, + 0.787, + 0.655 + ], + "angle": 0, + "content": "[60] Mishra, A., Shekhar, S., Singh, A.K., Chakraborty, A.: Ocr-vqa: Visual question answering by reading text in images. 2019 International Conference on Document Analysis and Recognition (ICDAR) pp. 947-952 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.228, + 0.656, + 0.645, + 0.671 + ], + "angle": 0, + "content": "[61] OpenAI: Chatgpt. https://chat.openai.com (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.228, + 0.672, + 0.724, + 0.687 + ], + "angle": 0, + "content": "[62] OpenAI: Gpt-4 technical report. ArXiv abs/2303.08774 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.228, + 0.688, + 0.787, + 0.718 + ], + "angle": 0, + "content": "[63] OpenAI: Vision - openai api. https://platform.openai.com/docs/guides/vision (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.228, + 0.719, + 0.787, + 0.779 + ], + "angle": 0, + "content": "[64] Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: DiNov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.228, + 0.78, + 0.787, + 0.841 + ], + "angle": 0, + "content": "[65] Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems 35, 27730-27744 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.227, + 0.146, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "20" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.127 + ], + "angle": 0, + "content": "Z. Lin et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.146, + 0.786, + 0.207 + ], + "angle": 0, + "content": "[66] Penedo, G., Malartic, Q., Hesslow, D., Cojocaru, R., Cappelli, A., Alobeidli, H., Pannier, B., Almazrouei, E., Launay, J.: The refined web dataset for falcon llm: outperforming curated corpora with web data, and web data only. arXiv preprint arXiv:2306.01116 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.208, + 0.786, + 0.237 + ], + "angle": 0, + "content": "[67] Peng, B., Li, C., He, P., Galley, M., Gao, J.: Instruction tuning with gpt-4. arXiv preprint arXiv:2304.03277 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.228, + 0.239, + 0.787, + 0.282 + ], + "angle": 0, + "content": "[68] Peng, Z., Wang, W., Dong, L., Hao, Y., Huang, S., Ma, S., Wei, F.: Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.228, + 0.284, + 0.786, + 0.342 + ], + "angle": 0, + "content": "[69] Plummer, B.A., Wang, L., Cervantes, C.M., Caicedo, J.C., Hockenmaier, J., Lazebnik, S.: Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. International Journal of Computer Vision 123, 74 - 93 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.228, + 0.344, + 0.786, + 0.403 + ], + "angle": 0, + "content": "[70] Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.228, + 0.405, + 0.786, + 0.433 + ], + "angle": 0, + "content": "[71] Radford, A., Narasimhan, K.: Improving language understanding by generative pre-training (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.228, + 0.435, + 0.786, + 0.479 + ], + "angle": 0, + "content": "[72] Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners. OpenAI blog 1(8), 9 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.228, + 0.48, + 0.786, + 0.508 + ], + "angle": 0, + "content": "[73] Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.228, + 0.51, + 0.786, + 0.569 + ], + "angle": 0, + "content": "[74] Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al.: Imagenet large scale visual recognition challenge. International journal of computer vision 115, 211-252 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.228, + 0.571, + 0.786, + 0.599 + ], + "angle": 0, + "content": "[75] Schuhmann, C., Köpf, A., Vencu, R., Coombes, T., Beaumont, R.: Laion-coco. https://laion.ai/blog/laion-coco/ (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.228, + 0.601, + 0.786, + 0.659 + ], + "angle": 0, + "content": "[76] Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.228, + 0.661, + 0.786, + 0.704 + ], + "angle": 0, + "content": "[77] Schwenk, D., Khandelwal, A., Clark, C., Marino, K., Mottaghi, R.: Akvqa: A benchmark for visual question answering using world knowledge. In: European Conference on Computer Vision (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.228, + 0.706, + 0.786, + 0.75 + ], + "angle": 0, + "content": "[78] Shao, W., Hu, Y., Gao, P., Lei, M., Zhang, K., Meng, F., Xu, P., Huang, S., Li, H., Qiao, Y., et al.: Tiny lvlm-ehub: Early multimodal experiments with bard. arXiv preprint arXiv:2308.03729 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.228, + 0.751, + 0.65, + 0.766 + ], + "angle": 0, + "content": "[79] ShareGPT: Sharegpt. https://sharegpt.com/ (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.228, + 0.767, + 0.786, + 0.809 + ], + "angle": 0, + "content": "[80] Shukor, M., Dancette, C., Rame, A., Cord, M.: Univalent: Unified model for image, video, audio and language tasks. Transactions on Machine Learning Research Journal (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.228, + 0.811, + 0.786, + 0.84 + ], + "angle": 0, + "content": "[81] Sidorov, O., Hu, R., Rohrbach, M., Singh, A.: Textcaps: a dataset for image captioning with reading comprehension. ArXiv abs/2003.12462 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.227, + 0.146, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.27, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "A Mixer of Weights, Visual Embeddings and Image Scales for MLLM" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "21" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.146, + 0.785, + 0.207 + ], + "angle": 0, + "content": "[82] Singh, A., Natarajan, V., Shah, M., Jiang, Y., Chen, X., Batra, D., Parikh, D., Rohrbach, M.: Towards vqa models that can read. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 8309-8318 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.209, + 0.785, + 0.239 + ], + "angle": 0, + "content": "[83] Su, Y., Lan, T., Li, H., Xu, J., Wang, Y., Cai, D.: Pandagpt: One model to instruction-follow them all. ArXiv abs/2305.16355 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.228, + 0.241, + 0.785, + 0.285 + ], + "angle": 0, + "content": "[84] Sung, Y.L., Li, L., Lin, K., Gan, Z., Bansal, M., Wang, L.: An empirical study of multimodal model merging. arXiv preprint arXiv:2304.14933 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.228, + 0.287, + 0.785, + 0.347 + ], + "angle": 0, + "content": "[85] Suvorov, R., Logacheva, E., Mashikhin, A., Remizova, A., Ashukha, A., Silvestrov, A., Kong, N., Goka, H., Park, K., Lempitsky, V.: Resolution-robust large mask inpainting with fourier convolutions. arXiv preprint arXiv:2109.07161 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.228, + 0.349, + 0.785, + 0.394 + ], + "angle": 0, + "content": "[86] Taori, R., Gulrajani, I., Zhang, T., Dubois, Y., Li, X., Guestrin, C., Liang, P., Hashimoto, T.B.: Stanford alpaca: An instruction-following llama model. https://github.com/tatsu-lab/stanford_alpaca (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.228, + 0.396, + 0.785, + 0.455 + ], + "angle": 0, + "content": "[87] Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., Lample, G.: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.228, + 0.457, + 0.785, + 0.516 + ], + "angle": 0, + "content": "[88] Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.228, + 0.518, + 0.785, + 0.563 + ], + "angle": 0, + "content": "[89] Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.228, + 0.565, + 0.785, + 0.609 + ], + "angle": 0, + "content": "[90] Wang, K., Ren, H., Zhou, A., Lu, Z., Luo, S., Shi, W., Zhang, R., Song, L., Zhan, M., Li, H.: Mathcoder: Seamless code integration in llms for enhanced mathematical reasoning. arXiv preprint arXiv:2310.03731 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.228, + 0.611, + 0.785, + 0.671 + ], + "angle": 0, + "content": "[91] Wang, P., Yang, A., Men, R., Lin, J., Bai, S., Li, Z., Ma, J., Zhou, C., Zhou, J., Yang, H.: Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In: International Conference on Machine Learning. pp. 23318-23340. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.228, + 0.673, + 0.785, + 0.733 + ], + "angle": 0, + "content": "[92] Wang, W., Chen, Z., Chen, X., Wu, J., Zhu, X., Zeng, G., Luo, P., Lu, T., Zhou, J., Qiao, Y., et al.: Visionllm: Large language model is also an open-ended decoder for vision-centric tasks. arXiv preprint arXiv:2305.11175 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.228, + 0.735, + 0.785, + 0.779 + ], + "angle": 0, + "content": "[93] Wen, S., Fang, G., Zhang, R., Gao, P., Dong, H., Metaxas, D.: Improving compositional text-to-image generation with large vision-language models. arXiv preprint arXiv:2310.06311 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.228, + 0.781, + 0.785, + 0.84 + ], + "angle": 0, + "content": "[94] Woo, S., Debnath, S., Hu, R., Chen, X., Liu, Z., Kweon, I.S., Xie, S.: Convnext v2: Co-designing and scaling convnets with masked autoencoders. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16133-16142 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.227, + 0.146, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "22" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.127 + ], + "angle": 0, + "content": "Z. Lin et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.146, + 0.787, + 0.223 + ], + "angle": 0, + "content": "[95] Wortsman, M., Ilharco, G., Gadre, S.Y., Roelofs, R., Gontijo-Lopes, R., Morcos, A.S., Namkoong, H., Farhadi, A., Carmon, Y., Kornblith, S., et al.: Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In: International Conference on Machine Learning. pp. 23965-23998. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.224, + 0.786, + 0.268 + ], + "angle": 0, + "content": "[96] Wu, C., Yin, S., Qi, W., Wang, X., Tang, Z., Duan, N.: Visual chatgpt: Talking, drawing and editing with visual foundation models. arXiv preprint arXiv:2303.04671 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.27, + 0.787, + 0.33 + ], + "angle": 0, + "content": "[97] Wu, C., Wang, T., Ge, Y., Lu, Z., Zhou, R., Shan, Y., Luo, P.: \\(\\pi\\)-tuning: Transferring multimodal foundation models with optimal multi-task interpolation. In: International Conference on Machine Learning. pp. 37713-37727. PMLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.332, + 0.787, + 0.377 + ], + "angle": 0, + "content": "[98] Xu, R., Wang, X., Wang, T., Chen, Y., Pang, J., Lin, D.: Pointllm: Empowering large language models to understand point clouds. ArXiv abs/2308.16911 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.378, + 0.787, + 0.438 + ], + "angle": 0, + "content": "[99] Yan, B., Jiang, Y., Wu, J., Wang, D., Luo, P., Yuan, Z., Lu, H.: Universal instance perception as object discovery and retrieval. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 15325-15336 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.44, + 0.786, + 0.485 + ], + "angle": 0, + "content": "[100] Yang, E., Wang, Z., Shen, L., Liu, S., Guo, G., Wang, X., Tao, D.: Adamerging: Adaptive model merging for multi-task learning. arXiv preprint arXiv:2310.02575 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.486, + 0.786, + 0.532 + ], + "angle": 0, + "content": "[101] Yang, Z., Li, L., Wang, J., Lin, K., Azarnasab, E., Ahmed, F., Liu, Z., Liu, C., Zeng, M., Wang, L.: Mm-react: Prompting chatgpt for multimodal reasoning and action. arXiv preprint arXiv:2303.11381 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.533, + 0.786, + 0.578 + ], + "angle": 0, + "content": "[102] Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.579, + 0.786, + 0.609 + ], + "angle": 0, + "content": "[103] Yu, F., Koltun, V.: Multi-scale context aggregation by dilated convolutions. arXiv preprint arXiv:1511.07122 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.61, + 0.786, + 0.655 + ], + "angle": 0, + "content": "[104] Yu, T., Feng, R., Feng, R., Liu, J., Jin, X., Zeng, W., Chen, Z.: Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.657, + 0.786, + 0.702 + ], + "angle": 0, + "content": "[105] Yu, W., Yang, Z., Li, L., Wang, J., Lin, K., Liu, Z., Wang, X., Wang, L.: Mm-vet: Evaluating large multimodal models for integrated capabilities. ArXiv abs/2308.02490 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.703, + 0.786, + 0.748 + ], + "angle": 0, + "content": "[106] Zhang, R., Han, J., Zhou, A., Hu, X., Yan, S., Lu, P., Li, H., Gao, P., Qiao, Y.: Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.749, + 0.786, + 0.794 + ], + "angle": 0, + "content": "[107] Zhang, R., Hu, X., Li, B., Huang, S., Deng, H., Li, H., Qiao, Y., Gao, P.: Prompt, generate, then cache: Cascade of foundation models makes strong few-shot learners. CVPR 2023 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.795, + 0.786, + 0.841 + ], + "angle": 0, + "content": "[108] Zhang, S., Roller, S., Goyal, N., Artetxe, M., Chen, M., Chen, S., Dewan, C., Diab, M., Li, X., Lin, X.V., et al.: Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.146, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/SPHINX_ A Mixer of Weights, Visual Embeddings and Image Scales for Multi-modal Large Language Models/cd141534-c355-432b-969a-838688ba26e3_origin.pdf b/2024/SPHINX_ A Mixer of Weights, Visual Embeddings and Image Scales for Multi-modal Large Language Models/cd141534-c355-432b-969a-838688ba26e3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6382ba0900a0e47fb5564d845fbcb7940929df46 --- /dev/null +++ b/2024/SPHINX_ A Mixer of Weights, Visual Embeddings and Image Scales for Multi-modal Large Language Models/cd141534-c355-432b-969a-838688ba26e3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:caece05a73fc679b313434340c9bedd718374434a582be893e49feb5328dc5b2 +size 3347424 diff --git a/2024/SPHINX_ A Mixer of Weights, Visual Embeddings and Image Scales for Multi-modal Large Language Models/full.md b/2024/SPHINX_ A Mixer of Weights, Visual Embeddings and Image Scales for Multi-modal Large Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..77b208f637e15606f4c2d04c2f459c19b3340580 --- /dev/null +++ b/2024/SPHINX_ A Mixer of Weights, Visual Embeddings and Image Scales for Multi-modal Large Language Models/full.md @@ -0,0 +1,396 @@ +# SPHINX: A Mixer of Weights, Visual Embeddings and Image Scales for Multi-modal Large Language Models + +Ziyi Lin $^{12*}$ , Dongyang Liu $^{2*}$ , Renrui Zhang $^{12*}$ , Peng Gao $^{2*\dagger\dagger}$ , Longtian Qiu $^{23*}$ , Han Xiao $^{2}$ , Han Qiu $^{2}$ , Wenqi Shao $^{2}$ , Keqin Chen $^{2}$ , Jiaming Han $^{12}$ , Siyuan Huang $^{2}$ , Yichi Zhang $^{2}$ , Xuming He $^{3}$ , Yu Qiao $^{2\dagger}$ , and Hongsheng Li $^{124\dagger}$ + +1 Multimedia Laboratory, The Chinese University of Hong Kong + +2 Shanghai AI Laboratory 3 ShanghaiTech University + +4 Centre for Perceptual and Interactive Intelligence Limited + +* Equal contribution † Equal advisory ‡ Project leader + +zylin@link.cuhk.edu.hk, gaopeng@pjlab.org.cn, hsli@ee.cuhk.edu.hk + +Abstract. We present SPHINX, a versatile multi-modal large language model (MLLM) with a joint mixing of model weights, visual embeddings and image scales. First, for stronger vision-language alignment, we unfreeze the large language model (LLM) during pre-training, and introduce a weight mix strategy between LLMs trained by real-world and synthetic data. By directly integrating the weights from two domains, the mixed LLM can efficiently incorporate diverse semantics with favorable robustness. Then, we propose to extract comprehensive visual embeddings from various network architectures, pre-training paradigms, and information granularity, providing language models with more robust image representations. We further propose an efficient strategy aiming to better capture fine-grained appearances of high-resolution images. With a mixing of different scales and high-resolution sub-images, SPHINX attains exceptional visual parsing and reasoning performance on existing evaluation benchmarks. Based on our proposed joint mixing, SPHINX exhibits superior multi-modal understanding capabilities on a wide range of applications, with highlighted fine-grained visual recognition abilities such as region-level understanding, caption grounding, document layout detection, and human pose estimation. We hope our work may cast a light on the exploration of joint mixing in future MLLM research. Code is released at https://github.com/Alpha-VLLM/LLaMA2-Accessory. + +# 1 Introduction + +Since the era of big data, large language models (LLMs) have attained tremendous strides (7; 61; 62; 87; 108), showcasing unprecedented application scenarios and generalization capabilities. Despite the phenomenal success, plain LLMs are still only reasoning models, and the lack of perception capability has limited their further evolution to autonomous agents with general intelligence. The mission + +Emergent Referring Pose Estimation + +![](images/dbffd25f30333c077f55064be515e1510236d599e46b160aa63ac9f463cc721b.jpg) + +Else: + - Select the keypoints of the person wearing a green suit. + +![](images/bff5837be91f9bf40bc591fea0853f6d5418f5359ce5201e9813799030586359.jpg) + +SPINDX: The detected key points are left ear[0.76,0.22], left shoulder[0.78,0.32], right shoulder[0.88,0.30], left elbow[0.74,0.49], right elbow[0.90,0.47], wrist[0.68,0.52], right wrist[0.81,0.46], left hip[0.80,0.65], right hip[0.90,0.64], knee[0.71,0.88], and right knee[0.84,0.88]. + +![](images/08c1336de65bd73b330d66b057276387e2bc846d3ae38023285ebd48b2597e58.jpg) + +Multi-level Dense Captioning + +User: Detect all objectives in this image. + +SHINE: There are 10 objects shown in the image, including a person [0.21, 0.39, 37.0, 76], a helmet[0.30,0.39,0.36,0.43], two sneakers [0.67,0.75,0.85,0.95,0.76], bicycle[0.03,0.53,0.43,0.76], a cell phone[0.22,0.40,0.33,0.42], and four polo shirt[0.50,0.59,0.59,0.59,0.48,0.44,0.48,0.43,0.43,0.16,0.59]. + +![](images/2a86f2ad90d5b69658e3d97b7061e2153230323078e6175baeb82fdd2e8e4d5d.jpg) + +Emergent Anomaly Detection + +User: Is there defect on the image? If yes detect it and plot the box on the image + +SPHINX: Yes, there is a defect on the image. The defect is a scratch on the metal object. The box on the image is $[0.450.46, 0.500.57]$ , which is the scratch on the metal object. + +![](images/e18427647bc9055a328fd86d13f8dd6acbd5e67a5a509d6c82b8bebaf215ccab.jpg) + +![](images/c2114a7b6ea442c886589f514ccb79c8d087bb187efbb3a1470bb3e487d1078c.jpg) + +![](images/b4df21ab2b24e733942d412936c1525ebe66843dfe4b9df8a389eca7072e2b88.jpg) + +Image Editing (w/ Stable Diffusion) + +SPEHIN: Detect road sign in this picture + +![](images/762006744f2e2d5a1a7b63a214c46295da302e9b16ad402b961b9b354d2c45de.jpg) + +SPHINX: Detect dog in this + +![](images/7be2eb41567a3321e3ccd14803ffa537885a1fbd4373ca7a0b43150e6229b5e1.jpg) + +SPHINX: Detect monocite in this picture. + +![](images/826bcd4a3347ca2b110f5f2e35362f10a3e0c83d4ea85f57723a2a6a48cd7211.jpg) + +Stable Diffusion: A light in the middle of the picture + +![](images/47720c29c7c4bc1c5a495a8f03e6a91cf657d56db06abfa9215db01d0e3003a2.jpg) +Fig. 1: Examples of multi-task visual reasoning on selected tasks by our proposed SPHINX models, which excels in diverse visual understanding and perception tasks, such as object detection, caption grounding, and region-level description, displaying some emergent capabilities. + +Stable Diffusion: A teddy bear sitting in a basket + +![](images/9bdf9cef7c90c7c2cb7054de5c4c616ab725176b36ce1eaaea76cc8b97957610.jpg) + +Stable Diffusion: A man riding a bicycle + +![](images/f25ee9bc17315cf3fb9aa75284d825782cde17bbbb09645737e6686c0e51f79c.jpg) + +User: Please provide a short description for this document. + +SPHINX: A blue bicycle + +User: Provide a one-sentence caption for the provided image + +SPHINK: A man riding a bike in a grassy field. + +User:Generates a detailed description about the image. + +SHANK: The image shows a man standing next to a blue mountain bike in a grassy area. He is standing on the right leg of the bike, with his right leg on the pedal. The man appears to be in the process of mounting the bike, as he is standing on the pedal and holding his foot. There are several yellow flowers scattered throughout the grassy area, adding a touch of color to the scene. This is wearing shorts, which are visible as the stands next to the bike. + +User: Please provide a short description for this region: + +SPHINK: A man in a blue shirt + +User: Provide a one-Sentence caption for the provided image + +SHINX: A man riding a bike while holding a cell phone. + +User: Generate a detailed description about the image. + +SHINK: The image shows a man sitting on a bicycle in a grassy area, and the grassy area is filled with people to be taking a selfie or recording a video while sitting on the bike. The grassy area is filled with people who he is prepared for a safe ride. There are several other bicycles in the scene, with one located to the left of the bike, and another located to the right, and a third one further back. The grassy area is scattered with people standing around, riding a couch on color to the scene. + +User: Please provide a short description of the region [0.37,4.40,97.80] + +SHINK: A cow licking a man's face + +User: Provide a one-sentence caphor for the provided image + +SPEEDON: A cow is sniffing a bicycle wheel. + +Uc:Generale a deblie du about the image. + +SPINDL: The image features a large brown and white cow standing in a grassy field. The cow is standing on the ground, looking at the edge of the field, and it appears to be sniffing a bicycle tire. Another cow is standing slightly to the right of the main cow. The scene captures the cow's perspective as it interacts with the bicycle tire. + +to integrate reasoning and perception has motivated the research on multi-modal large language models (MLLMs) (41; 50; 106; 109; 112), which make responses to not only human language, but also natural signals, among which visual signals are predominantly studied. + +Early works on MLLMs (22; 50; 102) focus on coarse-grained tasks, such as image captioning and question answering. While this enables visual instruction following or visual dialogue to some extent, fine-grained visual recognition capabilities, such as visual grounding, key-point detection and referring generation, might bring new outlook to the development of MLLMs: They make possible new applications such as precise actions in embodied AI, as well as potentially improved diversity and reduced hallucination on traditional tasks with better access to detailed visual clues. + +Besides fine-grained recognition, it is equally important for a foundational MLLM to maintain the performance on a wider domain. Despite the success of previous works such as Shikra (10) which tackles fine-grained tasks alone, it is still challenging to explore the scaling-up methods to utilize larger datasets, higher parameter counts, higher resolutions, etc. in order to equip the models with broader or even emergent abilities. + +We thus present the building of SPHINX, our proposed foundational MLLM with versatility and an emphasis on fine-grained visual recognition capabilities. Compared to previous works with similar goals, such as Qwen-VL (4) + +and MiniGPT-v2 (9), our pipeline can be highlighted with the following unique scaling-up techniques conveniently named as mixing strategies: + +- Mixed model weights. Vision-language data from particular domains might contain special semantics, e.g., synthetic captions (75) compared to real-world ones (76). Considering that directly mixing such data might confuse the MLLM, we introduce a weight-mixing strategy to efficiently combine such domain-specific knowledge. Based on the MLLM pre-trained on real-world data, we fine-tune it on the synthetic data, and then linearly combine the finetuned LLM's weights with the real-world ones. In this way, the two types of models would not be affected by contradictory data and our final SPHINX can effectively integrate knowledge from both synthetic and real-world domains, enabling effective use of datasets from diverse sources. + +- Mixed visual embeddings. To take the advantage of different encoders, we propose to mix the visual embeddings from various vision backbones (41; 64; 70) with different network architectures (CNN vs. ViT), pre-training paradigms (supervised vs. self-supervised), and information granularity (global vs. local). By mixing the different image tokens channel-wisely and sequence-wisely, SPHINX obtains stronger visual representations and leads to better vision-language alignment efficacy. + +- Mixed scales with high-resolution sub-images. To process high-resolution images as required for fine-grained visual recognition, we first spatially divide the input high-resolution image into multiple sub-images, and also downsample it into a low-resolution one. Then, we feed all the images concurrently into the mixed visual encoders, and concatenate the extracted multiple token groups to represent the entire high-resolution visual features. By mixing visual embeddings of different scales and sub-images, our SPHINX can adaptively explore more fine-grained visual semantics from the high resolution and multi-scale image representations, while maintaining encoding efficiency. + +By the proposed three-fold mixing strategy, SPHINX fine-tunes LLMs, e.g., LLaMA-2 (88), to be a powerful MLLM with superior visual instruction-following capacity. Our model excels in a variety of vision tasks, e.g., detecting different objects with remarkable precision and parsing their relations, or accurately interpreting the content within complicated figures. Importantly, SPHINX can achieve impressive fine-grained visual perception for high-resolution images, which exhibits state-of-the-art performance on extensive evaluation benchmarks, e.g., MMBench (52), MME (20), and POPE (45). + +# 2 Related works + +Large language models (LLMs). The field of Natural Language Processing (NLP) has witnessed significant progress over the years, particularly with the advent of LLMs. With Transformer (89) as the fundamental architecture, + +LLMs (61; 62; 72) have demonstrated unprecedented performance in modeling intricate language patterns over extensive contexts. Therein, BERT (15) showcases the benefits of pre-training on vast text corpora and fine-tuning on specific tasks, setting new standards on various benchmarks. OpenAI's GPT series (61; 62; 71; 72), especially GPT-3 (7), harness the power of massive model scaling, with billions and even trillions of parameters. To obtain better instruction following ability, InstructGPT (65) and ChatGPT (61) are presented to exhibit exceptional fluency and versatility in open-domain conversation tasks, ranging from text generation to question answering. Recently, the instruction tuning based on LLaMA (87) and LLaMA-2 (88) has gained great popularity as open-source LLMs in the community. Therein, Alpaca (86) and LLaMA-Adapter (106) respectively adopt full and parameter-efficient fine-tuning to acquire favorable instruction-following LLMs. Vicuna (12) and GPT-4-LLM (67) further showcase the improvement brought by higher-quality instruction datasets. Other efforts also extend LLMs for match problem solving (90; 111), visual model system (96; 101), and open-world recognition (107; 113). In this paper, we develop our SPHINX based on the superior language understanding of LLaMA-2 (88) and instruction tuning experience of LLaMA-Adapter series (22; 106), which introduce a three-fold mixer to extend the capability ceiling of instruction-following LLMs for multi-modal input. + +Multi-modal large language models (MLLMs). In addition to language instruction following, many efforts have been made to inject multi-modal conditions into LLMs for wider application scenarios. As prior attempts, VisualGPT (8) and BLIP series (14; 41; 42) indicate the potential of aligning LLMs with visual input for image captioning and question answering. Flamingo (3) and Kosmos-1 (32) further exhibit promising multi-modal understanding performance for image-text interleaved contexts. With large-scale pre-training and model sizes, GPT-4 (62) and Bard (24) both showcase remarkable proficiency in vision-language understanding and reasoning over diverse multi-modal tasks. In parallel, a bunch of works have been proposed to align LLaMA with vision modality for advanced visual instruction-following capabilities. LLaVA (50) and MiniGPT-4 (112) utilize a simple projection layer to connect vision encoders (41; 70) with LLMs. LLaMA-Adapter V2 (22) introduces zero-initialized attention mechanisms for efficient visual instruction tuning, and mPLUG-Owl (102) adopts delicately designed intermediate networks for cross-modal alignment. For more modality input, ImageBind-LLM (28) and PandaGPT (83) further incorporate audio and video conditions guided by ImageBind (23). Besides, recent MLLMs are also extended to region-level parsing (10; 68), in-context learning (37; 38), arbitrary image resolutions (6), text-to-image generation (16; 93), and 3D question answering (25; 30; 98). Different from previous works, our SPHINX aims for image-conditioned MLLM, and proposes a three-fold mixer, i.e., model weights, tuning tasks, and visual embeddings, attaining superior generalization capacity for multi-modal learning. + +Weight interpolation. Despite being commonly known to be highly nonlinear (40), neural networks have, under some conditions, exhibited surprisingly + +![](images/e50bfc0f53a05f22b325d62f9dba5b5fda3ebf8df5e46606a68345a80fa8ef69.jpg) +Fig. 2: The joint mixing paradigm of SPHINX. with mixed visual embeddings (a) and mixed model weights (b). + +good results with linearly interpolated weights in several cases, including higher-performing single-task models (11; 95), task interpolation (31; 97; 100), federated or low-communication distributed training (19; 43; 59). A few recent works (2; 80; 84) also investigated weight mixing under the MLLM context. On top of these works, we further demonstrate the effects of inter-domain weight mixing for image-c captioning pre-trained models, as well as how these benefits persist after further task-specific fine-tuning. + +High-resolution visual embeddings. Fine-grained visual recognition tasks, such as detection and segmentation, traditionally utilize higher-resolution image features to capture detailed information. Convolutional neural networks (CNNs) naturally scale to high-resolution inputs (53) as the convolution operation intrinsically supports varied input sizes, while for the recent transformer-based vision backbones, scaling input resolution is commonly achieved by interpolating the positional embeddings (18). There are also network components specifically designed to obtain high-resolution feature maps, e.g., (44; 46; 103; 110). Unlike those model-surgery-based methods, our mixed-scale solution instead explores the use of fine-tuned large language models to fuse the visual embeddings from multiple parts of the image while leaving the visual encoders themselves untouched. With such a zooming-in mechanism, our framework captures fine-grained visual clues while preserves capabilities of the visual encoders in an architecture-agnostic manner. Interestingly, the recently released GPT-4V API (63) shows a similar correspondence between the image resolution and the number of visual tokens. + +# 3 SPHINX + +In this section, we introduce a versatile MLLM, SPHINX, with an model overview in Section 3.1, the joint mixing of model weights and visual embeddings in Section 3.2 and the mixing of high-resolution sub-image tokens Section 3.3. Finally, in Section 3.4, we introduce several extended applications of SPHINX. + +# 3.1 Model overview + +The SPHINX model is composed of an LLM, e.g., LLaMA-2 (88), a mixture of vision encoders and linear projections connecting the vision encoder outputs to the language model inputs. For sequences with image input, visual tokens are extracted from the vision encoders, projected to the LM input space, and then concatenated to the beginning of the sequence. Sequences without image inputs are processed identically as in the language-only LLM. + +We adopt a two-stage training paradigm: the first pre-training stage for vision-language alignment, and the second fine-tuning stage for visual instruction-following learning. In both stages, the LM parameters are fully-tuned, together with the linear layers for visual-language feature alignment, but the visual encoders themselves are kept frozen. + +# 3.2 Pre-training: The joint mixing of visual embeddings and model weights + +The pre-training stage (or stage 1) involves two types of mixing: weight mixing and visual embedding mixing, as shown in Figure 2. Since the pre-training datasets are weakly annotated image-caption pairs where the textual information does not contain much detail, we do not use high-resolution images for this stage, thus bypassing the scale mixing. + +Mixed model weights of different domains. Vision-language data from different domains may exhibit distinct characteristics: For example, web-crawled captions like those in LAION-400M (76) tend to be diverse but noisy, but synthetic captions like those in LAION-COCO (75) tend to be coherent but of a relatively limited vocabulary. We propose a weight mixing strategy of domain-specifically tuned LLMs to integrate respective knowledge from real-world and synthetic data. We first utilize the most common domain data (LAION-400M (76)) for pre-training, which endows the MLLM with fundamental visual understanding capabilities. Then, we regard such a pre-trained model as the initial checkpoint to further fine-tune the LLM on synthetic domains, e.g., LAION-COCO (75). Finally, to take advantage of the best data domains, we directly conduct a weighted mixing of two LLMs' weights for semantic aggregation. In detail, we denote the parameters of the fundamental LLM as $\theta_{real}$ , and the fine-tuned parameters by synthetic data as $\theta_{syn}$ . The mixing process is formulated as + +$$ +\theta_ {m i x} = \beta \cdot \theta_ {r e a l} + (1 - \beta) \cdot \theta_ {s y n}, \tag {1} +$$ + +where $\beta$ denotes the mixing coefficient, and $\theta_{\text{mix}}$ represents the mixed LLM weights with aggregated semantics. Compared to fusing different domain data for joint pre-training, our weight mix strategy can encourage every MLLM to better learn domain-unique knowledge, and exhibit flexible scalability for any new data domains. + +Mixed embeddings for visual encoding. To capture robust visual representations from different aspects, we propose to ensemble a variety of vision + +![](images/00591b1905a8833dc4bbfc8b522dd169744614c00a7788e3377bd119572b6e9a.jpg) +Fig. 3: Pipeline of SPHINX for high-resolution images. We propose to further mix different scales and sub-images to better capture fine-grained semantics on high-resolution images. + +backbones for image encoding. The visual backbones with different characteristics are chosen as follows. 1) Different network architectures. As CNN (29) and ViT (18) mainly aggregate different types of visual appearances, i.e., neighboring dependencies and long-range interactions, we adopt CLIP (70) models respectively with ConvNeXt (94) and ViT image encoders. 2) Different pre-training paradigms. Supervised training can impose explicit semantic information from textual captions or category labels, while self-supervised learning enforces the model to explore implicit pretext task signals. Thus, we further employ the ViT self-supervised by DINOv2 (64) as well as the text-supervised vision encoders, CLIP. 3) Different information granularity. The aforementioned visual encoders all produce visual tokens in the patch level. To better capture global features, we also adopt Q-Former (41) to summarize visual embeddings via querying from the global context. After all the aforementioned encoding, we first channel-wisely concatenate the patch level visual tokens. The channel-wise concatenation allows for better alignment among spatially corresponding features, as well as keeps the number of tokens constant to the number of visual encoders. Then, by using two projection layers for dimension alignment, we spatial-wisely concatenate the representations between those of Q-Former and the other patch-level features. The obtained image tokens are directly placed in front of language instructions, which provide visual context for the language instructions. + +# 3.3 Fine-tuning: Mixing scales with high-resolution sub-images + +In the fine-tuning stage, or stage 2 training, the model from stage 1 will adapt for instruct following as well as high-resolution image recognition. Stage 2 introduces + +scale mixing to process high-resolution input images, which are crucial for fine-grained visual tasks and will be elaborated in this section. + +Mixed scales and high-resolution sub-images. State-of-the-art open-source MLLMs (9; 10; 22; 41; 50; 68) works adopt frozen image encoders during all training stages, in order to preserve the pre-trained visual semantics. Therefore, the image resolution of MLLMs is usually set as $224 \times 224$ , severely hindering their efficacy for fine-grained visual perception, especially region-level grounding and description. In our SPHINX, we extend the mixing of visual embeddings to more scales and high-resolution sub-images, allowing for efficient high-resolution image encoding. For an input high-resolution image, e.g., $448 \times 448$ , we construct five corresponding images of $224 \times 224$ , and feed them as independent images into our mixed vision encoders. Specifically, we first downsample the input image to $224 \times 224$ as an abstract representation, and also downsample the input image to $448 \times 448$ and crop four sub-images of $224 \times 224$ from the four corners of the $448 \times 448$ image, which preserve the detailed visual information. In this way, we enable MLLMs to not only capture fine-grained visual appearances with $224 \times 224$ positional encodings, but also achieve favorable computation efficiency. Afterwards, the five groups of image tokens are encoded and concatenated as a long sequence for feeding into LLM, where the first one group encodes global semantics, and the other four record fine-grained local features. Importantly, as the image tokens of different patches do not have interaction through the vision encoders, they are forced to interact within the LLM to obtain complete visual information. Such a strategy, in turn, motivates LLMs to parse the relations within visual conditions for better cross-modal learning. From this perspective, our SPHINX can be regarded as a new paradigm for similar to ViT (18), where the mixed vision encoders serve as a patch embedding layer, and the LLM plays the role for patch interaction as a vision decoder. On visual understanding tasks requiring higher resolutions, SPHINX achieves significant improvement with the mixed visual representations of scales and high-resolution sub-images. + +On comparison with scaling vision encoder input. The previous standard method to enlarge MLLM input resolutions is to enlarge the input of vision encoders themselves (e.g., by positional embedding interpolation for ViT) and make the vision encoder trainable (4; 9). Compared to vision encoder fine-tuning, we argue that our LLM-based patch fusion is the scalable choice when the vision encoder is large or consists of multiple models as is in our case, because (i) vision encoders can be slow to tune, as the sequence of the same length back-propagating through the vision encoders can be equally costly as backpropagating through the LM, and (ii) vision encoders can be unstable to tune, as different visual backbones may require different tuning configurations (17) given the vast difference in how they are built in the first place. As a result, we end up finding it prohibitively expensive to work out a training recipe for trainable interpolated vision encoders to achieve the same performance as our fixed-vision-encoder method. + +Efficiency. Mixing scales only incurs a $5\%$ overhead (SPHINX-1k v. SPHINX) during training because $(i)$ we limit mixed scale training to stage 2, which itself + +constitutes a smaller portion of the total training time than stage 1, and $(ii)$ high resolution training images are usually paired with longer text, further dragging down the portion of the overhead in the total training cost. We also measured the inference cost on actual hardware: On a A100 GPU with a batch size of 1 and flash decoding enabled, generating a 200-token paragraph without mixed scales takes 29GB of memory and 5.2s, and generating with mixed scales takes 33GB and 5.5s. The memory and time overhead is moderate because $(i)$ the memory is dominated by LLM parameters which are invariant to resolution and $(ii)$ processing image tokens is a prefetching process which utilizes GPU computing cores better than the decoding generation process. As visual encoders are frozen and much smaller (3B parameters combined) compared to the language model (13B parameters), we find mixing visual encoders brings negligible overhead expect for inference-time memory usage (up to 6GB, which is manageable on most modern GPUs). + +# 3.4 Extensions of SPHINX + +With the strong and general multi-modal understanding capability, SPHINX models can be integrated with other vision models to tackle even more challenging tasks, including language-referred segmentation with SAM (35), image inpainting (104) with LaMa (85) and image editing with Stable Diffusion (73). We have also explored SPHINX as supervised learners, e.g., full-data finetuning on ImageNet-1k (74). The details can be found in the supplementary materials. + +# 4 Experiments + +# 4.1 Implementation details + +We provide the important training and evaluation details in this section. More information can be found in the supplementary materials. + +Pre-training datasets. We use two image captioning datasets LAION-400M (76) and LAION-COCO (75) for multi-modal alignment. As we full-fine-tune the language model backbone for long steps, we also jointly train with a text-only dataset RefinedWeb (66) to avoid harming its text reasoning capability due to catastrophic forgetting. + +Pre-training configuration. We fine-tune the weight of the large language model and the randomly-initialized visual projections in the pre-training stage, The visual encoders are kept frozen. We only use the low-resolution view in the pre-training stage, as the web-crawled captions in the LAION datasets are relatively coarse. + +Fine-tuning datasets. In the multi-task fine-tuning phase, our objective is to equip the MLLM with the versatile needs of downstream tasks. Building upon insights from prior research (10; 14; 49; 50; 112), we include instruction following data such as LLaVA (50) and ShareGPT (79), exposing the model to tasks + +requiring explicit directives. For general Vision Question Answering (VQA), we leverage datasets like VQAv2 (1) and GQA (33). Expanding the scope to out-of-domain knowledge, we integrate datasets like OKVQA (58) and A-OKVQA (77), providing the model with information beyond the training data. Optical Character Recognition (OCR) datasets, such as OCRVQA (60) and TextCaps (81) are utilized to increase the text understanding ability of SPHINX. We introduce abundant general object detection and pose estimation datasets, such as COCO (47) and LVIS (26) to inspire the model's capabilities of localization, classification, and human pose estimation. To address grounding tasks, we incorporate RefCOCO (34) and VG (36) datasets, training the model to handle referring object localization. Additionally, Grounding Caption datasets, such as those from Flickr30k (69), further refine the understanding of descriptions in the context of image regions. We streamline the training by converting all datasets into a multi-turn conversation format. + +Fine-tuning configuration. The trained and frozen network components are identical as the pre-training stage. Training data are sampled from the mixture of datasets following their natural frequencies, i.e., the chance of a dataset being sampled from is proportional to its original size. + +Model variant definition. We mainly use 3 model variants for experiments: SPHINX, SPHINX-1k and SPHINX-2k, which take increasingly higher resolution images as inputs. The relevant details are listed in the supplementary materials. All models use LLaMA-2-13B (88) as the language model backbone unless otherwise specified, resulting in a total parameter count of around 16B with the 4 visual encoders. + +Evaluation prompt design. In our model evaluation, we prioritize aligning with each benchmark's desired output format. To achieve this, we employ distinct prompts tailored to benchmarks that necessitate long answers, short answers, and multiple-choice responses. The detailed information is provided in the supplementary materials. This approach ensures that our model is capable of handling diverse scenarios. + +# 4.2 Model evaluation + +We provide a comprehensive evaluation of SPHINX and showcase results across multiple benchmarks. Our evaluation encompasses both quantitative metrics (in this section) and qualitative assessments (see the supplementary materials, providing a holistic understanding of our VLM model's performance. + +Benchmarks on multi-modal large language models. We test our model on recently proposed MLLM benchmarks to comprehensively evaluation of the model's characteristic. As shown in Table 1, SPHINX obtains new state-of-the-art performance on 7 out of 11 MLLM benchmarks. Furthermore, the gap between SPHINX and SPHINX-1k on POPE suggests that the introduction of high-resolution sub-images significantly alleviates the visual hallucination problems. + +Visual question answering. We also evaluate our models on various classical visual question answering (VQA) benchmarks in Table 2. SPHINX achieves + +
MethodsPOPE (45)\(MME^P\) (21)\(MMEC^C\) (21)MMB (52)SEED (39)\(LLawa^W\) (50)MM-Vet (105)CCbench (13)MathVista (54)TinyLVLM (78)Touchstone (5)
BLIP-2 (41)85.31293.8--46.438.122.4--284.7-
InstructBLIP-7B (14)---36.053.460.926.212.125.3300.6552.4
InstructBLIP-13B (14)78.91212.8---58.225.6----
Shikra (10)---58.8-------
LLaMA-AdapterV2 (22)-1328.4356.4------229.2590.1
Qwen-VL-7B (4)---38.256.3--5.5---
Qwen-VL-7B-Chat (4)-1487.6360.760.658.2--39.3-316.8645.2
LLaVA1.5-7B (49)85.91510.7-64.358.663.430.516.4---
LLaVA1.5-13B (49)85.91531.3295.467.761.670.735.426.5---
SPHINX80.71476.1322.266.969.173.536.025.627.0-632.4
SPHINX-1k90.81560.2310.067.171.674.336.627.927.5288.9645.0
SPHINX-2k87.21470.6326.865.971.676.940.227.427.8-659.5
+ +Table 1: Comparison with state-of-the-art methods on MLLM benchmarks. + +
MethodsGeneral VQAText-Oriented VQA
OK-VQA (58)VQAv2 (1)VizWiz (27)GQA (33)VSR (48)ScienceQA (55)IconVQA (56)TextVQA (82)OCR-VQA (60)
BLIP-2 (41)45.9-19.641.050.9-40.6-40.6
InstructBLIP (14)--33.449.552.1-44.8-44.8
LLaMA-Adapter-v2 (22)49.670.739.845.1---37.4-
Shikra (10)47.277.4-------
Fuyu-8B (6)60.674.2-------
MiniGPT-v2 (9)57.8-53.660.162.9-51.5--
Qwen-VL-7B (4)58.679.535.259.363.867.1-63.875.7
Qwen-VL-7B-Chat (4)56.678.238.957.561.568.2-61.570.5
LLaVA1.5-7B (49)-78.550.062.0-66.8-58.2-
LLaVA1.5-13B (49)-80.053.663.3-71.6-61.3-
SPHINX62.178.139.962.658.569.350.451.666.0
SPHINX-1k62.280.246.862.965.469.152.758.870.0
SPHINX-2k62.680.744.963.157.170.650.561.267.8
+ +Table 2: Comparison with state-of-the-art methods on classical VQA benchmarks. + +comparative results across all benchmarks. Notably, SPHINX-1k and SPHINX-2k significantly outperform SPHINX on VQAv2 and text-oriented VQA which demand fine-grained visual information, showcasing the effectiveness of our scale mix-up approach for achieving high resolution without relying on a visual encoder trained specifically on high-resolution images. Although the performances of SPHINX on text-oriented VQA surpass strong baselines such as BLIP-2 and InstructBLIP, it falls behind Qwen-VL-7B, which we conjecture is due to the lack of text-related pre-training data. Introducing more such datasets for training may resolve the issue in the future. + +Visual grounding. Table 3 evaluates SPHINX on REferring Comprehension (REC) benchmarks with RefCOCO (34), RefCOCO+ (57), and RefCOCOg (57) datasets. SPHINX outperforms most state-of-the-art models, including specialist model G-DINO-L (51) and other visual-language generalist models. Compared to a recent strong baseline Qwen-VL-7B (4), which also leverages the large language model for visual understanding, our model still achieves better results across all splits by a large margin. Moreover, SPHINX-1k and SPHINX-2k + +
MethodsRefCOCO+RefCOCORefCOCOg
valtest-Atest-Bvaltest-Atest-Bval-utest-u
Specialist models
UNINEXT (99)85.2489.6379.7992.6494.3391.4688.7389.37
G-DINO-L (51)82.7588.9575.9290.5693.1988.2486.1387.02
Generalist models
VisionLLM-H (92)----86.70---
OFA-L (91)68.2976.0061.7579.9683.6776.3967.5767.58
Shikra 7B (10)81.6087.3672.1287.0190.6180.2482.2782.19
Shikra 13B (10)82.8987.7974.4187.8391.1181.8182.6483.16
MiniGPT-v2 7B (9)79.9785.1274.4588.6991.6585.3384.4484.66
MiniGPT-v2-7B-Chat (9)79.5885.5273.3288.0691.2984.3084.1984.31
Qwen-VL-7B (4)83.1288.2577.2189.3692.2685.3485.5885.48
Qwen-VL-7B-Chat (4)82.8288.5976.7988.5592.2784.5185.9686.32
SPHINX82.7787.2976.8589.1591.3785.1384.8783.65
SPHINX-1k86.6491.0880.3591.0592.6586.5688.1988.35
SPHINX-2k85.5190.6280.4591.1092.8887.0788.0788.65
+ +Table 3: Performance comparisons (Top-1 Accuracy@0.5) on the referring expression comprehension task. + +
patch feat.global feat.MMEGeneral VQAText-oriented VQARefCOCO+RefCOCORefCOCOg
Conv NeXtCLIP LargeDINO v2Q-FormerPCOK VQAViz WizGQAVSRText VQAOCR VQAvaltest-Atest-Bvaltest-Atest-Bvaltest
✓✓✓1392.9299.260.540.359.853.050.461.780.5286.2075.4187.8590.7383.8282.7783.10
✓✓1428.4287.160.347.959.151.649.561.674.7982.3067.4782.1787.8576.7076.4876.20
✓✓✓1246.7267.560.946.160.055.251.365.982.1486.5575.3788.7390.6983.7084.1283.22
✓✓✓1371.7298.959.740.259.751.950.667.079.6185.8173.2386.0289.8080.6581.0081.58
✓✓✓1379.2274.660.547.759.952.251.666.781.0886.5175.0087.7990.9482.5182.1182.14
✓1364.0263.960.740.158.853.450.464.837.2242.4532.0746.5250.2442.6839.8840.79
✓✓✓✓1476.1322.262.139.962.658.551.666.082.7787.2876.8589.1491.3785.1284.8683.64
+ +Table 4: SPHINX visual embedding ablation study. The best result in each column is marked bold and the significantly worse results are marked red. + +enable the use of high-resolution input images, leading to consecutive improvement over SPHINX and narrowing down the gap to the strong specialist model UNINEXT, which adopts a larger input image size. These results demonstrate the competitive capability of SPHINX for visual grounding. + +# 4.3 Ablation study + +On the influence of image-text joint pre-training. We show in Figure 4 the effects of jointly training with a text-only dataset (i.e., RefinedWeb) in addition to the image-captioning dataset (i.e., LAION-400M). For comparison, we run an experiment where only LAION-400M is used for training, and observe the text generation loss and image captioning loss as a measurement of its text and image modeling capability, respectively. The text generation loss gradually grows if the model is not trained with RefinedWeb, showing that our joint-training scheme + +
viewsMMEGeneral VQAText-oriented VQARefCOCO+RefCOCORefCOCOg
low. sub-res. imagesPCOK VQAViz WizGQAVSRText VQAOCR VQAvaltest-Atest-Bvaltest-Atest-Bvaltest
✓1476.1322.262.139.962.658.551.666.082.7787.2876.8589.1491.3785.1284.8683.64
✓1461.9255.361.846.162.054.057.569.169.8677.4562.7871.7277.0964.7469.1570.89
✓✓1560.2310.062.246.862.965.458.870.086.6491.0880.3591.0592.6586.5688.1988.35
+ +Table 5: Ablation study on mixed scales. For comparison, we remove either the low-resolution view or the sub-image views from SPHINX-1k (note that SPHINX-1k without sub-images is equivalent to SPHINX). The best results are marked bold and the second-best results are marked with underlines. + +is effective in preserving the text reasoning capability while adapting to cross-modal understanding. + +On the influence of image-text joint pre-training. We show in Figure 4 the effects of jointly training with a text-only dataset (i.e., RefinedWeb) in addition to the image captioning dataset (i.e., LAION-400M). For comparison, we run an experiment where the model is trained only with LAION-400M and without RefinedWeb, and observe the text generation loss and image captioning loss as a measurement of its text and image modeling capability, respectively. The text generation loss gradually grows if the model is not trained with RefinedWeb, showing that our joint-training scheme is effective in preserving the text reasoning capability while adapting to cross-modal understanding. + +On the influence of mixed visual embeddings. To validate the effectiveness of visual embedding mix-up, we conducted experiments on SPHINX employing different combinations of visual encoders. The results are presented in Table 4. Interestingly, mixing the encoders exhibits different behaviors on different types of tasks: The mixture of all image encoders do not steadily outperforms its subsets on the VQA benchmarks, which we conjecture is because some datasets are either saturated or out-of-domain for some encoders, while we have noticed a huge gain on visual grounding tasks with the mixture of encoders compared to using only the global features (Q-Former in our case), certifying the importance of spatially-structured features for such fine-grained tasks. We also observed that DINOv2, which is a self-supervised encoder and complementary to the remaining image-text-based encoders, tend to give better results on the visual grounding tsks. As adding visual encoders brings negligible overhead, as analyzed in Sec. 3.3, we choose to use all the encoders by default as they obtain the best performance on most tests. + +On the influence of mixed scales. To give some insights into our mixed-scale strategy, we perform multi-task fine-tuning with SPHINX-1k, but remove either the low-resolution view or the sub-images to see their respective effects on various benchmarks. The results in Table 5 demonstrate the importance of both types of views: The combination of the two obtains the best results on most benchmarks. Our observation also highlights the critical role of the low-resolution view: Using only the sub-images, the model under-performs the SPHINX baseline on MME and referring comprehension despite the much longer input se + +quence. This emphasizes the importance of the low-resolution view as a global representation which helps glue together the detailed information captured in the sub-images. + +![](images/ff3eb064e3f1f88ecc3a2d26faca562b386a172f29729c3af84f83dadc405c6d.jpg) +Fig. 4: Loss curve in the pre-training stage with and without optimizing on RefinedWeb. As we perform 1-epoch training on massive datasets, the curve can serve as both training and validation losses, as no sample is seen twice in the process. + +# 5 Conclusion + +In this paper, we propose SPHINX, a versatile multi-modal large language model (MLLM) with multi-purpose visual instruction-following capabilities. In our MLLM, we introduce a joint mixing of three different aspects: model weights of pre-trained LLMs by real-world and synthetic data, visual embeddings from different types of vision backbones, and multi-scale views of images with subimage embeddings. Via our proposed three-fold mixing strategy, SPHINX achieves impressive performance over a wide range of multi-modality evaluation benchmarks, and can serve as a strong vision generalist to tackle object detection, region-level captioning, and human pose estimation, etc. Our MLLM can also be integrated with other visual foundation models for wider functionalities, e.g., SAM (35) for language-referred segmentation and Stable Diffusion (73) for image editing. + +Limitations. Limited by the context length of LLaMA, it is currently difficult for our model to process multiple images, interleaved images, videos or very large input images (e.g., beyond $1000 \times 1000$ ). However, we believe this can be resolved in the near future with long-context LLMs under active research. + +Potential societal impacts. Our method does not pose new negative societal impacts other than those existing in the pre-train models and training datasets themselves to our knowledge. With all base models and datasets used in the paper being publicly available, necessary censorship can be made based on specific ethnicity, regulations, etc. Our method has been shown to reduce hallucination, which could be generally considered a positive societal impacts. + +Acknowledgements. This project is funded in part by National Key R&D Program of China Project 2022ZD0161100 and 2022ZD0160102, by National Natural Science Foundation of China (Grant No. 62206272), by the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK, by Smart Traffic Fund PSRI/76/2311/PR and by RGC General Research Fund Project 14204021. Hongsheng Li is a PI of CPII under the InnoHK. + +# Bibliography + +[1] Agrawal, A., Lu, J., Antol, S., Mitchell, M., Zitnick, C.L., Parikh, D., Ba-tra, D.: Vqa: Visual question answering. International Journal of Computer Vision 123, 4 - 31 (2015) +[2] Aiello, E., Yu, L., Nie, Y., Aghajanyan, A., Oguz, B.: Jointly training large autoregressive multimodal models. arXiv preprint arXiv:2309.15564 (2023) +[3] Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems 35, 23716-23736 (2022) +[4] Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-vl: A frontier large vision-language model with versatile abilities. ArXiv abs/2308.12966 (2023) +[5] Bai, S., Yang, S., Bai, J., Wang, P., Zhang, X., Lin, J., Wang, X., Zhou, C., Zhou, J.: Touchstone: Evaluating vision-language models by language models (2023) +[6] Bavishi, R., Elsen, E., Hawthorne, C., Nye, M., Odena, A., Somani, A., Taşirlar, S.: Introducing our multimodal models (2023), https://www. adept.ai/blog/fuyu-8b +[7] Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020) +[8] Chen, J., Guo, H., Yi, K., Li, B., Elhoseiny, M.: Visualgpt: Data-efficient adaptation of pretrained language models for image captioning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18030-18040 (2022) +[9] Chen, J., Li, D.Z.X.S.X., Zhang, Z.L.P., Xiong, R.K.V.C.Y., Elhoseiny, M.: Minigpt-v2: Large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478 (2023) + +[10] Chen, K., Zhang, Z., Zeng, W., Zhang, R., Zhu, F., Zhao, R.: Shikra: Unleashing multimodal lvm's referential dialogue magic. arXiv preprint arXiv:2306.15195 (2023) +[11] Chen, X., Wang, X., Changpinyo, S., Piergiovanni, A., Padlewski, P., Salz, D., Goodman, S., Grycner, A., Mustafa, B., Beyer, L., et al.: Pali: A jointly-scaled multilingual language-image model. arXiv preprint arXiv:2209.06794 (2022) +[12] Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., Stoica, I., Xing, E.P.: Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ chatgpt quality. https://lmsys.org/blog/2023-03-30-vicuna/ (March 2023) +[13] Contributors, O.: Opencompass: A universal evaluation platform for foundation models. https://github.com/open-compass/opencompass (2023) +[14] Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B.A., Fung, P., Hoi, S.C.H.: Instructclip: Towards general-purpose vision-language models with instruction tuning. ArXiv abs/2305.06500 (2023) +[15] Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018) +[16] Dong, R., Han, C., Peng, Y., Qi, Z., Ge, Z., Yang, J., Zhao, L., Sun, J., Zhou, H., Wei, H., et al.: Dreamllm: Synergistic multimodal comprehension and creation. arXiv preprint arXiv:2309.11499 (2023) +[17] Dong, X., Bao, J., Zhang, T., Chen, D., Gu, S., Zhang, W., Yuan, L., Chen, D., Wen, F., Yu, N.: Clip itself is a strong fine-tuner: Achieving 85.7% and 88.0% top-1 accuracy with vit-b and vit-l onImagenet. arXiv preprint arXiv:2212.06138 (2022) +[18] Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) +[19] Douillard, A., Feng, Q., Rusu, A.A., Chhaparia, R., Donchev, Y., Kuncoro, A., Ranzato, M., Szlam, A., Shen, J.: Diloco: Distributed low-communication training of language models. arXiv preprint arXiv:2311.08105 (2023) +[20] Fu, C., Chen, P., Shen, Y., Qin, Y., Zhang, M., Lin, X., Qiu, Z., Lin, W., Yang, J., Zheng, X., et al.: Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394 (2023) +[21] Fu, C., Chen, P., Shen, Y., Qin, Y., Zhang, M., Lin, X., Qiu, Z., Lin, W., Yang, J., Zheng, X., et al.: Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394 (2023) +[22] Gao, P., Han, J., Zhang, R., Lin, Z., Geng, S., Zhou, A., Zhang, W., Lu, P., He, C., Yue, X., et al.: Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010 (2023) + +[23] Girdhar, R., El-Nouby, A., Liu, Z., Singh, M., Alwala, K.V., Joulin, A., Misra, I.: Imagebind one embedding space to bind them all. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 15180-15190 (2023) +[24] Google: Bard. https://bard.google.com/ (2023) +[25] Guo, Z., Zhang, R., Zhu, X., Tang, Y., Ma, X., Han, J., Chen, K., Gao, P., Li, X., Li, H., Heng, P.A.: Point-bind & point-llm: Aligning point cloud with multi-modality for 3d understanding, generation, and instruction following. ArXiv abs/2309.00615 (2023) +[26] Gupta, A., Dollar, P., Girshick, R.: Lvis: A dataset for large vocabulary instance segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5356-5364 (2019) +[27] Gurari, D., Li, Q., Stangl, A., Guo, A., Lin, C., Grauman, K., Luo, J., Bigham, J.P.: Vizwiz grand challenge: Answering visual questions from blind people. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition pp. 3608-3617 (2018) +[28] Han, J., Zhang, R., Shao, W., Gao, P., Xu, P., Xiao, H., Zhang, K., Liu, C., Wen, S., Guo, Z., et al.: Imagebind-llm: Multi-modality instruction tuning. arXiv preprint arXiv:2309.03905 (2023) +[29] He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) +[30] Hong, Y., Zhen, H., Chen, P., Zheng, S., Du, Y., Chen, Z., Gan, C.: 3d-dllm: Injecting the 3d world into large language models. arXiv preprint arXiv:2307.12981 (2023) +[31] Huang, C., Liu, Q., Lin, B.Y., Pang, T., Du, C., Lin, M.: Lorahub: Efficient cross-task generalization via dynamic lora composition. arXiv preprint arXiv:2307.13269 (2023) +[32] Huang, S., Dong, L., Wang, W., Hao, Y., Singhal, S., Ma, S., Lv, T., Cui, L., Mohammed, O.K., Liu, Q., et al.: Language is not all you need: Aligning perception with language models. arXiv preprint arXiv:2302.14045 (2023) +[33] Hudson, D.A., Manning, C.D.: Gqa: A new dataset for real-world visual reasoning and compositional question answering. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 6693-6702 (2019) +[34] Kazemzadeh, S., Ordonez, V., andre Matten, M., Berg, T.L.: Referitgame: Referring to objects in photographs of natural scenes. In: Conference on Empirical Methods in Natural Language Processing (2014) +[35] Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023) +[36] Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., et al.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision 123, 32-73 (2017) + +[37] Li, B., Zhang, Y., Chen, L., Wang, J., Pu, F., Yang, J., Li, C., Liu, Z.: Mimic-it: Multi-modal in-context instruction tuning. ArXiv abs/2306.05425 (2023) +[38] Li, B., Zhang, Y., Chen, L., Wang, J., Yang, J., Liu, Z.: Otter: A multimodal model with in-context instruction tuning. ArXiv abs/2305.03726 (2023) +[39] Li, B., Wang, R., Wang, G., Ge, Y., Ge, Y., Shan, Y.: Seed-bench: Benchmarking multimodal llms with generative comprehension. ArXiv abs/2307.16125 (2023) +[40] Li, H., Xu, Z., Taylor, G., Studer, C., Goldstein, T.: Visualizing the loss landscape of neural nets. Advances in neural information processing systems 31 (2018) +[41] Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023) +[42] Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022) +[43] Li, M., Gururangan, S., Dettmers, T., Lewis, M., Althoff, T., Smith, N.A., Zettlemoyer, L.: Branch-train-merge: Embarrassingly parallel training of expert language models. arXiv preprint arXiv:2208.03306 (2022) +[44] Li, Y., Mao, H., Girshick, R., He, K.: Exploring plain vision transformer backbones for object detection. In: European Conference on Computer Vision. pp. 280-296. Springer (2022) +[45] Li, Y., Du, Y., Zhou, K., Wang, J., Zhao, W.X., Wen, J.R.: Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355 (2023) +[46] Lin, T.Y., Dollar, P., Girshick, R., He, K., Hariharan, B., Belongie, S.: Feature pyramid networks for object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2117-2125 (2017) +[47] Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dolar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014) +[48] Liu, F., Emerson, G.E.T., Collier, N.: Visual spatial reasoning. Transactions of the Association for Computational Linguistics (2023) +[49] Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. ArXiv abs/2310.03744 (2023) +[50] Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv preprint arXiv:2304.08485 (2023) +[51] Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., yue Li, C., Yang, J., Su, H., Zhu, J.J., Zhang, L.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. ArXiv abs/2303.05499 (2023) + +[52] Liu, Y., Duan, H., Zhang, Y., Li, B., Zhang, S., Zhao, W., Yuan, Y., Wang, J., He, C., Liu, Z., et al.: Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281 (2023) +[53] Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3431-3440 (2015) +[54] Lu, P., Bansal, H., Xia, T., Liu, J., yue Li, C., Hajishirzi, H., Cheng, H., Chang, K.W., Galley, M., Gao, J.: Mathvista: Evaluating math reasoning in visual contexts with gpt-4v, bard, and other large multimodal models. ArXiv abs/2310.02255 (2023) +[55] Lu, P., Mishra, S., Xia, T., Qiu, L., Chang, K.W., Zhu, S.C., Tafjord, O., Clark, P., Kalyan, A.: Learn to explain: Multimodal reasoning via thought chains for science question answering. In: The 36th Conference on Neural Information Processing Systems (NeurIPS) (2022) +[56] Lu, P., Qiu, L., Chen, J., Xia, T., Zhao, Y., Zhang, W., Yu, Z., Liang, X., Zhu, S.C.: Iconqa: A new benchmark for abstract diagram understanding and visual language reasoning. In: The 35th Conference on Neural Information Processing Systems (NeurIPS) Track on Datasets and Benchmarks (2021) +[57] Mao, J., Huang, J., Toshev, A., Camburu, O.M., Yuille, A.L., Murphy, K.P.: Generation and comprehension of unambiguous object descriptions. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) pp. 11-20 (2015) +[58] Marino, K., Rastegari, M., Farhadi, A., Mottaghi, R.: Ok-vqa: A visual question answering benchmark requiring external knowledge. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 3190-3199 (2019) +[59] McMahan, B., Moore, E., Ramage, D., Hampson, S., y Arcas, B.A.: Communication-efficient learning of deep networks from decentralized data. In: Artificial intelligence and statistics. pp. 1273-1282. PMLR (2017) +[60] Mishra, A., Shekhar, S., Singh, A.K., Chakraborty, A.: Ocr-vqa: Visual question answering by reading text in images. 2019 International Conference on Document Analysis and Recognition (ICDAR) pp. 947-952 (2019) +[61] OpenAI: Chatgpt. https://chat.openai.com (2023) +[62] OpenAI: Gpt-4 technical report. ArXiv abs/2303.08774 (2023) +[63] OpenAI: Vision - openai api. https://platform.openai.com/docs/guides/vision (2023) +[64] Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: DiNov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023) +[65] Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems 35, 27730-27744 (2022) + +[66] Penedo, G., Malartic, Q., Hesslow, D., Cojocaru, R., Cappelli, A., Alobeidli, H., Pannier, B., Almazrouei, E., Launay, J.: The refined web dataset for falcon llm: outperforming curated corpora with web data, and web data only. arXiv preprint arXiv:2306.01116 (2023) +[67] Peng, B., Li, C., He, P., Galley, M., Gao, J.: Instruction tuning with gpt-4. arXiv preprint arXiv:2304.03277 (2023) +[68] Peng, Z., Wang, W., Dong, L., Hao, Y., Huang, S., Ma, S., Wei, F.: Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824 (2023) +[69] Plummer, B.A., Wang, L., Cervantes, C.M., Caicedo, J.C., Hockenmaier, J., Lazebnik, S.: Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. International Journal of Computer Vision 123, 74 - 93 (2015) +[70] Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) +[71] Radford, A., Narasimhan, K.: Improving language understanding by generative pre-training (2018) +[72] Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners. OpenAI blog 1(8), 9 (2019) +[73] Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models (2021) +[74] Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al.: Imagenet large scale visual recognition challenge. International journal of computer vision 115, 211-252 (2015) +[75] Schuhmann, C., Köpf, A., Vencu, R., Coombes, T., Beaumont, R.: Laion-coco. https://laion.ai/blog/laion-coco/ (2022) +[76] Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114 (2021) +[77] Schwenk, D., Khandelwal, A., Clark, C., Marino, K., Mottaghi, R.: Akvqa: A benchmark for visual question answering using world knowledge. In: European Conference on Computer Vision (2022) +[78] Shao, W., Hu, Y., Gao, P., Lei, M., Zhang, K., Meng, F., Xu, P., Huang, S., Li, H., Qiao, Y., et al.: Tiny lvlm-ehub: Early multimodal experiments with bard. arXiv preprint arXiv:2308.03729 (2023) +[79] ShareGPT: Sharegpt. https://sharegpt.com/ (2023) +[80] Shukor, M., Dancette, C., Rame, A., Cord, M.: Univalent: Unified model for image, video, audio and language tasks. Transactions on Machine Learning Research Journal (2023) +[81] Sidorov, O., Hu, R., Rohrbach, M., Singh, A.: Textcaps: a dataset for image captioning with reading comprehension. ArXiv abs/2003.12462 (2020) + +[82] Singh, A., Natarajan, V., Shah, M., Jiang, Y., Chen, X., Batra, D., Parikh, D., Rohrbach, M.: Towards vqa models that can read. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 8309-8318 (2019) +[83] Su, Y., Lan, T., Li, H., Xu, J., Wang, Y., Cai, D.: Pandagpt: One model to instruction-follow them all. ArXiv abs/2305.16355 (2023) +[84] Sung, Y.L., Li, L., Lin, K., Gan, Z., Bansal, M., Wang, L.: An empirical study of multimodal model merging. arXiv preprint arXiv:2304.14933 (2023) +[85] Suvorov, R., Logacheva, E., Mashikhin, A., Remizova, A., Ashukha, A., Silvestrov, A., Kong, N., Goka, H., Park, K., Lempitsky, V.: Resolution-robust large mask inpainting with fourier convolutions. arXiv preprint arXiv:2109.07161 (2021) +[86] Taori, R., Gulrajani, I., Zhang, T., Dubois, Y., Li, X., Guestrin, C., Liang, P., Hashimoto, T.B.: Stanford alpaca: An instruction-following llama model. https://github.com/tatsu-lab/stanford_alpaca (2023) +[87] Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., Lample, G.: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023) +[88] Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023) +[89] Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017) +[90] Wang, K., Ren, H., Zhou, A., Lu, Z., Luo, S., Shi, W., Zhang, R., Song, L., Zhan, M., Li, H.: Mathcoder: Seamless code integration in llms for enhanced mathematical reasoning. arXiv preprint arXiv:2310.03731 (2023) +[91] Wang, P., Yang, A., Men, R., Lin, J., Bai, S., Li, Z., Ma, J., Zhou, C., Zhou, J., Yang, H.: Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In: International Conference on Machine Learning. pp. 23318-23340. PMLR (2022) +[92] Wang, W., Chen, Z., Chen, X., Wu, J., Zhu, X., Zeng, G., Luo, P., Lu, T., Zhou, J., Qiao, Y., et al.: Visionllm: Large language model is also an open-ended decoder for vision-centric tasks. arXiv preprint arXiv:2305.11175 (2023) +[93] Wen, S., Fang, G., Zhang, R., Gao, P., Dong, H., Metaxas, D.: Improving compositional text-to-image generation with large vision-language models. arXiv preprint arXiv:2310.06311 (2023) +[94] Woo, S., Debnath, S., Hu, R., Chen, X., Liu, Z., Kweon, I.S., Xie, S.: Convnext v2: Co-designing and scaling convnets with masked autoencoders. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16133-16142 (2023) + +[95] Wortsman, M., Ilharco, G., Gadre, S.Y., Roelofs, R., Gontijo-Lopes, R., Morcos, A.S., Namkoong, H., Farhadi, A., Carmon, Y., Kornblith, S., et al.: Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In: International Conference on Machine Learning. pp. 23965-23998. PMLR (2022) +[96] Wu, C., Yin, S., Qi, W., Wang, X., Tang, Z., Duan, N.: Visual chatgpt: Talking, drawing and editing with visual foundation models. arXiv preprint arXiv:2303.04671 (2023) +[97] Wu, C., Wang, T., Ge, Y., Lu, Z., Zhou, R., Shan, Y., Luo, P.: $\pi$ -tuning: Transferring multimodal foundation models with optimal multi-task interpolation. In: International Conference on Machine Learning. pp. 37713-37727. PMLR (2023) +[98] Xu, R., Wang, X., Wang, T., Chen, Y., Pang, J., Lin, D.: Pointllm: Empowering large language models to understand point clouds. ArXiv abs/2308.16911 (2023) +[99] Yan, B., Jiang, Y., Wu, J., Wang, D., Luo, P., Yuan, Z., Lu, H.: Universal instance perception as object discovery and retrieval. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 15325-15336 (2023) +[100] Yang, E., Wang, Z., Shen, L., Liu, S., Guo, G., Wang, X., Tao, D.: Adamerging: Adaptive model merging for multi-task learning. arXiv preprint arXiv:2310.02575 (2023) +[101] Yang, Z., Li, L., Wang, J., Lin, K., Azarnasab, E., Ahmed, F., Liu, Z., Liu, C., Zeng, M., Wang, L.: Mm-react: Prompting chatgpt for multimodal reasoning and action. arXiv preprint arXiv:2303.11381 (2023) +[102] Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178 (2023) +[103] Yu, F., Koltun, V.: Multi-scale context aggregation by dilated convolutions. arXiv preprint arXiv:1511.07122 (2015) +[104] Yu, T., Feng, R., Feng, R., Liu, J., Jin, X., Zeng, W., Chen, Z.: Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790 (2023) +[105] Yu, W., Yang, Z., Li, L., Wang, J., Lin, K., Liu, Z., Wang, X., Wang, L.: Mm-vet: Evaluating large multimodal models for integrated capabilities. ArXiv abs/2308.02490 (2023) +[106] Zhang, R., Han, J., Zhou, A., Hu, X., Yan, S., Lu, P., Li, H., Gao, P., Qiao, Y.: Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199 (2023) +[107] Zhang, R., Hu, X., Li, B., Huang, S., Deng, H., Li, H., Qiao, Y., Gao, P.: Prompt, generate, then cache: Cascade of foundation models makes strong few-shot learners. CVPR 2023 (2023) +[108] Zhang, S., Roller, S., Goyal, N., Artetxe, M., Chen, M., Chen, S., Dewan, C., Diab, M., Li, X., Lin, X.V., et al.: Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068 (2022) \ No newline at end of file diff --git a/2024/SPHINX_ A Mixer of Weights, Visual Embeddings and Image Scales for Multi-modal Large Language Models/images.zip b/2024/SPHINX_ A Mixer of Weights, Visual Embeddings and Image Scales for Multi-modal Large Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4e173ec4303ca4cdd4bdbe756a10a1fa987202d7 --- /dev/null +++ b/2024/SPHINX_ A Mixer of Weights, Visual Embeddings and Image Scales for Multi-modal Large Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:956d2f8d8f6eee32690cdd66ca2d20a161ecd59a830b06ae441363a97a1fbd87 +size 451848 diff --git a/2024/SPHINX_ A Mixer of Weights, Visual Embeddings and Image Scales for Multi-modal Large Language Models/layout.json b/2024/SPHINX_ A Mixer of Weights, Visual Embeddings and Image Scales for Multi-modal Large Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e1a4b124b9493ae6287b0fee06d8084fe97fcb7f --- /dev/null +++ b/2024/SPHINX_ A Mixer of Weights, Visual Embeddings and Image Scales for Multi-modal Large Language Models/layout.json @@ -0,0 +1,10863 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 143, + 111, + 471, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 111, + 471, + 166 + ], + "spans": [ + { + "bbox": [ + 143, + 111, + 471, + 166 + ], + "type": "text", + "content": "SPHINX: A Mixer of Weights, Visual Embeddings and Image Scales for Multi-modal Large Language Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 132, + 185, + 481, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 185, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 185, + 481, + 224 + ], + "type": "text", + "content": "Ziyi Lin" + }, + { + "bbox": [ + 132, + 185, + 481, + 224 + ], + "type": "inline_equation", + "content": "^{12*}" + }, + { + "bbox": [ + 132, + 185, + 481, + 224 + ], + "type": "text", + "content": ", Dongyang Liu" + }, + { + "bbox": [ + 132, + 185, + 481, + 224 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 132, + 185, + 481, + 224 + ], + "type": "text", + "content": ", Renrui Zhang" + }, + { + "bbox": [ + 132, + 185, + 481, + 224 + ], + "type": "inline_equation", + "content": "^{12*}" + }, + { + "bbox": [ + 132, + 185, + 481, + 224 + ], + "type": "text", + "content": ", Peng Gao" + }, + { + "bbox": [ + 132, + 185, + 481, + 224 + ], + "type": "inline_equation", + "content": "^{2*\\dagger\\dagger}" + }, + { + "bbox": [ + 132, + 185, + 481, + 224 + ], + "type": "text", + "content": ", Longtian Qiu" + }, + { + "bbox": [ + 132, + 185, + 481, + 224 + ], + "type": "inline_equation", + "content": "^{23*}" + }, + { + "bbox": [ + 132, + 185, + 481, + 224 + ], + "type": "text", + "content": ", Han Xiao" + }, + { + "bbox": [ + 132, + 185, + 481, + 224 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 132, + 185, + 481, + 224 + ], + "type": "text", + "content": ", Han Qiu" + }, + { + "bbox": [ + 132, + 185, + 481, + 224 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 132, + 185, + 481, + 224 + ], + "type": "text", + "content": ", Wenqi Shao" + }, + { + "bbox": [ + 132, + 185, + 481, + 224 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 132, + 185, + 481, + 224 + ], + "type": "text", + "content": ", Keqin Chen" + }, + { + "bbox": [ + 132, + 185, + 481, + 224 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 132, + 185, + 481, + 224 + ], + "type": "text", + "content": ", Jiaming Han" + }, + { + "bbox": [ + 132, + 185, + 481, + 224 + ], + "type": "inline_equation", + "content": "^{12}" + }, + { + "bbox": [ + 132, + 185, + 481, + 224 + ], + "type": "text", + "content": ", Siyuan Huang" + }, + { + "bbox": [ + 132, + 185, + 481, + 224 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 132, + 185, + 481, + 224 + ], + "type": "text", + "content": ", Yichi Zhang" + }, + { + "bbox": [ + 132, + 185, + 481, + 224 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 132, + 185, + 481, + 224 + ], + "type": "text", + "content": ", Xuming He" + }, + { + "bbox": [ + 132, + 185, + 481, + 224 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 132, + 185, + 481, + 224 + ], + "type": "text", + "content": ", Yu Qiao" + }, + { + "bbox": [ + 132, + 185, + 481, + 224 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 132, + 185, + 481, + 224 + ], + "type": "text", + "content": ", and Hongsheng Li" + }, + { + "bbox": [ + 132, + 185, + 481, + 224 + ], + "type": "inline_equation", + "content": "^{124\\dagger}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 175, + 231, + 441, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 231, + 441, + 243 + ], + "spans": [ + { + "bbox": [ + 175, + 231, + 441, + 243 + ], + "type": "text", + "content": "1 Multimedia Laboratory, The Chinese University of Hong Kong" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 192, + 243, + 422, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 243, + 422, + 254 + ], + "spans": [ + { + "bbox": [ + 192, + 243, + 422, + 254 + ], + "type": "text", + "content": "2 Shanghai AI Laboratory 3 ShanghaiTech University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 185, + 254, + 428, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 185, + 254, + 428, + 265 + ], + "spans": [ + { + "bbox": [ + 185, + 254, + 428, + 265 + ], + "type": "text", + "content": "4 Centre for Perceptual and Interactive Intelligence Limited" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 176, + 265, + 436, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 265, + 436, + 276 + ], + "spans": [ + { + "bbox": [ + 176, + 265, + 436, + 276 + ], + "type": "text", + "content": "* Equal contribution † Equal advisory ‡ Project leader" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 153, + 277, + 460, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 277, + 460, + 287 + ], + "spans": [ + { + "bbox": [ + 153, + 277, + 460, + 287 + ], + "type": "text", + "content": "zylin@link.cuhk.edu.hk, gaopeng@pjlab.org.cn, hsli@ee.cuhk.edu.hk" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 159, + 314, + 455, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 314, + 455, + 547 + ], + "spans": [ + { + "bbox": [ + 159, + 314, + 455, + 547 + ], + "type": "text", + "content": "Abstract. We present SPHINX, a versatile multi-modal large language model (MLLM) with a joint mixing of model weights, visual embeddings and image scales. First, for stronger vision-language alignment, we unfreeze the large language model (LLM) during pre-training, and introduce a weight mix strategy between LLMs trained by real-world and synthetic data. By directly integrating the weights from two domains, the mixed LLM can efficiently incorporate diverse semantics with favorable robustness. Then, we propose to extract comprehensive visual embeddings from various network architectures, pre-training paradigms, and information granularity, providing language models with more robust image representations. We further propose an efficient strategy aiming to better capture fine-grained appearances of high-resolution images. With a mixing of different scales and high-resolution sub-images, SPHINX attains exceptional visual parsing and reasoning performance on existing evaluation benchmarks. Based on our proposed joint mixing, SPHINX exhibits superior multi-modal understanding capabilities on a wide range of applications, with highlighted fine-grained visual recognition abilities such as region-level understanding, caption grounding, document layout detection, and human pose estimation. We hope our work may cast a light on the exploration of joint mixing in future MLLM research. Code is released at https://github.com/Alpha-VLLM/LLaMA2-Accessory." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 581, + 230, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 581, + 230, + 594 + ], + "spans": [ + { + "bbox": [ + 132, + 581, + 230, + 594 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Since the era of big data, large language models (LLMs) have attained tremendous strides (7; 61; 62; 87; 108), showcasing unprecedented application scenarios and generalization capabilities. Despite the phenomenal success, plain LLMs are still only reasoning models, and the lack of perception capability has limited their further evolution to autonomous agents with general intelligence. The mission" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 140, + 108, + 212, + 116 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 108, + 212, + 116 + ], + "spans": [ + { + "bbox": [ + 140, + 108, + 212, + 116 + ], + "type": "text", + "content": "Emergent Referring Pose Estimation" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 140, + 116, + 149, + 122 + ], + "blocks": [ + { + "bbox": [ + 140, + 116, + 149, + 122 + ], + "lines": [ + { + "bbox": [ + 140, + 116, + 149, + 122 + ], + "spans": [ + { + "bbox": [ + 140, + 116, + 149, + 122 + ], + "type": "image", + "image_path": "dbffd25f30333c077f55064be515e1510236d599e46b160aa63ac9f463cc721b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 149, + 117, + 226, + 130 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 117, + 226, + 130 + ], + "spans": [ + { + "bbox": [ + 149, + 117, + 226, + 130 + ], + "type": "text", + "content": "Else: \n - Select the keypoints of the person wearing a green suit." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 140, + 131, + 149, + 143 + ], + "blocks": [ + { + "bbox": [ + 140, + 131, + 149, + 143 + ], + "lines": [ + { + "bbox": [ + 140, + 131, + 149, + 143 + ], + "spans": [ + { + "bbox": [ + 140, + 131, + 149, + 143 + ], + "type": "image", + "image_path": "bff5837be91f9bf40bc591fea0853f6d5418f5359ce5201e9813799030586359.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 150, + 134, + 221, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 134, + 221, + 162 + ], + "spans": [ + { + "bbox": [ + 150, + 134, + 221, + 162 + ], + "type": "text", + "content": "SPINDX: The detected key points are left ear[0.76,0.22], left shoulder[0.78,0.32], right shoulder[0.88,0.30], left elbow[0.74,0.49], right elbow[0.90,0.47], wrist[0.68,0.52], right wrist[0.81,0.46], left hip[0.80,0.65], right hip[0.90,0.64], knee[0.71,0.88], and right knee[0.84,0.88]." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 228, + 110, + 296, + 159 + ], + "blocks": [ + { + "bbox": [ + 228, + 110, + 296, + 159 + ], + "lines": [ + { + "bbox": [ + 228, + 110, + 296, + 159 + ], + "spans": [ + { + "bbox": [ + 228, + 110, + 296, + 159 + ], + "type": "image", + "image_path": "08c1336de65bd73b330d66b057276387e2bc846d3ae38023285ebd48b2597e58.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 110, + 361, + 117 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 110, + 361, + 117 + ], + "spans": [ + { + "bbox": [ + 304, + 110, + 361, + 117 + ], + "type": "text", + "content": "Multi-level Dense Captioning" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 118, + 362, + 123 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 118, + 362, + 123 + ], + "spans": [ + { + "bbox": [ + 302, + 118, + 362, + 123 + ], + "type": "text", + "content": "User: Detect all objectives in this image." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 124, + 379, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 124, + 379, + 156 + ], + "spans": [ + { + "bbox": [ + 302, + 124, + 379, + 156 + ], + "type": "text", + "content": "SHINE: There are 10 objects shown in the image, including a person [0.21, 0.39, 37.0, 76], a helmet[0.30,0.39,0.36,0.43], two sneakers [0.67,0.75,0.85,0.95,0.76], bicycle[0.03,0.53,0.43,0.76], a cell phone[0.22,0.40,0.33,0.42], and four polo shirt[0.50,0.59,0.59,0.59,0.48,0.44,0.48,0.43,0.43,0.16,0.59]." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 382, + 110, + 473, + 170 + ], + "blocks": [ + { + "bbox": [ + 382, + 110, + 473, + 170 + ], + "lines": [ + { + "bbox": [ + 382, + 110, + 473, + 170 + ], + "spans": [ + { + "bbox": [ + 382, + 110, + 473, + 170 + ], + "type": "image", + "image_path": "2a86f2ad90d5b69658e3d97b7061e2153230323078e6175baeb82fdd2e8e4d5d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 172, + 214, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 172, + 214, + 178 + ], + "spans": [ + { + "bbox": [ + 141, + 172, + 214, + 178 + ], + "type": "text", + "content": "Emergent Anomaly Detection" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 180, + 200, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 180, + 200, + 196 + ], + "spans": [ + { + "bbox": [ + 141, + 180, + 200, + 196 + ], + "type": "text", + "content": "User: Is there defect on the image? If yes detect it and plot the box on the image" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 201, + 180, + 293, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 180, + 293, + 196 + ], + "spans": [ + { + "bbox": [ + 201, + 180, + 293, + 196 + ], + "type": "text", + "content": "SPHINX: Yes, there is a defect on the image. The defect is a scratch on the metal object. The box on the image is " + }, + { + "bbox": [ + 201, + 180, + 293, + 196 + ], + "type": "inline_equation", + "content": "[0.450.46, 0.500.57]" + }, + { + "bbox": [ + 201, + 180, + 293, + 196 + ], + "type": "text", + "content": ", which is the scratch on the metal object." + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 141, + 199, + 192, + 239 + ], + "blocks": [ + { + "bbox": [ + 141, + 199, + 192, + 239 + ], + "lines": [ + { + "bbox": [ + 141, + 199, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 141, + 199, + 192, + 239 + ], + "type": "image", + "image_path": "e18427647bc9055a328fd86d13f8dd6acbd5e67a5a509d6c82b8bebaf215ccab.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 194, + 199, + 243, + 239 + ], + "blocks": [ + { + "bbox": [ + 194, + 199, + 243, + 239 + ], + "lines": [ + { + "bbox": [ + 194, + 199, + 243, + 239 + ], + "spans": [ + { + "bbox": [ + 194, + 199, + 243, + 239 + ], + "type": "image", + "image_path": "c2114a7b6ea442c886589f514ccb79c8d087bb187efbb3a1470bb3e487d1078c.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 245, + 199, + 294, + 239 + ], + "blocks": [ + { + "bbox": [ + 245, + 199, + 294, + 239 + ], + "lines": [ + { + "bbox": [ + 245, + 199, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 245, + 199, + 294, + 239 + ], + "type": "image", + "image_path": "b4df21ab2b24e733942d412936c1525ebe66843dfe4b9df8a389eca7072e2b88.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 246, + 213, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 246, + 213, + 254 + ], + "spans": [ + { + "bbox": [ + 141, + 246, + 213, + 254 + ], + "type": "text", + "content": "Image Editing (w/ Stable Diffusion)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 145, + 257, + 190, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 257, + 190, + 266 + ], + "spans": [ + { + "bbox": [ + 145, + 257, + 190, + 266 + ], + "type": "text", + "content": "SPEHIN: Detect road sign in this picture" + } + ] + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 145, + 266, + 191, + 296 + ], + "blocks": [ + { + "bbox": [ + 145, + 266, + 191, + 296 + ], + "lines": [ + { + "bbox": [ + 145, + 266, + 191, + 296 + ], + "spans": [ + { + "bbox": [ + 145, + 266, + 191, + 296 + ], + "type": "image", + "image_path": "762006744f2e2d5a1a7b63a214c46295da302e9b16ad402b961b9b354d2c45de.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 194, + 257, + 240, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 257, + 240, + 263 + ], + "spans": [ + { + "bbox": [ + 194, + 257, + 240, + 263 + ], + "type": "text", + "content": "SPHINX: Detect dog in this" + } + ] + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 195, + 263, + 242, + 296 + ], + "blocks": [ + { + "bbox": [ + 195, + 263, + 242, + 296 + ], + "lines": [ + { + "bbox": [ + 195, + 263, + 242, + 296 + ], + "spans": [ + { + "bbox": [ + 195, + 263, + 242, + 296 + ], + "type": "image", + "image_path": "7be2eb41567a3321e3ccd14803ffa537885a1fbd4373ca7a0b43150e6229b5e1.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "bbox": [ + 246, + 257, + 291, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 257, + 291, + 263 + ], + "spans": [ + { + "bbox": [ + 246, + 257, + 291, + 263 + ], + "type": "text", + "content": "SPHINX: Detect monocite in this picture." + } + ] + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 246, + 263, + 292, + 296 + ], + "blocks": [ + { + "bbox": [ + 246, + 263, + 292, + 296 + ], + "lines": [ + { + "bbox": [ + 246, + 263, + 292, + 296 + ], + "spans": [ + { + "bbox": [ + 246, + 263, + 292, + 296 + ], + "type": "image", + "image_path": "826bcd4a3347ca2b110f5f2e35362f10a3e0c83d4ea85f57723a2a6a48cd7211.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 145, + 299, + 186, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 299, + 186, + 306 + ], + "spans": [ + { + "bbox": [ + 145, + 299, + 186, + 306 + ], + "type": "text", + "content": "Stable Diffusion: A light in the middle of the picture" + } + ] + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 145, + 306, + 191, + 337 + ], + "blocks": [ + { + "bbox": [ + 145, + 306, + 191, + 337 + ], + "lines": [ + { + "bbox": [ + 145, + 306, + 191, + 337 + ], + "spans": [ + { + "bbox": [ + 145, + 306, + 191, + 337 + ], + "type": "image", + "image_path": "47720c29c7c4bc1c5a495a8f03e6a91cf657d56db06abfa9215db01d0e3003a2.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 357, + 482, + 401 + ], + "lines": [ + { + "bbox": [ + 132, + 357, + 482, + 401 + ], + "spans": [ + { + "bbox": [ + 132, + 357, + 482, + 401 + ], + "type": "text", + "content": "Fig. 1: Examples of multi-task visual reasoning on selected tasks by our proposed SPHINX models, which excels in diverse visual understanding and perception tasks, such as object detection, caption grounding, and region-level description, displaying some emergent capabilities." + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "bbox": [ + 194, + 299, + 242, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 299, + 242, + 305 + ], + "spans": [ + { + "bbox": [ + 194, + 299, + 242, + 305 + ], + "type": "text", + "content": "Stable Diffusion: A teddy bear sitting in a basket" + } + ] + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 195, + 305, + 242, + 335 + ], + "blocks": [ + { + "bbox": [ + 195, + 305, + 242, + 335 + ], + "lines": [ + { + "bbox": [ + 195, + 305, + 242, + 335 + ], + "spans": [ + { + "bbox": [ + 195, + 305, + 242, + 335 + ], + "type": "image", + "image_path": "9bdf9cef7c90c7c2cb7054de5c4c616ab725176b36ce1eaaea76cc8b97957610.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "bbox": [ + 246, + 299, + 293, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 299, + 293, + 305 + ], + "spans": [ + { + "bbox": [ + 246, + 299, + 293, + 305 + ], + "type": "text", + "content": "Stable Diffusion: A man riding a bicycle" + } + ] + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 246, + 305, + 292, + 335 + ], + "blocks": [ + { + "bbox": [ + 246, + 305, + 292, + 335 + ], + "lines": [ + { + "bbox": [ + 246, + 305, + 292, + 335 + ], + "spans": [ + { + "bbox": [ + 246, + 305, + 292, + 335 + ], + "type": "image", + "image_path": "f25ee9bc17315cf3fb9aa75284d825782cde17bbbb09645737e6686c0e51f79c.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "bbox": [ + 304, + 175, + 345, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 175, + 345, + 187 + ], + "spans": [ + { + "bbox": [ + 304, + 175, + 345, + 187 + ], + "type": "text", + "content": "User: Please provide a short description for this document." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 304, + 188, + 342, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 188, + 342, + 194 + ], + "spans": [ + { + "bbox": [ + 304, + 188, + 342, + 194 + ], + "type": "text", + "content": "SPHINX: A blue bicycle" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 304, + 196, + 356, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 196, + 356, + 205 + ], + "spans": [ + { + "bbox": [ + 304, + 196, + 356, + 205 + ], + "type": "text", + "content": "User: Provide a one-sentence caption for the provided image" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 304, + 207, + 356, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 207, + 356, + 217 + ], + "spans": [ + { + "bbox": [ + 304, + 207, + 356, + 217 + ], + "type": "text", + "content": "SPHINK: A man riding a bike in a grassy field." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 304, + 219, + 356, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 219, + 356, + 229 + ], + "spans": [ + { + "bbox": [ + 304, + 219, + 356, + 229 + ], + "type": "text", + "content": "User:Generates a detailed description about the image." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 304, + 230, + 357, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 230, + 357, + 287 + ], + "spans": [ + { + "bbox": [ + 304, + 230, + 357, + 287 + ], + "type": "text", + "content": "SHANK: The image shows a man standing next to a blue mountain bike in a grassy area. He is standing on the right leg of the bike, with his right leg on the pedal. The man appears to be in the process of mounting the bike, as he is standing on the pedal and holding his foot. There are several yellow flowers scattered throughout the grassy area, adding a touch of color to the scene. This is wearing shorts, which are visible as the stands next to the bike." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 363, + 175, + 420, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 175, + 420, + 186 + ], + "spans": [ + { + "bbox": [ + 363, + 175, + 420, + 186 + ], + "type": "text", + "content": "User: Please provide a short description for this region:" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 363, + 187, + 411, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 187, + 411, + 194 + ], + "spans": [ + { + "bbox": [ + 363, + 187, + 411, + 194 + ], + "type": "text", + "content": "SPHINK: A man in a blue shirt" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 363, + 196, + 415, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 196, + 415, + 204 + ], + "spans": [ + { + "bbox": [ + 363, + 196, + 415, + 204 + ], + "type": "text", + "content": "User: Provide a one-Sentence caption for the provided image" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 363, + 207, + 416, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 207, + 416, + 217 + ], + "spans": [ + { + "bbox": [ + 363, + 207, + 416, + 217 + ], + "type": "text", + "content": "SHINX: A man riding a bike while holding a cell phone." + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 363, + 218, + 416, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 218, + 416, + 225 + ], + "spans": [ + { + "bbox": [ + 363, + 218, + 416, + 225 + ], + "type": "text", + "content": "User: Generate a detailed description about the image." + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 363, + 228, + 416, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 228, + 416, + 281 + ], + "spans": [ + { + "bbox": [ + 363, + 228, + 416, + 281 + ], + "type": "text", + "content": "SHINK: The image shows a man sitting on a bicycle in a grassy area, and the grassy area is filled with people to be taking a selfie or recording a video while sitting on the bike. The grassy area is filled with people who he is prepared for a safe ride. There are several other bicycles in the scene, with one located to the left of the bike, and another located to the right, and a third one further back. The grassy area is scattered with people standing around, riding a couch on color to the scene." + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 419, + 175, + 473, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 175, + 473, + 187 + ], + "spans": [ + { + "bbox": [ + 419, + 175, + 473, + 187 + ], + "type": "text", + "content": "User: Please provide a short description of the region [0.37,4.40,97.80]" + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 419, + 188, + 471, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 188, + 471, + 198 + ], + "spans": [ + { + "bbox": [ + 419, + 188, + 471, + 198 + ], + "type": "text", + "content": "SHINK: A cow licking a man's face" + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 419, + 200, + 472, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 200, + 472, + 209 + ], + "spans": [ + { + "bbox": [ + 419, + 200, + 472, + 209 + ], + "type": "text", + "content": "User: Provide a one-sentence caphor for the provided image" + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 419, + 210, + 464, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 210, + 464, + 219 + ], + "spans": [ + { + "bbox": [ + 419, + 210, + 464, + 219 + ], + "type": "text", + "content": "SPEEDON: A cow is sniffing a bicycle wheel." + } + ] + } + ], + "index": 46 + }, + { + "bbox": [ + 419, + 220, + 460, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 220, + 460, + 230 + ], + "spans": [ + { + "bbox": [ + 419, + 220, + 460, + 230 + ], + "type": "text", + "content": "Uc:Generale a deblie du about the image." + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 419, + 232, + 471, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 232, + 471, + 277 + ], + "spans": [ + { + "bbox": [ + 419, + 232, + 471, + 277 + ], + "type": "text", + "content": "SPINDL: The image features a large brown and white cow standing in a grassy field. The cow is standing on the ground, looking at the edge of the field, and it appears to be sniffing a bicycle tire. Another cow is standing slightly to the right of the main cow. The scene captures the cow's perspective as it interacts with the bicycle tire." + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 130, + 412, + 480, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 412, + 480, + 459 + ], + "spans": [ + { + "bbox": [ + 130, + 412, + 480, + 459 + ], + "type": "text", + "content": "to integrate reasoning and perception has motivated the research on multi-modal large language models (MLLMs) (41; 50; 106; 109; 112), which make responses to not only human language, but also natural signals, among which visual signals are predominantly studied." + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 130, + 460, + 481, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 481, + 555 + ], + "type": "text", + "content": "Early works on MLLMs (22; 50; 102) focus on coarse-grained tasks, such as image captioning and question answering. While this enables visual instruction following or visual dialogue to some extent, fine-grained visual recognition capabilities, such as visual grounding, key-point detection and referring generation, might bring new outlook to the development of MLLMs: They make possible new applications such as precise actions in embodied AI, as well as potentially improved diversity and reduced hallucination on traditional tasks with better access to detailed visual clues." + } + ] + } + ], + "index": 51 + }, + { + "bbox": [ + 130, + 557, + 481, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 481, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 481, + 628 + ], + "type": "text", + "content": "Besides fine-grained recognition, it is equally important for a foundational MLLM to maintain the performance on a wider domain. Despite the success of previous works such as Shikra (10) which tackles fine-grained tasks alone, it is still challenging to explore the scaling-up methods to utilize larger datasets, higher parameter counts, higher resolutions, etc. in order to equip the models with broader or even emergent abilities." + } + ] + } + ], + "index": 52 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "We thus present the building of SPHINX, our proposed foundational MLLM with versatility and an emphasis on fine-grained visual recognition capabilities. Compared to previous works with similar goals, such as Qwen-VL (4)" + } + ] + } + ], + "index": 53 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "text", + "content": "Z. Lin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "and MiniGPT-v2 (9), our pipeline can be highlighted with the following unique scaling-up techniques conveniently named as mixing strategies:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 152, + 481, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 152, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 138, + 152, + 481, + 270 + ], + "type": "text", + "content": "- Mixed model weights. Vision-language data from particular domains might contain special semantics, e.g., synthetic captions (75) compared to real-world ones (76). Considering that directly mixing such data might confuse the MLLM, we introduce a weight-mixing strategy to efficiently combine such domain-specific knowledge. Based on the MLLM pre-trained on real-world data, we fine-tune it on the synthetic data, and then linearly combine the finetuned LLM's weights with the real-world ones. In this way, the two types of models would not be affected by contradictory data and our final SPHINX can effectively integrate knowledge from both synthetic and real-world domains, enabling effective use of datasets from diverse sources." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 271, + 487, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 271, + 487, + 354 + ], + "spans": [ + { + "bbox": [ + 138, + 271, + 487, + 354 + ], + "type": "text", + "content": "- Mixed visual embeddings. To take the advantage of different encoders, we propose to mix the visual embeddings from various vision backbones (41; 64; 70) with different network architectures (CNN vs. ViT), pre-training paradigms (supervised vs. self-supervised), and information granularity (global vs. local). By mixing the different image tokens channel-wisely and sequence-wisely, SPHINX obtains stronger visual representations and leads to better vision-language alignment efficacy." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 356, + 490, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 356, + 490, + 475 + ], + "spans": [ + { + "bbox": [ + 138, + 356, + 490, + 475 + ], + "type": "text", + "content": "- Mixed scales with high-resolution sub-images. To process high-resolution images as required for fine-grained visual recognition, we first spatially divide the input high-resolution image into multiple sub-images, and also downsample it into a low-resolution one. Then, we feed all the images concurrently into the mixed visual encoders, and concatenate the extracted multiple token groups to represent the entire high-resolution visual features. By mixing visual embeddings of different scales and sub-images, our SPHINX can adaptively explore more fine-grained visual semantics from the high resolution and multi-scale image representations, while maintaining encoding efficiency." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 486, + 485, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 485, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 485, + 582 + ], + "type": "text", + "content": "By the proposed three-fold mixing strategy, SPHINX fine-tunes LLMs, e.g., LLaMA-2 (88), to be a powerful MLLM with superior visual instruction-following capacity. Our model excels in a variety of vision tasks, e.g., detecting different objects with remarkable precision and parsing their relations, or accurately interpreting the content within complicated figures. Importantly, SPHINX can achieve impressive fine-grained visual perception for high-resolution images, which exhibits state-of-the-art performance on extensive evaluation benchmarks, e.g., MMBench (52), MME (20), and POPE (45)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 602, + 239, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 602, + 239, + 615 + ], + "spans": [ + { + "bbox": [ + 131, + 602, + 239, + 615 + ], + "type": "text", + "content": "2 Related works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "content": "Large language models (LLMs). The field of Natural Language Processing (NLP) has witnessed significant progress over the years, particularly with the advent of LLMs. With Transformer (89) as the fundamental architecture," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "type": "text", + "content": "A Mixer of Weights, Visual Embeddings and Image Scales for MLLM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 355 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 355 + ], + "type": "text", + "content": "LLMs (61; 62; 72) have demonstrated unprecedented performance in modeling intricate language patterns over extensive contexts. Therein, BERT (15) showcases the benefits of pre-training on vast text corpora and fine-tuning on specific tasks, setting new standards on various benchmarks. OpenAI's GPT series (61; 62; 71; 72), especially GPT-3 (7), harness the power of massive model scaling, with billions and even trillions of parameters. To obtain better instruction following ability, InstructGPT (65) and ChatGPT (61) are presented to exhibit exceptional fluency and versatility in open-domain conversation tasks, ranging from text generation to question answering. Recently, the instruction tuning based on LLaMA (87) and LLaMA-2 (88) has gained great popularity as open-source LLMs in the community. Therein, Alpaca (86) and LLaMA-Adapter (106) respectively adopt full and parameter-efficient fine-tuning to acquire favorable instruction-following LLMs. Vicuna (12) and GPT-4-LLM (67) further showcase the improvement brought by higher-quality instruction datasets. Other efforts also extend LLMs for match problem solving (90; 111), visual model system (96; 101), and open-world recognition (107; 113). In this paper, we develop our SPHINX based on the superior language understanding of LLaMA-2 (88) and instruction tuning experience of LLaMA-Adapter series (22; 106), which introduce a three-fold mixer to extend the capability ceiling of instruction-following LLMs for multi-modal input." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 361, + 482, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 482, + 636 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 482, + 636 + ], + "type": "text", + "content": "Multi-modal large language models (MLLMs). In addition to language instruction following, many efforts have been made to inject multi-modal conditions into LLMs for wider application scenarios. As prior attempts, VisualGPT (8) and BLIP series (14; 41; 42) indicate the potential of aligning LLMs with visual input for image captioning and question answering. Flamingo (3) and Kosmos-1 (32) further exhibit promising multi-modal understanding performance for image-text interleaved contexts. With large-scale pre-training and model sizes, GPT-4 (62) and Bard (24) both showcase remarkable proficiency in vision-language understanding and reasoning over diverse multi-modal tasks. In parallel, a bunch of works have been proposed to align LLaMA with vision modality for advanced visual instruction-following capabilities. LLaVA (50) and MiniGPT-4 (112) utilize a simple projection layer to connect vision encoders (41; 70) with LLMs. LLaMA-Adapter V2 (22) introduces zero-initialized attention mechanisms for efficient visual instruction tuning, and mPLUG-Owl (102) adopts delicately designed intermediate networks for cross-modal alignment. For more modality input, ImageBind-LLM (28) and PandaGPT (83) further incorporate audio and video conditions guided by ImageBind (23). Besides, recent MLLMs are also extended to region-level parsing (10; 68), in-context learning (37; 38), arbitrary image resolutions (6), text-to-image generation (16; 93), and 3D question answering (25; 30; 98). Different from previous works, our SPHINX aims for image-conditioned MLLM, and proposes a three-fold mixer, i.e., model weights, tuning tasks, and visual embeddings, attaining superior generalization capacity for multi-modal learning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Weight interpolation. Despite being commonly known to be highly nonlinear (40), neural networks have, under some conditions, exhibited surprisingly" + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "text", + "content": "Z. Lin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 152, + 118, + 455, + 240 + ], + "blocks": [ + { + "bbox": [ + 152, + 118, + 455, + 240 + ], + "lines": [ + { + "bbox": [ + 152, + 118, + 455, + 240 + ], + "spans": [ + { + "bbox": [ + 152, + 118, + 455, + 240 + ], + "type": "image", + "image_path": "e50bfc0f53a05f22b325d62f9dba5b5fda3ebf8df5e46606a68345a80fa8ef69.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 257, + 480, + 280 + ], + "lines": [ + { + "bbox": [ + 132, + 257, + 480, + 280 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 480, + 280 + ], + "type": "text", + "content": "Fig. 2: The joint mixing paradigm of SPHINX. with mixed visual embeddings (a) and mixed model weights (b)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 301, + 480, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 301, + 480, + 386 + ], + "spans": [ + { + "bbox": [ + 130, + 301, + 480, + 386 + ], + "type": "text", + "content": "good results with linearly interpolated weights in several cases, including higher-performing single-task models (11; 95), task interpolation (31; 97; 100), federated or low-communication distributed training (19; 43; 59). A few recent works (2; 80; 84) also investigated weight mixing under the MLLM context. On top of these works, we further demonstrate the effects of inter-domain weight mixing for image-c captioning pre-trained models, as well as how these benefits persist after further task-specific fine-tuning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 386, + 481, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 386, + 481, + 577 + ], + "spans": [ + { + "bbox": [ + 130, + 386, + 481, + 577 + ], + "type": "text", + "content": "High-resolution visual embeddings. Fine-grained visual recognition tasks, such as detection and segmentation, traditionally utilize higher-resolution image features to capture detailed information. Convolutional neural networks (CNNs) naturally scale to high-resolution inputs (53) as the convolution operation intrinsically supports varied input sizes, while for the recent transformer-based vision backbones, scaling input resolution is commonly achieved by interpolating the positional embeddings (18). There are also network components specifically designed to obtain high-resolution feature maps, e.g., (44; 46; 103; 110). Unlike those model-surgery-based methods, our mixed-scale solution instead explores the use of fine-tuned large language models to fuse the visual embeddings from multiple parts of the image while leaving the visual encoders themselves untouched. With such a zooming-in mechanism, our framework captures fine-grained visual clues while preserves capabilities of the visual encoders in an architecture-agnostic manner. Interestingly, the recently released GPT-4V API (63) shows a similar correspondence between the image resolution and the number of visual tokens." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 594, + 208, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 594, + 208, + 605 + ], + "spans": [ + { + "bbox": [ + 132, + 594, + 208, + 605 + ], + "type": "text", + "content": "3 SPHINX" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "type": "text", + "content": "In this section, we introduce a versatile MLLM, SPHINX, with an model overview in Section 3.1, the joint mixing of model weights and visual embeddings in Section 3.2 and the mixing of high-resolution sub-image tokens Section 3.3. Finally, in Section 3.4, we introduce several extended applications of SPHINX." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "type": "text", + "content": "A Mixer of Weights, Visual Embeddings and Image Scales for MLLM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 242, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 242, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 242, + 128 + ], + "type": "text", + "content": "3.1 Model overview" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 482, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 482, + 209 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 482, + 209 + ], + "type": "text", + "content": "The SPHINX model is composed of an LLM, e.g., LLaMA-2 (88), a mixture of vision encoders and linear projections connecting the vision encoder outputs to the language model inputs. For sequences with image input, visual tokens are extracted from the vision encoders, projected to the LM input space, and then concatenated to the beginning of the sequence. Sequences without image inputs are processed identically as in the language-only LLM." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 209, + 483, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 209, + 483, + 269 + ], + "spans": [ + { + "bbox": [ + 130, + 209, + 483, + 269 + ], + "type": "text", + "content": "We adopt a two-stage training paradigm: the first pre-training stage for vision-language alignment, and the second fine-tuning stage for visual instruction-following learning. In both stages, the LM parameters are fully-tuned, together with the linear layers for visual-language feature alignment, but the visual encoders themselves are kept frozen." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 288, + 479, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 288, + 479, + 312 + ], + "spans": [ + { + "bbox": [ + 131, + 288, + 479, + 312 + ], + "type": "text", + "content": "3.2 Pre-training: The joint mixing of visual embeddings and model weights" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 321, + 482, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 321, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 321, + 482, + 380 + ], + "type": "text", + "content": "The pre-training stage (or stage 1) involves two types of mixing: weight mixing and visual embedding mixing, as shown in Figure 2. Since the pre-training datasets are weakly annotated image-caption pairs where the textual information does not contain much detail, we do not use high-resolution images for this stage, thus bypassing the scale mixing." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 381, + 487, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 381, + 487, + 549 + ], + "spans": [ + { + "bbox": [ + 130, + 381, + 487, + 549 + ], + "type": "text", + "content": "Mixed model weights of different domains. Vision-language data from different domains may exhibit distinct characteristics: For example, web-crawled captions like those in LAION-400M (76) tend to be diverse but noisy, but synthetic captions like those in LAION-COCO (75) tend to be coherent but of a relatively limited vocabulary. We propose a weight mixing strategy of domain-specifically tuned LLMs to integrate respective knowledge from real-world and synthetic data. We first utilize the most common domain data (LAION-400M (76)) for pre-training, which endows the MLLM with fundamental visual understanding capabilities. Then, we regard such a pre-trained model as the initial checkpoint to further fine-tune the LLM on synthetic domains, e.g., LAION-COCO (75). Finally, to take advantage of the best data domains, we directly conduct a weighted mixing of two LLMs' weights for semantic aggregation. In detail, we denote the parameters of the fundamental LLM as " + }, + { + "bbox": [ + 130, + 381, + 487, + 549 + ], + "type": "inline_equation", + "content": "\\theta_{real}" + }, + { + "bbox": [ + 130, + 381, + 487, + 549 + ], + "type": "text", + "content": ", and the fine-tuned parameters by synthetic data as " + }, + { + "bbox": [ + 130, + 381, + 487, + 549 + ], + "type": "inline_equation", + "content": "\\theta_{syn}" + }, + { + "bbox": [ + 130, + 381, + 487, + 549 + ], + "type": "text", + "content": ". The mixing process is formulated as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 238, + 559, + 481, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 559, + 481, + 572 + ], + "spans": [ + { + "bbox": [ + 238, + 559, + 481, + 572 + ], + "type": "interline_equation", + "content": "\\theta_ {m i x} = \\beta \\cdot \\theta_ {r e a l} + (1 - \\beta) \\cdot \\theta_ {s y n}, \\tag {1}", + "image_path": "91d685701804aff5ab1833d12adea4b9ab9b89545d571affc7bca288d3a4e933.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 581, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 640 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 581, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 581, + 482, + 640 + ], + "type": "text", + "content": " denotes the mixing coefficient, and " + }, + { + "bbox": [ + 130, + 581, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\theta_{\\text{mix}}" + }, + { + "bbox": [ + 130, + 581, + 482, + 640 + ], + "type": "text", + "content": " represents the mixed LLM weights with aggregated semantics. Compared to fusing different domain data for joint pre-training, our weight mix strategy can encourage every MLLM to better learn domain-unique knowledge, and exhibit flexible scalability for any new data domains." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "content": "Mixed embeddings for visual encoding. To capture robust visual representations from different aspects, we propose to ensemble a variety of vision" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "Z. Lin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 171, + 113, + 449, + 285 + ], + "blocks": [ + { + "bbox": [ + 171, + 113, + 449, + 285 + ], + "lines": [ + { + "bbox": [ + 171, + 113, + 449, + 285 + ], + "spans": [ + { + "bbox": [ + 171, + 113, + 449, + 285 + ], + "type": "image", + "image_path": "00591b1905a8833dc4bbfc8b522dd169744614c00a7788e3377bd119572b6e9a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 293, + 482, + 327 + ], + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 327 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 327 + ], + "type": "text", + "content": "Fig. 3: Pipeline of SPHINX for high-resolution images. We propose to further mix different scales and sub-images to better capture fine-grained semantics on high-resolution images." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 353, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 353, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 353, + 482, + 594 + ], + "type": "text", + "content": "backbones for image encoding. The visual backbones with different characteristics are chosen as follows. 1) Different network architectures. As CNN (29) and ViT (18) mainly aggregate different types of visual appearances, i.e., neighboring dependencies and long-range interactions, we adopt CLIP (70) models respectively with ConvNeXt (94) and ViT image encoders. 2) Different pre-training paradigms. Supervised training can impose explicit semantic information from textual captions or category labels, while self-supervised learning enforces the model to explore implicit pretext task signals. Thus, we further employ the ViT self-supervised by DINOv2 (64) as well as the text-supervised vision encoders, CLIP. 3) Different information granularity. The aforementioned visual encoders all produce visual tokens in the patch level. To better capture global features, we also adopt Q-Former (41) to summarize visual embeddings via querying from the global context. After all the aforementioned encoding, we first channel-wisely concatenate the patch level visual tokens. The channel-wise concatenation allows for better alignment among spatially corresponding features, as well as keeps the number of tokens constant to the number of visual encoders. Then, by using two projection layers for dimension alignment, we spatial-wisely concatenate the representations between those of Q-Former and the other patch-level features. The obtained image tokens are directly placed in front of language instructions, which provide visual context for the language instructions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 616, + 458, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 616, + 458, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 616, + 458, + 628 + ], + "type": "text", + "content": "3.3 Fine-tuning: Mixing scales with high-resolution sub-images" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "In the fine-tuning stage, or stage 2 training, the model from stage 1 will adapt for instruct following as well as high-resolution image recognition. Stage 2 introduces" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "type": "text", + "content": "A Mixer of Weights, Visual Embeddings and Image Scales for MLLM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "content": "scale mixing to process high-resolution input images, which are crucial for fine-grained visual tasks and will be elaborated in this section." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 144, + 482, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 144, + 482, + 467 + ], + "spans": [ + { + "bbox": [ + 132, + 144, + 482, + 467 + ], + "type": "text", + "content": "Mixed scales and high-resolution sub-images. State-of-the-art open-source MLLMs (9; 10; 22; 41; 50; 68) works adopt frozen image encoders during all training stages, in order to preserve the pre-trained visual semantics. Therefore, the image resolution of MLLMs is usually set as " + }, + { + "bbox": [ + 132, + 144, + 482, + 467 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 132, + 144, + 482, + 467 + ], + "type": "text", + "content": ", severely hindering their efficacy for fine-grained visual perception, especially region-level grounding and description. In our SPHINX, we extend the mixing of visual embeddings to more scales and high-resolution sub-images, allowing for efficient high-resolution image encoding. For an input high-resolution image, e.g., " + }, + { + "bbox": [ + 132, + 144, + 482, + 467 + ], + "type": "inline_equation", + "content": "448 \\times 448" + }, + { + "bbox": [ + 132, + 144, + 482, + 467 + ], + "type": "text", + "content": ", we construct five corresponding images of " + }, + { + "bbox": [ + 132, + 144, + 482, + 467 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 132, + 144, + 482, + 467 + ], + "type": "text", + "content": ", and feed them as independent images into our mixed vision encoders. Specifically, we first downsample the input image to " + }, + { + "bbox": [ + 132, + 144, + 482, + 467 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 132, + 144, + 482, + 467 + ], + "type": "text", + "content": " as an abstract representation, and also downsample the input image to " + }, + { + "bbox": [ + 132, + 144, + 482, + 467 + ], + "type": "inline_equation", + "content": "448 \\times 448" + }, + { + "bbox": [ + 132, + 144, + 482, + 467 + ], + "type": "text", + "content": " and crop four sub-images of " + }, + { + "bbox": [ + 132, + 144, + 482, + 467 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 132, + 144, + 482, + 467 + ], + "type": "text", + "content": " from the four corners of the " + }, + { + "bbox": [ + 132, + 144, + 482, + 467 + ], + "type": "inline_equation", + "content": "448 \\times 448" + }, + { + "bbox": [ + 132, + 144, + 482, + 467 + ], + "type": "text", + "content": " image, which preserve the detailed visual information. In this way, we enable MLLMs to not only capture fine-grained visual appearances with " + }, + { + "bbox": [ + 132, + 144, + 482, + 467 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 132, + 144, + 482, + 467 + ], + "type": "text", + "content": " positional encodings, but also achieve favorable computation efficiency. Afterwards, the five groups of image tokens are encoded and concatenated as a long sequence for feeding into LLM, where the first one group encodes global semantics, and the other four record fine-grained local features. Importantly, as the image tokens of different patches do not have interaction through the vision encoders, they are forced to interact within the LLM to obtain complete visual information. Such a strategy, in turn, motivates LLMs to parse the relations within visual conditions for better cross-modal learning. From this perspective, our SPHINX can be regarded as a new paradigm for similar to ViT (18), where the mixed vision encoders serve as a patch embedding layer, and the LLM plays the role for patch interaction as a vision decoder. On visual understanding tasks requiring higher resolutions, SPHINX achieves significant improvement with the mixed visual representations of scales and high-resolution sub-images." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 470, + 482, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 470, + 482, + 637 + ], + "spans": [ + { + "bbox": [ + 132, + 470, + 482, + 637 + ], + "type": "text", + "content": "On comparison with scaling vision encoder input. The previous standard method to enlarge MLLM input resolutions is to enlarge the input of vision encoders themselves (e.g., by positional embedding interpolation for ViT) and make the vision encoder trainable (4; 9). Compared to vision encoder fine-tuning, we argue that our LLM-based patch fusion is the scalable choice when the vision encoder is large or consists of multiple models as is in our case, because (i) vision encoders can be slow to tune, as the sequence of the same length back-propagating through the vision encoders can be equally costly as backpropagating through the LM, and (ii) vision encoders can be unstable to tune, as different visual backbones may require different tuning configurations (17) given the vast difference in how they are built in the first place. As a result, we end up finding it prohibitively expensive to work out a training recipe for trainable interpolated vision encoders to achieve the same performance as our fixed-vision-encoder method." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 641, + 491, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 491, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 491, + 666 + ], + "type": "text", + "content": "Efficiency. Mixing scales only incurs a " + }, + { + "bbox": [ + 132, + 641, + 491, + 666 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 132, + 641, + 491, + 666 + ], + "type": "text", + "content": " overhead (SPHINX-1k v. SPHINX) during training because " + }, + { + "bbox": [ + 132, + 641, + 491, + 666 + ], + "type": "inline_equation", + "content": "(i)" + }, + { + "bbox": [ + 132, + 641, + 491, + 666 + ], + "type": "text", + "content": " we limit mixed scale training to stage 2, which itself" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Z. Lin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 283 + ], + "type": "text", + "content": "constitutes a smaller portion of the total training time than stage 1, and " + }, + { + "bbox": [ + 130, + 116, + 482, + 283 + ], + "type": "inline_equation", + "content": "(ii)" + }, + { + "bbox": [ + 130, + 116, + 482, + 283 + ], + "type": "text", + "content": " high resolution training images are usually paired with longer text, further dragging down the portion of the overhead in the total training cost. We also measured the inference cost on actual hardware: On a A100 GPU with a batch size of 1 and flash decoding enabled, generating a 200-token paragraph without mixed scales takes 29GB of memory and 5.2s, and generating with mixed scales takes 33GB and 5.5s. The memory and time overhead is moderate because " + }, + { + "bbox": [ + 130, + 116, + 482, + 283 + ], + "type": "inline_equation", + "content": "(i)" + }, + { + "bbox": [ + 130, + 116, + 482, + 283 + ], + "type": "text", + "content": " the memory is dominated by LLM parameters which are invariant to resolution and " + }, + { + "bbox": [ + 130, + 116, + 482, + 283 + ], + "type": "inline_equation", + "content": "(ii)" + }, + { + "bbox": [ + 130, + 116, + 482, + 283 + ], + "type": "text", + "content": " processing image tokens is a prefetching process which utilizes GPU computing cores better than the decoding generation process. As visual encoders are frozen and much smaller (3B parameters combined) compared to the language model (13B parameters), we find mixing visual encoders brings negligible overhead expect for inference-time memory usage (up to 6GB, which is manageable on most modern GPUs)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 302, + 278, + 314 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 302, + 278, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 302, + 278, + 314 + ], + "type": "text", + "content": "3.4 Extensions of SPHINX" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 323, + 482, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 323, + 482, + 406 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 482, + 406 + ], + "type": "text", + "content": "With the strong and general multi-modal understanding capability, SPHINX models can be integrated with other vision models to tackle even more challenging tasks, including language-referred segmentation with SAM (35), image inpainting (104) with LaMa (85) and image editing with Stable Diffusion (73). We have also explored SPHINX as supervised learners, e.g., full-data finetuning on ImageNet-1k (74). The details can be found in the supplementary materials." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 426, + 230, + 440 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 426, + 230, + 440 + ], + "spans": [ + { + "bbox": [ + 132, + 426, + 230, + 440 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 453, + 278, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 278, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 278, + 464 + ], + "type": "text", + "content": "4.1 Implementation details" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 473, + 481, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 481, + 497 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 481, + 497 + ], + "type": "text", + "content": "We provide the important training and evaluation details in this section. More information can be found in the supplementary materials." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 498, + 487, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 487, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 487, + 556 + ], + "type": "text", + "content": "Pre-training datasets. We use two image captioning datasets LAION-400M (76) and LAION-COCO (75) for multi-modal alignment. As we full-fine-tune the language model backbone for long steps, we also jointly train with a text-only dataset RefinedWeb (66) to avoid harming its text reasoning capability due to catastrophic forgetting." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 557, + 481, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 481, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 481, + 616 + ], + "type": "text", + "content": "Pre-training configuration. We fine-tune the weight of the large language model and the randomly-initialized visual projections in the pre-training stage, The visual encoders are kept frozen. We only use the low-resolution view in the pre-training stage, as the web-crawled captions in the LAION datasets are relatively coarse." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "Fine-tuning datasets. In the multi-task fine-tuning phase, our objective is to equip the MLLM with the versatile needs of downstream tasks. Building upon insights from prior research (10; 14; 49; 50; 112), we include instruction following data such as LLaVA (50) and ShareGPT (79), exposing the model to tasks" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "type": "text", + "content": "A Mixer of Weights, Visual Embeddings and Image Scales for MLLM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "text", + "content": "requiring explicit directives. For general Vision Question Answering (VQA), we leverage datasets like VQAv2 (1) and GQA (33). Expanding the scope to out-of-domain knowledge, we integrate datasets like OKVQA (58) and A-OKVQA (77), providing the model with information beyond the training data. Optical Character Recognition (OCR) datasets, such as OCRVQA (60) and TextCaps (81) are utilized to increase the text understanding ability of SPHINX. We introduce abundant general object detection and pose estimation datasets, such as COCO (47) and LVIS (26) to inspire the model's capabilities of localization, classification, and human pose estimation. To address grounding tasks, we incorporate RefCOCO (34) and VG (36) datasets, training the model to handle referring object localization. Additionally, Grounding Caption datasets, such as those from Flickr30k (69), further refine the understanding of descriptions in the context of image regions. We streamline the training by converting all datasets into a multi-turn conversation format." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 283, + 482, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 283, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 482, + 331 + ], + "type": "text", + "content": "Fine-tuning configuration. The trained and frozen network components are identical as the pre-training stage. Training data are sampled from the mixture of datasets following their natural frequencies, i.e., the chance of a dataset being sampled from is proportional to its original size." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 331, + 482, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 482, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 482, + 402 + ], + "type": "text", + "content": "Model variant definition. We mainly use 3 model variants for experiments: SPHINX, SPHINX-1k and SPHINX-2k, which take increasingly higher resolution images as inputs. The relevant details are listed in the supplementary materials. All models use LLaMA-2-13B (88) as the language model backbone unless otherwise specified, resulting in a total parameter count of around 16B with the 4 visual encoders." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 403, + 482, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 475 + ], + "type": "text", + "content": "Evaluation prompt design. In our model evaluation, we prioritize aligning with each benchmark's desired output format. To achieve this, we employ distinct prompts tailored to benchmarks that necessitate long answers, short answers, and multiple-choice responses. The detailed information is provided in the supplementary materials. This approach ensures that our model is capable of handling diverse scenarios." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 491, + 248, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 491, + 248, + 502 + ], + "spans": [ + { + "bbox": [ + 131, + 491, + 248, + 502 + ], + "type": "text", + "content": "4.2 Model evaluation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 510, + 482, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 557 + ], + "type": "text", + "content": "We provide a comprehensive evaluation of SPHINX and showcase results across multiple benchmarks. Our evaluation encompasses both quantitative metrics (in this section) and qualitative assessments (see the supplementary materials, providing a holistic understanding of our VLM model's performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 558, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 640 + ], + "type": "text", + "content": "Benchmarks on multi-modal large language models. We test our model on recently proposed MLLM benchmarks to comprehensively evaluation of the model's characteristic. As shown in Table 1, SPHINX obtains new state-of-the-art performance on 7 out of 11 MLLM benchmarks. Furthermore, the gap between SPHINX and SPHINX-1k on POPE suggests that the introduction of high-resolution sub-images significantly alleviates the visual hallucination problems." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 641, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 482, + 665 + ], + "type": "text", + "content": "Visual question answering. We also evaluate our models on various classical visual question answering (VQA) benchmarks in Table 2. SPHINX achieves" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "text", + "content": "Z. Lin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 114, + 479, + 228 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 479, + 228 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 479, + 228 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 479, + 228 + ], + "type": "table", + "html": "
MethodsPOPE (45)\\(MME^P\\) (21)\\(MMEC^C\\) (21)MMB (52)SEED (39)\\(LLawa^W\\) (50)MM-Vet (105)CCbench (13)MathVista (54)TinyLVLM (78)Touchstone (5)
BLIP-2 (41)85.31293.8--46.438.122.4--284.7-
InstructBLIP-7B (14)---36.053.460.926.212.125.3300.6552.4
InstructBLIP-13B (14)78.91212.8---58.225.6----
Shikra (10)---58.8-------
LLaMA-AdapterV2 (22)-1328.4356.4------229.2590.1
Qwen-VL-7B (4)---38.256.3--5.5---
Qwen-VL-7B-Chat (4)-1487.6360.760.658.2--39.3-316.8645.2
LLaVA1.5-7B (49)85.91510.7-64.358.663.430.516.4---
LLaVA1.5-13B (49)85.91531.3295.467.761.670.735.426.5---
SPHINX80.71476.1322.266.969.173.536.025.627.0-632.4
SPHINX-1k90.81560.2310.067.171.674.336.627.927.5288.9645.0
SPHINX-2k87.21470.6326.865.971.676.940.227.427.8-659.5
", + "image_path": "cd5c70083ad1cbb97cc3f63da6156179b38d91ccd20fb1d13b1d3483e7316da0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 134, + 264, + 479, + 412 + ], + "blocks": [ + { + "bbox": [ + 133, + 228, + 479, + 239 + ], + "lines": [ + { + "bbox": [ + 133, + 228, + 479, + 239 + ], + "spans": [ + { + "bbox": [ + 133, + 228, + 479, + 239 + ], + "type": "text", + "content": "Table 1: Comparison with state-of-the-art methods on MLLM benchmarks." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 264, + 479, + 412 + ], + "lines": [ + { + "bbox": [ + 134, + 264, + 479, + 412 + ], + "spans": [ + { + "bbox": [ + 134, + 264, + 479, + 412 + ], + "type": "table", + "html": "
MethodsGeneral VQAText-Oriented VQA
OK-VQA (58)VQAv2 (1)VizWiz (27)GQA (33)VSR (48)ScienceQA (55)IconVQA (56)TextVQA (82)OCR-VQA (60)
BLIP-2 (41)45.9-19.641.050.9-40.6-40.6
InstructBLIP (14)--33.449.552.1-44.8-44.8
LLaMA-Adapter-v2 (22)49.670.739.845.1---37.4-
Shikra (10)47.277.4-------
Fuyu-8B (6)60.674.2-------
MiniGPT-v2 (9)57.8-53.660.162.9-51.5--
Qwen-VL-7B (4)58.679.535.259.363.867.1-63.875.7
Qwen-VL-7B-Chat (4)56.678.238.957.561.568.2-61.570.5
LLaVA1.5-7B (49)-78.550.062.0-66.8-58.2-
LLaVA1.5-13B (49)-80.053.663.3-71.6-61.3-
SPHINX62.178.139.962.658.569.350.451.666.0
SPHINX-1k62.280.246.862.965.469.152.758.870.0
SPHINX-2k62.680.744.963.157.170.650.561.267.8
", + "image_path": "213e9e0329e3906acbe654d23be34a860c038de2c64cdeebd6a964443bac6ad1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 413, + 479, + 433 + ], + "lines": [ + { + "bbox": [ + 132, + 413, + 479, + 433 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 479, + 433 + ], + "type": "text", + "content": "Table 2: Comparison with state-of-the-art methods on classical VQA benchmarks." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 472, + 488, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 488, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 488, + 579 + ], + "type": "text", + "content": "comparative results across all benchmarks. Notably, SPHINX-1k and SPHINX-2k significantly outperform SPHINX on VQAv2 and text-oriented VQA which demand fine-grained visual information, showcasing the effectiveness of our scale mix-up approach for achieving high resolution without relying on a visual encoder trained specifically on high-resolution images. Although the performances of SPHINX on text-oriented VQA surpass strong baselines such as BLIP-2 and InstructBLIP, it falls behind Qwen-VL-7B, which we conjecture is due to the lack of text-related pre-training data. Introducing more such datasets for training may resolve the issue in the future." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": "Visual grounding. Table 3 evaluates SPHINX on REferring Comprehension (REC) benchmarks with RefCOCO (34), RefCOCO+ (57), and RefCOCOg (57) datasets. SPHINX outperforms most state-of-the-art models, including specialist model G-DINO-L (51) and other visual-language generalist models. Compared to a recent strong baseline Qwen-VL-7B (4), which also leverages the large language model for visual understanding, our model still achieves better results across all splits by a large margin. Moreover, SPHINX-1k and SPHINX-2k" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "type": "text", + "content": "A Mixer of Weights, Visual Embeddings and Image Scales for MLLM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 151, + 114, + 462, + 300 + ], + "blocks": [ + { + "bbox": [ + 151, + 114, + 462, + 300 + ], + "lines": [ + { + "bbox": [ + 151, + 114, + 462, + 300 + ], + "spans": [ + { + "bbox": [ + 151, + 114, + 462, + 300 + ], + "type": "table", + "html": "
MethodsRefCOCO+RefCOCORefCOCOg
valtest-Atest-Bvaltest-Atest-Bval-utest-u
Specialist models
UNINEXT (99)85.2489.6379.7992.6494.3391.4688.7389.37
G-DINO-L (51)82.7588.9575.9290.5693.1988.2486.1387.02
Generalist models
VisionLLM-H (92)----86.70---
OFA-L (91)68.2976.0061.7579.9683.6776.3967.5767.58
Shikra 7B (10)81.6087.3672.1287.0190.6180.2482.2782.19
Shikra 13B (10)82.8987.7974.4187.8391.1181.8182.6483.16
MiniGPT-v2 7B (9)79.9785.1274.4588.6991.6585.3384.4484.66
MiniGPT-v2-7B-Chat (9)79.5885.5273.3288.0691.2984.3084.1984.31
Qwen-VL-7B (4)83.1288.2577.2189.3692.2685.3485.5885.48
Qwen-VL-7B-Chat (4)82.8288.5976.7988.5592.2784.5185.9686.32
SPHINX82.7787.2976.8589.1591.3785.1384.8783.65
SPHINX-1k86.6491.0880.3591.0592.6586.5688.1988.35
SPHINX-2k85.5190.6280.4591.1092.8887.0788.0788.65
", + "image_path": "885c4c1687d4e89ba7d06026b4a57337d420bf87901ff91e1fe87d206413f55e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 136, + 346, + 479, + 428 + ], + "blocks": [ + { + "bbox": [ + 132, + 302, + 481, + 324 + ], + "lines": [ + { + "bbox": [ + 132, + 302, + 481, + 324 + ], + "spans": [ + { + "bbox": [ + 132, + 302, + 481, + 324 + ], + "type": "text", + "content": "Table 3: Performance comparisons (Top-1 Accuracy@0.5) on the referring expression comprehension task." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 346, + 479, + 428 + ], + "lines": [ + { + "bbox": [ + 136, + 346, + 479, + 428 + ], + "spans": [ + { + "bbox": [ + 136, + 346, + 479, + 428 + ], + "type": "table", + "html": "
patch feat.global feat.MMEGeneral VQAText-oriented VQARefCOCO+RefCOCORefCOCOg
Conv NeXtCLIP LargeDINO v2Q-FormerPCOK VQAViz WizGQAVSRText VQAOCR VQAvaltest-Atest-Bvaltest-Atest-Bvaltest
✓✓✓1392.9299.260.540.359.853.050.461.780.5286.2075.4187.8590.7383.8282.7783.10
✓✓1428.4287.160.347.959.151.649.561.674.7982.3067.4782.1787.8576.7076.4876.20
✓✓✓1246.7267.560.946.160.055.251.365.982.1486.5575.3788.7390.6983.7084.1283.22
✓✓✓1371.7298.959.740.259.751.950.667.079.6185.8173.2386.0289.8080.6581.0081.58
✓✓✓1379.2274.660.547.759.952.251.666.781.0886.5175.0087.7990.9482.5182.1182.14
✓1364.0263.960.740.158.853.450.464.837.2242.4532.0746.5250.2442.6839.8840.79
✓✓✓✓1476.1322.262.139.962.658.551.666.082.7787.2876.8589.1491.3785.1284.8683.64
", + "image_path": "e524ddd56c8f032e3352abab11ad43d06ab8cc4facb0ca4da72aab21a8ca3fd2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 430, + 481, + 451 + ], + "lines": [ + { + "bbox": [ + 132, + 430, + 481, + 451 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 481, + 451 + ], + "type": "text", + "content": "Table 4: SPHINX visual embedding ablation study. The best result in each column is marked bold and the significantly worse results are marked red." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 487, + 481, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 481, + 536 + ], + "type": "text", + "content": "enable the use of high-resolution input images, leading to consecutive improvement over SPHINX and narrowing down the gap to the strong specialist model UNINEXT, which adopts a larger input image size. These results demonstrate the competitive capability of SPHINX for visual grounding." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 557, + 236, + 569 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 236, + 569 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 236, + 569 + ], + "type": "text", + "content": "4.3 Ablation study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "On the influence of image-text joint pre-training. We show in Figure 4 the effects of jointly training with a text-only dataset (i.e., RefinedWeb) in addition to the image-captioning dataset (i.e., LAION-400M). For comparison, we run an experiment where only LAION-400M is used for training, and observe the text generation loss and image captioning loss as a measurement of its text and image modeling capability, respectively. The text generation loss gradually grows if the model is not trained with RefinedWeb, showing that our joint-training scheme" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "text", + "content": "Z. Lin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 114, + 479, + 178 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 479, + 178 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 479, + 178 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 479, + 178 + ], + "type": "table", + "html": "
viewsMMEGeneral VQAText-oriented VQARefCOCO+RefCOCORefCOCOg
low. sub-res. imagesPCOK VQAViz WizGQAVSRText VQAOCR VQAvaltest-Atest-Bvaltest-Atest-Bvaltest
✓1476.1322.262.139.962.658.551.666.082.7787.2876.8589.1491.3785.1284.8683.64
✓1461.9255.361.846.162.054.057.569.169.8677.4562.7871.7277.0964.7469.1570.89
✓✓1560.2310.062.246.862.965.458.870.086.6491.0880.3591.0592.6586.5688.1988.35
", + "image_path": "ede0b422c28622221e07117d519dc8672c092f629d6136092c21b0b8736a5f14.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 178, + 481, + 222 + ], + "lines": [ + { + "bbox": [ + 132, + 178, + 481, + 222 + ], + "spans": [ + { + "bbox": [ + 132, + 178, + 481, + 222 + ], + "type": "text", + "content": "Table 5: Ablation study on mixed scales. For comparison, we remove either the low-resolution view or the sub-image views from SPHINX-1k (note that SPHINX-1k without sub-images is equivalent to SPHINX). The best results are marked bold and the second-best results are marked with underlines." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 257, + 479, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 257, + 479, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 257, + 479, + 281 + ], + "type": "text", + "content": "is effective in preserving the text reasoning capability while adapting to cross-modal understanding." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 281, + 480, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 480, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 480, + 389 + ], + "type": "text", + "content": "On the influence of image-text joint pre-training. We show in Figure 4 the effects of jointly training with a text-only dataset (i.e., RefinedWeb) in addition to the image captioning dataset (i.e., LAION-400M). For comparison, we run an experiment where the model is trained only with LAION-400M and without RefinedWeb, and observe the text generation loss and image captioning loss as a measurement of its text and image modeling capability, respectively. The text generation loss gradually grows if the model is not trained with RefinedWeb, showing that our joint-training scheme is effective in preserving the text reasoning capability while adapting to cross-modal understanding." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 389, + 481, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 389, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 481, + 568 + ], + "type": "text", + "content": "On the influence of mixed visual embeddings. To validate the effectiveness of visual embedding mix-up, we conducted experiments on SPHINX employing different combinations of visual encoders. The results are presented in Table 4. Interestingly, mixing the encoders exhibits different behaviors on different types of tasks: The mixture of all image encoders do not steadily outperforms its subsets on the VQA benchmarks, which we conjecture is because some datasets are either saturated or out-of-domain for some encoders, while we have noticed a huge gain on visual grounding tasks with the mixture of encoders compared to using only the global features (Q-Former in our case), certifying the importance of spatially-structured features for such fine-grained tasks. We also observed that DINOv2, which is a self-supervised encoder and complementary to the remaining image-text-based encoders, tend to give better results on the visual grounding tsks. As adding visual encoders brings negligible overhead, as analyzed in Sec. 3.3, we choose to use all the encoders by default as they obtain the best performance on most tests." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "On the influence of mixed scales. To give some insights into our mixed-scale strategy, we perform multi-task fine-tuning with SPHINX-1k, but remove either the low-resolution view or the sub-images to see their respective effects on various benchmarks. The results in Table 5 demonstrate the importance of both types of views: The combination of the two obtains the best results on most benchmarks. Our observation also highlights the critical role of the low-resolution view: Using only the sub-images, the model under-performs the SPHINX baseline on MME and referring comprehension despite the much longer input se" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "type": "text", + "content": "A Mixer of Weights, Visual Embeddings and Image Scales for MLLM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "type": "text", + "content": "quence. This emphasizes the importance of the low-resolution view as a global representation which helps glue together the detailed information captured in the sub-images." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 205, + 175, + 410, + 287 + ], + "blocks": [ + { + "bbox": [ + 205, + 175, + 410, + 287 + ], + "lines": [ + { + "bbox": [ + 205, + 175, + 410, + 287 + ], + "spans": [ + { + "bbox": [ + 205, + 175, + 410, + 287 + ], + "type": "image", + "image_path": "ff3eb064e3f1f88ecc3a2d26faca562b386a172f29729c3af84f83dadc405c6d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 298, + 482, + 333 + ], + "lines": [ + { + "bbox": [ + 130, + 298, + 482, + 333 + ], + "spans": [ + { + "bbox": [ + 130, + 298, + 482, + 333 + ], + "type": "text", + "content": "Fig. 4: Loss curve in the pre-training stage with and without optimizing on RefinedWeb. As we perform 1-epoch training on massive datasets, the curve can serve as both training and validation losses, as no sample is seen twice in the process." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 375, + 220, + 388 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 375, + 220, + 388 + ], + "spans": [ + { + "bbox": [ + 131, + 375, + 220, + 388 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 402, + 481, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 481, + 544 + ], + "type": "text", + "content": "In this paper, we propose SPHINX, a versatile multi-modal large language model (MLLM) with multi-purpose visual instruction-following capabilities. In our MLLM, we introduce a joint mixing of three different aspects: model weights of pre-trained LLMs by real-world and synthetic data, visual embeddings from different types of vision backbones, and multi-scale views of images with subimage embeddings. Via our proposed three-fold mixing strategy, SPHINX achieves impressive performance over a wide range of multi-modality evaluation benchmarks, and can serve as a strong vision generalist to tackle object detection, region-level captioning, and human pose estimation, etc. Our MLLM can also be integrated with other visual foundation models for wider functionalities, e.g., SAM (35) for language-referred segmentation and Stable Diffusion (73) for image editing." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 545, + 481, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 481, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 481, + 593 + ], + "type": "text", + "content": "Limitations. Limited by the context length of LLaMA, it is currently difficult for our model to process multiple images, interleaved images, videos or very large input images (e.g., beyond " + }, + { + "bbox": [ + 130, + 545, + 481, + 593 + ], + "type": "inline_equation", + "content": "1000 \\times 1000" + }, + { + "bbox": [ + 130, + 545, + 481, + 593 + ], + "type": "text", + "content": "). However, we believe this can be resolved in the near future with long-context LLMs under active research." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 481, + 666 + ], + "type": "text", + "content": "Potential societal impacts. Our method does not pose new negative societal impacts other than those existing in the pre-train models and training datasets themselves to our knowledge. With all base models and datasets used in the paper being publicly available, necessary censorship can be made based on specific ethnicity, regulations, etc. Our method has been shown to reduce hallucination, which could be generally considered a positive societal impacts." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "Z. Lin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "Acknowledgements. This project is funded in part by National Key R&D Program of China Project 2022ZD0161100 and 2022ZD0160102, by National Natural Science Foundation of China (Grant No. 62206272), by the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK, by Smart Traffic Fund PSRI/76/2311/PR and by RGC General Research Fund Project 14204021. Hongsheng Li is a PI of CPII under the InnoHK." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 261, + 246, + 353, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 246, + 353, + 264 + ], + "spans": [ + { + "bbox": [ + 261, + 246, + 353, + 264 + ], + "type": "text", + "content": "Bibliography" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 143, + 301, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 143, + 301, + 480, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 301, + 480, + 338 + ], + "spans": [ + { + "bbox": [ + 143, + 301, + 480, + 338 + ], + "type": "text", + "content": "[1] Agrawal, A., Lu, J., Antol, S., Mitchell, M., Zitnick, C.L., Parikh, D., Ba-tra, D.: Vqa: Visual question answering. International Journal of Computer Vision 123, 4 - 31 (2015)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 143, + 340, + 480, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 340, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 143, + 340, + 480, + 365 + ], + "type": "text", + "content": "[2] Aiello, E., Yu, L., Nie, Y., Aghajanyan, A., Oguz, B.: Jointly training large autoregressive multimodal models. arXiv preprint arXiv:2309.15564 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 366, + 480, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 366, + 480, + 414 + ], + "spans": [ + { + "bbox": [ + 143, + 366, + 480, + 414 + ], + "type": "text", + "content": "[3] Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems 35, 23716-23736 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 415, + 480, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 415, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 143, + 415, + 480, + 453 + ], + "type": "text", + "content": "[4] Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-vl: A frontier large vision-language model with versatile abilities. ArXiv abs/2308.12966 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 453, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 453, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 143, + 453, + 480, + 490 + ], + "type": "text", + "content": "[5] Bai, S., Yang, S., Bai, J., Wang, P., Zhang, X., Lin, J., Wang, X., Zhou, C., Zhou, J.: Touchstone: Evaluating vision-language models by language models (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 491, + 480, + 528 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 491, + 480, + 528 + ], + "spans": [ + { + "bbox": [ + 143, + 491, + 480, + 528 + ], + "type": "text", + "content": "[6] Bavishi, R., Elsen, E., Hawthorne, C., Nye, M., Odena, A., Somani, A., Taşirlar, S.: Introducing our multimodal models (2023), https://www. adept.ai/blog/fuyu-8b" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 529, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 529, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 143, + 529, + 480, + 578 + ], + "type": "text", + "content": "[7] Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 579, + 480, + 628 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 579, + 480, + 628 + ], + "spans": [ + { + "bbox": [ + 143, + 579, + 480, + 628 + ], + "type": "text", + "content": "[8] Chen, J., Guo, H., Yi, K., Li, B., Elhoseiny, M.: Visualgpt: Data-efficient adaptation of pretrained language models for image captioning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18030-18040 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 143, + 628, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 628, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 143, + 628, + 480, + 666 + ], + "type": "text", + "content": "[9] Chen, J., Li, D.Z.X.S.X., Zhang, Z.L.P., Xiong, R.K.V.C.Y., Elhoseiny, M.: Minigpt-v2: Large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478 (2023)" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 115, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 138, + 115, + 480, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 115, + 480, + 151 + ], + "spans": [ + { + "bbox": [ + 138, + 115, + 480, + 151 + ], + "type": "text", + "content": "[10] Chen, K., Zhang, Z., Zeng, W., Zhang, R., Zhu, F., Zhao, R.: Shikra: Unleashing multimodal lvm's referential dialogue magic. arXiv preprint arXiv:2306.15195 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 152, + 481, + 200 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 152, + 481, + 200 + ], + "spans": [ + { + "bbox": [ + 138, + 152, + 481, + 200 + ], + "type": "text", + "content": "[11] Chen, X., Wang, X., Changpinyo, S., Piergiovanni, A., Padlewski, P., Salz, D., Goodman, S., Grycner, A., Mustafa, B., Beyer, L., et al.: Pali: A jointly-scaled multilingual language-image model. arXiv preprint arXiv:2209.06794 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 201, + 480, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 201, + 480, + 250 + ], + "spans": [ + { + "bbox": [ + 138, + 201, + 480, + 250 + ], + "type": "text", + "content": "[12] Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., Stoica, I., Xing, E.P.: Vicuna: An open-source chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 138, + 201, + 480, + 250 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 138, + 201, + 480, + 250 + ], + "type": "text", + "content": " chatgpt quality. https://lmsys.org/blog/2023-03-30-vicuna/ (March 2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 251, + 480, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 251, + 480, + 275 + ], + "spans": [ + { + "bbox": [ + 138, + 251, + 480, + 275 + ], + "type": "text", + "content": "[13] Contributors, O.: Opencompass: A universal evaluation platform for foundation models. https://github.com/open-compass/opencompass (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 276, + 480, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 276, + 480, + 311 + ], + "spans": [ + { + "bbox": [ + 138, + 276, + 480, + 311 + ], + "type": "text", + "content": "[14] Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B.A., Fung, P., Hoi, S.C.H.: Instructclip: Towards general-purpose vision-language models with instruction tuning. ArXiv abs/2305.06500 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 312, + 480, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 312, + 480, + 348 + ], + "spans": [ + { + "bbox": [ + 138, + 312, + 480, + 348 + ], + "type": "text", + "content": "[15] Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 349, + 480, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 349, + 480, + 384 + ], + "spans": [ + { + "bbox": [ + 138, + 349, + 480, + 384 + ], + "type": "text", + "content": "[16] Dong, R., Han, C., Peng, Y., Qi, Z., Ge, Z., Yang, J., Zhao, L., Sun, J., Zhou, H., Wei, H., et al.: Dreamllm: Synergistic multimodal comprehension and creation. arXiv preprint arXiv:2309.11499 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 386, + 480, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 386, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 138, + 386, + 480, + 434 + ], + "type": "text", + "content": "[17] Dong, X., Bao, J., Zhang, T., Chen, D., Gu, S., Zhang, W., Yuan, L., Chen, D., Wen, F., Yu, N.: Clip itself is a strong fine-tuner: Achieving 85.7% and 88.0% top-1 accuracy with vit-b and vit-l onImagenet. arXiv preprint arXiv:2212.06138 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 434, + 480, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 434, + 480, + 482 + ], + "spans": [ + { + "bbox": [ + 138, + 434, + 480, + 482 + ], + "type": "text", + "content": "[18] Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 483, + 480, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 483, + 480, + 531 + ], + "spans": [ + { + "bbox": [ + 138, + 483, + 480, + 531 + ], + "type": "text", + "content": "[19] Douillard, A., Feng, Q., Rusu, A.A., Chhaparia, R., Donchev, Y., Kuncoro, A., Ranzato, M., Szlam, A., Shen, J.: Diloco: Distributed low-communication training of language models. arXiv preprint arXiv:2311.08105 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 532, + 480, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 532, + 480, + 580 + ], + "spans": [ + { + "bbox": [ + 138, + 532, + 480, + 580 + ], + "type": "text", + "content": "[20] Fu, C., Chen, P., Shen, Y., Qin, Y., Zhang, M., Lin, X., Qiu, Z., Lin, W., Yang, J., Zheng, X., et al.: Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 581, + 480, + 628 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 581, + 480, + 628 + ], + "spans": [ + { + "bbox": [ + 138, + 581, + 480, + 628 + ], + "type": "text", + "content": "[21] Fu, C., Chen, P., Shen, Y., Qin, Y., Zhang, M., Lin, X., Qiu, Z., Lin, W., Yang, J., Zheng, X., et al.: Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 629, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 629, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 138, + 629, + 480, + 666 + ], + "type": "text", + "content": "[22] Gao, P., Han, J., Zhang, R., Lin, Z., Geng, S., Zhou, A., Zhang, W., Lu, P., He, C., Yue, X., et al.: Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010 (2023)" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "text", + "content": "Z. Lin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 115, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 138, + 115, + 481, + 164 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 115, + 481, + 164 + ], + "spans": [ + { + "bbox": [ + 138, + 115, + 481, + 164 + ], + "type": "text", + "content": "[23] Girdhar, R., El-Nouby, A., Liu, Z., Singh, M., Alwala, K.V., Joulin, A., Misra, I.: Imagebind one embedding space to bind them all. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 15180-15190 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 165, + 381, + 177 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 165, + 381, + 177 + ], + "spans": [ + { + "bbox": [ + 138, + 165, + 381, + 177 + ], + "type": "text", + "content": "[24] Google: Bard. https://bard.google.com/ (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 139, + 178, + 481, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 178, + 481, + 225 + ], + "spans": [ + { + "bbox": [ + 139, + 178, + 481, + 225 + ], + "type": "text", + "content": "[25] Guo, Z., Zhang, R., Zhu, X., Tang, Y., Ma, X., Han, J., Chen, K., Gao, P., Li, X., Li, H., Heng, P.A.: Point-bind & point-llm: Aligning point cloud with multi-modality for 3d understanding, generation, and instruction following. ArXiv abs/2309.00615 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 139, + 227, + 481, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 227, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 139, + 227, + 481, + 262 + ], + "type": "text", + "content": "[26] Gupta, A., Dollar, P., Girshick, R.: Lvis: A dataset for large vocabulary instance segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5356-5364 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 139, + 263, + 481, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 263, + 481, + 311 + ], + "spans": [ + { + "bbox": [ + 139, + 263, + 481, + 311 + ], + "type": "text", + "content": "[27] Gurari, D., Li, Q., Stangl, A., Guo, A., Lin, C., Grauman, K., Luo, J., Bigham, J.P.: Vizwiz grand challenge: Answering visual questions from blind people. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition pp. 3608-3617 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 139, + 312, + 481, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 312, + 481, + 348 + ], + "spans": [ + { + "bbox": [ + 139, + 312, + 481, + 348 + ], + "type": "text", + "content": "[28] Han, J., Zhang, R., Shao, W., Gao, P., Xu, P., Xiao, H., Zhang, K., Liu, C., Wen, S., Guo, Z., et al.: Imagebind-llm: Multi-modality instruction tuning. arXiv preprint arXiv:2309.03905 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 139, + 349, + 481, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 349, + 481, + 384 + ], + "spans": [ + { + "bbox": [ + 139, + 349, + 481, + 384 + ], + "type": "text", + "content": "[29] He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 139, + 386, + 481, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 386, + 481, + 421 + ], + "spans": [ + { + "bbox": [ + 139, + 386, + 481, + 421 + ], + "type": "text", + "content": "[30] Hong, Y., Zhen, H., Chen, P., Zheng, S., Du, Y., Chen, Z., Gan, C.: 3d-dllm: Injecting the 3d world into large language models. arXiv preprint arXiv:2307.12981 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 139, + 422, + 481, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 422, + 481, + 458 + ], + "spans": [ + { + "bbox": [ + 139, + 422, + 481, + 458 + ], + "type": "text", + "content": "[31] Huang, C., Liu, Q., Lin, B.Y., Pang, T., Du, C., Lin, M.: Lorahub: Efficient cross-task generalization via dynamic lora composition. arXiv preprint arXiv:2307.13269 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 139, + 459, + 481, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 459, + 481, + 495 + ], + "spans": [ + { + "bbox": [ + 139, + 459, + 481, + 495 + ], + "type": "text", + "content": "[32] Huang, S., Dong, L., Wang, W., Hao, Y., Singhal, S., Ma, S., Lv, T., Cui, L., Mohammed, O.K., Liu, Q., et al.: Language is not all you need: Aligning perception with language models. arXiv preprint arXiv:2302.14045 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 139, + 496, + 481, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 496, + 481, + 543 + ], + "spans": [ + { + "bbox": [ + 139, + 496, + 481, + 543 + ], + "type": "text", + "content": "[33] Hudson, D.A., Manning, C.D.: Gqa: A new dataset for real-world visual reasoning and compositional question answering. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 6693-6702 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 139, + 544, + 481, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 544, + 481, + 581 + ], + "spans": [ + { + "bbox": [ + 139, + 544, + 481, + 581 + ], + "type": "text", + "content": "[34] Kazemzadeh, S., Ordonez, V., andre Matten, M., Berg, T.L.: Referitgame: Referring to objects in photographs of natural scenes. In: Conference on Empirical Methods in Natural Language Processing (2014)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 139, + 582, + 481, + 616 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 582, + 481, + 616 + ], + "spans": [ + { + "bbox": [ + 139, + 582, + 481, + 616 + ], + "type": "text", + "content": "[35] Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 139, + 618, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 618, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 139, + 618, + 481, + 666 + ], + "type": "text", + "content": "[36] Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., et al.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision 123, 32-73 (2017)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 165, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 90, + 447, + 102 + ], + "type": "text", + "content": "A Mixer of Weights, Visual Embeddings and Image Scales for MLLM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 115, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 138, + 115, + 480, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 115, + 480, + 152 + ], + "spans": [ + { + "bbox": [ + 138, + 115, + 480, + 152 + ], + "type": "text", + "content": "[37] Li, B., Zhang, Y., Chen, L., Wang, J., Pu, F., Yang, J., Li, C., Liu, Z.: Mimic-it: Multi-modal in-context instruction tuning. ArXiv abs/2306.05425 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 152, + 480, + 187 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 152, + 480, + 187 + ], + "spans": [ + { + "bbox": [ + 138, + 152, + 480, + 187 + ], + "type": "text", + "content": "[38] Li, B., Zhang, Y., Chen, L., Wang, J., Yang, J., Liu, Z.: Otter: A multimodal model with in-context instruction tuning. ArXiv abs/2305.03726 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 187, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 187, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 138, + 187, + 480, + 224 + ], + "type": "text", + "content": "[39] Li, B., Wang, R., Wang, G., Ge, Y., Ge, Y., Shan, Y.: Seed-bench: Benchmarking multimodal llms with generative comprehension. ArXiv abs/2307.16125 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 224, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 224, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 138, + 224, + 480, + 258 + ], + "type": "text", + "content": "[40] Li, H., Xu, Z., Taylor, G., Studer, C., Goldstein, T.: Visualizing the loss landscape of neural nets. Advances in neural information processing systems 31 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 259, + 480, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 259, + 480, + 295 + ], + "spans": [ + { + "bbox": [ + 138, + 259, + 480, + 295 + ], + "type": "text", + "content": "[41] Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 295, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 295, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 295, + 480, + 342 + ], + "type": "text", + "content": "[42] Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 342, + 480, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 342, + 480, + 378 + ], + "spans": [ + { + "bbox": [ + 138, + 342, + 480, + 378 + ], + "type": "text", + "content": "[43] Li, M., Gururangan, S., Dettmers, T., Lewis, M., Althoff, T., Smith, N.A., Zettlemoyer, L.: Branch-train-merge: Embarrassingly parallel training of expert language models. arXiv preprint arXiv:2208.03306 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 379, + 480, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 379, + 480, + 415 + ], + "spans": [ + { + "bbox": [ + 138, + 379, + 480, + 415 + ], + "type": "text", + "content": "[44] Li, Y., Mao, H., Girshick, R., He, K.: Exploring plain vision transformer backbones for object detection. In: European Conference on Computer Vision. pp. 280-296. Springer (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 139, + 415, + 480, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 415, + 480, + 450 + ], + "spans": [ + { + "bbox": [ + 139, + 415, + 480, + 450 + ], + "type": "text", + "content": "[45] Li, Y., Du, Y., Zhou, K., Wang, J., Zhao, W.X., Wen, J.R.: Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 139, + 450, + 480, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 450, + 480, + 498 + ], + "spans": [ + { + "bbox": [ + 139, + 450, + 480, + 498 + ], + "type": "text", + "content": "[46] Lin, T.Y., Dollar, P., Girshick, R., He, K., Hariharan, B., Belongie, S.: Feature pyramid networks for object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2117-2125 (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 139, + 498, + 480, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 498, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 139, + 498, + 480, + 557 + ], + "type": "text", + "content": "[47] Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dolar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 139, + 558, + 480, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 558, + 480, + 581 + ], + "spans": [ + { + "bbox": [ + 139, + 558, + 480, + 581 + ], + "type": "text", + "content": "[48] Liu, F., Emerson, G.E.T., Collier, N.: Visual spatial reasoning. Transactions of the Association for Computational Linguistics (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 139, + 582, + 480, + 605 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 582, + 480, + 605 + ], + "spans": [ + { + "bbox": [ + 139, + 582, + 480, + 605 + ], + "type": "text", + "content": "[49] Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. ArXiv abs/2310.03744 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 139, + 605, + 480, + 628 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 605, + 480, + 628 + ], + "spans": [ + { + "bbox": [ + 139, + 605, + 480, + 628 + ], + "type": "text", + "content": "[50] Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv preprint arXiv:2304.08485 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 139, + 629, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 629, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 139, + 629, + 480, + 666 + ], + "type": "text", + "content": "[51] Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., yue Li, C., Yang, J., Su, H., Zhu, J.J., Zhang, L.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. ArXiv abs/2303.05499 (2023)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "text", + "content": "Z. Lin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 115, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 138, + 115, + 481, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 115, + 481, + 152 + ], + "spans": [ + { + "bbox": [ + 138, + 115, + 481, + 152 + ], + "type": "text", + "content": "[52] Liu, Y., Duan, H., Zhang, Y., Li, B., Zhang, S., Zhao, W., Yuan, Y., Wang, J., He, C., Liu, Z., et al.: Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 153, + 481, + 189 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 153, + 481, + 189 + ], + "spans": [ + { + "bbox": [ + 138, + 153, + 481, + 189 + ], + "type": "text", + "content": "[53] Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3431-3440 (2015)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 190, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 190, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 138, + 190, + 481, + 237 + ], + "type": "text", + "content": "[54] Lu, P., Bansal, H., Xia, T., Liu, J., yue Li, C., Hajishirzi, H., Cheng, H., Chang, K.W., Galley, M., Gao, J.: Mathvista: Evaluating math reasoning in visual contexts with gpt-4v, bard, and other large multimodal models. ArXiv abs/2310.02255 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 238, + 481, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 238, + 481, + 286 + ], + "spans": [ + { + "bbox": [ + 138, + 238, + 481, + 286 + ], + "type": "text", + "content": "[55] Lu, P., Mishra, S., Xia, T., Qiu, L., Chang, K.W., Zhu, S.C., Tafjord, O., Clark, P., Kalyan, A.: Learn to explain: Multimodal reasoning via thought chains for science question answering. In: The 36th Conference on Neural Information Processing Systems (NeurIPS) (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 287, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 287, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 138, + 287, + 481, + 346 + ], + "type": "text", + "content": "[56] Lu, P., Qiu, L., Chen, J., Xia, T., Zhao, Y., Zhang, W., Yu, Z., Liang, X., Zhu, S.C.: Iconqa: A new benchmark for abstract diagram understanding and visual language reasoning. In: The 35th Conference on Neural Information Processing Systems (NeurIPS) Track on Datasets and Benchmarks (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 347, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 347, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 138, + 347, + 481, + 396 + ], + "type": "text", + "content": "[57] Mao, J., Huang, J., Toshev, A., Camburu, O.M., Yuille, A.L., Murphy, K.P.: Generation and comprehension of unambiguous object descriptions. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) pp. 11-20 (2015)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 396, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 396, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 138, + 396, + 481, + 445 + ], + "type": "text", + "content": "[58] Marino, K., Rastegari, M., Farhadi, A., Mottaghi, R.: Ok-vqa: A visual question answering benchmark requiring external knowledge. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 3190-3199 (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 139, + 445, + 481, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 445, + 481, + 481 + ], + "spans": [ + { + "bbox": [ + 139, + 445, + 481, + 481 + ], + "type": "text", + "content": "[59] McMahan, B., Moore, E., Ramage, D., Hampson, S., y Arcas, B.A.: Communication-efficient learning of deep networks from decentralized data. In: Artificial intelligence and statistics. pp. 1273-1282. PMLR (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 139, + 482, + 481, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 482, + 481, + 518 + ], + "spans": [ + { + "bbox": [ + 139, + 482, + 481, + 518 + ], + "type": "text", + "content": "[60] Mishra, A., Shekhar, S., Singh, A.K., Chakraborty, A.: Ocr-vqa: Visual question answering by reading text in images. 2019 International Conference on Document Analysis and Recognition (ICDAR) pp. 947-952 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 139, + 519, + 394, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 519, + 394, + 531 + ], + "spans": [ + { + "bbox": [ + 139, + 519, + 394, + 531 + ], + "type": "text", + "content": "[61] OpenAI: Chatgpt. https://chat.openai.com (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 139, + 532, + 443, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 532, + 443, + 544 + ], + "spans": [ + { + "bbox": [ + 139, + 532, + 443, + 544 + ], + "type": "text", + "content": "[62] OpenAI: Gpt-4 technical report. ArXiv abs/2303.08774 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 139, + 544, + 481, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 544, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 139, + 544, + 481, + 568 + ], + "type": "text", + "content": "[63] OpenAI: Vision - openai api. https://platform.openai.com/docs/guides/vision (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 139, + 569, + 481, + 616 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 569, + 481, + 616 + ], + "spans": [ + { + "bbox": [ + 139, + 569, + 481, + 616 + ], + "type": "text", + "content": "[64] Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: DiNov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 139, + 617, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 139, + 617, + 481, + 666 + ], + "type": "text", + "content": "[65] Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems 35, 27730-27744 (2022)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 165, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 90, + 447, + 102 + ], + "type": "text", + "content": "A Mixer of Weights, Visual Embeddings and Image Scales for MLLM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 115, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 115, + 481, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 115, + 481, + 163 + ], + "spans": [ + { + "bbox": [ + 138, + 115, + 481, + 163 + ], + "type": "text", + "content": "[66] Penedo, G., Malartic, Q., Hesslow, D., Cojocaru, R., Cappelli, A., Alobeidli, H., Pannier, B., Almazrouei, E., Launay, J.: The refined web dataset for falcon llm: outperforming curated corpora with web data, and web data only. arXiv preprint arXiv:2306.01116 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 164, + 481, + 187 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 164, + 481, + 187 + ], + "spans": [ + { + "bbox": [ + 138, + 164, + 481, + 187 + ], + "type": "text", + "content": "[67] Peng, B., Li, C., He, P., Galley, M., Gao, J.: Instruction tuning with gpt-4. arXiv preprint arXiv:2304.03277 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 139, + 189, + 481, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 189, + 481, + 223 + ], + "spans": [ + { + "bbox": [ + 139, + 189, + 481, + 223 + ], + "type": "text", + "content": "[68] Peng, Z., Wang, W., Dong, L., Hao, Y., Huang, S., Ma, S., Wei, F.: Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 139, + 224, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 224, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 139, + 224, + 481, + 270 + ], + "type": "text", + "content": "[69] Plummer, B.A., Wang, L., Cervantes, C.M., Caicedo, J.C., Hockenmaier, J., Lazebnik, S.: Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. International Journal of Computer Vision 123, 74 - 93 (2015)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 139, + 272, + 481, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 272, + 481, + 319 + ], + "spans": [ + { + "bbox": [ + 139, + 272, + 481, + 319 + ], + "type": "text", + "content": "[70] Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 139, + 320, + 481, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 320, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 139, + 320, + 481, + 342 + ], + "type": "text", + "content": "[71] Radford, A., Narasimhan, K.: Improving language understanding by generative pre-training (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 139, + 344, + 481, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 344, + 481, + 379 + ], + "spans": [ + { + "bbox": [ + 139, + 344, + 481, + 379 + ], + "type": "text", + "content": "[72] Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners. OpenAI blog 1(8), 9 (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 139, + 380, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 380, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 139, + 380, + 481, + 402 + ], + "type": "text", + "content": "[73] Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 139, + 403, + 481, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 403, + 481, + 450 + ], + "spans": [ + { + "bbox": [ + 139, + 403, + 481, + 450 + ], + "type": "text", + "content": "[74] Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al.: Imagenet large scale visual recognition challenge. International journal of computer vision 115, 211-252 (2015)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 139, + 452, + 481, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 452, + 481, + 474 + ], + "spans": [ + { + "bbox": [ + 139, + 452, + 481, + 474 + ], + "type": "text", + "content": "[75] Schuhmann, C., Köpf, A., Vencu, R., Coombes, T., Beaumont, R.: Laion-coco. https://laion.ai/blog/laion-coco/ (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 139, + 475, + 481, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 475, + 481, + 521 + ], + "spans": [ + { + "bbox": [ + 139, + 475, + 481, + 521 + ], + "type": "text", + "content": "[76] Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 139, + 523, + 481, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 523, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 139, + 523, + 481, + 557 + ], + "type": "text", + "content": "[77] Schwenk, D., Khandelwal, A., Clark, C., Marino, K., Mottaghi, R.: Akvqa: A benchmark for visual question answering using world knowledge. In: European Conference on Computer Vision (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 139, + 559, + 481, + 594 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 559, + 481, + 594 + ], + "spans": [ + { + "bbox": [ + 139, + 559, + 481, + 594 + ], + "type": "text", + "content": "[78] Shao, W., Hu, Y., Gao, P., Lei, M., Zhang, K., Meng, F., Xu, P., Huang, S., Li, H., Qiao, Y., et al.: Tiny lvlm-ehub: Early multimodal experiments with bard. arXiv preprint arXiv:2308.03729 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 139, + 594, + 397, + 606 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 594, + 397, + 606 + ], + "spans": [ + { + "bbox": [ + 139, + 594, + 397, + 606 + ], + "type": "text", + "content": "[79] ShareGPT: Sharegpt. https://sharegpt.com/ (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 139, + 607, + 481, + 640 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 607, + 481, + 640 + ], + "spans": [ + { + "bbox": [ + 139, + 607, + 481, + 640 + ], + "type": "text", + "content": "[80] Shukor, M., Dancette, C., Rame, A., Cord, M.: Univalent: Unified model for image, video, audio and language tasks. Transactions on Machine Learning Research Journal (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 139, + 642, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 642, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 139, + 642, + 481, + 665 + ], + "type": "text", + "content": "[81] Sidorov, O., Hu, R., Rohrbach, M., Singh, A.: Textcaps: a dataset for image captioning with reading comprehension. ArXiv abs/2003.12462 (2020)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "20" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "text", + "content": "Z. Lin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 19 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 115, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 138, + 115, + 480, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 115, + 480, + 163 + ], + "spans": [ + { + "bbox": [ + 138, + 115, + 480, + 163 + ], + "type": "text", + "content": "[82] Singh, A., Natarajan, V., Shah, M., Jiang, Y., Chen, X., Batra, D., Parikh, D., Rohrbach, M.: Towards vqa models that can read. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 8309-8318 (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 165, + 480, + 189 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 165, + 480, + 189 + ], + "spans": [ + { + "bbox": [ + 138, + 165, + 480, + 189 + ], + "type": "text", + "content": "[83] Su, Y., Lan, T., Li, H., Xu, J., Wang, Y., Cai, D.: Pandagpt: One model to instruction-follow them all. ArXiv abs/2305.16355 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 139, + 190, + 480, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 190, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 139, + 190, + 480, + 225 + ], + "type": "text", + "content": "[84] Sung, Y.L., Li, L., Lin, K., Gan, Z., Bansal, M., Wang, L.: An empirical study of multimodal model merging. arXiv preprint arXiv:2304.14933 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 139, + 227, + 480, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 227, + 480, + 274 + ], + "spans": [ + { + "bbox": [ + 139, + 227, + 480, + 274 + ], + "type": "text", + "content": "[85] Suvorov, R., Logacheva, E., Mashikhin, A., Remizova, A., Ashukha, A., Silvestrov, A., Kong, N., Goka, H., Park, K., Lempitsky, V.: Resolution-robust large mask inpainting with fourier convolutions. arXiv preprint arXiv:2109.07161 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 139, + 276, + 480, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 276, + 480, + 312 + ], + "spans": [ + { + "bbox": [ + 139, + 276, + 480, + 312 + ], + "type": "text", + "content": "[86] Taori, R., Gulrajani, I., Zhang, T., Dubois, Y., Li, X., Guestrin, C., Liang, P., Hashimoto, T.B.: Stanford alpaca: An instruction-following llama model. https://github.com/tatsu-lab/stanford_alpaca (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 139, + 313, + 480, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 313, + 480, + 360 + ], + "spans": [ + { + "bbox": [ + 139, + 313, + 480, + 360 + ], + "type": "text", + "content": "[87] Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., Lample, G.: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 139, + 361, + 480, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 361, + 480, + 408 + ], + "spans": [ + { + "bbox": [ + 139, + 361, + 480, + 408 + ], + "type": "text", + "content": "[88] Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 139, + 410, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 410, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 139, + 410, + 480, + 445 + ], + "type": "text", + "content": "[89] Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 139, + 447, + 480, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 447, + 480, + 482 + ], + "spans": [ + { + "bbox": [ + 139, + 447, + 480, + 482 + ], + "type": "text", + "content": "[90] Wang, K., Ren, H., Zhou, A., Lu, Z., Luo, S., Shi, W., Zhang, R., Song, L., Zhan, M., Li, H.: Mathcoder: Seamless code integration in llms for enhanced mathematical reasoning. arXiv preprint arXiv:2310.03731 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 139, + 483, + 480, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 483, + 480, + 531 + ], + "spans": [ + { + "bbox": [ + 139, + 483, + 480, + 531 + ], + "type": "text", + "content": "[91] Wang, P., Yang, A., Men, R., Lin, J., Bai, S., Li, Z., Ma, J., Zhou, C., Zhou, J., Yang, H.: Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In: International Conference on Machine Learning. pp. 23318-23340. PMLR (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 139, + 533, + 480, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 533, + 480, + 580 + ], + "spans": [ + { + "bbox": [ + 139, + 533, + 480, + 580 + ], + "type": "text", + "content": "[92] Wang, W., Chen, Z., Chen, X., Wu, J., Zhu, X., Zeng, G., Luo, P., Lu, T., Zhou, J., Qiao, Y., et al.: Visionllm: Large language model is also an open-ended decoder for vision-centric tasks. arXiv preprint arXiv:2305.11175 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 139, + 582, + 480, + 616 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 582, + 480, + 616 + ], + "spans": [ + { + "bbox": [ + 139, + 582, + 480, + 616 + ], + "type": "text", + "content": "[93] Wen, S., Fang, G., Zhang, R., Gao, P., Dong, H., Metaxas, D.: Improving compositional text-to-image generation with large vision-language models. arXiv preprint arXiv:2310.06311 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 139, + 618, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 618, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 139, + 618, + 480, + 665 + ], + "type": "text", + "content": "[94] Woo, S., Debnath, S., Hu, R., Chen, X., Liu, Z., Kweon, I.S., Xie, S.: Convnext v2: Co-designing and scaling convnets with masked autoencoders. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16133-16142 (2023)" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 165, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 90, + 447, + 102 + ], + "type": "text", + "content": "A Mixer of Weights, Visual Embeddings and Image Scales for MLLM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "21" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 20 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 115, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 138, + 115, + 481, + 176 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 115, + 481, + 176 + ], + "spans": [ + { + "bbox": [ + 138, + 115, + 481, + 176 + ], + "type": "text", + "content": "[95] Wortsman, M., Ilharco, G., Gadre, S.Y., Roelofs, R., Gontijo-Lopes, R., Morcos, A.S., Namkoong, H., Farhadi, A., Carmon, Y., Kornblith, S., et al.: Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In: International Conference on Machine Learning. pp. 23965-23998. PMLR (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 177, + 481, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 177, + 481, + 212 + ], + "spans": [ + { + "bbox": [ + 138, + 177, + 481, + 212 + ], + "type": "text", + "content": "[96] Wu, C., Yin, S., Qi, W., Wang, X., Tang, Z., Duan, N.: Visual chatgpt: Talking, drawing and editing with visual foundation models. arXiv preprint arXiv:2303.04671 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 213, + 481, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 213, + 481, + 261 + ], + "spans": [ + { + "bbox": [ + 138, + 213, + 481, + 261 + ], + "type": "text", + "content": "[97] Wu, C., Wang, T., Ge, Y., Lu, Z., Zhou, R., Shan, Y., Luo, P.: " + }, + { + "bbox": [ + 138, + 213, + 481, + 261 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 138, + 213, + 481, + 261 + ], + "type": "text", + "content": "-tuning: Transferring multimodal foundation models with optimal multi-task interpolation. In: International Conference on Machine Learning. pp. 37713-37727. PMLR (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 262, + 481, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 262, + 481, + 298 + ], + "spans": [ + { + "bbox": [ + 138, + 262, + 481, + 298 + ], + "type": "text", + "content": "[98] Xu, R., Wang, X., Wang, T., Chen, Y., Pang, J., Lin, D.: Pointllm: Empowering large language models to understand point clouds. ArXiv abs/2308.16911 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 299, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 299, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 138, + 299, + 481, + 346 + ], + "type": "text", + "content": "[99] Yan, B., Jiang, Y., Wu, J., Wang, D., Luo, P., Yuan, Z., Lu, H.: Universal instance perception as object discovery and retrieval. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 15325-15336 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 348, + 481, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 348, + 481, + 384 + ], + "spans": [ + { + "bbox": [ + 133, + 348, + 481, + 384 + ], + "type": "text", + "content": "[100] Yang, E., Wang, Z., Shen, L., Liu, S., Guo, G., Wang, X., Tao, D.: Adamerging: Adaptive model merging for multi-task learning. arXiv preprint arXiv:2310.02575 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 384, + 481, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 384, + 481, + 421 + ], + "spans": [ + { + "bbox": [ + 133, + 384, + 481, + 421 + ], + "type": "text", + "content": "[101] Yang, Z., Li, L., Wang, J., Lin, K., Azarnasab, E., Ahmed, F., Liu, Z., Liu, C., Zeng, M., Wang, L.: Mm-react: Prompting chatgpt for multimodal reasoning and action. arXiv preprint arXiv:2303.11381 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 422, + 481, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 422, + 481, + 457 + ], + "spans": [ + { + "bbox": [ + 133, + 422, + 481, + 457 + ], + "type": "text", + "content": "[102] Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 458, + 481, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 458, + 481, + 482 + ], + "spans": [ + { + "bbox": [ + 133, + 458, + 481, + 482 + ], + "type": "text", + "content": "[103] Yu, F., Koltun, V.: Multi-scale context aggregation by dilated convolutions. arXiv preprint arXiv:1511.07122 (2015)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 483, + 481, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 483, + 481, + 518 + ], + "spans": [ + { + "bbox": [ + 133, + 483, + 481, + 518 + ], + "type": "text", + "content": "[104] Yu, T., Feng, R., Feng, R., Liu, J., Jin, X., Zeng, W., Chen, Z.: Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 520, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 520, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 133, + 520, + 481, + 555 + ], + "type": "text", + "content": "[105] Yu, W., Yang, Z., Li, L., Wang, J., Lin, K., Liu, Z., Wang, X., Wang, L.: Mm-vet: Evaluating large multimodal models for integrated capabilities. ArXiv abs/2308.02490 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 556, + 481, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 556, + 481, + 592 + ], + "spans": [ + { + "bbox": [ + 133, + 556, + 481, + 592 + ], + "type": "text", + "content": "[106] Zhang, R., Han, J., Zhou, A., Hu, X., Yan, S., Lu, P., Li, H., Gao, P., Qiao, Y.: Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 593, + 481, + 628 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 593, + 481, + 628 + ], + "spans": [ + { + "bbox": [ + 133, + 593, + 481, + 628 + ], + "type": "text", + "content": "[107] Zhang, R., Hu, X., Li, B., Huang, S., Deng, H., Li, H., Qiao, Y., Gao, P.: Prompt, generate, then cache: Cascade of foundation models makes strong few-shot learners. CVPR 2023 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 629, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 629, + 481, + 666 + ], + "type": "text", + "content": "[108] Zhang, S., Roller, S., Goyal, N., Artetxe, M., Chen, M., Chen, S., Dewan, C., Diab, M., Li, X., Lin, X.V., et al.: Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068 (2022)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "22" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "text", + "content": "Z. Lin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 21 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/SPIN_ Hierarchical Segmentation with Subpart Granularity in Natural Images/88afeb42-c98d-4785-a2a2-08cde2fa67ab_content_list.json b/2024/SPIN_ Hierarchical Segmentation with Subpart Granularity in Natural Images/88afeb42-c98d-4785-a2a2-08cde2fa67ab_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b8877c8b1521642a825e045444261f0f3217e655 --- /dev/null +++ b/2024/SPIN_ Hierarchical Segmentation with Subpart Granularity in Natural Images/88afeb42-c98d-4785-a2a2-08cde2fa67ab_content_list.json @@ -0,0 +1,1754 @@ +[ + { + "type": "text", + "text": "SPIN: Hierarchical Segmentation with Subpart Granularity in Natural Images", + "text_level": 1, + "bbox": [ + 232, + 140, + 772, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Josh Myers-Dean $^{1}$ , Jarek Reynolds $^{1}$ , Brian Price $^{2}$ , Yifei Fan $^{2}$ , and Danna Gurari $^{1,3}$", + "bbox": [ + 259, + 212, + 743, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 University of Colorado Boulder, 2 Adobe, 3 University of Texas at Austin", + "bbox": [ + 246, + 253, + 759, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Hierarchical segmentation entails creating segmentations at varying levels of granularity. We introduce the first hierarchical semantic segmentation dataset with subpart annotations for natural images, which we call SPIN (SubPartImageNet). We also introduce two novel evaluation metrics to evaluate how well algorithms capture spatial and semantic relationships across hierarchical levels. We benchmark modern models across three different tasks and analyze their strengths and weaknesses across objects, parts, and subparts. To facilitate community-wide progress, we publicly release our dataset at https://joshmyersdean.github.io/spin/index.html.", + "bbox": [ + 259, + 308, + 743, + 449 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 474, + 375, + 491 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "When people discuss hierarchical image analysis tasks they are typically talking about one of two approaches: is-a relationships [27,28], which treat categories at different abstraction levels (e.g., a Subaru is a car), or is-part-of relationships, which concentrate on dividing objects into their constituent parts [13, 14, 19] (e.g., a door is part of a car). While the former has been widely explored, the latter—focusing on object decomposition—has received limited attention in computer vision research. Within segmentation research, the focus of this paper, is-part-of relationships have primarily been explored only for part-whole hierarchies, ignoring finer-grained details, such as subparts (i.e., parts of parts).", + "bbox": [ + 212, + 507, + 787, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Research to subpart-level segmentation granularity is hindered by a scarcity of data. While synthetic 3D datasets [14, 19, 35, 63] could be used to infer hierarchical segmentations to subpart granularity, it is well-known that models developed with synthetic data typically generalize poorly to natural images (i.e., images taken by a camera) [8]. This lack of annotated natural data has meant that the few models designed to generate subpart granularity segmentations [11, 52, 60] could only be evaluated qualitatively on a small number of examples (rather than quantitatively at scale).", + "bbox": [ + 212, + 643, + 787, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Our work facilitates the development of algorithms for subpart granularity in three key ways. First, we collect over 102,000 subpart segmentations across 203 diverse subpart categories to expand upon PartImageNet [17] and create the SubPartImageNet (SPIN) dataset—the first to offer subpart annotations for natural images. We release this dataset publicly to foster community progress.", + "bbox": [ + 212, + 765, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/430260e209de2ae28810f30abe916d3f821a96710724b757d244b2066cb0766e.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 313, + 218 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a9a305fa1738e86babc7dab2b438910eef6fd0c0fc3ac9d79eedbf22688291cd.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 313, + 145, + 406, + 218 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d9bd6d7ea9e6f65fb3986e731fb31ec67a4de54295a7844d9b3ec78976525854.jpg", + "image_caption": [ + "(c)", + "Fig. 1: Overview of the diversity of SPIN. Panels (a) and (b) depict subparts unique to specific object class members, such as a roll cage in a car and a shell in a turtle. Panels (c) and (d) illustrate the variability in the number of subparts per object of the same class, with examples of 13 and 6 subparts. Panels (e) and (f) highlight the disparity in image area coverage by different subparts, such as a bottle label (large) versus quadruped claws (tiny)." + ], + "image_footnote": [], + "bbox": [ + 408, + 145, + 500, + 218 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b63a593494b211050330e7ceab3339405c8447b513a34c88bb3c137b2b6ed62a.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 501, + 145, + 596, + 218 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0bdb3c52360b4667684fcb75b04100d30678e1454ab7d7a7fb22f7502d1b00cf.jpg", + "image_caption": [ + "(e)" + ], + "image_footnote": [], + "bbox": [ + 596, + 145, + 687, + 218 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/624f1cc272c0a9e7caea6832adeb66eb3860a38a51e6c2b77ac9b7337db41618.jpg", + "image_caption": [ + "(f)" + ], + "image_footnote": [], + "bbox": [ + 691, + 145, + 782, + 218 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Examples of annotated subparts are shown in Fig. 1. Second, we introduce two novel metrics to address a shortcoming of the prevailing evaluation methods for hierarchical segmentation, which only assess each granularity level in isolation. Our newly proposed metrics are intended to work alongside these traditional metrics (e.g., Intersection over Union) and assess how well models capture spatial and semantic relationships across hierarchical levels. Our third key contribution is that we benchmark modern models on SPIN across objects, parts, and subparts, thereby presenting the first comprehensive quantitative evaluation of subpart performance and key areas needed for future improvement.", + "bbox": [ + 212, + 356, + 787, + 492 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Advancing hierarchical segmentation to subpart granularity could significantly benefit both research and societal applications. One potential use is in generating more detailed image descriptions, such as for augmented reality experiences, answering visual questions, captioning images, or visual storytelling. Similarly, this work could be the foundation for individuals to (1) interactively learn how to speak about finer-grained entities for the first time (for children) or in a new language and (2) recall forgotten words, such as when facing temporary or chronic memory injuries or disabilities. More deeply nested, hierarchical segmentation representations can also enrich tactile visual discovery [25, 42], including for use in screen readers employed by visually impaired individuals (e.g., Apple's object-based exploration feature [1]). Similarly, we believe finer-grained information could facilitate improvements for related problems, including image/video retrieval, image/video editing, automatic magnification, and robotics.", + "bbox": [ + 212, + 493, + 787, + 690 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 714, + 397, + 729 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Datasets with Hierarchically Segmented Objects. Several datasets provide segmentations showing how objects are hierarchically decomposed into their recursively nested parts. A few do so without semantic labels specifying the categories of segmented content, including the pioneering Berkeley dataset published in 2001 [36] and the large-scale SA-1B dataset released in 2023 [22]. Our work, in contrast, focuses on hierarchically segmenting objects with category", + "bbox": [ + 212, + 748, + 787, + 842 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Myers-Dean et al.", + "bbox": [ + 271, + 114, + 392, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "labels. Already, numerous datasets semantically decompose objects and their parts in natural images [7, 14, 15, 17, 21, 31, 45, 50, 56, 61, 67-69] and synthetic images [5, 19, 20, 35, 39]. One dataset, (i.e., ADE20K [69]) even provides subpart annotations for $10\\%$ of its objects, however its lack of exhaustive annotations for objects and their associated parts [52] impedes it benefit for evaluating prediction models. Extending prior work, we introduce the first exhaustively labelled hierarchical semantic segmentation dataset with subpart annotations for natural images by extending one of the largest part-based datasets [17].", + "bbox": [ + 212, + 146, + 787, + 268 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Hierarchical Semantic Segmentation Evaluation. Most evaluation protocols for hierarchical segmentation algorithms assess each hierarchy level independently, such as by reporting for object and part categories separately their Intersection over Union scores [9, 23, 31, 32, 37, 40, 52, 55, 57, 58, 61, 65], Average Precision scores [45, 51], and Panoptic Quality scores [14, 29, 30, 52, 60]. The one exception is Hierarchical Panoptic Quality [52] which recursively measures the extent to which a predicted hierarchy is complete (e.g., a wheel is a part of a car). Complementing prior work, we introduce evaluation metrics for assessing how well a model captures spatial and semantic hierarchical relationships, such as if a part is perfectly contained in it's whole.", + "bbox": [ + 212, + 281, + 789, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Hierarchical Semantic Segmentation Algorithms. Thus far, the focus of methods that predict all entities in a hierarchical decomposition largely has centered on predicting just objects and parts [16, 51, 58], aligning with existing datasets which only recently increasingly have added part segmentations to object segmentations. Only a few works, HIPIE [60], VDT [11], Semantic-SAM [26], and ViRReq [52] $^{1}$ , examined predicting more layers of a hierarchy to also include subparts. However, these works relied on qualitative assessments for subpart predictions due to a lack of annotated datasets supporting quantitative assessments. Complementing prior work, we quantitatively evaluate HIPIE and over 20 model variants to show their performance on our new dataset challenge across all three layers of the hierarchy (i.e., object, parts, and subparts).", + "bbox": [ + 212, + 446, + 789, + 612 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 SPIN Dataset", + "text_level": 1, + "bbox": [ + 214, + 633, + 387, + 651 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We now introduce and characterize our new subpart semantic segmentation dataset that we call SPIN, which is short for SubPart ImageNet.", + "bbox": [ + 212, + 666, + 785, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1 Dataset Creation", + "text_level": 1, + "bbox": [ + 214, + 717, + 405, + 732 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Data Source. Our dataset extends an existing dataset that provides both object and associated part annotations: PartImageNet [17]. It includes 24,080 natural images across 158 ImageNet categories (112 non-rigid and 38 rigid, e.g. animals and vehicles respectively). The authors of PartImageNet used the", + "bbox": [ + 212, + 741, + 787, + 803 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "SPIN", + "bbox": [ + 692, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "1 No code is publicly available for VDT, and ViRReq does not offer complete code. At the time of writing, Semantic-SAM has not released their semantic prediction code.", + "bbox": [ + 217, + 810, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "WordNet [38] taxonomy to establish these categories and their 11 parent supercategories. The authors then identified 40 part categories pertinent to the supercategories and annotated these in all images. Every image in PartImageNet contains one semantically segmented object, and we used only those with at least one segmented part. Due to our limited annotation budget, we capped each supercategory at 1,200 images. This resulted in the following amounts per supercategory: 311 - Aeroplane, 483 - Bottle, 559 - Boat, 634 - Bicycle, and 1,200 each for Biped, Bird, Car, Fish, Quadruped, Reptile, and Snake.", + "bbox": [ + 212, + 146, + 787, + 268 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Subpart Category Selection. We identified 206 subpart categories to segment for 34 PartImageNet part categories that we identified as being decomposable into subparts. Of these, 168 subpart categories generalize across many part categories and 38 categories pertain to only a few. For example, in our dataset, mouths generalize while shells don't, with most reptiles having heads with mouths but only one type of reptile having a body with a shell (i.e., turtles). We show examples of specific subparts in Fig. 1(a-b). A histogram showing how many subparts are assigned to each of the 34 part categories is shown in Fig. 2, and a full list is provided in the supplementary materials.", + "bbox": [ + 212, + 286, + 787, + 421 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We created our subpart taxonomy through a multi-step process. We identified the general subpart categories by prompting GPT-4 [2] to list expected subparts for every object-part category in PartImageNet, and then three authors edited the list to exclude non-visible subparts (e.g., a skull). We identified specific subpart categories through manual inspection of at least 15 images per object category, excluding any categories that had ambiguous subpart decompositions; for example, it's not obvious to a lay person what boundaries to use when decomposing a tail into a tip, shaft, and base.", + "bbox": [ + 212, + 422, + 787, + 544 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Subpart Annotation Task Design. We created a task interface for collecting all subpart annotations. It presents each image-object pair alongside one of the object's nested parts, outlined with an overlaid polygon. For each part, the task", + "bbox": [ + 212, + 563, + 787, + 609 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/929d7239c260f06f0a3fd6cef87c8d6ccb03ae0f99ab0a0413ed4c1f6452ead6.jpg", + "image_caption": [ + "Fig. 2: Histogram the number of unique subpart category labels for each of the 34 part categories. (Aero=Aeroplane; Quad=Quadruped)" + ], + "image_footnote": [], + "bbox": [ + 215, + 681, + 785, + 797 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Myers-Dean et al.", + "bbox": [ + 271, + 114, + 392, + 128 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "2 We prompted GPT-4 with \"Please list the canonical subparts of a $<$ object $\\succ -\\text{part}>$ . Only include subparts that are clearly visible and recognizable to a layperson.\"", + "bbox": [ + 217, + 619, + 785, + 650 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/ef24e8a325b06588baa3f3e3f2f319c4836e6feb84a65f80ef1f2068d85f43ea.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
#Images#Object Cat#Part Cat#Sub-Part Cat
10,38711 (10,387)40 (29,818)203 (106,324)
", + "bbox": [ + 232, + 143, + 771, + 186 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1: SPIN Composition: Each column lists category counts alongside the total number of annotated entities within those categories. (Cat=Categories)", + "bbox": [ + 215, + 186, + 785, + 214 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "interface has multiple steps. First, users are asked, \"Can you locate the -- enclosed in a polygon?\" If the target content is present, annotators are then asked for each possible subpart, \"Can you locate any subpart> on the --?\" If the response is yes, annotators segment the subpart. The interface supports segmentation by collecting a series of clicked points to create a connected polygon. It also supports annotating multiple polygons in three scenarios: (1) multiple instances of a subpart (e.g., two eyes), (2) subparts with holes (e.g., coiled snakes), and (3) occlusions breaking a subpart into multiple, disconnected pieces.", + "bbox": [ + 212, + 258, + 787, + 395 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Annotation Collection. We hired 18 highly trusted annotators from Amazon Mechanical Turk (AMT) who we had vetted through repeated employment for previous segmentation tasks. During data collection, we ensured high annotation quality through five methods: an onboarding qualification test, detailed instructions, live \"office hours\" during annotation periods, phased task roll-out for worker feedback, and continuous inspection of submitted results.", + "bbox": [ + 212, + 412, + 787, + 503 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset Splits. Our dataset adheres to the PartImageNet training, validation, and testing splits of $85\\%$ , $5\\%$ , and $10\\%$ of the data, respectively. This results in 8,828 training, 519 validation, and 1,040 test images in our SPIN dataset.", + "bbox": [ + 212, + 521, + 785, + 568 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Dataset Analysis", + "text_level": 1, + "bbox": [ + 214, + 593, + 401, + 608 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We now characterize SPIN's overall composition as well as its subparts.", + "bbox": [ + 214, + 621, + 730, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Overall Dataset Composition. We characterize SPIN with respect to the following seven factors: number of images, number of object categories, total number of annotated objects, number of part categories, total number of annotated parts, number of subpart categories, and total number of annotated subparts. Results are shown in Table 1. As shown, our contribution of 203 entity categories is over a five-order-of-magnitude increase compared to the 40 categories added for PartImageNet [17], while providing nearly an order of magnitude more annotations (i.e. 106,324 semantic annotations versus 11,960 part segmentations in PartImageNet).", + "bbox": [ + 212, + 655, + 787, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Subpart Statistics. We next compute the mean number of subparts per part and characterize the typical appearance of subparts with respect to five metrics:", + "bbox": [ + 214, + 809, + 785, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "SPIN", + "bbox": [ + 692, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e27bf80c8da0872c26717dfabd74cd446b5750be21a6decec092c84e034bd0f2.jpg", + "image_caption": [ + "(a)InstancesPerPart" + ], + "image_footnote": [], + "bbox": [ + 215, + 145, + 496, + 188 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f3631d77a8daa27e9b50bcdbe6533ddaeff8f554f0d8d498f9056f68815eb1c0.jpg", + "image_caption": [ + "(d) Image Coverage" + ], + "image_footnote": [], + "bbox": [ + 500, + 145, + 784, + 186 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/02c1d19954c0b8a8be2278aa0bfb6aeee8435c36109c719e6cf6ad2b26aa2a24.jpg", + "image_caption": [ + "(b) Extent" + ], + "image_footnote": [], + "bbox": [ + 215, + 199, + 496, + 241 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e4a9c85556fe4b6ab9637225151d7bd2a5aff75e6a00e8d29ce8011715a9cd65.jpg", + "image_caption": [ + "(e) Object Coverage" + ], + "image_footnote": [], + "bbox": [ + 500, + 199, + 782, + 239 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ba7dced9c2690df2deb057d8007b7976724e45c71f87630bde54df26fc2e5d07.jpg", + "image_caption": [ + "(c) Boundary Complexity" + ], + "image_footnote": [], + "bbox": [ + 215, + 251, + 496, + 311 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5949391db2cbc1db7b5857255d943d57df4500982fb2a298c541327733e7f32e.jpg", + "image_caption": [ + "(f) Part Coverage", + "Fig. 3: Boxplots showing the distribution of subpart image occupation (left) and boundary complexities per part, per object (right). The blue lines represent medians, bottoms and tops of each box represent the 25th and 75th percentile values respectively, and whiskers represent the most extreme data points not considered outliers. Overall, SPIN's subparts take up a relatively small number of pixels per image, while featuring a range of geometric complexity. (Aero=Aeroplane; Quad=Quadruped)" + ], + "image_footnote": [], + "bbox": [ + 500, + 251, + 782, + 316 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Boundary complexity: ratio of a subpart's area to the length of its perimeter (i.e., isoperimetric quotient). Values range from 0 (highly jagged boundary) to 1 (circular).", + "- Extent: the ratio of the area of a contour to the area of its bounding box. Values are in $(0,1]$ , where values approaching 0 mean that a contour occupies little area in its bounding box (e.g., a thin diagonal line) and 1 means that a contour is perfectly contained (e.g., a square).", + "- Image coverage: fraction of image pixels occupied by the subpart.", + "- Object coverage: fraction of parent object pixels occupied by the subpart.", + "- Part coverage: fraction of parent part pixels occupied by the subpart." + ], + "bbox": [ + 225, + 454, + 784, + 612 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For all six factors, we report both the overall mean as well as the mean with respect to each object category. Results are shown in Fig. 3.", + "bbox": [ + 212, + 626, + 782, + 656 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Objects tend to have 2 to 5 subparts per part (i.e. Fig. 3a; 25th to 75th percentile values), with a total of 4 to 15 subparts. There is diversity across object categories with aeroplanes typically having small variability with only 1 to 2 subparts per part (i.e. 25th to 75th percentile range) while car typically exhibits great variability with 2 to 8 subparts per part (i.e. 25th to 75th percentile range), as exemplified in Fig. 1(c-d). We attribute the large range of subpart per part counts to two factors: the intrinsic diversity in cars and the viewing angle. For example, a go-cart viewed from the front will have no windshield, no windows, and two tires, while a bus viewed from a front-right angle will have a windshield, tires, and windows. Such object-specific patterns around the expected prevalence of subparts could be predictive cues for models decomposing detected objects into their recursively nested subparts.", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Myers-Dean et al.", + "bbox": [ + 271, + 114, + 392, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In terms of shape, most subparts' contours occupy the majority of their bounding box (i.e., Fig. 3b, median extent value $\\sim 0.7$ ) with moderate boundary complexity (i.e., Fig. 3c, median boundary complexity values around 0.5). Still, we observe high variability for both metrics that we again attribute to the intrinsic diversity in each object category as well as the viewing angle. For example, a bottle cap or a bicycle tire viewed from the side will be an ellipse versus a nearly perfect circle when viewed from the top (i.e., occupying more of its bounding box). Altogether, these findings highlight that our dataset encourages the design of models that can account for a large range of subpart shapes.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Regarding the relative area occupied by each subpart, we find that all subparts occupy small portions of the parent parts, parent objects, and entire image. For example, across every object category, subparts occupy less than $6\\%$ of the image for at least $75\\%$ of subpart instances (Fig. 3d). Additionally, the subparts tend to occupy less than $20\\%$ of their parent parts (Fig. 3e; Overall) and less than $5\\%$ of their parent objects (Fig. 3f; Overall). When adopting size thresholds introduced for the MSCOCO dataset [33], where $32^2$ and $96^2$ are thresholds determining whether an object is small, medium, or large, we find $54.10\\%$ (57,525) of SPIN's subpart annotations qualify as small, $38.08\\%$ (40,488) as medium, and only $7.82\\%$ (8,311) as large. Most subparts are already considered small according to mainstream research despite that we annotated subparts for a dataset known to typically contain a single, large prominent object in the image. Consequently, embedded in our problem is the well-known challenge of small entity detection [53], which will only grow as subparts are explored for more complex scenes where less prominent objects must also be hierarchically decomposed.", + "bbox": [ + 212, + 285, + 787, + 512 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "When comparing subparts belonging to the rigid object categories (i.e., human-made) versus non-rigid object categories (i.e., animals), we found no clear distinction. For instance, the mean number of subparts per part (Fig. 3a) is the smallest amount with least variability in aeroplanes and the greatest amount with greatest variability in cars. While prior work has emphasized distinguishing between these two extremes [17], our analysis suggests that rigid and non-rigid objects appear quite similar at the subpart level.", + "bbox": [ + 212, + 513, + 787, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Evaluating Hierarchical Consistency", + "text_level": 1, + "bbox": [ + 214, + 656, + 602, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We facilitate assessing how well relationships across varying levels of granularity are captured for hierarchical segmentation by introducing two metrics that account for spatial and semantic relationships respectively.", + "bbox": [ + 212, + 700, + 785, + 746 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We first introduce the Spatial Consistency Score (SpCS), which focuses on structural hierarchical alignment rather than semantic alignment. It indicates the proportion of a child's segmentation (e.g., subparts) that is contained in its super-region's segmentation (e.g., part, object), with values ranging from 0 (no containment) to 1 (perfect containment). It leverages binarized segmentations distinguishing the foreground entity from the background. Formally, it is com", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "SPIN", + "bbox": [ + 692, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "puted as the average containment ratio across all pairs as follows:", + "bbox": [ + 215, + 146, + 689, + 162 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nS p C S = \\frac {1}{| \\mathcal {R} |} \\sum_ {\\text {(c h i l d , p a r e n t)} \\in \\mathcal {R}} \\underbrace {\\frac {| \\text {c h i l d} \\cap \\text {p a r e n t} |}{| \\text {c h i l d} |}} _ {\\text {c o n t a i n m e n t r a t i o}} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 348, + 170, + 785, + 220 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathcal{R}$ denotes the set of prediction pairs with a ground truth parent-child relationship within a hierarchical structure. $^3 |\\mathcal{R}|$ denotes the total number of parent-child pairs, $|\\mathrm{child}\\cap \\mathrm{parent}|$ is the number of pixels in the intersection of the child segmentation with the parent, $|\\mathrm{child}|$ is the number of pixels in the child segmentation.", + "bbox": [ + 212, + 227, + 787, + 303 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We next introduce the Semantic Consistency Score (SeCS), which focuses on semantic alignment by measuring how well predictions at different levels (subpart, part, object) logically correspond. For example, an eye (subpart) should correspond to a head (part) and, by extension, to a quadruped (object), rather than illogical associations like a windshield (subpart) with a head (part) or a bottle (object). Let $FG_{S}$ , $FG_{P}$ , and $FG_{O}$ be the sets of non-background predictions in subpart, part, and object predictions, respectively. Define $X = FG_{S} \\cap FG_{P} \\cap FG_{O}$ as the intersection of these predictions, representing pixels with consistent foreground labels across levels. For each pixel $x \\in X$ , with predicted categories $C(S_x)$ , $C(P_x)$ , and $C(O_x)$ for subparts $(S_x)$ , parts $(P_x)$ , and objects $(O_x)$ , respectively, category entailment is evaluated using $\\mathbb{1}[C(S_x), C(P_x)]$ and $\\mathbb{1}[C(P_x), C(O_x)]$ , where $\\mathbb{1}$ is an indicator function returning 1 if the hierarchical entailment is correct, using ground truth relations, otherwise 0. We compute the entailment measure $M(x)$ as:", + "bbox": [ + 212, + 303, + 787, + 517 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nM (x) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} \\mathbb {1} [ C (S _ {x}), C (P _ {x}) ] \\wedge \\mathbb {1} [ C (P _ {x}), C (O _ {x}) ], \\\\ 0 & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 316, + 525, + 785, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\wedge$ is the logical and operator. SeCS is the mean of $M(x)$ across all $x\\in X$ . Values range from 0 (no semantic coherence) to 1 (perfect semantic coherence).", + "bbox": [ + 214, + 571, + 787, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5 Algorithm Benchmarking", + "text_level": 1, + "bbox": [ + 215, + 625, + 501, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We now assess modern models' ability to account for hierarchical decompositions to the subpart granularity.", + "bbox": [ + 214, + 652, + 787, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.1 Open-Vocabulary Localization in a Zero-Shot Setting", + "text_level": 1, + "bbox": [ + 214, + 704, + 702, + 722 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We first evaluate models for open-vocabulary localization in zero-shot mode across three hierarchy levels: object, part, and subpart. This approach avoids the drawbacks of the alternative options of training models from scratch or finetuning, including the computational expense and risks of overfitting to the target task and so diminishing generalizable knowledge.", + "bbox": [ + 212, + 728, + 787, + 805 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Myers-Dean et al.", + "bbox": [ + 271, + 114, + 392, + 128 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "3 For a quadruped, for instance, pairs such as $\\{(\\text{eyes}, \\text{head}), (\\text{chest}, \\text{torso}), (\\text{torso}, \\text{quadruped})\\}$ could be present in $\\mathcal{R}$ .", + "bbox": [ + 217, + 810, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Models. We benchmark 11 model variants.", + "bbox": [ + 215, + 146, + 535, + 160 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Most related to our work is HIPIE [60], which is designed for open-vocabulary hierarchical segmentation to the subpart level. It takes as input a list of all candidate categories that can appear and then predicts which categories are present where in the given image. We test two variants with different backbones: ResNet-50 [18] and ViT-H [12].", + "bbox": [ + 212, + 161, + 787, + 236 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We also evaluate 10 variants of foundation models that, by design, are intended for use on a variety of downstream tasks. Five produce pixel-wise segmentation masks: LISA-7B [24], LISA-13B [24], GLaMM [46], and PixelLLM-7B [48], and PixelLLM-13B [48]. The other five can only produce bounding boxes: CoGVLM [59], Ferret-7B [64], Ferret-13B [64], Shikra [6], and Kosmos2 [43]. For these models, we analyze their upper bound of their potential performance by notifying the model what object category needs to be located in a given image. We examine two types of requests: \"Can you please locate the in the image?\" for objects, and \"Can you please locate the subpart> of the in the image?\"4 for each part and subpart. We evaluate with these prompts a Specific category by using the original WordNet category from the 158 ImageNet categories as well as a General category by referring to the 11 super-categories. Consequently, we test each model with four prompts.", + "bbox": [ + 212, + 238, + 787, + 434 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Evaluation Metrics. We use two evaluation metrics for all models: (1) Intersection over Union (IoU) to assess segmentation performance for objects, parts, and subparts independently and (2) our Spatial Consistency Score (SpCS) to assess structural coherence across object, part, and subpart levels. We also use our new Semantic Consistency Score (SeCS) to assess semantic coherence between predicted hierarchical levels for the only relevant model, HIPIE (it predicts from a list of candidate categories rather than for the target, single category).", + "bbox": [ + 212, + 450, + 787, + 556 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Localization Results (mIoU). Results are shown in Table 2.", + "text_level": 1, + "bbox": [ + 214, + 573, + 674, + 588 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We observe a consistent trend across all models: localization performance is best for object segmentation, followed by increasingly worse performance for more granular categories of parts and then subparts. The best score for localizing objects is 86 followed by 40 for parts and 14 for subparts. The worst scores come from the only model directly designed for hierarchical segmentation, HIPIE. It does worse across all granularity levels than the foundation models; e.g., for subpart localization, 1 vs 3 from the worst-performing foundation model, Kosmos2. We attribute the poor performance to HIPIE's limited training data (e.g., limited part vocabulary), which only included part-level categories from Pascal Parts [7]. Moreover, HIPIE's heuristic grouping assumes that parts are the sum of their subparts, which is not the case in SPIN (e.g., unlabelled \"space\" exists between a cheek an eye in a head).", + "bbox": [ + 212, + 589, + 787, + 768 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We explored to what extent entity size is correlated to the resulting IoU scores by using linear regression for the better-performing foundation models.", + "bbox": [ + 214, + 771, + 785, + 801 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "SPIN", + "bbox": [ + 692, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_footnote", + "text": "4 For subparts, we refer to the parent object rather than the parent part, as broader context aids in processing finer details [47].", + "bbox": [ + 217, + 810, + 785, + 840 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/30b76fd5695f262381be2edcff54132a4ae1453f033fc563b31eeb36a653ad73.jpg", + "table_caption": [], + "table_footnote": [ + "Table 2: Performance of modern models with respect to localization at three granularity levels—subparts (mIoUs), parts (mIoUP), objects (mIoUO)—and hierarchical consistency at two levels—subparts to parts (SpCS-S2P), parts to objects (SpCS-P2O), and their mean (SpCS - Avg). Results are shown for hierarchical semantic segmentation models (described in Section 5.1), hierarchical object detection models (Section 5.1), and interactive segmentation models that perform category-agnostic localization (Section 5.2). The method families are separated by dashed lines in that order. Best performing scores are shown in bold." + ], + "table_body": "
Method# Params.mIoUsmIoUPmIoUOSpCS - AvgSpCS - S2PSpCS - P2O
SpecificGeneralSpecificGeneralSpecificGeneralSpecificGeneralSpecificGeneralSpecificGeneral
HIPIE [60] (R-50)200M0.800.108.057.2251.3613.8797.8298.7710010095.6497.54
HIPIE [60] (ViT-H)800M0.900.927.218.2366.7721.6998.0499.2510098.5096.0896.06
PixelLLM [48]7B9.089.2432.3731.3783.9379.8773.7982.4669.5779.2089.4990.73
PixelLLM [48]13B9.5310.1332.0432.9682.9079.9684.8683.0682.1380.8792.6088.90
LISA [24]7B9.619.1427.7227.2886.2383.3679.3077.4474.6172.9392.1089.78
LISA [24]13B11.5211.5532.2931.3687.7885.4586.2879.9882.8775.7096.0292.41
GLaMM [46]7B11.0311.0039.4140.0086.2986.3187.1287.7584.2184.9695.7295.38
GLaMM - FT7B24.2524.5659.3760.7686.4291.0887.1387.9575.9385.1690.2396.04
Ferret [64]7B7.227.3025.6926.5447.8447.9878.7073.0246.5065.6893.6594.60
Ferret [64]13B6.376.2523.6723.6547.9947.5081.0477.3048.7070.9894.1795.40
CoGVLM [59]17B13.9414.2938.1338.5646.4743.6576.3377.2072.3773.3286.1987.59
Shikra [6]7B8.418.7226.5027.4245.5029.1468.5767.0163.8266.7681.8567.89
Kosmos2 [43]2.6B3.453.5919.0019.4848.6348.8982.7481.6078.3477.0396.0195.58
SAM [22]630M49.6149.6169.2369.2390.0690.0686.4286.4283.8583.8592.3092.30
SAM + CoGVLM630M19.8819.9850.9350.9077.9475.0966.7167.3380.5080.7360.6561.44
", + "bbox": [ + 215, + 143, + 785, + 292 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Due to space constraints, we provide further experimental details and results in the supplementary materials. In summary though, our findings suggest that size has little correlation to performance for objects but greater correlations for more granular levels of parts and subparts. We suspect that one contributing factor is the sensitivity of the IoU metric for entities occupying a small number of pixels, as small errors are especially prone to yield dramatic changes in the IoU scores for such cases.", + "bbox": [ + 212, + 455, + 787, + 561 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We next explored the impact of a prompt's category specificity on model performance by comparing performance for when general versus specific terms are employed. Most models perform worse when trying to locate objects categorized under general terms (e.g., 'quadruped') rather than specific ones (e.g., 'dog'). For example, CoGVLM performs 2.82 percentage points (pp) worse with general categories compared to specific categories. Similarly, the 13B variant of LISA has a 2.33pp performance drop and its 7B counterpart has a 2.87pp performance drop. The exceptions to this are GLaMM, Ferret, and Kosmos2, whose performance are relatively stable and unaffected by the level of category specificity. A potential reason is that these models utilize training data that explicitly use both general and specific terms. For example, GCG [46], a source of training data for GLaMM, refers to entities with both abstract and specific terms (e.g., person and toddler). Overall, this variance highlights the challenge in choosing an appropriate level of category specificity for objects. However, the story changes when trying to locate parts and subparts, with the performance gap between general and specific categories falling to within 1pp. An interesting direction for future work is better understanding how and why an entity's granularity level influences models' abilities to overcome abstraction challenges.", + "bbox": [ + 212, + 568, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Myers-Dean et al.", + "bbox": [ + 271, + 114, + 390, + 128 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We also examine how the training data of the Large Language Models (LLM) Llama [54], since it serves as the basis for the majority of foundation models evaluated in our study. Specifically, we analyze the empirical frequency (i.e., total occurrence) of uni-grams (e.g., eye) and n-grams (e.g., eye of the quadruped) within the dataset using the $\\infty$ -gram API [34], which uses byte array pattern matching to find all occurrences of a given BPE tokenized [49] $n$ -gram. Due to space constraints, results are provided in the supplementary materials. In summary though, we observe a decline in average uni-gram occurrence frequency when increasing granularity (e.g., part to subpart). An interesting direction for future work is to remedy this imbalance, such as by designing parameter-efficient methods to restructure the text embedding space of LLMs to recognize hierarchical relationships, akin to MERU [10] restructuring CLIP [44]-space for image-level is-a relationships.", + "bbox": [ + 212, + 145, + 787, + 343 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Performance When Training on Our SPIN Dataset. We next examine the potential benefit of our SPIN training data for modern models. To do so, we fine-tuned the top-performing GLaMM algorithm [46] for all hierarchy levels in SPIN, and we refer to this variant as GLaMM-FT. Results are shown in Table 2.", + "bbox": [ + 212, + 377, + 787, + 438 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Overall, we observe a considerable performance boost. For example, when prompting with general categories, we observe a 13.56 percentage point (pp) increase (123% relative increase) for subparts, a 20.76pp increase (51.90% relative increase) for parts, and a 4.77pp increase for objects (5.51% relative increase). While such performance gains are promising, we still observe very low absolute scores for subparts, indicating their challenge for modern models. We suspect a promising direction for future work is providing cost-effective ways to further increase the amount of available training data, such as through synthetic data creation and 2D projections from 3D models.", + "bbox": [ + 212, + 441, + 787, + 580 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Hierarchical Consistency Results. SpCS results are shown in Table 2. Similar trends are observed for part-object and subpart-part relationships. Part-object pairs consistently exhibit high SpCS-P2O scores, indicating models tend to correctly predict parts within their corresponding whole objects. This provides evidence that models' have some understanding of spatial containment among part-object relationships. A potential reason for this is that models are indirectly trained on part-whole relationships, because they are already taught that an object is a part of a scene. The aforementioned trend is slightly worse when examining consistency between subparts and parts (i.e., SpCS-S2P), with scores varying from 46.50 (i.e., Ferret 7B) to 84.21 (i.e., GLaMM) when employing general terms for objects and from 65.68 (i.e., Ferret 7B) to 84.96 (i.e., GLaMM) when employing more specific categories for objects. This variability, coupled with low mIoU scores for subparts, reveals that models often predict subparts outside their intended part boundaries, underscoring poor performance in properly contextualizing subparts within an object.", + "bbox": [ + 212, + 613, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "SPIN", + "bbox": [ + 692, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.2 Interactive Segmentation for Vocabulary-Free Localization", + "text_level": 1, + "bbox": [ + 215, + 146, + 743, + 161 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We next assess to what extent models can localize entities at all hierarchy levels in our dataset if they are notified where to look in an image. This sets an upper bound of what is possible. To do so, we benchmark the state-of-the-art interactive segmentation model, SAM [22], which takes as input a coarse marking on an image specifying where to create the segmentation. We use bounding boxes as our coarse image markings, because they are shown to outperform alternatives [41] when using SAM. We explore the upper bound performance for this model by using the ground truth object, part, and subpart segmentations to generate the bounding boxes (i.e., by using $[x_{min}, y_{min}, x_{max}, y_{max}]$ for each entity's pixels). Results are shown in Table 2.", + "bbox": [ + 212, + 172, + 787, + 321 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Despite the idealized guidance from the bounding box prompts, SAM's performance is still imperfect. As granularity increases, localization performance consistently decreases by about 20pp when going down a granularity level (e.g., objects to parts), revealing that SAM struggles more as segmentation requests become more fine-grained. Still, our findings demonstrate SAM's potential for yielding considerably higher quality hierarchical segmentations than observed from existing localization models benchmarked in the previous section. Compared to the top-performing zero-shot segmentation-based model (i.e., GLaMM), SAM achieves a boost of 38.61pp for subparts, 20.23pp for parts, and 3.75pp for objects, alongside similar SpCS scores.", + "bbox": [ + 212, + 323, + 787, + 474 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We also evaluate a fully-automated system with SAM, by incorporating bounding box predictions from the top-performing bounding box producing model, CoGVLM. Results are shown in Table 2. We similarly observe that it outperforms the leading zero-shot segmentation method, GLaMM, in segmenting subparts and parts (9.24 percentage point (pp) boost for subparts and 10.87pp for parts). We attribute these performance gains to SAM's ability to leverage a location prior (e.g., bounding box), whereas models like GLaMM rely solely on language priors. In contrast, this approach performs worse than the leading zero-shot segmentation method, GLaMM, for objects, with a 11.22pp decrease. We attribute this performance drop to CoGVLM's inferior performance in consistently and accurately detecting bounding boxes.", + "bbox": [ + 212, + 474, + 787, + 640 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.3 Recognizing Hierarchical Semantics in a Zero-Shot Setting", + "text_level": 1, + "bbox": [ + 215, + 662, + 743, + 678 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We finally evaluate the ability of models to recognize hierarchical semantic labels for objects in images (i.e., without localization). We assess this in an idealized scenario in which we notify models what to look for as well as where to look in the image. This sets an upper bound of what may be possible with such models as they are explicitly told where to look.", + "bbox": [ + 212, + 688, + 787, + 763 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Models. We evaluate eight off-the-shelf model variants: CoGVLM [59], Ferret-7B [64], Ferret-13B [64], Shikra [6], Osprey [66], Kosmos2 [43], ViP-Llava-7B [4], and ViP-Llava-13B [4]. All prompts to the models resemble the following basic request: \"Is there a in the ?\" for objects and \"Is there", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Myers-Dean et al.", + "bbox": [ + 271, + 114, + 390, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/608b20079d52782dd62c32f0e8176d111f05f40cb53f1018a4bc593e5936b556.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method# Params.Prompt\\( \\mathrm{mACC}_S \\)\\( \\mathrm{mACC}_{SS} \\)\\( \\mathrm{mACC}_P \\)\\( \\mathrm{mACC}_{PS} \\)\\( \\mathrm{mACC}_O \\)\\( \\mathrm{mACC}_{OS} \\)
Shikra [6]7BBbox66.1261.2879.6774.5677.0482.05
Ferret [64]7BBbox59.0157.1959.9566.1786.7889.21
Ferret [64]13BBbox50.3169.4261.7267.8488.9389.02
CoGVLM [59]17BCoT42.1645.5481.7583.3187.1583.67
Kosmos2 [43]2.6BSG79.1264.3274.2368.5384.5578.78
Osprey [66]7BMask77.4355.9586.3081.2096.4384.85
ViP-Llava [4]7BMask53.5758.5264.1364.9671.6265.25
ViP-Llava [4]13BMask94.7897.4399.9099.8098.3395.06
", + "bbox": [ + 215, + 143, + 785, + 261 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 3: Accuracy metrics for object hierarchical decomposition: general (mACC $S$ , $\\mathrm{mACC}_P$ , $\\mathrm{mACC}_O$ ) and specific ( $\\mathrm{mACC}_{SS}$ , $\\mathrm{mACC}_{PS}$ , $\\mathrm{mACC}_{OS}$ ) categories across subparts, parts, and objects. (CoT = Chain of Thought; SG = Self-Grounding)", + "bbox": [ + 215, + 262, + 785, + 304 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "a $<(sub)part>$ of the $$ in the $?$ \" for parts and subparts. Regions are specified either with segmentations (ViP-Llava, Ferret, Osprey) or bounding boxes (CoGVLM, Shikra, Kosmos2). We generate these regions directly from the ground truth annotations (e.g., bounding boxes as described in Section 5.2) for objects, parts, and subparts. We prompt models for each category known to be in the region, one at a time. An answer is 'yes' if 'yes' is present in the model's output, and otherwise 'no'. As done in Section 5.1, we test both general and specific categories in the prompts.", + "bbox": [ + 212, + 345, + 787, + 467 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Evaluation Metrics. We use accuracy to evaluate at each granularity level.", + "bbox": [ + 215, + 483, + 772, + 498 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Results. Results are reported in Table 3.", + "text_level": 1, + "bbox": [ + 215, + 513, + 519, + 527 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "While most models struggle overall to recognize categories in images (despite being told where to look in the images), one model has nearly perfect accuracy: ViP-Llava-13B. One hypothesis for this atypically strong performance is the model lacks understanding and instead always answers yes when prompted about the presence of categories. We explored this with further experiments using adversarial examples that are instead expected to always answer \"no\". However, our findings from these experiments are none-conclusive. We observe the model can answer \"no\", but it also suffers in this case from a significant decrease in accuracy for the adversarial prompts. This performance drop could be due to a separate issue that the model struggles with negation (i.e., the reversal curse [3]). Altogether, this evaluation represents the best-case scenario for performance; it is likely to fall further in real-world applications due to imperfect human inputs or model-generated segmentations, both of which we have shown to exhibit poorer performance at finer granularity levels, as discussed in Sections 5.1 and 5.2.", + "bbox": [ + 212, + 529, + 787, + 739 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We observed mixed outcomes regarding which granularity levels are most difficult to recognize. For Ferret-13B and Osprey, an increase in granularity correlates with improved accuracy; e.g., transitioning from subparts to objects using general category names results in a 28.62pp increase for Ferret-13B and a 19.00pp", + "bbox": [ + 212, + 741, + 787, + 803 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "SPIN", + "bbox": [ + 692, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_footnote", + "text": "5 Results are shown in the supplementary materials for two prompts asking \"Is the category not present\" and \"Is the [different category] present\".", + "bbox": [ + 217, + 810, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "increase for Osprey. Conversely, for most other models, object identification consistently outperforms subpart identification. The differential performance across granularity levels may partly stem from the disparate prevalence of tokens in the training data across different granularity levels, particularly within the Llama dataset (as previously discussed). For parts and subparts, the mixed results observed—where parts sometimes outperform subparts (e.g., CoGVLM) and vice versa (e.g., Ferret-7B)—could be attributed to their relative frequency in the training data. Future research could focus on examining how foundation models differentiate responses based on prompt granularity and on enriching datasets like GRIT [64] with hierarchical relationships to improve model understanding of intra-entity relationships. Additionally, future work could explore enhancing hierarchical-relationship understanding so success at different hierarchical levels can facilitate finding entities at other granularity levels.", + "bbox": [ + 212, + 146, + 787, + 343 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We also observed mixed outcomes when comparing the efficacy of specific versus general categories. Half the models performed better using general terms for subparts, with a similar pattern observed in parts and objects. A potential reason for this discrepancy, compared to the relatively consistent results in Sec 5.1, is that localization models often rely on feature similarities, such as CLIP [44] features. Even if a model does not capture every nuance of a category query, it still attempts to find a match based on overall similarity to known categories (e.g., mini-van vs. van) and is less likely to reject user queries [62]. Conversely, interactive understanding models, designed to maximize accuracy, may strictly search for precise category matches within their trained representations. Given our task's structure (i.e., answers are restricted to 'yes' or 'no'), these models are predisposed to outright reject the query (i.e., answer 'no') if an exact match is absent, rather than proposing close or related category matches.", + "bbox": [ + 212, + 343, + 787, + 541 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 561, + 359, + 578 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We introduce SPIN, the first dataset challenge for hierarchical segmentation at the subpart granularity in natural images. Our analysis reveals SPIN's characteristics, our two new evaluation metrics enable gauging algorithmic proficiency in capturing the spatial and semantic relationships across hierarchy levels, and our benchmarking across three tasks demonstrates that models struggle to recognize and segment subparts in SPIN, even under idealized guidance. We publicly release SPIN to encourage further advancements in hierarchical segmentation, sharing the dataset at https://joshmyersdean.github.io/spin/index.html.", + "bbox": [ + 212, + 594, + 787, + 715 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments. This work was supported by Adobe Research Gift Funds and utilized the Blanca condo computing resource at the University of Colorado Boulder. Josh Myers-Dean is supported by a NSF GRFP fellowship (#1917573). We thank the crowdworkers for contributing their time for the construction of SPIN and the authors of our benchmarked models for open-sourcing their work.", + "bbox": [ + 212, + 729, + 787, + 806 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Myers-Dean et al.", + "bbox": [ + 271, + 114, + 390, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Explore images. https://support.apple.com/guide/iphone/use-voiceover-for-images-and-videos-iph37e6b3844/ios", + "2. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)", + "3. Berglund, L., Tong, M., Kaufmann, M., Balesni, M., Stickland, A.C., Korbak, T., Evans, O.: The reversal curse: Llms trained on\" a is b\" fail to learn\" b is a\". arXiv preprint arXiv:2309.12288 (2023)", + "4. Cai, M., Liu, H., Mustikovela, S.K., Meyer, G.P., Chai, Y., Park, D., Lee, Y.J.: Making large multimodal models understand arbitrary visual prompts. In: IEEE Conference on Computer Vision and Pattern Recognition (2024)", + "5. Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., et al.: Shapenet: An information-rich 3d moel repository. arXiv preprint arXiv:1512.03012 (2015)", + "6. Chen, K., Zhang, Z., Zeng, W., Zhang, R., Zhu, F., Zhao, R.: Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195 (2023)", + "7. Chen, X., Mottaghi, R., Liu, X., Fidler, S., Urtasun, R., Yuille, A.: Detect what you can: Detecting and representing objects using holistic models and body parts. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1971-1978 (2014)", + "8. Deitke, M., Han, W., Herrasti, A., Kembhavi, A., Kolve, E., Mottaghi, R., Salvador, J., Schwenk, D., VanderBilt, E., Wallingford, M., Weihs, L., Yatskar, M., Farhadi, A.: Robothor: An open simulation-to-real embodied ai platform. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)", + "9. Deng, B., Genova, K., Yazdani, S., Bouaziz, S., Hinton, G., Tagliasacchi, A.: Cvxnet: Learnable convex decomposition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 31-44 (2020)", + "10. Desai, K., Nickel, M., Rajpurohit, T., Johnson, J., Vedantam, S.R.: Hyperbolic image-text representations. In: International Conference on Machine Learning. pp. 7694-7731. PMLR (2023)", + "1. Ding, M., Shen, Y., Fan, L., Chen, Z., Chen, Z., Luo, P., Tenenbaum, J.B., Gan, C.: Visual dependency transformers: Dependency tree emerges from reversed attention. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14528-14539 (2023)", + "2. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)", + "3. Geng, H., Xu, H., Zhao, C., Xu, C., Yi, L., Huang, S., Wang, H.: Gapartnet: Cross-category domain-generalizable object perception and manipulation via generalizable and actionable parts. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7081-7091 (2023)", + "4. de Geus, D., Meletis, P., Lu, C., Wen, X., Dubbelman, G.: Part-aware panoptic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5485-5494 (2021)" + ], + "bbox": [ + 225, + 179, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "SPIN", + "bbox": [ + 692, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Gong, K., Liang, X., Li, Y., Chen, Y., Yang, M., Lin, L.: Instance-level human parsing via part grouping network. In: Proceedings of the European conference on computer vision (ECCV). pp. 770-785 (2018)", + "16. He, J., Chen, J., Lin, M.X., Yu, Q., Yuille, A.L.: Compositor: Bottom-up clustering and compositing for robust part and object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11259-11268 (June 2023)", + "17. He, J., Yang, S., Yang, S., Kortylewski, A., Yuan, X., Chen, J.N., Liu, S., Yang, C., Yu, Q., Yuille, A.: Partimagenet: A large, high-quality dataset of parts. In: European Conference on Computer Vision. pp. 128-145. Springer (2022)", + "18. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)", + "19. Hong, Y., Li, Q., Zhu, S.C., Huang, S.: VLgrammer: Grounded grammar induction of vision and language. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1665-1674 (2021)", + "20. Hong, Y., Yi, L., Tenenbaum, J., Torralba, A., Gan, C.: PTR: A benchmark for part-based conceptual, relational, and physical reasoning. Advances in Neural Information Processing Systems 34, 17427-17440 (2021)", + "21. Jia, M., Shi, M., Sirotenko, M., Cui, Y., Cardie, C., Hariharan, B., Adam, H., Belongie, S.: Fashionpedia: Ontology, segmentation, and an attribute localization dataset. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16. pp. 316-332. Springer (2020)", + "22. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 4015-4026 (October 2023)", + "23. Koo, J., Huang, I., Achlioptas, P., Guibas, L.J., Sung, M.: Partglot: Learning shape part segmentation from language reference games. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16505-16514 (2022)", + "24. Lai, X., Tian, Z., Chen, Y., Li, Y., Yuan, Y., Liu, S., Jia, J.: Lisa: Reasoning segmentation via large language model. arXiv preprint arXiv:2308.00692 (2023)", + "25. Lee, J., Peng, Y.H., Herskovitz, J., Guo, A.: Image explorer: Multi-layered touch exploration to make images accessible. In: Proceedings of the 23rd International ACM SIGACCESS Conference on Computers and Accessibility. ASSETS '21, Association for Computing Machinery, New York, NY, USA (2021). https://doi.org/10.1145/3441852.3476548, https://doi.org/10.1145/3441852.3476548", + "26. Li, F., Zhang, H., Sun, P., Zou, X., Liu, S., Yang, J., Li, C., Zhang, L., Gao, J.: Semantic-sam: Segment and recognize anything at any granularity. arXiv preprint arXiv:2307.04767 (2023)", + "27. Li, L., Zhou, T., Wang, W., Li, J., Yang, Y.: Deep hierarchical semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1246-1257 (2022)", + "28. Li, T., Gupta, V., Mehta, M., Srikumar, V.: A logic-driven framework for consistency of neural models. arXiv preprint arXiv:1909.00126 (2019)", + "29. Li, X., Xu, S., Yang, Y., Cheng, G., Tong, Y., Tao, D.: Panoptic-partformer: Learning a unified model for panoptic part segmentation. In: European Conference on Computer Vision. pp. 729-747. Springer (2022)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Myers-Dean et al.", + "bbox": [ + 271, + 114, + 390, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "30. Li, X., Xu, S., Yang, Y., Yuan, H., Cheng, G., Tong, Y., Lin, Z., Yang, M.H., Tao, D.: Panopticpartformer++: A unified and decoupled view for panoptic part segmentation. arXiv preprint arXiv:2301.00954 (2023)", + "31. Liang, X., Gong, K., Shen, X., Lin, L.: Look into person: Joint body parsing & pose estimation network and a new benchmark. IEEE transactions on pattern analysis and machine intelligence 41(4), 871-885 (2018)", + "32. Liang, X., Shen, X., Feng, J., Lin, L., Yan, S.: Semantic object parsing with graph lstm. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14. pp. 125-143. Springer (2016)", + "33. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)", + "34. Liu, J., Min, S., Zettlemoyer, L., Choi, Y., Hajishirzi, H.: Infini-gram: Scaling unbounded n-gram language models to a trillion tokens. arXiv preprint arXiv:2401.17377 (2024)", + "35. Liu, Q., Kortylewski, A., Zhang, Z., Li, Z., Guo, M., Liu, Q., Yuan, X., Mu, J., Qiu, W., Yuille, A.: Cgpart: A part segmentation dataset based on 3d computer graphics models. arXiv preprint arXiv:2103.14098 (2021)", + "36. Martin, D., Fowlkes, C., Tal, D., Malik, J.: A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In: Proceedings Eighth IEEE International Conference on Computer Vision. ICCV 2001. vol. 2, pp. 416-423. IEEE (2001)", + "37. Michieli, U., Zanuttigh, P.: Edge-aware graph matching network for part-based semantic segmentation. International Journal of Computer Vision 130(11), 2797-2821 (2022)", + "38. Miller, G.A.: Wordnet: a lexical database for english. Commun. ACM 38(11), 39-41 (nov 1995). https://doi.org/10.1145/219717.219748, https://doi.org/10.1145/219717.219748", + "39. Mo, K., Zhu, S., Chang, A.X., Yi, L., Tripathi, S., Guibas, L.J., Su, H.: Partnet: A large-scale benchmark for fine-grained and hierarchical part-level 3d object understanding. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 909-918 (2019)", + "40. Mo, K., Zhu, S., Chang, A.X., Yi, L., Tripathi, S., Guibas, L.J., Su, H.: PartNet: A large-scale benchmark for fine-grained and hierarchical part-level 3D object understanding. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019)", + "41. Myers-Dean, J., Fan, Y., Price, B., Chan, W., Gurari, D.: Interactive segmentation for diverse gesture types without context. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 7198-7208 (January 2024)", + "42. Nair, V., Zhu, H.H., Smith, B.A.: Imageassist: Tools for enhancing touchscreen-based image exploration systems for blind and low vision users. In: Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems. CHI '23, Association for Computing Machinery, New York, NY, USA (2023). https://doi.org/10.1145/3544548.3581302, https://doi.org/10.1145/3544548.3581302", + "43. Peng, Z., Wang, W., Dong, L., Hao, Y., Huang, S., Ma, S., Wei, F.: Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824 (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "SPIN", + "bbox": [ + 692, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "44. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)", + "45. Ramanathan, V., Kalia, A., Petrovic, V., Wen, Y., Zheng, B., Guo, B., Wang, R., Marquez, A., Kovvuri, R., Kadian, A., et al.: Paco: Parts and attributes of common objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7141-7151 (2023)", + "46. Rasheed, H., Maaz, M., Shaji, S., Shaker, A., Khan, S., Cholakkal, H., Anwer, R.M., Xing, E., Yang, M.H., Khan, F.S.: Glamm: Pixel grounding large multimodal model. arXiv preprint arXiv:2311.03356 (2023)", + "47. RAYMOND W. GIBBS, J., Matlock, T.: Psycholinguistics and mental representations (2000)", + "48. Ren, Z., Huang, Z., Wei, Y., Zhao, Y., Fu, D., Feng, J., Jin, X.: Pixelm: Pixel reasoning with large multimodal model (2023)", + "49. Sennrich, R., Haddow, B., Birch, A.: Neural machine translation of rare words with subword units. arXiv preprint arXiv:1508.07909 (2015)", + "50. Song, X., Wang, P., Zhou, D., Zhu, R., Guan, C., Dai, Y., Su, H., Li, H., Yang, R.: Apollocar3d: A large 3d car instance understanding benchmark for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5452-5462 (2019)", + "51. Sun, P., Chen, S., Zhu, C., Xiao, F., Luo, P., Xie, S., Yan, Z.: Going denser with open-vocabulary part segmentation. 2023 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 15407-15419 (2023), https://api.sementicscholar.org/CorpusID:258762519", + "52. Tang, C., Xie, L., Zhang, X., Hu, X., Tian, Q.: Visual recognition by request. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15265-15274 (2023)", + "53. Tong, K., Wu, Y.: Deep learning-based detection from the perspective of small or tiny objects: A survey. Image and Vision Computing 123, 104471 (2022)", + "54. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., Lample, G.: Llama: Open and efficient foundation language models (2023)", + "55. Tsogkas, S., Kokkinos, I., Papandreou, G., Vedaldi, A.: Deep learning for semantic part segmentation with high-level guidance. arXiv preprint arXiv:1505.02438 (2015)", + "56. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The caltech-ucsd birds-200-2011 dataset (2011)", + "57. Wang, J., Yuille, A.L.: Semantic part segmentation using compositional model combining shape and appearance. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1788-1797 (2015)", + "58. Wang, P., Shen, X., Lin, Z., Cohen, S., Price, B., Yuille, A.L.: Joint object and part segmentation using deep learned potentials. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 1573-1581 (2015)", + "59. Wang, W., Lv, Q., Yu, W., Hong, W., Qi, J., Wang, Y., Ji, J., Yang, Z., Zhao, L., Song, X., et al.: Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079 (2023)", + "60. Wang, X., Li, S., Kallidromitis, K., Kato, Y., Kozuka, K., Darrell, T.: Hierarchical open-vocabulary universal image segmentation. Advances in Neural Information Processing Systems 36 (2024)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Myers-Dean et al.", + "bbox": [ + 271, + 114, + 390, + 128 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "61. Wei, M., Yue, X., Zhang, W., Kong, S., Liu, X., Pang, J.: Ov-parts: Towards open-vocabulary part segmentation. Advances in Neural Information Processing Systems 36 (2024)", + "62. Wu, T.H., Biamby, G., Chan, D., Dunlap, L., Gupta, R., Wang, X., Gonzalez, J.E., Darrell, T.: See, say, and segment: Teaching lmms to overcome false premises. arXiv preprint arXiv:2312.08366 (2023)", + "63. Xiang, F., Qin, Y., Mo, K., Xia, Y., Zhu, H., Liu, F., Liu, M., Jiang, H., Yuan, Y., Wang, H., Yi, L., Chang, A.X., Guibas, L.J., Su, H.: SAPIEN: A simulated part-based interactive environment. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)", + "64. You, H., Zhang, H., Gan, Z., Du, X., Zhang, B., Wang, Z., Cao, L., Chang, S.F., Yang, Y.: Ferret: Refer and ground anything anywhere at any granularity. arXiv preprint arXiv:2310.07704 (2023)", + "65. Yu, F., Liu, K., Zhang, Y., Zhu, C., Xu, K.: Partnet: A recursive part decomposition network for fine-grained and hierarchical shape segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9491-9500 (2019)", + "66. Yuan, Y., Li, W., Liu, J., Tang, D., Luo, X., Qin, C., Zhang, L., Zhu, J.: Osprey: Pixel understanding with visual instruction tuning (2023)", + "67. Zhao, J., Li, J., Cheng, Y., Sim, T., Yan, S., Feng, J.: Understanding humans in crowded scenes: Deep nested adversarial learning and a new benchmark for multi-human parsing. In: Proceedings of the 26th ACM international conference on Multimedia. pp. 792-800 (2018)", + "68. Zheng, S., Yang, F., Kiapour, M.H., Piramuthu, R.: Modanet: A large-scale street fashion dataset with polygon annotations. In: Proceedings of the 26th ACM international conference on Multimedia. pp. 1670-1678 (2018)", + "69. Zhou, B., Zhao, H., Puig, X., Fidler, S., Barriuso, A., Torralba, A.: Scene parsing through ade20k dataset. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 633-641 (2017)" + ], + "bbox": [ + 212, + 146, + 787, + 550 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "SPIN", + "bbox": [ + 692, + 114, + 730, + 126 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/SPIN_ Hierarchical Segmentation with Subpart Granularity in Natural Images/88afeb42-c98d-4785-a2a2-08cde2fa67ab_model.json b/2024/SPIN_ Hierarchical Segmentation with Subpart Granularity in Natural Images/88afeb42-c98d-4785-a2a2-08cde2fa67ab_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ca9dc09359fcc14e5d487495c1bf6733f02f453b --- /dev/null +++ b/2024/SPIN_ Hierarchical Segmentation with Subpart Granularity in Natural Images/88afeb42-c98d-4785-a2a2-08cde2fa67ab_model.json @@ -0,0 +1,2614 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.233, + 0.141, + 0.773, + 0.187 + ], + "angle": 0, + "content": "SPIN: Hierarchical Segmentation with Subpart Granularity in Natural Images" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.213, + 0.744, + 0.244 + ], + "angle": 0, + "content": "Josh Myers-Dean\\(^{1}\\), Jarek Reynolds\\(^{1}\\), Brian Price\\(^{2}\\), Yifei Fan\\(^{2}\\), and Danna Gurari\\(^{1,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.247, + 0.254, + 0.761, + 0.27 + ], + "angle": 0, + "content": "1 University of Colorado Boulder, 2 Adobe, 3 University of Texas at Austin" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.309, + 0.744, + 0.45 + ], + "angle": 0, + "content": "Abstract. Hierarchical segmentation entails creating segmentations at varying levels of granularity. We introduce the first hierarchical semantic segmentation dataset with subpart annotations for natural images, which we call SPIN (SubPartImageNet). We also introduce two novel evaluation metrics to evaluate how well algorithms capture spatial and semantic relationships across hierarchical levels. We benchmark modern models across three different tasks and analyze their strengths and weaknesses across objects, parts, and subparts. To facilitate community-wide progress, we publicly release our dataset at https://joshmyersdean.github.io/spin/index.html." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.475, + 0.377, + 0.492 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.644 + ], + "angle": 0, + "content": "When people discuss hierarchical image analysis tasks they are typically talking about one of two approaches: is-a relationships [27,28], which treat categories at different abstraction levels (e.g., a Subaru is a car), or is-part-of relationships, which concentrate on dividing objects into their constituent parts [13, 14, 19] (e.g., a door is part of a car). While the former has been widely explored, the latter—focusing on object decomposition—has received limited attention in computer vision research. Within segmentation research, the focus of this paper, is-part-of relationships have primarily been explored only for part-whole hierarchies, ignoring finer-grained details, such as subparts (i.e., parts of parts)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.788, + 0.765 + ], + "angle": 0, + "content": "Research to subpart-level segmentation granularity is hindered by a scarcity of data. While synthetic 3D datasets [14, 19, 35, 63] could be used to infer hierarchical segmentations to subpart granularity, it is well-known that models developed with synthetic data typically generalize poorly to natural images (i.e., images taken by a camera) [8]. This lack of annotated natural data has meant that the few models designed to generate subpart granularity segmentations [11, 52, 60] could only be evaluated qualitatively on a small number of examples (rather than quantitatively at scale)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Our work facilitates the development of algorithms for subpart granularity in three key ways. First, we collect over 102,000 subpart segmentations across 203 diverse subpart categories to expand upon PartImageNet [17] and create the SubPartImageNet (SPIN) dataset—the first to offer subpart annotations for natural images. We release this dataset publicly to foster community progress." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.129 + ], + "angle": 0, + "content": "Myers-Dean et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.315, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.255, + 0.22, + 0.272, + 0.231 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.146, + 0.407, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.352, + 0.22, + 0.369, + 0.231 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.146, + 0.501, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.446, + 0.22, + 0.462, + 0.231 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.146, + 0.597, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.541, + 0.22, + 0.558, + 0.231 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.146, + 0.689, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.635, + 0.22, + 0.651, + 0.231 + ], + "angle": 0, + "content": "(e)" + }, + { + "type": "image", + "bbox": [ + 0.692, + 0.146, + 0.784, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.728, + 0.22, + 0.744, + 0.231 + ], + "angle": 0, + "content": "(f)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.242, + 0.788, + 0.327 + ], + "angle": 0, + "content": "Fig. 1: Overview of the diversity of SPIN. Panels (a) and (b) depict subparts unique to specific object class members, such as a roll cage in a car and a shell in a turtle. Panels (c) and (d) illustrate the variability in the number of subparts per object of the same class, with examples of 13 and 6 subparts. Panels (e) and (f) highlight the disparity in image area coverage by different subparts, such as a bottle label (large) versus quadruped claws (tiny)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.357, + 0.788, + 0.493 + ], + "angle": 0, + "content": "Examples of annotated subparts are shown in Fig. 1. Second, we introduce two novel metrics to address a shortcoming of the prevailing evaluation methods for hierarchical segmentation, which only assess each granularity level in isolation. Our newly proposed metrics are intended to work alongside these traditional metrics (e.g., Intersection over Union) and assess how well models capture spatial and semantic relationships across hierarchical levels. Our third key contribution is that we benchmark modern models on SPIN across objects, parts, and subparts, thereby presenting the first comprehensive quantitative evaluation of subpart performance and key areas needed for future improvement." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.494, + 0.788, + 0.691 + ], + "angle": 0, + "content": "Advancing hierarchical segmentation to subpart granularity could significantly benefit both research and societal applications. One potential use is in generating more detailed image descriptions, such as for augmented reality experiences, answering visual questions, captioning images, or visual storytelling. Similarly, this work could be the foundation for individuals to (1) interactively learn how to speak about finer-grained entities for the first time (for children) or in a new language and (2) recall forgotten words, such as when facing temporary or chronic memory injuries or disabilities. More deeply nested, hierarchical segmentation representations can also enrich tactile visual discovery [25, 42], including for use in screen readers employed by visually impaired individuals (e.g., Apple's object-based exploration feature [1]). Similarly, we believe finer-grained information could facilitate improvements for related problems, including image/video retrieval, image/video editing, automatic magnification, and robotics." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.715, + 0.398, + 0.731 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.749, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Datasets with Hierarchically Segmented Objects. Several datasets provide segmentations showing how objects are hierarchically decomposed into their recursively nested parts. A few do so without semantic labels specifying the categories of segmented content, including the pioneering Berkeley dataset published in 2001 [36] and the large-scale SA-1B dataset released in 2023 [22]. Our work, in contrast, focuses on hierarchically segmenting objects with category" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.693, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SPIN" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.269 + ], + "angle": 0, + "content": "labels. Already, numerous datasets semantically decompose objects and their parts in natural images [7, 14, 15, 17, 21, 31, 45, 50, 56, 61, 67-69] and synthetic images [5, 19, 20, 35, 39]. One dataset, (i.e., ADE20K [69]) even provides subpart annotations for \\(10\\%\\) of its objects, however its lack of exhaustive annotations for objects and their associated parts [52] impedes it benefit for evaluating prediction models. Extending prior work, we introduce the first exhaustively labelled hierarchical semantic segmentation dataset with subpart annotations for natural images by extending one of the largest part-based datasets [17]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.282, + 0.79, + 0.433 + ], + "angle": 0, + "content": "Hierarchical Semantic Segmentation Evaluation. Most evaluation protocols for hierarchical segmentation algorithms assess each hierarchy level independently, such as by reporting for object and part categories separately their Intersection over Union scores [9, 23, 31, 32, 37, 40, 52, 55, 57, 58, 61, 65], Average Precision scores [45, 51], and Panoptic Quality scores [14, 29, 30, 52, 60]. The one exception is Hierarchical Panoptic Quality [52] which recursively measures the extent to which a predicted hierarchy is complete (e.g., a wheel is a part of a car). Complementing prior work, we introduce evaluation metrics for assessing how well a model captures spatial and semantic hierarchical relationships, such as if a part is perfectly contained in it's whole." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.447, + 0.79, + 0.613 + ], + "angle": 0, + "content": "Hierarchical Semantic Segmentation Algorithms. Thus far, the focus of methods that predict all entities in a hierarchical decomposition largely has centered on predicting just objects and parts [16, 51, 58], aligning with existing datasets which only recently increasingly have added part segmentations to object segmentations. Only a few works, HIPIE [60], VDT [11], Semantic-SAM [26], and ViRReq [52]\\(^{1}\\), examined predicting more layers of a hierarchy to also include subparts. However, these works relied on qualitative assessments for subpart predictions due to a lack of annotated datasets supporting quantitative assessments. Complementing prior work, we quantitatively evaluate HIPIE and over 20 model variants to show their performance on our new dataset challenge across all three layers of the hierarchy (i.e., object, parts, and subparts)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.635, + 0.388, + 0.652 + ], + "angle": 0, + "content": "3 SPIN Dataset" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.667, + 0.787, + 0.697 + ], + "angle": 0, + "content": "We now introduce and characterize our new subpart semantic segmentation dataset that we call SPIN, which is short for SubPart ImageNet." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.718, + 0.406, + 0.733 + ], + "angle": 0, + "content": "3.1 Dataset Creation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.742, + 0.788, + 0.804 + ], + "angle": 0, + "content": "Data Source. Our dataset extends an existing dataset that provides both object and associated part annotations: PartImageNet [17]. It includes 24,080 natural images across 158 ImageNet categories (112 non-rigid and 38 rigid, e.g. animals and vehicles respectively). The authors of PartImageNet used the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.788, + 0.842 + ], + "angle": 0, + "content": "1 No code is publicly available for VDT, and ViRReq does not offer complete code. At the time of writing, Semantic-SAM has not released their semantic prediction code." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.129 + ], + "angle": 0, + "content": "Myers-Dean et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.269 + ], + "angle": 0, + "content": "WordNet [38] taxonomy to establish these categories and their 11 parent supercategories. The authors then identified 40 part categories pertinent to the supercategories and annotated these in all images. Every image in PartImageNet contains one semantically segmented object, and we used only those with at least one segmented part. Due to our limited annotation budget, we capped each supercategory at 1,200 images. This resulted in the following amounts per supercategory: 311 - Aeroplane, 483 - Bottle, 559 - Boat, 634 - Bicycle, and 1,200 each for Biped, Bird, Car, Fish, Quadruped, Reptile, and Snake." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.287, + 0.788, + 0.422 + ], + "angle": 0, + "content": "Subpart Category Selection. We identified 206 subpart categories to segment for 34 PartImageNet part categories that we identified as being decomposable into subparts. Of these, 168 subpart categories generalize across many part categories and 38 categories pertain to only a few. For example, in our dataset, mouths generalize while shells don't, with most reptiles having heads with mouths but only one type of reptile having a body with a shell (i.e., turtles). We show examples of specific subparts in Fig. 1(a-b). A histogram showing how many subparts are assigned to each of the 34 part categories is shown in Fig. 2, and a full list is provided in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.424, + 0.788, + 0.545 + ], + "angle": 0, + "content": "We created our subpart taxonomy through a multi-step process. We identified the general subpart categories by prompting GPT-4 [2] to list expected subparts for every object-part category in PartImageNet, and then three authors edited the list to exclude non-visible subparts (e.g., a skull). We identified specific subpart categories through manual inspection of at least 15 images per object category, excluding any categories that had ambiguous subpart decompositions; for example, it's not obvious to a lay person what boundaries to use when decomposing a tail into a tip, shaft, and base." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.564, + 0.788, + 0.61 + ], + "angle": 0, + "content": "Subpart Annotation Task Design. We created a task interface for collecting all subpart annotations. It presents each image-object pair alongside one of the object's nested parts, outlined with an overlaid polygon. For each part, the task" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.621, + 0.787, + 0.651 + ], + "angle": 0, + "content": "2 We prompted GPT-4 with \"Please list the canonical subparts of a \\(<\\) object \\(\\succ -\\text{part}>\\). Only include subparts that are clearly visible and recognizable to a layperson.\"" + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.682, + 0.787, + 0.798 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.808, + 0.787, + 0.838 + ], + "angle": 0, + "content": "Fig. 2: Histogram the number of unique subpart category labels for each of the 34 part categories. (Aero=Aeroplane; Quad=Quadruped)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.693, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SPIN" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "table", + "bbox": [ + 0.233, + 0.144, + 0.772, + 0.187 + ], + "angle": 0, + "content": "
#Images#Object Cat#Part Cat#Sub-Part Cat
10,38711 (10,387)40 (29,818)203 (106,324)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.187, + 0.786, + 0.215 + ], + "angle": 0, + "content": "Table 1: SPIN Composition: Each column lists category counts alongside the total number of annotated entities within those categories. (Cat=Categories)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.259, + 0.788, + 0.396 + ], + "angle": 0, + "content": "interface has multiple steps. First, users are asked, \"Can you locate the -- enclosed in a polygon?\" If the target content is present, annotators are then asked for each possible subpart, \"Can you locate any subpart> on the --?\" If the response is yes, annotators segment the subpart. The interface supports segmentation by collecting a series of clicked points to create a connected polygon. It also supports annotating multiple polygons in three scenarios: (1) multiple instances of a subpart (e.g., two eyes), (2) subparts with holes (e.g., coiled snakes), and (3) occlusions breaking a subpart into multiple, disconnected pieces." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.413, + 0.788, + 0.504 + ], + "angle": 0, + "content": "Annotation Collection. We hired 18 highly trusted annotators from Amazon Mechanical Turk (AMT) who we had vetted through repeated employment for previous segmentation tasks. During data collection, we ensured high annotation quality through five methods: an onboarding qualification test, detailed instructions, live \"office hours\" during annotation periods, phased task roll-out for worker feedback, and continuous inspection of submitted results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.522, + 0.787, + 0.569 + ], + "angle": 0, + "content": "Dataset Splits. Our dataset adheres to the PartImageNet training, validation, and testing splits of \\(85\\%\\), \\(5\\%\\), and \\(10\\%\\) of the data, respectively. This results in 8,828 training, 519 validation, and 1,040 test images in our SPIN dataset." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.594, + 0.403, + 0.609 + ], + "angle": 0, + "content": "3.2 Dataset Analysis" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.622, + 0.731, + 0.639 + ], + "angle": 0, + "content": "We now characterize SPIN's overall composition as well as its subparts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.656, + 0.788, + 0.793 + ], + "angle": 0, + "content": "Overall Dataset Composition. We characterize SPIN with respect to the following seven factors: number of images, number of object categories, total number of annotated objects, number of part categories, total number of annotated parts, number of subpart categories, and total number of annotated subparts. Results are shown in Table 1. As shown, our contribution of 203 entity categories is over a five-order-of-magnitude increase compared to the 40 categories added for PartImageNet [17], while providing nearly an order of magnitude more annotations (i.e. 106,324 semantic annotations versus 11,960 part segmentations in PartImageNet)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Subpart Statistics. We next compute the mean number of subparts per part and characterize the typical appearance of subparts with respect to five metrics:" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.129 + ], + "angle": 0, + "content": "Myers-Dean et al." + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.146, + 0.498, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.301, + 0.189, + 0.418, + 0.199 + ], + "angle": 0, + "content": "(a)InstancesPerPart" + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.146, + 0.785, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.594, + 0.189, + 0.698, + 0.199 + ], + "angle": 0, + "content": "(d) Image Coverage" + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.2, + 0.498, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.332, + 0.242, + 0.388, + 0.251 + ], + "angle": 0, + "content": "(b) Extent" + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.2, + 0.783, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.595, + 0.242, + 0.701, + 0.251 + ], + "angle": 0, + "content": "(e) Object Coverage" + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.252, + 0.498, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.291, + 0.314, + 0.434, + 0.325 + ], + "angle": 0, + "content": "(c) Boundary Complexity" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.252, + 0.783, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.606, + 0.317, + 0.705, + 0.327 + ], + "angle": 0, + "content": "(f) Part Coverage" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.339, + 0.788, + 0.424 + ], + "angle": 0, + "content": "Fig. 3: Boxplots showing the distribution of subpart image occupation (left) and boundary complexities per part, per object (right). The blue lines represent medians, bottoms and tops of each box represent the 25th and 75th percentile values respectively, and whiskers represent the most extreme data points not considered outliers. Overall, SPIN's subparts take up a relatively small number of pixels per image, while featuring a range of geometric complexity. (Aero=Aeroplane; Quad=Quadruped)" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.455, + 0.785, + 0.5 + ], + "angle": 0, + "content": "- Boundary complexity: ratio of a subpart's area to the length of its perimeter (i.e., isoperimetric quotient). Values range from 0 (highly jagged boundary) to 1 (circular)." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.502, + 0.785, + 0.563 + ], + "angle": 0, + "content": "- Extent: the ratio of the area of a contour to the area of its bounding box. Values are in \\((0,1]\\), where values approaching 0 mean that a contour occupies little area in its bounding box (e.g., a thin diagonal line) and 1 means that a contour is perfectly contained (e.g., a square)." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.564, + 0.731, + 0.58 + ], + "angle": 0, + "content": "- Image coverage: fraction of image pixels occupied by the subpart." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.58, + 0.784, + 0.596 + ], + "angle": 0, + "content": "- Object coverage: fraction of parent object pixels occupied by the subpart." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.597, + 0.757, + 0.613 + ], + "angle": 0, + "content": "- Part coverage: fraction of parent part pixels occupied by the subpart." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.455, + 0.785, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.627, + 0.784, + 0.657 + ], + "angle": 0, + "content": "For all six factors, we report both the overall mean as well as the mean with respect to each object category. Results are shown in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Objects tend to have 2 to 5 subparts per part (i.e. Fig. 3a; 25th to 75th percentile values), with a total of 4 to 15 subparts. There is diversity across object categories with aeroplanes typically having small variability with only 1 to 2 subparts per part (i.e. 25th to 75th percentile range) while car typically exhibits great variability with 2 to 8 subparts per part (i.e. 25th to 75th percentile range), as exemplified in Fig. 1(c-d). We attribute the large range of subpart per part counts to two factors: the intrinsic diversity in cars and the viewing angle. For example, a go-cart viewed from the front will have no windshield, no windows, and two tires, while a bus viewed from a front-right angle will have a windshield, tires, and windows. Such object-specific patterns around the expected prevalence of subparts could be predictive cues for models decomposing detected objects into their recursively nested subparts." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.693, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SPIN" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.284 + ], + "angle": 0, + "content": "In terms of shape, most subparts' contours occupy the majority of their bounding box (i.e., Fig. 3b, median extent value \\(\\sim 0.7\\)) with moderate boundary complexity (i.e., Fig. 3c, median boundary complexity values around 0.5). Still, we observe high variability for both metrics that we again attribute to the intrinsic diversity in each object category as well as the viewing angle. For example, a bottle cap or a bicycle tire viewed from the side will be an ellipse versus a nearly perfect circle when viewed from the top (i.e., occupying more of its bounding box). Altogether, these findings highlight that our dataset encourages the design of models that can account for a large range of subpart shapes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.286, + 0.788, + 0.513 + ], + "angle": 0, + "content": "Regarding the relative area occupied by each subpart, we find that all subparts occupy small portions of the parent parts, parent objects, and entire image. For example, across every object category, subparts occupy less than \\(6\\%\\) of the image for at least \\(75\\%\\) of subpart instances (Fig. 3d). Additionally, the subparts tend to occupy less than \\(20\\%\\) of their parent parts (Fig. 3e; Overall) and less than \\(5\\%\\) of their parent objects (Fig. 3f; Overall). When adopting size thresholds introduced for the MSCOCO dataset [33], where \\(32^2\\) and \\(96^2\\) are thresholds determining whether an object is small, medium, or large, we find \\(54.10\\%\\) (57,525) of SPIN's subpart annotations qualify as small, \\(38.08\\%\\) (40,488) as medium, and only \\(7.82\\%\\) (8,311) as large. Most subparts are already considered small according to mainstream research despite that we annotated subparts for a dataset known to typically contain a single, large prominent object in the image. Consequently, embedded in our problem is the well-known challenge of small entity detection [53], which will only grow as subparts are explored for more complex scenes where less prominent objects must also be hierarchically decomposed." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.515, + 0.788, + 0.621 + ], + "angle": 0, + "content": "When comparing subparts belonging to the rigid object categories (i.e., human-made) versus non-rigid object categories (i.e., animals), we found no clear distinction. For instance, the mean number of subparts per part (Fig. 3a) is the smallest amount with least variability in aeroplanes and the greatest amount with greatest variability in cars. While prior work has emphasized distinguishing between these two extremes [17], our analysis suggests that rigid and non-rigid objects appear quite similar at the subpart level." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.657, + 0.603, + 0.674 + ], + "angle": 0, + "content": "4 Evaluating Hierarchical Consistency" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.702, + 0.787, + 0.747 + ], + "angle": 0, + "content": "We facilitate assessing how well relationships across varying levels of granularity are captured for hierarchical segmentation by introducing two metrics that account for spatial and semantic relationships respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We first introduce the Spatial Consistency Score (SpCS), which focuses on structural hierarchical alignment rather than semantic alignment. It indicates the proportion of a child's segmentation (e.g., subparts) that is contained in its super-region's segmentation (e.g., part, object), with values ranging from 0 (no containment) to 1 (perfect containment). It leverages binarized segmentations distinguishing the foreground entity from the background. Formally, it is com" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.129 + ], + "angle": 0, + "content": "Myers-Dean et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.69, + 0.163 + ], + "angle": 0, + "content": "puted as the average containment ratio across all pairs as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.349, + 0.171, + 0.786, + 0.222 + ], + "angle": 0, + "content": "\\[\nS p C S = \\frac {1}{| \\mathcal {R} |} \\sum_ {\\text {(c h i l d , p a r e n t)} \\in \\mathcal {R}} \\underbrace {\\frac {| \\text {c h i l d} \\cap \\text {p a r e n t} |}{| \\text {c h i l d} |}} _ {\\text {c o n t a i n m e n t r a t i o}} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.228, + 0.789, + 0.304 + ], + "angle": 0, + "content": "where \\(\\mathcal{R}\\) denotes the set of prediction pairs with a ground truth parent-child relationship within a hierarchical structure. \\(^3 |\\mathcal{R}|\\) denotes the total number of parent-child pairs, \\(|\\mathrm{child}\\cap \\mathrm{parent}|\\) is the number of pixels in the intersection of the child segmentation with the parent, \\(|\\mathrm{child}|\\) is the number of pixels in the child segmentation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.304, + 0.789, + 0.518 + ], + "angle": 0, + "content": "We next introduce the Semantic Consistency Score (SeCS), which focuses on semantic alignment by measuring how well predictions at different levels (subpart, part, object) logically correspond. For example, an eye (subpart) should correspond to a head (part) and, by extension, to a quadruped (object), rather than illogical associations like a windshield (subpart) with a head (part) or a bottle (object). Let \\( FG_{S} \\), \\( FG_{P} \\), and \\( FG_{O} \\) be the sets of non-background predictions in subpart, part, and object predictions, respectively. Define \\( X = FG_{S} \\cap FG_{P} \\cap FG_{O} \\) as the intersection of these predictions, representing pixels with consistent foreground labels across levels. For each pixel \\( x \\in X \\), with predicted categories \\( C(S_x) \\), \\( C(P_x) \\), and \\( C(O_x) \\) for subparts \\( (S_x) \\), parts \\( (P_x) \\), and objects \\( (O_x) \\), respectively, category entailment is evaluated using \\( \\mathbb{1}[C(S_x), C(P_x)] \\) and \\( \\mathbb{1}[C(P_x), C(O_x)] \\), where \\( \\mathbb{1} \\) is an indicator function returning 1 if the hierarchical entailment is correct, using ground truth relations, otherwise 0. We compute the entailment measure \\( M(x) \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.318, + 0.526, + 0.786, + 0.567 + ], + "angle": 0, + "content": "\\[\nM (x) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} \\mathbb {1} [ C (S _ {x}), C (P _ {x}) ] \\wedge \\mathbb {1} [ C (P _ {x}), C (O _ {x}) ], \\\\ 0 & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.573, + 0.788, + 0.605 + ], + "angle": 0, + "content": "where \\(\\wedge\\) is the logical and operator. SeCS is the mean of \\(M(x)\\) across all \\(x\\in X\\). Values range from 0 (no semantic coherence) to 1 (perfect semantic coherence)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.625, + 0.502, + 0.645 + ], + "angle": 0, + "content": "5 Algorithm Benchmarking" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.654, + 0.788, + 0.687 + ], + "angle": 0, + "content": "We now assess modern models' ability to account for hierarchical decompositions to the subpart granularity." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.705, + 0.703, + 0.723 + ], + "angle": 0, + "content": "5.1 Open-Vocabulary Localization in a Zero-Shot Setting" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.729, + 0.788, + 0.806 + ], + "angle": 0, + "content": "We first evaluate models for open-vocabulary localization in zero-shot mode across three hierarchy levels: object, part, and subpart. This approach avoids the drawbacks of the alternative options of training models from scratch or finetuning, including the computational expense and risks of overfitting to the target task and so diminishing generalizable knowledge." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.788, + 0.842 + ], + "angle": 0, + "content": "3 For a quadruped, for instance, pairs such as \\(\\{(\\text{eyes}, \\text{head}), (\\text{chest}, \\text{torso}), (\\text{torso}, \\text{quadruped})\\}\\) could be present in \\(\\mathcal{R}\\)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.693, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SPIN" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.536, + 0.161 + ], + "angle": 0, + "content": "Models. We benchmark 11 model variants." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.162, + 0.788, + 0.237 + ], + "angle": 0, + "content": "Most related to our work is HIPIE [60], which is designed for open-vocabulary hierarchical segmentation to the subpart level. It takes as input a list of all candidate categories that can appear and then predicts which categories are present where in the given image. We test two variants with different backbones: ResNet-50 [18] and ViT-H [12]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.239, + 0.788, + 0.435 + ], + "angle": 0, + "content": "We also evaluate 10 variants of foundation models that, by design, are intended for use on a variety of downstream tasks. Five produce pixel-wise segmentation masks: LISA-7B [24], LISA-13B [24], GLaMM [46], and PixelLLM-7B [48], and PixelLLM-13B [48]. The other five can only produce bounding boxes: CoGVLM [59], Ferret-7B [64], Ferret-13B [64], Shikra [6], and Kosmos2 [43]. For these models, we analyze their upper bound of their potential performance by notifying the model what object category needs to be located in a given image. We examine two types of requests: \"Can you please locate the in the image?\" for objects, and \"Can you please locate the subpart> of the in the image?\"4 for each part and subpart. We evaluate with these prompts a Specific category by using the original WordNet category from the 158 ImageNet categories as well as a General category by referring to the 11 super-categories. Consequently, we test each model with four prompts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.452, + 0.788, + 0.558 + ], + "angle": 0, + "content": "Evaluation Metrics. We use two evaluation metrics for all models: (1) Intersection over Union (IoU) to assess segmentation performance for objects, parts, and subparts independently and (2) our Spatial Consistency Score (SpCS) to assess structural coherence across object, part, and subpart levels. We also use our new Semantic Consistency Score (SeCS) to assess semantic coherence between predicted hierarchical levels for the only relevant model, HIPIE (it predicts from a list of candidate categories rather than for the target, single category)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.574, + 0.676, + 0.589 + ], + "angle": 0, + "content": "Localization Results (mIoU). Results are shown in Table 2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.59, + 0.788, + 0.77 + ], + "angle": 0, + "content": "We observe a consistent trend across all models: localization performance is best for object segmentation, followed by increasingly worse performance for more granular categories of parts and then subparts. The best score for localizing objects is 86 followed by 40 for parts and 14 for subparts. The worst scores come from the only model directly designed for hierarchical segmentation, HIPIE. It does worse across all granularity levels than the foundation models; e.g., for subpart localization, 1 vs 3 from the worst-performing foundation model, Kosmos2. We attribute the poor performance to HIPIE's limited training data (e.g., limited part vocabulary), which only included part-level categories from Pascal Parts [7]. Moreover, HIPIE's heuristic grouping assumes that parts are the sum of their subparts, which is not the case in SPIN (e.g., unlabelled \"space\" exists between a cheek an eye in a head)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.772, + 0.787, + 0.802 + ], + "angle": 0, + "content": "We explored to what extent entity size is correlated to the resulting IoU scores by using linear regression for the better-performing foundation models." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.787, + 0.841 + ], + "angle": 0, + "content": "4 For subparts, we refer to the parent object rather than the parent part, as broader context aids in processing finer details [47]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.392, + 0.129 + ], + "angle": 0, + "content": "Myers-Dean et al." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.144, + 0.787, + 0.293 + ], + "angle": 0, + "content": "
Method# Params.mIoUsmIoUPmIoUOSpCS - AvgSpCS - S2PSpCS - P2O
SpecificGeneralSpecificGeneralSpecificGeneralSpecificGeneralSpecificGeneralSpecificGeneral
HIPIE [60] (R-50)200M0.800.108.057.2251.3613.8797.8298.7710010095.6497.54
HIPIE [60] (ViT-H)800M0.900.927.218.2366.7721.6998.0499.2510098.5096.0896.06
PixelLLM [48]7B9.089.2432.3731.3783.9379.8773.7982.4669.5779.2089.4990.73
PixelLLM [48]13B9.5310.1332.0432.9682.9079.9684.8683.0682.1380.8792.6088.90
LISA [24]7B9.619.1427.7227.2886.2383.3679.3077.4474.6172.9392.1089.78
LISA [24]13B11.5211.5532.2931.3687.7885.4586.2879.9882.8775.7096.0292.41
GLaMM [46]7B11.0311.0039.4140.0086.2986.3187.1287.7584.2184.9695.7295.38
GLaMM - FT7B24.2524.5659.3760.7686.4291.0887.1387.9575.9385.1690.2396.04
Ferret [64]7B7.227.3025.6926.5447.8447.9878.7073.0246.5065.6893.6594.60
Ferret [64]13B6.376.2523.6723.6547.9947.5081.0477.3048.7070.9894.1795.40
CoGVLM [59]17B13.9414.2938.1338.5646.4743.6576.3377.2072.3773.3286.1987.59
Shikra [6]7B8.418.7226.5027.4245.5029.1468.5767.0163.8266.7681.8567.89
Kosmos2 [43]2.6B3.453.5919.0019.4848.6348.8982.7481.6078.3477.0396.0195.58
SAM [22]630M49.6149.6169.2369.2390.0690.0686.4286.4283.8583.8592.3092.30
SAM + CoGVLM630M19.8819.9850.9350.9077.9475.0966.7167.3380.5080.7360.6561.44
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.214, + 0.294, + 0.788, + 0.404 + ], + "angle": 0, + "content": "Table 2: Performance of modern models with respect to localization at three granularity levels—subparts (mIoUs), parts (mIoUP), objects (mIoUO)—and hierarchical consistency at two levels—subparts to parts (SpCS-S2P), parts to objects (SpCS-P2O), and their mean (SpCS - Avg). Results are shown for hierarchical semantic segmentation models (described in Section 5.1), hierarchical object detection models (Section 5.1), and interactive segmentation models that perform category-agnostic localization (Section 5.2). The method families are separated by dashed lines in that order. Best performing scores are shown in bold." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.457, + 0.788, + 0.562 + ], + "angle": 0, + "content": "Due to space constraints, we provide further experimental details and results in the supplementary materials. In summary though, our findings suggest that size has little correlation to performance for objects but greater correlations for more granular levels of parts and subparts. We suspect that one contributing factor is the sensitivity of the IoU metric for entities occupying a small number of pixels, as small errors are especially prone to yield dramatic changes in the IoU scores for such cases." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.789, + 0.842 + ], + "angle": 0, + "content": "We next explored the impact of a prompt's category specificity on model performance by comparing performance for when general versus specific terms are employed. Most models perform worse when trying to locate objects categorized under general terms (e.g., 'quadruped') rather than specific ones (e.g., 'dog'). For example, CoGVLM performs 2.82 percentage points (pp) worse with general categories compared to specific categories. Similarly, the 13B variant of LISA has a 2.33pp performance drop and its 7B counterpart has a 2.87pp performance drop. The exceptions to this are GLaMM, Ferret, and Kosmos2, whose performance are relatively stable and unaffected by the level of category specificity. A potential reason is that these models utilize training data that explicitly use both general and specific terms. For example, GCG [46], a source of training data for GLaMM, refers to entities with both abstract and specific terms (e.g., person and toddler). Overall, this variance highlights the challenge in choosing an appropriate level of category specificity for objects. However, the story changes when trying to locate parts and subparts, with the performance gap between general and specific categories falling to within 1pp. An interesting direction for future work is better understanding how and why an entity's granularity level influences models' abilities to overcome abstraction challenges." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.693, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SPIN" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.344 + ], + "angle": 0, + "content": "We also examine how the training data of the Large Language Models (LLM) Llama [54], since it serves as the basis for the majority of foundation models evaluated in our study. Specifically, we analyze the empirical frequency (i.e., total occurrence) of uni-grams (e.g., eye) and n-grams (e.g., eye of the quadruped) within the dataset using the \\(\\infty\\)-gram API [34], which uses byte array pattern matching to find all occurrences of a given BPE tokenized [49] \\(n\\)-gram. Due to space constraints, results are provided in the supplementary materials. In summary though, we observe a decline in average uni-gram occurrence frequency when increasing granularity (e.g., part to subpart). An interesting direction for future work is to remedy this imbalance, such as by designing parameter-efficient methods to restructure the text embedding space of LLMs to recognize hierarchical relationships, akin to MERU [10] restructuring CLIP [44]-space for image-level is-a relationships." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.378, + 0.788, + 0.439 + ], + "angle": 0, + "content": "Performance When Training on Our SPIN Dataset. We next examine the potential benefit of our SPIN training data for modern models. To do so, we fine-tuned the top-performing GLaMM algorithm [46] for all hierarchy levels in SPIN, and we refer to this variant as GLaMM-FT. Results are shown in Table 2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.443, + 0.788, + 0.581 + ], + "angle": 0, + "content": "Overall, we observe a considerable performance boost. For example, when prompting with general categories, we observe a 13.56 percentage point (pp) increase (123% relative increase) for subparts, a 20.76pp increase (51.90% relative increase) for parts, and a 4.77pp increase for objects (5.51% relative increase). While such performance gains are promising, we still observe very low absolute scores for subparts, indicating their challenge for modern models. We suspect a promising direction for future work is providing cost-effective ways to further increase the amount of available training data, such as through synthetic data creation and 2D projections from 3D models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Hierarchical Consistency Results. SpCS results are shown in Table 2. Similar trends are observed for part-object and subpart-part relationships. Part-object pairs consistently exhibit high SpCS-P2O scores, indicating models tend to correctly predict parts within their corresponding whole objects. This provides evidence that models' have some understanding of spatial containment among part-object relationships. A potential reason for this is that models are indirectly trained on part-whole relationships, because they are already taught that an object is a part of a scene. The aforementioned trend is slightly worse when examining consistency between subparts and parts (i.e., SpCS-S2P), with scores varying from 46.50 (i.e., Ferret 7B) to 84.21 (i.e., GLaMM) when employing general terms for objects and from 65.68 (i.e., Ferret 7B) to 84.96 (i.e., GLaMM) when employing more specific categories for objects. This variability, coupled with low mIoU scores for subparts, reveals that models often predict subparts outside their intended part boundaries, underscoring poor performance in properly contextualizing subparts within an object." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.392, + 0.129 + ], + "angle": 0, + "content": "Myers-Dean et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.744, + 0.162 + ], + "angle": 0, + "content": "5.2 Interactive Segmentation for Vocabulary-Free Localization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.173, + 0.788, + 0.323 + ], + "angle": 0, + "content": "We next assess to what extent models can localize entities at all hierarchy levels in our dataset if they are notified where to look in an image. This sets an upper bound of what is possible. To do so, we benchmark the state-of-the-art interactive segmentation model, SAM [22], which takes as input a coarse marking on an image specifying where to create the segmentation. We use bounding boxes as our coarse image markings, because they are shown to outperform alternatives [41] when using SAM. We explore the upper bound performance for this model by using the ground truth object, part, and subpart segmentations to generate the bounding boxes (i.e., by using \\([x_{min}, y_{min}, x_{max}, y_{max}]\\) for each entity's pixels). Results are shown in Table 2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.324, + 0.788, + 0.475 + ], + "angle": 0, + "content": "Despite the idealized guidance from the bounding box prompts, SAM's performance is still imperfect. As granularity increases, localization performance consistently decreases by about 20pp when going down a granularity level (e.g., objects to parts), revealing that SAM struggles more as segmentation requests become more fine-grained. Still, our findings demonstrate SAM's potential for yielding considerably higher quality hierarchical segmentations than observed from existing localization models benchmarked in the previous section. Compared to the top-performing zero-shot segmentation-based model (i.e., GLaMM), SAM achieves a boost of 38.61pp for subparts, 20.23pp for parts, and 3.75pp for objects, alongside similar SpCS scores." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.476, + 0.789, + 0.641 + ], + "angle": 0, + "content": "We also evaluate a fully-automated system with SAM, by incorporating bounding box predictions from the top-performing bounding box producing model, CoGVLM. Results are shown in Table 2. We similarly observe that it outperforms the leading zero-shot segmentation method, GLaMM, in segmenting subparts and parts (9.24 percentage point (pp) boost for subparts and 10.87pp for parts). We attribute these performance gains to SAM's ability to leverage a location prior (e.g., bounding box), whereas models like GLaMM rely solely on language priors. In contrast, this approach performs worse than the leading zero-shot segmentation method, GLaMM, for objects, with a 11.22pp decrease. We attribute this performance drop to CoGVLM's inferior performance in consistently and accurately detecting bounding boxes." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.664, + 0.744, + 0.679 + ], + "angle": 0, + "content": "5.3 Recognizing Hierarchical Semantics in a Zero-Shot Setting" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.765 + ], + "angle": 0, + "content": "We finally evaluate the ability of models to recognize hierarchical semantic labels for objects in images (i.e., without localization). We assess this in an idealized scenario in which we notify models what to look for as well as where to look in the image. This sets an upper bound of what may be possible with such models as they are explicitly told where to look." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Models. We evaluate eight off-the-shelf model variants: CoGVLM [59], Ferret-7B [64], Ferret-13B [64], Shikra [6], Osprey [66], Kosmos2 [43], ViP-Llava-7B [4], and ViP-Llava-13B [4]. All prompts to the models resemble the following basic request: \"Is there a in the ?\" for objects and \"Is there" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.693, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SPIN" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.145, + 0.787, + 0.262 + ], + "angle": 0, + "content": "
Method# Params.Prompt\\( \\mathrm{mACC}_S \\)\\( \\mathrm{mACC}_{SS} \\)\\( \\mathrm{mACC}_P \\)\\( \\mathrm{mACC}_{PS} \\)\\( \\mathrm{mACC}_O \\)\\( \\mathrm{mACC}_{OS} \\)
Shikra [6]7BBbox66.1261.2879.6774.5677.0482.05
Ferret [64]7BBbox59.0157.1959.9566.1786.7889.21
Ferret [64]13BBbox50.3169.4261.7267.8488.9389.02
CoGVLM [59]17BCoT42.1645.5481.7583.3187.1583.67
Kosmos2 [43]2.6BSG79.1264.3274.2368.5384.5578.78
Osprey [66]7BMask77.4355.9586.3081.2096.4384.85
ViP-Llava [4]7BMask53.5758.5264.1364.9671.6265.25
ViP-Llava [4]13BMask94.7897.4399.9099.8098.3395.06
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.263, + 0.787, + 0.305 + ], + "angle": 0, + "content": "Table 3: Accuracy metrics for object hierarchical decomposition: general (mACC \\(S\\), \\(\\mathrm{mACC}_P\\), \\(\\mathrm{mACC}_O\\)) and specific (\\(\\mathrm{mACC}_{SS}\\), \\(\\mathrm{mACC}_{PS}\\), \\(\\mathrm{mACC}_{OS}\\)) categories across subparts, parts, and objects. (CoT = Chain of Thought; SG = Self-Grounding)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.347, + 0.788, + 0.468 + ], + "angle": 0, + "content": "a \\(<(sub)part>\\) of the \\(\\) in the \\(?\\)\" for parts and subparts. Regions are specified either with segmentations (ViP-Llava, Ferret, Osprey) or bounding boxes (CoGVLM, Shikra, Kosmos2). We generate these regions directly from the ground truth annotations (e.g., bounding boxes as described in Section 5.2) for objects, parts, and subparts. We prompt models for each category known to be in the region, one at a time. An answer is 'yes' if 'yes' is present in the model's output, and otherwise 'no'. As done in Section 5.1, we test both general and specific categories in the prompts." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.484, + 0.774, + 0.499 + ], + "angle": 0, + "content": "Evaluation Metrics. We use accuracy to evaluate at each granularity level." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.515, + 0.521, + 0.529 + ], + "angle": 0, + "content": "Results. Results are reported in Table 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.53, + 0.788, + 0.741 + ], + "angle": 0, + "content": "While most models struggle overall to recognize categories in images (despite being told where to look in the images), one model has nearly perfect accuracy: ViP-Llava-13B. One hypothesis for this atypically strong performance is the model lacks understanding and instead always answers yes when prompted about the presence of categories. We explored this with further experiments using adversarial examples that are instead expected to always answer \"no\". However, our findings from these experiments are none-conclusive. We observe the model can answer \"no\", but it also suffers in this case from a significant decrease in accuracy for the adversarial prompts. This performance drop could be due to a separate issue that the model struggles with negation (i.e., the reversal curse [3]). Altogether, this evaluation represents the best-case scenario for performance; it is likely to fall further in real-world applications due to imperfect human inputs or model-generated segmentations, both of which we have shown to exhibit poorer performance at finer granularity levels, as discussed in Sections 5.1 and 5.2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.742, + 0.788, + 0.804 + ], + "angle": 0, + "content": "We observed mixed outcomes regarding which granularity levels are most difficult to recognize. For Ferret-13B and Osprey, an increase in granularity correlates with improved accuracy; e.g., transitioning from subparts to objects using general category names results in a 28.62pp increase for Ferret-13B and a 19.00pp" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.787, + 0.84 + ], + "angle": 0, + "content": "5 Results are shown in the supplementary materials for two prompts asking \"Is the category not present\" and \"Is the [different category] present\"." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.392, + 0.129 + ], + "angle": 0, + "content": "Myers-Dean et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.344 + ], + "angle": 0, + "content": "increase for Osprey. Conversely, for most other models, object identification consistently outperforms subpart identification. The differential performance across granularity levels may partly stem from the disparate prevalence of tokens in the training data across different granularity levels, particularly within the Llama dataset (as previously discussed). For parts and subparts, the mixed results observed—where parts sometimes outperform subparts (e.g., CoGVLM) and vice versa (e.g., Ferret-7B)—could be attributed to their relative frequency in the training data. Future research could focus on examining how foundation models differentiate responses based on prompt granularity and on enriching datasets like GRIT [64] with hierarchical relationships to improve model understanding of intra-entity relationships. Additionally, future work could explore enhancing hierarchical-relationship understanding so success at different hierarchical levels can facilitate finding entities at other granularity levels." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.344, + 0.788, + 0.542 + ], + "angle": 0, + "content": "We also observed mixed outcomes when comparing the efficacy of specific versus general categories. Half the models performed better using general terms for subparts, with a similar pattern observed in parts and objects. A potential reason for this discrepancy, compared to the relatively consistent results in Sec 5.1, is that localization models often rely on feature similarities, such as CLIP [44] features. Even if a model does not capture every nuance of a category query, it still attempts to find a match based on overall similarity to known categories (e.g., mini-van vs. van) and is less likely to reject user queries [62]. Conversely, interactive understanding models, designed to maximize accuracy, may strictly search for precise category matches within their trained representations. Given our task's structure (i.e., answers are restricted to 'yes' or 'no'), these models are predisposed to outright reject the query (i.e., answer 'no') if an exact match is absent, rather than proposing close or related category matches." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.563, + 0.36, + 0.579 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.595, + 0.788, + 0.716 + ], + "angle": 0, + "content": "We introduce SPIN, the first dataset challenge for hierarchical segmentation at the subpart granularity in natural images. Our analysis reveals SPIN's characteristics, our two new evaluation metrics enable gauging algorithmic proficiency in capturing the spatial and semantic relationships across hierarchy levels, and our benchmarking across three tasks demonstrates that models struggle to recognize and segment subparts in SPIN, even under idealized guidance. We publicly release SPIN to encourage further advancements in hierarchical segmentation, sharing the dataset at https://joshmyersdean.github.io/spin/index.html." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.731, + 0.788, + 0.807 + ], + "angle": 0, + "content": "Acknowledgments. This work was supported by Adobe Research Gift Funds and utilized the Blanca condo computing resource at the University of Colorado Boulder. Josh Myers-Dean is supported by a NSF GRFP fellowship (#1917573). We thank the crowdworkers for contributing their time for the construction of SPIN and the authors of our benchmarked models for open-sourcing their work." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.693, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SPIN" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.18, + 0.785, + 0.208 + ], + "angle": 0, + "content": "1. Explore images. https://support.apple.com/guide/iphone/use-voiceover-for-images-and-videos-iph37e6b3844/ios" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.209, + 0.785, + 0.25 + ], + "angle": 0, + "content": "2. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.251, + 0.785, + 0.293 + ], + "angle": 0, + "content": "3. Berglund, L., Tong, M., Kaufmann, M., Balesni, M., Stickland, A.C., Korbak, T., Evans, O.: The reversal curse: Llms trained on\" a is b\" fail to learn\" b is a\". arXiv preprint arXiv:2309.12288 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.294, + 0.785, + 0.335 + ], + "angle": 0, + "content": "4. Cai, M., Liu, H., Mustikovela, S.K., Meyer, G.P., Chai, Y., Park, D., Lee, Y.J.: Making large multimodal models understand arbitrary visual prompts. In: IEEE Conference on Computer Vision and Pattern Recognition (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.336, + 0.785, + 0.378 + ], + "angle": 0, + "content": "5. Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., et al.: Shapenet: An information-rich 3d moel repository. arXiv preprint arXiv:1512.03012 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.379, + 0.785, + 0.42 + ], + "angle": 0, + "content": "6. Chen, K., Zhang, Z., Zeng, W., Zhang, R., Zhu, F., Zhao, R.: Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.421, + 0.785, + 0.475 + ], + "angle": 0, + "content": "7. Chen, X., Mottaghi, R., Liu, X., Fidler, S., Urtasun, R., Yuille, A.: Detect what you can: Detecting and representing objects using holistic models and body parts. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1971-1978 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.476, + 0.785, + 0.545 + ], + "angle": 0, + "content": "8. Deitke, M., Han, W., Herrasti, A., Kembhavi, A., Kolve, E., Mottaghi, R., Salvador, J., Schwenk, D., VanderBilt, E., Wallingford, M., Weihs, L., Yatskar, M., Farhadi, A.: Robothor: An open simulation-to-real embodied ai platform. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.546, + 0.785, + 0.588 + ], + "angle": 0, + "content": "9. Deng, B., Genova, K., Yazdani, S., Bouaziz, S., Hinton, G., Tagliasacchi, A.: Cvxnet: Learnable convex decomposition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 31-44 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.589, + 0.785, + 0.63 + ], + "angle": 0, + "content": "10. Desai, K., Nickel, M., Rajpurohit, T., Johnson, J., Vedantam, S.R.: Hyperbolic image-text representations. In: International Conference on Machine Learning. pp. 7694-7731. PMLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.631, + 0.785, + 0.686 + ], + "angle": 0, + "content": "1. Ding, M., Shen, Y., Fan, L., Chen, Z., Chen, Z., Luo, P., Tenenbaum, J.B., Gan, C.: Visual dependency transformers: Dependency tree emerges from reversed attention. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14528-14539 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.687, + 0.785, + 0.741 + ], + "angle": 0, + "content": "2. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.742, + 0.785, + 0.798 + ], + "angle": 0, + "content": "3. Geng, H., Xu, H., Zhao, C., Xu, C., Yi, L., Huang, S., Wang, H.: Gapartnet: Cross-category domain-generalizable object perception and manipulation via generalizable and actionable parts. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7081-7091 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "4. de Geus, D., Meletis, P., Lu, C., Wen, X., Dubbelman, G.: Part-aware panoptic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5485-5494 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.18, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.392, + 0.128 + ], + "angle": 0, + "content": "Myers-Dean et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "15. Gong, K., Liang, X., Li, Y., Chen, Y., Yang, M., Lin, L.: Instance-level human parsing via part grouping network. In: Proceedings of the European conference on computer vision (ECCV). pp. 770-785 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.247 + ], + "angle": 0, + "content": "16. He, J., Chen, J., Lin, M.X., Yu, Q., Yuille, A.L.: Compositor: Bottom-up clustering and compositing for robust part and object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11259-11268 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.248, + 0.785, + 0.288 + ], + "angle": 0, + "content": "17. He, J., Yang, S., Yang, S., Kortylewski, A., Yuan, X., Chen, J.N., Liu, S., Yang, C., Yu, Q., Yuille, A.: Partimagenet: A large, high-quality dataset of parts. In: European Conference on Computer Vision. pp. 128-145. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.29, + 0.785, + 0.33 + ], + "angle": 0, + "content": "18. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.332, + 0.785, + 0.374 + ], + "angle": 0, + "content": "19. Hong, Y., Li, Q., Zhu, S.C., Huang, S.: VLgrammer: Grounded grammar induction of vision and language. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1665-1674 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.375, + 0.785, + 0.415 + ], + "angle": 0, + "content": "20. Hong, Y., Yi, L., Tenenbaum, J., Torralba, A., Gan, C.: PTR: A benchmark for part-based conceptual, relational, and physical reasoning. Advances in Neural Information Processing Systems 34, 17427-17440 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.417, + 0.785, + 0.473 + ], + "angle": 0, + "content": "21. Jia, M., Shi, M., Sirotenko, M., Cui, Y., Cardie, C., Hariharan, B., Adam, H., Belongie, S.: Fashionpedia: Ontology, segmentation, and an attribute localization dataset. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16. pp. 316-332. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.474, + 0.785, + 0.529 + ], + "angle": 0, + "content": "22. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 4015-4026 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.53, + 0.785, + 0.585 + ], + "angle": 0, + "content": "23. Koo, J., Huang, I., Achlioptas, P., Guibas, L.J., Sung, M.: Partglot: Learning shape part segmentation from language reference games. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16505-16514 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.587, + 0.785, + 0.613 + ], + "angle": 0, + "content": "24. Lai, X., Tian, Z., Chen, Y., Li, Y., Yuan, Y., Liu, S., Jia, J.: Lisa: Reasoning segmentation via large language model. arXiv preprint arXiv:2308.00692 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.615, + 0.785, + 0.684 + ], + "angle": 0, + "content": "25. Lee, J., Peng, Y.H., Herskovitz, J., Guo, A.: Image explorer: Multi-layered touch exploration to make images accessible. In: Proceedings of the 23rd International ACM SIGACCESS Conference on Computers and Accessibility. ASSETS '21, Association for Computing Machinery, New York, NY, USA (2021). https://doi.org/10.1145/3441852.3476548, https://doi.org/10.1145/3441852.3476548" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.685, + 0.785, + 0.727 + ], + "angle": 0, + "content": "26. Li, F., Zhang, H., Sun, P., Zou, X., Liu, S., Yang, J., Li, C., Zhang, L., Gao, J.: Semantic-sam: Segment and recognize anything at any granularity. arXiv preprint arXiv:2307.04767 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.728, + 0.785, + 0.769 + ], + "angle": 0, + "content": "27. Li, L., Zhou, T., Wang, W., Li, J., Yang, Y.: Deep hierarchical semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1246-1257 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.77, + 0.785, + 0.798 + ], + "angle": 0, + "content": "28. Li, T., Gupta, V., Mehta, M., Srikumar, V.: A logic-driven framework for consistency of neural models. arXiv preprint arXiv:1909.00126 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "29. Li, X., Xu, S., Yang, Y., Cheng, G., Tong, Y., Tao, D.: Panoptic-partformer: Learning a unified model for panoptic part segmentation. In: European Conference on Computer Vision. pp. 729-747. Springer (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.693, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SPIN" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "30. Li, X., Xu, S., Yang, Y., Yuan, H., Cheng, G., Tong, Y., Lin, Z., Yang, M.H., Tao, D.: Panopticpartformer++: A unified and decoupled view for panoptic part segmentation. arXiv preprint arXiv:2301.00954 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "31. Liang, X., Gong, K., Shen, X., Lin, L.: Look into person: Joint body parsing & pose estimation network and a new benchmark. IEEE transactions on pattern analysis and machine intelligence 41(4), 871-885 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.232, + 0.787, + 0.286 + ], + "angle": 0, + "content": "32. Liang, X., Shen, X., Feng, J., Lin, L., Yan, S.: Semantic object parsing with graph lstm. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14. pp. 125-143. Springer (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.286, + 0.787, + 0.342 + ], + "angle": 0, + "content": "33. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.342, + 0.787, + 0.383 + ], + "angle": 0, + "content": "34. Liu, J., Min, S., Zettlemoyer, L., Choi, Y., Hajishirzi, H.: Infini-gram: Scaling unbounded n-gram language models to a trillion tokens. arXiv preprint arXiv:2401.17377 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.383, + 0.787, + 0.425 + ], + "angle": 0, + "content": "35. Liu, Q., Kortylewski, A., Zhang, Z., Li, Z., Guo, M., Liu, Q., Yuan, X., Mu, J., Qiu, W., Yuille, A.: Cgpart: A part segmentation dataset based on 3d computer graphics models. arXiv preprint arXiv:2103.14098 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.425, + 0.787, + 0.48 + ], + "angle": 0, + "content": "36. Martin, D., Fowlkes, C., Tal, D., Malik, J.: A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In: Proceedings Eighth IEEE International Conference on Computer Vision. ICCV 2001. vol. 2, pp. 416-423. IEEE (2001)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.48, + 0.787, + 0.522 + ], + "angle": 0, + "content": "37. Michieli, U., Zanuttigh, P.: Edge-aware graph matching network for part-based semantic segmentation. International Journal of Computer Vision 130(11), 2797-2821 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.522, + 0.787, + 0.562 + ], + "angle": 0, + "content": "38. Miller, G.A.: Wordnet: a lexical database for english. Commun. ACM 38(11), 39-41 (nov 1995). https://doi.org/10.1145/219717.219748, https://doi.org/10.1145/219717.219748" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.562, + 0.787, + 0.619 + ], + "angle": 0, + "content": "39. Mo, K., Zhu, S., Chang, A.X., Yi, L., Tripathi, S., Guibas, L.J., Su, H.: Partnet: A large-scale benchmark for fine-grained and hierarchical part-level 3d object understanding. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 909-918 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.619, + 0.787, + 0.674 + ], + "angle": 0, + "content": "40. Mo, K., Zhu, S., Chang, A.X., Yi, L., Tripathi, S., Guibas, L.J., Su, H.: PartNet: A large-scale benchmark for fine-grained and hierarchical part-level 3D object understanding. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.674, + 0.787, + 0.729 + ], + "angle": 0, + "content": "41. Myers-Dean, J., Fan, Y., Price, B., Chan, W., Gurari, D.: Interactive segmentation for diverse gesture types without context. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 7198-7208 (January 2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.729, + 0.787, + 0.799 + ], + "angle": 0, + "content": "42. Nair, V., Zhu, H.H., Smith, B.A.: Imageassist: Tools for enhancing touchscreen-based image exploration systems for blind and low vision users. In: Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems. CHI '23, Association for Computing Machinery, New York, NY, USA (2023). https://doi.org/10.1145/3544548.3581302, https://doi.org/10.1145/3544548.3581302" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "43. Peng, Z., Wang, W., Dong, L., Hao, Y., Huang, S., Ma, S., Wei, F.: Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.392, + 0.129 + ], + "angle": 0, + "content": "Myers-Dean et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.203 + ], + "angle": 0, + "content": "44. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.787, + 0.259 + ], + "angle": 0, + "content": "45. Ramanathan, V., Kalia, A., Petrovic, V., Wen, Y., Zheng, B., Guo, B., Wang, R., Marquez, A., Kovvuri, R., Kadian, A., et al.: Paco: Parts and attributes of common objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7141-7151 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.787, + 0.3 + ], + "angle": 0, + "content": "46. Rasheed, H., Maaz, M., Shaji, S., Shaker, A., Khan, S., Cholakkal, H., Anwer, R.M., Xing, E., Yang, M.H., Khan, F.S.: Glamm: Pixel grounding large multimodal model. arXiv preprint arXiv:2311.03356 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.301, + 0.787, + 0.328 + ], + "angle": 0, + "content": "47. RAYMOND W. GIBBS, J., Matlock, T.: Psycholinguistics and mental representations (2000)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.787, + 0.356 + ], + "angle": 0, + "content": "48. Ren, Z., Huang, Z., Wei, Y., Zhao, Y., Fu, D., Feng, J., Jin, X.: Pixelm: Pixel reasoning with large multimodal model (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.787, + 0.383 + ], + "angle": 0, + "content": "49. Sennrich, R., Haddow, B., Birch, A.: Neural machine translation of rare words with subword units. arXiv preprint arXiv:1508.07909 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.787, + 0.438 + ], + "angle": 0, + "content": "50. Song, X., Wang, P., Zhou, D., Zhu, R., Guan, C., Dai, Y., Su, H., Li, H., Yang, R.: Apollocar3d: A large 3d car instance understanding benchmark for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5452-5462 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.787, + 0.494 + ], + "angle": 0, + "content": "51. Sun, P., Chen, S., Zhu, C., Xiao, F., Luo, P., Xie, S., Yan, Z.: Going denser with open-vocabulary part segmentation. 2023 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 15407-15419 (2023), https://api.sementicscholar.org/CorpusID:258762519" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.787, + 0.536 + ], + "angle": 0, + "content": "52. Tang, C., Xie, L., Zhang, X., Hu, X., Tian, Q.: Visual recognition by request. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15265-15274 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.537, + 0.787, + 0.563 + ], + "angle": 0, + "content": "53. Tong, K., Wu, Y.: Deep learning-based detection from the perspective of small or tiny objects: A survey. Image and Vision Computing 123, 104471 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.787, + 0.605 + ], + "angle": 0, + "content": "54. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., Lample, G.: Llama: Open and efficient foundation language models (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.787, + 0.646 + ], + "angle": 0, + "content": "55. Tsogkas, S., Kokkinos, I., Papandreou, G., Vedaldi, A.: Deep learning for semantic part segmentation with high-level guidance. arXiv preprint arXiv:1505.02438 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.787, + 0.674 + ], + "angle": 0, + "content": "56. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The caltech-ucsd birds-200-2011 dataset (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.787, + 0.715 + ], + "angle": 0, + "content": "57. Wang, J., Yuille, A.L.: Semantic part segmentation using compositional model combining shape and appearance. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1788-1797 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.787, + 0.757 + ], + "angle": 0, + "content": "58. Wang, P., Shen, X., Lin, Z., Cohen, S., Price, B., Yuille, A.L.: Joint object and part segmentation using deep learned potentials. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 1573-1581 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.787, + 0.799 + ], + "angle": 0, + "content": "59. Wang, W., Lv, Q., Yu, W., Hong, W., Qi, J., Wang, Y., Ji, J., Yang, Z., Zhao, L., Song, X., et al.: Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "60. Wang, X., Li, S., Kallidromitis, K., Kato, Y., Kozuka, K., Darrell, T.: Hierarchical open-vocabulary universal image segmentation. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.693, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SPIN" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "61. Wei, M., Yue, X., Zhang, W., Kong, S., Liu, X., Pang, J.: Ov-parts: Towards open-vocabulary part segmentation. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.788, + 0.231 + ], + "angle": 0, + "content": "62. Wu, T.H., Biamby, G., Chan, D., Dunlap, L., Gupta, R., Wang, X., Gonzalez, J.E., Darrell, T.: See, say, and segment: Teaching lmms to overcome false premises. arXiv preprint arXiv:2312.08366 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.231, + 0.788, + 0.286 + ], + "angle": 0, + "content": "63. Xiang, F., Qin, Y., Mo, K., Xia, Y., Zhu, H., Liu, F., Liu, M., Jiang, H., Yuan, Y., Wang, H., Yi, L., Chang, A.X., Guibas, L.J., Su, H.: SAPIEN: A simulated part-based interactive environment. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.286, + 0.788, + 0.328 + ], + "angle": 0, + "content": "64. You, H., Zhang, H., Gan, Z., Du, X., Zhang, B., Wang, Z., Cao, L., Chang, S.F., Yang, Y.: Ferret: Refer and ground anything anywhere at any granularity. arXiv preprint arXiv:2310.07704 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.328, + 0.788, + 0.383 + ], + "angle": 0, + "content": "65. Yu, F., Liu, K., Zhang, Y., Zhu, C., Xu, K.: Partnet: A recursive part decomposition network for fine-grained and hierarchical shape segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9491-9500 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.383, + 0.788, + 0.411 + ], + "angle": 0, + "content": "66. Yuan, Y., Li, W., Liu, J., Tang, D., Luo, X., Qin, C., Zhang, L., Zhu, J.: Osprey: Pixel understanding with visual instruction tuning (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.411, + 0.788, + 0.466 + ], + "angle": 0, + "content": "67. Zhao, J., Li, J., Cheng, Y., Sim, T., Yan, S., Feng, J.: Understanding humans in crowded scenes: Deep nested adversarial learning and a new benchmark for multi-human parsing. In: Proceedings of the 26th ACM international conference on Multimedia. pp. 792-800 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.466, + 0.788, + 0.508 + ], + "angle": 0, + "content": "68. Zheng, S., Yang, F., Kiapour, M.H., Piramuthu, R.: Modanet: A large-scale street fashion dataset with polygon annotations. In: Proceedings of the 26th ACM international conference on Multimedia. pp. 1670-1678 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.551 + ], + "angle": 0, + "content": "69. Zhou, B., Zhao, H., Puig, X., Fidler, S., Barriuso, A., Torralba, A.: Scene parsing through ade20k dataset. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 633-641 (2017)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.551 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/SPIN_ Hierarchical Segmentation with Subpart Granularity in Natural Images/88afeb42-c98d-4785-a2a2-08cde2fa67ab_origin.pdf b/2024/SPIN_ Hierarchical Segmentation with Subpart Granularity in Natural Images/88afeb42-c98d-4785-a2a2-08cde2fa67ab_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..773b24a82f257d148f75c9c3eb7d2d4f5fa9357a --- /dev/null +++ b/2024/SPIN_ Hierarchical Segmentation with Subpart Granularity in Natural Images/88afeb42-c98d-4785-a2a2-08cde2fa67ab_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a3b853bc05b9cad96ef1c009220ede068abb69153cc0af5a083cbcecfb1ef5c5 +size 6963649 diff --git a/2024/SPIN_ Hierarchical Segmentation with Subpart Granularity in Natural Images/full.md b/2024/SPIN_ Hierarchical Segmentation with Subpart Granularity in Natural Images/full.md new file mode 100644 index 0000000000000000000000000000000000000000..938d98c7f31f2a375b9533039ccfd2b0e9c8121f --- /dev/null +++ b/2024/SPIN_ Hierarchical Segmentation with Subpart Granularity in Natural Images/full.md @@ -0,0 +1,294 @@ +# SPIN: Hierarchical Segmentation with Subpart Granularity in Natural Images + +Josh Myers-Dean $^{1}$ , Jarek Reynolds $^{1}$ , Brian Price $^{2}$ , Yifei Fan $^{2}$ , and Danna Gurari $^{1,3}$ + +1 University of Colorado Boulder, 2 Adobe, 3 University of Texas at Austin + +Abstract. Hierarchical segmentation entails creating segmentations at varying levels of granularity. We introduce the first hierarchical semantic segmentation dataset with subpart annotations for natural images, which we call SPIN (SubPartImageNet). We also introduce two novel evaluation metrics to evaluate how well algorithms capture spatial and semantic relationships across hierarchical levels. We benchmark modern models across three different tasks and analyze their strengths and weaknesses across objects, parts, and subparts. To facilitate community-wide progress, we publicly release our dataset at https://joshmyersdean.github.io/spin/index.html. + +# 1 Introduction + +When people discuss hierarchical image analysis tasks they are typically talking about one of two approaches: is-a relationships [27,28], which treat categories at different abstraction levels (e.g., a Subaru is a car), or is-part-of relationships, which concentrate on dividing objects into their constituent parts [13, 14, 19] (e.g., a door is part of a car). While the former has been widely explored, the latter—focusing on object decomposition—has received limited attention in computer vision research. Within segmentation research, the focus of this paper, is-part-of relationships have primarily been explored only for part-whole hierarchies, ignoring finer-grained details, such as subparts (i.e., parts of parts). + +Research to subpart-level segmentation granularity is hindered by a scarcity of data. While synthetic 3D datasets [14, 19, 35, 63] could be used to infer hierarchical segmentations to subpart granularity, it is well-known that models developed with synthetic data typically generalize poorly to natural images (i.e., images taken by a camera) [8]. This lack of annotated natural data has meant that the few models designed to generate subpart granularity segmentations [11, 52, 60] could only be evaluated qualitatively on a small number of examples (rather than quantitatively at scale). + +Our work facilitates the development of algorithms for subpart granularity in three key ways. First, we collect over 102,000 subpart segmentations across 203 diverse subpart categories to expand upon PartImageNet [17] and create the SubPartImageNet (SPIN) dataset—the first to offer subpart annotations for natural images. We release this dataset publicly to foster community progress. + +![](images/430260e209de2ae28810f30abe916d3f821a96710724b757d244b2066cb0766e.jpg) +(a) + +![](images/a9a305fa1738e86babc7dab2b438910eef6fd0c0fc3ac9d79eedbf22688291cd.jpg) +(b) + +![](images/d9bd6d7ea9e6f65fb3986e731fb31ec67a4de54295a7844d9b3ec78976525854.jpg) +(c) +Fig. 1: Overview of the diversity of SPIN. Panels (a) and (b) depict subparts unique to specific object class members, such as a roll cage in a car and a shell in a turtle. Panels (c) and (d) illustrate the variability in the number of subparts per object of the same class, with examples of 13 and 6 subparts. Panels (e) and (f) highlight the disparity in image area coverage by different subparts, such as a bottle label (large) versus quadruped claws (tiny). + +![](images/b63a593494b211050330e7ceab3339405c8447b513a34c88bb3c137b2b6ed62a.jpg) +(d) + +![](images/0bdb3c52360b4667684fcb75b04100d30678e1454ab7d7a7fb22f7502d1b00cf.jpg) +(e) + +![](images/624f1cc272c0a9e7caea6832adeb66eb3860a38a51e6c2b77ac9b7337db41618.jpg) +(f) + +Examples of annotated subparts are shown in Fig. 1. Second, we introduce two novel metrics to address a shortcoming of the prevailing evaluation methods for hierarchical segmentation, which only assess each granularity level in isolation. Our newly proposed metrics are intended to work alongside these traditional metrics (e.g., Intersection over Union) and assess how well models capture spatial and semantic relationships across hierarchical levels. Our third key contribution is that we benchmark modern models on SPIN across objects, parts, and subparts, thereby presenting the first comprehensive quantitative evaluation of subpart performance and key areas needed for future improvement. + +Advancing hierarchical segmentation to subpart granularity could significantly benefit both research and societal applications. One potential use is in generating more detailed image descriptions, such as for augmented reality experiences, answering visual questions, captioning images, or visual storytelling. Similarly, this work could be the foundation for individuals to (1) interactively learn how to speak about finer-grained entities for the first time (for children) or in a new language and (2) recall forgotten words, such as when facing temporary or chronic memory injuries or disabilities. More deeply nested, hierarchical segmentation representations can also enrich tactile visual discovery [25, 42], including for use in screen readers employed by visually impaired individuals (e.g., Apple's object-based exploration feature [1]). Similarly, we believe finer-grained information could facilitate improvements for related problems, including image/video retrieval, image/video editing, automatic magnification, and robotics. + +# 2 Related Works + +Datasets with Hierarchically Segmented Objects. Several datasets provide segmentations showing how objects are hierarchically decomposed into their recursively nested parts. A few do so without semantic labels specifying the categories of segmented content, including the pioneering Berkeley dataset published in 2001 [36] and the large-scale SA-1B dataset released in 2023 [22]. Our work, in contrast, focuses on hierarchically segmenting objects with category + +labels. Already, numerous datasets semantically decompose objects and their parts in natural images [7, 14, 15, 17, 21, 31, 45, 50, 56, 61, 67-69] and synthetic images [5, 19, 20, 35, 39]. One dataset, (i.e., ADE20K [69]) even provides subpart annotations for $10\%$ of its objects, however its lack of exhaustive annotations for objects and their associated parts [52] impedes it benefit for evaluating prediction models. Extending prior work, we introduce the first exhaustively labelled hierarchical semantic segmentation dataset with subpart annotations for natural images by extending one of the largest part-based datasets [17]. + +Hierarchical Semantic Segmentation Evaluation. Most evaluation protocols for hierarchical segmentation algorithms assess each hierarchy level independently, such as by reporting for object and part categories separately their Intersection over Union scores [9, 23, 31, 32, 37, 40, 52, 55, 57, 58, 61, 65], Average Precision scores [45, 51], and Panoptic Quality scores [14, 29, 30, 52, 60]. The one exception is Hierarchical Panoptic Quality [52] which recursively measures the extent to which a predicted hierarchy is complete (e.g., a wheel is a part of a car). Complementing prior work, we introduce evaluation metrics for assessing how well a model captures spatial and semantic hierarchical relationships, such as if a part is perfectly contained in it's whole. + +Hierarchical Semantic Segmentation Algorithms. Thus far, the focus of methods that predict all entities in a hierarchical decomposition largely has centered on predicting just objects and parts [16, 51, 58], aligning with existing datasets which only recently increasingly have added part segmentations to object segmentations. Only a few works, HIPIE [60], VDT [11], Semantic-SAM [26], and ViRReq [52] $^{1}$ , examined predicting more layers of a hierarchy to also include subparts. However, these works relied on qualitative assessments for subpart predictions due to a lack of annotated datasets supporting quantitative assessments. Complementing prior work, we quantitatively evaluate HIPIE and over 20 model variants to show their performance on our new dataset challenge across all three layers of the hierarchy (i.e., object, parts, and subparts). + +# 3 SPIN Dataset + +We now introduce and characterize our new subpart semantic segmentation dataset that we call SPIN, which is short for SubPart ImageNet. + +# 3.1 Dataset Creation + +Data Source. Our dataset extends an existing dataset that provides both object and associated part annotations: PartImageNet [17]. It includes 24,080 natural images across 158 ImageNet categories (112 non-rigid and 38 rigid, e.g. animals and vehicles respectively). The authors of PartImageNet used the + +WordNet [38] taxonomy to establish these categories and their 11 parent supercategories. The authors then identified 40 part categories pertinent to the supercategories and annotated these in all images. Every image in PartImageNet contains one semantically segmented object, and we used only those with at least one segmented part. Due to our limited annotation budget, we capped each supercategory at 1,200 images. This resulted in the following amounts per supercategory: 311 - Aeroplane, 483 - Bottle, 559 - Boat, 634 - Bicycle, and 1,200 each for Biped, Bird, Car, Fish, Quadruped, Reptile, and Snake. + +Subpart Category Selection. We identified 206 subpart categories to segment for 34 PartImageNet part categories that we identified as being decomposable into subparts. Of these, 168 subpart categories generalize across many part categories and 38 categories pertain to only a few. For example, in our dataset, mouths generalize while shells don't, with most reptiles having heads with mouths but only one type of reptile having a body with a shell (i.e., turtles). We show examples of specific subparts in Fig. 1(a-b). A histogram showing how many subparts are assigned to each of the 34 part categories is shown in Fig. 2, and a full list is provided in the supplementary materials. + +We created our subpart taxonomy through a multi-step process. We identified the general subpart categories by prompting GPT-4 [2] to list expected subparts for every object-part category in PartImageNet, and then three authors edited the list to exclude non-visible subparts (e.g., a skull). We identified specific subpart categories through manual inspection of at least 15 images per object category, excluding any categories that had ambiguous subpart decompositions; for example, it's not obvious to a lay person what boundaries to use when decomposing a tail into a tip, shaft, and base. + +Subpart Annotation Task Design. We created a task interface for collecting all subpart annotations. It presents each image-object pair alongside one of the object's nested parts, outlined with an overlaid polygon. For each part, the task + +![](images/929d7239c260f06f0a3fd6cef87c8d6ccb03ae0f99ab0a0413ed4c1f6452ead6.jpg) +Fig. 2: Histogram the number of unique subpart category labels for each of the 34 part categories. (Aero=Aeroplane; Quad=Quadruped) + +
#Images#Object Cat#Part Cat#Sub-Part Cat
10,38711 (10,387)40 (29,818)203 (106,324)
+ +Table 1: SPIN Composition: Each column lists category counts alongside the total number of annotated entities within those categories. (Cat=Categories) + +interface has multiple steps. First, users are asked, "Can you locate the -- enclosed in a polygon?" If the target content is present, annotators are then asked for each possible subpart, "Can you locate any subpart> on the --?" If the response is yes, annotators segment the subpart. The interface supports segmentation by collecting a series of clicked points to create a connected polygon. It also supports annotating multiple polygons in three scenarios: (1) multiple instances of a subpart (e.g., two eyes), (2) subparts with holes (e.g., coiled snakes), and (3) occlusions breaking a subpart into multiple, disconnected pieces. + +Annotation Collection. We hired 18 highly trusted annotators from Amazon Mechanical Turk (AMT) who we had vetted through repeated employment for previous segmentation tasks. During data collection, we ensured high annotation quality through five methods: an onboarding qualification test, detailed instructions, live "office hours" during annotation periods, phased task roll-out for worker feedback, and continuous inspection of submitted results. + +Dataset Splits. Our dataset adheres to the PartImageNet training, validation, and testing splits of $85\%$ , $5\%$ , and $10\%$ of the data, respectively. This results in 8,828 training, 519 validation, and 1,040 test images in our SPIN dataset. + +# 3.2 Dataset Analysis + +We now characterize SPIN's overall composition as well as its subparts. + +Overall Dataset Composition. We characterize SPIN with respect to the following seven factors: number of images, number of object categories, total number of annotated objects, number of part categories, total number of annotated parts, number of subpart categories, and total number of annotated subparts. Results are shown in Table 1. As shown, our contribution of 203 entity categories is over a five-order-of-magnitude increase compared to the 40 categories added for PartImageNet [17], while providing nearly an order of magnitude more annotations (i.e. 106,324 semantic annotations versus 11,960 part segmentations in PartImageNet). + +Subpart Statistics. We next compute the mean number of subparts per part and characterize the typical appearance of subparts with respect to five metrics: + +![](images/e27bf80c8da0872c26717dfabd74cd446b5750be21a6decec092c84e034bd0f2.jpg) +(a)InstancesPerPart + +![](images/f3631d77a8daa27e9b50bcdbe6533ddaeff8f554f0d8d498f9056f68815eb1c0.jpg) +(d) Image Coverage + +![](images/02c1d19954c0b8a8be2278aa0bfb6aeee8435c36109c719e6cf6ad2b26aa2a24.jpg) +(b) Extent + +![](images/e4a9c85556fe4b6ab9637225151d7bd2a5aff75e6a00e8d29ce8011715a9cd65.jpg) +(e) Object Coverage + +![](images/ba7dced9c2690df2deb057d8007b7976724e45c71f87630bde54df26fc2e5d07.jpg) +(c) Boundary Complexity + +![](images/5949391db2cbc1db7b5857255d943d57df4500982fb2a298c541327733e7f32e.jpg) +(f) Part Coverage +Fig. 3: Boxplots showing the distribution of subpart image occupation (left) and boundary complexities per part, per object (right). The blue lines represent medians, bottoms and tops of each box represent the 25th and 75th percentile values respectively, and whiskers represent the most extreme data points not considered outliers. Overall, SPIN's subparts take up a relatively small number of pixels per image, while featuring a range of geometric complexity. (Aero=Aeroplane; Quad=Quadruped) + +- Boundary complexity: ratio of a subpart's area to the length of its perimeter (i.e., isoperimetric quotient). Values range from 0 (highly jagged boundary) to 1 (circular). +- Extent: the ratio of the area of a contour to the area of its bounding box. Values are in $(0,1]$ , where values approaching 0 mean that a contour occupies little area in its bounding box (e.g., a thin diagonal line) and 1 means that a contour is perfectly contained (e.g., a square). +- Image coverage: fraction of image pixels occupied by the subpart. +- Object coverage: fraction of parent object pixels occupied by the subpart. +- Part coverage: fraction of parent part pixels occupied by the subpart. + +For all six factors, we report both the overall mean as well as the mean with respect to each object category. Results are shown in Fig. 3. + +Objects tend to have 2 to 5 subparts per part (i.e. Fig. 3a; 25th to 75th percentile values), with a total of 4 to 15 subparts. There is diversity across object categories with aeroplanes typically having small variability with only 1 to 2 subparts per part (i.e. 25th to 75th percentile range) while car typically exhibits great variability with 2 to 8 subparts per part (i.e. 25th to 75th percentile range), as exemplified in Fig. 1(c-d). We attribute the large range of subpart per part counts to two factors: the intrinsic diversity in cars and the viewing angle. For example, a go-cart viewed from the front will have no windshield, no windows, and two tires, while a bus viewed from a front-right angle will have a windshield, tires, and windows. Such object-specific patterns around the expected prevalence of subparts could be predictive cues for models decomposing detected objects into their recursively nested subparts. + +In terms of shape, most subparts' contours occupy the majority of their bounding box (i.e., Fig. 3b, median extent value $\sim 0.7$ ) with moderate boundary complexity (i.e., Fig. 3c, median boundary complexity values around 0.5). Still, we observe high variability for both metrics that we again attribute to the intrinsic diversity in each object category as well as the viewing angle. For example, a bottle cap or a bicycle tire viewed from the side will be an ellipse versus a nearly perfect circle when viewed from the top (i.e., occupying more of its bounding box). Altogether, these findings highlight that our dataset encourages the design of models that can account for a large range of subpart shapes. + +Regarding the relative area occupied by each subpart, we find that all subparts occupy small portions of the parent parts, parent objects, and entire image. For example, across every object category, subparts occupy less than $6\%$ of the image for at least $75\%$ of subpart instances (Fig. 3d). Additionally, the subparts tend to occupy less than $20\%$ of their parent parts (Fig. 3e; Overall) and less than $5\%$ of their parent objects (Fig. 3f; Overall). When adopting size thresholds introduced for the MSCOCO dataset [33], where $32^2$ and $96^2$ are thresholds determining whether an object is small, medium, or large, we find $54.10\%$ (57,525) of SPIN's subpart annotations qualify as small, $38.08\%$ (40,488) as medium, and only $7.82\%$ (8,311) as large. Most subparts are already considered small according to mainstream research despite that we annotated subparts for a dataset known to typically contain a single, large prominent object in the image. Consequently, embedded in our problem is the well-known challenge of small entity detection [53], which will only grow as subparts are explored for more complex scenes where less prominent objects must also be hierarchically decomposed. + +When comparing subparts belonging to the rigid object categories (i.e., human-made) versus non-rigid object categories (i.e., animals), we found no clear distinction. For instance, the mean number of subparts per part (Fig. 3a) is the smallest amount with least variability in aeroplanes and the greatest amount with greatest variability in cars. While prior work has emphasized distinguishing between these two extremes [17], our analysis suggests that rigid and non-rigid objects appear quite similar at the subpart level. + +# 4 Evaluating Hierarchical Consistency + +We facilitate assessing how well relationships across varying levels of granularity are captured for hierarchical segmentation by introducing two metrics that account for spatial and semantic relationships respectively. + +We first introduce the Spatial Consistency Score (SpCS), which focuses on structural hierarchical alignment rather than semantic alignment. It indicates the proportion of a child's segmentation (e.g., subparts) that is contained in its super-region's segmentation (e.g., part, object), with values ranging from 0 (no containment) to 1 (perfect containment). It leverages binarized segmentations distinguishing the foreground entity from the background. Formally, it is com + +puted as the average containment ratio across all pairs as follows: + +$$ +S p C S = \frac {1}{| \mathcal {R} |} \sum_ {\text {(c h i l d , p a r e n t)} \in \mathcal {R}} \underbrace {\frac {| \text {c h i l d} \cap \text {p a r e n t} |}{| \text {c h i l d} |}} _ {\text {c o n t a i n m e n t r a t i o}} \tag {1} +$$ + +where $\mathcal{R}$ denotes the set of prediction pairs with a ground truth parent-child relationship within a hierarchical structure. $^3 |\mathcal{R}|$ denotes the total number of parent-child pairs, $|\mathrm{child}\cap \mathrm{parent}|$ is the number of pixels in the intersection of the child segmentation with the parent, $|\mathrm{child}|$ is the number of pixels in the child segmentation. + +We next introduce the Semantic Consistency Score (SeCS), which focuses on semantic alignment by measuring how well predictions at different levels (subpart, part, object) logically correspond. For example, an eye (subpart) should correspond to a head (part) and, by extension, to a quadruped (object), rather than illogical associations like a windshield (subpart) with a head (part) or a bottle (object). Let $FG_{S}$ , $FG_{P}$ , and $FG_{O}$ be the sets of non-background predictions in subpart, part, and object predictions, respectively. Define $X = FG_{S} \cap FG_{P} \cap FG_{O}$ as the intersection of these predictions, representing pixels with consistent foreground labels across levels. For each pixel $x \in X$ , with predicted categories $C(S_x)$ , $C(P_x)$ , and $C(O_x)$ for subparts $(S_x)$ , parts $(P_x)$ , and objects $(O_x)$ , respectively, category entailment is evaluated using $\mathbb{1}[C(S_x), C(P_x)]$ and $\mathbb{1}[C(P_x), C(O_x)]$ , where $\mathbb{1}$ is an indicator function returning 1 if the hierarchical entailment is correct, using ground truth relations, otherwise 0. We compute the entailment measure $M(x)$ as: + +$$ +M (x) = \left\{ \begin{array}{l l} 1 & \text {i f} \mathbb {1} [ C (S _ {x}), C (P _ {x}) ] \wedge \mathbb {1} [ C (P _ {x}), C (O _ {x}) ], \\ 0 & \text {o t h e r w i s e}, \end{array} \right. \tag {2} +$$ + +where $\wedge$ is the logical and operator. SeCS is the mean of $M(x)$ across all $x\in X$ . Values range from 0 (no semantic coherence) to 1 (perfect semantic coherence). + +# 5 Algorithm Benchmarking + +We now assess modern models' ability to account for hierarchical decompositions to the subpart granularity. + +# 5.1 Open-Vocabulary Localization in a Zero-Shot Setting + +We first evaluate models for open-vocabulary localization in zero-shot mode across three hierarchy levels: object, part, and subpart. This approach avoids the drawbacks of the alternative options of training models from scratch or finetuning, including the computational expense and risks of overfitting to the target task and so diminishing generalizable knowledge. + +Models. We benchmark 11 model variants. + +Most related to our work is HIPIE [60], which is designed for open-vocabulary hierarchical segmentation to the subpart level. It takes as input a list of all candidate categories that can appear and then predicts which categories are present where in the given image. We test two variants with different backbones: ResNet-50 [18] and ViT-H [12]. + +We also evaluate 10 variants of foundation models that, by design, are intended for use on a variety of downstream tasks. Five produce pixel-wise segmentation masks: LISA-7B [24], LISA-13B [24], GLaMM [46], and PixelLLM-7B [48], and PixelLLM-13B [48]. The other five can only produce bounding boxes: CoGVLM [59], Ferret-7B [64], Ferret-13B [64], Shikra [6], and Kosmos2 [43]. For these models, we analyze their upper bound of their potential performance by notifying the model what object category needs to be located in a given image. We examine two types of requests: "Can you please locate the in the image?" for objects, and "Can you please locate the subpart> of the in the image?"4 for each part and subpart. We evaluate with these prompts a Specific category by using the original WordNet category from the 158 ImageNet categories as well as a General category by referring to the 11 super-categories. Consequently, we test each model with four prompts. + +Evaluation Metrics. We use two evaluation metrics for all models: (1) Intersection over Union (IoU) to assess segmentation performance for objects, parts, and subparts independently and (2) our Spatial Consistency Score (SpCS) to assess structural coherence across object, part, and subpart levels. We also use our new Semantic Consistency Score (SeCS) to assess semantic coherence between predicted hierarchical levels for the only relevant model, HIPIE (it predicts from a list of candidate categories rather than for the target, single category). + +# Localization Results (mIoU). Results are shown in Table 2. + +We observe a consistent trend across all models: localization performance is best for object segmentation, followed by increasingly worse performance for more granular categories of parts and then subparts. The best score for localizing objects is 86 followed by 40 for parts and 14 for subparts. The worst scores come from the only model directly designed for hierarchical segmentation, HIPIE. It does worse across all granularity levels than the foundation models; e.g., for subpart localization, 1 vs 3 from the worst-performing foundation model, Kosmos2. We attribute the poor performance to HIPIE's limited training data (e.g., limited part vocabulary), which only included part-level categories from Pascal Parts [7]. Moreover, HIPIE's heuristic grouping assumes that parts are the sum of their subparts, which is not the case in SPIN (e.g., unlabelled "space" exists between a cheek an eye in a head). + +We explored to what extent entity size is correlated to the resulting IoU scores by using linear regression for the better-performing foundation models. + +
Method# Params.mIoUsmIoUPmIoUOSpCS - AvgSpCS - S2PSpCS - P2O
SpecificGeneralSpecificGeneralSpecificGeneralSpecificGeneralSpecificGeneralSpecificGeneral
HIPIE [60] (R-50)200M0.800.108.057.2251.3613.8797.8298.7710010095.6497.54
HIPIE [60] (ViT-H)800M0.900.927.218.2366.7721.6998.0499.2510098.5096.0896.06
PixelLLM [48]7B9.089.2432.3731.3783.9379.8773.7982.4669.5779.2089.4990.73
PixelLLM [48]13B9.5310.1332.0432.9682.9079.9684.8683.0682.1380.8792.6088.90
LISA [24]7B9.619.1427.7227.2886.2383.3679.3077.4474.6172.9392.1089.78
LISA [24]13B11.5211.5532.2931.3687.7885.4586.2879.9882.8775.7096.0292.41
GLaMM [46]7B11.0311.0039.4140.0086.2986.3187.1287.7584.2184.9695.7295.38
GLaMM - FT7B24.2524.5659.3760.7686.4291.0887.1387.9575.9385.1690.2396.04
Ferret [64]7B7.227.3025.6926.5447.8447.9878.7073.0246.5065.6893.6594.60
Ferret [64]13B6.376.2523.6723.6547.9947.5081.0477.3048.7070.9894.1795.40
CoGVLM [59]17B13.9414.2938.1338.5646.4743.6576.3377.2072.3773.3286.1987.59
Shikra [6]7B8.418.7226.5027.4245.5029.1468.5767.0163.8266.7681.8567.89
Kosmos2 [43]2.6B3.453.5919.0019.4848.6348.8982.7481.6078.3477.0396.0195.58
SAM [22]630M49.6149.6169.2369.2390.0690.0686.4286.4283.8583.8592.3092.30
SAM + CoGVLM630M19.8819.9850.9350.9077.9475.0966.7167.3380.5080.7360.6561.44
+ +Table 2: Performance of modern models with respect to localization at three granularity levels—subparts (mIoUs), parts (mIoUP), objects (mIoUO)—and hierarchical consistency at two levels—subparts to parts (SpCS-S2P), parts to objects (SpCS-P2O), and their mean (SpCS - Avg). Results are shown for hierarchical semantic segmentation models (described in Section 5.1), hierarchical object detection models (Section 5.1), and interactive segmentation models that perform category-agnostic localization (Section 5.2). The method families are separated by dashed lines in that order. Best performing scores are shown in bold. + +Due to space constraints, we provide further experimental details and results in the supplementary materials. In summary though, our findings suggest that size has little correlation to performance for objects but greater correlations for more granular levels of parts and subparts. We suspect that one contributing factor is the sensitivity of the IoU metric for entities occupying a small number of pixels, as small errors are especially prone to yield dramatic changes in the IoU scores for such cases. + +We next explored the impact of a prompt's category specificity on model performance by comparing performance for when general versus specific terms are employed. Most models perform worse when trying to locate objects categorized under general terms (e.g., 'quadruped') rather than specific ones (e.g., 'dog'). For example, CoGVLM performs 2.82 percentage points (pp) worse with general categories compared to specific categories. Similarly, the 13B variant of LISA has a 2.33pp performance drop and its 7B counterpart has a 2.87pp performance drop. The exceptions to this are GLaMM, Ferret, and Kosmos2, whose performance are relatively stable and unaffected by the level of category specificity. A potential reason is that these models utilize training data that explicitly use both general and specific terms. For example, GCG [46], a source of training data for GLaMM, refers to entities with both abstract and specific terms (e.g., person and toddler). Overall, this variance highlights the challenge in choosing an appropriate level of category specificity for objects. However, the story changes when trying to locate parts and subparts, with the performance gap between general and specific categories falling to within 1pp. An interesting direction for future work is better understanding how and why an entity's granularity level influences models' abilities to overcome abstraction challenges. + +We also examine how the training data of the Large Language Models (LLM) Llama [54], since it serves as the basis for the majority of foundation models evaluated in our study. Specifically, we analyze the empirical frequency (i.e., total occurrence) of uni-grams (e.g., eye) and n-grams (e.g., eye of the quadruped) within the dataset using the $\infty$ -gram API [34], which uses byte array pattern matching to find all occurrences of a given BPE tokenized [49] $n$ -gram. Due to space constraints, results are provided in the supplementary materials. In summary though, we observe a decline in average uni-gram occurrence frequency when increasing granularity (e.g., part to subpart). An interesting direction for future work is to remedy this imbalance, such as by designing parameter-efficient methods to restructure the text embedding space of LLMs to recognize hierarchical relationships, akin to MERU [10] restructuring CLIP [44]-space for image-level is-a relationships. + +Performance When Training on Our SPIN Dataset. We next examine the potential benefit of our SPIN training data for modern models. To do so, we fine-tuned the top-performing GLaMM algorithm [46] for all hierarchy levels in SPIN, and we refer to this variant as GLaMM-FT. Results are shown in Table 2. + +Overall, we observe a considerable performance boost. For example, when prompting with general categories, we observe a 13.56 percentage point (pp) increase (123% relative increase) for subparts, a 20.76pp increase (51.90% relative increase) for parts, and a 4.77pp increase for objects (5.51% relative increase). While such performance gains are promising, we still observe very low absolute scores for subparts, indicating their challenge for modern models. We suspect a promising direction for future work is providing cost-effective ways to further increase the amount of available training data, such as through synthetic data creation and 2D projections from 3D models. + +Hierarchical Consistency Results. SpCS results are shown in Table 2. Similar trends are observed for part-object and subpart-part relationships. Part-object pairs consistently exhibit high SpCS-P2O scores, indicating models tend to correctly predict parts within their corresponding whole objects. This provides evidence that models' have some understanding of spatial containment among part-object relationships. A potential reason for this is that models are indirectly trained on part-whole relationships, because they are already taught that an object is a part of a scene. The aforementioned trend is slightly worse when examining consistency between subparts and parts (i.e., SpCS-S2P), with scores varying from 46.50 (i.e., Ferret 7B) to 84.21 (i.e., GLaMM) when employing general terms for objects and from 65.68 (i.e., Ferret 7B) to 84.96 (i.e., GLaMM) when employing more specific categories for objects. This variability, coupled with low mIoU scores for subparts, reveals that models often predict subparts outside their intended part boundaries, underscoring poor performance in properly contextualizing subparts within an object. + +# 5.2 Interactive Segmentation for Vocabulary-Free Localization + +We next assess to what extent models can localize entities at all hierarchy levels in our dataset if they are notified where to look in an image. This sets an upper bound of what is possible. To do so, we benchmark the state-of-the-art interactive segmentation model, SAM [22], which takes as input a coarse marking on an image specifying where to create the segmentation. We use bounding boxes as our coarse image markings, because they are shown to outperform alternatives [41] when using SAM. We explore the upper bound performance for this model by using the ground truth object, part, and subpart segmentations to generate the bounding boxes (i.e., by using $[x_{min}, y_{min}, x_{max}, y_{max}]$ for each entity's pixels). Results are shown in Table 2. + +Despite the idealized guidance from the bounding box prompts, SAM's performance is still imperfect. As granularity increases, localization performance consistently decreases by about 20pp when going down a granularity level (e.g., objects to parts), revealing that SAM struggles more as segmentation requests become more fine-grained. Still, our findings demonstrate SAM's potential for yielding considerably higher quality hierarchical segmentations than observed from existing localization models benchmarked in the previous section. Compared to the top-performing zero-shot segmentation-based model (i.e., GLaMM), SAM achieves a boost of 38.61pp for subparts, 20.23pp for parts, and 3.75pp for objects, alongside similar SpCS scores. + +We also evaluate a fully-automated system with SAM, by incorporating bounding box predictions from the top-performing bounding box producing model, CoGVLM. Results are shown in Table 2. We similarly observe that it outperforms the leading zero-shot segmentation method, GLaMM, in segmenting subparts and parts (9.24 percentage point (pp) boost for subparts and 10.87pp for parts). We attribute these performance gains to SAM's ability to leverage a location prior (e.g., bounding box), whereas models like GLaMM rely solely on language priors. In contrast, this approach performs worse than the leading zero-shot segmentation method, GLaMM, for objects, with a 11.22pp decrease. We attribute this performance drop to CoGVLM's inferior performance in consistently and accurately detecting bounding boxes. + +# 5.3 Recognizing Hierarchical Semantics in a Zero-Shot Setting + +We finally evaluate the ability of models to recognize hierarchical semantic labels for objects in images (i.e., without localization). We assess this in an idealized scenario in which we notify models what to look for as well as where to look in the image. This sets an upper bound of what may be possible with such models as they are explicitly told where to look. + +Models. We evaluate eight off-the-shelf model variants: CoGVLM [59], Ferret-7B [64], Ferret-13B [64], Shikra [6], Osprey [66], Kosmos2 [43], ViP-Llava-7B [4], and ViP-Llava-13B [4]. All prompts to the models resemble the following basic request: "Is there a in the ?" for objects and "Is there + +
Method# Params.Prompt\( \mathrm{mACC}_S \)\( \mathrm{mACC}_{SS} \)\( \mathrm{mACC}_P \)\( \mathrm{mACC}_{PS} \)\( \mathrm{mACC}_O \)\( \mathrm{mACC}_{OS} \)
Shikra [6]7BBbox66.1261.2879.6774.5677.0482.05
Ferret [64]7BBbox59.0157.1959.9566.1786.7889.21
Ferret [64]13BBbox50.3169.4261.7267.8488.9389.02
CoGVLM [59]17BCoT42.1645.5481.7583.3187.1583.67
Kosmos2 [43]2.6BSG79.1264.3274.2368.5384.5578.78
Osprey [66]7BMask77.4355.9586.3081.2096.4384.85
ViP-Llava [4]7BMask53.5758.5264.1364.9671.6265.25
ViP-Llava [4]13BMask94.7897.4399.9099.8098.3395.06
+ +Table 3: Accuracy metrics for object hierarchical decomposition: general (mACC $S$ , $\mathrm{mACC}_P$ , $\mathrm{mACC}_O$ ) and specific ( $\mathrm{mACC}_{SS}$ , $\mathrm{mACC}_{PS}$ , $\mathrm{mACC}_{OS}$ ) categories across subparts, parts, and objects. (CoT = Chain of Thought; SG = Self-Grounding) + +a $<(sub)part>$ of the $$ in the $?$ " for parts and subparts. Regions are specified either with segmentations (ViP-Llava, Ferret, Osprey) or bounding boxes (CoGVLM, Shikra, Kosmos2). We generate these regions directly from the ground truth annotations (e.g., bounding boxes as described in Section 5.2) for objects, parts, and subparts. We prompt models for each category known to be in the region, one at a time. An answer is 'yes' if 'yes' is present in the model's output, and otherwise 'no'. As done in Section 5.1, we test both general and specific categories in the prompts. + +Evaluation Metrics. We use accuracy to evaluate at each granularity level. + +# Results. Results are reported in Table 3. + +While most models struggle overall to recognize categories in images (despite being told where to look in the images), one model has nearly perfect accuracy: ViP-Llava-13B. One hypothesis for this atypically strong performance is the model lacks understanding and instead always answers yes when prompted about the presence of categories. We explored this with further experiments using adversarial examples that are instead expected to always answer "no". However, our findings from these experiments are none-conclusive. We observe the model can answer "no", but it also suffers in this case from a significant decrease in accuracy for the adversarial prompts. This performance drop could be due to a separate issue that the model struggles with negation (i.e., the reversal curse [3]). Altogether, this evaluation represents the best-case scenario for performance; it is likely to fall further in real-world applications due to imperfect human inputs or model-generated segmentations, both of which we have shown to exhibit poorer performance at finer granularity levels, as discussed in Sections 5.1 and 5.2. + +We observed mixed outcomes regarding which granularity levels are most difficult to recognize. For Ferret-13B and Osprey, an increase in granularity correlates with improved accuracy; e.g., transitioning from subparts to objects using general category names results in a 28.62pp increase for Ferret-13B and a 19.00pp + +increase for Osprey. Conversely, for most other models, object identification consistently outperforms subpart identification. The differential performance across granularity levels may partly stem from the disparate prevalence of tokens in the training data across different granularity levels, particularly within the Llama dataset (as previously discussed). For parts and subparts, the mixed results observed—where parts sometimes outperform subparts (e.g., CoGVLM) and vice versa (e.g., Ferret-7B)—could be attributed to their relative frequency in the training data. Future research could focus on examining how foundation models differentiate responses based on prompt granularity and on enriching datasets like GRIT [64] with hierarchical relationships to improve model understanding of intra-entity relationships. Additionally, future work could explore enhancing hierarchical-relationship understanding so success at different hierarchical levels can facilitate finding entities at other granularity levels. + +We also observed mixed outcomes when comparing the efficacy of specific versus general categories. Half the models performed better using general terms for subparts, with a similar pattern observed in parts and objects. A potential reason for this discrepancy, compared to the relatively consistent results in Sec 5.1, is that localization models often rely on feature similarities, such as CLIP [44] features. Even if a model does not capture every nuance of a category query, it still attempts to find a match based on overall similarity to known categories (e.g., mini-van vs. van) and is less likely to reject user queries [62]. Conversely, interactive understanding models, designed to maximize accuracy, may strictly search for precise category matches within their trained representations. Given our task's structure (i.e., answers are restricted to 'yes' or 'no'), these models are predisposed to outright reject the query (i.e., answer 'no') if an exact match is absent, rather than proposing close or related category matches. + +# 6 Conclusion + +We introduce SPIN, the first dataset challenge for hierarchical segmentation at the subpart granularity in natural images. Our analysis reveals SPIN's characteristics, our two new evaluation metrics enable gauging algorithmic proficiency in capturing the spatial and semantic relationships across hierarchy levels, and our benchmarking across three tasks demonstrates that models struggle to recognize and segment subparts in SPIN, even under idealized guidance. We publicly release SPIN to encourage further advancements in hierarchical segmentation, sharing the dataset at https://joshmyersdean.github.io/spin/index.html. + +Acknowledgments. This work was supported by Adobe Research Gift Funds and utilized the Blanca condo computing resource at the University of Colorado Boulder. Josh Myers-Dean is supported by a NSF GRFP fellowship (#1917573). We thank the crowdworkers for contributing their time for the construction of SPIN and the authors of our benchmarked models for open-sourcing their work. + +# References + +1. Explore images. https://support.apple.com/guide/iphone/use-voiceover-for-images-and-videos-iph37e6b3844/ios +2. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023) +3. Berglund, L., Tong, M., Kaufmann, M., Balesni, M., Stickland, A.C., Korbak, T., Evans, O.: The reversal curse: Llms trained on" a is b" fail to learn" b is a". arXiv preprint arXiv:2309.12288 (2023) +4. Cai, M., Liu, H., Mustikovela, S.K., Meyer, G.P., Chai, Y., Park, D., Lee, Y.J.: Making large multimodal models understand arbitrary visual prompts. In: IEEE Conference on Computer Vision and Pattern Recognition (2024) +5. Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., et al.: Shapenet: An information-rich 3d moel repository. arXiv preprint arXiv:1512.03012 (2015) +6. Chen, K., Zhang, Z., Zeng, W., Zhang, R., Zhu, F., Zhao, R.: Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195 (2023) +7. Chen, X., Mottaghi, R., Liu, X., Fidler, S., Urtasun, R., Yuille, A.: Detect what you can: Detecting and representing objects using holistic models and body parts. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1971-1978 (2014) +8. Deitke, M., Han, W., Herrasti, A., Kembhavi, A., Kolve, E., Mottaghi, R., Salvador, J., Schwenk, D., VanderBilt, E., Wallingford, M., Weihs, L., Yatskar, M., Farhadi, A.: Robothor: An open simulation-to-real embodied ai platform. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020) +9. Deng, B., Genova, K., Yazdani, S., Bouaziz, S., Hinton, G., Tagliasacchi, A.: Cvxnet: Learnable convex decomposition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 31-44 (2020) +10. Desai, K., Nickel, M., Rajpurohit, T., Johnson, J., Vedantam, S.R.: Hyperbolic image-text representations. In: International Conference on Machine Learning. pp. 7694-7731. PMLR (2023) +1. Ding, M., Shen, Y., Fan, L., Chen, Z., Chen, Z., Luo, P., Tenenbaum, J.B., Gan, C.: Visual dependency transformers: Dependency tree emerges from reversed attention. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14528-14539 (2023) +2. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) +3. Geng, H., Xu, H., Zhao, C., Xu, C., Yi, L., Huang, S., Wang, H.: Gapartnet: Cross-category domain-generalizable object perception and manipulation via generalizable and actionable parts. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7081-7091 (2023) +4. de Geus, D., Meletis, P., Lu, C., Wen, X., Dubbelman, G.: Part-aware panoptic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5485-5494 (2021) + +15. Gong, K., Liang, X., Li, Y., Chen, Y., Yang, M., Lin, L.: Instance-level human parsing via part grouping network. In: Proceedings of the European conference on computer vision (ECCV). pp. 770-785 (2018) +16. He, J., Chen, J., Lin, M.X., Yu, Q., Yuille, A.L.: Compositor: Bottom-up clustering and compositing for robust part and object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11259-11268 (June 2023) +17. He, J., Yang, S., Yang, S., Kortylewski, A., Yuan, X., Chen, J.N., Liu, S., Yang, C., Yu, Q., Yuille, A.: Partimagenet: A large, high-quality dataset of parts. In: European Conference on Computer Vision. pp. 128-145. Springer (2022) +18. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) +19. Hong, Y., Li, Q., Zhu, S.C., Huang, S.: VLgrammer: Grounded grammar induction of vision and language. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1665-1674 (2021) +20. Hong, Y., Yi, L., Tenenbaum, J., Torralba, A., Gan, C.: PTR: A benchmark for part-based conceptual, relational, and physical reasoning. Advances in Neural Information Processing Systems 34, 17427-17440 (2021) +21. Jia, M., Shi, M., Sirotenko, M., Cui, Y., Cardie, C., Hariharan, B., Adam, H., Belongie, S.: Fashionpedia: Ontology, segmentation, and an attribute localization dataset. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16. pp. 316-332. Springer (2020) +22. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 4015-4026 (October 2023) +23. Koo, J., Huang, I., Achlioptas, P., Guibas, L.J., Sung, M.: Partglot: Learning shape part segmentation from language reference games. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16505-16514 (2022) +24. Lai, X., Tian, Z., Chen, Y., Li, Y., Yuan, Y., Liu, S., Jia, J.: Lisa: Reasoning segmentation via large language model. arXiv preprint arXiv:2308.00692 (2023) +25. Lee, J., Peng, Y.H., Herskovitz, J., Guo, A.: Image explorer: Multi-layered touch exploration to make images accessible. In: Proceedings of the 23rd International ACM SIGACCESS Conference on Computers and Accessibility. ASSETS '21, Association for Computing Machinery, New York, NY, USA (2021). https://doi.org/10.1145/3441852.3476548, https://doi.org/10.1145/3441852.3476548 +26. Li, F., Zhang, H., Sun, P., Zou, X., Liu, S., Yang, J., Li, C., Zhang, L., Gao, J.: Semantic-sam: Segment and recognize anything at any granularity. arXiv preprint arXiv:2307.04767 (2023) +27. Li, L., Zhou, T., Wang, W., Li, J., Yang, Y.: Deep hierarchical semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1246-1257 (2022) +28. Li, T., Gupta, V., Mehta, M., Srikumar, V.: A logic-driven framework for consistency of neural models. arXiv preprint arXiv:1909.00126 (2019) +29. Li, X., Xu, S., Yang, Y., Cheng, G., Tong, Y., Tao, D.: Panoptic-partformer: Learning a unified model for panoptic part segmentation. In: European Conference on Computer Vision. pp. 729-747. Springer (2022) + +30. Li, X., Xu, S., Yang, Y., Yuan, H., Cheng, G., Tong, Y., Lin, Z., Yang, M.H., Tao, D.: Panopticpartformer++: A unified and decoupled view for panoptic part segmentation. arXiv preprint arXiv:2301.00954 (2023) +31. Liang, X., Gong, K., Shen, X., Lin, L.: Look into person: Joint body parsing & pose estimation network and a new benchmark. IEEE transactions on pattern analysis and machine intelligence 41(4), 871-885 (2018) +32. Liang, X., Shen, X., Feng, J., Lin, L., Yan, S.: Semantic object parsing with graph lstm. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14. pp. 125-143. Springer (2016) +33. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014) +34. Liu, J., Min, S., Zettlemoyer, L., Choi, Y., Hajishirzi, H.: Infini-gram: Scaling unbounded n-gram language models to a trillion tokens. arXiv preprint arXiv:2401.17377 (2024) +35. Liu, Q., Kortylewski, A., Zhang, Z., Li, Z., Guo, M., Liu, Q., Yuan, X., Mu, J., Qiu, W., Yuille, A.: Cgpart: A part segmentation dataset based on 3d computer graphics models. arXiv preprint arXiv:2103.14098 (2021) +36. Martin, D., Fowlkes, C., Tal, D., Malik, J.: A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In: Proceedings Eighth IEEE International Conference on Computer Vision. ICCV 2001. vol. 2, pp. 416-423. IEEE (2001) +37. Michieli, U., Zanuttigh, P.: Edge-aware graph matching network for part-based semantic segmentation. International Journal of Computer Vision 130(11), 2797-2821 (2022) +38. Miller, G.A.: Wordnet: a lexical database for english. Commun. ACM 38(11), 39-41 (nov 1995). https://doi.org/10.1145/219717.219748, https://doi.org/10.1145/219717.219748 +39. Mo, K., Zhu, S., Chang, A.X., Yi, L., Tripathi, S., Guibas, L.J., Su, H.: Partnet: A large-scale benchmark for fine-grained and hierarchical part-level 3d object understanding. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 909-918 (2019) +40. Mo, K., Zhu, S., Chang, A.X., Yi, L., Tripathi, S., Guibas, L.J., Su, H.: PartNet: A large-scale benchmark for fine-grained and hierarchical part-level 3D object understanding. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019) +41. Myers-Dean, J., Fan, Y., Price, B., Chan, W., Gurari, D.: Interactive segmentation for diverse gesture types without context. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 7198-7208 (January 2024) +42. Nair, V., Zhu, H.H., Smith, B.A.: Imageassist: Tools for enhancing touchscreen-based image exploration systems for blind and low vision users. In: Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems. CHI '23, Association for Computing Machinery, New York, NY, USA (2023). https://doi.org/10.1145/3544548.3581302, https://doi.org/10.1145/3544548.3581302 +43. Peng, Z., Wang, W., Dong, L., Hao, Y., Huang, S., Ma, S., Wei, F.: Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824 (2023) + +44. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) +45. Ramanathan, V., Kalia, A., Petrovic, V., Wen, Y., Zheng, B., Guo, B., Wang, R., Marquez, A., Kovvuri, R., Kadian, A., et al.: Paco: Parts and attributes of common objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7141-7151 (2023) +46. Rasheed, H., Maaz, M., Shaji, S., Shaker, A., Khan, S., Cholakkal, H., Anwer, R.M., Xing, E., Yang, M.H., Khan, F.S.: Glamm: Pixel grounding large multimodal model. arXiv preprint arXiv:2311.03356 (2023) +47. RAYMOND W. GIBBS, J., Matlock, T.: Psycholinguistics and mental representations (2000) +48. Ren, Z., Huang, Z., Wei, Y., Zhao, Y., Fu, D., Feng, J., Jin, X.: Pixelm: Pixel reasoning with large multimodal model (2023) +49. Sennrich, R., Haddow, B., Birch, A.: Neural machine translation of rare words with subword units. arXiv preprint arXiv:1508.07909 (2015) +50. Song, X., Wang, P., Zhou, D., Zhu, R., Guan, C., Dai, Y., Su, H., Li, H., Yang, R.: Apollocar3d: A large 3d car instance understanding benchmark for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5452-5462 (2019) +51. Sun, P., Chen, S., Zhu, C., Xiao, F., Luo, P., Xie, S., Yan, Z.: Going denser with open-vocabulary part segmentation. 2023 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 15407-15419 (2023), https://api.sementicscholar.org/CorpusID:258762519 +52. Tang, C., Xie, L., Zhang, X., Hu, X., Tian, Q.: Visual recognition by request. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15265-15274 (2023) +53. Tong, K., Wu, Y.: Deep learning-based detection from the perspective of small or tiny objects: A survey. Image and Vision Computing 123, 104471 (2022) +54. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., Lample, G.: Llama: Open and efficient foundation language models (2023) +55. Tsogkas, S., Kokkinos, I., Papandreou, G., Vedaldi, A.: Deep learning for semantic part segmentation with high-level guidance. arXiv preprint arXiv:1505.02438 (2015) +56. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The caltech-ucsd birds-200-2011 dataset (2011) +57. Wang, J., Yuille, A.L.: Semantic part segmentation using compositional model combining shape and appearance. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1788-1797 (2015) +58. Wang, P., Shen, X., Lin, Z., Cohen, S., Price, B., Yuille, A.L.: Joint object and part segmentation using deep learned potentials. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 1573-1581 (2015) +59. Wang, W., Lv, Q., Yu, W., Hong, W., Qi, J., Wang, Y., Ji, J., Yang, Z., Zhao, L., Song, X., et al.: Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079 (2023) +60. Wang, X., Li, S., Kallidromitis, K., Kato, Y., Kozuka, K., Darrell, T.: Hierarchical open-vocabulary universal image segmentation. Advances in Neural Information Processing Systems 36 (2024) + +61. Wei, M., Yue, X., Zhang, W., Kong, S., Liu, X., Pang, J.: Ov-parts: Towards open-vocabulary part segmentation. Advances in Neural Information Processing Systems 36 (2024) +62. Wu, T.H., Biamby, G., Chan, D., Dunlap, L., Gupta, R., Wang, X., Gonzalez, J.E., Darrell, T.: See, say, and segment: Teaching lmms to overcome false premises. arXiv preprint arXiv:2312.08366 (2023) +63. Xiang, F., Qin, Y., Mo, K., Xia, Y., Zhu, H., Liu, F., Liu, M., Jiang, H., Yuan, Y., Wang, H., Yi, L., Chang, A.X., Guibas, L.J., Su, H.: SAPIEN: A simulated part-based interactive environment. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020) +64. You, H., Zhang, H., Gan, Z., Du, X., Zhang, B., Wang, Z., Cao, L., Chang, S.F., Yang, Y.: Ferret: Refer and ground anything anywhere at any granularity. arXiv preprint arXiv:2310.07704 (2023) +65. Yu, F., Liu, K., Zhang, Y., Zhu, C., Xu, K.: Partnet: A recursive part decomposition network for fine-grained and hierarchical shape segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9491-9500 (2019) +66. Yuan, Y., Li, W., Liu, J., Tang, D., Luo, X., Qin, C., Zhang, L., Zhu, J.: Osprey: Pixel understanding with visual instruction tuning (2023) +67. Zhao, J., Li, J., Cheng, Y., Sim, T., Yan, S., Feng, J.: Understanding humans in crowded scenes: Deep nested adversarial learning and a new benchmark for multi-human parsing. In: Proceedings of the 26th ACM international conference on Multimedia. pp. 792-800 (2018) +68. Zheng, S., Yang, F., Kiapour, M.H., Piramuthu, R.: Modanet: A large-scale street fashion dataset with polygon annotations. In: Proceedings of the 26th ACM international conference on Multimedia. pp. 1670-1678 (2018) +69. Zhou, B., Zhao, H., Puig, X., Fidler, S., Barriuso, A., Torralba, A.: Scene parsing through ade20k dataset. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 633-641 (2017) \ No newline at end of file diff --git a/2024/SPIN_ Hierarchical Segmentation with Subpart Granularity in Natural Images/images.zip b/2024/SPIN_ Hierarchical Segmentation with Subpart Granularity in Natural Images/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..bdcb95331fcc657837efae8d68d5f7481e99ef98 --- /dev/null +++ b/2024/SPIN_ Hierarchical Segmentation with Subpart Granularity in Natural Images/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3a52a85882089a6012d7868ee9889b4d14b3a31a8351cdb5f741f2bfc32eba70 +size 308342 diff --git a/2024/SPIN_ Hierarchical Segmentation with Subpart Granularity in Natural Images/layout.json b/2024/SPIN_ Hierarchical Segmentation with Subpart Granularity in Natural Images/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..89888e92cbc70c7aaba53ee98dabd8f7ffc6a226 --- /dev/null +++ b/2024/SPIN_ Hierarchical Segmentation with Subpart Granularity in Natural Images/layout.json @@ -0,0 +1,9189 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 142, + 111, + 473, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 111, + 473, + 148 + ], + "spans": [ + { + "bbox": [ + 142, + 111, + 473, + 148 + ], + "type": "text", + "content": "SPIN: Hierarchical Segmentation with Subpart Granularity in Natural Images" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 159, + 168, + 455, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 168, + 455, + 193 + ], + "spans": [ + { + "bbox": [ + 159, + 168, + 455, + 193 + ], + "type": "text", + "content": "Josh Myers-Dean" + }, + { + "bbox": [ + 159, + 168, + 455, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 159, + 168, + 455, + 193 + ], + "type": "text", + "content": ", Jarek Reynolds" + }, + { + "bbox": [ + 159, + 168, + 455, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 159, + 168, + 455, + 193 + ], + "type": "text", + "content": ", Brian Price" + }, + { + "bbox": [ + 159, + 168, + 455, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 159, + 168, + 455, + 193 + ], + "type": "text", + "content": ", Yifei Fan" + }, + { + "bbox": [ + 159, + 168, + 455, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 159, + 168, + 455, + 193 + ], + "type": "text", + "content": ", and Danna Gurari" + }, + { + "bbox": [ + 159, + 168, + 455, + 193 + ], + "type": "inline_equation", + "content": "^{1,3}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 151, + 201, + 465, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 201, + 465, + 213 + ], + "spans": [ + { + "bbox": [ + 151, + 201, + 465, + 213 + ], + "type": "text", + "content": "1 University of Colorado Boulder, 2 Adobe, 3 University of Texas at Austin" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 159, + 244, + 455, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 244, + 455, + 356 + ], + "spans": [ + { + "bbox": [ + 159, + 244, + 455, + 356 + ], + "type": "text", + "content": "Abstract. Hierarchical segmentation entails creating segmentations at varying levels of granularity. We introduce the first hierarchical semantic segmentation dataset with subpart annotations for natural images, which we call SPIN (SubPartImageNet). We also introduce two novel evaluation metrics to evaluate how well algorithms capture spatial and semantic relationships across hierarchical levels. We benchmark modern models across three different tasks and analyze their strengths and weaknesses across objects, parts, and subparts. To facilitate community-wide progress, we publicly release our dataset at https://joshmyersdean.github.io/spin/index.html." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 376, + 230, + 389 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 376, + 230, + 389 + ], + "spans": [ + { + "bbox": [ + 132, + 376, + 230, + 389 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 402, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 510 + ], + "type": "text", + "content": "When people discuss hierarchical image analysis tasks they are typically talking about one of two approaches: is-a relationships [27,28], which treat categories at different abstraction levels (e.g., a Subaru is a car), or is-part-of relationships, which concentrate on dividing objects into their constituent parts [13, 14, 19] (e.g., a door is part of a car). While the former has been widely explored, the latter—focusing on object decomposition—has received limited attention in computer vision research. Within segmentation research, the focus of this paper, is-part-of relationships have primarily been explored only for part-whole hierarchies, ignoring finer-grained details, such as subparts (i.e., parts of parts)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 510, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 605 + ], + "type": "text", + "content": "Research to subpart-level segmentation granularity is hindered by a scarcity of data. While synthetic 3D datasets [14, 19, 35, 63] could be used to infer hierarchical segmentations to subpart granularity, it is well-known that models developed with synthetic data typically generalize poorly to natural images (i.e., images taken by a camera) [8]. This lack of annotated natural data has meant that the few models designed to generate subpart granularity segmentations [11, 52, 60] could only be evaluated qualitatively on a small number of examples (rather than quantitatively at scale)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "content": "Our work facilitates the development of algorithms for subpart granularity in three key ways. First, we collect over 102,000 subpart segmentations across 203 diverse subpart categories to expand upon PartImageNet [17] and create the SubPartImageNet (SPIN) dataset—the first to offer subpart annotations for natural images. We release this dataset publicly to foster community progress." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 192, + 173 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 192, + 173 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 192, + 173 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 192, + 173 + ], + "type": "image", + "image_path": "430260e209de2ae28810f30abe916d3f821a96710724b757d244b2066cb0766e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 156, + 174, + 166, + 182 + ], + "lines": [ + { + "bbox": [ + 156, + 174, + 166, + 182 + ], + "spans": [ + { + "bbox": [ + 156, + 174, + 166, + 182 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 192, + 115, + 249, + 173 + ], + "blocks": [ + { + "bbox": [ + 192, + 115, + 249, + 173 + ], + "lines": [ + { + "bbox": [ + 192, + 115, + 249, + 173 + ], + "spans": [ + { + "bbox": [ + 192, + 115, + 249, + 173 + ], + "type": "image", + "image_path": "a9a305fa1738e86babc7dab2b438910eef6fd0c0fc3ac9d79eedbf22688291cd.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 215, + 174, + 225, + 182 + ], + "lines": [ + { + "bbox": [ + 215, + 174, + 225, + 182 + ], + "spans": [ + { + "bbox": [ + 215, + 174, + 225, + 182 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 250, + 115, + 306, + 173 + ], + "blocks": [ + { + "bbox": [ + 250, + 115, + 306, + 173 + ], + "lines": [ + { + "bbox": [ + 250, + 115, + 306, + 173 + ], + "spans": [ + { + "bbox": [ + 250, + 115, + 306, + 173 + ], + "type": "image", + "image_path": "d9bd6d7ea9e6f65fb3986e731fb31ec67a4de54295a7844d9b3ec78976525854.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 272, + 174, + 282, + 182 + ], + "lines": [ + { + "bbox": [ + 272, + 174, + 282, + 182 + ], + "spans": [ + { + "bbox": [ + 272, + 174, + 282, + 182 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 191, + 482, + 258 + ], + "lines": [ + { + "bbox": [ + 130, + 191, + 482, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 191, + 482, + 258 + ], + "type": "text", + "content": "Fig. 1: Overview of the diversity of SPIN. Panels (a) and (b) depict subparts unique to specific object class members, such as a roll cage in a car and a shell in a turtle. Panels (c) and (d) illustrate the variability in the number of subparts per object of the same class, with examples of 13 and 6 subparts. Panels (e) and (f) highlight the disparity in image area coverage by different subparts, such as a bottle label (large) versus quadruped claws (tiny)." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 307, + 115, + 365, + 173 + ], + "blocks": [ + { + "bbox": [ + 307, + 115, + 365, + 173 + ], + "lines": [ + { + "bbox": [ + 307, + 115, + 365, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 115, + 365, + 173 + ], + "type": "image", + "image_path": "b63a593494b211050330e7ceab3339405c8447b513a34c88bb3c137b2b6ed62a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 174, + 341, + 182 + ], + "lines": [ + { + "bbox": [ + 331, + 174, + 341, + 182 + ], + "spans": [ + { + "bbox": [ + 331, + 174, + 341, + 182 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 365, + 115, + 421, + 173 + ], + "blocks": [ + { + "bbox": [ + 365, + 115, + 421, + 173 + ], + "lines": [ + { + "bbox": [ + 365, + 115, + 421, + 173 + ], + "spans": [ + { + "bbox": [ + 365, + 115, + 421, + 173 + ], + "type": "image", + "image_path": "0bdb3c52360b4667684fcb75b04100d30678e1454ab7d7a7fb22f7502d1b00cf.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 388, + 174, + 398, + 182 + ], + "lines": [ + { + "bbox": [ + 388, + 174, + 398, + 182 + ], + "spans": [ + { + "bbox": [ + 388, + 174, + 398, + 182 + ], + "type": "text", + "content": "(e)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 423, + 115, + 479, + 173 + ], + "blocks": [ + { + "bbox": [ + 423, + 115, + 479, + 173 + ], + "lines": [ + { + "bbox": [ + 423, + 115, + 479, + 173 + ], + "spans": [ + { + "bbox": [ + 423, + 115, + 479, + 173 + ], + "type": "image", + "image_path": "624f1cc272c0a9e7caea6832adeb66eb3860a38a51e6c2b77ac9b7337db41618.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 445, + 174, + 455, + 182 + ], + "lines": [ + { + "bbox": [ + 445, + 174, + 455, + 182 + ], + "spans": [ + { + "bbox": [ + 445, + 174, + 455, + 182 + ], + "type": "text", + "content": "(f)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 282, + 482, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 282, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 282, + 482, + 390 + ], + "type": "text", + "content": "Examples of annotated subparts are shown in Fig. 1. Second, we introduce two novel metrics to address a shortcoming of the prevailing evaluation methods for hierarchical segmentation, which only assess each granularity level in isolation. Our newly proposed metrics are intended to work alongside these traditional metrics (e.g., Intersection over Union) and assess how well models capture spatial and semantic relationships across hierarchical levels. Our third key contribution is that we benchmark modern models on SPIN across objects, parts, and subparts, thereby presenting the first comprehensive quantitative evaluation of subpart performance and key areas needed for future improvement." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 391, + 482, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 391, + 482, + 547 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 482, + 547 + ], + "type": "text", + "content": "Advancing hierarchical segmentation to subpart granularity could significantly benefit both research and societal applications. One potential use is in generating more detailed image descriptions, such as for augmented reality experiences, answering visual questions, captioning images, or visual storytelling. Similarly, this work could be the foundation for individuals to (1) interactively learn how to speak about finer-grained entities for the first time (for children) or in a new language and (2) recall forgotten words, such as when facing temporary or chronic memory injuries or disabilities. More deeply nested, hierarchical segmentation representations can also enrich tactile visual discovery [25, 42], including for use in screen readers employed by visually impaired individuals (e.g., Apple's object-based exploration feature [1]). Similarly, we believe finer-grained information could facilitate improvements for related problems, including image/video retrieval, image/video editing, automatic magnification, and robotics." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 566, + 243, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 566, + 243, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 566, + 243, + 578 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "text", + "content": "Datasets with Hierarchically Segmented Objects. Several datasets provide segmentations showing how objects are hierarchically decomposed into their recursively nested parts. A few do so without semantic labels specifying the categories of segmented content, including the pioneering Berkeley dataset published in 2001 [36] and the large-scale SA-1B dataset released in 2023 [22]. Our work, in contrast, focuses on hierarchically segmenting objects with category" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "type": "text", + "content": "Myers-Dean et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "labels. Already, numerous datasets semantically decompose objects and their parts in natural images [7, 14, 15, 17, 21, 31, 45, 50, 56, 61, 67-69] and synthetic images [5, 19, 20, 35, 39]. One dataset, (i.e., ADE20K [69]) even provides subpart annotations for " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " of its objects, however its lack of exhaustive annotations for objects and their associated parts [52] impedes it benefit for evaluating prediction models. Extending prior work, we introduce the first exhaustively labelled hierarchical semantic segmentation dataset with subpart annotations for natural images by extending one of the largest part-based datasets [17]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 223, + 483, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 223, + 483, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 223, + 483, + 342 + ], + "type": "text", + "content": "Hierarchical Semantic Segmentation Evaluation. Most evaluation protocols for hierarchical segmentation algorithms assess each hierarchy level independently, such as by reporting for object and part categories separately their Intersection over Union scores [9, 23, 31, 32, 37, 40, 52, 55, 57, 58, 61, 65], Average Precision scores [45, 51], and Panoptic Quality scores [14, 29, 30, 52, 60]. The one exception is Hierarchical Panoptic Quality [52] which recursively measures the extent to which a predicted hierarchy is complete (e.g., a wheel is a part of a car). Complementing prior work, we introduce evaluation metrics for assessing how well a model captures spatial and semantic hierarchical relationships, such as if a part is perfectly contained in it's whole." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 354, + 483, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 483, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 483, + 485 + ], + "type": "text", + "content": "Hierarchical Semantic Segmentation Algorithms. Thus far, the focus of methods that predict all entities in a hierarchical decomposition largely has centered on predicting just objects and parts [16, 51, 58], aligning with existing datasets which only recently increasingly have added part segmentations to object segmentations. Only a few works, HIPIE [60], VDT [11], Semantic-SAM [26], and ViRReq [52]" + }, + { + "bbox": [ + 130, + 354, + 483, + 485 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 130, + 354, + 483, + 485 + ], + "type": "text", + "content": ", examined predicting more layers of a hierarchy to also include subparts. However, these works relied on qualitative assessments for subpart predictions due to a lack of annotated datasets supporting quantitative assessments. Complementing prior work, we quantitatively evaluate HIPIE and over 20 model variants to show their performance on our new dataset challenge across all three layers of the hierarchy (i.e., object, parts, and subparts)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 502, + 237, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 502, + 237, + 516 + ], + "spans": [ + { + "bbox": [ + 131, + 502, + 237, + 516 + ], + "type": "text", + "content": "3 SPIN Dataset" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 528, + 481, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 528, + 481, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 528, + 481, + 552 + ], + "type": "text", + "content": "We now introduce and characterize our new subpart semantic segmentation dataset that we call SPIN, which is short for SubPart ImageNet." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 568, + 248, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 568, + 248, + 580 + ], + "spans": [ + { + "bbox": [ + 131, + 568, + 248, + 580 + ], + "type": "text", + "content": "3.1 Dataset Creation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 587, + 482, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 587, + 482, + 636 + ], + "spans": [ + { + "bbox": [ + 130, + 587, + 482, + 636 + ], + "type": "text", + "content": "Data Source. Our dataset extends an existing dataset that provides both object and associated part annotations: PartImageNet [17]. It includes 24,080 natural images across 158 ImageNet categories (112 non-rigid and 38 rigid, e.g. animals and vehicles respectively). The authors of PartImageNet used the" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 424, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 424, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 424, + 91, + 447, + 100 + ], + "type": "text", + "content": "SPIN" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 642, + 482, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 482, + 666 + ], + "type": "text", + "content": "1 No code is publicly available for VDT, and ViRReq does not offer complete code. At the time of writing, Semantic-SAM has not released their semantic prediction code." + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "WordNet [38] taxonomy to establish these categories and their 11 parent supercategories. The authors then identified 40 part categories pertinent to the supercategories and annotated these in all images. Every image in PartImageNet contains one semantically segmented object, and we used only those with at least one segmented part. Due to our limited annotation budget, we capped each supercategory at 1,200 images. This resulted in the following amounts per supercategory: 311 - Aeroplane, 483 - Bottle, 559 - Boat, 634 - Bicycle, and 1,200 each for Biped, Bird, Car, Fish, Quadruped, Reptile, and Snake." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 227, + 482, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 227, + 482, + 334 + ], + "spans": [ + { + "bbox": [ + 130, + 227, + 482, + 334 + ], + "type": "text", + "content": "Subpart Category Selection. We identified 206 subpart categories to segment for 34 PartImageNet part categories that we identified as being decomposable into subparts. Of these, 168 subpart categories generalize across many part categories and 38 categories pertain to only a few. For example, in our dataset, mouths generalize while shells don't, with most reptiles having heads with mouths but only one type of reptile having a body with a shell (i.e., turtles). We show examples of specific subparts in Fig. 1(a-b). A histogram showing how many subparts are assigned to each of the 34 part categories is shown in Fig. 2, and a full list is provided in the supplementary materials." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 335, + 482, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 482, + 431 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 482, + 431 + ], + "type": "text", + "content": "We created our subpart taxonomy through a multi-step process. We identified the general subpart categories by prompting GPT-4 [2] to list expected subparts for every object-part category in PartImageNet, and then three authors edited the list to exclude non-visible subparts (e.g., a skull). We identified specific subpart categories through manual inspection of at least 15 images per object category, excluding any categories that had ambiguous subpart decompositions; for example, it's not obvious to a lay person what boundaries to use when decomposing a tail into a tip, shaft, and base." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 446, + 482, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 446, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 446, + 482, + 483 + ], + "type": "text", + "content": "Subpart Annotation Task Design. We created a task interface for collecting all subpart annotations. It presents each image-object pair alongside one of the object's nested parts, outlined with an overlaid polygon. For each part, the task" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 132, + 540, + 481, + 632 + ], + "blocks": [ + { + "bbox": [ + 132, + 540, + 481, + 632 + ], + "lines": [ + { + "bbox": [ + 132, + 540, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 540, + 481, + 632 + ], + "type": "image", + "image_path": "929d7239c260f06f0a3fd6cef87c8d6ccb03ae0f99ab0a0413ed4c1f6452ead6.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 639, + 481, + 663 + ], + "lines": [ + { + "bbox": [ + 130, + 639, + 481, + 663 + ], + "spans": [ + { + "bbox": [ + 130, + 639, + 481, + 663 + ], + "type": "text", + "content": "Fig. 2: Histogram the number of unique subpart category labels for each of the 34 part categories. (Aero=Aeroplane; Quad=Quadruped)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "type": "text", + "content": "Myers-Dean et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 491, + 481, + 515 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 491, + 481, + 515 + ], + "spans": [ + { + "bbox": [ + 133, + 491, + 481, + 515 + ], + "type": "text", + "content": "2 We prompted GPT-4 with \"Please list the canonical subparts of a " + }, + { + "bbox": [ + 133, + 491, + 481, + 515 + ], + "type": "inline_equation", + "content": "<" + }, + { + "bbox": [ + 133, + 491, + 481, + 515 + ], + "type": "text", + "content": " object " + }, + { + "bbox": [ + 133, + 491, + 481, + 515 + ], + "type": "inline_equation", + "content": "\\succ -\\text{part}>" + }, + { + "bbox": [ + 133, + 491, + 481, + 515 + ], + "type": "text", + "content": ". Only include subparts that are clearly visible and recognizable to a layperson.\"" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 142, + 114, + 472, + 148 + ], + "blocks": [ + { + "bbox": [ + 142, + 114, + 472, + 148 + ], + "lines": [ + { + "bbox": [ + 142, + 114, + 472, + 148 + ], + "spans": [ + { + "bbox": [ + 142, + 114, + 472, + 148 + ], + "type": "table", + "html": "
#Images#Object Cat#Part Cat#Sub-Part Cat
10,38711 (10,387)40 (29,818)203 (106,324)
", + "image_path": "ef24e8a325b06588baa3f3e3f2f319c4836e6feb84a65f80ef1f2068d85f43ea.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 148, + 481, + 170 + ], + "lines": [ + { + "bbox": [ + 132, + 148, + 481, + 170 + ], + "spans": [ + { + "bbox": [ + 132, + 148, + 481, + 170 + ], + "type": "text", + "content": "Table 1: SPIN Composition: Each column lists category counts alongside the total number of annotated entities within those categories. (Cat=Categories)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 205, + 482, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 482, + 313 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 482, + 313 + ], + "type": "text", + "content": "interface has multiple steps. First, users are asked, \"Can you locate the -- enclosed in a polygon?\" If the target content is present, annotators are then asked for each possible subpart, \"Can you locate any subpart> on the --?\" If the response is yes, annotators segment the subpart. The interface supports segmentation by collecting a series of clicked points to create a connected polygon. It also supports annotating multiple polygons in three scenarios: (1) multiple instances of a subpart (e.g., two eyes), (2) subparts with holes (e.g., coiled snakes), and (3) occlusions breaking a subpart into multiple, disconnected pieces." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 327, + 482, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 482, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 482, + 399 + ], + "type": "text", + "content": "Annotation Collection. We hired 18 highly trusted annotators from Amazon Mechanical Turk (AMT) who we had vetted through repeated employment for previous segmentation tasks. During data collection, we ensured high annotation quality through five methods: an onboarding qualification test, detailed instructions, live \"office hours\" during annotation periods, phased task roll-out for worker feedback, and continuous inspection of submitted results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 413, + 481, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 481, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 481, + 450 + ], + "type": "text", + "content": "Dataset Splits. Our dataset adheres to the PartImageNet training, validation, and testing splits of " + }, + { + "bbox": [ + 130, + 413, + 481, + 450 + ], + "type": "inline_equation", + "content": "85\\%" + }, + { + "bbox": [ + 130, + 413, + 481, + 450 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 413, + 481, + 450 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 130, + 413, + 481, + 450 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 413, + 481, + 450 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 413, + 481, + 450 + ], + "type": "text", + "content": " of the data, respectively. This results in 8,828 training, 519 validation, and 1,040 test images in our SPIN dataset." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 470, + 246, + 482 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 470, + 246, + 482 + ], + "spans": [ + { + "bbox": [ + 131, + 470, + 246, + 482 + ], + "type": "text", + "content": "3.2 Dataset Analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 492, + 447, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 492, + 447, + 506 + ], + "spans": [ + { + "bbox": [ + 131, + 492, + 447, + 506 + ], + "type": "text", + "content": "We now characterize SPIN's overall composition as well as its subparts." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 519, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 519, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 519, + 482, + 628 + ], + "type": "text", + "content": "Overall Dataset Composition. We characterize SPIN with respect to the following seven factors: number of images, number of object categories, total number of annotated objects, number of part categories, total number of annotated parts, number of subpart categories, and total number of annotated subparts. Results are shown in Table 1. As shown, our contribution of 203 entity categories is over a five-order-of-magnitude increase compared to the 40 categories added for PartImageNet [17], while providing nearly an order of magnitude more annotations (i.e. 106,324 semantic annotations versus 11,960 part segmentations in PartImageNet)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": "Subpart Statistics. We next compute the mean number of subparts per part and characterize the typical appearance of subparts with respect to five metrics:" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 424, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 424, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 424, + 91, + 447, + 100 + ], + "type": "text", + "content": "SPIN" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 115, + 304, + 149 + ], + "blocks": [ + { + "bbox": [ + 132, + 115, + 304, + 149 + ], + "lines": [ + { + "bbox": [ + 132, + 115, + 304, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 304, + 149 + ], + "type": "image", + "image_path": "e27bf80c8da0872c26717dfabd74cd446b5750be21a6decec092c84e034bd0f2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 184, + 149, + 255, + 157 + ], + "lines": [ + { + "bbox": [ + 184, + 149, + 255, + 157 + ], + "spans": [ + { + "bbox": [ + 184, + 149, + 255, + 157 + ], + "type": "text", + "content": "(a)InstancesPerPart" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 306, + 115, + 480, + 148 + ], + "blocks": [ + { + "bbox": [ + 306, + 115, + 480, + 148 + ], + "lines": [ + { + "bbox": [ + 306, + 115, + 480, + 148 + ], + "spans": [ + { + "bbox": [ + 306, + 115, + 480, + 148 + ], + "type": "image", + "image_path": "f3631d77a8daa27e9b50bcdbe6533ddaeff8f554f0d8d498f9056f68815eb1c0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 363, + 149, + 427, + 157 + ], + "lines": [ + { + "bbox": [ + 363, + 149, + 427, + 157 + ], + "spans": [ + { + "bbox": [ + 363, + 149, + 427, + 157 + ], + "type": "text", + "content": "(d) Image Coverage" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 132, + 158, + 304, + 191 + ], + "blocks": [ + { + "bbox": [ + 132, + 158, + 304, + 191 + ], + "lines": [ + { + "bbox": [ + 132, + 158, + 304, + 191 + ], + "spans": [ + { + "bbox": [ + 132, + 158, + 304, + 191 + ], + "type": "image", + "image_path": "02c1d19954c0b8a8be2278aa0bfb6aeee8435c36109c719e6cf6ad2b26aa2a24.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 203, + 191, + 237, + 198 + ], + "lines": [ + { + "bbox": [ + 203, + 191, + 237, + 198 + ], + "spans": [ + { + "bbox": [ + 203, + 191, + 237, + 198 + ], + "type": "text", + "content": "(b) Extent" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 306, + 158, + 479, + 190 + ], + "blocks": [ + { + "bbox": [ + 306, + 158, + 479, + 190 + ], + "lines": [ + { + "bbox": [ + 306, + 158, + 479, + 190 + ], + "spans": [ + { + "bbox": [ + 306, + 158, + 479, + 190 + ], + "type": "image", + "image_path": "e4a9c85556fe4b6ab9637225151d7bd2a5aff75e6a00e8d29ce8011715a9cd65.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 364, + 191, + 429, + 198 + ], + "lines": [ + { + "bbox": [ + 364, + 191, + 429, + 198 + ], + "spans": [ + { + "bbox": [ + 364, + 191, + 429, + 198 + ], + "type": "text", + "content": "(e) Object Coverage" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 132, + 199, + 304, + 247 + ], + "blocks": [ + { + "bbox": [ + 132, + 199, + 304, + 247 + ], + "lines": [ + { + "bbox": [ + 132, + 199, + 304, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 199, + 304, + 247 + ], + "type": "image", + "image_path": "ba7dced9c2690df2deb057d8007b7976724e45c71f87630bde54df26fc2e5d07.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 178, + 248, + 265, + 257 + ], + "lines": [ + { + "bbox": [ + 178, + 248, + 265, + 257 + ], + "spans": [ + { + "bbox": [ + 178, + 248, + 265, + 257 + ], + "type": "text", + "content": "(c) Boundary Complexity" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 306, + 199, + 479, + 251 + ], + "blocks": [ + { + "bbox": [ + 306, + 199, + 479, + 251 + ], + "lines": [ + { + "bbox": [ + 306, + 199, + 479, + 251 + ], + "spans": [ + { + "bbox": [ + 306, + 199, + 479, + 251 + ], + "type": "image", + "image_path": "5949391db2cbc1db7b5857255d943d57df4500982fb2a298c541327733e7f32e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 370, + 251, + 431, + 258 + ], + "lines": [ + { + "bbox": [ + 370, + 251, + 431, + 258 + ], + "spans": [ + { + "bbox": [ + 370, + 251, + 431, + 258 + ], + "type": "text", + "content": "(f) Part Coverage" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 268, + 482, + 335 + ], + "lines": [ + { + "bbox": [ + 130, + 268, + 482, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 268, + 482, + 335 + ], + "type": "text", + "content": "Fig. 3: Boxplots showing the distribution of subpart image occupation (left) and boundary complexities per part, per object (right). The blue lines represent medians, bottoms and tops of each box represent the 25th and 75th percentile values respectively, and whiskers represent the most extreme data points not considered outliers. Overall, SPIN's subparts take up a relatively small number of pixels per image, while featuring a range of geometric complexity. (Aero=Aeroplane; Quad=Quadruped)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 360, + 480, + 485 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 360, + 480, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 360, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 138, + 360, + 480, + 396 + ], + "type": "text", + "content": "- Boundary complexity: ratio of a subpart's area to the length of its perimeter (i.e., isoperimetric quotient). Values range from 0 (highly jagged boundary) to 1 (circular)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 397, + 480, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 397, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 138, + 397, + 480, + 445 + ], + "type": "text", + "content": "- Extent: the ratio of the area of a contour to the area of its bounding box. Values are in " + }, + { + "bbox": [ + 138, + 397, + 480, + 445 + ], + "type": "inline_equation", + "content": "(0,1]" + }, + { + "bbox": [ + 138, + 397, + 480, + 445 + ], + "type": "text", + "content": ", where values approaching 0 mean that a contour occupies little area in its bounding box (e.g., a thin diagonal line) and 1 means that a contour is perfectly contained (e.g., a square)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 446, + 447, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 446, + 447, + 459 + ], + "spans": [ + { + "bbox": [ + 138, + 446, + 447, + 459 + ], + "type": "text", + "content": "- Image coverage: fraction of image pixels occupied by the subpart." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 459, + 479, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 459, + 479, + 472 + ], + "spans": [ + { + "bbox": [ + 138, + 459, + 479, + 472 + ], + "type": "text", + "content": "- Object coverage: fraction of parent object pixels occupied by the subpart." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 472, + 463, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 472, + 463, + 485 + ], + "spans": [ + { + "bbox": [ + 138, + 472, + 463, + 485 + ], + "type": "text", + "content": "- Part coverage: fraction of parent part pixels occupied by the subpart." + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 496, + 479, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 496, + 479, + 520 + ], + "spans": [ + { + "bbox": [ + 130, + 496, + 479, + 520 + ], + "type": "text", + "content": "For all six factors, we report both the overall mean as well as the mean with respect to each object category. Results are shown in Fig. 3." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "Objects tend to have 2 to 5 subparts per part (i.e. Fig. 3a; 25th to 75th percentile values), with a total of 4 to 15 subparts. There is diversity across object categories with aeroplanes typically having small variability with only 1 to 2 subparts per part (i.e. 25th to 75th percentile range) while car typically exhibits great variability with 2 to 8 subparts per part (i.e. 25th to 75th percentile range), as exemplified in Fig. 1(c-d). We attribute the large range of subpart per part counts to two factors: the intrinsic diversity in cars and the viewing angle. For example, a go-cart viewed from the front will have no windshield, no windows, and two tires, while a bus viewed from a front-right angle will have a windshield, tires, and windows. Such object-specific patterns around the expected prevalence of subparts could be predictive cues for models decomposing detected objects into their recursively nested subparts." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "type": "text", + "content": "Myers-Dean et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "In terms of shape, most subparts' contours occupy the majority of their bounding box (i.e., Fig. 3b, median extent value " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\sim 0.7" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": ") with moderate boundary complexity (i.e., Fig. 3c, median boundary complexity values around 0.5). Still, we observe high variability for both metrics that we again attribute to the intrinsic diversity in each object category as well as the viewing angle. For example, a bottle cap or a bicycle tire viewed from the side will be an ellipse versus a nearly perfect circle when viewed from the top (i.e., occupying more of its bounding box). Altogether, these findings highlight that our dataset encourages the design of models that can account for a large range of subpart shapes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 226, + 482, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 226, + 482, + 406 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 482, + 406 + ], + "type": "text", + "content": "Regarding the relative area occupied by each subpart, we find that all subparts occupy small portions of the parent parts, parent objects, and entire image. For example, across every object category, subparts occupy less than " + }, + { + "bbox": [ + 130, + 226, + 482, + 406 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 130, + 226, + 482, + 406 + ], + "type": "text", + "content": " of the image for at least " + }, + { + "bbox": [ + 130, + 226, + 482, + 406 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 130, + 226, + 482, + 406 + ], + "type": "text", + "content": " of subpart instances (Fig. 3d). Additionally, the subparts tend to occupy less than " + }, + { + "bbox": [ + 130, + 226, + 482, + 406 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 226, + 482, + 406 + ], + "type": "text", + "content": " of their parent parts (Fig. 3e; Overall) and less than " + }, + { + "bbox": [ + 130, + 226, + 482, + 406 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 130, + 226, + 482, + 406 + ], + "type": "text", + "content": " of their parent objects (Fig. 3f; Overall). When adopting size thresholds introduced for the MSCOCO dataset [33], where " + }, + { + "bbox": [ + 130, + 226, + 482, + 406 + ], + "type": "inline_equation", + "content": "32^2" + }, + { + "bbox": [ + 130, + 226, + 482, + 406 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 226, + 482, + 406 + ], + "type": "inline_equation", + "content": "96^2" + }, + { + "bbox": [ + 130, + 226, + 482, + 406 + ], + "type": "text", + "content": " are thresholds determining whether an object is small, medium, or large, we find " + }, + { + "bbox": [ + 130, + 226, + 482, + 406 + ], + "type": "inline_equation", + "content": "54.10\\%" + }, + { + "bbox": [ + 130, + 226, + 482, + 406 + ], + "type": "text", + "content": " (57,525) of SPIN's subpart annotations qualify as small, " + }, + { + "bbox": [ + 130, + 226, + 482, + 406 + ], + "type": "inline_equation", + "content": "38.08\\%" + }, + { + "bbox": [ + 130, + 226, + 482, + 406 + ], + "type": "text", + "content": " (40,488) as medium, and only " + }, + { + "bbox": [ + 130, + 226, + 482, + 406 + ], + "type": "inline_equation", + "content": "7.82\\%" + }, + { + "bbox": [ + 130, + 226, + 482, + 406 + ], + "type": "text", + "content": " (8,311) as large. Most subparts are already considered small according to mainstream research despite that we annotated subparts for a dataset known to typically contain a single, large prominent object in the image. Consequently, embedded in our problem is the well-known challenge of small entity detection [53], which will only grow as subparts are explored for more complex scenes where less prominent objects must also be hierarchically decomposed." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 407, + 482, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 482, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 482, + 491 + ], + "type": "text", + "content": "When comparing subparts belonging to the rigid object categories (i.e., human-made) versus non-rigid object categories (i.e., animals), we found no clear distinction. For instance, the mean number of subparts per part (Fig. 3a) is the smallest amount with least variability in aeroplanes and the greatest amount with greatest variability in cars. While prior work has emphasized distinguishing between these two extremes [17], our analysis suggests that rigid and non-rigid objects appear quite similar at the subpart level." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 520, + 369, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 520, + 369, + 533 + ], + "spans": [ + { + "bbox": [ + 131, + 520, + 369, + 533 + ], + "type": "text", + "content": "4 Evaluating Hierarchical Consistency" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 555, + 481, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 555, + 481, + 591 + ], + "spans": [ + { + "bbox": [ + 130, + 555, + 481, + 591 + ], + "type": "text", + "content": "We facilitate assessing how well relationships across varying levels of granularity are captured for hierarchical segmentation by introducing two metrics that account for spatial and semantic relationships respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "We first introduce the Spatial Consistency Score (SpCS), which focuses on structural hierarchical alignment rather than semantic alignment. It indicates the proportion of a child's segmentation (e.g., subparts) that is contained in its super-region's segmentation (e.g., part, object), with values ranging from 0 (no containment) to 1 (perfect containment). It leverages binarized segmentations distinguishing the foreground entity from the background. Formally, it is com" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 424, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 424, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 424, + 91, + 447, + 100 + ], + "type": "text", + "content": "SPIN" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 422, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 422, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 422, + 129 + ], + "type": "text", + "content": "puted as the average containment ratio across all pairs as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 213, + 135, + 481, + 175 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 213, + 135, + 481, + 175 + ], + "spans": [ + { + "bbox": [ + 213, + 135, + 481, + 175 + ], + "type": "interline_equation", + "content": "S p C S = \\frac {1}{| \\mathcal {R} |} \\sum_ {\\text {(c h i l d , p a r e n t)} \\in \\mathcal {R}} \\underbrace {\\frac {| \\text {c h i l d} \\cap \\text {p a r e n t} |}{| \\text {c h i l d} |}} _ {\\text {c o n t a i n m e n t r a t i o}} \\tag {1}", + "image_path": "8f3e4afb85da973f28096d216fd3ac53001ad66359496686f2883b48e384099c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 180, + 482, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 180, + 482, + 240 + ], + "spans": [ + { + "bbox": [ + 130, + 180, + 482, + 240 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 180, + 482, + 240 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 130, + 180, + 482, + 240 + ], + "type": "text", + "content": " denotes the set of prediction pairs with a ground truth parent-child relationship within a hierarchical structure. " + }, + { + "bbox": [ + 130, + 180, + 482, + 240 + ], + "type": "inline_equation", + "content": "^3 |\\mathcal{R}|" + }, + { + "bbox": [ + 130, + 180, + 482, + 240 + ], + "type": "text", + "content": " denotes the total number of parent-child pairs, " + }, + { + "bbox": [ + 130, + 180, + 482, + 240 + ], + "type": "inline_equation", + "content": "|\\mathrm{child}\\cap \\mathrm{parent}|" + }, + { + "bbox": [ + 130, + 180, + 482, + 240 + ], + "type": "text", + "content": " is the number of pixels in the intersection of the child segmentation with the parent, " + }, + { + "bbox": [ + 130, + 180, + 482, + 240 + ], + "type": "inline_equation", + "content": "|\\mathrm{child}|" + }, + { + "bbox": [ + 130, + 180, + 482, + 240 + ], + "type": "text", + "content": " is the number of pixels in the child segmentation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 240, + 482, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 240, + 482, + 410 + ], + "spans": [ + { + "bbox": [ + 130, + 240, + 482, + 410 + ], + "type": "text", + "content": "We next introduce the Semantic Consistency Score (SeCS), which focuses on semantic alignment by measuring how well predictions at different levels (subpart, part, object) logically correspond. For example, an eye (subpart) should correspond to a head (part) and, by extension, to a quadruped (object), rather than illogical associations like a windshield (subpart) with a head (part) or a bottle (object). Let " + }, + { + "bbox": [ + 130, + 240, + 482, + 410 + ], + "type": "inline_equation", + "content": "FG_{S}" + }, + { + "bbox": [ + 130, + 240, + 482, + 410 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 240, + 482, + 410 + ], + "type": "inline_equation", + "content": "FG_{P}" + }, + { + "bbox": [ + 130, + 240, + 482, + 410 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 240, + 482, + 410 + ], + "type": "inline_equation", + "content": "FG_{O}" + }, + { + "bbox": [ + 130, + 240, + 482, + 410 + ], + "type": "text", + "content": " be the sets of non-background predictions in subpart, part, and object predictions, respectively. Define " + }, + { + "bbox": [ + 130, + 240, + 482, + 410 + ], + "type": "inline_equation", + "content": "X = FG_{S} \\cap FG_{P} \\cap FG_{O}" + }, + { + "bbox": [ + 130, + 240, + 482, + 410 + ], + "type": "text", + "content": " as the intersection of these predictions, representing pixels with consistent foreground labels across levels. For each pixel " + }, + { + "bbox": [ + 130, + 240, + 482, + 410 + ], + "type": "inline_equation", + "content": "x \\in X" + }, + { + "bbox": [ + 130, + 240, + 482, + 410 + ], + "type": "text", + "content": ", with predicted categories " + }, + { + "bbox": [ + 130, + 240, + 482, + 410 + ], + "type": "inline_equation", + "content": "C(S_x)" + }, + { + "bbox": [ + 130, + 240, + 482, + 410 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 240, + 482, + 410 + ], + "type": "inline_equation", + "content": "C(P_x)" + }, + { + "bbox": [ + 130, + 240, + 482, + 410 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 240, + 482, + 410 + ], + "type": "inline_equation", + "content": "C(O_x)" + }, + { + "bbox": [ + 130, + 240, + 482, + 410 + ], + "type": "text", + "content": " for subparts " + }, + { + "bbox": [ + 130, + 240, + 482, + 410 + ], + "type": "inline_equation", + "content": "(S_x)" + }, + { + "bbox": [ + 130, + 240, + 482, + 410 + ], + "type": "text", + "content": ", parts " + }, + { + "bbox": [ + 130, + 240, + 482, + 410 + ], + "type": "inline_equation", + "content": "(P_x)" + }, + { + "bbox": [ + 130, + 240, + 482, + 410 + ], + "type": "text", + "content": ", and objects " + }, + { + "bbox": [ + 130, + 240, + 482, + 410 + ], + "type": "inline_equation", + "content": "(O_x)" + }, + { + "bbox": [ + 130, + 240, + 482, + 410 + ], + "type": "text", + "content": ", respectively, category entailment is evaluated using " + }, + { + "bbox": [ + 130, + 240, + 482, + 410 + ], + "type": "inline_equation", + "content": "\\mathbb{1}[C(S_x), C(P_x)]" + }, + { + "bbox": [ + 130, + 240, + 482, + 410 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 240, + 482, + 410 + ], + "type": "inline_equation", + "content": "\\mathbb{1}[C(P_x), C(O_x)]" + }, + { + "bbox": [ + 130, + 240, + 482, + 410 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 240, + 482, + 410 + ], + "type": "inline_equation", + "content": "\\mathbb{1}" + }, + { + "bbox": [ + 130, + 240, + 482, + 410 + ], + "type": "text", + "content": " is an indicator function returning 1 if the hierarchical entailment is correct, using ground truth relations, otherwise 0. We compute the entailment measure " + }, + { + "bbox": [ + 130, + 240, + 482, + 410 + ], + "type": "inline_equation", + "content": "M(x)" + }, + { + "bbox": [ + 130, + 240, + 482, + 410 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 194, + 416, + 481, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 416, + 481, + 449 + ], + "spans": [ + { + "bbox": [ + 194, + 416, + 481, + 449 + ], + "type": "interline_equation", + "content": "M (x) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} \\mathbb {1} [ C (S _ {x}), C (P _ {x}) ] \\wedge \\mathbb {1} [ C (P _ {x}), C (O _ {x}) ], \\\\ 0 & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {2}", + "image_path": "6d7959cc04ebd4a40f062508e881ef0271f35d1dbf5c96e45f1ad7ab9ae32b4b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 453, + 482, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 453, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 131, + 453, + 482, + 479 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 453, + 482, + 479 + ], + "type": "inline_equation", + "content": "\\wedge" + }, + { + "bbox": [ + 131, + 453, + 482, + 479 + ], + "type": "text", + "content": " is the logical and operator. SeCS is the mean of " + }, + { + "bbox": [ + 131, + 453, + 482, + 479 + ], + "type": "inline_equation", + "content": "M(x)" + }, + { + "bbox": [ + 131, + 453, + 482, + 479 + ], + "type": "text", + "content": " across all " + }, + { + "bbox": [ + 131, + 453, + 482, + 479 + ], + "type": "inline_equation", + "content": "x\\in X" + }, + { + "bbox": [ + 131, + 453, + 482, + 479 + ], + "type": "text", + "content": ". Values range from 0 (no semantic coherence) to 1 (perfect semantic coherence)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 495, + 307, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 495, + 307, + 510 + ], + "spans": [ + { + "bbox": [ + 132, + 495, + 307, + 510 + ], + "type": "text", + "content": "5 Algorithm Benchmarking" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 517, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 517, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 131, + 517, + 482, + 544 + ], + "type": "text", + "content": "We now assess modern models' ability to account for hierarchical decompositions to the subpart granularity." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 558, + 430, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 558, + 430, + 572 + ], + "spans": [ + { + "bbox": [ + 131, + 558, + 430, + 572 + ], + "type": "text", + "content": "5.1 Open-Vocabulary Localization in a Zero-Shot Setting" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 577, + 482, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 577, + 482, + 638 + ], + "spans": [ + { + "bbox": [ + 130, + 577, + 482, + 638 + ], + "type": "text", + "content": "We first evaluate models for open-vocabulary localization in zero-shot mode across three hierarchy levels: object, part, and subpart. This approach avoids the drawbacks of the alternative options of training models from scratch or finetuning, including the computational expense and risks of overfitting to the target task and so diminishing generalizable knowledge." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "type": "text", + "content": "Myers-Dean et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 642, + 482, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 482, + 666 + ], + "type": "text", + "content": "3 For a quadruped, for instance, pairs such as " + }, + { + "bbox": [ + 133, + 642, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\{(\\text{eyes}, \\text{head}), (\\text{chest}, \\text{torso}), (\\text{torso}, \\text{quadruped})\\}" + }, + { + "bbox": [ + 133, + 642, + 482, + 666 + ], + "type": "text", + "content": " could be present in " + }, + { + "bbox": [ + 133, + 642, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 133, + 642, + 482, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 328, + 127 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 328, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 328, + 127 + ], + "type": "text", + "content": "Models. We benchmark 11 model variants." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 128, + 482, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 128, + 482, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 128, + 482, + 187 + ], + "type": "text", + "content": "Most related to our work is HIPIE [60], which is designed for open-vocabulary hierarchical segmentation to the subpart level. It takes as input a list of all candidate categories that can appear and then predicts which categories are present where in the given image. We test two variants with different backbones: ResNet-50 [18] and ViT-H [12]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 189, + 482, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 189, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 189, + 482, + 344 + ], + "type": "text", + "content": "We also evaluate 10 variants of foundation models that, by design, are intended for use on a variety of downstream tasks. Five produce pixel-wise segmentation masks: LISA-7B [24], LISA-13B [24], GLaMM [46], and PixelLLM-7B [48], and PixelLLM-13B [48]. The other five can only produce bounding boxes: CoGVLM [59], Ferret-7B [64], Ferret-13B [64], Shikra [6], and Kosmos2 [43]. For these models, we analyze their upper bound of their potential performance by notifying the model what object category needs to be located in a given image. We examine two types of requests: \"Can you please locate the in the image?\" for objects, and \"Can you please locate the subpart> of the in the image?\"4 for each part and subpart. We evaluate with these prompts a Specific category by using the original WordNet category from the 158 ImageNet categories as well as a General category by referring to the 11 super-categories. Consequently, we test each model with four prompts." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 357, + 482, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 357, + 482, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 357, + 482, + 441 + ], + "type": "text", + "content": "Evaluation Metrics. We use two evaluation metrics for all models: (1) Intersection over Union (IoU) to assess segmentation performance for objects, parts, and subparts independently and (2) our Spatial Consistency Score (SpCS) to assess structural coherence across object, part, and subpart levels. We also use our new Semantic Consistency Score (SeCS) to assess semantic coherence between predicted hierarchical levels for the only relevant model, HIPIE (it predicts from a list of candidate categories rather than for the target, single category)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 454, + 413, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 454, + 413, + 466 + ], + "spans": [ + { + "bbox": [ + 131, + 454, + 413, + 466 + ], + "type": "text", + "content": "Localization Results (mIoU). Results are shown in Table 2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 467, + 482, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 467, + 482, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 467, + 482, + 609 + ], + "type": "text", + "content": "We observe a consistent trend across all models: localization performance is best for object segmentation, followed by increasingly worse performance for more granular categories of parts and then subparts. The best score for localizing objects is 86 followed by 40 for parts and 14 for subparts. The worst scores come from the only model directly designed for hierarchical segmentation, HIPIE. It does worse across all granularity levels than the foundation models; e.g., for subpart localization, 1 vs 3 from the worst-performing foundation model, Kosmos2. We attribute the poor performance to HIPIE's limited training data (e.g., limited part vocabulary), which only included part-level categories from Pascal Parts [7]. Moreover, HIPIE's heuristic grouping assumes that parts are the sum of their subparts, which is not the case in SPIN (e.g., unlabelled \"space\" exists between a cheek an eye in a head)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 611, + 481, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 611, + 481, + 635 + ], + "spans": [ + { + "bbox": [ + 131, + 611, + 481, + 635 + ], + "type": "text", + "content": "We explored to what extent entity size is correlated to the resulting IoU scores by using linear regression for the better-performing foundation models." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 424, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 424, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 424, + 91, + 447, + 100 + ], + "type": "text", + "content": "SPIN" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 642, + 481, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 481, + 666 + ], + "type": "text", + "content": "4 For subparts, we refer to the parent object rather than the parent part, as broader context aids in processing finer details [47]." + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 114, + 481, + 232 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 232 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 232 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 232 + ], + "type": "table", + "html": "
Method# Params.mIoUsmIoUPmIoUOSpCS - AvgSpCS - S2PSpCS - P2O
SpecificGeneralSpecificGeneralSpecificGeneralSpecificGeneralSpecificGeneralSpecificGeneral
HIPIE [60] (R-50)200M0.800.108.057.2251.3613.8797.8298.7710010095.6497.54
HIPIE [60] (ViT-H)800M0.900.927.218.2366.7721.6998.0499.2510098.5096.0896.06
PixelLLM [48]7B9.089.2432.3731.3783.9379.8773.7982.4669.5779.2089.4990.73
PixelLLM [48]13B9.5310.1332.0432.9682.9079.9684.8683.0682.1380.8792.6088.90
LISA [24]7B9.619.1427.7227.2886.2383.3679.3077.4474.6172.9392.1089.78
LISA [24]13B11.5211.5532.2931.3687.7885.4586.2879.9882.8775.7096.0292.41
GLaMM [46]7B11.0311.0039.4140.0086.2986.3187.1287.7584.2184.9695.7295.38
GLaMM - FT7B24.2524.5659.3760.7686.4291.0887.1387.9575.9385.1690.2396.04
Ferret [64]7B7.227.3025.6926.5447.8447.9878.7073.0246.5065.6893.6594.60
Ferret [64]13B6.376.2523.6723.6547.9947.5081.0477.3048.7070.9894.1795.40
CoGVLM [59]17B13.9414.2938.1338.5646.4743.6576.3377.2072.3773.3286.1987.59
Shikra [6]7B8.418.7226.5027.4245.5029.1468.5767.0163.8266.7681.8567.89
Kosmos2 [43]2.6B3.453.5919.0019.4848.6348.8982.7481.6078.3477.0396.0195.58
SAM [22]630M49.6149.6169.2369.2390.0690.0686.4286.4283.8583.8592.3092.30
SAM + CoGVLM630M19.8819.9850.9350.9077.9475.0966.7167.3380.5080.7360.6561.44
", + "image_path": "30b76fd5695f262381be2edcff54132a4ae1453f033fc563b31eeb36a653ad73.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 130, + 232, + 482, + 319 + ], + "lines": [ + { + "bbox": [ + 130, + 232, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 232, + 482, + 319 + ], + "type": "text", + "content": "Table 2: Performance of modern models with respect to localization at three granularity levels—subparts (mIoUs), parts (mIoUP), objects (mIoUO)—and hierarchical consistency at two levels—subparts to parts (SpCS-S2P), parts to objects (SpCS-P2O), and their mean (SpCS - Avg). Results are shown for hierarchical semantic segmentation models (described in Section 5.1), hierarchical object detection models (Section 5.1), and interactive segmentation models that perform category-agnostic localization (Section 5.2). The method families are separated by dashed lines in that order. Best performing scores are shown in bold." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 361, + 482, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 482, + 445 + ], + "type": "text", + "content": "Due to space constraints, we provide further experimental details and results in the supplementary materials. In summary though, our findings suggest that size has little correlation to performance for objects but greater correlations for more granular levels of parts and subparts. We suspect that one contributing factor is the sensitivity of the IoU metric for entities occupying a small number of pixels, as small errors are especially prone to yield dramatic changes in the IoU scores for such cases." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "content": "We next explored the impact of a prompt's category specificity on model performance by comparing performance for when general versus specific terms are employed. Most models perform worse when trying to locate objects categorized under general terms (e.g., 'quadruped') rather than specific ones (e.g., 'dog'). For example, CoGVLM performs 2.82 percentage points (pp) worse with general categories compared to specific categories. Similarly, the 13B variant of LISA has a 2.33pp performance drop and its 7B counterpart has a 2.87pp performance drop. The exceptions to this are GLaMM, Ferret, and Kosmos2, whose performance are relatively stable and unaffected by the level of category specificity. A potential reason is that these models utilize training data that explicitly use both general and specific terms. For example, GCG [46], a source of training data for GLaMM, refers to entities with both abstract and specific terms (e.g., person and toddler). Overall, this variance highlights the challenge in choosing an appropriate level of category specificity for objects. However, the story changes when trying to locate parts and subparts, with the performance gap between general and specific categories falling to within 1pp. An interesting direction for future work is better understanding how and why an entity's granularity level influences models' abilities to overcome abstraction challenges." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 239, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 239, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 239, + 102 + ], + "type": "text", + "content": "Myers-Dean et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 272 + ], + "type": "text", + "content": "We also examine how the training data of the Large Language Models (LLM) Llama [54], since it serves as the basis for the majority of foundation models evaluated in our study. Specifically, we analyze the empirical frequency (i.e., total occurrence) of uni-grams (e.g., eye) and n-grams (e.g., eye of the quadruped) within the dataset using the " + }, + { + "bbox": [ + 130, + 115, + 482, + 272 + ], + "type": "inline_equation", + "content": "\\infty" + }, + { + "bbox": [ + 130, + 115, + 482, + 272 + ], + "type": "text", + "content": "-gram API [34], which uses byte array pattern matching to find all occurrences of a given BPE tokenized [49] " + }, + { + "bbox": [ + 130, + 115, + 482, + 272 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 115, + 482, + 272 + ], + "type": "text", + "content": "-gram. Due to space constraints, results are provided in the supplementary materials. In summary though, we observe a decline in average uni-gram occurrence frequency when increasing granularity (e.g., part to subpart). An interesting direction for future work is to remedy this imbalance, such as by designing parameter-efficient methods to restructure the text embedding space of LLMs to recognize hierarchical relationships, akin to MERU [10] restructuring CLIP [44]-space for image-level is-a relationships." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 299, + 482, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 299, + 482, + 347 + ], + "spans": [ + { + "bbox": [ + 130, + 299, + 482, + 347 + ], + "type": "text", + "content": "Performance When Training on Our SPIN Dataset. We next examine the potential benefit of our SPIN training data for modern models. To do so, we fine-tuned the top-performing GLaMM algorithm [46] for all hierarchy levels in SPIN, and we refer to this variant as GLaMM-FT. Results are shown in Table 2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 350, + 482, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 482, + 460 + ], + "type": "text", + "content": "Overall, we observe a considerable performance boost. For example, when prompting with general categories, we observe a 13.56 percentage point (pp) increase (123% relative increase) for subparts, a 20.76pp increase (51.90% relative increase) for parts, and a 4.77pp increase for objects (5.51% relative increase). While such performance gains are promising, we still observe very low absolute scores for subparts, indicating their challenge for modern models. We suspect a promising direction for future work is providing cost-effective ways to further increase the amount of available training data, such as through synthetic data creation and 2D projections from 3D models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": "Hierarchical Consistency Results. SpCS results are shown in Table 2. Similar trends are observed for part-object and subpart-part relationships. Part-object pairs consistently exhibit high SpCS-P2O scores, indicating models tend to correctly predict parts within their corresponding whole objects. This provides evidence that models' have some understanding of spatial containment among part-object relationships. A potential reason for this is that models are indirectly trained on part-whole relationships, because they are already taught that an object is a part of a scene. The aforementioned trend is slightly worse when examining consistency between subparts and parts (i.e., SpCS-S2P), with scores varying from 46.50 (i.e., Ferret 7B) to 84.21 (i.e., GLaMM) when employing general terms for objects and from 65.68 (i.e., Ferret 7B) to 84.96 (i.e., GLaMM) when employing more specific categories for objects. This variability, coupled with low mIoU scores for subparts, reveals that models often predict subparts outside their intended part boundaries, underscoring poor performance in properly contextualizing subparts within an object." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 424, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 424, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 424, + 91, + 447, + 100 + ], + "type": "text", + "content": "SPIN" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 455, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 455, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 455, + 128 + ], + "type": "text", + "content": "5.2 Interactive Segmentation for Vocabulary-Free Localization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 482, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 482, + 255 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 482, + 255 + ], + "type": "text", + "content": "We next assess to what extent models can localize entities at all hierarchy levels in our dataset if they are notified where to look in an image. This sets an upper bound of what is possible. To do so, we benchmark the state-of-the-art interactive segmentation model, SAM [22], which takes as input a coarse marking on an image specifying where to create the segmentation. We use bounding boxes as our coarse image markings, because they are shown to outperform alternatives [41] when using SAM. We explore the upper bound performance for this model by using the ground truth object, part, and subpart segmentations to generate the bounding boxes (i.e., by using " + }, + { + "bbox": [ + 130, + 137, + 482, + 255 + ], + "type": "inline_equation", + "content": "[x_{min}, y_{min}, x_{max}, y_{max}]" + }, + { + "bbox": [ + 130, + 137, + 482, + 255 + ], + "type": "text", + "content": " for each entity's pixels). Results are shown in Table 2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 256, + 482, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 256, + 482, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 256, + 482, + 376 + ], + "type": "text", + "content": "Despite the idealized guidance from the bounding box prompts, SAM's performance is still imperfect. As granularity increases, localization performance consistently decreases by about 20pp when going down a granularity level (e.g., objects to parts), revealing that SAM struggles more as segmentation requests become more fine-grained. Still, our findings demonstrate SAM's potential for yielding considerably higher quality hierarchical segmentations than observed from existing localization models benchmarked in the previous section. Compared to the top-performing zero-shot segmentation-based model (i.e., GLaMM), SAM achieves a boost of 38.61pp for subparts, 20.23pp for parts, and 3.75pp for objects, alongside similar SpCS scores." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 376, + 482, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 507 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 507 + ], + "type": "text", + "content": "We also evaluate a fully-automated system with SAM, by incorporating bounding box predictions from the top-performing bounding box producing model, CoGVLM. Results are shown in Table 2. We similarly observe that it outperforms the leading zero-shot segmentation method, GLaMM, in segmenting subparts and parts (9.24 percentage point (pp) boost for subparts and 10.87pp for parts). We attribute these performance gains to SAM's ability to leverage a location prior (e.g., bounding box), whereas models like GLaMM rely solely on language priors. In contrast, this approach performs worse than the leading zero-shot segmentation method, GLaMM, for objects, with a 11.22pp decrease. We attribute this performance drop to CoGVLM's inferior performance in consistently and accurately detecting bounding boxes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 525, + 455, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 455, + 537 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 455, + 537 + ], + "type": "text", + "content": "5.3 Recognizing Hierarchical Semantics in a Zero-Shot Setting" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 545, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 605 + ], + "type": "text", + "content": "We finally evaluate the ability of models to recognize hierarchical semantic labels for objects in images (i.e., without localization). We assess this in an idealized scenario in which we notify models what to look for as well as where to look in the image. This sets an upper bound of what may be possible with such models as they are explicitly told where to look." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Models. We evaluate eight off-the-shelf model variants: CoGVLM [59], Ferret-7B [64], Ferret-13B [64], Shikra [6], Osprey [66], Kosmos2 [43], ViP-Llava-7B [4], and ViP-Llava-13B [4]. All prompts to the models resemble the following basic request: \"Is there a in the ?\" for objects and \"Is there" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 239, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 239, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 239, + 102 + ], + "type": "text", + "content": "Myers-Dean et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 114, + 481, + 207 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 207 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 207 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 207 + ], + "type": "table", + "html": "
Method# Params.Prompt\\( \\mathrm{mACC}_S \\)\\( \\mathrm{mACC}_{SS} \\)\\( \\mathrm{mACC}_P \\)\\( \\mathrm{mACC}_{PS} \\)\\( \\mathrm{mACC}_O \\)\\( \\mathrm{mACC}_{OS} \\)
Shikra [6]7BBbox66.1261.2879.6774.5677.0482.05
Ferret [64]7BBbox59.0157.1959.9566.1786.7889.21
Ferret [64]13BBbox50.3169.4261.7267.8488.9389.02
CoGVLM [59]17BCoT42.1645.5481.7583.3187.1583.67
Kosmos2 [43]2.6BSG79.1264.3274.2368.5384.5578.78
Osprey [66]7BMask77.4355.9586.3081.2096.4384.85
ViP-Llava [4]7BMask53.5758.5264.1364.9671.6265.25
ViP-Llava [4]13BMask94.7897.4399.9099.8098.3395.06
", + "image_path": "608b20079d52782dd62c32f0e8176d111f05f40cb53f1018a4bc593e5936b556.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 208, + 481, + 241 + ], + "lines": [ + { + "bbox": [ + 132, + 208, + 481, + 241 + ], + "spans": [ + { + "bbox": [ + 132, + 208, + 481, + 241 + ], + "type": "text", + "content": "Table 3: Accuracy metrics for object hierarchical decomposition: general (mACC " + }, + { + "bbox": [ + 132, + 208, + 481, + 241 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 132, + 208, + 481, + 241 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 208, + 481, + 241 + ], + "type": "inline_equation", + "content": "\\mathrm{mACC}_P" + }, + { + "bbox": [ + 132, + 208, + 481, + 241 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 208, + 481, + 241 + ], + "type": "inline_equation", + "content": "\\mathrm{mACC}_O" + }, + { + "bbox": [ + 132, + 208, + 481, + 241 + ], + "type": "text", + "content": ") and specific (" + }, + { + "bbox": [ + 132, + 208, + 481, + 241 + ], + "type": "inline_equation", + "content": "\\mathrm{mACC}_{SS}" + }, + { + "bbox": [ + 132, + 208, + 481, + 241 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 208, + 481, + 241 + ], + "type": "inline_equation", + "content": "\\mathrm{mACC}_{PS}" + }, + { + "bbox": [ + 132, + 208, + 481, + 241 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 208, + 481, + 241 + ], + "type": "inline_equation", + "content": "\\mathrm{mACC}_{OS}" + }, + { + "bbox": [ + 132, + 208, + 481, + 241 + ], + "type": "text", + "content": ") categories across subparts, parts, and objects. (CoT = Chain of Thought; SG = Self-Grounding)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 274, + 482, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 274, + 482, + 370 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 482, + 370 + ], + "type": "text", + "content": "a " + }, + { + "bbox": [ + 130, + 274, + 482, + 370 + ], + "type": "inline_equation", + "content": "<(sub)part>" + }, + { + "bbox": [ + 130, + 274, + 482, + 370 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 130, + 274, + 482, + 370 + ], + "type": "inline_equation", + "content": "" + }, + { + "bbox": [ + 130, + 274, + 482, + 370 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 130, + 274, + 482, + 370 + ], + "type": "inline_equation", + "content": "?" + }, + { + "bbox": [ + 130, + 274, + 482, + 370 + ], + "type": "text", + "content": "\" for parts and subparts. Regions are specified either with segmentations (ViP-Llava, Ferret, Osprey) or bounding boxes (CoGVLM, Shikra, Kosmos2). We generate these regions directly from the ground truth annotations (e.g., bounding boxes as described in Section 5.2) for objects, parts, and subparts. We prompt models for each category known to be in the region, one at a time. An answer is 'yes' if 'yes' is present in the model's output, and otherwise 'no'. As done in Section 5.1, we test both general and specific categories in the prompts." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 383, + 473, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 383, + 473, + 395 + ], + "spans": [ + { + "bbox": [ + 132, + 383, + 473, + 395 + ], + "type": "text", + "content": "Evaluation Metrics. We use accuracy to evaluate at each granularity level." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 407, + 318, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 318, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 318, + 418 + ], + "type": "text", + "content": "Results. Results are reported in Table 3." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 419, + 482, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 419, + 482, + 586 + ], + "spans": [ + { + "bbox": [ + 130, + 419, + 482, + 586 + ], + "type": "text", + "content": "While most models struggle overall to recognize categories in images (despite being told where to look in the images), one model has nearly perfect accuracy: ViP-Llava-13B. One hypothesis for this atypically strong performance is the model lacks understanding and instead always answers yes when prompted about the presence of categories. We explored this with further experiments using adversarial examples that are instead expected to always answer \"no\". However, our findings from these experiments are none-conclusive. We observe the model can answer \"no\", but it also suffers in this case from a significant decrease in accuracy for the adversarial prompts. This performance drop could be due to a separate issue that the model struggles with negation (i.e., the reversal curse [3]). Altogether, this evaluation represents the best-case scenario for performance; it is likely to fall further in real-world applications due to imperfect human inputs or model-generated segmentations, both of which we have shown to exhibit poorer performance at finer granularity levels, as discussed in Sections 5.1 and 5.2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 587, + 482, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 587, + 482, + 636 + ], + "spans": [ + { + "bbox": [ + 130, + 587, + 482, + 636 + ], + "type": "text", + "content": "We observed mixed outcomes regarding which granularity levels are most difficult to recognize. For Ferret-13B and Osprey, an increase in granularity correlates with improved accuracy; e.g., transitioning from subparts to objects using general category names results in a 28.62pp increase for Ferret-13B and a 19.00pp" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 424, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 424, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 424, + 91, + 447, + 100 + ], + "type": "text", + "content": "SPIN" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "type": "text", + "content": "5 Results are shown in the supplementary materials for two prompts asking \"Is the category not present\" and \"Is the [different category] present\"." + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": "increase for Osprey. Conversely, for most other models, object identification consistently outperforms subpart identification. The differential performance across granularity levels may partly stem from the disparate prevalence of tokens in the training data across different granularity levels, particularly within the Llama dataset (as previously discussed). For parts and subparts, the mixed results observed—where parts sometimes outperform subparts (e.g., CoGVLM) and vice versa (e.g., Ferret-7B)—could be attributed to their relative frequency in the training data. Future research could focus on examining how foundation models differentiate responses based on prompt granularity and on enriching datasets like GRIT [64] with hierarchical relationships to improve model understanding of intra-entity relationships. Additionally, future work could explore enhancing hierarchical-relationship understanding so success at different hierarchical levels can facilitate finding entities at other granularity levels." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 272, + 482, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 272, + 482, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 482, + 429 + ], + "type": "text", + "content": "We also observed mixed outcomes when comparing the efficacy of specific versus general categories. Half the models performed better using general terms for subparts, with a similar pattern observed in parts and objects. A potential reason for this discrepancy, compared to the relatively consistent results in Sec 5.1, is that localization models often rely on feature similarities, such as CLIP [44] features. Even if a model does not capture every nuance of a category query, it still attempts to find a match based on overall similarity to known categories (e.g., mini-van vs. van) and is less likely to reject user queries [62]. Conversely, interactive understanding models, designed to maximize accuracy, may strictly search for precise category matches within their trained representations. Given our task's structure (i.e., answers are restricted to 'yes' or 'no'), these models are predisposed to outright reject the query (i.e., answer 'no') if an exact match is absent, rather than proposing close or related category matches." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 445, + 220, + 458 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 445, + 220, + 458 + ], + "spans": [ + { + "bbox": [ + 132, + 445, + 220, + 458 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 471, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 471, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 471, + 482, + 567 + ], + "type": "text", + "content": "We introduce SPIN, the first dataset challenge for hierarchical segmentation at the subpart granularity in natural images. Our analysis reveals SPIN's characteristics, our two new evaluation metrics enable gauging algorithmic proficiency in capturing the spatial and semantic relationships across hierarchy levels, and our benchmarking across three tasks demonstrates that models struggle to recognize and segment subparts in SPIN, even under idealized guidance. We publicly release SPIN to encourage further advancements in hierarchical segmentation, sharing the dataset at https://joshmyersdean.github.io/spin/index.html." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 578, + 482, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 578, + 482, + 639 + ], + "spans": [ + { + "bbox": [ + 130, + 578, + 482, + 639 + ], + "type": "text", + "content": "Acknowledgments. This work was supported by Adobe Research Gift Funds and utilized the Blanca condo computing resource at the University of Colorado Boulder. Josh Myers-Dean is supported by a NSF GRFP fellowship (#1917573). We thank the crowdworkers for contributing their time for the construction of SPIN and the authors of our benchmarked models for open-sourcing their work." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 239, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 239, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 239, + 102 + ], + "type": "text", + "content": "Myers-Dean et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 142, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 138, + 142, + 480, + 164 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 142, + 480, + 164 + ], + "spans": [ + { + "bbox": [ + 138, + 142, + 480, + 164 + ], + "type": "text", + "content": "1. Explore images. https://support.apple.com/guide/iphone/use-voiceover-for-images-and-videos-iph37e6b3844/ios" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 165, + 480, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 165, + 480, + 198 + ], + "spans": [ + { + "bbox": [ + 138, + 165, + 480, + 198 + ], + "type": "text", + "content": "2. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 198, + 480, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 198, + 480, + 232 + ], + "spans": [ + { + "bbox": [ + 138, + 198, + 480, + 232 + ], + "type": "text", + "content": "3. Berglund, L., Tong, M., Kaufmann, M., Balesni, M., Stickland, A.C., Korbak, T., Evans, O.: The reversal curse: Llms trained on\" a is b\" fail to learn\" b is a\". arXiv preprint arXiv:2309.12288 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 232, + 480, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 232, + 480, + 265 + ], + "spans": [ + { + "bbox": [ + 138, + 232, + 480, + 265 + ], + "type": "text", + "content": "4. Cai, M., Liu, H., Mustikovela, S.K., Meyer, G.P., Chai, Y., Park, D., Lee, Y.J.: Making large multimodal models understand arbitrary visual prompts. In: IEEE Conference on Computer Vision and Pattern Recognition (2024)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 266, + 480, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 266, + 480, + 299 + ], + "spans": [ + { + "bbox": [ + 138, + 266, + 480, + 299 + ], + "type": "text", + "content": "5. Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., et al.: Shapenet: An information-rich 3d moel repository. arXiv preprint arXiv:1512.03012 (2015)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 300, + 480, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 300, + 480, + 332 + ], + "spans": [ + { + "bbox": [ + 138, + 300, + 480, + 332 + ], + "type": "text", + "content": "6. Chen, K., Zhang, Z., Zeng, W., Zhang, R., Zhu, F., Zhao, R.: Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 333, + 480, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 333, + 480, + 376 + ], + "spans": [ + { + "bbox": [ + 138, + 333, + 480, + 376 + ], + "type": "text", + "content": "7. Chen, X., Mottaghi, R., Liu, X., Fidler, S., Urtasun, R., Yuille, A.: Detect what you can: Detecting and representing objects using holistic models and body parts. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1971-1978 (2014)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 376, + 480, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 376, + 480, + 431 + ], + "spans": [ + { + "bbox": [ + 138, + 376, + 480, + 431 + ], + "type": "text", + "content": "8. Deitke, M., Han, W., Herrasti, A., Kembhavi, A., Kolve, E., Mottaghi, R., Salvador, J., Schwenk, D., VanderBilt, E., Wallingford, M., Weihs, L., Yatskar, M., Farhadi, A.: Robothor: An open simulation-to-real embodied ai platform. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 432, + 480, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 432, + 480, + 465 + ], + "spans": [ + { + "bbox": [ + 138, + 432, + 480, + 465 + ], + "type": "text", + "content": "9. Deng, B., Genova, K., Yazdani, S., Bouaziz, S., Hinton, G., Tagliasacchi, A.: Cvxnet: Learnable convex decomposition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 31-44 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 466, + 480, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 466, + 480, + 498 + ], + "spans": [ + { + "bbox": [ + 138, + 466, + 480, + 498 + ], + "type": "text", + "content": "10. Desai, K., Nickel, M., Rajpurohit, T., Johnson, J., Vedantam, S.R.: Hyperbolic image-text representations. In: International Conference on Machine Learning. pp. 7694-7731. PMLR (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 499, + 480, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 499, + 480, + 543 + ], + "spans": [ + { + "bbox": [ + 138, + 499, + 480, + 543 + ], + "type": "text", + "content": "1. Ding, M., Shen, Y., Fan, L., Chen, Z., Chen, Z., Luo, P., Tenenbaum, J.B., Gan, C.: Visual dependency transformers: Dependency tree emerges from reversed attention. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14528-14539 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 544, + 480, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 544, + 480, + 586 + ], + "spans": [ + { + "bbox": [ + 138, + 544, + 480, + 586 + ], + "type": "text", + "content": "2. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 587, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 587, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 587, + 480, + 632 + ], + "type": "text", + "content": "3. Geng, H., Xu, H., Zhao, C., Xu, C., Yi, L., Huang, S., Wang, H.: Gapartnet: Cross-category domain-generalizable object perception and manipulation via generalizable and actionable parts. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7081-7091 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "text", + "content": "4. de Geus, D., Meletis, P., Lu, C., Wen, X., Dubbelman, G.: Part-aware panoptic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5485-5494 (2021)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 424, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 424, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 424, + 91, + 447, + 100 + ], + "type": "text", + "content": "SPIN" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "15. Gong, K., Liang, X., Li, Y., Chen, Y., Yang, M., Lin, L.: Instance-level human parsing via part grouping network. In: Proceedings of the European conference on computer vision (ECCV). pp. 770-785 (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 195 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 195 + ], + "type": "text", + "content": "16. He, J., Chen, J., Lin, M.X., Yu, Q., Yuille, A.L.: Compositor: Bottom-up clustering and compositing for robust part and object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11259-11268 (June 2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 196, + 480, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 196, + 480, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 196, + 480, + 228 + ], + "type": "text", + "content": "17. He, J., Yang, S., Yang, S., Kortylewski, A., Yuan, X., Chen, J.N., Liu, S., Yang, C., Yu, Q., Yuille, A.: Partimagenet: A large, high-quality dataset of parts. In: European Conference on Computer Vision. pp. 128-145. Springer (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 229, + 480, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 229, + 480, + 261 + ], + "spans": [ + { + "bbox": [ + 132, + 229, + 480, + 261 + ], + "type": "text", + "content": "18. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 262, + 480, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 262, + 480, + 296 + ], + "spans": [ + { + "bbox": [ + 132, + 262, + 480, + 296 + ], + "type": "text", + "content": "19. Hong, Y., Li, Q., Zhu, S.C., Huang, S.: VLgrammer: Grounded grammar induction of vision and language. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1665-1674 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 297, + 480, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 297, + 480, + 328 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 480, + 328 + ], + "type": "text", + "content": "20. Hong, Y., Yi, L., Tenenbaum, J., Torralba, A., Gan, C.: PTR: A benchmark for part-based conceptual, relational, and physical reasoning. Advances in Neural Information Processing Systems 34, 17427-17440 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 330, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 330, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 330, + 480, + 374 + ], + "type": "text", + "content": "21. Jia, M., Shi, M., Sirotenko, M., Cui, Y., Cardie, C., Hariharan, B., Adam, H., Belongie, S.: Fashionpedia: Ontology, segmentation, and an attribute localization dataset. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16. pp. 316-332. Springer (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 375, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 375, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 375, + 480, + 418 + ], + "type": "text", + "content": "22. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 4015-4026 (October 2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 419, + 480, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 419, + 480, + 463 + ], + "spans": [ + { + "bbox": [ + 132, + 419, + 480, + 463 + ], + "type": "text", + "content": "23. Koo, J., Huang, I., Achlioptas, P., Guibas, L.J., Sung, M.: Partglot: Learning shape part segmentation from language reference games. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16505-16514 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 464, + 480, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 480, + 485 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 480, + 485 + ], + "type": "text", + "content": "24. Lai, X., Tian, Z., Chen, Y., Li, Y., Yuan, Y., Liu, S., Jia, J.: Lisa: Reasoning segmentation via large language model. arXiv preprint arXiv:2308.00692 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 487, + 480, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 480, + 541 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 480, + 541 + ], + "type": "text", + "content": "25. Lee, J., Peng, Y.H., Herskovitz, J., Guo, A.: Image explorer: Multi-layered touch exploration to make images accessible. In: Proceedings of the 23rd International ACM SIGACCESS Conference on Computers and Accessibility. ASSETS '21, Association for Computing Machinery, New York, NY, USA (2021). https://doi.org/10.1145/3441852.3476548, https://doi.org/10.1145/3441852.3476548" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 542, + 480, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 542, + 480, + 575 + ], + "spans": [ + { + "bbox": [ + 132, + 542, + 480, + 575 + ], + "type": "text", + "content": "26. Li, F., Zhang, H., Sun, P., Zou, X., Liu, S., Yang, J., Li, C., Zhang, L., Gao, J.: Semantic-sam: Segment and recognize anything at any granularity. arXiv preprint arXiv:2307.04767 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 576, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 576, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 576, + 480, + 609 + ], + "type": "text", + "content": "27. Li, L., Zhou, T., Wang, W., Li, J., Yang, Y.: Deep hierarchical semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1246-1257 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 609, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 480, + 632 + ], + "type": "text", + "content": "28. Li, T., Gupta, V., Mehta, M., Srikumar, V.: A logic-driven framework for consistency of neural models. arXiv preprint arXiv:1909.00126 (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "29. Li, X., Xu, S., Yang, Y., Cheng, G., Tong, Y., Tao, D.: Panoptic-partformer: Learning a unified model for panoptic part segmentation. In: European Conference on Computer Vision. pp. 729-747. Springer (2022)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 239, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 239, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 239, + 101 + ], + "type": "text", + "content": "Myers-Dean et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "30. Li, X., Xu, S., Yang, Y., Yuan, H., Cheng, G., Tong, Y., Lin, Z., Yang, M.H., Tao, D.: Panopticpartformer++: A unified and decoupled view for panoptic part segmentation. arXiv preprint arXiv:2301.00954 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 481, + 183 + ], + "type": "text", + "content": "31. Liang, X., Gong, K., Shen, X., Lin, L.: Look into person: Joint body parsing & pose estimation network and a new benchmark. IEEE transactions on pattern analysis and machine intelligence 41(4), 871-885 (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 481, + 226 + ], + "type": "text", + "content": "32. Liang, X., Shen, X., Feng, J., Lin, L., Yan, S.: Semantic object parsing with graph lstm. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14. pp. 125-143. Springer (2016)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 226, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 226, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 481, + 270 + ], + "type": "text", + "content": "33. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 270, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 481, + 303 + ], + "type": "text", + "content": "34. Liu, J., Min, S., Zettlemoyer, L., Choi, Y., Hajishirzi, H.: Infini-gram: Scaling unbounded n-gram language models to a trillion tokens. arXiv preprint arXiv:2401.17377 (2024)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 303, + 481, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 303, + 481, + 336 + ], + "spans": [ + { + "bbox": [ + 130, + 303, + 481, + 336 + ], + "type": "text", + "content": "35. Liu, Q., Kortylewski, A., Zhang, Z., Li, Z., Guo, M., Liu, Q., Yuan, X., Mu, J., Qiu, W., Yuille, A.: Cgpart: A part segmentation dataset based on 3d computer graphics models. arXiv preprint arXiv:2103.14098 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 336, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 336, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 336, + 481, + 380 + ], + "type": "text", + "content": "36. Martin, D., Fowlkes, C., Tal, D., Malik, J.: A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In: Proceedings Eighth IEEE International Conference on Computer Vision. ICCV 2001. vol. 2, pp. 416-423. IEEE (2001)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 380, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 481, + 413 + ], + "type": "text", + "content": "37. Michieli, U., Zanuttigh, P.: Edge-aware graph matching network for part-based semantic segmentation. International Journal of Computer Vision 130(11), 2797-2821 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 413, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 481, + 445 + ], + "type": "text", + "content": "38. Miller, G.A.: Wordnet: a lexical database for english. Commun. ACM 38(11), 39-41 (nov 1995). https://doi.org/10.1145/219717.219748, https://doi.org/10.1145/219717.219748" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 445, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 445, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 481, + 490 + ], + "type": "text", + "content": "39. Mo, K., Zhu, S., Chang, A.X., Yi, L., Tripathi, S., Guibas, L.J., Su, H.: Partnet: A large-scale benchmark for fine-grained and hierarchical part-level 3d object understanding. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 909-918 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 490, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 490, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 490, + 481, + 533 + ], + "type": "text", + "content": "40. Mo, K., Zhu, S., Chang, A.X., Yi, L., Tripathi, S., Guibas, L.J., Su, H.: PartNet: A large-scale benchmark for fine-grained and hierarchical part-level 3D object understanding. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 533, + 481, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 481, + 577 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 481, + 577 + ], + "type": "text", + "content": "41. Myers-Dean, J., Fan, Y., Price, B., Chan, W., Gurari, D.: Interactive segmentation for diverse gesture types without context. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 7198-7208 (January 2024)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 577, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 577, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 577, + 481, + 632 + ], + "type": "text", + "content": "42. Nair, V., Zhu, H.H., Smith, B.A.: Imageassist: Tools for enhancing touchscreen-based image exploration systems for blind and low vision users. In: Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems. CHI '23, Association for Computing Machinery, New York, NY, USA (2023). https://doi.org/10.1145/3544548.3581302, https://doi.org/10.1145/3544548.3581302" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 632, + 481, + 665 + ], + "type": "text", + "content": "43. Peng, Z., Wang, W., Dong, L., Hao, Y., Huang, S., Ma, S., Wei, F.: Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824 (2023)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 424, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 424, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 424, + 91, + 447, + 100 + ], + "type": "text", + "content": "SPIN" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "text", + "content": "44. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 161, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 481, + 205 + ], + "type": "text", + "content": "45. Ramanathan, V., Kalia, A., Petrovic, V., Wen, Y., Zheng, B., Guo, B., Wang, R., Marquez, A., Kovvuri, R., Kadian, A., et al.: Paco: Parts and attributes of common objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7141-7151 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "text", + "content": "46. Rasheed, H., Maaz, M., Shaji, S., Shaker, A., Khan, S., Cholakkal, H., Anwer, R.M., Xing, E., Yang, M.H., Khan, F.S.: Glamm: Pixel grounding large multimodal model. arXiv preprint arXiv:2311.03356 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 238, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 481, + 259 + ], + "type": "text", + "content": "47. RAYMOND W. GIBBS, J., Matlock, T.: Psycholinguistics and mental representations (2000)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "type": "text", + "content": "48. Ren, Z., Huang, Z., Wei, Y., Zhao, Y., Fu, D., Feng, J., Jin, X.: Pixelm: Pixel reasoning with large multimodal model (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 282, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 303 + ], + "type": "text", + "content": "49. Sennrich, R., Haddow, B., Birch, A.: Neural machine translation of rare words with subword units. arXiv preprint arXiv:1508.07909 (2015)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 346 + ], + "type": "text", + "content": "50. Song, X., Wang, P., Zhou, D., Zhu, R., Guan, C., Dai, Y., Su, H., Li, H., Yang, R.: Apollocar3d: A large 3d car instance understanding benchmark for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5452-5462 (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 347, + 481, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 391 + ], + "type": "text", + "content": "51. Sun, P., Chen, S., Zhu, C., Xiao, F., Luo, P., Xie, S., Yan, Z.: Going denser with open-vocabulary part segmentation. 2023 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 15407-15419 (2023), https://api.sementicscholar.org/CorpusID:258762519" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 392, + 481, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 481, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 481, + 424 + ], + "type": "text", + "content": "52. Tang, C., Xie, L., Zhang, X., Hu, X., Tian, Q.: Visual recognition by request. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15265-15274 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 425, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 425, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 425, + 481, + 445 + ], + "type": "text", + "content": "53. Tong, K., Wu, Y.: Deep learning-based detection from the perspective of small or tiny objects: A survey. Image and Vision Computing 123, 104471 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 447, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 481, + 479 + ], + "type": "text", + "content": "54. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., Lample, G.: Llama: Open and efficient foundation language models (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "text", + "content": "55. Tsogkas, S., Kokkinos, I., Papandreou, G., Vedaldi, A.: Deep learning for semantic part segmentation with high-level guidance. arXiv preprint arXiv:1505.02438 (2015)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "type": "text", + "content": "56. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The caltech-ucsd birds-200-2011 dataset (2011)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "type": "text", + "content": "57. Wang, J., Yuille, A.L.: Semantic part segmentation using compositional model combining shape and appearance. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1788-1797 (2015)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "type": "text", + "content": "58. Wang, P., Shen, X., Lin, Z., Cohen, S., Price, B., Yuille, A.L.: Joint object and part segmentation using deep learned potentials. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 1573-1581 (2015)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "type": "text", + "content": "59. Wang, W., Lv, Q., Yu, W., Hong, W., Qi, J., Wang, Y., Ji, J., Yang, Z., Zhao, L., Song, X., et al.: Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "text", + "content": "60. Wang, X., Li, S., Kallidromitis, K., Kato, Y., Kozuka, K., Darrell, T.: Hierarchical open-vocabulary universal image segmentation. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 239, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 239, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 239, + 102 + ], + "type": "text", + "content": "Myers-Dean et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 436 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "61. Wei, M., Yue, X., Zhang, W., Kong, S., Liu, X., Pang, J.: Ov-parts: Towards open-vocabulary part segmentation. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "type": "text", + "content": "62. Wu, T.H., Biamby, G., Chan, D., Dunlap, L., Gupta, R., Wang, X., Gonzalez, J.E., Darrell, T.: See, say, and segment: Teaching lmms to overcome false premises. arXiv preprint arXiv:2312.08366 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 182, + 482, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 482, + 226 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 482, + 226 + ], + "type": "text", + "content": "63. Xiang, F., Qin, Y., Mo, K., Xia, Y., Zhu, H., Liu, F., Liu, M., Jiang, H., Yuan, Y., Wang, H., Yi, L., Chang, A.X., Guibas, L.J., Su, H.: SAPIEN: A simulated part-based interactive environment. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 226, + 482, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 226, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 482, + 259 + ], + "type": "text", + "content": "64. You, H., Zhang, H., Gan, Z., Du, X., Zhang, B., Wang, Z., Cao, L., Chang, S.F., Yang, Y.: Ferret: Refer and ground anything anywhere at any granularity. arXiv preprint arXiv:2310.07704 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 259, + 482, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 482, + 303 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 482, + 303 + ], + "type": "text", + "content": "65. Yu, F., Liu, K., Zhang, Y., Zhu, C., Xu, K.: Partnet: A recursive part decomposition network for fine-grained and hierarchical shape segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9491-9500 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 303, + 482, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 303, + 482, + 325 + ], + "spans": [ + { + "bbox": [ + 130, + 303, + 482, + 325 + ], + "type": "text", + "content": "66. Yuan, Y., Li, W., Liu, J., Tang, D., Luo, X., Qin, C., Zhang, L., Zhu, J.: Osprey: Pixel understanding with visual instruction tuning (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 325, + 482, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 325, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 325, + 482, + 369 + ], + "type": "text", + "content": "67. Zhao, J., Li, J., Cheng, Y., Sim, T., Yan, S., Feng, J.: Understanding humans in crowded scenes: Deep nested adversarial learning and a new benchmark for multi-human parsing. In: Proceedings of the 26th ACM international conference on Multimedia. pp. 792-800 (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 369, + 482, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 482, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 482, + 402 + ], + "type": "text", + "content": "68. Zheng, S., Yang, F., Kiapour, M.H., Piramuthu, R.: Modanet: A large-scale street fashion dataset with polygon annotations. In: Proceedings of the 26th ACM international conference on Multimedia. pp. 1670-1678 (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 402, + 482, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 436 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 436 + ], + "type": "text", + "content": "69. Zhou, B., Zhao, H., Puig, X., Fidler, S., Barriuso, A., Torralba, A.: Scene parsing through ade20k dataset. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 633-641 (2017)" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 424, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 424, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 424, + 91, + 447, + 100 + ], + "type": "text", + "content": "SPIN" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/SPIRE_ Semantic Prompt-Driven Image Restoration/0f07d6c7-f604-4c26-b13e-8eef2743217e_content_list.json b/2024/SPIRE_ Semantic Prompt-Driven Image Restoration/0f07d6c7-f604-4c26-b13e-8eef2743217e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..171e77ec4a6e030381eb55283b8c576af1056f3b --- /dev/null +++ b/2024/SPIRE_ Semantic Prompt-Driven Image Restoration/0f07d6c7-f604-4c26-b13e-8eef2743217e_content_list.json @@ -0,0 +1,1959 @@ +[ + { + "type": "text", + "text": "SPIRE: Semantic Prompt-Driven Image Restoration", + "text_level": 1, + "bbox": [ + 272, + 140, + 732, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chenyang Qi $^{1,2*}$ , Zhengzhong Tu $^{1}$ , Keren Ye $^{1}$ , Maurizio Delbracio $^{1}$ , Peyman Milanfar $^{1}$ , Qifeng Chen $^{2}$ , and Hossein Talebi $^{1}$", + "bbox": [ + 235, + 212, + 767, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Google Research 2 HKUST", + "bbox": [ + 436, + 253, + 563, + 282 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/87bac0dffcf974577e47bdf79cb79aef0d83d7e7446a26a027cd557068a560f0.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 217, + 301, + 312, + 375 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d15147ed5a186f3928ae2c99ae961fecedfd982337ec7cb55baeeb2847920387.jpg", + "image_caption": [ + "\"Remove all degradation\"" + ], + "image_footnote": [], + "bbox": [ + 313, + 301, + 406, + 375 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4bc8cc66a65671967468e08441026a8e939bb4cc2d52cbf068519985e6bd32f1.jpg", + "image_caption": [ + "\"Remove all degradation\"" + ], + "image_footnote": [], + "bbox": [ + 406, + 301, + 498, + 375 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f3ba838f50d8cd2753e78188424782a28daf0f479aaa2090f3c5953b64b5f94b.jpg", + "image_caption": [ + "\"Upsample...denoise..." + ], + "image_footnote": [], + "bbox": [ + 500, + 301, + 591, + 375 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5477461e3917d0f66a58645f62be7077c34bff1d2c656e28979dab7e1fad73cc.jpg", + "image_caption": [ + "'Upsample...denoise...'" + ], + "image_footnote": [], + "bbox": [ + 591, + 301, + 684, + 375 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/db2f022c99661c46e9278bfff823cadb716c1574182b3f6727dfd83db7fd181e.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 686, + 301, + 779, + 375 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ample..., denoise...", + "bbox": [ + 625, + 376, + 674, + 383 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/533ae31e3b666ae6a20b4cf74980c90237eeedca88d214a3ca7301384534a786.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 217, + 388, + 310, + 462 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/670b4166754aa4503b6d0466d2dc19929cac7f1221541ced98ca7cb42ba4404f.jpg", + "image_caption": [ + "\"Remove all degradation\"" + ], + "image_footnote": [], + "bbox": [ + 313, + 388, + 406, + 462 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/51c8ef5cde5d49efe49a1ca7f128b92711040cc67e294d5d7a1e15775ab4465c.jpg", + "image_caption": [ + "\"Remove all degradation\"" + ], + "image_footnote": [], + "bbox": [ + 406, + 388, + 498, + 462 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/480ac784c13b5c29093dd41675616b439a9a33ac49ea4c461b48707fa3929693.jpg", + "image_caption": [ + "\"Deblur...denoise...\"" + ], + "image_footnote": [], + "bbox": [ + 500, + 388, + 591, + 462 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/39176c35dc533ef2e0bd694774883084084ea39412df211341e344ab9ad30bf5.jpg", + "image_caption": [ + "\"Deblur...denoise...\"" + ], + "image_footnote": [], + "bbox": [ + 591, + 388, + 684, + 462 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bb9ac5602ac7ac436890df22ffb4eebed637e32e13124a80ab3350f8910b8847.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 686, + 388, + 779, + 462 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b6cf43e0d633afc09beefad697e1af572861e45414f02e78d1ba9297b28a7c97.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 217, + 477, + 310, + 550 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7cb67d3420abd0d8021549e8b2a23d5d84dd3bcf0b200d8c5ceb2a96281553b5.jpg", + "image_caption": [ + "$\\because {a}_{4} = {2}^{3} \\cdot {2}^{2}$" + ], + "image_footnote": [], + "bbox": [ + 313, + 477, + 406, + 550 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/888052140f6dbc026b92c03600d9936ef2673866e50dcddc4fe7e4c1b08cf5ff.jpg", + "image_caption": [ + "\"DANGER...\"" + ], + "image_footnote": [], + "bbox": [ + 406, + 477, + 498, + 550 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fc2d2f3fe4cfc2197285d8c9ea412eb69693c85971f9eeb31d332b6ff5e44f02.jpg", + "image_caption": [ + "\"EXIT...\"", + "Fig. 1: We present SPIRE: Semantic Prompt-Driven Image Restoration, a text-based foundation model for all-in-one, instructed image restoration. SPIRE allows users to flexibly leverage either semantic-level content prompt, or quantitative degradation-aware restoration prompt, or both, to obtain their desired enhancement results based on personal preferences. In other words, SPIRE can be easily prompted to conduct blind restoration, semantic restoration, or task-specific granular treatment. Our framework also enables a new paradigm of instruction-based image restoration, providing a reliable evaluation benchmark to facilitate vision-language models for low-level computational photography applications." + ], + "image_footnote": [], + "bbox": [ + 500, + 477, + 591, + 550 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/807ff203776fab889a2b54635f0969f4016753ae9231555252fe55568a52c377.jpg", + "image_caption": [ + "\"GO...\"", + "\"STOP..." + ], + "image_footnote": [], + "bbox": [ + 591, + 477, + 684, + 550 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Text-driven diffusion models have become increasingly popular for various image editing tasks, including inpainting, stylization, and object replacement. However, it still remains an open research problem to adopt this language-vision paradigm for more fine-level image processing tasks, such as denoising, super-resolution, deblurring, and compression artifact removal. In this paper, we develop SPIRE, a Semantic and restoration Prompt-driven Image Restoration framework that leverages", + "bbox": [ + 259, + 714, + 740, + 811 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*This work was done during an internship at Google Research.", + "bbox": [ + 230, + 825, + 650, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "natural language as a user-friendly interface to control the image restoration process. We consider the capacity of prompt information in two dimensions. First, we use content-related prompts to enhance the semantic alignment, effectively alleviating identity ambiguity in the restoration outcomes. Second, our approach is the first framework that supports fine-level instruction through language-based quantitative specification of the restoration strength, without the need for explicit task-specific design. In addition, we introduce a novel fusion mechanism that augments the existing ControlNet architecture by learning to rescale the generative prior, thereby achieving better restoration fidelity. Our extensive experiments demonstrate the superior restoration performance of SPIRE compared to the state of the arts, alongside offering the flexibility of text-based control over the restoration effects.", + "bbox": [ + 259, + 147, + 743, + 327 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 354, + 375, + 369 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image restoration or enhancement aims to recover high-quality, pixel-level details from a degraded image, while preserving as much the original semantic information as possible. Although neural network models [11, 12, 30, 63, 67, 71, 73, 74] have marked significant progress, it still remains challenging to design an effective task-conditioning mechanism instead of building multiple individual models for each task (such as denoise, deblur, compression artifact removal) in practical scenarios. The advancement of text-driven diffusion models [48, 50, 53] has unveiled the potential of natural language as a universal input condition for a broad range of image processing challenges, which improves interactivity with users and reduces the cost of task-specific fine-tuning. However, the existing applications of natural language in stylization [38, 44], and inpainting [3, 4, 10, 21, 44, 64] predominantly focus on high-level semantic editing, whereas the uniqueness and challenges for low-level image processing have been less explored.", + "bbox": [ + 212, + 386, + 787, + 582 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Natural language text prompts in image restoration can play two crucial roles: alleviating semantic ambiguities and resolving degradation type ambiguities. Firstly, the same degraded image can correspond to different visual objects, leading to ambiguities in content interpretation (e.g., discerning whether the blurred animal in Fig. 1 is a horse or a zebra). Typically, unconditional image-to-image restoration leads to a random or average estimation of the identities, resulting in neither fish nor fowl. Secondly, certain photographic effects that are deliberately introduced for aesthetic purposes, such as bokeh with a soft out-of-focus background, can be misconstrued as blur distortions by many existing deblurring models, leading to unwanted artifacts or severe hallucinations in the outputs. Although blind restoration methods can produce clean images by either leveraging frozen generative priors [67,70,77], or using end-to-end regression [71], they do not consider the aforementioned ambiguities.", + "bbox": [ + 212, + 584, + 787, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we introduce SPIRE—a Semantic Prompt-Driven Image REstoration framework that provides a user-friendly interface to fully control both the restored image semantics and the enhancement granularity using natural language instructions. Traditional blind restoration without prompts (Fig", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Qi et al.", + "bbox": [ + 271, + 114, + 328, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ure 1) has limited flexibility in solving semantic and degradation ambiguity, thus tends to generate average blurry images. Leveraging the semantic and quantitative restoration prompts, we show that the visual quality can be significantly improved. Moreover, our framework can also be used as blind restoration with null prompts. In addition to higher perceptual quality, our framework provides the flexibility for users to generate more than one plausible result. Specifically, in imaging scenarios with large levels of degradation there are multiple plausible solutions. The proposed interactive framework has the capability to personalize the output by leveraging input prompts according to user preference. In concurrent studies, the focus has typically been isolated to one of three areas: either blind restoration [34,67], only semantic prompting [78], or only discrete restoration types [9,36]. To the best of our knowledge, this is the first unified model to support the following three distinct features simultaneously:", + "bbox": [ + 212, + 146, + 787, + 343 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Blind Restoration: When instructed with a general restoration prompt \"remove all degradations\" and empty semantic prompt \"\", SPIRE operates as a blind restoration model (\"Ours w/o text\" in Tab. 1).", + "2. Semantic Restoration: When provided with a text description of the (desired) visual content, SPIRE concentrates on restoring the specified identity of the uncertain or ambiguous objects in the degraded image.", + "3. Quantitative Task-Specific Restoration: Receiving specific restoration type hints (e.g., \"deblur...\", \"deblur... denoise...\"), SPIRE transforms into a task-specific model (e.g., deblur, denoise, or both). Moreover, it understands the numeric nuances of different degradation parameters in language prompts so that the users can control the restoration strength (e.g., \"deblur with sigma 3.0\" in Fig. 7)." + ], + "bbox": [ + 222, + 356, + 785, + 539 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To train SPIRE, we first build a synthetic data generation pipeline on a large-scale text-image dataset [8] upon the second-order degradation process proposed in Real-ESRGAN [71]. Additionally, we embed the degradation parameters into the restoration prompts to encode finer-grained degradation information (e.g., \"deblur with sigma 3.5\"). Our approach provides richer degradation-specific information compared to contemporary works [26,36] which only employ the degradation types (e.g., \"gaussian blur\"). We then finetune a ControlNet adaptor [84]—which learns a parameter-efficient, parallel branch on top of the latent diffusion models (LDMs) [50]—on a mixture of diverse restoration tasks (see Fig. 1). The semantic text prompts are processed through the LDM backbone, aligning with LDM's text-to-image pretraining. The degradation text prompts and image-conditioning are implemented in the ControlNet branch, as these aspects are not inherently addressed by the vanilla LDM. We further improve the ControlNet by introducing a new modulation connection that adaptively fuses the degradation condition with the generative prior with only a few extra trainable parameters yet showing impressive performance gain.", + "bbox": [ + 212, + 551, + 787, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our extensive experiments demonstrate remarkable restoration quality achieved by our proposed SPIRE method. Moreover, SPIRE offers additional freedom for both content and restoration prompt engineering. For example, the semantic", + "bbox": [ + 212, + 795, + 802, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "SPIRE: Semantic Prompt-Driven Image Restoration", + "bbox": [ + 380, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "prompt \"a very large giraffe eating leaves\" helps to resolve the semantic ambiguity in Fig. 2; the degradation prompt \"deblur with sigma 3.0\" reduces the gaussian blur while maintaining the intentional motion blur in Fig. 3. We also find out that our model learns a continuous latent space of the restoration strength in Fig. 7, even though we do not explicitly fine-tune the CLIP embedding. Our contribution can be summarized as follows:", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce the first unified text-driven image restoration model that supports both semantic prompts and restoration instructions. Our experiments demonstrate that incorporating semantic prompts and restoration instructions significantly enhances the restoration quality.", + "- Our proposed paradigm empowers users to fully control the semantic outcome of the restored image using different semantic prompts during test time.", + "- Our proposed approach provides a mechanism for users to adjust the category and strength of the restoration effect based on their subjective preferences.", + "- We demonstrate that text can serve as universal guidance control for low-level image restoration, eliminating the need for task-specific model design." + ], + "bbox": [ + 212, + 244, + 782, + 392 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 214, + 416, + 387, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "First, we review the literature on image restoration. Our proposed framework aims to address some of the limitations of the existing restoration methods. Next, we compare multiple text-guided diffusion editing methods. Our method is designed to deal with two different text ambiguities concurrently, which is an unprecedented challenge for existing text-driven diffusion methods. Finally, we discuss the parameter-efficient fine-tuning of diffusion models, which further motivates our proposed modulation connection.", + "bbox": [ + 212, + 446, + 787, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Image Restoration is the task of recovering a high-resolution, clean image from a degraded input. Pioneering works in the fields of super-resolution [14,32], motion and defocus deblurring [1,61,79], denoising [82], and JPEG and artifact removal [16, 45] primarily utilize deep neural network architectures to address specific tasks. Later, Transformer-based [30,63,73] and adversarial-based [71] formulations were explored, demonstrating state-of-the-art performance with unified model architecture. Recently there has been a focus on exploiting iterative restorations, such as the ones from diffusion models, to generate images with much higher perceptual quality [11,49,52,54,74,75,78].", + "bbox": [ + 212, + 553, + 787, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "There have been some recent works attempting to control the enhancement strength of a single model, such as the scaling factor condition in arbitrary super-resolution [68, 69] and noise-level map for image denoising [83]. However, these methods require hand-crafting dedicated architectures for each task, which limits their scalability. In contrast to the task-specific models, other approaches seek to train a single blind model by building a pipeline composed of several classical degradation process, such as BSRGAN [80] and Real-ESRGAN [71].", + "bbox": [ + 212, + 689, + 787, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Text-guided Diffusion Editing. Denoising diffusion models are trained to reverse the noisiging process [23, 57, 58]. Early methods focused on unconditioned generation [13], but recent trends pivot more to conditioned generation, such as", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Qi et al.", + "bbox": [ + 271, + 114, + 328, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/271b8020dfdd93e0071501f87af5c4adb0823db53ee96ea6e256577234cfe22d.jpg", + "image_caption": [ + "Fig. 2: Framework of SPIRE. In the training phase, we begin by synthesizing a degraded version $y$ , of a clean image $x$ . Our degradation synthesis pipeline also creates a restoration prompt $c_{r}$ , which contains numeric parameters that reflects the intensity of the degradation introduced. Then, we inject the synthetic restoration prompt into a ControlNet adaptor, which uses our proposed modulation fusion blocks $(\\gamma, \\beta)$ to connect with the frozen backbone driven by the semantic prompt $c_{s}$ . During test time, the users can employ the SPIRE framework as either a blind restoration model with restoration prompt \"Remove all degradation\" and empty semantic prompt $\\varnothing$ , or manually adjust the restoration $c_{r}$ and semantic prompts $c_{s}$ to obtain what they want." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 781, + 279 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "image-to-image translation [52, 55] and text-to-image generation [50, 53]. Latent Diffusion [50] is a groundbreaking approach that introduces a versatile framework for improving both training and sampling efficiency, while flexible enough for general conditioning inputs. Subsequent works have built upon their pretrained text-to-image checkpoints, and designed customized architectures for different tasks like text-guided image editing. For instance, SDEdit [37] generates content for a new prompt by adding noise to the input image. Attention-based methods [21, 44, 64, 81] show that images can be edited via reweighting and replacing the cross-attention map of different prompts. Aside from the previously mentioned approaches driven by target prompts, instruction-based editing [7, 17] entail modifying a source image based on specific instructions. These textual instructions are typically synthesized using large language models [41, 47]. As CLIP serves as a cornerstone for bridging vision and language, several studies have extensively investigated the representation space of CLIP embeddings [46]. Instead of directly applying CLIP to fundamental discriminative tasks [19, 85], these works either tailor CLIP to specific applications, such as image quality assessment [28, 31], serve as semantic guidance in generative adversarial network [5] or aligning image embedding with degradation types in the feature space [26, 36]. To enhance CLIP's ability to understand numerical information, some works [42, 43] finetune it using contrastive learning on synthetic numerical data, and then incorporate the fine-tuned text embedding into diffusion training.", + "bbox": [ + 212, + 433, + 787, + 752 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Parameter-Efficient Diffusion Model Finetuning. To leverage the powerful generative prior in pretrained diffusion models, parameter-efficient components such as text embedding [15], low-rank approximations of model weights [20, 25], and cross attention layers [29] can be finetuned to personalize the pretrained model. Adaptor-based finetuning paradigms [39, 84] propose to keep the original", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "SPIRE: Semantic Prompt-Driven Image Restoration", + "bbox": [ + 382, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "UNet weights frozen, and add new image or text conditions. This adaptor generates residual features that are subsequently added to the frozen UNet backbone.", + "bbox": [ + 215, + 146, + 782, + 176 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 200, + 330, + 217 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We introduce a universal approach to combine the above mentioned task-specific, strength-aware, and blind restoration methods within a unified framework (illustrated in Sec. 3.2). We further propose to decouple the learning of content and restoration prompts to better preserve the pre-trained prior while injecting new conditions, as unfolded in Sec. 3.3. Sec. 3.4 details our design on how to accurately control the restoration type and strength, as well as our proposed modulation fusion layer that adaptively fuses the restoration features back to the frozen backbone. We start with some preliminaries in the following section.", + "bbox": [ + 212, + 233, + 782, + 354 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.1 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 377, + 372, + 391 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Latent Diffusion Models (LDMs) [50] are probabilistic generative models that learn the underlying data distribution by iteratively removing Gaussian noise in the latent space, which is typically learned using a VAE autoencoder. Formally, the VAE encoder $\\mathcal{E}$ compresses an input image $\\pmb{x}$ into a compact latent representation $\\pmb{z} = \\mathcal{E}(\\pmb{x})$ , which can be later decoded back to the pixel space using the coupled VAE decoder $\\mathcal{D}$ , often learned under an image reconstruction objective: $\\mathcal{D}(\\mathcal{E}(\\pmb{x})) \\approx \\pmb{x}$ . During training stage, the output of a UNet [51] $\\epsilon_{\\theta}(z_t, t, \\pmb{y})$ conditioned on $\\pmb{y}$ (such as text, images, etc.) is parameterized [13, 56] to remove Gaussian noise $\\pmb{\\epsilon}$ in the latent space $\\pmb{z}_t$ as:", + "bbox": [ + 212, + 402, + 784, + 539 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} \\mathbb {E} _ {\\left(\\boldsymbol {z} _ {0}, \\boldsymbol {y}\\right) \\sim p _ {\\text {d a t a}}, \\epsilon \\sim \\mathcal {N} (0, I), t} \\| \\boldsymbol {\\epsilon} - \\boldsymbol {\\epsilon} _ {\\theta} \\left(\\boldsymbol {z} _ {t}, t, \\boldsymbol {y}\\right) \\| _ {2} ^ {2}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 339, + 550, + 784, + 573 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{z}_t$ is a noisy sample of $\\mathbf{z}_0$ at sampled timestep $t$ . The condition $\\mathbf{y}$ is randomly dropped out as $\\varnothing$ to make the model unconditional.", + "bbox": [ + 212, + 584, + 782, + 613 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "At test time, deterministic DDIM sampling [57] is utilized to convert a random noise $z_{T}$ to a clean latent $z_{0}$ which is decoded as final result $\\mathcal{D}(z_0)$ . In each timestep $t$ , classifier-free guidance [13, 24] can be applied to trade-off sample quality and condition alignment:", + "bbox": [ + 212, + 614, + 782, + 675 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\overline {{\\epsilon}} _ {\\theta} \\left(\\boldsymbol {z} _ {t}, t, \\boldsymbol {y}\\right) = \\epsilon_ {\\theta} \\left(\\boldsymbol {z} _ {t}, t, \\varnothing\\right) + w \\left(\\epsilon_ {\\theta} \\left(\\boldsymbol {z} _ {t}, t, \\boldsymbol {y}\\right) - \\epsilon_ {\\theta} \\left(\\boldsymbol {z} _ {t}, t, \\varnothing\\right)\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 300, + 688, + 692, + 704 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $w$ is a scalar to adjust the guidance strength of $y$ . Note that the estimated noise $\\overline{\\epsilon}_{\\theta}(\\pmb{z}_t, t, \\pmb{y})$ is used to update $z_t$ to $z_{t-1}$ , which is a approximation of the distribution gradient score [60] as $\\nabla_{z_t} \\log p(\\pmb{z}_t | \\pmb{y}) \\propto \\overline{\\epsilon}_{\\theta}(\\pmb{z}_t, t, \\pmb{y})$ .", + "bbox": [ + 212, + 715, + 782, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Text-driven Image Restoration", + "text_level": 1, + "bbox": [ + 215, + 784, + 517, + 799 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Based on the LDM framework, we propose a new restoration paradigm—text-driven image restoration. Our method targets to restore images $(\\pmb{x}$ or $\\pmb{z}_0)$ based", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Qi et al.", + "bbox": [ + 271, + 114, + 328, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "on conditions $\\{\\pmb{y},\\pmb{c}_s,\\pmb{c}_r\\}$ . Specifically: $\\pmb{y}$ denotes the degraded image condition, $\\pmb{c}_s$ is the semantic prompt describing the clean image $\\pmb{x}$ (e.g., \"a panda is sitting by the bamboo\" or \"a panda\"), and $\\pmb{c}_r$ is the restoration prompt that describes the details of the degradation in terms of both the operation and parameter (e.g., \"deblur with sigma 3.0\"). We use $\\pmb{y} = \\mathrm{Deg}(\\pmb{x},\\pmb{c}_r)$ to denote the degradation process which turns the clean image $\\pmb{x}$ into its degraded counterpart $\\pmb{y}_i$ .", + "bbox": [ + 212, + 146, + 782, + 236 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The above text-driven image restoration model $p(\\pmb{z}_t|\\{\\pmb{y},\\pmb{c}_s,\\pmb{c}_r\\})$ can be trained using paired data. We use the text-image dataset [8] so that each clean image $\\pmb{x}$ is paired with a semantic prompt $\\pmb{c}_s$ (the paired alt-text). Then $\\pmb{y} = \\mathrm{Deg}(\\pmb{x},\\pmb{c}_r)$ is simulated using the synthetic degradation [71] pipeline (Sec. 4.1), yielding the final paired training data $(\\pmb{x} \\text{ or } \\pmb{z}_0, \\{\\pmb{y},\\pmb{c}_s,\\pmb{c}_r\\}) \\sim p_{data}$ .", + "bbox": [ + 212, + 237, + 787, + 314 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison to Blind Restoration. Recent existing blind restoration models [34, 55, 67] also leverage diffusion priors to generate high-quality images. However, these methods are prone to hallucinate unwanted, unnatural, or oversharpened textures given that the semantic- and degradation-ambiguities persist everywhere. Fig. 3 provides an example, where the input image is degraded simultaneously with Gaussian blur and motion blur. The fully blind restoration method removes all of the degradation, leaving no way to steer the model to remove only the Gaussian blur while preserving the motion effect for aesthetic purposes. Therefore, having the flexibility to control the restoration behavior and its strength at test time is a crucial requirement. Our text-driven formulation of the problem introduces additional controlling ability using both the semantic content and the restoration instruction. In our SPIRE design, $\\pmb{c}_r =$ \"Remove all degradation\" is also a plausible restoration prompt which makes our SPIRE to be compatible with the blind restoration. Further, our text-driven restoration takes advantages of the pretrained content-aware LDM using the additional semantic prompt $\\pmb{c}_s$ , hence able to better handle semantic ambiguities on the noisy, blurry regions whose visual objects are less recognizable.", + "bbox": [ + 212, + 316, + 787, + 574 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Decoupling Semantic and Restoration Prompts", + "text_level": 1, + "bbox": [ + 215, + 597, + 651, + 614 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To effectively learn the latent distribution $p(\\pmb{z}_t | \\pmb{y}, \\pmb{c}_s, \\pmb{c}_r)$ , we further decouple the conditions $\\{\\pmb{y}, \\pmb{c}_s, \\pmb{c}_r\\}$ into two groups—one for the text-to-image prior $(\\pmb{c}_s \\rightarrow \\pmb{z}_t)$ already imbued in the pretrained LDM model, and the other for both the image-to-image $(\\pmb{y} \\rightarrow \\pmb{z}_t)$ and restoration-to-image $(\\pmb{c}_r \\rightarrow \\pmb{z}_t)$ that needs to be learnt from the synthetic data. This decoupling strategy prevents catastrophic forgetting in the pretrained diffusion model and enables independent training of the restoration-aware model, whose score function [60] is expressed as:", + "bbox": [ + 212, + 625, + 787, + 729 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {z _ {t}} \\log p (\\boldsymbol {z} _ {t} | \\boldsymbol {y}, \\boldsymbol {c} _ {s}, \\boldsymbol {c} _ {r}) \\approx \\underbrace {\\nabla_ {z _ {t}} \\log p (\\boldsymbol {z} _ {t} | \\boldsymbol {c} _ {s})} _ {\\text {S e m a n t i c - a w a r e (f r o z e n)}} + \\underbrace {\\nabla_ {z _ {t}} \\log p (\\boldsymbol {y} | \\boldsymbol {z} _ {t} , \\boldsymbol {c} _ {r})} _ {\\text {R e s t o r a t i o n - a w a r e (l e a r n a b l e)}}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 243, + 731, + 785, + 768 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the above equation, the first part $\\nabla_{\\boldsymbol{z}_t}\\log p(\\boldsymbol{z}_t|\\boldsymbol{c}_s)$ aligns with the text-to-image prior inherent in the pretrained LDM. The second term $\\nabla_{\\boldsymbol{z}_t}\\log p(\\boldsymbol{y}|\\boldsymbol{z}_t,\\boldsymbol{c}_r)$ approximates the consistency constraint with degraded image $\\boldsymbol{y}$ , meaning the latent image $\\boldsymbol{z}_t$ should be inferred by the degraded image $\\boldsymbol{y}$ and the degradation", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "SPIRE: Semantic Prompt-Driven Image Restoration", + "bbox": [ + 380, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "details $\\pmb{c}_r$ . This is somewhat similar to the reverse process of $\\pmb{y} = \\mathrm{Deg}(\\pmb{x}_t, \\pmb{c}_r)$ in the pixel space. More derivation is provided in our supplement. Providing degradation information through a restoration text prompt $\\pmb{c}_r$ can largely reduce the uncertainty of estimation $p(\\pmb{y} | \\pmb{z}_t, \\pmb{c}_r)$ , thereby alleviating the ambiguity of the restoration task and leading to better results that are less prone to hallucinations.", + "bbox": [ + 212, + 146, + 787, + 223 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Learning to Control the Restoration.", + "text_level": 1, + "bbox": [ + 214, + 229, + 565, + 244 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Inspired by ControlNet [84], we employ an adaptor to model the restoration-aware branch in Eq. (2). First, the pretrained latent diffusion UNet $\\epsilon_{\\theta}(\\pmb{z}_t,t,\\pmb{c}_s)$ is frozen during training to preserve text-to-image prior $\\nabla_{\\pmb{z}_t}\\log p(\\pmb{z}_t|\\pmb{c}_s)$ . Secondly, an encoder initialized from pretrained UNet is finetuned to learn the restoration condi", + "bbox": [ + 212, + 250, + 431, + 431 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cd4cdd9227e8c76a2fe883bd0d71f902ce64ed16a21d140fb49458a3d8c0dad6.jpg", + "image_caption": [ + "Fig. 3: Degradation ambiguities in real-world problems. By adjusting the restoration prompt, our method can preserve the motion effect that is coupled with Gaussian blur, while fully blind restoration models do not provide this flexibility." + ], + "image_footnote": [], + "bbox": [ + 442, + 263, + 787, + 343 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tion $\\nabla_{z_t}\\log p(\\pmb {y}|\\pmb {z}_t,\\pmb {c}_r)$ , which takes degradation image $\\pmb{y}$ , current noisy latents $\\pmb{z}_{t}$ and the restoration prompt $\\pmb{c}_{r}$ as input. The output of encoder is fusion features $\\pmb{f}_{con}$ for frozen UNet. In this way, the choice of ControlNet implementation follows our decoupling analysis in Eq. (2).", + "bbox": [ + 212, + 431, + 787, + 492 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Conditioning on the Degraded Image $\\pmb{y}$ . We apply downsample layers on $\\pmb{y}$ to make the feature map matching to the shape of the $\\pmb{z}_t$ , denoted as $\\mathcal{E}'(\\pmb{y})$ . $\\mathcal{E}'$ is more compact [84] than VAE encoder $\\mathcal{E}$ , and its trainable parameters allow model to perceive and adapt to the degraded images from our synthetic data pipeline. Then, we concatenate the downsampled feature map $\\mathcal{E}'(\\pmb{y})$ with the latent $\\pmb{z}_t$ , and feed them to the ControlNet encoder.", + "bbox": [ + 212, + 492, + 787, + 582 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Conditioning on the Restoration Prompt $\\pmb{c}_r$ . Following [50], we first use the vision-language CLIP [46] to infer the text sequence embedding $\\pmb{e}_r \\in \\mathbb{R}^{M \\times d}$ for $\\pmb{c}_r$ , where $M$ is the number of tokens, and $d = 768$ is the dimension of the image-text features. Specifically, given a restoration instruction \"deblur with sigma 3.0\", CLIP first tokenizes it into a sequence of tokens (e.g., \"3.0\" is tokenized into [\"3\", \".\", \"0)]) and embeds them into $M$ distinct token embeddings. Then, stacked causal transformer layers [65] are applied on the embedding sequence to contextualize the token embeddings, resulting in $\\pmb{e}_r$ . Hence, tokens such as \"deblur\" may arm with the numeric information such as \"3\", \".\", and \"0\", making it possible for the model to distinguish the semantic difference of strength definition in different restoration tasks. Cross-attention process ensures that the information from $\\pmb{e}_r$ is propagated to the ControlNet feature $f_{con}$ . While [42, 43] finetuned CLIP using contrastive learning, while we found frozen CLIP still works. The reason is that the learnable cross-attention layers can somehow ensemble the observations from frozen CLIP and squeeze useful information from it.", + "bbox": [ + 212, + 583, + 787, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Modulation Fusion Layer. The decoupling of learning two branches (Sec. 3.3), despite offering the benefit of effective separation of prior and conditioning,", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Qi et al.", + "bbox": [ + 271, + 114, + 328, + 128 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1e3f3902d4d30784056e664e6ad875526750512d1255cfc216e8fb51a40a35ab.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodPromptsParameterized Degradation with synthesized crReal-ESRGAN Degradation without cr
Sem ResFID↓LPIPS↓PSNR↑SSIM↑CLIP-I↑CLIP-T↑FID↓LPIPS↓PSNR↑SSIM↑CLIP-I↑CLIP-T↑
SwinIR [30]XX43.220.42324.400.7170.8560.28548.370.44923.450.6990.8420.284
StableSR [67]XX20.550.31321.030.6130.8860.29825.750.36420.420.5810.8640.298
DiffBIR [34]XX17.260.30222.160.6040.9120.29719.170.33021.480.5870.8980.298
ControlNet-SR [84]XX13.650.22223.750.6690.9380.30016.990.26922.950.6280.9240.299
Ours w/o textXX12.700.22123.840.6710.9390.29916.250.26223.150.6360.9290.300
DiffBIR [34] + SDEdit [37]✓X19.360.36219.390.5270.8910.30517.510.37519.150.5210.8870.308
DiffBIR [34] + CLIP [46]✓X18.460.36520.500.5260.8960.30820.310.37420.450.5390.8850.307
ControlNet-SR + CLIP [46]✓X13.000.24123.180.6480.9370.30715.160.28622.450.6100.9260.308
Ours✓✓11.340.21923.610.6650.9430.30614.420.26223.140.6330.9350.308
", + "bbox": [ + 217, + 143, + 782, + 258 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/ea171af9dab562f30e82091fe411d74eaebd406121aa6c4a112c5e17995cf4c4.jpg", + "table_caption": [ + "Table 1: Quantitative results on the MS-COCO dataset (with $c_{s}$ ) using our parameterized degradation (left) and Real-ESRGAN degradation (right). We also denote the prompt choice at test time. 'Sem' stands for semantic prompt; 'Res' stands for restoration prompt. The first group of baselines are tested without prompt. The second group are combined with semantic prompt in zero-shot way." + ], + "table_footnote": [], + "table_body": "
Degradation ProcessParameterizedReal-ESRGAN
p(choose)Restoration PromptDegradation Processp(choose)Restoration Prompt
Gaussian Blur0.5Deblur with {sigma∈[0.2, 3.0]} or DeblurBlur1.0
Downsample0.5Upsample to {resizing factor∈[1.0, 7.0]} or UpsampleResize1.0Remove all degradation
Gaussian Noise0.5Denoise with {sigma∈[0.0, 0.12]} or DenoiseNoise1.0
JPEG0.5Dejpeg with quality {quality factor∈[30, 92]} or DejpegJPEG1.0
", + "bbox": [ + 217, + 344, + 782, + 401 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 2: Our training degradation is randomly sampled in these two pipeline with $50\\%$ each. (1) Images generated by our parameterized pipeline are paired with either a restoration type (e.g., \"Deblur\") or a restoration parameter prompt (e.g., \"Deblur with sigma 0.3.\") (2) In other $50\\%$ iterations, degraded images $\\pmb{y}$ synthesized by Real-ESRGAN are paired with the same restoration prompt $\\pmb{c}_r = \\text{\"Remove all degradation\"}$", + "bbox": [ + 212, + 402, + 785, + 470 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "may cause distribution shift in the learned representations between the frozen backbone feature $\\pmb{f}_{\\mathrm{skip}}$ and ControlNet feature $\\pmb{f}_{\\mathrm{control}}$ . To allow for adaptive alignment of the above features, here we propose a new modulation fusion layer to fuse the multi-scale features from ControlNet to the skip connection features of the frozen backbone: $\\hat{\\pmb{f}}_{\\mathrm{skip}} = (1 + \\gamma)\\pmb{f}_{\\mathrm{skip}} + \\beta$ ; $\\gamma, \\beta = \\mathcal{M}(\\pmb{f}_{\\mathrm{con}})$ , where $\\gamma, \\beta$ are the scaling and bias parameters from $\\mathcal{M}$ a lightweight $1\\times 1$ zero-initialized convolution layer.", + "bbox": [ + 212, + 488, + 787, + 595 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 604, + 375, + 622 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Text-based Training Data and Benchmarks", + "text_level": 1, + "bbox": [ + 215, + 633, + 616, + 648 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Our method opens up an entirely new paradigm of instruction-based image restoration. However, existing image restoration datasets such as DIV2K [2] and Flickr2K [40] do not provide high-quality semantic prompts, and Real-ESRGAN degradation is for blind restoration. To address this, we construct a new setting including training data generation and test benchmarks for evaluation.", + "bbox": [ + 212, + 657, + 784, + 733 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Our parameterized degradation pipeline is based on the Real-ESRGAN [71] which contains a large number of degradation types. Since the original full Real-ESRGAN degradation is difficult to be parameterized and represented as user-friendly natural language (e.g., not every one understands \"anisotropic plateau blur with beta $2.0^{\\prime \\prime}$ ), we choose the 4 most general degradations which are practical for parameterization to support degradation parameter-aware restoration, as shown in Tab. 2. Our parameterized pipeline skips each degradation stage", + "bbox": [ + 212, + 733, + 785, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "SPIRE: Semantic Prompt-Driven Image Restoration", + "bbox": [ + 380, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/66e49157111f2c87df01cd69bcc0e613dd7448addd03ada21cab04e71c31a9f9.jpg", + "image_caption": [ + "Fig. 4: Visual Comparison with other baselines. Our method of integrating both the semantic prompt $\\pmb{c}_{s}$ and the restoration prompt $\\pmb{c}_{r}$ outperforms image-to-image restoration (DiffBIR, Retrained ControlNet-SR) and naive zero-shot combination with semantic prompt. It achieves more sharp and clean results while maintaining consistency with the degraded image." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 782, + 421 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "with a $50\\%$ probability to increase the diversity of restoration prompts and support single task restoration. The descriptions of the selected degradations are appended following the order of the image degradations to synthesize $\\pmb{c}_r$ . In real application, users can control the restoration type and its strength by modifying restoration prompt. As shown in third column of Fig. 3, driven by restoration prompt \"Deblur with sigma $3.0^{\\prime \\prime}$ , our framework removes the Gaussian blur, while preserving motion blur. The representation of the restoration strength is also general enough to decouple different tasks in natural language as shown in Fig. 7. More visual results are presented in our supplement.", + "bbox": [ + 212, + 508, + 787, + 647 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Training data construction. In the training stage, we sample near 100 million text-image pairs from internal data source and use the alternative text label with highest relevance as the semantic prompt $\\pmb{c}_s$ in the framework. We drop out the semantic prompt to empty string $\\varnothing$ by a probability of $10\\%$ to support semantic-agnostic restoration. We mix the Real-ESRGAN degradation pipeline and our parameterized pipeline by $50\\%$ each in training iterations.", + "bbox": [ + 212, + 651, + 787, + 743 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation Setting. We randomly choose 3000 pairs of image $\\pmb{x}$ and semantic prompt $\\pmb{c}_{s}$ from the MS-COCO [33] validation set. Then, we build two test sets on parameterized and the Real-ESRGAN processes separately. At the left half of Tab. 1, images are sent to the parameterized degradation to form the synthesize degradation image $\\pmb{y}$ with restoration prompt $\\pmb{c}_{r}$ . In this setting, we expect \"Ours\" model to fully utilize the $\\pmb{c}_{s}$ describing the semantic and the $\\pmb{c}_{r}$", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Qi et al.", + "bbox": [ + 271, + 114, + 328, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/4b37870e6839052683d07b4ef8fe6dec31d8d89f67cdfe5fd9a3b27f68f26705.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodFID↓LPIPS↓PSNR↑SSIM↑CLIP-I↑
Real-ESRGAN [72]32.370.31222.520.6460.683
DiffBIR [34] (zero-shot)30.710.35422.010.5260.921
StableSR24.440.31121.620.5330.928
Ours w/o text (zero-shot)28.800.35221.680.5490.927
Ours w/o text (finetuned)22.450.32121.380.5320.932
", + "bbox": [ + 218, + 143, + 612, + 224 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 3: Numerical results on the DIV2K testset without any prompt.", + "bbox": [ + 616, + 160, + 764, + 217 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "describing the degradations. On the right half, the same images are fed to the Real-ESRGAN degradation—so the degradation is aligned with open-sourced checkpoints [30, 34, 67]. Because not every model can utilize both the semantic $(c_{s})$ and restoration $(c_{r})$ information, Tab. 1 provides a “Prompts” column denoting what information the model relied on. In summary, our constructed benchmarks based on MS-COCO can serve multiple purposes—from fully-blind, content-prompted, to all-prompted restoration.", + "bbox": [ + 212, + 239, + 787, + 347 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Evaluation Metrics. We use the generative metric FID [22] and perceptual metric LPIPS [86] for quantitative evaluation of image quality. PSNR, and SSIM are reported for reference. We also evaluate similarity scores in the clip embedding space with ground truth image (CLIP-I) and caption (CLIP-T).", + "bbox": [ + 212, + 349, + 787, + 411 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Comparing with baselines", + "text_level": 1, + "bbox": [ + 215, + 417, + 475, + 434 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We compare our method with three categories of baselines:", + "bbox": [ + 215, + 441, + 640, + 458 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Open-sourced image restoration checkpoints, including the regression model SwinIR [30] and the latent diffusion models StableSR [67] and DiffBIR [34].", + "Combining the state-of-the-art method [34] with the zero-shot post-editing [37] (DiffBIR+SDEdit) or zero-shot injection through CLIP [46] (DiffBIR + CLIP).", + "- ControlNet-SR (retrained): We adapt the image-to-image ControlNet model [84] for super-resolution. For consistency, we maintain the same training iterations as our own model. We also evaluate its zero-shot capabilities with CLIP guidance (ControlNet-SR + CLIP)." + ], + "bbox": [ + 215, + 465, + 795, + 590 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Quantitative comparison with baselines is presented in Tab. 1. On the left, we evaluate our full model with the parameterized degradation test set. Since open-sourced baselines are pretrained on Real-ESRGAN degradation, not perfectly aligned with our parameterized degradation, we also evaluate our full model on MS-COCO with Real-ESRGAN degradation by setting our restoration prompt to \"Remove all degradation\". Thanks to our prompts guidance and architecture improvements, our full model achieves best FID, LPIPS, CLIP-Image score, which means better image quality and semantic restoration. \"Ours w/o text\" is the same checkpoint conditioned on only degradation image and has high pixel-level similarity (the second best PSNR and SSIM) with GT. Although SwinIR has highest PSNR, its visual results are blurry. Although combining semantic prompt in zero-shot way can bring marginal improvement in FID and CLIP similarity with caption, it deteriorates the image restoration capacity and results in worse LPIPS and CLIP-Image. In contrast, semantic prompt guidance improves the CLIP image similarity of our full model. Tab. 3 shows the evaluation results of our model on DIV2K test set provided by StableSR. In zero-shot", + "bbox": [ + 212, + 598, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "SPIRE: Semantic Prompt-Driven Image Restoration", + "bbox": [ + 380, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/6a56b5abf87c9b569a850dcd4be82b8553d20c5dc8c202a19b17cd96a09623f0.jpg", + "image_caption": [ + "Fig. 5: Real-world image restorations. Our framework enhances real-world low-quality images following semantic prompts provided by language models or user input." + ], + "image_footnote": [], + "bbox": [ + 220, + 132, + 315, + 255 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/05822db3db402b76fd6021faf22a65a78216020cd316b8037478e56b720c551d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 316, + 132, + 408, + 255 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/6e844126cba331decf626d3ed980c7474cde6a85d75940a1b6a24bdf21dbd48b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 408, + 132, + 501, + 255 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/3be6e31444462b912c1799d410d299cfd82cc11a88d9255a710182a071a9cfc1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 132, + 594, + 255 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/92af50cf89b62bb95f8a749107424572e5ee301e75240dac4e9f8201fcda6fd2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 594, + 132, + 687, + 255 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/3af1353d07f3adb4aa2500e3acd0378d57a4618fc6c1f6c4cc2aae72441eb47e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 687, + 132, + 781, + 255 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/3cbbeef7554bbac8caf8f61b520e12c68da362a9a81df55ddb9b992ec51424ed.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MetricSwinIR [30]Diffbir [34]Ablation of restoration prompt crAblation of semantic prompt cs
Sem c_sXXXXshortunmatchedfull
Res crXX"Deblur 0.6"... 1.0"... 2.0"... 5.0""Remove all degradation"
CLIP-IQA↑0.61900.69830.40740.41590.48310.47970.58180.57500.64600.6610
MUSIQ↑63.6569.6935.8239.8749.9848.8161.6860.2767.5467.82
MANIQA↑0.40350.26190.22990.24870.29700.28210.36940.36370.43550.4474
", + "bbox": [ + 217, + 284, + 784, + 349 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 4: Quantitative results on real-world images. Our aesthetic quality is improved when increasing restoration strength or using a more accurate and detailed prompt.", + "bbox": [ + 215, + 351, + 782, + 378 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "test setting, our model has lower FID and LPIPS than DiffBIR. To fairly compare with StableSR, we finetune our model on their training set. Our finetuned model also outperforms StableSR in FID and CLIP-I.", + "bbox": [ + 212, + 392, + 784, + 436 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Qualitative comparison with baselines is presented in Fig. 4, where the corresponding semantic prompt is provided below each row. Image-to-image baselines such as DiffBir and the retrained ControlNet-SR can easily generate artifacts (e.g., hydrant in the first row) and blurry images (e.g., hands example in third row) that may look close to an \"average\" estimation. Besides, naive combination with semantic prompt in the zero-shot approach results in heavy hallucinations and may fail to preserve the object identity in the input image (e.g., giraffes on the second row). Unlike the existing methods, our full model considers semantic prompt, degradation image and restoration prompt in both training and test stages, which makes its results more aligned with all conditions.", + "bbox": [ + 212, + 441, + 784, + 592 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Real-world restoration", + "text_level": 1, + "bbox": [ + 215, + 617, + 446, + 630 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Besides evaluation on synthetic degradations, we present an analysis on real-world images in Tab. 4 and Fig. 5. We choose RealPhoto60 [78] as our real-world test-set, and use non-reference metrics CLIP-IQA [66], MUSIQ [27] and MANIQA [76] scores to study the influence of the semantic and restoration prompts. First, we fix the semantic prompt to the empty string $\\varnothing$ and conduct an ablation study on the restoration prompt. As the deblurring strength increases from 0.6 to 2.0, the image quality is improved consistently. A blind prompt \"Remove all degradation\" can provide better results than the manually controlled restoration prompts. This showcases that the proposed model is flexible to work with generic restoration prompts, especially when coming up with the perfect prompt is challenging. We also fix the blind restoration prompt and change the semantic prompts. Following [78], we feed degraded images to LLaVA [35] and generate corresponding semantic prompts. Our experiments", + "bbox": [ + 212, + 643, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Qi et al.", + "bbox": [ + 271, + 114, + 328, + 127 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/daa0906e8b3bd0788da4ba27554dc6e6bac528f5d521d1c2e6758519bfe02e20.jpg", + "image_caption": [ + "Fig. 6: Test-time semantic prompting. Our framework restores degraded images guided by flexible semantic prompts, while unrelated background elements and global tones remain aligned with the degraded input conditioning." + ], + "image_footnote": [], + "bbox": [ + 220, + 143, + 782, + 304 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/7de3d5beec6f58dec1bbbf7714b1fe10b6a26feadada030da3323936d31e7f47.jpg", + "image_caption": [ + "Fig.7: Prompt space walking visualization for the restoration prompt. Given the same degraded input (upper left) and empty semantic prompt $\\varnothing$ , our method can decouple the restoration direction and strength via only prompting the quantitative number in natural language. An interesting finding is that our model learns a continuous range of restoration strength from discrete language tokens." + ], + "image_footnote": [], + "bbox": [ + 217, + 359, + 586, + 606 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "show that detailed full descriptions (about 60 words) of the degraded image from LLaVA can be more effective to improve the visual results than short prompts that have less than 20 words. Unmatched prompts simulate the case when a user provides wrong prompts by randomly re-sampling unmatched long prompts from the data set. An interesting observation is that unmatched long prompts also improve the aesthetic quality of images against a short prompt or a blind restoration. Nonetheless, it can lead to more \"hallucination\" when the semantic input prompt is unmatched with the real ground truth. In contrast to concurrent work [62] that leverages implicit clip embedding, our text-based formulation provides better control of hallucinations by adjusting the LLM-synthesized or manually-designed prompt at test-time. Note that many studies [6,18] find current image restoration metrics are biased towards specific datasets and not fully aligned with human's visual perception. The quantitative results of previous works [34,71] are provided for reference. More results and discussions are provided in the supplements.", + "bbox": [ + 212, + 628, + 787, + 857 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "SPIRE: Semantic Prompt-Driven Image Restoration", + "bbox": [ + 382, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/738f300f4302f7b4c4e52c23a9e229763b0b817b25f1daef50abcaafb696346b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodSem csRes crFID↓LPIPS↓PSNR↑SSIM↑CLIP-I↑CLIP-T↑
ControlNet-SRXX13.650.22223.750.6690.9380.300
ControlNet-SR✓✓12.140.22323.500.6600.9400.306
Ours✓wo param11.580.22323.610.6650.9420.305
Ours✓✓11.340.21923.610.6650.9430.305
", + "bbox": [ + 246, + 143, + 754, + 228 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Table 5: Ablation of architecture and degradation strength in ${c}_{r}$", + "bbox": [ + 281, + 228, + 720, + 241 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "4.4 Prompting the SPIRE", + "text_level": 1, + "bbox": [ + 215, + 246, + 447, + 262 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Semantic Prompting. Our model can be driven by user-defined semantic prompts. As shown in Fig. 6, the blind restoration with empty string generates ambiguous and unrealistic identity. In contrast, our framework reconstructs sharp and realistic results. The object identity accurately follows user's semantic prompt, while the global layout and color tone remain consistent with input.", + "bbox": [ + 212, + 267, + 785, + 343 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Restoration Prompting. Users also have the freedom to adjust the degradation types and strengths in the restoration prompts. As shown in Fig. 7, the denoising and deblurring restoration are well decoupled. Using the task-specific prompts, our method can get a clean but blur (second row), or sharp but noisy (third row) result. In addition, our model learns the continuous space of restoration as the result becomes cleaner progressively if we modify the \"Denoise with sigma 0.06\" to \"0.24\". Beyond purely relying on numerical description, our model \"understand\" the difference of strength definition for tasks, such as the range difference of denoise sigma and deblur sigma, which makes it promising to be a universal representation for restoration strength control.", + "bbox": [ + 212, + 343, + 787, + 494 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "4.5 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 512, + 388, + 527 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We conduct several ablations in Tab. 5. Providing both prompts $c_s$ , $c_r$ to \"ControlNet-SR\" baseline improves the FID by 1.5. Adding the modulation fusion layer (\"Ours\") further improves the generative quality (0.8 for FID) and pixel-level similarity (0.11 dB). Embedding degradation parameters in the restoration prompt, not only enables restoration strength prompting (Fig. 7), but also improves image quality (FID and LPIPS). More details are in the supplement.", + "bbox": [ + 212, + 534, + 785, + 626 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 645, + 359, + 662 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We have presented a unified text-driven framework for instructed image restoration using semantic and restoration prompts. We design our model in a decoupled way to better preserve the semantic text-to-image generative prior while efficiently learning to control both the restoration direction and its strength. To the best of our knowledge, this is the first framework to support both semantic and parameter-embedded restoration instructions simultaneously, allowing users to flexibly prompt-tune the results up to their expectations. Our extensive experiments have shown that our method significantly outperforms prior works in terms of both quantitative and qualitative results. Our model and evaluation benchmark have established a new paradigm of instruction-based image restoration, paving the way for further multi-modality generative imaging applications.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Qi et al.", + "bbox": [ + 271, + 114, + 330, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Abuolaim, A., Delbracio, M., Kelly, D., Brown, M.S., Milanfar, P.: Learning to reduce defocus blur by realistically modeling dual-pixel data. In: ICCV. pp. 2289-2298 (2021)", + "2. Agustsson, E., Timofte, R.: NTIRE 2017 challenge on single image superresolution: Dataset and study. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition Workshops, CVPR Workshops 2017, Honolulu, HI, USA, July 21-26, 2017. pp. 1122-1131. IEEE Computer Society (2017). https://doi.org/10.1109/CVPRW.2017.150, https://doi.org/10.1109/CVPRW.2017.150", + "3. Avrahami, O., Fried, O., Lischinski, D.: Blended latent diffusion. ACM Transactions on Graphics (2023)", + "4. Avrahami, O., Lischinski, D., Fried, O.: Blended diffusion for text-driven editing of natural images. In: CVPR. pp. 18208-18218 (2022)", + "5. Bai, Y., Wang, C., Xie, S., Dong, C., Yuan, C., Wang, Z.: Textir: A simple framework for text-based editable image restoration. arXiv preprint arXiv:2302.14736 (2023)", + "6. Blau, Y., Michaeli, T.: The perception-distortion tradeoff. In: CVPR. pp. 6228-6237 (2018)", + "7. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: CVPR (2023)", + "8. Chen, X., Wang, X., Changpinyo, S., Piergiovanni, A., Padlewski, P., Salz, D., Goodman, S.A., Grycner, A., Mustafa, B., Beyer, L., Kolesnikov, A., Puigcerver, J., Ding, N., Rong, K., Akbari, H., Mishra, G., Xue, L., Thapliyal, A., Bradbury, J., Kuo, W., Seyedhosseini, M., Jia, C., Ayan, B.K., Riquelme, C., Steiner, A., Angelova, A., Zhai, X., Houlsby, N., Soricut, R.: Pali: A jointly-scaled multilingual language-image model. In: ICLR (2023), https://arxiv.org/abs/2209.06794", + "9. Chen, Z., Zhang, Y., Gu, J., Yuan, X., Kong, L., Chen, G., Yang, X.: Image superresolution with text prompt diffusion. arXiv preprint arXiv:2303.06373 (2023)", + "0. Couairon, G., Verbeek, J., Schwenk, H., Cord, M.: Diffedit: Diffusion-based semantic image editing with mask guidance. In: ICLR (2022)", + "1. Delbracio, M., Milanfar, P.: Inversion by direct iteration: An alternative to denoising diffusion for image restoration. Transactions on Machine Learning Research (2023), https://openreview.net/forum?id=VmyFF5lL3F, featured Certification", + "2. Delbracio, M., Talebei, H., Milanfar, P.: Projected distribution loss for image enhancement. In: 2021 IEEE International Conference on Computational Photography (ICCP). pp. 1-12. IEEE (2021)", + "3. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Neural Information Processing Systems (2021)", + "4. Dong, C., Loy, C.C., He, K., Tang, X.: Image super-resolution using deep convolutional networks. IEEE Trans. Pattern Anal. Mach. Intell. 38(2), 295-307 (2016)", + "5. Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A.H., Chechik, G., Cohen-Or, D.: An image is worth one word: Personalizing text-to-image generation using textual inversion. In: ICLR (2023)", + "6. Galteri, L., Seidenari, L., Bertini, M., Bimbo, A.: Deep generative adversarial compression artifact removal. In: ICCV (2017)", + "7. Geng, Z., Yang, B., Hang, T., Li, C., Gu, S., Zhang, T., Bao, J., Zhang, Z., Hu, H., Chen, D., Guo, B.: Instructdiffusion: A generalist modeling interface for vision tasks. CoRR abs/2309.03895 (2023). https://doi.org/10.48550/arXiv.2309.03895, https://doi.org/10.48550/arXiv.2309.03895" + ], + "bbox": [ + 225, + 176, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "SPIRE: Semantic Prompt-Driven Image Restoration", + "bbox": [ + 382, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Gu, J., Cai, H., Dong, C., Ren, J.S., Timofte, R., Gong, Y., Lao, S., Shi, S., Wang, J., Yang, S., Wu, T., Xia, W., Yang, Y., Cao, M., Heng, C., Fu, L., Zhang, R., Zhang, Y., Wang, H., Song, H., Wang, J., Fan, H., Hou, X., Sun, M., Li, M., Zhao, K., Yuan, K., Kong, Z., Wu, M., Zheng, C., Conde, M.V., Burchi, M., Feng, L., Zhang, T., Li, Y., Xu, J., Wang, H., Liao, Y., Li, J., Xu, K., Sun, T., Xiong, Y., Keshari, A., Komal, Thakur, S., Jakhetiya, V., Subudhi, B.N., Yang, H.H., Chang, H.E., Huang, Z.K., Chen, W.T., Kuo, S.Y., Dutta, S., Das, S.D., Shah, N.A., Tiwari, A.K.: Ntire 2022 challenge on perceptual image quality assessment. In: CVPRW. pp. 951-967 (June 2022)", + "19. Gu, X., Lin, T.Y., Kuo, W., Cui, Y.: Open-vocabulary object detection via vision and language knowledge distillation. In: ICLR (2021)", + "20. Han, L., Li, Y., Zhang, H., Milanfar, P., Metaxas, D., Yang, F.: Svdiff: Compact parameter space for diffusion fine-tuning. In: ICCV. pp. 7323-7334 (October 2023)", + "21. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-or, D.: Prompt-to-prompt image editing with cross-attention control. In: ICLR (2023), https://openreview.net/forum?id=CDixzkzeyb", + "22. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: NeurIPS (2017)", + "23. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. NeurIPS 33, 6840-6851 (2020)", + "24. Ho, J., Salimans, T.: Classifier-free diffusion guidance. In: NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications (2021), https://openreview.net/forum?id=qw8AKxfYbI", + "25. Hu, E.J., yelong shen, Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-rank adaptation of large language models. In: ICLR (2022), https://openreview.net/forum?id=nZeVKeeFYf9", + "26. Jiang, Y., Zhang, Z., Xue, T., Gu, J.: Autodir: Automatic all-in-one image restoration with latent diffusion. arXiv preprint arXiv:2310.10123 (2023)", + "27. Ke, J., Wang, Q., Wang, Y., Milanfar, P., Yang, F.: Musiq: Multi-scale image quality transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5148-5157 (2021)", + "28. Ke, J., Ye, K., Yu, J., Wu, Y., Milanfar, P., Yang, F.: Vila: Learning image aesthetics from user comments with vision-language pretraining. In: CVPR. pp. 10041-10051 (2023)", + "29. Kumari, N., Zhang, B., Zhang, R., Shechtman, E., Zhu, J.Y.: Multi-concept customization of text-to-image diffusion. In: CVPR. pp. 1931-1941 (2023)", + "30. Liang, J., Cao, J., Sun, G., Zhang, K., Gool, L.V., Timofte, R.: Swinir: Image restoration using swin transformer. In: Proceedings of ICCV Workshops (2021)", + "31. Liang, Z., Li, C., Zhou, S., Feng, R., Loy, C.C.: Iterative prompt learning for unsupervised backlit image enhancement. In: ICCV. pp. 8094-8103 (2023)", + "32. Lim, B., Son, S., Kim, H., Nah, S., Lee, K.M.: Enhanced deep residual networks for single image super-resolution. In: Proceedings of CVPR Workshops (2017)", + "33. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, L.: Microsoft coco: Common objects in context. In: ECCV. ECCV (September 2014), https://www.microsoft.com/en-us/research/publication/microsoft-coco-common-objects-in-context/", + "34. Lin, X., He, J., Chen, Z., Lyu, Z., Fei, B., Dai, B., Ouyang, W., Qiao, Y., Dong, C.: Diffbir: Towards blind image restoration with generative diffusion prior. arXiv preprint arXiv:2308.15070 (2023)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Qi et al.", + "bbox": [ + 271, + 114, + 328, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "35. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: NeurIPS (2023)", + "36. Luo, Z., Gustafsson, F.K., Zhao, Z., Sjolund, J., Schon, T.B.: Controlling vision-language models for multi-task image restoration. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=t3vnnLeajU", + "37. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: SDEdit: Guided image synthesis and editing with stochastic differential equations. In: ICLR (2022)", + "38. Mokady, R., Hertz, A., Aberman, K., Pritch, Y., Cohen-Or, D.: Null-text inversion for editing real images using guided diffusion models. In: CVPR. pp. 6038-6047 (June 2023)", + "39. Mou, C., Wang, X., Xie, L., Zhang, J., Qi, Z., Shan, Y., Qie, X.: T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453 (2023)", + "40. Nah, S., Timofte, R., Baik, S., Hong, S., Moon, G., Son, S., Lee, K.M., Wang, X., Chan, K.C.K., Yu, K., Dong, C., Loy, C.C., Fan, Y., Yu, J., Liu, D., Huang, T.S., Sim, H., Kim, M., Park, D., Kim, J., Chun, S.Y., Haris, M., Shakhnarovich, G., Ukita, N., Zamir, S.W., Arora, A., Khan, S.H., Khan, F.S., Shao, L., Gupta, R.K., Chudasama, V.M., Patel, H., Upla, K.P., Fan, H., Li, G., Zhang, Y., Li, X., Zhang, W., He, Q., Purohit, K., Rajagopalan, A.N., Kim, J., Tofighi, M., Guo, T., Monga, V.: NTIRE 2019 challenge on video deblurring: Methods and results. In: IEEE Conference on Computer Vision and Pattern Recognition Workshops, CVPR Workshops 2019, Long Beach, CA, USA, June 16-20, 2019. pp. 1974-1984. Computer Vision Foundation / IEEE (2019). https://doi.org/10.1109/CVPRW.2019.00249, http://openaccess.thecvf.com/content_CVPRW_2019/html/NTIRE/Nah_NTIRE_2019_Challenge_on_Video_Deblurring_Methods_and_Results_CVPRW_2019_paper.html", + "41. OpenAI: Gpt-4 technical report (2023)", + "42. Paiss, R., Chefer, H., Wolf, L.: No token left behind: Explainability-aided image classification and generation. In: ECCV (2022)", + "43. Paiss, R., Ephrat, A., Tov, O., Zada, S., Mosseri, I., Irani, M., Dekel, T.: Teaching clip to count to ten. In: ICCV (2023)", + "44. Parmar, G., Kumar Singh, K., Zhang, R., Li, Y., Lu, J., Zhu, J.Y.: Zero-shot image-to-image translation. In: ACM SIGGRAPH 2023 Conference Proceedings. pp. 1-11 (2023)", + "45. Prakash, M., Delbracio, M., Milanfar, P., Jug, F.: Interpretable unsupervised diversity denoising and artefact removal. In: ICLR (2022), https://openreview.net/forum?id=DfMqlBOPXjM", + "46. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)", + "47. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners. OpenAI blog 1(8), 9 (2019)", + "48. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022)", + "49. Ren, M., Delbracio, M., Talebi, H., Gerig, G., Milanfar, P.: Multiscale structure guided diffusion for image deblurring. In: ICCV. pp. 10721-10733 (2023)", + "50. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10684-10695 (2022)" + ], + "bbox": [ + 215, + 146, + 784, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "SPIRE: Semantic Prompt-Driven Image Restoration", + "bbox": [ + 382, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "51. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015)", + "52. Sahara, C., Chan, W., Chang, H., Lee, C., Ho, J., Salimans, T., Fleet, D., Norouzi, M.: Palette: Image-to-image diffusion models. In: ACM SIGGRAPH 2022 Conference Proceedings. pp. 1-10 (2022)", + "53. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS (2022)", + "54. Sahara, C., Ho, J., Chan, W., Salimans, T., Fleet, D.J., Norouzi, M.: Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(4), 4713-4726 (2022)", + "55. Sahara, C., Ho, J., Chan, W., Salimans, T., Fleet, D.J., Norouzi, M.: Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(4), 4713-4726 (2022)", + "56. Salimans, T., Ho, J.: Progressive distillation for fast sampling of diffusion models. In: ICLR. OpenReview.net (2022), https://openreview.net/forum?id=TIdIXIpzhoI", + "57. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)", + "58. Song, Y., Ermon, S.: Generative modeling by estimating gradients of the data distribution. In: NeurIPS. pp. 11895-11907 (2019)", + "59. Song, Y., Shen, L., Xing, L., Ermon, S.: Solving inverse problems in medical imaging with score-based generative models. In: ICLR. OpenReview.net (2022)", + "60. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. In: ICLR (2021), https://openreview.net/forum?id=PxTIG12RRHS", + "61. Su, S., Delbracio, M., Wang, J., Sapiro, G., Heidrich, W., Wang, O.: Deep video deblurring for hand-held cameras. In: CVPR. pp. 1279-1288 (2017)", + "62. Sun, H., Li, W., Liu, J., Chen, H., Pei, R., Zou, X., Yan, Y., Yang, Y.: Coser: Bridging image and language for cognitive super-resolution. arXiv preprint arXiv:2311.16512 (2023)", + "63. Tu, Z., Talebi, H., Zhang, H., Yang, F., Milanfar, P., Bovik, A.C., Li, Y.: MAXIM: multi-axis MLP for image processing. In: CVPR (2022)", + "64. Tumanyan, N., Geyer, M., Bagon, S., Dekel, T.: Plug-and-play diffusion features for text-driven image-to-image translation. In: CVPR. pp. 1921-1930 (2023)", + "65. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. NeurIPS 30 (2017)", + "66. Wang, J., Chan, K.C., Loy, C.C.: Exploring clip for assessing the look and feel of images. In: AAAI (2023)", + "67. Wang, J., Yue, Z., Zhou, S., Chan, K.C., Loy, C.C.: Exploiting diffusion prior for real-world image super-resolution. arXiv preprint arXiv:2305.07015 (2023)", + "68. Wang, L., Wang, Y., Lin, Z., Yang, J., An, W., Guo, Y.: Learning a single network for scale-arbitrary super-resolution. In: ICCV. pp. 4801-4810 (2021)", + "69. Wang, X., Chen, X., Ni, B., Wang, H., Tong, Z., Liu, Y.: Deep arbitrary-scale image super-resolution via scale-equivariance pursuit. In: CVPR (2023)", + "70. Wang, X., Li, Y., Zhang, H., Shan, Y.: Towards real-world blind face restoration with generative facial prior. In: CVPR (2021)", + "71. Wang, X., Xie, L., Dong, C., Shan, Y.: Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In: ICCV. pp. 1905-1914 (2021)" + ], + "bbox": [ + 215, + 147, + 785, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Qi et al.", + "bbox": [ + 271, + 114, + 328, + 128 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "72. Wang, X., Yu, K., Wu, S., Gu, J., Liu, Y., Dong, C., Qiao, Y., Loy, C.C.: ESRGAN: enhanced super-resolution generative adversarial networks. In: Proceedings of ECCV Workshops (2018)", + "73. Wang, Z., Cun, X., Bao, J., Zhou, W., Liu, J., Li, H.: Uformer: A general u-shaped transformer for image restoration. In: CVPR (2022)", + "74. Whang, J., Delbracio, M., Talebi, H., Saharia, C., Dimakis, A.G., Milanfar, P.: Deblurring via stochastic refinement. In: CVPR. pp. 16293-16303 (2022)", + "75. Wu, R., Yang, T., Sun, L., Zhang, Z., Li, S., Zhang, L.: Seesr: Towards semantics-aware real-world image super-resolution. In: CVPR (2024)", + "76. Yang, S., Wu, T., Shi, S., Lao, S., Gong, Y., Cao, M., Wang, J., Yang, Y.: Maniaq: Multi-dimension attention network for no-reference image quality assessment. In: CVPR. pp. 1191-1200 (2022)", + "77. Yang, T., Ren, P., Xie, X., Zhang, L.: Gan prior embedded network for blind face restoration in the wild. In: CVPR (2021)", + "78. Yu, F., Gu, J., Li, Z., Hu, J., Kong, X., Wang, X., He, J., Qiao, Y., Dong, C.: Scaling up to excellence: Practicing model scaling for photo-realistic image restoration in the wild. In: CVPR (2024)", + "79. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Multi-stage progressive image restoration. In: CVPR (2021)", + "80. Zhang, K., Liang, J., Van Gool, L., Timofte, R.: Designing a practical degradation model for deep blind image super-resolution. In: ICCV. pp. 4791-4800 (2021)", + "81. Zhang, K., Mo, L., Chen, W., Sun, H., Su, Y.: Magicbrush: A manually annotated dataset for instruction-guided image editing. Advances in Neural Information Processing Systems 36 (2024)", + "82. Zhang, K., Zuo, W., Chen, Y., Meng, D., Zhang, L.: Beyond a Gaussian denoiser: Residual learning of deep CNN for image denoising. IEEE Transactions on Image Processing 26(7), 3142-3155 (2017)", + "83. Zhang, K., Zuo, W., Zhang, L.: Ffdnet: Toward a fast and flexible solution for CNN based image denoising. IEEE Transactions on Image Processing (2018)", + "84. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: ICCV (2023)", + "85. Zhang, R., Fang, R., Gao, P., Zhang, W., Li, K., Dai, J., Qiao, Y., Li, H.: Tip-adapter: Training-free clip-adapter for better vision-language modeling. In: ECCV (2022)", + "86. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)" + ], + "bbox": [ + 215, + 146, + 785, + 646 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "SPIRE: Semantic Prompt-Driven Image Restoration", + "bbox": [ + 382, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/SPIRE_ Semantic Prompt-Driven Image Restoration/0f07d6c7-f604-4c26-b13e-8eef2743217e_model.json b/2024/SPIRE_ Semantic Prompt-Driven Image Restoration/0f07d6c7-f604-4c26-b13e-8eef2743217e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b12b6f174bc08eeb7bc379f9a38c13f52a8a1262 --- /dev/null +++ b/2024/SPIRE_ Semantic Prompt-Driven Image Restoration/0f07d6c7-f604-4c26-b13e-8eef2743217e_model.json @@ -0,0 +1,3087 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.273, + 0.141, + 0.733, + 0.185 + ], + "angle": 0, + "content": "SPIRE: Semantic Prompt-Driven Image Restoration" + }, + { + "type": "text", + "bbox": [ + 0.236, + 0.213, + 0.768, + 0.246 + ], + "angle": 0, + "content": "Chenyang Qi\\(^{1,2*}\\), Zhengzhong Tu\\(^{1}\\), Keren Ye\\(^{1}\\), Maurizio Delbracio\\(^{1}\\), Peyman Milanfar\\(^{1}\\), Qifeng Chen\\(^{2}\\), and Hossein Talebi\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.437, + 0.255, + 0.565, + 0.283 + ], + "angle": 0, + "content": "1 Google Research 2 HKUST" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.302, + 0.313, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.253, + 0.378, + 0.274, + 0.387 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.314, + 0.303, + 0.407, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.319, + 0.377, + 0.402, + 0.385 + ], + "angle": 0, + "content": "\"Remove all degradation\"" + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.303, + 0.499, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.413, + 0.377, + 0.495, + 0.385 + ], + "angle": 0, + "content": "\"Remove all degradation\"" + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.303, + 0.592, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.377, + 0.581, + 0.384 + ], + "angle": 0, + "content": "\"Upsample...denoise..." + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.303, + 0.686, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.603, + 0.377, + 0.676, + 0.384 + ], + "angle": 0, + "content": "'Upsample...denoise...'" + }, + { + "type": "image", + "bbox": [ + 0.687, + 0.303, + 0.78, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.626, + 0.377, + 0.676, + 0.384 + ], + "angle": 0, + "content": "ample..., denoise..." + }, + { + "type": "image_caption", + "bbox": [ + 0.709, + 0.378, + 0.756, + 0.387 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.39, + 0.312, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.253, + 0.466, + 0.274, + 0.474 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.314, + 0.39, + 0.407, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.319, + 0.465, + 0.402, + 0.472 + ], + "angle": 0, + "content": "\"Remove all degradation\"" + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.39, + 0.499, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.413, + 0.465, + 0.496, + 0.472 + ], + "angle": 0, + "content": "\"Remove all degradation\"" + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.39, + 0.592, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.511, + 0.465, + 0.58, + 0.472 + ], + "angle": 0, + "content": "\"Deblur...denoise...\"" + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.39, + 0.686, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.606, + 0.465, + 0.674, + 0.472 + ], + "angle": 0, + "content": "\"Deblur...denoise...\"" + }, + { + "type": "image", + "bbox": [ + 0.687, + 0.39, + 0.78, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.71, + 0.466, + 0.757, + 0.474 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.478, + 0.312, + 0.551 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.253, + 0.553, + 0.274, + 0.561 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.314, + 0.478, + 0.407, + 0.551 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.354, + 0.553, + 0.364, + 0.559 + ], + "angle": 0, + "content": "\\( \\because {a}_{4} = {2}^{3} \\cdot {2}^{2} \\)" + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.478, + 0.499, + 0.551 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.433, + 0.553, + 0.475, + 0.56 + ], + "angle": 0, + "content": "\"DANGER...\"" + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.478, + 0.593, + 0.551 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.531, + 0.553, + 0.562, + 0.56 + ], + "angle": 0, + "content": "\"EXIT...\"" + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.478, + 0.686, + 0.551 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.629, + 0.553, + 0.651, + 0.56 + ], + "angle": 0, + "content": "\"GO...\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.718, + 0.553, + 0.749, + 0.56 + ], + "angle": 0, + "content": "\"STOP..." + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.565, + 0.788, + 0.69 + ], + "angle": 0, + "content": "Fig. 1: We present SPIRE: Semantic Prompt-Driven Image Restoration, a text-based foundation model for all-in-one, instructed image restoration. SPIRE allows users to flexibly leverage either semantic-level content prompt, or quantitative degradation-aware restoration prompt, or both, to obtain their desired enhancement results based on personal preferences. In other words, SPIRE can be easily prompted to conduct blind restoration, semantic restoration, or task-specific granular treatment. Our framework also enables a new paradigm of instruction-based image restoration, providing a reliable evaluation benchmark to facilitate vision-language models for low-level computational photography applications." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.715, + 0.741, + 0.813 + ], + "angle": 0, + "content": "Abstract. Text-driven diffusion models have become increasingly popular for various image editing tasks, including inpainting, stylization, and object replacement. However, it still remains an open research problem to adopt this language-vision paradigm for more fine-level image processing tasks, such as denoising, super-resolution, deblurring, and compression artifact removal. In this paper, we develop SPIRE, a Semantic and restoration Prompt-driven Image Restoration framework that leverages" + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.826, + 0.651, + 0.84 + ], + "angle": 0, + "content": "*This work was done during an internship at Google Research." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.33, + 0.129 + ], + "angle": 0, + "content": "Qi et al." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.148, + 0.744, + 0.328 + ], + "angle": 0, + "content": "natural language as a user-friendly interface to control the image restoration process. We consider the capacity of prompt information in two dimensions. First, we use content-related prompts to enhance the semantic alignment, effectively alleviating identity ambiguity in the restoration outcomes. Second, our approach is the first framework that supports fine-level instruction through language-based quantitative specification of the restoration strength, without the need for explicit task-specific design. In addition, we introduce a novel fusion mechanism that augments the existing ControlNet architecture by learning to rescale the generative prior, thereby achieving better restoration fidelity. Our extensive experiments demonstrate the superior restoration performance of SPIRE compared to the state of the arts, alongside offering the flexibility of text-based control over the restoration effects." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.355, + 0.377, + 0.371 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.387, + 0.788, + 0.583 + ], + "angle": 0, + "content": "Image restoration or enhancement aims to recover high-quality, pixel-level details from a degraded image, while preserving as much the original semantic information as possible. Although neural network models [11, 12, 30, 63, 67, 71, 73, 74] have marked significant progress, it still remains challenging to design an effective task-conditioning mechanism instead of building multiple individual models for each task (such as denoise, deblur, compression artifact removal) in practical scenarios. The advancement of text-driven diffusion models [48, 50, 53] has unveiled the potential of natural language as a universal input condition for a broad range of image processing challenges, which improves interactivity with users and reduces the cost of task-specific fine-tuning. However, the existing applications of natural language in stylization [38, 44], and inpainting [3, 4, 10, 21, 44, 64] predominantly focus on high-level semantic editing, whereas the uniqueness and challenges for low-level image processing have been less explored." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.585, + 0.788, + 0.78 + ], + "angle": 0, + "content": "Natural language text prompts in image restoration can play two crucial roles: alleviating semantic ambiguities and resolving degradation type ambiguities. Firstly, the same degraded image can correspond to different visual objects, leading to ambiguities in content interpretation (e.g., discerning whether the blurred animal in Fig. 1 is a horse or a zebra). Typically, unconditional image-to-image restoration leads to a random or average estimation of the identities, resulting in neither fish nor fowl. Secondly, certain photographic effects that are deliberately introduced for aesthetic purposes, such as bokeh with a soft out-of-focus background, can be misconstrued as blur distortions by many existing deblurring models, leading to unwanted artifacts or severe hallucinations in the outputs. Although blind restoration methods can produce clean images by either leveraging frozen generative priors [67,70,77], or using end-to-end regression [71], they do not consider the aforementioned ambiguities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "In this paper, we introduce SPIRE—a Semantic Prompt-Driven Image REstoration framework that provides a user-friendly interface to fully control both the restored image semantics and the enhancement granularity using natural language instructions. Traditional blind restoration without prompts (Fig" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.382, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SPIRE: Semantic Prompt-Driven Image Restoration" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.344 + ], + "angle": 0, + "content": "ure 1) has limited flexibility in solving semantic and degradation ambiguity, thus tends to generate average blurry images. Leveraging the semantic and quantitative restoration prompts, we show that the visual quality can be significantly improved. Moreover, our framework can also be used as blind restoration with null prompts. In addition to higher perceptual quality, our framework provides the flexibility for users to generate more than one plausible result. Specifically, in imaging scenarios with large levels of degradation there are multiple plausible solutions. The proposed interactive framework has the capability to personalize the output by leveraging input prompts according to user preference. In concurrent studies, the focus has typically been isolated to one of three areas: either blind restoration [34,67], only semantic prompting [78], or only discrete restoration types [9,36]. To the best of our knowledge, this is the first unified model to support the following three distinct features simultaneously:" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.357, + 0.785, + 0.401 + ], + "angle": 0, + "content": "1. Blind Restoration: When instructed with a general restoration prompt \"remove all degradations\" and empty semantic prompt \"\", SPIRE operates as a blind restoration model (\"Ours w/o text\" in Tab. 1)." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.403, + 0.785, + 0.447 + ], + "angle": 0, + "content": "2. Semantic Restoration: When provided with a text description of the (desired) visual content, SPIRE concentrates on restoring the specified identity of the uncertain or ambiguous objects in the degraded image." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.45, + 0.787, + 0.54 + ], + "angle": 0, + "content": "3. Quantitative Task-Specific Restoration: Receiving specific restoration type hints (e.g., \"deblur...\", \"deblur... denoise...\"), SPIRE transforms into a task-specific model (e.g., deblur, denoise, or both). Moreover, it understands the numeric nuances of different degradation parameters in language prompts so that the users can control the restoration strength (e.g., \"deblur with sigma 3.0\" in Fig. 7)." + }, + { + "type": "list", + "bbox": [ + 0.223, + 0.357, + 0.787, + 0.54 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.788, + 0.794 + ], + "angle": 0, + "content": "To train SPIRE, we first build a synthetic data generation pipeline on a large-scale text-image dataset [8] upon the second-order degradation process proposed in Real-ESRGAN [71]. Additionally, we embed the degradation parameters into the restoration prompts to encode finer-grained degradation information (e.g., \"deblur with sigma 3.5\"). Our approach provides richer degradation-specific information compared to contemporary works [26,36] which only employ the degradation types (e.g., \"gaussian blur\"). We then finetune a ControlNet adaptor [84]—which learns a parameter-efficient, parallel branch on top of the latent diffusion models (LDMs) [50]—on a mixture of diverse restoration tasks (see Fig. 1). The semantic text prompts are processed through the LDM backbone, aligning with LDM's text-to-image pretraining. The degradation text prompts and image-conditioning are implemented in the ControlNet branch, as these aspects are not inherently addressed by the vanilla LDM. We further improve the ControlNet by introducing a new modulation connection that adaptively fuses the degradation condition with the generative prior with only a few extra trainable parameters yet showing impressive performance gain." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.803, + 0.84 + ], + "angle": 0, + "content": "Our extensive experiments demonstrate remarkable restoration quality achieved by our proposed SPIRE method. Moreover, SPIRE offers additional freedom for both content and restoration prompt engineering. For example, the semantic" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.33, + 0.129 + ], + "angle": 0, + "content": "Qi et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "prompt \"a very large giraffe eating leaves\" helps to resolve the semantic ambiguity in Fig. 2; the degradation prompt \"deblur with sigma 3.0\" reduces the gaussian blur while maintaining the intentional motion blur in Fig. 3. We also find out that our model learns a continuous latent space of the restoration strength in Fig. 7, even though we do not explicitly fine-tune the CLIP embedding. Our contribution can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.246, + 0.784, + 0.303 + ], + "angle": 0, + "content": "- We introduce the first unified text-driven image restoration model that supports both semantic prompts and restoration instructions. Our experiments demonstrate that incorporating semantic prompts and restoration instructions significantly enhances the restoration quality." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.306, + 0.784, + 0.333 + ], + "angle": 0, + "content": "- Our proposed paradigm empowers users to fully control the semantic outcome of the restored image using different semantic prompts during test time." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.336, + 0.784, + 0.362 + ], + "angle": 0, + "content": "- Our proposed approach provides a mechanism for users to adjust the category and strength of the restoration effect based on their subjective preferences." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.366, + 0.784, + 0.393 + ], + "angle": 0, + "content": "- We demonstrate that text can serve as universal guidance control for low-level image restoration, eliminating the need for task-specific model design." + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.246, + 0.784, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.417, + 0.388, + 0.433 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.788, + 0.553 + ], + "angle": 0, + "content": "First, we review the literature on image restoration. Our proposed framework aims to address some of the limitations of the existing restoration methods. Next, we compare multiple text-guided diffusion editing methods. Our method is designed to deal with two different text ambiguities concurrently, which is an unprecedented challenge for existing text-driven diffusion methods. Finally, we discuss the parameter-efficient fine-tuning of diffusion models, which further motivates our proposed modulation connection." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.554, + 0.788, + 0.689 + ], + "angle": 0, + "content": "Image Restoration is the task of recovering a high-resolution, clean image from a degraded input. Pioneering works in the fields of super-resolution [14,32], motion and defocus deblurring [1,61,79], denoising [82], and JPEG and artifact removal [16, 45] primarily utilize deep neural network architectures to address specific tasks. Later, Transformer-based [30,63,73] and adversarial-based [71] formulations were explored, demonstrating state-of-the-art performance with unified model architecture. Recently there has been a focus on exploiting iterative restorations, such as the ones from diffusion models, to generate images with much higher perceptual quality [11,49,52,54,74,75,78]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.788, + 0.795 + ], + "angle": 0, + "content": "There have been some recent works attempting to control the enhancement strength of a single model, such as the scaling factor condition in arbitrary super-resolution [68, 69] and noise-level map for image denoising [83]. However, these methods require hand-crafting dedicated architectures for each task, which limits their scalability. In contrast to the task-specific models, other approaches seek to train a single blind model by building a pipeline composed of several classical degradation process, such as BSRGAN [80] and Real-ESRGAN [71]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Text-guided Diffusion Editing. Denoising diffusion models are trained to reverse the noisiging process [23, 57, 58]. Early methods focused on unconditioned generation [13], but recent trends pivot more to conditioned generation, such as" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.383, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SPIRE: Semantic Prompt-Driven Image Restoration" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.145, + 0.782, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.283, + 0.788, + 0.409 + ], + "angle": 0, + "content": "Fig. 2: Framework of SPIRE. In the training phase, we begin by synthesizing a degraded version \\( y \\), of a clean image \\( x \\). Our degradation synthesis pipeline also creates a restoration prompt \\( c_{r} \\), which contains numeric parameters that reflects the intensity of the degradation introduced. Then, we inject the synthetic restoration prompt into a ControlNet adaptor, which uses our proposed modulation fusion blocks \\( (\\gamma, \\beta) \\) to connect with the frozen backbone driven by the semantic prompt \\( c_{s} \\). During test time, the users can employ the SPIRE framework as either a blind restoration model with restoration prompt \"Remove all degradation\" and empty semantic prompt \\( \\varnothing \\), or manually adjust the restoration \\( c_{r} \\) and semantic prompts \\( c_{s} \\) to obtain what they want." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.434, + 0.789, + 0.753 + ], + "angle": 0, + "content": "image-to-image translation [52, 55] and text-to-image generation [50, 53]. Latent Diffusion [50] is a groundbreaking approach that introduces a versatile framework for improving both training and sampling efficiency, while flexible enough for general conditioning inputs. Subsequent works have built upon their pretrained text-to-image checkpoints, and designed customized architectures for different tasks like text-guided image editing. For instance, SDEdit [37] generates content for a new prompt by adding noise to the input image. Attention-based methods [21, 44, 64, 81] show that images can be edited via reweighting and replacing the cross-attention map of different prompts. Aside from the previously mentioned approaches driven by target prompts, instruction-based editing [7, 17] entail modifying a source image based on specific instructions. These textual instructions are typically synthesized using large language models [41, 47]. As CLIP serves as a cornerstone for bridging vision and language, several studies have extensively investigated the representation space of CLIP embeddings [46]. Instead of directly applying CLIP to fundamental discriminative tasks [19, 85], these works either tailor CLIP to specific applications, such as image quality assessment [28, 31], serve as semantic guidance in generative adversarial network [5] or aligning image embedding with degradation types in the feature space [26, 36]. To enhance CLIP's ability to understand numerical information, some works [42, 43] finetune it using contrastive learning on synthetic numerical data, and then incorporate the fine-tuned text embedding into diffusion training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Parameter-Efficient Diffusion Model Finetuning. To leverage the powerful generative prior in pretrained diffusion models, parameter-efficient components such as text embedding [15], low-rank approximations of model weights [20, 25], and cross attention layers [29] can be finetuned to personalize the pretrained model. Adaptor-based finetuning paradigms [39, 84] propose to keep the original" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.33, + 0.129 + ], + "angle": 0, + "content": "Qi et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "UNet weights frozen, and add new image or text conditions. This adaptor generates residual features that are subsequently added to the frozen UNet backbone." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.201, + 0.331, + 0.218 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.234, + 0.784, + 0.355 + ], + "angle": 0, + "content": "We introduce a universal approach to combine the above mentioned task-specific, strength-aware, and blind restoration methods within a unified framework (illustrated in Sec. 3.2). We further propose to decouple the learning of content and restoration prompts to better preserve the pre-trained prior while injecting new conditions, as unfolded in Sec. 3.3. Sec. 3.4 details our design on how to accurately control the restoration type and strength, as well as our proposed modulation fusion layer that adaptively fuses the restoration features back to the frozen backbone. We start with some preliminaries in the following section." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.378, + 0.373, + 0.392 + ], + "angle": 0, + "content": "3.1 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.403, + 0.785, + 0.54 + ], + "angle": 0, + "content": "Latent Diffusion Models (LDMs) [50] are probabilistic generative models that learn the underlying data distribution by iteratively removing Gaussian noise in the latent space, which is typically learned using a VAE autoencoder. Formally, the VAE encoder \\(\\mathcal{E}\\) compresses an input image \\(\\pmb{x}\\) into a compact latent representation \\(\\pmb{z} = \\mathcal{E}(\\pmb{x})\\), which can be later decoded back to the pixel space using the coupled VAE decoder \\(\\mathcal{D}\\), often learned under an image reconstruction objective: \\(\\mathcal{D}(\\mathcal{E}(\\pmb{x})) \\approx \\pmb{x}\\). During training stage, the output of a UNet [51] \\(\\epsilon_{\\theta}(z_t, t, \\pmb{y})\\) conditioned on \\(\\pmb{y}\\) (such as text, images, etc.) is parameterized [13, 56] to remove Gaussian noise \\(\\pmb{\\epsilon}\\) in the latent space \\(\\pmb{z}_t\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.34, + 0.551, + 0.785, + 0.574 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} \\mathbb {E} _ {\\left(\\boldsymbol {z} _ {0}, \\boldsymbol {y}\\right) \\sim p _ {\\text {d a t a}}, \\epsilon \\sim \\mathcal {N} (0, I), t} \\| \\boldsymbol {\\epsilon} - \\boldsymbol {\\epsilon} _ {\\theta} \\left(\\boldsymbol {z} _ {t}, t, \\boldsymbol {y}\\right) \\| _ {2} ^ {2}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.585, + 0.784, + 0.614 + ], + "angle": 0, + "content": "where \\( \\mathbf{z}_t \\) is a noisy sample of \\( \\mathbf{z}_0 \\) at sampled timestep \\( t \\). The condition \\( \\mathbf{y} \\) is randomly dropped out as \\( \\varnothing \\) to make the model unconditional." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.615, + 0.784, + 0.676 + ], + "angle": 0, + "content": "At test time, deterministic DDIM sampling [57] is utilized to convert a random noise \\( z_{T} \\) to a clean latent \\( z_{0} \\) which is decoded as final result \\( \\mathcal{D}(z_0) \\). In each timestep \\( t \\), classifier-free guidance [13, 24] can be applied to trade-off sample quality and condition alignment:" + }, + { + "type": "equation", + "bbox": [ + 0.302, + 0.689, + 0.694, + 0.705 + ], + "angle": 0, + "content": "\\[\n\\overline {{\\epsilon}} _ {\\theta} \\left(\\boldsymbol {z} _ {t}, t, \\boldsymbol {y}\\right) = \\epsilon_ {\\theta} \\left(\\boldsymbol {z} _ {t}, t, \\varnothing\\right) + w \\left(\\epsilon_ {\\theta} \\left(\\boldsymbol {z} _ {t}, t, \\boldsymbol {y}\\right) - \\epsilon_ {\\theta} \\left(\\boldsymbol {z} _ {t}, t, \\varnothing\\right)\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.716, + 0.784, + 0.763 + ], + "angle": 0, + "content": "where \\( w \\) is a scalar to adjust the guidance strength of \\( y \\). Note that the estimated noise \\( \\overline{\\epsilon}_{\\theta}(\\pmb{z}_t, t, \\pmb{y}) \\) is used to update \\( z_t \\) to \\( z_{t-1} \\), which is a approximation of the distribution gradient score [60] as \\( \\nabla_{z_t} \\log p(\\pmb{z}_t | \\pmb{y}) \\propto \\overline{\\epsilon}_{\\theta}(\\pmb{z}_t, t, \\pmb{y}) \\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.785, + 0.518, + 0.8 + ], + "angle": 0, + "content": "3.2 Text-driven Image Restoration" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Based on the LDM framework, we propose a new restoration paradigm—text-driven image restoration. Our method targets to restore images \\((\\pmb{x}\\) or \\(\\pmb{z}_0)\\) based" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.382, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "SPIRE: Semantic Prompt-Driven Image Restoration" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.237 + ], + "angle": 0, + "content": "on conditions \\(\\{\\pmb{y},\\pmb{c}_s,\\pmb{c}_r\\}\\). Specifically: \\(\\pmb{y}\\) denotes the degraded image condition, \\(\\pmb{c}_s\\) is the semantic prompt describing the clean image \\(\\pmb{x}\\) (e.g., \"a panda is sitting by the bamboo\" or \"a panda\"), and \\(\\pmb{c}_r\\) is the restoration prompt that describes the details of the degradation in terms of both the operation and parameter (e.g., \"deblur with sigma 3.0\"). We use \\(\\pmb{y} = \\mathrm{Deg}(\\pmb{x},\\pmb{c}_r)\\) to denote the degradation process which turns the clean image \\(\\pmb{x}\\) into its degraded counterpart \\(\\pmb{y}_i\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.788, + 0.315 + ], + "angle": 0, + "content": "The above text-driven image restoration model \\( p(\\pmb{z}_t|\\{\\pmb{y},\\pmb{c}_s,\\pmb{c}_r\\}) \\) can be trained using paired data. We use the text-image dataset [8] so that each clean image \\( \\pmb{x} \\) is paired with a semantic prompt \\( \\pmb{c}_s \\) (the paired alt-text). Then \\( \\pmb{y} = \\mathrm{Deg}(\\pmb{x},\\pmb{c}_r) \\) is simulated using the synthetic degradation [71] pipeline (Sec. 4.1), yielding the final paired training data \\( (\\pmb{x} \\text{ or } \\pmb{z}_0, \\{\\pmb{y},\\pmb{c}_s,\\pmb{c}_r\\}) \\sim p_{data} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.317, + 0.789, + 0.575 + ], + "angle": 0, + "content": "Comparison to Blind Restoration. Recent existing blind restoration models [34, 55, 67] also leverage diffusion priors to generate high-quality images. However, these methods are prone to hallucinate unwanted, unnatural, or oversharpened textures given that the semantic- and degradation-ambiguities persist everywhere. Fig. 3 provides an example, where the input image is degraded simultaneously with Gaussian blur and motion blur. The fully blind restoration method removes all of the degradation, leaving no way to steer the model to remove only the Gaussian blur while preserving the motion effect for aesthetic purposes. Therefore, having the flexibility to control the restoration behavior and its strength at test time is a crucial requirement. Our text-driven formulation of the problem introduces additional controlling ability using both the semantic content and the restoration instruction. In our SPIRE design, \\( \\pmb{c}_r = \\) \"Remove all degradation\" is also a plausible restoration prompt which makes our SPIRE to be compatible with the blind restoration. Further, our text-driven restoration takes advantages of the pretrained content-aware LDM using the additional semantic prompt \\( \\pmb{c}_s \\), hence able to better handle semantic ambiguities on the noisy, blurry regions whose visual objects are less recognizable." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.598, + 0.652, + 0.615 + ], + "angle": 0, + "content": "3.3 Decoupling Semantic and Restoration Prompts" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.625, + 0.788, + 0.731 + ], + "angle": 0, + "content": "To effectively learn the latent distribution \\( p(\\pmb{z}_t | \\pmb{y}, \\pmb{c}_s, \\pmb{c}_r) \\), we further decouple the conditions \\( \\{\\pmb{y}, \\pmb{c}_s, \\pmb{c}_r\\} \\) into two groups—one for the text-to-image prior \\( (\\pmb{c}_s \\rightarrow \\pmb{z}_t) \\) already imbued in the pretrained LDM model, and the other for both the image-to-image \\( (\\pmb{y} \\rightarrow \\pmb{z}_t) \\) and restoration-to-image \\( (\\pmb{c}_r \\rightarrow \\pmb{z}_t) \\) that needs to be learnt from the synthetic data. This decoupling strategy prevents catastrophic forgetting in the pretrained diffusion model and enables independent training of the restoration-aware model, whose score function [60] is expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.244, + 0.732, + 0.786, + 0.769 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {z _ {t}} \\log p (\\boldsymbol {z} _ {t} | \\boldsymbol {y}, \\boldsymbol {c} _ {s}, \\boldsymbol {c} _ {r}) \\approx \\underbrace {\\nabla_ {z _ {t}} \\log p (\\boldsymbol {z} _ {t} | \\boldsymbol {c} _ {s})} _ {\\text {S e m a n t i c - a w a r e (f r o z e n)}} + \\underbrace {\\nabla_ {z _ {t}} \\log p (\\boldsymbol {y} | \\boldsymbol {z} _ {t} , \\boldsymbol {c} _ {r})} _ {\\text {R e s t o r a t i o n - a w a r e (l e a r n a b l e)}}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In the above equation, the first part \\(\\nabla_{\\boldsymbol{z}_t}\\log p(\\boldsymbol{z}_t|\\boldsymbol{c}_s)\\) aligns with the text-to-image prior inherent in the pretrained LDM. The second term \\(\\nabla_{\\boldsymbol{z}_t}\\log p(\\boldsymbol{y}|\\boldsymbol{z}_t,\\boldsymbol{c}_r)\\) approximates the consistency constraint with degraded image \\(\\boldsymbol{y}\\), meaning the latent image \\(\\boldsymbol{z}_t\\) should be inferred by the degraded image \\(\\boldsymbol{y}\\) and the degradation" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.33, + 0.129 + ], + "angle": 0, + "content": "Qi et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "details \\( \\pmb{c}_r \\). This is somewhat similar to the reverse process of \\( \\pmb{y} = \\mathrm{Deg}(\\pmb{x}_t, \\pmb{c}_r) \\) in the pixel space. More derivation is provided in our supplement. Providing degradation information through a restoration text prompt \\( \\pmb{c}_r \\) can largely reduce the uncertainty of estimation \\( p(\\pmb{y} | \\pmb{z}_t, \\pmb{c}_r) \\), thereby alleviating the ambiguity of the restoration task and leading to better results that are less prone to hallucinations." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.23, + 0.566, + 0.245 + ], + "angle": 0, + "content": "3.4 Learning to Control the Restoration." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.251, + 0.432, + 0.432 + ], + "angle": 0, + "content": "Inspired by ControlNet [84], we employ an adaptor to model the restoration-aware branch in Eq. (2). First, the pretrained latent diffusion UNet \\(\\epsilon_{\\theta}(\\pmb{z}_t,t,\\pmb{c}_s)\\) is frozen during training to preserve text-to-image prior \\(\\nabla_{\\pmb{z}_t}\\log p(\\pmb{z}_t|\\pmb{c}_s)\\). Secondly, an encoder initialized from pretrained UNet is finetuned to learn the restoration condi" + }, + { + "type": "image", + "bbox": [ + 0.443, + 0.264, + 0.788, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.441, + 0.35, + 0.788, + 0.421 + ], + "angle": 0, + "content": "Fig. 3: Degradation ambiguities in real-world problems. By adjusting the restoration prompt, our method can preserve the motion effect that is coupled with Gaussian blur, while fully blind restoration models do not provide this flexibility." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.433, + 0.788, + 0.493 + ], + "angle": 0, + "content": "tion \\(\\nabla_{z_t}\\log p(\\pmb {y}|\\pmb {z}_t,\\pmb {c}_r)\\), which takes degradation image \\(\\pmb{y}\\), current noisy latents \\(\\pmb{z}_{t}\\) and the restoration prompt \\(\\pmb{c}_{r}\\) as input. The output of encoder is fusion features \\(\\pmb{f}_{con}\\) for frozen UNet. In this way, the choice of ControlNet implementation follows our decoupling analysis in Eq. (2)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.788, + 0.583 + ], + "angle": 0, + "content": "Conditioning on the Degraded Image \\( \\pmb{y} \\). We apply downsample layers on \\( \\pmb{y} \\) to make the feature map matching to the shape of the \\( \\pmb{z}_t \\), denoted as \\( \\mathcal{E}'(\\pmb{y}) \\). \\( \\mathcal{E}' \\) is more compact [84] than VAE encoder \\( \\mathcal{E} \\), and its trainable parameters allow model to perceive and adapt to the degraded images from our synthetic data pipeline. Then, we concatenate the downsampled feature map \\( \\mathcal{E}'(\\pmb{y}) \\) with the latent \\( \\pmb{z}_t \\), and feed them to the ControlNet encoder." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.788, + 0.81 + ], + "angle": 0, + "content": "Conditioning on the Restoration Prompt \\( \\pmb{c}_r \\). Following [50], we first use the vision-language CLIP [46] to infer the text sequence embedding \\( \\pmb{e}_r \\in \\mathbb{R}^{M \\times d} \\) for \\( \\pmb{c}_r \\), where \\( M \\) is the number of tokens, and \\( d = 768 \\) is the dimension of the image-text features. Specifically, given a restoration instruction \"deblur with sigma 3.0\", CLIP first tokenizes it into a sequence of tokens (e.g., \"3.0\" is tokenized into [\"3\", \".\", \"0)]) and embeds them into \\( M \\) distinct token embeddings. Then, stacked causal transformer layers [65] are applied on the embedding sequence to contextualize the token embeddings, resulting in \\( \\pmb{e}_r \\). Hence, tokens such as \"deblur\" may arm with the numeric information such as \"3\", \".\", and \"0\", making it possible for the model to distinguish the semantic difference of strength definition in different restoration tasks. Cross-attention process ensures that the information from \\( \\pmb{e}_r \\) is propagated to the ControlNet feature \\( f_{con} \\). While [42, 43] finetuned CLIP using contrastive learning, while we found frozen CLIP still works. The reason is that the learnable cross-attention layers can somehow ensemble the observations from frozen CLIP and squeeze useful information from it." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Modulation Fusion Layer. The decoupling of learning two branches (Sec. 3.3), despite offering the benefit of effective separation of prior and conditioning," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.382, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SPIRE: Semantic Prompt-Driven Image Restoration" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.144, + 0.784, + 0.26 + ], + "angle": 0, + "content": "
MethodPromptsParameterized Degradation with synthesized crReal-ESRGAN Degradation without cr
Sem ResFID↓LPIPS↓PSNR↑SSIM↑CLIP-I↑CLIP-T↑FID↓LPIPS↓PSNR↑SSIM↑CLIP-I↑CLIP-T↑
SwinIR [30]XX43.220.42324.400.7170.8560.28548.370.44923.450.6990.8420.284
StableSR [67]XX20.550.31321.030.6130.8860.29825.750.36420.420.5810.8640.298
DiffBIR [34]XX17.260.30222.160.6040.9120.29719.170.33021.480.5870.8980.298
ControlNet-SR [84]XX13.650.22223.750.6690.9380.30016.990.26922.950.6280.9240.299
Ours w/o textXX12.700.22123.840.6710.9390.29916.250.26223.150.6360.9290.300
DiffBIR [34] + SDEdit [37]✓X19.360.36219.390.5270.8910.30517.510.37519.150.5210.8870.308
DiffBIR [34] + CLIP [46]✓X18.460.36520.500.5260.8960.30820.310.37420.450.5390.8850.307
ControlNet-SR + CLIP [46]✓X13.000.24123.180.6480.9370.30715.160.28622.450.6100.9260.308
Ours✓✓11.340.21923.610.6650.9430.30614.420.26223.140.6330.9350.308
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.262, + 0.785, + 0.331 + ], + "angle": 0, + "content": "Table 1: Quantitative results on the MS-COCO dataset (with \\( c_{s} \\)) using our parameterized degradation (left) and Real-ESRGAN degradation (right). We also denote the prompt choice at test time. 'Sem' stands for semantic prompt; 'Res' stands for restoration prompt. The first group of baselines are tested without prompt. The second group are combined with semantic prompt in zero-shot way." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.345, + 0.784, + 0.402 + ], + "angle": 0, + "content": "
Degradation ProcessParameterizedReal-ESRGAN
p(choose)Restoration PromptDegradation Processp(choose)Restoration Prompt
Gaussian Blur0.5Deblur with {sigma∈[0.2, 3.0]} or DeblurBlur1.0
Downsample0.5Upsample to {resizing factor∈[1.0, 7.0]} or UpsampleResize1.0Remove all degradation
Gaussian Noise0.5Denoise with {sigma∈[0.0, 0.12]} or DenoiseNoise1.0
JPEG0.5Dejpeg with quality {quality factor∈[30, 92]} or DejpegJPEG1.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.403, + 0.787, + 0.472 + ], + "angle": 0, + "content": "Table 2: Our training degradation is randomly sampled in these two pipeline with \\(50\\%\\) each. (1) Images generated by our parameterized pipeline are paired with either a restoration type (e.g., \"Deblur\") or a restoration parameter prompt (e.g., \"Deblur with sigma 0.3.\") (2) In other \\(50\\%\\) iterations, degraded images \\(\\pmb{y}\\) synthesized by Real-ESRGAN are paired with the same restoration prompt \\(\\pmb{c}_r = \\text{\"Remove all degradation\"}\\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.489, + 0.788, + 0.597 + ], + "angle": 0, + "content": "may cause distribution shift in the learned representations between the frozen backbone feature \\( \\pmb{f}_{\\mathrm{skip}} \\) and ControlNet feature \\( \\pmb{f}_{\\mathrm{control}} \\). To allow for adaptive alignment of the above features, here we propose a new modulation fusion layer to fuse the multi-scale features from ControlNet to the skip connection features of the frozen backbone: \\( \\hat{\\pmb{f}}_{\\mathrm{skip}} = (1 + \\gamma)\\pmb{f}_{\\mathrm{skip}} + \\beta \\); \\( \\gamma, \\beta = \\mathcal{M}(\\pmb{f}_{\\mathrm{con}}) \\), where \\( \\gamma, \\beta \\) are the scaling and bias parameters from \\( \\mathcal{M} \\) a lightweight \\( 1\\times 1 \\) zero-initialized convolution layer." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.605, + 0.376, + 0.623 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.635, + 0.617, + 0.65 + ], + "angle": 0, + "content": "4.1 Text-based Training Data and Benchmarks" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.785, + 0.734 + ], + "angle": 0, + "content": "Our method opens up an entirely new paradigm of instruction-based image restoration. However, existing image restoration datasets such as DIV2K [2] and Flickr2K [40] do not provide high-quality semantic prompts, and Real-ESRGAN degradation is for blind restoration. To address this, we construct a new setting including training data generation and test benchmarks for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Our parameterized degradation pipeline is based on the Real-ESRGAN [71] which contains a large number of degradation types. Since the original full Real-ESRGAN degradation is difficult to be parameterized and represented as user-friendly natural language (e.g., not every one understands \"anisotropic plateau blur with beta \\(2.0^{\\prime \\prime}\\)), we choose the 4 most general degradations which are practical for parameterization to support degradation parameter-aware restoration, as shown in Tab. 2. Our parameterized pipeline skips each degradation stage" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.33, + 0.129 + ], + "angle": 0, + "content": "Qi et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.144, + 0.784, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.423, + 0.788, + 0.493 + ], + "angle": 0, + "content": "Fig. 4: Visual Comparison with other baselines. Our method of integrating both the semantic prompt \\( \\pmb{c}_{s} \\) and the restoration prompt \\( \\pmb{c}_{r} \\) outperforms image-to-image restoration (DiffBIR, Retrained ControlNet-SR) and naive zero-shot combination with semantic prompt. It achieves more sharp and clean results while maintaining consistency with the degraded image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.51, + 0.788, + 0.648 + ], + "angle": 0, + "content": "with a \\(50\\%\\) probability to increase the diversity of restoration prompts and support single task restoration. The descriptions of the selected degradations are appended following the order of the image degradations to synthesize \\(\\pmb{c}_r\\). In real application, users can control the restoration type and its strength by modifying restoration prompt. As shown in third column of Fig. 3, driven by restoration prompt \"Deblur with sigma \\(3.0^{\\prime \\prime}\\), our framework removes the Gaussian blur, while preserving motion blur. The representation of the restoration strength is also general enough to decouple different tasks in natural language as shown in Fig. 7. More visual results are presented in our supplement." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.652, + 0.788, + 0.744 + ], + "angle": 0, + "content": "Training data construction. In the training stage, we sample near 100 million text-image pairs from internal data source and use the alternative text label with highest relevance as the semantic prompt \\( \\pmb{c}_s \\) in the framework. We drop out the semantic prompt to empty string \\( \\varnothing \\) by a probability of \\( 10\\% \\) to support semantic-agnostic restoration. We mix the Real-ESRGAN degradation pipeline and our parameterized pipeline by \\( 50\\% \\) each in training iterations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Evaluation Setting. We randomly choose 3000 pairs of image \\( \\pmb{x} \\) and semantic prompt \\( \\pmb{c}_{s} \\) from the MS-COCO [33] validation set. Then, we build two test sets on parameterized and the Real-ESRGAN processes separately. At the left half of Tab. 1, images are sent to the parameterized degradation to form the synthesize degradation image \\( \\pmb{y} \\) with restoration prompt \\( \\pmb{c}_{r} \\). In this setting, we expect \"Ours\" model to fully utilize the \\( \\pmb{c}_{s} \\) describing the semantic and the \\( \\pmb{c}_{r} \\)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.382, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SPIRE: Semantic Prompt-Driven Image Restoration" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.144, + 0.614, + 0.226 + ], + "angle": 0, + "content": "
MethodFID↓LPIPS↓PSNR↑SSIM↑CLIP-I↑
Real-ESRGAN [72]32.370.31222.520.6460.683
DiffBIR [34] (zero-shot)30.710.35422.010.5260.921
StableSR24.440.31121.620.5330.928
Ours w/o text (zero-shot)28.800.35221.680.5490.927
Ours w/o text (finetuned)22.450.32121.380.5320.932
" + }, + { + "type": "table_caption", + "bbox": [ + 0.617, + 0.161, + 0.765, + 0.218 + ], + "angle": 0, + "content": "Table 3: Numerical results on the DIV2K testset without any prompt." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.241, + 0.788, + 0.348 + ], + "angle": 0, + "content": "describing the degradations. On the right half, the same images are fed to the Real-ESRGAN degradation—so the degradation is aligned with open-sourced checkpoints [30, 34, 67]. Because not every model can utilize both the semantic \\((c_{s})\\) and restoration \\((c_{r})\\) information, Tab. 1 provides a “Prompts” column denoting what information the model relied on. In summary, our constructed benchmarks based on MS-COCO can serve multiple purposes—from fully-blind, content-prompted, to all-prompted restoration." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.35, + 0.788, + 0.412 + ], + "angle": 0, + "content": "Evaluation Metrics. We use the generative metric FID [22] and perceptual metric LPIPS [86] for quantitative evaluation of image quality. PSNR, and SSIM are reported for reference. We also evaluate similarity scores in the clip embedding space with ground truth image (CLIP-I) and caption (CLIP-T)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.419, + 0.476, + 0.435 + ], + "angle": 0, + "content": "4.2 Comparing with baselines" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.443, + 0.642, + 0.459 + ], + "angle": 0, + "content": "We compare our method with three categories of baselines:" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.466, + 0.786, + 0.496 + ], + "angle": 0, + "content": "- Open-sourced image restoration checkpoints, including the regression model SwinIR [30] and the latent diffusion models StableSR [67] and DiffBIR [34]." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.498, + 0.788, + 0.528 + ], + "angle": 0, + "content": "Combining the state-of-the-art method [34] with the zero-shot post-editing [37] (DiffBIR+SDEdit) or zero-shot injection through CLIP [46] (DiffBIR + CLIP)." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.53, + 0.796, + 0.591 + ], + "angle": 0, + "content": "- ControlNet-SR (retrained): We adapt the image-to-image ControlNet model [84] for super-resolution. For consistency, we maintain the same training iterations as our own model. We also evaluate its zero-shot capabilities with CLIP guidance (ControlNet-SR + CLIP)." + }, + { + "type": "list", + "bbox": [ + 0.216, + 0.466, + 0.796, + 0.591 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Quantitative comparison with baselines is presented in Tab. 1. On the left, we evaluate our full model with the parameterized degradation test set. Since open-sourced baselines are pretrained on Real-ESRGAN degradation, not perfectly aligned with our parameterized degradation, we also evaluate our full model on MS-COCO with Real-ESRGAN degradation by setting our restoration prompt to \"Remove all degradation\". Thanks to our prompts guidance and architecture improvements, our full model achieves best FID, LPIPS, CLIP-Image score, which means better image quality and semantic restoration. \"Ours w/o text\" is the same checkpoint conditioned on only degradation image and has high pixel-level similarity (the second best PSNR and SSIM) with GT. Although SwinIR has highest PSNR, its visual results are blurry. Although combining semantic prompt in zero-shot way can bring marginal improvement in FID and CLIP similarity with caption, it deteriorates the image restoration capacity and results in worse LPIPS and CLIP-Image. In contrast, semantic prompt guidance improves the CLIP image similarity of our full model. Tab. 3 shows the evaluation results of our model on DIV2K test set provided by StableSR. In zero-shot" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.329, + 0.128 + ], + "angle": 0, + "content": "Qi et al." + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.133, + 0.316, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.317, + 0.133, + 0.409, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.133, + 0.503, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.133, + 0.596, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.596, + 0.133, + 0.688, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.689, + 0.133, + 0.782, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.255, + 0.784, + 0.281 + ], + "angle": 0, + "content": "Fig. 5: Real-world image restorations. Our framework enhances real-world low-quality images following semantic prompts provided by language models or user input." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.285, + 0.785, + 0.351 + ], + "angle": 0, + "content": "
MetricSwinIR [30]Diffbir [34]Ablation of restoration prompt crAblation of semantic prompt cs
Sem c_sXXXXshortunmatchedfull
Res crXX"Deblur 0.6"... 1.0"... 2.0"... 5.0""Remove all degradation"
CLIP-IQA↑0.61900.69830.40740.41590.48310.47970.58180.57500.64600.6610
MUSIQ↑63.6569.6935.8239.8749.9848.8161.6860.2767.5467.82
MANIQA↑0.40350.26190.22990.24870.29700.28210.36940.36370.43550.4474
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.352, + 0.784, + 0.379 + ], + "angle": 0, + "content": "Table 4: Quantitative results on real-world images. Our aesthetic quality is improved when increasing restoration strength or using a more accurate and detailed prompt." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.393, + 0.785, + 0.438 + ], + "angle": 0, + "content": "test setting, our model has lower FID and LPIPS than DiffBIR. To fairly compare with StableSR, we finetune our model on their training set. Our finetuned model also outperforms StableSR in FID and CLIP-I." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.443, + 0.785, + 0.593 + ], + "angle": 0, + "content": "Qualitative comparison with baselines is presented in Fig. 4, where the corresponding semantic prompt is provided below each row. Image-to-image baselines such as DiffBir and the retrained ControlNet-SR can easily generate artifacts (e.g., hydrant in the first row) and blurry images (e.g., hands example in third row) that may look close to an \"average\" estimation. Besides, naive combination with semantic prompt in the zero-shot approach results in heavy hallucinations and may fail to preserve the object identity in the input image (e.g., giraffes on the second row). Unlike the existing methods, our full model considers semantic prompt, degradation image and restoration prompt in both training and test stages, which makes its results more aligned with all conditions." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.618, + 0.447, + 0.631 + ], + "angle": 0, + "content": "4.3 Real-world restoration" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Besides evaluation on synthetic degradations, we present an analysis on real-world images in Tab. 4 and Fig. 5. We choose RealPhoto60 [78] as our real-world test-set, and use non-reference metrics CLIP-IQA [66], MUSIQ [27] and MANIQA [76] scores to study the influence of the semantic and restoration prompts. First, we fix the semantic prompt to the empty string \\(\\varnothing\\) and conduct an ablation study on the restoration prompt. As the deblurring strength increases from 0.6 to 2.0, the image quality is improved consistently. A blind prompt \"Remove all degradation\" can provide better results than the manually controlled restoration prompts. This showcases that the proposed model is flexible to work with generic restoration prompts, especially when coming up with the perfect prompt is challenging. We also fix the blind restoration prompt and change the semantic prompts. Following [78], we feed degraded images to LLaVA [35] and generate corresponding semantic prompts. Our experiments" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.383, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "SPIRE: Semantic Prompt-Driven Image Restoration" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.144, + 0.784, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.306, + 0.785, + 0.348 + ], + "angle": 0, + "content": "Fig. 6: Test-time semantic prompting. Our framework restores degraded images guided by flexible semantic prompts, while unrelated background elements and global tones remain aligned with the degraded input conditioning." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.361, + 0.588, + 0.607 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.59, + 0.364, + 0.788, + 0.584 + ], + "angle": 0, + "content": "Fig.7: Prompt space walking visualization for the restoration prompt. Given the same degraded input (upper left) and empty semantic prompt \\(\\varnothing\\), our method can decouple the restoration direction and strength via only prompting the quantitative number in natural language. An interesting finding is that our model learns a continuous range of restoration strength from discrete language tokens." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.788, + 0.858 + ], + "angle": 0, + "content": "show that detailed full descriptions (about 60 words) of the degraded image from LLaVA can be more effective to improve the visual results than short prompts that have less than 20 words. Unmatched prompts simulate the case when a user provides wrong prompts by randomly re-sampling unmatched long prompts from the data set. An interesting observation is that unmatched long prompts also improve the aesthetic quality of images against a short prompt or a blind restoration. Nonetheless, it can lead to more \"hallucination\" when the semantic input prompt is unmatched with the real ground truth. In contrast to concurrent work [62] that leverages implicit clip embedding, our text-based formulation provides better control of hallucinations by adjusting the LLM-synthesized or manually-designed prompt at test-time. Note that many studies [6,18] find current image restoration metrics are biased towards specific datasets and not fully aligned with human's visual perception. The quantitative results of previous works [34,71] are provided for reference. More results and discussions are provided in the supplements." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.331, + 0.129 + ], + "angle": 0, + "content": "Qi et al." + }, + { + "type": "table", + "bbox": [ + 0.248, + 0.144, + 0.756, + 0.229 + ], + "angle": 0, + "content": "
MethodSem csRes crFID↓LPIPS↓PSNR↑SSIM↑CLIP-I↑CLIP-T↑
ControlNet-SRXX13.650.22223.750.6690.9380.300
ControlNet-SR✓✓12.140.22323.500.6600.9400.306
Ours✓wo param11.580.22323.610.6650.9420.305
Ours✓✓11.340.21923.610.6650.9430.305
" + }, + { + "type": "table_caption", + "bbox": [ + 0.282, + 0.229, + 0.721, + 0.242 + ], + "angle": 0, + "content": "Table 5: Ablation of architecture and degradation strength in \\( {c}_{r} \\)" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.247, + 0.449, + 0.263 + ], + "angle": 0, + "content": "4.4 Prompting the SPIRE" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.268, + 0.787, + 0.344 + ], + "angle": 0, + "content": "Semantic Prompting. Our model can be driven by user-defined semantic prompts. As shown in Fig. 6, the blind restoration with empty string generates ambiguous and unrealistic identity. In contrast, our framework reconstructs sharp and realistic results. The object identity accurately follows user's semantic prompt, while the global layout and color tone remain consistent with input." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.344, + 0.788, + 0.495 + ], + "angle": 0, + "content": "Restoration Prompting. Users also have the freedom to adjust the degradation types and strengths in the restoration prompts. As shown in Fig. 7, the denoising and deblurring restoration are well decoupled. Using the task-specific prompts, our method can get a clean but blur (second row), or sharp but noisy (third row) result. In addition, our model learns the continuous space of restoration as the result becomes cleaner progressively if we modify the \"Denoise with sigma 0.06\" to \"0.24\". Beyond purely relying on numerical description, our model \"understand\" the difference of strength definition for tasks, such as the range difference of denoise sigma and deblur sigma, which makes it promising to be a universal representation for restoration strength control." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.513, + 0.39, + 0.529 + ], + "angle": 0, + "content": "4.5 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.535, + 0.787, + 0.627 + ], + "angle": 0, + "content": "We conduct several ablations in Tab. 5. Providing both prompts \\( c_s \\), \\( c_r \\) to \"ControlNet-SR\" baseline improves the FID by 1.5. Adding the modulation fusion layer (\"Ours\") further improves the generative quality (0.8 for FID) and pixel-level similarity (0.11 dB). Embedding degradation parameters in the restoration prompt, not only enables restoration strength prompting (Fig. 7), but also improves image quality (FID and LPIPS). More details are in the supplement." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.646, + 0.36, + 0.663 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.842 + ], + "angle": 0, + "content": "We have presented a unified text-driven framework for instructed image restoration using semantic and restoration prompts. We design our model in a decoupled way to better preserve the semantic text-to-image generative prior while efficiently learning to control both the restoration direction and its strength. To the best of our knowledge, this is the first framework to support both semantic and parameter-embedded restoration instructions simultaneously, allowing users to flexibly prompt-tune the results up to their expectations. Our extensive experiments have shown that our method significantly outperforms prior works in terms of both quantitative and qualitative results. Our model and evaluation benchmark have established a new paradigm of instruction-based image restoration, paving the way for further multi-modality generative imaging applications." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.383, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SPIRE: Semantic Prompt-Driven Image Restoration" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.218 + ], + "angle": 0, + "content": "1. Abuolaim, A., Delbracio, M., Kelly, D., Brown, M.S., Milanfar, P.: Learning to reduce defocus blur by realistically modeling dual-pixel data. In: ICCV. pp. 2289-2298 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.219, + 0.785, + 0.287 + ], + "angle": 0, + "content": "2. Agustsson, E., Timofte, R.: NTIRE 2017 challenge on single image superresolution: Dataset and study. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition Workshops, CVPR Workshops 2017, Honolulu, HI, USA, July 21-26, 2017. pp. 1122-1131. IEEE Computer Society (2017). https://doi.org/10.1109/CVPRW.2017.150, https://doi.org/10.1109/CVPRW.2017.150" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.288, + 0.785, + 0.315 + ], + "angle": 0, + "content": "3. Avrahami, O., Fried, O., Lischinski, D.: Blended latent diffusion. ACM Transactions on Graphics (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.316, + 0.785, + 0.343 + ], + "angle": 0, + "content": "4. Avrahami, O., Lischinski, D., Fried, O.: Blended diffusion for text-driven editing of natural images. In: CVPR. pp. 18208-18218 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.343, + 0.785, + 0.383 + ], + "angle": 0, + "content": "5. Bai, Y., Wang, C., Xie, S., Dong, C., Yuan, C., Wang, Z.: Textir: A simple framework for text-based editable image restoration. arXiv preprint arXiv:2302.14736 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.384, + 0.785, + 0.412 + ], + "angle": 0, + "content": "6. Blau, Y., Michaeli, T.: The perception-distortion tradeoff. In: CVPR. pp. 6228-6237 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.412, + 0.785, + 0.439 + ], + "angle": 0, + "content": "7. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.44, + 0.785, + 0.522 + ], + "angle": 0, + "content": "8. Chen, X., Wang, X., Changpinyo, S., Piergiovanni, A., Padlewski, P., Salz, D., Goodman, S.A., Grycner, A., Mustafa, B., Beyer, L., Kolesnikov, A., Puigcerver, J., Ding, N., Rong, K., Akbari, H., Mishra, G., Xue, L., Thapliyal, A., Bradbury, J., Kuo, W., Seyedhosseini, M., Jia, C., Ayan, B.K., Riquelme, C., Steiner, A., Angelova, A., Zhai, X., Houlsby, N., Soricut, R.: Pali: A jointly-scaled multilingual language-image model. In: ICLR (2023), https://arxiv.org/abs/2209.06794" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.523, + 0.785, + 0.551 + ], + "angle": 0, + "content": "9. Chen, Z., Zhang, Y., Gu, J., Yuan, X., Kong, L., Chen, G., Yang, X.: Image superresolution with text prompt diffusion. arXiv preprint arXiv:2303.06373 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.551, + 0.785, + 0.578 + ], + "angle": 0, + "content": "0. Couairon, G., Verbeek, J., Schwenk, H., Cord, M.: Diffedit: Diffusion-based semantic image editing with mask guidance. In: ICLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.579, + 0.785, + 0.619 + ], + "angle": 0, + "content": "1. Delbracio, M., Milanfar, P.: Inversion by direct iteration: An alternative to denoising diffusion for image restoration. Transactions on Machine Learning Research (2023), https://openreview.net/forum?id=VmyFF5lL3F, featured Certification" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.62, + 0.785, + 0.66 + ], + "angle": 0, + "content": "2. Delbracio, M., Talebei, H., Milanfar, P.: Projected distribution loss for image enhancement. In: 2021 IEEE International Conference on Computational Photography (ICCP). pp. 1-12. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.661, + 0.785, + 0.688 + ], + "angle": 0, + "content": "3. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Neural Information Processing Systems (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.689, + 0.785, + 0.716 + ], + "angle": 0, + "content": "4. Dong, C., Loy, C.C., He, K., Tang, X.: Image super-resolution using deep convolutional networks. IEEE Trans. Pattern Anal. Mach. Intell. 38(2), 295-307 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.717, + 0.785, + 0.757 + ], + "angle": 0, + "content": "5. Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A.H., Chechik, G., Cohen-Or, D.: An image is worth one word: Personalizing text-to-image generation using textual inversion. In: ICLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.758, + 0.785, + 0.785 + ], + "angle": 0, + "content": "6. Galteri, L., Seidenari, L., Bertini, M., Bimbo, A.: Deep generative adversarial compression artifact removal. In: ICCV (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.786, + 0.785, + 0.84 + ], + "angle": 0, + "content": "7. Geng, Z., Yang, B., Hang, T., Li, C., Gu, S., Zhang, T., Bao, J., Zhang, Z., Hu, H., Chen, D., Guo, B.: Instructdiffusion: A generalist modeling interface for vision tasks. CoRR abs/2309.03895 (2023). https://doi.org/10.48550/arXiv.2309.03895, https://doi.org/10.48550/arXiv.2309.03895" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.329, + 0.129 + ], + "angle": 0, + "content": "Qi et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.272 + ], + "angle": 0, + "content": "18. Gu, J., Cai, H., Dong, C., Ren, J.S., Timofte, R., Gong, Y., Lao, S., Shi, S., Wang, J., Yang, S., Wu, T., Xia, W., Yang, Y., Cao, M., Heng, C., Fu, L., Zhang, R., Zhang, Y., Wang, H., Song, H., Wang, J., Fan, H., Hou, X., Sun, M., Li, M., Zhao, K., Yuan, K., Kong, Z., Wu, M., Zheng, C., Conde, M.V., Burchi, M., Feng, L., Zhang, T., Li, Y., Xu, J., Wang, H., Liao, Y., Li, J., Xu, K., Sun, T., Xiong, Y., Keshari, A., Komal, Thakur, S., Jakhetiya, V., Subudhi, B.N., Yang, H.H., Chang, H.E., Huang, Z.K., Chen, W.T., Kuo, S.Y., Dutta, S., Das, S.D., Shah, N.A., Tiwari, A.K.: Ntire 2022 challenge on perceptual image quality assessment. In: CVPRW. pp. 951-967 (June 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.273, + 0.786, + 0.3 + ], + "angle": 0, + "content": "19. Gu, X., Lin, T.Y., Kuo, W., Cui, Y.: Open-vocabulary object detection via vision and language knowledge distillation. In: ICLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.786, + 0.328 + ], + "angle": 0, + "content": "20. Han, L., Li, Y., Zhang, H., Milanfar, P., Metaxas, D., Yang, F.: Svdiff: Compact parameter space for diffusion fine-tuning. In: ICCV. pp. 7323-7334 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.786, + 0.37 + ], + "angle": 0, + "content": "21. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-or, D.: Prompt-to-prompt image editing with cross-attention control. In: ICLR (2023), https://openreview.net/forum?id=CDixzkzeyb" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.37, + 0.786, + 0.411 + ], + "angle": 0, + "content": "22. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: NeurIPS (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.786, + 0.438 + ], + "angle": 0, + "content": "23. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. NeurIPS 33, 6840-6851 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.786, + 0.481 + ], + "angle": 0, + "content": "24. Ho, J., Salimans, T.: Classifier-free diffusion guidance. In: NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications (2021), https://openreview.net/forum?id=qw8AKxfYbI" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.786, + 0.522 + ], + "angle": 0, + "content": "25. Hu, E.J., yelong shen, Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-rank adaptation of large language models. In: ICLR (2022), https://openreview.net/forum?id=nZeVKeeFYf9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.786, + 0.55 + ], + "angle": 0, + "content": "26. Jiang, Y., Zhang, Z., Xue, T., Gu, J.: Autodir: Automatic all-in-one image restoration with latent diffusion. arXiv preprint arXiv:2310.10123 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.786, + 0.591 + ], + "angle": 0, + "content": "27. Ke, J., Wang, Q., Wang, Y., Milanfar, P., Yang, F.: Musiq: Multi-scale image quality transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5148-5157 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.591, + 0.786, + 0.632 + ], + "angle": 0, + "content": "28. Ke, J., Ye, K., Yu, J., Wu, Y., Milanfar, P., Yang, F.: Vila: Learning image aesthetics from user comments with vision-language pretraining. In: CVPR. pp. 10041-10051 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.632, + 0.786, + 0.661 + ], + "angle": 0, + "content": "29. Kumari, N., Zhang, B., Zhang, R., Shechtman, E., Zhu, J.Y.: Multi-concept customization of text-to-image diffusion. In: CVPR. pp. 1931-1941 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.786, + 0.688 + ], + "angle": 0, + "content": "30. Liang, J., Cao, J., Sun, G., Zhang, K., Gool, L.V., Timofte, R.: Swinir: Image restoration using swin transformer. In: Proceedings of ICCV Workshops (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.786, + 0.716 + ], + "angle": 0, + "content": "31. Liang, Z., Li, C., Zhou, S., Feng, R., Loy, C.C.: Iterative prompt learning for unsupervised backlit image enhancement. In: ICCV. pp. 8094-8103 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.786, + 0.744 + ], + "angle": 0, + "content": "32. Lim, B., Son, S., Kim, H., Nah, S., Lee, K.M.: Enhanced deep residual networks for single image super-resolution. In: Proceedings of CVPR Workshops (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.786, + 0.799 + ], + "angle": 0, + "content": "33. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, L.: Microsoft coco: Common objects in context. In: ECCV. ECCV (September 2014), https://www.microsoft.com/en-us/research/publication/microsoft-coco-common-objects-in-context/" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.786, + 0.841 + ], + "angle": 0, + "content": "34. Lin, X., He, J., Chen, Z., Lyu, Z., Fei, B., Dai, B., Ouyang, W., Qiao, Y., Dong, C.: Diffbir: Towards blind image restoration with generative diffusion prior. arXiv preprint arXiv:2308.15070 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.383, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SPIRE: Semantic Prompt-Driven Image Restoration" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.777, + 0.161 + ], + "angle": 0, + "content": "35. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.162, + 0.785, + 0.217 + ], + "angle": 0, + "content": "36. Luo, Z., Gustafsson, F.K., Zhao, Z., Sjolund, J., Schon, T.B.: Controlling vision-language models for multi-task image restoration. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=t3vnnLeajU" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.244 + ], + "angle": 0, + "content": "37. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: SDEdit: Guided image synthesis and editing with stochastic differential equations. In: ICLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.785, + 0.285 + ], + "angle": 0, + "content": "38. Mokady, R., Hertz, A., Aberman, K., Pritch, Y., Cohen-Or, D.: Null-text inversion for editing real images using guided diffusion models. In: CVPR. pp. 6038-6047 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.327 + ], + "angle": 0, + "content": "39. Mou, C., Wang, X., Xie, L., Zhang, J., Qi, Z., Shan, Y., Qie, X.: T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.508 + ], + "angle": 0, + "content": "40. Nah, S., Timofte, R., Baik, S., Hong, S., Moon, G., Son, S., Lee, K.M., Wang, X., Chan, K.C.K., Yu, K., Dong, C., Loy, C.C., Fan, Y., Yu, J., Liu, D., Huang, T.S., Sim, H., Kim, M., Park, D., Kim, J., Chun, S.Y., Haris, M., Shakhnarovich, G., Ukita, N., Zamir, S.W., Arora, A., Khan, S.H., Khan, F.S., Shao, L., Gupta, R.K., Chudasama, V.M., Patel, H., Upla, K.P., Fan, H., Li, G., Zhang, Y., Li, X., Zhang, W., He, Q., Purohit, K., Rajagopalan, A.N., Kim, J., Tofighi, M., Guo, T., Monga, V.: NTIRE 2019 challenge on video deblurring: Methods and results. In: IEEE Conference on Computer Vision and Pattern Recognition Workshops, CVPR Workshops 2019, Long Beach, CA, USA, June 16-20, 2019. pp. 1974-1984. Computer Vision Foundation / IEEE (2019). https://doi.org/10.1109/CVPRW.2019.00249, http://openaccess.thecvf.com/content_CVPRW_2019/html/NTIRE/Nah_NTIRE_2019_Challenge_on_Video_Deblurring_Methods_and_Results_CVPRW_2019_paper.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.508, + 0.503, + 0.522 + ], + "angle": 0, + "content": "41. OpenAI: Gpt-4 technical report (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.785, + 0.549 + ], + "angle": 0, + "content": "42. Paiss, R., Chefer, H., Wolf, L.: No token left behind: Explainability-aided image classification and generation. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.549, + 0.785, + 0.577 + ], + "angle": 0, + "content": "43. Paiss, R., Ephrat, A., Tov, O., Zada, S., Mosseri, I., Irani, M., Dekel, T.: Teaching clip to count to ten. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.577, + 0.785, + 0.618 + ], + "angle": 0, + "content": "44. Parmar, G., Kumar Singh, K., Zhang, R., Li, Y., Lu, J., Zhu, J.Y.: Zero-shot image-to-image translation. In: ACM SIGGRAPH 2023 Conference Proceedings. pp. 1-11 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.785, + 0.66 + ], + "angle": 0, + "content": "45. Prakash, M., Delbracio, M., Milanfar, P., Jug, F.: Interpretable unsupervised diversity denoising and artefact removal. In: ICLR (2022), https://openreview.net/forum?id=DfMqlBOPXjM" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.66, + 0.785, + 0.715 + ], + "angle": 0, + "content": "46. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.715, + 0.785, + 0.743 + ], + "angle": 0, + "content": "47. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners. OpenAI blog 1(8), 9 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.743, + 0.785, + 0.785 + ], + "angle": 0, + "content": "48. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.785, + 0.813 + ], + "angle": 0, + "content": "49. Ren, M., Delbracio, M., Talebi, H., Gerig, G., Milanfar, P.: Multiscale structure guided diffusion for image deblurring. In: ICCV. pp. 10721-10733 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.841 + ], + "angle": 0, + "content": "50. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10684-10695 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.33, + 0.129 + ], + "angle": 0, + "content": "Qi et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.787, + 0.203 + ], + "angle": 0, + "content": "51. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.786, + 0.244 + ], + "angle": 0, + "content": "52. Sahara, C., Chan, W., Chang, H., Lee, C., Ho, J., Salimans, T., Fleet, D., Norouzi, M.: Palette: Image-to-image diffusion models. In: ACM SIGGRAPH 2022 Conference Proceedings. pp. 1-10 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.787, + 0.284 + ], + "angle": 0, + "content": "53. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.285, + 0.787, + 0.325 + ], + "angle": 0, + "content": "54. Sahara, C., Ho, J., Chan, W., Salimans, T., Fleet, D.J., Norouzi, M.: Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(4), 4713-4726 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.326, + 0.787, + 0.367 + ], + "angle": 0, + "content": "55. Sahara, C., Ho, J., Chan, W., Salimans, T., Fleet, D.J., Norouzi, M.: Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(4), 4713-4726 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.367, + 0.787, + 0.407 + ], + "angle": 0, + "content": "56. Salimans, T., Ho, J.: Progressive distillation for fast sampling of diffusion models. In: ICLR. OpenReview.net (2022), https://openreview.net/forum?id=TIdIXIpzhoI" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.408, + 0.787, + 0.434 + ], + "angle": 0, + "content": "57. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.435, + 0.787, + 0.462 + ], + "angle": 0, + "content": "58. Song, Y., Ermon, S.: Generative modeling by estimating gradients of the data distribution. In: NeurIPS. pp. 11895-11907 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.463, + 0.787, + 0.489 + ], + "angle": 0, + "content": "59. Song, Y., Shen, L., Xing, L., Ermon, S.: Solving inverse problems in medical imaging with score-based generative models. In: ICLR. OpenReview.net (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.489, + 0.787, + 0.529 + ], + "angle": 0, + "content": "60. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. In: ICLR (2021), https://openreview.net/forum?id=PxTIG12RRHS" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.53, + 0.787, + 0.556 + ], + "angle": 0, + "content": "61. Su, S., Delbracio, M., Wang, J., Sapiro, G., Heidrich, W., Wang, O.: Deep video deblurring for hand-held cameras. In: CVPR. pp. 1279-1288 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.557, + 0.787, + 0.597 + ], + "angle": 0, + "content": "62. Sun, H., Li, W., Liu, J., Chen, H., Pei, R., Zou, X., Yan, Y., Yang, Y.: Coser: Bridging image and language for cognitive super-resolution. arXiv preprint arXiv:2311.16512 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.598, + 0.787, + 0.624 + ], + "angle": 0, + "content": "63. Tu, Z., Talebi, H., Zhang, H., Yang, F., Milanfar, P., Bovik, A.C., Li, Y.: MAXIM: multi-axis MLP for image processing. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.625, + 0.787, + 0.651 + ], + "angle": 0, + "content": "64. Tumanyan, N., Geyer, M., Bagon, S., Dekel, T.: Plug-and-play diffusion features for text-driven image-to-image translation. In: CVPR. pp. 1921-1930 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.652, + 0.787, + 0.678 + ], + "angle": 0, + "content": "65. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. NeurIPS 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.679, + 0.787, + 0.704 + ], + "angle": 0, + "content": "66. Wang, J., Chan, K.C., Loy, C.C.: Exploring clip for assessing the look and feel of images. In: AAAI (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.705, + 0.787, + 0.732 + ], + "angle": 0, + "content": "67. Wang, J., Yue, Z., Zhou, S., Chan, K.C., Loy, C.C.: Exploiting diffusion prior for real-world image super-resolution. arXiv preprint arXiv:2305.07015 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.733, + 0.787, + 0.759 + ], + "angle": 0, + "content": "68. Wang, L., Wang, Y., Lin, Z., Yang, J., An, W., Guo, Y.: Learning a single network for scale-arbitrary super-resolution. In: ICCV. pp. 4801-4810 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.76, + 0.787, + 0.786 + ], + "angle": 0, + "content": "69. Wang, X., Chen, X., Ni, B., Wang, H., Tong, Z., Liu, Y.: Deep arbitrary-scale image super-resolution via scale-equivariance pursuit. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.787, + 0.787, + 0.813 + ], + "angle": 0, + "content": "70. Wang, X., Li, Y., Zhang, H., Shan, Y.: Towards real-world blind face restoration with generative facial prior. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.814, + 0.787, + 0.84 + ], + "angle": 0, + "content": "71. Wang, X., Xie, L., Dong, C., Shan, Y.: Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In: ICCV. pp. 1905-1914 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.383, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SPIRE: Semantic Prompt-Driven Image Restoration" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.786, + 0.189 + ], + "angle": 0, + "content": "72. Wang, X., Yu, K., Wu, S., Gu, J., Liu, Y., Dong, C., Qiao, Y., Loy, C.C.: ESRGAN: enhanced super-resolution generative adversarial networks. In: Proceedings of ECCV Workshops (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.218 + ], + "angle": 0, + "content": "73. Wang, Z., Cun, X., Bao, J., Zhou, W., Liu, J., Li, H.: Uformer: A general u-shaped transformer for image restoration. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.786, + 0.245 + ], + "angle": 0, + "content": "74. Whang, J., Delbracio, M., Talebi, H., Saharia, C., Dimakis, A.G., Milanfar, P.: Deblurring via stochastic refinement. In: CVPR. pp. 16293-16303 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.786, + 0.273 + ], + "angle": 0, + "content": "75. Wu, R., Yang, T., Sun, L., Zhang, Z., Li, S., Zhang, L.: Seesr: Towards semantics-aware real-world image super-resolution. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.786, + 0.314 + ], + "angle": 0, + "content": "76. Yang, S., Wu, T., Shi, S., Lao, S., Gong, Y., Cao, M., Wang, J., Yang, Y.: Maniaq: Multi-dimension attention network for no-reference image quality assessment. In: CVPR. pp. 1191-1200 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.786, + 0.342 + ], + "angle": 0, + "content": "77. Yang, T., Ren, P., Xie, X., Zhang, L.: Gan prior embedded network for blind face restoration in the wild. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.786, + 0.383 + ], + "angle": 0, + "content": "78. Yu, F., Gu, J., Li, Z., Hu, J., Kong, X., Wang, X., He, J., Qiao, Y., Dong, C.: Scaling up to excellence: Practicing model scaling for photo-realistic image restoration in the wild. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.383, + 0.786, + 0.411 + ], + "angle": 0, + "content": "79. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Multi-stage progressive image restoration. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.786, + 0.439 + ], + "angle": 0, + "content": "80. Zhang, K., Liang, J., Van Gool, L., Timofte, R.: Designing a practical degradation model for deep blind image super-resolution. In: ICCV. pp. 4791-4800 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.786, + 0.481 + ], + "angle": 0, + "content": "81. Zhang, K., Mo, L., Chen, W., Sun, H., Su, Y.: Magicbrush: A manually annotated dataset for instruction-guided image editing. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.786, + 0.522 + ], + "angle": 0, + "content": "82. Zhang, K., Zuo, W., Chen, Y., Meng, D., Zhang, L.: Beyond a Gaussian denoiser: Residual learning of deep CNN for image denoising. IEEE Transactions on Image Processing 26(7), 3142-3155 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.786, + 0.551 + ], + "angle": 0, + "content": "83. Zhang, K., Zuo, W., Zhang, L.: Ffdnet: Toward a fast and flexible solution for CNN based image denoising. IEEE Transactions on Image Processing (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.551, + 0.786, + 0.577 + ], + "angle": 0, + "content": "84. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.577, + 0.786, + 0.618 + ], + "angle": 0, + "content": "85. Zhang, R., Fang, R., Gao, P., Zhang, W., Li, K., Dai, J., Qiao, Y., Li, H.: Tip-adapter: Training-free clip-adapter for better vision-language modeling. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.786, + 0.647 + ], + "angle": 0, + "content": "86. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.647 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/SPIRE_ Semantic Prompt-Driven Image Restoration/0f07d6c7-f604-4c26-b13e-8eef2743217e_origin.pdf b/2024/SPIRE_ Semantic Prompt-Driven Image Restoration/0f07d6c7-f604-4c26-b13e-8eef2743217e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e7b129be4600d38942379c73ac3298631d23af18 --- /dev/null +++ b/2024/SPIRE_ Semantic Prompt-Driven Image Restoration/0f07d6c7-f604-4c26-b13e-8eef2743217e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5e4c3570e73d1de145b2161696576c82d676a75613992bcdbbe49c6f79dfa967 +size 8197963 diff --git a/2024/SPIRE_ Semantic Prompt-Driven Image Restoration/full.md b/2024/SPIRE_ Semantic Prompt-Driven Image Restoration/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8751cb386806f4a9885dfa6878e0db8c54feef96 --- /dev/null +++ b/2024/SPIRE_ Semantic Prompt-Driven Image Restoration/full.md @@ -0,0 +1,356 @@ +# SPIRE: Semantic Prompt-Driven Image Restoration + +Chenyang Qi $^{1,2*}$ , Zhengzhong Tu $^{1}$ , Keren Ye $^{1}$ , Maurizio Delbracio $^{1}$ , Peyman Milanfar $^{1}$ , Qifeng Chen $^{2}$ , and Hossein Talebi $^{1}$ + +1 Google Research 2 HKUST + +![](images/87bac0dffcf974577e47bdf79cb79aef0d83d7e7446a26a027cd557068a560f0.jpg) +Input + +![](images/d15147ed5a186f3928ae2c99ae961fecedfd982337ec7cb55baeeb2847920387.jpg) +"Remove all degradation" + +![](images/4bc8cc66a65671967468e08441026a8e939bb4cc2d52cbf068519985e6bd32f1.jpg) +"Remove all degradation" + +![](images/f3ba838f50d8cd2753e78188424782a28daf0f479aaa2090f3c5953b64b5f94b.jpg) +"Upsample...denoise... + +![](images/5477461e3917d0f66a58645f62be7077c34bff1d2c656e28979dab7e1fad73cc.jpg) +'Upsample...denoise...' + +![](images/db2f022c99661c46e9278bfff823cadb716c1574182b3f6727dfd83db7fd181e.jpg) +Ground Truth + +ample..., denoise... + +![](images/533ae31e3b666ae6a20b4cf74980c90237eeedca88d214a3ca7301384534a786.jpg) +Input + +![](images/670b4166754aa4503b6d0466d2dc19929cac7f1221541ced98ca7cb42ba4404f.jpg) +"Remove all degradation" + +![](images/51c8ef5cde5d49efe49a1ca7f128b92711040cc67e294d5d7a1e15775ab4465c.jpg) +"Remove all degradation" + +![](images/480ac784c13b5c29093dd41675616b439a9a33ac49ea4c461b48707fa3929693.jpg) +"Deblur...denoise..." + +![](images/39176c35dc533ef2e0bd694774883084084ea39412df211341e344ab9ad30bf5.jpg) +"Deblur...denoise..." + +![](images/bb9ac5602ac7ac436890df22ffb4eebed637e32e13124a80ab3350f8910b8847.jpg) +Ground Truth + +![](images/b6cf43e0d633afc09beefad697e1af572861e45414f02e78d1ba9297b28a7c97.jpg) +Input + +![](images/7cb67d3420abd0d8021549e8b2a23d5d84dd3bcf0b200d8c5ceb2a96281553b5.jpg) +$\because {a}_{4} = {2}^{3} \cdot {2}^{2}$ + +![](images/888052140f6dbc026b92c03600d9936ef2673866e50dcddc4fe7e4c1b08cf5ff.jpg) +"DANGER..." + +![](images/fc2d2f3fe4cfc2197285d8c9ea412eb69693c85971f9eeb31d332b6ff5e44f02.jpg) +"EXIT..." +Fig. 1: We present SPIRE: Semantic Prompt-Driven Image Restoration, a text-based foundation model for all-in-one, instructed image restoration. SPIRE allows users to flexibly leverage either semantic-level content prompt, or quantitative degradation-aware restoration prompt, or both, to obtain their desired enhancement results based on personal preferences. In other words, SPIRE can be easily prompted to conduct blind restoration, semantic restoration, or task-specific granular treatment. Our framework also enables a new paradigm of instruction-based image restoration, providing a reliable evaluation benchmark to facilitate vision-language models for low-level computational photography applications. + +![](images/807ff203776fab889a2b54635f0969f4016753ae9231555252fe55568a52c377.jpg) +"GO..." +"STOP... + +Abstract. Text-driven diffusion models have become increasingly popular for various image editing tasks, including inpainting, stylization, and object replacement. However, it still remains an open research problem to adopt this language-vision paradigm for more fine-level image processing tasks, such as denoising, super-resolution, deblurring, and compression artifact removal. In this paper, we develop SPIRE, a Semantic and restoration Prompt-driven Image Restoration framework that leverages + +natural language as a user-friendly interface to control the image restoration process. We consider the capacity of prompt information in two dimensions. First, we use content-related prompts to enhance the semantic alignment, effectively alleviating identity ambiguity in the restoration outcomes. Second, our approach is the first framework that supports fine-level instruction through language-based quantitative specification of the restoration strength, without the need for explicit task-specific design. In addition, we introduce a novel fusion mechanism that augments the existing ControlNet architecture by learning to rescale the generative prior, thereby achieving better restoration fidelity. Our extensive experiments demonstrate the superior restoration performance of SPIRE compared to the state of the arts, alongside offering the flexibility of text-based control over the restoration effects. + +# 1 Introduction + +Image restoration or enhancement aims to recover high-quality, pixel-level details from a degraded image, while preserving as much the original semantic information as possible. Although neural network models [11, 12, 30, 63, 67, 71, 73, 74] have marked significant progress, it still remains challenging to design an effective task-conditioning mechanism instead of building multiple individual models for each task (such as denoise, deblur, compression artifact removal) in practical scenarios. The advancement of text-driven diffusion models [48, 50, 53] has unveiled the potential of natural language as a universal input condition for a broad range of image processing challenges, which improves interactivity with users and reduces the cost of task-specific fine-tuning. However, the existing applications of natural language in stylization [38, 44], and inpainting [3, 4, 10, 21, 44, 64] predominantly focus on high-level semantic editing, whereas the uniqueness and challenges for low-level image processing have been less explored. + +Natural language text prompts in image restoration can play two crucial roles: alleviating semantic ambiguities and resolving degradation type ambiguities. Firstly, the same degraded image can correspond to different visual objects, leading to ambiguities in content interpretation (e.g., discerning whether the blurred animal in Fig. 1 is a horse or a zebra). Typically, unconditional image-to-image restoration leads to a random or average estimation of the identities, resulting in neither fish nor fowl. Secondly, certain photographic effects that are deliberately introduced for aesthetic purposes, such as bokeh with a soft out-of-focus background, can be misconstrued as blur distortions by many existing deblurring models, leading to unwanted artifacts or severe hallucinations in the outputs. Although blind restoration methods can produce clean images by either leveraging frozen generative priors [67,70,77], or using end-to-end regression [71], they do not consider the aforementioned ambiguities. + +In this paper, we introduce SPIRE—a Semantic Prompt-Driven Image REstoration framework that provides a user-friendly interface to fully control both the restored image semantics and the enhancement granularity using natural language instructions. Traditional blind restoration without prompts (Fig + +ure 1) has limited flexibility in solving semantic and degradation ambiguity, thus tends to generate average blurry images. Leveraging the semantic and quantitative restoration prompts, we show that the visual quality can be significantly improved. Moreover, our framework can also be used as blind restoration with null prompts. In addition to higher perceptual quality, our framework provides the flexibility for users to generate more than one plausible result. Specifically, in imaging scenarios with large levels of degradation there are multiple plausible solutions. The proposed interactive framework has the capability to personalize the output by leveraging input prompts according to user preference. In concurrent studies, the focus has typically been isolated to one of three areas: either blind restoration [34,67], only semantic prompting [78], or only discrete restoration types [9,36]. To the best of our knowledge, this is the first unified model to support the following three distinct features simultaneously: + +1. Blind Restoration: When instructed with a general restoration prompt "remove all degradations" and empty semantic prompt "", SPIRE operates as a blind restoration model ("Ours w/o text" in Tab. 1). +2. Semantic Restoration: When provided with a text description of the (desired) visual content, SPIRE concentrates on restoring the specified identity of the uncertain or ambiguous objects in the degraded image. +3. Quantitative Task-Specific Restoration: Receiving specific restoration type hints (e.g., "deblur...", "deblur... denoise..."), SPIRE transforms into a task-specific model (e.g., deblur, denoise, or both). Moreover, it understands the numeric nuances of different degradation parameters in language prompts so that the users can control the restoration strength (e.g., "deblur with sigma 3.0" in Fig. 7). + +To train SPIRE, we first build a synthetic data generation pipeline on a large-scale text-image dataset [8] upon the second-order degradation process proposed in Real-ESRGAN [71]. Additionally, we embed the degradation parameters into the restoration prompts to encode finer-grained degradation information (e.g., "deblur with sigma 3.5"). Our approach provides richer degradation-specific information compared to contemporary works [26,36] which only employ the degradation types (e.g., "gaussian blur"). We then finetune a ControlNet adaptor [84]—which learns a parameter-efficient, parallel branch on top of the latent diffusion models (LDMs) [50]—on a mixture of diverse restoration tasks (see Fig. 1). The semantic text prompts are processed through the LDM backbone, aligning with LDM's text-to-image pretraining. The degradation text prompts and image-conditioning are implemented in the ControlNet branch, as these aspects are not inherently addressed by the vanilla LDM. We further improve the ControlNet by introducing a new modulation connection that adaptively fuses the degradation condition with the generative prior with only a few extra trainable parameters yet showing impressive performance gain. + +Our extensive experiments demonstrate remarkable restoration quality achieved by our proposed SPIRE method. Moreover, SPIRE offers additional freedom for both content and restoration prompt engineering. For example, the semantic + +prompt "a very large giraffe eating leaves" helps to resolve the semantic ambiguity in Fig. 2; the degradation prompt "deblur with sigma 3.0" reduces the gaussian blur while maintaining the intentional motion blur in Fig. 3. We also find out that our model learns a continuous latent space of the restoration strength in Fig. 7, even though we do not explicitly fine-tune the CLIP embedding. Our contribution can be summarized as follows: + +- We introduce the first unified text-driven image restoration model that supports both semantic prompts and restoration instructions. Our experiments demonstrate that incorporating semantic prompts and restoration instructions significantly enhances the restoration quality. +- Our proposed paradigm empowers users to fully control the semantic outcome of the restored image using different semantic prompts during test time. +- Our proposed approach provides a mechanism for users to adjust the category and strength of the restoration effect based on their subjective preferences. +- We demonstrate that text can serve as universal guidance control for low-level image restoration, eliminating the need for task-specific model design. + +# 2 Related Work + +First, we review the literature on image restoration. Our proposed framework aims to address some of the limitations of the existing restoration methods. Next, we compare multiple text-guided diffusion editing methods. Our method is designed to deal with two different text ambiguities concurrently, which is an unprecedented challenge for existing text-driven diffusion methods. Finally, we discuss the parameter-efficient fine-tuning of diffusion models, which further motivates our proposed modulation connection. + +Image Restoration is the task of recovering a high-resolution, clean image from a degraded input. Pioneering works in the fields of super-resolution [14,32], motion and defocus deblurring [1,61,79], denoising [82], and JPEG and artifact removal [16, 45] primarily utilize deep neural network architectures to address specific tasks. Later, Transformer-based [30,63,73] and adversarial-based [71] formulations were explored, demonstrating state-of-the-art performance with unified model architecture. Recently there has been a focus on exploiting iterative restorations, such as the ones from diffusion models, to generate images with much higher perceptual quality [11,49,52,54,74,75,78]. + +There have been some recent works attempting to control the enhancement strength of a single model, such as the scaling factor condition in arbitrary super-resolution [68, 69] and noise-level map for image denoising [83]. However, these methods require hand-crafting dedicated architectures for each task, which limits their scalability. In contrast to the task-specific models, other approaches seek to train a single blind model by building a pipeline composed of several classical degradation process, such as BSRGAN [80] and Real-ESRGAN [71]. + +Text-guided Diffusion Editing. Denoising diffusion models are trained to reverse the noisiging process [23, 57, 58]. Early methods focused on unconditioned generation [13], but recent trends pivot more to conditioned generation, such as + +![](images/271b8020dfdd93e0071501f87af5c4adb0823db53ee96ea6e256577234cfe22d.jpg) +Fig. 2: Framework of SPIRE. In the training phase, we begin by synthesizing a degraded version $y$ , of a clean image $x$ . Our degradation synthesis pipeline also creates a restoration prompt $c_{r}$ , which contains numeric parameters that reflects the intensity of the degradation introduced. Then, we inject the synthetic restoration prompt into a ControlNet adaptor, which uses our proposed modulation fusion blocks $(\gamma, \beta)$ to connect with the frozen backbone driven by the semantic prompt $c_{s}$ . During test time, the users can employ the SPIRE framework as either a blind restoration model with restoration prompt "Remove all degradation" and empty semantic prompt $\varnothing$ , or manually adjust the restoration $c_{r}$ and semantic prompts $c_{s}$ to obtain what they want. + +image-to-image translation [52, 55] and text-to-image generation [50, 53]. Latent Diffusion [50] is a groundbreaking approach that introduces a versatile framework for improving both training and sampling efficiency, while flexible enough for general conditioning inputs. Subsequent works have built upon their pretrained text-to-image checkpoints, and designed customized architectures for different tasks like text-guided image editing. For instance, SDEdit [37] generates content for a new prompt by adding noise to the input image. Attention-based methods [21, 44, 64, 81] show that images can be edited via reweighting and replacing the cross-attention map of different prompts. Aside from the previously mentioned approaches driven by target prompts, instruction-based editing [7, 17] entail modifying a source image based on specific instructions. These textual instructions are typically synthesized using large language models [41, 47]. As CLIP serves as a cornerstone for bridging vision and language, several studies have extensively investigated the representation space of CLIP embeddings [46]. Instead of directly applying CLIP to fundamental discriminative tasks [19, 85], these works either tailor CLIP to specific applications, such as image quality assessment [28, 31], serve as semantic guidance in generative adversarial network [5] or aligning image embedding with degradation types in the feature space [26, 36]. To enhance CLIP's ability to understand numerical information, some works [42, 43] finetune it using contrastive learning on synthetic numerical data, and then incorporate the fine-tuned text embedding into diffusion training. + +Parameter-Efficient Diffusion Model Finetuning. To leverage the powerful generative prior in pretrained diffusion models, parameter-efficient components such as text embedding [15], low-rank approximations of model weights [20, 25], and cross attention layers [29] can be finetuned to personalize the pretrained model. Adaptor-based finetuning paradigms [39, 84] propose to keep the original + +UNet weights frozen, and add new image or text conditions. This adaptor generates residual features that are subsequently added to the frozen UNet backbone. + +# 3 Method + +We introduce a universal approach to combine the above mentioned task-specific, strength-aware, and blind restoration methods within a unified framework (illustrated in Sec. 3.2). We further propose to decouple the learning of content and restoration prompts to better preserve the pre-trained prior while injecting new conditions, as unfolded in Sec. 3.3. Sec. 3.4 details our design on how to accurately control the restoration type and strength, as well as our proposed modulation fusion layer that adaptively fuses the restoration features back to the frozen backbone. We start with some preliminaries in the following section. + +# 3.1 Preliminaries + +Latent Diffusion Models (LDMs) [50] are probabilistic generative models that learn the underlying data distribution by iteratively removing Gaussian noise in the latent space, which is typically learned using a VAE autoencoder. Formally, the VAE encoder $\mathcal{E}$ compresses an input image $\pmb{x}$ into a compact latent representation $\pmb{z} = \mathcal{E}(\pmb{x})$ , which can be later decoded back to the pixel space using the coupled VAE decoder $\mathcal{D}$ , often learned under an image reconstruction objective: $\mathcal{D}(\mathcal{E}(\pmb{x})) \approx \pmb{x}$ . During training stage, the output of a UNet [51] $\epsilon_{\theta}(z_t, t, \pmb{y})$ conditioned on $\pmb{y}$ (such as text, images, etc.) is parameterized [13, 56] to remove Gaussian noise $\pmb{\epsilon}$ in the latent space $\pmb{z}_t$ as: + +$$ +\min _ {\theta} \mathbb {E} _ {\left(\boldsymbol {z} _ {0}, \boldsymbol {y}\right) \sim p _ {\text {d a t a}}, \epsilon \sim \mathcal {N} (0, I), t} \| \boldsymbol {\epsilon} - \boldsymbol {\epsilon} _ {\theta} \left(\boldsymbol {z} _ {t}, t, \boldsymbol {y}\right) \| _ {2} ^ {2}, \tag {1} +$$ + +where $\mathbf{z}_t$ is a noisy sample of $\mathbf{z}_0$ at sampled timestep $t$ . The condition $\mathbf{y}$ is randomly dropped out as $\varnothing$ to make the model unconditional. + +At test time, deterministic DDIM sampling [57] is utilized to convert a random noise $z_{T}$ to a clean latent $z_{0}$ which is decoded as final result $\mathcal{D}(z_0)$ . In each timestep $t$ , classifier-free guidance [13, 24] can be applied to trade-off sample quality and condition alignment: + +$$ +\overline {{\epsilon}} _ {\theta} \left(\boldsymbol {z} _ {t}, t, \boldsymbol {y}\right) = \epsilon_ {\theta} \left(\boldsymbol {z} _ {t}, t, \varnothing\right) + w \left(\epsilon_ {\theta} \left(\boldsymbol {z} _ {t}, t, \boldsymbol {y}\right) - \epsilon_ {\theta} \left(\boldsymbol {z} _ {t}, t, \varnothing\right)\right), +$$ + +where $w$ is a scalar to adjust the guidance strength of $y$ . Note that the estimated noise $\overline{\epsilon}_{\theta}(\pmb{z}_t, t, \pmb{y})$ is used to update $z_t$ to $z_{t-1}$ , which is a approximation of the distribution gradient score [60] as $\nabla_{z_t} \log p(\pmb{z}_t | \pmb{y}) \propto \overline{\epsilon}_{\theta}(\pmb{z}_t, t, \pmb{y})$ . + +# 3.2 Text-driven Image Restoration + +Based on the LDM framework, we propose a new restoration paradigm—text-driven image restoration. Our method targets to restore images $(\pmb{x}$ or $\pmb{z}_0)$ based + +on conditions $\{\pmb{y},\pmb{c}_s,\pmb{c}_r\}$ . Specifically: $\pmb{y}$ denotes the degraded image condition, $\pmb{c}_s$ is the semantic prompt describing the clean image $\pmb{x}$ (e.g., "a panda is sitting by the bamboo" or "a panda"), and $\pmb{c}_r$ is the restoration prompt that describes the details of the degradation in terms of both the operation and parameter (e.g., "deblur with sigma 3.0"). We use $\pmb{y} = \mathrm{Deg}(\pmb{x},\pmb{c}_r)$ to denote the degradation process which turns the clean image $\pmb{x}$ into its degraded counterpart $\pmb{y}_i$ . + +The above text-driven image restoration model $p(\pmb{z}_t|\{\pmb{y},\pmb{c}_s,\pmb{c}_r\})$ can be trained using paired data. We use the text-image dataset [8] so that each clean image $\pmb{x}$ is paired with a semantic prompt $\pmb{c}_s$ (the paired alt-text). Then $\pmb{y} = \mathrm{Deg}(\pmb{x},\pmb{c}_r)$ is simulated using the synthetic degradation [71] pipeline (Sec. 4.1), yielding the final paired training data $(\pmb{x} \text{ or } \pmb{z}_0, \{\pmb{y},\pmb{c}_s,\pmb{c}_r\}) \sim p_{data}$ . + +Comparison to Blind Restoration. Recent existing blind restoration models [34, 55, 67] also leverage diffusion priors to generate high-quality images. However, these methods are prone to hallucinate unwanted, unnatural, or oversharpened textures given that the semantic- and degradation-ambiguities persist everywhere. Fig. 3 provides an example, where the input image is degraded simultaneously with Gaussian blur and motion blur. The fully blind restoration method removes all of the degradation, leaving no way to steer the model to remove only the Gaussian blur while preserving the motion effect for aesthetic purposes. Therefore, having the flexibility to control the restoration behavior and its strength at test time is a crucial requirement. Our text-driven formulation of the problem introduces additional controlling ability using both the semantic content and the restoration instruction. In our SPIRE design, $\pmb{c}_r =$ "Remove all degradation" is also a plausible restoration prompt which makes our SPIRE to be compatible with the blind restoration. Further, our text-driven restoration takes advantages of the pretrained content-aware LDM using the additional semantic prompt $\pmb{c}_s$ , hence able to better handle semantic ambiguities on the noisy, blurry regions whose visual objects are less recognizable. + +# 3.3 Decoupling Semantic and Restoration Prompts + +To effectively learn the latent distribution $p(\pmb{z}_t | \pmb{y}, \pmb{c}_s, \pmb{c}_r)$ , we further decouple the conditions $\{\pmb{y}, \pmb{c}_s, \pmb{c}_r\}$ into two groups—one for the text-to-image prior $(\pmb{c}_s \rightarrow \pmb{z}_t)$ already imbued in the pretrained LDM model, and the other for both the image-to-image $(\pmb{y} \rightarrow \pmb{z}_t)$ and restoration-to-image $(\pmb{c}_r \rightarrow \pmb{z}_t)$ that needs to be learnt from the synthetic data. This decoupling strategy prevents catastrophic forgetting in the pretrained diffusion model and enables independent training of the restoration-aware model, whose score function [60] is expressed as: + +$$ +\nabla_ {z _ {t}} \log p (\boldsymbol {z} _ {t} | \boldsymbol {y}, \boldsymbol {c} _ {s}, \boldsymbol {c} _ {r}) \approx \underbrace {\nabla_ {z _ {t}} \log p (\boldsymbol {z} _ {t} | \boldsymbol {c} _ {s})} _ {\text {S e m a n t i c - a w a r e (f r o z e n)}} + \underbrace {\nabla_ {z _ {t}} \log p (\boldsymbol {y} | \boldsymbol {z} _ {t} , \boldsymbol {c} _ {r})} _ {\text {R e s t o r a t i o n - a w a r e (l e a r n a b l e)}}. \tag {2} +$$ + +In the above equation, the first part $\nabla_{\boldsymbol{z}_t}\log p(\boldsymbol{z}_t|\boldsymbol{c}_s)$ aligns with the text-to-image prior inherent in the pretrained LDM. The second term $\nabla_{\boldsymbol{z}_t}\log p(\boldsymbol{y}|\boldsymbol{z}_t,\boldsymbol{c}_r)$ approximates the consistency constraint with degraded image $\boldsymbol{y}$ , meaning the latent image $\boldsymbol{z}_t$ should be inferred by the degraded image $\boldsymbol{y}$ and the degradation + +details $\pmb{c}_r$ . This is somewhat similar to the reverse process of $\pmb{y} = \mathrm{Deg}(\pmb{x}_t, \pmb{c}_r)$ in the pixel space. More derivation is provided in our supplement. Providing degradation information through a restoration text prompt $\pmb{c}_r$ can largely reduce the uncertainty of estimation $p(\pmb{y} | \pmb{z}_t, \pmb{c}_r)$ , thereby alleviating the ambiguity of the restoration task and leading to better results that are less prone to hallucinations. + +# 3.4 Learning to Control the Restoration. + +Inspired by ControlNet [84], we employ an adaptor to model the restoration-aware branch in Eq. (2). First, the pretrained latent diffusion UNet $\epsilon_{\theta}(\pmb{z}_t,t,\pmb{c}_s)$ is frozen during training to preserve text-to-image prior $\nabla_{\pmb{z}_t}\log p(\pmb{z}_t|\pmb{c}_s)$ . Secondly, an encoder initialized from pretrained UNet is finetuned to learn the restoration condi + +![](images/cd4cdd9227e8c76a2fe883bd0d71f902ce64ed16a21d140fb49458a3d8c0dad6.jpg) +Fig. 3: Degradation ambiguities in real-world problems. By adjusting the restoration prompt, our method can preserve the motion effect that is coupled with Gaussian blur, while fully blind restoration models do not provide this flexibility. + +tion $\nabla_{z_t}\log p(\pmb {y}|\pmb {z}_t,\pmb {c}_r)$ , which takes degradation image $\pmb{y}$ , current noisy latents $\pmb{z}_{t}$ and the restoration prompt $\pmb{c}_{r}$ as input. The output of encoder is fusion features $\pmb{f}_{con}$ for frozen UNet. In this way, the choice of ControlNet implementation follows our decoupling analysis in Eq. (2). + +Conditioning on the Degraded Image $\pmb{y}$ . We apply downsample layers on $\pmb{y}$ to make the feature map matching to the shape of the $\pmb{z}_t$ , denoted as $\mathcal{E}'(\pmb{y})$ . $\mathcal{E}'$ is more compact [84] than VAE encoder $\mathcal{E}$ , and its trainable parameters allow model to perceive and adapt to the degraded images from our synthetic data pipeline. Then, we concatenate the downsampled feature map $\mathcal{E}'(\pmb{y})$ with the latent $\pmb{z}_t$ , and feed them to the ControlNet encoder. + +Conditioning on the Restoration Prompt $\pmb{c}_r$ . Following [50], we first use the vision-language CLIP [46] to infer the text sequence embedding $\pmb{e}_r \in \mathbb{R}^{M \times d}$ for $\pmb{c}_r$ , where $M$ is the number of tokens, and $d = 768$ is the dimension of the image-text features. Specifically, given a restoration instruction "deblur with sigma 3.0", CLIP first tokenizes it into a sequence of tokens (e.g., "3.0" is tokenized into ["3", ".", "0)]) and embeds them into $M$ distinct token embeddings. Then, stacked causal transformer layers [65] are applied on the embedding sequence to contextualize the token embeddings, resulting in $\pmb{e}_r$ . Hence, tokens such as "deblur" may arm with the numeric information such as "3", ".", and "0", making it possible for the model to distinguish the semantic difference of strength definition in different restoration tasks. Cross-attention process ensures that the information from $\pmb{e}_r$ is propagated to the ControlNet feature $f_{con}$ . While [42, 43] finetuned CLIP using contrastive learning, while we found frozen CLIP still works. The reason is that the learnable cross-attention layers can somehow ensemble the observations from frozen CLIP and squeeze useful information from it. + +Modulation Fusion Layer. The decoupling of learning two branches (Sec. 3.3), despite offering the benefit of effective separation of prior and conditioning, + +
MethodPromptsParameterized Degradation with synthesized crReal-ESRGAN Degradation without cr
Sem ResFID↓LPIPS↓PSNR↑SSIM↑CLIP-I↑CLIP-T↑FID↓LPIPS↓PSNR↑SSIM↑CLIP-I↑CLIP-T↑
SwinIR [30]XX43.220.42324.400.7170.8560.28548.370.44923.450.6990.8420.284
StableSR [67]XX20.550.31321.030.6130.8860.29825.750.36420.420.5810.8640.298
DiffBIR [34]XX17.260.30222.160.6040.9120.29719.170.33021.480.5870.8980.298
ControlNet-SR [84]XX13.650.22223.750.6690.9380.30016.990.26922.950.6280.9240.299
Ours w/o textXX12.700.22123.840.6710.9390.29916.250.26223.150.6360.9290.300
DiffBIR [34] + SDEdit [37]✓X19.360.36219.390.5270.8910.30517.510.37519.150.5210.8870.308
DiffBIR [34] + CLIP [46]✓X18.460.36520.500.5260.8960.30820.310.37420.450.5390.8850.307
ControlNet-SR + CLIP [46]✓X13.000.24123.180.6480.9370.30715.160.28622.450.6100.9260.308
Ours✓✓11.340.21923.610.6650.9430.30614.420.26223.140.6330.9350.308
+ +Table 1: Quantitative results on the MS-COCO dataset (with $c_{s}$ ) using our parameterized degradation (left) and Real-ESRGAN degradation (right). We also denote the prompt choice at test time. 'Sem' stands for semantic prompt; 'Res' stands for restoration prompt. The first group of baselines are tested without prompt. The second group are combined with semantic prompt in zero-shot way. + +
Degradation ProcessParameterizedReal-ESRGAN
p(choose)Restoration PromptDegradation Processp(choose)Restoration Prompt
Gaussian Blur0.5Deblur with {sigma∈[0.2, 3.0]} or DeblurBlur1.0
Downsample0.5Upsample to {resizing factor∈[1.0, 7.0]} or UpsampleResize1.0Remove all degradation
Gaussian Noise0.5Denoise with {sigma∈[0.0, 0.12]} or DenoiseNoise1.0
JPEG0.5Dejpeg with quality {quality factor∈[30, 92]} or DejpegJPEG1.0
+ +Table 2: Our training degradation is randomly sampled in these two pipeline with $50\%$ each. (1) Images generated by our parameterized pipeline are paired with either a restoration type (e.g., "Deblur") or a restoration parameter prompt (e.g., "Deblur with sigma 0.3.") (2) In other $50\%$ iterations, degraded images $\pmb{y}$ synthesized by Real-ESRGAN are paired with the same restoration prompt $\pmb{c}_r = \text{"Remove all degradation"}$ + +may cause distribution shift in the learned representations between the frozen backbone feature $\pmb{f}_{\mathrm{skip}}$ and ControlNet feature $\pmb{f}_{\mathrm{control}}$ . To allow for adaptive alignment of the above features, here we propose a new modulation fusion layer to fuse the multi-scale features from ControlNet to the skip connection features of the frozen backbone: $\hat{\pmb{f}}_{\mathrm{skip}} = (1 + \gamma)\pmb{f}_{\mathrm{skip}} + \beta$ ; $\gamma, \beta = \mathcal{M}(\pmb{f}_{\mathrm{con}})$ , where $\gamma, \beta$ are the scaling and bias parameters from $\mathcal{M}$ a lightweight $1\times 1$ zero-initialized convolution layer. + +# 4 Experiments + +# 4.1 Text-based Training Data and Benchmarks + +Our method opens up an entirely new paradigm of instruction-based image restoration. However, existing image restoration datasets such as DIV2K [2] and Flickr2K [40] do not provide high-quality semantic prompts, and Real-ESRGAN degradation is for blind restoration. To address this, we construct a new setting including training data generation and test benchmarks for evaluation. + +Our parameterized degradation pipeline is based on the Real-ESRGAN [71] which contains a large number of degradation types. Since the original full Real-ESRGAN degradation is difficult to be parameterized and represented as user-friendly natural language (e.g., not every one understands "anisotropic plateau blur with beta $2.0^{\prime \prime}$ ), we choose the 4 most general degradations which are practical for parameterization to support degradation parameter-aware restoration, as shown in Tab. 2. Our parameterized pipeline skips each degradation stage + +![](images/66e49157111f2c87df01cd69bcc0e613dd7448addd03ada21cab04e71c31a9f9.jpg) +Fig. 4: Visual Comparison with other baselines. Our method of integrating both the semantic prompt $\pmb{c}_{s}$ and the restoration prompt $\pmb{c}_{r}$ outperforms image-to-image restoration (DiffBIR, Retrained ControlNet-SR) and naive zero-shot combination with semantic prompt. It achieves more sharp and clean results while maintaining consistency with the degraded image. + +with a $50\%$ probability to increase the diversity of restoration prompts and support single task restoration. The descriptions of the selected degradations are appended following the order of the image degradations to synthesize $\pmb{c}_r$ . In real application, users can control the restoration type and its strength by modifying restoration prompt. As shown in third column of Fig. 3, driven by restoration prompt "Deblur with sigma $3.0^{\prime \prime}$ , our framework removes the Gaussian blur, while preserving motion blur. The representation of the restoration strength is also general enough to decouple different tasks in natural language as shown in Fig. 7. More visual results are presented in our supplement. + +Training data construction. In the training stage, we sample near 100 million text-image pairs from internal data source and use the alternative text label with highest relevance as the semantic prompt $\pmb{c}_s$ in the framework. We drop out the semantic prompt to empty string $\varnothing$ by a probability of $10\%$ to support semantic-agnostic restoration. We mix the Real-ESRGAN degradation pipeline and our parameterized pipeline by $50\%$ each in training iterations. + +Evaluation Setting. We randomly choose 3000 pairs of image $\pmb{x}$ and semantic prompt $\pmb{c}_{s}$ from the MS-COCO [33] validation set. Then, we build two test sets on parameterized and the Real-ESRGAN processes separately. At the left half of Tab. 1, images are sent to the parameterized degradation to form the synthesize degradation image $\pmb{y}$ with restoration prompt $\pmb{c}_{r}$ . In this setting, we expect "Ours" model to fully utilize the $\pmb{c}_{s}$ describing the semantic and the $\pmb{c}_{r}$ + +
MethodFID↓LPIPS↓PSNR↑SSIM↑CLIP-I↑
Real-ESRGAN [72]32.370.31222.520.6460.683
DiffBIR [34] (zero-shot)30.710.35422.010.5260.921
StableSR24.440.31121.620.5330.928
Ours w/o text (zero-shot)28.800.35221.680.5490.927
Ours w/o text (finetuned)22.450.32121.380.5320.932
+ +Table 3: Numerical results on the DIV2K testset without any prompt. + +describing the degradations. On the right half, the same images are fed to the Real-ESRGAN degradation—so the degradation is aligned with open-sourced checkpoints [30, 34, 67]. Because not every model can utilize both the semantic $(c_{s})$ and restoration $(c_{r})$ information, Tab. 1 provides a “Prompts” column denoting what information the model relied on. In summary, our constructed benchmarks based on MS-COCO can serve multiple purposes—from fully-blind, content-prompted, to all-prompted restoration. + +Evaluation Metrics. We use the generative metric FID [22] and perceptual metric LPIPS [86] for quantitative evaluation of image quality. PSNR, and SSIM are reported for reference. We also evaluate similarity scores in the clip embedding space with ground truth image (CLIP-I) and caption (CLIP-T). + +# 4.2 Comparing with baselines + +We compare our method with three categories of baselines: + +- Open-sourced image restoration checkpoints, including the regression model SwinIR [30] and the latent diffusion models StableSR [67] and DiffBIR [34]. +Combining the state-of-the-art method [34] with the zero-shot post-editing [37] (DiffBIR+SDEdit) or zero-shot injection through CLIP [46] (DiffBIR + CLIP). +- ControlNet-SR (retrained): We adapt the image-to-image ControlNet model [84] for super-resolution. For consistency, we maintain the same training iterations as our own model. We also evaluate its zero-shot capabilities with CLIP guidance (ControlNet-SR + CLIP). + +Quantitative comparison with baselines is presented in Tab. 1. On the left, we evaluate our full model with the parameterized degradation test set. Since open-sourced baselines are pretrained on Real-ESRGAN degradation, not perfectly aligned with our parameterized degradation, we also evaluate our full model on MS-COCO with Real-ESRGAN degradation by setting our restoration prompt to "Remove all degradation". Thanks to our prompts guidance and architecture improvements, our full model achieves best FID, LPIPS, CLIP-Image score, which means better image quality and semantic restoration. "Ours w/o text" is the same checkpoint conditioned on only degradation image and has high pixel-level similarity (the second best PSNR and SSIM) with GT. Although SwinIR has highest PSNR, its visual results are blurry. Although combining semantic prompt in zero-shot way can bring marginal improvement in FID and CLIP similarity with caption, it deteriorates the image restoration capacity and results in worse LPIPS and CLIP-Image. In contrast, semantic prompt guidance improves the CLIP image similarity of our full model. Tab. 3 shows the evaluation results of our model on DIV2K test set provided by StableSR. In zero-shot + +![](images/6a56b5abf87c9b569a850dcd4be82b8553d20c5dc8c202a19b17cd96a09623f0.jpg) +Fig. 5: Real-world image restorations. Our framework enhances real-world low-quality images following semantic prompts provided by language models or user input. + +![](images/05822db3db402b76fd6021faf22a65a78216020cd316b8037478e56b720c551d.jpg) + +![](images/6e844126cba331decf626d3ed980c7474cde6a85d75940a1b6a24bdf21dbd48b.jpg) + +![](images/3be6e31444462b912c1799d410d299cfd82cc11a88d9255a710182a071a9cfc1.jpg) + +![](images/92af50cf89b62bb95f8a749107424572e5ee301e75240dac4e9f8201fcda6fd2.jpg) + +![](images/3af1353d07f3adb4aa2500e3acd0378d57a4618fc6c1f6c4cc2aae72441eb47e.jpg) + +
MetricSwinIR [30]Diffbir [34]Ablation of restoration prompt crAblation of semantic prompt cs
Sem c_sXXXXshortunmatchedfull
Res crXX"Deblur 0.6"... 1.0"... 2.0"... 5.0""Remove all degradation"
CLIP-IQA↑0.61900.69830.40740.41590.48310.47970.58180.57500.64600.6610
MUSIQ↑63.6569.6935.8239.8749.9848.8161.6860.2767.5467.82
MANIQA↑0.40350.26190.22990.24870.29700.28210.36940.36370.43550.4474
+ +Table 4: Quantitative results on real-world images. Our aesthetic quality is improved when increasing restoration strength or using a more accurate and detailed prompt. + +test setting, our model has lower FID and LPIPS than DiffBIR. To fairly compare with StableSR, we finetune our model on their training set. Our finetuned model also outperforms StableSR in FID and CLIP-I. + +Qualitative comparison with baselines is presented in Fig. 4, where the corresponding semantic prompt is provided below each row. Image-to-image baselines such as DiffBir and the retrained ControlNet-SR can easily generate artifacts (e.g., hydrant in the first row) and blurry images (e.g., hands example in third row) that may look close to an "average" estimation. Besides, naive combination with semantic prompt in the zero-shot approach results in heavy hallucinations and may fail to preserve the object identity in the input image (e.g., giraffes on the second row). Unlike the existing methods, our full model considers semantic prompt, degradation image and restoration prompt in both training and test stages, which makes its results more aligned with all conditions. + +# 4.3 Real-world restoration + +Besides evaluation on synthetic degradations, we present an analysis on real-world images in Tab. 4 and Fig. 5. We choose RealPhoto60 [78] as our real-world test-set, and use non-reference metrics CLIP-IQA [66], MUSIQ [27] and MANIQA [76] scores to study the influence of the semantic and restoration prompts. First, we fix the semantic prompt to the empty string $\varnothing$ and conduct an ablation study on the restoration prompt. As the deblurring strength increases from 0.6 to 2.0, the image quality is improved consistently. A blind prompt "Remove all degradation" can provide better results than the manually controlled restoration prompts. This showcases that the proposed model is flexible to work with generic restoration prompts, especially when coming up with the perfect prompt is challenging. We also fix the blind restoration prompt and change the semantic prompts. Following [78], we feed degraded images to LLaVA [35] and generate corresponding semantic prompts. Our experiments + +![](images/daa0906e8b3bd0788da4ba27554dc6e6bac528f5d521d1c2e6758519bfe02e20.jpg) +Fig. 6: Test-time semantic prompting. Our framework restores degraded images guided by flexible semantic prompts, while unrelated background elements and global tones remain aligned with the degraded input conditioning. + +![](images/7de3d5beec6f58dec1bbbf7714b1fe10b6a26feadada030da3323936d31e7f47.jpg) +Fig.7: Prompt space walking visualization for the restoration prompt. Given the same degraded input (upper left) and empty semantic prompt $\varnothing$ , our method can decouple the restoration direction and strength via only prompting the quantitative number in natural language. An interesting finding is that our model learns a continuous range of restoration strength from discrete language tokens. + +show that detailed full descriptions (about 60 words) of the degraded image from LLaVA can be more effective to improve the visual results than short prompts that have less than 20 words. Unmatched prompts simulate the case when a user provides wrong prompts by randomly re-sampling unmatched long prompts from the data set. An interesting observation is that unmatched long prompts also improve the aesthetic quality of images against a short prompt or a blind restoration. Nonetheless, it can lead to more "hallucination" when the semantic input prompt is unmatched with the real ground truth. In contrast to concurrent work [62] that leverages implicit clip embedding, our text-based formulation provides better control of hallucinations by adjusting the LLM-synthesized or manually-designed prompt at test-time. Note that many studies [6,18] find current image restoration metrics are biased towards specific datasets and not fully aligned with human's visual perception. The quantitative results of previous works [34,71] are provided for reference. More results and discussions are provided in the supplements. + +
MethodSem csRes crFID↓LPIPS↓PSNR↑SSIM↑CLIP-I↑CLIP-T↑
ControlNet-SRXX13.650.22223.750.6690.9380.300
ControlNet-SR✓✓12.140.22323.500.6600.9400.306
Ours✓wo param11.580.22323.610.6650.9420.305
Ours✓✓11.340.21923.610.6650.9430.305
+ +Table 5: Ablation of architecture and degradation strength in ${c}_{r}$ + +# 4.4 Prompting the SPIRE + +Semantic Prompting. Our model can be driven by user-defined semantic prompts. As shown in Fig. 6, the blind restoration with empty string generates ambiguous and unrealistic identity. In contrast, our framework reconstructs sharp and realistic results. The object identity accurately follows user's semantic prompt, while the global layout and color tone remain consistent with input. + +Restoration Prompting. Users also have the freedom to adjust the degradation types and strengths in the restoration prompts. As shown in Fig. 7, the denoising and deblurring restoration are well decoupled. Using the task-specific prompts, our method can get a clean but blur (second row), or sharp but noisy (third row) result. In addition, our model learns the continuous space of restoration as the result becomes cleaner progressively if we modify the "Denoise with sigma 0.06" to "0.24". Beyond purely relying on numerical description, our model "understand" the difference of strength definition for tasks, such as the range difference of denoise sigma and deblur sigma, which makes it promising to be a universal representation for restoration strength control. + +# 4.5 Ablation Study + +We conduct several ablations in Tab. 5. Providing both prompts $c_s$ , $c_r$ to "ControlNet-SR" baseline improves the FID by 1.5. Adding the modulation fusion layer ("Ours") further improves the generative quality (0.8 for FID) and pixel-level similarity (0.11 dB). Embedding degradation parameters in the restoration prompt, not only enables restoration strength prompting (Fig. 7), but also improves image quality (FID and LPIPS). More details are in the supplement. + +# 5 Conclusion + +We have presented a unified text-driven framework for instructed image restoration using semantic and restoration prompts. We design our model in a decoupled way to better preserve the semantic text-to-image generative prior while efficiently learning to control both the restoration direction and its strength. To the best of our knowledge, this is the first framework to support both semantic and parameter-embedded restoration instructions simultaneously, allowing users to flexibly prompt-tune the results up to their expectations. Our extensive experiments have shown that our method significantly outperforms prior works in terms of both quantitative and qualitative results. Our model and evaluation benchmark have established a new paradigm of instruction-based image restoration, paving the way for further multi-modality generative imaging applications. + +# References + +1. Abuolaim, A., Delbracio, M., Kelly, D., Brown, M.S., Milanfar, P.: Learning to reduce defocus blur by realistically modeling dual-pixel data. In: ICCV. pp. 2289-2298 (2021) +2. Agustsson, E., Timofte, R.: NTIRE 2017 challenge on single image superresolution: Dataset and study. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition Workshops, CVPR Workshops 2017, Honolulu, HI, USA, July 21-26, 2017. pp. 1122-1131. IEEE Computer Society (2017). https://doi.org/10.1109/CVPRW.2017.150, https://doi.org/10.1109/CVPRW.2017.150 +3. Avrahami, O., Fried, O., Lischinski, D.: Blended latent diffusion. ACM Transactions on Graphics (2023) +4. Avrahami, O., Lischinski, D., Fried, O.: Blended diffusion for text-driven editing of natural images. In: CVPR. pp. 18208-18218 (2022) +5. Bai, Y., Wang, C., Xie, S., Dong, C., Yuan, C., Wang, Z.: Textir: A simple framework for text-based editable image restoration. arXiv preprint arXiv:2302.14736 (2023) +6. Blau, Y., Michaeli, T.: The perception-distortion tradeoff. In: CVPR. pp. 6228-6237 (2018) +7. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: CVPR (2023) +8. Chen, X., Wang, X., Changpinyo, S., Piergiovanni, A., Padlewski, P., Salz, D., Goodman, S.A., Grycner, A., Mustafa, B., Beyer, L., Kolesnikov, A., Puigcerver, J., Ding, N., Rong, K., Akbari, H., Mishra, G., Xue, L., Thapliyal, A., Bradbury, J., Kuo, W., Seyedhosseini, M., Jia, C., Ayan, B.K., Riquelme, C., Steiner, A., Angelova, A., Zhai, X., Houlsby, N., Soricut, R.: Pali: A jointly-scaled multilingual language-image model. In: ICLR (2023), https://arxiv.org/abs/2209.06794 +9. Chen, Z., Zhang, Y., Gu, J., Yuan, X., Kong, L., Chen, G., Yang, X.: Image superresolution with text prompt diffusion. arXiv preprint arXiv:2303.06373 (2023) +0. Couairon, G., Verbeek, J., Schwenk, H., Cord, M.: Diffedit: Diffusion-based semantic image editing with mask guidance. In: ICLR (2022) +1. Delbracio, M., Milanfar, P.: Inversion by direct iteration: An alternative to denoising diffusion for image restoration. Transactions on Machine Learning Research (2023), https://openreview.net/forum?id=VmyFF5lL3F, featured Certification +2. Delbracio, M., Talebei, H., Milanfar, P.: Projected distribution loss for image enhancement. In: 2021 IEEE International Conference on Computational Photography (ICCP). pp. 1-12. IEEE (2021) +3. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Neural Information Processing Systems (2021) +4. Dong, C., Loy, C.C., He, K., Tang, X.: Image super-resolution using deep convolutional networks. IEEE Trans. Pattern Anal. Mach. Intell. 38(2), 295-307 (2016) +5. Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A.H., Chechik, G., Cohen-Or, D.: An image is worth one word: Personalizing text-to-image generation using textual inversion. In: ICLR (2023) +6. Galteri, L., Seidenari, L., Bertini, M., Bimbo, A.: Deep generative adversarial compression artifact removal. In: ICCV (2017) +7. Geng, Z., Yang, B., Hang, T., Li, C., Gu, S., Zhang, T., Bao, J., Zhang, Z., Hu, H., Chen, D., Guo, B.: Instructdiffusion: A generalist modeling interface for vision tasks. CoRR abs/2309.03895 (2023). https://doi.org/10.48550/arXiv.2309.03895, https://doi.org/10.48550/arXiv.2309.03895 + +18. Gu, J., Cai, H., Dong, C., Ren, J.S., Timofte, R., Gong, Y., Lao, S., Shi, S., Wang, J., Yang, S., Wu, T., Xia, W., Yang, Y., Cao, M., Heng, C., Fu, L., Zhang, R., Zhang, Y., Wang, H., Song, H., Wang, J., Fan, H., Hou, X., Sun, M., Li, M., Zhao, K., Yuan, K., Kong, Z., Wu, M., Zheng, C., Conde, M.V., Burchi, M., Feng, L., Zhang, T., Li, Y., Xu, J., Wang, H., Liao, Y., Li, J., Xu, K., Sun, T., Xiong, Y., Keshari, A., Komal, Thakur, S., Jakhetiya, V., Subudhi, B.N., Yang, H.H., Chang, H.E., Huang, Z.K., Chen, W.T., Kuo, S.Y., Dutta, S., Das, S.D., Shah, N.A., Tiwari, A.K.: Ntire 2022 challenge on perceptual image quality assessment. In: CVPRW. pp. 951-967 (June 2022) +19. Gu, X., Lin, T.Y., Kuo, W., Cui, Y.: Open-vocabulary object detection via vision and language knowledge distillation. In: ICLR (2021) +20. Han, L., Li, Y., Zhang, H., Milanfar, P., Metaxas, D., Yang, F.: Svdiff: Compact parameter space for diffusion fine-tuning. In: ICCV. pp. 7323-7334 (October 2023) +21. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-or, D.: Prompt-to-prompt image editing with cross-attention control. In: ICLR (2023), https://openreview.net/forum?id=CDixzkzeyb +22. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: NeurIPS (2017) +23. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. NeurIPS 33, 6840-6851 (2020) +24. Ho, J., Salimans, T.: Classifier-free diffusion guidance. In: NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications (2021), https://openreview.net/forum?id=qw8AKxfYbI +25. Hu, E.J., yelong shen, Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-rank adaptation of large language models. In: ICLR (2022), https://openreview.net/forum?id=nZeVKeeFYf9 +26. Jiang, Y., Zhang, Z., Xue, T., Gu, J.: Autodir: Automatic all-in-one image restoration with latent diffusion. arXiv preprint arXiv:2310.10123 (2023) +27. Ke, J., Wang, Q., Wang, Y., Milanfar, P., Yang, F.: Musiq: Multi-scale image quality transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5148-5157 (2021) +28. Ke, J., Ye, K., Yu, J., Wu, Y., Milanfar, P., Yang, F.: Vila: Learning image aesthetics from user comments with vision-language pretraining. In: CVPR. pp. 10041-10051 (2023) +29. Kumari, N., Zhang, B., Zhang, R., Shechtman, E., Zhu, J.Y.: Multi-concept customization of text-to-image diffusion. In: CVPR. pp. 1931-1941 (2023) +30. Liang, J., Cao, J., Sun, G., Zhang, K., Gool, L.V., Timofte, R.: Swinir: Image restoration using swin transformer. In: Proceedings of ICCV Workshops (2021) +31. Liang, Z., Li, C., Zhou, S., Feng, R., Loy, C.C.: Iterative prompt learning for unsupervised backlit image enhancement. In: ICCV. pp. 8094-8103 (2023) +32. Lim, B., Son, S., Kim, H., Nah, S., Lee, K.M.: Enhanced deep residual networks for single image super-resolution. In: Proceedings of CVPR Workshops (2017) +33. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, L.: Microsoft coco: Common objects in context. In: ECCV. ECCV (September 2014), https://www.microsoft.com/en-us/research/publication/microsoft-coco-common-objects-in-context/ +34. Lin, X., He, J., Chen, Z., Lyu, Z., Fei, B., Dai, B., Ouyang, W., Qiao, Y., Dong, C.: Diffbir: Towards blind image restoration with generative diffusion prior. arXiv preprint arXiv:2308.15070 (2023) + +35. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: NeurIPS (2023) +36. Luo, Z., Gustafsson, F.K., Zhao, Z., Sjolund, J., Schon, T.B.: Controlling vision-language models for multi-task image restoration. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=t3vnnLeajU +37. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: SDEdit: Guided image synthesis and editing with stochastic differential equations. In: ICLR (2022) +38. Mokady, R., Hertz, A., Aberman, K., Pritch, Y., Cohen-Or, D.: Null-text inversion for editing real images using guided diffusion models. In: CVPR. pp. 6038-6047 (June 2023) +39. Mou, C., Wang, X., Xie, L., Zhang, J., Qi, Z., Shan, Y., Qie, X.: T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453 (2023) +40. Nah, S., Timofte, R., Baik, S., Hong, S., Moon, G., Son, S., Lee, K.M., Wang, X., Chan, K.C.K., Yu, K., Dong, C., Loy, C.C., Fan, Y., Yu, J., Liu, D., Huang, T.S., Sim, H., Kim, M., Park, D., Kim, J., Chun, S.Y., Haris, M., Shakhnarovich, G., Ukita, N., Zamir, S.W., Arora, A., Khan, S.H., Khan, F.S., Shao, L., Gupta, R.K., Chudasama, V.M., Patel, H., Upla, K.P., Fan, H., Li, G., Zhang, Y., Li, X., Zhang, W., He, Q., Purohit, K., Rajagopalan, A.N., Kim, J., Tofighi, M., Guo, T., Monga, V.: NTIRE 2019 challenge on video deblurring: Methods and results. In: IEEE Conference on Computer Vision and Pattern Recognition Workshops, CVPR Workshops 2019, Long Beach, CA, USA, June 16-20, 2019. pp. 1974-1984. Computer Vision Foundation / IEEE (2019). https://doi.org/10.1109/CVPRW.2019.00249, http://openaccess.thecvf.com/content_CVPRW_2019/html/NTIRE/Nah_NTIRE_2019_Challenge_on_Video_Deblurring_Methods_and_Results_CVPRW_2019_paper.html +41. OpenAI: Gpt-4 technical report (2023) +42. Paiss, R., Chefer, H., Wolf, L.: No token left behind: Explainability-aided image classification and generation. In: ECCV (2022) +43. Paiss, R., Ephrat, A., Tov, O., Zada, S., Mosseri, I., Irani, M., Dekel, T.: Teaching clip to count to ten. In: ICCV (2023) +44. Parmar, G., Kumar Singh, K., Zhang, R., Li, Y., Lu, J., Zhu, J.Y.: Zero-shot image-to-image translation. In: ACM SIGGRAPH 2023 Conference Proceedings. pp. 1-11 (2023) +45. Prakash, M., Delbracio, M., Milanfar, P., Jug, F.: Interpretable unsupervised diversity denoising and artefact removal. In: ICLR (2022), https://openreview.net/forum?id=DfMqlBOPXjM +46. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) +47. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners. OpenAI blog 1(8), 9 (2019) +48. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022) +49. Ren, M., Delbracio, M., Talebi, H., Gerig, G., Milanfar, P.: Multiscale structure guided diffusion for image deblurring. In: ICCV. pp. 10721-10733 (2023) +50. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10684-10695 (2022) + +51. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015) +52. Sahara, C., Chan, W., Chang, H., Lee, C., Ho, J., Salimans, T., Fleet, D., Norouzi, M.: Palette: Image-to-image diffusion models. In: ACM SIGGRAPH 2022 Conference Proceedings. pp. 1-10 (2022) +53. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS (2022) +54. Sahara, C., Ho, J., Chan, W., Salimans, T., Fleet, D.J., Norouzi, M.: Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(4), 4713-4726 (2022) +55. Sahara, C., Ho, J., Chan, W., Salimans, T., Fleet, D.J., Norouzi, M.: Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(4), 4713-4726 (2022) +56. Salimans, T., Ho, J.: Progressive distillation for fast sampling of diffusion models. In: ICLR. OpenReview.net (2022), https://openreview.net/forum?id=TIdIXIpzhoI +57. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020) +58. Song, Y., Ermon, S.: Generative modeling by estimating gradients of the data distribution. In: NeurIPS. pp. 11895-11907 (2019) +59. Song, Y., Shen, L., Xing, L., Ermon, S.: Solving inverse problems in medical imaging with score-based generative models. In: ICLR. OpenReview.net (2022) +60. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. In: ICLR (2021), https://openreview.net/forum?id=PxTIG12RRHS +61. Su, S., Delbracio, M., Wang, J., Sapiro, G., Heidrich, W., Wang, O.: Deep video deblurring for hand-held cameras. In: CVPR. pp. 1279-1288 (2017) +62. Sun, H., Li, W., Liu, J., Chen, H., Pei, R., Zou, X., Yan, Y., Yang, Y.: Coser: Bridging image and language for cognitive super-resolution. arXiv preprint arXiv:2311.16512 (2023) +63. Tu, Z., Talebi, H., Zhang, H., Yang, F., Milanfar, P., Bovik, A.C., Li, Y.: MAXIM: multi-axis MLP for image processing. In: CVPR (2022) +64. Tumanyan, N., Geyer, M., Bagon, S., Dekel, T.: Plug-and-play diffusion features for text-driven image-to-image translation. In: CVPR. pp. 1921-1930 (2023) +65. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. NeurIPS 30 (2017) +66. Wang, J., Chan, K.C., Loy, C.C.: Exploring clip for assessing the look and feel of images. In: AAAI (2023) +67. Wang, J., Yue, Z., Zhou, S., Chan, K.C., Loy, C.C.: Exploiting diffusion prior for real-world image super-resolution. arXiv preprint arXiv:2305.07015 (2023) +68. Wang, L., Wang, Y., Lin, Z., Yang, J., An, W., Guo, Y.: Learning a single network for scale-arbitrary super-resolution. In: ICCV. pp. 4801-4810 (2021) +69. Wang, X., Chen, X., Ni, B., Wang, H., Tong, Z., Liu, Y.: Deep arbitrary-scale image super-resolution via scale-equivariance pursuit. In: CVPR (2023) +70. Wang, X., Li, Y., Zhang, H., Shan, Y.: Towards real-world blind face restoration with generative facial prior. In: CVPR (2021) +71. Wang, X., Xie, L., Dong, C., Shan, Y.: Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In: ICCV. pp. 1905-1914 (2021) + +72. Wang, X., Yu, K., Wu, S., Gu, J., Liu, Y., Dong, C., Qiao, Y., Loy, C.C.: ESRGAN: enhanced super-resolution generative adversarial networks. In: Proceedings of ECCV Workshops (2018) +73. Wang, Z., Cun, X., Bao, J., Zhou, W., Liu, J., Li, H.: Uformer: A general u-shaped transformer for image restoration. In: CVPR (2022) +74. Whang, J., Delbracio, M., Talebi, H., Saharia, C., Dimakis, A.G., Milanfar, P.: Deblurring via stochastic refinement. In: CVPR. pp. 16293-16303 (2022) +75. Wu, R., Yang, T., Sun, L., Zhang, Z., Li, S., Zhang, L.: Seesr: Towards semantics-aware real-world image super-resolution. In: CVPR (2024) +76. Yang, S., Wu, T., Shi, S., Lao, S., Gong, Y., Cao, M., Wang, J., Yang, Y.: Maniaq: Multi-dimension attention network for no-reference image quality assessment. In: CVPR. pp. 1191-1200 (2022) +77. Yang, T., Ren, P., Xie, X., Zhang, L.: Gan prior embedded network for blind face restoration in the wild. In: CVPR (2021) +78. Yu, F., Gu, J., Li, Z., Hu, J., Kong, X., Wang, X., He, J., Qiao, Y., Dong, C.: Scaling up to excellence: Practicing model scaling for photo-realistic image restoration in the wild. In: CVPR (2024) +79. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Multi-stage progressive image restoration. In: CVPR (2021) +80. Zhang, K., Liang, J., Van Gool, L., Timofte, R.: Designing a practical degradation model for deep blind image super-resolution. In: ICCV. pp. 4791-4800 (2021) +81. Zhang, K., Mo, L., Chen, W., Sun, H., Su, Y.: Magicbrush: A manually annotated dataset for instruction-guided image editing. Advances in Neural Information Processing Systems 36 (2024) +82. Zhang, K., Zuo, W., Chen, Y., Meng, D., Zhang, L.: Beyond a Gaussian denoiser: Residual learning of deep CNN for image denoising. IEEE Transactions on Image Processing 26(7), 3142-3155 (2017) +83. Zhang, K., Zuo, W., Zhang, L.: Ffdnet: Toward a fast and flexible solution for CNN based image denoising. IEEE Transactions on Image Processing (2018) +84. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: ICCV (2023) +85. Zhang, R., Fang, R., Gao, P., Zhang, W., Li, K., Dai, J., Qiao, Y., Li, H.: Tip-adapter: Training-free clip-adapter for better vision-language modeling. In: ECCV (2022) +86. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018) \ No newline at end of file diff --git a/2024/SPIRE_ Semantic Prompt-Driven Image Restoration/images.zip b/2024/SPIRE_ Semantic Prompt-Driven Image Restoration/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..cc7ab7e8f76b8d9363a1f7aa5a1de937e8ada83c --- /dev/null +++ b/2024/SPIRE_ Semantic Prompt-Driven Image Restoration/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5d59b34965522896931d2abb0b6a87ba21973737830316c8d869642af153cde4 +size 677953 diff --git a/2024/SPIRE_ Semantic Prompt-Driven Image Restoration/layout.json b/2024/SPIRE_ Semantic Prompt-Driven Image Restoration/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..edc2c9d4e81118d25205498593d4b27541c692a2 --- /dev/null +++ b/2024/SPIRE_ Semantic Prompt-Driven Image Restoration/layout.json @@ -0,0 +1,12148 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 167, + 111, + 448, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 111, + 448, + 146 + ], + "spans": [ + { + "bbox": [ + 167, + 111, + 448, + 146 + ], + "type": "text", + "content": "SPIRE: Semantic Prompt-Driven Image Restoration" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 144, + 168, + 470, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 168, + 470, + 194 + ], + "spans": [ + { + "bbox": [ + 144, + 168, + 470, + 194 + ], + "type": "text", + "content": "Chenyang Qi" + }, + { + "bbox": [ + 144, + 168, + 470, + 194 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 144, + 168, + 470, + 194 + ], + "type": "text", + "content": ", Zhengzhong Tu" + }, + { + "bbox": [ + 144, + 168, + 470, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 144, + 168, + 470, + 194 + ], + "type": "text", + "content": ", Keren Ye" + }, + { + "bbox": [ + 144, + 168, + 470, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 144, + 168, + 470, + 194 + ], + "type": "text", + "content": ", Maurizio Delbracio" + }, + { + "bbox": [ + 144, + 168, + 470, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 144, + 168, + 470, + 194 + ], + "type": "text", + "content": ", Peyman Milanfar" + }, + { + "bbox": [ + 144, + 168, + 470, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 144, + 168, + 470, + 194 + ], + "type": "text", + "content": ", Qifeng Chen" + }, + { + "bbox": [ + 144, + 168, + 470, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 144, + 168, + 470, + 194 + ], + "type": "text", + "content": ", and Hossein Talebi" + }, + { + "bbox": [ + 144, + 168, + 470, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 267, + 201, + 345, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 267, + 201, + 345, + 224 + ], + "spans": [ + { + "bbox": [ + 267, + 201, + 345, + 224 + ], + "type": "text", + "content": "1 Google Research 2 HKUST" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 239, + 191, + 297 + ], + "blocks": [ + { + "bbox": [ + 133, + 239, + 191, + 297 + ], + "lines": [ + { + "bbox": [ + 133, + 239, + 191, + 297 + ], + "spans": [ + { + "bbox": [ + 133, + 239, + 191, + 297 + ], + "type": "image", + "image_path": "87bac0dffcf974577e47bdf79cb79aef0d83d7e7446a26a027cd557068a560f0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 154, + 299, + 167, + 306 + ], + "lines": [ + { + "bbox": [ + 154, + 299, + 167, + 306 + ], + "spans": [ + { + "bbox": [ + 154, + 299, + 167, + 306 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 192, + 239, + 249, + 297 + ], + "blocks": [ + { + "bbox": [ + 192, + 239, + 249, + 297 + ], + "lines": [ + { + "bbox": [ + 192, + 239, + 249, + 297 + ], + "spans": [ + { + "bbox": [ + 192, + 239, + 249, + 297 + ], + "type": "image", + "image_path": "d15147ed5a186f3928ae2c99ae961fecedfd982337ec7cb55baeeb2847920387.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 195, + 298, + 246, + 304 + ], + "lines": [ + { + "bbox": [ + 195, + 298, + 246, + 304 + ], + "spans": [ + { + "bbox": [ + 195, + 298, + 246, + 304 + ], + "type": "text", + "content": "\"Remove all degradation\"" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 249, + 239, + 305, + 297 + ], + "blocks": [ + { + "bbox": [ + 249, + 239, + 305, + 297 + ], + "lines": [ + { + "bbox": [ + 249, + 239, + 305, + 297 + ], + "spans": [ + { + "bbox": [ + 249, + 239, + 305, + 297 + ], + "type": "image", + "image_path": "4bc8cc66a65671967468e08441026a8e939bb4cc2d52cbf068519985e6bd32f1.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 252, + 298, + 302, + 304 + ], + "lines": [ + { + "bbox": [ + 252, + 298, + 302, + 304 + ], + "spans": [ + { + "bbox": [ + 252, + 298, + 302, + 304 + ], + "type": "text", + "content": "\"Remove all degradation\"" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 306, + 239, + 362, + 297 + ], + "blocks": [ + { + "bbox": [ + 306, + 239, + 362, + 297 + ], + "lines": [ + { + "bbox": [ + 306, + 239, + 362, + 297 + ], + "spans": [ + { + "bbox": [ + 306, + 239, + 362, + 297 + ], + "type": "image", + "image_path": "f3ba838f50d8cd2753e78188424782a28daf0f479aaa2090f3c5953b64b5f94b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 298, + 355, + 304 + ], + "lines": [ + { + "bbox": [ + 310, + 298, + 355, + 304 + ], + "spans": [ + { + "bbox": [ + 310, + 298, + 355, + 304 + ], + "type": "text", + "content": "\"Upsample...denoise..." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 362, + 239, + 419, + 297 + ], + "blocks": [ + { + "bbox": [ + 362, + 239, + 419, + 297 + ], + "lines": [ + { + "bbox": [ + 362, + 239, + 419, + 297 + ], + "spans": [ + { + "bbox": [ + 362, + 239, + 419, + 297 + ], + "type": "image", + "image_path": "5477461e3917d0f66a58645f62be7077c34bff1d2c656e28979dab7e1fad73cc.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 369, + 298, + 413, + 304 + ], + "lines": [ + { + "bbox": [ + 369, + 298, + 413, + 304 + ], + "spans": [ + { + "bbox": [ + 369, + 298, + 413, + 304 + ], + "type": "text", + "content": "'Upsample...denoise...'" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 420, + 239, + 477, + 297 + ], + "blocks": [ + { + "bbox": [ + 420, + 239, + 477, + 297 + ], + "lines": [ + { + "bbox": [ + 420, + 239, + 477, + 297 + ], + "spans": [ + { + "bbox": [ + 420, + 239, + 477, + 297 + ], + "type": "image", + "image_path": "db2f022c99661c46e9278bfff823cadb716c1574182b3f6727dfd83db7fd181e.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 433, + 299, + 462, + 306 + ], + "lines": [ + { + "bbox": [ + 433, + 299, + 462, + 306 + ], + "spans": [ + { + "bbox": [ + 433, + 299, + 462, + 306 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 383, + 298, + 413, + 304 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 298, + 413, + 304 + ], + "spans": [ + { + "bbox": [ + 383, + 298, + 413, + 304 + ], + "type": "text", + "content": "ample..., denoise..." + } + ] + } + ], + "index": 14, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 133, + 308, + 190, + 366 + ], + "blocks": [ + { + "bbox": [ + 133, + 308, + 190, + 366 + ], + "lines": [ + { + "bbox": [ + 133, + 308, + 190, + 366 + ], + "spans": [ + { + "bbox": [ + 133, + 308, + 190, + 366 + ], + "type": "image", + "image_path": "533ae31e3b666ae6a20b4cf74980c90237eeedca88d214a3ca7301384534a786.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 154, + 369, + 167, + 375 + ], + "lines": [ + { + "bbox": [ + 154, + 369, + 167, + 375 + ], + "spans": [ + { + "bbox": [ + 154, + 369, + 167, + 375 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 192, + 308, + 249, + 366 + ], + "blocks": [ + { + "bbox": [ + 192, + 308, + 249, + 366 + ], + "lines": [ + { + "bbox": [ + 192, + 308, + 249, + 366 + ], + "spans": [ + { + "bbox": [ + 192, + 308, + 249, + 366 + ], + "type": "image", + "image_path": "670b4166754aa4503b6d0466d2dc19929cac7f1221541ced98ca7cb42ba4404f.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 195, + 368, + 246, + 373 + ], + "lines": [ + { + "bbox": [ + 195, + 368, + 246, + 373 + ], + "spans": [ + { + "bbox": [ + 195, + 368, + 246, + 373 + ], + "type": "text", + "content": "\"Remove all degradation\"" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 249, + 308, + 305, + 366 + ], + "blocks": [ + { + "bbox": [ + 249, + 308, + 305, + 366 + ], + "lines": [ + { + "bbox": [ + 249, + 308, + 305, + 366 + ], + "spans": [ + { + "bbox": [ + 249, + 308, + 305, + 366 + ], + "type": "image", + "image_path": "51c8ef5cde5d49efe49a1ca7f128b92711040cc67e294d5d7a1e15775ab4465c.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 252, + 368, + 303, + 373 + ], + "lines": [ + { + "bbox": [ + 252, + 368, + 303, + 373 + ], + "spans": [ + { + "bbox": [ + 252, + 368, + 303, + 373 + ], + "type": "text", + "content": "\"Remove all degradation\"" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 306, + 308, + 362, + 366 + ], + "blocks": [ + { + "bbox": [ + 306, + 308, + 362, + 366 + ], + "lines": [ + { + "bbox": [ + 306, + 308, + 362, + 366 + ], + "spans": [ + { + "bbox": [ + 306, + 308, + 362, + 366 + ], + "type": "image", + "image_path": "480ac784c13b5c29093dd41675616b439a9a33ac49ea4c461b48707fa3929693.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 368, + 354, + 373 + ], + "lines": [ + { + "bbox": [ + 312, + 368, + 354, + 373 + ], + "spans": [ + { + "bbox": [ + 312, + 368, + 354, + 373 + ], + "type": "text", + "content": "\"Deblur...denoise...\"" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 362, + 308, + 419, + 366 + ], + "blocks": [ + { + "bbox": [ + 362, + 308, + 419, + 366 + ], + "lines": [ + { + "bbox": [ + 362, + 308, + 419, + 366 + ], + "spans": [ + { + "bbox": [ + 362, + 308, + 419, + 366 + ], + "type": "image", + "image_path": "39176c35dc533ef2e0bd694774883084084ea39412df211341e344ab9ad30bf5.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 370, + 368, + 412, + 373 + ], + "lines": [ + { + "bbox": [ + 370, + 368, + 412, + 373 + ], + "spans": [ + { + "bbox": [ + 370, + 368, + 412, + 373 + ], + "type": "text", + "content": "\"Deblur...denoise...\"" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 420, + 308, + 477, + 366 + ], + "blocks": [ + { + "bbox": [ + 420, + 308, + 477, + 366 + ], + "lines": [ + { + "bbox": [ + 420, + 308, + 477, + 366 + ], + "spans": [ + { + "bbox": [ + 420, + 308, + 477, + 366 + ], + "type": "image", + "image_path": "bb9ac5602ac7ac436890df22ffb4eebed637e32e13124a80ab3350f8910b8847.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 434, + 369, + 463, + 375 + ], + "lines": [ + { + "bbox": [ + 434, + 369, + 463, + 375 + ], + "spans": [ + { + "bbox": [ + 434, + 369, + 463, + 375 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 133, + 378, + 190, + 436 + ], + "blocks": [ + { + "bbox": [ + 133, + 378, + 190, + 436 + ], + "lines": [ + { + "bbox": [ + 133, + 378, + 190, + 436 + ], + "spans": [ + { + "bbox": [ + 133, + 378, + 190, + 436 + ], + "type": "image", + "image_path": "b6cf43e0d633afc09beefad697e1af572861e45414f02e78d1ba9297b28a7c97.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 154, + 437, + 167, + 444 + ], + "lines": [ + { + "bbox": [ + 154, + 437, + 167, + 444 + ], + "spans": [ + { + "bbox": [ + 154, + 437, + 167, + 444 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 192, + 378, + 249, + 436 + ], + "blocks": [ + { + "bbox": [ + 192, + 378, + 249, + 436 + ], + "lines": [ + { + "bbox": [ + 192, + 378, + 249, + 436 + ], + "spans": [ + { + "bbox": [ + 192, + 378, + 249, + 436 + ], + "type": "image", + "image_path": "7cb67d3420abd0d8021549e8b2a23d5d84dd3bcf0b200d8c5ceb2a96281553b5.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 216, + 437, + 222, + 442 + ], + "lines": [ + { + "bbox": [ + 216, + 437, + 222, + 442 + ], + "spans": [ + { + "bbox": [ + 216, + 437, + 222, + 442 + ], + "type": "inline_equation", + "content": "\\because {a}_{4} = {2}^{3} \\cdot {2}^{2}" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 249, + 378, + 305, + 436 + ], + "blocks": [ + { + "bbox": [ + 249, + 378, + 305, + 436 + ], + "lines": [ + { + "bbox": [ + 249, + 378, + 305, + 436 + ], + "spans": [ + { + "bbox": [ + 249, + 378, + 305, + 436 + ], + "type": "image", + "image_path": "888052140f6dbc026b92c03600d9936ef2673866e50dcddc4fe7e4c1b08cf5ff.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 264, + 437, + 290, + 443 + ], + "lines": [ + { + "bbox": [ + 264, + 437, + 290, + 443 + ], + "spans": [ + { + "bbox": [ + 264, + 437, + 290, + 443 + ], + "type": "text", + "content": "\"DANGER...\"" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 306, + 378, + 362, + 436 + ], + "blocks": [ + { + "bbox": [ + 306, + 378, + 362, + 436 + ], + "lines": [ + { + "bbox": [ + 306, + 378, + 362, + 436 + ], + "spans": [ + { + "bbox": [ + 306, + 378, + 362, + 436 + ], + "type": "image", + "image_path": "fc2d2f3fe4cfc2197285d8c9ea412eb69693c85971f9eeb31d332b6ff5e44f02.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 324, + 437, + 343, + 443 + ], + "lines": [ + { + "bbox": [ + 324, + 437, + 343, + 443 + ], + "spans": [ + { + "bbox": [ + 324, + 437, + 343, + 443 + ], + "type": "text", + "content": "\"EXIT...\"" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 447, + 482, + 546 + ], + "lines": [ + { + "bbox": [ + 130, + 447, + 482, + 546 + ], + "spans": [ + { + "bbox": [ + 130, + 447, + 482, + 546 + ], + "type": "text", + "content": "Fig. 1: We present SPIRE: Semantic Prompt-Driven Image Restoration, a text-based foundation model for all-in-one, instructed image restoration. SPIRE allows users to flexibly leverage either semantic-level content prompt, or quantitative degradation-aware restoration prompt, or both, to obtain their desired enhancement results based on personal preferences. In other words, SPIRE can be easily prompted to conduct blind restoration, semantic restoration, or task-specific granular treatment. Our framework also enables a new paradigm of instruction-based image restoration, providing a reliable evaluation benchmark to facilitate vision-language models for low-level computational photography applications." + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 362, + 378, + 419, + 436 + ], + "blocks": [ + { + "bbox": [ + 362, + 378, + 419, + 436 + ], + "lines": [ + { + "bbox": [ + 362, + 378, + 419, + 436 + ], + "spans": [ + { + "bbox": [ + 362, + 378, + 419, + 436 + ], + "type": "image", + "image_path": "807ff203776fab889a2b54635f0969f4016753ae9231555252fe55568a52c377.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 384, + 437, + 398, + 443 + ], + "lines": [ + { + "bbox": [ + 384, + 437, + 398, + 443 + ], + "spans": [ + { + "bbox": [ + 384, + 437, + 398, + 443 + ], + "type": "text", + "content": "\"GO...\"" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 439, + 437, + 458, + 443 + ], + "lines": [ + { + "bbox": [ + 439, + 437, + 458, + 443 + ], + "spans": [ + { + "bbox": [ + 439, + 437, + 458, + 443 + ], + "type": "text", + "content": "\"STOP..." + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "bbox": [ + 159, + 566, + 453, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 566, + 453, + 643 + ], + "spans": [ + { + "bbox": [ + 159, + 566, + 453, + 643 + ], + "type": "text", + "content": "Abstract. Text-driven diffusion models have become increasingly popular for various image editing tasks, including inpainting, stylization, and object replacement. However, it still remains an open research problem to adopt this language-vision paradigm for more fine-level image processing tasks, such as denoising, super-resolution, deblurring, and compression artifact removal. In this paper, we develop SPIRE, a Semantic and restoration Prompt-driven Image Restoration framework that leverages" + } + ] + } + ], + "index": 40 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 654, + 398, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 654, + 398, + 665 + ], + "spans": [ + { + "bbox": [ + 141, + 654, + 398, + 665 + ], + "type": "text", + "content": "*This work was done during an internship at Google Research." + } + ] + } + ], + "index": 41 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 159, + 117, + 455, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 117, + 455, + 259 + ], + "spans": [ + { + "bbox": [ + 159, + 117, + 455, + 259 + ], + "type": "text", + "content": "natural language as a user-friendly interface to control the image restoration process. We consider the capacity of prompt information in two dimensions. First, we use content-related prompts to enhance the semantic alignment, effectively alleviating identity ambiguity in the restoration outcomes. Second, our approach is the first framework that supports fine-level instruction through language-based quantitative specification of the restoration strength, without the need for explicit task-specific design. In addition, we introduce a novel fusion mechanism that augments the existing ControlNet architecture by learning to rescale the generative prior, thereby achieving better restoration fidelity. Our extensive experiments demonstrate the superior restoration performance of SPIRE compared to the state of the arts, alongside offering the flexibility of text-based control over the restoration effects." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 281, + 230, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 281, + 230, + 293 + ], + "spans": [ + { + "bbox": [ + 133, + 281, + 230, + 293 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 306, + 482, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 306, + 482, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 306, + 482, + 461 + ], + "type": "text", + "content": "Image restoration or enhancement aims to recover high-quality, pixel-level details from a degraded image, while preserving as much the original semantic information as possible. Although neural network models [11, 12, 30, 63, 67, 71, 73, 74] have marked significant progress, it still remains challenging to design an effective task-conditioning mechanism instead of building multiple individual models for each task (such as denoise, deblur, compression artifact removal) in practical scenarios. The advancement of text-driven diffusion models [48, 50, 53] has unveiled the potential of natural language as a universal input condition for a broad range of image processing challenges, which improves interactivity with users and reduces the cost of task-specific fine-tuning. However, the existing applications of natural language in stylization [38, 44], and inpainting [3, 4, 10, 21, 44, 64] predominantly focus on high-level semantic editing, whereas the uniqueness and challenges for low-level image processing have been less explored." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 463, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 463, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 463, + 482, + 617 + ], + "type": "text", + "content": "Natural language text prompts in image restoration can play two crucial roles: alleviating semantic ambiguities and resolving degradation type ambiguities. Firstly, the same degraded image can correspond to different visual objects, leading to ambiguities in content interpretation (e.g., discerning whether the blurred animal in Fig. 1 is a horse or a zebra). Typically, unconditional image-to-image restoration leads to a random or average estimation of the identities, resulting in neither fish nor fowl. Secondly, certain photographic effects that are deliberately introduced for aesthetic purposes, such as bokeh with a soft out-of-focus background, can be misconstrued as blur distortions by many existing deblurring models, leading to unwanted artifacts or severe hallucinations in the outputs. Although blind restoration methods can produce clean images by either leveraging frozen generative priors [67,70,77], or using end-to-end regression [71], they do not consider the aforementioned ambiguities." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we introduce SPIRE—a Semantic Prompt-Driven Image REstoration framework that provides a user-friendly interface to fully control both the restored image semantics and the enhancement granularity using natural language instructions. Traditional blind restoration without prompts (Fig" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 201, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 201, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 201, + 102 + ], + "type": "text", + "content": "Qi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": "ure 1) has limited flexibility in solving semantic and degradation ambiguity, thus tends to generate average blurry images. Leveraging the semantic and quantitative restoration prompts, we show that the visual quality can be significantly improved. Moreover, our framework can also be used as blind restoration with null prompts. In addition to higher perceptual quality, our framework provides the flexibility for users to generate more than one plausible result. Specifically, in imaging scenarios with large levels of degradation there are multiple plausible solutions. The proposed interactive framework has the capability to personalize the output by leveraging input prompts according to user preference. In concurrent studies, the focus has typically been isolated to one of three areas: either blind restoration [34,67], only semantic prompting [78], or only discrete restoration types [9,36]. To the best of our knowledge, this is the first unified model to support the following three distinct features simultaneously:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 136, + 282, + 481, + 427 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 136, + 282, + 480, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 282, + 480, + 317 + ], + "spans": [ + { + "bbox": [ + 136, + 282, + 480, + 317 + ], + "type": "text", + "content": "1. Blind Restoration: When instructed with a general restoration prompt \"remove all degradations\" and empty semantic prompt \"\", SPIRE operates as a blind restoration model (\"Ours w/o text\" in Tab. 1)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 136, + 319, + 480, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 319, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 136, + 319, + 480, + 354 + ], + "type": "text", + "content": "2. Semantic Restoration: When provided with a text description of the (desired) visual content, SPIRE concentrates on restoring the specified identity of the uncertain or ambiguous objects in the degraded image." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 136, + 356, + 481, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 356, + 481, + 427 + ], + "spans": [ + { + "bbox": [ + 136, + 356, + 481, + 427 + ], + "type": "text", + "content": "3. Quantitative Task-Specific Restoration: Receiving specific restoration type hints (e.g., \"deblur...\", \"deblur... denoise...\"), SPIRE transforms into a task-specific model (e.g., deblur, denoise, or both). Moreover, it understands the numeric nuances of different degradation parameters in language prompts so that the users can control the restoration strength (e.g., \"deblur with sigma 3.0\" in Fig. 7)." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 437, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 628 + ], + "type": "text", + "content": "To train SPIRE, we first build a synthetic data generation pipeline on a large-scale text-image dataset [8] upon the second-order degradation process proposed in Real-ESRGAN [71]. Additionally, we embed the degradation parameters into the restoration prompts to encode finer-grained degradation information (e.g., \"deblur with sigma 3.5\"). Our approach provides richer degradation-specific information compared to contemporary works [26,36] which only employ the degradation types (e.g., \"gaussian blur\"). We then finetune a ControlNet adaptor [84]—which learns a parameter-efficient, parallel branch on top of the latent diffusion models (LDMs) [50]—on a mixture of diverse restoration tasks (see Fig. 1). The semantic text prompts are processed through the LDM backbone, aligning with LDM's text-to-image pretraining. The degradation text prompts and image-conditioning are implemented in the ControlNet branch, as these aspects are not inherently addressed by the vanilla LDM. We further improve the ControlNet by introducing a new modulation connection that adaptively fuses the degradation condition with the generative prior with only a few extra trainable parameters yet showing impressive performance gain." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 630, + 491, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 491, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 491, + 665 + ], + "type": "text", + "content": "Our extensive experiments demonstrate remarkable restoration quality achieved by our proposed SPIRE method. Moreover, SPIRE offers additional freedom for both content and restoration prompt engineering. For example, the semantic" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 233, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 233, + 91, + 447, + 102 + ], + "type": "text", + "content": "SPIRE: Semantic Prompt-Driven Image Restoration" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "prompt \"a very large giraffe eating leaves\" helps to resolve the semantic ambiguity in Fig. 2; the degradation prompt \"deblur with sigma 3.0\" reduces the gaussian blur while maintaining the intentional motion blur in Fig. 3. We also find out that our model learns a continuous latent space of the restoration strength in Fig. 7, even though we do not explicitly fine-tune the CLIP embedding. Our contribution can be summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 194, + 479, + 311 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 131, + 194, + 479, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 194, + 479, + 239 + ], + "spans": [ + { + "bbox": [ + 131, + 194, + 479, + 239 + ], + "type": "text", + "content": "- We introduce the first unified text-driven image restoration model that supports both semantic prompts and restoration instructions. Our experiments demonstrate that incorporating semantic prompts and restoration instructions significantly enhances the restoration quality." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 242, + 479, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 242, + 479, + 263 + ], + "spans": [ + { + "bbox": [ + 130, + 242, + 479, + 263 + ], + "type": "text", + "content": "- Our proposed paradigm empowers users to fully control the semantic outcome of the restored image using different semantic prompts during test time." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 266, + 479, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 266, + 479, + 286 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 479, + 286 + ], + "type": "text", + "content": "- Our proposed approach provides a mechanism for users to adjust the category and strength of the restoration effect based on their subjective preferences." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 289, + 479, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 479, + 311 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 479, + 311 + ], + "type": "text", + "content": "- We demonstrate that text can serve as universal guidance control for low-level image restoration, eliminating the need for task-specific model design." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 131, + 330, + 237, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 330, + 237, + 342 + ], + "spans": [ + { + "bbox": [ + 131, + 330, + 237, + 342 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 354, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 437 + ], + "type": "text", + "content": "First, we review the literature on image restoration. Our proposed framework aims to address some of the limitations of the existing restoration methods. Next, we compare multiple text-guided diffusion editing methods. Our method is designed to deal with two different text ambiguities concurrently, which is an unprecedented challenge for existing text-driven diffusion methods. Finally, we discuss the parameter-efficient fine-tuning of diffusion models, which further motivates our proposed modulation connection." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 438, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 482, + 545 + ], + "type": "text", + "content": "Image Restoration is the task of recovering a high-resolution, clean image from a degraded input. Pioneering works in the fields of super-resolution [14,32], motion and defocus deblurring [1,61,79], denoising [82], and JPEG and artifact removal [16, 45] primarily utilize deep neural network architectures to address specific tasks. Later, Transformer-based [30,63,73] and adversarial-based [71] formulations were explored, demonstrating state-of-the-art performance with unified model architecture. Recently there has been a focus on exploiting iterative restorations, such as the ones from diffusion models, to generate images with much higher perceptual quality [11,49,52,54,74,75,78]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 546, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 629 + ], + "type": "text", + "content": "There have been some recent works attempting to control the enhancement strength of a single model, such as the scaling factor condition in arbitrary super-resolution [68, 69] and noise-level map for image denoising [83]. However, these methods require hand-crafting dedicated architectures for each task, which limits their scalability. In contrast to the task-specific models, other approaches seek to train a single blind model by building a pipeline composed of several classical degradation process, such as BSRGAN [80] and Real-ESRGAN [71]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "Text-guided Diffusion Editing. Denoising diffusion models are trained to reverse the noisiging process [23, 57, 58]. Early methods focused on unconditioned generation [13], but recent trends pivot more to conditioned generation, such as" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 201, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 201, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 201, + 102 + ], + "type": "text", + "content": "Qi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 478, + 221 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 478, + 221 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 478, + 221 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 478, + 221 + ], + "type": "image", + "image_path": "271b8020dfdd93e0071501f87af5c4adb0823db53ee96ea6e256577234cfe22d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 224, + 482, + 323 + ], + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 323 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 323 + ], + "type": "text", + "content": "Fig. 2: Framework of SPIRE. In the training phase, we begin by synthesizing a degraded version " + }, + { + "bbox": [ + 130, + 224, + 482, + 323 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 224, + 482, + 323 + ], + "type": "text", + "content": ", of a clean image " + }, + { + "bbox": [ + 130, + 224, + 482, + 323 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 224, + 482, + 323 + ], + "type": "text", + "content": ". Our degradation synthesis pipeline also creates a restoration prompt " + }, + { + "bbox": [ + 130, + 224, + 482, + 323 + ], + "type": "inline_equation", + "content": "c_{r}" + }, + { + "bbox": [ + 130, + 224, + 482, + 323 + ], + "type": "text", + "content": ", which contains numeric parameters that reflects the intensity of the degradation introduced. Then, we inject the synthetic restoration prompt into a ControlNet adaptor, which uses our proposed modulation fusion blocks " + }, + { + "bbox": [ + 130, + 224, + 482, + 323 + ], + "type": "inline_equation", + "content": "(\\gamma, \\beta)" + }, + { + "bbox": [ + 130, + 224, + 482, + 323 + ], + "type": "text", + "content": " to connect with the frozen backbone driven by the semantic prompt " + }, + { + "bbox": [ + 130, + 224, + 482, + 323 + ], + "type": "inline_equation", + "content": "c_{s}" + }, + { + "bbox": [ + 130, + 224, + 482, + 323 + ], + "type": "text", + "content": ". During test time, the users can employ the SPIRE framework as either a blind restoration model with restoration prompt \"Remove all degradation\" and empty semantic prompt " + }, + { + "bbox": [ + 130, + 224, + 482, + 323 + ], + "type": "inline_equation", + "content": "\\varnothing" + }, + { + "bbox": [ + 130, + 224, + 482, + 323 + ], + "type": "text", + "content": ", or manually adjust the restoration " + }, + { + "bbox": [ + 130, + 224, + 482, + 323 + ], + "type": "inline_equation", + "content": "c_{r}" + }, + { + "bbox": [ + 130, + 224, + 482, + 323 + ], + "type": "text", + "content": " and semantic prompts " + }, + { + "bbox": [ + 130, + 224, + 482, + 323 + ], + "type": "inline_equation", + "content": "c_{s}" + }, + { + "bbox": [ + 130, + 224, + 482, + 323 + ], + "type": "text", + "content": " to obtain what they want." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 343, + 482, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 343, + 482, + 596 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 482, + 596 + ], + "type": "text", + "content": "image-to-image translation [52, 55] and text-to-image generation [50, 53]. Latent Diffusion [50] is a groundbreaking approach that introduces a versatile framework for improving both training and sampling efficiency, while flexible enough for general conditioning inputs. Subsequent works have built upon their pretrained text-to-image checkpoints, and designed customized architectures for different tasks like text-guided image editing. For instance, SDEdit [37] generates content for a new prompt by adding noise to the input image. Attention-based methods [21, 44, 64, 81] show that images can be edited via reweighting and replacing the cross-attention map of different prompts. Aside from the previously mentioned approaches driven by target prompts, instruction-based editing [7, 17] entail modifying a source image based on specific instructions. These textual instructions are typically synthesized using large language models [41, 47]. As CLIP serves as a cornerstone for bridging vision and language, several studies have extensively investigated the representation space of CLIP embeddings [46]. Instead of directly applying CLIP to fundamental discriminative tasks [19, 85], these works either tailor CLIP to specific applications, such as image quality assessment [28, 31], serve as semantic guidance in generative adversarial network [5] or aligning image embedding with degradation types in the feature space [26, 36]. To enhance CLIP's ability to understand numerical information, some works [42, 43] finetune it using contrastive learning on synthetic numerical data, and then incorporate the fine-tuned text embedding into diffusion training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Parameter-Efficient Diffusion Model Finetuning. To leverage the powerful generative prior in pretrained diffusion models, parameter-efficient components such as text embedding [15], low-rank approximations of model weights [20, 25], and cross attention layers [29] can be finetuned to personalize the pretrained model. Adaptor-based finetuning paradigms [39, 84] propose to keep the original" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 234, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 234, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 234, + 91, + 447, + 102 + ], + "type": "text", + "content": "SPIRE: Semantic Prompt-Driven Image Restoration" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "type": "text", + "content": "UNet weights frozen, and add new image or text conditions. This adaptor generates residual features that are subsequently added to the frozen UNet backbone." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 159, + 202, + 172 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 159, + 202, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 159, + 202, + 172 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 185, + 479, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 185, + 479, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 185, + 479, + 281 + ], + "type": "text", + "content": "We introduce a universal approach to combine the above mentioned task-specific, strength-aware, and blind restoration methods within a unified framework (illustrated in Sec. 3.2). We further propose to decouple the learning of content and restoration prompts to better preserve the pre-trained prior while injecting new conditions, as unfolded in Sec. 3.3. Sec. 3.4 details our design on how to accurately control the restoration type and strength, as well as our proposed modulation fusion layer that adaptively fuses the restoration features back to the frozen backbone. We start with some preliminaries in the following section." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 299, + 228, + 310 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 299, + 228, + 310 + ], + "spans": [ + { + "bbox": [ + 132, + 299, + 228, + 310 + ], + "type": "text", + "content": "3.1 Preliminaries" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 319, + 480, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 480, + 427 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 480, + 427 + ], + "type": "text", + "content": "Latent Diffusion Models (LDMs) [50] are probabilistic generative models that learn the underlying data distribution by iteratively removing Gaussian noise in the latent space, which is typically learned using a VAE autoencoder. Formally, the VAE encoder " + }, + { + "bbox": [ + 130, + 319, + 480, + 427 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 130, + 319, + 480, + 427 + ], + "type": "text", + "content": " compresses an input image " + }, + { + "bbox": [ + 130, + 319, + 480, + 427 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 130, + 319, + 480, + 427 + ], + "type": "text", + "content": " into a compact latent representation " + }, + { + "bbox": [ + 130, + 319, + 480, + 427 + ], + "type": "inline_equation", + "content": "\\pmb{z} = \\mathcal{E}(\\pmb{x})" + }, + { + "bbox": [ + 130, + 319, + 480, + 427 + ], + "type": "text", + "content": ", which can be later decoded back to the pixel space using the coupled VAE decoder " + }, + { + "bbox": [ + 130, + 319, + 480, + 427 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 319, + 480, + 427 + ], + "type": "text", + "content": ", often learned under an image reconstruction objective: " + }, + { + "bbox": [ + 130, + 319, + 480, + 427 + ], + "type": "inline_equation", + "content": "\\mathcal{D}(\\mathcal{E}(\\pmb{x})) \\approx \\pmb{x}" + }, + { + "bbox": [ + 130, + 319, + 480, + 427 + ], + "type": "text", + "content": ". During training stage, the output of a UNet [51] " + }, + { + "bbox": [ + 130, + 319, + 480, + 427 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(z_t, t, \\pmb{y})" + }, + { + "bbox": [ + 130, + 319, + 480, + 427 + ], + "type": "text", + "content": " conditioned on " + }, + { + "bbox": [ + 130, + 319, + 480, + 427 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 130, + 319, + 480, + 427 + ], + "type": "text", + "content": " (such as text, images, etc.) is parameterized [13, 56] to remove Gaussian noise " + }, + { + "bbox": [ + 130, + 319, + 480, + 427 + ], + "type": "inline_equation", + "content": "\\pmb{\\epsilon}" + }, + { + "bbox": [ + 130, + 319, + 480, + 427 + ], + "type": "text", + "content": " in the latent space " + }, + { + "bbox": [ + 130, + 319, + 480, + 427 + ], + "type": "inline_equation", + "content": "\\pmb{z}_t" + }, + { + "bbox": [ + 130, + 319, + 480, + 427 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 208, + 436, + 480, + 454 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 436, + 480, + 454 + ], + "spans": [ + { + "bbox": [ + 208, + 436, + 480, + 454 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} \\mathbb {E} _ {\\left(\\boldsymbol {z} _ {0}, \\boldsymbol {y}\\right) \\sim p _ {\\text {d a t a}}, \\epsilon \\sim \\mathcal {N} (0, I), t} \\| \\boldsymbol {\\epsilon} - \\boldsymbol {\\epsilon} _ {\\theta} \\left(\\boldsymbol {z} _ {t}, t, \\boldsymbol {y}\\right) \\| _ {2} ^ {2}, \\tag {1}", + "image_path": "6d05a84ba4d5e0b02288cd76b3bf166d43e7c136f4bdadeeb341d3e7bb8545d6.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 463, + 479, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 463, + 479, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 463, + 479, + 486 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 463, + 479, + 486 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 130, + 463, + 479, + 486 + ], + "type": "text", + "content": " is a noisy sample of " + }, + { + "bbox": [ + 130, + 463, + 479, + 486 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0" + }, + { + "bbox": [ + 130, + 463, + 479, + 486 + ], + "type": "text", + "content": " at sampled timestep " + }, + { + "bbox": [ + 130, + 463, + 479, + 486 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 463, + 479, + 486 + ], + "type": "text", + "content": ". The condition " + }, + { + "bbox": [ + 130, + 463, + 479, + 486 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 130, + 463, + 479, + 486 + ], + "type": "text", + "content": " is randomly dropped out as " + }, + { + "bbox": [ + 130, + 463, + 479, + 486 + ], + "type": "inline_equation", + "content": "\\varnothing" + }, + { + "bbox": [ + 130, + 463, + 479, + 486 + ], + "type": "text", + "content": " to make the model unconditional." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 487, + 479, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 479, + 535 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 479, + 535 + ], + "type": "text", + "content": "At test time, deterministic DDIM sampling [57] is utilized to convert a random noise " + }, + { + "bbox": [ + 130, + 487, + 479, + 535 + ], + "type": "inline_equation", + "content": "z_{T}" + }, + { + "bbox": [ + 130, + 487, + 479, + 535 + ], + "type": "text", + "content": " to a clean latent " + }, + { + "bbox": [ + 130, + 487, + 479, + 535 + ], + "type": "inline_equation", + "content": "z_{0}" + }, + { + "bbox": [ + 130, + 487, + 479, + 535 + ], + "type": "text", + "content": " which is decoded as final result " + }, + { + "bbox": [ + 130, + 487, + 479, + 535 + ], + "type": "inline_equation", + "content": "\\mathcal{D}(z_0)" + }, + { + "bbox": [ + 130, + 487, + 479, + 535 + ], + "type": "text", + "content": ". In each timestep " + }, + { + "bbox": [ + 130, + 487, + 479, + 535 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 487, + 479, + 535 + ], + "type": "text", + "content": ", classifier-free guidance [13, 24] can be applied to trade-off sample quality and condition alignment:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 184, + 545, + 424, + 558 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 545, + 424, + 558 + ], + "spans": [ + { + "bbox": [ + 184, + 545, + 424, + 558 + ], + "type": "interline_equation", + "content": "\\overline {{\\epsilon}} _ {\\theta} \\left(\\boldsymbol {z} _ {t}, t, \\boldsymbol {y}\\right) = \\epsilon_ {\\theta} \\left(\\boldsymbol {z} _ {t}, t, \\varnothing\\right) + w \\left(\\epsilon_ {\\theta} \\left(\\boldsymbol {z} _ {t}, t, \\boldsymbol {y}\\right) - \\epsilon_ {\\theta} \\left(\\boldsymbol {z} _ {t}, t, \\varnothing\\right)\\right),", + "image_path": "e288a700fc917d3d4678d62673d461c0b2df4b953fe432203303fbeb3ebf757b.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 567, + 479, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 567, + 479, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 567, + 479, + 604 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 567, + 479, + 604 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 130, + 567, + 479, + 604 + ], + "type": "text", + "content": " is a scalar to adjust the guidance strength of " + }, + { + "bbox": [ + 130, + 567, + 479, + 604 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 567, + 479, + 604 + ], + "type": "text", + "content": ". Note that the estimated noise " + }, + { + "bbox": [ + 130, + 567, + 479, + 604 + ], + "type": "inline_equation", + "content": "\\overline{\\epsilon}_{\\theta}(\\pmb{z}_t, t, \\pmb{y})" + }, + { + "bbox": [ + 130, + 567, + 479, + 604 + ], + "type": "text", + "content": " is used to update " + }, + { + "bbox": [ + 130, + 567, + 479, + 604 + ], + "type": "inline_equation", + "content": "z_t" + }, + { + "bbox": [ + 130, + 567, + 479, + 604 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 567, + 479, + 604 + ], + "type": "inline_equation", + "content": "z_{t-1}" + }, + { + "bbox": [ + 130, + 567, + 479, + 604 + ], + "type": "text", + "content": ", which is a approximation of the distribution gradient score [60] as " + }, + { + "bbox": [ + 130, + 567, + 479, + 604 + ], + "type": "inline_equation", + "content": "\\nabla_{z_t} \\log p(\\pmb{z}_t | \\pmb{y}) \\propto \\overline{\\epsilon}_{\\theta}(\\pmb{z}_t, t, \\pmb{y})" + }, + { + "bbox": [ + 130, + 567, + 479, + 604 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 621, + 317, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 317, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 317, + 633 + ], + "type": "text", + "content": "3.2 Text-driven Image Restoration" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "Based on the LDM framework, we propose a new restoration paradigm—text-driven image restoration. Our method targets to restore images " + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "inline_equation", + "content": "(\\pmb{x}" + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\pmb{z}_0)" + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": " based" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 201, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 201, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 201, + 102 + ], + "type": "text", + "content": "Qi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "text", + "content": "on conditions " + }, + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "inline_equation", + "content": "\\{\\pmb{y},\\pmb{c}_s,\\pmb{c}_r\\}" + }, + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "text", + "content": ". Specifically: " + }, + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "text", + "content": " denotes the degraded image condition, " + }, + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "inline_equation", + "content": "\\pmb{c}_s" + }, + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "text", + "content": " is the semantic prompt describing the clean image " + }, + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "text", + "content": " (e.g., \"a panda is sitting by the bamboo\" or \"a panda\"), and " + }, + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "inline_equation", + "content": "\\pmb{c}_r" + }, + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "text", + "content": " is the restoration prompt that describes the details of the degradation in terms of both the operation and parameter (e.g., \"deblur with sigma 3.0\"). We use " + }, + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "inline_equation", + "content": "\\pmb{y} = \\mathrm{Deg}(\\pmb{x},\\pmb{c}_r)" + }, + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "text", + "content": " to denote the degradation process which turns the clean image " + }, + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "text", + "content": " into its degraded counterpart " + }, + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "inline_equation", + "content": "\\pmb{y}_i" + }, + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 482, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 482, + 249 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 482, + 249 + ], + "type": "text", + "content": "The above text-driven image restoration model " + }, + { + "bbox": [ + 130, + 188, + 482, + 249 + ], + "type": "inline_equation", + "content": "p(\\pmb{z}_t|\\{\\pmb{y},\\pmb{c}_s,\\pmb{c}_r\\})" + }, + { + "bbox": [ + 130, + 188, + 482, + 249 + ], + "type": "text", + "content": " can be trained using paired data. We use the text-image dataset [8] so that each clean image " + }, + { + "bbox": [ + 130, + 188, + 482, + 249 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 130, + 188, + 482, + 249 + ], + "type": "text", + "content": " is paired with a semantic prompt " + }, + { + "bbox": [ + 130, + 188, + 482, + 249 + ], + "type": "inline_equation", + "content": "\\pmb{c}_s" + }, + { + "bbox": [ + 130, + 188, + 482, + 249 + ], + "type": "text", + "content": " (the paired alt-text). Then " + }, + { + "bbox": [ + 130, + 188, + 482, + 249 + ], + "type": "inline_equation", + "content": "\\pmb{y} = \\mathrm{Deg}(\\pmb{x},\\pmb{c}_r)" + }, + { + "bbox": [ + 130, + 188, + 482, + 249 + ], + "type": "text", + "content": " is simulated using the synthetic degradation [71] pipeline (Sec. 4.1), yielding the final paired training data " + }, + { + "bbox": [ + 130, + 188, + 482, + 249 + ], + "type": "inline_equation", + "content": "(\\pmb{x} \\text{ or } \\pmb{z}_0, \\{\\pmb{y},\\pmb{c}_s,\\pmb{c}_r\\}) \\sim p_{data}" + }, + { + "bbox": [ + 130, + 188, + 482, + 249 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 251, + 482, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 251, + 482, + 455 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 482, + 455 + ], + "type": "text", + "content": "Comparison to Blind Restoration. Recent existing blind restoration models [34, 55, 67] also leverage diffusion priors to generate high-quality images. However, these methods are prone to hallucinate unwanted, unnatural, or oversharpened textures given that the semantic- and degradation-ambiguities persist everywhere. Fig. 3 provides an example, where the input image is degraded simultaneously with Gaussian blur and motion blur. The fully blind restoration method removes all of the degradation, leaving no way to steer the model to remove only the Gaussian blur while preserving the motion effect for aesthetic purposes. Therefore, having the flexibility to control the restoration behavior and its strength at test time is a crucial requirement. Our text-driven formulation of the problem introduces additional controlling ability using both the semantic content and the restoration instruction. In our SPIRE design, " + }, + { + "bbox": [ + 130, + 251, + 482, + 455 + ], + "type": "inline_equation", + "content": "\\pmb{c}_r =" + }, + { + "bbox": [ + 130, + 251, + 482, + 455 + ], + "type": "text", + "content": " \"Remove all degradation\" is also a plausible restoration prompt which makes our SPIRE to be compatible with the blind restoration. Further, our text-driven restoration takes advantages of the pretrained content-aware LDM using the additional semantic prompt " + }, + { + "bbox": [ + 130, + 251, + 482, + 455 + ], + "type": "inline_equation", + "content": "\\pmb{c}_s" + }, + { + "bbox": [ + 130, + 251, + 482, + 455 + ], + "type": "text", + "content": ", hence able to better handle semantic ambiguities on the noisy, blurry regions whose visual objects are less recognizable." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 473, + 399, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 473, + 399, + 487 + ], + "spans": [ + { + "bbox": [ + 132, + 473, + 399, + 487 + ], + "type": "text", + "content": "3.3 Decoupling Semantic and Restoration Prompts" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 495, + 482, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 482, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 482, + 578 + ], + "type": "text", + "content": "To effectively learn the latent distribution " + }, + { + "bbox": [ + 130, + 495, + 482, + 578 + ], + "type": "inline_equation", + "content": "p(\\pmb{z}_t | \\pmb{y}, \\pmb{c}_s, \\pmb{c}_r)" + }, + { + "bbox": [ + 130, + 495, + 482, + 578 + ], + "type": "text", + "content": ", we further decouple the conditions " + }, + { + "bbox": [ + 130, + 495, + 482, + 578 + ], + "type": "inline_equation", + "content": "\\{\\pmb{y}, \\pmb{c}_s, \\pmb{c}_r\\}" + }, + { + "bbox": [ + 130, + 495, + 482, + 578 + ], + "type": "text", + "content": " into two groups—one for the text-to-image prior " + }, + { + "bbox": [ + 130, + 495, + 482, + 578 + ], + "type": "inline_equation", + "content": "(\\pmb{c}_s \\rightarrow \\pmb{z}_t)" + }, + { + "bbox": [ + 130, + 495, + 482, + 578 + ], + "type": "text", + "content": " already imbued in the pretrained LDM model, and the other for both the image-to-image " + }, + { + "bbox": [ + 130, + 495, + 482, + 578 + ], + "type": "inline_equation", + "content": "(\\pmb{y} \\rightarrow \\pmb{z}_t)" + }, + { + "bbox": [ + 130, + 495, + 482, + 578 + ], + "type": "text", + "content": " and restoration-to-image " + }, + { + "bbox": [ + 130, + 495, + 482, + 578 + ], + "type": "inline_equation", + "content": "(\\pmb{c}_r \\rightarrow \\pmb{z}_t)" + }, + { + "bbox": [ + 130, + 495, + 482, + 578 + ], + "type": "text", + "content": " that needs to be learnt from the synthetic data. This decoupling strategy prevents catastrophic forgetting in the pretrained diffusion model and enables independent training of the restoration-aware model, whose score function [60] is expressed as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 149, + 579, + 481, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 579, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 149, + 579, + 481, + 609 + ], + "type": "interline_equation", + "content": "\\nabla_ {z _ {t}} \\log p (\\boldsymbol {z} _ {t} | \\boldsymbol {y}, \\boldsymbol {c} _ {s}, \\boldsymbol {c} _ {r}) \\approx \\underbrace {\\nabla_ {z _ {t}} \\log p (\\boldsymbol {z} _ {t} | \\boldsymbol {c} _ {s})} _ {\\text {S e m a n t i c - a w a r e (f r o z e n)}} + \\underbrace {\\nabla_ {z _ {t}} \\log p (\\boldsymbol {y} | \\boldsymbol {z} _ {t} , \\boldsymbol {c} _ {r})} _ {\\text {R e s t o r a t i o n - a w a r e (l e a r n a b l e)}}. \\tag {2}", + "image_path": "d510e224793316b4e31150264fd196998c9d08db12504615c6075dd72449b5c0.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "In the above equation, the first part " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\nabla_{\\boldsymbol{z}_t}\\log p(\\boldsymbol{z}_t|\\boldsymbol{c}_s)" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " aligns with the text-to-image prior inherent in the pretrained LDM. The second term " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\nabla_{\\boldsymbol{z}_t}\\log p(\\boldsymbol{y}|\\boldsymbol{z}_t,\\boldsymbol{c}_r)" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " approximates the consistency constraint with degraded image " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\boldsymbol{y}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ", meaning the latent image " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\boldsymbol{z}_t" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " should be inferred by the degraded image " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\boldsymbol{y}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " and the degradation" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 233, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 233, + 91, + 448, + 102 + ], + "type": "text", + "content": "SPIRE: Semantic Prompt-Driven Image Restoration" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": "details " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "\\pmb{c}_r" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": ". This is somewhat similar to the reverse process of " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "\\pmb{y} = \\mathrm{Deg}(\\pmb{x}_t, \\pmb{c}_r)" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": " in the pixel space. More derivation is provided in our supplement. Providing degradation information through a restoration text prompt " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "\\pmb{c}_r" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": " can largely reduce the uncertainty of estimation " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "p(\\pmb{y} | \\pmb{z}_t, \\pmb{c}_r)" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": ", thereby alleviating the ambiguity of the restoration task and leading to better results that are less prone to hallucinations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 182, + 346, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 182, + 346, + 194 + ], + "spans": [ + { + "bbox": [ + 131, + 182, + 346, + 194 + ], + "type": "text", + "content": "3.4 Learning to Control the Restoration." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 198, + 264, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 198, + 264, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 198, + 264, + 342 + ], + "type": "text", + "content": "Inspired by ControlNet [84], we employ an adaptor to model the restoration-aware branch in Eq. (2). First, the pretrained latent diffusion UNet " + }, + { + "bbox": [ + 130, + 198, + 264, + 342 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\pmb{z}_t,t,\\pmb{c}_s)" + }, + { + "bbox": [ + 130, + 198, + 264, + 342 + ], + "type": "text", + "content": " is frozen during training to preserve text-to-image prior " + }, + { + "bbox": [ + 130, + 198, + 264, + 342 + ], + "type": "inline_equation", + "content": "\\nabla_{\\pmb{z}_t}\\log p(\\pmb{z}_t|\\pmb{c}_s)" + }, + { + "bbox": [ + 130, + 198, + 264, + 342 + ], + "type": "text", + "content": ". Secondly, an encoder initialized from pretrained UNet is finetuned to learn the restoration condi" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 271, + 209, + 482, + 272 + ], + "blocks": [ + { + "bbox": [ + 271, + 209, + 482, + 272 + ], + "lines": [ + { + "bbox": [ + 271, + 209, + 482, + 272 + ], + "spans": [ + { + "bbox": [ + 271, + 209, + 482, + 272 + ], + "type": "image", + "image_path": "cd4cdd9227e8c76a2fe883bd0d71f902ce64ed16a21d140fb49458a3d8c0dad6.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 269, + 277, + 482, + 333 + ], + "lines": [ + { + "bbox": [ + 269, + 277, + 482, + 333 + ], + "spans": [ + { + "bbox": [ + 269, + 277, + 482, + 333 + ], + "type": "text", + "content": "Fig. 3: Degradation ambiguities in real-world problems. By adjusting the restoration prompt, our method can preserve the motion effect that is coupled with Gaussian blur, while fully blind restoration models do not provide this flexibility." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 342, + 482, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 390 + ], + "type": "text", + "content": "tion " + }, + { + "bbox": [ + 130, + 342, + 482, + 390 + ], + "type": "inline_equation", + "content": "\\nabla_{z_t}\\log p(\\pmb {y}|\\pmb {z}_t,\\pmb {c}_r)" + }, + { + "bbox": [ + 130, + 342, + 482, + 390 + ], + "type": "text", + "content": ", which takes degradation image " + }, + { + "bbox": [ + 130, + 342, + 482, + 390 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 130, + 342, + 482, + 390 + ], + "type": "text", + "content": ", current noisy latents " + }, + { + "bbox": [ + 130, + 342, + 482, + 390 + ], + "type": "inline_equation", + "content": "\\pmb{z}_{t}" + }, + { + "bbox": [ + 130, + 342, + 482, + 390 + ], + "type": "text", + "content": " and the restoration prompt " + }, + { + "bbox": [ + 130, + 342, + 482, + 390 + ], + "type": "inline_equation", + "content": "\\pmb{c}_{r}" + }, + { + "bbox": [ + 130, + 342, + 482, + 390 + ], + "type": "text", + "content": " as input. The output of encoder is fusion features " + }, + { + "bbox": [ + 130, + 342, + 482, + 390 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{con}" + }, + { + "bbox": [ + 130, + 342, + 482, + 390 + ], + "type": "text", + "content": " for frozen UNet. In this way, the choice of ControlNet implementation follows our decoupling analysis in Eq. (2)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 390, + 482, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 482, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 482, + 461 + ], + "type": "text", + "content": "Conditioning on the Degraded Image " + }, + { + "bbox": [ + 130, + 390, + 482, + 461 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 130, + 390, + 482, + 461 + ], + "type": "text", + "content": ". We apply downsample layers on " + }, + { + "bbox": [ + 130, + 390, + 482, + 461 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 130, + 390, + 482, + 461 + ], + "type": "text", + "content": " to make the feature map matching to the shape of the " + }, + { + "bbox": [ + 130, + 390, + 482, + 461 + ], + "type": "inline_equation", + "content": "\\pmb{z}_t" + }, + { + "bbox": [ + 130, + 390, + 482, + 461 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 130, + 390, + 482, + 461 + ], + "type": "inline_equation", + "content": "\\mathcal{E}'(\\pmb{y})" + }, + { + "bbox": [ + 130, + 390, + 482, + 461 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 390, + 482, + 461 + ], + "type": "inline_equation", + "content": "\\mathcal{E}'" + }, + { + "bbox": [ + 130, + 390, + 482, + 461 + ], + "type": "text", + "content": " is more compact [84] than VAE encoder " + }, + { + "bbox": [ + 130, + 390, + 482, + 461 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 130, + 390, + 482, + 461 + ], + "type": "text", + "content": ", and its trainable parameters allow model to perceive and adapt to the degraded images from our synthetic data pipeline. Then, we concatenate the downsampled feature map " + }, + { + "bbox": [ + 130, + 390, + 482, + 461 + ], + "type": "inline_equation", + "content": "\\mathcal{E}'(\\pmb{y})" + }, + { + "bbox": [ + 130, + 390, + 482, + 461 + ], + "type": "text", + "content": " with the latent " + }, + { + "bbox": [ + 130, + 390, + 482, + 461 + ], + "type": "inline_equation", + "content": "\\pmb{z}_t" + }, + { + "bbox": [ + 130, + 390, + 482, + 461 + ], + "type": "text", + "content": ", and feed them to the ControlNet encoder." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 462, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 641 + ], + "type": "text", + "content": "Conditioning on the Restoration Prompt " + }, + { + "bbox": [ + 130, + 462, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\pmb{c}_r" + }, + { + "bbox": [ + 130, + 462, + 482, + 641 + ], + "type": "text", + "content": ". Following [50], we first use the vision-language CLIP [46] to infer the text sequence embedding " + }, + { + "bbox": [ + 130, + 462, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\pmb{e}_r \\in \\mathbb{R}^{M \\times d}" + }, + { + "bbox": [ + 130, + 462, + 482, + 641 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 130, + 462, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\pmb{c}_r" + }, + { + "bbox": [ + 130, + 462, + 482, + 641 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 462, + 482, + 641 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 462, + 482, + 641 + ], + "type": "text", + "content": " is the number of tokens, and " + }, + { + "bbox": [ + 130, + 462, + 482, + 641 + ], + "type": "inline_equation", + "content": "d = 768" + }, + { + "bbox": [ + 130, + 462, + 482, + 641 + ], + "type": "text", + "content": " is the dimension of the image-text features. Specifically, given a restoration instruction \"deblur with sigma 3.0\", CLIP first tokenizes it into a sequence of tokens (e.g., \"3.0\" is tokenized into [\"3\", \".\", \"0)]) and embeds them into " + }, + { + "bbox": [ + 130, + 462, + 482, + 641 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 462, + 482, + 641 + ], + "type": "text", + "content": " distinct token embeddings. Then, stacked causal transformer layers [65] are applied on the embedding sequence to contextualize the token embeddings, resulting in " + }, + { + "bbox": [ + 130, + 462, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\pmb{e}_r" + }, + { + "bbox": [ + 130, + 462, + 482, + 641 + ], + "type": "text", + "content": ". Hence, tokens such as \"deblur\" may arm with the numeric information such as \"3\", \".\", and \"0\", making it possible for the model to distinguish the semantic difference of strength definition in different restoration tasks. Cross-attention process ensures that the information from " + }, + { + "bbox": [ + 130, + 462, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\pmb{e}_r" + }, + { + "bbox": [ + 130, + 462, + 482, + 641 + ], + "type": "text", + "content": " is propagated to the ControlNet feature " + }, + { + "bbox": [ + 130, + 462, + 482, + 641 + ], + "type": "inline_equation", + "content": "f_{con}" + }, + { + "bbox": [ + 130, + 462, + 482, + 641 + ], + "type": "text", + "content": ". While [42, 43] finetuned CLIP using contrastive learning, while we found frozen CLIP still works. The reason is that the learnable cross-attention layers can somehow ensemble the observations from frozen CLIP and squeeze useful information from it." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Modulation Fusion Layer. The decoupling of learning two branches (Sec. 3.3), despite offering the benefit of effective separation of prior and conditioning," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 201, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 201, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 201, + 102 + ], + "type": "text", + "content": "Qi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 114, + 479, + 205 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 479, + 205 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 479, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 479, + 205 + ], + "type": "table", + "html": "
MethodPromptsParameterized Degradation with synthesized crReal-ESRGAN Degradation without cr
Sem ResFID↓LPIPS↓PSNR↑SSIM↑CLIP-I↑CLIP-T↑FID↓LPIPS↓PSNR↑SSIM↑CLIP-I↑CLIP-T↑
SwinIR [30]XX43.220.42324.400.7170.8560.28548.370.44923.450.6990.8420.284
StableSR [67]XX20.550.31321.030.6130.8860.29825.750.36420.420.5810.8640.298
DiffBIR [34]XX17.260.30222.160.6040.9120.29719.170.33021.480.5870.8980.298
ControlNet-SR [84]XX13.650.22223.750.6690.9380.30016.990.26922.950.6280.9240.299
Ours w/o textXX12.700.22123.840.6710.9390.29916.250.26223.150.6360.9290.300
DiffBIR [34] + SDEdit [37]✓X19.360.36219.390.5270.8910.30517.510.37519.150.5210.8870.308
DiffBIR [34] + CLIP [46]✓X18.460.36520.500.5260.8960.30820.310.37420.450.5390.8850.307
ControlNet-SR + CLIP [46]✓X13.000.24123.180.6480.9370.30715.160.28622.450.6100.9260.308
Ours✓✓11.340.21923.610.6650.9430.30614.420.26223.140.6330.9350.308
", + "image_path": "1e3f3902d4d30784056e664e6ad875526750512d1255cfc216e8fb51a40a35ab.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 133, + 273, + 479, + 318 + ], + "blocks": [ + { + "bbox": [ + 130, + 207, + 480, + 262 + ], + "lines": [ + { + "bbox": [ + 130, + 207, + 480, + 262 + ], + "spans": [ + { + "bbox": [ + 130, + 207, + 480, + 262 + ], + "type": "text", + "content": "Table 1: Quantitative results on the MS-COCO dataset (with " + }, + { + "bbox": [ + 130, + 207, + 480, + 262 + ], + "type": "inline_equation", + "content": "c_{s}" + }, + { + "bbox": [ + 130, + 207, + 480, + 262 + ], + "type": "text", + "content": ") using our parameterized degradation (left) and Real-ESRGAN degradation (right). We also denote the prompt choice at test time. 'Sem' stands for semantic prompt; 'Res' stands for restoration prompt. The first group of baselines are tested without prompt. The second group are combined with semantic prompt in zero-shot way." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 273, + 479, + 318 + ], + "lines": [ + { + "bbox": [ + 133, + 273, + 479, + 318 + ], + "spans": [ + { + "bbox": [ + 133, + 273, + 479, + 318 + ], + "type": "table", + "html": "
Degradation ProcessParameterizedReal-ESRGAN
p(choose)Restoration PromptDegradation Processp(choose)Restoration Prompt
Gaussian Blur0.5Deblur with {sigma∈[0.2, 3.0]} or DeblurBlur1.0
Downsample0.5Upsample to {resizing factor∈[1.0, 7.0]} or UpsampleResize1.0Remove all degradation
Gaussian Noise0.5Denoise with {sigma∈[0.0, 0.12]} or DenoiseNoise1.0
JPEG0.5Dejpeg with quality {quality factor∈[30, 92]} or DejpegJPEG1.0
", + "image_path": "ea171af9dab562f30e82091fe411d74eaebd406121aa6c4a112c5e17995cf4c4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 319, + 481, + 373 + ], + "lines": [ + { + "bbox": [ + 130, + 319, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 481, + 373 + ], + "type": "text", + "content": "Table 2: Our training degradation is randomly sampled in these two pipeline with " + }, + { + "bbox": [ + 130, + 319, + 481, + 373 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 319, + 481, + 373 + ], + "type": "text", + "content": " each. (1) Images generated by our parameterized pipeline are paired with either a restoration type (e.g., \"Deblur\") or a restoration parameter prompt (e.g., \"Deblur with sigma 0.3.\") (2) In other " + }, + { + "bbox": [ + 130, + 319, + 481, + 373 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 319, + 481, + 373 + ], + "type": "text", + "content": " iterations, degraded images " + }, + { + "bbox": [ + 130, + 319, + 481, + 373 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 130, + 319, + 481, + 373 + ], + "type": "text", + "content": " synthesized by Real-ESRGAN are paired with the same restoration prompt " + }, + { + "bbox": [ + 130, + 319, + 481, + 373 + ], + "type": "inline_equation", + "content": "\\pmb{c}_r = \\text{\"Remove all degradation\"}" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 387, + 482, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 387, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 387, + 482, + 472 + ], + "type": "text", + "content": "may cause distribution shift in the learned representations between the frozen backbone feature " + }, + { + "bbox": [ + 130, + 387, + 482, + 472 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{\\mathrm{skip}}" + }, + { + "bbox": [ + 130, + 387, + 482, + 472 + ], + "type": "text", + "content": " and ControlNet feature " + }, + { + "bbox": [ + 130, + 387, + 482, + 472 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{\\mathrm{control}}" + }, + { + "bbox": [ + 130, + 387, + 482, + 472 + ], + "type": "text", + "content": ". To allow for adaptive alignment of the above features, here we propose a new modulation fusion layer to fuse the multi-scale features from ControlNet to the skip connection features of the frozen backbone: " + }, + { + "bbox": [ + 130, + 387, + 482, + 472 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{f}}_{\\mathrm{skip}} = (1 + \\gamma)\\pmb{f}_{\\mathrm{skip}} + \\beta" + }, + { + "bbox": [ + 130, + 387, + 482, + 472 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 130, + 387, + 482, + 472 + ], + "type": "inline_equation", + "content": "\\gamma, \\beta = \\mathcal{M}(\\pmb{f}_{\\mathrm{con}})" + }, + { + "bbox": [ + 130, + 387, + 482, + 472 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 387, + 482, + 472 + ], + "type": "inline_equation", + "content": "\\gamma, \\beta" + }, + { + "bbox": [ + 130, + 387, + 482, + 472 + ], + "type": "text", + "content": " are the scaling and bias parameters from " + }, + { + "bbox": [ + 130, + 387, + 482, + 472 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 130, + 387, + 482, + 472 + ], + "type": "text", + "content": " a lightweight " + }, + { + "bbox": [ + 130, + 387, + 482, + 472 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 130, + 387, + 482, + 472 + ], + "type": "text", + "content": " zero-initialized convolution layer." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 479, + 230, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 230, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 230, + 493 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 502, + 377, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 377, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 377, + 514 + ], + "type": "text", + "content": "4.1 Text-based Training Data and Benchmarks" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 521, + 480, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 480, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 480, + 581 + ], + "type": "text", + "content": "Our method opens up an entirely new paradigm of instruction-based image restoration. However, existing image restoration datasets such as DIV2K [2] and Flickr2K [40] do not provide high-quality semantic prompts, and Real-ESRGAN degradation is for blind restoration. To address this, we construct a new setting including training data generation and test benchmarks for evaluation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 581, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 481, + 666 + ], + "type": "text", + "content": "Our parameterized degradation pipeline is based on the Real-ESRGAN [71] which contains a large number of degradation types. Since the original full Real-ESRGAN degradation is difficult to be parameterized and represented as user-friendly natural language (e.g., not every one understands \"anisotropic plateau blur with beta " + }, + { + "bbox": [ + 130, + 581, + 481, + 666 + ], + "type": "inline_equation", + "content": "2.0^{\\prime \\prime}" + }, + { + "bbox": [ + 130, + 581, + 481, + 666 + ], + "type": "text", + "content": "), we choose the 4 most general degradations which are practical for parameterization to support degradation parameter-aware restoration, as shown in Tab. 2. Our parameterized pipeline skips each degradation stage" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 233, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 233, + 91, + 447, + 102 + ], + "type": "text", + "content": "SPIRE: Semantic Prompt-Driven Image Restoration" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 479, + 334 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 479, + 334 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 479, + 334 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 479, + 334 + ], + "type": "image", + "image_path": "66e49157111f2c87df01cd69bcc0e613dd7448addd03ada21cab04e71c31a9f9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 335, + 482, + 390 + ], + "lines": [ + { + "bbox": [ + 130, + 335, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 482, + 390 + ], + "type": "text", + "content": "Fig. 4: Visual Comparison with other baselines. Our method of integrating both the semantic prompt " + }, + { + "bbox": [ + 130, + 335, + 482, + 390 + ], + "type": "inline_equation", + "content": "\\pmb{c}_{s}" + }, + { + "bbox": [ + 130, + 335, + 482, + 390 + ], + "type": "text", + "content": " and the restoration prompt " + }, + { + "bbox": [ + 130, + 335, + 482, + 390 + ], + "type": "inline_equation", + "content": "\\pmb{c}_{r}" + }, + { + "bbox": [ + 130, + 335, + 482, + 390 + ], + "type": "text", + "content": " outperforms image-to-image restoration (DiffBIR, Retrained ControlNet-SR) and naive zero-shot combination with semantic prompt. It achieves more sharp and clean results while maintaining consistency with the degraded image." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 403, + 482, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 513 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 513 + ], + "type": "text", + "content": "with a " + }, + { + "bbox": [ + 130, + 403, + 482, + 513 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 403, + 482, + 513 + ], + "type": "text", + "content": " probability to increase the diversity of restoration prompts and support single task restoration. The descriptions of the selected degradations are appended following the order of the image degradations to synthesize " + }, + { + "bbox": [ + 130, + 403, + 482, + 513 + ], + "type": "inline_equation", + "content": "\\pmb{c}_r" + }, + { + "bbox": [ + 130, + 403, + 482, + 513 + ], + "type": "text", + "content": ". In real application, users can control the restoration type and its strength by modifying restoration prompt. As shown in third column of Fig. 3, driven by restoration prompt \"Deblur with sigma " + }, + { + "bbox": [ + 130, + 403, + 482, + 513 + ], + "type": "inline_equation", + "content": "3.0^{\\prime \\prime}" + }, + { + "bbox": [ + 130, + 403, + 482, + 513 + ], + "type": "text", + "content": ", our framework removes the Gaussian blur, while preserving motion blur. The representation of the restoration strength is also general enough to decouple different tasks in natural language as shown in Fig. 7. More visual results are presented in our supplement." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 516, + 482, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 516, + 482, + 589 + ], + "spans": [ + { + "bbox": [ + 130, + 516, + 482, + 589 + ], + "type": "text", + "content": "Training data construction. In the training stage, we sample near 100 million text-image pairs from internal data source and use the alternative text label with highest relevance as the semantic prompt " + }, + { + "bbox": [ + 130, + 516, + 482, + 589 + ], + "type": "inline_equation", + "content": "\\pmb{c}_s" + }, + { + "bbox": [ + 130, + 516, + 482, + 589 + ], + "type": "text", + "content": " in the framework. We drop out the semantic prompt to empty string " + }, + { + "bbox": [ + 130, + 516, + 482, + 589 + ], + "type": "inline_equation", + "content": "\\varnothing" + }, + { + "bbox": [ + 130, + 516, + 482, + 589 + ], + "type": "text", + "content": " by a probability of " + }, + { + "bbox": [ + 130, + 516, + 482, + 589 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 516, + 482, + 589 + ], + "type": "text", + "content": " to support semantic-agnostic restoration. We mix the Real-ESRGAN degradation pipeline and our parameterized pipeline by " + }, + { + "bbox": [ + 130, + 516, + 482, + 589 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 516, + 482, + 589 + ], + "type": "text", + "content": " each in training iterations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Evaluation Setting. We randomly choose 3000 pairs of image " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " and semantic prompt " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{c}_{s}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " from the MS-COCO [33] validation set. Then, we build two test sets on parameterized and the Real-ESRGAN processes separately. At the left half of Tab. 1, images are sent to the parameterized degradation to form the synthesize degradation image " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " with restoration prompt " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{c}_{r}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ". In this setting, we expect \"Ours\" model to fully utilize the " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{c}_{s}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " describing the semantic and the " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{c}_{r}" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 201, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 201, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 201, + 102 + ], + "type": "text", + "content": "Qi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 114, + 375, + 178 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 375, + 178 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 375, + 178 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 375, + 178 + ], + "type": "table", + "html": "
MethodFID↓LPIPS↓PSNR↑SSIM↑CLIP-I↑
Real-ESRGAN [72]32.370.31222.520.6460.683
DiffBIR [34] (zero-shot)30.710.35422.010.5260.921
StableSR24.440.31121.620.5330.928
Ours w/o text (zero-shot)28.800.35221.680.5490.927
Ours w/o text (finetuned)22.450.32121.380.5320.932
", + "image_path": "4b37870e6839052683d07b4ef8fe6dec31d8d89f67cdfe5fd9a3b27f68f26705.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 377, + 127, + 468, + 172 + ], + "lines": [ + { + "bbox": [ + 377, + 127, + 468, + 172 + ], + "spans": [ + { + "bbox": [ + 377, + 127, + 468, + 172 + ], + "type": "text", + "content": "Table 3: Numerical results on the DIV2K testset without any prompt." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 190, + 482, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 190, + 482, + 275 + ], + "spans": [ + { + "bbox": [ + 130, + 190, + 482, + 275 + ], + "type": "text", + "content": "describing the degradations. On the right half, the same images are fed to the Real-ESRGAN degradation—so the degradation is aligned with open-sourced checkpoints [30, 34, 67]. Because not every model can utilize both the semantic " + }, + { + "bbox": [ + 130, + 190, + 482, + 275 + ], + "type": "inline_equation", + "content": "(c_{s})" + }, + { + "bbox": [ + 130, + 190, + 482, + 275 + ], + "type": "text", + "content": " and restoration " + }, + { + "bbox": [ + 130, + 190, + 482, + 275 + ], + "type": "inline_equation", + "content": "(c_{r})" + }, + { + "bbox": [ + 130, + 190, + 482, + 275 + ], + "type": "text", + "content": " information, Tab. 1 provides a “Prompts” column denoting what information the model relied on. In summary, our constructed benchmarks based on MS-COCO can serve multiple purposes—from fully-blind, content-prompted, to all-prompted restoration." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 277, + 482, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 482, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 482, + 326 + ], + "type": "text", + "content": "Evaluation Metrics. We use the generative metric FID [22] and perceptual metric LPIPS [86] for quantitative evaluation of image quality. PSNR, and SSIM are reported for reference. We also evaluate similarity scores in the clip embedding space with ground truth image (CLIP-I) and caption (CLIP-T)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 331, + 291, + 344 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 331, + 291, + 344 + ], + "spans": [ + { + "bbox": [ + 132, + 331, + 291, + 344 + ], + "type": "text", + "content": "4.2 Comparing with baselines" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 350, + 392, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 350, + 392, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 350, + 392, + 363 + ], + "type": "text", + "content": "We compare our method with three categories of baselines:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 369, + 487, + 468 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 132, + 369, + 481, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 481, + 392 + ], + "type": "text", + "content": "- Open-sourced image restoration checkpoints, including the regression model SwinIR [30] and the latent diffusion models StableSR [67] and DiffBIR [34]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 394, + 482, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 394, + 482, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 394, + 482, + 418 + ], + "type": "text", + "content": "Combining the state-of-the-art method [34] with the zero-shot post-editing [37] (DiffBIR+SDEdit) or zero-shot injection through CLIP [46] (DiffBIR + CLIP)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 419, + 487, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 419, + 487, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 419, + 487, + 468 + ], + "type": "text", + "content": "- ControlNet-SR (retrained): We adapt the image-to-image ControlNet model [84] for super-resolution. For consistency, we maintain the same training iterations as our own model. We also evaluate its zero-shot capabilities with CLIP guidance (ControlNet-SR + CLIP)." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "content": "Quantitative comparison with baselines is presented in Tab. 1. On the left, we evaluate our full model with the parameterized degradation test set. Since open-sourced baselines are pretrained on Real-ESRGAN degradation, not perfectly aligned with our parameterized degradation, we also evaluate our full model on MS-COCO with Real-ESRGAN degradation by setting our restoration prompt to \"Remove all degradation\". Thanks to our prompts guidance and architecture improvements, our full model achieves best FID, LPIPS, CLIP-Image score, which means better image quality and semantic restoration. \"Ours w/o text\" is the same checkpoint conditioned on only degradation image and has high pixel-level similarity (the second best PSNR and SSIM) with GT. Although SwinIR has highest PSNR, its visual results are blurry. Although combining semantic prompt in zero-shot way can bring marginal improvement in FID and CLIP similarity with caption, it deteriorates the image restoration capacity and results in worse LPIPS and CLIP-Image. In contrast, semantic prompt guidance improves the CLIP image similarity of our full model. Tab. 3 shows the evaluation results of our model on DIV2K test set provided by StableSR. In zero-shot" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 233, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 233, + 91, + 447, + 102 + ], + "type": "text", + "content": "SPIRE: Semantic Prompt-Driven Image Restoration" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 105, + 193, + 202 + ], + "blocks": [ + { + "bbox": [ + 135, + 105, + 193, + 202 + ], + "lines": [ + { + "bbox": [ + 135, + 105, + 193, + 202 + ], + "spans": [ + { + "bbox": [ + 135, + 105, + 193, + 202 + ], + "type": "image", + "image_path": "6a56b5abf87c9b569a850dcd4be82b8553d20c5dc8c202a19b17cd96a09623f0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 201, + 479, + 222 + ], + "lines": [ + { + "bbox": [ + 132, + 201, + 479, + 222 + ], + "spans": [ + { + "bbox": [ + 132, + 201, + 479, + 222 + ], + "type": "text", + "content": "Fig. 5: Real-world image restorations. Our framework enhances real-world low-quality images following semantic prompts provided by language models or user input." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 194, + 105, + 250, + 202 + ], + "blocks": [ + { + "bbox": [ + 194, + 105, + 250, + 202 + ], + "lines": [ + { + "bbox": [ + 194, + 105, + 250, + 202 + ], + "spans": [ + { + "bbox": [ + 194, + 105, + 250, + 202 + ], + "type": "image", + "image_path": "05822db3db402b76fd6021faf22a65a78216020cd316b8037478e56b720c551d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 250, + 105, + 307, + 202 + ], + "blocks": [ + { + "bbox": [ + 250, + 105, + 307, + 202 + ], + "lines": [ + { + "bbox": [ + 250, + 105, + 307, + 202 + ], + "spans": [ + { + "bbox": [ + 250, + 105, + 307, + 202 + ], + "type": "image", + "image_path": "6e844126cba331decf626d3ed980c7474cde6a85d75940a1b6a24bdf21dbd48b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 105, + 364, + 202 + ], + "blocks": [ + { + "bbox": [ + 307, + 105, + 364, + 202 + ], + "lines": [ + { + "bbox": [ + 307, + 105, + 364, + 202 + ], + "spans": [ + { + "bbox": [ + 307, + 105, + 364, + 202 + ], + "type": "image", + "image_path": "3be6e31444462b912c1799d410d299cfd82cc11a88d9255a710182a071a9cfc1.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 364, + 105, + 421, + 202 + ], + "blocks": [ + { + "bbox": [ + 364, + 105, + 421, + 202 + ], + "lines": [ + { + "bbox": [ + 364, + 105, + 421, + 202 + ], + "spans": [ + { + "bbox": [ + 364, + 105, + 421, + 202 + ], + "type": "image", + "image_path": "92af50cf89b62bb95f8a749107424572e5ee301e75240dac4e9f8201fcda6fd2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 421, + 105, + 478, + 202 + ], + "blocks": [ + { + "bbox": [ + 421, + 105, + 478, + 202 + ], + "lines": [ + { + "bbox": [ + 421, + 105, + 478, + 202 + ], + "spans": [ + { + "bbox": [ + 421, + 105, + 478, + 202 + ], + "type": "image", + "image_path": "3af1353d07f3adb4aa2500e3acd0378d57a4618fc6c1f6c4cc2aae72441eb47e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 133, + 225, + 480, + 277 + ], + "blocks": [ + { + "bbox": [ + 133, + 225, + 480, + 277 + ], + "lines": [ + { + "bbox": [ + 133, + 225, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 133, + 225, + 480, + 277 + ], + "type": "table", + "html": "
MetricSwinIR [30]Diffbir [34]Ablation of restoration prompt crAblation of semantic prompt cs
Sem c_sXXXXshortunmatchedfull
Res crXX"Deblur 0.6"... 1.0"... 2.0"... 5.0""Remove all degradation"
CLIP-IQA↑0.61900.69830.40740.41590.48310.47970.58180.57500.64600.6610
MUSIQ↑63.6569.6935.8239.8749.9848.8161.6860.2767.5467.82
MANIQA↑0.40350.26190.22990.24870.29700.28210.36940.36370.43550.4474
", + "image_path": "3cbbeef7554bbac8caf8f61b520e12c68da362a9a81df55ddb9b992ec51424ed.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 278, + 479, + 300 + ], + "lines": [ + { + "bbox": [ + 132, + 278, + 479, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 479, + 300 + ], + "type": "text", + "content": "Table 4: Quantitative results on real-world images. Our aesthetic quality is improved when increasing restoration strength or using a more accurate and detailed prompt." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 311, + 480, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 311, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 311, + 480, + 346 + ], + "type": "text", + "content": "test setting, our model has lower FID and LPIPS than DiffBIR. To fairly compare with StableSR, we finetune our model on their training set. Our finetuned model also outperforms StableSR in FID and CLIP-I." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 350, + 480, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 480, + 469 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 480, + 469 + ], + "type": "text", + "content": "Qualitative comparison with baselines is presented in Fig. 4, where the corresponding semantic prompt is provided below each row. Image-to-image baselines such as DiffBir and the retrained ControlNet-SR can easily generate artifacts (e.g., hydrant in the first row) and blurry images (e.g., hands example in third row) that may look close to an \"average\" estimation. Besides, naive combination with semantic prompt in the zero-shot approach results in heavy hallucinations and may fail to preserve the object identity in the input image (e.g., giraffes on the second row). Unlike the existing methods, our full model considers semantic prompt, degradation image and restoration prompt in both training and test stages, which makes its results more aligned with all conditions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 489, + 273, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 489, + 273, + 499 + ], + "spans": [ + { + "bbox": [ + 132, + 489, + 273, + 499 + ], + "type": "text", + "content": "4.3 Real-world restoration" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 510, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 481, + 666 + ], + "type": "text", + "content": "Besides evaluation on synthetic degradations, we present an analysis on real-world images in Tab. 4 and Fig. 5. We choose RealPhoto60 [78] as our real-world test-set, and use non-reference metrics CLIP-IQA [66], MUSIQ [27] and MANIQA [76] scores to study the influence of the semantic and restoration prompts. First, we fix the semantic prompt to the empty string " + }, + { + "bbox": [ + 130, + 510, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\varnothing" + }, + { + "bbox": [ + 130, + 510, + 481, + 666 + ], + "type": "text", + "content": " and conduct an ablation study on the restoration prompt. As the deblurring strength increases from 0.6 to 2.0, the image quality is improved consistently. A blind prompt \"Remove all degradation\" can provide better results than the manually controlled restoration prompts. This showcases that the proposed model is flexible to work with generic restoration prompts, especially when coming up with the perfect prompt is challenging. We also fix the blind restoration prompt and change the semantic prompts. Following [78], we feed degraded images to LLaVA [35] and generate corresponding semantic prompts. Our experiments" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 201, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 201, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 201, + 101 + ], + "type": "text", + "content": "Qi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 114, + 479, + 241 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 479, + 241 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 479, + 241 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 479, + 241 + ], + "type": "image", + "image_path": "daa0906e8b3bd0788da4ba27554dc6e6bac528f5d521d1c2e6758519bfe02e20.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 242, + 480, + 275 + ], + "lines": [ + { + "bbox": [ + 132, + 242, + 480, + 275 + ], + "spans": [ + { + "bbox": [ + 132, + 242, + 480, + 275 + ], + "type": "text", + "content": "Fig. 6: Test-time semantic prompting. Our framework restores degraded images guided by flexible semantic prompts, while unrelated background elements and global tones remain aligned with the degraded input conditioning." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 285, + 359, + 480 + ], + "blocks": [ + { + "bbox": [ + 133, + 285, + 359, + 480 + ], + "lines": [ + { + "bbox": [ + 133, + 285, + 359, + 480 + ], + "spans": [ + { + "bbox": [ + 133, + 285, + 359, + 480 + ], + "type": "image", + "image_path": "7de3d5beec6f58dec1bbbf7714b1fe10b6a26feadada030da3323936d31e7f47.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 361, + 288, + 482, + 462 + ], + "lines": [ + { + "bbox": [ + 361, + 288, + 482, + 462 + ], + "spans": [ + { + "bbox": [ + 361, + 288, + 482, + 462 + ], + "type": "text", + "content": "Fig.7: Prompt space walking visualization for the restoration prompt. Given the same degraded input (upper left) and empty semantic prompt " + }, + { + "bbox": [ + 361, + 288, + 482, + 462 + ], + "type": "inline_equation", + "content": "\\varnothing" + }, + { + "bbox": [ + 361, + 288, + 482, + 462 + ], + "type": "text", + "content": ", our method can decouple the restoration direction and strength via only prompting the quantitative number in natural language. An interesting finding is that our model learns a continuous range of restoration strength from discrete language tokens." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 498, + 482, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 679 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 679 + ], + "type": "text", + "content": "show that detailed full descriptions (about 60 words) of the degraded image from LLaVA can be more effective to improve the visual results than short prompts that have less than 20 words. Unmatched prompts simulate the case when a user provides wrong prompts by randomly re-sampling unmatched long prompts from the data set. An interesting observation is that unmatched long prompts also improve the aesthetic quality of images against a short prompt or a blind restoration. Nonetheless, it can lead to more \"hallucination\" when the semantic input prompt is unmatched with the real ground truth. In contrast to concurrent work [62] that leverages implicit clip embedding, our text-based formulation provides better control of hallucinations by adjusting the LLM-synthesized or manually-designed prompt at test-time. Note that many studies [6,18] find current image restoration metrics are biased towards specific datasets and not fully aligned with human's visual perception. The quantitative results of previous works [34,71] are provided for reference. More results and discussions are provided in the supplements." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 234, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 234, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 234, + 91, + 447, + 102 + ], + "type": "text", + "content": "SPIRE: Semantic Prompt-Driven Image Restoration" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 151, + 114, + 462, + 181 + ], + "blocks": [ + { + "bbox": [ + 151, + 114, + 462, + 181 + ], + "lines": [ + { + "bbox": [ + 151, + 114, + 462, + 181 + ], + "spans": [ + { + "bbox": [ + 151, + 114, + 462, + 181 + ], + "type": "table", + "html": "
MethodSem csRes crFID↓LPIPS↓PSNR↑SSIM↑CLIP-I↑CLIP-T↑
ControlNet-SRXX13.650.22223.750.6690.9380.300
ControlNet-SR✓✓12.140.22323.500.6600.9400.306
Ours✓wo param11.580.22323.610.6650.9420.305
Ours✓✓11.340.21923.610.6650.9430.305
", + "image_path": "738f300f4302f7b4c4e52c23a9e229763b0b817b25f1daef50abcaafb696346b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 172, + 181, + 441, + 191 + ], + "lines": [ + { + "bbox": [ + 172, + 181, + 441, + 191 + ], + "spans": [ + { + "bbox": [ + 172, + 181, + 441, + 191 + ], + "type": "text", + "content": "Table 5: Ablation of architecture and degradation strength in " + }, + { + "bbox": [ + 172, + 181, + 441, + 191 + ], + "type": "inline_equation", + "content": "{c}_{r}" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 132, + 195, + 274, + 208 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 195, + 274, + 208 + ], + "spans": [ + { + "bbox": [ + 132, + 195, + 274, + 208 + ], + "type": "text", + "content": "4.4 Prompting the SPIRE" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 212, + 481, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 212, + 481, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 481, + 272 + ], + "type": "text", + "content": "Semantic Prompting. Our model can be driven by user-defined semantic prompts. As shown in Fig. 6, the blind restoration with empty string generates ambiguous and unrealistic identity. In contrast, our framework reconstructs sharp and realistic results. The object identity accurately follows user's semantic prompt, while the global layout and color tone remain consistent with input." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 272, + 482, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 272, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 482, + 392 + ], + "type": "text", + "content": "Restoration Prompting. Users also have the freedom to adjust the degradation types and strengths in the restoration prompts. As shown in Fig. 7, the denoising and deblurring restoration are well decoupled. Using the task-specific prompts, our method can get a clean but blur (second row), or sharp but noisy (third row) result. In addition, our model learns the continuous space of restoration as the result becomes cleaner progressively if we modify the \"Denoise with sigma 0.06\" to \"0.24\". Beyond purely relying on numerical description, our model \"understand\" the difference of strength definition for tasks, such as the range difference of denoise sigma and deblur sigma, which makes it promising to be a universal representation for restoration strength control." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 406, + 238, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 406, + 238, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 406, + 238, + 418 + ], + "type": "text", + "content": "4.5 Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 423, + 481, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 481, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 481, + 496 + ], + "type": "text", + "content": "We conduct several ablations in Tab. 5. Providing both prompts " + }, + { + "bbox": [ + 130, + 423, + 481, + 496 + ], + "type": "inline_equation", + "content": "c_s" + }, + { + "bbox": [ + 130, + 423, + 481, + 496 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 423, + 481, + 496 + ], + "type": "inline_equation", + "content": "c_r" + }, + { + "bbox": [ + 130, + 423, + 481, + 496 + ], + "type": "text", + "content": " to \"ControlNet-SR\" baseline improves the FID by 1.5. Adding the modulation fusion layer (\"Ours\") further improves the generative quality (0.8 for FID) and pixel-level similarity (0.11 dB). Embedding degradation parameters in the restoration prompt, not only enables restoration strength prompting (Fig. 7), but also improves image quality (FID and LPIPS). More details are in the supplement." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 511, + 220, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 511, + 220, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 511, + 220, + 525 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "We have presented a unified text-driven framework for instructed image restoration using semantic and restoration prompts. We design our model in a decoupled way to better preserve the semantic text-to-image generative prior while efficiently learning to control both the restoration direction and its strength. To the best of our knowledge, this is the first framework to support both semantic and parameter-embedded restoration instructions simultaneously, allowing users to flexibly prompt-tune the results up to their expectations. Our extensive experiments have shown that our method significantly outperforms prior works in terms of both quantitative and qualitative results. Our model and evaluation benchmark have established a new paradigm of instruction-based image restoration, paving the way for further multi-modality generative imaging applications." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 202, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 202, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 202, + 102 + ], + "type": "text", + "content": "Qi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "type": "text", + "content": "1. Abuolaim, A., Delbracio, M., Kelly, D., Brown, M.S., Milanfar, P.: Learning to reduce defocus blur by realistically modeling dual-pixel data. In: ICCV. pp. 2289-2298 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 173, + 480, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 173, + 480, + 227 + ], + "spans": [ + { + "bbox": [ + 138, + 173, + 480, + 227 + ], + "type": "text", + "content": "2. Agustsson, E., Timofte, R.: NTIRE 2017 challenge on single image superresolution: Dataset and study. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition Workshops, CVPR Workshops 2017, Honolulu, HI, USA, July 21-26, 2017. pp. 1122-1131. IEEE Computer Society (2017). https://doi.org/10.1109/CVPRW.2017.150, https://doi.org/10.1109/CVPRW.2017.150" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 228, + 480, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 480, + 249 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 480, + 249 + ], + "type": "text", + "content": "3. Avrahami, O., Fried, O., Lischinski, D.: Blended latent diffusion. ACM Transactions on Graphics (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 250, + 480, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 250, + 480, + 271 + ], + "spans": [ + { + "bbox": [ + 138, + 250, + 480, + 271 + ], + "type": "text", + "content": "4. Avrahami, O., Lischinski, D., Fried, O.: Blended diffusion for text-driven editing of natural images. In: CVPR. pp. 18208-18218 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 271, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 271, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 138, + 271, + 480, + 303 + ], + "type": "text", + "content": "5. Bai, Y., Wang, C., Xie, S., Dong, C., Yuan, C., Wang, Z.: Textir: A simple framework for text-based editable image restoration. arXiv preprint arXiv:2302.14736 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 304, + 480, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 304, + 480, + 326 + ], + "spans": [ + { + "bbox": [ + 138, + 304, + 480, + 326 + ], + "type": "text", + "content": "6. Blau, Y., Michaeli, T.: The perception-distortion tradeoff. In: CVPR. pp. 6228-6237 (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 326, + 480, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 326, + 480, + 347 + ], + "spans": [ + { + "bbox": [ + 138, + 326, + 480, + 347 + ], + "type": "text", + "content": "7. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: CVPR (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 348, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 348, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 138, + 348, + 480, + 413 + ], + "type": "text", + "content": "8. Chen, X., Wang, X., Changpinyo, S., Piergiovanni, A., Padlewski, P., Salz, D., Goodman, S.A., Grycner, A., Mustafa, B., Beyer, L., Kolesnikov, A., Puigcerver, J., Ding, N., Rong, K., Akbari, H., Mishra, G., Xue, L., Thapliyal, A., Bradbury, J., Kuo, W., Seyedhosseini, M., Jia, C., Ayan, B.K., Riquelme, C., Steiner, A., Angelova, A., Zhai, X., Houlsby, N., Soricut, R.: Pali: A jointly-scaled multilingual language-image model. In: ICLR (2023), https://arxiv.org/abs/2209.06794" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 414, + 480, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 414, + 480, + 436 + ], + "spans": [ + { + "bbox": [ + 138, + 414, + 480, + 436 + ], + "type": "text", + "content": "9. Chen, Z., Zhang, Y., Gu, J., Yuan, X., Kong, L., Chen, G., Yang, X.: Image superresolution with text prompt diffusion. arXiv preprint arXiv:2303.06373 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 436, + 480, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 436, + 480, + 457 + ], + "spans": [ + { + "bbox": [ + 138, + 436, + 480, + 457 + ], + "type": "text", + "content": "0. Couairon, G., Verbeek, J., Schwenk, H., Cord, M.: Diffedit: Diffusion-based semantic image editing with mask guidance. In: ICLR (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 458, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 458, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 138, + 458, + 480, + 490 + ], + "type": "text", + "content": "1. Delbracio, M., Milanfar, P.: Inversion by direct iteration: An alternative to denoising diffusion for image restoration. Transactions on Machine Learning Research (2023), https://openreview.net/forum?id=VmyFF5lL3F, featured Certification" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 491, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 491, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 138, + 491, + 480, + 522 + ], + "type": "text", + "content": "2. Delbracio, M., Talebei, H., Milanfar, P.: Projected distribution loss for image enhancement. In: 2021 IEEE International Conference on Computational Photography (ICCP). pp. 1-12. IEEE (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 523, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 523, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 523, + 480, + 544 + ], + "type": "text", + "content": "3. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Neural Information Processing Systems (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 545, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 545, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 545, + 480, + 567 + ], + "type": "text", + "content": "4. Dong, C., Loy, C.C., He, K., Tang, X.: Image super-resolution using deep convolutional networks. IEEE Trans. Pattern Anal. Mach. Intell. 38(2), 295-307 (2016)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 567, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 480, + 599 + ], + "type": "text", + "content": "5. Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A.H., Chechik, G., Cohen-Or, D.: An image is worth one word: Personalizing text-to-image generation using textual inversion. In: ICLR (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 600, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 600, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 138, + 600, + 480, + 621 + ], + "type": "text", + "content": "6. Galteri, L., Seidenari, L., Bertini, M., Bimbo, A.: Deep generative adversarial compression artifact removal. In: ICCV (2017)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 622, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 622, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 622, + 480, + 665 + ], + "type": "text", + "content": "7. Geng, Z., Yang, B., Hang, T., Li, C., Gu, S., Zhang, T., Bao, J., Zhang, Z., Hu, H., Chen, D., Guo, B.: Instructdiffusion: A generalist modeling interface for vision tasks. CoRR abs/2309.03895 (2023). https://doi.org/10.48550/arXiv.2309.03895, https://doi.org/10.48550/arXiv.2309.03895" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 234, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 234, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 234, + 91, + 447, + 102 + ], + "type": "text", + "content": "SPIRE: Semantic Prompt-Driven Image Restoration" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 215 + ], + "type": "text", + "content": "18. Gu, J., Cai, H., Dong, C., Ren, J.S., Timofte, R., Gong, Y., Lao, S., Shi, S., Wang, J., Yang, S., Wu, T., Xia, W., Yang, Y., Cao, M., Heng, C., Fu, L., Zhang, R., Zhang, Y., Wang, H., Song, H., Wang, J., Fan, H., Hou, X., Sun, M., Li, M., Zhao, K., Yuan, K., Kong, Z., Wu, M., Zheng, C., Conde, M.V., Burchi, M., Feng, L., Zhang, T., Li, Y., Xu, J., Wang, H., Liao, Y., Li, J., Xu, K., Sun, T., Xiong, Y., Keshari, A., Komal, Thakur, S., Jakhetiya, V., Subudhi, B.N., Yang, H.H., Chang, H.E., Huang, Z.K., Chen, W.T., Kuo, S.Y., Dutta, S., Das, S.D., Shah, N.A., Tiwari, A.K.: Ntire 2022 challenge on perceptual image quality assessment. In: CVPRW. pp. 951-967 (June 2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 216, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 216, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 133, + 216, + 481, + 237 + ], + "type": "text", + "content": "19. Gu, X., Lin, T.Y., Kuo, W., Cui, Y.: Open-vocabulary object detection via vision and language knowledge distillation. In: ICLR (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 237, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 481, + 259 + ], + "type": "text", + "content": "20. Han, L., Li, Y., Zhang, H., Milanfar, P., Metaxas, D., Yang, F.: Svdiff: Compact parameter space for diffusion fine-tuning. In: ICCV. pp. 7323-7334 (October 2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "text", + "content": "21. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-or, D.: Prompt-to-prompt image editing with cross-attention control. In: ICLR (2023), https://openreview.net/forum?id=CDixzkzeyb" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 293, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 325 + ], + "type": "text", + "content": "22. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: NeurIPS (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 325, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 481, + 346 + ], + "type": "text", + "content": "23. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. NeurIPS 33, 6840-6851 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "text", + "content": "24. Ho, J., Salimans, T.: Classifier-free diffusion guidance. In: NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications (2021), https://openreview.net/forum?id=qw8AKxfYbI" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "text", + "content": "25. Hu, E.J., yelong shen, Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-rank adaptation of large language models. In: ICLR (2022), https://openreview.net/forum?id=nZeVKeeFYf9" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 413, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 481, + 435 + ], + "type": "text", + "content": "26. Jiang, Y., Zhang, Z., Xue, T., Gu, J.: Autodir: Automatic all-in-one image restoration with latent diffusion. arXiv preprint arXiv:2310.10123 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "type": "text", + "content": "27. Ke, J., Wang, Q., Wang, Y., Milanfar, P., Yang, F.: Musiq: Multi-scale image quality transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5148-5157 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 468, + 481, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 500 + ], + "type": "text", + "content": "28. Ke, J., Ye, K., Yu, J., Wu, Y., Milanfar, P., Yang, F.: Vila: Learning image aesthetics from user comments with vision-language pretraining. In: CVPR. pp. 10041-10051 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 500, + 481, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 500, + 481, + 523 + ], + "spans": [ + { + "bbox": [ + 132, + 500, + 481, + 523 + ], + "type": "text", + "content": "29. Kumari, N., Zhang, B., Zhang, R., Shechtman, E., Zhu, J.Y.: Multi-concept customization of text-to-image diffusion. In: CVPR. pp. 1931-1941 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 523, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 481, + 544 + ], + "type": "text", + "content": "30. Liang, J., Cao, J., Sun, G., Zhang, K., Gool, L.V., Timofte, R.: Swinir: Image restoration using swin transformer. In: Proceedings of ICCV Workshops (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 545, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 481, + 567 + ], + "type": "text", + "content": "31. Liang, Z., Li, C., Zhou, S., Feng, R., Loy, C.C.: Iterative prompt learning for unsupervised backlit image enhancement. In: ICCV. pp. 8094-8103 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 567, + 481, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 589 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 589 + ], + "type": "text", + "content": "32. Lim, B., Son, S., Kim, H., Nah, S., Lee, K.M.: Enhanced deep residual networks for single image super-resolution. In: Proceedings of CVPR Workshops (2017)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 589, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 481, + 632 + ], + "type": "text", + "content": "33. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, L.: Microsoft coco: Common objects in context. In: ECCV. ECCV (September 2014), https://www.microsoft.com/en-us/research/publication/microsoft-coco-common-objects-in-context/" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "34. Lin, X., He, J., Chen, Z., Lyu, Z., Fei, B., Dai, B., Ouyang, W., Qiao, Y., Dong, C.: Diffbir: Towards blind image restoration with generative diffusion prior. arXiv preprint arXiv:2308.15070 (2023)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 201, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 201, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 201, + 102 + ], + "type": "text", + "content": "Qi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 132, + 116, + 475, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 475, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 475, + 127 + ], + "type": "text", + "content": "35. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: NeurIPS (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 128, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 128, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 128, + 480, + 171 + ], + "type": "text", + "content": "36. Luo, Z., Gustafsson, F.K., Zhao, Z., Sjolund, J., Schon, T.B.: Controlling vision-language models for multi-task image restoration. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=t3vnnLeajU" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "type": "text", + "content": "37. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: SDEdit: Guided image synthesis and editing with stochastic differential equations. In: ICLR (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 225 + ], + "type": "text", + "content": "38. Mokady, R., Hertz, A., Aberman, K., Pritch, Y., Cohen-Or, D.: Null-text inversion for editing real images using guided diffusion models. In: CVPR. pp. 6038-6047 (June 2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "text", + "content": "39. Mou, C., Wang, X., Xie, L., Zhang, J., Qi, Z., Shan, Y., Qie, X.: T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 402 + ], + "type": "text", + "content": "40. Nah, S., Timofte, R., Baik, S., Hong, S., Moon, G., Son, S., Lee, K.M., Wang, X., Chan, K.C.K., Yu, K., Dong, C., Loy, C.C., Fan, Y., Yu, J., Liu, D., Huang, T.S., Sim, H., Kim, M., Park, D., Kim, J., Chun, S.Y., Haris, M., Shakhnarovich, G., Ukita, N., Zamir, S.W., Arora, A., Khan, S.H., Khan, F.S., Shao, L., Gupta, R.K., Chudasama, V.M., Patel, H., Upla, K.P., Fan, H., Li, G., Zhang, Y., Li, X., Zhang, W., He, Q., Purohit, K., Rajagopalan, A.N., Kim, J., Tofighi, M., Guo, T., Monga, V.: NTIRE 2019 challenge on video deblurring: Methods and results. In: IEEE Conference on Computer Vision and Pattern Recognition Workshops, CVPR Workshops 2019, Long Beach, CA, USA, June 16-20, 2019. pp. 1974-1984. Computer Vision Foundation / IEEE (2019). https://doi.org/10.1109/CVPRW.2019.00249, http://openaccess.thecvf.com/content_CVPRW_2019/html/NTIRE/Nah_NTIRE_2019_Challenge_on_Video_Deblurring_Methods_and_Results_CVPRW_2019_paper.html" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 402, + 307, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 307, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 307, + 413 + ], + "type": "text", + "content": "41. OpenAI: Gpt-4 technical report (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 413, + 480, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 480, + 434 + ], + "type": "text", + "content": "42. Paiss, R., Chefer, H., Wolf, L.: No token left behind: Explainability-aided image classification and generation. In: ECCV (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 434, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 434, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 434, + 480, + 456 + ], + "type": "text", + "content": "43. Paiss, R., Ephrat, A., Tov, O., Zada, S., Mosseri, I., Irani, M., Dekel, T.: Teaching clip to count to ten. In: ICCV (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 456, + 480, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 456, + 480, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 456, + 480, + 489 + ], + "type": "text", + "content": "44. Parmar, G., Kumar Singh, K., Zhang, R., Li, Y., Lu, J., Zhu, J.Y.: Zero-shot image-to-image translation. In: ACM SIGGRAPH 2023 Conference Proceedings. pp. 1-11 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 490, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 480, + 522 + ], + "type": "text", + "content": "45. Prakash, M., Delbracio, M., Milanfar, P., Jug, F.: Interpretable unsupervised diversity denoising and artefact removal. In: ICLR (2022), https://openreview.net/forum?id=DfMqlBOPXjM" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 522, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 480, + 566 + ], + "type": "text", + "content": "46. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 566, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 566, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 566, + 480, + 588 + ], + "type": "text", + "content": "47. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners. OpenAI blog 1(8), 9 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 588, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 588, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 480, + 621 + ], + "type": "text", + "content": "48. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 621, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 480, + 643 + ], + "type": "text", + "content": "49. Ren, M., Delbracio, M., Talebi, H., Gerig, G., Milanfar, P.: Multiscale structure guided diffusion for image deblurring. In: ICCV. pp. 10721-10733 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 643, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 666 + ], + "type": "text", + "content": "50. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10684-10695 (2022)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 234, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 234, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 234, + 91, + 447, + 102 + ], + "type": "text", + "content": "SPIRE: Semantic Prompt-Driven Image Restoration" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 133, + 117, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 481, + 160 + ], + "type": "text", + "content": "51. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 161, + 481, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 481, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 481, + 193 + ], + "type": "text", + "content": "52. Sahara, C., Chan, W., Chang, H., Lee, C., Ho, J., Salimans, T., Fleet, D., Norouzi, M.: Palette: Image-to-image diffusion models. In: ACM SIGGRAPH 2022 Conference Proceedings. pp. 1-10 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 224 + ], + "type": "text", + "content": "53. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 225, + 481, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 225, + 481, + 257 + ], + "spans": [ + { + "bbox": [ + 132, + 225, + 481, + 257 + ], + "type": "text", + "content": "54. Sahara, C., Ho, J., Chan, W., Salimans, T., Fleet, D.J., Norouzi, M.: Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(4), 4713-4726 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 258, + 481, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 258, + 481, + 290 + ], + "spans": [ + { + "bbox": [ + 132, + 258, + 481, + 290 + ], + "type": "text", + "content": "55. Sahara, C., Ho, J., Chan, W., Salimans, T., Fleet, D.J., Norouzi, M.: Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(4), 4713-4726 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 290, + 481, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 290, + 481, + 322 + ], + "spans": [ + { + "bbox": [ + 132, + 290, + 481, + 322 + ], + "type": "text", + "content": "56. Salimans, T., Ho, J.: Progressive distillation for fast sampling of diffusion models. In: ICLR. OpenReview.net (2022), https://openreview.net/forum?id=TIdIXIpzhoI" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 323, + 481, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 323, + 481, + 343 + ], + "spans": [ + { + "bbox": [ + 132, + 323, + 481, + 343 + ], + "type": "text", + "content": "57. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 344, + 481, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 344, + 481, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 344, + 481, + 365 + ], + "type": "text", + "content": "58. Song, Y., Ermon, S.: Generative modeling by estimating gradients of the data distribution. In: NeurIPS. pp. 11895-11907 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 366, + 481, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 366, + 481, + 387 + ], + "spans": [ + { + "bbox": [ + 132, + 366, + 481, + 387 + ], + "type": "text", + "content": "59. Song, Y., Shen, L., Xing, L., Ermon, S.: Solving inverse problems in medical imaging with score-based generative models. In: ICLR. OpenReview.net (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 387, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 387, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 387, + 481, + 418 + ], + "type": "text", + "content": "60. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. In: ICLR (2021), https://openreview.net/forum?id=PxTIG12RRHS" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 419, + 481, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 419, + 481, + 440 + ], + "spans": [ + { + "bbox": [ + 132, + 419, + 481, + 440 + ], + "type": "text", + "content": "61. Su, S., Delbracio, M., Wang, J., Sapiro, G., Heidrich, W., Wang, O.: Deep video deblurring for hand-held cameras. In: CVPR. pp. 1279-1288 (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 441, + 481, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 441, + 481, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 441, + 481, + 472 + ], + "type": "text", + "content": "62. Sun, H., Li, W., Liu, J., Chen, H., Pei, R., Zou, X., Yan, Y., Yang, Y.: Coser: Bridging image and language for cognitive super-resolution. arXiv preprint arXiv:2311.16512 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 473, + 481, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 473, + 481, + 494 + ], + "spans": [ + { + "bbox": [ + 132, + 473, + 481, + 494 + ], + "type": "text", + "content": "63. Tu, Z., Talebi, H., Zhang, H., Yang, F., Milanfar, P., Bovik, A.C., Li, Y.: MAXIM: multi-axis MLP for image processing. In: CVPR (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 495, + 481, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 495, + 481, + 515 + ], + "spans": [ + { + "bbox": [ + 132, + 495, + 481, + 515 + ], + "type": "text", + "content": "64. Tumanyan, N., Geyer, M., Bagon, S., Dekel, T.: Plug-and-play diffusion features for text-driven image-to-image translation. In: CVPR. pp. 1921-1930 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 516, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 516, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 516, + 481, + 536 + ], + "type": "text", + "content": "65. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. NeurIPS 30 (2017)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 537, + 481, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 537, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 537, + 481, + 557 + ], + "type": "text", + "content": "66. Wang, J., Chan, K.C., Loy, C.C.: Exploring clip for assessing the look and feel of images. In: AAAI (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 558, + 481, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 558, + 481, + 579 + ], + "spans": [ + { + "bbox": [ + 132, + 558, + 481, + 579 + ], + "type": "text", + "content": "67. Wang, J., Yue, Z., Zhou, S., Chan, K.C., Loy, C.C.: Exploiting diffusion prior for real-world image super-resolution. arXiv preprint arXiv:2305.07015 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 580, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 580, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 580, + 481, + 601 + ], + "type": "text", + "content": "68. Wang, L., Wang, Y., Lin, Z., Yang, J., An, W., Guo, Y.: Learning a single network for scale-arbitrary super-resolution. In: ICCV. pp. 4801-4810 (2021)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 601, + 481, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 481, + 622 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 481, + 622 + ], + "type": "text", + "content": "69. Wang, X., Chen, X., Ni, B., Wang, H., Tong, Z., Liu, Y.: Deep arbitrary-scale image super-resolution via scale-equivariance pursuit. In: CVPR (2023)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 623, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 623, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 623, + 481, + 643 + ], + "type": "text", + "content": "70. Wang, X., Li, Y., Zhang, H., Shan, Y.: Towards real-world blind face restoration with generative facial prior. In: CVPR (2021)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "type": "text", + "content": "71. Wang, X., Xie, L., Dong, C., Shan, Y.: Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In: ICCV. pp. 1905-1914 (2021)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 201, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 201, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 201, + 102 + ], + "type": "text", + "content": "Qi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 512 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "type": "text", + "content": "72. Wang, X., Yu, K., Wu, S., Gu, J., Liu, Y., Dong, C., Qiao, Y., Loy, C.C.: ESRGAN: enhanced super-resolution generative adversarial networks. In: Proceedings of ECCV Workshops (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 172 + ], + "type": "text", + "content": "73. Wang, Z., Cun, X., Bao, J., Zhou, W., Liu, J., Li, H.: Uformer: A general u-shaped transformer for image restoration. In: CVPR (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 194 + ], + "type": "text", + "content": "74. Whang, J., Delbracio, M., Talebi, H., Saharia, C., Dimakis, A.G., Milanfar, P.: Deblurring via stochastic refinement. In: CVPR. pp. 16293-16303 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "text", + "content": "75. Wu, R., Yang, T., Sun, L., Zhang, Z., Li, S., Zhang, L.: Seesr: Towards semantics-aware real-world image super-resolution. In: CVPR (2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "text", + "content": "76. Yang, S., Wu, T., Shi, S., Lao, S., Gong, Y., Cao, M., Wang, J., Yang, Y.: Maniaq: Multi-dimension attention network for no-reference image quality assessment. In: CVPR. pp. 1191-1200 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "type": "text", + "content": "77. Yang, T., Ren, P., Xie, X., Zhang, L.: Gan prior embedded network for blind face restoration in the wild. In: CVPR (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 270, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 481, + 303 + ], + "type": "text", + "content": "78. Yu, F., Gu, J., Li, Z., Hu, J., Kong, X., Wang, X., He, J., Qiao, Y., Dong, C.: Scaling up to excellence: Practicing model scaling for photo-realistic image restoration in the wild. In: CVPR (2024)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 303, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 303, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 303, + 481, + 325 + ], + "type": "text", + "content": "79. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Multi-stage progressive image restoration. In: CVPR (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 325, + 481, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 481, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 481, + 347 + ], + "type": "text", + "content": "80. Zhang, K., Liang, J., Van Gool, L., Timofte, R.: Designing a practical degradation model for deep blind image super-resolution. In: ICCV. pp. 4791-4800 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "text", + "content": "81. Zhang, K., Mo, L., Chen, W., Sun, H., Su, Y.: Magicbrush: A manually annotated dataset for instruction-guided image editing. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "text", + "content": "82. Zhang, K., Zuo, W., Chen, Y., Meng, D., Zhang, L.: Beyond a Gaussian denoiser: Residual learning of deep CNN for image denoising. IEEE Transactions on Image Processing 26(7), 3142-3155 (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 413, + 481, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 481, + 436 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 481, + 436 + ], + "type": "text", + "content": "83. Zhang, K., Zuo, W., Zhang, L.: Ffdnet: Toward a fast and flexible solution for CNN based image denoising. IEEE Transactions on Image Processing (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 436, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 436, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 436, + 481, + 456 + ], + "type": "text", + "content": "84. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: ICCV (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 456, + 481, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 456, + 481, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 456, + 481, + 489 + ], + "type": "text", + "content": "85. Zhang, R., Fang, R., Gao, P., Zhang, W., Li, K., Dai, J., Qiao, Y., Li, H.: Tip-adapter: Training-free clip-adapter for better vision-language modeling. In: ECCV (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 490, + 481, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 481, + 512 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 481, + 512 + ], + "type": "text", + "content": "86. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 234, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 234, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 234, + 91, + 447, + 102 + ], + "type": "text", + "content": "SPIRE: Semantic Prompt-Driven Image Restoration" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/SPVLoc_ Semantic Panoramic Viewport Matching for 6D Camera Localization in Unseen Environments/a0f540fa-1e67-4e90-a2ae-34590f8ccef2_content_list.json b/2024/SPVLoc_ Semantic Panoramic Viewport Matching for 6D Camera Localization in Unseen Environments/a0f540fa-1e67-4e90-a2ae-34590f8ccef2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..add1c323ae10c5bdffadb520950d5c076cab1949 --- /dev/null +++ b/2024/SPVLoc_ Semantic Panoramic Viewport Matching for 6D Camera Localization in Unseen Environments/a0f540fa-1e67-4e90-a2ae-34590f8ccef2_content_list.json @@ -0,0 +1,1584 @@ +[ + { + "type": "text", + "text": "SPVLoc: Semantic Panoramic Viewport Matching for 6D Camera Localization in Unseen Environments", + "text_level": 1, + "bbox": [ + 217, + 140, + 785, + 205 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Niklas Gard $^{1,2}$ , Anna Hilsmann $^{1}$ , and Peter Eisert $^{1,2}$", + "bbox": [ + 292, + 233, + 707, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Fraunhofer Heinrich Hertz Institute, HHI, Berlin, Germany", + "bbox": [ + 294, + 261, + 704, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{first.last}@hhi.fraunhofer.de", + "bbox": [ + 383, + 277, + 614, + 289 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Humboldt University of Berlin, Germany", + "bbox": [ + 357, + 290, + 643, + 304 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. In this paper, we present SPVLoc, a global indoor localization method that accurately determines the six-dimensional (6D) camera pose of a query image and requires minimal scene-specific prior knowledge and no scene-specific training. Our approach employs a novel matching procedure to localize the perspective camera's viewpoint, given as an RGB image, within a set of panoramic semantic layout representations of the indoor environment. The panoramas are rendered from an textured 3D reference model, which only comprises approximate structural information about room shapes, along with door and window annotations. We demonstrate that a straightforward convolutional network structure can successfully achieve image-to-panorama and ultimately image-to-model matching. Through a viewpoint classification score, we rank reference panoramas and select the best match for the query image. Then, a 6D relative pose is estimated between the chosen panorama and query image. Our experiments demonstrate that this approach not only efficiently bridges the domain gap but also generalizes well to previously unseen scenes that are not part of the training data. Moreover, it achieves superior localization accuracy compared to the state of the art methods and also estimates more degrees of freedom of the camera pose. Our source code is publicly available at: https://fraunhoferhhi.github.io/spvloc.", + "bbox": [ + 261, + 333, + 743, + 612 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Localization $\\cdot$ 6D pose estimation $\\cdot$ Multimodal matching", + "bbox": [ + 261, + 623, + 732, + 638 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 660, + 375, + 676 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Indoor camera localization is a fundamental challenge of computer vision, aiming to precisely determine a camera's position and orientation within an indoor environment. Applications include navigation [38,42], augmented reality (AR) [2], building management, and building digitization [26]. Current methods often depend on extensive scene-specific data, acquired through manual creation of databases with localized images, depth information, or point clouds [22,38,43]. However, this limits their applicability, especially in situations where exploring the building is impractical or data acquisition is costly and time-consuming.", + "bbox": [ + 212, + 688, + 787, + 809 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent methods incorporate synthetic data during training, but these localization models require retraining or fine-tuning for new scenes using detailed BIM", + "bbox": [ + 212, + 810, + 787, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/102ed9dd0baa77f7c8c51552017da18c2ccdcc84dea05456ec39e50272a1495d.jpg", + "image_caption": [ + "Fig. 1: SPVLoc for 6D indoor localization. Our method calculates the indoor 6D camera pose by determining the image position and orientation relative to synthetic panoramas. The best panoramic match is found through semantic viewpoint matching." + ], + "image_footnote": [], + "bbox": [ + 305, + 143, + 699, + 361 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(Building Information Modeling) models [1,3]. The adoption of the Structured3D format [46] streamlines and standardizes 3D structure annotation by simplifying complex models into their most fundamental geometric representations. Methods that match a query image to this simplified structural representation exhibit robustness in unexplored, dynamic environments [17,30]. However, existing methods [9,30] mainly focus on 2D perspective camera localization with respect to floor plans, missing two degrees of freedom for rotation and one degree of freedom for position. Also, some methods solely rely on panoramic queries [16,17,23]. These queries are typically captured under controlled conditions, involving a tripod-mounted camera with a precise horizontal alignment, and a fixed camera height [17]. However, in AR applications or scenarios where archived photos must be accurately registered, e.g. to annotate visible building details into a digital building model, idealized capture conditions cannot always be guaranteed. In such cases, the demand for precise 6D localization becomes imperative.", + "bbox": [ + 212, + 446, + 787, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we present SPVLoc, a novel method for indoor 6D camera localization in unseen environments. It represents a 3D scene with a set of synthetically generated omnidirectional panorama images depicting fundamental semantic labels such as door, window, wall, floor, and ceiling. The core idea involves correlating the deep embedding of a perspective RGB camera image with the embeddings of panoramas to estimate the camera viewpoint - the precise region visible to the camera within the panorama. Leveraging a RetinaNet-like architecture [25], the 2D-bounding box around the viewpoint is predicted and classified. Retrieval of the closest panorama is performed by selecting the panorama with the globally highest classification score. The exact 6D pose is then found via relative pose estimation with an MLP starting from the panorama's position. Our method offers several advantages. It relies solely on a minimalist", + "bbox": [ + 212, + 659, + 789, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "N. Gard et al.", + "bbox": [ + 271, + 114, + 367, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D semantic building model [46], which could even be automatically extracted from simple 2D plans [8,28,34]. By bridging the domain gap between semantic representation and real images, it facilitates matching in unseen environments. Our approach (Fig. 1) learns wide baseline relative pose estimation and accurately predicts poses with few reference renderings.", + "bbox": [ + 212, + 146, + 787, + 222 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In our experiments, we evaluate our method using publicly available datasets, including Structured3D [46] and Zillow Indoor [10], encompassing both photorealistic synthetic data and real-world data. By sampling perspective query images from equirectangular panoramas, we efficiently augment the dataset, generating an almost infinite amount of training data. Compared to the state of the art, our method excels in estimating all degrees of freedom of the camera pose, enabling precise localization even in uncontrolled recording conditions. To summarise, we make the following contributions.", + "bbox": [ + 212, + 222, + 787, + 344 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We introduce a model-based 6D camera pose estimation system for unseen indoor environments without the need for scene-specific training.", + "2. We present a novel perspective-to-panoramic image matching concept with high retrieval accuracy even under wide camera baselines.", + "3. Demonstrating superior localization accuracy compared to the state of the art, our method estimates more degrees of freedom concurrently." + ], + "bbox": [ + 220, + 357, + 782, + 446 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 473, + 387, + 489 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Indoor Camera Localization. Indoor camera localization is a longstanding challenge in the computer vision community and in recent years, with the advance of neural networks, several deep learning based approaches have been proposed. Typically, these methods rely on prior scene acquisition to built a detailed scene representation and estimate a relation to a query image. Some methods use a 3D model generated via Structure-from-Motion [5] from registered camera images of the indoor environment and calculate the pose from 2D-3D correspondences, which might be either sparse [27,43] or dense [38]. Alternatively, encoding the entire scene within a neural network and performing regression to determine the absolute pose is a frequently used strategy [31].", + "bbox": [ + 212, + 507, + 787, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "By leveraging synthetic renderings from a detailed 3D BIM model, Acharya et al. eliminated the need for real-world reference data. Their approaches achieve real-time indoor localization [1] and an accuracy of approximately 1 meter [4] but are limited to the building of their origin, lacking generalizability to other structures. Liu et al. [26] pioneered floor plan-based image localization, predicting the facing wall within a cuboidal layout without estimating a full 3D pose. Howard et al. achieved 2D panorama localization in unseen scenes by aligning inferred room layouts with panoramic reference renderings [17] or encoded floor plans [16].", + "bbox": [ + 212, + 657, + 787, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Several methods estimate the 2D position and 1D rotation of a perspective camera relative to a floor plan. PF-net [19] uses a recurrent neural network optimized for a differentiable particle filter, designed for sequential updating but suboptimal for single images [30]. F $^3$ Loc [9] improves sequential 2D localization", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "SPVLoc: Semantic Panoramic Viewport Matching for 6D Localization", + "bbox": [ + 263, + 114, + 730, + 130 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "against a floorplan by integrating multiview geometry cues. LASER [30] localizes the camera with respect to a semantic floor plan, and considers both perspective and panoramic queries. The rendering code book scheme efficiently encodes floor plan information but is limited to 2D data, which constrains the network's ability to estimate the remaining degrees of freedom.", + "bbox": [ + 212, + 146, + 787, + 222 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To sum up, existing methods either depend on accurate scene specific data for a detailed 3D representation or do not estimate the 6D camera pose. In contrast, our method neither depends on real-world reference images, a detailed 3D model, or scene specific training but is still able to estimate the 6D perspective camera pose by introducing a new cross-domain panoramic-to-perspective image registration approach between RGB images and semantic panoramas.", + "bbox": [ + 212, + 222, + 787, + 313 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Relative Pose Regression. Relative pose regression (RPR) aims at estimating the relative motion between a query image and reference images with known poses. These frameworks, relying on image comparisons, often enhance generalizability. Typically, they directly utilize localized examples at inference time [12].", + "bbox": [ + 212, + 327, + 787, + 387 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Retrieval-based methods like NetVLAD [7] and DenseVLAD [40] start by identifying the most similar images in a database, providing a starting point for RPR. PixLoc [35] predicts multiscale deep features for direct alignment of reference and query image with a CNN and uses geometric optimization for calculating the pose. Turkoglu et al. [41] employ a GNN to re-localize cameras with respect to multiple frames. A graph links the query image to training counterparts, and a GNN refines representations for a consistent camera pose estimate. Our method shares similarities with [41] but goes further by addressing the domain gap between simplified synthetic and real data, and by incorporating multimodal matching between panoramic reference and perspective query images.", + "bbox": [ + 212, + 388, + 787, + 540 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Perspective-to-Panorama Matching. Matching between perspective query and panoramic reference images is a crucial aspect of localization, yet it has received limited attention in the literature. Most systems avoid multimodal matching and instead either match perspective crops to perspective queries [7,40] or panoramic to panoramic images [45].", + "bbox": [ + 212, + 553, + 787, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While the past discussed geometric fundamentals for integrating catadioptric and perspective camera systems into hybrid stereo setups, including concepts like epipolar constraints [37] and calibration [14], these ideas have not yet found their way into contemporary AI-based localization systems. In the domain of CNNs and panoramic image processing, distortion-aware convolutional filters were proposed by Tateno et al. [39] and further refined by Fernandez et al. [13], particularly for indoor equirectangular images. Unlike traditional convolutions, these distortion-aware filters target the differences between perspective and panoramic images, making them particularly valuable for cross-modal matching.", + "bbox": [ + 212, + 628, + 787, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recently, perspective-to-panoramic matching advanced, due to sliding window approaches in feature space [32, 36]. It improved visual recognition compared to methods that simply sampled perspective reference images from panoramas [44]. Orhan et al. [33] explored leveraging semantic segmentation similarity to address appearance variations between reference and query images. Our approach", + "bbox": [ + 212, + 765, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "N. Gard et al.", + "bbox": [ + 271, + 114, + 367, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8e197c49d9fe4be59ae1989312daf454efebb43d8af1543d8c47f784d72bc9c5.jpg", + "image_caption": [ + "Fig. 2: Network architecture overview. The information from the query branch is fed into the panorama branch by depth-wise correlation. Three task heads predict the corresponding viewpoint, one task head predicts the relative pose offset." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "enhances perspective-to-panorama matching using virtual scene renderings as panoramic references. Inspired by multimodal zero-shot retrieval [18] and object detection, particularly Mercier et al.'s work [29], it utilizes the panorama as a global context for query matching.", + "bbox": [ + 212, + 402, + 787, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 487, + 328, + 503 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SPVLoc conducts 6D indoor localization of a 2D RGB query image within a simple semantic textureless 3D scene model. The model adheres to the Structured3D annotation format [46], containing walls, doors, windows, ceiling, and floor as planes. It can be derived from straightforward from building models [46], or from floor plans using contemporary AI tools [8, 28, 34]. SPVLoc first estimates the image's viewpoint within rendered panoramas from the semantic scene model using cross-domain image-to-panorama matching (Sec. 3.1). It then ranks these predictions based on classification accuracy (Sec. 3.4) and performs relative 6D pose regression with respect to the best-matching reference panorama (Sec. 3.2).", + "bbox": [ + 212, + 518, + 787, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Semantic Panoramic Viewport Matching", + "text_level": 1, + "bbox": [ + 215, + 678, + 598, + 694 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Problem Definition. Our objective is to precisely estimate the 6D camera pose denoted as $P$ for a given query RGB image $I_{q}$ relative to a semantic 3D model $S$ . To achieve this, we reformulate the 6D indoor localization of $I_{q}$ within $S$ as a cross-domain image-to-panorama matching problem. The model is represented by a set of $N$ equirectangular reference layouts $L = \\{L_{i} \\mid i = 1, \\dots, N\\}$ consisting of semantic labels in a single channel. Each rendering comprehensively captures the entire environment around a specific vantage point. The viewpoint $V$ is a distinct cutout section from a panorama, defining the exact region a camera positioned nearby can see. The perspective distortion of $V_{P_{i}}$ contains information about", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "SPVLoc: Semantic Panoramic Viewport Matching for 6D Localization", + "bbox": [ + 264, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the relative pose offset $P_{i}$ between $L_{i}$ and $I_{q}$ . Scoring and ranking viewpoints helps in choosing the optimal panorama $L^{\\star}$ . The absolute pose $P$ is then formed by concatenating $P^{\\star}$ with $P_{i}$ . Identifying the viewpoint and the corresponding pose offset poses several challenges. First, with increasing distance between the two cameras the perspective distortion between image and panorama increases. Second, the matching involves a severe domain gap as a photographic RGB image is matched against a rendered semantic panorama.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Viewport Prediction. Semantic panorama reference renderings are created utilizing a ray-tracing-based renderer [24]. The viewport mask $V_{P}$ is a binary image that signifies the visibility of points within the equirectangular image from the view of the perspective camera. The calculation involves projecting all 3D points from the panorama in the perspective view. The projection is performed based on the known underlying layout during the training phase. The mask $V_{P}$ is determined by identifying pixels projecting into the perspective image, taking into account the 3D camera pose, and excluding those covered by occlusion.", + "bbox": [ + 212, + 266, + 787, + 387 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Additionally, the 2D bounding box $v_{bb}$ is the minimal axis-aligned frame around $V_P$ . This bounding box is circularly defined within the $360^\\circ$ panorama, allowing the overlap on the right image side to re-enter on the left side of the panorama. Each predicted bounding box is associated with a confidence score $c_{bb}$ , which quantifies the probability of the bounding box being a valid match. The network's objective is to predict $V_P$ , $v_{bb}$ and $c_{bb}$ for each pair of input semantic panorama and query RGB image. The optimal viewpoint corresponds to the predicted bounding box with the highest overall confidence score, $c_{bb}$ , when presented with a set of panorama images.", + "bbox": [ + 212, + 388, + 787, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Network Architecture. Our matching is inspired by the object detection strategy of DTOID [29]. Figure 2 illustrates the network architecture. Initially, $I_{q}$ is encoded with an EfficientNet-S backbone. The features obtained after the fourth and the last downsampling layer are preserved. Subsequently, both outputs are rescaled to a tensor of spatial size of $7 \\times 7$ and concatenated to a tensor $F$ .", + "bbox": [ + 212, + 537, + 787, + 612 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$L$ is encoded using a DenseNet-121 backbone, extending up to the fifth downsampling block to extract features $R$ . The information extracted from the input image guides the panorama branch to estimate the viewport mask and bounding box. This guidance is achieved through the application of a depth-wise correlation technique (denoted as $\\star_{\\mathbf{d}}$ ) [6,29]. $F$ is processed through two consecutive convolution blocks, each incorporating convolution, batch normalization, and ELU activation, resulting in a set of $3 \\times 3$ filters $\\hat{F}$ used during the depth-wise correlation $R \\star_{\\mathbf{d}} \\hat{F}$ .", + "bbox": [ + 212, + 613, + 787, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Furthermore, $F$ undergoes compression via global max pooling and is both multiplied and subtracted from $R$ , resulting in three feature tensors with injected image information. The features are individually processed with an additional convolution block and then concatenated. The concatenated features further undergo a final convolution block, leading to the correlated features $R^{\\star}$ .", + "bbox": [ + 212, + 734, + 787, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A decoder estimates the viewpoint of the perspective camera projected in the layout of the omnidirectional panorama. The viewpoint in the panorama is", + "bbox": [ + 214, + 809, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "N. Gard et al.", + "bbox": [ + 271, + 114, + 367, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "estimated as a bounding box with a BBox Class. and BBox Reg. head, similar to RetinaNet [25]. The BBox Class. head's output is optimized using focal loss $\\mathcal{L}_1^{VP}$ , while the BBox Reg. head is optimized with a smooth $l_{1}$ loss $\\mathcal{L}_2^{VP}$ . The BBox Class. provides a score for a given estimation, ensuring that the bounding box with the highest probability corresponds to the best estimate. Additionally, $V_{P}$ representing the exact viewpoint of the perspective camera is estimated with the aid of another convolutional decoder head and optimized with binary cross entropy loss $\\mathcal{L}_3^{VP}$ .", + "bbox": [ + 214, + 146, + 787, + 268 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Perspective Supervision. The encoded image features $I_{E}$ are connected to a convolutional decoder, outputting the underlying semantic and normal map to ensure that features closely corresponding to the image content are extracted. The semantics are learned with cross entropy loss $\\mathcal{L}_4^{VP}$ and the normals are learned with cosine loss $\\mathcal{L}_5^{VP}$ .", + "bbox": [ + 214, + 277, + 789, + 354 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Feature-Correlation-based Pose Regression", + "text_level": 1, + "bbox": [ + 214, + 372, + 614, + 387 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The correlated features, denoted as $R^{\\star}$ , encode information about the viewpoint of $I_{q}$ within a $L_{i}$ , but should also lead to relative pose offset $P_{i}$ between them. Given that $L$ is synthetically generated, the precise position of $L_{i}$ is known. Therefore, determining the relative position of $I_{q}$ in relation to $L_{i}$ is equivalent to establishing its absolute position. All rendered panoramas exhibit identity orientation, due to horizontal alignment and a northward orientation. Consequently, the rotation delta directly corresponds to the absolute rotation of the camera.", + "bbox": [ + 212, + 393, + 787, + 498 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "An additional Pose head processes $R^{\\star}$ to estimate the relative pose offset of the camera concerning a panorama rendering (Fig. 3). We first utilize three convolution blocks, whereby the first two are followed by a max-pooling layer, to compress the features. The flattened features form an 1D embedding $r^{\\star} \\in \\mathbb{R}^{C}$ . Here, $C = 640$ for a panorama of dimensions $256 \\times 128$ . Optionally, to handle different camera field of views during training and testing, $r^{\\star}$ is normalized and multiplied by the horizontal field of view angle in radians. The result is fed into a four-layer MLP, producing the 3D offset to the query image location $\\hat{t} i \\in \\mathbb{R}^{3}$ and a rotation given as the logarithm of the unit quaternion $\\hat{r} i = \\log (\\hat{q})_{ij} \\in \\mathbb{R}^{3}$ [41].", + "bbox": [ + 212, + 500, + 787, + 635 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The MLP is evaluated using two loss functions. $\\mathcal{L}_1^P$ assesses the precision of the relative translation, and $\\mathcal{L}_2^P$ evaluates the accuracy of the final rotation estimation. Both losses are computed using the $l_{1}$ distance.", + "bbox": [ + 212, + 635, + 787, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Optimization", + "text_level": 1, + "bbox": [ + 214, + 699, + 372, + 714 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The training loss is the sum of the two losses of the MLP and the five losses of the viewpoint estimation. With seven losses, our multi-task learning has to be balanced by using a different weight for each loss. As in [20, 21], we use an exponential mapping and weight multiple loss functions by considering the homoscedastic uncertainty of each task and learned weighting factors. The final loss is given as:", + "bbox": [ + 212, + 720, + 787, + 796 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\sum_ {i = 1} ^ {2} \\mathcal {L} _ {i} ^ {P} e ^ {- \\beta_ {i}} + \\beta_ {i} + \\sum_ {i = 1} ^ {5} \\mathcal {L} _ {i} ^ {V P} e ^ {- \\gamma_ {i}} + \\gamma_ {i} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 356, + 803, + 787, + 844 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "SPVLoc: Semantic Panoramic Viewport Matching for 6D Localization", + "bbox": [ + 263, + 114, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a3821fff45ec52102ba1cadb4f1ced6a630caebc8364cc0d4b3fa3174c85a14b.jpg", + "image_caption": [ + "Fig. 3: The Pose head utilizes convolutions, an MLP, and a FoV-based side input." + ], + "image_footnote": [], + "bbox": [ + 315, + 143, + 687, + 268 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Inference", + "text_level": 1, + "bbox": [ + 215, + 321, + 339, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "During inference, panorama positions are determined from a fixed 2D grid superimposed on the floor plan, with virtual cameras set at a constant height denoted as $h_{pano}$ , and the grid's dimensions specified as $xy_{pano}$ . Virtual cameras are exclusively sampled above floors to ensure valid reference images inside rooms. Setting $xy_{pano}$ as large as possible is advantageous in order to minimize the number of references, given that feature correlation and bounding box regression are executed for every panorama.", + "bbox": [ + 212, + 343, + 787, + 448 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For pose estimation, we select reference positions with the top- $n$ highest bounding box classification scores. The absolute pose is determined by the result of the Pose head, where the selected reference position is added to the translation. To eliminate the necessity for rendering during inference, we propose pre-generating reference panoramas for a floor plan. Additionally, the encoding of these panoramas can be precomputed. This ensures that only the encoded features $R$ need to be stored, significantly streamlining the inference process.", + "bbox": [ + 212, + 449, + 787, + 555 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Refinement. Without altering the network architecture or training process, SPVLoc can enhance an estimated pose by rendering a new reference panorama at the estimated position and re-executing the relative pose regression. The smaller step to the actual target position in this second iteration simplifies the problem, allowing for increased precision in the refined pose estimation.", + "bbox": [ + 212, + 566, + 787, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 662, + 375, + 680 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To validate our method we conduct several experiments and make use of two publicly available and frequently used datasets.", + "bbox": [ + 212, + 691, + 785, + 722 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Structured3D (S3D) [46] consists of 3500 near-photorealistic professionally modeled indoor environments, each annotated with ground truth 3D structure information. It encompasses 21835 panoramic images, each provided in both furnished and unfurnished configurations. The authors define a unified representation that describes scenes using geometric primitives such as planes, lines, junction points, and semantic annotations including floor, ceiling, wall, door, and window classes.", + "bbox": [ + 212, + 733, + 787, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "N. Gard et al.", + "bbox": [ + 271, + 114, + 367, + 126 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0f30960556da9c1e4a2ddd24240ce607f2f35d2cf3a963cead6b61282d4b06f9.jpg", + "image_caption": [ + "Fig. 4: ZinD data preparation. Annotations generate 3D reference models (left), while resampled panoramas create perspective train and test images (right)." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 787, + 236 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Zillow Indoor (ZInD) [10] includes 67,448 panorama images from 1,575 unfurnished residential homes, globally aligned and registered to floor plans. The 3D annotations of panoramas within a scene are adequate for transforming them into the S3D representation format. Before training, we transform all data to that format. Figure 4 shows an example reference model generated from ZInD annotations. Unlike S3D, ZInD includes room openings such as corridors connected to rooms without doors, enhancing structural variation in semantic renderings.", + "bbox": [ + 212, + 305, + 787, + 426 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For both datasets, we employ the official train-test scene split. To enhance control over the field of view and angle variation, we sample perspective test images from the RGB panoramas. They are registered in the 3D model, so the semantics are known for each perspective image. We compile test-sets suitable for 6D pose estimation and structured evaluation. The test sets consist of square perspective images with a viewing angle of 60, 90, and $120^{\\circ}$ , respectively, along with variations in roll and pitch, and random yaw angles for ZinD and S3D. To ensure meaningful views, we apply two criteria on the sampled yaw. First, the images contain a minimum of three semantic classes, like wall-floor-ceiling or wall-window-floor. Second, the camera faces inside the room rather than towards the next wall. The criteria are applied to the $60^{\\circ}$ images (4420 for ZiND and 1133 for S3D). Subsequent variations in opening angles and roll-pitch configurations are generated from this starting point.", + "bbox": [ + 212, + 430, + 789, + 627 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Details", + "text_level": 1, + "bbox": [ + 215, + 648, + 320, + 662 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For evaluation, we report the fraction of frames localized correctly within a threshold $\\tau$ . We also list the median rotation and translation error, as well as the fraction of frames localized within $\\tau$ and an additional angle threshold $v$ . In the retrieval process, the top-3 best matches are selected, and we report recall metrics for both top-1 and the best among the top-3.", + "bbox": [ + 212, + 674, + 787, + 750 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Training. For the state of the art comparison, we train a model with a variable field of view, while other experiments use a model trained with a fixed opening angle of $90^{\\circ}$ . The model with variable viewing angles randomly samples views between 45 and $135^{\\circ}$ during training. We consistently render $s$ randomly positioned panoramas around the query within a radius of $\\pm r_1$ in $xy$ direction and", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "SPVLoc: Semantic Panoramic Viewport Matching for 6D Localization", + "bbox": [ + 263, + 114, + 732, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/5ece2e871f34bff2e6e6f2d4b5da189235d2e87488b0bb0c0ca29fdadbc0c0cc.jpg", + "table_caption": [ + "Table 1: Comparison with state-of-the-art baseline on ZInD and S3D: Median translation $(\\mathrm{terr}_{\\mathrm{xy}})$ in cm, rotation error (rerrayw) in deg. for instances localized under 1m. Reporting recall at various translation accuracy levels, recall for inliers (<1m and <30°), and top-3 recall at 1m. Ours† refers to our model tested with a slightly different testset, incorporating pitch and roll rotation variation ( $\\pm 10^{\\circ}$ )." + ], + "table_footnote": [], + "table_body": "
ZInDStructured3D (Furnishing-Level : Full)
QueryMethod<1m med terr (cm)<1m med rerr (deg)10cm (%)50cm (%)1m (%)1m & 30° (%)top-3 1m (%)<1m med terr (cm)<1m med rerr (deg)10cm (%)50cm (%)1m (%)1m & 30° (%)top-3 1m(%)
Perspective 60° FoVLASER [30]33.531.034.0547.2464.5263.5378.6417.600.8416.4663.2268.3967.3486.08
Ours17.161.5021.0481.3386.7686.0994.1212.251.2134.9487.1389.4988.7996.15
\\( Ours^† \\)17.771.6319.5981.1387.1086.4394.8613.311.2430.8284.6888.1887.4895.80
Perspective 90° FoVLASER [30]26.280.778.6967.0180.9080.2589.7314.600.6723.4772.7776.8076.1891.94
Ours13.631.3430.5785.5990.8490.3496.369.170.8350.0989.6792.1291.8697.37
\\( Ours^† \\)14.251.3729.5985.9091.3390.9596.9710.090.9145.6289.3292.1291.9496.85
Perspective 120° FoVLASER [30]23.100.7812.6975.0985.6385.2092.3812.900.7629.0777.7682.1481.3594.13
Ours14.781.4529.0585.1491.7291.2796.9211.690.8739.4988.0092.1291.9497.37
\\( Ours^† \\)15.131.5328.1286.1392.7192.3597.2411.770.8838.1889.6793.0892.9197.55
", + "bbox": [ + 215, + 226, + 782, + 343 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "$\\pm r_2$ in $z$ (upward) direction. For a single query, the viewpoint and relative pose offset is therefore calculated against multiple references. In our experimental setup, we set $s$ to 4, $r_1$ to 1.4 meters, and $r_2$ to 30 centimeters. Additionally, one random negative example is generated outside $r_1$ , likely in a different room, with the target to enhance the network's ability to distinguish subtle room differences. Images are sampled with random yaw and $\\pm 10^{\\circ}$ random pitch and roll angle. The batch size is set to 40, comprising 40 query images and 200 panoramas, and the training is performed on a single NVIDIA A100 GPU. Query images with fewer than three semantic classes are disregarded during the loss calculation process. Training runs for approximately 42,000 steps with an initial learning rate of $2.5 \\times 10^{-4}$ , halved two times during training.", + "bbox": [ + 212, + 369, + 789, + 537 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Testing. During testing, panoramas are sampled on a $1.2 \\times 1.2$ m grid, requiring significantly fewer samples than other rendering-based approaches such as LaLaLoc [17], which uses a $0.5 \\times 0.5$ m grid. Additionally, for a scene, all panoramic views can be encoded initially, and during inference, only the query image needs to be encoded. If not otherwise noted the studies are conducted on ZInD with fixed 90 degree FoV, reporting the 3D rotation and translation error.", + "bbox": [ + 212, + 551, + 787, + 643 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Comparison with the State of the Art", + "text_level": 1, + "bbox": [ + 214, + 664, + 576, + 679 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To our knowledge, SPVLoc is the first method to perform 6D localization in unseen scenes with respect to semantic 3D models. The state of the-art method for scene-independent global localization, LASER [30], takes perspective views as input and matches them against a semantic floor plan. Hence, it estimates only two position and one rotation degree of freedom. Nonetheless, it serves as a strong baseline to assess the localization accuracy of our method.", + "bbox": [ + 212, + 688, + 787, + 777 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 1 shows the results on ZInD and S3D. Given that we compare against a 2D method, our evaluation also includes recall based on 2D distance and angular error derived from the yaw angle. The yaw angle is extracted from the estimated rotation matrix. Notably, SPVLoc without refinement outperforms LASER in", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "N. Gard et al.", + "bbox": [ + 271, + 114, + 367, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/686541602b623601a940925072dcd7d9407c7f516994ccf6f02e7aa8e0dda318.jpg", + "table_caption": [ + "Table 2: Ablation study. SPVLoc variations compared on ZInD with $90^{\\circ}$ FoV. In contrast to Tab. 1, terr denotes 3D position errors, and rerr denotes 3D rotation errors." + ], + "table_footnote": [], + "table_body": "
Abl. TypeMethod<1m med terr (cm)<1m med rerr (deg)10cm (%)50cm (%)1m (%)1m & 30° (%)top-3 1m (%)
(Ours)base14.312.0526.8886.9291.9091.6196.56
+ refine12.321.9633.6990.8492.7691.74-
Loss- query dec.15.702.2321.9785.5090.7990.4396.45
- view dec.15.192.2123.6085.3690.4589.8296.49
- view/query dec.16.222.1921.5684.5089.5989.1296.36
Training- neg. samples15.062.2822.6076.6582.1981.7492.99
Arch.Δ Resnet18 (query)15.562.5620.8886.5690.7290.3296.56
Δ Equiconv (pano)14.982.1325.0787.4491.7091.3196.97
Input+ normal14.682.2825.0985.9390.8490.2396.47
+ depth14.062.1027.6088.0592.1591.7996.74
", + "bbox": [ + 297, + 184, + 692, + 313 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "terms of localization accuracy, exhibiting higher recall across all categories, an increased inlier rate, and significantly reduced median translation error. In terms of 2D rotation, our method is slightly less precise. This is because LASER estimates yaw directly while we estimate a full rotation matrix. Estimating camera height and roll and pitch angle is an advantage of our method. In each last row $(Ours^{\\dagger})$ , we evaluate our method on an alternate test set with roll and pitch variations $(\\pm 10^{\\circ})$ . The results suggest that these variations do not substantially impact the accuracy of our method for 2D localization.", + "bbox": [ + 212, + 340, + 787, + 460 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Regarding further methods, LASER has showcased a significant performance improvement over PF-net [19] and non-AI-based Monte-Carlo Localization (MCL) [11]. PF-net, originally designed for sequential time updating, does not perform well in single-query localization scenarios. MCL neglects semantic annotations, uses LIDAR instead of images, which also has been shown to drastically reduce performance [30]. Our substantial improvement over LASER therefore also illustrates a performance gain over aforementioned methods. Another experiment using the pre-rendered perspective images of S3D, including a comparison to $\\mathrm{F}^3\\mathrm{Loc}$ [9], is part of the supplementary material.", + "bbox": [ + 212, + 460, + 789, + 599 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Performance Studies", + "text_level": 1, + "bbox": [ + 214, + 619, + 433, + 633 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Ablation Study. We commence our experimentation with the full network (base) configuration. Subsequently, we modify the network by removing specific components (Tab. 2). Initially, we test the influence of perspective supervision (-query dec.) and the view segment task head (-view dec.) by excluding them during training. Their exclusion reduces network performance, likely because the network's ability to understand the image content and establish correspondences is diminished. Removing both (-view/query dec.) further reduces accuracy.", + "bbox": [ + 212, + 643, + 787, + 750 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We then exclude the negative samples from different rooms during training (-neg. samples). This significantly reduces the amount of correctly localized images. It supports the hypothesis that teaching the network to recognize spaces where there is no match helps to improve the accuracy of predictions.", + "bbox": [ + 212, + 750, + 787, + 809 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "To modify the architecture, the image encoder EfficientNet-S is replaced with a smaller ResNet-18 [15] (Δ Resnet18), resulting in a performance drop. This", + "bbox": [ + 214, + 809, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "SPVLoc: Semantic Panoramic Viewport Matching for 6D Localization", + "bbox": [ + 263, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/1e784445864227badce7d8b33244ad5e0d2e554920c1a55c0a59cedd36937147.jpg", + "image_caption": [ + "Fig. 5: Qualitative localization results. Top to bottom - query, rendering with top-1 estimated pose, panorama with estimated viewport, map. Green box: success for top-1 match. Yellow box: success for top-2 match. Red box: failure case." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 329 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "suggests an increased demand for network capacity to analyze the images. In the panorama encoder, replacing all convolutional layers with Equiconv [13] (ΔEquiconv), a specialized convolution for panorama images that remaps query locations from Euclidean to spherical space, does not result in a performance gain. This suggests that the capacity of the small DenseNet-121 is sufficient for analyzing the low-resolution semantic inputs.", + "bbox": [ + 212, + 412, + 787, + 503 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Lastly, we incorporate additional input modalities for the panoramas into the panorama encoder. While the addition of world coordinate normals (+normals) did not improve the result, including an extra depth map slightly enhanced the outcome (+depth). This suggests that there are cases where additional information is encoded in the depth that cannot be inferred from semantics alone.", + "bbox": [ + 212, + 503, + 789, + 580 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Qualitative Results. Figure 5 shows selected SPVLoc outcomes. The top row displays camera images, and the second row shows renderings from the semantic 3D model with estimated top-1 poses. All green boxed images are correctly localized for the top-1 match. In the yellow box, the top-1 match fails, but the top-2 match is accurate. The red box example cannot be localized. Across all examples, a pose with high visual similarity is found. Ambiguous examples arise from similar layouts occurring in multiple rooms. Failure cases are attributed to either images lacking sufficient structural information for matching or outliers in the data, such as those taken on balconies or in dark cellars. The third row displays the chosen reference panorama, the selected viewpoint bounding box corresponding to the camera image, and a semi-transparent overlay representing the estimated viewpoint mask. Additional detailed visualizations are part of the supplementary material.", + "bbox": [ + 212, + 595, + 787, + 792 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Influence of Sampling Grid Sizes and Refinement. Reference positions result from overlaying a grid onto the floorplan, and the grid density significantly", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "N. Gard et al.", + "bbox": [ + 271, + 114, + 367, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/3708d75da9aed7390a3951a2fb970fb20da92ec08a745d5400f25d184e7174bd.jpg", + "table_caption": [ + "Table 3: Grid size vs. accuracy: Deviation of $10\\mathrm{cm}$ and $1\\mathrm{m}$ recall from (boxed) baseline." + ], + "table_footnote": [], + "table_body": "
Sample dist.Samples per \\( m^2 \\)Adv.10cm (%)10cm + refine1m (%)1m + refine
0.72.07+1.00+7.49+1.86+2.38
0.91.25+0.95+6.99+1.79+2.08
1.20.70\\( \\boxed{26.88} \\)+6.81\\( \\boxed{91.90} \\)+0.86
1.20.48✓+0.57+6.95-0.63+0.61
1.50.45-1.00+6.63-2.47-1.02
1.50.31✓-1.04+6.47-2.67-0.38
2.00.26-4.19+3.60-9.05-5.79
2.00.20✓-1.76+5.81-3.91-1.18
", + "bbox": [ + 220, + 184, + 503, + 308 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/91b10ff04ac2a7894072d24db67c926cf32755479e03db8746e94c5cda1c653d.jpg", + "image_caption": [ + "Fig. 6: Reference positions for global grid of size $0.7\\mathrm{m}$ , $1.5\\mathrm{m}$ and local grid of size $1.5\\mathrm{m}$ (from left to right)." + ], + "image_footnote": [], + "bbox": [ + 532, + 150, + 604, + 237 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/af39788a9556d7dbfe55ad0eecd8b086b179e09c3c8c42015b9e761cf7d1158c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 150, + 689, + 238 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/5ec39db21124c7bbea1d4ec2e108d18a8e26f0b23d85a2075210fdf383184c46.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 705, + 151, + 772, + 238 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "affects SPVLoc's efficiency and inference time. Minimizing references reduces rendering effort and requires the model to estimate larger offsets. E.g. to increase the grid from $0.7m$ to $2m$ reduces the number of samples per $m^2$ by a factor of 10. We vary the grid size to analyze its impact on performance. We also experiment with a local grid per room substituting the global grid (Adv.) reducing the risk of missing rooms entirely. The local grid is centered on the room outline and additionally places one camera in the middle of each opening. Figure 6 shows different grids for an example plan. Table 3 illustrates the relationships among grid size, global and local sampling, and refinement. Refinement notably improves $10\\mathrm{cm}$ recall across configurations. For the largest local grid with refinement, the $1\\mathrm{m}$ recall only decreases by $1.18\\%$ compared to the baseline configuration. In summary, a trade-off between refinement and dense/sparse scene sampling exists, yet our method consistently delivers strong results with very sparse sampling, making it suitable for large scenes.", + "bbox": [ + 212, + 337, + 787, + 547 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Influence of Focal Length. Table 4 depicts the performance difference between a variant trained with a known camera focal length and a focal length-independent network trained with varying focal lengths (ranging from 45 to 135 degrees). The fixed focal length network performs slightly better with matching images but loses precision with test images of different focal lengths. In contrast, the focal length-independent network demonstrates the capability to handle such differences. We argue that the significantly enhanced flexibility of the latter justifies the minor performance degradation in practical, real-world scenarios.", + "bbox": [ + 212, + 565, + 789, + 688 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Influence of Camera Rotation Angles. The range of pitch and roll angles the network is able to handle is determined by the variations of angles seen during training. We train the network for ZInD using a angle variation of $\\pm 22^{\\circ}$ and test with images where the range of variation increases from $0^{\\circ}$ to $20^{\\circ}$ in steps of $\\pm 5^{\\circ}$ . Table 5 shows only a slight increase in median errors for rotation and translation as the rotation angles in test images increase. The chosen ranges represent a significant coverage and offer substantial improvements over existing methods. In indoor scenes, larger camera tilt often results in images dominated by the ceiling or floor, rendering them unsuitable for semantic matching.", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "SPVLoc: Semantic Panoramic Viewport Matching for 6D Localization", + "bbox": [ + 263, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/07ae866ab6ecb22088efd7fd62cf434753c24edfc429a9a920ec03af075e9a6d.jpg", + "table_caption": [ + "Table 4: The fixed FoV network excels at a specific angle but loses precision with changing FoV. The proposed network remains stable." + ], + "table_footnote": [], + "table_body": "
Input FoVTrain FoV<1m med terr (cm)<1m med rerr (deg)10cm (%)1m (%)
90°90°14.312.0526.8891.90
45 - 135°16.112.2021.8191.45
80°90°37.512.371.6790.14
45 - 135°16.752.1919.3090.50
", + "bbox": [ + 228, + 198, + 496, + 292 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/592afff64bda853bbba4136d8cbf09ea012544bd02f4583734452413404e12e7.jpg", + "table_caption": [ + "Table 5: Effect of increasing angle variation in test data." + ], + "table_footnote": [], + "table_body": "
MetricRoll/Pitch Variation
±0°±5°±10°±15°±20°
<1m med. terr14.4614.8814.8815.2115.91
<1m med. verr2.352.362.382.512.60
10cm (%)24.6423.9424.4623.6922.74
1m (%)91.4591.4791.9992.3892.60
", + "bbox": [ + 532, + 194, + 779, + 282 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/811c6a3a4d41398bd98d4f4fb6ace48ba700416048fb375106718f055b490243.jpg", + "table_caption": [ + "Table 6: Timing. Performance on ZInD using a NVIDIA Tesla V100, following the configurations outlined in Sec. 4.3, and compared against results from [30]." + ], + "table_footnote": [], + "table_body": "
MethodSampling Time (s)Query FpsQuery Module Time (ms)
PF-net [19]48.95 ± 38.955.06 ± 1.77Eff.Net V2 S4.4
LASER [30]0.97 ± 1.098.31 ± 0.64Correlation35.7
Ours1.66 ± 1.0628.63 ± 11.54Pose1.1
", + "bbox": [ + 297, + 347, + 691, + 397 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations. In large spaces with repeating room layouts, the approach's effectiveness may be constrained by the limited level of detail in the semantic reference model. However, enhancing the S3D format by adding more semantic classes and structural information, such as staircases or sanitary installations, could reduce ambiguities.", + "bbox": [ + 212, + 426, + 787, + 503 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Timing. We split the measured time into scene sampling/preparation time, including position sampling, panorama rendering and encoding, and inference/query time, including query image encoding, feature correlation with bounding box classification, and the Pose task head, but no refinement. Table 6 shows that our method exhibits a considerably lower query time than LASER [30] and PF-net [19]. This difference places our approach close to real-time for moderately sized scenes. Currently, scene sampling time is slower than [30], mainly due to rendering on the CPU without parallelization [24]. Parallelizing this process using GPU or CPU could significantly accelerate scene preparation.", + "bbox": [ + 212, + 521, + 787, + 657 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 683, + 359, + 699 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This paper introduces a method for scene-independent model-based 6D localization for indoor scenes, involving a novel method for multimodal image matching (panorama to perspective, RGB to semantic). The matching and retrieval is efficient and scalable under sparse reference sampling. The localization accuracy and inference speed outperform state of the art methods and the inclusion of the 3D-model, reduces ambiguities to precisely estimate 6D poses. Future work involves combining localization and image analysis to enhance digital building models or exploring applications in augmented reality scenarios.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "N. Gard et al.", + "bbox": [ + 271, + 114, + 367, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements. This work is partially funded by the German Federal Ministry of Economic Affairs and Climate Action (BIMKIT, 01MK21001H) and the German Federal Ministry for Digital and Transport (EConoM, 19OI22009C).", + "bbox": [ + 215, + 146, + 787, + 193 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 212, + 323, + 228 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Acharya, D., Khoshelham, K., Winter, S.: BIM-PoseNet: Indoor camera localisation using a 3D indoor model and deep learning from synthetic images. ISPRS journal of photogrammetry and remote sensing 150, 245-258 (2019)", + "2. Acharya, D., Ramezani, M., Khoshelham, K., Winter, S.: Bim-tracker: A model-based visual tracking approach for indoor localisation using a 3d building model. ISPRS Journal of Photogrammetry and Remote Sensing 150, 157-171 (2019)", + "3. Acharya, D., Tatli, C.J., Khoshelham, K.: Synthetic-real image domain adaptation for indoor camera pose regression using a 3d model. ISPRS Journal of Photogrammetry and Remote Sensing 202, 405-421 (2023)", + "4. Acharya, D., Tennakoon, R., Muthu, S., Khoshelham, K., Hoseinnezhad, R., Bab-Hadiashar, A.: Single-image localisation using 3D models: Combining hierarchical edge maps and semantic segmentation for domain adaptation. Automation in Construction 136, 104-152 (2022)", + "5. Agarwal, S., Furukawa, Y., Snavely, N., Simon, I., Curless, B., Seitz, S.M., Szeliski, R.: Building rome in a day. Communications of the ACM 54(10), 105-112 (2011)", + "6. Ammirato, P., Fu, C.Y., Shvets, M., Kosecka, J., Berg, A.C.: Target driven instance detection. arXiv preprint arXiv:1803.04610 (2018)", + "7. Arandjelovic, R., Gronat, P., Torii, A., Pajdla, T., Sivic, J.: Netvlad: Cnn architecture for weakly supervised place recognition. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2016)", + "8. Cambeiro, A.B., Trzeciakiewicz, M., Hilsmann, A., Eisert, P.: Automatic reconstruction of semantic 3d models from 2d floor plans. In: Proc. Int. Conf. on Machine Vision Applications (MVA). Hamamatsu, Japan (July 2023)", + "9. Chen, C., Wang, R., Vogel, C., Pollefeys, M.: F3Loc: Fusion and Filtering for Floorplan Localization. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2024)", + "0. Cruz, S., Hutchcroft, W., Li, Y., Khosravan, N., Boyadzhiev, I., Kang, S.B.: Zillow indoor dataset: Annotated floor plans with 360deg panoramas and 3d room layouts. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2133-2143 (2021)", + "1. Dellaert, F., Fox, D., Burgard, W., Thrun, S.: Monte carlo localization for mobile robots. In: Proceedings IEEE International Conference on Robotics and Automation. vol. 2, pp. 1322-1328 (1999)", + "2. Ding, M., Wang, Z., Sun, J., Shi, J., Luo, P.: CamNet: Coarse-to-fine retrieval for camera re-localization. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2871-2880 (2019)", + "3. Fernandez-Labrador, C., Facil, J.M., Perez-Yus, A., Demonceaux, C., Civera, J., Guerrero, J.J.: Corners for layout: End-to-end layout recovery from 360 images. IEEE Robotics and Automation Letters 5(2), 1255-1262 (2020)", + "4. He, B., Chen, Z., Li, Y.: Calibration method for a central catadioptric-perspective camera system. JOSA A 29(11), 2514-2524 (2012)", + "5. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 770-778 (2016)" + ], + "bbox": [ + 225, + 239, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "SPVLoc: Semantic Panoramic Viewport Matching for 6D Localization", + "bbox": [ + 263, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Howard-Jenkins, H., Prisacariu, V.A.: LaLaLoc++: Global Floor Plan Comprehension for Layout Localisation in Unvisited Environments. In: Proc. European Conference on Computer Vision (ECCV). pp. 693-709. Springer (2022)", + "17. Howard-Jenkins, H., Ruiz-Sarmiento, J.R., Prisacariu, V.A.: LaLaLoc: Latent Layout Localisation in Dynamic, Unvisited Environments. In: Proc. IEEE/CVF International Conference on Computer Vision (ICCV) (2021)", + "18. Janik, M., Gard, N., Hilsmann, A., Eisert, P.: Zero in on Shape: A Generic 2D-3D Instance Similarity Metric Learned from Synthetic Data. In: Proc. IEEE International Conference on Image Processing (ICIP). pp. 2638-2642. IEEE (2021)", + "19. Karkus, P., Hsu, D., Lee, W.S.: Particle filter networks with application to visual localization. In: Conference on robot learning. pp. 169-178. PMLR (2018)", + "20. Kendall, A., Cipolla, R.: Geometric loss functions for camera pose regression with deep learning. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5974-5983 (2017)", + "21. Kendall, A., Gal, Y., Cipolla, R.: Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7482-7491 (2018)", + "22. Kendall, A., Grimes, M., Cipolla, R.: Posenet: A convolutional network for real-time 6-dof camera relocalization. In: Proc. IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2938-2946 (2015)", + "23. Kim, J., Jeong, J., Kim, Y.M.: Fully geometric panoramic localization. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 20827-20837 (2024)", + "24. Li, T.M., Aittala, M., Durand, F., Lehtinen, J.: Differentiable Monte Carlo Ray Tracing through Edge Sampling. ACM Trans. Graph. (Proc. SIGGRAPH Asia) 37(6), 222:1-222:11 (2018)", + "25. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proc. IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2980-2988 (2017)", + "26. Liu, C., Schwing, A., Kundu, K., Urtasun, R., Fidler, S.: Rent3d: Floor-plan priors for monocular layout estimation. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2015)", + "27. Liu, L., Li, H., Dai, Y.: Efficient global 2d-3d matching for camera localization in a large-scale 3d map. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2372-2381 (2017)", + "28. Lv, X., Zhao, S., Yu, X., Zhao, B.: Residential floor plan recognition and reconstruction. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16717-16726 (June 2021)", + "29. Mercier, J.P., Garon, M., Giguere, P., Lalonde, J.F.: Deep Template-Based Object Instance Detection. In: Proc. WACV. pp. 1507-1516 (January 2021)", + "30. Min, Z., Khosravan, N., Bessinger, Z., Narayana, M., Kang, S.B., Dunn, E., Boyadzhiev, I.: Laser: Latent space rendering for 2d visual localization. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11122-11131 (2022)", + "31. Naseer, T., Burgard, W.: Deep regression for monocular camera-based 6-dof global localization in outdoor environments. In: Proc. International Conference on Intelligent Robot Systems (IROS). pp. 1525-1530. IEEE (2017)", + "32. Orhan, S., Bastanlar, Y.: Efficient search in a panoramic image database for long-term visual localization. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1727-1734 (2021)" + ], + "bbox": [ + 215, + 147, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "N. Gard et al.", + "bbox": [ + 271, + 114, + 367, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "33. Orhan, S., Guerrero, J.J., Bastanlar, Y.: Semantic Pose Verification for Outdoor Visual Localization With Self-Supervised Contrastive Learning. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops. pp. 3989-3998 (June 2022)", + "34. Park, S., Kim, H.: 3dplannet: generating 3d models from 2d floor plan images using ensemble methods. Electronics 10(22), 2729 (2021)", + "35. Sarlin, P.E., Unagar, A., Larsson, M., Germain, H., Toft, C., Larsson, V., Pollefeys, M., Lepetit, V., Hammarstrand, L., Kahl, F., et al.: Back to the feature: Learning robust camera localization from pixels to pose. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3247-3257 (2021)", + "36. Shi, Z., Shi, H., Yang, K., Yin, Z., Lin, Y., Wang, K.: PanoVPR: Towards Unified Perspective-to-Equirectangular Visual Place Recognition via Sliding Windows across the Panoramic View. In: 2023 IEEE 26th International Conference on Intelligent Transportation Systems (ITSC). pp. 1333-1340 (2023)", + "37. Sturm, P.: Mixing catadioptric and perspective cameras. In: Proceedings of the IEEE Workshop on Omnidirectional Vision 2002. Held in conjunction with ECCV'02. pp. 37-44. IEEE (2002)", + "38. Taira, H., Okutomi, M., Sattler, T., Cimpoi, M., Pollefeys, M., Sivic, J., Pajdla, T., Torii, A.: InLoc: Indoor visual localization with dense matching and view synthesis. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7199-7209 (2018)", + "39. Tateno, K., Navab, N., Tombari, F.: Distortion-aware convolutional filters for dense prediction in panoramic images. In: Proc. European Conference on Computer Vision (ECCV). pp. 707-722 (2018)", + "40. Torii, A., Arandjelovic, R., Sivic, J., Okutomi, M., Pajdla, T.: 24/7 place recognition by view synthesis. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1808-1817 (2015)", + "41. Türkoglu, M.Ö., Brachmann, E., Schindler, K., Brostow, G., Monszpart, A.: Visual Camera Re-Localization Using Graph Neural Networks and Relative Pose Supervision. In: Proc. International Conference on 3D Vision (3DV) (2021)", + "42. Wang, S., Fidler, S., Urtasun, R.: Lost shopping! monocular localization in large indoor spaces. In: Proc. IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2695-2703 (2015)", + "43. Xia, J., Gong, J.: Precise indoor localization with 3d facility scan data. Computer-Aided Civil and Infrastructure Engineering 37(10), 1243-1259 (2022)", + "44. Zamir, A.R., Shah, M.: Accurate image localization based on google maps street view. In: Proc. European Conference on Computer Vision (ECCV). pp. 255-268. Springer, Heraklion, Greece (Sep 2010)", + "45. Zhang, C., Budvytis, I., Liwicki, S., Cipolla, R.: Lifted semantic graph embedding for omnidirectional place recognition. In: Proc. International Conference on 3D Vision (3DV). pp. 1401-1410. IEEE (2021)", + "46. Zheng, J., Zhang, J., Li, J., Tang, R., Gao, S., Zhou, Z.: Structured3D: A large photo-realistic dataset for structured 3d modeling. In: Proc. European Conference on Computer Vision (ECCV). pp. 519-535. Springer, Glasgow, UK (Aug 2020)" + ], + "bbox": [ + 215, + 146, + 787, + 757 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "SPVLoc: Semantic Panoramic Viewport Matching for 6D Localization", + "bbox": [ + 264, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/SPVLoc_ Semantic Panoramic Viewport Matching for 6D Camera Localization in Unseen Environments/a0f540fa-1e67-4e90-a2ae-34590f8ccef2_model.json b/2024/SPVLoc_ Semantic Panoramic Viewport Matching for 6D Camera Localization in Unseen Environments/a0f540fa-1e67-4e90-a2ae-34590f8ccef2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c403bc80516731aa66b0bad6e66a87336eca711c --- /dev/null +++ b/2024/SPVLoc_ Semantic Panoramic Viewport Matching for 6D Camera Localization in Unseen Environments/a0f540fa-1e67-4e90-a2ae-34590f8ccef2_model.json @@ -0,0 +1,2159 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.218, + 0.141, + 0.787, + 0.207 + ], + "angle": 0, + "content": "SPVLoc: Semantic Panoramic Viewport Matching for 6D Camera Localization in Unseen Environments" + }, + { + "type": "text", + "bbox": [ + 0.293, + 0.234, + 0.709, + 0.251 + ], + "angle": 0, + "content": "Niklas Gard\\(^{1,2}\\), Anna Hilsmann\\(^{1}\\), and Peter Eisert\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.295, + 0.262, + 0.705, + 0.277 + ], + "angle": 0, + "content": "1 Fraunhofer Heinrich Hertz Institute, HHI, Berlin, Germany" + }, + { + "type": "text", + "bbox": [ + 0.385, + 0.278, + 0.616, + 0.29 + ], + "angle": 0, + "content": "{first.last}@hhi.fraunhofer.de" + }, + { + "type": "text", + "bbox": [ + 0.358, + 0.291, + 0.645, + 0.305 + ], + "angle": 0, + "content": "2 Humboldt University of Berlin, Germany" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.334, + 0.744, + 0.613 + ], + "angle": 0, + "content": "Abstract. In this paper, we present SPVLoc, a global indoor localization method that accurately determines the six-dimensional (6D) camera pose of a query image and requires minimal scene-specific prior knowledge and no scene-specific training. Our approach employs a novel matching procedure to localize the perspective camera's viewpoint, given as an RGB image, within a set of panoramic semantic layout representations of the indoor environment. The panoramas are rendered from an textured 3D reference model, which only comprises approximate structural information about room shapes, along with door and window annotations. We demonstrate that a straightforward convolutional network structure can successfully achieve image-to-panorama and ultimately image-to-model matching. Through a viewpoint classification score, we rank reference panoramas and select the best match for the query image. Then, a 6D relative pose is estimated between the chosen panorama and query image. Our experiments demonstrate that this approach not only efficiently bridges the domain gap but also generalizes well to previously unseen scenes that are not part of the training data. Moreover, it achieves superior localization accuracy compared to the state of the art methods and also estimates more degrees of freedom of the camera pose. Our source code is publicly available at: https://fraunhoferhhi.github.io/spvloc." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.624, + 0.733, + 0.639 + ], + "angle": 0, + "content": "Keywords: Localization \\(\\cdot\\) 6D pose estimation \\(\\cdot\\) Multimodal matching" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.661, + 0.377, + 0.677 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.81 + ], + "angle": 0, + "content": "Indoor camera localization is a fundamental challenge of computer vision, aiming to precisely determine a camera's position and orientation within an indoor environment. Applications include navigation [38,42], augmented reality (AR) [2], building management, and building digitization [26]. Current methods often depend on extensive scene-specific data, acquired through manual creation of databases with localized images, depth information, or point clouds [22,38,43]. However, this limits their applicability, especially in situations where exploring the building is impractical or data acquisition is costly and time-consuming." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.811, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Recent methods incorporate synthetic data during training, but these localization models require retraining or fine-tuning for new scenes using detailed BIM" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.127 + ], + "angle": 0, + "content": "N. Gard et al." + }, + { + "type": "image", + "bbox": [ + 0.306, + 0.145, + 0.7, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.373, + 0.788, + 0.416 + ], + "angle": 0, + "content": "Fig. 1: SPVLoc for 6D indoor localization. Our method calculates the indoor 6D camera pose by determining the image position and orientation relative to synthetic panoramas. The best panoramic match is found through semantic viewpoint matching." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.447, + 0.789, + 0.658 + ], + "angle": 0, + "content": "(Building Information Modeling) models [1,3]. The adoption of the Structured3D format [46] streamlines and standardizes 3D structure annotation by simplifying complex models into their most fundamental geometric representations. Methods that match a query image to this simplified structural representation exhibit robustness in unexplored, dynamic environments [17,30]. However, existing methods [9,30] mainly focus on 2D perspective camera localization with respect to floor plans, missing two degrees of freedom for rotation and one degree of freedom for position. Also, some methods solely rely on panoramic queries [16,17,23]. These queries are typically captured under controlled conditions, involving a tripod-mounted camera with a precise horizontal alignment, and a fixed camera height [17]. However, in AR applications or scenarios where archived photos must be accurately registered, e.g. to annotate visible building details into a digital building model, idealized capture conditions cannot always be guaranteed. In such cases, the demand for precise 6D localization becomes imperative." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.66, + 0.79, + 0.841 + ], + "angle": 0, + "content": "In this paper, we present SPVLoc, a novel method for indoor 6D camera localization in unseen environments. It represents a 3D scene with a set of synthetically generated omnidirectional panorama images depicting fundamental semantic labels such as door, window, wall, floor, and ceiling. The core idea involves correlating the deep embedding of a perspective RGB camera image with the embeddings of panoramas to estimate the camera viewpoint - the precise region visible to the camera within the panorama. Leveraging a RetinaNet-like architecture [25], the 2D-bounding box around the viewpoint is predicted and classified. Retrieval of the closest panorama is performed by selecting the panorama with the globally highest classification score. The exact 6D pose is then found via relative pose estimation with an MLP starting from the panorama's position. Our method offers several advantages. It relies solely on a minimalist" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.264, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "SPVLoc: Semantic Panoramic Viewport Matching for 6D Localization" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.223 + ], + "angle": 0, + "content": "3D semantic building model [46], which could even be automatically extracted from simple 2D plans [8,28,34]. By bridging the domain gap between semantic representation and real images, it facilitates matching in unseen environments. Our approach (Fig. 1) learns wide baseline relative pose estimation and accurately predicts poses with few reference renderings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.788, + 0.345 + ], + "angle": 0, + "content": "In our experiments, we evaluate our method using publicly available datasets, including Structured3D [46] and Zillow Indoor [10], encompassing both photorealistic synthetic data and real-world data. By sampling perspective query images from equirectangular panoramas, we efficiently augment the dataset, generating an almost infinite amount of training data. Compared to the state of the art, our method excels in estimating all degrees of freedom of the camera pose, enabling precise localization even in uncontrolled recording conditions. To summarise, we make the following contributions." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.358, + 0.784, + 0.386 + ], + "angle": 0, + "content": "1. We introduce a model-based 6D camera pose estimation system for unseen indoor environments without the need for scene-specific training." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.388, + 0.784, + 0.417 + ], + "angle": 0, + "content": "2. We present a novel perspective-to-panoramic image matching concept with high retrieval accuracy even under wide camera baselines." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.419, + 0.784, + 0.448 + ], + "angle": 0, + "content": "3. Demonstrating superior localization accuracy compared to the state of the art, our method estimates more degrees of freedom concurrently." + }, + { + "type": "list", + "bbox": [ + 0.222, + 0.358, + 0.784, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.474, + 0.388, + 0.49 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.658 + ], + "angle": 0, + "content": "Indoor Camera Localization. Indoor camera localization is a longstanding challenge in the computer vision community and in recent years, with the advance of neural networks, several deep learning based approaches have been proposed. Typically, these methods rely on prior scene acquisition to built a detailed scene representation and estimate a relation to a query image. Some methods use a 3D model generated via Structure-from-Motion [5] from registered camera images of the indoor environment and calculate the pose from 2D-3D correspondences, which might be either sparse [27,43] or dense [38]. Alternatively, encoding the entire scene within a neural network and performing regression to determine the absolute pose is a frequently used strategy [31]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.78 + ], + "angle": 0, + "content": "By leveraging synthetic renderings from a detailed 3D BIM model, Acharya et al. eliminated the need for real-world reference data. Their approaches achieve real-time indoor localization [1] and an accuracy of approximately 1 meter [4] but are limited to the building of their origin, lacking generalizability to other structures. Liu et al. [26] pioneered floor plan-based image localization, predicting the facing wall within a cuboidal layout without estimating a full 3D pose. Howard et al. achieved 2D panorama localization in unseen scenes by aligning inferred room layouts with panoramic reference renderings [17] or encoded floor plans [16]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Several methods estimate the 2D position and 1D rotation of a perspective camera relative to a floor plan. PF-net [19] uses a recurrent neural network optimized for a differentiable particle filter, designed for sequential updating but suboptimal for single images [30]. F\\(^3\\)Loc [9] improves sequential 2D localization" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.128 + ], + "angle": 0, + "content": "N. Gard et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.223 + ], + "angle": 0, + "content": "against a floorplan by integrating multiview geometry cues. LASER [30] localizes the camera with respect to a semantic floor plan, and considers both perspective and panoramic queries. The rendering code book scheme efficiently encodes floor plan information but is limited to 2D data, which constrains the network's ability to estimate the remaining degrees of freedom." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.788, + 0.314 + ], + "angle": 0, + "content": "To sum up, existing methods either depend on accurate scene specific data for a detailed 3D representation or do not estimate the 6D camera pose. In contrast, our method neither depends on real-world reference images, a detailed 3D model, or scene specific training but is still able to estimate the 6D perspective camera pose by introducing a new cross-domain panoramic-to-perspective image registration approach between RGB images and semantic panoramas." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.328, + 0.788, + 0.388 + ], + "angle": 0, + "content": "Relative Pose Regression. Relative pose regression (RPR) aims at estimating the relative motion between a query image and reference images with known poses. These frameworks, relying on image comparisons, often enhance generalizability. Typically, they directly utilize localized examples at inference time [12]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.389, + 0.788, + 0.541 + ], + "angle": 0, + "content": "Retrieval-based methods like NetVLAD [7] and DenseVLAD [40] start by identifying the most similar images in a database, providing a starting point for RPR. PixLoc [35] predicts multiscale deep features for direct alignment of reference and query image with a CNN and uses geometric optimization for calculating the pose. Turkoglu et al. [41] employ a GNN to re-localize cameras with respect to multiple frames. A graph links the query image to training counterparts, and a GNN refines representations for a consistent camera pose estimate. Our method shares similarities with [41] but goes further by addressing the domain gap between simplified synthetic and real data, and by incorporating multimodal matching between panoramic reference and perspective query images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.554, + 0.788, + 0.63 + ], + "angle": 0, + "content": "Perspective-to-Panorama Matching. Matching between perspective query and panoramic reference images is a crucial aspect of localization, yet it has received limited attention in the literature. Most systems avoid multimodal matching and instead either match perspective crops to perspective queries [7,40] or panoramic to panoramic images [45]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.788, + 0.765 + ], + "angle": 0, + "content": "While the past discussed geometric fundamentals for integrating catadioptric and perspective camera systems into hybrid stereo setups, including concepts like epipolar constraints [37] and calibration [14], these ideas have not yet found their way into contemporary AI-based localization systems. In the domain of CNNs and panoramic image processing, distortion-aware convolutional filters were proposed by Tateno et al. [39] and further refined by Fernandez et al. [13], particularly for indoor equirectangular images. Unlike traditional convolutions, these distortion-aware filters target the differences between perspective and panoramic images, making them particularly valuable for cross-modal matching." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Recently, perspective-to-panoramic matching advanced, due to sliding window approaches in feature space [32, 36]. It improved visual recognition compared to methods that simply sampled perspective reference images from panoramas [44]. Orhan et al. [33] explored leveraging semantic segmentation similarity to address appearance variations between reference and query images. Our approach" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.265, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "SPVLoc: Semantic Panoramic Viewport Matching for 6D Localization" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.787, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.332, + 0.788, + 0.375 + ], + "angle": 0, + "content": "Fig. 2: Network architecture overview. The information from the query branch is fed into the panorama branch by depth-wise correlation. Three task heads predict the corresponding viewpoint, one task head predicts the relative pose offset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.404, + 0.788, + 0.466 + ], + "angle": 0, + "content": "enhances perspective-to-panorama matching using virtual scene renderings as panoramic references. Inspired by multimodal zero-shot retrieval [18] and object detection, particularly Mercier et al.'s work [29], it utilizes the panorama as a global context for query matching." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.488, + 0.33, + 0.504 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.52, + 0.789, + 0.658 + ], + "angle": 0, + "content": "SPVLoc conducts 6D indoor localization of a 2D RGB query image within a simple semantic textureless 3D scene model. The model adheres to the Structured3D annotation format [46], containing walls, doors, windows, ceiling, and floor as planes. It can be derived from straightforward from building models [46], or from floor plans using contemporary AI tools [8, 28, 34]. SPVLoc first estimates the image's viewpoint within rendered panoramas from the semantic scene model using cross-domain image-to-panorama matching (Sec. 3.1). It then ranks these predictions based on classification accuracy (Sec. 3.4) and performs relative 6D pose regression with respect to the best-matching reference panorama (Sec. 3.2)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.679, + 0.599, + 0.695 + ], + "angle": 0, + "content": "3.1 Semantic Panoramic Viewport Matching" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Problem Definition. Our objective is to precisely estimate the 6D camera pose denoted as \\( P \\) for a given query RGB image \\( I_{q} \\) relative to a semantic 3D model \\( S \\). To achieve this, we reformulate the 6D indoor localization of \\( I_{q} \\) within \\( S \\) as a cross-domain image-to-panorama matching problem. The model is represented by a set of \\( N \\) equirectangular reference layouts \\( L = \\{L_{i} \\mid i = 1, \\dots, N\\} \\) consisting of semantic labels in a single channel. Each rendering comprehensively captures the entire environment around a specific vantage point. The viewpoint \\( V \\) is a distinct cutout section from a panorama, defining the exact region a camera positioned nearby can see. The perspective distortion of \\( V_{P_{i}} \\) contains information about" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.128 + ], + "angle": 0, + "content": "N. Gard et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "the relative pose offset \\( P_{i} \\) between \\( L_{i} \\) and \\( I_{q} \\). Scoring and ranking viewpoints helps in choosing the optimal panorama \\( L^{\\star} \\). The absolute pose \\( P \\) is then formed by concatenating \\( P^{\\star} \\) with \\( P_{i} \\). Identifying the viewpoint and the corresponding pose offset poses several challenges. First, with increasing distance between the two cameras the perspective distortion between image and panorama increases. Second, the matching involves a severe domain gap as a photographic RGB image is matched against a rendered semantic panorama." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.267, + 0.788, + 0.388 + ], + "angle": 0, + "content": "Viewport Prediction. Semantic panorama reference renderings are created utilizing a ray-tracing-based renderer [24]. The viewport mask \\( V_{P} \\) is a binary image that signifies the visibility of points within the equirectangular image from the view of the perspective camera. The calculation involves projecting all 3D points from the panorama in the perspective view. The projection is performed based on the known underlying layout during the training phase. The mask \\( V_{P} \\) is determined by identifying pixels projecting into the perspective image, taking into account the 3D camera pose, and excluding those covered by occlusion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.389, + 0.788, + 0.525 + ], + "angle": 0, + "content": "Additionally, the 2D bounding box \\( v_{bb} \\) is the minimal axis-aligned frame around \\( V_P \\). This bounding box is circularly defined within the \\( 360^\\circ \\) panorama, allowing the overlap on the right image side to re-enter on the left side of the panorama. Each predicted bounding box is associated with a confidence score \\( c_{bb} \\), which quantifies the probability of the bounding box being a valid match. The network's objective is to predict \\( V_P \\), \\( v_{bb} \\) and \\( c_{bb} \\) for each pair of input semantic panorama and query RGB image. The optimal viewpoint corresponds to the predicted bounding box with the highest overall confidence score, \\( c_{bb} \\), when presented with a set of panorama images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.788, + 0.613 + ], + "angle": 0, + "content": "Network Architecture. Our matching is inspired by the object detection strategy of DTOID [29]. Figure 2 illustrates the network architecture. Initially, \\( I_{q} \\) is encoded with an EfficientNet-S backbone. The features obtained after the fourth and the last downsampling layer are preserved. Subsequently, both outputs are rescaled to a tensor of spatial size of \\( 7 \\times 7 \\) and concatenated to a tensor \\( F \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.734 + ], + "angle": 0, + "content": "\\(L\\) is encoded using a DenseNet-121 backbone, extending up to the fifth downsampling block to extract features \\(R\\). The information extracted from the input image guides the panorama branch to estimate the viewport mask and bounding box. This guidance is achieved through the application of a depth-wise correlation technique (denoted as \\(\\star_{\\mathbf{d}}\\)) [6,29]. \\(F\\) is processed through two consecutive convolution blocks, each incorporating convolution, batch normalization, and ELU activation, resulting in a set of \\(3 \\times 3\\) filters \\(\\hat{F}\\) used during the depth-wise correlation \\(R \\star_{\\mathbf{d}} \\hat{F}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.809 + ], + "angle": 0, + "content": "Furthermore, \\( F \\) undergoes compression via global max pooling and is both multiplied and subtracted from \\( R \\), resulting in three feature tensors with injected image information. The features are individually processed with an additional convolution block and then concatenated. The concatenated features further undergo a final convolution block, leading to the correlated features \\( R^{\\star} \\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.788, + 0.841 + ], + "angle": 0, + "content": "A decoder estimates the viewpoint of the perspective camera projected in the layout of the omnidirectional panorama. The viewpoint in the panorama is" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.264, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "SPVLoc: Semantic Panoramic Viewport Matching for 6D Localization" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.788, + 0.269 + ], + "angle": 0, + "content": "estimated as a bounding box with a BBox Class. and BBox Reg. head, similar to RetinaNet [25]. The BBox Class. head's output is optimized using focal loss \\(\\mathcal{L}_1^{VP}\\), while the BBox Reg. head is optimized with a smooth \\(l_{1}\\) loss \\(\\mathcal{L}_2^{VP}\\). The BBox Class. provides a score for a given estimation, ensuring that the bounding box with the highest probability corresponds to the best estimate. Additionally, \\(V_{P}\\) representing the exact viewpoint of the perspective camera is estimated with the aid of another convolutional decoder head and optimized with binary cross entropy loss \\(\\mathcal{L}_3^{VP}\\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.279, + 0.79, + 0.356 + ], + "angle": 0, + "content": "Perspective Supervision. The encoded image features \\( I_{E} \\) are connected to a convolutional decoder, outputting the underlying semantic and normal map to ensure that features closely corresponding to the image content are extracted. The semantics are learned with cross entropy loss \\( \\mathcal{L}_4^{VP} \\) and the normals are learned with cosine loss \\( \\mathcal{L}_5^{VP} \\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.373, + 0.616, + 0.388 + ], + "angle": 0, + "content": "3.2 Feature-Correlation-based Pose Regression" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.394, + 0.789, + 0.499 + ], + "angle": 0, + "content": "The correlated features, denoted as \\( R^{\\star} \\), encode information about the viewpoint of \\( I_{q} \\) within a \\( L_{i} \\), but should also lead to relative pose offset \\( P_{i} \\) between them. Given that \\( L \\) is synthetically generated, the precise position of \\( L_{i} \\) is known. Therefore, determining the relative position of \\( I_{q} \\) in relation to \\( L_{i} \\) is equivalent to establishing its absolute position. All rendered panoramas exhibit identity orientation, due to horizontal alignment and a northward orientation. Consequently, the rotation delta directly corresponds to the absolute rotation of the camera." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.501, + 0.789, + 0.636 + ], + "angle": 0, + "content": "An additional Pose head processes \\( R^{\\star} \\) to estimate the relative pose offset of the camera concerning a panorama rendering (Fig. 3). We first utilize three convolution blocks, whereby the first two are followed by a max-pooling layer, to compress the features. The flattened features form an 1D embedding \\( r^{\\star} \\in \\mathbb{R}^{C} \\). Here, \\( C = 640 \\) for a panorama of dimensions \\( 256 \\times 128 \\). Optionally, to handle different camera field of views during training and testing, \\( r^{\\star} \\) is normalized and multiplied by the horizontal field of view angle in radians. The result is fed into a four-layer MLP, producing the 3D offset to the query image location \\( \\hat{t} i \\in \\mathbb{R}^{3} \\) and a rotation given as the logarithm of the unit quaternion \\( \\hat{r} i = \\log (\\hat{q})_{ij} \\in \\mathbb{R}^{3} \\) [41]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.636, + 0.788, + 0.682 + ], + "angle": 0, + "content": "The MLP is evaluated using two loss functions. \\(\\mathcal{L}_1^P\\) assesses the precision of the relative translation, and \\(\\mathcal{L}_2^P\\) evaluates the accuracy of the final rotation estimation. Both losses are computed using the \\(l_{1}\\) distance." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.7, + 0.373, + 0.715 + ], + "angle": 0, + "content": "3.3 Optimization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.721, + 0.788, + 0.797 + ], + "angle": 0, + "content": "The training loss is the sum of the two losses of the MLP and the five losses of the viewpoint estimation. With seven losses, our multi-task learning has to be balanced by using a different weight for each loss. As in [20, 21], we use an exponential mapping and weight multiple loss functions by considering the homoscedastic uncertainty of each task and learned weighting factors. The final loss is given as:" + }, + { + "type": "equation", + "bbox": [ + 0.357, + 0.804, + 0.788, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\sum_ {i = 1} ^ {2} \\mathcal {L} _ {i} ^ {P} e ^ {- \\beta_ {i}} + \\beta_ {i} + \\sum_ {i = 1} ^ {5} \\mathcal {L} _ {i} ^ {V P} e ^ {- \\gamma_ {i}} + \\gamma_ {i} \\tag {1}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.127 + ], + "angle": 0, + "content": "N. Gard et al." + }, + { + "type": "image", + "bbox": [ + 0.316, + 0.144, + 0.688, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.225, + 0.28, + 0.776, + 0.296 + ], + "angle": 0, + "content": "Fig. 3: The Pose head utilizes convolutions, an MLP, and a FoV-based side input." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.322, + 0.341, + 0.336 + ], + "angle": 0, + "content": "3.4 Inference" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.344, + 0.788, + 0.449 + ], + "angle": 0, + "content": "During inference, panorama positions are determined from a fixed 2D grid superimposed on the floor plan, with virtual cameras set at a constant height denoted as \\( h_{pano} \\), and the grid's dimensions specified as \\( xy_{pano} \\). Virtual cameras are exclusively sampled above floors to ensure valid reference images inside rooms. Setting \\( xy_{pano} \\) as large as possible is advantageous in order to minimize the number of references, given that feature correlation and bounding box regression are executed for every panorama." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.45, + 0.788, + 0.556 + ], + "angle": 0, + "content": "For pose estimation, we select reference positions with the top-\\(n\\) highest bounding box classification scores. The absolute pose is determined by the result of the Pose head, where the selected reference position is added to the translation. To eliminate the necessity for rendering during inference, we propose pre-generating reference panoramas for a floor plan. Additionally, the encoding of these panoramas can be precomputed. This ensures that only the encoded features \\(R\\) need to be stored, significantly streamlining the inference process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.788, + 0.645 + ], + "angle": 0, + "content": "Refinement. Without altering the network architecture or training process, SPVLoc can enhance an estimated pose by rendering a new reference panorama at the estimated position and re-executing the relative pose regression. The smaller step to the actual target position in this second iteration simplifies the problem, allowing for increased precision in the refined pose estimation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.664, + 0.377, + 0.681 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.692, + 0.787, + 0.723 + ], + "angle": 0, + "content": "To validate our method we conduct several experiments and make use of two publicly available and frequently used datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Structured3D (S3D) [46] consists of 3500 near-photorealistic professionally modeled indoor environments, each annotated with ground truth 3D structure information. It encompasses 21835 panoramic images, each provided in both furnished and unfurnished configurations. The authors define a unified representation that describes scenes using geometric primitives such as planes, lines, junction points, and semantic annotations including floor, ceiling, wall, door, and window classes." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.264, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "SPVLoc: Semantic Panoramic Viewport Matching for 6D Localization" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.144, + 0.788, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.248, + 0.788, + 0.277 + ], + "angle": 0, + "content": "Fig. 4: ZinD data preparation. Annotations generate 3D reference models (left), while resampled panoramas create perspective train and test images (right)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.306, + 0.789, + 0.428 + ], + "angle": 0, + "content": "Zillow Indoor (ZInD) [10] includes 67,448 panorama images from 1,575 unfurnished residential homes, globally aligned and registered to floor plans. The 3D annotations of panoramas within a scene are adequate for transforming them into the S3D representation format. Before training, we transform all data to that format. Figure 4 shows an example reference model generated from ZInD annotations. Unlike S3D, ZInD includes room openings such as corridors connected to rooms without doors, enhancing structural variation in semantic renderings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.431, + 0.79, + 0.628 + ], + "angle": 0, + "content": "For both datasets, we employ the official train-test scene split. To enhance control over the field of view and angle variation, we sample perspective test images from the RGB panoramas. They are registered in the 3D model, so the semantics are known for each perspective image. We compile test-sets suitable for 6D pose estimation and structured evaluation. The test sets consist of square perspective images with a viewing angle of 60, 90, and \\(120^{\\circ}\\), respectively, along with variations in roll and pitch, and random yaw angles for ZinD and S3D. To ensure meaningful views, we apply two criteria on the sampled yaw. First, the images contain a minimum of three semantic classes, like wall-floor-ceiling or wall-window-floor. Second, the camera faces inside the room rather than towards the next wall. The criteria are applied to the \\(60^{\\circ}\\) images (4420 for ZiND and 1133 for S3D). Subsequent variations in opening angles and roll-pitch configurations are generated from this starting point." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.65, + 0.321, + 0.663 + ], + "angle": 0, + "content": "4.1 Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.788, + 0.751 + ], + "angle": 0, + "content": "For evaluation, we report the fraction of frames localized correctly within a threshold \\(\\tau\\). We also list the median rotation and translation error, as well as the fraction of frames localized within \\(\\tau\\) and an additional angle threshold \\(v\\). In the retrieval process, the top-3 best matches are selected, and we report recall metrics for both top-1 and the best among the top-3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Training. For the state of the art comparison, we train a model with a variable field of view, while other experiments use a model trained with a fixed opening angle of \\(90^{\\circ}\\). The model with variable viewing angles randomly samples views between 45 and \\(135^{\\circ}\\) during training. We consistently render \\(s\\) randomly positioned panoramas around the query within a radius of \\(\\pm r_1\\) in \\(xy\\) direction and" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.127 + ], + "angle": 0, + "content": "N. Gard et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 1: Comparison with state-of-the-art baseline on ZInD and S3D: Median translation \\((\\mathrm{terr}_{\\mathrm{xy}})\\) in cm, rotation error (rerrayw) in deg. for instances localized under 1m. Reporting recall at various translation accuracy levels, recall for inliers (<1m and <30°), and top-3 recall at 1m. Ours† refers to our model tested with a slightly different testset, incorporating pitch and roll rotation variation (\\(\\pm 10^{\\circ}\\))." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.227, + 0.784, + 0.344 + ], + "angle": 0, + "content": "
ZInDStructured3D (Furnishing-Level : Full)
QueryMethod<1m med terr (cm)<1m med rerr (deg)10cm (%)50cm (%)1m (%)1m & 30° (%)top-3 1m (%)<1m med terr (cm)<1m med rerr (deg)10cm (%)50cm (%)1m (%)1m & 30° (%)top-3 1m(%)
Perspective 60° FoVLASER [30]33.531.034.0547.2464.5263.5378.6417.600.8416.4663.2268.3967.3486.08
Ours17.161.5021.0481.3386.7686.0994.1212.251.2134.9487.1389.4988.7996.15
\\( Ours^† \\)17.771.6319.5981.1387.1086.4394.8613.311.2430.8284.6888.1887.4895.80
Perspective 90° FoVLASER [30]26.280.778.6967.0180.9080.2589.7314.600.6723.4772.7776.8076.1891.94
Ours13.631.3430.5785.5990.8490.3496.369.170.8350.0989.6792.1291.8697.37
\\( Ours^† \\)14.251.3729.5985.9091.3390.9596.9710.090.9145.6289.3292.1291.9496.85
Perspective 120° FoVLASER [30]23.100.7812.6975.0985.6385.2092.3812.900.7629.0777.7682.1481.3594.13
Ours14.781.4529.0585.1491.7291.2796.9211.690.8739.4988.0092.1291.9497.37
\\( Ours^† \\)15.131.5328.1286.1392.7192.3597.2411.770.8838.1889.6793.0892.9197.55
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.371, + 0.79, + 0.538 + ], + "angle": 0, + "content": "\\(\\pm r_2\\) in \\(z\\) (upward) direction. For a single query, the viewpoint and relative pose offset is therefore calculated against multiple references. In our experimental setup, we set \\(s\\) to 4, \\(r_1\\) to 1.4 meters, and \\(r_2\\) to 30 centimeters. Additionally, one random negative example is generated outside \\(r_1\\), likely in a different room, with the target to enhance the network's ability to distinguish subtle room differences. Images are sampled with random yaw and \\(\\pm 10^{\\circ}\\) random pitch and roll angle. The batch size is set to 40, comprising 40 query images and 200 panoramas, and the training is performed on a single NVIDIA A100 GPU. Query images with fewer than three semantic classes are disregarded during the loss calculation process. Training runs for approximately 42,000 steps with an initial learning rate of \\(2.5 \\times 10^{-4}\\), halved two times during training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.552, + 0.789, + 0.645 + ], + "angle": 0, + "content": "Testing. During testing, panoramas are sampled on a \\(1.2 \\times 1.2\\) m grid, requiring significantly fewer samples than other rendering-based approaches such as LaLaLoc [17], which uses a \\(0.5 \\times 0.5\\) m grid. Additionally, for a scene, all panoramic views can be encoded initially, and during inference, only the query image needs to be encoded. If not otherwise noted the studies are conducted on ZInD with fixed 90 degree FoV, reporting the 3D rotation and translation error." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.665, + 0.577, + 0.68 + ], + "angle": 0, + "content": "4.2 Comparison with the State of the Art" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.779 + ], + "angle": 0, + "content": "To our knowledge, SPVLoc is the first method to perform 6D localization in unseen scenes with respect to semantic 3D models. The state of the-art method for scene-independent global localization, LASER [30], takes perspective views as input and matches them against a semantic floor plan. Hence, it estimates only two position and one rotation degree of freedom. Nonetheless, it serves as a strong baseline to assess the localization accuracy of our method." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Table 1 shows the results on ZInD and S3D. Given that we compare against a 2D method, our evaluation also includes recall based on 2D distance and angular error derived from the yaw angle. The yaw angle is extracted from the estimated rotation matrix. Notably, SPVLoc without refinement outperforms LASER in" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.264, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "SPVLoc: Semantic Panoramic Viewport Matching for 6D Localization" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 2: Ablation study. SPVLoc variations compared on ZInD with \\(90^{\\circ}\\) FoV. In contrast to Tab. 1, terr denotes 3D position errors, and rerr denotes 3D rotation errors." + }, + { + "type": "table", + "bbox": [ + 0.299, + 0.185, + 0.694, + 0.314 + ], + "angle": 0, + "content": "
Abl. TypeMethod<1m med terr (cm)<1m med rerr (deg)10cm (%)50cm (%)1m (%)1m & 30° (%)top-3 1m (%)
(Ours)base14.312.0526.8886.9291.9091.6196.56
+ refine12.321.9633.6990.8492.7691.74-
Loss- query dec.15.702.2321.9785.5090.7990.4396.45
- view dec.15.192.2123.6085.3690.4589.8296.49
- view/query dec.16.222.1921.5684.5089.5989.1296.36
Training- neg. samples15.062.2822.6076.6582.1981.7492.99
Arch.Δ Resnet18 (query)15.562.5620.8886.5690.7290.3296.56
Δ Equiconv (pano)14.982.1325.0787.4491.7091.3196.97
Input+ normal14.682.2825.0985.9390.8490.2396.47
+ depth14.062.1027.6088.0592.1591.7996.74
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.341, + 0.788, + 0.461 + ], + "angle": 0, + "content": "terms of localization accuracy, exhibiting higher recall across all categories, an increased inlier rate, and significantly reduced median translation error. In terms of 2D rotation, our method is slightly less precise. This is because LASER estimates yaw directly while we estimate a full rotation matrix. Estimating camera height and roll and pitch angle is an advantage of our method. In each last row \\((Ours^{\\dagger})\\), we evaluate our method on an alternate test set with roll and pitch variations \\((\\pm 10^{\\circ})\\). The results suggest that these variations do not substantially impact the accuracy of our method for 2D localization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.79, + 0.6 + ], + "angle": 0, + "content": "Regarding further methods, LASER has showcased a significant performance improvement over PF-net [19] and non-AI-based Monte-Carlo Localization (MCL) [11]. PF-net, originally designed for sequential time updating, does not perform well in single-query localization scenarios. MCL neglects semantic annotations, uses LIDAR instead of images, which also has been shown to drastically reduce performance [30]. Our substantial improvement over LASER therefore also illustrates a performance gain over aforementioned methods. Another experiment using the pre-rendered perspective images of S3D, including a comparison to \\(\\mathrm{F}^3\\mathrm{Loc}\\) [9], is part of the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.62, + 0.434, + 0.634 + ], + "angle": 0, + "content": "4.3 Performance Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.75 + ], + "angle": 0, + "content": "Ablation Study. We commence our experimentation with the full network (base) configuration. Subsequently, we modify the network by removing specific components (Tab. 2). Initially, we test the influence of perspective supervision (-query dec.) and the view segment task head (-view dec.) by excluding them during training. Their exclusion reduces network performance, likely because the network's ability to understand the image content and establish correspondences is diminished. Removing both (-view/query dec.) further reduces accuracy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.81 + ], + "angle": 0, + "content": "We then exclude the negative samples from different rooms during training (-neg. samples). This significantly reduces the amount of correctly localized images. It supports the hypothesis that teaching the network to recognize spaces where there is no match helps to improve the accuracy of predictions." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "To modify the architecture, the image encoder EfficientNet-S is replaced with a smaller ResNet-18 [15] (Δ Resnet18), resulting in a performance drop. This" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.127 + ], + "angle": 0, + "content": "N. Gard et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.787, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.341, + 0.788, + 0.385 + ], + "angle": 0, + "content": "Fig. 5: Qualitative localization results. Top to bottom - query, rendering with top-1 estimated pose, panorama with estimated viewport, map. Green box: success for top-1 match. Yellow box: success for top-2 match. Red box: failure case." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.414, + 0.788, + 0.504 + ], + "angle": 0, + "content": "suggests an increased demand for network capacity to analyze the images. In the panorama encoder, replacing all convolutional layers with Equiconv [13] (ΔEquiconv), a specialized convolution for panorama images that remaps query locations from Euclidean to spherical space, does not result in a performance gain. This suggests that the capacity of the small DenseNet-121 is sufficient for analyzing the low-resolution semantic inputs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.505, + 0.79, + 0.581 + ], + "angle": 0, + "content": "Lastly, we incorporate additional input modalities for the panoramas into the panorama encoder. While the addition of world coordinate normals (+normals) did not improve the result, including an extra depth map slightly enhanced the outcome (+depth). This suggests that there are cases where additional information is encoded in the depth that cannot be inferred from semantics alone." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.597, + 0.789, + 0.793 + ], + "angle": 0, + "content": "Qualitative Results. Figure 5 shows selected SPVLoc outcomes. The top row displays camera images, and the second row shows renderings from the semantic 3D model with estimated top-1 poses. All green boxed images are correctly localized for the top-1 match. In the yellow box, the top-1 match fails, but the top-2 match is accurate. The red box example cannot be localized. Across all examples, a pose with high visual similarity is found. Ambiguous examples arise from similar layouts occurring in multiple rooms. Failure cases are attributed to either images lacking sufficient structural information for matching or outliers in the data, such as those taken on balconies or in dark cellars. The third row displays the chosen reference panorama, the selected viewpoint bounding box corresponding to the camera image, and a semi-transparent overlay representing the estimated viewpoint mask. Additional detailed visualizations are part of the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Influence of Sampling Grid Sizes and Refinement. Reference positions result from overlaying a grid onto the floorplan, and the grid density significantly" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.264, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SPVLoc: Semantic Panoramic Viewport Matching for 6D Localization" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.52, + 0.173 + ], + "angle": 0, + "content": "Table 3: Grid size vs. accuracy: Deviation of \\(10\\mathrm{cm}\\) and \\(1\\mathrm{m}\\) recall from (boxed) baseline." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.185, + 0.504, + 0.309 + ], + "angle": 0, + "content": "
Sample dist.Samples per \\( m^2 \\)Adv.10cm (%)10cm + refine1m (%)1m + refine
0.72.07+1.00+7.49+1.86+2.38
0.91.25+0.95+6.99+1.79+2.08
1.20.70\\( \\boxed{26.88} \\)+6.81\\( \\boxed{91.90} \\)+0.86
1.20.48✓+0.57+6.95-0.63+0.61
1.50.45-1.00+6.63-2.47-1.02
1.50.31✓-1.04+6.47-2.67-0.38
2.00.26-4.19+3.60-9.05-5.79
2.00.20✓-1.76+5.81-3.91-1.18
" + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.151, + 0.605, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.151, + 0.69, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.152, + 0.774, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.532, + 0.245, + 0.787, + 0.288 + ], + "angle": 0, + "content": "Fig. 6: Reference positions for global grid of size \\(0.7\\mathrm{m}\\), \\(1.5\\mathrm{m}\\) and local grid of size \\(1.5\\mathrm{m}\\) (from left to right)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.338, + 0.789, + 0.549 + ], + "angle": 0, + "content": "affects SPVLoc's efficiency and inference time. Minimizing references reduces rendering effort and requires the model to estimate larger offsets. E.g. to increase the grid from \\(0.7m\\) to \\(2m\\) reduces the number of samples per \\(m^2\\) by a factor of 10. We vary the grid size to analyze its impact on performance. We also experiment with a local grid per room substituting the global grid (Adv.) reducing the risk of missing rooms entirely. The local grid is centered on the room outline and additionally places one camera in the middle of each opening. Figure 6 shows different grids for an example plan. Table 3 illustrates the relationships among grid size, global and local sampling, and refinement. Refinement notably improves \\(10\\mathrm{cm}\\) recall across configurations. For the largest local grid with refinement, the \\(1\\mathrm{m}\\) recall only decreases by \\(1.18\\%\\) compared to the baseline configuration. In summary, a trade-off between refinement and dense/sparse scene sampling exists, yet our method consistently delivers strong results with very sparse sampling, making it suitable for large scenes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.566, + 0.79, + 0.689 + ], + "angle": 0, + "content": "Influence of Focal Length. Table 4 depicts the performance difference between a variant trained with a known camera focal length and a focal length-independent network trained with varying focal lengths (ranging from 45 to 135 degrees). The fixed focal length network performs slightly better with matching images but loses precision with test images of different focal lengths. In contrast, the focal length-independent network demonstrates the capability to handle such differences. We argue that the significantly enhanced flexibility of the latter justifies the minor performance degradation in practical, real-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Influence of Camera Rotation Angles. The range of pitch and roll angles the network is able to handle is determined by the variations of angles seen during training. We train the network for ZInD using a angle variation of \\(\\pm 22^{\\circ}\\) and test with images where the range of variation increases from \\(0^{\\circ}\\) to \\(20^{\\circ}\\) in steps of \\(\\pm 5^{\\circ}\\). Table 5 shows only a slight increase in median errors for rotation and translation as the rotation angles in test images increase. The chosen ranges represent a significant coverage and offer substantial improvements over existing methods. In indoor scenes, larger camera tilt often results in images dominated by the ceiling or floor, rendering them unsuitable for semantic matching." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.128 + ], + "angle": 0, + "content": "N. Gard et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.522, + 0.187 + ], + "angle": 0, + "content": "Table 4: The fixed FoV network excels at a specific angle but loses precision with changing FoV. The proposed network remains stable." + }, + { + "type": "table", + "bbox": [ + 0.229, + 0.199, + 0.498, + 0.294 + ], + "angle": 0, + "content": "
Input FoVTrain FoV<1m med terr (cm)<1m med rerr (deg)10cm (%)1m (%)
90°90°14.312.0526.8891.90
45 - 135°16.112.2021.8191.45
80°90°37.512.371.6790.14
45 - 135°16.752.1919.3090.50
" + }, + { + "type": "table_caption", + "bbox": [ + 0.532, + 0.154, + 0.788, + 0.181 + ], + "angle": 0, + "content": "Table 5: Effect of increasing angle variation in test data." + }, + { + "type": "table", + "bbox": [ + 0.533, + 0.195, + 0.78, + 0.283 + ], + "angle": 0, + "content": "
MetricRoll/Pitch Variation
±0°±5°±10°±15°±20°
<1m med. terr14.4614.8814.8815.2115.91
<1m med. verr2.352.362.382.512.60
10cm (%)24.6423.9424.4623.6922.74
1m (%)91.4591.4791.9992.3892.60
" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.308, + 0.787, + 0.337 + ], + "angle": 0, + "content": "Table 6: Timing. Performance on ZInD using a NVIDIA Tesla V100, following the configurations outlined in Sec. 4.3, and compared against results from [30]." + }, + { + "type": "table", + "bbox": [ + 0.299, + 0.348, + 0.692, + 0.398 + ], + "angle": 0, + "content": "
MethodSampling Time (s)Query FpsQuery Module Time (ms)
PF-net [19]48.95 ± 38.955.06 ± 1.77Eff.Net V2 S4.4
LASER [30]0.97 ± 1.098.31 ± 0.64Correlation35.7
Ours1.66 ± 1.0628.63 ± 11.54Pose1.1
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.427, + 0.788, + 0.504 + ], + "angle": 0, + "content": "Limitations. In large spaces with repeating room layouts, the approach's effectiveness may be constrained by the limited level of detail in the semantic reference model. However, enhancing the S3D format by adding more semantic classes and structural information, such as staircases or sanitary installations, could reduce ambiguities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.522, + 0.788, + 0.658 + ], + "angle": 0, + "content": "Timing. We split the measured time into scene sampling/preparation time, including position sampling, panorama rendering and encoding, and inference/query time, including query image encoding, feature correlation with bounding box classification, and the Pose task head, but no refinement. Table 6 shows that our method exhibits a considerably lower query time than LASER [30] and PF-net [19]. This difference places our approach close to real-time for moderately sized scenes. Currently, scene sampling time is slower than [30], mainly due to rendering on the CPU without parallelization [24]. Parallelizing this process using GPU or CPU could significantly accelerate scene preparation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.684, + 0.36, + 0.7 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "This paper introduces a method for scene-independent model-based 6D localization for indoor scenes, involving a novel method for multimodal image matching (panorama to perspective, RGB to semantic). The matching and retrieval is efficient and scalable under sparse reference sampling. The localization accuracy and inference speed outperform state of the art methods and the inclusion of the 3D-model, reduces ambiguities to precisely estimate 6D poses. Future work involves combining localization and image analysis to enhance digital building models or exploring applications in augmented reality scenarios." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.264, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "SPVLoc: Semantic Panoramic Viewport Matching for 6D Localization" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.788, + 0.194 + ], + "angle": 0, + "content": "Acknowledgements. This work is partially funded by the German Federal Ministry of Economic Affairs and Climate Action (BIMKIT, 01MK21001H) and the German Federal Ministry for Digital and Transport (EConoM, 19OI22009C)." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.213, + 0.324, + 0.229 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.241, + 0.785, + 0.283 + ], + "angle": 0, + "content": "1. Acharya, D., Khoshelham, K., Winter, S.: BIM-PoseNet: Indoor camera localisation using a 3D indoor model and deep learning from synthetic images. ISPRS journal of photogrammetry and remote sensing 150, 245-258 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.283, + 0.787, + 0.323 + ], + "angle": 0, + "content": "2. Acharya, D., Ramezani, M., Khoshelham, K., Winter, S.: Bim-tracker: A model-based visual tracking approach for indoor localisation using a 3d building model. ISPRS Journal of Photogrammetry and Remote Sensing 150, 157-171 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.323, + 0.787, + 0.364 + ], + "angle": 0, + "content": "3. Acharya, D., Tatli, C.J., Khoshelham, K.: Synthetic-real image domain adaptation for indoor camera pose regression using a 3d model. ISPRS Journal of Photogrammetry and Remote Sensing 202, 405-421 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.364, + 0.787, + 0.419 + ], + "angle": 0, + "content": "4. Acharya, D., Tennakoon, R., Muthu, S., Khoshelham, K., Hoseinnezhad, R., Bab-Hadiashar, A.: Single-image localisation using 3D models: Combining hierarchical edge maps and semantic segmentation for domain adaptation. Automation in Construction 136, 104-152 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.419, + 0.787, + 0.446 + ], + "angle": 0, + "content": "5. Agarwal, S., Furukawa, Y., Snavely, N., Simon, I., Curless, B., Seitz, S.M., Szeliski, R.: Building rome in a day. Communications of the ACM 54(10), 105-112 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.446, + 0.787, + 0.472 + ], + "angle": 0, + "content": "6. Ammirato, P., Fu, C.Y., Shvets, M., Kosecka, J., Berg, A.C.: Target driven instance detection. arXiv preprint arXiv:1803.04610 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.472, + 0.787, + 0.513 + ], + "angle": 0, + "content": "7. Arandjelovic, R., Gronat, P., Torii, A., Pajdla, T., Sivic, J.: Netvlad: Cnn architecture for weakly supervised place recognition. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.513, + 0.787, + 0.554 + ], + "angle": 0, + "content": "8. Cambeiro, A.B., Trzeciakiewicz, M., Hilsmann, A., Eisert, P.: Automatic reconstruction of semantic 3d models from 2d floor plans. In: Proc. Int. Conf. on Machine Vision Applications (MVA). Hamamatsu, Japan (July 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.554, + 0.787, + 0.595 + ], + "angle": 0, + "content": "9. Chen, C., Wang, R., Vogel, C., Pollefeys, M.: F3Loc: Fusion and Filtering for Floorplan Localization. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.595, + 0.787, + 0.65 + ], + "angle": 0, + "content": "0. Cruz, S., Hutchcroft, W., Li, Y., Khosravan, N., Boyadzhiev, I., Kang, S.B.: Zillow indoor dataset: Annotated floor plans with 360deg panoramas and 3d room layouts. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2133-2143 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.65, + 0.787, + 0.691 + ], + "angle": 0, + "content": "1. Dellaert, F., Fox, D., Burgard, W., Thrun, S.: Monte carlo localization for mobile robots. In: Proceedings IEEE International Conference on Robotics and Automation. vol. 2, pp. 1322-1328 (1999)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.691, + 0.787, + 0.732 + ], + "angle": 0, + "content": "2. Ding, M., Wang, Z., Sun, J., Shi, J., Luo, P.: CamNet: Coarse-to-fine retrieval for camera re-localization. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2871-2880 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.732, + 0.787, + 0.772 + ], + "angle": 0, + "content": "3. Fernandez-Labrador, C., Facil, J.M., Perez-Yus, A., Demonceaux, C., Civera, J., Guerrero, J.J.: Corners for layout: End-to-end layout recovery from 360 images. IEEE Robotics and Automation Letters 5(2), 1255-1262 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.787, + 0.799 + ], + "angle": 0, + "content": "4. He, B., Chen, Z., Li, Y.: Calibration method for a central catadioptric-perspective camera system. JOSA A 29(11), 2514-2524 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "5. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 770-778 (2016)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.241, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.127 + ], + "angle": 0, + "content": "N. Gard et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.787, + 0.189 + ], + "angle": 0, + "content": "16. Howard-Jenkins, H., Prisacariu, V.A.: LaLaLoc++: Global Floor Plan Comprehension for Layout Localisation in Unvisited Environments. In: Proc. European Conference on Computer Vision (ECCV). pp. 693-709. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "17. Howard-Jenkins, H., Ruiz-Sarmiento, J.R., Prisacariu, V.A.: LaLaLoc: Latent Layout Localisation in Dynamic, Unvisited Environments. In: Proc. IEEE/CVF International Conference on Computer Vision (ICCV) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.232, + 0.787, + 0.272 + ], + "angle": 0, + "content": "18. Janik, M., Gard, N., Hilsmann, A., Eisert, P.: Zero in on Shape: A Generic 2D-3D Instance Similarity Metric Learned from Synthetic Data. In: Proc. IEEE International Conference on Image Processing (ICIP). pp. 2638-2642. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.273, + 0.787, + 0.3 + ], + "angle": 0, + "content": "19. Karkus, P., Hsu, D., Lee, W.S.: Particle filter networks with application to visual localization. In: Conference on robot learning. pp. 169-178. PMLR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.301, + 0.787, + 0.342 + ], + "angle": 0, + "content": "20. Kendall, A., Cipolla, R.: Geometric loss functions for camera pose regression with deep learning. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5974-5983 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.787, + 0.384 + ], + "angle": 0, + "content": "21. Kendall, A., Gal, Y., Cipolla, R.: Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7482-7491 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.787, + 0.425 + ], + "angle": 0, + "content": "22. Kendall, A., Grimes, M., Cipolla, R.: Posenet: A convolutional network for real-time 6-dof camera relocalization. In: Proc. IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2938-2946 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.787, + 0.467 + ], + "angle": 0, + "content": "23. Kim, J., Jeong, J., Kim, Y.M.: Fully geometric panoramic localization. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 20827-20837 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.468, + 0.787, + 0.509 + ], + "angle": 0, + "content": "24. Li, T.M., Aittala, M., Durand, F., Lehtinen, J.: Differentiable Monte Carlo Ray Tracing through Edge Sampling. ACM Trans. Graph. (Proc. SIGGRAPH Asia) 37(6), 222:1-222:11 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.787, + 0.55 + ], + "angle": 0, + "content": "25. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proc. IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2980-2988 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.787, + 0.592 + ], + "angle": 0, + "content": "26. Liu, C., Schwing, A., Kundu, K., Urtasun, R., Fidler, S.: Rent3d: Floor-plan priors for monocular layout estimation. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.787, + 0.633 + ], + "angle": 0, + "content": "27. Liu, L., Li, H., Dai, Y.: Efficient global 2d-3d matching for camera localization in a large-scale 3d map. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2372-2381 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.787, + 0.675 + ], + "angle": 0, + "content": "28. Lv, X., Zhao, S., Yu, X., Zhao, B.: Residential floor plan recognition and reconstruction. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16717-16726 (June 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.787, + 0.702 + ], + "angle": 0, + "content": "29. Mercier, J.P., Garon, M., Giguere, P., Lalonde, J.F.: Deep Template-Based Object Instance Detection. In: Proc. WACV. pp. 1507-1516 (January 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.787, + 0.758 + ], + "angle": 0, + "content": "30. Min, Z., Khosravan, N., Bessinger, Z., Narayana, M., Kang, S.B., Dunn, E., Boyadzhiev, I.: Laser: Latent space rendering for 2d visual localization. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11122-11131 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.787, + 0.8 + ], + "angle": 0, + "content": "31. Naseer, T., Burgard, W.: Deep regression for monocular camera-based 6-dof global localization in outdoor environments. In: Proc. International Conference on Intelligent Robot Systems (IROS). pp. 1525-1530. IEEE (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.787, + 0.842 + ], + "angle": 0, + "content": "32. Orhan, S., Bastanlar, Y.: Efficient search in a panoramic image database for long-term visual localization. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1727-1734 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.842 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.265, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SPVLoc: Semantic Panoramic Viewport Matching for 6D Localization" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.203 + ], + "angle": 0, + "content": "33. Orhan, S., Guerrero, J.J., Bastanlar, Y.: Semantic Pose Verification for Outdoor Visual Localization With Self-Supervised Contrastive Learning. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops. pp. 3989-3998 (June 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.787, + 0.231 + ], + "angle": 0, + "content": "34. Park, S., Kim, H.: 3dplannet: generating 3d models from 2d floor plan images using ensemble methods. Electronics 10(22), 2729 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.787, + 0.286 + ], + "angle": 0, + "content": "35. Sarlin, P.E., Unagar, A., Larsson, M., Germain, H., Toft, C., Larsson, V., Pollefeys, M., Lepetit, V., Hammarstrand, L., Kahl, F., et al.: Back to the feature: Learning robust camera localization from pixels to pose. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3247-3257 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.787, + 0.342 + ], + "angle": 0, + "content": "36. Shi, Z., Shi, H., Yang, K., Yin, Z., Lin, Y., Wang, K.: PanoVPR: Towards Unified Perspective-to-Equirectangular Visual Place Recognition via Sliding Windows across the Panoramic View. In: 2023 IEEE 26th International Conference on Intelligent Transportation Systems (ITSC). pp. 1333-1340 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.787, + 0.383 + ], + "angle": 0, + "content": "37. Sturm, P.: Mixing catadioptric and perspective cameras. In: Proceedings of the IEEE Workshop on Omnidirectional Vision 2002. Held in conjunction with ECCV'02. pp. 37-44. IEEE (2002)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.787, + 0.438 + ], + "angle": 0, + "content": "38. Taira, H., Okutomi, M., Sattler, T., Cimpoi, M., Pollefeys, M., Sivic, J., Pajdla, T., Torii, A.: InLoc: Indoor visual localization with dense matching and view synthesis. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7199-7209 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.787, + 0.481 + ], + "angle": 0, + "content": "39. Tateno, K., Navab, N., Tombari, F.: Distortion-aware convolutional filters for dense prediction in panoramic images. In: Proc. European Conference on Computer Vision (ECCV). pp. 707-722 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.787, + 0.522 + ], + "angle": 0, + "content": "40. Torii, A., Arandjelovic, R., Sivic, J., Okutomi, M., Pajdla, T.: 24/7 place recognition by view synthesis. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1808-1817 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.787, + 0.563 + ], + "angle": 0, + "content": "41. Türkoglu, M.Ö., Brachmann, E., Schindler, K., Brostow, G., Monszpart, A.: Visual Camera Re-Localization Using Graph Neural Networks and Relative Pose Supervision. In: Proc. International Conference on 3D Vision (3DV) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.787, + 0.605 + ], + "angle": 0, + "content": "42. Wang, S., Fidler, S., Urtasun, R.: Lost shopping! monocular localization in large indoor spaces. In: Proc. IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2695-2703 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.787, + 0.633 + ], + "angle": 0, + "content": "43. Xia, J., Gong, J.: Precise indoor localization with 3d facility scan data. Computer-Aided Civil and Infrastructure Engineering 37(10), 1243-1259 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.787, + 0.674 + ], + "angle": 0, + "content": "44. Zamir, A.R., Shah, M.: Accurate image localization based on google maps street view. In: Proc. European Conference on Computer Vision (ECCV). pp. 255-268. Springer, Heraklion, Greece (Sep 2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.787, + 0.715 + ], + "angle": 0, + "content": "45. Zhang, C., Budvytis, I., Liwicki, S., Cipolla, R.: Lifted semantic graph embedding for omnidirectional place recognition. In: Proc. International Conference on 3D Vision (3DV). pp. 1401-1410. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.787, + 0.758 + ], + "angle": 0, + "content": "46. Zheng, J., Zhang, J., Li, J., Tang, R., Gao, S., Zhou, Z.: Structured3D: A large photo-realistic dataset for structured 3d modeling. In: Proc. European Conference on Computer Vision (ECCV). pp. 519-535. Springer, Glasgow, UK (Aug 2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.758 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/SPVLoc_ Semantic Panoramic Viewport Matching for 6D Camera Localization in Unseen Environments/a0f540fa-1e67-4e90-a2ae-34590f8ccef2_origin.pdf b/2024/SPVLoc_ Semantic Panoramic Viewport Matching for 6D Camera Localization in Unseen Environments/a0f540fa-1e67-4e90-a2ae-34590f8ccef2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..413fb65fd731502b722c8d7f3ca5d431d1257783 --- /dev/null +++ b/2024/SPVLoc_ Semantic Panoramic Viewport Matching for 6D Camera Localization in Unseen Environments/a0f540fa-1e67-4e90-a2ae-34590f8ccef2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf0d0f97c67c597d8f7f6c5ab003a9a3db164f38b77e0632c4a02bfc29b0d618 +size 3123131 diff --git a/2024/SPVLoc_ Semantic Panoramic Viewport Matching for 6D Camera Localization in Unseen Environments/full.md b/2024/SPVLoc_ Semantic Panoramic Viewport Matching for 6D Camera Localization in Unseen Environments/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9b6aeac25126658dae77a9b9287a7861df9dbae5 --- /dev/null +++ b/2024/SPVLoc_ Semantic Panoramic Viewport Matching for 6D Camera Localization in Unseen Environments/full.md @@ -0,0 +1,264 @@ +# SPVLoc: Semantic Panoramic Viewport Matching for 6D Camera Localization in Unseen Environments + +Niklas Gard $^{1,2}$ , Anna Hilsmann $^{1}$ , and Peter Eisert $^{1,2}$ + +1 Fraunhofer Heinrich Hertz Institute, HHI, Berlin, Germany + +{first.last}@hhi.fraunhofer.de + +2 Humboldt University of Berlin, Germany + +Abstract. In this paper, we present SPVLoc, a global indoor localization method that accurately determines the six-dimensional (6D) camera pose of a query image and requires minimal scene-specific prior knowledge and no scene-specific training. Our approach employs a novel matching procedure to localize the perspective camera's viewpoint, given as an RGB image, within a set of panoramic semantic layout representations of the indoor environment. The panoramas are rendered from an textured 3D reference model, which only comprises approximate structural information about room shapes, along with door and window annotations. We demonstrate that a straightforward convolutional network structure can successfully achieve image-to-panorama and ultimately image-to-model matching. Through a viewpoint classification score, we rank reference panoramas and select the best match for the query image. Then, a 6D relative pose is estimated between the chosen panorama and query image. Our experiments demonstrate that this approach not only efficiently bridges the domain gap but also generalizes well to previously unseen scenes that are not part of the training data. Moreover, it achieves superior localization accuracy compared to the state of the art methods and also estimates more degrees of freedom of the camera pose. Our source code is publicly available at: https://fraunhoferhhi.github.io/spvloc. + +Keywords: Localization $\cdot$ 6D pose estimation $\cdot$ Multimodal matching + +# 1 Introduction + +Indoor camera localization is a fundamental challenge of computer vision, aiming to precisely determine a camera's position and orientation within an indoor environment. Applications include navigation [38,42], augmented reality (AR) [2], building management, and building digitization [26]. Current methods often depend on extensive scene-specific data, acquired through manual creation of databases with localized images, depth information, or point clouds [22,38,43]. However, this limits their applicability, especially in situations where exploring the building is impractical or data acquisition is costly and time-consuming. + +Recent methods incorporate synthetic data during training, but these localization models require retraining or fine-tuning for new scenes using detailed BIM + +![](images/102ed9dd0baa77f7c8c51552017da18c2ccdcc84dea05456ec39e50272a1495d.jpg) +Fig. 1: SPVLoc for 6D indoor localization. Our method calculates the indoor 6D camera pose by determining the image position and orientation relative to synthetic panoramas. The best panoramic match is found through semantic viewpoint matching. + +(Building Information Modeling) models [1,3]. The adoption of the Structured3D format [46] streamlines and standardizes 3D structure annotation by simplifying complex models into their most fundamental geometric representations. Methods that match a query image to this simplified structural representation exhibit robustness in unexplored, dynamic environments [17,30]. However, existing methods [9,30] mainly focus on 2D perspective camera localization with respect to floor plans, missing two degrees of freedom for rotation and one degree of freedom for position. Also, some methods solely rely on panoramic queries [16,17,23]. These queries are typically captured under controlled conditions, involving a tripod-mounted camera with a precise horizontal alignment, and a fixed camera height [17]. However, in AR applications or scenarios where archived photos must be accurately registered, e.g. to annotate visible building details into a digital building model, idealized capture conditions cannot always be guaranteed. In such cases, the demand for precise 6D localization becomes imperative. + +In this paper, we present SPVLoc, a novel method for indoor 6D camera localization in unseen environments. It represents a 3D scene with a set of synthetically generated omnidirectional panorama images depicting fundamental semantic labels such as door, window, wall, floor, and ceiling. The core idea involves correlating the deep embedding of a perspective RGB camera image with the embeddings of panoramas to estimate the camera viewpoint - the precise region visible to the camera within the panorama. Leveraging a RetinaNet-like architecture [25], the 2D-bounding box around the viewpoint is predicted and classified. Retrieval of the closest panorama is performed by selecting the panorama with the globally highest classification score. The exact 6D pose is then found via relative pose estimation with an MLP starting from the panorama's position. Our method offers several advantages. It relies solely on a minimalist + +3D semantic building model [46], which could even be automatically extracted from simple 2D plans [8,28,34]. By bridging the domain gap between semantic representation and real images, it facilitates matching in unseen environments. Our approach (Fig. 1) learns wide baseline relative pose estimation and accurately predicts poses with few reference renderings. + +In our experiments, we evaluate our method using publicly available datasets, including Structured3D [46] and Zillow Indoor [10], encompassing both photorealistic synthetic data and real-world data. By sampling perspective query images from equirectangular panoramas, we efficiently augment the dataset, generating an almost infinite amount of training data. Compared to the state of the art, our method excels in estimating all degrees of freedom of the camera pose, enabling precise localization even in uncontrolled recording conditions. To summarise, we make the following contributions. + +1. We introduce a model-based 6D camera pose estimation system for unseen indoor environments without the need for scene-specific training. +2. We present a novel perspective-to-panoramic image matching concept with high retrieval accuracy even under wide camera baselines. +3. Demonstrating superior localization accuracy compared to the state of the art, our method estimates more degrees of freedom concurrently. + +# 2 Related Work + +Indoor Camera Localization. Indoor camera localization is a longstanding challenge in the computer vision community and in recent years, with the advance of neural networks, several deep learning based approaches have been proposed. Typically, these methods rely on prior scene acquisition to built a detailed scene representation and estimate a relation to a query image. Some methods use a 3D model generated via Structure-from-Motion [5] from registered camera images of the indoor environment and calculate the pose from 2D-3D correspondences, which might be either sparse [27,43] or dense [38]. Alternatively, encoding the entire scene within a neural network and performing regression to determine the absolute pose is a frequently used strategy [31]. + +By leveraging synthetic renderings from a detailed 3D BIM model, Acharya et al. eliminated the need for real-world reference data. Their approaches achieve real-time indoor localization [1] and an accuracy of approximately 1 meter [4] but are limited to the building of their origin, lacking generalizability to other structures. Liu et al. [26] pioneered floor plan-based image localization, predicting the facing wall within a cuboidal layout without estimating a full 3D pose. Howard et al. achieved 2D panorama localization in unseen scenes by aligning inferred room layouts with panoramic reference renderings [17] or encoded floor plans [16]. + +Several methods estimate the 2D position and 1D rotation of a perspective camera relative to a floor plan. PF-net [19] uses a recurrent neural network optimized for a differentiable particle filter, designed for sequential updating but suboptimal for single images [30]. F $^3$ Loc [9] improves sequential 2D localization + +against a floorplan by integrating multiview geometry cues. LASER [30] localizes the camera with respect to a semantic floor plan, and considers both perspective and panoramic queries. The rendering code book scheme efficiently encodes floor plan information but is limited to 2D data, which constrains the network's ability to estimate the remaining degrees of freedom. + +To sum up, existing methods either depend on accurate scene specific data for a detailed 3D representation or do not estimate the 6D camera pose. In contrast, our method neither depends on real-world reference images, a detailed 3D model, or scene specific training but is still able to estimate the 6D perspective camera pose by introducing a new cross-domain panoramic-to-perspective image registration approach between RGB images and semantic panoramas. + +Relative Pose Regression. Relative pose regression (RPR) aims at estimating the relative motion between a query image and reference images with known poses. These frameworks, relying on image comparisons, often enhance generalizability. Typically, they directly utilize localized examples at inference time [12]. + +Retrieval-based methods like NetVLAD [7] and DenseVLAD [40] start by identifying the most similar images in a database, providing a starting point for RPR. PixLoc [35] predicts multiscale deep features for direct alignment of reference and query image with a CNN and uses geometric optimization for calculating the pose. Turkoglu et al. [41] employ a GNN to re-localize cameras with respect to multiple frames. A graph links the query image to training counterparts, and a GNN refines representations for a consistent camera pose estimate. Our method shares similarities with [41] but goes further by addressing the domain gap between simplified synthetic and real data, and by incorporating multimodal matching between panoramic reference and perspective query images. + +Perspective-to-Panorama Matching. Matching between perspective query and panoramic reference images is a crucial aspect of localization, yet it has received limited attention in the literature. Most systems avoid multimodal matching and instead either match perspective crops to perspective queries [7,40] or panoramic to panoramic images [45]. + +While the past discussed geometric fundamentals for integrating catadioptric and perspective camera systems into hybrid stereo setups, including concepts like epipolar constraints [37] and calibration [14], these ideas have not yet found their way into contemporary AI-based localization systems. In the domain of CNNs and panoramic image processing, distortion-aware convolutional filters were proposed by Tateno et al. [39] and further refined by Fernandez et al. [13], particularly for indoor equirectangular images. Unlike traditional convolutions, these distortion-aware filters target the differences between perspective and panoramic images, making them particularly valuable for cross-modal matching. + +Recently, perspective-to-panoramic matching advanced, due to sliding window approaches in feature space [32, 36]. It improved visual recognition compared to methods that simply sampled perspective reference images from panoramas [44]. Orhan et al. [33] explored leveraging semantic segmentation similarity to address appearance variations between reference and query images. Our approach + +![](images/8e197c49d9fe4be59ae1989312daf454efebb43d8af1543d8c47f784d72bc9c5.jpg) +Fig. 2: Network architecture overview. The information from the query branch is fed into the panorama branch by depth-wise correlation. Three task heads predict the corresponding viewpoint, one task head predicts the relative pose offset. + +enhances perspective-to-panorama matching using virtual scene renderings as panoramic references. Inspired by multimodal zero-shot retrieval [18] and object detection, particularly Mercier et al.'s work [29], it utilizes the panorama as a global context for query matching. + +# 3 Method + +SPVLoc conducts 6D indoor localization of a 2D RGB query image within a simple semantic textureless 3D scene model. The model adheres to the Structured3D annotation format [46], containing walls, doors, windows, ceiling, and floor as planes. It can be derived from straightforward from building models [46], or from floor plans using contemporary AI tools [8, 28, 34]. SPVLoc first estimates the image's viewpoint within rendered panoramas from the semantic scene model using cross-domain image-to-panorama matching (Sec. 3.1). It then ranks these predictions based on classification accuracy (Sec. 3.4) and performs relative 6D pose regression with respect to the best-matching reference panorama (Sec. 3.2). + +# 3.1 Semantic Panoramic Viewport Matching + +Problem Definition. Our objective is to precisely estimate the 6D camera pose denoted as $P$ for a given query RGB image $I_{q}$ relative to a semantic 3D model $S$ . To achieve this, we reformulate the 6D indoor localization of $I_{q}$ within $S$ as a cross-domain image-to-panorama matching problem. The model is represented by a set of $N$ equirectangular reference layouts $L = \{L_{i} \mid i = 1, \dots, N\}$ consisting of semantic labels in a single channel. Each rendering comprehensively captures the entire environment around a specific vantage point. The viewpoint $V$ is a distinct cutout section from a panorama, defining the exact region a camera positioned nearby can see. The perspective distortion of $V_{P_{i}}$ contains information about + +the relative pose offset $P_{i}$ between $L_{i}$ and $I_{q}$ . Scoring and ranking viewpoints helps in choosing the optimal panorama $L^{\star}$ . The absolute pose $P$ is then formed by concatenating $P^{\star}$ with $P_{i}$ . Identifying the viewpoint and the corresponding pose offset poses several challenges. First, with increasing distance between the two cameras the perspective distortion between image and panorama increases. Second, the matching involves a severe domain gap as a photographic RGB image is matched against a rendered semantic panorama. + +Viewport Prediction. Semantic panorama reference renderings are created utilizing a ray-tracing-based renderer [24]. The viewport mask $V_{P}$ is a binary image that signifies the visibility of points within the equirectangular image from the view of the perspective camera. The calculation involves projecting all 3D points from the panorama in the perspective view. The projection is performed based on the known underlying layout during the training phase. The mask $V_{P}$ is determined by identifying pixels projecting into the perspective image, taking into account the 3D camera pose, and excluding those covered by occlusion. + +Additionally, the 2D bounding box $v_{bb}$ is the minimal axis-aligned frame around $V_P$ . This bounding box is circularly defined within the $360^\circ$ panorama, allowing the overlap on the right image side to re-enter on the left side of the panorama. Each predicted bounding box is associated with a confidence score $c_{bb}$ , which quantifies the probability of the bounding box being a valid match. The network's objective is to predict $V_P$ , $v_{bb}$ and $c_{bb}$ for each pair of input semantic panorama and query RGB image. The optimal viewpoint corresponds to the predicted bounding box with the highest overall confidence score, $c_{bb}$ , when presented with a set of panorama images. + +Network Architecture. Our matching is inspired by the object detection strategy of DTOID [29]. Figure 2 illustrates the network architecture. Initially, $I_{q}$ is encoded with an EfficientNet-S backbone. The features obtained after the fourth and the last downsampling layer are preserved. Subsequently, both outputs are rescaled to a tensor of spatial size of $7 \times 7$ and concatenated to a tensor $F$ . + +$L$ is encoded using a DenseNet-121 backbone, extending up to the fifth downsampling block to extract features $R$ . The information extracted from the input image guides the panorama branch to estimate the viewport mask and bounding box. This guidance is achieved through the application of a depth-wise correlation technique (denoted as $\star_{\mathbf{d}}$ ) [6,29]. $F$ is processed through two consecutive convolution blocks, each incorporating convolution, batch normalization, and ELU activation, resulting in a set of $3 \times 3$ filters $\hat{F}$ used during the depth-wise correlation $R \star_{\mathbf{d}} \hat{F}$ . + +Furthermore, $F$ undergoes compression via global max pooling and is both multiplied and subtracted from $R$ , resulting in three feature tensors with injected image information. The features are individually processed with an additional convolution block and then concatenated. The concatenated features further undergo a final convolution block, leading to the correlated features $R^{\star}$ . + +A decoder estimates the viewpoint of the perspective camera projected in the layout of the omnidirectional panorama. The viewpoint in the panorama is + +estimated as a bounding box with a BBox Class. and BBox Reg. head, similar to RetinaNet [25]. The BBox Class. head's output is optimized using focal loss $\mathcal{L}_1^{VP}$ , while the BBox Reg. head is optimized with a smooth $l_{1}$ loss $\mathcal{L}_2^{VP}$ . The BBox Class. provides a score for a given estimation, ensuring that the bounding box with the highest probability corresponds to the best estimate. Additionally, $V_{P}$ representing the exact viewpoint of the perspective camera is estimated with the aid of another convolutional decoder head and optimized with binary cross entropy loss $\mathcal{L}_3^{VP}$ . + +Perspective Supervision. The encoded image features $I_{E}$ are connected to a convolutional decoder, outputting the underlying semantic and normal map to ensure that features closely corresponding to the image content are extracted. The semantics are learned with cross entropy loss $\mathcal{L}_4^{VP}$ and the normals are learned with cosine loss $\mathcal{L}_5^{VP}$ . + +# 3.2 Feature-Correlation-based Pose Regression + +The correlated features, denoted as $R^{\star}$ , encode information about the viewpoint of $I_{q}$ within a $L_{i}$ , but should also lead to relative pose offset $P_{i}$ between them. Given that $L$ is synthetically generated, the precise position of $L_{i}$ is known. Therefore, determining the relative position of $I_{q}$ in relation to $L_{i}$ is equivalent to establishing its absolute position. All rendered panoramas exhibit identity orientation, due to horizontal alignment and a northward orientation. Consequently, the rotation delta directly corresponds to the absolute rotation of the camera. + +An additional Pose head processes $R^{\star}$ to estimate the relative pose offset of the camera concerning a panorama rendering (Fig. 3). We first utilize three convolution blocks, whereby the first two are followed by a max-pooling layer, to compress the features. The flattened features form an 1D embedding $r^{\star} \in \mathbb{R}^{C}$ . Here, $C = 640$ for a panorama of dimensions $256 \times 128$ . Optionally, to handle different camera field of views during training and testing, $r^{\star}$ is normalized and multiplied by the horizontal field of view angle in radians. The result is fed into a four-layer MLP, producing the 3D offset to the query image location $\hat{t} i \in \mathbb{R}^{3}$ and a rotation given as the logarithm of the unit quaternion $\hat{r} i = \log (\hat{q})_{ij} \in \mathbb{R}^{3}$ [41]. + +The MLP is evaluated using two loss functions. $\mathcal{L}_1^P$ assesses the precision of the relative translation, and $\mathcal{L}_2^P$ evaluates the accuracy of the final rotation estimation. Both losses are computed using the $l_{1}$ distance. + +# 3.3 Optimization + +The training loss is the sum of the two losses of the MLP and the five losses of the viewpoint estimation. With seven losses, our multi-task learning has to be balanced by using a different weight for each loss. As in [20, 21], we use an exponential mapping and weight multiple loss functions by considering the homoscedastic uncertainty of each task and learned weighting factors. The final loss is given as: + +$$ +\mathcal {L} = \sum_ {i = 1} ^ {2} \mathcal {L} _ {i} ^ {P} e ^ {- \beta_ {i}} + \beta_ {i} + \sum_ {i = 1} ^ {5} \mathcal {L} _ {i} ^ {V P} e ^ {- \gamma_ {i}} + \gamma_ {i} \tag {1} +$$ + +![](images/a3821fff45ec52102ba1cadb4f1ced6a630caebc8364cc0d4b3fa3174c85a14b.jpg) +Fig. 3: The Pose head utilizes convolutions, an MLP, and a FoV-based side input. + +# 3.4 Inference + +During inference, panorama positions are determined from a fixed 2D grid superimposed on the floor plan, with virtual cameras set at a constant height denoted as $h_{pano}$ , and the grid's dimensions specified as $xy_{pano}$ . Virtual cameras are exclusively sampled above floors to ensure valid reference images inside rooms. Setting $xy_{pano}$ as large as possible is advantageous in order to minimize the number of references, given that feature correlation and bounding box regression are executed for every panorama. + +For pose estimation, we select reference positions with the top- $n$ highest bounding box classification scores. The absolute pose is determined by the result of the Pose head, where the selected reference position is added to the translation. To eliminate the necessity for rendering during inference, we propose pre-generating reference panoramas for a floor plan. Additionally, the encoding of these panoramas can be precomputed. This ensures that only the encoded features $R$ need to be stored, significantly streamlining the inference process. + +Refinement. Without altering the network architecture or training process, SPVLoc can enhance an estimated pose by rendering a new reference panorama at the estimated position and re-executing the relative pose regression. The smaller step to the actual target position in this second iteration simplifies the problem, allowing for increased precision in the refined pose estimation. + +# 4 Experiments + +To validate our method we conduct several experiments and make use of two publicly available and frequently used datasets. + +Structured3D (S3D) [46] consists of 3500 near-photorealistic professionally modeled indoor environments, each annotated with ground truth 3D structure information. It encompasses 21835 panoramic images, each provided in both furnished and unfurnished configurations. The authors define a unified representation that describes scenes using geometric primitives such as planes, lines, junction points, and semantic annotations including floor, ceiling, wall, door, and window classes. + +![](images/0f30960556da9c1e4a2ddd24240ce607f2f35d2cf3a963cead6b61282d4b06f9.jpg) +Fig. 4: ZinD data preparation. Annotations generate 3D reference models (left), while resampled panoramas create perspective train and test images (right). + +Zillow Indoor (ZInD) [10] includes 67,448 panorama images from 1,575 unfurnished residential homes, globally aligned and registered to floor plans. The 3D annotations of panoramas within a scene are adequate for transforming them into the S3D representation format. Before training, we transform all data to that format. Figure 4 shows an example reference model generated from ZInD annotations. Unlike S3D, ZInD includes room openings such as corridors connected to rooms without doors, enhancing structural variation in semantic renderings. + +For both datasets, we employ the official train-test scene split. To enhance control over the field of view and angle variation, we sample perspective test images from the RGB panoramas. They are registered in the 3D model, so the semantics are known for each perspective image. We compile test-sets suitable for 6D pose estimation and structured evaluation. The test sets consist of square perspective images with a viewing angle of 60, 90, and $120^{\circ}$ , respectively, along with variations in roll and pitch, and random yaw angles for ZinD and S3D. To ensure meaningful views, we apply two criteria on the sampled yaw. First, the images contain a minimum of three semantic classes, like wall-floor-ceiling or wall-window-floor. Second, the camera faces inside the room rather than towards the next wall. The criteria are applied to the $60^{\circ}$ images (4420 for ZiND and 1133 for S3D). Subsequent variations in opening angles and roll-pitch configurations are generated from this starting point. + +# 4.1 Details + +For evaluation, we report the fraction of frames localized correctly within a threshold $\tau$ . We also list the median rotation and translation error, as well as the fraction of frames localized within $\tau$ and an additional angle threshold $v$ . In the retrieval process, the top-3 best matches are selected, and we report recall metrics for both top-1 and the best among the top-3. + +Training. For the state of the art comparison, we train a model with a variable field of view, while other experiments use a model trained with a fixed opening angle of $90^{\circ}$ . The model with variable viewing angles randomly samples views between 45 and $135^{\circ}$ during training. We consistently render $s$ randomly positioned panoramas around the query within a radius of $\pm r_1$ in $xy$ direction and + +Table 1: Comparison with state-of-the-art baseline on ZInD and S3D: Median translation $(\mathrm{terr}_{\mathrm{xy}})$ in cm, rotation error (rerrayw) in deg. for instances localized under 1m. Reporting recall at various translation accuracy levels, recall for inliers (<1m and <30°), and top-3 recall at 1m. Ours† refers to our model tested with a slightly different testset, incorporating pitch and roll rotation variation ( $\pm 10^{\circ}$ ). + +
ZInDStructured3D (Furnishing-Level : Full)
QueryMethod<1m med terr (cm)<1m med rerr (deg)10cm (%)50cm (%)1m (%)1m & 30° (%)top-3 1m (%)<1m med terr (cm)<1m med rerr (deg)10cm (%)50cm (%)1m (%)1m & 30° (%)top-3 1m(%)
Perspective 60° FoVLASER [30]33.531.034.0547.2464.5263.5378.6417.600.8416.4663.2268.3967.3486.08
Ours17.161.5021.0481.3386.7686.0994.1212.251.2134.9487.1389.4988.7996.15
\( Ours^† \)17.771.6319.5981.1387.1086.4394.8613.311.2430.8284.6888.1887.4895.80
Perspective 90° FoVLASER [30]26.280.778.6967.0180.9080.2589.7314.600.6723.4772.7776.8076.1891.94
Ours13.631.3430.5785.5990.8490.3496.369.170.8350.0989.6792.1291.8697.37
\( Ours^† \)14.251.3729.5985.9091.3390.9596.9710.090.9145.6289.3292.1291.9496.85
Perspective 120° FoVLASER [30]23.100.7812.6975.0985.6385.2092.3812.900.7629.0777.7682.1481.3594.13
Ours14.781.4529.0585.1491.7291.2796.9211.690.8739.4988.0092.1291.9497.37
\( Ours^† \)15.131.5328.1286.1392.7192.3597.2411.770.8838.1889.6793.0892.9197.55
+ +$\pm r_2$ in $z$ (upward) direction. For a single query, the viewpoint and relative pose offset is therefore calculated against multiple references. In our experimental setup, we set $s$ to 4, $r_1$ to 1.4 meters, and $r_2$ to 30 centimeters. Additionally, one random negative example is generated outside $r_1$ , likely in a different room, with the target to enhance the network's ability to distinguish subtle room differences. Images are sampled with random yaw and $\pm 10^{\circ}$ random pitch and roll angle. The batch size is set to 40, comprising 40 query images and 200 panoramas, and the training is performed on a single NVIDIA A100 GPU. Query images with fewer than three semantic classes are disregarded during the loss calculation process. Training runs for approximately 42,000 steps with an initial learning rate of $2.5 \times 10^{-4}$ , halved two times during training. + +Testing. During testing, panoramas are sampled on a $1.2 \times 1.2$ m grid, requiring significantly fewer samples than other rendering-based approaches such as LaLaLoc [17], which uses a $0.5 \times 0.5$ m grid. Additionally, for a scene, all panoramic views can be encoded initially, and during inference, only the query image needs to be encoded. If not otherwise noted the studies are conducted on ZInD with fixed 90 degree FoV, reporting the 3D rotation and translation error. + +# 4.2 Comparison with the State of the Art + +To our knowledge, SPVLoc is the first method to perform 6D localization in unseen scenes with respect to semantic 3D models. The state of the-art method for scene-independent global localization, LASER [30], takes perspective views as input and matches them against a semantic floor plan. Hence, it estimates only two position and one rotation degree of freedom. Nonetheless, it serves as a strong baseline to assess the localization accuracy of our method. + +Table 1 shows the results on ZInD and S3D. Given that we compare against a 2D method, our evaluation also includes recall based on 2D distance and angular error derived from the yaw angle. The yaw angle is extracted from the estimated rotation matrix. Notably, SPVLoc without refinement outperforms LASER in + +Table 2: Ablation study. SPVLoc variations compared on ZInD with $90^{\circ}$ FoV. In contrast to Tab. 1, terr denotes 3D position errors, and rerr denotes 3D rotation errors. + +
Abl. TypeMethod<1m med terr (cm)<1m med rerr (deg)10cm (%)50cm (%)1m (%)1m & 30° (%)top-3 1m (%)
(Ours)base14.312.0526.8886.9291.9091.6196.56
+ refine12.321.9633.6990.8492.7691.74-
Loss- query dec.15.702.2321.9785.5090.7990.4396.45
- view dec.15.192.2123.6085.3690.4589.8296.49
- view/query dec.16.222.1921.5684.5089.5989.1296.36
Training- neg. samples15.062.2822.6076.6582.1981.7492.99
Arch.Δ Resnet18 (query)15.562.5620.8886.5690.7290.3296.56
Δ Equiconv (pano)14.982.1325.0787.4491.7091.3196.97
Input+ normal14.682.2825.0985.9390.8490.2396.47
+ depth14.062.1027.6088.0592.1591.7996.74
+ +terms of localization accuracy, exhibiting higher recall across all categories, an increased inlier rate, and significantly reduced median translation error. In terms of 2D rotation, our method is slightly less precise. This is because LASER estimates yaw directly while we estimate a full rotation matrix. Estimating camera height and roll and pitch angle is an advantage of our method. In each last row $(Ours^{\dagger})$ , we evaluate our method on an alternate test set with roll and pitch variations $(\pm 10^{\circ})$ . The results suggest that these variations do not substantially impact the accuracy of our method for 2D localization. + +Regarding further methods, LASER has showcased a significant performance improvement over PF-net [19] and non-AI-based Monte-Carlo Localization (MCL) [11]. PF-net, originally designed for sequential time updating, does not perform well in single-query localization scenarios. MCL neglects semantic annotations, uses LIDAR instead of images, which also has been shown to drastically reduce performance [30]. Our substantial improvement over LASER therefore also illustrates a performance gain over aforementioned methods. Another experiment using the pre-rendered perspective images of S3D, including a comparison to $\mathrm{F}^3\mathrm{Loc}$ [9], is part of the supplementary material. + +# 4.3 Performance Studies + +Ablation Study. We commence our experimentation with the full network (base) configuration. Subsequently, we modify the network by removing specific components (Tab. 2). Initially, we test the influence of perspective supervision (-query dec.) and the view segment task head (-view dec.) by excluding them during training. Their exclusion reduces network performance, likely because the network's ability to understand the image content and establish correspondences is diminished. Removing both (-view/query dec.) further reduces accuracy. + +We then exclude the negative samples from different rooms during training (-neg. samples). This significantly reduces the amount of correctly localized images. It supports the hypothesis that teaching the network to recognize spaces where there is no match helps to improve the accuracy of predictions. + +To modify the architecture, the image encoder EfficientNet-S is replaced with a smaller ResNet-18 [15] (Δ Resnet18), resulting in a performance drop. This + +![](images/1e784445864227badce7d8b33244ad5e0d2e554920c1a55c0a59cedd36937147.jpg) +Fig. 5: Qualitative localization results. Top to bottom - query, rendering with top-1 estimated pose, panorama with estimated viewport, map. Green box: success for top-1 match. Yellow box: success for top-2 match. Red box: failure case. + +suggests an increased demand for network capacity to analyze the images. In the panorama encoder, replacing all convolutional layers with Equiconv [13] (ΔEquiconv), a specialized convolution for panorama images that remaps query locations from Euclidean to spherical space, does not result in a performance gain. This suggests that the capacity of the small DenseNet-121 is sufficient for analyzing the low-resolution semantic inputs. + +Lastly, we incorporate additional input modalities for the panoramas into the panorama encoder. While the addition of world coordinate normals (+normals) did not improve the result, including an extra depth map slightly enhanced the outcome (+depth). This suggests that there are cases where additional information is encoded in the depth that cannot be inferred from semantics alone. + +Qualitative Results. Figure 5 shows selected SPVLoc outcomes. The top row displays camera images, and the second row shows renderings from the semantic 3D model with estimated top-1 poses. All green boxed images are correctly localized for the top-1 match. In the yellow box, the top-1 match fails, but the top-2 match is accurate. The red box example cannot be localized. Across all examples, a pose with high visual similarity is found. Ambiguous examples arise from similar layouts occurring in multiple rooms. Failure cases are attributed to either images lacking sufficient structural information for matching or outliers in the data, such as those taken on balconies or in dark cellars. The third row displays the chosen reference panorama, the selected viewpoint bounding box corresponding to the camera image, and a semi-transparent overlay representing the estimated viewpoint mask. Additional detailed visualizations are part of the supplementary material. + +Influence of Sampling Grid Sizes and Refinement. Reference positions result from overlaying a grid onto the floorplan, and the grid density significantly + +Table 3: Grid size vs. accuracy: Deviation of $10\mathrm{cm}$ and $1\mathrm{m}$ recall from (boxed) baseline. + +
Sample dist.Samples per \( m^2 \)Adv.10cm (%)10cm + refine1m (%)1m + refine
0.72.07+1.00+7.49+1.86+2.38
0.91.25+0.95+6.99+1.79+2.08
1.20.70\( \boxed{26.88} \)+6.81\( \boxed{91.90} \)+0.86
1.20.48✓+0.57+6.95-0.63+0.61
1.50.45-1.00+6.63-2.47-1.02
1.50.31✓-1.04+6.47-2.67-0.38
2.00.26-4.19+3.60-9.05-5.79
2.00.20✓-1.76+5.81-3.91-1.18
+ +![](images/91b10ff04ac2a7894072d24db67c926cf32755479e03db8746e94c5cda1c653d.jpg) +Fig. 6: Reference positions for global grid of size $0.7\mathrm{m}$ , $1.5\mathrm{m}$ and local grid of size $1.5\mathrm{m}$ (from left to right). + +![](images/af39788a9556d7dbfe55ad0eecd8b086b179e09c3c8c42015b9e761cf7d1158c.jpg) + +![](images/5ec39db21124c7bbea1d4ec2e108d18a8e26f0b23d85a2075210fdf383184c46.jpg) + +affects SPVLoc's efficiency and inference time. Minimizing references reduces rendering effort and requires the model to estimate larger offsets. E.g. to increase the grid from $0.7m$ to $2m$ reduces the number of samples per $m^2$ by a factor of 10. We vary the grid size to analyze its impact on performance. We also experiment with a local grid per room substituting the global grid (Adv.) reducing the risk of missing rooms entirely. The local grid is centered on the room outline and additionally places one camera in the middle of each opening. Figure 6 shows different grids for an example plan. Table 3 illustrates the relationships among grid size, global and local sampling, and refinement. Refinement notably improves $10\mathrm{cm}$ recall across configurations. For the largest local grid with refinement, the $1\mathrm{m}$ recall only decreases by $1.18\%$ compared to the baseline configuration. In summary, a trade-off between refinement and dense/sparse scene sampling exists, yet our method consistently delivers strong results with very sparse sampling, making it suitable for large scenes. + +Influence of Focal Length. Table 4 depicts the performance difference between a variant trained with a known camera focal length and a focal length-independent network trained with varying focal lengths (ranging from 45 to 135 degrees). The fixed focal length network performs slightly better with matching images but loses precision with test images of different focal lengths. In contrast, the focal length-independent network demonstrates the capability to handle such differences. We argue that the significantly enhanced flexibility of the latter justifies the minor performance degradation in practical, real-world scenarios. + +Influence of Camera Rotation Angles. The range of pitch and roll angles the network is able to handle is determined by the variations of angles seen during training. We train the network for ZInD using a angle variation of $\pm 22^{\circ}$ and test with images where the range of variation increases from $0^{\circ}$ to $20^{\circ}$ in steps of $\pm 5^{\circ}$ . Table 5 shows only a slight increase in median errors for rotation and translation as the rotation angles in test images increase. The chosen ranges represent a significant coverage and offer substantial improvements over existing methods. In indoor scenes, larger camera tilt often results in images dominated by the ceiling or floor, rendering them unsuitable for semantic matching. + +Table 4: The fixed FoV network excels at a specific angle but loses precision with changing FoV. The proposed network remains stable. + +
Input FoVTrain FoV<1m med terr (cm)<1m med rerr (deg)10cm (%)1m (%)
90°90°14.312.0526.8891.90
45 - 135°16.112.2021.8191.45
80°90°37.512.371.6790.14
45 - 135°16.752.1919.3090.50
+ +Table 5: Effect of increasing angle variation in test data. + +
MetricRoll/Pitch Variation
±0°±5°±10°±15°±20°
<1m med. terr14.4614.8814.8815.2115.91
<1m med. verr2.352.362.382.512.60
10cm (%)24.6423.9424.4623.6922.74
1m (%)91.4591.4791.9992.3892.60
+ +Table 6: Timing. Performance on ZInD using a NVIDIA Tesla V100, following the configurations outlined in Sec. 4.3, and compared against results from [30]. + +
MethodSampling Time (s)Query FpsQuery Module Time (ms)
PF-net [19]48.95 ± 38.955.06 ± 1.77Eff.Net V2 S4.4
LASER [30]0.97 ± 1.098.31 ± 0.64Correlation35.7
Ours1.66 ± 1.0628.63 ± 11.54Pose1.1
+ +Limitations. In large spaces with repeating room layouts, the approach's effectiveness may be constrained by the limited level of detail in the semantic reference model. However, enhancing the S3D format by adding more semantic classes and structural information, such as staircases or sanitary installations, could reduce ambiguities. + +Timing. We split the measured time into scene sampling/preparation time, including position sampling, panorama rendering and encoding, and inference/query time, including query image encoding, feature correlation with bounding box classification, and the Pose task head, but no refinement. Table 6 shows that our method exhibits a considerably lower query time than LASER [30] and PF-net [19]. This difference places our approach close to real-time for moderately sized scenes. Currently, scene sampling time is slower than [30], mainly due to rendering on the CPU without parallelization [24]. Parallelizing this process using GPU or CPU could significantly accelerate scene preparation. + +# 5 Conclusion + +This paper introduces a method for scene-independent model-based 6D localization for indoor scenes, involving a novel method for multimodal image matching (panorama to perspective, RGB to semantic). The matching and retrieval is efficient and scalable under sparse reference sampling. The localization accuracy and inference speed outperform state of the art methods and the inclusion of the 3D-model, reduces ambiguities to precisely estimate 6D poses. Future work involves combining localization and image analysis to enhance digital building models or exploring applications in augmented reality scenarios. + +Acknowledgements. This work is partially funded by the German Federal Ministry of Economic Affairs and Climate Action (BIMKIT, 01MK21001H) and the German Federal Ministry for Digital and Transport (EConoM, 19OI22009C). + +# References + +1. Acharya, D., Khoshelham, K., Winter, S.: BIM-PoseNet: Indoor camera localisation using a 3D indoor model and deep learning from synthetic images. ISPRS journal of photogrammetry and remote sensing 150, 245-258 (2019) +2. Acharya, D., Ramezani, M., Khoshelham, K., Winter, S.: Bim-tracker: A model-based visual tracking approach for indoor localisation using a 3d building model. ISPRS Journal of Photogrammetry and Remote Sensing 150, 157-171 (2019) +3. Acharya, D., Tatli, C.J., Khoshelham, K.: Synthetic-real image domain adaptation for indoor camera pose regression using a 3d model. ISPRS Journal of Photogrammetry and Remote Sensing 202, 405-421 (2023) +4. Acharya, D., Tennakoon, R., Muthu, S., Khoshelham, K., Hoseinnezhad, R., Bab-Hadiashar, A.: Single-image localisation using 3D models: Combining hierarchical edge maps and semantic segmentation for domain adaptation. Automation in Construction 136, 104-152 (2022) +5. Agarwal, S., Furukawa, Y., Snavely, N., Simon, I., Curless, B., Seitz, S.M., Szeliski, R.: Building rome in a day. Communications of the ACM 54(10), 105-112 (2011) +6. Ammirato, P., Fu, C.Y., Shvets, M., Kosecka, J., Berg, A.C.: Target driven instance detection. arXiv preprint arXiv:1803.04610 (2018) +7. Arandjelovic, R., Gronat, P., Torii, A., Pajdla, T., Sivic, J.: Netvlad: Cnn architecture for weakly supervised place recognition. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2016) +8. Cambeiro, A.B., Trzeciakiewicz, M., Hilsmann, A., Eisert, P.: Automatic reconstruction of semantic 3d models from 2d floor plans. In: Proc. Int. Conf. on Machine Vision Applications (MVA). Hamamatsu, Japan (July 2023) +9. Chen, C., Wang, R., Vogel, C., Pollefeys, M.: F3Loc: Fusion and Filtering for Floorplan Localization. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2024) +0. Cruz, S., Hutchcroft, W., Li, Y., Khosravan, N., Boyadzhiev, I., Kang, S.B.: Zillow indoor dataset: Annotated floor plans with 360deg panoramas and 3d room layouts. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2133-2143 (2021) +1. Dellaert, F., Fox, D., Burgard, W., Thrun, S.: Monte carlo localization for mobile robots. In: Proceedings IEEE International Conference on Robotics and Automation. vol. 2, pp. 1322-1328 (1999) +2. Ding, M., Wang, Z., Sun, J., Shi, J., Luo, P.: CamNet: Coarse-to-fine retrieval for camera re-localization. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2871-2880 (2019) +3. Fernandez-Labrador, C., Facil, J.M., Perez-Yus, A., Demonceaux, C., Civera, J., Guerrero, J.J.: Corners for layout: End-to-end layout recovery from 360 images. IEEE Robotics and Automation Letters 5(2), 1255-1262 (2020) +4. He, B., Chen, Z., Li, Y.: Calibration method for a central catadioptric-perspective camera system. JOSA A 29(11), 2514-2524 (2012) +5. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 770-778 (2016) + +16. Howard-Jenkins, H., Prisacariu, V.A.: LaLaLoc++: Global Floor Plan Comprehension for Layout Localisation in Unvisited Environments. In: Proc. European Conference on Computer Vision (ECCV). pp. 693-709. Springer (2022) +17. Howard-Jenkins, H., Ruiz-Sarmiento, J.R., Prisacariu, V.A.: LaLaLoc: Latent Layout Localisation in Dynamic, Unvisited Environments. In: Proc. IEEE/CVF International Conference on Computer Vision (ICCV) (2021) +18. Janik, M., Gard, N., Hilsmann, A., Eisert, P.: Zero in on Shape: A Generic 2D-3D Instance Similarity Metric Learned from Synthetic Data. In: Proc. IEEE International Conference on Image Processing (ICIP). pp. 2638-2642. IEEE (2021) +19. Karkus, P., Hsu, D., Lee, W.S.: Particle filter networks with application to visual localization. In: Conference on robot learning. pp. 169-178. PMLR (2018) +20. Kendall, A., Cipolla, R.: Geometric loss functions for camera pose regression with deep learning. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5974-5983 (2017) +21. Kendall, A., Gal, Y., Cipolla, R.: Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7482-7491 (2018) +22. Kendall, A., Grimes, M., Cipolla, R.: Posenet: A convolutional network for real-time 6-dof camera relocalization. In: Proc. IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2938-2946 (2015) +23. Kim, J., Jeong, J., Kim, Y.M.: Fully geometric panoramic localization. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 20827-20837 (2024) +24. Li, T.M., Aittala, M., Durand, F., Lehtinen, J.: Differentiable Monte Carlo Ray Tracing through Edge Sampling. ACM Trans. Graph. (Proc. SIGGRAPH Asia) 37(6), 222:1-222:11 (2018) +25. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proc. IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2980-2988 (2017) +26. Liu, C., Schwing, A., Kundu, K., Urtasun, R., Fidler, S.: Rent3d: Floor-plan priors for monocular layout estimation. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2015) +27. Liu, L., Li, H., Dai, Y.: Efficient global 2d-3d matching for camera localization in a large-scale 3d map. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2372-2381 (2017) +28. Lv, X., Zhao, S., Yu, X., Zhao, B.: Residential floor plan recognition and reconstruction. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16717-16726 (June 2021) +29. Mercier, J.P., Garon, M., Giguere, P., Lalonde, J.F.: Deep Template-Based Object Instance Detection. In: Proc. WACV. pp. 1507-1516 (January 2021) +30. Min, Z., Khosravan, N., Bessinger, Z., Narayana, M., Kang, S.B., Dunn, E., Boyadzhiev, I.: Laser: Latent space rendering for 2d visual localization. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11122-11131 (2022) +31. Naseer, T., Burgard, W.: Deep regression for monocular camera-based 6-dof global localization in outdoor environments. In: Proc. International Conference on Intelligent Robot Systems (IROS). pp. 1525-1530. IEEE (2017) +32. Orhan, S., Bastanlar, Y.: Efficient search in a panoramic image database for long-term visual localization. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1727-1734 (2021) + +33. Orhan, S., Guerrero, J.J., Bastanlar, Y.: Semantic Pose Verification for Outdoor Visual Localization With Self-Supervised Contrastive Learning. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops. pp. 3989-3998 (June 2022) +34. Park, S., Kim, H.: 3dplannet: generating 3d models from 2d floor plan images using ensemble methods. Electronics 10(22), 2729 (2021) +35. Sarlin, P.E., Unagar, A., Larsson, M., Germain, H., Toft, C., Larsson, V., Pollefeys, M., Lepetit, V., Hammarstrand, L., Kahl, F., et al.: Back to the feature: Learning robust camera localization from pixels to pose. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3247-3257 (2021) +36. Shi, Z., Shi, H., Yang, K., Yin, Z., Lin, Y., Wang, K.: PanoVPR: Towards Unified Perspective-to-Equirectangular Visual Place Recognition via Sliding Windows across the Panoramic View. In: 2023 IEEE 26th International Conference on Intelligent Transportation Systems (ITSC). pp. 1333-1340 (2023) +37. Sturm, P.: Mixing catadioptric and perspective cameras. In: Proceedings of the IEEE Workshop on Omnidirectional Vision 2002. Held in conjunction with ECCV'02. pp. 37-44. IEEE (2002) +38. Taira, H., Okutomi, M., Sattler, T., Cimpoi, M., Pollefeys, M., Sivic, J., Pajdla, T., Torii, A.: InLoc: Indoor visual localization with dense matching and view synthesis. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7199-7209 (2018) +39. Tateno, K., Navab, N., Tombari, F.: Distortion-aware convolutional filters for dense prediction in panoramic images. In: Proc. European Conference on Computer Vision (ECCV). pp. 707-722 (2018) +40. Torii, A., Arandjelovic, R., Sivic, J., Okutomi, M., Pajdla, T.: 24/7 place recognition by view synthesis. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1808-1817 (2015) +41. Türkoglu, M.Ö., Brachmann, E., Schindler, K., Brostow, G., Monszpart, A.: Visual Camera Re-Localization Using Graph Neural Networks and Relative Pose Supervision. In: Proc. International Conference on 3D Vision (3DV) (2021) +42. Wang, S., Fidler, S., Urtasun, R.: Lost shopping! monocular localization in large indoor spaces. In: Proc. IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2695-2703 (2015) +43. Xia, J., Gong, J.: Precise indoor localization with 3d facility scan data. Computer-Aided Civil and Infrastructure Engineering 37(10), 1243-1259 (2022) +44. Zamir, A.R., Shah, M.: Accurate image localization based on google maps street view. In: Proc. European Conference on Computer Vision (ECCV). pp. 255-268. Springer, Heraklion, Greece (Sep 2010) +45. Zhang, C., Budvytis, I., Liwicki, S., Cipolla, R.: Lifted semantic graph embedding for omnidirectional place recognition. In: Proc. International Conference on 3D Vision (3DV). pp. 1401-1410. IEEE (2021) +46. Zheng, J., Zhang, J., Li, J., Tang, R., Gao, S., Zhou, Z.: Structured3D: A large photo-realistic dataset for structured 3d modeling. In: Proc. European Conference on Computer Vision (ECCV). pp. 519-535. Springer, Glasgow, UK (Aug 2020) \ No newline at end of file diff --git a/2024/SPVLoc_ Semantic Panoramic Viewport Matching for 6D Camera Localization in Unseen Environments/images.zip b/2024/SPVLoc_ Semantic Panoramic Viewport Matching for 6D Camera Localization in Unseen Environments/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..cdc52ca63724d261a9a2a180440ee7c193dd9cf8 --- /dev/null +++ b/2024/SPVLoc_ Semantic Panoramic Viewport Matching for 6D Camera Localization in Unseen Environments/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c100a8594e6096be8e1bf6a82b613dc7849fb7a07f77cba0443d89e0b3b404dd +size 410456 diff --git a/2024/SPVLoc_ Semantic Panoramic Viewport Matching for 6D Camera Localization in Unseen Environments/layout.json b/2024/SPVLoc_ Semantic Panoramic Viewport Matching for 6D Camera Localization in Unseen Environments/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a14066845e012452ce72c9546d818f6fdcfff89d --- /dev/null +++ b/2024/SPVLoc_ Semantic Panoramic Viewport Matching for 6D Camera Localization in Unseen Environments/layout.json @@ -0,0 +1,8970 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 133, + 111, + 481, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 111, + 481, + 163 + ], + "spans": [ + { + "bbox": [ + 133, + 111, + 481, + 163 + ], + "type": "text", + "content": "SPVLoc: Semantic Panoramic Viewport Matching for 6D Camera Localization in Unseen Environments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 179, + 185, + 433, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 185, + 433, + 198 + ], + "spans": [ + { + "bbox": [ + 179, + 185, + 433, + 198 + ], + "type": "text", + "content": "Niklas Gard" + }, + { + "bbox": [ + 179, + 185, + 433, + 198 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 179, + 185, + 433, + 198 + ], + "type": "text", + "content": ", Anna Hilsmann" + }, + { + "bbox": [ + 179, + 185, + 433, + 198 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 179, + 185, + 433, + 198 + ], + "type": "text", + "content": ", and Peter Eisert" + }, + { + "bbox": [ + 179, + 185, + 433, + 198 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 180, + 207, + 431, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 180, + 207, + 431, + 219 + ], + "spans": [ + { + "bbox": [ + 180, + 207, + 431, + 219 + ], + "type": "text", + "content": "1 Fraunhofer Heinrich Hertz Institute, HHI, Berlin, Germany" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 235, + 220, + 376, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 220, + 376, + 229 + ], + "spans": [ + { + "bbox": [ + 235, + 220, + 376, + 229 + ], + "type": "text", + "content": "{first.last}@hhi.fraunhofer.de" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 219, + 230, + 394, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 230, + 394, + 241 + ], + "spans": [ + { + "bbox": [ + 219, + 230, + 394, + 241 + ], + "type": "text", + "content": "2 Humboldt University of Berlin, Germany" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 264, + 455, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 264, + 455, + 485 + ], + "spans": [ + { + "bbox": [ + 160, + 264, + 455, + 485 + ], + "type": "text", + "content": "Abstract. In this paper, we present SPVLoc, a global indoor localization method that accurately determines the six-dimensional (6D) camera pose of a query image and requires minimal scene-specific prior knowledge and no scene-specific training. Our approach employs a novel matching procedure to localize the perspective camera's viewpoint, given as an RGB image, within a set of panoramic semantic layout representations of the indoor environment. The panoramas are rendered from an textured 3D reference model, which only comprises approximate structural information about room shapes, along with door and window annotations. We demonstrate that a straightforward convolutional network structure can successfully achieve image-to-panorama and ultimately image-to-model matching. Through a viewpoint classification score, we rank reference panoramas and select the best match for the query image. Then, a 6D relative pose is estimated between the chosen panorama and query image. Our experiments demonstrate that this approach not only efficiently bridges the domain gap but also generalizes well to previously unseen scenes that are not part of the training data. Moreover, it achieves superior localization accuracy compared to the state of the art methods and also estimates more degrees of freedom of the camera pose. Our source code is publicly available at: https://fraunhoferhhi.github.io/spvloc." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 160, + 494, + 448, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 494, + 448, + 506 + ], + "spans": [ + { + "bbox": [ + 160, + 494, + 448, + 506 + ], + "type": "text", + "content": "Keywords: Localization " + }, + { + "bbox": [ + 160, + 494, + 448, + 506 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 494, + 448, + 506 + ], + "type": "text", + "content": " 6D pose estimation " + }, + { + "bbox": [ + 160, + 494, + 448, + 506 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 494, + 448, + 506 + ], + "type": "text", + "content": " Multimodal matching" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 523, + 230, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 230, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 230, + 536 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 641 + ], + "type": "text", + "content": "Indoor camera localization is a fundamental challenge of computer vision, aiming to precisely determine a camera's position and orientation within an indoor environment. Applications include navigation [38,42], augmented reality (AR) [2], building management, and building digitization [26]. Current methods often depend on extensive scene-specific data, acquired through manual creation of databases with localized images, depth information, or point clouds [22,38,43]. However, this limits their applicability, especially in situations where exploring the building is impractical or data acquisition is costly and time-consuming." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 642, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 642, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 642, + 482, + 665 + ], + "type": "text", + "content": "Recent methods incorporate synthetic data during training, but these localization models require retraining or fine-tuning for new scenes using detailed BIM" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 187, + 114, + 428, + 286 + ], + "blocks": [ + { + "bbox": [ + 187, + 114, + 428, + 286 + ], + "lines": [ + { + "bbox": [ + 187, + 114, + 428, + 286 + ], + "spans": [ + { + "bbox": [ + 187, + 114, + 428, + 286 + ], + "type": "image", + "image_path": "102ed9dd0baa77f7c8c51552017da18c2ccdcc84dea05456ec39e50272a1495d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 295, + 482, + 329 + ], + "lines": [ + { + "bbox": [ + 131, + 295, + 482, + 329 + ], + "spans": [ + { + "bbox": [ + 131, + 295, + 482, + 329 + ], + "type": "text", + "content": "Fig. 1: SPVLoc for 6D indoor localization. Our method calculates the indoor 6D camera pose by determining the image position and orientation relative to synthetic panoramas. The best panoramic match is found through semantic viewpoint matching." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 354, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 521 + ], + "type": "text", + "content": "(Building Information Modeling) models [1,3]. The adoption of the Structured3D format [46] streamlines and standardizes 3D structure annotation by simplifying complex models into their most fundamental geometric representations. Methods that match a query image to this simplified structural representation exhibit robustness in unexplored, dynamic environments [17,30]. However, existing methods [9,30] mainly focus on 2D perspective camera localization with respect to floor plans, missing two degrees of freedom for rotation and one degree of freedom for position. Also, some methods solely rely on panoramic queries [16,17,23]. These queries are typically captured under controlled conditions, involving a tripod-mounted camera with a precise horizontal alignment, and a fixed camera height [17]. However, in AR applications or scenarios where archived photos must be accurately registered, e.g. to annotate visible building details into a digital building model, idealized capture conditions cannot always be guaranteed. In such cases, the demand for precise 6D localization becomes imperative." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 522, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 483, + 666 + ], + "type": "text", + "content": "In this paper, we present SPVLoc, a novel method for indoor 6D camera localization in unseen environments. It represents a 3D scene with a set of synthetically generated omnidirectional panorama images depicting fundamental semantic labels such as door, window, wall, floor, and ceiling. The core idea involves correlating the deep embedding of a perspective RGB camera image with the embeddings of panoramas to estimate the camera viewpoint - the precise region visible to the camera within the panorama. Leveraging a RetinaNet-like architecture [25], the 2D-bounding box around the viewpoint is predicted and classified. Retrieval of the closest panorama is performed by selecting the panorama with the globally highest classification score. The exact 6D pose is then found via relative pose estimation with an MLP starting from the panorama's position. Our method offers several advantages. It relies solely on a minimalist" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "text", + "content": "N. Gard et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "content": "3D semantic building model [46], which could even be automatically extracted from simple 2D plans [8,28,34]. By bridging the domain gap between semantic representation and real images, it facilitates matching in unseen environments. Our approach (Fig. 1) learns wide baseline relative pose estimation and accurately predicts poses with few reference renderings." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 482, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 482, + 273 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 482, + 273 + ], + "type": "text", + "content": "In our experiments, we evaluate our method using publicly available datasets, including Structured3D [46] and Zillow Indoor [10], encompassing both photorealistic synthetic data and real-world data. By sampling perspective query images from equirectangular panoramas, we efficiently augment the dataset, generating an almost infinite amount of training data. Compared to the state of the art, our method excels in estimating all degrees of freedom of the camera pose, enabling precise localization even in uncontrolled recording conditions. To summarise, we make the following contributions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 135, + 283, + 479, + 354 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 136, + 283, + 479, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 283, + 479, + 305 + ], + "spans": [ + { + "bbox": [ + 136, + 283, + 479, + 305 + ], + "type": "text", + "content": "1. We introduce a model-based 6D camera pose estimation system for unseen indoor environments without the need for scene-specific training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 135, + 307, + 479, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 307, + 479, + 330 + ], + "spans": [ + { + "bbox": [ + 135, + 307, + 479, + 330 + ], + "type": "text", + "content": "2. We present a novel perspective-to-panoramic image matching concept with high retrieval accuracy even under wide camera baselines." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 135, + 331, + 479, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 331, + 479, + 354 + ], + "spans": [ + { + "bbox": [ + 135, + 331, + 479, + 354 + ], + "type": "text", + "content": "3. Demonstrating superior localization accuracy compared to the state of the art, our method estimates more degrees of freedom concurrently." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 375, + 237, + 388 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 375, + 237, + 388 + ], + "spans": [ + { + "bbox": [ + 132, + 375, + 237, + 388 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 402, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 521 + ], + "type": "text", + "content": "Indoor Camera Localization. Indoor camera localization is a longstanding challenge in the computer vision community and in recent years, with the advance of neural networks, several deep learning based approaches have been proposed. Typically, these methods rely on prior scene acquisition to built a detailed scene representation and estimate a relation to a query image. Some methods use a 3D model generated via Structure-from-Motion [5] from registered camera images of the indoor environment and calculate the pose from 2D-3D correspondences, which might be either sparse [27,43] or dense [38]. Alternatively, encoding the entire scene within a neural network and performing regression to determine the absolute pose is a frequently used strategy [31]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 521, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 617 + ], + "type": "text", + "content": "By leveraging synthetic renderings from a detailed 3D BIM model, Acharya et al. eliminated the need for real-world reference data. Their approaches achieve real-time indoor localization [1] and an accuracy of approximately 1 meter [4] but are limited to the building of their origin, lacking generalizability to other structures. Liu et al. [26] pioneered floor plan-based image localization, predicting the facing wall within a cuboidal layout without estimating a full 3D pose. Howard et al. achieved 2D panorama localization in unseen scenes by aligning inferred room layouts with panoramic reference renderings [17] or encoded floor plans [16]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Several methods estimate the 2D position and 1D rotation of a perspective camera relative to a floor plan. PF-net [19] uses a recurrent neural network optimized for a differentiable particle filter, designed for sequential updating but suboptimal for single images [30]. F" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Loc [9] improves sequential 2D localization" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 161, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 161, + 91, + 447, + 103 + ], + "type": "text", + "content": "SPVLoc: Semantic Panoramic Viewport Matching for 6D Localization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "content": "against a floorplan by integrating multiview geometry cues. LASER [30] localizes the camera with respect to a semantic floor plan, and considers both perspective and panoramic queries. The rendering code book scheme efficiently encodes floor plan information but is limited to 2D data, which constrains the network's ability to estimate the remaining degrees of freedom." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 482, + 248 + ], + "type": "text", + "content": "To sum up, existing methods either depend on accurate scene specific data for a detailed 3D representation or do not estimate the 6D camera pose. In contrast, our method neither depends on real-world reference images, a detailed 3D model, or scene specific training but is still able to estimate the 6D perspective camera pose by introducing a new cross-domain panoramic-to-perspective image registration approach between RGB images and semantic panoramas." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 259, + 482, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 482, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 482, + 307 + ], + "type": "text", + "content": "Relative Pose Regression. Relative pose regression (RPR) aims at estimating the relative motion between a query image and reference images with known poses. These frameworks, relying on image comparisons, often enhance generalizability. Typically, they directly utilize localized examples at inference time [12]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 308, + 482, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 482, + 428 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 482, + 428 + ], + "type": "text", + "content": "Retrieval-based methods like NetVLAD [7] and DenseVLAD [40] start by identifying the most similar images in a database, providing a starting point for RPR. PixLoc [35] predicts multiscale deep features for direct alignment of reference and query image with a CNN and uses geometric optimization for calculating the pose. Turkoglu et al. [41] employ a GNN to re-localize cameras with respect to multiple frames. A graph links the query image to training counterparts, and a GNN refines representations for a consistent camera pose estimate. Our method shares similarities with [41] but goes further by addressing the domain gap between simplified synthetic and real data, and by incorporating multimodal matching between panoramic reference and perspective query images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 438, + 482, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 482, + 498 + ], + "type": "text", + "content": "Perspective-to-Panorama Matching. Matching between perspective query and panoramic reference images is a crucial aspect of localization, yet it has received limited attention in the literature. Most systems avoid multimodal matching and instead either match perspective crops to perspective queries [7,40] or panoramic to panoramic images [45]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "type": "text", + "content": "While the past discussed geometric fundamentals for integrating catadioptric and perspective camera systems into hybrid stereo setups, including concepts like epipolar constraints [37] and calibration [14], these ideas have not yet found their way into contemporary AI-based localization systems. In the domain of CNNs and panoramic image processing, distortion-aware convolutional filters were proposed by Tateno et al. [39] and further refined by Fernandez et al. [13], particularly for indoor equirectangular images. Unlike traditional convolutions, these distortion-aware filters target the differences between perspective and panoramic images, making them particularly valuable for cross-modal matching." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "content": "Recently, perspective-to-panoramic matching advanced, due to sliding window approaches in feature space [32, 36]. It improved visual recognition compared to methods that simply sampled perspective reference images from panoramas [44]. Orhan et al. [33] explored leveraging semantic segmentation similarity to address appearance variations between reference and query images. Our approach" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "text", + "content": "N. Gard et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 255 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 255 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 255 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 255 + ], + "type": "image", + "image_path": "8e197c49d9fe4be59ae1989312daf454efebb43d8af1543d8c47f784d72bc9c5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 262, + 482, + 297 + ], + "lines": [ + { + "bbox": [ + 130, + 262, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 482, + 297 + ], + "type": "text", + "content": "Fig. 2: Network architecture overview. The information from the query branch is fed into the panorama branch by depth-wise correlation. Three task heads predict the corresponding viewpoint, one task head predicts the relative pose offset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 319, + 482, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 482, + 369 + ], + "type": "text", + "content": "enhances perspective-to-panorama matching using virtual scene renderings as panoramic references. Inspired by multimodal zero-shot retrieval [18] and object detection, particularly Mercier et al.'s work [29], it utilizes the panorama as a global context for query matching." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 386, + 201, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 201, + 399 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 201, + 399 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 411, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 482, + 521 + ], + "type": "text", + "content": "SPVLoc conducts 6D indoor localization of a 2D RGB query image within a simple semantic textureless 3D scene model. The model adheres to the Structured3D annotation format [46], containing walls, doors, windows, ceiling, and floor as planes. It can be derived from straightforward from building models [46], or from floor plans using contemporary AI tools [8, 28, 34]. SPVLoc first estimates the image's viewpoint within rendered panoramas from the semantic scene model using cross-domain image-to-panorama matching (Sec. 3.1). It then ranks these predictions based on classification accuracy (Sec. 3.4) and performs relative 6D pose regression with respect to the best-matching reference panorama (Sec. 3.2)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 537, + 366, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 537, + 366, + 550 + ], + "spans": [ + { + "bbox": [ + 132, + 537, + 366, + 550 + ], + "type": "text", + "content": "3.1 Semantic Panoramic Viewport Matching" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "Problem Definition. Our objective is to precisely estimate the 6D camera pose denoted as " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " for a given query RGB image " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "I_{q}" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " relative to a semantic 3D model " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": ". To achieve this, we reformulate the 6D indoor localization of " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "I_{q}" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " within " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " as a cross-domain image-to-panorama matching problem. The model is represented by a set of " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " equirectangular reference layouts " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "L = \\{L_{i} \\mid i = 1, \\dots, N\\}" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " consisting of semantic labels in a single channel. Each rendering comprehensively captures the entire environment around a specific vantage point. The viewpoint " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " is a distinct cutout section from a panorama, defining the exact region a camera positioned nearby can see. The perspective distortion of " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "V_{P_{i}}" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " contains information about" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 162, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 162, + 91, + 447, + 102 + ], + "type": "text", + "content": "SPVLoc: Semantic Panoramic Viewport Matching for 6D Localization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "the relative pose offset " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "P_{i}" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "L_{i}" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "I_{q}" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": ". Scoring and ranking viewpoints helps in choosing the optimal panorama " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "L^{\\star}" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": ". The absolute pose " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " is then formed by concatenating " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "P^{\\star}" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "P_{i}" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": ". Identifying the viewpoint and the corresponding pose offset poses several challenges. First, with increasing distance between the two cameras the perspective distortion between image and panorama increases. Second, the matching involves a severe domain gap as a photographic RGB image is matched against a rendered semantic panorama." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 211, + 482, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 211, + 482, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 211, + 482, + 307 + ], + "type": "text", + "content": "Viewport Prediction. Semantic panorama reference renderings are created utilizing a ray-tracing-based renderer [24]. The viewport mask " + }, + { + "bbox": [ + 130, + 211, + 482, + 307 + ], + "type": "inline_equation", + "content": "V_{P}" + }, + { + "bbox": [ + 130, + 211, + 482, + 307 + ], + "type": "text", + "content": " is a binary image that signifies the visibility of points within the equirectangular image from the view of the perspective camera. The calculation involves projecting all 3D points from the panorama in the perspective view. The projection is performed based on the known underlying layout during the training phase. The mask " + }, + { + "bbox": [ + 130, + 211, + 482, + 307 + ], + "type": "inline_equation", + "content": "V_{P}" + }, + { + "bbox": [ + 130, + 211, + 482, + 307 + ], + "type": "text", + "content": " is determined by identifying pixels projecting into the perspective image, taking into account the 3D camera pose, and excluding those covered by occlusion." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 308, + 482, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 482, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 482, + 415 + ], + "type": "text", + "content": "Additionally, the 2D bounding box " + }, + { + "bbox": [ + 130, + 308, + 482, + 415 + ], + "type": "inline_equation", + "content": "v_{bb}" + }, + { + "bbox": [ + 130, + 308, + 482, + 415 + ], + "type": "text", + "content": " is the minimal axis-aligned frame around " + }, + { + "bbox": [ + 130, + 308, + 482, + 415 + ], + "type": "inline_equation", + "content": "V_P" + }, + { + "bbox": [ + 130, + 308, + 482, + 415 + ], + "type": "text", + "content": ". This bounding box is circularly defined within the " + }, + { + "bbox": [ + 130, + 308, + 482, + 415 + ], + "type": "inline_equation", + "content": "360^\\circ" + }, + { + "bbox": [ + 130, + 308, + 482, + 415 + ], + "type": "text", + "content": " panorama, allowing the overlap on the right image side to re-enter on the left side of the panorama. Each predicted bounding box is associated with a confidence score " + }, + { + "bbox": [ + 130, + 308, + 482, + 415 + ], + "type": "inline_equation", + "content": "c_{bb}" + }, + { + "bbox": [ + 130, + 308, + 482, + 415 + ], + "type": "text", + "content": ", which quantifies the probability of the bounding box being a valid match. The network's objective is to predict " + }, + { + "bbox": [ + 130, + 308, + 482, + 415 + ], + "type": "inline_equation", + "content": "V_P" + }, + { + "bbox": [ + 130, + 308, + 482, + 415 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 308, + 482, + 415 + ], + "type": "inline_equation", + "content": "v_{bb}" + }, + { + "bbox": [ + 130, + 308, + 482, + 415 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 308, + 482, + 415 + ], + "type": "inline_equation", + "content": "c_{bb}" + }, + { + "bbox": [ + 130, + 308, + 482, + 415 + ], + "type": "text", + "content": " for each pair of input semantic panorama and query RGB image. The optimal viewpoint corresponds to the predicted bounding box with the highest overall confidence score, " + }, + { + "bbox": [ + 130, + 308, + 482, + 415 + ], + "type": "inline_equation", + "content": "c_{bb}" + }, + { + "bbox": [ + 130, + 308, + 482, + 415 + ], + "type": "text", + "content": ", when presented with a set of panorama images." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 426, + 482, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 485 + ], + "type": "text", + "content": "Network Architecture. Our matching is inspired by the object detection strategy of DTOID [29]. Figure 2 illustrates the network architecture. Initially, " + }, + { + "bbox": [ + 130, + 426, + 482, + 485 + ], + "type": "inline_equation", + "content": "I_{q}" + }, + { + "bbox": [ + 130, + 426, + 482, + 485 + ], + "type": "text", + "content": " is encoded with an EfficientNet-S backbone. The features obtained after the fourth and the last downsampling layer are preserved. Subsequently, both outputs are rescaled to a tensor of spatial size of " + }, + { + "bbox": [ + 130, + 426, + 482, + 485 + ], + "type": "inline_equation", + "content": "7 \\times 7" + }, + { + "bbox": [ + 130, + 426, + 482, + 485 + ], + "type": "text", + "content": " and concatenated to a tensor " + }, + { + "bbox": [ + 130, + 426, + 482, + 485 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 426, + 482, + 485 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 486, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 581 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 486, + 482, + 581 + ], + "type": "text", + "content": " is encoded using a DenseNet-121 backbone, extending up to the fifth downsampling block to extract features " + }, + { + "bbox": [ + 130, + 486, + 482, + 581 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 130, + 486, + 482, + 581 + ], + "type": "text", + "content": ". The information extracted from the input image guides the panorama branch to estimate the viewport mask and bounding box. This guidance is achieved through the application of a depth-wise correlation technique (denoted as " + }, + { + "bbox": [ + 130, + 486, + 482, + 581 + ], + "type": "inline_equation", + "content": "\\star_{\\mathbf{d}}" + }, + { + "bbox": [ + 130, + 486, + 482, + 581 + ], + "type": "text", + "content": ") [6,29]. " + }, + { + "bbox": [ + 130, + 486, + 482, + 581 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 486, + 482, + 581 + ], + "type": "text", + "content": " is processed through two consecutive convolution blocks, each incorporating convolution, batch normalization, and ELU activation, resulting in a set of " + }, + { + "bbox": [ + 130, + 486, + 482, + 581 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 486, + 482, + 581 + ], + "type": "text", + "content": " filters " + }, + { + "bbox": [ + 130, + 486, + 482, + 581 + ], + "type": "inline_equation", + "content": "\\hat{F}" + }, + { + "bbox": [ + 130, + 486, + 482, + 581 + ], + "type": "text", + "content": " used during the depth-wise correlation " + }, + { + "bbox": [ + 130, + 486, + 482, + 581 + ], + "type": "inline_equation", + "content": "R \\star_{\\mathbf{d}} \\hat{F}" + }, + { + "bbox": [ + 130, + 486, + 482, + 581 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 582, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 640 + ], + "type": "text", + "content": "Furthermore, " + }, + { + "bbox": [ + 130, + 582, + 482, + 640 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 582, + 482, + 640 + ], + "type": "text", + "content": " undergoes compression via global max pooling and is both multiplied and subtracted from " + }, + { + "bbox": [ + 130, + 582, + 482, + 640 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 130, + 582, + 482, + 640 + ], + "type": "text", + "content": ", resulting in three feature tensors with injected image information. The features are individually processed with an additional convolution block and then concatenated. The concatenated features further undergo a final convolution block, leading to the correlated features " + }, + { + "bbox": [ + 130, + 582, + 482, + 640 + ], + "type": "inline_equation", + "content": "R^{\\star}" + }, + { + "bbox": [ + 130, + 582, + 482, + 640 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": "A decoder estimates the viewpoint of the perspective camera projected in the layout of the omnidirectional panorama. The viewpoint in the panorama is" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "text", + "content": "N. Gard et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 482, + 213 + ], + "type": "text", + "content": "estimated as a bounding box with a BBox Class. and BBox Reg. head, similar to RetinaNet [25]. The BBox Class. head's output is optimized using focal loss " + }, + { + "bbox": [ + 131, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1^{VP}" + }, + { + "bbox": [ + 131, + 116, + 482, + 213 + ], + "type": "text", + "content": ", while the BBox Reg. head is optimized with a smooth " + }, + { + "bbox": [ + 131, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "l_{1}" + }, + { + "bbox": [ + 131, + 116, + 482, + 213 + ], + "type": "text", + "content": " loss " + }, + { + "bbox": [ + 131, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_2^{VP}" + }, + { + "bbox": [ + 131, + 116, + 482, + 213 + ], + "type": "text", + "content": ". The BBox Class. provides a score for a given estimation, ensuring that the bounding box with the highest probability corresponds to the best estimate. Additionally, " + }, + { + "bbox": [ + 131, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "V_{P}" + }, + { + "bbox": [ + 131, + 116, + 482, + 213 + ], + "type": "text", + "content": " representing the exact viewpoint of the perspective camera is estimated with the aid of another convolutional decoder head and optimized with binary cross entropy loss " + }, + { + "bbox": [ + 131, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_3^{VP}" + }, + { + "bbox": [ + 131, + 116, + 482, + 213 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 220, + 483, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 220, + 483, + 281 + ], + "spans": [ + { + "bbox": [ + 131, + 220, + 483, + 281 + ], + "type": "text", + "content": "Perspective Supervision. The encoded image features " + }, + { + "bbox": [ + 131, + 220, + 483, + 281 + ], + "type": "inline_equation", + "content": "I_{E}" + }, + { + "bbox": [ + 131, + 220, + 483, + 281 + ], + "type": "text", + "content": " are connected to a convolutional decoder, outputting the underlying semantic and normal map to ensure that features closely corresponding to the image content are extracted. The semantics are learned with cross entropy loss " + }, + { + "bbox": [ + 131, + 220, + 483, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_4^{VP}" + }, + { + "bbox": [ + 131, + 220, + 483, + 281 + ], + "type": "text", + "content": " and the normals are learned with cosine loss " + }, + { + "bbox": [ + 131, + 220, + 483, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_5^{VP}" + }, + { + "bbox": [ + 131, + 220, + 483, + 281 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 295, + 376, + 307 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 295, + 376, + 307 + ], + "spans": [ + { + "bbox": [ + 131, + 295, + 376, + 307 + ], + "type": "text", + "content": "3.2 Feature-Correlation-based Pose Regression" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 312, + 482, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 482, + 395 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 482, + 395 + ], + "type": "text", + "content": "The correlated features, denoted as " + }, + { + "bbox": [ + 130, + 312, + 482, + 395 + ], + "type": "inline_equation", + "content": "R^{\\star}" + }, + { + "bbox": [ + 130, + 312, + 482, + 395 + ], + "type": "text", + "content": ", encode information about the viewpoint of " + }, + { + "bbox": [ + 130, + 312, + 482, + 395 + ], + "type": "inline_equation", + "content": "I_{q}" + }, + { + "bbox": [ + 130, + 312, + 482, + 395 + ], + "type": "text", + "content": " within a " + }, + { + "bbox": [ + 130, + 312, + 482, + 395 + ], + "type": "inline_equation", + "content": "L_{i}" + }, + { + "bbox": [ + 130, + 312, + 482, + 395 + ], + "type": "text", + "content": ", but should also lead to relative pose offset " + }, + { + "bbox": [ + 130, + 312, + 482, + 395 + ], + "type": "inline_equation", + "content": "P_{i}" + }, + { + "bbox": [ + 130, + 312, + 482, + 395 + ], + "type": "text", + "content": " between them. Given that " + }, + { + "bbox": [ + 130, + 312, + 482, + 395 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 312, + 482, + 395 + ], + "type": "text", + "content": " is synthetically generated, the precise position of " + }, + { + "bbox": [ + 130, + 312, + 482, + 395 + ], + "type": "inline_equation", + "content": "L_{i}" + }, + { + "bbox": [ + 130, + 312, + 482, + 395 + ], + "type": "text", + "content": " is known. Therefore, determining the relative position of " + }, + { + "bbox": [ + 130, + 312, + 482, + 395 + ], + "type": "inline_equation", + "content": "I_{q}" + }, + { + "bbox": [ + 130, + 312, + 482, + 395 + ], + "type": "text", + "content": " in relation to " + }, + { + "bbox": [ + 130, + 312, + 482, + 395 + ], + "type": "inline_equation", + "content": "L_{i}" + }, + { + "bbox": [ + 130, + 312, + 482, + 395 + ], + "type": "text", + "content": " is equivalent to establishing its absolute position. All rendered panoramas exhibit identity orientation, due to horizontal alignment and a northward orientation. Consequently, the rotation delta directly corresponds to the absolute rotation of the camera." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 396, + 482, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 396, + 482, + 503 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 482, + 503 + ], + "type": "text", + "content": "An additional Pose head processes " + }, + { + "bbox": [ + 130, + 396, + 482, + 503 + ], + "type": "inline_equation", + "content": "R^{\\star}" + }, + { + "bbox": [ + 130, + 396, + 482, + 503 + ], + "type": "text", + "content": " to estimate the relative pose offset of the camera concerning a panorama rendering (Fig. 3). We first utilize three convolution blocks, whereby the first two are followed by a max-pooling layer, to compress the features. The flattened features form an 1D embedding " + }, + { + "bbox": [ + 130, + 396, + 482, + 503 + ], + "type": "inline_equation", + "content": "r^{\\star} \\in \\mathbb{R}^{C}" + }, + { + "bbox": [ + 130, + 396, + 482, + 503 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 130, + 396, + 482, + 503 + ], + "type": "inline_equation", + "content": "C = 640" + }, + { + "bbox": [ + 130, + 396, + 482, + 503 + ], + "type": "text", + "content": " for a panorama of dimensions " + }, + { + "bbox": [ + 130, + 396, + 482, + 503 + ], + "type": "inline_equation", + "content": "256 \\times 128" + }, + { + "bbox": [ + 130, + 396, + 482, + 503 + ], + "type": "text", + "content": ". Optionally, to handle different camera field of views during training and testing, " + }, + { + "bbox": [ + 130, + 396, + 482, + 503 + ], + "type": "inline_equation", + "content": "r^{\\star}" + }, + { + "bbox": [ + 130, + 396, + 482, + 503 + ], + "type": "text", + "content": " is normalized and multiplied by the horizontal field of view angle in radians. The result is fed into a four-layer MLP, producing the 3D offset to the query image location " + }, + { + "bbox": [ + 130, + 396, + 482, + 503 + ], + "type": "inline_equation", + "content": "\\hat{t} i \\in \\mathbb{R}^{3}" + }, + { + "bbox": [ + 130, + 396, + 482, + 503 + ], + "type": "text", + "content": " and a rotation given as the logarithm of the unit quaternion " + }, + { + "bbox": [ + 130, + 396, + 482, + 503 + ], + "type": "inline_equation", + "content": "\\hat{r} i = \\log (\\hat{q})_{ij} \\in \\mathbb{R}^{3}" + }, + { + "bbox": [ + 130, + 396, + 482, + 503 + ], + "type": "text", + "content": " [41]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 503, + 482, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 503, + 482, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 503, + 482, + 540 + ], + "type": "text", + "content": "The MLP is evaluated using two loss functions. " + }, + { + "bbox": [ + 130, + 503, + 482, + 540 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1^P" + }, + { + "bbox": [ + 130, + 503, + 482, + 540 + ], + "type": "text", + "content": " assesses the precision of the relative translation, and " + }, + { + "bbox": [ + 130, + 503, + 482, + 540 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_2^P" + }, + { + "bbox": [ + 130, + 503, + 482, + 540 + ], + "type": "text", + "content": " evaluates the accuracy of the final rotation estimation. Both losses are computed using the " + }, + { + "bbox": [ + 130, + 503, + 482, + 540 + ], + "type": "inline_equation", + "content": "l_{1}" + }, + { + "bbox": [ + 130, + 503, + 482, + 540 + ], + "type": "text", + "content": " distance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 554, + 228, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 554, + 228, + 566 + ], + "spans": [ + { + "bbox": [ + 131, + 554, + 228, + 566 + ], + "type": "text", + "content": "3.3 Optimization" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 571, + 482, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 571, + 482, + 631 + ], + "spans": [ + { + "bbox": [ + 130, + 571, + 482, + 631 + ], + "type": "text", + "content": "The training loss is the sum of the two losses of the MLP and the five losses of the viewpoint estimation. With seven losses, our multi-task learning has to be balanced by using a different weight for each loss. As in [20, 21], we use an exponential mapping and weight multiple loss functions by considering the homoscedastic uncertainty of each task and learned weighting factors. The final loss is given as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 218, + 636, + 482, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 218, + 636, + 482, + 669 + ], + "spans": [ + { + "bbox": [ + 218, + 636, + 482, + 669 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\sum_ {i = 1} ^ {2} \\mathcal {L} _ {i} ^ {P} e ^ {- \\beta_ {i}} + \\beta_ {i} + \\sum_ {i = 1} ^ {5} \\mathcal {L} _ {i} ^ {V P} e ^ {- \\gamma_ {i}} + \\gamma_ {i} \\tag {1}", + "image_path": "96bcbc20eb94b296df845eebb32041e26a1b177192d344b3a88c140761b0ac66.jpg" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 161, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 161, + 91, + 448, + 103 + ], + "type": "text", + "content": "SPVLoc: Semantic Panoramic Viewport Matching for 6D Localization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 193, + 114, + 421, + 213 + ], + "blocks": [ + { + "bbox": [ + 193, + 114, + 421, + 213 + ], + "lines": [ + { + "bbox": [ + 193, + 114, + 421, + 213 + ], + "spans": [ + { + "bbox": [ + 193, + 114, + 421, + 213 + ], + "type": "image", + "image_path": "a3821fff45ec52102ba1cadb4f1ced6a630caebc8364cc0d4b3fa3174c85a14b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 137, + 221, + 474, + 234 + ], + "lines": [ + { + "bbox": [ + 137, + 221, + 474, + 234 + ], + "spans": [ + { + "bbox": [ + 137, + 221, + 474, + 234 + ], + "type": "text", + "content": "Fig. 3: The Pose head utilizes convolutions, an MLP, and a FoV-based side input." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 255, + 208, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 255, + 208, + 266 + ], + "spans": [ + { + "bbox": [ + 132, + 255, + 208, + 266 + ], + "type": "text", + "content": "3.4 Inference" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 272, + 482, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 272, + 482, + 355 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 482, + 355 + ], + "type": "text", + "content": "During inference, panorama positions are determined from a fixed 2D grid superimposed on the floor plan, with virtual cameras set at a constant height denoted as " + }, + { + "bbox": [ + 130, + 272, + 482, + 355 + ], + "type": "inline_equation", + "content": "h_{pano}" + }, + { + "bbox": [ + 130, + 272, + 482, + 355 + ], + "type": "text", + "content": ", and the grid's dimensions specified as " + }, + { + "bbox": [ + 130, + 272, + 482, + 355 + ], + "type": "inline_equation", + "content": "xy_{pano}" + }, + { + "bbox": [ + 130, + 272, + 482, + 355 + ], + "type": "text", + "content": ". Virtual cameras are exclusively sampled above floors to ensure valid reference images inside rooms. Setting " + }, + { + "bbox": [ + 130, + 272, + 482, + 355 + ], + "type": "inline_equation", + "content": "xy_{pano}" + }, + { + "bbox": [ + 130, + 272, + 482, + 355 + ], + "type": "text", + "content": " as large as possible is advantageous in order to minimize the number of references, given that feature correlation and bounding box regression are executed for every panorama." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 356, + 482, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 356, + 482, + 440 + ], + "spans": [ + { + "bbox": [ + 130, + 356, + 482, + 440 + ], + "type": "text", + "content": "For pose estimation, we select reference positions with the top-" + }, + { + "bbox": [ + 130, + 356, + 482, + 440 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 356, + 482, + 440 + ], + "type": "text", + "content": " highest bounding box classification scores. The absolute pose is determined by the result of the Pose head, where the selected reference position is added to the translation. To eliminate the necessity for rendering during inference, we propose pre-generating reference panoramas for a floor plan. Additionally, the encoding of these panoramas can be precomputed. This ensures that only the encoded features " + }, + { + "bbox": [ + 130, + 356, + 482, + 440 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 130, + 356, + 482, + 440 + ], + "type": "text", + "content": " need to be stored, significantly streamlining the inference process." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 449, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 510 + ], + "type": "text", + "content": "Refinement. Without altering the network architecture or training process, SPVLoc can enhance an estimated pose by rendering a new reference panorama at the estimated position and re-executing the relative pose regression. The smaller step to the actual target position in this second iteration simplifies the problem, allowing for increased precision in the refined pose estimation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 525, + 230, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 230, + 539 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 230, + 539 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 548, + 481, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 548, + 481, + 572 + ], + "spans": [ + { + "bbox": [ + 130, + 548, + 481, + 572 + ], + "type": "text", + "content": "To validate our method we conduct several experiments and make use of two publicly available and frequently used datasets." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "type": "text", + "content": "Structured3D (S3D) [46] consists of 3500 near-photorealistic professionally modeled indoor environments, each annotated with ground truth 3D structure information. It encompasses 21835 panoramic images, each provided in both furnished and unfurnished configurations. The authors define a unified representation that describes scenes using geometric primitives such as planes, lines, junction points, and semantic annotations including floor, ceiling, wall, door, and window classes." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "text", + "content": "N. Gard et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 482, + 187 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 187 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 187 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 187 + ], + "type": "image", + "image_path": "0f30960556da9c1e4a2ddd24240ce607f2f35d2cf3a963cead6b61282d4b06f9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 196, + 482, + 219 + ], + "lines": [ + { + "bbox": [ + 130, + 196, + 482, + 219 + ], + "spans": [ + { + "bbox": [ + 130, + 196, + 482, + 219 + ], + "type": "text", + "content": "Fig. 4: ZinD data preparation. Annotations generate 3D reference models (left), while resampled panoramas create perspective train and test images (right)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 242, + 482, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 242, + 482, + 338 + ], + "spans": [ + { + "bbox": [ + 130, + 242, + 482, + 338 + ], + "type": "text", + "content": "Zillow Indoor (ZInD) [10] includes 67,448 panorama images from 1,575 unfurnished residential homes, globally aligned and registered to floor plans. The 3D annotations of panoramas within a scene are adequate for transforming them into the S3D representation format. Before training, we transform all data to that format. Figure 4 shows an example reference model generated from ZInD annotations. Unlike S3D, ZInD includes room openings such as corridors connected to rooms without doors, enhancing structural variation in semantic renderings." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 341, + 483, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 341, + 483, + 497 + ], + "spans": [ + { + "bbox": [ + 130, + 341, + 483, + 497 + ], + "type": "text", + "content": "For both datasets, we employ the official train-test scene split. To enhance control over the field of view and angle variation, we sample perspective test images from the RGB panoramas. They are registered in the 3D model, so the semantics are known for each perspective image. We compile test-sets suitable for 6D pose estimation and structured evaluation. The test sets consist of square perspective images with a viewing angle of 60, 90, and " + }, + { + "bbox": [ + 130, + 341, + 483, + 497 + ], + "type": "inline_equation", + "content": "120^{\\circ}" + }, + { + "bbox": [ + 130, + 341, + 483, + 497 + ], + "type": "text", + "content": ", respectively, along with variations in roll and pitch, and random yaw angles for ZinD and S3D. To ensure meaningful views, we apply two criteria on the sampled yaw. First, the images contain a minimum of three semantic classes, like wall-floor-ceiling or wall-window-floor. Second, the camera faces inside the room rather than towards the next wall. The criteria are applied to the " + }, + { + "bbox": [ + 130, + 341, + 483, + 497 + ], + "type": "inline_equation", + "content": "60^{\\circ}" + }, + { + "bbox": [ + 130, + 341, + 483, + 497 + ], + "type": "text", + "content": " images (4420 for ZiND and 1133 for S3D). Subsequent variations in opening angles and roll-pitch configurations are generated from this starting point." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 514, + 196, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 196, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 196, + 525 + ], + "type": "text", + "content": "4.1 Details" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 534, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 594 + ], + "type": "text", + "content": "For evaluation, we report the fraction of frames localized correctly within a threshold " + }, + { + "bbox": [ + 130, + 534, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 534, + 482, + 594 + ], + "type": "text", + "content": ". We also list the median rotation and translation error, as well as the fraction of frames localized within " + }, + { + "bbox": [ + 130, + 534, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 534, + 482, + 594 + ], + "type": "text", + "content": " and an additional angle threshold " + }, + { + "bbox": [ + 130, + 534, + 482, + 594 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 130, + 534, + 482, + 594 + ], + "type": "text", + "content": ". In the retrieval process, the top-3 best matches are selected, and we report recall metrics for both top-1 and the best among the top-3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Training. For the state of the art comparison, we train a model with a variable field of view, while other experiments use a model trained with a fixed opening angle of " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "90^{\\circ}" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": ". The model with variable viewing angles randomly samples views between 45 and " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "135^{\\circ}" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " during training. We consistently render " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " randomly positioned panoramas around the query within a radius of " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pm r_1" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "xy" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " direction and" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 161, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 161, + 91, + 448, + 103 + ], + "type": "text", + "content": "SPVLoc: Semantic Panoramic Viewport Matching for 6D Localization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 179, + 479, + 272 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 1: Comparison with state-of-the-art baseline on ZInD and S3D: Median translation " + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "(\\mathrm{terr}_{\\mathrm{xy}})" + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": " in cm, rotation error (rerrayw) in deg. for instances localized under 1m. Reporting recall at various translation accuracy levels, recall for inliers (<1m and <30°), and top-3 recall at 1m. Ours† refers to our model tested with a slightly different testset, incorporating pitch and roll rotation variation (" + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "\\pm 10^{\\circ}" + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 179, + 479, + 272 + ], + "lines": [ + { + "bbox": [ + 132, + 179, + 479, + 272 + ], + "spans": [ + { + "bbox": [ + 132, + 179, + 479, + 272 + ], + "type": "table", + "html": "
ZInDStructured3D (Furnishing-Level : Full)
QueryMethod<1m med terr (cm)<1m med rerr (deg)10cm (%)50cm (%)1m (%)1m & 30° (%)top-3 1m (%)<1m med terr (cm)<1m med rerr (deg)10cm (%)50cm (%)1m (%)1m & 30° (%)top-3 1m(%)
Perspective 60° FoVLASER [30]33.531.034.0547.2464.5263.5378.6417.600.8416.4663.2268.3967.3486.08
Ours17.161.5021.0481.3386.7686.0994.1212.251.2134.9487.1389.4988.7996.15
\\( Ours^† \\)17.771.6319.5981.1387.1086.4394.8613.311.2430.8284.6888.1887.4895.80
Perspective 90° FoVLASER [30]26.280.778.6967.0180.9080.2589.7314.600.6723.4772.7776.8076.1891.94
Ours13.631.3430.5785.5990.8490.3496.369.170.8350.0989.6792.1291.8697.37
\\( Ours^† \\)14.251.3729.5985.9091.3390.9596.9710.090.9145.6289.3292.1291.9496.85
Perspective 120° FoVLASER [30]23.100.7812.6975.0985.6385.2092.3812.900.7629.0777.7682.1481.3594.13
Ours14.781.4529.0585.1491.7291.2796.9211.690.8739.4988.0092.1291.9497.37
\\( Ours^† \\)15.131.5328.1286.1392.7192.3597.2411.770.8838.1889.6793.0892.9197.55
", + "image_path": "5ece2e871f34bff2e6e6f2d4b5da189235d2e87488b0bb0c0ca29fdadbc0c0cc.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 293, + 483, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 483, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 483, + 426 + ], + "type": "inline_equation", + "content": "\\pm r_2" + }, + { + "bbox": [ + 130, + 293, + 483, + 426 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 130, + 293, + 483, + 426 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 293, + 483, + 426 + ], + "type": "text", + "content": " (upward) direction. For a single query, the viewpoint and relative pose offset is therefore calculated against multiple references. In our experimental setup, we set " + }, + { + "bbox": [ + 130, + 293, + 483, + 426 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 293, + 483, + 426 + ], + "type": "text", + "content": " to 4, " + }, + { + "bbox": [ + 130, + 293, + 483, + 426 + ], + "type": "inline_equation", + "content": "r_1" + }, + { + "bbox": [ + 130, + 293, + 483, + 426 + ], + "type": "text", + "content": " to 1.4 meters, and " + }, + { + "bbox": [ + 130, + 293, + 483, + 426 + ], + "type": "inline_equation", + "content": "r_2" + }, + { + "bbox": [ + 130, + 293, + 483, + 426 + ], + "type": "text", + "content": " to 30 centimeters. Additionally, one random negative example is generated outside " + }, + { + "bbox": [ + 130, + 293, + 483, + 426 + ], + "type": "inline_equation", + "content": "r_1" + }, + { + "bbox": [ + 130, + 293, + 483, + 426 + ], + "type": "text", + "content": ", likely in a different room, with the target to enhance the network's ability to distinguish subtle room differences. Images are sampled with random yaw and " + }, + { + "bbox": [ + 130, + 293, + 483, + 426 + ], + "type": "inline_equation", + "content": "\\pm 10^{\\circ}" + }, + { + "bbox": [ + 130, + 293, + 483, + 426 + ], + "type": "text", + "content": " random pitch and roll angle. The batch size is set to 40, comprising 40 query images and 200 panoramas, and the training is performed on a single NVIDIA A100 GPU. Query images with fewer than three semantic classes are disregarded during the loss calculation process. Training runs for approximately 42,000 steps with an initial learning rate of " + }, + { + "bbox": [ + 130, + 293, + 483, + 426 + ], + "type": "inline_equation", + "content": "2.5 \\times 10^{-4}" + }, + { + "bbox": [ + 130, + 293, + 483, + 426 + ], + "type": "text", + "content": ", halved two times during training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 437, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 510 + ], + "type": "text", + "content": "Testing. During testing, panoramas are sampled on a " + }, + { + "bbox": [ + 130, + 437, + 482, + 510 + ], + "type": "inline_equation", + "content": "1.2 \\times 1.2" + }, + { + "bbox": [ + 130, + 437, + 482, + 510 + ], + "type": "text", + "content": " m grid, requiring significantly fewer samples than other rendering-based approaches such as LaLaLoc [17], which uses a " + }, + { + "bbox": [ + 130, + 437, + 482, + 510 + ], + "type": "inline_equation", + "content": "0.5 \\times 0.5" + }, + { + "bbox": [ + 130, + 437, + 482, + 510 + ], + "type": "text", + "content": " m grid. Additionally, for a scene, all panoramic views can be encoded initially, and during inference, only the query image needs to be encoded. If not otherwise noted the studies are conducted on ZInD with fixed 90 degree FoV, reporting the 3D rotation and translation error." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 526, + 353, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 526, + 353, + 538 + ], + "spans": [ + { + "bbox": [ + 131, + 526, + 353, + 538 + ], + "type": "text", + "content": "4.2 Comparison with the State of the Art" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 545, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 616 + ], + "type": "text", + "content": "To our knowledge, SPVLoc is the first method to perform 6D localization in unseen scenes with respect to semantic 3D models. The state of the-art method for scene-independent global localization, LASER [30], takes perspective views as input and matches them against a semantic floor plan. Hence, it estimates only two position and one rotation degree of freedom. Nonetheless, it serves as a strong baseline to assess the localization accuracy of our method." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Table 1 shows the results on ZInD and S3D. Given that we compare against a 2D method, our evaluation also includes recall based on 2D distance and angular error derived from the yaw angle. The yaw angle is extracted from the estimated rotation matrix. Notably, SPVLoc without refinement outperforms LASER in" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "text", + "content": "N. Gard et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 182, + 146, + 424, + 248 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 2: Ablation study. SPVLoc variations compared on ZInD with " + }, + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "type": "inline_equation", + "content": "90^{\\circ}" + }, + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "type": "text", + "content": " FoV. In contrast to Tab. 1, terr denotes 3D position errors, and rerr denotes 3D rotation errors." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 182, + 146, + 424, + 248 + ], + "lines": [ + { + "bbox": [ + 182, + 146, + 424, + 248 + ], + "spans": [ + { + "bbox": [ + 182, + 146, + 424, + 248 + ], + "type": "table", + "html": "
Abl. TypeMethod<1m med terr (cm)<1m med rerr (deg)10cm (%)50cm (%)1m (%)1m & 30° (%)top-3 1m (%)
(Ours)base14.312.0526.8886.9291.9091.6196.56
+ refine12.321.9633.6990.8492.7691.74-
Loss- query dec.15.702.2321.9785.5090.7990.4396.45
- view dec.15.192.2123.6085.3690.4589.8296.49
- view/query dec.16.222.1921.5684.5089.5989.1296.36
Training- neg. samples15.062.2822.6076.6582.1981.7492.99
Arch.Δ Resnet18 (query)15.562.5620.8886.5690.7290.3296.56
Δ Equiconv (pano)14.982.1325.0787.4491.7091.3196.97
Input+ normal14.682.2825.0985.9390.8490.2396.47
+ depth14.062.1027.6088.0592.1591.7996.74
", + "image_path": "686541602b623601a940925072dcd7d9407c7f516994ccf6f02e7aa8e0dda318.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 270, + 482, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 365 + ], + "type": "text", + "content": "terms of localization accuracy, exhibiting higher recall across all categories, an increased inlier rate, and significantly reduced median translation error. In terms of 2D rotation, our method is slightly less precise. This is because LASER estimates yaw directly while we estimate a full rotation matrix. Estimating camera height and roll and pitch angle is an advantage of our method. In each last row " + }, + { + "bbox": [ + 130, + 270, + 482, + 365 + ], + "type": "inline_equation", + "content": "(Ours^{\\dagger})" + }, + { + "bbox": [ + 130, + 270, + 482, + 365 + ], + "type": "text", + "content": ", we evaluate our method on an alternate test set with roll and pitch variations " + }, + { + "bbox": [ + 130, + 270, + 482, + 365 + ], + "type": "inline_equation", + "content": "(\\pm 10^{\\circ})" + }, + { + "bbox": [ + 130, + 270, + 482, + 365 + ], + "type": "text", + "content": ". The results suggest that these variations do not substantially impact the accuracy of our method for 2D localization." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 365, + 483, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 483, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 483, + 475 + ], + "type": "text", + "content": "Regarding further methods, LASER has showcased a significant performance improvement over PF-net [19] and non-AI-based Monte-Carlo Localization (MCL) [11]. PF-net, originally designed for sequential time updating, does not perform well in single-query localization scenarios. MCL neglects semantic annotations, uses LIDAR instead of images, which also has been shown to drastically reduce performance [30]. Our substantial improvement over LASER therefore also illustrates a performance gain over aforementioned methods. Another experiment using the pre-rendered perspective images of S3D, including a comparison to " + }, + { + "bbox": [ + 130, + 365, + 483, + 475 + ], + "type": "inline_equation", + "content": "\\mathrm{F}^3\\mathrm{Loc}" + }, + { + "bbox": [ + 130, + 365, + 483, + 475 + ], + "type": "text", + "content": " [9], is part of the supplementary material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 491, + 265, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 491, + 265, + 502 + ], + "spans": [ + { + "bbox": [ + 131, + 491, + 265, + 502 + ], + "type": "text", + "content": "4.3 Performance Studies" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 510, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 594 + ], + "type": "text", + "content": "Ablation Study. We commence our experimentation with the full network (base) configuration. Subsequently, we modify the network by removing specific components (Tab. 2). Initially, we test the influence of perspective supervision (-query dec.) and the view segment task head (-view dec.) by excluding them during training. Their exclusion reduces network performance, likely because the network's ability to understand the image content and establish correspondences is diminished. Removing both (-view/query dec.) further reduces accuracy." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 641 + ], + "type": "text", + "content": "We then exclude the negative samples from different rooms during training (-neg. samples). This significantly reduces the amount of correctly localized images. It supports the hypothesis that teaching the network to recognize spaces where there is no match helps to improve the accuracy of predictions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": "To modify the architecture, the image encoder EfficientNet-S is replaced with a smaller ResNet-18 [15] (Δ Resnet18), resulting in a performance drop. This" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 161, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 161, + 91, + 447, + 102 + ], + "type": "text", + "content": "SPVLoc: Semantic Panoramic Viewport Matching for 6D Localization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 261 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 261 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 261 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 261 + ], + "type": "image", + "image_path": "1e784445864227badce7d8b33244ad5e0d2e554920c1a55c0a59cedd36937147.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 270, + 482, + 304 + ], + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 304 + ], + "type": "text", + "content": "Fig. 5: Qualitative localization results. Top to bottom - query, rendering with top-1 estimated pose, panorama with estimated viewport, map. Green box: success for top-1 match. Yellow box: success for top-2 match. Red box: failure case." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 327, + 482, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 482, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 482, + 399 + ], + "type": "text", + "content": "suggests an increased demand for network capacity to analyze the images. In the panorama encoder, replacing all convolutional layers with Equiconv [13] (ΔEquiconv), a specialized convolution for panorama images that remaps query locations from Euclidean to spherical space, does not result in a performance gain. This suggests that the capacity of the small DenseNet-121 is sufficient for analyzing the low-resolution semantic inputs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 399, + 483, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 483, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 483, + 460 + ], + "type": "text", + "content": "Lastly, we incorporate additional input modalities for the panoramas into the panorama encoder. While the addition of world coordinate normals (+normals) did not improve the result, including an extra depth map slightly enhanced the outcome (+depth). This suggests that there are cases where additional information is encoded in the depth that cannot be inferred from semantics alone." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 472, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 628 + ], + "type": "text", + "content": "Qualitative Results. Figure 5 shows selected SPVLoc outcomes. The top row displays camera images, and the second row shows renderings from the semantic 3D model with estimated top-1 poses. All green boxed images are correctly localized for the top-1 match. In the yellow box, the top-1 match fails, but the top-2 match is accurate. The red box example cannot be localized. Across all examples, a pose with high visual similarity is found. Ambiguous examples arise from similar layouts occurring in multiple rooms. Failure cases are attributed to either images lacking sufficient structural information for matching or outliers in the data, such as those taken on balconies or in dark cellars. The third row displays the chosen reference panorama, the selected viewpoint bounding box corresponding to the camera image, and a semi-transparent overlay representing the estimated viewpoint mask. Additional detailed visualizations are part of the supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Influence of Sampling Grid Sizes and Refinement. Reference positions result from overlaying a grid onto the floorplan, and the grid density significantly" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "text", + "content": "N. Gard et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 146, + 308, + 244 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 318, + 137 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 318, + 137 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 318, + 137 + ], + "type": "text", + "content": "Table 3: Grid size vs. accuracy: Deviation of " + }, + { + "bbox": [ + 131, + 114, + 318, + 137 + ], + "type": "inline_equation", + "content": "10\\mathrm{cm}" + }, + { + "bbox": [ + 131, + 114, + 318, + 137 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 114, + 318, + 137 + ], + "type": "inline_equation", + "content": "1\\mathrm{m}" + }, + { + "bbox": [ + 131, + 114, + 318, + 137 + ], + "type": "text", + "content": " recall from (boxed) baseline." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 146, + 308, + 244 + ], + "lines": [ + { + "bbox": [ + 135, + 146, + 308, + 244 + ], + "spans": [ + { + "bbox": [ + 135, + 146, + 308, + 244 + ], + "type": "table", + "html": "
Sample dist.Samples per \\( m^2 \\)Adv.10cm (%)10cm + refine1m (%)1m + refine
0.72.07+1.00+7.49+1.86+2.38
0.91.25+0.95+6.99+1.79+2.08
1.20.70\\( \\boxed{26.88} \\)+6.81\\( \\boxed{91.90} \\)+0.86
1.20.48✓+0.57+6.95-0.63+0.61
1.50.45-1.00+6.63-2.47-1.02
1.50.31✓-1.04+6.47-2.67-0.38
2.00.26-4.19+3.60-9.05-5.79
2.00.20✓-1.76+5.81-3.91-1.18
", + "image_path": "3708d75da9aed7390a3951a2fb970fb20da92ec08a745d5400f25d184e7174bd.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 326, + 119, + 370, + 188 + ], + "blocks": [ + { + "bbox": [ + 326, + 119, + 370, + 188 + ], + "lines": [ + { + "bbox": [ + 326, + 119, + 370, + 188 + ], + "spans": [ + { + "bbox": [ + 326, + 119, + 370, + 188 + ], + "type": "image", + "image_path": "91b10ff04ac2a7894072d24db67c926cf32755479e03db8746e94c5cda1c653d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 325, + 194, + 481, + 228 + ], + "lines": [ + { + "bbox": [ + 325, + 194, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 325, + 194, + 481, + 228 + ], + "type": "text", + "content": "Fig. 6: Reference positions for global grid of size " + }, + { + "bbox": [ + 325, + 194, + 481, + 228 + ], + "type": "inline_equation", + "content": "0.7\\mathrm{m}" + }, + { + "bbox": [ + 325, + 194, + 481, + 228 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 325, + 194, + 481, + 228 + ], + "type": "inline_equation", + "content": "1.5\\mathrm{m}" + }, + { + "bbox": [ + 325, + 194, + 481, + 228 + ], + "type": "text", + "content": " and local grid of size " + }, + { + "bbox": [ + 325, + 194, + 481, + 228 + ], + "type": "inline_equation", + "content": "1.5\\mathrm{m}" + }, + { + "bbox": [ + 325, + 194, + 481, + 228 + ], + "type": "text", + "content": " (from left to right)." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 378, + 119, + 422, + 189 + ], + "blocks": [ + { + "bbox": [ + 378, + 119, + 422, + 189 + ], + "lines": [ + { + "bbox": [ + 378, + 119, + 422, + 189 + ], + "spans": [ + { + "bbox": [ + 378, + 119, + 422, + 189 + ], + "type": "image", + "image_path": "af39788a9556d7dbfe55ad0eecd8b086b179e09c3c8c42015b9e761cf7d1158c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 432, + 120, + 473, + 189 + ], + "blocks": [ + { + "bbox": [ + 432, + 120, + 473, + 189 + ], + "lines": [ + { + "bbox": [ + 432, + 120, + 473, + 189 + ], + "spans": [ + { + "bbox": [ + 432, + 120, + 473, + 189 + ], + "type": "image", + "image_path": "5ec39db21124c7bbea1d4ec2e108d18a8e26f0b23d85a2075210fdf383184c46.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 267, + 482, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 267, + 482, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 267, + 482, + 434 + ], + "type": "text", + "content": "affects SPVLoc's efficiency and inference time. Minimizing references reduces rendering effort and requires the model to estimate larger offsets. E.g. to increase the grid from " + }, + { + "bbox": [ + 130, + 267, + 482, + 434 + ], + "type": "inline_equation", + "content": "0.7m" + }, + { + "bbox": [ + 130, + 267, + 482, + 434 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 267, + 482, + 434 + ], + "type": "inline_equation", + "content": "2m" + }, + { + "bbox": [ + 130, + 267, + 482, + 434 + ], + "type": "text", + "content": " reduces the number of samples per " + }, + { + "bbox": [ + 130, + 267, + 482, + 434 + ], + "type": "inline_equation", + "content": "m^2" + }, + { + "bbox": [ + 130, + 267, + 482, + 434 + ], + "type": "text", + "content": " by a factor of 10. We vary the grid size to analyze its impact on performance. We also experiment with a local grid per room substituting the global grid (Adv.) reducing the risk of missing rooms entirely. The local grid is centered on the room outline and additionally places one camera in the middle of each opening. Figure 6 shows different grids for an example plan. Table 3 illustrates the relationships among grid size, global and local sampling, and refinement. Refinement notably improves " + }, + { + "bbox": [ + 130, + 267, + 482, + 434 + ], + "type": "inline_equation", + "content": "10\\mathrm{cm}" + }, + { + "bbox": [ + 130, + 267, + 482, + 434 + ], + "type": "text", + "content": " recall across configurations. For the largest local grid with refinement, the " + }, + { + "bbox": [ + 130, + 267, + 482, + 434 + ], + "type": "inline_equation", + "content": "1\\mathrm{m}" + }, + { + "bbox": [ + 130, + 267, + 482, + 434 + ], + "type": "text", + "content": " recall only decreases by " + }, + { + "bbox": [ + 130, + 267, + 482, + 434 + ], + "type": "inline_equation", + "content": "1.18\\%" + }, + { + "bbox": [ + 130, + 267, + 482, + 434 + ], + "type": "text", + "content": " compared to the baseline configuration. In summary, a trade-off between refinement and dense/sparse scene sampling exists, yet our method consistently delivers strong results with very sparse sampling, making it suitable for large scenes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 448, + 483, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 448, + 483, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 448, + 483, + 545 + ], + "type": "text", + "content": "Influence of Focal Length. Table 4 depicts the performance difference between a variant trained with a known camera focal length and a focal length-independent network trained with varying focal lengths (ranging from 45 to 135 degrees). The fixed focal length network performs slightly better with matching images but loses precision with test images of different focal lengths. In contrast, the focal length-independent network demonstrates the capability to handle such differences. We argue that the significantly enhanced flexibility of the latter justifies the minor performance degradation in practical, real-world scenarios." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "Influence of Camera Rotation Angles. The range of pitch and roll angles the network is able to handle is determined by the variations of angles seen during training. We train the network for ZInD using a angle variation of " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pm 22^{\\circ}" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": " and test with images where the range of variation increases from " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "0^{\\circ}" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "20^{\\circ}" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": " in steps of " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pm 5^{\\circ}" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": ". Table 5 shows only a slight increase in median errors for rotation and translation as the rotation angles in test images increase. The chosen ranges represent a significant coverage and offer substantial improvements over existing methods. In indoor scenes, larger camera tilt often results in images dominated by the ceiling or floor, rendering them unsuitable for semantic matching." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 161, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 161, + 91, + 447, + 102 + ], + "type": "text", + "content": "SPVLoc: Semantic Panoramic Viewport Matching for 6D Localization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 140, + 157, + 304, + 232 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 319, + 148 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 319, + 148 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 319, + 148 + ], + "type": "text", + "content": "Table 4: The fixed FoV network excels at a specific angle but loses precision with changing FoV. The proposed network remains stable." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 140, + 157, + 304, + 232 + ], + "lines": [ + { + "bbox": [ + 140, + 157, + 304, + 232 + ], + "spans": [ + { + "bbox": [ + 140, + 157, + 304, + 232 + ], + "type": "table", + "html": "
Input FoVTrain FoV<1m med terr (cm)<1m med rerr (deg)10cm (%)1m (%)
90°90°14.312.0526.8891.90
45 - 135°16.112.2021.8191.45
80°90°37.512.371.6790.14
45 - 135°16.752.1919.3090.50
", + "image_path": "07ae866ab6ecb22088efd7fd62cf434753c24edfc429a9a920ec03af075e9a6d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 326, + 154, + 477, + 224 + ], + "blocks": [ + { + "bbox": [ + 325, + 121, + 482, + 143 + ], + "lines": [ + { + "bbox": [ + 325, + 121, + 482, + 143 + ], + "spans": [ + { + "bbox": [ + 325, + 121, + 482, + 143 + ], + "type": "text", + "content": "Table 5: Effect of increasing angle variation in test data." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 326, + 154, + 477, + 224 + ], + "lines": [ + { + "bbox": [ + 326, + 154, + 477, + 224 + ], + "spans": [ + { + "bbox": [ + 326, + 154, + 477, + 224 + ], + "type": "table", + "html": "
MetricRoll/Pitch Variation
±0°±5°±10°±15°±20°
<1m med. terr14.4614.8814.8815.2115.91
<1m med. verr2.352.362.382.512.60
10cm (%)24.6423.9424.4623.6922.74
1m (%)91.4591.4791.9992.3892.60
", + "image_path": "592afff64bda853bbba4136d8cbf09ea012544bd02f4583734452413404e12e7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 182, + 275, + 423, + 315 + ], + "blocks": [ + { + "bbox": [ + 131, + 243, + 481, + 266 + ], + "lines": [ + { + "bbox": [ + 131, + 243, + 481, + 266 + ], + "spans": [ + { + "bbox": [ + 131, + 243, + 481, + 266 + ], + "type": "text", + "content": "Table 6: Timing. Performance on ZInD using a NVIDIA Tesla V100, following the configurations outlined in Sec. 4.3, and compared against results from [30]." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 182, + 275, + 423, + 315 + ], + "lines": [ + { + "bbox": [ + 182, + 275, + 423, + 315 + ], + "spans": [ + { + "bbox": [ + 182, + 275, + 423, + 315 + ], + "type": "table", + "html": "
MethodSampling Time (s)Query FpsQuery Module Time (ms)
PF-net [19]48.95 ± 38.955.06 ± 1.77Eff.Net V2 S4.4
LASER [30]0.97 ± 1.098.31 ± 0.64Correlation35.7
Ours1.66 ± 1.0628.63 ± 11.54Pose1.1
", + "image_path": "811c6a3a4d41398bd98d4f4fb6ace48ba700416048fb375106718f055b490243.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 338, + 482, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 338, + 482, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 482, + 399 + ], + "type": "text", + "content": "Limitations. In large spaces with repeating room layouts, the approach's effectiveness may be constrained by the limited level of detail in the semantic reference model. However, enhancing the S3D format by adding more semantic classes and structural information, such as staircases or sanitary installations, could reduce ambiguities." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 413, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 482, + 521 + ], + "type": "text", + "content": "Timing. We split the measured time into scene sampling/preparation time, including position sampling, panorama rendering and encoding, and inference/query time, including query image encoding, feature correlation with bounding box classification, and the Pose task head, but no refinement. Table 6 shows that our method exhibits a considerably lower query time than LASER [30] and PF-net [19]. This difference places our approach close to real-time for moderately sized scenes. Currently, scene sampling time is slower than [30], mainly due to rendering on the CPU without parallelization [24]. Parallelizing this process using GPU or CPU could significantly accelerate scene preparation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 541, + 220, + 554 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 541, + 220, + 554 + ], + "spans": [ + { + "bbox": [ + 132, + 541, + 220, + 554 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "This paper introduces a method for scene-independent model-based 6D localization for indoor scenes, involving a novel method for multimodal image matching (panorama to perspective, RGB to semantic). The matching and retrieval is efficient and scalable under sparse reference sampling. The localization accuracy and inference speed outperform state of the art methods and the inclusion of the 3D-model, reduces ambiguities to precisely estimate 6D poses. Future work involves combining localization and image analysis to enhance digital building models or exploring applications in augmented reality scenarios." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "text", + "content": "N. Gard et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 153 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 153 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 153 + ], + "type": "text", + "content": "Acknowledgements. This work is partially funded by the German Federal Ministry of Economic Affairs and Climate Action (BIMKIT, 01MK21001H) and the German Federal Ministry for Digital and Transport (EConoM, 19OI22009C)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 168, + 198, + 181 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 168, + 198, + 181 + ], + "spans": [ + { + "bbox": [ + 132, + 168, + 198, + 181 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 190, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 190, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 190, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 138, + 190, + 480, + 224 + ], + "type": "text", + "content": "1. Acharya, D., Khoshelham, K., Winter, S.: BIM-PoseNet: Indoor camera localisation using a 3D indoor model and deep learning from synthetic images. ISPRS journal of photogrammetry and remote sensing 150, 245-258 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 224, + 481, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 224, + 481, + 255 + ], + "spans": [ + { + "bbox": [ + 138, + 224, + 481, + 255 + ], + "type": "text", + "content": "2. Acharya, D., Ramezani, M., Khoshelham, K., Winter, S.: Bim-tracker: A model-based visual tracking approach for indoor localisation using a 3d building model. ISPRS Journal of Photogrammetry and Remote Sensing 150, 157-171 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 255, + 481, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 255, + 481, + 288 + ], + "spans": [ + { + "bbox": [ + 138, + 255, + 481, + 288 + ], + "type": "text", + "content": "3. Acharya, D., Tatli, C.J., Khoshelham, K.: Synthetic-real image domain adaptation for indoor camera pose regression using a 3d model. ISPRS Journal of Photogrammetry and Remote Sensing 202, 405-421 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 288, + 481, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 288, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 138, + 288, + 481, + 331 + ], + "type": "text", + "content": "4. Acharya, D., Tennakoon, R., Muthu, S., Khoshelham, K., Hoseinnezhad, R., Bab-Hadiashar, A.: Single-image localisation using 3D models: Combining hierarchical edge maps and semantic segmentation for domain adaptation. Automation in Construction 136, 104-152 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 331, + 481, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 331, + 481, + 353 + ], + "spans": [ + { + "bbox": [ + 138, + 331, + 481, + 353 + ], + "type": "text", + "content": "5. Agarwal, S., Furukawa, Y., Snavely, N., Simon, I., Curless, B., Seitz, S.M., Szeliski, R.: Building rome in a day. Communications of the ACM 54(10), 105-112 (2011)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 353, + 481, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 353, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 138, + 353, + 481, + 373 + ], + "type": "text", + "content": "6. Ammirato, P., Fu, C.Y., Shvets, M., Kosecka, J., Berg, A.C.: Target driven instance detection. arXiv preprint arXiv:1803.04610 (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 373, + 481, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 373, + 481, + 406 + ], + "spans": [ + { + "bbox": [ + 138, + 373, + 481, + 406 + ], + "type": "text", + "content": "7. Arandjelovic, R., Gronat, P., Torii, A., Pajdla, T., Sivic, J.: Netvlad: Cnn architecture for weakly supervised place recognition. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2016)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 406, + 481, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 406, + 481, + 438 + ], + "spans": [ + { + "bbox": [ + 138, + 406, + 481, + 438 + ], + "type": "text", + "content": "8. Cambeiro, A.B., Trzeciakiewicz, M., Hilsmann, A., Eisert, P.: Automatic reconstruction of semantic 3d models from 2d floor plans. In: Proc. Int. Conf. on Machine Vision Applications (MVA). Hamamatsu, Japan (July 2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 438, + 481, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 438, + 481, + 471 + ], + "spans": [ + { + "bbox": [ + 138, + 438, + 481, + 471 + ], + "type": "text", + "content": "9. Chen, C., Wang, R., Vogel, C., Pollefeys, M.: F3Loc: Fusion and Filtering for Floorplan Localization. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2024)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 471, + 481, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 471, + 481, + 514 + ], + "spans": [ + { + "bbox": [ + 138, + 471, + 481, + 514 + ], + "type": "text", + "content": "0. Cruz, S., Hutchcroft, W., Li, Y., Khosravan, N., Boyadzhiev, I., Kang, S.B.: Zillow indoor dataset: Annotated floor plans with 360deg panoramas and 3d room layouts. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2133-2143 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 514, + 481, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 514, + 481, + 547 + ], + "spans": [ + { + "bbox": [ + 138, + 514, + 481, + 547 + ], + "type": "text", + "content": "1. Dellaert, F., Fox, D., Burgard, W., Thrun, S.: Monte carlo localization for mobile robots. In: Proceedings IEEE International Conference on Robotics and Automation. vol. 2, pp. 1322-1328 (1999)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 547, + 481, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 547, + 481, + 579 + ], + "spans": [ + { + "bbox": [ + 138, + 547, + 481, + 579 + ], + "type": "text", + "content": "2. Ding, M., Wang, Z., Sun, J., Shi, J., Luo, P.: CamNet: Coarse-to-fine retrieval for camera re-localization. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2871-2880 (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 579, + 481, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 579, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 138, + 579, + 481, + 611 + ], + "type": "text", + "content": "3. Fernandez-Labrador, C., Facil, J.M., Perez-Yus, A., Demonceaux, C., Civera, J., Guerrero, J.J.: Corners for layout: End-to-end layout recovery from 360 images. IEEE Robotics and Automation Letters 5(2), 1255-1262 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 611, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 481, + 632 + ], + "type": "text", + "content": "4. He, B., Chen, Z., Li, Y.: Calibration method for a central catadioptric-perspective camera system. JOSA A 29(11), 2514-2524 (2012)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "text", + "content": "5. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 770-778 (2016)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 161, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 161, + 91, + 447, + 102 + ], + "type": "text", + "content": "SPVLoc: Semantic Panoramic Viewport Matching for 6D Localization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 117, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 481, + 149 + ], + "type": "text", + "content": "16. Howard-Jenkins, H., Prisacariu, V.A.: LaLaLoc++: Global Floor Plan Comprehension for Layout Localisation in Unvisited Environments. In: Proc. European Conference on Computer Vision (ECCV). pp. 693-709. Springer (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "type": "text", + "content": "17. Howard-Jenkins, H., Ruiz-Sarmiento, J.R., Prisacariu, V.A.: LaLaLoc: Latent Layout Localisation in Dynamic, Unvisited Environments. In: Proc. IEEE/CVF International Conference on Computer Vision (ICCV) (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 183, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 183, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 133, + 183, + 481, + 215 + ], + "type": "text", + "content": "18. Janik, M., Gard, N., Hilsmann, A., Eisert, P.: Zero in on Shape: A Generic 2D-3D Instance Similarity Metric Learned from Synthetic Data. In: Proc. IEEE International Conference on Image Processing (ICIP). pp. 2638-2642. IEEE (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 216, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 216, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 133, + 216, + 481, + 237 + ], + "type": "text", + "content": "19. Karkus, P., Hsu, D., Lee, W.S.: Particle filter networks with application to visual localization. In: Conference on robot learning. pp. 169-178. PMLR (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "type": "text", + "content": "20. Kendall, A., Cipolla, R.: Geometric loss functions for camera pose regression with deep learning. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5974-5983 (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 271, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 481, + 304 + ], + "type": "text", + "content": "21. Kendall, A., Gal, Y., Cipolla, R.: Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7482-7491 (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 481, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 336 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 336 + ], + "type": "text", + "content": "22. Kendall, A., Grimes, M., Cipolla, R.: Posenet: A convolutional network for real-time 6-dof camera relocalization. In: Proc. IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2938-2946 (2015)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "text", + "content": "23. Kim, J., Jeong, J., Kim, Y.M.: Fully geometric panoramic localization. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 20827-20837 (2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 370, + 481, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 370, + 481, + 403 + ], + "spans": [ + { + "bbox": [ + 132, + 370, + 481, + 403 + ], + "type": "text", + "content": "24. Li, T.M., Aittala, M., Durand, F., Lehtinen, J.: Differentiable Monte Carlo Ray Tracing through Edge Sampling. ACM Trans. Graph. (Proc. SIGGRAPH Asia) 37(6), 222:1-222:11 (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "type": "text", + "content": "25. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proc. IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2980-2988 (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "type": "text", + "content": "26. Liu, C., Schwing, A., Kundu, K., Urtasun, R., Fidler, S.: Rent3d: Floor-plan priors for monocular layout estimation. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2015)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "type": "text", + "content": "27. Liu, L., Li, H., Dai, Y.: Efficient global 2d-3d matching for camera localization in a large-scale 3d map. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2372-2381 (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 501, + 481, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 481, + 534 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 481, + 534 + ], + "type": "text", + "content": "28. Lv, X., Zhao, S., Yu, X., Zhao, B.: Residential floor plan recognition and reconstruction. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16717-16726 (June 2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 534, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 555 + ], + "type": "text", + "content": "29. Mercier, J.P., Garon, M., Giguere, P., Lalonde, J.F.: Deep Template-Based Object Instance Detection. In: Proc. WACV. pp. 1507-1516 (January 2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 555, + 481, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 481, + 600 + ], + "type": "text", + "content": "30. Min, Z., Khosravan, N., Bessinger, Z., Narayana, M., Kang, S.B., Dunn, E., Boyadzhiev, I.: Laser: Latent space rendering for 2d visual localization. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11122-11131 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 600, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 481, + 633 + ], + "type": "text", + "content": "31. Naseer, T., Burgard, W.: Deep regression for monocular camera-based 6-dof global localization in outdoor environments. In: Proc. International Conference on Intelligent Robot Systems (IROS). pp. 1525-1530. IEEE (2017)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "type": "text", + "content": "32. Orhan, S., Bastanlar, Y.: Efficient search in a panoramic image database for long-term visual localization. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1727-1734 (2021)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "text", + "content": "N. Gard et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 600 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 160 + ], + "type": "text", + "content": "33. Orhan, S., Guerrero, J.J., Bastanlar, Y.: Semantic Pose Verification for Outdoor Visual Localization With Self-Supervised Contrastive Learning. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops. pp. 3989-3998 (June 2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 161, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 481, + 182 + ], + "type": "text", + "content": "34. Park, S., Kim, H.: 3dplannet: generating 3d models from 2d floor plan images using ensemble methods. Electronics 10(22), 2729 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 226 + ], + "type": "text", + "content": "35. Sarlin, P.E., Unagar, A., Larsson, M., Germain, H., Toft, C., Larsson, V., Pollefeys, M., Lepetit, V., Hammarstrand, L., Kahl, F., et al.: Back to the feature: Learning robust camera localization from pixels to pose. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3247-3257 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 270 + ], + "type": "text", + "content": "36. Shi, Z., Shi, H., Yang, K., Yin, Z., Lin, Y., Wang, K.: PanoVPR: Towards Unified Perspective-to-Equirectangular Visual Place Recognition via Sliding Windows across the Panoramic View. In: 2023 IEEE 26th International Conference on Intelligent Transportation Systems (ITSC). pp. 1333-1340 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "type": "text", + "content": "37. Sturm, P.: Mixing catadioptric and perspective cameras. In: Proceedings of the IEEE Workshop on Omnidirectional Vision 2002. Held in conjunction with ECCV'02. pp. 37-44. IEEE (2002)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 304, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 346 + ], + "type": "text", + "content": "38. Taira, H., Okutomi, M., Sattler, T., Cimpoi, M., Pollefeys, M., Sivic, J., Pajdla, T., Torii, A.: InLoc: Indoor visual localization with dense matching and view synthesis. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7199-7209 (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "text", + "content": "39. Tateno, K., Navab, N., Tombari, F.: Distortion-aware convolutional filters for dense prediction in panoramic images. In: Proc. European Conference on Computer Vision (ECCV). pp. 707-722 (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "text", + "content": "40. Torii, A., Arandjelovic, R., Sivic, J., Okutomi, M., Pajdla, T.: 24/7 place recognition by view synthesis. In: Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1808-1817 (2015)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 413, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 481, + 445 + ], + "type": "text", + "content": "41. Türkoglu, M.Ö., Brachmann, E., Schindler, K., Brostow, G., Monszpart, A.: Visual Camera Re-Localization Using Graph Neural Networks and Relative Pose Supervision. In: Proc. International Conference on 3D Vision (3DV) (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 447, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 481, + 479 + ], + "type": "text", + "content": "42. Wang, S., Fidler, S., Urtasun, R.: Lost shopping! monocular localization in large indoor spaces. In: Proc. IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2695-2703 (2015)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 479, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 501 + ], + "type": "text", + "content": "43. Xia, J., Gong, J.: Precise indoor localization with 3d facility scan data. Computer-Aided Civil and Infrastructure Engineering 37(10), 1243-1259 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 501, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 481, + 533 + ], + "type": "text", + "content": "44. Zamir, A.R., Shah, M.: Accurate image localization based on google maps street view. In: Proc. European Conference on Computer Vision (ECCV). pp. 255-268. Springer, Heraklion, Greece (Sep 2010)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "type": "text", + "content": "45. Zhang, C., Budvytis, I., Liwicki, S., Cipolla, R.: Lifted semantic graph embedding for omnidirectional place recognition. In: Proc. International Conference on 3D Vision (3DV). pp. 1401-1410. IEEE (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 567, + 481, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 600 + ], + "type": "text", + "content": "46. Zheng, J., Zhang, J., Li, J., Tang, R., Gao, S., Zhou, Z.: Structured3D: A large photo-realistic dataset for structured 3d modeling. In: Proc. European Conference on Computer Vision (ECCV). pp. 519-535. Springer, Glasgow, UK (Aug 2020)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 162, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 162, + 91, + 447, + 102 + ], + "type": "text", + "content": "SPVLoc: Semantic Panoramic Viewport Matching for 6D Localization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/SQ-LLaVA_ Self-Questioning for Large Vision-Language Assistant/3909e8d0-9729-43b7-abd9-e1c4bba247cf_content_list.json b/2024/SQ-LLaVA_ Self-Questioning for Large Vision-Language Assistant/3909e8d0-9729-43b7-abd9-e1c4bba247cf_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..325ab5dd7574ba95a9549d753ce2af152c1b2d17 --- /dev/null +++ b/2024/SQ-LLaVA_ Self-Questioning for Large Vision-Language Assistant/3909e8d0-9729-43b7-abd9-e1c4bba247cf_content_list.json @@ -0,0 +1,1832 @@ +[ + { + "type": "text", + "text": "SQ-LLaVA: Self-Questioning for Large Vision-Language Assistant", + "text_level": 1, + "bbox": [ + 282, + 140, + 722, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Guohao Sun1, Can Qin2, Jiamian Wang1, Zeyuan Chen2, Ran Xu2, and Zhiqiang Tao1", + "bbox": [ + 272, + 210, + 730, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Rochester Institute of Technology, Rochester, NY, US $^{2}$ Salesforce AI Research, CA, US", + "bbox": [ + 310, + 253, + 689, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Recent advances in vision-language models have shown notable generalization in broad tasks through visual instruction tuning. However, bridging the gap between the pre-trained vision encoder and the large language models (LLMs) becomes the whole network's bottleneck. To improve cross-modality alignment, existing works usually consider more visual instruction data covering a broader range of vision tasks to fine-tune the model for question-answering, which, however, is costly to obtain and has not thoroughly explored the rich contextual information contained in images. This paper first attempts to harness the overlooked context within visual instruction data, training the model to self-supervised \"learning\" how to ask high-quality questions. In this way, we introduce a novel framework named SQ-LLaVA: Self-Questioning for Large Vision-Language Assistant. SQ-LLaVA exhibits proficiency in generating flexible and meaningful image-related questions while analyzing the visual clue and prior language knowledge, signifying an advanced level of generalized visual understanding. Moreover, fine-tuning SQ-LLaVA on higher-quality instruction data shows a performance improvement compared with traditional visual-instruction tuning methods. This improvement highlights the efficacy of self-questioning techniques in achieving a deeper and more nuanced comprehension of visual content across various contexts. Our code is available at https://github.com/heliossun/SQ-LLaVA.", + "bbox": [ + 259, + 316, + 743, + 608 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Vision-Language Understanding $\\cdot$ Multi-modal LLM $\\cdot$ Instruction Tuning", + "bbox": [ + 259, + 619, + 743, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 672, + 375, + 690 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The recently emerging large vision-language methods, such as large language-and-vision assistant (LLaVA) and its variants [3,5,21,22,54], fine-tune large language models (LLM) [3,54,57] on visual instruction data [18,55,56,58] to realize diverse open-world multimodal understanding, demonstrating a surprising efficacy of visual instruction tuning - the LLM learns to perform complex vision tasks by conditioning on a prompt containing image and text clues. Existing visual instruction data is mainly built upon conversations (e.g., ChatGPT/GPT4-V [1]), consisting of images and multiple question-answer pairs. Building high-quality visual instruction data usually requires images and texts from different tasks to", + "bbox": [ + 212, + 704, + 789, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/25892de293d2dfaee7812478ca6e319cee90bd564ffd006db9b542b01f00afde.jpg", + "image_caption": [ + "Fig. 1: (a) Comparison between visual instruction tuning and visual self-questioning (ours) for vision-language assistant. (b) The proposed SQ-LLaVA achieves state-of-the-art performance on 9 out of 10 tasks compared with other open-ended models." + ], + "image_footnote": [], + "bbox": [ + 233, + 143, + 774, + 318 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "generate diverse questions, such as \"Please provide the bounding box coordinate of the region this sentence describes A dead leaf on the ground\" for the object detection task. Empirically, by increasing the diversity of questions, LLaVA has achieved better performance on GQA and VizWiz tasks (26% and 45% over previous state-of-the-art methods [8]). This evidence strongly suggests the advantage of training models on a broad spectrum and diverse array of tasks for enriching general vision-language understanding.", + "bbox": [ + 212, + 401, + 787, + 507 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "LLaVA [21] model family usually consists of a pre-trained vision encoder (e.g., CLIP-ViT [33]), a large generative language model (LLMs like Vicuna [57], LLaMA [54], QWen [3], etc.), and a vision-to-language projector implemented by a few linear layers. However, the modality gap between the pre-trained vision encoder and the language model restricts both sides' generalization ability and feature representation. To overcome this challenge, various techniques have been proposed to align the vision and language domains, which could be roughly categorized into three groups: 1) build a more robust image feature extractor [8, 54, 58], 2) collect more high-quality training data [5, 8, 21, 58], and 3) fully fine-tune the vision and language models simultaneously during the pre-training stage [5]. While these methods have shown good progress in mitigating the domain gap, they inevitably bring higher computational costs and more expensive data collection and may also require sophisticated manual designs as well as heavy annotating efforts. Moreover, images generally encompass rich information, including color, context, and the relationships between objects, but most existing visual instruction datasets capture only a fraction. We posit that leveraging such under-explored knowledge could significantly aid in vision-language understanding. In this study, we propose a new visual self-questioning approach (Fig. 1a) by training the LLM to ask questions and discover vision clues without collecting extra data from other sources.", + "bbox": [ + 212, + 507, + 787, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unlike existing visual instruction tuning methods that focus solely on answer prediction, the proposed visual self-questioning aims to extract relevant question", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "G. Sun et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9542d76fdec00731d017521f54070ffb255aca4d0dd2b26ef4310d0f163c853f.jpg", + "image_caption": [ + "Fig. 2: Left: Questions spread more samples around a higher mean CLIPScore than answers. Right: Example of highly image-relevant questions within the visual instructional dataset for training a Vision-Language assistant." + ], + "image_footnote": [], + "bbox": [ + 222, + 145, + 367, + 224 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/752976e3d1808ce7bd5632448ae11110285fe1be37b01ecc68f2b49655174b3d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 369, + 146, + 493, + 224 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/dbaf444cb5668a56e37c500aa32db41ae18770b8868f82fd94fba2cd6b5c035c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 148, + 570, + 220 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/afaceb6cfb129dcd4a175fc60429cfcdea1e56df1f3b14e90898b116acf979f1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 573, + 148, + 777, + 220 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "context. As illustrated in Fig. 2 right, the questions could contain more image-related information than the answers, inspiring us that aligning images and related questions may further improve the model's vision-language understanding capacity. Quantitatively, we compute the CLIPScore [33] (a higher value means better visual-text relevance) for all the image-question/-answer pairs on two visual instruction datasets (LLaVA-instruct [21] and ShareGPT4V-instruct [5]). By comparison, the questions' mean CLIPScore $\\mu_q = 0.184$ is larger than the answers' $\\mu_a = 0.183$ on LLaVA-instruct; $\\mu_q = 0.186$ is larger than the answers' $\\mu_a = 0.184$ on ShareGPT4V-instruct. Fig. 2 left shows the CLIPScore distribution, indicating questions have a similar and even better visual relevance than the answers. Therefore, questions in current visual instruction data can be used to fine-tune instructional LLMs owing to their diverse semantic information.", + "bbox": [ + 212, + 311, + 787, + 492 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This work introduces a self-questioning LLaVA, namely SQ-LLaVA, to fully utilize questions within the instruction data as an additional learning resource for training instructional LLMs and empowering the model's curiosity (questioning ability). To efficiently align the vision and language domains, we apply LoRAs [12] to optimize both the vision encoder and the instructional LLM within the SQ-LLaVA. Plus, we develop a prototype extractor to enhance visual representation by leveraging learned clusters with meaningful semantic information to improve vision-language alignment further. Extensive experiments demonstrate that SQ-LLaVA surpasses existing visual instruction tuning methods in general vision understanding (see Fig. 1b). We summarize the contributions as follows.", + "bbox": [ + 212, + 494, + 787, + 645 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel training technique, visual self-questioning for vision-language assistants (SQ-LLaVA), by leveraging highly relevant question contexts in instruction data. This SQ learning task promotes instructional LLMs to understand the relationship between images and questions, enhancing vision-language alignment without needing new data collection.", + "- We design and develop a lightweight tuning architecture for SQ-LLaVA, consisting of ViT-LoRA, LLM-LoRA, and a prototype extractor. The prototype extractor enhances vision embeddings, while ViT-LoRA and LLM-LoRA efficiently align vision and language domains during training.", + "- Extensive experimental results show that the proposed SQ-LLaVA leads to better performance in several tasks, including visual question-answering, visual instruction benchmarks, and zero-shot image captioning." + ], + "bbox": [ + 223, + 657, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "SQ-LLaVA: Self-Questioning for Large Vision-Language Assistant", + "bbox": [ + 292, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 143, + 388, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.1 Instruction Tuning", + "text_level": 1, + "bbox": [ + 215, + 172, + 418, + 188 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Instruction tuning emerged as a pivotal methodology within the realm of natural language processing (NLP), facilitating Large Language Models (LLMs) such as GPT-3 [4], PaLM [7], and LLaMA [42] to interpret and execute human language instructions across a spectrum of NLP tasks. This approach diverges from traditional fine-tuning mechanisms by incorporating a specialized data structure, termed instruction-following data [50], which is instrumental in the fine-tuning process of LLMs. Generally, there are two main categories regarding instruction tuning methods - 1) closed-domain and 2) open-domain. The closed-domain instruction tuning [35,50,51] studies engaged LLMs with a comprehensive assortment of publicly accessible datasets and subsequently assessed their performance across diverse NLP tasks [44]. The empirical evidence from these inquiries consistently indicated that integrating varied NLP task instructions significantly augments the LLMs' efficacy in navigating novel tasks. Nonetheless, LLMs calibrated with such closed-form instructions exhibited limitations in real-world user scenarios, prompting the development of an alternative approach. To address these constraints, the concept of open-domain instruction tuning [30, 49] is conceived. OpenAI pioneered this approach by employing human annotators to compile a corpus of real-world question-answer datasets. These datasets form the foundation for training a reward model through reinforcement learning methodologies. The trained reward model then functions as a supervisory mechanism for further training instruction-oriented language models, such as InstructGPT [31] and Vicuna [57]. This innovation marks a significant advancement in the field, aiming to bridge the gap between LLM performance and real-world applicability by leveraging instruction data derived from authentic user interactions.", + "bbox": [ + 212, + 195, + 789, + 559 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Large Vision Language Models", + "text_level": 1, + "bbox": [ + 215, + 575, + 517, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As the field of LLMs and instruction tuning undergoes rapid advancements, the academic research community is progressively focusing on integrating visual information into LLMs to facilitate visual instruction tuning. This emerging area of research has witnessed the development of various methodologies [3,5,21,22,54] based on foundational vision-language models [19,33,40,47] and diverse LLM architectures [3,43,57]. In particular, LLaVA [22] pioneers the integration of an LLM with a CLIP vision encoder to construct a vision language model, demonstrating remarkable capabilities in image-text dialogue tasks through pretraining alignment strategies and targeted instruction tuning. Subsequent investigations have sought to refine visual instruction tuning by enhancing the quality and variety of the datasets used during the pre-training and fine-tuning phases. Building upon these advancements, recent studies like LLaVA-v1.5 [21] and ShareGPT4V [5] have achieved notable success in general vision-language comprehension, showcasing their ability to undertake complex question-answering tasks. This progression underscores the importance of sophisticated data handling and model tuning strategies in developing effective vision-language models.", + "bbox": [ + 212, + 598, + 789, + 843 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "G. Sun et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b7b262649a963a2c222fa00777c5e37a3722f0df3764ae09d3ef66c868ccb104.jpg", + "image_caption": [ + "Fig. 3: Model architecture of SQ-LLaVA. We propose prototype extractor to extract visual clustering information to enhance the visual embedding encoded by the visual encoder. SQ-LLaVA defines a new token [vusr] as specific instruction for LLM to perform visual self-questioning. Besides question answering, SQ-LLaVA treats questioning as another training objective." + ], + "image_footnote": [], + "bbox": [ + 215, + 142, + 782, + 300 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 412, + 328, + 429 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Architecture Overview", + "text_level": 1, + "bbox": [ + 215, + 445, + 450, + 459 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The proposed SQ-LLaVA model (see Fig. 3) consists of four main components: 1) A pre-trained vision encoder CLIP-ViT [33] that extracts a sequence embedding of image tokens $Z_{v}$ for an input image $X_{v}$ ; 2) A prototype extractor $\\phi(\\cdot)$ learning visual clusters to enhance the original image tokens; 3) A trainable projection block $W(\\cdot)$ with two linear layers to map the enhanced image tokens to the language domain tokens $H_{v}$ , handling the dimension misalignment between the vision and language domain; and 4) Our LLM backbone $f(\\cdot)$ implemented by the pre-trained Vicuna [57] to predict the next token upon the previous embedding sequence. Given the input question $X_{q}$ and answer $X_{a}$ , a word embedding matrix is used to map them to contextual embeddings $H_{q}$ and $H_{a}$ , and the distribution over $H_{a}^{(i+1)}$ can be obtained following the auto-regressive model as:", + "bbox": [ + 212, + 470, + 787, + 642 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(H _ {a} ^ {(i + 1)} \\mid H _ {v}, H _ {q}, H _ {a} ^ {(1: i)}\\right) = \\sigma \\left(f \\left(H _ {v}, H _ {q}, H _ {a} ^ {(1: i)}\\right)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 318, + 650, + 787, + 671 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\theta$ represents all the trainable parameters in our model, $\\sigma(\\cdot)$ is a softmax function, and $f(\\cdot)$ outputs the last token embedding of the whole sequence. We denote $p_{\\theta}$ as the prediction probability for the anticipated answer token $H_{a}^{(i+1)}$ at the position $i+1$ , conditioning on the input image token embedding $H_{v}$ , the question token embedding $H_{q}$ , and the previous answer token embeddings $H_{a}^{(1:i)}$ . As shown in Eq. (1), the proposed SQ-LLaVA applies the language model $f(\\cdot)$ to model $p_{\\theta}$ given by $H_{v}$ , $H_{q}$ , and $H_{a}^{(1:i)}$ . Existing visual instruction tuning methods [5,21] are only able to predict $H_{a}$ , yet cannot fully exploit the rich semantic clues within $H_{q}$ . In this study, we propose visual self-questioning instructions to guide LLMs in capturing the visual content within questions.", + "bbox": [ + 212, + 680, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "SQ-LLaVA: Self-Questioning for Large Vision-Language Assistant", + "bbox": [ + 292, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "System-message", + "bbox": [ + 246, + 148, + 349, + 161 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} _ {c} ^ {(1)}: [ \\mathrm {u s r} ]: \\mathbf {X} _ {v} \\backslash \\mathrm {n} \\mathbf {X} _ {q} ^ {(1)} [ \\mathrm {a s w r} ]: \\mathbf {X} _ {a} ^ {(1)} < o ^ {d} > \\quad \\mathbf {X} _ {c} ^ {(2)}: [ \\mathrm {v u s r} ]: \\mathbf {X} _ {q} ^ {(2)} < o ^ {d} > [ \\mathrm {a s w r} ]: \\mathbf {X} _ {a} ^ {(2)} < o ^ {d} >\n$$\n", + "text_format": "latex", + "bbox": [ + 246, + 162, + 728, + 184 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} _ {c} ^ {(3)}: [ \\text {u s r} ]: \\mathbf {X} _ {q} ^ {(3)} [ \\text {a s w r} ]: \\mathbf {X} _ {a} ^ {(3)} < _ {O} d > \\quad \\dots \\quad \\mathbf {X} _ {c} ^ {(P)}: [ \\text {v u s r} ]: \\mathbf {X} _ {q} ^ {(P)} < _ {O} d > [ \\text {a s w r} ]: \\mathbf {X} _ {a} ^ {(P)} < _ {O} d >\n$$\n", + "text_format": "latex", + "bbox": [ + 246, + 186, + 728, + 207 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 4: The input sequence used to train SQ-LLaVA. Our model is trained to predict question, answer, and where to stop. We use tokens to represent learnable tokens, where $X_{q}$ is the question, $X_{a}$ is the answer, and $< o^{d}>$ is the delimiter token. In SQ-LLaVA, the System-message = \"The assistant gives helpful, detailed, and polite answers to the user's questions. Also, the assistant is a curious virtual user can ask complex questions that are relevant to the content in the image.\"", + "bbox": [ + 212, + 223, + 789, + 306 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Visual Self-questioning Instruction", + "text_level": 1, + "bbox": [ + 214, + 339, + 547, + 356 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In broad real-world scenarios, proactively asking a question requires more understanding and background knowledge than answering [41]. Similarly, this work proposes visual self-questioning to encourage the LLM to discover deeper vision-language alignment and improve the overall instruction-following performance. Particularly, SQ-LLaVA treats questioning as a new learning objective beyond answering, which, to the best of our knowledge, is the first practice in the field of visual instruction tuning. While the current vision-language model can ask questions [1], such skill is still learned from question-answering through instruction tuning. However, our proposed method shows that the decoder-based LLM has the potential to learn more skills, such as how to ask questions spontaneously when a unique instruction token is given (e.g., we define [vusi] in our work). Furthermore, visual self-questioning can potentially improve general vision-language understanding. To be specific, as shown in Fig. 1, there are a certain amount of questions containing more meaningful image-related information than answers in the existing visual instruction data [5, 21]. Thus, we hypothesize that the vision-language understanding can be improved once the LLM learns how to predict relevant questions about a given image.", + "bbox": [ + 212, + 369, + 789, + 627 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Self-Questioning Prompt Design. We provide ground-truth content for the visual self-questioning, restricting SQ-LLaVA from asking image-related questions. To this end, we leverage questions as another learnable resource and follow the regular auto-regressive training objective. As shown in Fig. 4, the training data for SQ-LLaVA is designed in a format with a pre-defined template. To be specific, the system message as a fixed prompt is added at the beginning of each instruction data, indicating a general job description (e.g., gives helpful answers and asks complex questions) for the LLM. Existing visual instruction tuning methods utilize the unique tokens ([usr], [aswr]) to give the LLM a particular instruction (e.g., questioning understanding, answer prediction, etc.) and apply the delimiter token $< \\sigma^d>$ to mark the ending position. In this work, we propose a new special token [vusr], indicating a specific instruction - asking questions. Combining with the delimiter, we can construct instructions for self-questioning as a new training task.", + "bbox": [ + 212, + 628, + 789, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "G. Sun et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Each sample of current visual instruction data consists of one image $X_{v}$ and $P$ question-answer pairs $(X_{q}^{(1)},X_{a}^{(1)},\\dots ,X_{q}^{(P)},X_{a}^{(P)})$ . We collect one question $X_{q}^{(j)}$ and its answer $X_{a}^{(j)}$ with special tokens to construct the $j^{th}$ turn conversation as", + "bbox": [ + 212, + 146, + 782, + 196 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nX _ {c} ^ {j} = \\left\\{ \\begin{array}{l l} \\left([ \\mathrm {u s r} ], X _ {q} ^ {(j)}, [ \\mathrm {a s w r} ], X _ {a} ^ {(j)}\\right) & j = 1 \\text {o r} j > 1, R < \\delta \\\\ \\left([ \\mathrm {v u s r} ], X _ {q} ^ {(j)}, [ \\mathrm {a s w r} ], X _ {a} ^ {(j)}\\right) & j > 1, R > \\delta \\end{array} , \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 279, + 205, + 785, + 247 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $R \\in [0,1]$ is a random number and $\\delta = 0.5$ is a threshold that sets the proportion of self-questioning pairs in the conversations. Finally, the full-text input $X_{c}$ will be mapped to textual embedding $H_{c}$ through word embedding.", + "bbox": [ + 212, + 258, + 782, + 303 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "SQ-LLaVA performs zero-shot questioning without any in-context knowledge or human language instruction since the only instructional prompt is a unique token [vusr]. After visual self-questioning on instruction data with various question formats $X_{q}$ , the questions sampled by SQ-LLaVA are more diversified than GPT4-V (as shown in Fig. 5), since the LLM has learned alignment between image and questions, which is different from previous works [1, 48]. Specifically, previous general-purpose vision language models such as GPT4-V [1] can generate questions based on a given image, but it requires explicit language instruction such as \"Ask complex questions that are relevant to the content in the image\". Accordingly, the quality of generated questions highly relies on prompt engineering. Also, self-instruct [48] utilizes in-context learning to prompt the LLM to ask specific questions.", + "bbox": [ + 212, + 304, + 784, + 484 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Enhanced Visual Representation", + "text_level": 1, + "bbox": [ + 214, + 508, + 529, + 523 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Unlike previous visual instruction tuning methods, SQ-LLaVA jointly benefits from visual self-questioning and question-answering. For better visual self-questioning, we develop a prototype extractor that recognizes and groups similar patterns of visual information from the latent space. Our primary goal is to enhance visual representation through prototype learning.", + "bbox": [ + 212, + 534, + 782, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Specifically, we utilize clustering to extract centroid representations of image tokens $Z_{v}$ , where each cluster center is treated as a prototype, which, in return, will be distributed to each of the original image token embeddings. Our proposed prototype extractor $\\phi(\\cdot)$ is a lightweight design involving two parts: 1) cluster center optimization and 2) prototype information distribution. Following [14, 45], we randomly initialize $K = 256$ cluster centers $C$ and deploy the iterative Expectation-Maximization (EM) clustering process to capture representative semantics in the latent space by", + "bbox": [ + 212, + 609, + 784, + 729 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\operatorname {E} \\text {- s t e p}: \\mathcal {M} ^ {(t)} = \\sigma \\left(q \\left(C ^ {(t)}\\right) * k \\left(Z _ {v}\\right) ^ {\\top}\\right), \\\\ \\text {M - s t e p}: C ^ {(t + 1)} = \\mathcal {M} ^ {(t)} * v \\left(Z _ {v}\\right), \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 366, + 741, + 785, + 781 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathcal{M}^{(t)}\\in [0,1]$ denotes a soft cluster assignment map at the $t^{th}$ iteration $\\sigma (\\cdot)$ is a softmax function, and $t\\in \\{1,\\dots ,T\\}$ indexes the iteration of EM step with $T = 2$ in this work. Three trainable linear layers $q,k$ , and $v$ are used in (3.3)", + "bbox": [ + 212, + 792, + 782, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "SQ-LLaVA: Self-Questioning for Large Vision-Language Assistant", + "bbox": [ + 292, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $q(\\cdot)$ projects the prototype $C$ to a query vector, and $k(\\cdot)$ and $v(\\cdot)$ project $H_{v}$ into key and value vectors, followed by a normalization layer, respectively. The prototype extractor iteratively updates cluster map $\\mathcal{M}$ and centers $C$ .", + "bbox": [ + 212, + 146, + 787, + 191 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "After the prototype extraction, we train a linear layer $z(\\cdot)$ to adaptively map the visual cluster information to the raw image embedding $Z_{v}$ . For the $i^{th}$ token embedding $Z_{v}^{(i)}$ , we update it as", + "bbox": [ + 212, + 191, + 787, + 239 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nZ _ {v} ^ {(i)} = Z _ {v} ^ {(i)} + z \\left(\\frac {1}{K} \\sum_ {j = 1} ^ {K} S _ {c} \\left(C _ {j}, Z _ {v} ^ {(i)}\\right) \\times C _ {j}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 250, + 787, + 277 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $S_{c}(\\cdot, \\cdot)$ is a normalized cosine similarity function. The weighted sum over prototypes in Eq. (4) emerges as an indispensable step for contextual understanding from image tokens, recognizing and grouping similar patterns and semantics. It clusters image tokens as prototypes that display homogeneity in semantics, such as \"grass\" and \"dog\". The prototypes can describe the intrinsic semantic meanings by aggregating entities that exhibit shared attributes. Finally, we map the image sequence embedding $Z_{v}$ to language domain $H_{v}$ with a two-layer linear projector $W(\\cdot)$ .", + "bbox": [ + 212, + 285, + 787, + 407 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Model Training", + "text_level": 1, + "bbox": [ + 214, + 426, + 392, + 444 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Stage1: Pre-training for Vision-Language Alignment. Unlike text-only LLMs, the vision-language model also fine-tunes LLM using image tokens as input (see Fig. 3). Therefore, the pre-training stage aims to optimize the LLM by explicitly executing the visual instruction. The proposed SQ-LLaVA adopts Vicuna [57] as its instruction LLM, pre-training on massive text corpora to predict the next text token given the previous context, not only containing text but also visual instructions. To achieve this, we organize the pre-training data as $D_{PT} = \\{[X_v^{(1)}, X_a^{(1)}], \\dots, [X_v^{(N)}, X_a^{(N)}]\\}$ , where $N$ is the total number of training samples, and each sample has an image and its related descriptions. Each image and text input pair will be mapped to sequence embeddings ( $H_v$ and $H_a$ ) as elaborated in Section 3.1. During pre-training, we freeze the vision encoder and LLM and mainly train the prototype extractor $\\phi$ and the vision-to-language projector $W$ . The pre-training goal is to maximize the probability of the predicted image description $H_a$ given an image $H_v$ . When training a visual instructional LLM, we follow the negative log-likelihood objective function as", + "bbox": [ + 212, + 450, + 787, + 680 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\sum_ {v, a \\in D _ {P T}} - \\log p _ {\\theta} (H _ {a} \\mid H _ {v}) = \\sum_ {v, a \\in D _ {P T}} \\sum_ {i = 1} ^ {L} - \\log p _ {\\theta} (H _ {a} ^ {(i + 1)} \\mid H _ {v}, H _ {a} ^ {(1: i)}), \\quad (5)\n$$\n", + "text_format": "latex", + "bbox": [ + 261, + 689, + 787, + 732 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $L$ denotes the sequence length of answer tokens in $H_{a}$ , $\\theta$ is the total trainable parameter of $\\phi$ and $W$ , $p(H_{a} \\mid H_{v})$ can be computed by Eq. (1), and $H_{a}^{(1:i)}$ represents all the answer tokens before the current prediction $H_{a}^{(i+1)}$ .", + "bbox": [ + 214, + 739, + 787, + 789 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Stage2: Fine-tuning. Existing methods, such as LLaVA [5,22], mainly update the vision-to-language projector (usually a couple of linear layers) and the", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "G. Sun et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "language model during fine-tuning. Nevertheless, the projector might be too weak to capture the relationship between the image and the questions. Following the previous multi-modal understanding method [33], we unfreeze the vision encoder and LLM during fine-tuning for a joint optimization further to eliminate the gap between the vision and language domain.", + "bbox": [ + 212, + 146, + 782, + 222 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To mitigate the heavy computational overhead, we take advantage of LoRA [12] as a lightweight tuning option that can achieve similar (even better) performance to fully fine-tuning when training large models on a relatively small dataset. We add LoRA in both the vision encoder and LLM. Thus, the learnable parameters $\\theta$ of the proposed SQ-LLaVA during fine-tuning represent a combination of all the parameters of LLM-LoRA, ViT-LoRA, prototype extractor $\\phi$ , and the vision-to-language projector $W$ . Given the instruction tuning data $D_{IT} = \\{[X_v^{(1)}, X_c^{(1)}], \\dots, [X_v^{(N)}, X_c^{(N)}]\\}$ , we take the conversational data $X_c$ and the image $X_v$ as input, mapping them to sequence embedding $(H_c$ and $H_v)$ as elaborated in Section 3.1, and minimize the negative log-likelihood loss for the self-questioning and answering tasks as follows", + "bbox": [ + 212, + 220, + 787, + 390 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\text {S e l f - q u e s t i o n i n g}: \\sum_ {v, c \\in D _ {I T}} - \\log p _ {\\theta} \\left(H _ {q} ^ {(j + 1)} \\mid H _ {v}, H _ {c} ^ {(1: j)}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 274, + 402, + 787, + 436 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\text {A n s w e r i n g}: \\sum_ {v, c \\in D _ {I T}} - \\log p _ {\\theta} \\left(H _ {a} ^ {(j + 1)} \\mid H _ {v}, H _ {c} ^ {(1: j)}, H _ {q} ^ {(j + 1)}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 313, + 443, + 785, + 476 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $j \\in \\{1, \\dots, P\\}$ , indicating the index of question or answer within the conversational data $X_{c}^{(*)}$ . Notably, previous works [5,21,58] only involve answering tasks, but we introduce visual self-questioning as an additional training task for visual instruction tuning. Eventually, SQ-LLaVA, as a vision-language assistant, not only executes human instructions by optimizing the objective function in Eq. (7) but can raise questions out of the given image after optimizing the Eq. (6). This capability potentially yields more diverse question-answer guidance and enhances multi-modal understanding.", + "bbox": [ + 212, + 487, + 787, + 609 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 633, + 375, + 648 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Experimental Setting", + "text_level": 1, + "bbox": [ + 214, + 664, + 439, + 680 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Dataset. Our work uses the open-source visual instruction dataset provided by LLaVA [21] and ShareGPT4V [5] for training. Each dataset has large-scale image-text paired data for pre-training and instruction-following data for finetuning. Specifically, the instruction data proposed by LLaVA is a comprehensive mixture of COCO, GQA [13], OCR-VQA [27], TextVQA [39], VisualGenome [17], RefCOCO [15], and image from InstructBLIP [8] involves multiple reasoning, spatial understanding, multi-step inference, optical character recognition, and grounding of visual concepts to language. The ShareGPT4V dataset consists of the same mixture image as LLaVA but enrolls more images from other datasets such as SAM [16] and Data [29,36].", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "SQ-LLaVA: Self-Questioning for Large Vision-Language Assistant", + "bbox": [ + 292, + 114, + 730, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We evaluate our model on general vision-language understanding tasks. Specifically, we use ten visual oriented question answering benchmarks: $\\mathrm{VQA}^{v2}$ [10]; GQA [13]; VizWiz [11]; $\\mathrm{SQA}^I$ [25]: ScienceQA-IMG; $\\mathrm{VQA}^T$ : TextVQA [39]; POPE [20]; MM-Vet [53]; $\\mathrm{LLaVA}^W$ [22]: LLaVA (in the wild); MMB: MMBBenchmark [23]; $\\mathrm{MMB}^{CN}$ : MMBench-Chinese [23]. We report the prediction accuracy on all the benchmarks. We also evaluate our model for visual information discovery through captioning. To be specific, we employ testing images from four datasets, i.e., COCO [6], Flickr30K [52], Conceptual [37], and Nocaps [2]. Following [9], we evaluate the proposed method with regular image captioning metrics, e.g., BLEU [32] and CIDEr [46]. For the open-world methods LLaVA-v1.5 [21], ShareGPT4V [5] and the proposed SQ-LLaVA, we utilize greedy search for caption generation with a prompt of \"Provide a brief description of the given image, your answer should be in one sentence.\" The generated caption is used to evaluate the performance of image captioning.", + "bbox": [ + 212, + 146, + 787, + 358 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Implementation", + "text_level": 1, + "bbox": [ + 215, + 378, + 392, + 393 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We adopt Vicuna [57] as the pre-trained language generative model and CLIP-ViT [33] as the vision encoder. We pre-train the prototype extractor and the vision-to-language projector using AdamW [24] optimizer with a learning rate of $2 \\times 10^{-3}$ and a constant scheduler for one epoch. Following previous work [21], we keep the global batch size as 256 for pre-training and 128 for fine-tuning. During fine-tuning, we insert LoRA [12] with rank = 128 and $\\alpha = 256$ into the language model (LLM-LoRA) and LoRA with rank = 32 and $\\alpha = 64$ into the vision encoder (ViT-LoRA). We optimize LoRA modules, the prototype extractor, and vision-to-language projector for one epoch. We set the learning rate to $2 \\times 10^{-4}$ for LoRA, and $2 \\times 10^{-5}$ for the other layers. All the weights of the pre-trained language model and vision encoder remain fixed during fine-tuning.", + "bbox": [ + 212, + 402, + 787, + 568 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Zero-shot Multilingual Capability", + "text_level": 1, + "bbox": [ + 215, + 589, + 537, + 606 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We evaluate SQ-LLaVA on ten benchmarks, covering a range of academic Visual Question Answering (VQA) tasks and recent instruction tuning tasks designed for large vision language models. The academic VQA includes VQA-v2 [10] and VizWiz [11]. GQA [13] is a fine-grained real-world visual reasoning and question-answering benchmark. ScienceQA [25] is a benchmark with rich subjects (natural science, language science, and social science). TextVQA [39] requires the model to recognize the texts in the image. LLaVA (in the wild) and MM-Vet [53] use GPT4 to assess the capability of the models for testing. With manually designed questions, MM-Bench and MMBench-CN [23] evaluate the model's vision-related reasoning and perception for English and Chinese, respectively. POPE [20] is a benchmark for evaluating object hallucination [34].", + "bbox": [ + 212, + 613, + 787, + 779 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In Table 1, we quantitatively compare between SQ-LLaVA and existing models. SQ-LLaVA-7B and SQ-LLaVA-13B trained on two instruction datasets [5,21] outperform previous methods in six out of ten visual instruction tuning tasks. To be specific, SQ-LLaVA-7B achieves $17.2\\%$ improvement over LLaVA-v1.5-7B", + "bbox": [ + 212, + 779, + 787, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "G. Sun et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/2b2598b2b7c3ec6ac7e34f5ef69676fe089dfb07523f27f0086b623aafbc1b52.jpg", + "table_caption": [ + "Table 1: Comparison with state-of-the-art methods on ten benchmarks. After training on the same instruction data, SQ-LLaVA train on [21] and SQ-LLaVA* trained on [5] surpass their baseline model LLaVA-v1.5 and ShareGPT4V on 9 out of 10 and 6 out of 10 benchmarks in the 7B scale, and 8 out of 10 and 6 out of 10 in the 13B scale. The best results are bold and the second-best results are underlined." + ], + "table_footnote": [], + "table_body": "
LLMModelVQA v2GQAVizWizSQA1VQA TPOPEMM-VetLLaVA WMMBMMB CN
7BInstructBLIP [8]-49.234.560.550.1-26.260.936.023.7
Qwen-VL [3]78.859.335.267.163.8---38.27.4
Qwen-VL-chat [3]78.257.538.968.261.5---60.656.7
LLaVA-v1.5 [21]78.562.050.066.858.285.930.563.464.358.3
ShareGPT4V [5]80.663.357.268.460.486.837.672.668.862.2
SQ-LLaVA79.262.854.068.958.687.732.566.366.258.1
SQ-LLaVA*80.363.755.370.560.587.237.674.366.660.0
13BInstructBLIP [8]-49.533.463.150.778.925.658.2--
LLaVA-v1.5 [21]80.063.353.671.658.285.935.470.767.763.6
ShareGPT4V [5]81.064.855.671.262.2-43.179.968.563.7
SQ-LLaVA80.163.654.669.860.287.735.574.668.762.0
SQ-LLaVA*81.365.058.271.561.987.439.780.768.562.5
", + "bbox": [ + 222, + 226, + 784, + 388 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "on the LLaVA (in the wild) benchmark, indicating the superior capabilities of our model in tasks such as detailed description and complex reasoning. Also, SQ-LLavA-7B improves over previous methods on ScienceQA, indicating that our model excels in understanding and reasoning over scientific content and can effectively handle multi-modal information. The improvement in ScienceQA suggests strong capabilities in multi-hop reasoning, comprehension of complex scientific concepts, and the ability to utilize context and explanations to derive correct answers. SQ-LLaVA-7B has a steady improvement over LLaVA-v1.5-7B and ShareGPT4V-7B on the POPE benchmark, and the $2\\%$ and $1\\%$ improvement indicates that our proposed method has better reliability and trustworthiness since POPE is a task designed to evaluate the phenomenon of object hallucination [20, 34]. In the bottom section of Table 1, the proposed SQ-LLaVA-13B surpasses previous works in six out of ten benchmarks, indicating the scalability of our method on larger LLM. Notably, the performance inconsistency on some datasets might be due to the unsupervised prototype extractor (lacking pixel-wise guidance) in our model. To mitigate this issue, we could leverage pseudo object masks (e.g., given by the pre-trained segment anything) in learning prototypes. Despite this limitation, all the improvements are achieved with significantly fewer trainable parameters compared to other methods [3, 5, 21].", + "bbox": [ + 212, + 417, + 787, + 705 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.4 Visual Information Discovery", + "text_level": 1, + "bbox": [ + 214, + 724, + 503, + 739 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In this experiment, we showcase the diversity and reliability of the proposed SQ-LLaVA through various qualitative applications, including detailed image description, visual information summary, and visual self-questioning. We also present quantitative results on the task of image captioning.", + "bbox": [ + 212, + 750, + 787, + 809 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Abilities of SQ-LLaVA Through Qualitative Samples. SQ-LLaVA exhibits numerous advanced capabilities compared to traditional vision-language models.", + "bbox": [ + 212, + 809, + 787, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "SQ-LLaVA: Self-Questioning for Large Vision-Language Assistant", + "bbox": [ + 292, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/136c42a072a95726e4d9809d77f134ccfcb9ee16686c2c4f4b300baa8370bb1d.jpg", + "image_caption": [ + "GPT4-V Generated" + ], + "image_footnote": [], + "bbox": [ + 225, + 145, + 316, + 194 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "$Q_{1}$ : What makes this particular toothbrush stand out among others in the image? $Q_{2}$ : What color is the background?", + "Q2: What color is the background? \nQ3: What is this product?" + ], + "bbox": [ + 320, + 159, + 480, + 195 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "SQ-LLaVA Visual Self-Questioning", + "text_level": 1, + "bbox": [ + 284, + 198, + 424, + 205 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "$SQ_{1}$ : What are some of the key aspects that this image focuses on?", + "$S Q_{2}$ : What is the primary color of the toothbrush?", + "$S Q_{3}$ : What is the expression of the face drawn on the toothbrush?" + ], + "bbox": [ + 228, + 208, + 485, + 232 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/ece03bb4b799ad9b0528600b4496451126f4c03af0cddf96704da72e8299f59b.jpg", + "image_caption": [ + "GPT4-V Generated" + ], + "image_footnote": [], + "bbox": [ + 227, + 250, + 325, + 301 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "$Q_{1}$ : What is the setting of the image? $Q_{2}$ : What are some activities that people are engaging in within the image? $Q_{3}$ : What are some benefits of spending time outdoors in a beachside setting, especially on a sunny day?", + "bbox": [ + 326, + 263, + 480, + 313 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "SQ-LLaVA Visual Self-Questioning", + "text_level": 1, + "bbox": [ + 282, + 315, + 424, + 324 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "$S_{Q_1}$ : What is the hair color of the man that is looking up at the ocean?", + "$SQ_{2}$ : What color is the man's shirt?", + "$SQ_{3}$ : In which part is the bench, the top or the bottom?" + ], + "bbox": [ + 228, + 325, + 464, + 359 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/70b7a2b0d2bb65837892e832bb74fd4f3e7053a174d41ef5baa6557c83736bc7.jpg", + "image_caption": [ + "GPT4-V Generated" + ], + "image_footnote": [], + "bbox": [ + 500, + 146, + 591, + 193 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "$Q_{1}$ : What type of environment can the sheep be seen in?", + "bbox": [ + 591, + 157, + 763, + 172 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "$Q_{2}$ : What are the sheep doing in the image? $Q_{3}$ How do sheep contribute to the ecosystem of a meadow or field?", + "bbox": [ + 593, + 174, + 761, + 196 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "SQ-LLaVA Visual Self-Questioning", + "text_level": 1, + "bbox": [ + 560, + 199, + 702, + 208 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "$SQ_{1}$ : In the vicinity of this group of sheep, how many horses are there? a) Five b) Four c) One d) Zero", + "$S_{Q2}$ : In this image, what kind of animals can you see in the green field? $S_{Q3}$ : Are there any cows in the picture?" + ], + "bbox": [ + 500, + 210, + 771, + 244 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/de62dffe85e775b4f6f0a44f297a821f4725bd1c697cceac3544561065315f45.jpg", + "image_caption": [ + "GPT4-V Generated", + "Fig.5: Visual self-questioning of SQ-LLaVA-7B. Comparing to the question data provided by GPT4-V (data collected by LLaVA-v1.5 [21]), SQ-LLaVA can generate questions with higher diversity i.e. multiple choice and tricky questions." + ], + "image_footnote": [], + "bbox": [ + 500, + 252, + 594, + 301 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "SQ-LLaVA Visual Self-Questioning", + "text_level": 1, + "bbox": [ + 560, + 315, + 702, + 324 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "$S_{Q1}$ : The two small children are sitting on a tricycle. Which place is it? Town or village?", + "$SQ_{2}$ : What can you tell me about the child's age?", + "$S_{Q3}$ : What does the boy's action of swinging a bat suggest about his play or sport?" + ], + "bbox": [ + 500, + 325, + 754, + 367 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/7767a9ca24062f40c2beb45e431a9f380e948cedb5236f6a433b1f8111f07b48.jpg", + "table_caption": [ + "Table 2: Comparison of image captioning on four datasets. SQ-LLaVA is trained on data collected by [21]. SQ-LLaVA* is trained on data collected by [5]." + ], + "table_footnote": [ + "(a) Zero-shot image captioning" + ], + "table_body": "
ModelFlickr30kNocapsoutConceptual
B@4CIDErB@4CIDErB@4CIDEr
ClipCap [28]17.2141.6520.3251.741.4723.74
DiscriTune [9]18.4844.7824.1057.061.7128.01
LLaVA-v1.5 [21]28.6781.2735.78103.562.7939.20
ShareGPT4V [5]31.0086.1737.19107.452.7837.86
SQ-LLaVA29.8883.5136.21105.422.9039.49
SQ-LLaVA*31.4983.1437.20107.422.9141.24
", + "bbox": [ + 220, + 479, + 534, + 588 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/38aaa20ea581ef709118c303c92d0712b61998d81e6b4499b3e234b21d51b191.jpg", + "table_caption": [], + "table_footnote": [ + "(b) Image captioning on COCO" + ], + "table_body": "
ModelZero-shotFine-tune
B@4CIDErB@4CIDEr
ClipCap8.5037.0332.60108.55
DiscriTune13.9953.2032.31105.40
eP-ALM [38]29.4797.2233.35111.63
MAPL [26]12.3054.3036.45125.20
LLaVA-v1.529.96111.46--
SQ-LLaVA29.85110.7740.76136.78
", + "bbox": [ + 537, + 479, + 769, + 587 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Notably, SQ-LLaVA effectively mitigates object hallucination [20, 34], resulting in predictions that are more trustworthy. Additionally, SQ-LLaVA is capable of generating diverse and meaningful questions about a given image without requiring human textual instructions. As illustrated in Fig. 5, the generated questions include multiple-choice, reasoning, binary choice questions, and the content within each question, all of which are either highly correlated or tricky to the given image, such as \"What is the hair color of the man that is looking up at the ocean?\" from the third image and \"Are there cows in the picture?\" from the second image. These examples highlight SQ-LLaVA's ability to produce questions with high diversity and quality.", + "bbox": [ + 212, + 642, + 787, + 794 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The detailed image description example in Fig. 6 demonstrates that SQ-LLaVA can generate visual information with concrete concepts within the image, such as \"Hyundai\", \"small white dog\", \"Nikky Stephen\". In contrast, LLaVA-v1.5 can", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "G. Sun et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/507a676ade812af4fb871183548354e4209fede95d1848284cab55c6f5f0276f.jpg", + "image_caption": [ + "Detailed Image Description", + "Fig. 6: A qualitative evaluation of detailed image descriptions from three models. We highlight the words and sentences that represent how each model describes the main object in the image." + ], + "image_footnote": [], + "bbox": [ + 223, + 146, + 325, + 262 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "SQ-LLaVA-7B (Ours): In this black and white photograph, a man is engrossed in the task of inspecting the engine of a Hyundai car. The car, with its hood open, reveals the intricate details of its engine. The man, dressed in a casual attire of a t-shirt and jeans, is leaning over the car, his attention completely absorbed by the engine. The car's license plate, reading \"VK8-LQ17\", is clearly visible. Adding a touch of companionship to the scene, a small white dog is seen standing next to the car, seemingly curious about the man's actions. The photograph is credited to Nikky Stephen, as indicated by the text at the top of the image. The overall scene paints a picture of a typical day in the life of a car enthusiast.", + "bbox": [ + 330, + 157, + 495, + 258 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "LLaVA-v1.5-7B: The image features a man standing next to a car, with the hood open. He is leaning over the car, possibly working on it or checking something under the hood. A dog is standing nearby, attentively watching the man's actions. The car is parked in a driveway, and the scene appears to be a casual, everyday moment between the man and his dog.", + "bbox": [ + 227, + 266, + 493, + 301 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "ShareGPT4V-7B: The man, dressed in a casual attire of a t-shirt and khaki pants, stands to the right of the car, his attention completely absorbed by the task at hand. The car's license plate, reading 'YKL8-O17', is clearly visible. Adding a touch of companionship to the scene, a white dog stands obediently at the man's feet, perhaps waiting for a command or simply enjoying the company. The photograph is taken from a low angle, giving a unique perspective to the scene. The man, the car, and the dog are the main subjects of this image, each contributing to the narrative of a moment frozen in time. The black and white color scheme adds a timeless quality to the photograph, highlighting the contrast between the man, the car, and the dog.", + "bbox": [ + 227, + 308, + 493, + 369 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/24fb97b1af7b0ebf481d7b8082670b2d571ae71e0da48c4761af0547fafc2cea.jpg", + "image_caption": [ + "Detailed Image Description" + ], + "image_footnote": [], + "bbox": [ + 504, + 145, + 609, + 253 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "SQ-LLA-7B (Ours): The image captures a scene on a city street. Dominating the foreground is a traffic sign, painted in striking red and white colors. The sign features a black silhouette of a duck, a universal symbol for caution. Above the duck, there's a yellow light, indicating a stop signal. The background reveals a modern building with large windows, reflecting the urban environment. A tree, standing in front of the building, adds a touch of nature to the scene. The image does not contain any discernible text. The relative positions of the objects suggest that the sign is closer to the viewer, while the building and the tree are further away.", + "bbox": [ + 614, + 159, + 772, + 253 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "LLaVA-v1.5-7B: The image features a street scene with a unique combination of traffic signs. There is a stop sign with a duck on it, which is an unusual sight. The duck is positioned in the middle of the stop sign, making it stand out. In the background, there are several cars parked or driving along the street. The cars are of various sizes and are scattered throughout the scene. The presence of the duck on the stop sign adds a touch of humor to the otherwise ordinary street scene.", + "bbox": [ + 506, + 261, + 771, + 303 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "ShareGPT4V-7B: The image captures a scene on a city sidewalk. Dominating the foreground is a traffic sign, painted in a striking red and white color scheme. The sign is triangular, a common shape for traffic signs, and it features a black duck in the center, adding a touch of whimsy to the otherwise mundane object. Just behind the duck sign, there's another sign. This one is diamond-shaped and also red and white, but it's adorned with a green traffic light. The green light stands out against the red and white background, drawing attention to its presence. The signs are affixed to a gray pole, which is firmly planted in the sidewalk.", + "bbox": [ + 504, + 313, + 772, + 368 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "only describe the image from a general concept, such as \"car\", \"dog\", or \"everyday moment\". Even though ShareGPT4V can generate detailed descriptions, it still suffers from the issue of object hallucination, such as \"perhaps waiting\" and \"behind the duck sign\". By observation, SQ-LLaVA can describe the image with less unintended text, yielding higher reliability.", + "bbox": [ + 212, + 455, + 787, + 532 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Quantitative Analysis. SQ-LLaVA serves as a general-purpose vision-language model and enables zero-shot image captioning. As indicated in Table 2a, SQ-LLaVA achieves $73\\%$ and $66\\%$ averaged improvement over ClipCap and DiscriTune on all datasets, indicating the effectiveness of visual instruction tuning. Compared with the baseline model LLaVA-V1.5 [21], SQ-LLaVA achieves $2\\%$ averaged improvement on all datasets, indicating the effectiveness of visual self-questioning. Also, we find SQ-LLaVA* surpasses the baseline model ShareGPT4V [5] on Nocaps out and Conceptual dataset, demonstrating the adaptability of the proposed method on unseen testing images from new domains. Moreover, as shown in Table 2b, SQ-LLaVA can be easily adapted to COCO captioning via instruction tuning on short descriptions.", + "bbox": [ + 212, + 532, + 789, + 700 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.5 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 722, + 388, + 738 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In Table 3, we conduct experiments with different architecture designs and training strategies on five question-answering benchmarks. For a fair comparison, we train the baseline models on our local machine with the same training recipe of LLaVA-LoRA [21]. Specifically, we present our full model and three ablated models by removing one component each time. We adopt the dataset [21] with $558\\mathrm{k}$ for pre-training (PT) and $665\\mathrm{k}$ for fine-tuning (FT). As compared, self", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "SQ-LLaVA: Self-Questioning for Large Vision-Language Assistant", + "bbox": [ + 292, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/e74a3a3c94ca7648145653e0fb90d9ed91b84155f960cd721737b6fff669c19f.jpg", + "table_caption": [ + "Table 3: Ablation study of training strategy on visual instruction tasks. All models are in 7B scale with three components of ViT-LoRA (V-LoRA), self-questioning (SQ), and prototype extractor (Proto). We provide experiments on two instruction datasets [5,21] with different pre-training (PT) and instruction tuning (IT) data scales." + ], + "table_footnote": [], + "table_body": "
PTITV-LoRASQProtoVizWiz\\( \\mathrm{SQA}^I \\)\\( \\mathrm{VQAT} \\)POPE\\( \\mathrm{LLaVA}^W \\)Avg.
558K665KXXX49.468.458.286.567.165.9
✓X✓52.467.958.687.765.666.4
✓✓X52.668.457.888.267.366.9
X✓✓53.469.358.187.967.967.3
✓✓✓54.068.958.687.768.167.5
1200K700KXXX51.568.958.986.872.167.6
✓X✓54.068.960.287.271.668.4
✓✓X55.469.259.586.877.369.6
X✓✓54.270.360.587.572.769.0
✓✓✓55.370.560.587.274.369.6
", + "bbox": [ + 271, + 212, + 736, + 359 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "questioning (SQ) brings a consistent performance boost on all the five benchmarks, indicating the effectiveness of visual self-questioning on improving visual language understanding. Besides, we introduce the prototype extractor (Proto) to enhance visual representation, achieving $0.9\\%$ improvement in average accuracy among five benchmark. With all three components incorporated, we observe a $2.4\\%$ improvement in average accuracy.", + "bbox": [ + 212, + 387, + 787, + 477 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "As shown by the bottom block of Table 3, we conduct experiments with the same ablation settings but with a larger scale of the visual instruction data [5] (i.e., for both PT and IT). Overall, SQ-LLaVA achieves $2.4\\%$ improvement over the baseline model after training on the smaller dataset and achieves $3.0\\%$ improvement after training on the larger dataset.", + "bbox": [ + 212, + 478, + 787, + 553 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusions", + "text_level": 1, + "bbox": [ + 214, + 577, + 369, + 593 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This work has introduced SQ-LLaVA, a new visual instruction tuning method that enhances general-purpose vision-language understanding and image-oriented question answering through visual self-questioning. Our experiments demonstrate that SQ-LLaVA achieves superior performance with fewer training parameters and instructional data. We have also conducted a comprehensive study on visual discovery/reasoning tasks and found that SQ-LLaVA generalizes well to a wide range of unseen tasks and outperforms several state-of-the-art methods. Qualitative assessments show that SQ-LLaVA strengthens visual representation and domain alignment, effectively reducing object hallucination and improving the semantic interpretation of images. Our findings highlight the potential of visual self-questioning as a powerful training strategy for the visual instruction tuning framework, paving the way for realizing more efficient and effective large vision-language models. Particularly, SQ-LLaVA frames questioning as an intrinsic goal of tuning LLMs, encouraging the exploration of the model's curiosity (the ability to ask questions proactively) in solving complex problems.", + "bbox": [ + 212, + 608, + 787, + 837 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "G. Sun et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 323, + 160 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Gpt-4v(ison) system card (2023)", + "2. Agrawal, H., Desai, K., Wang, Y., Chen, X., Jain, R., Johnson, M., Batra, D., Parikh, D., Lee, S., Anderson, P.: Nocaps: Novel object captioning at scale. In: ICCV (2019)", + "3. Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. In: ArXiv (2023)", + "4. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., Ziegler, D., Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Chess, B., Clark, J., Berner, C., McCandlish, S., Radford, A., Sutskever, I., Amodei, D.: Language models are few-shot learners. In: NeurIPS (2020)", + "5. Chen, L., Li, J., wen Dong, X., Zhang, P., He, C., Wang, J., Zhao, F., Lin, D.: Sharegpt4v: Improving large multi-modal models with better captions. In: ArXiv (2023)", + "6. Chen, X., Fang, H., Lin, T.Y., Vedantam, R., Gupta, S., Dollar, P., Zitnick, C.L.: Microsoft coco captions: Data collection and evaluation server. In: ArXiv (2015)", + "7. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S., Schuh, P., Shi, K., Tsvyashchenko, S., Maynez, J., Rao, A., Barnes, P., Tay, Y., Shazeer, N.M., Prabhakaran, V., Reif, E., Du, N., Hutchinson, B.C., Pope, R., Bradbury, J., Austin, J., Isard, M., Gur-Ari, G., Yin, P., Duke, T., Levskaya, A., Ghemawat, S., Dev, S., Michalewski, H., Garcia, X., Misra, V., Robinson, K., Fedus, L., Zhou, D., Ippolito, D., Luan, D., Lim, H., Zoph, B., Spiridonov, A., Sepassi, R., Dohan, D., Agrawal, S., Omernick, M., Dai, A.M., Pillai, T.S., Pellat, M., Lewkowycz, A., Moreira, E., Child, R., Polozov, O., Lee, K., Zhou, Z., Wang, X., Saeta, B., Diaz, M., Firat, O., Catasta, M., Wei, J., Meier-Hellstern, K.S., Eck, D., Dean, J., Petrov, S., Fiedel, N.: Palm: Scaling language modeling with pathways. In: J. Mach. Learn. Res. (2022)", + "8. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B.A., Fung, P., Hoi, S.C.H.: Instructclip: Towards general-purpose vision-language models with instruction tuning. In: NeurIPS (2023)", + "9. Dessi, R., Bevilacqua, M., Gualdoni, E., Rakotonirina, N.C., Franzon, F., Baroni, M.: Cross-domain image captioning with discriminative finetuning. In: CVPR (2023)", + "10. Goyal, Y., Khot, T., Summers-Stay, D., Batra, D., Parikh, D.: Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In: CVPR (2017)", + "11. Gurari, D., Li, Q., Stangl, A., Guo, A., Lin, C., Grauman, K., Luo, J., Bigham, J.P.: Vizwiz grand challenge: Answering visual questions from blind people. In: CVPR (2018)", + "12. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. In: ICLR (2022)", + "13. Hudson, D.A., Manning, C.D.: Gqa: A new dataset for real-world visual reasoning and compositional question answering. In: CVPR (2019)", + "14. Ikotun, A.M., Ezugwu, A.E., Abualigah, L., Abuhaija, B., Heming, J.: K-means clustering algorithms: A comprehensive review, variants analysis, and advances in the era of big data. In: Information Sciences (2023)" + ], + "bbox": [ + 218, + 179, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "SQ-LLaVA: Self-Questioning for Large Vision-Language Assistant", + "bbox": [ + 292, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Kazemzadeh, S., Ordonez, V., Matten, M., Berg, T.: ReferItGame: Referring to objects in photographs of natural scenes. In: EMNLP (2014)", + "16. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.B.: Segment anything. In: ICCV (2023)", + "17. Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., Bernstein, M.S., Fei-Fei, L.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. In: IJCV (2016)", + "18. Li, J., Pan, K., Ge, Z., Gao, M., Zhang, H., Ji, W., Zhang, W., Chua, T.S., Tang, S., Zhuang, Y.: Empowering vision-language models to follow interleaved vision-language instructions. In: ICLR (2024)", + "19. Li, J., Li, D., Xiong, C., Hoi, S.C.H.: Blip: Bootstrapping language-image pretraining for unified vision-language understanding and generation. In: International Conference on Machine Learning (2022)", + "20. Li, Y., Du, Y., Zhou, K., Wang, J., Zhao, X., Wen, J.R.: Evaluating object hallucination in large vision-language models. In: EMNLP (2023)", + "21. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. In: ArXiv (2023)", + "22. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: NeurIPS (2023)", + "23. Liu, Y., Duan, H., Zhang, Y., Li, B., Zhang, S., Zhao, W., Yuan, Y., Wang, J., He, C., Liu, Z., Chen, K., Lin, D.: Mmbench: Is your multi-modal model an all-around player? In: ArXiv (2023)", + "24. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: ICLR (2019)", + "25. Lu, P., Mishra, S., Xia, T., Qiu, L., Chang, K.W., Zhu, S.C., Tafjord, O., Clark, P., Kalyan, A.: Learn to explain: Multimodal reasoning via thought chains for science question answering. In: NeurIPS (2022)", + "26. Manas, O., López, P.R., Ahmadi, S., Nematzadeh, A., Goyal, Y., Agrawal, A.: Mapl: Parameter-efficient adaptation of unimodal pre-trained models for vision-language few-shot prompting. In: EACL (2023)", + "27. Mishra, A., Shekhar, S., Singh, A.K., Chakraborty, A.: Ocr-vqa: Visual question answering by reading text in images. In: ICDAR (2019)", + "28. Mokady, R.: Clipcap: Clip prefix for image captioning. In: ArXiv (2021)", + "29. Ordonez, V., Kulkarni, G., Berg, T.: Im2text: Describing images using 1 million captioned photographs. In: NeurIPS (2011)", + "30. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C.L., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., Schulman, J., Hilton, J., Kelton, F., Miller, L.E., Simens, M., Askell, A., Welinder, P., Christiano, P.F., Leike, J., Lowe, R.J.: Training language models to follow instructions with human feedback. In: NeurIPS (2022)", + "31. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C.L., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., Schulman, J., Hilton, J., Kelton, F., Miller, L.E., Simens, M., Askell, A., Welinder, P., Christiano, P.F., Leike, J., Lowe, R.J.: Training language models to follow instructions with human feedback. In: NeurIPS (2022)", + "32. Papineni, K., Roukos, S., Ward, T., Zhu, W.J.: Bleu: A method for automatic evaluation of machine translation. In: ACL (2002)", + "33. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: ICML (2021)" + ], + "bbox": [ + 215, + 147, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "G. Sun et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "34. Rohrbach, A., Hendricks, L.A., Burns, K., Darrell, T., Saenko, K.: Object hallucination in image captioning. In: EMNLP (2018)", + "35. Sanh, V., Webson, A., Raffel, C., Bach, S.H., Sutawika, L., Alyafeai, Z., Chaffin, A., Stiegler, A., Scao, T.L., Raja, A., Dey, M., Bari, M.S., Xu, C., Thakker, U., Sharma, S., Szczechla, E., Kim, T., Chhablani, G., Nayak, N.V., Datta, D., Chang, J.D., Jiang, M.T.J., Wang, H., Manica, M., Shen, S., Yong, Z.X., Pandey, H., Bawden, R., Wang, T., Neeraj, T., Rozen, J., Sharma, A., Santilli, A., Févry, T., Fries, J.A., Teehan, R., Biderman, S., Gao, L., Bers, T., Wolf, T., Rush, A.M.: Multitask prompted training enables zero-shot task generalization. In: ICLR (2022)", + "36. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. In: NeurIPS Workshop (2021)", + "37. Sharma, P., Ding, N., Goodman, S., Soricut, R.: Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In: Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) (2018)", + "38. Shukor, M., Dancette, C., Cord, M.: ep-alm: Efficient perceptual augmentation of language models. In: ICCV (2023)", + "39. Singh, A., Natarajan, V., Shah, M., Jiang, Y., Chen, X., Batra, D., Parikh, D., Rohrbach, M.: Towards vqa models that can read. In: CVPR (2019)", + "40. Sun, G., Bai, Y., Yang, X., Fang, Y., Fu, Y., Tao, Z.: Aligning out-of-distribution web images and caption semantics via evidential learning. In: Proceedings of the ACM on Web Conference 2024 (2024)", + "41. Tofade, T., Elsner, J., Haines, S.: Best practice strategies for effective use of questions as a teaching tool. In: American journal of pharmaceutical education (2013)", + "42. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., Lample, G.: Llama: Open and efficient foundation language models. In: ArXiv (2023)", + "43. Touvron, H., Martin, L., Stone, K.R., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., Bikel, D.M., Blecher, L., Ferrer, C.C., Chen, M., Cucurull, G., Esiobu, D., Fernandes, J., Fu, J., Fu, W., Fuller, B., Gao, C., Goswami, V., Goyal, N., Hartshorn, A.S., Hosseini, S., Hou, R., Inan, H., Kardas, M., Kerkez, V., Khabsa, M., Kloumann, I.M., Korenev, A.V., Koura, P.S., Lachaux, M.A., Lavril, T., Lee, J., Liskovich, D., Lu, Y., Mao, Y., Martinet, X., Mihaylov, T., Mishra, P., Molybog, I., Nie, Y., Poulton, A., Reizenstein, J., Rungta, R., Saladi, K., Schelten, A., Silva, R., Smith, E.M., Subramanian, R., Tan, X., Tang, B., Taylor, R., Williams, A., Kuan, J.X., Xu, P., Yan, Z., Zarov, I., Zhang, Y., Fan, A., Kambadur, M., Narang, S., Rodriguez, A., Stojnic, R., Edunov, S., Scialom, T.: Llama 2: Open foundation and fine-tuned chat models. In: ArXiv (2023)", + "44. Triantafillou, E., Zhu, T.L., Dumoulin, V., Lamblin, P., Xu, K., Goroshin, R., Gelada, C., Swersky, K., Manzagol, P.A., Larochelle, H.: Meta-dataset: A dataset of datasets for learning to learn from few examples. In: ICLR (2019)", + "45. Vattani, A.: K-means requires exponentially many iterations even in the plane. In: Annual Symposium on Computational Geometry (2009)", + "46. Vedantam, R., Zitnick, C.L., Parikh, D.: Cider: Consensus-based image description evaluation. In: CVPR (2015)", + "47. Wang, J., Sun, G., Wang, P., Liu, D., Dianat, S., Rabbani, M., Rao, R., Tao, Z.: Text is mass: Modeling as stochastic embedding for text-video retrieval. In: CVPR (2024)" + ], + "bbox": [ + 215, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "SQ-LLaVA: Self-Questioning for Large Vision-Language Assistant", + "bbox": [ + 292, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "48. Wang, Y., Kordi, Y., Mishra, S., Liu, A., Smith, N.A., Khashabi, D., Hajishirzi, H.: Self-instruct: Aligning language models with self-generated instructions. In: ACL (2022)", + "49. Wang, Y., Mishra, S., Alipoormolabashi, P., Kordi, Y., Mirzaei, A., Arunkumar, A., Ashok, A., Dhanasekaran, A.S., Naik, A., Stap, D., Pathak, E., Karamanolakis, G., Lai, H.G., Purohit, I., Mondal, I., Anderson, J., Kuznia, K., Doshi, K., Patel, M., Pal, K.K., Moradshahi, M., Parmar, M., Purohit, M., Varshney, N., Kaza, P.R., Verma, P., Puri, R.S., Karia, R., Sampat, S.K., Doshi, S., Mishra, S.D., Reddy, S., Patro, S., Dixit, T., Shen, X., Baral, C., Choi, Y., Smith, N.A., Hajishirzi, H., Khashabi, D.: Super-naturalinstructions: Generalization via declarative instructions on 1600+ nlp tasks. In: EMNLP (2022)", + "50. Wei, J., Bosma, M., Zhao, V., Guu, K., Yu, A.W., Lester, B., Du, N., Dai, A.M., Le, Q.V.: Finetuned language models are zero-shot learners. In: ICLR (2022)", + "51. Xu, C., Sun, Q., Zheng, K., Geng, X., Zhao, P., Feng, J., Tao, C., Jiang, D.: Wizardlm: Empowering large language models to follow complex instructions. In: ICLR (2024)", + "52. Young, P., Lai, A., Hodosh, M., Hockenmaier, J.: From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. In: TACL (2014)", + "53. Yu, W., Yang, Z., Li, L., Wang, J., Lin, K., Liu, Z., Wang, X., Wang, L.: Mm-vet: Evaluating large multimodal models for integrated capabilities. In: ArXiv (2023)", + "54. Zhang, R., Han, J., Zhou, A., Hu, X., Yan, S., Lu, P., Li, H., Gao, P., Qiao, Y.: Llama-adapter: Efficient fine-tuning of language models with zero-init attention. In: ICLR (2024)", + "55. Zhang, Y., Zhang, R., Gu, J., Zhou, Y., Lipka, N., Yang, D., Sun, T.: Llavar: Enhanced visual instruction tuning for text-rich image understanding. In: ArXiv (2023)", + "56. Zhao, B., Wu, B., Huang, T.: Svit: Scaling up visual instruction tuning. In: ArXiv (2023)", + "57. Zheng, L., Chiang, W.L., Sheng, Y., Zhuang, S., Wu, Z., Zhuang, Y., Lin, Z., Li, Z., Li, D., Xing, E.P., Zhang, H., Gonzalez, J., Stoica, I.: Judging llm-as-a-judge with mt-bench and chatbot arena. In: NeurIPS (2023)", + "58. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. In: ICLR (2024)" + ], + "bbox": [ + 215, + 146, + 785, + 619 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "G. Sun et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/SQ-LLaVA_ Self-Questioning for Large Vision-Language Assistant/3909e8d0-9729-43b7-abd9-e1c4bba247cf_model.json b/2024/SQ-LLaVA_ Self-Questioning for Large Vision-Language Assistant/3909e8d0-9729-43b7-abd9-e1c4bba247cf_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4935c0f15526f7a0f9b87c9c3ae0452fd1e73c21 --- /dev/null +++ b/2024/SQ-LLaVA_ Self-Questioning for Large Vision-Language Assistant/3909e8d0-9729-43b7-abd9-e1c4bba247cf_model.json @@ -0,0 +1,2667 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.283, + 0.141, + 0.723, + 0.187 + ], + "angle": 0, + "content": "SQ-LLaVA: Self-Questioning for Large Vision-Language Assistant" + }, + { + "type": "text", + "bbox": [ + 0.273, + 0.212, + 0.732, + 0.245 + ], + "angle": 0, + "content": "Guohao Sun1, Can Qin2, Jiamian Wang1, Zeyuan Chen2, Ran Xu2, and Zhiqiang Tao1" + }, + { + "type": "text", + "bbox": [ + 0.312, + 0.255, + 0.69, + 0.283 + ], + "angle": 0, + "content": "\\(^{1}\\) Rochester Institute of Technology, Rochester, NY, US \\(^{2}\\) Salesforce AI Research, CA, US" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.317, + 0.744, + 0.609 + ], + "angle": 0, + "content": "Abstract. Recent advances in vision-language models have shown notable generalization in broad tasks through visual instruction tuning. However, bridging the gap between the pre-trained vision encoder and the large language models (LLMs) becomes the whole network's bottleneck. To improve cross-modality alignment, existing works usually consider more visual instruction data covering a broader range of vision tasks to fine-tune the model for question-answering, which, however, is costly to obtain and has not thoroughly explored the rich contextual information contained in images. This paper first attempts to harness the overlooked context within visual instruction data, training the model to self-supervised \"learning\" how to ask high-quality questions. In this way, we introduce a novel framework named SQ-LLaVA: Self-Questioning for Large Vision-Language Assistant. SQ-LLaVA exhibits proficiency in generating flexible and meaningful image-related questions while analyzing the visual clue and prior language knowledge, signifying an advanced level of generalized visual understanding. Moreover, fine-tuning SQ-LLaVA on higher-quality instruction data shows a performance improvement compared with traditional visual-instruction tuning methods. This improvement highlights the efficacy of self-questioning techniques in achieving a deeper and more nuanced comprehension of visual content across various contexts. Our code is available at https://github.com/heliossun/SQ-LLaVA." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.621, + 0.744, + 0.65 + ], + "angle": 0, + "content": "Keywords: Vision-Language Understanding \\(\\cdot\\) Multi-modal LLM \\(\\cdot\\) Instruction Tuning" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.674, + 0.377, + 0.691 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.79, + 0.841 + ], + "angle": 0, + "content": "The recently emerging large vision-language methods, such as large language-and-vision assistant (LLaVA) and its variants [3,5,21,22,54], fine-tune large language models (LLM) [3,54,57] on visual instruction data [18,55,56,58] to realize diverse open-world multimodal understanding, demonstrating a surprising efficacy of visual instruction tuning - the LLM learns to perform complex vision tasks by conditioning on a prompt containing image and text clues. Existing visual instruction data is mainly built upon conversations (e.g., ChatGPT/GPT4-V [1]), consisting of images and multiple question-answer pairs. Building high-quality visual instruction data usually requires images and texts from different tasks to" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "G. Sun et al." + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.145, + 0.775, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.328, + 0.79, + 0.371 + ], + "angle": 0, + "content": "Fig. 1: (a) Comparison between visual instruction tuning and visual self-questioning (ours) for vision-language assistant. (b) The proposed SQ-LLaVA achieves state-of-the-art performance on 9 out of 10 tasks compared with other open-ended models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.402, + 0.789, + 0.508 + ], + "angle": 0, + "content": "generate diverse questions, such as \"Please provide the bounding box coordinate of the region this sentence describes A dead leaf on the ground\" for the object detection task. Empirically, by increasing the diversity of questions, LLaVA has achieved better performance on GQA and VizWiz tasks (26% and 45% over previous state-of-the-art methods [8]). This evidence strongly suggests the advantage of training models on a broad spectrum and diverse array of tasks for enriching general vision-language understanding." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.789, + 0.809 + ], + "angle": 0, + "content": "LLaVA [21] model family usually consists of a pre-trained vision encoder (e.g., CLIP-ViT [33]), a large generative language model (LLMs like Vicuna [57], LLaMA [54], QWen [3], etc.), and a vision-to-language projector implemented by a few linear layers. However, the modality gap between the pre-trained vision encoder and the language model restricts both sides' generalization ability and feature representation. To overcome this challenge, various techniques have been proposed to align the vision and language domains, which could be roughly categorized into three groups: 1) build a more robust image feature extractor [8, 54, 58], 2) collect more high-quality training data [5, 8, 21, 58], and 3) fully fine-tune the vision and language models simultaneously during the pre-training stage [5]. While these methods have shown good progress in mitigating the domain gap, they inevitably bring higher computational costs and more expensive data collection and may also require sophisticated manual designs as well as heavy annotating efforts. Moreover, images generally encompass rich information, including color, context, and the relationships between objects, but most existing visual instruction datasets capture only a fraction. We posit that leveraging such under-explored knowledge could significantly aid in vision-language understanding. In this study, we propose a new visual self-questioning approach (Fig. 1a) by training the LLM to ask questions and discover vision clues without collecting extra data from other sources." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Unlike existing visual instruction tuning methods that focus solely on answer prediction, the proposed visual self-questioning aims to extract relevant question" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.293, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "SQ-LLaVA: Self-Questioning for Large Vision-Language Assistant" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.146, + 0.369, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.37, + 0.147, + 0.495, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.15, + 0.571, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.574, + 0.15, + 0.779, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.24, + 0.788, + 0.282 + ], + "angle": 0, + "content": "Fig. 2: Left: Questions spread more samples around a higher mean CLIPScore than answers. Right: Example of highly image-relevant questions within the visual instructional dataset for training a Vision-Language assistant." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.313, + 0.789, + 0.493 + ], + "angle": 0, + "content": "context. As illustrated in Fig. 2 right, the questions could contain more image-related information than the answers, inspiring us that aligning images and related questions may further improve the model's vision-language understanding capacity. Quantitatively, we compute the CLIPScore [33] (a higher value means better visual-text relevance) for all the image-question/-answer pairs on two visual instruction datasets (LLaVA-instruct [21] and ShareGPT4V-instruct [5]). By comparison, the questions' mean CLIPScore \\(\\mu_q = 0.184\\) is larger than the answers' \\(\\mu_a = 0.183\\) on LLaVA-instruct; \\(\\mu_q = 0.186\\) is larger than the answers' \\(\\mu_a = 0.184\\) on ShareGPT4V-instruct. Fig. 2 left shows the CLIPScore distribution, indicating questions have a similar and even better visual relevance than the answers. Therefore, questions in current visual instruction data can be used to fine-tune instructional LLMs owing to their diverse semantic information." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.495, + 0.789, + 0.646 + ], + "angle": 0, + "content": "This work introduces a self-questioning LLaVA, namely SQ-LLaVA, to fully utilize questions within the instruction data as an additional learning resource for training instructional LLMs and empowering the model's curiosity (questioning ability). To efficiently align the vision and language domains, we apply LoRAs [12] to optimize both the vision encoder and the instructional LLM within the SQ-LLaVA. Plus, we develop a prototype extractor to enhance visual representation by leveraging learned clusters with meaningful semantic information to improve vision-language alignment further. Extensive experiments demonstrate that SQ-LLaVA surpasses existing visual instruction tuning methods in general vision understanding (see Fig. 1b). We summarize the contributions as follows." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.658, + 0.787, + 0.734 + ], + "angle": 0, + "content": "- We propose a novel training technique, visual self-questioning for vision-language assistants (SQ-LLaVA), by leveraging highly relevant question contexts in instruction data. This SQ learning task promotes instructional LLMs to understand the relationship between images and questions, enhancing vision-language alignment without needing new data collection." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.734, + 0.787, + 0.795 + ], + "angle": 0, + "content": "- We design and develop a lightweight tuning architecture for SQ-LLaVA, consisting of ViT-LoRA, LLM-LoRA, and a prototype extractor. The prototype extractor enhances vision embeddings, while ViT-LoRA and LLM-LoRA efficiently align vision and language domains during training." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "- Extensive experimental results show that the proposed SQ-LLaVA leads to better performance in several tasks, including visual question-answering, visual instruction benchmarks, and zero-shot image captioning." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.658, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "G. Sun et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.389, + 0.161 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.173, + 0.419, + 0.189 + ], + "angle": 0, + "content": "2.1 Instruction Tuning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.196, + 0.79, + 0.56 + ], + "angle": 0, + "content": "Instruction tuning emerged as a pivotal methodology within the realm of natural language processing (NLP), facilitating Large Language Models (LLMs) such as GPT-3 [4], PaLM [7], and LLaMA [42] to interpret and execute human language instructions across a spectrum of NLP tasks. This approach diverges from traditional fine-tuning mechanisms by incorporating a specialized data structure, termed instruction-following data [50], which is instrumental in the fine-tuning process of LLMs. Generally, there are two main categories regarding instruction tuning methods - 1) closed-domain and 2) open-domain. The closed-domain instruction tuning [35,50,51] studies engaged LLMs with a comprehensive assortment of publicly accessible datasets and subsequently assessed their performance across diverse NLP tasks [44]. The empirical evidence from these inquiries consistently indicated that integrating varied NLP task instructions significantly augments the LLMs' efficacy in navigating novel tasks. Nonetheless, LLMs calibrated with such closed-form instructions exhibited limitations in real-world user scenarios, prompting the development of an alternative approach. To address these constraints, the concept of open-domain instruction tuning [30, 49] is conceived. OpenAI pioneered this approach by employing human annotators to compile a corpus of real-world question-answer datasets. These datasets form the foundation for training a reward model through reinforcement learning methodologies. The trained reward model then functions as a supervisory mechanism for further training instruction-oriented language models, such as InstructGPT [31] and Vicuna [57]. This innovation marks a significant advancement in the field, aiming to bridge the gap between LLM performance and real-world applicability by leveraging instruction data derived from authentic user interactions." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.577, + 0.518, + 0.592 + ], + "angle": 0, + "content": "2.2 Large Vision Language Models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.79, + 0.844 + ], + "angle": 0, + "content": "As the field of LLMs and instruction tuning undergoes rapid advancements, the academic research community is progressively focusing on integrating visual information into LLMs to facilitate visual instruction tuning. This emerging area of research has witnessed the development of various methodologies [3,5,21,22,54] based on foundational vision-language models [19,33,40,47] and diverse LLM architectures [3,43,57]. In particular, LLaVA [22] pioneers the integration of an LLM with a CLIP vision encoder to construct a vision language model, demonstrating remarkable capabilities in image-text dialogue tasks through pretraining alignment strategies and targeted instruction tuning. Subsequent investigations have sought to refine visual instruction tuning by enhancing the quality and variety of the datasets used during the pre-training and fine-tuning phases. Building upon these advancements, recent studies like LLaVA-v1.5 [21] and ShareGPT4V [5] have achieved notable success in general vision-language comprehension, showcasing their ability to undertake complex question-answering tasks. This progression underscores the importance of sophisticated data handling and model tuning strategies in developing effective vision-language models." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.293, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SQ-LLaVA: Self-Questioning for Large Vision-Language Assistant" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.143, + 0.784, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.316, + 0.788, + 0.384 + ], + "angle": 0, + "content": "Fig. 3: Model architecture of SQ-LLaVA. We propose prototype extractor to extract visual clustering information to enhance the visual embedding encoded by the visual encoder. SQ-LLaVA defines a new token [vusr] as specific instruction for LLM to perform visual self-questioning. Besides question answering, SQ-LLaVA treats questioning as another training objective." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.414, + 0.33, + 0.43 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.446, + 0.451, + 0.46 + ], + "angle": 0, + "content": "3.1 Architecture Overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.472, + 0.789, + 0.643 + ], + "angle": 0, + "content": "The proposed SQ-LLaVA model (see Fig. 3) consists of four main components: 1) A pre-trained vision encoder CLIP-ViT [33] that extracts a sequence embedding of image tokens \\( Z_{v} \\) for an input image \\( X_{v} \\); 2) A prototype extractor \\( \\phi(\\cdot) \\) learning visual clusters to enhance the original image tokens; 3) A trainable projection block \\( W(\\cdot) \\) with two linear layers to map the enhanced image tokens to the language domain tokens \\( H_{v} \\), handling the dimension misalignment between the vision and language domain; and 4) Our LLM backbone \\( f(\\cdot) \\) implemented by the pre-trained Vicuna [57] to predict the next token upon the previous embedding sequence. Given the input question \\( X_{q} \\) and answer \\( X_{a} \\), a word embedding matrix is used to map them to contextual embeddings \\( H_{q} \\) and \\( H_{a} \\), and the distribution over \\( H_{a}^{(i+1)} \\) can be obtained following the auto-regressive model as:" + }, + { + "type": "equation", + "bbox": [ + 0.32, + 0.651, + 0.788, + 0.672 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(H _ {a} ^ {(i + 1)} \\mid H _ {v}, H _ {q}, H _ {a} ^ {(1: i)}\\right) = \\sigma \\left(f \\left(H _ {v}, H _ {q}, H _ {a} ^ {(1: i)}\\right)\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.681, + 0.789, + 0.841 + ], + "angle": 0, + "content": "where \\(\\theta\\) represents all the trainable parameters in our model, \\(\\sigma(\\cdot)\\) is a softmax function, and \\(f(\\cdot)\\) outputs the last token embedding of the whole sequence. We denote \\(p_{\\theta}\\) as the prediction probability for the anticipated answer token \\(H_{a}^{(i+1)}\\) at the position \\(i+1\\), conditioning on the input image token embedding \\(H_{v}\\), the question token embedding \\(H_{q}\\), and the previous answer token embeddings \\(H_{a}^{(1:i)}\\). As shown in Eq. (1), the proposed SQ-LLaVA applies the language model \\(f(\\cdot)\\) to model \\(p_{\\theta}\\) given by \\(H_{v}\\), \\(H_{q}\\), and \\(H_{a}^{(1:i)}\\). Existing visual instruction tuning methods [5,21] are only able to predict \\(H_{a}\\), yet cannot fully exploit the rich semantic clues within \\(H_{q}\\). In this study, we propose visual self-questioning instructions to guide LLMs in capturing the visual content within questions." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "G. Sun et al." + }, + { + "type": "text", + "bbox": [ + 0.248, + 0.149, + 0.351, + 0.162 + ], + "angle": 0, + "content": "System-message" + }, + { + "type": "equation", + "bbox": [ + 0.248, + 0.164, + 0.73, + 0.185 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} _ {c} ^ {(1)}: [ \\mathrm {u s r} ]: \\mathbf {X} _ {v} \\backslash \\mathrm {n} \\mathbf {X} _ {q} ^ {(1)} [ \\mathrm {a s w r} ]: \\mathbf {X} _ {a} ^ {(1)} < o ^ {d} > \\quad \\mathbf {X} _ {c} ^ {(2)}: [ \\mathrm {v u s r} ]: \\mathbf {X} _ {q} ^ {(2)} < o ^ {d} > [ \\mathrm {a s w r} ]: \\mathbf {X} _ {a} ^ {(2)} < o ^ {d} >\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.248, + 0.187, + 0.73, + 0.208 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} _ {c} ^ {(3)}: [ \\text {u s r} ]: \\mathbf {X} _ {q} ^ {(3)} [ \\text {a s w r} ]: \\mathbf {X} _ {a} ^ {(3)} < _ {O} d > \\quad \\dots \\quad \\mathbf {X} _ {c} ^ {(P)}: [ \\text {v u s r} ]: \\mathbf {X} _ {q} ^ {(P)} < _ {O} d > [ \\text {a s w r} ]: \\mathbf {X} _ {a} ^ {(P)} < _ {O} d >\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.224, + 0.79, + 0.308 + ], + "angle": 0, + "content": "Fig. 4: The input sequence used to train SQ-LLaVA. Our model is trained to predict question, answer, and where to stop. We use tokens to represent learnable tokens, where \\( X_{q} \\) is the question, \\( X_{a} \\) is the answer, and \\( < o^{d}> \\) is the delimiter token. In SQ-LLaVA, the System-message = \"The assistant gives helpful, detailed, and polite answers to the user's questions. Also, the assistant is a curious virtual user can ask complex questions that are relevant to the content in the image.\"" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.34, + 0.548, + 0.357 + ], + "angle": 0, + "content": "3.2 Visual Self-questioning Instruction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.371, + 0.79, + 0.628 + ], + "angle": 0, + "content": "In broad real-world scenarios, proactively asking a question requires more understanding and background knowledge than answering [41]. Similarly, this work proposes visual self-questioning to encourage the LLM to discover deeper vision-language alignment and improve the overall instruction-following performance. Particularly, SQ-LLaVA treats questioning as a new learning objective beyond answering, which, to the best of our knowledge, is the first practice in the field of visual instruction tuning. While the current vision-language model can ask questions [1], such skill is still learned from question-answering through instruction tuning. However, our proposed method shows that the decoder-based LLM has the potential to learn more skills, such as how to ask questions spontaneously when a unique instruction token is given (e.g., we define [vusi] in our work). Furthermore, visual self-questioning can potentially improve general vision-language understanding. To be specific, as shown in Fig. 1, there are a certain amount of questions containing more meaningful image-related information than answers in the existing visual instruction data [5, 21]. Thus, we hypothesize that the vision-language understanding can be improved once the LLM learns how to predict relevant questions about a given image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.79, + 0.841 + ], + "angle": 0, + "content": "Self-Questioning Prompt Design. We provide ground-truth content for the visual self-questioning, restricting SQ-LLaVA from asking image-related questions. To this end, we leverage questions as another learnable resource and follow the regular auto-regressive training objective. As shown in Fig. 4, the training data for SQ-LLaVA is designed in a format with a pre-defined template. To be specific, the system message as a fixed prompt is added at the beginning of each instruction data, indicating a general job description (e.g., gives helpful answers and asks complex questions) for the LLM. Existing visual instruction tuning methods utilize the unique tokens ([usr], [aswr]) to give the LLM a particular instruction (e.g., questioning understanding, answer prediction, etc.) and apply the delimiter token \\(< \\sigma^d>\\) to mark the ending position. In this work, we propose a new special token [vusr], indicating a specific instruction - asking questions. Combining with the delimiter, we can construct instructions for self-questioning as a new training task." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.293, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "SQ-LLaVA: Self-Questioning for Large Vision-Language Assistant" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.197 + ], + "angle": 0, + "content": "Each sample of current visual instruction data consists of one image \\(X_{v}\\) and \\(P\\) question-answer pairs \\((X_{q}^{(1)},X_{a}^{(1)},\\dots ,X_{q}^{(P)},X_{a}^{(P)})\\). We collect one question \\(X_{q}^{(j)}\\) and its answer \\(X_{a}^{(j)}\\) with special tokens to construct the \\(j^{th}\\) turn conversation as" + }, + { + "type": "equation", + "bbox": [ + 0.281, + 0.207, + 0.786, + 0.248 + ], + "angle": 0, + "content": "\\[\nX _ {c} ^ {j} = \\left\\{ \\begin{array}{l l} \\left([ \\mathrm {u s r} ], X _ {q} ^ {(j)}, [ \\mathrm {a s w r} ], X _ {a} ^ {(j)}\\right) & j = 1 \\text {o r} j > 1, R < \\delta \\\\ \\left([ \\mathrm {v u s r} ], X _ {q} ^ {(j)}, [ \\mathrm {a s w r} ], X _ {a} ^ {(j)}\\right) & j > 1, R > \\delta \\end{array} , \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.259, + 0.784, + 0.304 + ], + "angle": 0, + "content": "where \\( R \\in [0,1] \\) is a random number and \\( \\delta = 0.5 \\) is a threshold that sets the proportion of self-questioning pairs in the conversations. Finally, the full-text input \\( X_{c} \\) will be mapped to textual embedding \\( H_{c} \\) through word embedding." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.305, + 0.785, + 0.485 + ], + "angle": 0, + "content": "SQ-LLaVA performs zero-shot questioning without any in-context knowledge or human language instruction since the only instructional prompt is a unique token [vusr]. After visual self-questioning on instruction data with various question formats \\( X_{q} \\), the questions sampled by SQ-LLaVA are more diversified than GPT4-V (as shown in Fig. 5), since the LLM has learned alignment between image and questions, which is different from previous works [1, 48]. Specifically, previous general-purpose vision language models such as GPT4-V [1] can generate questions based on a given image, but it requires explicit language instruction such as \"Ask complex questions that are relevant to the content in the image\". Accordingly, the quality of generated questions highly relies on prompt engineering. Also, self-instruct [48] utilizes in-context learning to prompt the LLM to ask specific questions." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.509, + 0.53, + 0.525 + ], + "angle": 0, + "content": "3.3 Enhanced Visual Representation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.535, + 0.784, + 0.61 + ], + "angle": 0, + "content": "Unlike previous visual instruction tuning methods, SQ-LLaVA jointly benefits from visual self-questioning and question-answering. For better visual self-questioning, we develop a prototype extractor that recognizes and groups similar patterns of visual information from the latent space. Our primary goal is to enhance visual representation through prototype learning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.611, + 0.785, + 0.731 + ], + "angle": 0, + "content": "Specifically, we utilize clustering to extract centroid representations of image tokens \\( Z_{v} \\), where each cluster center is treated as a prototype, which, in return, will be distributed to each of the original image token embeddings. Our proposed prototype extractor \\( \\phi(\\cdot) \\) is a lightweight design involving two parts: 1) cluster center optimization and 2) prototype information distribution. Following [14, 45], we randomly initialize \\( K = 256 \\) cluster centers \\( C \\) and deploy the iterative Expectation-Maximization (EM) clustering process to capture representative semantics in the latent space by" + }, + { + "type": "equation", + "bbox": [ + 0.367, + 0.742, + 0.786, + 0.782 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\operatorname {E} \\text {- s t e p}: \\mathcal {M} ^ {(t)} = \\sigma \\left(q \\left(C ^ {(t)}\\right) * k \\left(Z _ {v}\\right) ^ {\\top}\\right), \\\\ \\text {M - s t e p}: C ^ {(t + 1)} = \\mathcal {M} ^ {(t)} * v \\left(Z _ {v}\\right), \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.794, + 0.784, + 0.841 + ], + "angle": 0, + "content": "where \\(\\mathcal{M}^{(t)}\\in [0,1]\\) denotes a soft cluster assignment map at the \\(t^{th}\\) iteration \\(\\sigma (\\cdot)\\) is a softmax function, and \\(t\\in \\{1,\\dots ,T\\}\\) indexes the iteration of EM step with \\(T = 2\\) in this work. Three trainable linear layers \\(q,k\\), and \\(v\\) are used in (3.3)" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "G. Sun et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.192 + ], + "angle": 0, + "content": "where \\( q(\\cdot) \\) projects the prototype \\( C \\) to a query vector, and \\( k(\\cdot) \\) and \\( v(\\cdot) \\) project \\( H_{v} \\) into key and value vectors, followed by a normalization layer, respectively. The prototype extractor iteratively updates cluster map \\( \\mathcal{M} \\) and centers \\( C \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.788, + 0.24 + ], + "angle": 0, + "content": "After the prototype extraction, we train a linear layer \\( z(\\cdot) \\) to adaptively map the visual cluster information to the raw image embedding \\( Z_{v} \\). For the \\( i^{th} \\) token embedding \\( Z_{v}^{(i)} \\), we update it as" + }, + { + "type": "equation", + "bbox": [ + 0.345, + 0.25, + 0.788, + 0.278 + ], + "angle": 0, + "content": "\\[\nZ _ {v} ^ {(i)} = Z _ {v} ^ {(i)} + z \\left(\\frac {1}{K} \\sum_ {j = 1} ^ {K} S _ {c} \\left(C _ {j}, Z _ {v} ^ {(i)}\\right) \\times C _ {j}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.286, + 0.789, + 0.409 + ], + "angle": 0, + "content": "where \\( S_{c}(\\cdot, \\cdot) \\) is a normalized cosine similarity function. The weighted sum over prototypes in Eq. (4) emerges as an indispensable step for contextual understanding from image tokens, recognizing and grouping similar patterns and semantics. It clusters image tokens as prototypes that display homogeneity in semantics, such as \"grass\" and \"dog\". The prototypes can describe the intrinsic semantic meanings by aggregating entities that exhibit shared attributes. Finally, we map the image sequence embedding \\( Z_{v} \\) to language domain \\( H_{v} \\) with a two-layer linear projector \\( W(\\cdot) \\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.428, + 0.393, + 0.445 + ], + "angle": 0, + "content": "3.4 Model Training" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.452, + 0.789, + 0.681 + ], + "angle": 0, + "content": "Stage1: Pre-training for Vision-Language Alignment. Unlike text-only LLMs, the vision-language model also fine-tunes LLM using image tokens as input (see Fig. 3). Therefore, the pre-training stage aims to optimize the LLM by explicitly executing the visual instruction. The proposed SQ-LLaVA adopts Vicuna [57] as its instruction LLM, pre-training on massive text corpora to predict the next text token given the previous context, not only containing text but also visual instructions. To achieve this, we organize the pre-training data as \\( D_{PT} = \\{[X_v^{(1)}, X_a^{(1)}], \\dots, [X_v^{(N)}, X_a^{(N)}]\\} \\), where \\( N \\) is the total number of training samples, and each sample has an image and its related descriptions. Each image and text input pair will be mapped to sequence embeddings (\\( H_v \\) and \\( H_a \\)) as elaborated in Section 3.1. During pre-training, we freeze the vision encoder and LLM and mainly train the prototype extractor \\( \\phi \\) and the vision-to-language projector \\( W \\). The pre-training goal is to maximize the probability of the predicted image description \\( H_a \\) given an image \\( H_v \\). When training a visual instructional LLM, we follow the negative log-likelihood objective function as" + }, + { + "type": "equation", + "bbox": [ + 0.263, + 0.69, + 0.788, + 0.733 + ], + "angle": 0, + "content": "\\[\n\\sum_ {v, a \\in D _ {P T}} - \\log p _ {\\theta} (H _ {a} \\mid H _ {v}) = \\sum_ {v, a \\in D _ {P T}} \\sum_ {i = 1} ^ {L} - \\log p _ {\\theta} (H _ {a} ^ {(i + 1)} \\mid H _ {v}, H _ {a} ^ {(1: i)}), \\quad (5)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.741, + 0.788, + 0.79 + ], + "angle": 0, + "content": "where \\( L \\) denotes the sequence length of answer tokens in \\( H_{a} \\), \\( \\theta \\) is the total trainable parameter of \\( \\phi \\) and \\( W \\), \\( p(H_{a} \\mid H_{v}) \\) can be computed by Eq. (1), and \\( H_{a}^{(1:i)} \\) represents all the answer tokens before the current prediction \\( H_{a}^{(i+1)} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Stage2: Fine-tuning. Existing methods, such as LLaVA [5,22], mainly update the vision-to-language projector (usually a couple of linear layers) and the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.293, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "SQ-LLaVA: Self-Questioning for Large Vision-Language Assistant" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.223 + ], + "angle": 0, + "content": "language model during fine-tuning. Nevertheless, the projector might be too weak to capture the relationship between the image and the questions. Following the previous multi-modal understanding method [33], we unfreeze the vision encoder and LLM during fine-tuning for a joint optimization further to eliminate the gap between the vision and language domain." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.222, + 0.788, + 0.391 + ], + "angle": 0, + "content": "To mitigate the heavy computational overhead, we take advantage of LoRA [12] as a lightweight tuning option that can achieve similar (even better) performance to fully fine-tuning when training large models on a relatively small dataset. We add LoRA in both the vision encoder and LLM. Thus, the learnable parameters \\(\\theta\\) of the proposed SQ-LLaVA during fine-tuning represent a combination of all the parameters of LLM-LoRA, ViT-LoRA, prototype extractor \\(\\phi\\), and the vision-to-language projector \\(W\\). Given the instruction tuning data \\(D_{IT} = \\{[X_v^{(1)}, X_c^{(1)}], \\dots, [X_v^{(N)}, X_c^{(N)}]\\}\\), we take the conversational data \\(X_c\\) and the image \\(X_v\\) as input, mapping them to sequence embedding \\((H_c\\) and \\(H_v)\\) as elaborated in Section 3.1, and minimize the negative log-likelihood loss for the self-questioning and answering tasks as follows" + }, + { + "type": "equation", + "bbox": [ + 0.276, + 0.404, + 0.788, + 0.437 + ], + "angle": 0, + "content": "\\[\n\\text {S e l f - q u e s t i o n i n g}: \\sum_ {v, c \\in D _ {I T}} - \\log p _ {\\theta} \\left(H _ {q} ^ {(j + 1)} \\mid H _ {v}, H _ {c} ^ {(1: j)}\\right), \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.315, + 0.444, + 0.787, + 0.477 + ], + "angle": 0, + "content": "\\[\n\\text {A n s w e r i n g}: \\sum_ {v, c \\in D _ {I T}} - \\log p _ {\\theta} \\left(H _ {a} ^ {(j + 1)} \\mid H _ {v}, H _ {c} ^ {(1: j)}, H _ {q} ^ {(j + 1)}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.488, + 0.788, + 0.611 + ], + "angle": 0, + "content": "where \\( j \\in \\{1, \\dots, P\\} \\), indicating the index of question or answer within the conversational data \\( X_{c}^{(*)} \\). Notably, previous works [5,21,58] only involve answering tasks, but we introduce visual self-questioning as an additional training task for visual instruction tuning. Eventually, SQ-LLaVA, as a vision-language assistant, not only executes human instructions by optimizing the objective function in Eq. (7) but can raise questions out of the given image after optimizing the Eq. (6). This capability potentially yields more diverse question-answer guidance and enhances multi-modal understanding." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.634, + 0.376, + 0.65 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.665, + 0.44, + 0.681 + ], + "angle": 0, + "content": "4.1 Experimental Setting" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Dataset. Our work uses the open-source visual instruction dataset provided by LLaVA [21] and ShareGPT4V [5] for training. Each dataset has large-scale image-text paired data for pre-training and instruction-following data for finetuning. Specifically, the instruction data proposed by LLaVA is a comprehensive mixture of COCO, GQA [13], OCR-VQA [27], TextVQA [39], VisualGenome [17], RefCOCO [15], and image from InstructBLIP [8] involves multiple reasoning, spatial understanding, multi-step inference, optical character recognition, and grounding of visual concepts to language. The ShareGPT4V dataset consists of the same mixture image as LLaVA but enrolls more images from other datasets such as SAM [16] and Data [29,36]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.359, + 0.127 + ], + "angle": 0, + "content": "G. Sun et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.359 + ], + "angle": 0, + "content": "We evaluate our model on general vision-language understanding tasks. Specifically, we use ten visual oriented question answering benchmarks: \\(\\mathrm{VQA}^{v2}\\) [10]; GQA [13]; VizWiz [11]; \\(\\mathrm{SQA}^I\\) [25]: ScienceQA-IMG; \\(\\mathrm{VQA}^T\\): TextVQA [39]; POPE [20]; MM-Vet [53]; \\(\\mathrm{LLaVA}^W\\) [22]: LLaVA (in the wild); MMB: MMBBenchmark [23]; \\(\\mathrm{MMB}^{CN}\\): MMBench-Chinese [23]. We report the prediction accuracy on all the benchmarks. We also evaluate our model for visual information discovery through captioning. To be specific, we employ testing images from four datasets, i.e., COCO [6], Flickr30K [52], Conceptual [37], and Nocaps [2]. Following [9], we evaluate the proposed method with regular image captioning metrics, e.g., BLEU [32] and CIDEr [46]. For the open-world methods LLaVA-v1.5 [21], ShareGPT4V [5] and the proposed SQ-LLaVA, we utilize greedy search for caption generation with a prompt of \"Provide a brief description of the given image, your answer should be in one sentence.\" The generated caption is used to evaluate the performance of image captioning." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.38, + 0.393, + 0.394 + ], + "angle": 0, + "content": "4.2 Implementation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.403, + 0.788, + 0.569 + ], + "angle": 0, + "content": "We adopt Vicuna [57] as the pre-trained language generative model and CLIP-ViT [33] as the vision encoder. We pre-train the prototype extractor and the vision-to-language projector using AdamW [24] optimizer with a learning rate of \\( 2 \\times 10^{-3} \\) and a constant scheduler for one epoch. Following previous work [21], we keep the global batch size as 256 for pre-training and 128 for fine-tuning. During fine-tuning, we insert LoRA [12] with rank = 128 and \\( \\alpha = 256 \\) into the language model (LLM-LoRA) and LoRA with rank = 32 and \\( \\alpha = 64 \\) into the vision encoder (ViT-LoRA). We optimize LoRA modules, the prototype extractor, and vision-to-language projector for one epoch. We set the learning rate to \\( 2 \\times 10^{-4} \\) for LoRA, and \\( 2 \\times 10^{-5} \\) for the other layers. All the weights of the pre-trained language model and vision encoder remain fixed during fine-tuning." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.59, + 0.538, + 0.607 + ], + "angle": 0, + "content": "4.3 Zero-shot Multilingual Capability" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.78 + ], + "angle": 0, + "content": "We evaluate SQ-LLaVA on ten benchmarks, covering a range of academic Visual Question Answering (VQA) tasks and recent instruction tuning tasks designed for large vision language models. The academic VQA includes VQA-v2 [10] and VizWiz [11]. GQA [13] is a fine-grained real-world visual reasoning and question-answering benchmark. ScienceQA [25] is a benchmark with rich subjects (natural science, language science, and social science). TextVQA [39] requires the model to recognize the texts in the image. LLaVA (in the wild) and MM-Vet [53] use GPT4 to assess the capability of the models for testing. With manually designed questions, MM-Bench and MMBench-CN [23] evaluate the model's vision-related reasoning and perception for English and Chinese, respectively. POPE [20] is a benchmark for evaluating object hallucination [34]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.84 + ], + "angle": 0, + "content": "In Table 1, we quantitatively compare between SQ-LLaVA and existing models. SQ-LLaVA-7B and SQ-LLaVA-13B trained on two instruction datasets [5,21] outperform previous methods in six out of ten visual instruction tuning tasks. To be specific, SQ-LLaVA-7B achieves \\(17.2\\%\\) improvement over LLaVA-v1.5-7B" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.293, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "SQ-LLaVA: Self-Questioning for Large Vision-Language Assistant" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 1: Comparison with state-of-the-art methods on ten benchmarks. After training on the same instruction data, SQ-LLaVA train on [21] and SQ-LLaVA* trained on [5] surpass their baseline model LLaVA-v1.5 and ShareGPT4V on 9 out of 10 and 6 out of 10 benchmarks in the 7B scale, and 8 out of 10 and 6 out of 10 in the 13B scale. The best results are bold and the second-best results are underlined." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.227, + 0.785, + 0.39 + ], + "angle": 0, + "content": "
LLMModelVQA v2GQAVizWizSQA1VQA TPOPEMM-VetLLaVA WMMBMMB CN
7BInstructBLIP [8]-49.234.560.550.1-26.260.936.023.7
Qwen-VL [3]78.859.335.267.163.8---38.27.4
Qwen-VL-chat [3]78.257.538.968.261.5---60.656.7
LLaVA-v1.5 [21]78.562.050.066.858.285.930.563.464.358.3
ShareGPT4V [5]80.663.357.268.460.486.837.672.668.862.2
SQ-LLaVA79.262.854.068.958.687.732.566.366.258.1
SQ-LLaVA*80.363.755.370.560.587.237.674.366.660.0
13BInstructBLIP [8]-49.533.463.150.778.925.658.2--
LLaVA-v1.5 [21]80.063.353.671.658.285.935.470.767.763.6
ShareGPT4V [5]81.064.855.671.262.2-43.179.968.563.7
SQ-LLaVA80.163.654.669.860.287.735.574.668.762.0
SQ-LLaVA*81.365.058.271.561.987.439.780.768.562.5
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.418, + 0.789, + 0.706 + ], + "angle": 0, + "content": "on the LLaVA (in the wild) benchmark, indicating the superior capabilities of our model in tasks such as detailed description and complex reasoning. Also, SQ-LLavA-7B improves over previous methods on ScienceQA, indicating that our model excels in understanding and reasoning over scientific content and can effectively handle multi-modal information. The improvement in ScienceQA suggests strong capabilities in multi-hop reasoning, comprehension of complex scientific concepts, and the ability to utilize context and explanations to derive correct answers. SQ-LLaVA-7B has a steady improvement over LLaVA-v1.5-7B and ShareGPT4V-7B on the POPE benchmark, and the \\(2\\%\\) and \\(1\\%\\) improvement indicates that our proposed method has better reliability and trustworthiness since POPE is a task designed to evaluate the phenomenon of object hallucination [20, 34]. In the bottom section of Table 1, the proposed SQ-LLaVA-13B surpasses previous works in six out of ten benchmarks, indicating the scalability of our method on larger LLM. Notably, the performance inconsistency on some datasets might be due to the unsupervised prototype extractor (lacking pixel-wise guidance) in our model. To mitigate this issue, we could leverage pseudo object masks (e.g., given by the pre-trained segment anything) in learning prototypes. Despite this limitation, all the improvements are achieved with significantly fewer trainable parameters compared to other methods [3, 5, 21]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.726, + 0.504, + 0.741 + ], + "angle": 0, + "content": "4.4 Visual Information Discovery" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.81 + ], + "angle": 0, + "content": "In this experiment, we showcase the diversity and reliability of the proposed SQ-LLaVA through various qualitative applications, including detailed image description, visual information summary, and visual self-questioning. We also present quantitative results on the task of image captioning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Abilities of SQ-LLaVA Through Qualitative Samples. SQ-LLaVA exhibits numerous advanced capabilities compared to traditional vision-language models." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.359, + 0.127 + ], + "angle": 0, + "content": "G. Sun et al." + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.146, + 0.318, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.361, + 0.149, + 0.44, + 0.157 + ], + "angle": 0, + "content": "GPT4-V Generated" + }, + { + "type": "text", + "bbox": [ + 0.321, + 0.16, + 0.482, + 0.179 + ], + "angle": 0, + "content": "\\(Q_{1}\\) : What makes this particular toothbrush stand out among others in the image? \\(Q_{2}\\) : What color is the background?" + }, + { + "type": "text", + "bbox": [ + 0.321, + 0.179, + 0.457, + 0.196 + ], + "angle": 0, + "content": "Q2: What color is the background? \nQ3: What is this product?" + }, + { + "type": "list", + "bbox": [ + 0.321, + 0.16, + 0.482, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.285, + 0.199, + 0.426, + 0.207 + ], + "angle": 0, + "content": "SQ-LLaVA Visual Self-Questioning" + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.209, + 0.486, + 0.219 + ], + "angle": 0, + "content": "\\(SQ_{1}\\): What are some of the key aspects that this image focuses on?" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.219, + 0.417, + 0.226 + ], + "angle": 0, + "content": "\\(S Q_{2}\\) : What is the primary color of the toothbrush?" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.226, + 0.479, + 0.233 + ], + "angle": 0, + "content": "\\(S Q_{3}\\): What is the expression of the face drawn on the toothbrush?" + }, + { + "type": "list", + "bbox": [ + 0.229, + 0.209, + 0.486, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.228, + 0.25, + 0.326, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.366, + 0.254, + 0.444, + 0.262 + ], + "angle": 0, + "content": "GPT4-V Generated" + }, + { + "type": "text", + "bbox": [ + 0.328, + 0.265, + 0.482, + 0.314 + ], + "angle": 0, + "content": "\\(Q_{1}\\) : What is the setting of the image? \\(Q_{2}\\) : What are some activities that people are engaging in within the image? \\(Q_{3}\\) : What are some benefits of spending time outdoors in a beachside setting, especially on a sunny day?" + }, + { + "type": "title", + "bbox": [ + 0.284, + 0.316, + 0.426, + 0.325 + ], + "angle": 0, + "content": "SQ-LLaVA Visual Self-Questioning" + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.327, + 0.465, + 0.344 + ], + "angle": 0, + "content": "\\(S_{Q_1}\\): What is the hair color of the man that is looking up at the ocean?" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.344, + 0.363, + 0.351 + ], + "angle": 0, + "content": "\\(SQ_{2}\\): What color is the man's shirt?" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.352, + 0.436, + 0.36 + ], + "angle": 0, + "content": "\\(SQ_{3}\\): In which part is the bench, the top or the bottom?" + }, + { + "type": "list", + "bbox": [ + 0.229, + 0.327, + 0.465, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.147, + 0.592, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.639, + 0.148, + 0.718, + 0.156 + ], + "angle": 0, + "content": "GPT4-V Generated" + }, + { + "type": "text", + "bbox": [ + 0.593, + 0.158, + 0.764, + 0.174 + ], + "angle": 0, + "content": "\\(Q_{1}\\): What type of environment can the sheep be seen in?" + }, + { + "type": "text", + "bbox": [ + 0.594, + 0.175, + 0.762, + 0.198 + ], + "angle": 0, + "content": "\\(Q_{2}\\) : What are the sheep doing in the image? \\(Q_{3}\\) How do sheep contribute to the ecosystem of a meadow or field?" + }, + { + "type": "title", + "bbox": [ + 0.562, + 0.2, + 0.704, + 0.209 + ], + "angle": 0, + "content": "SQ-LLaVA Visual Self-Questioning" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.212, + 0.772, + 0.229 + ], + "angle": 0, + "content": "\\(SQ_{1}\\): In the vicinity of this group of sheep, how many horses are there? a) Five b) Four c) One d) Zero" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.229, + 0.772, + 0.245 + ], + "angle": 0, + "content": "\\(S_{Q2}\\): In this image, what kind of animals can you see in the green field? \\(S_{Q3}\\): Are there any cows in the picture?" + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.212, + 0.772, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.253, + 0.596, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.645, + 0.255, + 0.722, + 0.262 + ], + "angle": 0, + "content": "GPT4-V Generated" + }, + { + "type": "title", + "bbox": [ + 0.562, + 0.316, + 0.704, + 0.325 + ], + "angle": 0, + "content": "SQ-LLaVA Visual Self-Questioning" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.327, + 0.756, + 0.344 + ], + "angle": 0, + "content": "\\(S_{Q1}\\): The two small children are sitting on a tricycle. Which place is it? Town or village?" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.344, + 0.69, + 0.352 + ], + "angle": 0, + "content": "\\(SQ_{2}\\): What can you tell me about the child's age?" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.352, + 0.75, + 0.368 + ], + "angle": 0, + "content": "\\(S_{Q3}\\): What does the boy's action of swinging a bat suggest about his play or sport?" + }, + { + "type": "list", + "bbox": [ + 0.501, + 0.327, + 0.756, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.383, + 0.787, + 0.426 + ], + "angle": 0, + "content": "Fig.5: Visual self-questioning of SQ-LLaVA-7B. Comparing to the question data provided by GPT4-V (data collected by LLaVA-v1.5 [21]), SQ-LLaVA can generate questions with higher diversity i.e. multiple choice and tricky questions." + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.44, + 0.787, + 0.47 + ], + "angle": 0, + "content": "Table 2: Comparison of image captioning on four datasets. SQ-LLaVA is trained on data collected by [21]. SQ-LLaVA* is trained on data collected by [5]." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.48, + 0.535, + 0.589 + ], + "angle": 0, + "content": "
ModelFlickr30kNocapsoutConceptual
B@4CIDErB@4CIDErB@4CIDEr
ClipCap [28]17.2141.6520.3251.741.4723.74
DiscriTune [9]18.4844.7824.1057.061.7128.01
LLaVA-v1.5 [21]28.6781.2735.78103.562.7939.20
ShareGPT4V [5]31.0086.1737.19107.452.7837.86
SQ-LLaVA29.8883.5136.21105.422.9039.49
SQ-LLaVA*31.4983.1437.20107.422.9141.24
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.276, + 0.59, + 0.462, + 0.602 + ], + "angle": 0, + "content": "(a) Zero-shot image captioning" + }, + { + "type": "table", + "bbox": [ + 0.539, + 0.48, + 0.771, + 0.588 + ], + "angle": 0, + "content": "
ModelZero-shotFine-tune
B@4CIDErB@4CIDEr
ClipCap8.5037.0332.60108.55
DiscriTune13.9953.2032.31105.40
eP-ALM [38]29.4797.2233.35111.63
MAPL [26]12.3054.3036.45125.20
LLaVA-v1.529.96111.46--
SQ-LLaVA29.85110.7740.76136.78
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.556, + 0.59, + 0.746, + 0.602 + ], + "angle": 0, + "content": "(b) Image captioning on COCO" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.788, + 0.795 + ], + "angle": 0, + "content": "Notably, SQ-LLaVA effectively mitigates object hallucination [20, 34], resulting in predictions that are more trustworthy. Additionally, SQ-LLaVA is capable of generating diverse and meaningful questions about a given image without requiring human textual instructions. As illustrated in Fig. 5, the generated questions include multiple-choice, reasoning, binary choice questions, and the content within each question, all of which are either highly correlated or tricky to the given image, such as \"What is the hair color of the man that is looking up at the ocean?\" from the third image and \"Are there cows in the picture?\" from the second image. These examples highlight SQ-LLaVA's ability to produce questions with high diversity and quality." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.841 + ], + "angle": 0, + "content": "The detailed image description example in Fig. 6 demonstrates that SQ-LLaVA can generate visual information with concrete concepts within the image, such as \"Hyundai\", \"small white dog\", \"Nikky Stephen\". In contrast, LLaVA-v1.5 can" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.293, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "SQ-LLaVA: Self-Questioning for Large Vision-Language Assistant" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.147, + 0.326, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.354, + 0.148, + 0.471, + 0.157 + ], + "angle": 0, + "content": "Detailed Image Description" + }, + { + "type": "text", + "bbox": [ + 0.331, + 0.159, + 0.496, + 0.26 + ], + "angle": 0, + "content": "SQ-LLaVA-7B (Ours): In this black and white photograph, a man is engrossed in the task of inspecting the engine of a Hyundai car. The car, with its hood open, reveals the intricate details of its engine. The man, dressed in a casual attire of a t-shirt and jeans, is leaning over the car, his attention completely absorbed by the engine. The car's license plate, reading \"VK8-LQ17\", is clearly visible. Adding a touch of companionship to the scene, a small white dog is seen standing next to the car, seemingly curious about the man's actions. The photograph is credited to Nikky Stephen, as indicated by the text at the top of the image. The overall scene paints a picture of a typical day in the life of a car enthusiast." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.267, + 0.495, + 0.303 + ], + "angle": 0, + "content": "LLaVA-v1.5-7B: The image features a man standing next to a car, with the hood open. He is leaning over the car, possibly working on it or checking something under the hood. A dog is standing nearby, attentively watching the man's actions. The car is parked in a driveway, and the scene appears to be a casual, everyday moment between the man and his dog." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.309, + 0.495, + 0.371 + ], + "angle": 0, + "content": "ShareGPT4V-7B: The man, dressed in a casual attire of a t-shirt and khaki pants, stands to the right of the car, his attention completely absorbed by the task at hand. The car's license plate, reading 'YKL8-O17', is clearly visible. Adding a touch of companionship to the scene, a white dog stands obediently at the man's feet, perhaps waiting for a command or simply enjoying the company. The photograph is taken from a low angle, giving a unique perspective to the scene. The man, the car, and the dog are the main subjects of this image, each contributing to the narrative of a moment frozen in time. The black and white color scheme adds a timeless quality to the photograph, highlighting the contrast between the man, the car, and the dog." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.146, + 0.611, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.637, + 0.148, + 0.753, + 0.157 + ], + "angle": 0, + "content": "Detailed Image Description" + }, + { + "type": "text", + "bbox": [ + 0.615, + 0.16, + 0.773, + 0.255 + ], + "angle": 0, + "content": "SQ-LLA-7B (Ours): The image captures a scene on a city street. Dominating the foreground is a traffic sign, painted in striking red and white colors. The sign features a black silhouette of a duck, a universal symbol for caution. Above the duck, there's a yellow light, indicating a stop signal. The background reveals a modern building with large windows, reflecting the urban environment. A tree, standing in front of the building, adds a touch of nature to the scene. The image does not contain any discernible text. The relative positions of the objects suggest that the sign is closer to the viewer, while the building and the tree are further away." + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.262, + 0.772, + 0.304 + ], + "angle": 0, + "content": "LLaVA-v1.5-7B: The image features a street scene with a unique combination of traffic signs. There is a stop sign with a duck on it, which is an unusual sight. The duck is positioned in the middle of the stop sign, making it stand out. In the background, there are several cars parked or driving along the street. The cars are of various sizes and are scattered throughout the scene. The presence of the duck on the stop sign adds a touch of humor to the otherwise ordinary street scene." + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.314, + 0.773, + 0.369 + ], + "angle": 0, + "content": "ShareGPT4V-7B: The image captures a scene on a city sidewalk. Dominating the foreground is a traffic sign, painted in a striking red and white color scheme. The sign is triangular, a common shape for traffic signs, and it features a black duck in the center, adding a touch of whimsy to the otherwise mundane object. Just behind the duck sign, there's another sign. This one is diamond-shaped and also red and white, but it's adorned with a green traffic light. The green light stands out against the red and white background, drawing attention to its presence. The signs are affixed to a gray pole, which is firmly planted in the sidewalk." + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.384, + 0.788, + 0.427 + ], + "angle": 0, + "content": "Fig. 6: A qualitative evaluation of detailed image descriptions from three models. We highlight the words and sentences that represent how each model describes the main object in the image." + }, + { + "type": "text", + "bbox": [ + 0.213, + 0.457, + 0.789, + 0.533 + ], + "angle": 0, + "content": "only describe the image from a general concept, such as \"car\", \"dog\", or \"everyday moment\". Even though ShareGPT4V can generate detailed descriptions, it still suffers from the issue of object hallucination, such as \"perhaps waiting\" and \"behind the duck sign\". By observation, SQ-LLaVA can describe the image with less unintended text, yielding higher reliability." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.534, + 0.79, + 0.701 + ], + "angle": 0, + "content": "Quantitative Analysis. SQ-LLaVA serves as a general-purpose vision-language model and enables zero-shot image captioning. As indicated in Table 2a, SQ-LLaVA achieves \\(73\\%\\) and \\(66\\%\\) averaged improvement over ClipCap and DiscriTune on all datasets, indicating the effectiveness of visual instruction tuning. Compared with the baseline model LLaVA-V1.5 [21], SQ-LLaVA achieves \\(2\\%\\) averaged improvement on all datasets, indicating the effectiveness of visual self-questioning. Also, we find SQ-LLaVA* surpasses the baseline model ShareGPT4V [5] on Nocaps out and Conceptual dataset, demonstrating the adaptability of the proposed method on unseen testing images from new domains. Moreover, as shown in Table 2b, SQ-LLaVA can be easily adapted to COCO captioning via instruction tuning on short descriptions." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.723, + 0.39, + 0.739 + ], + "angle": 0, + "content": "4.5 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.789, + 0.841 + ], + "angle": 0, + "content": "In Table 3, we conduct experiments with different architecture designs and training strategies on five question-answering benchmarks. For a fair comparison, we train the baseline models on our local machine with the same training recipe of LLaVA-LoRA [21]. Specifically, we present our full model and three ablated models by removing one component each time. We adopt the dataset [21] with \\(558\\mathrm{k}\\) for pre-training (PT) and \\(665\\mathrm{k}\\) for fine-tuning (FT). As compared, self" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.359, + 0.127 + ], + "angle": 0, + "content": "G. Sun et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.203 + ], + "angle": 0, + "content": "Table 3: Ablation study of training strategy on visual instruction tasks. All models are in 7B scale with three components of ViT-LoRA (V-LoRA), self-questioning (SQ), and prototype extractor (Proto). We provide experiments on two instruction datasets [5,21] with different pre-training (PT) and instruction tuning (IT) data scales." + }, + { + "type": "table", + "bbox": [ + 0.272, + 0.213, + 0.737, + 0.361 + ], + "angle": 0, + "content": "
PTITV-LoRASQProtoVizWiz\\( \\mathrm{SQA}^I \\)\\( \\mathrm{VQAT} \\)POPE\\( \\mathrm{LLaVA}^W \\)Avg.
558K665KXXX49.468.458.286.567.165.9
✓X✓52.467.958.687.765.666.4
✓✓X52.668.457.888.267.366.9
X✓✓53.469.358.187.967.967.3
✓✓✓54.068.958.687.768.167.5
1200K700KXXX51.568.958.986.872.167.6
✓X✓54.068.960.287.271.668.4
✓✓X55.469.259.586.877.369.6
X✓✓54.270.360.587.572.769.0
✓✓✓55.370.560.587.274.369.6
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.388, + 0.788, + 0.478 + ], + "angle": 0, + "content": "questioning (SQ) brings a consistent performance boost on all the five benchmarks, indicating the effectiveness of visual self-questioning on improving visual language understanding. Besides, we introduce the prototype extractor (Proto) to enhance visual representation, achieving \\(0.9\\%\\) improvement in average accuracy among five benchmark. With all three components incorporated, we observe a \\(2.4\\%\\) improvement in average accuracy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.479, + 0.788, + 0.554 + ], + "angle": 0, + "content": "As shown by the bottom block of Table 3, we conduct experiments with the same ablation settings but with a larger scale of the visual instruction data [5] (i.e., for both PT and IT). Overall, SQ-LLaVA achieves \\(2.4\\%\\) improvement over the baseline model after training on the smaller dataset and achieves \\(3.0\\%\\) improvement after training on the larger dataset." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.578, + 0.37, + 0.594 + ], + "angle": 0, + "content": "5 Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.609, + 0.789, + 0.838 + ], + "angle": 0, + "content": "This work has introduced SQ-LLaVA, a new visual instruction tuning method that enhances general-purpose vision-language understanding and image-oriented question answering through visual self-questioning. Our experiments demonstrate that SQ-LLaVA achieves superior performance with fewer training parameters and instructional data. We have also conducted a comprehensive study on visual discovery/reasoning tasks and found that SQ-LLaVA generalizes well to a wide range of unseen tasks and outperforms several state-of-the-art methods. Qualitative assessments show that SQ-LLaVA strengthens visual representation and domain alignment, effectively reducing object hallucination and improving the semantic interpretation of images. Our findings highlight the potential of visual self-questioning as a powerful training strategy for the visual instruction tuning framework, paving the way for realizing more efficient and effective large vision-language models. Particularly, SQ-LLaVA frames questioning as an intrinsic goal of tuning LLMs, encouraging the exploration of the model's curiosity (the ability to ask questions proactively) in solving complex problems." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.293, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SQ-LLaVA: Self-Questioning for Large Vision-Language Assistant" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.324, + 0.161 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.18, + 0.466, + 0.194 + ], + "angle": 0, + "content": "1. Gpt-4v(ison) system card (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.196, + 0.787, + 0.236 + ], + "angle": 0, + "content": "2. Agrawal, H., Desai, K., Wang, Y., Chen, X., Jain, R., Johnson, M., Batra, D., Parikh, D., Lee, S., Anderson, P.: Nocaps: Novel object captioning at scale. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.239, + 0.787, + 0.279 + ], + "angle": 0, + "content": "3. Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. In: ArXiv (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.281, + 0.787, + 0.363 + ], + "angle": 0, + "content": "4. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., Ziegler, D., Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Chess, B., Clark, J., Berner, C., McCandlish, S., Radford, A., Sutskever, I., Amodei, D.: Language models are few-shot learners. In: NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.364, + 0.787, + 0.404 + ], + "angle": 0, + "content": "5. Chen, L., Li, J., wen Dong, X., Zhang, P., He, C., Wang, J., Zhao, F., Lin, D.: Sharegpt4v: Improving large multi-modal models with better captions. In: ArXiv (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.406, + 0.787, + 0.433 + ], + "angle": 0, + "content": "6. Chen, X., Fang, H., Lin, T.Y., Vedantam, R., Gupta, S., Dollar, P., Zitnick, C.L.: Microsoft coco captions: Data collection and evaluation server. In: ArXiv (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.435, + 0.787, + 0.586 + ], + "angle": 0, + "content": "7. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S., Schuh, P., Shi, K., Tsvyashchenko, S., Maynez, J., Rao, A., Barnes, P., Tay, Y., Shazeer, N.M., Prabhakaran, V., Reif, E., Du, N., Hutchinson, B.C., Pope, R., Bradbury, J., Austin, J., Isard, M., Gur-Ari, G., Yin, P., Duke, T., Levskaya, A., Ghemawat, S., Dev, S., Michalewski, H., Garcia, X., Misra, V., Robinson, K., Fedus, L., Zhou, D., Ippolito, D., Luan, D., Lim, H., Zoph, B., Spiridonov, A., Sepassi, R., Dohan, D., Agrawal, S., Omernick, M., Dai, A.M., Pillai, T.S., Pellat, M., Lewkowycz, A., Moreira, E., Child, R., Polozov, O., Lee, K., Zhou, Z., Wang, X., Saeta, B., Diaz, M., Firat, O., Catasta, M., Wei, J., Meier-Hellstern, K.S., Eck, D., Dean, J., Petrov, S., Fiedel, N.: Palm: Scaling language modeling with pathways. In: J. Mach. Learn. Res. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.588, + 0.787, + 0.628 + ], + "angle": 0, + "content": "8. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B.A., Fung, P., Hoi, S.C.H.: Instructclip: Towards general-purpose vision-language models with instruction tuning. In: NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.63, + 0.787, + 0.657 + ], + "angle": 0, + "content": "9. Dessi, R., Bevilacqua, M., Gualdoni, E., Rakotonirina, N.C., Franzon, F., Baroni, M.: Cross-domain image captioning with discriminative finetuning. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.659, + 0.787, + 0.699 + ], + "angle": 0, + "content": "10. Goyal, Y., Khot, T., Summers-Stay, D., Batra, D., Parikh, D.: Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In: CVPR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.701, + 0.787, + 0.741 + ], + "angle": 0, + "content": "11. Gurari, D., Li, Q., Stangl, A., Guo, A., Lin, C., Grauman, K., Luo, J., Bigham, J.P.: Vizwiz grand challenge: Answering visual questions from blind people. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.743, + 0.787, + 0.77 + ], + "angle": 0, + "content": "12. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. In: ICLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.771, + 0.787, + 0.798 + ], + "angle": 0, + "content": "13. Hudson, D.A., Manning, C.D.: Gqa: A new dataset for real-world visual reasoning and compositional question answering. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.8, + 0.787, + 0.84 + ], + "angle": 0, + "content": "14. Ikotun, A.M., Ezugwu, A.E., Abualigah, L., Abuhaija, B., Heming, J.: K-means clustering algorithms: A comprehensive review, variants analysis, and advances in the era of big data. In: Information Sciences (2023)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.18, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.359, + 0.127 + ], + "angle": 0, + "content": "G. Sun et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "15. Kazemzadeh, S., Ordonez, V., Matten, M., Berg, T.: ReferItGame: Referring to objects in photographs of natural scenes. In: EMNLP (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.787, + 0.217 + ], + "angle": 0, + "content": "16. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.B.: Segment anything. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.787, + 0.272 + ], + "angle": 0, + "content": "17. Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., Bernstein, M.S., Fei-Fei, L.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. In: IJCV (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.273, + 0.787, + 0.314 + ], + "angle": 0, + "content": "18. Li, J., Pan, K., Ge, Z., Gao, M., Zhang, H., Ji, W., Zhang, W., Chua, T.S., Tang, S., Zhuang, Y.: Empowering vision-language models to follow interleaved vision-language instructions. In: ICLR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.315, + 0.787, + 0.355 + ], + "angle": 0, + "content": "19. Li, J., Li, D., Xiong, C., Hoi, S.C.H.: Blip: Bootstrapping language-image pretraining for unified vision-language understanding and generation. In: International Conference on Machine Learning (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.787, + 0.384 + ], + "angle": 0, + "content": "20. Li, Y., Du, Y., Zhou, K., Wang, J., Zhao, X., Wen, J.R.: Evaluating object hallucination in large vision-language models. In: EMNLP (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.385, + 0.787, + 0.411 + ], + "angle": 0, + "content": "21. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. In: ArXiv (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.787, + 0.425 + ], + "angle": 0, + "content": "22. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.426, + 0.787, + 0.466 + ], + "angle": 0, + "content": "23. Liu, Y., Duan, H., Zhang, Y., Li, B., Zhang, S., Zhao, W., Yuan, Y., Wang, J., He, C., Liu, Z., Chen, K., Lin, D.: Mmbench: Is your multi-modal model an all-around player? In: ArXiv (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.787, + 0.48 + ], + "angle": 0, + "content": "24. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: ICLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.481, + 0.787, + 0.522 + ], + "angle": 0, + "content": "25. Lu, P., Mishra, S., Xia, T., Qiu, L., Chang, K.W., Zhu, S.C., Tafjord, O., Clark, P., Kalyan, A.: Learn to explain: Multimodal reasoning via thought chains for science question answering. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.523, + 0.787, + 0.563 + ], + "angle": 0, + "content": "26. Manas, O., López, P.R., Ahmadi, S., Nematzadeh, A., Goyal, Y., Agrawal, A.: Mapl: Parameter-efficient adaptation of unimodal pre-trained models for vision-language few-shot prompting. In: EACL (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.787, + 0.591 + ], + "angle": 0, + "content": "27. Mishra, A., Shekhar, S., Singh, A.K., Chakraborty, A.: Ocr-vqa: Visual question answering by reading text in images. In: ICDAR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.722, + 0.605 + ], + "angle": 0, + "content": "28. Mokady, R.: Clipcap: Clip prefix for image captioning. In: ArXiv (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.606, + 0.787, + 0.633 + ], + "angle": 0, + "content": "29. Ordonez, V., Kulkarni, G., Berg, T.: Im2text: Describing images using 1 million captioned photographs. In: NeurIPS (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.634, + 0.787, + 0.701 + ], + "angle": 0, + "content": "30. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C.L., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., Schulman, J., Hilton, J., Kelton, F., Miller, L.E., Simens, M., Askell, A., Welinder, P., Christiano, P.F., Leike, J., Lowe, R.J.: Training language models to follow instructions with human feedback. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.702, + 0.787, + 0.77 + ], + "angle": 0, + "content": "31. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C.L., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., Schulman, J., Hilton, J., Kelton, F., Miller, L.E., Simens, M., Askell, A., Welinder, P., Christiano, P.F., Leike, J., Lowe, R.J.: Training language models to follow instructions with human feedback. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.787, + 0.799 + ], + "angle": 0, + "content": "32. Papineni, K., Roukos, S., Ward, T., Zhu, W.J.: Bleu: A method for automatic evaluation of machine translation. In: ACL (2002)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.8, + 0.787, + 0.84 + ], + "angle": 0, + "content": "33. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: ICML (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.293, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SQ-LLaVA: Self-Questioning for Large Vision-Language Assistant" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.175 + ], + "angle": 0, + "content": "34. Rohrbach, A., Hendricks, L.A., Burns, K., Darrell, T., Saenko, K.: Object hallucination in image captioning. In: EMNLP (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.175, + 0.788, + 0.271 + ], + "angle": 0, + "content": "35. Sanh, V., Webson, A., Raffel, C., Bach, S.H., Sutawika, L., Alyafeai, Z., Chaffin, A., Stiegler, A., Scao, T.L., Raja, A., Dey, M., Bari, M.S., Xu, C., Thakker, U., Sharma, S., Szczechla, E., Kim, T., Chhablani, G., Nayak, N.V., Datta, D., Chang, J.D., Jiang, M.T.J., Wang, H., Manica, M., Shen, S., Yong, Z.X., Pandey, H., Bawden, R., Wang, T., Neeraj, T., Rozen, J., Sharma, A., Santilli, A., Févry, T., Fries, J.A., Teehan, R., Biderman, S., Gao, L., Bers, T., Wolf, T., Rush, A.M.: Multitask prompted training enables zero-shot task generalization. In: ICLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.271, + 0.788, + 0.312 + ], + "angle": 0, + "content": "36. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. In: NeurIPS Workshop (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.312, + 0.788, + 0.367 + ], + "angle": 0, + "content": "37. Sharma, P., Ding, N., Goodman, S., Soricut, R.: Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In: Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.366, + 0.788, + 0.393 + ], + "angle": 0, + "content": "38. Shukor, M., Dancette, C., Cord, M.: ep-alm: Efficient perceptual augmentation of language models. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.393, + 0.788, + 0.419 + ], + "angle": 0, + "content": "39. Singh, A., Natarajan, V., Shah, M., Jiang, Y., Chen, X., Batra, D., Parikh, D., Rohrbach, M.: Towards vqa models that can read. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.419, + 0.788, + 0.46 + ], + "angle": 0, + "content": "40. Sun, G., Bai, Y., Yang, X., Fang, Y., Fu, Y., Tao, Z.: Aligning out-of-distribution web images and caption semantics via evidential learning. In: Proceedings of the ACM on Web Conference 2024 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.46, + 0.788, + 0.499 + ], + "angle": 0, + "content": "41. Tofade, T., Elsner, J., Haines, S.: Best practice strategies for effective use of questions as a teaching tool. In: American journal of pharmaceutical education (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.499, + 0.788, + 0.554 + ], + "angle": 0, + "content": "42. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., Lample, G.: Llama: Open and efficient foundation language models. In: ArXiv (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.554, + 0.788, + 0.705 + ], + "angle": 0, + "content": "43. Touvron, H., Martin, L., Stone, K.R., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., Bikel, D.M., Blecher, L., Ferrer, C.C., Chen, M., Cucurull, G., Esiobu, D., Fernandes, J., Fu, J., Fu, W., Fuller, B., Gao, C., Goswami, V., Goyal, N., Hartshorn, A.S., Hosseini, S., Hou, R., Inan, H., Kardas, M., Kerkez, V., Khabsa, M., Kloumann, I.M., Korenev, A.V., Koura, P.S., Lachaux, M.A., Lavril, T., Lee, J., Liskovich, D., Lu, Y., Mao, Y., Martinet, X., Mihaylov, T., Mishra, P., Molybog, I., Nie, Y., Poulton, A., Reizenstein, J., Rungta, R., Saladi, K., Schelten, A., Silva, R., Smith, E.M., Subramanian, R., Tan, X., Tang, B., Taylor, R., Williams, A., Kuan, J.X., Xu, P., Yan, Z., Zarov, I., Zhang, Y., Fan, A., Kambadur, M., Narang, S., Rodriguez, A., Stojnic, R., Edunov, S., Scialom, T.: Llama 2: Open foundation and fine-tuned chat models. In: ArXiv (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.705, + 0.788, + 0.746 + ], + "angle": 0, + "content": "44. Triantafillou, E., Zhu, T.L., Dumoulin, V., Lamblin, P., Xu, K., Goroshin, R., Gelada, C., Swersky, K., Manzagol, P.A., Larochelle, H.: Meta-dataset: A dataset of datasets for learning to learn from few examples. In: ICLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.746, + 0.788, + 0.773 + ], + "angle": 0, + "content": "45. Vattani, A.: K-means requires exponentially many iterations even in the plane. In: Annual Symposium on Computational Geometry (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.773, + 0.788, + 0.799 + ], + "angle": 0, + "content": "46. Vedantam, R., Zitnick, C.L., Parikh, D.: Cider: Consensus-based image description evaluation. In: CVPR (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.788, + 0.84 + ], + "angle": 0, + "content": "47. Wang, J., Sun, G., Wang, P., Liu, D., Dianat, S., Rabbani, M., Rao, R., Tao, Z.: Text is mass: Modeling as stochastic embedding for text-video retrieval. In: CVPR (2024)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.359, + 0.127 + ], + "angle": 0, + "content": "G. Sun et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "48. Wang, Y., Kordi, Y., Mishra, S., Liu, A., Smith, N.A., Khashabi, D., Hajishirzi, H.: Self-instruct: Aligning language models with self-generated instructions. In: ACL (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.3 + ], + "angle": 0, + "content": "49. Wang, Y., Mishra, S., Alipoormolabashi, P., Kordi, Y., Mirzaei, A., Arunkumar, A., Ashok, A., Dhanasekaran, A.S., Naik, A., Stap, D., Pathak, E., Karamanolakis, G., Lai, H.G., Purohit, I., Mondal, I., Anderson, J., Kuznia, K., Doshi, K., Patel, M., Pal, K.K., Moradshahi, M., Parmar, M., Purohit, M., Varshney, N., Kaza, P.R., Verma, P., Puri, R.S., Karia, R., Sampat, S.K., Doshi, S., Mishra, S.D., Reddy, S., Patro, S., Dixit, T., Shen, X., Baral, C., Choi, Y., Smith, N.A., Hajishirzi, H., Khashabi, D.: Super-naturalinstructions: Generalization via declarative instructions on 1600+ nlp tasks. In: EMNLP (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.787, + 0.328 + ], + "angle": 0, + "content": "50. Wei, J., Bosma, M., Zhao, V., Guu, K., Yu, A.W., Lester, B., Du, N., Dai, A.M., Le, Q.V.: Finetuned language models are zero-shot learners. In: ICLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.787, + 0.37 + ], + "angle": 0, + "content": "51. Xu, C., Sun, Q., Zheng, K., Geng, X., Zhao, P., Feng, J., Tao, C., Jiang, D.: Wizardlm: Empowering large language models to follow complex instructions. In: ICLR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.37, + 0.787, + 0.411 + ], + "angle": 0, + "content": "52. Young, P., Lai, A., Hodosh, M., Hockenmaier, J.: From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. In: TACL (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.787, + 0.439 + ], + "angle": 0, + "content": "53. Yu, W., Yang, Z., Li, L., Wang, J., Lin, K., Liu, Z., Wang, X., Wang, L.: Mm-vet: Evaluating large multimodal models for integrated capabilities. In: ArXiv (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.787, + 0.481 + ], + "angle": 0, + "content": "54. Zhang, R., Han, J., Zhou, A., Hu, X., Yan, S., Lu, P., Li, H., Gao, P., Qiao, Y.: Llama-adapter: Efficient fine-tuning of language models with zero-init attention. In: ICLR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.787, + 0.522 + ], + "angle": 0, + "content": "55. Zhang, Y., Zhang, R., Gu, J., Zhou, Y., Lipka, N., Yang, D., Sun, T.: Llavar: Enhanced visual instruction tuning for text-rich image understanding. In: ArXiv (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.787, + 0.55 + ], + "angle": 0, + "content": "56. Zhao, B., Wu, B., Huang, T.: Svit: Scaling up visual instruction tuning. In: ArXiv (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.787, + 0.591 + ], + "angle": 0, + "content": "57. Zheng, L., Chiang, W.L., Sheng, Y., Zhuang, S., Wu, Z., Zhuang, Y., Lin, Z., Li, Z., Li, D., Xing, E.P., Zhang, H., Gonzalez, J., Stoica, I.: Judging llm-as-a-judge with mt-bench and chatbot arena. In: NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.591, + 0.787, + 0.62 + ], + "angle": 0, + "content": "58. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. In: ICLR (2024)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.62 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/SQ-LLaVA_ Self-Questioning for Large Vision-Language Assistant/3909e8d0-9729-43b7-abd9-e1c4bba247cf_origin.pdf b/2024/SQ-LLaVA_ Self-Questioning for Large Vision-Language Assistant/3909e8d0-9729-43b7-abd9-e1c4bba247cf_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..40f0d17cfa25e2f9842f992848ef77350e403353 --- /dev/null +++ b/2024/SQ-LLaVA_ Self-Questioning for Large Vision-Language Assistant/3909e8d0-9729-43b7-abd9-e1c4bba247cf_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:38edc473ca00f5493ecb94cf2436b7583953ccbdc02eba952a819034e7b565fe +size 1746308 diff --git a/2024/SQ-LLaVA_ Self-Questioning for Large Vision-Language Assistant/full.md b/2024/SQ-LLaVA_ Self-Questioning for Large Vision-Language Assistant/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7a263da8b8767a3dee63e7bdd3aa5fae10ef217f --- /dev/null +++ b/2024/SQ-LLaVA_ Self-Questioning for Large Vision-Language Assistant/full.md @@ -0,0 +1,333 @@ +# SQ-LLaVA: Self-Questioning for Large Vision-Language Assistant + +Guohao Sun1, Can Qin2, Jiamian Wang1, Zeyuan Chen2, Ran Xu2, and Zhiqiang Tao1 + +$^{1}$ Rochester Institute of Technology, Rochester, NY, US $^{2}$ Salesforce AI Research, CA, US + +Abstract. Recent advances in vision-language models have shown notable generalization in broad tasks through visual instruction tuning. However, bridging the gap between the pre-trained vision encoder and the large language models (LLMs) becomes the whole network's bottleneck. To improve cross-modality alignment, existing works usually consider more visual instruction data covering a broader range of vision tasks to fine-tune the model for question-answering, which, however, is costly to obtain and has not thoroughly explored the rich contextual information contained in images. This paper first attempts to harness the overlooked context within visual instruction data, training the model to self-supervised "learning" how to ask high-quality questions. In this way, we introduce a novel framework named SQ-LLaVA: Self-Questioning for Large Vision-Language Assistant. SQ-LLaVA exhibits proficiency in generating flexible and meaningful image-related questions while analyzing the visual clue and prior language knowledge, signifying an advanced level of generalized visual understanding. Moreover, fine-tuning SQ-LLaVA on higher-quality instruction data shows a performance improvement compared with traditional visual-instruction tuning methods. This improvement highlights the efficacy of self-questioning techniques in achieving a deeper and more nuanced comprehension of visual content across various contexts. Our code is available at https://github.com/heliossun/SQ-LLaVA. + +Keywords: Vision-Language Understanding $\cdot$ Multi-modal LLM $\cdot$ Instruction Tuning + +# 1 Introduction + +The recently emerging large vision-language methods, such as large language-and-vision assistant (LLaVA) and its variants [3,5,21,22,54], fine-tune large language models (LLM) [3,54,57] on visual instruction data [18,55,56,58] to realize diverse open-world multimodal understanding, demonstrating a surprising efficacy of visual instruction tuning - the LLM learns to perform complex vision tasks by conditioning on a prompt containing image and text clues. Existing visual instruction data is mainly built upon conversations (e.g., ChatGPT/GPT4-V [1]), consisting of images and multiple question-answer pairs. Building high-quality visual instruction data usually requires images and texts from different tasks to + +![](images/25892de293d2dfaee7812478ca6e319cee90bd564ffd006db9b542b01f00afde.jpg) +Fig. 1: (a) Comparison between visual instruction tuning and visual self-questioning (ours) for vision-language assistant. (b) The proposed SQ-LLaVA achieves state-of-the-art performance on 9 out of 10 tasks compared with other open-ended models. + +generate diverse questions, such as "Please provide the bounding box coordinate of the region this sentence describes A dead leaf on the ground" for the object detection task. Empirically, by increasing the diversity of questions, LLaVA has achieved better performance on GQA and VizWiz tasks (26% and 45% over previous state-of-the-art methods [8]). This evidence strongly suggests the advantage of training models on a broad spectrum and diverse array of tasks for enriching general vision-language understanding. + +LLaVA [21] model family usually consists of a pre-trained vision encoder (e.g., CLIP-ViT [33]), a large generative language model (LLMs like Vicuna [57], LLaMA [54], QWen [3], etc.), and a vision-to-language projector implemented by a few linear layers. However, the modality gap between the pre-trained vision encoder and the language model restricts both sides' generalization ability and feature representation. To overcome this challenge, various techniques have been proposed to align the vision and language domains, which could be roughly categorized into three groups: 1) build a more robust image feature extractor [8, 54, 58], 2) collect more high-quality training data [5, 8, 21, 58], and 3) fully fine-tune the vision and language models simultaneously during the pre-training stage [5]. While these methods have shown good progress in mitigating the domain gap, they inevitably bring higher computational costs and more expensive data collection and may also require sophisticated manual designs as well as heavy annotating efforts. Moreover, images generally encompass rich information, including color, context, and the relationships between objects, but most existing visual instruction datasets capture only a fraction. We posit that leveraging such under-explored knowledge could significantly aid in vision-language understanding. In this study, we propose a new visual self-questioning approach (Fig. 1a) by training the LLM to ask questions and discover vision clues without collecting extra data from other sources. + +Unlike existing visual instruction tuning methods that focus solely on answer prediction, the proposed visual self-questioning aims to extract relevant question + +![](images/9542d76fdec00731d017521f54070ffb255aca4d0dd2b26ef4310d0f163c853f.jpg) +Fig. 2: Left: Questions spread more samples around a higher mean CLIPScore than answers. Right: Example of highly image-relevant questions within the visual instructional dataset for training a Vision-Language assistant. + +![](images/752976e3d1808ce7bd5632448ae11110285fe1be37b01ecc68f2b49655174b3d.jpg) + +![](images/dbaf444cb5668a56e37c500aa32db41ae18770b8868f82fd94fba2cd6b5c035c.jpg) + +![](images/afaceb6cfb129dcd4a175fc60429cfcdea1e56df1f3b14e90898b116acf979f1.jpg) + +context. As illustrated in Fig. 2 right, the questions could contain more image-related information than the answers, inspiring us that aligning images and related questions may further improve the model's vision-language understanding capacity. Quantitatively, we compute the CLIPScore [33] (a higher value means better visual-text relevance) for all the image-question/-answer pairs on two visual instruction datasets (LLaVA-instruct [21] and ShareGPT4V-instruct [5]). By comparison, the questions' mean CLIPScore $\mu_q = 0.184$ is larger than the answers' $\mu_a = 0.183$ on LLaVA-instruct; $\mu_q = 0.186$ is larger than the answers' $\mu_a = 0.184$ on ShareGPT4V-instruct. Fig. 2 left shows the CLIPScore distribution, indicating questions have a similar and even better visual relevance than the answers. Therefore, questions in current visual instruction data can be used to fine-tune instructional LLMs owing to their diverse semantic information. + +This work introduces a self-questioning LLaVA, namely SQ-LLaVA, to fully utilize questions within the instruction data as an additional learning resource for training instructional LLMs and empowering the model's curiosity (questioning ability). To efficiently align the vision and language domains, we apply LoRAs [12] to optimize both the vision encoder and the instructional LLM within the SQ-LLaVA. Plus, we develop a prototype extractor to enhance visual representation by leveraging learned clusters with meaningful semantic information to improve vision-language alignment further. Extensive experiments demonstrate that SQ-LLaVA surpasses existing visual instruction tuning methods in general vision understanding (see Fig. 1b). We summarize the contributions as follows. + +- We propose a novel training technique, visual self-questioning for vision-language assistants (SQ-LLaVA), by leveraging highly relevant question contexts in instruction data. This SQ learning task promotes instructional LLMs to understand the relationship between images and questions, enhancing vision-language alignment without needing new data collection. +- We design and develop a lightweight tuning architecture for SQ-LLaVA, consisting of ViT-LoRA, LLM-LoRA, and a prototype extractor. The prototype extractor enhances vision embeddings, while ViT-LoRA and LLM-LoRA efficiently align vision and language domains during training. +- Extensive experimental results show that the proposed SQ-LLaVA leads to better performance in several tasks, including visual question-answering, visual instruction benchmarks, and zero-shot image captioning. + +# 2 Related Work + +# 2.1 Instruction Tuning + +Instruction tuning emerged as a pivotal methodology within the realm of natural language processing (NLP), facilitating Large Language Models (LLMs) such as GPT-3 [4], PaLM [7], and LLaMA [42] to interpret and execute human language instructions across a spectrum of NLP tasks. This approach diverges from traditional fine-tuning mechanisms by incorporating a specialized data structure, termed instruction-following data [50], which is instrumental in the fine-tuning process of LLMs. Generally, there are two main categories regarding instruction tuning methods - 1) closed-domain and 2) open-domain. The closed-domain instruction tuning [35,50,51] studies engaged LLMs with a comprehensive assortment of publicly accessible datasets and subsequently assessed their performance across diverse NLP tasks [44]. The empirical evidence from these inquiries consistently indicated that integrating varied NLP task instructions significantly augments the LLMs' efficacy in navigating novel tasks. Nonetheless, LLMs calibrated with such closed-form instructions exhibited limitations in real-world user scenarios, prompting the development of an alternative approach. To address these constraints, the concept of open-domain instruction tuning [30, 49] is conceived. OpenAI pioneered this approach by employing human annotators to compile a corpus of real-world question-answer datasets. These datasets form the foundation for training a reward model through reinforcement learning methodologies. The trained reward model then functions as a supervisory mechanism for further training instruction-oriented language models, such as InstructGPT [31] and Vicuna [57]. This innovation marks a significant advancement in the field, aiming to bridge the gap between LLM performance and real-world applicability by leveraging instruction data derived from authentic user interactions. + +# 2.2 Large Vision Language Models + +As the field of LLMs and instruction tuning undergoes rapid advancements, the academic research community is progressively focusing on integrating visual information into LLMs to facilitate visual instruction tuning. This emerging area of research has witnessed the development of various methodologies [3,5,21,22,54] based on foundational vision-language models [19,33,40,47] and diverse LLM architectures [3,43,57]. In particular, LLaVA [22] pioneers the integration of an LLM with a CLIP vision encoder to construct a vision language model, demonstrating remarkable capabilities in image-text dialogue tasks through pretraining alignment strategies and targeted instruction tuning. Subsequent investigations have sought to refine visual instruction tuning by enhancing the quality and variety of the datasets used during the pre-training and fine-tuning phases. Building upon these advancements, recent studies like LLaVA-v1.5 [21] and ShareGPT4V [5] have achieved notable success in general vision-language comprehension, showcasing their ability to undertake complex question-answering tasks. This progression underscores the importance of sophisticated data handling and model tuning strategies in developing effective vision-language models. + +![](images/b7b262649a963a2c222fa00777c5e37a3722f0df3764ae09d3ef66c868ccb104.jpg) +Fig. 3: Model architecture of SQ-LLaVA. We propose prototype extractor to extract visual clustering information to enhance the visual embedding encoded by the visual encoder. SQ-LLaVA defines a new token [vusr] as specific instruction for LLM to perform visual self-questioning. Besides question answering, SQ-LLaVA treats questioning as another training objective. + +# 3 Method + +# 3.1 Architecture Overview + +The proposed SQ-LLaVA model (see Fig. 3) consists of four main components: 1) A pre-trained vision encoder CLIP-ViT [33] that extracts a sequence embedding of image tokens $Z_{v}$ for an input image $X_{v}$ ; 2) A prototype extractor $\phi(\cdot)$ learning visual clusters to enhance the original image tokens; 3) A trainable projection block $W(\cdot)$ with two linear layers to map the enhanced image tokens to the language domain tokens $H_{v}$ , handling the dimension misalignment between the vision and language domain; and 4) Our LLM backbone $f(\cdot)$ implemented by the pre-trained Vicuna [57] to predict the next token upon the previous embedding sequence. Given the input question $X_{q}$ and answer $X_{a}$ , a word embedding matrix is used to map them to contextual embeddings $H_{q}$ and $H_{a}$ , and the distribution over $H_{a}^{(i+1)}$ can be obtained following the auto-regressive model as: + +$$ +p _ {\theta} \left(H _ {a} ^ {(i + 1)} \mid H _ {v}, H _ {q}, H _ {a} ^ {(1: i)}\right) = \sigma \left(f \left(H _ {v}, H _ {q}, H _ {a} ^ {(1: i)}\right)\right), \tag {1} +$$ + +where $\theta$ represents all the trainable parameters in our model, $\sigma(\cdot)$ is a softmax function, and $f(\cdot)$ outputs the last token embedding of the whole sequence. We denote $p_{\theta}$ as the prediction probability for the anticipated answer token $H_{a}^{(i+1)}$ at the position $i+1$ , conditioning on the input image token embedding $H_{v}$ , the question token embedding $H_{q}$ , and the previous answer token embeddings $H_{a}^{(1:i)}$ . As shown in Eq. (1), the proposed SQ-LLaVA applies the language model $f(\cdot)$ to model $p_{\theta}$ given by $H_{v}$ , $H_{q}$ , and $H_{a}^{(1:i)}$ . Existing visual instruction tuning methods [5,21] are only able to predict $H_{a}$ , yet cannot fully exploit the rich semantic clues within $H_{q}$ . In this study, we propose visual self-questioning instructions to guide LLMs in capturing the visual content within questions. + +System-message + +$$ +\mathbf {X} _ {c} ^ {(1)}: [ \mathrm {u s r} ]: \mathbf {X} _ {v} \backslash \mathrm {n} \mathbf {X} _ {q} ^ {(1)} [ \mathrm {a s w r} ]: \mathbf {X} _ {a} ^ {(1)} < o ^ {d} > \quad \mathbf {X} _ {c} ^ {(2)}: [ \mathrm {v u s r} ]: \mathbf {X} _ {q} ^ {(2)} < o ^ {d} > [ \mathrm {a s w r} ]: \mathbf {X} _ {a} ^ {(2)} < o ^ {d} > +$$ + +$$ +\mathbf {X} _ {c} ^ {(3)}: [ \text {u s r} ]: \mathbf {X} _ {q} ^ {(3)} [ \text {a s w r} ]: \mathbf {X} _ {a} ^ {(3)} < _ {O} d > \quad \dots \quad \mathbf {X} _ {c} ^ {(P)}: [ \text {v u s r} ]: \mathbf {X} _ {q} ^ {(P)} < _ {O} d > [ \text {a s w r} ]: \mathbf {X} _ {a} ^ {(P)} < _ {O} d > +$$ + +Fig. 4: The input sequence used to train SQ-LLaVA. Our model is trained to predict question, answer, and where to stop. We use tokens to represent learnable tokens, where $X_{q}$ is the question, $X_{a}$ is the answer, and $< o^{d}>$ is the delimiter token. In SQ-LLaVA, the System-message = "The assistant gives helpful, detailed, and polite answers to the user's questions. Also, the assistant is a curious virtual user can ask complex questions that are relevant to the content in the image." + +# 3.2 Visual Self-questioning Instruction + +In broad real-world scenarios, proactively asking a question requires more understanding and background knowledge than answering [41]. Similarly, this work proposes visual self-questioning to encourage the LLM to discover deeper vision-language alignment and improve the overall instruction-following performance. Particularly, SQ-LLaVA treats questioning as a new learning objective beyond answering, which, to the best of our knowledge, is the first practice in the field of visual instruction tuning. While the current vision-language model can ask questions [1], such skill is still learned from question-answering through instruction tuning. However, our proposed method shows that the decoder-based LLM has the potential to learn more skills, such as how to ask questions spontaneously when a unique instruction token is given (e.g., we define [vusi] in our work). Furthermore, visual self-questioning can potentially improve general vision-language understanding. To be specific, as shown in Fig. 1, there are a certain amount of questions containing more meaningful image-related information than answers in the existing visual instruction data [5, 21]. Thus, we hypothesize that the vision-language understanding can be improved once the LLM learns how to predict relevant questions about a given image. + +Self-Questioning Prompt Design. We provide ground-truth content for the visual self-questioning, restricting SQ-LLaVA from asking image-related questions. To this end, we leverage questions as another learnable resource and follow the regular auto-regressive training objective. As shown in Fig. 4, the training data for SQ-LLaVA is designed in a format with a pre-defined template. To be specific, the system message as a fixed prompt is added at the beginning of each instruction data, indicating a general job description (e.g., gives helpful answers and asks complex questions) for the LLM. Existing visual instruction tuning methods utilize the unique tokens ([usr], [aswr]) to give the LLM a particular instruction (e.g., questioning understanding, answer prediction, etc.) and apply the delimiter token $< \sigma^d>$ to mark the ending position. In this work, we propose a new special token [vusr], indicating a specific instruction - asking questions. Combining with the delimiter, we can construct instructions for self-questioning as a new training task. + +Each sample of current visual instruction data consists of one image $X_{v}$ and $P$ question-answer pairs $(X_{q}^{(1)},X_{a}^{(1)},\dots ,X_{q}^{(P)},X_{a}^{(P)})$ . We collect one question $X_{q}^{(j)}$ and its answer $X_{a}^{(j)}$ with special tokens to construct the $j^{th}$ turn conversation as + +$$ +X _ {c} ^ {j} = \left\{ \begin{array}{l l} \left([ \mathrm {u s r} ], X _ {q} ^ {(j)}, [ \mathrm {a s w r} ], X _ {a} ^ {(j)}\right) & j = 1 \text {o r} j > 1, R < \delta \\ \left([ \mathrm {v u s r} ], X _ {q} ^ {(j)}, [ \mathrm {a s w r} ], X _ {a} ^ {(j)}\right) & j > 1, R > \delta \end{array} , \right. \tag {2} +$$ + +where $R \in [0,1]$ is a random number and $\delta = 0.5$ is a threshold that sets the proportion of self-questioning pairs in the conversations. Finally, the full-text input $X_{c}$ will be mapped to textual embedding $H_{c}$ through word embedding. + +SQ-LLaVA performs zero-shot questioning without any in-context knowledge or human language instruction since the only instructional prompt is a unique token [vusr]. After visual self-questioning on instruction data with various question formats $X_{q}$ , the questions sampled by SQ-LLaVA are more diversified than GPT4-V (as shown in Fig. 5), since the LLM has learned alignment between image and questions, which is different from previous works [1, 48]. Specifically, previous general-purpose vision language models such as GPT4-V [1] can generate questions based on a given image, but it requires explicit language instruction such as "Ask complex questions that are relevant to the content in the image". Accordingly, the quality of generated questions highly relies on prompt engineering. Also, self-instruct [48] utilizes in-context learning to prompt the LLM to ask specific questions. + +# 3.3 Enhanced Visual Representation + +Unlike previous visual instruction tuning methods, SQ-LLaVA jointly benefits from visual self-questioning and question-answering. For better visual self-questioning, we develop a prototype extractor that recognizes and groups similar patterns of visual information from the latent space. Our primary goal is to enhance visual representation through prototype learning. + +Specifically, we utilize clustering to extract centroid representations of image tokens $Z_{v}$ , where each cluster center is treated as a prototype, which, in return, will be distributed to each of the original image token embeddings. Our proposed prototype extractor $\phi(\cdot)$ is a lightweight design involving two parts: 1) cluster center optimization and 2) prototype information distribution. Following [14, 45], we randomly initialize $K = 256$ cluster centers $C$ and deploy the iterative Expectation-Maximization (EM) clustering process to capture representative semantics in the latent space by + +$$ +\begin{array}{l} \operatorname {E} \text {- s t e p}: \mathcal {M} ^ {(t)} = \sigma \left(q \left(C ^ {(t)}\right) * k \left(Z _ {v}\right) ^ {\top}\right), \\ \text {M - s t e p}: C ^ {(t + 1)} = \mathcal {M} ^ {(t)} * v \left(Z _ {v}\right), \tag {3} \\ \end{array} +$$ + +where $\mathcal{M}^{(t)}\in [0,1]$ denotes a soft cluster assignment map at the $t^{th}$ iteration $\sigma (\cdot)$ is a softmax function, and $t\in \{1,\dots ,T\}$ indexes the iteration of EM step with $T = 2$ in this work. Three trainable linear layers $q,k$ , and $v$ are used in (3.3) + +where $q(\cdot)$ projects the prototype $C$ to a query vector, and $k(\cdot)$ and $v(\cdot)$ project $H_{v}$ into key and value vectors, followed by a normalization layer, respectively. The prototype extractor iteratively updates cluster map $\mathcal{M}$ and centers $C$ . + +After the prototype extraction, we train a linear layer $z(\cdot)$ to adaptively map the visual cluster information to the raw image embedding $Z_{v}$ . For the $i^{th}$ token embedding $Z_{v}^{(i)}$ , we update it as + +$$ +Z _ {v} ^ {(i)} = Z _ {v} ^ {(i)} + z \left(\frac {1}{K} \sum_ {j = 1} ^ {K} S _ {c} \left(C _ {j}, Z _ {v} ^ {(i)}\right) \times C _ {j}\right), \tag {4} +$$ + +where $S_{c}(\cdot, \cdot)$ is a normalized cosine similarity function. The weighted sum over prototypes in Eq. (4) emerges as an indispensable step for contextual understanding from image tokens, recognizing and grouping similar patterns and semantics. It clusters image tokens as prototypes that display homogeneity in semantics, such as "grass" and "dog". The prototypes can describe the intrinsic semantic meanings by aggregating entities that exhibit shared attributes. Finally, we map the image sequence embedding $Z_{v}$ to language domain $H_{v}$ with a two-layer linear projector $W(\cdot)$ . + +# 3.4 Model Training + +Stage1: Pre-training for Vision-Language Alignment. Unlike text-only LLMs, the vision-language model also fine-tunes LLM using image tokens as input (see Fig. 3). Therefore, the pre-training stage aims to optimize the LLM by explicitly executing the visual instruction. The proposed SQ-LLaVA adopts Vicuna [57] as its instruction LLM, pre-training on massive text corpora to predict the next text token given the previous context, not only containing text but also visual instructions. To achieve this, we organize the pre-training data as $D_{PT} = \{[X_v^{(1)}, X_a^{(1)}], \dots, [X_v^{(N)}, X_a^{(N)}]\}$ , where $N$ is the total number of training samples, and each sample has an image and its related descriptions. Each image and text input pair will be mapped to sequence embeddings ( $H_v$ and $H_a$ ) as elaborated in Section 3.1. During pre-training, we freeze the vision encoder and LLM and mainly train the prototype extractor $\phi$ and the vision-to-language projector $W$ . The pre-training goal is to maximize the probability of the predicted image description $H_a$ given an image $H_v$ . When training a visual instructional LLM, we follow the negative log-likelihood objective function as + +$$ +\sum_ {v, a \in D _ {P T}} - \log p _ {\theta} (H _ {a} \mid H _ {v}) = \sum_ {v, a \in D _ {P T}} \sum_ {i = 1} ^ {L} - \log p _ {\theta} (H _ {a} ^ {(i + 1)} \mid H _ {v}, H _ {a} ^ {(1: i)}), \quad (5) +$$ + +where $L$ denotes the sequence length of answer tokens in $H_{a}$ , $\theta$ is the total trainable parameter of $\phi$ and $W$ , $p(H_{a} \mid H_{v})$ can be computed by Eq. (1), and $H_{a}^{(1:i)}$ represents all the answer tokens before the current prediction $H_{a}^{(i+1)}$ . + +Stage2: Fine-tuning. Existing methods, such as LLaVA [5,22], mainly update the vision-to-language projector (usually a couple of linear layers) and the + +language model during fine-tuning. Nevertheless, the projector might be too weak to capture the relationship between the image and the questions. Following the previous multi-modal understanding method [33], we unfreeze the vision encoder and LLM during fine-tuning for a joint optimization further to eliminate the gap between the vision and language domain. + +To mitigate the heavy computational overhead, we take advantage of LoRA [12] as a lightweight tuning option that can achieve similar (even better) performance to fully fine-tuning when training large models on a relatively small dataset. We add LoRA in both the vision encoder and LLM. Thus, the learnable parameters $\theta$ of the proposed SQ-LLaVA during fine-tuning represent a combination of all the parameters of LLM-LoRA, ViT-LoRA, prototype extractor $\phi$ , and the vision-to-language projector $W$ . Given the instruction tuning data $D_{IT} = \{[X_v^{(1)}, X_c^{(1)}], \dots, [X_v^{(N)}, X_c^{(N)}]\}$ , we take the conversational data $X_c$ and the image $X_v$ as input, mapping them to sequence embedding $(H_c$ and $H_v)$ as elaborated in Section 3.1, and minimize the negative log-likelihood loss for the self-questioning and answering tasks as follows + +$$ +\text {S e l f - q u e s t i o n i n g}: \sum_ {v, c \in D _ {I T}} - \log p _ {\theta} \left(H _ {q} ^ {(j + 1)} \mid H _ {v}, H _ {c} ^ {(1: j)}\right), \tag {6} +$$ + +$$ +\text {A n s w e r i n g}: \sum_ {v, c \in D _ {I T}} - \log p _ {\theta} \left(H _ {a} ^ {(j + 1)} \mid H _ {v}, H _ {c} ^ {(1: j)}, H _ {q} ^ {(j + 1)}\right), \tag {7} +$$ + +where $j \in \{1, \dots, P\}$ , indicating the index of question or answer within the conversational data $X_{c}^{(*)}$ . Notably, previous works [5,21,58] only involve answering tasks, but we introduce visual self-questioning as an additional training task for visual instruction tuning. Eventually, SQ-LLaVA, as a vision-language assistant, not only executes human instructions by optimizing the objective function in Eq. (7) but can raise questions out of the given image after optimizing the Eq. (6). This capability potentially yields more diverse question-answer guidance and enhances multi-modal understanding. + +# 4 Experiments + +# 4.1 Experimental Setting + +Dataset. Our work uses the open-source visual instruction dataset provided by LLaVA [21] and ShareGPT4V [5] for training. Each dataset has large-scale image-text paired data for pre-training and instruction-following data for finetuning. Specifically, the instruction data proposed by LLaVA is a comprehensive mixture of COCO, GQA [13], OCR-VQA [27], TextVQA [39], VisualGenome [17], RefCOCO [15], and image from InstructBLIP [8] involves multiple reasoning, spatial understanding, multi-step inference, optical character recognition, and grounding of visual concepts to language. The ShareGPT4V dataset consists of the same mixture image as LLaVA but enrolls more images from other datasets such as SAM [16] and Data [29,36]. + +We evaluate our model on general vision-language understanding tasks. Specifically, we use ten visual oriented question answering benchmarks: $\mathrm{VQA}^{v2}$ [10]; GQA [13]; VizWiz [11]; $\mathrm{SQA}^I$ [25]: ScienceQA-IMG; $\mathrm{VQA}^T$ : TextVQA [39]; POPE [20]; MM-Vet [53]; $\mathrm{LLaVA}^W$ [22]: LLaVA (in the wild); MMB: MMBBenchmark [23]; $\mathrm{MMB}^{CN}$ : MMBench-Chinese [23]. We report the prediction accuracy on all the benchmarks. We also evaluate our model for visual information discovery through captioning. To be specific, we employ testing images from four datasets, i.e., COCO [6], Flickr30K [52], Conceptual [37], and Nocaps [2]. Following [9], we evaluate the proposed method with regular image captioning metrics, e.g., BLEU [32] and CIDEr [46]. For the open-world methods LLaVA-v1.5 [21], ShareGPT4V [5] and the proposed SQ-LLaVA, we utilize greedy search for caption generation with a prompt of "Provide a brief description of the given image, your answer should be in one sentence." The generated caption is used to evaluate the performance of image captioning. + +# 4.2 Implementation + +We adopt Vicuna [57] as the pre-trained language generative model and CLIP-ViT [33] as the vision encoder. We pre-train the prototype extractor and the vision-to-language projector using AdamW [24] optimizer with a learning rate of $2 \times 10^{-3}$ and a constant scheduler for one epoch. Following previous work [21], we keep the global batch size as 256 for pre-training and 128 for fine-tuning. During fine-tuning, we insert LoRA [12] with rank = 128 and $\alpha = 256$ into the language model (LLM-LoRA) and LoRA with rank = 32 and $\alpha = 64$ into the vision encoder (ViT-LoRA). We optimize LoRA modules, the prototype extractor, and vision-to-language projector for one epoch. We set the learning rate to $2 \times 10^{-4}$ for LoRA, and $2 \times 10^{-5}$ for the other layers. All the weights of the pre-trained language model and vision encoder remain fixed during fine-tuning. + +# 4.3 Zero-shot Multilingual Capability + +We evaluate SQ-LLaVA on ten benchmarks, covering a range of academic Visual Question Answering (VQA) tasks and recent instruction tuning tasks designed for large vision language models. The academic VQA includes VQA-v2 [10] and VizWiz [11]. GQA [13] is a fine-grained real-world visual reasoning and question-answering benchmark. ScienceQA [25] is a benchmark with rich subjects (natural science, language science, and social science). TextVQA [39] requires the model to recognize the texts in the image. LLaVA (in the wild) and MM-Vet [53] use GPT4 to assess the capability of the models for testing. With manually designed questions, MM-Bench and MMBench-CN [23] evaluate the model's vision-related reasoning and perception for English and Chinese, respectively. POPE [20] is a benchmark for evaluating object hallucination [34]. + +In Table 1, we quantitatively compare between SQ-LLaVA and existing models. SQ-LLaVA-7B and SQ-LLaVA-13B trained on two instruction datasets [5,21] outperform previous methods in six out of ten visual instruction tuning tasks. To be specific, SQ-LLaVA-7B achieves $17.2\%$ improvement over LLaVA-v1.5-7B + +Table 1: Comparison with state-of-the-art methods on ten benchmarks. After training on the same instruction data, SQ-LLaVA train on [21] and SQ-LLaVA* trained on [5] surpass their baseline model LLaVA-v1.5 and ShareGPT4V on 9 out of 10 and 6 out of 10 benchmarks in the 7B scale, and 8 out of 10 and 6 out of 10 in the 13B scale. The best results are bold and the second-best results are underlined. + +
LLMModelVQA v2GQAVizWizSQA1VQA TPOPEMM-VetLLaVA WMMBMMB CN
7BInstructBLIP [8]-49.234.560.550.1-26.260.936.023.7
Qwen-VL [3]78.859.335.267.163.8---38.27.4
Qwen-VL-chat [3]78.257.538.968.261.5---60.656.7
LLaVA-v1.5 [21]78.562.050.066.858.285.930.563.464.358.3
ShareGPT4V [5]80.663.357.268.460.486.837.672.668.862.2
SQ-LLaVA79.262.854.068.958.687.732.566.366.258.1
SQ-LLaVA*80.363.755.370.560.587.237.674.366.660.0
13BInstructBLIP [8]-49.533.463.150.778.925.658.2--
LLaVA-v1.5 [21]80.063.353.671.658.285.935.470.767.763.6
ShareGPT4V [5]81.064.855.671.262.2-43.179.968.563.7
SQ-LLaVA80.163.654.669.860.287.735.574.668.762.0
SQ-LLaVA*81.365.058.271.561.987.439.780.768.562.5
+ +on the LLaVA (in the wild) benchmark, indicating the superior capabilities of our model in tasks such as detailed description and complex reasoning. Also, SQ-LLavA-7B improves over previous methods on ScienceQA, indicating that our model excels in understanding and reasoning over scientific content and can effectively handle multi-modal information. The improvement in ScienceQA suggests strong capabilities in multi-hop reasoning, comprehension of complex scientific concepts, and the ability to utilize context and explanations to derive correct answers. SQ-LLaVA-7B has a steady improvement over LLaVA-v1.5-7B and ShareGPT4V-7B on the POPE benchmark, and the $2\%$ and $1\%$ improvement indicates that our proposed method has better reliability and trustworthiness since POPE is a task designed to evaluate the phenomenon of object hallucination [20, 34]. In the bottom section of Table 1, the proposed SQ-LLaVA-13B surpasses previous works in six out of ten benchmarks, indicating the scalability of our method on larger LLM. Notably, the performance inconsistency on some datasets might be due to the unsupervised prototype extractor (lacking pixel-wise guidance) in our model. To mitigate this issue, we could leverage pseudo object masks (e.g., given by the pre-trained segment anything) in learning prototypes. Despite this limitation, all the improvements are achieved with significantly fewer trainable parameters compared to other methods [3, 5, 21]. + +# 4.4 Visual Information Discovery + +In this experiment, we showcase the diversity and reliability of the proposed SQ-LLaVA through various qualitative applications, including detailed image description, visual information summary, and visual self-questioning. We also present quantitative results on the task of image captioning. + +Abilities of SQ-LLaVA Through Qualitative Samples. SQ-LLaVA exhibits numerous advanced capabilities compared to traditional vision-language models. + +![](images/136c42a072a95726e4d9809d77f134ccfcb9ee16686c2c4f4b300baa8370bb1d.jpg) +GPT4-V Generated + +$Q_{1}$ : What makes this particular toothbrush stand out among others in the image? $Q_{2}$ : What color is the background? +Q2: What color is the background? +Q3: What is this product? + +# SQ-LLaVA Visual Self-Questioning + +$SQ_{1}$ : What are some of the key aspects that this image focuses on? +$S Q_{2}$ : What is the primary color of the toothbrush? +$S Q_{3}$ : What is the expression of the face drawn on the toothbrush? + +![](images/ece03bb4b799ad9b0528600b4496451126f4c03af0cddf96704da72e8299f59b.jpg) +GPT4-V Generated + +$Q_{1}$ : What is the setting of the image? $Q_{2}$ : What are some activities that people are engaging in within the image? $Q_{3}$ : What are some benefits of spending time outdoors in a beachside setting, especially on a sunny day? + +# SQ-LLaVA Visual Self-Questioning + +$S_{Q_1}$ : What is the hair color of the man that is looking up at the ocean? +$SQ_{2}$ : What color is the man's shirt? +$SQ_{3}$ : In which part is the bench, the top or the bottom? + +![](images/70b7a2b0d2bb65837892e832bb74fd4f3e7053a174d41ef5baa6557c83736bc7.jpg) +GPT4-V Generated + +$Q_{1}$ : What type of environment can the sheep be seen in? + +$Q_{2}$ : What are the sheep doing in the image? $Q_{3}$ How do sheep contribute to the ecosystem of a meadow or field? + +# SQ-LLaVA Visual Self-Questioning + +$SQ_{1}$ : In the vicinity of this group of sheep, how many horses are there? a) Five b) Four c) One d) Zero +$S_{Q2}$ : In this image, what kind of animals can you see in the green field? $S_{Q3}$ : Are there any cows in the picture? + +![](images/de62dffe85e775b4f6f0a44f297a821f4725bd1c697cceac3544561065315f45.jpg) +GPT4-V Generated +Fig.5: Visual self-questioning of SQ-LLaVA-7B. Comparing to the question data provided by GPT4-V (data collected by LLaVA-v1.5 [21]), SQ-LLaVA can generate questions with higher diversity i.e. multiple choice and tricky questions. + +# SQ-LLaVA Visual Self-Questioning + +$S_{Q1}$ : The two small children are sitting on a tricycle. Which place is it? Town or village? +$SQ_{2}$ : What can you tell me about the child's age? +$S_{Q3}$ : What does the boy's action of swinging a bat suggest about his play or sport? + +Table 2: Comparison of image captioning on four datasets. SQ-LLaVA is trained on data collected by [21]. SQ-LLaVA* is trained on data collected by [5]. + +
ModelFlickr30kNocapsoutConceptual
B@4CIDErB@4CIDErB@4CIDEr
ClipCap [28]17.2141.6520.3251.741.4723.74
DiscriTune [9]18.4844.7824.1057.061.7128.01
LLaVA-v1.5 [21]28.6781.2735.78103.562.7939.20
ShareGPT4V [5]31.0086.1737.19107.452.7837.86
SQ-LLaVA29.8883.5136.21105.422.9039.49
SQ-LLaVA*31.4983.1437.20107.422.9141.24
+ +(a) Zero-shot image captioning + +
ModelZero-shotFine-tune
B@4CIDErB@4CIDEr
ClipCap8.5037.0332.60108.55
DiscriTune13.9953.2032.31105.40
eP-ALM [38]29.4797.2233.35111.63
MAPL [26]12.3054.3036.45125.20
LLaVA-v1.529.96111.46--
SQ-LLaVA29.85110.7740.76136.78
+ +(b) Image captioning on COCO + +Notably, SQ-LLaVA effectively mitigates object hallucination [20, 34], resulting in predictions that are more trustworthy. Additionally, SQ-LLaVA is capable of generating diverse and meaningful questions about a given image without requiring human textual instructions. As illustrated in Fig. 5, the generated questions include multiple-choice, reasoning, binary choice questions, and the content within each question, all of which are either highly correlated or tricky to the given image, such as "What is the hair color of the man that is looking up at the ocean?" from the third image and "Are there cows in the picture?" from the second image. These examples highlight SQ-LLaVA's ability to produce questions with high diversity and quality. + +The detailed image description example in Fig. 6 demonstrates that SQ-LLaVA can generate visual information with concrete concepts within the image, such as "Hyundai", "small white dog", "Nikky Stephen". In contrast, LLaVA-v1.5 can + +![](images/507a676ade812af4fb871183548354e4209fede95d1848284cab55c6f5f0276f.jpg) +Detailed Image Description +Fig. 6: A qualitative evaluation of detailed image descriptions from three models. We highlight the words and sentences that represent how each model describes the main object in the image. + +SQ-LLaVA-7B (Ours): In this black and white photograph, a man is engrossed in the task of inspecting the engine of a Hyundai car. The car, with its hood open, reveals the intricate details of its engine. The man, dressed in a casual attire of a t-shirt and jeans, is leaning over the car, his attention completely absorbed by the engine. The car's license plate, reading "VK8-LQ17", is clearly visible. Adding a touch of companionship to the scene, a small white dog is seen standing next to the car, seemingly curious about the man's actions. The photograph is credited to Nikky Stephen, as indicated by the text at the top of the image. The overall scene paints a picture of a typical day in the life of a car enthusiast. + +LLaVA-v1.5-7B: The image features a man standing next to a car, with the hood open. He is leaning over the car, possibly working on it or checking something under the hood. A dog is standing nearby, attentively watching the man's actions. The car is parked in a driveway, and the scene appears to be a casual, everyday moment between the man and his dog. + +ShareGPT4V-7B: The man, dressed in a casual attire of a t-shirt and khaki pants, stands to the right of the car, his attention completely absorbed by the task at hand. The car's license plate, reading 'YKL8-O17', is clearly visible. Adding a touch of companionship to the scene, a white dog stands obediently at the man's feet, perhaps waiting for a command or simply enjoying the company. The photograph is taken from a low angle, giving a unique perspective to the scene. The man, the car, and the dog are the main subjects of this image, each contributing to the narrative of a moment frozen in time. The black and white color scheme adds a timeless quality to the photograph, highlighting the contrast between the man, the car, and the dog. + +![](images/24fb97b1af7b0ebf481d7b8082670b2d571ae71e0da48c4761af0547fafc2cea.jpg) +Detailed Image Description + +SQ-LLA-7B (Ours): The image captures a scene on a city street. Dominating the foreground is a traffic sign, painted in striking red and white colors. The sign features a black silhouette of a duck, a universal symbol for caution. Above the duck, there's a yellow light, indicating a stop signal. The background reveals a modern building with large windows, reflecting the urban environment. A tree, standing in front of the building, adds a touch of nature to the scene. The image does not contain any discernible text. The relative positions of the objects suggest that the sign is closer to the viewer, while the building and the tree are further away. + +LLaVA-v1.5-7B: The image features a street scene with a unique combination of traffic signs. There is a stop sign with a duck on it, which is an unusual sight. The duck is positioned in the middle of the stop sign, making it stand out. In the background, there are several cars parked or driving along the street. The cars are of various sizes and are scattered throughout the scene. The presence of the duck on the stop sign adds a touch of humor to the otherwise ordinary street scene. + +ShareGPT4V-7B: The image captures a scene on a city sidewalk. Dominating the foreground is a traffic sign, painted in a striking red and white color scheme. The sign is triangular, a common shape for traffic signs, and it features a black duck in the center, adding a touch of whimsy to the otherwise mundane object. Just behind the duck sign, there's another sign. This one is diamond-shaped and also red and white, but it's adorned with a green traffic light. The green light stands out against the red and white background, drawing attention to its presence. The signs are affixed to a gray pole, which is firmly planted in the sidewalk. + +only describe the image from a general concept, such as "car", "dog", or "everyday moment". Even though ShareGPT4V can generate detailed descriptions, it still suffers from the issue of object hallucination, such as "perhaps waiting" and "behind the duck sign". By observation, SQ-LLaVA can describe the image with less unintended text, yielding higher reliability. + +Quantitative Analysis. SQ-LLaVA serves as a general-purpose vision-language model and enables zero-shot image captioning. As indicated in Table 2a, SQ-LLaVA achieves $73\%$ and $66\%$ averaged improvement over ClipCap and DiscriTune on all datasets, indicating the effectiveness of visual instruction tuning. Compared with the baseline model LLaVA-V1.5 [21], SQ-LLaVA achieves $2\%$ averaged improvement on all datasets, indicating the effectiveness of visual self-questioning. Also, we find SQ-LLaVA* surpasses the baseline model ShareGPT4V [5] on Nocaps out and Conceptual dataset, demonstrating the adaptability of the proposed method on unseen testing images from new domains. Moreover, as shown in Table 2b, SQ-LLaVA can be easily adapted to COCO captioning via instruction tuning on short descriptions. + +# 4.5 Ablation Study + +In Table 3, we conduct experiments with different architecture designs and training strategies on five question-answering benchmarks. For a fair comparison, we train the baseline models on our local machine with the same training recipe of LLaVA-LoRA [21]. Specifically, we present our full model and three ablated models by removing one component each time. We adopt the dataset [21] with $558\mathrm{k}$ for pre-training (PT) and $665\mathrm{k}$ for fine-tuning (FT). As compared, self + +Table 3: Ablation study of training strategy on visual instruction tasks. All models are in 7B scale with three components of ViT-LoRA (V-LoRA), self-questioning (SQ), and prototype extractor (Proto). We provide experiments on two instruction datasets [5,21] with different pre-training (PT) and instruction tuning (IT) data scales. + +
PTITV-LoRASQProtoVizWiz\( \mathrm{SQA}^I \)\( \mathrm{VQAT} \)POPE\( \mathrm{LLaVA}^W \)Avg.
558K665KXXX49.468.458.286.567.165.9
✓X✓52.467.958.687.765.666.4
✓✓X52.668.457.888.267.366.9
X✓✓53.469.358.187.967.967.3
✓✓✓54.068.958.687.768.167.5
1200K700KXXX51.568.958.986.872.167.6
✓X✓54.068.960.287.271.668.4
✓✓X55.469.259.586.877.369.6
X✓✓54.270.360.587.572.769.0
✓✓✓55.370.560.587.274.369.6
+ +questioning (SQ) brings a consistent performance boost on all the five benchmarks, indicating the effectiveness of visual self-questioning on improving visual language understanding. Besides, we introduce the prototype extractor (Proto) to enhance visual representation, achieving $0.9\%$ improvement in average accuracy among five benchmark. With all three components incorporated, we observe a $2.4\%$ improvement in average accuracy. + +As shown by the bottom block of Table 3, we conduct experiments with the same ablation settings but with a larger scale of the visual instruction data [5] (i.e., for both PT and IT). Overall, SQ-LLaVA achieves $2.4\%$ improvement over the baseline model after training on the smaller dataset and achieves $3.0\%$ improvement after training on the larger dataset. + +# 5 Conclusions + +This work has introduced SQ-LLaVA, a new visual instruction tuning method that enhances general-purpose vision-language understanding and image-oriented question answering through visual self-questioning. Our experiments demonstrate that SQ-LLaVA achieves superior performance with fewer training parameters and instructional data. We have also conducted a comprehensive study on visual discovery/reasoning tasks and found that SQ-LLaVA generalizes well to a wide range of unseen tasks and outperforms several state-of-the-art methods. Qualitative assessments show that SQ-LLaVA strengthens visual representation and domain alignment, effectively reducing object hallucination and improving the semantic interpretation of images. Our findings highlight the potential of visual self-questioning as a powerful training strategy for the visual instruction tuning framework, paving the way for realizing more efficient and effective large vision-language models. Particularly, SQ-LLaVA frames questioning as an intrinsic goal of tuning LLMs, encouraging the exploration of the model's curiosity (the ability to ask questions proactively) in solving complex problems. + +# References + +1. Gpt-4v(ison) system card (2023) +2. Agrawal, H., Desai, K., Wang, Y., Chen, X., Jain, R., Johnson, M., Batra, D., Parikh, D., Lee, S., Anderson, P.: Nocaps: Novel object captioning at scale. In: ICCV (2019) +3. Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. In: ArXiv (2023) +4. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., Ziegler, D., Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Chess, B., Clark, J., Berner, C., McCandlish, S., Radford, A., Sutskever, I., Amodei, D.: Language models are few-shot learners. In: NeurIPS (2020) +5. Chen, L., Li, J., wen Dong, X., Zhang, P., He, C., Wang, J., Zhao, F., Lin, D.: Sharegpt4v: Improving large multi-modal models with better captions. In: ArXiv (2023) +6. Chen, X., Fang, H., Lin, T.Y., Vedantam, R., Gupta, S., Dollar, P., Zitnick, C.L.: Microsoft coco captions: Data collection and evaluation server. In: ArXiv (2015) +7. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S., Schuh, P., Shi, K., Tsvyashchenko, S., Maynez, J., Rao, A., Barnes, P., Tay, Y., Shazeer, N.M., Prabhakaran, V., Reif, E., Du, N., Hutchinson, B.C., Pope, R., Bradbury, J., Austin, J., Isard, M., Gur-Ari, G., Yin, P., Duke, T., Levskaya, A., Ghemawat, S., Dev, S., Michalewski, H., Garcia, X., Misra, V., Robinson, K., Fedus, L., Zhou, D., Ippolito, D., Luan, D., Lim, H., Zoph, B., Spiridonov, A., Sepassi, R., Dohan, D., Agrawal, S., Omernick, M., Dai, A.M., Pillai, T.S., Pellat, M., Lewkowycz, A., Moreira, E., Child, R., Polozov, O., Lee, K., Zhou, Z., Wang, X., Saeta, B., Diaz, M., Firat, O., Catasta, M., Wei, J., Meier-Hellstern, K.S., Eck, D., Dean, J., Petrov, S., Fiedel, N.: Palm: Scaling language modeling with pathways. In: J. Mach. Learn. Res. (2022) +8. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B.A., Fung, P., Hoi, S.C.H.: Instructclip: Towards general-purpose vision-language models with instruction tuning. In: NeurIPS (2023) +9. Dessi, R., Bevilacqua, M., Gualdoni, E., Rakotonirina, N.C., Franzon, F., Baroni, M.: Cross-domain image captioning with discriminative finetuning. In: CVPR (2023) +10. Goyal, Y., Khot, T., Summers-Stay, D., Batra, D., Parikh, D.: Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In: CVPR (2017) +11. Gurari, D., Li, Q., Stangl, A., Guo, A., Lin, C., Grauman, K., Luo, J., Bigham, J.P.: Vizwiz grand challenge: Answering visual questions from blind people. In: CVPR (2018) +12. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. In: ICLR (2022) +13. Hudson, D.A., Manning, C.D.: Gqa: A new dataset for real-world visual reasoning and compositional question answering. In: CVPR (2019) +14. Ikotun, A.M., Ezugwu, A.E., Abualigah, L., Abuhaija, B., Heming, J.: K-means clustering algorithms: A comprehensive review, variants analysis, and advances in the era of big data. In: Information Sciences (2023) + +15. Kazemzadeh, S., Ordonez, V., Matten, M., Berg, T.: ReferItGame: Referring to objects in photographs of natural scenes. In: EMNLP (2014) +16. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.B.: Segment anything. In: ICCV (2023) +17. Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., Bernstein, M.S., Fei-Fei, L.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. In: IJCV (2016) +18. Li, J., Pan, K., Ge, Z., Gao, M., Zhang, H., Ji, W., Zhang, W., Chua, T.S., Tang, S., Zhuang, Y.: Empowering vision-language models to follow interleaved vision-language instructions. In: ICLR (2024) +19. Li, J., Li, D., Xiong, C., Hoi, S.C.H.: Blip: Bootstrapping language-image pretraining for unified vision-language understanding and generation. In: International Conference on Machine Learning (2022) +20. Li, Y., Du, Y., Zhou, K., Wang, J., Zhao, X., Wen, J.R.: Evaluating object hallucination in large vision-language models. In: EMNLP (2023) +21. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. In: ArXiv (2023) +22. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: NeurIPS (2023) +23. Liu, Y., Duan, H., Zhang, Y., Li, B., Zhang, S., Zhao, W., Yuan, Y., Wang, J., He, C., Liu, Z., Chen, K., Lin, D.: Mmbench: Is your multi-modal model an all-around player? In: ArXiv (2023) +24. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: ICLR (2019) +25. Lu, P., Mishra, S., Xia, T., Qiu, L., Chang, K.W., Zhu, S.C., Tafjord, O., Clark, P., Kalyan, A.: Learn to explain: Multimodal reasoning via thought chains for science question answering. In: NeurIPS (2022) +26. Manas, O., López, P.R., Ahmadi, S., Nematzadeh, A., Goyal, Y., Agrawal, A.: Mapl: Parameter-efficient adaptation of unimodal pre-trained models for vision-language few-shot prompting. In: EACL (2023) +27. Mishra, A., Shekhar, S., Singh, A.K., Chakraborty, A.: Ocr-vqa: Visual question answering by reading text in images. In: ICDAR (2019) +28. Mokady, R.: Clipcap: Clip prefix for image captioning. In: ArXiv (2021) +29. Ordonez, V., Kulkarni, G., Berg, T.: Im2text: Describing images using 1 million captioned photographs. In: NeurIPS (2011) +30. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C.L., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., Schulman, J., Hilton, J., Kelton, F., Miller, L.E., Simens, M., Askell, A., Welinder, P., Christiano, P.F., Leike, J., Lowe, R.J.: Training language models to follow instructions with human feedback. In: NeurIPS (2022) +31. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C.L., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., Schulman, J., Hilton, J., Kelton, F., Miller, L.E., Simens, M., Askell, A., Welinder, P., Christiano, P.F., Leike, J., Lowe, R.J.: Training language models to follow instructions with human feedback. In: NeurIPS (2022) +32. Papineni, K., Roukos, S., Ward, T., Zhu, W.J.: Bleu: A method for automatic evaluation of machine translation. In: ACL (2002) +33. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: ICML (2021) + +34. Rohrbach, A., Hendricks, L.A., Burns, K., Darrell, T., Saenko, K.: Object hallucination in image captioning. In: EMNLP (2018) +35. Sanh, V., Webson, A., Raffel, C., Bach, S.H., Sutawika, L., Alyafeai, Z., Chaffin, A., Stiegler, A., Scao, T.L., Raja, A., Dey, M., Bari, M.S., Xu, C., Thakker, U., Sharma, S., Szczechla, E., Kim, T., Chhablani, G., Nayak, N.V., Datta, D., Chang, J.D., Jiang, M.T.J., Wang, H., Manica, M., Shen, S., Yong, Z.X., Pandey, H., Bawden, R., Wang, T., Neeraj, T., Rozen, J., Sharma, A., Santilli, A., Févry, T., Fries, J.A., Teehan, R., Biderman, S., Gao, L., Bers, T., Wolf, T., Rush, A.M.: Multitask prompted training enables zero-shot task generalization. In: ICLR (2022) +36. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. In: NeurIPS Workshop (2021) +37. Sharma, P., Ding, N., Goodman, S., Soricut, R.: Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In: Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) (2018) +38. Shukor, M., Dancette, C., Cord, M.: ep-alm: Efficient perceptual augmentation of language models. In: ICCV (2023) +39. Singh, A., Natarajan, V., Shah, M., Jiang, Y., Chen, X., Batra, D., Parikh, D., Rohrbach, M.: Towards vqa models that can read. In: CVPR (2019) +40. Sun, G., Bai, Y., Yang, X., Fang, Y., Fu, Y., Tao, Z.: Aligning out-of-distribution web images and caption semantics via evidential learning. In: Proceedings of the ACM on Web Conference 2024 (2024) +41. Tofade, T., Elsner, J., Haines, S.: Best practice strategies for effective use of questions as a teaching tool. In: American journal of pharmaceutical education (2013) +42. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., Lample, G.: Llama: Open and efficient foundation language models. In: ArXiv (2023) +43. Touvron, H., Martin, L., Stone, K.R., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., Bikel, D.M., Blecher, L., Ferrer, C.C., Chen, M., Cucurull, G., Esiobu, D., Fernandes, J., Fu, J., Fu, W., Fuller, B., Gao, C., Goswami, V., Goyal, N., Hartshorn, A.S., Hosseini, S., Hou, R., Inan, H., Kardas, M., Kerkez, V., Khabsa, M., Kloumann, I.M., Korenev, A.V., Koura, P.S., Lachaux, M.A., Lavril, T., Lee, J., Liskovich, D., Lu, Y., Mao, Y., Martinet, X., Mihaylov, T., Mishra, P., Molybog, I., Nie, Y., Poulton, A., Reizenstein, J., Rungta, R., Saladi, K., Schelten, A., Silva, R., Smith, E.M., Subramanian, R., Tan, X., Tang, B., Taylor, R., Williams, A., Kuan, J.X., Xu, P., Yan, Z., Zarov, I., Zhang, Y., Fan, A., Kambadur, M., Narang, S., Rodriguez, A., Stojnic, R., Edunov, S., Scialom, T.: Llama 2: Open foundation and fine-tuned chat models. In: ArXiv (2023) +44. Triantafillou, E., Zhu, T.L., Dumoulin, V., Lamblin, P., Xu, K., Goroshin, R., Gelada, C., Swersky, K., Manzagol, P.A., Larochelle, H.: Meta-dataset: A dataset of datasets for learning to learn from few examples. In: ICLR (2019) +45. Vattani, A.: K-means requires exponentially many iterations even in the plane. In: Annual Symposium on Computational Geometry (2009) +46. Vedantam, R., Zitnick, C.L., Parikh, D.: Cider: Consensus-based image description evaluation. In: CVPR (2015) +47. Wang, J., Sun, G., Wang, P., Liu, D., Dianat, S., Rabbani, M., Rao, R., Tao, Z.: Text is mass: Modeling as stochastic embedding for text-video retrieval. In: CVPR (2024) + +48. Wang, Y., Kordi, Y., Mishra, S., Liu, A., Smith, N.A., Khashabi, D., Hajishirzi, H.: Self-instruct: Aligning language models with self-generated instructions. In: ACL (2022) +49. Wang, Y., Mishra, S., Alipoormolabashi, P., Kordi, Y., Mirzaei, A., Arunkumar, A., Ashok, A., Dhanasekaran, A.S., Naik, A., Stap, D., Pathak, E., Karamanolakis, G., Lai, H.G., Purohit, I., Mondal, I., Anderson, J., Kuznia, K., Doshi, K., Patel, M., Pal, K.K., Moradshahi, M., Parmar, M., Purohit, M., Varshney, N., Kaza, P.R., Verma, P., Puri, R.S., Karia, R., Sampat, S.K., Doshi, S., Mishra, S.D., Reddy, S., Patro, S., Dixit, T., Shen, X., Baral, C., Choi, Y., Smith, N.A., Hajishirzi, H., Khashabi, D.: Super-naturalinstructions: Generalization via declarative instructions on 1600+ nlp tasks. In: EMNLP (2022) +50. Wei, J., Bosma, M., Zhao, V., Guu, K., Yu, A.W., Lester, B., Du, N., Dai, A.M., Le, Q.V.: Finetuned language models are zero-shot learners. In: ICLR (2022) +51. Xu, C., Sun, Q., Zheng, K., Geng, X., Zhao, P., Feng, J., Tao, C., Jiang, D.: Wizardlm: Empowering large language models to follow complex instructions. In: ICLR (2024) +52. Young, P., Lai, A., Hodosh, M., Hockenmaier, J.: From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. In: TACL (2014) +53. Yu, W., Yang, Z., Li, L., Wang, J., Lin, K., Liu, Z., Wang, X., Wang, L.: Mm-vet: Evaluating large multimodal models for integrated capabilities. In: ArXiv (2023) +54. Zhang, R., Han, J., Zhou, A., Hu, X., Yan, S., Lu, P., Li, H., Gao, P., Qiao, Y.: Llama-adapter: Efficient fine-tuning of language models with zero-init attention. In: ICLR (2024) +55. Zhang, Y., Zhang, R., Gu, J., Zhou, Y., Lipka, N., Yang, D., Sun, T.: Llavar: Enhanced visual instruction tuning for text-rich image understanding. In: ArXiv (2023) +56. Zhao, B., Wu, B., Huang, T.: Svit: Scaling up visual instruction tuning. In: ArXiv (2023) +57. Zheng, L., Chiang, W.L., Sheng, Y., Zhuang, S., Wu, Z., Zhuang, Y., Lin, Z., Li, Z., Li, D., Xing, E.P., Zhang, H., Gonzalez, J., Stoica, I.: Judging llm-as-a-judge with mt-bench and chatbot arena. In: NeurIPS (2023) +58. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. In: ICLR (2024) \ No newline at end of file diff --git a/2024/SQ-LLaVA_ Self-Questioning for Large Vision-Language Assistant/images.zip b/2024/SQ-LLaVA_ Self-Questioning for Large Vision-Language Assistant/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f211a85dd9a10448762ed0c12386b0ab736ef3c5 --- /dev/null +++ b/2024/SQ-LLaVA_ Self-Questioning for Large Vision-Language Assistant/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a04c6b279ff1fd919c526728b23bfcc5d4ae55f342e34f49edca07a7ecb654b3 +size 448662 diff --git a/2024/SQ-LLaVA_ Self-Questioning for Large Vision-Language Assistant/layout.json b/2024/SQ-LLaVA_ Self-Questioning for Large Vision-Language Assistant/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..56f83e038b4f281f8b198126dead8e5f985ec895 --- /dev/null +++ b/2024/SQ-LLaVA_ Self-Questioning for Large Vision-Language Assistant/layout.json @@ -0,0 +1,10825 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 173, + 111, + 442, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 111, + 442, + 148 + ], + "spans": [ + { + "bbox": [ + 173, + 111, + 442, + 148 + ], + "type": "text", + "content": "SQ-LLaVA: Self-Questioning for Large Vision-Language Assistant" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 167, + 167, + 447, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 167, + 447, + 194 + ], + "spans": [ + { + "bbox": [ + 167, + 167, + 447, + 194 + ], + "type": "text", + "content": "Guohao Sun1, Can Qin2, Jiamian Wang1, Zeyuan Chen2, Ran Xu2, and Zhiqiang Tao1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 190, + 201, + 422, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 201, + 422, + 224 + ], + "spans": [ + { + "bbox": [ + 190, + 201, + 422, + 224 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 190, + 201, + 422, + 224 + ], + "type": "text", + "content": " Rochester Institute of Technology, Rochester, NY, US " + }, + { + "bbox": [ + 190, + 201, + 422, + 224 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 190, + 201, + 422, + 224 + ], + "type": "text", + "content": " Salesforce AI Research, CA, US" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 159, + 251, + 455, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 251, + 455, + 482 + ], + "spans": [ + { + "bbox": [ + 159, + 251, + 455, + 482 + ], + "type": "text", + "content": "Abstract. Recent advances in vision-language models have shown notable generalization in broad tasks through visual instruction tuning. However, bridging the gap between the pre-trained vision encoder and the large language models (LLMs) becomes the whole network's bottleneck. To improve cross-modality alignment, existing works usually consider more visual instruction data covering a broader range of vision tasks to fine-tune the model for question-answering, which, however, is costly to obtain and has not thoroughly explored the rich contextual information contained in images. This paper first attempts to harness the overlooked context within visual instruction data, training the model to self-supervised \"learning\" how to ask high-quality questions. In this way, we introduce a novel framework named SQ-LLaVA: Self-Questioning for Large Vision-Language Assistant. SQ-LLaVA exhibits proficiency in generating flexible and meaningful image-related questions while analyzing the visual clue and prior language knowledge, signifying an advanced level of generalized visual understanding. Moreover, fine-tuning SQ-LLaVA on higher-quality instruction data shows a performance improvement compared with traditional visual-instruction tuning methods. This improvement highlights the efficacy of self-questioning techniques in achieving a deeper and more nuanced comprehension of visual content across various contexts. Our code is available at https://github.com/heliossun/SQ-LLaVA." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 491, + 455, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 491, + 455, + 514 + ], + "spans": [ + { + "bbox": [ + 159, + 491, + 455, + 514 + ], + "type": "text", + "content": "Keywords: Vision-Language Understanding " + }, + { + "bbox": [ + 159, + 491, + 455, + 514 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 491, + 455, + 514 + ], + "type": "text", + "content": " Multi-modal LLM " + }, + { + "bbox": [ + 159, + 491, + 455, + 514 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 491, + 455, + 514 + ], + "type": "text", + "content": " Instruction Tuning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 533, + 230, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 230, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 230, + 547 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 558, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 483, + 666 + ], + "type": "text", + "content": "The recently emerging large vision-language methods, such as large language-and-vision assistant (LLaVA) and its variants [3,5,21,22,54], fine-tune large language models (LLM) [3,54,57] on visual instruction data [18,55,56,58] to realize diverse open-world multimodal understanding, demonstrating a surprising efficacy of visual instruction tuning - the LLM learns to perform complex vision tasks by conditioning on a prompt containing image and text clues. Existing visual instruction data is mainly built upon conversations (e.g., ChatGPT/GPT4-V [1]), consisting of images and multiple question-answer pairs. Building high-quality visual instruction data usually requires images and texts from different tasks to" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 114, + 474, + 252 + ], + "blocks": [ + { + "bbox": [ + 143, + 114, + 474, + 252 + ], + "lines": [ + { + "bbox": [ + 143, + 114, + 474, + 252 + ], + "spans": [ + { + "bbox": [ + 143, + 114, + 474, + 252 + ], + "type": "image", + "image_path": "25892de293d2dfaee7812478ca6e319cee90bd564ffd006db9b542b01f00afde.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 259, + 483, + 293 + ], + "lines": [ + { + "bbox": [ + 130, + 259, + 483, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 483, + 293 + ], + "type": "text", + "content": "Fig. 1: (a) Comparison between visual instruction tuning and visual self-questioning (ours) for vision-language assistant. (b) The proposed SQ-LLaVA achieves state-of-the-art performance on 9 out of 10 tasks compared with other open-ended models." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 318, + 482, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 318, + 482, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 318, + 482, + 402 + ], + "type": "text", + "content": "generate diverse questions, such as \"Please provide the bounding box coordinate of the region this sentence describes A dead leaf on the ground\" for the object detection task. Empirically, by increasing the diversity of questions, LLaVA has achieved better performance on GQA and VizWiz tasks (26% and 45% over previous state-of-the-art methods [8]). This evidence strongly suggests the advantage of training models on a broad spectrum and diverse array of tasks for enriching general vision-language understanding." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 402, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 640 + ], + "type": "text", + "content": "LLaVA [21] model family usually consists of a pre-trained vision encoder (e.g., CLIP-ViT [33]), a large generative language model (LLMs like Vicuna [57], LLaMA [54], QWen [3], etc.), and a vision-to-language projector implemented by a few linear layers. However, the modality gap between the pre-trained vision encoder and the language model restricts both sides' generalization ability and feature representation. To overcome this challenge, various techniques have been proposed to align the vision and language domains, which could be roughly categorized into three groups: 1) build a more robust image feature extractor [8, 54, 58], 2) collect more high-quality training data [5, 8, 21, 58], and 3) fully fine-tune the vision and language models simultaneously during the pre-training stage [5]. While these methods have shown good progress in mitigating the domain gap, they inevitably bring higher computational costs and more expensive data collection and may also require sophisticated manual designs as well as heavy annotating efforts. Moreover, images generally encompass rich information, including color, context, and the relationships between objects, but most existing visual instruction datasets capture only a fraction. We posit that leveraging such under-explored knowledge could significantly aid in vision-language understanding. In this study, we propose a new visual self-questioning approach (Fig. 1a) by training the LLM to ask questions and discover vision clues without collecting extra data from other sources." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Unlike existing visual instruction tuning methods that focus solely on answer prediction, the proposed visual self-questioning aims to extract relevant question" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "G. Sun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 115, + 225, + 178 + ], + "blocks": [ + { + "bbox": [ + 136, + 115, + 225, + 178 + ], + "lines": [ + { + "bbox": [ + 136, + 115, + 225, + 178 + ], + "spans": [ + { + "bbox": [ + 136, + 115, + 225, + 178 + ], + "type": "image", + "image_path": "9542d76fdec00731d017521f54070ffb255aca4d0dd2b26ef4310d0f163c853f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 190, + 482, + 223 + ], + "lines": [ + { + "bbox": [ + 131, + 190, + 482, + 223 + ], + "spans": [ + { + "bbox": [ + 131, + 190, + 482, + 223 + ], + "type": "text", + "content": "Fig. 2: Left: Questions spread more samples around a higher mean CLIPScore than answers. Right: Example of highly image-relevant questions within the visual instructional dataset for training a Vision-Language assistant." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 226, + 116, + 302, + 178 + ], + "blocks": [ + { + "bbox": [ + 226, + 116, + 302, + 178 + ], + "lines": [ + { + "bbox": [ + 226, + 116, + 302, + 178 + ], + "spans": [ + { + "bbox": [ + 226, + 116, + 302, + 178 + ], + "type": "image", + "image_path": "752976e3d1808ce7bd5632448ae11110285fe1be37b01ecc68f2b49655174b3d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 307, + 118, + 349, + 175 + ], + "blocks": [ + { + "bbox": [ + 307, + 118, + 349, + 175 + ], + "lines": [ + { + "bbox": [ + 307, + 118, + 349, + 175 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 349, + 175 + ], + "type": "image", + "image_path": "dbaf444cb5668a56e37c500aa32db41ae18770b8868f82fd94fba2cd6b5c035c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 351, + 118, + 476, + 175 + ], + "blocks": [ + { + "bbox": [ + 351, + 118, + 476, + 175 + ], + "lines": [ + { + "bbox": [ + 351, + 118, + 476, + 175 + ], + "spans": [ + { + "bbox": [ + 351, + 118, + 476, + 175 + ], + "type": "image", + "image_path": "afaceb6cfb129dcd4a175fc60429cfcdea1e56df1f3b14e90898b116acf979f1.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 247, + 482, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 390 + ], + "type": "text", + "content": "context. As illustrated in Fig. 2 right, the questions could contain more image-related information than the answers, inspiring us that aligning images and related questions may further improve the model's vision-language understanding capacity. Quantitatively, we compute the CLIPScore [33] (a higher value means better visual-text relevance) for all the image-question/-answer pairs on two visual instruction datasets (LLaVA-instruct [21] and ShareGPT4V-instruct [5]). By comparison, the questions' mean CLIPScore " + }, + { + "bbox": [ + 130, + 247, + 482, + 390 + ], + "type": "inline_equation", + "content": "\\mu_q = 0.184" + }, + { + "bbox": [ + 130, + 247, + 482, + 390 + ], + "type": "text", + "content": " is larger than the answers' " + }, + { + "bbox": [ + 130, + 247, + 482, + 390 + ], + "type": "inline_equation", + "content": "\\mu_a = 0.183" + }, + { + "bbox": [ + 130, + 247, + 482, + 390 + ], + "type": "text", + "content": " on LLaVA-instruct; " + }, + { + "bbox": [ + 130, + 247, + 482, + 390 + ], + "type": "inline_equation", + "content": "\\mu_q = 0.186" + }, + { + "bbox": [ + 130, + 247, + 482, + 390 + ], + "type": "text", + "content": " is larger than the answers' " + }, + { + "bbox": [ + 130, + 247, + 482, + 390 + ], + "type": "inline_equation", + "content": "\\mu_a = 0.184" + }, + { + "bbox": [ + 130, + 247, + 482, + 390 + ], + "type": "text", + "content": " on ShareGPT4V-instruct. Fig. 2 left shows the CLIPScore distribution, indicating questions have a similar and even better visual relevance than the answers. Therefore, questions in current visual instruction data can be used to fine-tune instructional LLMs owing to their diverse semantic information." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 392, + 482, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 482, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 482, + 511 + ], + "type": "text", + "content": "This work introduces a self-questioning LLaVA, namely SQ-LLaVA, to fully utilize questions within the instruction data as an additional learning resource for training instructional LLMs and empowering the model's curiosity (questioning ability). To efficiently align the vision and language domains, we apply LoRAs [12] to optimize both the vision encoder and the instructional LLM within the SQ-LLaVA. Plus, we develop a prototype extractor to enhance visual representation by leveraging learned clusters with meaningful semantic information to improve vision-language alignment further. Extensive experiments demonstrate that SQ-LLaVA surpasses existing visual instruction tuning methods in general vision understanding (see Fig. 1b). We summarize the contributions as follows." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 137, + 521, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 137, + 521, + 481, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 521, + 481, + 581 + ], + "spans": [ + { + "bbox": [ + 137, + 521, + 481, + 581 + ], + "type": "text", + "content": "- We propose a novel training technique, visual self-questioning for vision-language assistants (SQ-LLaVA), by leveraging highly relevant question contexts in instruction data. This SQ learning task promotes instructional LLMs to understand the relationship between images and questions, enhancing vision-language alignment without needing new data collection." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 137, + 581, + 481, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 581, + 481, + 629 + ], + "spans": [ + { + "bbox": [ + 137, + 581, + 481, + 629 + ], + "type": "text", + "content": "- We design and develop a lightweight tuning architecture for SQ-LLaVA, consisting of ViT-LoRA, LLM-LoRA, and a prototype extractor. The prototype extractor enhances vision embeddings, while ViT-LoRA and LLM-LoRA efficiently align vision and language domains during training." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 137, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 137, + 629, + 481, + 665 + ], + "type": "text", + "content": "- Extensive experimental results show that the proposed SQ-LLaVA leads to better performance in several tasks, including visual question-answering, visual instruction benchmarks, and zero-shot image captioning." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "type": "text", + "content": "SQ-LLaVA: Self-Questioning for Large Vision-Language Assistant" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 137, + 256, + 149 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 137, + 256, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 137, + 256, + 149 + ], + "type": "text", + "content": "2.1 Instruction Tuning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 155, + 483, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 155, + 483, + 443 + ], + "spans": [ + { + "bbox": [ + 130, + 155, + 483, + 443 + ], + "type": "text", + "content": "Instruction tuning emerged as a pivotal methodology within the realm of natural language processing (NLP), facilitating Large Language Models (LLMs) such as GPT-3 [4], PaLM [7], and LLaMA [42] to interpret and execute human language instructions across a spectrum of NLP tasks. This approach diverges from traditional fine-tuning mechanisms by incorporating a specialized data structure, termed instruction-following data [50], which is instrumental in the fine-tuning process of LLMs. Generally, there are two main categories regarding instruction tuning methods - 1) closed-domain and 2) open-domain. The closed-domain instruction tuning [35,50,51] studies engaged LLMs with a comprehensive assortment of publicly accessible datasets and subsequently assessed their performance across diverse NLP tasks [44]. The empirical evidence from these inquiries consistently indicated that integrating varied NLP task instructions significantly augments the LLMs' efficacy in navigating novel tasks. Nonetheless, LLMs calibrated with such closed-form instructions exhibited limitations in real-world user scenarios, prompting the development of an alternative approach. To address these constraints, the concept of open-domain instruction tuning [30, 49] is conceived. OpenAI pioneered this approach by employing human annotators to compile a corpus of real-world question-answer datasets. These datasets form the foundation for training a reward model through reinforcement learning methodologies. The trained reward model then functions as a supervisory mechanism for further training instruction-oriented language models, such as InstructGPT [31] and Vicuna [57]. This innovation marks a significant advancement in the field, aiming to bridge the gap between LLM performance and real-world applicability by leveraging instruction data derived from authentic user interactions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 456, + 317, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 456, + 317, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 456, + 317, + 468 + ], + "type": "text", + "content": "2.2 Large Vision Language Models" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 474, + 483, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 483, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 483, + 668 + ], + "type": "text", + "content": "As the field of LLMs and instruction tuning undergoes rapid advancements, the academic research community is progressively focusing on integrating visual information into LLMs to facilitate visual instruction tuning. This emerging area of research has witnessed the development of various methodologies [3,5,21,22,54] based on foundational vision-language models [19,33,40,47] and diverse LLM architectures [3,43,57]. In particular, LLaVA [22] pioneers the integration of an LLM with a CLIP vision encoder to construct a vision language model, demonstrating remarkable capabilities in image-text dialogue tasks through pretraining alignment strategies and targeted instruction tuning. Subsequent investigations have sought to refine visual instruction tuning by enhancing the quality and variety of the datasets used during the pre-training and fine-tuning phases. Building upon these advancements, recent studies like LLaVA-v1.5 [21] and ShareGPT4V [5] have achieved notable success in general vision-language comprehension, showcasing their ability to undertake complex question-answering tasks. This progression underscores the importance of sophisticated data handling and model tuning strategies in developing effective vision-language models." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "G. Sun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 113, + 479, + 238 + ], + "blocks": [ + { + "bbox": [ + 132, + 113, + 479, + 238 + ], + "lines": [ + { + "bbox": [ + 132, + 113, + 479, + 238 + ], + "spans": [ + { + "bbox": [ + 132, + 113, + 479, + 238 + ], + "type": "image", + "image_path": "b7b262649a963a2c222fa00777c5e37a3722f0df3764ae09d3ef66c868ccb104.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 250, + 482, + 304 + ], + "lines": [ + { + "bbox": [ + 130, + 250, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 250, + 482, + 304 + ], + "type": "text", + "content": "Fig. 3: Model architecture of SQ-LLaVA. We propose prototype extractor to extract visual clustering information to enhance the visual embedding encoded by the visual encoder. SQ-LLaVA defines a new token [vusr] as specific instruction for LLM to perform visual self-questioning. Besides question answering, SQ-LLaVA treats questioning as another training objective." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 327, + 201, + 340 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 327, + 201, + 340 + ], + "spans": [ + { + "bbox": [ + 132, + 327, + 201, + 340 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 353, + 276, + 364 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 353, + 276, + 364 + ], + "spans": [ + { + "bbox": [ + 132, + 353, + 276, + 364 + ], + "type": "text", + "content": "3.1 Architecture Overview" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 373, + 482, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 373, + 482, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 482, + 509 + ], + "type": "text", + "content": "The proposed SQ-LLaVA model (see Fig. 3) consists of four main components: 1) A pre-trained vision encoder CLIP-ViT [33] that extracts a sequence embedding of image tokens " + }, + { + "bbox": [ + 130, + 373, + 482, + 509 + ], + "type": "inline_equation", + "content": "Z_{v}" + }, + { + "bbox": [ + 130, + 373, + 482, + 509 + ], + "type": "text", + "content": " for an input image " + }, + { + "bbox": [ + 130, + 373, + 482, + 509 + ], + "type": "inline_equation", + "content": "X_{v}" + }, + { + "bbox": [ + 130, + 373, + 482, + 509 + ], + "type": "text", + "content": "; 2) A prototype extractor " + }, + { + "bbox": [ + 130, + 373, + 482, + 509 + ], + "type": "inline_equation", + "content": "\\phi(\\cdot)" + }, + { + "bbox": [ + 130, + 373, + 482, + 509 + ], + "type": "text", + "content": " learning visual clusters to enhance the original image tokens; 3) A trainable projection block " + }, + { + "bbox": [ + 130, + 373, + 482, + 509 + ], + "type": "inline_equation", + "content": "W(\\cdot)" + }, + { + "bbox": [ + 130, + 373, + 482, + 509 + ], + "type": "text", + "content": " with two linear layers to map the enhanced image tokens to the language domain tokens " + }, + { + "bbox": [ + 130, + 373, + 482, + 509 + ], + "type": "inline_equation", + "content": "H_{v}" + }, + { + "bbox": [ + 130, + 373, + 482, + 509 + ], + "type": "text", + "content": ", handling the dimension misalignment between the vision and language domain; and 4) Our LLM backbone " + }, + { + "bbox": [ + 130, + 373, + 482, + 509 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 130, + 373, + 482, + 509 + ], + "type": "text", + "content": " implemented by the pre-trained Vicuna [57] to predict the next token upon the previous embedding sequence. Given the input question " + }, + { + "bbox": [ + 130, + 373, + 482, + 509 + ], + "type": "inline_equation", + "content": "X_{q}" + }, + { + "bbox": [ + 130, + 373, + 482, + 509 + ], + "type": "text", + "content": " and answer " + }, + { + "bbox": [ + 130, + 373, + 482, + 509 + ], + "type": "inline_equation", + "content": "X_{a}" + }, + { + "bbox": [ + 130, + 373, + 482, + 509 + ], + "type": "text", + "content": ", a word embedding matrix is used to map them to contextual embeddings " + }, + { + "bbox": [ + 130, + 373, + 482, + 509 + ], + "type": "inline_equation", + "content": "H_{q}" + }, + { + "bbox": [ + 130, + 373, + 482, + 509 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 373, + 482, + 509 + ], + "type": "inline_equation", + "content": "H_{a}" + }, + { + "bbox": [ + 130, + 373, + 482, + 509 + ], + "type": "text", + "content": ", and the distribution over " + }, + { + "bbox": [ + 130, + 373, + 482, + 509 + ], + "type": "inline_equation", + "content": "H_{a}^{(i+1)}" + }, + { + "bbox": [ + 130, + 373, + 482, + 509 + ], + "type": "text", + "content": " can be obtained following the auto-regressive model as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 195, + 515, + 482, + 532 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 515, + 482, + 532 + ], + "spans": [ + { + "bbox": [ + 195, + 515, + 482, + 532 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(H _ {a} ^ {(i + 1)} \\mid H _ {v}, H _ {q}, H _ {a} ^ {(1: i)}\\right) = \\sigma \\left(f \\left(H _ {v}, H _ {q}, H _ {a} ^ {(1: i)}\\right)\\right), \\tag {1}", + "image_path": "c9d1dbec444d8b06b6fd8510ccb4d03db4e0c1f20933fa982ce0901c3df6abeb.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 539, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 539, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 539, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 539, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 539, + 482, + 666 + ], + "type": "text", + "content": " represents all the trainable parameters in our model, " + }, + { + "bbox": [ + 130, + 539, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\sigma(\\cdot)" + }, + { + "bbox": [ + 130, + 539, + 482, + 666 + ], + "type": "text", + "content": " is a softmax function, and " + }, + { + "bbox": [ + 130, + 539, + 482, + 666 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 130, + 539, + 482, + 666 + ], + "type": "text", + "content": " outputs the last token embedding of the whole sequence. We denote " + }, + { + "bbox": [ + 130, + 539, + 482, + 666 + ], + "type": "inline_equation", + "content": "p_{\\theta}" + }, + { + "bbox": [ + 130, + 539, + 482, + 666 + ], + "type": "text", + "content": " as the prediction probability for the anticipated answer token " + }, + { + "bbox": [ + 130, + 539, + 482, + 666 + ], + "type": "inline_equation", + "content": "H_{a}^{(i+1)}" + }, + { + "bbox": [ + 130, + 539, + 482, + 666 + ], + "type": "text", + "content": " at the position " + }, + { + "bbox": [ + 130, + 539, + 482, + 666 + ], + "type": "inline_equation", + "content": "i+1" + }, + { + "bbox": [ + 130, + 539, + 482, + 666 + ], + "type": "text", + "content": ", conditioning on the input image token embedding " + }, + { + "bbox": [ + 130, + 539, + 482, + 666 + ], + "type": "inline_equation", + "content": "H_{v}" + }, + { + "bbox": [ + 130, + 539, + 482, + 666 + ], + "type": "text", + "content": ", the question token embedding " + }, + { + "bbox": [ + 130, + 539, + 482, + 666 + ], + "type": "inline_equation", + "content": "H_{q}" + }, + { + "bbox": [ + 130, + 539, + 482, + 666 + ], + "type": "text", + "content": ", and the previous answer token embeddings " + }, + { + "bbox": [ + 130, + 539, + 482, + 666 + ], + "type": "inline_equation", + "content": "H_{a}^{(1:i)}" + }, + { + "bbox": [ + 130, + 539, + 482, + 666 + ], + "type": "text", + "content": ". As shown in Eq. (1), the proposed SQ-LLaVA applies the language model " + }, + { + "bbox": [ + 130, + 539, + 482, + 666 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 130, + 539, + 482, + 666 + ], + "type": "text", + "content": " to model " + }, + { + "bbox": [ + 130, + 539, + 482, + 666 + ], + "type": "inline_equation", + "content": "p_{\\theta}" + }, + { + "bbox": [ + 130, + 539, + 482, + 666 + ], + "type": "text", + "content": " given by " + }, + { + "bbox": [ + 130, + 539, + 482, + 666 + ], + "type": "inline_equation", + "content": "H_{v}" + }, + { + "bbox": [ + 130, + 539, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 539, + 482, + 666 + ], + "type": "inline_equation", + "content": "H_{q}" + }, + { + "bbox": [ + 130, + 539, + 482, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 539, + 482, + 666 + ], + "type": "inline_equation", + "content": "H_{a}^{(1:i)}" + }, + { + "bbox": [ + 130, + 539, + 482, + 666 + ], + "type": "text", + "content": ". Existing visual instruction tuning methods [5,21] are only able to predict " + }, + { + "bbox": [ + 130, + 539, + 482, + 666 + ], + "type": "inline_equation", + "content": "H_{a}" + }, + { + "bbox": [ + 130, + 539, + 482, + 666 + ], + "type": "text", + "content": ", yet cannot fully exploit the rich semantic clues within " + }, + { + "bbox": [ + 130, + 539, + 482, + 666 + ], + "type": "inline_equation", + "content": "H_{q}" + }, + { + "bbox": [ + 130, + 539, + 482, + 666 + ], + "type": "text", + "content": ". In this study, we propose visual self-questioning instructions to guide LLMs in capturing the visual content within questions." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "type": "text", + "content": "SQ-LLaVA: Self-Questioning for Large Vision-Language Assistant" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 151, + 118, + 214, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 118, + 214, + 128 + ], + "spans": [ + { + "bbox": [ + 151, + 118, + 214, + 128 + ], + "type": "text", + "content": "System-message" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 151, + 129, + 446, + 146 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 129, + 446, + 146 + ], + "spans": [ + { + "bbox": [ + 151, + 129, + 446, + 146 + ], + "type": "interline_equation", + "content": "\\mathbf {X} _ {c} ^ {(1)}: [ \\mathrm {u s r} ]: \\mathbf {X} _ {v} \\backslash \\mathrm {n} \\mathbf {X} _ {q} ^ {(1)} [ \\mathrm {a s w r} ]: \\mathbf {X} _ {a} ^ {(1)} < o ^ {d} > \\quad \\mathbf {X} _ {c} ^ {(2)}: [ \\mathrm {v u s r} ]: \\mathbf {X} _ {q} ^ {(2)} < o ^ {d} > [ \\mathrm {a s w r} ]: \\mathbf {X} _ {a} ^ {(2)} < o ^ {d} >", + "image_path": "4c2491c23e88f4c633682a0d8010d83ccdc3eeb0d86b02eb4c14b0785f6a58b0.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 148, + 446, + 164 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 148, + 446, + 164 + ], + "spans": [ + { + "bbox": [ + 151, + 148, + 446, + 164 + ], + "type": "interline_equation", + "content": "\\mathbf {X} _ {c} ^ {(3)}: [ \\text {u s r} ]: \\mathbf {X} _ {q} ^ {(3)} [ \\text {a s w r} ]: \\mathbf {X} _ {a} ^ {(3)} < _ {O} d > \\quad \\dots \\quad \\mathbf {X} _ {c} ^ {(P)}: [ \\text {v u s r} ]: \\mathbf {X} _ {q} ^ {(P)} < _ {O} d > [ \\text {a s w r} ]: \\mathbf {X} _ {a} ^ {(P)} < _ {O} d >", + "image_path": "1e63da0ac070ab5131807e14c79baf0629a5f42d20fcc72e3d324349f744f02d.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 177, + 483, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 177, + 483, + 243 + ], + "spans": [ + { + "bbox": [ + 130, + 177, + 483, + 243 + ], + "type": "text", + "content": "Fig. 4: The input sequence used to train SQ-LLaVA. Our model is trained to predict question, answer, and where to stop. We use tokens to represent learnable tokens, where " + }, + { + "bbox": [ + 130, + 177, + 483, + 243 + ], + "type": "inline_equation", + "content": "X_{q}" + }, + { + "bbox": [ + 130, + 177, + 483, + 243 + ], + "type": "text", + "content": " is the question, " + }, + { + "bbox": [ + 130, + 177, + 483, + 243 + ], + "type": "inline_equation", + "content": "X_{a}" + }, + { + "bbox": [ + 130, + 177, + 483, + 243 + ], + "type": "text", + "content": " is the answer, and " + }, + { + "bbox": [ + 130, + 177, + 483, + 243 + ], + "type": "inline_equation", + "content": "< o^{d}>" + }, + { + "bbox": [ + 130, + 177, + 483, + 243 + ], + "type": "text", + "content": " is the delimiter token. In SQ-LLaVA, the System-message = \"The assistant gives helpful, detailed, and polite answers to the user's questions. Also, the assistant is a curious virtual user can ask complex questions that are relevant to the content in the image.\"" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 269, + 335, + 282 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 269, + 335, + 282 + ], + "spans": [ + { + "bbox": [ + 131, + 269, + 335, + 282 + ], + "type": "text", + "content": "3.2 Visual Self-questioning Instruction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 293, + 483, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 483, + 497 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 483, + 497 + ], + "type": "text", + "content": "In broad real-world scenarios, proactively asking a question requires more understanding and background knowledge than answering [41]. Similarly, this work proposes visual self-questioning to encourage the LLM to discover deeper vision-language alignment and improve the overall instruction-following performance. Particularly, SQ-LLaVA treats questioning as a new learning objective beyond answering, which, to the best of our knowledge, is the first practice in the field of visual instruction tuning. While the current vision-language model can ask questions [1], such skill is still learned from question-answering through instruction tuning. However, our proposed method shows that the decoder-based LLM has the potential to learn more skills, such as how to ask questions spontaneously when a unique instruction token is given (e.g., we define [vusi] in our work). Furthermore, visual self-questioning can potentially improve general vision-language understanding. To be specific, as shown in Fig. 1, there are a certain amount of questions containing more meaningful image-related information than answers in the existing visual instruction data [5, 21]. Thus, we hypothesize that the vision-language understanding can be improved once the LLM learns how to predict relevant questions about a given image." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 498, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 483, + 666 + ], + "type": "text", + "content": "Self-Questioning Prompt Design. We provide ground-truth content for the visual self-questioning, restricting SQ-LLaVA from asking image-related questions. To this end, we leverage questions as another learnable resource and follow the regular auto-regressive training objective. As shown in Fig. 4, the training data for SQ-LLaVA is designed in a format with a pre-defined template. To be specific, the system message as a fixed prompt is added at the beginning of each instruction data, indicating a general job description (e.g., gives helpful answers and asks complex questions) for the LLM. Existing visual instruction tuning methods utilize the unique tokens ([usr], [aswr]) to give the LLM a particular instruction (e.g., questioning understanding, answer prediction, etc.) and apply the delimiter token " + }, + { + "bbox": [ + 130, + 498, + 483, + 666 + ], + "type": "inline_equation", + "content": "< \\sigma^d>" + }, + { + "bbox": [ + 130, + 498, + 483, + 666 + ], + "type": "text", + "content": " to mark the ending position. In this work, we propose a new special token [vusr], indicating a specific instruction - asking questions. Combining with the delimiter, we can construct instructions for self-questioning as a new training task." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "G. Sun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 156 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 156 + ], + "type": "text", + "content": "Each sample of current visual instruction data consists of one image " + }, + { + "bbox": [ + 130, + 116, + 479, + 156 + ], + "type": "inline_equation", + "content": "X_{v}" + }, + { + "bbox": [ + 130, + 116, + 479, + 156 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 479, + 156 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 130, + 116, + 479, + 156 + ], + "type": "text", + "content": " question-answer pairs " + }, + { + "bbox": [ + 130, + 116, + 479, + 156 + ], + "type": "inline_equation", + "content": "(X_{q}^{(1)},X_{a}^{(1)},\\dots ,X_{q}^{(P)},X_{a}^{(P)})" + }, + { + "bbox": [ + 130, + 116, + 479, + 156 + ], + "type": "text", + "content": ". We collect one question " + }, + { + "bbox": [ + 130, + 116, + 479, + 156 + ], + "type": "inline_equation", + "content": "X_{q}^{(j)}" + }, + { + "bbox": [ + 130, + 116, + 479, + 156 + ], + "type": "text", + "content": " and its answer " + }, + { + "bbox": [ + 130, + 116, + 479, + 156 + ], + "type": "inline_equation", + "content": "X_{a}^{(j)}" + }, + { + "bbox": [ + 130, + 116, + 479, + 156 + ], + "type": "text", + "content": " with special tokens to construct the " + }, + { + "bbox": [ + 130, + 116, + 479, + 156 + ], + "type": "inline_equation", + "content": "j^{th}" + }, + { + "bbox": [ + 130, + 116, + 479, + 156 + ], + "type": "text", + "content": " turn conversation as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 171, + 163, + 481, + 196 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 163, + 481, + 196 + ], + "spans": [ + { + "bbox": [ + 171, + 163, + 481, + 196 + ], + "type": "interline_equation", + "content": "X _ {c} ^ {j} = \\left\\{ \\begin{array}{l l} \\left([ \\mathrm {u s r} ], X _ {q} ^ {(j)}, [ \\mathrm {a s w r} ], X _ {a} ^ {(j)}\\right) & j = 1 \\text {o r} j > 1, R < \\delta \\\\ \\left([ \\mathrm {v u s r} ], X _ {q} ^ {(j)}, [ \\mathrm {a s w r} ], X _ {a} ^ {(j)}\\right) & j > 1, R > \\delta \\end{array} , \\right. \\tag {2}", + "image_path": "859fd38eaf6209594cd0bc42e5e19b5ed4f52d6c319e16714f84f736a3cf355e.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 205, + 479, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 479, + 240 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 479, + 240 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 205, + 479, + 240 + ], + "type": "inline_equation", + "content": "R \\in [0,1]" + }, + { + "bbox": [ + 130, + 205, + 479, + 240 + ], + "type": "text", + "content": " is a random number and " + }, + { + "bbox": [ + 130, + 205, + 479, + 240 + ], + "type": "inline_equation", + "content": "\\delta = 0.5" + }, + { + "bbox": [ + 130, + 205, + 479, + 240 + ], + "type": "text", + "content": " is a threshold that sets the proportion of self-questioning pairs in the conversations. Finally, the full-text input " + }, + { + "bbox": [ + 130, + 205, + 479, + 240 + ], + "type": "inline_equation", + "content": "X_{c}" + }, + { + "bbox": [ + 130, + 205, + 479, + 240 + ], + "type": "text", + "content": " will be mapped to textual embedding " + }, + { + "bbox": [ + 130, + 205, + 479, + 240 + ], + "type": "inline_equation", + "content": "H_{c}" + }, + { + "bbox": [ + 130, + 205, + 479, + 240 + ], + "type": "text", + "content": " through word embedding." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 241, + 480, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 241, + 480, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 241, + 480, + 384 + ], + "type": "text", + "content": "SQ-LLaVA performs zero-shot questioning without any in-context knowledge or human language instruction since the only instructional prompt is a unique token [vusr]. After visual self-questioning on instruction data with various question formats " + }, + { + "bbox": [ + 130, + 241, + 480, + 384 + ], + "type": "inline_equation", + "content": "X_{q}" + }, + { + "bbox": [ + 130, + 241, + 480, + 384 + ], + "type": "text", + "content": ", the questions sampled by SQ-LLaVA are more diversified than GPT4-V (as shown in Fig. 5), since the LLM has learned alignment between image and questions, which is different from previous works [1, 48]. Specifically, previous general-purpose vision language models such as GPT4-V [1] can generate questions based on a given image, but it requires explicit language instruction such as \"Ask complex questions that are relevant to the content in the image\". Accordingly, the quality of generated questions highly relies on prompt engineering. Also, self-instruct [48] utilizes in-context learning to prompt the LLM to ask specific questions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 403, + 324, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 403, + 324, + 415 + ], + "spans": [ + { + "bbox": [ + 131, + 403, + 324, + 415 + ], + "type": "text", + "content": "3.3 Enhanced Visual Representation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 423, + 479, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 479, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 479, + 483 + ], + "type": "text", + "content": "Unlike previous visual instruction tuning methods, SQ-LLaVA jointly benefits from visual self-questioning and question-answering. For better visual self-questioning, we develop a prototype extractor that recognizes and groups similar patterns of visual information from the latent space. Our primary goal is to enhance visual representation through prototype learning." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 483, + 480, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 483, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 483, + 480, + 578 + ], + "type": "text", + "content": "Specifically, we utilize clustering to extract centroid representations of image tokens " + }, + { + "bbox": [ + 130, + 483, + 480, + 578 + ], + "type": "inline_equation", + "content": "Z_{v}" + }, + { + "bbox": [ + 130, + 483, + 480, + 578 + ], + "type": "text", + "content": ", where each cluster center is treated as a prototype, which, in return, will be distributed to each of the original image token embeddings. Our proposed prototype extractor " + }, + { + "bbox": [ + 130, + 483, + 480, + 578 + ], + "type": "inline_equation", + "content": "\\phi(\\cdot)" + }, + { + "bbox": [ + 130, + 483, + 480, + 578 + ], + "type": "text", + "content": " is a lightweight design involving two parts: 1) cluster center optimization and 2) prototype information distribution. Following [14, 45], we randomly initialize " + }, + { + "bbox": [ + 130, + 483, + 480, + 578 + ], + "type": "inline_equation", + "content": "K = 256" + }, + { + "bbox": [ + 130, + 483, + 480, + 578 + ], + "type": "text", + "content": " cluster centers " + }, + { + "bbox": [ + 130, + 483, + 480, + 578 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 483, + 480, + 578 + ], + "type": "text", + "content": " and deploy the iterative Expectation-Maximization (EM) clustering process to capture representative semantics in the latent space by" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 224, + 587, + 481, + 619 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 587, + 481, + 619 + ], + "spans": [ + { + "bbox": [ + 224, + 587, + 481, + 619 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\operatorname {E} \\text {- s t e p}: \\mathcal {M} ^ {(t)} = \\sigma \\left(q \\left(C ^ {(t)}\\right) * k \\left(Z _ {v}\\right) ^ {\\top}\\right), \\\\ \\text {M - s t e p}: C ^ {(t + 1)} = \\mathcal {M} ^ {(t)} * v \\left(Z _ {v}\\right), \\tag {3} \\\\ \\end{array}", + "image_path": "ea0281f22e18c0369d1377fd029bf3e01fee81fbb1e8a72bc48186ab006ce50f.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 628, + 479, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 628, + 479, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 628, + 479, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 628, + 479, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^{(t)}\\in [0,1]" + }, + { + "bbox": [ + 130, + 628, + 479, + 666 + ], + "type": "text", + "content": " denotes a soft cluster assignment map at the " + }, + { + "bbox": [ + 130, + 628, + 479, + 666 + ], + "type": "inline_equation", + "content": "t^{th}" + }, + { + "bbox": [ + 130, + 628, + 479, + 666 + ], + "type": "text", + "content": " iteration " + }, + { + "bbox": [ + 130, + 628, + 479, + 666 + ], + "type": "inline_equation", + "content": "\\sigma (\\cdot)" + }, + { + "bbox": [ + 130, + 628, + 479, + 666 + ], + "type": "text", + "content": " is a softmax function, and " + }, + { + "bbox": [ + 130, + 628, + 479, + 666 + ], + "type": "inline_equation", + "content": "t\\in \\{1,\\dots ,T\\}" + }, + { + "bbox": [ + 130, + 628, + 479, + 666 + ], + "type": "text", + "content": " indexes the iteration of EM step with " + }, + { + "bbox": [ + 130, + 628, + 479, + 666 + ], + "type": "inline_equation", + "content": "T = 2" + }, + { + "bbox": [ + 130, + 628, + 479, + 666 + ], + "type": "text", + "content": " in this work. Three trainable linear layers " + }, + { + "bbox": [ + 130, + 628, + 479, + 666 + ], + "type": "inline_equation", + "content": "q,k" + }, + { + "bbox": [ + 130, + 628, + 479, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 628, + 479, + 666 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 130, + 628, + 479, + 666 + ], + "type": "text", + "content": " are used in (3.3)" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "type": "text", + "content": "SQ-LLaVA: Self-Questioning for Large Vision-Language Assistant" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "q(\\cdot)" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": " projects the prototype " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": " to a query vector, and " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "k(\\cdot)" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "v(\\cdot)" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": " project " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "H_{v}" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": " into key and value vectors, followed by a normalization layer, respectively. The prototype extractor iteratively updates cluster map " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": " and centers " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 482, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 482, + 190 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 482, + 190 + ], + "type": "text", + "content": "After the prototype extraction, we train a linear layer " + }, + { + "bbox": [ + 130, + 152, + 482, + 190 + ], + "type": "inline_equation", + "content": "z(\\cdot)" + }, + { + "bbox": [ + 130, + 152, + 482, + 190 + ], + "type": "text", + "content": " to adaptively map the visual cluster information to the raw image embedding " + }, + { + "bbox": [ + 130, + 152, + 482, + 190 + ], + "type": "inline_equation", + "content": "Z_{v}" + }, + { + "bbox": [ + 130, + 152, + 482, + 190 + ], + "type": "text", + "content": ". For the " + }, + { + "bbox": [ + 130, + 152, + 482, + 190 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 130, + 152, + 482, + 190 + ], + "type": "text", + "content": " token embedding " + }, + { + "bbox": [ + 130, + 152, + 482, + 190 + ], + "type": "inline_equation", + "content": "Z_{v}^{(i)}" + }, + { + "bbox": [ + 130, + 152, + 482, + 190 + ], + "type": "text", + "content": ", we update it as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 211, + 198, + 482, + 220 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 198, + 482, + 220 + ], + "spans": [ + { + "bbox": [ + 211, + 198, + 482, + 220 + ], + "type": "interline_equation", + "content": "Z _ {v} ^ {(i)} = Z _ {v} ^ {(i)} + z \\left(\\frac {1}{K} \\sum_ {j = 1} ^ {K} S _ {c} \\left(C _ {j}, Z _ {v} ^ {(i)}\\right) \\times C _ {j}\\right), \\tag {4}", + "image_path": "457b87856cdf14e6038d08c22d24d9e5c8fca532e764e0ddee64377b89a2ad17.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 226, + 482, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 226, + 482, + 323 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 482, + 323 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 226, + 482, + 323 + ], + "type": "inline_equation", + "content": "S_{c}(\\cdot, \\cdot)" + }, + { + "bbox": [ + 130, + 226, + 482, + 323 + ], + "type": "text", + "content": " is a normalized cosine similarity function. The weighted sum over prototypes in Eq. (4) emerges as an indispensable step for contextual understanding from image tokens, recognizing and grouping similar patterns and semantics. It clusters image tokens as prototypes that display homogeneity in semantics, such as \"grass\" and \"dog\". The prototypes can describe the intrinsic semantic meanings by aggregating entities that exhibit shared attributes. Finally, we map the image sequence embedding " + }, + { + "bbox": [ + 130, + 226, + 482, + 323 + ], + "type": "inline_equation", + "content": "Z_{v}" + }, + { + "bbox": [ + 130, + 226, + 482, + 323 + ], + "type": "text", + "content": " to language domain " + }, + { + "bbox": [ + 130, + 226, + 482, + 323 + ], + "type": "inline_equation", + "content": "H_{v}" + }, + { + "bbox": [ + 130, + 226, + 482, + 323 + ], + "type": "text", + "content": " with a two-layer linear projector " + }, + { + "bbox": [ + 130, + 226, + 482, + 323 + ], + "type": "inline_equation", + "content": "W(\\cdot)" + }, + { + "bbox": [ + 130, + 226, + 482, + 323 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 338, + 240, + 352 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 338, + 240, + 352 + ], + "spans": [ + { + "bbox": [ + 131, + 338, + 240, + 352 + ], + "type": "text", + "content": "3.4 Model Training" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 357, + 482, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 357, + 482, + 539 + ], + "spans": [ + { + "bbox": [ + 130, + 357, + 482, + 539 + ], + "type": "text", + "content": "Stage1: Pre-training for Vision-Language Alignment. Unlike text-only LLMs, the vision-language model also fine-tunes LLM using image tokens as input (see Fig. 3). Therefore, the pre-training stage aims to optimize the LLM by explicitly executing the visual instruction. The proposed SQ-LLaVA adopts Vicuna [57] as its instruction LLM, pre-training on massive text corpora to predict the next text token given the previous context, not only containing text but also visual instructions. To achieve this, we organize the pre-training data as " + }, + { + "bbox": [ + 130, + 357, + 482, + 539 + ], + "type": "inline_equation", + "content": "D_{PT} = \\{[X_v^{(1)}, X_a^{(1)}], \\dots, [X_v^{(N)}, X_a^{(N)}]\\}" + }, + { + "bbox": [ + 130, + 357, + 482, + 539 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 357, + 482, + 539 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 357, + 482, + 539 + ], + "type": "text", + "content": " is the total number of training samples, and each sample has an image and its related descriptions. Each image and text input pair will be mapped to sequence embeddings (" + }, + { + "bbox": [ + 130, + 357, + 482, + 539 + ], + "type": "inline_equation", + "content": "H_v" + }, + { + "bbox": [ + 130, + 357, + 482, + 539 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 357, + 482, + 539 + ], + "type": "inline_equation", + "content": "H_a" + }, + { + "bbox": [ + 130, + 357, + 482, + 539 + ], + "type": "text", + "content": ") as elaborated in Section 3.1. During pre-training, we freeze the vision encoder and LLM and mainly train the prototype extractor " + }, + { + "bbox": [ + 130, + 357, + 482, + 539 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 357, + 482, + 539 + ], + "type": "text", + "content": " and the vision-to-language projector " + }, + { + "bbox": [ + 130, + 357, + 482, + 539 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 357, + 482, + 539 + ], + "type": "text", + "content": ". The pre-training goal is to maximize the probability of the predicted image description " + }, + { + "bbox": [ + 130, + 357, + 482, + 539 + ], + "type": "inline_equation", + "content": "H_a" + }, + { + "bbox": [ + 130, + 357, + 482, + 539 + ], + "type": "text", + "content": " given an image " + }, + { + "bbox": [ + 130, + 357, + 482, + 539 + ], + "type": "inline_equation", + "content": "H_v" + }, + { + "bbox": [ + 130, + 357, + 482, + 539 + ], + "type": "text", + "content": ". When training a visual instructional LLM, we follow the negative log-likelihood objective function as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 160, + 546, + 482, + 580 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 546, + 482, + 580 + ], + "spans": [ + { + "bbox": [ + 160, + 546, + 482, + 580 + ], + "type": "interline_equation", + "content": "\\sum_ {v, a \\in D _ {P T}} - \\log p _ {\\theta} (H _ {a} \\mid H _ {v}) = \\sum_ {v, a \\in D _ {P T}} \\sum_ {i = 1} ^ {L} - \\log p _ {\\theta} (H _ {a} ^ {(i + 1)} \\mid H _ {v}, H _ {a} ^ {(1: i)}), \\quad (5)", + "image_path": "f0d09f09f1194bfd49e0dd9b460137f86655cbbe8b873de9807b4010112f078b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 586, + 482, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 586, + 482, + 625 + ], + "spans": [ + { + "bbox": [ + 131, + 586, + 482, + 625 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 586, + 482, + 625 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 131, + 586, + 482, + 625 + ], + "type": "text", + "content": " denotes the sequence length of answer tokens in " + }, + { + "bbox": [ + 131, + 586, + 482, + 625 + ], + "type": "inline_equation", + "content": "H_{a}" + }, + { + "bbox": [ + 131, + 586, + 482, + 625 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 586, + 482, + 625 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 131, + 586, + 482, + 625 + ], + "type": "text", + "content": " is the total trainable parameter of " + }, + { + "bbox": [ + 131, + 586, + 482, + 625 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 131, + 586, + 482, + 625 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 586, + 482, + 625 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 131, + 586, + 482, + 625 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 586, + 482, + 625 + ], + "type": "inline_equation", + "content": "p(H_{a} \\mid H_{v})" + }, + { + "bbox": [ + 131, + 586, + 482, + 625 + ], + "type": "text", + "content": " can be computed by Eq. (1), and " + }, + { + "bbox": [ + 131, + 586, + 482, + 625 + ], + "type": "inline_equation", + "content": "H_{a}^{(1:i)}" + }, + { + "bbox": [ + 131, + 586, + 482, + 625 + ], + "type": "text", + "content": " represents all the answer tokens before the current prediction " + }, + { + "bbox": [ + 131, + 586, + 482, + 625 + ], + "type": "inline_equation", + "content": "H_{a}^{(i+1)}" + }, + { + "bbox": [ + 131, + 586, + 482, + 625 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Stage2: Fine-tuning. Existing methods, such as LLaVA [5,22], mainly update the vision-to-language projector (usually a couple of linear layers) and the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "G. Sun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "text", + "content": "language model during fine-tuning. Nevertheless, the projector might be too weak to capture the relationship between the image and the questions. Following the previous multi-modal understanding method [33], we unfreeze the vision encoder and LLM during fine-tuning for a joint optimization further to eliminate the gap between the vision and language domain." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 175, + 482, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 175, + 482, + 309 + ], + "spans": [ + { + "bbox": [ + 130, + 175, + 482, + 309 + ], + "type": "text", + "content": "To mitigate the heavy computational overhead, we take advantage of LoRA [12] as a lightweight tuning option that can achieve similar (even better) performance to fully fine-tuning when training large models on a relatively small dataset. We add LoRA in both the vision encoder and LLM. Thus, the learnable parameters " + }, + { + "bbox": [ + 130, + 175, + 482, + 309 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 175, + 482, + 309 + ], + "type": "text", + "content": " of the proposed SQ-LLaVA during fine-tuning represent a combination of all the parameters of LLM-LoRA, ViT-LoRA, prototype extractor " + }, + { + "bbox": [ + 130, + 175, + 482, + 309 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 175, + 482, + 309 + ], + "type": "text", + "content": ", and the vision-to-language projector " + }, + { + "bbox": [ + 130, + 175, + 482, + 309 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 175, + 482, + 309 + ], + "type": "text", + "content": ". Given the instruction tuning data " + }, + { + "bbox": [ + 130, + 175, + 482, + 309 + ], + "type": "inline_equation", + "content": "D_{IT} = \\{[X_v^{(1)}, X_c^{(1)}], \\dots, [X_v^{(N)}, X_c^{(N)}]\\}" + }, + { + "bbox": [ + 130, + 175, + 482, + 309 + ], + "type": "text", + "content": ", we take the conversational data " + }, + { + "bbox": [ + 130, + 175, + 482, + 309 + ], + "type": "inline_equation", + "content": "X_c" + }, + { + "bbox": [ + 130, + 175, + 482, + 309 + ], + "type": "text", + "content": " and the image " + }, + { + "bbox": [ + 130, + 175, + 482, + 309 + ], + "type": "inline_equation", + "content": "X_v" + }, + { + "bbox": [ + 130, + 175, + 482, + 309 + ], + "type": "text", + "content": " as input, mapping them to sequence embedding " + }, + { + "bbox": [ + 130, + 175, + 482, + 309 + ], + "type": "inline_equation", + "content": "(H_c" + }, + { + "bbox": [ + 130, + 175, + 482, + 309 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 175, + 482, + 309 + ], + "type": "inline_equation", + "content": "H_v)" + }, + { + "bbox": [ + 130, + 175, + 482, + 309 + ], + "type": "text", + "content": " as elaborated in Section 3.1, and minimize the negative log-likelihood loss for the self-questioning and answering tasks as follows" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 168, + 319, + 482, + 346 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 319, + 482, + 346 + ], + "spans": [ + { + "bbox": [ + 168, + 319, + 482, + 346 + ], + "type": "interline_equation", + "content": "\\text {S e l f - q u e s t i o n i n g}: \\sum_ {v, c \\in D _ {I T}} - \\log p _ {\\theta} \\left(H _ {q} ^ {(j + 1)} \\mid H _ {v}, H _ {c} ^ {(1: j)}\\right), \\tag {6}", + "image_path": "eabe92cbc722b138897d30cce2afb59017647409c42d3fa60a58e1659b73e82d.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 192, + 351, + 481, + 377 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 351, + 481, + 377 + ], + "spans": [ + { + "bbox": [ + 192, + 351, + 481, + 377 + ], + "type": "interline_equation", + "content": "\\text {A n s w e r i n g}: \\sum_ {v, c \\in D _ {I T}} - \\log p _ {\\theta} \\left(H _ {a} ^ {(j + 1)} \\mid H _ {v}, H _ {c} ^ {(1: j)}, H _ {q} ^ {(j + 1)}\\right), \\tag {7}", + "image_path": "f57cb8592233c58eb10673a1be24b138b4a2492d86a8adb8d0f67e63d7da11f6.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 386, + 482, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 386, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 386, + 482, + 483 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 386, + 482, + 483 + ], + "type": "inline_equation", + "content": "j \\in \\{1, \\dots, P\\}" + }, + { + "bbox": [ + 130, + 386, + 482, + 483 + ], + "type": "text", + "content": ", indicating the index of question or answer within the conversational data " + }, + { + "bbox": [ + 130, + 386, + 482, + 483 + ], + "type": "inline_equation", + "content": "X_{c}^{(*)}" + }, + { + "bbox": [ + 130, + 386, + 482, + 483 + ], + "type": "text", + "content": ". Notably, previous works [5,21,58] only involve answering tasks, but we introduce visual self-questioning as an additional training task for visual instruction tuning. Eventually, SQ-LLaVA, as a vision-language assistant, not only executes human instructions by optimizing the objective function in Eq. (7) but can raise questions out of the given image after optimizing the Eq. (6). This capability potentially yields more diverse question-answer guidance and enhances multi-modal understanding." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 502, + 230, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 502, + 230, + 514 + ], + "spans": [ + { + "bbox": [ + 131, + 502, + 230, + 514 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 526, + 269, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 526, + 269, + 539 + ], + "spans": [ + { + "bbox": [ + 131, + 526, + 269, + 539 + ], + "type": "text", + "content": "4.1 Experimental Setting" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Dataset. Our work uses the open-source visual instruction dataset provided by LLaVA [21] and ShareGPT4V [5] for training. Each dataset has large-scale image-text paired data for pre-training and instruction-following data for finetuning. Specifically, the instruction data proposed by LLaVA is a comprehensive mixture of COCO, GQA [13], OCR-VQA [27], TextVQA [39], VisualGenome [17], RefCOCO [15], and image from InstructBLIP [8] involves multiple reasoning, spatial understanding, multi-step inference, optical character recognition, and grounding of visual concepts to language. The ShareGPT4V dataset consists of the same mixture image as LLaVA but enrolls more images from other datasets such as SAM [16] and Data [29,36]." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 179, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 179, + 91, + 447, + 103 + ], + "type": "text", + "content": "SQ-LLaVA: Self-Questioning for Large Vision-Language Assistant" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": "We evaluate our model on general vision-language understanding tasks. Specifically, we use ten visual oriented question answering benchmarks: " + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "inline_equation", + "content": "\\mathrm{VQA}^{v2}" + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": " [10]; GQA [13]; VizWiz [11]; " + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "inline_equation", + "content": "\\mathrm{SQA}^I" + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": " [25]: ScienceQA-IMG; " + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "inline_equation", + "content": "\\mathrm{VQA}^T" + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": ": TextVQA [39]; POPE [20]; MM-Vet [53]; " + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "inline_equation", + "content": "\\mathrm{LLaVA}^W" + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": " [22]: LLaVA (in the wild); MMB: MMBBenchmark [23]; " + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "inline_equation", + "content": "\\mathrm{MMB}^{CN}" + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": ": MMBench-Chinese [23]. We report the prediction accuracy on all the benchmarks. We also evaluate our model for visual information discovery through captioning. To be specific, we employ testing images from four datasets, i.e., COCO [6], Flickr30K [52], Conceptual [37], and Nocaps [2]. Following [9], we evaluate the proposed method with regular image captioning metrics, e.g., BLEU [32] and CIDEr [46]. For the open-world methods LLaVA-v1.5 [21], ShareGPT4V [5] and the proposed SQ-LLaVA, we utilize greedy search for caption generation with a prompt of \"Provide a brief description of the given image, your answer should be in one sentence.\" The generated caption is used to evaluate the performance of image captioning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 300, + 240, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 240, + 312 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 240, + 312 + ], + "type": "text", + "content": "4.2 Implementation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 319, + 482, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 482, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 482, + 450 + ], + "type": "text", + "content": "We adopt Vicuna [57] as the pre-trained language generative model and CLIP-ViT [33] as the vision encoder. We pre-train the prototype extractor and the vision-to-language projector using AdamW [24] optimizer with a learning rate of " + }, + { + "bbox": [ + 130, + 319, + 482, + 450 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-3}" + }, + { + "bbox": [ + 130, + 319, + 482, + 450 + ], + "type": "text", + "content": " and a constant scheduler for one epoch. Following previous work [21], we keep the global batch size as 256 for pre-training and 128 for fine-tuning. During fine-tuning, we insert LoRA [12] with rank = 128 and " + }, + { + "bbox": [ + 130, + 319, + 482, + 450 + ], + "type": "inline_equation", + "content": "\\alpha = 256" + }, + { + "bbox": [ + 130, + 319, + 482, + 450 + ], + "type": "text", + "content": " into the language model (LLM-LoRA) and LoRA with rank = 32 and " + }, + { + "bbox": [ + 130, + 319, + 482, + 450 + ], + "type": "inline_equation", + "content": "\\alpha = 64" + }, + { + "bbox": [ + 130, + 319, + 482, + 450 + ], + "type": "text", + "content": " into the vision encoder (ViT-LoRA). We optimize LoRA modules, the prototype extractor, and vision-to-language projector for one epoch. We set the learning rate to " + }, + { + "bbox": [ + 130, + 319, + 482, + 450 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-4}" + }, + { + "bbox": [ + 130, + 319, + 482, + 450 + ], + "type": "text", + "content": " for LoRA, and " + }, + { + "bbox": [ + 130, + 319, + 482, + 450 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-5}" + }, + { + "bbox": [ + 130, + 319, + 482, + 450 + ], + "type": "text", + "content": " for the other layers. All the weights of the pre-trained language model and vision encoder remain fixed during fine-tuning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 467, + 329, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 467, + 329, + 480 + ], + "spans": [ + { + "bbox": [ + 132, + 467, + 329, + 480 + ], + "type": "text", + "content": "4.3 Zero-shot Multilingual Capability" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 486, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 617 + ], + "type": "text", + "content": "We evaluate SQ-LLaVA on ten benchmarks, covering a range of academic Visual Question Answering (VQA) tasks and recent instruction tuning tasks designed for large vision language models. The academic VQA includes VQA-v2 [10] and VizWiz [11]. GQA [13] is a fine-grained real-world visual reasoning and question-answering benchmark. ScienceQA [25] is a benchmark with rich subjects (natural science, language science, and social science). TextVQA [39] requires the model to recognize the texts in the image. LLaVA (in the wild) and MM-Vet [53] use GPT4 to assess the capability of the models for testing. With manually designed questions, MM-Bench and MMBench-CN [23] evaluate the model's vision-related reasoning and perception for English and Chinese, respectively. POPE [20] is a benchmark for evaluating object hallucination [34]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "content": "In Table 1, we quantitatively compare between SQ-LLaVA and existing models. SQ-LLaVA-7B and SQ-LLaVA-13B trained on two instruction datasets [5,21] outperform previous methods in six out of ten visual instruction tuning tasks. To be specific, SQ-LLaVA-7B achieves " + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "inline_equation", + "content": "17.2\\%" + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "content": " improvement over LLaVA-v1.5-7B" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "G. Sun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 179, + 480, + 308 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 1: Comparison with state-of-the-art methods on ten benchmarks. After training on the same instruction data, SQ-LLaVA train on [21] and SQ-LLaVA* trained on [5] surpass their baseline model LLaVA-v1.5 and ShareGPT4V on 9 out of 10 and 6 out of 10 benchmarks in the 7B scale, and 8 out of 10 and 6 out of 10 in the 13B scale. The best results are bold and the second-best results are underlined." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 179, + 480, + 308 + ], + "lines": [ + { + "bbox": [ + 136, + 179, + 480, + 308 + ], + "spans": [ + { + "bbox": [ + 136, + 179, + 480, + 308 + ], + "type": "table", + "html": "
LLMModelVQA v2GQAVizWizSQA1VQA TPOPEMM-VetLLaVA WMMBMMB CN
7BInstructBLIP [8]-49.234.560.550.1-26.260.936.023.7
Qwen-VL [3]78.859.335.267.163.8---38.27.4
Qwen-VL-chat [3]78.257.538.968.261.5---60.656.7
LLaVA-v1.5 [21]78.562.050.066.858.285.930.563.464.358.3
ShareGPT4V [5]80.663.357.268.460.486.837.672.668.862.2
SQ-LLaVA79.262.854.068.958.687.732.566.366.258.1
SQ-LLaVA*80.363.755.370.560.587.237.674.366.660.0
13BInstructBLIP [8]-49.533.463.150.778.925.658.2--
LLaVA-v1.5 [21]80.063.353.671.658.285.935.470.767.763.6
ShareGPT4V [5]81.064.855.671.262.2-43.179.968.563.7
SQ-LLaVA80.163.654.669.860.287.735.574.668.762.0
SQ-LLaVA*81.365.058.271.561.987.439.780.768.562.5
", + "image_path": "2b2598b2b7c3ec6ac7e34f5ef69676fe089dfb07523f27f0086b623aafbc1b52.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 331, + 482, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 482, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 482, + 559 + ], + "type": "text", + "content": "on the LLaVA (in the wild) benchmark, indicating the superior capabilities of our model in tasks such as detailed description and complex reasoning. Also, SQ-LLavA-7B improves over previous methods on ScienceQA, indicating that our model excels in understanding and reasoning over scientific content and can effectively handle multi-modal information. The improvement in ScienceQA suggests strong capabilities in multi-hop reasoning, comprehension of complex scientific concepts, and the ability to utilize context and explanations to derive correct answers. SQ-LLaVA-7B has a steady improvement over LLaVA-v1.5-7B and ShareGPT4V-7B on the POPE benchmark, and the " + }, + { + "bbox": [ + 130, + 331, + 482, + 559 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 130, + 331, + 482, + 559 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 331, + 482, + 559 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 331, + 482, + 559 + ], + "type": "text", + "content": " improvement indicates that our proposed method has better reliability and trustworthiness since POPE is a task designed to evaluate the phenomenon of object hallucination [20, 34]. In the bottom section of Table 1, the proposed SQ-LLaVA-13B surpasses previous works in six out of ten benchmarks, indicating the scalability of our method on larger LLM. Notably, the performance inconsistency on some datasets might be due to the unsupervised prototype extractor (lacking pixel-wise guidance) in our model. To mitigate this issue, we could leverage pseudo object masks (e.g., given by the pre-trained segment anything) in learning prototypes. Despite this limitation, all the improvements are achieved with significantly fewer trainable parameters compared to other methods [3, 5, 21]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 574, + 308, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 574, + 308, + 586 + ], + "spans": [ + { + "bbox": [ + 131, + 574, + 308, + 586 + ], + "type": "text", + "content": "4.4 Visual Information Discovery" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 594, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 641 + ], + "type": "text", + "content": "In this experiment, we showcase the diversity and reliability of the proposed SQ-LLaVA through various qualitative applications, including detailed image description, visual information summary, and visual self-questioning. We also present quantitative results on the task of image captioning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "content": "Abilities of SQ-LLaVA Through Qualitative Samples. SQ-LLaVA exhibits numerous advanced capabilities compared to traditional vision-language models." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "type": "text", + "content": "SQ-LLaVA: Self-Questioning for Large Vision-Language Assistant" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 115, + 194, + 154 + ], + "blocks": [ + { + "bbox": [ + 138, + 115, + 194, + 154 + ], + "lines": [ + { + "bbox": [ + 138, + 115, + 194, + 154 + ], + "spans": [ + { + "bbox": [ + 138, + 115, + 194, + 154 + ], + "type": "image", + "image_path": "136c42a072a95726e4d9809d77f134ccfcb9ee16686c2c4f4b300baa8370bb1d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 118, + 269, + 124 + ], + "lines": [ + { + "bbox": [ + 220, + 118, + 269, + 124 + ], + "spans": [ + { + "bbox": [ + 220, + 118, + 269, + 124 + ], + "type": "text", + "content": "GPT4-V Generated" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 196, + 126, + 294, + 155 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 196, + 126, + 294, + 141 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 126, + 294, + 141 + ], + "spans": [ + { + "bbox": [ + 196, + 126, + 294, + 141 + ], + "type": "inline_equation", + "content": "Q_{1}" + }, + { + "bbox": [ + 196, + 126, + 294, + 141 + ], + "type": "text", + "content": " : What makes this particular toothbrush stand out among others in the image? " + }, + { + "bbox": [ + 196, + 126, + 294, + 141 + ], + "type": "inline_equation", + "content": "Q_{2}" + }, + { + "bbox": [ + 196, + 126, + 294, + 141 + ], + "type": "text", + "content": " : What color is the background?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 196, + 141, + 279, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 141, + 279, + 155 + ], + "spans": [ + { + "bbox": [ + 196, + 141, + 279, + 155 + ], + "type": "text", + "content": "Q2: What color is the background? \nQ3: What is this product?" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 174, + 157, + 260, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 157, + 260, + 163 + ], + "spans": [ + { + "bbox": [ + 174, + 157, + 260, + 163 + ], + "type": "text", + "content": "SQ-LLaVA Visual Self-Questioning" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 140, + 165, + 297, + 184 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 140, + 165, + 297, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 165, + 297, + 173 + ], + "spans": [ + { + "bbox": [ + 140, + 165, + 297, + 173 + ], + "type": "inline_equation", + "content": "SQ_{1}" + }, + { + "bbox": [ + 140, + 165, + 297, + 173 + ], + "type": "text", + "content": ": What are some of the key aspects that this image focuses on?" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 140, + 173, + 255, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 173, + 255, + 178 + ], + "spans": [ + { + "bbox": [ + 140, + 173, + 255, + 178 + ], + "type": "inline_equation", + "content": "S Q_{2}" + }, + { + "bbox": [ + 140, + 173, + 255, + 178 + ], + "type": "text", + "content": " : What is the primary color of the toothbrush?" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 140, + 178, + 293, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 178, + 293, + 184 + ], + "spans": [ + { + "bbox": [ + 140, + 178, + 293, + 184 + ], + "type": "inline_equation", + "content": "S Q_{3}" + }, + { + "bbox": [ + 140, + 178, + 293, + 184 + ], + "type": "text", + "content": ": What is the expression of the face drawn on the toothbrush?" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 139, + 198, + 199, + 239 + ], + "blocks": [ + { + "bbox": [ + 139, + 198, + 199, + 239 + ], + "lines": [ + { + "bbox": [ + 139, + 198, + 199, + 239 + ], + "spans": [ + { + "bbox": [ + 139, + 198, + 199, + 239 + ], + "type": "image", + "image_path": "ece03bb4b799ad9b0528600b4496451126f4c03af0cddf96704da72e8299f59b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 223, + 201, + 271, + 207 + ], + "lines": [ + { + "bbox": [ + 223, + 201, + 271, + 207 + ], + "spans": [ + { + "bbox": [ + 223, + 201, + 271, + 207 + ], + "type": "text", + "content": "GPT4-V Generated" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 200, + 209, + 294, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 209, + 294, + 248 + ], + "spans": [ + { + "bbox": [ + 200, + 209, + 294, + 248 + ], + "type": "inline_equation", + "content": "Q_{1}" + }, + { + "bbox": [ + 200, + 209, + 294, + 248 + ], + "type": "text", + "content": " : What is the setting of the image? " + }, + { + "bbox": [ + 200, + 209, + 294, + 248 + ], + "type": "inline_equation", + "content": "Q_{2}" + }, + { + "bbox": [ + 200, + 209, + 294, + 248 + ], + "type": "text", + "content": " : What are some activities that people are engaging in within the image? " + }, + { + "bbox": [ + 200, + 209, + 294, + 248 + ], + "type": "inline_equation", + "content": "Q_{3}" + }, + { + "bbox": [ + 200, + 209, + 294, + 248 + ], + "type": "text", + "content": " : What are some benefits of spending time outdoors in a beachside setting, especially on a sunny day?" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 173, + 250, + 260, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 250, + 260, + 257 + ], + "spans": [ + { + "bbox": [ + 173, + 250, + 260, + 257 + ], + "type": "text", + "content": "SQ-LLaVA Visual Self-Questioning" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 140, + 258, + 284, + 285 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 140, + 258, + 284, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 258, + 284, + 272 + ], + "spans": [ + { + "bbox": [ + 140, + 258, + 284, + 272 + ], + "type": "inline_equation", + "content": "S_{Q_1}" + }, + { + "bbox": [ + 140, + 258, + 284, + 272 + ], + "type": "text", + "content": ": What is the hair color of the man that is looking up at the ocean?" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 140, + 272, + 222, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 272, + 222, + 277 + ], + "spans": [ + { + "bbox": [ + 140, + 272, + 222, + 277 + ], + "type": "inline_equation", + "content": "SQ_{2}" + }, + { + "bbox": [ + 140, + 272, + 222, + 277 + ], + "type": "text", + "content": ": What color is the man's shirt?" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 140, + 278, + 266, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 278, + 266, + 285 + ], + "spans": [ + { + "bbox": [ + 140, + 278, + 266, + 285 + ], + "type": "inline_equation", + "content": "SQ_{3}" + }, + { + "bbox": [ + 140, + 278, + 266, + 285 + ], + "type": "text", + "content": ": In which part is the bench, the top or the bottom?" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 306, + 116, + 362, + 153 + ], + "blocks": [ + { + "bbox": [ + 306, + 116, + 362, + 153 + ], + "lines": [ + { + "bbox": [ + 306, + 116, + 362, + 153 + ], + "spans": [ + { + "bbox": [ + 306, + 116, + 362, + 153 + ], + "type": "image", + "image_path": "70b7a2b0d2bb65837892e832bb74fd4f3e7053a174d41ef5baa6557c83736bc7.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 391, + 117, + 439, + 123 + ], + "lines": [ + { + "bbox": [ + 391, + 117, + 439, + 123 + ], + "spans": [ + { + "bbox": [ + 391, + 117, + 439, + 123 + ], + "type": "text", + "content": "GPT4-V Generated" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 362, + 125, + 467, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 125, + 467, + 137 + ], + "spans": [ + { + "bbox": [ + 362, + 125, + 467, + 137 + ], + "type": "inline_equation", + "content": "Q_{1}" + }, + { + "bbox": [ + 362, + 125, + 467, + 137 + ], + "type": "text", + "content": ": What type of environment can the sheep be seen in?" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 363, + 138, + 466, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 138, + 466, + 156 + ], + "spans": [ + { + "bbox": [ + 363, + 138, + 466, + 156 + ], + "type": "inline_equation", + "content": "Q_{2}" + }, + { + "bbox": [ + 363, + 138, + 466, + 156 + ], + "type": "text", + "content": " : What are the sheep doing in the image? " + }, + { + "bbox": [ + 363, + 138, + 466, + 156 + ], + "type": "inline_equation", + "content": "Q_{3}" + }, + { + "bbox": [ + 363, + 138, + 466, + 156 + ], + "type": "text", + "content": " How do sheep contribute to the ecosystem of a meadow or field?" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 343, + 158, + 430, + 165 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 158, + 430, + 165 + ], + "spans": [ + { + "bbox": [ + 343, + 158, + 430, + 165 + ], + "type": "text", + "content": "SQ-LLaVA Visual Self-Questioning" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 167, + 472, + 194 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 306, + 167, + 472, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 167, + 472, + 181 + ], + "spans": [ + { + "bbox": [ + 306, + 167, + 472, + 181 + ], + "type": "inline_equation", + "content": "SQ_{1}" + }, + { + "bbox": [ + 306, + 167, + 472, + 181 + ], + "type": "text", + "content": ": In the vicinity of this group of sheep, how many horses are there? a) Five b) Four c) One d) Zero" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 181, + 472, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 181, + 472, + 194 + ], + "spans": [ + { + "bbox": [ + 306, + 181, + 472, + 194 + ], + "type": "inline_equation", + "content": "S_{Q2}" + }, + { + "bbox": [ + 306, + 181, + 472, + 194 + ], + "type": "text", + "content": ": In this image, what kind of animals can you see in the green field? " + }, + { + "bbox": [ + 306, + 181, + 472, + 194 + ], + "type": "inline_equation", + "content": "S_{Q3}" + }, + { + "bbox": [ + 306, + 181, + 472, + 194 + ], + "type": "text", + "content": ": Are there any cows in the picture?" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 306, + 200, + 364, + 239 + ], + "blocks": [ + { + "bbox": [ + 306, + 200, + 364, + 239 + ], + "lines": [ + { + "bbox": [ + 306, + 200, + 364, + 239 + ], + "spans": [ + { + "bbox": [ + 306, + 200, + 364, + 239 + ], + "type": "image", + "image_path": "de62dffe85e775b4f6f0a44f297a821f4725bd1c697cceac3544561065315f45.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 394, + 201, + 441, + 207 + ], + "lines": [ + { + "bbox": [ + 394, + 201, + 441, + 207 + ], + "spans": [ + { + "bbox": [ + 394, + 201, + 441, + 207 + ], + "type": "text", + "content": "GPT4-V Generated" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 303, + 481, + 337 + ], + "lines": [ + { + "bbox": [ + 131, + 303, + 481, + 337 + ], + "spans": [ + { + "bbox": [ + 131, + 303, + 481, + 337 + ], + "type": "text", + "content": "Fig.5: Visual self-questioning of SQ-LLaVA-7B. Comparing to the question data provided by GPT4-V (data collected by LLaVA-v1.5 [21]), SQ-LLaVA can generate questions with higher diversity i.e. multiple choice and tricky questions." + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "bbox": [ + 343, + 250, + 430, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 250, + 430, + 257 + ], + "spans": [ + { + "bbox": [ + 343, + 250, + 430, + 257 + ], + "type": "text", + "content": "SQ-LLaVA Visual Self-Questioning" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 306, + 258, + 462, + 291 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 306, + 258, + 462, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 258, + 462, + 272 + ], + "spans": [ + { + "bbox": [ + 306, + 258, + 462, + 272 + ], + "type": "inline_equation", + "content": "S_{Q1}" + }, + { + "bbox": [ + 306, + 258, + 462, + 272 + ], + "type": "text", + "content": ": The two small children are sitting on a tricycle. Which place is it? Town or village?" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 306, + 272, + 422, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 272, + 422, + 278 + ], + "spans": [ + { + "bbox": [ + 306, + 272, + 422, + 278 + ], + "type": "inline_equation", + "content": "SQ_{2}" + }, + { + "bbox": [ + 306, + 272, + 422, + 278 + ], + "type": "text", + "content": ": What can you tell me about the child's age?" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 306, + 278, + 459, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 278, + 459, + 291 + ], + "spans": [ + { + "bbox": [ + 306, + 278, + 459, + 291 + ], + "type": "inline_equation", + "content": "S_{Q3}" + }, + { + "bbox": [ + 306, + 278, + 459, + 291 + ], + "type": "text", + "content": ": What does the boy's action of swinging a bat suggest about his play or sport?" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "text" + }, + { + "type": "table", + "bbox": [ + 135, + 380, + 327, + 466 + ], + "blocks": [ + { + "bbox": [ + 131, + 348, + 481, + 372 + ], + "lines": [ + { + "bbox": [ + 131, + 348, + 481, + 372 + ], + "spans": [ + { + "bbox": [ + 131, + 348, + 481, + 372 + ], + "type": "text", + "content": "Table 2: Comparison of image captioning on four datasets. SQ-LLaVA is trained on data collected by [21]. SQ-LLaVA* is trained on data collected by [5]." + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 380, + 327, + 466 + ], + "lines": [ + { + "bbox": [ + 135, + 380, + 327, + 466 + ], + "spans": [ + { + "bbox": [ + 135, + 380, + 327, + 466 + ], + "type": "table", + "html": "
ModelFlickr30kNocapsoutConceptual
B@4CIDErB@4CIDErB@4CIDEr
ClipCap [28]17.2141.6520.3251.741.4723.74
DiscriTune [9]18.4844.7824.1057.061.7128.01
LLaVA-v1.5 [21]28.6781.2735.78103.562.7939.20
ShareGPT4V [5]31.0086.1737.19107.452.7837.86
SQ-LLaVA29.8883.5136.21105.422.9039.49
SQ-LLaVA*31.4983.1437.20107.422.9141.24
", + "image_path": "7767a9ca24062f40c2beb45e431a9f380e948cedb5236f6a433b1f8111f07b48.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 168, + 467, + 282, + 476 + ], + "lines": [ + { + "bbox": [ + 168, + 467, + 282, + 476 + ], + "spans": [ + { + "bbox": [ + 168, + 467, + 282, + 476 + ], + "type": "text", + "content": "(a) Zero-shot image captioning" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 37 + }, + { + "type": "table", + "bbox": [ + 329, + 380, + 471, + 465 + ], + "blocks": [ + { + "bbox": [ + 329, + 380, + 471, + 465 + ], + "lines": [ + { + "bbox": [ + 329, + 380, + 471, + 465 + ], + "spans": [ + { + "bbox": [ + 329, + 380, + 471, + 465 + ], + "type": "table", + "html": "
ModelZero-shotFine-tune
B@4CIDErB@4CIDEr
ClipCap8.5037.0332.60108.55
DiscriTune13.9953.2032.31105.40
eP-ALM [38]29.4797.2233.35111.63
MAPL [26]12.3054.3036.45125.20
LLaVA-v1.529.96111.46--
SQ-LLaVA29.85110.7740.76136.78
", + "image_path": "38aaa20ea581ef709118c303c92d0712b61998d81e6b4499b3e234b21d51b191.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 340, + 467, + 456, + 476 + ], + "lines": [ + { + "bbox": [ + 340, + 467, + 456, + 476 + ], + "spans": [ + { + "bbox": [ + 340, + 467, + 456, + 476 + ], + "type": "text", + "content": "(b) Image captioning on COCO" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 39 + }, + { + "bbox": [ + 130, + 509, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 482, + 629 + ], + "type": "text", + "content": "Notably, SQ-LLaVA effectively mitigates object hallucination [20, 34], resulting in predictions that are more trustworthy. Additionally, SQ-LLaVA is capable of generating diverse and meaningful questions about a given image without requiring human textual instructions. As illustrated in Fig. 5, the generated questions include multiple-choice, reasoning, binary choice questions, and the content within each question, all of which are either highly correlated or tricky to the given image, such as \"What is the hair color of the man that is looking up at the ocean?\" from the third image and \"Are there cows in the picture?\" from the second image. These examples highlight SQ-LLaVA's ability to produce questions with high diversity and quality." + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "The detailed image description example in Fig. 6 demonstrates that SQ-LLaVA can generate visual information with concrete concepts within the image, such as \"Hyundai\", \"small white dog\", \"Nikky Stephen\". In contrast, LLaVA-v1.5 can" + } + ] + } + ], + "index": 42 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "G. Sun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 116, + 199, + 208 + ], + "blocks": [ + { + "bbox": [ + 137, + 116, + 199, + 208 + ], + "lines": [ + { + "bbox": [ + 137, + 116, + 199, + 208 + ], + "spans": [ + { + "bbox": [ + 137, + 116, + 199, + 208 + ], + "type": "image", + "image_path": "507a676ade812af4fb871183548354e4209fede95d1848284cab55c6f5f0276f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 216, + 117, + 288, + 124 + ], + "lines": [ + { + "bbox": [ + 216, + 117, + 288, + 124 + ], + "spans": [ + { + "bbox": [ + 216, + 117, + 288, + 124 + ], + "type": "text", + "content": "Detailed Image Description" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 304, + 482, + 338 + ], + "lines": [ + { + "bbox": [ + 130, + 304, + 482, + 338 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 482, + 338 + ], + "type": "text", + "content": "Fig. 6: A qualitative evaluation of detailed image descriptions from three models. We highlight the words and sentences that represent how each model describes the main object in the image." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 202, + 125, + 303, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 125, + 303, + 205 + ], + "spans": [ + { + "bbox": [ + 202, + 125, + 303, + 205 + ], + "type": "text", + "content": "SQ-LLaVA-7B (Ours): In this black and white photograph, a man is engrossed in the task of inspecting the engine of a Hyundai car. The car, with its hood open, reveals the intricate details of its engine. The man, dressed in a casual attire of a t-shirt and jeans, is leaning over the car, his attention completely absorbed by the engine. The car's license plate, reading \"VK8-LQ17\", is clearly visible. Adding a touch of companionship to the scene, a small white dog is seen standing next to the car, seemingly curious about the man's actions. The photograph is credited to Nikky Stephen, as indicated by the text at the top of the image. The overall scene paints a picture of a typical day in the life of a car enthusiast." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 139, + 211, + 302, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 211, + 302, + 239 + ], + "spans": [ + { + "bbox": [ + 139, + 211, + 302, + 239 + ], + "type": "text", + "content": "LLaVA-v1.5-7B: The image features a man standing next to a car, with the hood open. He is leaning over the car, possibly working on it or checking something under the hood. A dog is standing nearby, attentively watching the man's actions. The car is parked in a driveway, and the scene appears to be a casual, everyday moment between the man and his dog." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 139, + 244, + 302, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 244, + 302, + 293 + ], + "spans": [ + { + "bbox": [ + 139, + 244, + 302, + 293 + ], + "type": "text", + "content": "ShareGPT4V-7B: The man, dressed in a casual attire of a t-shirt and khaki pants, stands to the right of the car, his attention completely absorbed by the task at hand. The car's license plate, reading 'YKL8-O17', is clearly visible. Adding a touch of companionship to the scene, a white dog stands obediently at the man's feet, perhaps waiting for a command or simply enjoying the company. The photograph is taken from a low angle, giving a unique perspective to the scene. The man, the car, and the dog are the main subjects of this image, each contributing to the narrative of a moment frozen in time. The black and white color scheme adds a timeless quality to the photograph, highlighting the contrast between the man, the car, and the dog." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 309, + 115, + 373, + 201 + ], + "blocks": [ + { + "bbox": [ + 309, + 115, + 373, + 201 + ], + "lines": [ + { + "bbox": [ + 309, + 115, + 373, + 201 + ], + "spans": [ + { + "bbox": [ + 309, + 115, + 373, + 201 + ], + "type": "image", + "image_path": "24fb97b1af7b0ebf481d7b8082670b2d571ae71e0da48c4761af0547fafc2cea.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 389, + 117, + 460, + 124 + ], + "lines": [ + { + "bbox": [ + 389, + 117, + 460, + 124 + ], + "spans": [ + { + "bbox": [ + 389, + 117, + 460, + 124 + ], + "type": "text", + "content": "Detailed Image Description" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 376, + 126, + 473, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 126, + 473, + 201 + ], + "spans": [ + { + "bbox": [ + 376, + 126, + 473, + 201 + ], + "type": "text", + "content": "SQ-LLA-7B (Ours): The image captures a scene on a city street. Dominating the foreground is a traffic sign, painted in striking red and white colors. The sign features a black silhouette of a duck, a universal symbol for caution. Above the duck, there's a yellow light, indicating a stop signal. The background reveals a modern building with large windows, reflecting the urban environment. A tree, standing in front of the building, adds a touch of nature to the scene. The image does not contain any discernible text. The relative positions of the objects suggest that the sign is closer to the viewer, while the building and the tree are further away." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 310, + 207, + 472, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 207, + 472, + 240 + ], + "spans": [ + { + "bbox": [ + 310, + 207, + 472, + 240 + ], + "type": "text", + "content": "LLaVA-v1.5-7B: The image features a street scene with a unique combination of traffic signs. There is a stop sign with a duck on it, which is an unusual sight. The duck is positioned in the middle of the stop sign, making it stand out. In the background, there are several cars parked or driving along the street. The cars are of various sizes and are scattered throughout the scene. The presence of the duck on the stop sign adds a touch of humor to the otherwise ordinary street scene." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 309, + 248, + 473, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 248, + 473, + 292 + ], + "spans": [ + { + "bbox": [ + 309, + 248, + 473, + 292 + ], + "type": "text", + "content": "ShareGPT4V-7B: The image captures a scene on a city sidewalk. Dominating the foreground is a traffic sign, painted in a striking red and white color scheme. The sign is triangular, a common shape for traffic signs, and it features a black duck in the center, adding a touch of whimsy to the otherwise mundane object. Just behind the duck sign, there's another sign. This one is diamond-shaped and also red and white, but it's adorned with a green traffic light. The green light stands out against the red and white background, drawing attention to its presence. The signs are affixed to a gray pole, which is firmly planted in the sidewalk." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 361, + 482, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 482, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 482, + 422 + ], + "type": "text", + "content": "only describe the image from a general concept, such as \"car\", \"dog\", or \"everyday moment\". Even though ShareGPT4V can generate detailed descriptions, it still suffers from the issue of object hallucination, such as \"perhaps waiting\" and \"behind the duck sign\". By observation, SQ-LLaVA can describe the image with less unintended text, yielding higher reliability." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 422, + 483, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 483, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 483, + 555 + ], + "type": "text", + "content": "Quantitative Analysis. SQ-LLaVA serves as a general-purpose vision-language model and enables zero-shot image captioning. As indicated in Table 2a, SQ-LLaVA achieves " + }, + { + "bbox": [ + 130, + 422, + 483, + 555 + ], + "type": "inline_equation", + "content": "73\\%" + }, + { + "bbox": [ + 130, + 422, + 483, + 555 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 422, + 483, + 555 + ], + "type": "inline_equation", + "content": "66\\%" + }, + { + "bbox": [ + 130, + 422, + 483, + 555 + ], + "type": "text", + "content": " averaged improvement over ClipCap and DiscriTune on all datasets, indicating the effectiveness of visual instruction tuning. Compared with the baseline model LLaVA-V1.5 [21], SQ-LLaVA achieves " + }, + { + "bbox": [ + 130, + 422, + 483, + 555 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 130, + 422, + 483, + 555 + ], + "type": "text", + "content": " averaged improvement on all datasets, indicating the effectiveness of visual self-questioning. Also, we find SQ-LLaVA* surpasses the baseline model ShareGPT4V [5] on Nocaps out and Conceptual dataset, demonstrating the adaptability of the proposed method on unseen testing images from new domains. Moreover, as shown in Table 2b, SQ-LLaVA can be easily adapted to COCO captioning via instruction tuning on short descriptions." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 572, + 238, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 572, + 238, + 585 + ], + "spans": [ + { + "bbox": [ + 132, + 572, + 238, + 585 + ], + "type": "text", + "content": "4.5 Ablation Study" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "In Table 3, we conduct experiments with different architecture designs and training strategies on five question-answering benchmarks. For a fair comparison, we train the baseline models on our local machine with the same training recipe of LLaVA-LoRA [21]. Specifically, we present our full model and three ablated models by removing one component each time. We adopt the dataset [21] with " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "558\\mathrm{k}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " for pre-training (PT) and " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "665\\mathrm{k}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " for fine-tuning (FT). As compared, self" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "type": "text", + "content": "SQ-LLaVA: Self-Questioning for Large Vision-Language Assistant" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 166, + 168, + 451, + 285 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "text", + "content": "Table 3: Ablation study of training strategy on visual instruction tasks. All models are in 7B scale with three components of ViT-LoRA (V-LoRA), self-questioning (SQ), and prototype extractor (Proto). We provide experiments on two instruction datasets [5,21] with different pre-training (PT) and instruction tuning (IT) data scales." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 166, + 168, + 451, + 285 + ], + "lines": [ + { + "bbox": [ + 166, + 168, + 451, + 285 + ], + "spans": [ + { + "bbox": [ + 166, + 168, + 451, + 285 + ], + "type": "table", + "html": "
PTITV-LoRASQProtoVizWiz\\( \\mathrm{SQA}^I \\)\\( \\mathrm{VQAT} \\)POPE\\( \\mathrm{LLaVA}^W \\)Avg.
558K665KXXX49.468.458.286.567.165.9
✓X✓52.467.958.687.765.666.4
✓✓X52.668.457.888.267.366.9
X✓✓53.469.358.187.967.967.3
✓✓✓54.068.958.687.768.167.5
1200K700KXXX51.568.958.986.872.167.6
✓X✓54.068.960.287.271.668.4
✓✓X55.469.259.586.877.369.6
X✓✓54.270.360.587.572.769.0
✓✓✓55.370.560.587.274.369.6
", + "image_path": "e74a3a3c94ca7648145653e0fb90d9ed91b84155f960cd721737b6fff669c19f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 307, + 482, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 307, + 482, + 378 + ], + "spans": [ + { + "bbox": [ + 130, + 307, + 482, + 378 + ], + "type": "text", + "content": "questioning (SQ) brings a consistent performance boost on all the five benchmarks, indicating the effectiveness of visual self-questioning on improving visual language understanding. Besides, we introduce the prototype extractor (Proto) to enhance visual representation, achieving " + }, + { + "bbox": [ + 130, + 307, + 482, + 378 + ], + "type": "inline_equation", + "content": "0.9\\%" + }, + { + "bbox": [ + 130, + 307, + 482, + 378 + ], + "type": "text", + "content": " improvement in average accuracy among five benchmark. With all three components incorporated, we observe a " + }, + { + "bbox": [ + 130, + 307, + 482, + 378 + ], + "type": "inline_equation", + "content": "2.4\\%" + }, + { + "bbox": [ + 130, + 307, + 482, + 378 + ], + "type": "text", + "content": " improvement in average accuracy." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 379, + 482, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 379, + 482, + 438 + ], + "spans": [ + { + "bbox": [ + 130, + 379, + 482, + 438 + ], + "type": "text", + "content": "As shown by the bottom block of Table 3, we conduct experiments with the same ablation settings but with a larger scale of the visual instruction data [5] (i.e., for both PT and IT). Overall, SQ-LLaVA achieves " + }, + { + "bbox": [ + 130, + 379, + 482, + 438 + ], + "type": "inline_equation", + "content": "2.4\\%" + }, + { + "bbox": [ + 130, + 379, + 482, + 438 + ], + "type": "text", + "content": " improvement over the baseline model after training on the smaller dataset and achieves " + }, + { + "bbox": [ + 130, + 379, + 482, + 438 + ], + "type": "inline_equation", + "content": "3.0\\%" + }, + { + "bbox": [ + 130, + 379, + 482, + 438 + ], + "type": "text", + "content": " improvement after training on the larger dataset." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 457, + 226, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 457, + 226, + 470 + ], + "spans": [ + { + "bbox": [ + 131, + 457, + 226, + 470 + ], + "type": "text", + "content": "5 Conclusions" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 482, + 482, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 482, + 482, + 663 + ], + "spans": [ + { + "bbox": [ + 130, + 482, + 482, + 663 + ], + "type": "text", + "content": "This work has introduced SQ-LLaVA, a new visual instruction tuning method that enhances general-purpose vision-language understanding and image-oriented question answering through visual self-questioning. Our experiments demonstrate that SQ-LLaVA achieves superior performance with fewer training parameters and instructional data. We have also conducted a comprehensive study on visual discovery/reasoning tasks and found that SQ-LLaVA generalizes well to a wide range of unseen tasks and outperforms several state-of-the-art methods. Qualitative assessments show that SQ-LLaVA strengthens visual representation and domain alignment, effectively reducing object hallucination and improving the semantic interpretation of images. Our findings highlight the potential of visual self-questioning as a powerful training strategy for the visual instruction tuning framework, paving the way for realizing more efficient and effective large vision-language models. Particularly, SQ-LLaVA frames questioning as an intrinsic goal of tuning LLMs, encouraging the exploration of the model's curiosity (the ability to ask questions proactively) in solving complex problems." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "G. Sun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 142, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 137, + 142, + 285, + 153 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 142, + 285, + 153 + ], + "spans": [ + { + "bbox": [ + 137, + 142, + 285, + 153 + ], + "type": "text", + "content": "1. Gpt-4v(ison) system card (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 137, + 155, + 481, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 155, + 481, + 186 + ], + "spans": [ + { + "bbox": [ + 137, + 155, + 481, + 186 + ], + "type": "text", + "content": "2. Agrawal, H., Desai, K., Wang, Y., Chen, X., Jain, R., Johnson, M., Batra, D., Parikh, D., Lee, S., Anderson, P.: Nocaps: Novel object captioning at scale. In: ICCV (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 189, + 481, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 189, + 481, + 220 + ], + "spans": [ + { + "bbox": [ + 138, + 189, + 481, + 220 + ], + "type": "text", + "content": "3. Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. In: ArXiv (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 222, + 481, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 222, + 481, + 287 + ], + "spans": [ + { + "bbox": [ + 138, + 222, + 481, + 287 + ], + "type": "text", + "content": "4. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., Ziegler, D., Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Chess, B., Clark, J., Berner, C., McCandlish, S., Radford, A., Sutskever, I., Amodei, D.: Language models are few-shot learners. In: NeurIPS (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 288, + 481, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 288, + 481, + 319 + ], + "spans": [ + { + "bbox": [ + 138, + 288, + 481, + 319 + ], + "type": "text", + "content": "5. Chen, L., Li, J., wen Dong, X., Zhang, P., He, C., Wang, J., Zhao, F., Lin, D.: Sharegpt4v: Improving large multi-modal models with better captions. In: ArXiv (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 321, + 481, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 321, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 321, + 481, + 342 + ], + "type": "text", + "content": "6. Chen, X., Fang, H., Lin, T.Y., Vedantam, R., Gupta, S., Dollar, P., Zitnick, C.L.: Microsoft coco captions: Data collection and evaluation server. In: ArXiv (2015)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 344, + 481, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 344, + 481, + 464 + ], + "spans": [ + { + "bbox": [ + 138, + 344, + 481, + 464 + ], + "type": "text", + "content": "7. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S., Schuh, P., Shi, K., Tsvyashchenko, S., Maynez, J., Rao, A., Barnes, P., Tay, Y., Shazeer, N.M., Prabhakaran, V., Reif, E., Du, N., Hutchinson, B.C., Pope, R., Bradbury, J., Austin, J., Isard, M., Gur-Ari, G., Yin, P., Duke, T., Levskaya, A., Ghemawat, S., Dev, S., Michalewski, H., Garcia, X., Misra, V., Robinson, K., Fedus, L., Zhou, D., Ippolito, D., Luan, D., Lim, H., Zoph, B., Spiridonov, A., Sepassi, R., Dohan, D., Agrawal, S., Omernick, M., Dai, A.M., Pillai, T.S., Pellat, M., Lewkowycz, A., Moreira, E., Child, R., Polozov, O., Lee, K., Zhou, Z., Wang, X., Saeta, B., Diaz, M., Firat, O., Catasta, M., Wei, J., Meier-Hellstern, K.S., Eck, D., Dean, J., Petrov, S., Fiedel, N.: Palm: Scaling language modeling with pathways. In: J. Mach. Learn. Res. (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 465, + 481, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 465, + 481, + 497 + ], + "spans": [ + { + "bbox": [ + 138, + 465, + 481, + 497 + ], + "type": "text", + "content": "8. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B.A., Fung, P., Hoi, S.C.H.: Instructclip: Towards general-purpose vision-language models with instruction tuning. In: NeurIPS (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 498, + 481, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 498, + 481, + 520 + ], + "spans": [ + { + "bbox": [ + 138, + 498, + 481, + 520 + ], + "type": "text", + "content": "9. Dessi, R., Bevilacqua, M., Gualdoni, E., Rakotonirina, N.C., Franzon, F., Baroni, M.: Cross-domain image captioning with discriminative finetuning. In: CVPR (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 134, + 521, + 481, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 521, + 481, + 553 + ], + "spans": [ + { + "bbox": [ + 134, + 521, + 481, + 553 + ], + "type": "text", + "content": "10. Goyal, Y., Khot, T., Summers-Stay, D., Batra, D., Parikh, D.: Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In: CVPR (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 555, + 481, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 555, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 134, + 555, + 481, + 586 + ], + "type": "text", + "content": "11. Gurari, D., Li, Q., Stangl, A., Guo, A., Lin, C., Grauman, K., Luo, J., Bigham, J.P.: Vizwiz grand challenge: Answering visual questions from blind people. In: CVPR (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 588, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 588, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 134, + 588, + 481, + 609 + ], + "type": "text", + "content": "12. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. In: ICLR (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 610, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 610, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 610, + 481, + 632 + ], + "type": "text", + "content": "13. Hudson, D.A., Manning, C.D.: Gqa: A new dataset for real-world visual reasoning and compositional question answering. In: CVPR (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 633, + 481, + 665 + ], + "type": "text", + "content": "14. Ikotun, A.M., Ezugwu, A.E., Abualigah, L., Abuhaija, B., Heming, J.: K-means clustering algorithms: A comprehensive review, variants analysis, and advances in the era of big data. In: Information Sciences (2023)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "type": "text", + "content": "SQ-LLaVA: Self-Questioning for Large Vision-Language Assistant" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "15. Kazemzadeh, S., Ordonez, V., Matten, M., Berg, T.: ReferItGame: Referring to objects in photographs of natural scenes. In: EMNLP (2014)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 481, + 171 + ], + "type": "text", + "content": "16. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.B.: Segment anything. In: ICCV (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 481, + 215 + ], + "type": "text", + "content": "17. Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., Bernstein, M.S., Fei-Fei, L.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. In: IJCV (2016)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 216, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 216, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 133, + 216, + 481, + 248 + ], + "type": "text", + "content": "18. Li, J., Pan, K., Ge, Z., Gao, M., Zhang, H., Ji, W., Zhang, W., Chua, T.S., Tang, S., Zhuang, Y.: Empowering vision-language models to follow interleaved vision-language instructions. In: ICLR (2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 249, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 249, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 133, + 249, + 481, + 281 + ], + "type": "text", + "content": "19. Li, J., Li, D., Xiong, C., Hoi, S.C.H.: Blip: Bootstrapping language-image pretraining for unified vision-language understanding and generation. In: International Conference on Machine Learning (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 282, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 304 + ], + "type": "text", + "content": "20. Li, Y., Du, Y., Zhou, K., Wang, J., Zhao, X., Wen, J.R.: Evaluating object hallucination in large vision-language models. In: EMNLP (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "type": "text", + "content": "21. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. In: ArXiv (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 326, + 481, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 336 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 336 + ], + "type": "text", + "content": "22. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: NeurIPS (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 337, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 337, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 133, + 337, + 481, + 369 + ], + "type": "text", + "content": "23. Liu, Y., Duan, H., Zhang, Y., Li, B., Zhang, S., Zhao, W., Yuan, Y., Wang, J., He, C., Liu, Z., Chen, K., Lin, D.: Mmbench: Is your multi-modal model an all-around player? In: ArXiv (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 369, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 481, + 380 + ], + "type": "text", + "content": "24. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: ICLR (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 380, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 380, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 133, + 380, + 481, + 413 + ], + "type": "text", + "content": "25. Lu, P., Mishra, S., Xia, T., Qiu, L., Chang, K.W., Zhu, S.C., Tafjord, O., Clark, P., Kalyan, A.: Learn to explain: Multimodal reasoning via thought chains for science question answering. In: NeurIPS (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 414, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 414, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 133, + 414, + 481, + 445 + ], + "type": "text", + "content": "26. Manas, O., López, P.R., Ahmadi, S., Nematzadeh, A., Goyal, Y., Agrawal, A.: Mapl: Parameter-efficient adaptation of unimodal pre-trained models for vision-language few-shot prompting. In: EACL (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 447, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 481, + 468 + ], + "type": "text", + "content": "27. Mishra, A., Shekhar, S., Singh, A.K., Chakraborty, A.: Ocr-vqa: Visual question answering by reading text in images. In: ICDAR (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 468, + 441, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 441, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 441, + 479 + ], + "type": "text", + "content": "28. Mokady, R.: Clipcap: Clip prefix for image captioning. In: ArXiv (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 479, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 479, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 133, + 479, + 481, + 501 + ], + "type": "text", + "content": "29. Ordonez, V., Kulkarni, G., Berg, T.: Im2text: Describing images using 1 million captioned photographs. In: NeurIPS (2011)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 502, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 502, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 133, + 502, + 481, + 555 + ], + "type": "text", + "content": "30. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C.L., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., Schulman, J., Hilton, J., Kelton, F., Miller, L.E., Simens, M., Askell, A., Welinder, P., Christiano, P.F., Leike, J., Lowe, R.J.: Training language models to follow instructions with human feedback. In: NeurIPS (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 555, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 555, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 133, + 555, + 481, + 609 + ], + "type": "text", + "content": "31. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C.L., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., Schulman, J., Hilton, J., Kelton, F., Miller, L.E., Simens, M., Askell, A., Welinder, P., Christiano, P.F., Leike, J., Lowe, R.J.: Training language models to follow instructions with human feedback. In: NeurIPS (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 611, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 481, + 632 + ], + "type": "text", + "content": "32. Papineni, K., Roukos, S., Ward, T., Zhu, W.J.: Bleu: A method for automatic evaluation of machine translation. In: ACL (2002)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 133, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 633, + 481, + 665 + ], + "type": "text", + "content": "33. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: ICML (2021)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "G. Sun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 138 + ], + "type": "text", + "content": "34. Rohrbach, A., Hendricks, L.A., Burns, K., Darrell, T., Saenko, K.: Object hallucination in image captioning. In: EMNLP (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 138, + 482, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 138, + 482, + 214 + ], + "spans": [ + { + "bbox": [ + 132, + 138, + 482, + 214 + ], + "type": "text", + "content": "35. Sanh, V., Webson, A., Raffel, C., Bach, S.H., Sutawika, L., Alyafeai, Z., Chaffin, A., Stiegler, A., Scao, T.L., Raja, A., Dey, M., Bari, M.S., Xu, C., Thakker, U., Sharma, S., Szczechla, E., Kim, T., Chhablani, G., Nayak, N.V., Datta, D., Chang, J.D., Jiang, M.T.J., Wang, H., Manica, M., Shen, S., Yong, Z.X., Pandey, H., Bawden, R., Wang, T., Neeraj, T., Rozen, J., Sharma, A., Santilli, A., Févry, T., Fries, J.A., Teehan, R., Biderman, S., Gao, L., Bers, T., Wolf, T., Rush, A.M.: Multitask prompted training enables zero-shot task generalization. In: ICLR (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 214, + 482, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 214, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 214, + 482, + 247 + ], + "type": "text", + "content": "36. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. In: NeurIPS Workshop (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 247, + 482, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 247, + 482, + 290 + ], + "spans": [ + { + "bbox": [ + 132, + 247, + 482, + 290 + ], + "type": "text", + "content": "37. Sharma, P., Ding, N., Goodman, S., Soricut, R.: Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In: Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 289, + 482, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 482, + 311 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 482, + 311 + ], + "type": "text", + "content": "38. Shukor, M., Dancette, C., Cord, M.: ep-alm: Efficient perceptual augmentation of language models. In: ICCV (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 311, + 482, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 311, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 132, + 311, + 482, + 331 + ], + "type": "text", + "content": "39. Singh, A., Natarajan, V., Shah, M., Jiang, Y., Chen, X., Batra, D., Parikh, D., Rohrbach, M.: Towards vqa models that can read. In: CVPR (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 331, + 482, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 331, + 482, + 364 + ], + "spans": [ + { + "bbox": [ + 132, + 331, + 482, + 364 + ], + "type": "text", + "content": "40. Sun, G., Bai, Y., Yang, X., Fang, Y., Fu, Y., Tao, Z.: Aligning out-of-distribution web images and caption semantics via evidential learning. In: Proceedings of the ACM on Web Conference 2024 (2024)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 364, + 482, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 482, + 395 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 482, + 395 + ], + "type": "text", + "content": "41. Tofade, T., Elsner, J., Haines, S.: Best practice strategies for effective use of questions as a teaching tool. In: American journal of pharmaceutical education (2013)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 395, + 482, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 395, + 482, + 438 + ], + "spans": [ + { + "bbox": [ + 132, + 395, + 482, + 438 + ], + "type": "text", + "content": "42. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., Lample, G.: Llama: Open and efficient foundation language models. In: ArXiv (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 438, + 482, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 438, + 482, + 558 + ], + "spans": [ + { + "bbox": [ + 132, + 438, + 482, + 558 + ], + "type": "text", + "content": "43. Touvron, H., Martin, L., Stone, K.R., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., Bikel, D.M., Blecher, L., Ferrer, C.C., Chen, M., Cucurull, G., Esiobu, D., Fernandes, J., Fu, J., Fu, W., Fuller, B., Gao, C., Goswami, V., Goyal, N., Hartshorn, A.S., Hosseini, S., Hou, R., Inan, H., Kardas, M., Kerkez, V., Khabsa, M., Kloumann, I.M., Korenev, A.V., Koura, P.S., Lachaux, M.A., Lavril, T., Lee, J., Liskovich, D., Lu, Y., Mao, Y., Martinet, X., Mihaylov, T., Mishra, P., Molybog, I., Nie, Y., Poulton, A., Reizenstein, J., Rungta, R., Saladi, K., Schelten, A., Silva, R., Smith, E.M., Subramanian, R., Tan, X., Tang, B., Taylor, R., Williams, A., Kuan, J.X., Xu, P., Yan, Z., Zarov, I., Zhang, Y., Fan, A., Kambadur, M., Narang, S., Rodriguez, A., Stojnic, R., Edunov, S., Scialom, T.: Llama 2: Open foundation and fine-tuned chat models. In: ArXiv (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 558, + 482, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 558, + 482, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 558, + 482, + 590 + ], + "type": "text", + "content": "44. Triantafillou, E., Zhu, T.L., Dumoulin, V., Lamblin, P., Xu, K., Goroshin, R., Gelada, C., Swersky, K., Manzagol, P.A., Larochelle, H.: Meta-dataset: A dataset of datasets for learning to learn from few examples. In: ICLR (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 590, + 482, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 482, + 612 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 482, + 612 + ], + "type": "text", + "content": "45. Vattani, A.: K-means requires exponentially many iterations even in the plane. In: Annual Symposium on Computational Geometry (2009)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 612, + 482, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 612, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 612, + 482, + 632 + ], + "type": "text", + "content": "46. Vedantam, R., Zitnick, C.L., Parikh, D.: Cider: Consensus-based image description evaluation. In: CVPR (2015)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 632, + 482, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 482, + 665 + ], + "type": "text", + "content": "47. Wang, J., Sun, G., Wang, P., Liu, D., Dianat, S., Rabbani, M., Rao, R., Tao, Z.: Text is mass: Modeling as stochastic embedding for text-video retrieval. In: CVPR (2024)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "type": "text", + "content": "SQ-LLaVA: Self-Questioning for Large Vision-Language Assistant" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 491 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "48. Wang, Y., Kordi, Y., Mishra, S., Liu, A., Smith, N.A., Khashabi, D., Hajishirzi, H.: Self-instruct: Aligning language models with self-generated instructions. In: ACL (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 237 + ], + "type": "text", + "content": "49. Wang, Y., Mishra, S., Alipoormolabashi, P., Kordi, Y., Mirzaei, A., Arunkumar, A., Ashok, A., Dhanasekaran, A.S., Naik, A., Stap, D., Pathak, E., Karamanolakis, G., Lai, H.G., Purohit, I., Mondal, I., Anderson, J., Kuznia, K., Doshi, K., Patel, M., Pal, K.K., Moradshahi, M., Parmar, M., Purohit, M., Varshney, N., Kaza, P.R., Verma, P., Puri, R.S., Karia, R., Sampat, S.K., Doshi, S., Mishra, S.D., Reddy, S., Patro, S., Dixit, T., Shen, X., Baral, C., Choi, Y., Smith, N.A., Hajishirzi, H., Khashabi, D.: Super-naturalinstructions: Generalization via declarative instructions on 1600+ nlp tasks. In: EMNLP (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 237, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 481, + 259 + ], + "type": "text", + "content": "50. Wei, J., Bosma, M., Zhao, V., Guu, K., Yu, A.W., Lester, B., Du, N., Dai, A.M., Le, Q.V.: Finetuned language models are zero-shot learners. In: ICLR (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "text", + "content": "51. Xu, C., Sun, Q., Zheng, K., Geng, X., Zhao, P., Feng, J., Tao, C., Jiang, D.: Wizardlm: Empowering large language models to follow complex instructions. In: ICLR (2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 293, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 325 + ], + "type": "text", + "content": "52. Young, P., Lai, A., Hodosh, M., Hockenmaier, J.: From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. In: TACL (2014)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 325, + 481, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 481, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 481, + 347 + ], + "type": "text", + "content": "53. Yu, W., Yang, Z., Li, L., Wang, J., Lin, K., Liu, Z., Wang, X., Wang, L.: Mm-vet: Evaluating large multimodal models for integrated capabilities. In: ArXiv (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "text", + "content": "54. Zhang, R., Han, J., Zhou, A., Hu, X., Yan, S., Lu, P., Li, H., Gao, P., Qiao, Y.: Llama-adapter: Efficient fine-tuning of language models with zero-init attention. In: ICLR (2024)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "text", + "content": "55. Zhang, Y., Zhang, R., Gu, J., Zhou, Y., Lipka, N., Yang, D., Sun, T.: Llavar: Enhanced visual instruction tuning for text-rich image understanding. In: ArXiv (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 413, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 481, + 435 + ], + "type": "text", + "content": "56. Zhao, B., Wu, B., Huang, T.: Svit: Scaling up visual instruction tuning. In: ArXiv (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "type": "text", + "content": "57. Zheng, L., Chiang, W.L., Sheng, Y., Zhuang, S., Wu, Z., Zhuang, Y., Lin, Z., Li, Z., Li, D., Xing, E.P., Zhang, H., Gonzalez, J., Stoica, I.: Judging llm-as-a-judge with mt-bench and chatbot arena. In: NeurIPS (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 468, + 481, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 491 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 491 + ], + "type": "text", + "content": "58. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. In: ICLR (2024)" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "G. Sun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/SRPose_ Two-view Relative Pose Estimation with Sparse Keypoints/e429b1d6-1778-4297-b226-264e6bf36452_content_list.json b/2024/SRPose_ Two-view Relative Pose Estimation with Sparse Keypoints/e429b1d6-1778-4297-b226-264e6bf36452_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..67889e525208a25ae059b5e888d966f2d8de548b --- /dev/null +++ b/2024/SRPose_ Two-view Relative Pose Estimation with Sparse Keypoints/e429b1d6-1778-4297-b226-264e6bf36452_content_list.json @@ -0,0 +1,1943 @@ +[ + { + "type": "text", + "text": "SRPose: Two-view Relative Pose Estimation with Sparse Keypoints", + "text_level": 1, + "bbox": [ + 250, + 140, + 751, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rui Yin $^{1}$ , Yulun Zhang $^{2}$ , Zherong Pan $^{3}$ , Jianjun Zhu $^{1}$ , Cheng Wang $^{1}$ , and Biao Jia $^{1^{\\star}}$", + "bbox": [ + 303, + 210, + 697, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Hanglok-Tech, China \n $^{2}$ MoE Key Lab of Artificial Intelligence, Shanghai Jiao Tong University, China \n $^{3}$ LightSpeed Studios, Tencent America", + "bbox": [ + 233, + 260, + 767, + 305 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Two-view pose estimation is essential for map-free visual relocalization and object pose tracking tasks. However, traditional matching methods suffer from time-consuming robust estimators, while deep learning-based pose regressors only cater to camera-to-world pose estimation, lacking generalizability to different image sizes and camera intrinsics. In this paper, we propose SRPose, a sparse keypoint-based framework for two-view relative pose estimation in camera-to-world and object-to-camera scenarios. SRPose consists of a sparse keypoint detector, an intrinsic-calibration position encoder, and promptable prior knowledge-guided attention layers. Given two RGB images of a fixed scene or a moving object, SRPose estimates the relative camera or 6D object pose transformation. Extensive experiments demonstrate that SRPose achieves competitive or superior performance compared to state-of-the-art methods in terms of accuracy and speed, showing generalizability to both scenarios. It is robust to different image sizes and camera intrinsics, and can be deployed with low computing resources. Project page: https://frickyinn.github.io/srpose.", + "bbox": [ + 261, + 342, + 743, + 580 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Relative Pose Estimation $\\cdot$ 6D Object Pose Estimation", + "bbox": [ + 261, + 590, + 718, + 607 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 638, + 375, + 654 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Relative pose estimation between two images plays a crucial role in many 3D vision tasks, including visual odometry and map-free visual relocalization [3], 6D object pose tracking [66,67], etc. These tasks can be categorized into two scenarios: camera-to-world and object-to-camera estimation. In the camera-to-world scenario, such as in map-free visual relocalization, we estimate the pose transformation from the reference image to the query image, predicting the change in camera extrinsic between the two views. In the object-to-camera scenario, taking object pose tracking for example, we aim to track the 6D pose of the target object in the videos by estimating its relative transformation in the camera coordinate system between two adjacent frames. While these two scenarios may", + "bbox": [ + 212, + 664, + 787, + 815 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* indicates corresponding author.", + "bbox": [ + 217, + 825, + 444, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/57982706270fbec43987e0e4adc8405d20f72effda6bd14de5d15e8045a2bd53.jpg", + "image_caption": [ + "(a) Reference" + ], + "image_footnote": [], + "bbox": [ + 256, + 143, + 416, + 237 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/dd0be0ec8e50da40d3a979a4bdf49612a77eab57bbc12296dd31468d2f3e82a4.jpg", + "image_caption": [ + "(b) Query" + ], + "image_footnote": [], + "bbox": [ + 418, + 143, + 576, + 237 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/fc65be99503e69315faa02432f7da2197878de4c04f541f27c4a8c55fdfee04e.jpg", + "image_caption": [ + "(c) Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 581, + 143, + 741, + 237 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1c61be74d7c10000943d1aa9116753914ccd4d0124232d6a2992b21506ce517a.jpg", + "image_caption": [ + "(d) Reference" + ], + "image_footnote": [], + "bbox": [ + 254, + 253, + 415, + 347 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a891af73eec98b5499f70449c883f9e7b126b56bce20048dd270102aa4bf64ca.jpg", + "image_caption": [ + "(e) Query" + ], + "image_footnote": [], + "bbox": [ + 418, + 253, + 576, + 347 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3205b13a541666eaaf8e29d45dd60f8c971be081b889f0990944615ca73331ce.jpg", + "image_caption": [ + "(f) Ground Truth", + "Fig. 1: Relative pose estimation by SRPose. Dots drawn in the figures visualize the cross-attention scores of sparse keypoints across the two images, with brighter dots representing higher attention. Camera-to-world: (a), (b), (c) visualize the epipolar lines, representing the connections between the nine corresponding points across two views. Higher attention is shown to the overlap of the scenes. Object-to-Camera: (d), (e), (f) show the relative 6D pose estimation in the query image with only one accessible object prompt in the reference image. Higher attention is shown to the target object. SRPose establishes implicit correspondences." + ], + "image_footnote": [], + "bbox": [ + 581, + 253, + 741, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "seem distinct, they both entail the prediction of the pose matrix from the two RGB images depicting the scene or object, which in turn represents the relative pose transformation from the first image to the second.", + "bbox": [ + 212, + 484, + 785, + 531 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Traditional pose estimation approaches typically involve detecting and matching keypoints or local features in the two images to establish point-to-point correspondences, then use a robust estimator to denoise the outliers and recover the relative pose from the essential matrix by solving a two-view geometry problem, i.e., the epipolar constraint equation [48]. These approaches produce accurate results regardless of image sizes and camera intrinsicics, thanks to the intrinsic calibration performed before solving the equation. Given the segmentation of the target object in the two images, these approaches can also yield acceptable predictions of the relative object pose transformation. However, on the downside, robust estimators can be computationally expensive. Indeed, the process of eliminating mismatched correspondences and solving the equations using robust estimators is significantly slower compared to detecting and matching keypoints or features, making it prohibitive in the real-time application. Moreover, real-time object pose tracking requires video object segmentation models to eliminate off-target keypoints [66, 67], which can also introduce additional overhead.", + "bbox": [ + 212, + 534, + 787, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The maturity of deep learning offers the advantage of directly regressing the relative pose transformation from two RGB inputs, significantly boosting the runtime performance. However, existing regression methods lack of generalizability or precision to images with varying sizes and camera intrinsics [3, 38, 53, 64, 74, 77]. They typically employ image encoders that only accept", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "R. Yin et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "fixed-size inputs during batched training. And unlike traditional approaches, most deep learning-based regressors lack awareness of the original camera intrinsic parameters, unable to utilize two-view geometry to achieve higher precision in pose estimation. Furthermore, state-of-the-art deep regressors only apply to camera-to-world pose estimation, despite the common solution shared by both the camera-to-world and object-to-camera tasks.", + "bbox": [ + 212, + 146, + 782, + 234 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address the aforementioned challenges, we propose SRPose: a Sparse keypoint-based framework for Relative Pose estimation. SRPose estimates the relative pose matrix based on two-view geometry by implicitly solving the epipolar constraint, as shown in Fig. 1. First, SRPose employs a point detector to extract sparse keypoints and associated descriptors to form the candidate set of implicit correspondences. Then, we use an intrinsic-calibration (IC) position encoder to modulate the keypoints before it computes the position embeddings, accounting for varying image sizes and camera intrinsics. The keypoint descriptors and the position embeddings are then fused into the prior knowledge-guided attention layers, which leverage the prior knowledge of keypoint similarities to establish implicit cross-view correspondences. With the mechanism of the promptable cross-attention in these layers, SRPose only requires an accessible object prompt in one of the two views to compute the relative 6D pose transformation of the target object, eliminating the need for elusive object segmentations.", + "bbox": [ + 212, + 237, + 784, + 448 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The main contribution of this paper can be summarized as follows:", + "bbox": [ + 238, + 449, + 720, + 463 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel framework using sparse keypoints, SRPose, for two-view relative pose estimation. To the best of our knowledge, it is the first attempt to directly regress relative poses from sparse keypoints for this task.", + "- We introduce an intrinsic-calibration position encoder to adapt SRPose to different image sizes and camera intrinsics.", + "- We enable object-to-camera estimation in addition to camera-to-world estimation by utilizing one accessible object prompt.", + "- Our SRPose greatly reduces estimation time by replacing robust estimators with direct regression, achieving state-of-the-art performance in accuracy and speed in camera-to-world and object-to-camera scenarios." + ], + "bbox": [ + 223, + 467, + 782, + 616 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related works", + "text_level": 1, + "bbox": [ + 215, + 636, + 390, + 651 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Relative Pose Estimation: In relative pose estimation, traditional approaches recover the rotation and translation from the essential matrix [27, 42]. By using the epipolar constraint [48], the essential matrix is estimated based on the cross-view correspondences established through the matching of sparse keypoints or dense features. Using this pipeline, sparse keypoint-based approaches typically start by detecting sparse keypoints and their associated descriptors, including classic hand-crafted detectors [5, 44, 54], and deep learning-based detectors [6, 15-17, 24, 35, 40, 45, 52, 63, 72, 75, 76]. Recent methods further employ deep learning-based matchers [12, 39, 55, 56, 70] to establish keypoint matches, while classical approaches rely on nearest neighbor search. In contrast to using sparse keypoints, dense matchers [11, 13, 18, 19, 36, 47, 59, 62, 65, 71, 73, 79] are detector-free and perform pixel-wise dense feature matching. Once the correspondences", + "bbox": [ + 212, + 657, + 784, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "SRPose: Two-view Relative Pose Estimation with Sparse Keypoints", + "bbox": [ + 281, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "are established through matching, a robust estimator such as RANSAC [23] is employed to estimate the essential matrix and recover the relative pose. Overall, while the traditional matcher-based approaches have shown promising results, they still face challenges with their time-consuming robust estimators. Furthermore, dense matchers, while providing exceptional accuracy, may suffer from slow performance due to their resource-intensive design.", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Relative Pose Regression: Deep learning regressors offer an alternative approach by directly predicting the relative pose using dedicated neural networks, without explicitly establishing correspondences. Two-view relative pose estimation focuses on frame-to-frame pose transformations, aiming at visual odometry and map-free relocalization [1,3,4,21,34,46,68]. Leveraging prior knowledge is an important aspect in two-view estimation. For instance, SparsePlanes [32], PlaneFormer [8] and NOPE-SAC [61] utilize geometry relations of the planar surfaces to enhance performance. Unfortunately, these approaches rely on image encoders designed for images of fixed sizes and intrinsics, lacking generalizability across different settings. Sparse-view relative pose estimation emphasizes global pose optimization or end-to-end SfM involving multiple views [38,57,64,74]. Although these sparse-view methods can be extended to two-view estimation, and may even predict unknown intrinsics, their adaptability to varying camera settings depends on multi-view information. In contrast, SRPose leverages sparse keypoints and two-view geometry to achieve higher precision and generalizability across different image sizes and camera intrinsics.", + "bbox": [ + 212, + 239, + 787, + 482 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Object Pose Estimation: While most of the aforementioned approaches focus on camera pose estimation in a fixed scene, some tasks, such as 6D object pose tracking, require relative object pose estimation as well. Most frameworks predict the relative object pose using the same pipeline of keypoint correspondences. BundleTrack [66] and BundleSDF [67] track the object poses in videos by matching keypoints between two frames, and conducting global pose graph optimization. POPE [22] combines advanced foundation models of segmentation [33] and matching [49] to enable zero-shot object pose estimation. Robust as these classic methods may be, they require dedicated object masks to focus on the objects, which are provided by other resource-consuming segmentation models. To get rid of object segmentation, OnePose [60], OnePose++ [29], and Gen6D [41] rely on multiple views of the object, and ZeroShot [25] establishes semantic correspondences to mute the background, and employs depth map to enable pose estimation. Nonetheless, these methods require additional high-quality information, and they restrict the querying view to only accommodate a single object. As a remedy, SRPose employs an accessible object prompt to focus on the target object, enabling effective depth-free and mask-free two-view relative object estimation with superior performance in the object-to-camera scenario.", + "bbox": [ + 212, + 484, + 787, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 214, + 761, + 331, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given two images $I_1, I_2$ , SRPose estimates the relative pose consisting of a rotation $R \\in \\mathcal{SO}(3)$ , and a translation $t \\in \\mathbb{R}^3$ based on the two-view geometry. Detailed problem definitions can be found in the supplementary materials. Traditional approaches use the epipolar constraint [48] to estimate the following", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "R. Yin et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/689444cd282daa516213b93b0c16f3427b7984e595ca9f7796569d6f21b6f37d.jpg", + "image_caption": [ + "Fig. 2: Overview. SRPose comprises four main components: 1) The sparse keypoint detector detects keypoints associated with descriptors separately sfrom the two images; 2) The intrinsic-calibration (IC) position encoder modulates the keypoints' coordinates with camera intrinsics, and encodes their position information; 3) Guided by the prior knowledge of keypoint similarities, along with the object prompt, the attention layers establish implicit cross-view correspondences; 4) The regressor estimates relative pose $R, t$ under the constraints of implicit correspondences." + ], + "image_footnote": [], + "bbox": [ + 228, + 146, + 767, + 265 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "essential matrix $E$ , from which $R, t$ are recovered:", + "bbox": [ + 215, + 372, + 576, + 387 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left(K _ {1} ^ {- 1} q _ {1}\\right) ^ {\\top} E K _ {2} ^ {- 1} q _ {2} = 0, \\quad \\forall \\left(q _ {1}, q _ {2}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 374, + 395, + 785, + 414 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nE = [ t ] _ {\\times} R, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 460, + 416, + 784, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $q_{1}, q_{2}$ are two arbitrary corresponding points in $I_{1}, I_{2}$ , and $K_{1}, K_{2}$ are the camera intrinsics of the two images. SRPose builds and solves the constraint equation Eq. (1) with neural networks implicitly. With SRPose denoted as $\\mathcal{SR}$ , the function of our framework is represented as:", + "bbox": [ + 214, + 441, + 784, + 502 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n[ R | t ] = \\mathcal {S R} (I _ {1}, I _ {2}). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 431, + 511, + 784, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 2 summarizes SRPose, involving four components: a sparse keypoint detector, an intrinsic-calibration position encoder, the promptable prior knowledge-guided attention layers, and an MLP regressor.", + "bbox": [ + 214, + 537, + 782, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Sparse Keypoint Detector", + "text_level": 1, + "bbox": [ + 215, + 588, + 477, + 603 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Instead of employing an image encoder to extract image features, we use a detector to extract sparse keypoints and associated descriptors from the two images. To this end, we can use one of the existing techniques developed over the years, which include classic methods such as SIFT [44], and deep learning-based methods such as SuperPoint [16], ALIKE [76], DISK [63], etc. Given an image $I$ , our detector, denoted as $\\mathcal{F}^{kd}$ , detects keypoints consisting of $N$ coordinates $P^k \\subset \\mathbb{R}^2$ in image space, and associated descriptors $D \\subset \\mathbb{R}^d$ , describing the local features in the image, where $d$ denotes the dimension of the keypoints descriptors in the network. To summarize, our detector is defined as the following function:", + "bbox": [ + 212, + 604, + 787, + 741 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{P ^ {k}, D \\right\\} = \\mathcal {F} ^ {k d} (I). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 429, + 750, + 785, + 767 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In practice, we resize all images to a standardized size to enable parallel batched detection, then rescale the keypoints' coordinates $P^k$ to their original positions. The sparse keypoint detector yields two sets of keypoints $P_1^k, P_2^k$ from $I_1, I_2$ , from which a set of correspondences $\\{(q_1, q_2)\\} \\subset P_1^k \\times P_2^k$ will be established implicitly later for solving the constraint Eq. (1).", + "bbox": [ + 214, + 775, + 787, + 852 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "SRPose: Two-view Relative Pose Estimation with Sparse Keypoints", + "bbox": [ + 281, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Intrinsic-Calibration (IC) Position Encoder", + "text_level": 1, + "bbox": [ + 215, + 146, + 619, + 161 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "According to Eq. (1), for the keypoints detected from an image $I$ , their coordinates should be first calibrated or normalized by the camera intrinsics $K$ :", + "bbox": [ + 212, + 162, + 785, + 193 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nP ^ {c} = K ^ {- 1} P ^ {k}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 452, + 204, + 785, + 220 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Enforcing the intrinsic calibration (IC) offers a crucial improvement of both the accuracy and robustness [48]. IC allows SRPose to adapt images captured by different cameras with varying intrinsic parameters. Previous regressors [3, 53] extract implicit image features with image encoders trained on fixed-size images, assuming the camera intrinsics of all inputs are the same by default. While the sparse keypoints in SRPose are modulated by intrinsic calibration, which provides the accurate positions, i.e. $P^c$ , of all keypoints in a unified camera coordinate system across different intrinsics.", + "bbox": [ + 212, + 229, + 782, + 349 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After the intrinsic calibration, the keypoints' coordinates are encoded to address the keypoints based on their positions. Due to the possible position distortion caused by relative position encoding [39, 59], we employ absolute position encoding that maintains the precise coordinate values, which is crucial to the two-view geometry problems. With a single-layer fully connected network, denoted as the following function $\\mathcal{F}^{pe}$ , the position encoder maps 2D calibrated coordinates to high-dimension position embeddings:", + "bbox": [ + 212, + 352, + 784, + 455 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nP ^ {e} = \\mathcal {F} ^ {p e} \\left(K ^ {- 1} P ^ {k}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 433, + 468, + 785, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The resulting encoded position embeddings, denoted as $P^e$ , are then incorporated with the keypoint descriptors and descriptor embeddings, denoted as $D \\oplus P^e$ , which provides calibrated and normalized position information for solving the constraint Eq. (1) using a set of attention layers.", + "bbox": [ + 214, + 494, + 782, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Promptable Prior-Knowledge-Guided Attention", + "text_level": 1, + "bbox": [ + 214, + 561, + 656, + 577 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given the position embeddings and keypoint descriptors, we use a multi-layer attention network to exploit semantic information and establish implicit correspondences. Our network consists of $M$ consecutive attention layers, with each layer comprising a self-attention module and a promptable prior knowledge-guided cross-attention module. We denote $X^{m} \\subset \\mathbb{R}^{d}$ as the keypoint descriptor embeddings being processed in the $m^{th}$ layer, in which $m \\in M$ , with $X^{m}$ being either $X_{1}^{m}$ or $X_{2}^{m}$ from the image $I_{1}$ or $I_{2}$ . Each layer takes $X^{m-1}$ as inputs, then computes the deeper embeddings $X^{m}$ . In summary, the entire multi-layer network inputs keypoint descriptors $D_{1}D_{2}$ , and outputs $X_{1}^{M}, X_{2}^{M}$ after $M$ layers, i.e., we define $D$ as the initial state of embeddings $X^{0}$ . Prior to each attention layer, $X_{1}^{m-1}, X_{2}^{m-1}$ are incorporated with the associated position embeddings $P_{1}^{e}, P_{2}^{e}$ . Following [28], residual connections are applied to every self- and cross-attention module. The architecture of the attention layers is outlined in Fig. 3.", + "bbox": [ + 212, + 579, + 784, + 775 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Self-Attention: The self-attention module captures the comprehensive semantic information of all keypoints within an entire image. In this module, $X^{m - 1} \\oplus P^e$ of both the images is first mapped to three vectors by a linear transformation $W_{s}^{m}$ , denoted as the query $q$ , the key $k$ , and value $v$ respectively, where", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "R. Yin et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bf8095184d857ed796812d2f567ac18f8cefee0318e694ba40bb390b968be5fc.jpg", + "image_caption": [ + "Fig. 3: Overview of the prior knowledge-guided attention layers. Each layer contains a self-attention and a cross-attention module. A similarity matrix $S$ of the keypoint descriptors is utilized as the prior knowledge to readjust the cross-attention scores, guiding more attention to cross-view keypoint pairs with implicit correspondences. The object prompt and the residual connections are omitted in the figure." + ], + "image_footnote": [], + "bbox": [ + 233, + 146, + 759, + 276 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$q, k, v \\subset \\mathbb{R}^d$ . Then $k$ and $q$ are used to compute the self-attention scores $A_s$ , which are then scaled by the dimension of descriptors and embeddings $d$ :", + "bbox": [ + 214, + 357, + 782, + 388 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n[ q | k | v ] = W _ {s} ^ {M} \\left(X ^ {m - 1} \\oplus P ^ {e}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 398, + 397, + 785, + 417 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nA _ {s} = \\frac {q k ^ {\\top}}{\\sqrt {d}}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 459, + 417, + 785, + 453 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "After being normalized by the Softmax function, the self-attention scores $A_{s}$ are then used to extract the highly attended information from $v$ :", + "bbox": [ + 214, + 460, + 782, + 491 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nX _ {s} ^ {m - 1} = \\operatorname {S o f t m a x} \\left(A _ {s}\\right) v. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 413, + 501, + 785, + 520 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the self-attention module, keypoint position embeddings and descriptor embeddings are computed together to summarize the semantic information within one entire image before deeper cross-view correspondences are established.", + "bbox": [ + 214, + 529, + 782, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Prior Knowledge-guided Cross-Attention: Guided by the prior knowledge of keypoint similarities, the cross-attention module exploits the mutual information across the two images to establish implicit correspondences. Similar to the self-attention, cross-attention maps the input embeddings $X_{s}^{m - 1}$ to two keys $k_{1},k_{2}$ and values $v_{1},v_{2}$ of the two images respectively through another linear transformation $W_{c}^{m}$ layer. Following [50], we compute the bidirectional attention scores across the two images from the resulting embeddings of the previous self-attention module, using the cross-view keys $k_{1},k_{2}$ :", + "bbox": [ + 214, + 578, + 784, + 700 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ k _ {1} \\mid v _ {1} \\right] = W _ {c} ^ {m} X _ {s, 1} ^ {m - 1}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 424, + 709, + 785, + 727 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ k _ {2} \\mid v _ {2} \\right] = W _ {c} ^ {m} X _ {s, 2} ^ {m - 1}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 729, + 785, + 750 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nA _ {2 1} ^ {\\top} = A _ {1 2} = \\frac {k _ {1} k _ {2} ^ {\\top}}{\\sqrt {d}}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 429, + 751, + 785, + 785 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The module recognizes the overlapping areas of the scene or object between the two views by identifying the mutuality between two sets of keypoint embeddings from the computed cross-attention scores. To facilitate the process, prior", + "bbox": [ + 214, + 794, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "SRPose: Two-view Relative Pose Estimation with Sparse Keypoints", + "bbox": [ + 281, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "knowledge is used to guide and enhance the attention on highly correlated keypoints, by adjusting the cross-attention scores with a keypoint similarity matrix $S \\in \\mathbb{R}^{N_1 \\times N_2}$ . Each entry $s_{ij}$ of the matrix is computed as the cosine similarity between the descriptors of every cross-view keypoint pair:", + "bbox": [ + 215, + 146, + 787, + 208 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\ns _ {i j} = \\left(\\frac {d _ {i} \\cdot d _ {j}}{\\| d _ {i} \\| \\| d _ {j} \\|} + 1\\right) / 2, \\quad \\{(d _ {i}, d _ {j}) \\} \\subseteq D _ {1} \\times D _ {2}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 330, + 218, + 785, + 251 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Keypoints with similar local features tend to possess similar descriptors, describing correlated semantic information of similar views. Thus, the cosine similarity captures the semantic similarities between two keypoints, and guides the framework to recognize the overlapping parts between the images.", + "bbox": [ + 214, + 260, + 782, + 319 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We normalize the matrix $S$ to [0,1] and element-wise multiply it with the cross-attention scores. Finally, the scores are used to extract and exchange crucial cross-view information from each others' values $v_{1}, v_{2}$ :", + "bbox": [ + 214, + 320, + 784, + 366 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nA _ {2 1} ^ {\\prime \\top} = A _ {1 2} ^ {\\prime} = A _ {1 2} \\cdot S, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 376, + 784, + 393 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nX _ {1} ^ {m - 1} = \\operatorname {S o f t m a x} \\left(A _ {1 2} ^ {\\prime}\\right) v _ {2}, \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 410, + 398, + 784, + 415 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nX _ {2} ^ {m - 1} = \\operatorname {S o f t m a x} \\left(A _ {2 1} ^ {\\prime}\\right) v _ {1}. \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 410, + 417, + 784, + 436 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "By capturing and incorporating relevant information from each image into the other, the implicit correspondences are established across the two images. As shown in Figs. 1a and 1b, with brighter dots representing higher cross-attention scores, SRPose establishes implicitly corresponding points densely in the overlapping areas of the two images, which is utilized to construct and solve an implicit version of the epipolar constraint Eq. (1).", + "bbox": [ + 214, + 446, + 782, + 537 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Object Prompt: In the object-to-camera estimation, we employ an accessible user-provided object prompt in only one image to identify the target object. In contrast, traditional matching approaches require high-quality object segmentations in both images. An object prompt $b$ is a bounding box consisting of the two coordinates of top-left and bottom-right points, which bounds the object $o$ in one of the images $I_1$ , making $I_1$ as the reference view. Once the keypoints are extracted from $I_1$ , the object prompt removes the keypoints outside of the bounding box in the reference view, eliminating most irrelevant keypoints. Thus, the descriptors and positions of keypoints within bounds, focusing on the target object $o$ , will be later processed in the Similarity Matrix and other following modules. With the guided cross-attention modules, SRPose learns to identify highly correlated keypoints lying on the same object $o$ from the entire query view $I_2$ , excluding irrelevant background keypoints, as shown in Figs. 1d and 1e. With the object prompt, SRPose searches the implicit corresponding points on the target object and computes the relative 6D object pose transformation in the same process as the camera-to-world scenario.", + "bbox": [ + 214, + 541, + 787, + 784 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Relative Pose Regressor", + "text_level": 1, + "bbox": [ + 215, + 791, + 460, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "After the embeddings $X_1^M, X_2^M$ are computed via the previous modules, a MultiLayer Perceptron (MLP) is employed to regress the relative pose as our last step.", + "bbox": [ + 214, + 808, + 785, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "R. Yin et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The outputs of the previous attention layers $X_1^M, X_2^M$ first undergo 2D average pooling, and then are concatenated together, resulting in the joint embedding $f \\in \\mathbb{R}^{2d}$ , which carries the deep features and implicit correspondences across the two images. Finally, embedding $f$ is input to the MLP, and the regressor solves the epipolar constraint Eq. (1) implicitly and regresses the relative pose, consisting of a rotation $R$ , and a translation $t$ . Following [78], we set the rotation output of the regressor to be a 6d vector $r \\in \\mathbb{R}^6$ , which is then transformed into a rotation matrix $R \\in \\mathbb{R}^{3 \\times 3}$ through a partial Gram-Schmidt process. This technique, as argued in [78], enables a continuous representation of rotations.", + "bbox": [ + 212, + 145, + 787, + 282 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.5 Loss Function", + "text_level": 1, + "bbox": [ + 214, + 290, + 377, + 304 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "SRPose is trained in a supervised manner. With the ground-truth rotation $R_{gt}$ , we compute the Huber Loss [31] of the error in the rotation angle, which minimizes the angular error between the predicted and ground-truth rotations:", + "bbox": [ + 212, + 308, + 785, + 354 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {R} \\left(I _ {1}, I _ {2}\\right) = \\mathcal {H} \\left(\\operatorname {a r c c o s} \\left(\\frac {\\operatorname {T r} \\left(R ^ {\\top} R _ {g t} - 1\\right)}{2}\\right)\\right), \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 341, + 364, + 785, + 398 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\mathcal{H}$ denotes the Huber Loss function. Similarly, given the ground-truth translation $t_{gt}$ , the error is computed in both normalized and unnormalized forms. To further enhance accuracy, we also incorporate the angular error of translation, resulting in the following three loss terms:", + "bbox": [ + 214, + 407, + 787, + 469 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {t} \\left(I _ {1}, I _ {2}\\right) = \\mathcal {H} \\left(t - t _ {g t}\\right), \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 419, + 482, + 784, + 498 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {t _ {n}} \\left(I _ {1}, I _ {2}\\right) = \\mathcal {H} \\left(\\frac {t}{\\| t \\|} - \\frac {t _ {g t}}{\\| t _ {g t} \\|}\\right), \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 502, + 784, + 534 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {t _ {a}} \\left(I _ {1}, I _ {2}\\right) = \\mathcal {H} \\left(\\arccos \\left(\\frac {t \\cdot t _ {g t}}{\\| t \\| \\| t _ {g t} \\|}\\right)\\right). \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 364, + 537, + 784, + 570 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The above three loss functions supervise both the scale and direction of translations. We compute the final loss by the weighted sum of the three functions:", + "bbox": [ + 214, + 579, + 785, + 609 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} \\left(I _ {1}, I _ {2}\\right) = \\mathcal {L} _ {R} + \\lambda_ {t} \\mathcal {L} _ {t} + \\lambda_ {t _ {n}} \\mathcal {L} _ {t _ {n}} + \\lambda_ {t _ {a}} \\mathcal {L} _ {t _ {a}}, \\tag {21}\n$$\n", + "text_format": "latex", + "bbox": [ + 354, + 625, + 784, + 642 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\lambda_t, \\lambda_{t_n}, \\lambda_{t_a}$ are scalars used to balance the weights of four different losses.", + "bbox": [ + 212, + 647, + 782, + 664 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.6 Implementation Details", + "text_level": 1, + "bbox": [ + 214, + 670, + 455, + 685 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "SRPose is first trained on ScanNet [14] and then fine-tuned on other datasets to achieve better performance. For all datasets, we resize the images to $640 \\times 480$ , and extract 1,024 sparse keypoints using SuperPoint [16] for training and 2,048 for evaluation. Other sparse keypoint detectors can also be used for this step. We utilize the pre-trained SuperPoint provided by the official source and freeze its weights in the training stage. The remaining learnable modules are initialized with random weights. In practice, SRPose consists of 6 guided attention layers, with 4 attention heads in each attention module. The dimension $d$ of keypoint descriptors and embeddings is set to 256. The rotation and translation are estimated using two separate 3-layer MLPs. All three balancing scalars $\\lambda_{t}, \\lambda_{t_{n}}, \\lambda_{t_{a}}$", + "bbox": [ + 212, + 688, + 787, + 843 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "SRPose: Two-view Relative Pose Estimation with Sparse Keypoints", + "bbox": [ + 279, + 114, + 732, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/8e762daccbb5361e8dc0327197506b15e988cca63fee8abb4793ebaf17ff9779.jpg", + "table_caption": [ + "Table 1: Relative Pose Estimation on Matterport [10]. Without depth, matcher-based approaches are incapable of scaled translation estimation, while SRPose achieves high accuracy in translation scales using only RGB inputs." + ], + "table_footnote": [], + "table_body": "
CategoryMethodRot. (°)Trans. (m)
Med.↓Avg.↓≤ 30° ↑Med.↓Avg.↓≤ 1m↑
SparseSuperGlue [55]3.8824.1777.8n/an/an/a
SGMNet [12]1.0621.3280.9n/an/an/a
LightGlue [39]1.3222.4580.0n/an/an/a
DenseLoFTR [59]0.7111.1190.5n/an/an/a
ASpanFormer [13]3.7331.4575.7n/an/an/a
DKM [18]0.4612.8989.0n/an/an/a
RegressorSparsePlanes [32]7.3322.7883.40.631.2566.6
PlaneFormers [2]5.9622.2083.80.661.1966.8
8point [53]8.0119.1385.40.641.0167.4
NOPE-SAC [61]2.7714.3789.00.520.9473.2
SRPose2.6511.1291.60.270.6183.7
", + "bbox": [ + 256, + 188, + 741, + 348 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "are set to 1. On ScanNet, our framework is trained using AdamW optimizer [43] and follows 1cycle learning rate policy [58] with a maximum learning rate of $1 \\times 10^{-4}$ for 500 epochs. We fix the batch size to 32 for training. The pretraining requires 120 hours utilizing 8 RTX 3090 GPUs. More implementation details can be found in the supplementary materials.", + "bbox": [ + 212, + 366, + 785, + 444 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 454, + 375, + 472 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We evaluate the performance of SRPose for two-view relative pose estimation in both the camera-to-world and object-to-camera scenarios. Additionally, we assess its effectiveness in the application of map-free relocalization. Finally, we conduct an ablation study to analyze the roles of the components we proposed.", + "bbox": [ + 212, + 481, + 785, + 544 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Camera-to-World Pose Estimation", + "text_level": 1, + "bbox": [ + 215, + 556, + 545, + 571 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Setup: In the camera-to-world scenario, we evaluate our framework's performance on Matterport [10] and ScanNet [14]. Following [32], for Matterport, we report the median and mean of the translation and rotation error, and the translation/rotation accuracy at the thresholds of $1\\mathrm{m} / 30^{\\circ}$ , respectively. For ScanNet, we adopt the metrics from [55] to compute the AUC of the pose error at $5^{\\circ}$ , $10^{\\circ}$ , and $20^{\\circ}$ . The pose error here is the maximum angular error in rotation and translation. We also report an analysis of the time consumption on the ScanNet-1500 test set. For all matcher-based approaches, we use the implementation of RANSAC [23] from OpenCV [7] to recover relative poses. All methods are compared on a device with RTX 4090, i9-13900K, and 128GB memory.", + "bbox": [ + 212, + 580, + 785, + 732 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Baselines: We choose three categories of approaches as our baselines, including sparse keypoint-based matchers, dense matchers, and deep learning-based relative pose regressors. Results of all the baselines are reported according to the official papers or implementations. All sparse matchers and SRPose employ SuperPoint [16] as their sparse keypoint detector. All regressors are trained on the specific dataset, i.e. Matterport or ScanNet, before evaluation. LightGlue [39] is trained on MegaDepth [37], while all other matchers are trained on ScanNet.", + "bbox": [ + 212, + 744, + 785, + 853 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "R. Yin et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/15e1b6b67f7edc068549a9a99ddd86a987d89ad18fd728a478b3f6b2e2534f08.jpg", + "table_caption": [ + "Table 2: Relative Pose Estimation on ScanNet [14]. SRPose outperforms state-of-the-art regressors and sparse matchers in terms of various metrics and reduces computational time greatly due to the riddance of robust estimators." + ], + "table_footnote": [], + "table_body": "
CategoryMethodPose estimationAUCTime (ms)
@5°@10°@20°MatchTotal
SparseSuperGlue [55]16.233.851.8102.5361.9
SGMNet [12]15.432.148.3111.7439.7
LightGlue [39]16.433.650.231.5269.7
DenseLoFTR [59]22.040.857.644.4292.4
ASpanFormer [13]25.646.063.359.6370.9
DKM [18]29.450.768.3235.0491.6
RegressorMap-free [3]2.7011.529.0n/a12.7
SRPose13.334.356.8n/a28.5
", + "bbox": [ + 250, + 188, + 748, + 319 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/ab9c7615a46a7b7e4a9ff956854a87d6ac510e75ae64c2395e1639f864738284.jpg", + "image_caption": [ + "Fig. 4: Time consumption comparison on ScanNet [14]. Regressors, including SRPose, achieve much higher computational efficacy than all matchers." + ], + "image_footnote": [], + "bbox": [ + 243, + 321, + 769, + 463 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Results: Table 1 demonstrates that SRPose greatly outperforms the existing deep learning-based regressors across all metrics on Matterport. Our framework achieves the most accurate scaled translation estimation, which is unattainable for matcher-based approaches lacking depth information. Table 2 shows SRPose achieves competitive or superior performance compared to all sparse matchers and the regressor and on ScanNet. Although dense matchers outperform our framework in accuracy, SRPose offers a significant advantage in computational speed compared to all the matcher-based approaches by replacing robust estimators with direct regression. As shown in Fig. 4, matcher-based approaches spend a substantial proportion of time on recovering pose using robust estimators after matching. In contrast, SRPose consumes even less time than the matching stage of all matchers, saving at least 200ms in the total time.", + "bbox": [ + 212, + 508, + 787, + 691 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Object-to-Camera Pose Estimation", + "text_level": 1, + "bbox": [ + 214, + 700, + 552, + 715 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Setup: In the object-to-camera scenario, we evaluate SRPose on HO3D [26]. HO3D contains videos of hand-holding objects of 21 categories from the YCB dataset [9]. The depth captured by the stereo depth camera and the object segmentation of each frame in the videos are also provided. We randomly select image pairs with small object rotation transformations for training and evaluation. More details can be found in the supplementary materials. Following [30, 69], we adopt the widely-used Average Distance (ADD, ADD-S) metric for evaluation. To measure the smaller scales in object pose transformation, we set the", + "bbox": [ + 212, + 718, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "SRPose: Two-view Relative Pose Estimation with Sparse Keypoints", + "bbox": [ + 279, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/54d1452d1ed87108a884e295106dc986a82fc85917e002d8e532946aec3df694.jpg", + "table_caption": [ + "Table 3: Relative object pose estimation on HO3D [26]. SRPose achieves state-of-the-art performance in object-to-camera estimation without additional depth information, enabling scale pose estimation from only RGB inputs." + ], + "table_footnote": [], + "table_body": "
MethodRot. (°)Trans. (cm)ADD↑ADD-S↑
Med.↓Avg.↓≤ 30°↑Med.↓Avg.↓≤ 10cm↑
SGMNet [12]24.129.062.713.917.536.816.831.3
LightGlue [39]24.428.761.713.516.338.217.231.2
LoFTR [59]29.257.950.913.916.435.616.730.9
DKM [18]23.325.664.813.615.439.817.833.3
SRPose8.911.495.45.98.073.936.456.8
", + "bbox": [ + 243, + 189, + 754, + 284 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "threshold of metrics to $10\\mathrm{cm}$ , which differs from the 1m threshold used in the camera-to-world scenario. Specifically, we calculate the AUC using both ADD and ADD-S with the threshold set to $10\\mathrm{cm}$ . We also report the median, mean error of translation/rotation, and their accuracy at $10\\mathrm{cm} / 30^{\\circ}$ .", + "bbox": [ + 212, + 295, + 782, + 357 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Baselines: We only choose matchers listed in Sec. 4.1 as our baselines, for no deep learning-based regressors are available in this two-view relative object pose estimation task. We utilize the stereo depth offered in the dataset to enable matcher-based approaches to estimate scaled relative poses via Orthogonal Procrustes [20], while SRPose only requires RGB images as inputs. The segmentations in the dataset are adopted for matcher-based approaches to identify the target object in both images, while SRPose utilizes the bounding box of the segmentation in the first image as the object prompt.", + "bbox": [ + 212, + 363, + 784, + 486 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Results: As shown in Table 3, SRPose outperforms state-of-the-art matcher-based approaches significantly. SRPose exhibits superior accuracy in relative 6D object pose estimation between two images using the object prompts in the reference view. The low quality of stereo depth leads to limited performance in scaled pose estimation of the matcher-based approaches, while SRPose can learn scaled information to achieve lower estimation errors.", + "bbox": [ + 212, + 493, + 784, + 583 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Map-Free Visual Relocalization", + "text_level": 1, + "bbox": [ + 215, + 590, + 521, + 606 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We evaluate the application of SRPose on Niantic [3], a dataset for the map-free visual relocalization task. Following [3], we adopt the Virtual Correspondence Reprojection Error (VCRE) and the median error of translation and rotation as the metrics. Specifically, we compute the precision of VCRE at 90 pixels, and the precision of relocalization at $0.25\\mathrm{m}$ and $5^{\\circ}$ .", + "bbox": [ + 212, + 608, + 784, + 684 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Baselines: We still select the three categories of approaches as our baselines. For **matching-based approaches**, we estimate the pose by recovering from the essential matrix. Then we use the depth estimated by DPT [51], a depth prediction model, to provide the matchers with the scale information.", + "bbox": [ + 212, + 691, + 784, + 753 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Results: Table 4 shows SRPose outperforms other sparse matchers and the regression method in map-free visual relocalization. Although the state-of-the-art dense matcher RoMa achieves higher precision in relocalization using scales estimated by an additional depth prediction model, SRPose still outperforms in VCRE. Our framework exhibits competitive performance compared to state-of-the-art approaches in the map-free visual relocalization task.", + "bbox": [ + 212, + 761, + 785, + 852 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "R. Yin et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/687710288f4779109cf2a99d5ad755d24cab511e0473bfa8d1f3a956e6b5f00a.jpg", + "table_caption": [ + "Table 4: Map-free visual relocalization results on Niantic [3]. SRPose achieves competitive performance compared to state-of-the-art approaches without relying on an additional depth prediction model to assist with scaled estimation." + ], + "table_footnote": [], + "table_body": "
CategoryMethodVCRE \nPrec.↑ / Med.↓Med. (m / °) \nTrans.↓ / Rot.↓Prec.↑
SparseSIFT [44]25.0 / 222.82.93 / 61.410.3
+DPT [51]SuperGlue [55]36.1 / 160.31.88 / 25.416.8
DenseLoFTR [59]34.7 / 167.61.98 / 30.515.4
+DPT [51]RoMa [19]45.6 / 128.81.23 / 11.122.8
RegressorMap-free [3]40.2 / 147.11.68 / 22.96.0
SRPose46.4 / 127.71.37 / 17.216.9
", + "bbox": [ + 282, + 181, + 712, + 292 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/0c9b21b3f8fe1b76bd0092991fa89c502354f15e0b098ba03348acdcce0ab3dc.jpg", + "table_caption": [ + "Table 5: Ablation study on Matterport [10]. The prior knowledge-guided attention layers assist in establishing implicit cross-view correspondences." + ], + "table_footnote": [], + "table_body": "
MethodTrans. (m)Rot. (°)
Avg.↓≤ 1m↑Avg.↓≤ 30°↑
8point [53]1.0167.419.1385.4
SRPose (full)0.8474.214.3288.9
1) no prior knowledge guidance0.9769.616.9186.8
2) replace all cross-attn. with self-attn.1.9026.148.042.0
3) no position encoding1.1761.021.3981.3
", + "bbox": [ + 282, + 325, + 718, + 420 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Insights", + "text_level": 1, + "bbox": [ + 215, + 430, + 328, + 445 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Ablation study: We evaluate the effectiveness of the different components in SRPose by Matterport [10] and MegaDepth [37] datasets. Specifically, we train four different variants initialized with random weights on Matterport by removing the three different designs we proposed to validate their effectiveness. As shown in Table 5: 1) Removing the prior knowledge", + "bbox": [ + 212, + 446, + 486, + 598 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/6ba9e38bf474ea0eeb9f5f16239374e848bff2cb4d51d9b14fc867b5897104f5.jpg", + "table_caption": [ + "Table 6: Ablation study on MegaDepth [37]. Intrinsic-calibration position encoder enables adaption on varying image sizes and camera intrinsics." + ], + "table_footnote": [], + "table_body": "
MethodPose est.AUC
@5°@10°@20°
Map-free [3]2.69.322.9
SRPose (full)16.636.058.0
1) no intrinsic calibration1.58.524.3
2) no position encoding1.06.218.9
", + "bbox": [ + 501, + 505, + 781, + 588 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "edge guidance of keypoint similarities in the cross-attention modules results in a moderate decrease in accuracy, as it assists SRPose in seeking the cross-view correspondences. 2) Replacing all cross-attention modules with self-attention modules leads to a significant accuracy drop, as the capability to establish implicit cross-view correspondences through cross-attention is eliminated. 3) Removing the entire position encoder leads to a moderately lower accuracy due to the absence of keypoint position information. The experiment highlights the crucial roles of the three designs of SRPose in accurate relative pose estimation.", + "bbox": [ + 212, + 598, + 787, + 718 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Moreover, we train three different variants initialized with random weights on MegaDepth, a dataset that consists of images with different sizes and camera intrinsics, to further validate the effectiveness of the intrinsic-calibration position encoder, with results shown in Table 6. 1) Without enforcing intrinsic calibration on keypoint coordinates, SRPose only achieves comparable performance to Map-free, both inadaptable to varying image sizes and camera intrinsics in MegaDepth. 2) Removing the entire IC Position Encoder results in a significant drop in accuracy as expected owing to the elimination of critical position infor", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "SRPose: Two-view Relative Pose Estimation with Sparse Keypoints", + "bbox": [ + 279, + 113, + 732, + 130 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "mation. By comparing the full framework with the variants lacking IC and the entire position encoder, we demonstrate the essential roles of the two designs in adapting different image sizes and camera intrinsics.", + "bbox": [ + 212, + 146, + 787, + 191 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Visualizing attention: In Fig. 5, we visualize the cross-attention scores $A_{12}^{\\prime}$ and compare SRPose without prior knowledge guidance to its full designs. We also draw the connections of highly-attended cross-view keypoint pairs, which represent the implicit correspondences established by SRPose.", + "bbox": [ + 212, + 196, + 787, + 257 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/8b3dd6dce6dd434d7266560f41d224db217cd6ec6f8aa3d4a57462c2037643ad.jpg", + "image_caption": [ + "Fig.5: Visualization of implicit cross-view correspondences. Dots and lines drawn with brighter colors represent higher cross-view attention scores. The prior knowledge guidance enhances attention to the overlapping areas, removing the irrelevant cross-view connections, and assisting in establishing implicit correspondences." + ], + "image_footnote": [], + "bbox": [ + 232, + 266, + 764, + 571 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 638, + 359, + 652 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This paper presents SRPose, a novel deep learning-based regressor utilizing sparse keypoints for two-view relative pose estimation in camera-to-world and object-to-camera scenarios. As our key innovations, SRPose extracts keypoints with sparse keypoint detectors and employs an intrinsic-calibration position encoder to enable adaptability to different image sizes and camera intrinsics. Further, our proposed promptable prior knowledge-guided attention layers establish implicit correspondences to estimate the relative pose under the epipolar constraint in both scenarios. By directly regressing the rotation and translation, SRPose achieves a significant decrease in computational time, with a minimum time reduction of 200ms. Extensive experiments demonstrate that SRPose achieves state-of-the-art performance in terms of accuracy and speed in the two scenarios, as well as in the map-free visual relocalization task.", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "R. Yin et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported by Shanghai Municipal Science and Technology Major Project (2021SHZDZX0102) and the Fundamental Research Funds for the Central Universities.", + "bbox": [ + 212, + 174, + 785, + 219 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 244, + 321, + 262 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Abouelnaga, Y., Bui, M., Ilic, S.: Distillpose: lightweight camera localization using auxiliary learning. In: IROS (2021)", + "2. Agarwala, S., Jin, L., Rockwell, C., Fouhey, D.F.: Planeformers: From sparse view planes to 3d reconstruction. In: ECCV (2022)", + "3. Arnold, E., Wynn, J., Vicente, S., Garcia-Hernando, G., Monszpart, A., Prisacariu, V., Turmukhambetov, D., Brachmann, E.: Map-free visual relocalization: Metric pose relative to a single image. In: ECCV (2022)", + "4. Balntas, V., Li, S., Prisacariu, V.: Relocnet: Continuous metric learning relocalisation using neural nets. In: ECCV (2018)", + "5. Bay, H., Tuytelaars, T., Van Gool, L.: Surf: Speeded up robust features. In: ECCV (2006)", + "6. Bhowmik, A., Gumhold, S., Rother, C., Brachmann, E.: Reinforced feature points: Optimizing feature detection and description for a high-level task. In: CVPR (2020)", + "7. Bradski, G.: The OpenCV Library. Dr. Dobb's Journal of Software Tools (2000)", + "8. Cai, R., Hariharan, B., Snavely, N., Averbuch-Elor, H.: Extreme rotation estimation using dense correlation volumes. In: CVPR (2021)", + "9. Calli, B., Singh, A., Walsman, A., Srinivasa, S., Abbeel, P., Dollar, A.M.: The ycb object and model set: Towards common benchmarks for manipulation research. In: ICAR (2015)", + "10. Chang, A., Dai, A., Funkhouser, T., Halber, M., Niebner, M., Savva, M., Song, S., Zeng, A., Zhang, Y.: Matterport3d: Learning from rgb-d data in indoor environments. In: 3DV (2017)", + "11. Chang, J., Yu, J., Zhang, T.: Structured epipolar matcher for local feature matching. In: CVPR (2023)", + "12. Chen, H., Luo, Z., Zhang, J., Zhou, L., Bai, X., Hu, Z., Tai, C.L., Quan, L.: Learning to match features with seeded graph matching network. In: ICCV (2021)", + "13. Chen, H., Luo, Z., Zhou, L., Tian, Y., Zhen, M., Fang, T., McKinnon, D., Tsin, Y., Quan, L.: Aspanformer: Detector-free image matching with adaptive span transformer. In: ECCV (2022)", + "14. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: CVPR (2017)", + "15. DeTone, D., Malisiewicz, T., Rabinovich, A.: Toward geometric deep slam. arXiv preprint arXiv:1707.07410 (2017)", + "16. DeTone, D., Malisiewicz, T., Rabinovich, A.: Superpoint: Self-supervised interest point detection and description. In: CVPR (2018)", + "17. Dusmanu, M., Rocco, I., Pajdla, T., Pollefeys, M., Sivic, J., Torii, A., Sattler, T.: D2-net: A trainable cnn for joint description and detection of local features. In: CVPR (2019)", + "18. Edstedt, J., Athanasiadis, I., Wadenback, M., Felsberg, M.: Dkm: Dense kernelized feature matching for geometry estimation. In: CVPR (2023)" + ], + "bbox": [ + 217, + 277, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "SRPose: Two-view Relative Pose Estimation with Sparse Keypoints", + "bbox": [ + 281, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "19. Edstedt, J., Sun, Q., Bokman, G., Wadenback, M., Felsberg, M.: RoMa: Robust Dense Feature Matching. arXiv preprint arXiv:2305.15404 (2023)", + "20. Eggert, D.W., Lorusso, A., Fisher, R.B.: Estimating 3-d rigid body transformations: a comparison of four major algorithms. Machine vision and applications (1997)", + "21. En, S., Lechery, A., Jurie, F.: Rpnet: An end-to-end network for relative camera pose estimation. In: ECCV (2018)", + "22. Fan, Z., Pan, P., Wang, P., Jiang, Y., Xu, D., Jiang, H., Wang, Z.: Pope: 6-dof promptable pose estimation of any object, in any scene, with one reference. arXiv preprint arXiv:2305.15727 (2023)", + "23. Fischler, M.A., Bolles, R.C.: Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. COMMUN ACM (1981)", + "24. Gleize, P., Wang, W., Feiszli, M.: Silk: Simple learned keypoints. In: ICCV (2023)", + "25. Goodwin, W., Vaze, S., Havoutis, I., Posner, I.: Zero-shot category-level object pose estimation. In: ECCV (2022)", + "26. Hampali, S., Rad, M., Oberweger, M., Lepetit, V.: Honnotate: A method for 3d annotation of hand and object poses. In: CVPR (2020)", + "27. Hartley, R.I.: In defense of the eight-point algorithm. TPAMI (1997)", + "28. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016)", + "29. He, X., Sun, J., Wang, Y., Huang, D., Bao, H., Zhou, X.: Onepose++: Keypoint-free one-shot object pose estimation without cad models. NeurIPS (2022)", + "30. Hinterstoisser, S., Lepetit, V., Ilic, S., Holzer, S., Bradski, G., Konolige, K., Navab, N.: Model based training, detection and pose estimation of texture-less 3d objects in heavily cluttered scenes. In: ACCV (2013)", + "31. Huber, P.J.: Robust estimation of a location parameter. The Annals of Mathematical Statistics (1964)", + "32. Jin, L., Qian, S., Owens, A., Fouhey, D.F.: Planar surface reconstruction from sparse views. In: ICCV (2021)", + "33. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)", + "34. Laskar, Z., Melekhov, I., Kalia, S., Kannala, J.: Camera relocalization by computing pairwise relative poses using convolutional neural network. In: ICCVW (2017)", + "35. Li, K., Wang, L., Liu, L., Ran, Q., Xu, K., Guo, Y.: Decoupling makes weakly supervised local feature better. In: CVPR (2022)", + "36. Li, X., Han, K., Li, S., Prisacariu, V.: Dual-resolution correspondence networks. NeurIPS (2020)", + "37. Li, Z., Snavely, N.: Megadepth: Learning single-view depth prediction from internet photos. In: CVPR (2018)", + "38. Lin, A., Zhang, J.Y., Ramanan, D., Tulsiani, S.: Relpose++: Recovering 6d poses from sparse-view observations. In: 2024 International Conference on 3D Vision (3DV). pp. 106-115. IEEE (2024)", + "39. Lindenberger, P., Sarlin, P.E., Pollefeys, M.: LightGlue: Local Feature Matching at Light Speed. In: ICCV (2023)", + "40. Liu, Y., Shen, Z., Lin, Z., Peng, S., Bao, H., Zhou, X.: Gift: Learning transformation-invariant dense visual descriptors via group cnns. NeurIPS (2019)", + "41. Liu, Y., Wen, Y., Peng, S., Lin, C., Long, X., Komura, T., Wang, W.: Gen6d: Generalizable model-free 6-dof object pose estimation from rgb images. In: ECCV (2022)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "R. Yin et al.", + "bbox": [ + 271, + 114, + 356, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "42. Longuet-Higgins, H.C.: A computer algorithm for reconstructing a scene from two projections. Nature (1981)", + "43. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: ICLR (2018)", + "44. Lowe, D.G.: Distinctive image features from scale-invariant keypoints. IJCV (2004)", + "45. Luo, Z., Zhou, L., Bai, X., Chen, H., Zhang, J., Yao, Y., Li, S., Fang, T., Quan, L.: Aslfeat: Learning local features of accurate shape and localization. In: CVPR (2020)", + "46. Melekhov, I., Ylioinas, J., Kannala, J., Rahtu, E.: Relative camera pose estimation using convolutional neural networks. In: ACIVS (2017)", + "47. Ni, J., Li, Y., Huang, Z., Li, H., Bao, H., Cui, Z., Zhang, G.: Pats: Patch area transportation with subdivision for local feature matching. In: CVPR (2023)", + "48. Nister, D.: An efficient solution to the five-point relative pose problem. TPAMI (2004)", + "49. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023)", + "50. Phil, W.: Bidirectional cross attention. https://github.com/lucidrains/bidirectional-cross-attention (2022)", + "51. Ranftl, R., Lasinger, K., Hafner, D., Schindler, K., Koltun, V.: Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. TPAMI (2020)", + "52. Revaud, J., Weinzaepfel, P., de Souza, C.R., Humenberger, M.: R2D2: repeatable and reliable detector and descriptor. In: NeurIPS (2019)", + "53. Rockwell, C., Johnson, J., Fouhey, D.F.: The 8-point algorithm as an inductive bias for relative pose prediction by vits. In: 3DV (2022)", + "54. Rublee, E., Rabaud, V., Konolige, K., Bradski, G.: Orb: An efficient alternative to sift or surf. In: ICCV (2011)", + "55. Sarlin, P.E., DeTone, D., Malisiewicz, T., Rabinovich, A.: SuperGlue: Learning feature matching with graph neural networks. In: CVPR (2020)", + "56. Shi, Y., Cai, J.X., Shavit, Y., Mu, T.J., Feng, W., Zhang, K.: Clustergnn: Cluster-based coarse-to-fine graph neural network for efficient feature matching. In: CVPR (2022)", + "57. Sinha, S., Zhang, J.Y., Tagliasacchi, A., Gilitschenski, I., Lindell, D.B.: Sparse-pose: Sparse-view camera pose regression and refinement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21349–21359 (2023)", + "58. Smith, L.N., Topin, N.: Super-convergence: Very fast training of neural networks using large learning rates. In: Artificial intelligence and machine learning for multidomain operations applications (2019)", + "59. Sun, J., Shen, Z., Wang, Y., Bao, H., Zhou, X.: LoFTR: Detector-free local feature matching with transformers. CVPR (2021)", + "60. Sun, J., Wang, Z., Zhang, S., He, X., Zhao, H., Zhang, G., Zhou, X.: One-positron object pose estimation without cad models. In: CVPR (2022)", + "61. Tan, B., Xue, N., Wu, T., Xia, G.S.: Nope-sac: Neural one-plane ransac for sparse-view planar 3d reconstruction. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)", + "62. Tang, S., Zhang, J., Zhu, S., Tan, P.: Quadtree attention for vision transformers. ICLR (2022)", + "63. Tyszzkiewicz, M., Fua, P., Trulls, E.: Disk: Learning local features with policy gradient. NeurIPS (2020)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "SRPose: Two-view Relative Pose Estimation with Sparse Keypoints", + "bbox": [ + 281, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "64. Wang, J., Rupprecht, C., Novotny, D.: Posediffusion: Solving pose estimation via diffusion-aided bundle adjustment. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9773-9783 (2023)", + "65. Wang, Q., Zhang, J., Yang, K., Peng, K., Stiefelhagen, R.: Matchformer: Interleaving attention in transformers for feature matching. In: ACCV (2022)", + "66. Wen, B., Bekris, K.: Bundletrack: 6d pose tracking for novel objects without instance or category-level 3d models. In: IROS (2021)", + "67. Wen, B., Tremblay, J., Blukis, V., Tyree, S., Müller, T., Evans, A., Fox, D., Kautz, J., Birchfield, S.: Bundlesdf: Neural 6-dof tracking and 3d reconstruction of unknown objects. In: CVPR (2023)", + "68. Winkelbauer, D., Denninger, M., Triebel, R.: Learning to localize in new environments from synthetic training data. In: ICRA (2021)", + "69. Xiang, Y., Schmidt, T., Narayanan, V., Fox, D.: PoseCNN: A convolutional neural network for 6d object pose estimation in cluttered scenes. Robotics: Science and Systems XIV (2018)", + "70. Xue, F., Budvytis, I., Cipolla, R.: Imp: Iterative matching and pose estimation with adaptive pooling. In: CVPR (2023)", + "71. Xue, F., Budvytis, I., Cipolla, R.: Sfd2: Semantic-guided feature detection and description. In: CVPR (2023)", + "72. Yi, K.M., Trulls, E., Lepetit, V., Fua, P.: Lift: Learned invariant feature transform. In: ECCV (2016)", + "73. Yu, J., Chang, J., He, J., Zhang, T., Yu, J., Wu, F.: Adaptive spot-guided transformer for consistent local feature matching. In: CVPR (2023)", + "74. Zhang, J.Y., Ramanan, D., Tulsiani, S.: Relpose: Predicting probabilistic relative rotation for single objects in the wild. In: European Conference on Computer Vision. pp. 592-611. Springer (2022)", + "75. Zhao, X., Wu, X., Chen, W., Chen, P.C.Y., Xu, Q., Li, Z.: Aliked: A lighter keypoint and descriptor extraction network via deformable transformation. TIM (2023)", + "76. Zhao, X., Wu, X., Miao, J., Chen, W., Chen, P.C., Li, Z.: Alike: Accurate and lightweight keypoint detection and descriptor extraction. TMM (2022)", + "77. Zhou, Q., Sattler, T., Pollefeys, M., Leal-Taixe, L.: To learn or not to learn: Visual localization from essential matrices. In: ICRA (2020)", + "78. Zhou, Y., Barnes, C., Lu, J., Yang, J., Li, H.: On the continuity of rotation representations in neural networks. In: CVPR (2019)", + "79. Zhu, S., Liu, X.: Pmatch: Paired masked image modeling for dense geometric matching. In: CVPR (2023)" + ], + "bbox": [ + 215, + 147, + 784, + 646 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "R. Yin et al.", + "bbox": [ + 271, + 114, + 356, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/SRPose_ Two-view Relative Pose Estimation with Sparse Keypoints/e429b1d6-1778-4297-b226-264e6bf36452_model.json b/2024/SRPose_ Two-view Relative Pose Estimation with Sparse Keypoints/e429b1d6-1778-4297-b226-264e6bf36452_model.json new file mode 100644 index 0000000000000000000000000000000000000000..cd25d7cd7b2b1452daaf75ca33f939ad16ab497d --- /dev/null +++ b/2024/SRPose_ Two-view Relative Pose Estimation with Sparse Keypoints/e429b1d6-1778-4297-b226-264e6bf36452_model.json @@ -0,0 +1,2876 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.25, + 0.141, + 0.753, + 0.187 + ], + "angle": 0, + "content": "SRPose: Two-view Relative Pose Estimation with Sparse Keypoints" + }, + { + "type": "text", + "bbox": [ + 0.304, + 0.212, + 0.699, + 0.252 + ], + "angle": 0, + "content": "Rui Yin\\(^{1}\\), Yulun Zhang\\(^{2}\\), Zherong Pan\\(^{3}\\), Jianjun Zhu\\(^{1}\\), Cheng Wang\\(^{1}\\), and Biao Jia\\(^{1^{\\star}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.261, + 0.768, + 0.306 + ], + "angle": 0, + "content": "\\(^{1}\\) Hanglok-Tech, China \n\\(^{2}\\) MoE Key Lab of Artificial Intelligence, Shanghai Jiao Tong University, China \n\\(^{3}\\) LightSpeed Studios, Tencent America" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.343, + 0.744, + 0.581 + ], + "angle": 0, + "content": "Abstract. Two-view pose estimation is essential for map-free visual relocalization and object pose tracking tasks. However, traditional matching methods suffer from time-consuming robust estimators, while deep learning-based pose regressors only cater to camera-to-world pose estimation, lacking generalizability to different image sizes and camera intrinsics. In this paper, we propose SRPose, a sparse keypoint-based framework for two-view relative pose estimation in camera-to-world and object-to-camera scenarios. SRPose consists of a sparse keypoint detector, an intrinsic-calibration position encoder, and promptable prior knowledge-guided attention layers. Given two RGB images of a fixed scene or a moving object, SRPose estimates the relative camera or 6D object pose transformation. Extensive experiments demonstrate that SRPose achieves competitive or superior performance compared to state-of-the-art methods in terms of accuracy and speed, showing generalizability to both scenarios. It is robust to different image sizes and camera intrinsics, and can be deployed with low computing resources. Project page: https://frickyinn.github.io/srpose." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.592, + 0.719, + 0.608 + ], + "angle": 0, + "content": "Keywords: Relative Pose Estimation \\(\\cdot\\) 6D Object Pose Estimation" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.639, + 0.377, + 0.655 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.665, + 0.788, + 0.816 + ], + "angle": 0, + "content": "Relative pose estimation between two images plays a crucial role in many 3D vision tasks, including visual odometry and map-free visual relocalization [3], 6D object pose tracking [66,67], etc. These tasks can be categorized into two scenarios: camera-to-world and object-to-camera estimation. In the camera-to-world scenario, such as in map-free visual relocalization, we estimate the pose transformation from the reference image to the query image, predicting the change in camera extrinsic between the two views. In the object-to-camera scenario, taking object pose tracking for example, we aim to track the 6D pose of the target object in the videos by estimating its relative transformation in the camera coordinate system between two adjacent frames. While these two scenarios may" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.826, + 0.446, + 0.841 + ], + "angle": 0, + "content": "* indicates corresponding author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "R. Yin et al." + }, + { + "type": "image", + "bbox": [ + 0.257, + 0.145, + 0.417, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.297, + 0.24, + 0.377, + 0.251 + ], + "angle": 0, + "content": "(a) Reference" + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.144, + 0.578, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.469, + 0.24, + 0.53, + 0.251 + ], + "angle": 0, + "content": "(b) Query" + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.144, + 0.742, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.611, + 0.24, + 0.714, + 0.251 + ], + "angle": 0, + "content": "(c) Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.256, + 0.255, + 0.416, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.296, + 0.35, + 0.378, + 0.361 + ], + "angle": 0, + "content": "(d) Reference" + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.255, + 0.578, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.47, + 0.35, + 0.53, + 0.361 + ], + "angle": 0, + "content": "(e) Query" + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.255, + 0.742, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.61, + 0.35, + 0.714, + 0.361 + ], + "angle": 0, + "content": "(f) Ground Truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.365, + 0.788, + 0.476 + ], + "angle": 0, + "content": "Fig. 1: Relative pose estimation by SRPose. Dots drawn in the figures visualize the cross-attention scores of sparse keypoints across the two images, with brighter dots representing higher attention. Camera-to-world: (a), (b), (c) visualize the epipolar lines, representing the connections between the nine corresponding points across two views. Higher attention is shown to the overlap of the scenes. Object-to-Camera: (d), (e), (f) show the relative 6D pose estimation in the query image with only one accessible object prompt in the reference image. Higher attention is shown to the target object. SRPose establishes implicit correspondences." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.486, + 0.787, + 0.532 + ], + "angle": 0, + "content": "seem distinct, they both entail the prediction of the pose matrix from the two RGB images depicting the scene or object, which in turn represents the relative pose transformation from the first image to the second." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.535, + 0.788, + 0.763 + ], + "angle": 0, + "content": "Traditional pose estimation approaches typically involve detecting and matching keypoints or local features in the two images to establish point-to-point correspondences, then use a robust estimator to denoise the outliers and recover the relative pose from the essential matrix by solving a two-view geometry problem, i.e., the epipolar constraint equation [48]. These approaches produce accurate results regardless of image sizes and camera intrinsicics, thanks to the intrinsic calibration performed before solving the equation. Given the segmentation of the target object in the two images, these approaches can also yield acceptable predictions of the relative object pose transformation. However, on the downside, robust estimators can be computationally expensive. Indeed, the process of eliminating mismatched correspondences and solving the equations using robust estimators is significantly slower compared to detecting and matching keypoints or features, making it prohibitive in the real-time application. Moreover, real-time object pose tracking requires video object segmentation models to eliminate off-target keypoints [66, 67], which can also introduce additional overhead." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.842 + ], + "angle": 0, + "content": "The maturity of deep learning offers the advantage of directly regressing the relative pose transformation from two RGB inputs, significantly boosting the runtime performance. However, existing regression methods lack of generalizability or precision to images with varying sizes and camera intrinsics [3, 38, 53, 64, 74, 77]. They typically employ image encoders that only accept" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.282, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SRPose: Two-view Relative Pose Estimation with Sparse Keypoints" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.236 + ], + "angle": 0, + "content": "fixed-size inputs during batched training. And unlike traditional approaches, most deep learning-based regressors lack awareness of the original camera intrinsic parameters, unable to utilize two-view geometry to achieve higher precision in pose estimation. Furthermore, state-of-the-art deep regressors only apply to camera-to-world pose estimation, despite the common solution shared by both the camera-to-world and object-to-camera tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.785, + 0.449 + ], + "angle": 0, + "content": "To address the aforementioned challenges, we propose SRPose: a Sparse keypoint-based framework for Relative Pose estimation. SRPose estimates the relative pose matrix based on two-view geometry by implicitly solving the epipolar constraint, as shown in Fig. 1. First, SRPose employs a point detector to extract sparse keypoints and associated descriptors to form the candidate set of implicit correspondences. Then, we use an intrinsic-calibration (IC) position encoder to modulate the keypoints before it computes the position embeddings, accounting for varying image sizes and camera intrinsics. The keypoint descriptors and the position embeddings are then fused into the prior knowledge-guided attention layers, which leverage the prior knowledge of keypoint similarities to establish implicit cross-view correspondences. With the mechanism of the promptable cross-attention in these layers, SRPose only requires an accessible object prompt in one of the two views to compute the relative 6D pose transformation of the target object, eliminating the need for elusive object segmentations." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.45, + 0.722, + 0.464 + ], + "angle": 0, + "content": "The main contribution of this paper can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.468, + 0.784, + 0.511 + ], + "angle": 0, + "content": "- We propose a novel framework using sparse keypoints, SRPose, for two-view relative pose estimation. To the best of our knowledge, it is the first attempt to directly regress relative poses from sparse keypoints for this task." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.514, + 0.784, + 0.54 + ], + "angle": 0, + "content": "- We introduce an intrinsic-calibration position encoder to adapt SRPose to different image sizes and camera intrinsics." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.544, + 0.784, + 0.57 + ], + "angle": 0, + "content": "- We enable object-to-camera estimation in addition to camera-to-world estimation by utilizing one accessible object prompt." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.575, + 0.784, + 0.617 + ], + "angle": 0, + "content": "- Our SRPose greatly reduces estimation time by replacing robust estimators with direct regression, achieving state-of-the-art performance in accuracy and speed in camera-to-world and object-to-camera scenarios." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.468, + 0.784, + 0.617 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.637, + 0.391, + 0.652 + ], + "angle": 0, + "content": "2 Related works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Relative Pose Estimation: In relative pose estimation, traditional approaches recover the rotation and translation from the essential matrix [27, 42]. By using the epipolar constraint [48], the essential matrix is estimated based on the cross-view correspondences established through the matching of sparse keypoints or dense features. Using this pipeline, sparse keypoint-based approaches typically start by detecting sparse keypoints and their associated descriptors, including classic hand-crafted detectors [5, 44, 54], and deep learning-based detectors [6, 15-17, 24, 35, 40, 45, 52, 63, 72, 75, 76]. Recent methods further employ deep learning-based matchers [12, 39, 55, 56, 70] to establish keypoint matches, while classical approaches rely on nearest neighbor search. In contrast to using sparse keypoints, dense matchers [11, 13, 18, 19, 36, 47, 59, 62, 65, 71, 73, 79] are detector-free and perform pixel-wise dense feature matching. Once the correspondences" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "R. Yin et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "are established through matching, a robust estimator such as RANSAC [23] is employed to estimate the essential matrix and recover the relative pose. Overall, while the traditional matcher-based approaches have shown promising results, they still face challenges with their time-consuming robust estimators. Furthermore, dense matchers, while providing exceptional accuracy, may suffer from slow performance due to their resource-intensive design." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.241, + 0.788, + 0.483 + ], + "angle": 0, + "content": "Relative Pose Regression: Deep learning regressors offer an alternative approach by directly predicting the relative pose using dedicated neural networks, without explicitly establishing correspondences. Two-view relative pose estimation focuses on frame-to-frame pose transformations, aiming at visual odometry and map-free relocalization [1,3,4,21,34,46,68]. Leveraging prior knowledge is an important aspect in two-view estimation. For instance, SparsePlanes [32], PlaneFormer [8] and NOPE-SAC [61] utilize geometry relations of the planar surfaces to enhance performance. Unfortunately, these approaches rely on image encoders designed for images of fixed sizes and intrinsics, lacking generalizability across different settings. Sparse-view relative pose estimation emphasizes global pose optimization or end-to-end SfM involving multiple views [38,57,64,74]. Although these sparse-view methods can be extended to two-view estimation, and may even predict unknown intrinsics, their adaptability to varying camera settings depends on multi-view information. In contrast, SRPose leverages sparse keypoints and two-view geometry to achieve higher precision and generalizability across different image sizes and camera intrinsics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.486, + 0.788, + 0.759 + ], + "angle": 0, + "content": "Object Pose Estimation: While most of the aforementioned approaches focus on camera pose estimation in a fixed scene, some tasks, such as 6D object pose tracking, require relative object pose estimation as well. Most frameworks predict the relative object pose using the same pipeline of keypoint correspondences. BundleTrack [66] and BundleSDF [67] track the object poses in videos by matching keypoints between two frames, and conducting global pose graph optimization. POPE [22] combines advanced foundation models of segmentation [33] and matching [49] to enable zero-shot object pose estimation. Robust as these classic methods may be, they require dedicated object masks to focus on the objects, which are provided by other resource-consuming segmentation models. To get rid of object segmentation, OnePose [60], OnePose++ [29], and Gen6D [41] rely on multiple views of the object, and ZeroShot [25] establishes semantic correspondences to mute the background, and employs depth map to enable pose estimation. Nonetheless, these methods require additional high-quality information, and they restrict the querying view to only accommodate a single object. As a remedy, SRPose employs an accessible object prompt to focus on the target object, enabling effective depth-free and mask-free two-view relative object estimation with superior performance in the object-to-camera scenario." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.762, + 0.332, + 0.778 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Given two images \\( I_1, I_2 \\), SRPose estimates the relative pose consisting of a rotation \\( R \\in \\mathcal{SO}(3) \\), and a translation \\( t \\in \\mathbb{R}^3 \\) based on the two-view geometry. Detailed problem definitions can be found in the supplementary materials. Traditional approaches use the epipolar constraint [48] to estimate the following" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.282, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "SRPose: Two-view Relative Pose Estimation with Sparse Keypoints" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.147, + 0.768, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.269, + 0.788, + 0.367 + ], + "angle": 0, + "content": "Fig. 2: Overview. SRPose comprises four main components: 1) The sparse keypoint detector detects keypoints associated with descriptors separately sfrom the two images; 2) The intrinsic-calibration (IC) position encoder modulates the keypoints' coordinates with camera intrinsics, and encodes their position information; 3) Guided by the prior knowledge of keypoint similarities, along with the object prompt, the attention layers establish implicit cross-view correspondences; 4) The regressor estimates relative pose \\( R, t \\) under the constraints of implicit correspondences." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.373, + 0.578, + 0.388 + ], + "angle": 0, + "content": "essential matrix \\(E\\), from which \\(R, t\\) are recovered:" + }, + { + "type": "equation", + "bbox": [ + 0.375, + 0.396, + 0.786, + 0.415 + ], + "angle": 0, + "content": "\\[\n\\left(K _ {1} ^ {- 1} q _ {1}\\right) ^ {\\top} E K _ {2} ^ {- 1} q _ {2} = 0, \\quad \\forall \\left(q _ {1}, q _ {2}\\right), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.461, + 0.417, + 0.785, + 0.433 + ], + "angle": 0, + "content": "\\[\nE = [ t ] _ {\\times} R, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.442, + 0.785, + 0.503 + ], + "angle": 0, + "content": "where \\( q_{1}, q_{2} \\) are two arbitrary corresponding points in \\( I_{1}, I_{2} \\), and \\( K_{1}, K_{2} \\) are the camera intrinsics of the two images. SRPose builds and solves the constraint equation Eq. (1) with neural networks implicitly. With SRPose denoted as \\( \\mathcal{SR} \\), the function of our framework is represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.433, + 0.512, + 0.785, + 0.529 + ], + "angle": 0, + "content": "\\[\n[ R | t ] = \\mathcal {S R} (I _ {1}, I _ {2}). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.538, + 0.784, + 0.584 + ], + "angle": 0, + "content": "Figure 2 summarizes SRPose, involving four components: a sparse keypoint detector, an intrinsic-calibration position encoder, the promptable prior knowledge-guided attention layers, and an MLP regressor." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.589, + 0.478, + 0.604 + ], + "angle": 0, + "content": "3.1 Sparse Keypoint Detector" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.605, + 0.788, + 0.742 + ], + "angle": 0, + "content": "Instead of employing an image encoder to extract image features, we use a detector to extract sparse keypoints and associated descriptors from the two images. To this end, we can use one of the existing techniques developed over the years, which include classic methods such as SIFT [44], and deep learning-based methods such as SuperPoint [16], ALIKE [76], DISK [63], etc. Given an image \\(I\\), our detector, denoted as \\(\\mathcal{F}^{kd}\\), detects keypoints consisting of \\(N\\) coordinates \\(P^k \\subset \\mathbb{R}^2\\) in image space, and associated descriptors \\(D \\subset \\mathbb{R}^d\\), describing the local features in the image, where \\(d\\) denotes the dimension of the keypoints descriptors in the network. To summarize, our detector is defined as the following function:" + }, + { + "type": "equation", + "bbox": [ + 0.431, + 0.75, + 0.786, + 0.768 + ], + "angle": 0, + "content": "\\[\n\\left\\{P ^ {k}, D \\right\\} = \\mathcal {F} ^ {k d} (I). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.776, + 0.788, + 0.853 + ], + "angle": 0, + "content": "In practice, we resize all images to a standardized size to enable parallel batched detection, then rescale the keypoints' coordinates \\( P^k \\) to their original positions. The sparse keypoint detector yields two sets of keypoints \\( P_1^k, P_2^k \\) from \\( I_1, I_2 \\), from which a set of correspondences \\( \\{(q_1, q_2)\\} \\subset P_1^k \\times P_2^k \\) will be established implicitly later for solving the constraint Eq. (1)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "R. Yin et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.62, + 0.162 + ], + "angle": 0, + "content": "3.2 Intrinsic-Calibration (IC) Position Encoder" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.164, + 0.786, + 0.194 + ], + "angle": 0, + "content": "According to Eq. (1), for the keypoints detected from an image \\(I\\), their coordinates should be first calibrated or normalized by the camera intrinsics \\(K\\):" + }, + { + "type": "equation", + "bbox": [ + 0.454, + 0.205, + 0.786, + 0.222 + ], + "angle": 0, + "content": "\\[\nP ^ {c} = K ^ {- 1} P ^ {k}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.231, + 0.784, + 0.351 + ], + "angle": 0, + "content": "Enforcing the intrinsic calibration (IC) offers a crucial improvement of both the accuracy and robustness [48]. IC allows SRPose to adapt images captured by different cameras with varying intrinsic parameters. Previous regressors [3, 53] extract implicit image features with image encoders trained on fixed-size images, assuming the camera intrinsics of all inputs are the same by default. While the sparse keypoints in SRPose are modulated by intrinsic calibration, which provides the accurate positions, i.e. \\( P^c \\), of all keypoints in a unified camera coordinate system across different intrinsics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.353, + 0.785, + 0.457 + ], + "angle": 0, + "content": "After the intrinsic calibration, the keypoints' coordinates are encoded to address the keypoints based on their positions. Due to the possible position distortion caused by relative position encoding [39, 59], we employ absolute position encoding that maintains the precise coordinate values, which is crucial to the two-view geometry problems. With a single-layer fully connected network, denoted as the following function \\(\\mathcal{F}^{pe}\\), the position encoder maps 2D calibrated coordinates to high-dimension position embeddings:" + }, + { + "type": "equation", + "bbox": [ + 0.434, + 0.469, + 0.786, + 0.485 + ], + "angle": 0, + "content": "\\[\nP ^ {e} = \\mathcal {F} ^ {p e} \\left(K ^ {- 1} P ^ {k}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.495, + 0.784, + 0.557 + ], + "angle": 0, + "content": "The resulting encoded position embeddings, denoted as \\( P^e \\), are then incorporated with the keypoint descriptors and descriptor embeddings, denoted as \\( D \\oplus P^e \\), which provides calibrated and normalized position information for solving the constraint Eq. (1) using a set of attention layers." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.563, + 0.657, + 0.578 + ], + "angle": 0, + "content": "3.3 Promptable Prior-Knowledge-Guided Attention" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.58, + 0.785, + 0.776 + ], + "angle": 0, + "content": "Given the position embeddings and keypoint descriptors, we use a multi-layer attention network to exploit semantic information and establish implicit correspondences. Our network consists of \\(M\\) consecutive attention layers, with each layer comprising a self-attention module and a promptable prior knowledge-guided cross-attention module. We denote \\(X^{m} \\subset \\mathbb{R}^{d}\\) as the keypoint descriptor embeddings being processed in the \\(m^{th}\\) layer, in which \\(m \\in M\\), with \\(X^{m}\\) being either \\(X_{1}^{m}\\) or \\(X_{2}^{m}\\) from the image \\(I_{1}\\) or \\(I_{2}\\). Each layer takes \\(X^{m-1}\\) as inputs, then computes the deeper embeddings \\(X^{m}\\). In summary, the entire multi-layer network inputs keypoint descriptors \\(D_{1}D_{2}\\), and outputs \\(X_{1}^{M}, X_{2}^{M}\\) after \\(M\\) layers, i.e., we define \\(D\\) as the initial state of embeddings \\(X^{0}\\). Prior to each attention layer, \\(X_{1}^{m-1}, X_{2}^{m-1}\\) are incorporated with the associated position embeddings \\(P_{1}^{e}, P_{2}^{e}\\). Following [28], residual connections are applied to every self- and cross-attention module. The architecture of the attention layers is outlined in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.786, + 0.841 + ], + "angle": 0, + "content": "Self-Attention: The self-attention module captures the comprehensive semantic information of all keypoints within an entire image. In this module, \\( X^{m - 1} \\oplus P^e \\) of both the images is first mapped to three vectors by a linear transformation \\( W_{s}^{m} \\), denoted as the query \\( q \\), the key \\( k \\), and value \\( v \\) respectively, where" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.282, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "SRPose: Two-view Relative Pose Estimation with Sparse Keypoints" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.147, + 0.761, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.283, + 0.785, + 0.352 + ], + "angle": 0, + "content": "Fig. 3: Overview of the prior knowledge-guided attention layers. Each layer contains a self-attention and a cross-attention module. A similarity matrix \\( S \\) of the keypoint descriptors is utilized as the prior knowledge to readjust the cross-attention scores, guiding more attention to cross-view keypoint pairs with implicit correspondences. The object prompt and the residual connections are omitted in the figure." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.358, + 0.784, + 0.389 + ], + "angle": 0, + "content": "\\(q, k, v \\subset \\mathbb{R}^d\\). Then \\(k\\) and \\(q\\) are used to compute the self-attention scores \\(A_s\\), which are then scaled by the dimension of descriptors and embeddings \\(d\\):" + }, + { + "type": "equation", + "bbox": [ + 0.4, + 0.398, + 0.786, + 0.418 + ], + "angle": 0, + "content": "\\[\n[ q | k | v ] = W _ {s} ^ {M} \\left(X ^ {m - 1} \\oplus P ^ {e}\\right), \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.46, + 0.419, + 0.786, + 0.454 + ], + "angle": 0, + "content": "\\[\nA _ {s} = \\frac {q k ^ {\\top}}{\\sqrt {d}}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.462, + 0.784, + 0.492 + ], + "angle": 0, + "content": "After being normalized by the Softmax function, the self-attention scores \\( A_{s} \\) are then used to extract the highly attended information from \\( v \\):" + }, + { + "type": "equation", + "bbox": [ + 0.414, + 0.502, + 0.786, + 0.521 + ], + "angle": 0, + "content": "\\[\nX _ {s} ^ {m - 1} = \\operatorname {S o f t m a x} \\left(A _ {s}\\right) v. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.53, + 0.784, + 0.576 + ], + "angle": 0, + "content": "In the self-attention module, keypoint position embeddings and descriptor embeddings are computed together to summarize the semantic information within one entire image before deeper cross-view correspondences are established." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.579, + 0.785, + 0.701 + ], + "angle": 0, + "content": "Prior Knowledge-guided Cross-Attention: Guided by the prior knowledge of keypoint similarities, the cross-attention module exploits the mutual information across the two images to establish implicit correspondences. Similar to the self-attention, cross-attention maps the input embeddings \\( X_{s}^{m - 1} \\) to two keys \\( k_{1},k_{2} \\) and values \\( v_{1},v_{2} \\) of the two images respectively through another linear transformation \\( W_{c}^{m} \\) layer. Following [50], we compute the bidirectional attention scores across the two images from the resulting embeddings of the previous self-attention module, using the cross-view keys \\( k_{1},k_{2} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.426, + 0.71, + 0.786, + 0.728 + ], + "angle": 0, + "content": "\\[\n\\left[ k _ {1} \\mid v _ {1} \\right] = W _ {c} ^ {m} X _ {s, 1} ^ {m - 1}, \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.427, + 0.731, + 0.786, + 0.75 + ], + "angle": 0, + "content": "\\[\n\\left[ k _ {2} \\mid v _ {2} \\right] = W _ {c} ^ {m} X _ {s, 2} ^ {m - 1}, \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.43, + 0.752, + 0.786, + 0.786 + ], + "angle": 0, + "content": "\\[\nA _ {2 1} ^ {\\top} = A _ {1 2} = \\frac {k _ {1} k _ {2} ^ {\\top}}{\\sqrt {d}}. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.795, + 0.784, + 0.84 + ], + "angle": 0, + "content": "The module recognizes the overlapping areas of the scene or object between the two views by identifying the mutuality between two sets of keypoint embeddings from the computed cross-attention scores. To facilitate the process, prior" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "R. Yin et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.788, + 0.209 + ], + "angle": 0, + "content": "knowledge is used to guide and enhance the attention on highly correlated keypoints, by adjusting the cross-attention scores with a keypoint similarity matrix \\( S \\in \\mathbb{R}^{N_1 \\times N_2} \\). Each entry \\( s_{ij} \\) of the matrix is computed as the cosine similarity between the descriptors of every cross-view keypoint pair:" + }, + { + "type": "equation", + "bbox": [ + 0.331, + 0.219, + 0.786, + 0.252 + ], + "angle": 0, + "content": "\\[\ns _ {i j} = \\left(\\frac {d _ {i} \\cdot d _ {j}}{\\| d _ {i} \\| \\| d _ {j} \\|} + 1\\right) / 2, \\quad \\{(d _ {i}, d _ {j}) \\} \\subseteq D _ {1} \\times D _ {2}. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.261, + 0.784, + 0.32 + ], + "angle": 0, + "content": "Keypoints with similar local features tend to possess similar descriptors, describing correlated semantic information of similar views. Thus, the cosine similarity captures the semantic similarities between two keypoints, and guides the framework to recognize the overlapping parts between the images." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.321, + 0.785, + 0.367 + ], + "angle": 0, + "content": "We normalize the matrix \\( S \\) to [0,1] and element-wise multiply it with the cross-attention scores. Finally, the scores are used to extract and exchange crucial cross-view information from each others' values \\( v_{1}, v_{2} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.428, + 0.377, + 0.785, + 0.395 + ], + "angle": 0, + "content": "\\[\nA _ {2 1} ^ {\\prime \\top} = A _ {1 2} ^ {\\prime} = A _ {1 2} \\cdot S, \\tag {14}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.411, + 0.399, + 0.785, + 0.416 + ], + "angle": 0, + "content": "\\[\nX _ {1} ^ {m - 1} = \\operatorname {S o f t m a x} \\left(A _ {1 2} ^ {\\prime}\\right) v _ {2}, \\tag {15}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.411, + 0.419, + 0.785, + 0.437 + ], + "angle": 0, + "content": "\\[\nX _ {2} ^ {m - 1} = \\operatorname {S o f t m a x} \\left(A _ {2 1} ^ {\\prime}\\right) v _ {1}. \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.448, + 0.784, + 0.538 + ], + "angle": 0, + "content": "By capturing and incorporating relevant information from each image into the other, the implicit correspondences are established across the two images. As shown in Figs. 1a and 1b, with brighter dots representing higher cross-attention scores, SRPose establishes implicitly corresponding points densely in the overlapping areas of the two images, which is utilized to construct and solve an implicit version of the epipolar constraint Eq. (1)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.542, + 0.788, + 0.785 + ], + "angle": 0, + "content": "Object Prompt: In the object-to-camera estimation, we employ an accessible user-provided object prompt in only one image to identify the target object. In contrast, traditional matching approaches require high-quality object segmentations in both images. An object prompt \\( b \\) is a bounding box consisting of the two coordinates of top-left and bottom-right points, which bounds the object \\( o \\) in one of the images \\( I_1 \\), making \\( I_1 \\) as the reference view. Once the keypoints are extracted from \\( I_1 \\), the object prompt removes the keypoints outside of the bounding box in the reference view, eliminating most irrelevant keypoints. Thus, the descriptors and positions of keypoints within bounds, focusing on the target object \\( o \\), will be later processed in the Similarity Matrix and other following modules. With the guided cross-attention modules, SRPose learns to identify highly correlated keypoints lying on the same object \\( o \\) from the entire query view \\( I_2 \\), excluding irrelevant background keypoints, as shown in Figs. 1d and 1e. With the object prompt, SRPose searches the implicit corresponding points on the target object and computes the relative 6D object pose transformation in the same process as the camera-to-world scenario." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.792, + 0.462, + 0.808 + ], + "angle": 0, + "content": "3.4 Relative Pose Regressor" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.809, + 0.787, + 0.842 + ], + "angle": 0, + "content": "After the embeddings \\( X_1^M, X_2^M \\) are computed via the previous modules, a MultiLayer Perceptron (MLP) is employed to regress the relative pose as our last step." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.281, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "SRPose: Two-view Relative Pose Estimation with Sparse Keypoints" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.284 + ], + "angle": 0, + "content": "The outputs of the previous attention layers \\(X_1^M, X_2^M\\) first undergo 2D average pooling, and then are concatenated together, resulting in the joint embedding \\(f \\in \\mathbb{R}^{2d}\\), which carries the deep features and implicit correspondences across the two images. Finally, embedding \\(f\\) is input to the MLP, and the regressor solves the epipolar constraint Eq. (1) implicitly and regresses the relative pose, consisting of a rotation \\(R\\), and a translation \\(t\\). Following [78], we set the rotation output of the regressor to be a 6d vector \\(r \\in \\mathbb{R}^6\\), which is then transformed into a rotation matrix \\(R \\in \\mathbb{R}^{3 \\times 3}\\) through a partial Gram-Schmidt process. This technique, as argued in [78], enables a continuous representation of rotations." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.291, + 0.379, + 0.305 + ], + "angle": 0, + "content": "3.5 Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.309, + 0.787, + 0.356 + ], + "angle": 0, + "content": "SRPose is trained in a supervised manner. With the ground-truth rotation \\( R_{gt} \\), we compute the Huber Loss [31] of the error in the rotation angle, which minimizes the angular error between the predicted and ground-truth rotations:" + }, + { + "type": "equation", + "bbox": [ + 0.342, + 0.366, + 0.786, + 0.4 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {R} \\left(I _ {1}, I _ {2}\\right) = \\mathcal {H} \\left(\\operatorname {a r c c o s} \\left(\\frac {\\operatorname {T r} \\left(R ^ {\\top} R _ {g t} - 1\\right)}{2}\\right)\\right), \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.409, + 0.788, + 0.47 + ], + "angle": 0, + "content": "where \\(\\mathcal{H}\\) denotes the Huber Loss function. Similarly, given the ground-truth translation \\(t_{gt}\\), the error is computed in both normalized and unnormalized forms. To further enhance accuracy, we also incorporate the angular error of translation, resulting in the following three loss terms:" + }, + { + "type": "equation", + "bbox": [ + 0.421, + 0.483, + 0.785, + 0.499 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {t} \\left(I _ {1}, I _ {2}\\right) = \\mathcal {H} \\left(t - t _ {g t}\\right), \\tag {18}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.389, + 0.503, + 0.785, + 0.535 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {t _ {n}} \\left(I _ {1}, I _ {2}\\right) = \\mathcal {H} \\left(\\frac {t}{\\| t \\|} - \\frac {t _ {g t}}{\\| t _ {g t} \\|}\\right), \\tag {19}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.365, + 0.538, + 0.785, + 0.571 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {t _ {a}} \\left(I _ {1}, I _ {2}\\right) = \\mathcal {H} \\left(\\arccos \\left(\\frac {t \\cdot t _ {g t}}{\\| t \\| \\| t _ {g t} \\|}\\right)\\right). \\tag {20}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.58, + 0.787, + 0.611 + ], + "angle": 0, + "content": "The above three loss functions supervise both the scale and direction of translations. We compute the final loss by the weighted sum of the three functions:" + }, + { + "type": "equation", + "bbox": [ + 0.355, + 0.626, + 0.785, + 0.643 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} \\left(I _ {1}, I _ {2}\\right) = \\mathcal {L} _ {R} + \\lambda_ {t} \\mathcal {L} _ {t} + \\lambda_ {t _ {n}} \\mathcal {L} _ {t _ {n}} + \\lambda_ {t _ {a}} \\mathcal {L} _ {t _ {a}}, \\tag {21}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.648, + 0.784, + 0.665 + ], + "angle": 0, + "content": "where \\(\\lambda_t, \\lambda_{t_n}, \\lambda_{t_a}\\) are scalars used to balance the weights of four different losses." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.671, + 0.457, + 0.686 + ], + "angle": 0, + "content": "3.6 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.844 + ], + "angle": 0, + "content": "SRPose is first trained on ScanNet [14] and then fine-tuned on other datasets to achieve better performance. For all datasets, we resize the images to \\(640 \\times 480\\), and extract 1,024 sparse keypoints using SuperPoint [16] for training and 2,048 for evaluation. Other sparse keypoint detectors can also be used for this step. We utilize the pre-trained SuperPoint provided by the official source and freeze its weights in the training stage. The remaining learnable modules are initialized with random weights. In practice, SRPose consists of 6 guided attention layers, with 4 attention heads in each attention module. The dimension \\(d\\) of keypoint descriptors and embeddings is set to 256. The rotation and translation are estimated using two separate 3-layer MLPs. All three balancing scalars \\(\\lambda_{t}, \\lambda_{t_{n}}, \\lambda_{t_{a}}\\)" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "R. Yin et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.787, + 0.187 + ], + "angle": 0, + "content": "Table 1: Relative Pose Estimation on Matterport [10]. Without depth, matcher-based approaches are incapable of scaled translation estimation, while SRPose achieves high accuracy in translation scales using only RGB inputs." + }, + { + "type": "table", + "bbox": [ + 0.258, + 0.189, + 0.742, + 0.349 + ], + "angle": 0, + "content": "
CategoryMethodRot. (°)Trans. (m)
Med.↓Avg.↓≤ 30° ↑Med.↓Avg.↓≤ 1m↑
SparseSuperGlue [55]3.8824.1777.8n/an/an/a
SGMNet [12]1.0621.3280.9n/an/an/a
LightGlue [39]1.3222.4580.0n/an/an/a
DenseLoFTR [59]0.7111.1190.5n/an/an/a
ASpanFormer [13]3.7331.4575.7n/an/an/a
DKM [18]0.4612.8989.0n/an/an/a
RegressorSparsePlanes [32]7.3322.7883.40.631.2566.6
PlaneFormers [2]5.9622.2083.80.661.1966.8
8point [53]8.0119.1385.40.641.0167.4
NOPE-SAC [61]2.7714.3789.00.520.9473.2
SRPose2.6511.1291.60.270.6183.7
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.367, + 0.787, + 0.445 + ], + "angle": 0, + "content": "are set to 1. On ScanNet, our framework is trained using AdamW optimizer [43] and follows 1cycle learning rate policy [58] with a maximum learning rate of \\(1 \\times 10^{-4}\\) for 500 epochs. We fix the batch size to 32 for training. The pretraining requires 120 hours utilizing 8 RTX 3090 GPUs. More implementation details can be found in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.455, + 0.377, + 0.473 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.482, + 0.787, + 0.545 + ], + "angle": 0, + "content": "We evaluate the performance of SRPose for two-view relative pose estimation in both the camera-to-world and object-to-camera scenarios. Additionally, we assess its effectiveness in the application of map-free relocalization. Finally, we conduct an ablation study to analyze the roles of the components we proposed." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.557, + 0.547, + 0.572 + ], + "angle": 0, + "content": "4.1 Camera-to-World Pose Estimation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.581, + 0.787, + 0.733 + ], + "angle": 0, + "content": "Setup: In the camera-to-world scenario, we evaluate our framework's performance on Matterport [10] and ScanNet [14]. Following [32], for Matterport, we report the median and mean of the translation and rotation error, and the translation/rotation accuracy at the thresholds of \\( 1\\mathrm{m} / 30^{\\circ} \\), respectively. For ScanNet, we adopt the metrics from [55] to compute the AUC of the pose error at \\( 5^{\\circ} \\), \\( 10^{\\circ} \\), and \\( 20^{\\circ} \\). The pose error here is the maximum angular error in rotation and translation. We also report an analysis of the time consumption on the ScanNet-1500 test set. For all matcher-based approaches, we use the implementation of RANSAC [23] from OpenCV [7] to recover relative poses. All methods are compared on a device with RTX 4090, i9-13900K, and 128GB memory." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.746, + 0.787, + 0.854 + ], + "angle": 0, + "content": "Baselines: We choose three categories of approaches as our baselines, including sparse keypoint-based matchers, dense matchers, and deep learning-based relative pose regressors. Results of all the baselines are reported according to the official papers or implementations. All sparse matchers and SRPose employ SuperPoint [16] as their sparse keypoint detector. All regressors are trained on the specific dataset, i.e. Matterport or ScanNet, before evaluation. LightGlue [39] is trained on MegaDepth [37], while all other matchers are trained on ScanNet." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.281, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SRPose: Two-view Relative Pose Estimation with Sparse Keypoints" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.784, + 0.186 + ], + "angle": 0, + "content": "Table 2: Relative Pose Estimation on ScanNet [14]. SRPose outperforms state-of-the-art regressors and sparse matchers in terms of various metrics and reduces computational time greatly due to the riddance of robust estimators." + }, + { + "type": "table", + "bbox": [ + 0.251, + 0.189, + 0.749, + 0.32 + ], + "angle": 0, + "content": "
CategoryMethodPose estimationAUCTime (ms)
@5°@10°@20°MatchTotal
SparseSuperGlue [55]16.233.851.8102.5361.9
SGMNet [12]15.432.148.3111.7439.7
LightGlue [39]16.433.650.231.5269.7
DenseLoFTR [59]22.040.857.644.4292.4
ASpanFormer [13]25.646.063.359.6370.9
DKM [18]29.450.768.3235.0491.6
RegressorMap-free [3]2.7011.529.0n/a12.7
SRPose13.334.356.8n/a28.5
" + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.323, + 0.77, + 0.464 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.466, + 0.785, + 0.494 + ], + "angle": 0, + "content": "Fig. 4: Time consumption comparison on ScanNet [14]. Regressors, including SRPose, achieve much higher computational efficacy than all matchers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.51, + 0.788, + 0.692 + ], + "angle": 0, + "content": "Results: Table 1 demonstrates that SRPose greatly outperforms the existing deep learning-based regressors across all metrics on Matterport. Our framework achieves the most accurate scaled translation estimation, which is unattainable for matcher-based approaches lacking depth information. Table 2 shows SRPose achieves competitive or superior performance compared to all sparse matchers and the regressor and on ScanNet. Although dense matchers outperform our framework in accuracy, SRPose offers a significant advantage in computational speed compared to all the matcher-based approaches by replacing robust estimators with direct regression. As shown in Fig. 4, matcher-based approaches spend a substantial proportion of time on recovering pose using robust estimators after matching. In contrast, SRPose consumes even less time than the matching stage of all matchers, saving at least 200ms in the total time." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.701, + 0.553, + 0.716 + ], + "angle": 0, + "content": "4.2 Object-to-Camera Pose Estimation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Setup: In the object-to-camera scenario, we evaluate SRPose on HO3D [26]. HO3D contains videos of hand-holding objects of 21 categories from the YCB dataset [9]. The depth captured by the stereo depth camera and the object segmentation of each frame in the videos are also provided. We randomly select image pairs with small object rotation transformations for training and evaluation. More details can be found in the supplementary materials. Following [30, 69], we adopt the widely-used Average Distance (ADD, ADD-S) metric for evaluation. To measure the smaller scales in object pose transformation, we set the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "R. Yin et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.787, + 0.187 + ], + "angle": 0, + "content": "Table 3: Relative object pose estimation on HO3D [26]. SRPose achieves state-of-the-art performance in object-to-camera estimation without additional depth information, enabling scale pose estimation from only RGB inputs." + }, + { + "type": "table", + "bbox": [ + 0.244, + 0.19, + 0.756, + 0.285 + ], + "angle": 0, + "content": "
MethodRot. (°)Trans. (cm)ADD↑ADD-S↑
Med.↓Avg.↓≤ 30°↑Med.↓Avg.↓≤ 10cm↑
SGMNet [12]24.129.062.713.917.536.816.831.3
LightGlue [39]24.428.761.713.516.338.217.231.2
LoFTR [59]29.257.950.913.916.435.616.730.9
DKM [18]23.325.664.813.615.439.817.833.3
SRPose8.911.495.45.98.073.936.456.8
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.296, + 0.784, + 0.358 + ], + "angle": 0, + "content": "threshold of metrics to \\(10\\mathrm{cm}\\), which differs from the 1m threshold used in the camera-to-world scenario. Specifically, we calculate the AUC using both ADD and ADD-S with the threshold set to \\(10\\mathrm{cm}\\). We also report the median, mean error of translation/rotation, and their accuracy at \\(10\\mathrm{cm} / 30^{\\circ}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.364, + 0.785, + 0.487 + ], + "angle": 0, + "content": "Baselines: We only choose matchers listed in Sec. 4.1 as our baselines, for no deep learning-based regressors are available in this two-view relative object pose estimation task. We utilize the stereo depth offered in the dataset to enable matcher-based approaches to estimate scaled relative poses via Orthogonal Procrustes [20], while SRPose only requires RGB images as inputs. The segmentations in the dataset are adopted for matcher-based approaches to identify the target object in both images, while SRPose utilizes the bounding box of the segmentation in the first image as the object prompt." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.494, + 0.785, + 0.584 + ], + "angle": 0, + "content": "Results: As shown in Table 3, SRPose outperforms state-of-the-art matcher-based approaches significantly. SRPose exhibits superior accuracy in relative 6D object pose estimation between two images using the object prompts in the reference view. The low quality of stereo depth leads to limited performance in scaled pose estimation of the matcher-based approaches, while SRPose can learn scaled information to achieve lower estimation errors." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.592, + 0.522, + 0.607 + ], + "angle": 0, + "content": "4.3 Map-Free Visual Relocalization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.609, + 0.785, + 0.685 + ], + "angle": 0, + "content": "We evaluate the application of SRPose on Niantic [3], a dataset for the map-free visual relocalization task. Following [3], we adopt the Virtual Correspondence Reprojection Error (VCRE) and the median error of translation and rotation as the metrics. Specifically, we compute the precision of VCRE at 90 pixels, and the precision of relocalization at \\(0.25\\mathrm{m}\\) and \\(5^{\\circ}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.693, + 0.785, + 0.754 + ], + "angle": 0, + "content": "Baselines: We still select the three categories of approaches as our baselines. For **matching-based approaches**, we estimate the pose by recovering from the essential matrix. Then we use the depth estimated by DPT [51], a depth prediction model, to provide the matchers with the scale information." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.762, + 0.787, + 0.853 + ], + "angle": 0, + "content": "Results: Table 4 shows SRPose outperforms other sparse matchers and the regression method in map-free visual relocalization. Although the state-of-the-art dense matcher RoMa achieves higher precision in relocalization using scales estimated by an additional depth prediction model, SRPose still outperforms in VCRE. Our framework exhibits competitive performance compared to state-of-the-art approaches in the map-free visual relocalization task." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.281, + 0.114, + 0.733, + 0.131 + ], + "angle": 0, + "content": "SRPose: Two-view Relative Pose Estimation with Sparse Keypoints" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.137, + 0.788, + 0.179 + ], + "angle": 0, + "content": "Table 4: Map-free visual relocalization results on Niantic [3]. SRPose achieves competitive performance compared to state-of-the-art approaches without relying on an additional depth prediction model to assist with scaled estimation." + }, + { + "type": "table", + "bbox": [ + 0.283, + 0.182, + 0.714, + 0.293 + ], + "angle": 0, + "content": "
CategoryMethodVCRE \nPrec.↑ / Med.↓Med. (m / °) \nTrans.↓ / Rot.↓Prec.↑
SparseSIFT [44]25.0 / 222.82.93 / 61.410.3
+DPT [51]SuperGlue [55]36.1 / 160.31.88 / 25.416.8
DenseLoFTR [59]34.7 / 167.61.98 / 30.515.4
+DPT [51]RoMa [19]45.6 / 128.81.23 / 11.122.8
RegressorMap-free [3]40.2 / 147.11.68 / 22.96.0
SRPose46.4 / 127.71.37 / 17.216.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.296, + 0.787, + 0.325 + ], + "angle": 0, + "content": "Table 5: Ablation study on Matterport [10]. The prior knowledge-guided attention layers assist in establishing implicit cross-view correspondences." + }, + { + "type": "table", + "bbox": [ + 0.283, + 0.327, + 0.719, + 0.421 + ], + "angle": 0, + "content": "
MethodTrans. (m)Rot. (°)
Avg.↓≤ 1m↑Avg.↓≤ 30°↑
8point [53]1.0167.419.1385.4
SRPose (full)0.8474.214.3288.9
1) no prior knowledge guidance0.9769.616.9186.8
2) replace all cross-attn. with self-attn.1.9026.148.042.0
3) no position encoding1.1761.021.3981.3
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.431, + 0.329, + 0.446 + ], + "angle": 0, + "content": "4.4 Insights" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.487, + 0.599 + ], + "angle": 0, + "content": "Ablation study: We evaluate the effectiveness of the different components in SRPose by Matterport [10] and MegaDepth [37] datasets. Specifically, we train four different variants initialized with random weights on Matterport by removing the three different designs we proposed to validate their effectiveness. As shown in Table 5: 1) Removing the prior knowledge" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.449, + 0.788, + 0.505 + ], + "angle": 0, + "content": "Table 6: Ablation study on MegaDepth [37]. Intrinsic-calibration position encoder enables adaption on varying image sizes and camera intrinsics." + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.506, + 0.782, + 0.589 + ], + "angle": 0, + "content": "
MethodPose est.AUC
@5°@10°@20°
Map-free [3]2.69.322.9
SRPose (full)16.636.058.0
1) no intrinsic calibration1.58.524.3
2) no position encoding1.06.218.9
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.788, + 0.719 + ], + "angle": 0, + "content": "edge guidance of keypoint similarities in the cross-attention modules results in a moderate decrease in accuracy, as it assists SRPose in seeking the cross-view correspondences. 2) Replacing all cross-attention modules with self-attention modules leads to a significant accuracy drop, as the capability to establish implicit cross-view correspondences through cross-attention is eliminated. 3) Removing the entire position encoder leads to a moderately lower accuracy due to the absence of keypoint position information. The experiment highlights the crucial roles of the three designs of SRPose in accurate relative pose estimation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Moreover, we train three different variants initialized with random weights on MegaDepth, a dataset that consists of images with different sizes and camera intrinsics, to further validate the effectiveness of the intrinsic-calibration position encoder, with results shown in Table 6. 1) Without enforcing intrinsic calibration on keypoint coordinates, SRPose only achieves comparable performance to Map-free, both inadaptable to varying image sizes and camera intrinsics in MegaDepth. 2) Removing the entire IC Position Encoder results in a significant drop in accuracy as expected owing to the elimination of critical position infor" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "R. Yin et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.193 + ], + "angle": 0, + "content": "mation. By comparing the full framework with the variants lacking IC and the entire position encoder, we demonstrate the essential roles of the two designs in adapting different image sizes and camera intrinsics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.197, + 0.788, + 0.258 + ], + "angle": 0, + "content": "Visualizing attention: In Fig. 5, we visualize the cross-attention scores \\( A_{12}^{\\prime} \\) and compare SRPose without prior knowledge guidance to its full designs. We also draw the connections of highly-attended cross-view keypoint pairs, which represent the implicit correspondences established by SRPose." + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.267, + 0.766, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.576, + 0.788, + 0.634 + ], + "angle": 0, + "content": "Fig.5: Visualization of implicit cross-view correspondences. Dots and lines drawn with brighter colors represent higher cross-view attention scores. The prior knowledge guidance enhances attention to the overlapping areas, removing the irrelevant cross-view connections, and assisting in establishing implicit correspondences." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.639, + 0.36, + 0.654 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.789, + 0.841 + ], + "angle": 0, + "content": "This paper presents SRPose, a novel deep learning-based regressor utilizing sparse keypoints for two-view relative pose estimation in camera-to-world and object-to-camera scenarios. As our key innovations, SRPose extracts keypoints with sparse keypoint detectors and employs an intrinsic-calibration position encoder to enable adaptability to different image sizes and camera intrinsics. Further, our proposed promptable prior knowledge-guided attention layers establish implicit correspondences to estimate the relative pose under the epipolar constraint in both scenarios. By directly regressing the rotation and translation, SRPose achieves a significant decrease in computational time, with a minimum time reduction of 200ms. Extensive experiments demonstrate that SRPose achieves state-of-the-art performance in terms of accuracy and speed in the two scenarios, as well as in the map-free visual relocalization task." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.282, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SRPose: Two-view Relative Pose Estimation with Sparse Keypoints" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.175, + 0.787, + 0.22 + ], + "angle": 0, + "content": "This work was supported by Shanghai Municipal Science and Technology Major Project (2021SHZDZX0102) and the Fundamental Research Funds for the Central Universities." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.246, + 0.323, + 0.263 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.279, + 0.785, + 0.308 + ], + "angle": 0, + "content": "1. Abouelnaga, Y., Bui, M., Ilic, S.: Distillpose: lightweight camera localization using auxiliary learning. In: IROS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.309, + 0.785, + 0.336 + ], + "angle": 0, + "content": "2. Agarwala, S., Jin, L., Rockwell, C., Fouhey, D.F.: Planeformers: From sparse view planes to 3d reconstruction. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.337, + 0.785, + 0.378 + ], + "angle": 0, + "content": "3. Arnold, E., Wynn, J., Vicente, S., Garcia-Hernando, G., Monszpart, A., Prisacariu, V., Turmukhambetov, D., Brachmann, E.: Map-free visual relocalization: Metric pose relative to a single image. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.379, + 0.785, + 0.406 + ], + "angle": 0, + "content": "4. Balntas, V., Li, S., Prisacariu, V.: Relocnet: Continuous metric learning relocalisation using neural nets. In: ECCV (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.407, + 0.785, + 0.434 + ], + "angle": 0, + "content": "5. Bay, H., Tuytelaars, T., Van Gool, L.: Surf: Speeded up robust features. In: ECCV (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.435, + 0.785, + 0.462 + ], + "angle": 0, + "content": "6. Bhowmik, A., Gumhold, S., Rother, C., Brachmann, E.: Reinforced feature points: Optimizing feature detection and description for a high-level task. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.463, + 0.785, + 0.476 + ], + "angle": 0, + "content": "7. Bradski, G.: The OpenCV Library. Dr. Dobb's Journal of Software Tools (2000)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.477, + 0.785, + 0.504 + ], + "angle": 0, + "content": "8. Cai, R., Hariharan, B., Snavely, N., Averbuch-Elor, H.: Extreme rotation estimation using dense correlation volumes. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.505, + 0.785, + 0.545 + ], + "angle": 0, + "content": "9. Calli, B., Singh, A., Walsman, A., Srinivasa, S., Abbeel, P., Dollar, A.M.: The ycb object and model set: Towards common benchmarks for manipulation research. In: ICAR (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.547, + 0.785, + 0.588 + ], + "angle": 0, + "content": "10. Chang, A., Dai, A., Funkhouser, T., Halber, M., Niebner, M., Savva, M., Song, S., Zeng, A., Zhang, Y.: Matterport3d: Learning from rgb-d data in indoor environments. In: 3DV (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.589, + 0.785, + 0.616 + ], + "angle": 0, + "content": "11. Chang, J., Yu, J., Zhang, T.: Structured epipolar matcher for local feature matching. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.617, + 0.785, + 0.644 + ], + "angle": 0, + "content": "12. Chen, H., Luo, Z., Zhang, J., Zhou, L., Bai, X., Hu, Z., Tai, C.L., Quan, L.: Learning to match features with seeded graph matching network. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.645, + 0.785, + 0.685 + ], + "angle": 0, + "content": "13. Chen, H., Luo, Z., Zhou, L., Tian, Y., Zhen, M., Fang, T., McKinnon, D., Tsin, Y., Quan, L.: Aspanformer: Detector-free image matching with adaptive span transformer. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.687, + 0.785, + 0.714 + ], + "angle": 0, + "content": "14. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: CVPR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.715, + 0.785, + 0.742 + ], + "angle": 0, + "content": "15. DeTone, D., Malisiewicz, T., Rabinovich, A.: Toward geometric deep slam. arXiv preprint arXiv:1707.07410 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.743, + 0.785, + 0.771 + ], + "angle": 0, + "content": "16. DeTone, D., Malisiewicz, T., Rabinovich, A.: Superpoint: Self-supervised interest point detection and description. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.772, + 0.785, + 0.812 + ], + "angle": 0, + "content": "17. Dusmanu, M., Rocco, I., Pajdla, T., Pollefeys, M., Sivic, J., Torii, A., Sattler, T.: D2-net: A trainable cnn for joint description and detection of local features. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "18. Edstedt, J., Athanasiadis, I., Wadenback, M., Felsberg, M.: Dkm: Dense kernelized feature matching for geometry estimation. In: CVPR (2023)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.279, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.357, + 0.127 + ], + "angle": 0, + "content": "R. Yin et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "19. Edstedt, J., Sun, Q., Bokman, G., Wadenback, M., Felsberg, M.: RoMa: Robust Dense Feature Matching. arXiv preprint arXiv:2305.15404 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.785, + 0.215 + ], + "angle": 0, + "content": "20. Eggert, D.W., Lorusso, A., Fisher, R.B.: Estimating 3-d rigid body transformations: a comparison of four major algorithms. Machine vision and applications (1997)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.217, + 0.785, + 0.242 + ], + "angle": 0, + "content": "21. En, S., Lechery, A., Jurie, F.: Rpnet: An end-to-end network for relative camera pose estimation. In: ECCV (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.243, + 0.785, + 0.284 + ], + "angle": 0, + "content": "22. Fan, Z., Pan, P., Wang, P., Jiang, Y., Xu, D., Jiang, H., Wang, Z.: Pope: 6-dof promptable pose estimation of any object, in any scene, with one reference. arXiv preprint arXiv:2305.15727 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.285, + 0.785, + 0.324 + ], + "angle": 0, + "content": "23. Fischler, M.A., Bolles, R.C.: Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. COMMUN ACM (1981)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.326, + 0.785, + 0.338 + ], + "angle": 0, + "content": "24. Gleize, P., Wang, W., Feiszli, M.: Silk: Simple learned keypoints. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.339, + 0.785, + 0.365 + ], + "angle": 0, + "content": "25. Goodwin, W., Vaze, S., Havoutis, I., Posner, I.: Zero-shot category-level object pose estimation. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.367, + 0.785, + 0.392 + ], + "angle": 0, + "content": "26. Hampali, S., Rad, M., Oberweger, M., Lepetit, V.: Honnotate: A method for 3d annotation of hand and object poses. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.393, + 0.699, + 0.405 + ], + "angle": 0, + "content": "27. Hartley, R.I.: In defense of the eight-point algorithm. TPAMI (1997)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.407, + 0.785, + 0.432 + ], + "angle": 0, + "content": "28. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.434, + 0.785, + 0.459 + ], + "angle": 0, + "content": "29. He, X., Sun, J., Wang, Y., Huang, D., Bao, H., Zhou, X.: Onepose++: Keypoint-free one-shot object pose estimation without cad models. NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.46, + 0.785, + 0.5 + ], + "angle": 0, + "content": "30. Hinterstoisser, S., Lepetit, V., Ilic, S., Holzer, S., Bradski, G., Konolige, K., Navab, N.: Model based training, detection and pose estimation of texture-less 3d objects in heavily cluttered scenes. In: ACCV (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.785, + 0.527 + ], + "angle": 0, + "content": "31. Huber, P.J.: Robust estimation of a location parameter. The Annals of Mathematical Statistics (1964)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.528, + 0.785, + 0.554 + ], + "angle": 0, + "content": "32. Jin, L., Qian, S., Owens, A., Fouhey, D.F.: Planar surface reconstruction from sparse views. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.555, + 0.785, + 0.595 + ], + "angle": 0, + "content": "33. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.597, + 0.785, + 0.623 + ], + "angle": 0, + "content": "34. Laskar, Z., Melekhov, I., Kalia, S., Kannala, J.: Camera relocalization by computing pairwise relative poses using convolutional neural network. In: ICCVW (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.624, + 0.785, + 0.649 + ], + "angle": 0, + "content": "35. Li, K., Wang, L., Liu, L., Ran, Q., Xu, K., Guo, Y.: Decoupling makes weakly supervised local feature better. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.65, + 0.785, + 0.676 + ], + "angle": 0, + "content": "36. Li, X., Han, K., Li, S., Prisacariu, V.: Dual-resolution correspondence networks. NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.677, + 0.785, + 0.703 + ], + "angle": 0, + "content": "37. Li, Z., Snavely, N.: Megadepth: Learning single-view depth prediction from internet photos. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.704, + 0.785, + 0.744 + ], + "angle": 0, + "content": "38. Lin, A., Zhang, J.Y., Ramanan, D., Tulsiani, S.: Relpose++: Recovering 6d poses from sparse-view observations. In: 2024 International Conference on 3D Vision (3DV). pp. 106-115. IEEE (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.745, + 0.785, + 0.771 + ], + "angle": 0, + "content": "39. Lindenberger, P., Sarlin, P.E., Pollefeys, M.: LightGlue: Local Feature Matching at Light Speed. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.799 + ], + "angle": 0, + "content": "40. Liu, Y., Shen, Z., Lin, Z., Peng, S., Bao, H., Zhou, X.: Gift: Learning transformation-invariant dense visual descriptors via group cnns. NeurIPS (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "41. Liu, Y., Wen, Y., Peng, S., Lin, C., Long, X., Komura, T., Wang, W.: Gen6d: Generalizable model-free 6-dof object pose estimation from rgb images. In: ECCV (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.282, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SRPose: Two-view Relative Pose Estimation with Sparse Keypoints" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "42. Longuet-Higgins, H.C.: A computer algorithm for reconstructing a scene from two projections. Nature (1981)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.19 + ], + "angle": 0, + "content": "43. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: ICLR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.191, + 0.785, + 0.204 + ], + "angle": 0, + "content": "44. Lowe, D.G.: Distinctive image features from scale-invariant keypoints. IJCV (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.205, + 0.785, + 0.245 + ], + "angle": 0, + "content": "45. Luo, Z., Zhou, L., Bai, X., Chen, H., Zhang, J., Yao, Y., Li, S., Fang, T., Quan, L.: Aslfeat: Learning local features of accurate shape and localization. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.246, + 0.785, + 0.273 + ], + "angle": 0, + "content": "46. Melekhov, I., Ylioinas, J., Kannala, J., Rahtu, E.: Relative camera pose estimation using convolutional neural networks. In: ACIVS (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.274, + 0.785, + 0.3 + ], + "angle": 0, + "content": "47. Ni, J., Li, Y., Huang, Z., Li, H., Bao, H., Cui, Z., Zhang, G.: Pats: Patch area transportation with subdivision for local feature matching. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.301, + 0.785, + 0.327 + ], + "angle": 0, + "content": "48. Nister, D.: An efficient solution to the five-point relative pose problem. TPAMI (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.328, + 0.785, + 0.37 + ], + "angle": 0, + "content": "49. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.371, + 0.785, + 0.397 + ], + "angle": 0, + "content": "50. Phil, W.: Bidirectional cross attention. https://github.com/lucidrains/bidirectional-cross-attention (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.398, + 0.785, + 0.438 + ], + "angle": 0, + "content": "51. Ranftl, R., Lasinger, K., Hafner, D., Schindler, K., Koltun, V.: Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. TPAMI (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.439, + 0.785, + 0.466 + ], + "angle": 0, + "content": "52. Revaud, J., Weinzaepfel, P., de Souza, C.R., Humenberger, M.: R2D2: repeatable and reliable detector and descriptor. In: NeurIPS (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.467, + 0.785, + 0.494 + ], + "angle": 0, + "content": "53. Rockwell, C., Johnson, J., Fouhey, D.F.: The 8-point algorithm as an inductive bias for relative pose prediction by vits. In: 3DV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.495, + 0.785, + 0.522 + ], + "angle": 0, + "content": "54. Rublee, E., Rabaud, V., Konolige, K., Bradski, G.: Orb: An efficient alternative to sift or surf. In: ICCV (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.523, + 0.785, + 0.55 + ], + "angle": 0, + "content": "55. Sarlin, P.E., DeTone, D., Malisiewicz, T., Rabinovich, A.: SuperGlue: Learning feature matching with graph neural networks. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.551, + 0.785, + 0.59 + ], + "angle": 0, + "content": "56. Shi, Y., Cai, J.X., Shavit, Y., Mu, T.J., Feng, W., Zhang, K.: Clustergnn: Cluster-based coarse-to-fine graph neural network for efficient feature matching. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.591, + 0.785, + 0.646 + ], + "angle": 0, + "content": "57. Sinha, S., Zhang, J.Y., Tagliasacchi, A., Gilitschenski, I., Lindell, D.B.: Sparse-pose: Sparse-view camera pose regression and refinement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21349–21359 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.647, + 0.785, + 0.688 + ], + "angle": 0, + "content": "58. Smith, L.N., Topin, N.: Super-convergence: Very fast training of neural networks using large learning rates. In: Artificial intelligence and machine learning for multidomain operations applications (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.689, + 0.785, + 0.716 + ], + "angle": 0, + "content": "59. Sun, J., Shen, Z., Wang, Y., Bao, H., Zhou, X.: LoFTR: Detector-free local feature matching with transformers. CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.717, + 0.785, + 0.743 + ], + "angle": 0, + "content": "60. Sun, J., Wang, Z., Zhang, S., He, X., Zhao, H., Zhang, G., Zhou, X.: One-positron object pose estimation without cad models. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.744, + 0.785, + 0.785 + ], + "angle": 0, + "content": "61. Tan, B., Xue, N., Wu, T., Xia, G.S.: Nope-sac: Neural one-plane ransac for sparse-view planar 3d reconstruction. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.786, + 0.785, + 0.813 + ], + "angle": 0, + "content": "62. Tang, S., Zhang, J., Zhu, S., Tan, P.: Quadtree attention for vision transformers. ICLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "63. Tyszzkiewicz, M., Fua, P., Trulls, E.: Disk: Learning local features with policy gradient. NeurIPS (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.357, + 0.127 + ], + "angle": 0, + "content": "R. Yin et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "64. Wang, J., Rupprecht, C., Novotny, D.: Posediffusion: Solving pose estimation via diffusion-aided bundle adjustment. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9773-9783 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.217 + ], + "angle": 0, + "content": "65. Wang, Q., Zhang, J., Yang, K., Peng, K., Stiefelhagen, R.: Matchformer: Interleaving attention in transformers for feature matching. In: ACCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.244 + ], + "angle": 0, + "content": "66. Wen, B., Bekris, K.: Bundletrack: 6d pose tracking for novel objects without instance or category-level 3d models. In: IROS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.785, + 0.286 + ], + "angle": 0, + "content": "67. Wen, B., Tremblay, J., Blukis, V., Tyree, S., Müller, T., Evans, A., Fox, D., Kautz, J., Birchfield, S.: Bundlesdf: Neural 6-dof tracking and 3d reconstruction of unknown objects. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.314 + ], + "angle": 0, + "content": "68. Winkelbauer, D., Denninger, M., Triebel, R.: Learning to localize in new environments from synthetic training data. In: ICRA (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.355 + ], + "angle": 0, + "content": "69. Xiang, Y., Schmidt, T., Narayanan, V., Fox, D.: PoseCNN: A convolutional neural network for 6d object pose estimation in cluttered scenes. Robotics: Science and Systems XIV (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.785, + 0.383 + ], + "angle": 0, + "content": "70. Xue, F., Budvytis, I., Cipolla, R.: Imp: Iterative matching and pose estimation with adaptive pooling. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.411 + ], + "angle": 0, + "content": "71. Xue, F., Budvytis, I., Cipolla, R.: Sfd2: Semantic-guided feature detection and description. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.785, + 0.438 + ], + "angle": 0, + "content": "72. Yi, K.M., Trulls, E., Lepetit, V., Fua, P.: Lift: Learned invariant feature transform. In: ECCV (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.466 + ], + "angle": 0, + "content": "73. Yu, J., Chang, J., He, J., Zhang, T., Yu, J., Wu, F.: Adaptive spot-guided transformer for consistent local feature matching. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.508 + ], + "angle": 0, + "content": "74. Zhang, J.Y., Ramanan, D., Tulsiani, S.: Relpose: Predicting probabilistic relative rotation for single objects in the wild. In: European Conference on Computer Vision. pp. 592-611. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.508, + 0.785, + 0.536 + ], + "angle": 0, + "content": "75. Zhao, X., Wu, X., Chen, W., Chen, P.C.Y., Xu, Q., Li, Z.: Aliked: A lighter keypoint and descriptor extraction network via deformable transformation. TIM (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.785, + 0.563 + ], + "angle": 0, + "content": "76. Zhao, X., Wu, X., Miao, J., Chen, W., Chen, P.C., Li, Z.: Alike: Accurate and lightweight keypoint detection and descriptor extraction. TMM (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.785, + 0.591 + ], + "angle": 0, + "content": "77. Zhou, Q., Sattler, T., Pollefeys, M., Leal-Taixe, L.: To learn or not to learn: Visual localization from essential matrices. In: ICRA (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.591, + 0.785, + 0.619 + ], + "angle": 0, + "content": "78. Zhou, Y., Barnes, C., Lu, J., Yang, J., Li, H.: On the continuity of rotation representations in neural networks. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.785, + 0.647 + ], + "angle": 0, + "content": "79. Zhu, S., Liu, X.: Pmatch: Paired masked image modeling for dense geometric matching. In: CVPR (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.647 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/SRPose_ Two-view Relative Pose Estimation with Sparse Keypoints/e429b1d6-1778-4297-b226-264e6bf36452_origin.pdf b/2024/SRPose_ Two-view Relative Pose Estimation with Sparse Keypoints/e429b1d6-1778-4297-b226-264e6bf36452_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2d4fdd43370d717fc90447675182c6b3a01b64c6 --- /dev/null +++ b/2024/SRPose_ Two-view Relative Pose Estimation with Sparse Keypoints/e429b1d6-1778-4297-b226-264e6bf36452_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2bb760fb9427bc685feecdcc263efd7792fb5b84e0f5afc762fb0423adf52af6 +size 10653895 diff --git a/2024/SRPose_ Two-view Relative Pose Estimation with Sparse Keypoints/full.md b/2024/SRPose_ Two-view Relative Pose Estimation with Sparse Keypoints/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b7d249886f2da74686de126117ab80b8fff5db4c --- /dev/null +++ b/2024/SRPose_ Two-view Relative Pose Estimation with Sparse Keypoints/full.md @@ -0,0 +1,392 @@ +# SRPose: Two-view Relative Pose Estimation with Sparse Keypoints + +Rui Yin $^{1}$ , Yulun Zhang $^{2}$ , Zherong Pan $^{3}$ , Jianjun Zhu $^{1}$ , Cheng Wang $^{1}$ , and Biao Jia $^{1^{\star}}$ + +$^{1}$ Hanglok-Tech, China + $^{2}$ MoE Key Lab of Artificial Intelligence, Shanghai Jiao Tong University, China + $^{3}$ LightSpeed Studios, Tencent America + +Abstract. Two-view pose estimation is essential for map-free visual relocalization and object pose tracking tasks. However, traditional matching methods suffer from time-consuming robust estimators, while deep learning-based pose regressors only cater to camera-to-world pose estimation, lacking generalizability to different image sizes and camera intrinsics. In this paper, we propose SRPose, a sparse keypoint-based framework for two-view relative pose estimation in camera-to-world and object-to-camera scenarios. SRPose consists of a sparse keypoint detector, an intrinsic-calibration position encoder, and promptable prior knowledge-guided attention layers. Given two RGB images of a fixed scene or a moving object, SRPose estimates the relative camera or 6D object pose transformation. Extensive experiments demonstrate that SRPose achieves competitive or superior performance compared to state-of-the-art methods in terms of accuracy and speed, showing generalizability to both scenarios. It is robust to different image sizes and camera intrinsics, and can be deployed with low computing resources. Project page: https://frickyinn.github.io/srpose. + +Keywords: Relative Pose Estimation $\cdot$ 6D Object Pose Estimation + +# 1 Introduction + +Relative pose estimation between two images plays a crucial role in many 3D vision tasks, including visual odometry and map-free visual relocalization [3], 6D object pose tracking [66,67], etc. These tasks can be categorized into two scenarios: camera-to-world and object-to-camera estimation. In the camera-to-world scenario, such as in map-free visual relocalization, we estimate the pose transformation from the reference image to the query image, predicting the change in camera extrinsic between the two views. In the object-to-camera scenario, taking object pose tracking for example, we aim to track the 6D pose of the target object in the videos by estimating its relative transformation in the camera coordinate system between two adjacent frames. While these two scenarios may + +![](images/57982706270fbec43987e0e4adc8405d20f72effda6bd14de5d15e8045a2bd53.jpg) +(a) Reference + +![](images/dd0be0ec8e50da40d3a979a4bdf49612a77eab57bbc12296dd31468d2f3e82a4.jpg) +(b) Query + +![](images/fc65be99503e69315faa02432f7da2197878de4c04f541f27c4a8c55fdfee04e.jpg) +(c) Ground Truth + +![](images/1c61be74d7c10000943d1aa9116753914ccd4d0124232d6a2992b21506ce517a.jpg) +(d) Reference + +![](images/a891af73eec98b5499f70449c883f9e7b126b56bce20048dd270102aa4bf64ca.jpg) +(e) Query + +![](images/3205b13a541666eaaf8e29d45dd60f8c971be081b889f0990944615ca73331ce.jpg) +(f) Ground Truth +Fig. 1: Relative pose estimation by SRPose. Dots drawn in the figures visualize the cross-attention scores of sparse keypoints across the two images, with brighter dots representing higher attention. Camera-to-world: (a), (b), (c) visualize the epipolar lines, representing the connections between the nine corresponding points across two views. Higher attention is shown to the overlap of the scenes. Object-to-Camera: (d), (e), (f) show the relative 6D pose estimation in the query image with only one accessible object prompt in the reference image. Higher attention is shown to the target object. SRPose establishes implicit correspondences. + +seem distinct, they both entail the prediction of the pose matrix from the two RGB images depicting the scene or object, which in turn represents the relative pose transformation from the first image to the second. + +Traditional pose estimation approaches typically involve detecting and matching keypoints or local features in the two images to establish point-to-point correspondences, then use a robust estimator to denoise the outliers and recover the relative pose from the essential matrix by solving a two-view geometry problem, i.e., the epipolar constraint equation [48]. These approaches produce accurate results regardless of image sizes and camera intrinsicics, thanks to the intrinsic calibration performed before solving the equation. Given the segmentation of the target object in the two images, these approaches can also yield acceptable predictions of the relative object pose transformation. However, on the downside, robust estimators can be computationally expensive. Indeed, the process of eliminating mismatched correspondences and solving the equations using robust estimators is significantly slower compared to detecting and matching keypoints or features, making it prohibitive in the real-time application. Moreover, real-time object pose tracking requires video object segmentation models to eliminate off-target keypoints [66, 67], which can also introduce additional overhead. + +The maturity of deep learning offers the advantage of directly regressing the relative pose transformation from two RGB inputs, significantly boosting the runtime performance. However, existing regression methods lack of generalizability or precision to images with varying sizes and camera intrinsics [3, 38, 53, 64, 74, 77]. They typically employ image encoders that only accept + +fixed-size inputs during batched training. And unlike traditional approaches, most deep learning-based regressors lack awareness of the original camera intrinsic parameters, unable to utilize two-view geometry to achieve higher precision in pose estimation. Furthermore, state-of-the-art deep regressors only apply to camera-to-world pose estimation, despite the common solution shared by both the camera-to-world and object-to-camera tasks. + +To address the aforementioned challenges, we propose SRPose: a Sparse keypoint-based framework for Relative Pose estimation. SRPose estimates the relative pose matrix based on two-view geometry by implicitly solving the epipolar constraint, as shown in Fig. 1. First, SRPose employs a point detector to extract sparse keypoints and associated descriptors to form the candidate set of implicit correspondences. Then, we use an intrinsic-calibration (IC) position encoder to modulate the keypoints before it computes the position embeddings, accounting for varying image sizes and camera intrinsics. The keypoint descriptors and the position embeddings are then fused into the prior knowledge-guided attention layers, which leverage the prior knowledge of keypoint similarities to establish implicit cross-view correspondences. With the mechanism of the promptable cross-attention in these layers, SRPose only requires an accessible object prompt in one of the two views to compute the relative 6D pose transformation of the target object, eliminating the need for elusive object segmentations. + +The main contribution of this paper can be summarized as follows: + +- We propose a novel framework using sparse keypoints, SRPose, for two-view relative pose estimation. To the best of our knowledge, it is the first attempt to directly regress relative poses from sparse keypoints for this task. +- We introduce an intrinsic-calibration position encoder to adapt SRPose to different image sizes and camera intrinsics. +- We enable object-to-camera estimation in addition to camera-to-world estimation by utilizing one accessible object prompt. +- Our SRPose greatly reduces estimation time by replacing robust estimators with direct regression, achieving state-of-the-art performance in accuracy and speed in camera-to-world and object-to-camera scenarios. + +# 2 Related works + +Relative Pose Estimation: In relative pose estimation, traditional approaches recover the rotation and translation from the essential matrix [27, 42]. By using the epipolar constraint [48], the essential matrix is estimated based on the cross-view correspondences established through the matching of sparse keypoints or dense features. Using this pipeline, sparse keypoint-based approaches typically start by detecting sparse keypoints and their associated descriptors, including classic hand-crafted detectors [5, 44, 54], and deep learning-based detectors [6, 15-17, 24, 35, 40, 45, 52, 63, 72, 75, 76]. Recent methods further employ deep learning-based matchers [12, 39, 55, 56, 70] to establish keypoint matches, while classical approaches rely on nearest neighbor search. In contrast to using sparse keypoints, dense matchers [11, 13, 18, 19, 36, 47, 59, 62, 65, 71, 73, 79] are detector-free and perform pixel-wise dense feature matching. Once the correspondences + +are established through matching, a robust estimator such as RANSAC [23] is employed to estimate the essential matrix and recover the relative pose. Overall, while the traditional matcher-based approaches have shown promising results, they still face challenges with their time-consuming robust estimators. Furthermore, dense matchers, while providing exceptional accuracy, may suffer from slow performance due to their resource-intensive design. + +Relative Pose Regression: Deep learning regressors offer an alternative approach by directly predicting the relative pose using dedicated neural networks, without explicitly establishing correspondences. Two-view relative pose estimation focuses on frame-to-frame pose transformations, aiming at visual odometry and map-free relocalization [1,3,4,21,34,46,68]. Leveraging prior knowledge is an important aspect in two-view estimation. For instance, SparsePlanes [32], PlaneFormer [8] and NOPE-SAC [61] utilize geometry relations of the planar surfaces to enhance performance. Unfortunately, these approaches rely on image encoders designed for images of fixed sizes and intrinsics, lacking generalizability across different settings. Sparse-view relative pose estimation emphasizes global pose optimization or end-to-end SfM involving multiple views [38,57,64,74]. Although these sparse-view methods can be extended to two-view estimation, and may even predict unknown intrinsics, their adaptability to varying camera settings depends on multi-view information. In contrast, SRPose leverages sparse keypoints and two-view geometry to achieve higher precision and generalizability across different image sizes and camera intrinsics. + +Object Pose Estimation: While most of the aforementioned approaches focus on camera pose estimation in a fixed scene, some tasks, such as 6D object pose tracking, require relative object pose estimation as well. Most frameworks predict the relative object pose using the same pipeline of keypoint correspondences. BundleTrack [66] and BundleSDF [67] track the object poses in videos by matching keypoints between two frames, and conducting global pose graph optimization. POPE [22] combines advanced foundation models of segmentation [33] and matching [49] to enable zero-shot object pose estimation. Robust as these classic methods may be, they require dedicated object masks to focus on the objects, which are provided by other resource-consuming segmentation models. To get rid of object segmentation, OnePose [60], OnePose++ [29], and Gen6D [41] rely on multiple views of the object, and ZeroShot [25] establishes semantic correspondences to mute the background, and employs depth map to enable pose estimation. Nonetheless, these methods require additional high-quality information, and they restrict the querying view to only accommodate a single object. As a remedy, SRPose employs an accessible object prompt to focus on the target object, enabling effective depth-free and mask-free two-view relative object estimation with superior performance in the object-to-camera scenario. + +# 3 Method + +Given two images $I_1, I_2$ , SRPose estimates the relative pose consisting of a rotation $R \in \mathcal{SO}(3)$ , and a translation $t \in \mathbb{R}^3$ based on the two-view geometry. Detailed problem definitions can be found in the supplementary materials. Traditional approaches use the epipolar constraint [48] to estimate the following + +![](images/689444cd282daa516213b93b0c16f3427b7984e595ca9f7796569d6f21b6f37d.jpg) +Fig. 2: Overview. SRPose comprises four main components: 1) The sparse keypoint detector detects keypoints associated with descriptors separately sfrom the two images; 2) The intrinsic-calibration (IC) position encoder modulates the keypoints' coordinates with camera intrinsics, and encodes their position information; 3) Guided by the prior knowledge of keypoint similarities, along with the object prompt, the attention layers establish implicit cross-view correspondences; 4) The regressor estimates relative pose $R, t$ under the constraints of implicit correspondences. + +essential matrix $E$ , from which $R, t$ are recovered: + +$$ +\left(K _ {1} ^ {- 1} q _ {1}\right) ^ {\top} E K _ {2} ^ {- 1} q _ {2} = 0, \quad \forall \left(q _ {1}, q _ {2}\right), \tag {1} +$$ + +$$ +E = [ t ] _ {\times} R, \tag {2} +$$ + +where $q_{1}, q_{2}$ are two arbitrary corresponding points in $I_{1}, I_{2}$ , and $K_{1}, K_{2}$ are the camera intrinsics of the two images. SRPose builds and solves the constraint equation Eq. (1) with neural networks implicitly. With SRPose denoted as $\mathcal{SR}$ , the function of our framework is represented as: + +$$ +[ R | t ] = \mathcal {S R} (I _ {1}, I _ {2}). \tag {3} +$$ + +Figure 2 summarizes SRPose, involving four components: a sparse keypoint detector, an intrinsic-calibration position encoder, the promptable prior knowledge-guided attention layers, and an MLP regressor. + +# 3.1 Sparse Keypoint Detector + +Instead of employing an image encoder to extract image features, we use a detector to extract sparse keypoints and associated descriptors from the two images. To this end, we can use one of the existing techniques developed over the years, which include classic methods such as SIFT [44], and deep learning-based methods such as SuperPoint [16], ALIKE [76], DISK [63], etc. Given an image $I$ , our detector, denoted as $\mathcal{F}^{kd}$ , detects keypoints consisting of $N$ coordinates $P^k \subset \mathbb{R}^2$ in image space, and associated descriptors $D \subset \mathbb{R}^d$ , describing the local features in the image, where $d$ denotes the dimension of the keypoints descriptors in the network. To summarize, our detector is defined as the following function: + +$$ +\left\{P ^ {k}, D \right\} = \mathcal {F} ^ {k d} (I). \tag {4} +$$ + +In practice, we resize all images to a standardized size to enable parallel batched detection, then rescale the keypoints' coordinates $P^k$ to their original positions. The sparse keypoint detector yields two sets of keypoints $P_1^k, P_2^k$ from $I_1, I_2$ , from which a set of correspondences $\{(q_1, q_2)\} \subset P_1^k \times P_2^k$ will be established implicitly later for solving the constraint Eq. (1). + +# 3.2 Intrinsic-Calibration (IC) Position Encoder + +According to Eq. (1), for the keypoints detected from an image $I$ , their coordinates should be first calibrated or normalized by the camera intrinsics $K$ : + +$$ +P ^ {c} = K ^ {- 1} P ^ {k}. \tag {5} +$$ + +Enforcing the intrinsic calibration (IC) offers a crucial improvement of both the accuracy and robustness [48]. IC allows SRPose to adapt images captured by different cameras with varying intrinsic parameters. Previous regressors [3, 53] extract implicit image features with image encoders trained on fixed-size images, assuming the camera intrinsics of all inputs are the same by default. While the sparse keypoints in SRPose are modulated by intrinsic calibration, which provides the accurate positions, i.e. $P^c$ , of all keypoints in a unified camera coordinate system across different intrinsics. + +After the intrinsic calibration, the keypoints' coordinates are encoded to address the keypoints based on their positions. Due to the possible position distortion caused by relative position encoding [39, 59], we employ absolute position encoding that maintains the precise coordinate values, which is crucial to the two-view geometry problems. With a single-layer fully connected network, denoted as the following function $\mathcal{F}^{pe}$ , the position encoder maps 2D calibrated coordinates to high-dimension position embeddings: + +$$ +P ^ {e} = \mathcal {F} ^ {p e} \left(K ^ {- 1} P ^ {k}\right). \tag {6} +$$ + +The resulting encoded position embeddings, denoted as $P^e$ , are then incorporated with the keypoint descriptors and descriptor embeddings, denoted as $D \oplus P^e$ , which provides calibrated and normalized position information for solving the constraint Eq. (1) using a set of attention layers. + +# 3.3 Promptable Prior-Knowledge-Guided Attention + +Given the position embeddings and keypoint descriptors, we use a multi-layer attention network to exploit semantic information and establish implicit correspondences. Our network consists of $M$ consecutive attention layers, with each layer comprising a self-attention module and a promptable prior knowledge-guided cross-attention module. We denote $X^{m} \subset \mathbb{R}^{d}$ as the keypoint descriptor embeddings being processed in the $m^{th}$ layer, in which $m \in M$ , with $X^{m}$ being either $X_{1}^{m}$ or $X_{2}^{m}$ from the image $I_{1}$ or $I_{2}$ . Each layer takes $X^{m-1}$ as inputs, then computes the deeper embeddings $X^{m}$ . In summary, the entire multi-layer network inputs keypoint descriptors $D_{1}D_{2}$ , and outputs $X_{1}^{M}, X_{2}^{M}$ after $M$ layers, i.e., we define $D$ as the initial state of embeddings $X^{0}$ . Prior to each attention layer, $X_{1}^{m-1}, X_{2}^{m-1}$ are incorporated with the associated position embeddings $P_{1}^{e}, P_{2}^{e}$ . Following [28], residual connections are applied to every self- and cross-attention module. The architecture of the attention layers is outlined in Fig. 3. + +Self-Attention: The self-attention module captures the comprehensive semantic information of all keypoints within an entire image. In this module, $X^{m - 1} \oplus P^e$ of both the images is first mapped to three vectors by a linear transformation $W_{s}^{m}$ , denoted as the query $q$ , the key $k$ , and value $v$ respectively, where + +![](images/bf8095184d857ed796812d2f567ac18f8cefee0318e694ba40bb390b968be5fc.jpg) +Fig. 3: Overview of the prior knowledge-guided attention layers. Each layer contains a self-attention and a cross-attention module. A similarity matrix $S$ of the keypoint descriptors is utilized as the prior knowledge to readjust the cross-attention scores, guiding more attention to cross-view keypoint pairs with implicit correspondences. The object prompt and the residual connections are omitted in the figure. + +$q, k, v \subset \mathbb{R}^d$ . Then $k$ and $q$ are used to compute the self-attention scores $A_s$ , which are then scaled by the dimension of descriptors and embeddings $d$ : + +$$ +[ q | k | v ] = W _ {s} ^ {M} \left(X ^ {m - 1} \oplus P ^ {e}\right), \tag {7} +$$ + +$$ +A _ {s} = \frac {q k ^ {\top}}{\sqrt {d}}. \tag {8} +$$ + +After being normalized by the Softmax function, the self-attention scores $A_{s}$ are then used to extract the highly attended information from $v$ : + +$$ +X _ {s} ^ {m - 1} = \operatorname {S o f t m a x} \left(A _ {s}\right) v. \tag {9} +$$ + +In the self-attention module, keypoint position embeddings and descriptor embeddings are computed together to summarize the semantic information within one entire image before deeper cross-view correspondences are established. + +Prior Knowledge-guided Cross-Attention: Guided by the prior knowledge of keypoint similarities, the cross-attention module exploits the mutual information across the two images to establish implicit correspondences. Similar to the self-attention, cross-attention maps the input embeddings $X_{s}^{m - 1}$ to two keys $k_{1},k_{2}$ and values $v_{1},v_{2}$ of the two images respectively through another linear transformation $W_{c}^{m}$ layer. Following [50], we compute the bidirectional attention scores across the two images from the resulting embeddings of the previous self-attention module, using the cross-view keys $k_{1},k_{2}$ : + +$$ +\left[ k _ {1} \mid v _ {1} \right] = W _ {c} ^ {m} X _ {s, 1} ^ {m - 1}, \tag {10} +$$ + +$$ +\left[ k _ {2} \mid v _ {2} \right] = W _ {c} ^ {m} X _ {s, 2} ^ {m - 1}, \tag {11} +$$ + +$$ +A _ {2 1} ^ {\top} = A _ {1 2} = \frac {k _ {1} k _ {2} ^ {\top}}{\sqrt {d}}. \tag {12} +$$ + +The module recognizes the overlapping areas of the scene or object between the two views by identifying the mutuality between two sets of keypoint embeddings from the computed cross-attention scores. To facilitate the process, prior + +knowledge is used to guide and enhance the attention on highly correlated keypoints, by adjusting the cross-attention scores with a keypoint similarity matrix $S \in \mathbb{R}^{N_1 \times N_2}$ . Each entry $s_{ij}$ of the matrix is computed as the cosine similarity between the descriptors of every cross-view keypoint pair: + +$$ +s _ {i j} = \left(\frac {d _ {i} \cdot d _ {j}}{\| d _ {i} \| \| d _ {j} \|} + 1\right) / 2, \quad \{(d _ {i}, d _ {j}) \} \subseteq D _ {1} \times D _ {2}. \tag {13} +$$ + +Keypoints with similar local features tend to possess similar descriptors, describing correlated semantic information of similar views. Thus, the cosine similarity captures the semantic similarities between two keypoints, and guides the framework to recognize the overlapping parts between the images. + +We normalize the matrix $S$ to [0,1] and element-wise multiply it with the cross-attention scores. Finally, the scores are used to extract and exchange crucial cross-view information from each others' values $v_{1}, v_{2}$ : + +$$ +A _ {2 1} ^ {\prime \top} = A _ {1 2} ^ {\prime} = A _ {1 2} \cdot S, \tag {14} +$$ + +$$ +X _ {1} ^ {m - 1} = \operatorname {S o f t m a x} \left(A _ {1 2} ^ {\prime}\right) v _ {2}, \tag {15} +$$ + +$$ +X _ {2} ^ {m - 1} = \operatorname {S o f t m a x} \left(A _ {2 1} ^ {\prime}\right) v _ {1}. \tag {16} +$$ + +By capturing and incorporating relevant information from each image into the other, the implicit correspondences are established across the two images. As shown in Figs. 1a and 1b, with brighter dots representing higher cross-attention scores, SRPose establishes implicitly corresponding points densely in the overlapping areas of the two images, which is utilized to construct and solve an implicit version of the epipolar constraint Eq. (1). + +Object Prompt: In the object-to-camera estimation, we employ an accessible user-provided object prompt in only one image to identify the target object. In contrast, traditional matching approaches require high-quality object segmentations in both images. An object prompt $b$ is a bounding box consisting of the two coordinates of top-left and bottom-right points, which bounds the object $o$ in one of the images $I_1$ , making $I_1$ as the reference view. Once the keypoints are extracted from $I_1$ , the object prompt removes the keypoints outside of the bounding box in the reference view, eliminating most irrelevant keypoints. Thus, the descriptors and positions of keypoints within bounds, focusing on the target object $o$ , will be later processed in the Similarity Matrix and other following modules. With the guided cross-attention modules, SRPose learns to identify highly correlated keypoints lying on the same object $o$ from the entire query view $I_2$ , excluding irrelevant background keypoints, as shown in Figs. 1d and 1e. With the object prompt, SRPose searches the implicit corresponding points on the target object and computes the relative 6D object pose transformation in the same process as the camera-to-world scenario. + +# 3.4 Relative Pose Regressor + +After the embeddings $X_1^M, X_2^M$ are computed via the previous modules, a MultiLayer Perceptron (MLP) is employed to regress the relative pose as our last step. + +The outputs of the previous attention layers $X_1^M, X_2^M$ first undergo 2D average pooling, and then are concatenated together, resulting in the joint embedding $f \in \mathbb{R}^{2d}$ , which carries the deep features and implicit correspondences across the two images. Finally, embedding $f$ is input to the MLP, and the regressor solves the epipolar constraint Eq. (1) implicitly and regresses the relative pose, consisting of a rotation $R$ , and a translation $t$ . Following [78], we set the rotation output of the regressor to be a 6d vector $r \in \mathbb{R}^6$ , which is then transformed into a rotation matrix $R \in \mathbb{R}^{3 \times 3}$ through a partial Gram-Schmidt process. This technique, as argued in [78], enables a continuous representation of rotations. + +# 3.5 Loss Function + +SRPose is trained in a supervised manner. With the ground-truth rotation $R_{gt}$ , we compute the Huber Loss [31] of the error in the rotation angle, which minimizes the angular error between the predicted and ground-truth rotations: + +$$ +\mathcal {L} _ {R} \left(I _ {1}, I _ {2}\right) = \mathcal {H} \left(\operatorname {a r c c o s} \left(\frac {\operatorname {T r} \left(R ^ {\top} R _ {g t} - 1\right)}{2}\right)\right), \tag {17} +$$ + +where $\mathcal{H}$ denotes the Huber Loss function. Similarly, given the ground-truth translation $t_{gt}$ , the error is computed in both normalized and unnormalized forms. To further enhance accuracy, we also incorporate the angular error of translation, resulting in the following three loss terms: + +$$ +\mathcal {L} _ {t} \left(I _ {1}, I _ {2}\right) = \mathcal {H} \left(t - t _ {g t}\right), \tag {18} +$$ + +$$ +\mathcal {L} _ {t _ {n}} \left(I _ {1}, I _ {2}\right) = \mathcal {H} \left(\frac {t}{\| t \|} - \frac {t _ {g t}}{\| t _ {g t} \|}\right), \tag {19} +$$ + +$$ +\mathcal {L} _ {t _ {a}} \left(I _ {1}, I _ {2}\right) = \mathcal {H} \left(\arccos \left(\frac {t \cdot t _ {g t}}{\| t \| \| t _ {g t} \|}\right)\right). \tag {20} +$$ + +The above three loss functions supervise both the scale and direction of translations. We compute the final loss by the weighted sum of the three functions: + +$$ +\mathcal {L} \left(I _ {1}, I _ {2}\right) = \mathcal {L} _ {R} + \lambda_ {t} \mathcal {L} _ {t} + \lambda_ {t _ {n}} \mathcal {L} _ {t _ {n}} + \lambda_ {t _ {a}} \mathcal {L} _ {t _ {a}}, \tag {21} +$$ + +where $\lambda_t, \lambda_{t_n}, \lambda_{t_a}$ are scalars used to balance the weights of four different losses. + +# 3.6 Implementation Details + +SRPose is first trained on ScanNet [14] and then fine-tuned on other datasets to achieve better performance. For all datasets, we resize the images to $640 \times 480$ , and extract 1,024 sparse keypoints using SuperPoint [16] for training and 2,048 for evaluation. Other sparse keypoint detectors can also be used for this step. We utilize the pre-trained SuperPoint provided by the official source and freeze its weights in the training stage. The remaining learnable modules are initialized with random weights. In practice, SRPose consists of 6 guided attention layers, with 4 attention heads in each attention module. The dimension $d$ of keypoint descriptors and embeddings is set to 256. The rotation and translation are estimated using two separate 3-layer MLPs. All three balancing scalars $\lambda_{t}, \lambda_{t_{n}}, \lambda_{t_{a}}$ + +Table 1: Relative Pose Estimation on Matterport [10]. Without depth, matcher-based approaches are incapable of scaled translation estimation, while SRPose achieves high accuracy in translation scales using only RGB inputs. + +
CategoryMethodRot. (°)Trans. (m)
Med.↓Avg.↓≤ 30° ↑Med.↓Avg.↓≤ 1m↑
SparseSuperGlue [55]3.8824.1777.8n/an/an/a
SGMNet [12]1.0621.3280.9n/an/an/a
LightGlue [39]1.3222.4580.0n/an/an/a
DenseLoFTR [59]0.7111.1190.5n/an/an/a
ASpanFormer [13]3.7331.4575.7n/an/an/a
DKM [18]0.4612.8989.0n/an/an/a
RegressorSparsePlanes [32]7.3322.7883.40.631.2566.6
PlaneFormers [2]5.9622.2083.80.661.1966.8
8point [53]8.0119.1385.40.641.0167.4
NOPE-SAC [61]2.7714.3789.00.520.9473.2
SRPose2.6511.1291.60.270.6183.7
+ +are set to 1. On ScanNet, our framework is trained using AdamW optimizer [43] and follows 1cycle learning rate policy [58] with a maximum learning rate of $1 \times 10^{-4}$ for 500 epochs. We fix the batch size to 32 for training. The pretraining requires 120 hours utilizing 8 RTX 3090 GPUs. More implementation details can be found in the supplementary materials. + +# 4 Experiments + +We evaluate the performance of SRPose for two-view relative pose estimation in both the camera-to-world and object-to-camera scenarios. Additionally, we assess its effectiveness in the application of map-free relocalization. Finally, we conduct an ablation study to analyze the roles of the components we proposed. + +# 4.1 Camera-to-World Pose Estimation + +Setup: In the camera-to-world scenario, we evaluate our framework's performance on Matterport [10] and ScanNet [14]. Following [32], for Matterport, we report the median and mean of the translation and rotation error, and the translation/rotation accuracy at the thresholds of $1\mathrm{m} / 30^{\circ}$ , respectively. For ScanNet, we adopt the metrics from [55] to compute the AUC of the pose error at $5^{\circ}$ , $10^{\circ}$ , and $20^{\circ}$ . The pose error here is the maximum angular error in rotation and translation. We also report an analysis of the time consumption on the ScanNet-1500 test set. For all matcher-based approaches, we use the implementation of RANSAC [23] from OpenCV [7] to recover relative poses. All methods are compared on a device with RTX 4090, i9-13900K, and 128GB memory. + +Baselines: We choose three categories of approaches as our baselines, including sparse keypoint-based matchers, dense matchers, and deep learning-based relative pose regressors. Results of all the baselines are reported according to the official papers or implementations. All sparse matchers and SRPose employ SuperPoint [16] as their sparse keypoint detector. All regressors are trained on the specific dataset, i.e. Matterport or ScanNet, before evaluation. LightGlue [39] is trained on MegaDepth [37], while all other matchers are trained on ScanNet. + +Table 2: Relative Pose Estimation on ScanNet [14]. SRPose outperforms state-of-the-art regressors and sparse matchers in terms of various metrics and reduces computational time greatly due to the riddance of robust estimators. + +
CategoryMethodPose estimationAUCTime (ms)
@5°@10°@20°MatchTotal
SparseSuperGlue [55]16.233.851.8102.5361.9
SGMNet [12]15.432.148.3111.7439.7
LightGlue [39]16.433.650.231.5269.7
DenseLoFTR [59]22.040.857.644.4292.4
ASpanFormer [13]25.646.063.359.6370.9
DKM [18]29.450.768.3235.0491.6
RegressorMap-free [3]2.7011.529.0n/a12.7
SRPose13.334.356.8n/a28.5
+ +![](images/ab9c7615a46a7b7e4a9ff956854a87d6ac510e75ae64c2395e1639f864738284.jpg) +Fig. 4: Time consumption comparison on ScanNet [14]. Regressors, including SRPose, achieve much higher computational efficacy than all matchers. + +Results: Table 1 demonstrates that SRPose greatly outperforms the existing deep learning-based regressors across all metrics on Matterport. Our framework achieves the most accurate scaled translation estimation, which is unattainable for matcher-based approaches lacking depth information. Table 2 shows SRPose achieves competitive or superior performance compared to all sparse matchers and the regressor and on ScanNet. Although dense matchers outperform our framework in accuracy, SRPose offers a significant advantage in computational speed compared to all the matcher-based approaches by replacing robust estimators with direct regression. As shown in Fig. 4, matcher-based approaches spend a substantial proportion of time on recovering pose using robust estimators after matching. In contrast, SRPose consumes even less time than the matching stage of all matchers, saving at least 200ms in the total time. + +# 4.2 Object-to-Camera Pose Estimation + +Setup: In the object-to-camera scenario, we evaluate SRPose on HO3D [26]. HO3D contains videos of hand-holding objects of 21 categories from the YCB dataset [9]. The depth captured by the stereo depth camera and the object segmentation of each frame in the videos are also provided. We randomly select image pairs with small object rotation transformations for training and evaluation. More details can be found in the supplementary materials. Following [30, 69], we adopt the widely-used Average Distance (ADD, ADD-S) metric for evaluation. To measure the smaller scales in object pose transformation, we set the + +Table 3: Relative object pose estimation on HO3D [26]. SRPose achieves state-of-the-art performance in object-to-camera estimation without additional depth information, enabling scale pose estimation from only RGB inputs. + +
MethodRot. (°)Trans. (cm)ADD↑ADD-S↑
Med.↓Avg.↓≤ 30°↑Med.↓Avg.↓≤ 10cm↑
SGMNet [12]24.129.062.713.917.536.816.831.3
LightGlue [39]24.428.761.713.516.338.217.231.2
LoFTR [59]29.257.950.913.916.435.616.730.9
DKM [18]23.325.664.813.615.439.817.833.3
SRPose8.911.495.45.98.073.936.456.8
+ +threshold of metrics to $10\mathrm{cm}$ , which differs from the 1m threshold used in the camera-to-world scenario. Specifically, we calculate the AUC using both ADD and ADD-S with the threshold set to $10\mathrm{cm}$ . We also report the median, mean error of translation/rotation, and their accuracy at $10\mathrm{cm} / 30^{\circ}$ . + +Baselines: We only choose matchers listed in Sec. 4.1 as our baselines, for no deep learning-based regressors are available in this two-view relative object pose estimation task. We utilize the stereo depth offered in the dataset to enable matcher-based approaches to estimate scaled relative poses via Orthogonal Procrustes [20], while SRPose only requires RGB images as inputs. The segmentations in the dataset are adopted for matcher-based approaches to identify the target object in both images, while SRPose utilizes the bounding box of the segmentation in the first image as the object prompt. + +Results: As shown in Table 3, SRPose outperforms state-of-the-art matcher-based approaches significantly. SRPose exhibits superior accuracy in relative 6D object pose estimation between two images using the object prompts in the reference view. The low quality of stereo depth leads to limited performance in scaled pose estimation of the matcher-based approaches, while SRPose can learn scaled information to achieve lower estimation errors. + +# 4.3 Map-Free Visual Relocalization + +We evaluate the application of SRPose on Niantic [3], a dataset for the map-free visual relocalization task. Following [3], we adopt the Virtual Correspondence Reprojection Error (VCRE) and the median error of translation and rotation as the metrics. Specifically, we compute the precision of VCRE at 90 pixels, and the precision of relocalization at $0.25\mathrm{m}$ and $5^{\circ}$ . + +Baselines: We still select the three categories of approaches as our baselines. For **matching-based approaches**, we estimate the pose by recovering from the essential matrix. Then we use the depth estimated by DPT [51], a depth prediction model, to provide the matchers with the scale information. + +Results: Table 4 shows SRPose outperforms other sparse matchers and the regression method in map-free visual relocalization. Although the state-of-the-art dense matcher RoMa achieves higher precision in relocalization using scales estimated by an additional depth prediction model, SRPose still outperforms in VCRE. Our framework exhibits competitive performance compared to state-of-the-art approaches in the map-free visual relocalization task. + +Table 4: Map-free visual relocalization results on Niantic [3]. SRPose achieves competitive performance compared to state-of-the-art approaches without relying on an additional depth prediction model to assist with scaled estimation. + +
CategoryMethodVCRE +Prec.↑ / Med.↓Med. (m / °) +Trans.↓ / Rot.↓Prec.↑
SparseSIFT [44]25.0 / 222.82.93 / 61.410.3
+DPT [51]SuperGlue [55]36.1 / 160.31.88 / 25.416.8
DenseLoFTR [59]34.7 / 167.61.98 / 30.515.4
+DPT [51]RoMa [19]45.6 / 128.81.23 / 11.122.8
RegressorMap-free [3]40.2 / 147.11.68 / 22.96.0
SRPose46.4 / 127.71.37 / 17.216.9
+ +Table 5: Ablation study on Matterport [10]. The prior knowledge-guided attention layers assist in establishing implicit cross-view correspondences. + +
MethodTrans. (m)Rot. (°)
Avg.↓≤ 1m↑Avg.↓≤ 30°↑
8point [53]1.0167.419.1385.4
SRPose (full)0.8474.214.3288.9
1) no prior knowledge guidance0.9769.616.9186.8
2) replace all cross-attn. with self-attn.1.9026.148.042.0
3) no position encoding1.1761.021.3981.3
+ +# 4.4 Insights + +Ablation study: We evaluate the effectiveness of the different components in SRPose by Matterport [10] and MegaDepth [37] datasets. Specifically, we train four different variants initialized with random weights on Matterport by removing the three different designs we proposed to validate their effectiveness. As shown in Table 5: 1) Removing the prior knowledge + +Table 6: Ablation study on MegaDepth [37]. Intrinsic-calibration position encoder enables adaption on varying image sizes and camera intrinsics. + +
MethodPose est.AUC
@5°@10°@20°
Map-free [3]2.69.322.9
SRPose (full)16.636.058.0
1) no intrinsic calibration1.58.524.3
2) no position encoding1.06.218.9
+ +edge guidance of keypoint similarities in the cross-attention modules results in a moderate decrease in accuracy, as it assists SRPose in seeking the cross-view correspondences. 2) Replacing all cross-attention modules with self-attention modules leads to a significant accuracy drop, as the capability to establish implicit cross-view correspondences through cross-attention is eliminated. 3) Removing the entire position encoder leads to a moderately lower accuracy due to the absence of keypoint position information. The experiment highlights the crucial roles of the three designs of SRPose in accurate relative pose estimation. + +Moreover, we train three different variants initialized with random weights on MegaDepth, a dataset that consists of images with different sizes and camera intrinsics, to further validate the effectiveness of the intrinsic-calibration position encoder, with results shown in Table 6. 1) Without enforcing intrinsic calibration on keypoint coordinates, SRPose only achieves comparable performance to Map-free, both inadaptable to varying image sizes and camera intrinsics in MegaDepth. 2) Removing the entire IC Position Encoder results in a significant drop in accuracy as expected owing to the elimination of critical position infor + +mation. By comparing the full framework with the variants lacking IC and the entire position encoder, we demonstrate the essential roles of the two designs in adapting different image sizes and camera intrinsics. + +Visualizing attention: In Fig. 5, we visualize the cross-attention scores $A_{12}^{\prime}$ and compare SRPose without prior knowledge guidance to its full designs. We also draw the connections of highly-attended cross-view keypoint pairs, which represent the implicit correspondences established by SRPose. + +![](images/8b3dd6dce6dd434d7266560f41d224db217cd6ec6f8aa3d4a57462c2037643ad.jpg) +Fig.5: Visualization of implicit cross-view correspondences. Dots and lines drawn with brighter colors represent higher cross-view attention scores. The prior knowledge guidance enhances attention to the overlapping areas, removing the irrelevant cross-view connections, and assisting in establishing implicit correspondences. + +# 5 Conclusion + +This paper presents SRPose, a novel deep learning-based regressor utilizing sparse keypoints for two-view relative pose estimation in camera-to-world and object-to-camera scenarios. As our key innovations, SRPose extracts keypoints with sparse keypoint detectors and employs an intrinsic-calibration position encoder to enable adaptability to different image sizes and camera intrinsics. Further, our proposed promptable prior knowledge-guided attention layers establish implicit correspondences to estimate the relative pose under the epipolar constraint in both scenarios. By directly regressing the rotation and translation, SRPose achieves a significant decrease in computational time, with a minimum time reduction of 200ms. Extensive experiments demonstrate that SRPose achieves state-of-the-art performance in terms of accuracy and speed in the two scenarios, as well as in the map-free visual relocalization task. + +# Acknowledgements + +This work was supported by Shanghai Municipal Science and Technology Major Project (2021SHZDZX0102) and the Fundamental Research Funds for the Central Universities. + +# References + +1. Abouelnaga, Y., Bui, M., Ilic, S.: Distillpose: lightweight camera localization using auxiliary learning. In: IROS (2021) +2. Agarwala, S., Jin, L., Rockwell, C., Fouhey, D.F.: Planeformers: From sparse view planes to 3d reconstruction. In: ECCV (2022) +3. Arnold, E., Wynn, J., Vicente, S., Garcia-Hernando, G., Monszpart, A., Prisacariu, V., Turmukhambetov, D., Brachmann, E.: Map-free visual relocalization: Metric pose relative to a single image. In: ECCV (2022) +4. Balntas, V., Li, S., Prisacariu, V.: Relocnet: Continuous metric learning relocalisation using neural nets. In: ECCV (2018) +5. Bay, H., Tuytelaars, T., Van Gool, L.: Surf: Speeded up robust features. In: ECCV (2006) +6. Bhowmik, A., Gumhold, S., Rother, C., Brachmann, E.: Reinforced feature points: Optimizing feature detection and description for a high-level task. In: CVPR (2020) +7. Bradski, G.: The OpenCV Library. Dr. Dobb's Journal of Software Tools (2000) +8. Cai, R., Hariharan, B., Snavely, N., Averbuch-Elor, H.: Extreme rotation estimation using dense correlation volumes. In: CVPR (2021) +9. Calli, B., Singh, A., Walsman, A., Srinivasa, S., Abbeel, P., Dollar, A.M.: The ycb object and model set: Towards common benchmarks for manipulation research. In: ICAR (2015) +10. Chang, A., Dai, A., Funkhouser, T., Halber, M., Niebner, M., Savva, M., Song, S., Zeng, A., Zhang, Y.: Matterport3d: Learning from rgb-d data in indoor environments. In: 3DV (2017) +11. Chang, J., Yu, J., Zhang, T.: Structured epipolar matcher for local feature matching. In: CVPR (2023) +12. Chen, H., Luo, Z., Zhang, J., Zhou, L., Bai, X., Hu, Z., Tai, C.L., Quan, L.: Learning to match features with seeded graph matching network. In: ICCV (2021) +13. Chen, H., Luo, Z., Zhou, L., Tian, Y., Zhen, M., Fang, T., McKinnon, D., Tsin, Y., Quan, L.: Aspanformer: Detector-free image matching with adaptive span transformer. In: ECCV (2022) +14. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: CVPR (2017) +15. DeTone, D., Malisiewicz, T., Rabinovich, A.: Toward geometric deep slam. arXiv preprint arXiv:1707.07410 (2017) +16. DeTone, D., Malisiewicz, T., Rabinovich, A.: Superpoint: Self-supervised interest point detection and description. In: CVPR (2018) +17. Dusmanu, M., Rocco, I., Pajdla, T., Pollefeys, M., Sivic, J., Torii, A., Sattler, T.: D2-net: A trainable cnn for joint description and detection of local features. In: CVPR (2019) +18. Edstedt, J., Athanasiadis, I., Wadenback, M., Felsberg, M.: Dkm: Dense kernelized feature matching for geometry estimation. In: CVPR (2023) + +19. Edstedt, J., Sun, Q., Bokman, G., Wadenback, M., Felsberg, M.: RoMa: Robust Dense Feature Matching. arXiv preprint arXiv:2305.15404 (2023) +20. Eggert, D.W., Lorusso, A., Fisher, R.B.: Estimating 3-d rigid body transformations: a comparison of four major algorithms. Machine vision and applications (1997) +21. En, S., Lechery, A., Jurie, F.: Rpnet: An end-to-end network for relative camera pose estimation. In: ECCV (2018) +22. Fan, Z., Pan, P., Wang, P., Jiang, Y., Xu, D., Jiang, H., Wang, Z.: Pope: 6-dof promptable pose estimation of any object, in any scene, with one reference. arXiv preprint arXiv:2305.15727 (2023) +23. Fischler, M.A., Bolles, R.C.: Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. COMMUN ACM (1981) +24. Gleize, P., Wang, W., Feiszli, M.: Silk: Simple learned keypoints. In: ICCV (2023) +25. Goodwin, W., Vaze, S., Havoutis, I., Posner, I.: Zero-shot category-level object pose estimation. In: ECCV (2022) +26. Hampali, S., Rad, M., Oberweger, M., Lepetit, V.: Honnotate: A method for 3d annotation of hand and object poses. In: CVPR (2020) +27. Hartley, R.I.: In defense of the eight-point algorithm. TPAMI (1997) +28. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016) +29. He, X., Sun, J., Wang, Y., Huang, D., Bao, H., Zhou, X.: Onepose++: Keypoint-free one-shot object pose estimation without cad models. NeurIPS (2022) +30. Hinterstoisser, S., Lepetit, V., Ilic, S., Holzer, S., Bradski, G., Konolige, K., Navab, N.: Model based training, detection and pose estimation of texture-less 3d objects in heavily cluttered scenes. In: ACCV (2013) +31. Huber, P.J.: Robust estimation of a location parameter. The Annals of Mathematical Statistics (1964) +32. Jin, L., Qian, S., Owens, A., Fouhey, D.F.: Planar surface reconstruction from sparse views. In: ICCV (2021) +33. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023) +34. Laskar, Z., Melekhov, I., Kalia, S., Kannala, J.: Camera relocalization by computing pairwise relative poses using convolutional neural network. In: ICCVW (2017) +35. Li, K., Wang, L., Liu, L., Ran, Q., Xu, K., Guo, Y.: Decoupling makes weakly supervised local feature better. In: CVPR (2022) +36. Li, X., Han, K., Li, S., Prisacariu, V.: Dual-resolution correspondence networks. NeurIPS (2020) +37. Li, Z., Snavely, N.: Megadepth: Learning single-view depth prediction from internet photos. In: CVPR (2018) +38. Lin, A., Zhang, J.Y., Ramanan, D., Tulsiani, S.: Relpose++: Recovering 6d poses from sparse-view observations. In: 2024 International Conference on 3D Vision (3DV). pp. 106-115. IEEE (2024) +39. Lindenberger, P., Sarlin, P.E., Pollefeys, M.: LightGlue: Local Feature Matching at Light Speed. In: ICCV (2023) +40. Liu, Y., Shen, Z., Lin, Z., Peng, S., Bao, H., Zhou, X.: Gift: Learning transformation-invariant dense visual descriptors via group cnns. NeurIPS (2019) +41. Liu, Y., Wen, Y., Peng, S., Lin, C., Long, X., Komura, T., Wang, W.: Gen6d: Generalizable model-free 6-dof object pose estimation from rgb images. In: ECCV (2022) + +42. Longuet-Higgins, H.C.: A computer algorithm for reconstructing a scene from two projections. Nature (1981) +43. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: ICLR (2018) +44. Lowe, D.G.: Distinctive image features from scale-invariant keypoints. IJCV (2004) +45. Luo, Z., Zhou, L., Bai, X., Chen, H., Zhang, J., Yao, Y., Li, S., Fang, T., Quan, L.: Aslfeat: Learning local features of accurate shape and localization. In: CVPR (2020) +46. Melekhov, I., Ylioinas, J., Kannala, J., Rahtu, E.: Relative camera pose estimation using convolutional neural networks. In: ACIVS (2017) +47. Ni, J., Li, Y., Huang, Z., Li, H., Bao, H., Cui, Z., Zhang, G.: Pats: Patch area transportation with subdivision for local feature matching. In: CVPR (2023) +48. Nister, D.: An efficient solution to the five-point relative pose problem. TPAMI (2004) +49. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023) +50. Phil, W.: Bidirectional cross attention. https://github.com/lucidrains/bidirectional-cross-attention (2022) +51. Ranftl, R., Lasinger, K., Hafner, D., Schindler, K., Koltun, V.: Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. TPAMI (2020) +52. Revaud, J., Weinzaepfel, P., de Souza, C.R., Humenberger, M.: R2D2: repeatable and reliable detector and descriptor. In: NeurIPS (2019) +53. Rockwell, C., Johnson, J., Fouhey, D.F.: The 8-point algorithm as an inductive bias for relative pose prediction by vits. In: 3DV (2022) +54. Rublee, E., Rabaud, V., Konolige, K., Bradski, G.: Orb: An efficient alternative to sift or surf. In: ICCV (2011) +55. Sarlin, P.E., DeTone, D., Malisiewicz, T., Rabinovich, A.: SuperGlue: Learning feature matching with graph neural networks. In: CVPR (2020) +56. Shi, Y., Cai, J.X., Shavit, Y., Mu, T.J., Feng, W., Zhang, K.: Clustergnn: Cluster-based coarse-to-fine graph neural network for efficient feature matching. In: CVPR (2022) +57. Sinha, S., Zhang, J.Y., Tagliasacchi, A., Gilitschenski, I., Lindell, D.B.: Sparse-pose: Sparse-view camera pose regression and refinement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21349–21359 (2023) +58. Smith, L.N., Topin, N.: Super-convergence: Very fast training of neural networks using large learning rates. In: Artificial intelligence and machine learning for multidomain operations applications (2019) +59. Sun, J., Shen, Z., Wang, Y., Bao, H., Zhou, X.: LoFTR: Detector-free local feature matching with transformers. CVPR (2021) +60. Sun, J., Wang, Z., Zhang, S., He, X., Zhao, H., Zhang, G., Zhou, X.: One-positron object pose estimation without cad models. In: CVPR (2022) +61. Tan, B., Xue, N., Wu, T., Xia, G.S.: Nope-sac: Neural one-plane ransac for sparse-view planar 3d reconstruction. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023) +62. Tang, S., Zhang, J., Zhu, S., Tan, P.: Quadtree attention for vision transformers. ICLR (2022) +63. Tyszzkiewicz, M., Fua, P., Trulls, E.: Disk: Learning local features with policy gradient. NeurIPS (2020) + +64. Wang, J., Rupprecht, C., Novotny, D.: Posediffusion: Solving pose estimation via diffusion-aided bundle adjustment. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9773-9783 (2023) +65. Wang, Q., Zhang, J., Yang, K., Peng, K., Stiefelhagen, R.: Matchformer: Interleaving attention in transformers for feature matching. In: ACCV (2022) +66. Wen, B., Bekris, K.: Bundletrack: 6d pose tracking for novel objects without instance or category-level 3d models. In: IROS (2021) +67. Wen, B., Tremblay, J., Blukis, V., Tyree, S., Müller, T., Evans, A., Fox, D., Kautz, J., Birchfield, S.: Bundlesdf: Neural 6-dof tracking and 3d reconstruction of unknown objects. In: CVPR (2023) +68. Winkelbauer, D., Denninger, M., Triebel, R.: Learning to localize in new environments from synthetic training data. In: ICRA (2021) +69. Xiang, Y., Schmidt, T., Narayanan, V., Fox, D.: PoseCNN: A convolutional neural network for 6d object pose estimation in cluttered scenes. Robotics: Science and Systems XIV (2018) +70. Xue, F., Budvytis, I., Cipolla, R.: Imp: Iterative matching and pose estimation with adaptive pooling. In: CVPR (2023) +71. Xue, F., Budvytis, I., Cipolla, R.: Sfd2: Semantic-guided feature detection and description. In: CVPR (2023) +72. Yi, K.M., Trulls, E., Lepetit, V., Fua, P.: Lift: Learned invariant feature transform. In: ECCV (2016) +73. Yu, J., Chang, J., He, J., Zhang, T., Yu, J., Wu, F.: Adaptive spot-guided transformer for consistent local feature matching. In: CVPR (2023) +74. Zhang, J.Y., Ramanan, D., Tulsiani, S.: Relpose: Predicting probabilistic relative rotation for single objects in the wild. In: European Conference on Computer Vision. pp. 592-611. Springer (2022) +75. Zhao, X., Wu, X., Chen, W., Chen, P.C.Y., Xu, Q., Li, Z.: Aliked: A lighter keypoint and descriptor extraction network via deformable transformation. TIM (2023) +76. Zhao, X., Wu, X., Miao, J., Chen, W., Chen, P.C., Li, Z.: Alike: Accurate and lightweight keypoint detection and descriptor extraction. TMM (2022) +77. Zhou, Q., Sattler, T., Pollefeys, M., Leal-Taixe, L.: To learn or not to learn: Visual localization from essential matrices. In: ICRA (2020) +78. Zhou, Y., Barnes, C., Lu, J., Yang, J., Li, H.: On the continuity of rotation representations in neural networks. In: CVPR (2019) +79. Zhu, S., Liu, X.: Pmatch: Paired masked image modeling for dense geometric matching. In: CVPR (2023) \ No newline at end of file diff --git a/2024/SRPose_ Two-view Relative Pose Estimation with Sparse Keypoints/images.zip b/2024/SRPose_ Two-view Relative Pose Estimation with Sparse Keypoints/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..66d2e00ab8b17ee83cad634493abc9bf86713b00 --- /dev/null +++ b/2024/SRPose_ Two-view Relative Pose Estimation with Sparse Keypoints/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:654ce4582fd90f334e7c514673f4a451ebe80f262ca3445d9b1659b9953209f0 +size 591650 diff --git a/2024/SRPose_ Two-view Relative Pose Estimation with Sparse Keypoints/layout.json b/2024/SRPose_ Two-view Relative Pose Estimation with Sparse Keypoints/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e881d538d1a59aa68a762e7ecbeb77df937ccb78 --- /dev/null +++ b/2024/SRPose_ Two-view Relative Pose Estimation with Sparse Keypoints/layout.json @@ -0,0 +1,11002 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 153, + 111, + 460, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 111, + 460, + 148 + ], + "spans": [ + { + "bbox": [ + 153, + 111, + 460, + 148 + ], + "type": "text", + "content": "SRPose: Two-view Relative Pose Estimation with Sparse Keypoints" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 186, + 167, + 427, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 167, + 427, + 199 + ], + "spans": [ + { + "bbox": [ + 186, + 167, + 427, + 199 + ], + "type": "text", + "content": "Rui Yin" + }, + { + "bbox": [ + 186, + 167, + 427, + 199 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 186, + 167, + 427, + 199 + ], + "type": "text", + "content": ", Yulun Zhang" + }, + { + "bbox": [ + 186, + 167, + 427, + 199 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 186, + 167, + 427, + 199 + ], + "type": "text", + "content": ", Zherong Pan" + }, + { + "bbox": [ + 186, + 167, + 427, + 199 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 186, + 167, + 427, + 199 + ], + "type": "text", + "content": ", Jianjun Zhu" + }, + { + "bbox": [ + 186, + 167, + 427, + 199 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 186, + 167, + 427, + 199 + ], + "type": "text", + "content": ", Cheng Wang" + }, + { + "bbox": [ + 186, + 167, + 427, + 199 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 186, + 167, + 427, + 199 + ], + "type": "text", + "content": ", and Biao Jia" + }, + { + "bbox": [ + 186, + 167, + 427, + 199 + ], + "type": "inline_equation", + "content": "^{1^{\\star}}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 143, + 206, + 470, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 206, + 470, + 242 + ], + "spans": [ + { + "bbox": [ + 143, + 206, + 470, + 242 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 143, + 206, + 470, + 242 + ], + "type": "text", + "content": " Hanglok-Tech, China \n" + }, + { + "bbox": [ + 143, + 206, + 470, + 242 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 143, + 206, + 470, + 242 + ], + "type": "text", + "content": " MoE Key Lab of Artificial Intelligence, Shanghai Jiao Tong University, China \n" + }, + { + "bbox": [ + 143, + 206, + 470, + 242 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 143, + 206, + 470, + 242 + ], + "type": "text", + "content": " LightSpeed Studios, Tencent America" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 271, + 455, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 271, + 455, + 460 + ], + "spans": [ + { + "bbox": [ + 160, + 271, + 455, + 460 + ], + "type": "text", + "content": "Abstract. Two-view pose estimation is essential for map-free visual relocalization and object pose tracking tasks. However, traditional matching methods suffer from time-consuming robust estimators, while deep learning-based pose regressors only cater to camera-to-world pose estimation, lacking generalizability to different image sizes and camera intrinsics. In this paper, we propose SRPose, a sparse keypoint-based framework for two-view relative pose estimation in camera-to-world and object-to-camera scenarios. SRPose consists of a sparse keypoint detector, an intrinsic-calibration position encoder, and promptable prior knowledge-guided attention layers. Given two RGB images of a fixed scene or a moving object, SRPose estimates the relative camera or 6D object pose transformation. Extensive experiments demonstrate that SRPose achieves competitive or superior performance compared to state-of-the-art methods in terms of accuracy and speed, showing generalizability to both scenarios. It is robust to different image sizes and camera intrinsics, and can be deployed with low computing resources. Project page: https://frickyinn.github.io/srpose." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 468, + 440, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 468, + 440, + 481 + ], + "spans": [ + { + "bbox": [ + 160, + 468, + 440, + 481 + ], + "type": "text", + "content": "Keywords: Relative Pose Estimation " + }, + { + "bbox": [ + 160, + 468, + 440, + 481 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 468, + 440, + 481 + ], + "type": "text", + "content": " 6D Object Pose Estimation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 506, + 230, + 518 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 506, + 230, + 518 + ], + "spans": [ + { + "bbox": [ + 132, + 506, + 230, + 518 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 526, + 482, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 526, + 482, + 646 + ], + "spans": [ + { + "bbox": [ + 130, + 526, + 482, + 646 + ], + "type": "text", + "content": "Relative pose estimation between two images plays a crucial role in many 3D vision tasks, including visual odometry and map-free visual relocalization [3], 6D object pose tracking [66,67], etc. These tasks can be categorized into two scenarios: camera-to-world and object-to-camera estimation. In the camera-to-world scenario, such as in map-free visual relocalization, we estimate the pose transformation from the reference image to the query image, predicting the change in camera extrinsic between the two views. In the object-to-camera scenario, taking object pose tracking for example, we aim to track the 6D pose of the target object in the videos by estimating its relative transformation in the camera coordinate system between two adjacent frames. While these two scenarios may" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 654, + 272, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 654, + 272, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 654, + 272, + 666 + ], + "type": "text", + "content": "* indicates corresponding author." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 157, + 114, + 255, + 188 + ], + "blocks": [ + { + "bbox": [ + 157, + 114, + 255, + 188 + ], + "lines": [ + { + "bbox": [ + 157, + 114, + 255, + 188 + ], + "spans": [ + { + "bbox": [ + 157, + 114, + 255, + 188 + ], + "type": "image", + "image_path": "57982706270fbec43987e0e4adc8405d20f72effda6bd14de5d15e8045a2bd53.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 181, + 190, + 230, + 198 + ], + "lines": [ + { + "bbox": [ + 181, + 190, + 230, + 198 + ], + "spans": [ + { + "bbox": [ + 181, + 190, + 230, + 198 + ], + "type": "text", + "content": "(a) Reference" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 256, + 114, + 353, + 188 + ], + "blocks": [ + { + "bbox": [ + 256, + 114, + 353, + 188 + ], + "lines": [ + { + "bbox": [ + 256, + 114, + 353, + 188 + ], + "spans": [ + { + "bbox": [ + 256, + 114, + 353, + 188 + ], + "type": "image", + "image_path": "dd0be0ec8e50da40d3a979a4bdf49612a77eab57bbc12296dd31468d2f3e82a4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 287, + 190, + 324, + 198 + ], + "lines": [ + { + "bbox": [ + 287, + 190, + 324, + 198 + ], + "spans": [ + { + "bbox": [ + 287, + 190, + 324, + 198 + ], + "type": "text", + "content": "(b) Query" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 356, + 114, + 454, + 188 + ], + "blocks": [ + { + "bbox": [ + 356, + 114, + 454, + 188 + ], + "lines": [ + { + "bbox": [ + 356, + 114, + 454, + 188 + ], + "spans": [ + { + "bbox": [ + 356, + 114, + 454, + 188 + ], + "type": "image", + "image_path": "fc65be99503e69315faa02432f7da2197878de4c04f541f27c4a8c55fdfee04e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 373, + 190, + 436, + 198 + ], + "lines": [ + { + "bbox": [ + 373, + 190, + 436, + 198 + ], + "spans": [ + { + "bbox": [ + 373, + 190, + 436, + 198 + ], + "type": "text", + "content": "(c) Ground Truth" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 156, + 201, + 254, + 275 + ], + "blocks": [ + { + "bbox": [ + 156, + 201, + 254, + 275 + ], + "lines": [ + { + "bbox": [ + 156, + 201, + 254, + 275 + ], + "spans": [ + { + "bbox": [ + 156, + 201, + 254, + 275 + ], + "type": "image", + "image_path": "1c61be74d7c10000943d1aa9116753914ccd4d0124232d6a2992b21506ce517a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 181, + 277, + 231, + 285 + ], + "lines": [ + { + "bbox": [ + 181, + 277, + 231, + 285 + ], + "spans": [ + { + "bbox": [ + 181, + 277, + 231, + 285 + ], + "type": "text", + "content": "(d) Reference" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 256, + 201, + 353, + 275 + ], + "blocks": [ + { + "bbox": [ + 256, + 201, + 353, + 275 + ], + "lines": [ + { + "bbox": [ + 256, + 201, + 353, + 275 + ], + "spans": [ + { + "bbox": [ + 256, + 201, + 353, + 275 + ], + "type": "image", + "image_path": "a891af73eec98b5499f70449c883f9e7b126b56bce20048dd270102aa4bf64ca.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 287, + 277, + 324, + 285 + ], + "lines": [ + { + "bbox": [ + 287, + 277, + 324, + 285 + ], + "spans": [ + { + "bbox": [ + 287, + 277, + 324, + 285 + ], + "type": "text", + "content": "(e) Query" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 356, + 201, + 454, + 275 + ], + "blocks": [ + { + "bbox": [ + 356, + 201, + 454, + 275 + ], + "lines": [ + { + "bbox": [ + 356, + 201, + 454, + 275 + ], + "spans": [ + { + "bbox": [ + 356, + 201, + 454, + 275 + ], + "type": "image", + "image_path": "3205b13a541666eaaf8e29d45dd60f8c971be081b889f0990944615ca73331ce.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 373, + 277, + 436, + 285 + ], + "lines": [ + { + "bbox": [ + 373, + 277, + 436, + 285 + ], + "spans": [ + { + "bbox": [ + 373, + 277, + 436, + 285 + ], + "type": "text", + "content": "(f) Ground Truth" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 289, + 482, + 376 + ], + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 376 + ], + "type": "text", + "content": "Fig. 1: Relative pose estimation by SRPose. Dots drawn in the figures visualize the cross-attention scores of sparse keypoints across the two images, with brighter dots representing higher attention. Camera-to-world: (a), (b), (c) visualize the epipolar lines, representing the connections between the nine corresponding points across two views. Higher attention is shown to the overlap of the scenes. Object-to-Camera: (d), (e), (f) show the relative 6D pose estimation in the query image with only one accessible object prompt in the reference image. Higher attention is shown to the target object. SRPose establishes implicit correspondences." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 384, + 481, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 481, + 421 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 481, + 421 + ], + "type": "text", + "content": "seem distinct, they both entail the prediction of the pose matrix from the two RGB images depicting the scene or object, which in turn represents the relative pose transformation from the first image to the second." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 423, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 482, + 604 + ], + "type": "text", + "content": "Traditional pose estimation approaches typically involve detecting and matching keypoints or local features in the two images to establish point-to-point correspondences, then use a robust estimator to denoise the outliers and recover the relative pose from the essential matrix by solving a two-view geometry problem, i.e., the epipolar constraint equation [48]. These approaches produce accurate results regardless of image sizes and camera intrinsicics, thanks to the intrinsic calibration performed before solving the equation. Given the segmentation of the target object in the two images, these approaches can also yield acceptable predictions of the relative object pose transformation. However, on the downside, robust estimators can be computationally expensive. Indeed, the process of eliminating mismatched correspondences and solving the equations using robust estimators is significantly slower compared to detecting and matching keypoints or features, making it prohibitive in the real-time application. Moreover, real-time object pose tracking requires video object segmentation models to eliminate off-target keypoints [66, 67], which can also introduce additional overhead." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "The maturity of deep learning offers the advantage of directly regressing the relative pose transformation from two RGB inputs, significantly boosting the runtime performance. However, existing regression methods lack of generalizability or precision to images with varying sizes and camera intrinsics [3, 38, 53, 64, 74, 77]. They typically employ image encoders that only accept" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "R. Yin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "content": "fixed-size inputs during batched training. And unlike traditional approaches, most deep learning-based regressors lack awareness of the original camera intrinsic parameters, unable to utilize two-view geometry to achieve higher precision in pose estimation. Furthermore, state-of-the-art deep regressors only apply to camera-to-world pose estimation, despite the common solution shared by both the camera-to-world and object-to-camera tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 480, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 480, + 355 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 480, + 355 + ], + "type": "text", + "content": "To address the aforementioned challenges, we propose SRPose: a Sparse keypoint-based framework for Relative Pose estimation. SRPose estimates the relative pose matrix based on two-view geometry by implicitly solving the epipolar constraint, as shown in Fig. 1. First, SRPose employs a point detector to extract sparse keypoints and associated descriptors to form the candidate set of implicit correspondences. Then, we use an intrinsic-calibration (IC) position encoder to modulate the keypoints before it computes the position embeddings, accounting for varying image sizes and camera intrinsics. The keypoint descriptors and the position embeddings are then fused into the prior knowledge-guided attention layers, which leverage the prior knowledge of keypoint similarities to establish implicit cross-view correspondences. With the mechanism of the promptable cross-attention in these layers, SRPose only requires an accessible object prompt in one of the two views to compute the relative 6D pose transformation of the target object, eliminating the need for elusive object segmentations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 356, + 441, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 356, + 441, + 367 + ], + "spans": [ + { + "bbox": [ + 146, + 356, + 441, + 367 + ], + "type": "text", + "content": "The main contribution of this paper can be summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 370, + 479, + 488 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 138, + 370, + 479, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 370, + 479, + 404 + ], + "spans": [ + { + "bbox": [ + 138, + 370, + 479, + 404 + ], + "type": "text", + "content": "- We propose a novel framework using sparse keypoints, SRPose, for two-view relative pose estimation. To the best of our knowledge, it is the first attempt to directly regress relative poses from sparse keypoints for this task." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 407, + 479, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 407, + 479, + 427 + ], + "spans": [ + { + "bbox": [ + 138, + 407, + 479, + 427 + ], + "type": "text", + "content": "- We introduce an intrinsic-calibration position encoder to adapt SRPose to different image sizes and camera intrinsics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 430, + 479, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 430, + 479, + 451 + ], + "spans": [ + { + "bbox": [ + 137, + 430, + 479, + 451 + ], + "type": "text", + "content": "- We enable object-to-camera estimation in addition to camera-to-world estimation by utilizing one accessible object prompt." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 137, + 455, + 479, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 455, + 479, + 488 + ], + "spans": [ + { + "bbox": [ + 137, + 455, + 479, + 488 + ], + "type": "text", + "content": "- Our SRPose greatly reduces estimation time by replacing robust estimators with direct regression, achieving state-of-the-art performance in accuracy and speed in camera-to-world and object-to-camera scenarios." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 504, + 239, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 239, + 516 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 239, + 516 + ], + "type": "text", + "content": "2 Related works" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 521, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 480, + 665 + ], + "type": "text", + "content": "Relative Pose Estimation: In relative pose estimation, traditional approaches recover the rotation and translation from the essential matrix [27, 42]. By using the epipolar constraint [48], the essential matrix is estimated based on the cross-view correspondences established through the matching of sparse keypoints or dense features. Using this pipeline, sparse keypoint-based approaches typically start by detecting sparse keypoints and their associated descriptors, including classic hand-crafted detectors [5, 44, 54], and deep learning-based detectors [6, 15-17, 24, 35, 40, 45, 52, 63, 72, 75, 76]. Recent methods further employ deep learning-based matchers [12, 39, 55, 56, 70] to establish keypoint matches, while classical approaches rely on nearest neighbor search. In contrast to using sparse keypoints, dense matchers [11, 13, 18, 19, 36, 47, 59, 62, 65, 71, 73, 79] are detector-free and perform pixel-wise dense feature matching. Once the correspondences" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 172, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 172, + 91, + 447, + 102 + ], + "type": "text", + "content": "SRPose: Two-view Relative Pose Estimation with Sparse Keypoints" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "are established through matching, a robust estimator such as RANSAC [23] is employed to estimate the essential matrix and recover the relative pose. Overall, while the traditional matcher-based approaches have shown promising results, they still face challenges with their time-consuming robust estimators. Furthermore, dense matchers, while providing exceptional accuracy, may suffer from slow performance due to their resource-intensive design." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 190, + 482, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 190, + 482, + 382 + ], + "spans": [ + { + "bbox": [ + 130, + 190, + 482, + 382 + ], + "type": "text", + "content": "Relative Pose Regression: Deep learning regressors offer an alternative approach by directly predicting the relative pose using dedicated neural networks, without explicitly establishing correspondences. Two-view relative pose estimation focuses on frame-to-frame pose transformations, aiming at visual odometry and map-free relocalization [1,3,4,21,34,46,68]. Leveraging prior knowledge is an important aspect in two-view estimation. For instance, SparsePlanes [32], PlaneFormer [8] and NOPE-SAC [61] utilize geometry relations of the planar surfaces to enhance performance. Unfortunately, these approaches rely on image encoders designed for images of fixed sizes and intrinsics, lacking generalizability across different settings. Sparse-view relative pose estimation emphasizes global pose optimization or end-to-end SfM involving multiple views [38,57,64,74]. Although these sparse-view methods can be extended to two-view estimation, and may even predict unknown intrinsics, their adaptability to varying camera settings depends on multi-view information. In contrast, SRPose leverages sparse keypoints and two-view geometry to achieve higher precision and generalizability across different image sizes and camera intrinsics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 384, + 482, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 482, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 482, + 601 + ], + "type": "text", + "content": "Object Pose Estimation: While most of the aforementioned approaches focus on camera pose estimation in a fixed scene, some tasks, such as 6D object pose tracking, require relative object pose estimation as well. Most frameworks predict the relative object pose using the same pipeline of keypoint correspondences. BundleTrack [66] and BundleSDF [67] track the object poses in videos by matching keypoints between two frames, and conducting global pose graph optimization. POPE [22] combines advanced foundation models of segmentation [33] and matching [49] to enable zero-shot object pose estimation. Robust as these classic methods may be, they require dedicated object masks to focus on the objects, which are provided by other resource-consuming segmentation models. To get rid of object segmentation, OnePose [60], OnePose++ [29], and Gen6D [41] rely on multiple views of the object, and ZeroShot [25] establishes semantic correspondences to mute the background, and employs depth map to enable pose estimation. Nonetheless, these methods require additional high-quality information, and they restrict the querying view to only accommodate a single object. As a remedy, SRPose employs an accessible object prompt to focus on the target object, enabling effective depth-free and mask-free two-view relative object estimation with superior performance in the object-to-camera scenario." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 603, + 203, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 603, + 203, + 616 + ], + "spans": [ + { + "bbox": [ + 131, + 603, + 203, + 616 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "Given two images " + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "I_1, I_2" + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": ", SRPose estimates the relative pose consisting of a rotation " + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "R \\in \\mathcal{SO}(3)" + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": ", and a translation " + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "t \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": " based on the two-view geometry. Detailed problem definitions can be found in the supplementary materials. Traditional approaches use the epipolar constraint [48] to estimate the following" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "R. Yin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 116, + 470, + 210 + ], + "blocks": [ + { + "bbox": [ + 140, + 116, + 470, + 210 + ], + "lines": [ + { + "bbox": [ + 140, + 116, + 470, + 210 + ], + "spans": [ + { + "bbox": [ + 140, + 116, + 470, + 210 + ], + "type": "image", + "image_path": "689444cd282daa516213b93b0c16f3427b7984e595ca9f7796569d6f21b6f37d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 213, + 482, + 290 + ], + "lines": [ + { + "bbox": [ + 131, + 213, + 482, + 290 + ], + "spans": [ + { + "bbox": [ + 131, + 213, + 482, + 290 + ], + "type": "text", + "content": "Fig. 2: Overview. SRPose comprises four main components: 1) The sparse keypoint detector detects keypoints associated with descriptors separately sfrom the two images; 2) The intrinsic-calibration (IC) position encoder modulates the keypoints' coordinates with camera intrinsics, and encodes their position information; 3) Guided by the prior knowledge of keypoint similarities, along with the object prompt, the attention layers establish implicit cross-view correspondences; 4) The regressor estimates relative pose " + }, + { + "bbox": [ + 131, + 213, + 482, + 290 + ], + "type": "inline_equation", + "content": "R, t" + }, + { + "bbox": [ + 131, + 213, + 482, + 290 + ], + "type": "text", + "content": " under the constraints of implicit correspondences." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 295, + 353, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 295, + 353, + 307 + ], + "spans": [ + { + "bbox": [ + 132, + 295, + 353, + 307 + ], + "type": "text", + "content": "essential matrix " + }, + { + "bbox": [ + 132, + 295, + 353, + 307 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 132, + 295, + 353, + 307 + ], + "type": "text", + "content": ", from which " + }, + { + "bbox": [ + 132, + 295, + 353, + 307 + ], + "type": "inline_equation", + "content": "R, t" + }, + { + "bbox": [ + 132, + 295, + 353, + 307 + ], + "type": "text", + "content": " are recovered:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 229, + 313, + 481, + 328 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 229, + 313, + 481, + 328 + ], + "spans": [ + { + "bbox": [ + 229, + 313, + 481, + 328 + ], + "type": "interline_equation", + "content": "\\left(K _ {1} ^ {- 1} q _ {1}\\right) ^ {\\top} E K _ {2} ^ {- 1} q _ {2} = 0, \\quad \\forall \\left(q _ {1}, q _ {2}\\right), \\tag {1}", + "image_path": "a820e7e397adaccd7b70a32ebd6e9fb39ccba4122a6a5eb5890b4dab6c3d9280.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 282, + 330, + 480, + 342 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 282, + 330, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 282, + 330, + 480, + 342 + ], + "type": "interline_equation", + "content": "E = [ t ] _ {\\times} R, \\tag {2}", + "image_path": "eb1f9dd28ce4fed18d88c83f95e22aa24b6eb8a11407dcff27ff2c4c2a96ff91.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 350, + 480, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 350, + 480, + 398 + ], + "spans": [ + { + "bbox": [ + 131, + 350, + 480, + 398 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 350, + 480, + 398 + ], + "type": "inline_equation", + "content": "q_{1}, q_{2}" + }, + { + "bbox": [ + 131, + 350, + 480, + 398 + ], + "type": "text", + "content": " are two arbitrary corresponding points in " + }, + { + "bbox": [ + 131, + 350, + 480, + 398 + ], + "type": "inline_equation", + "content": "I_{1}, I_{2}" + }, + { + "bbox": [ + 131, + 350, + 480, + 398 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 350, + 480, + 398 + ], + "type": "inline_equation", + "content": "K_{1}, K_{2}" + }, + { + "bbox": [ + 131, + 350, + 480, + 398 + ], + "type": "text", + "content": " are the camera intrinsics of the two images. SRPose builds and solves the constraint equation Eq. (1) with neural networks implicitly. With SRPose denoted as " + }, + { + "bbox": [ + 131, + 350, + 480, + 398 + ], + "type": "inline_equation", + "content": "\\mathcal{SR}" + }, + { + "bbox": [ + 131, + 350, + 480, + 398 + ], + "type": "text", + "content": ", the function of our framework is represented as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 264, + 405, + 480, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 405, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 264, + 405, + 480, + 418 + ], + "type": "interline_equation", + "content": "[ R | t ] = \\mathcal {S R} (I _ {1}, I _ {2}). \\tag {3}", + "image_path": "52db44ab05229124ac02beb1f92fda822fc954ac07a4f18e4282378c6b324959.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 426, + 479, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 426, + 479, + 462 + ], + "spans": [ + { + "bbox": [ + 131, + 426, + 479, + 462 + ], + "type": "text", + "content": "Figure 2 summarizes SRPose, involving four components: a sparse keypoint detector, an intrinsic-calibration position encoder, the promptable prior knowledge-guided attention layers, and an MLP regressor." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 466, + 292, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 466, + 292, + 478 + ], + "spans": [ + { + "bbox": [ + 132, + 466, + 292, + 478 + ], + "type": "text", + "content": "3.1 Sparse Keypoint Detector" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 479, + 482, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 482, + 587 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 482, + 587 + ], + "type": "text", + "content": "Instead of employing an image encoder to extract image features, we use a detector to extract sparse keypoints and associated descriptors from the two images. To this end, we can use one of the existing techniques developed over the years, which include classic methods such as SIFT [44], and deep learning-based methods such as SuperPoint [16], ALIKE [76], DISK [63], etc. Given an image " + }, + { + "bbox": [ + 130, + 479, + 482, + 587 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 479, + 482, + 587 + ], + "type": "text", + "content": ", our detector, denoted as " + }, + { + "bbox": [ + 130, + 479, + 482, + 587 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{kd}" + }, + { + "bbox": [ + 130, + 479, + 482, + 587 + ], + "type": "text", + "content": ", detects keypoints consisting of " + }, + { + "bbox": [ + 130, + 479, + 482, + 587 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 479, + 482, + 587 + ], + "type": "text", + "content": " coordinates " + }, + { + "bbox": [ + 130, + 479, + 482, + 587 + ], + "type": "inline_equation", + "content": "P^k \\subset \\mathbb{R}^2" + }, + { + "bbox": [ + 130, + 479, + 482, + 587 + ], + "type": "text", + "content": " in image space, and associated descriptors " + }, + { + "bbox": [ + 130, + 479, + 482, + 587 + ], + "type": "inline_equation", + "content": "D \\subset \\mathbb{R}^d" + }, + { + "bbox": [ + 130, + 479, + 482, + 587 + ], + "type": "text", + "content": ", describing the local features in the image, where " + }, + { + "bbox": [ + 130, + 479, + 482, + 587 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 479, + 482, + 587 + ], + "type": "text", + "content": " denotes the dimension of the keypoints descriptors in the network. To summarize, our detector is defined as the following function:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 263, + 594, + 481, + 608 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 263, + 594, + 481, + 608 + ], + "spans": [ + { + "bbox": [ + 263, + 594, + 481, + 608 + ], + "type": "interline_equation", + "content": "\\left\\{P ^ {k}, D \\right\\} = \\mathcal {F} ^ {k d} (I). \\tag {4}", + "image_path": "704e0287388f7babd0f8ed1bbe202c0dc8a86c06a26b62b86842db1412d9ff80.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 614, + 482, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 614, + 482, + 675 + ], + "spans": [ + { + "bbox": [ + 131, + 614, + 482, + 675 + ], + "type": "text", + "content": "In practice, we resize all images to a standardized size to enable parallel batched detection, then rescale the keypoints' coordinates " + }, + { + "bbox": [ + 131, + 614, + 482, + 675 + ], + "type": "inline_equation", + "content": "P^k" + }, + { + "bbox": [ + 131, + 614, + 482, + 675 + ], + "type": "text", + "content": " to their original positions. The sparse keypoint detector yields two sets of keypoints " + }, + { + "bbox": [ + 131, + 614, + 482, + 675 + ], + "type": "inline_equation", + "content": "P_1^k, P_2^k" + }, + { + "bbox": [ + 131, + 614, + 482, + 675 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 131, + 614, + 482, + 675 + ], + "type": "inline_equation", + "content": "I_1, I_2" + }, + { + "bbox": [ + 131, + 614, + 482, + 675 + ], + "type": "text", + "content": ", from which a set of correspondences " + }, + { + "bbox": [ + 131, + 614, + 482, + 675 + ], + "type": "inline_equation", + "content": "\\{(q_1, q_2)\\} \\subset P_1^k \\times P_2^k" + }, + { + "bbox": [ + 131, + 614, + 482, + 675 + ], + "type": "text", + "content": " will be established implicitly later for solving the constraint Eq. (1)." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 172, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 172, + 91, + 448, + 102 + ], + "type": "text", + "content": "SRPose: Two-view Relative Pose Estimation with Sparse Keypoints" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 379, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 379, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 379, + 128 + ], + "type": "text", + "content": "3.2 Intrinsic-Calibration (IC) Position Encoder" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 129, + 481, + 153 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 129, + 481, + 153 + ], + "spans": [ + { + "bbox": [ + 130, + 129, + 481, + 153 + ], + "type": "text", + "content": "According to Eq. (1), for the keypoints detected from an image " + }, + { + "bbox": [ + 130, + 129, + 481, + 153 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 129, + 481, + 153 + ], + "type": "text", + "content": ", their coordinates should be first calibrated or normalized by the camera intrinsics " + }, + { + "bbox": [ + 130, + 129, + 481, + 153 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 129, + 481, + 153 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 277, + 162, + 481, + 175 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 162, + 481, + 175 + ], + "spans": [ + { + "bbox": [ + 277, + 162, + 481, + 175 + ], + "type": "interline_equation", + "content": "P ^ {c} = K ^ {- 1} P ^ {k}. \\tag {5}", + "image_path": "79dd83881688bdab8450571257f686021c6926fafcdff0313b713b49d13a8827.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 182, + 479, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 479, + 277 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 479, + 277 + ], + "type": "text", + "content": "Enforcing the intrinsic calibration (IC) offers a crucial improvement of both the accuracy and robustness [48]. IC allows SRPose to adapt images captured by different cameras with varying intrinsic parameters. Previous regressors [3, 53] extract implicit image features with image encoders trained on fixed-size images, assuming the camera intrinsics of all inputs are the same by default. While the sparse keypoints in SRPose are modulated by intrinsic calibration, which provides the accurate positions, i.e. " + }, + { + "bbox": [ + 130, + 182, + 479, + 277 + ], + "type": "inline_equation", + "content": "P^c" + }, + { + "bbox": [ + 130, + 182, + 479, + 277 + ], + "type": "text", + "content": ", of all keypoints in a unified camera coordinate system across different intrinsics." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 279, + 480, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 279, + 480, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 279, + 480, + 361 + ], + "type": "text", + "content": "After the intrinsic calibration, the keypoints' coordinates are encoded to address the keypoints based on their positions. Due to the possible position distortion caused by relative position encoding [39, 59], we employ absolute position encoding that maintains the precise coordinate values, which is crucial to the two-view geometry problems. With a single-layer fully connected network, denoted as the following function " + }, + { + "bbox": [ + 130, + 279, + 480, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{pe}" + }, + { + "bbox": [ + 130, + 279, + 480, + 361 + ], + "type": "text", + "content": ", the position encoder maps 2D calibrated coordinates to high-dimension position embeddings:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 265, + 371, + 481, + 384 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 265, + 371, + 481, + 384 + ], + "spans": [ + { + "bbox": [ + 265, + 371, + 481, + 384 + ], + "type": "interline_equation", + "content": "P ^ {e} = \\mathcal {F} ^ {p e} \\left(K ^ {- 1} P ^ {k}\\right). \\tag {6}", + "image_path": "6678b866409a042dd1bfe184bb675f2d8a7610ff1a97300c1ed27408d05ae535.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 392, + 479, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 392, + 479, + 441 + ], + "spans": [ + { + "bbox": [ + 131, + 392, + 479, + 441 + ], + "type": "text", + "content": "The resulting encoded position embeddings, denoted as " + }, + { + "bbox": [ + 131, + 392, + 479, + 441 + ], + "type": "inline_equation", + "content": "P^e" + }, + { + "bbox": [ + 131, + 392, + 479, + 441 + ], + "type": "text", + "content": ", are then incorporated with the keypoint descriptors and descriptor embeddings, denoted as " + }, + { + "bbox": [ + 131, + 392, + 479, + 441 + ], + "type": "inline_equation", + "content": "D \\oplus P^e" + }, + { + "bbox": [ + 131, + 392, + 479, + 441 + ], + "type": "text", + "content": ", which provides calibrated and normalized position information for solving the constraint Eq. (1) using a set of attention layers." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 445, + 402, + 457 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 445, + 402, + 457 + ], + "spans": [ + { + "bbox": [ + 131, + 445, + 402, + 457 + ], + "type": "text", + "content": "3.3 Promptable Prior-Knowledge-Guided Attention" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 459, + 480, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 459, + 480, + 614 + ], + "spans": [ + { + "bbox": [ + 130, + 459, + 480, + 614 + ], + "type": "text", + "content": "Given the position embeddings and keypoint descriptors, we use a multi-layer attention network to exploit semantic information and establish implicit correspondences. Our network consists of " + }, + { + "bbox": [ + 130, + 459, + 480, + 614 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 459, + 480, + 614 + ], + "type": "text", + "content": " consecutive attention layers, with each layer comprising a self-attention module and a promptable prior knowledge-guided cross-attention module. We denote " + }, + { + "bbox": [ + 130, + 459, + 480, + 614 + ], + "type": "inline_equation", + "content": "X^{m} \\subset \\mathbb{R}^{d}" + }, + { + "bbox": [ + 130, + 459, + 480, + 614 + ], + "type": "text", + "content": " as the keypoint descriptor embeddings being processed in the " + }, + { + "bbox": [ + 130, + 459, + 480, + 614 + ], + "type": "inline_equation", + "content": "m^{th}" + }, + { + "bbox": [ + 130, + 459, + 480, + 614 + ], + "type": "text", + "content": " layer, in which " + }, + { + "bbox": [ + 130, + 459, + 480, + 614 + ], + "type": "inline_equation", + "content": "m \\in M" + }, + { + "bbox": [ + 130, + 459, + 480, + 614 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 130, + 459, + 480, + 614 + ], + "type": "inline_equation", + "content": "X^{m}" + }, + { + "bbox": [ + 130, + 459, + 480, + 614 + ], + "type": "text", + "content": " being either " + }, + { + "bbox": [ + 130, + 459, + 480, + 614 + ], + "type": "inline_equation", + "content": "X_{1}^{m}" + }, + { + "bbox": [ + 130, + 459, + 480, + 614 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 130, + 459, + 480, + 614 + ], + "type": "inline_equation", + "content": "X_{2}^{m}" + }, + { + "bbox": [ + 130, + 459, + 480, + 614 + ], + "type": "text", + "content": " from the image " + }, + { + "bbox": [ + 130, + 459, + 480, + 614 + ], + "type": "inline_equation", + "content": "I_{1}" + }, + { + "bbox": [ + 130, + 459, + 480, + 614 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 130, + 459, + 480, + 614 + ], + "type": "inline_equation", + "content": "I_{2}" + }, + { + "bbox": [ + 130, + 459, + 480, + 614 + ], + "type": "text", + "content": ". Each layer takes " + }, + { + "bbox": [ + 130, + 459, + 480, + 614 + ], + "type": "inline_equation", + "content": "X^{m-1}" + }, + { + "bbox": [ + 130, + 459, + 480, + 614 + ], + "type": "text", + "content": " as inputs, then computes the deeper embeddings " + }, + { + "bbox": [ + 130, + 459, + 480, + 614 + ], + "type": "inline_equation", + "content": "X^{m}" + }, + { + "bbox": [ + 130, + 459, + 480, + 614 + ], + "type": "text", + "content": ". In summary, the entire multi-layer network inputs keypoint descriptors " + }, + { + "bbox": [ + 130, + 459, + 480, + 614 + ], + "type": "inline_equation", + "content": "D_{1}D_{2}" + }, + { + "bbox": [ + 130, + 459, + 480, + 614 + ], + "type": "text", + "content": ", and outputs " + }, + { + "bbox": [ + 130, + 459, + 480, + 614 + ], + "type": "inline_equation", + "content": "X_{1}^{M}, X_{2}^{M}" + }, + { + "bbox": [ + 130, + 459, + 480, + 614 + ], + "type": "text", + "content": " after " + }, + { + "bbox": [ + 130, + 459, + 480, + 614 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 459, + 480, + 614 + ], + "type": "text", + "content": " layers, i.e., we define " + }, + { + "bbox": [ + 130, + 459, + 480, + 614 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 459, + 480, + 614 + ], + "type": "text", + "content": " as the initial state of embeddings " + }, + { + "bbox": [ + 130, + 459, + 480, + 614 + ], + "type": "inline_equation", + "content": "X^{0}" + }, + { + "bbox": [ + 130, + 459, + 480, + 614 + ], + "type": "text", + "content": ". Prior to each attention layer, " + }, + { + "bbox": [ + 130, + 459, + 480, + 614 + ], + "type": "inline_equation", + "content": "X_{1}^{m-1}, X_{2}^{m-1}" + }, + { + "bbox": [ + 130, + 459, + 480, + 614 + ], + "type": "text", + "content": " are incorporated with the associated position embeddings " + }, + { + "bbox": [ + 130, + 459, + 480, + 614 + ], + "type": "inline_equation", + "content": "P_{1}^{e}, P_{2}^{e}" + }, + { + "bbox": [ + 130, + 459, + 480, + 614 + ], + "type": "text", + "content": ". Following [28], residual connections are applied to every self- and cross-attention module. The architecture of the attention layers is outlined in Fig. 3." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "Self-Attention: The self-attention module captures the comprehensive semantic information of all keypoints within an entire image. In this module, " + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "X^{m - 1} \\oplus P^e" + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": " of both the images is first mapped to three vectors by a linear transformation " + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "W_{s}^{m}" + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": ", denoted as the query " + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": ", the key " + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": ", and value " + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": " respectively, where" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "R. Yin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 116, + 465, + 219 + ], + "blocks": [ + { + "bbox": [ + 143, + 116, + 465, + 219 + ], + "lines": [ + { + "bbox": [ + 143, + 116, + 465, + 219 + ], + "spans": [ + { + "bbox": [ + 143, + 116, + 465, + 219 + ], + "type": "image", + "image_path": "bf8095184d857ed796812d2f567ac18f8cefee0318e694ba40bb390b968be5fc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 224, + 480, + 278 + ], + "lines": [ + { + "bbox": [ + 131, + 224, + 480, + 278 + ], + "spans": [ + { + "bbox": [ + 131, + 224, + 480, + 278 + ], + "type": "text", + "content": "Fig. 3: Overview of the prior knowledge-guided attention layers. Each layer contains a self-attention and a cross-attention module. A similarity matrix " + }, + { + "bbox": [ + 131, + 224, + 480, + 278 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 131, + 224, + 480, + 278 + ], + "type": "text", + "content": " of the keypoint descriptors is utilized as the prior knowledge to readjust the cross-attention scores, guiding more attention to cross-view keypoint pairs with implicit correspondences. The object prompt and the residual connections are omitted in the figure." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 283, + 479, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 283, + 479, + 308 + ], + "spans": [ + { + "bbox": [ + 131, + 283, + 479, + 308 + ], + "type": "inline_equation", + "content": "q, k, v \\subset \\mathbb{R}^d" + }, + { + "bbox": [ + 131, + 283, + 479, + 308 + ], + "type": "text", + "content": ". Then " + }, + { + "bbox": [ + 131, + 283, + 479, + 308 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 131, + 283, + 479, + 308 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 283, + 479, + 308 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 131, + 283, + 479, + 308 + ], + "type": "text", + "content": " are used to compute the self-attention scores " + }, + { + "bbox": [ + 131, + 283, + 479, + 308 + ], + "type": "inline_equation", + "content": "A_s" + }, + { + "bbox": [ + 131, + 283, + 479, + 308 + ], + "type": "text", + "content": ", which are then scaled by the dimension of descriptors and embeddings " + }, + { + "bbox": [ + 131, + 283, + 479, + 308 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 131, + 283, + 479, + 308 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 244, + 315, + 481, + 331 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 315, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 244, + 315, + 481, + 331 + ], + "type": "interline_equation", + "content": "[ q | k | v ] = W _ {s} ^ {M} \\left(X ^ {m - 1} \\oplus P ^ {e}\\right), \\tag {7}", + "image_path": "279c4bb745d66fa46bdcf095bf7080175608a4ccb9a2a2fec9b992581dcb59a1.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 281, + 331, + 481, + 359 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 281, + 331, + 481, + 359 + ], + "spans": [ + { + "bbox": [ + 281, + 331, + 481, + 359 + ], + "type": "interline_equation", + "content": "A _ {s} = \\frac {q k ^ {\\top}}{\\sqrt {d}}. \\tag {8}", + "image_path": "93a5ffe0f13d40f2f60679f84a850b83cd1cbd70247a46e9b5948fa7e2ffa931.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 365, + 479, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 365, + 479, + 389 + ], + "spans": [ + { + "bbox": [ + 131, + 365, + 479, + 389 + ], + "type": "text", + "content": "After being normalized by the Softmax function, the self-attention scores " + }, + { + "bbox": [ + 131, + 365, + 479, + 389 + ], + "type": "inline_equation", + "content": "A_{s}" + }, + { + "bbox": [ + 131, + 365, + 479, + 389 + ], + "type": "text", + "content": " are then used to extract the highly attended information from " + }, + { + "bbox": [ + 131, + 365, + 479, + 389 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 131, + 365, + 479, + 389 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 253, + 397, + 481, + 412 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 397, + 481, + 412 + ], + "spans": [ + { + "bbox": [ + 253, + 397, + 481, + 412 + ], + "type": "interline_equation", + "content": "X _ {s} ^ {m - 1} = \\operatorname {S o f t m a x} \\left(A _ {s}\\right) v. \\tag {9}", + "image_path": "1cb14c55bf7f2e984857f54be627d6d7439c30da6b73a3ca213235109c5a9b69.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 419, + 479, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 419, + 479, + 456 + ], + "spans": [ + { + "bbox": [ + 131, + 419, + 479, + 456 + ], + "type": "text", + "content": "In the self-attention module, keypoint position embeddings and descriptor embeddings are computed together to summarize the semantic information within one entire image before deeper cross-view correspondences are established." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 458, + 480, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 458, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 131, + 458, + 480, + 555 + ], + "type": "text", + "content": "Prior Knowledge-guided Cross-Attention: Guided by the prior knowledge of keypoint similarities, the cross-attention module exploits the mutual information across the two images to establish implicit correspondences. Similar to the self-attention, cross-attention maps the input embeddings " + }, + { + "bbox": [ + 131, + 458, + 480, + 555 + ], + "type": "inline_equation", + "content": "X_{s}^{m - 1}" + }, + { + "bbox": [ + 131, + 458, + 480, + 555 + ], + "type": "text", + "content": " to two keys " + }, + { + "bbox": [ + 131, + 458, + 480, + 555 + ], + "type": "inline_equation", + "content": "k_{1},k_{2}" + }, + { + "bbox": [ + 131, + 458, + 480, + 555 + ], + "type": "text", + "content": " and values " + }, + { + "bbox": [ + 131, + 458, + 480, + 555 + ], + "type": "inline_equation", + "content": "v_{1},v_{2}" + }, + { + "bbox": [ + 131, + 458, + 480, + 555 + ], + "type": "text", + "content": " of the two images respectively through another linear transformation " + }, + { + "bbox": [ + 131, + 458, + 480, + 555 + ], + "type": "inline_equation", + "content": "W_{c}^{m}" + }, + { + "bbox": [ + 131, + 458, + 480, + 555 + ], + "type": "text", + "content": " layer. Following [50], we compute the bidirectional attention scores across the two images from the resulting embeddings of the previous self-attention module, using the cross-view keys " + }, + { + "bbox": [ + 131, + 458, + 480, + 555 + ], + "type": "inline_equation", + "content": "k_{1},k_{2}" + }, + { + "bbox": [ + 131, + 458, + 480, + 555 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 260, + 562, + 481, + 576 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 562, + 481, + 576 + ], + "spans": [ + { + "bbox": [ + 260, + 562, + 481, + 576 + ], + "type": "interline_equation", + "content": "\\left[ k _ {1} \\mid v _ {1} \\right] = W _ {c} ^ {m} X _ {s, 1} ^ {m - 1}, \\tag {10}", + "image_path": "f4f2ce637f04bb2874af521d70781bbe7fa6c042a5c0393dc87c837ff72f8dff.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 261, + 578, + 481, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 578, + 481, + 594 + ], + "spans": [ + { + "bbox": [ + 261, + 578, + 481, + 594 + ], + "type": "interline_equation", + "content": "\\left[ k _ {2} \\mid v _ {2} \\right] = W _ {c} ^ {m} X _ {s, 2} ^ {m - 1}, \\tag {11}", + "image_path": "d70268d8104042093519051bb02aabfd1892b1c97833932c272dccda4343ac54.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 263, + 595, + 481, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 263, + 595, + 481, + 622 + ], + "spans": [ + { + "bbox": [ + 263, + 595, + 481, + 622 + ], + "type": "interline_equation", + "content": "A _ {2 1} ^ {\\top} = A _ {1 2} = \\frac {k _ {1} k _ {2} ^ {\\top}}{\\sqrt {d}}. \\tag {12}", + "image_path": "6aadc5fb32ef4b0a41e5c15caeb30d95dd28112d809e623ef568215cbd1ef6ee.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 629, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 629, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 629, + 479, + 665 + ], + "type": "text", + "content": "The module recognizes the overlapping areas of the scene or object between the two views by identifying the mutuality between two sets of keypoint embeddings from the computed cross-attention scores. To facilitate the process, prior" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 172, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 172, + 91, + 448, + 102 + ], + "type": "text", + "content": "SRPose: Two-view Relative Pose Estimation with Sparse Keypoints" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 165 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 165 + ], + "type": "text", + "content": "knowledge is used to guide and enhance the attention on highly correlated keypoints, by adjusting the cross-attention scores with a keypoint similarity matrix " + }, + { + "bbox": [ + 132, + 116, + 482, + 165 + ], + "type": "inline_equation", + "content": "S \\in \\mathbb{R}^{N_1 \\times N_2}" + }, + { + "bbox": [ + 132, + 116, + 482, + 165 + ], + "type": "text", + "content": ". Each entry " + }, + { + "bbox": [ + 132, + 116, + 482, + 165 + ], + "type": "inline_equation", + "content": "s_{ij}" + }, + { + "bbox": [ + 132, + 116, + 482, + 165 + ], + "type": "text", + "content": " of the matrix is computed as the cosine similarity between the descriptors of every cross-view keypoint pair:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 202, + 173, + 481, + 199 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 173, + 481, + 199 + ], + "spans": [ + { + "bbox": [ + 202, + 173, + 481, + 199 + ], + "type": "interline_equation", + "content": "s _ {i j} = \\left(\\frac {d _ {i} \\cdot d _ {j}}{\\| d _ {i} \\| \\| d _ {j} \\|} + 1\\right) / 2, \\quad \\{(d _ {i}, d _ {j}) \\} \\subseteq D _ {1} \\times D _ {2}. \\tag {13}", + "image_path": "87efeecdf41fa961daeaccd7d0669fddf28e5d03d92f5cbf8594a72c61ade26e.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 206, + 479, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 206, + 479, + 253 + ], + "spans": [ + { + "bbox": [ + 131, + 206, + 479, + 253 + ], + "type": "text", + "content": "Keypoints with similar local features tend to possess similar descriptors, describing correlated semantic information of similar views. Thus, the cosine similarity captures the semantic similarities between two keypoints, and guides the framework to recognize the overlapping parts between the images." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 254, + 480, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 254, + 480, + 290 + ], + "spans": [ + { + "bbox": [ + 131, + 254, + 480, + 290 + ], + "type": "text", + "content": "We normalize the matrix " + }, + { + "bbox": [ + 131, + 254, + 480, + 290 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 131, + 254, + 480, + 290 + ], + "type": "text", + "content": " to [0,1] and element-wise multiply it with the cross-attention scores. Finally, the scores are used to extract and exchange crucial cross-view information from each others' values " + }, + { + "bbox": [ + 131, + 254, + 480, + 290 + ], + "type": "inline_equation", + "content": "v_{1}, v_{2}" + }, + { + "bbox": [ + 131, + 254, + 480, + 290 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 261, + 298, + 480, + 312 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 298, + 480, + 312 + ], + "spans": [ + { + "bbox": [ + 261, + 298, + 480, + 312 + ], + "type": "interline_equation", + "content": "A _ {2 1} ^ {\\prime \\top} = A _ {1 2} ^ {\\prime} = A _ {1 2} \\cdot S, \\tag {14}", + "image_path": "84d313b5a6385f881fba0a83ff7d1d8cee966a6d743edf73d15be3c40240984d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 251, + 316, + 480, + 329 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 316, + 480, + 329 + ], + "spans": [ + { + "bbox": [ + 251, + 316, + 480, + 329 + ], + "type": "interline_equation", + "content": "X _ {1} ^ {m - 1} = \\operatorname {S o f t m a x} \\left(A _ {1 2} ^ {\\prime}\\right) v _ {2}, \\tag {15}", + "image_path": "6ed3bf2b1beee349454900f825d79a9c358e64e2f5e31792fd09f53b5b87951c.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 251, + 331, + 480, + 346 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 331, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 251, + 331, + 480, + 346 + ], + "type": "interline_equation", + "content": "X _ {2} ^ {m - 1} = \\operatorname {S o f t m a x} \\left(A _ {2 1} ^ {\\prime}\\right) v _ {1}. \\tag {16}", + "image_path": "720d1692a4c58d6da25454349e347e6681fe05725a980c5d13f1d78b8058f7cb.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 354, + 479, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 354, + 479, + 426 + ], + "spans": [ + { + "bbox": [ + 131, + 354, + 479, + 426 + ], + "type": "text", + "content": "By capturing and incorporating relevant information from each image into the other, the implicit correspondences are established across the two images. As shown in Figs. 1a and 1b, with brighter dots representing higher cross-attention scores, SRPose establishes implicitly corresponding points densely in the overlapping areas of the two images, which is utilized to construct and solve an implicit version of the epipolar constraint Eq. (1)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 429, + 482, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 429, + 482, + 621 + ], + "spans": [ + { + "bbox": [ + 131, + 429, + 482, + 621 + ], + "type": "text", + "content": "Object Prompt: In the object-to-camera estimation, we employ an accessible user-provided object prompt in only one image to identify the target object. In contrast, traditional matching approaches require high-quality object segmentations in both images. An object prompt " + }, + { + "bbox": [ + 131, + 429, + 482, + 621 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 131, + 429, + 482, + 621 + ], + "type": "text", + "content": " is a bounding box consisting of the two coordinates of top-left and bottom-right points, which bounds the object " + }, + { + "bbox": [ + 131, + 429, + 482, + 621 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 131, + 429, + 482, + 621 + ], + "type": "text", + "content": " in one of the images " + }, + { + "bbox": [ + 131, + 429, + 482, + 621 + ], + "type": "inline_equation", + "content": "I_1" + }, + { + "bbox": [ + 131, + 429, + 482, + 621 + ], + "type": "text", + "content": ", making " + }, + { + "bbox": [ + 131, + 429, + 482, + 621 + ], + "type": "inline_equation", + "content": "I_1" + }, + { + "bbox": [ + 131, + 429, + 482, + 621 + ], + "type": "text", + "content": " as the reference view. Once the keypoints are extracted from " + }, + { + "bbox": [ + 131, + 429, + 482, + 621 + ], + "type": "inline_equation", + "content": "I_1" + }, + { + "bbox": [ + 131, + 429, + 482, + 621 + ], + "type": "text", + "content": ", the object prompt removes the keypoints outside of the bounding box in the reference view, eliminating most irrelevant keypoints. Thus, the descriptors and positions of keypoints within bounds, focusing on the target object " + }, + { + "bbox": [ + 131, + 429, + 482, + 621 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 131, + 429, + 482, + 621 + ], + "type": "text", + "content": ", will be later processed in the Similarity Matrix and other following modules. With the guided cross-attention modules, SRPose learns to identify highly correlated keypoints lying on the same object " + }, + { + "bbox": [ + 131, + 429, + 482, + 621 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 131, + 429, + 482, + 621 + ], + "type": "text", + "content": " from the entire query view " + }, + { + "bbox": [ + 131, + 429, + 482, + 621 + ], + "type": "inline_equation", + "content": "I_2" + }, + { + "bbox": [ + 131, + 429, + 482, + 621 + ], + "type": "text", + "content": ", excluding irrelevant background keypoints, as shown in Figs. 1d and 1e. With the object prompt, SRPose searches the implicit corresponding points on the target object and computes the relative 6D object pose transformation in the same process as the camera-to-world scenario." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 627, + 282, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 627, + 282, + 639 + ], + "spans": [ + { + "bbox": [ + 132, + 627, + 282, + 639 + ], + "type": "text", + "content": "3.4 Relative Pose Regressor" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 640, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 640, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 640, + 481, + 666 + ], + "type": "text", + "content": "After the embeddings " + }, + { + "bbox": [ + 131, + 640, + 481, + 666 + ], + "type": "inline_equation", + "content": "X_1^M, X_2^M" + }, + { + "bbox": [ + 131, + 640, + 481, + 666 + ], + "type": "text", + "content": " are computed via the previous modules, a MultiLayer Perceptron (MLP) is employed to regress the relative pose as our last step." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "R. Yin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "text", + "content": "The outputs of the previous attention layers " + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "inline_equation", + "content": "X_1^M, X_2^M" + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "text", + "content": " first undergo 2D average pooling, and then are concatenated together, resulting in the joint embedding " + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "inline_equation", + "content": "f \\in \\mathbb{R}^{2d}" + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "text", + "content": ", which carries the deep features and implicit correspondences across the two images. Finally, embedding " + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "text", + "content": " is input to the MLP, and the regressor solves the epipolar constraint Eq. (1) implicitly and regresses the relative pose, consisting of a rotation " + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "text", + "content": ", and a translation " + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "text", + "content": ". Following [78], we set the rotation output of the regressor to be a 6d vector " + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "inline_equation", + "content": "r \\in \\mathbb{R}^6" + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "text", + "content": ", which is then transformed into a rotation matrix " + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "inline_equation", + "content": "R \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "text", + "content": " through a partial Gram-Schmidt process. This technique, as argued in [78], enables a continuous representation of rotations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 230, + 231, + 241 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 230, + 231, + 241 + ], + "spans": [ + { + "bbox": [ + 131, + 230, + 231, + 241 + ], + "type": "text", + "content": "3.5 Loss Function" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 244, + 481, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 244, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 244, + 481, + 281 + ], + "type": "text", + "content": "SRPose is trained in a supervised manner. With the ground-truth rotation " + }, + { + "bbox": [ + 130, + 244, + 481, + 281 + ], + "type": "inline_equation", + "content": "R_{gt}" + }, + { + "bbox": [ + 130, + 244, + 481, + 281 + ], + "type": "text", + "content": ", we compute the Huber Loss [31] of the error in the rotation angle, which minimizes the angular error between the predicted and ground-truth rotations:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 209, + 289, + 481, + 316 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 289, + 481, + 316 + ], + "spans": [ + { + "bbox": [ + 209, + 289, + 481, + 316 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {R} \\left(I _ {1}, I _ {2}\\right) = \\mathcal {H} \\left(\\operatorname {a r c c o s} \\left(\\frac {\\operatorname {T r} \\left(R ^ {\\top} R _ {g t} - 1\\right)}{2}\\right)\\right), \\tag {17}", + "image_path": "8c14ae5b5fbbc03b65b5304a83ab9aed4b72290f64d13e273a1195a9ae0a25e5.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 323, + 482, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 323, + 482, + 372 + ], + "spans": [ + { + "bbox": [ + 131, + 323, + 482, + 372 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 323, + 482, + 372 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 131, + 323, + 482, + 372 + ], + "type": "text", + "content": " denotes the Huber Loss function. Similarly, given the ground-truth translation " + }, + { + "bbox": [ + 131, + 323, + 482, + 372 + ], + "type": "inline_equation", + "content": "t_{gt}" + }, + { + "bbox": [ + 131, + 323, + 482, + 372 + ], + "type": "text", + "content": ", the error is computed in both normalized and unnormalized forms. To further enhance accuracy, we also incorporate the angular error of translation, resulting in the following three loss terms:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 257, + 382, + 480, + 395 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 382, + 480, + 395 + ], + "spans": [ + { + "bbox": [ + 257, + 382, + 480, + 395 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {t} \\left(I _ {1}, I _ {2}\\right) = \\mathcal {H} \\left(t - t _ {g t}\\right), \\tag {18}", + "image_path": "8ea4bb78e74a067ff89a17478db3dfa417f05f4042abd0161c1d5cdd6310542e.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 238, + 398, + 480, + 423 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 398, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 238, + 398, + 480, + 423 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {t _ {n}} \\left(I _ {1}, I _ {2}\\right) = \\mathcal {H} \\left(\\frac {t}{\\| t \\|} - \\frac {t _ {g t}}{\\| t _ {g t} \\|}\\right), \\tag {19}", + "image_path": "fc7741eaeb6352aa582f76230df2297bc0716388c346304db3cb6c3719d61c2f.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 223, + 426, + 480, + 452 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 426, + 480, + 452 + ], + "spans": [ + { + "bbox": [ + 223, + 426, + 480, + 452 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {t _ {a}} \\left(I _ {1}, I _ {2}\\right) = \\mathcal {H} \\left(\\arccos \\left(\\frac {t \\cdot t _ {g t}}{\\| t \\| \\| t _ {g t} \\|}\\right)\\right). \\tag {20}", + "image_path": "e1c5a95c8d31e0ff24f1989f026ac5124a403cf6567f5c041170f27dd386cf05.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 459, + 481, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 459, + 481, + 483 + ], + "spans": [ + { + "bbox": [ + 131, + 459, + 481, + 483 + ], + "type": "text", + "content": "The above three loss functions supervise both the scale and direction of translations. We compute the final loss by the weighted sum of the three functions:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 217, + 495, + 480, + 509 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 495, + 480, + 509 + ], + "spans": [ + { + "bbox": [ + 217, + 495, + 480, + 509 + ], + "type": "interline_equation", + "content": "\\mathcal {L} \\left(I _ {1}, I _ {2}\\right) = \\mathcal {L} _ {R} + \\lambda_ {t} \\mathcal {L} _ {t} + \\lambda_ {t _ {n}} \\mathcal {L} _ {t _ {n}} + \\lambda_ {t _ {a}} \\mathcal {L} _ {t _ {a}}, \\tag {21}", + "image_path": "a725d1eb7144b5dff8f41e5a698f047bae7fc545959d2e72445bc44e572cd5e8.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 513, + 479, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 513, + 479, + 526 + ], + "spans": [ + { + "bbox": [ + 130, + 513, + 479, + 526 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 513, + 479, + 526 + ], + "type": "inline_equation", + "content": "\\lambda_t, \\lambda_{t_n}, \\lambda_{t_a}" + }, + { + "bbox": [ + 130, + 513, + 479, + 526 + ], + "type": "text", + "content": " are scalars used to balance the weights of four different losses." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 531, + 279, + 543 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 531, + 279, + 543 + ], + "spans": [ + { + "bbox": [ + 131, + 531, + 279, + 543 + ], + "type": "text", + "content": "3.6 Implementation Details" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "text", + "content": "SRPose is first trained on ScanNet [14] and then fine-tuned on other datasets to achieve better performance. For all datasets, we resize the images to " + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "inline_equation", + "content": "640 \\times 480" + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "text", + "content": ", and extract 1,024 sparse keypoints using SuperPoint [16] for training and 2,048 for evaluation. Other sparse keypoint detectors can also be used for this step. We utilize the pre-trained SuperPoint provided by the official source and freeze its weights in the training stage. The remaining learnable modules are initialized with random weights. In practice, SRPose consists of 6 guided attention layers, with 4 attention heads in each attention module. The dimension " + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "text", + "content": " of keypoint descriptors and embeddings is set to 256. The rotation and translation are estimated using two separate 3-layer MLPs. All three balancing scalars " + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\lambda_{t}, \\lambda_{t_{n}}, \\lambda_{t_{a}}" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 171, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 171, + 91, + 448, + 103 + ], + "type": "text", + "content": "SRPose: Two-view Relative Pose Estimation with Sparse Keypoints" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 157, + 149, + 454, + 276 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 148 + ], + "type": "text", + "content": "Table 1: Relative Pose Estimation on Matterport [10]. Without depth, matcher-based approaches are incapable of scaled translation estimation, while SRPose achieves high accuracy in translation scales using only RGB inputs." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 157, + 149, + 454, + 276 + ], + "lines": [ + { + "bbox": [ + 157, + 149, + 454, + 276 + ], + "spans": [ + { + "bbox": [ + 157, + 149, + 454, + 276 + ], + "type": "table", + "html": "
CategoryMethodRot. (°)Trans. (m)
Med.↓Avg.↓≤ 30° ↑Med.↓Avg.↓≤ 1m↑
SparseSuperGlue [55]3.8824.1777.8n/an/an/a
SGMNet [12]1.0621.3280.9n/an/an/a
LightGlue [39]1.3222.4580.0n/an/an/a
DenseLoFTR [59]0.7111.1190.5n/an/an/a
ASpanFormer [13]3.7331.4575.7n/an/an/a
DKM [18]0.4612.8989.0n/an/an/a
RegressorSparsePlanes [32]7.3322.7883.40.631.2566.6
PlaneFormers [2]5.9622.2083.80.661.1966.8
8point [53]8.0119.1385.40.641.0167.4
NOPE-SAC [61]2.7714.3789.00.520.9473.2
SRPose2.6511.1291.60.270.6183.7
", + "image_path": "8e762daccbb5361e8dc0327197506b15e988cca63fee8abb4793ebaf17ff9779.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 290, + 481, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 290, + 481, + 352 + ], + "spans": [ + { + "bbox": [ + 130, + 290, + 481, + 352 + ], + "type": "text", + "content": "are set to 1. On ScanNet, our framework is trained using AdamW optimizer [43] and follows 1cycle learning rate policy [58] with a maximum learning rate of " + }, + { + "bbox": [ + 130, + 290, + 481, + 352 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 130, + 290, + 481, + 352 + ], + "type": "text", + "content": " for 500 epochs. We fix the batch size to 32 for training. The pretraining requires 120 hours utilizing 8 RTX 3090 GPUs. More implementation details can be found in the supplementary materials." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 360, + 230, + 374 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 360, + 230, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 360, + 230, + 374 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 381, + 481, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 381, + 481, + 431 + ], + "spans": [ + { + "bbox": [ + 130, + 381, + 481, + 431 + ], + "type": "text", + "content": "We evaluate the performance of SRPose for two-view relative pose estimation in both the camera-to-world and object-to-camera scenarios. Additionally, we assess its effectiveness in the application of map-free relocalization. Finally, we conduct an ablation study to analyze the roles of the components we proposed." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 441, + 334, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 441, + 334, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 441, + 334, + 453 + ], + "type": "text", + "content": "4.1 Camera-to-World Pose Estimation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 460, + 481, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 481, + 580 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 481, + 580 + ], + "type": "text", + "content": "Setup: In the camera-to-world scenario, we evaluate our framework's performance on Matterport [10] and ScanNet [14]. Following [32], for Matterport, we report the median and mean of the translation and rotation error, and the translation/rotation accuracy at the thresholds of " + }, + { + "bbox": [ + 130, + 460, + 481, + 580 + ], + "type": "inline_equation", + "content": "1\\mathrm{m} / 30^{\\circ}" + }, + { + "bbox": [ + 130, + 460, + 481, + 580 + ], + "type": "text", + "content": ", respectively. For ScanNet, we adopt the metrics from [55] to compute the AUC of the pose error at " + }, + { + "bbox": [ + 130, + 460, + 481, + 580 + ], + "type": "inline_equation", + "content": "5^{\\circ}" + }, + { + "bbox": [ + 130, + 460, + 481, + 580 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 460, + 481, + 580 + ], + "type": "inline_equation", + "content": "10^{\\circ}" + }, + { + "bbox": [ + 130, + 460, + 481, + 580 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 460, + 481, + 580 + ], + "type": "inline_equation", + "content": "20^{\\circ}" + }, + { + "bbox": [ + 130, + 460, + 481, + 580 + ], + "type": "text", + "content": ". The pose error here is the maximum angular error in rotation and translation. We also report an analysis of the time consumption on the ScanNet-1500 test set. For all matcher-based approaches, we use the implementation of RANSAC [23] from OpenCV [7] to recover relative poses. All methods are compared on a device with RTX 4090, i9-13900K, and 128GB memory." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 590, + 481, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 590, + 481, + 676 + ], + "spans": [ + { + "bbox": [ + 130, + 590, + 481, + 676 + ], + "type": "text", + "content": "Baselines: We choose three categories of approaches as our baselines, including sparse keypoint-based matchers, dense matchers, and deep learning-based relative pose regressors. Results of all the baselines are reported according to the official papers or implementations. All sparse matchers and SRPose employ SuperPoint [16] as their sparse keypoint detector. All regressors are trained on the specific dataset, i.e. Matterport or ScanNet, before evaluation. LightGlue [39] is trained on MegaDepth [37], while all other matchers are trained on ScanNet." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "R. Yin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 153, + 149, + 458, + 253 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 479, + 147 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 479, + 147 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 479, + 147 + ], + "type": "text", + "content": "Table 2: Relative Pose Estimation on ScanNet [14]. SRPose outperforms state-of-the-art regressors and sparse matchers in terms of various metrics and reduces computational time greatly due to the riddance of robust estimators." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 153, + 149, + 458, + 253 + ], + "lines": [ + { + "bbox": [ + 153, + 149, + 458, + 253 + ], + "spans": [ + { + "bbox": [ + 153, + 149, + 458, + 253 + ], + "type": "table", + "html": "
CategoryMethodPose estimationAUCTime (ms)
@5°@10°@20°MatchTotal
SparseSuperGlue [55]16.233.851.8102.5361.9
SGMNet [12]15.432.148.3111.7439.7
LightGlue [39]16.433.650.231.5269.7
DenseLoFTR [59]22.040.857.644.4292.4
ASpanFormer [13]25.646.063.359.6370.9
DKM [18]29.450.768.3235.0491.6
RegressorMap-free [3]2.7011.529.0n/a12.7
SRPose13.334.356.8n/a28.5
", + "image_path": "15e1b6b67f7edc068549a9a99ddd86a987d89ad18fd728a478b3f6b2e2534f08.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 149, + 255, + 471, + 367 + ], + "blocks": [ + { + "bbox": [ + 149, + 255, + 471, + 367 + ], + "lines": [ + { + "bbox": [ + 149, + 255, + 471, + 367 + ], + "spans": [ + { + "bbox": [ + 149, + 255, + 471, + 367 + ], + "type": "image", + "image_path": "ab9c7615a46a7b7e4a9ff956854a87d6ac510e75ae64c2395e1639f864738284.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 369, + 480, + 391 + ], + "lines": [ + { + "bbox": [ + 130, + 369, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 480, + 391 + ], + "type": "text", + "content": "Fig. 4: Time consumption comparison on ScanNet [14]. Regressors, including SRPose, achieve much higher computational efficacy than all matchers." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 403, + 482, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 548 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 548 + ], + "type": "text", + "content": "Results: Table 1 demonstrates that SRPose greatly outperforms the existing deep learning-based regressors across all metrics on Matterport. Our framework achieves the most accurate scaled translation estimation, which is unattainable for matcher-based approaches lacking depth information. Table 2 shows SRPose achieves competitive or superior performance compared to all sparse matchers and the regressor and on ScanNet. Although dense matchers outperform our framework in accuracy, SRPose offers a significant advantage in computational speed compared to all the matcher-based approaches by replacing robust estimators with direct regression. As shown in Fig. 4, matcher-based approaches spend a substantial proportion of time on recovering pose using robust estimators after matching. In contrast, SRPose consumes even less time than the matching stage of all matchers, saving at least 200ms in the total time." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 555, + 338, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 555, + 338, + 567 + ], + "spans": [ + { + "bbox": [ + 131, + 555, + 338, + 567 + ], + "type": "text", + "content": "4.2 Object-to-Camera Pose Estimation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": "Setup: In the object-to-camera scenario, we evaluate SRPose on HO3D [26]. HO3D contains videos of hand-holding objects of 21 categories from the YCB dataset [9]. The depth captured by the stereo depth camera and the object segmentation of each frame in the videos are also provided. We randomly select image pairs with small object rotation transformations for training and evaluation. More details can be found in the supplementary materials. Following [30, 69], we adopt the widely-used Average Distance (ADD, ADD-S) metric for evaluation. To measure the smaller scales in object pose transformation, we set the" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 171, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 171, + 91, + 447, + 102 + ], + "type": "text", + "content": "SRPose: Two-view Relative Pose Estimation with Sparse Keypoints" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 149, + 150, + 462, + 225 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 148 + ], + "type": "text", + "content": "Table 3: Relative object pose estimation on HO3D [26]. SRPose achieves state-of-the-art performance in object-to-camera estimation without additional depth information, enabling scale pose estimation from only RGB inputs." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 149, + 150, + 462, + 225 + ], + "lines": [ + { + "bbox": [ + 149, + 150, + 462, + 225 + ], + "spans": [ + { + "bbox": [ + 149, + 150, + 462, + 225 + ], + "type": "table", + "html": "
MethodRot. (°)Trans. (cm)ADD↑ADD-S↑
Med.↓Avg.↓≤ 30°↑Med.↓Avg.↓≤ 10cm↑
SGMNet [12]24.129.062.713.917.536.816.831.3
LightGlue [39]24.428.761.713.516.338.217.231.2
LoFTR [59]29.257.950.913.916.435.616.730.9
DKM [18]23.325.664.813.615.439.817.833.3
SRPose8.911.495.45.98.073.936.456.8
", + "image_path": "54d1452d1ed87108a884e295106dc986a82fc85917e002d8e532946aec3df694.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 234, + 479, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 234, + 479, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 234, + 479, + 283 + ], + "type": "text", + "content": "threshold of metrics to " + }, + { + "bbox": [ + 130, + 234, + 479, + 283 + ], + "type": "inline_equation", + "content": "10\\mathrm{cm}" + }, + { + "bbox": [ + 130, + 234, + 479, + 283 + ], + "type": "text", + "content": ", which differs from the 1m threshold used in the camera-to-world scenario. Specifically, we calculate the AUC using both ADD and ADD-S with the threshold set to " + }, + { + "bbox": [ + 130, + 234, + 479, + 283 + ], + "type": "inline_equation", + "content": "10\\mathrm{cm}" + }, + { + "bbox": [ + 130, + 234, + 479, + 283 + ], + "type": "text", + "content": ". We also report the median, mean error of translation/rotation, and their accuracy at " + }, + { + "bbox": [ + 130, + 234, + 479, + 283 + ], + "type": "inline_equation", + "content": "10\\mathrm{cm} / 30^{\\circ}" + }, + { + "bbox": [ + 130, + 234, + 479, + 283 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 288, + 480, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 288, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 130, + 288, + 480, + 385 + ], + "type": "text", + "content": "Baselines: We only choose matchers listed in Sec. 4.1 as our baselines, for no deep learning-based regressors are available in this two-view relative object pose estimation task. We utilize the stereo depth offered in the dataset to enable matcher-based approaches to estimate scaled relative poses via Orthogonal Procrustes [20], while SRPose only requires RGB images as inputs. The segmentations in the dataset are adopted for matcher-based approaches to identify the target object in both images, while SRPose utilizes the bounding box of the segmentation in the first image as the object prompt." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 391, + 480, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 391, + 480, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 480, + 462 + ], + "type": "text", + "content": "Results: As shown in Table 3, SRPose outperforms state-of-the-art matcher-based approaches significantly. SRPose exhibits superior accuracy in relative 6D object pose estimation between two images using the object prompts in the reference view. The low quality of stereo depth leads to limited performance in scaled pose estimation of the matcher-based approaches, while SRPose can learn scaled information to achieve lower estimation errors." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 468, + 319, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 319, + 480 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 319, + 480 + ], + "type": "text", + "content": "4.3 Map-Free Visual Relocalization" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 482, + 480, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 482, + 480, + 542 + ], + "spans": [ + { + "bbox": [ + 130, + 482, + 480, + 542 + ], + "type": "text", + "content": "We evaluate the application of SRPose on Niantic [3], a dataset for the map-free visual relocalization task. Following [3], we adopt the Virtual Correspondence Reprojection Error (VCRE) and the median error of translation and rotation as the metrics. Specifically, we compute the precision of VCRE at 90 pixels, and the precision of relocalization at " + }, + { + "bbox": [ + 130, + 482, + 480, + 542 + ], + "type": "inline_equation", + "content": "0.25\\mathrm{m}" + }, + { + "bbox": [ + 130, + 482, + 480, + 542 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 482, + 480, + 542 + ], + "type": "inline_equation", + "content": "5^{\\circ}" + }, + { + "bbox": [ + 130, + 482, + 480, + 542 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 548, + 480, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 548, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 130, + 548, + 480, + 597 + ], + "type": "text", + "content": "Baselines: We still select the three categories of approaches as our baselines. For **matching-based approaches**, we estimate the pose by recovering from the essential matrix. Then we use the depth estimated by DPT [51], a depth prediction model, to provide the matchers with the scale information." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 603, + 481, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 603, + 481, + 675 + ], + "spans": [ + { + "bbox": [ + 130, + 603, + 481, + 675 + ], + "type": "text", + "content": "Results: Table 4 shows SRPose outperforms other sparse matchers and the regression method in map-free visual relocalization. Although the state-of-the-art dense matcher RoMa achieves higher precision in relocalization using scales estimated by an additional depth prediction model, SRPose still outperforms in VCRE. Our framework exhibits competitive performance compared to state-of-the-art approaches in the map-free visual relocalization task." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "R. Yin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 173, + 144, + 436, + 232 + ], + "blocks": [ + { + "bbox": [ + 130, + 108, + 482, + 141 + ], + "lines": [ + { + "bbox": [ + 130, + 108, + 482, + 141 + ], + "spans": [ + { + "bbox": [ + 130, + 108, + 482, + 141 + ], + "type": "text", + "content": "Table 4: Map-free visual relocalization results on Niantic [3]. SRPose achieves competitive performance compared to state-of-the-art approaches without relying on an additional depth prediction model to assist with scaled estimation." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 173, + 144, + 436, + 232 + ], + "lines": [ + { + "bbox": [ + 173, + 144, + 436, + 232 + ], + "spans": [ + { + "bbox": [ + 173, + 144, + 436, + 232 + ], + "type": "table", + "html": "
CategoryMethodVCRE \nPrec.↑ / Med.↓Med. (m / °) \nTrans.↓ / Rot.↓Prec.↑
SparseSIFT [44]25.0 / 222.82.93 / 61.410.3
+DPT [51]SuperGlue [55]36.1 / 160.31.88 / 25.416.8
DenseLoFTR [59]34.7 / 167.61.98 / 30.515.4
+DPT [51]RoMa [19]45.6 / 128.81.23 / 11.122.8
RegressorMap-free [3]40.2 / 147.11.68 / 22.96.0
SRPose46.4 / 127.71.37 / 17.216.9
", + "image_path": "687710288f4779109cf2a99d5ad755d24cab511e0473bfa8d1f3a956e6b5f00a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 173, + 258, + 440, + 333 + ], + "blocks": [ + { + "bbox": [ + 132, + 234, + 481, + 257 + ], + "lines": [ + { + "bbox": [ + 132, + 234, + 481, + 257 + ], + "spans": [ + { + "bbox": [ + 132, + 234, + 481, + 257 + ], + "type": "text", + "content": "Table 5: Ablation study on Matterport [10]. The prior knowledge-guided attention layers assist in establishing implicit cross-view correspondences." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 173, + 258, + 440, + 333 + ], + "lines": [ + { + "bbox": [ + 173, + 258, + 440, + 333 + ], + "spans": [ + { + "bbox": [ + 173, + 258, + 440, + 333 + ], + "type": "table", + "html": "
MethodTrans. (m)Rot. (°)
Avg.↓≤ 1m↑Avg.↓≤ 30°↑
8point [53]1.0167.419.1385.4
SRPose (full)0.8474.214.3288.9
1) no prior knowledge guidance0.9769.616.9186.8
2) replace all cross-attn. with self-attn.1.9026.148.042.0
3) no position encoding1.1761.021.3981.3
", + "image_path": "0c9b21b3f8fe1b76bd0092991fa89c502354f15e0b098ba03348acdcce0ab3dc.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 341, + 201, + 353 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 341, + 201, + 353 + ], + "spans": [ + { + "bbox": [ + 132, + 341, + 201, + 353 + ], + "type": "text", + "content": "4.4 Insights" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 354, + 298, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 298, + 474 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 298, + 474 + ], + "type": "text", + "content": "Ablation study: We evaluate the effectiveness of the different components in SRPose by Matterport [10] and MegaDepth [37] datasets. Specifically, we train four different variants initialized with random weights on Matterport by removing the three different designs we proposed to validate their effectiveness. As shown in Table 5: 1) Removing the prior knowledge" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 307, + 400, + 478, + 466 + ], + "blocks": [ + { + "bbox": [ + 304, + 355, + 482, + 399 + ], + "lines": [ + { + "bbox": [ + 304, + 355, + 482, + 399 + ], + "spans": [ + { + "bbox": [ + 304, + 355, + 482, + 399 + ], + "type": "text", + "content": "Table 6: Ablation study on MegaDepth [37]. Intrinsic-calibration position encoder enables adaption on varying image sizes and camera intrinsics." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 400, + 478, + 466 + ], + "lines": [ + { + "bbox": [ + 307, + 400, + 478, + 466 + ], + "spans": [ + { + "bbox": [ + 307, + 400, + 478, + 466 + ], + "type": "table", + "html": "
MethodPose est.AUC
@5°@10°@20°
Map-free [3]2.69.322.9
SRPose (full)16.636.058.0
1) no intrinsic calibration1.58.524.3
2) no position encoding1.06.218.9
", + "image_path": "6ba9e38bf474ea0eeb9f5f16239374e848bff2cb4d51d9b14fc867b5897104f5.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 474, + 482, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 569 + ], + "type": "text", + "content": "edge guidance of keypoint similarities in the cross-attention modules results in a moderate decrease in accuracy, as it assists SRPose in seeking the cross-view correspondences. 2) Replacing all cross-attention modules with self-attention modules leads to a significant accuracy drop, as the capability to establish implicit cross-view correspondences through cross-attention is eliminated. 3) Removing the entire position encoder leads to a moderately lower accuracy due to the absence of keypoint position information. The experiment highlights the crucial roles of the three designs of SRPose in accurate relative pose estimation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Moreover, we train three different variants initialized with random weights on MegaDepth, a dataset that consists of images with different sizes and camera intrinsics, to further validate the effectiveness of the intrinsic-calibration position encoder, with results shown in Table 6. 1) Without enforcing intrinsic calibration on keypoint coordinates, SRPose only achieves comparable performance to Map-free, both inadaptable to varying image sizes and camera intrinsics in MegaDepth. 2) Removing the entire IC Position Encoder results in a significant drop in accuracy as expected owing to the elimination of critical position infor" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 171, + 90, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 90, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 171, + 90, + 448, + 103 + ], + "type": "text", + "content": "SRPose: Two-view Relative Pose Estimation with Sparse Keypoints" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": "mation. By comparing the full framework with the variants lacking IC and the entire position encoder, we demonstrate the essential roles of the two designs in adapting different image sizes and camera intrinsics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 156, + 482, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 156, + 482, + 204 + ], + "spans": [ + { + "bbox": [ + 130, + 156, + 482, + 204 + ], + "type": "text", + "content": "Visualizing attention: In Fig. 5, we visualize the cross-attention scores " + }, + { + "bbox": [ + 130, + 156, + 482, + 204 + ], + "type": "inline_equation", + "content": "A_{12}^{\\prime}" + }, + { + "bbox": [ + 130, + 156, + 482, + 204 + ], + "type": "text", + "content": " and compare SRPose without prior knowledge guidance to its full designs. We also draw the connections of highly-attended cross-view keypoint pairs, which represent the implicit correspondences established by SRPose." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 142, + 211, + 468, + 453 + ], + "blocks": [ + { + "bbox": [ + 142, + 211, + 468, + 453 + ], + "lines": [ + { + "bbox": [ + 142, + 211, + 468, + 453 + ], + "spans": [ + { + "bbox": [ + 142, + 211, + 468, + 453 + ], + "type": "image", + "image_path": "8b3dd6dce6dd434d7266560f41d224db217cd6ec6f8aa3d4a57462c2037643ad.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 456, + 482, + 502 + ], + "lines": [ + { + "bbox": [ + 130, + 456, + 482, + 502 + ], + "spans": [ + { + "bbox": [ + 130, + 456, + 482, + 502 + ], + "type": "text", + "content": "Fig.5: Visualization of implicit cross-view correspondences. Dots and lines drawn with brighter colors represent higher cross-view attention scores. The prior knowledge guidance enhances attention to the overlapping areas, removing the irrelevant cross-view connections, and assisting in establishing implicit correspondences." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 506, + 220, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 506, + 220, + 517 + ], + "spans": [ + { + "bbox": [ + 131, + 506, + 220, + 517 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "This paper presents SRPose, a novel deep learning-based regressor utilizing sparse keypoints for two-view relative pose estimation in camera-to-world and object-to-camera scenarios. As our key innovations, SRPose extracts keypoints with sparse keypoint detectors and employs an intrinsic-calibration position encoder to enable adaptability to different image sizes and camera intrinsics. Further, our proposed promptable prior knowledge-guided attention layers establish implicit correspondences to estimate the relative pose under the epipolar constraint in both scenarios. By directly regressing the rotation and translation, SRPose achieves a significant decrease in computational time, with a minimum time reduction of 200ms. Extensive experiments demonstrate that SRPose achieves state-of-the-art performance in terms of accuracy and speed in the two scenarios, as well as in the map-free visual relocalization task." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "R. Yin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 481, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 481, + 174 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 481, + 174 + ], + "type": "text", + "content": "This work was supported by Shanghai Municipal Science and Technology Major Project (2021SHZDZX0102) and the Fundamental Research Funds for the Central Universities." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 194, + 197, + 208 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 194, + 197, + 208 + ], + "spans": [ + { + "bbox": [ + 133, + 194, + 197, + 208 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 220, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 138, + 220, + 480, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 220, + 480, + 243 + ], + "spans": [ + { + "bbox": [ + 138, + 220, + 480, + 243 + ], + "type": "text", + "content": "1. Abouelnaga, Y., Bui, M., Ilic, S.: Distillpose: lightweight camera localization using auxiliary learning. In: IROS (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 244, + 480, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 244, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 138, + 244, + 480, + 266 + ], + "type": "text", + "content": "2. Agarwala, S., Jin, L., Rockwell, C., Fouhey, D.F.: Planeformers: From sparse view planes to 3d reconstruction. In: ECCV (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 266, + 480, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 266, + 480, + 299 + ], + "spans": [ + { + "bbox": [ + 138, + 266, + 480, + 299 + ], + "type": "text", + "content": "3. Arnold, E., Wynn, J., Vicente, S., Garcia-Hernando, G., Monszpart, A., Prisacariu, V., Turmukhambetov, D., Brachmann, E.: Map-free visual relocalization: Metric pose relative to a single image. In: ECCV (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 300, + 480, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 300, + 480, + 321 + ], + "spans": [ + { + "bbox": [ + 138, + 300, + 480, + 321 + ], + "type": "text", + "content": "4. Balntas, V., Li, S., Prisacariu, V.: Relocnet: Continuous metric learning relocalisation using neural nets. In: ECCV (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 322, + 480, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 322, + 480, + 343 + ], + "spans": [ + { + "bbox": [ + 138, + 322, + 480, + 343 + ], + "type": "text", + "content": "5. Bay, H., Tuytelaars, T., Van Gool, L.: Surf: Speeded up robust features. In: ECCV (2006)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 344, + 480, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 344, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 138, + 344, + 480, + 365 + ], + "type": "text", + "content": "6. Bhowmik, A., Gumhold, S., Rother, C., Brachmann, E.: Reinforced feature points: Optimizing feature detection and description for a high-level task. In: CVPR (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 366, + 480, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 366, + 480, + 376 + ], + "spans": [ + { + "bbox": [ + 138, + 366, + 480, + 376 + ], + "type": "text", + "content": "7. Bradski, G.: The OpenCV Library. Dr. Dobb's Journal of Software Tools (2000)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 377, + 480, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 377, + 480, + 399 + ], + "spans": [ + { + "bbox": [ + 138, + 377, + 480, + 399 + ], + "type": "text", + "content": "8. Cai, R., Hariharan, B., Snavely, N., Averbuch-Elor, H.: Extreme rotation estimation using dense correlation volumes. In: CVPR (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 399, + 480, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 399, + 480, + 431 + ], + "spans": [ + { + "bbox": [ + 138, + 399, + 480, + 431 + ], + "type": "text", + "content": "9. Calli, B., Singh, A., Walsman, A., Srinivasa, S., Abbeel, P., Dollar, A.M.: The ycb object and model set: Towards common benchmarks for manipulation research. In: ICAR (2015)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 433, + 480, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 433, + 480, + 465 + ], + "spans": [ + { + "bbox": [ + 133, + 433, + 480, + 465 + ], + "type": "text", + "content": "10. Chang, A., Dai, A., Funkhouser, T., Halber, M., Niebner, M., Savva, M., Song, S., Zeng, A., Zhang, Y.: Matterport3d: Learning from rgb-d data in indoor environments. In: 3DV (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 466, + 480, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 466, + 480, + 487 + ], + "spans": [ + { + "bbox": [ + 133, + 466, + 480, + 487 + ], + "type": "text", + "content": "11. Chang, J., Yu, J., Zhang, T.: Structured epipolar matcher for local feature matching. In: CVPR (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 488, + 480, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 488, + 480, + 510 + ], + "spans": [ + { + "bbox": [ + 133, + 488, + 480, + 510 + ], + "type": "text", + "content": "12. Chen, H., Luo, Z., Zhang, J., Zhou, L., Bai, X., Hu, Z., Tai, C.L., Quan, L.: Learning to match features with seeded graph matching network. In: ICCV (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 510, + 480, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 510, + 480, + 542 + ], + "spans": [ + { + "bbox": [ + 133, + 510, + 480, + 542 + ], + "type": "text", + "content": "13. Chen, H., Luo, Z., Zhou, L., Tian, Y., Zhen, M., Fang, T., McKinnon, D., Tsin, Y., Quan, L.: Aspanformer: Detector-free image matching with adaptive span transformer. In: ECCV (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 544, + 480, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 544, + 480, + 565 + ], + "spans": [ + { + "bbox": [ + 133, + 544, + 480, + 565 + ], + "type": "text", + "content": "14. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: CVPR (2017)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 133, + 566, + 480, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 566, + 480, + 587 + ], + "spans": [ + { + "bbox": [ + 133, + 566, + 480, + 587 + ], + "type": "text", + "content": "15. DeTone, D., Malisiewicz, T., Rabinovich, A.: Toward geometric deep slam. arXiv preprint arXiv:1707.07410 (2017)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 133, + 588, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 588, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 133, + 588, + 480, + 610 + ], + "type": "text", + "content": "16. DeTone, D., Malisiewicz, T., Rabinovich, A.: Superpoint: Self-supervised interest point detection and description. In: CVPR (2018)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 133, + 611, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 611, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 133, + 611, + 480, + 643 + ], + "type": "text", + "content": "17. Dusmanu, M., Rocco, I., Pajdla, T., Pollefeys, M., Sivic, J., Torii, A., Sattler, T.: D2-net: A trainable cnn for joint description and detection of local features. In: CVPR (2019)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 133, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 643, + 480, + 665 + ], + "type": "text", + "content": "18. Edstedt, J., Athanasiadis, I., Wadenback, M., Felsberg, M.: Dkm: Dense kernelized feature matching for geometry estimation. In: CVPR (2023)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 172, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 172, + 91, + 447, + 102 + ], + "type": "text", + "content": "SRPose: Two-view Relative Pose Estimation with Sparse Keypoints" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "19. Edstedt, J., Sun, Q., Bokman, G., Wadenback, M., Felsberg, M.: RoMa: Robust Dense Feature Matching. arXiv preprint arXiv:2305.15404 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 480, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 480, + 170 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 480, + 170 + ], + "type": "text", + "content": "20. Eggert, D.W., Lorusso, A., Fisher, R.B.: Estimating 3-d rigid body transformations: a comparison of four major algorithms. Machine vision and applications (1997)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 171, + 480, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 171, + 480, + 191 + ], + "spans": [ + { + "bbox": [ + 132, + 171, + 480, + 191 + ], + "type": "text", + "content": "21. En, S., Lechery, A., Jurie, F.: Rpnet: An end-to-end network for relative camera pose estimation. In: ECCV (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 192, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 192, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 192, + 480, + 224 + ], + "type": "text", + "content": "22. Fan, Z., Pan, P., Wang, P., Jiang, Y., Xu, D., Jiang, H., Wang, Z.: Pope: 6-dof promptable pose estimation of any object, in any scene, with one reference. arXiv preprint arXiv:2305.15727 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 225, + 480, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 225, + 480, + 256 + ], + "spans": [ + { + "bbox": [ + 132, + 225, + 480, + 256 + ], + "type": "text", + "content": "23. Fischler, M.A., Bolles, R.C.: Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. COMMUN ACM (1981)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 258, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 258, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 258, + 480, + 267 + ], + "type": "text", + "content": "24. Gleize, P., Wang, W., Feiszli, M.: Silk: Simple learned keypoints. In: ICCV (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 268, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 268, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 268, + 480, + 289 + ], + "type": "text", + "content": "25. Goodwin, W., Vaze, S., Havoutis, I., Posner, I.: Zero-shot category-level object pose estimation. In: ECCV (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 290, + 480, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 290, + 480, + 310 + ], + "spans": [ + { + "bbox": [ + 132, + 290, + 480, + 310 + ], + "type": "text", + "content": "26. Hampali, S., Rad, M., Oberweger, M., Lepetit, V.: Honnotate: A method for 3d annotation of hand and object poses. In: CVPR (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 311, + 427, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 311, + 427, + 320 + ], + "spans": [ + { + "bbox": [ + 132, + 311, + 427, + 320 + ], + "type": "text", + "content": "27. Hartley, R.I.: In defense of the eight-point algorithm. TPAMI (1997)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 322, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 322, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 322, + 480, + 342 + ], + "type": "text", + "content": "28. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 343, + 480, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 480, + 363 + ], + "type": "text", + "content": "29. He, X., Sun, J., Wang, Y., Huang, D., Bao, H., Zhou, X.: Onepose++: Keypoint-free one-shot object pose estimation without cad models. NeurIPS (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 364, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 480, + 396 + ], + "type": "text", + "content": "30. Hinterstoisser, S., Lepetit, V., Ilic, S., Holzer, S., Bradski, G., Konolige, K., Navab, N.: Model based training, detection and pose estimation of texture-less 3d objects in heavily cluttered scenes. In: ACCV (2013)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 396, + 480, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 480, + 417 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 480, + 417 + ], + "type": "text", + "content": "31. Huber, P.J.: Robust estimation of a location parameter. The Annals of Mathematical Statistics (1964)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 418, + 480, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 480, + 438 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 480, + 438 + ], + "type": "text", + "content": "32. Jin, L., Qian, S., Owens, A., Fouhey, D.F.: Planar surface reconstruction from sparse views. In: ICCV (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 439, + 480, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 439, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 132, + 439, + 480, + 471 + ], + "type": "text", + "content": "33. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 472, + 480, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 480, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 480, + 493 + ], + "type": "text", + "content": "34. Laskar, Z., Melekhov, I., Kalia, S., Kannala, J.: Camera relocalization by computing pairwise relative poses using convolutional neural network. In: ICCVW (2017)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 494, + 480, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 494, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 494, + 480, + 514 + ], + "type": "text", + "content": "35. Li, K., Wang, L., Liu, L., Ran, Q., Xu, K., Guo, Y.: Decoupling makes weakly supervised local feature better. In: CVPR (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 514, + 480, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 480, + 535 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 480, + 535 + ], + "type": "text", + "content": "36. Li, X., Han, K., Li, S., Prisacariu, V.: Dual-resolution correspondence networks. NeurIPS (2020)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 536, + 480, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 480, + 556 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 480, + 556 + ], + "type": "text", + "content": "37. Li, Z., Snavely, N.: Megadepth: Learning single-view depth prediction from internet photos. In: CVPR (2018)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 557, + 480, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 480, + 589 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 480, + 589 + ], + "type": "text", + "content": "38. Lin, A., Zhang, J.Y., Ramanan, D., Tulsiani, S.: Relpose++: Recovering 6d poses from sparse-view observations. In: 2024 International Conference on 3D Vision (3DV). pp. 106-115. IEEE (2024)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 590, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 480, + 610 + ], + "type": "text", + "content": "39. Lindenberger, P., Sarlin, P.E., Pollefeys, M.: LightGlue: Local Feature Matching at Light Speed. In: ICCV (2023)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "text", + "content": "40. Liu, Y., Shen, Z., Lin, Z., Peng, S., Bao, H., Zhou, X.: Gift: Learning transformation-invariant dense visual descriptors via group cnns. NeurIPS (2019)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "text", + "content": "41. Liu, Y., Wen, Y., Peng, S., Lin, C., Long, X., Komura, T., Wang, W.: Gen6d: Generalizable model-free 6-dof object pose estimation from rgb images. In: ECCV (2022)" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 218, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 218, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 218, + 100 + ], + "type": "text", + "content": "R. Yin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "text", + "content": "42. Longuet-Higgins, H.C.: A computer algorithm for reconstructing a scene from two projections. Nature (1981)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 150 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 150 + ], + "type": "text", + "content": "43. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: ICLR (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 151, + 480, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 151, + 480, + 161 + ], + "spans": [ + { + "bbox": [ + 133, + 151, + 480, + 161 + ], + "type": "text", + "content": "44. Lowe, D.G.: Distinctive image features from scale-invariant keypoints. IJCV (2004)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 162, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 162, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 162, + 480, + 194 + ], + "type": "text", + "content": "45. Luo, Z., Zhou, L., Bai, X., Chen, H., Zhang, J., Yao, Y., Li, S., Fang, T., Quan, L.: Aslfeat: Learning local features of accurate shape and localization. In: CVPR (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 194, + 480, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 194, + 480, + 216 + ], + "spans": [ + { + "bbox": [ + 133, + 194, + 480, + 216 + ], + "type": "text", + "content": "46. Melekhov, I., Ylioinas, J., Kannala, J., Rahtu, E.: Relative camera pose estimation using convolutional neural networks. In: ACIVS (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 217, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 217, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 133, + 217, + 480, + 237 + ], + "type": "text", + "content": "47. Ni, J., Li, Y., Huang, Z., Li, H., Bao, H., Cui, Z., Zhang, G.: Pats: Patch area transportation with subdivision for local feature matching. In: CVPR (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 238, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 238, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 133, + 238, + 480, + 258 + ], + "type": "text", + "content": "48. Nister, D.: An efficient solution to the five-point relative pose problem. TPAMI (2004)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 259, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 259, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 133, + 259, + 480, + 293 + ], + "type": "text", + "content": "49. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 293, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 293, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 133, + 293, + 480, + 314 + ], + "type": "text", + "content": "50. Phil, W.: Bidirectional cross attention. https://github.com/lucidrains/bidirectional-cross-attention (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 315, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 315, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 133, + 315, + 480, + 346 + ], + "type": "text", + "content": "51. Ranftl, R., Lasinger, K., Hafner, D., Schindler, K., Koltun, V.: Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. TPAMI (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 347, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 347, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 133, + 347, + 480, + 369 + ], + "type": "text", + "content": "52. Revaud, J., Weinzaepfel, P., de Souza, C.R., Humenberger, M.: R2D2: repeatable and reliable detector and descriptor. In: NeurIPS (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 369, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 369, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 133, + 369, + 480, + 391 + ], + "type": "text", + "content": "53. Rockwell, C., Johnson, J., Fouhey, D.F.: The 8-point algorithm as an inductive bias for relative pose prediction by vits. In: 3DV (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 392, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 392, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 133, + 392, + 480, + 413 + ], + "type": "text", + "content": "54. Rublee, E., Rabaud, V., Konolige, K., Bradski, G.: Orb: An efficient alternative to sift or surf. In: ICCV (2011)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 414, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 414, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 133, + 414, + 480, + 435 + ], + "type": "text", + "content": "55. Sarlin, P.E., DeTone, D., Malisiewicz, T., Rabinovich, A.: SuperGlue: Learning feature matching with graph neural networks. In: CVPR (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 436, + 480, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 436, + 480, + 467 + ], + "spans": [ + { + "bbox": [ + 133, + 436, + 480, + 467 + ], + "type": "text", + "content": "56. Shi, Y., Cai, J.X., Shavit, Y., Mu, T.J., Feng, W., Zhang, K.: Clustergnn: Cluster-based coarse-to-fine graph neural network for efficient feature matching. In: CVPR (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 468, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 468, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 133, + 468, + 480, + 511 + ], + "type": "text", + "content": "57. Sinha, S., Zhang, J.Y., Tagliasacchi, A., Gilitschenski, I., Lindell, D.B.: Sparse-pose: Sparse-view camera pose regression and refinement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21349–21359 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 512, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 512, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 133, + 512, + 480, + 544 + ], + "type": "text", + "content": "58. Smith, L.N., Topin, N.: Super-convergence: Very fast training of neural networks using large learning rates. In: Artificial intelligence and machine learning for multidomain operations applications (2019)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 133, + 545, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 545, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 133, + 545, + 480, + 567 + ], + "type": "text", + "content": "59. Sun, J., Shen, Z., Wang, Y., Bao, H., Zhou, X.: LoFTR: Detector-free local feature matching with transformers. CVPR (2021)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 133, + 567, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 567, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 133, + 567, + 480, + 588 + ], + "type": "text", + "content": "60. Sun, J., Wang, Z., Zhang, S., He, X., Zhao, H., Zhang, G., Zhou, X.: One-positron object pose estimation without cad models. In: CVPR (2022)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 133, + 589, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 589, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 133, + 589, + 480, + 621 + ], + "type": "text", + "content": "61. Tan, B., Xue, N., Wu, T., Xia, G.S.: Nope-sac: Neural one-plane ransac for sparse-view planar 3d reconstruction. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 133, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 133, + 622, + 480, + 643 + ], + "type": "text", + "content": "62. Tang, S., Zhang, J., Zhu, S., Tan, P.: Quadtree attention for vision transformers. ICLR (2022)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 133, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 644, + 480, + 665 + ], + "type": "text", + "content": "63. Tyszzkiewicz, M., Fua, P., Trulls, E.: Disk: Learning local features with policy gradient. NeurIPS (2020)" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 172, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 172, + 91, + 447, + 102 + ], + "type": "text", + "content": "SRPose: Two-view Relative Pose Estimation with Sparse Keypoints" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 512 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 480, + 149 + ], + "type": "text", + "content": "64. Wang, J., Rupprecht, C., Novotny, D.: Posediffusion: Solving pose estimation via diffusion-aided bundle adjustment. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9773-9783 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 171 + ], + "type": "text", + "content": "65. Wang, Q., Zhang, J., Yang, K., Peng, K., Stiefelhagen, R.: Matchformer: Interleaving attention in transformers for feature matching. In: ACCV (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "type": "text", + "content": "66. Wen, B., Bekris, K.: Bundletrack: 6d pose tracking for novel objects without instance or category-level 3d models. In: IROS (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "text", + "content": "67. Wen, B., Tremblay, J., Blukis, V., Tyree, S., Müller, T., Evans, A., Fox, D., Kautz, J., Birchfield, S.: Bundlesdf: Neural 6-dof tracking and 3d reconstruction of unknown objects. In: CVPR (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "text", + "content": "68. Winkelbauer, D., Denninger, M., Triebel, R.: Learning to localize in new environments from synthetic training data. In: ICRA (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "type": "text", + "content": "69. Xiang, Y., Schmidt, T., Narayanan, V., Fox, D.: PoseCNN: A convolutional neural network for 6d object pose estimation in cluttered scenes. Robotics: Science and Systems XIV (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 281, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 480, + 303 + ], + "type": "text", + "content": "70. Xue, F., Budvytis, I., Cipolla, R.: Imp: Iterative matching and pose estimation with adaptive pooling. In: CVPR (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "text", + "content": "71. Xue, F., Budvytis, I., Cipolla, R.: Sfd2: Semantic-guided feature detection and description. In: CVPR (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 325, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 480, + 346 + ], + "type": "text", + "content": "72. Yi, K.M., Trulls, E., Lepetit, V., Fua, P.: Lift: Learned invariant feature transform. In: ECCV (2016)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "type": "text", + "content": "73. Yu, J., Chang, J., He, J., Zhang, T., Yu, J., Wu, F.: Adaptive spot-guided transformer for consistent local feature matching. In: CVPR (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "text", + "content": "74. Zhang, J.Y., Ramanan, D., Tulsiani, S.: Relpose: Predicting probabilistic relative rotation for single objects in the wild. In: European Conference on Computer Vision. pp. 592-611. Springer (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 402, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 480, + 424 + ], + "type": "text", + "content": "75. Zhao, X., Wu, X., Chen, W., Chen, P.C.Y., Xu, Q., Li, Z.: Aliked: A lighter keypoint and descriptor extraction network via deformable transformation. TIM (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 424, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 480, + 445 + ], + "type": "text", + "content": "76. Zhao, X., Wu, X., Miao, J., Chen, W., Chen, P.C., Li, Z.: Alike: Accurate and lightweight keypoint detection and descriptor extraction. TMM (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 446, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 480, + 468 + ], + "type": "text", + "content": "77. Zhou, Q., Sattler, T., Pollefeys, M., Leal-Taixe, L.: To learn or not to learn: Visual localization from essential matrices. In: ICRA (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "type": "text", + "content": "78. Zhou, Y., Barnes, C., Lu, J., Yang, J., Li, H.: On the continuity of rotation representations in neural networks. In: CVPR (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 490, + 480, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 480, + 512 + ], + "type": "text", + "content": "79. Zhu, S., Liu, X.: Pmatch: Paired masked image modeling for dense geometric matching. In: CVPR (2023)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 218, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 218, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 218, + 100 + ], + "type": "text", + "content": "R. Yin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/SSL-Cleanse_ Trojan Detection and Mitigation in Self-Supervised Learning/a42be677-fdb6-4ca1-8003-8fcf32b7a42b_content_list.json b/2024/SSL-Cleanse_ Trojan Detection and Mitigation in Self-Supervised Learning/a42be677-fdb6-4ca1-8003-8fcf32b7a42b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5c05af5cd5c8f8f5fdd2ad61ed967dc87627d7c4 --- /dev/null +++ b/2024/SSL-Cleanse_ Trojan Detection and Mitigation in Self-Supervised Learning/a42be677-fdb6-4ca1-8003-8fcf32b7a42b_content_list.json @@ -0,0 +1,1421 @@ +[ + { + "type": "text", + "text": "SSL-Cleanse: Trojan Detection and Mitigation in Self-Supervised Learning", + "text_level": 1, + "bbox": [ + 220, + 140, + 781, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mengxin Zheng $^{12*}$ , Jiaqi Xue $^{1*}$ , Zihao Wang $^{2}$ , Xun Chen $^{3}$ , Qian Lou $^{1}$ , Lei Jiang $^{2}$ , and Xiaofeng Wang $^{2}$", + "bbox": [ + 290, + 210, + 712, + 244 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1 University of Central Florida, Orlando, Florida, USA", + "2 Indiana University Bloomington, Bloomington, Indiana, USA", + "$^{3}$ Samsung Research America, Mountain View, California, USA" + ], + "bbox": [ + 287, + 253, + 712, + 296 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Self-supervised learning (SSL) is a prevalent approach for encoding data representations. Using a pre-trained SSL image encoder and subsequently training a downstream classifier, impressive performance can be achieved on various tasks with very little labeled data. The growing adoption of SSL has led to an increase in security research on SSL encoders and associated Trojan attacks. Trojan attacks embedded in SSL encoders can operate covertly, spreading across multiple users and devices. The presence of backdoor behavior in Trojaned encoders can inadvertently be inherited by downstream classifiers, making it even more difficult to detect and mitigate the threat. Although current Trojan detection methods in supervised learning can potentially safeguard SSL downstream classifiers, identifying and addressing triggers in the SSL encoder before its widespread dissemination is a challenging task. This challenge arises because downstream tasks might be unknown, dataset labels may be unavailable, and the original unlabeled training dataset might be inaccessible during Trojan detection in SSL encoders. We introduce SSL-Cleanse as a solution to identify and mitigate backdoor threats in SSL encoders. We evaluated SSL-Cleanse on various datasets using 1200 encoders, achieving an average detection success rate of $82.2\\%$ on ImageNet-100. After mitigating backdoors, on average, backdoored encoders achieve $0.3\\%$ attack success rate without great accuracy loss, proving the effectiveness of SSL-Cleanse. The source code of SSL-Cleanse is available at https://github.com/UCF-ML-Research/SSL-Cleanse.", + "bbox": [ + 261, + 335, + 743, + 654 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Self-supervised learning $\\cdot$ Backdoor attack $\\cdot$ Detection", + "bbox": [ + 261, + 667, + 710, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 709, + 375, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Self-supervised learning (SSL) has seen remarkable advancements, particularly in computer vision applications [3,4,16,22]. This is particularly evident when labeled examples are scarce. Unlike supervised learning, SSL sidesteps the labor-intensive labeling process, training on pretext tasks generalizable to many downstream", + "bbox": [ + 212, + 741, + 787, + 801 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* These authors contributed equally to this work.", + "bbox": [ + 217, + 811, + 552, + 825 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Corresponding Author Email: mengxin.zheng@ucf.edu.", + "bbox": [ + 233, + 825, + 596, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cfc5b204c318dd1b856b71eb29b9249a94f48ae2bad9bc99f8626505d2365b8d.jpg", + "image_caption": [ + "Fig. 1: The overview of SSL-Cleanse. SSL-Cleanse has two components, Detector and Mitigator, aiming to remove the malicious behavior of Trojaned SSL encoders." + ], + "image_footnote": [], + "bbox": [ + 277, + 137, + 727, + 292 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tasks [4,5,10]. Several studies have demonstrated that SSL can achieve comparable [10] and in some cases even superior, performance in few-shot learning [31,33]. The extensive use of SSL has spurred security research and vulnerability exploration in SSL encoders, as evidenced by the emergence of various Trojan attacks [14, 19, 20, 25, 32, 35-37].", + "bbox": [ + 212, + 351, + 787, + 428 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Malicious backdoor (a.k.a, Trojan) attacks [1, 14, 19, 23, 25] in inputs, making the compromised model classify them into a predefined target class with high confidence. If the trigger is removed from the input, the backdoored model will still exhibit normal behaviors with almost the same accuracy as its clean counterpart. One direction of SSL backdoor attacks assumes the attacker can control the training phase and modify the loss function to achieve training-control SSL backdoor [14, 32] with high attack effects. In contrast, another popular SSL backdoor direction is training-agnostic [19, 25, 36] where SSL backdoor attacks are executed in three phases. The first stage involves poisoning unlabeled datasets by adding triggers into a small fraction of target-class images. The second phase entails training the SSL encoder on the poisoned dataset to establish a connection between trigger and target-class images. In the final step, any downstream classifiers that are fine-tuned on the backdoored encoder inherit the backdoor behavior. The current training-agnostic backdoor attacks have demonstrated an attack success rate of over $98\\%$ on the ImageNet-100 dataset [19]. Our goal is to scan SSL encoders and mitigate backdoor threats against such attacks.", + "bbox": [ + 212, + 429, + 787, + 671 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Trojan attacks in SSL encoders [19, 25] are perilous not only lies in their competitive attack success rates but also their covert functionality and broad reach across users and devices. Firstly, pre-trained SSL encoders are typically spread out in real-world scenarios and subsequently fine-tuned for downstream classifiers. However, these downstream classifiers may inadvertently inherit the backdoor behaviors of Trojaned encoders. While current popular Trojan detection techniques [15, 21, 30] in supervised learning may have the potential to protect SSL downstream classifiers, detecting and mitigating triggers in the SSL encoder prior to its wide distribution is a complex undertaking. Recent encoder scanning techniques, as mentioned in [8], recognize their inability to detect these Trojaned SSL encoders [19, 25] because of the distinct covert attack characteristics. We", + "bbox": [ + 212, + 672, + 789, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "M. Zheng et al.", + "bbox": [ + 271, + 114, + 377, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "contend that detecting and mitigating Trojans in SSL encoders is crucial since it can impede the malicious distribution of Trojaned encoders. However, there is a research gap to bridge the popular backdoor defense methods in supervised learning with an SSL encoder. Detecting Trojans in SSL encoders is challenging due to unknown downstream tasks, unavailable dataset labels, and limited access to the original training dataset. Even the linear probe strategy, which builds downstream classifiers, fails to detect these threats, as elaborated in the Limitations of Related Backdoor Defense section. So it is crucial to implement effective detection and defense against such backdoor attacks on SSL encoders.", + "bbox": [ + 212, + 146, + 787, + 281 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This paper introduces SSL-Cleanse, a novel backdoor defense method as illustrated in Figure 1. The proposed approach comprises two main components, namely Detector and Mitigator. The Detector is responsible for identifying the presence of Trojan in an SSL encoder, and if found, the Mitigator can mitigate the backdoor attack effect. Our SSL-Cleanse overcomes the challenges of backdoor detection without knowing the labeled data and the downstream tasks.", + "bbox": [ + 212, + 282, + 787, + 372 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 238, + 373, + 591, + 387 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We design a framework to detect training-agnostic attacks in SSL encoders without downstream labels. We reveal it is possible to prevent the dissemination of Trojaned SSL encoders via our SSL-Cleanse.", + "- We introduce the Sliding Window Kneedle (SWK) algorithm to auto-estimate cluster counts in unlabeled datasets, aiding representation clustering. We also present the representation-oriented trigger reverse (ROTR) method for SSL trigger inversion, alongside the Self-supervised Clustering Unlearning (SCU) algorithm to mitigate SSL encoder backdoors.", + "- We validate the effectiveness of the proposed SSL-Cleanse with extensive experiments on 1200 encoders." + ], + "bbox": [ + 225, + 395, + 785, + 545 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Background and Related Works", + "text_level": 1, + "bbox": [ + 215, + 566, + 560, + 585 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Self-Supervised Learning. Leveraging the large unlabeled data available in the real world is essential. Self-Supervised Learning (SSL) is the most popular method to learn representations from complex unlabeled data [3, 16]. Pre-training an encoder with significant unlabeled data and fine-tuning it with a small amount of labeled data has been shown to achieve comparable performance to using large labeled datasets with supervised learning methods for various downstream tasks [3, 4, 11, 13, 16, 18, 22]. Furthermore, SSL techniques that rely on instance discrimination, such as SimCLR [3] and MoCo V2 [4], have become increasingly popular for learning competitive visual representations through the use of contrastive loss. Typically, an SSL classification task involves pre-training an image encoder, constructing a classifier, and subsequent fine-tuning.", + "bbox": [ + 212, + 598, + 787, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SSL Backdoor Attacks. In SSL backdoor attacks, the first line of research [14] links the trigger to the downstream-task label, necessitating triggers to be appended to varying class inputs. Compared to other directions, it assumes a stronger threat model in which the adversary dominates the training process, e.g., modifying the loss functions. Another methodology [20,32] focuses on specific", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "SSL-Cleanse: Trojan Detection and Mitigation in Self-Supervised Learning", + "bbox": [ + 233, + 113, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/4cda52b3b06d5401afe2d67ac45635716ac4982d2a974048fa7efbb1b8b7c4c8.jpg", + "table_caption": [ + "Table 1: Limitations of current backdoor detectors in assessing SSL encoders, such as SSL-Backdoor [25] and CTRL [19]. The encoder undergoes pre-training on CIFAR-10, while the built classifiers are evaluated on CIFAR-10, STL-10, and GTSRB. These datasets have labels that overlap, partially overlap, or do not overlap with the encoder. If NC [30] Index $>2.0$ and ABS [21] REASR $>0.88$ , the model is seen as Trojaned." + ], + "table_footnote": [], + "table_body": "
SSL Attack MethodDownstream Task (Linear probe)NCABS
Anomaly IndexREASR
SSL-BackdoorCIFAR-102.050.89
STL-101.420.34
GTSRB1.680.29
CTRLCIFAR-101.520.52
STL-101.280.44
GTSRB1.160.37
", + "bbox": [ + 326, + 226, + 669, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "input sets, formulating poisoned data by combining these inputs with their corresponding reference representations. However, it functions only with predetermined specific inputs and not with any inputs that contain embedded triggers. A distinct approach [19, 25] avoids modifying the training phase, achieving a high attack success rate by merely connecting the trigger with desired unlabeled samples. Compared to the first approach, these SSL attacks are much more challenging to discern since triggers are solely attached to the unlabeled targets, and combining triggers with other targets does not consistently yield similar representations [8]. In our study, we aim to detect these third-category attacks.", + "bbox": [ + 212, + 353, + 787, + 489 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Limitations of Related Backdoor Defense. Prior scanners like NC [30] and ABS [21] face challenges in detecting backdoors in SSL encoders. The primary reasons include the often-unknown downstream tasks/labels. While the linear probe method, i.e., constructing downstream classifiers from various datasets using pre-trained encoders, offers an alternative, it is not efficacious in scanning SSL encoders. For example, we built a backdoored encoder using CIFAR-10 [17] with a specific label, airplane, as the target. This encoder was then utilized to train three distinct downstream classifiers on CIFAR-10, STL-10 [6], and GTSRB [28]. The results from applying NC and ABS to these classifiers are shown in Table 1. For scenarios where the encoder and the downstream task share the same dataset, both NC and ABS can detect Trojaned classifiers and hence the backdoored encoders for SSL-Backdoor [25], with encoder's Anomaly Index of 2.05 $>2.00$ in NC and a REASR of $0.89 > 0.88$ in ABS. Yet, the detection capability requires the knowledge of the downstream tasks and the detection capability diminishes when the datasets (like STL-10 or GTSRB) either only partially overlap or do not align at all with CIFAR-10. For the CTRL instance, both tools fail in detection, even when there's label congruence between the encoder and classifiers. The introduction of global frequency triggers in CTRL exacerbates the detection difficulty. While the recent study DEGREE [8] effectively detects backdoors in training-controlled attackers, like BadEncoder [14], it acknowledges the failure in identifying stealthy training-agnostic attacks [19, 25], e.g., $\\sim 50\\%$ detection accuracy. Also, scanners such as PatchSearch [29] and ASSET [24] are orthogonal to our method since they are introduced to identify poisoned samples", + "bbox": [ + 212, + 492, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "M. Zheng et al.", + "bbox": [ + 271, + 114, + 375, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ff4c4b85912ab744193ecda9406d15d94789cb49f783e625a74164df4b606c65.jpg", + "image_caption": [ + "Fig. 2: The workflow of SSL-Cleanse detector. Step 1: Unlabeled data samples are processed through the SSL encoder to compute their representations. The SWK algorithm is then utilized to process representations and determine the number of clusters. Step 2: Using K-Means with the derived cluster number $(K)$ and representation, $K$ clusters are established. Then, the Representation Oriented Trigger Reverse algorithm is employed to generate $K$ trigger patterns. Step 3: Accessing if any of the $K$ triggers are outliers in terms of their size and norm. The identified outlier indicates the encoder is Trojaned." + ], + "image_footnote": [], + "bbox": [ + 220, + 146, + 781, + 220 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "in a training dataset. SSL-ABD [34] employs adversarial simulation of trigger patterns and subsequent removal of the backdoor from the compromised encoder through feature embedding distillation. However, as the defender lacks knowledge of the target class, simulating the trigger pattern becomes challenging, hindering effective backdoor removal. In our threat model, the SSL encoder is required but there's no requirement for the poisoned training dataset.", + "bbox": [ + 212, + 349, + 787, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 SSL-Cleanse Design Overview", + "text_level": 1, + "bbox": [ + 215, + 465, + 540, + 483 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Defense Assumptions and Goals. We assume that the defender has access to a pre-trained SSL encoder, a small portion of the unlabeled dataset (which could be distinct from the training set, i.e., SSL-Cleanse does not require attackers to disclose their poisoned dataset).", + "bbox": [ + 212, + 500, + 785, + 559 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Goals. We have two specific goals:", + "bbox": [ + 215, + 560, + 470, + 575 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Detecting backdoor: Our aim is to make a binary determination regarding the potential backdoor infection of a given SSL encoder. If infected, we also want to identify the potential target classes of the backdoor attack.", + "- Mitigating backdoor: We plan to reversely generate the trigger used by the attack. Our ultimate goal is to deactivate the backdoor, making it ineffective. This requires eliminating the backdoor while preserving the classification performance of the SSL encoder for normal inputs." + ], + "bbox": [ + 225, + 585, + 787, + 694 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Challenges and Motivation. An inherent challenge of SSL backdoor detection lies in the indeterminate nature of the target class. While this can be addressed by iteratively searching through all classes, this method is not suitable for SSL encoders due to the uncertainty in class numbers for unlabeled datasets in SSL. Traditional clustering methods, like K-Means [27], necessitate manual predetermination of cluster counts $K$ . We observe that though automatic cluster number determination methods, such as the Kneedle algorithm [26] exist, they struggle with accurate class number prediction for SSL encoders. For instance, the Kneedle algorithm identifies only 20 classes within the ImageNet-100 dataset.", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "SSL-Cleanse: Trojan Detection and Mitigation in Self-Supervised Learning", + "bbox": [ + 233, + 114, + 732, + 130 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This discrepancy could arise from noisy samples affecting the clustering of SSL representations. Other methods like DBSCAN [2] and Mean Shift [9] rely on predefined parameters like density radius and kernel width. Determining optimal values for these parameters presents a challenge. To enhance cluster number prediction accuracy, we propose the Sliding-Window Kneedle (SWK) algorithm, which addresses noise influence by averaging adjacent Kneedle scores within a sliding window. And this approach is adaptable across various clustering methods.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Another challenge is how to efficiently generate triggers and identify outlier triggers given predicted cluster numbers, encoder, and sampled unlabeled dataset. This challenge is exaggerated when we target to detect attacks [19] based on frequency-domain global triggers since one cannot identify outlier triggers according to the trigger size. For these reasons, we are motivated to propose Representation-Oriented Trigger Reverse (ROTR) method to generate triggers via clustered representations. For identifying patch-wise trigger SSL attacks, specifically SSL-Backdoor as referenced in [25], we continue to rely on trigger size as a key outlier detection metric. However, when detecting the frequency-domain SSL attack, CTRL as cited in [19], we propose to use the magnitude of the trigger as the pivotal criterion. These detection methodologies for both attack types are unified under RORP. Upon identifying the triggers, it becomes feasible to pinpoint their related target-class clusters. Following this, we introduce the Self-supervised Clustering Unlearning (SCU) approach to mitigate backdoor threats in SSL encoders, resulting in a cleansed encoder.", + "bbox": [ + 212, + 253, + 789, + 479 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 SSL-Cleanse Detector", + "text_level": 1, + "bbox": [ + 215, + 510, + 465, + 526 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Workflow. The workflow of the SSL-Cleanse detector, as shown in Figure 2, consists of three steps. First, the SSL encoder processes a small amount of unlabeled data to generate representations. Using these, the Sliding Window Kneedle (SWK) algorithm determines the cluster count $K$ . Subsequently, K-Means creates $K$ clusters, from which the Representation Oriented Trigger Reverse algorithm derives $K$ trigger patterns. Finally, if any trigger significantly deviates in size or magnitude/norm, the encoder is deemed Trojaned. The details of each step and the associated algorithms are elaborated below.", + "bbox": [ + 212, + 547, + 787, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Sliding Window Kneedle. Class numbers for unlabeled training samples in SSL encoders are often unknown, even to model developers.", + "bbox": [ + 212, + 670, + 785, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "At first glance, clustering appears as an intuitive approach to determining class numbers. Yet, conventional methods such as K-Means [27] demand a predetermined value for $K$ . Opting for an automated determination of the cluster number $K$ offers a more flexible solution. To this end, we initially investigated existing automatic methods, with a focus on the Kneedle algorithm [26].", + "bbox": [ + 212, + 703, + 787, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given the SSL samples $D$ and encoder $f$ , the Kneedle algorithm starts by initializing a list with potential $K$ values to examine. For each $K$ in this list, the K-Means method clusters the representation $f(D)$ , producing a clustering outcome. The silhouette score for this clustering ranges from -1 to 1. A high", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "M. Zheng et al.", + "bbox": [ + 271, + 114, + 377, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2355fc664457791348c0ce16a48bb8869ede0cc06b6679d02560e9ee06f3ce34.jpg", + "image_caption": [ + "Fig. 3: Comparison of our SWK method and direct search (Kneedle) method on ImageNet-100 dataset. Our SWK method yields more stable and accurate K." + ], + "image_footnote": [], + "bbox": [ + 289, + 143, + 712, + 305 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "value signifies that the data point is well-suited to its own cluster and has a poor fit with neighboring clusters.", + "bbox": [ + 212, + 358, + 782, + 388 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The silhouette score for a particular data point is computed using: $s(i) = \\frac{b(i) - a(i)}{\\max\\{a(i), b(i)\\}}$ where $s(i)$ represents the silhouette coefficient for data point $i$ , $a(i)$ denotes the mean distance between the $i^{th}$ data point and other points within the same cluster, $b(i)$ is the smallest mean distance between the $i^{th}$ data point and points in a different cluster, minimized over all clusters. The clusters' overall silhouette score is the mean silhouette coefficient of all instances.", + "bbox": [ + 212, + 388, + 784, + 483 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The Knee/elbow point on the plot, which represents silhouette scores over varying $K$ values, indicates the optimal position. To identify the Knee point, one should normalize the silhouette curve, adjust the origin to [0, 0] and ensure the endpoint aligns with [1, 1]. Next, measure the distance of each point on the curve from the direct line connecting the origin [0, 0] to the endpoint [1, 1]. The point that is furthest from this direct line is recognized as the knee point of the curve", + "bbox": [ + 212, + 484, + 784, + 575 + ], + "page_idx": 6 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [ + "Algorithm 1 Sliding Window Kneedle for SSL Cluster Num." + ], + "code_body": "Input: SSL samples $D$ encoder $f$ pre-defined K_list Output: predicted cluster number $K$ initialize clusters_list,s_list,padded_s_list,d_list $\\equiv []$ for $i = 0$ to len(K_list) do clusters_list.append(kmeans(f(D),K_list[i])) s_list.append(sinhouette(f(D),clusters_list[i])) \nend for \ninitialize window size $w$ as a small odd number, e.g., 3 initialize swk_s_list to zero values of s_list's structure padded_s_list $\\leftarrow$ pad $\\frac{w - 1}{2}$ zeros to head and tail of s_list \nfor $i = 1$ to len(s_list) do swk_s_list[i] $\\equiv \\frac{1}{w}\\sum_{j = 0}^{w}$ padded_s_list[i+j] d_list.append(norm(swk_s_list[i])-norm(K_list)) \nend for \n $K\\gets$ index of maximum entry in (d_list)", + "guess_lang": "txt", + "bbox": [ + 215, + 621, + 630, + 834 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "SSL-Cleanse: Trojan Detection and Mitigation in Self-Supervised Learning", + "bbox": [ + 233, + 113, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Figure 3, we illustrate the silhouette curvature distance for each point $K$ . The line corresponding to the direct search showcases the application of the Kneedle algorithm. The most considerable distance is observed when $K = 20$ , indicating the elbow point of the Silhouette score is 20. Nonetheless, this estimation might not be precise, given that the SSL encoder $f$ is trained on the ImageNet-100 dataset. This discrepancy may arise from the impact of noisy samples on the clustering of SSL representations.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To address the high-dimensional noise and varied representations in the encoder's outputs, we enhanced the direct search (Kneedle) approach by introducing a sliding window technique with a window size of $w$ . The underlying idea is to compute the average silhouette scores for neighboring $K$ values, aiming to refine the silhouette curvature. As depicted in Algorithm 1, our new algorithm, Sliding Window Kneedle (SWK), ingests the encoder $f$ , SSL samples $D$ , and a pre-set K_list. The output is the anticipated cluster count, $K$ . The foundation of SWK relies on the Kneedle algorithm's methodology to determine K-Means clustering clusters_list[i] and silhouette scores for every specified K_list[i]. Post silhouette score list (s_list) computation, SWK introduces a window dimension to derive the mean silhouette score, symbolized as swk_s_list, for the neighboring $w$ scores. This process necessitates padding zeros to s_list, resulting in padded_s_list. The silhouette curvature distance, d_list[i], is deduced from the disparity between the normalized swk_s_list and K_list, with normalization between [0,1]. The position of the largest value in d_list dictates the predicted $K$ . Figure 3 illustrates that the evolved SWK curve offers better clarity in pinpointing the elbow juncture, diminishes noise, and assures a more precise cluster estimation.", + "bbox": [ + 212, + 252, + 789, + 508 + ], + "page_idx": 7 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 2 Representation-Oriented Trigger Reverse" + ], + "code_body": "Input: clustered samples $D_{i\\in [1,K]}$ , encoder $f$ , cluster number $K$ , epoch number $E$ \nOutput: masks $m_i^1,m_i^2$ perturbation $\\varDelta_{i}^{1},\\varDelta_{i}^{2}$ of $K$ clusters \nfor $i = 1$ to $K$ do initialize masks $m_i^1,m_i^2$ and perturbation $\\varDelta_{i}^{1},\\varDelta_{i}^{2}$ \nfor $e = 1$ to $E$ do $x_{i}\\gets$ randomly sample an image from $D_{i}\\#$ target $x_{j}\\gets$ randomly sample an image from $D_{j\\neq i}$ $x_{j}^{1}\\gets (1 - m_{i}^{1})\\cdot x_{j} + m_{i}^{1}\\cdot \\varDelta_{i}^{1}$ $x_{j}^{2}\\gets (1 - m_{i}^{2})\\cdot x_{j} + m_{i}^{2}\\cdot \\varDelta_{i}^{2}$ losssize $\\leftarrow \\mathcal{L}_{MSE}^{size}(f(x_i),f(x_j^1))$ lossnorm $\\leftarrow \\mathcal{L}_{MSE}^{norm}(f(x_i),f(x_j^2))$ $m_i^1,\\varDelta_i^1\\gets$ update $(m_i^1,\\varDelta_i^1,\\text{loss\\_size})$ $m_i^2,\\varDelta_i^2\\gets$ update $(m_i^2,\\varDelta_i^2,\\text{loss\\_norm})$ \nend for \nend for", + "bbox": [ + 215, + 551, + 785, + 765 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Representation-Oriented Trigger Reverse. Once the cluster count $K$ is determined, K-Means can be readily employed to produce $K$ clustered samples, denoted by $D_{i\\in [1,K]}$ , using the provided $D$ . The subsequent goal is to backtrack", + "bbox": [ + 212, + 794, + 787, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "M. Zheng et al.", + "bbox": [ + 271, + 114, + 377, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and identify Trojan triggers that are either compact in size or have a minimal $l_{1}$ -norm magnitude that can mimic the representations of target-class samples. Algorithm 2 presents the ROTR approach. The core concept involves creating two triggers for each cluster representation. It then determines which representation clusters can yield outlier triggers, either having smaller patch-based trigger sizes $|m_i^1|$ for patch-based trigger attacks or lesser trigger magnitudes $m_i^2 \\cdot \\Delta_i^2$ for global-trigger attacks. Initially, we designate trigger masks $m_{i}$ to specify the location of the trigger pixel, while $\\varDelta_{i}$ signifies the associated pixel value for the $i$ -th trigger. Consequently, the trigger $r_i$ is computed as $m_{i} \\cdot \\varDelta_{i}$ . To identify both patch-based triggers and frequency-domain global triggers, we suggest initiating two distinct trigger sets: $r_i^1 = m_i^1 \\cdot \\varDelta_i^1$ and $r_i^2 = m_i^2 \\cdot \\varDelta_i^2$ , where $m_{i}^{1}$ and $m_{i}^{2}$ represent masks and $\\varDelta_i^1$ and $\\varDelta_i^2$ delineate the pixel values of the triggers. For each cluster, we conduct $E$ epochs to produce these twin trigger sets. Specifically, we randomly select an image from $D_{i}$ to as the clean sample $x_{i}$ and subsequently get sample to add trigger, termed $S_{j}$ , from $D_{j\\neq i}$ . We then affix the pre-established two trigger sets to $x_{j}$ , resulting in $x_j^1$ and $x_j^2$ respectively.", + "bbox": [ + 212, + 146, + 787, + 388 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The clean image $x_{i}$ and images with trigger $x_{j}^{1}$ and $x_{j}^{2}$ are then sent to an encoder $f$ , and two separate loss functions are employed to update the trigger such that its representation can have more similarity with $x_{i}$ 's feature. The loss functions used in the ROTR to optimize $m_{i}^{1}, m_{i}^{2}$ and $\\Delta_{i}^{1}, \\Delta_{i}^{2}$ on the $i^{th}$ cluster are given by Equation 1 and Equation 2, respectively. Each loss function comprises two components. The foremost objective of the first term is to guarantee the similarity between the image patched with the trigger and the target class image in the feature space. The second term is responsible for constraining the size or the trigger magnitude/norm of the reversed trigger, we adapted $\\lambda$ dynamically during optimization in our experiment. The dynamic scheduler is described in the supplementary material. Finally, the resulting trigger patterns are sent to an outlier detector for determination.", + "bbox": [ + 212, + 388, + 787, + 570 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {M S E} ^ {\\text {s i z e}} \\left(f \\left(x _ {i}\\right), f \\left(x _ {j} ^ {1}\\right)\\right) = - \\frac {< f \\left(x _ {i}\\right) , f \\left(x _ {j} ^ {1}\\right) >}{\\left| \\left| f \\left(x _ {i}\\right) \\right| \\right| \\cdot \\left| \\left| f \\left(x _ {j} ^ {1}\\right) \\right| \\right|} + \\lambda \\cdot \\left| m _ {i} ^ {1} \\right| \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 580, + 787, + 613 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {M S E} ^ {\\text {n o r m}} \\left(f \\left(x _ {i}\\right), f \\left(x _ {j} ^ {2}\\right)\\right) = - \\frac {< f \\left(x _ {i}\\right) , f \\left(x _ {j} ^ {2}\\right) >}{\\left| \\left| f \\left(x _ {i}\\right) \\right| \\right| \\cdot \\left| \\left| f \\left(x _ {j} ^ {2}\\right) \\right| \\right|} + \\lambda \\cdot \\left| m _ {i} ^ {2} \\cdot \\Delta_ {i} ^ {2} \\right| \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 310, + 618, + 785, + 651 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Size-Norm Trigger Outlier Detector. Beyond just using trigger size to identify outliers for patch-wise trigger attacks, e.g., SSL-Backdoor [25], we recognize trigger magnitude/norm as a key criterion for global-trigger attacks stemming from frequency-domain disturbances, e.g., CTRL [19]. Our proposed Size-Norm Trigger Outlier Detection (STOD) is designed to detect both patch-based and global-based triggers. In particular, given a trigger size list $[(m_1^1,\\varDelta_1^1),\\dots,(m_K^1,\\varDelta_K^1)]$ and a trigger norm list $[(m_1^2,\\varDelta_1^2),\\dots,(m_K^2,\\varDelta_K^2)]$ , the STOD outputs the detection result, i.e., benign or Trojaned with a trigger dictionary $t_s$ . For $K$ cluster, we iteratively check if the trigger size $|m_i^1|$ and trigger norm $|m_i^2\\cdot \\varDelta_i^2|$ is outlier in the list of $|m_{[1:K]}^{1}|$ and $|m_{[1:K]}^{2}\\cdot \\varDelta_{[1:K]}^{2}|$ using the function is_outlier $(x_{i},\\pmb {x})$ . Here is_outlier $(x_{i},\\pmb {x})$ returns True if $M(x_{i},\\pmb {x}) > 2$ , otherwise False; The Anomaly Index function $M(x_{i},\\pmb {x}) = \\frac{|x_{i} - \\text{median}(\\pmb{x})|}{c\\cdot\\text{median}(|x - \\text{median}(\\pmb{x})|)}$ is used to ascertain if $x_{i}$ is an", + "bbox": [ + 212, + 654, + 787, + 843 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "SSL-Cleanse: Trojan Detection and Mitigation in Self-Supervised Learning", + "bbox": [ + 233, + 113, + 732, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "anomaly. $c$ is a constant estimator which equals 1.4826. If we discover a trigger that is substantially smaller than the other candidates, we classify it as an outlier and store its cluster number $i$ and itself in a dictionary $t_s$ , i.e., $t_s[i] = m_i \\cdot \\Delta_i$ . If $t_s$ remains empty, we conclude that the encoder is benign. However, if $t_s$ contains any triggers, we classify the encoder as a Trojaned encoder.", + "bbox": [ + 212, + 146, + 787, + 223 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/bc56b363772a253943844dc9dabfbe9edcce2c04392b4854681f0faa81a89402.jpg", + "image_caption": [ + "Fig. 4: Illustration of Self-supervised Clustering Unlearning (SCU). The image $x$ is sampled from a cluster distinct from the cluster producing trigger $t$ ." + ], + "image_footnote": [], + "bbox": [ + 272, + 237, + 725, + 335 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 SSL-Cleanse Mitigator", + "text_level": 1, + "bbox": [ + 215, + 410, + 472, + 428 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Once we have the triggers $t_s$ generated by our detector, their corresponding target-class clusters become identifiable. Subsequently, we introduce a Self-supervised Clustering Unlearning (SCU) strategy to mitigate SSL encoder backdoor threats, leading to a purified encoder. The mitigation approach is detailed in Algorithm 3. This method accepts cluster samples $D_{i\\in [1,k]}$ , a Trojaned encoder $f$ , and the trigger list $t_s$ as inputs, producing a purified encoder $f'$ in return. Since $K$ clusters are generated, we need to clean them one by one. For each cluster $i$ , we iteratively select clean image $x$ from each cluster samples $D_i$ , and augment the image to create a new training sample consisting of the augmented images $x_1$ and $x_2$ . Subsequently, we randomly select a trigger from $t_s$ excluding $t_s[i]$ to make sure the image $x$ is sampled from a cluster distinct from the cluster producing trigger $t$ . Then we attach trigger $t$ to $x_2$ or directly use $x_{i2}$ without adding a trigger. The probability is $50\\%$ which is the meaning of equal Sample in Algorithm 3. The trigger insertion rate is critical for balancing attack mitigation with the preservation of clean accuracy. A $50\\%$ rate is shown to be optimal—it sufficiently counters attacks without compromising clean accuracy. Notice that here we pass these new training samples through the Trojaned encoder $f$ to obtain their respective representations. We then optimize the similarity between the representations using a loss function by fixing the model $f$ and updating the encoder $f'$ to eliminate the Trojan trigger effects, resulting in a clean encoder. Figure 4 illustrates the mitigation process, where a clean image of a cat $x$ is augmented to generate two cat images $x_1$ and $x_2$ , one that is the same and another that has a $50\\%$ chance of being attached with a Trojan trigger. The clean image is sent to a reference model $f$ , which maintains its parameters and representations unchanged. In contrast, the other model $f'$ is updated with the new training sample to remove the Trojan trigger effect. The updated encoder", + "bbox": [ + 212, + 446, + 787, + 843 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "M. Zheng et al.", + "bbox": [ + 271, + 114, + 377, + 128 + ], + "page_idx": 9 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [ + "Algorithm 3 Self-supervised Clustering Unlearning" + ], + "code_body": "Input: $D_{i\\in [1,k]}$ ,Trojaned encoder $f$ , trigger list $t_s$ \nOutput: a clean encoder $f^{\\prime}$ \nInitialize $f^{\\prime}\\gets f$ \nfor $i = 1$ to $K$ do \nfor $x$ in $D_{i}$ do \n $x_{1}\\gets aug_{1}(x);x_{2}\\gets aug_{2}(x)$ $t\\gets$ randomly selected from $t_s$ excluding $t_s[i]$ $\\hat{x}_2\\gets$ equalSample{aug $_2(x\\oplus t),aug_2(x)}$ $z,z^{\\prime} = f(x_{1}),f^{\\prime}(\\hat{x}_{2})$ \nloss $\\leftarrow -s i m i l a r i t y(z,z^{\\prime})$ $f^{\\prime}\\gets$ update(f,loss) \nend for \nend for", + "guess_lang": "txt", + "bbox": [ + 218, + 164, + 568, + 344 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "is deemed to be clean after it is able to accurately classify data, even in the presence of the Trojan trigger.", + "bbox": [ + 212, + 376, + 784, + 407 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "6 Experimental Methodology", + "text_level": 1, + "bbox": [ + 214, + 430, + 514, + 448 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Dataset. Our experiments were conducted on benchmark datasets: CIFAR-10 [17] and ImageNet-100 [7]. ImageNet-100 is a random subset of 100 classes from the larger ImageNet dataset and contains around 127,000 training images, which is widely used in prior SSL attacks [19,25].", + "bbox": [ + 212, + 460, + 787, + 522 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "SSL Attacks and Encoders. To assess the effectiveness of our detector against various attack methods, we evaluated it against two backdoor attacks, namely SSL-Backdoor [25] and CTRL [19] over BYOL [10], SimCLR [3], and MoCo V2 [4], respectively. We created 50 benign encoders and 50 Trojaned encoders for each backdoor attack setting, resulting in 1200 encoders. We follow the above attack's setting to use ResNet-18 [12] as encoder architecture.", + "bbox": [ + 212, + 523, + 787, + 612 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Experimental Settings. Our experiments are performed on two Nvidia GeForce RTX-3090 GPUs, each with a memory capacity of 24 GB. For the detector, the initial value of $\\lambda$ is set up as 0.01. For detection and mitigation running overhead, the detection method with $10\\%$ of the ImageNet-100 training data consumes roughly 20 minutes, and the mitigation process requires approximately 7 minutes.", + "bbox": [ + 212, + 613, + 787, + 688 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Evaluation Metrics. We define the following evaluation metrics to study the efficiency and effectiveness of our SSL-Cleanse. Detection Accuracy (DACC) is detection accuracy which is the ratio of correctly identified encoder types (either Benign or Trojan) relative to the total count of encoders. Attack Success Rate (ASR) is defined as the ratio of images that contain the trigger and are misclassified as the target class, to the total number of evaluated images. Accuracy (ACC) is the percentage of input images without triggers classified into their corresponding correct classes. TP indicates the true positive count, referring to Trojaned encoder numbers detected by our detector. FP represents false positives, indicating clean encoders misclassified as Trojaned encoders by our detector.", + "bbox": [ + 212, + 688, + 787, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "SSL-Cleanse: Trojan Detection and Mitigation in Self-Supervised Learning", + "bbox": [ + 233, + 113, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "7 SSL-Cleanse Results", + "text_level": 1, + "bbox": [ + 215, + 143, + 450, + 160 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Detection. In Table 2, we present the performance of our detector on two training-agnostic SSL attacks including SSL-Backdoor and CTRL on three SSL methods and two datasets. In total, our detection accuracy (DACC) across 1200 (600 Trojaned and 600 benign) encoders stands at $81.33\\%$ , illustrating the efficacy of our backdoor detection capabilities. In particular, for the ImageNet-100 dataset, the DACC is $>77\\%$ and the average DACC is $82.17\\%$ . For the CIFAR-10 dataset, the DACC is $>76\\%$ and the average DACC is $80.5\\%$ .", + "bbox": [ + 212, + 186, + 789, + 295 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/470a65721bc343b3289c157349c7f37ee3de805af1914feda20900a196c00ebc.jpg", + "table_caption": [ + "Table 2: The detection performance of our SSL-Cleanse." + ], + "table_footnote": [], + "table_body": "
DatasetMethodSSL-BackdoorCTRL
TPFPDACC(%)TPFPDACC(%)
CIFAR-10BYOL3558036482
SimCLR3347939584
MoCo V23157637582
ImageNet-100BYOL3888043885
SimCLR3477746888
MoCo V23677942884
", + "bbox": [ + 308, + 338, + 687, + 449 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 2 further indicates that our SSL-Cleanse can detect not only the patch-based trigger SSL attack SSL-Backdoor but also the frequency-based SSL attack CTRL. This capability stems from our method's utilization of trigger size and magnitude for patch-based trigger detection in tandem with frequency-domain detection. In particular, For SSL-Backdoor detection, SSL-Clease achieves $78.5\\%$ DACC on average for both CIFAR-10 and ImageNet-100 datasets. For CTRL, it obtains $84.17\\%$ average DACC. Against the CTRL, our detector identifies 46 TP from 50 Trojaned SimCLR encoders on ImageNet and registers 8 FP among 50 clean SimCLR encoders. Our SSL-Cleanse consistently exhibits reliable detection performance across prevalent SSL training techniques such as BYOL, SimCLR, and MoCo V2. In particular, our detector registers an average DACC of $81.75\\%$ , $82\\%$ , and $80.25\\%$ for BYOL, SimCLR, and MoCo V2, respectively.", + "bbox": [ + 212, + 472, + 789, + 654 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/f7e4bfe14df66679eb498d10206da5d9fb1caace76645670be9ae43f5d634372.jpg", + "table_caption": [ + "Table 3: The mitigation performance of our SSL-Cleanse." + ], + "table_footnote": [], + "table_body": "
DatasetMethodSSL-BackdoorCTRL
Before mitigationAfter mitigationBefore mitigationAfter mitigation
ACC(%)ASR(%)ACC(%)ASR(%)ACC(%)ASR(%)ACC(%)ASR(%)
CIFAR-10BYOL83.4248.3282.141.1483.1960.4782.591.96
SimCLR84.8842.1983.530.5880.7481.8479.601.15
MoCo V281.0237.9580.160.9281.4277.5180.031.62
ImageNet-100BYOL60.5733.2160.240.1453.3345.1052.650.35
SimCLR60.1831.8558.580.6252.9044.9851.040.33
MoCo V261.5735.0660.100.1750.6235.7248.880.17
", + "bbox": [ + 222, + 710, + 777, + 837 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "M. Zheng et al.", + "bbox": [ + 271, + 114, + 377, + 128 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Mitigation. Table 3 compares the attack success rate (ASR) and clean accuracy (ACC) before and after applying the mitigator against SSL-Backdoor and CTRL on the CIFAR-10 and ImageNet-100 dataset. On average, the ASR experiences a marked reduction to below $2\\%$ , while the ACC declines to approximately $1\\%$ before and after implementing mitigation.", + "bbox": [ + 212, + 146, + 787, + 222 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In the case of patch-based SSL-Backdoor, our mitigation approach significantly reduces the ASR to below $1.2\\%$ , demonstrating its successful eradication of backdoor effects. Additionally, the ACC experiences an average decrease of $1.15\\%$ in backdoored models after applying our mitigator. When utilizing the BYOL training method on the ImageNet-100 dataset, ACC remains relatively stable.", + "bbox": [ + 212, + 222, + 787, + 297 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Regarding global frequency-based SSL attack CTRL, our mitigation strategy substantially decreases the ASR to less than $2\\%$ , underscoring its effective elimination of backdoor attacks. Furthermore, the ACC has an average decline of $1.06\\%$ in backdoored models after mitigation. Importantly, when employing the BYOL training technique on both CIFAR-10 and ImageNet-100 datasets, ACC remains fairly consistent.", + "bbox": [ + 212, + 299, + 787, + 388 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Efficiency on Downstream Tasks. SSL-Cleanse demonstrates effective generalization across various downstream tasks. As indicated in Table 4, SSL-Cleanse reliably brings the ASR below $2\\%$ for different downstream tasks such as CIFAR-10, STL-10, and GTSRB datasets, when the SSL encoders were backdoored on CIFAR-10 dataset. In addition, SSL-Cleanse works well across all targeted task labels. For CIFAR-10, it is possible to introduce a backdoor into any label from 0 to 9. SSL-Cleanse has been tested and proven to identify attacks on each label with a detection accuracy (DACC) exceeding $75\\%$ , and the average DACC across 10 labels is $\\sim 80\\%$ .", + "bbox": [ + 212, + 390, + 789, + 526 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/17e9b77778883f0308cfe3f1ad1e5c07ecaa8acf8a65c463db2a9c4a46b802af.jpg", + "table_caption": [ + "Table 4: SSL-Cleanse performance on different downstream tasks." + ], + "table_footnote": [], + "table_body": "
Downstream TaskBefore MitigationAfter Mitigation
ACC(%)ASR(%)ACC(%)ASR(%)
CIFAR-1083.4248.3282.141.14
STL-1077.5336.5575.041.84
GTSRB72.0138.2870.981.03
", + "bbox": [ + 330, + 565, + 663, + 640 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The SWK Effects. Figure 5 presents an ablation study on our SWK approach. Employing SWK on ImageNet100 yields a predicted number of 100, in contrast to 20 as obtained using the direct search (Kneedle) method, as depicted in Figure 3. Consequently, SWK delivers a detection accuracy of $80\\%$ , marking a $30\\%$ DACC improvement. Data Ratio Effects. We analyze the influence of different dataset ratios on", + "bbox": [ + 212, + 657, + 488, + 823 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/f2d0feaf62d15828b1992283a267a6363761e64dcf70c0afb242baaf87cc88d4.jpg", + "image_caption": [ + "Fig.5: A comparison of detection accuracy between SSL-Cleanse using the SWK method and the direct search (Kneedle)on ImageNet-100." + ], + "image_footnote": [], + "bbox": [ + 496, + 661, + 785, + 750 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "detection performance using CTRL attack and BYOL training methods. We", + "bbox": [ + 212, + 824, + 785, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "SSL-Cleanse: Trojan Detection and Mitigation in Self-Supervised Learning", + "bbox": [ + 233, + 113, + 732, + 130 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "employ varied image proportions, namely, $5\\%$ , $8\\%$ , and $10\\%$ . As illustrated in Table 5, the results showcase an improvement in detection performance as the proportion of the training dataset increases. The DACC for the CIFAR-10 dataset shows a $1\\%$ increase between the $8\\%$ and $10\\%$ training datasets. Moreover, for the ImageNet-100 dataset, there is an enhancement from $83\\%$ to $85\\%$ in the DACC. Further details regarding the hyperparameters $\\lambda$ , along with nuances about the influence of attacking trigger size and perturbation norm, can be found in our supplementary material.", + "bbox": [ + 212, + 146, + 787, + 268 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/be85a190817ddf566fc9c81debd4b3cb8a19289526c22fc2bb925291f6bbee81.jpg", + "table_caption": [ + "Table 5: Influence of the ratio of the unlabeled dataset to the entire training dataset on the detection efficacy. A larger ratio usually introduces a higher DACC." + ], + "table_footnote": [], + "table_body": "
Data ratio (%)CIFAR-10ImageNet-100
TPFPDACC(%)TPFPDACC(%)
52877138682
83787940783
103888043885
", + "bbox": [ + 326, + 321, + 669, + 397 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Outlier Detection Effects. Figure 6 shows our Size-Norm Trigger Outlier Detector can successfully distinguish the outlier triggers from the $K$ -cluster trigger list for the trojaned encoders based on SSL-Backdoor and CTRL. In particular, each box bar plots the trigger size $|m_i^1|$ and trigger norm $|m_i^2 \\cdot \\Delta_i^2|$ for SSL-Backdoor and CTRL, respectively. Our Size-Norm trigger outlier detection can scan collectively, without the presumption that the defender has prior knowledge of the attack type.", + "bbox": [ + 212, + 421, + 789, + 527 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/975c6f7e5ba5469c14adc53a5675f7207604d0a463a3b9d47bb6018519772511.jpg", + "image_caption": [ + "Fig. 6: The Size-Norm trigger outlier detection criteria is able to identify both patch-based SSL-Backdoor and frequency-domain global trigger in CTRL. The box plot shows min/max and quartiles." + ], + "image_footnote": [], + "bbox": [ + 300, + 534, + 699, + 656 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "8 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 736, + 359, + 752 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This paper introduces SSL-Cleanse, a novel work to detect and mitigate Trojan attacks in SSL encoders without accessing any downstream labels. We evaluated SSL-Cleanse on various datasets using 1200 models, achieving an average detection success rate of $82.2\\%$ on ImageNet-100. After mitigating backdoors, backdoored encoders achieve $0.3\\%$ average attack success rate without great accuracy loss.", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "M. Zheng et al.", + "bbox": [ + 271, + 114, + 377, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Al Ghanim, M., Santriaji, M., Lou, Q., Solihin, Y.: Trojbits: A hardware aware inference-time attack on transformer-based language models. In: ECAI 2023, pp. 60-68. IOS Press (2023) 2", + "2. Campello, R.J., Moulavi, D., Zimek, A., Sander, J.: Hierarchical density estimates for data clustering, visualization, and outlier detection. ACM Transactions on Knowledge Discovery from Data (TKDD) 10(1), 1-51 (2015) 6", + "3. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International conference on machine learning. pp. 1597-1607. PMLR (2020) 1, 3, 11", + "4. Chen, X., Fan, H., Girshick, R., He, K.: Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297 (2020) 1, 2, 3, 11", + "5. Chen, X., He, K.: Exploring simple siamese representation learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15750-15758 (2021) 2", + "6. Coates, A., Ng, A., Lee, H.: An analysis of single-layer networks in unsupervised feature learning. In: Proceedings of the fourteenth international conference on artificial intelligence and statistics. pp. 215-223. JMLR Workshop and Conference Proceedings (2011) 4", + "7. Deng, J., Dong, W., Socher, R., Li, L., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2009), 20-25 June 2009, Miami, Florida, USA. pp. 248-255. IEEE Computer Society (2009). https://doi.org/10.1109/CVPR.2009.5206848, https://doi.org/10.1109/CVPR.2009.5206848 11", + "8. Feng, S., Tao, G., Cheng, S., Shen, G., Xu, X., Liu, Y., Zhang, K., Ma, S., Zhang, X.: Detecting backdoors in pre-trained encoders. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16352-16362 (2023) 2, 4", + "9. Ghassabeh, Y.A.: A sufficient condition for the convergence of the mean shift algorithm with gaussian kernel. Journal of Multivariate Analysis 135, 1-10 (2015) 6", + "10. Grill, J.B., Strub, F., Altché, F., Tallec, C., Richemond, P., Buchatskaya, E., Doersch, C., Avila Pires, B., Guo, Z., Gheshlaghi Azar, M., et al.: Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems 33, 21271-21284 (2020) 2, 11", + "11. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729-9738 (2020) 3", + "12. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016. pp. 770-778. IEEE Computer Society (2016). https://doi.org/10.1109/CVPR.2016.90, https://doi.org/10.1109/CVPR.2016.90 11", + "13. Jaiswal, A., Babu, A.R., Zadeh, M.Z., Banerjee, D., Makedon, F.: A survey on contrastive self-supervised learning. Technologies 9(1), 2 (2020) 3", + "14. Jia, J., Liu, Y., Gong, N.Z.: Badencoder: Backdoor attacks to pre-trained encoders in self-supervised learning. arXiv preprint arXiv:2108.00352 (2021) 2, 3, 4", + "15. Kolouri, S., Saha, A., Pirsiavash, H., Hoffmann, H.: Universal litmus patterns: Revealing backdoor attacks in cnns. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 301-310 (2020) 2" + ], + "bbox": [ + 218, + 172, + 787, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "SSL-Cleanse: Trojan Detection and Mitigation in Self-Supervised Learning", + "bbox": [ + 233, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Krishnan, R., Rajpurkar, P., Topol, E.J.: Self-supervised learning in medicine and healthcare. Nature Biomedical Engineering pp. 1-7 (2022) 1, 3", + "17. Krizhevsky, A.: Learning multiple layers of features from tiny images. Tech. rep., University of Toronto (2009) 4, 11", + "18. Lan, Z., Chen, M., Goodman, S., Gimpel, K., Sharma, P., Soricut, R.: Albert: A lite bert for self-supervised learning of language representations. arXiv preprint arXiv:1909.11942 (2019) 3", + "19. Li, C., Pang, R., Xi, Z., Du, T., Ji, S., Yao, Y., Wang, T.: Demystifying self-supervised trojan attacks. arXiv preprint arXiv:2210.07346 (2022) 2, 4, 6, 9, 11", + "20. Liu, H., Jia, J., Gong, N.Z.: Poisonedencoder: Poisoning the unlabeled pre-training data in contrastive learning. arXiv preprint arXiv:2205.06401 (2022) 2, 3", + "21. Liu, Y., Lee, W.C., Tao, G., Ma, S., Aafer, Y., Zhang, X.: Abs: Scanning neural networks for back-doors by artificial brain stimulation. In: Proceedings of the 2019 ACM SIGSAC Conference on Computer and Communications Security. pp. 1265-1282 (2019) 2, 4", + "22. Liu, Y., Jin, M., Pan, S., Zhou, C., Zheng, Y., Xia, F., Yu, P.: Graph self-supervised learning: A survey. IEEE Transactions on Knowledge and Data Engineering (2022) 1, 3", + "23. Lou, Q., Liu, Y., Feng, B.: Trojtext: Test-time invisible textual Trojan insertion. In: The Eleventh International Conference on Learning Representations (2023) 2", + "24. Pan, M., Zeng, Y., Lyu, L., Lin, X., Jia, R.: Asset: Robust backdoor data detection across a multiplicity of deep learning paradigms. arXiv preprint arXiv:2302.11408 (2023) 4", + "25. Saha, A., Tejankar, A., Koohpayegani, S.A., Pirsiavash, H.: Backdoor attacks on self-supervised learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13337-13346 (2022) 2, 4, 6, 9, 11", + "26. Satopaa, V., Albrecht, J.R., Irwin, D.E., Raghavan, B.: Finding a \"kneedle\" in a haystack: Detecting knee points in system behavior. In: 31st IEEE International Conference on Distributed Computing Systems Workshops (ICDCS 2011 Workshops), 20-24 June 2011, Minneapolis, Minnesota, USA. pp. 166-171. IEEE Computer Society (2011). https://doi.org/10.1109/ICDCSW.2011.20, https://doi.org/10.1109/ICDCSW.2011.205, 6", + "27. Sinaga, K.P., Yang, M.S.: Unsupervised k-means clustering algorithm. IEEE access 8, 80716-80727 (2020) 5, 6", + "28. Stallkamp, J., Schlipsing, M., Salmen, J., Igel, C.: The german traffic sign recognition benchmark: a multi-class classification competition. In: The 2011 international joint conference on neural networks. pp. 1453-1460. IEEE (2011) 4", + "29. Tejankar, A., Sanjabi, M., Wang, Q., Wang, S., Firooz, H., Pirsiavash, H., Tan, L.: Defending against patch-based backdoor attacks on self-supervised learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12239-12249 (2023) 4", + "30. Wang, B., Yao, Y., Shan, S., Li, H., Viswanath, B., Zheng, H., Zhao, B.Y.: Neural cleansse: Identifying and mitigating backdoor attacks in neural networks. In: Proceedings of the IEEE Symposium on Security and Privacy (IEEE S&P). San Francisco, CA (2019) 2, 4", + "31. Wu, J., Zhang, T., Zha, Z.J., Luo, J., Zhang, Y., Wu, F.: Self-supervised domain-aware generative network for generalized zero-shot learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12767-12776 (2020) 2" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "M. Zheng et al.", + "bbox": [ + 271, + 114, + 375, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "32. Xue, J., Lou, Q.: Estas: Effective and stable trojan attacks in self-supervised encoders with one target unlabelled sample. arXiv preprint arXiv:2211.10908 (2022) 2, 3", + "33. Yaman, B., Hosseini, S.A.H., Akçakaya, M.: Zero-shot self-supervised learning for mri reconstruction. arXiv preprint arXiv:2102.07737 (2021) 2", + "34. Yang, H., Yang, R., Cai, H., Zhang, X., Pei, Q., Wang, S., Yan, H.: Ssl-abd: An adversarial defense method against backdoor attacks in self-supervised learning. In: International Conference on Artificial Intelligence Security and Privacy. pp. 456-467. Springer (2023) 5", + "35. Zhang, J., Liu, H., Jia, J., Gong, N.Z.: Corruptencoder: Data poisoning based backdoor attacks to contrastive learning. arXiv preprint arXiv:2211.08229 (2022) 2", + "36. Zheng, M., Lou, Q., Jiang, L.: Trojvit: Trojan insertion in vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4025-4034 (2023) 2", + "37. Zheng, M., Xue, J., Sheng, Y., Yang, L., Lou, Q., Jiang, L.: Trojfair: Trojan fairness attacks. arXiv preprint arXiv:2312.10508 (2023) 2" + ], + "bbox": [ + 215, + 146, + 787, + 369 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "SSL-Cleanse: Trojan Detection and Mitigation in Self-Supervised Learning", + "bbox": [ + 233, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/SSL-Cleanse_ Trojan Detection and Mitigation in Self-Supervised Learning/a42be677-fdb6-4ca1-8003-8fcf32b7a42b_model.json b/2024/SSL-Cleanse_ Trojan Detection and Mitigation in Self-Supervised Learning/a42be677-fdb6-4ca1-8003-8fcf32b7a42b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..22a3be1123aa7e2e4874efb4ffde16be2266373f --- /dev/null +++ b/2024/SSL-Cleanse_ Trojan Detection and Mitigation in Self-Supervised Learning/a42be677-fdb6-4ca1-8003-8fcf32b7a42b_model.json @@ -0,0 +1,1972 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.222, + 0.141, + 0.782, + 0.187 + ], + "angle": 0, + "content": "SSL-Cleanse: Trojan Detection and Mitigation in Self-Supervised Learning" + }, + { + "type": "text", + "bbox": [ + 0.291, + 0.212, + 0.713, + 0.245 + ], + "angle": 0, + "content": "Mengxin Zheng\\(^{12*}\\), Jiaqi Xue\\(^{1*}\\), Zihao Wang\\(^{2}\\), Xun Chen\\(^{3}\\), Qian Lou\\(^{1}\\), Lei Jiang\\(^{2}\\), and Xiaofeng Wang\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.315, + 0.255, + 0.686, + 0.269 + ], + "angle": 0, + "content": "1 University of Central Florida, Orlando, Florida, USA" + }, + { + "type": "text", + "bbox": [ + 0.289, + 0.27, + 0.712, + 0.283 + ], + "angle": 0, + "content": "2 Indiana University Bloomington, Bloomington, Indiana, USA" + }, + { + "type": "text", + "bbox": [ + 0.289, + 0.284, + 0.713, + 0.297 + ], + "angle": 0, + "content": "\\(^{3}\\) Samsung Research America, Mountain View, California, USA" + }, + { + "type": "list", + "bbox": [ + 0.289, + 0.255, + 0.713, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.336, + 0.744, + 0.655 + ], + "angle": 0, + "content": "Abstract. Self-supervised learning (SSL) is a prevalent approach for encoding data representations. Using a pre-trained SSL image encoder and subsequently training a downstream classifier, impressive performance can be achieved on various tasks with very little labeled data. The growing adoption of SSL has led to an increase in security research on SSL encoders and associated Trojan attacks. Trojan attacks embedded in SSL encoders can operate covertly, spreading across multiple users and devices. The presence of backdoor behavior in Trojaned encoders can inadvertently be inherited by downstream classifiers, making it even more difficult to detect and mitigate the threat. Although current Trojan detection methods in supervised learning can potentially safeguard SSL downstream classifiers, identifying and addressing triggers in the SSL encoder before its widespread dissemination is a challenging task. This challenge arises because downstream tasks might be unknown, dataset labels may be unavailable, and the original unlabeled training dataset might be inaccessible during Trojan detection in SSL encoders. We introduce SSL-Cleanse as a solution to identify and mitigate backdoor threats in SSL encoders. We evaluated SSL-Cleanse on various datasets using 1200 encoders, achieving an average detection success rate of \\(82.2\\%\\) on ImageNet-100. After mitigating backdoors, on average, backdoored encoders achieve \\(0.3\\%\\) attack success rate without great accuracy loss, proving the effectiveness of SSL-Cleanse. The source code of SSL-Cleanse is available at https://github.com/UCF-ML-Research/SSL-Cleanse." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.669, + 0.711, + 0.683 + ], + "angle": 0, + "content": "Keywords: Self-supervised learning \\(\\cdot\\) Backdoor attack \\(\\cdot\\) Detection" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.71, + 0.376, + 0.725 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.742, + 0.788, + 0.803 + ], + "angle": 0, + "content": "Self-supervised learning (SSL) has seen remarkable advancements, particularly in computer vision applications [3,4,16,22]. This is particularly evident when labeled examples are scarce. Unlike supervised learning, SSL sidesteps the labor-intensive labeling process, training on pretext tasks generalizable to many downstream" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.812, + 0.553, + 0.826 + ], + "angle": 0, + "content": "* These authors contributed equally to this work." + }, + { + "type": "page_footnote", + "bbox": [ + 0.234, + 0.827, + 0.598, + 0.84 + ], + "angle": 0, + "content": "Corresponding Author Email: mengxin.zheng@ucf.edu." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.129 + ], + "angle": 0, + "content": "M. Zheng et al." + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.138, + 0.728, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.303, + 0.789, + 0.334 + ], + "angle": 0, + "content": "Fig. 1: The overview of SSL-Cleanse. SSL-Cleanse has two components, Detector and Mitigator, aiming to remove the malicious behavior of Trojaned SSL encoders." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.352, + 0.789, + 0.429 + ], + "angle": 0, + "content": "tasks [4,5,10]. Several studies have demonstrated that SSL can achieve comparable [10] and in some cases even superior, performance in few-shot learning [31,33]. The extensive use of SSL has spurred security research and vulnerability exploration in SSL encoders, as evidenced by the emergence of various Trojan attacks [14, 19, 20, 25, 32, 35-37]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.789, + 0.672 + ], + "angle": 0, + "content": "Malicious backdoor (a.k.a, Trojan) attacks [1, 14, 19, 23, 25] in inputs, making the compromised model classify them into a predefined target class with high confidence. If the trigger is removed from the input, the backdoored model will still exhibit normal behaviors with almost the same accuracy as its clean counterpart. One direction of SSL backdoor attacks assumes the attacker can control the training phase and modify the loss function to achieve training-control SSL backdoor [14, 32] with high attack effects. In contrast, another popular SSL backdoor direction is training-agnostic [19, 25, 36] where SSL backdoor attacks are executed in three phases. The first stage involves poisoning unlabeled datasets by adding triggers into a small fraction of target-class images. The second phase entails training the SSL encoder on the poisoned dataset to establish a connection between trigger and target-class images. In the final step, any downstream classifiers that are fine-tuned on the backdoored encoder inherit the backdoor behavior. The current training-agnostic backdoor attacks have demonstrated an attack success rate of over \\(98\\%\\) on the ImageNet-100 dataset [19]. Our goal is to scan SSL encoders and mitigate backdoor threats against such attacks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.79, + 0.842 + ], + "angle": 0, + "content": "Trojan attacks in SSL encoders [19, 25] are perilous not only lies in their competitive attack success rates but also their covert functionality and broad reach across users and devices. Firstly, pre-trained SSL encoders are typically spread out in real-world scenarios and subsequently fine-tuned for downstream classifiers. However, these downstream classifiers may inadvertently inherit the backdoor behaviors of Trojaned encoders. While current popular Trojan detection techniques [15, 21, 30] in supervised learning may have the potential to protect SSL downstream classifiers, detecting and mitigating triggers in the SSL encoder prior to its wide distribution is a complex undertaking. Recent encoder scanning techniques, as mentioned in [8], recognize their inability to detect these Trojaned SSL encoders [19, 25] because of the distinct covert attack characteristics. We" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.235, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SSL-Cleanse: Trojan Detection and Mitigation in Self-Supervised Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.282 + ], + "angle": 0, + "content": "contend that detecting and mitigating Trojans in SSL encoders is crucial since it can impede the malicious distribution of Trojaned encoders. However, there is a research gap to bridge the popular backdoor defense methods in supervised learning with an SSL encoder. Detecting Trojans in SSL encoders is challenging due to unknown downstream tasks, unavailable dataset labels, and limited access to the original training dataset. Even the linear probe strategy, which builds downstream classifiers, fails to detect these threats, as elaborated in the Limitations of Related Backdoor Defense section. So it is crucial to implement effective detection and defense against such backdoor attacks on SSL encoders." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.788, + 0.373 + ], + "angle": 0, + "content": "This paper introduces SSL-Cleanse, a novel backdoor defense method as illustrated in Figure 1. The proposed approach comprises two main components, namely Detector and Mitigator. The Detector is responsible for identifying the presence of Trojan in an SSL encoder, and if found, the Mitigator can mitigate the backdoor attack effect. Our SSL-Cleanse overcomes the challenges of backdoor detection without knowing the labeled data and the downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.374, + 0.593, + 0.388 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.396, + 0.786, + 0.44 + ], + "angle": 0, + "content": "- We design a framework to detect training-agnostic attacks in SSL encoders without downstream labels. We reveal it is possible to prevent the dissemination of Trojaned SSL encoders via our SSL-Cleanse." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.442, + 0.786, + 0.516 + ], + "angle": 0, + "content": "- We introduce the Sliding Window Kneedle (SWK) algorithm to auto-estimate cluster counts in unlabeled datasets, aiding representation clustering. We also present the representation-oriented trigger reverse (ROTR) method for SSL trigger inversion, alongside the Self-supervised Clustering Unlearning (SCU) algorithm to mitigate SSL encoder backdoors." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.517, + 0.786, + 0.546 + ], + "angle": 0, + "content": "- We validate the effectiveness of the proposed SSL-Cleanse with extensive experiments on 1200 encoders." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.396, + 0.786, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.568, + 0.561, + 0.586 + ], + "angle": 0, + "content": "2 Background and Related Works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.788, + 0.765 + ], + "angle": 0, + "content": "Self-Supervised Learning. Leveraging the large unlabeled data available in the real world is essential. Self-Supervised Learning (SSL) is the most popular method to learn representations from complex unlabeled data [3, 16]. Pre-training an encoder with significant unlabeled data and fine-tuning it with a small amount of labeled data has been shown to achieve comparable performance to using large labeled datasets with supervised learning methods for various downstream tasks [3, 4, 11, 13, 16, 18, 22]. Furthermore, SSL techniques that rely on instance discrimination, such as SimCLR [3] and MoCo V2 [4], have become increasingly popular for learning competitive visual representations through the use of contrastive loss. Typically, an SSL classification task involves pre-training an image encoder, constructing a classifier, and subsequent fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "SSL Backdoor Attacks. In SSL backdoor attacks, the first line of research [14] links the trigger to the downstream-task label, necessitating triggers to be appended to varying class inputs. Compared to other directions, it assumes a stronger threat model in which the adversary dominates the training process, e.g., modifying the loss functions. Another methodology [20,32] focuses on specific" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.377, + 0.129 + ], + "angle": 0, + "content": "M. Zheng et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 1: Limitations of current backdoor detectors in assessing SSL encoders, such as SSL-Backdoor [25] and CTRL [19]. The encoder undergoes pre-training on CIFAR-10, while the built classifiers are evaluated on CIFAR-10, STL-10, and GTSRB. These datasets have labels that overlap, partially overlap, or do not overlap with the encoder. If NC [30] Index \\(>2.0\\) and ABS [21] REASR \\(>0.88\\), the model is seen as Trojaned." + }, + { + "type": "table", + "bbox": [ + 0.327, + 0.227, + 0.67, + 0.337 + ], + "angle": 0, + "content": "
SSL Attack MethodDownstream Task (Linear probe)NCABS
Anomaly IndexREASR
SSL-BackdoorCIFAR-102.050.89
STL-101.420.34
GTSRB1.680.29
CTRLCIFAR-101.520.52
STL-101.280.44
GTSRB1.160.37
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.354, + 0.789, + 0.49 + ], + "angle": 0, + "content": "input sets, formulating poisoned data by combining these inputs with their corresponding reference representations. However, it functions only with predetermined specific inputs and not with any inputs that contain embedded triggers. A distinct approach [19, 25] avoids modifying the training phase, achieving a high attack success rate by merely connecting the trigger with desired unlabeled samples. Compared to the first approach, these SSL attacks are much more challenging to discern since triggers are solely attached to the unlabeled targets, and combining triggers with other targets does not consistently yield similar representations [8]. In our study, we aim to detect these third-category attacks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Limitations of Related Backdoor Defense. Prior scanners like NC [30] and ABS [21] face challenges in detecting backdoors in SSL encoders. The primary reasons include the often-unknown downstream tasks/labels. While the linear probe method, i.e., constructing downstream classifiers from various datasets using pre-trained encoders, offers an alternative, it is not efficacious in scanning SSL encoders. For example, we built a backdoored encoder using CIFAR-10 [17] with a specific label, airplane, as the target. This encoder was then utilized to train three distinct downstream classifiers on CIFAR-10, STL-10 [6], and GTSRB [28]. The results from applying NC and ABS to these classifiers are shown in Table 1. For scenarios where the encoder and the downstream task share the same dataset, both NC and ABS can detect Trojaned classifiers and hence the backdoored encoders for SSL-Backdoor [25], with encoder's Anomaly Index of 2.05 \\(>2.00\\) in NC and a REASR of \\(0.89 > 0.88\\) in ABS. Yet, the detection capability requires the knowledge of the downstream tasks and the detection capability diminishes when the datasets (like STL-10 or GTSRB) either only partially overlap or do not align at all with CIFAR-10. For the CTRL instance, both tools fail in detection, even when there's label congruence between the encoder and classifiers. The introduction of global frequency triggers in CTRL exacerbates the detection difficulty. While the recent study DEGREE [8] effectively detects backdoors in training-controlled attackers, like BadEncoder [14], it acknowledges the failure in identifying stealthy training-agnostic attacks [19, 25], e.g., \\(\\sim 50\\%\\) detection accuracy. Also, scanners such as PatchSearch [29] and ASSET [24] are orthogonal to our method since they are introduced to identify poisoned samples" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.235, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "SSL-Cleanse: Trojan Detection and Mitigation in Self-Supervised Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.147, + 0.782, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.231, + 0.788, + 0.329 + ], + "angle": 0, + "content": "Fig. 2: The workflow of SSL-Cleanse detector. Step 1: Unlabeled data samples are processed through the SSL encoder to compute their representations. The SWK algorithm is then utilized to process representations and determine the number of clusters. Step 2: Using K-Means with the derived cluster number \\((K)\\) and representation, \\(K\\) clusters are established. Then, the Representation Oriented Trigger Reverse algorithm is employed to generate \\(K\\) trigger patterns. Step 3: Accessing if any of the \\(K\\) triggers are outliers in terms of their size and norm. The identified outlier indicates the encoder is Trojaned." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.35, + 0.788, + 0.442 + ], + "angle": 0, + "content": "in a training dataset. SSL-ABD [34] employs adversarial simulation of trigger patterns and subsequent removal of the backdoor from the compromised encoder through feature embedding distillation. However, as the defender lacks knowledge of the target class, simulating the trigger pattern becomes challenging, hindering effective backdoor removal. In our threat model, the SSL encoder is required but there's no requirement for the poisoned training dataset." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.467, + 0.542, + 0.484 + ], + "angle": 0, + "content": "3 SSL-Cleanse Design Overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.5, + 0.787, + 0.56 + ], + "angle": 0, + "content": "Defense Assumptions and Goals. We assume that the defender has access to a pre-trained SSL encoder, a small portion of the unlabeled dataset (which could be distinct from the training set, i.e., SSL-Cleanse does not require attackers to disclose their poisoned dataset)." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.561, + 0.471, + 0.576 + ], + "angle": 0, + "content": "Goals. We have two specific goals:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.587, + 0.787, + 0.632 + ], + "angle": 0, + "content": "- Detecting backdoor: Our aim is to make a binary determination regarding the potential backdoor infection of a given SSL encoder. If infected, we also want to identify the potential target classes of the backdoor attack." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.633, + 0.788, + 0.695 + ], + "angle": 0, + "content": "- Mitigating backdoor: We plan to reversely generate the trigger used by the attack. Our ultimate goal is to deactivate the backdoor, making it ineffective. This requires eliminating the backdoor while preserving the classification performance of the SSL encoder for normal inputs." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.587, + 0.788, + 0.695 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Challenges and Motivation. An inherent challenge of SSL backdoor detection lies in the indeterminate nature of the target class. While this can be addressed by iteratively searching through all classes, this method is not suitable for SSL encoders due to the uncertainty in class numbers for unlabeled datasets in SSL. Traditional clustering methods, like K-Means [27], necessitate manual predetermination of cluster counts \\(K\\). We observe that though automatic cluster number determination methods, such as the Kneedle algorithm [26] exist, they struggle with accurate class number prediction for SSL encoders. For instance, the Kneedle algorithm identifies only 20 classes within the ImageNet-100 dataset." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.129 + ], + "angle": 0, + "content": "M. Zheng et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "This discrepancy could arise from noisy samples affecting the clustering of SSL representations. Other methods like DBSCAN [2] and Mean Shift [9] rely on predefined parameters like density radius and kernel width. Determining optimal values for these parameters presents a challenge. To enhance cluster number prediction accuracy, we propose the Sliding-Window Kneedle (SWK) algorithm, which addresses noise influence by averaging adjacent Kneedle scores within a sliding window. And this approach is adaptable across various clustering methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.255, + 0.79, + 0.481 + ], + "angle": 0, + "content": "Another challenge is how to efficiently generate triggers and identify outlier triggers given predicted cluster numbers, encoder, and sampled unlabeled dataset. This challenge is exaggerated when we target to detect attacks [19] based on frequency-domain global triggers since one cannot identify outlier triggers according to the trigger size. For these reasons, we are motivated to propose Representation-Oriented Trigger Reverse (ROTR) method to generate triggers via clustered representations. For identifying patch-wise trigger SSL attacks, specifically SSL-Backdoor as referenced in [25], we continue to rely on trigger size as a key outlier detection metric. However, when detecting the frequency-domain SSL attack, CTRL as cited in [19], we propose to use the magnitude of the trigger as the pivotal criterion. These detection methodologies for both attack types are unified under RORP. Upon identifying the triggers, it becomes feasible to pinpoint their related target-class clusters. Following this, we introduce the Self-supervised Clustering Unlearning (SCU) approach to mitigate backdoor threats in SSL encoders, resulting in a cleansed encoder." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.511, + 0.466, + 0.527 + ], + "angle": 0, + "content": "4 SSL-Cleanse Detector" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.549, + 0.788, + 0.669 + ], + "angle": 0, + "content": "Workflow. The workflow of the SSL-Cleanse detector, as shown in Figure 2, consists of three steps. First, the SSL encoder processes a small amount of unlabeled data to generate representations. Using these, the Sliding Window Kneedle (SWK) algorithm determines the cluster count \\( K \\). Subsequently, K-Means creates \\( K \\) clusters, from which the Representation Oriented Trigger Reverse algorithm derives \\( K \\) trigger patterns. Finally, if any trigger significantly deviates in size or magnitude/norm, the encoder is deemed Trojaned. The details of each step and the associated algorithms are elaborated below." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.671, + 0.787, + 0.702 + ], + "angle": 0, + "content": "Sliding Window Kneedle. Class numbers for unlabeled training samples in SSL encoders are often unknown, even to model developers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.779 + ], + "angle": 0, + "content": "At first glance, clustering appears as an intuitive approach to determining class numbers. Yet, conventional methods such as K-Means [27] demand a predetermined value for \\(K\\). Opting for an automated determination of the cluster number \\(K\\) offers a more flexible solution. To this end, we initially investigated existing automatic methods, with a focus on the Kneedle algorithm [26]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Given the SSL samples \\( D \\) and encoder \\( f \\), the Kneedle algorithm starts by initializing a list with potential \\( K \\) values to examine. For each \\( K \\) in this list, the K-Means method clusters the representation \\( f(D) \\), producing a clustering outcome. The silhouette score for this clustering ranges from -1 to 1. A high" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.235, + 0.114, + 0.733, + 0.131 + ], + "angle": 0, + "content": "SSL-Cleanse: Trojan Detection and Mitigation in Self-Supervised Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.144, + 0.713, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.318, + 0.784, + 0.348 + ], + "angle": 0, + "content": "Fig. 3: Comparison of our SWK method and direct search (Kneedle) method on ImageNet-100 dataset. Our SWK method yields more stable and accurate K." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.359, + 0.784, + 0.389 + ], + "angle": 0, + "content": "value signifies that the data point is well-suited to its own cluster and has a poor fit with neighboring clusters." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.39, + 0.785, + 0.484 + ], + "angle": 0, + "content": "The silhouette score for a particular data point is computed using: \\( s(i) = \\frac{b(i) - a(i)}{\\max\\{a(i), b(i)\\}} \\) where \\( s(i) \\) represents the silhouette coefficient for data point \\( i \\), \\( a(i) \\) denotes the mean distance between the \\( i^{th} \\) data point and other points within the same cluster, \\( b(i) \\) is the smallest mean distance between the \\( i^{th} \\) data point and points in a different cluster, minimized over all clusters. The clusters' overall silhouette score is the mean silhouette coefficient of all instances." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.486, + 0.785, + 0.577 + ], + "angle": 0, + "content": "The Knee/elbow point on the plot, which represents silhouette scores over varying \\( K \\) values, indicates the optimal position. To identify the Knee point, one should normalize the silhouette curve, adjust the origin to [0, 0] and ensure the endpoint aligns with [1, 1]. Next, measure the distance of each point on the curve from the direct line connecting the origin [0, 0] to the endpoint [1, 1]. The point that is furthest from this direct line is recognized as the knee point of the curve" + }, + { + "type": "code_caption", + "bbox": [ + 0.216, + 0.605, + 0.663, + 0.621 + ], + "angle": 0, + "content": "Algorithm 1 Sliding Window Kneedle for SSL Cluster Num." + }, + { + "type": "code", + "bbox": [ + 0.216, + 0.622, + 0.631, + 0.835 + ], + "angle": 0, + "content": "Input: SSL samples \\(D\\) encoder \\(f\\) pre-defined K_list Output: predicted cluster number \\(K\\) initialize clusters_list,s_list,padded_s_list,d_list \\(\\equiv []\\) for \\(i = 0\\) to len(K_list) do clusters_list.append(kmeans(f(D),K_list[i])) s_list.append(sinhouette(f(D),clusters_list[i])) \nend for \ninitialize window size \\(w\\) as a small odd number, e.g., 3 initialize swk_s_list to zero values of s_list's structure padded_s_list \\(\\leftarrow\\) pad \\(\\frac{w - 1}{2}\\) zeros to head and tail of s_list \nfor \\(i = 1\\) to len(s_list) do swk_s_list[i] \\(\\equiv \\frac{1}{w}\\sum_{j = 0}^{w}\\) padded_s_list[i+j] d_list.append(norm(swk_s_list[i])-norm(K_list)) \nend for \n\\(K\\gets\\) index of maximum entry in (d_list)" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.129 + ], + "angle": 0, + "content": "M. Zheng et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "In Figure 3, we illustrate the silhouette curvature distance for each point \\( K \\). The line corresponding to the direct search showcases the application of the Kneedle algorithm. The most considerable distance is observed when \\( K = 20 \\), indicating the elbow point of the Silhouette score is 20. Nonetheless, this estimation might not be precise, given that the SSL encoder \\( f \\) is trained on the ImageNet-100 dataset. This discrepancy may arise from the impact of noisy samples on the clustering of SSL representations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.79, + 0.51 + ], + "angle": 0, + "content": "To address the high-dimensional noise and varied representations in the encoder's outputs, we enhanced the direct search (Kneedle) approach by introducing a sliding window technique with a window size of \\( w \\). The underlying idea is to compute the average silhouette scores for neighboring \\( K \\) values, aiming to refine the silhouette curvature. As depicted in Algorithm 1, our new algorithm, Sliding Window Kneedle (SWK), ingests the encoder \\( f \\), SSL samples \\( D \\), and a pre-set K_list. The output is the anticipated cluster count, \\( K \\). The foundation of SWK relies on the Kneedle algorithm's methodology to determine K-Means clustering clusters_list[i] and silhouette scores for every specified K_list[i]. Post silhouette score list (s_list) computation, SWK introduces a window dimension to derive the mean silhouette score, symbolized as swk_s_list, for the neighboring \\( w \\) scores. This process necessitates padding zeros to s_list, resulting in padded_s_list. The silhouette curvature distance, d_list[i], is deduced from the disparity between the normalized swk_s_list and K_list, with normalization between [0,1]. The position of the largest value in d_list dictates the predicted \\( K \\). Figure 3 illustrates that the evolved SWK curve offers better clarity in pinpointing the elbow juncture, diminishes noise, and assures a more precise cluster estimation." + }, + { + "type": "code_caption", + "bbox": [ + 0.218, + 0.537, + 0.619, + 0.551 + ], + "angle": 0, + "content": "Algorithm 2 Representation-Oriented Trigger Reverse" + }, + { + "type": "algorithm", + "bbox": [ + 0.217, + 0.553, + 0.787, + 0.766 + ], + "angle": 0, + "content": "Input: clustered samples \\(D_{i\\in [1,K]}\\) , encoder \\(f\\) , cluster number \\(K\\) , epoch number \\(E\\) \nOutput: masks \\(m_i^1,m_i^2\\) perturbation \\(\\varDelta_{i}^{1},\\varDelta_{i}^{2}\\) of \\(K\\) clusters \nfor \\(i = 1\\) to \\(K\\) do initialize masks \\(m_i^1,m_i^2\\) and perturbation \\(\\varDelta_{i}^{1},\\varDelta_{i}^{2}\\) \nfor \\(e = 1\\) to \\(E\\) do \\(x_{i}\\gets\\) randomly sample an image from \\(D_{i}\\#\\) target \\(x_{j}\\gets\\) randomly sample an image from \\(D_{j\\neq i}\\) \\(x_{j}^{1}\\gets (1 - m_{i}^{1})\\cdot x_{j} + m_{i}^{1}\\cdot \\varDelta_{i}^{1}\\) \\(x_{j}^{2}\\gets (1 - m_{i}^{2})\\cdot x_{j} + m_{i}^{2}\\cdot \\varDelta_{i}^{2}\\) losssize \\(\\leftarrow \\mathcal{L}_{MSE}^{size}(f(x_i),f(x_j^1))\\) lossnorm \\(\\leftarrow \\mathcal{L}_{MSE}^{norm}(f(x_i),f(x_j^2))\\) \\(m_i^1,\\varDelta_i^1\\gets\\) update \\((m_i^1,\\varDelta_i^1,\\text{loss\\_size})\\) \\(m_i^2,\\varDelta_i^2\\gets\\) update \\((m_i^2,\\varDelta_i^2,\\text{loss\\_norm})\\) \nend for \nend for" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.789, + 0.843 + ], + "angle": 0, + "content": "Representation-Oriented Trigger Reverse. Once the cluster count \\( K \\) is determined, K-Means can be readily employed to produce \\( K \\) clustered samples, denoted by \\( D_{i\\in [1,K]} \\), using the provided \\( D \\). The subsequent goal is to backtrack" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.235, + 0.114, + 0.733, + 0.131 + ], + "angle": 0, + "content": "SSL-Cleanse: Trojan Detection and Mitigation in Self-Supervised Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.39 + ], + "angle": 0, + "content": "and identify Trojan triggers that are either compact in size or have a minimal \\( l_{1} \\)-norm magnitude that can mimic the representations of target-class samples. Algorithm 2 presents the ROTR approach. The core concept involves creating two triggers for each cluster representation. It then determines which representation clusters can yield outlier triggers, either having smaller patch-based trigger sizes \\( |m_i^1| \\) for patch-based trigger attacks or lesser trigger magnitudes \\( m_i^2 \\cdot \\Delta_i^2 \\) for global-trigger attacks. Initially, we designate trigger masks \\( m_{i} \\) to specify the location of the trigger pixel, while \\( \\varDelta_{i} \\) signifies the associated pixel value for the \\( i \\)-th trigger. Consequently, the trigger \\( r_i \\) is computed as \\( m_{i} \\cdot \\varDelta_{i} \\). To identify both patch-based triggers and frequency-domain global triggers, we suggest initiating two distinct trigger sets: \\( r_i^1 = m_i^1 \\cdot \\varDelta_i^1 \\) and \\( r_i^2 = m_i^2 \\cdot \\varDelta_i^2 \\), where \\( m_{i}^{1} \\) and \\( m_{i}^{2} \\) represent masks and \\( \\varDelta_i^1 \\) and \\( \\varDelta_i^2 \\) delineate the pixel values of the triggers. For each cluster, we conduct \\( E \\) epochs to produce these twin trigger sets. Specifically, we randomly select an image from \\( D_{i} \\) to as the clean sample \\( x_{i} \\) and subsequently get sample to add trigger, termed \\( S_{j} \\), from \\( D_{j\\neq i} \\). We then affix the pre-established two trigger sets to \\( x_{j} \\), resulting in \\( x_j^1 \\) and \\( x_j^2 \\) respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.39, + 0.788, + 0.571 + ], + "angle": 0, + "content": "The clean image \\( x_{i} \\) and images with trigger \\( x_{j}^{1} \\) and \\( x_{j}^{2} \\) are then sent to an encoder \\( f \\), and two separate loss functions are employed to update the trigger such that its representation can have more similarity with \\( x_{i} \\)'s feature. The loss functions used in the ROTR to optimize \\( m_{i}^{1}, m_{i}^{2} \\) and \\( \\Delta_{i}^{1}, \\Delta_{i}^{2} \\) on the \\( i^{th} \\) cluster are given by Equation 1 and Equation 2, respectively. Each loss function comprises two components. The foremost objective of the first term is to guarantee the similarity between the image patched with the trigger and the target class image in the feature space. The second term is responsible for constraining the size or the trigger magnitude/norm of the reversed trigger, we adapted \\( \\lambda \\) dynamically during optimization in our experiment. The dynamic scheduler is described in the supplementary material. Finally, the resulting trigger patterns are sent to an outlier detector for determination." + }, + { + "type": "equation", + "bbox": [ + 0.328, + 0.582, + 0.788, + 0.614 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {M S E} ^ {\\text {s i z e}} \\left(f \\left(x _ {i}\\right), f \\left(x _ {j} ^ {1}\\right)\\right) = - \\frac {< f \\left(x _ {i}\\right) , f \\left(x _ {j} ^ {1}\\right) >}{\\left| \\left| f \\left(x _ {i}\\right) \\right| \\right| \\cdot \\left| \\left| f \\left(x _ {j} ^ {1}\\right) \\right| \\right|} + \\lambda \\cdot \\left| m _ {i} ^ {1} \\right| \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.312, + 0.619, + 0.787, + 0.652 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {M S E} ^ {\\text {n o r m}} \\left(f \\left(x _ {i}\\right), f \\left(x _ {j} ^ {2}\\right)\\right) = - \\frac {< f \\left(x _ {i}\\right) , f \\left(x _ {j} ^ {2}\\right) >}{\\left| \\left| f \\left(x _ {i}\\right) \\right| \\right| \\cdot \\left| \\left| f \\left(x _ {j} ^ {2}\\right) \\right| \\right|} + \\lambda \\cdot \\left| m _ {i} ^ {2} \\cdot \\Delta_ {i} ^ {2} \\right| \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.655, + 0.788, + 0.844 + ], + "angle": 0, + "content": "Size-Norm Trigger Outlier Detector. Beyond just using trigger size to identify outliers for patch-wise trigger attacks, e.g., SSL-Backdoor [25], we recognize trigger magnitude/norm as a key criterion for global-trigger attacks stemming from frequency-domain disturbances, e.g., CTRL [19]. Our proposed Size-Norm Trigger Outlier Detection (STOD) is designed to detect both patch-based and global-based triggers. In particular, given a trigger size list \\([(m_1^1,\\varDelta_1^1),\\dots,(m_K^1,\\varDelta_K^1)]\\) and a trigger norm list \\([(m_1^2,\\varDelta_1^2),\\dots,(m_K^2,\\varDelta_K^2)]\\), the STOD outputs the detection result, i.e., benign or Trojaned with a trigger dictionary \\(t_s\\). For \\(K\\) cluster, we iteratively check if the trigger size \\(|m_i^1|\\) and trigger norm \\(|m_i^2\\cdot \\varDelta_i^2|\\) is outlier in the list of \\(|m_{[1:K]}^{1}|\\) and \\(|m_{[1:K]}^{2}\\cdot \\varDelta_{[1:K]}^{2}|\\) using the function is_outlier \\((x_{i},\\pmb {x})\\). Here is_outlier \\((x_{i},\\pmb {x})\\) returns True if \\(M(x_{i},\\pmb {x}) > 2\\), otherwise False; The Anomaly Index function \\(M(x_{i},\\pmb {x}) = \\frac{|x_{i} - \\text{median}(\\pmb{x})|}{c\\cdot\\text{median}(|x - \\text{median}(\\pmb{x})|)}\\) is used to ascertain if \\(x_{i}\\) is an" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.129 + ], + "angle": 0, + "content": "M. Zheng et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "anomaly. \\( c \\) is a constant estimator which equals 1.4826. If we discover a trigger that is substantially smaller than the other candidates, we classify it as an outlier and store its cluster number \\( i \\) and itself in a dictionary \\( t_s \\), i.e., \\( t_s[i] = m_i \\cdot \\Delta_i \\). If \\( t_s \\) remains empty, we conclude that the encoder is benign. However, if \\( t_s \\) contains any triggers, we classify the encoder as a Trojaned encoder." + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.238, + 0.727, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.341, + 0.79, + 0.371 + ], + "angle": 0, + "content": "Fig. 4: Illustration of Self-supervised Clustering Unlearning (SCU). The image \\( x \\) is sampled from a cluster distinct from the cluster producing trigger \\( t \\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.411, + 0.473, + 0.429 + ], + "angle": 0, + "content": "5 SSL-Cleanse Mitigator" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.789, + 0.844 + ], + "angle": 0, + "content": "Once we have the triggers \\( t_s \\) generated by our detector, their corresponding target-class clusters become identifiable. Subsequently, we introduce a Self-supervised Clustering Unlearning (SCU) strategy to mitigate SSL encoder backdoor threats, leading to a purified encoder. The mitigation approach is detailed in Algorithm 3. This method accepts cluster samples \\( D_{i\\in [1,k]} \\), a Trojaned encoder \\( f \\), and the trigger list \\( t_s \\) as inputs, producing a purified encoder \\( f' \\) in return. Since \\( K \\) clusters are generated, we need to clean them one by one. For each cluster \\( i \\), we iteratively select clean image \\( x \\) from each cluster samples \\( D_i \\), and augment the image to create a new training sample consisting of the augmented images \\( x_1 \\) and \\( x_2 \\). Subsequently, we randomly select a trigger from \\( t_s \\) excluding \\( t_s[i] \\) to make sure the image \\( x \\) is sampled from a cluster distinct from the cluster producing trigger \\( t \\). Then we attach trigger \\( t \\) to \\( x_2 \\) or directly use \\( x_{i2} \\) without adding a trigger. The probability is \\( 50\\% \\) which is the meaning of equal Sample in Algorithm 3. The trigger insertion rate is critical for balancing attack mitigation with the preservation of clean accuracy. A \\( 50\\% \\) rate is shown to be optimal—it sufficiently counters attacks without compromising clean accuracy. Notice that here we pass these new training samples through the Trojaned encoder \\( f \\) to obtain their respective representations. We then optimize the similarity between the representations using a loss function by fixing the model \\( f \\) and updating the encoder \\( f' \\) to eliminate the Trojan trigger effects, resulting in a clean encoder. Figure 4 illustrates the mitigation process, where a clean image of a cat \\( x \\) is augmented to generate two cat images \\( x_1 \\) and \\( x_2 \\), one that is the same and another that has a \\( 50\\% \\) chance of being attached with a Trojan trigger. The clean image is sent to a reference model \\( f \\), which maintains its parameters and representations unchanged. In contrast, the other model \\( f' \\) is updated with the new training sample to remove the Trojan trigger effect. The updated encoder" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.235, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SSL-Cleanse: Trojan Detection and Mitigation in Self-Supervised Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "code_caption", + "bbox": [ + 0.218, + 0.146, + 0.594, + 0.162 + ], + "angle": 0, + "content": "Algorithm 3 Self-supervised Clustering Unlearning" + }, + { + "type": "code", + "bbox": [ + 0.219, + 0.165, + 0.57, + 0.345 + ], + "angle": 0, + "content": "Input: \\(D_{i\\in [1,k]}\\) ,Trojaned encoder \\(f\\) , trigger list \\(t_s\\) \nOutput: a clean encoder \\(f^{\\prime}\\) \nInitialize \\(f^{\\prime}\\gets f\\) \nfor \\(i = 1\\) to \\(K\\) do \nfor \\(x\\) in \\(D_{i}\\) do \n\\(x_{1}\\gets aug_{1}(x);x_{2}\\gets aug_{2}(x)\\) \n\\(t\\gets\\) randomly selected from \\(t_s\\) excluding \\(t_s[i]\\) \n\\(\\hat{x}_2\\gets\\) equalSample{aug \\(_2(x\\oplus t),aug_2(x)}\\) \n\\(z,z^{\\prime} = f(x_{1}),f^{\\prime}(\\hat{x}_{2})\\) \nloss \\(\\leftarrow -s i m i l a r i t y(z,z^{\\prime})\\) \n\\(f^{\\prime}\\gets\\) update(f,loss) \nend for \nend for" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.377, + 0.785, + 0.408 + ], + "angle": 0, + "content": "is deemed to be clean after it is able to accurately classify data, even in the presence of the Trojan trigger." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.431, + 0.516, + 0.449 + ], + "angle": 0, + "content": "6 Experimental Methodology" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.788, + 0.523 + ], + "angle": 0, + "content": "Dataset. Our experiments were conducted on benchmark datasets: CIFAR-10 [17] and ImageNet-100 [7]. ImageNet-100 is a random subset of 100 classes from the larger ImageNet dataset and contains around 127,000 training images, which is widely used in prior SSL attacks [19,25]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.524, + 0.788, + 0.613 + ], + "angle": 0, + "content": "SSL Attacks and Encoders. To assess the effectiveness of our detector against various attack methods, we evaluated it against two backdoor attacks, namely SSL-Backdoor [25] and CTRL [19] over BYOL [10], SimCLR [3], and MoCo V2 [4], respectively. We created 50 benign encoders and 50 Trojaned encoders for each backdoor attack setting, resulting in 1200 encoders. We follow the above attack's setting to use ResNet-18 [12] as encoder architecture." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.689 + ], + "angle": 0, + "content": "Experimental Settings. Our experiments are performed on two Nvidia GeForce RTX-3090 GPUs, each with a memory capacity of 24 GB. For the detector, the initial value of \\(\\lambda\\) is set up as 0.01. For detection and mitigation running overhead, the detection method with \\(10\\%\\) of the ImageNet-100 training data consumes roughly 20 minutes, and the mitigation process requires approximately 7 minutes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Evaluation Metrics. We define the following evaluation metrics to study the efficiency and effectiveness of our SSL-Cleanse. Detection Accuracy (DACC) is detection accuracy which is the ratio of correctly identified encoder types (either Benign or Trojan) relative to the total count of encoders. Attack Success Rate (ASR) is defined as the ratio of images that contain the trigger and are misclassified as the target class, to the total number of evaluated images. Accuracy (ACC) is the percentage of input images without triggers classified into their corresponding correct classes. TP indicates the true positive count, referring to Trojaned encoder numbers detected by our detector. FP represents false positives, indicating clean encoders misclassified as Trojaned encoders by our detector." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.129 + ], + "angle": 0, + "content": "M. Zheng et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.452, + 0.161 + ], + "angle": 0, + "content": "7 SSL-Cleanse Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.188, + 0.79, + 0.296 + ], + "angle": 0, + "content": "Detection. In Table 2, we present the performance of our detector on two training-agnostic SSL attacks including SSL-Backdoor and CTRL on three SSL methods and two datasets. In total, our detection accuracy (DACC) across 1200 (600 Trojaned and 600 benign) encoders stands at \\(81.33\\%\\), illustrating the efficacy of our backdoor detection capabilities. In particular, for the ImageNet-100 dataset, the DACC is \\(>77\\%\\) and the average DACC is \\(82.17\\%\\). For the CIFAR-10 dataset, the DACC is \\(>76\\%\\) and the average DACC is \\(80.5\\%\\)." + }, + { + "type": "table_caption", + "bbox": [ + 0.309, + 0.312, + 0.694, + 0.326 + ], + "angle": 0, + "content": "Table 2: The detection performance of our SSL-Cleanse." + }, + { + "type": "table", + "bbox": [ + 0.309, + 0.339, + 0.688, + 0.45 + ], + "angle": 0, + "content": "
DatasetMethodSSL-BackdoorCTRL
TPFPDACC(%)TPFPDACC(%)
CIFAR-10BYOL3558036482
SimCLR3347939584
MoCo V23157637582
ImageNet-100BYOL3888043885
SimCLR3477746888
MoCo V23677942884
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.473, + 0.79, + 0.655 + ], + "angle": 0, + "content": "Table 2 further indicates that our SSL-Cleanse can detect not only the patch-based trigger SSL attack SSL-Backdoor but also the frequency-based SSL attack CTRL. This capability stems from our method's utilization of trigger size and magnitude for patch-based trigger detection in tandem with frequency-domain detection. In particular, For SSL-Backdoor detection, SSL-Clease achieves \\(78.5\\%\\) DACC on average for both CIFAR-10 and ImageNet-100 datasets. For CTRL, it obtains \\(84.17\\%\\) average DACC. Against the CTRL, our detector identifies 46 TP from 50 Trojaned SimCLR encoders on ImageNet and registers 8 FP among 50 clean SimCLR encoders. Our SSL-Cleanse consistently exhibits reliable detection performance across prevalent SSL training techniques such as BYOL, SimCLR, and MoCo V2. In particular, our detector registers an average DACC of \\(81.75\\%\\), \\(82\\%\\), and \\(80.25\\%\\) for BYOL, SimCLR, and MoCo V2, respectively." + }, + { + "type": "table_caption", + "bbox": [ + 0.305, + 0.685, + 0.697, + 0.7 + ], + "angle": 0, + "content": "Table 3: The mitigation performance of our SSL-Cleanse." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.712, + 0.778, + 0.838 + ], + "angle": 0, + "content": "
DatasetMethodSSL-BackdoorCTRL
Before mitigationAfter mitigationBefore mitigationAfter mitigation
ACC(%)ASR(%)ACC(%)ASR(%)ACC(%)ASR(%)ACC(%)ASR(%)
CIFAR-10BYOL83.4248.3282.141.1483.1960.4782.591.96
SimCLR84.8842.1983.530.5880.7481.8479.601.15
MoCo V281.0237.9580.160.9281.4277.5180.031.62
ImageNet-100BYOL60.5733.2160.240.1453.3345.1052.650.35
SimCLR60.1831.8558.580.6252.9044.9851.040.33
MoCo V261.5735.0660.100.1750.6235.7248.880.17
" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.235, + 0.114, + 0.733, + 0.131 + ], + "angle": 0, + "content": "SSL-Cleanse: Trojan Detection and Mitigation in Self-Supervised Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.223 + ], + "angle": 0, + "content": "Mitigation. Table 3 compares the attack success rate (ASR) and clean accuracy (ACC) before and after applying the mitigator against SSL-Backdoor and CTRL on the CIFAR-10 and ImageNet-100 dataset. On average, the ASR experiences a marked reduction to below \\(2\\%\\), while the ACC declines to approximately \\(1\\%\\) before and after implementing mitigation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.788, + 0.299 + ], + "angle": 0, + "content": "In the case of patch-based SSL-Backdoor, our mitigation approach significantly reduces the ASR to below \\(1.2\\%\\), demonstrating its successful eradication of backdoor effects. Additionally, the ACC experiences an average decrease of \\(1.15\\%\\) in backdoored models after applying our mitigator. When utilizing the BYOL training method on the ImageNet-100 dataset, ACC remains relatively stable." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.3, + 0.788, + 0.39 + ], + "angle": 0, + "content": "Regarding global frequency-based SSL attack CTRL, our mitigation strategy substantially decreases the ASR to less than \\(2\\%\\), underscoring its effective elimination of backdoor attacks. Furthermore, the ACC has an average decline of \\(1.06\\%\\) in backdoored models after mitigation. Importantly, when employing the BYOL training technique on both CIFAR-10 and ImageNet-100 datasets, ACC remains fairly consistent." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.391, + 0.79, + 0.527 + ], + "angle": 0, + "content": "Efficiency on Downstream Tasks. SSL-Cleanse demonstrates effective generalization across various downstream tasks. As indicated in Table 4, SSL-Cleanse reliably brings the ASR below \\(2\\%\\) for different downstream tasks such as CIFAR-10, STL-10, and GTSRB datasets, when the SSL encoders were backdoored on CIFAR-10 dataset. In addition, SSL-Cleanse works well across all targeted task labels. For CIFAR-10, it is possible to introduce a backdoor into any label from 0 to 9. SSL-Cleanse has been tested and proven to identify attacks on each label with a detection accuracy (DACC) exceeding \\(75\\%\\), and the average DACC across 10 labels is \\(\\sim 80\\%\\)." + }, + { + "type": "table_caption", + "bbox": [ + 0.277, + 0.54, + 0.724, + 0.555 + ], + "angle": 0, + "content": "Table 4: SSL-Cleanse performance on different downstream tasks." + }, + { + "type": "table", + "bbox": [ + 0.331, + 0.566, + 0.665, + 0.641 + ], + "angle": 0, + "content": "
Downstream TaskBefore MitigationAfter Mitigation
ACC(%)ASR(%)ACC(%)ASR(%)
CIFAR-1083.4248.3282.141.14
STL-1077.5336.5575.041.84
GTSRB72.0138.2870.981.03
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.658, + 0.49, + 0.824 + ], + "angle": 0, + "content": "The SWK Effects. Figure 5 presents an ablation study on our SWK approach. Employing SWK on ImageNet100 yields a predicted number of 100, in contrast to 20 as obtained using the direct search (Kneedle) method, as depicted in Figure 3. Consequently, SWK delivers a detection accuracy of \\(80\\%\\), marking a \\(30\\%\\) DACC improvement. Data Ratio Effects. We analyze the influence of different dataset ratios on" + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.662, + 0.787, + 0.751 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.762, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Fig.5: A comparison of detection accuracy between SSL-Cleanse using the SWK method and the direct search (Kneedle)on ImageNet-100." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.825, + 0.787, + 0.841 + ], + "angle": 0, + "content": "detection performance using CTRL attack and BYOL training methods. We" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.129 + ], + "angle": 0, + "content": "M. Zheng et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.269 + ], + "angle": 0, + "content": "employ varied image proportions, namely, \\(5\\%\\), \\(8\\%\\), and \\(10\\%\\). As illustrated in Table 5, the results showcase an improvement in detection performance as the proportion of the training dataset increases. The DACC for the CIFAR-10 dataset shows a \\(1\\%\\) increase between the \\(8\\%\\) and \\(10\\%\\) training datasets. Moreover, for the ImageNet-100 dataset, there is an enhancement from \\(83\\%\\) to \\(85\\%\\) in the DACC. Further details regarding the hyperparameters \\(\\lambda\\), along with nuances about the influence of attacking trigger size and perturbation norm, can be found in our supplementary material." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.279, + 0.788, + 0.309 + ], + "angle": 0, + "content": "Table 5: Influence of the ratio of the unlabeled dataset to the entire training dataset on the detection efficacy. A larger ratio usually introduces a higher DACC." + }, + { + "type": "table", + "bbox": [ + 0.327, + 0.322, + 0.67, + 0.398 + ], + "angle": 0, + "content": "
Data ratio (%)CIFAR-10ImageNet-100
TPFPDACC(%)TPFPDACC(%)
52877138682
83787940783
103888043885
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.422, + 0.79, + 0.528 + ], + "angle": 0, + "content": "Outlier Detection Effects. Figure 6 shows our Size-Norm Trigger Outlier Detector can successfully distinguish the outlier triggers from the \\(K\\)-cluster trigger list for the trojaned encoders based on SSL-Backdoor and CTRL. In particular, each box bar plots the trigger size \\(|m_i^1|\\) and trigger norm \\(|m_i^2 \\cdot \\Delta_i^2|\\) for SSL-Backdoor and CTRL, respectively. Our Size-Norm trigger outlier detection can scan collectively, without the presumption that the defender has prior knowledge of the attack type." + }, + { + "type": "image", + "bbox": [ + 0.302, + 0.535, + 0.7, + 0.657 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.664, + 0.788, + 0.708 + ], + "angle": 0, + "content": "Fig. 6: The Size-Norm trigger outlier detection criteria is able to identify both patch-based SSL-Backdoor and frequency-domain global trigger in CTRL. The box plot shows min/max and quartiles." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.737, + 0.36, + 0.753 + ], + "angle": 0, + "content": "8 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "This paper introduces SSL-Cleanse, a novel work to detect and mitigate Trojan attacks in SSL encoders without accessing any downstream labels. We evaluated SSL-Cleanse on various datasets using 1200 models, achieving an average detection success rate of \\(82.2\\%\\) on ImageNet-100. After mitigating backdoors, backdoored encoders achieve \\(0.3\\%\\) average attack success rate without great accuracy loss." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.235, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SSL-Cleanse: Trojan Detection and Mitigation in Self-Supervised Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.173, + 0.788, + 0.214 + ], + "angle": 0, + "content": "1. Al Ghanim, M., Santriaji, M., Lou, Q., Solihin, Y.: Trojbits: A hardware aware inference-time attack on transformer-based language models. In: ECAI 2023, pp. 60-68. IOS Press (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.215, + 0.787, + 0.255 + ], + "angle": 0, + "content": "2. Campello, R.J., Moulavi, D., Zimek, A., Sander, J.: Hierarchical density estimates for data clustering, visualization, and outlier detection. ACM Transactions on Knowledge Discovery from Data (TKDD) 10(1), 1-51 (2015) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.256, + 0.787, + 0.296 + ], + "angle": 0, + "content": "3. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International conference on machine learning. pp. 1597-1607. PMLR (2020) 1, 3, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.296, + 0.787, + 0.323 + ], + "angle": 0, + "content": "4. Chen, X., Fan, H., Girshick, R., He, K.: Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297 (2020) 1, 2, 3, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.323, + 0.787, + 0.363 + ], + "angle": 0, + "content": "5. Chen, X., He, K.: Exploring simple siamese representation learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15750-15758 (2021) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.364, + 0.787, + 0.417 + ], + "angle": 0, + "content": "6. Coates, A., Ng, A., Lee, H.: An analysis of single-layer networks in unsupervised feature learning. In: Proceedings of the fourteenth international conference on artificial intelligence and statistics. pp. 215-223. JMLR Workshop and Conference Proceedings (2011) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.418, + 0.787, + 0.486 + ], + "angle": 0, + "content": "7. Deng, J., Dong, W., Socher, R., Li, L., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2009), 20-25 June 2009, Miami, Florida, USA. pp. 248-255. IEEE Computer Society (2009). https://doi.org/10.1109/CVPR.2009.5206848, https://doi.org/10.1109/CVPR.2009.5206848 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.487, + 0.787, + 0.54 + ], + "angle": 0, + "content": "8. Feng, S., Tao, G., Cheng, S., Shen, G., Xu, X., Liu, Y., Zhang, K., Ma, S., Zhang, X.: Detecting backdoors in pre-trained encoders. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16352-16362 (2023) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.541, + 0.787, + 0.579 + ], + "angle": 0, + "content": "9. Ghassabeh, Y.A.: A sufficient condition for the convergence of the mean shift algorithm with gaussian kernel. Journal of Multivariate Analysis 135, 1-10 (2015) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.581, + 0.787, + 0.636 + ], + "angle": 0, + "content": "10. Grill, J.B., Strub, F., Altché, F., Tallec, C., Richemond, P., Buchatskaya, E., Doersch, C., Avila Pires, B., Guo, Z., Gheshlaghi Azar, M., et al.: Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems 33, 21271-21284 (2020) 2, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.636, + 0.787, + 0.677 + ], + "angle": 0, + "content": "11. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729-9738 (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.677, + 0.787, + 0.745 + ], + "angle": 0, + "content": "12. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016. pp. 770-778. IEEE Computer Society (2016). https://doi.org/10.1109/CVPR.2016.90, https://doi.org/10.1109/CVPR.2016.90 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.745, + 0.787, + 0.772 + ], + "angle": 0, + "content": "13. Jaiswal, A., Babu, A.R., Zadeh, M.Z., Banerjee, D., Makedon, F.: A survey on contrastive self-supervised learning. Technologies 9(1), 2 (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.772, + 0.787, + 0.799 + ], + "angle": 0, + "content": "14. Jia, J., Liu, Y., Gong, N.Z.: Badencoder: Backdoor attacks to pre-trained encoders in self-supervised learning. arXiv preprint arXiv:2108.00352 (2021) 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "15. Kolouri, S., Saha, A., Pirsiavash, H., Hoffmann, H.: Universal litmus patterns: Revealing backdoor attacks in cnns. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 301-310 (2020) 2" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.173, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.377, + 0.129 + ], + "angle": 0, + "content": "M. Zheng et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "16. Krishnan, R., Rajpurkar, P., Topol, E.J.: Self-supervised learning in medicine and healthcare. Nature Biomedical Engineering pp. 1-7 (2022) 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.787, + 0.204 + ], + "angle": 0, + "content": "17. Krizhevsky, A.: Learning multiple layers of features from tiny images. Tech. rep., University of Toronto (2009) 4, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.206, + 0.785, + 0.247 + ], + "angle": 0, + "content": "18. Lan, Z., Chen, M., Goodman, S., Gimpel, K., Sharma, P., Soricut, R.: Albert: A lite bert for self-supervised learning of language representations. arXiv preprint arXiv:1909.11942 (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.249, + 0.785, + 0.275 + ], + "angle": 0, + "content": "19. Li, C., Pang, R., Xi, Z., Du, T., Ji, S., Yao, Y., Wang, T.: Demystifying self-supervised trojan attacks. arXiv preprint arXiv:2210.07346 (2022) 2, 4, 6, 9, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.277, + 0.785, + 0.304 + ], + "angle": 0, + "content": "20. Liu, H., Jia, J., Gong, N.Z.: Poisonedencoder: Poisoning the unlabeled pre-training data in contrastive learning. arXiv preprint arXiv:2205.06401 (2022) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.305, + 0.787, + 0.36 + ], + "angle": 0, + "content": "21. Liu, Y., Lee, W.C., Tao, G., Ma, S., Aafer, Y., Zhang, X.: Abs: Scanning neural networks for back-doors by artificial brain stimulation. In: Proceedings of the 2019 ACM SIGSAC Conference on Computer and Communications Security. pp. 1265-1282 (2019) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.361, + 0.785, + 0.401 + ], + "angle": 0, + "content": "22. Liu, Y., Jin, M., Pan, S., Zhou, C., Zheng, Y., Xia, F., Yu, P.: Graph self-supervised learning: A survey. IEEE Transactions on Knowledge and Data Engineering (2022) 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.403, + 0.787, + 0.431 + ], + "angle": 0, + "content": "23. Lou, Q., Liu, Y., Feng, B.: Trojtext: Test-time invisible textual Trojan insertion. In: The Eleventh International Conference on Learning Representations (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.433, + 0.785, + 0.473 + ], + "angle": 0, + "content": "24. Pan, M., Zeng, Y., Lyu, L., Lin, X., Jia, R.: Asset: Robust backdoor data detection across a multiplicity of deep learning paradigms. arXiv preprint arXiv:2302.11408 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.475, + 0.785, + 0.516 + ], + "angle": 0, + "content": "25. Saha, A., Tejankar, A., Koohpayegani, S.A., Pirsiavash, H.: Backdoor attacks on self-supervised learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13337-13346 (2022) 2, 4, 6, 9, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.517, + 0.787, + 0.6 + ], + "angle": 0, + "content": "26. Satopaa, V., Albrecht, J.R., Irwin, D.E., Raghavan, B.: Finding a \"kneedle\" in a haystack: Detecting knee points in system behavior. In: 31st IEEE International Conference on Distributed Computing Systems Workshops (ICDCS 2011 Workshops), 20-24 June 2011, Minneapolis, Minnesota, USA. pp. 166-171. IEEE Computer Society (2011). https://doi.org/10.1109/ICDCSW.2011.20, https://doi.org/10.1109/ICDCSW.2011.205, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.601, + 0.785, + 0.628 + ], + "angle": 0, + "content": "27. Sinaga, K.P., Yang, M.S.: Unsupervised k-means clustering algorithm. IEEE access 8, 80716-80727 (2020) 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.63, + 0.785, + 0.671 + ], + "angle": 0, + "content": "28. Stallkamp, J., Schlipsing, M., Salmen, J., Igel, C.: The german traffic sign recognition benchmark: a multi-class classification competition. In: The 2011 international joint conference on neural networks. pp. 1453-1460. IEEE (2011) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.672, + 0.787, + 0.728 + ], + "angle": 0, + "content": "29. Tejankar, A., Sanjabi, M., Wang, Q., Wang, S., Firooz, H., Pirsiavash, H., Tan, L.: Defending against patch-based backdoor attacks on self-supervised learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12239-12249 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.729, + 0.787, + 0.784 + ], + "angle": 0, + "content": "30. Wang, B., Yao, Y., Shan, S., Li, H., Viswanath, B., Zheng, H., Zhao, B.Y.: Neural cleansse: Identifying and mitigating backdoor attacks in neural networks. In: Proceedings of the IEEE Symposium on Security and Privacy (IEEE S&P). San Francisco, CA (2019) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.787, + 0.84 + ], + "angle": 0, + "content": "31. Wu, J., Zhang, T., Zha, Z.J., Luo, J., Zhang, Y., Wu, F.: Self-supervised domain-aware generative network for generalized zero-shot learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12767-12776 (2020) 2" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.235, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SSL-Cleanse: Trojan Detection and Mitigation in Self-Supervised Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.188 + ], + "angle": 0, + "content": "32. Xue, J., Lou, Q.: Estas: Effective and stable trojan attacks in self-supervised encoders with one target unlabelled sample. arXiv preprint arXiv:2211.10908 (2022) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.189, + 0.788, + 0.217 + ], + "angle": 0, + "content": "33. Yaman, B., Hosseini, S.A.H., Akçakaya, M.: Zero-shot self-supervised learning for mri reconstruction. arXiv preprint arXiv:2102.07737 (2021) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.788, + 0.272 + ], + "angle": 0, + "content": "34. Yang, H., Yang, R., Cai, H., Zhang, X., Pei, Q., Wang, S., Yan, H.: Ssl-abd: An adversarial defense method against backdoor attacks in self-supervised learning. In: International Conference on Artificial Intelligence Security and Privacy. pp. 456-467. Springer (2023) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.788, + 0.3 + ], + "angle": 0, + "content": "35. Zhang, J., Liu, H., Jia, J., Gong, N.Z.: Corruptencoder: Data poisoning based backdoor attacks to contrastive learning. arXiv preprint arXiv:2211.08229 (2022) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.301, + 0.788, + 0.342 + ], + "angle": 0, + "content": "36. Zheng, M., Lou, Q., Jiang, L.: Trojvit: Trojan insertion in vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4025-4034 (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.788, + 0.37 + ], + "angle": 0, + "content": "37. Zheng, M., Xue, J., Sheng, Y., Yang, L., Lou, Q., Jiang, L.: Trojfair: Trojan fairness attacks. arXiv preprint arXiv:2312.10508 (2023) 2" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.37 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/SSL-Cleanse_ Trojan Detection and Mitigation in Self-Supervised Learning/a42be677-fdb6-4ca1-8003-8fcf32b7a42b_origin.pdf b/2024/SSL-Cleanse_ Trojan Detection and Mitigation in Self-Supervised Learning/a42be677-fdb6-4ca1-8003-8fcf32b7a42b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b7524b313a1d7334612dbedf4963a9ed5090f475 --- /dev/null +++ b/2024/SSL-Cleanse_ Trojan Detection and Mitigation in Self-Supervised Learning/a42be677-fdb6-4ca1-8003-8fcf32b7a42b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:adaede6fe0447b4e6af33c72588b22eec19542a7cba77dcea909f6d9c0fc028c +size 1006972 diff --git a/2024/SSL-Cleanse_ Trojan Detection and Mitigation in Self-Supervised Learning/full.md b/2024/SSL-Cleanse_ Trojan Detection and Mitigation in Self-Supervised Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..26db93951f4c920f057bd9564d9bf5d77acc8a86 --- /dev/null +++ b/2024/SSL-Cleanse_ Trojan Detection and Mitigation in Self-Supervised Learning/full.md @@ -0,0 +1,249 @@ +# SSL-Cleanse: Trojan Detection and Mitigation in Self-Supervised Learning + +Mengxin Zheng $^{12*}$ , Jiaqi Xue $^{1*}$ , Zihao Wang $^{2}$ , Xun Chen $^{3}$ , Qian Lou $^{1}$ , Lei Jiang $^{2}$ , and Xiaofeng Wang $^{2}$ + +1 University of Central Florida, Orlando, Florida, USA +2 Indiana University Bloomington, Bloomington, Indiana, USA +$^{3}$ Samsung Research America, Mountain View, California, USA + +Abstract. Self-supervised learning (SSL) is a prevalent approach for encoding data representations. Using a pre-trained SSL image encoder and subsequently training a downstream classifier, impressive performance can be achieved on various tasks with very little labeled data. The growing adoption of SSL has led to an increase in security research on SSL encoders and associated Trojan attacks. Trojan attacks embedded in SSL encoders can operate covertly, spreading across multiple users and devices. The presence of backdoor behavior in Trojaned encoders can inadvertently be inherited by downstream classifiers, making it even more difficult to detect and mitigate the threat. Although current Trojan detection methods in supervised learning can potentially safeguard SSL downstream classifiers, identifying and addressing triggers in the SSL encoder before its widespread dissemination is a challenging task. This challenge arises because downstream tasks might be unknown, dataset labels may be unavailable, and the original unlabeled training dataset might be inaccessible during Trojan detection in SSL encoders. We introduce SSL-Cleanse as a solution to identify and mitigate backdoor threats in SSL encoders. We evaluated SSL-Cleanse on various datasets using 1200 encoders, achieving an average detection success rate of $82.2\%$ on ImageNet-100. After mitigating backdoors, on average, backdoored encoders achieve $0.3\%$ attack success rate without great accuracy loss, proving the effectiveness of SSL-Cleanse. The source code of SSL-Cleanse is available at https://github.com/UCF-ML-Research/SSL-Cleanse. + +Keywords: Self-supervised learning $\cdot$ Backdoor attack $\cdot$ Detection + +# 1 Introduction + +Self-supervised learning (SSL) has seen remarkable advancements, particularly in computer vision applications [3,4,16,22]. This is particularly evident when labeled examples are scarce. Unlike supervised learning, SSL sidesteps the labor-intensive labeling process, training on pretext tasks generalizable to many downstream + +![](images/cfc5b204c318dd1b856b71eb29b9249a94f48ae2bad9bc99f8626505d2365b8d.jpg) +Fig. 1: The overview of SSL-Cleanse. SSL-Cleanse has two components, Detector and Mitigator, aiming to remove the malicious behavior of Trojaned SSL encoders. + +tasks [4,5,10]. Several studies have demonstrated that SSL can achieve comparable [10] and in some cases even superior, performance in few-shot learning [31,33]. The extensive use of SSL has spurred security research and vulnerability exploration in SSL encoders, as evidenced by the emergence of various Trojan attacks [14, 19, 20, 25, 32, 35-37]. + +Malicious backdoor (a.k.a, Trojan) attacks [1, 14, 19, 23, 25] in inputs, making the compromised model classify them into a predefined target class with high confidence. If the trigger is removed from the input, the backdoored model will still exhibit normal behaviors with almost the same accuracy as its clean counterpart. One direction of SSL backdoor attacks assumes the attacker can control the training phase and modify the loss function to achieve training-control SSL backdoor [14, 32] with high attack effects. In contrast, another popular SSL backdoor direction is training-agnostic [19, 25, 36] where SSL backdoor attacks are executed in three phases. The first stage involves poisoning unlabeled datasets by adding triggers into a small fraction of target-class images. The second phase entails training the SSL encoder on the poisoned dataset to establish a connection between trigger and target-class images. In the final step, any downstream classifiers that are fine-tuned on the backdoored encoder inherit the backdoor behavior. The current training-agnostic backdoor attacks have demonstrated an attack success rate of over $98\%$ on the ImageNet-100 dataset [19]. Our goal is to scan SSL encoders and mitigate backdoor threats against such attacks. + +Trojan attacks in SSL encoders [19, 25] are perilous not only lies in their competitive attack success rates but also their covert functionality and broad reach across users and devices. Firstly, pre-trained SSL encoders are typically spread out in real-world scenarios and subsequently fine-tuned for downstream classifiers. However, these downstream classifiers may inadvertently inherit the backdoor behaviors of Trojaned encoders. While current popular Trojan detection techniques [15, 21, 30] in supervised learning may have the potential to protect SSL downstream classifiers, detecting and mitigating triggers in the SSL encoder prior to its wide distribution is a complex undertaking. Recent encoder scanning techniques, as mentioned in [8], recognize their inability to detect these Trojaned SSL encoders [19, 25] because of the distinct covert attack characteristics. We + +contend that detecting and mitigating Trojans in SSL encoders is crucial since it can impede the malicious distribution of Trojaned encoders. However, there is a research gap to bridge the popular backdoor defense methods in supervised learning with an SSL encoder. Detecting Trojans in SSL encoders is challenging due to unknown downstream tasks, unavailable dataset labels, and limited access to the original training dataset. Even the linear probe strategy, which builds downstream classifiers, fails to detect these threats, as elaborated in the Limitations of Related Backdoor Defense section. So it is crucial to implement effective detection and defense against such backdoor attacks on SSL encoders. + +This paper introduces SSL-Cleanse, a novel backdoor defense method as illustrated in Figure 1. The proposed approach comprises two main components, namely Detector and Mitigator. The Detector is responsible for identifying the presence of Trojan in an SSL encoder, and if found, the Mitigator can mitigate the backdoor attack effect. Our SSL-Cleanse overcomes the challenges of backdoor detection without knowing the labeled data and the downstream tasks. + +Our contributions can be summarized as follows: + +- We design a framework to detect training-agnostic attacks in SSL encoders without downstream labels. We reveal it is possible to prevent the dissemination of Trojaned SSL encoders via our SSL-Cleanse. +- We introduce the Sliding Window Kneedle (SWK) algorithm to auto-estimate cluster counts in unlabeled datasets, aiding representation clustering. We also present the representation-oriented trigger reverse (ROTR) method for SSL trigger inversion, alongside the Self-supervised Clustering Unlearning (SCU) algorithm to mitigate SSL encoder backdoors. +- We validate the effectiveness of the proposed SSL-Cleanse with extensive experiments on 1200 encoders. + +# 2 Background and Related Works + +Self-Supervised Learning. Leveraging the large unlabeled data available in the real world is essential. Self-Supervised Learning (SSL) is the most popular method to learn representations from complex unlabeled data [3, 16]. Pre-training an encoder with significant unlabeled data and fine-tuning it with a small amount of labeled data has been shown to achieve comparable performance to using large labeled datasets with supervised learning methods for various downstream tasks [3, 4, 11, 13, 16, 18, 22]. Furthermore, SSL techniques that rely on instance discrimination, such as SimCLR [3] and MoCo V2 [4], have become increasingly popular for learning competitive visual representations through the use of contrastive loss. Typically, an SSL classification task involves pre-training an image encoder, constructing a classifier, and subsequent fine-tuning. + +SSL Backdoor Attacks. In SSL backdoor attacks, the first line of research [14] links the trigger to the downstream-task label, necessitating triggers to be appended to varying class inputs. Compared to other directions, it assumes a stronger threat model in which the adversary dominates the training process, e.g., modifying the loss functions. Another methodology [20,32] focuses on specific + +Table 1: Limitations of current backdoor detectors in assessing SSL encoders, such as SSL-Backdoor [25] and CTRL [19]. The encoder undergoes pre-training on CIFAR-10, while the built classifiers are evaluated on CIFAR-10, STL-10, and GTSRB. These datasets have labels that overlap, partially overlap, or do not overlap with the encoder. If NC [30] Index $>2.0$ and ABS [21] REASR $>0.88$ , the model is seen as Trojaned. + +
SSL Attack MethodDownstream Task (Linear probe)NCABS
Anomaly IndexREASR
SSL-BackdoorCIFAR-102.050.89
STL-101.420.34
GTSRB1.680.29
CTRLCIFAR-101.520.52
STL-101.280.44
GTSRB1.160.37
+ +input sets, formulating poisoned data by combining these inputs with their corresponding reference representations. However, it functions only with predetermined specific inputs and not with any inputs that contain embedded triggers. A distinct approach [19, 25] avoids modifying the training phase, achieving a high attack success rate by merely connecting the trigger with desired unlabeled samples. Compared to the first approach, these SSL attacks are much more challenging to discern since triggers are solely attached to the unlabeled targets, and combining triggers with other targets does not consistently yield similar representations [8]. In our study, we aim to detect these third-category attacks. + +Limitations of Related Backdoor Defense. Prior scanners like NC [30] and ABS [21] face challenges in detecting backdoors in SSL encoders. The primary reasons include the often-unknown downstream tasks/labels. While the linear probe method, i.e., constructing downstream classifiers from various datasets using pre-trained encoders, offers an alternative, it is not efficacious in scanning SSL encoders. For example, we built a backdoored encoder using CIFAR-10 [17] with a specific label, airplane, as the target. This encoder was then utilized to train three distinct downstream classifiers on CIFAR-10, STL-10 [6], and GTSRB [28]. The results from applying NC and ABS to these classifiers are shown in Table 1. For scenarios where the encoder and the downstream task share the same dataset, both NC and ABS can detect Trojaned classifiers and hence the backdoored encoders for SSL-Backdoor [25], with encoder's Anomaly Index of 2.05 $>2.00$ in NC and a REASR of $0.89 > 0.88$ in ABS. Yet, the detection capability requires the knowledge of the downstream tasks and the detection capability diminishes when the datasets (like STL-10 or GTSRB) either only partially overlap or do not align at all with CIFAR-10. For the CTRL instance, both tools fail in detection, even when there's label congruence between the encoder and classifiers. The introduction of global frequency triggers in CTRL exacerbates the detection difficulty. While the recent study DEGREE [8] effectively detects backdoors in training-controlled attackers, like BadEncoder [14], it acknowledges the failure in identifying stealthy training-agnostic attacks [19, 25], e.g., $\sim 50\%$ detection accuracy. Also, scanners such as PatchSearch [29] and ASSET [24] are orthogonal to our method since they are introduced to identify poisoned samples + +![](images/ff4c4b85912ab744193ecda9406d15d94789cb49f783e625a74164df4b606c65.jpg) +Fig. 2: The workflow of SSL-Cleanse detector. Step 1: Unlabeled data samples are processed through the SSL encoder to compute their representations. The SWK algorithm is then utilized to process representations and determine the number of clusters. Step 2: Using K-Means with the derived cluster number $(K)$ and representation, $K$ clusters are established. Then, the Representation Oriented Trigger Reverse algorithm is employed to generate $K$ trigger patterns. Step 3: Accessing if any of the $K$ triggers are outliers in terms of their size and norm. The identified outlier indicates the encoder is Trojaned. + +in a training dataset. SSL-ABD [34] employs adversarial simulation of trigger patterns and subsequent removal of the backdoor from the compromised encoder through feature embedding distillation. However, as the defender lacks knowledge of the target class, simulating the trigger pattern becomes challenging, hindering effective backdoor removal. In our threat model, the SSL encoder is required but there's no requirement for the poisoned training dataset. + +# 3 SSL-Cleanse Design Overview + +Defense Assumptions and Goals. We assume that the defender has access to a pre-trained SSL encoder, a small portion of the unlabeled dataset (which could be distinct from the training set, i.e., SSL-Cleanse does not require attackers to disclose their poisoned dataset). + +Goals. We have two specific goals: + +- Detecting backdoor: Our aim is to make a binary determination regarding the potential backdoor infection of a given SSL encoder. If infected, we also want to identify the potential target classes of the backdoor attack. +- Mitigating backdoor: We plan to reversely generate the trigger used by the attack. Our ultimate goal is to deactivate the backdoor, making it ineffective. This requires eliminating the backdoor while preserving the classification performance of the SSL encoder for normal inputs. + +Challenges and Motivation. An inherent challenge of SSL backdoor detection lies in the indeterminate nature of the target class. While this can be addressed by iteratively searching through all classes, this method is not suitable for SSL encoders due to the uncertainty in class numbers for unlabeled datasets in SSL. Traditional clustering methods, like K-Means [27], necessitate manual predetermination of cluster counts $K$ . We observe that though automatic cluster number determination methods, such as the Kneedle algorithm [26] exist, they struggle with accurate class number prediction for SSL encoders. For instance, the Kneedle algorithm identifies only 20 classes within the ImageNet-100 dataset. + +This discrepancy could arise from noisy samples affecting the clustering of SSL representations. Other methods like DBSCAN [2] and Mean Shift [9] rely on predefined parameters like density radius and kernel width. Determining optimal values for these parameters presents a challenge. To enhance cluster number prediction accuracy, we propose the Sliding-Window Kneedle (SWK) algorithm, which addresses noise influence by averaging adjacent Kneedle scores within a sliding window. And this approach is adaptable across various clustering methods. + +Another challenge is how to efficiently generate triggers and identify outlier triggers given predicted cluster numbers, encoder, and sampled unlabeled dataset. This challenge is exaggerated when we target to detect attacks [19] based on frequency-domain global triggers since one cannot identify outlier triggers according to the trigger size. For these reasons, we are motivated to propose Representation-Oriented Trigger Reverse (ROTR) method to generate triggers via clustered representations. For identifying patch-wise trigger SSL attacks, specifically SSL-Backdoor as referenced in [25], we continue to rely on trigger size as a key outlier detection metric. However, when detecting the frequency-domain SSL attack, CTRL as cited in [19], we propose to use the magnitude of the trigger as the pivotal criterion. These detection methodologies for both attack types are unified under RORP. Upon identifying the triggers, it becomes feasible to pinpoint their related target-class clusters. Following this, we introduce the Self-supervised Clustering Unlearning (SCU) approach to mitigate backdoor threats in SSL encoders, resulting in a cleansed encoder. + +# 4 SSL-Cleanse Detector + +Workflow. The workflow of the SSL-Cleanse detector, as shown in Figure 2, consists of three steps. First, the SSL encoder processes a small amount of unlabeled data to generate representations. Using these, the Sliding Window Kneedle (SWK) algorithm determines the cluster count $K$ . Subsequently, K-Means creates $K$ clusters, from which the Representation Oriented Trigger Reverse algorithm derives $K$ trigger patterns. Finally, if any trigger significantly deviates in size or magnitude/norm, the encoder is deemed Trojaned. The details of each step and the associated algorithms are elaborated below. + +Sliding Window Kneedle. Class numbers for unlabeled training samples in SSL encoders are often unknown, even to model developers. + +At first glance, clustering appears as an intuitive approach to determining class numbers. Yet, conventional methods such as K-Means [27] demand a predetermined value for $K$ . Opting for an automated determination of the cluster number $K$ offers a more flexible solution. To this end, we initially investigated existing automatic methods, with a focus on the Kneedle algorithm [26]. + +Given the SSL samples $D$ and encoder $f$ , the Kneedle algorithm starts by initializing a list with potential $K$ values to examine. For each $K$ in this list, the K-Means method clusters the representation $f(D)$ , producing a clustering outcome. The silhouette score for this clustering ranges from -1 to 1. A high + +![](images/2355fc664457791348c0ce16a48bb8869ede0cc06b6679d02560e9ee06f3ce34.jpg) +Fig. 3: Comparison of our SWK method and direct search (Kneedle) method on ImageNet-100 dataset. Our SWK method yields more stable and accurate K. + +value signifies that the data point is well-suited to its own cluster and has a poor fit with neighboring clusters. + +The silhouette score for a particular data point is computed using: $s(i) = \frac{b(i) - a(i)}{\max\{a(i), b(i)\}}$ where $s(i)$ represents the silhouette coefficient for data point $i$ , $a(i)$ denotes the mean distance between the $i^{th}$ data point and other points within the same cluster, $b(i)$ is the smallest mean distance between the $i^{th}$ data point and points in a different cluster, minimized over all clusters. The clusters' overall silhouette score is the mean silhouette coefficient of all instances. + +The Knee/elbow point on the plot, which represents silhouette scores over varying $K$ values, indicates the optimal position. To identify the Knee point, one should normalize the silhouette curve, adjust the origin to [0, 0] and ensure the endpoint aligns with [1, 1]. Next, measure the distance of each point on the curve from the direct line connecting the origin [0, 0] to the endpoint [1, 1]. The point that is furthest from this direct line is recognized as the knee point of the curve + +Algorithm 1 Sliding Window Kneedle for SSL Cluster Num. +```txt +Input: SSL samples $D$ encoder $f$ pre-defined K_list Output: predicted cluster number $K$ initialize clusters_list,s_list,padded_s_list,d_list $\equiv []$ for $i = 0$ to len(K_list) do clusters_list.append(kmeans(f(D),K_list[i])) s_list.append(sinhouette(f(D),clusters_list[i])) +end for +initialize window size $w$ as a small odd number, e.g., 3 initialize swk_s_list to zero values of s_list's structure padded_s_list $\leftarrow$ pad $\frac{w - 1}{2}$ zeros to head and tail of s_list +for $i = 1$ to len(s_list) do swk_s_list[i] $\equiv \frac{1}{w}\sum_{j = 0}^{w}$ padded_s_list[i+j] d_list.append(norm(swk_s_list[i])-norm(K_list)) +end for + $K\gets$ index of maximum entry in (d_list) +``` + +In Figure 3, we illustrate the silhouette curvature distance for each point $K$ . The line corresponding to the direct search showcases the application of the Kneedle algorithm. The most considerable distance is observed when $K = 20$ , indicating the elbow point of the Silhouette score is 20. Nonetheless, this estimation might not be precise, given that the SSL encoder $f$ is trained on the ImageNet-100 dataset. This discrepancy may arise from the impact of noisy samples on the clustering of SSL representations. + +To address the high-dimensional noise and varied representations in the encoder's outputs, we enhanced the direct search (Kneedle) approach by introducing a sliding window technique with a window size of $w$ . The underlying idea is to compute the average silhouette scores for neighboring $K$ values, aiming to refine the silhouette curvature. As depicted in Algorithm 1, our new algorithm, Sliding Window Kneedle (SWK), ingests the encoder $f$ , SSL samples $D$ , and a pre-set K_list. The output is the anticipated cluster count, $K$ . The foundation of SWK relies on the Kneedle algorithm's methodology to determine K-Means clustering clusters_list[i] and silhouette scores for every specified K_list[i]. Post silhouette score list (s_list) computation, SWK introduces a window dimension to derive the mean silhouette score, symbolized as swk_s_list, for the neighboring $w$ scores. This process necessitates padding zeros to s_list, resulting in padded_s_list. The silhouette curvature distance, d_list[i], is deduced from the disparity between the normalized swk_s_list and K_list, with normalization between [0,1]. The position of the largest value in d_list dictates the predicted $K$ . Figure 3 illustrates that the evolved SWK curve offers better clarity in pinpointing the elbow juncture, diminishes noise, and assures a more precise cluster estimation. + +Algorithm 2 Representation-Oriented Trigger Reverse +Input: clustered samples $D_{i\in [1,K]}$ , encoder $f$ , cluster number $K$ , epoch number $E$ +Output: masks $m_i^1,m_i^2$ perturbation $\varDelta_{i}^{1},\varDelta_{i}^{2}$ of $K$ clusters +for $i = 1$ to $K$ do initialize masks $m_i^1,m_i^2$ and perturbation $\varDelta_{i}^{1},\varDelta_{i}^{2}$ +for $e = 1$ to $E$ do $x_{i}\gets$ randomly sample an image from $D_{i}\#$ target $x_{j}\gets$ randomly sample an image from $D_{j\neq i}$ $x_{j}^{1}\gets (1 - m_{i}^{1})\cdot x_{j} + m_{i}^{1}\cdot \varDelta_{i}^{1}$ $x_{j}^{2}\gets (1 - m_{i}^{2})\cdot x_{j} + m_{i}^{2}\cdot \varDelta_{i}^{2}$ losssize $\leftarrow \mathcal{L}_{MSE}^{size}(f(x_i),f(x_j^1))$ lossnorm $\leftarrow \mathcal{L}_{MSE}^{norm}(f(x_i),f(x_j^2))$ $m_i^1,\varDelta_i^1\gets$ update $(m_i^1,\varDelta_i^1,\text{loss\_size})$ $m_i^2,\varDelta_i^2\gets$ update $(m_i^2,\varDelta_i^2,\text{loss\_norm})$ +end for +end for + +Representation-Oriented Trigger Reverse. Once the cluster count $K$ is determined, K-Means can be readily employed to produce $K$ clustered samples, denoted by $D_{i\in [1,K]}$ , using the provided $D$ . The subsequent goal is to backtrack + +and identify Trojan triggers that are either compact in size or have a minimal $l_{1}$ -norm magnitude that can mimic the representations of target-class samples. Algorithm 2 presents the ROTR approach. The core concept involves creating two triggers for each cluster representation. It then determines which representation clusters can yield outlier triggers, either having smaller patch-based trigger sizes $|m_i^1|$ for patch-based trigger attacks or lesser trigger magnitudes $m_i^2 \cdot \Delta_i^2$ for global-trigger attacks. Initially, we designate trigger masks $m_{i}$ to specify the location of the trigger pixel, while $\varDelta_{i}$ signifies the associated pixel value for the $i$ -th trigger. Consequently, the trigger $r_i$ is computed as $m_{i} \cdot \varDelta_{i}$ . To identify both patch-based triggers and frequency-domain global triggers, we suggest initiating two distinct trigger sets: $r_i^1 = m_i^1 \cdot \varDelta_i^1$ and $r_i^2 = m_i^2 \cdot \varDelta_i^2$ , where $m_{i}^{1}$ and $m_{i}^{2}$ represent masks and $\varDelta_i^1$ and $\varDelta_i^2$ delineate the pixel values of the triggers. For each cluster, we conduct $E$ epochs to produce these twin trigger sets. Specifically, we randomly select an image from $D_{i}$ to as the clean sample $x_{i}$ and subsequently get sample to add trigger, termed $S_{j}$ , from $D_{j\neq i}$ . We then affix the pre-established two trigger sets to $x_{j}$ , resulting in $x_j^1$ and $x_j^2$ respectively. + +The clean image $x_{i}$ and images with trigger $x_{j}^{1}$ and $x_{j}^{2}$ are then sent to an encoder $f$ , and two separate loss functions are employed to update the trigger such that its representation can have more similarity with $x_{i}$ 's feature. The loss functions used in the ROTR to optimize $m_{i}^{1}, m_{i}^{2}$ and $\Delta_{i}^{1}, \Delta_{i}^{2}$ on the $i^{th}$ cluster are given by Equation 1 and Equation 2, respectively. Each loss function comprises two components. The foremost objective of the first term is to guarantee the similarity between the image patched with the trigger and the target class image in the feature space. The second term is responsible for constraining the size or the trigger magnitude/norm of the reversed trigger, we adapted $\lambda$ dynamically during optimization in our experiment. The dynamic scheduler is described in the supplementary material. Finally, the resulting trigger patterns are sent to an outlier detector for determination. + +$$ +\mathcal {L} _ {M S E} ^ {\text {s i z e}} \left(f \left(x _ {i}\right), f \left(x _ {j} ^ {1}\right)\right) = - \frac {< f \left(x _ {i}\right) , f \left(x _ {j} ^ {1}\right) >}{\left| \left| f \left(x _ {i}\right) \right| \right| \cdot \left| \left| f \left(x _ {j} ^ {1}\right) \right| \right|} + \lambda \cdot \left| m _ {i} ^ {1} \right| \tag {1} +$$ + +$$ +\mathcal {L} _ {M S E} ^ {\text {n o r m}} \left(f \left(x _ {i}\right), f \left(x _ {j} ^ {2}\right)\right) = - \frac {< f \left(x _ {i}\right) , f \left(x _ {j} ^ {2}\right) >}{\left| \left| f \left(x _ {i}\right) \right| \right| \cdot \left| \left| f \left(x _ {j} ^ {2}\right) \right| \right|} + \lambda \cdot \left| m _ {i} ^ {2} \cdot \Delta_ {i} ^ {2} \right| \tag {2} +$$ + +Size-Norm Trigger Outlier Detector. Beyond just using trigger size to identify outliers for patch-wise trigger attacks, e.g., SSL-Backdoor [25], we recognize trigger magnitude/norm as a key criterion for global-trigger attacks stemming from frequency-domain disturbances, e.g., CTRL [19]. Our proposed Size-Norm Trigger Outlier Detection (STOD) is designed to detect both patch-based and global-based triggers. In particular, given a trigger size list $[(m_1^1,\varDelta_1^1),\dots,(m_K^1,\varDelta_K^1)]$ and a trigger norm list $[(m_1^2,\varDelta_1^2),\dots,(m_K^2,\varDelta_K^2)]$ , the STOD outputs the detection result, i.e., benign or Trojaned with a trigger dictionary $t_s$ . For $K$ cluster, we iteratively check if the trigger size $|m_i^1|$ and trigger norm $|m_i^2\cdot \varDelta_i^2|$ is outlier in the list of $|m_{[1:K]}^{1}|$ and $|m_{[1:K]}^{2}\cdot \varDelta_{[1:K]}^{2}|$ using the function is_outlier $(x_{i},\pmb {x})$ . Here is_outlier $(x_{i},\pmb {x})$ returns True if $M(x_{i},\pmb {x}) > 2$ , otherwise False; The Anomaly Index function $M(x_{i},\pmb {x}) = \frac{|x_{i} - \text{median}(\pmb{x})|}{c\cdot\text{median}(|x - \text{median}(\pmb{x})|)}$ is used to ascertain if $x_{i}$ is an + +anomaly. $c$ is a constant estimator which equals 1.4826. If we discover a trigger that is substantially smaller than the other candidates, we classify it as an outlier and store its cluster number $i$ and itself in a dictionary $t_s$ , i.e., $t_s[i] = m_i \cdot \Delta_i$ . If $t_s$ remains empty, we conclude that the encoder is benign. However, if $t_s$ contains any triggers, we classify the encoder as a Trojaned encoder. + +![](images/bc56b363772a253943844dc9dabfbe9edcce2c04392b4854681f0faa81a89402.jpg) +Fig. 4: Illustration of Self-supervised Clustering Unlearning (SCU). The image $x$ is sampled from a cluster distinct from the cluster producing trigger $t$ . + +# 5 SSL-Cleanse Mitigator + +Once we have the triggers $t_s$ generated by our detector, their corresponding target-class clusters become identifiable. Subsequently, we introduce a Self-supervised Clustering Unlearning (SCU) strategy to mitigate SSL encoder backdoor threats, leading to a purified encoder. The mitigation approach is detailed in Algorithm 3. This method accepts cluster samples $D_{i\in [1,k]}$ , a Trojaned encoder $f$ , and the trigger list $t_s$ as inputs, producing a purified encoder $f'$ in return. Since $K$ clusters are generated, we need to clean them one by one. For each cluster $i$ , we iteratively select clean image $x$ from each cluster samples $D_i$ , and augment the image to create a new training sample consisting of the augmented images $x_1$ and $x_2$ . Subsequently, we randomly select a trigger from $t_s$ excluding $t_s[i]$ to make sure the image $x$ is sampled from a cluster distinct from the cluster producing trigger $t$ . Then we attach trigger $t$ to $x_2$ or directly use $x_{i2}$ without adding a trigger. The probability is $50\%$ which is the meaning of equal Sample in Algorithm 3. The trigger insertion rate is critical for balancing attack mitigation with the preservation of clean accuracy. A $50\%$ rate is shown to be optimal—it sufficiently counters attacks without compromising clean accuracy. Notice that here we pass these new training samples through the Trojaned encoder $f$ to obtain their respective representations. We then optimize the similarity between the representations using a loss function by fixing the model $f$ and updating the encoder $f'$ to eliminate the Trojan trigger effects, resulting in a clean encoder. Figure 4 illustrates the mitigation process, where a clean image of a cat $x$ is augmented to generate two cat images $x_1$ and $x_2$ , one that is the same and another that has a $50\%$ chance of being attached with a Trojan trigger. The clean image is sent to a reference model $f$ , which maintains its parameters and representations unchanged. In contrast, the other model $f'$ is updated with the new training sample to remove the Trojan trigger effect. The updated encoder + +Algorithm 3 Self-supervised Clustering Unlearning +```txt +Input: $D_{i\in [1,k]}$ ,Trojaned encoder $f$ , trigger list $t_s$ +Output: a clean encoder $f^{\prime}$ +Initialize $f^{\prime}\gets f$ +for $i = 1$ to $K$ do +for $x$ in $D_{i}$ do + $x_{1}\gets aug_{1}(x);x_{2}\gets aug_{2}(x)$ $t\gets$ randomly selected from $t_s$ excluding $t_s[i]$ $\hat{x}_2\gets$ equalSample{aug $_2(x\oplus t),aug_2(x)}$ $z,z^{\prime} = f(x_{1}),f^{\prime}(\hat{x}_{2})$ +loss $\leftarrow -s i m i l a r i t y(z,z^{\prime})$ $f^{\prime}\gets$ update(f,loss) +end for +end for +``` + +is deemed to be clean after it is able to accurately classify data, even in the presence of the Trojan trigger. + +# 6 Experimental Methodology + +Dataset. Our experiments were conducted on benchmark datasets: CIFAR-10 [17] and ImageNet-100 [7]. ImageNet-100 is a random subset of 100 classes from the larger ImageNet dataset and contains around 127,000 training images, which is widely used in prior SSL attacks [19,25]. + +SSL Attacks and Encoders. To assess the effectiveness of our detector against various attack methods, we evaluated it against two backdoor attacks, namely SSL-Backdoor [25] and CTRL [19] over BYOL [10], SimCLR [3], and MoCo V2 [4], respectively. We created 50 benign encoders and 50 Trojaned encoders for each backdoor attack setting, resulting in 1200 encoders. We follow the above attack's setting to use ResNet-18 [12] as encoder architecture. + +Experimental Settings. Our experiments are performed on two Nvidia GeForce RTX-3090 GPUs, each with a memory capacity of 24 GB. For the detector, the initial value of $\lambda$ is set up as 0.01. For detection and mitigation running overhead, the detection method with $10\%$ of the ImageNet-100 training data consumes roughly 20 minutes, and the mitigation process requires approximately 7 minutes. + +Evaluation Metrics. We define the following evaluation metrics to study the efficiency and effectiveness of our SSL-Cleanse. Detection Accuracy (DACC) is detection accuracy which is the ratio of correctly identified encoder types (either Benign or Trojan) relative to the total count of encoders. Attack Success Rate (ASR) is defined as the ratio of images that contain the trigger and are misclassified as the target class, to the total number of evaluated images. Accuracy (ACC) is the percentage of input images without triggers classified into their corresponding correct classes. TP indicates the true positive count, referring to Trojaned encoder numbers detected by our detector. FP represents false positives, indicating clean encoders misclassified as Trojaned encoders by our detector. + +# 7 SSL-Cleanse Results + +Detection. In Table 2, we present the performance of our detector on two training-agnostic SSL attacks including SSL-Backdoor and CTRL on three SSL methods and two datasets. In total, our detection accuracy (DACC) across 1200 (600 Trojaned and 600 benign) encoders stands at $81.33\%$ , illustrating the efficacy of our backdoor detection capabilities. In particular, for the ImageNet-100 dataset, the DACC is $>77\%$ and the average DACC is $82.17\%$ . For the CIFAR-10 dataset, the DACC is $>76\%$ and the average DACC is $80.5\%$ . + +Table 2: The detection performance of our SSL-Cleanse. + +
DatasetMethodSSL-BackdoorCTRL
TPFPDACC(%)TPFPDACC(%)
CIFAR-10BYOL3558036482
SimCLR3347939584
MoCo V23157637582
ImageNet-100BYOL3888043885
SimCLR3477746888
MoCo V23677942884
+ +Table 2 further indicates that our SSL-Cleanse can detect not only the patch-based trigger SSL attack SSL-Backdoor but also the frequency-based SSL attack CTRL. This capability stems from our method's utilization of trigger size and magnitude for patch-based trigger detection in tandem with frequency-domain detection. In particular, For SSL-Backdoor detection, SSL-Clease achieves $78.5\%$ DACC on average for both CIFAR-10 and ImageNet-100 datasets. For CTRL, it obtains $84.17\%$ average DACC. Against the CTRL, our detector identifies 46 TP from 50 Trojaned SimCLR encoders on ImageNet and registers 8 FP among 50 clean SimCLR encoders. Our SSL-Cleanse consistently exhibits reliable detection performance across prevalent SSL training techniques such as BYOL, SimCLR, and MoCo V2. In particular, our detector registers an average DACC of $81.75\%$ , $82\%$ , and $80.25\%$ for BYOL, SimCLR, and MoCo V2, respectively. + +Table 3: The mitigation performance of our SSL-Cleanse. + +
DatasetMethodSSL-BackdoorCTRL
Before mitigationAfter mitigationBefore mitigationAfter mitigation
ACC(%)ASR(%)ACC(%)ASR(%)ACC(%)ASR(%)ACC(%)ASR(%)
CIFAR-10BYOL83.4248.3282.141.1483.1960.4782.591.96
SimCLR84.8842.1983.530.5880.7481.8479.601.15
MoCo V281.0237.9580.160.9281.4277.5180.031.62
ImageNet-100BYOL60.5733.2160.240.1453.3345.1052.650.35
SimCLR60.1831.8558.580.6252.9044.9851.040.33
MoCo V261.5735.0660.100.1750.6235.7248.880.17
+ +Mitigation. Table 3 compares the attack success rate (ASR) and clean accuracy (ACC) before and after applying the mitigator against SSL-Backdoor and CTRL on the CIFAR-10 and ImageNet-100 dataset. On average, the ASR experiences a marked reduction to below $2\%$ , while the ACC declines to approximately $1\%$ before and after implementing mitigation. + +In the case of patch-based SSL-Backdoor, our mitigation approach significantly reduces the ASR to below $1.2\%$ , demonstrating its successful eradication of backdoor effects. Additionally, the ACC experiences an average decrease of $1.15\%$ in backdoored models after applying our mitigator. When utilizing the BYOL training method on the ImageNet-100 dataset, ACC remains relatively stable. + +Regarding global frequency-based SSL attack CTRL, our mitigation strategy substantially decreases the ASR to less than $2\%$ , underscoring its effective elimination of backdoor attacks. Furthermore, the ACC has an average decline of $1.06\%$ in backdoored models after mitigation. Importantly, when employing the BYOL training technique on both CIFAR-10 and ImageNet-100 datasets, ACC remains fairly consistent. + +Efficiency on Downstream Tasks. SSL-Cleanse demonstrates effective generalization across various downstream tasks. As indicated in Table 4, SSL-Cleanse reliably brings the ASR below $2\%$ for different downstream tasks such as CIFAR-10, STL-10, and GTSRB datasets, when the SSL encoders were backdoored on CIFAR-10 dataset. In addition, SSL-Cleanse works well across all targeted task labels. For CIFAR-10, it is possible to introduce a backdoor into any label from 0 to 9. SSL-Cleanse has been tested and proven to identify attacks on each label with a detection accuracy (DACC) exceeding $75\%$ , and the average DACC across 10 labels is $\sim 80\%$ . + +Table 4: SSL-Cleanse performance on different downstream tasks. + +
Downstream TaskBefore MitigationAfter Mitigation
ACC(%)ASR(%)ACC(%)ASR(%)
CIFAR-1083.4248.3282.141.14
STL-1077.5336.5575.041.84
GTSRB72.0138.2870.981.03
+ +The SWK Effects. Figure 5 presents an ablation study on our SWK approach. Employing SWK on ImageNet100 yields a predicted number of 100, in contrast to 20 as obtained using the direct search (Kneedle) method, as depicted in Figure 3. Consequently, SWK delivers a detection accuracy of $80\%$ , marking a $30\%$ DACC improvement. Data Ratio Effects. We analyze the influence of different dataset ratios on + +![](images/f2d0feaf62d15828b1992283a267a6363761e64dcf70c0afb242baaf87cc88d4.jpg) +Fig.5: A comparison of detection accuracy between SSL-Cleanse using the SWK method and the direct search (Kneedle)on ImageNet-100. + +detection performance using CTRL attack and BYOL training methods. We + +employ varied image proportions, namely, $5\%$ , $8\%$ , and $10\%$ . As illustrated in Table 5, the results showcase an improvement in detection performance as the proportion of the training dataset increases. The DACC for the CIFAR-10 dataset shows a $1\%$ increase between the $8\%$ and $10\%$ training datasets. Moreover, for the ImageNet-100 dataset, there is an enhancement from $83\%$ to $85\%$ in the DACC. Further details regarding the hyperparameters $\lambda$ , along with nuances about the influence of attacking trigger size and perturbation norm, can be found in our supplementary material. + +Table 5: Influence of the ratio of the unlabeled dataset to the entire training dataset on the detection efficacy. A larger ratio usually introduces a higher DACC. + +
Data ratio (%)CIFAR-10ImageNet-100
TPFPDACC(%)TPFPDACC(%)
52877138682
83787940783
103888043885
+ +Outlier Detection Effects. Figure 6 shows our Size-Norm Trigger Outlier Detector can successfully distinguish the outlier triggers from the $K$ -cluster trigger list for the trojaned encoders based on SSL-Backdoor and CTRL. In particular, each box bar plots the trigger size $|m_i^1|$ and trigger norm $|m_i^2 \cdot \Delta_i^2|$ for SSL-Backdoor and CTRL, respectively. Our Size-Norm trigger outlier detection can scan collectively, without the presumption that the defender has prior knowledge of the attack type. + +![](images/975c6f7e5ba5469c14adc53a5675f7207604d0a463a3b9d47bb6018519772511.jpg) +Fig. 6: The Size-Norm trigger outlier detection criteria is able to identify both patch-based SSL-Backdoor and frequency-domain global trigger in CTRL. The box plot shows min/max and quartiles. + +# 8 Conclusion + +This paper introduces SSL-Cleanse, a novel work to detect and mitigate Trojan attacks in SSL encoders without accessing any downstream labels. We evaluated SSL-Cleanse on various datasets using 1200 models, achieving an average detection success rate of $82.2\%$ on ImageNet-100. After mitigating backdoors, backdoored encoders achieve $0.3\%$ average attack success rate without great accuracy loss. + +# References + +1. Al Ghanim, M., Santriaji, M., Lou, Q., Solihin, Y.: Trojbits: A hardware aware inference-time attack on transformer-based language models. In: ECAI 2023, pp. 60-68. IOS Press (2023) 2 +2. Campello, R.J., Moulavi, D., Zimek, A., Sander, J.: Hierarchical density estimates for data clustering, visualization, and outlier detection. ACM Transactions on Knowledge Discovery from Data (TKDD) 10(1), 1-51 (2015) 6 +3. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International conference on machine learning. pp. 1597-1607. PMLR (2020) 1, 3, 11 +4. Chen, X., Fan, H., Girshick, R., He, K.: Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297 (2020) 1, 2, 3, 11 +5. Chen, X., He, K.: Exploring simple siamese representation learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15750-15758 (2021) 2 +6. Coates, A., Ng, A., Lee, H.: An analysis of single-layer networks in unsupervised feature learning. In: Proceedings of the fourteenth international conference on artificial intelligence and statistics. pp. 215-223. JMLR Workshop and Conference Proceedings (2011) 4 +7. Deng, J., Dong, W., Socher, R., Li, L., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2009), 20-25 June 2009, Miami, Florida, USA. pp. 248-255. IEEE Computer Society (2009). https://doi.org/10.1109/CVPR.2009.5206848, https://doi.org/10.1109/CVPR.2009.5206848 11 +8. Feng, S., Tao, G., Cheng, S., Shen, G., Xu, X., Liu, Y., Zhang, K., Ma, S., Zhang, X.: Detecting backdoors in pre-trained encoders. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16352-16362 (2023) 2, 4 +9. Ghassabeh, Y.A.: A sufficient condition for the convergence of the mean shift algorithm with gaussian kernel. Journal of Multivariate Analysis 135, 1-10 (2015) 6 +10. Grill, J.B., Strub, F., Altché, F., Tallec, C., Richemond, P., Buchatskaya, E., Doersch, C., Avila Pires, B., Guo, Z., Gheshlaghi Azar, M., et al.: Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems 33, 21271-21284 (2020) 2, 11 +11. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729-9738 (2020) 3 +12. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016. pp. 770-778. IEEE Computer Society (2016). https://doi.org/10.1109/CVPR.2016.90, https://doi.org/10.1109/CVPR.2016.90 11 +13. Jaiswal, A., Babu, A.R., Zadeh, M.Z., Banerjee, D., Makedon, F.: A survey on contrastive self-supervised learning. Technologies 9(1), 2 (2020) 3 +14. Jia, J., Liu, Y., Gong, N.Z.: Badencoder: Backdoor attacks to pre-trained encoders in self-supervised learning. arXiv preprint arXiv:2108.00352 (2021) 2, 3, 4 +15. Kolouri, S., Saha, A., Pirsiavash, H., Hoffmann, H.: Universal litmus patterns: Revealing backdoor attacks in cnns. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 301-310 (2020) 2 + +16. Krishnan, R., Rajpurkar, P., Topol, E.J.: Self-supervised learning in medicine and healthcare. Nature Biomedical Engineering pp. 1-7 (2022) 1, 3 +17. Krizhevsky, A.: Learning multiple layers of features from tiny images. Tech. rep., University of Toronto (2009) 4, 11 +18. Lan, Z., Chen, M., Goodman, S., Gimpel, K., Sharma, P., Soricut, R.: Albert: A lite bert for self-supervised learning of language representations. arXiv preprint arXiv:1909.11942 (2019) 3 +19. Li, C., Pang, R., Xi, Z., Du, T., Ji, S., Yao, Y., Wang, T.: Demystifying self-supervised trojan attacks. arXiv preprint arXiv:2210.07346 (2022) 2, 4, 6, 9, 11 +20. Liu, H., Jia, J., Gong, N.Z.: Poisonedencoder: Poisoning the unlabeled pre-training data in contrastive learning. arXiv preprint arXiv:2205.06401 (2022) 2, 3 +21. Liu, Y., Lee, W.C., Tao, G., Ma, S., Aafer, Y., Zhang, X.: Abs: Scanning neural networks for back-doors by artificial brain stimulation. In: Proceedings of the 2019 ACM SIGSAC Conference on Computer and Communications Security. pp. 1265-1282 (2019) 2, 4 +22. Liu, Y., Jin, M., Pan, S., Zhou, C., Zheng, Y., Xia, F., Yu, P.: Graph self-supervised learning: A survey. IEEE Transactions on Knowledge and Data Engineering (2022) 1, 3 +23. Lou, Q., Liu, Y., Feng, B.: Trojtext: Test-time invisible textual Trojan insertion. In: The Eleventh International Conference on Learning Representations (2023) 2 +24. Pan, M., Zeng, Y., Lyu, L., Lin, X., Jia, R.: Asset: Robust backdoor data detection across a multiplicity of deep learning paradigms. arXiv preprint arXiv:2302.11408 (2023) 4 +25. Saha, A., Tejankar, A., Koohpayegani, S.A., Pirsiavash, H.: Backdoor attacks on self-supervised learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13337-13346 (2022) 2, 4, 6, 9, 11 +26. Satopaa, V., Albrecht, J.R., Irwin, D.E., Raghavan, B.: Finding a "kneedle" in a haystack: Detecting knee points in system behavior. In: 31st IEEE International Conference on Distributed Computing Systems Workshops (ICDCS 2011 Workshops), 20-24 June 2011, Minneapolis, Minnesota, USA. pp. 166-171. IEEE Computer Society (2011). https://doi.org/10.1109/ICDCSW.2011.20, https://doi.org/10.1109/ICDCSW.2011.205, 6 +27. Sinaga, K.P., Yang, M.S.: Unsupervised k-means clustering algorithm. IEEE access 8, 80716-80727 (2020) 5, 6 +28. Stallkamp, J., Schlipsing, M., Salmen, J., Igel, C.: The german traffic sign recognition benchmark: a multi-class classification competition. In: The 2011 international joint conference on neural networks. pp. 1453-1460. IEEE (2011) 4 +29. Tejankar, A., Sanjabi, M., Wang, Q., Wang, S., Firooz, H., Pirsiavash, H., Tan, L.: Defending against patch-based backdoor attacks on self-supervised learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12239-12249 (2023) 4 +30. Wang, B., Yao, Y., Shan, S., Li, H., Viswanath, B., Zheng, H., Zhao, B.Y.: Neural cleansse: Identifying and mitigating backdoor attacks in neural networks. In: Proceedings of the IEEE Symposium on Security and Privacy (IEEE S&P). San Francisco, CA (2019) 2, 4 +31. Wu, J., Zhang, T., Zha, Z.J., Luo, J., Zhang, Y., Wu, F.: Self-supervised domain-aware generative network for generalized zero-shot learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12767-12776 (2020) 2 + +32. Xue, J., Lou, Q.: Estas: Effective and stable trojan attacks in self-supervised encoders with one target unlabelled sample. arXiv preprint arXiv:2211.10908 (2022) 2, 3 +33. Yaman, B., Hosseini, S.A.H., Akçakaya, M.: Zero-shot self-supervised learning for mri reconstruction. arXiv preprint arXiv:2102.07737 (2021) 2 +34. Yang, H., Yang, R., Cai, H., Zhang, X., Pei, Q., Wang, S., Yan, H.: Ssl-abd: An adversarial defense method against backdoor attacks in self-supervised learning. In: International Conference on Artificial Intelligence Security and Privacy. pp. 456-467. Springer (2023) 5 +35. Zhang, J., Liu, H., Jia, J., Gong, N.Z.: Corruptencoder: Data poisoning based backdoor attacks to contrastive learning. arXiv preprint arXiv:2211.08229 (2022) 2 +36. Zheng, M., Lou, Q., Jiang, L.: Trojvit: Trojan insertion in vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4025-4034 (2023) 2 +37. Zheng, M., Xue, J., Sheng, Y., Yang, L., Lou, Q., Jiang, L.: Trojfair: Trojan fairness attacks. arXiv preprint arXiv:2312.10508 (2023) 2 \ No newline at end of file diff --git a/2024/SSL-Cleanse_ Trojan Detection and Mitigation in Self-Supervised Learning/images.zip b/2024/SSL-Cleanse_ Trojan Detection and Mitigation in Self-Supervised Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..342f90fcb669e0ee582606d1f802a3135fcc0bcb --- /dev/null +++ b/2024/SSL-Cleanse_ Trojan Detection and Mitigation in Self-Supervised Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:331a393206de791c07af00785c6a7c549f677cdf7828eed43b4e8285424aa841 +size 294835 diff --git a/2024/SSL-Cleanse_ Trojan Detection and Mitigation in Self-Supervised Learning/layout.json b/2024/SSL-Cleanse_ Trojan Detection and Mitigation in Self-Supervised Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9a3db749fb606411ab9ba47d3702a4f38e0daac0 --- /dev/null +++ b/2024/SSL-Cleanse_ Trojan Detection and Mitigation in Self-Supervised Learning/layout.json @@ -0,0 +1,10507 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 135, + 111, + 478, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 111, + 478, + 148 + ], + "spans": [ + { + "bbox": [ + 135, + 111, + 478, + 148 + ], + "type": "text", + "content": "SSL-Cleanse: Trojan Detection and Mitigation in Self-Supervised Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 178, + 167, + 436, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 167, + 436, + 194 + ], + "spans": [ + { + "bbox": [ + 178, + 167, + 436, + 194 + ], + "type": "text", + "content": "Mengxin Zheng" + }, + { + "bbox": [ + 178, + 167, + 436, + 194 + ], + "type": "inline_equation", + "content": "^{12*}" + }, + { + "bbox": [ + 178, + 167, + 436, + 194 + ], + "type": "text", + "content": ", Jiaqi Xue" + }, + { + "bbox": [ + 178, + 167, + 436, + 194 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 178, + 167, + 436, + 194 + ], + "type": "text", + "content": ", Zihao Wang" + }, + { + "bbox": [ + 178, + 167, + 436, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 178, + 167, + 436, + 194 + ], + "type": "text", + "content": ", Xun Chen" + }, + { + "bbox": [ + 178, + 167, + 436, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 178, + 167, + 436, + 194 + ], + "type": "text", + "content": ", Qian Lou" + }, + { + "bbox": [ + 178, + 167, + 436, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 178, + 167, + 436, + 194 + ], + "type": "text", + "content": ", Lei Jiang" + }, + { + "bbox": [ + 178, + 167, + 436, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 178, + 167, + 436, + 194 + ], + "type": "text", + "content": ", and Xiaofeng Wang" + }, + { + "bbox": [ + 178, + 167, + 436, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 176, + 201, + 436, + 235 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 192, + 201, + 419, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 201, + 419, + 213 + ], + "spans": [ + { + "bbox": [ + 192, + 201, + 419, + 213 + ], + "type": "text", + "content": "1 University of Central Florida, Orlando, Florida, USA" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 176, + 213, + 435, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 213, + 435, + 224 + ], + "spans": [ + { + "bbox": [ + 176, + 213, + 435, + 224 + ], + "type": "text", + "content": "2 Indiana University Bloomington, Bloomington, Indiana, USA" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 176, + 224, + 436, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 224, + 436, + 235 + ], + "spans": [ + { + "bbox": [ + 176, + 224, + 436, + 235 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 176, + 224, + 436, + 235 + ], + "type": "text", + "content": " Samsung Research America, Mountain View, California, USA" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 160, + 266, + 455, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 266, + 455, + 518 + ], + "spans": [ + { + "bbox": [ + 160, + 266, + 455, + 518 + ], + "type": "text", + "content": "Abstract. Self-supervised learning (SSL) is a prevalent approach for encoding data representations. Using a pre-trained SSL image encoder and subsequently training a downstream classifier, impressive performance can be achieved on various tasks with very little labeled data. The growing adoption of SSL has led to an increase in security research on SSL encoders and associated Trojan attacks. Trojan attacks embedded in SSL encoders can operate covertly, spreading across multiple users and devices. The presence of backdoor behavior in Trojaned encoders can inadvertently be inherited by downstream classifiers, making it even more difficult to detect and mitigate the threat. Although current Trojan detection methods in supervised learning can potentially safeguard SSL downstream classifiers, identifying and addressing triggers in the SSL encoder before its widespread dissemination is a challenging task. This challenge arises because downstream tasks might be unknown, dataset labels may be unavailable, and the original unlabeled training dataset might be inaccessible during Trojan detection in SSL encoders. We introduce SSL-Cleanse as a solution to identify and mitigate backdoor threats in SSL encoders. We evaluated SSL-Cleanse on various datasets using 1200 encoders, achieving an average detection success rate of " + }, + { + "bbox": [ + 160, + 266, + 455, + 518 + ], + "type": "inline_equation", + "content": "82.2\\%" + }, + { + "bbox": [ + 160, + 266, + 455, + 518 + ], + "type": "text", + "content": " on ImageNet-100. After mitigating backdoors, on average, backdoored encoders achieve " + }, + { + "bbox": [ + 160, + 266, + 455, + 518 + ], + "type": "inline_equation", + "content": "0.3\\%" + }, + { + "bbox": [ + 160, + 266, + 455, + 518 + ], + "type": "text", + "content": " attack success rate without great accuracy loss, proving the effectiveness of SSL-Cleanse. The source code of SSL-Cleanse is available at https://github.com/UCF-ML-Research/SSL-Cleanse." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 160, + 529, + 435, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 529, + 435, + 540 + ], + "spans": [ + { + "bbox": [ + 160, + 529, + 435, + 540 + ], + "type": "text", + "content": "Keywords: Self-supervised learning " + }, + { + "bbox": [ + 160, + 529, + 435, + 540 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 529, + 435, + 540 + ], + "type": "text", + "content": " Backdoor attack " + }, + { + "bbox": [ + 160, + 529, + 435, + 540 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 529, + 435, + 540 + ], + "type": "text", + "content": " Detection" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 562, + 230, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 562, + 230, + 574 + ], + "spans": [ + { + "bbox": [ + 132, + 562, + 230, + 574 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 587, + 482, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 587, + 482, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 587, + 482, + 635 + ], + "type": "text", + "content": "Self-supervised learning (SSL) has seen remarkable advancements, particularly in computer vision applications [3,4,16,22]. This is particularly evident when labeled examples are scarce. Unlike supervised learning, SSL sidesteps the labor-intensive labeling process, training on pretext tasks generalizable to many downstream" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 643, + 338, + 654 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 643, + 338, + 654 + ], + "spans": [ + { + "bbox": [ + 133, + 643, + 338, + 654 + ], + "type": "text", + "content": "* These authors contributed equally to this work." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 143, + 654, + 365, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 654, + 365, + 665 + ], + "spans": [ + { + "bbox": [ + 143, + 654, + 365, + 665 + ], + "type": "text", + "content": "Corresponding Author Email: mengxin.zheng@ucf.edu." + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 170, + 109, + 445, + 232 + ], + "blocks": [ + { + "bbox": [ + 170, + 109, + 445, + 232 + ], + "lines": [ + { + "bbox": [ + 170, + 109, + 445, + 232 + ], + "spans": [ + { + "bbox": [ + 170, + 109, + 445, + 232 + ], + "type": "image", + "image_path": "cfc5b204c318dd1b856b71eb29b9249a94f48ae2bad9bc99f8626505d2365b8d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 239, + 482, + 264 + ], + "lines": [ + { + "bbox": [ + 131, + 239, + 482, + 264 + ], + "spans": [ + { + "bbox": [ + 131, + 239, + 482, + 264 + ], + "type": "text", + "content": "Fig. 1: The overview of SSL-Cleanse. SSL-Cleanse has two components, Detector and Mitigator, aiming to remove the malicious behavior of Trojaned SSL encoders." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 278, + 482, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 278, + 482, + 339 + ], + "spans": [ + { + "bbox": [ + 130, + 278, + 482, + 339 + ], + "type": "text", + "content": "tasks [4,5,10]. Several studies have demonstrated that SSL can achieve comparable [10] and in some cases even superior, performance in few-shot learning [31,33]. The extensive use of SSL has spurred security research and vulnerability exploration in SSL encoders, as evidenced by the emergence of various Trojan attacks [14, 19, 20, 25, 32, 35-37]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 340, + 482, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 482, + 532 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 482, + 532 + ], + "type": "text", + "content": "Malicious backdoor (a.k.a, Trojan) attacks [1, 14, 19, 23, 25] in inputs, making the compromised model classify them into a predefined target class with high confidence. If the trigger is removed from the input, the backdoored model will still exhibit normal behaviors with almost the same accuracy as its clean counterpart. One direction of SSL backdoor attacks assumes the attacker can control the training phase and modify the loss function to achieve training-control SSL backdoor [14, 32] with high attack effects. In contrast, another popular SSL backdoor direction is training-agnostic [19, 25, 36] where SSL backdoor attacks are executed in three phases. The first stage involves poisoning unlabeled datasets by adding triggers into a small fraction of target-class images. The second phase entails training the SSL encoder on the poisoned dataset to establish a connection between trigger and target-class images. In the final step, any downstream classifiers that are fine-tuned on the backdoored encoder inherit the backdoor behavior. The current training-agnostic backdoor attacks have demonstrated an attack success rate of over " + }, + { + "bbox": [ + 130, + 340, + 482, + 532 + ], + "type": "inline_equation", + "content": "98\\%" + }, + { + "bbox": [ + 130, + 340, + 482, + 532 + ], + "type": "text", + "content": " on the ImageNet-100 dataset [19]. Our goal is to scan SSL encoders and mitigate backdoor threats against such attacks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "type": "text", + "content": "Trojan attacks in SSL encoders [19, 25] are perilous not only lies in their competitive attack success rates but also their covert functionality and broad reach across users and devices. Firstly, pre-trained SSL encoders are typically spread out in real-world scenarios and subsequently fine-tuned for downstream classifiers. However, these downstream classifiers may inadvertently inherit the backdoor behaviors of Trojaned encoders. While current popular Trojan detection techniques [15, 21, 30] in supervised learning may have the potential to protect SSL downstream classifiers, detecting and mitigating triggers in the SSL encoder prior to its wide distribution is a complex undertaking. Recent encoder scanning techniques, as mentioned in [8], recognize their inability to detect these Trojaned SSL encoders [19, 25] because of the distinct covert attack characteristics. We" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "text", + "content": "M. Zheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 223 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 223 + ], + "type": "text", + "content": "contend that detecting and mitigating Trojans in SSL encoders is crucial since it can impede the malicious distribution of Trojaned encoders. However, there is a research gap to bridge the popular backdoor defense methods in supervised learning with an SSL encoder. Detecting Trojans in SSL encoders is challenging due to unknown downstream tasks, unavailable dataset labels, and limited access to the original training dataset. Even the linear probe strategy, which builds downstream classifiers, fails to detect these threats, as elaborated in the Limitations of Related Backdoor Defense section. So it is crucial to implement effective detection and defense against such backdoor attacks on SSL encoders." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 224, + 482, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 295 + ], + "type": "text", + "content": "This paper introduces SSL-Cleanse, a novel backdoor defense method as illustrated in Figure 1. The proposed approach comprises two main components, namely Detector and Mitigator. The Detector is responsible for identifying the presence of Trojan in an SSL encoder, and if found, the Mitigator can mitigate the backdoor attack effect. Our SSL-Cleanse overcomes the challenges of backdoor detection without knowing the labeled data and the downstream tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 296, + 362, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 296, + 362, + 307 + ], + "spans": [ + { + "bbox": [ + 146, + 296, + 362, + 307 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 313, + 481, + 432 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 313, + 481, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 313, + 481, + 348 + ], + "spans": [ + { + "bbox": [ + 138, + 313, + 481, + 348 + ], + "type": "text", + "content": "- We design a framework to detect training-agnostic attacks in SSL encoders without downstream labels. We reveal it is possible to prevent the dissemination of Trojaned SSL encoders via our SSL-Cleanse." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 350, + 481, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 350, + 481, + 408 + ], + "spans": [ + { + "bbox": [ + 138, + 350, + 481, + 408 + ], + "type": "text", + "content": "- We introduce the Sliding Window Kneedle (SWK) algorithm to auto-estimate cluster counts in unlabeled datasets, aiding representation clustering. We also present the representation-oriented trigger reverse (ROTR) method for SSL trigger inversion, alongside the Self-supervised Clustering Unlearning (SCU) algorithm to mitigate SSL encoder backdoors." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 409, + 481, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 409, + 481, + 432 + ], + "spans": [ + { + "bbox": [ + 138, + 409, + 481, + 432 + ], + "type": "text", + "content": "- We validate the effectiveness of the proposed SSL-Cleanse with extensive experiments on 1200 encoders." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 449, + 343, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 449, + 343, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 449, + 343, + 464 + ], + "type": "text", + "content": "2 Background and Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 474, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 605 + ], + "type": "text", + "content": "Self-Supervised Learning. Leveraging the large unlabeled data available in the real world is essential. Self-Supervised Learning (SSL) is the most popular method to learn representations from complex unlabeled data [3, 16]. Pre-training an encoder with significant unlabeled data and fine-tuning it with a small amount of labeled data has been shown to achieve comparable performance to using large labeled datasets with supervised learning methods for various downstream tasks [3, 4, 11, 13, 16, 18, 22]. Furthermore, SSL techniques that rely on instance discrimination, such as SimCLR [3] and MoCo V2 [4], have become increasingly popular for learning competitive visual representations through the use of contrastive loss. Typically, an SSL classification task involves pre-training an image encoder, constructing a classifier, and subsequent fine-tuning." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "SSL Backdoor Attacks. In SSL backdoor attacks, the first line of research [14] links the trigger to the downstream-task label, necessitating triggers to be appended to varying class inputs. Compared to other directions, it assumes a stronger threat model in which the adversary dominates the training process, e.g., modifying the loss functions. Another methodology [20,32] focuses on specific" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 143, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 143, + 90, + 447, + 102 + ], + "type": "text", + "content": "SSL-Cleanse: Trojan Detection and Mitigation in Self-Supervised Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 200, + 179, + 410, + 266 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 1: Limitations of current backdoor detectors in assessing SSL encoders, such as SSL-Backdoor [25] and CTRL [19]. The encoder undergoes pre-training on CIFAR-10, while the built classifiers are evaluated on CIFAR-10, STL-10, and GTSRB. These datasets have labels that overlap, partially overlap, or do not overlap with the encoder. If NC [30] Index " + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": ">2.0" + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": " and ABS [21] REASR " + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": ">0.88" + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": ", the model is seen as Trojaned." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 200, + 179, + 410, + 266 + ], + "lines": [ + { + "bbox": [ + 200, + 179, + 410, + 266 + ], + "spans": [ + { + "bbox": [ + 200, + 179, + 410, + 266 + ], + "type": "table", + "html": "
SSL Attack MethodDownstream Task (Linear probe)NCABS
Anomaly IndexREASR
SSL-BackdoorCIFAR-102.050.89
STL-101.420.34
GTSRB1.680.29
CTRLCIFAR-101.520.52
STL-101.280.44
GTSRB1.160.37
", + "image_path": "4cda52b3b06d5401afe2d67ac45635716ac4982d2a974048fa7efbb1b8b7c4c8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 280, + 482, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 280, + 482, + 388 + ], + "spans": [ + { + "bbox": [ + 130, + 280, + 482, + 388 + ], + "type": "text", + "content": "input sets, formulating poisoned data by combining these inputs with their corresponding reference representations. However, it functions only with predetermined specific inputs and not with any inputs that contain embedded triggers. A distinct approach [19, 25] avoids modifying the training phase, achieving a high attack success rate by merely connecting the trigger with desired unlabeled samples. Compared to the first approach, these SSL attacks are much more challenging to discern since triggers are solely attached to the unlabeled targets, and combining triggers with other targets does not consistently yield similar representations [8]. In our study, we aim to detect these third-category attacks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 390, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 482, + 666 + ], + "type": "text", + "content": "Limitations of Related Backdoor Defense. Prior scanners like NC [30] and ABS [21] face challenges in detecting backdoors in SSL encoders. The primary reasons include the often-unknown downstream tasks/labels. While the linear probe method, i.e., constructing downstream classifiers from various datasets using pre-trained encoders, offers an alternative, it is not efficacious in scanning SSL encoders. For example, we built a backdoored encoder using CIFAR-10 [17] with a specific label, airplane, as the target. This encoder was then utilized to train three distinct downstream classifiers on CIFAR-10, STL-10 [6], and GTSRB [28]. The results from applying NC and ABS to these classifiers are shown in Table 1. For scenarios where the encoder and the downstream task share the same dataset, both NC and ABS can detect Trojaned classifiers and hence the backdoored encoders for SSL-Backdoor [25], with encoder's Anomaly Index of 2.05 " + }, + { + "bbox": [ + 130, + 390, + 482, + 666 + ], + "type": "inline_equation", + "content": ">2.00" + }, + { + "bbox": [ + 130, + 390, + 482, + 666 + ], + "type": "text", + "content": " in NC and a REASR of " + }, + { + "bbox": [ + 130, + 390, + 482, + 666 + ], + "type": "inline_equation", + "content": "0.89 > 0.88" + }, + { + "bbox": [ + 130, + 390, + 482, + 666 + ], + "type": "text", + "content": " in ABS. Yet, the detection capability requires the knowledge of the downstream tasks and the detection capability diminishes when the datasets (like STL-10 or GTSRB) either only partially overlap or do not align at all with CIFAR-10. For the CTRL instance, both tools fail in detection, even when there's label congruence between the encoder and classifiers. The introduction of global frequency triggers in CTRL exacerbates the detection difficulty. While the recent study DEGREE [8] effectively detects backdoors in training-controlled attackers, like BadEncoder [14], it acknowledges the failure in identifying stealthy training-agnostic attacks [19, 25], e.g., " + }, + { + "bbox": [ + 130, + 390, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\sim 50\\%" + }, + { + "bbox": [ + 130, + 390, + 482, + 666 + ], + "type": "text", + "content": " detection accuracy. Also, scanners such as PatchSearch [29] and ASSET [24] are orthogonal to our method since they are introduced to identify poisoned samples" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "text", + "content": "M. Zheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 116, + 478, + 175 + ], + "blocks": [ + { + "bbox": [ + 135, + 116, + 478, + 175 + ], + "lines": [ + { + "bbox": [ + 135, + 116, + 478, + 175 + ], + "spans": [ + { + "bbox": [ + 135, + 116, + 478, + 175 + ], + "type": "image", + "image_path": "ff4c4b85912ab744193ecda9406d15d94789cb49f783e625a74164df4b606c65.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 182, + 482, + 260 + ], + "lines": [ + { + "bbox": [ + 130, + 182, + 482, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 482, + 260 + ], + "type": "text", + "content": "Fig. 2: The workflow of SSL-Cleanse detector. Step 1: Unlabeled data samples are processed through the SSL encoder to compute their representations. The SWK algorithm is then utilized to process representations and determine the number of clusters. Step 2: Using K-Means with the derived cluster number " + }, + { + "bbox": [ + 130, + 182, + 482, + 260 + ], + "type": "inline_equation", + "content": "(K)" + }, + { + "bbox": [ + 130, + 182, + 482, + 260 + ], + "type": "text", + "content": " and representation, " + }, + { + "bbox": [ + 130, + 182, + 482, + 260 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 182, + 482, + 260 + ], + "type": "text", + "content": " clusters are established. Then, the Representation Oriented Trigger Reverse algorithm is employed to generate " + }, + { + "bbox": [ + 130, + 182, + 482, + 260 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 182, + 482, + 260 + ], + "type": "text", + "content": " trigger patterns. Step 3: Accessing if any of the " + }, + { + "bbox": [ + 130, + 182, + 482, + 260 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 182, + 482, + 260 + ], + "type": "text", + "content": " triggers are outliers in terms of their size and norm. The identified outlier indicates the encoder is Trojaned." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 277, + 482, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 482, + 350 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 482, + 350 + ], + "type": "text", + "content": "in a training dataset. SSL-ABD [34] employs adversarial simulation of trigger patterns and subsequent removal of the backdoor from the compromised encoder through feature embedding distillation. However, as the defender lacks knowledge of the target class, simulating the trigger pattern becomes challenging, hindering effective backdoor removal. In our threat model, the SSL encoder is required but there's no requirement for the poisoned training dataset." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 369, + 331, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 331, + 383 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 331, + 383 + ], + "type": "text", + "content": "3 SSL-Cleanse Design Overview" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 396, + 481, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 396, + 481, + 443 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 481, + 443 + ], + "type": "text", + "content": "Defense Assumptions and Goals. We assume that the defender has access to a pre-trained SSL encoder, a small portion of the unlabeled dataset (which could be distinct from the training set, i.e., SSL-Cleanse does not require attackers to disclose their poisoned dataset)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 444, + 288, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 444, + 288, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 444, + 288, + 456 + ], + "type": "text", + "content": "Goals. We have two specific goals:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 464, + 482, + 550 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 138, + 464, + 481, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 464, + 481, + 500 + ], + "spans": [ + { + "bbox": [ + 138, + 464, + 481, + 500 + ], + "type": "text", + "content": "- Detecting backdoor: Our aim is to make a binary determination regarding the potential backdoor infection of a given SSL encoder. If infected, we also want to identify the potential target classes of the backdoor attack." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 501, + 482, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 501, + 482, + 550 + ], + "spans": [ + { + "bbox": [ + 138, + 501, + 482, + 550 + ], + "type": "text", + "content": "- Mitigating backdoor: We plan to reversely generate the trigger used by the attack. Our ultimate goal is to deactivate the backdoor, making it ineffective. This requires eliminating the backdoor while preserving the classification performance of the SSL encoder for normal inputs." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "Challenges and Motivation. An inherent challenge of SSL backdoor detection lies in the indeterminate nature of the target class. While this can be addressed by iteratively searching through all classes, this method is not suitable for SSL encoders due to the uncertainty in class numbers for unlabeled datasets in SSL. Traditional clustering methods, like K-Means [27], necessitate manual predetermination of cluster counts " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": ". We observe that though automatic cluster number determination methods, such as the Kneedle algorithm [26] exist, they struggle with accurate class number prediction for SSL encoders. For instance, the Kneedle algorithm identifies only 20 classes within the ImageNet-100 dataset." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 143, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 143, + 91, + 448, + 103 + ], + "type": "text", + "content": "SSL-Cleanse: Trojan Detection and Mitigation in Self-Supervised Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "This discrepancy could arise from noisy samples affecting the clustering of SSL representations. Other methods like DBSCAN [2] and Mean Shift [9] rely on predefined parameters like density radius and kernel width. Determining optimal values for these parameters presents a challenge. To enhance cluster number prediction accuracy, we propose the Sliding-Window Kneedle (SWK) algorithm, which addresses noise influence by averaging adjacent Kneedle scores within a sliding window. And this approach is adaptable across various clustering methods." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 201, + 483, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 201, + 483, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 201, + 483, + 380 + ], + "type": "text", + "content": "Another challenge is how to efficiently generate triggers and identify outlier triggers given predicted cluster numbers, encoder, and sampled unlabeled dataset. This challenge is exaggerated when we target to detect attacks [19] based on frequency-domain global triggers since one cannot identify outlier triggers according to the trigger size. For these reasons, we are motivated to propose Representation-Oriented Trigger Reverse (ROTR) method to generate triggers via clustered representations. For identifying patch-wise trigger SSL attacks, specifically SSL-Backdoor as referenced in [25], we continue to rely on trigger size as a key outlier detection metric. However, when detecting the frequency-domain SSL attack, CTRL as cited in [19], we propose to use the magnitude of the trigger as the pivotal criterion. These detection methodologies for both attack types are unified under RORP. Upon identifying the triggers, it becomes feasible to pinpoint their related target-class clusters. Following this, we introduce the Self-supervised Clustering Unlearning (SCU) approach to mitigate backdoor threats in SSL encoders, resulting in a cleansed encoder." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 404, + 285, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 404, + 285, + 417 + ], + "spans": [ + { + "bbox": [ + 132, + 404, + 285, + 417 + ], + "type": "text", + "content": "4 SSL-Cleanse Detector" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 434, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 482, + 529 + ], + "type": "text", + "content": "Workflow. The workflow of the SSL-Cleanse detector, as shown in Figure 2, consists of three steps. First, the SSL encoder processes a small amount of unlabeled data to generate representations. Using these, the Sliding Window Kneedle (SWK) algorithm determines the cluster count " + }, + { + "bbox": [ + 130, + 434, + 482, + 529 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 434, + 482, + 529 + ], + "type": "text", + "content": ". Subsequently, K-Means creates " + }, + { + "bbox": [ + 130, + 434, + 482, + 529 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 434, + 482, + 529 + ], + "type": "text", + "content": " clusters, from which the Representation Oriented Trigger Reverse algorithm derives " + }, + { + "bbox": [ + 130, + 434, + 482, + 529 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 434, + 482, + 529 + ], + "type": "text", + "content": " trigger patterns. Finally, if any trigger significantly deviates in size or magnitude/norm, the encoder is deemed Trojaned. The details of each step and the associated algorithms are elaborated below." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 531, + 481, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 531, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 531, + 481, + 555 + ], + "type": "text", + "content": "Sliding Window Kneedle. Class numbers for unlabeled training samples in SSL encoders are often unknown, even to model developers." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 557, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 616 + ], + "type": "text", + "content": "At first glance, clustering appears as an intuitive approach to determining class numbers. Yet, conventional methods such as K-Means [27] demand a predetermined value for " + }, + { + "bbox": [ + 130, + 557, + 482, + 616 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 557, + 482, + 616 + ], + "type": "text", + "content": ". Opting for an automated determination of the cluster number " + }, + { + "bbox": [ + 130, + 557, + 482, + 616 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 557, + 482, + 616 + ], + "type": "text", + "content": " offers a more flexible solution. To this end, we initially investigated existing automatic methods, with a focus on the Kneedle algorithm [26]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Given the SSL samples " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " and encoder " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ", the Kneedle algorithm starts by initializing a list with potential " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " values to examine. For each " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " in this list, the K-Means method clusters the representation " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "f(D)" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ", producing a clustering outcome. The silhouette score for this clustering ranges from -1 to 1. A high" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "text", + "content": "M. Zheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 177, + 114, + 436, + 242 + ], + "blocks": [ + { + "bbox": [ + 177, + 114, + 436, + 242 + ], + "lines": [ + { + "bbox": [ + 177, + 114, + 436, + 242 + ], + "spans": [ + { + "bbox": [ + 177, + 114, + 436, + 242 + ], + "type": "image", + "image_path": "2355fc664457791348c0ce16a48bb8869ede0cc06b6679d02560e9ee06f3ce34.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 251, + 479, + 275 + ], + "lines": [ + { + "bbox": [ + 130, + 251, + 479, + 275 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 479, + 275 + ], + "type": "text", + "content": "Fig. 3: Comparison of our SWK method and direct search (Kneedle) method on ImageNet-100 dataset. Our SWK method yields more stable and accurate K." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 284, + 479, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 284, + 479, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 284, + 479, + 308 + ], + "type": "text", + "content": "value signifies that the data point is well-suited to its own cluster and has a poor fit with neighboring clusters." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 308, + 480, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 480, + 383 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 480, + 383 + ], + "type": "text", + "content": "The silhouette score for a particular data point is computed using: " + }, + { + "bbox": [ + 130, + 308, + 480, + 383 + ], + "type": "inline_equation", + "content": "s(i) = \\frac{b(i) - a(i)}{\\max\\{a(i), b(i)\\}}" + }, + { + "bbox": [ + 130, + 308, + 480, + 383 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 308, + 480, + 383 + ], + "type": "inline_equation", + "content": "s(i)" + }, + { + "bbox": [ + 130, + 308, + 480, + 383 + ], + "type": "text", + "content": " represents the silhouette coefficient for data point " + }, + { + "bbox": [ + 130, + 308, + 480, + 383 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 308, + 480, + 383 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 308, + 480, + 383 + ], + "type": "inline_equation", + "content": "a(i)" + }, + { + "bbox": [ + 130, + 308, + 480, + 383 + ], + "type": "text", + "content": " denotes the mean distance between the " + }, + { + "bbox": [ + 130, + 308, + 480, + 383 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 130, + 308, + 480, + 383 + ], + "type": "text", + "content": " data point and other points within the same cluster, " + }, + { + "bbox": [ + 130, + 308, + 480, + 383 + ], + "type": "inline_equation", + "content": "b(i)" + }, + { + "bbox": [ + 130, + 308, + 480, + 383 + ], + "type": "text", + "content": " is the smallest mean distance between the " + }, + { + "bbox": [ + 130, + 308, + 480, + 383 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 130, + 308, + 480, + 383 + ], + "type": "text", + "content": " data point and points in a different cluster, minimized over all clusters. The clusters' overall silhouette score is the mean silhouette coefficient of all instances." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 384, + 480, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 480, + 456 + ], + "type": "text", + "content": "The Knee/elbow point on the plot, which represents silhouette scores over varying " + }, + { + "bbox": [ + 130, + 384, + 480, + 456 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 384, + 480, + 456 + ], + "type": "text", + "content": " values, indicates the optimal position. To identify the Knee point, one should normalize the silhouette curve, adjust the origin to [0, 0] and ensure the endpoint aligns with [1, 1]. Next, measure the distance of each point on the curve from the direct line connecting the origin [0, 0] to the endpoint [1, 1]. The point that is furthest from this direct line is recognized as the knee point of the curve" + } + ] + } + ], + "index": 6 + }, + { + "type": "code", + "bbox": [ + 132, + 492, + 386, + 661 + ], + "blocks": [ + { + "bbox": [ + 132, + 479, + 405, + 491 + ], + "lines": [ + { + "bbox": [ + 132, + 479, + 405, + 491 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 405, + 491 + ], + "type": "text", + "content": "Algorithm 1 Sliding Window Kneedle for SSL Cluster Num." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 132, + 492, + 386, + 661 + ], + "lines": [ + { + "bbox": [ + 132, + 492, + 386, + 661 + ], + "spans": [ + { + "bbox": [ + 132, + 492, + 386, + 661 + ], + "type": "text", + "content": "Input: SSL samples " + }, + { + "bbox": [ + 132, + 492, + 386, + 661 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 132, + 492, + 386, + 661 + ], + "type": "text", + "content": " encoder " + }, + { + "bbox": [ + 132, + 492, + 386, + 661 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 132, + 492, + 386, + 661 + ], + "type": "text", + "content": " pre-defined K_list Output: predicted cluster number " + }, + { + "bbox": [ + 132, + 492, + 386, + 661 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 132, + 492, + 386, + 661 + ], + "type": "text", + "content": " initialize clusters_list,s_list,padded_s_list,d_list " + }, + { + "bbox": [ + 132, + 492, + 386, + 661 + ], + "type": "inline_equation", + "content": "\\equiv []" + }, + { + "bbox": [ + 132, + 492, + 386, + 661 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 132, + 492, + 386, + 661 + ], + "type": "inline_equation", + "content": "i = 0" + }, + { + "bbox": [ + 132, + 492, + 386, + 661 + ], + "type": "text", + "content": " to len(K_list) do clusters_list.append(kmeans(f(D),K_list[i])) s_list.append(sinhouette(f(D),clusters_list[i])) \nend for \ninitialize window size " + }, + { + "bbox": [ + 132, + 492, + 386, + 661 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 132, + 492, + 386, + 661 + ], + "type": "text", + "content": " as a small odd number, e.g., 3 initialize swk_s_list to zero values of s_list's structure padded_s_list " + }, + { + "bbox": [ + 132, + 492, + 386, + 661 + ], + "type": "inline_equation", + "content": "\\leftarrow" + }, + { + "bbox": [ + 132, + 492, + 386, + 661 + ], + "type": "text", + "content": " pad " + }, + { + "bbox": [ + 132, + 492, + 386, + 661 + ], + "type": "inline_equation", + "content": "\\frac{w - 1}{2}" + }, + { + "bbox": [ + 132, + 492, + 386, + 661 + ], + "type": "text", + "content": " zeros to head and tail of s_list \nfor " + }, + { + "bbox": [ + 132, + 492, + 386, + 661 + ], + "type": "inline_equation", + "content": "i = 1" + }, + { + "bbox": [ + 132, + 492, + 386, + 661 + ], + "type": "text", + "content": " to len(s_list) do swk_s_list[i] " + }, + { + "bbox": [ + 132, + 492, + 386, + 661 + ], + "type": "inline_equation", + "content": "\\equiv \\frac{1}{w}\\sum_{j = 0}^{w}" + }, + { + "bbox": [ + 132, + 492, + 386, + 661 + ], + "type": "text", + "content": " padded_s_list[i+j] d_list.append(norm(swk_s_list[i])-norm(K_list)) \nend for \n" + }, + { + "bbox": [ + 132, + 492, + 386, + 661 + ], + "type": "inline_equation", + "content": "K\\gets" + }, + { + "bbox": [ + 132, + 492, + 386, + 661 + ], + "type": "text", + "content": " index of maximum entry in (d_list)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "code_body" + } + ], + "index": 8, + "sub_type": "code", + "guess_lang": "txt" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 143, + 90, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 90, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 143, + 90, + 448, + 103 + ], + "type": "text", + "content": "SSL-Cleanse: Trojan Detection and Mitigation in Self-Supervised Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "In Figure 3, we illustrate the silhouette curvature distance for each point " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": ". The line corresponding to the direct search showcases the application of the Kneedle algorithm. The most considerable distance is observed when " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "K = 20" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": ", indicating the elbow point of the Silhouette score is 20. Nonetheless, this estimation might not be precise, given that the SSL encoder " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " is trained on the ImageNet-100 dataset. This discrepancy may arise from the impact of noisy samples on the clustering of SSL representations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 483, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 483, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 483, + 403 + ], + "type": "text", + "content": "To address the high-dimensional noise and varied representations in the encoder's outputs, we enhanced the direct search (Kneedle) approach by introducing a sliding window technique with a window size of " + }, + { + "bbox": [ + 130, + 200, + 483, + 403 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 130, + 200, + 483, + 403 + ], + "type": "text", + "content": ". The underlying idea is to compute the average silhouette scores for neighboring " + }, + { + "bbox": [ + 130, + 200, + 483, + 403 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 200, + 483, + 403 + ], + "type": "text", + "content": " values, aiming to refine the silhouette curvature. As depicted in Algorithm 1, our new algorithm, Sliding Window Kneedle (SWK), ingests the encoder " + }, + { + "bbox": [ + 130, + 200, + 483, + 403 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 200, + 483, + 403 + ], + "type": "text", + "content": ", SSL samples " + }, + { + "bbox": [ + 130, + 200, + 483, + 403 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 200, + 483, + 403 + ], + "type": "text", + "content": ", and a pre-set K_list. The output is the anticipated cluster count, " + }, + { + "bbox": [ + 130, + 200, + 483, + 403 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 200, + 483, + 403 + ], + "type": "text", + "content": ". The foundation of SWK relies on the Kneedle algorithm's methodology to determine K-Means clustering clusters_list[i] and silhouette scores for every specified K_list[i]. Post silhouette score list (s_list) computation, SWK introduces a window dimension to derive the mean silhouette score, symbolized as swk_s_list, for the neighboring " + }, + { + "bbox": [ + 130, + 200, + 483, + 403 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 130, + 200, + 483, + 403 + ], + "type": "text", + "content": " scores. This process necessitates padding zeros to s_list, resulting in padded_s_list. The silhouette curvature distance, d_list[i], is deduced from the disparity between the normalized swk_s_list and K_list, with normalization between [0,1]. The position of the largest value in d_list dictates the predicted " + }, + { + "bbox": [ + 130, + 200, + 483, + 403 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 200, + 483, + 403 + ], + "type": "text", + "content": ". Figure 3 illustrates that the evolved SWK curve offers better clarity in pinpointing the elbow juncture, diminishes noise, and assures a more precise cluster estimation." + } + ] + } + ], + "index": 3 + }, + { + "type": "code", + "bbox": [ + 132, + 437, + 481, + 606 + ], + "blocks": [ + { + "bbox": [ + 133, + 425, + 378, + 436 + ], + "lines": [ + { + "bbox": [ + 133, + 425, + 378, + 436 + ], + "spans": [ + { + "bbox": [ + 133, + 425, + 378, + 436 + ], + "type": "text", + "content": "Algorithm 2 Representation-Oriented Trigger Reverse" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 132, + 437, + 481, + 606 + ], + "lines": [ + { + "bbox": [ + 132, + 437, + 481, + 606 + ], + "spans": [ + { + "bbox": [ + 132, + 437, + 481, + 606 + ], + "type": "text", + "content": "Input: clustered samples " + }, + { + "bbox": [ + 132, + 437, + 481, + 606 + ], + "type": "inline_equation", + "content": "D_{i\\in [1,K]}" + }, + { + "bbox": [ + 132, + 437, + 481, + 606 + ], + "type": "text", + "content": " , encoder " + }, + { + "bbox": [ + 132, + 437, + 481, + 606 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 132, + 437, + 481, + 606 + ], + "type": "text", + "content": " , cluster number " + }, + { + "bbox": [ + 132, + 437, + 481, + 606 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 132, + 437, + 481, + 606 + ], + "type": "text", + "content": " , epoch number " + }, + { + "bbox": [ + 132, + 437, + 481, + 606 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 132, + 437, + 481, + 606 + ], + "type": "text", + "content": " \nOutput: masks " + }, + { + "bbox": [ + 132, + 437, + 481, + 606 + ], + "type": "inline_equation", + "content": "m_i^1,m_i^2" + }, + { + "bbox": [ + 132, + 437, + 481, + 606 + ], + "type": "text", + "content": " perturbation " + }, + { + "bbox": [ + 132, + 437, + 481, + 606 + ], + "type": "inline_equation", + "content": "\\varDelta_{i}^{1},\\varDelta_{i}^{2}" + }, + { + "bbox": [ + 132, + 437, + 481, + 606 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 132, + 437, + 481, + 606 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 132, + 437, + 481, + 606 + ], + "type": "text", + "content": " clusters \nfor " + }, + { + "bbox": [ + 132, + 437, + 481, + 606 + ], + "type": "inline_equation", + "content": "i = 1" + }, + { + "bbox": [ + 132, + 437, + 481, + 606 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 132, + 437, + 481, + 606 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 132, + 437, + 481, + 606 + ], + "type": "text", + "content": " do initialize masks " + }, + { + "bbox": [ + 132, + 437, + 481, + 606 + ], + "type": "inline_equation", + "content": "m_i^1,m_i^2" + }, + { + "bbox": [ + 132, + 437, + 481, + 606 + ], + "type": "text", + "content": " and perturbation " + }, + { + "bbox": [ + 132, + 437, + 481, + 606 + ], + "type": "inline_equation", + "content": "\\varDelta_{i}^{1},\\varDelta_{i}^{2}" + }, + { + "bbox": [ + 132, + 437, + 481, + 606 + ], + "type": "text", + "content": " \nfor " + }, + { + "bbox": [ + 132, + 437, + 481, + 606 + ], + "type": "inline_equation", + "content": "e = 1" + }, + { + "bbox": [ + 132, + 437, + 481, + 606 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 132, + 437, + 481, + 606 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 132, + 437, + 481, + 606 + ], + "type": "text", + "content": " do " + }, + { + "bbox": [ + 132, + 437, + 481, + 606 + ], + "type": "inline_equation", + "content": "x_{i}\\gets" + }, + { + "bbox": [ + 132, + 437, + 481, + 606 + ], + "type": "text", + "content": " randomly sample an image from " + }, + { + "bbox": [ + 132, + 437, + 481, + 606 + ], + "type": "inline_equation", + "content": "D_{i}\\#" + }, + { + "bbox": [ + 132, + 437, + 481, + 606 + ], + "type": "text", + "content": " target " + }, + { + "bbox": [ + 132, + 437, + 481, + 606 + ], + "type": "inline_equation", + "content": "x_{j}\\gets" + }, + { + "bbox": [ + 132, + 437, + 481, + 606 + ], + "type": "text", + "content": " randomly sample an image from " + }, + { + "bbox": [ + 132, + 437, + 481, + 606 + ], + "type": "inline_equation", + "content": "D_{j\\neq i}" + }, + { + "bbox": [ + 132, + 437, + 481, + 606 + ], + "type": "inline_equation", + "content": "x_{j}^{1}\\gets (1 - m_{i}^{1})\\cdot x_{j} + m_{i}^{1}\\cdot \\varDelta_{i}^{1}" + }, + { + "bbox": [ + 132, + 437, + 481, + 606 + ], + "type": "inline_equation", + "content": "x_{j}^{2}\\gets (1 - m_{i}^{2})\\cdot x_{j} + m_{i}^{2}\\cdot \\varDelta_{i}^{2}" + }, + { + "bbox": [ + 132, + 437, + 481, + 606 + ], + "type": "text", + "content": " losssize " + }, + { + "bbox": [ + 132, + 437, + 481, + 606 + ], + "type": "inline_equation", + "content": "\\leftarrow \\mathcal{L}_{MSE}^{size}(f(x_i),f(x_j^1))" + }, + { + "bbox": [ + 132, + 437, + 481, + 606 + ], + "type": "text", + "content": " lossnorm " + }, + { + "bbox": [ + 132, + 437, + 481, + 606 + ], + "type": "inline_equation", + "content": "\\leftarrow \\mathcal{L}_{MSE}^{norm}(f(x_i),f(x_j^2))" + }, + { + "bbox": [ + 132, + 437, + 481, + 606 + ], + "type": "inline_equation", + "content": "m_i^1,\\varDelta_i^1\\gets" + }, + { + "bbox": [ + 132, + 437, + 481, + 606 + ], + "type": "text", + "content": " update " + }, + { + "bbox": [ + 132, + 437, + 481, + 606 + ], + "type": "inline_equation", + "content": "(m_i^1,\\varDelta_i^1,\\text{loss\\_size})" + }, + { + "bbox": [ + 132, + 437, + 481, + 606 + ], + "type": "inline_equation", + "content": "m_i^2,\\varDelta_i^2\\gets" + }, + { + "bbox": [ + 132, + 437, + 481, + 606 + ], + "type": "text", + "content": " update " + }, + { + "bbox": [ + 132, + 437, + 481, + 606 + ], + "type": "inline_equation", + "content": "(m_i^2,\\varDelta_i^2,\\text{loss\\_norm})" + }, + { + "bbox": [ + 132, + 437, + 481, + 606 + ], + "type": "text", + "content": " \nend for \nend for" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "code_body" + } + ], + "index": 5, + "sub_type": "algorithm" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "Representation-Oriented Trigger Reverse. Once the cluster count " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " is determined, K-Means can be readily employed to produce " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " clustered samples, denoted by " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "D_{i\\in [1,K]}" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": ", using the provided " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": ". The subsequent goal is to backtrack" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "text", + "content": "M. Zheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "content": "and identify Trojan triggers that are either compact in size or have a minimal " + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "inline_equation", + "content": "l_{1}" + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "content": "-norm magnitude that can mimic the representations of target-class samples. Algorithm 2 presents the ROTR approach. The core concept involves creating two triggers for each cluster representation. It then determines which representation clusters can yield outlier triggers, either having smaller patch-based trigger sizes " + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "inline_equation", + "content": "|m_i^1|" + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "content": " for patch-based trigger attacks or lesser trigger magnitudes " + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "inline_equation", + "content": "m_i^2 \\cdot \\Delta_i^2" + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "content": " for global-trigger attacks. Initially, we designate trigger masks " + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "inline_equation", + "content": "m_{i}" + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "content": " to specify the location of the trigger pixel, while " + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\varDelta_{i}" + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "content": " signifies the associated pixel value for the " + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "content": "-th trigger. Consequently, the trigger " + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "content": " is computed as " + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "inline_equation", + "content": "m_{i} \\cdot \\varDelta_{i}" + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "content": ". To identify both patch-based triggers and frequency-domain global triggers, we suggest initiating two distinct trigger sets: " + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "inline_equation", + "content": "r_i^1 = m_i^1 \\cdot \\varDelta_i^1" + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "inline_equation", + "content": "r_i^2 = m_i^2 \\cdot \\varDelta_i^2" + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "inline_equation", + "content": "m_{i}^{1}" + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "inline_equation", + "content": "m_{i}^{2}" + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "content": " represent masks and " + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\varDelta_i^1" + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\varDelta_i^2" + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "content": " delineate the pixel values of the triggers. For each cluster, we conduct " + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "content": " epochs to produce these twin trigger sets. Specifically, we randomly select an image from " + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "inline_equation", + "content": "D_{i}" + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "content": " to as the clean sample " + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "content": " and subsequently get sample to add trigger, termed " + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "inline_equation", + "content": "S_{j}" + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "content": ", from " + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "inline_equation", + "content": "D_{j\\neq i}" + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "content": ". We then affix the pre-established two trigger sets to " + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "inline_equation", + "content": "x_{j}" + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "content": ", resulting in " + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "inline_equation", + "content": "x_j^1" + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "inline_equation", + "content": "x_j^2" + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "content": " respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 308, + 482, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 482, + 452 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 482, + 452 + ], + "type": "text", + "content": "The clean image " + }, + { + "bbox": [ + 130, + 308, + 482, + 452 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 130, + 308, + 482, + 452 + ], + "type": "text", + "content": " and images with trigger " + }, + { + "bbox": [ + 130, + 308, + 482, + 452 + ], + "type": "inline_equation", + "content": "x_{j}^{1}" + }, + { + "bbox": [ + 130, + 308, + 482, + 452 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 308, + 482, + 452 + ], + "type": "inline_equation", + "content": "x_{j}^{2}" + }, + { + "bbox": [ + 130, + 308, + 482, + 452 + ], + "type": "text", + "content": " are then sent to an encoder " + }, + { + "bbox": [ + 130, + 308, + 482, + 452 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 308, + 482, + 452 + ], + "type": "text", + "content": ", and two separate loss functions are employed to update the trigger such that its representation can have more similarity with " + }, + { + "bbox": [ + 130, + 308, + 482, + 452 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 130, + 308, + 482, + 452 + ], + "type": "text", + "content": "'s feature. The loss functions used in the ROTR to optimize " + }, + { + "bbox": [ + 130, + 308, + 482, + 452 + ], + "type": "inline_equation", + "content": "m_{i}^{1}, m_{i}^{2}" + }, + { + "bbox": [ + 130, + 308, + 482, + 452 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 308, + 482, + 452 + ], + "type": "inline_equation", + "content": "\\Delta_{i}^{1}, \\Delta_{i}^{2}" + }, + { + "bbox": [ + 130, + 308, + 482, + 452 + ], + "type": "text", + "content": " on the " + }, + { + "bbox": [ + 130, + 308, + 482, + 452 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 130, + 308, + 482, + 452 + ], + "type": "text", + "content": " cluster are given by Equation 1 and Equation 2, respectively. Each loss function comprises two components. The foremost objective of the first term is to guarantee the similarity between the image patched with the trigger and the target class image in the feature space. The second term is responsible for constraining the size or the trigger magnitude/norm of the reversed trigger, we adapted " + }, + { + "bbox": [ + 130, + 308, + 482, + 452 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 308, + 482, + 452 + ], + "type": "text", + "content": " dynamically during optimization in our experiment. The dynamic scheduler is described in the supplementary material. Finally, the resulting trigger patterns are sent to an outlier detector for determination." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 200, + 460, + 482, + 486 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 460, + 482, + 486 + ], + "spans": [ + { + "bbox": [ + 200, + 460, + 482, + 486 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {M S E} ^ {\\text {s i z e}} \\left(f \\left(x _ {i}\\right), f \\left(x _ {j} ^ {1}\\right)\\right) = - \\frac {< f \\left(x _ {i}\\right) , f \\left(x _ {j} ^ {1}\\right) >}{\\left| \\left| f \\left(x _ {i}\\right) \\right| \\right| \\cdot \\left| \\left| f \\left(x _ {j} ^ {1}\\right) \\right| \\right|} + \\lambda \\cdot \\left| m _ {i} ^ {1} \\right| \\tag {1}", + "image_path": "58a56b45bafa34fb1aa64c6216ad7280cac3d6ebb8d525341db80d5cbcc804d3.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 190, + 490, + 481, + 516 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 490, + 481, + 516 + ], + "spans": [ + { + "bbox": [ + 190, + 490, + 481, + 516 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {M S E} ^ {\\text {n o r m}} \\left(f \\left(x _ {i}\\right), f \\left(x _ {j} ^ {2}\\right)\\right) = - \\frac {< f \\left(x _ {i}\\right) , f \\left(x _ {j} ^ {2}\\right) >}{\\left| \\left| f \\left(x _ {i}\\right) \\right| \\right| \\cdot \\left| \\left| f \\left(x _ {j} ^ {2}\\right) \\right| \\right|} + \\lambda \\cdot \\left| m _ {i} ^ {2} \\cdot \\Delta_ {i} ^ {2} \\right| \\tag {2}", + "image_path": "81b869336b51f9555a5b9f2f84417a9cca562b2a7ac993472a8beffe982372cb.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 518, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 518, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 518, + 482, + 668 + ], + "type": "text", + "content": "Size-Norm Trigger Outlier Detector. Beyond just using trigger size to identify outliers for patch-wise trigger attacks, e.g., SSL-Backdoor [25], we recognize trigger magnitude/norm as a key criterion for global-trigger attacks stemming from frequency-domain disturbances, e.g., CTRL [19]. Our proposed Size-Norm Trigger Outlier Detection (STOD) is designed to detect both patch-based and global-based triggers. In particular, given a trigger size list " + }, + { + "bbox": [ + 130, + 518, + 482, + 668 + ], + "type": "inline_equation", + "content": "[(m_1^1,\\varDelta_1^1),\\dots,(m_K^1,\\varDelta_K^1)]" + }, + { + "bbox": [ + 130, + 518, + 482, + 668 + ], + "type": "text", + "content": " and a trigger norm list " + }, + { + "bbox": [ + 130, + 518, + 482, + 668 + ], + "type": "inline_equation", + "content": "[(m_1^2,\\varDelta_1^2),\\dots,(m_K^2,\\varDelta_K^2)]" + }, + { + "bbox": [ + 130, + 518, + 482, + 668 + ], + "type": "text", + "content": ", the STOD outputs the detection result, i.e., benign or Trojaned with a trigger dictionary " + }, + { + "bbox": [ + 130, + 518, + 482, + 668 + ], + "type": "inline_equation", + "content": "t_s" + }, + { + "bbox": [ + 130, + 518, + 482, + 668 + ], + "type": "text", + "content": ". For " + }, + { + "bbox": [ + 130, + 518, + 482, + 668 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 518, + 482, + 668 + ], + "type": "text", + "content": " cluster, we iteratively check if the trigger size " + }, + { + "bbox": [ + 130, + 518, + 482, + 668 + ], + "type": "inline_equation", + "content": "|m_i^1|" + }, + { + "bbox": [ + 130, + 518, + 482, + 668 + ], + "type": "text", + "content": " and trigger norm " + }, + { + "bbox": [ + 130, + 518, + 482, + 668 + ], + "type": "inline_equation", + "content": "|m_i^2\\cdot \\varDelta_i^2|" + }, + { + "bbox": [ + 130, + 518, + 482, + 668 + ], + "type": "text", + "content": " is outlier in the list of " + }, + { + "bbox": [ + 130, + 518, + 482, + 668 + ], + "type": "inline_equation", + "content": "|m_{[1:K]}^{1}|" + }, + { + "bbox": [ + 130, + 518, + 482, + 668 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 518, + 482, + 668 + ], + "type": "inline_equation", + "content": "|m_{[1:K]}^{2}\\cdot \\varDelta_{[1:K]}^{2}|" + }, + { + "bbox": [ + 130, + 518, + 482, + 668 + ], + "type": "text", + "content": " using the function is_outlier " + }, + { + "bbox": [ + 130, + 518, + 482, + 668 + ], + "type": "inline_equation", + "content": "(x_{i},\\pmb {x})" + }, + { + "bbox": [ + 130, + 518, + 482, + 668 + ], + "type": "text", + "content": ". Here is_outlier " + }, + { + "bbox": [ + 130, + 518, + 482, + 668 + ], + "type": "inline_equation", + "content": "(x_{i},\\pmb {x})" + }, + { + "bbox": [ + 130, + 518, + 482, + 668 + ], + "type": "text", + "content": " returns True if " + }, + { + "bbox": [ + 130, + 518, + 482, + 668 + ], + "type": "inline_equation", + "content": "M(x_{i},\\pmb {x}) > 2" + }, + { + "bbox": [ + 130, + 518, + 482, + 668 + ], + "type": "text", + "content": ", otherwise False; The Anomaly Index function " + }, + { + "bbox": [ + 130, + 518, + 482, + 668 + ], + "type": "inline_equation", + "content": "M(x_{i},\\pmb {x}) = \\frac{|x_{i} - \\text{median}(\\pmb{x})|}{c\\cdot\\text{median}(|x - \\text{median}(\\pmb{x})|)}" + }, + { + "bbox": [ + 130, + 518, + 482, + 668 + ], + "type": "text", + "content": " is used to ascertain if " + }, + { + "bbox": [ + 130, + 518, + 482, + 668 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 130, + 518, + 482, + 668 + ], + "type": "text", + "content": " is an" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 143, + 90, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 90, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 143, + 90, + 448, + 103 + ], + "type": "text", + "content": "SSL-Cleanse: Trojan Detection and Mitigation in Self-Supervised Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": "anomaly. " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": " is a constant estimator which equals 1.4826. If we discover a trigger that is substantially smaller than the other candidates, we classify it as an outlier and store its cluster number " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": " and itself in a dictionary " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "t_s" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "t_s[i] = m_i \\cdot \\Delta_i" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": ". If " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "t_s" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": " remains empty, we conclude that the encoder is benign. However, if " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "t_s" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": " contains any triggers, we classify the encoder as a Trojaned encoder." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 167, + 188, + 444, + 266 + ], + "blocks": [ + { + "bbox": [ + 167, + 188, + 444, + 266 + ], + "lines": [ + { + "bbox": [ + 167, + 188, + 444, + 266 + ], + "spans": [ + { + "bbox": [ + 167, + 188, + 444, + 266 + ], + "type": "image", + "image_path": "bc56b363772a253943844dc9dabfbe9edcce2c04392b4854681f0faa81a89402.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 270, + 483, + 293 + ], + "lines": [ + { + "bbox": [ + 130, + 270, + 483, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 483, + 293 + ], + "type": "text", + "content": "Fig. 4: Illustration of Self-supervised Clustering Unlearning (SCU). The image " + }, + { + "bbox": [ + 130, + 270, + 483, + 293 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 270, + 483, + 293 + ], + "type": "text", + "content": " is sampled from a cluster distinct from the cluster producing trigger " + }, + { + "bbox": [ + 130, + 270, + 483, + 293 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 270, + 483, + 293 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 325, + 289, + 339 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 289, + 339 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 289, + 339 + ], + "type": "text", + "content": "5 SSL-Cleanse Mitigator" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "text", + "content": "Once we have the triggers " + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "inline_equation", + "content": "t_s" + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "text", + "content": " generated by our detector, their corresponding target-class clusters become identifiable. Subsequently, we introduce a Self-supervised Clustering Unlearning (SCU) strategy to mitigate SSL encoder backdoor threats, leading to a purified encoder. The mitigation approach is detailed in Algorithm 3. This method accepts cluster samples " + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "inline_equation", + "content": "D_{i\\in [1,k]}" + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "text", + "content": ", a Trojaned encoder " + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "text", + "content": ", and the trigger list " + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "inline_equation", + "content": "t_s" + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "text", + "content": " as inputs, producing a purified encoder " + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "inline_equation", + "content": "f'" + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "text", + "content": " in return. Since " + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "text", + "content": " clusters are generated, we need to clean them one by one. For each cluster " + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "text", + "content": ", we iteratively select clean image " + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "text", + "content": " from each cluster samples " + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "inline_equation", + "content": "D_i" + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "text", + "content": ", and augment the image to create a new training sample consisting of the augmented images " + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "inline_equation", + "content": "x_1" + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "inline_equation", + "content": "x_2" + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "text", + "content": ". Subsequently, we randomly select a trigger from " + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "inline_equation", + "content": "t_s" + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "text", + "content": " excluding " + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "inline_equation", + "content": "t_s[i]" + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "text", + "content": " to make sure the image " + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "text", + "content": " is sampled from a cluster distinct from the cluster producing trigger " + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "text", + "content": ". Then we attach trigger " + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "inline_equation", + "content": "x_2" + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "text", + "content": " or directly use " + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "inline_equation", + "content": "x_{i2}" + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "text", + "content": " without adding a trigger. The probability is " + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "text", + "content": " which is the meaning of equal Sample in Algorithm 3. The trigger insertion rate is critical for balancing attack mitigation with the preservation of clean accuracy. A " + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "text", + "content": " rate is shown to be optimal—it sufficiently counters attacks without compromising clean accuracy. Notice that here we pass these new training samples through the Trojaned encoder " + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "text", + "content": " to obtain their respective representations. We then optimize the similarity between the representations using a loss function by fixing the model " + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "text", + "content": " and updating the encoder " + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "inline_equation", + "content": "f'" + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "text", + "content": " to eliminate the Trojan trigger effects, resulting in a clean encoder. Figure 4 illustrates the mitigation process, where a clean image of a cat " + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "text", + "content": " is augmented to generate two cat images " + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "inline_equation", + "content": "x_1" + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "inline_equation", + "content": "x_2" + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "text", + "content": ", one that is the same and another that has a " + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "text", + "content": " chance of being attached with a Trojan trigger. The clean image is sent to a reference model " + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "text", + "content": ", which maintains its parameters and representations unchanged. In contrast, the other model " + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "inline_equation", + "content": "f'" + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "text", + "content": " is updated with the new training sample to remove the Trojan trigger effect. The updated encoder" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "text", + "content": "M. Zheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 134, + 130, + 348, + 273 + ], + "blocks": [ + { + "bbox": [ + 133, + 115, + 363, + 128 + ], + "lines": [ + { + "bbox": [ + 133, + 115, + 363, + 128 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 363, + 128 + ], + "type": "text", + "content": "Algorithm 3 Self-supervised Clustering Unlearning" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 134, + 130, + 348, + 273 + ], + "lines": [ + { + "bbox": [ + 134, + 130, + 348, + 273 + ], + "spans": [ + { + "bbox": [ + 134, + 130, + 348, + 273 + ], + "type": "text", + "content": "Input: " + }, + { + "bbox": [ + 134, + 130, + 348, + 273 + ], + "type": "inline_equation", + "content": "D_{i\\in [1,k]}" + }, + { + "bbox": [ + 134, + 130, + 348, + 273 + ], + "type": "text", + "content": " ,Trojaned encoder " + }, + { + "bbox": [ + 134, + 130, + 348, + 273 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 134, + 130, + 348, + 273 + ], + "type": "text", + "content": " , trigger list " + }, + { + "bbox": [ + 134, + 130, + 348, + 273 + ], + "type": "inline_equation", + "content": "t_s" + }, + { + "bbox": [ + 134, + 130, + 348, + 273 + ], + "type": "text", + "content": " \nOutput: a clean encoder " + }, + { + "bbox": [ + 134, + 130, + 348, + 273 + ], + "type": "inline_equation", + "content": "f^{\\prime}" + }, + { + "bbox": [ + 134, + 130, + 348, + 273 + ], + "type": "text", + "content": " \nInitialize " + }, + { + "bbox": [ + 134, + 130, + 348, + 273 + ], + "type": "inline_equation", + "content": "f^{\\prime}\\gets f" + }, + { + "bbox": [ + 134, + 130, + 348, + 273 + ], + "type": "text", + "content": " \nfor " + }, + { + "bbox": [ + 134, + 130, + 348, + 273 + ], + "type": "inline_equation", + "content": "i = 1" + }, + { + "bbox": [ + 134, + 130, + 348, + 273 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 134, + 130, + 348, + 273 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 134, + 130, + 348, + 273 + ], + "type": "text", + "content": " do \nfor " + }, + { + "bbox": [ + 134, + 130, + 348, + 273 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 134, + 130, + 348, + 273 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 134, + 130, + 348, + 273 + ], + "type": "inline_equation", + "content": "D_{i}" + }, + { + "bbox": [ + 134, + 130, + 348, + 273 + ], + "type": "text", + "content": " do \n" + }, + { + "bbox": [ + 134, + 130, + 348, + 273 + ], + "type": "inline_equation", + "content": "x_{1}\\gets aug_{1}(x);x_{2}\\gets aug_{2}(x)" + }, + { + "bbox": [ + 134, + 130, + 348, + 273 + ], + "type": "inline_equation", + "content": "t\\gets" + }, + { + "bbox": [ + 134, + 130, + 348, + 273 + ], + "type": "text", + "content": " randomly selected from " + }, + { + "bbox": [ + 134, + 130, + 348, + 273 + ], + "type": "inline_equation", + "content": "t_s" + }, + { + "bbox": [ + 134, + 130, + 348, + 273 + ], + "type": "text", + "content": " excluding " + }, + { + "bbox": [ + 134, + 130, + 348, + 273 + ], + "type": "inline_equation", + "content": "t_s[i]" + }, + { + "bbox": [ + 134, + 130, + 348, + 273 + ], + "type": "inline_equation", + "content": "\\hat{x}_2\\gets" + }, + { + "bbox": [ + 134, + 130, + 348, + 273 + ], + "type": "text", + "content": " equalSample{aug " + }, + { + "bbox": [ + 134, + 130, + 348, + 273 + ], + "type": "inline_equation", + "content": "_2(x\\oplus t),aug_2(x)}" + }, + { + "bbox": [ + 134, + 130, + 348, + 273 + ], + "type": "inline_equation", + "content": "z,z^{\\prime} = f(x_{1}),f^{\\prime}(\\hat{x}_{2})" + }, + { + "bbox": [ + 134, + 130, + 348, + 273 + ], + "type": "text", + "content": " \nloss " + }, + { + "bbox": [ + 134, + 130, + 348, + 273 + ], + "type": "inline_equation", + "content": "\\leftarrow -s i m i l a r i t y(z,z^{\\prime})" + }, + { + "bbox": [ + 134, + 130, + 348, + 273 + ], + "type": "inline_equation", + "content": "f^{\\prime}\\gets" + }, + { + "bbox": [ + 134, + 130, + 348, + 273 + ], + "type": "text", + "content": " update(f,loss) \nend for \nend for" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "code", + "guess_lang": "txt" + }, + { + "bbox": [ + 130, + 298, + 480, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 298, + 480, + 323 + ], + "spans": [ + { + "bbox": [ + 130, + 298, + 480, + 323 + ], + "type": "text", + "content": "is deemed to be clean after it is able to accurately classify data, even in the presence of the Trojan trigger." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 341, + 315, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 341, + 315, + 355 + ], + "spans": [ + { + "bbox": [ + 131, + 341, + 315, + 355 + ], + "type": "text", + "content": "6 Experimental Methodology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 365, + 482, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 414 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 414 + ], + "type": "text", + "content": "Dataset. Our experiments were conducted on benchmark datasets: CIFAR-10 [17] and ImageNet-100 [7]. ImageNet-100 is a random subset of 100 classes from the larger ImageNet dataset and contains around 127,000 training images, which is widely used in prior SSL attacks [19,25]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 415, + 482, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 482, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 482, + 485 + ], + "type": "text", + "content": "SSL Attacks and Encoders. To assess the effectiveness of our detector against various attack methods, we evaluated it against two backdoor attacks, namely SSL-Backdoor [25] and CTRL [19] over BYOL [10], SimCLR [3], and MoCo V2 [4], respectively. We created 50 benign encoders and 50 Trojaned encoders for each backdoor attack setting, resulting in 1200 encoders. We follow the above attack's setting to use ResNet-18 [12] as encoder architecture." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 486, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 545 + ], + "type": "text", + "content": "Experimental Settings. Our experiments are performed on two Nvidia GeForce RTX-3090 GPUs, each with a memory capacity of 24 GB. For the detector, the initial value of " + }, + { + "bbox": [ + 130, + 486, + 482, + 545 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 486, + 482, + 545 + ], + "type": "text", + "content": " is set up as 0.01. For detection and mitigation running overhead, the detection method with " + }, + { + "bbox": [ + 130, + 486, + 482, + 545 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 486, + 482, + 545 + ], + "type": "text", + "content": " of the ImageNet-100 training data consumes roughly 20 minutes, and the mitigation process requires approximately 7 minutes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "content": "Evaluation Metrics. We define the following evaluation metrics to study the efficiency and effectiveness of our SSL-Cleanse. Detection Accuracy (DACC) is detection accuracy which is the ratio of correctly identified encoder types (either Benign or Trojan) relative to the total count of encoders. Attack Success Rate (ASR) is defined as the ratio of images that contain the trigger and are misclassified as the target class, to the total number of evaluated images. Accuracy (ACC) is the percentage of input images without triggers classified into their corresponding correct classes. TP indicates the true positive count, referring to Trojaned encoder numbers detected by our detector. FP represents false positives, indicating clean encoders misclassified as Trojaned encoders by our detector." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 143, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 143, + 90, + 447, + 102 + ], + "type": "text", + "content": "SSL-Cleanse: Trojan Detection and Mitigation in Self-Supervised Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 276, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 276, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 276, + 127 + ], + "type": "text", + "content": "7 SSL-Cleanse Results" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 148, + 483, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 148, + 483, + 234 + ], + "spans": [ + { + "bbox": [ + 130, + 148, + 483, + 234 + ], + "type": "text", + "content": "Detection. In Table 2, we present the performance of our detector on two training-agnostic SSL attacks including SSL-Backdoor and CTRL on three SSL methods and two datasets. In total, our detection accuracy (DACC) across 1200 (600 Trojaned and 600 benign) encoders stands at " + }, + { + "bbox": [ + 130, + 148, + 483, + 234 + ], + "type": "inline_equation", + "content": "81.33\\%" + }, + { + "bbox": [ + 130, + 148, + 483, + 234 + ], + "type": "text", + "content": ", illustrating the efficacy of our backdoor detection capabilities. In particular, for the ImageNet-100 dataset, the DACC is " + }, + { + "bbox": [ + 130, + 148, + 483, + 234 + ], + "type": "inline_equation", + "content": ">77\\%" + }, + { + "bbox": [ + 130, + 148, + 483, + 234 + ], + "type": "text", + "content": " and the average DACC is " + }, + { + "bbox": [ + 130, + 148, + 483, + 234 + ], + "type": "inline_equation", + "content": "82.17\\%" + }, + { + "bbox": [ + 130, + 148, + 483, + 234 + ], + "type": "text", + "content": ". For the CIFAR-10 dataset, the DACC is " + }, + { + "bbox": [ + 130, + 148, + 483, + 234 + ], + "type": "inline_equation", + "content": ">76\\%" + }, + { + "bbox": [ + 130, + 148, + 483, + 234 + ], + "type": "text", + "content": " and the average DACC is " + }, + { + "bbox": [ + 130, + 148, + 483, + 234 + ], + "type": "inline_equation", + "content": "80.5\\%" + }, + { + "bbox": [ + 130, + 148, + 483, + 234 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 189, + 268, + 421, + 356 + ], + "blocks": [ + { + "bbox": [ + 189, + 247, + 424, + 258 + ], + "lines": [ + { + "bbox": [ + 189, + 247, + 424, + 258 + ], + "spans": [ + { + "bbox": [ + 189, + 247, + 424, + 258 + ], + "type": "text", + "content": "Table 2: The detection performance of our SSL-Cleanse." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 189, + 268, + 421, + 356 + ], + "lines": [ + { + "bbox": [ + 189, + 268, + 421, + 356 + ], + "spans": [ + { + "bbox": [ + 189, + 268, + 421, + 356 + ], + "type": "table", + "html": "
DatasetMethodSSL-BackdoorCTRL
TPFPDACC(%)TPFPDACC(%)
CIFAR-10BYOL3558036482
SimCLR3347939584
MoCo V23157637582
ImageNet-100BYOL3888043885
SimCLR3477746888
MoCo V23677942884
", + "image_path": "470a65721bc343b3289c157349c7f37ee3de805af1914feda20900a196c00ebc.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 374, + 483, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 374, + 483, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 483, + 518 + ], + "type": "text", + "content": "Table 2 further indicates that our SSL-Cleanse can detect not only the patch-based trigger SSL attack SSL-Backdoor but also the frequency-based SSL attack CTRL. This capability stems from our method's utilization of trigger size and magnitude for patch-based trigger detection in tandem with frequency-domain detection. In particular, For SSL-Backdoor detection, SSL-Clease achieves " + }, + { + "bbox": [ + 130, + 374, + 483, + 518 + ], + "type": "inline_equation", + "content": "78.5\\%" + }, + { + "bbox": [ + 130, + 374, + 483, + 518 + ], + "type": "text", + "content": " DACC on average for both CIFAR-10 and ImageNet-100 datasets. For CTRL, it obtains " + }, + { + "bbox": [ + 130, + 374, + 483, + 518 + ], + "type": "inline_equation", + "content": "84.17\\%" + }, + { + "bbox": [ + 130, + 374, + 483, + 518 + ], + "type": "text", + "content": " average DACC. Against the CTRL, our detector identifies 46 TP from 50 Trojaned SimCLR encoders on ImageNet and registers 8 FP among 50 clean SimCLR encoders. Our SSL-Cleanse consistently exhibits reliable detection performance across prevalent SSL training techniques such as BYOL, SimCLR, and MoCo V2. In particular, our detector registers an average DACC of " + }, + { + "bbox": [ + 130, + 374, + 483, + 518 + ], + "type": "inline_equation", + "content": "81.75\\%" + }, + { + "bbox": [ + 130, + 374, + 483, + 518 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 374, + 483, + 518 + ], + "type": "inline_equation", + "content": "82\\%" + }, + { + "bbox": [ + 130, + 374, + 483, + 518 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 374, + 483, + 518 + ], + "type": "inline_equation", + "content": "80.25\\%" + }, + { + "bbox": [ + 130, + 374, + 483, + 518 + ], + "type": "text", + "content": " for BYOL, SimCLR, and MoCo V2, respectively." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 136, + 563, + 476, + 663 + ], + "blocks": [ + { + "bbox": [ + 186, + 542, + 426, + 554 + ], + "lines": [ + { + "bbox": [ + 186, + 542, + 426, + 554 + ], + "spans": [ + { + "bbox": [ + 186, + 542, + 426, + 554 + ], + "type": "text", + "content": "Table 3: The mitigation performance of our SSL-Cleanse." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 563, + 476, + 663 + ], + "lines": [ + { + "bbox": [ + 136, + 563, + 476, + 663 + ], + "spans": [ + { + "bbox": [ + 136, + 563, + 476, + 663 + ], + "type": "table", + "html": "
DatasetMethodSSL-BackdoorCTRL
Before mitigationAfter mitigationBefore mitigationAfter mitigation
ACC(%)ASR(%)ACC(%)ASR(%)ACC(%)ASR(%)ACC(%)ASR(%)
CIFAR-10BYOL83.4248.3282.141.1483.1960.4782.591.96
SimCLR84.8842.1983.530.5880.7481.8479.601.15
MoCo V281.0237.9580.160.9281.4277.5180.031.62
ImageNet-100BYOL60.5733.2160.240.1453.3345.1052.650.35
SimCLR60.1831.8558.580.6252.9044.9851.040.33
MoCo V261.5735.0660.100.1750.6235.7248.880.17
", + "image_path": "f7e4bfe14df66679eb498d10206da5d9fb1caace76645670be9ae43f5d634372.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "text", + "content": "M. Zheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "content": "Mitigation. Table 3 compares the attack success rate (ASR) and clean accuracy (ACC) before and after applying the mitigator against SSL-Backdoor and CTRL on the CIFAR-10 and ImageNet-100 dataset. On average, the ASR experiences a marked reduction to below " + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "content": ", while the ACC declines to approximately " + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "content": " before and after implementing mitigation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 482, + 236 + ], + "type": "text", + "content": "In the case of patch-based SSL-Backdoor, our mitigation approach significantly reduces the ASR to below " + }, + { + "bbox": [ + 130, + 176, + 482, + 236 + ], + "type": "inline_equation", + "content": "1.2\\%" + }, + { + "bbox": [ + 130, + 176, + 482, + 236 + ], + "type": "text", + "content": ", demonstrating its successful eradication of backdoor effects. Additionally, the ACC experiences an average decrease of " + }, + { + "bbox": [ + 130, + 176, + 482, + 236 + ], + "type": "inline_equation", + "content": "1.15\\%" + }, + { + "bbox": [ + 130, + 176, + 482, + 236 + ], + "type": "text", + "content": " in backdoored models after applying our mitigator. When utilizing the BYOL training method on the ImageNet-100 dataset, ACC remains relatively stable." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 237, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 482, + 308 + ], + "type": "text", + "content": "Regarding global frequency-based SSL attack CTRL, our mitigation strategy substantially decreases the ASR to less than " + }, + { + "bbox": [ + 130, + 237, + 482, + 308 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 130, + 237, + 482, + 308 + ], + "type": "text", + "content": ", underscoring its effective elimination of backdoor attacks. Furthermore, the ACC has an average decline of " + }, + { + "bbox": [ + 130, + 237, + 482, + 308 + ], + "type": "inline_equation", + "content": "1.06\\%" + }, + { + "bbox": [ + 130, + 237, + 482, + 308 + ], + "type": "text", + "content": " in backdoored models after mitigation. Importantly, when employing the BYOL training technique on both CIFAR-10 and ImageNet-100 datasets, ACC remains fairly consistent." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 309, + 483, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 309, + 483, + 417 + ], + "spans": [ + { + "bbox": [ + 130, + 309, + 483, + 417 + ], + "type": "text", + "content": "Efficiency on Downstream Tasks. SSL-Cleanse demonstrates effective generalization across various downstream tasks. As indicated in Table 4, SSL-Cleanse reliably brings the ASR below " + }, + { + "bbox": [ + 130, + 309, + 483, + 417 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 130, + 309, + 483, + 417 + ], + "type": "text", + "content": " for different downstream tasks such as CIFAR-10, STL-10, and GTSRB datasets, when the SSL encoders were backdoored on CIFAR-10 dataset. In addition, SSL-Cleanse works well across all targeted task labels. For CIFAR-10, it is possible to introduce a backdoor into any label from 0 to 9. SSL-Cleanse has been tested and proven to identify attacks on each label with a detection accuracy (DACC) exceeding " + }, + { + "bbox": [ + 130, + 309, + 483, + 417 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 130, + 309, + 483, + 417 + ], + "type": "text", + "content": ", and the average DACC across 10 labels is " + }, + { + "bbox": [ + 130, + 309, + 483, + 417 + ], + "type": "inline_equation", + "content": "\\sim 80\\%" + }, + { + "bbox": [ + 130, + 309, + 483, + 417 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 202, + 448, + 406, + 507 + ], + "blocks": [ + { + "bbox": [ + 169, + 427, + 443, + 439 + ], + "lines": [ + { + "bbox": [ + 169, + 427, + 443, + 439 + ], + "spans": [ + { + "bbox": [ + 169, + 427, + 443, + 439 + ], + "type": "text", + "content": "Table 4: SSL-Cleanse performance on different downstream tasks." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 202, + 448, + 406, + 507 + ], + "lines": [ + { + "bbox": [ + 202, + 448, + 406, + 507 + ], + "spans": [ + { + "bbox": [ + 202, + 448, + 406, + 507 + ], + "type": "table", + "html": "
Downstream TaskBefore MitigationAfter Mitigation
ACC(%)ASR(%)ACC(%)ASR(%)
CIFAR-1083.4248.3282.141.14
STL-1077.5336.5575.041.84
GTSRB72.0138.2870.981.03
", + "image_path": "17e9b77778883f0308cfe3f1ad1e5c07ecaa8acf8a65c463db2a9c4a46b802af.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 521, + 299, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 299, + 652 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 299, + 652 + ], + "type": "text", + "content": "The SWK Effects. Figure 5 presents an ablation study on our SWK approach. Employing SWK on ImageNet100 yields a predicted number of 100, in contrast to 20 as obtained using the direct search (Kneedle) method, as depicted in Figure 3. Consequently, SWK delivers a detection accuracy of " + }, + { + "bbox": [ + 130, + 521, + 299, + 652 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 130, + 521, + 299, + 652 + ], + "type": "text", + "content": ", marking a " + }, + { + "bbox": [ + 130, + 521, + 299, + 652 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 130, + 521, + 299, + 652 + ], + "type": "text", + "content": " DACC improvement. Data Ratio Effects. We analyze the influence of different dataset ratios on" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 304, + 524, + 481, + 594 + ], + "blocks": [ + { + "bbox": [ + 304, + 524, + 481, + 594 + ], + "lines": [ + { + "bbox": [ + 304, + 524, + 481, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 524, + 481, + 594 + ], + "type": "image", + "image_path": "f2d0feaf62d15828b1992283a267a6363761e64dcf70c0afb242baaf87cc88d4.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 603, + 482, + 647 + ], + "lines": [ + { + "bbox": [ + 304, + 603, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 603, + 482, + 647 + ], + "type": "text", + "content": "Fig.5: A comparison of detection accuracy between SSL-Cleanse using the SWK method and the direct search (Kneedle)on ImageNet-100." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 653, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 653, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 653, + 481, + 666 + ], + "type": "text", + "content": "detection performance using CTRL attack and BYOL training methods. We" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 143, + 90, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 90, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 143, + 90, + 448, + 103 + ], + "type": "text", + "content": "SSL-Cleanse: Trojan Detection and Mitigation in Self-Supervised Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "employ varied image proportions, namely, " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "8\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": ". As illustrated in Table 5, the results showcase an improvement in detection performance as the proportion of the training dataset increases. The DACC for the CIFAR-10 dataset shows a " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " increase between the " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "8\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " training datasets. Moreover, for the ImageNet-100 dataset, there is an enhancement from " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "83\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "85\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " in the DACC. Further details regarding the hyperparameters " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": ", along with nuances about the influence of attacking trigger size and perturbation norm, can be found in our supplementary material." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 200, + 255, + 410, + 315 + ], + "blocks": [ + { + "bbox": [ + 130, + 220, + 482, + 244 + ], + "lines": [ + { + "bbox": [ + 130, + 220, + 482, + 244 + ], + "spans": [ + { + "bbox": [ + 130, + 220, + 482, + 244 + ], + "type": "text", + "content": "Table 5: Influence of the ratio of the unlabeled dataset to the entire training dataset on the detection efficacy. A larger ratio usually introduces a higher DACC." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 200, + 255, + 410, + 315 + ], + "lines": [ + { + "bbox": [ + 200, + 255, + 410, + 315 + ], + "spans": [ + { + "bbox": [ + 200, + 255, + 410, + 315 + ], + "type": "table", + "html": "
Data ratio (%)CIFAR-10ImageNet-100
TPFPDACC(%)TPFPDACC(%)
52877138682
83787940783
103888043885
", + "image_path": "be85a190817ddf566fc9c81debd4b3cb8a19289526c22fc2bb925291f6bbee81.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 334, + 483, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 334, + 483, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 334, + 483, + 418 + ], + "type": "text", + "content": "Outlier Detection Effects. Figure 6 shows our Size-Norm Trigger Outlier Detector can successfully distinguish the outlier triggers from the " + }, + { + "bbox": [ + 130, + 334, + 483, + 418 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 334, + 483, + 418 + ], + "type": "text", + "content": "-cluster trigger list for the trojaned encoders based on SSL-Backdoor and CTRL. In particular, each box bar plots the trigger size " + }, + { + "bbox": [ + 130, + 334, + 483, + 418 + ], + "type": "inline_equation", + "content": "|m_i^1|" + }, + { + "bbox": [ + 130, + 334, + 483, + 418 + ], + "type": "text", + "content": " and trigger norm " + }, + { + "bbox": [ + 130, + 334, + 483, + 418 + ], + "type": "inline_equation", + "content": "|m_i^2 \\cdot \\Delta_i^2|" + }, + { + "bbox": [ + 130, + 334, + 483, + 418 + ], + "type": "text", + "content": " for SSL-Backdoor and CTRL, respectively. Our Size-Norm trigger outlier detection can scan collectively, without the presumption that the defender has prior knowledge of the attack type." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 184, + 423, + 428, + 520 + ], + "blocks": [ + { + "bbox": [ + 184, + 423, + 428, + 520 + ], + "lines": [ + { + "bbox": [ + 184, + 423, + 428, + 520 + ], + "spans": [ + { + "bbox": [ + 184, + 423, + 428, + 520 + ], + "type": "image", + "image_path": "975c6f7e5ba5469c14adc53a5675f7207604d0a463a3b9d47bb6018519772511.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 525, + 482, + 560 + ], + "lines": [ + { + "bbox": [ + 130, + 525, + 482, + 560 + ], + "spans": [ + { + "bbox": [ + 130, + 525, + 482, + 560 + ], + "type": "text", + "content": "Fig. 6: The Size-Norm trigger outlier detection criteria is able to identify both patch-based SSL-Backdoor and frequency-domain global trigger in CTRL. The box plot shows min/max and quartiles." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 583, + 220, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 583, + 220, + 596 + ], + "spans": [ + { + "bbox": [ + 132, + 583, + 220, + 596 + ], + "type": "text", + "content": "8 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "This paper introduces SSL-Cleanse, a novel work to detect and mitigate Trojan attacks in SSL encoders without accessing any downstream labels. We evaluated SSL-Cleanse on various datasets using 1200 models, achieving an average detection success rate of " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "82.2\\%" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " on ImageNet-100. After mitigating backdoors, backdoored encoders achieve " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "0.3\\%" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " average attack success rate without great accuracy loss." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "text", + "content": "M. Zheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 137, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 137, + 137, + 482, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 137, + 482, + 169 + ], + "spans": [ + { + "bbox": [ + 137, + 137, + 482, + 169 + ], + "type": "text", + "content": "1. Al Ghanim, M., Santriaji, M., Lou, Q., Solihin, Y.: Trojbits: A hardware aware inference-time attack on transformer-based language models. In: ECAI 2023, pp. 60-68. IOS Press (2023) 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 137, + 170, + 481, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 170, + 481, + 201 + ], + "spans": [ + { + "bbox": [ + 137, + 170, + 481, + 201 + ], + "type": "text", + "content": "2. Campello, R.J., Moulavi, D., Zimek, A., Sander, J.: Hierarchical density estimates for data clustering, visualization, and outlier detection. ACM Transactions on Knowledge Discovery from Data (TKDD) 10(1), 1-51 (2015) 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 202, + 481, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 202, + 481, + 234 + ], + "spans": [ + { + "bbox": [ + 137, + 202, + 481, + 234 + ], + "type": "text", + "content": "3. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International conference on machine learning. pp. 1597-1607. PMLR (2020) 1, 3, 11" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 234, + 481, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 234, + 481, + 255 + ], + "spans": [ + { + "bbox": [ + 138, + 234, + 481, + 255 + ], + "type": "text", + "content": "4. Chen, X., Fan, H., Girshick, R., He, K.: Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297 (2020) 1, 2, 3, 11" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 255, + 481, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 255, + 481, + 287 + ], + "spans": [ + { + "bbox": [ + 138, + 255, + 481, + 287 + ], + "type": "text", + "content": "5. Chen, X., He, K.: Exploring simple siamese representation learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15750-15758 (2021) 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 288, + 481, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 288, + 481, + 330 + ], + "spans": [ + { + "bbox": [ + 138, + 288, + 481, + 330 + ], + "type": "text", + "content": "6. Coates, A., Ng, A., Lee, H.: An analysis of single-layer networks in unsupervised feature learning. In: Proceedings of the fourteenth international conference on artificial intelligence and statistics. pp. 215-223. JMLR Workshop and Conference Proceedings (2011) 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 331, + 481, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 331, + 481, + 384 + ], + "spans": [ + { + "bbox": [ + 138, + 331, + 481, + 384 + ], + "type": "text", + "content": "7. Deng, J., Dong, W., Socher, R., Li, L., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2009), 20-25 June 2009, Miami, Florida, USA. pp. 248-255. IEEE Computer Society (2009). https://doi.org/10.1109/CVPR.2009.5206848, https://doi.org/10.1109/CVPR.2009.5206848 11" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 385, + 481, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 385, + 481, + 427 + ], + "spans": [ + { + "bbox": [ + 138, + 385, + 481, + 427 + ], + "type": "text", + "content": "8. Feng, S., Tao, G., Cheng, S., Shen, G., Xu, X., Liu, Y., Zhang, K., Ma, S., Zhang, X.: Detecting backdoors in pre-trained encoders. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16352-16362 (2023) 2, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 428, + 481, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 428, + 481, + 458 + ], + "spans": [ + { + "bbox": [ + 138, + 428, + 481, + 458 + ], + "type": "text", + "content": "9. Ghassabeh, Y.A.: A sufficient condition for the convergence of the mean shift algorithm with gaussian kernel. Journal of Multivariate Analysis 135, 1-10 (2015) 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 134, + 460, + 481, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 460, + 481, + 503 + ], + "spans": [ + { + "bbox": [ + 134, + 460, + 481, + 503 + ], + "type": "text", + "content": "10. Grill, J.B., Strub, F., Altché, F., Tallec, C., Richemond, P., Buchatskaya, E., Doersch, C., Avila Pires, B., Guo, Z., Gheshlaghi Azar, M., et al.: Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems 33, 21271-21284 (2020) 2, 11" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 503, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 503, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 134, + 503, + 481, + 536 + ], + "type": "text", + "content": "11. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729-9738 (2020) 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 536, + 481, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 536, + 481, + 590 + ], + "spans": [ + { + "bbox": [ + 134, + 536, + 481, + 590 + ], + "type": "text", + "content": "12. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016. pp. 770-778. IEEE Computer Society (2016). https://doi.org/10.1109/CVPR.2016.90, https://doi.org/10.1109/CVPR.2016.90 11" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 590, + 481, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 590, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 134, + 590, + 481, + 611 + ], + "type": "text", + "content": "13. Jaiswal, A., Babu, A.R., Zadeh, M.Z., Banerjee, D., Makedon, F.: A survey on contrastive self-supervised learning. Technologies 9(1), 2 (2020) 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 611, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 611, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 611, + 481, + 632 + ], + "type": "text", + "content": "14. Jia, J., Liu, Y., Gong, N.Z.: Badencoder: Backdoor attacks to pre-trained encoders in self-supervised learning. arXiv preprint arXiv:2108.00352 (2021) 2, 3, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 632, + 481, + 665 + ], + "type": "text", + "content": "15. Kolouri, S., Saha, A., Pirsiavash, H., Hoffmann, H.: Universal litmus patterns: Revealing backdoor attacks in cnns. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 301-310 (2020) 2" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 143, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 143, + 91, + 447, + 102 + ], + "type": "text", + "content": "SSL-Cleanse: Trojan Detection and Mitigation in Self-Supervised Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "16. Krishnan, R., Rajpurkar, P., Topol, E.J.: Self-supervised learning in medicine and healthcare. Nature Biomedical Engineering pp. 1-7 (2022) 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 481, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 481, + 161 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 481, + 161 + ], + "type": "text", + "content": "17. Krizhevsky, A.: Learning multiple layers of features from tiny images. Tech. rep., University of Toronto (2009) 4, 11" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 134, + 163, + 480, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 163, + 480, + 195 + ], + "spans": [ + { + "bbox": [ + 134, + 163, + 480, + 195 + ], + "type": "text", + "content": "18. Lan, Z., Chen, M., Goodman, S., Gimpel, K., Sharma, P., Soricut, R.: Albert: A lite bert for self-supervised learning of language representations. arXiv preprint arXiv:1909.11942 (2019) 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 197, + 480, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 197, + 480, + 217 + ], + "spans": [ + { + "bbox": [ + 134, + 197, + 480, + 217 + ], + "type": "text", + "content": "19. Li, C., Pang, R., Xi, Z., Du, T., Ji, S., Yao, Y., Wang, T.: Demystifying self-supervised trojan attacks. arXiv preprint arXiv:2210.07346 (2022) 2, 4, 6, 9, 11" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 219, + 480, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 219, + 480, + 240 + ], + "spans": [ + { + "bbox": [ + 133, + 219, + 480, + 240 + ], + "type": "text", + "content": "20. Liu, H., Jia, J., Gong, N.Z.: Poisonedencoder: Poisoning the unlabeled pre-training data in contrastive learning. arXiv preprint arXiv:2205.06401 (2022) 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 241, + 481, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 241, + 481, + 285 + ], + "spans": [ + { + "bbox": [ + 132, + 241, + 481, + 285 + ], + "type": "text", + "content": "21. Liu, Y., Lee, W.C., Tao, G., Ma, S., Aafer, Y., Zhang, X.: Abs: Scanning neural networks for back-doors by artificial brain stimulation. In: Proceedings of the 2019 ACM SIGSAC Conference on Computer and Communications Security. pp. 1265-1282 (2019) 2, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 285, + 480, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 480, + 317 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 480, + 317 + ], + "type": "text", + "content": "22. Liu, Y., Jin, M., Pan, S., Zhou, C., Zheng, Y., Xia, F., Yu, P.: Graph self-supervised learning: A survey. IEEE Transactions on Knowledge and Data Engineering (2022) 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 319, + 481, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 319, + 481, + 341 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 481, + 341 + ], + "type": "text", + "content": "23. Lou, Q., Liu, Y., Feng, B.: Trojtext: Test-time invisible textual Trojan insertion. In: The Eleventh International Conference on Learning Representations (2023) 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 342, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 480, + 374 + ], + "type": "text", + "content": "24. Pan, M., Zeng, Y., Lyu, L., Lin, X., Jia, R.: Asset: Robust backdoor data detection across a multiplicity of deep learning paradigms. arXiv preprint arXiv:2302.11408 (2023) 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 376, + 480, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 376, + 480, + 408 + ], + "spans": [ + { + "bbox": [ + 132, + 376, + 480, + 408 + ], + "type": "text", + "content": "25. Saha, A., Tejankar, A., Koohpayegani, S.A., Pirsiavash, H.: Backdoor attacks on self-supervised learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13337-13346 (2022) 2, 4, 6, 9, 11" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 409, + 481, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 409, + 481, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 409, + 481, + 475 + ], + "type": "text", + "content": "26. Satopaa, V., Albrecht, J.R., Irwin, D.E., Raghavan, B.: Finding a \"kneedle\" in a haystack: Detecting knee points in system behavior. In: 31st IEEE International Conference on Distributed Computing Systems Workshops (ICDCS 2011 Workshops), 20-24 June 2011, Minneapolis, Minnesota, USA. pp. 166-171. IEEE Computer Society (2011). https://doi.org/10.1109/ICDCSW.2011.20, https://doi.org/10.1109/ICDCSW.2011.205, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 475, + 480, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 480, + 497 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 480, + 497 + ], + "type": "text", + "content": "27. Sinaga, K.P., Yang, M.S.: Unsupervised k-means clustering algorithm. IEEE access 8, 80716-80727 (2020) 5, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 498, + 480, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 480, + 531 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 480, + 531 + ], + "type": "text", + "content": "28. Stallkamp, J., Schlipsing, M., Salmen, J., Igel, C.: The german traffic sign recognition benchmark: a multi-class classification competition. In: The 2011 international joint conference on neural networks. pp. 1453-1460. IEEE (2011) 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 532, + 481, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 532, + 481, + 576 + ], + "spans": [ + { + "bbox": [ + 132, + 532, + 481, + 576 + ], + "type": "text", + "content": "29. Tejankar, A., Sanjabi, M., Wang, Q., Wang, S., Firooz, H., Pirsiavash, H., Tan, L.: Defending against patch-based backdoor attacks on self-supervised learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12239-12249 (2023) 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 577, + 481, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 577, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 577, + 481, + 620 + ], + "type": "text", + "content": "30. Wang, B., Yao, Y., Shan, S., Li, H., Viswanath, B., Zheng, H., Zhao, B.Y.: Neural cleansse: Identifying and mitigating backdoor attacks in neural networks. In: Proceedings of the IEEE Symposium on Security and Privacy (IEEE S&P). San Francisco, CA (2019) 2, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "text", + "content": "31. Wu, J., Zhang, T., Zha, Z.J., Luo, J., Zhang, Y., Wu, F.: Self-supervised domain-aware generative network for generalized zero-shot learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12767-12776 (2020) 2" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "text", + "content": "M. Zheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 293 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 148 + ], + "type": "text", + "content": "32. Xue, J., Lou, Q.: Estas: Effective and stable trojan attacks in self-supervised encoders with one target unlabelled sample. arXiv preprint arXiv:2211.10908 (2022) 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 149, + 482, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 482, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 482, + 171 + ], + "type": "text", + "content": "33. Yaman, B., Hosseini, S.A.H., Akçakaya, M.: Zero-shot self-supervised learning for mri reconstruction. arXiv preprint arXiv:2102.07737 (2021) 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 482, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 482, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 482, + 215 + ], + "type": "text", + "content": "34. Yang, H., Yang, R., Cai, H., Zhang, X., Pei, Q., Wang, S., Yan, H.: Ssl-abd: An adversarial defense method against backdoor attacks in self-supervised learning. In: International Conference on Artificial Intelligence Security and Privacy. pp. 456-467. Springer (2023) 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 482, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 482, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 482, + 237 + ], + "type": "text", + "content": "35. Zhang, J., Liu, H., Jia, J., Gong, N.Z.: Corruptencoder: Data poisoning based backdoor attacks to contrastive learning. arXiv preprint arXiv:2211.08229 (2022) 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 238, + 482, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 482, + 270 + ], + "type": "text", + "content": "36. Zheng, M., Lou, Q., Jiang, L.: Trojvit: Trojan insertion in vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4025-4034 (2023) 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 271, + 482, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 482, + 293 + ], + "type": "text", + "content": "37. Zheng, M., Xue, J., Sheng, Y., Yang, L., Lou, Q., Jiang, L.: Trojfair: Trojan fairness attacks. arXiv preprint arXiv:2312.10508 (2023) 2" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 143, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 143, + 91, + 447, + 102 + ], + "type": "text", + "content": "SSL-Cleanse: Trojan Detection and Mitigation in Self-Supervised Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/ST-LDM_ A Universal Framework for Text-Grounded Object Generation in Real Images/ea78a6df-2485-4a42-a86e-90a6085c52c4_content_list.json b/2024/ST-LDM_ A Universal Framework for Text-Grounded Object Generation in Real Images/ea78a6df-2485-4a42-a86e-90a6085c52c4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b541338d0a715ceb3b4ebb6c2a50fd34d8a4d61e --- /dev/null +++ b/2024/ST-LDM_ A Universal Framework for Text-Grounded Object Generation in Real Images/ea78a6df-2485-4a42-a86e-90a6085c52c4_content_list.json @@ -0,0 +1,1992 @@ +[ + { + "type": "text", + "text": "ST-LDM: A Universal Framework for Text-Grounded Object Generation in Real Images", + "text_level": 1, + "bbox": [ + 217, + 140, + 787, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiangtian Xue $^{1,2,4}$ , Jiasong Wu $^{1,2,4\\star}$ , Youyong Kong $^{1,2,4}$ , Lotfi Senhadji $^{2,3,4}$ , and Huazhong Shu $^{1,2,4}$", + "bbox": [ + 225, + 210, + 776, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Key Laboratory of New Generation Artificial Intelligence Technology and Its Interdisciplinary Applications (Southeast University), Nanjing 210096, China", + "bbox": [ + 238, + 253, + 763, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Jiangsu Provincial Joint International Research Laboratory of Medical Information Processing, Southeast University, Nanjing 210096, China", + "bbox": [ + 217, + 282, + 782, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Univ-Rennes, INSERM, LTSI-UMR 1099, Rennes F-3502, France", + "bbox": [ + 276, + 309, + 725, + 323 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 Centre de Recherche en Information Biomédicale Sino-français (CRIBs), Rennes F-3502, France", + "bbox": [ + 228, + 324, + 774, + 351 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. We present a novel image editing scenario termed Text-grounded", + "bbox": [ + 261, + 377, + 764, + 390 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Object Generation (TOG), defined as generating a new object in the real image spatially conditioned by textual descriptions. Existing diffusion models exhibit limitations of spatial perception in complex real-world scenes, relying on additional modalities to enforce constraints, and TOG imposes heightened challenges on scene comprehension under the weak supervision of linguistic information. We propose a universal framework ST-LDM based on Swin-Transformer, which can be integrated into any latent diffusion model with training-free backward guidance. ST-LDM encompasses a global-perceptual autoencoder with adaptable compression scales and hierarchical visual features, parallel with deformable multimodal transformer to generate region-wise guidance for the subsequent denoising process. We transcend the limitation of traditional attention mechanisms that only focus on existing visual features by introducing deformable feature alignment to hierarchically refine spatial positioning fused with multi-scale visual and linguistic information. Extensive Experiments demonstrate that our model enhances the localization of attention mechanisms while preserving the generative capabilities inherent to diffusion models.", + "bbox": [ + 261, + 390, + 738, + 638 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Instruction-based image editing. Text-grounded object generation. Deformable feature alignment. Training-free guidance", + "bbox": [ + 259, + 652, + 740, + 680 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 700, + 375, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Controllable text2img editing has demonstrated remarkable performance with ongoing exploration of diffusion models [10,34,48], and recent studies have been predominantly directed towards enhancing the controllability of language in editing the appearance of objects [15,19]. As for generating new objects in real images, additional modalities such as bounding boxes and keypoints have to be introduced to explicitly regulate their spatial placement [24].", + "bbox": [ + 212, + 727, + 787, + 819 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author: jswu@seu.edu.cn", + "bbox": [ + 218, + 825, + 500, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1149bd161f7dd216ce202a159e0947ca6a9546892f74212b5b7ed75acdc3a1fc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 143, + 310, + 215 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/017cbcec5e4a50e6f38375ab4f04b341194add502b0b99631de7fe4d8637102d.jpg", + "image_caption": [ + "wooden bench" + ], + "image_footnote": [], + "bbox": [ + 312, + 143, + 403, + 215 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d4d218a80485e5f80497ed8ebaf2018a1c35247fba7a07595fc4f669fcf27349.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 405, + 143, + 496, + 215 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1d2ebbf36d6c497a8c0ccf264adf4dbf3af28e4bac6bd07809545592990a0f51.jpg", + "image_caption": [ + "[bottom left corner]", + "fanny pack, black and white empty area on the table]", + "Fig. 1: Text-grounded object generation in real images. Our method can generate the rational object within real-world input images based on textual descriptions encompassing its visual characteristic and spatial attribute. The generated object seamlessly integrates with the original image, exhibiting a visually harmonious presence while preserving the contextual semantics of the original scene (e.g. consistency in shadow (top left), geometry (top right), illumination (bottom left), occlusion (bottom right))." + ], + "image_footnote": [], + "bbox": [ + 312, + 227, + 403, + 297 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a2bd42d721a52ecee83e9b7e5d9a9ede21a11e7b68f5613e71743bf2dbf02072.jpg", + "image_caption": [ + "old newspaper [on the \"MAP\" magazine]" + ], + "image_footnote": [], + "bbox": [ + 406, + 227, + 496, + 297 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/765641b48cbdc2a32f7b63d5459d384f0eb5fca5df663d14ffec3a819ad951ca.jpg", + "image_caption": [ + "" + ], + "image_footnote": [], + "bbox": [ + 501, + 143, + 594, + 215 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3cc44ec75256194edb1c517b28a576b2f018b2ab4d6d5b3c638b715ebff7d315.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 222, + 594, + 295 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/591e76cccaf19e5c83ef7a7113383cb4c304bf477429d3efbcb50b27f68f4422.jpg", + "image_caption": [ + "copper vase, with pattern" + ], + "image_footnote": [], + "bbox": [ + 596, + 143, + 687, + 215 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/de8724c805cbbd62287dfd631ee26ca23a58b9f51715bd4b7f13332fdb0409c3.jpg", + "image_caption": [ + "blue sedan [between bus and green car]" + ], + "image_footnote": [], + "bbox": [ + 596, + 222, + 687, + 295 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/cc7d9315efc524984b7d034b6975e52fe219d3c6b12ca0db8954d37fb68db29e.jpg", + "image_caption": [ + "woman with back turned" + ], + "image_footnote": [], + "bbox": [ + 689, + 143, + 782, + 215 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9ed3a5d1d8e057d6160b072ecec2e296196e8ff8b20259764217525c99aa34fe.jpg", + "image_caption": [ + "board, white letters on blue [on the roof of the bus]" + ], + "image_footnote": [], + "bbox": [ + 691, + 222, + 779, + 295 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The aforementioned modalities provide precise spatial constraints, but it is noteworthy that language-based instruction serves as the most fundamental and convenient interactive modality in real-world scenarios (e.g. chatting with a robot commonly using language as the sole mode of interaction, conducting similar manipulations on extensive image datasets and dynamic video sequences). Moreover, language affords a considerable degree of creative flexibility while adhering to the principles of spatial logic. This enhances the naturalness of edited images, as the model can estimate the most suitable generation location by comprehending multi-modal information.", + "bbox": [ + 212, + 417, + 784, + 551 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose a new image editing scenario named Text-grounded Object Generation (TOG), which aims to generate objects in real scenarios according to natural language expressions (Fig. 1). Concretely, given an image in conjunction with textual descriptions detailing visual and spatial attributes of the generated object, the task entails generating a rational object in complex scenarios. TOG embodies the dual characteristics of text2img generation and editing tasks. The generated object must adhere to textual descriptions, while simultaneously ensuring that the edited image maintains high fidelity with irrelevant areas remaining consistent with original semantics.", + "bbox": [ + 212, + 553, + 787, + 688 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although diffusion models have demonstrated notable creative capabilities in image generation, they fall short in spatial perception and orientation, particularly when dealing with intricate real-world scenes [41]. The prevalent approaches to address this limitation involve introducing additional conditions during the denoising process to control the placement of the generated object. However, a fundamental paradox arises in our specific scenario: how to locate an object which is absent from the original image. This emphasizes the primary challenge, that is determining the optimal area for object generation under the weak supervision of texts. The objective is to ensure that the generated object aligns semantically with the contextual information contained within the image.", + "bbox": [ + 212, + 689, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "X. Xue et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we present ST-LDM, a universal framework with hierarchical Swin-Transformer (ST) blocks [28], amenable for integration into latent diffusion models (LDM) via training-free backward guidance. ST-LDM comprises two parallel branches with similar structures, dedicated respectively to adaptable latent representation and spatial attention guidance. For latent representation, the Transformer-based autoencoder facilitates multi-resolution compression through a low-cost trainable module while preserving the formidable modeling capacity inherent to the pre-trained Swin-T. For spatial guidance, we propose deformable feature alignment in multimodal fusion Transformer to enhance the refinement of grounded regions by incorporating information from multi-scale visual and linguistic features. ST-LDM provides visually coherent location guidance in pixel space which can be integrated to the denoising U-Net of LDMs with region-wise backpropagation.", + "bbox": [ + 212, + 146, + 787, + 343 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To quantify the effectiveness of our proposed approach, we compare it with text-prompted image editing models with diverse conditions based on our constructed benchmark. Experiments demonstrate that our model enhances the localization capability of Swin-Transformer while preserving the generative proficiency inherent in diffusion models. The main contributions of our work are highlighted as follows:", + "bbox": [ + 212, + 343, + 787, + 434 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present a novel image editing scenario termed Text-grounded Object Generation (TOG), which is defined as generating objects in real images grounded by textual descriptions.", + "- We further propose a universal framework ST-LDM with deformable feature alignment to hierarchically refine spatial guidance, which can be integrated into diffusion models with training-free region-wise backpropagation.", + "- Quantitative comparisons with related works on our constructed benchmark and qualitative visualizations substantiate the robust editing and generative capabilities according to textual conditions." + ], + "bbox": [ + 225, + 439, + 784, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 593, + 387, + 608 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Spatially Controllable Object Generation", + "text_level": 1, + "bbox": [ + 215, + 621, + 599, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent studies have employed mandatory modalities to exert control over the spatial distribution of generated objects [6,12,43]. Make-A-Scene [13] introduces implicit conditioning over optionally controlled scene tokens derived from segmentation maps with fixed labels to control the layout of the image. SpaText [4] further proposes a spatial-textual representation that encapsulates the semantic attributes and structural characteristics of each segment. GLIGEN [24] extends grounding conditions to bounding boxes, reference images, keypoints, etc., which are injected into new trainable layers of the denoising process via gated self-attention. Nevertheless, these modalities are cumbersome when applied to large-scale operations involving numerous images and videos. The utilization of text-based spatial position descriptions serves to mitigate rigid constraints, thereby affording models a heightened capacity for comprehending visual information.", + "bbox": [ + 212, + 643, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "ST-LDM", + "bbox": [ + 668, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 Diffusion-Based Semantic Image Editing", + "text_level": 1, + "bbox": [ + 215, + 146, + 593, + 162 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Image editing emphasizes modifying the visual entity consistent with textual descriptions while preserving the entity-irrelevant regions. Diffusion models can be easily adapted for inpainting when given a mask input since they iteratively refine an image starting from random noise [30,31]. Recent approaches endeavor to obviate the necessity of additional inputs such as image masks or additional views of the object [6,16,37]. DiffEdit [7] automatically generates a mask accentuating regions of the input image necessitating modification, by contrasting predictions of a diffusion model conditioned on different text prompts. Imagic [19] produces a text embedding that aligns with both the input image and the target text while fine-tuning the diffusion model to capture the image-specific appearance. These attempts are experimented on images with obvious foreground objects and simple semantics. However, when applied to complex scenes characterized by a multiplicity of objects and spatial relationships, their efficacy in localizing objects exhibits a discernible diminishment.", + "bbox": [ + 212, + 169, + 787, + 381 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Vision-and-Language Transformer", + "text_level": 1, + "bbox": [ + 215, + 402, + 540, + 417 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The initial phase of our proposed method involves locating visual regions of interest guided by language, and Transformer [36] has demonstrated noteworthy performance at the interaction of vision and language. Ding et al. [9] introduce Transformer and multi-head attention to construct an encoder-decoder attention mechanism architecture. LAVT [42] achieves better cross-modal alignments through the early fusion of linguistic and visual features in a vision Transformer encoder network. SLViT [32] employs the variance of cross-modal correlations between adjacent stages to identify regions of high uncertainty and refine features of these regions with complementing information from multiple stages. While Transformer facilitates the integration of linguistic and visual modalities through cross-attention mechanisms, it encounters challenges stemming from the absence of target objects within the image, which is a distinctive characteristic of the TOG task.", + "bbox": [ + 212, + 425, + 787, + 621 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 643, + 330, + 659 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we present our proposed framework, denoted as ST-LDM, and the overall pipeline is depicted in Fig. 2.", + "bbox": [ + 212, + 672, + 785, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In pixel space (Sec. 3.1), visual feature extraction and multi-modal localization are concurrently performed during the downsampling process, leveraging Swin Transformer [28] as the backbone. The feature extraction module acts as an autoencoder to concurrently reconstruct the image into a lower-dimensional latent space while furnishing hierarchical visual features for spatial localization. In parallel, the multi-modal Transformer module leverages the interplay between image and language to provide positional guidance for the denoising process. We innovatively introduce deformable feature alignment to dynamically adjust spatial constraints, ensuring that the generated object aligns with semantic context", + "bbox": [ + 212, + 704, + 787, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "X. Xue et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ae8d2f13f143c733564f7063e8b6254fea02db3001d623d519ee2d066f500634.jpg", + "image_caption": [ + "Fig. 2: The basic framework of the proposed ST-LDM. Given a real image and text descriptions, we initially encode them into visual features $V_{i}$ and text embedding $L$ , where $L$ comprises appearance prompt $L_{a}$ and spatial conditions $L_{s}$ . We employ the Swin-T structure to hierarchically integrate $L$ into $V_{i}$ , where the multimodal map $M_{i}$ is deformed by learning multi-scale visual and semantic relationships. The generated spatial guidance attention map $G$ and textual prompt $L_{a}$ are integrated into latent diffusion models with training-free guidance. The autoencoder-decoder in pixel space and the latent diffusion model in latent space retain original structure and parameters, and these components are frozen throughout the training process." + ], + "image_footnote": [], + "bbox": [ + 220, + 145, + 782, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "of the input image. The congruence in interaction between two branches is optimized through the application of a similar network structure (SwinT) to both of the aforementioned modules.", + "bbox": [ + 212, + 532, + 784, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In latent space (Sec. 3.2), appearance-based text prompt is separately mapped into denoising U-Net via cross attention. Subsequently, a region-wise backpropagation scheme is introduced to iteratively update the activations in the network to match the desired layout specified in pixel space. The fundamental structure and parameters of diffusion models remain constant, with trainable parameter layers exclusively concentrated in the pixel space. Consequently, the proposed ST-LDM emerges as a universal framework adaptable to a spectrum of pretrained LDMs.", + "bbox": [ + 212, + 578, + 787, + 699 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Hierarchical Feature Extraction", + "text_level": 1, + "bbox": [ + 214, + 722, + 522, + 736 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Autoencoder with Swin-Transformer. LDMs [33] leverages an individually pretrained autoencoding model for perceptual image compression, but its quantitative performance exhibits significant variability across different downsampling factors due to the inherent instability of its encoding structure. Recently, Swin-Transformer has emerged as a generic visual feature extractor, owing to its robust modeling capabilities [28]. It initially splits an input RGB image into", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "ST-LDM", + "bbox": [ + 668, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "non-overlapping patches with a $4 \\times 4$ size, and the resolution of subsequent layers' output continually decreases accordingly. However, excessive information compression will induce distortion of noisy images in the denoising process.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To address this issue, we add a trainable patch expanding layer to convert the selected feature map to the specified size. The original structure of SwinTransformer is retained for hierarchical feature extraction, and an intermediate feature map $V_{i}$ is selected and upsampled to high-dimensional resolution via patch expanding layer. It can increase the resolution by a factor of 4 while simultaneously reducing the dimension to a configurable value $(H_{i}\\times W_{i}\\times C_{i}\\to 4H_{i}\\times 4W_{i}\\times C_{t},C_{i}\\geq 16C_{t})$ . The layer consists of two operations, a linear layer to convert the feature dimension $(H_{i}\\times W_{i}\\times C_{i}\\to H_{i}\\times W_{i}\\times 16C_{t})$ , and a rearrange operation to expand the resolution to $4\\times$ and reduce the dimension to one-sixteenth $(H_{i}\\times W_{i}\\times 16C_{t}\\to 4H_{i}\\times 4W_{i}\\times C_{t})$ .", + "bbox": [ + 212, + 191, + 784, + 342 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The extracted feature map is optional within $\\{V_2, V_3, V_4\\}$ contingent upon the depth of compression perception. As the initial patch size in Swin-Transformer is $4 \\times 4$ , the resulting output resolution for patch expansion remains fixed at $4 \\times$ relative to its initial dimensions. Consequently, $\\{V_2, V_3, V_4\\}$ corresponds to down-sampling factors $f = \\{2, 4, 8\\}$ . Due to the varied channel configurations across model variants, the output channel dimension remains adaptable. The encoding-decoding process in pixel space exhibits a symmetrical structure, wherein the linear layer in patch expanding and patch partition undergoes low-cost training following [33] to cater to specific downstream tasks while retaining the formidable modeling capabilities of pretrained models.", + "bbox": [ + 212, + 342, + 785, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Multimodal Transformer. We jointly embed language and vision features during visual encoding with Swin-Transformer to generate semantically coherent positional guidance for object generation. Consistent with previous works [42], we densely integrate spatial-based language features $L_{s} \\in R^{T_{L_{s}} \\times C_{L_{s}}}$ into image features $M_{i} \\in R^{H_{i} \\times W_{i} \\times C_{i}}$ via cross attention implementing", + "bbox": [ + 212, + 515, + 784, + 592 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A t t e n t i o n} \\left(M _ {i} ^ {q}, L _ {s} ^ {k}, L _ {s} ^ {v}\\right) = \\operatorname {s o f t m a x} \\left(\\frac {M _ {i} ^ {q} L _ {s} ^ {k T}}{\\sqrt {d}}\\right) \\cdot L _ {s} ^ {v} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 601, + 784, + 641 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "with $L_{s}^{k} = W^{k}L_{s}$ , $L_{s}^{v} = W^{v}L_{s}$ , $M_{i}^{q} = \\text{flatens}(W^{q}M_{i})$ , and $L_{s}^{k}, L_{s}^{v} \\in R^{T_{L_{s}} \\times C_{i}}$ , $M_{i}^{q} \\in R^{H_{i}W_{i} \\times C_{i}}$ . $W^{k}, W^{v}, W^{q}$ are implemented as a $1 \\times 1$ convolution with $C_{i}$ number of output channels.", + "bbox": [ + 214, + 650, + 782, + 696 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Deformable Feature Alignment. The TOG task presents a unique challenge in that the given image lacks the referred object, which is adverse to aforementioned attention mechanism. The input language encompasses both appearance and spatial attributes. If exclusively relying on spatial information, the model can only provide a generic localization of the region due to its limited grasp of object concepts, consequently compromising the semantic rationality of edited images (e.g. a generated mouse being larger than the existing elephant). However, when appearance attributes are incorporated into the input, the cross-attention", + "bbox": [ + 212, + 718, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "X. Xue et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/02720657c05cfc12c4222f6b1a9070a19be0a25695f6d194554d9d0ca168a4a0.jpg", + "image_caption": [ + "Fig. 3: An illustration of deformable feature alignment. Global visual feature $V_{i}$ , multimodal feature $M_{i}$ , and text feature $L$ are concatenated and inputted into the Transformer layer to learn offsets and modulation scalar. The multi-head cross attention is then performed to generate the refined attention map with sampled queries in $M_{i}$ and language keys and values." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 381 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "module faces a dilemma. It is tasked with reconciling the existing spatial attributes with the absent appearance features in the image, leading to conflicting information. This predicament significantly hampers the model's performance and generalization ability, especially in cases where the described appearance features are already present in the given image.", + "bbox": [ + 212, + 491, + 784, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Inspired by deformation mechanisms [8,39,47], we propose deformable feature alignment to mitigate potential conflicts by shifting candidate queries towards regions of interest in multi-head cross attention, as illustrated in Fig. 3. The multimodal map $M_{i}$ focuses on visually existing positional features by capturing the correlation between spatial-based textual and visual features, and it undergoes deformation with the alignment of object descriptions and multi-scale image features.", + "bbox": [ + 212, + 566, + 785, + 671 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Offset Generation. Initially, we add a sinusoidal 1D positional encoding $e_{pos} \\in R^{T_L \\times C_L}$ to the word embedding $L$ , updating as $L \\gets L + e_{pos}$ . A uniform grid of points $p \\in R^{H_i / \\gamma \\times W_i / \\gamma \\times 2}$ are generated as references on $M_i$ , where $\\gamma$ is the sampling factor. $M_i$ , $V_i$ , and $L$ are normalized to the same dimension through a linear layer and then concatenated as inputs into a transformer layer to generate patch-wise multimodal offset features. Subsequently, a linear projection is applied to modify the dimension to $H_i / \\gamma \\times W_i / \\gamma \\times 3$ , where the first two channels correspond to the learned 2D offsets $\\Delta p$ constrained with maximum offset value $s_i$ ( $\\Delta p \\gets s_i \\tanh(\\Delta p)$ ), and the last channel is fed to a sigmoid layer to obtain modulation scalar $\\Delta m$ , which is introduced to exclude redundant context of which the elements lie in the range [0, 1].", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "ST-LDM", + "bbox": [ + 668, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Multi-head Cross Attention. For reference points $p$ on $M_{i}$ , 2D offsets $\\Delta p$ and modulation scalar $\\Delta m$ are generated via offset generation module. $\\Delta p$ refines the spatial position according to textual descriptions, and $\\Delta m$ introduces an additional dimension of freedom to exclude irrelevant context. Both of them are learned from the same transformer layer and activated with distinct modules. We designate $\\Delta p_{k}$ and $\\Delta m_{k}$ as the offset and scalar for the $k$ -th reference point $p_{k}$ , where $k \\in \\{0,1,2,\\dots,H_iW_i / \\gamma^2\\}$ . As $p_{k} + \\Delta p_{k}$ is fractional, a bilinear interpolation $\\phi (I,p_k)$ is applied as", + "bbox": [ + 212, + 146, + 787, + 268 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\phi (I, p _ {k}) = \\sum_ {q} g \\left(p _ {k x}, q _ {x}\\right) \\cdot g \\left(p _ {k y}, q _ {y}\\right) \\cdot I (q), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 349, + 277, + 785, + 310 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $g(a, b) = \\max(0, 1 - |a - b|)$ and $q$ indexes the 4 integral points closest to $p_k$ . The sampling queries are deformed as", + "bbox": [ + 212, + 320, + 785, + 351 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nM _ {i} ^ {q} = \\operatorname {f l a t t e n} \\left(W ^ {q} \\phi \\left(M _ {i}, p + \\Delta p\\right)\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 362, + 785, + 378 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The sampling visual queries are attended to the language keys with modulate attention scores $\\Delta m$ . A matrix completion function $\\psi(A)$ is then introduced to fill in the attention score of unsampled queries and restore the dimension of the attention map to $H_i W_i \\times T_{L_s}$ . For each unsampled query $a_u$ in $A$ , the attention score of $a_u$ is the weighted sum of scores of all the sampled queries $a_s$ in the range $\\mathbb{R}$ around $a_u$ . The number of $a_s$ in the range $\\mathbb{R}$ , denoted as $\\operatorname{card}(a_s)$ , reflects the degree of deformation near $a_u$ and is introduced as a modulation factor. $\\operatorname{Avg}(\\operatorname{card}(a_s))$ refers to the average number of $\\operatorname{card}(a_s)$ for all the $a_u$ in $A$ .", + "bbox": [ + 212, + 388, + 787, + 521 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\psi \\left(a _ {u}\\right) = \\frac {\\sum_ {a _ {s} \\in \\mathbb {R}} w _ {s} a _ {s}}{\\sum w _ {s}} \\cdot \\frac {\\operatorname {c a r d} \\left(a _ {s}\\right)}{\\operatorname {A v g} \\left(\\operatorname {c a r d} \\left(a _ {s}\\right)\\right)}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 364, + 521, + 785, + 555 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nw _ {s} = \\frac {1}{\\mathrm {d} \\left(a _ {u} , a _ {s}\\right) + \\epsilon}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 433, + 564, + 785, + 594 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathrm{d}(a_u, a_s)$ refers to the Euclidean distance of $a_u$ and $a_s$ , and $\\epsilon$ is a modulation constant. The output of the $j$ -th attention head is formulated as", + "bbox": [ + 212, + 599, + 785, + 630 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nZ ^ {(j)} = \\operatorname {s o f t m a x} \\left(\\psi \\left(\\frac {M _ {i} ^ {q} L _ {s} ^ {k T}}{\\sqrt {d}} \\cdot \\varDelta m\\right)\\right) \\cdot L _ {s} ^ {v}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 354, + 638, + 785, + 680 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We preserve the original structure of cross attention in the initial fusion phase and utilize spatial conditions $L_{s}$ as the key and value to generate a coarse-grained attention region. In subsequent stages, we introduce deformable feature alignment with $s_i$ set as $\\{8,4,2\\}$ to progressively refine spatial positions during the process of deep feature extraction. This strategy can bias the attention to spatial position information rather than image local features, simultaneously with fast convergence and high computational efficiency. In the last layer, we directly extract the attention score in the multi-head cross attention and then convert it to a two-dimensional matrix averaged by the channel axis, denoted as $G$ , as location guidance in denoising U-Net.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "X. Xue et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.2 Training-free Spatial Guidance", + "text_level": 1, + "bbox": [ + 215, + 146, + 514, + 161 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Location guidance $G$ encompasses spatial attention generated in pixel space, delineating the approximate boundary of the generation position. Despite the refinement conducted through deformable feature alignment which adaptively adjusts spatial localization based on multimodal information, the numerical difference of attention scores in $G$ may influence the shape of the generated object, compared to a manually defined bounding box. Thus, we utilize region-wise backward guidance to avoid potential overly aggressive intervention.", + "bbox": [ + 212, + 178, + 782, + 282 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "LDMs [33] augment underlying UNet backbone with spatial transformer, where the core module is cross-attention with $L_{k}^{(\\iota)} = W_{k}L_{a}$ , $L_{v}^{(\\iota)} = W_{v}L_{a}$ , $V_{q}^{(\\iota)} = \\text{ flatten}(W_{q}z_{t})$ , where $\\iota$ is the index of the relevant layer in the network, $L_{a}$ refers to appearance-based text prompt, and $z_{t}$ is the latent representation of the input image. The attention weight matrix is described as:", + "bbox": [ + 214, + 284, + 784, + 364 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nS ^ {(\\iota)} = \\operatorname {s o f t m a x} \\left(\\frac {V _ {q} ^ {(\\iota)} L _ {k} ^ {(\\iota) ^ {T}}}{\\sqrt {d}}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 393, + 378, + 784, + 426 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We extract the first cross-attention block in the up-sampling branch to compute backpropagation, which yields superior performance while incurring minimal computational time overhead [6]. The corresponding attention score at timestamp $t$ is denoted as $S_{t}$ , and it has a similar compression scale with $G$ , which is advantageous in subsequent calculations pertaining to the energy function. The energy function is defined as", + "bbox": [ + 214, + 440, + 782, + 531 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nE \\left(S _ {t}, G\\right) = \\left(1 - \\frac {\\sum \\left(\\theta \\left(\\hat {G} , \\beta\\right) \\odot S _ {t}\\right)}{\\sum S _ {t}}\\right) ^ {2}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 366, + 545, + 784, + 588 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $G$ is converted to the same resolution as $S_{t}$ , denoted as $\\hat{G}$ . $\\theta (\\hat{G},\\beta)$ is an activation function with two steps, defined as", + "bbox": [ + 214, + 602, + 782, + 633 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\theta (\\hat {G}, \\beta) = \\operatorname {D i l a t e} (\\hat {G}) \\leftarrow \\hat {G} _ {m n} = \\left\\{ \\begin{array}{l l} 1 & \\hat {G} _ {m n} > \\beta \\\\ 0 & \\hat {G} _ {m n} \\leq \\beta \\end{array} , \\right. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 647, + 784, + 688 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where the element $\\hat{G}_{mn}$ is first activated with the threshold value $\\beta$ , and then $\\hat{G}$ is dilated to contour external maxmatrix to eliminate the numerical interference on the shape of the object. At each denoising U-Net, the latent $\\mathbf{z}_t$ is updated with the gradient of Eq. 8.", + "bbox": [ + 214, + 702, + 782, + 763 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {z} _ {t} \\leftarrow \\boldsymbol {z} _ {t} - \\eta \\sqrt {\\left(1 - \\alpha_ {t}\\right) / \\alpha_ {t}} \\nabla_ {\\boldsymbol {z} _ {t}} E (S _ {t}, G), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 362, + 781, + 784, + 800 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\eta > 0$ defines the strength of guidance and $\\alpha_{t}$ refers to the decreasing sequence of noise schedule in stable diffusion.", + "bbox": [ + 214, + 809, + 782, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "ST-LDM", + "bbox": [ + 668, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 143, + 375, + 162 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Dataset", + "text_level": 1, + "bbox": [ + 215, + 176, + 328, + 191 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Our primary focus of chosen datasets is the linguistic diversity of spatial conditions rather than text2img prompts since our framework is explicitly tailored for spatial guidance with the parameter layers of LDMs held frozen.", + "bbox": [ + 212, + 202, + 782, + 244 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "VrR-VG. VrR-VG [25] is constructed from Visual Genome [21] and contains 58,983 real images with 23,375 relation pairs. Excluding positional and statistically biased relationships, It encompasses more visually relevant relationships, like \"sit on\", \"other side of\", \"fly in\", etc. We select spatial relationships and restructure linguistic expressions to align with our input format. For instance, for annotations involving attributes (fire hydrant, yellow), attributes (man, standing), and beside (man, fire hydrant), we reformulate them as \"standing man [beside yellow fire hydrant]\". Subsequently, We remove the specific object and complete images employing image inpainting techniques [11] as the input image, and the ground truth is the original position of the removed object.", + "bbox": [ + 212, + 247, + 784, + 397 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "ComCOCO. ComCOCO [40] is derived from RefCOCO + [20] and consists of 136,495 referring expressions for 34,615 objects in 23,951 image pairs. It is a synthetic dataset and the spatial relationship is delineated in the process of manual inspection. The annotations within ComCOCO encompass both appearance and spatial attributes, and we combine them in a manner analogous to the approach detailed above. We utilize real images in ComCOCO as the input image, and the corresponding synthetic images provide the ground-truth position.", + "bbox": [ + 212, + 398, + 785, + 503 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 527, + 455, + 542 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We utilize SwinV2-T [27] with layer number $= \\{2, 2, 6, 2\\}$ pretained on ADE20K [46] as the autoencoder in pixel space. Text embedding is initialized with official weights of ALBERT Base v2 [23] with 11M parameters and 12 layers. Our framework is adaptable to various latent diffusion models, and we conduct experiments by building upon a pretrained Stable-Diffusion (SD) V-1.5 [33] trained on the LAION-5B dataset [35], unless stated otherwise.", + "bbox": [ + 212, + 551, + 802, + 643 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The training process is separated into two distinct phases. First, we train patch expanding and patch partition module on OpenImages [22] with KL-regularization following [33] to ensure optimal scaling properties of the pretrained autoencoder concerning spatial dimensionality when applied to LDMs. Subsequently, we train multimodal fusion transformer layers with the backbone of Swin-T on a combined dataset comprising a modified version of VrR-VG and ComCOCO. VrR-VG contributes a diverse and balanced set of spatial relationships, albeit with potential inpainting artifacts. To mitigate this limitation, we introduce ComCOCO which employs real images as its input source. Moreover, spatial descriptions in ComCOCO predominantly revolve around the holistic perspective of an image, whereas VrR-VG is specifically directed toward capturing and describing the intricate relationships between objects within the visual content.", + "bbox": [ + 212, + 643, + 785, + 838 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "X. Xue et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "All the experiments are implemented on 4 V100 GPUs of 32G RAM. We set the downsampling factor $f = 4$ and channel dimension $c = 3$ in the autoencoder. Backward guidance is performed during the initial 10 steps of the diffusion process and repeated 3 times at each step with the strength factor $\\eta$ set to 35. The input image is scaled to $480 \\times 480$ , and the data augmentation strategy includes random cropping and flipping. The optimizer is AdamW [29] with $\\beta_{1} = 0.85$ , $\\beta_{2} = 0.91$ . The weight decay factor and the learning rate of the initial exponential decay are set to 0.003 and 0.00024, respectively, and the remaining parameters are randomly initialized. The model is trained for 172 rounds with the batch size set to 32 to achieve optimal performance.", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Qualitative Evaluation", + "text_level": 1, + "bbox": [ + 215, + 320, + 446, + 335 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We apply our method on a multitude of real images from various domains, as exhibited in Fig. 1. We further compare the localization ability in Fig. 4 and generation ability in Fig. 5. Figure 4 shows the qualitative evaluation of ablative comparison of appearance and spatial properties. The deformable feature alignment module exhibits a remarkable ability to adjust spatial positions in accordance with linguistic descriptions and visual context. This dynamic adjustment imparts a direct influence on the image editing process within the latent diffusion model. As illustrated in Fig. 5, our method maintains the robust generative capabilities intrinsic to diffusion models with the preservation of semantics coherence to the original scene. The spatial constraints exerted on the denoising process have no discernible impact on the diversity and fidelity of the generated objects owing to region-wise backward guidance. More visualizations are provided in Appendix C.", + "bbox": [ + 212, + 347, + 787, + 544 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/f4fd301c86199d473189b65968815287a43fbf96ad595419b264779e3231db17.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 569, + 331, + 657 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/d3744cd16794620b204ef968e68e95c07e6956281e5710ecdffdc7440e46dacf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 334, + 569, + 447, + 657 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/0dfee3714abbba89db6a23c00153c90d134102a743dc4b80b41409249f1e1f81.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 449, + 569, + 560, + 657 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/5ae5482ab908094ecb016182e6ea3bae4768d36501ad60c5cd9cca038cfb96b4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 560, + 569, + 669, + 657 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/1ed9ee9a8cbf41e1aa65360fc5f9d28be27d9bf03cd8a9310e55999205b27c81.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 569, + 784, + 657 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/391fdfc2858759940eebe066417b7eddfb1fa3a648cc68b7934b12dd52754a3b.jpg", + "image_caption": [ + "Fig. 4: The first column corresponds to the input real image, and we visualize the guidance attention map $G$ and corresponding edited results under different prompts with identical conditions (top) and identical prompts with different conditions (bottom)." + ], + "image_footnote": [], + "bbox": [ + 217, + 675, + 331, + 763 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/1eac37b2a3679326598843ce777434bd8cc877db5748a21f7c3e5d140450640c.jpg", + "image_caption": [ + "beef burger with ketchup on the nearby plate" + ], + "image_footnote": [], + "bbox": [ + 334, + 675, + 447, + 762 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/f389749c01cd816b21580115703a97ecb24e71940739e3ff5f9eee65772986be.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 449, + 675, + 560, + 762 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/1bcba744b19c479e865617d342e9376515cbef8380c04e5fb89655efa7f058d5.jpg", + "image_caption": [ + "beef burger with ketchup on the more distant plate" + ], + "image_footnote": [], + "bbox": [ + 560, + 675, + 671, + 762 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/366a04d5421400277d6ecfef5c9276b33ce1b8c9e9455c1cc0e4e591715dafd1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 671, + 675, + 784, + 762 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "ST-LDM", + "bbox": [ + 668, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/6e5fccfce4d3076a0de15f78d1957168ba8b78e481ad38890ad2059d1d9e6ca6.jpg", + "image_caption": [ + "a blue luxury car with a streamlined body and a shiny paint job [to the right of the bus]" + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 331, + 231 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/77ccf27d14991410aa3e7aec9558bda056903c8db74cbfc5fcd3a9c586f22d3a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 334, + 143, + 447, + 231 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/0ae96e1d50b3b3f75d4c40a823d1111285245d51ce38ab432f3fc27f79dcc2b6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 447, + 143, + 558, + 231 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/9c46e48637fc8f6dbd1f600d2c0dae642202a1a071f3d24e8bdb58c1a5eac9e8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 558, + 143, + 660, + 231 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/69fcacd1b12763f8cdc47c93b3ef7313fa524a3ad9b07409a6ff523853f7c17a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 660, + 143, + 782, + 231 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/3486a468f8e0436bf52fb9291cc80c8d2394ac2279409044e53ebcee890deb90.jpg", + "image_caption": [ + "Fig. 5: Various results produced by ST-LDM, showcasing diverse appearances while adhering to the provided text descriptions." + ], + "image_footnote": [], + "bbox": [ + 218, + 243, + 331, + 330 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/ce6a1228682cbc9b9014a7f81cddf3d00c857f6739c1ce428f78cf9d01a79e5b.jpg", + "image_caption": [ + "a cute dinosaur with round face, animation style [standing in front of the door]" + ], + "image_footnote": [], + "bbox": [ + 334, + 243, + 450, + 330 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/b699748339e401ea9bc10a70fc17193429cb25d2a6ab95cfcc188059ae962d42.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 450, + 243, + 562, + 330 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/38a635fd2cfaed842fcd5e0c829069487bbeab5c3cc5b7b80aa5b6029c9e18cd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 562, + 243, + 673, + 330 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/378271a66860bc610209cca6db9a86be9373695c327efc750372ae0ce3a0d192.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 673, + 243, + 782, + 330 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Quantitative Comparison", + "text_level": 1, + "bbox": [ + 215, + 411, + 472, + 428 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Evaluation Metrics. We conduct quantitative assessments of the quality of edited images by examining two facets. (1) Object alignment: We introduce VISOR [14] to score spatial object relationships according to text descriptions and text2img metric CLIPScore [17] to measure the consistency of generated objects with the input text. (2) Image fidelity: We utilize full-reference image quality assessments FID [44] and LPIPS [45] to measure image fidelity. HPS [38] is also introduced to align with human preference.", + "bbox": [ + 212, + 435, + 787, + 541 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Benchmark Dataset. We construct a benchmark dataset for quantitative analysis and compare ST-LDM with related tasks. We find a distinct bias in the textual descriptions associated with the presence or absence of the specific object within the image. When the object is present in an image, the descriptions tend to emphasize distinctive features that differentiate the object from other visual entities. Conversely, when generating an object that does not exist in the original image, the focus shifts towards the detailed attributes of the object. Therefore, we randomly select 1,500 images from COCO2014 [26], and manually describe the object required to be generated in each image, simultaneously grounding the ideal position using bounding boxes. The detailed construction process is illustrated in Appendix B.", + "bbox": [ + 212, + 561, + 787, + 729 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Results. We compare our method with editing models conditioned by diverse modalities in Tab. 1. Existing text-guided editing models exhibit limitations in generating new objects within complex scenes, and our model outperforms by a significant margin benefiting from explicit guidance generated from the deformable feature alignment. When compared to approaches employing forced modalities to precisely confine generated objects to designated regions,", + "bbox": [ + 212, + 750, + 787, + 842 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "X. Xue et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "our method exhibits the capacity to edit local region of real images under the weak supervision of language, and the generated objects accurately adhere to the spatial relationships as specified in the text conditions.", + "bbox": [ + 212, + 146, + 787, + 191 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/3dac86d9a00ec746e9f7c23857f25ba98eca3e42e608b6ef7225faef9b5fdd52.jpg", + "table_caption": [ + "Table 1: Comparison with editing models conditioned by diverse modalities." + ], + "table_footnote": [], + "table_body": "
ConditionModelFID↓LPIPS↓HPS↑CLIPScore↑VISOR↑
TextInstructPix2Pix [5]22.2530.9341.5860.4572.15
TextSmartEdit [18]18.4727.1147.4069.0277.86
MaskBlendedLatentDiffusion [3]12.9318.0554.3380.9688.12
MaskPaintByWord [1]9.7220.6850.2681.2387.41
BBoxBreakAScene [2]7.8916.2459.2581.1091.95
BBoxGLIGEN [24]6.2417.0962.4180.6290.53
TextST-LDM6.1615.5464.2881.9792.39
", + "bbox": [ + 217, + 234, + 785, + 369 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.5 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 411, + 388, + 426 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Deformable Feature Alignment. We replace the feature alignment of multimodal fusion Transformer with our deformable mechanism at different stages, resulting in a direct impact on the modification of generated regions, as depicted in Fig. 6. We introduce FID and LPIPS to assess the extent of alteration in the original image. A user study is also conducted to assess the appropriateness of object position, which serves as a more persuasive and reference-worthy metric. As delineated in Tab. 2, the integration of deformable feature alignment refines the attention region, resulting in a numerical improvement across all metrics. However, premature intervention of DFA in the initial stage interferes with spatial perception (since the input $M_{1}$ of DFA will be equal to $V_{1}$ if we introduce DFA in Stage1). Actually, in the initial stage of modal fusion, the model generates a preliminary spatial positioning via cross attention, and DFA is designed as a refining strategy. More illustrations are provided in Appendix A.", + "bbox": [ + 212, + 436, + 787, + 635 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/f1f2b5f93a62236a571c74f5ceee8fe2f6062579e0f988c9aa60eacea8a7f3a3.jpg", + "table_caption": [ + "Table 2: Ablation study of deformable feature alignment (DFA) applied at different stages. Users are required to estimate the conformance of the object's position (True or False), and the score is calculated based on the overall accuracy across all the images. IT refers to the inference time (sec/image)." + ], + "table_footnote": [], + "table_body": "
Stages w/ DFAUserScore↑FID↓LPIPS↓IT
S-1S-2S-3S-4
✓✓✓✓89.37.5016.23~7.0
✓✓✓92.46.1615.54~6.5
✓✓91.19.1420.46~6.3
✓82.515.9725.43~6.1
75.927.0334.86~6.0
", + "bbox": [ + 272, + 729, + 732, + 837 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "ST-LDM", + "bbox": [ + 668, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/98a01e03cf7a630875dd1e9b44d1ca816c2a95f6d147a35be63d8f2e2e3f88ee.jpg", + "image_caption": [ + "Fig. 6: Ablative visualization of attention maps at diverse stages with DFA." + ], + "image_footnote": [], + "bbox": [ + 215, + 142, + 785, + 311 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Backward guidance. We further assess the effectiveness of backward guidance in Tab. 3. We replace guidance maps with textual conditions during the denoising process, and the notable decline indicates that linguistic information inadequately constrains the spatial placement of generated objects. We employ an activation function to eliminate potential discrepancies in attention scores, followed by dilation to mitigate the morphological interference of the attention region to the object.", + "bbox": [ + 212, + 364, + 787, + 470 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/d8d6c7a44d4359f36a28a97d6e0845581c8fec5505b37f90ffef67c557882c47.jpg", + "table_caption": [ + "Table 3: Ablation study of backward guidance. Users are required to assess the matching score between the edited image and spatial descriptions. IT refers to the inference time (sec/image)." + ], + "table_footnote": [], + "table_body": "
UserScoreCLIPScoreVISORIT
uncondcond
w/o guidance58.7355.0711.5460.45~4.1
w/o activation80.8669.0137.2690.94~6.0
w/o dilation83.9870.5239.1490.37~6.2
Ours89.1181.9743.4192.39~6.5
", + "bbox": [ + 261, + 551, + 736, + 665 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 700, + 359, + 717 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we explore the text-grounded object generation in real-world images. A universal framework with deformable feature alignment is proposed to hierarchically refine the attention region conforming to vision and language features. The generated guidance map can be incorporated into any latent diffusion model with training-free backward guidance. Our model addresses the inherent limitation of diffusion models in spatial perception under textual supervision, and offers valuable insights for subsequent research.", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "X. Xue et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported in part by the National Science and Technology Major Project under Grant 2021ZD0113202, in part by the National Natural Science Foundation of China under Grants 62171125, 61876037, in part by the innovation project of Jiangsu Province under grants BZ2023042, BY2022564. The authors would like to thank the Big Data Computing Center of Southeast University for providing the facility support on the numerical calculations in this paper.", + "bbox": [ + 212, + 176, + 787, + 268 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 292, + 321, + 308 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Andonian, A., Osmany, S., Cui, A., Park, Y., Jahanian, A., Torralba, A., Bau, D.: Paint by word", + "2. Avrahami, O., Aberman, K., Fried, O., Cohen-Or, D., Lischinski, D.: Break-a-scene: Extracting multiple concepts from a single image. arXiv preprint arXiv:2305.16311 (2023)", + "3. Avrahami, O., Fried, O., Lischinski, D.: Blended latent diffusion. ACM Transactions on Graphics 42(4), 1-11 (Jul 2023)", + "4. Avrahami, O., Hayes, T., Gafni, O., Gupta, S., Taigman, Y., Parikh, D., Lischinski, D., Fried, O., Yin, X.: Spatext: Spatio-textual representation for controllable image generation. In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE (Jun 2023)", + "5. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18392-18402 (2023)", + "6. Chen, M., Laina, I., Vedaldi, A.: Training-free layout control with cross-attention guidance (2023)", + "7. Couairon, G., Verbeek, J., Schwenk, H., Cord, M.: Diffedit: Diffusion-based semantic image editing with mask guidance. In: The Eleventh International Conference on Learning Representations (2023)", + "8. Dai, J., Qi, H., Xiong, Y., Li, Y., Zhang, G., Hu, H., Wei, Y.: Deformable convolutional networks. In: 2017 IEEE International Conference on Computer Vision (ICCV). pp. 764-773 (2017)", + "9. Ding, H., Liu, C., Wang, S., Jiang, X.: Vision-language transformer and query generation for referring segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 16321-16330 (October 2021)", + "10. Ding, Z., Zhang, X., Xia, Z., Jebe, L., Tu, Z., Zhang, X.: Diffusionrig: Learning personalized priors for facial appearance editing. In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 12736-12746 (2023)", + "11. Dong, Q., Cao, C., Fu, Y.: Incremental transformer structure enhanced image inpainting with masking positional encoding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022)", + "12. Fan, W.C., Chen, Y.C., Chen, D., Cheng, Y., Yuan, L., Wang, Y.C.F.: Frido: Feature pyramid diffusion for complex scene image synthesis. In: AAAI (2023)", + "13. Gafni, O., Polyak, A., Ashual, O., Sheynin, S., Parikh, D., Taigman, Y.: Make-a-scene: Scene-based text-to-image generation with human priors. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022. pp. 89-106. Springer Nature Switzerland, Cham (2022)" + ], + "bbox": [ + 218, + 324, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "ST-LDM", + "bbox": [ + 668, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Gokhale, T., Palangi, H., Nushi, B., Vineet, V., Horvitz, E., Kamar, E., Baral, C., Yang, Y.: Benchmarking spatial relationships in text-to-image generation (2022)", + "15. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control (2022)", + "16. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-or, D.: Prompt-to-prompt image editing with cross-attention control. In: The Eleventh International Conference on Learning Representations (2023)", + "17. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: CLIPScore: a reference-free evaluation metric for image captioning. In: EMNLP (2021)", + "18. Huang, Y., Xie, L., Wang, X., Yuan, Z., Cun, X., Ge, Y., Zhou, J., Dong, C., Huang, R., Zhang, R., et al.: Smartedit: Exploring complex instruction-based image editing with multimodal large language models. arXiv preprint arXiv:2312.06739 (2023)", + "19. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 6007-6017 (June 2023)", + "20. Kazemzadeh, S., Ordonez, V., Matten, M., Berg, T.: Referitgame: Referring to objects in photographs of natural scenes. In: Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP). pp. 787-798 (2014)", + "21. Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., et al.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision 123, 32-73 (2017)", + "22. Kuznetsova, A., Rom, H., Alldrin, N., Uijlings, J., Krasin, I., Pont-Tuset, J., Kamali, S., Popov, S., Malloci, M., Kolesnikov, A., et al.: The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision 128(7), 1956-1981 (2020)", + "23. Lan, Z., Chen, M., Goodman, S., Gimpel, K., Sharma, P., Soricut, R.: ALBERT: A lite BERT for self-supervised learning of language representations. In: 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net (2020)", + "24. Li, Y., Liu, H., Wu, Q., Mu, F., Yang, J., Gao, J., Li, C., Lee, Y.J.: Gligen: Open-set grounded text-to-image generation. CVPR (2023)", + "25. Liang, Y., Bai, Y., Zhang, W., Qian, X., Zhu, L., Mei, T.: Vrr-vg: Refocusing visually-relevant relationships. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10403-10412 (2019)", + "26. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: European conference on computer vision. pp. 740-755. Springer (2014)", + "27. Liu, Z., Hu, H., Lin, Y., Yao, Z., Xie, Z., Wei, Y., Ning, J., Cao, Y., Zhang, Z., Dong, L., et al.: Swin transformer v2: Scaling up capacity and resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12009-12019 (2022)", + "28. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 10012-10022 (October 2021)", + "29. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: International Conference on Learning Representations (2017)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "X. Xue et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "30. Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Van Gool, L.: Repaint: Inpainting using denoising diffusion probabilistic models. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11451-11461 (2022)", + "31. Nichol, A.Q., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., Mcgrew, B., Sutskever, I., Chen, M.: GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. In: Chaudhuri, K., Jegelka, S., Song, L., Szepesvari, C., Niu, G., Sabato, S. (eds.) Proceedings of the 39th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 162, pp. 16784-16804. PMLR (17-23 Jul 2022)", + "32. OuYang, S.R., Wang, H., Xie, S., Niu, Z., Tong, R., Chen, Y.W., Lin, L.: Slvit: Scale-wise language-guided vision transformer for referring image segmentation. In: International Joint Conference on Artificial Intelligence (2023)", + "33. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022)", + "34. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 22500-22510 (June 2023)", + "35. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., Schramowski, P., Kundurthy, S., Crowson, K., Schmidt, L., Kaczmarczyk, R., Jitsev, J.: Laion-5b: An open large-scale dataset for training next generation image-text models. In: Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., Oh, A. (eds.) Advances in Neural Information Processing Systems. vol. 35, pp. 25278-25294. Curran Associates, Inc. (2022)", + "36. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L.u., Polosukhin, I.: Attention is all you need. In: Guyon, I., Luxburg, U.V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., Garnett, R. (eds.) Advances in Neural Information Processing Systems. vol. 30. Curran Associates, Inc. (2017)", + "37. Wang, J., Lu, G., Xu, H., Li, Z., Xu, C., Fu, Y.: Manitrans: Entity-level text-guided image manipulation via token-wise semantic alignment and generation. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10697-10707 (2022)", + "38. Wu, X., Sun, K., Zhu, F., Zhao, R., Li, H.: Human preference score: Better aligning text-to-image models with human preference. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2096-2105 (October 2023)", + "39. Xia, Z., Pan, X., Song, S., Li, L.E., Huang, G.: Vision transformer with deformable attention. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4794-4803 (June 2022)", + "40. Xue, X., Wu, J., Kong, Y., Senhadji, L., Shu, H.: Rethinking referring object removal (2024)", + "41. Yang, L., Zhang, Z., Song, Y., Hong, S., Xu, R., Zhao, Y., Shao, Y., Zhang, W., Cui, B., Yang, M.H.: Diffusion models: A comprehensive survey of methods and applications. arXiv preprint arXiv:2209.00796 (2022)", + "42. Yang, Z., Wang, J., Tang, Y., Chen, K., Zhao, H., Torr, P.H.: Lavo: Language-aware vision transformer for referring image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 18155-18165 (June 2022)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "ST-LDM", + "bbox": [ + 668, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "43. Yang, Z., Wang, J., Gan, Z., Li, L., Lin, K., Wu, C., Duan, N., Liu, Z., Liu, C., Zeng, M., Wang, L.: Reco: Region-controlled text-to-image generation. In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)", + "44. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)", + "45. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2018)", + "46. Zhou, B., Zhao, H., Puig, X., Xiao, T., Fidler, S., Barriuso, A., Torralba, A.: Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision 127, 302-321 (2019)", + "47. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. In: 9th International Conference on Learning Representations, ICLR (2021)", + "48. Zhu, Y., Liu, H., Song, Y., Yuan, Z., Han, X., Yuan, C., Chen, Q., Wang, J.: One model to edit them all: Free-form text-driven image manipulation with semantic modulations. In: Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., Oh, A. (eds.) Advances in Neural Information Processing Systems. pp. 25146-25159. Curran Associates, Inc." + ], + "bbox": [ + 215, + 146, + 785, + 436 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "X. Xue et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/ST-LDM_ A Universal Framework for Text-Grounded Object Generation in Real Images/ea78a6df-2485-4a42-a86e-90a6085c52c4_model.json b/2024/ST-LDM_ A Universal Framework for Text-Grounded Object Generation in Real Images/ea78a6df-2485-4a42-a86e-90a6085c52c4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f22f58490ced981f85cb156aa6193d77a615d555 --- /dev/null +++ b/2024/ST-LDM_ A Universal Framework for Text-Grounded Object Generation in Real Images/ea78a6df-2485-4a42-a86e-90a6085c52c4_model.json @@ -0,0 +1,2623 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.218, + 0.141, + 0.788, + 0.187 + ], + "angle": 0, + "content": "ST-LDM: A Universal Framework for Text-Grounded Object Generation in Real Images" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.211, + 0.777, + 0.244 + ], + "angle": 0, + "content": "Xiangtian Xue\\(^{1,2,4}\\), Jiasong Wu\\(^{1,2,4\\star}\\), Youyong Kong\\(^{1,2,4}\\), Lotfi Senhadji\\(^{2,3,4}\\), and Huazhong Shu\\(^{1,2,4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.255, + 0.764, + 0.284 + ], + "angle": 0, + "content": "1 Key Laboratory of New Generation Artificial Intelligence Technology and Its Interdisciplinary Applications (Southeast University), Nanjing 210096, China" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.284, + 0.784, + 0.31 + ], + "angle": 0, + "content": "\\(^{2}\\) Jiangsu Provincial Joint International Research Laboratory of Medical Information Processing, Southeast University, Nanjing 210096, China" + }, + { + "type": "text", + "bbox": [ + 0.277, + 0.31, + 0.726, + 0.324 + ], + "angle": 0, + "content": "3 Univ-Rennes, INSERM, LTSI-UMR 1099, Rennes F-3502, France" + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.325, + 0.775, + 0.352 + ], + "angle": 0, + "content": "4 Centre de Recherche en Information Biomédicale Sino-français (CRIBs), Rennes F-3502, France" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.378, + 0.765, + 0.391 + ], + "angle": 0, + "content": "Abstract. We present a novel image editing scenario termed Text-grounded" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.391, + 0.74, + 0.64 + ], + "angle": 0, + "content": "Object Generation (TOG), defined as generating a new object in the real image spatially conditioned by textual descriptions. Existing diffusion models exhibit limitations of spatial perception in complex real-world scenes, relying on additional modalities to enforce constraints, and TOG imposes heightened challenges on scene comprehension under the weak supervision of linguistic information. We propose a universal framework ST-LDM based on Swin-Transformer, which can be integrated into any latent diffusion model with training-free backward guidance. ST-LDM encompasses a global-perceptual autoencoder with adaptable compression scales and hierarchical visual features, parallel with deformable multimodal transformer to generate region-wise guidance for the subsequent denoising process. We transcend the limitation of traditional attention mechanisms that only focus on existing visual features by introducing deformable feature alignment to hierarchically refine spatial positioning fused with multi-scale visual and linguistic information. Extensive Experiments demonstrate that our model enhances the localization of attention mechanisms while preserving the generative capabilities inherent to diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.653, + 0.741, + 0.681 + ], + "angle": 0, + "content": "Keywords: Instruction-based image editing. Text-grounded object generation. Deformable feature alignment. Training-free guidance" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.701, + 0.377, + 0.717 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.728, + 0.788, + 0.82 + ], + "angle": 0, + "content": "Controllable text2img editing has demonstrated remarkable performance with ongoing exploration of diffusion models [10,34,48], and recent studies have been predominantly directed towards enhancing the controllability of language in editing the appearance of objects [15,19]. As for generating new objects in real images, additional modalities such as bounding boxes and keypoints have to be introduced to explicitly regulate their spatial placement [24]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.22, + 0.826, + 0.501, + 0.84 + ], + "angle": 0, + "content": "* Corresponding author: jswu@seu.edu.cn" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "X. Xue et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.311, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.313, + 0.144, + 0.404, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.144, + 0.498, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.328, + 0.216, + 0.385, + 0.223 + ], + "angle": 0, + "content": "wooden bench" + }, + { + "type": "image_caption", + "bbox": [ + 0.328, + 0.223, + 0.388, + 0.228 + ], + "angle": 0, + "content": "[bottom left corner]" + }, + { + "type": "image", + "bbox": [ + 0.313, + 0.228, + 0.404, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.315, + 0.298, + 0.402, + 0.31 + ], + "angle": 0, + "content": "fanny pack, black and white empty area on the table]" + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.228, + 0.498, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.411, + 0.298, + 0.495, + 0.31 + ], + "angle": 0, + "content": "old newspaper [on the \"MAP\" magazine]" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.144, + 0.596, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.504, + 0.216, + 0.596, + 0.223 + ], + "angle": 0, + "content": "" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.223, + 0.596, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.144, + 0.688, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.602, + 0.216, + 0.68, + 0.223 + ], + "angle": 0, + "content": "copper vase, with pattern" + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.223, + 0.689, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.6, + 0.296, + 0.688, + 0.31 + ], + "angle": 0, + "content": "blue sedan [between bus and green car]" + }, + { + "type": "image", + "bbox": [ + 0.691, + 0.144, + 0.783, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.692, + 0.216, + 0.78, + 0.223 + ], + "angle": 0, + "content": "woman with back turned" + }, + { + "type": "image", + "bbox": [ + 0.692, + 0.223, + 0.78, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.699, + 0.296, + 0.785, + 0.309 + ], + "angle": 0, + "content": "board, white letters on blue [on the roof of the bus]" + }, + { + "type": "image_caption", + "bbox": [ + 0.217, + 0.322, + 0.788, + 0.407 + ], + "angle": 0, + "content": "Fig. 1: Text-grounded object generation in real images. Our method can generate the rational object within real-world input images based on textual descriptions encompassing its visual characteristic and spatial attribute. The generated object seamlessly integrates with the original image, exhibiting a visually harmonious presence while preserving the contextual semantics of the original scene (e.g. consistency in shadow (top left), geometry (top right), illumination (bottom left), occlusion (bottom right))." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.418, + 0.785, + 0.552 + ], + "angle": 0, + "content": "The aforementioned modalities provide precise spatial constraints, but it is noteworthy that language-based instruction serves as the most fundamental and convenient interactive modality in real-world scenarios (e.g. chatting with a robot commonly using language as the sole mode of interaction, conducting similar manipulations on extensive image datasets and dynamic video sequences). Moreover, language affords a considerable degree of creative flexibility while adhering to the principles of spatial logic. This enhances the naturalness of edited images, as the model can estimate the most suitable generation location by comprehending multi-modal information." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.554, + 0.788, + 0.689 + ], + "angle": 0, + "content": "We propose a new image editing scenario named Text-grounded Object Generation (TOG), which aims to generate objects in real scenarios according to natural language expressions (Fig. 1). Concretely, given an image in conjunction with textual descriptions detailing visual and spatial attributes of the generated object, the task entails generating a rational object in complex scenarios. TOG embodies the dual characteristics of text2img generation and editing tasks. The generated object must adhere to textual descriptions, while simultaneously ensuring that the edited image maintains high fidelity with irrelevant areas remaining consistent with original semantics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Although diffusion models have demonstrated notable creative capabilities in image generation, they fall short in spatial perception and orientation, particularly when dealing with intricate real-world scenes [41]. The prevalent approaches to address this limitation involve introducing additional conditions during the denoising process to control the placement of the generated object. However, a fundamental paradox arises in our specific scenario: how to locate an object which is absent from the original image. This emphasizes the primary challenge, that is determining the optimal area for object generation under the weak supervision of texts. The objective is to ensure that the generated object aligns semantically with the contextual information contained within the image." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.669, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ST-LDM" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.344 + ], + "angle": 0, + "content": "In this paper, we present ST-LDM, a universal framework with hierarchical Swin-Transformer (ST) blocks [28], amenable for integration into latent diffusion models (LDM) via training-free backward guidance. ST-LDM comprises two parallel branches with similar structures, dedicated respectively to adaptable latent representation and spatial attention guidance. For latent representation, the Transformer-based autoencoder facilitates multi-resolution compression through a low-cost trainable module while preserving the formidable modeling capacity inherent to the pre-trained Swin-T. For spatial guidance, we propose deformable feature alignment in multimodal fusion Transformer to enhance the refinement of grounded regions by incorporating information from multi-scale visual and linguistic features. ST-LDM provides visually coherent location guidance in pixel space which can be integrated to the denoising U-Net of LDMs with region-wise backpropagation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.344, + 0.788, + 0.435 + ], + "angle": 0, + "content": "To quantify the effectiveness of our proposed approach, we compare it with text-prompted image editing models with diverse conditions based on our constructed benchmark. Experiments demonstrate that our model enhances the localization capability of Swin-Transformer while preserving the generative proficiency inherent in diffusion models. The main contributions of our work are highlighted as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.44, + 0.784, + 0.484 + ], + "angle": 0, + "content": "- We present a novel image editing scenario termed Text-grounded Object Generation (TOG), which is defined as generating objects in real images grounded by textual descriptions." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.484, + 0.785, + 0.529 + ], + "angle": 0, + "content": "- We further propose a universal framework ST-LDM with deformable feature alignment to hierarchically refine spatial guidance, which can be integrated into diffusion models with training-free region-wise backpropagation." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.53, + 0.785, + 0.573 + ], + "angle": 0, + "content": "- Quantitative comparisons with related works on our constructed benchmark and qualitative visualizations substantiate the robust editing and generative capabilities according to textual conditions." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.44, + 0.785, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.594, + 0.388, + 0.609 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.622, + 0.601, + 0.639 + ], + "angle": 0, + "content": "2.1 Spatially Controllable Object Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Recent studies have employed mandatory modalities to exert control over the spatial distribution of generated objects [6,12,43]. Make-A-Scene [13] introduces implicit conditioning over optionally controlled scene tokens derived from segmentation maps with fixed labels to control the layout of the image. SpaText [4] further proposes a spatial-textual representation that encapsulates the semantic attributes and structural characteristics of each segment. GLIGEN [24] extends grounding conditions to bounding boxes, reference images, keypoints, etc., which are injected into new trainable layers of the denoising process via gated self-attention. Nevertheless, these modalities are cumbersome when applied to large-scale operations involving numerous images and videos. The utilization of text-based spatial position descriptions serves to mitigate rigid constraints, thereby affording models a heightened capacity for comprehending visual information." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "X. Xue et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.594, + 0.163 + ], + "angle": 0, + "content": "2.2 Diffusion-Based Semantic Image Editing" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.17, + 0.788, + 0.382 + ], + "angle": 0, + "content": "Image editing emphasizes modifying the visual entity consistent with textual descriptions while preserving the entity-irrelevant regions. Diffusion models can be easily adapted for inpainting when given a mask input since they iteratively refine an image starting from random noise [30,31]. Recent approaches endeavor to obviate the necessity of additional inputs such as image masks or additional views of the object [6,16,37]. DiffEdit [7] automatically generates a mask accentuating regions of the input image necessitating modification, by contrasting predictions of a diffusion model conditioned on different text prompts. Imagic [19] produces a text embedding that aligns with both the input image and the target text while fine-tuning the diffusion model to capture the image-specific appearance. These attempts are experimented on images with obvious foreground objects and simple semantics. However, when applied to complex scenes characterized by a multiplicity of objects and spatial relationships, their efficacy in localizing objects exhibits a discernible diminishment." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.403, + 0.542, + 0.418 + ], + "angle": 0, + "content": "2.3 Vision-and-Language Transformer" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.426, + 0.788, + 0.622 + ], + "angle": 0, + "content": "The initial phase of our proposed method involves locating visual regions of interest guided by language, and Transformer [36] has demonstrated noteworthy performance at the interaction of vision and language. Ding et al. [9] introduce Transformer and multi-head attention to construct an encoder-decoder attention mechanism architecture. LAVT [42] achieves better cross-modal alignments through the early fusion of linguistic and visual features in a vision Transformer encoder network. SLViT [32] employs the variance of cross-modal correlations between adjacent stages to identify regions of high uncertainty and refine features of these regions with complementing information from multiple stages. While Transformer facilitates the integration of linguistic and visual modalities through cross-attention mechanisms, it encounters challenges stemming from the absence of target objects within the image, which is a distinctive characteristic of the TOG task." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.644, + 0.331, + 0.66 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.787, + 0.704 + ], + "angle": 0, + "content": "In this section, we present our proposed framework, denoted as ST-LDM, and the overall pipeline is depicted in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.84 + ], + "angle": 0, + "content": "In pixel space (Sec. 3.1), visual feature extraction and multi-modal localization are concurrently performed during the downsampling process, leveraging Swin Transformer [28] as the backbone. The feature extraction module acts as an autoencoder to concurrently reconstruct the image into a lower-dimensional latent space while furnishing hierarchical visual features for spatial localization. In parallel, the multi-modal Transformer module leverages the interplay between image and language to provide positional guidance for the denoising process. We innovatively introduce deformable feature alignment to dynamically adjust spatial constraints, ensuring that the generated object aligns with semantic context" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.669, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ST-LDM" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.146, + 0.784, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.378, + 0.788, + 0.504 + ], + "angle": 0, + "content": "Fig. 2: The basic framework of the proposed ST-LDM. Given a real image and text descriptions, we initially encode them into visual features \\( V_{i} \\) and text embedding \\( L \\), where \\( L \\) comprises appearance prompt \\( L_{a} \\) and spatial conditions \\( L_{s} \\). We employ the Swin-T structure to hierarchically integrate \\( L \\) into \\( V_{i} \\), where the multimodal map \\( M_{i} \\) is deformed by learning multi-scale visual and semantic relationships. The generated spatial guidance attention map \\( G \\) and textual prompt \\( L_{a} \\) are integrated into latent diffusion models with training-free guidance. The autoencoder-decoder in pixel space and the latent diffusion model in latent space retain original structure and parameters, and these components are frozen throughout the training process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.533, + 0.785, + 0.577 + ], + "angle": 0, + "content": "of the input image. The congruence in interaction between two branches is optimized through the application of a similar network structure (SwinT) to both of the aforementioned modules." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.579, + 0.788, + 0.7 + ], + "angle": 0, + "content": "In latent space (Sec. 3.2), appearance-based text prompt is separately mapped into denoising U-Net via cross attention. Subsequently, a region-wise backpropagation scheme is introduced to iteratively update the activations in the network to match the desired layout specified in pixel space. The fundamental structure and parameters of diffusion models remain constant, with trainable parameter layers exclusively concentrated in the pixel space. Consequently, the proposed ST-LDM emerges as a universal framework adaptable to a spectrum of pretrained LDMs." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.723, + 0.524, + 0.737 + ], + "angle": 0, + "content": "3.1 Hierarchical Feature Extraction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Autoencoder with Swin-Transformer. LDMs [33] leverages an individually pretrained autoencoding model for perceptual image compression, but its quantitative performance exhibits significant variability across different downsampling factors due to the inherent instability of its encoding structure. Recently, Swin-Transformer has emerged as a generic visual feature extractor, owing to its robust modeling capabilities [28]. It initially splits an input RGB image into" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "X. Xue et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "non-overlapping patches with a \\(4 \\times 4\\) size, and the resolution of subsequent layers' output continually decreases accordingly. However, excessive information compression will induce distortion of noisy images in the denoising process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.192, + 0.785, + 0.343 + ], + "angle": 0, + "content": "To address this issue, we add a trainable patch expanding layer to convert the selected feature map to the specified size. The original structure of SwinTransformer is retained for hierarchical feature extraction, and an intermediate feature map \\( V_{i} \\) is selected and upsampled to high-dimensional resolution via patch expanding layer. It can increase the resolution by a factor of 4 while simultaneously reducing the dimension to a configurable value \\( (H_{i}\\times W_{i}\\times C_{i}\\to 4H_{i}\\times 4W_{i}\\times C_{t},C_{i}\\geq 16C_{t}) \\). The layer consists of two operations, a linear layer to convert the feature dimension \\( (H_{i}\\times W_{i}\\times C_{i}\\to H_{i}\\times W_{i}\\times 16C_{t}) \\), and a rearrange operation to expand the resolution to \\( 4\\times \\) and reduce the dimension to one-sixteenth \\( (H_{i}\\times W_{i}\\times 16C_{t}\\to 4H_{i}\\times 4W_{i}\\times C_{t}) \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.343, + 0.787, + 0.495 + ], + "angle": 0, + "content": "The extracted feature map is optional within \\(\\{V_2, V_3, V_4\\}\\) contingent upon the depth of compression perception. As the initial patch size in Swin-Transformer is \\(4 \\times 4\\), the resulting output resolution for patch expansion remains fixed at \\(4 \\times\\) relative to its initial dimensions. Consequently, \\(\\{V_2, V_3, V_4\\}\\) corresponds to down-sampling factors \\(f = \\{2, 4, 8\\}\\). Due to the varied channel configurations across model variants, the output channel dimension remains adaptable. The encoding-decoding process in pixel space exhibits a symmetrical structure, wherein the linear layer in patch expanding and patch partition undergoes low-cost training following [33] to cater to specific downstream tasks while retaining the formidable modeling capabilities of pretrained models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.516, + 0.785, + 0.593 + ], + "angle": 0, + "content": "Multimodal Transformer. We jointly embed language and vision features during visual encoding with Swin-Transformer to generate semantically coherent positional guidance for object generation. Consistent with previous works [42], we densely integrate spatial-based language features \\( L_{s} \\in R^{T_{L_{s}} \\times C_{L_{s}}} \\) into image features \\( M_{i} \\in R^{H_{i} \\times W_{i} \\times C_{i}} \\) via cross attention implementing" + }, + { + "type": "equation", + "bbox": [ + 0.327, + 0.602, + 0.785, + 0.642 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A t t e n t i o n} \\left(M _ {i} ^ {q}, L _ {s} ^ {k}, L _ {s} ^ {v}\\right) = \\operatorname {s o f t m a x} \\left(\\frac {M _ {i} ^ {q} L _ {s} ^ {k T}}{\\sqrt {d}}\\right) \\cdot L _ {s} ^ {v} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.651, + 0.784, + 0.698 + ], + "angle": 0, + "content": "with \\( L_{s}^{k} = W^{k}L_{s} \\), \\( L_{s}^{v} = W^{v}L_{s} \\), \\( M_{i}^{q} = \\text{flatens}(W^{q}M_{i}) \\), and \\( L_{s}^{k}, L_{s}^{v} \\in R^{T_{L_{s}} \\times C_{i}} \\), \\( M_{i}^{q} \\in R^{H_{i}W_{i} \\times C_{i}} \\). \\( W^{k}, W^{v}, W^{q} \\) are implemented as a \\( 1 \\times 1 \\) convolution with \\( C_{i} \\) number of output channels." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.786, + 0.841 + ], + "angle": 0, + "content": "Deformable Feature Alignment. The TOG task presents a unique challenge in that the given image lacks the referred object, which is adverse to aforementioned attention mechanism. The input language encompasses both appearance and spatial attributes. If exclusively relying on spatial information, the model can only provide a generic localization of the region due to its limited grasp of object concepts, consequently compromising the semantic rationality of edited images (e.g. a generated mouse being larger than the existing elephant). However, when appearance attributes are incorporated into the input, the cross-attention" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.669, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ST-LDM" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.787, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.391, + 0.788, + 0.462 + ], + "angle": 0, + "content": "Fig. 3: An illustration of deformable feature alignment. Global visual feature \\( V_{i} \\), multimodal feature \\( M_{i} \\), and text feature \\( L \\) are concatenated and inputted into the Transformer layer to learn offsets and modulation scalar. The multi-head cross attention is then performed to generate the refined attention map with sampled queries in \\( M_{i} \\) and language keys and values." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.492, + 0.785, + 0.568 + ], + "angle": 0, + "content": "module faces a dilemma. It is tasked with reconciling the existing spatial attributes with the absent appearance features in the image, leading to conflicting information. This predicament significantly hampers the model's performance and generalization ability, especially in cases where the described appearance features are already present in the given image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.787, + 0.672 + ], + "angle": 0, + "content": "Inspired by deformation mechanisms [8,39,47], we propose deformable feature alignment to mitigate potential conflicts by shifting candidate queries towards regions of interest in multi-head cross attention, as illustrated in Fig. 3. The multimodal map \\( M_{i} \\) focuses on visually existing positional features by capturing the correlation between spatial-based textual and visual features, and it undergoes deformation with the alignment of object descriptions and multi-scale image features." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Offset Generation. Initially, we add a sinusoidal 1D positional encoding \\( e_{pos} \\in R^{T_L \\times C_L} \\) to the word embedding \\( L \\), updating as \\( L \\gets L + e_{pos} \\). A uniform grid of points \\( p \\in R^{H_i / \\gamma \\times W_i / \\gamma \\times 2} \\) are generated as references on \\( M_i \\), where \\( \\gamma \\) is the sampling factor. \\( M_i \\), \\( V_i \\), and \\( L \\) are normalized to the same dimension through a linear layer and then concatenated as inputs into a transformer layer to generate patch-wise multimodal offset features. Subsequently, a linear projection is applied to modify the dimension to \\( H_i / \\gamma \\times W_i / \\gamma \\times 3 \\), where the first two channels correspond to the learned 2D offsets \\( \\Delta p \\) constrained with maximum offset value \\( s_i \\) (\\( \\Delta p \\gets s_i \\tanh(\\Delta p) \\)), and the last channel is fed to a sigmoid layer to obtain modulation scalar \\( \\Delta m \\), which is introduced to exclude redundant context of which the elements lie in the range [0, 1]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "X. Xue et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.269 + ], + "angle": 0, + "content": "Multi-head Cross Attention. For reference points \\( p \\) on \\( M_{i} \\), 2D offsets \\( \\Delta p \\) and modulation scalar \\( \\Delta m \\) are generated via offset generation module. \\( \\Delta p \\) refines the spatial position according to textual descriptions, and \\( \\Delta m \\) introduces an additional dimension of freedom to exclude irrelevant context. Both of them are learned from the same transformer layer and activated with distinct modules. We designate \\( \\Delta p_{k} \\) and \\( \\Delta m_{k} \\) as the offset and scalar for the \\( k \\)-th reference point \\( p_{k} \\), where \\( k \\in \\{0,1,2,\\dots,H_iW_i / \\gamma^2\\} \\). As \\( p_{k} + \\Delta p_{k} \\) is fractional, a bilinear interpolation \\( \\phi (I,p_k) \\) is applied as" + }, + { + "type": "equation", + "bbox": [ + 0.351, + 0.279, + 0.786, + 0.311 + ], + "angle": 0, + "content": "\\[\n\\phi (I, p _ {k}) = \\sum_ {q} g \\left(p _ {k x}, q _ {x}\\right) \\cdot g \\left(p _ {k y}, q _ {y}\\right) \\cdot I (q), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.321, + 0.786, + 0.352 + ], + "angle": 0, + "content": "where \\( g(a, b) = \\max(0, 1 - |a - b|) \\) and \\( q \\) indexes the 4 integral points closest to \\( p_k \\). The sampling queries are deformed as" + }, + { + "type": "equation", + "bbox": [ + 0.385, + 0.363, + 0.786, + 0.38 + ], + "angle": 0, + "content": "\\[\nM _ {i} ^ {q} = \\operatorname {f l a t t e n} \\left(W ^ {q} \\phi \\left(M _ {i}, p + \\Delta p\\right)\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.389, + 0.788, + 0.522 + ], + "angle": 0, + "content": "The sampling visual queries are attended to the language keys with modulate attention scores \\(\\Delta m\\). A matrix completion function \\(\\psi(A)\\) is then introduced to fill in the attention score of unsampled queries and restore the dimension of the attention map to \\(H_i W_i \\times T_{L_s}\\). For each unsampled query \\(a_u\\) in \\(A\\), the attention score of \\(a_u\\) is the weighted sum of scores of all the sampled queries \\(a_s\\) in the range \\(\\mathbb{R}\\) around \\(a_u\\). The number of \\(a_s\\) in the range \\(\\mathbb{R}\\), denoted as \\(\\operatorname{card}(a_s)\\), reflects the degree of deformation near \\(a_u\\) and is introduced as a modulation factor. \\(\\operatorname{Avg}(\\operatorname{card}(a_s))\\) refers to the average number of \\(\\operatorname{card}(a_s)\\) for all the \\(a_u\\) in \\(A\\)." + }, + { + "type": "equation", + "bbox": [ + 0.366, + 0.522, + 0.786, + 0.556 + ], + "angle": 0, + "content": "\\[\n\\psi \\left(a _ {u}\\right) = \\frac {\\sum_ {a _ {s} \\in \\mathbb {R}} w _ {s} a _ {s}}{\\sum w _ {s}} \\cdot \\frac {\\operatorname {c a r d} \\left(a _ {s}\\right)}{\\operatorname {A v g} \\left(\\operatorname {c a r d} \\left(a _ {s}\\right)\\right)}, \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.434, + 0.565, + 0.786, + 0.595 + ], + "angle": 0, + "content": "\\[\nw _ {s} = \\frac {1}{\\mathrm {d} \\left(a _ {u} , a _ {s}\\right) + \\epsilon}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.6, + 0.787, + 0.631 + ], + "angle": 0, + "content": "where \\( \\mathrm{d}(a_u, a_s) \\) refers to the Euclidean distance of \\( a_u \\) and \\( a_s \\), and \\( \\epsilon \\) is a modulation constant. The output of the \\( j \\)-th attention head is formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.356, + 0.64, + 0.787, + 0.681 + ], + "angle": 0, + "content": "\\[\nZ ^ {(j)} = \\operatorname {s o f t m a x} \\left(\\psi \\left(\\frac {M _ {i} ^ {q} L _ {s} ^ {k T}}{\\sqrt {d}} \\cdot \\varDelta m\\right)\\right) \\cdot L _ {s} ^ {v}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We preserve the original structure of cross attention in the initial fusion phase and utilize spatial conditions \\( L_{s} \\) as the key and value to generate a coarse-grained attention region. In subsequent stages, we introduce deformable feature alignment with \\( s_i \\) set as \\( \\{8,4,2\\} \\) to progressively refine spatial positions during the process of deep feature extraction. This strategy can bias the attention to spatial position information rather than image local features, simultaneously with fast convergence and high computational efficiency. In the last layer, we directly extract the attention score in the multi-head cross attention and then convert it to a two-dimensional matrix averaged by the channel axis, denoted as \\( G \\), as location guidance in denoising U-Net." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.669, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ST-LDM" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.516, + 0.162 + ], + "angle": 0, + "content": "3.2 Training-free Spatial Guidance" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.784, + 0.284 + ], + "angle": 0, + "content": "Location guidance \\( G \\) encompasses spatial attention generated in pixel space, delineating the approximate boundary of the generation position. Despite the refinement conducted through deformable feature alignment which adaptively adjusts spatial localization based on multimodal information, the numerical difference of attention scores in \\( G \\) may influence the shape of the generated object, compared to a manually defined bounding box. Thus, we utilize region-wise backward guidance to avoid potential overly aggressive intervention." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.285, + 0.785, + 0.366 + ], + "angle": 0, + "content": "LDMs [33] augment underlying UNet backbone with spatial transformer, where the core module is cross-attention with \\( L_{k}^{(\\iota)} = W_{k}L_{a} \\), \\( L_{v}^{(\\iota)} = W_{v}L_{a} \\), \\( V_{q}^{(\\iota)} = \\text{ flatten}(W_{q}z_{t}) \\), where \\( \\iota \\) is the index of the relevant layer in the network, \\( L_{a} \\) refers to appearance-based text prompt, and \\( z_{t} \\) is the latent representation of the input image. The attention weight matrix is described as:" + }, + { + "type": "equation", + "bbox": [ + 0.395, + 0.379, + 0.785, + 0.428 + ], + "angle": 0, + "content": "\\[\nS ^ {(\\iota)} = \\operatorname {s o f t m a x} \\left(\\frac {V _ {q} ^ {(\\iota)} L _ {k} ^ {(\\iota) ^ {T}}}{\\sqrt {d}}\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.441, + 0.784, + 0.532 + ], + "angle": 0, + "content": "We extract the first cross-attention block in the up-sampling branch to compute backpropagation, which yields superior performance while incurring minimal computational time overhead [6]. The corresponding attention score at timestamp \\( t \\) is denoted as \\( S_{t} \\), and it has a similar compression scale with \\( G \\), which is advantageous in subsequent calculations pertaining to the energy function. The energy function is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.367, + 0.546, + 0.785, + 0.589 + ], + "angle": 0, + "content": "\\[\nE \\left(S _ {t}, G\\right) = \\left(1 - \\frac {\\sum \\left(\\theta \\left(\\hat {G} , \\beta\\right) \\odot S _ {t}\\right)}{\\sum S _ {t}}\\right) ^ {2}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.603, + 0.784, + 0.634 + ], + "angle": 0, + "content": "where \\(G\\) is converted to the same resolution as \\(S_{t}\\), denoted as \\(\\hat{G}\\). \\(\\theta (\\hat{G},\\beta)\\) is an activation function with two steps, defined as" + }, + { + "type": "equation", + "bbox": [ + 0.327, + 0.648, + 0.785, + 0.689 + ], + "angle": 0, + "content": "\\[\n\\theta (\\hat {G}, \\beta) = \\operatorname {D i l a t e} (\\hat {G}) \\leftarrow \\hat {G} _ {m n} = \\left\\{ \\begin{array}{l l} 1 & \\hat {G} _ {m n} > \\beta \\\\ 0 & \\hat {G} _ {m n} \\leq \\beta \\end{array} , \\right. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.703, + 0.784, + 0.764 + ], + "angle": 0, + "content": "where the element \\(\\hat{G}_{mn}\\) is first activated with the threshold value \\(\\beta\\), and then \\(\\hat{G}\\) is dilated to contour external maxmatrix to eliminate the numerical interference on the shape of the object. At each denoising U-Net, the latent \\(\\mathbf{z}_t\\) is updated with the gradient of Eq. 8." + }, + { + "type": "equation", + "bbox": [ + 0.363, + 0.782, + 0.785, + 0.801 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {z} _ {t} \\leftarrow \\boldsymbol {z} _ {t} - \\eta \\sqrt {\\left(1 - \\alpha_ {t}\\right) / \\alpha_ {t}} \\nabla_ {\\boldsymbol {z} _ {t}} E (S _ {t}, G), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "where \\(\\eta > 0\\) defines the strength of guidance and \\(\\alpha_{t}\\) refers to the decreasing sequence of noise schedule in stable diffusion." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "X. Xue et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.377, + 0.163 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.178, + 0.329, + 0.192 + ], + "angle": 0, + "content": "4.1 Dataset" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.203, + 0.784, + 0.246 + ], + "angle": 0, + "content": "Our primary focus of chosen datasets is the linguistic diversity of spatial conditions rather than text2img prompts since our framework is explicitly tailored for spatial guidance with the parameter layers of LDMs held frozen." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.248, + 0.785, + 0.398 + ], + "angle": 0, + "content": "VrR-VG. VrR-VG [25] is constructed from Visual Genome [21] and contains 58,983 real images with 23,375 relation pairs. Excluding positional and statistically biased relationships, It encompasses more visually relevant relationships, like \"sit on\", \"other side of\", \"fly in\", etc. We select spatial relationships and restructure linguistic expressions to align with our input format. For instance, for annotations involving attributes (fire hydrant, yellow), attributes (man, standing), and beside (man, fire hydrant), we reformulate them as \"standing man [beside yellow fire hydrant]\". Subsequently, We remove the specific object and complete images employing image inpainting techniques [11] as the input image, and the ground truth is the original position of the removed object." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.4, + 0.787, + 0.505 + ], + "angle": 0, + "content": "ComCOCO. ComCOCO [40] is derived from RefCOCO + [20] and consists of 136,495 referring expressions for 34,615 objects in 23,951 image pairs. It is a synthetic dataset and the spatial relationship is delineated in the process of manual inspection. The annotations within ComCOCO encompass both appearance and spatial attributes, and we combine them in a manner analogous to the approach detailed above. We utilize real images in ComCOCO as the input image, and the corresponding synthetic images provide the ground-truth position." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.528, + 0.457, + 0.543 + ], + "angle": 0, + "content": "4.2 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.803, + 0.644 + ], + "angle": 0, + "content": "We utilize SwinV2-T [27] with layer number \\(= \\{2, 2, 6, 2\\}\\) pretained on ADE20K [46] as the autoencoder in pixel space. Text embedding is initialized with official weights of ALBERT Base v2 [23] with 11M parameters and 12 layers. Our framework is adaptable to various latent diffusion models, and we conduct experiments by building upon a pretrained Stable-Diffusion (SD) V-1.5 [33] trained on the LAION-5B dataset [35], unless stated otherwise." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.787, + 0.839 + ], + "angle": 0, + "content": "The training process is separated into two distinct phases. First, we train patch expanding and patch partition module on OpenImages [22] with KL-regularization following [33] to ensure optimal scaling properties of the pretrained autoencoder concerning spatial dimensionality when applied to LDMs. Subsequently, we train multimodal fusion transformer layers with the backbone of Swin-T on a combined dataset comprising a modified version of VrR-VG and ComCOCO. VrR-VG contributes a diverse and balanced set of spatial relationships, albeit with potential inpainting artifacts. To mitigate this limitation, we introduce ComCOCO which employs real images as its input source. Moreover, spatial descriptions in ComCOCO predominantly revolve around the holistic perspective of an image, whereas VrR-VG is specifically directed toward capturing and describing the intricate relationships between objects within the visual content." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.669, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ST-LDM" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.298 + ], + "angle": 0, + "content": "All the experiments are implemented on 4 V100 GPUs of 32G RAM. We set the downsampling factor \\( f = 4 \\) and channel dimension \\( c = 3 \\) in the autoencoder. Backward guidance is performed during the initial 10 steps of the diffusion process and repeated 3 times at each step with the strength factor \\( \\eta \\) set to 35. The input image is scaled to \\( 480 \\times 480 \\), and the data augmentation strategy includes random cropping and flipping. The optimizer is AdamW [29] with \\( \\beta_{1} = 0.85 \\), \\( \\beta_{2} = 0.91 \\). The weight decay factor and the learning rate of the initial exponential decay are set to 0.003 and 0.00024, respectively, and the remaining parameters are randomly initialized. The model is trained for 172 rounds with the batch size set to 32 to achieve optimal performance." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.321, + 0.447, + 0.336 + ], + "angle": 0, + "content": "4.3 Qualitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.348, + 0.788, + 0.545 + ], + "angle": 0, + "content": "We apply our method on a multitude of real images from various domains, as exhibited in Fig. 1. We further compare the localization ability in Fig. 4 and generation ability in Fig. 5. Figure 4 shows the qualitative evaluation of ablative comparison of appearance and spatial properties. The deformable feature alignment module exhibits a remarkable ability to adjust spatial positions in accordance with linguistic descriptions and visual context. This dynamic adjustment imparts a direct influence on the image editing process within the latent diffusion model. As illustrated in Fig. 5, our method maintains the robust generative capabilities intrinsic to diffusion models with the preservation of semantics coherence to the original scene. The spatial constraints exerted on the denoising process have no discernible impact on the diversity and fidelity of the generated objects owing to region-wise backward guidance. More visualizations are provided in Appendix C." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.57, + 0.333, + 0.659 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.57, + 0.449, + 0.659 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.45, + 0.57, + 0.561, + 0.659 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.562, + 0.57, + 0.671, + 0.659 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.57, + 0.785, + 0.659 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.676, + 0.333, + 0.764 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.676, + 0.449, + 0.763 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.45, + 0.676, + 0.561, + 0.763 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.562, + 0.676, + 0.673, + 0.763 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.673, + 0.676, + 0.785, + 0.763 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.383, + 0.763, + 0.514, + 0.782 + ], + "angle": 0, + "content": "beef burger with ketchup on the nearby plate" + }, + { + "type": "image_caption", + "bbox": [ + 0.606, + 0.763, + 0.738, + 0.782 + ], + "angle": 0, + "content": "beef burger with ketchup on the more distant plate" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.794, + 0.787, + 0.837 + ], + "angle": 0, + "content": "Fig. 4: The first column corresponds to the input real image, and we visualize the guidance attention map \\( G \\) and corresponding edited results under different prompts with identical conditions (top) and identical prompts with different conditions (bottom)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "X. Xue et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.145, + 0.333, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.145, + 0.448, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.448, + 0.145, + 0.559, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.559, + 0.145, + 0.661, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.661, + 0.145, + 0.783, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.331, + 0.232, + 0.785, + 0.244 + ], + "angle": 0, + "content": "a blue luxury car with a streamlined body and a shiny paint job [to the right of the bus]" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.244, + 0.333, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.244, + 0.452, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.452, + 0.244, + 0.563, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.563, + 0.244, + 0.674, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.674, + 0.244, + 0.784, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.355, + 0.332, + 0.768, + 0.344 + ], + "angle": 0, + "content": "a cute dinosaur with round face, animation style [standing in front of the door]" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.354, + 0.788, + 0.384 + ], + "angle": 0, + "content": "Fig. 5: Various results produced by ST-LDM, showcasing diverse appearances while adhering to the provided text descriptions." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.412, + 0.473, + 0.429 + ], + "angle": 0, + "content": "4.4 Quantitative Comparison" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.436, + 0.788, + 0.542 + ], + "angle": 0, + "content": "Evaluation Metrics. We conduct quantitative assessments of the quality of edited images by examining two facets. (1) Object alignment: We introduce VISOR [14] to score spatial object relationships according to text descriptions and text2img metric CLIPScore [17] to measure the consistency of generated objects with the input text. (2) Image fidelity: We utilize full-reference image quality assessments FID [44] and LPIPS [45] to measure image fidelity. HPS [38] is also introduced to align with human preference." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.563, + 0.788, + 0.73 + ], + "angle": 0, + "content": "Benchmark Dataset. We construct a benchmark dataset for quantitative analysis and compare ST-LDM with related tasks. We find a distinct bias in the textual descriptions associated with the presence or absence of the specific object within the image. When the object is present in an image, the descriptions tend to emphasize distinctive features that differentiate the object from other visual entities. Conversely, when generating an object that does not exist in the original image, the focus shifts towards the detailed attributes of the object. Therefore, we randomly select 1,500 images from COCO2014 [26], and manually describe the object required to be generated in each image, simultaneously grounding the ideal position using bounding boxes. The detailed construction process is illustrated in Appendix B." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Results. We compare our method with editing models conditioned by diverse modalities in Tab. 1. Existing text-guided editing models exhibit limitations in generating new objects within complex scenes, and our model outperforms by a significant margin benefiting from explicit guidance generated from the deformable feature alignment. When compared to approaches employing forced modalities to precisely confine generated objects to designated regions," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.669, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ST-LDM" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.193 + ], + "angle": 0, + "content": "our method exhibits the capacity to edit local region of real images under the weak supervision of language, and the generated objects accurately adhere to the spatial relationships as specified in the text conditions." + }, + { + "type": "table_caption", + "bbox": [ + 0.243, + 0.209, + 0.757, + 0.224 + ], + "angle": 0, + "content": "Table 1: Comparison with editing models conditioned by diverse modalities." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.235, + 0.787, + 0.37 + ], + "angle": 0, + "content": "
ConditionModelFID↓LPIPS↓HPS↑CLIPScore↑VISOR↑
TextInstructPix2Pix [5]22.2530.9341.5860.4572.15
TextSmartEdit [18]18.4727.1147.4069.0277.86
MaskBlendedLatentDiffusion [3]12.9318.0554.3380.9688.12
MaskPaintByWord [1]9.7220.6850.2681.2387.41
BBoxBreakAScene [2]7.8916.2459.2581.1091.95
BBoxGLIGEN [24]6.2417.0962.4180.6290.53
TextST-LDM6.1615.5464.2881.9792.39
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.412, + 0.39, + 0.428 + ], + "angle": 0, + "content": "4.5 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.438, + 0.789, + 0.636 + ], + "angle": 0, + "content": "Deformable Feature Alignment. We replace the feature alignment of multimodal fusion Transformer with our deformable mechanism at different stages, resulting in a direct impact on the modification of generated regions, as depicted in Fig. 6. We introduce FID and LPIPS to assess the extent of alteration in the original image. A user study is also conducted to assess the appropriateness of object position, which serves as a more persuasive and reference-worthy metric. As delineated in Tab. 2, the integration of deformable feature alignment refines the attention region, resulting in a numerical improvement across all metrics. However, premature intervention of DFA in the initial stage interferes with spatial perception (since the input \\( M_{1} \\) of DFA will be equal to \\( V_{1} \\) if we introduce DFA in Stage1). Actually, in the initial stage of modal fusion, the model generates a preliminary spatial positioning via cross attention, and DFA is designed as a refining strategy. More illustrations are provided in Appendix A." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.661, + 0.788, + 0.718 + ], + "angle": 0, + "content": "Table 2: Ablation study of deformable feature alignment (DFA) applied at different stages. Users are required to estimate the conformance of the object's position (True or False), and the score is calculated based on the overall accuracy across all the images. IT refers to the inference time (sec/image)." + }, + { + "type": "table", + "bbox": [ + 0.273, + 0.73, + 0.733, + 0.838 + ], + "angle": 0, + "content": "
Stages w/ DFAUserScore↑FID↓LPIPS↓IT
S-1S-2S-3S-4
✓✓✓✓89.37.5016.23~7.0
✓✓✓92.46.1615.54~6.5
✓✓91.19.1420.46~6.3
✓82.515.9725.43~6.1
75.927.0334.86~6.0
" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "X. Xue et al." + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.143, + 0.787, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.246, + 0.324, + 0.757, + 0.34 + ], + "angle": 0, + "content": "Fig. 6: Ablative visualization of attention maps at diverse stages with DFA." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.365, + 0.788, + 0.472 + ], + "angle": 0, + "content": "Backward guidance. We further assess the effectiveness of backward guidance in Tab. 3. We replace guidance maps with textual conditions during the denoising process, and the notable decline indicates that linguistic information inadequately constrains the spatial placement of generated objects. We employ an activation function to eliminate potential discrepancies in attention scores, followed by dilation to mitigate the morphological interference of the attention region to the object." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.498, + 0.789, + 0.542 + ], + "angle": 0, + "content": "Table 3: Ablation study of backward guidance. Users are required to assess the matching score between the edited image and spatial descriptions. IT refers to the inference time (sec/image)." + }, + { + "type": "table", + "bbox": [ + 0.263, + 0.552, + 0.737, + 0.666 + ], + "angle": 0, + "content": "
UserScoreCLIPScoreVISORIT
uncondcond
w/o guidance58.7355.0711.5460.45~4.1
w/o activation80.8669.0137.2690.94~6.0
w/o dilation83.9870.5239.1490.37~6.2
Ours89.1181.9743.4192.39~6.5
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.701, + 0.36, + 0.718 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.842 + ], + "angle": 0, + "content": "In this paper, we explore the text-grounded object generation in real-world images. A universal framework with deformable feature alignment is proposed to hierarchically refine the attention region conforming to vision and language features. The generated guidance map can be incorporated into any latent diffusion model with training-free backward guidance. Our model addresses the inherent limitation of diffusion models in spatial perception under textual supervision, and offers valuable insights for subsequent research." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.669, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ST-LDM" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.27 + ], + "angle": 0, + "content": "This work was supported in part by the National Science and Technology Major Project under Grant 2021ZD0113202, in part by the National Natural Science Foundation of China under Grants 62171125, 61876037, in part by the innovation project of Jiangsu Province under grants BZ2023042, BY2022564. The authors would like to thank the Big Data Computing Center of Southeast University for providing the facility support on the numerical calculations in this paper." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.293, + 0.323, + 0.309 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.325, + 0.784, + 0.351 + ], + "angle": 0, + "content": "1. Andonian, A., Osmany, S., Cui, A., Park, Y., Jahanian, A., Torralba, A., Bau, D.: Paint by word" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.354, + 0.785, + 0.393 + ], + "angle": 0, + "content": "2. Avrahami, O., Aberman, K., Fried, O., Cohen-Or, D., Lischinski, D.: Break-a-scene: Extracting multiple concepts from a single image. arXiv preprint arXiv:2305.16311 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.396, + 0.785, + 0.422 + ], + "angle": 0, + "content": "3. Avrahami, O., Fried, O., Lischinski, D.: Blended latent diffusion. ACM Transactions on Graphics 42(4), 1-11 (Jul 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.424, + 0.785, + 0.477 + ], + "angle": 0, + "content": "4. Avrahami, O., Hayes, T., Gafni, O., Gupta, S., Taigman, Y., Parikh, D., Lischinski, D., Fried, O., Yin, X.: Spatext: Spatio-textual representation for controllable image generation. In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE (Jun 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.479, + 0.785, + 0.52 + ], + "angle": 0, + "content": "5. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18392-18402 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.521, + 0.785, + 0.547 + ], + "angle": 0, + "content": "6. Chen, M., Laina, I., Vedaldi, A.: Training-free layout control with cross-attention guidance (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.549, + 0.785, + 0.589 + ], + "angle": 0, + "content": "7. Couairon, G., Verbeek, J., Schwenk, H., Cord, M.: Diffedit: Diffusion-based semantic image editing with mask guidance. In: The Eleventh International Conference on Learning Representations (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.591, + 0.785, + 0.631 + ], + "angle": 0, + "content": "8. Dai, J., Qi, H., Xiong, Y., Li, Y., Zhang, G., Hu, H., Wei, Y.: Deformable convolutional networks. In: 2017 IEEE International Conference on Computer Vision (ICCV). pp. 764-773 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.633, + 0.785, + 0.672 + ], + "angle": 0, + "content": "9. Ding, H., Liu, C., Wang, S., Jiang, X.: Vision-language transformer and query generation for referring segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 16321-16330 (October 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.674, + 0.785, + 0.715 + ], + "angle": 0, + "content": "10. Ding, Z., Zhang, X., Xia, Z., Jebe, L., Tu, Z., Zhang, X.: Diffusionrig: Learning personalized priors for facial appearance editing. In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 12736-12746 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.716, + 0.785, + 0.756 + ], + "angle": 0, + "content": "11. Dong, Q., Cao, C., Fu, Y.: Incremental transformer structure enhanced image inpainting with masking positional encoding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.757, + 0.785, + 0.784 + ], + "angle": 0, + "content": "12. Fan, W.C., Chen, Y.C., Chen, D., Cheng, Y., Yuan, L., Wang, Y.C.F.: Frido: Feature pyramid diffusion for complex scene image synthesis. In: AAAI (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.785, + 0.785, + 0.84 + ], + "angle": 0, + "content": "13. Gafni, O., Polyak, A., Ashual, O., Sheynin, S., Parikh, D., Taigman, Y.: Make-a-scene: Scene-based text-to-image generation with human priors. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022. pp. 89-106. Springer Nature Switzerland, Cham (2022)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.325, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.36, + 0.127 + ], + "angle": 0, + "content": "X. Xue et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "14. Gokhale, T., Palangi, H., Nushi, B., Vineet, V., Horvitz, E., Kamar, E., Baral, C., Yang, Y.: Benchmarking spatial relationships in text-to-image generation (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.785, + 0.204 + ], + "angle": 0, + "content": "15. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.206, + 0.785, + 0.247 + ], + "angle": 0, + "content": "16. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-or, D.: Prompt-to-prompt image editing with cross-attention control. In: The Eleventh International Conference on Learning Representations (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.249, + 0.785, + 0.275 + ], + "angle": 0, + "content": "17. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: CLIPScore: a reference-free evaluation metric for image captioning. In: EMNLP (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.277, + 0.785, + 0.318 + ], + "angle": 0, + "content": "18. Huang, Y., Xie, L., Wang, X., Yuan, Z., Cun, X., Ge, Y., Zhou, J., Dong, C., Huang, R., Zhang, R., et al.: Smartedit: Exploring complex instruction-based image editing with multimodal large language models. arXiv preprint arXiv:2312.06739 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.32, + 0.785, + 0.374 + ], + "angle": 0, + "content": "19. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 6007-6017 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.376, + 0.785, + 0.417 + ], + "angle": 0, + "content": "20. Kazemzadeh, S., Ordonez, V., Matten, M., Berg, T.: Referitgame: Referring to objects in photographs of natural scenes. In: Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP). pp. 787-798 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.418, + 0.785, + 0.473 + ], + "angle": 0, + "content": "21. Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., et al.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision 123, 32-73 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.475, + 0.785, + 0.529 + ], + "angle": 0, + "content": "22. Kuznetsova, A., Rom, H., Alldrin, N., Uijlings, J., Krasin, I., Pont-Tuset, J., Kamali, S., Popov, S., Malloci, M., Kolesnikov, A., et al.: The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision 128(7), 1956-1981 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.531, + 0.785, + 0.585 + ], + "angle": 0, + "content": "23. Lan, Z., Chen, M., Goodman, S., Gimpel, K., Sharma, P., Soricut, R.: ALBERT: A lite BERT for self-supervised learning of language representations. In: 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.587, + 0.785, + 0.614 + ], + "angle": 0, + "content": "24. Li, Y., Liu, H., Wu, Q., Mu, F., Yang, J., Gao, J., Li, C., Lee, Y.J.: Gligen: Open-set grounded text-to-image generation. CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.616, + 0.785, + 0.656 + ], + "angle": 0, + "content": "25. Liang, Y., Bai, Y., Zhang, W., Qian, X., Zhu, L., Mei, T.: Vrr-vg: Refocusing visually-relevant relationships. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10403-10412 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.658, + 0.785, + 0.699 + ], + "angle": 0, + "content": "26. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: European conference on computer vision. pp. 740-755. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.7, + 0.785, + 0.755 + ], + "angle": 0, + "content": "27. Liu, Z., Hu, H., Lin, Y., Yao, Z., Xie, Z., Wei, Y., Ning, J., Cao, Y., Zhang, Z., Dong, L., et al.: Swin transformer v2: Scaling up capacity and resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12009-12019 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.757, + 0.785, + 0.811 + ], + "angle": 0, + "content": "28. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 10012-10022 (October 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "29. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: International Conference on Learning Representations (2017)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.669, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ST-LDM" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.203 + ], + "angle": 0, + "content": "30. Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Van Gool, L.: Repaint: Inpainting using denoising diffusion probabilistic models. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11451-11461 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.203, + 0.788, + 0.285 + ], + "angle": 0, + "content": "31. Nichol, A.Q., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., Mcgrew, B., Sutskever, I., Chen, M.: GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. In: Chaudhuri, K., Jegelka, S., Song, L., Szepesvari, C., Niu, G., Sabato, S. (eds.) Proceedings of the 39th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 162, pp. 16784-16804. PMLR (17-23 Jul 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.285, + 0.788, + 0.326 + ], + "angle": 0, + "content": "32. OuYang, S.R., Wang, H., Xie, S., Niu, Z., Tong, R., Chen, Y.W., Lin, L.: Slvit: Scale-wise language-guided vision transformer for referring image segmentation. In: International Joint Conference on Artificial Intelligence (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.326, + 0.788, + 0.366 + ], + "angle": 0, + "content": "33. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.366, + 0.788, + 0.42 + ], + "angle": 0, + "content": "34. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 22500-22510 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.42, + 0.788, + 0.515 + ], + "angle": 0, + "content": "35. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., Schramowski, P., Kundurthy, S., Crowson, K., Schmidt, L., Kaczmarczyk, R., Jitsev, J.: Laion-5b: An open large-scale dataset for training next generation image-text models. In: Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., Oh, A. (eds.) Advances in Neural Information Processing Systems. vol. 35, pp. 25278-25294. Curran Associates, Inc. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.515, + 0.788, + 0.57 + ], + "angle": 0, + "content": "36. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L.u., Polosukhin, I.: Attention is all you need. In: Guyon, I., Luxburg, U.V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., Garnett, R. (eds.) Advances in Neural Information Processing Systems. vol. 30. Curran Associates, Inc. (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.57, + 0.788, + 0.624 + ], + "angle": 0, + "content": "37. Wang, J., Lu, G., Xu, H., Li, Z., Xu, C., Fu, Y.: Manitrans: Entity-level text-guided image manipulation via token-wise semantic alignment and generation. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10697-10707 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.624, + 0.788, + 0.678 + ], + "angle": 0, + "content": "38. Wu, X., Sun, K., Zhu, F., Zhao, R., Li, H.: Human preference score: Better aligning text-to-image models with human preference. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2096-2105 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.678, + 0.788, + 0.718 + ], + "angle": 0, + "content": "39. Xia, Z., Pan, X., Song, S., Li, L.E., Huang, G.: Vision transformer with deformable attention. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4794-4803 (June 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.718, + 0.788, + 0.745 + ], + "angle": 0, + "content": "40. Xue, X., Wu, J., Kong, Y., Senhadji, L., Shu, H.: Rethinking referring object removal (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.745, + 0.788, + 0.786 + ], + "angle": 0, + "content": "41. Yang, L., Zhang, Z., Song, Y., Hong, S., Xu, R., Zhao, Y., Shao, Y., Zhang, W., Cui, B., Yang, M.H.: Diffusion models: A comprehensive survey of methods and applications. arXiv preprint arXiv:2209.00796 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.786, + 0.788, + 0.84 + ], + "angle": 0, + "content": "42. Yang, Z., Wang, J., Tang, Y., Chen, K., Zhao, H., Torr, P.H.: Lavo: Language-aware vision transformer for referring image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 18155-18165 (June 2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "X. Xue et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.204 + ], + "angle": 0, + "content": "43. Yang, Z., Wang, J., Gan, Z., Li, L., Lin, K., Wu, C., Duan, N., Liu, Z., Liu, C., Zeng, M., Wang, L.: Reco: Region-controlled text-to-image generation. In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.205, + 0.787, + 0.245 + ], + "angle": 0, + "content": "44. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.787, + 0.286 + ], + "angle": 0, + "content": "45. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.787, + 0.327 + ], + "angle": 0, + "content": "46. Zhou, B., Zhao, H., Puig, X., Xiao, T., Fidler, S., Barriuso, A., Torralba, A.: Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision 127, 302-321 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.787, + 0.37 + ], + "angle": 0, + "content": "47. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. In: 9th International Conference on Learning Representations, ICLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.787, + 0.438 + ], + "angle": 0, + "content": "48. Zhu, Y., Liu, H., Song, Y., Yuan, Z., Han, X., Yuan, C., Chen, Q., Wang, J.: One model to edit them all: Free-form text-driven image manipulation with semantic modulations. In: Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., Oh, A. (eds.) Advances in Neural Information Processing Systems. pp. 25146-25159. Curran Associates, Inc." + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.438 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/ST-LDM_ A Universal Framework for Text-Grounded Object Generation in Real Images/ea78a6df-2485-4a42-a86e-90a6085c52c4_origin.pdf b/2024/ST-LDM_ A Universal Framework for Text-Grounded Object Generation in Real Images/ea78a6df-2485-4a42-a86e-90a6085c52c4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6e44ad56894c1338928e48e1a2d41f8b4d5c1e74 --- /dev/null +++ b/2024/ST-LDM_ A Universal Framework for Text-Grounded Object Generation in Real Images/ea78a6df-2485-4a42-a86e-90a6085c52c4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:55234795136270b960a336f4d5a615cc135827954f4984880d0ef245deaad96a +size 8392282 diff --git a/2024/ST-LDM_ A Universal Framework for Text-Grounded Object Generation in Real Images/full.md b/2024/ST-LDM_ A Universal Framework for Text-Grounded Object Generation in Real Images/full.md new file mode 100644 index 0000000000000000000000000000000000000000..341b51eddde66799e33ebecafbb7c1a2125dc784 --- /dev/null +++ b/2024/ST-LDM_ A Universal Framework for Text-Grounded Object Generation in Real Images/full.md @@ -0,0 +1,343 @@ +# ST-LDM: A Universal Framework for Text-Grounded Object Generation in Real Images + +Xiangtian Xue $^{1,2,4}$ , Jiasong Wu $^{1,2,4\star}$ , Youyong Kong $^{1,2,4}$ , Lotfi Senhadji $^{2,3,4}$ , and Huazhong Shu $^{1,2,4}$ + +1 Key Laboratory of New Generation Artificial Intelligence Technology and Its Interdisciplinary Applications (Southeast University), Nanjing 210096, China + +$^{2}$ Jiangsu Provincial Joint International Research Laboratory of Medical Information Processing, Southeast University, Nanjing 210096, China + +3 Univ-Rennes, INSERM, LTSI-UMR 1099, Rennes F-3502, France + +4 Centre de Recherche en Information Biomédicale Sino-français (CRIBs), Rennes F-3502, France + +Abstract. We present a novel image editing scenario termed Text-grounded + +Object Generation (TOG), defined as generating a new object in the real image spatially conditioned by textual descriptions. Existing diffusion models exhibit limitations of spatial perception in complex real-world scenes, relying on additional modalities to enforce constraints, and TOG imposes heightened challenges on scene comprehension under the weak supervision of linguistic information. We propose a universal framework ST-LDM based on Swin-Transformer, which can be integrated into any latent diffusion model with training-free backward guidance. ST-LDM encompasses a global-perceptual autoencoder with adaptable compression scales and hierarchical visual features, parallel with deformable multimodal transformer to generate region-wise guidance for the subsequent denoising process. We transcend the limitation of traditional attention mechanisms that only focus on existing visual features by introducing deformable feature alignment to hierarchically refine spatial positioning fused with multi-scale visual and linguistic information. Extensive Experiments demonstrate that our model enhances the localization of attention mechanisms while preserving the generative capabilities inherent to diffusion models. + +Keywords: Instruction-based image editing. Text-grounded object generation. Deformable feature alignment. Training-free guidance + +# 1 Introduction + +Controllable text2img editing has demonstrated remarkable performance with ongoing exploration of diffusion models [10,34,48], and recent studies have been predominantly directed towards enhancing the controllability of language in editing the appearance of objects [15,19]. As for generating new objects in real images, additional modalities such as bounding boxes and keypoints have to be introduced to explicitly regulate their spatial placement [24]. + +![](images/1149bd161f7dd216ce202a159e0947ca6a9546892f74212b5b7ed75acdc3a1fc.jpg) + +![](images/017cbcec5e4a50e6f38375ab4f04b341194add502b0b99631de7fe4d8637102d.jpg) +wooden bench + +![](images/d4d218a80485e5f80497ed8ebaf2018a1c35247fba7a07595fc4f669fcf27349.jpg) + +![](images/1d2ebbf36d6c497a8c0ccf264adf4dbf3af28e4bac6bd07809545592990a0f51.jpg) +[bottom left corner] +fanny pack, black and white empty area on the table] +Fig. 1: Text-grounded object generation in real images. Our method can generate the rational object within real-world input images based on textual descriptions encompassing its visual characteristic and spatial attribute. The generated object seamlessly integrates with the original image, exhibiting a visually harmonious presence while preserving the contextual semantics of the original scene (e.g. consistency in shadow (top left), geometry (top right), illumination (bottom left), occlusion (bottom right)). + +![](images/a2bd42d721a52ecee83e9b7e5d9a9ede21a11e7b68f5613e71743bf2dbf02072.jpg) +old newspaper [on the "MAP" magazine] + +![](images/765641b48cbdc2a32f7b63d5459d384f0eb5fca5df663d14ffec3a819ad951ca.jpg) + +![](images/3cc44ec75256194edb1c517b28a576b2f018b2ab4d6d5b3c638b715ebff7d315.jpg) + +![](images/591e76cccaf19e5c83ef7a7113383cb4c304bf477429d3efbcb50b27f68f4422.jpg) +copper vase, with pattern + +![](images/de8724c805cbbd62287dfd631ee26ca23a58b9f51715bd4b7f13332fdb0409c3.jpg) +blue sedan [between bus and green car] + +![](images/cc7d9315efc524984b7d034b6975e52fe219d3c6b12ca0db8954d37fb68db29e.jpg) +woman with back turned + +![](images/9ed3a5d1d8e057d6160b072ecec2e296196e8ff8b20259764217525c99aa34fe.jpg) +board, white letters on blue [on the roof of the bus] + +The aforementioned modalities provide precise spatial constraints, but it is noteworthy that language-based instruction serves as the most fundamental and convenient interactive modality in real-world scenarios (e.g. chatting with a robot commonly using language as the sole mode of interaction, conducting similar manipulations on extensive image datasets and dynamic video sequences). Moreover, language affords a considerable degree of creative flexibility while adhering to the principles of spatial logic. This enhances the naturalness of edited images, as the model can estimate the most suitable generation location by comprehending multi-modal information. + +We propose a new image editing scenario named Text-grounded Object Generation (TOG), which aims to generate objects in real scenarios according to natural language expressions (Fig. 1). Concretely, given an image in conjunction with textual descriptions detailing visual and spatial attributes of the generated object, the task entails generating a rational object in complex scenarios. TOG embodies the dual characteristics of text2img generation and editing tasks. The generated object must adhere to textual descriptions, while simultaneously ensuring that the edited image maintains high fidelity with irrelevant areas remaining consistent with original semantics. + +Although diffusion models have demonstrated notable creative capabilities in image generation, they fall short in spatial perception and orientation, particularly when dealing with intricate real-world scenes [41]. The prevalent approaches to address this limitation involve introducing additional conditions during the denoising process to control the placement of the generated object. However, a fundamental paradox arises in our specific scenario: how to locate an object which is absent from the original image. This emphasizes the primary challenge, that is determining the optimal area for object generation under the weak supervision of texts. The objective is to ensure that the generated object aligns semantically with the contextual information contained within the image. + +In this paper, we present ST-LDM, a universal framework with hierarchical Swin-Transformer (ST) blocks [28], amenable for integration into latent diffusion models (LDM) via training-free backward guidance. ST-LDM comprises two parallel branches with similar structures, dedicated respectively to adaptable latent representation and spatial attention guidance. For latent representation, the Transformer-based autoencoder facilitates multi-resolution compression through a low-cost trainable module while preserving the formidable modeling capacity inherent to the pre-trained Swin-T. For spatial guidance, we propose deformable feature alignment in multimodal fusion Transformer to enhance the refinement of grounded regions by incorporating information from multi-scale visual and linguistic features. ST-LDM provides visually coherent location guidance in pixel space which can be integrated to the denoising U-Net of LDMs with region-wise backpropagation. + +To quantify the effectiveness of our proposed approach, we compare it with text-prompted image editing models with diverse conditions based on our constructed benchmark. Experiments demonstrate that our model enhances the localization capability of Swin-Transformer while preserving the generative proficiency inherent in diffusion models. The main contributions of our work are highlighted as follows: + +- We present a novel image editing scenario termed Text-grounded Object Generation (TOG), which is defined as generating objects in real images grounded by textual descriptions. +- We further propose a universal framework ST-LDM with deformable feature alignment to hierarchically refine spatial guidance, which can be integrated into diffusion models with training-free region-wise backpropagation. +- Quantitative comparisons with related works on our constructed benchmark and qualitative visualizations substantiate the robust editing and generative capabilities according to textual conditions. + +# 2 Related Work + +# 2.1 Spatially Controllable Object Generation + +Recent studies have employed mandatory modalities to exert control over the spatial distribution of generated objects [6,12,43]. Make-A-Scene [13] introduces implicit conditioning over optionally controlled scene tokens derived from segmentation maps with fixed labels to control the layout of the image. SpaText [4] further proposes a spatial-textual representation that encapsulates the semantic attributes and structural characteristics of each segment. GLIGEN [24] extends grounding conditions to bounding boxes, reference images, keypoints, etc., which are injected into new trainable layers of the denoising process via gated self-attention. Nevertheless, these modalities are cumbersome when applied to large-scale operations involving numerous images and videos. The utilization of text-based spatial position descriptions serves to mitigate rigid constraints, thereby affording models a heightened capacity for comprehending visual information. + +# 2.2 Diffusion-Based Semantic Image Editing + +Image editing emphasizes modifying the visual entity consistent with textual descriptions while preserving the entity-irrelevant regions. Diffusion models can be easily adapted for inpainting when given a mask input since they iteratively refine an image starting from random noise [30,31]. Recent approaches endeavor to obviate the necessity of additional inputs such as image masks or additional views of the object [6,16,37]. DiffEdit [7] automatically generates a mask accentuating regions of the input image necessitating modification, by contrasting predictions of a diffusion model conditioned on different text prompts. Imagic [19] produces a text embedding that aligns with both the input image and the target text while fine-tuning the diffusion model to capture the image-specific appearance. These attempts are experimented on images with obvious foreground objects and simple semantics. However, when applied to complex scenes characterized by a multiplicity of objects and spatial relationships, their efficacy in localizing objects exhibits a discernible diminishment. + +# 2.3 Vision-and-Language Transformer + +The initial phase of our proposed method involves locating visual regions of interest guided by language, and Transformer [36] has demonstrated noteworthy performance at the interaction of vision and language. Ding et al. [9] introduce Transformer and multi-head attention to construct an encoder-decoder attention mechanism architecture. LAVT [42] achieves better cross-modal alignments through the early fusion of linguistic and visual features in a vision Transformer encoder network. SLViT [32] employs the variance of cross-modal correlations between adjacent stages to identify regions of high uncertainty and refine features of these regions with complementing information from multiple stages. While Transformer facilitates the integration of linguistic and visual modalities through cross-attention mechanisms, it encounters challenges stemming from the absence of target objects within the image, which is a distinctive characteristic of the TOG task. + +# 3 Method + +In this section, we present our proposed framework, denoted as ST-LDM, and the overall pipeline is depicted in Fig. 2. + +In pixel space (Sec. 3.1), visual feature extraction and multi-modal localization are concurrently performed during the downsampling process, leveraging Swin Transformer [28] as the backbone. The feature extraction module acts as an autoencoder to concurrently reconstruct the image into a lower-dimensional latent space while furnishing hierarchical visual features for spatial localization. In parallel, the multi-modal Transformer module leverages the interplay between image and language to provide positional guidance for the denoising process. We innovatively introduce deformable feature alignment to dynamically adjust spatial constraints, ensuring that the generated object aligns with semantic context + +![](images/ae8d2f13f143c733564f7063e8b6254fea02db3001d623d519ee2d066f500634.jpg) +Fig. 2: The basic framework of the proposed ST-LDM. Given a real image and text descriptions, we initially encode them into visual features $V_{i}$ and text embedding $L$ , where $L$ comprises appearance prompt $L_{a}$ and spatial conditions $L_{s}$ . We employ the Swin-T structure to hierarchically integrate $L$ into $V_{i}$ , where the multimodal map $M_{i}$ is deformed by learning multi-scale visual and semantic relationships. The generated spatial guidance attention map $G$ and textual prompt $L_{a}$ are integrated into latent diffusion models with training-free guidance. The autoencoder-decoder in pixel space and the latent diffusion model in latent space retain original structure and parameters, and these components are frozen throughout the training process. + +of the input image. The congruence in interaction between two branches is optimized through the application of a similar network structure (SwinT) to both of the aforementioned modules. + +In latent space (Sec. 3.2), appearance-based text prompt is separately mapped into denoising U-Net via cross attention. Subsequently, a region-wise backpropagation scheme is introduced to iteratively update the activations in the network to match the desired layout specified in pixel space. The fundamental structure and parameters of diffusion models remain constant, with trainable parameter layers exclusively concentrated in the pixel space. Consequently, the proposed ST-LDM emerges as a universal framework adaptable to a spectrum of pretrained LDMs. + +# 3.1 Hierarchical Feature Extraction + +Autoencoder with Swin-Transformer. LDMs [33] leverages an individually pretrained autoencoding model for perceptual image compression, but its quantitative performance exhibits significant variability across different downsampling factors due to the inherent instability of its encoding structure. Recently, Swin-Transformer has emerged as a generic visual feature extractor, owing to its robust modeling capabilities [28]. It initially splits an input RGB image into + +non-overlapping patches with a $4 \times 4$ size, and the resolution of subsequent layers' output continually decreases accordingly. However, excessive information compression will induce distortion of noisy images in the denoising process. + +To address this issue, we add a trainable patch expanding layer to convert the selected feature map to the specified size. The original structure of SwinTransformer is retained for hierarchical feature extraction, and an intermediate feature map $V_{i}$ is selected and upsampled to high-dimensional resolution via patch expanding layer. It can increase the resolution by a factor of 4 while simultaneously reducing the dimension to a configurable value $(H_{i}\times W_{i}\times C_{i}\to 4H_{i}\times 4W_{i}\times C_{t},C_{i}\geq 16C_{t})$ . The layer consists of two operations, a linear layer to convert the feature dimension $(H_{i}\times W_{i}\times C_{i}\to H_{i}\times W_{i}\times 16C_{t})$ , and a rearrange operation to expand the resolution to $4\times$ and reduce the dimension to one-sixteenth $(H_{i}\times W_{i}\times 16C_{t}\to 4H_{i}\times 4W_{i}\times C_{t})$ . + +The extracted feature map is optional within $\{V_2, V_3, V_4\}$ contingent upon the depth of compression perception. As the initial patch size in Swin-Transformer is $4 \times 4$ , the resulting output resolution for patch expansion remains fixed at $4 \times$ relative to its initial dimensions. Consequently, $\{V_2, V_3, V_4\}$ corresponds to down-sampling factors $f = \{2, 4, 8\}$ . Due to the varied channel configurations across model variants, the output channel dimension remains adaptable. The encoding-decoding process in pixel space exhibits a symmetrical structure, wherein the linear layer in patch expanding and patch partition undergoes low-cost training following [33] to cater to specific downstream tasks while retaining the formidable modeling capabilities of pretrained models. + +Multimodal Transformer. We jointly embed language and vision features during visual encoding with Swin-Transformer to generate semantically coherent positional guidance for object generation. Consistent with previous works [42], we densely integrate spatial-based language features $L_{s} \in R^{T_{L_{s}} \times C_{L_{s}}}$ into image features $M_{i} \in R^{H_{i} \times W_{i} \times C_{i}}$ via cross attention implementing + +$$ +\operatorname {A t t e n t i o n} \left(M _ {i} ^ {q}, L _ {s} ^ {k}, L _ {s} ^ {v}\right) = \operatorname {s o f t m a x} \left(\frac {M _ {i} ^ {q} L _ {s} ^ {k T}}{\sqrt {d}}\right) \cdot L _ {s} ^ {v} \tag {1} +$$ + +with $L_{s}^{k} = W^{k}L_{s}$ , $L_{s}^{v} = W^{v}L_{s}$ , $M_{i}^{q} = \text{flatens}(W^{q}M_{i})$ , and $L_{s}^{k}, L_{s}^{v} \in R^{T_{L_{s}} \times C_{i}}$ , $M_{i}^{q} \in R^{H_{i}W_{i} \times C_{i}}$ . $W^{k}, W^{v}, W^{q}$ are implemented as a $1 \times 1$ convolution with $C_{i}$ number of output channels. + +Deformable Feature Alignment. The TOG task presents a unique challenge in that the given image lacks the referred object, which is adverse to aforementioned attention mechanism. The input language encompasses both appearance and spatial attributes. If exclusively relying on spatial information, the model can only provide a generic localization of the region due to its limited grasp of object concepts, consequently compromising the semantic rationality of edited images (e.g. a generated mouse being larger than the existing elephant). However, when appearance attributes are incorporated into the input, the cross-attention + +![](images/02720657c05cfc12c4222f6b1a9070a19be0a25695f6d194554d9d0ca168a4a0.jpg) +Fig. 3: An illustration of deformable feature alignment. Global visual feature $V_{i}$ , multimodal feature $M_{i}$ , and text feature $L$ are concatenated and inputted into the Transformer layer to learn offsets and modulation scalar. The multi-head cross attention is then performed to generate the refined attention map with sampled queries in $M_{i}$ and language keys and values. + +module faces a dilemma. It is tasked with reconciling the existing spatial attributes with the absent appearance features in the image, leading to conflicting information. This predicament significantly hampers the model's performance and generalization ability, especially in cases where the described appearance features are already present in the given image. + +Inspired by deformation mechanisms [8,39,47], we propose deformable feature alignment to mitigate potential conflicts by shifting candidate queries towards regions of interest in multi-head cross attention, as illustrated in Fig. 3. The multimodal map $M_{i}$ focuses on visually existing positional features by capturing the correlation between spatial-based textual and visual features, and it undergoes deformation with the alignment of object descriptions and multi-scale image features. + +Offset Generation. Initially, we add a sinusoidal 1D positional encoding $e_{pos} \in R^{T_L \times C_L}$ to the word embedding $L$ , updating as $L \gets L + e_{pos}$ . A uniform grid of points $p \in R^{H_i / \gamma \times W_i / \gamma \times 2}$ are generated as references on $M_i$ , where $\gamma$ is the sampling factor. $M_i$ , $V_i$ , and $L$ are normalized to the same dimension through a linear layer and then concatenated as inputs into a transformer layer to generate patch-wise multimodal offset features. Subsequently, a linear projection is applied to modify the dimension to $H_i / \gamma \times W_i / \gamma \times 3$ , where the first two channels correspond to the learned 2D offsets $\Delta p$ constrained with maximum offset value $s_i$ ( $\Delta p \gets s_i \tanh(\Delta p)$ ), and the last channel is fed to a sigmoid layer to obtain modulation scalar $\Delta m$ , which is introduced to exclude redundant context of which the elements lie in the range [0, 1]. + +Multi-head Cross Attention. For reference points $p$ on $M_{i}$ , 2D offsets $\Delta p$ and modulation scalar $\Delta m$ are generated via offset generation module. $\Delta p$ refines the spatial position according to textual descriptions, and $\Delta m$ introduces an additional dimension of freedom to exclude irrelevant context. Both of them are learned from the same transformer layer and activated with distinct modules. We designate $\Delta p_{k}$ and $\Delta m_{k}$ as the offset and scalar for the $k$ -th reference point $p_{k}$ , where $k \in \{0,1,2,\dots,H_iW_i / \gamma^2\}$ . As $p_{k} + \Delta p_{k}$ is fractional, a bilinear interpolation $\phi (I,p_k)$ is applied as + +$$ +\phi (I, p _ {k}) = \sum_ {q} g \left(p _ {k x}, q _ {x}\right) \cdot g \left(p _ {k y}, q _ {y}\right) \cdot I (q), \tag {2} +$$ + +where $g(a, b) = \max(0, 1 - |a - b|)$ and $q$ indexes the 4 integral points closest to $p_k$ . The sampling queries are deformed as + +$$ +M _ {i} ^ {q} = \operatorname {f l a t t e n} \left(W ^ {q} \phi \left(M _ {i}, p + \Delta p\right)\right). \tag {3} +$$ + +The sampling visual queries are attended to the language keys with modulate attention scores $\Delta m$ . A matrix completion function $\psi(A)$ is then introduced to fill in the attention score of unsampled queries and restore the dimension of the attention map to $H_i W_i \times T_{L_s}$ . For each unsampled query $a_u$ in $A$ , the attention score of $a_u$ is the weighted sum of scores of all the sampled queries $a_s$ in the range $\mathbb{R}$ around $a_u$ . The number of $a_s$ in the range $\mathbb{R}$ , denoted as $\operatorname{card}(a_s)$ , reflects the degree of deformation near $a_u$ and is introduced as a modulation factor. $\operatorname{Avg}(\operatorname{card}(a_s))$ refers to the average number of $\operatorname{card}(a_s)$ for all the $a_u$ in $A$ . + +$$ +\psi \left(a _ {u}\right) = \frac {\sum_ {a _ {s} \in \mathbb {R}} w _ {s} a _ {s}}{\sum w _ {s}} \cdot \frac {\operatorname {c a r d} \left(a _ {s}\right)}{\operatorname {A v g} \left(\operatorname {c a r d} \left(a _ {s}\right)\right)}, \tag {4} +$$ + +$$ +w _ {s} = \frac {1}{\mathrm {d} \left(a _ {u} , a _ {s}\right) + \epsilon}, \tag {5} +$$ + +where $\mathrm{d}(a_u, a_s)$ refers to the Euclidean distance of $a_u$ and $a_s$ , and $\epsilon$ is a modulation constant. The output of the $j$ -th attention head is formulated as + +$$ +Z ^ {(j)} = \operatorname {s o f t m a x} \left(\psi \left(\frac {M _ {i} ^ {q} L _ {s} ^ {k T}}{\sqrt {d}} \cdot \varDelta m\right)\right) \cdot L _ {s} ^ {v}. \tag {6} +$$ + +We preserve the original structure of cross attention in the initial fusion phase and utilize spatial conditions $L_{s}$ as the key and value to generate a coarse-grained attention region. In subsequent stages, we introduce deformable feature alignment with $s_i$ set as $\{8,4,2\}$ to progressively refine spatial positions during the process of deep feature extraction. This strategy can bias the attention to spatial position information rather than image local features, simultaneously with fast convergence and high computational efficiency. In the last layer, we directly extract the attention score in the multi-head cross attention and then convert it to a two-dimensional matrix averaged by the channel axis, denoted as $G$ , as location guidance in denoising U-Net. + +# 3.2 Training-free Spatial Guidance + +Location guidance $G$ encompasses spatial attention generated in pixel space, delineating the approximate boundary of the generation position. Despite the refinement conducted through deformable feature alignment which adaptively adjusts spatial localization based on multimodal information, the numerical difference of attention scores in $G$ may influence the shape of the generated object, compared to a manually defined bounding box. Thus, we utilize region-wise backward guidance to avoid potential overly aggressive intervention. + +LDMs [33] augment underlying UNet backbone with spatial transformer, where the core module is cross-attention with $L_{k}^{(\iota)} = W_{k}L_{a}$ , $L_{v}^{(\iota)} = W_{v}L_{a}$ , $V_{q}^{(\iota)} = \text{ flatten}(W_{q}z_{t})$ , where $\iota$ is the index of the relevant layer in the network, $L_{a}$ refers to appearance-based text prompt, and $z_{t}$ is the latent representation of the input image. The attention weight matrix is described as: + +$$ +S ^ {(\iota)} = \operatorname {s o f t m a x} \left(\frac {V _ {q} ^ {(\iota)} L _ {k} ^ {(\iota) ^ {T}}}{\sqrt {d}}\right). \tag {7} +$$ + +We extract the first cross-attention block in the up-sampling branch to compute backpropagation, which yields superior performance while incurring minimal computational time overhead [6]. The corresponding attention score at timestamp $t$ is denoted as $S_{t}$ , and it has a similar compression scale with $G$ , which is advantageous in subsequent calculations pertaining to the energy function. The energy function is defined as + +$$ +E \left(S _ {t}, G\right) = \left(1 - \frac {\sum \left(\theta \left(\hat {G} , \beta\right) \odot S _ {t}\right)}{\sum S _ {t}}\right) ^ {2}, \tag {8} +$$ + +where $G$ is converted to the same resolution as $S_{t}$ , denoted as $\hat{G}$ . $\theta (\hat{G},\beta)$ is an activation function with two steps, defined as + +$$ +\theta (\hat {G}, \beta) = \operatorname {D i l a t e} (\hat {G}) \leftarrow \hat {G} _ {m n} = \left\{ \begin{array}{l l} 1 & \hat {G} _ {m n} > \beta \\ 0 & \hat {G} _ {m n} \leq \beta \end{array} , \right. \tag {9} +$$ + +where the element $\hat{G}_{mn}$ is first activated with the threshold value $\beta$ , and then $\hat{G}$ is dilated to contour external maxmatrix to eliminate the numerical interference on the shape of the object. At each denoising U-Net, the latent $\mathbf{z}_t$ is updated with the gradient of Eq. 8. + +$$ +\boldsymbol {z} _ {t} \leftarrow \boldsymbol {z} _ {t} - \eta \sqrt {\left(1 - \alpha_ {t}\right) / \alpha_ {t}} \nabla_ {\boldsymbol {z} _ {t}} E (S _ {t}, G), \tag {10} +$$ + +where $\eta > 0$ defines the strength of guidance and $\alpha_{t}$ refers to the decreasing sequence of noise schedule in stable diffusion. + +# 4 Experiments + +# 4.1 Dataset + +Our primary focus of chosen datasets is the linguistic diversity of spatial conditions rather than text2img prompts since our framework is explicitly tailored for spatial guidance with the parameter layers of LDMs held frozen. + +VrR-VG. VrR-VG [25] is constructed from Visual Genome [21] and contains 58,983 real images with 23,375 relation pairs. Excluding positional and statistically biased relationships, It encompasses more visually relevant relationships, like "sit on", "other side of", "fly in", etc. We select spatial relationships and restructure linguistic expressions to align with our input format. For instance, for annotations involving attributes (fire hydrant, yellow), attributes (man, standing), and beside (man, fire hydrant), we reformulate them as "standing man [beside yellow fire hydrant]". Subsequently, We remove the specific object and complete images employing image inpainting techniques [11] as the input image, and the ground truth is the original position of the removed object. + +ComCOCO. ComCOCO [40] is derived from RefCOCO + [20] and consists of 136,495 referring expressions for 34,615 objects in 23,951 image pairs. It is a synthetic dataset and the spatial relationship is delineated in the process of manual inspection. The annotations within ComCOCO encompass both appearance and spatial attributes, and we combine them in a manner analogous to the approach detailed above. We utilize real images in ComCOCO as the input image, and the corresponding synthetic images provide the ground-truth position. + +# 4.2 Implementation Details + +We utilize SwinV2-T [27] with layer number $= \{2, 2, 6, 2\}$ pretained on ADE20K [46] as the autoencoder in pixel space. Text embedding is initialized with official weights of ALBERT Base v2 [23] with 11M parameters and 12 layers. Our framework is adaptable to various latent diffusion models, and we conduct experiments by building upon a pretrained Stable-Diffusion (SD) V-1.5 [33] trained on the LAION-5B dataset [35], unless stated otherwise. + +The training process is separated into two distinct phases. First, we train patch expanding and patch partition module on OpenImages [22] with KL-regularization following [33] to ensure optimal scaling properties of the pretrained autoencoder concerning spatial dimensionality when applied to LDMs. Subsequently, we train multimodal fusion transformer layers with the backbone of Swin-T on a combined dataset comprising a modified version of VrR-VG and ComCOCO. VrR-VG contributes a diverse and balanced set of spatial relationships, albeit with potential inpainting artifacts. To mitigate this limitation, we introduce ComCOCO which employs real images as its input source. Moreover, spatial descriptions in ComCOCO predominantly revolve around the holistic perspective of an image, whereas VrR-VG is specifically directed toward capturing and describing the intricate relationships between objects within the visual content. + +All the experiments are implemented on 4 V100 GPUs of 32G RAM. We set the downsampling factor $f = 4$ and channel dimension $c = 3$ in the autoencoder. Backward guidance is performed during the initial 10 steps of the diffusion process and repeated 3 times at each step with the strength factor $\eta$ set to 35. The input image is scaled to $480 \times 480$ , and the data augmentation strategy includes random cropping and flipping. The optimizer is AdamW [29] with $\beta_{1} = 0.85$ , $\beta_{2} = 0.91$ . The weight decay factor and the learning rate of the initial exponential decay are set to 0.003 and 0.00024, respectively, and the remaining parameters are randomly initialized. The model is trained for 172 rounds with the batch size set to 32 to achieve optimal performance. + +# 4.3 Qualitative Evaluation + +We apply our method on a multitude of real images from various domains, as exhibited in Fig. 1. We further compare the localization ability in Fig. 4 and generation ability in Fig. 5. Figure 4 shows the qualitative evaluation of ablative comparison of appearance and spatial properties. The deformable feature alignment module exhibits a remarkable ability to adjust spatial positions in accordance with linguistic descriptions and visual context. This dynamic adjustment imparts a direct influence on the image editing process within the latent diffusion model. As illustrated in Fig. 5, our method maintains the robust generative capabilities intrinsic to diffusion models with the preservation of semantics coherence to the original scene. The spatial constraints exerted on the denoising process have no discernible impact on the diversity and fidelity of the generated objects owing to region-wise backward guidance. More visualizations are provided in Appendix C. + +![](images/f4fd301c86199d473189b65968815287a43fbf96ad595419b264779e3231db17.jpg) + +![](images/d3744cd16794620b204ef968e68e95c07e6956281e5710ecdffdc7440e46dacf.jpg) + +![](images/0dfee3714abbba89db6a23c00153c90d134102a743dc4b80b41409249f1e1f81.jpg) + +![](images/5ae5482ab908094ecb016182e6ea3bae4768d36501ad60c5cd9cca038cfb96b4.jpg) + +![](images/1ed9ee9a8cbf41e1aa65360fc5f9d28be27d9bf03cd8a9310e55999205b27c81.jpg) + +![](images/391fdfc2858759940eebe066417b7eddfb1fa3a648cc68b7934b12dd52754a3b.jpg) +Fig. 4: The first column corresponds to the input real image, and we visualize the guidance attention map $G$ and corresponding edited results under different prompts with identical conditions (top) and identical prompts with different conditions (bottom). + +![](images/1eac37b2a3679326598843ce777434bd8cc877db5748a21f7c3e5d140450640c.jpg) +beef burger with ketchup on the nearby plate + +![](images/f389749c01cd816b21580115703a97ecb24e71940739e3ff5f9eee65772986be.jpg) + +![](images/1bcba744b19c479e865617d342e9376515cbef8380c04e5fb89655efa7f058d5.jpg) +beef burger with ketchup on the more distant plate + +![](images/366a04d5421400277d6ecfef5c9276b33ce1b8c9e9455c1cc0e4e591715dafd1.jpg) + +![](images/6e5fccfce4d3076a0de15f78d1957168ba8b78e481ad38890ad2059d1d9e6ca6.jpg) +a blue luxury car with a streamlined body and a shiny paint job [to the right of the bus] + +![](images/77ccf27d14991410aa3e7aec9558bda056903c8db74cbfc5fcd3a9c586f22d3a.jpg) + +![](images/0ae96e1d50b3b3f75d4c40a823d1111285245d51ce38ab432f3fc27f79dcc2b6.jpg) + +![](images/9c46e48637fc8f6dbd1f600d2c0dae642202a1a071f3d24e8bdb58c1a5eac9e8.jpg) + +![](images/69fcacd1b12763f8cdc47c93b3ef7313fa524a3ad9b07409a6ff523853f7c17a.jpg) + +![](images/3486a468f8e0436bf52fb9291cc80c8d2394ac2279409044e53ebcee890deb90.jpg) +Fig. 5: Various results produced by ST-LDM, showcasing diverse appearances while adhering to the provided text descriptions. + +![](images/ce6a1228682cbc9b9014a7f81cddf3d00c857f6739c1ce428f78cf9d01a79e5b.jpg) +a cute dinosaur with round face, animation style [standing in front of the door] + +![](images/b699748339e401ea9bc10a70fc17193429cb25d2a6ab95cfcc188059ae962d42.jpg) + +![](images/38a635fd2cfaed842fcd5e0c829069487bbeab5c3cc5b7b80aa5b6029c9e18cd.jpg) + +![](images/378271a66860bc610209cca6db9a86be9373695c327efc750372ae0ce3a0d192.jpg) + +# 4.4 Quantitative Comparison + +Evaluation Metrics. We conduct quantitative assessments of the quality of edited images by examining two facets. (1) Object alignment: We introduce VISOR [14] to score spatial object relationships according to text descriptions and text2img metric CLIPScore [17] to measure the consistency of generated objects with the input text. (2) Image fidelity: We utilize full-reference image quality assessments FID [44] and LPIPS [45] to measure image fidelity. HPS [38] is also introduced to align with human preference. + +Benchmark Dataset. We construct a benchmark dataset for quantitative analysis and compare ST-LDM with related tasks. We find a distinct bias in the textual descriptions associated with the presence or absence of the specific object within the image. When the object is present in an image, the descriptions tend to emphasize distinctive features that differentiate the object from other visual entities. Conversely, when generating an object that does not exist in the original image, the focus shifts towards the detailed attributes of the object. Therefore, we randomly select 1,500 images from COCO2014 [26], and manually describe the object required to be generated in each image, simultaneously grounding the ideal position using bounding boxes. The detailed construction process is illustrated in Appendix B. + +Results. We compare our method with editing models conditioned by diverse modalities in Tab. 1. Existing text-guided editing models exhibit limitations in generating new objects within complex scenes, and our model outperforms by a significant margin benefiting from explicit guidance generated from the deformable feature alignment. When compared to approaches employing forced modalities to precisely confine generated objects to designated regions, + +our method exhibits the capacity to edit local region of real images under the weak supervision of language, and the generated objects accurately adhere to the spatial relationships as specified in the text conditions. + +Table 1: Comparison with editing models conditioned by diverse modalities. + +
ConditionModelFID↓LPIPS↓HPS↑CLIPScore↑VISOR↑
TextInstructPix2Pix [5]22.2530.9341.5860.4572.15
TextSmartEdit [18]18.4727.1147.4069.0277.86
MaskBlendedLatentDiffusion [3]12.9318.0554.3380.9688.12
MaskPaintByWord [1]9.7220.6850.2681.2387.41
BBoxBreakAScene [2]7.8916.2459.2581.1091.95
BBoxGLIGEN [24]6.2417.0962.4180.6290.53
TextST-LDM6.1615.5464.2881.9792.39
+ +# 4.5 Ablation Study + +Deformable Feature Alignment. We replace the feature alignment of multimodal fusion Transformer with our deformable mechanism at different stages, resulting in a direct impact on the modification of generated regions, as depicted in Fig. 6. We introduce FID and LPIPS to assess the extent of alteration in the original image. A user study is also conducted to assess the appropriateness of object position, which serves as a more persuasive and reference-worthy metric. As delineated in Tab. 2, the integration of deformable feature alignment refines the attention region, resulting in a numerical improvement across all metrics. However, premature intervention of DFA in the initial stage interferes with spatial perception (since the input $M_{1}$ of DFA will be equal to $V_{1}$ if we introduce DFA in Stage1). Actually, in the initial stage of modal fusion, the model generates a preliminary spatial positioning via cross attention, and DFA is designed as a refining strategy. More illustrations are provided in Appendix A. + +Table 2: Ablation study of deformable feature alignment (DFA) applied at different stages. Users are required to estimate the conformance of the object's position (True or False), and the score is calculated based on the overall accuracy across all the images. IT refers to the inference time (sec/image). + +
Stages w/ DFAUserScore↑FID↓LPIPS↓IT
S-1S-2S-3S-4
✓✓✓✓89.37.5016.23~7.0
✓✓✓92.46.1615.54~6.5
✓✓91.19.1420.46~6.3
✓82.515.9725.43~6.1
75.927.0334.86~6.0
+ +![](images/98a01e03cf7a630875dd1e9b44d1ca816c2a95f6d147a35be63d8f2e2e3f88ee.jpg) +Fig. 6: Ablative visualization of attention maps at diverse stages with DFA. + +Backward guidance. We further assess the effectiveness of backward guidance in Tab. 3. We replace guidance maps with textual conditions during the denoising process, and the notable decline indicates that linguistic information inadequately constrains the spatial placement of generated objects. We employ an activation function to eliminate potential discrepancies in attention scores, followed by dilation to mitigate the morphological interference of the attention region to the object. + +Table 3: Ablation study of backward guidance. Users are required to assess the matching score between the edited image and spatial descriptions. IT refers to the inference time (sec/image). + +
UserScoreCLIPScoreVISORIT
uncondcond
w/o guidance58.7355.0711.5460.45~4.1
w/o activation80.8669.0137.2690.94~6.0
w/o dilation83.9870.5239.1490.37~6.2
Ours89.1181.9743.4192.39~6.5
+ +# 5 Conclusion + +In this paper, we explore the text-grounded object generation in real-world images. A universal framework with deformable feature alignment is proposed to hierarchically refine the attention region conforming to vision and language features. The generated guidance map can be incorporated into any latent diffusion model with training-free backward guidance. Our model addresses the inherent limitation of diffusion models in spatial perception under textual supervision, and offers valuable insights for subsequent research. + +# Acknowledgements + +This work was supported in part by the National Science and Technology Major Project under Grant 2021ZD0113202, in part by the National Natural Science Foundation of China under Grants 62171125, 61876037, in part by the innovation project of Jiangsu Province under grants BZ2023042, BY2022564. The authors would like to thank the Big Data Computing Center of Southeast University for providing the facility support on the numerical calculations in this paper. + +# References + +1. Andonian, A., Osmany, S., Cui, A., Park, Y., Jahanian, A., Torralba, A., Bau, D.: Paint by word +2. Avrahami, O., Aberman, K., Fried, O., Cohen-Or, D., Lischinski, D.: Break-a-scene: Extracting multiple concepts from a single image. arXiv preprint arXiv:2305.16311 (2023) +3. Avrahami, O., Fried, O., Lischinski, D.: Blended latent diffusion. ACM Transactions on Graphics 42(4), 1-11 (Jul 2023) +4. Avrahami, O., Hayes, T., Gafni, O., Gupta, S., Taigman, Y., Parikh, D., Lischinski, D., Fried, O., Yin, X.: Spatext: Spatio-textual representation for controllable image generation. In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE (Jun 2023) +5. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18392-18402 (2023) +6. Chen, M., Laina, I., Vedaldi, A.: Training-free layout control with cross-attention guidance (2023) +7. Couairon, G., Verbeek, J., Schwenk, H., Cord, M.: Diffedit: Diffusion-based semantic image editing with mask guidance. In: The Eleventh International Conference on Learning Representations (2023) +8. Dai, J., Qi, H., Xiong, Y., Li, Y., Zhang, G., Hu, H., Wei, Y.: Deformable convolutional networks. In: 2017 IEEE International Conference on Computer Vision (ICCV). pp. 764-773 (2017) +9. Ding, H., Liu, C., Wang, S., Jiang, X.: Vision-language transformer and query generation for referring segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 16321-16330 (October 2021) +10. Ding, Z., Zhang, X., Xia, Z., Jebe, L., Tu, Z., Zhang, X.: Diffusionrig: Learning personalized priors for facial appearance editing. In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 12736-12746 (2023) +11. Dong, Q., Cao, C., Fu, Y.: Incremental transformer structure enhanced image inpainting with masking positional encoding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022) +12. Fan, W.C., Chen, Y.C., Chen, D., Cheng, Y., Yuan, L., Wang, Y.C.F.: Frido: Feature pyramid diffusion for complex scene image synthesis. In: AAAI (2023) +13. Gafni, O., Polyak, A., Ashual, O., Sheynin, S., Parikh, D., Taigman, Y.: Make-a-scene: Scene-based text-to-image generation with human priors. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022. pp. 89-106. Springer Nature Switzerland, Cham (2022) + +14. Gokhale, T., Palangi, H., Nushi, B., Vineet, V., Horvitz, E., Kamar, E., Baral, C., Yang, Y.: Benchmarking spatial relationships in text-to-image generation (2022) +15. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control (2022) +16. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-or, D.: Prompt-to-prompt image editing with cross-attention control. In: The Eleventh International Conference on Learning Representations (2023) +17. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: CLIPScore: a reference-free evaluation metric for image captioning. In: EMNLP (2021) +18. Huang, Y., Xie, L., Wang, X., Yuan, Z., Cun, X., Ge, Y., Zhou, J., Dong, C., Huang, R., Zhang, R., et al.: Smartedit: Exploring complex instruction-based image editing with multimodal large language models. arXiv preprint arXiv:2312.06739 (2023) +19. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 6007-6017 (June 2023) +20. Kazemzadeh, S., Ordonez, V., Matten, M., Berg, T.: Referitgame: Referring to objects in photographs of natural scenes. In: Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP). pp. 787-798 (2014) +21. Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., et al.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision 123, 32-73 (2017) +22. Kuznetsova, A., Rom, H., Alldrin, N., Uijlings, J., Krasin, I., Pont-Tuset, J., Kamali, S., Popov, S., Malloci, M., Kolesnikov, A., et al.: The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision 128(7), 1956-1981 (2020) +23. Lan, Z., Chen, M., Goodman, S., Gimpel, K., Sharma, P., Soricut, R.: ALBERT: A lite BERT for self-supervised learning of language representations. In: 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net (2020) +24. Li, Y., Liu, H., Wu, Q., Mu, F., Yang, J., Gao, J., Li, C., Lee, Y.J.: Gligen: Open-set grounded text-to-image generation. CVPR (2023) +25. Liang, Y., Bai, Y., Zhang, W., Qian, X., Zhu, L., Mei, T.: Vrr-vg: Refocusing visually-relevant relationships. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10403-10412 (2019) +26. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: European conference on computer vision. pp. 740-755. Springer (2014) +27. Liu, Z., Hu, H., Lin, Y., Yao, Z., Xie, Z., Wei, Y., Ning, J., Cao, Y., Zhang, Z., Dong, L., et al.: Swin transformer v2: Scaling up capacity and resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12009-12019 (2022) +28. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 10012-10022 (October 2021) +29. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: International Conference on Learning Representations (2017) + +30. Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Van Gool, L.: Repaint: Inpainting using denoising diffusion probabilistic models. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11451-11461 (2022) +31. Nichol, A.Q., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., Mcgrew, B., Sutskever, I., Chen, M.: GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. In: Chaudhuri, K., Jegelka, S., Song, L., Szepesvari, C., Niu, G., Sabato, S. (eds.) Proceedings of the 39th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 162, pp. 16784-16804. PMLR (17-23 Jul 2022) +32. OuYang, S.R., Wang, H., Xie, S., Niu, Z., Tong, R., Chen, Y.W., Lin, L.: Slvit: Scale-wise language-guided vision transformer for referring image segmentation. In: International Joint Conference on Artificial Intelligence (2023) +33. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022) +34. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 22500-22510 (June 2023) +35. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., Schramowski, P., Kundurthy, S., Crowson, K., Schmidt, L., Kaczmarczyk, R., Jitsev, J.: Laion-5b: An open large-scale dataset for training next generation image-text models. In: Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., Oh, A. (eds.) Advances in Neural Information Processing Systems. vol. 35, pp. 25278-25294. Curran Associates, Inc. (2022) +36. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L.u., Polosukhin, I.: Attention is all you need. In: Guyon, I., Luxburg, U.V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., Garnett, R. (eds.) Advances in Neural Information Processing Systems. vol. 30. Curran Associates, Inc. (2017) +37. Wang, J., Lu, G., Xu, H., Li, Z., Xu, C., Fu, Y.: Manitrans: Entity-level text-guided image manipulation via token-wise semantic alignment and generation. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10697-10707 (2022) +38. Wu, X., Sun, K., Zhu, F., Zhao, R., Li, H.: Human preference score: Better aligning text-to-image models with human preference. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2096-2105 (October 2023) +39. Xia, Z., Pan, X., Song, S., Li, L.E., Huang, G.: Vision transformer with deformable attention. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4794-4803 (June 2022) +40. Xue, X., Wu, J., Kong, Y., Senhadji, L., Shu, H.: Rethinking referring object removal (2024) +41. Yang, L., Zhang, Z., Song, Y., Hong, S., Xu, R., Zhao, Y., Shao, Y., Zhang, W., Cui, B., Yang, M.H.: Diffusion models: A comprehensive survey of methods and applications. arXiv preprint arXiv:2209.00796 (2022) +42. Yang, Z., Wang, J., Tang, Y., Chen, K., Zhao, H., Torr, P.H.: Lavo: Language-aware vision transformer for referring image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 18155-18165 (June 2022) + +43. Yang, Z., Wang, J., Gan, Z., Li, L., Lin, K., Wu, C., Duan, N., Liu, Z., Liu, C., Zeng, M., Wang, L.: Reco: Region-controlled text-to-image generation. In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023) +44. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018) +45. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2018) +46. Zhou, B., Zhao, H., Puig, X., Xiao, T., Fidler, S., Barriuso, A., Torralba, A.: Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision 127, 302-321 (2019) +47. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. In: 9th International Conference on Learning Representations, ICLR (2021) +48. Zhu, Y., Liu, H., Song, Y., Yuan, Z., Han, X., Yuan, C., Chen, Q., Wang, J.: One model to edit them all: Free-form text-driven image manipulation with semantic modulations. In: Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., Oh, A. (eds.) Advances in Neural Information Processing Systems. pp. 25146-25159. Curran Associates, Inc. \ No newline at end of file diff --git a/2024/ST-LDM_ A Universal Framework for Text-Grounded Object Generation in Real Images/images.zip b/2024/ST-LDM_ A Universal Framework for Text-Grounded Object Generation in Real Images/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a48405efcd20d513c199e8cd39ee8f06e372af43 --- /dev/null +++ b/2024/ST-LDM_ A Universal Framework for Text-Grounded Object Generation in Real Images/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:724892abe6e79c8ced8c4b0ad9ca4c79e2eaa40d73ddf4aa21fb7be102472c27 +size 596065 diff --git a/2024/ST-LDM_ A Universal Framework for Text-Grounded Object Generation in Real Images/layout.json b/2024/ST-LDM_ A Universal Framework for Text-Grounded Object Generation in Real Images/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..387aaf19aae1acae17e8a561549e7d7e615e3b3b --- /dev/null +++ b/2024/ST-LDM_ A Universal Framework for Text-Grounded Object Generation in Real Images/layout.json @@ -0,0 +1,10881 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 133, + 111, + 482, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 111, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 133, + 111, + 482, + 148 + ], + "type": "text", + "content": "ST-LDM: A Universal Framework for Text-Grounded Object Generation in Real Images" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 138, + 167, + 475, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 167, + 475, + 193 + ], + "spans": [ + { + "bbox": [ + 138, + 167, + 475, + 193 + ], + "type": "text", + "content": "Xiangtian Xue" + }, + { + "bbox": [ + 138, + 167, + 475, + 193 + ], + "type": "inline_equation", + "content": "^{1,2,4}" + }, + { + "bbox": [ + 138, + 167, + 475, + 193 + ], + "type": "text", + "content": ", Jiasong Wu" + }, + { + "bbox": [ + 138, + 167, + 475, + 193 + ], + "type": "inline_equation", + "content": "^{1,2,4\\star}" + }, + { + "bbox": [ + 138, + 167, + 475, + 193 + ], + "type": "text", + "content": ", Youyong Kong" + }, + { + "bbox": [ + 138, + 167, + 475, + 193 + ], + "type": "inline_equation", + "content": "^{1,2,4}" + }, + { + "bbox": [ + 138, + 167, + 475, + 193 + ], + "type": "text", + "content": ", Lotfi Senhadji" + }, + { + "bbox": [ + 138, + 167, + 475, + 193 + ], + "type": "inline_equation", + "content": "^{2,3,4}" + }, + { + "bbox": [ + 138, + 167, + 475, + 193 + ], + "type": "text", + "content": ", and Huazhong Shu" + }, + { + "bbox": [ + 138, + 167, + 475, + 193 + ], + "type": "inline_equation", + "content": "^{1,2,4}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 146, + 201, + 467, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 201, + 467, + 224 + ], + "spans": [ + { + "bbox": [ + 146, + 201, + 467, + 224 + ], + "type": "text", + "content": "1 Key Laboratory of New Generation Artificial Intelligence Technology and Its Interdisciplinary Applications (Southeast University), Nanjing 210096, China" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 224, + 479, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 224, + 479, + 245 + ], + "spans": [ + { + "bbox": [ + 133, + 224, + 479, + 245 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 133, + 224, + 479, + 245 + ], + "type": "text", + "content": " Jiangsu Provincial Joint International Research Laboratory of Medical Information Processing, Southeast University, Nanjing 210096, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 169, + 245, + 444, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 245, + 444, + 256 + ], + "spans": [ + { + "bbox": [ + 169, + 245, + 444, + 256 + ], + "type": "text", + "content": "3 Univ-Rennes, INSERM, LTSI-UMR 1099, Rennes F-3502, France" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 140, + 257, + 474, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 257, + 474, + 278 + ], + "spans": [ + { + "bbox": [ + 140, + 257, + 474, + 278 + ], + "type": "text", + "content": "4 Centre de Recherche en Information Biomédicale Sino-français (CRIBs), Rennes F-3502, France" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 160, + 299, + 468, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 299, + 468, + 309 + ], + "spans": [ + { + "bbox": [ + 160, + 299, + 468, + 309 + ], + "type": "text", + "content": "Abstract. We present a novel image editing scenario termed Text-grounded" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 160, + 309, + 452, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 309, + 452, + 506 + ], + "spans": [ + { + "bbox": [ + 160, + 309, + 452, + 506 + ], + "type": "text", + "content": "Object Generation (TOG), defined as generating a new object in the real image spatially conditioned by textual descriptions. Existing diffusion models exhibit limitations of spatial perception in complex real-world scenes, relying on additional modalities to enforce constraints, and TOG imposes heightened challenges on scene comprehension under the weak supervision of linguistic information. We propose a universal framework ST-LDM based on Swin-Transformer, which can be integrated into any latent diffusion model with training-free backward guidance. ST-LDM encompasses a global-perceptual autoencoder with adaptable compression scales and hierarchical visual features, parallel with deformable multimodal transformer to generate region-wise guidance for the subsequent denoising process. We transcend the limitation of traditional attention mechanisms that only focus on existing visual features by introducing deformable feature alignment to hierarchically refine spatial positioning fused with multi-scale visual and linguistic information. Extensive Experiments demonstrate that our model enhances the localization of attention mechanisms while preserving the generative capabilities inherent to diffusion models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 159, + 517, + 453, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 517, + 453, + 539 + ], + "spans": [ + { + "bbox": [ + 159, + 517, + 453, + 539 + ], + "type": "text", + "content": "Keywords: Instruction-based image editing. Text-grounded object generation. Deformable feature alignment. Training-free guidance" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 555, + 230, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 230, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 230, + 567 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 576, + 482, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 576, + 482, + 649 + ], + "spans": [ + { + "bbox": [ + 130, + 576, + 482, + 649 + ], + "type": "text", + "content": "Controllable text2img editing has demonstrated remarkable performance with ongoing exploration of diffusion models [10,34,48], and recent studies have been predominantly directed towards enhancing the controllability of language in editing the appearance of objects [15,19]. As for generating new objects in real images, additional modalities such as bounding boxes and keypoints have to be introduced to explicitly regulate their spatial placement [24]." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 654, + 306, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 654, + 306, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 654, + 306, + 665 + ], + "type": "text", + "content": "* Corresponding author: jswu@seu.edu.cn" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 190, + 171 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 190, + 171 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 190, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 190, + 171 + ], + "type": "image", + "image_path": "1149bd161f7dd216ce202a159e0947ca6a9546892f74212b5b7ed75acdc3a1fc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 191, + 114, + 247, + 171 + ], + "blocks": [ + { + "bbox": [ + 191, + 114, + 247, + 171 + ], + "lines": [ + { + "bbox": [ + 191, + 114, + 247, + 171 + ], + "spans": [ + { + "bbox": [ + 191, + 114, + 247, + 171 + ], + "type": "image", + "image_path": "017cbcec5e4a50e6f38375ab4f04b341194add502b0b99631de7fe4d8637102d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 200, + 171, + 235, + 176 + ], + "lines": [ + { + "bbox": [ + 200, + 171, + 235, + 176 + ], + "spans": [ + { + "bbox": [ + 200, + 171, + 235, + 176 + ], + "type": "text", + "content": "wooden bench" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 248, + 114, + 304, + 171 + ], + "blocks": [ + { + "bbox": [ + 248, + 114, + 304, + 171 + ], + "lines": [ + { + "bbox": [ + 248, + 114, + 304, + 171 + ], + "spans": [ + { + "bbox": [ + 248, + 114, + 304, + 171 + ], + "type": "image", + "image_path": "d4d218a80485e5f80497ed8ebaf2018a1c35247fba7a07595fc4f669fcf27349.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 191, + 180, + 247, + 236 + ], + "blocks": [ + { + "bbox": [ + 200, + 176, + 237, + 180 + ], + "lines": [ + { + "bbox": [ + 200, + 176, + 237, + 180 + ], + "spans": [ + { + "bbox": [ + 200, + 176, + 237, + 180 + ], + "type": "text", + "content": "[bottom left corner]" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 191, + 180, + 247, + 236 + ], + "lines": [ + { + "bbox": [ + 191, + 180, + 247, + 236 + ], + "spans": [ + { + "bbox": [ + 191, + 180, + 247, + 236 + ], + "type": "image", + "image_path": "1d2ebbf36d6c497a8c0ccf264adf4dbf3af28e4bac6bd07809545592990a0f51.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 192, + 236, + 246, + 245 + ], + "lines": [ + { + "bbox": [ + 192, + 236, + 246, + 245 + ], + "spans": [ + { + "bbox": [ + 192, + 236, + 246, + 245 + ], + "type": "text", + "content": "fanny pack, black and white empty area on the table]" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 255, + 482, + 322 + ], + "lines": [ + { + "bbox": [ + 132, + 255, + 482, + 322 + ], + "spans": [ + { + "bbox": [ + 132, + 255, + 482, + 322 + ], + "type": "text", + "content": "Fig. 1: Text-grounded object generation in real images. Our method can generate the rational object within real-world input images based on textual descriptions encompassing its visual characteristic and spatial attribute. The generated object seamlessly integrates with the original image, exhibiting a visually harmonious presence while preserving the contextual semantics of the original scene (e.g. consistency in shadow (top left), geometry (top right), illumination (bottom left), occlusion (bottom right))." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 249, + 180, + 304, + 236 + ], + "blocks": [ + { + "bbox": [ + 249, + 180, + 304, + 236 + ], + "lines": [ + { + "bbox": [ + 249, + 180, + 304, + 236 + ], + "spans": [ + { + "bbox": [ + 249, + 180, + 304, + 236 + ], + "type": "image", + "image_path": "a2bd42d721a52ecee83e9b7e5d9a9ede21a11e7b68f5613e71743bf2dbf02072.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 251, + 236, + 302, + 245 + ], + "lines": [ + { + "bbox": [ + 251, + 236, + 302, + 245 + ], + "spans": [ + { + "bbox": [ + 251, + 236, + 302, + 245 + ], + "type": "text", + "content": "old newspaper [on the \"MAP\" magazine]" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 114, + 364, + 171 + ], + "blocks": [ + { + "bbox": [ + 307, + 114, + 364, + 171 + ], + "lines": [ + { + "bbox": [ + 307, + 114, + 364, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 114, + 364, + 171 + ], + "type": "image", + "image_path": "765641b48cbdc2a32f7b63d5459d384f0eb5fca5df663d14ffec3a819ad951ca.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 308, + 171, + 364, + 176 + ], + "lines": [ + { + "bbox": [ + 308, + 171, + 364, + 176 + ], + "spans": [ + { + "bbox": [ + 308, + 171, + 364, + 176 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 308, + 176, + 364, + 234 + ], + "blocks": [ + { + "bbox": [ + 308, + 176, + 364, + 234 + ], + "lines": [ + { + "bbox": [ + 308, + 176, + 364, + 234 + ], + "spans": [ + { + "bbox": [ + 308, + 176, + 364, + 234 + ], + "type": "image", + "image_path": "3cc44ec75256194edb1c517b28a576b2f018b2ab4d6d5b3c638b715ebff7d315.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 365, + 114, + 421, + 171 + ], + "blocks": [ + { + "bbox": [ + 365, + 114, + 421, + 171 + ], + "lines": [ + { + "bbox": [ + 365, + 114, + 421, + 171 + ], + "spans": [ + { + "bbox": [ + 365, + 114, + 421, + 171 + ], + "type": "image", + "image_path": "591e76cccaf19e5c83ef7a7113383cb4c304bf477429d3efbcb50b27f68f4422.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 368, + 171, + 416, + 176 + ], + "lines": [ + { + "bbox": [ + 368, + 171, + 416, + 176 + ], + "spans": [ + { + "bbox": [ + 368, + 171, + 416, + 176 + ], + "type": "text", + "content": "copper vase, with pattern" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 365, + 176, + 421, + 234 + ], + "blocks": [ + { + "bbox": [ + 365, + 176, + 421, + 234 + ], + "lines": [ + { + "bbox": [ + 365, + 176, + 421, + 234 + ], + "spans": [ + { + "bbox": [ + 365, + 176, + 421, + 234 + ], + "type": "image", + "image_path": "de8724c805cbbd62287dfd631ee26ca23a58b9f51715bd4b7f13332fdb0409c3.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 367, + 234, + 421, + 245 + ], + "lines": [ + { + "bbox": [ + 367, + 234, + 421, + 245 + ], + "spans": [ + { + "bbox": [ + 367, + 234, + 421, + 245 + ], + "type": "text", + "content": "blue sedan [between bus and green car]" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 422, + 114, + 479, + 171 + ], + "blocks": [ + { + "bbox": [ + 422, + 114, + 479, + 171 + ], + "lines": [ + { + "bbox": [ + 422, + 114, + 479, + 171 + ], + "spans": [ + { + "bbox": [ + 422, + 114, + 479, + 171 + ], + "type": "image", + "image_path": "cc7d9315efc524984b7d034b6975e52fe219d3c6b12ca0db8954d37fb68db29e.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 423, + 171, + 477, + 176 + ], + "lines": [ + { + "bbox": [ + 423, + 171, + 477, + 176 + ], + "spans": [ + { + "bbox": [ + 423, + 171, + 477, + 176 + ], + "type": "text", + "content": "woman with back turned" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 423, + 176, + 477, + 234 + ], + "blocks": [ + { + "bbox": [ + 423, + 176, + 477, + 234 + ], + "lines": [ + { + "bbox": [ + 423, + 176, + 477, + 234 + ], + "spans": [ + { + "bbox": [ + 423, + 176, + 477, + 234 + ], + "type": "image", + "image_path": "9ed3a5d1d8e057d6160b072ecec2e296196e8ff8b20259764217525c99aa34fe.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 427, + 234, + 480, + 244 + ], + "lines": [ + { + "bbox": [ + 427, + 234, + 480, + 244 + ], + "spans": [ + { + "bbox": [ + 427, + 234, + 480, + 244 + ], + "type": "text", + "content": "board, white letters on blue [on the roof of the bus]" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 331, + 480, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 480, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 480, + 437 + ], + "type": "text", + "content": "The aforementioned modalities provide precise spatial constraints, but it is noteworthy that language-based instruction serves as the most fundamental and convenient interactive modality in real-world scenarios (e.g. chatting with a robot commonly using language as the sole mode of interaction, conducting similar manipulations on extensive image datasets and dynamic video sequences). Moreover, language affords a considerable degree of creative flexibility while adhering to the principles of spatial logic. This enhances the naturalness of edited images, as the model can estimate the most suitable generation location by comprehending multi-modal information." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 130, + 438, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 482, + 545 + ], + "type": "text", + "content": "We propose a new image editing scenario named Text-grounded Object Generation (TOG), which aims to generate objects in real scenarios according to natural language expressions (Fig. 1). Concretely, given an image in conjunction with textual descriptions detailing visual and spatial attributes of the generated object, the task entails generating a rational object in complex scenarios. TOG embodies the dual characteristics of text2img generation and editing tasks. The generated object must adhere to textual descriptions, while simultaneously ensuring that the edited image maintains high fidelity with irrelevant areas remaining consistent with original semantics." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": "Although diffusion models have demonstrated notable creative capabilities in image generation, they fall short in spatial perception and orientation, particularly when dealing with intricate real-world scenes [41]. The prevalent approaches to address this limitation involve introducing additional conditions during the denoising process to control the placement of the generated object. However, a fundamental paradox arises in our specific scenario: how to locate an object which is absent from the original image. This emphasizes the primary challenge, that is determining the optimal area for object generation under the weak supervision of texts. The objective is to ensure that the generated object aligns semantically with the contextual information contained within the image." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "X. Xue et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": "In this paper, we present ST-LDM, a universal framework with hierarchical Swin-Transformer (ST) blocks [28], amenable for integration into latent diffusion models (LDM) via training-free backward guidance. ST-LDM comprises two parallel branches with similar structures, dedicated respectively to adaptable latent representation and spatial attention guidance. For latent representation, the Transformer-based autoencoder facilitates multi-resolution compression through a low-cost trainable module while preserving the formidable modeling capacity inherent to the pre-trained Swin-T. For spatial guidance, we propose deformable feature alignment in multimodal fusion Transformer to enhance the refinement of grounded regions by incorporating information from multi-scale visual and linguistic features. ST-LDM provides visually coherent location guidance in pixel space which can be integrated to the denoising U-Net of LDMs with region-wise backpropagation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 272, + 482, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 272, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 482, + 344 + ], + "type": "text", + "content": "To quantify the effectiveness of our proposed approach, we compare it with text-prompted image editing models with diverse conditions based on our constructed benchmark. Experiments demonstrate that our model enhances the localization capability of Swin-Transformer while preserving the generative proficiency inherent in diffusion models. The main contributions of our work are highlighted as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 348, + 480, + 453 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 138, + 348, + 479, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 348, + 479, + 383 + ], + "spans": [ + { + "bbox": [ + 138, + 348, + 479, + 383 + ], + "type": "text", + "content": "- We present a novel image editing scenario termed Text-grounded Object Generation (TOG), which is defined as generating objects in real images grounded by textual descriptions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 383, + 480, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 383, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 138, + 383, + 480, + 418 + ], + "type": "text", + "content": "- We further propose a universal framework ST-LDM with deformable feature alignment to hierarchically refine spatial guidance, which can be integrated into diffusion models with training-free region-wise backpropagation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 419, + 480, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 419, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 138, + 419, + 480, + 453 + ], + "type": "text", + "content": "- Quantitative comparisons with related works on our constructed benchmark and qualitative visualizations substantiate the robust editing and generative capabilities according to textual conditions." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 470, + 237, + 482 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 470, + 237, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 470, + 237, + 482 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 492, + 367, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 492, + 367, + 506 + ], + "spans": [ + { + "bbox": [ + 132, + 492, + 367, + 506 + ], + "type": "text", + "content": "2.1 Spatially Controllable Object Generation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "text", + "content": "Recent studies have employed mandatory modalities to exert control over the spatial distribution of generated objects [6,12,43]. Make-A-Scene [13] introduces implicit conditioning over optionally controlled scene tokens derived from segmentation maps with fixed labels to control the layout of the image. SpaText [4] further proposes a spatial-textual representation that encapsulates the semantic attributes and structural characteristics of each segment. GLIGEN [24] extends grounding conditions to bounding boxes, reference images, keypoints, etc., which are injected into new trainable layers of the denoising process via gated self-attention. Nevertheless, these modalities are cumbersome when applied to large-scale operations involving numerous images and videos. The utilization of text-based spatial position descriptions serves to mitigate rigid constraints, thereby affording models a heightened capacity for comprehending visual information." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 409, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 409, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 409, + 91, + 447, + 100 + ], + "type": "text", + "content": "ST-LDM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 363, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 363, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 363, + 129 + ], + "type": "text", + "content": "2.2 Diffusion-Based Semantic Image Editing" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 134, + 482, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 134, + 482, + 302 + ], + "spans": [ + { + "bbox": [ + 130, + 134, + 482, + 302 + ], + "type": "text", + "content": "Image editing emphasizes modifying the visual entity consistent with textual descriptions while preserving the entity-irrelevant regions. Diffusion models can be easily adapted for inpainting when given a mask input since they iteratively refine an image starting from random noise [30,31]. Recent approaches endeavor to obviate the necessity of additional inputs such as image masks or additional views of the object [6,16,37]. DiffEdit [7] automatically generates a mask accentuating regions of the input image necessitating modification, by contrasting predictions of a diffusion model conditioned on different text prompts. Imagic [19] produces a text embedding that aligns with both the input image and the target text while fine-tuning the diffusion model to capture the image-specific appearance. These attempts are experimented on images with obvious foreground objects and simple semantics. However, when applied to complex scenes characterized by a multiplicity of objects and spatial relationships, their efficacy in localizing objects exhibits a discernible diminishment." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 319, + 331, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 319, + 331, + 331 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 331, + 331 + ], + "type": "text", + "content": "2.3 Vision-and-Language Transformer" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 337, + 482, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 337, + 482, + 492 + ], + "spans": [ + { + "bbox": [ + 130, + 337, + 482, + 492 + ], + "type": "text", + "content": "The initial phase of our proposed method involves locating visual regions of interest guided by language, and Transformer [36] has demonstrated noteworthy performance at the interaction of vision and language. Ding et al. [9] introduce Transformer and multi-head attention to construct an encoder-decoder attention mechanism architecture. LAVT [42] achieves better cross-modal alignments through the early fusion of linguistic and visual features in a vision Transformer encoder network. SLViT [32] employs the variance of cross-modal correlations between adjacent stages to identify regions of high uncertainty and refine features of these regions with complementing information from multiple stages. While Transformer facilitates the integration of linguistic and visual modalities through cross-attention mechanisms, it encounters challenges stemming from the absence of target objects within the image, which is a distinctive characteristic of the TOG task." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 510, + 202, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 510, + 202, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 510, + 202, + 522 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 481, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 481, + 557 + ], + "type": "text", + "content": "In this section, we present our proposed framework, denoted as ST-LDM, and the overall pipeline is depicted in Fig. 2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": "In pixel space (Sec. 3.1), visual feature extraction and multi-modal localization are concurrently performed during the downsampling process, leveraging Swin Transformer [28] as the backbone. The feature extraction module acts as an autoencoder to concurrently reconstruct the image into a lower-dimensional latent space while furnishing hierarchical visual features for spatial localization. In parallel, the multi-modal Transformer module leverages the interplay between image and language to provide positional guidance for the denoising process. We innovatively introduce deformable feature alignment to dynamically adjust spatial constraints, ensuring that the generated object aligns with semantic context" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "X. Xue et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 115, + 479, + 289 + ], + "blocks": [ + { + "bbox": [ + 135, + 115, + 479, + 289 + ], + "lines": [ + { + "bbox": [ + 135, + 115, + 479, + 289 + ], + "spans": [ + { + "bbox": [ + 135, + 115, + 479, + 289 + ], + "type": "image", + "image_path": "ae8d2f13f143c733564f7063e8b6254fea02db3001d623d519ee2d066f500634.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 299, + 482, + 399 + ], + "lines": [ + { + "bbox": [ + 130, + 299, + 482, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 299, + 482, + 399 + ], + "type": "text", + "content": "Fig. 2: The basic framework of the proposed ST-LDM. Given a real image and text descriptions, we initially encode them into visual features " + }, + { + "bbox": [ + 130, + 299, + 482, + 399 + ], + "type": "inline_equation", + "content": "V_{i}" + }, + { + "bbox": [ + 130, + 299, + 482, + 399 + ], + "type": "text", + "content": " and text embedding " + }, + { + "bbox": [ + 130, + 299, + 482, + 399 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 299, + 482, + 399 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 299, + 482, + 399 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 299, + 482, + 399 + ], + "type": "text", + "content": " comprises appearance prompt " + }, + { + "bbox": [ + 130, + 299, + 482, + 399 + ], + "type": "inline_equation", + "content": "L_{a}" + }, + { + "bbox": [ + 130, + 299, + 482, + 399 + ], + "type": "text", + "content": " and spatial conditions " + }, + { + "bbox": [ + 130, + 299, + 482, + 399 + ], + "type": "inline_equation", + "content": "L_{s}" + }, + { + "bbox": [ + 130, + 299, + 482, + 399 + ], + "type": "text", + "content": ". We employ the Swin-T structure to hierarchically integrate " + }, + { + "bbox": [ + 130, + 299, + 482, + 399 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 299, + 482, + 399 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 130, + 299, + 482, + 399 + ], + "type": "inline_equation", + "content": "V_{i}" + }, + { + "bbox": [ + 130, + 299, + 482, + 399 + ], + "type": "text", + "content": ", where the multimodal map " + }, + { + "bbox": [ + 130, + 299, + 482, + 399 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 130, + 299, + 482, + 399 + ], + "type": "text", + "content": " is deformed by learning multi-scale visual and semantic relationships. The generated spatial guidance attention map " + }, + { + "bbox": [ + 130, + 299, + 482, + 399 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 299, + 482, + 399 + ], + "type": "text", + "content": " and textual prompt " + }, + { + "bbox": [ + 130, + 299, + 482, + 399 + ], + "type": "inline_equation", + "content": "L_{a}" + }, + { + "bbox": [ + 130, + 299, + 482, + 399 + ], + "type": "text", + "content": " are integrated into latent diffusion models with training-free guidance. The autoencoder-decoder in pixel space and the latent diffusion model in latent space retain original structure and parameters, and these components are frozen throughout the training process." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 422, + 480, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 480, + 456 + ], + "type": "text", + "content": "of the input image. The congruence in interaction between two branches is optimized through the application of a similar network structure (SwinT) to both of the aforementioned modules." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 458, + 482, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 458, + 482, + 554 + ], + "spans": [ + { + "bbox": [ + 130, + 458, + 482, + 554 + ], + "type": "text", + "content": "In latent space (Sec. 3.2), appearance-based text prompt is separately mapped into denoising U-Net via cross attention. Subsequently, a region-wise backpropagation scheme is introduced to iteratively update the activations in the network to match the desired layout specified in pixel space. The fundamental structure and parameters of diffusion models remain constant, with trainable parameter layers exclusively concentrated in the pixel space. Consequently, the proposed ST-LDM emerges as a universal framework adaptable to a spectrum of pretrained LDMs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 572, + 320, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 572, + 320, + 583 + ], + "spans": [ + { + "bbox": [ + 131, + 572, + 320, + 583 + ], + "type": "text", + "content": "3.1 Hierarchical Feature Extraction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Autoencoder with Swin-Transformer. LDMs [33] leverages an individually pretrained autoencoding model for perceptual image compression, but its quantitative performance exhibits significant variability across different downsampling factors due to the inherent instability of its encoding structure. Recently, Swin-Transformer has emerged as a generic visual feature extractor, owing to its robust modeling capabilities [28]. It initially splits an input RGB image into" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 409, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 409, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 409, + 91, + 447, + 100 + ], + "type": "text", + "content": "ST-LDM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "non-overlapping patches with a " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "4 \\times 4" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " size, and the resolution of subsequent layers' output continually decreases accordingly. However, excessive information compression will induce distortion of noisy images in the denoising process." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 480, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 480, + 271 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 480, + 271 + ], + "type": "text", + "content": "To address this issue, we add a trainable patch expanding layer to convert the selected feature map to the specified size. The original structure of SwinTransformer is retained for hierarchical feature extraction, and an intermediate feature map " + }, + { + "bbox": [ + 130, + 152, + 480, + 271 + ], + "type": "inline_equation", + "content": "V_{i}" + }, + { + "bbox": [ + 130, + 152, + 480, + 271 + ], + "type": "text", + "content": " is selected and upsampled to high-dimensional resolution via patch expanding layer. It can increase the resolution by a factor of 4 while simultaneously reducing the dimension to a configurable value " + }, + { + "bbox": [ + 130, + 152, + 480, + 271 + ], + "type": "inline_equation", + "content": "(H_{i}\\times W_{i}\\times C_{i}\\to 4H_{i}\\times 4W_{i}\\times C_{t},C_{i}\\geq 16C_{t})" + }, + { + "bbox": [ + 130, + 152, + 480, + 271 + ], + "type": "text", + "content": ". The layer consists of two operations, a linear layer to convert the feature dimension " + }, + { + "bbox": [ + 130, + 152, + 480, + 271 + ], + "type": "inline_equation", + "content": "(H_{i}\\times W_{i}\\times C_{i}\\to H_{i}\\times W_{i}\\times 16C_{t})" + }, + { + "bbox": [ + 130, + 152, + 480, + 271 + ], + "type": "text", + "content": ", and a rearrange operation to expand the resolution to " + }, + { + "bbox": [ + 130, + 152, + 480, + 271 + ], + "type": "inline_equation", + "content": "4\\times" + }, + { + "bbox": [ + 130, + 152, + 480, + 271 + ], + "type": "text", + "content": " and reduce the dimension to one-sixteenth " + }, + { + "bbox": [ + 130, + 152, + 480, + 271 + ], + "type": "inline_equation", + "content": "(H_{i}\\times W_{i}\\times 16C_{t}\\to 4H_{i}\\times 4W_{i}\\times C_{t})" + }, + { + "bbox": [ + 130, + 152, + 480, + 271 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 271, + 481, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 271, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 271, + 481, + 392 + ], + "type": "text", + "content": "The extracted feature map is optional within " + }, + { + "bbox": [ + 130, + 271, + 481, + 392 + ], + "type": "inline_equation", + "content": "\\{V_2, V_3, V_4\\}" + }, + { + "bbox": [ + 130, + 271, + 481, + 392 + ], + "type": "text", + "content": " contingent upon the depth of compression perception. As the initial patch size in Swin-Transformer is " + }, + { + "bbox": [ + 130, + 271, + 481, + 392 + ], + "type": "inline_equation", + "content": "4 \\times 4" + }, + { + "bbox": [ + 130, + 271, + 481, + 392 + ], + "type": "text", + "content": ", the resulting output resolution for patch expansion remains fixed at " + }, + { + "bbox": [ + 130, + 271, + 481, + 392 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 130, + 271, + 481, + 392 + ], + "type": "text", + "content": " relative to its initial dimensions. Consequently, " + }, + { + "bbox": [ + 130, + 271, + 481, + 392 + ], + "type": "inline_equation", + "content": "\\{V_2, V_3, V_4\\}" + }, + { + "bbox": [ + 130, + 271, + 481, + 392 + ], + "type": "text", + "content": " corresponds to down-sampling factors " + }, + { + "bbox": [ + 130, + 271, + 481, + 392 + ], + "type": "inline_equation", + "content": "f = \\{2, 4, 8\\}" + }, + { + "bbox": [ + 130, + 271, + 481, + 392 + ], + "type": "text", + "content": ". Due to the varied channel configurations across model variants, the output channel dimension remains adaptable. The encoding-decoding process in pixel space exhibits a symmetrical structure, wherein the linear layer in patch expanding and patch partition undergoes low-cost training following [33] to cater to specific downstream tasks while retaining the formidable modeling capabilities of pretrained models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 408, + 480, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 408, + 480, + 469 + ], + "spans": [ + { + "bbox": [ + 130, + 408, + 480, + 469 + ], + "type": "text", + "content": "Multimodal Transformer. We jointly embed language and vision features during visual encoding with Swin-Transformer to generate semantically coherent positional guidance for object generation. Consistent with previous works [42], we densely integrate spatial-based language features " + }, + { + "bbox": [ + 130, + 408, + 480, + 469 + ], + "type": "inline_equation", + "content": "L_{s} \\in R^{T_{L_{s}} \\times C_{L_{s}}}" + }, + { + "bbox": [ + 130, + 408, + 480, + 469 + ], + "type": "text", + "content": " into image features " + }, + { + "bbox": [ + 130, + 408, + 480, + 469 + ], + "type": "inline_equation", + "content": "M_{i} \\in R^{H_{i} \\times W_{i} \\times C_{i}}" + }, + { + "bbox": [ + 130, + 408, + 480, + 469 + ], + "type": "text", + "content": " via cross attention implementing" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 200, + 476, + 480, + 508 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 476, + 480, + 508 + ], + "spans": [ + { + "bbox": [ + 200, + 476, + 480, + 508 + ], + "type": "interline_equation", + "content": "\\operatorname {A t t e n t i o n} \\left(M _ {i} ^ {q}, L _ {s} ^ {k}, L _ {s} ^ {v}\\right) = \\operatorname {s o f t m a x} \\left(\\frac {M _ {i} ^ {q} L _ {s} ^ {k T}}{\\sqrt {d}}\\right) \\cdot L _ {s} ^ {v} \\tag {1}", + "image_path": "fd42c0fa9f196137e3ef8c55c1ebf48f73e7876dbf83048ce9d01d9cafee86bd.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 515, + 479, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 515, + 479, + 552 + ], + "spans": [ + { + "bbox": [ + 131, + 515, + 479, + 552 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 131, + 515, + 479, + 552 + ], + "type": "inline_equation", + "content": "L_{s}^{k} = W^{k}L_{s}" + }, + { + "bbox": [ + 131, + 515, + 479, + 552 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 515, + 479, + 552 + ], + "type": "inline_equation", + "content": "L_{s}^{v} = W^{v}L_{s}" + }, + { + "bbox": [ + 131, + 515, + 479, + 552 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 515, + 479, + 552 + ], + "type": "inline_equation", + "content": "M_{i}^{q} = \\text{flatens}(W^{q}M_{i})" + }, + { + "bbox": [ + 131, + 515, + 479, + 552 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 515, + 479, + 552 + ], + "type": "inline_equation", + "content": "L_{s}^{k}, L_{s}^{v} \\in R^{T_{L_{s}} \\times C_{i}}" + }, + { + "bbox": [ + 131, + 515, + 479, + 552 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 515, + 479, + 552 + ], + "type": "inline_equation", + "content": "M_{i}^{q} \\in R^{H_{i}W_{i} \\times C_{i}}" + }, + { + "bbox": [ + 131, + 515, + 479, + 552 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 131, + 515, + 479, + 552 + ], + "type": "inline_equation", + "content": "W^{k}, W^{v}, W^{q}" + }, + { + "bbox": [ + 131, + 515, + 479, + 552 + ], + "type": "text", + "content": " are implemented as a " + }, + { + "bbox": [ + 131, + 515, + 479, + 552 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 131, + 515, + 479, + 552 + ], + "type": "text", + "content": " convolution with " + }, + { + "bbox": [ + 131, + 515, + 479, + 552 + ], + "type": "inline_equation", + "content": "C_{i}" + }, + { + "bbox": [ + 131, + 515, + 479, + 552 + ], + "type": "text", + "content": " number of output channels." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 569, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 481, + 666 + ], + "type": "text", + "content": "Deformable Feature Alignment. The TOG task presents a unique challenge in that the given image lacks the referred object, which is adverse to aforementioned attention mechanism. The input language encompasses both appearance and spatial attributes. If exclusively relying on spatial information, the model can only provide a generic localization of the region due to its limited grasp of object concepts, consequently compromising the semantic rationality of edited images (e.g. a generated mouse being larger than the existing elephant). However, when appearance attributes are incorporated into the input, the cross-attention" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "X. Xue et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 302 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 302 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 302 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 302 + ], + "type": "image", + "image_path": "02720657c05cfc12c4222f6b1a9070a19be0a25695f6d194554d9d0ca168a4a0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 309, + 482, + 365 + ], + "lines": [ + { + "bbox": [ + 130, + 309, + 482, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 309, + 482, + 365 + ], + "type": "text", + "content": "Fig. 3: An illustration of deformable feature alignment. Global visual feature " + }, + { + "bbox": [ + 130, + 309, + 482, + 365 + ], + "type": "inline_equation", + "content": "V_{i}" + }, + { + "bbox": [ + 130, + 309, + 482, + 365 + ], + "type": "text", + "content": ", multimodal feature " + }, + { + "bbox": [ + 130, + 309, + 482, + 365 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 130, + 309, + 482, + 365 + ], + "type": "text", + "content": ", and text feature " + }, + { + "bbox": [ + 130, + 309, + 482, + 365 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 309, + 482, + 365 + ], + "type": "text", + "content": " are concatenated and inputted into the Transformer layer to learn offsets and modulation scalar. The multi-head cross attention is then performed to generate the refined attention map with sampled queries in " + }, + { + "bbox": [ + 130, + 309, + 482, + 365 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 130, + 309, + 482, + 365 + ], + "type": "text", + "content": " and language keys and values." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 389, + 480, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 389, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 480, + 449 + ], + "type": "text", + "content": "module faces a dilemma. It is tasked with reconciling the existing spatial attributes with the absent appearance features in the image, leading to conflicting information. This predicament significantly hampers the model's performance and generalization ability, especially in cases where the described appearance features are already present in the given image." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 449, + 481, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 481, + 532 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 481, + 532 + ], + "type": "text", + "content": "Inspired by deformation mechanisms [8,39,47], we propose deformable feature alignment to mitigate potential conflicts by shifting candidate queries towards regions of interest in multi-head cross attention, as illustrated in Fig. 3. The multimodal map " + }, + { + "bbox": [ + 130, + 449, + 481, + 532 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 130, + 449, + 481, + 532 + ], + "type": "text", + "content": " focuses on visually existing positional features by capturing the correlation between spatial-based textual and visual features, and it undergoes deformation with the alignment of object descriptions and multi-scale image features." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Offset Generation. Initially, we add a sinusoidal 1D positional encoding " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "e_{pos} \\in R^{T_L \\times C_L}" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " to the word embedding " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": ", updating as " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "L \\gets L + e_{pos}" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": ". A uniform grid of points " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "p \\in R^{H_i / \\gamma \\times W_i / \\gamma \\times 2}" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " are generated as references on " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "M_i" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " is the sampling factor. " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "M_i" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "V_i" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " are normalized to the same dimension through a linear layer and then concatenated as inputs into a transformer layer to generate patch-wise multimodal offset features. Subsequently, a linear projection is applied to modify the dimension to " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "H_i / \\gamma \\times W_i / \\gamma \\times 3" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": ", where the first two channels correspond to the learned 2D offsets " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\Delta p" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " constrained with maximum offset value " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\Delta p \\gets s_i \\tanh(\\Delta p)" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "), and the last channel is fed to a sigmoid layer to obtain modulation scalar " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\Delta m" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": ", which is introduced to exclude redundant context of which the elements lie in the range [0, 1]." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 409, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 409, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 409, + 91, + 447, + 100 + ], + "type": "text", + "content": "ST-LDM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "Multi-head Cross Attention. For reference points " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": ", 2D offsets " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\Delta p" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " and modulation scalar " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\Delta m" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " are generated via offset generation module. " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\Delta p" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " refines the spatial position according to textual descriptions, and " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\Delta m" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " introduces an additional dimension of freedom to exclude irrelevant context. Both of them are learned from the same transformer layer and activated with distinct modules. We designate " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\Delta p_{k}" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\Delta m_{k}" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " as the offset and scalar for the " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "-th reference point " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "p_{k}" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "k \\in \\{0,1,2,\\dots,H_iW_i / \\gamma^2\\}" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": ". As " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "p_{k} + \\Delta p_{k}" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " is fractional, a bilinear interpolation " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\phi (I,p_k)" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " is applied as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 214, + 220, + 481, + 246 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 220, + 481, + 246 + ], + "spans": [ + { + "bbox": [ + 214, + 220, + 481, + 246 + ], + "type": "interline_equation", + "content": "\\phi (I, p _ {k}) = \\sum_ {q} g \\left(p _ {k x}, q _ {x}\\right) \\cdot g \\left(p _ {k y}, q _ {y}\\right) \\cdot I (q), \\tag {2}", + "image_path": "fa91b89e1758ae0675951e4c1726f6d8acb9d3a9522e26e27f8df0fc93c20dc0.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 254, + 481, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 254, + 481, + 278 + ], + "spans": [ + { + "bbox": [ + 130, + 254, + 481, + 278 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 254, + 481, + 278 + ], + "type": "inline_equation", + "content": "g(a, b) = \\max(0, 1 - |a - b|)" + }, + { + "bbox": [ + 130, + 254, + 481, + 278 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 254, + 481, + 278 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 130, + 254, + 481, + 278 + ], + "type": "text", + "content": " indexes the 4 integral points closest to " + }, + { + "bbox": [ + 130, + 254, + 481, + 278 + ], + "type": "inline_equation", + "content": "p_k" + }, + { + "bbox": [ + 130, + 254, + 481, + 278 + ], + "type": "text", + "content": ". The sampling queries are deformed as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 235, + 287, + 481, + 300 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 287, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 235, + 287, + 481, + 300 + ], + "type": "interline_equation", + "content": "M _ {i} ^ {q} = \\operatorname {f l a t t e n} \\left(W ^ {q} \\phi \\left(M _ {i}, p + \\Delta p\\right)\\right). \\tag {3}", + "image_path": "1c49539f95942178af4d0105c2a8c76fbf473ba85ca8b7c12edb6df6fc2a12da.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 308, + 482, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 482, + 413 + ], + "type": "text", + "content": "The sampling visual queries are attended to the language keys with modulate attention scores " + }, + { + "bbox": [ + 130, + 308, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\Delta m" + }, + { + "bbox": [ + 130, + 308, + 482, + 413 + ], + "type": "text", + "content": ". A matrix completion function " + }, + { + "bbox": [ + 130, + 308, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\psi(A)" + }, + { + "bbox": [ + 130, + 308, + 482, + 413 + ], + "type": "text", + "content": " is then introduced to fill in the attention score of unsampled queries and restore the dimension of the attention map to " + }, + { + "bbox": [ + 130, + 308, + 482, + 413 + ], + "type": "inline_equation", + "content": "H_i W_i \\times T_{L_s}" + }, + { + "bbox": [ + 130, + 308, + 482, + 413 + ], + "type": "text", + "content": ". For each unsampled query " + }, + { + "bbox": [ + 130, + 308, + 482, + 413 + ], + "type": "inline_equation", + "content": "a_u" + }, + { + "bbox": [ + 130, + 308, + 482, + 413 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 130, + 308, + 482, + 413 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 130, + 308, + 482, + 413 + ], + "type": "text", + "content": ", the attention score of " + }, + { + "bbox": [ + 130, + 308, + 482, + 413 + ], + "type": "inline_equation", + "content": "a_u" + }, + { + "bbox": [ + 130, + 308, + 482, + 413 + ], + "type": "text", + "content": " is the weighted sum of scores of all the sampled queries " + }, + { + "bbox": [ + 130, + 308, + 482, + 413 + ], + "type": "inline_equation", + "content": "a_s" + }, + { + "bbox": [ + 130, + 308, + 482, + 413 + ], + "type": "text", + "content": " in the range " + }, + { + "bbox": [ + 130, + 308, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\mathbb{R}" + }, + { + "bbox": [ + 130, + 308, + 482, + 413 + ], + "type": "text", + "content": " around " + }, + { + "bbox": [ + 130, + 308, + 482, + 413 + ], + "type": "inline_equation", + "content": "a_u" + }, + { + "bbox": [ + 130, + 308, + 482, + 413 + ], + "type": "text", + "content": ". The number of " + }, + { + "bbox": [ + 130, + 308, + 482, + 413 + ], + "type": "inline_equation", + "content": "a_s" + }, + { + "bbox": [ + 130, + 308, + 482, + 413 + ], + "type": "text", + "content": " in the range " + }, + { + "bbox": [ + 130, + 308, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\mathbb{R}" + }, + { + "bbox": [ + 130, + 308, + 482, + 413 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 130, + 308, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\operatorname{card}(a_s)" + }, + { + "bbox": [ + 130, + 308, + 482, + 413 + ], + "type": "text", + "content": ", reflects the degree of deformation near " + }, + { + "bbox": [ + 130, + 308, + 482, + 413 + ], + "type": "inline_equation", + "content": "a_u" + }, + { + "bbox": [ + 130, + 308, + 482, + 413 + ], + "type": "text", + "content": " and is introduced as a modulation factor. " + }, + { + "bbox": [ + 130, + 308, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\operatorname{Avg}(\\operatorname{card}(a_s))" + }, + { + "bbox": [ + 130, + 308, + 482, + 413 + ], + "type": "text", + "content": " refers to the average number of " + }, + { + "bbox": [ + 130, + 308, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\operatorname{card}(a_s)" + }, + { + "bbox": [ + 130, + 308, + 482, + 413 + ], + "type": "text", + "content": " for all the " + }, + { + "bbox": [ + 130, + 308, + 482, + 413 + ], + "type": "inline_equation", + "content": "a_u" + }, + { + "bbox": [ + 130, + 308, + 482, + 413 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 130, + 308, + 482, + 413 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 130, + 308, + 482, + 413 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 223, + 413, + 481, + 440 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 413, + 481, + 440 + ], + "spans": [ + { + "bbox": [ + 223, + 413, + 481, + 440 + ], + "type": "interline_equation", + "content": "\\psi \\left(a _ {u}\\right) = \\frac {\\sum_ {a _ {s} \\in \\mathbb {R}} w _ {s} a _ {s}}{\\sum w _ {s}} \\cdot \\frac {\\operatorname {c a r d} \\left(a _ {s}\\right)}{\\operatorname {A v g} \\left(\\operatorname {c a r d} \\left(a _ {s}\\right)\\right)}, \\tag {4}", + "image_path": "f7e5f20a0eb02f189f1194f2aebac245f20f22a8f9adcf982e75956f7b6c410c.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 265, + 447, + 481, + 471 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 265, + 447, + 481, + 471 + ], + "spans": [ + { + "bbox": [ + 265, + 447, + 481, + 471 + ], + "type": "interline_equation", + "content": "w _ {s} = \\frac {1}{\\mathrm {d} \\left(a _ {u} , a _ {s}\\right) + \\epsilon}, \\tag {5}", + "image_path": "be9ffe25f20670db2206f84d2cfcf608c6207101827d7ac5956c7e10c8782ad9.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 475, + 481, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 475, + 481, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 481, + 499 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 475, + 481, + 499 + ], + "type": "inline_equation", + "content": "\\mathrm{d}(a_u, a_s)" + }, + { + "bbox": [ + 130, + 475, + 481, + 499 + ], + "type": "text", + "content": " refers to the Euclidean distance of " + }, + { + "bbox": [ + 130, + 475, + 481, + 499 + ], + "type": "inline_equation", + "content": "a_u" + }, + { + "bbox": [ + 130, + 475, + 481, + 499 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 475, + 481, + 499 + ], + "type": "inline_equation", + "content": "a_s" + }, + { + "bbox": [ + 130, + 475, + 481, + 499 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 475, + 481, + 499 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 475, + 481, + 499 + ], + "type": "text", + "content": " is a modulation constant. The output of the " + }, + { + "bbox": [ + 130, + 475, + 481, + 499 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 475, + 481, + 499 + ], + "type": "text", + "content": "-th attention head is formulated as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 217, + 506, + 481, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 506, + 481, + 539 + ], + "spans": [ + { + "bbox": [ + 217, + 506, + 481, + 539 + ], + "type": "interline_equation", + "content": "Z ^ {(j)} = \\operatorname {s o f t m a x} \\left(\\psi \\left(\\frac {M _ {i} ^ {q} L _ {s} ^ {k T}}{\\sqrt {d}} \\cdot \\varDelta m\\right)\\right) \\cdot L _ {s} ^ {v}. \\tag {6}", + "image_path": "eccae2306349ea6615d2a99d0b64e8f19b360c851d4a12e642139d2b5a34dc81.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "We preserve the original structure of cross attention in the initial fusion phase and utilize spatial conditions " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "L_{s}" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " as the key and value to generate a coarse-grained attention region. In subsequent stages, we introduce deformable feature alignment with " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " set as " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\{8,4,2\\}" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " to progressively refine spatial positions during the process of deep feature extraction. This strategy can bias the attention to spatial position information rather than image local features, simultaneously with fast convergence and high computational efficiency. In the last layer, we directly extract the attention score in the multi-head cross attention and then convert it to a two-dimensional matrix averaged by the channel axis, denoted as " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": ", as location guidance in denoising U-Net." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "X. Xue et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 315, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 315, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 315, + 128 + ], + "type": "text", + "content": "3.2 Training-free Spatial Guidance" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 479, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 479, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 479, + 224 + ], + "type": "text", + "content": "Location guidance " + }, + { + "bbox": [ + 130, + 141, + 479, + 224 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 141, + 479, + 224 + ], + "type": "text", + "content": " encompasses spatial attention generated in pixel space, delineating the approximate boundary of the generation position. Despite the refinement conducted through deformable feature alignment which adaptively adjusts spatial localization based on multimodal information, the numerical difference of attention scores in " + }, + { + "bbox": [ + 130, + 141, + 479, + 224 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 141, + 479, + 224 + ], + "type": "text", + "content": " may influence the shape of the generated object, compared to a manually defined bounding box. Thus, we utilize region-wise backward guidance to avoid potential overly aggressive intervention." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 225, + 480, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 225, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 131, + 225, + 480, + 289 + ], + "type": "text", + "content": "LDMs [33] augment underlying UNet backbone with spatial transformer, where the core module is cross-attention with " + }, + { + "bbox": [ + 131, + 225, + 480, + 289 + ], + "type": "inline_equation", + "content": "L_{k}^{(\\iota)} = W_{k}L_{a}" + }, + { + "bbox": [ + 131, + 225, + 480, + 289 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 225, + 480, + 289 + ], + "type": "inline_equation", + "content": "L_{v}^{(\\iota)} = W_{v}L_{a}" + }, + { + "bbox": [ + 131, + 225, + 480, + 289 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 225, + 480, + 289 + ], + "type": "inline_equation", + "content": "V_{q}^{(\\iota)} = \\text{ flatten}(W_{q}z_{t})" + }, + { + "bbox": [ + 131, + 225, + 480, + 289 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 131, + 225, + 480, + 289 + ], + "type": "inline_equation", + "content": "\\iota" + }, + { + "bbox": [ + 131, + 225, + 480, + 289 + ], + "type": "text", + "content": " is the index of the relevant layer in the network, " + }, + { + "bbox": [ + 131, + 225, + 480, + 289 + ], + "type": "inline_equation", + "content": "L_{a}" + }, + { + "bbox": [ + 131, + 225, + 480, + 289 + ], + "type": "text", + "content": " refers to appearance-based text prompt, and " + }, + { + "bbox": [ + 131, + 225, + 480, + 289 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 131, + 225, + 480, + 289 + ], + "type": "text", + "content": " is the latent representation of the input image. The attention weight matrix is described as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 241, + 300, + 480, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 300, + 480, + 338 + ], + "spans": [ + { + "bbox": [ + 241, + 300, + 480, + 338 + ], + "type": "interline_equation", + "content": "S ^ {(\\iota)} = \\operatorname {s o f t m a x} \\left(\\frac {V _ {q} ^ {(\\iota)} L _ {k} ^ {(\\iota) ^ {T}}}{\\sqrt {d}}\\right). \\tag {7}", + "image_path": "fd34e78664cb8410bbd386fb7343135b6dda0f3c07fc818b20e2c625e17e74c0.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 349, + 479, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 349, + 479, + 421 + ], + "spans": [ + { + "bbox": [ + 131, + 349, + 479, + 421 + ], + "type": "text", + "content": "We extract the first cross-attention block in the up-sampling branch to compute backpropagation, which yields superior performance while incurring minimal computational time overhead [6]. The corresponding attention score at timestamp " + }, + { + "bbox": [ + 131, + 349, + 479, + 421 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 349, + 479, + 421 + ], + "type": "text", + "content": " is denoted as " + }, + { + "bbox": [ + 131, + 349, + 479, + 421 + ], + "type": "inline_equation", + "content": "S_{t}" + }, + { + "bbox": [ + 131, + 349, + 479, + 421 + ], + "type": "text", + "content": ", and it has a similar compression scale with " + }, + { + "bbox": [ + 131, + 349, + 479, + 421 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 131, + 349, + 479, + 421 + ], + "type": "text", + "content": ", which is advantageous in subsequent calculations pertaining to the energy function. The energy function is defined as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 224, + 432, + 480, + 466 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 432, + 480, + 466 + ], + "spans": [ + { + "bbox": [ + 224, + 432, + 480, + 466 + ], + "type": "interline_equation", + "content": "E \\left(S _ {t}, G\\right) = \\left(1 - \\frac {\\sum \\left(\\theta \\left(\\hat {G} , \\beta\\right) \\odot S _ {t}\\right)}{\\sum S _ {t}}\\right) ^ {2}, \\tag {8}", + "image_path": "57bfb5cf618b7fce8bab910fe012b1ac717e3c7e5a846369788a52cbc0bf54a3.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 477, + 479, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 477, + 479, + 502 + ], + "spans": [ + { + "bbox": [ + 131, + 477, + 479, + 502 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 477, + 479, + 502 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 131, + 477, + 479, + 502 + ], + "type": "text", + "content": " is converted to the same resolution as " + }, + { + "bbox": [ + 131, + 477, + 479, + 502 + ], + "type": "inline_equation", + "content": "S_{t}" + }, + { + "bbox": [ + 131, + 477, + 479, + 502 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 131, + 477, + 479, + 502 + ], + "type": "inline_equation", + "content": "\\hat{G}" + }, + { + "bbox": [ + 131, + 477, + 479, + 502 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 131, + 477, + 479, + 502 + ], + "type": "inline_equation", + "content": "\\theta (\\hat{G},\\beta)" + }, + { + "bbox": [ + 131, + 477, + 479, + 502 + ], + "type": "text", + "content": " is an activation function with two steps, defined as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 200, + 513, + 480, + 545 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 513, + 480, + 545 + ], + "spans": [ + { + "bbox": [ + 200, + 513, + 480, + 545 + ], + "type": "interline_equation", + "content": "\\theta (\\hat {G}, \\beta) = \\operatorname {D i l a t e} (\\hat {G}) \\leftarrow \\hat {G} _ {m n} = \\left\\{ \\begin{array}{l l} 1 & \\hat {G} _ {m n} > \\beta \\\\ 0 & \\hat {G} _ {m n} \\leq \\beta \\end{array} , \\right. \\tag {9}", + "image_path": "f2d1b8a6ee6b760125dd627fa38767f099dbfbeec4e0bc3cd4492d10f96f0ae4.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 556, + 479, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 556, + 479, + 605 + ], + "spans": [ + { + "bbox": [ + 131, + 556, + 479, + 605 + ], + "type": "text", + "content": "where the element " + }, + { + "bbox": [ + 131, + 556, + 479, + 605 + ], + "type": "inline_equation", + "content": "\\hat{G}_{mn}" + }, + { + "bbox": [ + 131, + 556, + 479, + 605 + ], + "type": "text", + "content": " is first activated with the threshold value " + }, + { + "bbox": [ + 131, + 556, + 479, + 605 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 131, + 556, + 479, + 605 + ], + "type": "text", + "content": ", and then " + }, + { + "bbox": [ + 131, + 556, + 479, + 605 + ], + "type": "inline_equation", + "content": "\\hat{G}" + }, + { + "bbox": [ + 131, + 556, + 479, + 605 + ], + "type": "text", + "content": " is dilated to contour external maxmatrix to eliminate the numerical interference on the shape of the object. At each denoising U-Net, the latent " + }, + { + "bbox": [ + 131, + 556, + 479, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 131, + 556, + 479, + 605 + ], + "type": "text", + "content": " is updated with the gradient of Eq. 8." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 222, + 619, + 480, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 619, + 480, + 634 + ], + "spans": [ + { + "bbox": [ + 222, + 619, + 480, + 634 + ], + "type": "interline_equation", + "content": "\\boldsymbol {z} _ {t} \\leftarrow \\boldsymbol {z} _ {t} - \\eta \\sqrt {\\left(1 - \\alpha_ {t}\\right) / \\alpha_ {t}} \\nabla_ {\\boldsymbol {z} _ {t}} E (S _ {t}, G), \\tag {10}", + "image_path": "3cccdaf4bcbbcaacea46a38b95b91895928b901898fe16c47bb0e11c70c0c587.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\eta > 0" + }, + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "content": " defines the strength of guidance and " + }, + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "content": " refers to the decreasing sequence of noise schedule in stable diffusion." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 409, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 409, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 409, + 91, + 447, + 100 + ], + "type": "text", + "content": "ST-LDM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 201, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 201, + 152 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 201, + 152 + ], + "type": "text", + "content": "4.1 Dataset" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 160, + 479, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 160, + 479, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 160, + 479, + 194 + ], + "type": "text", + "content": "Our primary focus of chosen datasets is the linguistic diversity of spatial conditions rather than text2img prompts since our framework is explicitly tailored for spatial guidance with the parameter layers of LDMs held frozen." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 196, + 480, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 196, + 480, + 315 + ], + "spans": [ + { + "bbox": [ + 130, + 196, + 480, + 315 + ], + "type": "text", + "content": "VrR-VG. VrR-VG [25] is constructed from Visual Genome [21] and contains 58,983 real images with 23,375 relation pairs. Excluding positional and statistically biased relationships, It encompasses more visually relevant relationships, like \"sit on\", \"other side of\", \"fly in\", etc. We select spatial relationships and restructure linguistic expressions to align with our input format. For instance, for annotations involving attributes (fire hydrant, yellow), attributes (man, standing), and beside (man, fire hydrant), we reformulate them as \"standing man [beside yellow fire hydrant]\". Subsequently, We remove the specific object and complete images employing image inpainting techniques [11] as the input image, and the ground truth is the original position of the removed object." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 316, + 481, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 481, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 481, + 399 + ], + "type": "text", + "content": "ComCOCO. ComCOCO [40] is derived from RefCOCO + [20] and consists of 136,495 referring expressions for 34,615 objects in 23,951 image pairs. It is a synthetic dataset and the spatial relationship is delineated in the process of manual inspection. The annotations within ComCOCO encompass both appearance and spatial attributes, and we combine them in a manner analogous to the approach detailed above. We utilize real images in ComCOCO as the input image, and the corresponding synthetic images provide the ground-truth position." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 418, + 279, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 279, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 279, + 430 + ], + "type": "text", + "content": "4.2 Implementation Details" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 437, + 491, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 491, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 491, + 510 + ], + "type": "text", + "content": "We utilize SwinV2-T [27] with layer number " + }, + { + "bbox": [ + 130, + 437, + 491, + 510 + ], + "type": "inline_equation", + "content": "= \\{2, 2, 6, 2\\}" + }, + { + "bbox": [ + 130, + 437, + 491, + 510 + ], + "type": "text", + "content": " pretained on ADE20K [46] as the autoencoder in pixel space. Text embedding is initialized with official weights of ALBERT Base v2 [23] with 11M parameters and 12 layers. Our framework is adaptable to various latent diffusion models, and we conduct experiments by building upon a pretrained Stable-Diffusion (SD) V-1.5 [33] trained on the LAION-5B dataset [35], unless stated otherwise." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 510, + 481, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 481, + 664 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 481, + 664 + ], + "type": "text", + "content": "The training process is separated into two distinct phases. First, we train patch expanding and patch partition module on OpenImages [22] with KL-regularization following [33] to ensure optimal scaling properties of the pretrained autoencoder concerning spatial dimensionality when applied to LDMs. Subsequently, we train multimodal fusion transformer layers with the backbone of Swin-T on a combined dataset comprising a modified version of VrR-VG and ComCOCO. VrR-VG contributes a diverse and balanced set of spatial relationships, albeit with potential inpainting artifacts. To mitigate this limitation, we introduce ComCOCO which employs real images as its input source. Moreover, spatial descriptions in ComCOCO predominantly revolve around the holistic perspective of an image, whereas VrR-VG is specifically directed toward capturing and describing the intricate relationships between objects within the visual content." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "X. Xue et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "All the experiments are implemented on 4 V100 GPUs of 32G RAM. We set the downsampling factor " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "f = 4" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": " and channel dimension " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "c = 3" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": " in the autoencoder. Backward guidance is performed during the initial 10 steps of the diffusion process and repeated 3 times at each step with the strength factor " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": " set to 35. The input image is scaled to " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "480 \\times 480" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": ", and the data augmentation strategy includes random cropping and flipping. The optimizer is AdamW [29] with " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.85" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.91" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": ". The weight decay factor and the learning rate of the initial exponential decay are set to 0.003 and 0.00024, respectively, and the remaining parameters are randomly initialized. The model is trained for 172 rounds with the batch size set to 32 to achieve optimal performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 254, + 273, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 254, + 273, + 266 + ], + "spans": [ + { + "bbox": [ + 132, + 254, + 273, + 266 + ], + "type": "text", + "content": "4.3 Qualitative Evaluation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 275, + 482, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 275, + 482, + 431 + ], + "spans": [ + { + "bbox": [ + 130, + 275, + 482, + 431 + ], + "type": "text", + "content": "We apply our method on a multitude of real images from various domains, as exhibited in Fig. 1. We further compare the localization ability in Fig. 4 and generation ability in Fig. 5. Figure 4 shows the qualitative evaluation of ablative comparison of appearance and spatial properties. The deformable feature alignment module exhibits a remarkable ability to adjust spatial positions in accordance with linguistic descriptions and visual context. This dynamic adjustment imparts a direct influence on the image editing process within the latent diffusion model. As illustrated in Fig. 5, our method maintains the robust generative capabilities intrinsic to diffusion models with the preservation of semantics coherence to the original scene. The spatial constraints exerted on the denoising process have no discernible impact on the diversity and fidelity of the generated objects owing to region-wise backward guidance. More visualizations are provided in Appendix C." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 133, + 451, + 203, + 521 + ], + "blocks": [ + { + "bbox": [ + 133, + 451, + 203, + 521 + ], + "lines": [ + { + "bbox": [ + 133, + 451, + 203, + 521 + ], + "spans": [ + { + "bbox": [ + 133, + 451, + 203, + 521 + ], + "type": "image", + "image_path": "f4fd301c86199d473189b65968815287a43fbf96ad595419b264779e3231db17.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 205, + 451, + 274, + 521 + ], + "blocks": [ + { + "bbox": [ + 205, + 451, + 274, + 521 + ], + "lines": [ + { + "bbox": [ + 205, + 451, + 274, + 521 + ], + "spans": [ + { + "bbox": [ + 205, + 451, + 274, + 521 + ], + "type": "image", + "image_path": "d3744cd16794620b204ef968e68e95c07e6956281e5710ecdffdc7440e46dacf.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 275, + 451, + 343, + 521 + ], + "blocks": [ + { + "bbox": [ + 275, + 451, + 343, + 521 + ], + "lines": [ + { + "bbox": [ + 275, + 451, + 343, + 521 + ], + "spans": [ + { + "bbox": [ + 275, + 451, + 343, + 521 + ], + "type": "image", + "image_path": "0dfee3714abbba89db6a23c00153c90d134102a743dc4b80b41409249f1e1f81.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 343, + 451, + 410, + 521 + ], + "blocks": [ + { + "bbox": [ + 343, + 451, + 410, + 521 + ], + "lines": [ + { + "bbox": [ + 343, + 451, + 410, + 521 + ], + "spans": [ + { + "bbox": [ + 343, + 451, + 410, + 521 + ], + "type": "image", + "image_path": "5ae5482ab908094ecb016182e6ea3bae4768d36501ad60c5cd9cca038cfb96b4.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 410, + 451, + 480, + 521 + ], + "blocks": [ + { + "bbox": [ + 410, + 451, + 480, + 521 + ], + "lines": [ + { + "bbox": [ + 410, + 451, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 410, + 451, + 480, + 521 + ], + "type": "image", + "image_path": "1ed9ee9a8cbf41e1aa65360fc5f9d28be27d9bf03cd8a9310e55999205b27c81.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 133, + 535, + 203, + 605 + ], + "blocks": [ + { + "bbox": [ + 133, + 535, + 203, + 605 + ], + "lines": [ + { + "bbox": [ + 133, + 535, + 203, + 605 + ], + "spans": [ + { + "bbox": [ + 133, + 535, + 203, + 605 + ], + "type": "image", + "image_path": "391fdfc2858759940eebe066417b7eddfb1fa3a648cc68b7934b12dd52754a3b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 628, + 481, + 662 + ], + "lines": [ + { + "bbox": [ + 130, + 628, + 481, + 662 + ], + "spans": [ + { + "bbox": [ + 130, + 628, + 481, + 662 + ], + "type": "text", + "content": "Fig. 4: The first column corresponds to the input real image, and we visualize the guidance attention map " + }, + { + "bbox": [ + 130, + 628, + 481, + 662 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 628, + 481, + 662 + ], + "type": "text", + "content": " and corresponding edited results under different prompts with identical conditions (top) and identical prompts with different conditions (bottom)." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 205, + 535, + 274, + 604 + ], + "blocks": [ + { + "bbox": [ + 205, + 535, + 274, + 604 + ], + "lines": [ + { + "bbox": [ + 205, + 535, + 274, + 604 + ], + "spans": [ + { + "bbox": [ + 205, + 535, + 274, + 604 + ], + "type": "image", + "image_path": "1eac37b2a3679326598843ce777434bd8cc877db5748a21f7c3e5d140450640c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 234, + 604, + 314, + 619 + ], + "lines": [ + { + "bbox": [ + 234, + 604, + 314, + 619 + ], + "spans": [ + { + "bbox": [ + 234, + 604, + 314, + 619 + ], + "type": "text", + "content": "beef burger with ketchup on the nearby plate" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 275, + 535, + 343, + 604 + ], + "blocks": [ + { + "bbox": [ + 275, + 535, + 343, + 604 + ], + "lines": [ + { + "bbox": [ + 275, + 535, + 343, + 604 + ], + "spans": [ + { + "bbox": [ + 275, + 535, + 343, + 604 + ], + "type": "image", + "image_path": "f389749c01cd816b21580115703a97ecb24e71940739e3ff5f9eee65772986be.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 343, + 535, + 411, + 604 + ], + "blocks": [ + { + "bbox": [ + 343, + 535, + 411, + 604 + ], + "lines": [ + { + "bbox": [ + 343, + 535, + 411, + 604 + ], + "spans": [ + { + "bbox": [ + 343, + 535, + 411, + 604 + ], + "type": "image", + "image_path": "1bcba744b19c479e865617d342e9376515cbef8380c04e5fb89655efa7f058d5.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 370, + 604, + 451, + 619 + ], + "lines": [ + { + "bbox": [ + 370, + 604, + 451, + 619 + ], + "spans": [ + { + "bbox": [ + 370, + 604, + 451, + 619 + ], + "type": "text", + "content": "beef burger with ketchup on the more distant plate" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 411, + 535, + 480, + 604 + ], + "blocks": [ + { + "bbox": [ + 411, + 535, + 480, + 604 + ], + "lines": [ + { + "bbox": [ + 411, + 535, + 480, + 604 + ], + "spans": [ + { + "bbox": [ + 411, + 535, + 480, + 604 + ], + "type": "image", + "image_path": "366a04d5421400277d6ecfef5c9276b33ce1b8c9e9455c1cc0e4e591715dafd1.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 409, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 409, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 409, + 91, + 447, + 100 + ], + "type": "text", + "content": "ST-LDM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 203, + 183 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 203, + 183 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 203, + 183 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 203, + 183 + ], + "type": "image", + "image_path": "6e5fccfce4d3076a0de15f78d1957168ba8b78e481ad38890ad2059d1d9e6ca6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 202, + 183, + 480, + 193 + ], + "lines": [ + { + "bbox": [ + 202, + 183, + 480, + 193 + ], + "spans": [ + { + "bbox": [ + 202, + 183, + 480, + 193 + ], + "type": "text", + "content": "a blue luxury car with a streamlined body and a shiny paint job [to the right of the bus]" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 205, + 114, + 274, + 183 + ], + "blocks": [ + { + "bbox": [ + 205, + 114, + 274, + 183 + ], + "lines": [ + { + "bbox": [ + 205, + 114, + 274, + 183 + ], + "spans": [ + { + "bbox": [ + 205, + 114, + 274, + 183 + ], + "type": "image", + "image_path": "77ccf27d14991410aa3e7aec9558bda056903c8db74cbfc5fcd3a9c586f22d3a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 274, + 114, + 342, + 183 + ], + "blocks": [ + { + "bbox": [ + 274, + 114, + 342, + 183 + ], + "lines": [ + { + "bbox": [ + 274, + 114, + 342, + 183 + ], + "spans": [ + { + "bbox": [ + 274, + 114, + 342, + 183 + ], + "type": "image", + "image_path": "0ae96e1d50b3b3f75d4c40a823d1111285245d51ce38ab432f3fc27f79dcc2b6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 342, + 114, + 404, + 183 + ], + "blocks": [ + { + "bbox": [ + 342, + 114, + 404, + 183 + ], + "lines": [ + { + "bbox": [ + 342, + 114, + 404, + 183 + ], + "spans": [ + { + "bbox": [ + 342, + 114, + 404, + 183 + ], + "type": "image", + "image_path": "9c46e48637fc8f6dbd1f600d2c0dae642202a1a071f3d24e8bdb58c1a5eac9e8.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 404, + 114, + 479, + 183 + ], + "blocks": [ + { + "bbox": [ + 404, + 114, + 479, + 183 + ], + "lines": [ + { + "bbox": [ + 404, + 114, + 479, + 183 + ], + "spans": [ + { + "bbox": [ + 404, + 114, + 479, + 183 + ], + "type": "image", + "image_path": "69fcacd1b12763f8cdc47c93b3ef7313fa524a3ad9b07409a6ff523853f7c17a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 134, + 193, + 203, + 262 + ], + "blocks": [ + { + "bbox": [ + 134, + 193, + 203, + 262 + ], + "lines": [ + { + "bbox": [ + 134, + 193, + 203, + 262 + ], + "spans": [ + { + "bbox": [ + 134, + 193, + 203, + 262 + ], + "type": "image", + "image_path": "3486a468f8e0436bf52fb9291cc80c8d2394ac2279409044e53ebcee890deb90.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 280, + 482, + 304 + ], + "lines": [ + { + "bbox": [ + 131, + 280, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 131, + 280, + 482, + 304 + ], + "type": "text", + "content": "Fig. 5: Various results produced by ST-LDM, showcasing diverse appearances while adhering to the provided text descriptions." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 205, + 193, + 276, + 262 + ], + "blocks": [ + { + "bbox": [ + 205, + 193, + 276, + 262 + ], + "lines": [ + { + "bbox": [ + 205, + 193, + 276, + 262 + ], + "spans": [ + { + "bbox": [ + 205, + 193, + 276, + 262 + ], + "type": "image", + "image_path": "ce6a1228682cbc9b9014a7f81cddf3d00c857f6739c1ce428f78cf9d01a79e5b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 217, + 262, + 470, + 272 + ], + "lines": [ + { + "bbox": [ + 217, + 262, + 470, + 272 + ], + "spans": [ + { + "bbox": [ + 217, + 262, + 470, + 272 + ], + "type": "text", + "content": "a cute dinosaur with round face, animation style [standing in front of the door]" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 276, + 193, + 344, + 262 + ], + "blocks": [ + { + "bbox": [ + 276, + 193, + 344, + 262 + ], + "lines": [ + { + "bbox": [ + 276, + 193, + 344, + 262 + ], + "spans": [ + { + "bbox": [ + 276, + 193, + 344, + 262 + ], + "type": "image", + "image_path": "b699748339e401ea9bc10a70fc17193429cb25d2a6ab95cfcc188059ae962d42.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 344, + 193, + 412, + 262 + ], + "blocks": [ + { + "bbox": [ + 344, + 193, + 412, + 262 + ], + "lines": [ + { + "bbox": [ + 344, + 193, + 412, + 262 + ], + "spans": [ + { + "bbox": [ + 344, + 193, + 412, + 262 + ], + "type": "image", + "image_path": "38a635fd2cfaed842fcd5e0c829069487bbeab5c3cc5b7b80aa5b6029c9e18cd.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 412, + 193, + 479, + 262 + ], + "blocks": [ + { + "bbox": [ + 412, + 193, + 479, + 262 + ], + "lines": [ + { + "bbox": [ + 412, + 193, + 479, + 262 + ], + "spans": [ + { + "bbox": [ + 412, + 193, + 479, + 262 + ], + "type": "image", + "image_path": "378271a66860bc610209cca6db9a86be9373695c327efc750372ae0ce3a0d192.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 326, + 289, + 339 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 289, + 339 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 289, + 339 + ], + "type": "text", + "content": "4.4 Quantitative Comparison" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 345, + 482, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 345, + 482, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 345, + 482, + 429 + ], + "type": "text", + "content": "Evaluation Metrics. We conduct quantitative assessments of the quality of edited images by examining two facets. (1) Object alignment: We introduce VISOR [14] to score spatial object relationships according to text descriptions and text2img metric CLIPScore [17] to measure the consistency of generated objects with the input text. (2) Image fidelity: We utilize full-reference image quality assessments FID [44] and LPIPS [45] to measure image fidelity. HPS [38] is also introduced to align with human preference." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 445, + 482, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 445, + 482, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 482, + 578 + ], + "type": "text", + "content": "Benchmark Dataset. We construct a benchmark dataset for quantitative analysis and compare ST-LDM with related tasks. We find a distinct bias in the textual descriptions associated with the presence or absence of the specific object within the image. When the object is present in an image, the descriptions tend to emphasize distinctive features that differentiate the object from other visual entities. Conversely, when generating an object that does not exist in the original image, the focus shifts towards the detailed attributes of the object. Therefore, we randomly select 1,500 images from COCO2014 [26], and manually describe the object required to be generated in each image, simultaneously grounding the ideal position using bounding boxes. The detailed construction process is illustrated in Appendix B." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": "Results. We compare our method with editing models conditioned by diverse modalities in Tab. 1. Existing text-guided editing models exhibit limitations in generating new objects within complex scenes, and our model outperforms by a significant margin benefiting from explicit guidance generated from the deformable feature alignment. When compared to approaches employing forced modalities to precisely confine generated objects to designated regions," + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "X. Xue et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": "our method exhibits the capacity to edit local region of real images under the weak supervision of language, and the generated objects accurately adhere to the spatial relationships as specified in the text conditions." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 133, + 186, + 481, + 293 + ], + "blocks": [ + { + "bbox": [ + 148, + 165, + 463, + 177 + ], + "lines": [ + { + "bbox": [ + 148, + 165, + 463, + 177 + ], + "spans": [ + { + "bbox": [ + 148, + 165, + 463, + 177 + ], + "type": "text", + "content": "Table 1: Comparison with editing models conditioned by diverse modalities." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 186, + 481, + 293 + ], + "lines": [ + { + "bbox": [ + 133, + 186, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 133, + 186, + 481, + 293 + ], + "type": "table", + "html": "
ConditionModelFID↓LPIPS↓HPS↑CLIPScore↑VISOR↑
TextInstructPix2Pix [5]22.2530.9341.5860.4572.15
TextSmartEdit [18]18.4727.1147.4069.0277.86
MaskBlendedLatentDiffusion [3]12.9318.0554.3380.9688.12
MaskPaintByWord [1]9.7220.6850.2681.2387.41
BBoxBreakAScene [2]7.8916.2459.2581.1091.95
BBoxGLIGEN [24]6.2417.0962.4180.6290.53
TextST-LDM6.1615.5464.2881.9792.39
", + "image_path": "3dac86d9a00ec746e9f7c23857f25ba98eca3e42e608b6ef7225faef9b5fdd52.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 326, + 238, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 238, + 338 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 238, + 338 + ], + "type": "text", + "content": "4.5 Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 346, + 482, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 482, + 503 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 482, + 503 + ], + "type": "text", + "content": "Deformable Feature Alignment. We replace the feature alignment of multimodal fusion Transformer with our deformable mechanism at different stages, resulting in a direct impact on the modification of generated regions, as depicted in Fig. 6. We introduce FID and LPIPS to assess the extent of alteration in the original image. A user study is also conducted to assess the appropriateness of object position, which serves as a more persuasive and reference-worthy metric. As delineated in Tab. 2, the integration of deformable feature alignment refines the attention region, resulting in a numerical improvement across all metrics. However, premature intervention of DFA in the initial stage interferes with spatial perception (since the input " + }, + { + "bbox": [ + 130, + 346, + 482, + 503 + ], + "type": "inline_equation", + "content": "M_{1}" + }, + { + "bbox": [ + 130, + 346, + 482, + 503 + ], + "type": "text", + "content": " of DFA will be equal to " + }, + { + "bbox": [ + 130, + 346, + 482, + 503 + ], + "type": "inline_equation", + "content": "V_{1}" + }, + { + "bbox": [ + 130, + 346, + 482, + 503 + ], + "type": "text", + "content": " if we introduce DFA in Stage1). Actually, in the initial stage of modal fusion, the model generates a preliminary spatial positioning via cross attention, and DFA is designed as a refining strategy. More illustrations are provided in Appendix A." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 167, + 578, + 448, + 663 + ], + "blocks": [ + { + "bbox": [ + 130, + 523, + 482, + 568 + ], + "lines": [ + { + "bbox": [ + 130, + 523, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 523, + 482, + 568 + ], + "type": "text", + "content": "Table 2: Ablation study of deformable feature alignment (DFA) applied at different stages. Users are required to estimate the conformance of the object's position (True or False), and the score is calculated based on the overall accuracy across all the images. IT refers to the inference time (sec/image)." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 167, + 578, + 448, + 663 + ], + "lines": [ + { + "bbox": [ + 167, + 578, + 448, + 663 + ], + "spans": [ + { + "bbox": [ + 167, + 578, + 448, + 663 + ], + "type": "table", + "html": "
Stages w/ DFAUserScore↑FID↓LPIPS↓IT
S-1S-2S-3S-4
✓✓✓✓89.37.5016.23~7.0
✓✓✓92.46.1615.54~6.5
✓✓91.19.1420.46~6.3
✓82.515.9725.43~6.1
75.927.0334.86~6.0
", + "image_path": "f1f2b5f93a62236a571c74f5ceee8fe2f6062579e0f988c9aa60eacea8a7f3a3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 409, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 409, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 409, + 91, + 447, + 100 + ], + "type": "text", + "content": "ST-LDM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 113, + 481, + 247 + ], + "blocks": [ + { + "bbox": [ + 132, + 113, + 481, + 247 + ], + "lines": [ + { + "bbox": [ + 132, + 113, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 113, + 481, + 247 + ], + "type": "image", + "image_path": "98a01e03cf7a630875dd1e9b44d1ca816c2a95f6d147a35be63d8f2e2e3f88ee.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 150, + 256, + 463, + 269 + ], + "lines": [ + { + "bbox": [ + 150, + 256, + 463, + 269 + ], + "spans": [ + { + "bbox": [ + 150, + 256, + 463, + 269 + ], + "type": "text", + "content": "Fig. 6: Ablative visualization of attention maps at diverse stages with DFA." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 289, + 482, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 373 + ], + "type": "text", + "content": "Backward guidance. We further assess the effectiveness of backward guidance in Tab. 3. We replace guidance maps with textual conditions during the denoising process, and the notable decline indicates that linguistic information inadequately constrains the spatial placement of generated objects. We employ an activation function to eliminate potential discrepancies in attention scores, followed by dilation to mitigate the morphological interference of the attention region to the object." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 160, + 437, + 451, + 527 + ], + "blocks": [ + { + "bbox": [ + 130, + 394, + 482, + 429 + ], + "lines": [ + { + "bbox": [ + 130, + 394, + 482, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 394, + 482, + 429 + ], + "type": "text", + "content": "Table 3: Ablation study of backward guidance. Users are required to assess the matching score between the edited image and spatial descriptions. IT refers to the inference time (sec/image)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 160, + 437, + 451, + 527 + ], + "lines": [ + { + "bbox": [ + 160, + 437, + 451, + 527 + ], + "spans": [ + { + "bbox": [ + 160, + 437, + 451, + 527 + ], + "type": "table", + "html": "
UserScoreCLIPScoreVISORIT
uncondcond
w/o guidance58.7355.0711.5460.45~4.1
w/o activation80.8669.0137.2690.94~6.0
w/o dilation83.9870.5239.1490.37~6.2
Ours89.1181.9743.4192.39~6.5
", + "image_path": "d8d6c7a44d4359f36a28a97d6e0845581c8fec5505b37f90ffef67c557882c47.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 555, + 220, + 568 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 220, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 220, + 568 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we explore the text-grounded object generation in real-world images. A universal framework with deformable feature alignment is proposed to hierarchically refine the attention region conforming to vision and language features. The generated guidance map can be incorporated into any latent diffusion model with training-free backward guidance. Our model addresses the inherent limitation of diffusion models in spatial perception under textual supervision, and offers valuable insights for subsequent research." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "X. Xue et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 213 + ], + "type": "text", + "content": "This work was supported in part by the National Science and Technology Major Project under Grant 2021ZD0113202, in part by the National Natural Science Foundation of China under Grants 62171125, 61876037, in part by the innovation project of Jiangsu Province under grants BZ2023042, BY2022564. The authors would like to thank the Big Data Computing Center of Southeast University for providing the facility support on the numerical calculations in this paper." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 232, + 197, + 244 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 232, + 197, + 244 + ], + "spans": [ + { + "bbox": [ + 132, + 232, + 197, + 244 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 257, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 257, + 479, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 257, + 479, + 277 + ], + "spans": [ + { + "bbox": [ + 138, + 257, + 479, + 277 + ], + "type": "text", + "content": "1. Andonian, A., Osmany, S., Cui, A., Park, Y., Jahanian, A., Torralba, A., Bau, D.: Paint by word" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 280, + 480, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 280, + 480, + 311 + ], + "spans": [ + { + "bbox": [ + 138, + 280, + 480, + 311 + ], + "type": "text", + "content": "2. Avrahami, O., Aberman, K., Fried, O., Cohen-Or, D., Lischinski, D.: Break-a-scene: Extracting multiple concepts from a single image. arXiv preprint arXiv:2305.16311 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 313, + 480, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 313, + 480, + 334 + ], + "spans": [ + { + "bbox": [ + 138, + 313, + 480, + 334 + ], + "type": "text", + "content": "3. Avrahami, O., Fried, O., Lischinski, D.: Blended latent diffusion. ACM Transactions on Graphics 42(4), 1-11 (Jul 2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 335, + 480, + 377 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 335, + 480, + 377 + ], + "spans": [ + { + "bbox": [ + 138, + 335, + 480, + 377 + ], + "type": "text", + "content": "4. Avrahami, O., Hayes, T., Gafni, O., Gupta, S., Taigman, Y., Parikh, D., Lischinski, D., Fried, O., Yin, X.: Spatext: Spatio-textual representation for controllable image generation. In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE (Jun 2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 379, + 480, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 379, + 480, + 411 + ], + "spans": [ + { + "bbox": [ + 138, + 379, + 480, + 411 + ], + "type": "text", + "content": "5. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18392-18402 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 412, + 480, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 412, + 480, + 433 + ], + "spans": [ + { + "bbox": [ + 138, + 412, + 480, + 433 + ], + "type": "text", + "content": "6. Chen, M., Laina, I., Vedaldi, A.: Training-free layout control with cross-attention guidance (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 434, + 480, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 434, + 480, + 466 + ], + "spans": [ + { + "bbox": [ + 138, + 434, + 480, + 466 + ], + "type": "text", + "content": "7. Couairon, G., Verbeek, J., Schwenk, H., Cord, M.: Diffedit: Diffusion-based semantic image editing with mask guidance. In: The Eleventh International Conference on Learning Representations (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 468, + 480, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 468, + 480, + 499 + ], + "spans": [ + { + "bbox": [ + 138, + 468, + 480, + 499 + ], + "type": "text", + "content": "8. Dai, J., Qi, H., Xiong, Y., Li, Y., Zhang, G., Hu, H., Wei, Y.: Deformable convolutional networks. In: 2017 IEEE International Conference on Computer Vision (ICCV). pp. 764-773 (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 501, + 480, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 501, + 480, + 532 + ], + "spans": [ + { + "bbox": [ + 138, + 501, + 480, + 532 + ], + "type": "text", + "content": "9. Ding, H., Liu, C., Wang, S., Jiang, X.: Vision-language transformer and query generation for referring segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 16321-16330 (October 2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 533, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 533, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 134, + 533, + 480, + 566 + ], + "type": "text", + "content": "10. Ding, Z., Zhang, X., Xia, Z., Jebe, L., Tu, Z., Zhang, X.: Diffusionrig: Learning personalized priors for facial appearance editing. In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 12736-12746 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 567, + 480, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 567, + 480, + 598 + ], + "spans": [ + { + "bbox": [ + 134, + 567, + 480, + 598 + ], + "type": "text", + "content": "11. Dong, Q., Cao, C., Fu, Y.: Incremental transformer structure enhanced image inpainting with masking positional encoding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 599, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 599, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 134, + 599, + 480, + 620 + ], + "type": "text", + "content": "12. Fan, W.C., Chen, Y.C., Chen, D., Cheng, Y., Yuan, L., Wang, Y.C.F.: Frido: Feature pyramid diffusion for complex scene image synthesis. In: AAAI (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 621, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 621, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 621, + 480, + 665 + ], + "type": "text", + "content": "13. Gafni, O., Polyak, A., Ashual, O., Sheynin, S., Parikh, D., Taigman, Y.: Make-a-scene: Scene-based text-to-image generation with human priors. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022. pp. 89-106. Springer Nature Switzerland, Cham (2022)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 409, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 409, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 409, + 91, + 447, + 100 + ], + "type": "text", + "content": "ST-LDM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "14. Gokhale, T., Palangi, H., Nushi, B., Vineet, V., Horvitz, E., Kamar, E., Baral, C., Yang, Y.: Benchmarking spatial relationships in text-to-image generation (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 480, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 480, + 161 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 480, + 161 + ], + "type": "text", + "content": "15. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 163, + 480, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 163, + 480, + 195 + ], + "spans": [ + { + "bbox": [ + 133, + 163, + 480, + 195 + ], + "type": "text", + "content": "16. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-or, D.: Prompt-to-prompt image editing with cross-attention control. In: The Eleventh International Conference on Learning Representations (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 197, + 480, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 197, + 480, + 217 + ], + "spans": [ + { + "bbox": [ + 133, + 197, + 480, + 217 + ], + "type": "text", + "content": "17. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: CLIPScore: a reference-free evaluation metric for image captioning. In: EMNLP (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 219, + 480, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 219, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 133, + 219, + 480, + 251 + ], + "type": "text", + "content": "18. Huang, Y., Xie, L., Wang, X., Yuan, Z., Cun, X., Ge, Y., Zhou, J., Dong, C., Huang, R., Zhang, R., et al.: Smartedit: Exploring complex instruction-based image editing with multimodal large language models. arXiv preprint arXiv:2312.06739 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 253, + 480, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 253, + 480, + 296 + ], + "spans": [ + { + "bbox": [ + 133, + 253, + 480, + 296 + ], + "type": "text", + "content": "19. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 6007-6017 (June 2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 297, + 480, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 297, + 480, + 330 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 480, + 330 + ], + "type": "text", + "content": "20. Kazemzadeh, S., Ordonez, V., Matten, M., Berg, T.: Referitgame: Referring to objects in photographs of natural scenes. In: Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP). pp. 787-798 (2014)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 331, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 331, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 331, + 480, + 374 + ], + "type": "text", + "content": "21. Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., et al.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision 123, 32-73 (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 376, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 376, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 376, + 480, + 418 + ], + "type": "text", + "content": "22. Kuznetsova, A., Rom, H., Alldrin, N., Uijlings, J., Krasin, I., Pont-Tuset, J., Kamali, S., Popov, S., Malloci, M., Kolesnikov, A., et al.: The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision 128(7), 1956-1981 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 420, + 480, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 420, + 480, + 463 + ], + "spans": [ + { + "bbox": [ + 132, + 420, + 480, + 463 + ], + "type": "text", + "content": "23. Lan, Z., Chen, M., Goodman, S., Gimpel, K., Sharma, P., Soricut, R.: ALBERT: A lite BERT for self-supervised learning of language representations. In: 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 464, + 480, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 480, + 486 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 480, + 486 + ], + "type": "text", + "content": "24. Li, Y., Liu, H., Wu, Q., Mu, F., Yang, J., Gao, J., Li, C., Lee, Y.J.: Gligen: Open-set grounded text-to-image generation. CVPR (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 487, + 480, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 480, + 519 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 480, + 519 + ], + "type": "text", + "content": "25. Liang, Y., Bai, Y., Zhang, W., Qian, X., Zhu, L., Mei, T.: Vrr-vg: Refocusing visually-relevant relationships. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10403-10412 (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 521, + 480, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 521, + 480, + 553 + ], + "spans": [ + { + "bbox": [ + 132, + 521, + 480, + 553 + ], + "type": "text", + "content": "26. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: European conference on computer vision. pp. 740-755. Springer (2014)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 554, + 480, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 554, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 554, + 480, + 597 + ], + "type": "text", + "content": "27. Liu, Z., Hu, H., Lin, Y., Yao, Z., Xie, Z., Wei, Y., Ning, J., Cao, Y., Zhang, Z., Dong, L., et al.: Swin transformer v2: Scaling up capacity and resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12009-12019 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 599, + 480, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 480, + 642 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 480, + 642 + ], + "type": "text", + "content": "28. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 10012-10022 (October 2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "29. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: International Conference on Learning Representations (2017)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "X. Xue et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "text", + "content": "30. Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Van Gool, L.: Repaint: Inpainting using denoising diffusion probabilistic models. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11451-11461 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 160, + 482, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 160, + 482, + 225 + ], + "spans": [ + { + "bbox": [ + 130, + 160, + 482, + 225 + ], + "type": "text", + "content": "31. Nichol, A.Q., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., Mcgrew, B., Sutskever, I., Chen, M.: GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. In: Chaudhuri, K., Jegelka, S., Song, L., Szepesvari, C., Niu, G., Sabato, S. (eds.) Proceedings of the 39th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 162, pp. 16784-16804. PMLR (17-23 Jul 2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 225, + 482, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 225, + 482, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 225, + 482, + 258 + ], + "type": "text", + "content": "32. OuYang, S.R., Wang, H., Xie, S., Niu, Z., Tong, R., Chen, Y.W., Lin, L.: Slvit: Scale-wise language-guided vision transformer for referring image segmentation. In: International Joint Conference on Artificial Intelligence (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 258, + 482, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 482, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 482, + 289 + ], + "type": "text", + "content": "33. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 289, + 482, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 332 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 332 + ], + "type": "text", + "content": "34. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 22500-22510 (June 2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 332, + 482, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 332, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 332, + 482, + 407 + ], + "type": "text", + "content": "35. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., Schramowski, P., Kundurthy, S., Crowson, K., Schmidt, L., Kaczmarczyk, R., Jitsev, J.: Laion-5b: An open large-scale dataset for training next generation image-text models. In: Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., Oh, A. (eds.) Advances in Neural Information Processing Systems. vol. 35, pp. 25278-25294. Curran Associates, Inc. (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 407, + 482, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 482, + 451 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 482, + 451 + ], + "type": "text", + "content": "36. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L.u., Polosukhin, I.: Attention is all you need. In: Guyon, I., Luxburg, U.V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., Garnett, R. (eds.) Advances in Neural Information Processing Systems. vol. 30. Curran Associates, Inc. (2017)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 451, + 482, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 451, + 482, + 494 + ], + "spans": [ + { + "bbox": [ + 130, + 451, + 482, + 494 + ], + "type": "text", + "content": "37. Wang, J., Lu, G., Xu, H., Li, Z., Xu, C., Fu, Y.: Manitrans: Entity-level text-guided image manipulation via token-wise semantic alignment and generation. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10697-10707 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 494, + 482, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 494, + 482, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 494, + 482, + 536 + ], + "type": "text", + "content": "38. Wu, X., Sun, K., Zhu, F., Zhao, R., Li, H.: Human preference score: Better aligning text-to-image models with human preference. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2096-2105 (October 2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 536, + 482, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 536, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 536, + 482, + 568 + ], + "type": "text", + "content": "39. Xia, Z., Pan, X., Song, S., Li, L.E., Huang, G.: Vision transformer with deformable attention. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4794-4803 (June 2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 568, + 482, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 568, + 482, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 568, + 482, + 590 + ], + "type": "text", + "content": "40. Xue, X., Wu, J., Kong, Y., Senhadji, L., Shu, H.: Rethinking referring object removal (2024)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 590, + 482, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 590, + 482, + 622 + ], + "spans": [ + { + "bbox": [ + 130, + 590, + 482, + 622 + ], + "type": "text", + "content": "41. Yang, L., Zhang, Z., Song, Y., Hong, S., Xu, R., Zhao, Y., Shao, Y., Zhang, W., Cui, B., Yang, M.H.: Diffusion models: A comprehensive survey of methods and applications. arXiv preprint arXiv:2209.00796 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 622, + 482, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 622, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 622, + 482, + 665 + ], + "type": "text", + "content": "42. Yang, Z., Wang, J., Tang, Y., Chen, K., Zhao, H., Torr, P.H.: Lavo: Language-aware vision transformer for referring image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 18155-18165 (June 2022)" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 409, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 409, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 409, + 91, + 447, + 100 + ], + "type": "text", + "content": "ST-LDM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 346 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "text", + "content": "43. Yang, Z., Wang, J., Gan, Z., Li, L., Lin, K., Wu, C., Duan, N., Liu, Z., Liu, C., Zeng, M., Wang, L.: Reco: Region-controlled text-to-image generation. In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 162, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 162, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 162, + 481, + 194 + ], + "type": "text", + "content": "44. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "text", + "content": "45. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "type": "text", + "content": "46. Zhou, B., Zhao, H., Puig, X., Xiao, T., Fidler, S., Barriuso, A., Torralba, A.: Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision 127, 302-321 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "text", + "content": "47. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. In: 9th International Conference on Learning Representations, ICLR (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 346 + ], + "type": "text", + "content": "48. Zhu, Y., Liu, H., Song, Y., Yuan, Z., Han, X., Yuan, C., Chen, Q., Wang, J.: One model to edit them all: Free-form text-driven image manipulation with semantic modulations. In: Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., Oh, A. (eds.) Advances in Neural Information Processing Systems. pp. 25146-25159. Curran Associates, Inc." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "X. Xue et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/ST-LLM_ Large Language Models Are Effective Temporal Learners/f824889a-c655-4483-9dd5-c5cb83e43784_content_list.json b/2024/ST-LLM_ Large Language Models Are Effective Temporal Learners/f824889a-c655-4483-9dd5-c5cb83e43784_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b9daece90fa99b41ab57cdbeaa381ee5773f2949 --- /dev/null +++ b/2024/ST-LLM_ Large Language Models Are Effective Temporal Learners/f824889a-c655-4483-9dd5-c5cb83e43784_content_list.json @@ -0,0 +1,1554 @@ +[ + { + "type": "text", + "text": "ST-LLM: Large Language Models Are Effective Temporal Learners", + "text_level": 1, + "bbox": [ + 228, + 140, + 774, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ruyang Liu\\*1,3 Chen Li\\*2 Haoran Tang1,3 Yixiao Ge2 Ying Shan2 Ge Li", + "bbox": [ + 215, + 212, + 784, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 School of Electronic and Computer Engineering, Shenzhen Graduate School,", + "bbox": [ + 238, + 239, + 761, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Peking University {ruyang@stu,hrtang@geli@ece}.pku.edu.cn", + "bbox": [ + 274, + 253, + 728, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2ARC Lab, Tencent PCG 3 Peng Cheng Laboratory", + "bbox": [ + 308, + 267, + 692, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{palchenli,yixiaoge,yingsshan}@tencent.com", + "bbox": [ + 338, + 282, + 663, + 296 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Large Language Models (LLMs) have showcased impressive capabilities in text comprehension and generation, prompting research efforts towards video LLMs to facilitate human-AI interaction at the video level. However, how to effectively encode and understand videos in video-based dialogue systems remains to be solved. In this paper, we investigate a straightforward yet unexplored question: Can we feed all spatial-temporal tokens into the LLM, thus delegating the task of video sequence modeling to the LLMs? Surprisingly, this simple approach yields significant improvements in video understanding. Based upon this, we propose ST-LLM, an effective video-LLM baseline with Spatial-Temporal sequence modeling inside LLM. Furthermore, to address the overhead and stability issues introduced by uncompressed video tokens within LLMs, we develop a dynamic masking strategy with tailormade training objectives. For particularly long videos, we have also designed a global-local input module to balance efficiency and effectiveness. Consequently, we harness LLM for proficient spatial-temporal modeling, while upholding efficiency and stability. Extensive experimental results attest to the effectiveness of our method. Through a more concise model and training pipeline, ST-LLM establishes a new state-of-the-art result on VideoChatGPT-Bench and MVBench. Codes have been available at https://github.com/TencentARC/ST-LLM.", + "bbox": [ + 259, + 337, + 743, + 630 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: ST-LLM $\\cdot$ Dynamic Masking $\\cdot$ Global-Local Input", + "bbox": [ + 261, + 642, + 686, + 656 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 684, + 375, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large Language Models (LLMs), such as GPT [19], PaLM [316] and LLaMA [43,44], have achieved remarkable success owing to their formidable language understanding and generation abilities, signaling a promising advancement towards artificial general intelligence. Driven by the widespread adoption of LLMs, research on Large Visual-Language Models (LVLMs) [14,29,55] has emerged to extend the capabilities of LLMs to process visual signals, which have revealed proficiency in image-based conversations. However, in contrast to image inputs, videos feature heavier input and additional temporal information. Developing a", + "bbox": [ + 212, + 718, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f59a836b920cc3b1c4307ed8c08f2b860e20ab2f43f53d9e34b006a741086425.jpg", + "image_caption": [ + "Fig. 1: Qualitative comparisons among top-performance video LLMs. We illustrate two cases from MVBench [25] (above) and one case from a YouTube video (below)." + ], + "image_footnote": [], + "bbox": [ + 223, + 146, + 779, + 385 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "large video-language model capable of effectively extracting meaningful spatial-temporal information from intricate video signals presents a significant challenge.", + "bbox": [ + 212, + 465, + 787, + 496 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, several preliminary attempts have surfaced aimed at extending LLMs to facilitate video conversation [19,20,24,25,27,31,33,53], which unlock the capability of generating generic summaries of video content. Despite these efforts, existing video LLMs continue to exhibit shortcomings in achieving satisfactory performance in video comprehension, particularly in the realm of understanding content that relies heavily on temporal dynamics. As depicted in Fig. 11, contemporary top-performing video LLMs demonstrate robust understanding capabilities for actions reliant on static contexts. However, their proficiency in comprehending scenes involving motion remains constrained. They encounter challenges in discerning even the most basic direction of object movement, let alone the complex motion or scene transition.", + "bbox": [ + 212, + 502, + 787, + 667 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This issue may stem from the inherent difficulty of temporal modeling within conversational systems. Historically, achieving robust video encoding has often required substantially more time and memory resources than images, which is impractical based on LLMs. As illustrated by Fig. 2(a), early video LLMs tend to adopt mean pooling on temporal dimension [24,32,33,53], a method that, while efficient, is inadequate for handling dynamic temporal sequences. Therefore, recent models frequently integrate additional structures for temporal sampling and modeling [19,20,25,27]. While these methods offer improved effectiveness compared to average pooling, they entail increased storage and demand extensive GPU time for training from scratch, often involving two or even three-stage pretraining to align new modules.", + "bbox": [ + 212, + 672, + 789, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Liu. et al.", + "bbox": [ + 271, + 114, + 339, + 126 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4c8d15fb042557baabf867ecd3c28e164eee717b27b8545061fc64a3e77a9d27.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 220, + 143, + 377, + 215 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/358cc03e5bbc1bc108e1a8b7d3f1bee6bcfee4fe6fb96c1e0a4b2226c201984f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 382, + 143, + 542, + 215 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0ae16f48316405930a7404358e1f3e1a766a004fa8736d4530498f554722a377.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 218, + 218, + 377, + 276 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/904b434c78bffa8a9a4d7df9a0171f05f4d91c38c25748d4f6761dabd240329f.jpg", + "image_caption": [ + "Fig. 2: (a) Comparison among different video LLMs based on the way of inputting visual tokens. (b) Quantitative comparison of representative video LLMs, presenting the average results of all 20 metrics and temporal-sensitive motion-related metrics from MVBench, which encompass Moving Direction, Moving Count, and Moving Attribute." + ], + "image_footnote": [], + "bbox": [ + 382, + 218, + 542, + 276 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/49c933261dc499f7ed1f56945e79dbdec09abcc195a1ab4a7f3ce799365ed549.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 555, + 148, + 771, + 271 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inspired by the recent success of joint spatial-temporal-text modeling in visual generation [8,36], in this work, we investigate a simple yet unexplored idea: Given the robust sequence modeling capabilities inherent to LLMs, what if we input all visual tokens into the LLM and delegate the task of modeling spatial-temporal sequences to the LLM itself? Meanwhile, there are two additional challenges: (1) The inclusion of all visual tokens significantly increases the context length within LLMs, particularly for long videos, rendering the processing of a large number of frames unaffordable. (2) LLM may struggle to handle videos of varying lengths, potentially leading to hallucinations when there is a discrepancy between the number of frames in testing and training.", + "bbox": [ + 212, + 383, + 787, + 535 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address these issues, we propose ST-LLM, a simple but powerful baseline of video LLM. As depicted in Fig. 2(a), with raw spatial-temporal tokens inside LLM, we harness the LLMs' robust sequence modeling capabilities for effective temporal modeling. Meanwhile, we introduce a novel dynamic video token masking strategy along with masking video modeling during training. With this approach, we reduce the length of sequences input to the LLM while significantly improving the robustness of videos of varying lengths during inference. To accommodate particularly long videos, we have devised a unique global-local input mechanism. This involves employing mean pooling of a large number of frames to generate residual input for a smaller subset of frames. Through this asymmetric design, we can process input from a large number of video frames while preserving the operations of modeling video tokens within LLMs.", + "bbox": [ + 212, + 536, + 787, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As a result, we maintain input tokens of comparable length to most video LLMs while effectively harnessing the LLM's capacity to comprehend videos. Furthermore, as ST-LLM does not introduce additional modules, it does not necessitate expensive alignment pre-training. This enables ST-LLM to directly leverage existing state-of-the-art image conversational models, resulting in significantly reduced GPU time compared to other state-of-the-art video LLMs. Extensive experiments are conducted to verify the effectiveness of ST-LLM. Qualitatively, as depicted in Fig. 1, ST-LLM demonstrates a superior ability to", + "bbox": [ + 212, + 719, + 787, + 842 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "ST-LLM: Large Language Models Are Effective Temporal Learners", + "bbox": [ + 279, + 113, + 727, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "understand dynamics compared to other video LLMs. Quantitatively, ST-LLM achieves new state-of-the-art performance across various contemporary video benchmarks, including MVBench [25], VideoChatGPT-Bench [33] and zero-shot QA-Evaluation. Specifically, as illustrated in Fig. 2(b), the superiority of ST-LLM is particularly evident in metrics related to the temporal-sensitive motion, which underscores the expertise of our model in temporal understanding.", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 262, + 397, + 277 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "LLMs and Image LLMs. Recent years have witnessed remarkable advancements in the evolution of Large Language Models (LLMs). Inspired by InstructGPT [35] and the widely recognized commercial model ChatGPT [34], the academic community has seen a proliferation of open-source LLMs, such as LLaMA [43], Alpaca [41], Vicuna [13], and LLaMA 2 [44], which have become foundational components for a myriad of research endeavors. Meanwhile, LLMs, being large-scale transformer models, inherently possess strong sequence modeling capabilities. This inspired us to explore the feasibility of entrusting the task of spatial-temporal sequence modeling to LLMs. The success of LLMs has spurred increasing interest in the development of Multimodal Large Language Models (MLLMs). Notable breakthroughs include works such as Flamingo [2], BLIP2 [23], and PaLM-E [16], which have successfully bridged the gap between vision models and LLMs. Inspired by the concept of instruction tuning for LLMs, a series of works [12, 14, 29, 55] leveraged open-source chatbots for visual instruction tuning, thus facilitating support for image-based conversations. It is noteworthy that in the pre-LLM era, video models were commonly constructed based on pre-trained 2D image models [7, 11, 30, 45]. This may be attributed to the significant computational demands of video processing, combined with the relatively limited scale and diversity of video data compared to images. Therefore, we believe that a well-pretrained image dialogue model could similarly offer training cost and performance benefits to video LLMs.", + "bbox": [ + 212, + 295, + 787, + 611 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Video LLMs. The emergence of MLLMs quickly extended into the domain of video as well [19,20,24,25,27,31,33,53]. Early models like VideoChat [24], VideoChatGPT [33] and Valley [32] generate video instruction tuning data through GPT to enable video conversations. At the model level, these approaches typically involve the use of mean pooling to aggregate the encoding results of individual frames before feeding them into the LLMs. While efficient, it is evident that this method is inadequate for effective temporal modeling. Hence, some subsequent models have initiated exploration into adaptations of image models to video-specific requirements. For instance, BT-Adapter [31] proposed a lightweight adapter to extend the capabilities of image LLMs. Chat-UniVi [20] introduces DPC-KNN to cluster dynamic visual tokens. Moreover, VideoChat2 [25] deviated from using CLIP [38] and directly introduced a dedicated video encoder. Although these methods offer improvements over mean pooling, the inclusion of additional modules typically necessitates intensive training. For example, in the case of VideoChat2, aligning the newly introduced video encoder entails", + "bbox": [ + 212, + 613, + 795, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Liu. et al.", + "bbox": [ + 271, + 114, + 339, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7be2628aadc0f8c2d2bc5d6887ec862b3413425dbadc3730c2945f67a03faab4.jpg", + "image_caption": [ + "Fig.3: The overview of ST-LLM for generating responses on the input video and instructions. ST-LLM directly feeds the spatial-temporal sequence into the LLM and employs dynamic masking and global-local input for efficiency and robustness." + ], + "image_footnote": [], + "bbox": [ + 246, + 148, + 748, + 324 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "a three-stage training involving 30 million visual-language samples. In contrast, ST-LLM adopts a more direct approach by entrusting the task of visual sequence modeling to LLMs, which is characterized by its conciseness, reduced training requirements, and, according to our experimental findings, superior effectiveness. Joint Spatial-Temporal Modeling. Prior to the emergence of LLMs, joint spatial-temporal modeling had demonstrated effectiveness in both video-only and video-language pretraining [26,42,49]. More recently, inspired by DIT [36], the popular video generation model Sora has shown success in simultaneously processing video tokens and text using the transformer. Despite its potential, joint spatial-temporal-text modeling remains relatively uncommon in video LLMs. While VideoChat2 employed a joint-ST video encoder, it necessitated extensive pretraining. A closer relative, InstructBLIP incorporated joint spatial-temporal sequences into LLM for zero-shot evaluation on video datasets. However, this approach led to significant hallucinations due to the absence of video instruction tuning. To the best of our knowledge, ST-LLM is the first open-source video LLM to adopt joint spatial-temporal-text modeling. Moreover, ST-LLM introduces a series of design innovations to mitigate the limitations associated with incorporating all video tokens within LLM.", + "bbox": [ + 212, + 411, + 787, + 684 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 215, + 712, + 380, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we elaborate on the structure and training methodology of ST-LLM. Firstly, we present a description of the model architecture, which incorporates all spatial-temporal tokens within the LLM, in Section 3.1. Subsequently, we delve into the masking mechanism and training objectives in Section 3.2. Finally, we introduce the global-local input approach in Section 3.3. The overarching framework of our method is illustrated in Fig. 3.", + "bbox": [ + 212, + 750, + 787, + 842 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "ST-LLM: Large Language Models Are Effective Temporal Learners", + "bbox": [ + 279, + 114, + 727, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Video Tokens Inside LLM", + "text_level": 1, + "bbox": [ + 215, + 146, + 477, + 161 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As depicted in Fig. 3, visually, ST-LLM closely resembles an image chatbot, consisting of a visual encoder, a linear projection, and an LLM. Instead of opting for the commonly used CLIP-L/14 [38], we have selected BLIP-2 [23] as our visual encoder, whose Q-Former efficiently compresses redundant image tokens into fewer visual tokens. To facilitate the input of raw videos for the image encoder, we adopt the strategy of treating each frame as an individual image. Given a video $v$ with $T$ frames, we assume that the image encoder $f_{\\mathrm{v}}$ encodes each frame into $K$ tokens. Hence, we obtain all visual tokens $V$ to be fed into the LLM as follows:", + "bbox": [ + 212, + 169, + 787, + 303 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nV = f _ {\\mathrm {p}} (f _ {\\mathrm {v}} (v)), \\mathrm {w h e r e} V = \\{V _ {i} \\} _ {i = 1} ^ {T}, V _ {i} = \\{v _ {(i, j)} \\} _ {j = 1} ^ {K}, \\qquad (1)\n$$\n", + "text_format": "latex", + "bbox": [ + 307, + 313, + 785, + 333 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $f_{\\mathrm{p}}$ is the visual projection layer. Then, we concatenate all visual tokens to form a joint spatial-temporal [7] visual sequence $V \\in \\mathbb{R}^{(T*K)\\times D}$ , where $D$ is the number of embedding dimensions. After being passed through the word embedding layers, the text tokens $C = \\{c_i\\}_{i=1}^N$ are directly concatenated with all spatial-temporal tokens, to organize the input $I$ to the LLM as follows:", + "bbox": [ + 212, + 339, + 787, + 416 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nI = \\left[ V: C \\right] = \\left\\{v _ {(1, 1)}, v _ {(1, 2)}, \\dots , v _ {(1, K)}, v _ {(2, 1)}, \\dots , v _ {(T, K)}, c _ {1}, c _ {2}, \\dots , c _ {N} \\right\\}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 272, + 425, + 785, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Indeed, several alternative strategies exist for organizing the input tokens before feeding them into the LLM. These include: (1) Adding separator tokens between frame-wise tokens or visual-text tokens, to demarcate the boundaries between modalities or frames. (2) Adding spatial-temporal position embeddings specifically for visual tokens, to facilitate better spatial-temporal relationship understandings. However, our experiments indicate that simplicity often yields the best. Adding numerous separator tokens introduces additional overhead without necessarily improving performance. Moreover, LLMs come with Rotary Position Embeddings [39] that are already effective in distinguishing the positions of all spatial-temporal tokens and text tokens.", + "bbox": [ + 212, + 449, + 787, + 602 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Training with Dynamic Masking", + "text_level": 1, + "bbox": [ + 215, + 619, + 532, + 637 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "With video tokens modeling inside LLM, we can leverage the potent sequence modeling capabilities inherent in LLMs to understand the information encapsulated within spatial-temporal sequences. However, integrating all video tokens into LLM significantly increases the context length, imposing a considerable computational burden and rendering it unfeasible to input additional frames for lengthy videos. Meanwhile, experimental findings suggest that during testing, utilizing uncompressed tokens is more sensitive to variations in the number of frames than mean pooling. Consequently, when a substantial dissonance exists between the number of frames in training and testing, a noticeable degradation in performance ensues.", + "bbox": [ + 212, + 643, + 787, + 794 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To tackle the aforementioned issues, we first propose masking the visual tokens during training. Masking modeling has achieved significant success in natural language processing [15] and video-language pretraining [26,31,42]. Although", + "bbox": [ + 212, + 795, + 787, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Liu. et al.", + "bbox": [ + 271, + 114, + 339, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "masking is rarely employed in autoregressive LLMs, given our aim to leverage LLMs for encoding spatial-temporal sequences, we can still utilize masking modeling. Specifically, we keep the text tokens in $I$ unchanged, while applying a mask to the video tokens. This mask randomly masks $\\rho$ of all video tokens, irrespective of their position across different frames. Different from previous works [26,31], we adopt a dynamic masking strategy, where the masking rate is randomly sampled from a normal distribution as follows:", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\rho \\sim \\mathbb {N} (0. 5, \\sigma), \\quad 0. 3 \\leq \\rho \\leq 0. 7, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 397, + 263, + 785, + 280 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\sigma$ represents the variance of the normal distribution. This strategy ensures that the length of the spatial-temporal sequence varies continuously while maintaining a consistently high average masking rate of $50\\%$ . As a result, it minimizes training costs and notably enhances robustness during inference.", + "bbox": [ + 212, + 290, + 787, + 349 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Furthermore, building upon dynamic masking, we have formulated the Masked Video Modeling (MVM) objective to encourage the LLM to grasp spatial-temporal dependencies. Unlike the expensive masked token recovery [6,37], we have opted for unmasked token reconstruction. Specifically, besides the masked sequence $I$ , we conduct an extra non-grad forward pass for the unmasked sequence $\\hat{I} = [\\hat{V}:C]$ , serving as the reference output. Then, we select unmasked tokens from $f_{llm}(I)$ and the corresponding tokens from $f_{llm}(\\hat{I})$ based on their positions, computing the Mean Squared Error (MSE) between the selected pairs as follows:", + "bbox": [ + 212, + 351, + 792, + 470 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {m v m} = \\frac {1}{(1 - \\rho) K T} \\sum_ {i = 1} ^ {T} \\sum_ {j = 1} ^ {K} \\left(v _ {i, j} ^ {- 1} - \\hat {v} _ {i, j} ^ {- 1}\\right) ^ {2}, (i, j) \\notin M, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 310, + 484, + 785, + 523 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $M$ denotes the set of masked tokens' indices, $v_{i,j}^{-1}$ is from $f_{llm}(I)$ , and $\\mathcal{L}_{mvm}$ represents our MVM objective. Finally, our overall loss $\\mathcal{L}$ is composed of $\\mathcal{L}_{mvm}$ and the LLM decoder loss $\\mathcal{L}_{llm}$ , yielding $\\mathcal{L} = \\mathcal{L}_{mvm} + \\mathcal{L}_{llm}$ . By integrating these two loss components, we can encourage the LLM to effectively respond to questions derived from the video content while improving its ability to model temporal and spatial dependencies.", + "bbox": [ + 214, + 536, + 787, + 628 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Global-Local Input", + "text_level": 1, + "bbox": [ + 215, + 648, + 419, + 666 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Although dynamic masking partially mitigates the challenge of processing lengthy input sequences, it fails to fully resolve the issue for extremely long videos that entail inputting numerous frames, rendering the context length still impractical to handle. Therefore, we have devised an additional module to tackle the issue of excessively long videos. To elaborate, given a lengthy video with a large $T$ , we still commence by encoding each frame individually to derive $V$ . Subsequently, we proceed to derive the global video representation $V_{0}$ through average pooling on the frame-wise tokens:", + "bbox": [ + 212, + 672, + 787, + 792 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nv _ {(0, j)} = \\frac {1}{T} \\sum_ {i = 1} ^ {T} v _ {(i, j)}, \\quad V _ {0} = \\left\\{v _ {(0, j)} \\right\\} _ {j = 1} ^ {K}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 364, + 803, + 785, + 843 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "ST-LLM: Large Language Models Are Effective Temporal Learners", + "bbox": [ + 277, + 114, + 727, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Next, we average-sample $\\overline{T}$ frames from the T total frames. All tokens from these $\\overline{T}$ frames are concatenated to produce a joint spatial-temporal sequence $\\overline{V} \\in \\mathbb{R}^{(\\overline{T}*K) \\times D}$ , which serves as the local video representation. Finally, the global-local input for the LLM, denoted as $\\overline{I}$ , is constructed as follows:", + "bbox": [ + 212, + 146, + 787, + 209 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {I} = \\left[ \\bar {V} + f _ {m} \\left(V _ {0}\\right): C \\right], \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 424, + 220, + 785, + 238 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $f_{m}$ is a simple MLP projector with upsampling projection initialized with zeros. With this global-local input design, the low fps spatial-temporal sequences within the LLM can gradually incorporate information from the high fps branch. This approach allows the model to benefit from the LLM's ability to model temporal sequences within a limited context while also considering the global information of long videos.", + "bbox": [ + 212, + 250, + 787, + 340 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 364, + 375, + 382 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we have performed comprehensive experimental evaluations of ST-LLM, covering crucial settings, comparisons, and ablation studies. For a more detailed account of experimental settings, ablation studies, visualizations, and limitations analysis, please consult the appendix.", + "bbox": [ + 212, + 397, + 789, + 459 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Experiment Setup", + "text_level": 1, + "bbox": [ + 214, + 481, + 413, + 497 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Benchmarks. To evaluate the video understanding capabilities of ST-LLM, we primarily conduct assessments on three benchmarks: MVBench [25], VideoChat-GPT Bench [33], and zero-shot video QA benchmark. MVBench comprises 20 challenging video tasks, each consisting of 200 samples in the form of multiple-choice questions. These tasks provide a comprehensive and objective assessment of a model's ability to understand videos. VideoChatGPT-Bench gathers videos from ActivityNet [10] and uses GPT to evaluate the quality of video conversations across five dimensions. Additionally, the zero-shot video QA benchmark entails GPT-based assessments of various open-source video QA datasets.", + "bbox": [ + 212, + 507, + 787, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Implementation Details. In our study, we adopt state-of-the-art image dialogue models as the baseline and proceed directly with video instruction tuning. Unless otherwise specified, our model is initialized with InstructBLIP [14] in all experiments, whose LLM is Vicuna-v1.1 [13] with 7B parameters. Additionally, we employ Minigpt4-v1 [55] as a basic model to assess the effectiveness across various image models in Table 6. Following previous research [20, 27, 28], we adopt full-finetuning on the LLM to ensure a fair comparison. The results of LoRA are also included as ablation results in Table 6. $\\sigma$ is set as 0.1, ensuring that over $95\\%$ of the mask rate falls within the range of 0.3 to 0.7. During training, we sample 16 frames per video. Considering dynamic masking, an average of 256 tokens per video are input into the LLM, which is the same as LLaVA [29]. For all benchmarks during inference, frames are sampled at a frame rate of 1, and the number of local frames is kept between 4 and 16. This strategy closely", + "bbox": [ + 212, + 643, + 789, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Liu. et al.", + "bbox": [ + 271, + 114, + 339, + 127 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d084cb1979183c3dded291089c47e447b719cc4161f89b7bd248809ab8edc491.jpg", + "table_caption": [ + "Table 1: Comparisons on MVBench. Except for BLIP2 and Otter, all models are built upon LLaMA-1-7B for fair comparisons. \"Avg\" denotes the average of all 20 metrics. The leaderboards for each task can be found in the Appendix. The compared methods include: (1)Rd: random guesses; (2) mOI: mPLUG-Owl-I 51; (3) LMA: LLaMA-Adapter 54 (4)B2: BLIP-2 23; (5)OI: Otter-I 22; (6)MG: MiniGPT-4 55; (7)IB: InstructBLIP 14; (8)LV: LLaVA 29; (9)OV: Otter-V 22; (10)mOV: mPLUG-Owl-V 51; (11)VCG: VideoChatGPT 33; (12)VLM: VideoLLaMA 53; (13)VC: VideoChat 24; (14)VC2: VideoChat2 25." + ], + "table_footnote": [], + "table_body": "
MetricRdmOILMAB2OIMGIBLVOVmOVVCGVLMVCVC2Ours
AS25.025.023.024.534.516.020.028.023.022.023.527.533.566.066.0
AP25.020.028.029.032.018.016.539.523.028.026.025.526.547.553.5
AA33.344.551.033.539.526.046.063.027.534.062.051.050.083.584.0
FA25.027.030.017.030.521.524.530.527.029.022.529.033.549.544.0
UA25.023.533.042.038.516.046.039.029.529.026.539.040.560.058.5
OE33.336.053.551.548.529.551.053.053.040.554.048.053.058.080.5
OI25.024.032.526.044.025.526.041.028.027.028.040.540.571.573.5
OS33.334.033.531.029.513.037.541.533.031.540.038.030.042.538.5
MD25.023.025.525.519.011.522.023.024.527.023.022.525.523.042.5
AL25.024.021.526.025.512.023.020.523.523.020.022.527.023.031.0
ST25.034.530.532.555.09.546.545.027.529.031.043.048.588.586.5
AC33.334.529.025.520.032.542.534.026.031.530.534.035.039.036.5
MC25.022.022.530.032.515.526.520.528.527.025.522.520.542.056.5
MA33.331.541.540.028.58.040.538.518.040.039.532.542.558.578.5
SC33.340.039.542.039.034.032.047.038.544.048.545.546.044.043.0
FP25.024.025.027.028.026.025.525.022.024.029.032.526.549.044.5
CO33.337.031.530.027.029.530.536.022.031.033.040.041.036.546.5
EN25.025.522.526.032.019.030.527.023.526.029.530.023.535.034.5
ER20.021.028.037.036.59.930.526.519.520.526.021.023.540.541.5
CI30.937.032.031.036.53.038.042.019.029.535.537.036.065.558.5
Avg27.329.431.731.433.518.832.536.026.829.732.734.135.551.154.9
", + "bbox": [ + 220, + 267, + 781, + 584 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "aligns with practical scenarios and involves fewer frames than most video LLMs. For VideoChatGPT-Bench and the video QA benchmark, which may involve videos spanning several minutes, we utilize the global-local input approach. In this configuration, the global frame number is set to 64. All videos are resized into $224^{*}224$ . We train ST-LLM using a learning rate of 2e-5 and a batch size of 128 for 2 epochs.", + "bbox": [ + 212, + 613, + 787, + 703 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Training Datasets. Citing findings from Jin et al. [20], the sequential process of first conducting image instruction tuning, followed by video instruction tuning and image-video joint pretraining, had a negligible impact on video dialogue performance. Hence, in light of InstructBLIP's prior training on extensive image data, we exclusively focus on video instruction tuning. Following [25], we leverage video instruction data sourced from a variety of datasets, including VideoChatGPT-100k [33], VideoChat-11k [24], Webvid [5], NExT-QA [47], CLEVRER [52], Kinetics-710 [21], and Something-Something-2 [17]. All samples are uniformly formatted in accordance with [25]. More details about the", + "bbox": [ + 212, + 704, + 787, + 842 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "ST-LLM: Large Language Models Are Effective Temporal Learners", + "bbox": [ + 279, + 114, + 727, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/429a145e438720229240fc03027e0438e3fa4942fc5f6f5fc0976df9e719cd45.jpg", + "table_caption": [ + "Table 2: Comparisons on video-based generative performance benchmarking." + ], + "table_footnote": [], + "table_body": "
MethodCorrectDetailContextTemporalConsistMean Score
VideoLLaMA 531.962.182.161.821.791.98
LLaMA-Adapter 542.032.322.301.982.152.16
VideoChat 242.232.502.531.942.242.29
VideoChatGPT 332.402.522.621.982.372.38
BT-Adapter 312.682.693.272.342.462.69
VTimeLLM 192.783.103.402.492.472.85
Chat-UniVi 202.892.913.462.892.812.99
LLaMA-VID-7B 272.963.003.532.462.512.89
LLaMA-VID-13B 273.073.053.602.582.632.99
VideoChat2 253.022.883.512.662.812.98
ST-LLM (Ours)3.233.053.742.932.813.15
", + "bbox": [ + 233, + 171, + 764, + 349 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "dataset and the organization of instructions can be found in the appendix. An experiment on joint image-video training is also presented in the appendix.", + "bbox": [ + 212, + 375, + 787, + 405 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Quantitative Result", + "text_level": 1, + "bbox": [ + 214, + 424, + 426, + 439 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "MVBench Performance. The evaluation results on MVBench are presented in Table 1. The evaluation metrics included Action Sequence (AS), Action Prediction (AP), Action Antonym (AA), Fine-grained Action (FA), Unexpected Action (UA), Object Existence (OE), Object Interaction (OI), Object Shuffle (OS), Moving Direction (MD), Action Localization (AL), Scene Transition (ST), Action Count (AC), Moving Count (MC), Moving Attribute (MA), State Change (SC), Fine-grained Pose (FP), Character Order (CO), Egocentric Navigation (EN), Episodic Reasoning (ER), Counterfactual Inference (CI), and the average of all 20 metrics (Avg). It is evident that on such a challenging benchmark, the performance of most multimodal LLMs falls significantly short of satisfactory levels, even barely surpassing random performance by less than $10\\%$ . Our only formidable competitor is VideoChat2 [25]. Through direct substitution of CLIP with a dedicated video encoder, coupled with a three-stage training process and comprehensive instruction tuning for the visual encoder, Q-Former, and LLM, VideoChat2 has yielded results that significantly surpass those of previous models. In contrast, we solely carried out single-stage video-only instruction tuning while maintaining the weights of CLIP and Q-Former frozen. Even under these conditions, ST-LLM still attained the top performance, with an average score surpassing VideoChat2 by a notable margin of $3.8\\%$ . Specifically, ST-LLM showcased leading results across all 11 tasks, exhibiting a particularly significant advantage in motion-related tasks where other models struggled. In three motion-related tasks, ST-LLM achieved an average score of $59.2\\%$ , far outperforming VideoChat2's $36.3\\%$ . However, ST-LLM encounters challenges in fine-grained tasks, such as Fine-grained Action (FA) and Fine-grained Pose (FP). This could be attributed to CLIP's limitations as an image encoder, which may struggle to capture low-level spatial-temporal features required for these tasks.", + "bbox": [ + 212, + 446, + 787, + 843 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Liu. et al.", + "bbox": [ + 271, + 114, + 339, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/f0663376c0d98cf89a848533b2a89b0d4e529e0702909502b4f8c8328bf44d52.jpg", + "table_caption": [ + "Table 3: Comparisons on zero-shot question-answering, including MSVD-QA [46], MSRVTT-QA [48], and ActivityNet-QA [10]." + ], + "table_footnote": [], + "table_body": "
MethodLLMMSVD-QAMSRVTT-QAActivityNet-QA
AccScoreAccScoreAccScore
FrozenBiLM 50DeBERTa-V232.2-16.8-24.7-
VideoLLaMA 53Vicuna-7B51.62.529.61.812.41.1
LLaMA-Adapter 54LLaMA-7B54.93.143.82.734.22.7
VideoChat 24Vicuna-7B56.32.845.02.526.52.2
VideoChatGPT 33Vicuna-7B64.93.349.32.835.22.7
BT-Adapter 31Vicuna-7B67.53.757.03.245.73.2
Chat-UniVi 20Vicuna-7B65.03.654.63.145.83.2
LLaMA-VID 27Vicuna-7B69.73.757.73.247.43.3
LLaMA-VID 27Vicuna-13B70.03.758.93.347.53.3
VideoChat2 25Vicuna-7B70.03.954.13.349.13.3
ST-LLM (Ours)Vicuna-7B74.63.963.23.450.93.3
", + "bbox": [ + 251, + 185, + 750, + 354 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "VideoChatGPT-Bench Performance. VideoChatGPT-Bench, also referred to as the Video-Based Generative Performance Benchmark, primarily evaluates the capability of video conversation. Utilizing GPT-3.5 [34], it assigns a score to the generated content across five aspects: Correctness of Information (Correct), Detail Orientation (Detail), Contextual Understanding (Context), Temporal Understanding (Temporal), and Consistency (Consist). As depicted in Table 2, the recently proposed state-of-the-art models [20,25,27] exhibit similar performance levels on this benchmark. However, due to the inherent instability of GPT evaluation, there is considerable variation in the performance of each model across different tasks. Despite this variability, ST-LLM consistently performs admirably across all tasks, while demonstrating significant advantages in terms of the mean score compared to previous models.", + "bbox": [ + 212, + 383, + 787, + 564 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Zero-shot Video-question Answering Performance. In Table 3 we present the zero-shot video question-answering performance on several commonly used video-text datasets, including MSVD-QA [46], MSRVTT-QA [48], and ActivityNet-QA [10]. Following the methodology outlined in [33], we utilize GPT-3.5 to evaluate the accuracy and score of the generated results. As is depicted, STLLM demonstrates significant advantages on MSVD and MSRVTT, surpassing the previous SOTAs by $4.6\\%$ and $4.3\\%$ respectively. However, STLLM exhibits a slightly smaller advantage compared to VideoChat2 on the ActivityNet. This could be attributed to the requirement for a potent video encoder for this dataset.", + "bbox": [ + 212, + 565, + 800, + 702 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Ablations and Analysis", + "text_level": 1, + "bbox": [ + 215, + 724, + 454, + 739 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Ways of Inputting video tokens. The core of this work lies in modeling joint spatial-temporal video tokens using LLM. Table 4 presents the results of different methods of inputting video tokens. As depicted, without video tuning, the approach of joint spatial-temporal input is considerably less effective compared to directly averaging the input over the temporal dimension. This discrepancy might stem from inconsistencies between training and testing. However, after", + "bbox": [ + 212, + 748, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "ST-LLM: Large Language Models Are Effective Temporal Learners", + "bbox": [ + 277, + 113, + 727, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/7524b810a2dbbf3141325600a173f20aaf66a8b12f2fd243717feea30eb674f3.jpg", + "table_caption": [ + "Table 4: The ablation study on the methods of inputting video tokens to LLMs. The baseline configurations include no instruction tuning and mean pooling input." + ], + "table_footnote": [], + "table_body": "
MethodAvg on MV bench
Instructclip Baseline
Mean Pooling42.0
S-T Tokens in LLM35.5
Video Instruction Tuning
Mean Pooling47.8
S-T Tokens in LLM53.8
+Masking & MVM Loss54.9
", + "bbox": [ + 220, + 219, + 503, + 335 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/d51c3b50eefedfff5883c76f9d933ac6618712042adc1ba0dff8d821206d2ff2.jpg", + "image_caption": [ + "Fig. 4: The ablation study on the effect of dynamic masking and MVM loss." + ], + "image_footnote": [], + "bbox": [ + 517, + 143, + 781, + 296 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/bbf33dae587544bdf54c022b3c9f3d0b5c5930ddfd6a5f667ff3399ba79ffe1a.jpg", + "table_caption": [ + "Table 5: The ablation study on global-local input module. The \"Local\" and \"Global\" mean the joint temporal-temporal sequence and average pooling sequence respectively." + ], + "table_footnote": [], + "table_body": "
MethodAvg on MVBenchVideo-ChatGPT Bench
CorrectDetailContextTemporalConsist
Global Only48.33.012.923.582.612.61
Local Only54.93.082.993.632.752.69
Local+Global (simply add)50.13.072.883.582.592.50
Local+Global (adapter)54.73.233.053.742.932.81
", + "bbox": [ + 220, + 392, + 782, + 477 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "training, the effectiveness of the joint spatial-temporal input significantly surpasses that of the mean pooling input, indicating that LLMs can effectively model spatial-temporal sequences, and directly inputting all tokens is a superior approach compared to compression before input. Furthermore, incorporating dynamic masking and the MVM loss based on the joint spatial-temporal sequence has further improved the performance of our model.", + "bbox": [ + 212, + 507, + 787, + 597 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Dynamic Masking Strategy and MVM Loss. In Fig. 4, we conduct a more detailed ablation study on the effect of dynamic masking and MVM loss. It is evident that when the training frame count remains fixed, the performance of inputting all video tokens is not very robust. Particularly, when there is a discrepancy between the training and testing frame, the performance noticeably decreases. Meanwhile, when we randomly mask video tokens at a static $50\\%$ rate, this issue persists, with effectiveness akin to training with a fixed 8-frame setup. However, employing dynamic masking shows some improvement in overall performance, and notably, the robustness to varying testing frame counts is significantly enhanced. Furthermore, the inclusion of the MVM loss built upon the masking further enhances performance, with greater improvements observed as the sequence length increases.", + "bbox": [ + 212, + 598, + 787, + 777 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Global-Local Input Module. The ablation results on the global-local module are presented in Table 5 where we compare the performance of global features extracted from more frames and local features from fewer frames, as well as their combination. As shown, on MVBench, the performance of global features", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Liu. et al.", + "bbox": [ + 271, + 114, + 339, + 127 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/2009037448447e6fd39ffb4cacebb25750228b3d9192dfaaf19507444b6f6ac2.jpg", + "table_caption": [ + "Table 6: The ablation study on the basic image models and the fine-tuning methodologies of LLM. We report the Avg on MVBench." + ], + "table_footnote": [], + "table_body": "
Baseline ModelLLMLLM ParametersMeanpooling BaselineST-LLM Full Model
MiniGPT-4 55Vicuna-7B v0LoRA40.945.6
Full Fine-tuning46.450.5
InstructBLIP 14Vicuna-7B v1.1LoRA45.251.9
Full Fine-tuning47.854.9
", + "bbox": [ + 220, + 185, + 782, + 253 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/373904aeac22fcd8ea70df2e6c853eff7f98a02544de22620029c18a2492b4f5.jpg", + "table_caption": [ + "Table 7: The analysis on the designs of inputting S-T video tokens into LLM." + ], + "table_footnote": [], + "table_body": "
MethodAvg on MVBench
S-T Tokens in LLM (baseline)54.9
+frame-frame separators54.0
+video-text separators54.8
+S-T position embeddings52.2
", + "bbox": [ + 220, + 311, + 521, + 385 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/6fe0f45bc7f2e23d18f8316b75ffab4ff1fa7bf246f12444ec11b25ac1b5095f.jpg", + "table_caption": [ + "Table 8: The analysis on the designs of dynamic masking." + ], + "table_footnote": [], + "table_body": "
MethodAvg on MVBatch
w/o masking (baseline)53.8
ρ ~ U(0.3, 0.7)54.2
ρ ~ N(0.5, 0.2)54.3
ρ ~ N(0.5, 0.1)54.9
", + "bbox": [ + 531, + 311, + 785, + 385 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "is significantly inferior to the joint S-T features, so combining global and local features did not lead to improvement. This may be due to the fact that MVBench mainly consists of short videos lasting a few to a dozen seconds, where a large number of frames might not be necessary. However, on VideoChatGPT Bench, where videos are parsed at the minute level, the discrepancy between global and local features has notably diminished, with their amalgamation resulting in a discernible enhancement. This underscores the efficacy of the Global-Local Input module in processing extended video content. Moreover, the direct incorporation of global and local features resulted in a performance decline, presumably due to its interference with the representation learning of the LLM.", + "bbox": [ + 212, + 414, + 787, + 566 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Basic Image Models & LLMs. In Table 6, we conduct ablations on both the basic image chatbots and the LLM fine-tuning strategy. As depicted in the table, regardless of whether it's LoRA [18] tuning or full fine-tuning, or different basic image models, ST-LLM demonstrates notable enhancements compared to the mean pooling baseline. This suggests that our approach can be adeptly applied across diverse scenarios, such as employing more resource-efficient LoRA tuning on larger models, or transferring to more potent image-dialogue models.", + "bbox": [ + 212, + 566, + 787, + 672 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Designs of ST-LLM. In Table 7 and Table 8, we investigated several alternative design possibilities for ST-LLM. Primarily, we explored various separator designs. It's noteworthy that almost all cross-modal LLMs incorporate separator designs, with some models even integrating separators between multiple images or frames [2,4]. However, we discovered that when employing LLMs to model spatial-temporal sequences, these separators are superfluous. Subsequently, we address positional encodings. Numerous approaches integrate temporal positional encodings before inputting video tokens into LLMs [32,33,53]. However, in our methodology, we observed that incorporating absolute positional encodings to all visual tokens significantly impairs performance. Lastly, we consider the design of dynamic masking. Our findings suggest that all distributions", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "ST-LLM: Large Language Models Are Effective Temporal Learners", + "bbox": [ + 279, + 113, + 727, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/42f1b8056b5f8ef0be156a6cf389d6863825d969515d934706bc96954dbd8798.jpg", + "image_caption": [ + "Fig. 5: Qualitative results from real-world videos." + ], + "image_footnote": [], + "bbox": [ + 223, + 146, + 779, + 433 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "of masking rates can yield some improvement, yet distributions with smaller standard deviations exhibit superior performance.", + "bbox": [ + 212, + 489, + 787, + 521 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "4.4 Qualitative Results", + "text_level": 1, + "bbox": [ + 214, + 540, + 421, + 555 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In Fig. 5, we present a qualitative comparison. As illustrated, ST-LLM excels in adhering to instructions and delivering precise responses. More importantly, ST-LLM showcases superior sensitivity to temporal sequences and actions, thereby validating the effectiveness of our modeling approach. For further quantitative results and failure cases, please refer to the appendix.", + "bbox": [ + 212, + 563, + 787, + 640 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 659, + 359, + 676 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we present ST-LLM, a straightforward yet robust video large language model. Our aim is to achieve effective video comprehension by leveraging LLM to model video tokens, marking the inception of a joint spatial-temporal-text modeling paradigm. Additionally, we introduce a dynamic masking strategy and a global-local input module to enhance this framework. Remarkably, ST-LLM excels in encoding and comprehending spatial-temporal sequences while addressing concerns related to efficiency, stability, and modeling lengthy videos, all with reduced training resource requirements. Through extensive experimental analysis, we demonstrate the effectiveness of ST-LLM and its internal design, resulting in state-of-the-art performance across multiple video LLM benchmarks.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Liu. et al.", + "bbox": [ + 271, + 114, + 339, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported by National Natural Science Foundation of China (No. 62172021). We are also grateful to the support from Guangdong Provincial Key Laboratory of Ultra High Definition Immersive Media Technology.", + "bbox": [ + 212, + 146, + 787, + 205 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "*: Equal Contribution, work done during internship at ARC Lab, Tencent PCG.", + "bbox": [ + 240, + 207, + 774, + 220 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 239, + 321, + 256 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)", + "2. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems 35, 23716-23736 (2022)", + "3. Anil, R., Dai, A.M., First, O., Johnson, M., Lepikhin, D., Passos, A., Shakeri, S., Taropa, E., Bailey, P., Chen, Z., et al.: Palm 2 technical report. arXiv preprint arXiv:2305.10403 (2023)", + "4. Awadalla, A., Gao, I., Gardner, J., Hessel, J., Hanafy, Y., Zhu, W., Marathe, K., Bitton, Y., Gadre, S., Sagawa, S., et al.: Openflamingo: An open-source framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390 (2023)", + "5. Bain, M., Nagrani, A., Varol, G., Zisserman, A.: Frozen in time: A joint video and image encoder for end-to-end retrieval. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1728-1738 (2021)", + "6. Bao, H., Dong, L., Piao, S., Wei, F.: Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254 (2021)", + "7. Bertasius, G., Wang, H., Torresani, L.: Is space-time attention all you need for video understanding? In: ICML. vol. 2, p. 4 (2021)", + "8. Brooks, T., Peebles, B., Holmes, C., DePue, W., Guo, Y., Jing, L., Schnurr, D., Taylor, J., Luhman, T., Luhman, E., Ng, C., Wang, R., Ramesh, A.: Video generation models as world simulators (2024), https://openai.com/research/video-generation-models-as-world-simulators", + "9. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)", + "0. Caba Heilbron, F., Escorcia, V., Ghanem, B., Carlos Niebles, J.: Activitynet: A large-scale video benchmark for human activity understanding. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 961-970 (2015)", + "1. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6299-6308 (2017)", + "2. Chen, J., Zhu, D., Shen, X., Li, X., Liu, Z., Zhang, P., Krishnamoorthi, R., Chandra, V., Xiong, Y., Elhoseiny, M.: Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478 (2023)", + "3. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., et al.: Vicuna: An open-source chatbot impressing gpt-4 with $90\\%$ chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023) (2023)" + ], + "bbox": [ + 225, + 267, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "ST-LLM: Large Language Models Are Effective Temporal Learners", + "bbox": [ + 279, + 114, + 727, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Dai, W., Li, J., Li, D., Tiong, A., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: towards general-purpose vision-language models with instruction tuning. arxiv. Preprint posted online on June 15, 2023 (2023)", + "15. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)", + "16. Driess, D., Xia, F., Sajjadi, M.S., Lynch, C., Chowdhery, A., Ichter, B., Wahid, A., Tompson, J., Vuong, Q., Yu, T., et al.: Palm-e: An embodied multimodal language model. arXiv preprint arXiv:2303.03378 (2023)", + "17. Goyal, R., Ebrahimi Kahou, S., Michalski, V., Materzynska, J., Westphal, S., Kim, H., Haenel, V., Fruend, I., Yianilos, P., Mueller-Freitag, M., et al.: The\" something something\" video database for learning and evaluating visual common sense. In: Proceedings of the IEEE international conference on computer vision. pp. 5842-5850 (2017)", + "18. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models (2021)", + "19. Huang, B., Wang, X., Chen, H., Song, Z., Zhu, W.: Vtimellm: Empower llm to grasp video moments. arXiv preprint arXiv:2311.18445 (2023)", + "20. Jin, P., Takanobu, R., Zhang, C., Cao, X., Yuan, L.: Chat-univi: Unified visual representation empowers large language models with image and video understanding. arXiv preprint arXiv:2311.08046 (2023)", + "21. Kay, W., Carreira, J., Simonyan, K., Zhang, B., Hillier, C., Vijayanarasimhan, S., Viola, F., Green, T., Back, T., Natsev, P., et al.: The kinetics human action video dataset. arXiv preprint arXiv:1705.06950 (2017)", + "22. Li, B., Zhang, Y., Chen, L., Wang, J., Yang, J., Liu, Z.: Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726 (2023)", + "23. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023)", + "24. Li, K., He, Y., Wang, Y., Li, Y., Wang, W., Luo, P., Wang, Y., Wang, L., Qiao, Y.: Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355 (2023)", + "25. Li, K., Wang, Y., He, Y., Li, Y., Wang, Y., Liu, Y., Wang, Z., Xu, J., Chen, G., Luo, P., et al.: Mvbench: A comprehensive multi-modal video understanding benchmark. arXiv preprint arXiv:2311.17005 (2023)", + "26. Li, K., Wang, Y., Li, Y., Wang, Y., He, Y., Wang, L., Qiao, Y.: Unmasked teacher: Towards training-efficient video foundation models. arXiv preprint arXiv:2303.16058 (2023)", + "27. Li, Y., Wang, C., Jia, J.: Llama-vid: An image is worth 2 tokens in large language models. arXiv preprint arXiv:2311.17043 (2023)", + "28. Lin, B., Zhu, B., Ye, Y., Ning, M., Jin, P., Yuan, L.: Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122 (2023)", + "29. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv preprint arXiv:2304.08485 (2023)", + "30. Liu, R., Huang, J., Li, G., Feng, J., Wu, X., Li, T.H.: Revisiting temporal modeling for clip-based image-to-video knowledge transferring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6555-6564 (2023)", + "31. Liu, R., Li, C., Ge, Y., Shan, Y., Li, T.H., Li, G.: One for all: Video conversation is feasible without video instruction tuning. arXiv preprint arXiv:2309.15785 (2023)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Liu. et al.", + "bbox": [ + 271, + 114, + 336, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "32. Luo, R., Zhao, Z., Yang, M., Dong, J., Qiu, M., Lu, P., Wang, T., Wei, Z.: Valley: Video assistant with large language model enhanced ability. arXiv preprint arXiv:2306.07207 (2023)", + "33. Maaz, M., Rasheed, H., Khan, S., Khan, F.S.: Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424 (2023)", + "34. OpenAI: Chatgpt. https://openai.com/blog/chatgpt (2023)", + "35. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems 35, 27730-27744 (2022)", + "36. Peebles, W., Xie, S.: Scalable diffusion models with transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4195-4205 (2023)", + "37. Peng, Z., Dong, L., Bao, H., Ye, Q., Wei, F.: Beit v2: Masked image modeling with vector-quantized visual tokenizers. arXiv preprint arXiv:2208.06366 (2022)", + "38. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)", + "39. Su, J., Ahmed, M., Lu, Y., Pan, S., Bo, W., Liu, Y.: Roformer: Enhanced transformer with rotary position embedding. Neurocomputing 568, 127063 (2024)", + "40. Sun, Q., Fang, Y., Wu, L., Wang, X., Cao, Y.: Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389 (2023)", + "41. Taori, R., Gulrajani, I., Zhang, T., Dubois, Y., Li, X., Guestrin, C., Liang, P., Hashimoto, T.B.: Stanford alpaca: An instruction-following llama model (2023)", + "42. Tong, Z., Song, Y., Wang, J., Wang, L.: Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. Advances in neural information processing systems 35, 10078-10093 (2022)", + "43. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)", + "44. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023)", + "45. Tran, D., Bourdev, L., Fergus, R., Torresani, L., Paluri, M.: Learning spatiotemporal features with 3d convolutional networks. In: Proceedings of the IEEE international conference on computer vision. pp. 4489-4497 (2015)", + "46. Wu, Z., Yao, T., Fu, Y., Jiang, Y.G.: Deep learning for video classification and captioning. In: Frontiers of multimedia research, pp. 3-29. ACM (2017)", + "47. Xiao, J., Shang, X., Yao, A., Chua, T.S.: Next-qa: Next phase of question-answering to explaining temporal actions. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9777-9786 (2021)", + "48. Xu, J., Mei, T., Yao, T., Rui, Y.: Msr-vtt: A large video description dataset for bridging video and language. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5288-5296 (2016)", + "49. Xue, H., Sun, Y., Liu, B., Fu, J., Song, R., Li, H., Luo, J.: Clip-vip: Adapting pre-trained image-text model to video-language representation alignment. arXiv preprint arXiv:2209.06430 (2022)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "ST-LLM: Large Language Models Are Effective Temporal Learners", + "bbox": [ + 279, + 114, + 727, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "50. Yang, A., Miech, A., Sivic, J., Laptev, I., Schmid, C.: Zero-shot video question answering via frozen bidirectional language models. Advances in Neural Information Processing Systems 35, 124-141 (2022)", + "51. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178 (2023)", + "52. Yi, K., Gan, C., Li, Y., Kohli, P., Wu, J., Torralba, A., Tenenbaum, J.B.: Clevrer: Collision events for video representation and reasoning. arXiv preprint arXiv:1910.01442 (2019)", + "53. Zhang, H., Li, X., Bing, L.: Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858 (2023)", + "54. Zhang, R., Han, J., Zhou, A., Hu, X., Yan, S., Lu, P., Li, H., Gao, P., Qiao, Y.: Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199 (2023)", + "55. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023)" + ], + "bbox": [ + 215, + 146, + 787, + 383 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Liu. et al.", + "bbox": [ + 271, + 114, + 336, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/ST-LLM_ Large Language Models Are Effective Temporal Learners/f824889a-c655-4483-9dd5-c5cb83e43784_model.json b/2024/ST-LLM_ Large Language Models Are Effective Temporal Learners/f824889a-c655-4483-9dd5-c5cb83e43784_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2a9b2d1b23e2e8b790cdbb40c41bd939880d6c66 --- /dev/null +++ b/2024/ST-LLM_ Large Language Models Are Effective Temporal Learners/f824889a-c655-4483-9dd5-c5cb83e43784_model.json @@ -0,0 +1,2205 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.23, + 0.141, + 0.776, + 0.187 + ], + "angle": 0, + "content": "ST-LLM: Large Language Models Are Effective Temporal Learners" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.213, + 0.785, + 0.229 + ], + "angle": 0, + "content": "Ruyang Liu\\*1,3 Chen Li\\*2 Haoran Tang1,3 Yixiao Ge2 Ying Shan2 Ge Li" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.24, + 0.763, + 0.254 + ], + "angle": 0, + "content": "1 School of Electronic and Computer Engineering, Shenzhen Graduate School," + }, + { + "type": "text", + "bbox": [ + 0.275, + 0.255, + 0.729, + 0.268 + ], + "angle": 0, + "content": "Peking University {ruyang@stu,hrtang@geli@ece}.pku.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.309, + 0.268, + 0.694, + 0.282 + ], + "angle": 0, + "content": "2ARC Lab, Tencent PCG 3 Peng Cheng Laboratory" + }, + { + "type": "text", + "bbox": [ + 0.339, + 0.283, + 0.665, + 0.297 + ], + "angle": 0, + "content": "{palchenli,yixiaoge,yingsshan}@tencent.com" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.338, + 0.744, + 0.631 + ], + "angle": 0, + "content": "Abstract. Large Language Models (LLMs) have showcased impressive capabilities in text comprehension and generation, prompting research efforts towards video LLMs to facilitate human-AI interaction at the video level. However, how to effectively encode and understand videos in video-based dialogue systems remains to be solved. In this paper, we investigate a straightforward yet unexplored question: Can we feed all spatial-temporal tokens into the LLM, thus delegating the task of video sequence modeling to the LLMs? Surprisingly, this simple approach yields significant improvements in video understanding. Based upon this, we propose ST-LLM, an effective video-LLM baseline with Spatial-Temporal sequence modeling inside LLM. Furthermore, to address the overhead and stability issues introduced by uncompressed video tokens within LLMs, we develop a dynamic masking strategy with tailormade training objectives. For particularly long videos, we have also designed a global-local input module to balance efficiency and effectiveness. Consequently, we harness LLM for proficient spatial-temporal modeling, while upholding efficiency and stability. Extensive experimental results attest to the effectiveness of our method. Through a more concise model and training pipeline, ST-LLM establishes a new state-of-the-art result on VideoChatGPT-Bench and MVBench. Codes have been available at https://github.com/TencentARC/ST-LLM." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.643, + 0.687, + 0.657 + ], + "angle": 0, + "content": "Keywords: ST-LLM \\(\\cdot\\) Dynamic Masking \\(\\cdot\\) Global-Local Input" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.685, + 0.377, + 0.701 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Large Language Models (LLMs), such as GPT [19], PaLM [316] and LLaMA [43,44], have achieved remarkable success owing to their formidable language understanding and generation abilities, signaling a promising advancement towards artificial general intelligence. Driven by the widespread adoption of LLMs, research on Large Visual-Language Models (LVLMs) [14,29,55] has emerged to extend the capabilities of LLMs to process visual signals, which have revealed proficiency in image-based conversations. However, in contrast to image inputs, videos feature heavier input and additional temporal information. Developing a" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.34, + 0.127 + ], + "angle": 0, + "content": "Liu. et al." + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.147, + 0.78, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.395, + 0.789, + 0.427 + ], + "angle": 0, + "content": "Fig. 1: Qualitative comparisons among top-performance video LLMs. We illustrate two cases from MVBench [25] (above) and one case from a YouTube video (below)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.466, + 0.789, + 0.497 + ], + "angle": 0, + "content": "large video-language model capable of effectively extracting meaningful spatial-temporal information from intricate video signals presents a significant challenge." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.503, + 0.789, + 0.669 + ], + "angle": 0, + "content": "Recently, several preliminary attempts have surfaced aimed at extending LLMs to facilitate video conversation [19,20,24,25,27,31,33,53], which unlock the capability of generating generic summaries of video content. Despite these efforts, existing video LLMs continue to exhibit shortcomings in achieving satisfactory performance in video comprehension, particularly in the realm of understanding content that relies heavily on temporal dynamics. As depicted in Fig. 11, contemporary top-performing video LLMs demonstrate robust understanding capabilities for actions reliant on static contexts. However, their proficiency in comprehending scenes involving motion remains constrained. They encounter challenges in discerning even the most basic direction of object movement, let alone the complex motion or scene transition." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.79, + 0.841 + ], + "angle": 0, + "content": "This issue may stem from the inherent difficulty of temporal modeling within conversational systems. Historically, achieving robust video encoding has often required substantially more time and memory resources than images, which is impractical based on LLMs. As illustrated by Fig. 2(a), early video LLMs tend to adopt mean pooling on temporal dimension [24,32,33,53], a method that, while efficient, is inadequate for handling dynamic temporal sequences. Therefore, recent models frequently integrate additional structures for temporal sampling and modeling [19,20,25,27]. While these methods offer improved effectiveness compared to average pooling, they entail increased storage and demand extensive GPU time for training from scratch, often involving two or even three-stage pretraining to align new modules." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.28, + 0.114, + 0.728, + 0.129 + ], + "angle": 0, + "content": "ST-LLM: Large Language Models Are Effective Temporal Learners" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.144, + 0.379, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.383, + 0.144, + 0.544, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.219, + 0.379, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.383, + 0.219, + 0.544, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.377, + 0.273, + 0.391, + 0.282 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.149, + 0.772, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.671, + 0.274, + 0.687, + 0.283 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.295, + 0.789, + 0.352 + ], + "angle": 0, + "content": "Fig. 2: (a) Comparison among different video LLMs based on the way of inputting visual tokens. (b) Quantitative comparison of representative video LLMs, presenting the average results of all 20 metrics and temporal-sensitive motion-related metrics from MVBench, which encompass Moving Direction, Moving Count, and Moving Attribute." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.385, + 0.789, + 0.536 + ], + "angle": 0, + "content": "Inspired by the recent success of joint spatial-temporal-text modeling in visual generation [8,36], in this work, we investigate a simple yet unexplored idea: Given the robust sequence modeling capabilities inherent to LLMs, what if we input all visual tokens into the LLM and delegate the task of modeling spatial-temporal sequences to the LLM itself? Meanwhile, there are two additional challenges: (1) The inclusion of all visual tokens significantly increases the context length within LLMs, particularly for long videos, rendering the processing of a large number of frames unaffordable. (2) LLM may struggle to handle videos of varying lengths, potentially leading to hallucinations when there is a discrepancy between the number of frames in testing and training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.537, + 0.789, + 0.718 + ], + "angle": 0, + "content": "To address these issues, we propose ST-LLM, a simple but powerful baseline of video LLM. As depicted in Fig. 2(a), with raw spatial-temporal tokens inside LLM, we harness the LLMs' robust sequence modeling capabilities for effective temporal modeling. Meanwhile, we introduce a novel dynamic video token masking strategy along with masking video modeling during training. With this approach, we reduce the length of sequences input to the LLM while significantly improving the robustness of videos of varying lengths during inference. To accommodate particularly long videos, we have devised a unique global-local input mechanism. This involves employing mean pooling of a large number of frames to generate residual input for a smaller subset of frames. Through this asymmetric design, we can process input from a large number of video frames while preserving the operations of modeling video tokens within LLMs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.843 + ], + "angle": 0, + "content": "As a result, we maintain input tokens of comparable length to most video LLMs while effectively harnessing the LLM's capacity to comprehend videos. Furthermore, as ST-LLM does not introduce additional modules, it does not necessitate expensive alignment pre-training. This enables ST-LLM to directly leverage existing state-of-the-art image conversational models, resulting in significantly reduced GPU time compared to other state-of-the-art video LLMs. Extensive experiments are conducted to verify the effectiveness of ST-LLM. Qualitatively, as depicted in Fig. 1, ST-LLM demonstrates a superior ability to" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.34, + 0.128 + ], + "angle": 0, + "content": "Liu. et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "understand dynamics compared to other video LLMs. Quantitatively, ST-LLM achieves new state-of-the-art performance across various contemporary video benchmarks, including MVBench [25], VideoChatGPT-Bench [33] and zero-shot QA-Evaluation. Specifically, as illustrated in Fig. 2(b), the superiority of ST-LLM is particularly evident in metrics related to the temporal-sensitive motion, which underscores the expertise of our model in temporal understanding." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.263, + 0.398, + 0.279 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.296, + 0.789, + 0.612 + ], + "angle": 0, + "content": "LLMs and Image LLMs. Recent years have witnessed remarkable advancements in the evolution of Large Language Models (LLMs). Inspired by InstructGPT [35] and the widely recognized commercial model ChatGPT [34], the academic community has seen a proliferation of open-source LLMs, such as LLaMA [43], Alpaca [41], Vicuna [13], and LLaMA 2 [44], which have become foundational components for a myriad of research endeavors. Meanwhile, LLMs, being large-scale transformer models, inherently possess strong sequence modeling capabilities. This inspired us to explore the feasibility of entrusting the task of spatial-temporal sequence modeling to LLMs. The success of LLMs has spurred increasing interest in the development of Multimodal Large Language Models (MLLMs). Notable breakthroughs include works such as Flamingo [2], BLIP2 [23], and PaLM-E [16], which have successfully bridged the gap between vision models and LLMs. Inspired by the concept of instruction tuning for LLMs, a series of works [12, 14, 29, 55] leveraged open-source chatbots for visual instruction tuning, thus facilitating support for image-based conversations. It is noteworthy that in the pre-LLM era, video models were commonly constructed based on pre-trained 2D image models [7, 11, 30, 45]. This may be attributed to the significant computational demands of video processing, combined with the relatively limited scale and diversity of video data compared to images. Therefore, we believe that a well-pretrained image dialogue model could similarly offer training cost and performance benefits to video LLMs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.797, + 0.84 + ], + "angle": 0, + "content": "Video LLMs. The emergence of MLLMs quickly extended into the domain of video as well [19,20,24,25,27,31,33,53]. Early models like VideoChat [24], VideoChatGPT [33] and Valley [32] generate video instruction tuning data through GPT to enable video conversations. At the model level, these approaches typically involve the use of mean pooling to aggregate the encoding results of individual frames before feeding them into the LLMs. While efficient, it is evident that this method is inadequate for effective temporal modeling. Hence, some subsequent models have initiated exploration into adaptations of image models to video-specific requirements. For instance, BT-Adapter [31] proposed a lightweight adapter to extend the capabilities of image LLMs. Chat-UniVi [20] introduces DPC-KNN to cluster dynamic visual tokens. Moreover, VideoChat2 [25] deviated from using CLIP [38] and directly introduced a dedicated video encoder. Although these methods offer improvements over mean pooling, the inclusion of additional modules typically necessitates intensive training. For example, in the case of VideoChat2, aligning the newly introduced video encoder entails" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.28, + 0.115, + 0.728, + 0.129 + ], + "angle": 0, + "content": "ST-LLM: Large Language Models Are Effective Temporal Learners" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.149, + 0.749, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.337, + 0.788, + 0.381 + ], + "angle": 0, + "content": "Fig.3: The overview of ST-LLM for generating responses on the input video and instructions. ST-LLM directly feeds the spatial-temporal sequence into the LLM and employs dynamic masking and global-local input for efficiency and robustness." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.412, + 0.788, + 0.685 + ], + "angle": 0, + "content": "a three-stage training involving 30 million visual-language samples. In contrast, ST-LLM adopts a more direct approach by entrusting the task of visual sequence modeling to LLMs, which is characterized by its conciseness, reduced training requirements, and, according to our experimental findings, superior effectiveness. Joint Spatial-Temporal Modeling. Prior to the emergence of LLMs, joint spatial-temporal modeling had demonstrated effectiveness in both video-only and video-language pretraining [26,42,49]. More recently, inspired by DIT [36], the popular video generation model Sora has shown success in simultaneously processing video tokens and text using the transformer. Despite its potential, joint spatial-temporal-text modeling remains relatively uncommon in video LLMs. While VideoChat2 employed a joint-ST video encoder, it necessitated extensive pretraining. A closer relative, InstructBLIP incorporated joint spatial-temporal sequences into LLM for zero-shot evaluation on video datasets. However, this approach led to significant hallucinations due to the absence of video instruction tuning. To the best of our knowledge, ST-LLM is the first open-source video LLM to adopt joint spatial-temporal-text modeling. Moreover, ST-LLM introduces a series of design innovations to mitigate the limitations associated with incorporating all video tokens within LLM." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.713, + 0.381, + 0.731 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.843 + ], + "angle": 0, + "content": "In this section, we elaborate on the structure and training methodology of ST-LLM. Firstly, we present a description of the model architecture, which incorporates all spatial-temporal tokens within the LLM, in Section 3.1. Subsequently, we delve into the masking mechanism and training objectives in Section 3.2. Finally, we introduce the global-local input approach in Section 3.3. The overarching framework of our method is illustrated in Fig. 3." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.34, + 0.128 + ], + "angle": 0, + "content": "Liu. et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.478, + 0.162 + ], + "angle": 0, + "content": "3.1 Video Tokens Inside LLM" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.17, + 0.788, + 0.304 + ], + "angle": 0, + "content": "As depicted in Fig. 3, visually, ST-LLM closely resembles an image chatbot, consisting of a visual encoder, a linear projection, and an LLM. Instead of opting for the commonly used CLIP-L/14 [38], we have selected BLIP-2 [23] as our visual encoder, whose Q-Former efficiently compresses redundant image tokens into fewer visual tokens. To facilitate the input of raw videos for the image encoder, we adopt the strategy of treating each frame as an individual image. Given a video \\( v \\) with \\( T \\) frames, we assume that the image encoder \\( f_{\\mathrm{v}} \\) encodes each frame into \\( K \\) tokens. Hence, we obtain all visual tokens \\( V \\) to be fed into the LLM as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.308, + 0.314, + 0.786, + 0.334 + ], + "angle": 0, + "content": "\\[\nV = f _ {\\mathrm {p}} (f _ {\\mathrm {v}} (v)), \\mathrm {w h e r e} V = \\{V _ {i} \\} _ {i = 1} ^ {T}, V _ {i} = \\{v _ {(i, j)} \\} _ {j = 1} ^ {K}, \\qquad (1)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.34, + 0.788, + 0.417 + ], + "angle": 0, + "content": "where \\( f_{\\mathrm{p}} \\) is the visual projection layer. Then, we concatenate all visual tokens to form a joint spatial-temporal [7] visual sequence \\( V \\in \\mathbb{R}^{(T*K)\\times D} \\), where \\( D \\) is the number of embedding dimensions. After being passed through the word embedding layers, the text tokens \\( C = \\{c_i\\}_{i=1}^N \\) are directly concatenated with all spatial-temporal tokens, to organize the input \\( I \\) to the LLM as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.273, + 0.426, + 0.786, + 0.443 + ], + "angle": 0, + "content": "\\[\nI = \\left[ V: C \\right] = \\left\\{v _ {(1, 1)}, v _ {(1, 2)}, \\dots , v _ {(1, K)}, v _ {(2, 1)}, \\dots , v _ {(T, K)}, c _ {1}, c _ {2}, \\dots , c _ {N} \\right\\}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.45, + 0.789, + 0.603 + ], + "angle": 0, + "content": "Indeed, several alternative strategies exist for organizing the input tokens before feeding them into the LLM. These include: (1) Adding separator tokens between frame-wise tokens or visual-text tokens, to demarcate the boundaries between modalities or frames. (2) Adding spatial-temporal position embeddings specifically for visual tokens, to facilitate better spatial-temporal relationship understandings. However, our experiments indicate that simplicity often yields the best. Adding numerous separator tokens introduces additional overhead without necessarily improving performance. Moreover, LLMs come with Rotary Position Embeddings [39] that are already effective in distinguishing the positions of all spatial-temporal tokens and text tokens." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.621, + 0.533, + 0.638 + ], + "angle": 0, + "content": "3.2 Training with Dynamic Masking" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.795 + ], + "angle": 0, + "content": "With video tokens modeling inside LLM, we can leverage the potent sequence modeling capabilities inherent in LLMs to understand the information encapsulated within spatial-temporal sequences. However, integrating all video tokens into LLM significantly increases the context length, imposing a considerable computational burden and rendering it unfeasible to input additional frames for lengthy videos. Meanwhile, experimental findings suggest that during testing, utilizing uncompressed tokens is more sensitive to variations in the number of frames than mean pooling. Consequently, when a substantial dissonance exists between the number of frames in training and testing, a noticeable degradation in performance ensues." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.843 + ], + "angle": 0, + "content": "To tackle the aforementioned issues, we first propose masking the visual tokens during training. Masking modeling has achieved significant success in natural language processing [15] and video-language pretraining [26,31,42]. Although" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.279, + 0.115, + 0.728, + 0.131 + ], + "angle": 0, + "content": "ST-LLM: Large Language Models Are Effective Temporal Learners" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "masking is rarely employed in autoregressive LLMs, given our aim to leverage LLMs for encoding spatial-temporal sequences, we can still utilize masking modeling. Specifically, we keep the text tokens in \\(I\\) unchanged, while applying a mask to the video tokens. This mask randomly masks \\(\\rho\\) of all video tokens, irrespective of their position across different frames. Different from previous works [26,31], we adopt a dynamic masking strategy, where the masking rate is randomly sampled from a normal distribution as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.398, + 0.265, + 0.786, + 0.281 + ], + "angle": 0, + "content": "\\[\n\\rho \\sim \\mathbb {N} (0. 5, \\sigma), \\quad 0. 3 \\leq \\rho \\leq 0. 7, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.291, + 0.789, + 0.351 + ], + "angle": 0, + "content": "where \\(\\sigma\\) represents the variance of the normal distribution. This strategy ensures that the length of the spatial-temporal sequence varies continuously while maintaining a consistently high average masking rate of \\(50\\%\\). As a result, it minimizes training costs and notably enhances robustness during inference." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.352, + 0.794, + 0.472 + ], + "angle": 0, + "content": "Furthermore, building upon dynamic masking, we have formulated the Masked Video Modeling (MVM) objective to encourage the LLM to grasp spatial-temporal dependencies. Unlike the expensive masked token recovery [6,37], we have opted for unmasked token reconstruction. Specifically, besides the masked sequence \\( I \\), we conduct an extra non-grad forward pass for the unmasked sequence \\( \\hat{I} = [\\hat{V}:C] \\), serving as the reference output. Then, we select unmasked tokens from \\( f_{llm}(I) \\) and the corresponding tokens from \\( f_{llm}(\\hat{I}) \\) based on their positions, computing the Mean Squared Error (MSE) between the selected pairs as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.312, + 0.485, + 0.786, + 0.525 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {m v m} = \\frac {1}{(1 - \\rho) K T} \\sum_ {i = 1} ^ {T} \\sum_ {j = 1} ^ {K} \\left(v _ {i, j} ^ {- 1} - \\hat {v} _ {i, j} ^ {- 1}\\right) ^ {2}, (i, j) \\notin M, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.537, + 0.788, + 0.629 + ], + "angle": 0, + "content": "where \\(M\\) denotes the set of masked tokens' indices, \\(v_{i,j}^{-1}\\) is from \\(f_{llm}(I)\\), and \\(\\mathcal{L}_{mvm}\\) represents our MVM objective. Finally, our overall loss \\(\\mathcal{L}\\) is composed of \\(\\mathcal{L}_{mvm}\\) and the LLM decoder loss \\(\\mathcal{L}_{llm}\\), yielding \\(\\mathcal{L} = \\mathcal{L}_{mvm} + \\mathcal{L}_{llm}\\). By integrating these two loss components, we can encourage the LLM to effectively respond to questions derived from the video content while improving its ability to model temporal and spatial dependencies." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.65, + 0.421, + 0.667 + ], + "angle": 0, + "content": "3.3 Global-Local Input" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.794 + ], + "angle": 0, + "content": "Although dynamic masking partially mitigates the challenge of processing lengthy input sequences, it fails to fully resolve the issue for extremely long videos that entail inputting numerous frames, rendering the context length still impractical to handle. Therefore, we have devised an additional module to tackle the issue of excessively long videos. To elaborate, given a lengthy video with a large \\( T \\), we still commence by encoding each frame individually to derive \\( V \\). Subsequently, we proceed to derive the global video representation \\( V_{0} \\) through average pooling on the frame-wise tokens:" + }, + { + "type": "equation", + "bbox": [ + 0.366, + 0.804, + 0.786, + 0.844 + ], + "angle": 0, + "content": "\\[\nv _ {(0, j)} = \\frac {1}{T} \\sum_ {i = 1} ^ {T} v _ {(i, j)}, \\quad V _ {0} = \\left\\{v _ {(0, j)} \\right\\} _ {j = 1} ^ {K}. \\tag {5}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.34, + 0.128 + ], + "angle": 0, + "content": "Liu. et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.21 + ], + "angle": 0, + "content": "Next, we average-sample \\(\\overline{T}\\) frames from the T total frames. All tokens from these \\(\\overline{T}\\) frames are concatenated to produce a joint spatial-temporal sequence \\(\\overline{V} \\in \\mathbb{R}^{(\\overline{T}*K) \\times D}\\), which serves as the local video representation. Finally, the global-local input for the LLM, denoted as \\(\\overline{I}\\), is constructed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.425, + 0.222, + 0.786, + 0.239 + ], + "angle": 0, + "content": "\\[\n\\bar {I} = \\left[ \\bar {V} + f _ {m} \\left(V _ {0}\\right): C \\right], \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.25, + 0.788, + 0.342 + ], + "angle": 0, + "content": "where \\( f_{m} \\) is a simple MLP projector with upsampling projection initialized with zeros. With this global-local input design, the low fps spatial-temporal sequences within the LLM can gradually incorporate information from the high fps branch. This approach allows the model to benefit from the LLM's ability to model temporal sequences within a limited context while also considering the global information of long videos." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.365, + 0.377, + 0.383 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.398, + 0.79, + 0.46 + ], + "angle": 0, + "content": "In this section, we have performed comprehensive experimental evaluations of ST-LLM, covering crucial settings, comparisons, and ablation studies. For a more detailed account of experimental settings, ablation studies, visualizations, and limitations analysis, please consult the appendix." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.482, + 0.414, + 0.498 + ], + "angle": 0, + "content": "4.1 Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.789, + 0.644 + ], + "angle": 0, + "content": "Benchmarks. To evaluate the video understanding capabilities of ST-LLM, we primarily conduct assessments on three benchmarks: MVBench [25], VideoChat-GPT Bench [33], and zero-shot video QA benchmark. MVBench comprises 20 challenging video tasks, each consisting of 200 samples in the form of multiple-choice questions. These tasks provide a comprehensive and objective assessment of a model's ability to understand videos. VideoChatGPT-Bench gathers videos from ActivityNet [10] and uses GPT to evaluate the quality of video conversations across five dimensions. Additionally, the zero-shot video QA benchmark entails GPT-based assessments of various open-source video QA datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.79, + 0.842 + ], + "angle": 0, + "content": "Implementation Details. In our study, we adopt state-of-the-art image dialogue models as the baseline and proceed directly with video instruction tuning. Unless otherwise specified, our model is initialized with InstructBLIP [14] in all experiments, whose LLM is Vicuna-v1.1 [13] with 7B parameters. Additionally, we employ Minigpt4-v1 [55] as a basic model to assess the effectiveness across various image models in Table 6. Following previous research [20, 27, 28], we adopt full-finetuning on the LLM to ensure a fair comparison. The results of LoRA are also included as ablation results in Table 6. \\(\\sigma\\) is set as 0.1, ensuring that over \\(95\\%\\) of the mask rate falls within the range of 0.3 to 0.7. During training, we sample 16 frames per video. Considering dynamic masking, an average of 256 tokens per video are input into the LLM, which is the same as LLaVA [29]. For all benchmarks during inference, frames are sampled at a frame rate of 1, and the number of local frames is kept between 4 and 16. This strategy closely" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.28, + 0.115, + 0.728, + 0.129 + ], + "angle": 0, + "content": "ST-LLM: Large Language Models Are Effective Temporal Learners" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.26 + ], + "angle": 0, + "content": "Table 1: Comparisons on MVBench. Except for BLIP2 and Otter, all models are built upon LLaMA-1-7B for fair comparisons. \"Avg\" denotes the average of all 20 metrics. The leaderboards for each task can be found in the Appendix. The compared methods include: (1)Rd: random guesses; (2) mOI: mPLUG-Owl-I 51; (3) LMA: LLaMA-Adapter 54 (4)B2: BLIP-2 23; (5)OI: Otter-I 22; (6)MG: MiniGPT-4 55; (7)IB: InstructBLIP 14; (8)LV: LLaVA 29; (9)OV: Otter-V 22; (10)mOV: mPLUG-Owl-V 51; (11)VCG: VideoChatGPT 33; (12)VLM: VideoLLaMA 53; (13)VC: VideoChat 24; (14)VC2: VideoChat2 25." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.268, + 0.782, + 0.585 + ], + "angle": 0, + "content": "
MetricRdmOILMAB2OIMGIBLVOVmOVVCGVLMVCVC2Ours
AS25.025.023.024.534.516.020.028.023.022.023.527.533.566.066.0
AP25.020.028.029.032.018.016.539.523.028.026.025.526.547.553.5
AA33.344.551.033.539.526.046.063.027.534.062.051.050.083.584.0
FA25.027.030.017.030.521.524.530.527.029.022.529.033.549.544.0
UA25.023.533.042.038.516.046.039.029.529.026.539.040.560.058.5
OE33.336.053.551.548.529.551.053.053.040.554.048.053.058.080.5
OI25.024.032.526.044.025.526.041.028.027.028.040.540.571.573.5
OS33.334.033.531.029.513.037.541.533.031.540.038.030.042.538.5
MD25.023.025.525.519.011.522.023.024.527.023.022.525.523.042.5
AL25.024.021.526.025.512.023.020.523.523.020.022.527.023.031.0
ST25.034.530.532.555.09.546.545.027.529.031.043.048.588.586.5
AC33.334.529.025.520.032.542.534.026.031.530.534.035.039.036.5
MC25.022.022.530.032.515.526.520.528.527.025.522.520.542.056.5
MA33.331.541.540.028.58.040.538.518.040.039.532.542.558.578.5
SC33.340.039.542.039.034.032.047.038.544.048.545.546.044.043.0
FP25.024.025.027.028.026.025.525.022.024.029.032.526.549.044.5
CO33.337.031.530.027.029.530.536.022.031.033.040.041.036.546.5
EN25.025.522.526.032.019.030.527.023.526.029.530.023.535.034.5
ER20.021.028.037.036.59.930.526.519.520.526.021.023.540.541.5
CI30.937.032.031.036.53.038.042.019.029.535.537.036.065.558.5
Avg27.329.431.731.433.518.832.536.026.829.732.734.135.551.154.9
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.704 + ], + "angle": 0, + "content": "aligns with practical scenarios and involves fewer frames than most video LLMs. For VideoChatGPT-Bench and the video QA benchmark, which may involve videos spanning several minutes, we utilize the global-local input approach. In this configuration, the global frame number is set to 64. All videos are resized into \\(224^{*}224\\). We train ST-LLM using a learning rate of 2e-5 and a batch size of 128 for 2 epochs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Training Datasets. Citing findings from Jin et al. [20], the sequential process of first conducting image instruction tuning, followed by video instruction tuning and image-video joint pretraining, had a negligible impact on video dialogue performance. Hence, in light of InstructBLIP's prior training on extensive image data, we exclusively focus on video instruction tuning. Following [25], we leverage video instruction data sourced from a variety of datasets, including VideoChatGPT-100k [33], VideoChat-11k [24], Webvid [5], NExT-QA [47], CLEVRER [52], Kinetics-710 [21], and Something-Something-2 [17]. All samples are uniformly formatted in accordance with [25]. More details about the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.34, + 0.127 + ], + "angle": 0, + "content": "Liu. et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.241, + 0.145, + 0.761, + 0.16 + ], + "angle": 0, + "content": "Table 2: Comparisons on video-based generative performance benchmarking." + }, + { + "type": "table", + "bbox": [ + 0.234, + 0.172, + 0.766, + 0.35 + ], + "angle": 0, + "content": "
MethodCorrectDetailContextTemporalConsistMean Score
VideoLLaMA 531.962.182.161.821.791.98
LLaMA-Adapter 542.032.322.301.982.152.16
VideoChat 242.232.502.531.942.242.29
VideoChatGPT 332.402.522.621.982.372.38
BT-Adapter 312.682.693.272.342.462.69
VTimeLLM 192.783.103.402.492.472.85
Chat-UniVi 202.892.913.462.892.812.99
LLaMA-VID-7B 272.963.003.532.462.512.89
LLaMA-VID-13B 273.073.053.602.582.632.99
VideoChat2 253.022.883.512.662.812.98
ST-LLM (Ours)3.233.053.742.932.813.15
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.375, + 0.788, + 0.406 + ], + "angle": 0, + "content": "dataset and the organization of instructions can be found in the appendix. An experiment on joint image-video training is also presented in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.425, + 0.427, + 0.44 + ], + "angle": 0, + "content": "4.2 Quantitative Result" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.789, + 0.844 + ], + "angle": 0, + "content": "MVBench Performance. The evaluation results on MVBench are presented in Table 1. The evaluation metrics included Action Sequence (AS), Action Prediction (AP), Action Antonym (AA), Fine-grained Action (FA), Unexpected Action (UA), Object Existence (OE), Object Interaction (OI), Object Shuffle (OS), Moving Direction (MD), Action Localization (AL), Scene Transition (ST), Action Count (AC), Moving Count (MC), Moving Attribute (MA), State Change (SC), Fine-grained Pose (FP), Character Order (CO), Egocentric Navigation (EN), Episodic Reasoning (ER), Counterfactual Inference (CI), and the average of all 20 metrics (Avg). It is evident that on such a challenging benchmark, the performance of most multimodal LLMs falls significantly short of satisfactory levels, even barely surpassing random performance by less than \\(10\\%\\). Our only formidable competitor is VideoChat2 [25]. Through direct substitution of CLIP with a dedicated video encoder, coupled with a three-stage training process and comprehensive instruction tuning for the visual encoder, Q-Former, and LLM, VideoChat2 has yielded results that significantly surpass those of previous models. In contrast, we solely carried out single-stage video-only instruction tuning while maintaining the weights of CLIP and Q-Former frozen. Even under these conditions, ST-LLM still attained the top performance, with an average score surpassing VideoChat2 by a notable margin of \\(3.8\\%\\). Specifically, ST-LLM showcased leading results across all 11 tasks, exhibiting a particularly significant advantage in motion-related tasks where other models struggled. In three motion-related tasks, ST-LLM achieved an average score of \\(59.2\\%\\), far outperforming VideoChat2's \\(36.3\\%\\). However, ST-LLM encounters challenges in fine-grained tasks, such as Fine-grained Action (FA) and Fine-grained Pose (FP). This could be attributed to CLIP's limitations as an image encoder, which may struggle to capture low-level spatial-temporal features required for these tasks." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.279, + 0.114, + 0.728, + 0.129 + ], + "angle": 0, + "content": "ST-LLM: Large Language Models Are Effective Temporal Learners" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.144, + 0.786, + 0.176 + ], + "angle": 0, + "content": "Table 3: Comparisons on zero-shot question-answering, including MSVD-QA [46], MSRVTT-QA [48], and ActivityNet-QA [10]." + }, + { + "type": "table", + "bbox": [ + 0.253, + 0.186, + 0.751, + 0.356 + ], + "angle": 0, + "content": "
MethodLLMMSVD-QAMSRVTT-QAActivityNet-QA
AccScoreAccScoreAccScore
FrozenBiLM 50DeBERTa-V232.2-16.8-24.7-
VideoLLaMA 53Vicuna-7B51.62.529.61.812.41.1
LLaMA-Adapter 54LLaMA-7B54.93.143.82.734.22.7
VideoChat 24Vicuna-7B56.32.845.02.526.52.2
VideoChatGPT 33Vicuna-7B64.93.349.32.835.22.7
BT-Adapter 31Vicuna-7B67.53.757.03.245.73.2
Chat-UniVi 20Vicuna-7B65.03.654.63.145.83.2
LLaMA-VID 27Vicuna-7B69.73.757.73.247.43.3
LLaMA-VID 27Vicuna-13B70.03.758.93.347.53.3
VideoChat2 25Vicuna-7B70.03.954.13.349.13.3
ST-LLM (Ours)Vicuna-7B74.63.963.23.450.93.3
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.384, + 0.788, + 0.565 + ], + "angle": 0, + "content": "VideoChatGPT-Bench Performance. VideoChatGPT-Bench, also referred to as the Video-Based Generative Performance Benchmark, primarily evaluates the capability of video conversation. Utilizing GPT-3.5 [34], it assigns a score to the generated content across five aspects: Correctness of Information (Correct), Detail Orientation (Detail), Contextual Understanding (Context), Temporal Understanding (Temporal), and Consistency (Consist). As depicted in Table 2, the recently proposed state-of-the-art models [20,25,27] exhibit similar performance levels on this benchmark. However, due to the inherent instability of GPT evaluation, there is considerable variation in the performance of each model across different tasks. Despite this variability, ST-LLM consistently performs admirably across all tasks, while demonstrating significant advantages in terms of the mean score compared to previous models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.566, + 0.801, + 0.703 + ], + "angle": 0, + "content": "Zero-shot Video-question Answering Performance. In Table 3 we present the zero-shot video question-answering performance on several commonly used video-text datasets, including MSVD-QA [46], MSRVTT-QA [48], and ActivityNet-QA [10]. Following the methodology outlined in [33], we utilize GPT-3.5 to evaluate the accuracy and score of the generated results. As is depicted, STLLM demonstrates significant advantages on MSVD and MSRVTT, surpassing the previous SOTAs by \\(4.6\\%\\) and \\(4.3\\%\\) respectively. However, STLLM exhibits a slightly smaller advantage compared to VideoChat2 on the ActivityNet. This could be attributed to the requirement for a potent video encoder for this dataset." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.725, + 0.455, + 0.74 + ], + "angle": 0, + "content": "4.3 Ablations and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.749, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Ways of Inputting video tokens. The core of this work lies in modeling joint spatial-temporal video tokens using LLM. Table 4 presents the results of different methods of inputting video tokens. As depicted, without video tuning, the approach of joint spatial-temporal input is considerably less effective compared to directly averaging the input over the temporal dimension. This discrepancy might stem from inconsistencies between training and testing. However, after" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.34, + 0.128 + ], + "angle": 0, + "content": "Liu. et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.163, + 0.505, + 0.219 + ], + "angle": 0, + "content": "Table 4: The ablation study on the methods of inputting video tokens to LLMs. The baseline configurations include no instruction tuning and mean pooling input." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.22, + 0.504, + 0.337 + ], + "angle": 0, + "content": "
MethodAvg on MV bench
Instructclip Baseline
Mean Pooling42.0
S-T Tokens in LLM35.5
Video Instruction Tuning
Mean Pooling47.8
S-T Tokens in LLM53.8
+Masking & MVM Loss54.9
" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.144, + 0.782, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.515, + 0.308, + 0.788, + 0.337 + ], + "angle": 0, + "content": "Fig. 4: The ablation study on the effect of dynamic masking and MVM loss." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.351, + 0.788, + 0.382 + ], + "angle": 0, + "content": "Table 5: The ablation study on global-local input module. The \"Local\" and \"Global\" mean the joint temporal-temporal sequence and average pooling sequence respectively." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.393, + 0.784, + 0.478 + ], + "angle": 0, + "content": "
MethodAvg on MVBenchVideo-ChatGPT Bench
CorrectDetailContextTemporalConsist
Global Only48.33.012.923.582.612.61
Local Only54.93.082.993.632.752.69
Local+Global (simply add)50.13.072.883.582.592.50
Local+Global (adapter)54.73.233.053.742.932.81
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.598 + ], + "angle": 0, + "content": "training, the effectiveness of the joint spatial-temporal input significantly surpasses that of the mean pooling input, indicating that LLMs can effectively model spatial-temporal sequences, and directly inputting all tokens is a superior approach compared to compression before input. Furthermore, incorporating dynamic masking and the MVM loss based on the joint spatial-temporal sequence has further improved the performance of our model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.788, + 0.779 + ], + "angle": 0, + "content": "Dynamic Masking Strategy and MVM Loss. In Fig. 4, we conduct a more detailed ablation study on the effect of dynamic masking and MVM loss. It is evident that when the training frame count remains fixed, the performance of inputting all video tokens is not very robust. Particularly, when there is a discrepancy between the training and testing frame, the performance noticeably decreases. Meanwhile, when we randomly mask video tokens at a static \\(50\\%\\) rate, this issue persists, with effectiveness akin to training with a fixed 8-frame setup. However, employing dynamic masking shows some improvement in overall performance, and notably, the robustness to varying testing frame counts is significantly enhanced. Furthermore, the inclusion of the MVM loss built upon the masking further enhances performance, with greater improvements observed as the sequence length increases." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Global-Local Input Module. The ablation results on the global-local module are presented in Table 5 where we compare the performance of global features extracted from more frames and local features from fewer frames, as well as their combination. As shown, on MVBench, the performance of global features" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.28, + 0.114, + 0.728, + 0.129 + ], + "angle": 0, + "content": "ST-LLM: Large Language Models Are Effective Temporal Learners" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.785, + 0.173 + ], + "angle": 0, + "content": "Table 6: The ablation study on the basic image models and the fine-tuning methodologies of LLM. We report the Avg on MVBench." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.186, + 0.784, + 0.254 + ], + "angle": 0, + "content": "
Baseline ModelLLMLLM ParametersMeanpooling BaselineST-LLM Full Model
MiniGPT-4 55Vicuna-7B v0LoRA40.945.6
Full Fine-tuning46.450.5
InstructBLIP 14Vicuna-7B v1.1LoRA45.251.9
Full Fine-tuning47.854.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.283, + 0.52, + 0.31 + ], + "angle": 0, + "content": "Table 7: The analysis on the designs of inputting S-T video tokens into LLM." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.312, + 0.522, + 0.386 + ], + "angle": 0, + "content": "
MethodAvg on MVBench
S-T Tokens in LLM (baseline)54.9
+frame-frame separators54.0
+video-text separators54.8
+S-T position embeddings52.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.527, + 0.284, + 0.786, + 0.311 + ], + "angle": 0, + "content": "Table 8: The analysis on the designs of dynamic masking." + }, + { + "type": "table", + "bbox": [ + 0.532, + 0.313, + 0.787, + 0.386 + ], + "angle": 0, + "content": "
MethodAvg on MVBatch
w/o masking (baseline)53.8
ρ ~ U(0.3, 0.7)54.2
ρ ~ N(0.5, 0.2)54.3
ρ ~ N(0.5, 0.1)54.9
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.415, + 0.788, + 0.567 + ], + "angle": 0, + "content": "is significantly inferior to the joint S-T features, so combining global and local features did not lead to improvement. This may be due to the fact that MVBench mainly consists of short videos lasting a few to a dozen seconds, where a large number of frames might not be necessary. However, on VideoChatGPT Bench, where videos are parsed at the minute level, the discrepancy between global and local features has notably diminished, with their amalgamation resulting in a discernible enhancement. This underscores the efficacy of the Global-Local Input module in processing extended video content. Moreover, the direct incorporation of global and local features resulted in a performance decline, presumably due to its interference with the representation learning of the LLM." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.788, + 0.674 + ], + "angle": 0, + "content": "Basic Image Models & LLMs. In Table 6, we conduct ablations on both the basic image chatbots and the LLM fine-tuning strategy. As depicted in the table, regardless of whether it's LoRA [18] tuning or full fine-tuning, or different basic image models, ST-LLM demonstrates notable enhancements compared to the mean pooling baseline. This suggests that our approach can be adeptly applied across diverse scenarios, such as employing more resource-efficient LoRA tuning on larger models, or transferring to more potent image-dialogue models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Designs of ST-LLM. In Table 7 and Table 8, we investigated several alternative design possibilities for ST-LLM. Primarily, we explored various separator designs. It's noteworthy that almost all cross-modal LLMs incorporate separator designs, with some models even integrating separators between multiple images or frames [2,4]. However, we discovered that when employing LLMs to model spatial-temporal sequences, these separators are superfluous. Subsequently, we address positional encodings. Numerous approaches integrate temporal positional encodings before inputting video tokens into LLMs [32,33,53]. However, in our methodology, we observed that incorporating absolute positional encodings to all visual tokens significantly impairs performance. Lastly, we consider the design of dynamic masking. Our findings suggest that all distributions" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.34, + 0.127 + ], + "angle": 0, + "content": "Liu. et al." + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.147, + 0.78, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.334, + 0.449, + 0.669, + 0.464 + ], + "angle": 0, + "content": "Fig. 5: Qualitative results from real-world videos." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.491, + 0.789, + 0.522 + ], + "angle": 0, + "content": "of masking rates can yield some improvement, yet distributions with smaller standard deviations exhibit superior performance." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.541, + 0.423, + 0.556 + ], + "angle": 0, + "content": "4.4 Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.564, + 0.789, + 0.641 + ], + "angle": 0, + "content": "In Fig. 5, we present a qualitative comparison. As illustrated, ST-LLM excels in adhering to instructions and delivering precise responses. More importantly, ST-LLM showcases superior sensitivity to temporal sequences and actions, thereby validating the effectiveness of our modeling approach. For further quantitative results and failure cases, please refer to the appendix." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.66, + 0.36, + 0.677 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.841 + ], + "angle": 0, + "content": "In this paper, we present ST-LLM, a straightforward yet robust video large language model. Our aim is to achieve effective video comprehension by leveraging LLM to model video tokens, marking the inception of a joint spatial-temporal-text modeling paradigm. Additionally, we introduce a dynamic masking strategy and a global-local input module to enhance this framework. Remarkably, ST-LLM excels in encoding and comprehending spatial-temporal sequences while addressing concerns related to efficiency, stability, and modeling lengthy videos, all with reduced training resource requirements. Through extensive experimental analysis, we demonstrate the effectiveness of ST-LLM and its internal design, resulting in state-of-the-art performance across multiple video LLM benchmarks." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.28, + 0.115, + 0.728, + 0.129 + ], + "angle": 0, + "content": "ST-LLM: Large Language Models Are Effective Temporal Learners" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.207 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported by National Natural Science Foundation of China (No. 62172021). We are also grateful to the support from Guangdong Provincial Key Laboratory of Ultra High Definition Immersive Media Technology." + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.208, + 0.775, + 0.221 + ], + "angle": 0, + "content": "*: Equal Contribution, work done during internship at ARC Lab, Tencent PCG." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.241, + 0.323, + 0.257 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.268, + 0.786, + 0.31 + ], + "angle": 0, + "content": "1. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.311, + 0.787, + 0.365 + ], + "angle": 0, + "content": "2. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems 35, 23716-23736 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.365, + 0.787, + 0.405 + ], + "angle": 0, + "content": "3. Anil, R., Dai, A.M., First, O., Johnson, M., Lepikhin, D., Passos, A., Shakeri, S., Taropa, E., Bailey, P., Chen, Z., et al.: Palm 2 technical report. arXiv preprint arXiv:2305.10403 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.406, + 0.787, + 0.46 + ], + "angle": 0, + "content": "4. Awadalla, A., Gao, I., Gardner, J., Hessel, J., Hanafy, Y., Zhu, W., Marathe, K., Bitton, Y., Gadre, S., Sagawa, S., et al.: Openflamingo: An open-source framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.46, + 0.787, + 0.5 + ], + "angle": 0, + "content": "5. Bain, M., Nagrani, A., Varol, G., Zisserman, A.: Frozen in time: A joint video and image encoder for end-to-end retrieval. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1728-1738 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.501, + 0.787, + 0.528 + ], + "angle": 0, + "content": "6. Bao, H., Dong, L., Piao, S., Wei, F.: Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.528, + 0.787, + 0.554 + ], + "angle": 0, + "content": "7. Bertasius, G., Wang, H., Torresani, L.: Is space-time attention all you need for video understanding? In: ICML. vol. 2, p. 4 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.555, + 0.787, + 0.609 + ], + "angle": 0, + "content": "8. Brooks, T., Peebles, B., Holmes, C., DePue, W., Guo, Y., Jing, L., Schnurr, D., Taylor, J., Luhman, T., Luhman, E., Ng, C., Wang, R., Ramesh, A.: Video generation models as world simulators (2024), https://openai.com/research/video-generation-models-as-world-simulators" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.609, + 0.787, + 0.65 + ], + "angle": 0, + "content": "9. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.65, + 0.787, + 0.69 + ], + "angle": 0, + "content": "0. Caba Heilbron, F., Escorcia, V., Ghanem, B., Carlos Niebles, J.: Activitynet: A large-scale video benchmark for human activity understanding. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 961-970 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.691, + 0.787, + 0.731 + ], + "angle": 0, + "content": "1. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6299-6308 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.731, + 0.787, + 0.785 + ], + "angle": 0, + "content": "2. Chen, J., Zhu, D., Shen, X., Li, X., Liu, Z., Zhang, P., Krishnamoorthi, R., Chandra, V., Xiong, Y., Elhoseiny, M.: Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.787, + 0.84 + ], + "angle": 0, + "content": "3. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., et al.: Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\) chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023) (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.268, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.338, + 0.127 + ], + "angle": 0, + "content": "Liu. et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "14. Dai, W., Li, J., Li, D., Tiong, A., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: towards general-purpose vision-language models with instruction tuning. arxiv. Preprint posted online on June 15, 2023 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.787, + 0.23 + ], + "angle": 0, + "content": "15. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.23, + 0.787, + 0.271 + ], + "angle": 0, + "content": "16. Driess, D., Xia, F., Sajjadi, M.S., Lynch, C., Chowdhery, A., Ichter, B., Wahid, A., Tompson, J., Vuong, Q., Yu, T., et al.: Palm-e: An embodied multimodal language model. arXiv preprint arXiv:2303.03378 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.271, + 0.787, + 0.339 + ], + "angle": 0, + "content": "17. Goyal, R., Ebrahimi Kahou, S., Michalski, V., Materzynska, J., Westphal, S., Kim, H., Haenel, V., Fruend, I., Yianilos, P., Mueller-Freitag, M., et al.: The\" something something\" video database for learning and evaluating visual common sense. In: Proceedings of the IEEE international conference on computer vision. pp. 5842-5850 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.339, + 0.787, + 0.367 + ], + "angle": 0, + "content": "18. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.367, + 0.787, + 0.394 + ], + "angle": 0, + "content": "19. Huang, B., Wang, X., Chen, H., Song, Z., Zhu, W.: Vtimellm: Empower llm to grasp video moments. arXiv preprint arXiv:2311.18445 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.394, + 0.787, + 0.434 + ], + "angle": 0, + "content": "20. Jin, P., Takanobu, R., Zhang, C., Cao, X., Yuan, L.: Chat-univi: Unified visual representation empowers large language models with image and video understanding. arXiv preprint arXiv:2311.08046 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.434, + 0.787, + 0.474 + ], + "angle": 0, + "content": "21. Kay, W., Carreira, J., Simonyan, K., Zhang, B., Hillier, C., Vijayanarasimhan, S., Viola, F., Green, T., Back, T., Natsev, P., et al.: The kinetics human action video dataset. arXiv preprint arXiv:1705.06950 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.474, + 0.787, + 0.502 + ], + "angle": 0, + "content": "22. Li, B., Zhang, Y., Chen, L., Wang, J., Yang, J., Liu, Z.: Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.502, + 0.787, + 0.542 + ], + "angle": 0, + "content": "23. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.542, + 0.787, + 0.583 + ], + "angle": 0, + "content": "24. Li, K., He, Y., Wang, Y., Li, Y., Wang, W., Luo, P., Wang, Y., Wang, L., Qiao, Y.: Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.583, + 0.787, + 0.624 + ], + "angle": 0, + "content": "25. Li, K., Wang, Y., He, Y., Li, Y., Wang, Y., Liu, Y., Wang, Z., Xu, J., Chen, G., Luo, P., et al.: Mvbench: A comprehensive multi-modal video understanding benchmark. arXiv preprint arXiv:2311.17005 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.624, + 0.787, + 0.664 + ], + "angle": 0, + "content": "26. Li, K., Wang, Y., Li, Y., Wang, Y., He, Y., Wang, L., Qiao, Y.: Unmasked teacher: Towards training-efficient video foundation models. arXiv preprint arXiv:2303.16058 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.664, + 0.787, + 0.692 + ], + "angle": 0, + "content": "27. Li, Y., Wang, C., Jia, J.: Llama-vid: An image is worth 2 tokens in large language models. arXiv preprint arXiv:2311.17043 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.692, + 0.787, + 0.732 + ], + "angle": 0, + "content": "28. Lin, B., Zhu, B., Ye, Y., Ning, M., Jin, P., Yuan, L.: Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.732, + 0.787, + 0.759 + ], + "angle": 0, + "content": "29. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv preprint arXiv:2304.08485 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.787, + 0.813 + ], + "angle": 0, + "content": "30. Liu, R., Huang, J., Li, G., Feng, J., Wu, X., Li, T.H.: Revisiting temporal modeling for clip-based image-to-video knowledge transferring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6555-6564 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.787, + 0.841 + ], + "angle": 0, + "content": "31. Liu, R., Li, C., Ge, Y., Shan, Y., Li, T.H., Li, G.: One for all: Video conversation is feasible without video instruction tuning. arXiv preprint arXiv:2309.15785 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.28, + 0.115, + 0.728, + 0.129 + ], + "angle": 0, + "content": "ST-LLM: Large Language Models Are Effective Temporal Learners" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.786, + 0.189 + ], + "angle": 0, + "content": "32. Luo, R., Zhao, Z., Yang, M., Dong, J., Qiu, M., Lu, P., Wang, T., Wei, Z.: Valley: Video assistant with large language model enhanced ability. arXiv preprint arXiv:2306.07207 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "33. Maaz, M., Rasheed, H., Khan, S., Khan, F.S.: Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.233, + 0.658, + 0.248 + ], + "angle": 0, + "content": "34. OpenAI: Chatgpt. https://openai.com/blog/chatgpt (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.248, + 0.786, + 0.303 + ], + "angle": 0, + "content": "35. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems 35, 27730-27744 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.304, + 0.786, + 0.346 + ], + "angle": 0, + "content": "36. Peebles, W., Xie, S.: Scalable diffusion models with transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4195-4205 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.346, + 0.786, + 0.373 + ], + "angle": 0, + "content": "37. Peng, Z., Dong, L., Bao, H., Ye, Q., Wei, F.: Beit v2: Masked image modeling with vector-quantized visual tokenizers. arXiv preprint arXiv:2208.06366 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.374, + 0.786, + 0.43 + ], + "angle": 0, + "content": "38. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.431, + 0.786, + 0.459 + ], + "angle": 0, + "content": "39. Su, J., Ahmed, M., Lu, Y., Pan, S., Bo, W., Liu, Y.: Roformer: Enhanced transformer with rotary position embedding. Neurocomputing 568, 127063 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.459, + 0.786, + 0.487 + ], + "angle": 0, + "content": "40. Sun, Q., Fang, Y., Wu, L., Wang, X., Cao, Y.: Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.786, + 0.515 + ], + "angle": 0, + "content": "41. Taori, R., Gulrajani, I., Zhang, T., Dubois, Y., Li, X., Guestrin, C., Liang, P., Hashimoto, T.B.: Stanford alpaca: An instruction-following llama model (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.516, + 0.786, + 0.558 + ], + "angle": 0, + "content": "42. Tong, Z., Song, Y., Wang, J., Wang, L.: Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. Advances in neural information processing systems 35, 10078-10093 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.559, + 0.786, + 0.601 + ], + "angle": 0, + "content": "43. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.601, + 0.786, + 0.643 + ], + "angle": 0, + "content": "44. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.643, + 0.786, + 0.685 + ], + "angle": 0, + "content": "45. Tran, D., Bourdev, L., Fergus, R., Torresani, L., Paluri, M.: Learning spatiotemporal features with 3d convolutional networks. In: Proceedings of the IEEE international conference on computer vision. pp. 4489-4497 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.686, + 0.786, + 0.713 + ], + "angle": 0, + "content": "46. Wu, Z., Yao, T., Fu, Y., Jiang, Y.G.: Deep learning for video classification and captioning. In: Frontiers of multimedia research, pp. 3-29. ACM (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.714, + 0.786, + 0.756 + ], + "angle": 0, + "content": "47. Xiao, J., Shang, X., Yao, A., Chua, T.S.: Next-qa: Next phase of question-answering to explaining temporal actions. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9777-9786 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.757, + 0.786, + 0.798 + ], + "angle": 0, + "content": "48. Xu, J., Mei, T., Yao, T., Rui, Y.: Msr-vtt: A large video description dataset for bridging video and language. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5288-5296 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.786, + 0.841 + ], + "angle": 0, + "content": "49. Xue, H., Sun, Y., Liu, B., Fu, J., Song, R., Li, H., Luo, J.: Clip-vip: Adapting pre-trained image-text model to video-language representation alignment. arXiv preprint arXiv:2209.06430 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.338, + 0.127 + ], + "angle": 0, + "content": "Liu. et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "50. Yang, A., Miech, A., Sivic, J., Laptev, I., Schmid, C.: Zero-shot video question answering via frozen bidirectional language models. Advances in Neural Information Processing Systems 35, 124-141 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.788, + 0.231 + ], + "angle": 0, + "content": "51. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.231, + 0.787, + 0.272 + ], + "angle": 0, + "content": "52. Yi, K., Gan, C., Li, Y., Kohli, P., Wu, J., Torralba, A., Tenenbaum, J.B.: Clevrer: Collision events for video representation and reasoning. arXiv preprint arXiv:1910.01442 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.787, + 0.3 + ], + "angle": 0, + "content": "53. Zhang, H., Li, X., Bing, L.: Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.787, + 0.342 + ], + "angle": 0, + "content": "54. Zhang, R., Han, J., Zhou, A., Hu, X., Yan, S., Lu, P., Li, H., Gao, P., Qiao, Y.: Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.787, + 0.384 + ], + "angle": 0, + "content": "55. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.384 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/ST-LLM_ Large Language Models Are Effective Temporal Learners/f824889a-c655-4483-9dd5-c5cb83e43784_origin.pdf b/2024/ST-LLM_ Large Language Models Are Effective Temporal Learners/f824889a-c655-4483-9dd5-c5cb83e43784_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4b94eb64bd1b13cab45c08e1d203bad0f90168ed --- /dev/null +++ b/2024/ST-LLM_ Large Language Models Are Effective Temporal Learners/f824889a-c655-4483-9dd5-c5cb83e43784_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dd7c52436cb01970f879cb4a6cbb24bc4ed491c56a41e5e62725d04bf2a2ec88 +size 4716181 diff --git a/2024/ST-LLM_ Large Language Models Are Effective Temporal Learners/full.md b/2024/ST-LLM_ Large Language Models Are Effective Temporal Learners/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c7b9bc8caaca29bc2ceab54a3ccfba4c6e846bf9 --- /dev/null +++ b/2024/ST-LLM_ Large Language Models Are Effective Temporal Learners/full.md @@ -0,0 +1,271 @@ +# ST-LLM: Large Language Models Are Effective Temporal Learners + +Ruyang Liu\*1,3 Chen Li\*2 Haoran Tang1,3 Yixiao Ge2 Ying Shan2 Ge Li + +1 School of Electronic and Computer Engineering, Shenzhen Graduate School, + +Peking University {ruyang@stu,hrtang@geli@ece}.pku.edu.cn + +2ARC Lab, Tencent PCG 3 Peng Cheng Laboratory + +{palchenli,yixiaoge,yingsshan}@tencent.com + +Abstract. Large Language Models (LLMs) have showcased impressive capabilities in text comprehension and generation, prompting research efforts towards video LLMs to facilitate human-AI interaction at the video level. However, how to effectively encode and understand videos in video-based dialogue systems remains to be solved. In this paper, we investigate a straightforward yet unexplored question: Can we feed all spatial-temporal tokens into the LLM, thus delegating the task of video sequence modeling to the LLMs? Surprisingly, this simple approach yields significant improvements in video understanding. Based upon this, we propose ST-LLM, an effective video-LLM baseline with Spatial-Temporal sequence modeling inside LLM. Furthermore, to address the overhead and stability issues introduced by uncompressed video tokens within LLMs, we develop a dynamic masking strategy with tailormade training objectives. For particularly long videos, we have also designed a global-local input module to balance efficiency and effectiveness. Consequently, we harness LLM for proficient spatial-temporal modeling, while upholding efficiency and stability. Extensive experimental results attest to the effectiveness of our method. Through a more concise model and training pipeline, ST-LLM establishes a new state-of-the-art result on VideoChatGPT-Bench and MVBench. Codes have been available at https://github.com/TencentARC/ST-LLM. + +Keywords: ST-LLM $\cdot$ Dynamic Masking $\cdot$ Global-Local Input + +# 1 Introduction + +Large Language Models (LLMs), such as GPT [19], PaLM [316] and LLaMA [43,44], have achieved remarkable success owing to their formidable language understanding and generation abilities, signaling a promising advancement towards artificial general intelligence. Driven by the widespread adoption of LLMs, research on Large Visual-Language Models (LVLMs) [14,29,55] has emerged to extend the capabilities of LLMs to process visual signals, which have revealed proficiency in image-based conversations. However, in contrast to image inputs, videos feature heavier input and additional temporal information. Developing a + +![](images/f59a836b920cc3b1c4307ed8c08f2b860e20ab2f43f53d9e34b006a741086425.jpg) +Fig. 1: Qualitative comparisons among top-performance video LLMs. We illustrate two cases from MVBench [25] (above) and one case from a YouTube video (below). + +large video-language model capable of effectively extracting meaningful spatial-temporal information from intricate video signals presents a significant challenge. + +Recently, several preliminary attempts have surfaced aimed at extending LLMs to facilitate video conversation [19,20,24,25,27,31,33,53], which unlock the capability of generating generic summaries of video content. Despite these efforts, existing video LLMs continue to exhibit shortcomings in achieving satisfactory performance in video comprehension, particularly in the realm of understanding content that relies heavily on temporal dynamics. As depicted in Fig. 11, contemporary top-performing video LLMs demonstrate robust understanding capabilities for actions reliant on static contexts. However, their proficiency in comprehending scenes involving motion remains constrained. They encounter challenges in discerning even the most basic direction of object movement, let alone the complex motion or scene transition. + +This issue may stem from the inherent difficulty of temporal modeling within conversational systems. Historically, achieving robust video encoding has often required substantially more time and memory resources than images, which is impractical based on LLMs. As illustrated by Fig. 2(a), early video LLMs tend to adopt mean pooling on temporal dimension [24,32,33,53], a method that, while efficient, is inadequate for handling dynamic temporal sequences. Therefore, recent models frequently integrate additional structures for temporal sampling and modeling [19,20,25,27]. While these methods offer improved effectiveness compared to average pooling, they entail increased storage and demand extensive GPU time for training from scratch, often involving two or even three-stage pretraining to align new modules. + +![](images/4c8d15fb042557baabf867ecd3c28e164eee717b27b8545061fc64a3e77a9d27.jpg) + +![](images/358cc03e5bbc1bc108e1a8b7d3f1bee6bcfee4fe6fb96c1e0a4b2226c201984f.jpg) + +![](images/0ae16f48316405930a7404358e1f3e1a766a004fa8736d4530498f554722a377.jpg) +(a) + +![](images/904b434c78bffa8a9a4d7df9a0171f05f4d91c38c25748d4f6761dabd240329f.jpg) +Fig. 2: (a) Comparison among different video LLMs based on the way of inputting visual tokens. (b) Quantitative comparison of representative video LLMs, presenting the average results of all 20 metrics and temporal-sensitive motion-related metrics from MVBench, which encompass Moving Direction, Moving Count, and Moving Attribute. + +![](images/49c933261dc499f7ed1f56945e79dbdec09abcc195a1ab4a7f3ce799365ed549.jpg) +(b) + +Inspired by the recent success of joint spatial-temporal-text modeling in visual generation [8,36], in this work, we investigate a simple yet unexplored idea: Given the robust sequence modeling capabilities inherent to LLMs, what if we input all visual tokens into the LLM and delegate the task of modeling spatial-temporal sequences to the LLM itself? Meanwhile, there are two additional challenges: (1) The inclusion of all visual tokens significantly increases the context length within LLMs, particularly for long videos, rendering the processing of a large number of frames unaffordable. (2) LLM may struggle to handle videos of varying lengths, potentially leading to hallucinations when there is a discrepancy between the number of frames in testing and training. + +To address these issues, we propose ST-LLM, a simple but powerful baseline of video LLM. As depicted in Fig. 2(a), with raw spatial-temporal tokens inside LLM, we harness the LLMs' robust sequence modeling capabilities for effective temporal modeling. Meanwhile, we introduce a novel dynamic video token masking strategy along with masking video modeling during training. With this approach, we reduce the length of sequences input to the LLM while significantly improving the robustness of videos of varying lengths during inference. To accommodate particularly long videos, we have devised a unique global-local input mechanism. This involves employing mean pooling of a large number of frames to generate residual input for a smaller subset of frames. Through this asymmetric design, we can process input from a large number of video frames while preserving the operations of modeling video tokens within LLMs. + +As a result, we maintain input tokens of comparable length to most video LLMs while effectively harnessing the LLM's capacity to comprehend videos. Furthermore, as ST-LLM does not introduce additional modules, it does not necessitate expensive alignment pre-training. This enables ST-LLM to directly leverage existing state-of-the-art image conversational models, resulting in significantly reduced GPU time compared to other state-of-the-art video LLMs. Extensive experiments are conducted to verify the effectiveness of ST-LLM. Qualitatively, as depicted in Fig. 1, ST-LLM demonstrates a superior ability to + +understand dynamics compared to other video LLMs. Quantitatively, ST-LLM achieves new state-of-the-art performance across various contemporary video benchmarks, including MVBench [25], VideoChatGPT-Bench [33] and zero-shot QA-Evaluation. Specifically, as illustrated in Fig. 2(b), the superiority of ST-LLM is particularly evident in metrics related to the temporal-sensitive motion, which underscores the expertise of our model in temporal understanding. + +# 2 Related Works + +LLMs and Image LLMs. Recent years have witnessed remarkable advancements in the evolution of Large Language Models (LLMs). Inspired by InstructGPT [35] and the widely recognized commercial model ChatGPT [34], the academic community has seen a proliferation of open-source LLMs, such as LLaMA [43], Alpaca [41], Vicuna [13], and LLaMA 2 [44], which have become foundational components for a myriad of research endeavors. Meanwhile, LLMs, being large-scale transformer models, inherently possess strong sequence modeling capabilities. This inspired us to explore the feasibility of entrusting the task of spatial-temporal sequence modeling to LLMs. The success of LLMs has spurred increasing interest in the development of Multimodal Large Language Models (MLLMs). Notable breakthroughs include works such as Flamingo [2], BLIP2 [23], and PaLM-E [16], which have successfully bridged the gap between vision models and LLMs. Inspired by the concept of instruction tuning for LLMs, a series of works [12, 14, 29, 55] leveraged open-source chatbots for visual instruction tuning, thus facilitating support for image-based conversations. It is noteworthy that in the pre-LLM era, video models were commonly constructed based on pre-trained 2D image models [7, 11, 30, 45]. This may be attributed to the significant computational demands of video processing, combined with the relatively limited scale and diversity of video data compared to images. Therefore, we believe that a well-pretrained image dialogue model could similarly offer training cost and performance benefits to video LLMs. + +Video LLMs. The emergence of MLLMs quickly extended into the domain of video as well [19,20,24,25,27,31,33,53]. Early models like VideoChat [24], VideoChatGPT [33] and Valley [32] generate video instruction tuning data through GPT to enable video conversations. At the model level, these approaches typically involve the use of mean pooling to aggregate the encoding results of individual frames before feeding them into the LLMs. While efficient, it is evident that this method is inadequate for effective temporal modeling. Hence, some subsequent models have initiated exploration into adaptations of image models to video-specific requirements. For instance, BT-Adapter [31] proposed a lightweight adapter to extend the capabilities of image LLMs. Chat-UniVi [20] introduces DPC-KNN to cluster dynamic visual tokens. Moreover, VideoChat2 [25] deviated from using CLIP [38] and directly introduced a dedicated video encoder. Although these methods offer improvements over mean pooling, the inclusion of additional modules typically necessitates intensive training. For example, in the case of VideoChat2, aligning the newly introduced video encoder entails + +![](images/7be2628aadc0f8c2d2bc5d6887ec862b3413425dbadc3730c2945f67a03faab4.jpg) +Fig.3: The overview of ST-LLM for generating responses on the input video and instructions. ST-LLM directly feeds the spatial-temporal sequence into the LLM and employs dynamic masking and global-local input for efficiency and robustness. + +a three-stage training involving 30 million visual-language samples. In contrast, ST-LLM adopts a more direct approach by entrusting the task of visual sequence modeling to LLMs, which is characterized by its conciseness, reduced training requirements, and, according to our experimental findings, superior effectiveness. Joint Spatial-Temporal Modeling. Prior to the emergence of LLMs, joint spatial-temporal modeling had demonstrated effectiveness in both video-only and video-language pretraining [26,42,49]. More recently, inspired by DIT [36], the popular video generation model Sora has shown success in simultaneously processing video tokens and text using the transformer. Despite its potential, joint spatial-temporal-text modeling remains relatively uncommon in video LLMs. While VideoChat2 employed a joint-ST video encoder, it necessitated extensive pretraining. A closer relative, InstructBLIP incorporated joint spatial-temporal sequences into LLM for zero-shot evaluation on video datasets. However, this approach led to significant hallucinations due to the absence of video instruction tuning. To the best of our knowledge, ST-LLM is the first open-source video LLM to adopt joint spatial-temporal-text modeling. Moreover, ST-LLM introduces a series of design innovations to mitigate the limitations associated with incorporating all video tokens within LLM. + +# 3 Methodology + +In this section, we elaborate on the structure and training methodology of ST-LLM. Firstly, we present a description of the model architecture, which incorporates all spatial-temporal tokens within the LLM, in Section 3.1. Subsequently, we delve into the masking mechanism and training objectives in Section 3.2. Finally, we introduce the global-local input approach in Section 3.3. The overarching framework of our method is illustrated in Fig. 3. + +# 3.1 Video Tokens Inside LLM + +As depicted in Fig. 3, visually, ST-LLM closely resembles an image chatbot, consisting of a visual encoder, a linear projection, and an LLM. Instead of opting for the commonly used CLIP-L/14 [38], we have selected BLIP-2 [23] as our visual encoder, whose Q-Former efficiently compresses redundant image tokens into fewer visual tokens. To facilitate the input of raw videos for the image encoder, we adopt the strategy of treating each frame as an individual image. Given a video $v$ with $T$ frames, we assume that the image encoder $f_{\mathrm{v}}$ encodes each frame into $K$ tokens. Hence, we obtain all visual tokens $V$ to be fed into the LLM as follows: + +$$ +V = f _ {\mathrm {p}} (f _ {\mathrm {v}} (v)), \mathrm {w h e r e} V = \{V _ {i} \} _ {i = 1} ^ {T}, V _ {i} = \{v _ {(i, j)} \} _ {j = 1} ^ {K}, \qquad (1) +$$ + +where $f_{\mathrm{p}}$ is the visual projection layer. Then, we concatenate all visual tokens to form a joint spatial-temporal [7] visual sequence $V \in \mathbb{R}^{(T*K)\times D}$ , where $D$ is the number of embedding dimensions. After being passed through the word embedding layers, the text tokens $C = \{c_i\}_{i=1}^N$ are directly concatenated with all spatial-temporal tokens, to organize the input $I$ to the LLM as follows: + +$$ +I = \left[ V: C \right] = \left\{v _ {(1, 1)}, v _ {(1, 2)}, \dots , v _ {(1, K)}, v _ {(2, 1)}, \dots , v _ {(T, K)}, c _ {1}, c _ {2}, \dots , c _ {N} \right\}. \tag {2} +$$ + +Indeed, several alternative strategies exist for organizing the input tokens before feeding them into the LLM. These include: (1) Adding separator tokens between frame-wise tokens or visual-text tokens, to demarcate the boundaries between modalities or frames. (2) Adding spatial-temporal position embeddings specifically for visual tokens, to facilitate better spatial-temporal relationship understandings. However, our experiments indicate that simplicity often yields the best. Adding numerous separator tokens introduces additional overhead without necessarily improving performance. Moreover, LLMs come with Rotary Position Embeddings [39] that are already effective in distinguishing the positions of all spatial-temporal tokens and text tokens. + +# 3.2 Training with Dynamic Masking + +With video tokens modeling inside LLM, we can leverage the potent sequence modeling capabilities inherent in LLMs to understand the information encapsulated within spatial-temporal sequences. However, integrating all video tokens into LLM significantly increases the context length, imposing a considerable computational burden and rendering it unfeasible to input additional frames for lengthy videos. Meanwhile, experimental findings suggest that during testing, utilizing uncompressed tokens is more sensitive to variations in the number of frames than mean pooling. Consequently, when a substantial dissonance exists between the number of frames in training and testing, a noticeable degradation in performance ensues. + +To tackle the aforementioned issues, we first propose masking the visual tokens during training. Masking modeling has achieved significant success in natural language processing [15] and video-language pretraining [26,31,42]. Although + +masking is rarely employed in autoregressive LLMs, given our aim to leverage LLMs for encoding spatial-temporal sequences, we can still utilize masking modeling. Specifically, we keep the text tokens in $I$ unchanged, while applying a mask to the video tokens. This mask randomly masks $\rho$ of all video tokens, irrespective of their position across different frames. Different from previous works [26,31], we adopt a dynamic masking strategy, where the masking rate is randomly sampled from a normal distribution as follows: + +$$ +\rho \sim \mathbb {N} (0. 5, \sigma), \quad 0. 3 \leq \rho \leq 0. 7, \tag {3} +$$ + +where $\sigma$ represents the variance of the normal distribution. This strategy ensures that the length of the spatial-temporal sequence varies continuously while maintaining a consistently high average masking rate of $50\%$ . As a result, it minimizes training costs and notably enhances robustness during inference. + +Furthermore, building upon dynamic masking, we have formulated the Masked Video Modeling (MVM) objective to encourage the LLM to grasp spatial-temporal dependencies. Unlike the expensive masked token recovery [6,37], we have opted for unmasked token reconstruction. Specifically, besides the masked sequence $I$ , we conduct an extra non-grad forward pass for the unmasked sequence $\hat{I} = [\hat{V}:C]$ , serving as the reference output. Then, we select unmasked tokens from $f_{llm}(I)$ and the corresponding tokens from $f_{llm}(\hat{I})$ based on their positions, computing the Mean Squared Error (MSE) between the selected pairs as follows: + +$$ +\mathcal {L} _ {m v m} = \frac {1}{(1 - \rho) K T} \sum_ {i = 1} ^ {T} \sum_ {j = 1} ^ {K} \left(v _ {i, j} ^ {- 1} - \hat {v} _ {i, j} ^ {- 1}\right) ^ {2}, (i, j) \notin M, \tag {4} +$$ + +where $M$ denotes the set of masked tokens' indices, $v_{i,j}^{-1}$ is from $f_{llm}(I)$ , and $\mathcal{L}_{mvm}$ represents our MVM objective. Finally, our overall loss $\mathcal{L}$ is composed of $\mathcal{L}_{mvm}$ and the LLM decoder loss $\mathcal{L}_{llm}$ , yielding $\mathcal{L} = \mathcal{L}_{mvm} + \mathcal{L}_{llm}$ . By integrating these two loss components, we can encourage the LLM to effectively respond to questions derived from the video content while improving its ability to model temporal and spatial dependencies. + +# 3.3 Global-Local Input + +Although dynamic masking partially mitigates the challenge of processing lengthy input sequences, it fails to fully resolve the issue for extremely long videos that entail inputting numerous frames, rendering the context length still impractical to handle. Therefore, we have devised an additional module to tackle the issue of excessively long videos. To elaborate, given a lengthy video with a large $T$ , we still commence by encoding each frame individually to derive $V$ . Subsequently, we proceed to derive the global video representation $V_{0}$ through average pooling on the frame-wise tokens: + +$$ +v _ {(0, j)} = \frac {1}{T} \sum_ {i = 1} ^ {T} v _ {(i, j)}, \quad V _ {0} = \left\{v _ {(0, j)} \right\} _ {j = 1} ^ {K}. \tag {5} +$$ + +Next, we average-sample $\overline{T}$ frames from the T total frames. All tokens from these $\overline{T}$ frames are concatenated to produce a joint spatial-temporal sequence $\overline{V} \in \mathbb{R}^{(\overline{T}*K) \times D}$ , which serves as the local video representation. Finally, the global-local input for the LLM, denoted as $\overline{I}$ , is constructed as follows: + +$$ +\bar {I} = \left[ \bar {V} + f _ {m} \left(V _ {0}\right): C \right], \tag {6} +$$ + +where $f_{m}$ is a simple MLP projector with upsampling projection initialized with zeros. With this global-local input design, the low fps spatial-temporal sequences within the LLM can gradually incorporate information from the high fps branch. This approach allows the model to benefit from the LLM's ability to model temporal sequences within a limited context while also considering the global information of long videos. + +# 4 Experiments + +In this section, we have performed comprehensive experimental evaluations of ST-LLM, covering crucial settings, comparisons, and ablation studies. For a more detailed account of experimental settings, ablation studies, visualizations, and limitations analysis, please consult the appendix. + +# 4.1 Experiment Setup + +Benchmarks. To evaluate the video understanding capabilities of ST-LLM, we primarily conduct assessments on three benchmarks: MVBench [25], VideoChat-GPT Bench [33], and zero-shot video QA benchmark. MVBench comprises 20 challenging video tasks, each consisting of 200 samples in the form of multiple-choice questions. These tasks provide a comprehensive and objective assessment of a model's ability to understand videos. VideoChatGPT-Bench gathers videos from ActivityNet [10] and uses GPT to evaluate the quality of video conversations across five dimensions. Additionally, the zero-shot video QA benchmark entails GPT-based assessments of various open-source video QA datasets. + +Implementation Details. In our study, we adopt state-of-the-art image dialogue models as the baseline and proceed directly with video instruction tuning. Unless otherwise specified, our model is initialized with InstructBLIP [14] in all experiments, whose LLM is Vicuna-v1.1 [13] with 7B parameters. Additionally, we employ Minigpt4-v1 [55] as a basic model to assess the effectiveness across various image models in Table 6. Following previous research [20, 27, 28], we adopt full-finetuning on the LLM to ensure a fair comparison. The results of LoRA are also included as ablation results in Table 6. $\sigma$ is set as 0.1, ensuring that over $95\%$ of the mask rate falls within the range of 0.3 to 0.7. During training, we sample 16 frames per video. Considering dynamic masking, an average of 256 tokens per video are input into the LLM, which is the same as LLaVA [29]. For all benchmarks during inference, frames are sampled at a frame rate of 1, and the number of local frames is kept between 4 and 16. This strategy closely + +Table 1: Comparisons on MVBench. Except for BLIP2 and Otter, all models are built upon LLaMA-1-7B for fair comparisons. "Avg" denotes the average of all 20 metrics. The leaderboards for each task can be found in the Appendix. The compared methods include: (1)Rd: random guesses; (2) mOI: mPLUG-Owl-I 51; (3) LMA: LLaMA-Adapter 54 (4)B2: BLIP-2 23; (5)OI: Otter-I 22; (6)MG: MiniGPT-4 55; (7)IB: InstructBLIP 14; (8)LV: LLaVA 29; (9)OV: Otter-V 22; (10)mOV: mPLUG-Owl-V 51; (11)VCG: VideoChatGPT 33; (12)VLM: VideoLLaMA 53; (13)VC: VideoChat 24; (14)VC2: VideoChat2 25. + +
MetricRdmOILMAB2OIMGIBLVOVmOVVCGVLMVCVC2Ours
AS25.025.023.024.534.516.020.028.023.022.023.527.533.566.066.0
AP25.020.028.029.032.018.016.539.523.028.026.025.526.547.553.5
AA33.344.551.033.539.526.046.063.027.534.062.051.050.083.584.0
FA25.027.030.017.030.521.524.530.527.029.022.529.033.549.544.0
UA25.023.533.042.038.516.046.039.029.529.026.539.040.560.058.5
OE33.336.053.551.548.529.551.053.053.040.554.048.053.058.080.5
OI25.024.032.526.044.025.526.041.028.027.028.040.540.571.573.5
OS33.334.033.531.029.513.037.541.533.031.540.038.030.042.538.5
MD25.023.025.525.519.011.522.023.024.527.023.022.525.523.042.5
AL25.024.021.526.025.512.023.020.523.523.020.022.527.023.031.0
ST25.034.530.532.555.09.546.545.027.529.031.043.048.588.586.5
AC33.334.529.025.520.032.542.534.026.031.530.534.035.039.036.5
MC25.022.022.530.032.515.526.520.528.527.025.522.520.542.056.5
MA33.331.541.540.028.58.040.538.518.040.039.532.542.558.578.5
SC33.340.039.542.039.034.032.047.038.544.048.545.546.044.043.0
FP25.024.025.027.028.026.025.525.022.024.029.032.526.549.044.5
CO33.337.031.530.027.029.530.536.022.031.033.040.041.036.546.5
EN25.025.522.526.032.019.030.527.023.526.029.530.023.535.034.5
ER20.021.028.037.036.59.930.526.519.520.526.021.023.540.541.5
CI30.937.032.031.036.53.038.042.019.029.535.537.036.065.558.5
Avg27.329.431.731.433.518.832.536.026.829.732.734.135.551.154.9
+ +aligns with practical scenarios and involves fewer frames than most video LLMs. For VideoChatGPT-Bench and the video QA benchmark, which may involve videos spanning several minutes, we utilize the global-local input approach. In this configuration, the global frame number is set to 64. All videos are resized into $224^{*}224$ . We train ST-LLM using a learning rate of 2e-5 and a batch size of 128 for 2 epochs. + +Training Datasets. Citing findings from Jin et al. [20], the sequential process of first conducting image instruction tuning, followed by video instruction tuning and image-video joint pretraining, had a negligible impact on video dialogue performance. Hence, in light of InstructBLIP's prior training on extensive image data, we exclusively focus on video instruction tuning. Following [25], we leverage video instruction data sourced from a variety of datasets, including VideoChatGPT-100k [33], VideoChat-11k [24], Webvid [5], NExT-QA [47], CLEVRER [52], Kinetics-710 [21], and Something-Something-2 [17]. All samples are uniformly formatted in accordance with [25]. More details about the + +Table 2: Comparisons on video-based generative performance benchmarking. + +
MethodCorrectDetailContextTemporalConsistMean Score
VideoLLaMA 531.962.182.161.821.791.98
LLaMA-Adapter 542.032.322.301.982.152.16
VideoChat 242.232.502.531.942.242.29
VideoChatGPT 332.402.522.621.982.372.38
BT-Adapter 312.682.693.272.342.462.69
VTimeLLM 192.783.103.402.492.472.85
Chat-UniVi 202.892.913.462.892.812.99
LLaMA-VID-7B 272.963.003.532.462.512.89
LLaMA-VID-13B 273.073.053.602.582.632.99
VideoChat2 253.022.883.512.662.812.98
ST-LLM (Ours)3.233.053.742.932.813.15
+ +dataset and the organization of instructions can be found in the appendix. An experiment on joint image-video training is also presented in the appendix. + +# 4.2 Quantitative Result + +MVBench Performance. The evaluation results on MVBench are presented in Table 1. The evaluation metrics included Action Sequence (AS), Action Prediction (AP), Action Antonym (AA), Fine-grained Action (FA), Unexpected Action (UA), Object Existence (OE), Object Interaction (OI), Object Shuffle (OS), Moving Direction (MD), Action Localization (AL), Scene Transition (ST), Action Count (AC), Moving Count (MC), Moving Attribute (MA), State Change (SC), Fine-grained Pose (FP), Character Order (CO), Egocentric Navigation (EN), Episodic Reasoning (ER), Counterfactual Inference (CI), and the average of all 20 metrics (Avg). It is evident that on such a challenging benchmark, the performance of most multimodal LLMs falls significantly short of satisfactory levels, even barely surpassing random performance by less than $10\%$ . Our only formidable competitor is VideoChat2 [25]. Through direct substitution of CLIP with a dedicated video encoder, coupled with a three-stage training process and comprehensive instruction tuning for the visual encoder, Q-Former, and LLM, VideoChat2 has yielded results that significantly surpass those of previous models. In contrast, we solely carried out single-stage video-only instruction tuning while maintaining the weights of CLIP and Q-Former frozen. Even under these conditions, ST-LLM still attained the top performance, with an average score surpassing VideoChat2 by a notable margin of $3.8\%$ . Specifically, ST-LLM showcased leading results across all 11 tasks, exhibiting a particularly significant advantage in motion-related tasks where other models struggled. In three motion-related tasks, ST-LLM achieved an average score of $59.2\%$ , far outperforming VideoChat2's $36.3\%$ . However, ST-LLM encounters challenges in fine-grained tasks, such as Fine-grained Action (FA) and Fine-grained Pose (FP). This could be attributed to CLIP's limitations as an image encoder, which may struggle to capture low-level spatial-temporal features required for these tasks. + +Table 3: Comparisons on zero-shot question-answering, including MSVD-QA [46], MSRVTT-QA [48], and ActivityNet-QA [10]. + +
MethodLLMMSVD-QAMSRVTT-QAActivityNet-QA
AccScoreAccScoreAccScore
FrozenBiLM 50DeBERTa-V232.2-16.8-24.7-
VideoLLaMA 53Vicuna-7B51.62.529.61.812.41.1
LLaMA-Adapter 54LLaMA-7B54.93.143.82.734.22.7
VideoChat 24Vicuna-7B56.32.845.02.526.52.2
VideoChatGPT 33Vicuna-7B64.93.349.32.835.22.7
BT-Adapter 31Vicuna-7B67.53.757.03.245.73.2
Chat-UniVi 20Vicuna-7B65.03.654.63.145.83.2
LLaMA-VID 27Vicuna-7B69.73.757.73.247.43.3
LLaMA-VID 27Vicuna-13B70.03.758.93.347.53.3
VideoChat2 25Vicuna-7B70.03.954.13.349.13.3
ST-LLM (Ours)Vicuna-7B74.63.963.23.450.93.3
+ +VideoChatGPT-Bench Performance. VideoChatGPT-Bench, also referred to as the Video-Based Generative Performance Benchmark, primarily evaluates the capability of video conversation. Utilizing GPT-3.5 [34], it assigns a score to the generated content across five aspects: Correctness of Information (Correct), Detail Orientation (Detail), Contextual Understanding (Context), Temporal Understanding (Temporal), and Consistency (Consist). As depicted in Table 2, the recently proposed state-of-the-art models [20,25,27] exhibit similar performance levels on this benchmark. However, due to the inherent instability of GPT evaluation, there is considerable variation in the performance of each model across different tasks. Despite this variability, ST-LLM consistently performs admirably across all tasks, while demonstrating significant advantages in terms of the mean score compared to previous models. + +Zero-shot Video-question Answering Performance. In Table 3 we present the zero-shot video question-answering performance on several commonly used video-text datasets, including MSVD-QA [46], MSRVTT-QA [48], and ActivityNet-QA [10]. Following the methodology outlined in [33], we utilize GPT-3.5 to evaluate the accuracy and score of the generated results. As is depicted, STLLM demonstrates significant advantages on MSVD and MSRVTT, surpassing the previous SOTAs by $4.6\%$ and $4.3\%$ respectively. However, STLLM exhibits a slightly smaller advantage compared to VideoChat2 on the ActivityNet. This could be attributed to the requirement for a potent video encoder for this dataset. + +# 4.3 Ablations and Analysis + +Ways of Inputting video tokens. The core of this work lies in modeling joint spatial-temporal video tokens using LLM. Table 4 presents the results of different methods of inputting video tokens. As depicted, without video tuning, the approach of joint spatial-temporal input is considerably less effective compared to directly averaging the input over the temporal dimension. This discrepancy might stem from inconsistencies between training and testing. However, after + +Table 4: The ablation study on the methods of inputting video tokens to LLMs. The baseline configurations include no instruction tuning and mean pooling input. + +
MethodAvg on MV bench
Instructclip Baseline
Mean Pooling42.0
S-T Tokens in LLM35.5
Video Instruction Tuning
Mean Pooling47.8
S-T Tokens in LLM53.8
+Masking & MVM Loss54.9
+ +![](images/d51c3b50eefedfff5883c76f9d933ac6618712042adc1ba0dff8d821206d2ff2.jpg) +Fig. 4: The ablation study on the effect of dynamic masking and MVM loss. + +Table 5: The ablation study on global-local input module. The "Local" and "Global" mean the joint temporal-temporal sequence and average pooling sequence respectively. + +
MethodAvg on MVBenchVideo-ChatGPT Bench
CorrectDetailContextTemporalConsist
Global Only48.33.012.923.582.612.61
Local Only54.93.082.993.632.752.69
Local+Global (simply add)50.13.072.883.582.592.50
Local+Global (adapter)54.73.233.053.742.932.81
+ +training, the effectiveness of the joint spatial-temporal input significantly surpasses that of the mean pooling input, indicating that LLMs can effectively model spatial-temporal sequences, and directly inputting all tokens is a superior approach compared to compression before input. Furthermore, incorporating dynamic masking and the MVM loss based on the joint spatial-temporal sequence has further improved the performance of our model. + +Dynamic Masking Strategy and MVM Loss. In Fig. 4, we conduct a more detailed ablation study on the effect of dynamic masking and MVM loss. It is evident that when the training frame count remains fixed, the performance of inputting all video tokens is not very robust. Particularly, when there is a discrepancy between the training and testing frame, the performance noticeably decreases. Meanwhile, when we randomly mask video tokens at a static $50\%$ rate, this issue persists, with effectiveness akin to training with a fixed 8-frame setup. However, employing dynamic masking shows some improvement in overall performance, and notably, the robustness to varying testing frame counts is significantly enhanced. Furthermore, the inclusion of the MVM loss built upon the masking further enhances performance, with greater improvements observed as the sequence length increases. + +Global-Local Input Module. The ablation results on the global-local module are presented in Table 5 where we compare the performance of global features extracted from more frames and local features from fewer frames, as well as their combination. As shown, on MVBench, the performance of global features + +Table 6: The ablation study on the basic image models and the fine-tuning methodologies of LLM. We report the Avg on MVBench. + +
Baseline ModelLLMLLM ParametersMeanpooling BaselineST-LLM Full Model
MiniGPT-4 55Vicuna-7B v0LoRA40.945.6
Full Fine-tuning46.450.5
InstructBLIP 14Vicuna-7B v1.1LoRA45.251.9
Full Fine-tuning47.854.9
+ +Table 7: The analysis on the designs of inputting S-T video tokens into LLM. + +
MethodAvg on MVBench
S-T Tokens in LLM (baseline)54.9
+frame-frame separators54.0
+video-text separators54.8
+S-T position embeddings52.2
+ +Table 8: The analysis on the designs of dynamic masking. + +
MethodAvg on MVBatch
w/o masking (baseline)53.8
ρ ~ U(0.3, 0.7)54.2
ρ ~ N(0.5, 0.2)54.3
ρ ~ N(0.5, 0.1)54.9
+ +is significantly inferior to the joint S-T features, so combining global and local features did not lead to improvement. This may be due to the fact that MVBench mainly consists of short videos lasting a few to a dozen seconds, where a large number of frames might not be necessary. However, on VideoChatGPT Bench, where videos are parsed at the minute level, the discrepancy between global and local features has notably diminished, with their amalgamation resulting in a discernible enhancement. This underscores the efficacy of the Global-Local Input module in processing extended video content. Moreover, the direct incorporation of global and local features resulted in a performance decline, presumably due to its interference with the representation learning of the LLM. + +Basic Image Models & LLMs. In Table 6, we conduct ablations on both the basic image chatbots and the LLM fine-tuning strategy. As depicted in the table, regardless of whether it's LoRA [18] tuning or full fine-tuning, or different basic image models, ST-LLM demonstrates notable enhancements compared to the mean pooling baseline. This suggests that our approach can be adeptly applied across diverse scenarios, such as employing more resource-efficient LoRA tuning on larger models, or transferring to more potent image-dialogue models. + +Designs of ST-LLM. In Table 7 and Table 8, we investigated several alternative design possibilities for ST-LLM. Primarily, we explored various separator designs. It's noteworthy that almost all cross-modal LLMs incorporate separator designs, with some models even integrating separators between multiple images or frames [2,4]. However, we discovered that when employing LLMs to model spatial-temporal sequences, these separators are superfluous. Subsequently, we address positional encodings. Numerous approaches integrate temporal positional encodings before inputting video tokens into LLMs [32,33,53]. However, in our methodology, we observed that incorporating absolute positional encodings to all visual tokens significantly impairs performance. Lastly, we consider the design of dynamic masking. Our findings suggest that all distributions + +![](images/42f1b8056b5f8ef0be156a6cf389d6863825d969515d934706bc96954dbd8798.jpg) +Fig. 5: Qualitative results from real-world videos. + +of masking rates can yield some improvement, yet distributions with smaller standard deviations exhibit superior performance. + +# 4.4 Qualitative Results + +In Fig. 5, we present a qualitative comparison. As illustrated, ST-LLM excels in adhering to instructions and delivering precise responses. More importantly, ST-LLM showcases superior sensitivity to temporal sequences and actions, thereby validating the effectiveness of our modeling approach. For further quantitative results and failure cases, please refer to the appendix. + +# 5 Conclusion + +In this paper, we present ST-LLM, a straightforward yet robust video large language model. Our aim is to achieve effective video comprehension by leveraging LLM to model video tokens, marking the inception of a joint spatial-temporal-text modeling paradigm. Additionally, we introduce a dynamic masking strategy and a global-local input module to enhance this framework. Remarkably, ST-LLM excels in encoding and comprehending spatial-temporal sequences while addressing concerns related to efficiency, stability, and modeling lengthy videos, all with reduced training resource requirements. Through extensive experimental analysis, we demonstrate the effectiveness of ST-LLM and its internal design, resulting in state-of-the-art performance across multiple video LLM benchmarks. + +Acknowledgements. This work was supported by National Natural Science Foundation of China (No. 62172021). We are also grateful to the support from Guangdong Provincial Key Laboratory of Ultra High Definition Immersive Media Technology. + +*: Equal Contribution, work done during internship at ARC Lab, Tencent PCG. + +# References + +1. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023) +2. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems 35, 23716-23736 (2022) +3. Anil, R., Dai, A.M., First, O., Johnson, M., Lepikhin, D., Passos, A., Shakeri, S., Taropa, E., Bailey, P., Chen, Z., et al.: Palm 2 technical report. arXiv preprint arXiv:2305.10403 (2023) +4. Awadalla, A., Gao, I., Gardner, J., Hessel, J., Hanafy, Y., Zhu, W., Marathe, K., Bitton, Y., Gadre, S., Sagawa, S., et al.: Openflamingo: An open-source framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390 (2023) +5. Bain, M., Nagrani, A., Varol, G., Zisserman, A.: Frozen in time: A joint video and image encoder for end-to-end retrieval. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1728-1738 (2021) +6. Bao, H., Dong, L., Piao, S., Wei, F.: Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254 (2021) +7. Bertasius, G., Wang, H., Torresani, L.: Is space-time attention all you need for video understanding? In: ICML. vol. 2, p. 4 (2021) +8. Brooks, T., Peebles, B., Holmes, C., DePue, W., Guo, Y., Jing, L., Schnurr, D., Taylor, J., Luhman, T., Luhman, E., Ng, C., Wang, R., Ramesh, A.: Video generation models as world simulators (2024), https://openai.com/research/video-generation-models-as-world-simulators +9. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020) +0. Caba Heilbron, F., Escorcia, V., Ghanem, B., Carlos Niebles, J.: Activitynet: A large-scale video benchmark for human activity understanding. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 961-970 (2015) +1. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6299-6308 (2017) +2. Chen, J., Zhu, D., Shen, X., Li, X., Liu, Z., Zhang, P., Krishnamoorthi, R., Chandra, V., Xiong, Y., Elhoseiny, M.: Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478 (2023) +3. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., et al.: Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023) (2023) + +14. Dai, W., Li, J., Li, D., Tiong, A., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: towards general-purpose vision-language models with instruction tuning. arxiv. Preprint posted online on June 15, 2023 (2023) +15. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018) +16. Driess, D., Xia, F., Sajjadi, M.S., Lynch, C., Chowdhery, A., Ichter, B., Wahid, A., Tompson, J., Vuong, Q., Yu, T., et al.: Palm-e: An embodied multimodal language model. arXiv preprint arXiv:2303.03378 (2023) +17. Goyal, R., Ebrahimi Kahou, S., Michalski, V., Materzynska, J., Westphal, S., Kim, H., Haenel, V., Fruend, I., Yianilos, P., Mueller-Freitag, M., et al.: The" something something" video database for learning and evaluating visual common sense. In: Proceedings of the IEEE international conference on computer vision. pp. 5842-5850 (2017) +18. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models (2021) +19. Huang, B., Wang, X., Chen, H., Song, Z., Zhu, W.: Vtimellm: Empower llm to grasp video moments. arXiv preprint arXiv:2311.18445 (2023) +20. Jin, P., Takanobu, R., Zhang, C., Cao, X., Yuan, L.: Chat-univi: Unified visual representation empowers large language models with image and video understanding. arXiv preprint arXiv:2311.08046 (2023) +21. Kay, W., Carreira, J., Simonyan, K., Zhang, B., Hillier, C., Vijayanarasimhan, S., Viola, F., Green, T., Back, T., Natsev, P., et al.: The kinetics human action video dataset. arXiv preprint arXiv:1705.06950 (2017) +22. Li, B., Zhang, Y., Chen, L., Wang, J., Yang, J., Liu, Z.: Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726 (2023) +23. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023) +24. Li, K., He, Y., Wang, Y., Li, Y., Wang, W., Luo, P., Wang, Y., Wang, L., Qiao, Y.: Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355 (2023) +25. Li, K., Wang, Y., He, Y., Li, Y., Wang, Y., Liu, Y., Wang, Z., Xu, J., Chen, G., Luo, P., et al.: Mvbench: A comprehensive multi-modal video understanding benchmark. arXiv preprint arXiv:2311.17005 (2023) +26. Li, K., Wang, Y., Li, Y., Wang, Y., He, Y., Wang, L., Qiao, Y.: Unmasked teacher: Towards training-efficient video foundation models. arXiv preprint arXiv:2303.16058 (2023) +27. Li, Y., Wang, C., Jia, J.: Llama-vid: An image is worth 2 tokens in large language models. arXiv preprint arXiv:2311.17043 (2023) +28. Lin, B., Zhu, B., Ye, Y., Ning, M., Jin, P., Yuan, L.: Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122 (2023) +29. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv preprint arXiv:2304.08485 (2023) +30. Liu, R., Huang, J., Li, G., Feng, J., Wu, X., Li, T.H.: Revisiting temporal modeling for clip-based image-to-video knowledge transferring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6555-6564 (2023) +31. Liu, R., Li, C., Ge, Y., Shan, Y., Li, T.H., Li, G.: One for all: Video conversation is feasible without video instruction tuning. arXiv preprint arXiv:2309.15785 (2023) + +32. Luo, R., Zhao, Z., Yang, M., Dong, J., Qiu, M., Lu, P., Wang, T., Wei, Z.: Valley: Video assistant with large language model enhanced ability. arXiv preprint arXiv:2306.07207 (2023) +33. Maaz, M., Rasheed, H., Khan, S., Khan, F.S.: Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424 (2023) +34. OpenAI: Chatgpt. https://openai.com/blog/chatgpt (2023) +35. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems 35, 27730-27744 (2022) +36. Peebles, W., Xie, S.: Scalable diffusion models with transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4195-4205 (2023) +37. Peng, Z., Dong, L., Bao, H., Ye, Q., Wei, F.: Beit v2: Masked image modeling with vector-quantized visual tokenizers. arXiv preprint arXiv:2208.06366 (2022) +38. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) +39. Su, J., Ahmed, M., Lu, Y., Pan, S., Bo, W., Liu, Y.: Roformer: Enhanced transformer with rotary position embedding. Neurocomputing 568, 127063 (2024) +40. Sun, Q., Fang, Y., Wu, L., Wang, X., Cao, Y.: Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389 (2023) +41. Taori, R., Gulrajani, I., Zhang, T., Dubois, Y., Li, X., Guestrin, C., Liang, P., Hashimoto, T.B.: Stanford alpaca: An instruction-following llama model (2023) +42. Tong, Z., Song, Y., Wang, J., Wang, L.: Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. Advances in neural information processing systems 35, 10078-10093 (2022) +43. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023) +44. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023) +45. Tran, D., Bourdev, L., Fergus, R., Torresani, L., Paluri, M.: Learning spatiotemporal features with 3d convolutional networks. In: Proceedings of the IEEE international conference on computer vision. pp. 4489-4497 (2015) +46. Wu, Z., Yao, T., Fu, Y., Jiang, Y.G.: Deep learning for video classification and captioning. In: Frontiers of multimedia research, pp. 3-29. ACM (2017) +47. Xiao, J., Shang, X., Yao, A., Chua, T.S.: Next-qa: Next phase of question-answering to explaining temporal actions. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9777-9786 (2021) +48. Xu, J., Mei, T., Yao, T., Rui, Y.: Msr-vtt: A large video description dataset for bridging video and language. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5288-5296 (2016) +49. Xue, H., Sun, Y., Liu, B., Fu, J., Song, R., Li, H., Luo, J.: Clip-vip: Adapting pre-trained image-text model to video-language representation alignment. arXiv preprint arXiv:2209.06430 (2022) + +50. Yang, A., Miech, A., Sivic, J., Laptev, I., Schmid, C.: Zero-shot video question answering via frozen bidirectional language models. Advances in Neural Information Processing Systems 35, 124-141 (2022) +51. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178 (2023) +52. Yi, K., Gan, C., Li, Y., Kohli, P., Wu, J., Torralba, A., Tenenbaum, J.B.: Clevrer: Collision events for video representation and reasoning. arXiv preprint arXiv:1910.01442 (2019) +53. Zhang, H., Li, X., Bing, L.: Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858 (2023) +54. Zhang, R., Han, J., Zhou, A., Hu, X., Yan, S., Lu, P., Li, H., Gao, P., Qiao, Y.: Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199 (2023) +55. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023) \ No newline at end of file diff --git a/2024/ST-LLM_ Large Language Models Are Effective Temporal Learners/images.zip b/2024/ST-LLM_ Large Language Models Are Effective Temporal Learners/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1df28334db85777277b937b958b2ed2d42fe7727 --- /dev/null +++ b/2024/ST-LLM_ Large Language Models Are Effective Temporal Learners/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c279ff8c737d78c6bd19b68aa3cfb4754e4df529f85b933329457a8c30e627ec +size 740079 diff --git a/2024/ST-LLM_ Large Language Models Are Effective Temporal Learners/layout.json b/2024/ST-LLM_ Large Language Models Are Effective Temporal Learners/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7de7526e7a2885f7d95b0548b7848aae098cafb1 --- /dev/null +++ b/2024/ST-LLM_ Large Language Models Are Effective Temporal Learners/layout.json @@ -0,0 +1,7786 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 140, + 111, + 474, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 111, + 474, + 148 + ], + "spans": [ + { + "bbox": [ + 140, + 111, + 474, + 148 + ], + "type": "text", + "content": "ST-LLM: Large Language Models Are Effective Temporal Learners" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 132, + 168, + 480, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 168, + 480, + 181 + ], + "spans": [ + { + "bbox": [ + 132, + 168, + 480, + 181 + ], + "type": "text", + "content": "Ruyang Liu\\*1,3 Chen Li\\*2 Haoran Tang1,3 Yixiao Ge2 Ying Shan2 Ge Li" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 146, + 190, + 466, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 190, + 466, + 201 + ], + "spans": [ + { + "bbox": [ + 146, + 190, + 466, + 201 + ], + "type": "text", + "content": "1 School of Electronic and Computer Engineering, Shenzhen Graduate School," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 168, + 201, + 446, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 201, + 446, + 212 + ], + "spans": [ + { + "bbox": [ + 168, + 201, + 446, + 212 + ], + "type": "text", + "content": "Peking University {ruyang@stu,hrtang@geli@ece}.pku.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 189, + 212, + 424, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 212, + 424, + 223 + ], + "spans": [ + { + "bbox": [ + 189, + 212, + 424, + 223 + ], + "type": "text", + "content": "2ARC Lab, Tencent PCG 3 Peng Cheng Laboratory" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 207, + 224, + 406, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 224, + 406, + 235 + ], + "spans": [ + { + "bbox": [ + 207, + 224, + 406, + 235 + ], + "type": "text", + "content": "{palchenli,yixiaoge,yingsshan}@tencent.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 267, + 455, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 267, + 455, + 499 + ], + "spans": [ + { + "bbox": [ + 159, + 267, + 455, + 499 + ], + "type": "text", + "content": "Abstract. Large Language Models (LLMs) have showcased impressive capabilities in text comprehension and generation, prompting research efforts towards video LLMs to facilitate human-AI interaction at the video level. However, how to effectively encode and understand videos in video-based dialogue systems remains to be solved. In this paper, we investigate a straightforward yet unexplored question: Can we feed all spatial-temporal tokens into the LLM, thus delegating the task of video sequence modeling to the LLMs? Surprisingly, this simple approach yields significant improvements in video understanding. Based upon this, we propose ST-LLM, an effective video-LLM baseline with Spatial-Temporal sequence modeling inside LLM. Furthermore, to address the overhead and stability issues introduced by uncompressed video tokens within LLMs, we develop a dynamic masking strategy with tailormade training objectives. For particularly long videos, we have also designed a global-local input module to balance efficiency and effectiveness. Consequently, we harness LLM for proficient spatial-temporal modeling, while upholding efficiency and stability. Extensive experimental results attest to the effectiveness of our method. Through a more concise model and training pipeline, ST-LLM establishes a new state-of-the-art result on VideoChatGPT-Bench and MVBench. Codes have been available at https://github.com/TencentARC/ST-LLM." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 160, + 509, + 420, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 509, + 420, + 520 + ], + "spans": [ + { + "bbox": [ + 160, + 509, + 420, + 520 + ], + "type": "text", + "content": "Keywords: ST-LLM " + }, + { + "bbox": [ + 160, + 509, + 420, + 520 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 509, + 420, + 520 + ], + "type": "text", + "content": " Dynamic Masking " + }, + { + "bbox": [ + 160, + 509, + 420, + 520 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 509, + 420, + 520 + ], + "type": "text", + "content": " Global-Local Input" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 542, + 230, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 542, + 230, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 542, + 230, + 555 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": "Large Language Models (LLMs), such as GPT [19], PaLM [316] and LLaMA [43,44], have achieved remarkable success owing to their formidable language understanding and generation abilities, signaling a promising advancement towards artificial general intelligence. Driven by the widespread adoption of LLMs, research on Large Visual-Language Models (LVLMs) [14,29,55] has emerged to extend the capabilities of LLMs to process visual signals, which have revealed proficiency in image-based conversations. However, in contrast to image inputs, videos feature heavier input and additional temporal information. Developing a" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 116, + 477, + 305 + ], + "blocks": [ + { + "bbox": [ + 137, + 116, + 477, + 305 + ], + "lines": [ + { + "bbox": [ + 137, + 116, + 477, + 305 + ], + "spans": [ + { + "bbox": [ + 137, + 116, + 477, + 305 + ], + "type": "image", + "image_path": "f59a836b920cc3b1c4307ed8c08f2b860e20ab2f43f53d9e34b006a741086425.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 312, + 482, + 338 + ], + "lines": [ + { + "bbox": [ + 130, + 312, + 482, + 338 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 482, + 338 + ], + "type": "text", + "content": "Fig. 1: Qualitative comparisons among top-performance video LLMs. We illustrate two cases from MVBench [25] (above) and one case from a YouTube video (below)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 369, + 482, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 482, + 393 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 482, + 393 + ], + "type": "text", + "content": "large video-language model capable of effectively extracting meaningful spatial-temporal information from intricate video signals presents a significant challenge." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 398, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 398, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 398, + 482, + 529 + ], + "type": "text", + "content": "Recently, several preliminary attempts have surfaced aimed at extending LLMs to facilitate video conversation [19,20,24,25,27,31,33,53], which unlock the capability of generating generic summaries of video content. Despite these efforts, existing video LLMs continue to exhibit shortcomings in achieving satisfactory performance in video comprehension, particularly in the realm of understanding content that relies heavily on temporal dynamics. As depicted in Fig. 11, contemporary top-performing video LLMs demonstrate robust understanding capabilities for actions reliant on static contexts. However, their proficiency in comprehending scenes involving motion remains constrained. They encounter challenges in discerning even the most basic direction of object movement, let alone the complex motion or scene transition." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "type": "text", + "content": "This issue may stem from the inherent difficulty of temporal modeling within conversational systems. Historically, achieving robust video encoding has often required substantially more time and memory resources than images, which is impractical based on LLMs. As illustrated by Fig. 2(a), early video LLMs tend to adopt mean pooling on temporal dimension [24,32,33,53], a method that, while efficient, is inadequate for handling dynamic temporal sequences. Therefore, recent models frequently integrate additional structures for temporal sampling and modeling [19,20,25,27]. While these methods offer improved effectiveness compared to average pooling, they entail increased storage and demand extensive GPU time for training from scratch, often involving two or even three-stage pretraining to align new modules." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "text", + "content": "Liu. et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 114, + 231, + 171 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 231, + 171 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 231, + 171 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 231, + 171 + ], + "type": "image", + "image_path": "4c8d15fb042557baabf867ecd3c28e164eee717b27b8545061fc64a3e77a9d27.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 234, + 114, + 332, + 171 + ], + "blocks": [ + { + "bbox": [ + 234, + 114, + 332, + 171 + ], + "lines": [ + { + "bbox": [ + 234, + 114, + 332, + 171 + ], + "spans": [ + { + "bbox": [ + 234, + 114, + 332, + 171 + ], + "type": "image", + "image_path": "358cc03e5bbc1bc108e1a8b7d3f1bee6bcfee4fe6fb96c1e0a4b2226c201984f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 134, + 173, + 231, + 219 + ], + "blocks": [ + { + "bbox": [ + 134, + 173, + 231, + 219 + ], + "lines": [ + { + "bbox": [ + 134, + 173, + 231, + 219 + ], + "spans": [ + { + "bbox": [ + 134, + 173, + 231, + 219 + ], + "type": "image", + "image_path": "0ae16f48316405930a7404358e1f3e1a766a004fa8736d4530498f554722a377.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 230, + 216, + 239, + 223 + ], + "lines": [ + { + "bbox": [ + 230, + 216, + 239, + 223 + ], + "spans": [ + { + "bbox": [ + 230, + 216, + 239, + 223 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 234, + 173, + 332, + 219 + ], + "blocks": [ + { + "bbox": [ + 234, + 173, + 332, + 219 + ], + "lines": [ + { + "bbox": [ + 234, + 173, + 332, + 219 + ], + "spans": [ + { + "bbox": [ + 234, + 173, + 332, + 219 + ], + "type": "image", + "image_path": "904b434c78bffa8a9a4d7df9a0171f05f4d91c38c25748d4f6761dabd240329f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 233, + 482, + 278 + ], + "lines": [ + { + "bbox": [ + 131, + 233, + 482, + 278 + ], + "spans": [ + { + "bbox": [ + 131, + 233, + 482, + 278 + ], + "type": "text", + "content": "Fig. 2: (a) Comparison among different video LLMs based on the way of inputting visual tokens. (b) Quantitative comparison of representative video LLMs, presenting the average results of all 20 metrics and temporal-sensitive motion-related metrics from MVBench, which encompass Moving Direction, Moving Count, and Moving Attribute." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 340, + 118, + 472, + 215 + ], + "blocks": [ + { + "bbox": [ + 340, + 118, + 472, + 215 + ], + "lines": [ + { + "bbox": [ + 340, + 118, + 472, + 215 + ], + "spans": [ + { + "bbox": [ + 340, + 118, + 472, + 215 + ], + "type": "image", + "image_path": "49c933261dc499f7ed1f56945e79dbdec09abcc195a1ab4a7f3ce799365ed549.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 410, + 217, + 420, + 224 + ], + "lines": [ + { + "bbox": [ + 410, + 217, + 420, + 224 + ], + "spans": [ + { + "bbox": [ + 410, + 217, + 420, + 224 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 304, + 482, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 482, + 424 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 482, + 424 + ], + "type": "text", + "content": "Inspired by the recent success of joint spatial-temporal-text modeling in visual generation [8,36], in this work, we investigate a simple yet unexplored idea: Given the robust sequence modeling capabilities inherent to LLMs, what if we input all visual tokens into the LLM and delegate the task of modeling spatial-temporal sequences to the LLM itself? Meanwhile, there are two additional challenges: (1) The inclusion of all visual tokens significantly increases the context length within LLMs, particularly for long videos, rendering the processing of a large number of frames unaffordable. (2) LLM may struggle to handle videos of varying lengths, potentially leading to hallucinations when there is a discrepancy between the number of frames in testing and training." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 425, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 425, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 425, + 482, + 568 + ], + "type": "text", + "content": "To address these issues, we propose ST-LLM, a simple but powerful baseline of video LLM. As depicted in Fig. 2(a), with raw spatial-temporal tokens inside LLM, we harness the LLMs' robust sequence modeling capabilities for effective temporal modeling. Meanwhile, we introduce a novel dynamic video token masking strategy along with masking video modeling during training. With this approach, we reduce the length of sequences input to the LLM while significantly improving the robustness of videos of varying lengths during inference. To accommodate particularly long videos, we have devised a unique global-local input mechanism. This involves employing mean pooling of a large number of frames to generate residual input for a smaller subset of frames. Through this asymmetric design, we can process input from a large number of video frames while preserving the operations of modeling video tokens within LLMs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": "As a result, we maintain input tokens of comparable length to most video LLMs while effectively harnessing the LLM's capacity to comprehend videos. Furthermore, as ST-LLM does not introduce additional modules, it does not necessitate expensive alignment pre-training. This enables ST-LLM to directly leverage existing state-of-the-art image conversational models, resulting in significantly reduced GPU time compared to other state-of-the-art video LLMs. Extensive experiments are conducted to verify the effectiveness of ST-LLM. Qualitatively, as depicted in Fig. 1, ST-LLM demonstrates a superior ability to" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 171, + 90, + 445, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 90, + 445, + 102 + ], + "spans": [ + { + "bbox": [ + 171, + 90, + 445, + 102 + ], + "type": "text", + "content": "ST-LLM: Large Language Models Are Effective Temporal Learners" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "understand dynamics compared to other video LLMs. Quantitatively, ST-LLM achieves new state-of-the-art performance across various contemporary video benchmarks, including MVBench [25], VideoChatGPT-Bench [33] and zero-shot QA-Evaluation. Specifically, as illustrated in Fig. 2(b), the superiority of ST-LLM is particularly evident in metrics related to the temporal-sensitive motion, which underscores the expertise of our model in temporal understanding." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 208, + 243, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 208, + 243, + 220 + ], + "spans": [ + { + "bbox": [ + 132, + 208, + 243, + 220 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 234, + 482, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 234, + 482, + 484 + ], + "spans": [ + { + "bbox": [ + 130, + 234, + 482, + 484 + ], + "type": "text", + "content": "LLMs and Image LLMs. Recent years have witnessed remarkable advancements in the evolution of Large Language Models (LLMs). Inspired by InstructGPT [35] and the widely recognized commercial model ChatGPT [34], the academic community has seen a proliferation of open-source LLMs, such as LLaMA [43], Alpaca [41], Vicuna [13], and LLaMA 2 [44], which have become foundational components for a myriad of research endeavors. Meanwhile, LLMs, being large-scale transformer models, inherently possess strong sequence modeling capabilities. This inspired us to explore the feasibility of entrusting the task of spatial-temporal sequence modeling to LLMs. The success of LLMs has spurred increasing interest in the development of Multimodal Large Language Models (MLLMs). Notable breakthroughs include works such as Flamingo [2], BLIP2 [23], and PaLM-E [16], which have successfully bridged the gap between vision models and LLMs. Inspired by the concept of instruction tuning for LLMs, a series of works [12, 14, 29, 55] leveraged open-source chatbots for visual instruction tuning, thus facilitating support for image-based conversations. It is noteworthy that in the pre-LLM era, video models were commonly constructed based on pre-trained 2D image models [7, 11, 30, 45]. This may be attributed to the significant computational demands of video processing, combined with the relatively limited scale and diversity of video data compared to images. Therefore, we believe that a well-pretrained image dialogue model could similarly offer training cost and performance benefits to video LLMs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 486, + 487, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 487, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 487, + 665 + ], + "type": "text", + "content": "Video LLMs. The emergence of MLLMs quickly extended into the domain of video as well [19,20,24,25,27,31,33,53]. Early models like VideoChat [24], VideoChatGPT [33] and Valley [32] generate video instruction tuning data through GPT to enable video conversations. At the model level, these approaches typically involve the use of mean pooling to aggregate the encoding results of individual frames before feeding them into the LLMs. While efficient, it is evident that this method is inadequate for effective temporal modeling. Hence, some subsequent models have initiated exploration into adaptations of image models to video-specific requirements. For instance, BT-Adapter [31] proposed a lightweight adapter to extend the capabilities of image LLMs. Chat-UniVi [20] introduces DPC-KNN to cluster dynamic visual tokens. Moreover, VideoChat2 [25] deviated from using CLIP [38] and directly introduced a dedicated video encoder. Although these methods offer improvements over mean pooling, the inclusion of additional modules typically necessitates intensive training. For example, in the case of VideoChat2, aligning the newly introduced video encoder entails" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "type": "text", + "content": "Liu. et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 151, + 118, + 458, + 257 + ], + "blocks": [ + { + "bbox": [ + 151, + 118, + 458, + 257 + ], + "lines": [ + { + "bbox": [ + 151, + 118, + 458, + 257 + ], + "spans": [ + { + "bbox": [ + 151, + 118, + 458, + 257 + ], + "type": "image", + "image_path": "7be2628aadc0f8c2d2bc5d6887ec862b3413425dbadc3730c2945f67a03faab4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 266, + 482, + 301 + ], + "lines": [ + { + "bbox": [ + 130, + 266, + 482, + 301 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 482, + 301 + ], + "type": "text", + "content": "Fig.3: The overview of ST-LLM for generating responses on the input video and instructions. ST-LLM directly feeds the spatial-temporal sequence into the LLM and employs dynamic masking and global-local input for efficiency and robustness." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 326, + 482, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 326, + 482, + 542 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 482, + 542 + ], + "type": "text", + "content": "a three-stage training involving 30 million visual-language samples. In contrast, ST-LLM adopts a more direct approach by entrusting the task of visual sequence modeling to LLMs, which is characterized by its conciseness, reduced training requirements, and, according to our experimental findings, superior effectiveness. Joint Spatial-Temporal Modeling. Prior to the emergence of LLMs, joint spatial-temporal modeling had demonstrated effectiveness in both video-only and video-language pretraining [26,42,49]. More recently, inspired by DIT [36], the popular video generation model Sora has shown success in simultaneously processing video tokens and text using the transformer. Despite its potential, joint spatial-temporal-text modeling remains relatively uncommon in video LLMs. While VideoChat2 employed a joint-ST video encoder, it necessitated extensive pretraining. A closer relative, InstructBLIP incorporated joint spatial-temporal sequences into LLM for zero-shot evaluation on video datasets. However, this approach led to significant hallucinations due to the absence of video instruction tuning. To the best of our knowledge, ST-LLM is the first open-source video LLM to adopt joint spatial-temporal-text modeling. Moreover, ST-LLM introduces a series of design innovations to mitigate the limitations associated with incorporating all video tokens within LLM." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 564, + 233, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 564, + 233, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 564, + 233, + 578 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": "In this section, we elaborate on the structure and training methodology of ST-LLM. Firstly, we present a description of the model architecture, which incorporates all spatial-temporal tokens within the LLM, in Section 3.1. Subsequently, we delve into the masking mechanism and training objectives in Section 3.2. Finally, we introduce the global-local input approach in Section 3.3. The overarching framework of our method is illustrated in Fig. 3." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 171, + 91, + 445, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 91, + 445, + 102 + ], + "spans": [ + { + "bbox": [ + 171, + 91, + 445, + 102 + ], + "type": "text", + "content": "ST-LLM: Large Language Models Are Effective Temporal Learners" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 292, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 292, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 292, + 128 + ], + "type": "text", + "content": "3.1 Video Tokens Inside LLM" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 134, + 482, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 134, + 482, + 240 + ], + "spans": [ + { + "bbox": [ + 130, + 134, + 482, + 240 + ], + "type": "text", + "content": "As depicted in Fig. 3, visually, ST-LLM closely resembles an image chatbot, consisting of a visual encoder, a linear projection, and an LLM. Instead of opting for the commonly used CLIP-L/14 [38], we have selected BLIP-2 [23] as our visual encoder, whose Q-Former efficiently compresses redundant image tokens into fewer visual tokens. To facilitate the input of raw videos for the image encoder, we adopt the strategy of treating each frame as an individual image. Given a video " + }, + { + "bbox": [ + 130, + 134, + 482, + 240 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 130, + 134, + 482, + 240 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 134, + 482, + 240 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 134, + 482, + 240 + ], + "type": "text", + "content": " frames, we assume that the image encoder " + }, + { + "bbox": [ + 130, + 134, + 482, + 240 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{v}}" + }, + { + "bbox": [ + 130, + 134, + 482, + 240 + ], + "type": "text", + "content": " encodes each frame into " + }, + { + "bbox": [ + 130, + 134, + 482, + 240 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 134, + 482, + 240 + ], + "type": "text", + "content": " tokens. Hence, we obtain all visual tokens " + }, + { + "bbox": [ + 130, + 134, + 482, + 240 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 134, + 482, + 240 + ], + "type": "text", + "content": " to be fed into the LLM as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 188, + 248, + 481, + 264 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 248, + 481, + 264 + ], + "spans": [ + { + "bbox": [ + 188, + 248, + 481, + 264 + ], + "type": "interline_equation", + "content": "V = f _ {\\mathrm {p}} (f _ {\\mathrm {v}} (v)), \\mathrm {w h e r e} V = \\{V _ {i} \\} _ {i = 1} ^ {T}, V _ {i} = \\{v _ {(i, j)} \\} _ {j = 1} ^ {K}, \\qquad (1)", + "image_path": "1b47e4ec094efdb8f06aa323a613917d4b63926a99b77735585fe426c422b06f.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 269, + 482, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 269, + 482, + 330 + ], + "spans": [ + { + "bbox": [ + 130, + 269, + 482, + 330 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 269, + 482, + 330 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{p}}" + }, + { + "bbox": [ + 130, + 269, + 482, + 330 + ], + "type": "text", + "content": " is the visual projection layer. Then, we concatenate all visual tokens to form a joint spatial-temporal [7] visual sequence " + }, + { + "bbox": [ + 130, + 269, + 482, + 330 + ], + "type": "inline_equation", + "content": "V \\in \\mathbb{R}^{(T*K)\\times D}" + }, + { + "bbox": [ + 130, + 269, + 482, + 330 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 269, + 482, + 330 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 269, + 482, + 330 + ], + "type": "text", + "content": " is the number of embedding dimensions. After being passed through the word embedding layers, the text tokens " + }, + { + "bbox": [ + 130, + 269, + 482, + 330 + ], + "type": "inline_equation", + "content": "C = \\{c_i\\}_{i=1}^N" + }, + { + "bbox": [ + 130, + 269, + 482, + 330 + ], + "type": "text", + "content": " are directly concatenated with all spatial-temporal tokens, to organize the input " + }, + { + "bbox": [ + 130, + 269, + 482, + 330 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 269, + 482, + 330 + ], + "type": "text", + "content": " to the LLM as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 167, + 337, + 481, + 350 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 337, + 481, + 350 + ], + "spans": [ + { + "bbox": [ + 167, + 337, + 481, + 350 + ], + "type": "interline_equation", + "content": "I = \\left[ V: C \\right] = \\left\\{v _ {(1, 1)}, v _ {(1, 2)}, \\dots , v _ {(1, K)}, v _ {(2, 1)}, \\dots , v _ {(T, K)}, c _ {1}, c _ {2}, \\dots , c _ {N} \\right\\}. \\tag {2}", + "image_path": "77a12747dda4015b4bb8a238b9f9d88567b784fc508e61b43834b463a440f4ff.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 356, + 482, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 356, + 482, + 477 + ], + "spans": [ + { + "bbox": [ + 130, + 356, + 482, + 477 + ], + "type": "text", + "content": "Indeed, several alternative strategies exist for organizing the input tokens before feeding them into the LLM. These include: (1) Adding separator tokens between frame-wise tokens or visual-text tokens, to demarcate the boundaries between modalities or frames. (2) Adding spatial-temporal position embeddings specifically for visual tokens, to facilitate better spatial-temporal relationship understandings. However, our experiments indicate that simplicity often yields the best. Adding numerous separator tokens introduces additional overhead without necessarily improving performance. Moreover, LLMs come with Rotary Position Embeddings [39] that are already effective in distinguishing the positions of all spatial-temporal tokens and text tokens." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 491, + 326, + 505 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 326, + 505 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 326, + 505 + ], + "type": "text", + "content": "3.2 Training with Dynamic Masking" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 510, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 629 + ], + "type": "text", + "content": "With video tokens modeling inside LLM, we can leverage the potent sequence modeling capabilities inherent in LLMs to understand the information encapsulated within spatial-temporal sequences. However, integrating all video tokens into LLM significantly increases the context length, imposing a considerable computational burden and rendering it unfeasible to input additional frames for lengthy videos. Meanwhile, experimental findings suggest that during testing, utilizing uncompressed tokens is more sensitive to variations in the number of frames than mean pooling. Consequently, when a substantial dissonance exists between the number of frames in training and testing, a noticeable degradation in performance ensues." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 630, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 667 + ], + "type": "text", + "content": "To tackle the aforementioned issues, we first propose masking the visual tokens during training. Masking modeling has achieved significant success in natural language processing [15] and video-language pretraining [26,31,42]. Although" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "type": "text", + "content": "Liu. et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "masking is rarely employed in autoregressive LLMs, given our aim to leverage LLMs for encoding spatial-temporal sequences, we can still utilize masking modeling. Specifically, we keep the text tokens in " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " unchanged, while applying a mask to the video tokens. This mask randomly masks " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " of all video tokens, irrespective of their position across different frames. Different from previous works [26,31], we adopt a dynamic masking strategy, where the masking rate is randomly sampled from a normal distribution as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 243, + 209, + 481, + 222 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 243, + 209, + 481, + 222 + ], + "spans": [ + { + "bbox": [ + 243, + 209, + 481, + 222 + ], + "type": "interline_equation", + "content": "\\rho \\sim \\mathbb {N} (0. 5, \\sigma), \\quad 0. 3 \\leq \\rho \\leq 0. 7, \\tag {3}", + "image_path": "f3c79588a6d0d97c3d3fa50cd5850dd7de21d98ef61f4d82c4f681168998637d.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 230, + 482, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 230, + 482, + 277 + ], + "spans": [ + { + "bbox": [ + 130, + 230, + 482, + 277 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 230, + 482, + 277 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 230, + 482, + 277 + ], + "type": "text", + "content": " represents the variance of the normal distribution. This strategy ensures that the length of the spatial-temporal sequence varies continuously while maintaining a consistently high average masking rate of " + }, + { + "bbox": [ + 130, + 230, + 482, + 277 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 230, + 482, + 277 + ], + "type": "text", + "content": ". As a result, it minimizes training costs and notably enhances robustness during inference." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 278, + 485, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 278, + 485, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 278, + 485, + 373 + ], + "type": "text", + "content": "Furthermore, building upon dynamic masking, we have formulated the Masked Video Modeling (MVM) objective to encourage the LLM to grasp spatial-temporal dependencies. Unlike the expensive masked token recovery [6,37], we have opted for unmasked token reconstruction. Specifically, besides the masked sequence " + }, + { + "bbox": [ + 130, + 278, + 485, + 373 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 278, + 485, + 373 + ], + "type": "text", + "content": ", we conduct an extra non-grad forward pass for the unmasked sequence " + }, + { + "bbox": [ + 130, + 278, + 485, + 373 + ], + "type": "inline_equation", + "content": "\\hat{I} = [\\hat{V}:C]" + }, + { + "bbox": [ + 130, + 278, + 485, + 373 + ], + "type": "text", + "content": ", serving as the reference output. Then, we select unmasked tokens from " + }, + { + "bbox": [ + 130, + 278, + 485, + 373 + ], + "type": "inline_equation", + "content": "f_{llm}(I)" + }, + { + "bbox": [ + 130, + 278, + 485, + 373 + ], + "type": "text", + "content": " and the corresponding tokens from " + }, + { + "bbox": [ + 130, + 278, + 485, + 373 + ], + "type": "inline_equation", + "content": "f_{llm}(\\hat{I})" + }, + { + "bbox": [ + 130, + 278, + 485, + 373 + ], + "type": "text", + "content": " based on their positions, computing the Mean Squared Error (MSE) between the selected pairs as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 190, + 384, + 481, + 415 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 384, + 481, + 415 + ], + "spans": [ + { + "bbox": [ + 190, + 384, + 481, + 415 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {m v m} = \\frac {1}{(1 - \\rho) K T} \\sum_ {i = 1} ^ {T} \\sum_ {j = 1} ^ {K} \\left(v _ {i, j} ^ {- 1} - \\hat {v} _ {i, j} ^ {- 1}\\right) ^ {2}, (i, j) \\notin M, \\tag {4}", + "image_path": "2b49a0b9683cec43c952fb393a0fe44cd36d3aa2d91f5df6df6e12113398b268.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 425, + 482, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 425, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 131, + 425, + 482, + 498 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 425, + 482, + 498 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 131, + 425, + 482, + 498 + ], + "type": "text", + "content": " denotes the set of masked tokens' indices, " + }, + { + "bbox": [ + 131, + 425, + 482, + 498 + ], + "type": "inline_equation", + "content": "v_{i,j}^{-1}" + }, + { + "bbox": [ + 131, + 425, + 482, + 498 + ], + "type": "text", + "content": " is from " + }, + { + "bbox": [ + 131, + 425, + 482, + 498 + ], + "type": "inline_equation", + "content": "f_{llm}(I)" + }, + { + "bbox": [ + 131, + 425, + 482, + 498 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 425, + 482, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{mvm}" + }, + { + "bbox": [ + 131, + 425, + 482, + 498 + ], + "type": "text", + "content": " represents our MVM objective. Finally, our overall loss " + }, + { + "bbox": [ + 131, + 425, + 482, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 131, + 425, + 482, + 498 + ], + "type": "text", + "content": " is composed of " + }, + { + "bbox": [ + 131, + 425, + 482, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{mvm}" + }, + { + "bbox": [ + 131, + 425, + 482, + 498 + ], + "type": "text", + "content": " and the LLM decoder loss " + }, + { + "bbox": [ + 131, + 425, + 482, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{llm}" + }, + { + "bbox": [ + 131, + 425, + 482, + 498 + ], + "type": "text", + "content": ", yielding " + }, + { + "bbox": [ + 131, + 425, + 482, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\mathcal{L}_{mvm} + \\mathcal{L}_{llm}" + }, + { + "bbox": [ + 131, + 425, + 482, + 498 + ], + "type": "text", + "content": ". By integrating these two loss components, we can encourage the LLM to effectively respond to questions derived from the video content while improving its ability to model temporal and spatial dependencies." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 514, + 257, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 257, + 528 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 257, + 528 + ], + "type": "text", + "content": "3.3 Global-Local Input" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 533, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 628 + ], + "type": "text", + "content": "Although dynamic masking partially mitigates the challenge of processing lengthy input sequences, it fails to fully resolve the issue for extremely long videos that entail inputting numerous frames, rendering the context length still impractical to handle. Therefore, we have devised an additional module to tackle the issue of excessively long videos. To elaborate, given a lengthy video with a large " + }, + { + "bbox": [ + 130, + 533, + 482, + 628 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 533, + 482, + 628 + ], + "type": "text", + "content": ", we still commence by encoding each frame individually to derive " + }, + { + "bbox": [ + 130, + 533, + 482, + 628 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 533, + 482, + 628 + ], + "type": "text", + "content": ". Subsequently, we proceed to derive the global video representation " + }, + { + "bbox": [ + 130, + 533, + 482, + 628 + ], + "type": "inline_equation", + "content": "V_{0}" + }, + { + "bbox": [ + 130, + 533, + 482, + 628 + ], + "type": "text", + "content": " through average pooling on the frame-wise tokens:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 223, + 636, + 481, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 636, + 481, + 668 + ], + "spans": [ + { + "bbox": [ + 223, + 636, + 481, + 668 + ], + "type": "interline_equation", + "content": "v _ {(0, j)} = \\frac {1}{T} \\sum_ {i = 1} ^ {T} v _ {(i, j)}, \\quad V _ {0} = \\left\\{v _ {(0, j)} \\right\\} _ {j = 1} ^ {K}. \\tag {5}", + "image_path": "ecfdee856b4fafd73964445f2ed49b561e8fc0bd8c94539cf0eb9a20784a60c5.jpg" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 170, + 91, + 445, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 91, + 445, + 103 + ], + "spans": [ + { + "bbox": [ + 170, + 91, + 445, + 103 + ], + "type": "text", + "content": "ST-LLM: Large Language Models Are Effective Temporal Learners" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 166 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 166 + ], + "type": "text", + "content": "Next, we average-sample " + }, + { + "bbox": [ + 130, + 116, + 482, + 166 + ], + "type": "inline_equation", + "content": "\\overline{T}" + }, + { + "bbox": [ + 130, + 116, + 482, + 166 + ], + "type": "text", + "content": " frames from the T total frames. All tokens from these " + }, + { + "bbox": [ + 130, + 116, + 482, + 166 + ], + "type": "inline_equation", + "content": "\\overline{T}" + }, + { + "bbox": [ + 130, + 116, + 482, + 166 + ], + "type": "text", + "content": " frames are concatenated to produce a joint spatial-temporal sequence " + }, + { + "bbox": [ + 130, + 116, + 482, + 166 + ], + "type": "inline_equation", + "content": "\\overline{V} \\in \\mathbb{R}^{(\\overline{T}*K) \\times D}" + }, + { + "bbox": [ + 130, + 116, + 482, + 166 + ], + "type": "text", + "content": ", which serves as the local video representation. Finally, the global-local input for the LLM, denoted as " + }, + { + "bbox": [ + 130, + 116, + 482, + 166 + ], + "type": "inline_equation", + "content": "\\overline{I}" + }, + { + "bbox": [ + 130, + 116, + 482, + 166 + ], + "type": "text", + "content": ", is constructed as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 260, + 175, + 481, + 189 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 175, + 481, + 189 + ], + "spans": [ + { + "bbox": [ + 260, + 175, + 481, + 189 + ], + "type": "interline_equation", + "content": "\\bar {I} = \\left[ \\bar {V} + f _ {m} \\left(V _ {0}\\right): C \\right], \\tag {6}", + "image_path": "f7feef5c9f4df42db98ca994af4731afa1b88550a06524693ea282b595d60156.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 198, + 482, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 198, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 198, + 482, + 270 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 198, + 482, + 270 + ], + "type": "inline_equation", + "content": "f_{m}" + }, + { + "bbox": [ + 130, + 198, + 482, + 270 + ], + "type": "text", + "content": " is a simple MLP projector with upsampling projection initialized with zeros. With this global-local input design, the low fps spatial-temporal sequences within the LLM can gradually incorporate information from the high fps branch. This approach allows the model to benefit from the LLM's ability to model temporal sequences within a limited context while also considering the global information of long videos." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 289, + 230, + 303 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 289, + 230, + 303 + ], + "spans": [ + { + "bbox": [ + 131, + 289, + 230, + 303 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 315, + 483, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 315, + 483, + 364 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 483, + 364 + ], + "type": "text", + "content": "In this section, we have performed comprehensive experimental evaluations of ST-LLM, covering crucial settings, comparisons, and ablation studies. For a more detailed account of experimental settings, ablation studies, visualizations, and limitations analysis, please consult the appendix." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 381, + 253, + 394 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 381, + 253, + 394 + ], + "spans": [ + { + "bbox": [ + 131, + 381, + 253, + 394 + ], + "type": "text", + "content": "4.1 Experiment Setup" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 402, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 510 + ], + "type": "text", + "content": "Benchmarks. To evaluate the video understanding capabilities of ST-LLM, we primarily conduct assessments on three benchmarks: MVBench [25], VideoChat-GPT Bench [33], and zero-shot video QA benchmark. MVBench comprises 20 challenging video tasks, each consisting of 200 samples in the form of multiple-choice questions. These tasks provide a comprehensive and objective assessment of a model's ability to understand videos. VideoChatGPT-Bench gathers videos from ActivityNet [10] and uses GPT to evaluate the quality of video conversations across five dimensions. Additionally, the zero-shot video QA benchmark entails GPT-based assessments of various open-source video QA datasets." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 510, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 483, + 666 + ], + "type": "text", + "content": "Implementation Details. In our study, we adopt state-of-the-art image dialogue models as the baseline and proceed directly with video instruction tuning. Unless otherwise specified, our model is initialized with InstructBLIP [14] in all experiments, whose LLM is Vicuna-v1.1 [13] with 7B parameters. Additionally, we employ Minigpt4-v1 [55] as a basic model to assess the effectiveness across various image models in Table 6. Following previous research [20, 27, 28], we adopt full-finetuning on the LLM to ensure a fair comparison. The results of LoRA are also included as ablation results in Table 6. " + }, + { + "bbox": [ + 130, + 510, + 483, + 666 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 510, + 483, + 666 + ], + "type": "text", + "content": " is set as 0.1, ensuring that over " + }, + { + "bbox": [ + 130, + 510, + 483, + 666 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 130, + 510, + 483, + 666 + ], + "type": "text", + "content": " of the mask rate falls within the range of 0.3 to 0.7. During training, we sample 16 frames per video. Considering dynamic masking, an average of 256 tokens per video are input into the LLM, which is the same as LLaVA [29]. For all benchmarks during inference, frames are sampled at a frame rate of 1, and the number of local frames is kept between 4 and 16. This strategy closely" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "type": "text", + "content": "Liu. et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 212, + 478, + 463 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 205 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 205 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 205 + ], + "type": "text", + "content": "Table 1: Comparisons on MVBench. Except for BLIP2 and Otter, all models are built upon LLaMA-1-7B for fair comparisons. \"Avg\" denotes the average of all 20 metrics. The leaderboards for each task can be found in the Appendix. The compared methods include: (1)Rd: random guesses; (2) mOI: mPLUG-Owl-I 51; (3) LMA: LLaMA-Adapter 54 (4)B2: BLIP-2 23; (5)OI: Otter-I 22; (6)MG: MiniGPT-4 55; (7)IB: InstructBLIP 14; (8)LV: LLaVA 29; (9)OV: Otter-V 22; (10)mOV: mPLUG-Owl-V 51; (11)VCG: VideoChatGPT 33; (12)VLM: VideoLLaMA 53; (13)VC: VideoChat 24; (14)VC2: VideoChat2 25." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 212, + 478, + 463 + ], + "lines": [ + { + "bbox": [ + 135, + 212, + 478, + 463 + ], + "spans": [ + { + "bbox": [ + 135, + 212, + 478, + 463 + ], + "type": "table", + "html": "
MetricRdmOILMAB2OIMGIBLVOVmOVVCGVLMVCVC2Ours
AS25.025.023.024.534.516.020.028.023.022.023.527.533.566.066.0
AP25.020.028.029.032.018.016.539.523.028.026.025.526.547.553.5
AA33.344.551.033.539.526.046.063.027.534.062.051.050.083.584.0
FA25.027.030.017.030.521.524.530.527.029.022.529.033.549.544.0
UA25.023.533.042.038.516.046.039.029.529.026.539.040.560.058.5
OE33.336.053.551.548.529.551.053.053.040.554.048.053.058.080.5
OI25.024.032.526.044.025.526.041.028.027.028.040.540.571.573.5
OS33.334.033.531.029.513.037.541.533.031.540.038.030.042.538.5
MD25.023.025.525.519.011.522.023.024.527.023.022.525.523.042.5
AL25.024.021.526.025.512.023.020.523.523.020.022.527.023.031.0
ST25.034.530.532.555.09.546.545.027.529.031.043.048.588.586.5
AC33.334.529.025.520.032.542.534.026.031.530.534.035.039.036.5
MC25.022.022.530.032.515.526.520.528.527.025.522.520.542.056.5
MA33.331.541.540.028.58.040.538.518.040.039.532.542.558.578.5
SC33.340.039.542.039.034.032.047.038.544.048.545.546.044.043.0
FP25.024.025.027.028.026.025.525.022.024.029.032.526.549.044.5
CO33.337.031.530.027.029.530.536.022.031.033.040.041.036.546.5
EN25.025.522.526.032.019.030.527.023.526.029.530.023.535.034.5
ER20.021.028.037.036.59.930.526.519.520.526.021.023.540.541.5
CI30.937.032.031.036.53.038.042.019.029.535.537.036.065.558.5
Avg27.329.431.731.433.518.832.536.026.829.732.734.135.551.154.9
", + "image_path": "d084cb1979183c3dded291089c47e447b719cc4161f89b7bd248809ab8edc491.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 486, + 482, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 557 + ], + "type": "text", + "content": "aligns with practical scenarios and involves fewer frames than most video LLMs. For VideoChatGPT-Bench and the video QA benchmark, which may involve videos spanning several minutes, we utilize the global-local input approach. In this configuration, the global frame number is set to 64. All videos are resized into " + }, + { + "bbox": [ + 130, + 486, + 482, + 557 + ], + "type": "inline_equation", + "content": "224^{*}224" + }, + { + "bbox": [ + 130, + 486, + 482, + 557 + ], + "type": "text", + "content": ". We train ST-LLM using a learning rate of 2e-5 and a batch size of 128 for 2 epochs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 558, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 667 + ], + "type": "text", + "content": "Training Datasets. Citing findings from Jin et al. [20], the sequential process of first conducting image instruction tuning, followed by video instruction tuning and image-video joint pretraining, had a negligible impact on video dialogue performance. Hence, in light of InstructBLIP's prior training on extensive image data, we exclusively focus on video instruction tuning. Following [25], we leverage video instruction data sourced from a variety of datasets, including VideoChatGPT-100k [33], VideoChat-11k [24], Webvid [5], NExT-QA [47], CLEVRER [52], Kinetics-710 [21], and Something-Something-2 [17]. All samples are uniformly formatted in accordance with [25]. More details about the" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 171, + 91, + 445, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 91, + 445, + 102 + ], + "spans": [ + { + "bbox": [ + 171, + 91, + 445, + 102 + ], + "type": "text", + "content": "ST-LLM: Large Language Models Are Effective Temporal Learners" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 143, + 136, + 468, + 277 + ], + "blocks": [ + { + "bbox": [ + 147, + 114, + 465, + 126 + ], + "lines": [ + { + "bbox": [ + 147, + 114, + 465, + 126 + ], + "spans": [ + { + "bbox": [ + 147, + 114, + 465, + 126 + ], + "type": "text", + "content": "Table 2: Comparisons on video-based generative performance benchmarking." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 143, + 136, + 468, + 277 + ], + "lines": [ + { + "bbox": [ + 143, + 136, + 468, + 277 + ], + "spans": [ + { + "bbox": [ + 143, + 136, + 468, + 277 + ], + "type": "table", + "html": "
MethodCorrectDetailContextTemporalConsistMean Score
VideoLLaMA 531.962.182.161.821.791.98
LLaMA-Adapter 542.032.322.301.982.152.16
VideoChat 242.232.502.531.942.242.29
VideoChatGPT 332.402.522.621.982.372.38
BT-Adapter 312.682.693.272.342.462.69
VTimeLLM 192.783.103.402.492.472.85
Chat-UniVi 202.892.913.462.892.812.99
LLaMA-VID-7B 272.963.003.532.462.512.89
LLaMA-VID-13B 273.073.053.602.582.632.99
VideoChat2 253.022.883.512.662.812.98
ST-LLM (Ours)3.233.053.742.932.813.15
", + "image_path": "429a145e438720229240fc03027e0438e3fa4942fc5f6f5fc0976df9e719cd45.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 297, + 482, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 482, + 321 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 482, + 321 + ], + "type": "text", + "content": "dataset and the organization of instructions can be found in the appendix. An experiment on joint image-video training is also presented in the appendix." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 336, + 261, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 336, + 261, + 348 + ], + "spans": [ + { + "bbox": [ + 131, + 336, + 261, + 348 + ], + "type": "text", + "content": "4.2 Quantitative Result" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "text", + "content": "MVBench Performance. The evaluation results on MVBench are presented in Table 1. The evaluation metrics included Action Sequence (AS), Action Prediction (AP), Action Antonym (AA), Fine-grained Action (FA), Unexpected Action (UA), Object Existence (OE), Object Interaction (OI), Object Shuffle (OS), Moving Direction (MD), Action Localization (AL), Scene Transition (ST), Action Count (AC), Moving Count (MC), Moving Attribute (MA), State Change (SC), Fine-grained Pose (FP), Character Order (CO), Egocentric Navigation (EN), Episodic Reasoning (ER), Counterfactual Inference (CI), and the average of all 20 metrics (Avg). It is evident that on such a challenging benchmark, the performance of most multimodal LLMs falls significantly short of satisfactory levels, even barely surpassing random performance by less than " + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "text", + "content": ". Our only formidable competitor is VideoChat2 [25]. Through direct substitution of CLIP with a dedicated video encoder, coupled with a three-stage training process and comprehensive instruction tuning for the visual encoder, Q-Former, and LLM, VideoChat2 has yielded results that significantly surpass those of previous models. In contrast, we solely carried out single-stage video-only instruction tuning while maintaining the weights of CLIP and Q-Former frozen. Even under these conditions, ST-LLM still attained the top performance, with an average score surpassing VideoChat2 by a notable margin of " + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "inline_equation", + "content": "3.8\\%" + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "text", + "content": ". Specifically, ST-LLM showcased leading results across all 11 tasks, exhibiting a particularly significant advantage in motion-related tasks where other models struggled. In three motion-related tasks, ST-LLM achieved an average score of " + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "inline_equation", + "content": "59.2\\%" + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "text", + "content": ", far outperforming VideoChat2's " + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "inline_equation", + "content": "36.3\\%" + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "text", + "content": ". However, ST-LLM encounters challenges in fine-grained tasks, such as Fine-grained Action (FA) and Fine-grained Pose (FP). This could be attributed to CLIP's limitations as an image encoder, which may struggle to capture low-level spatial-temporal features required for these tasks." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "text", + "content": "Liu. et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 154, + 147, + 459, + 281 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 139 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 139 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 139 + ], + "type": "text", + "content": "Table 3: Comparisons on zero-shot question-answering, including MSVD-QA [46], MSRVTT-QA [48], and ActivityNet-QA [10]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 154, + 147, + 459, + 281 + ], + "lines": [ + { + "bbox": [ + 154, + 147, + 459, + 281 + ], + "spans": [ + { + "bbox": [ + 154, + 147, + 459, + 281 + ], + "type": "table", + "html": "
MethodLLMMSVD-QAMSRVTT-QAActivityNet-QA
AccScoreAccScoreAccScore
FrozenBiLM 50DeBERTa-V232.2-16.8-24.7-
VideoLLaMA 53Vicuna-7B51.62.529.61.812.41.1
LLaMA-Adapter 54LLaMA-7B54.93.143.82.734.22.7
VideoChat 24Vicuna-7B56.32.845.02.526.52.2
VideoChatGPT 33Vicuna-7B64.93.349.32.835.22.7
BT-Adapter 31Vicuna-7B67.53.757.03.245.73.2
Chat-UniVi 20Vicuna-7B65.03.654.63.145.83.2
LLaMA-VID 27Vicuna-7B69.73.757.73.247.43.3
LLaMA-VID 27Vicuna-13B70.03.758.93.347.53.3
VideoChat2 25Vicuna-7B70.03.954.13.349.13.3
ST-LLM (Ours)Vicuna-7B74.63.963.23.450.93.3
", + "image_path": "f0663376c0d98cf89a848533b2a89b0d4e529e0702909502b4f8c8328bf44d52.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 304, + 482, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 482, + 447 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 482, + 447 + ], + "type": "text", + "content": "VideoChatGPT-Bench Performance. VideoChatGPT-Bench, also referred to as the Video-Based Generative Performance Benchmark, primarily evaluates the capability of video conversation. Utilizing GPT-3.5 [34], it assigns a score to the generated content across five aspects: Correctness of Information (Correct), Detail Orientation (Detail), Contextual Understanding (Context), Temporal Understanding (Temporal), and Consistency (Consist). As depicted in Table 2, the recently proposed state-of-the-art models [20,25,27] exhibit similar performance levels on this benchmark. However, due to the inherent instability of GPT evaluation, there is considerable variation in the performance of each model across different tasks. Despite this variability, ST-LLM consistently performs admirably across all tasks, while demonstrating significant advantages in terms of the mean score compared to previous models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 448, + 490, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 448, + 490, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 448, + 490, + 556 + ], + "type": "text", + "content": "Zero-shot Video-question Answering Performance. In Table 3 we present the zero-shot video question-answering performance on several commonly used video-text datasets, including MSVD-QA [46], MSRVTT-QA [48], and ActivityNet-QA [10]. Following the methodology outlined in [33], we utilize GPT-3.5 to evaluate the accuracy and score of the generated results. As is depicted, STLLM demonstrates significant advantages on MSVD and MSRVTT, surpassing the previous SOTAs by " + }, + { + "bbox": [ + 130, + 448, + 490, + 556 + ], + "type": "inline_equation", + "content": "4.6\\%" + }, + { + "bbox": [ + 130, + 448, + 490, + 556 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 448, + 490, + 556 + ], + "type": "inline_equation", + "content": "4.3\\%" + }, + { + "bbox": [ + 130, + 448, + 490, + 556 + ], + "type": "text", + "content": " respectively. However, STLLM exhibits a slightly smaller advantage compared to VideoChat2 on the ActivityNet. This could be attributed to the requirement for a potent video encoder for this dataset." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 574, + 278, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 574, + 278, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 574, + 278, + 586 + ], + "type": "text", + "content": "4.3 Ablations and Analysis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "content": "Ways of Inputting video tokens. The core of this work lies in modeling joint spatial-temporal video tokens using LLM. Table 4 presents the results of different methods of inputting video tokens. As depicted, without video tuning, the approach of joint spatial-temporal input is considerably less effective compared to directly averaging the input over the temporal dimension. This discrepancy might stem from inconsistencies between training and testing. However, after" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 170, + 90, + 445, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 90, + 445, + 102 + ], + "spans": [ + { + "bbox": [ + 170, + 90, + 445, + 102 + ], + "type": "text", + "content": "ST-LLM: Large Language Models Are Effective Temporal Learners" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 174, + 308, + 266 + ], + "blocks": [ + { + "bbox": [ + 132, + 129, + 309, + 173 + ], + "lines": [ + { + "bbox": [ + 132, + 129, + 309, + 173 + ], + "spans": [ + { + "bbox": [ + 132, + 129, + 309, + 173 + ], + "type": "text", + "content": "Table 4: The ablation study on the methods of inputting video tokens to LLMs. The baseline configurations include no instruction tuning and mean pooling input." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 174, + 308, + 266 + ], + "lines": [ + { + "bbox": [ + 135, + 174, + 308, + 266 + ], + "spans": [ + { + "bbox": [ + 135, + 174, + 308, + 266 + ], + "type": "table", + "html": "
MethodAvg on MV bench
Instructclip Baseline
Mean Pooling42.0
S-T Tokens in LLM35.5
Video Instruction Tuning
Mean Pooling47.8
S-T Tokens in LLM53.8
+Masking & MVM Loss54.9
", + "image_path": "7524b810a2dbbf3141325600a173f20aaf66a8b12f2fd243717feea30eb674f3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 317, + 114, + 478, + 235 + ], + "blocks": [ + { + "bbox": [ + 317, + 114, + 478, + 235 + ], + "lines": [ + { + "bbox": [ + 317, + 114, + 478, + 235 + ], + "spans": [ + { + "bbox": [ + 317, + 114, + 478, + 235 + ], + "type": "image", + "image_path": "d51c3b50eefedfff5883c76f9d933ac6618712042adc1ba0dff8d821206d2ff2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 243, + 482, + 266 + ], + "lines": [ + { + "bbox": [ + 315, + 243, + 482, + 266 + ], + "spans": [ + { + "bbox": [ + 315, + 243, + 482, + 266 + ], + "type": "text", + "content": "Fig. 4: The ablation study on the effect of dynamic masking and MVM loss." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 135, + 311, + 479, + 378 + ], + "blocks": [ + { + "bbox": [ + 132, + 277, + 482, + 302 + ], + "lines": [ + { + "bbox": [ + 132, + 277, + 482, + 302 + ], + "spans": [ + { + "bbox": [ + 132, + 277, + 482, + 302 + ], + "type": "text", + "content": "Table 5: The ablation study on global-local input module. The \"Local\" and \"Global\" mean the joint temporal-temporal sequence and average pooling sequence respectively." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 311, + 479, + 378 + ], + "lines": [ + { + "bbox": [ + 135, + 311, + 479, + 378 + ], + "spans": [ + { + "bbox": [ + 135, + 311, + 479, + 378 + ], + "type": "table", + "html": "
MethodAvg on MVBenchVideo-ChatGPT Bench
CorrectDetailContextTemporalConsist
Global Only48.33.012.923.582.612.61
Local Only54.93.082.993.632.752.69
Local+Global (simply add)50.13.072.883.582.592.50
Local+Global (adapter)54.73.233.053.742.932.81
", + "image_path": "bbf33dae587544bdf54c022b3c9f3d0b5c5930ddfd6a5f667ff3399ba79ffe1a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 402, + 482, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 473 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 473 + ], + "type": "text", + "content": "training, the effectiveness of the joint spatial-temporal input significantly surpasses that of the mean pooling input, indicating that LLMs can effectively model spatial-temporal sequences, and directly inputting all tokens is a superior approach compared to compression before input. Furthermore, incorporating dynamic masking and the MVM loss based on the joint spatial-temporal sequence has further improved the performance of our model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 474, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 616 + ], + "type": "text", + "content": "Dynamic Masking Strategy and MVM Loss. In Fig. 4, we conduct a more detailed ablation study on the effect of dynamic masking and MVM loss. It is evident that when the training frame count remains fixed, the performance of inputting all video tokens is not very robust. Particularly, when there is a discrepancy between the training and testing frame, the performance noticeably decreases. Meanwhile, when we randomly mask video tokens at a static " + }, + { + "bbox": [ + 130, + 474, + 482, + 616 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 474, + 482, + 616 + ], + "type": "text", + "content": " rate, this issue persists, with effectiveness akin to training with a fixed 8-frame setup. However, employing dynamic masking shows some improvement in overall performance, and notably, the robustness to varying testing frame counts is significantly enhanced. Furthermore, the inclusion of the MVM loss built upon the masking further enhances performance, with greater improvements observed as the sequence length increases." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Global-Local Input Module. The ablation results on the global-local module are presented in Table 5 where we compare the performance of global features extracted from more frames and local features from fewer frames, as well as their combination. As shown, on MVBench, the performance of global features" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "type": "text", + "content": "Liu. et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 147, + 479, + 201 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 6: The ablation study on the basic image models and the fine-tuning methodologies of LLM. We report the Avg on MVBench." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 147, + 479, + 201 + ], + "lines": [ + { + "bbox": [ + 135, + 147, + 479, + 201 + ], + "spans": [ + { + "bbox": [ + 135, + 147, + 479, + 201 + ], + "type": "table", + "html": "
Baseline ModelLLMLLM ParametersMeanpooling BaselineST-LLM Full Model
MiniGPT-4 55Vicuna-7B v0LoRA40.945.6
Full Fine-tuning46.450.5
InstructBLIP 14Vicuna-7B v1.1LoRA45.251.9
Full Fine-tuning47.854.9
", + "image_path": "2009037448447e6fd39ffb4cacebb25750228b3d9192dfaaf19507444b6f6ac2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 135, + 247, + 319, + 305 + ], + "blocks": [ + { + "bbox": [ + 132, + 224, + 318, + 245 + ], + "lines": [ + { + "bbox": [ + 132, + 224, + 318, + 245 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 318, + 245 + ], + "type": "text", + "content": "Table 7: The analysis on the designs of inputting S-T video tokens into LLM." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 247, + 319, + 305 + ], + "lines": [ + { + "bbox": [ + 135, + 247, + 319, + 305 + ], + "spans": [ + { + "bbox": [ + 135, + 247, + 319, + 305 + ], + "type": "table", + "html": "
MethodAvg on MVBench
S-T Tokens in LLM (baseline)54.9
+frame-frame separators54.0
+video-text separators54.8
+S-T position embeddings52.2
", + "image_path": "373904aeac22fcd8ea70df2e6c853eff7f98a02544de22620029c18a2492b4f5.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 325, + 247, + 481, + 305 + ], + "blocks": [ + { + "bbox": [ + 322, + 224, + 481, + 246 + ], + "lines": [ + { + "bbox": [ + 322, + 224, + 481, + 246 + ], + "spans": [ + { + "bbox": [ + 322, + 224, + 481, + 246 + ], + "type": "text", + "content": "Table 8: The analysis on the designs of dynamic masking." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 325, + 247, + 481, + 305 + ], + "lines": [ + { + "bbox": [ + 325, + 247, + 481, + 305 + ], + "spans": [ + { + "bbox": [ + 325, + 247, + 481, + 305 + ], + "type": "table", + "html": "
MethodAvg on MVBatch
w/o masking (baseline)53.8
ρ ~ U(0.3, 0.7)54.2
ρ ~ N(0.5, 0.2)54.3
ρ ~ N(0.5, 0.1)54.9
", + "image_path": "6fe0f45bc7f2e23d18f8316b75ffab4ff1fa7bf246f12444ec11b25ac1b5095f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 328, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 328, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 328, + 482, + 449 + ], + "type": "text", + "content": "is significantly inferior to the joint S-T features, so combining global and local features did not lead to improvement. This may be due to the fact that MVBench mainly consists of short videos lasting a few to a dozen seconds, where a large number of frames might not be necessary. However, on VideoChatGPT Bench, where videos are parsed at the minute level, the discrepancy between global and local features has notably diminished, with their amalgamation resulting in a discernible enhancement. This underscores the efficacy of the Global-Local Input module in processing extended video content. Moreover, the direct incorporation of global and local features resulted in a performance decline, presumably due to its interference with the representation learning of the LLM." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 449, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 533 + ], + "type": "text", + "content": "Basic Image Models & LLMs. In Table 6, we conduct ablations on both the basic image chatbots and the LLM fine-tuning strategy. As depicted in the table, regardless of whether it's LoRA [18] tuning or full fine-tuning, or different basic image models, ST-LLM demonstrates notable enhancements compared to the mean pooling baseline. This suggests that our approach can be adeptly applied across diverse scenarios, such as employing more resource-efficient LoRA tuning on larger models, or transferring to more potent image-dialogue models." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Designs of ST-LLM. In Table 7 and Table 8, we investigated several alternative design possibilities for ST-LLM. Primarily, we explored various separator designs. It's noteworthy that almost all cross-modal LLMs incorporate separator designs, with some models even integrating separators between multiple images or frames [2,4]. However, we discovered that when employing LLMs to model spatial-temporal sequences, these separators are superfluous. Subsequently, we address positional encodings. Numerous approaches integrate temporal positional encodings before inputting video tokens into LLMs [32,33,53]. However, in our methodology, we observed that incorporating absolute positional encodings to all visual tokens significantly impairs performance. Lastly, we consider the design of dynamic masking. Our findings suggest that all distributions" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 171, + 90, + 445, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 90, + 445, + 102 + ], + "spans": [ + { + "bbox": [ + 171, + 90, + 445, + 102 + ], + "type": "text", + "content": "ST-LLM: Large Language Models Are Effective Temporal Learners" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 116, + 477, + 343 + ], + "blocks": [ + { + "bbox": [ + 137, + 116, + 477, + 343 + ], + "lines": [ + { + "bbox": [ + 137, + 116, + 477, + 343 + ], + "spans": [ + { + "bbox": [ + 137, + 116, + 477, + 343 + ], + "type": "image", + "image_path": "42f1b8056b5f8ef0be156a6cf389d6863825d969515d934706bc96954dbd8798.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 204, + 355, + 409, + 367 + ], + "lines": [ + { + "bbox": [ + 204, + 355, + 409, + 367 + ], + "spans": [ + { + "bbox": [ + 204, + 355, + 409, + 367 + ], + "type": "text", + "content": "Fig. 5: Qualitative results from real-world videos." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 388, + 482, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 482, + 413 + ], + "type": "text", + "content": "of masking rates can yield some improvement, yet distributions with smaller standard deviations exhibit superior performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 428, + 258, + 440 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 428, + 258, + 440 + ], + "spans": [ + { + "bbox": [ + 131, + 428, + 258, + 440 + ], + "type": "text", + "content": "4.4 Qualitative Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 446, + 482, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 446, + 482, + 507 + ], + "spans": [ + { + "bbox": [ + 130, + 446, + 482, + 507 + ], + "type": "text", + "content": "In Fig. 5, we present a qualitative comparison. As illustrated, ST-LLM excels in adhering to instructions and delivering precise responses. More importantly, ST-LLM showcases superior sensitivity to temporal sequences and actions, thereby validating the effectiveness of our modeling approach. For further quantitative results and failure cases, please refer to the appendix." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 522, + 220, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 522, + 220, + 536 + ], + "spans": [ + { + "bbox": [ + 131, + 522, + 220, + 536 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we present ST-LLM, a straightforward yet robust video large language model. Our aim is to achieve effective video comprehension by leveraging LLM to model video tokens, marking the inception of a joint spatial-temporal-text modeling paradigm. Additionally, we introduce a dynamic masking strategy and a global-local input module to enhance this framework. Remarkably, ST-LLM excels in encoding and comprehending spatial-temporal sequences while addressing concerns related to efficiency, stability, and modeling lengthy videos, all with reduced training resource requirements. Through extensive experimental analysis, we demonstrate the effectiveness of ST-LLM and its internal design, resulting in state-of-the-art performance across multiple video LLM benchmarks." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "text", + "content": "Liu. et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": "Acknowledgements. This work was supported by National Natural Science Foundation of China (No. 62172021). We are also grateful to the support from Guangdong Provincial Key Laboratory of Ultra High Definition Immersive Media Technology." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 147, + 164, + 474, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 164, + 474, + 175 + ], + "spans": [ + { + "bbox": [ + 147, + 164, + 474, + 175 + ], + "type": "text", + "content": "*: Equal Contribution, work done during internship at ARC Lab, Tencent PCG." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 190, + 197, + 203 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 190, + 197, + 203 + ], + "spans": [ + { + "bbox": [ + 133, + 190, + 197, + 203 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 212, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 212, + 481, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 212, + 481, + 245 + ], + "spans": [ + { + "bbox": [ + 138, + 212, + 481, + 245 + ], + "type": "text", + "content": "1. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 246, + 481, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 246, + 481, + 289 + ], + "spans": [ + { + "bbox": [ + 138, + 246, + 481, + 289 + ], + "type": "text", + "content": "2. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems 35, 23716-23736 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 289, + 481, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 289, + 481, + 320 + ], + "spans": [ + { + "bbox": [ + 138, + 289, + 481, + 320 + ], + "type": "text", + "content": "3. Anil, R., Dai, A.M., First, O., Johnson, M., Lepikhin, D., Passos, A., Shakeri, S., Taropa, E., Bailey, P., Chen, Z., et al.: Palm 2 technical report. arXiv preprint arXiv:2305.10403 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 321, + 481, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 321, + 481, + 364 + ], + "spans": [ + { + "bbox": [ + 138, + 321, + 481, + 364 + ], + "type": "text", + "content": "4. Awadalla, A., Gao, I., Gardner, J., Hessel, J., Hanafy, Y., Zhu, W., Marathe, K., Bitton, Y., Gadre, S., Sagawa, S., et al.: Openflamingo: An open-source framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 364, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 364, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 138, + 364, + 481, + 396 + ], + "type": "text", + "content": "5. Bain, M., Nagrani, A., Varol, G., Zisserman, A.: Frozen in time: A joint video and image encoder for end-to-end retrieval. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1728-1738 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 396, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 396, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 138, + 396, + 481, + 418 + ], + "type": "text", + "content": "6. Bao, H., Dong, L., Piao, S., Wei, F.: Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 418, + 481, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 418, + 481, + 438 + ], + "spans": [ + { + "bbox": [ + 138, + 418, + 481, + 438 + ], + "type": "text", + "content": "7. Bertasius, G., Wang, H., Torresani, L.: Is space-time attention all you need for video understanding? In: ICML. vol. 2, p. 4 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 439, + 481, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 439, + 481, + 482 + ], + "spans": [ + { + "bbox": [ + 138, + 439, + 481, + 482 + ], + "type": "text", + "content": "8. Brooks, T., Peebles, B., Holmes, C., DePue, W., Guo, Y., Jing, L., Schnurr, D., Taylor, J., Luhman, T., Luhman, E., Ng, C., Wang, R., Ramesh, A.: Video generation models as world simulators (2024), https://openai.com/research/video-generation-models-as-world-simulators" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 482, + 481, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 482, + 481, + 514 + ], + "spans": [ + { + "bbox": [ + 138, + 482, + 481, + 514 + ], + "type": "text", + "content": "9. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 514, + 481, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 514, + 481, + 546 + ], + "spans": [ + { + "bbox": [ + 138, + 514, + 481, + 546 + ], + "type": "text", + "content": "0. Caba Heilbron, F., Escorcia, V., Ghanem, B., Carlos Niebles, J.: Activitynet: A large-scale video benchmark for human activity understanding. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 961-970 (2015)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 547, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 547, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 547, + 481, + 578 + ], + "type": "text", + "content": "1. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6299-6308 (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 578, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 481, + 621 + ], + "type": "text", + "content": "2. Chen, J., Zhu, D., Shen, X., Li, X., Liu, Z., Zhang, P., Krishnamoorthi, R., Chandra, V., Xiong, Y., Elhoseiny, M.: Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 621, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 481, + 665 + ], + "type": "text", + "content": "3. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., et al.: Vicuna: An open-source chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 138, + 621, + 481, + 665 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 138, + 621, + 481, + 665 + ], + "type": "text", + "content": " chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023) (2023)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 171, + 91, + 445, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 91, + 445, + 102 + ], + "spans": [ + { + "bbox": [ + 171, + 91, + 445, + 102 + ], + "type": "text", + "content": "ST-LLM: Large Language Models Are Effective Temporal Learners" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "14. Dai, W., Li, J., Li, D., Tiong, A., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: towards general-purpose vision-language models with instruction tuning. arxiv. Preprint posted online on June 15, 2023 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 481, + 182 + ], + "type": "text", + "content": "15. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 182, + 481, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 182, + 481, + 214 + ], + "spans": [ + { + "bbox": [ + 133, + 182, + 481, + 214 + ], + "type": "text", + "content": "16. Driess, D., Xia, F., Sajjadi, M.S., Lynch, C., Chowdhery, A., Ichter, B., Wahid, A., Tompson, J., Vuong, Q., Yu, T., et al.: Palm-e: An embodied multimodal language model. arXiv preprint arXiv:2303.03378 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 214, + 481, + 268 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 214, + 481, + 268 + ], + "spans": [ + { + "bbox": [ + 133, + 214, + 481, + 268 + ], + "type": "text", + "content": "17. Goyal, R., Ebrahimi Kahou, S., Michalski, V., Materzynska, J., Westphal, S., Kim, H., Haenel, V., Fruend, I., Yianilos, P., Mueller-Freitag, M., et al.: The\" something something\" video database for learning and evaluating visual common sense. In: Proceedings of the IEEE international conference on computer vision. pp. 5842-5850 (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 268, + 481, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 268, + 481, + 290 + ], + "spans": [ + { + "bbox": [ + 133, + 268, + 481, + 290 + ], + "type": "text", + "content": "18. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 290, + 481, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 290, + 481, + 312 + ], + "spans": [ + { + "bbox": [ + 133, + 290, + 481, + 312 + ], + "type": "text", + "content": "19. Huang, B., Wang, X., Chen, H., Song, Z., Zhu, W.: Vtimellm: Empower llm to grasp video moments. arXiv preprint arXiv:2311.18445 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 312, + 481, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 312, + 481, + 343 + ], + "spans": [ + { + "bbox": [ + 132, + 312, + 481, + 343 + ], + "type": "text", + "content": "20. Jin, P., Takanobu, R., Zhang, C., Cao, X., Yuan, L.: Chat-univi: Unified visual representation empowers large language models with image and video understanding. arXiv preprint arXiv:2311.08046 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 343, + 481, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 481, + 375 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 481, + 375 + ], + "type": "text", + "content": "21. Kay, W., Carreira, J., Simonyan, K., Zhang, B., Hillier, C., Vijayanarasimhan, S., Viola, F., Green, T., Back, T., Natsev, P., et al.: The kinetics human action video dataset. arXiv preprint arXiv:1705.06950 (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 375, + 481, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 375, + 481, + 397 + ], + "spans": [ + { + "bbox": [ + 132, + 375, + 481, + 397 + ], + "type": "text", + "content": "22. Li, B., Zhang, Y., Chen, L., Wang, J., Yang, J., Liu, Z.: Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 397, + 481, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 481, + 429 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 481, + 429 + ], + "type": "text", + "content": "23. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 429, + 481, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 429, + 481, + 461 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 481, + 461 + ], + "type": "text", + "content": "24. Li, K., He, Y., Wang, Y., Li, Y., Wang, W., Luo, P., Wang, Y., Wang, L., Qiao, Y.: Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 461, + 481, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 461, + 481, + 494 + ], + "spans": [ + { + "bbox": [ + 132, + 461, + 481, + 494 + ], + "type": "text", + "content": "25. Li, K., Wang, Y., He, Y., Li, Y., Wang, Y., Liu, Y., Wang, Z., Xu, J., Chen, G., Luo, P., et al.: Mvbench: A comprehensive multi-modal video understanding benchmark. arXiv preprint arXiv:2311.17005 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 494, + 481, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 494, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 494, + 481, + 525 + ], + "type": "text", + "content": "26. Li, K., Wang, Y., Li, Y., Wang, Y., He, Y., Wang, L., Qiao, Y.: Unmasked teacher: Towards training-efficient video foundation models. arXiv preprint arXiv:2303.16058 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 525, + 481, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 481, + 548 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 481, + 548 + ], + "type": "text", + "content": "27. Li, Y., Wang, C., Jia, J.: Llama-vid: An image is worth 2 tokens in large language models. arXiv preprint arXiv:2311.17043 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 548, + 481, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 548, + 481, + 579 + ], + "spans": [ + { + "bbox": [ + 132, + 548, + 481, + 579 + ], + "type": "text", + "content": "28. Lin, B., Zhu, B., Ye, Y., Ning, M., Jin, P., Yuan, L.: Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 579, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 579, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 579, + 481, + 601 + ], + "type": "text", + "content": "29. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv preprint arXiv:2304.08485 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 601, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 481, + 643 + ], + "type": "text", + "content": "30. Liu, R., Huang, J., Li, G., Feng, J., Wu, X., Li, T.H.: Revisiting temporal modeling for clip-based image-to-video knowledge transferring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6555-6564 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "text", + "content": "31. Liu, R., Li, C., Ge, Y., Shan, Y., Li, T.H., Li, G.: One for all: Video conversation is feasible without video instruction tuning. arXiv preprint arXiv:2309.15785 (2023)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 206, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 206, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 206, + 100 + ], + "type": "text", + "content": "Liu. et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "type": "text", + "content": "32. Luo, R., Zhao, Z., Yang, M., Dong, J., Qiu, M., Lu, P., Wang, T., Wei, Z.: Valley: Video assistant with large language model enhanced ability. arXiv preprint arXiv:2306.07207 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "type": "text", + "content": "33. Maaz, M., Rasheed, H., Khan, S., Khan, F.S.: Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 184, + 402, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 184, + 402, + 196 + ], + "spans": [ + { + "bbox": [ + 132, + 184, + 402, + 196 + ], + "type": "text", + "content": "34. OpenAI: Chatgpt. https://openai.com/blog/chatgpt (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 196, + 481, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 196, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 196, + 481, + 239 + ], + "type": "text", + "content": "35. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems 35, 27730-27744 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 240, + 481, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 240, + 481, + 274 + ], + "spans": [ + { + "bbox": [ + 132, + 240, + 481, + 274 + ], + "type": "text", + "content": "36. Peebles, W., Xie, S.: Scalable diffusion models with transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4195-4205 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 274, + 481, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 481, + 295 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 481, + 295 + ], + "type": "text", + "content": "37. Peng, Z., Dong, L., Bao, H., Ye, Q., Wei, F.: Beit v2: Masked image modeling with vector-quantized visual tokenizers. arXiv preprint arXiv:2208.06366 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 296, + 481, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 296, + 481, + 340 + ], + "spans": [ + { + "bbox": [ + 132, + 296, + 481, + 340 + ], + "type": "text", + "content": "38. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 341, + 481, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 341, + 481, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 341, + 481, + 363 + ], + "type": "text", + "content": "39. Su, J., Ahmed, M., Lu, Y., Pan, S., Bo, W., Liu, Y.: Roformer: Enhanced transformer with rotary position embedding. Neurocomputing 568, 127063 (2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 363, + 481, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 363, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 363, + 481, + 385 + ], + "type": "text", + "content": "40. Sun, Q., Fang, Y., Wu, L., Wang, X., Cao, Y.: Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 386, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 481, + 407 + ], + "type": "text", + "content": "41. Taori, R., Gulrajani, I., Zhang, T., Dubois, Y., Li, X., Guestrin, C., Liang, P., Hashimoto, T.B.: Stanford alpaca: An instruction-following llama model (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 408, + 481, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 408, + 481, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 408, + 481, + 441 + ], + "type": "text", + "content": "42. Tong, Z., Song, Y., Wang, J., Wang, L.: Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. Advances in neural information processing systems 35, 10078-10093 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 442, + 481, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 442, + 481, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 442, + 481, + 475 + ], + "type": "text", + "content": "43. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 475, + 481, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 481, + 509 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 481, + 509 + ], + "type": "text", + "content": "44. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 509, + 481, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 509, + 481, + 542 + ], + "spans": [ + { + "bbox": [ + 132, + 509, + 481, + 542 + ], + "type": "text", + "content": "45. Tran, D., Bourdev, L., Fergus, R., Torresani, L., Paluri, M.: Learning spatiotemporal features with 3d convolutional networks. In: Proceedings of the IEEE international conference on computer vision. pp. 4489-4497 (2015)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 543, + 481, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 543, + 481, + 564 + ], + "spans": [ + { + "bbox": [ + 132, + 543, + 481, + 564 + ], + "type": "text", + "content": "46. Wu, Z., Yao, T., Fu, Y., Jiang, Y.G.: Deep learning for video classification and captioning. In: Frontiers of multimedia research, pp. 3-29. ACM (2017)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 565, + 481, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 565, + 481, + 598 + ], + "spans": [ + { + "bbox": [ + 132, + 565, + 481, + 598 + ], + "type": "text", + "content": "47. Xiao, J., Shang, X., Yao, A., Chua, T.S.: Next-qa: Next phase of question-answering to explaining temporal actions. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9777-9786 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "type": "text", + "content": "48. Xu, J., Mei, T., Yao, T., Rui, Y.: Msr-vtt: A large video description dataset for bridging video and language. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5288-5296 (2016)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "49. Xue, H., Sun, Y., Liu, B., Fu, J., Song, R., Li, H., Luo, J.: Clip-vip: Adapting pre-trained image-text model to video-language representation alignment. arXiv preprint arXiv:2209.06430 (2022)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 171, + 91, + 445, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 91, + 445, + 102 + ], + "spans": [ + { + "bbox": [ + 171, + 91, + 445, + 102 + ], + "type": "text", + "content": "ST-LLM: Large Language Models Are Effective Temporal Learners" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 304 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "text", + "content": "50. Yang, A., Miech, A., Sivic, J., Laptev, I., Schmid, C.: Zero-shot video question answering via frozen bidirectional language models. Advances in Neural Information Processing Systems 35, 124-141 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 482, + 182 + ], + "type": "text", + "content": "51. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 481, + 215 + ], + "type": "text", + "content": "52. Yi, K., Gan, C., Li, Y., Kohli, P., Wu, J., Torralba, A., Tenenbaum, J.B.: Clevrer: Collision events for video representation and reasoning. arXiv preprint arXiv:1910.01442 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "type": "text", + "content": "53. Zhang, H., Li, X., Bing, L.: Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "type": "text", + "content": "54. Zhang, R., Han, J., Zhou, A., Hu, X., Yan, S., Lu, P., Li, H., Gao, P., Qiao, Y.: Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 270, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 481, + 304 + ], + "type": "text", + "content": "55. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023)" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 206, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 206, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 206, + 100 + ], + "type": "text", + "content": "Liu. et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/STAG4D_ Spatial-Temporal Anchored Generative 4D Gaussians/51eb6cce-864e-4035-8901-6f032e5306da_content_list.json b/2024/STAG4D_ Spatial-Temporal Anchored Generative 4D Gaussians/51eb6cce-864e-4035-8901-6f032e5306da_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5328a67ae6dab82bd92885257cbdd2854cd5f51f --- /dev/null +++ b/2024/STAG4D_ Spatial-Temporal Anchored Generative 4D Gaussians/51eb6cce-864e-4035-8901-6f032e5306da_content_list.json @@ -0,0 +1,1545 @@ +[ + { + "type": "text", + "text": "STAG4D: Spatial-Temporal Anchored Generative 4D Gaussians", + "text_level": 1, + "bbox": [ + 218, + 140, + 782, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yifei Zeng $^{1}$ , Yanqin Jiang $^{2}$ , Siyu Zhu $^{3}$ , Yuanxun Lu $^{1}$ , Youtian Lin $^{1}$ , Hao Zhu $^{1}$ , Weiming Hu $^{2}$ , Xun Cao $^{1}$ , and Yao Yao $^{1}$", + "bbox": [ + 215, + 210, + 785, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Nanjing University \n $^{2}$ CASIA", + "bbox": [ + 429, + 253, + 573, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Fudan University", + "bbox": [ + 434, + 282, + 566, + 296 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Recent progress in pre-trained diffusion models and 3D generation have spurred interest in 4D content creation. However, achieving high-fidelity 4D generation with spatial-temporal consistency remains a challenge. In this work, we propose STAG4D, a novel framework that combines pre-trained diffusion models with dynamic 3D Gaussian splatting for high-fidelity 4D generation. Drawing inspiration from 3D generation techniques, we utilize a multi-view diffusion model to initialize multiview images anchoring on the input video frames, where the video can be either real-world captured or generated by a video diffusion model. To ensure the temporal consistency of the multi-view sequence initialization, we introduce a simple yet effective fusion strategy to leverage the first frame as a temporal anchor in the self-attention computation. With the almost consistent multi-view sequences, we then apply the score distillation sampling to optimize the 4D Gaussian point cloud. The 4D Gaussian spatting is specially crafted for the generation task, where an adaptive densification strategy is proposed to mitigate the unstable Gaussian gradient for robust optimization. Notably, the proposed pipeline does not require any pre-training or fine-tuning of diffusion networks, offering a more accessible and practical solution for the 4D generation task. Extensive experiments demonstrate that our method outperforms prior 4D generation works in rendering quality, spatial-temporal consistency, and generation robustness, setting a new state-of-the-art for 4D generation from diverse inputs, including text, image, and video.", + "bbox": [ + 261, + 332, + 743, + 652 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: 4D Generation $\\cdot$ 3D Gaussian Splatting $\\cdot$ Diffusion Model", + "bbox": [ + 261, + 662, + 735, + 678 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 703, + 375, + 719 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advancements in large-scale pre-trained diffusion models have shown remarkable progress in producing high-quality and diverse visual content, including images, videos, and 3D asserts [5,13,16,22,24,27,31,38]. The progress naturally extends to the realm of dynamic 3D generation [1,6,12,28,32,36,39,40]. This endeavor has gained prominence in computer vision and generative AI research, as high-quality 4D content generation is key to a broad range of applications such as autonomous driving simulation, game and film industries, digital Avatar", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/947e1e25b9801a2e83c615582ff67dac2fb4bc1fb83e9d9ddf34dd56b2213027.jpg", + "image_caption": [ + "Fig. 1: Visualization of the generated 4D asserts. Our approach can generate diverse 4D content from various inputs, including text, image, and video." + ], + "image_footnote": [], + "bbox": [ + 217, + 145, + 514, + 268 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c7c283703a6144e5e95f8156d0823c0f101035b83cc3d4ce347ada2cca08b41d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 522, + 145, + 787, + 258 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "creation, and spatial video production. However, previous 4D generation methods face challenges including blurry rendering, spatial-temporal inconsistency, and slow generation speed. Generating high-quality 4D content efficiently and practically remains a significant challenge.", + "bbox": [ + 212, + 348, + 782, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given the promising strides in 3D generation from pre-trained diffusion models, there has been a concerted focus on generalized dynamic 3D generation from text or uncalibrated monocular video [1,6,12,28,32,40]. MAV3D [28] is the pioneering work of text-to-4D generation which creates 4D scenes from textual descriptions through Score Distillation Sampling (SDS) from a video diffusion model. Alternatively, Consistent4D [6] takes a video as input, and focuses on the task of video-to-4D generation. The pipeline is more flexible in that it can take advantage of high-quality input videos, dividing the difficult problem of realistic 4D content creation into a practical pipeline of text-to-video and video-to-4D generation. However, current video-to-4D generation methods still struggle to generate high-fidelity 4D asserts with spatial and temporal rendering consistency. Meanwhile, the commonly used Neural Radiance Field(NeRF) representation would suffer from an over-saturated appearance and a long optimization time.", + "bbox": [ + 212, + 412, + 784, + 622 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we address two important factors in 4D content creation: a suitable 4D representation and spatial-temporal consistency imposed during pseudoprojection generation (i.e., multi-view video generation). Our approach employs a 4D Gaussian splatting approach specially tailored for the generation task. In the generation process, we first generate the multi-view images at all timestamps of the 4D scene. To enforce the 4D consistency of the multi-view video initialization, we design a direct fusion approach for attention computation in both spatial and temporal domains. This approach implicitly addresses texture degradation and geometric misalignment, thereby avoiding the need for adding an explicit multi-view or temporal consistency loss during the Score Distillation Sampling(SDS) [22] optimization. Then, the reference photometric loss and the SDS loss conditioned on the initial multi-view video are applied for 4D Gaussian point cloud optimization. To tackle the instability in optimizing the 4D Gaussian points, we introduce an adaptive densification strategy that is informed by the", + "bbox": [ + 212, + 628, + 787, + 843 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Zeng et al.", + "bbox": [ + 271, + 114, + 344, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Gaussian gradient distribution. This approach yields high-quality and robust 4D scene generation from a monocular video input.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We have conducted comprehensive experiments to demonstrate the effectiveness of each component of the proposed method. Our approach achieves a 2x faster generation speed compared with the previous video-to-4D approaches (e.g., [6]) and a significantly better generation quality than prior state-of-the-art methods. It is also noteworthy that the 4D content generated by our method can be rendered in real-time, opening up a wide range of possibilities for practical applications. In summary, our primary contributions are as follows:", + "bbox": [ + 212, + 176, + 784, + 282 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a holistic 4D generation pipeline that streamlines the generation process into sequential stages: video generation, multi-view video initialization, and 4D optimization using multi-view video conditioned SDS.", + "- We harness the dynamic 3D Gaussian representation for 4D generation, complemented by an adaptive densification strategy. This combination enables highly precise and efficient 4D generation from monocular video inputs.", + "- We present a novel training-free attention fusion module to effectively integrate temporal anchor frames into the multi-view diffusion process, significantly enhancing the 4D consistency of the generated multi-view videos.", + "- Our method outperforms previous approaches in optimization efficiency, rendering quality, and 4D consistency, establishing a new benchmark for 4D generation across various input types, such as text, images, and videos." + ], + "bbox": [ + 225, + 297, + 782, + 477 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related work", + "text_level": 1, + "bbox": [ + 214, + 506, + 382, + 522 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 3D Generation", + "text_level": 1, + "bbox": [ + 214, + 540, + 385, + 554 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dreamfusion [22] first proposes the SDS loss [22, 30] for NeRF optimization, which stands for the most prevalent technique for nowadays text-to-3D approaches. To mitigate the multi-view Janus problem, Zero123 [13] and SycnDreammer [14] fine-tune 2D diffusion models to grant the image generator with the ability of viewpoint control, while later works [15, 17, 26, 27] explicitly generate fixed multi-view images in one diffusion pass. Building upon the success of multi-view diffusion methods, we integrate a multi-view generation module into the 4D generation task, with meticulous handling of 4D spatial-temporal consistency for enhanced 4D generation quality.", + "bbox": [ + 212, + 566, + 784, + 703 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Except for equipping diffusion models with multi-view or depth awareness, another direction of progress focuses on 3D representations. Following the idea of differentiable optimization, Magic3D [10] applies a two-stage generation pipeline and changes the 3D representation to instant-NGP [20] and DMTet [25], achieving faster runtime and better generation quality. Direct2.5 [17] applies an explicit mesh representation and uses differentiable rasterization for fast mesh optimization. Recently, with the newly developed 3DGS, DreamGaussian [29] demonstrates the ability to generate 3D objects within several minutes by substituting NeRF with the 3D Gaussian representation, showing the potential of using an", + "bbox": [ + 212, + 704, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "STAG4D", + "bbox": [ + 666, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "explicit point-based representation in 3D generation tasks. In this work, we introduce a novel 4D Gaussian representation and a specially tailored optimization scheme for the 4D generation task.", + "bbox": [ + 215, + 146, + 784, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 4D Reconstruction", + "text_level": 1, + "bbox": [ + 215, + 215, + 415, + 231 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Dynamic 3D reconstruction is yet another heated research topic in the field of computer vision and graphics. By extending the static NeRF [19] framework to dynamic scenes, dynamic NeRFs [3, 4, 9, 34] have demonstrated remarkable progress on dynamic 3D reconstruction. However, limited by the implicit neural representation and the complex nature of dynamic 3D information, dynamic NeRFs still suffer from slow optimization speed and low reconstruction quality. With the development of 3DGS [7], researchers have quickly extended the 3D Gaussian to dynamic scene representation, which achieves faster training and rendering speeds compared to D-NeRFs based on implicit neural representations. Dynamic 3D Gaussian [18] applies the per-frame 3DGS optimization for 4D scene reconstruction. Some studies [33, 35] attempt to amalgamate the explicit point-based 3DGS with an implicit neural field for dynamic information modeling, while Gaussian-Flow [11] introduces an explicit per-point motion model to represent a 4D scene without using implicit neural networks. It is noteworthy that it is non-trivial to extend the Gaussian reconstruction to 4D reconstruction as each reconstruction pipeline requires heuristic tuning of hyperparameters at each stage. In this paper, we target combining the 4DGS for the challenging task of 4D generation.", + "bbox": [ + 215, + 243, + 787, + 517 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 4D Generation", + "text_level": 1, + "bbox": [ + 215, + 540, + 383, + 555 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recent developments in 3D content generation and video diffusion technologies have triggered researchers' interest in exploring 4D content generation from various input conditionings. MAV3D [28] presents the early attempt at text-to-4D generation. By utilizing score distillation sampling derived from video diffusion models, MAV3D optimizes a dynamic NeRF based on textual prompts. In a parallel vein, Consistent4D [6] has introduced the task of video-to-4D. This method capitalizes on the pre-trained knowledge from image diffusion models to optimize dynamic NeRFs via SDS optimization, showcasing the potential of high-quality 4D content generation from pre-trained 2D diffusion models. Concurrently with our work, 4DGen [36] has introduced a framework for generating dynamic 3D models, which employs spatial-temporal pseudo labels on keyframes within a multi-view diffusion model. However, the overall quality of its generation can be further improved, as its rendering fidelity is limited by the implicit neural representation and inadequate spatial-temporal information exchange exists in the diffusion generation process. In contrast, our method applies an efficient 4D Gaussian representation with spatial-temporal attention from reference- and the first-frame anchors, achieving substantially better generation results than previous works.", + "bbox": [ + 215, + 568, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Zeng et al.", + "bbox": [ + 271, + 114, + 344, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c7e10cdc452231c0f2744e8812fb2453d8d06434cdcc68be42bc0d9174837572.jpg", + "image_caption": [ + "Fig. 2: Overall pipeline. Given a video input, we apply a multi-view diffusion model to produce coherent multi-view sequences, which serve as spatial and temporal anchors. Next, we train a deformable 3D Gaussian using multi-view SDS loss and reference loss. For a text-to-4D generation, our pipeline can be naturally extended to accept the text input by integrating with an off-the-shelf text-to-video module." + ], + "image_footnote": [], + "bbox": [ + 222, + 146, + 781, + 343 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 453, + 328, + 468 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The framework of the proposed approach, as depicted in Figure 2, outlines the generation and optimization process of a dynamic scene from a given video input. Also, our approach can be easily extended to a text-to-4D generation pipeline by utilizing an off-the-shelf video generation module. Section 3.1 presents the 4D representation used in the proposed method, along with an adaptive densification strategy that adjusts the densification threshold based on the relative motion gradient of the scene points. Subsequently, Section 3.2 provides a brief overview of the multi-view diffusion model used in the pipeline and introduces a novel spatial-temporal attention fusion mechanism designed to produce the almost consistent multi-view image sequences. Finally, as detailed in Section 3.3, the generated multi-view sequences are leveraged with spatial and temporal anchor frames for consistent 4D generation with the SDS optimization.", + "bbox": [ + 212, + 484, + 787, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 4D Representation", + "text_level": 1, + "bbox": [ + 215, + 686, + 418, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4D Gaussian Splatting The concept of 3D Gaussian Splatting was initially introduced in the work of [7], where an explicit point-based representation is used to model the 3D scene. Specifically, the 3D Gaussian point cloud, denoted as $\\mathcal{S}$ for a static scene, is characterized by the tuple:", + "bbox": [ + 212, + 710, + 787, + 772 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} = [ \\mathcal {X}, s, r, \\sigma , \\zeta ], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 436, + 782, + 785, + 799 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{X} = (x,y,z)$ denotes the positions of the 3D Gaussian points, and $s$ , $r$ , $\\sigma$ and $\\zeta$ represent the scale, rotation, opacity, and spherical harmonics (SH) coeffi-", + "bbox": [ + 214, + 809, + 785, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "STAG4D", + "bbox": [ + 668, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "cients of the radiance, respectively. In 3D reconstruction, the 3D Gaussian point cloud will be optimized through a point-based differentiable volume rendering.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This representation can be naturally extended to 4D scenes by presenting the continuous scene dynamics as a 3D motion field. Inspired by [33], we present a 4D scene as a 3D Gaussian point cloud with a hex-plane-based deformation field, denoted as $\\mathcal{F}(\\mathcal{S},t)$ . The 3D Gaussian point cloud at time $t$ can be expressed as:", + "bbox": [ + 212, + 178, + 784, + 238 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} (\\mathcal {S}, t) = \\left[ \\mathcal {X} _ {t}, s _ {t}, r _ {t}, \\sigma , \\zeta \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 413, + 253, + 784, + 270 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{X}_t = (x_t, y_t, z_t)$ , $s_t$ and $r_t$ represent the updated information of Gaussian position, scale, and rotation at time $t$ .", + "bbox": [ + 212, + 282, + 782, + 313 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Adaptive Densification The standard 3D Gaussian Splatting technique typically employs a point cloud densification control strategy to dynamically adjust the number of Gaussians and their density within a unit volume. This adaptive approach allows for the transition from an initial sparse Gaussian set to a denser configuration to better represent the 3D scene. The 4D Gaussian Splatting method introduced in the work [33] utilizes a densification strategy similar to that presented in vanilla 3DGS [7], which involves the use of a fixed densification threshold based on the view-space position gradient. However, while the fixed gradient threshold strategy demonstrates efficacy in the context of reconstruction, particularly when multiple views afford robust and redundant coverage of the target scene, it performs sub-optimally in the generative setting. This limitation arises from the constraints imposed by the input single image or monocular videos, leading to significant uncertainty in the spatial and scale dimensions for each training object. This will lead to different optimal thresholds for different cases, as illustrated in the ablation study.", + "bbox": [ + 212, + 340, + 784, + 568 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To address the issue, we propose an adaptive threshold approach, wherein only candidates with relatively large gradients are selected for densification. Our approach is based on the statistical analysis of the accumulated gradient of each point, which follows a log distribution of similar shapes throughout the training process. We apply an adaptive threshold that filters out Gaussians with a relatively small gradient and only densifies those with a large gradient. The adaptive threshold is set to select a fixed percentage (top $\\lambda\\%$ ) of points with the highest gradient in each densification operation. This ensures that the threshold adapts to the distribution of the gradient and maintains a stable relative position. Experimental results demonstrate that the proposed simple strategy can significantly enhance the quality and robustness of the 4D generation.", + "bbox": [ + 212, + 569, + 785, + 734 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Temporal and Multi-view Consistent Diffusion", + "text_level": 1, + "bbox": [ + 212, + 763, + 645, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Multi-view Consistent Diffusion Our pipeline adopts Score Distillation Sampling(SDS) [22] from image diffusion models for the optimization of dynamic Gaussian. Specifically, Image-to-image diffusion model Zero123 [13] is utilized", + "bbox": [ + 212, + 794, + 782, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Zeng et al.", + "bbox": [ + 271, + 114, + 344, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and the SDS loss gradient could be formulated as follows:", + "bbox": [ + 215, + 146, + 630, + 161 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\theta} \\mathcal {L} _ {S D S} (\\phi , \\mathbf {x}) = \\mathbb {E} _ {t, \\epsilon} \\left[ \\omega (t) \\left(\\hat {\\epsilon} _ {\\theta} \\left(\\mathbf {z} _ {t}; \\mathbf {I} _ {i n}, \\mathbf {R}, \\mathbf {T}, t\\right) - \\epsilon\\right) \\frac {\\partial \\mathbf {x}}{\\partial \\theta} \\right], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 305, + 170, + 784, + 204 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {z} = z _ {t} - \\sigma_ {t} \\hat {\\epsilon} _ {t} (z; \\mathbf {I} _ {i n}, \\mathbf {R}, \\mathbf {T}).\n$$\n", + "text_format": "latex", + "bbox": [ + 398, + 205, + 584, + 220 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\theta$ represents the parameters of the 3D representation, $\\mathbf{x}$ the rendered image at the current view, $t$ the timestamp in the diffusion process, $\\epsilon$ the ground truth noise, $\\hat{\\epsilon}$ the predicted noise from the noisy image $\\mathbf{z}_{\\mathbf{t}}$ conditioned on an initial input $\\mathbf{I}_{in}$ , and the relative camera pose between input view and target view $(\\mathbf{R},\\mathbf{T})$ . Zero123 can generate the target view image at any relative camera location but only one target view at a time, thus good at optimizing the 3D object from all views whilst bad at generating spatially consistent images from multiple target views. In contrast, Zero123++ [26] leverages reference attention [37] to model the relationship of the images from multiple target views as well as the input view, resulting in multi-view consistency output yet at the cost of fixed target camera locations. We combine the advantages of both models by taking the multi-view consistent output of Zero123++ as the input images of Zero123 when calculating SDS loss.", + "bbox": [ + 215, + 229, + 787, + 426 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Temporally Consistent Diffusion Following the view generation philosophy of multi-view diffusion based 3D generation, a practical solution to 4D content creation is to generate the multi-view videos of the 4D scene, and then apply the multi-view video conditioned SDS loss to optimize the scene. However, it is rather difficult to generate temporally consistent multi-view videos by the separate per-frame generation. Inspired by zero-shot video generation method [8], we propose a training-free temporal attention module to enable Zero123++ [26] with temporal-awareness.", + "bbox": [ + 212, + 446, + 787, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In our design, we obtain the attention information during denoising the multiview images of the first frame. Then we apply the recorded attention to the later denoising process. Specifically, during the denoising process of the latent code, the self-attention layer computes three essential components—queries $(Q)$ , keys $(K)$ , and values $(V)$ , which is formulated as:", + "bbox": [ + 212, + 568, + 787, + 643 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S e l f - A t t n} (Q, K, V) = \\operatorname {S o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {c}}\\right) V. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 652, + 784, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We first compute the key $K_{0}$ and value $V_{0}$ for the initial frame, $T_{0}$ . Subsequently, for each subsequent frame $T_{t}$ , where $t \\in \\{1,\\dots,N\\}$ , we perform a mixing operation on the key $K_{t}$ and value $V_{t}$ with the initial key $K_{0}$ and value $V_{0}$ derived from the first frame $T_{0}$ . This process can be formally represented as follows:", + "bbox": [ + 215, + 693, + 787, + 753 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} K _ {t} = \\gamma \\mathbf {K} _ {0} + (1 - \\gamma) \\mathbf {K} _ {t} \\\\ V _ {t} = \\gamma \\mathbf {V} _ {0} + (1 - \\gamma) \\mathbf {V} _ {t}, \\end{array} \\right. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 763, + 784, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where the parameter $\\gamma$ serves as a weighting factor that governs the influence of the initial frame, as illustrated in the ablation part. Additionally, we employ", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "STAG4D", + "bbox": [ + 666, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/03e8900e2be9d449b0b6bb36f4e8a47fb662986816617e26cfa5b1cf096879d8.jpg", + "image_caption": [ + "Fig. 3: Illustration of the proposed spatial and temporal attention fusion during multiview sequence inference." + ], + "image_footnote": [], + "bbox": [ + 228, + 148, + 777, + 344 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "reference attention following the approach outlined in [26,37], to extract local conditioning from the input image. In this reference attention mechanism, both the key and the value are obtained through the concatenation of features from the reference image and the noisy input image. Our design of temporal attention module is simple yet effective with superior temporal consistency, image quality and 3D consistency to vanilla cross-frame attention proposed in [8], as demonstrated in experiment section.", + "bbox": [ + 212, + 419, + 787, + 526 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 Training Objectives", + "text_level": 1, + "bbox": [ + 215, + 550, + 424, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Following the generation of multi-view sequences from a monocular reference video, we can acquire 6 anchor views $\\{I_t^i\\}_{i\\in \\{1\\dots 6\\}}$ and a reference view $I_{t}^{ref}$ at each timestep $t$ . In our optimization process, we employ multi-view score distillation sampling (SDS) utilizing the generated images $\\{I_t^i\\}_{i = 1\\dots 6}$ in conjunction with the reference images $I_{t}^{ref}$ . The multi-view score distillation loss function $\\mathcal{L}_{MVSDS}$ can be defined as:", + "bbox": [ + 214, + 578, + 787, + 672 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {M V S D S} = \\lambda_ {1} \\mathcal {L} _ {S D S} ^ {i} + \\lambda_ {2} \\mathcal {L} _ {S D S} ^ {r e f} \\\\ = \\lambda_ {1} \\mathcal {L} _ {S D S} \\left(\\phi , I _ {t} ^ {i}\\right) + \\lambda_ {2} \\mathcal {L} _ {S D S} \\left(\\phi , I _ {t} ^ {\\text {r e f}}\\right), \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 333, + 681, + 784, + 722 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\lambda_{1}$ and $\\lambda_{2}$ are two weighting factors, the index $i$ is determined based on the proximity of the rendering viewpoint to the viewpoint of the generated images. This selection process, which we refer to as multi-view score distillation sampling, involves choosing the nearest reference image to the rendered camera view for calculating the SDS loss.", + "bbox": [ + 212, + 733, + 787, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Following the approach [29], we utilize the reference image to compute both the reconstruction loss $\\mathcal{L}_{rec}$ and the foreground mask loss $\\mathcal{L}_{mask}$ . Thus, the final", + "bbox": [ + 214, + 809, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Zeng et al.", + "bbox": [ + 271, + 114, + 344, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "optimization objective is:", + "bbox": [ + 215, + 146, + 401, + 161 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {M V S D S} + \\lambda_ {3} \\mathcal {L} _ {r e c} + \\lambda_ {4} \\mathcal {L} _ {\\text {m a s k}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 178, + 785, + 193 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\lambda_3$ and $\\lambda_4$ are the weighting parameters. During training, we first use $\\mathcal{L}$ to supervise a fixed frame to get static canonical 3D Gaussian. Then we use all anchors and reference images to train the dynamic 4D Gaussian point cloud.", + "bbox": [ + 214, + 208, + 785, + 253 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.4 Text-to-4D Extension", + "text_level": 1, + "bbox": [ + 215, + 280, + 441, + 294 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Our method is designed to be readily adaptable to text and image inputs, offering novel capabilities such as directly generating video sequences from textual descriptions or static images. We initiate this process using a 2D diffusion model SDXL [21], which generates the image from textual input. This 2D output is then transformed into a video sequence via a video diffusion model, such as SVD [2], adding temporal coherence and motion to the static image. Finally, the aforementioned pipeline further lifts the video sequence to a 4D scene, introducing an extra spatial dimension that creates immersive experiences beyond conventional 2D or 3D media. The generation pipeline with the given text, image, and video inputs is visualized in Fig. 2.", + "bbox": [ + 214, + 309, + 787, + 460 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiment", + "text_level": 1, + "bbox": [ + 215, + 488, + 366, + 505 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Experiment Setup", + "text_level": 1, + "bbox": [ + 215, + 523, + 413, + 539 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Dataset For the video-to-4D task, we utilize the dataset provided by Consistent4D [6] for quantitative evaluation, which comprises multi-view videos depicting 7 dynamic objects. Additionally, for qualitative evaluation, we curate a set of challenging videos from online sources to assess the robustness and generalization capabilities of each method. For text/image-to-4D tasks, we follow the data settings in 4Dfy [1] and DreamGaussian4D [23] to create the corresponding results. Notably, we use the same input image and input video with DreamGaussian4D for a fair comparison.", + "bbox": [ + 214, + 551, + 787, + 672 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Evaluation metrics We employ CLIP, LPIPS, and FVD for Video-to-4D evaluation as in the previous approach [6]. Specifically, CLIP and LPIPS serve as image-level metrics, assessing the semantic similarity between rendered images and ground truths; FVD is a video-level metric commonly used in video generation tasks, considering not only single-frame quality but also temporal coherence; we also consider the FID-VID metric to measure the temporal consistency. As 4D-Gen [36] does not support a 30-frame setting we report its FVD under a 16-frame reconstruction, namely FVD-16 as an additional metrics. For text-to-4D generation, we follow previous works and provide a user study for quantitative comparisons between different methods. We also provide a user study for video-to-4D and both studies are detailed in the supplementary material.", + "bbox": [ + 214, + 672, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "STAG4D", + "bbox": [ + 666, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/739a36b04b574fcb05f5e51cdedbcc347c55f2297d53625a5e939a38ef6e8997.jpg", + "image_caption": [ + "Fig. 4: Qualitative comparison on video-to-4D generation. Our method achieves better faithfulness to the input video with higher quality on the overall reconstruction." + ], + "image_footnote": [], + "bbox": [ + 220, + 143, + 784, + 327 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Baselines For video-to-4D generation, we compare our method with two baselines: Consistent4D [6] and the concurrent work 4DGen [36]. We utilize the official code released by the authors to generate comparison results. For text&image-to-4D generation, we make comparisons with two state-of-art models that are open-sourced: 4Dfy [1] and DreamGaussian4D [23]. Their results are generated using codes from their official GitHub repository.", + "bbox": [ + 212, + 405, + 782, + 497 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 530, + 455, + 545 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In the training of the video-to-4D generative model, we adopt a two-stage approach. Initially, the model is initialized in canonical space and trained for 1000 steps. Subsequently, the deformation fields are learned over 7000 additional steps to accurately capture the dynamic scene. For the deformation process, we employ multi-layer perceptrons (MLPs) with 64 hidden layers and 32 hidden features per layer. The initial learning rate for the deformation MLPs is set to $1.6 \\times 10^{-4}$ and is decayed to $1.6 \\times 10^{-6}$ by the end of the training. In the context of adaptive densification, we opt to densify the top $2.5\\%$ of points with the most accumulated gradient. In the generative phase, we utilize the spatial-temporal consistent videos for the multiview score distillation sampling optimization. Empirically, we set $\\lambda$ to 0.5 for our temporally consistent diffusion. Regarding the loss functions, we maintain the SDS loss weight at 1 and adjust the weights of other losses accordingly. Specifically, the reconstruction loss is assigned to $4 \\times 10^{4}$ , the mask loss is weighted at $1 \\times 10^{4}$ . The training process requires approximately 1 hour on an RTX 3090 GPU, and the rendering process can be performed at 150 FPS in real time. For Text&Image-to-4D, we use SDXL [21] for image generation, and apply SVD [2] to the generated image to create a corresponding video. Then we use the same setting as Video-to-4D to transfer the video into 4D content.", + "bbox": [ + 212, + 568, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Zeng et al.", + "bbox": [ + 271, + 114, + 344, + 127 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/1ed653f04bfc9197b1adfa0edb704b373dc56507d1104fd7559a2fb1a44a4531.jpg", + "image_caption": [ + "Fig. 5: Qualitative comparison on text-to-4D generation. Our method achieves the best results on the text alignment and visual quality." + ], + "image_footnote": [], + "bbox": [ + 218, + 146, + 781, + 354 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Comparison with State-of-the-art Approaches", + "text_level": 1, + "bbox": [ + 215, + 424, + 638, + 439 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In this section, we conduct a comprehensive comparison with the aforementioned baselines using synthetic and in-the-wild data. The superior performance demonstrated in both quantitative and qualitative evaluations serves to substantiate the effectiveness and robustness of our proposed method.", + "bbox": [ + 212, + 446, + 784, + 507 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Quantitative Results on Video-to-4D The experimental results presented in Table 1 offer a comparative analysis of our method against state-of-the-art techniques, namely Consistent4D and 4DGen, across several metrics. Our method outperforms Consistent4D and 4DGen in CLIP, showing better semantic consistency with the target content. It also achieves the lowest LPIPS score, meaning it generates more realistic images than the others. For the video quality and smoothness, our method beats 4DGen in FID-VID and FVD-16, and Consistent4D in FVD, implying that our videos are closer to real videos and have less temporal artifacts. Overall, the numbers confirm that our method is superior in generating semantically aligned and perceptually convincing videos with high fidelity and coherence.", + "bbox": [ + 212, + 507, + 785, + 672 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Qualitative Comparison on Video-to-4D In our investigation of the video-to-4D task, we have employed various methodologies to conduct qualitative comparisons. Figure 4 shows the rendered outcomes at distinct temporal intervals and perspectives. A critical observation of our findings reveals that Gaussian splatting is inadequate in producing credible and temporally consistent images during per-frame reconstruction. Furthermore, we have observed that Consistent4D often produces over-saturated and unrealistic patterns, which may stem from the over-restrictive geometry representation in their cascade Dynerf. Additionally, our analysis indicates that 4DGen struggles to generate reasonable motion for the Gaussians, leading to its points lacking deformation over time. Moreover, 4DGen fails to generate detailed textures for the surface, which appear", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "STAG4D", + "bbox": [ + 668, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/97944a9d0d08c11f6788a824ee1ef039af965363aefed34324797c2ce81367de.jpg", + "image_caption": [ + "Fig. 6: Ablation on spatial and temporal attention. We evaluate the effect of spatial and temporal attention on the reconstruction results." + ], + "image_footnote": [], + "bbox": [ + 220, + 146, + 493, + 340 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/4b3a9d7d1b1b49693c72fc53048b60ece35c286e205dce6cb43f93927a8cbeaa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 152, + 787, + 338 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "to be over-smoothed in some cases. Conversely, our methodology demonstrates substantial enhancements in comparison to existing strategies, particularly in aspects of reconstruction fidelity and stability. Notably, our technique facilitates seamless color transitions, as exemplified in the Fish case, and renders a convincing low-poly effect, as demonstrated in the Trump case, underscoring our method's adaptability to diverse dynamic environments.", + "bbox": [ + 212, + 429, + 787, + 518 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Qualitative Comparison on Text-to-4D We also evaluate our method on the Text&Image-to-4D task, which aims to generate 4D content from text and image inputs. Figure 5 shows the 4D generation results of our method and two baselines, 4Dfy [1] and DreamGaussian4D [23], under different views. As can be seen, 4Dfy [1] produces coarse results and fails to capture the complex semantics of the text input, e.g. in the Cat case. DreamGaussian4D [23] generates meshes with fine-grained details in the front view, but suffers from severe artifacts and distortions in the side view or back view. In contrast, our method synthesizes Gaussians with smooth and consistent geometry and realistic texture across different views, thanks to our attention mechanism that leverages spatial and temporal information. Our method demonstrates its superiority and versatility on the Text&Image-to-4D task, which can be seen as a direct application of our Video-to-4D framework.", + "bbox": [ + 212, + 523, + 789, + 720 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/621be71528fb35961b53ee3fe9764fa9a36b4794d80f029f0f9a23d9cbb256c7.jpg", + "table_caption": [ + "Table 1: Comparison with state-of-the-art methods." + ], + "table_footnote": [], + "table_body": "
CLIP ↑LPIPS ↓FID-VID ↓FVD ↓FVD-16 ↓
Consistent4D0.8770.134/1133.93/
4DGen0.8940.13071.99/1005.72
Ours0.9090.12652.58992.21952.41
", + "bbox": [ + 320, + 779, + 676, + 837 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Zeng et al.", + "bbox": [ + 271, + 114, + 344, + 128 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 146, + 388, + 161 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The experimental analysis presented in Table 2 provides a comprehensive evaluation of the impact of various components on the performance of a diffusion model with a 4D representation. The components under consideration include baseline diffusion, spatial anchor, spatial-temporal anchor, and adaptive densification. In the first configuration of Table 2, we regard the model utilizing the diffusion process [13] with adaptive densification as the baseline.", + "bbox": [ + 212, + 174, + 787, + 267 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/413b9323868bf125b22eff1e5fe43843061d12907047d46e9f525514c22ac813.jpg", + "table_caption": [ + "Table 2: Ablation study on effectiveness of each novel component." + ], + "table_footnote": [], + "table_body": "
Diffusion Model4D Gaussian \nAdaptive Dens.Evaluation Metrics
BaselineSpat.Temp.CLIP ↑LPIPS ↓FID-VID ↓FVD ↓
✓✓0.8950.13577.881369.65
✓✓✓0.8990.13067.041198.23
✓✓✓0.8900.13690.301453.53
✓✓✓✓0.9090.12652.58992.21
", + "bbox": [ + 215, + 333, + 542, + 402 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/0e2e05c78a84716fd4fa3e1901371869caaac689a3f60f86cfeaddda3670d794.jpg", + "table_caption": [ + "Table 3: Ablation study on different attention mechanisms." + ], + "table_footnote": [], + "table_body": "
Evaluation Metrics
CLIPLPIPSFID-VIDFVD
Spat. only0.8990.13067.041198.23
Cross-Frame0.9010.12963.321053.25
Spat.-Temp.0.9090.12652.58992.21
", + "bbox": [ + 553, + 333, + 787, + 395 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Spatial and Temporal Anchors Table 2 shows that adding the spatial anchor to the baseline model leads to a marginal improvement in all metrics, especially FID-VID and FVD. The final model, which uses both spatial-temporal anchor and adaptive densification, improves all metrics significantly. It has the best CLIP, LPIPS, FID-VID, and FVD scores, meaning it generates the most semantically consistent, realistic, and smooth videos.", + "bbox": [ + 212, + 430, + 782, + 518 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Figure 6 shows the importance of our spatial-temporal anchor for improving the 4D content quality. The baseline 4D Gaussian method, labeled as \"w./o. spatial,\" fails to deform the point clouds properly, leading to unreasonable shapes without spatial-temporal constraints. A model that uses only the spatial anchor, \"w./o. temporal,\" struggles to deform solid shapes, as seen in the rabbit's leg and the frog's back. Our method, which combines both spatial and temporal anchors, overcomes these issues and produces visually consistent and accurate 4D content.", + "bbox": [ + 212, + 522, + 787, + 641 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Adaptive Densification We present the statistical analysis of the adaptive densification module in the third and final configurations of Table 2. The results show that the adaptive densification module improves the overall quality of the generative model, especially in terms of the FID-VID and FVD metrics. We also compare the proposed adaptive densification with a fixed threshold approach in Figure 7b. We observe that a fixed threshold can lead to either under-densification or over-densification, depending on the case. For example, in the Bird case, the fixed threshold is too low for densification, which results in excessive Gaussian points. On the other hand, in the Duck case, the fixed threshold is too high for densification, which causes the foot to be blurry due to insufficient Gaussian points. Our method overcomes this limitation by adapting the gradient threshold to different cases, thus producing robust and stable 4D Gaussian results.", + "bbox": [ + 212, + 643, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "STAG4D", + "bbox": [ + 666, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/e1442b050157ece5ea308ebdf48253713e9f5707d1a4febf8799d5f677f773f4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 145, + 779, + 362 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Different Attention Mechanisms In our study, we evaluated three distinct attention mechanisms within the context of images generated by our diffusion model: spatial attention, cross-frame attention, and spatial-temporal attention. The visual outcomes of these varied approaches are illustrated in Figure 7a. We employed a parameter $\\lambda$ to modulate the extent of spatial and temporal attention exerted, as delineated in Equation 5. Our findings indicate that exclusive reliance on spatial attention does not guarantee consistency across sequential frames. While cross-frame attention enhances frame-to-frame coherence, it fails to achieve a satisfactory alignment with reference images. Conversely, the integration of spatial-temporal attention ensures consistency across both dimensions. We quantitatively assessed the quality of images generated under different attention regimes using four established metrics: CLIP, LPIPS, FID-VID, and FVD. The comparative results are tabulated in Table 3. It was observed that images generated with spatial-temporal attention outperformed others across all evaluation metrics, which underscores the efficacy of our proposed mechanism.", + "bbox": [ + 212, + 391, + 787, + 618 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 642, + 359, + 657 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The paper presents a novel approach for dynamic 3D content generation from monocular videos, addressing the challenges of 4D representation and spatial-temporal consistency. By leveraging specially tailored 4D Gaussian splatting and a novel information fusion module, the proposed method achieves high-quality and robust 4D scene generation. Comprehensive experiments demonstrate the method's effectiveness, showcasing a faster generation speed and significant improvements in rendering quality and temporal consistency compared to prior state-of-the-art methods. Overall, the proposed method sets a new benchmark for training speed, rendering quality, and 4D consistency in dynamic 3D content generation from monocular videos, opening up possibilities for real-world applications.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Zeng et al.", + "bbox": [ + 271, + 114, + 344, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported by the National Key R&D Program of China (2022YFF0902200) and NSFC grant 62441204.", + "bbox": [ + 215, + 172, + 727, + 204 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 224, + 323, + 239 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bahmani, S., Skorokhodov, I., Rong, V., Wetzstein, G., Guibas, L., Wonka, P., Tulyakov, S., Park, J.J., Tagliasacchi, A., Lindell, D.B.: 4d-fy: Text-to-4d generation using hybrid score distillation sampling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7996-8006 (2024)", + "2. Blattmann, A., Dockhorn, T., Kulal, S., Mendelevitch, D., Kilian, M., Lorenz, D., Levi, Y., English, Z., Voleti, V., Letts, A., Jampani, V., Rombach, R.: Stable video diffusion: Scaling latent video diffusion models to large datasets (2023)", + "3. Du, Y., Zhang, Y., Yu, H.X., Tenenbaum, J.B., Wu, J.: Neural radiance flow for 4d view synthesis and video processing. In: 2021 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 14304-14314. IEEE Computer Society (2021)", + "4. Gao, C., Saraf, A., Kopf, J., Huang, J.B.: Dynamic view synthesis from dynamic monocular video. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5712-5721 (2021)", + "5. Ho, J., Chan, W., Sahara, C., Whang, J., Gao, R., Gritsenko, A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., et al.:Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303 (2022)", + "6. Jiang, Y., Zhang, L., Gao, J., Hu, W., Yao, Y.: Consistent4d: Consistent 360 $\\{\\backslash \\mathrm{deg}\\}$ dynamic object generation from monocular video. arXiv preprint arXiv:2311.02848 (2023)", + "7. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG) 42(4), 1-14 (2023)", + "8. Khachatryan, L., Movsisyan, A., Tadevosyan, V., Henschel, R., Wang, Z., Navasardyan, S., Shi, H.: Text2video-zero: Text-to-image diffusion models are zero-shot video generators. arXiv preprint arXiv:2303.13439 (2023)", + "9. Li, Z., Niklaus, S., Snavely, N., Wang, O.: Neural scene flow fields for space-time view synthesis of dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2021)", + "0. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3d: High-resolution text-to-3d content creation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 300–309 (2023)", + "1. Lin, Y., Dai, Z., Zhu, S., Yao, Y.: Gaussian-flow: 4d reconstruction with dynamic 3d gaussian particle. arXiv:2312.03431 (2023)", + "2. Ling, H., Kim, S.W., Torralba, A., Fidler, S., Kreis, K.: Align your gaussians: Text-to-4d with dynamic 3d gaussians and composed diffusion models. arXiv preprint arXiv:2312.13763 (2023)", + "3. Liu, R., Wu, R., Van Hoorick, B., Tokmakov, P., Zakharov, S., Vondrick, C.: Zero-1-to-3: Zero-shot one image to 3d object. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9298-9309 (2023)", + "4. Liu, Y., Lin, C., Zeng, Z., Long, X., Liu, L., Komura, T., Wang, W.: Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453 (2023)" + ], + "bbox": [ + 225, + 253, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "STAG4D", + "bbox": [ + 666, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Long, X., Guo, Y.C., Lin, C., Liu, Y., Dou, Z., Liu, L., Ma, Y., Zhang, S.H., Habermann, M., Theobalt, C., et al.: Wonder3d: Single image to 3d using cross-domain diffusion. arXiv preprint arXiv:2310.15008 (2023)", + "16. Lu, Y., Zhang, J., Li, S., Fang, T., McKinnon, D., Tsin, Y., Quan, L., Cao, X., Yao, Y.: Direct2. 5: Diverse text-to-3d generation via multi-view 2.5 d diffusion. arXiv preprint arXiv:2311.15980 (2023)", + "17. Lu, Y., Zhang, J., Li, S., Fang, T., McKinnon, D., Tsin, Y., Quan, L., Cao, X., Yao, Y.: Direct2.5: Diverse text-to-3d generation via multi-view 2.5d diffusion (2023)", + "18. Luiten, J., Kopanas, G., Leibe, B., Ramanan, D.: Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713 (2023)", + "19. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)", + "20. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph. 41(4), 102:1-102:15 (Jul 2022). https://doi.org/10.1145/3528223.3530127, https://doi.org/10.1145/3528223.3530127", + "21. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis (2023)", + "22. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988 (2022)", + "23. Ren, J., Pan, L., Tang, J., Zhang, C., Cao, A., Zeng, G., Liu, Z.: Dreamgaussian4d: Generative 4d gaussian splatting. arXiv preprint arXiv:2312.17142 (2023)", + "24. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2022)", + "25. Shen, T., Gao, J., Yin, K., Liu, M.Y., Fidler, S.: Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. In: Advances in Neural Information Processing Systems (NeurIPS) (2021)", + "26. Shi, R., Chen, H., Zhang, Z., Liu, M., Xu, C., Wei, X., Chen, L., Zeng, C., Su, H.: Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110 (2023)", + "27. Shi, Y., Wang, P., Ye, J., Mai, L., Li, K., Yang, X.: Mvdream: Multi-view diffusion for 3d generation. arXiv:2308.16512 (2023)", + "28. Singer, U., Sheynin, S., Polyak, A., Ashual, O., Makarov, I., Kokkinos, F., Goyal, N., Vedaldi, A., Parikh, D., Johnson, J., et al.: Text-to-4d dynamic scene generation. arXiv preprint arXiv:2301.11280 (2023)", + "29. Tang, J., Ren, J., Zhou, H., Liu, Z., Zeng, G.: Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653 (2023)", + "30. Wang, H., Du, X., Li, J., Yeh, R.A., Shakhnarovich, G.: Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12619-12629 (2023)", + "31. Wang, J., Yuan, H., Chen, D., Zhang, Y., Wang, X., Zhang, S.: Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571 (2023)", + "32. Wang, X., Wang, Y., Ye, J., Wang, Z., Sun, F., Liu, P., Wang, L., Sun, K., Wang, X., He, B.: Animatabledreamer: Text-guided non-rigid 3d model generation and reconstruction with canonical score distillation. arXiv preprint arXiv:2312.03795 (2023)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Zeng et al.", + "bbox": [ + 271, + 114, + 344, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "33. Wu, G., Yi, T., Fang, J., Xie, L., Zhang, X., Wei, W., Liu, W., Tian, Q., Wang, X.: 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528 (2023)", + "34. Xian, W., Huang, J.B., Kopf, J., Kim, C.: Space-time neural irradiance fields for free-viewpoint video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9421-9431 (2021)", + "35. Yang, Z., Gao, X., Zhou, W., Jiao, S., Zhang, Y., Jin, X.: Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. arXiv preprint arXiv:2309.13101 (2023)", + "36. Yin, Y., Xu, D., Wang, Z., Zhao, Y., Wei, Y.: 4dgen: Grounded 4d content generation with spatial-temporal consistency. arXiv preprint arXiv:2312.17225 (2023)", + "37. Zhang, L.: Reference-only control (2023), https://github.com/Mikubill/sd-webui-controlnet/discussions/1236", + "38. Zhang, S., Wang, J., Zhang, Y., Zhao, K., Yuan, H., Qing, Z., Wang, X., Zhao, D., Zhou, J.: I2vgen-xl: High-quality image-to-video synthesis via cascaded diffusion models (2023)", + "39. Zhao, Y., Yan, Z., Xie, E., Hong, L., Li, Z., Lee, G.H.:Animate124: Animating one image to 4d dynamic scene. arXiv preprint arXiv:2311.14603 (2023)", + "40. Zheng, Y., Li, X., Nagano, K., Liu, S., Hilliges, O., De Mello, S.: A unified approach for text-and image-guided 4d scene generation. arXiv preprint arXiv:2311.16854 (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 438 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "STAG4D", + "bbox": [ + 666, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/STAG4D_ Spatial-Temporal Anchored Generative 4D Gaussians/51eb6cce-864e-4035-8901-6f032e5306da_model.json b/2024/STAG4D_ Spatial-Temporal Anchored Generative 4D Gaussians/51eb6cce-864e-4035-8901-6f032e5306da_model.json new file mode 100644 index 0000000000000000000000000000000000000000..dc39e2d1520913bb7b6fd8c23ef4c8e8384969c5 --- /dev/null +++ b/2024/STAG4D_ Spatial-Temporal Anchored Generative 4D Gaussians/51eb6cce-864e-4035-8901-6f032e5306da_model.json @@ -0,0 +1,2038 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.219, + 0.141, + 0.784, + 0.185 + ], + "angle": 0, + "content": "STAG4D: Spatial-Temporal Anchored Generative 4D Gaussians" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.212, + 0.786, + 0.244 + ], + "angle": 0, + "content": "Yifei Zeng\\(^{1}\\), Yanqin Jiang\\(^{2}\\), Siyu Zhu\\(^{3}\\), Yuanxun Lu\\(^{1}\\), Youtian Lin\\(^{1}\\), Hao Zhu\\(^{1}\\), Weiming Hu\\(^{2}\\), Xun Cao\\(^{1}\\), and Yao Yao\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.43, + 0.255, + 0.574, + 0.282 + ], + "angle": 0, + "content": "\\(^{1}\\) Nanjing University \n\\(^{2}\\) CASIA" + }, + { + "type": "text", + "bbox": [ + 0.435, + 0.283, + 0.568, + 0.297 + ], + "angle": 0, + "content": "3 Fudan University" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.333, + 0.744, + 0.653 + ], + "angle": 0, + "content": "Abstract. Recent progress in pre-trained diffusion models and 3D generation have spurred interest in 4D content creation. However, achieving high-fidelity 4D generation with spatial-temporal consistency remains a challenge. In this work, we propose STAG4D, a novel framework that combines pre-trained diffusion models with dynamic 3D Gaussian splatting for high-fidelity 4D generation. Drawing inspiration from 3D generation techniques, we utilize a multi-view diffusion model to initialize multiview images anchoring on the input video frames, where the video can be either real-world captured or generated by a video diffusion model. To ensure the temporal consistency of the multi-view sequence initialization, we introduce a simple yet effective fusion strategy to leverage the first frame as a temporal anchor in the self-attention computation. With the almost consistent multi-view sequences, we then apply the score distillation sampling to optimize the 4D Gaussian point cloud. The 4D Gaussian spatting is specially crafted for the generation task, where an adaptive densification strategy is proposed to mitigate the unstable Gaussian gradient for robust optimization. Notably, the proposed pipeline does not require any pre-training or fine-tuning of diffusion networks, offering a more accessible and practical solution for the 4D generation task. Extensive experiments demonstrate that our method outperforms prior 4D generation works in rendering quality, spatial-temporal consistency, and generation robustness, setting a new state-of-the-art for 4D generation from diverse inputs, including text, image, and video." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.664, + 0.736, + 0.679 + ], + "angle": 0, + "content": "Keywords: 4D Generation \\(\\cdot\\) 3D Gaussian Splatting \\(\\cdot\\) Diffusion Model" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.704, + 0.377, + 0.72 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Recent advancements in large-scale pre-trained diffusion models have shown remarkable progress in producing high-quality and diverse visual content, including images, videos, and 3D asserts [5,13,16,22,24,27,31,38]. The progress naturally extends to the realm of dynamic 3D generation [1,6,12,28,32,36,39,40]. This endeavor has gained prominence in computer vision and generative AI research, as high-quality 4D content generation is key to a broad range of applications such as autonomous driving simulation, game and film industries, digital Avatar" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.345, + 0.129 + ], + "angle": 0, + "content": "Zeng et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.146, + 0.516, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.524, + 0.146, + 0.788, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.282, + 0.788, + 0.311 + ], + "angle": 0, + "content": "Fig. 1: Visualization of the generated 4D asserts. Our approach can generate diverse 4D content from various inputs, including text, image, and video." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.349, + 0.784, + 0.41 + ], + "angle": 0, + "content": "creation, and spatial video production. However, previous 4D generation methods face challenges including blurry rendering, spatial-temporal inconsistency, and slow generation speed. Generating high-quality 4D content efficiently and practically remains a significant challenge." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.413, + 0.785, + 0.623 + ], + "angle": 0, + "content": "Given the promising strides in 3D generation from pre-trained diffusion models, there has been a concerted focus on generalized dynamic 3D generation from text or uncalibrated monocular video [1,6,12,28,32,40]. MAV3D [28] is the pioneering work of text-to-4D generation which creates 4D scenes from textual descriptions through Score Distillation Sampling (SDS) from a video diffusion model. Alternatively, Consistent4D [6] takes a video as input, and focuses on the task of video-to-4D generation. The pipeline is more flexible in that it can take advantage of high-quality input videos, dividing the difficult problem of realistic 4D content creation into a practical pipeline of text-to-video and video-to-4D generation. However, current video-to-4D generation methods still struggle to generate high-fidelity 4D asserts with spatial and temporal rendering consistency. Meanwhile, the commonly used Neural Radiance Field(NeRF) representation would suffer from an over-saturated appearance and a long optimization time." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.844 + ], + "angle": 0, + "content": "In this paper, we address two important factors in 4D content creation: a suitable 4D representation and spatial-temporal consistency imposed during pseudoprojection generation (i.e., multi-view video generation). Our approach employs a 4D Gaussian splatting approach specially tailored for the generation task. In the generation process, we first generate the multi-view images at all timestamps of the 4D scene. To enforce the 4D consistency of the multi-view video initialization, we design a direct fusion approach for attention computation in both spatial and temporal domains. This approach implicitly addresses texture degradation and geometric misalignment, thereby avoiding the need for adding an explicit multi-view or temporal consistency loss during the Score Distillation Sampling(SDS) [22] optimization. Then, the reference photometric loss and the SDS loss conditioned on the initial multi-view video are applied for 4D Gaussian point cloud optimization. To tackle the instability in optimizing the 4D Gaussian points, we introduce an adaptive densification strategy that is informed by the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.668, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "STAG4D" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "Gaussian gradient distribution. This approach yields high-quality and robust 4D scene generation from a monocular video input." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.785, + 0.283 + ], + "angle": 0, + "content": "We have conducted comprehensive experiments to demonstrate the effectiveness of each component of the proposed method. Our approach achieves a 2x faster generation speed compared with the previous video-to-4D approaches (e.g., [6]) and a significantly better generation quality than prior state-of-the-art methods. It is also noteworthy that the 4D content generated by our method can be rendered in real-time, opening up a wide range of possibilities for practical applications. In summary, our primary contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.298, + 0.784, + 0.341 + ], + "angle": 0, + "content": "- We introduce a holistic 4D generation pipeline that streamlines the generation process into sequential stages: video generation, multi-view video initialization, and 4D optimization using multi-view video conditioned SDS." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.344, + 0.784, + 0.387 + ], + "angle": 0, + "content": "- We harness the dynamic 3D Gaussian representation for 4D generation, complemented by an adaptive densification strategy. This combination enables highly precise and efficient 4D generation from monocular video inputs." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.39, + 0.784, + 0.432 + ], + "angle": 0, + "content": "- We present a novel training-free attention fusion module to effectively integrate temporal anchor frames into the multi-view diffusion process, significantly enhancing the 4D consistency of the generated multi-view videos." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.436, + 0.784, + 0.478 + ], + "angle": 0, + "content": "- Our method outperforms previous approaches in optimization efficiency, rendering quality, and 4D consistency, establishing a new benchmark for 4D generation across various input types, such as text, images, and videos." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.298, + 0.784, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.507, + 0.383, + 0.523 + ], + "angle": 0, + "content": "2 Related work" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.541, + 0.386, + 0.555 + ], + "angle": 0, + "content": "2.1 3D Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.785, + 0.704 + ], + "angle": 0, + "content": "Dreamfusion [22] first proposes the SDS loss [22, 30] for NeRF optimization, which stands for the most prevalent technique for nowadays text-to-3D approaches. To mitigate the multi-view Janus problem, Zero123 [13] and SycnDreammer [14] fine-tune 2D diffusion models to grant the image generator with the ability of viewpoint control, while later works [15, 17, 26, 27] explicitly generate fixed multi-view images in one diffusion pass. Building upon the success of multi-view diffusion methods, we integrate a multi-view generation module into the 4D generation task, with meticulous handling of 4D spatial-temporal consistency for enhanced 4D generation quality." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Except for equipping diffusion models with multi-view or depth awareness, another direction of progress focuses on 3D representations. Following the idea of differentiable optimization, Magic3D [10] applies a two-stage generation pipeline and changes the 3D representation to instant-NGP [20] and DMTet [25], achieving faster runtime and better generation quality. Direct2.5 [17] applies an explicit mesh representation and uses differentiable rasterization for fast mesh optimization. Recently, with the newly developed 3DGS, DreamGaussian [29] demonstrates the ability to generate 3D objects within several minutes by substituting NeRF with the 3D Gaussian representation, showing the potential of using an" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.345, + 0.129 + ], + "angle": 0, + "content": "Zeng et al." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.193 + ], + "angle": 0, + "content": "explicit point-based representation in 3D generation tasks. In this work, we introduce a novel 4D Gaussian representation and a specially tailored optimization scheme for the 4D generation task." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.217, + 0.416, + 0.232 + ], + "angle": 0, + "content": "2.2 4D Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.244, + 0.788, + 0.518 + ], + "angle": 0, + "content": "Dynamic 3D reconstruction is yet another heated research topic in the field of computer vision and graphics. By extending the static NeRF [19] framework to dynamic scenes, dynamic NeRFs [3, 4, 9, 34] have demonstrated remarkable progress on dynamic 3D reconstruction. However, limited by the implicit neural representation and the complex nature of dynamic 3D information, dynamic NeRFs still suffer from slow optimization speed and low reconstruction quality. With the development of 3DGS [7], researchers have quickly extended the 3D Gaussian to dynamic scene representation, which achieves faster training and rendering speeds compared to D-NeRFs based on implicit neural representations. Dynamic 3D Gaussian [18] applies the per-frame 3DGS optimization for 4D scene reconstruction. Some studies [33, 35] attempt to amalgamate the explicit point-based 3DGS with an implicit neural field for dynamic information modeling, while Gaussian-Flow [11] introduces an explicit per-point motion model to represent a 4D scene without using implicit neural networks. It is noteworthy that it is non-trivial to extend the Gaussian reconstruction to 4D reconstruction as each reconstruction pipeline requires heuristic tuning of hyperparameters at each stage. In this paper, we target combining the 4DGS for the challenging task of 4D generation." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.541, + 0.384, + 0.556 + ], + "angle": 0, + "content": "2.3 4D Generation" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.569, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Recent developments in 3D content generation and video diffusion technologies have triggered researchers' interest in exploring 4D content generation from various input conditionings. MAV3D [28] presents the early attempt at text-to-4D generation. By utilizing score distillation sampling derived from video diffusion models, MAV3D optimizes a dynamic NeRF based on textual prompts. In a parallel vein, Consistent4D [6] has introduced the task of video-to-4D. This method capitalizes on the pre-trained knowledge from image diffusion models to optimize dynamic NeRFs via SDS optimization, showcasing the potential of high-quality 4D content generation from pre-trained 2D diffusion models. Concurrently with our work, 4DGen [36] has introduced a framework for generating dynamic 3D models, which employs spatial-temporal pseudo labels on keyframes within a multi-view diffusion model. However, the overall quality of its generation can be further improved, as its rendering fidelity is limited by the implicit neural representation and inadequate spatial-temporal information exchange exists in the diffusion generation process. In contrast, our method applies an efficient 4D Gaussian representation with spatial-temporal attention from reference- and the first-frame anchors, achieving substantially better generation results than previous works." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.669, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "STAG4D" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.147, + 0.782, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.357, + 0.788, + 0.427 + ], + "angle": 0, + "content": "Fig. 2: Overall pipeline. Given a video input, we apply a multi-view diffusion model to produce coherent multi-view sequences, which serve as spatial and temporal anchors. Next, we train a deformable 3D Gaussian using multi-view SDS loss and reference loss. For a text-to-4D generation, our pipeline can be naturally extended to accept the text input by integrating with an off-the-shelf text-to-video module." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.454, + 0.33, + 0.469 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.485, + 0.789, + 0.668 + ], + "angle": 0, + "content": "The framework of the proposed approach, as depicted in Figure 2, outlines the generation and optimization process of a dynamic scene from a given video input. Also, our approach can be easily extended to a text-to-4D generation pipeline by utilizing an off-the-shelf video generation module. Section 3.1 presents the 4D representation used in the proposed method, along with an adaptive densification strategy that adjusts the densification threshold based on the relative motion gradient of the scene points. Subsequently, Section 3.2 provides a brief overview of the multi-view diffusion model used in the pipeline and introduces a novel spatial-temporal attention fusion mechanism designed to produce the almost consistent multi-view image sequences. Finally, as detailed in Section 3.3, the generated multi-view sequences are leveraged with spatial and temporal anchor frames for consistent 4D generation with the SDS optimization." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.688, + 0.419, + 0.703 + ], + "angle": 0, + "content": "3.1 4D Representation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.712, + 0.788, + 0.773 + ], + "angle": 0, + "content": "4D Gaussian Splatting The concept of 3D Gaussian Splatting was initially introduced in the work of [7], where an explicit point-based representation is used to model the 3D scene. Specifically, the 3D Gaussian point cloud, denoted as \\(\\mathcal{S}\\) for a static scene, is characterized by the tuple:" + }, + { + "type": "equation", + "bbox": [ + 0.437, + 0.784, + 0.786, + 0.8 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} = [ \\mathcal {X}, s, r, \\sigma , \\zeta ], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "where \\(\\mathcal{X} = (x,y,z)\\) denotes the positions of the 3D Gaussian points, and \\(s\\), \\(r\\), \\(\\sigma\\) and \\(\\zeta\\) represent the scale, rotation, opacity, and spherical harmonics (SH) coeffi-" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.345, + 0.129 + ], + "angle": 0, + "content": "Zeng et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "cients of the radiance, respectively. In 3D reconstruction, the 3D Gaussian point cloud will be optimized through a point-based differentiable volume rendering." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.785, + 0.239 + ], + "angle": 0, + "content": "This representation can be naturally extended to 4D scenes by presenting the continuous scene dynamics as a 3D motion field. Inspired by [33], we present a 4D scene as a 3D Gaussian point cloud with a hex-plane-based deformation field, denoted as \\(\\mathcal{F}(\\mathcal{S},t)\\). The 3D Gaussian point cloud at time \\(t\\) can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.414, + 0.255, + 0.785, + 0.271 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} (\\mathcal {S}, t) = \\left[ \\mathcal {X} _ {t}, s _ {t}, r _ {t}, \\sigma , \\zeta \\right], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.284, + 0.784, + 0.314 + ], + "angle": 0, + "content": "where \\(\\mathcal{X}_t = (x_t, y_t, z_t)\\), \\(s_t\\) and \\(r_t\\) represent the updated information of Gaussian position, scale, and rotation at time \\(t\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.342, + 0.785, + 0.569 + ], + "angle": 0, + "content": "Adaptive Densification The standard 3D Gaussian Splatting technique typically employs a point cloud densification control strategy to dynamically adjust the number of Gaussians and their density within a unit volume. This adaptive approach allows for the transition from an initial sparse Gaussian set to a denser configuration to better represent the 3D scene. The 4D Gaussian Splatting method introduced in the work [33] utilizes a densification strategy similar to that presented in vanilla 3DGS [7], which involves the use of a fixed densification threshold based on the view-space position gradient. However, while the fixed gradient threshold strategy demonstrates efficacy in the context of reconstruction, particularly when multiple views afford robust and redundant coverage of the target scene, it performs sub-optimally in the generative setting. This limitation arises from the constraints imposed by the input single image or monocular videos, leading to significant uncertainty in the spatial and scale dimensions for each training object. This will lead to different optimal thresholds for different cases, as illustrated in the ablation study." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.57, + 0.786, + 0.736 + ], + "angle": 0, + "content": "To address the issue, we propose an adaptive threshold approach, wherein only candidates with relatively large gradients are selected for densification. Our approach is based on the statistical analysis of the accumulated gradient of each point, which follows a log distribution of similar shapes throughout the training process. We apply an adaptive threshold that filters out Gaussians with a relatively small gradient and only densifies those with a large gradient. The adaptive threshold is set to select a fixed percentage (top \\(\\lambda\\%\\)) of points with the highest gradient in each densification operation. This ensures that the threshold adapts to the distribution of the gradient and maintains a stable relative position. Experimental results demonstrate that the proposed simple strategy can significantly enhance the quality and robustness of the 4D generation." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.764, + 0.646, + 0.779 + ], + "angle": 0, + "content": "3.2 Temporal and Multi-view Consistent Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.784, + 0.842 + ], + "angle": 0, + "content": "Multi-view Consistent Diffusion Our pipeline adopts Score Distillation Sampling(SDS) [22] from image diffusion models for the optimization of dynamic Gaussian. Specifically, Image-to-image diffusion model Zero123 [13] is utilized" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.668, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "STAG4D" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.632, + 0.162 + ], + "angle": 0, + "content": "and the SDS loss gradient could be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.307, + 0.171, + 0.785, + 0.205 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\theta} \\mathcal {L} _ {S D S} (\\phi , \\mathbf {x}) = \\mathbb {E} _ {t, \\epsilon} \\left[ \\omega (t) \\left(\\hat {\\epsilon} _ {\\theta} \\left(\\mathbf {z} _ {t}; \\mathbf {I} _ {i n}, \\mathbf {R}, \\mathbf {T}, t\\right) - \\epsilon\\right) \\frac {\\partial \\mathbf {x}}{\\partial \\theta} \\right], \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.399, + 0.207, + 0.585, + 0.222 + ], + "angle": 0, + "content": "\\[\n\\hat {z} = z _ {t} - \\sigma_ {t} \\hat {\\epsilon} _ {t} (z; \\mathbf {I} _ {i n}, \\mathbf {R}, \\mathbf {T}).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.231, + 0.788, + 0.428 + ], + "angle": 0, + "content": "where \\(\\theta\\) represents the parameters of the 3D representation, \\(\\mathbf{x}\\) the rendered image at the current view, \\(t\\) the timestamp in the diffusion process, \\(\\epsilon\\) the ground truth noise, \\(\\hat{\\epsilon}\\) the predicted noise from the noisy image \\(\\mathbf{z}_{\\mathbf{t}}\\) conditioned on an initial input \\(\\mathbf{I}_{in}\\), and the relative camera pose between input view and target view \\((\\mathbf{R},\\mathbf{T})\\). Zero123 can generate the target view image at any relative camera location but only one target view at a time, thus good at optimizing the 3D object from all views whilst bad at generating spatially consistent images from multiple target views. In contrast, Zero123++ [26] leverages reference attention [37] to model the relationship of the images from multiple target views as well as the input view, resulting in multi-view consistency output yet at the cost of fixed target camera locations. We combine the advantages of both models by taking the multi-view consistent output of Zero123++ as the input images of Zero123 when calculating SDS loss." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.788, + 0.568 + ], + "angle": 0, + "content": "Temporally Consistent Diffusion Following the view generation philosophy of multi-view diffusion based 3D generation, a practical solution to 4D content creation is to generate the multi-view videos of the 4D scene, and then apply the multi-view video conditioned SDS loss to optimize the scene. However, it is rather difficult to generate temporally consistent multi-view videos by the separate per-frame generation. Inspired by zero-shot video generation method [8], we propose a training-free temporal attention module to enable Zero123++ [26] with temporal-awareness." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.788, + 0.645 + ], + "angle": 0, + "content": "In our design, we obtain the attention information during denoising the multiview images of the first frame. Then we apply the recorded attention to the later denoising process. Specifically, during the denoising process of the latent code, the self-attention layer computes three essential components—queries \\((Q)\\), keys \\((K)\\), and values \\((V)\\), which is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.353, + 0.654, + 0.785, + 0.687 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S e l f - A t t n} (Q, K, V) = \\operatorname {S o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {c}}\\right) V. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.694, + 0.788, + 0.755 + ], + "angle": 0, + "content": "We first compute the key \\( K_{0} \\) and value \\( V_{0} \\) for the initial frame, \\( T_{0} \\). Subsequently, for each subsequent frame \\( T_{t} \\), where \\( t \\in \\{1,\\dots,N\\} \\), we perform a mixing operation on the key \\( K_{t} \\) and value \\( V_{t} \\) with the initial key \\( K_{0} \\) and value \\( V_{0} \\) derived from the first frame \\( T_{0} \\). This process can be formally represented as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.407, + 0.764, + 0.785, + 0.803 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} K _ {t} = \\gamma \\mathbf {K} _ {0} + (1 - \\gamma) \\mathbf {K} _ {t} \\\\ V _ {t} = \\gamma \\mathbf {V} _ {0} + (1 - \\gamma) \\mathbf {V} _ {t}, \\end{array} \\right. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.841 + ], + "angle": 0, + "content": "where the parameter \\(\\gamma\\) serves as a weighting factor that governs the influence of the initial frame, as illustrated in the ablation part. Additionally, we employ" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.345, + 0.129 + ], + "angle": 0, + "content": "Zeng et al." + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.149, + 0.778, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.361, + 0.785, + 0.39 + ], + "angle": 0, + "content": "Fig. 3: Illustration of the proposed spatial and temporal attention fusion during multiview sequence inference." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.42, + 0.788, + 0.527 + ], + "angle": 0, + "content": "reference attention following the approach outlined in [26,37], to extract local conditioning from the input image. In this reference attention mechanism, both the key and the value are obtained through the concatenation of features from the reference image and the noisy input image. Our design of temporal attention module is simple yet effective with superior temporal consistency, image quality and 3D consistency to vanilla cross-frame attention proposed in [8], as demonstrated in experiment section." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.551, + 0.426, + 0.567 + ], + "angle": 0, + "content": "3.3 Training Objectives" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.579, + 0.788, + 0.673 + ], + "angle": 0, + "content": "Following the generation of multi-view sequences from a monocular reference video, we can acquire 6 anchor views \\(\\{I_t^i\\}_{i\\in \\{1\\dots 6\\}}\\) and a reference view \\(I_{t}^{ref}\\) at each timestep \\(t\\). In our optimization process, we employ multi-view score distillation sampling (SDS) utilizing the generated images \\(\\{I_t^i\\}_{i = 1\\dots 6}\\) in conjunction with the reference images \\(I_{t}^{ref}\\). The multi-view score distillation loss function \\(\\mathcal{L}_{MVSDS}\\) can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.334, + 0.682, + 0.785, + 0.723 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {M V S D S} = \\lambda_ {1} \\mathcal {L} _ {S D S} ^ {i} + \\lambda_ {2} \\mathcal {L} _ {S D S} ^ {r e f} \\\\ = \\lambda_ {1} \\mathcal {L} _ {S D S} \\left(\\phi , I _ {t} ^ {i}\\right) + \\lambda_ {2} \\mathcal {L} _ {S D S} \\left(\\phi , I _ {t} ^ {\\text {r e f}}\\right), \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.809 + ], + "angle": 0, + "content": "where \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) are two weighting factors, the index \\(i\\) is determined based on the proximity of the rendering viewpoint to the viewpoint of the generated images. This selection process, which we refer to as multi-view score distillation sampling, involves choosing the nearest reference image to the rendered camera view for calculating the SDS loss." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Following the approach [29], we utilize the reference image to compute both the reconstruction loss \\(\\mathcal{L}_{rec}\\) and the foreground mask loss \\(\\mathcal{L}_{mask}\\). Thus, the final" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.668, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "STAG4D" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.403, + 0.162 + ], + "angle": 0, + "content": "optimization objective is:" + }, + { + "type": "equation", + "bbox": [ + 0.381, + 0.179, + 0.786, + 0.194 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {M V S D S} + \\lambda_ {3} \\mathcal {L} _ {r e c} + \\lambda_ {4} \\mathcal {L} _ {\\text {m a s k}}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.209, + 0.787, + 0.255 + ], + "angle": 0, + "content": "where \\(\\lambda_3\\) and \\(\\lambda_4\\) are the weighting parameters. During training, we first use \\(\\mathcal{L}\\) to supervise a fixed frame to get static canonical 3D Gaussian. Then we use all anchors and reference images to train the dynamic 4D Gaussian point cloud." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.281, + 0.442, + 0.295 + ], + "angle": 0, + "content": "3.4 Text-to-4D Extension" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.31, + 0.789, + 0.462 + ], + "angle": 0, + "content": "Our method is designed to be readily adaptable to text and image inputs, offering novel capabilities such as directly generating video sequences from textual descriptions or static images. We initiate this process using a 2D diffusion model SDXL [21], which generates the image from textual input. This 2D output is then transformed into a video sequence via a video diffusion model, such as SVD [2], adding temporal coherence and motion to the static image. Finally, the aforementioned pipeline further lifts the video sequence to a 4D scene, introducing an extra spatial dimension that creates immersive experiences beyond conventional 2D or 3D media. The generation pipeline with the given text, image, and video inputs is visualized in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.489, + 0.367, + 0.506 + ], + "angle": 0, + "content": "4 Experiment" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.524, + 0.414, + 0.54 + ], + "angle": 0, + "content": "4.1 Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.553, + 0.788, + 0.673 + ], + "angle": 0, + "content": "Dataset For the video-to-4D task, we utilize the dataset provided by Consistent4D [6] for quantitative evaluation, which comprises multi-view videos depicting 7 dynamic objects. Additionally, for qualitative evaluation, we curate a set of challenging videos from online sources to assess the robustness and generalization capabilities of each method. For text/image-to-4D tasks, we follow the data settings in 4Dfy [1] and DreamGaussian4D [23] to create the corresponding results. Notably, we use the same input image and input video with DreamGaussian4D for a fair comparison." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.674, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Evaluation metrics We employ CLIP, LPIPS, and FVD for Video-to-4D evaluation as in the previous approach [6]. Specifically, CLIP and LPIPS serve as image-level metrics, assessing the semantic similarity between rendered images and ground truths; FVD is a video-level metric commonly used in video generation tasks, considering not only single-frame quality but also temporal coherence; we also consider the FID-VID metric to measure the temporal consistency. As 4D-Gen [36] does not support a 30-frame setting we report its FVD under a 16-frame reconstruction, namely FVD-16 as an additional metrics. For text-to-4D generation, we follow previous works and provide a user study for quantitative comparisons between different methods. We also provide a user study for video-to-4D and both studies are detailed in the supplementary material." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.345, + 0.128 + ], + "angle": 0, + "content": "Zeng et al." + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.145, + 0.785, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.343, + 0.788, + 0.371 + ], + "angle": 0, + "content": "Fig. 4: Qualitative comparison on video-to-4D generation. Our method achieves better faithfulness to the input video with higher quality on the overall reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.406, + 0.784, + 0.498 + ], + "angle": 0, + "content": "Baselines For video-to-4D generation, we compare our method with two baselines: Consistent4D [6] and the concurrent work 4DGen [36]. We utilize the official code released by the authors to generate comparison results. For text&image-to-4D generation, we make comparisons with two state-of-art models that are open-sourced: 4Dfy [1] and DreamGaussian4D [23]. Their results are generated using codes from their official GitHub repository." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.531, + 0.456, + 0.546 + ], + "angle": 0, + "content": "4.2 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In the training of the video-to-4D generative model, we adopt a two-stage approach. Initially, the model is initialized in canonical space and trained for 1000 steps. Subsequently, the deformation fields are learned over 7000 additional steps to accurately capture the dynamic scene. For the deformation process, we employ multi-layer perceptrons (MLPs) with 64 hidden layers and 32 hidden features per layer. The initial learning rate for the deformation MLPs is set to \\(1.6 \\times 10^{-4}\\) and is decayed to \\(1.6 \\times 10^{-6}\\) by the end of the training. In the context of adaptive densification, we opt to densify the top \\(2.5\\%\\) of points with the most accumulated gradient. In the generative phase, we utilize the spatial-temporal consistent videos for the multiview score distillation sampling optimization. Empirically, we set \\(\\lambda\\) to 0.5 for our temporally consistent diffusion. Regarding the loss functions, we maintain the SDS loss weight at 1 and adjust the weights of other losses accordingly. Specifically, the reconstruction loss is assigned to \\(4 \\times 10^{4}\\), the mask loss is weighted at \\(1 \\times 10^{4}\\). The training process requires approximately 1 hour on an RTX 3090 GPU, and the rendering process can be performed at 150 FPS in real time. For Text&Image-to-4D, we use SDXL [21] for image generation, and apply SVD [2] to the generated image to create a corresponding video. Then we use the same setting as Video-to-4D to transfer the video into 4D content." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.669, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "STAG4D" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.147, + 0.782, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.368, + 0.788, + 0.397 + ], + "angle": 0, + "content": "Fig. 5: Qualitative comparison on text-to-4D generation. Our method achieves the best results on the text alignment and visual quality." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.425, + 0.64, + 0.44 + ], + "angle": 0, + "content": "4.3 Comparison with State-of-the-art Approaches" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.785, + 0.508 + ], + "angle": 0, + "content": "In this section, we conduct a comprehensive comparison with the aforementioned baselines using synthetic and in-the-wild data. The superior performance demonstrated in both quantitative and qualitative evaluations serves to substantiate the effectiveness and robustness of our proposed method." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.787, + 0.674 + ], + "angle": 0, + "content": "Quantitative Results on Video-to-4D The experimental results presented in Table 1 offer a comparative analysis of our method against state-of-the-art techniques, namely Consistent4D and 4DGen, across several metrics. Our method outperforms Consistent4D and 4DGen in CLIP, showing better semantic consistency with the target content. It also achieves the lowest LPIPS score, meaning it generates more realistic images than the others. For the video quality and smoothness, our method beats 4DGen in FID-VID and FVD-16, and Consistent4D in FVD, implying that our videos are closer to real videos and have less temporal artifacts. Overall, the numbers confirm that our method is superior in generating semantically aligned and perceptually convincing videos with high fidelity and coherence." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Qualitative Comparison on Video-to-4D In our investigation of the video-to-4D task, we have employed various methodologies to conduct qualitative comparisons. Figure 4 shows the rendered outcomes at distinct temporal intervals and perspectives. A critical observation of our findings reveals that Gaussian splatting is inadequate in producing credible and temporally consistent images during per-frame reconstruction. Furthermore, we have observed that Consistent4D often produces over-saturated and unrealistic patterns, which may stem from the over-restrictive geometry representation in their cascade Dynerf. Additionally, our analysis indicates that 4DGen struggles to generate reasonable motion for the Gaussians, leading to its points lacking deformation over time. Moreover, 4DGen fails to generate detailed textures for the surface, which appear" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.345, + 0.129 + ], + "angle": 0, + "content": "Zeng et al." + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.147, + 0.495, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.153, + 0.788, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.362, + 0.789, + 0.391 + ], + "angle": 0, + "content": "Fig. 6: Ablation on spatial and temporal attention. We evaluate the effect of spatial and temporal attention on the reconstruction results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.789, + 0.52 + ], + "angle": 0, + "content": "to be over-smoothed in some cases. Conversely, our methodology demonstrates substantial enhancements in comparison to existing strategies, particularly in aspects of reconstruction fidelity and stability. Notably, our technique facilitates seamless color transitions, as exemplified in the Fish case, and renders a convincing low-poly effect, as demonstrated in the Trump case, underscoring our method's adaptability to diverse dynamic environments." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.524, + 0.79, + 0.721 + ], + "angle": 0, + "content": "Qualitative Comparison on Text-to-4D We also evaluate our method on the Text&Image-to-4D task, which aims to generate 4D content from text and image inputs. Figure 5 shows the 4D generation results of our method and two baselines, 4Dfy [1] and DreamGaussian4D [23], under different views. As can be seen, 4Dfy [1] produces coarse results and fails to capture the complex semantics of the text input, e.g. in the Cat case. DreamGaussian4D [23] generates meshes with fine-grained details in the front view, but suffers from severe artifacts and distortions in the side view or back view. In contrast, our method synthesizes Gaussians with smooth and consistent geometry and realistic texture across different views, thanks to our attention mechanism that leverages spatial and temporal information. Our method demonstrates its superiority and versatility on the Text&Image-to-4D task, which can be seen as a direct application of our Video-to-4D framework." + }, + { + "type": "table_caption", + "bbox": [ + 0.321, + 0.754, + 0.677, + 0.768 + ], + "angle": 0, + "content": "Table 1: Comparison with state-of-the-art methods." + }, + { + "type": "table", + "bbox": [ + 0.321, + 0.78, + 0.677, + 0.838 + ], + "angle": 0, + "content": "
CLIP ↑LPIPS ↓FID-VID ↓FVD ↓FVD-16 ↓
Consistent4D0.8770.134/1133.93/
4DGen0.8940.13071.99/1005.72
Ours0.9090.12652.58992.21952.41
" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.668, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "STAG4D" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.39, + 0.162 + ], + "angle": 0, + "content": "4.4 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.268 + ], + "angle": 0, + "content": "The experimental analysis presented in Table 2 provides a comprehensive evaluation of the impact of various components on the performance of a diffusion model with a 4D representation. The components under consideration include baseline diffusion, spatial anchor, spatial-temporal anchor, and adaptive densification. In the first configuration of Table 2, we regard the model utilizing the diffusion process [13] with adaptive densification as the baseline." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.294, + 0.544, + 0.322 + ], + "angle": 0, + "content": "Table 2: Ablation study on effectiveness of each novel component." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.334, + 0.544, + 0.403 + ], + "angle": 0, + "content": "
Diffusion Model4D Gaussian \nAdaptive Dens.Evaluation Metrics
BaselineSpat.Temp.CLIP ↑LPIPS ↓FID-VID ↓FVD ↓
✓✓0.8950.13577.881369.65
✓✓✓0.8990.13067.041198.23
✓✓✓0.8900.13690.301453.53
✓✓✓✓0.9090.12652.58992.21
" + }, + { + "type": "table_caption", + "bbox": [ + 0.555, + 0.294, + 0.786, + 0.321 + ], + "angle": 0, + "content": "Table 3: Ablation study on different attention mechanisms." + }, + { + "type": "table", + "bbox": [ + 0.555, + 0.334, + 0.788, + 0.396 + ], + "angle": 0, + "content": "
Evaluation Metrics
CLIPLPIPSFID-VIDFVD
Spat. only0.8990.13067.041198.23
Cross-Frame0.9010.12963.321053.25
Spat.-Temp.0.9090.12652.58992.21
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.431, + 0.784, + 0.52 + ], + "angle": 0, + "content": "Spatial and Temporal Anchors Table 2 shows that adding the spatial anchor to the baseline model leads to a marginal improvement in all metrics, especially FID-VID and FVD. The final model, which uses both spatial-temporal anchor and adaptive densification, improves all metrics significantly. It has the best CLIP, LPIPS, FID-VID, and FVD scores, meaning it generates the most semantically consistent, realistic, and smooth videos." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.788, + 0.642 + ], + "angle": 0, + "content": "Figure 6 shows the importance of our spatial-temporal anchor for improving the 4D content quality. The baseline 4D Gaussian method, labeled as \"w./o. spatial,\" fails to deform the point clouds properly, leading to unreasonable shapes without spatial-temporal constraints. A model that uses only the spatial anchor, \"w./o. temporal,\" struggles to deform solid shapes, as seen in the rabbit's leg and the frog's back. Our method, which combines both spatial and temporal anchors, overcomes these issues and produces visually consistent and accurate 4D content." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.789, + 0.84 + ], + "angle": 0, + "content": "Adaptive Densification We present the statistical analysis of the adaptive densification module in the third and final configurations of Table 2. The results show that the adaptive densification module improves the overall quality of the generative model, especially in terms of the FID-VID and FVD metrics. We also compare the proposed adaptive densification with a fixed threshold approach in Figure 7b. We observe that a fixed threshold can lead to either under-densification or over-densification, depending on the case. For example, in the Bird case, the fixed threshold is too low for densification, which results in excessive Gaussian points. On the other hand, in the Duck case, the fixed threshold is too high for densification, which causes the foot to be blurry due to insufficient Gaussian points. Our method overcomes this limitation by adapting the gradient threshold to different cases, thus producing robust and stable 4D Gaussian results." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.345, + 0.128 + ], + "angle": 0, + "content": "Zeng et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.146, + 0.781, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.392, + 0.788, + 0.619 + ], + "angle": 0, + "content": "Different Attention Mechanisms In our study, we evaluated three distinct attention mechanisms within the context of images generated by our diffusion model: spatial attention, cross-frame attention, and spatial-temporal attention. The visual outcomes of these varied approaches are illustrated in Figure 7a. We employed a parameter \\(\\lambda\\) to modulate the extent of spatial and temporal attention exerted, as delineated in Equation 5. Our findings indicate that exclusive reliance on spatial attention does not guarantee consistency across sequential frames. While cross-frame attention enhances frame-to-frame coherence, it fails to achieve a satisfactory alignment with reference images. Conversely, the integration of spatial-temporal attention ensures consistency across both dimensions. We quantitatively assessed the quality of images generated under different attention regimes using four established metrics: CLIP, LPIPS, FID-VID, and FVD. The comparative results are tabulated in Table 3. It was observed that images generated with spatial-temporal attention outperformed others across all evaluation metrics, which underscores the efficacy of our proposed mechanism." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.643, + 0.36, + 0.658 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The paper presents a novel approach for dynamic 3D content generation from monocular videos, addressing the challenges of 4D representation and spatial-temporal consistency. By leveraging specially tailored 4D Gaussian splatting and a novel information fusion module, the proposed method achieves high-quality and robust 4D scene generation. Comprehensive experiments demonstrate the method's effectiveness, showcasing a faster generation speed and significant improvements in rendering quality and temporal consistency compared to prior state-of-the-art methods. Overall, the proposed method sets a new benchmark for training speed, rendering quality, and 4D consistency in dynamic 3D content generation from monocular videos, opening up possibilities for real-world applications." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.668, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "STAG4D" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.174, + 0.728, + 0.205 + ], + "angle": 0, + "content": "This work was supported by the National Key R&D Program of China (2022YFF0902200) and NSFC grant 62441204." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.225, + 0.324, + 0.241 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.254, + 0.786, + 0.308 + ], + "angle": 0, + "content": "1. Bahmani, S., Skorokhodov, I., Rong, V., Wetzstein, G., Guibas, L., Wonka, P., Tulyakov, S., Park, J.J., Tagliasacchi, A., Lindell, D.B.: 4d-fy: Text-to-4d generation using hybrid score distillation sampling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7996-8006 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.31, + 0.786, + 0.35 + ], + "angle": 0, + "content": "2. Blattmann, A., Dockhorn, T., Kulal, S., Mendelevitch, D., Kilian, M., Lorenz, D., Levi, Y., English, Z., Voleti, V., Letts, A., Jampani, V., Rombach, R.: Stable video diffusion: Scaling latent video diffusion models to large datasets (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.35, + 0.786, + 0.391 + ], + "angle": 0, + "content": "3. Du, Y., Zhang, Y., Yu, H.X., Tenenbaum, J.B., Wu, J.: Neural radiance flow for 4d view synthesis and video processing. In: 2021 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 14304-14314. IEEE Computer Society (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.391, + 0.786, + 0.432 + ], + "angle": 0, + "content": "4. Gao, C., Saraf, A., Kopf, J., Huang, J.B.: Dynamic view synthesis from dynamic monocular video. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5712-5721 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.432, + 0.786, + 0.472 + ], + "angle": 0, + "content": "5. Ho, J., Chan, W., Sahara, C., Whang, J., Gao, R., Gritsenko, A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., et al.:Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.473, + 0.786, + 0.514 + ], + "angle": 0, + "content": "6. Jiang, Y., Zhang, L., Gao, J., Hu, W., Yao, Y.: Consistent4d: Consistent 360 \\(\\{\\backslash \\mathrm{deg}\\}\\) dynamic object generation from monocular video. arXiv preprint arXiv:2311.02848 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.514, + 0.786, + 0.554 + ], + "angle": 0, + "content": "7. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG) 42(4), 1-14 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.554, + 0.786, + 0.595 + ], + "angle": 0, + "content": "8. Khachatryan, L., Movsisyan, A., Tadevosyan, V., Henschel, R., Wang, Z., Navasardyan, S., Shi, H.: Text2video-zero: Text-to-image diffusion models are zero-shot video generators. arXiv preprint arXiv:2303.13439 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.595, + 0.786, + 0.636 + ], + "angle": 0, + "content": "9. Li, Z., Niklaus, S., Snavely, N., Wang, O.: Neural scene flow fields for space-time view synthesis of dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.636, + 0.786, + 0.69 + ], + "angle": 0, + "content": "0. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3d: High-resolution text-to-3d content creation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 300–309 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.691, + 0.786, + 0.718 + ], + "angle": 0, + "content": "1. Lin, Y., Dai, Z., Zhu, S., Yao, Y.: Gaussian-flow: 4d reconstruction with dynamic 3d gaussian particle. arXiv:2312.03431 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.718, + 0.786, + 0.758 + ], + "angle": 0, + "content": "2. Ling, H., Kim, S.W., Torralba, A., Fidler, S., Kreis, K.: Align your gaussians: Text-to-4d with dynamic 3d gaussians and composed diffusion models. arXiv preprint arXiv:2312.13763 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.759, + 0.786, + 0.8 + ], + "angle": 0, + "content": "3. Liu, R., Wu, R., Van Hoorick, B., Tokmakov, P., Zakharov, S., Vondrick, C.: Zero-1-to-3: Zero-shot one image to 3d object. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9298-9309 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.786, + 0.84 + ], + "angle": 0, + "content": "4. Liu, Y., Lin, C., Zeng, Z., Long, X., Liu, L., Komura, T., Wang, W.: Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.254, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.345, + 0.129 + ], + "angle": 0, + "content": "Zeng et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "15. Long, X., Guo, Y.C., Lin, C., Liu, Y., Dou, Z., Liu, L., Ma, Y., Zhang, S.H., Habermann, M., Theobalt, C., et al.: Wonder3d: Single image to 3d using cross-domain diffusion. arXiv preprint arXiv:2310.15008 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.232 + ], + "angle": 0, + "content": "16. Lu, Y., Zhang, J., Li, S., Fang, T., McKinnon, D., Tsin, Y., Quan, L., Cao, X., Yao, Y.: Direct2. 5: Diverse text-to-3d generation via multi-view 2.5 d diffusion. arXiv preprint arXiv:2311.15980 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.259 + ], + "angle": 0, + "content": "17. Lu, Y., Zhang, J., Li, S., Fang, T., McKinnon, D., Tsin, Y., Quan, L., Cao, X., Yao, Y.: Direct2.5: Diverse text-to-3d generation via multi-view 2.5d diffusion (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.785, + 0.286 + ], + "angle": 0, + "content": "18. Luiten, J., Kopanas, G., Leibe, B., Ramanan, D.: Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.327 + ], + "angle": 0, + "content": "19. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.383 + ], + "angle": 0, + "content": "20. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph. 41(4), 102:1-102:15 (Jul 2022). https://doi.org/10.1145/3528223.3530127, https://doi.org/10.1145/3528223.3530127" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.424 + ], + "angle": 0, + "content": "21. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.452 + ], + "angle": 0, + "content": "22. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.479 + ], + "angle": 0, + "content": "23. Ren, J., Pan, L., Tang, J., Zhang, C., Cao, A., Zeng, G., Liu, Z.: Dreamgaussian4d: Generative 4d gaussian splatting. arXiv preprint arXiv:2312.17142 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.522 + ], + "angle": 0, + "content": "24. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.785, + 0.563 + ], + "angle": 0, + "content": "25. Shen, T., Gao, J., Yin, K., Liu, M.Y., Fidler, S.: Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. In: Advances in Neural Information Processing Systems (NeurIPS) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.785, + 0.605 + ], + "angle": 0, + "content": "26. Shi, R., Chen, H., Zhang, Z., Liu, M., Xu, C., Wei, X., Chen, L., Zeng, C., Su, H.: Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.785, + 0.633 + ], + "angle": 0, + "content": "27. Shi, Y., Wang, P., Ye, J., Mai, L., Li, K., Yang, X.: Mvdream: Multi-view diffusion for 3d generation. arXiv:2308.16512 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.634, + 0.785, + 0.674 + ], + "angle": 0, + "content": "28. Singer, U., Sheynin, S., Polyak, A., Ashual, O., Makarov, I., Kokkinos, F., Goyal, N., Vedaldi, A., Parikh, D., Johnson, J., et al.: Text-to-4d dynamic scene generation. arXiv preprint arXiv:2301.11280 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.785, + 0.702 + ], + "angle": 0, + "content": "29. Tang, J., Ren, J., Zhou, H., Liu, Z., Zeng, G.: Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.785, + 0.757 + ], + "angle": 0, + "content": "30. Wang, H., Du, X., Li, J., Yeh, R.A., Shakhnarovich, G.: Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12619-12629 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.785 + ], + "angle": 0, + "content": "31. Wang, J., Yuan, H., Chen, D., Zhang, Y., Wang, X., Zhang, S.: Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.785, + 0.84 + ], + "angle": 0, + "content": "32. Wang, X., Wang, Y., Ye, J., Wang, Z., Sun, F., Liu, P., Wang, L., Sun, K., Wang, X., He, B.: Animatabledreamer: Text-guided non-rigid 3d model generation and reconstruction with canonical score distillation. arXiv preprint arXiv:2312.03795 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.668, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "STAG4D" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "33. Wu, G., Yi, T., Fang, J., Xie, L., Zhang, X., Wei, W., Liu, W., Tian, Q., Wang, X.: 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "34. Xian, W., Huang, J.B., Kopf, J., Kim, C.: Space-time neural irradiance fields for free-viewpoint video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9421-9431 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.232, + 0.788, + 0.272 + ], + "angle": 0, + "content": "35. Yang, Z., Gao, X., Zhou, W., Jiao, S., Zhang, Y., Jin, X.: Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. arXiv preprint arXiv:2309.13101 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.273, + 0.788, + 0.3 + ], + "angle": 0, + "content": "36. Yin, Y., Xu, D., Wang, Z., Zhao, Y., Wei, Y.: 4dgen: Grounded 4d content generation with spatial-temporal consistency. arXiv preprint arXiv:2312.17225 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.3, + 0.788, + 0.327 + ], + "angle": 0, + "content": "37. Zhang, L.: Reference-only control (2023), https://github.com/Mikubill/sd-webui-controlnet/discussions/1236" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.328, + 0.788, + 0.37 + ], + "angle": 0, + "content": "38. Zhang, S., Wang, J., Zhang, Y., Zhao, K., Yuan, H., Qing, Z., Wang, X., Zhao, D., Zhou, J.: I2vgen-xl: High-quality image-to-video synthesis via cascaded diffusion models (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.371, + 0.788, + 0.398 + ], + "angle": 0, + "content": "39. Zhao, Y., Yan, Z., Xie, E., Hong, L., Li, Z., Lee, G.H.:Animate124: Animating one image to 4d dynamic scene. arXiv preprint arXiv:2311.14603 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.398, + 0.788, + 0.439 + ], + "angle": 0, + "content": "40. Zheng, Y., Li, X., Nagano, K., Liu, S., Hilliges, O., De Mello, S.: A unified approach for text-and image-guided 4d scene generation. arXiv preprint arXiv:2311.16854 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.439 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/STAG4D_ Spatial-Temporal Anchored Generative 4D Gaussians/51eb6cce-864e-4035-8901-6f032e5306da_origin.pdf b/2024/STAG4D_ Spatial-Temporal Anchored Generative 4D Gaussians/51eb6cce-864e-4035-8901-6f032e5306da_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f9904901dcf6a35df6551dfef734157ed9aed458 --- /dev/null +++ b/2024/STAG4D_ Spatial-Temporal Anchored Generative 4D Gaussians/51eb6cce-864e-4035-8901-6f032e5306da_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:761f5649a5c3bf9882763e1e02d4fe7fbfcd6c8cd70bfb55e9e0cd6ca23a3cea +size 6073166 diff --git a/2024/STAG4D_ Spatial-Temporal Anchored Generative 4D Gaussians/full.md b/2024/STAG4D_ Spatial-Temporal Anchored Generative 4D Gaussians/full.md new file mode 100644 index 0000000000000000000000000000000000000000..31893195cbc8e0f0d09e05bee2f3304beef8a162 --- /dev/null +++ b/2024/STAG4D_ Spatial-Temporal Anchored Generative 4D Gaussians/full.md @@ -0,0 +1,263 @@ +# STAG4D: Spatial-Temporal Anchored Generative 4D Gaussians + +Yifei Zeng $^{1}$ , Yanqin Jiang $^{2}$ , Siyu Zhu $^{3}$ , Yuanxun Lu $^{1}$ , Youtian Lin $^{1}$ , Hao Zhu $^{1}$ , Weiming Hu $^{2}$ , Xun Cao $^{1}$ , and Yao Yao $^{1}$ + +$^{1}$ Nanjing University + $^{2}$ CASIA + +3 Fudan University + +Abstract. Recent progress in pre-trained diffusion models and 3D generation have spurred interest in 4D content creation. However, achieving high-fidelity 4D generation with spatial-temporal consistency remains a challenge. In this work, we propose STAG4D, a novel framework that combines pre-trained diffusion models with dynamic 3D Gaussian splatting for high-fidelity 4D generation. Drawing inspiration from 3D generation techniques, we utilize a multi-view diffusion model to initialize multiview images anchoring on the input video frames, where the video can be either real-world captured or generated by a video diffusion model. To ensure the temporal consistency of the multi-view sequence initialization, we introduce a simple yet effective fusion strategy to leverage the first frame as a temporal anchor in the self-attention computation. With the almost consistent multi-view sequences, we then apply the score distillation sampling to optimize the 4D Gaussian point cloud. The 4D Gaussian spatting is specially crafted for the generation task, where an adaptive densification strategy is proposed to mitigate the unstable Gaussian gradient for robust optimization. Notably, the proposed pipeline does not require any pre-training or fine-tuning of diffusion networks, offering a more accessible and practical solution for the 4D generation task. Extensive experiments demonstrate that our method outperforms prior 4D generation works in rendering quality, spatial-temporal consistency, and generation robustness, setting a new state-of-the-art for 4D generation from diverse inputs, including text, image, and video. + +Keywords: 4D Generation $\cdot$ 3D Gaussian Splatting $\cdot$ Diffusion Model + +# 1 Introduction + +Recent advancements in large-scale pre-trained diffusion models have shown remarkable progress in producing high-quality and diverse visual content, including images, videos, and 3D asserts [5,13,16,22,24,27,31,38]. The progress naturally extends to the realm of dynamic 3D generation [1,6,12,28,32,36,39,40]. This endeavor has gained prominence in computer vision and generative AI research, as high-quality 4D content generation is key to a broad range of applications such as autonomous driving simulation, game and film industries, digital Avatar + +![](images/947e1e25b9801a2e83c615582ff67dac2fb4bc1fb83e9d9ddf34dd56b2213027.jpg) +Fig. 1: Visualization of the generated 4D asserts. Our approach can generate diverse 4D content from various inputs, including text, image, and video. + +![](images/c7c283703a6144e5e95f8156d0823c0f101035b83cc3d4ce347ada2cca08b41d.jpg) + +creation, and spatial video production. However, previous 4D generation methods face challenges including blurry rendering, spatial-temporal inconsistency, and slow generation speed. Generating high-quality 4D content efficiently and practically remains a significant challenge. + +Given the promising strides in 3D generation from pre-trained diffusion models, there has been a concerted focus on generalized dynamic 3D generation from text or uncalibrated monocular video [1,6,12,28,32,40]. MAV3D [28] is the pioneering work of text-to-4D generation which creates 4D scenes from textual descriptions through Score Distillation Sampling (SDS) from a video diffusion model. Alternatively, Consistent4D [6] takes a video as input, and focuses on the task of video-to-4D generation. The pipeline is more flexible in that it can take advantage of high-quality input videos, dividing the difficult problem of realistic 4D content creation into a practical pipeline of text-to-video and video-to-4D generation. However, current video-to-4D generation methods still struggle to generate high-fidelity 4D asserts with spatial and temporal rendering consistency. Meanwhile, the commonly used Neural Radiance Field(NeRF) representation would suffer from an over-saturated appearance and a long optimization time. + +In this paper, we address two important factors in 4D content creation: a suitable 4D representation and spatial-temporal consistency imposed during pseudoprojection generation (i.e., multi-view video generation). Our approach employs a 4D Gaussian splatting approach specially tailored for the generation task. In the generation process, we first generate the multi-view images at all timestamps of the 4D scene. To enforce the 4D consistency of the multi-view video initialization, we design a direct fusion approach for attention computation in both spatial and temporal domains. This approach implicitly addresses texture degradation and geometric misalignment, thereby avoiding the need for adding an explicit multi-view or temporal consistency loss during the Score Distillation Sampling(SDS) [22] optimization. Then, the reference photometric loss and the SDS loss conditioned on the initial multi-view video are applied for 4D Gaussian point cloud optimization. To tackle the instability in optimizing the 4D Gaussian points, we introduce an adaptive densification strategy that is informed by the + +Gaussian gradient distribution. This approach yields high-quality and robust 4D scene generation from a monocular video input. + +We have conducted comprehensive experiments to demonstrate the effectiveness of each component of the proposed method. Our approach achieves a 2x faster generation speed compared with the previous video-to-4D approaches (e.g., [6]) and a significantly better generation quality than prior state-of-the-art methods. It is also noteworthy that the 4D content generated by our method can be rendered in real-time, opening up a wide range of possibilities for practical applications. In summary, our primary contributions are as follows: + +- We introduce a holistic 4D generation pipeline that streamlines the generation process into sequential stages: video generation, multi-view video initialization, and 4D optimization using multi-view video conditioned SDS. +- We harness the dynamic 3D Gaussian representation for 4D generation, complemented by an adaptive densification strategy. This combination enables highly precise and efficient 4D generation from monocular video inputs. +- We present a novel training-free attention fusion module to effectively integrate temporal anchor frames into the multi-view diffusion process, significantly enhancing the 4D consistency of the generated multi-view videos. +- Our method outperforms previous approaches in optimization efficiency, rendering quality, and 4D consistency, establishing a new benchmark for 4D generation across various input types, such as text, images, and videos. + +# 2 Related work + +# 2.1 3D Generation + +Dreamfusion [22] first proposes the SDS loss [22, 30] for NeRF optimization, which stands for the most prevalent technique for nowadays text-to-3D approaches. To mitigate the multi-view Janus problem, Zero123 [13] and SycnDreammer [14] fine-tune 2D diffusion models to grant the image generator with the ability of viewpoint control, while later works [15, 17, 26, 27] explicitly generate fixed multi-view images in one diffusion pass. Building upon the success of multi-view diffusion methods, we integrate a multi-view generation module into the 4D generation task, with meticulous handling of 4D spatial-temporal consistency for enhanced 4D generation quality. + +Except for equipping diffusion models with multi-view or depth awareness, another direction of progress focuses on 3D representations. Following the idea of differentiable optimization, Magic3D [10] applies a two-stage generation pipeline and changes the 3D representation to instant-NGP [20] and DMTet [25], achieving faster runtime and better generation quality. Direct2.5 [17] applies an explicit mesh representation and uses differentiable rasterization for fast mesh optimization. Recently, with the newly developed 3DGS, DreamGaussian [29] demonstrates the ability to generate 3D objects within several minutes by substituting NeRF with the 3D Gaussian representation, showing the potential of using an + +explicit point-based representation in 3D generation tasks. In this work, we introduce a novel 4D Gaussian representation and a specially tailored optimization scheme for the 4D generation task. + +# 2.2 4D Reconstruction + +Dynamic 3D reconstruction is yet another heated research topic in the field of computer vision and graphics. By extending the static NeRF [19] framework to dynamic scenes, dynamic NeRFs [3, 4, 9, 34] have demonstrated remarkable progress on dynamic 3D reconstruction. However, limited by the implicit neural representation and the complex nature of dynamic 3D information, dynamic NeRFs still suffer from slow optimization speed and low reconstruction quality. With the development of 3DGS [7], researchers have quickly extended the 3D Gaussian to dynamic scene representation, which achieves faster training and rendering speeds compared to D-NeRFs based on implicit neural representations. Dynamic 3D Gaussian [18] applies the per-frame 3DGS optimization for 4D scene reconstruction. Some studies [33, 35] attempt to amalgamate the explicit point-based 3DGS with an implicit neural field for dynamic information modeling, while Gaussian-Flow [11] introduces an explicit per-point motion model to represent a 4D scene without using implicit neural networks. It is noteworthy that it is non-trivial to extend the Gaussian reconstruction to 4D reconstruction as each reconstruction pipeline requires heuristic tuning of hyperparameters at each stage. In this paper, we target combining the 4DGS for the challenging task of 4D generation. + +# 2.3 4D Generation + +Recent developments in 3D content generation and video diffusion technologies have triggered researchers' interest in exploring 4D content generation from various input conditionings. MAV3D [28] presents the early attempt at text-to-4D generation. By utilizing score distillation sampling derived from video diffusion models, MAV3D optimizes a dynamic NeRF based on textual prompts. In a parallel vein, Consistent4D [6] has introduced the task of video-to-4D. This method capitalizes on the pre-trained knowledge from image diffusion models to optimize dynamic NeRFs via SDS optimization, showcasing the potential of high-quality 4D content generation from pre-trained 2D diffusion models. Concurrently with our work, 4DGen [36] has introduced a framework for generating dynamic 3D models, which employs spatial-temporal pseudo labels on keyframes within a multi-view diffusion model. However, the overall quality of its generation can be further improved, as its rendering fidelity is limited by the implicit neural representation and inadequate spatial-temporal information exchange exists in the diffusion generation process. In contrast, our method applies an efficient 4D Gaussian representation with spatial-temporal attention from reference- and the first-frame anchors, achieving substantially better generation results than previous works. + +![](images/c7e10cdc452231c0f2744e8812fb2453d8d06434cdcc68be42bc0d9174837572.jpg) +Fig. 2: Overall pipeline. Given a video input, we apply a multi-view diffusion model to produce coherent multi-view sequences, which serve as spatial and temporal anchors. Next, we train a deformable 3D Gaussian using multi-view SDS loss and reference loss. For a text-to-4D generation, our pipeline can be naturally extended to accept the text input by integrating with an off-the-shelf text-to-video module. + +# 3 Method + +The framework of the proposed approach, as depicted in Figure 2, outlines the generation and optimization process of a dynamic scene from a given video input. Also, our approach can be easily extended to a text-to-4D generation pipeline by utilizing an off-the-shelf video generation module. Section 3.1 presents the 4D representation used in the proposed method, along with an adaptive densification strategy that adjusts the densification threshold based on the relative motion gradient of the scene points. Subsequently, Section 3.2 provides a brief overview of the multi-view diffusion model used in the pipeline and introduces a novel spatial-temporal attention fusion mechanism designed to produce the almost consistent multi-view image sequences. Finally, as detailed in Section 3.3, the generated multi-view sequences are leveraged with spatial and temporal anchor frames for consistent 4D generation with the SDS optimization. + +# 3.1 4D Representation + +4D Gaussian Splatting The concept of 3D Gaussian Splatting was initially introduced in the work of [7], where an explicit point-based representation is used to model the 3D scene. Specifically, the 3D Gaussian point cloud, denoted as $\mathcal{S}$ for a static scene, is characterized by the tuple: + +$$ +\mathcal {S} = [ \mathcal {X}, s, r, \sigma , \zeta ], \tag {1} +$$ + +where $\mathcal{X} = (x,y,z)$ denotes the positions of the 3D Gaussian points, and $s$ , $r$ , $\sigma$ and $\zeta$ represent the scale, rotation, opacity, and spherical harmonics (SH) coeffi- + +cients of the radiance, respectively. In 3D reconstruction, the 3D Gaussian point cloud will be optimized through a point-based differentiable volume rendering. + +This representation can be naturally extended to 4D scenes by presenting the continuous scene dynamics as a 3D motion field. Inspired by [33], we present a 4D scene as a 3D Gaussian point cloud with a hex-plane-based deformation field, denoted as $\mathcal{F}(\mathcal{S},t)$ . The 3D Gaussian point cloud at time $t$ can be expressed as: + +$$ +\mathcal {F} (\mathcal {S}, t) = \left[ \mathcal {X} _ {t}, s _ {t}, r _ {t}, \sigma , \zeta \right], \tag {2} +$$ + +where $\mathcal{X}_t = (x_t, y_t, z_t)$ , $s_t$ and $r_t$ represent the updated information of Gaussian position, scale, and rotation at time $t$ . + +Adaptive Densification The standard 3D Gaussian Splatting technique typically employs a point cloud densification control strategy to dynamically adjust the number of Gaussians and their density within a unit volume. This adaptive approach allows for the transition from an initial sparse Gaussian set to a denser configuration to better represent the 3D scene. The 4D Gaussian Splatting method introduced in the work [33] utilizes a densification strategy similar to that presented in vanilla 3DGS [7], which involves the use of a fixed densification threshold based on the view-space position gradient. However, while the fixed gradient threshold strategy demonstrates efficacy in the context of reconstruction, particularly when multiple views afford robust and redundant coverage of the target scene, it performs sub-optimally in the generative setting. This limitation arises from the constraints imposed by the input single image or monocular videos, leading to significant uncertainty in the spatial and scale dimensions for each training object. This will lead to different optimal thresholds for different cases, as illustrated in the ablation study. + +To address the issue, we propose an adaptive threshold approach, wherein only candidates with relatively large gradients are selected for densification. Our approach is based on the statistical analysis of the accumulated gradient of each point, which follows a log distribution of similar shapes throughout the training process. We apply an adaptive threshold that filters out Gaussians with a relatively small gradient and only densifies those with a large gradient. The adaptive threshold is set to select a fixed percentage (top $\lambda\%$ ) of points with the highest gradient in each densification operation. This ensures that the threshold adapts to the distribution of the gradient and maintains a stable relative position. Experimental results demonstrate that the proposed simple strategy can significantly enhance the quality and robustness of the 4D generation. + +# 3.2 Temporal and Multi-view Consistent Diffusion + +Multi-view Consistent Diffusion Our pipeline adopts Score Distillation Sampling(SDS) [22] from image diffusion models for the optimization of dynamic Gaussian. Specifically, Image-to-image diffusion model Zero123 [13] is utilized + +and the SDS loss gradient could be formulated as follows: + +$$ +\nabla_ {\theta} \mathcal {L} _ {S D S} (\phi , \mathbf {x}) = \mathbb {E} _ {t, \epsilon} \left[ \omega (t) \left(\hat {\epsilon} _ {\theta} \left(\mathbf {z} _ {t}; \mathbf {I} _ {i n}, \mathbf {R}, \mathbf {T}, t\right) - \epsilon\right) \frac {\partial \mathbf {x}}{\partial \theta} \right], \tag {3} +$$ + +$$ +\hat {z} = z _ {t} - \sigma_ {t} \hat {\epsilon} _ {t} (z; \mathbf {I} _ {i n}, \mathbf {R}, \mathbf {T}). +$$ + +where $\theta$ represents the parameters of the 3D representation, $\mathbf{x}$ the rendered image at the current view, $t$ the timestamp in the diffusion process, $\epsilon$ the ground truth noise, $\hat{\epsilon}$ the predicted noise from the noisy image $\mathbf{z}_{\mathbf{t}}$ conditioned on an initial input $\mathbf{I}_{in}$ , and the relative camera pose between input view and target view $(\mathbf{R},\mathbf{T})$ . Zero123 can generate the target view image at any relative camera location but only one target view at a time, thus good at optimizing the 3D object from all views whilst bad at generating spatially consistent images from multiple target views. In contrast, Zero123++ [26] leverages reference attention [37] to model the relationship of the images from multiple target views as well as the input view, resulting in multi-view consistency output yet at the cost of fixed target camera locations. We combine the advantages of both models by taking the multi-view consistent output of Zero123++ as the input images of Zero123 when calculating SDS loss. + +Temporally Consistent Diffusion Following the view generation philosophy of multi-view diffusion based 3D generation, a practical solution to 4D content creation is to generate the multi-view videos of the 4D scene, and then apply the multi-view video conditioned SDS loss to optimize the scene. However, it is rather difficult to generate temporally consistent multi-view videos by the separate per-frame generation. Inspired by zero-shot video generation method [8], we propose a training-free temporal attention module to enable Zero123++ [26] with temporal-awareness. + +In our design, we obtain the attention information during denoising the multiview images of the first frame. Then we apply the recorded attention to the later denoising process. Specifically, during the denoising process of the latent code, the self-attention layer computes three essential components—queries $(Q)$ , keys $(K)$ , and values $(V)$ , which is formulated as: + +$$ +\operatorname {S e l f - A t t n} (Q, K, V) = \operatorname {S o f t m a x} \left(\frac {Q K ^ {T}}{\sqrt {c}}\right) V. \tag {4} +$$ + +We first compute the key $K_{0}$ and value $V_{0}$ for the initial frame, $T_{0}$ . Subsequently, for each subsequent frame $T_{t}$ , where $t \in \{1,\dots,N\}$ , we perform a mixing operation on the key $K_{t}$ and value $V_{t}$ with the initial key $K_{0}$ and value $V_{0}$ derived from the first frame $T_{0}$ . This process can be formally represented as follows: + +$$ +\left\{ \begin{array}{l} K _ {t} = \gamma \mathbf {K} _ {0} + (1 - \gamma) \mathbf {K} _ {t} \\ V _ {t} = \gamma \mathbf {V} _ {0} + (1 - \gamma) \mathbf {V} _ {t}, \end{array} \right. \tag {5} +$$ + +where the parameter $\gamma$ serves as a weighting factor that governs the influence of the initial frame, as illustrated in the ablation part. Additionally, we employ + +![](images/03e8900e2be9d449b0b6bb36f4e8a47fb662986816617e26cfa5b1cf096879d8.jpg) +Fig. 3: Illustration of the proposed spatial and temporal attention fusion during multiview sequence inference. + +reference attention following the approach outlined in [26,37], to extract local conditioning from the input image. In this reference attention mechanism, both the key and the value are obtained through the concatenation of features from the reference image and the noisy input image. Our design of temporal attention module is simple yet effective with superior temporal consistency, image quality and 3D consistency to vanilla cross-frame attention proposed in [8], as demonstrated in experiment section. + +# 3.3 Training Objectives + +Following the generation of multi-view sequences from a monocular reference video, we can acquire 6 anchor views $\{I_t^i\}_{i\in \{1\dots 6\}}$ and a reference view $I_{t}^{ref}$ at each timestep $t$ . In our optimization process, we employ multi-view score distillation sampling (SDS) utilizing the generated images $\{I_t^i\}_{i = 1\dots 6}$ in conjunction with the reference images $I_{t}^{ref}$ . The multi-view score distillation loss function $\mathcal{L}_{MVSDS}$ can be defined as: + +$$ +\begin{array}{l} \mathcal {L} _ {M V S D S} = \lambda_ {1} \mathcal {L} _ {S D S} ^ {i} + \lambda_ {2} \mathcal {L} _ {S D S} ^ {r e f} \\ = \lambda_ {1} \mathcal {L} _ {S D S} \left(\phi , I _ {t} ^ {i}\right) + \lambda_ {2} \mathcal {L} _ {S D S} \left(\phi , I _ {t} ^ {\text {r e f}}\right), \tag {6} \\ \end{array} +$$ + +where $\lambda_{1}$ and $\lambda_{2}$ are two weighting factors, the index $i$ is determined based on the proximity of the rendering viewpoint to the viewpoint of the generated images. This selection process, which we refer to as multi-view score distillation sampling, involves choosing the nearest reference image to the rendered camera view for calculating the SDS loss. + +Following the approach [29], we utilize the reference image to compute both the reconstruction loss $\mathcal{L}_{rec}$ and the foreground mask loss $\mathcal{L}_{mask}$ . Thus, the final + +optimization objective is: + +$$ +\mathcal {L} = \mathcal {L} _ {M V S D S} + \lambda_ {3} \mathcal {L} _ {r e c} + \lambda_ {4} \mathcal {L} _ {\text {m a s k}}, \tag {7} +$$ + +where $\lambda_3$ and $\lambda_4$ are the weighting parameters. During training, we first use $\mathcal{L}$ to supervise a fixed frame to get static canonical 3D Gaussian. Then we use all anchors and reference images to train the dynamic 4D Gaussian point cloud. + +# 3.4 Text-to-4D Extension + +Our method is designed to be readily adaptable to text and image inputs, offering novel capabilities such as directly generating video sequences from textual descriptions or static images. We initiate this process using a 2D diffusion model SDXL [21], which generates the image from textual input. This 2D output is then transformed into a video sequence via a video diffusion model, such as SVD [2], adding temporal coherence and motion to the static image. Finally, the aforementioned pipeline further lifts the video sequence to a 4D scene, introducing an extra spatial dimension that creates immersive experiences beyond conventional 2D or 3D media. The generation pipeline with the given text, image, and video inputs is visualized in Fig. 2. + +# 4 Experiment + +# 4.1 Experiment Setup + +Dataset For the video-to-4D task, we utilize the dataset provided by Consistent4D [6] for quantitative evaluation, which comprises multi-view videos depicting 7 dynamic objects. Additionally, for qualitative evaluation, we curate a set of challenging videos from online sources to assess the robustness and generalization capabilities of each method. For text/image-to-4D tasks, we follow the data settings in 4Dfy [1] and DreamGaussian4D [23] to create the corresponding results. Notably, we use the same input image and input video with DreamGaussian4D for a fair comparison. + +Evaluation metrics We employ CLIP, LPIPS, and FVD for Video-to-4D evaluation as in the previous approach [6]. Specifically, CLIP and LPIPS serve as image-level metrics, assessing the semantic similarity between rendered images and ground truths; FVD is a video-level metric commonly used in video generation tasks, considering not only single-frame quality but also temporal coherence; we also consider the FID-VID metric to measure the temporal consistency. As 4D-Gen [36] does not support a 30-frame setting we report its FVD under a 16-frame reconstruction, namely FVD-16 as an additional metrics. For text-to-4D generation, we follow previous works and provide a user study for quantitative comparisons between different methods. We also provide a user study for video-to-4D and both studies are detailed in the supplementary material. + +![](images/739a36b04b574fcb05f5e51cdedbcc347c55f2297d53625a5e939a38ef6e8997.jpg) +Fig. 4: Qualitative comparison on video-to-4D generation. Our method achieves better faithfulness to the input video with higher quality on the overall reconstruction. + +Baselines For video-to-4D generation, we compare our method with two baselines: Consistent4D [6] and the concurrent work 4DGen [36]. We utilize the official code released by the authors to generate comparison results. For text&image-to-4D generation, we make comparisons with two state-of-art models that are open-sourced: 4Dfy [1] and DreamGaussian4D [23]. Their results are generated using codes from their official GitHub repository. + +# 4.2 Implementation Details + +In the training of the video-to-4D generative model, we adopt a two-stage approach. Initially, the model is initialized in canonical space and trained for 1000 steps. Subsequently, the deformation fields are learned over 7000 additional steps to accurately capture the dynamic scene. For the deformation process, we employ multi-layer perceptrons (MLPs) with 64 hidden layers and 32 hidden features per layer. The initial learning rate for the deformation MLPs is set to $1.6 \times 10^{-4}$ and is decayed to $1.6 \times 10^{-6}$ by the end of the training. In the context of adaptive densification, we opt to densify the top $2.5\%$ of points with the most accumulated gradient. In the generative phase, we utilize the spatial-temporal consistent videos for the multiview score distillation sampling optimization. Empirically, we set $\lambda$ to 0.5 for our temporally consistent diffusion. Regarding the loss functions, we maintain the SDS loss weight at 1 and adjust the weights of other losses accordingly. Specifically, the reconstruction loss is assigned to $4 \times 10^{4}$ , the mask loss is weighted at $1 \times 10^{4}$ . The training process requires approximately 1 hour on an RTX 3090 GPU, and the rendering process can be performed at 150 FPS in real time. For Text&Image-to-4D, we use SDXL [21] for image generation, and apply SVD [2] to the generated image to create a corresponding video. Then we use the same setting as Video-to-4D to transfer the video into 4D content. + +![](images/1ed653f04bfc9197b1adfa0edb704b373dc56507d1104fd7559a2fb1a44a4531.jpg) +Fig. 5: Qualitative comparison on text-to-4D generation. Our method achieves the best results on the text alignment and visual quality. + +# 4.3 Comparison with State-of-the-art Approaches + +In this section, we conduct a comprehensive comparison with the aforementioned baselines using synthetic and in-the-wild data. The superior performance demonstrated in both quantitative and qualitative evaluations serves to substantiate the effectiveness and robustness of our proposed method. + +Quantitative Results on Video-to-4D The experimental results presented in Table 1 offer a comparative analysis of our method against state-of-the-art techniques, namely Consistent4D and 4DGen, across several metrics. Our method outperforms Consistent4D and 4DGen in CLIP, showing better semantic consistency with the target content. It also achieves the lowest LPIPS score, meaning it generates more realistic images than the others. For the video quality and smoothness, our method beats 4DGen in FID-VID and FVD-16, and Consistent4D in FVD, implying that our videos are closer to real videos and have less temporal artifacts. Overall, the numbers confirm that our method is superior in generating semantically aligned and perceptually convincing videos with high fidelity and coherence. + +Qualitative Comparison on Video-to-4D In our investigation of the video-to-4D task, we have employed various methodologies to conduct qualitative comparisons. Figure 4 shows the rendered outcomes at distinct temporal intervals and perspectives. A critical observation of our findings reveals that Gaussian splatting is inadequate in producing credible and temporally consistent images during per-frame reconstruction. Furthermore, we have observed that Consistent4D often produces over-saturated and unrealistic patterns, which may stem from the over-restrictive geometry representation in their cascade Dynerf. Additionally, our analysis indicates that 4DGen struggles to generate reasonable motion for the Gaussians, leading to its points lacking deformation over time. Moreover, 4DGen fails to generate detailed textures for the surface, which appear + +![](images/97944a9d0d08c11f6788a824ee1ef039af965363aefed34324797c2ce81367de.jpg) +Fig. 6: Ablation on spatial and temporal attention. We evaluate the effect of spatial and temporal attention on the reconstruction results. + +![](images/4b3a9d7d1b1b49693c72fc53048b60ece35c286e205dce6cb43f93927a8cbeaa.jpg) + +to be over-smoothed in some cases. Conversely, our methodology demonstrates substantial enhancements in comparison to existing strategies, particularly in aspects of reconstruction fidelity and stability. Notably, our technique facilitates seamless color transitions, as exemplified in the Fish case, and renders a convincing low-poly effect, as demonstrated in the Trump case, underscoring our method's adaptability to diverse dynamic environments. + +Qualitative Comparison on Text-to-4D We also evaluate our method on the Text&Image-to-4D task, which aims to generate 4D content from text and image inputs. Figure 5 shows the 4D generation results of our method and two baselines, 4Dfy [1] and DreamGaussian4D [23], under different views. As can be seen, 4Dfy [1] produces coarse results and fails to capture the complex semantics of the text input, e.g. in the Cat case. DreamGaussian4D [23] generates meshes with fine-grained details in the front view, but suffers from severe artifacts and distortions in the side view or back view. In contrast, our method synthesizes Gaussians with smooth and consistent geometry and realistic texture across different views, thanks to our attention mechanism that leverages spatial and temporal information. Our method demonstrates its superiority and versatility on the Text&Image-to-4D task, which can be seen as a direct application of our Video-to-4D framework. + +Table 1: Comparison with state-of-the-art methods. + +
CLIP ↑LPIPS ↓FID-VID ↓FVD ↓FVD-16 ↓
Consistent4D0.8770.134/1133.93/
4DGen0.8940.13071.99/1005.72
Ours0.9090.12652.58992.21952.41
+ +# 4.4 Ablation Study + +The experimental analysis presented in Table 2 provides a comprehensive evaluation of the impact of various components on the performance of a diffusion model with a 4D representation. The components under consideration include baseline diffusion, spatial anchor, spatial-temporal anchor, and adaptive densification. In the first configuration of Table 2, we regard the model utilizing the diffusion process [13] with adaptive densification as the baseline. + +Table 2: Ablation study on effectiveness of each novel component. + +
Diffusion Model4D Gaussian +Adaptive Dens.Evaluation Metrics
BaselineSpat.Temp.CLIP ↑LPIPS ↓FID-VID ↓FVD ↓
✓✓0.8950.13577.881369.65
✓✓✓0.8990.13067.041198.23
✓✓✓0.8900.13690.301453.53
✓✓✓✓0.9090.12652.58992.21
+ +Table 3: Ablation study on different attention mechanisms. + +
Evaluation Metrics
CLIPLPIPSFID-VIDFVD
Spat. only0.8990.13067.041198.23
Cross-Frame0.9010.12963.321053.25
Spat.-Temp.0.9090.12652.58992.21
+ +Spatial and Temporal Anchors Table 2 shows that adding the spatial anchor to the baseline model leads to a marginal improvement in all metrics, especially FID-VID and FVD. The final model, which uses both spatial-temporal anchor and adaptive densification, improves all metrics significantly. It has the best CLIP, LPIPS, FID-VID, and FVD scores, meaning it generates the most semantically consistent, realistic, and smooth videos. + +Figure 6 shows the importance of our spatial-temporal anchor for improving the 4D content quality. The baseline 4D Gaussian method, labeled as "w./o. spatial," fails to deform the point clouds properly, leading to unreasonable shapes without spatial-temporal constraints. A model that uses only the spatial anchor, "w./o. temporal," struggles to deform solid shapes, as seen in the rabbit's leg and the frog's back. Our method, which combines both spatial and temporal anchors, overcomes these issues and produces visually consistent and accurate 4D content. + +Adaptive Densification We present the statistical analysis of the adaptive densification module in the third and final configurations of Table 2. The results show that the adaptive densification module improves the overall quality of the generative model, especially in terms of the FID-VID and FVD metrics. We also compare the proposed adaptive densification with a fixed threshold approach in Figure 7b. We observe that a fixed threshold can lead to either under-densification or over-densification, depending on the case. For example, in the Bird case, the fixed threshold is too low for densification, which results in excessive Gaussian points. On the other hand, in the Duck case, the fixed threshold is too high for densification, which causes the foot to be blurry due to insufficient Gaussian points. Our method overcomes this limitation by adapting the gradient threshold to different cases, thus producing robust and stable 4D Gaussian results. + +![](images/e1442b050157ece5ea308ebdf48253713e9f5707d1a4febf8799d5f677f773f4.jpg) + +Different Attention Mechanisms In our study, we evaluated three distinct attention mechanisms within the context of images generated by our diffusion model: spatial attention, cross-frame attention, and spatial-temporal attention. The visual outcomes of these varied approaches are illustrated in Figure 7a. We employed a parameter $\lambda$ to modulate the extent of spatial and temporal attention exerted, as delineated in Equation 5. Our findings indicate that exclusive reliance on spatial attention does not guarantee consistency across sequential frames. While cross-frame attention enhances frame-to-frame coherence, it fails to achieve a satisfactory alignment with reference images. Conversely, the integration of spatial-temporal attention ensures consistency across both dimensions. We quantitatively assessed the quality of images generated under different attention regimes using four established metrics: CLIP, LPIPS, FID-VID, and FVD. The comparative results are tabulated in Table 3. It was observed that images generated with spatial-temporal attention outperformed others across all evaluation metrics, which underscores the efficacy of our proposed mechanism. + +# 5 Conclusion + +The paper presents a novel approach for dynamic 3D content generation from monocular videos, addressing the challenges of 4D representation and spatial-temporal consistency. By leveraging specially tailored 4D Gaussian splatting and a novel information fusion module, the proposed method achieves high-quality and robust 4D scene generation. Comprehensive experiments demonstrate the method's effectiveness, showcasing a faster generation speed and significant improvements in rendering quality and temporal consistency compared to prior state-of-the-art methods. Overall, the proposed method sets a new benchmark for training speed, rendering quality, and 4D consistency in dynamic 3D content generation from monocular videos, opening up possibilities for real-world applications. + +# Acknowledgements + +This work was supported by the National Key R&D Program of China (2022YFF0902200) and NSFC grant 62441204. + +# References + +1. Bahmani, S., Skorokhodov, I., Rong, V., Wetzstein, G., Guibas, L., Wonka, P., Tulyakov, S., Park, J.J., Tagliasacchi, A., Lindell, D.B.: 4d-fy: Text-to-4d generation using hybrid score distillation sampling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7996-8006 (2024) +2. Blattmann, A., Dockhorn, T., Kulal, S., Mendelevitch, D., Kilian, M., Lorenz, D., Levi, Y., English, Z., Voleti, V., Letts, A., Jampani, V., Rombach, R.: Stable video diffusion: Scaling latent video diffusion models to large datasets (2023) +3. Du, Y., Zhang, Y., Yu, H.X., Tenenbaum, J.B., Wu, J.: Neural radiance flow for 4d view synthesis and video processing. In: 2021 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 14304-14314. IEEE Computer Society (2021) +4. Gao, C., Saraf, A., Kopf, J., Huang, J.B.: Dynamic view synthesis from dynamic monocular video. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5712-5721 (2021) +5. Ho, J., Chan, W., Sahara, C., Whang, J., Gao, R., Gritsenko, A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., et al.:Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303 (2022) +6. Jiang, Y., Zhang, L., Gao, J., Hu, W., Yao, Y.: Consistent4d: Consistent 360 $\{\backslash \mathrm{deg}\}$ dynamic object generation from monocular video. arXiv preprint arXiv:2311.02848 (2023) +7. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG) 42(4), 1-14 (2023) +8. Khachatryan, L., Movsisyan, A., Tadevosyan, V., Henschel, R., Wang, Z., Navasardyan, S., Shi, H.: Text2video-zero: Text-to-image diffusion models are zero-shot video generators. arXiv preprint arXiv:2303.13439 (2023) +9. Li, Z., Niklaus, S., Snavely, N., Wang, O.: Neural scene flow fields for space-time view synthesis of dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2021) +0. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3d: High-resolution text-to-3d content creation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 300–309 (2023) +1. Lin, Y., Dai, Z., Zhu, S., Yao, Y.: Gaussian-flow: 4d reconstruction with dynamic 3d gaussian particle. arXiv:2312.03431 (2023) +2. Ling, H., Kim, S.W., Torralba, A., Fidler, S., Kreis, K.: Align your gaussians: Text-to-4d with dynamic 3d gaussians and composed diffusion models. arXiv preprint arXiv:2312.13763 (2023) +3. Liu, R., Wu, R., Van Hoorick, B., Tokmakov, P., Zakharov, S., Vondrick, C.: Zero-1-to-3: Zero-shot one image to 3d object. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9298-9309 (2023) +4. Liu, Y., Lin, C., Zeng, Z., Long, X., Liu, L., Komura, T., Wang, W.: Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453 (2023) + +15. Long, X., Guo, Y.C., Lin, C., Liu, Y., Dou, Z., Liu, L., Ma, Y., Zhang, S.H., Habermann, M., Theobalt, C., et al.: Wonder3d: Single image to 3d using cross-domain diffusion. arXiv preprint arXiv:2310.15008 (2023) +16. Lu, Y., Zhang, J., Li, S., Fang, T., McKinnon, D., Tsin, Y., Quan, L., Cao, X., Yao, Y.: Direct2. 5: Diverse text-to-3d generation via multi-view 2.5 d diffusion. arXiv preprint arXiv:2311.15980 (2023) +17. Lu, Y., Zhang, J., Li, S., Fang, T., McKinnon, D., Tsin, Y., Quan, L., Cao, X., Yao, Y.: Direct2.5: Diverse text-to-3d generation via multi-view 2.5d diffusion (2023) +18. Luiten, J., Kopanas, G., Leibe, B., Ramanan, D.: Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713 (2023) +19. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020) +20. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph. 41(4), 102:1-102:15 (Jul 2022). https://doi.org/10.1145/3528223.3530127, https://doi.org/10.1145/3528223.3530127 +21. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis (2023) +22. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988 (2022) +23. Ren, J., Pan, L., Tang, J., Zhang, C., Cao, A., Zeng, G., Liu, Z.: Dreamgaussian4d: Generative 4d gaussian splatting. arXiv preprint arXiv:2312.17142 (2023) +24. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2022) +25. Shen, T., Gao, J., Yin, K., Liu, M.Y., Fidler, S.: Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. In: Advances in Neural Information Processing Systems (NeurIPS) (2021) +26. Shi, R., Chen, H., Zhang, Z., Liu, M., Xu, C., Wei, X., Chen, L., Zeng, C., Su, H.: Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110 (2023) +27. Shi, Y., Wang, P., Ye, J., Mai, L., Li, K., Yang, X.: Mvdream: Multi-view diffusion for 3d generation. arXiv:2308.16512 (2023) +28. Singer, U., Sheynin, S., Polyak, A., Ashual, O., Makarov, I., Kokkinos, F., Goyal, N., Vedaldi, A., Parikh, D., Johnson, J., et al.: Text-to-4d dynamic scene generation. arXiv preprint arXiv:2301.11280 (2023) +29. Tang, J., Ren, J., Zhou, H., Liu, Z., Zeng, G.: Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653 (2023) +30. Wang, H., Du, X., Li, J., Yeh, R.A., Shakhnarovich, G.: Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12619-12629 (2023) +31. Wang, J., Yuan, H., Chen, D., Zhang, Y., Wang, X., Zhang, S.: Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571 (2023) +32. Wang, X., Wang, Y., Ye, J., Wang, Z., Sun, F., Liu, P., Wang, L., Sun, K., Wang, X., He, B.: Animatabledreamer: Text-guided non-rigid 3d model generation and reconstruction with canonical score distillation. arXiv preprint arXiv:2312.03795 (2023) + +33. Wu, G., Yi, T., Fang, J., Xie, L., Zhang, X., Wei, W., Liu, W., Tian, Q., Wang, X.: 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528 (2023) +34. Xian, W., Huang, J.B., Kopf, J., Kim, C.: Space-time neural irradiance fields for free-viewpoint video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9421-9431 (2021) +35. Yang, Z., Gao, X., Zhou, W., Jiao, S., Zhang, Y., Jin, X.: Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. arXiv preprint arXiv:2309.13101 (2023) +36. Yin, Y., Xu, D., Wang, Z., Zhao, Y., Wei, Y.: 4dgen: Grounded 4d content generation with spatial-temporal consistency. arXiv preprint arXiv:2312.17225 (2023) +37. Zhang, L.: Reference-only control (2023), https://github.com/Mikubill/sd-webui-controlnet/discussions/1236 +38. Zhang, S., Wang, J., Zhang, Y., Zhao, K., Yuan, H., Qing, Z., Wang, X., Zhao, D., Zhou, J.: I2vgen-xl: High-quality image-to-video synthesis via cascaded diffusion models (2023) +39. Zhao, Y., Yan, Z., Xie, E., Hong, L., Li, Z., Lee, G.H.:Animate124: Animating one image to 4d dynamic scene. arXiv preprint arXiv:2311.14603 (2023) +40. Zheng, Y., Li, X., Nagano, K., Liu, S., Hilliges, O., De Mello, S.: A unified approach for text-and image-guided 4d scene generation. arXiv preprint arXiv:2311.16854 (2023) \ No newline at end of file diff --git a/2024/STAG4D_ Spatial-Temporal Anchored Generative 4D Gaussians/images.zip b/2024/STAG4D_ Spatial-Temporal Anchored Generative 4D Gaussians/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c13acd73d51ff5cc074639e25bb43fd8c1e6b886 --- /dev/null +++ b/2024/STAG4D_ Spatial-Temporal Anchored Generative 4D Gaussians/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:14317313cd4ecd334d70fafeaa3a39aa42097691779fcfad35b5be3fcba5e6c2 +size 512979 diff --git a/2024/STAG4D_ Spatial-Temporal Anchored Generative 4D Gaussians/layout.json b/2024/STAG4D_ Spatial-Temporal Anchored Generative 4D Gaussians/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..de5f24399dce240a85d89e392c791871fb4f2e76 --- /dev/null +++ b/2024/STAG4D_ Spatial-Temporal Anchored Generative 4D Gaussians/layout.json @@ -0,0 +1,7657 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 134, + 111, + 479, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 111, + 479, + 146 + ], + "spans": [ + { + "bbox": [ + 134, + 111, + 479, + 146 + ], + "type": "text", + "content": "STAG4D: Spatial-Temporal Anchored Generative 4D Gaussians" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 132, + 167, + 481, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 167, + 481, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 167, + 481, + 193 + ], + "type": "text", + "content": "Yifei Zeng" + }, + { + "bbox": [ + 132, + 167, + 481, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 132, + 167, + 481, + 193 + ], + "type": "text", + "content": ", Yanqin Jiang" + }, + { + "bbox": [ + 132, + 167, + 481, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 132, + 167, + 481, + 193 + ], + "type": "text", + "content": ", Siyu Zhu" + }, + { + "bbox": [ + 132, + 167, + 481, + 193 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 132, + 167, + 481, + 193 + ], + "type": "text", + "content": ", Yuanxun Lu" + }, + { + "bbox": [ + 132, + 167, + 481, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 132, + 167, + 481, + 193 + ], + "type": "text", + "content": ", Youtian Lin" + }, + { + "bbox": [ + 132, + 167, + 481, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 132, + 167, + 481, + 193 + ], + "type": "text", + "content": ", Hao Zhu" + }, + { + "bbox": [ + 132, + 167, + 481, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 132, + 167, + 481, + 193 + ], + "type": "text", + "content": ", Weiming Hu" + }, + { + "bbox": [ + 132, + 167, + 481, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 132, + 167, + 481, + 193 + ], + "type": "text", + "content": ", Xun Cao" + }, + { + "bbox": [ + 132, + 167, + 481, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 132, + 167, + 481, + 193 + ], + "type": "text", + "content": ", and Yao Yao" + }, + { + "bbox": [ + 132, + 167, + 481, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 263, + 201, + 351, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 263, + 201, + 351, + 223 + ], + "spans": [ + { + "bbox": [ + 263, + 201, + 351, + 223 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 263, + 201, + 351, + 223 + ], + "type": "text", + "content": " Nanjing University \n" + }, + { + "bbox": [ + 263, + 201, + 351, + 223 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 263, + 201, + 351, + 223 + ], + "type": "text", + "content": " CASIA" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 266, + 224, + 347, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 224, + 347, + 235 + ], + "spans": [ + { + "bbox": [ + 266, + 224, + 347, + 235 + ], + "type": "text", + "content": "3 Fudan University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 263, + 455, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 263, + 455, + 517 + ], + "spans": [ + { + "bbox": [ + 160, + 263, + 455, + 517 + ], + "type": "text", + "content": "Abstract. Recent progress in pre-trained diffusion models and 3D generation have spurred interest in 4D content creation. However, achieving high-fidelity 4D generation with spatial-temporal consistency remains a challenge. In this work, we propose STAG4D, a novel framework that combines pre-trained diffusion models with dynamic 3D Gaussian splatting for high-fidelity 4D generation. Drawing inspiration from 3D generation techniques, we utilize a multi-view diffusion model to initialize multiview images anchoring on the input video frames, where the video can be either real-world captured or generated by a video diffusion model. To ensure the temporal consistency of the multi-view sequence initialization, we introduce a simple yet effective fusion strategy to leverage the first frame as a temporal anchor in the self-attention computation. With the almost consistent multi-view sequences, we then apply the score distillation sampling to optimize the 4D Gaussian point cloud. The 4D Gaussian spatting is specially crafted for the generation task, where an adaptive densification strategy is proposed to mitigate the unstable Gaussian gradient for robust optimization. Notably, the proposed pipeline does not require any pre-training or fine-tuning of diffusion networks, offering a more accessible and practical solution for the 4D generation task. Extensive experiments demonstrate that our method outperforms prior 4D generation works in rendering quality, spatial-temporal consistency, and generation robustness, setting a new state-of-the-art for 4D generation from diverse inputs, including text, image, and video." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 525, + 450, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 525, + 450, + 537 + ], + "spans": [ + { + "bbox": [ + 160, + 525, + 450, + 537 + ], + "type": "text", + "content": "Keywords: 4D Generation " + }, + { + "bbox": [ + 160, + 525, + 450, + 537 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 525, + 450, + 537 + ], + "type": "text", + "content": " 3D Gaussian Splatting " + }, + { + "bbox": [ + 160, + 525, + 450, + 537 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 525, + 450, + 537 + ], + "type": "text", + "content": " Diffusion Model" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 557, + 230, + 570 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 230, + 570 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 230, + 570 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Recent advancements in large-scale pre-trained diffusion models have shown remarkable progress in producing high-quality and diverse visual content, including images, videos, and 3D asserts [5,13,16,22,24,27,31,38]. The progress naturally extends to the realm of dynamic 3D generation [1,6,12,28,32,36,39,40]. This endeavor has gained prominence in computer vision and generative AI research, as high-quality 4D content generation is key to a broad range of applications such as autonomous driving simulation, game and film industries, digital Avatar" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 115, + 315, + 213 + ], + "blocks": [ + { + "bbox": [ + 133, + 115, + 315, + 213 + ], + "lines": [ + { + "bbox": [ + 133, + 115, + 315, + 213 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 315, + 213 + ], + "type": "image", + "image_path": "947e1e25b9801a2e83c615582ff67dac2fb4bc1fb83e9d9ddf34dd56b2213027.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 223, + 482, + 246 + ], + "lines": [ + { + "bbox": [ + 131, + 223, + 482, + 246 + ], + "spans": [ + { + "bbox": [ + 131, + 223, + 482, + 246 + ], + "type": "text", + "content": "Fig. 1: Visualization of the generated 4D asserts. Our approach can generate diverse 4D content from various inputs, including text, image, and video." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 320, + 115, + 482, + 205 + ], + "blocks": [ + { + "bbox": [ + 320, + 115, + 482, + 205 + ], + "lines": [ + { + "bbox": [ + 320, + 115, + 482, + 205 + ], + "spans": [ + { + "bbox": [ + 320, + 115, + 482, + 205 + ], + "type": "image", + "image_path": "c7c283703a6144e5e95f8156d0823c0f101035b83cc3d4ce347ada2cca08b41d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 276, + 479, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 276, + 479, + 324 + ], + "spans": [ + { + "bbox": [ + 130, + 276, + 479, + 324 + ], + "type": "text", + "content": "creation, and spatial video production. However, previous 4D generation methods face challenges including blurry rendering, spatial-temporal inconsistency, and slow generation speed. Generating high-quality 4D content efficiently and practically remains a significant challenge." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 327, + 480, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 480, + 493 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 480, + 493 + ], + "type": "text", + "content": "Given the promising strides in 3D generation from pre-trained diffusion models, there has been a concerted focus on generalized dynamic 3D generation from text or uncalibrated monocular video [1,6,12,28,32,40]. MAV3D [28] is the pioneering work of text-to-4D generation which creates 4D scenes from textual descriptions through Score Distillation Sampling (SDS) from a video diffusion model. Alternatively, Consistent4D [6] takes a video as input, and focuses on the task of video-to-4D generation. The pipeline is more flexible in that it can take advantage of high-quality input videos, dividing the difficult problem of realistic 4D content creation into a practical pipeline of text-to-video and video-to-4D generation. However, current video-to-4D generation methods still struggle to generate high-fidelity 4D asserts with spatial and temporal rendering consistency. Meanwhile, the commonly used Neural Radiance Field(NeRF) representation would suffer from an over-saturated appearance and a long optimization time." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "text", + "content": "In this paper, we address two important factors in 4D content creation: a suitable 4D representation and spatial-temporal consistency imposed during pseudoprojection generation (i.e., multi-view video generation). Our approach employs a 4D Gaussian splatting approach specially tailored for the generation task. In the generation process, we first generate the multi-view images at all timestamps of the 4D scene. To enforce the 4D consistency of the multi-view video initialization, we design a direct fusion approach for attention computation in both spatial and temporal domains. This approach implicitly addresses texture degradation and geometric misalignment, thereby avoiding the need for adding an explicit multi-view or temporal consistency loss during the Score Distillation Sampling(SDS) [22] optimization. Then, the reference photometric loss and the SDS loss conditioned on the initial multi-view video are applied for 4D Gaussian point cloud optimization. To tackle the instability in optimizing the 4D Gaussian points, we introduce an adaptive densification strategy that is informed by the" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 211, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 211, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 211, + 102 + ], + "type": "text", + "content": "Zeng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "Gaussian gradient distribution. This approach yields high-quality and robust 4D scene generation from a monocular video input." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 480, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 480, + 224 + ], + "type": "text", + "content": "We have conducted comprehensive experiments to demonstrate the effectiveness of each component of the proposed method. Our approach achieves a 2x faster generation speed compared with the previous video-to-4D approaches (e.g., [6]) and a significantly better generation quality than prior state-of-the-art methods. It is also noteworthy that the 4D content generated by our method can be rendered in real-time, opening up a wide range of possibilities for practical applications. In summary, our primary contributions are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 236, + 479, + 378 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 236, + 479, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 236, + 479, + 270 + ], + "spans": [ + { + "bbox": [ + 138, + 236, + 479, + 270 + ], + "type": "text", + "content": "- We introduce a holistic 4D generation pipeline that streamlines the generation process into sequential stages: video generation, multi-view video initialization, and 4D optimization using multi-view video conditioned SDS." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 272, + 479, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 272, + 479, + 306 + ], + "spans": [ + { + "bbox": [ + 138, + 272, + 479, + 306 + ], + "type": "text", + "content": "- We harness the dynamic 3D Gaussian representation for 4D generation, complemented by an adaptive densification strategy. This combination enables highly precise and efficient 4D generation from monocular video inputs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 308, + 479, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 308, + 479, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 308, + 479, + 342 + ], + "type": "text", + "content": "- We present a novel training-free attention fusion module to effectively integrate temporal anchor frames into the multi-view diffusion process, significantly enhancing the 4D consistency of the generated multi-view videos." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 345, + 479, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 345, + 479, + 378 + ], + "spans": [ + { + "bbox": [ + 138, + 345, + 479, + 378 + ], + "type": "text", + "content": "- Our method outperforms previous approaches in optimization efficiency, rendering quality, and 4D consistency, establishing a new benchmark for 4D generation across various input types, such as text, images, and videos." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 131, + 401, + 234, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 401, + 234, + 414 + ], + "spans": [ + { + "bbox": [ + 131, + 401, + 234, + 414 + ], + "type": "text", + "content": "2 Related work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 428, + 236, + 439 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 428, + 236, + 439 + ], + "spans": [ + { + "bbox": [ + 131, + 428, + 236, + 439 + ], + "type": "text", + "content": "2.1 3D Generation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 449, + 480, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 480, + 557 + ], + "type": "text", + "content": "Dreamfusion [22] first proposes the SDS loss [22, 30] for NeRF optimization, which stands for the most prevalent technique for nowadays text-to-3D approaches. To mitigate the multi-view Janus problem, Zero123 [13] and SycnDreammer [14] fine-tune 2D diffusion models to grant the image generator with the ability of viewpoint control, while later works [15, 17, 26, 27] explicitly generate fixed multi-view images in one diffusion pass. Building upon the success of multi-view diffusion methods, we integrate a multi-view generation module into the 4D generation task, with meticulous handling of 4D spatial-temporal consistency for enhanced 4D generation quality." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "type": "text", + "content": "Except for equipping diffusion models with multi-view or depth awareness, another direction of progress focuses on 3D representations. Following the idea of differentiable optimization, Magic3D [10] applies a two-stage generation pipeline and changes the 3D representation to instant-NGP [20] and DMTet [25], achieving faster runtime and better generation quality. Direct2.5 [17] applies an explicit mesh representation and uses differentiable rasterization for fast mesh optimization. Recently, with the newly developed 3DGS, DreamGaussian [29] demonstrates the ability to generate 3D objects within several minutes by substituting NeRF with the 3D Gaussian representation, showing the potential of using an" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "text", + "content": "STAG4D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 152 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 152 + ], + "type": "text", + "content": "explicit point-based representation in 3D generation tasks. In this work, we introduce a novel 4D Gaussian representation and a specially tailored optimization scheme for the 4D generation task." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 171, + 254, + 183 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 171, + 254, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 171, + 254, + 183 + ], + "type": "text", + "content": "2.2 4D Reconstruction" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 193, + 482, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 193, + 482, + 410 + ], + "spans": [ + { + "bbox": [ + 132, + 193, + 482, + 410 + ], + "type": "text", + "content": "Dynamic 3D reconstruction is yet another heated research topic in the field of computer vision and graphics. By extending the static NeRF [19] framework to dynamic scenes, dynamic NeRFs [3, 4, 9, 34] have demonstrated remarkable progress on dynamic 3D reconstruction. However, limited by the implicit neural representation and the complex nature of dynamic 3D information, dynamic NeRFs still suffer from slow optimization speed and low reconstruction quality. With the development of 3DGS [7], researchers have quickly extended the 3D Gaussian to dynamic scene representation, which achieves faster training and rendering speeds compared to D-NeRFs based on implicit neural representations. Dynamic 3D Gaussian [18] applies the per-frame 3DGS optimization for 4D scene reconstruction. Some studies [33, 35] attempt to amalgamate the explicit point-based 3DGS with an implicit neural field for dynamic information modeling, while Gaussian-Flow [11] introduces an explicit per-point motion model to represent a 4D scene without using implicit neural networks. It is noteworthy that it is non-trivial to extend the Gaussian reconstruction to 4D reconstruction as each reconstruction pipeline requires heuristic tuning of hyperparameters at each stage. In this paper, we target combining the 4DGS for the challenging task of 4D generation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 428, + 235, + 440 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 428, + 235, + 440 + ], + "spans": [ + { + "bbox": [ + 132, + 428, + 235, + 440 + ], + "type": "text", + "content": "2.3 4D Generation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 450, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 482, + 666 + ], + "type": "text", + "content": "Recent developments in 3D content generation and video diffusion technologies have triggered researchers' interest in exploring 4D content generation from various input conditionings. MAV3D [28] presents the early attempt at text-to-4D generation. By utilizing score distillation sampling derived from video diffusion models, MAV3D optimizes a dynamic NeRF based on textual prompts. In a parallel vein, Consistent4D [6] has introduced the task of video-to-4D. This method capitalizes on the pre-trained knowledge from image diffusion models to optimize dynamic NeRFs via SDS optimization, showcasing the potential of high-quality 4D content generation from pre-trained 2D diffusion models. Concurrently with our work, 4DGen [36] has introduced a framework for generating dynamic 3D models, which employs spatial-temporal pseudo labels on keyframes within a multi-view diffusion model. However, the overall quality of its generation can be further improved, as its rendering fidelity is limited by the implicit neural representation and inadequate spatial-temporal information exchange exists in the diffusion generation process. In contrast, our method applies an efficient 4D Gaussian representation with spatial-temporal attention from reference- and the first-frame anchors, achieving substantially better generation results than previous works." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 211, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 211, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 211, + 102 + ], + "type": "text", + "content": "Zeng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 116, + 478, + 272 + ], + "blocks": [ + { + "bbox": [ + 136, + 116, + 478, + 272 + ], + "lines": [ + { + "bbox": [ + 136, + 116, + 478, + 272 + ], + "spans": [ + { + "bbox": [ + 136, + 116, + 478, + 272 + ], + "type": "image", + "image_path": "c7e10cdc452231c0f2744e8812fb2453d8d06434cdcc68be42bc0d9174837572.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 282, + 482, + 338 + ], + "lines": [ + { + "bbox": [ + 131, + 282, + 482, + 338 + ], + "spans": [ + { + "bbox": [ + 131, + 282, + 482, + 338 + ], + "type": "text", + "content": "Fig. 2: Overall pipeline. Given a video input, we apply a multi-view diffusion model to produce coherent multi-view sequences, which serve as spatial and temporal anchors. Next, we train a deformable 3D Gaussian using multi-view SDS loss and reference loss. For a text-to-4D generation, our pipeline can be naturally extended to accept the text input by integrating with an off-the-shelf text-to-video module." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 359, + 201, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 359, + 201, + 371 + ], + "spans": [ + { + "bbox": [ + 132, + 359, + 201, + 371 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 384, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 482, + 529 + ], + "type": "text", + "content": "The framework of the proposed approach, as depicted in Figure 2, outlines the generation and optimization process of a dynamic scene from a given video input. Also, our approach can be easily extended to a text-to-4D generation pipeline by utilizing an off-the-shelf video generation module. Section 3.1 presents the 4D representation used in the proposed method, along with an adaptive densification strategy that adjusts the densification threshold based on the relative motion gradient of the scene points. Subsequently, Section 3.2 provides a brief overview of the multi-view diffusion model used in the pipeline and introduces a novel spatial-temporal attention fusion mechanism designed to produce the almost consistent multi-view image sequences. Finally, as detailed in Section 3.3, the generated multi-view sequences are leveraged with spatial and temporal anchor frames for consistent 4D generation with the SDS optimization." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 544, + 256, + 556 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 256, + 556 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 256, + 556 + ], + "type": "text", + "content": "3.1 4D Representation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 563, + 482, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 563, + 482, + 612 + ], + "spans": [ + { + "bbox": [ + 130, + 563, + 482, + 612 + ], + "type": "text", + "content": "4D Gaussian Splatting The concept of 3D Gaussian Splatting was initially introduced in the work of [7], where an explicit point-based representation is used to model the 3D scene. Specifically, the 3D Gaussian point cloud, denoted as " + }, + { + "bbox": [ + 130, + 563, + 482, + 612 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 130, + 563, + 482, + 612 + ], + "type": "text", + "content": " for a static scene, is characterized by the tuple:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 267, + 620, + 481, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 267, + 620, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 267, + 620, + 481, + 633 + ], + "type": "interline_equation", + "content": "\\mathcal {S} = [ \\mathcal {X}, s, r, \\sigma , \\zeta ], \\tag {1}", + "image_path": "1e54d64e30e112f307670d5314b26448da3b3d1108d6ffaa7421f939f3c2871b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{X} = (x,y,z)" + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": " denotes the positions of the 3D Gaussian points, and " + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\zeta" + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": " represent the scale, rotation, opacity, and spherical harmonics (SH) coeffi-" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 409, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 409, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 409, + 91, + 447, + 100 + ], + "type": "text", + "content": "STAG4D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "cients of the radiance, respectively. In 3D reconstruction, the 3D Gaussian point cloud will be optimized through a point-based differentiable volume rendering." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 480, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 480, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 480, + 189 + ], + "type": "text", + "content": "This representation can be naturally extended to 4D scenes by presenting the continuous scene dynamics as a 3D motion field. Inspired by [33], we present a 4D scene as a 3D Gaussian point cloud with a hex-plane-based deformation field, denoted as " + }, + { + "bbox": [ + 130, + 141, + 480, + 189 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\mathcal{S},t)" + }, + { + "bbox": [ + 130, + 141, + 480, + 189 + ], + "type": "text", + "content": ". The 3D Gaussian point cloud at time " + }, + { + "bbox": [ + 130, + 141, + 480, + 189 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 141, + 480, + 189 + ], + "type": "text", + "content": " can be expressed as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 253, + 201, + 480, + 214 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 201, + 480, + 214 + ], + "spans": [ + { + "bbox": [ + 253, + 201, + 480, + 214 + ], + "type": "interline_equation", + "content": "\\mathcal {F} (\\mathcal {S}, t) = \\left[ \\mathcal {X} _ {t}, s _ {t}, r _ {t}, \\sigma , \\zeta \\right], \\tag {2}", + "image_path": "ae7539b84c189c5a014fed9581e3f1c208af51e3b1d62d94e9cabb1243c80f03.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 224, + 479, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 479, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 479, + 248 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 224, + 479, + 248 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_t = (x_t, y_t, z_t)" + }, + { + "bbox": [ + 130, + 224, + 479, + 248 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 224, + 479, + 248 + ], + "type": "inline_equation", + "content": "s_t" + }, + { + "bbox": [ + 130, + 224, + 479, + 248 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 224, + 479, + 248 + ], + "type": "inline_equation", + "content": "r_t" + }, + { + "bbox": [ + 130, + 224, + 479, + 248 + ], + "type": "text", + "content": " represent the updated information of Gaussian position, scale, and rotation at time " + }, + { + "bbox": [ + 130, + 224, + 479, + 248 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 224, + 479, + 248 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 270, + 480, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 480, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 480, + 450 + ], + "type": "text", + "content": "Adaptive Densification The standard 3D Gaussian Splatting technique typically employs a point cloud densification control strategy to dynamically adjust the number of Gaussians and their density within a unit volume. This adaptive approach allows for the transition from an initial sparse Gaussian set to a denser configuration to better represent the 3D scene. The 4D Gaussian Splatting method introduced in the work [33] utilizes a densification strategy similar to that presented in vanilla 3DGS [7], which involves the use of a fixed densification threshold based on the view-space position gradient. However, while the fixed gradient threshold strategy demonstrates efficacy in the context of reconstruction, particularly when multiple views afford robust and redundant coverage of the target scene, it performs sub-optimally in the generative setting. This limitation arises from the constraints imposed by the input single image or monocular videos, leading to significant uncertainty in the spatial and scale dimensions for each training object. This will lead to different optimal thresholds for different cases, as illustrated in the ablation study." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 451, + 481, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 451, + 481, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 451, + 481, + 582 + ], + "type": "text", + "content": "To address the issue, we propose an adaptive threshold approach, wherein only candidates with relatively large gradients are selected for densification. Our approach is based on the statistical analysis of the accumulated gradient of each point, which follows a log distribution of similar shapes throughout the training process. We apply an adaptive threshold that filters out Gaussians with a relatively small gradient and only densifies those with a large gradient. The adaptive threshold is set to select a fixed percentage (top " + }, + { + "bbox": [ + 130, + 451, + 481, + 582 + ], + "type": "inline_equation", + "content": "\\lambda\\%" + }, + { + "bbox": [ + 130, + 451, + 481, + 582 + ], + "type": "text", + "content": ") of points with the highest gradient in each densification operation. This ensures that the threshold adapts to the distribution of the gradient and maintains a stable relative position. Experimental results demonstrate that the proposed simple strategy can significantly enhance the quality and robustness of the 4D generation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 395, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 395, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 395, + 616 + ], + "type": "text", + "content": "3.2 Temporal and Multi-view Consistent Diffusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 479, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 479, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 479, + 666 + ], + "type": "text", + "content": "Multi-view Consistent Diffusion Our pipeline adopts Score Distillation Sampling(SDS) [22] from image diffusion models for the optimization of dynamic Gaussian. Specifically, Image-to-image diffusion model Zero123 [13] is utilized" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 211, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 211, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 211, + 102 + ], + "type": "text", + "content": "Zeng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 386, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 386, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 386, + 128 + ], + "type": "text", + "content": "and the SDS loss gradient could be formulated as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 187, + 135, + 480, + 162 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 135, + 480, + 162 + ], + "spans": [ + { + "bbox": [ + 187, + 135, + 480, + 162 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\theta} \\mathcal {L} _ {S D S} (\\phi , \\mathbf {x}) = \\mathbb {E} _ {t, \\epsilon} \\left[ \\omega (t) \\left(\\hat {\\epsilon} _ {\\theta} \\left(\\mathbf {z} _ {t}; \\mathbf {I} _ {i n}, \\mathbf {R}, \\mathbf {T}, t\\right) - \\epsilon\\right) \\frac {\\partial \\mathbf {x}}{\\partial \\theta} \\right], \\tag {3}", + "image_path": "450192a87814c646b878d9585f5cd1d6a5c769c05a0335d317d2daa9a765b313.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 244, + 163, + 358, + 175 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 163, + 358, + 175 + ], + "spans": [ + { + "bbox": [ + 244, + 163, + 358, + 175 + ], + "type": "interline_equation", + "content": "\\hat {z} = z _ {t} - \\sigma_ {t} \\hat {\\epsilon} _ {t} (z; \\mathbf {I} _ {i n}, \\mathbf {R}, \\mathbf {T}).", + "image_path": "94bdc210975b833d8d587308d26f3616ea08edf5349084c93943384e087f605a.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 182, + 482, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 482, + 338 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 482, + 338 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 182, + 482, + 338 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 132, + 182, + 482, + 338 + ], + "type": "text", + "content": " represents the parameters of the 3D representation, " + }, + { + "bbox": [ + 132, + 182, + 482, + 338 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 132, + 182, + 482, + 338 + ], + "type": "text", + "content": " the rendered image at the current view, " + }, + { + "bbox": [ + 132, + 182, + 482, + 338 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 132, + 182, + 482, + 338 + ], + "type": "text", + "content": " the timestamp in the diffusion process, " + }, + { + "bbox": [ + 132, + 182, + 482, + 338 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 132, + 182, + 482, + 338 + ], + "type": "text", + "content": " the ground truth noise, " + }, + { + "bbox": [ + 132, + 182, + 482, + 338 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}" + }, + { + "bbox": [ + 132, + 182, + 482, + 338 + ], + "type": "text", + "content": " the predicted noise from the noisy image " + }, + { + "bbox": [ + 132, + 182, + 482, + 338 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{\\mathbf{t}}" + }, + { + "bbox": [ + 132, + 182, + 482, + 338 + ], + "type": "text", + "content": " conditioned on an initial input " + }, + { + "bbox": [ + 132, + 182, + 482, + 338 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{in}" + }, + { + "bbox": [ + 132, + 182, + 482, + 338 + ], + "type": "text", + "content": ", and the relative camera pose between input view and target view " + }, + { + "bbox": [ + 132, + 182, + 482, + 338 + ], + "type": "inline_equation", + "content": "(\\mathbf{R},\\mathbf{T})" + }, + { + "bbox": [ + 132, + 182, + 482, + 338 + ], + "type": "text", + "content": ". Zero123 can generate the target view image at any relative camera location but only one target view at a time, thus good at optimizing the 3D object from all views whilst bad at generating spatially consistent images from multiple target views. In contrast, Zero123++ [26] leverages reference attention [37] to model the relationship of the images from multiple target views as well as the input view, resulting in multi-view consistency output yet at the cost of fixed target camera locations. We combine the advantages of both models by taking the multi-view consistent output of Zero123++ as the input images of Zero123 when calculating SDS loss." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 354, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 449 + ], + "type": "text", + "content": "Temporally Consistent Diffusion Following the view generation philosophy of multi-view diffusion based 3D generation, a practical solution to 4D content creation is to generate the multi-view videos of the 4D scene, and then apply the multi-view video conditioned SDS loss to optimize the scene. However, it is rather difficult to generate temporally consistent multi-view videos by the separate per-frame generation. Inspired by zero-shot video generation method [8], we propose a training-free temporal attention module to enable Zero123++ [26] with temporal-awareness." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 450, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 510 + ], + "type": "text", + "content": "In our design, we obtain the attention information during denoising the multiview images of the first frame. Then we apply the recorded attention to the later denoising process. Specifically, during the denoising process of the latent code, the self-attention layer computes three essential components—queries " + }, + { + "bbox": [ + 130, + 450, + 482, + 510 + ], + "type": "inline_equation", + "content": "(Q)" + }, + { + "bbox": [ + 130, + 450, + 482, + 510 + ], + "type": "text", + "content": ", keys " + }, + { + "bbox": [ + 130, + 450, + 482, + 510 + ], + "type": "inline_equation", + "content": "(K)" + }, + { + "bbox": [ + 130, + 450, + 482, + 510 + ], + "type": "text", + "content": ", and values " + }, + { + "bbox": [ + 130, + 450, + 482, + 510 + ], + "type": "inline_equation", + "content": "(V)" + }, + { + "bbox": [ + 130, + 450, + 482, + 510 + ], + "type": "text", + "content": ", which is formulated as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 216, + 517, + 480, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 517, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 216, + 517, + 480, + 544 + ], + "type": "interline_equation", + "content": "\\operatorname {S e l f - A t t n} (Q, K, V) = \\operatorname {S o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {c}}\\right) V. \\tag {4}", + "image_path": "90a355ae479458397033f2ea12706fd3b57bf7a38a5335111781477860823c6a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 549, + 482, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 549, + 482, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 549, + 482, + 597 + ], + "type": "text", + "content": "We first compute the key " + }, + { + "bbox": [ + 132, + 549, + 482, + 597 + ], + "type": "inline_equation", + "content": "K_{0}" + }, + { + "bbox": [ + 132, + 549, + 482, + 597 + ], + "type": "text", + "content": " and value " + }, + { + "bbox": [ + 132, + 549, + 482, + 597 + ], + "type": "inline_equation", + "content": "V_{0}" + }, + { + "bbox": [ + 132, + 549, + 482, + 597 + ], + "type": "text", + "content": " for the initial frame, " + }, + { + "bbox": [ + 132, + 549, + 482, + 597 + ], + "type": "inline_equation", + "content": "T_{0}" + }, + { + "bbox": [ + 132, + 549, + 482, + 597 + ], + "type": "text", + "content": ". Subsequently, for each subsequent frame " + }, + { + "bbox": [ + 132, + 549, + 482, + 597 + ], + "type": "inline_equation", + "content": "T_{t}" + }, + { + "bbox": [ + 132, + 549, + 482, + 597 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 132, + 549, + 482, + 597 + ], + "type": "inline_equation", + "content": "t \\in \\{1,\\dots,N\\}" + }, + { + "bbox": [ + 132, + 549, + 482, + 597 + ], + "type": "text", + "content": ", we perform a mixing operation on the key " + }, + { + "bbox": [ + 132, + 549, + 482, + 597 + ], + "type": "inline_equation", + "content": "K_{t}" + }, + { + "bbox": [ + 132, + 549, + 482, + 597 + ], + "type": "text", + "content": " and value " + }, + { + "bbox": [ + 132, + 549, + 482, + 597 + ], + "type": "inline_equation", + "content": "V_{t}" + }, + { + "bbox": [ + 132, + 549, + 482, + 597 + ], + "type": "text", + "content": " with the initial key " + }, + { + "bbox": [ + 132, + 549, + 482, + 597 + ], + "type": "inline_equation", + "content": "K_{0}" + }, + { + "bbox": [ + 132, + 549, + 482, + 597 + ], + "type": "text", + "content": " and value " + }, + { + "bbox": [ + 132, + 549, + 482, + 597 + ], + "type": "inline_equation", + "content": "V_{0}" + }, + { + "bbox": [ + 132, + 549, + 482, + 597 + ], + "type": "text", + "content": " derived from the first frame " + }, + { + "bbox": [ + 132, + 549, + 482, + 597 + ], + "type": "inline_equation", + "content": "T_{0}" + }, + { + "bbox": [ + 132, + 549, + 482, + 597 + ], + "type": "text", + "content": ". This process can be formally represented as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 249, + 605, + 480, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 605, + 480, + 635 + ], + "spans": [ + { + "bbox": [ + 249, + 605, + 480, + 635 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} K _ {t} = \\gamma \\mathbf {K} _ {0} + (1 - \\gamma) \\mathbf {K} _ {t} \\\\ V _ {t} = \\gamma \\mathbf {V} _ {0} + (1 - \\gamma) \\mathbf {V} _ {t}, \\end{array} \\right. \\tag {5}", + "image_path": "1a55b678e422fff1ed15de3d77da03b766878b19af0e77a8a15fab780a5e2204.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "where the parameter " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": " serves as a weighting factor that governs the influence of the initial frame, as illustrated in the ablation part. Additionally, we employ" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "text", + "content": "STAG4D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 118, + 476, + 273 + ], + "blocks": [ + { + "bbox": [ + 140, + 118, + 476, + 273 + ], + "lines": [ + { + "bbox": [ + 140, + 118, + 476, + 273 + ], + "spans": [ + { + "bbox": [ + 140, + 118, + 476, + 273 + ], + "type": "image", + "image_path": "03e8900e2be9d449b0b6bb36f4e8a47fb662986816617e26cfa5b1cf096879d8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 285, + 480, + 308 + ], + "lines": [ + { + "bbox": [ + 131, + 285, + 480, + 308 + ], + "spans": [ + { + "bbox": [ + 131, + 285, + 480, + 308 + ], + "type": "text", + "content": "Fig. 3: Illustration of the proposed spatial and temporal attention fusion during multiview sequence inference." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 332, + 482, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 332, + 482, + 417 + ], + "spans": [ + { + "bbox": [ + 130, + 332, + 482, + 417 + ], + "type": "text", + "content": "reference attention following the approach outlined in [26,37], to extract local conditioning from the input image. In this reference attention mechanism, both the key and the value are obtained through the concatenation of features from the reference image and the noisy input image. Our design of temporal attention module is simple yet effective with superior temporal consistency, image quality and 3D consistency to vanilla cross-frame attention proposed in [8], as demonstrated in experiment section." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 436, + 260, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 436, + 260, + 449 + ], + "spans": [ + { + "bbox": [ + 132, + 436, + 260, + 449 + ], + "type": "text", + "content": "3.3 Training Objectives" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 458, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 458, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 131, + 458, + 482, + 533 + ], + "type": "text", + "content": "Following the generation of multi-view sequences from a monocular reference video, we can acquire 6 anchor views " + }, + { + "bbox": [ + 131, + 458, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\{I_t^i\\}_{i\\in \\{1\\dots 6\\}}" + }, + { + "bbox": [ + 131, + 458, + 482, + 533 + ], + "type": "text", + "content": " and a reference view " + }, + { + "bbox": [ + 131, + 458, + 482, + 533 + ], + "type": "inline_equation", + "content": "I_{t}^{ref}" + }, + { + "bbox": [ + 131, + 458, + 482, + 533 + ], + "type": "text", + "content": " at each timestep " + }, + { + "bbox": [ + 131, + 458, + 482, + 533 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 458, + 482, + 533 + ], + "type": "text", + "content": ". In our optimization process, we employ multi-view score distillation sampling (SDS) utilizing the generated images " + }, + { + "bbox": [ + 131, + 458, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\{I_t^i\\}_{i = 1\\dots 6}" + }, + { + "bbox": [ + 131, + 458, + 482, + 533 + ], + "type": "text", + "content": " in conjunction with the reference images " + }, + { + "bbox": [ + 131, + 458, + 482, + 533 + ], + "type": "inline_equation", + "content": "I_{t}^{ref}" + }, + { + "bbox": [ + 131, + 458, + 482, + 533 + ], + "type": "text", + "content": ". The multi-view score distillation loss function " + }, + { + "bbox": [ + 131, + 458, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{MVSDS}" + }, + { + "bbox": [ + 131, + 458, + 482, + 533 + ], + "type": "text", + "content": " can be defined as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 204, + 540, + 480, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 204, + 540, + 480, + 572 + ], + "spans": [ + { + "bbox": [ + 204, + 540, + 480, + 572 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {M V S D S} = \\lambda_ {1} \\mathcal {L} _ {S D S} ^ {i} + \\lambda_ {2} \\mathcal {L} _ {S D S} ^ {r e f} \\\\ = \\lambda_ {1} \\mathcal {L} _ {S D S} \\left(\\phi , I _ {t} ^ {i}\\right) + \\lambda_ {2} \\mathcal {L} _ {S D S} \\left(\\phi , I _ {t} ^ {\\text {r e f}}\\right), \\tag {6} \\\\ \\end{array}", + "image_path": "c14b1d41fe7a6c0676c00cedfc12232b7eaab8af2e2bb1a517bbdfc80a36952d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 581, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 640 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 581, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 130, + 581, + 482, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 581, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 130, + 581, + 482, + 640 + ], + "type": "text", + "content": " are two weighting factors, the index " + }, + { + "bbox": [ + 130, + 581, + 482, + 640 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 581, + 482, + 640 + ], + "type": "text", + "content": " is determined based on the proximity of the rendering viewpoint to the viewpoint of the generated images. This selection process, which we refer to as multi-view score distillation sampling, involves choosing the nearest reference image to the rendered camera view for calculating the SDS loss." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": "Following the approach [29], we utilize the reference image to compute both the reconstruction loss " + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rec}" + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": " and the foreground mask loss " + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{mask}" + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": ". Thus, the final" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 211, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 211, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 211, + 102 + ], + "type": "text", + "content": "Zeng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 246, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 246, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 246, + 128 + ], + "type": "text", + "content": "optimization objective is:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 233, + 141, + 481, + 153 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 141, + 481, + 153 + ], + "spans": [ + { + "bbox": [ + 233, + 141, + 481, + 153 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {M V S D S} + \\lambda_ {3} \\mathcal {L} _ {r e c} + \\lambda_ {4} \\mathcal {L} _ {\\text {m a s k}}, \\tag {7}", + "image_path": "8d77365c04321663f8eb28181f0c7bf15fd6b70fa5e24cbae642d60744ca7ddc.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 165, + 481, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 165, + 481, + 201 + ], + "spans": [ + { + "bbox": [ + 131, + 165, + 481, + 201 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 165, + 481, + 201 + ], + "type": "inline_equation", + "content": "\\lambda_3" + }, + { + "bbox": [ + 131, + 165, + 481, + 201 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 165, + 481, + 201 + ], + "type": "inline_equation", + "content": "\\lambda_4" + }, + { + "bbox": [ + 131, + 165, + 481, + 201 + ], + "type": "text", + "content": " are the weighting parameters. During training, we first use " + }, + { + "bbox": [ + 131, + 165, + 481, + 201 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 131, + 165, + 481, + 201 + ], + "type": "text", + "content": " to supervise a fixed frame to get static canonical 3D Gaussian. Then we use all anchors and reference images to train the dynamic 4D Gaussian point cloud." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 222, + 270, + 233 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 222, + 270, + 233 + ], + "spans": [ + { + "bbox": [ + 132, + 222, + 270, + 233 + ], + "type": "text", + "content": "3.4 Text-to-4D Extension" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 245, + 482, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 245, + 482, + 365 + ], + "spans": [ + { + "bbox": [ + 131, + 245, + 482, + 365 + ], + "type": "text", + "content": "Our method is designed to be readily adaptable to text and image inputs, offering novel capabilities such as directly generating video sequences from textual descriptions or static images. We initiate this process using a 2D diffusion model SDXL [21], which generates the image from textual input. This 2D output is then transformed into a video sequence via a video diffusion model, such as SVD [2], adding temporal coherence and motion to the static image. Finally, the aforementioned pipeline further lifts the video sequence to a 4D scene, introducing an extra spatial dimension that creates immersive experiences beyond conventional 2D or 3D media. The generation pipeline with the given text, image, and video inputs is visualized in Fig. 2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 387, + 224, + 400 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 387, + 224, + 400 + ], + "spans": [ + { + "bbox": [ + 132, + 387, + 224, + 400 + ], + "type": "text", + "content": "4 Experiment" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 415, + 253, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 415, + 253, + 427 + ], + "spans": [ + { + "bbox": [ + 132, + 415, + 253, + 427 + ], + "type": "text", + "content": "4.1 Experiment Setup" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 437, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 437, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 131, + 437, + 482, + 533 + ], + "type": "text", + "content": "Dataset For the video-to-4D task, we utilize the dataset provided by Consistent4D [6] for quantitative evaluation, which comprises multi-view videos depicting 7 dynamic objects. Additionally, for qualitative evaluation, we curate a set of challenging videos from online sources to assess the robustness and generalization capabilities of each method. For text/image-to-4D tasks, we follow the data settings in 4Dfy [1] and DreamGaussian4D [23] to create the corresponding results. Notably, we use the same input image and input video with DreamGaussian4D for a fair comparison." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 533, + 482, + 666 + ], + "type": "text", + "content": "Evaluation metrics We employ CLIP, LPIPS, and FVD for Video-to-4D evaluation as in the previous approach [6]. Specifically, CLIP and LPIPS serve as image-level metrics, assessing the semantic similarity between rendered images and ground truths; FVD is a video-level metric commonly used in video generation tasks, considering not only single-frame quality but also temporal coherence; we also consider the FID-VID metric to measure the temporal consistency. As 4D-Gen [36] does not support a 30-frame setting we report its FVD under a 16-frame reconstruction, namely FVD-16 as an additional metrics. For text-to-4D generation, we follow previous works and provide a user study for quantitative comparisons between different methods. We also provide a user study for video-to-4D and both studies are detailed in the supplementary material." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "text", + "content": "STAG4D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 114, + 480, + 259 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 480, + 259 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 480, + 259 + ], + "type": "image", + "image_path": "739a36b04b574fcb05f5e51cdedbcc347c55f2297d53625a5e939a38ef6e8997.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 271, + 482, + 293 + ], + "lines": [ + { + "bbox": [ + 132, + 271, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 482, + 293 + ], + "type": "text", + "content": "Fig. 4: Qualitative comparison on video-to-4D generation. Our method achieves better faithfulness to the input video with higher quality on the overall reconstruction." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 321, + 479, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 321, + 479, + 394 + ], + "spans": [ + { + "bbox": [ + 130, + 321, + 479, + 394 + ], + "type": "text", + "content": "Baselines For video-to-4D generation, we compare our method with two baselines: Consistent4D [6] and the concurrent work 4DGen [36]. We utilize the official code released by the authors to generate comparison results. For text&image-to-4D generation, we make comparisons with two state-of-art models that are open-sourced: 4Dfy [1] and DreamGaussian4D [23]. Their results are generated using codes from their official GitHub repository." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 420, + 279, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 420, + 279, + 432 + ], + "spans": [ + { + "bbox": [ + 132, + 420, + 279, + 432 + ], + "type": "text", + "content": "4.2 Implementation Details" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "content": "In the training of the video-to-4D generative model, we adopt a two-stage approach. Initially, the model is initialized in canonical space and trained for 1000 steps. Subsequently, the deformation fields are learned over 7000 additional steps to accurately capture the dynamic scene. For the deformation process, we employ multi-layer perceptrons (MLPs) with 64 hidden layers and 32 hidden features per layer. The initial learning rate for the deformation MLPs is set to " + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "inline_equation", + "content": "1.6 \\times 10^{-4}" + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "content": " and is decayed to " + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "inline_equation", + "content": "1.6 \\times 10^{-6}" + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "content": " by the end of the training. In the context of adaptive densification, we opt to densify the top " + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "inline_equation", + "content": "2.5\\%" + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "content": " of points with the most accumulated gradient. In the generative phase, we utilize the spatial-temporal consistent videos for the multiview score distillation sampling optimization. Empirically, we set " + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "content": " to 0.5 for our temporally consistent diffusion. Regarding the loss functions, we maintain the SDS loss weight at 1 and adjust the weights of other losses accordingly. Specifically, the reconstruction loss is assigned to " + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "inline_equation", + "content": "4 \\times 10^{4}" + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "content": ", the mask loss is weighted at " + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "inline_equation", + "content": "1 \\times 10^{4}" + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "content": ". The training process requires approximately 1 hour on an RTX 3090 GPU, and the rendering process can be performed at 150 FPS in real time. For Text&Image-to-4D, we use SDXL [21] for image generation, and apply SVD [2] to the generated image to create a corresponding video. Then we use the same setting as Video-to-4D to transfer the video into 4D content." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 211, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 211, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 211, + 101 + ], + "type": "text", + "content": "Zeng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 116, + 478, + 281 + ], + "blocks": [ + { + "bbox": [ + 134, + 116, + 478, + 281 + ], + "lines": [ + { + "bbox": [ + 134, + 116, + 478, + 281 + ], + "spans": [ + { + "bbox": [ + 134, + 116, + 478, + 281 + ], + "type": "image", + "image_path": "1ed653f04bfc9197b1adfa0edb704b373dc56507d1104fd7559a2fb1a44a4531.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 291, + 482, + 314 + ], + "lines": [ + { + "bbox": [ + 132, + 291, + 482, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 291, + 482, + 314 + ], + "type": "text", + "content": "Fig. 5: Qualitative comparison on text-to-4D generation. Our method achieves the best results on the text alignment and visual quality." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 336, + 391, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 391, + 348 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 391, + 348 + ], + "type": "text", + "content": "4.3 Comparison with State-of-the-art Approaches" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 354, + 480, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 480, + 402 + ], + "type": "text", + "content": "In this section, we conduct a comprehensive comparison with the aforementioned baselines using synthetic and in-the-wild data. The superior performance demonstrated in both quantitative and qualitative evaluations serves to substantiate the effectiveness and robustness of our proposed method." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 402, + 481, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 481, + 533 + ], + "type": "text", + "content": "Quantitative Results on Video-to-4D The experimental results presented in Table 1 offer a comparative analysis of our method against state-of-the-art techniques, namely Consistent4D and 4DGen, across several metrics. Our method outperforms Consistent4D and 4DGen in CLIP, showing better semantic consistency with the target content. It also achieves the lowest LPIPS score, meaning it generates more realistic images than the others. For the video quality and smoothness, our method beats 4DGen in FID-VID and FVD-16, and Consistent4D in FVD, implying that our videos are closer to real videos and have less temporal artifacts. Overall, the numbers confirm that our method is superior in generating semantically aligned and perceptually convincing videos with high fidelity and coherence." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Qualitative Comparison on Video-to-4D In our investigation of the video-to-4D task, we have employed various methodologies to conduct qualitative comparisons. Figure 4 shows the rendered outcomes at distinct temporal intervals and perspectives. A critical observation of our findings reveals that Gaussian splatting is inadequate in producing credible and temporally consistent images during per-frame reconstruction. Furthermore, we have observed that Consistent4D often produces over-saturated and unrealistic patterns, which may stem from the over-restrictive geometry representation in their cascade Dynerf. Additionally, our analysis indicates that 4DGen struggles to generate reasonable motion for the Gaussians, leading to its points lacking deformation over time. Moreover, 4DGen fails to generate detailed textures for the surface, which appear" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 409, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 409, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 409, + 91, + 447, + 100 + ], + "type": "text", + "content": "STAG4D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 116, + 302, + 270 + ], + "blocks": [ + { + "bbox": [ + 135, + 116, + 302, + 270 + ], + "lines": [ + { + "bbox": [ + 135, + 116, + 302, + 270 + ], + "spans": [ + { + "bbox": [ + 135, + 116, + 302, + 270 + ], + "type": "image", + "image_path": "97944a9d0d08c11f6788a824ee1ef039af965363aefed34324797c2ce81367de.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 286, + 482, + 309 + ], + "lines": [ + { + "bbox": [ + 130, + 286, + 482, + 309 + ], + "spans": [ + { + "bbox": [ + 130, + 286, + 482, + 309 + ], + "type": "text", + "content": "Fig. 6: Ablation on spatial and temporal attention. We evaluate the effect of spatial and temporal attention on the reconstruction results." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 309, + 121, + 482, + 268 + ], + "blocks": [ + { + "bbox": [ + 309, + 121, + 482, + 268 + ], + "lines": [ + { + "bbox": [ + 309, + 121, + 482, + 268 + ], + "spans": [ + { + "bbox": [ + 309, + 121, + 482, + 268 + ], + "type": "image", + "image_path": "4b3a9d7d1b1b49693c72fc53048b60ece35c286e205dce6cb43f93927a8cbeaa.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 340, + 482, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 482, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 482, + 411 + ], + "type": "text", + "content": "to be over-smoothed in some cases. Conversely, our methodology demonstrates substantial enhancements in comparison to existing strategies, particularly in aspects of reconstruction fidelity and stability. Notably, our technique facilitates seamless color transitions, as exemplified in the Fish case, and renders a convincing low-poly effect, as demonstrated in the Trump case, underscoring our method's adaptability to diverse dynamic environments." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 415, + 483, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 483, + 571 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 483, + 571 + ], + "type": "text", + "content": "Qualitative Comparison on Text-to-4D We also evaluate our method on the Text&Image-to-4D task, which aims to generate 4D content from text and image inputs. Figure 5 shows the 4D generation results of our method and two baselines, 4Dfy [1] and DreamGaussian4D [23], under different views. As can be seen, 4Dfy [1] produces coarse results and fails to capture the complex semantics of the text input, e.g. in the Cat case. DreamGaussian4D [23] generates meshes with fine-grained details in the front view, but suffers from severe artifacts and distortions in the side view or back view. In contrast, our method synthesizes Gaussians with smooth and consistent geometry and realistic texture across different views, thanks to our attention mechanism that leverages spatial and temporal information. Our method demonstrates its superiority and versatility on the Text&Image-to-4D task, which can be seen as a direct application of our Video-to-4D framework." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 196, + 617, + 414, + 663 + ], + "blocks": [ + { + "bbox": [ + 196, + 597, + 414, + 608 + ], + "lines": [ + { + "bbox": [ + 196, + 597, + 414, + 608 + ], + "spans": [ + { + "bbox": [ + 196, + 597, + 414, + 608 + ], + "type": "text", + "content": "Table 1: Comparison with state-of-the-art methods." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 196, + 617, + 414, + 663 + ], + "lines": [ + { + "bbox": [ + 196, + 617, + 414, + 663 + ], + "spans": [ + { + "bbox": [ + 196, + 617, + 414, + 663 + ], + "type": "table", + "html": "
CLIP ↑LPIPS ↓FID-VID ↓FVD ↓FVD-16 ↓
Consistent4D0.8770.134/1133.93/
4DGen0.8940.13071.99/1005.72
Ours0.9090.12652.58992.21952.41
", + "image_path": "621be71528fb35961b53ee3fe9764fa9a36b4794d80f029f0f9a23d9cbb256c7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 211, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 211, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 211, + 102 + ], + "type": "text", + "content": "Zeng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 238, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 238, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 238, + 128 + ], + "type": "text", + "content": "4.4 Ablation Study" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 212 + ], + "type": "text", + "content": "The experimental analysis presented in Table 2 provides a comprehensive evaluation of the impact of various components on the performance of a diffusion model with a 4D representation. The components under consideration include baseline diffusion, spatial anchor, spatial-temporal anchor, and adaptive densification. In the first configuration of Table 2, we regard the model utilizing the diffusion process [13] with adaptive densification as the baseline." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 132, + 264, + 332, + 319 + ], + "blocks": [ + { + "bbox": [ + 132, + 232, + 332, + 255 + ], + "lines": [ + { + "bbox": [ + 132, + 232, + 332, + 255 + ], + "spans": [ + { + "bbox": [ + 132, + 232, + 332, + 255 + ], + "type": "text", + "content": "Table 2: Ablation study on effectiveness of each novel component." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 264, + 332, + 319 + ], + "lines": [ + { + "bbox": [ + 132, + 264, + 332, + 319 + ], + "spans": [ + { + "bbox": [ + 132, + 264, + 332, + 319 + ], + "type": "table", + "html": "
Diffusion Model4D Gaussian \nAdaptive Dens.Evaluation Metrics
BaselineSpat.Temp.CLIP ↑LPIPS ↓FID-VID ↓FVD ↓
✓✓0.8950.13577.881369.65
✓✓✓0.8990.13067.041198.23
✓✓✓0.8900.13690.301453.53
✓✓✓✓0.9090.12652.58992.21
", + "image_path": "413b9323868bf125b22eff1e5fe43843061d12907047d46e9f525514c22ac813.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 339, + 264, + 482, + 313 + ], + "blocks": [ + { + "bbox": [ + 339, + 232, + 481, + 254 + ], + "lines": [ + { + "bbox": [ + 339, + 232, + 481, + 254 + ], + "spans": [ + { + "bbox": [ + 339, + 232, + 481, + 254 + ], + "type": "text", + "content": "Table 3: Ablation study on different attention mechanisms." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 339, + 264, + 482, + 313 + ], + "lines": [ + { + "bbox": [ + 339, + 264, + 482, + 313 + ], + "spans": [ + { + "bbox": [ + 339, + 264, + 482, + 313 + ], + "type": "table", + "html": "
Evaluation Metrics
CLIPLPIPSFID-VIDFVD
Spat. only0.8990.13067.041198.23
Cross-Frame0.9010.12963.321053.25
Spat.-Temp.0.9090.12652.58992.21
", + "image_path": "0e2e05c78a84716fd4fa3e1901371869caaac689a3f60f86cfeaddda3670d794.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 341, + 479, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 341, + 479, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 341, + 479, + 411 + ], + "type": "text", + "content": "Spatial and Temporal Anchors Table 2 shows that adding the spatial anchor to the baseline model leads to a marginal improvement in all metrics, especially FID-VID and FVD. The final model, which uses both spatial-temporal anchor and adaptive densification, improves all metrics significantly. It has the best CLIP, LPIPS, FID-VID, and FVD scores, meaning it generates the most semantically consistent, realistic, and smooth videos." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 414, + 482, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 482, + 508 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 482, + 508 + ], + "type": "text", + "content": "Figure 6 shows the importance of our spatial-temporal anchor for improving the 4D content quality. The baseline 4D Gaussian method, labeled as \"w./o. spatial,\" fails to deform the point clouds properly, leading to unreasonable shapes without spatial-temporal constraints. A model that uses only the spatial anchor, \"w./o. temporal,\" struggles to deform solid shapes, as seen in the rabbit's leg and the frog's back. Our method, which combines both spatial and temporal anchors, overcomes these issues and produces visually consistent and accurate 4D content." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "text", + "content": "Adaptive Densification We present the statistical analysis of the adaptive densification module in the third and final configurations of Table 2. The results show that the adaptive densification module improves the overall quality of the generative model, especially in terms of the FID-VID and FVD metrics. We also compare the proposed adaptive densification with a fixed threshold approach in Figure 7b. We observe that a fixed threshold can lead to either under-densification or over-densification, depending on the case. For example, in the Bird case, the fixed threshold is too low for densification, which results in excessive Gaussian points. On the other hand, in the Duck case, the fixed threshold is too high for densification, which causes the foot to be blurry due to insufficient Gaussian points. Our method overcomes this limitation by adapting the gradient threshold to different cases, thus producing robust and stable 4D Gaussian results." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "text", + "content": "STAG4D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 115, + 477, + 287 + ], + "blocks": [ + { + "bbox": [ + 133, + 115, + 477, + 287 + ], + "lines": [ + { + "bbox": [ + 133, + 115, + 477, + 287 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 477, + 287 + ], + "type": "image", + "image_path": "e1442b050157ece5ea308ebdf48253713e9f5707d1a4febf8799d5f677f773f4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 310, + 482, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 310, + 482, + 490 + ], + "spans": [ + { + "bbox": [ + 130, + 310, + 482, + 490 + ], + "type": "text", + "content": "Different Attention Mechanisms In our study, we evaluated three distinct attention mechanisms within the context of images generated by our diffusion model: spatial attention, cross-frame attention, and spatial-temporal attention. The visual outcomes of these varied approaches are illustrated in Figure 7a. We employed a parameter " + }, + { + "bbox": [ + 130, + 310, + 482, + 490 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 310, + 482, + 490 + ], + "type": "text", + "content": " to modulate the extent of spatial and temporal attention exerted, as delineated in Equation 5. Our findings indicate that exclusive reliance on spatial attention does not guarantee consistency across sequential frames. While cross-frame attention enhances frame-to-frame coherence, it fails to achieve a satisfactory alignment with reference images. Conversely, the integration of spatial-temporal attention ensures consistency across both dimensions. We quantitatively assessed the quality of images generated under different attention regimes using four established metrics: CLIP, LPIPS, FID-VID, and FVD. The comparative results are tabulated in Table 3. It was observed that images generated with spatial-temporal attention outperformed others across all evaluation metrics, which underscores the efficacy of our proposed mechanism." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 509, + 220, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 509, + 220, + 521 + ], + "spans": [ + { + "bbox": [ + 132, + 509, + 220, + 521 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "The paper presents a novel approach for dynamic 3D content generation from monocular videos, addressing the challenges of 4D representation and spatial-temporal consistency. By leveraging specially tailored 4D Gaussian splatting and a novel information fusion module, the proposed method achieves high-quality and robust 4D scene generation. Comprehensive experiments demonstrate the method's effectiveness, showcasing a faster generation speed and significant improvements in rendering quality and temporal consistency compared to prior state-of-the-art methods. Overall, the proposed method sets a new benchmark for training speed, rendering quality, and 4D consistency in dynamic 3D content generation from monocular videos, opening up possibilities for real-world applications." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 211, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 211, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 211, + 101 + ], + "type": "text", + "content": "Zeng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 137, + 445, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 137, + 445, + 162 + ], + "spans": [ + { + "bbox": [ + 132, + 137, + 445, + 162 + ], + "type": "text", + "content": "This work was supported by the National Key R&D Program of China (2022YFF0902200) and NSFC grant 62441204." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 178, + 198, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 178, + 198, + 190 + ], + "spans": [ + { + "bbox": [ + 132, + 178, + 198, + 190 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 201, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 201, + 481, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 201, + 481, + 243 + ], + "spans": [ + { + "bbox": [ + 138, + 201, + 481, + 243 + ], + "type": "text", + "content": "1. Bahmani, S., Skorokhodov, I., Rong, V., Wetzstein, G., Guibas, L., Wonka, P., Tulyakov, S., Park, J.J., Tagliasacchi, A., Lindell, D.B.: 4d-fy: Text-to-4d generation using hybrid score distillation sampling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7996-8006 (2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 245, + 481, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 245, + 481, + 277 + ], + "spans": [ + { + "bbox": [ + 138, + 245, + 481, + 277 + ], + "type": "text", + "content": "2. Blattmann, A., Dockhorn, T., Kulal, S., Mendelevitch, D., Kilian, M., Lorenz, D., Levi, Y., English, Z., Voleti, V., Letts, A., Jampani, V., Rombach, R.: Stable video diffusion: Scaling latent video diffusion models to large datasets (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 277, + 481, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 277, + 481, + 309 + ], + "spans": [ + { + "bbox": [ + 138, + 277, + 481, + 309 + ], + "type": "text", + "content": "3. Du, Y., Zhang, Y., Yu, H.X., Tenenbaum, J.B., Wu, J.: Neural radiance flow for 4d view synthesis and video processing. In: 2021 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 14304-14314. IEEE Computer Society (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 309, + 481, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 309, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 309, + 481, + 342 + ], + "type": "text", + "content": "4. Gao, C., Saraf, A., Kopf, J., Huang, J.B.: Dynamic view synthesis from dynamic monocular video. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5712-5721 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 342, + 481, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 342, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 138, + 342, + 481, + 373 + ], + "type": "text", + "content": "5. Ho, J., Chan, W., Sahara, C., Whang, J., Gao, R., Gritsenko, A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., et al.:Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 374, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 374, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 138, + 374, + 481, + 407 + ], + "type": "text", + "content": "6. Jiang, Y., Zhang, L., Gao, J., Hu, W., Yao, Y.: Consistent4d: Consistent 360 " + }, + { + "bbox": [ + 138, + 374, + 481, + 407 + ], + "type": "inline_equation", + "content": "\\{\\backslash \\mathrm{deg}\\}" + }, + { + "bbox": [ + 138, + 374, + 481, + 407 + ], + "type": "text", + "content": " dynamic object generation from monocular video. arXiv preprint arXiv:2311.02848 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 407, + 481, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 407, + 481, + 438 + ], + "spans": [ + { + "bbox": [ + 138, + 407, + 481, + 438 + ], + "type": "text", + "content": "7. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG) 42(4), 1-14 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 438, + 481, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 438, + 481, + 471 + ], + "spans": [ + { + "bbox": [ + 138, + 438, + 481, + 471 + ], + "type": "text", + "content": "8. Khachatryan, L., Movsisyan, A., Tadevosyan, V., Henschel, R., Wang, Z., Navasardyan, S., Shi, H.: Text2video-zero: Text-to-image diffusion models are zero-shot video generators. arXiv preprint arXiv:2303.13439 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 471, + 481, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 471, + 481, + 503 + ], + "spans": [ + { + "bbox": [ + 138, + 471, + 481, + 503 + ], + "type": "text", + "content": "9. Li, Z., Niklaus, S., Snavely, N., Wang, O.: Neural scene flow fields for space-time view synthesis of dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 503, + 481, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 503, + 481, + 546 + ], + "spans": [ + { + "bbox": [ + 138, + 503, + 481, + 546 + ], + "type": "text", + "content": "0. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3d: High-resolution text-to-3d content creation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 300–309 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 547, + 481, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 547, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 138, + 547, + 481, + 568 + ], + "type": "text", + "content": "1. Lin, Y., Dai, Z., Zhu, S., Yao, Y.: Gaussian-flow: 4d reconstruction with dynamic 3d gaussian particle. arXiv:2312.03431 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 568, + 481, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 568, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 138, + 568, + 481, + 600 + ], + "type": "text", + "content": "2. Ling, H., Kim, S.W., Torralba, A., Fidler, S., Kreis, K.: Align your gaussians: Text-to-4d with dynamic 3d gaussians and composed diffusion models. arXiv preprint arXiv:2312.13763 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 601, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 481, + 633 + ], + "type": "text", + "content": "3. Liu, R., Wu, R., Van Hoorick, B., Tokmakov, P., Zakharov, S., Vondrick, C.: Zero-1-to-3: Zero-shot one image to 3d object. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9298-9309 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "type": "text", + "content": "4. Liu, Y., Lin, C., Zeng, Z., Long, X., Liu, L., Komura, T., Wang, W.: Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453 (2023)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "text", + "content": "STAG4D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "15. Long, X., Guo, Y.C., Lin, C., Liu, Y., Dou, Z., Liu, L., Ma, Y., Zhang, S.H., Habermann, M., Theobalt, C., et al.: Wonder3d: Single image to 3d using cross-domain diffusion. arXiv preprint arXiv:2310.15008 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 183 + ], + "type": "text", + "content": "16. Lu, Y., Zhang, J., Li, S., Fang, T., McKinnon, D., Tsin, Y., Quan, L., Cao, X., Yao, Y.: Direct2. 5: Diverse text-to-3d generation via multi-view 2.5 d diffusion. arXiv preprint arXiv:2311.15980 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "type": "text", + "content": "17. Lu, Y., Zhang, J., Li, S., Fang, T., McKinnon, D., Tsin, Y., Quan, L., Cao, X., Yao, Y.: Direct2.5: Diverse text-to-3d generation via multi-view 2.5d diffusion (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "text", + "content": "18. Luiten, J., Kopanas, G., Leibe, B., Ramanan, D.: Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "text", + "content": "19. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 303 + ], + "type": "text", + "content": "20. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph. 41(4), 102:1-102:15 (Jul 2022). https://doi.org/10.1145/3528223.3530127, https://doi.org/10.1145/3528223.3530127" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "type": "text", + "content": "21. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "type": "text", + "content": "22. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "type": "text", + "content": "23. Ren, J., Pan, L., Tang, J., Zhang, C., Cao, A., Zeng, G., Liu, Z.: Dreamgaussian4d: Generative 4d gaussian splatting. arXiv preprint arXiv:2312.17142 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "text", + "content": "24. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 414, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 480, + 445 + ], + "type": "text", + "content": "25. Shen, T., Gao, J., Yin, K., Liu, M.Y., Fidler, S.: Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. In: Advances in Neural Information Processing Systems (NeurIPS) (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "type": "text", + "content": "26. Shi, R., Chen, H., Zhang, Z., Liu, M., Xu, C., Wei, X., Chen, L., Zeng, C., Su, H.: Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "type": "text", + "content": "27. Shi, Y., Wang, P., Ye, J., Mai, L., Li, K., Yang, X.: Mvdream: Multi-view diffusion for 3d generation. arXiv:2308.16512 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 502, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 480, + 533 + ], + "type": "text", + "content": "28. Singer, U., Sheynin, S., Polyak, A., Ashual, O., Makarov, I., Kokkinos, F., Goyal, N., Vedaldi, A., Parikh, D., Johnson, J., et al.: Text-to-4d dynamic scene generation. arXiv preprint arXiv:2301.11280 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "type": "text", + "content": "29. Tang, J., Ren, J., Zhou, H., Liu, Z., Zeng, G.: Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 555, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 480, + 599 + ], + "type": "text", + "content": "30. Wang, H., Du, X., Li, J., Yeh, R.A., Shakhnarovich, G.: Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12619-12629 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 600, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 621 + ], + "type": "text", + "content": "31. Wang, J., Yuan, H., Chen, D., Zhang, Y., Wang, X., Zhang, S.: Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 622, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 480, + 665 + ], + "type": "text", + "content": "32. Wang, X., Wang, Y., Ye, J., Wang, Z., Sun, F., Liu, P., Wang, L., Sun, K., Wang, X., He, B.: Animatabledreamer: Text-guided non-rigid 3d model generation and reconstruction with canonical score distillation. arXiv preprint arXiv:2312.03795 (2023)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 211, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 211, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 211, + 102 + ], + "type": "text", + "content": "Zeng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 347 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "33. Wu, G., Yi, T., Fang, J., Xie, L., Zhang, X., Wei, W., Liu, W., Tian, Q., Wang, X.: 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "text", + "content": "34. Xian, W., Huang, J.B., Kopf, J., Kim, C.: Space-time neural irradiance fields for free-viewpoint video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9421-9431 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 482, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 482, + 215 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 482, + 215 + ], + "type": "text", + "content": "35. Yang, Z., Gao, X., Zhou, W., Jiao, S., Zhang, Y., Jin, X.: Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. arXiv preprint arXiv:2309.13101 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 216, + 482, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 216, + 482, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 482, + 237 + ], + "type": "text", + "content": "36. Yin, Y., Xu, D., Wang, Z., Zhao, Y., Wei, Y.: 4dgen: Grounded 4d content generation with spatial-temporal consistency. arXiv preprint arXiv:2312.17225 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 237, + 482, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 482, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 482, + 258 + ], + "type": "text", + "content": "37. Zhang, L.: Reference-only control (2023), https://github.com/Mikubill/sd-webui-controlnet/discussions/1236" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "type": "text", + "content": "38. Zhang, S., Wang, J., Zhang, Y., Zhao, K., Yuan, H., Qing, Z., Wang, X., Zhao, D., Zhou, J.: I2vgen-xl: High-quality image-to-video synthesis via cascaded diffusion models (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 293, + 482, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 315 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 315 + ], + "type": "text", + "content": "39. Zhao, Y., Yan, Z., Xie, E., Hong, L., Li, Z., Lee, G.H.:Animate124: Animating one image to 4d dynamic scene. arXiv preprint arXiv:2311.14603 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 315, + 482, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 315, + 482, + 347 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 482, + 347 + ], + "type": "text", + "content": "40. Zheng, Y., Li, X., Nagano, K., Liu, S., Hilliges, O., De Mello, S.: A unified approach for text-and image-guided 4d scene generation. arXiv preprint arXiv:2311.16854 (2023)" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "text", + "content": "STAG4D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/STAMP_ Outlier-Aware Test-Time Adaptation with Stable Memory Replay/20b10b92-8cd4-4e35-ad35-a50ac0ff44cf_content_list.json b/2024/STAMP_ Outlier-Aware Test-Time Adaptation with Stable Memory Replay/20b10b92-8cd4-4e35-ad35-a50ac0ff44cf_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3d4ca5fb318eac11031b559eed928f8e338e0278 --- /dev/null +++ b/2024/STAMP_ Outlier-Aware Test-Time Adaptation with Stable Memory Replay/20b10b92-8cd4-4e35-ad35-a50ac0ff44cf_content_list.json @@ -0,0 +1,1631 @@ +[ + { + "type": "image", + "img_path": "images/2661f00c92b9e29c0adf43c07a615933123144781e6f9d47b5962a503944823d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 145, + 326, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "STAMP: Outlier-Aware Test-Time Annotation with Stable Memory Replay", + "text_level": 1, + "bbox": [ + 330, + 154, + 727, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yongcan Yu $^{1,2}$ , Lijun Sheng $^{1,3}$ , Ran He $^{1,2}$ , and Jian Liang $^{1,2\\star}$", + "bbox": [ + 254, + 223, + 746, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ NLPR & MAIS, Institute of Automation, Chinese Academy of Sciences", + "bbox": [ + 243, + 251, + 754, + 265 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ School of Artificial Intelligence, University of Chinese Academy of Sciences", + "bbox": [ + 243, + 265, + 756, + 279 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 University of Science and Technology of China", + "bbox": [ + 339, + 279, + 663, + 294 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{yuyongcan0223, liangjian92}@gmail.com", + "bbox": [ + 354, + 295, + 648, + 308 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Test-time adaptation (TTA) aims to address the distribution shift between the training and test data with only unlabeled data at test time. Existing TTA methods often focus on improving recognition performance specifically for test data associated with classes in the training set. However, during the open-world inference process, there are inevitably test data instances from unknown classes, commonly referred to as outliers. This paper pays attention to the problem that conducts both sample recognition and outlier rejection during inference while outliers exist. To address this problem, we propose a new approach called STAble Memory rePlay (STAMP), which performs optimization over a stable memory bank instead of the risky mini-batch. In particular, the memory bank is dynamically updated by selecting low-entropy and label-consistent samples in a class-balanced manner. In addition, we develop a self-weighted entropy minimization strategy that assigns higher weight to low-entropy samples. Extensive results demonstrate that STAMP outperforms existing TTA methods in terms of both recognition and outlier detection performance. The code is released at https://github.com/yuyongcan/STAMP.", + "bbox": [ + 259, + 345, + 743, + 595 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: online test-time adaptation $\\cdot$ outlier detection $\\cdot$ memory bank $\\cdot$ self-weighted entropy minimization", + "bbox": [ + 259, + 609, + 740, + 638 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 662, + 375, + 679 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep neural networks (DNNs) [46, 48] have achieved remarkable performance across various tasks. Nevertheless, the impressive performance is based on the assumption that training data and test data come from the same distribution. However, in real-world scenarios, the presence of distribution shifts is ubiquitous [13, 30], leading to a degradation in the prediction performance [7]. Taking autonomous driving [60] as an example, changing weather conditions causes systems trained on laboratory data to make incorrect decisions. To mitigate the impact of distribution shift, test-time adaptation (TTA) [29, 58] has emerged as", + "bbox": [ + 212, + 694, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author.", + "bbox": [ + 217, + 824, + 385, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/afdec46572993664b26a61c6f1c8adaaed02f7115bbb7ecdd56e1c0e1a343085.jpg", + "image_caption": [ + "Fig. 1: The illustration of outlier-aware test-time adaptation. The test data stream consists of both normal samples and outliers. When the test data arrives, the algorithm adapts the model and provides two outputs, a prediction and an OOD score. OOD score indicates the likelihood that a sample is an outlier. The detector in the deployment scenario chooses to keep the prediction or reject samples according to the OOD score." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 785, + 292 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "a promising approach, adapting models with unlabeled samples during inference. Nowadays, TTA has achieved remarkable results in various tasks, such as image classification [51, 62], semantic segmentation [3, 49], object detection [1].", + "bbox": [ + 212, + 428, + 782, + 473 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Previous TTA works [5, 38, 59] always assume that test data and training data share the same class space. Nevertheless, in realistic open-world scenarios, test data streams will inevitably contain samples from new semantic classes, also known as outliers. The introduction of outliers brings two additional risks to adaptation tasks: (i) Classifying outliers into known categories poses safety risks, especially in contexts like autonomous driving, where systems should request human intervention to handle unknown objects to avoid traffic accidents. (ii) Self-supervised loss (e.g., entropy minimization, pseudo-labeling) calculated with outliers may misguide the optimization process, decreasing the model's recognition ability on known classes.", + "bbox": [ + 212, + 474, + 787, + 626 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To better tackle the above risks, we pay attention to a more realistic TTA scenario termed outlier-aware test-time adaptation. Outlier-aware test-time adaptation aims to conduct recognition for normal samples and reject to make predictions for outliers. The overview of outlier-aware test-time adaptation is illustrated in Figure 1. When test data mixed with outliers arrives, the algorithm adapts the model and generates the prediction and an out-of-distribution (OOD) score for each sample. Then we decide to keep the prediction or reject the sample based on its OOD score. The sample with a higher OOD score is more likely to be considered an outlier.", + "bbox": [ + 212, + 627, + 787, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the outlier-aware test-time adaptation, we additionally design a novel algorithm called STAble Memory rePlay (STAMP). STAMP consists of three components: reliable class-balanced memory, self-weighted entropy minimization, and stable optimization strategy. Reliable class-balanced memory stores reliable samples for optimization by filtering samples based on their pre", + "bbox": [ + 212, + 763, + 787, + 842 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Yu et al.", + "bbox": [ + 271, + 114, + 331, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "diction's consistency and entropy. To maintain the class-balanced prediction of the model, it dynamically discards samples with the class that participates most frequently in optimization when the memory is full. Furthermore, we adopt self-weighted entropy minimization to assign greater weights to low-entropy samples in memory and adjust the weights for each sample adaptively. To ensure the stability of the optimization process, we dynamically adjust the optimization step size to quickly adapt to the target domain at the early stage and mitigate error accumulation later. Moreover, we apply the sharpness-aware minimization technique [14] to further reduce the impact of noisy gradients.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To demonstrate the effectiveness of STAMP, we compare it with eight state-of-the-art TTA algorithms. We select three commonly used datasets in TTA, including CIFAR10-C, CIFAR100-C [10, 20], and ImageNet-C [9, 20], alongside some outlier datasets (i.e., Noise, SVHN [37], LSUN [57], TinyImageNet [25], Textures [6], Places365 [64]). Experimental results show that our method achieves superior performance on recognition and outlier detection under the outlier-aware TTA setting. The experiment on conventional protocol indicates that STAMP also performs well in the scenario without outliers. In summary, we list our contributions as follows:", + "bbox": [ + 212, + 282, + 787, + 417 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce outlier-aware test-time adaptation, a TTA scenario considering both normal sample recognition and outlier rejection during inference.", + "- We propose a memory that dynamically stores reliable samples for replay. To prevent noisy samples from entering the memory, we propose a filtering strategy based on the consistency and entropy of the model's prediction.", + "- We design self-weighted entropy minimization, which guides the model to adaptively learn to assign higher weights to valuable samples.", + "- We propose a proper metric that considers both generalization and outlier detection performance. Extensive experiments compared with eight TTA algorithms on several benchmarks demonstrate the effectiveness of our method." + ], + "bbox": [ + 225, + 428, + 784, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 601, + 387, + 617 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Test-Time Adaptation", + "text_level": 1, + "bbox": [ + 215, + 633, + 444, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "TTA intends to adapt the model pre-trained on the source domain to the target domain while the distribution shift exists. Compared to domain generalization [47, 52] and domain adaptation [15, 53], TTA needs to access neither the source data nor the target labels, making it has a wide range of applications [11, 12, 45]. TTA methods can be mainly divided into two solutions: batch normalization (BN) calibration and self-training. In BN calibration, Nado et al. [36] utilize the BN statistics of the test batch $(\\mu_t, \\sigma_t)$ to conduct normalization in the BN layer, while Schneider [43] propose a more comprehensive approach considering statistics from both the source domain $(\\mu_s, \\sigma_s)$ and the target domain to calibrate the domain shifts. To conquer the correlation sampling problem, Yuan et al. [59] adopt robust batch normalization, which updates the statistic $(\\mu_t, \\sigma_t)$ in a moving average manner with the test data stream.", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Outlier-Aware Test-Time Adaptation with Stable Memory Replay", + "bbox": [ + 292, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Currently, the primary paradigm in TTA is self-training [2, 61], where Wang et al. [51] adopt entropy minimization as a self-supervised loss for optimization and Chen et al. [5] introduce contrastive loss to TTA. However, unsupervised loss may contain noise information. Niu et al. [38] first propose filtering out high-entropy samples, a mechanism adopted by subsequent works [17, 39]. Subsequently, more attention has been paid to addressing more realistic adaptation scenarios such as continually changing target domains [4, 54], label shift [41, 65], and temporally correlated shift [16, 63]. In our study, we focus on a more realistic scenario where test data contains samples with new semantic classes. Some existing works [17, 26, 28] discuss the TTA algorithm in scenarios containing outliers. However, Lee et al. [26] and Gong et al. [17] solely focus on the generalization ability to domain shift, neglecting the issue of rejecting the outliers. A recent work [28] discusses outlier detection performance based on TTA tasks, but it only provides binary predictions to indicate whether the test data is outlier. In contrast, outlier-aware test-time adaptation is more flexible, providing an OOD score, and different systems can conduct sample rejection based on security requirements and the score.", + "bbox": [ + 212, + 146, + 787, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Out-of-Distribution Detection", + "text_level": 1, + "bbox": [ + 215, + 433, + 508, + 448 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "OOD detection aims to identify instances that significantly deviate from the distribution of the training data [31]. Their primary focus is usually on detecting samples whose labels were not seen during the training phase. This area has been extensively researched, with one common approach being the utilization of confidence scores such as max softmax probability [21], energy score [33, 34], and others [27] derived from the model's output to detect OOD samples. Further improvements have been made by Liang et al. [32] through techniques like temperature scaling and input preprocessing to enhance the effectiveness of confidence scores. Another avenue for OOD sample detection lies in training-based methods [10, 50, 55]. These techniques involve training a branch using both indistribution data and OOD data to enable the model to distinguish OOD data during training. However, these methods alter the training process of the source model, which can not be directly applied to adaptation tasks based on a well-trained model. Moreover, OOD detection traditionally assumes that both the training and test data originate from the same domain, a condition not met in TTA scenarios. Under the realistic scenario, there may be a domain shift between test data and training data, which brings greater challenges to detection.", + "bbox": [ + 212, + 465, + 787, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 215, + 753, + 380, + 772 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we first introduce the outlier-aware test-time adaptation problem in Section 3.1. Then we describe our proposed method, including how to build a memory bank and conduct the optimization in Section 3.2.", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Yu et al.", + "bbox": [ + 271, + 114, + 330, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Outlier-Aware Test-Time Adaptation", + "text_level": 1, + "bbox": [ + 215, + 146, + 568, + 161 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Let $\\mathcal{D}_S$ be the source domain dataset with label space $\\mathcal{C}_S$ follows a distribution $P_S(x,y)$ . Similarly, let $\\mathcal{D}_T = \\{(x_i,y_i)\\}_{i=1}^{N_T}$ be the target domain dataset with label space $\\mathcal{C}_T$ follows a distribution $P_T(x,y)$ . Given a source model $f_{\\theta_s}(x)$ well trained on $\\mathcal{D}_S$ with parameter $\\theta_s$ , TTA aims to adapt the source model to the target domain while $P_S(x) \\neq P_T(x)$ .", + "bbox": [ + 212, + 169, + 782, + 243 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In traditional closed-set TTA, it is commonly assumed that the $\\mathcal{D}_S$ and $\\mathcal{D}_T$ share the same label space, i.e., $\\mathcal{C}_S = \\mathcal{C}_T$ . However, in outlier-aware test-time adaptation, we consider $\\mathcal{C}_S \\subseteq \\mathcal{C}_T$ while some outliers with new semantic classes will be encountered in $\\mathcal{D}_T$ . The illustration of outlier-aware test-time adaption is shown in Figure 1. In addition to adapting the model and providing the prediction, we also generate an OOD score $\\delta_i$ for each sample $x_i$ . Then we adopt an outlier detector $g(\\cdot)$ as follow:", + "bbox": [ + 212, + 244, + 784, + 349 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ng \\left(x _ {i}\\right) = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} \\delta_ {i} \\geq \\delta_ {t h r} \\\\ 0, & \\text {i f} \\delta_ {i} < \\delta_ {t h r}, \\end{array} \\right. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 405, + 358, + 784, + 398 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\delta_{thr}$ is a threshold that can be adjusted according to the requirements of different application scenarios. The samples with $g(x_i) = 1$ will be considered outliers and rejected for recognition. And we output the prediction for samples with $g(x_i) = 0$ . In our work, we take the entropy of the model prediction as our OOD score.", + "bbox": [ + 212, + 405, + 787, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Stable Memory Replay", + "text_level": 1, + "bbox": [ + 214, + 500, + 454, + 516 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we introduce stable memory replay (STAMP). STAMP first uses a filtering mechanism to filter out unreliable samples in the coming mini-batch and stores reliable samples in memory. Then, STAMP optimizes the model using in-memory samples by applying self-weighted entropy minimization. In order to further improve the stability of optimization, STAMP uses test-time step size decay and sharpness-aware minimization to guide optimization. The illustration of STAMP is shown in Figure 2.", + "bbox": [ + 212, + 522, + 782, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Reliable Class-Balanced Memory. Since excluding unreliable samples from optimization has shown great effectiveness [17, 38], we propose reliable class-balanced memory to store reliable samples for optimization. To have a more robust estimation of the sample's reliability, we get its prediction in an augmentation-averaged way. Moreover, we prevent unreliable samples from getting into the memory bank $\\mathcal{M}$ by employing consistency filtering and entropy filtering. To preserve the model's balanced class prior distribution, we dynamically discard samples with the highest class frequency from the memory bank.", + "bbox": [ + 212, + 628, + 782, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To ensure robust predictions, we adopt the augmentation-averaged method, which addresses the issue of deep neural networks being overconfident with out-of-distribution samples. This overconfidence often blurs the line between in-distribution and out-of-distribution samples. However, research has demonstrated that employing test-time augmentation [44, 61] enhances model calibration [40]. Thus, when presented with a test sample $x$ , we derive its prediction", + "bbox": [ + 212, + 750, + 784, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Outlier-Aware Test-Time Adaptation with Stable Memory Replay", + "bbox": [ + 292, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4241aedecb67cdbdce38260bb146b30410a5af6c66a8456efd2282f25356f565.jpg", + "image_caption": [ + "Fig. 2: The overview of STAMP. STAMP retains reliable samples for optimizing the network by maintaining a memory bank. When test samples arrive, STAMP generates an augmentation-averaged prediction. Two filtering mechanisms are designed to filter out inconsistent and high-entropy samples and add the remaining to the memory. A frequency vector is maintained to select samples to discard when the memory is full. STAMP leverages the reliable samples in the memory to optimize the model with the SAM optimizer alongside a decaying step strategy." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 782, + 372 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "through the following process:", + "bbox": [ + 215, + 532, + 434, + 547 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {p} (x) = \\frac {1}{K} \\sum_ {k = 1} ^ {K} f _ {\\theta_ {t}} \\left(\\tau_ {k} (x)\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 408, + 553, + 785, + 595 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\tau(\\cdot)$ is the operation of augmentation, $K$ is the total number of augmented views. By averaging predictions over augmented samples, the model can better eliminate the uncertainty inherent in its predictions, improving calibration and obtaining more reliable confidence estimates. In the supplementary material, we discuss the impact of the augmentation-averaged estimation.", + "bbox": [ + 212, + 602, + 787, + 678 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After acquiring the prediction $\\hat{p}$ , we implement filtering strategies to prevent unreliable samples from being added to memory for two reasons. Firstly, optimization using unreliable samples can result in error accumulation over time [38]. Secondly, we prevent outliers from entering the memory and reduce the entropy of normal samples in the memory. This action facilitates the model's ability to distinguish between normal samples and outliers. Since reliable samples that are beneficial to optimization tend to have the same pseudo-label as the source model, we employ consistency filtering to remove samples whose predictions are inconsistent with those of the source model:", + "bbox": [ + 212, + 679, + 787, + 813 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nm _ {c o n} (x) = \\mathbb {I} \\left(\\arg \\max _ {c} [ \\hat {p} (x) ] _ {c} = \\arg \\max _ {c} [ f _ {\\theta_ {s}} (x) ] _ {c}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 316, + 816, + 785, + 844 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Yu et al.", + "bbox": [ + 271, + 114, + 331, + 126 + ], + "page_idx": 5 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Reliable Class-Balanced Memory" + ], + "code_body": "Input: memory bank $\\mathcal{M}$ , a batch of data $\\{x_{i}\\}_{i = 1}^{B}$ , entropy threshold $\\mathcal{H}_{thr}$ , class frequency $\\{\\xi_c\\}_{c = 1}^{|C_S|}$ . \nfor $i = 1$ to $B$ do \nCalculate $\\hat{p}(x_i)$ by Equation (2) and $\\hat{y}_i = \\arg \\max_c [\\hat{p}(x_i)]_c$ . Calculate $m_{con}(x_i)$ , $m_{ent}(x_i)$ of $x_i$ by Equation (3), (4). \nif $m_{con}(x_i) = 1$ and $m_{ent}(x_i) = 1$ then \nif $\\mathcal{M}$ is full then \nDiscard one sample with highest class frequency $\\xi$ from $\\mathcal{M}$ . \nend if \nadd $(x_i, \\hat{y}_i)$ into $\\mathcal{M}$ . \nend if \nend for \nUpdate $\\xi$ with Equation (5).", + "bbox": [ + 217, + 164, + 787, + 352 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathbb{I}(\\cdot)$ is the indicator function. Moreover, including high entropy samples for optimization can harm the stability of the model [38, 39], while samples with low entropy always possess greater value and reliability. Therefore, we use entropy filtering to filter out samples with high entropy from entering the memory.", + "bbox": [ + 212, + 381, + 782, + 441 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nm _ {e n t} (x) = \\mathbb {I} \\left(\\mathcal {H} (\\hat {p} (x)) < \\mathcal {H} _ {t h r}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 385, + 450, + 785, + 477 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathcal{H}(q) = -\\sum_{c} q_c \\log q_c$ denotes the entropy of the prediction vector $q$ . We filter out samples with $\\mathcal{H}(\\hat{p}) \\geq \\mathcal{H}_{thr}$ , where $\\mathcal{H}_{thr}$ is a threshold. Considering both filtering strategies, we add sample $(x, \\hat{y})$ into memory bank $\\mathcal{M}$ if the sample satisfies consistency filtering and entropy filtering simultaneously, i.e., $m_{con}(x) = 1$ and $m_{ent}(x) = 1$ .", + "bbox": [ + 212, + 484, + 782, + 560 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Since the number of samples stored in memory is limited, we design a class-balanced strategy to choose a sample to discard when the memory is full. Since we adopt entropy minimization as our optimization objective, an issue arises when the model excessively optimizes samples from a single category over a short period. It leads to a bias towards predicting the dominant category. To counteract this bias, we dynamically discard samples from the memory that belong to the category with the highest update frequency. After each optimization of the model, we update the class frequencies in a moving average manner:", + "bbox": [ + 212, + 561, + 784, + 681 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\xi_ {c} ^ {t} = (1 - \\beta) \\xi_ {c} ^ {t - 1} + \\beta \\nu_ {c} ^ {t}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 413, + 691, + 785, + 710 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\nu_{c}^{t}$ represents the number of samples with class label $c$ in memory and $\\beta$ is moving average coefficient hyperparameter. While the memory is full, we discard a sample from the class with the highest $\\xi$ . We summarize the work process of reliable class-balanced memory in Algorithm 1.", + "bbox": [ + 212, + 718, + 782, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Self-Weighted Entropy Minimization. After updating the memory, we replay the reliable samples stored in the memory $\\mathcal{M}$ to optimize the model. Following [51], we employ entropy minimization as our loss function. This self-supervised loss encourages the model to exhibit higher confidence in its predic", + "bbox": [ + 212, + 779, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Outlier-Aware Test-Time Adaptation with Stable Memory Replay", + "bbox": [ + 292, + 113, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 2 The Pipeline of STAMP" + ], + "code_body": "Initialize: Memory bank $\\mathcal{M} = \\emptyset$ zero-initialized class frequency $\\{\\xi_c\\}_{c = 1}^{|C_S|}$ \nInput: Unlabeled test samples $\\mathcal{D}_T = \\{x_i\\}_{i = 1}^{NT}$ , a trained model $f_{\\theta_s}(\\cdot)$ \nfor each online mini-batch $X_{b}$ do Calculate $\\hat{p} (x)$ by Equation (2) and $\\mathcal{H}(\\hat{p} (x))$ for all $x\\in X_b$ . Obtain the prediction $\\hat{y} = \\arg \\max_{c}[\\hat{p} (x)]_{c}$ for all $x\\in X_b$ Update memory bank $\\mathcal{M}$ by Algorithm 1. Compute $\\mathcal{L}_{sem}$ with $\\mathcal{M}$ by Equation (6). Update $\\theta_t$ with SAM in Equation (7) and decay step size in Equation (8). \nend for \nOutput: The prediction and OOD score $\\{\\hat{y}_i,\\mathcal{H}(\\hat{p} (x_i))\\}_{i = 1}^{NT}$", + "bbox": [ + 228, + 165, + 743, + 310 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tions. In order to give more valuable samples a higher importance in optimization, we weight samples in $\\mathcal{M}$ in a self-weighted manner. Given samples $\\{x_{i}\\}_{i = 1}^{|M|}$ , we calculate the following loss function:", + "bbox": [ + 212, + 344, + 787, + 393 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s e m} (\\mathcal {M}, \\theta_ {t}) = \\sum_ {x _ {i} \\in \\mathcal {M}} \\frac {w _ {i}}{\\sum_ {j} w _ {j}} \\cdot \\mathcal {H} \\left(f _ {\\theta_ {t}} \\left(x _ {i}\\right)\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 354, + 402, + 785, + 438 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $w_{i} = e^{-\\mathcal{H}(f_{\\theta_{t}}(x_{i}))}$ denotes the weight of sample $x_{i}$ . It's worth noting that $w_{i}$ is also involved with the gradient back-propagation. The reason for this adaptive weighting is two-fold: (i) it gives greater weights to samples with lower entropy, which are considered more reliable. (ii) The model needs to learn the weight of each sample and adaptively give important samples higher weights, instead of solely minimizing entropy for all samples.", + "bbox": [ + 212, + 452, + 787, + 545 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Stable Optimization Strategy. To optimize the model in a stable way, we employ sharpness-aware minimization (SAM) [14] and step size decay in parameter updating. SAM is a technique that aims to improve generalization performance in deep learning models by minimizing a loss function while simultaneously controlling the sharpness of the loss landscape. Previous work [14] demonstrates that SAM exhibits resilience against noisy labels, which is also a prevalent challenge in TTA. Therefore, we consider SAM as a good choice of the optimizer in the adaptation scenario [17, 39]. Mathematically, this involves minimizing the following objective function:", + "bbox": [ + 212, + 545, + 789, + 681 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta_ {t}} \\mathcal {L} _ {s e m} ^ {S A} (\\mathcal {M}, \\theta_ {t}), \\text {w h e r e} \\mathcal {L} _ {s e m} ^ {S A} (\\mathcal {M}, \\theta_ {t}) \\triangleq \\max _ {| | \\epsilon | | _ {2} \\leq \\rho} \\mathcal {L} _ {s e m} (\\mathcal {M}, \\theta_ {t} + \\epsilon), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 267, + 693, + 785, + 720 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\rho$ represents the radius of the neighborhood. SAM minimizes the maximum loss within the parameters' vicinity.", + "bbox": [ + 212, + 732, + 787, + 762 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Recent research [26, 42] has revealed that under long-term adaptation, TTA algorithms may suffer from substantial error accumulation and consequent performance degradation. This issue becomes even more severe when test data contains outliers. To mitigate this problem, we introduce a test-time step size (a.k.a. learning rate) decay strategy. To achieve this, we update the model's parameters", + "bbox": [ + 212, + 763, + 787, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Yu et al.", + "bbox": [ + 271, + 114, + 331, + 127 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7b3035ccfaa5e959292840a993697c635d13c3ddee7018816b43cc80a585d197.jpg", + "table_caption": [ + "Table 1: Results (\\%) on CIFAR10-C and CIFAR100-C datasets under outlier-aware TTA setting across four outlier datasets. We report the accuracy (ACC), AUC score, and H-score. All the results are averaged over 15 different corruptions. The results in bold denote the best performance." + ], + "table_footnote": [], + "table_body": "
NoiseSVHN-CLSUN-CTinyImageNet-CAvg.
ACCAUCH-scoreACCAUCH-scoreACCAUCH-scoreACCAUCH-scoreACCAUCH-score
CIFAR10-CSource57.370.462.357.367.461.157.362.859.657.364.559.457.366.360.6
BN Stats [36]72.968.670.678.775.376.979.479.479.479.072.975.877.574.075.7
Tent [51]77.448.759.780.854.965.181.262.370.481.165.672.480.157.966.9
EATA [38]72.968.570.678.875.376.979.479.479.478.973.175.977.574.175.7
SAR [39]72.968.570.678.775.376.979.479.479.479.072.975.877.574.075.7
CoTTA [54]77.362.467.381.678.680.182.284.283.281.975.378.480.875.177.2
RoTTA [54]77.674.375.678.476.077.278.879.579.178.673.375.878.375.876.9
SoTTA [17]77.851.761.679.372.875.979.877.978.879.672.675.979.168.873.1
OWTTT [28]62.364.458.566.175.369.663.178.968.556.358.856.262.069.363.2
STAMP77.980.778.882.379.080.583.584.984.282.675.178.681.679.980.5
CIFAR100-CSource35.843.138.035.849.440.135.858.243.235.857.142.735.852.041.0
BN Stats [36]45.880.958.452.772.560.953.773.862.053.268.659.751.374.060.3
Tent [51]47.955.851.254.470.461.255.472.462.755.068.660.953.266.859.0
EATA [38]55.286.167.158.175.665.658.877.266.758.670.764.057.777.465.9
SAR [39]57.588.668.959.265.261.960.573.566.360.872.165.959.574.965.8
CoTTA [54]47.083.459.953.773.261.854.376.963.654.568.160.452.475.461.4
RoTTA [59]47.954.049.447.367.055.348.369.556.747.865.555.047.864.054.1
SoTTA [17]54.453.352.853.670.360.754.470.861.453.968.460.154.165.758.8
OWTTT [28]47.170.356.253.974.362.354.573.562.554.268.560.452.471.660.4
STAMP57.997.572.663.780.070.963.780.771.263.974.468.762.383.170.8
", + "bbox": [ + 233, + 212, + 769, + 450 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "$\\theta_{t}$ using cosine annealing decay step size:", + "bbox": [ + 214, + 479, + 511, + 493 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {t + 1} = \\theta_ {t} - \\alpha_ {t} \\nabla \\mathcal {L} _ {\\text {s e m}} ^ {S A} (\\mathcal {M}, \\theta_ {t}), \\text {w h e r e} \\alpha_ {t} = \\frac {1}{2} \\alpha_ {0} (1 + \\cos (\\frac {t}{T} \\pi)), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 277, + 503, + 784, + 532 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\alpha_{t}$ is the step size at timestamp $t$ and $T$ is a hyper-parameter to control the step size decay. While step size decay is commonly utilized in model training [56], we leverage it during testing for a distinct purpose. In the early stages of TTA, the model should adapt swiftly due to the significant disparity between the source and target domains. However, as adaptation progresses and the model acquires comprehensive distributional information about the target domain, it should reduce parameter updates to prevent error accumulation and enhance stability. We summarize the pipeline of STAMP in Algorithm 2.", + "bbox": [ + 212, + 542, + 787, + 666 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 689, + 375, + 707 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Setup", + "text_level": 1, + "bbox": [ + 214, + 722, + 310, + 738 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Datasets. Following previous TTA methods [16, 18, 54], we adopt three corruption datasets as normal datasets (i.e., in-distribution datasets). CIFAR10-C and CIFAR100-C [20, 24] stand out as the most widely used benchmarks in TTA tasks, comprising 10/100 categories with 15 types of corruptions respectively. Each type of corruption has 10,000 images. Similarly, ImageNet-C [9, 20] serves as a significant benchmark, consisting of 1,000 categories with the same 15 types", + "bbox": [ + 212, + 750, + 787, + 842 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Outlier-Aware Test-Time Adaptation with Stable Memory Replay", + "bbox": [ + 292, + 114, + 730, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/74968869a8c03d406750a6f97af169933a5c888734434c786b5c5b9d7b38130c.jpg", + "table_caption": [ + "Table 2: Results (\\%) on ImageNet-C under outlier-aware TTA setting across two outlier datasets. All the results are averaged over 15 different corruptions." + ], + "table_footnote": [], + "table_body": "
MethodPlaces365-CTextures-C
ACCAUCH-scoreACCAUCH-score
Source18.261.626.118.254.625.8
BN Stats [36]31.167.741.131.661.240.7
Tent [51]34.951.839.53948.642.0
EATA [38]46.472.656.046.462.252.8
SAR [39]44.973.355.045.667.054.0
CoTTA [54]33.866.943.534.260.742.8
RoTTA [59]36.668.646.537.065.346.5
SoTTA [17]41.767.850.741.860.348.8
OWTTT [28]9.154.013.99.459.414.6
STAMP46.476.257.246.570.055.6
", + "bbox": [ + 320, + 184, + 683, + 328 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "of corruption. For out-of-distribution datasets, we choose Noise, SVHN [37], LSUN [57], and TinyImageNet [25] for the CIFAR10/100-C datasets, and Textures [6] and Places365 [64] for ImageNet-C. To maintain consistency in domain shift between outliers and normal datasets, we apply the same corruptions $^4$ on outlier datasets. (e.g., SVHN-C refers to the corruptions version of SVHN dataset.) Throughout all experiments, we maintain a fixed ratio of 8:2 between normal samples and outliers. Further details of the outlier datasets are provided in the supplementary material.", + "bbox": [ + 212, + 357, + 787, + 478 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Baselines. To have a comprehensive comparison of existing TTA methods, we select eight TTA algorithms as our baselines. Source refers to the source pre-trained model. BN Stats [36] adjusts batch normalization (BN) layers [22] with the statistics of test data. Tent [51] optimizes the BN affine parameters in the model by entropy minimization. CoTTA [54] calculates weighted-average and augmentation-averaged pseudo-label to distill knowledge from the teacher model to the student model, and stochastically restores parameters to prevent the model from catastrophic forgetting. EATA [38] adopts the entropy and diversity weight with elastic weight consolidation [23] for anti-forgetting. SAR [39] excludes high-entropy samples from sharpness-aware entropy minimization and reset the model when it overfits. RoTTA [59] uses robust batch normalization and category-balanced sampling with timeliness and uncertainty to deal with the correlated label problem. SoTTA [17] applies high-confidence uniform-class sampling and entropy-sharpness minimization to alleviate the impact of noisy samples. OWTTT [28] proposes prototype expansion with prototype clustering and distribution alignment.", + "bbox": [ + 212, + 479, + 787, + 720 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation Metrics. We evaluate the above baselines for both generalization and outlier detection. For generalization, we adopt classification accuracy (ACC) as a metric. For outlier detection, we consider the area under the ROC curve (AUC) [8], a threshold-free metric to measure the binary classification problem. The ROC curve depicts the relationship between the true positive rate (TPR) and the false positive rate (FPR). Models with better detection ability", + "bbox": [ + 212, + 722, + 787, + 813 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Yu et al.", + "bbox": [ + 271, + 114, + 330, + 126 + ], + "page_idx": 9 + }, + { + "type": "page_footnote", + "text": "4 https://github.com/hendrycks/robustness", + "bbox": [ + 217, + 823, + 537, + 840 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/fb3b3cef1563476ffe91510603cef8fc13444f5ec1cd428e35d69a2f1332eb2f.jpg", + "table_caption": [ + "Table 3: Results (\\%) on CIFAR10/100-C datasets under outlier-aware TTA setting. We divide datasets into the normal part and the outlier part. In detail, we choose 8/80 classes as known samples and other 2/20 classes as outliers for CIFAR10/100-C. All the results are averaged over 15 different corruptions." + ], + "table_footnote": [], + "table_body": "
MethodCIFAR10-C (8:2)CIFAR100-C (80:20)
ACCAUCH-scoreACCAUCH-score
Source60.661.560.037.160.644.2
BN Stats [36]79.466.972.655.566.260.2
Tent [51]80.165.972.360.465.662.8
EATA [38]81.567.974.161.367.464.2
SAR [39]82.366.073.263.469.166.1
CoTTA [54]81.865.472.557.166.361.3
RoTTA [59]79.462.469.850.663.756.2
SoTTA [17]82.762.771.361.468.264.6
OWTTT [28]65.963.564.456.266.060.6
STAMP85.068.776.066.070.768.2
", + "bbox": [ + 333, + 212, + 671, + 345 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "will have a higher AUC score. To consider both generalization and detection performance simultaneously, we also report H-score:", + "bbox": [ + 212, + 371, + 782, + 402 + ], + "page_idx": 10 + }, + { + "type": "equation", + "text": "\n$$\nH - s c o r e = \\frac {2 \\cdot A C C \\cdot A U C}{A C C + A U C} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 401, + 409, + 785, + 440 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Implementation Details. We adopt the corruption with severity 5 in all experiments. We consider ResNet-18 [19] as backbones for CIFAR10-C, CIFAR100-C and ResNet-50 [19] for ImageNet-C. For CIFAR10-C and CIFAR100-C, we train the source model using standard cross-entropy loss with the initial learning rate of 0.1 and multi-step learning rate scheduling. For ImageNet-C, we use the released checkpoint from TorchVision library [35]. For a fair comparison, we set the batch size for all methods as 64 and the capacity of our memory bank is 64. Concerning the hyperparameters, we search the hyperparameters on a validation set and set the learning rate to $0.1 / 0.05 / 0.01$ , $\\mathcal{H}_{thr}$ to $0.25 / 0.9 / 0.8 \\times \\ln |\\mathcal{C}_S|$ , $T$ to $150 / 150 / 750$ for CIFAR10-C/CIFAR100-C/ImageNet-C. Moreover, we set the other hyperparameter the same in experiments with $\\beta = 0.1$ , $\\rho = 0.05$ , $K = 16$ . For parameters' updating, we select the affine parameters in the BN layer for optimization and freeze the top layer's parameter following [39]. More details of hyper-parameters searching and setting of baselines and STAMP can be found in the supplementary material.", + "bbox": [ + 212, + 448, + 792, + 676 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Results", + "text_level": 1, + "bbox": [ + 215, + 695, + 323, + 709 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Results under outlier-aware TTA setting. We first evaluate TTA methods on the datasets with outliers. The results for CIFAR datasets are illustrated in Table 1. We observe that a simple baseline method, Tent [51], has a competitive generalization performance on distribution shift (achieving an accuracy of $80.1\\%$ on CIFAR10-C). However, it exhibits substantial degradation in the AUC score compared to Source, for instance, dropping from $70.4\\%$ to $48.7\\%$ on CIFAR10-C with Noise outlier dataset. This decline could be attributed to Tent's indiscriminate minimization of entropy across all samples, leading to a gradual loss of its", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Outlier-Aware Test-Time Adaptation with Stable Memory Replay", + "bbox": [ + 292, + 114, + 730, + 130 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/2542c034bffd43d1a1d9f69c312cc7af93f238a7cf2befa283fddd2c70fe7a87.jpg", + "table_caption": [ + "Table 4: Classification accuracies (%) on CIFAR10-C and CIFAR100-C under closed-set TTA setting." + ], + "table_footnote": [], + "table_body": "
MethodGaussianshotinputdefocusglassrotationzonationshowfrostlogbrightnesscontrastelasticpixelatejpegAvg.
CIFAR10-CSource28.735.224.257.349.066.464.876.762.973.490.331.478.846.574.757.3
BN Stats [36]70.272.063.687.666.585.986.982.180.584.190.885.477.378.674.379.0
Tent [51]76.178.470.587.870.186.887.584.882.085.191.186.679.482.778.581.8
CoTTA [54]76.177.673.987.971.986.987.582.882.085.290.885.779.380.478.381.8
EATA [38]76.577.971.789.170.487.489.085.284.286.091.588.479.983.978.582.6
SAR [39]70.772.166.887.668.285.986.982.180.584.190.885.477.378.674.379.4
RoTTA [59]69.671.062.987.567.085.986.982.479.684.791.273.578.377.974.978.2
SoTTA [17]75.879.271.589.470.687.889.085.684.087.292.487.479.984.579.082.9
OWTTT [28]70.172.063.086.966.585.786.782.780.984.591.283.077.976.874.978.9
STAMP80.982.977.287.674.986.687.785.985.988.190.487.280.386.782.684.3
CIFAR100-CSource12.414.57.236.044.745.145.249.541.636.963.313.257.523.846.635.8
BN Stats [36]41.141.338.963.151.560.863.851.253.553.264.459.058.458.047.653.7
Tent [51]52.352.147.766.956.164.365.358.358.760.067.862.161.863.054.559.4
CoTTA [54]47.047.845.159.654.059.561.353.355.052.962.650.358.161.853.154.8
EATA [38]50.753.548.167.055.664.867.059.159.260.467.763.961.863.354.659.8
SAR [39]55.155.051.268.458.266.067.460.360.861.969.865.563.666.256.861.7
RoTTA [59]35.936.633.860.647.157.760.848.042.250.859.232.153.852.344.447.7
SoTTA [17]51.653.847.466.956.965.368.158.960.160.169.463.162.362.854.860.1
OWTTT [28]41.642.838.863.452.661.664.853.354.854.565.858.460.157.649.354.6
STAMP57.258.552.869.961.468.170.163.363.964.872.269.966.569.259.064.4
", + "bbox": [ + 233, + 184, + 769, + 431 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "ability to discern outliers. In addition, our experimental findings demonstrate that with judicious parameter selection, TTA methods can enhance not only the generalization capability against distribution shift but also the detection ability for outliers. Compared to existing TTA methods, STAMP achieves the most favorable performance on both classification and outlier detection. On CIFAR10-C, STAMP achieves the highest ACC, and $+4.1\\%$ huge improvements in AUC compared to the previous methods. Regarding CIFAR100-C, STAMP exhibits $+2.8\\%$ and $+5.7\\%$ improvements compared to the most effective TTA methods on ACC and AUC, respectively. We also report the results on the ImageNet-C in Table 2. STAMP still achieves the highest performance. Specifically, we outperform existing TTA methods with $+1.2\\%$ , $+1.6\\%$ on H-score with Places365-C and Textures-C outlier datasets, respectively.", + "bbox": [ + 212, + 460, + 787, + 641 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Additionally, we evaluate the TTA algorithm under a more challenging scenario involving homogeneous outlier datasets in Table 3, which makes it more difficult to identify outliers. STAMP also exhibits the most promising performance on CIFAR10-C and CIFAR100-C under this scenario.", + "bbox": [ + 212, + 642, + 787, + 702 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Concerning the above results, we achieve optimal results in most cases, affirming the robustness and effectiveness of STAMP.", + "bbox": [ + 212, + 703, + 787, + 733 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Results under closed-set TTA setting. Additionally, we assess our method in a scenario without outliers as most TTA works do. We present the classification accuracy on the clean dataset in Table 4. Remarkably, our method attains the highest accuracy on both CIFAR10-C and CIFAR100-C datasets, achieving an average improvement of $+2.9\\%$ over the second-best method (SoTTA). This shows that STAMP still has strong applicability even in the traditional situation without outliers.", + "bbox": [ + 212, + 734, + 802, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Yu et al.", + "bbox": [ + 271, + 114, + 330, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/7ab798d67bc1fe82b92f9f7727cda7668d26ff796bfbe5dee96e3da0a63577b7.jpg", + "table_caption": [ + "Table 5: Ablation studies on reliable class-balanced memory (RBM), self-weighted entropy minimization (S.W.), sharpness-aware minimization (S.A.), and decay step size (D.S.). Experiments are conducted on CIFAR100-C with the outlier dataset SVHN-C. The first row in the table reports the results of entropy minimization." + ], + "table_footnote": [], + "table_body": "
S.A.D.S.RBMS.W.ACCAUCH-score
----54.470.461.2
-✓--58.965.361.8
✓---56.472.863.5
✓✓--59.270.864.4
✓✓-✓59.377.767.1
✓✓✓-63.075.068.4
✓✓✓✓63.780.070.9
", + "bbox": [ + 361, + 212, + 643, + 295 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "To summarize, all the above results prove that our method achieves the best results under both outlier-aware TTA and conventional closed-set settings.", + "bbox": [ + 212, + 324, + 782, + 354 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.3 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 378, + 388, + 393 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effect of each component. We delve into the role of each component within our method as follows. First, we provide the results of different weight strategies of entropy minimization in Figure 3. We compare our adaptive weight with static weight and EATA weight [38]. EATA calculates the weight by $\\exp[\\mathcal{H}_{thr} - \\mathcal{H}(p(x))]$ and static weight calculates weight in the same way of Equation (6). Different from adaptive weight, both EATA and static weight don't involve any gradient back-propagation. As depicted in Figure 3, adaptive weight shows the best performance among the strategies.", + "bbox": [ + 212, + 405, + 787, + 526 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Furthermore, we conduct an ablation study on each component of our method, as summarized in Table 5. We observe that the effect of combining test-time step size decay and SAM is better than using either of them alone. Step size decay can", + "bbox": [ + 212, + 527, + 787, + 574 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/631e1fc72f95b0203dde42cd3276886b19169289d5e6f67a9cdb90b679ca47d6.jpg", + "image_caption": [ + "Fig. 3: Results for the different weight strategies on dataset CIFAR100-C with SVHN-C outlier dataset. Static weight represents that the $w_{i}$ in Equation (6) without gradient back-propagation. EATA weight represents the weight for reliable samples used in EATA [38]." + ], + "image_footnote": [], + "bbox": [ + 370, + 614, + 617, + 758 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Outlier-Aware Test-Time Adaptation with Stable Memory Replay", + "bbox": [ + 292, + 114, + 730, + 130 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/49a40b826ed8945b8d1a88d6924286e40193df24250940a61e21de6dc09fa88e.jpg", + "table_caption": [ + "Table 6: H-score (\\%) with the different ratios of outliers to total samples on CIFAR100-C with different outlier datasets." + ], + "table_footnote": [], + "table_body": "
SVHN-CLSUN-CTinyImageNet-C
5%68.470.268.3
10%70.170.968.8
20%70.971.268.7
33%70.370.168.1
50%69.368.867.4
", + "bbox": [ + 361, + 184, + 638, + 262 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "improve the ACC but result in AUC decline. However, combining SAM and step size decay mitigates the decline in AUC. In addition, after the introduction of self-weighted entropy minimization, the AUC of the model further increased to $77.7\\%$ . Notably, by removing reliable class-balanced memory from our method, the ACC decreases from $63.7\\%$ to $59.3\\%$ . It reveals that reliable class-balanced memory significantly contributes to enhancing the generalization ability to distribution shift. Moreover, by removing the self-weighted entropy minimization in STAMP, its AUC drops from $80.0\\%$ to $75.0\\%$ . Self-weighted entropy minimization plays an important role in outlier detection. Besides decreasing the sample entropy, self-weighted entropy minimization incorporates entropy into the optimization weight, allowing outliers that deviate from the majority of samples to receive lower weights, consequently increasing their entropy. This mechanism enhances outlier detection performance.", + "bbox": [ + 212, + 289, + 784, + 484 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Effect of outlier proportion. In the realistic scenario, the proportion of outliers can vary considerably. We investigate how this variability affects the performance of our method. Specifically, we manipulate the proportion of outliers, ranging from $5\\%$ to $50\\%$ , and present the corresponding H-scores in Table 6. The results show that the fluctuation of our method is within $3\\%$ . The findings suggest that STAMP exhibits robustness across different scenarios with varying proportions of outliers, as it does not experience significant degradation.", + "bbox": [ + 212, + 486, + 785, + 592 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 614, + 359, + 631 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this study, we research the challenge of handling outliers in test-time adaptation. To tackle this issue effectively, we introduce a scenario called outlier-aware test-time adaptation, which refers to conducting both recognition and outlier rejection simultaneously. Additionally, we propose STAMP, using augmentation-averaged estimation and a filtering mechanism to store reliable samples in memory. Moreover, we dynamically discard samples with the highest class frequency from memory to maintain a class-balanced prediction. Then we replay the samples stored in the memory with self-weighted entropy minimization. Furthermore, we leverage sharpness-aware minimization and learning rate decay to bolster the stability of our algorithm. Through extensive experiments across multiple benchmarks, we demonstrate that our method surpasses existing TTA methods, underscoring its effectiveness and versatility in open-world applications.", + "bbox": [ + 212, + 647, + 784, + 828 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Yu et al.", + "bbox": [ + 271, + 114, + 330, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 215, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "We thank Yuhe Ding and Zhengbo Wang for their valuable feedback on our work. This work was funded by the Beijing Municipal Science and Technology Project (No. Z231100010323005), the Beijing Nova Program (No. Z211100002121108), the Young Elite Scientists Sponsorship Program by CAST (2023QNRC001), and the National Natural Science Foundation of China under (No. 62276256).", + "bbox": [ + 212, + 175, + 787, + 251 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 273, + 321, + 289 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. An, S., Park, S., Kim, G., Baek, J., Lee, B., Kim, S.: Context enhanced transformer for single image object detection. arXiv preprint arXiv:2312.14492 (2023) 2", + "2. Boudiaf, M., Mueller, R., Ben Ayed, I., Bertinetto, L.: Parameter-free online test-time adaptation. In: Proc. CVPR (2022) 4", + "3. Cao, H., Xu, Y., Yang, J., Yin, P., Yuan, S., Xie, L.: Multi-modal continual test-time adaptation for 3d semantic segmentation. In: Proc. ICCV (2023) 2", + "4. Chakrabarty, G., Sreenivas, M., Biswas, S.: Santa: Source anchoring network and target alignment for continual test time adaptation. Transactions on Machine Learning Research (2023) 4", + "5. Chen, D., Wang, D., Darrell, T., Ebrahimi, S.: Contrastive test-time adaptation. In: Proc. CVPR (2022) 2, 4", + "6. Cimpoi, M., Maji, S., Kokkinos, I., Mohamed, S., Vedaldi, A.: Describing textures in the wild. In: Proc. CVPR (2014) 3, 10", + "7. Croce, F., Andriushchenko, M., Sehwag, V., Debenedetti, E., Flammarion, N., Chiang, M., Mittal, P., Hein, M.: Robustbench: a standardized adversarial robustness benchmark. In: Proc. NeurIPS (2021) 1", + "8. Davis, J., Goadrich, M.: The relationship between precision-recall and roc curves. In: Proc. ICML (2006) 10", + "9. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: Proc. CVPR (2009) 3, 9", + "0. DeVries, T., Taylor, G.W.: Learning confidence for out-of-distribution detection in neural networks. arXiv preprint arXiv:1802.04865 (2018) 3, 4", + "1. Ding, Y., Liang, J., Jiang, B., Zheng, A., He, R.: Maps: A noise-robust progressive learning approach for source-free domain adaptive keypoint detection. IEEE Transactions on Circuits and Systems for Video Technology (2023) 3", + "2. Ding, Y., Sheng, L., Liang, J., Zheng, A., He, R.: Proxymix: Proxy-based mixup training with label refinery for source-free domain adaptation. Neural Networks 167, 92-103 (2023) 3", + "3. Farahani, A., Voghoei, S., Rasheed, K., Arabnia, H.R.: A brief review of domain adaptation. Advances in Data Science and Information Engineering: Proceedings from ICDATA 2020 and IKE 2020 pp. 877-894 (2021) 1", + "4. Foret, P., Kleiner, A., Mobahi, H., Neyshabur, B.: Sharpness-aware minimization for efficiently improving generalization. In: Proc. ICLR (2021) 3, 8", + "5. Gebru, T., Hoffman, J., Fei-Fei, L.: Fine-grained recognition in the wild: A multitask domain adaptation approach. In: Proc. ICCV (2017) 3", + "6. Gong, T., Jeong, J., Kim, T., Kim, Y., Shin, J., Lee, S.J.: Note: Robust continual test-time adaptation against temporal correlation. In: Proc. NeurIPS (2022) 4, 9", + "7. Gong, T., Kim, Y., Lee, T., Chottananurak, S., Lee, S.J.: Sotta: Robust test-time adaptation on noisy data streams. In: Proc. NeurIPS (2023) 4, 5, 8, 9, 10, 11, 12" + ], + "bbox": [ + 225, + 303, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Outlier-Aware Test-Time Adaptation with Stable Memory Replay", + "bbox": [ + 292, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Goyal, S., Sun, M., Raghunathan, A., Kolter, J.Z.: Test time adaptation via conjugate pseudo-labels. In: Proc. NeurIPS (2022) 9", + "19. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proc. CVPR (2016) 11", + "20. Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. In: Proc. ICLR (2018) 3, 9", + "21. Hendrycks, D., Gimpel, K.: A baseline for detecting misclassified and out-of-distribution examples in neural networks. In: Proc. ICLR (2016) 4", + "22. Ioffe, S., Szegedy, C.: Batch normalization: Accelerating deep network training by reducing internal covariate shift. In: Proc. ICML (2015) 10", + "23. Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A.A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., et al.: Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences 114(13), 3521-3526 (2017) 10", + "24. Krizhevsky, A.: Learning multiple layers of features from tiny images. Master's thesis, University of Tront (2009) 9", + "25. Le, Y., Yang, X.: Tiny imagenet visual recognition challenge. CS 231N 7(7), 3 (2015) 3, 10", + "26. Lee, J., Das, D., Choo, J., Choi, S.: Towards open-set test-time adaptation utilizing the wisdom of crowds in entropy minimization. In: Proc. CVPR (2023) 4, 8", + "27. Lee, K., Lee, K., Lee, H., Shin, J.: A simple unified framework for detecting out-of-distribution samples and adversarial attacks. In: Proc. NeurIPS (2018) 4", + "28. Li, Y., Xu, X., Su, Y., Jia, K.: On the robustness of open-world test-time training: Self-training with dynamic prototype expansion. In: Proc. CVPR (2023) 4, 9, 10, 11, 12", + "29. Liang, J., He, R., Tan, T.: A comprehensive survey on test-time adaptation under distribution shifts. International Journal of Computer Vision (2024) 1", + "30. Liang, J., Hu, D., Feng, J.: Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In: Proc. ICML (2020) 1", + "31. Liang, J., Sheng, L., Wang, Z., He, R., Tan, T.: Realistic unsupervised clip finetuning with universal entropy optimization. In: Proc. ICML (2024) 4", + "32. Liang, S., Li, Y., Srikant, R.: Enhancing the reliability of out-of-distribution image detection in neural networks. In: Proc. ICLR (2018) 4", + "33. Lin, Z., Roy, S.D., Li, Y.: Mood: Multi-level out-of-distribution detection. In: Proc. CVPR (2021) 4", + "34. Liu, W., Wang, X., Owens, J., Li, Y.: Energy-based out-of-distribution detection. In: Proc. NeurIPS (2020) 4", + "35. Maintainers, T., Contributors: Torchvision: Pytorch's computer vision library (2016) 11", + "36. Nado, Z., Padhy, S., Sculley, D., D'Amour, A., Lakshminarayanan, B., Snoek, J.: Evaluating prediction-time batch normalization for robustness under covariate shift. In: Proc. ICML Workshops (2020) 3, 9, 10, 11, 12", + "37. Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B., Ng, A.Y.: Reading digits in natural images with unsupervised feature learning (2011) 3, 10", + "38. Niu, S., Wu, J., Zhang, Y., Chen, Y., Zheng, S., Zhao, P., Tan, M.: Efficient test-time model adaptation without forgetting. In: Proc. ICML (2022) 2, 4, 5, 6, 7, 9, 10, 11, 12, 13", + "39. Niu, S., Wu, J., Zhang, Y., Wen, Z., Chen, Y., Zhao, P., Tan, M.: Towards stable test-time adaptation in dynamic wild world. In: Proc. ICLR (2022) 4, 7, 8, 9, 10, 11, 12" + ], + "bbox": [ + 215, + 146, + 784, + 838 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Yu et al.", + "bbox": [ + 271, + 114, + 330, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "40. Nixon, J., Dusenberry, M.W., Zhang, L., Jerfel, G., Tran, D.: Measuring calibration in deep learning. In: Proc. CVPR Workshops (2019) 5", + "41. Park, S., Yang, S., Choo, J., Yun, S.: Label shift adapter for test-time adaptation under covariate and label shifts. In: Proc. CVPR (2023) 4", + "42. Press, O., Schneider, S., Kümmerer, M., Bethge, M.: Rdumb: A simple approach that questions our progress in continual test-time adaptation. In: Proc. NeurIPS (2023) 8", + "43. Schneider, S., Rusak, E., Eck, L., Bringmann, O., Brendel, W., Bethge, M.: Improving robustness against common corruptions by covariate shift adaptation. In: Proc. NeurIPS (2020) 3", + "44. Shanmugam, D., Blalock, D., Balakrishnan, G., Guttag, J.: Better aggregation in test-time augmentation. In: Proc. CVPR (2021) 5", + "45. Sheng, L., Liang, J., He, R., Wang, Z., Tan, T.: Adaptguard: Defending against universal attacks for model adaptation. In: Proc. ICCV (2023) 3", + "46. Tan, M., Le, Q.: Efficientnet: Rethinking model scaling for convolutional neural networks. In: Proc. ICML (2019) 1", + "47. Tong, P., Su, W., Li, H., Ding, J., Haoxiang, Z., Chen, S.X.: Distribution free domain generalization. In: Proc. ICML (2023) 3", + "48. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: Proc. NeurIPS (2017) 1", + "49. Volpi, R., De Jorge, P., Larlus, D., Csurka, G.: On the road to online adaptation for semantic image segmentation. In: Proc. CVPR (2022) 2", + "50. Vyas, A., Jammalamadaka, N., Zhu, X., Das, D., Kaul, B., Willke, T.L.: Out-of-distribution detection using an ensemble of self supervised leave-out classifiers. In: Proc. ECCV (2018) 4", + "51. Wang, D., Shelhamer, E., Liu, S., Olshausen, B., Darrell, T.: Tent: Fully test-time adaptation by entropy minimization. In: Proc. ICML (2020) 2, 4, 7, 9, 10, 11, 12", + "52. Wang, J., Lan, C., Liu, C., Ouyang, Y., Qin, T., Lu, W., Chen, Y., Zeng, W., Yu, P.S.: Generalizing to unseen domains: A survey on domain generalization. IEEE Transactions on Knowledge and Data Engineering 35(8), 8052-8072 (2023) 3", + "53. Wang, M., Deng, W.: Deep visual domain adaptation: A survey. Neurocomputing 312, 135-153 (2018) 3", + "54. Wang, Q., Fink, O., Van Gool, L., Dai, D.: Continual test-time domain adaptation. In: Proc. CVPR (2022) 4, 9, 10, 11, 12", + "55. Wang, Y., Li, B., Che, T., Zhou, K., Liu, Z., Li, D.: Energy-based open-world uncertainty modeling for confidence calibration. In: Proc. CVPR (2021) 4", + "56. You, K., Long, M., Wang, J., Jordan, M.I.: How does learning rate decay help modern neural networks? arXiv preprint arXiv:1908.01878 (2019) 9", + "57. Yu, F., Seff, A., Zhang, Y., Song, S., Funkhouser, T., Xiao, J.: Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365 (2015) 3, 10", + "58. Yu, Y., Sheng, L., He, R., Liang, J.: Benchmarking test-time adaptation against distribution shifts in image classification. arXiv preprint arXiv:2307.03133 (2023) 1", + "59. Yuan, L., Xie, B., Li, S.: Robust test-time adaptation in dynamic scenarios. In: Proc. CVPR (2023) 2, 3, 9, 10, 11, 12", + "60. Yurtsever, E., Lambert, J., Carballo, A., Takeda, K.: A survey of autonomous driving: Common practices and emerging technologies. IEEE Access 8, 58443-58469 (2020) 1", + "61. Zhang, M., Levine, S., Finn, C.: Memo: Test time robustness via adaptation and augmentation. In: Proc. NeurIPS (2022) 4, 5" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Outlier-Aware Test-Time Adaptation with Stable Memory Replay", + "bbox": [ + 292, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "62. Zhang, Y., Wang, X., Jin, K., Yuan, K., Zhang, Z., Wang, L., Jin, R., Tan, T.: Adanpc: Exploring non-parametric classifier for test-time adaptation. In: Proc. ICML (2023) 2", + "63. Zhao, H., Liu, Y., Alahi, A., Lin, T.: On pitfalls of test-time adaptation. In: Proc. ICML (2023) 4", + "64. Zhou, B., Lapedriza, A., Khosla, A., Oliva, A., Torralba, A.: Places: A 10 million image database for scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence 40(6), 1452-1464 (2017) 3, 10", + "65. Zhou, Z., Guo, L.Z., Jia, L.H., Zhang, D., Li, Y.F.: Ods: test-time adaptation in the presence of open-world data shift. In: Proc. ICML (2023) 4" + ], + "bbox": [ + 215, + 146, + 784, + 285 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Yu et al.", + "bbox": [ + 271, + 114, + 330, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/STAMP_ Outlier-Aware Test-Time Adaptation with Stable Memory Replay/20b10b92-8cd4-4e35-ad35-a50ac0ff44cf_model.json b/2024/STAMP_ Outlier-Aware Test-Time Adaptation with Stable Memory Replay/20b10b92-8cd4-4e35-ad35-a50ac0ff44cf_model.json new file mode 100644 index 0000000000000000000000000000000000000000..11cb10293f8200b69a07d8ee6e829b233ec7cbeb --- /dev/null +++ b/2024/STAMP_ Outlier-Aware Test-Time Adaptation with Stable Memory Replay/20b10b92-8cd4-4e35-ad35-a50ac0ff44cf_model.json @@ -0,0 +1,2392 @@ +[ + [ + { + "type": "image", + "bbox": [ + 0.275, + 0.146, + 0.327, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.331, + 0.155, + 0.728, + 0.2 + ], + "angle": 0, + "content": "STAMP: Outlier-Aware Test-Time Annotation with Stable Memory Replay" + }, + { + "type": "text", + "bbox": [ + 0.255, + 0.224, + 0.747, + 0.241 + ], + "angle": 0, + "content": "Yongcan Yu\\(^{1,2}\\), Lijun Sheng\\(^{1,3}\\), Ran He\\(^{1,2}\\), and Jian Liang\\(^{1,2\\star}\\)" + }, + { + "type": "text", + "bbox": [ + 0.244, + 0.252, + 0.756, + 0.266 + ], + "angle": 0, + "content": "\\(^{1}\\) NLPR & MAIS, Institute of Automation, Chinese Academy of Sciences" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.266, + 0.757, + 0.28 + ], + "angle": 0, + "content": "\\(^{2}\\) School of Artificial Intelligence, University of Chinese Academy of Sciences" + }, + { + "type": "text", + "bbox": [ + 0.34, + 0.28, + 0.664, + 0.295 + ], + "angle": 0, + "content": "3 University of Science and Technology of China" + }, + { + "type": "text", + "bbox": [ + 0.355, + 0.296, + 0.65, + 0.309 + ], + "angle": 0, + "content": "{yuyongcan0223, liangjian92}@gmail.com" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.346, + 0.744, + 0.596 + ], + "angle": 0, + "content": "Abstract. Test-time adaptation (TTA) aims to address the distribution shift between the training and test data with only unlabeled data at test time. Existing TTA methods often focus on improving recognition performance specifically for test data associated with classes in the training set. However, during the open-world inference process, there are inevitably test data instances from unknown classes, commonly referred to as outliers. This paper pays attention to the problem that conducts both sample recognition and outlier rejection during inference while outliers exist. To address this problem, we propose a new approach called STAble Memory rePlay (STAMP), which performs optimization over a stable memory bank instead of the risky mini-batch. In particular, the memory bank is dynamically updated by selecting low-entropy and label-consistent samples in a class-balanced manner. In addition, we develop a self-weighted entropy minimization strategy that assigns higher weight to low-entropy samples. Extensive results demonstrate that STAMP outperforms existing TTA methods in terms of both recognition and outlier detection performance. The code is released at https://github.com/yuyongcan/STAMP." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.61, + 0.741, + 0.639 + ], + "angle": 0, + "content": "Keywords: online test-time adaptation \\(\\cdot\\) outlier detection \\(\\cdot\\) memory bank \\(\\cdot\\) self-weighted entropy minimization" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.664, + 0.376, + 0.68 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.695, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Deep neural networks (DNNs) [46, 48] have achieved remarkable performance across various tasks. Nevertheless, the impressive performance is based on the assumption that training data and test data come from the same distribution. However, in real-world scenarios, the presence of distribution shifts is ubiquitous [13, 30], leading to a degradation in the prediction performance [7]. Taking autonomous driving [60] as an example, changing weather conditions causes systems trained on laboratory data to make incorrect decisions. To mitigate the impact of distribution shift, test-time adaptation (TTA) [29, 58] has emerged as" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.386, + 0.84 + ], + "angle": 0, + "content": "* Corresponding author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.333, + 0.128 + ], + "angle": 0, + "content": "Yu et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.144, + 0.787, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.303, + 0.789, + 0.377 + ], + "angle": 0, + "content": "Fig. 1: The illustration of outlier-aware test-time adaptation. The test data stream consists of both normal samples and outliers. When the test data arrives, the algorithm adapts the model and provides two outputs, a prediction and an OOD score. OOD score indicates the likelihood that a sample is an outlier. The detector in the deployment scenario chooses to keep the prediction or reject samples according to the OOD score." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.429, + 0.784, + 0.474 + ], + "angle": 0, + "content": "a promising approach, adapting models with unlabeled samples during inference. Nowadays, TTA has achieved remarkable results in various tasks, such as image classification [51, 62], semantic segmentation [3, 49], object detection [1]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.476, + 0.789, + 0.627 + ], + "angle": 0, + "content": "Previous TTA works [5, 38, 59] always assume that test data and training data share the same class space. Nevertheless, in realistic open-world scenarios, test data streams will inevitably contain samples from new semantic classes, also known as outliers. The introduction of outliers brings two additional risks to adaptation tasks: (i) Classifying outliers into known categories poses safety risks, especially in contexts like autonomous driving, where systems should request human intervention to handle unknown objects to avoid traffic accidents. (ii) Self-supervised loss (e.g., entropy minimization, pseudo-labeling) calculated with outliers may misguide the optimization process, decreasing the model's recognition ability on known classes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.789, + 0.763 + ], + "angle": 0, + "content": "To better tackle the above risks, we pay attention to a more realistic TTA scenario termed outlier-aware test-time adaptation. Outlier-aware test-time adaptation aims to conduct recognition for normal samples and reject to make predictions for outliers. The overview of outlier-aware test-time adaptation is illustrated in Figure 1. When test data mixed with outliers arrives, the algorithm adapts the model and generates the prediction and an out-of-distribution (OOD) score for each sample. Then we decide to keep the prediction or reject the sample based on its OOD score. The sample with a higher OOD score is more likely to be considered an outlier." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.789, + 0.843 + ], + "angle": 0, + "content": "To address the outlier-aware test-time adaptation, we additionally design a novel algorithm called STAble Memory rePlay (STAMP). STAMP consists of three components: reliable class-balanced memory, self-weighted entropy minimization, and stable optimization strategy. Reliable class-balanced memory stores reliable samples for optimization by filtering samples based on their pre" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.293, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Outlier-Aware Test-Time Adaptation with Stable Memory Replay" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.283 + ], + "angle": 0, + "content": "diction's consistency and entropy. To maintain the class-balanced prediction of the model, it dynamically discards samples with the class that participates most frequently in optimization when the memory is full. Furthermore, we adopt self-weighted entropy minimization to assign greater weights to low-entropy samples in memory and adjust the weights for each sample adaptively. To ensure the stability of the optimization process, we dynamically adjust the optimization step size to quickly adapt to the target domain at the early stage and mitigate error accumulation later. Moreover, we apply the sharpness-aware minimization technique [14] to further reduce the impact of noisy gradients." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.788, + 0.419 + ], + "angle": 0, + "content": "To demonstrate the effectiveness of STAMP, we compare it with eight state-of-the-art TTA algorithms. We select three commonly used datasets in TTA, including CIFAR10-C, CIFAR100-C [10, 20], and ImageNet-C [9, 20], alongside some outlier datasets (i.e., Noise, SVHN [37], LSUN [57], TinyImageNet [25], Textures [6], Places365 [64]). Experimental results show that our method achieves superior performance on recognition and outlier detection under the outlier-aware TTA setting. The experiment on conventional protocol indicates that STAMP also performs well in the scenario without outliers. In summary, we list our contributions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.429, + 0.784, + 0.458 + ], + "angle": 0, + "content": "- We introduce outlier-aware test-time adaptation, a TTA scenario considering both normal sample recognition and outlier rejection during inference." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.459, + 0.785, + 0.502 + ], + "angle": 0, + "content": "- We propose a memory that dynamically stores reliable samples for replay. To prevent noisy samples from entering the memory, we propose a filtering strategy based on the consistency and entropy of the model's prediction." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.504, + 0.784, + 0.533 + ], + "angle": 0, + "content": "- We design self-weighted entropy minimization, which guides the model to adaptively learn to assign higher weights to valuable samples." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.534, + 0.784, + 0.577 + ], + "angle": 0, + "content": "- We propose a proper metric that considers both generalization and outlier detection performance. Extensive experiments compared with eight TTA algorithms on several benchmarks demonstrate the effectiveness of our method." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.429, + 0.785, + 0.577 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.602, + 0.388, + 0.618 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.634, + 0.445, + 0.649 + ], + "angle": 0, + "content": "2.1 Test-Time Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.841 + ], + "angle": 0, + "content": "TTA intends to adapt the model pre-trained on the source domain to the target domain while the distribution shift exists. Compared to domain generalization [47, 52] and domain adaptation [15, 53], TTA needs to access neither the source data nor the target labels, making it has a wide range of applications [11, 12, 45]. TTA methods can be mainly divided into two solutions: batch normalization (BN) calibration and self-training. In BN calibration, Nado et al. [36] utilize the BN statistics of the test batch \\((\\mu_t, \\sigma_t)\\) to conduct normalization in the BN layer, while Schneider [43] propose a more comprehensive approach considering statistics from both the source domain \\((\\mu_s, \\sigma_s)\\) and the target domain to calibrate the domain shifts. To conquer the correlation sampling problem, Yuan et al. [59] adopt robust batch normalization, which updates the statistic \\((\\mu_t, \\sigma_t)\\) in a moving average manner with the test data stream." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.331, + 0.127 + ], + "angle": 0, + "content": "Yu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.404 + ], + "angle": 0, + "content": "Currently, the primary paradigm in TTA is self-training [2, 61], where Wang et al. [51] adopt entropy minimization as a self-supervised loss for optimization and Chen et al. [5] introduce contrastive loss to TTA. However, unsupervised loss may contain noise information. Niu et al. [38] first propose filtering out high-entropy samples, a mechanism adopted by subsequent works [17, 39]. Subsequently, more attention has been paid to addressing more realistic adaptation scenarios such as continually changing target domains [4, 54], label shift [41, 65], and temporally correlated shift [16, 63]. In our study, we focus on a more realistic scenario where test data contains samples with new semantic classes. Some existing works [17, 26, 28] discuss the TTA algorithm in scenarios containing outliers. However, Lee et al. [26] and Gong et al. [17] solely focus on the generalization ability to domain shift, neglecting the issue of rejecting the outliers. A recent work [28] discusses outlier detection performance based on TTA tasks, but it only provides binary predictions to indicate whether the test data is outlier. In contrast, outlier-aware test-time adaptation is more flexible, providing an OOD score, and different systems can conduct sample rejection based on security requirements and the score." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.434, + 0.509, + 0.449 + ], + "angle": 0, + "content": "2.2 Out-of-Distribution Detection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.467, + 0.788, + 0.725 + ], + "angle": 0, + "content": "OOD detection aims to identify instances that significantly deviate from the distribution of the training data [31]. Their primary focus is usually on detecting samples whose labels were not seen during the training phase. This area has been extensively researched, with one common approach being the utilization of confidence scores such as max softmax probability [21], energy score [33, 34], and others [27] derived from the model's output to detect OOD samples. Further improvements have been made by Liang et al. [32] through techniques like temperature scaling and input preprocessing to enhance the effectiveness of confidence scores. Another avenue for OOD sample detection lies in training-based methods [10, 50, 55]. These techniques involve training a branch using both indistribution data and OOD data to enable the model to distinguish OOD data during training. However, these methods alter the training process of the source model, which can not be directly applied to adaptation tasks based on a well-trained model. Moreover, OOD detection traditionally assumes that both the training and test data originate from the same domain, a condition not met in TTA scenarios. Under the realistic scenario, there may be a domain shift between test data and training data, which brings greater challenges to detection." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.755, + 0.381, + 0.773 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In this section, we first introduce the outlier-aware test-time adaptation problem in Section 3.1. Then we describe our proposed method, including how to build a memory bank and conduct the optimization in Section 3.2." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.293, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Outlier-Aware Test-Time Adaptation with Stable Memory Replay" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.569, + 0.162 + ], + "angle": 0, + "content": "3.1 Outlier-Aware Test-Time Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.17, + 0.784, + 0.244 + ], + "angle": 0, + "content": "Let \\(\\mathcal{D}_S\\) be the source domain dataset with label space \\(\\mathcal{C}_S\\) follows a distribution \\(P_S(x,y)\\). Similarly, let \\(\\mathcal{D}_T = \\{(x_i,y_i)\\}_{i=1}^{N_T}\\) be the target domain dataset with label space \\(\\mathcal{C}_T\\) follows a distribution \\(P_T(x,y)\\). Given a source model \\(f_{\\theta_s}(x)\\) well trained on \\(\\mathcal{D}_S\\) with parameter \\(\\theta_s\\), TTA aims to adapt the source model to the target domain while \\(P_S(x) \\neq P_T(x)\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.246, + 0.785, + 0.35 + ], + "angle": 0, + "content": "In traditional closed-set TTA, it is commonly assumed that the \\(\\mathcal{D}_S\\) and \\(\\mathcal{D}_T\\) share the same label space, i.e., \\(\\mathcal{C}_S = \\mathcal{C}_T\\). However, in outlier-aware test-time adaptation, we consider \\(\\mathcal{C}_S \\subseteq \\mathcal{C}_T\\) while some outliers with new semantic classes will be encountered in \\(\\mathcal{D}_T\\). The illustration of outlier-aware test-time adaption is shown in Figure 1. In addition to adapting the model and providing the prediction, we also generate an OOD score \\(\\delta_i\\) for each sample \\(x_i\\). Then we adopt an outlier detector \\(g(\\cdot)\\) as follow:" + }, + { + "type": "equation", + "bbox": [ + 0.406, + 0.359, + 0.785, + 0.4 + ], + "angle": 0, + "content": "\\[\ng \\left(x _ {i}\\right) = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} \\delta_ {i} \\geq \\delta_ {t h r} \\\\ 0, & \\text {i f} \\delta_ {i} < \\delta_ {t h r}, \\end{array} \\right. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.406, + 0.788, + 0.48 + ], + "angle": 0, + "content": "where \\(\\delta_{thr}\\) is a threshold that can be adjusted according to the requirements of different application scenarios. The samples with \\(g(x_i) = 1\\) will be considered outliers and rejected for recognition. And we output the prediction for samples with \\(g(x_i) = 0\\). In our work, we take the entropy of the model prediction as our OOD score." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.501, + 0.455, + 0.517 + ], + "angle": 0, + "content": "3.2 Stable Memory Replay" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.784, + 0.629 + ], + "angle": 0, + "content": "In this section, we introduce stable memory replay (STAMP). STAMP first uses a filtering mechanism to filter out unreliable samples in the coming mini-batch and stores reliable samples in memory. Then, STAMP optimizes the model using in-memory samples by applying self-weighted entropy minimization. In order to further improve the stability of optimization, STAMP uses test-time step size decay and sharpness-aware minimization to guide optimization. The illustration of STAMP is shown in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.784, + 0.75 + ], + "angle": 0, + "content": "Reliable Class-Balanced Memory. Since excluding unreliable samples from optimization has shown great effectiveness [17, 38], we propose reliable class-balanced memory to store reliable samples for optimization. To have a more robust estimation of the sample's reliability, we get its prediction in an augmentation-averaged way. Moreover, we prevent unreliable samples from getting into the memory bank \\(\\mathcal{M}\\) by employing consistency filtering and entropy filtering. To preserve the model's balanced class prior distribution, we dynamically discard samples with the highest class frequency from the memory bank." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.785, + 0.84 + ], + "angle": 0, + "content": "To ensure robust predictions, we adopt the augmentation-averaged method, which addresses the issue of deep neural networks being overconfident with out-of-distribution samples. This overconfidence often blurs the line between in-distribution and out-of-distribution samples. However, research has demonstrated that employing test-time augmentation [44, 61] enhances model calibration [40]. Thus, when presented with a test sample \\( x \\), we derive its prediction" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.333, + 0.127 + ], + "angle": 0, + "content": "Yu et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.145, + 0.784, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.386, + 0.788, + 0.485 + ], + "angle": 0, + "content": "Fig. 2: The overview of STAMP. STAMP retains reliable samples for optimizing the network by maintaining a memory bank. When test samples arrive, STAMP generates an augmentation-averaged prediction. Two filtering mechanisms are designed to filter out inconsistent and high-entropy samples and add the remaining to the memory. A frequency vector is maintained to select samples to discard when the memory is full. STAMP leverages the reliable samples in the memory to optimize the model with the SAM optimizer alongside a decaying step strategy." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.533, + 0.435, + 0.549 + ], + "angle": 0, + "content": "through the following process:" + }, + { + "type": "equation", + "bbox": [ + 0.41, + 0.554, + 0.787, + 0.597 + ], + "angle": 0, + "content": "\\[\n\\hat {p} (x) = \\frac {1}{K} \\sum_ {k = 1} ^ {K} f _ {\\theta_ {t}} \\left(\\tau_ {k} (x)\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.603, + 0.788, + 0.679 + ], + "angle": 0, + "content": "where \\(\\tau(\\cdot)\\) is the operation of augmentation, \\(K\\) is the total number of augmented views. By averaging predictions over augmented samples, the model can better eliminate the uncertainty inherent in its predictions, improving calibration and obtaining more reliable confidence estimates. In the supplementary material, we discuss the impact of the augmentation-averaged estimation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.68, + 0.788, + 0.814 + ], + "angle": 0, + "content": "After acquiring the prediction \\(\\hat{p}\\), we implement filtering strategies to prevent unreliable samples from being added to memory for two reasons. Firstly, optimization using unreliable samples can result in error accumulation over time [38]. Secondly, we prevent outliers from entering the memory and reduce the entropy of normal samples in the memory. This action facilitates the model's ability to distinguish between normal samples and outliers. Since reliable samples that are beneficial to optimization tend to have the same pseudo-label as the source model, we employ consistency filtering to remove samples whose predictions are inconsistent with those of the source model:" + }, + { + "type": "equation", + "bbox": [ + 0.317, + 0.818, + 0.787, + 0.845 + ], + "angle": 0, + "content": "\\[\nm _ {c o n} (x) = \\mathbb {I} \\left(\\arg \\max _ {c} [ \\hat {p} (x) ] _ {c} = \\arg \\max _ {c} [ f _ {\\theta_ {s}} (x) ] _ {c}\\right), \\tag {3}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.293, + 0.114, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Outlier-Aware Test-Time Adaptation with Stable Memory Replay" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "code_caption", + "bbox": [ + 0.218, + 0.146, + 0.561, + 0.162 + ], + "angle": 0, + "content": "Algorithm 1 Reliable Class-Balanced Memory" + }, + { + "type": "algorithm", + "bbox": [ + 0.218, + 0.165, + 0.788, + 0.353 + ], + "angle": 0, + "content": "Input: memory bank \\(\\mathcal{M}\\), a batch of data \\(\\{x_{i}\\}_{i = 1}^{B}\\), entropy threshold \\(\\mathcal{H}_{thr}\\), class frequency \\(\\{\\xi_c\\}_{c = 1}^{|C_S|}\\). \nfor \\(i = 1\\) to \\(B\\) do \nCalculate \\(\\hat{p}(x_i)\\) by Equation (2) and \\(\\hat{y}_i = \\arg \\max_c [\\hat{p}(x_i)]_c\\). Calculate \\(m_{con}(x_i)\\), \\(m_{ent}(x_i)\\) of \\(x_i\\) by Equation (3), (4). \nif \\(m_{con}(x_i) = 1\\) and \\(m_{ent}(x_i) = 1\\) then \nif \\(\\mathcal{M}\\) is full then \nDiscard one sample with highest class frequency \\(\\xi\\) from \\(\\mathcal{M}\\). \nend if \nadd \\((x_i, \\hat{y}_i)\\) into \\(\\mathcal{M}\\). \nend if \nend for \nUpdate \\(\\xi\\) with Equation (5)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.382, + 0.784, + 0.443 + ], + "angle": 0, + "content": "where \\(\\mathbb{I}(\\cdot)\\) is the indicator function. Moreover, including high entropy samples for optimization can harm the stability of the model [38, 39], while samples with low entropy always possess greater value and reliability. Therefore, we use entropy filtering to filter out samples with high entropy from entering the memory." + }, + { + "type": "equation", + "bbox": [ + 0.387, + 0.452, + 0.786, + 0.478 + ], + "angle": 0, + "content": "\\[\nm _ {e n t} (x) = \\mathbb {I} \\left(\\mathcal {H} (\\hat {p} (x)) < \\mathcal {H} _ {t h r}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.486, + 0.784, + 0.561 + ], + "angle": 0, + "content": "where \\(\\mathcal{H}(q) = -\\sum_{c} q_c \\log q_c\\) denotes the entropy of the prediction vector \\(q\\). We filter out samples with \\(\\mathcal{H}(\\hat{p}) \\geq \\mathcal{H}_{thr}\\), where \\(\\mathcal{H}_{thr}\\) is a threshold. Considering both filtering strategies, we add sample \\((x, \\hat{y})\\) into memory bank \\(\\mathcal{M}\\) if the sample satisfies consistency filtering and entropy filtering simultaneously, i.e., \\(m_{con}(x) = 1\\) and \\(m_{ent}(x) = 1\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.563, + 0.785, + 0.683 + ], + "angle": 0, + "content": "Since the number of samples stored in memory is limited, we design a class-balanced strategy to choose a sample to discard when the memory is full. Since we adopt entropy minimization as our optimization objective, an issue arises when the model excessively optimizes samples from a single category over a short period. It leads to a bias towards predicting the dominant category. To counteract this bias, we dynamically discard samples from the memory that belong to the category with the highest update frequency. After each optimization of the model, we update the class frequencies in a moving average manner:" + }, + { + "type": "equation", + "bbox": [ + 0.415, + 0.692, + 0.786, + 0.711 + ], + "angle": 0, + "content": "\\[\n\\xi_ {c} ^ {t} = (1 - \\beta) \\xi_ {c} ^ {t - 1} + \\beta \\nu_ {c} ^ {t}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.784, + 0.779 + ], + "angle": 0, + "content": "where \\(\\nu_{c}^{t}\\) represents the number of samples with class label \\(c\\) in memory and \\(\\beta\\) is moving average coefficient hyperparameter. While the memory is full, we discard a sample from the class with the highest \\(\\xi\\). We summarize the work process of reliable class-balanced memory in Algorithm 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Self-Weighted Entropy Minimization. After updating the memory, we replay the reliable samples stored in the memory \\(\\mathcal{M}\\) to optimize the model. Following [51], we employ entropy minimization as our loss function. This self-supervised loss encourages the model to exhibit higher confidence in its predic" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.333, + 0.128 + ], + "angle": 0, + "content": "Yu et al." + }, + { + "type": "code_caption", + "bbox": [ + 0.218, + 0.146, + 0.502, + 0.163 + ], + "angle": 0, + "content": "Algorithm 2 The Pipeline of STAMP" + }, + { + "type": "algorithm", + "bbox": [ + 0.229, + 0.166, + 0.744, + 0.311 + ], + "angle": 0, + "content": "Initialize: Memory bank \\(\\mathcal{M} = \\emptyset\\) zero-initialized class frequency \\(\\{\\xi_c\\}_{c = 1}^{|C_S|}\\) \nInput: Unlabeled test samples \\(\\mathcal{D}_T = \\{x_i\\}_{i = 1}^{NT}\\) , a trained model \\(f_{\\theta_s}(\\cdot)\\) \nfor each online mini-batch \\(X_{b}\\) do Calculate \\(\\hat{p} (x)\\) by Equation (2) and \\(\\mathcal{H}(\\hat{p} (x))\\) for all \\(x\\in X_b\\) . Obtain the prediction \\(\\hat{y} = \\arg \\max_{c}[\\hat{p} (x)]_{c}\\) for all \\(x\\in X_b\\) Update memory bank \\(\\mathcal{M}\\) by Algorithm 1. Compute \\(\\mathcal{L}_{sem}\\) with \\(\\mathcal{M}\\) by Equation (6). Update \\(\\theta_t\\) with SAM in Equation (7) and decay step size in Equation (8). \nend for \nOutput: The prediction and OOD score \\(\\{\\hat{y}_i,\\mathcal{H}(\\hat{p} (x_i))\\}_{i = 1}^{NT}\\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.345, + 0.789, + 0.394 + ], + "angle": 0, + "content": "tions. In order to give more valuable samples a higher importance in optimization, we weight samples in \\(\\mathcal{M}\\) in a self-weighted manner. Given samples \\(\\{x_{i}\\}_{i = 1}^{|M|}\\), we calculate the following loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.355, + 0.404, + 0.787, + 0.439 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s e m} (\\mathcal {M}, \\theta_ {t}) = \\sum_ {x _ {i} \\in \\mathcal {M}} \\frac {w _ {i}}{\\sum_ {j} w _ {j}} \\cdot \\mathcal {H} \\left(f _ {\\theta_ {t}} \\left(x _ {i}\\right)\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.453, + 0.789, + 0.546 + ], + "angle": 0, + "content": "where \\( w_{i} = e^{-\\mathcal{H}(f_{\\theta_{t}}(x_{i}))} \\) denotes the weight of sample \\( x_{i} \\). It's worth noting that \\( w_{i} \\) is also involved with the gradient back-propagation. The reason for this adaptive weighting is two-fold: (i) it gives greater weights to samples with lower entropy, which are considered more reliable. (ii) The model needs to learn the weight of each sample and adaptively give important samples higher weights, instead of solely minimizing entropy for all samples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.546, + 0.79, + 0.683 + ], + "angle": 0, + "content": "Stable Optimization Strategy. To optimize the model in a stable way, we employ sharpness-aware minimization (SAM) [14] and step size decay in parameter updating. SAM is a technique that aims to improve generalization performance in deep learning models by minimizing a loss function while simultaneously controlling the sharpness of the loss landscape. Previous work [14] demonstrates that SAM exhibits resilience against noisy labels, which is also a prevalent challenge in TTA. Therefore, we consider SAM as a good choice of the optimizer in the adaptation scenario [17, 39]. Mathematically, this involves minimizing the following objective function:" + }, + { + "type": "equation", + "bbox": [ + 0.269, + 0.694, + 0.787, + 0.722 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta_ {t}} \\mathcal {L} _ {s e m} ^ {S A} (\\mathcal {M}, \\theta_ {t}), \\text {w h e r e} \\mathcal {L} _ {s e m} ^ {S A} (\\mathcal {M}, \\theta_ {t}) \\triangleq \\max _ {| | \\epsilon | | _ {2} \\leq \\rho} \\mathcal {L} _ {s e m} (\\mathcal {M}, \\theta_ {t} + \\epsilon), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.733, + 0.788, + 0.763 + ], + "angle": 0, + "content": "where \\(\\rho\\) represents the radius of the neighborhood. SAM minimizes the maximum loss within the parameters' vicinity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.789, + 0.843 + ], + "angle": 0, + "content": "Recent research [26, 42] has revealed that under long-term adaptation, TTA algorithms may suffer from substantial error accumulation and consequent performance degradation. This issue becomes even more severe when test data contains outliers. To mitigate this problem, we introduce a test-time step size (a.k.a. learning rate) decay strategy. To achieve this, we update the model's parameters" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.293, + 0.115, + 0.731, + 0.131 + ], + "angle": 0, + "content": "Outlier-Aware Test-Time Adaptation with Stable Memory Replay" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 1: Results (\\%) on CIFAR10-C and CIFAR100-C datasets under outlier-aware TTA setting across four outlier datasets. We report the accuracy (ACC), AUC score, and H-score. All the results are averaged over 15 different corruptions. The results in bold denote the best performance." + }, + { + "type": "table", + "bbox": [ + 0.234, + 0.213, + 0.77, + 0.451 + ], + "angle": 0, + "content": "
NoiseSVHN-CLSUN-CTinyImageNet-CAvg.
ACCAUCH-scoreACCAUCH-scoreACCAUCH-scoreACCAUCH-scoreACCAUCH-score
CIFAR10-CSource57.370.462.357.367.461.157.362.859.657.364.559.457.366.360.6
BN Stats [36]72.968.670.678.775.376.979.479.479.479.072.975.877.574.075.7
Tent [51]77.448.759.780.854.965.181.262.370.481.165.672.480.157.966.9
EATA [38]72.968.570.678.875.376.979.479.479.478.973.175.977.574.175.7
SAR [39]72.968.570.678.775.376.979.479.479.479.072.975.877.574.075.7
CoTTA [54]77.362.467.381.678.680.182.284.283.281.975.378.480.875.177.2
RoTTA [54]77.674.375.678.476.077.278.879.579.178.673.375.878.375.876.9
SoTTA [17]77.851.761.679.372.875.979.877.978.879.672.675.979.168.873.1
OWTTT [28]62.364.458.566.175.369.663.178.968.556.358.856.262.069.363.2
STAMP77.980.778.882.379.080.583.584.984.282.675.178.681.679.980.5
CIFAR100-CSource35.843.138.035.849.440.135.858.243.235.857.142.735.852.041.0
BN Stats [36]45.880.958.452.772.560.953.773.862.053.268.659.751.374.060.3
Tent [51]47.955.851.254.470.461.255.472.462.755.068.660.953.266.859.0
EATA [38]55.286.167.158.175.665.658.877.266.758.670.764.057.777.465.9
SAR [39]57.588.668.959.265.261.960.573.566.360.872.165.959.574.965.8
CoTTA [54]47.083.459.953.773.261.854.376.963.654.568.160.452.475.461.4
RoTTA [59]47.954.049.447.367.055.348.369.556.747.865.555.047.864.054.1
SoTTA [17]54.453.352.853.670.360.754.470.861.453.968.460.154.165.758.8
OWTTT [28]47.170.356.253.974.362.354.573.562.554.268.560.452.471.660.4
STAMP57.997.572.663.780.070.963.780.771.263.974.468.762.383.170.8
" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.48, + 0.513, + 0.494 + ], + "angle": 0, + "content": "\\(\\theta_{t}\\) using cosine annealing decay step size:" + }, + { + "type": "equation", + "bbox": [ + 0.279, + 0.505, + 0.785, + 0.534 + ], + "angle": 0, + "content": "\\[\n\\theta_ {t + 1} = \\theta_ {t} - \\alpha_ {t} \\nabla \\mathcal {L} _ {\\text {s e m}} ^ {S A} (\\mathcal {M}, \\theta_ {t}), \\text {w h e r e} \\alpha_ {t} = \\frac {1}{2} \\alpha_ {0} (1 + \\cos (\\frac {t}{T} \\pi)), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.544, + 0.788, + 0.667 + ], + "angle": 0, + "content": "where \\(\\alpha_{t}\\) is the step size at timestamp \\(t\\) and \\(T\\) is a hyper-parameter to control the step size decay. While step size decay is commonly utilized in model training [56], we leverage it during testing for a distinct purpose. In the early stages of TTA, the model should adapt swiftly due to the significant disparity between the source and target domains. However, as adaptation progresses and the model acquires comprehensive distributional information about the target domain, it should reduce parameter updates to prevent error accumulation and enhance stability. We summarize the pipeline of STAMP in Algorithm 2." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.69, + 0.377, + 0.708 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.723, + 0.312, + 0.739 + ], + "angle": 0, + "content": "4.1 Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Datasets. Following previous TTA methods [16, 18, 54], we adopt three corruption datasets as normal datasets (i.e., in-distribution datasets). CIFAR10-C and CIFAR100-C [20, 24] stand out as the most widely used benchmarks in TTA tasks, comprising 10/100 categories with 15 types of corruptions respectively. Each type of corruption has 10,000 images. Similarly, ImageNet-C [9, 20] serves as a significant benchmark, consisting of 1,000 categories with the same 15 types" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.331, + 0.127 + ], + "angle": 0, + "content": "Yu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 2: Results (\\%) on ImageNet-C under outlier-aware TTA setting across two outlier datasets. All the results are averaged over 15 different corruptions." + }, + { + "type": "table", + "bbox": [ + 0.321, + 0.185, + 0.684, + 0.329 + ], + "angle": 0, + "content": "
MethodPlaces365-CTextures-C
ACCAUCH-scoreACCAUCH-score
Source18.261.626.118.254.625.8
BN Stats [36]31.167.741.131.661.240.7
Tent [51]34.951.839.53948.642.0
EATA [38]46.472.656.046.462.252.8
SAR [39]44.973.355.045.667.054.0
CoTTA [54]33.866.943.534.260.742.8
RoTTA [59]36.668.646.537.065.346.5
SoTTA [17]41.767.850.741.860.348.8
OWTTT [28]9.154.013.99.459.414.6
STAMP46.476.257.246.570.055.6
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.358, + 0.788, + 0.479 + ], + "angle": 0, + "content": "of corruption. For out-of-distribution datasets, we choose Noise, SVHN [37], LSUN [57], and TinyImageNet [25] for the CIFAR10/100-C datasets, and Textures [6] and Places365 [64] for ImageNet-C. To maintain consistency in domain shift between outliers and normal datasets, we apply the same corruptions \\(^4\\) on outlier datasets. (e.g., SVHN-C refers to the corruptions version of SVHN dataset.) Throughout all experiments, we maintain a fixed ratio of 8:2 between normal samples and outliers. Further details of the outlier datasets are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.481, + 0.789, + 0.721 + ], + "angle": 0, + "content": "Baselines. To have a comprehensive comparison of existing TTA methods, we select eight TTA algorithms as our baselines. Source refers to the source pre-trained model. BN Stats [36] adjusts batch normalization (BN) layers [22] with the statistics of test data. Tent [51] optimizes the BN affine parameters in the model by entropy minimization. CoTTA [54] calculates weighted-average and augmentation-averaged pseudo-label to distill knowledge from the teacher model to the student model, and stochastically restores parameters to prevent the model from catastrophic forgetting. EATA [38] adopts the entropy and diversity weight with elastic weight consolidation [23] for anti-forgetting. SAR [39] excludes high-entropy samples from sharpness-aware entropy minimization and reset the model when it overfits. RoTTA [59] uses robust batch normalization and category-balanced sampling with timeliness and uncertainty to deal with the correlated label problem. SoTTA [17] applies high-confidence uniform-class sampling and entropy-sharpness minimization to alleviate the impact of noisy samples. OWTTT [28] proposes prototype expansion with prototype clustering and distribution alignment." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.723, + 0.788, + 0.814 + ], + "angle": 0, + "content": "Evaluation Metrics. We evaluate the above baselines for both generalization and outlier detection. For generalization, we adopt classification accuracy (ACC) as a metric. For outlier detection, we consider the area under the ROC curve (AUC) [8], a threshold-free metric to measure the binary classification problem. The ROC curve depicts the relationship between the true positive rate (TPR) and the false positive rate (FPR). Models with better detection ability" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.824, + 0.538, + 0.841 + ], + "angle": 0, + "content": "4 https://github.com/hendrycks/robustness" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.293, + 0.115, + 0.731, + 0.131 + ], + "angle": 0, + "content": "Outlier-Aware Test-Time Adaptation with Stable Memory Replay" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 3: Results (\\%) on CIFAR10/100-C datasets under outlier-aware TTA setting. We divide datasets into the normal part and the outlier part. In detail, we choose 8/80 classes as known samples and other 2/20 classes as outliers for CIFAR10/100-C. All the results are averaged over 15 different corruptions." + }, + { + "type": "table", + "bbox": [ + 0.334, + 0.213, + 0.673, + 0.347 + ], + "angle": 0, + "content": "
MethodCIFAR10-C (8:2)CIFAR100-C (80:20)
ACCAUCH-scoreACCAUCH-score
Source60.661.560.037.160.644.2
BN Stats [36]79.466.972.655.566.260.2
Tent [51]80.165.972.360.465.662.8
EATA [38]81.567.974.161.367.464.2
SAR [39]82.366.073.263.469.166.1
CoTTA [54]81.865.472.557.166.361.3
RoTTA [59]79.462.469.850.663.756.2
SoTTA [17]82.762.771.361.468.264.6
OWTTT [28]65.963.564.456.266.060.6
STAMP85.068.776.066.070.768.2
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.372, + 0.784, + 0.403 + ], + "angle": 0, + "content": "will have a higher AUC score. To consider both generalization and detection performance simultaneously, we also report H-score:" + }, + { + "type": "equation", + "bbox": [ + 0.403, + 0.41, + 0.786, + 0.441 + ], + "angle": 0, + "content": "\\[\nH - s c o r e = \\frac {2 \\cdot A C C \\cdot A U C}{A C C + A U C} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.449, + 0.794, + 0.678 + ], + "angle": 0, + "content": "Implementation Details. We adopt the corruption with severity 5 in all experiments. We consider ResNet-18 [19] as backbones for CIFAR10-C, CIFAR100-C and ResNet-50 [19] for ImageNet-C. For CIFAR10-C and CIFAR100-C, we train the source model using standard cross-entropy loss with the initial learning rate of 0.1 and multi-step learning rate scheduling. For ImageNet-C, we use the released checkpoint from TorchVision library [35]. For a fair comparison, we set the batch size for all methods as 64 and the capacity of our memory bank is 64. Concerning the hyperparameters, we search the hyperparameters on a validation set and set the learning rate to \\(0.1 / 0.05 / 0.01\\), \\(\\mathcal{H}_{thr}\\) to \\(0.25 / 0.9 / 0.8 \\times \\ln |\\mathcal{C}_S|\\), \\(T\\) to \\(150 / 150 / 750\\) for CIFAR10-C/CIFAR100-C/ImageNet-C. Moreover, we set the other hyperparameter the same in experiments with \\(\\beta = 0.1\\), \\(\\rho = 0.05\\), \\(K = 16\\). For parameters' updating, we select the affine parameters in the BN layer for optimization and freeze the top layer's parameter following [39]. More details of hyper-parameters searching and setting of baselines and STAMP can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.696, + 0.325, + 0.71 + ], + "angle": 0, + "content": "4.2 Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Results under outlier-aware TTA setting. We first evaluate TTA methods on the datasets with outliers. The results for CIFAR datasets are illustrated in Table 1. We observe that a simple baseline method, Tent [51], has a competitive generalization performance on distribution shift (achieving an accuracy of \\(80.1\\%\\) on CIFAR10-C). However, it exhibits substantial degradation in the AUC score compared to Source, for instance, dropping from \\(70.4\\%\\) to \\(48.7\\%\\) on CIFAR10-C with Noise outlier dataset. This decline could be attributed to Tent's indiscriminate minimization of entropy across all samples, leading to a gradual loss of its" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.331, + 0.127 + ], + "angle": 0, + "content": "Yu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.144, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 4: Classification accuracies (%) on CIFAR10-C and CIFAR100-C under closed-set TTA setting." + }, + { + "type": "table", + "bbox": [ + 0.235, + 0.185, + 0.77, + 0.432 + ], + "angle": 0, + "content": "
MethodGaussianshotinputdefocusglassrotationzonationshowfrostlogbrightnesscontrastelasticpixelatejpegAvg.
CIFAR10-CSource28.735.224.257.349.066.464.876.762.973.490.331.478.846.574.757.3
BN Stats [36]70.272.063.687.666.585.986.982.180.584.190.885.477.378.674.379.0
Tent [51]76.178.470.587.870.186.887.584.882.085.191.186.679.482.778.581.8
CoTTA [54]76.177.673.987.971.986.987.582.882.085.290.885.779.380.478.381.8
EATA [38]76.577.971.789.170.487.489.085.284.286.091.588.479.983.978.582.6
SAR [39]70.772.166.887.668.285.986.982.180.584.190.885.477.378.674.379.4
RoTTA [59]69.671.062.987.567.085.986.982.479.684.791.273.578.377.974.978.2
SoTTA [17]75.879.271.589.470.687.889.085.684.087.292.487.479.984.579.082.9
OWTTT [28]70.172.063.086.966.585.786.782.780.984.591.283.077.976.874.978.9
STAMP80.982.977.287.674.986.687.785.985.988.190.487.280.386.782.684.3
CIFAR100-CSource12.414.57.236.044.745.145.249.541.636.963.313.257.523.846.635.8
BN Stats [36]41.141.338.963.151.560.863.851.253.553.264.459.058.458.047.653.7
Tent [51]52.352.147.766.956.164.365.358.358.760.067.862.161.863.054.559.4
CoTTA [54]47.047.845.159.654.059.561.353.355.052.962.650.358.161.853.154.8
EATA [38]50.753.548.167.055.664.867.059.159.260.467.763.961.863.354.659.8
SAR [39]55.155.051.268.458.266.067.460.360.861.969.865.563.666.256.861.7
RoTTA [59]35.936.633.860.647.157.760.848.042.250.859.232.153.852.344.447.7
SoTTA [17]51.653.847.466.956.965.368.158.960.160.169.463.162.362.854.860.1
OWTTT [28]41.642.838.863.452.661.664.853.354.854.565.858.460.157.649.354.6
STAMP57.258.552.869.961.468.170.163.363.964.872.269.966.569.259.064.4
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.461, + 0.788, + 0.642 + ], + "angle": 0, + "content": "ability to discern outliers. In addition, our experimental findings demonstrate that with judicious parameter selection, TTA methods can enhance not only the generalization capability against distribution shift but also the detection ability for outliers. Compared to existing TTA methods, STAMP achieves the most favorable performance on both classification and outlier detection. On CIFAR10-C, STAMP achieves the highest ACC, and \\(+4.1\\%\\) huge improvements in AUC compared to the previous methods. Regarding CIFAR100-C, STAMP exhibits \\(+2.8\\%\\) and \\(+5.7\\%\\) improvements compared to the most effective TTA methods on ACC and AUC, respectively. We also report the results on the ImageNet-C in Table 2. STAMP still achieves the highest performance. Specifically, we outperform existing TTA methods with \\(+1.2\\%\\), \\(+1.6\\%\\) on H-score with Places365-C and Textures-C outlier datasets, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.788, + 0.703 + ], + "angle": 0, + "content": "Additionally, we evaluate the TTA algorithm under a more challenging scenario involving homogeneous outlier datasets in Table 3, which makes it more difficult to identify outliers. STAMP also exhibits the most promising performance on CIFAR10-C and CIFAR100-C under this scenario." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.734 + ], + "angle": 0, + "content": "Concerning the above results, we achieve optimal results in most cases, affirming the robustness and effectiveness of STAMP." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.803, + 0.84 + ], + "angle": 0, + "content": "Results under closed-set TTA setting. Additionally, we assess our method in a scenario without outliers as most TTA works do. We present the classification accuracy on the clean dataset in Table 4. Remarkably, our method attains the highest accuracy on both CIFAR10-C and CIFAR100-C datasets, achieving an average improvement of \\(+2.9\\%\\) over the second-best method (SoTTA). This shows that STAMP still has strong applicability even in the traditional situation without outliers." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.293, + 0.115, + 0.731, + 0.131 + ], + "angle": 0, + "content": "Outlier-Aware Test-Time Adaptation with Stable Memory Replay" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.203 + ], + "angle": 0, + "content": "Table 5: Ablation studies on reliable class-balanced memory (RBM), self-weighted entropy minimization (S.W.), sharpness-aware minimization (S.A.), and decay step size (D.S.). Experiments are conducted on CIFAR100-C with the outlier dataset SVHN-C. The first row in the table reports the results of entropy minimization." + }, + { + "type": "table", + "bbox": [ + 0.362, + 0.213, + 0.644, + 0.296 + ], + "angle": 0, + "content": "
S.A.D.S.RBMS.W.ACCAUCH-score
----54.470.461.2
-✓--58.965.361.8
✓---56.472.863.5
✓✓--59.270.864.4
✓✓-✓59.377.767.1
✓✓✓-63.075.068.4
✓✓✓✓63.780.070.9
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.325, + 0.784, + 0.356 + ], + "angle": 0, + "content": "To summarize, all the above results prove that our method achieves the best results under both outlier-aware TTA and conventional closed-set settings." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.38, + 0.39, + 0.394 + ], + "angle": 0, + "content": "4.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.406, + 0.788, + 0.527 + ], + "angle": 0, + "content": "Effect of each component. We delve into the role of each component within our method as follows. First, we provide the results of different weight strategies of entropy minimization in Figure 3. We compare our adaptive weight with static weight and EATA weight [38]. EATA calculates the weight by \\( \\exp[\\mathcal{H}_{thr} - \\mathcal{H}(p(x))] \\) and static weight calculates weight in the same way of Equation (6). Different from adaptive weight, both EATA and static weight don't involve any gradient back-propagation. As depicted in Figure 3, adaptive weight shows the best performance among the strategies." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.528, + 0.788, + 0.575 + ], + "angle": 0, + "content": "Furthermore, we conduct an ablation study on each component of our method, as summarized in Table 5. We observe that the effect of combining test-time step size decay and SAM is better than using either of them alone. Step size decay can" + }, + { + "type": "image", + "bbox": [ + 0.371, + 0.616, + 0.619, + 0.759 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.776, + 0.788, + 0.833 + ], + "angle": 0, + "content": "Fig. 3: Results for the different weight strategies on dataset CIFAR100-C with SVHN-C outlier dataset. Static weight represents that the \\( w_{i} \\) in Equation (6) without gradient back-propagation. EATA weight represents the weight for reliable samples used in EATA [38]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.331, + 0.127 + ], + "angle": 0, + "content": "Yu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.785, + 0.172 + ], + "angle": 0, + "content": "Table 6: H-score (\\%) with the different ratios of outliers to total samples on CIFAR100-C with different outlier datasets." + }, + { + "type": "table", + "bbox": [ + 0.362, + 0.185, + 0.64, + 0.263 + ], + "angle": 0, + "content": "
SVHN-CLSUN-CTinyImageNet-C
5%68.470.268.3
10%70.170.968.8
20%70.971.268.7
33%70.370.168.1
50%69.368.867.4
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.29, + 0.785, + 0.486 + ], + "angle": 0, + "content": "improve the ACC but result in AUC decline. However, combining SAM and step size decay mitigates the decline in AUC. In addition, after the introduction of self-weighted entropy minimization, the AUC of the model further increased to \\(77.7\\%\\). Notably, by removing reliable class-balanced memory from our method, the ACC decreases from \\(63.7\\%\\) to \\(59.3\\%\\). It reveals that reliable class-balanced memory significantly contributes to enhancing the generalization ability to distribution shift. Moreover, by removing the self-weighted entropy minimization in STAMP, its AUC drops from \\(80.0\\%\\) to \\(75.0\\%\\). Self-weighted entropy minimization plays an important role in outlier detection. Besides decreasing the sample entropy, self-weighted entropy minimization incorporates entropy into the optimization weight, allowing outliers that deviate from the majority of samples to receive lower weights, consequently increasing their entropy. This mechanism enhances outlier detection performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.487, + 0.787, + 0.593 + ], + "angle": 0, + "content": "Effect of outlier proportion. In the realistic scenario, the proportion of outliers can vary considerably. We investigate how this variability affects the performance of our method. Specifically, we manipulate the proportion of outliers, ranging from \\(5\\%\\) to \\(50\\%\\), and present the corresponding H-scores in Table 6. The results show that the fluctuation of our method is within \\(3\\%\\). The findings suggest that STAMP exhibits robustness across different scenarios with varying proportions of outliers, as it does not experience significant degradation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.616, + 0.36, + 0.632 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.648, + 0.785, + 0.829 + ], + "angle": 0, + "content": "In this study, we research the challenge of handling outliers in test-time adaptation. To tackle this issue effectively, we introduce a scenario called outlier-aware test-time adaptation, which refers to conducting both recognition and outlier rejection simultaneously. Additionally, we propose STAMP, using augmentation-averaged estimation and a filtering mechanism to store reliable samples in memory. Moreover, we dynamically discard samples with the highest class frequency from memory to maintain a class-balanced prediction. Then we replay the samples stored in the memory with self-weighted entropy minimization. Furthermore, we leverage sharpness-aware minimization and learning rate decay to bolster the stability of our algorithm. Through extensive experiments across multiple benchmarks, we demonstrate that our method surpasses existing TTA methods, underscoring its effectiveness and versatility in open-world applications." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.293, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Outlier-Aware Test-Time Adaptation with Stable Memory Replay" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.788, + 0.252 + ], + "angle": 0, + "content": "We thank Yuhe Ding and Zhengbo Wang for their valuable feedback on our work. This work was funded by the Beijing Municipal Science and Technology Project (No. Z231100010323005), the Beijing Nova Program (No. Z211100002121108), the Young Elite Scientists Sponsorship Program by CAST (2023QNRC001), and the National Natural Science Foundation of China under (No. 62276256)." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.274, + 0.323, + 0.29 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.304, + 0.785, + 0.331 + ], + "angle": 0, + "content": "1. An, S., Park, S., Kim, G., Baek, J., Lee, B., Kim, S.: Context enhanced transformer for single image object detection. arXiv preprint arXiv:2312.14492 (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.333, + 0.785, + 0.358 + ], + "angle": 0, + "content": "2. Boudiaf, M., Mueller, R., Ben Ayed, I., Bertinetto, L.: Parameter-free online test-time adaptation. In: Proc. CVPR (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.36, + 0.785, + 0.386 + ], + "angle": 0, + "content": "3. Cao, H., Xu, Y., Yang, J., Yin, P., Yuan, S., Xie, L.: Multi-modal continual test-time adaptation for 3d semantic segmentation. In: Proc. ICCV (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.388, + 0.785, + 0.427 + ], + "angle": 0, + "content": "4. Chakrabarty, G., Sreenivas, M., Biswas, S.: Santa: Source anchoring network and target alignment for continual test time adaptation. Transactions on Machine Learning Research (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.429, + 0.785, + 0.455 + ], + "angle": 0, + "content": "5. Chen, D., Wang, D., Darrell, T., Ebrahimi, S.: Contrastive test-time adaptation. In: Proc. CVPR (2022) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.457, + 0.785, + 0.482 + ], + "angle": 0, + "content": "6. Cimpoi, M., Maji, S., Kokkinos, I., Mohamed, S., Vedaldi, A.: Describing textures in the wild. In: Proc. CVPR (2014) 3, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.484, + 0.785, + 0.524 + ], + "angle": 0, + "content": "7. Croce, F., Andriushchenko, M., Sehwag, V., Debenedetti, E., Flammarion, N., Chiang, M., Mittal, P., Hein, M.: Robustbench: a standardized adversarial robustness benchmark. In: Proc. NeurIPS (2021) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.526, + 0.785, + 0.551 + ], + "angle": 0, + "content": "8. Davis, J., Goadrich, M.: The relationship between precision-recall and roc curves. In: Proc. ICML (2006) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.553, + 0.785, + 0.579 + ], + "angle": 0, + "content": "9. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: Proc. CVPR (2009) 3, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.58, + 0.785, + 0.606 + ], + "angle": 0, + "content": "0. DeVries, T., Taylor, G.W.: Learning confidence for out-of-distribution detection in neural networks. arXiv preprint arXiv:1802.04865 (2018) 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.608, + 0.785, + 0.647 + ], + "angle": 0, + "content": "1. Ding, Y., Liang, J., Jiang, B., Zheng, A., He, R.: Maps: A noise-robust progressive learning approach for source-free domain adaptive keypoint detection. IEEE Transactions on Circuits and Systems for Video Technology (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.649, + 0.785, + 0.688 + ], + "angle": 0, + "content": "2. Ding, Y., Sheng, L., Liang, J., Zheng, A., He, R.: Proxymix: Proxy-based mixup training with label refinery for source-free domain adaptation. Neural Networks 167, 92-103 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.69, + 0.785, + 0.731 + ], + "angle": 0, + "content": "3. Farahani, A., Voghoei, S., Rasheed, K., Arabnia, H.R.: A brief review of domain adaptation. Advances in Data Science and Information Engineering: Proceedings from ICDATA 2020 and IKE 2020 pp. 877-894 (2021) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.732, + 0.785, + 0.757 + ], + "angle": 0, + "content": "4. Foret, P., Kleiner, A., Mobahi, H., Neyshabur, B.: Sharpness-aware minimization for efficiently improving generalization. In: Proc. ICLR (2021) 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.759, + 0.785, + 0.785 + ], + "angle": 0, + "content": "5. Gebru, T., Hoffman, J., Fei-Fei, L.: Fine-grained recognition in the wild: A multitask domain adaptation approach. In: Proc. ICCV (2017) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.786, + 0.785, + 0.812 + ], + "angle": 0, + "content": "6. Gong, T., Jeong, J., Kim, T., Kim, Y., Shin, J., Lee, S.J.: Note: Robust continual test-time adaptation against temporal correlation. In: Proc. NeurIPS (2022) 4, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "7. Gong, T., Kim, Y., Lee, T., Chottananurak, S., Lee, S.J.: Sotta: Robust test-time adaptation on noisy data streams. In: Proc. NeurIPS (2023) 4, 5, 8, 9, 10, 11, 12" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.304, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.331, + 0.127 + ], + "angle": 0, + "content": "Yu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "18. Goyal, S., Sun, M., Raghunathan, A., Kolter, J.Z.: Test time adaptation via conjugate pseudo-labels. In: Proc. NeurIPS (2022) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.785, + 0.203 + ], + "angle": 0, + "content": "19. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proc. CVPR (2016) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.785, + 0.231 + ], + "angle": 0, + "content": "20. Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. In: Proc. ICLR (2018) 3, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.259 + ], + "angle": 0, + "content": "21. Hendrycks, D., Gimpel, K.: A baseline for detecting misclassified and out-of-distribution examples in neural networks. In: Proc. ICLR (2016) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.785, + 0.286 + ], + "angle": 0, + "content": "22. Ioffe, S., Szegedy, C.: Batch normalization: Accelerating deep network training by reducing internal covariate shift. In: Proc. ICML (2015) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.342 + ], + "angle": 0, + "content": "23. Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A.A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., et al.: Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences 114(13), 3521-3526 (2017) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.785, + 0.37 + ], + "angle": 0, + "content": "24. Krizhevsky, A.: Learning multiple layers of features from tiny images. Master's thesis, University of Tront (2009) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.396 + ], + "angle": 0, + "content": "25. Le, Y., Yang, X.: Tiny imagenet visual recognition challenge. CS 231N 7(7), 3 (2015) 3, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.424 + ], + "angle": 0, + "content": "26. Lee, J., Das, D., Choo, J., Choi, S.: Towards open-set test-time adaptation utilizing the wisdom of crowds in entropy minimization. In: Proc. CVPR (2023) 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.452 + ], + "angle": 0, + "content": "27. Lee, K., Lee, K., Lee, H., Shin, J.: A simple unified framework for detecting out-of-distribution samples and adversarial attacks. In: Proc. NeurIPS (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.492 + ], + "angle": 0, + "content": "28. Li, Y., Xu, X., Su, Y., Jia, K.: On the robustness of open-world test-time training: Self-training with dynamic prototype expansion. In: Proc. CVPR (2023) 4, 9, 10, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.494, + 0.785, + 0.522 + ], + "angle": 0, + "content": "29. Liang, J., He, R., Tan, T.: A comprehensive survey on test-time adaptation under distribution shifts. International Journal of Computer Vision (2024) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.785, + 0.55 + ], + "angle": 0, + "content": "30. Liang, J., Hu, D., Feng, J.: Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In: Proc. ICML (2020) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.577 + ], + "angle": 0, + "content": "31. Liang, J., Sheng, L., Wang, Z., He, R., Tan, T.: Realistic unsupervised clip finetuning with universal entropy optimization. In: Proc. ICML (2024) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.604 + ], + "angle": 0, + "content": "32. Liang, S., Li, Y., Srikant, R.: Enhancing the reliability of out-of-distribution image detection in neural networks. In: Proc. ICLR (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.605, + 0.785, + 0.632 + ], + "angle": 0, + "content": "33. Lin, Z., Roy, S.D., Li, Y.: Mood: Multi-level out-of-distribution detection. In: Proc. CVPR (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.66 + ], + "angle": 0, + "content": "34. Liu, W., Wang, X., Owens, J., Li, Y.: Energy-based out-of-distribution detection. In: Proc. NeurIPS (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.785, + 0.688 + ], + "angle": 0, + "content": "35. Maintainers, T., Contributors: Torchvision: Pytorch's computer vision library (2016) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.785, + 0.729 + ], + "angle": 0, + "content": "36. Nado, Z., Padhy, S., Sculley, D., D'Amour, A., Lakshminarayanan, B., Snoek, J.: Evaluating prediction-time batch normalization for robustness under covariate shift. In: Proc. ICML Workshops (2020) 3, 9, 10, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.729, + 0.785, + 0.757 + ], + "angle": 0, + "content": "37. Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B., Ng, A.Y.: Reading digits in natural images with unsupervised feature learning (2011) 3, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.798 + ], + "angle": 0, + "content": "38. Niu, S., Wu, J., Zhang, Y., Chen, Y., Zheng, S., Zhao, P., Tan, M.: Efficient test-time model adaptation without forgetting. In: Proc. ICML (2022) 2, 4, 5, 6, 7, 9, 10, 11, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.839 + ], + "angle": 0, + "content": "39. Niu, S., Wu, J., Zhang, Y., Wen, Z., Chen, Y., Zhao, P., Tan, M.: Towards stable test-time adaptation in dynamic wild world. In: Proc. ICLR (2022) 4, 7, 8, 9, 10, 11, 12" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.839 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.293, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Outlier-Aware Test-Time Adaptation with Stable Memory Replay" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "40. Nixon, J., Dusenberry, M.W., Zhang, L., Jerfel, G., Tran, D.: Measuring calibration in deep learning. In: Proc. CVPR Workshops (2019) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.785, + 0.202 + ], + "angle": 0, + "content": "41. Park, S., Yang, S., Choo, J., Yun, S.: Label shift adapter for test-time adaptation under covariate and label shifts. In: Proc. CVPR (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.203, + 0.785, + 0.242 + ], + "angle": 0, + "content": "42. Press, O., Schneider, S., Kümmerer, M., Bethge, M.: Rdumb: A simple approach that questions our progress in continual test-time adaptation. In: Proc. NeurIPS (2023) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.243, + 0.785, + 0.283 + ], + "angle": 0, + "content": "43. Schneider, S., Rusak, E., Eck, L., Bringmann, O., Brendel, W., Bethge, M.: Improving robustness against common corruptions by covariate shift adaptation. In: Proc. NeurIPS (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.284, + 0.785, + 0.311 + ], + "angle": 0, + "content": "44. Shanmugam, D., Blalock, D., Balakrishnan, G., Guttag, J.: Better aggregation in test-time augmentation. In: Proc. CVPR (2021) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.312, + 0.785, + 0.338 + ], + "angle": 0, + "content": "45. Sheng, L., Liang, J., He, R., Wang, Z., Tan, T.: Adaptguard: Defending against universal attacks for model adaptation. In: Proc. ICCV (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.339, + 0.785, + 0.365 + ], + "angle": 0, + "content": "46. Tan, M., Le, Q.: Efficientnet: Rethinking model scaling for convolutional neural networks. In: Proc. ICML (2019) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.365, + 0.785, + 0.392 + ], + "angle": 0, + "content": "47. Tong, P., Su, W., Li, H., Ding, J., Haoxiang, Z., Chen, S.X.: Distribution free domain generalization. In: Proc. ICML (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.393, + 0.785, + 0.419 + ], + "angle": 0, + "content": "48. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: Proc. NeurIPS (2017) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.42, + 0.785, + 0.446 + ], + "angle": 0, + "content": "49. Volpi, R., De Jorge, P., Larlus, D., Csurka, G.: On the road to online adaptation for semantic image segmentation. In: Proc. CVPR (2022) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.447, + 0.785, + 0.487 + ], + "angle": 0, + "content": "50. Vyas, A., Jammalamadaka, N., Zhu, X., Das, D., Kaul, B., Willke, T.L.: Out-of-distribution detection using an ensemble of self supervised leave-out classifiers. In: Proc. ECCV (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.785, + 0.514 + ], + "angle": 0, + "content": "51. Wang, D., Shelhamer, E., Liu, S., Olshausen, B., Darrell, T.: Tent: Fully test-time adaptation by entropy minimization. In: Proc. ICML (2020) 2, 4, 7, 9, 10, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.515, + 0.785, + 0.555 + ], + "angle": 0, + "content": "52. Wang, J., Lan, C., Liu, C., Ouyang, Y., Qin, T., Lu, W., Chen, Y., Zeng, W., Yu, P.S.: Generalizing to unseen domains: A survey on domain generalization. IEEE Transactions on Knowledge and Data Engineering 35(8), 8052-8072 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.556, + 0.785, + 0.582 + ], + "angle": 0, + "content": "53. Wang, M., Deng, W.: Deep visual domain adaptation: A survey. Neurocomputing 312, 135-153 (2018) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.583, + 0.785, + 0.609 + ], + "angle": 0, + "content": "54. Wang, Q., Fink, O., Van Gool, L., Dai, D.: Continual test-time domain adaptation. In: Proc. CVPR (2022) 4, 9, 10, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.61, + 0.785, + 0.636 + ], + "angle": 0, + "content": "55. Wang, Y., Li, B., Che, T., Zhou, K., Liu, Z., Li, D.: Energy-based open-world uncertainty modeling for confidence calibration. In: Proc. CVPR (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.637, + 0.785, + 0.663 + ], + "angle": 0, + "content": "56. You, K., Long, M., Wang, J., Jordan, M.I.: How does learning rate decay help modern neural networks? arXiv preprint arXiv:1908.01878 (2019) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.664, + 0.785, + 0.704 + ], + "angle": 0, + "content": "57. Yu, F., Seff, A., Zhang, Y., Song, S., Funkhouser, T., Xiao, J.: Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365 (2015) 3, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.705, + 0.785, + 0.743 + ], + "angle": 0, + "content": "58. Yu, Y., Sheng, L., He, R., Liang, J.: Benchmarking test-time adaptation against distribution shifts in image classification. arXiv preprint arXiv:2307.03133 (2023) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.745, + 0.785, + 0.771 + ], + "angle": 0, + "content": "59. Yuan, L., Xie, B., Li, S.: Robust test-time adaptation in dynamic scenarios. In: Proc. CVPR (2023) 2, 3, 9, 10, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.812 + ], + "angle": 0, + "content": "60. Yurtsever, E., Lambert, J., Carballo, A., Takeda, K.: A survey of autonomous driving: Common practices and emerging technologies. IEEE Access 8, 58443-58469 (2020) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "61. Zhang, M., Levine, S., Finn, C.: Memo: Test time robustness via adaptation and augmentation. In: Proc. NeurIPS (2022) 4, 5" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.331, + 0.127 + ], + "angle": 0, + "content": "Yu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.188 + ], + "angle": 0, + "content": "62. Zhang, Y., Wang, X., Jin, K., Yuan, K., Zhang, Z., Wang, L., Jin, R., Tan, T.: Adanpc: Exploring non-parametric classifier for test-time adaptation. In: Proc. ICML (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.189, + 0.785, + 0.217 + ], + "angle": 0, + "content": "63. Zhao, H., Liu, Y., Alahi, A., Lin, T.: On pitfalls of test-time adaptation. In: Proc. ICML (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.259 + ], + "angle": 0, + "content": "64. Zhou, B., Lapedriza, A., Khosla, A., Oliva, A., Torralba, A.: Places: A 10 million image database for scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence 40(6), 1452-1464 (2017) 3, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.785, + 0.286 + ], + "angle": 0, + "content": "65. Zhou, Z., Guo, L.Z., Jia, L.H., Zhang, D., Li, Y.F.: Ods: test-time adaptation in the presence of open-world data shift. In: Proc. ICML (2023) 4" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.286 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/STAMP_ Outlier-Aware Test-Time Adaptation with Stable Memory Replay/20b10b92-8cd4-4e35-ad35-a50ac0ff44cf_origin.pdf b/2024/STAMP_ Outlier-Aware Test-Time Adaptation with Stable Memory Replay/20b10b92-8cd4-4e35-ad35-a50ac0ff44cf_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5bbb6dcc441ad596ff5f71b0a69cbbc145a2b3da --- /dev/null +++ b/2024/STAMP_ Outlier-Aware Test-Time Adaptation with Stable Memory Replay/20b10b92-8cd4-4e35-ad35-a50ac0ff44cf_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1eff8f6f01733393e917bbb9995882033b2117a024e33d8dd953eefe96736c4a +size 1349884 diff --git a/2024/STAMP_ Outlier-Aware Test-Time Adaptation with Stable Memory Replay/full.md b/2024/STAMP_ Outlier-Aware Test-Time Adaptation with Stable Memory Replay/full.md new file mode 100644 index 0000000000000000000000000000000000000000..955b1f338b744b4e9cdefb7140fa1be71615fb34 --- /dev/null +++ b/2024/STAMP_ Outlier-Aware Test-Time Adaptation with Stable Memory Replay/full.md @@ -0,0 +1,308 @@ +![](images/2661f00c92b9e29c0adf43c07a615933123144781e6f9d47b5962a503944823d.jpg) + +# STAMP: Outlier-Aware Test-Time Annotation with Stable Memory Replay + +Yongcan Yu $^{1,2}$ , Lijun Sheng $^{1,3}$ , Ran He $^{1,2}$ , and Jian Liang $^{1,2\star}$ + +$^{1}$ NLPR & MAIS, Institute of Automation, Chinese Academy of Sciences + +$^{2}$ School of Artificial Intelligence, University of Chinese Academy of Sciences + +3 University of Science and Technology of China + +{yuyongcan0223, liangjian92}@gmail.com + +Abstract. Test-time adaptation (TTA) aims to address the distribution shift between the training and test data with only unlabeled data at test time. Existing TTA methods often focus on improving recognition performance specifically for test data associated with classes in the training set. However, during the open-world inference process, there are inevitably test data instances from unknown classes, commonly referred to as outliers. This paper pays attention to the problem that conducts both sample recognition and outlier rejection during inference while outliers exist. To address this problem, we propose a new approach called STAble Memory rePlay (STAMP), which performs optimization over a stable memory bank instead of the risky mini-batch. In particular, the memory bank is dynamically updated by selecting low-entropy and label-consistent samples in a class-balanced manner. In addition, we develop a self-weighted entropy minimization strategy that assigns higher weight to low-entropy samples. Extensive results demonstrate that STAMP outperforms existing TTA methods in terms of both recognition and outlier detection performance. The code is released at https://github.com/yuyongcan/STAMP. + +Keywords: online test-time adaptation $\cdot$ outlier detection $\cdot$ memory bank $\cdot$ self-weighted entropy minimization + +# 1 Introduction + +Deep neural networks (DNNs) [46, 48] have achieved remarkable performance across various tasks. Nevertheless, the impressive performance is based on the assumption that training data and test data come from the same distribution. However, in real-world scenarios, the presence of distribution shifts is ubiquitous [13, 30], leading to a degradation in the prediction performance [7]. Taking autonomous driving [60] as an example, changing weather conditions causes systems trained on laboratory data to make incorrect decisions. To mitigate the impact of distribution shift, test-time adaptation (TTA) [29, 58] has emerged as + +![](images/afdec46572993664b26a61c6f1c8adaaed02f7115bbb7ecdd56e1c0e1a343085.jpg) +Fig. 1: The illustration of outlier-aware test-time adaptation. The test data stream consists of both normal samples and outliers. When the test data arrives, the algorithm adapts the model and provides two outputs, a prediction and an OOD score. OOD score indicates the likelihood that a sample is an outlier. The detector in the deployment scenario chooses to keep the prediction or reject samples according to the OOD score. + +a promising approach, adapting models with unlabeled samples during inference. Nowadays, TTA has achieved remarkable results in various tasks, such as image classification [51, 62], semantic segmentation [3, 49], object detection [1]. + +Previous TTA works [5, 38, 59] always assume that test data and training data share the same class space. Nevertheless, in realistic open-world scenarios, test data streams will inevitably contain samples from new semantic classes, also known as outliers. The introduction of outliers brings two additional risks to adaptation tasks: (i) Classifying outliers into known categories poses safety risks, especially in contexts like autonomous driving, where systems should request human intervention to handle unknown objects to avoid traffic accidents. (ii) Self-supervised loss (e.g., entropy minimization, pseudo-labeling) calculated with outliers may misguide the optimization process, decreasing the model's recognition ability on known classes. + +To better tackle the above risks, we pay attention to a more realistic TTA scenario termed outlier-aware test-time adaptation. Outlier-aware test-time adaptation aims to conduct recognition for normal samples and reject to make predictions for outliers. The overview of outlier-aware test-time adaptation is illustrated in Figure 1. When test data mixed with outliers arrives, the algorithm adapts the model and generates the prediction and an out-of-distribution (OOD) score for each sample. Then we decide to keep the prediction or reject the sample based on its OOD score. The sample with a higher OOD score is more likely to be considered an outlier. + +To address the outlier-aware test-time adaptation, we additionally design a novel algorithm called STAble Memory rePlay (STAMP). STAMP consists of three components: reliable class-balanced memory, self-weighted entropy minimization, and stable optimization strategy. Reliable class-balanced memory stores reliable samples for optimization by filtering samples based on their pre + +diction's consistency and entropy. To maintain the class-balanced prediction of the model, it dynamically discards samples with the class that participates most frequently in optimization when the memory is full. Furthermore, we adopt self-weighted entropy minimization to assign greater weights to low-entropy samples in memory and adjust the weights for each sample adaptively. To ensure the stability of the optimization process, we dynamically adjust the optimization step size to quickly adapt to the target domain at the early stage and mitigate error accumulation later. Moreover, we apply the sharpness-aware minimization technique [14] to further reduce the impact of noisy gradients. + +To demonstrate the effectiveness of STAMP, we compare it with eight state-of-the-art TTA algorithms. We select three commonly used datasets in TTA, including CIFAR10-C, CIFAR100-C [10, 20], and ImageNet-C [9, 20], alongside some outlier datasets (i.e., Noise, SVHN [37], LSUN [57], TinyImageNet [25], Textures [6], Places365 [64]). Experimental results show that our method achieves superior performance on recognition and outlier detection under the outlier-aware TTA setting. The experiment on conventional protocol indicates that STAMP also performs well in the scenario without outliers. In summary, we list our contributions as follows: + +- We introduce outlier-aware test-time adaptation, a TTA scenario considering both normal sample recognition and outlier rejection during inference. +- We propose a memory that dynamically stores reliable samples for replay. To prevent noisy samples from entering the memory, we propose a filtering strategy based on the consistency and entropy of the model's prediction. +- We design self-weighted entropy minimization, which guides the model to adaptively learn to assign higher weights to valuable samples. +- We propose a proper metric that considers both generalization and outlier detection performance. Extensive experiments compared with eight TTA algorithms on several benchmarks demonstrate the effectiveness of our method. + +# 2 Related Work + +# 2.1 Test-Time Adaptation + +TTA intends to adapt the model pre-trained on the source domain to the target domain while the distribution shift exists. Compared to domain generalization [47, 52] and domain adaptation [15, 53], TTA needs to access neither the source data nor the target labels, making it has a wide range of applications [11, 12, 45]. TTA methods can be mainly divided into two solutions: batch normalization (BN) calibration and self-training. In BN calibration, Nado et al. [36] utilize the BN statistics of the test batch $(\mu_t, \sigma_t)$ to conduct normalization in the BN layer, while Schneider [43] propose a more comprehensive approach considering statistics from both the source domain $(\mu_s, \sigma_s)$ and the target domain to calibrate the domain shifts. To conquer the correlation sampling problem, Yuan et al. [59] adopt robust batch normalization, which updates the statistic $(\mu_t, \sigma_t)$ in a moving average manner with the test data stream. + +Currently, the primary paradigm in TTA is self-training [2, 61], where Wang et al. [51] adopt entropy minimization as a self-supervised loss for optimization and Chen et al. [5] introduce contrastive loss to TTA. However, unsupervised loss may contain noise information. Niu et al. [38] first propose filtering out high-entropy samples, a mechanism adopted by subsequent works [17, 39]. Subsequently, more attention has been paid to addressing more realistic adaptation scenarios such as continually changing target domains [4, 54], label shift [41, 65], and temporally correlated shift [16, 63]. In our study, we focus on a more realistic scenario where test data contains samples with new semantic classes. Some existing works [17, 26, 28] discuss the TTA algorithm in scenarios containing outliers. However, Lee et al. [26] and Gong et al. [17] solely focus on the generalization ability to domain shift, neglecting the issue of rejecting the outliers. A recent work [28] discusses outlier detection performance based on TTA tasks, but it only provides binary predictions to indicate whether the test data is outlier. In contrast, outlier-aware test-time adaptation is more flexible, providing an OOD score, and different systems can conduct sample rejection based on security requirements and the score. + +# 2.2 Out-of-Distribution Detection + +OOD detection aims to identify instances that significantly deviate from the distribution of the training data [31]. Their primary focus is usually on detecting samples whose labels were not seen during the training phase. This area has been extensively researched, with one common approach being the utilization of confidence scores such as max softmax probability [21], energy score [33, 34], and others [27] derived from the model's output to detect OOD samples. Further improvements have been made by Liang et al. [32] through techniques like temperature scaling and input preprocessing to enhance the effectiveness of confidence scores. Another avenue for OOD sample detection lies in training-based methods [10, 50, 55]. These techniques involve training a branch using both indistribution data and OOD data to enable the model to distinguish OOD data during training. However, these methods alter the training process of the source model, which can not be directly applied to adaptation tasks based on a well-trained model. Moreover, OOD detection traditionally assumes that both the training and test data originate from the same domain, a condition not met in TTA scenarios. Under the realistic scenario, there may be a domain shift between test data and training data, which brings greater challenges to detection. + +# 3 Methodology + +In this section, we first introduce the outlier-aware test-time adaptation problem in Section 3.1. Then we describe our proposed method, including how to build a memory bank and conduct the optimization in Section 3.2. + +# 3.1 Outlier-Aware Test-Time Adaptation + +Let $\mathcal{D}_S$ be the source domain dataset with label space $\mathcal{C}_S$ follows a distribution $P_S(x,y)$ . Similarly, let $\mathcal{D}_T = \{(x_i,y_i)\}_{i=1}^{N_T}$ be the target domain dataset with label space $\mathcal{C}_T$ follows a distribution $P_T(x,y)$ . Given a source model $f_{\theta_s}(x)$ well trained on $\mathcal{D}_S$ with parameter $\theta_s$ , TTA aims to adapt the source model to the target domain while $P_S(x) \neq P_T(x)$ . + +In traditional closed-set TTA, it is commonly assumed that the $\mathcal{D}_S$ and $\mathcal{D}_T$ share the same label space, i.e., $\mathcal{C}_S = \mathcal{C}_T$ . However, in outlier-aware test-time adaptation, we consider $\mathcal{C}_S \subseteq \mathcal{C}_T$ while some outliers with new semantic classes will be encountered in $\mathcal{D}_T$ . The illustration of outlier-aware test-time adaption is shown in Figure 1. In addition to adapting the model and providing the prediction, we also generate an OOD score $\delta_i$ for each sample $x_i$ . Then we adopt an outlier detector $g(\cdot)$ as follow: + +$$ +g \left(x _ {i}\right) = \left\{ \begin{array}{l l} 1, & \text {i f} \delta_ {i} \geq \delta_ {t h r} \\ 0, & \text {i f} \delta_ {i} < \delta_ {t h r}, \end{array} \right. \tag {1} +$$ + +where $\delta_{thr}$ is a threshold that can be adjusted according to the requirements of different application scenarios. The samples with $g(x_i) = 1$ will be considered outliers and rejected for recognition. And we output the prediction for samples with $g(x_i) = 0$ . In our work, we take the entropy of the model prediction as our OOD score. + +# 3.2 Stable Memory Replay + +In this section, we introduce stable memory replay (STAMP). STAMP first uses a filtering mechanism to filter out unreliable samples in the coming mini-batch and stores reliable samples in memory. Then, STAMP optimizes the model using in-memory samples by applying self-weighted entropy minimization. In order to further improve the stability of optimization, STAMP uses test-time step size decay and sharpness-aware minimization to guide optimization. The illustration of STAMP is shown in Figure 2. + +Reliable Class-Balanced Memory. Since excluding unreliable samples from optimization has shown great effectiveness [17, 38], we propose reliable class-balanced memory to store reliable samples for optimization. To have a more robust estimation of the sample's reliability, we get its prediction in an augmentation-averaged way. Moreover, we prevent unreliable samples from getting into the memory bank $\mathcal{M}$ by employing consistency filtering and entropy filtering. To preserve the model's balanced class prior distribution, we dynamically discard samples with the highest class frequency from the memory bank. + +To ensure robust predictions, we adopt the augmentation-averaged method, which addresses the issue of deep neural networks being overconfident with out-of-distribution samples. This overconfidence often blurs the line between in-distribution and out-of-distribution samples. However, research has demonstrated that employing test-time augmentation [44, 61] enhances model calibration [40]. Thus, when presented with a test sample $x$ , we derive its prediction + +![](images/4241aedecb67cdbdce38260bb146b30410a5af6c66a8456efd2282f25356f565.jpg) +Fig. 2: The overview of STAMP. STAMP retains reliable samples for optimizing the network by maintaining a memory bank. When test samples arrive, STAMP generates an augmentation-averaged prediction. Two filtering mechanisms are designed to filter out inconsistent and high-entropy samples and add the remaining to the memory. A frequency vector is maintained to select samples to discard when the memory is full. STAMP leverages the reliable samples in the memory to optimize the model with the SAM optimizer alongside a decaying step strategy. + +through the following process: + +$$ +\hat {p} (x) = \frac {1}{K} \sum_ {k = 1} ^ {K} f _ {\theta_ {t}} \left(\tau_ {k} (x)\right), \tag {2} +$$ + +where $\tau(\cdot)$ is the operation of augmentation, $K$ is the total number of augmented views. By averaging predictions over augmented samples, the model can better eliminate the uncertainty inherent in its predictions, improving calibration and obtaining more reliable confidence estimates. In the supplementary material, we discuss the impact of the augmentation-averaged estimation. + +After acquiring the prediction $\hat{p}$ , we implement filtering strategies to prevent unreliable samples from being added to memory for two reasons. Firstly, optimization using unreliable samples can result in error accumulation over time [38]. Secondly, we prevent outliers from entering the memory and reduce the entropy of normal samples in the memory. This action facilitates the model's ability to distinguish between normal samples and outliers. Since reliable samples that are beneficial to optimization tend to have the same pseudo-label as the source model, we employ consistency filtering to remove samples whose predictions are inconsistent with those of the source model: + +$$ +m _ {c o n} (x) = \mathbb {I} \left(\arg \max _ {c} [ \hat {p} (x) ] _ {c} = \arg \max _ {c} [ f _ {\theta_ {s}} (x) ] _ {c}\right), \tag {3} +$$ + +Algorithm 1 Reliable Class-Balanced Memory +Input: memory bank $\mathcal{M}$ , a batch of data $\{x_{i}\}_{i = 1}^{B}$ , entropy threshold $\mathcal{H}_{thr}$ , class frequency $\{\xi_c\}_{c = 1}^{|C_S|}$ . +for $i = 1$ to $B$ do +Calculate $\hat{p}(x_i)$ by Equation (2) and $\hat{y}_i = \arg \max_c [\hat{p}(x_i)]_c$ . Calculate $m_{con}(x_i)$ , $m_{ent}(x_i)$ of $x_i$ by Equation (3), (4). +if $m_{con}(x_i) = 1$ and $m_{ent}(x_i) = 1$ then +if $\mathcal{M}$ is full then +Discard one sample with highest class frequency $\xi$ from $\mathcal{M}$ . +end if +add $(x_i, \hat{y}_i)$ into $\mathcal{M}$ . +end if +end for +Update $\xi$ with Equation (5). + +where $\mathbb{I}(\cdot)$ is the indicator function. Moreover, including high entropy samples for optimization can harm the stability of the model [38, 39], while samples with low entropy always possess greater value and reliability. Therefore, we use entropy filtering to filter out samples with high entropy from entering the memory. + +$$ +m _ {e n t} (x) = \mathbb {I} \left(\mathcal {H} (\hat {p} (x)) < \mathcal {H} _ {t h r}\right), \tag {4} +$$ + +where $\mathcal{H}(q) = -\sum_{c} q_c \log q_c$ denotes the entropy of the prediction vector $q$ . We filter out samples with $\mathcal{H}(\hat{p}) \geq \mathcal{H}_{thr}$ , where $\mathcal{H}_{thr}$ is a threshold. Considering both filtering strategies, we add sample $(x, \hat{y})$ into memory bank $\mathcal{M}$ if the sample satisfies consistency filtering and entropy filtering simultaneously, i.e., $m_{con}(x) = 1$ and $m_{ent}(x) = 1$ . + +Since the number of samples stored in memory is limited, we design a class-balanced strategy to choose a sample to discard when the memory is full. Since we adopt entropy minimization as our optimization objective, an issue arises when the model excessively optimizes samples from a single category over a short period. It leads to a bias towards predicting the dominant category. To counteract this bias, we dynamically discard samples from the memory that belong to the category with the highest update frequency. After each optimization of the model, we update the class frequencies in a moving average manner: + +$$ +\xi_ {c} ^ {t} = (1 - \beta) \xi_ {c} ^ {t - 1} + \beta \nu_ {c} ^ {t}, \tag {5} +$$ + +where $\nu_{c}^{t}$ represents the number of samples with class label $c$ in memory and $\beta$ is moving average coefficient hyperparameter. While the memory is full, we discard a sample from the class with the highest $\xi$ . We summarize the work process of reliable class-balanced memory in Algorithm 1. + +Self-Weighted Entropy Minimization. After updating the memory, we replay the reliable samples stored in the memory $\mathcal{M}$ to optimize the model. Following [51], we employ entropy minimization as our loss function. This self-supervised loss encourages the model to exhibit higher confidence in its predic + +Algorithm 2 The Pipeline of STAMP +Initialize: Memory bank $\mathcal{M} = \emptyset$ zero-initialized class frequency $\{\xi_c\}_{c = 1}^{|C_S|}$ +Input: Unlabeled test samples $\mathcal{D}_T = \{x_i\}_{i = 1}^{NT}$ , a trained model $f_{\theta_s}(\cdot)$ +for each online mini-batch $X_{b}$ do Calculate $\hat{p} (x)$ by Equation (2) and $\mathcal{H}(\hat{p} (x))$ for all $x\in X_b$ . Obtain the prediction $\hat{y} = \arg \max_{c}[\hat{p} (x)]_{c}$ for all $x\in X_b$ Update memory bank $\mathcal{M}$ by Algorithm 1. Compute $\mathcal{L}_{sem}$ with $\mathcal{M}$ by Equation (6). Update $\theta_t$ with SAM in Equation (7) and decay step size in Equation (8). +end for +Output: The prediction and OOD score $\{\hat{y}_i,\mathcal{H}(\hat{p} (x_i))\}_{i = 1}^{NT}$ + +tions. In order to give more valuable samples a higher importance in optimization, we weight samples in $\mathcal{M}$ in a self-weighted manner. Given samples $\{x_{i}\}_{i = 1}^{|M|}$ , we calculate the following loss function: + +$$ +\mathcal {L} _ {s e m} (\mathcal {M}, \theta_ {t}) = \sum_ {x _ {i} \in \mathcal {M}} \frac {w _ {i}}{\sum_ {j} w _ {j}} \cdot \mathcal {H} \left(f _ {\theta_ {t}} \left(x _ {i}\right)\right) \tag {6} +$$ + +where $w_{i} = e^{-\mathcal{H}(f_{\theta_{t}}(x_{i}))}$ denotes the weight of sample $x_{i}$ . It's worth noting that $w_{i}$ is also involved with the gradient back-propagation. The reason for this adaptive weighting is two-fold: (i) it gives greater weights to samples with lower entropy, which are considered more reliable. (ii) The model needs to learn the weight of each sample and adaptively give important samples higher weights, instead of solely minimizing entropy for all samples. + +Stable Optimization Strategy. To optimize the model in a stable way, we employ sharpness-aware minimization (SAM) [14] and step size decay in parameter updating. SAM is a technique that aims to improve generalization performance in deep learning models by minimizing a loss function while simultaneously controlling the sharpness of the loss landscape. Previous work [14] demonstrates that SAM exhibits resilience against noisy labels, which is also a prevalent challenge in TTA. Therefore, we consider SAM as a good choice of the optimizer in the adaptation scenario [17, 39]. Mathematically, this involves minimizing the following objective function: + +$$ +\min _ {\theta_ {t}} \mathcal {L} _ {s e m} ^ {S A} (\mathcal {M}, \theta_ {t}), \text {w h e r e} \mathcal {L} _ {s e m} ^ {S A} (\mathcal {M}, \theta_ {t}) \triangleq \max _ {| | \epsilon | | _ {2} \leq \rho} \mathcal {L} _ {s e m} (\mathcal {M}, \theta_ {t} + \epsilon), \tag {7} +$$ + +where $\rho$ represents the radius of the neighborhood. SAM minimizes the maximum loss within the parameters' vicinity. + +Recent research [26, 42] has revealed that under long-term adaptation, TTA algorithms may suffer from substantial error accumulation and consequent performance degradation. This issue becomes even more severe when test data contains outliers. To mitigate this problem, we introduce a test-time step size (a.k.a. learning rate) decay strategy. To achieve this, we update the model's parameters + +Table 1: Results (\%) on CIFAR10-C and CIFAR100-C datasets under outlier-aware TTA setting across four outlier datasets. We report the accuracy (ACC), AUC score, and H-score. All the results are averaged over 15 different corruptions. The results in bold denote the best performance. + +
NoiseSVHN-CLSUN-CTinyImageNet-CAvg.
ACCAUCH-scoreACCAUCH-scoreACCAUCH-scoreACCAUCH-scoreACCAUCH-score
CIFAR10-CSource57.370.462.357.367.461.157.362.859.657.364.559.457.366.360.6
BN Stats [36]72.968.670.678.775.376.979.479.479.479.072.975.877.574.075.7
Tent [51]77.448.759.780.854.965.181.262.370.481.165.672.480.157.966.9
EATA [38]72.968.570.678.875.376.979.479.479.478.973.175.977.574.175.7
SAR [39]72.968.570.678.775.376.979.479.479.479.072.975.877.574.075.7
CoTTA [54]77.362.467.381.678.680.182.284.283.281.975.378.480.875.177.2
RoTTA [54]77.674.375.678.476.077.278.879.579.178.673.375.878.375.876.9
SoTTA [17]77.851.761.679.372.875.979.877.978.879.672.675.979.168.873.1
OWTTT [28]62.364.458.566.175.369.663.178.968.556.358.856.262.069.363.2
STAMP77.980.778.882.379.080.583.584.984.282.675.178.681.679.980.5
CIFAR100-CSource35.843.138.035.849.440.135.858.243.235.857.142.735.852.041.0
BN Stats [36]45.880.958.452.772.560.953.773.862.053.268.659.751.374.060.3
Tent [51]47.955.851.254.470.461.255.472.462.755.068.660.953.266.859.0
EATA [38]55.286.167.158.175.665.658.877.266.758.670.764.057.777.465.9
SAR [39]57.588.668.959.265.261.960.573.566.360.872.165.959.574.965.8
CoTTA [54]47.083.459.953.773.261.854.376.963.654.568.160.452.475.461.4
RoTTA [59]47.954.049.447.367.055.348.369.556.747.865.555.047.864.054.1
SoTTA [17]54.453.352.853.670.360.754.470.861.453.968.460.154.165.758.8
OWTTT [28]47.170.356.253.974.362.354.573.562.554.268.560.452.471.660.4
STAMP57.997.572.663.780.070.963.780.771.263.974.468.762.383.170.8
+ +$\theta_{t}$ using cosine annealing decay step size: + +$$ +\theta_ {t + 1} = \theta_ {t} - \alpha_ {t} \nabla \mathcal {L} _ {\text {s e m}} ^ {S A} (\mathcal {M}, \theta_ {t}), \text {w h e r e} \alpha_ {t} = \frac {1}{2} \alpha_ {0} (1 + \cos (\frac {t}{T} \pi)), \tag {8} +$$ + +where $\alpha_{t}$ is the step size at timestamp $t$ and $T$ is a hyper-parameter to control the step size decay. While step size decay is commonly utilized in model training [56], we leverage it during testing for a distinct purpose. In the early stages of TTA, the model should adapt swiftly due to the significant disparity between the source and target domains. However, as adaptation progresses and the model acquires comprehensive distributional information about the target domain, it should reduce parameter updates to prevent error accumulation and enhance stability. We summarize the pipeline of STAMP in Algorithm 2. + +# 4 Experiments + +# 4.1 Setup + +Datasets. Following previous TTA methods [16, 18, 54], we adopt three corruption datasets as normal datasets (i.e., in-distribution datasets). CIFAR10-C and CIFAR100-C [20, 24] stand out as the most widely used benchmarks in TTA tasks, comprising 10/100 categories with 15 types of corruptions respectively. Each type of corruption has 10,000 images. Similarly, ImageNet-C [9, 20] serves as a significant benchmark, consisting of 1,000 categories with the same 15 types + +Table 2: Results (\%) on ImageNet-C under outlier-aware TTA setting across two outlier datasets. All the results are averaged over 15 different corruptions. + +
MethodPlaces365-CTextures-C
ACCAUCH-scoreACCAUCH-score
Source18.261.626.118.254.625.8
BN Stats [36]31.167.741.131.661.240.7
Tent [51]34.951.839.53948.642.0
EATA [38]46.472.656.046.462.252.8
SAR [39]44.973.355.045.667.054.0
CoTTA [54]33.866.943.534.260.742.8
RoTTA [59]36.668.646.537.065.346.5
SoTTA [17]41.767.850.741.860.348.8
OWTTT [28]9.154.013.99.459.414.6
STAMP46.476.257.246.570.055.6
+ +of corruption. For out-of-distribution datasets, we choose Noise, SVHN [37], LSUN [57], and TinyImageNet [25] for the CIFAR10/100-C datasets, and Textures [6] and Places365 [64] for ImageNet-C. To maintain consistency in domain shift between outliers and normal datasets, we apply the same corruptions $^4$ on outlier datasets. (e.g., SVHN-C refers to the corruptions version of SVHN dataset.) Throughout all experiments, we maintain a fixed ratio of 8:2 between normal samples and outliers. Further details of the outlier datasets are provided in the supplementary material. + +Baselines. To have a comprehensive comparison of existing TTA methods, we select eight TTA algorithms as our baselines. Source refers to the source pre-trained model. BN Stats [36] adjusts batch normalization (BN) layers [22] with the statistics of test data. Tent [51] optimizes the BN affine parameters in the model by entropy minimization. CoTTA [54] calculates weighted-average and augmentation-averaged pseudo-label to distill knowledge from the teacher model to the student model, and stochastically restores parameters to prevent the model from catastrophic forgetting. EATA [38] adopts the entropy and diversity weight with elastic weight consolidation [23] for anti-forgetting. SAR [39] excludes high-entropy samples from sharpness-aware entropy minimization and reset the model when it overfits. RoTTA [59] uses robust batch normalization and category-balanced sampling with timeliness and uncertainty to deal with the correlated label problem. SoTTA [17] applies high-confidence uniform-class sampling and entropy-sharpness minimization to alleviate the impact of noisy samples. OWTTT [28] proposes prototype expansion with prototype clustering and distribution alignment. + +Evaluation Metrics. We evaluate the above baselines for both generalization and outlier detection. For generalization, we adopt classification accuracy (ACC) as a metric. For outlier detection, we consider the area under the ROC curve (AUC) [8], a threshold-free metric to measure the binary classification problem. The ROC curve depicts the relationship between the true positive rate (TPR) and the false positive rate (FPR). Models with better detection ability + +Table 3: Results (\%) on CIFAR10/100-C datasets under outlier-aware TTA setting. We divide datasets into the normal part and the outlier part. In detail, we choose 8/80 classes as known samples and other 2/20 classes as outliers for CIFAR10/100-C. All the results are averaged over 15 different corruptions. + +
MethodCIFAR10-C (8:2)CIFAR100-C (80:20)
ACCAUCH-scoreACCAUCH-score
Source60.661.560.037.160.644.2
BN Stats [36]79.466.972.655.566.260.2
Tent [51]80.165.972.360.465.662.8
EATA [38]81.567.974.161.367.464.2
SAR [39]82.366.073.263.469.166.1
CoTTA [54]81.865.472.557.166.361.3
RoTTA [59]79.462.469.850.663.756.2
SoTTA [17]82.762.771.361.468.264.6
OWTTT [28]65.963.564.456.266.060.6
STAMP85.068.776.066.070.768.2
+ +will have a higher AUC score. To consider both generalization and detection performance simultaneously, we also report H-score: + +$$ +H - s c o r e = \frac {2 \cdot A C C \cdot A U C}{A C C + A U C} \tag {9} +$$ + +Implementation Details. We adopt the corruption with severity 5 in all experiments. We consider ResNet-18 [19] as backbones for CIFAR10-C, CIFAR100-C and ResNet-50 [19] for ImageNet-C. For CIFAR10-C and CIFAR100-C, we train the source model using standard cross-entropy loss with the initial learning rate of 0.1 and multi-step learning rate scheduling. For ImageNet-C, we use the released checkpoint from TorchVision library [35]. For a fair comparison, we set the batch size for all methods as 64 and the capacity of our memory bank is 64. Concerning the hyperparameters, we search the hyperparameters on a validation set and set the learning rate to $0.1 / 0.05 / 0.01$ , $\mathcal{H}_{thr}$ to $0.25 / 0.9 / 0.8 \times \ln |\mathcal{C}_S|$ , $T$ to $150 / 150 / 750$ for CIFAR10-C/CIFAR100-C/ImageNet-C. Moreover, we set the other hyperparameter the same in experiments with $\beta = 0.1$ , $\rho = 0.05$ , $K = 16$ . For parameters' updating, we select the affine parameters in the BN layer for optimization and freeze the top layer's parameter following [39]. More details of hyper-parameters searching and setting of baselines and STAMP can be found in the supplementary material. + +# 4.2 Results + +Results under outlier-aware TTA setting. We first evaluate TTA methods on the datasets with outliers. The results for CIFAR datasets are illustrated in Table 1. We observe that a simple baseline method, Tent [51], has a competitive generalization performance on distribution shift (achieving an accuracy of $80.1\%$ on CIFAR10-C). However, it exhibits substantial degradation in the AUC score compared to Source, for instance, dropping from $70.4\%$ to $48.7\%$ on CIFAR10-C with Noise outlier dataset. This decline could be attributed to Tent's indiscriminate minimization of entropy across all samples, leading to a gradual loss of its + +Table 4: Classification accuracies (%) on CIFAR10-C and CIFAR100-C under closed-set TTA setting. + +
MethodGaussianshotinputdefocusglassrotationzonationshowfrostlogbrightnesscontrastelasticpixelatejpegAvg.
CIFAR10-CSource28.735.224.257.349.066.464.876.762.973.490.331.478.846.574.757.3
BN Stats [36]70.272.063.687.666.585.986.982.180.584.190.885.477.378.674.379.0
Tent [51]76.178.470.587.870.186.887.584.882.085.191.186.679.482.778.581.8
CoTTA [54]76.177.673.987.971.986.987.582.882.085.290.885.779.380.478.381.8
EATA [38]76.577.971.789.170.487.489.085.284.286.091.588.479.983.978.582.6
SAR [39]70.772.166.887.668.285.986.982.180.584.190.885.477.378.674.379.4
RoTTA [59]69.671.062.987.567.085.986.982.479.684.791.273.578.377.974.978.2
SoTTA [17]75.879.271.589.470.687.889.085.684.087.292.487.479.984.579.082.9
OWTTT [28]70.172.063.086.966.585.786.782.780.984.591.283.077.976.874.978.9
STAMP80.982.977.287.674.986.687.785.985.988.190.487.280.386.782.684.3
CIFAR100-CSource12.414.57.236.044.745.145.249.541.636.963.313.257.523.846.635.8
BN Stats [36]41.141.338.963.151.560.863.851.253.553.264.459.058.458.047.653.7
Tent [51]52.352.147.766.956.164.365.358.358.760.067.862.161.863.054.559.4
CoTTA [54]47.047.845.159.654.059.561.353.355.052.962.650.358.161.853.154.8
EATA [38]50.753.548.167.055.664.867.059.159.260.467.763.961.863.354.659.8
SAR [39]55.155.051.268.458.266.067.460.360.861.969.865.563.666.256.861.7
RoTTA [59]35.936.633.860.647.157.760.848.042.250.859.232.153.852.344.447.7
SoTTA [17]51.653.847.466.956.965.368.158.960.160.169.463.162.362.854.860.1
OWTTT [28]41.642.838.863.452.661.664.853.354.854.565.858.460.157.649.354.6
STAMP57.258.552.869.961.468.170.163.363.964.872.269.966.569.259.064.4
+ +ability to discern outliers. In addition, our experimental findings demonstrate that with judicious parameter selection, TTA methods can enhance not only the generalization capability against distribution shift but also the detection ability for outliers. Compared to existing TTA methods, STAMP achieves the most favorable performance on both classification and outlier detection. On CIFAR10-C, STAMP achieves the highest ACC, and $+4.1\%$ huge improvements in AUC compared to the previous methods. Regarding CIFAR100-C, STAMP exhibits $+2.8\%$ and $+5.7\%$ improvements compared to the most effective TTA methods on ACC and AUC, respectively. We also report the results on the ImageNet-C in Table 2. STAMP still achieves the highest performance. Specifically, we outperform existing TTA methods with $+1.2\%$ , $+1.6\%$ on H-score with Places365-C and Textures-C outlier datasets, respectively. + +Additionally, we evaluate the TTA algorithm under a more challenging scenario involving homogeneous outlier datasets in Table 3, which makes it more difficult to identify outliers. STAMP also exhibits the most promising performance on CIFAR10-C and CIFAR100-C under this scenario. + +Concerning the above results, we achieve optimal results in most cases, affirming the robustness and effectiveness of STAMP. + +Results under closed-set TTA setting. Additionally, we assess our method in a scenario without outliers as most TTA works do. We present the classification accuracy on the clean dataset in Table 4. Remarkably, our method attains the highest accuracy on both CIFAR10-C and CIFAR100-C datasets, achieving an average improvement of $+2.9\%$ over the second-best method (SoTTA). This shows that STAMP still has strong applicability even in the traditional situation without outliers. + +Table 5: Ablation studies on reliable class-balanced memory (RBM), self-weighted entropy minimization (S.W.), sharpness-aware minimization (S.A.), and decay step size (D.S.). Experiments are conducted on CIFAR100-C with the outlier dataset SVHN-C. The first row in the table reports the results of entropy minimization. + +
S.A.D.S.RBMS.W.ACCAUCH-score
----54.470.461.2
-✓--58.965.361.8
✓---56.472.863.5
✓✓--59.270.864.4
✓✓-✓59.377.767.1
✓✓✓-63.075.068.4
✓✓✓✓63.780.070.9
+ +To summarize, all the above results prove that our method achieves the best results under both outlier-aware TTA and conventional closed-set settings. + +# 4.3 Ablation Study + +Effect of each component. We delve into the role of each component within our method as follows. First, we provide the results of different weight strategies of entropy minimization in Figure 3. We compare our adaptive weight with static weight and EATA weight [38]. EATA calculates the weight by $\exp[\mathcal{H}_{thr} - \mathcal{H}(p(x))]$ and static weight calculates weight in the same way of Equation (6). Different from adaptive weight, both EATA and static weight don't involve any gradient back-propagation. As depicted in Figure 3, adaptive weight shows the best performance among the strategies. + +Furthermore, we conduct an ablation study on each component of our method, as summarized in Table 5. We observe that the effect of combining test-time step size decay and SAM is better than using either of them alone. Step size decay can + +![](images/631e1fc72f95b0203dde42cd3276886b19169289d5e6f67a9cdb90b679ca47d6.jpg) +Fig. 3: Results for the different weight strategies on dataset CIFAR100-C with SVHN-C outlier dataset. Static weight represents that the $w_{i}$ in Equation (6) without gradient back-propagation. EATA weight represents the weight for reliable samples used in EATA [38]. + +Table 6: H-score (\%) with the different ratios of outliers to total samples on CIFAR100-C with different outlier datasets. + +
SVHN-CLSUN-CTinyImageNet-C
5%68.470.268.3
10%70.170.968.8
20%70.971.268.7
33%70.370.168.1
50%69.368.867.4
+ +improve the ACC but result in AUC decline. However, combining SAM and step size decay mitigates the decline in AUC. In addition, after the introduction of self-weighted entropy minimization, the AUC of the model further increased to $77.7\%$ . Notably, by removing reliable class-balanced memory from our method, the ACC decreases from $63.7\%$ to $59.3\%$ . It reveals that reliable class-balanced memory significantly contributes to enhancing the generalization ability to distribution shift. Moreover, by removing the self-weighted entropy minimization in STAMP, its AUC drops from $80.0\%$ to $75.0\%$ . Self-weighted entropy minimization plays an important role in outlier detection. Besides decreasing the sample entropy, self-weighted entropy minimization incorporates entropy into the optimization weight, allowing outliers that deviate from the majority of samples to receive lower weights, consequently increasing their entropy. This mechanism enhances outlier detection performance. + +Effect of outlier proportion. In the realistic scenario, the proportion of outliers can vary considerably. We investigate how this variability affects the performance of our method. Specifically, we manipulate the proportion of outliers, ranging from $5\%$ to $50\%$ , and present the corresponding H-scores in Table 6. The results show that the fluctuation of our method is within $3\%$ . The findings suggest that STAMP exhibits robustness across different scenarios with varying proportions of outliers, as it does not experience significant degradation. + +# 5 Conclusion + +In this study, we research the challenge of handling outliers in test-time adaptation. To tackle this issue effectively, we introduce a scenario called outlier-aware test-time adaptation, which refers to conducting both recognition and outlier rejection simultaneously. Additionally, we propose STAMP, using augmentation-averaged estimation and a filtering mechanism to store reliable samples in memory. Moreover, we dynamically discard samples with the highest class frequency from memory to maintain a class-balanced prediction. Then we replay the samples stored in the memory with self-weighted entropy minimization. Furthermore, we leverage sharpness-aware minimization and learning rate decay to bolster the stability of our algorithm. Through extensive experiments across multiple benchmarks, we demonstrate that our method surpasses existing TTA methods, underscoring its effectiveness and versatility in open-world applications. + +# Acknowledgements + +We thank Yuhe Ding and Zhengbo Wang for their valuable feedback on our work. This work was funded by the Beijing Municipal Science and Technology Project (No. Z231100010323005), the Beijing Nova Program (No. Z211100002121108), the Young Elite Scientists Sponsorship Program by CAST (2023QNRC001), and the National Natural Science Foundation of China under (No. 62276256). + +# References + +1. An, S., Park, S., Kim, G., Baek, J., Lee, B., Kim, S.: Context enhanced transformer for single image object detection. arXiv preprint arXiv:2312.14492 (2023) 2 +2. Boudiaf, M., Mueller, R., Ben Ayed, I., Bertinetto, L.: Parameter-free online test-time adaptation. In: Proc. CVPR (2022) 4 +3. Cao, H., Xu, Y., Yang, J., Yin, P., Yuan, S., Xie, L.: Multi-modal continual test-time adaptation for 3d semantic segmentation. In: Proc. ICCV (2023) 2 +4. Chakrabarty, G., Sreenivas, M., Biswas, S.: Santa: Source anchoring network and target alignment for continual test time adaptation. Transactions on Machine Learning Research (2023) 4 +5. Chen, D., Wang, D., Darrell, T., Ebrahimi, S.: Contrastive test-time adaptation. In: Proc. CVPR (2022) 2, 4 +6. Cimpoi, M., Maji, S., Kokkinos, I., Mohamed, S., Vedaldi, A.: Describing textures in the wild. In: Proc. CVPR (2014) 3, 10 +7. Croce, F., Andriushchenko, M., Sehwag, V., Debenedetti, E., Flammarion, N., Chiang, M., Mittal, P., Hein, M.: Robustbench: a standardized adversarial robustness benchmark. In: Proc. NeurIPS (2021) 1 +8. Davis, J., Goadrich, M.: The relationship between precision-recall and roc curves. In: Proc. ICML (2006) 10 +9. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: Proc. CVPR (2009) 3, 9 +0. DeVries, T., Taylor, G.W.: Learning confidence for out-of-distribution detection in neural networks. arXiv preprint arXiv:1802.04865 (2018) 3, 4 +1. Ding, Y., Liang, J., Jiang, B., Zheng, A., He, R.: Maps: A noise-robust progressive learning approach for source-free domain adaptive keypoint detection. IEEE Transactions on Circuits and Systems for Video Technology (2023) 3 +2. Ding, Y., Sheng, L., Liang, J., Zheng, A., He, R.: Proxymix: Proxy-based mixup training with label refinery for source-free domain adaptation. Neural Networks 167, 92-103 (2023) 3 +3. Farahani, A., Voghoei, S., Rasheed, K., Arabnia, H.R.: A brief review of domain adaptation. Advances in Data Science and Information Engineering: Proceedings from ICDATA 2020 and IKE 2020 pp. 877-894 (2021) 1 +4. Foret, P., Kleiner, A., Mobahi, H., Neyshabur, B.: Sharpness-aware minimization for efficiently improving generalization. In: Proc. ICLR (2021) 3, 8 +5. Gebru, T., Hoffman, J., Fei-Fei, L.: Fine-grained recognition in the wild: A multitask domain adaptation approach. In: Proc. ICCV (2017) 3 +6. Gong, T., Jeong, J., Kim, T., Kim, Y., Shin, J., Lee, S.J.: Note: Robust continual test-time adaptation against temporal correlation. In: Proc. NeurIPS (2022) 4, 9 +7. Gong, T., Kim, Y., Lee, T., Chottananurak, S., Lee, S.J.: Sotta: Robust test-time adaptation on noisy data streams. In: Proc. NeurIPS (2023) 4, 5, 8, 9, 10, 11, 12 + +18. Goyal, S., Sun, M., Raghunathan, A., Kolter, J.Z.: Test time adaptation via conjugate pseudo-labels. In: Proc. NeurIPS (2022) 9 +19. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proc. CVPR (2016) 11 +20. Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. In: Proc. ICLR (2018) 3, 9 +21. Hendrycks, D., Gimpel, K.: A baseline for detecting misclassified and out-of-distribution examples in neural networks. In: Proc. ICLR (2016) 4 +22. Ioffe, S., Szegedy, C.: Batch normalization: Accelerating deep network training by reducing internal covariate shift. In: Proc. ICML (2015) 10 +23. Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A.A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., et al.: Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences 114(13), 3521-3526 (2017) 10 +24. Krizhevsky, A.: Learning multiple layers of features from tiny images. Master's thesis, University of Tront (2009) 9 +25. Le, Y., Yang, X.: Tiny imagenet visual recognition challenge. CS 231N 7(7), 3 (2015) 3, 10 +26. Lee, J., Das, D., Choo, J., Choi, S.: Towards open-set test-time adaptation utilizing the wisdom of crowds in entropy minimization. In: Proc. CVPR (2023) 4, 8 +27. Lee, K., Lee, K., Lee, H., Shin, J.: A simple unified framework for detecting out-of-distribution samples and adversarial attacks. In: Proc. NeurIPS (2018) 4 +28. Li, Y., Xu, X., Su, Y., Jia, K.: On the robustness of open-world test-time training: Self-training with dynamic prototype expansion. In: Proc. CVPR (2023) 4, 9, 10, 11, 12 +29. Liang, J., He, R., Tan, T.: A comprehensive survey on test-time adaptation under distribution shifts. International Journal of Computer Vision (2024) 1 +30. Liang, J., Hu, D., Feng, J.: Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In: Proc. ICML (2020) 1 +31. Liang, J., Sheng, L., Wang, Z., He, R., Tan, T.: Realistic unsupervised clip finetuning with universal entropy optimization. In: Proc. ICML (2024) 4 +32. Liang, S., Li, Y., Srikant, R.: Enhancing the reliability of out-of-distribution image detection in neural networks. In: Proc. ICLR (2018) 4 +33. Lin, Z., Roy, S.D., Li, Y.: Mood: Multi-level out-of-distribution detection. In: Proc. CVPR (2021) 4 +34. Liu, W., Wang, X., Owens, J., Li, Y.: Energy-based out-of-distribution detection. In: Proc. NeurIPS (2020) 4 +35. Maintainers, T., Contributors: Torchvision: Pytorch's computer vision library (2016) 11 +36. Nado, Z., Padhy, S., Sculley, D., D'Amour, A., Lakshminarayanan, B., Snoek, J.: Evaluating prediction-time batch normalization for robustness under covariate shift. In: Proc. ICML Workshops (2020) 3, 9, 10, 11, 12 +37. Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B., Ng, A.Y.: Reading digits in natural images with unsupervised feature learning (2011) 3, 10 +38. Niu, S., Wu, J., Zhang, Y., Chen, Y., Zheng, S., Zhao, P., Tan, M.: Efficient test-time model adaptation without forgetting. In: Proc. ICML (2022) 2, 4, 5, 6, 7, 9, 10, 11, 12, 13 +39. Niu, S., Wu, J., Zhang, Y., Wen, Z., Chen, Y., Zhao, P., Tan, M.: Towards stable test-time adaptation in dynamic wild world. In: Proc. ICLR (2022) 4, 7, 8, 9, 10, 11, 12 + +40. Nixon, J., Dusenberry, M.W., Zhang, L., Jerfel, G., Tran, D.: Measuring calibration in deep learning. In: Proc. CVPR Workshops (2019) 5 +41. Park, S., Yang, S., Choo, J., Yun, S.: Label shift adapter for test-time adaptation under covariate and label shifts. In: Proc. CVPR (2023) 4 +42. Press, O., Schneider, S., Kümmerer, M., Bethge, M.: Rdumb: A simple approach that questions our progress in continual test-time adaptation. In: Proc. NeurIPS (2023) 8 +43. Schneider, S., Rusak, E., Eck, L., Bringmann, O., Brendel, W., Bethge, M.: Improving robustness against common corruptions by covariate shift adaptation. In: Proc. NeurIPS (2020) 3 +44. Shanmugam, D., Blalock, D., Balakrishnan, G., Guttag, J.: Better aggregation in test-time augmentation. In: Proc. CVPR (2021) 5 +45. Sheng, L., Liang, J., He, R., Wang, Z., Tan, T.: Adaptguard: Defending against universal attacks for model adaptation. In: Proc. ICCV (2023) 3 +46. Tan, M., Le, Q.: Efficientnet: Rethinking model scaling for convolutional neural networks. In: Proc. ICML (2019) 1 +47. Tong, P., Su, W., Li, H., Ding, J., Haoxiang, Z., Chen, S.X.: Distribution free domain generalization. In: Proc. ICML (2023) 3 +48. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: Proc. NeurIPS (2017) 1 +49. Volpi, R., De Jorge, P., Larlus, D., Csurka, G.: On the road to online adaptation for semantic image segmentation. In: Proc. CVPR (2022) 2 +50. Vyas, A., Jammalamadaka, N., Zhu, X., Das, D., Kaul, B., Willke, T.L.: Out-of-distribution detection using an ensemble of self supervised leave-out classifiers. In: Proc. ECCV (2018) 4 +51. Wang, D., Shelhamer, E., Liu, S., Olshausen, B., Darrell, T.: Tent: Fully test-time adaptation by entropy minimization. In: Proc. ICML (2020) 2, 4, 7, 9, 10, 11, 12 +52. Wang, J., Lan, C., Liu, C., Ouyang, Y., Qin, T., Lu, W., Chen, Y., Zeng, W., Yu, P.S.: Generalizing to unseen domains: A survey on domain generalization. IEEE Transactions on Knowledge and Data Engineering 35(8), 8052-8072 (2023) 3 +53. Wang, M., Deng, W.: Deep visual domain adaptation: A survey. Neurocomputing 312, 135-153 (2018) 3 +54. Wang, Q., Fink, O., Van Gool, L., Dai, D.: Continual test-time domain adaptation. In: Proc. CVPR (2022) 4, 9, 10, 11, 12 +55. Wang, Y., Li, B., Che, T., Zhou, K., Liu, Z., Li, D.: Energy-based open-world uncertainty modeling for confidence calibration. In: Proc. CVPR (2021) 4 +56. You, K., Long, M., Wang, J., Jordan, M.I.: How does learning rate decay help modern neural networks? arXiv preprint arXiv:1908.01878 (2019) 9 +57. Yu, F., Seff, A., Zhang, Y., Song, S., Funkhouser, T., Xiao, J.: Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365 (2015) 3, 10 +58. Yu, Y., Sheng, L., He, R., Liang, J.: Benchmarking test-time adaptation against distribution shifts in image classification. arXiv preprint arXiv:2307.03133 (2023) 1 +59. Yuan, L., Xie, B., Li, S.: Robust test-time adaptation in dynamic scenarios. In: Proc. CVPR (2023) 2, 3, 9, 10, 11, 12 +60. Yurtsever, E., Lambert, J., Carballo, A., Takeda, K.: A survey of autonomous driving: Common practices and emerging technologies. IEEE Access 8, 58443-58469 (2020) 1 +61. Zhang, M., Levine, S., Finn, C.: Memo: Test time robustness via adaptation and augmentation. In: Proc. NeurIPS (2022) 4, 5 + +62. Zhang, Y., Wang, X., Jin, K., Yuan, K., Zhang, Z., Wang, L., Jin, R., Tan, T.: Adanpc: Exploring non-parametric classifier for test-time adaptation. In: Proc. ICML (2023) 2 +63. Zhao, H., Liu, Y., Alahi, A., Lin, T.: On pitfalls of test-time adaptation. In: Proc. ICML (2023) 4 +64. Zhou, B., Lapedriza, A., Khosla, A., Oliva, A., Torralba, A.: Places: A 10 million image database for scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence 40(6), 1452-1464 (2017) 3, 10 +65. Zhou, Z., Guo, L.Z., Jia, L.H., Zhang, D., Li, Y.F.: Ods: test-time adaptation in the presence of open-world data shift. In: Proc. ICML (2023) 4 \ No newline at end of file diff --git a/2024/STAMP_ Outlier-Aware Test-Time Adaptation with Stable Memory Replay/images.zip b/2024/STAMP_ Outlier-Aware Test-Time Adaptation with Stable Memory Replay/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0d5eb05e25edbee7e48f4ce610168e4bdd1eec55 --- /dev/null +++ b/2024/STAMP_ Outlier-Aware Test-Time Adaptation with Stable Memory Replay/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:10ad31cfdb0f95e6dac27e2e1d11d41bb759b24cee2079779829275b8f7b46cd +size 508203 diff --git a/2024/STAMP_ Outlier-Aware Test-Time Adaptation with Stable Memory Replay/layout.json b/2024/STAMP_ Outlier-Aware Test-Time Adaptation with Stable Memory Replay/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6fa45fa7f5107504ff72a68fb0429dfdae06a020 --- /dev/null +++ b/2024/STAMP_ Outlier-Aware Test-Time Adaptation with Stable Memory Replay/layout.json @@ -0,0 +1,9762 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 168, + 115, + 200, + 140 + ], + "blocks": [ + { + "bbox": [ + 168, + 115, + 200, + 140 + ], + "lines": [ + { + "bbox": [ + 168, + 115, + 200, + 140 + ], + "spans": [ + { + "bbox": [ + 168, + 115, + 200, + 140 + ], + "type": "image", + "image_path": "2661f00c92b9e29c0adf43c07a615933123144781e6f9d47b5962a503944823d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 202, + 122, + 445, + 158 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 122, + 445, + 158 + ], + "spans": [ + { + "bbox": [ + 202, + 122, + 445, + 158 + ], + "type": "text", + "content": "STAMP: Outlier-Aware Test-Time Annotation with Stable Memory Replay" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 156, + 177, + 457, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 177, + 457, + 190 + ], + "spans": [ + { + "bbox": [ + 156, + 177, + 457, + 190 + ], + "type": "text", + "content": "Yongcan Yu" + }, + { + "bbox": [ + 156, + 177, + 457, + 190 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 156, + 177, + 457, + 190 + ], + "type": "text", + "content": ", Lijun Sheng" + }, + { + "bbox": [ + 156, + 177, + 457, + 190 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 156, + 177, + 457, + 190 + ], + "type": "text", + "content": ", Ran He" + }, + { + "bbox": [ + 156, + 177, + 457, + 190 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 156, + 177, + 457, + 190 + ], + "type": "text", + "content": ", and Jian Liang" + }, + { + "bbox": [ + 156, + 177, + 457, + 190 + ], + "type": "inline_equation", + "content": "^{1,2\\star}" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 149, + 199, + 462, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 199, + 462, + 210 + ], + "spans": [ + { + "bbox": [ + 149, + 199, + 462, + 210 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 149, + 199, + 462, + 210 + ], + "type": "text", + "content": " NLPR & MAIS, Institute of Automation, Chinese Academy of Sciences" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 149, + 210, + 463, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 210, + 463, + 221 + ], + "spans": [ + { + "bbox": [ + 149, + 210, + 463, + 221 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 149, + 210, + 463, + 221 + ], + "type": "text", + "content": " School of Artificial Intelligence, University of Chinese Academy of Sciences" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 208, + 221, + 406, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 221, + 406, + 233 + ], + "spans": [ + { + "bbox": [ + 208, + 221, + 406, + 233 + ], + "type": "text", + "content": "3 University of Science and Technology of China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 217, + 234, + 397, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 234, + 397, + 244 + ], + "spans": [ + { + "bbox": [ + 217, + 234, + 397, + 244 + ], + "type": "text", + "content": "{yuyongcan0223, liangjian92}@gmail.com" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 159, + 274, + 455, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 274, + 455, + 472 + ], + "spans": [ + { + "bbox": [ + 159, + 274, + 455, + 472 + ], + "type": "text", + "content": "Abstract. Test-time adaptation (TTA) aims to address the distribution shift between the training and test data with only unlabeled data at test time. Existing TTA methods often focus on improving recognition performance specifically for test data associated with classes in the training set. However, during the open-world inference process, there are inevitably test data instances from unknown classes, commonly referred to as outliers. This paper pays attention to the problem that conducts both sample recognition and outlier rejection during inference while outliers exist. To address this problem, we propose a new approach called STAble Memory rePlay (STAMP), which performs optimization over a stable memory bank instead of the risky mini-batch. In particular, the memory bank is dynamically updated by selecting low-entropy and label-consistent samples in a class-balanced manner. In addition, we develop a self-weighted entropy minimization strategy that assigns higher weight to low-entropy samples. Extensive results demonstrate that STAMP outperforms existing TTA methods in terms of both recognition and outlier detection performance. The code is released at https://github.com/yuyongcan/STAMP." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 159, + 483, + 453, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 483, + 453, + 506 + ], + "spans": [ + { + "bbox": [ + 159, + 483, + 453, + 506 + ], + "type": "text", + "content": "Keywords: online test-time adaptation " + }, + { + "bbox": [ + 159, + 483, + 453, + 506 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 483, + 453, + 506 + ], + "type": "text", + "content": " outlier detection " + }, + { + "bbox": [ + 159, + 483, + 453, + 506 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 483, + 453, + 506 + ], + "type": "text", + "content": " memory bank " + }, + { + "bbox": [ + 159, + 483, + 453, + 506 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 483, + 453, + 506 + ], + "type": "text", + "content": " self-weighted entropy minimization" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 525, + 230, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 230, + 538 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 230, + 538 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 550, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 550, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 550, + 482, + 647 + ], + "type": "text", + "content": "Deep neural networks (DNNs) [46, 48] have achieved remarkable performance across various tasks. Nevertheless, the impressive performance is based on the assumption that training data and test data come from the same distribution. However, in real-world scenarios, the presence of distribution shifts is ubiquitous [13, 30], leading to a degradation in the prediction performance [7]. Taking autonomous driving [60] as an example, changing weather conditions causes systems trained on laboratory data to make incorrect decisions. To mitigate the impact of distribution shift, test-time adaptation (TTA) [29, 58] has emerged as" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 236, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 236, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 236, + 665 + ], + "type": "text", + "content": "* Corresponding author." + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 481, + 232 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 481, + 232 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 481, + 232 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 481, + 232 + ], + "type": "image", + "image_path": "afdec46572993664b26a61c6f1c8adaaed02f7115bbb7ecdd56e1c0e1a343085.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 239, + 482, + 298 + ], + "lines": [ + { + "bbox": [ + 130, + 239, + 482, + 298 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 482, + 298 + ], + "type": "text", + "content": "Fig. 1: The illustration of outlier-aware test-time adaptation. The test data stream consists of both normal samples and outliers. When the test data arrives, the algorithm adapts the model and provides two outputs, a prediction and an OOD score. OOD score indicates the likelihood that a sample is an outlier. The detector in the deployment scenario chooses to keep the prediction or reject samples according to the OOD score." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 339, + 479, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 339, + 479, + 375 + ], + "spans": [ + { + "bbox": [ + 130, + 339, + 479, + 375 + ], + "type": "text", + "content": "a promising approach, adapting models with unlabeled samples during inference. Nowadays, TTA has achieved remarkable results in various tasks, such as image classification [51, 62], semantic segmentation [3, 49], object detection [1]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 376, + 482, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 496 + ], + "type": "text", + "content": "Previous TTA works [5, 38, 59] always assume that test data and training data share the same class space. Nevertheless, in realistic open-world scenarios, test data streams will inevitably contain samples from new semantic classes, also known as outliers. The introduction of outliers brings two additional risks to adaptation tasks: (i) Classifying outliers into known categories poses safety risks, especially in contexts like autonomous driving, where systems should request human intervention to handle unknown objects to avoid traffic accidents. (ii) Self-supervised loss (e.g., entropy minimization, pseudo-labeling) calculated with outliers may misguide the optimization process, decreasing the model's recognition ability on known classes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 497, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 482, + 604 + ], + "type": "text", + "content": "To better tackle the above risks, we pay attention to a more realistic TTA scenario termed outlier-aware test-time adaptation. Outlier-aware test-time adaptation aims to conduct recognition for normal samples and reject to make predictions for outliers. The overview of outlier-aware test-time adaptation is illustrated in Figure 1. When test data mixed with outliers arrives, the algorithm adapts the model and generates the prediction and an out-of-distribution (OOD) score for each sample. Then we decide to keep the prediction or reject the sample based on its OOD score. The sample with a higher OOD score is more likely to be considered an outlier." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": "To address the outlier-aware test-time adaptation, we additionally design a novel algorithm called STAble Memory rePlay (STAMP). STAMP consists of three components: reliable class-balanced memory, self-weighted entropy minimization, and stable optimization strategy. Reliable class-balanced memory stores reliable samples for optimization by filtering samples based on their pre" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 203, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 203, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 203, + 101 + ], + "type": "text", + "content": "Yu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "diction's consistency and entropy. To maintain the class-balanced prediction of the model, it dynamically discards samples with the class that participates most frequently in optimization when the memory is full. Furthermore, we adopt self-weighted entropy minimization to assign greater weights to low-entropy samples in memory and adjust the weights for each sample adaptively. To ensure the stability of the optimization process, we dynamically adjust the optimization step size to quickly adapt to the target domain at the early stage and mitigate error accumulation later. Moreover, we apply the sharpness-aware minimization technique [14] to further reduce the impact of noisy gradients." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 224, + 482, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 331 + ], + "type": "text", + "content": "To demonstrate the effectiveness of STAMP, we compare it with eight state-of-the-art TTA algorithms. We select three commonly used datasets in TTA, including CIFAR10-C, CIFAR100-C [10, 20], and ImageNet-C [9, 20], alongside some outlier datasets (i.e., Noise, SVHN [37], LSUN [57], TinyImageNet [25], Textures [6], Places365 [64]). Experimental results show that our method achieves superior performance on recognition and outlier detection under the outlier-aware TTA setting. The experiment on conventional protocol indicates that STAMP also performs well in the scenario without outliers. In summary, we list our contributions as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 339, + 480, + 456 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 339, + 479, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 339, + 479, + 362 + ], + "spans": [ + { + "bbox": [ + 138, + 339, + 479, + 362 + ], + "type": "text", + "content": "- We introduce outlier-aware test-time adaptation, a TTA scenario considering both normal sample recognition and outlier rejection during inference." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 363, + 480, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 363, + 480, + 397 + ], + "spans": [ + { + "bbox": [ + 138, + 363, + 480, + 397 + ], + "type": "text", + "content": "- We propose a memory that dynamically stores reliable samples for replay. To prevent noisy samples from entering the memory, we propose a filtering strategy based on the consistency and entropy of the model's prediction." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 399, + 479, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 399, + 479, + 422 + ], + "spans": [ + { + "bbox": [ + 138, + 399, + 479, + 422 + ], + "type": "text", + "content": "- We design self-weighted entropy minimization, which guides the model to adaptively learn to assign higher weights to valuable samples." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 422, + 479, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 422, + 479, + 456 + ], + "spans": [ + { + "bbox": [ + 138, + 422, + 479, + 456 + ], + "type": "text", + "content": "- We propose a proper metric that considers both generalization and outlier detection performance. Extensive experiments compared with eight TTA algorithms on several benchmarks demonstrate the effectiveness of our method." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 476, + 237, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 476, + 237, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 476, + 237, + 489 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 502, + 272, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 272, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 272, + 514 + ], + "type": "text", + "content": "2.1 Test-Time Adaptation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "TTA intends to adapt the model pre-trained on the source domain to the target domain while the distribution shift exists. Compared to domain generalization [47, 52] and domain adaptation [15, 53], TTA needs to access neither the source data nor the target labels, making it has a wide range of applications [11, 12, 45]. TTA methods can be mainly divided into two solutions: batch normalization (BN) calibration and self-training. In BN calibration, Nado et al. [36] utilize the BN statistics of the test batch " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "(\\mu_t, \\sigma_t)" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " to conduct normalization in the BN layer, while Schneider [43] propose a more comprehensive approach considering statistics from both the source domain " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "(\\mu_s, \\sigma_s)" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " and the target domain to calibrate the domain shifts. To conquer the correlation sampling problem, Yuan et al. [59] adopt robust batch normalization, which updates the statistic " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "(\\mu_t, \\sigma_t)" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " in a moving average manner with the test data stream." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "type": "text", + "content": "Outlier-Aware Test-Time Adaptation with Stable Memory Replay" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "text", + "content": "Currently, the primary paradigm in TTA is self-training [2, 61], where Wang et al. [51] adopt entropy minimization as a self-supervised loss for optimization and Chen et al. [5] introduce contrastive loss to TTA. However, unsupervised loss may contain noise information. Niu et al. [38] first propose filtering out high-entropy samples, a mechanism adopted by subsequent works [17, 39]. Subsequently, more attention has been paid to addressing more realistic adaptation scenarios such as continually changing target domains [4, 54], label shift [41, 65], and temporally correlated shift [16, 63]. In our study, we focus on a more realistic scenario where test data contains samples with new semantic classes. Some existing works [17, 26, 28] discuss the TTA algorithm in scenarios containing outliers. However, Lee et al. [26] and Gong et al. [17] solely focus on the generalization ability to domain shift, neglecting the issue of rejecting the outliers. A recent work [28] discusses outlier detection performance based on TTA tasks, but it only provides binary predictions to indicate whether the test data is outlier. In contrast, outlier-aware test-time adaptation is more flexible, providing an OOD score, and different systems can conduct sample rejection based on security requirements and the score." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 343, + 311, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 311, + 355 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 311, + 355 + ], + "type": "text", + "content": "2.2 Out-of-Distribution Detection" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 369, + 482, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 482, + 574 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 482, + 574 + ], + "type": "text", + "content": "OOD detection aims to identify instances that significantly deviate from the distribution of the training data [31]. Their primary focus is usually on detecting samples whose labels were not seen during the training phase. This area has been extensively researched, with one common approach being the utilization of confidence scores such as max softmax probability [21], energy score [33, 34], and others [27] derived from the model's output to detect OOD samples. Further improvements have been made by Liang et al. [32] through techniques like temperature scaling and input preprocessing to enhance the effectiveness of confidence scores. Another avenue for OOD sample detection lies in training-based methods [10, 50, 55]. These techniques involve training a branch using both indistribution data and OOD data to enable the model to distinguish OOD data during training. However, these methods alter the training process of the source model, which can not be directly applied to adaptation tasks based on a well-trained model. Moreover, OOD detection traditionally assumes that both the training and test data originate from the same domain, a condition not met in TTA scenarios. Under the realistic scenario, there may be a domain shift between test data and training data, which brings greater challenges to detection." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 597, + 233, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 597, + 233, + 612 + ], + "spans": [ + { + "bbox": [ + 132, + 597, + 233, + 612 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "In this section, we first introduce the outlier-aware test-time adaptation problem in Section 3.1. Then we describe our proposed method, including how to build a memory bank and conduct the optimization in Section 3.2." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "type": "text", + "content": "Yu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 348, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 348, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 348, + 128 + ], + "type": "text", + "content": "3.1 Outlier-Aware Test-Time Adaptation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 134, + 479, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 134, + 479, + 193 + ], + "spans": [ + { + "bbox": [ + 130, + 134, + 479, + 193 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 130, + 134, + 479, + 193 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_S" + }, + { + "bbox": [ + 130, + 134, + 479, + 193 + ], + "type": "text", + "content": " be the source domain dataset with label space " + }, + { + "bbox": [ + 130, + 134, + 479, + 193 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_S" + }, + { + "bbox": [ + 130, + 134, + 479, + 193 + ], + "type": "text", + "content": " follows a distribution " + }, + { + "bbox": [ + 130, + 134, + 479, + 193 + ], + "type": "inline_equation", + "content": "P_S(x,y)" + }, + { + "bbox": [ + 130, + 134, + 479, + 193 + ], + "type": "text", + "content": ". Similarly, let " + }, + { + "bbox": [ + 130, + 134, + 479, + 193 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_T = \\{(x_i,y_i)\\}_{i=1}^{N_T}" + }, + { + "bbox": [ + 130, + 134, + 479, + 193 + ], + "type": "text", + "content": " be the target domain dataset with label space " + }, + { + "bbox": [ + 130, + 134, + 479, + 193 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_T" + }, + { + "bbox": [ + 130, + 134, + 479, + 193 + ], + "type": "text", + "content": " follows a distribution " + }, + { + "bbox": [ + 130, + 134, + 479, + 193 + ], + "type": "inline_equation", + "content": "P_T(x,y)" + }, + { + "bbox": [ + 130, + 134, + 479, + 193 + ], + "type": "text", + "content": ". Given a source model " + }, + { + "bbox": [ + 130, + 134, + 479, + 193 + ], + "type": "inline_equation", + "content": "f_{\\theta_s}(x)" + }, + { + "bbox": [ + 130, + 134, + 479, + 193 + ], + "type": "text", + "content": " well trained on " + }, + { + "bbox": [ + 130, + 134, + 479, + 193 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_S" + }, + { + "bbox": [ + 130, + 134, + 479, + 193 + ], + "type": "text", + "content": " with parameter " + }, + { + "bbox": [ + 130, + 134, + 479, + 193 + ], + "type": "inline_equation", + "content": "\\theta_s" + }, + { + "bbox": [ + 130, + 134, + 479, + 193 + ], + "type": "text", + "content": ", TTA aims to adapt the source model to the target domain while " + }, + { + "bbox": [ + 130, + 134, + 479, + 193 + ], + "type": "inline_equation", + "content": "P_S(x) \\neq P_T(x)" + }, + { + "bbox": [ + 130, + 134, + 479, + 193 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 194, + 480, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 194, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 480, + 277 + ], + "type": "text", + "content": "In traditional closed-set TTA, it is commonly assumed that the " + }, + { + "bbox": [ + 130, + 194, + 480, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_S" + }, + { + "bbox": [ + 130, + 194, + 480, + 277 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 194, + 480, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_T" + }, + { + "bbox": [ + 130, + 194, + 480, + 277 + ], + "type": "text", + "content": " share the same label space, i.e., " + }, + { + "bbox": [ + 130, + 194, + 480, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_S = \\mathcal{C}_T" + }, + { + "bbox": [ + 130, + 194, + 480, + 277 + ], + "type": "text", + "content": ". However, in outlier-aware test-time adaptation, we consider " + }, + { + "bbox": [ + 130, + 194, + 480, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_S \\subseteq \\mathcal{C}_T" + }, + { + "bbox": [ + 130, + 194, + 480, + 277 + ], + "type": "text", + "content": " while some outliers with new semantic classes will be encountered in " + }, + { + "bbox": [ + 130, + 194, + 480, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_T" + }, + { + "bbox": [ + 130, + 194, + 480, + 277 + ], + "type": "text", + "content": ". The illustration of outlier-aware test-time adaption is shown in Figure 1. In addition to adapting the model and providing the prediction, we also generate an OOD score " + }, + { + "bbox": [ + 130, + 194, + 480, + 277 + ], + "type": "inline_equation", + "content": "\\delta_i" + }, + { + "bbox": [ + 130, + 194, + 480, + 277 + ], + "type": "text", + "content": " for each sample " + }, + { + "bbox": [ + 130, + 194, + 480, + 277 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 130, + 194, + 480, + 277 + ], + "type": "text", + "content": ". Then we adopt an outlier detector " + }, + { + "bbox": [ + 130, + 194, + 480, + 277 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 130, + 194, + 480, + 277 + ], + "type": "text", + "content": " as follow:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 248, + 284, + 480, + 316 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 248, + 284, + 480, + 316 + ], + "spans": [ + { + "bbox": [ + 248, + 284, + 480, + 316 + ], + "type": "interline_equation", + "content": "g \\left(x _ {i}\\right) = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} \\delta_ {i} \\geq \\delta_ {t h r} \\\\ 0, & \\text {i f} \\delta_ {i} < \\delta_ {t h r}, \\end{array} \\right. \\tag {1}", + "image_path": "bb1fb5095815d4f2b7cb87dcb9480a479b7dc886d2f8539d188af4619e306f26.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 321, + 482, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 321, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 321, + 482, + 380 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 321, + 482, + 380 + ], + "type": "inline_equation", + "content": "\\delta_{thr}" + }, + { + "bbox": [ + 130, + 321, + 482, + 380 + ], + "type": "text", + "content": " is a threshold that can be adjusted according to the requirements of different application scenarios. The samples with " + }, + { + "bbox": [ + 130, + 321, + 482, + 380 + ], + "type": "inline_equation", + "content": "g(x_i) = 1" + }, + { + "bbox": [ + 130, + 321, + 482, + 380 + ], + "type": "text", + "content": " will be considered outliers and rejected for recognition. And we output the prediction for samples with " + }, + { + "bbox": [ + 130, + 321, + 482, + 380 + ], + "type": "inline_equation", + "content": "g(x_i) = 0" + }, + { + "bbox": [ + 130, + 321, + 482, + 380 + ], + "type": "text", + "content": ". In our work, we take the entropy of the model prediction as our OOD score." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 396, + 278, + 409 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 396, + 278, + 409 + ], + "spans": [ + { + "bbox": [ + 131, + 396, + 278, + 409 + ], + "type": "text", + "content": "3.2 Stable Memory Replay" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 414, + 479, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 479, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 479, + 498 + ], + "type": "text", + "content": "In this section, we introduce stable memory replay (STAMP). STAMP first uses a filtering mechanism to filter out unreliable samples in the coming mini-batch and stores reliable samples in memory. Then, STAMP optimizes the model using in-memory samples by applying self-weighted entropy minimization. In order to further improve the stability of optimization, STAMP uses test-time step size decay and sharpness-aware minimization to guide optimization. The illustration of STAMP is shown in Figure 2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 498, + 479, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 479, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 479, + 594 + ], + "type": "text", + "content": "Reliable Class-Balanced Memory. Since excluding unreliable samples from optimization has shown great effectiveness [17, 38], we propose reliable class-balanced memory to store reliable samples for optimization. To have a more robust estimation of the sample's reliability, we get its prediction in an augmentation-averaged way. Moreover, we prevent unreliable samples from getting into the memory bank " + }, + { + "bbox": [ + 130, + 498, + 479, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 130, + 498, + 479, + 594 + ], + "type": "text", + "content": " by employing consistency filtering and entropy filtering. To preserve the model's balanced class prior distribution, we dynamically discard samples with the highest class frequency from the memory bank." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 594, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 480, + 665 + ], + "type": "text", + "content": "To ensure robust predictions, we adopt the augmentation-averaged method, which addresses the issue of deep neural networks being overconfident with out-of-distribution samples. This overconfidence often blurs the line between in-distribution and out-of-distribution samples. However, research has demonstrated that employing test-time augmentation [44, 61] enhances model calibration [40]. Thus, when presented with a test sample " + }, + { + "bbox": [ + 130, + 594, + 480, + 665 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 594, + 480, + 665 + ], + "type": "text", + "content": ", we derive its prediction" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "type": "text", + "content": "Outlier-Aware Test-Time Adaptation with Stable Memory Replay" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 479, + 295 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 479, + 295 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 479, + 295 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 479, + 295 + ], + "type": "image", + "image_path": "4241aedecb67cdbdce38260bb146b30410a5af6c66a8456efd2282f25356f565.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 305, + 482, + 384 + ], + "lines": [ + { + "bbox": [ + 130, + 305, + 482, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 305, + 482, + 384 + ], + "type": "text", + "content": "Fig. 2: The overview of STAMP. STAMP retains reliable samples for optimizing the network by maintaining a memory bank. When test samples arrive, STAMP generates an augmentation-averaged prediction. Two filtering mechanisms are designed to filter out inconsistent and high-entropy samples and add the remaining to the memory. A frequency vector is maintained to select samples to discard when the memory is full. STAMP leverages the reliable samples in the memory to optimize the model with the SAM optimizer alongside a decaying step strategy." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 422, + 266, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 422, + 266, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 422, + 266, + 434 + ], + "type": "text", + "content": "through the following process:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 250, + 438, + 481, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 438, + 481, + 472 + ], + "spans": [ + { + "bbox": [ + 250, + 438, + 481, + 472 + ], + "type": "interline_equation", + "content": "\\hat {p} (x) = \\frac {1}{K} \\sum_ {k = 1} ^ {K} f _ {\\theta_ {t}} \\left(\\tau_ {k} (x)\\right), \\tag {2}", + "image_path": "d8418d5d262a80ce2b1b1603cd7db6108ddad28ec9e8eade768f60df08f155a1.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 477, + 482, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 477, + 482, + 537 + ], + "spans": [ + { + "bbox": [ + 130, + 477, + 482, + 537 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 477, + 482, + 537 + ], + "type": "inline_equation", + "content": "\\tau(\\cdot)" + }, + { + "bbox": [ + 130, + 477, + 482, + 537 + ], + "type": "text", + "content": " is the operation of augmentation, " + }, + { + "bbox": [ + 130, + 477, + 482, + 537 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 477, + 482, + 537 + ], + "type": "text", + "content": " is the total number of augmented views. By averaging predictions over augmented samples, the model can better eliminate the uncertainty inherent in its predictions, improving calibration and obtaining more reliable confidence estimates. In the supplementary material, we discuss the impact of the augmentation-averaged estimation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 538, + 482, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 538, + 482, + 644 + ], + "spans": [ + { + "bbox": [ + 130, + 538, + 482, + 644 + ], + "type": "text", + "content": "After acquiring the prediction " + }, + { + "bbox": [ + 130, + 538, + 482, + 644 + ], + "type": "inline_equation", + "content": "\\hat{p}" + }, + { + "bbox": [ + 130, + 538, + 482, + 644 + ], + "type": "text", + "content": ", we implement filtering strategies to prevent unreliable samples from being added to memory for two reasons. Firstly, optimization using unreliable samples can result in error accumulation over time [38]. Secondly, we prevent outliers from entering the memory and reduce the entropy of normal samples in the memory. This action facilitates the model's ability to distinguish between normal samples and outliers. Since reliable samples that are beneficial to optimization tend to have the same pseudo-label as the source model, we employ consistency filtering to remove samples whose predictions are inconsistent with those of the source model:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 194, + 647, + 481, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 647, + 481, + 669 + ], + "spans": [ + { + "bbox": [ + 194, + 647, + 481, + 669 + ], + "type": "interline_equation", + "content": "m _ {c o n} (x) = \\mathbb {I} \\left(\\arg \\max _ {c} [ \\hat {p} (x) ] _ {c} = \\arg \\max _ {c} [ f _ {\\theta_ {s}} (x) ] _ {c}\\right), \\tag {3}", + "image_path": "6e813e56c40905ee5a3f81b4f13809a040c929ec4384661fcfafbeba9960f210.jpg" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 203, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 203, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 203, + 100 + ], + "type": "text", + "content": "Yu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 133, + 130, + 482, + 279 + ], + "blocks": [ + { + "bbox": [ + 133, + 115, + 343, + 128 + ], + "lines": [ + { + "bbox": [ + 133, + 115, + 343, + 128 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 343, + 128 + ], + "type": "text", + "content": "Algorithm 1 Reliable Class-Balanced Memory" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 133, + 130, + 482, + 279 + ], + "lines": [ + { + "bbox": [ + 133, + 130, + 482, + 279 + ], + "spans": [ + { + "bbox": [ + 133, + 130, + 482, + 279 + ], + "type": "text", + "content": "Input: memory bank " + }, + { + "bbox": [ + 133, + 130, + 482, + 279 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 133, + 130, + 482, + 279 + ], + "type": "text", + "content": ", a batch of data " + }, + { + "bbox": [ + 133, + 130, + 482, + 279 + ], + "type": "inline_equation", + "content": "\\{x_{i}\\}_{i = 1}^{B}" + }, + { + "bbox": [ + 133, + 130, + 482, + 279 + ], + "type": "text", + "content": ", entropy threshold " + }, + { + "bbox": [ + 133, + 130, + 482, + 279 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{thr}" + }, + { + "bbox": [ + 133, + 130, + 482, + 279 + ], + "type": "text", + "content": ", class frequency " + }, + { + "bbox": [ + 133, + 130, + 482, + 279 + ], + "type": "inline_equation", + "content": "\\{\\xi_c\\}_{c = 1}^{|C_S|}" + }, + { + "bbox": [ + 133, + 130, + 482, + 279 + ], + "type": "text", + "content": ". \nfor " + }, + { + "bbox": [ + 133, + 130, + 482, + 279 + ], + "type": "inline_equation", + "content": "i = 1" + }, + { + "bbox": [ + 133, + 130, + 482, + 279 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 133, + 130, + 482, + 279 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 133, + 130, + 482, + 279 + ], + "type": "text", + "content": " do \nCalculate " + }, + { + "bbox": [ + 133, + 130, + 482, + 279 + ], + "type": "inline_equation", + "content": "\\hat{p}(x_i)" + }, + { + "bbox": [ + 133, + 130, + 482, + 279 + ], + "type": "text", + "content": " by Equation (2) and " + }, + { + "bbox": [ + 133, + 130, + 482, + 279 + ], + "type": "inline_equation", + "content": "\\hat{y}_i = \\arg \\max_c [\\hat{p}(x_i)]_c" + }, + { + "bbox": [ + 133, + 130, + 482, + 279 + ], + "type": "text", + "content": ". Calculate " + }, + { + "bbox": [ + 133, + 130, + 482, + 279 + ], + "type": "inline_equation", + "content": "m_{con}(x_i)" + }, + { + "bbox": [ + 133, + 130, + 482, + 279 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 133, + 130, + 482, + 279 + ], + "type": "inline_equation", + "content": "m_{ent}(x_i)" + }, + { + "bbox": [ + 133, + 130, + 482, + 279 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 133, + 130, + 482, + 279 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 133, + 130, + 482, + 279 + ], + "type": "text", + "content": " by Equation (3), (4). \nif " + }, + { + "bbox": [ + 133, + 130, + 482, + 279 + ], + "type": "inline_equation", + "content": "m_{con}(x_i) = 1" + }, + { + "bbox": [ + 133, + 130, + 482, + 279 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 133, + 130, + 482, + 279 + ], + "type": "inline_equation", + "content": "m_{ent}(x_i) = 1" + }, + { + "bbox": [ + 133, + 130, + 482, + 279 + ], + "type": "text", + "content": " then \nif " + }, + { + "bbox": [ + 133, + 130, + 482, + 279 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 133, + 130, + 482, + 279 + ], + "type": "text", + "content": " is full then \nDiscard one sample with highest class frequency " + }, + { + "bbox": [ + 133, + 130, + 482, + 279 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 133, + 130, + 482, + 279 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 133, + 130, + 482, + 279 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 133, + 130, + 482, + 279 + ], + "type": "text", + "content": ". \nend if \nadd " + }, + { + "bbox": [ + 133, + 130, + 482, + 279 + ], + "type": "inline_equation", + "content": "(x_i, \\hat{y}_i)" + }, + { + "bbox": [ + 133, + 130, + 482, + 279 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 133, + 130, + 482, + 279 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 133, + 130, + 482, + 279 + ], + "type": "text", + "content": ". \nend if \nend for \nUpdate " + }, + { + "bbox": [ + 133, + 130, + 482, + 279 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 133, + 130, + 482, + 279 + ], + "type": "text", + "content": " with Equation (5)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "algorithm" + }, + { + "bbox": [ + 130, + 302, + 479, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 302, + 479, + 350 + ], + "spans": [ + { + "bbox": [ + 130, + 302, + 479, + 350 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 302, + 479, + 350 + ], + "type": "inline_equation", + "content": "\\mathbb{I}(\\cdot)" + }, + { + "bbox": [ + 130, + 302, + 479, + 350 + ], + "type": "text", + "content": " is the indicator function. Moreover, including high entropy samples for optimization can harm the stability of the model [38, 39], while samples with low entropy always possess greater value and reliability. Therefore, we use entropy filtering to filter out samples with high entropy from entering the memory." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 236, + 357, + 481, + 378 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 357, + 481, + 378 + ], + "spans": [ + { + "bbox": [ + 236, + 357, + 481, + 378 + ], + "type": "interline_equation", + "content": "m _ {e n t} (x) = \\mathbb {I} \\left(\\mathcal {H} (\\hat {p} (x)) < \\mathcal {H} _ {t h r}\\right), \\tag {4}", + "image_path": "20bc15f742169fcc0d66c2ed09169a22b825f6e06e3c55d06cc4f4de73241276.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 384, + 479, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 479, + 444 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 479, + 444 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 384, + 479, + 444 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(q) = -\\sum_{c} q_c \\log q_c" + }, + { + "bbox": [ + 130, + 384, + 479, + 444 + ], + "type": "text", + "content": " denotes the entropy of the prediction vector " + }, + { + "bbox": [ + 130, + 384, + 479, + 444 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 130, + 384, + 479, + 444 + ], + "type": "text", + "content": ". We filter out samples with " + }, + { + "bbox": [ + 130, + 384, + 479, + 444 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(\\hat{p}) \\geq \\mathcal{H}_{thr}" + }, + { + "bbox": [ + 130, + 384, + 479, + 444 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 384, + 479, + 444 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{thr}" + }, + { + "bbox": [ + 130, + 384, + 479, + 444 + ], + "type": "text", + "content": " is a threshold. Considering both filtering strategies, we add sample " + }, + { + "bbox": [ + 130, + 384, + 479, + 444 + ], + "type": "inline_equation", + "content": "(x, \\hat{y})" + }, + { + "bbox": [ + 130, + 384, + 479, + 444 + ], + "type": "text", + "content": " into memory bank " + }, + { + "bbox": [ + 130, + 384, + 479, + 444 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 130, + 384, + 479, + 444 + ], + "type": "text", + "content": " if the sample satisfies consistency filtering and entropy filtering simultaneously, i.e., " + }, + { + "bbox": [ + 130, + 384, + 479, + 444 + ], + "type": "inline_equation", + "content": "m_{con}(x) = 1" + }, + { + "bbox": [ + 130, + 384, + 479, + 444 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 384, + 479, + 444 + ], + "type": "inline_equation", + "content": "m_{ent}(x) = 1" + }, + { + "bbox": [ + 130, + 384, + 479, + 444 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 445, + 480, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 445, + 480, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 480, + 540 + ], + "type": "text", + "content": "Since the number of samples stored in memory is limited, we design a class-balanced strategy to choose a sample to discard when the memory is full. Since we adopt entropy minimization as our optimization objective, an issue arises when the model excessively optimizes samples from a single category over a short period. It leads to a bias towards predicting the dominant category. To counteract this bias, we dynamically discard samples from the memory that belong to the category with the highest update frequency. After each optimization of the model, we update the class frequencies in a moving average manner:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 253, + 548, + 481, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 548, + 481, + 563 + ], + "spans": [ + { + "bbox": [ + 253, + 548, + 481, + 563 + ], + "type": "interline_equation", + "content": "\\xi_ {c} ^ {t} = (1 - \\beta) \\xi_ {c} ^ {t - 1} + \\beta \\nu_ {c} ^ {t}, \\tag {5}", + "image_path": "15acb7f1c740930c928083a837aa0ae8157c0a7cf418f149f77e1160b16cd1d7.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 569, + 479, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 479, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 479, + 616 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 569, + 479, + 616 + ], + "type": "inline_equation", + "content": "\\nu_{c}^{t}" + }, + { + "bbox": [ + 130, + 569, + 479, + 616 + ], + "type": "text", + "content": " represents the number of samples with class label " + }, + { + "bbox": [ + 130, + 569, + 479, + 616 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 569, + 479, + 616 + ], + "type": "text", + "content": " in memory and " + }, + { + "bbox": [ + 130, + 569, + 479, + 616 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 569, + 479, + 616 + ], + "type": "text", + "content": " is moving average coefficient hyperparameter. While the memory is full, we discard a sample from the class with the highest " + }, + { + "bbox": [ + 130, + 569, + 479, + 616 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 130, + 569, + 479, + 616 + ], + "type": "text", + "content": ". We summarize the work process of reliable class-balanced memory in Algorithm 1." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "type": "text", + "content": "Self-Weighted Entropy Minimization. After updating the memory, we replay the reliable samples stored in the memory " + }, + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "type": "text", + "content": " to optimize the model. Following [51], we employ entropy minimization as our loss function. This self-supervised loss encourages the model to exhibit higher confidence in its predic" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 179, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 179, + 90, + 447, + 102 + ], + "type": "text", + "content": "Outlier-Aware Test-Time Adaptation with Stable Memory Replay" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 140, + 131, + 455, + 246 + ], + "blocks": [ + { + "bbox": [ + 133, + 115, + 307, + 129 + ], + "lines": [ + { + "bbox": [ + 133, + 115, + 307, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 307, + 129 + ], + "type": "text", + "content": "Algorithm 2 The Pipeline of STAMP" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 140, + 131, + 455, + 246 + ], + "lines": [ + { + "bbox": [ + 140, + 131, + 455, + 246 + ], + "spans": [ + { + "bbox": [ + 140, + 131, + 455, + 246 + ], + "type": "text", + "content": "Initialize: Memory bank " + }, + { + "bbox": [ + 140, + 131, + 455, + 246 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = \\emptyset" + }, + { + "bbox": [ + 140, + 131, + 455, + 246 + ], + "type": "text", + "content": " zero-initialized class frequency " + }, + { + "bbox": [ + 140, + 131, + 455, + 246 + ], + "type": "inline_equation", + "content": "\\{\\xi_c\\}_{c = 1}^{|C_S|}" + }, + { + "bbox": [ + 140, + 131, + 455, + 246 + ], + "type": "text", + "content": " \nInput: Unlabeled test samples " + }, + { + "bbox": [ + 140, + 131, + 455, + 246 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_T = \\{x_i\\}_{i = 1}^{NT}" + }, + { + "bbox": [ + 140, + 131, + 455, + 246 + ], + "type": "text", + "content": " , a trained model " + }, + { + "bbox": [ + 140, + 131, + 455, + 246 + ], + "type": "inline_equation", + "content": "f_{\\theta_s}(\\cdot)" + }, + { + "bbox": [ + 140, + 131, + 455, + 246 + ], + "type": "text", + "content": " \nfor each online mini-batch " + }, + { + "bbox": [ + 140, + 131, + 455, + 246 + ], + "type": "inline_equation", + "content": "X_{b}" + }, + { + "bbox": [ + 140, + 131, + 455, + 246 + ], + "type": "text", + "content": " do Calculate " + }, + { + "bbox": [ + 140, + 131, + 455, + 246 + ], + "type": "inline_equation", + "content": "\\hat{p} (x)" + }, + { + "bbox": [ + 140, + 131, + 455, + 246 + ], + "type": "text", + "content": " by Equation (2) and " + }, + { + "bbox": [ + 140, + 131, + 455, + 246 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(\\hat{p} (x))" + }, + { + "bbox": [ + 140, + 131, + 455, + 246 + ], + "type": "text", + "content": " for all " + }, + { + "bbox": [ + 140, + 131, + 455, + 246 + ], + "type": "inline_equation", + "content": "x\\in X_b" + }, + { + "bbox": [ + 140, + 131, + 455, + 246 + ], + "type": "text", + "content": " . Obtain the prediction " + }, + { + "bbox": [ + 140, + 131, + 455, + 246 + ], + "type": "inline_equation", + "content": "\\hat{y} = \\arg \\max_{c}[\\hat{p} (x)]_{c}" + }, + { + "bbox": [ + 140, + 131, + 455, + 246 + ], + "type": "text", + "content": " for all " + }, + { + "bbox": [ + 140, + 131, + 455, + 246 + ], + "type": "inline_equation", + "content": "x\\in X_b" + }, + { + "bbox": [ + 140, + 131, + 455, + 246 + ], + "type": "text", + "content": " Update memory bank " + }, + { + "bbox": [ + 140, + 131, + 455, + 246 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 140, + 131, + 455, + 246 + ], + "type": "text", + "content": " by Algorithm 1. Compute " + }, + { + "bbox": [ + 140, + 131, + 455, + 246 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{sem}" + }, + { + "bbox": [ + 140, + 131, + 455, + 246 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 140, + 131, + 455, + 246 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 140, + 131, + 455, + 246 + ], + "type": "text", + "content": " by Equation (6). Update " + }, + { + "bbox": [ + 140, + 131, + 455, + 246 + ], + "type": "inline_equation", + "content": "\\theta_t" + }, + { + "bbox": [ + 140, + 131, + 455, + 246 + ], + "type": "text", + "content": " with SAM in Equation (7) and decay step size in Equation (8). \nend for \nOutput: The prediction and OOD score " + }, + { + "bbox": [ + 140, + 131, + 455, + 246 + ], + "type": "inline_equation", + "content": "\\{\\hat{y}_i,\\mathcal{H}(\\hat{p} (x_i))\\}_{i = 1}^{NT}" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "algorithm" + }, + { + "bbox": [ + 130, + 273, + 482, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 273, + 482, + 312 + ], + "spans": [ + { + "bbox": [ + 130, + 273, + 482, + 312 + ], + "type": "text", + "content": "tions. In order to give more valuable samples a higher importance in optimization, we weight samples in " + }, + { + "bbox": [ + 130, + 273, + 482, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 130, + 273, + 482, + 312 + ], + "type": "text", + "content": " in a self-weighted manner. Given samples " + }, + { + "bbox": [ + 130, + 273, + 482, + 312 + ], + "type": "inline_equation", + "content": "\\{x_{i}\\}_{i = 1}^{|M|}" + }, + { + "bbox": [ + 130, + 273, + 482, + 312 + ], + "type": "text", + "content": ", we calculate the following loss function:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 217, + 319, + 481, + 347 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 319, + 481, + 347 + ], + "spans": [ + { + "bbox": [ + 217, + 319, + 481, + 347 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s e m} (\\mathcal {M}, \\theta_ {t}) = \\sum_ {x _ {i} \\in \\mathcal {M}} \\frac {w _ {i}}{\\sum_ {j} w _ {j}} \\cdot \\mathcal {H} \\left(f _ {\\theta_ {t}} \\left(x _ {i}\\right)\\right) \\tag {6}", + "image_path": "50064eccb99e04214611a3acd4966fef92ca668afd16e6a7c7a337462d324093.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 358, + 482, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 358, + 482, + 432 + ], + "spans": [ + { + "bbox": [ + 130, + 358, + 482, + 432 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 358, + 482, + 432 + ], + "type": "inline_equation", + "content": "w_{i} = e^{-\\mathcal{H}(f_{\\theta_{t}}(x_{i}))}" + }, + { + "bbox": [ + 130, + 358, + 482, + 432 + ], + "type": "text", + "content": " denotes the weight of sample " + }, + { + "bbox": [ + 130, + 358, + 482, + 432 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 130, + 358, + 482, + 432 + ], + "type": "text", + "content": ". It's worth noting that " + }, + { + "bbox": [ + 130, + 358, + 482, + 432 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 130, + 358, + 482, + 432 + ], + "type": "text", + "content": " is also involved with the gradient back-propagation. The reason for this adaptive weighting is two-fold: (i) it gives greater weights to samples with lower entropy, which are considered more reliable. (ii) The model needs to learn the weight of each sample and adaptively give important samples higher weights, instead of solely minimizing entropy for all samples." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 432, + 483, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 432, + 483, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 432, + 483, + 540 + ], + "type": "text", + "content": "Stable Optimization Strategy. To optimize the model in a stable way, we employ sharpness-aware minimization (SAM) [14] and step size decay in parameter updating. SAM is a technique that aims to improve generalization performance in deep learning models by minimizing a loss function while simultaneously controlling the sharpness of the loss landscape. Previous work [14] demonstrates that SAM exhibits resilience against noisy labels, which is also a prevalent challenge in TTA. Therefore, we consider SAM as a good choice of the optimizer in the adaptation scenario [17, 39]. Mathematically, this involves minimizing the following objective function:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 164, + 549, + 481, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 549, + 481, + 571 + ], + "spans": [ + { + "bbox": [ + 164, + 549, + 481, + 571 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta_ {t}} \\mathcal {L} _ {s e m} ^ {S A} (\\mathcal {M}, \\theta_ {t}), \\text {w h e r e} \\mathcal {L} _ {s e m} ^ {S A} (\\mathcal {M}, \\theta_ {t}) \\triangleq \\max _ {| | \\epsilon | | _ {2} \\leq \\rho} \\mathcal {L} _ {s e m} (\\mathcal {M}, \\theta_ {t} + \\epsilon), \\tag {7}", + "image_path": "7801deec33840cb3da4d23d53790bb38bd2f6407da7bbcdee73cf2580b525028.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 580, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 580, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 580, + 482, + 604 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 580, + 482, + 604 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 130, + 580, + 482, + 604 + ], + "type": "text", + "content": " represents the radius of the neighborhood. SAM minimizes the maximum loss within the parameters' vicinity." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": "Recent research [26, 42] has revealed that under long-term adaptation, TTA algorithms may suffer from substantial error accumulation and consequent performance degradation. This issue becomes even more severe when test data contains outliers. To mitigate this problem, we introduce a test-time step size (a.k.a. learning rate) decay strategy. To achieve this, we update the model's parameters" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 203, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 203, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 203, + 101 + ], + "type": "text", + "content": "Yu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 143, + 168, + 471, + 357 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 1: Results (\\%) on CIFAR10-C and CIFAR100-C datasets under outlier-aware TTA setting across four outlier datasets. We report the accuracy (ACC), AUC score, and H-score. All the results are averaged over 15 different corruptions. The results in bold denote the best performance." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 143, + 168, + 471, + 357 + ], + "lines": [ + { + "bbox": [ + 143, + 168, + 471, + 357 + ], + "spans": [ + { + "bbox": [ + 143, + 168, + 471, + 357 + ], + "type": "table", + "html": "
NoiseSVHN-CLSUN-CTinyImageNet-CAvg.
ACCAUCH-scoreACCAUCH-scoreACCAUCH-scoreACCAUCH-scoreACCAUCH-score
CIFAR10-CSource57.370.462.357.367.461.157.362.859.657.364.559.457.366.360.6
BN Stats [36]72.968.670.678.775.376.979.479.479.479.072.975.877.574.075.7
Tent [51]77.448.759.780.854.965.181.262.370.481.165.672.480.157.966.9
EATA [38]72.968.570.678.875.376.979.479.479.478.973.175.977.574.175.7
SAR [39]72.968.570.678.775.376.979.479.479.479.072.975.877.574.075.7
CoTTA [54]77.362.467.381.678.680.182.284.283.281.975.378.480.875.177.2
RoTTA [54]77.674.375.678.476.077.278.879.579.178.673.375.878.375.876.9
SoTTA [17]77.851.761.679.372.875.979.877.978.879.672.675.979.168.873.1
OWTTT [28]62.364.458.566.175.369.663.178.968.556.358.856.262.069.363.2
STAMP77.980.778.882.379.080.583.584.984.282.675.178.681.679.980.5
CIFAR100-CSource35.843.138.035.849.440.135.858.243.235.857.142.735.852.041.0
BN Stats [36]45.880.958.452.772.560.953.773.862.053.268.659.751.374.060.3
Tent [51]47.955.851.254.470.461.255.472.462.755.068.660.953.266.859.0
EATA [38]55.286.167.158.175.665.658.877.266.758.670.764.057.777.465.9
SAR [39]57.588.668.959.265.261.960.573.566.360.872.165.959.574.965.8
CoTTA [54]47.083.459.953.773.261.854.376.963.654.568.160.452.475.461.4
RoTTA [59]47.954.049.447.367.055.348.369.556.747.865.555.047.864.054.1
SoTTA [17]54.453.352.853.670.360.754.470.861.453.968.460.154.165.758.8
OWTTT [28]47.170.356.253.974.362.354.573.562.554.268.560.452.471.660.4
STAMP57.997.572.663.780.070.963.780.771.263.974.468.762.383.170.8
", + "image_path": "7b3035ccfaa5e959292840a993697c635d13c3ddee7018816b43cc80a585d197.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 380, + 313, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 380, + 313, + 391 + ], + "spans": [ + { + "bbox": [ + 131, + 380, + 313, + 391 + ], + "type": "inline_equation", + "content": "\\theta_{t}" + }, + { + "bbox": [ + 131, + 380, + 313, + 391 + ], + "type": "text", + "content": " using cosine annealing decay step size:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 170, + 399, + 480, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 399, + 480, + 422 + ], + "spans": [ + { + "bbox": [ + 170, + 399, + 480, + 422 + ], + "type": "interline_equation", + "content": "\\theta_ {t + 1} = \\theta_ {t} - \\alpha_ {t} \\nabla \\mathcal {L} _ {\\text {s e m}} ^ {S A} (\\mathcal {M}, \\theta_ {t}), \\text {w h e r e} \\alpha_ {t} = \\frac {1}{2} \\alpha_ {0} (1 + \\cos (\\frac {t}{T} \\pi)), \\tag {8}", + "image_path": "727b3bc0347c927b1e2ffc46743f6131aa4a75eba9fea8885ab597c89ae0cca5.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 430, + 482, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 482, + 528 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 482, + 528 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 430, + 482, + 528 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 130, + 430, + 482, + 528 + ], + "type": "text", + "content": " is the step size at timestamp " + }, + { + "bbox": [ + 130, + 430, + 482, + 528 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 430, + 482, + 528 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 430, + 482, + 528 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 430, + 482, + 528 + ], + "type": "text", + "content": " is a hyper-parameter to control the step size decay. While step size decay is commonly utilized in model training [56], we leverage it during testing for a distinct purpose. In the early stages of TTA, the model should adapt swiftly due to the significant disparity between the source and target domains. However, as adaptation progresses and the model acquires comprehensive distributional information about the target domain, it should reduce parameter updates to prevent error accumulation and enhance stability. We summarize the pipeline of STAMP in Algorithm 2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 546, + 230, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 546, + 230, + 560 + ], + "spans": [ + { + "bbox": [ + 131, + 546, + 230, + 560 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 572, + 190, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 572, + 190, + 585 + ], + "spans": [ + { + "bbox": [ + 131, + 572, + 190, + 585 + ], + "type": "text", + "content": "4.1 Setup" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": "Datasets. Following previous TTA methods [16, 18, 54], we adopt three corruption datasets as normal datasets (i.e., in-distribution datasets). CIFAR10-C and CIFAR100-C [20, 24] stand out as the most widely used benchmarks in TTA tasks, comprising 10/100 categories with 15 types of corruptions respectively. Each type of corruption has 10,000 images. Similarly, ImageNet-C [9, 20] serves as a significant benchmark, consisting of 1,000 categories with the same 15 types" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 179, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 179, + 91, + 447, + 103 + ], + "type": "text", + "content": "Outlier-Aware Test-Time Adaptation with Stable Memory Replay" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 196, + 146, + 418, + 260 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 2: Results (\\%) on ImageNet-C under outlier-aware TTA setting across two outlier datasets. All the results are averaged over 15 different corruptions." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 196, + 146, + 418, + 260 + ], + "lines": [ + { + "bbox": [ + 196, + 146, + 418, + 260 + ], + "spans": [ + { + "bbox": [ + 196, + 146, + 418, + 260 + ], + "type": "table", + "html": "
MethodPlaces365-CTextures-C
ACCAUCH-scoreACCAUCH-score
Source18.261.626.118.254.625.8
BN Stats [36]31.167.741.131.661.240.7
Tent [51]34.951.839.53948.642.0
EATA [38]46.472.656.046.462.252.8
SAR [39]44.973.355.045.667.054.0
CoTTA [54]33.866.943.534.260.742.8
RoTTA [59]36.668.646.537.065.346.5
SoTTA [17]41.767.850.741.860.348.8
OWTTT [28]9.154.013.99.459.414.6
STAMP46.476.257.246.570.055.6
", + "image_path": "74968869a8c03d406750a6f97af169933a5c888734434c786b5c5b9d7b38130c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 283, + 482, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 283, + 482, + 379 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 482, + 379 + ], + "type": "text", + "content": "of corruption. For out-of-distribution datasets, we choose Noise, SVHN [37], LSUN [57], and TinyImageNet [25] for the CIFAR10/100-C datasets, and Textures [6] and Places365 [64] for ImageNet-C. To maintain consistency in domain shift between outliers and normal datasets, we apply the same corruptions " + }, + { + "bbox": [ + 130, + 283, + 482, + 379 + ], + "type": "inline_equation", + "content": "^4" + }, + { + "bbox": [ + 130, + 283, + 482, + 379 + ], + "type": "text", + "content": " on outlier datasets. (e.g., SVHN-C refers to the corruptions version of SVHN dataset.) Throughout all experiments, we maintain a fixed ratio of 8:2 between normal samples and outliers. Further details of the outlier datasets are provided in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 380, + 482, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 571 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 571 + ], + "type": "text", + "content": "Baselines. To have a comprehensive comparison of existing TTA methods, we select eight TTA algorithms as our baselines. Source refers to the source pre-trained model. BN Stats [36] adjusts batch normalization (BN) layers [22] with the statistics of test data. Tent [51] optimizes the BN affine parameters in the model by entropy minimization. CoTTA [54] calculates weighted-average and augmentation-averaged pseudo-label to distill knowledge from the teacher model to the student model, and stochastically restores parameters to prevent the model from catastrophic forgetting. EATA [38] adopts the entropy and diversity weight with elastic weight consolidation [23] for anti-forgetting. SAR [39] excludes high-entropy samples from sharpness-aware entropy minimization and reset the model when it overfits. RoTTA [59] uses robust batch normalization and category-balanced sampling with timeliness and uncertainty to deal with the correlated label problem. SoTTA [17] applies high-confidence uniform-class sampling and entropy-sharpness minimization to alleviate the impact of noisy samples. OWTTT [28] proposes prototype expansion with prototype clustering and distribution alignment." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 572, + 482, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 572, + 482, + 644 + ], + "spans": [ + { + "bbox": [ + 130, + 572, + 482, + 644 + ], + "type": "text", + "content": "Evaluation Metrics. We evaluate the above baselines for both generalization and outlier detection. For generalization, we adopt classification accuracy (ACC) as a metric. For outlier detection, we consider the area under the ROC curve (AUC) [8], a threshold-free metric to measure the binary classification problem. The ROC curve depicts the relationship between the true positive rate (TPR) and the false positive rate (FPR). Models with better detection ability" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "type": "text", + "content": "Yu et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 652, + 329, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 652, + 329, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 652, + 329, + 666 + ], + "type": "text", + "content": "4 https://github.com/hendrycks/robustness" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 204, + 168, + 411, + 274 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 3: Results (\\%) on CIFAR10/100-C datasets under outlier-aware TTA setting. We divide datasets into the normal part and the outlier part. In detail, we choose 8/80 classes as known samples and other 2/20 classes as outliers for CIFAR10/100-C. All the results are averaged over 15 different corruptions." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 204, + 168, + 411, + 274 + ], + "lines": [ + { + "bbox": [ + 204, + 168, + 411, + 274 + ], + "spans": [ + { + "bbox": [ + 204, + 168, + 411, + 274 + ], + "type": "table", + "html": "
MethodCIFAR10-C (8:2)CIFAR100-C (80:20)
ACCAUCH-scoreACCAUCH-score
Source60.661.560.037.160.644.2
BN Stats [36]79.466.972.655.566.260.2
Tent [51]80.165.972.360.465.662.8
EATA [38]81.567.974.161.367.464.2
SAR [39]82.366.073.263.469.166.1
CoTTA [54]81.865.472.557.166.361.3
RoTTA [59]79.462.469.850.663.756.2
SoTTA [17]82.762.771.361.468.264.6
OWTTT [28]65.963.564.456.266.060.6
STAMP85.068.776.066.070.768.2
", + "image_path": "fb3b3cef1563476ffe91510603cef8fc13444f5ec1cd428e35d69a2f1332eb2f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 294, + 479, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 294, + 479, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 294, + 479, + 319 + ], + "type": "text", + "content": "will have a higher AUC score. To consider both generalization and detection performance simultaneously, we also report H-score:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 246, + 324, + 481, + 349 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 324, + 481, + 349 + ], + "spans": [ + { + "bbox": [ + 246, + 324, + 481, + 349 + ], + "type": "interline_equation", + "content": "H - s c o r e = \\frac {2 \\cdot A C C \\cdot A U C}{A C C + A U C} \\tag {9}", + "image_path": "46d06d0e4e48b8bc8bc9ce7910b9b9afc3b260f5392d15949a0e7dc213711640.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 355, + 485, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 355, + 485, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 355, + 485, + 536 + ], + "type": "text", + "content": "Implementation Details. We adopt the corruption with severity 5 in all experiments. We consider ResNet-18 [19] as backbones for CIFAR10-C, CIFAR100-C and ResNet-50 [19] for ImageNet-C. For CIFAR10-C and CIFAR100-C, we train the source model using standard cross-entropy loss with the initial learning rate of 0.1 and multi-step learning rate scheduling. For ImageNet-C, we use the released checkpoint from TorchVision library [35]. For a fair comparison, we set the batch size for all methods as 64 and the capacity of our memory bank is 64. Concerning the hyperparameters, we search the hyperparameters on a validation set and set the learning rate to " + }, + { + "bbox": [ + 130, + 355, + 485, + 536 + ], + "type": "inline_equation", + "content": "0.1 / 0.05 / 0.01" + }, + { + "bbox": [ + 130, + 355, + 485, + 536 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 355, + 485, + 536 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{thr}" + }, + { + "bbox": [ + 130, + 355, + 485, + 536 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 355, + 485, + 536 + ], + "type": "inline_equation", + "content": "0.25 / 0.9 / 0.8 \\times \\ln |\\mathcal{C}_S|" + }, + { + "bbox": [ + 130, + 355, + 485, + 536 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 355, + 485, + 536 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 355, + 485, + 536 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 355, + 485, + 536 + ], + "type": "inline_equation", + "content": "150 / 150 / 750" + }, + { + "bbox": [ + 130, + 355, + 485, + 536 + ], + "type": "text", + "content": " for CIFAR10-C/CIFAR100-C/ImageNet-C. Moreover, we set the other hyperparameter the same in experiments with " + }, + { + "bbox": [ + 130, + 355, + 485, + 536 + ], + "type": "inline_equation", + "content": "\\beta = 0.1" + }, + { + "bbox": [ + 130, + 355, + 485, + 536 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 355, + 485, + 536 + ], + "type": "inline_equation", + "content": "\\rho = 0.05" + }, + { + "bbox": [ + 130, + 355, + 485, + 536 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 355, + 485, + 536 + ], + "type": "inline_equation", + "content": "K = 16" + }, + { + "bbox": [ + 130, + 355, + 485, + 536 + ], + "type": "text", + "content": ". For parameters' updating, we select the affine parameters in the BN layer for optimization and freeze the top layer's parameter following [39]. More details of hyper-parameters searching and setting of baselines and STAMP can be found in the supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 551, + 198, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 551, + 198, + 562 + ], + "spans": [ + { + "bbox": [ + 132, + 551, + 198, + 562 + ], + "type": "text", + "content": "4.2 Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Results under outlier-aware TTA setting. We first evaluate TTA methods on the datasets with outliers. The results for CIFAR datasets are illustrated in Table 1. We observe that a simple baseline method, Tent [51], has a competitive generalization performance on distribution shift (achieving an accuracy of " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "80.1\\%" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " on CIFAR10-C). However, it exhibits substantial degradation in the AUC score compared to Source, for instance, dropping from " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "70.4\\%" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "48.7\\%" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " on CIFAR10-C with Noise outlier dataset. This decline could be attributed to Tent's indiscriminate minimization of entropy across all samples, leading to a gradual loss of its" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 179, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 179, + 91, + 447, + 103 + ], + "type": "text", + "content": "Outlier-Aware Test-Time Adaptation with Stable Memory Replay" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 143, + 146, + 471, + 342 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 4: Classification accuracies (%) on CIFAR10-C and CIFAR100-C under closed-set TTA setting." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 143, + 146, + 471, + 342 + ], + "lines": [ + { + "bbox": [ + 143, + 146, + 471, + 342 + ], + "spans": [ + { + "bbox": [ + 143, + 146, + 471, + 342 + ], + "type": "table", + "html": "
MethodGaussianshotinputdefocusglassrotationzonationshowfrostlogbrightnesscontrastelasticpixelatejpegAvg.
CIFAR10-CSource28.735.224.257.349.066.464.876.762.973.490.331.478.846.574.757.3
BN Stats [36]70.272.063.687.666.585.986.982.180.584.190.885.477.378.674.379.0
Tent [51]76.178.470.587.870.186.887.584.882.085.191.186.679.482.778.581.8
CoTTA [54]76.177.673.987.971.986.987.582.882.085.290.885.779.380.478.381.8
EATA [38]76.577.971.789.170.487.489.085.284.286.091.588.479.983.978.582.6
SAR [39]70.772.166.887.668.285.986.982.180.584.190.885.477.378.674.379.4
RoTTA [59]69.671.062.987.567.085.986.982.479.684.791.273.578.377.974.978.2
SoTTA [17]75.879.271.589.470.687.889.085.684.087.292.487.479.984.579.082.9
OWTTT [28]70.172.063.086.966.585.786.782.780.984.591.283.077.976.874.978.9
STAMP80.982.977.287.674.986.687.785.985.988.190.487.280.386.782.684.3
CIFAR100-CSource12.414.57.236.044.745.145.249.541.636.963.313.257.523.846.635.8
BN Stats [36]41.141.338.963.151.560.863.851.253.553.264.459.058.458.047.653.7
Tent [51]52.352.147.766.956.164.365.358.358.760.067.862.161.863.054.559.4
CoTTA [54]47.047.845.159.654.059.561.353.355.052.962.650.358.161.853.154.8
EATA [38]50.753.548.167.055.664.867.059.159.260.467.763.961.863.354.659.8
SAR [39]55.155.051.268.458.266.067.460.360.861.969.865.563.666.256.861.7
RoTTA [59]35.936.633.860.647.157.760.848.042.250.859.232.153.852.344.447.7
SoTTA [17]51.653.847.466.956.965.368.158.960.160.169.463.162.362.854.860.1
OWTTT [28]41.642.838.863.452.661.664.853.354.854.565.858.460.157.649.354.6
STAMP57.258.552.869.961.468.170.163.363.964.872.269.966.569.259.064.4
", + "image_path": "2542c034bffd43d1a1d9f69c312cc7af93f238a7cf2befa283fddd2c70fe7a87.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 365, + 482, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 508 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 508 + ], + "type": "text", + "content": "ability to discern outliers. In addition, our experimental findings demonstrate that with judicious parameter selection, TTA methods can enhance not only the generalization capability against distribution shift but also the detection ability for outliers. Compared to existing TTA methods, STAMP achieves the most favorable performance on both classification and outlier detection. On CIFAR10-C, STAMP achieves the highest ACC, and " + }, + { + "bbox": [ + 130, + 365, + 482, + 508 + ], + "type": "inline_equation", + "content": "+4.1\\%" + }, + { + "bbox": [ + 130, + 365, + 482, + 508 + ], + "type": "text", + "content": " huge improvements in AUC compared to the previous methods. Regarding CIFAR100-C, STAMP exhibits " + }, + { + "bbox": [ + 130, + 365, + 482, + 508 + ], + "type": "inline_equation", + "content": "+2.8\\%" + }, + { + "bbox": [ + 130, + 365, + 482, + 508 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 365, + 482, + 508 + ], + "type": "inline_equation", + "content": "+5.7\\%" + }, + { + "bbox": [ + 130, + 365, + 482, + 508 + ], + "type": "text", + "content": " improvements compared to the most effective TTA methods on ACC and AUC, respectively. We also report the results on the ImageNet-C in Table 2. STAMP still achieves the highest performance. Specifically, we outperform existing TTA methods with " + }, + { + "bbox": [ + 130, + 365, + 482, + 508 + ], + "type": "inline_equation", + "content": "+1.2\\%" + }, + { + "bbox": [ + 130, + 365, + 482, + 508 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 365, + 482, + 508 + ], + "type": "inline_equation", + "content": "+1.6\\%" + }, + { + "bbox": [ + 130, + 365, + 482, + 508 + ], + "type": "text", + "content": " on H-score with Places365-C and Textures-C outlier datasets, respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 509, + 482, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 482, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 482, + 556 + ], + "type": "text", + "content": "Additionally, we evaluate the TTA algorithm under a more challenging scenario involving homogeneous outlier datasets in Table 3, which makes it more difficult to identify outliers. STAMP also exhibits the most promising performance on CIFAR10-C and CIFAR100-C under this scenario." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 557, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 581 + ], + "type": "text", + "content": "Concerning the above results, we achieve optimal results in most cases, affirming the robustness and effectiveness of STAMP." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 582, + 491, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 491, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 491, + 665 + ], + "type": "text", + "content": "Results under closed-set TTA setting. Additionally, we assess our method in a scenario without outliers as most TTA works do. We present the classification accuracy on the clean dataset in Table 4. Remarkably, our method attains the highest accuracy on both CIFAR10-C and CIFAR100-C datasets, achieving an average improvement of " + }, + { + "bbox": [ + 130, + 582, + 491, + 665 + ], + "type": "inline_equation", + "content": "+2.9\\%" + }, + { + "bbox": [ + 130, + 582, + 491, + 665 + ], + "type": "text", + "content": " over the second-best method (SoTTA). This shows that STAMP still has strong applicability even in the traditional situation without outliers." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "type": "text", + "content": "Yu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 221, + 168, + 394, + 234 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "text", + "content": "Table 5: Ablation studies on reliable class-balanced memory (RBM), self-weighted entropy minimization (S.W.), sharpness-aware minimization (S.A.), and decay step size (D.S.). Experiments are conducted on CIFAR100-C with the outlier dataset SVHN-C. The first row in the table reports the results of entropy minimization." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 221, + 168, + 394, + 234 + ], + "lines": [ + { + "bbox": [ + 221, + 168, + 394, + 234 + ], + "spans": [ + { + "bbox": [ + 221, + 168, + 394, + 234 + ], + "type": "table", + "html": "
S.A.D.S.RBMS.W.ACCAUCH-score
----54.470.461.2
-✓--58.965.361.8
✓---56.472.863.5
✓✓--59.270.864.4
✓✓-✓59.377.767.1
✓✓✓-63.075.068.4
✓✓✓✓63.780.070.9
", + "image_path": "7ab798d67bc1fe82b92f9f7727cda7668d26ff796bfbe5dee96e3da0a63577b7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 257, + 479, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 257, + 479, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 257, + 479, + 281 + ], + "type": "text", + "content": "To summarize, all the above results prove that our method achieves the best results under both outlier-aware TTA and conventional closed-set settings." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 300, + 238, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 300, + 238, + 312 + ], + "spans": [ + { + "bbox": [ + 131, + 300, + 238, + 312 + ], + "type": "text", + "content": "4.3 Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 321, + 482, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 321, + 482, + 417 + ], + "spans": [ + { + "bbox": [ + 130, + 321, + 482, + 417 + ], + "type": "text", + "content": "Effect of each component. We delve into the role of each component within our method as follows. First, we provide the results of different weight strategies of entropy minimization in Figure 3. We compare our adaptive weight with static weight and EATA weight [38]. EATA calculates the weight by " + }, + { + "bbox": [ + 130, + 321, + 482, + 417 + ], + "type": "inline_equation", + "content": "\\exp[\\mathcal{H}_{thr} - \\mathcal{H}(p(x))]" + }, + { + "bbox": [ + 130, + 321, + 482, + 417 + ], + "type": "text", + "content": " and static weight calculates weight in the same way of Equation (6). Different from adaptive weight, both EATA and static weight don't involve any gradient back-propagation. As depicted in Figure 3, adaptive weight shows the best performance among the strategies." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 418, + 482, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 482, + 455 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 482, + 455 + ], + "type": "text", + "content": "Furthermore, we conduct an ablation study on each component of our method, as summarized in Table 5. We observe that the effect of combining test-time step size decay and SAM is better than using either of them alone. Step size decay can" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 227, + 487, + 378, + 601 + ], + "blocks": [ + { + "bbox": [ + 227, + 487, + 378, + 601 + ], + "lines": [ + { + "bbox": [ + 227, + 487, + 378, + 601 + ], + "spans": [ + { + "bbox": [ + 227, + 487, + 378, + 601 + ], + "type": "image", + "image_path": "631e1fc72f95b0203dde42cd3276886b19169289d5e6f67a9cdb90b679ca47d6.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 614, + 482, + 659 + ], + "lines": [ + { + "bbox": [ + 130, + 614, + 482, + 659 + ], + "spans": [ + { + "bbox": [ + 130, + 614, + 482, + 659 + ], + "type": "text", + "content": "Fig. 3: Results for the different weight strategies on dataset CIFAR100-C with SVHN-C outlier dataset. Static weight represents that the " + }, + { + "bbox": [ + 130, + 614, + 482, + 659 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 130, + 614, + 482, + 659 + ], + "type": "text", + "content": " in Equation (6) without gradient back-propagation. EATA weight represents the weight for reliable samples used in EATA [38]." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 179, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 179, + 91, + 447, + 103 + ], + "type": "text", + "content": "Outlier-Aware Test-Time Adaptation with Stable Memory Replay" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 221, + 146, + 391, + 208 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 136 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 136 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 136 + ], + "type": "text", + "content": "Table 6: H-score (\\%) with the different ratios of outliers to total samples on CIFAR100-C with different outlier datasets." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 221, + 146, + 391, + 208 + ], + "lines": [ + { + "bbox": [ + 221, + 146, + 391, + 208 + ], + "spans": [ + { + "bbox": [ + 221, + 146, + 391, + 208 + ], + "type": "table", + "html": "
SVHN-CLSUN-CTinyImageNet-C
5%68.470.268.3
10%70.170.968.8
20%70.971.268.7
33%70.370.168.1
50%69.368.867.4
", + "image_path": "49a40b826ed8945b8d1a88d6924286e40193df24250940a61e21de6dc09fa88e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 229, + 480, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 229, + 480, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 229, + 480, + 384 + ], + "type": "text", + "content": "improve the ACC but result in AUC decline. However, combining SAM and step size decay mitigates the decline in AUC. In addition, after the introduction of self-weighted entropy minimization, the AUC of the model further increased to " + }, + { + "bbox": [ + 130, + 229, + 480, + 384 + ], + "type": "inline_equation", + "content": "77.7\\%" + }, + { + "bbox": [ + 130, + 229, + 480, + 384 + ], + "type": "text", + "content": ". Notably, by removing reliable class-balanced memory from our method, the ACC decreases from " + }, + { + "bbox": [ + 130, + 229, + 480, + 384 + ], + "type": "inline_equation", + "content": "63.7\\%" + }, + { + "bbox": [ + 130, + 229, + 480, + 384 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 229, + 480, + 384 + ], + "type": "inline_equation", + "content": "59.3\\%" + }, + { + "bbox": [ + 130, + 229, + 480, + 384 + ], + "type": "text", + "content": ". It reveals that reliable class-balanced memory significantly contributes to enhancing the generalization ability to distribution shift. Moreover, by removing the self-weighted entropy minimization in STAMP, its AUC drops from " + }, + { + "bbox": [ + 130, + 229, + 480, + 384 + ], + "type": "inline_equation", + "content": "80.0\\%" + }, + { + "bbox": [ + 130, + 229, + 480, + 384 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 229, + 480, + 384 + ], + "type": "inline_equation", + "content": "75.0\\%" + }, + { + "bbox": [ + 130, + 229, + 480, + 384 + ], + "type": "text", + "content": ". Self-weighted entropy minimization plays an important role in outlier detection. Besides decreasing the sample entropy, self-weighted entropy minimization incorporates entropy into the optimization weight, allowing outliers that deviate from the majority of samples to receive lower weights, consequently increasing their entropy. This mechanism enhances outlier detection performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 385, + 481, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 385, + 481, + 469 + ], + "spans": [ + { + "bbox": [ + 130, + 385, + 481, + 469 + ], + "type": "text", + "content": "Effect of outlier proportion. In the realistic scenario, the proportion of outliers can vary considerably. We investigate how this variability affects the performance of our method. Specifically, we manipulate the proportion of outliers, ranging from " + }, + { + "bbox": [ + 130, + 385, + 481, + 469 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 130, + 385, + 481, + 469 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 385, + 481, + 469 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 385, + 481, + 469 + ], + "type": "text", + "content": ", and present the corresponding H-scores in Table 6. The results show that the fluctuation of our method is within " + }, + { + "bbox": [ + 130, + 385, + 481, + 469 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 130, + 385, + 481, + 469 + ], + "type": "text", + "content": ". The findings suggest that STAMP exhibits robustness across different scenarios with varying proportions of outliers, as it does not experience significant degradation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 487, + 220, + 500 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 220, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 220, + 500 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 513, + 480, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 513, + 480, + 656 + ], + "spans": [ + { + "bbox": [ + 130, + 513, + 480, + 656 + ], + "type": "text", + "content": "In this study, we research the challenge of handling outliers in test-time adaptation. To tackle this issue effectively, we introduce a scenario called outlier-aware test-time adaptation, which refers to conducting both recognition and outlier rejection simultaneously. Additionally, we propose STAMP, using augmentation-averaged estimation and a filtering mechanism to store reliable samples in memory. Moreover, we dynamically discard samples with the highest class frequency from memory to maintain a class-balanced prediction. Then we replay the samples stored in the memory with self-weighted entropy minimization. Furthermore, we leverage sharpness-aware minimization and learning rate decay to bolster the stability of our algorithm. Through extensive experiments across multiple benchmarks, we demonstrate that our method surpasses existing TTA methods, underscoring its effectiveness and versatility in open-world applications." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "type": "text", + "content": "Yu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 482, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 482, + 199 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 482, + 199 + ], + "type": "text", + "content": "We thank Yuhe Ding and Zhengbo Wang for their valuable feedback on our work. This work was funded by the Beijing Municipal Science and Technology Project (No. Z231100010323005), the Beijing Nova Program (No. Z211100002121108), the Young Elite Scientists Sponsorship Program by CAST (2023QNRC001), and the National Natural Science Foundation of China under (No. 62276256)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 217, + 197, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 217, + 197, + 229 + ], + "spans": [ + { + "bbox": [ + 132, + 217, + 197, + 229 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 240, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 138, + 240, + 480, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 240, + 480, + 262 + ], + "spans": [ + { + "bbox": [ + 138, + 240, + 480, + 262 + ], + "type": "text", + "content": "1. An, S., Park, S., Kim, G., Baek, J., Lee, B., Kim, S.: Context enhanced transformer for single image object detection. arXiv preprint arXiv:2312.14492 (2023) 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 263, + 480, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 263, + 480, + 283 + ], + "spans": [ + { + "bbox": [ + 138, + 263, + 480, + 283 + ], + "type": "text", + "content": "2. Boudiaf, M., Mueller, R., Ben Ayed, I., Bertinetto, L.: Parameter-free online test-time adaptation. In: Proc. CVPR (2022) 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 285, + 480, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 285, + 480, + 305 + ], + "spans": [ + { + "bbox": [ + 138, + 285, + 480, + 305 + ], + "type": "text", + "content": "3. Cao, H., Xu, Y., Yang, J., Yin, P., Yuan, S., Xie, L.: Multi-modal continual test-time adaptation for 3d semantic segmentation. In: Proc. ICCV (2023) 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 307, + 480, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 307, + 480, + 338 + ], + "spans": [ + { + "bbox": [ + 138, + 307, + 480, + 338 + ], + "type": "text", + "content": "4. Chakrabarty, G., Sreenivas, M., Biswas, S.: Santa: Source anchoring network and target alignment for continual test time adaptation. Transactions on Machine Learning Research (2023) 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 339, + 480, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 339, + 480, + 360 + ], + "spans": [ + { + "bbox": [ + 138, + 339, + 480, + 360 + ], + "type": "text", + "content": "5. Chen, D., Wang, D., Darrell, T., Ebrahimi, S.: Contrastive test-time adaptation. In: Proc. CVPR (2022) 2, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 361, + 480, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 361, + 480, + 381 + ], + "spans": [ + { + "bbox": [ + 138, + 361, + 480, + 381 + ], + "type": "text", + "content": "6. Cimpoi, M., Maji, S., Kokkinos, I., Mohamed, S., Vedaldi, A.: Describing textures in the wild. In: Proc. CVPR (2014) 3, 10" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 383, + 480, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 383, + 480, + 415 + ], + "spans": [ + { + "bbox": [ + 138, + 383, + 480, + 415 + ], + "type": "text", + "content": "7. Croce, F., Andriushchenko, M., Sehwag, V., Debenedetti, E., Flammarion, N., Chiang, M., Mittal, P., Hein, M.: Robustbench: a standardized adversarial robustness benchmark. In: Proc. NeurIPS (2021) 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 416, + 480, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 416, + 480, + 436 + ], + "spans": [ + { + "bbox": [ + 138, + 416, + 480, + 436 + ], + "type": "text", + "content": "8. Davis, J., Goadrich, M.: The relationship between precision-recall and roc curves. In: Proc. ICML (2006) 10" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 437, + 480, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 437, + 480, + 458 + ], + "spans": [ + { + "bbox": [ + 138, + 437, + 480, + 458 + ], + "type": "text", + "content": "9. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: Proc. CVPR (2009) 3, 9" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 459, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 459, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 138, + 459, + 480, + 479 + ], + "type": "text", + "content": "0. DeVries, T., Taylor, G.W.: Learning confidence for out-of-distribution detection in neural networks. arXiv preprint arXiv:1802.04865 (2018) 3, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 481, + 480, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 481, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 138, + 481, + 480, + 512 + ], + "type": "text", + "content": "1. Ding, Y., Liang, J., Jiang, B., Zheng, A., He, R.: Maps: A noise-robust progressive learning approach for source-free domain adaptive keypoint detection. IEEE Transactions on Circuits and Systems for Video Technology (2023) 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 514, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 514, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 514, + 480, + 544 + ], + "type": "text", + "content": "2. Ding, Y., Sheng, L., Liang, J., Zheng, A., He, R.: Proxymix: Proxy-based mixup training with label refinery for source-free domain adaptation. Neural Networks 167, 92-103 (2023) 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 546, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 546, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 546, + 480, + 578 + ], + "type": "text", + "content": "3. Farahani, A., Voghoei, S., Rasheed, K., Arabnia, H.R.: A brief review of domain adaptation. Advances in Data Science and Information Engineering: Proceedings from ICDATA 2020 and IKE 2020 pp. 877-894 (2021) 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 579, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 579, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 138, + 579, + 480, + 599 + ], + "type": "text", + "content": "4. Foret, P., Kleiner, A., Mobahi, H., Neyshabur, B.: Sharpness-aware minimization for efficiently improving generalization. In: Proc. ICLR (2021) 3, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 601, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 480, + 621 + ], + "type": "text", + "content": "5. Gebru, T., Hoffman, J., Fei-Fei, L.: Fine-grained recognition in the wild: A multitask domain adaptation approach. In: Proc. ICCV (2017) 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "type": "text", + "content": "6. Gong, T., Jeong, J., Kim, T., Kim, Y., Shin, J., Lee, S.J.: Note: Robust continual test-time adaptation against temporal correlation. In: Proc. NeurIPS (2022) 4, 9" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "type": "text", + "content": "7. Gong, T., Kim, Y., Lee, T., Chottananurak, S., Lee, S.J.: Sotta: Robust test-time adaptation on noisy data streams. In: Proc. NeurIPS (2023) 4, 5, 8, 9, 10, 11, 12" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "type": "text", + "content": "Outlier-Aware Test-Time Adaptation with Stable Memory Replay" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 664 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "18. Goyal, S., Sun, M., Raghunathan, A., Kolter, J.Z.: Test time adaptation via conjugate pseudo-labels. In: Proc. NeurIPS (2022) 9" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "type": "text", + "content": "19. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proc. CVPR (2016) 11" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "type": "text", + "content": "20. Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. In: Proc. ICLR (2018) 3, 9" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "type": "text", + "content": "21. Hendrycks, D., Gimpel, K.: A baseline for detecting misclassified and out-of-distribution examples in neural networks. In: Proc. ICLR (2016) 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "text", + "content": "22. Ioffe, S., Szegedy, C.: Batch normalization: Accelerating deep network training by reducing internal covariate shift. In: Proc. ICML (2015) 10" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 227, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 270 + ], + "type": "text", + "content": "23. Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A.A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., et al.: Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences 114(13), 3521-3526 (2017) 10" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 271, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 480, + 293 + ], + "type": "text", + "content": "24. Krizhevsky, A.: Learning multiple layers of features from tiny images. Master's thesis, University of Tront (2009) 9" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 293, + 480, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 313 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 313 + ], + "type": "text", + "content": "25. Le, Y., Yang, X.: Tiny imagenet visual recognition challenge. CS 231N 7(7), 3 (2015) 3, 10" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "type": "text", + "content": "26. Lee, J., Das, D., Choo, J., Choi, S.: Towards open-set test-time adaptation utilizing the wisdom of crowds in entropy minimization. In: Proc. CVPR (2023) 4, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "type": "text", + "content": "27. Lee, K., Lee, K., Lee, H., Shin, J.: A simple unified framework for detecting out-of-distribution samples and adversarial attacks. In: Proc. NeurIPS (2018) 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 358, + 480, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 389 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 389 + ], + "type": "text", + "content": "28. Li, Y., Xu, X., Su, Y., Jia, K.: On the robustness of open-world test-time training: Self-training with dynamic prototype expansion. In: Proc. CVPR (2023) 4, 9, 10, 11, 12" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 391, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 391, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 391, + 480, + 413 + ], + "type": "text", + "content": "29. Liang, J., He, R., Tan, T.: A comprehensive survey on test-time adaptation under distribution shifts. International Journal of Computer Vision (2024) 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 414, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 480, + 435 + ], + "type": "text", + "content": "30. Liang, J., Hu, D., Feng, J.: Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In: Proc. ICML (2020) 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "type": "text", + "content": "31. Liang, J., Sheng, L., Wang, Z., He, R., Tan, T.: Realistic unsupervised clip finetuning with universal entropy optimization. In: Proc. ICML (2024) 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 457, + 480, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 478 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 478 + ], + "type": "text", + "content": "32. Liang, S., Li, Y., Srikant, R.: Enhancing the reliability of out-of-distribution image detection in neural networks. In: Proc. ICLR (2018) 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 479, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 500 + ], + "type": "text", + "content": "33. Lin, Z., Roy, S.D., Li, Y.: Mood: Multi-level out-of-distribution detection. In: Proc. CVPR (2021) 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 501, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 522 + ], + "type": "text", + "content": "34. Liu, W., Wang, X., Owens, J., Li, Y.: Energy-based out-of-distribution detection. In: Proc. NeurIPS (2020) 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 523, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 480, + 544 + ], + "type": "text", + "content": "35. Maintainers, T., Contributors: Torchvision: Pytorch's computer vision library (2016) 11" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 545, + 480, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 480, + 577 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 480, + 577 + ], + "type": "text", + "content": "36. Nado, Z., Padhy, S., Sculley, D., D'Amour, A., Lakshminarayanan, B., Snoek, J.: Evaluating prediction-time batch normalization for robustness under covariate shift. In: Proc. ICML Workshops (2020) 3, 9, 10, 11, 12" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 577, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 577, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 577, + 480, + 599 + ], + "type": "text", + "content": "37. Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B., Ng, A.Y.: Reading digits in natural images with unsupervised feature learning (2011) 3, 10" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "text", + "content": "38. Niu, S., Wu, J., Zhang, Y., Chen, Y., Zheng, S., Zhao, P., Tan, M.: Efficient test-time model adaptation without forgetting. In: Proc. ICML (2022) 2, 4, 5, 6, 7, 9, 10, 11, 12, 13" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 632, + 480, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 664 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 664 + ], + "type": "text", + "content": "39. Niu, S., Wu, J., Zhang, Y., Wen, Z., Chen, Y., Zhao, P., Tan, M.: Towards stable test-time adaptation in dynamic wild world. In: Proc. ICLR (2022) 4, 7, 8, 9, 10, 11, 12" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "type": "text", + "content": "Yu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "40. Nixon, J., Dusenberry, M.W., Zhang, L., Jerfel, G., Tran, D.: Measuring calibration in deep learning. In: Proc. CVPR Workshops (2019) 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 480, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 480, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 480, + 159 + ], + "type": "text", + "content": "41. Park, S., Yang, S., Choo, J., Yun, S.: Label shift adapter for test-time adaptation under covariate and label shifts. In: Proc. CVPR (2023) 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 160, + 480, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 480, + 191 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 480, + 191 + ], + "type": "text", + "content": "42. Press, O., Schneider, S., Kümmerer, M., Bethge, M.: Rdumb: A simple approach that questions our progress in continual test-time adaptation. In: Proc. NeurIPS (2023) 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 192, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 192, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 192, + 480, + 224 + ], + "type": "text", + "content": "43. Schneider, S., Rusak, E., Eck, L., Bringmann, O., Brendel, W., Bethge, M.: Improving robustness against common corruptions by covariate shift adaptation. In: Proc. NeurIPS (2020) 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 224, + 480, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 480, + 246 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 480, + 246 + ], + "type": "text", + "content": "44. Shanmugam, D., Blalock, D., Balakrishnan, G., Guttag, J.: Better aggregation in test-time augmentation. In: Proc. CVPR (2021) 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 247, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 247, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 247, + 480, + 267 + ], + "type": "text", + "content": "45. Sheng, L., Liang, J., He, R., Wang, Z., Tan, T.: Adaptguard: Defending against universal attacks for model adaptation. In: Proc. ICCV (2023) 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 268, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 268, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 268, + 480, + 289 + ], + "type": "text", + "content": "46. Tan, M., Le, Q.: Efficientnet: Rethinking model scaling for convolutional neural networks. In: Proc. ICML (2019) 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 289, + 480, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 480, + 310 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 480, + 310 + ], + "type": "text", + "content": "47. Tong, P., Su, W., Li, H., Ding, J., Haoxiang, Z., Chen, S.X.: Distribution free domain generalization. In: Proc. ICML (2023) 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 311, + 480, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 311, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 132, + 311, + 480, + 331 + ], + "type": "text", + "content": "48. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: Proc. NeurIPS (2017) 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 332, + 480, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 480, + 353 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 480, + 353 + ], + "type": "text", + "content": "49. Volpi, R., De Jorge, P., Larlus, D., Csurka, G.: On the road to online adaptation for semantic image segmentation. In: Proc. CVPR (2022) 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 354, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 480, + 385 + ], + "type": "text", + "content": "50. Vyas, A., Jammalamadaka, N., Zhu, X., Das, D., Kaul, B., Willke, T.L.: Out-of-distribution detection using an ensemble of self supervised leave-out classifiers. In: Proc. ECCV (2018) 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 386, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 480, + 407 + ], + "type": "text", + "content": "51. Wang, D., Shelhamer, E., Liu, S., Olshausen, B., Darrell, T.: Tent: Fully test-time adaptation by entropy minimization. In: Proc. ICML (2020) 2, 4, 7, 9, 10, 11, 12" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 407, + 480, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 480, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 480, + 439 + ], + "type": "text", + "content": "52. Wang, J., Lan, C., Liu, C., Ouyang, Y., Qin, T., Lu, W., Chen, Y., Zeng, W., Yu, P.S.: Generalizing to unseen domains: A survey on domain generalization. IEEE Transactions on Knowledge and Data Engineering 35(8), 8052-8072 (2023) 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 440, + 480, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 440, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 440, + 480, + 460 + ], + "type": "text", + "content": "53. Wang, M., Deng, W.: Deep visual domain adaptation: A survey. Neurocomputing 312, 135-153 (2018) 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 461, + 480, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 461, + 480, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 461, + 480, + 482 + ], + "type": "text", + "content": "54. Wang, Q., Fink, O., Van Gool, L., Dai, D.: Continual test-time domain adaptation. In: Proc. CVPR (2022) 4, 9, 10, 11, 12" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 483, + 480, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 480, + 503 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 480, + 503 + ], + "type": "text", + "content": "55. Wang, Y., Li, B., Che, T., Zhou, K., Liu, Z., Li, D.: Energy-based open-world uncertainty modeling for confidence calibration. In: Proc. CVPR (2021) 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 504, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 480, + 525 + ], + "type": "text", + "content": "56. You, K., Long, M., Wang, J., Jordan, M.I.: How does learning rate decay help modern neural networks? arXiv preprint arXiv:1908.01878 (2019) 9" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 525, + 480, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 480, + 557 + ], + "type": "text", + "content": "57. Yu, F., Seff, A., Zhang, Y., Song, S., Funkhouser, T., Xiao, J.: Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365 (2015) 3, 10" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 558, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 558, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 558, + 480, + 588 + ], + "type": "text", + "content": "58. Yu, Y., Sheng, L., He, R., Liang, J.: Benchmarking test-time adaptation against distribution shifts in image classification. arXiv preprint arXiv:2307.03133 (2023) 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 590, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 480, + 610 + ], + "type": "text", + "content": "59. Yuan, L., Xie, B., Li, S.: Robust test-time adaptation in dynamic scenarios. In: Proc. CVPR (2023) 2, 3, 9, 10, 11, 12" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "text", + "content": "60. Yurtsever, E., Lambert, J., Carballo, A., Takeda, K.: A survey of autonomous driving: Common practices and emerging technologies. IEEE Access 8, 58443-58469 (2020) 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "61. Zhang, M., Levine, S., Finn, C.: Memo: Test time robustness via adaptation and augmentation. In: Proc. NeurIPS (2022) 4, 5" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "type": "text", + "content": "Outlier-Aware Test-Time Adaptation with Stable Memory Replay" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 226 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 148 + ], + "type": "text", + "content": "62. Zhang, Y., Wang, X., Jin, K., Yuan, K., Zhang, Z., Wang, L., Jin, R., Tan, T.: Adanpc: Exploring non-parametric classifier for test-time adaptation. In: Proc. ICML (2023) 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 149, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 480, + 171 + ], + "type": "text", + "content": "63. Zhao, H., Liu, Y., Alahi, A., Lin, T.: On pitfalls of test-time adaptation. In: Proc. ICML (2023) 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "type": "text", + "content": "64. Zhou, B., Lapedriza, A., Khosla, A., Oliva, A., Torralba, A.: Places: A 10 million image database for scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence 40(6), 1452-1464 (2017) 3, 10" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "text", + "content": "65. Zhou, Z., Guo, L.Z., Jia, L.H., Zhang, D., Li, Y.F.: Ods: test-time adaptation in the presence of open-world data shift. In: Proc. ICML (2023) 4" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "type": "text", + "content": "Yu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/STSP_ Spatial-Temporal Subspace Projection for Video Class-incremental Learning/9413ef29-ccb6-4b51-bb25-c4b8dee0d763_content_list.json b/2024/STSP_ Spatial-Temporal Subspace Projection for Video Class-incremental Learning/9413ef29-ccb6-4b51-bb25-c4b8dee0d763_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..db2577d5b64d8e88c759eb9513c24d945c901e72 --- /dev/null +++ b/2024/STSP_ Spatial-Temporal Subspace Projection for Video Class-incremental Learning/9413ef29-ccb6-4b51-bb25-c4b8dee0d763_content_list.json @@ -0,0 +1,1799 @@ +[ + { + "type": "text", + "text": "STSP: Spatial-Temporal Subspace Projection for Video Class-incremental Learning", + "text_level": 1, + "bbox": [ + 222, + 140, + 779, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hao Cheng $^{1,2}$ , Siyuan Yang $^{1}$ , Chong Wang $^{1}$ , Joey Tianyi Zhou $^{3,4}$ , Alex C. Kot $^{1}$ , and Bihan Wen $^{1*}$", + "bbox": [ + 220, + 210, + 781, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Nanyang Technological University, Singapore {siyuan.yang, wang1711, eackot, bihan.wen}@ntu.edu.sg", + "bbox": [ + 305, + 253, + 694, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Hebei University of Technology, Tianjin, China 2024025@hebut.edu.cn", + "bbox": [ + 334, + 282, + 665, + 308 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ CFAR, Agency for Science, Technology and Research, Singapore", + "bbox": [ + 279, + 309, + 720, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 IHPC, Agency for Science, Technology and Research, Singapore zhouty@cfar.a-star.edu.sg", + "bbox": [ + 281, + 324, + 720, + 352 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Video class-incremental learning (VCIL) aims to learn discriminative and generalized feature representations for video frames to mitigate catastrophic forgetting. Conventional VCIL methods often retain a subset of frames or features from prior tasks as exemplars for subsequent incremental learning stages. However, these strategies overlook the connection between base and novel classes, sometimes even leading to privacy leakage. To address this challenge, we introduce a Spatial-Temporal Subspace Projection (STSP) scheme for VCIL. Specifically, we propose a discriminative Temporal-based Subspace Classifier (TSC) that represents each class with an orthogonal subspace basis and adopts subspace projection loss for classification. Unlike typical classification methods relying on fully connected layers, our TSC discerns the spatial-temporal dynamics in video content, thereby enhancing the representation of each video sample. Additionally, we implement inter- and intra-class orthogonal constraints into TSC, ensuring each class occupies a unique orthogonal subspace, defined by its basis. To prevent catastrophic forgetting, we further employ a Spatial-based Gradient Projection (SGP) strategy, adjusting network gradients to align with the null space of the spatial feature set from previous tasks. Extensive experiments conducted on three benchmarks, namely HMDB51, UCF101, and Something-Something V2, demonstrate that our STSP method outperforms state-of-the-art comparison methods, evidencing its efficacy in VCIL.", + "bbox": [ + 261, + 378, + 743, + 684 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Video action recognition $\\cdot$ Class incremental learning $\\cdot$ Subspace Projection $\\cdot$ Gradient Projection", + "bbox": [ + 259, + 696, + 740, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 744, + 375, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep neural networks have demonstrated remarkable performance in modern computer vision tasks [7,17,32,37,46], particularly when trained on large, annotated datasets. However, the real-world challenge of incrementally incorporating", + "bbox": [ + 212, + 773, + 787, + 820 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author.", + "bbox": [ + 217, + 825, + 385, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/173df513cde9670e818f909bb4b61e1031d81913c37529a5d66ea38e36c0f6dc.jpg", + "image_caption": [ + "Fig. 1: Class incremental learning with (a) linear classifier and (b) subspace-based classifier. Incorporating new classes alters the decision hyperspace and modifies the representation of prior classes, leading to misalignment. In contrast, subspace-based classifiers, by constructing an orthogonal basis for each class, establish a unique geometric structure for each class, thereby reducing class misalignment." + ], + "image_footnote": [], + "bbox": [ + 238, + 145, + 493, + 241 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/eed59928631da623d347d013762b457e615a6324bed1440d775135688415c2fb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 146, + 769, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "data from new classes into a model initially trained on existing classes presents significant hurdles, known as class-incremental learning (CIL) [26]. Unlike typical classification tasks, CIL involves accessing novel classes and limited memories of prior tasks in incremental sessions, often due to privacy concerns or memory constraints. Consequently, merely updating the model with data from new classes can result in overfitting substantial performance degradation on previous classes, a phenomenon referred to as catastrophic forgetting [19, 20].", + "bbox": [ + 212, + 354, + 787, + 460 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent advancements in CIL methods [1,4,26,44] have shown remarkable performance in image classification tasks. However, recent CIL studies [1,21,25,42] demonstrate that straightforward fine-tuning of the network for novel classes may lead the subspace spanned by the novel classifier to deviate from the base one, as illustrated in Fig. 1 (a). The classification performance can be significantly affected by misalignment, particularly in cases where a notable distribution gap between novel and base classes. Moreover, when extending this task to Video Class-Incremental Learning (VCIL), unique challenges due to the inherent information redundancy within video samples, such as continuous actions with a static background. This necessitates methods to model temporal-spatial relationships between video frames to extract critical information for action recognition properly. However, prevalent approaches [22-24,35,36] in VCIL predominantly involve a learning pipeline that employs simple operations (e.g., mean) on the feature vectors of sampled frames to generate video representations for action recognition. While this methodology is effective for standard action recognition with sufficient data, it may struggle to ensure robust generalization in scenarios of incremental learning. The aforementioned observations prompt us to consider acquiring a more suitable representation for both the classifier and video features while also exploring their relationship. To address these challenges, several subspace learning-based methods [28,33,38,45] have shown effectiveness in reducing dimensionality and capturing intrinsic structure, thereby offering a potential solution to the issues associated with simplistic video representation methods.", + "bbox": [ + 212, + 460, + 787, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a spatial-temporal subspace projection method for VCIL including a temporal-based subspace classifier (TSC) and a spatial-based gradient projection (SGP) strategy. Specifically, we extend the existing linear", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "H. Cheng et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "classifiers based on fully connected (FC) layers to a TSC in which a subspace basis is adopted to model each class. Differing from linear classifiers that seek to distinguish classes in feature space through a linear decision boundary, the proposed orthogonality constraints offer a distinct geometric structure for each class. This characteristic not only facilitates a more comprehensive representation of each class but also allows for straightforward extension to novel classes. Furthermore, to avoid catastrophic forgetting, we utilize the SGP strategy to constrain the process of gradient update by projecting the gradient of network parameters in each layer onto the subspace of corresponding input features so that losses for the old task will not increase during the incremental sessions. Compared with existing VCIL methods [22-24,35,36] that rely on feature distillation or prompt-based fine-tuning strategies, SGP offers the advantage of fine-tuning all network parameters without necessitating regularizers. Furthermore, SGP obviates the need to explicitly retain class-relevant information, such as frames or prototypes, thereby contributing to privacy protection.", + "bbox": [ + 212, + 146, + 787, + 372 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, we make the following contributions: 1) We propose a spatial-temporal subspace projection (STSP) method, containing a temporal-based subspace classifier and spatial-based gradient projection strategy to better model feature and class representation for discriminative classification. 2) We employ a spatial-based gradient projection strategy to constrain the gradient update process without the need to store frames or prototypes from previous tasks. This not only enhances performance but also mitigates the risk of privacy leakage. 3) The proposed STSP method achieves superior results on three video action recognition incremental benchmarks over state-of-the-art methods.", + "bbox": [ + 212, + 373, + 787, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 530, + 387, + 545 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Video Action Recognition", + "text_level": 1, + "bbox": [ + 215, + 560, + 475, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With significant advances in deep learning, current video action recognition approaches aim to learn effective spatial-temporal representation by well-designed network architectures. A notable advancement by Simonyan and Zisserman [29] introduces a dual-stream architecture that analyzes RGB frames in one stream and optical flow in another, laying the foundation for subsequent enhancements. Recently, Lin et al. [17] proposed a Temporal Shift Module (TSM), which shifts some channels along the temporal dimension to perform temporal interaction between the features from adjacent frames. Concurrently, significant research [3, 13, 14, 31, 32, 41] has evolved 2D Convolutional Neural Networks (CNNs) into 3D structures to capture both spatial and temporal contexts within videos, a critical aspect of human action recognition. Notably, Tran et al. [31] introduced the C3D model, a 3D CNN designed to learn spatial-temporal features through an end-to-end learning approach. Carreira et al. [3] developed the Inflated 3D CNN, which extends a 2D CNN into the temporal domain by inflating its convolutional and pooling layers. Additionally, Feichtenhofer et al. [9] crafted a dual-pathway 3D CNN, featuring both slow and fast pathways that process RGB frames at varying frame rates to simultaneously capture semantic content and motion details.", + "bbox": [ + 212, + 583, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Spatial-Temporal Subspace Projection for VCIL", + "bbox": [ + 408, + 113, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Despite significant progress in action recognition, catastrophic forgetting has received relatively limited attention. This work aims to address this gap by developing a spatial-temporal subspace projection method, which aims to offer a robust solution to the challenges of incremental learning in action recognition.", + "bbox": [ + 212, + 146, + 787, + 208 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Class-Incremental Learning", + "text_level": 1, + "bbox": [ + 214, + 228, + 488, + 243 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "CIL addresses the challenge of adapting machine learning models to recognize new classes without forgetting previously learned information, a phenomenon known as catastrophic forgetting. Pioneering in this domain, Rebuffi et al. [26] introduced iCaRL, which combines representation learning with a memory mechanism to selectively retain the knowledge of previous classes. Following this, Hou et al. [12] proposed a UCIR strategy, which addresses the class imbalance in CIL by adjusting the classification layer to focus more on newer classes while still remembering old ones. Moreover, Gao et al. [10] introduced R-DFCIL, employing relation-guided representation learning to overcome catastrophic forgetting and enable the assimilation of new classes in a data-free scenario.", + "bbox": [ + 212, + 251, + 787, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While CIL has been well-studied in the image domain, its application to the video domain remains largely unexplored. Park et al. [22] introduced Time-Channel Distillation (TCD), leveraging knowledge distillation loss with time-channel importance masks. Pei et al. [23] proposed FrameMaker, a memory-efficient VCIL strategy that develops a condensed frame for each video to facilitate future incremental tasks with less memory usage. Pei et al. [24] introduced ST-Prompt, a novel space-time prompting framework tailored for VCIL, enhancing the model's adaptability to new classes with minimal memory increase.", + "bbox": [ + 212, + 402, + 787, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Existing VCIL approaches often rely on preserving a subset of frames or features from previous tasks as exemplars for subsequent incremental stages to prevent catastrophic forgetting. However, these strategies overlook the connection between base and novel classes and potentially lead to privacy leakage. In contrast, we employ a spatial-based gradient projection to regulate gradient updates, thereby mitigating catastrophic forgetting without these drawbacks.", + "bbox": [ + 212, + 523, + 787, + 614 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Subspace Representation", + "text_level": 1, + "bbox": [ + 214, + 633, + 470, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Subspace learning, by learning a suitable representation for high dimensional data, has shed light on various vision tasks [2, 18, 34, 40, 43]. Existing subspace methods such as Principal Component Analysis (PCA) [2, 18] and Linear Discriminant Analysis (LDA) [2, 40] adaptively transfer the original input data to low dimensional subspace for dimensionality reduction. For example, Tzimiropoulos et al. [34] formulated PCA on image gradient orientations, which demonstrated robust performance in face recognition. Recently, subspace learning has been further extended to regularize the implicit features or learned classifiers within deep networks [1,28,38,45]. Notably, Cheraghian et al. [5] constructed a mixture of subspaces to represent the feature clusters formed by visually similar samples. Such subspace representation enables a better approximation of the overall data distribution, consequently leading to superior performance in", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "H. Cheng et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/969ada04ac68c9d7743347b61d21d25e5a17aaa79b0ee3b7f0e24a433ec9a6c5.jpg", + "image_caption": [ + "Fig. 2: The overall framework of the proposed STSP method, including a temporal-based subspace classifier (TSC) and a spatial-based gradient projection (SGP). For each input video, spatial-temporal features are extracted using a backbone network. Subsequently, The TSC classifies by projecting temporal features into a subspace, where classification decisions are based on the distance to each class's subspace basis within this subspace. During parameter updates for the novel tasks, the SGP constrains gradient modifications for each specific layer, drawing on the mean covariance matrix $M_{t-1}$ of previous tasks' input features. Upon the current session completion, the mean covariance matrix is updated to $M_t$ for the gradient projection in the subsequent session." + ], + "image_footnote": [], + "bbox": [ + 228, + 152, + 767, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "CIL. To mitigate the catastrophic forgetting and overfitting, Simon et al. [28] formulated the dynamic classifiers via subspace modeling which demonstrated a significant robustness to noise in few-shot learning. Similarly, Akyurek et al. [1] proposed an additional subspace regularizer that pulls the weight vectors of the novel class close to the subspace spanned by the base classes.", + "bbox": [ + 212, + 428, + 782, + 503 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, existing methods lack sufficient constraints on the representation of each class in the classifier. In this paper, we model each class via an orthogonal subspace basis and leverage spatial-based gradient projection to update the entire network during incremental sessions.", + "bbox": [ + 212, + 506, + 784, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 604, + 328, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we start with an introduction of VCIL problem formulation. Then, the proposed STSP method is described and explained in detail, followed by the loss function of our method as well as the training and inference procedure.", + "bbox": [ + 212, + 651, + 782, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In VCIL, a series of distinct tasks $\\{\\mathcal{T}_1,\\mathcal{T}_2,\\dots ,\\mathcal{T}_k,\\dots \\}$ unfolds sequentially. Each task $\\mathcal{T}_k$ comprises a distinct subset of video-class pairs $\\{(D_i^k,y_i^k),y_i^k\\in$ $\\mathcal{C}_k\\} |_{i = 1}^{|\\mathcal{T}_k|}$ . Specifically, the novel class set $\\mathcal{C}_k$ of $\\mathcal{T}_k$ is distinct from previous classes, i.e., $(\\mathcal{C}_1\\cup \\dots \\cup \\mathcal{C}_{k - 1})\\cap \\mathcal{C}_k = \\emptyset$ . The primary objective is to incrementally train a deep neural network $f(\\cdot)$ capable of accurately classifying videos across all classes by preventing the loss of class information from previous tasks. In contrast to methods that have access to a limited exemplar set from historical tasks, our emphasis lies on exemplar-free VCIL, where the model is not permitted to store exemplars from previous tasks during the learning process of the current session.", + "bbox": [ + 212, + 700, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Spatial-Temporal Subspace Projection for VCIL", + "bbox": [ + 408, + 113, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Overall STSP Framework", + "text_level": 1, + "bbox": [ + 215, + 146, + 473, + 159 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given the problem formulation, we follow the standard learning protocol of previous methods [22, 23] based on knowledge distillation. During the $k^{th}$ incremental session, the network $f(\\cdot ;\\mathbf{w})$ parameterized by $\\mathbf{w}$ is updated from $\\mathbf{w}_{k - 1}$ to adapt to the novel class set $\\mathcal{C}_k$ . Fig. 2 illustrates the overall framework of our method. For an input video $D_{i}^{k}$ belonging to $\\mathcal{C}_k$ , we first extract $T$ frames as $V_{i}$ , which are then fed to the backbone network to generate the feature embeddings. Subsequently, we employ a TSC to differentiate between distinct classes. Moreover, our method stands out by employing an SGP strategy without restoring frames or prototypes from previous tasks. More in detail, the network parameters are optimized by projecting the gradients in each layer to the subspace corresponding to their input features. With the merit of this strategy, we can completely avoid the risk of privacy leakage, whereas existing approaches may suffer.", + "bbox": [ + 212, + 167, + 787, + 349 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Temporal-based Subspace Classifier", + "text_level": 1, + "bbox": [ + 215, + 369, + 555, + 383 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In accordance with the standardized procedure for VCIL, given an input video-class pair denoted as $(D_i,y_i)$ , we first randomly sampled $T$ frames $V_{i}^{T}$ . Subsequently, these frames are fed to the backbone network, yielding the feature embeddings as $X_{i}\\in \\mathbb{R}^{T\\times C\\times H\\times W}$ , where $C$ , $H$ , and $W$ represent the size of channel, height, and width, respectively. For the existing architectures [22-24] for VCIL, a prevalent technique for obtaining video representations is global average pooling (GAP), denoted as $Z_{i} = GAP(X_{i})\\in \\mathbb{R}^{C\\times T}$ . Following this, a linear classifier $L$ , comprising an FC layer with weights $\\mathbf{w}\\in \\mathbb{R}^{C\\times K}$ , is utilized to predict the potential class of each frame. The class predictions from individual frames are subsequently averaged to produce the overall classification result.", + "bbox": [ + 212, + 390, + 787, + 541 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Nevertheless, the temporal information holds the potential to encapsulate more intricate intra-frame structural details, warranting thoughtful consideration and seamless integration into the classification process. Additionally, in each incremental task, adjustments to the linear classifier may be necessary to modify the overall representation across all classes, which could influence the base classes for all previous incremental sessions. Conversely, a subspace-based classifier demonstrates heightened adaptability, enabling the modeling of each new class within its unique subspace without imposing substantial effects on the existing representations of previous classes.", + "bbox": [ + 212, + 542, + 787, + 676 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To this end, to properly model each class and learn a robust classifier, we construct a TSC based on subspaces spanned by an orthogonal basis of each class. Specifically, for $m$ -th class, we construct a subspace with orthogonal basis $P_{m} \\in \\mathbb{R}^{C \\times \\tau}$ with $P_{m}^{\\top} P_{m} = I_{\\tau}, \\tau \\ll C$ , where $C$ and $\\tau$ indicate the feature dimension and the number of subspace basis for $m$ -th class, respectively. Hence, for each task $\\mathcal{T}_k$ in the $k^{th}$ incremental session, we construct $|\\mathcal{C}_k|$ distinct subspaces for $|\\mathcal{C}_k|$ novel classes. We then adopt current subspaces to predict the class of $Z_{i}$ based on the subspace distance between the subspace spanned by temporal video features $Z_{i}$ and the orthogonal basis of each class as:", + "bbox": [ + 214, + 678, + 787, + 813 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nd _ {m} (Z _ {i}) = - \\left\\| Z _ {i} Z _ {i} ^ {\\top} - P _ {m} P _ {m} ^ {\\top} \\right\\| _ {F} ^ {2}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 385, + 821, + 784, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "H. Cheng et al.", + "bbox": [ + 271, + 114, + 374, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Then, the probability of $Z_{i}$ assigned to $m$ -th class can be calculated using the softmax function as follows:", + "bbox": [ + 214, + 146, + 785, + 175 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\np _ {m} (Z _ {i}) = \\frac {e ^ {d _ {m} (Z _ {i})}}{\\sum_ {m ^ {\\prime} = 1} ^ {M} e ^ {d _ {m ^ {\\prime}} (Z _ {i})}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 185, + 785, + 223 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $M$ is the total number of classes in the current session. Now, we can employ the cross-entropy loss on Eq. (2) to update the network as:", + "bbox": [ + 214, + 231, + 782, + 262 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nL _ {C E} = - \\sum_ {i = 1} ^ {J} y _ {i} \\log p _ {i} \\left(Z _ {i}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 410, + 273, + 785, + 314 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $J$ denotes the number of videos in each mini-batch.", + "bbox": [ + 214, + 324, + 633, + 338 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To encourage the network to capture discriminative information for each class and construct distinct subspaces corresponding to different classes, we impose orthogonal constraints on both the temporal features of the video and the weights of the TSC. Specifically, we employ orthogonal loss on the temporal features $Z_{i}$ of each video $D_{i}$ in the final layer as:", + "bbox": [ + 212, + 339, + 782, + 414 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nL _ {O F} = \\sum_ {i = 1} ^ {J} \\left\\| Z _ {i} ^ {\\top} Z _ {i} - I _ {T} \\right\\| _ {F} ^ {2}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 405, + 428, + 785, + 467 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $I_T \\in \\mathbb{R}^{T \\times T}$ is an identity matrix. This approach guarantees that every frame independently captures distinct, class-aware information, setting it apart uniquely from the rest. Furthermore, it facilitates the construction of a discriminative subspace for each corresponding class.", + "bbox": [ + 214, + 473, + 782, + 534 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For TSC, we impose intra- and inter-class orthogonal constraints on the subspace basis of each class, which is denoted by $L_{OW}$ as:", + "bbox": [ + 214, + 535, + 782, + 565 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nL _ {O W} = \\sum_ {m = 1} ^ {M} \\left(\\underbrace {\\left\\| P _ {m} ^ {\\top} P _ {m} - I _ {\\tau} \\right\\| _ {F} ^ {2}} _ {\\text {i n t r a - c l a s s}} + \\sum_ {q = m + 1} ^ {M} \\underbrace {\\left\\| P _ {m} ^ {\\top} P _ {q} \\right\\| _ {F} ^ {2}} _ {\\text {i n t e r - c l a s s}}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 577, + 785, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The intra-class orthogonal constraint is introduced to ensure independence among the basis vectors of subspaces, which is crucial for learning distinct and discriminative feature representations. Furthermore, the imposition of inter-class orthogonal constraint ensures that the basis vectors of distinct classes are orthogonal, signifying their mutual independence. This orthogonality enforcement serves to prevent confusion and safeguard the uniqueness of each basis vector, thus preserving the integrity of class information previously learned. Moreover, these orthogonal constraints enhance discriminability, guaranteeing that the basis vectors of subspaces related to different classes are clearly distinguishable.", + "bbox": [ + 214, + 633, + 782, + 768 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Therefore, the overall objective function for TSC is defined as:", + "bbox": [ + 238, + 770, + 689, + 785 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nL _ {t o t a l} = \\alpha L _ {C E} + \\beta L _ {O F} + \\gamma L _ {O W}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 797, + 785, + 813 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\alpha$ , $\\beta$ , and $\\gamma$ denote the weighting parameters of each losses, respectively.", + "bbox": [ + 214, + 824, + 782, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Spatial-Temporal Subspace Projection for VCIL", + "bbox": [ + 408, + 114, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 127 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Spatial-based Gradient Projection", + "text_level": 1, + "bbox": [ + 215, + 146, + 542, + 162 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In contrast to prior approaches [22-24] that rely on retaining exemplars or prompts from previous classes during incremental sessions, our focus is on addressing the challenging exemplar-free VCIL problems by imposing constraints on the network during parameter updates. To avoid catastrophic forgetting on previous classes in the incremental sessions, the network's parameter update should satisfy the following proposition:", + "bbox": [ + 212, + 176, + 787, + 267 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Proposition 1. Given the data $x_{k}$ from the previous task $\\mathcal{T}_k$ , the current network $f(\\cdot ;\\mathbf{w}_t)$ could preserve the historical information if it satisfies the following equation for all sessions:", + "bbox": [ + 214, + 277, + 789, + 324 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nf \\left(x _ {k}; \\mathbf {w} _ {t}\\right) = f \\left(x _ {k}; \\mathbf {w} _ {k}\\right), \\quad \\forall k < t. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 337, + 785, + 354 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Here, $\\mathbf{w}_t$ denotes the network parameters at task $T_{t}$", + "bbox": [ + 215, + 367, + 593, + 382 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To realize the objective presented in Proposition 1, it is imperative to confine the gradient update of network parameters to a specific direction—precisely, the gradient update should lie in the subspace spanned by the corresponding input spatial features. Specifically, the gradients in the linear layers, including FC and Convolution layers $^5$ should satisfy Proposition 2 [27,39].", + "bbox": [ + 212, + 393, + 787, + 470 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Proposition 2. Let $f^l(\\cdot; \\mathbf{w}_t)$ denotes the $l$ -th linear layer of the network trained on current task $\\mathcal{T}_t$ and its corresponding input feature $A_k^l$ from any previous task $\\mathcal{T}_k$ ( $k < t$ ). During each training step $s$ , if the linear layer's parameter update $\\Delta \\mathbf{w}_{t,s}^l$ lies in the null space of $A_k^l$ , i.e.,", + "bbox": [ + 214, + 484, + 789, + 547 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nA _ {k} ^ {l} \\Delta \\mathbf {w} _ {t, s} ^ {l} = 0, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 450, + 559, + 785, + 579 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "we could preserve the historical information as $f^{l}\\left(A_{k}^{l};\\mathbf{w}_{t}^{l}\\right) = f^{l}\\left(A_{k}^{l};\\mathbf{w}_{k}^{l}\\right)$ .", + "bbox": [ + 215, + 590, + 746, + 611 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Building upon Proposition 2, we further extend it to our proposed TSC with Eq. (1), as detailed in Proposition 3.", + "bbox": [ + 214, + 622, + 785, + 654 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Proposition 3. Let $f^{TSC}(\\cdot ;\\mathbf{w}_t)$ denotes the TSC of the network trained on current task $\\mathcal{T}_t$ and its corresponding input feature $Z_{k}$ from any previous task $\\mathcal{T}_k$ ( $k < t$ ). During each training step $s$ , if the TSC parameter update $\\Delta \\mathbf{w}_{t,s}^{TSC}$ lies in the null space of the subspace spanned by $Z_{k}$ , i.e.,", + "bbox": [ + 214, + 666, + 787, + 729 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nZ _ {k} Z _ {k} ^ {\\top} \\Delta \\mathbf {w} _ {t, s} ^ {T S C} = 0, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 431, + 739, + 785, + 760 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "similarly we have $f^{TSC}\\left(Z_k;\\mathbf{w}_t^{TSC}\\right) = f^{TSC}\\left(Z_k;\\mathbf{w}_k^{TSC}\\right)$ .", + "bbox": [ + 215, + 771, + 627, + 791 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Proof. Please refer to the supplemental material.", + "bbox": [ + 215, + 800, + 594, + 816 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "H. Cheng et al.", + "bbox": [ + 271, + 114, + 375, + 128 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "5 The convolution operation can be regarded as a specific type of linear operation.", + "bbox": [ + 217, + 824, + 767, + 840 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Inspired by previous work [27,39], we calculate the mean covariance matrix, denoted as $M_{t-1}$ , of the input feature for each specific layer in each mini-batch to enhance network stability and optimize memory efficiency. Here, let $Q$ represent the collective term for the current input feature at each specific layer, with $A_t^l$ for the $l$ -th linear layer and $Z_t Z_t^\\top$ for the TSC, respectively. To initialize, $M_0$ is set to 0 for the first session. At the end of each session $t$ , $M_t$ is updated as:", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nM _ {t} = M _ {t - 1} + \\sum_ {b = 1} ^ {\\mathcal {B}} \\frac {1}{J} Q _ {b} ^ {\\top} Q _ {b}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 244, + 785, + 287 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\mathcal{B}$ represents the number of mini-batches in the current task, and $J$ denotes the number of videos in each mini-batch.", + "bbox": [ + 212, + 291, + 784, + 320 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "During the feature update in task $\\mathcal{T}_t$ , we perform Singular Value Decomposition (SVD) to find the approximate null space of the given feature matrix $M_{t-1}$ corresponding to Proposition 2 and Proposition 3 as:", + "bbox": [ + 212, + 321, + 784, + 368 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ U _ {1}, U _ {2} \\right] \\left[ \\begin{array}{c c} \\Sigma_ {1} & 0 \\\\ 0 & \\Sigma_ {2} \\end{array} \\right] \\left[ \\begin{array}{l} U _ {1} ^ {\\top} \\\\ U _ {2} ^ {\\top} \\end{array} \\right] = \\operatorname {S V D} \\left(M _ {t - 1}\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 364, + 375, + 785, + 407 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $U_{1}, U_{2}$ represent unitary matrices that comprise orthonormal bases associated with the singular values in $\\Sigma_{1}, \\Sigma_{2}$ , respectively, with all zero singular values relegated to $\\Sigma_{2}$ . However, it is unrealistic to guarantee that there always exists zero singular values. Previous work [27, 39] constructs $U_{2}^{\\top}$ by selecting singular values restricted to a specific proportion. However, this strategy might not always provide the most compact representation, particularly when there is a significant difference in the scale of singular values. Hence, we employ a threshold strategy based on the energy retained in the matrix $M$ for the construction $U_{2}^{\\top}$ , corresponding to a segment of smallest singular values that satisfy $\\| \\Sigma_{2} \\|_{F}^{2} \\leq (1 - \\eta) \\left( \\| \\Sigma_{1} \\|_{F}^{2} + \\| \\Sigma_{2} \\|_{F}^{2} \\right)$ , with the proportion of retained energy denoted by $\\eta$ . When $\\eta \\approx 1$ , it is reasonable to approximate the null space of the matrix $M$ with $U_{2}^{\\top}$ . This enables the parameter update $\\Delta \\mathbf{w}$ from the gradient $\\mathbf{g}$ concerning the parameters $\\mathbf{w}$ through subspace projection based on $U_{2}$ . That is, during the training step $s$ within any incremental session $t$ , we have:", + "bbox": [ + 212, + 412, + 787, + 625 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\mathbf {w} _ {t, s} = U _ {2} U _ {2} ^ {\\top} \\mathbf {g} _ {t, s}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 434, + 631, + 785, + 648 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.4 Model Optimization and Inference", + "text_level": 1, + "bbox": [ + 214, + 667, + 547, + 683 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Our STSP method is summarized in Algorithm 1. In the training phase, the network parameters are updated conventionally during the initial session. After this initial session, considering parameter updates, the mean covariance of input features from the base classes is retained for each layer. These retained features are then utilized to update parameters according to Eq. (12) at each step in every subsequent incremental session. Additionally, after each incremental session, these parameters are further updated with features extracted from novel classes.", + "bbox": [ + 212, + 688, + 787, + 794 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For the testing phase, no additional memories or operations are necessary. We simply feed the video into the network and obtain classification results using the trained model.", + "bbox": [ + 212, + 795, + 787, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Spatial-Temporal Subspace Projection for VCIL", + "bbox": [ + 408, + 113, + 730, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 127 + ], + "page_idx": 8 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 STSP for class incremental learning" + ], + "code_body": "Input: Task series $\\{\\mathcal{T}_1,\\mathcal{T}_2,\\dots \\}$ , network $f(\\cdot ;\\mathbf{w})$ with parameters w for linear layers and TSC, learning rate $\\alpha$ and energy threshold $\\eta$ \nInitialize: $M_0 = 0$ for each linear layer and TSC. \n1: for $\\mathcal{T}_t\\in \\{\\mathcal{T}_1,\\mathcal{T}_2,\\dots \\}$ do \n2: $s = 0$ \n3: while not converge do \n4: Sample a mini-batch $\\{D^t,y^t\\} \\in \\mathcal{T}_t$ with videos $\\{D_i^t,y_i^t\\}_{j = 1}^J$ \n5: Compute $f(\\{D^{t},y^{t}\\} ;\\mathbf{w}_{t,s})$ and obtain the gradient $\\mathbf{g}_{t,s}$ \n6: if $t > 1$ then \n7: for $\\mathbf{w}_{t,s}$ in linear layers and TSC do \n8: Compute $U_{2}$ using $M_{t - 1}$ via Eq. (11) with a threshold $\\eta$ \n9: Update $\\Delta \\mathbf{w}_{t,s}$ using Eq. (12) \n10: Update $\\mathbf{w}_{t,s + 1} = \\mathbf{w}_{t,s} - \\alpha \\Delta \\mathbf{w}_{t,s}$ \n11: end for \n12: else \n13: Update $\\Delta \\mathbf{w}_{t,s} = \\mathbf{g}_{t,s}$ \n14: Update $\\mathbf{w}_{t,s + 1} = \\mathbf{w}_{t,s} - \\alpha \\Delta \\mathbf{w}_{t,s}$ \n15: end if \n16: $s = s + 1$ \n17: end while \n18: Get current input features $Q$ at each linear layer and TSC \n19: Update $M_t$ using Eq. (10) for later incremental session \n20: end for", + "bbox": [ + 215, + 164, + 785, + 484 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 511, + 375, + 529 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In this section, we compare our results against state-of-the-art techniques in class-incremental action recognition benchmarks. Additionally, we perform ablation studies to demonstrate the efficacy of our proposed method.", + "bbox": [ + 212, + 541, + 784, + 585 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 215, + 606, + 426, + 622 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Datasets. The evaluation of our proposed framework is conducted using the UCF101 [30], HMDB51 [15], and Something-Something V2 [11] datasets, which are recognized benchmarks in video action recognition. The UCF-101 dataset comprises 13,320 instances across 101 action categories, with the data divided into three splits for training (70%) and testing (30%). Similarly, the HMDB-51 dataset features 6,766 video clips distributed among 51 action categories, also with three splits for training and testing in the same proportions. For our evaluations, we selected the first split from both datasets. Distinct from UCF101 and HMDB51, the Something-Something V2 dataset, a crowd-sourced collection, includes 220,000 videos across 174 classes. This dataset uniquely maintains consistent objects and backgrounds across its action categories, thereby emphasizing the importance of a model's capability to discern subtle motion cues.", + "bbox": [ + 212, + 628, + 787, + 809 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation Protocol. Our methodology adheres to the class-incremental learning benchmarks frequently utilized in video analysis, as outlined in TCD [22].", + "bbox": [ + 215, + 809, + 785, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "H. Cheng et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/156348bf1b1ef7e062298c2d10a498541158b830d8fe77f3a263ef5b44ff12c1.jpg", + "table_caption": [ + "Table 1: Video class incremental learning results on UCF101 and HMDB51. The best result for each setting is emphasized in bold." + ], + "table_footnote": [], + "table_body": "
MethodUCF101HMDB51
10 × 5 stages5 × 10 stages2 × 25 stages5 × 5 stages1 × 25 stages
LwFMC [16]42.1425.5911.6826.8216.49
LwM [6]43.3926.0712.0826.9716.50
UCIR [12]74.3170.4263.2244.9037.04
PODNet [8]73.2671.5870.2844.3238.76
TCD [22]74.8973.4372.1945.3440.47
FrameMaker [23]78.1376.3875.7747.5442.65
STSP (Ours)81.1582.8479.2556.9949.19
", + "bbox": [ + 245, + 184, + 751, + 316 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "This approach begins with a model pre-trained on half of the total number of classes, with the remaining classes introduced sequentially at each incremental step. Specifically, for UCF101, the initial training involves 51 classes, after which the subsequent 50 classes are structured into incremental learning tasks of 5, 10, and 25 classes. In the case of HMDB51, the model begins with training on 26 classes, followed by the distribution of the remaining classes into groups of 5 and 25 for subsequent learning phases. For Something-Something V2, the initial training phase covers 84 classes, with the remaining classes organized into incremental learning groups of 10 and 5 classes.", + "bbox": [ + 212, + 349, + 787, + 487 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Implementation Details. Our framework leverages the official implementation of TSM [17], utilizing the PyTorch Library for development. We train a ResNet-34 TSM model on UCF101 for 50 epochs with a batch size of 32. For both HMDB51 and Something-Something V2, ResNet-50 TSM models are employed, undergoing training for 50 epochs with a batch size of 32.", + "bbox": [ + 212, + 489, + 787, + 566 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Main Results", + "text_level": 1, + "bbox": [ + 215, + 603, + 372, + 617 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "For a fair comparison, we employ the same model structure with the same pretraining initialization as the existing methods [22, 23]. Unlike prior approaches that incorporate exemplars for incremental learning and assess performance using both the standard classification protocol and Nearest Mean of Exemplars (NME) metrics [26], our method does not retain any exemplars. Consequently, we only report the standard classification results. We compare our proposed STSP method with the existing video-class incremental learning methods over three benchmarks, as shown in Tab. 1 and Tab. 2. From both tables, we observe that the proposed STSP method significantly outperforms all competing methods across all settings on each dataset. Notably, unlike other approaches that rely on retaining exemplars for incremental learning sessions, our method utilizes gradient projection to constrain model updates, effectively safeguarding against performance drops on previously learned classes.", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Spatial-Temporal Subspace Projection for VCIL", + "bbox": [ + 408, + 113, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/3656398c9388e07a216c98556f1bd19af1218fda393218e25564fe7c1bc7e7a1.jpg", + "table_caption": [ + "Table 2: Video class incremental learning results on Something-Something V2. The best result for each setting is emphasized in bold." + ], + "table_footnote": [], + "table_body": "
Method10 × 9 stages5 × 18 stages
UCIR [12]26.8420.69
PODNet [8]34.9726.95
TCD [22]35.7829.60
FrameMaker [23]37.2530.98
STSP (Ours)69.6870.87
", + "bbox": [ + 220, + 210, + 467, + 296 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/a0aeb7a36745e6fcbd155b0a8e458c7ea06ac247b4203bbdb23144bf96e9987b.jpg", + "table_caption": [ + "Table 3: Ablation studies for our proposed STSP method on UCF101 with $5 \\times 10$ stages and $2 \\times 25$ stages settings, respectively." + ], + "table_footnote": [], + "table_body": "
TSCSGPExemplarsUCF101
5 × 10 stages2 × 25 stages
(i)××5/class68.2857.39
(ii)✓×5/class71.5061.78
(iii)××017.6319.22
(iv)✓×08.7814.99
(v)×✓077.6674.02
(vi)✓✓082.8479.25
", + "bbox": [ + 501, + 198, + 784, + 299 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/f448247ea081d492cf4b2f6a773e7c6a8a7d713f0c9b73ed1d85be9c26a1a0f4.jpg", + "image_caption": [ + "(a) UCF101 (5 × 10 stages)" + ], + "image_footnote": [], + "bbox": [ + 220, + 313, + 498, + 450 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/07d7d1b95cf4a0e3ba7fbd1f8033c9086d61a878a5e6599d581b72cdecd51932.jpg", + "image_caption": [ + "(b) UCF101 (2 × 25 stages)", + "Fig. 3: Plots of accuracy over seen classes along with the incremental steps on UCF101." + ], + "image_footnote": [], + "bbox": [ + 509, + 313, + 787, + 450 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Ablation Study and Analysis", + "text_level": 1, + "bbox": [ + 215, + 518, + 501, + 536 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In this section, we perform ablation studies to investigate the effectiveness of the key components of the proposed STSP on the UCF101 dataset under both $2 \\times 25$ stages and $5 \\times 10$ stages settings.", + "bbox": [ + 212, + 545, + 787, + 592 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Effectiveness of gradient projection To evaluate the SGP's capability to mitigate catastrophic forgetting, we designed experiments with various configurations: utilizing the method with SGP, the method without SGP, and the method without SGP but incorporating exemplars. In the last configuration, we preserved 5 videos, each containing 8 frames, as exemplars for prior classes to be used in future training sessions. Furthermore, we extended these comparisons to assess performance by employing a linear classifier and the proposed TSC. The comparison results are shown in Tab. 3, where (i), (iii), and (v) utilize the linear classifier, while (ii), (iv), and (vi) implement the TSC. In cooperation with the comparative analysis of the network's performance as the number of seen classes increases, as shown in Fig. 3, it can be found that the network tends to forget most knowledge of previous classes when no strategy is adopted.", + "bbox": [ + 212, + 612, + 787, + 794 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Moreover, when comparing our SGP strategy against the exemplar-based distillation method, particularly through the comparison of (i) $v.s.$ (v) and the comparison of (ii) $v.s.$ (vi) in Tab. 3, we can find that our proposed SGP strategy", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "H. Cheng et al.", + "bbox": [ + 271, + 114, + 375, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/2d74f708b12972e5e4466687d76f8d33705e67f3d8b1d6ea10c1977c7fd714d7.jpg", + "table_caption": [ + "Table 4: Ablation study of the subspace basis size $\\tau$ for TSC on UCF101 with $5\\times 10$ stages setting." + ], + "table_footnote": [], + "table_body": "
τ=1τ=2τ=4τ=8τ=16
STSP (Ours)66.7567.8875.3982.8463.02
", + "bbox": [ + 330, + 184, + 668, + 220 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "significantly outperforms the exemplar-based distillation method. Specifically, regarding the performance of base classes during incremental learning sessions, as shown in Fig. 4, it is evident that our proposed SGP has minimal impact on the classification performance for the base classes. These findings indicate that our proposed SGP can uphold accurate classification across all classes without compromising performance on the base classes, emphasizing the effectiveness of SGP in mitigating knowledge degradation.", + "bbox": [ + 212, + 250, + 787, + 356 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Furthermore, based on the comparison in Fig. 3 and Fig. 4, we observe that all methods perform worse under the setting of $2 \\times 25$ stages, while our method with SGP still achieves the best performance. A plausible explanation is that the model may be susceptible to overfitting when dealing with a limited set of novel classes. Moreover, for our method, the computation of the null space becomes more unstable in these scenarios, especially when updating the mean covariance matrix is solely based on a few classes.", + "bbox": [ + 212, + 357, + 787, + 464 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Linear or Subspace-based Classifier? To assess the effectiveness of the proposed TSC, we compare it with a linear classifier based on an FC layer. Moreover, this ablation study extends to both the exemplar-based distillation method and our proposed SGP, with the comparison results shown in Tab. 3. For the exemplar-based distillation approach, the implementation of TSC leads to a performance increase of $3.22\\%$ and $4.39\\%$ on UCF101 at the 10-stage and 25-stage settings, respectively, as evidenced by the comparison between (i) and (ii). In the context of SGP, the comparison between (v) and (vi) reveals that the proposed TSC similarly improves the overall performance by $5.18\\%$ and $5.23\\%$ for the 10-stage and 25-stage settings on UCF101, respectively. Compared with the exemplar-based distillation method, our method employs SGP to constrain the classifier update through gradient subspace projection. The above findings align with the concept of TSC, collectively constraining the subspace-based classifier and further enhancing classification performance.", + "bbox": [ + 212, + 491, + 787, + 702 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "From the comparison in Fig. 4, another observation is that TSC can even boost the classification performance on the base classes, demonstrating its effectiveness in modeling class representations. However, we also observe that TSC may perform poorly when there are no constraints (i.e. without exemplars nor SGP) on the network. One plausible explanation for this phenomenon is that, during incremental sessions, the network tends to shift the feature space toward new tasks without constraints from historical information. Such adjustments can result in the original subspace deviating from its previous configuration, leading to significant misalignment that greatly impacts the performance of TSC.", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Spatial-Temporal Subspace Projection for VCIL", + "bbox": [ + 408, + 113, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/e12369de4149b125837e8d1cf7e00e6f1db461bf46d3b054d2e714a9fe1b1b61.jpg", + "image_caption": [ + "(a) UCF101 $(5\\times 10$ stages)" + ], + "image_footnote": [], + "bbox": [ + 222, + 143, + 488, + 281 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/0171a0332d9e8d19f03421da3bd4fc01d980d684a015deee4823b40c531b01f5.jpg", + "image_caption": [ + "(b) UCF101 (2 × 25 stages)", + "Fig. 4: Plots of accuracy over base classes along with the incremental steps on UCF101." + ], + "image_footnote": [], + "bbox": [ + 509, + 143, + 777, + 281 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The Size of the Subspace Basis We explore how the size of the subspace basis influences the classification accuracy of our proposed method, the comparison results are shown in Tab. 4. It can be found that large subspace basis sizes (i.e., $\\tau = 4$ or $\\tau = 8$ ) are more effective at constructing the subspace for each category, enabling the capture of more distinctive features, which in turn leads to better performance as opposed to smaller bases. This improvement, however, results in an increase in both the network's parameters and computational complexity. Additionally, we can observe that a subspace basis size of 16 yields the least favorable results. This could be attributed to the difficulty in constructing orthogonal subspaces for new classes when the subspace basis is excessively large, especially as $M \\times \\tau$ approaches the total number of classes $(\\mathcal{C})$ .", + "bbox": [ + 212, + 349, + 787, + 517 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 537, + 359, + 553 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we proposed STSP, a novel spatial-temporal subspace projection method designed for video class-incremental learning tasks. The objective is to learn discriminative and robust representations for both video and class, facilitating further incremental learning. STSP mainly consists of two key components, i.e., temporal-base subspace classifier (TSC) and spatial-based gradient projection (SGP). TSC employs a subspace to model each video and its class, incorporating intra-frame temporal information. On the other hand, SGP is designed to mitigate catastrophic forgetting by projecting the gradient of network parameters in each layer onto the subspace spanned by the corresponding input spatial features. Extensive experiments on three benchmarks demonstrate the effectiveness of our proposed STSP method.", + "bbox": [ + 212, + 566, + 787, + 734 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations. In contrast to previous video class incremental learning methods, our proposed STSP method incorporates gradient projection to prevent the forgetting of historical information from previous classes. This strategy removes the necessity of storing exemplars and prototypes of previously learned classes, yet it requires the storage and updating of input features for both linear layers and the introduced TSC. Such a requirement could pose challenges when employing large backbones, requiring further exploration and discussion.", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "H. Cheng et al.", + "bbox": [ + 271, + 114, + 375, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This research was carried out at the Rapid-Rich Object Search (ROSE) Lab at Nanyang Technological University in Singapore.", + "bbox": [ + 215, + 176, + 784, + 205 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 229, + 321, + 244 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Akyurek, A.F., Akyurek, E., Wijaya, D., Andreas, J.: Subspace regularizers for few-shot class incremental learning. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=boJy41J-tnQ", + "2. Belhumeur, P.N., Hespanha, J.P., Kriegman, D.J.: Eigenfaces vs. fisherfaces: Recognition using class specific linear projection. IEEE Transactions on pattern analysis and machine intelligence 19(7), 711-720 (1997)", + "3. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6299-6308 (2017)", + "4. Chen, H., Wang, Y., Hu, Q.: Multi-granularity regularized re-balancing for class incremental learning. IEEE Transactions on Knowledge and Data Engineering 35(7), 7263-7277 (2023)", + "5. Cheraghian, A., Rahman, S., Ramasinghe, S., Fang, P., Simon, C., Petersson, L., Harandi, M.: Synthesized feature based few-shot class-incremental learning on a mixture of subspaces. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 8661-8670 (2021)", + "6. Dhar, P., Singh, R.V., Peng, K.C., Wu, Z., Chellappa, R.: Learning without memorizing. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5138-5146 (2019)", + "7. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=YicbFdNTTy", + "8. Douillard, A., Cord, M., Ollion, C., Robert, T., Valle, E.: Podnet: Pooled outputs distillation for small-tasks incremental learning. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XX 16. pp. 86-102. Springer (2020)", + "9. Feichtenhofer, C., Fan, H., Malik, J., He, K.: Slowfast networks for video recognition. In: Proceedings of the IEEE international conference on computer vision. pp. 6202-6211 (2019)", + "0. Gao, Q., Zhao, C., Ghanem, B., Zhang, J.: R-dfcil: Relation-guided representation learning for data-free class incremental learning. In: European Conference on Computer Vision. pp. 423-439. Springer (2022)", + "1. Goyal, R., Ebrahimi Kahou, S., Michalski, V., Materzynska, J., Westphal, S., Kim, H., Haenel, V., Fruend, I., Yianilos, P., Mueller-Freitag, M., et al.: The\" something something\" video database for learning and evaluating visual common sense. In: Proceedings of the IEEE international conference on computer vision. pp. 5842-5850 (2017)", + "2. Hou, S., Pan, X., Loy, C.C., Wang, Z., Lin, D.: Learning a unified classifier incrementally via rebalancing. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 831-839 (2019)" + ], + "bbox": [ + 225, + 258, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Spatial-Temporal Subspace Projection for VCIL", + "bbox": [ + 408, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "13. Huang, H., Wang, Y., Hu, Q., Cheng, M.M.: Class-specific semantic reconstruction for open set recognition. IEEE transactions on pattern analysis and machine intelligence 45(4), 4214-4228 (2023)", + "14. Ji, S., Xu, W., Yang, M., Yu, K.: 3d convolutional neural networks for human action recognition. IEEE transactions on pattern analysis and machine intelligence 35(1), 221-231 (2012)", + "15. Kuehne, H., Jhuang, H., Garrote, E., Poggio, T., Serre, T.: Hmdb: a large video database for human motion recognition. In: 2011 International conference on computer vision. pp. 2556-2563. IEEE (2011)", + "16. Li, Z., Hoiem, D.: Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence 40(12), 2935-2947 (2017)", + "17. Lin, J., Gan, C., Han, S.: Temporal shift module for efficient video understanding. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7083-7093 (2019)", + "18. Liu, C.: Gabor-based kernel pca with fractional power polynomial models for face recognition. IEEE transactions on pattern analysis and machine intelligence 26(5), 572-581 (2004)", + "19. Masana, M., Liu, X., Twardowski, B., Menta, M., Bagdanov, A.D., Van De Weijer, J.: Class-incremental learning: survey and performance evaluation on image classification. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(5), 5513-5533 (2022)", + "20. McCloskey, M., Cohen, N.J.: Catastrophic interference in connectionist networks: The sequential learning problem. In: Psychology of learning and motivation, vol. 24, pp. 109-165. Elsevier (1989)", + "21. Papyan, V., Han, X., Donoho, D.L.: Prevalence of neural collapse during the terminal phase of deep learning training. Proceedings of the National Academy of Sciences 117(40), 24652-24663 (2020)", + "22. Park, J., Kang, M., Han, B.: Class-incremental learning for action recognition in videos. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 13698-13707 (2021)", + "23. Pei, Y., Qing, Z., Cen, J., Wang, X., Zhang, S., Wang, Y., Tang, M., Sang, N., Qian, X.: Learning a condensed frame for memory-efficient video class-incremental learning. Advances in Neural Information Processing Systems 35, 31002-31016 (2022)", + "24. Pei, Y., Qing, Z., Zhang, S., Wang, X., Zhang, Y., Zhao, D., Qian, X.: Space-time prompting for video class-incremental learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11932-11942 (2023)", + "25. Peifeng, G., Xu, Q., Wen, P., Yang, Z., Shao, H., Huang, Q.: Feature directions matter: Long-tailed learning via rotated balanced representation. In: International Conference on Machine Learning. pp. 27542-27563. PMLR (2023)", + "26. Rebuffi, S.A., Kolesnikov, A., Sperl, G., Lampert, C.H.: icarl: Incremental classifier and representation learning. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 2001-2010 (2017)", + "27. Saha, G., Garg, I., Roy, K.: Gradient projection memory for continual learning. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=3A0jORCNC2", + "28. Simon, C., Koniusz, P., Nock, R., Harandi, M.: Adaptive subspaces for few-shot learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4136-4145 (2020)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "H. Cheng et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "29. Simonyan, K., Zisserman, A.: Two-stream convolutional networks for action recognition in videos. In: Advances in neural information processing systems. pp. 568-576 (2014)", + "30. Soomro, K., Zamir, A.R., Shah, M.: Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402 (2012)", + "31. Tran, D., Bourdev, L., Fergus, R., Torresani, L., Paluri, M.: Learning spatiotemporal features with 3d convolutional networks. In: Proceedings of the IEEE international conference on computer vision. pp. 4489-4497 (2015)", + "32. Tran, D., Wang, H., Torresani, L., Ray, J., LeCun, Y., Paluri, M.: A closer look at spatiotemporal convolutions for action recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6450-6459 (2018)", + "33. Turk, M.A., Pentland, A.P.: Face recognition using eigenfaces. In: Proceedings. 1991 IEEE computer society conference on computer vision and pattern recognition. pp. 586-587. IEEE Computer Society (1991)", + "34. Tzimiropoulos, G., Zafeiriou, S., Pantic, M.: Subspace learning from image gradient orientations. IEEE transactions on pattern analysis and machine intelligence 34(12), 2454-2466 (2012)", + "35. Villa, A., Alcázar, J.L., Alfarra, M., Alhamoud, K., Hurtado, J., Heilbron, F.C., Soto, A., Ghanem, B.: Pivot: Prompting for video continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24214-24223 (2023)", + "36. Villa, A., Alhamoud, K., Escorcia, V., Caba, F., Alcazar, J.L., Ghanem, B.: vclimb: A novel video class incremental learning benchmark. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19035-19044 (2022)", + "37. Wang, C., Guo, L., Wang, Y., Cheng, H., Yu, Y., Wen, B.: Progressive divide-and-conquer via subsampling decomposition for accelerated mri. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 25128-25137 (2024)", + "38. Wang, J., Cherian, A.: Gods: Generalized one-class discriminative subspaces for anomaly detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8201-8211 (2019)", + "39. Wang, S., Li, X., Sun, J., Xu, Z.: Training networks in null space of feature covariance for continual learning. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 184-193 (2021)", + "40. Yang, J., Frangi, A.F., Yang, J.y., Zhang, D., Jin, Z.: Kpca plus lda: a complete kernel fisher discriminant framework for feature extraction and recognition. IEEE Transactions on pattern analysis and machine intelligence 27(2), 230-244 (2005)", + "41. Yang, S., Liu, J., Lu, S., Er, M.H., Kot, A.C.: Collaborative learning of gesture recognition and 3d hand pose estimation with multi-order feature analysis. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16. pp. 769-786. Springer (2020)", + "42. Yang, Y., Yuan, H., Li, X., Lin, Z., Torr, P., Tao, D.: Neural collapse inspired feature-classifier alignment for few-shot class-incremental learning. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=y5W8tpoJhtJ", + "43. Zhang, L., Fu, J., Wang, S., Zhang, D., Dong, Z., Chen, C.P.: Guide subspace learning for unsupervised domain adaptation. IEEE transactions on neural networks and learning systems 31(9), 3374-3388 (2019)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Spatial-Temporal Subspace Projection for VCIL", + "bbox": [ + 408, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "44. Zhao, B., Xiao, X., Gan, G., Zhang, B., Xia, S.T.: Maintaining discrimination and fairness in class incremental learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 13208-13217 (2020)", + "45. Zhu, H., Koniusz, P.: Ease: Unsupervised discriminant subspace learning for transductive few-shot learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9078-9088 (2022)", + "46. Zhu, P., Yao, X., Wang, Y., Hui, B., Du, D., Hu, Q.: Multiview deep subspace clustering networks. IEEE Transactions on Cybernetics (2024)" + ], + "bbox": [ + 215, + 146, + 787, + 258 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "H. Cheng et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/STSP_ Spatial-Temporal Subspace Projection for Video Class-incremental Learning/9413ef29-ccb6-4b51-bb25-c4b8dee0d763_model.json b/2024/STSP_ Spatial-Temporal Subspace Projection for Video Class-incremental Learning/9413ef29-ccb6-4b51-bb25-c4b8dee0d763_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1e9bdad1ed1c13449204fc1eba6475493c673862 --- /dev/null +++ b/2024/STSP_ Spatial-Temporal Subspace Projection for Video Class-incremental Learning/9413ef29-ccb6-4b51-bb25-c4b8dee0d763_model.json @@ -0,0 +1,2348 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.223, + 0.141, + 0.78, + 0.187 + ], + "angle": 0, + "content": "STSP: Spatial-Temporal Subspace Projection for Video Class-incremental Learning" + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.212, + 0.782, + 0.245 + ], + "angle": 0, + "content": "Hao Cheng\\(^{1,2}\\), Siyuan Yang\\(^{1}\\), Chong Wang\\(^{1}\\), Joey Tianyi Zhou\\(^{3,4}\\), Alex C. Kot\\(^{1}\\), and Bihan Wen\\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.307, + 0.255, + 0.696, + 0.283 + ], + "angle": 0, + "content": "1Nanyang Technological University, Singapore {siyuan.yang, wang1711, eackot, bihan.wen}@ntu.edu.sg" + }, + { + "type": "text", + "bbox": [ + 0.336, + 0.283, + 0.666, + 0.309 + ], + "angle": 0, + "content": "2 Hebei University of Technology, Tianjin, China 2024025@hebut.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.28, + 0.31, + 0.722, + 0.325 + ], + "angle": 0, + "content": "\\(^{3}\\) CFAR, Agency for Science, Technology and Research, Singapore" + }, + { + "type": "text", + "bbox": [ + 0.282, + 0.325, + 0.721, + 0.353 + ], + "angle": 0, + "content": "4 IHPC, Agency for Science, Technology and Research, Singapore zhouty@cfar.a-star.edu.sg" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.38, + 0.744, + 0.685 + ], + "angle": 0, + "content": "Abstract. Video class-incremental learning (VCIL) aims to learn discriminative and generalized feature representations for video frames to mitigate catastrophic forgetting. Conventional VCIL methods often retain a subset of frames or features from prior tasks as exemplars for subsequent incremental learning stages. However, these strategies overlook the connection between base and novel classes, sometimes even leading to privacy leakage. To address this challenge, we introduce a Spatial-Temporal Subspace Projection (STSP) scheme for VCIL. Specifically, we propose a discriminative Temporal-based Subspace Classifier (TSC) that represents each class with an orthogonal subspace basis and adopts subspace projection loss for classification. Unlike typical classification methods relying on fully connected layers, our TSC discerns the spatial-temporal dynamics in video content, thereby enhancing the representation of each video sample. Additionally, we implement inter- and intra-class orthogonal constraints into TSC, ensuring each class occupies a unique orthogonal subspace, defined by its basis. To prevent catastrophic forgetting, we further employ a Spatial-based Gradient Projection (SGP) strategy, adjusting network gradients to align with the null space of the spatial feature set from previous tasks. Extensive experiments conducted on three benchmarks, namely HMDB51, UCF101, and Something-Something V2, demonstrate that our STSP method outperforms state-of-the-art comparison methods, evidencing its efficacy in VCIL." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.697, + 0.741, + 0.725 + ], + "angle": 0, + "content": "Keywords: Video action recognition \\(\\cdot\\) Class incremental learning \\(\\cdot\\) Subspace Projection \\(\\cdot\\) Gradient Projection" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.746, + 0.376, + 0.763 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.774, + 0.788, + 0.821 + ], + "angle": 0, + "content": "Deep neural networks have demonstrated remarkable performance in modern computer vision tasks [7,17,32,37,46], particularly when trained on large, annotated datasets. However, the real-world challenge of incrementally incorporating" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.826, + 0.386, + 0.841 + ], + "angle": 0, + "content": "* Corresponding author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "H. Cheng et al." + }, + { + "type": "image", + "bbox": [ + 0.239, + 0.146, + 0.495, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.147, + 0.77, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.254, + 0.788, + 0.325 + ], + "angle": 0, + "content": "Fig. 1: Class incremental learning with (a) linear classifier and (b) subspace-based classifier. Incorporating new classes alters the decision hyperspace and modifies the representation of prior classes, leading to misalignment. In contrast, subspace-based classifiers, by constructing an orthogonal basis for each class, establish a unique geometric structure for each class, thereby reducing class misalignment." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.355, + 0.788, + 0.461 + ], + "angle": 0, + "content": "data from new classes into a model initially trained on existing classes presents significant hurdles, known as class-incremental learning (CIL) [26]. Unlike typical classification tasks, CIL involves accessing novel classes and limited memories of prior tasks in incremental sessions, often due to privacy concerns or memory constraints. Consequently, merely updating the model with data from new classes can result in overfitting substantial performance degradation on previous classes, a phenomenon referred to as catastrophic forgetting [19, 20]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.788, + 0.794 + ], + "angle": 0, + "content": "Recent advancements in CIL methods [1,4,26,44] have shown remarkable performance in image classification tasks. However, recent CIL studies [1,21,25,42] demonstrate that straightforward fine-tuning of the network for novel classes may lead the subspace spanned by the novel classifier to deviate from the base one, as illustrated in Fig. 1 (a). The classification performance can be significantly affected by misalignment, particularly in cases where a notable distribution gap between novel and base classes. Moreover, when extending this task to Video Class-Incremental Learning (VCIL), unique challenges due to the inherent information redundancy within video samples, such as continuous actions with a static background. This necessitates methods to model temporal-spatial relationships between video frames to extract critical information for action recognition properly. However, prevalent approaches [22-24,35,36] in VCIL predominantly involve a learning pipeline that employs simple operations (e.g., mean) on the feature vectors of sampled frames to generate video representations for action recognition. While this methodology is effective for standard action recognition with sufficient data, it may struggle to ensure robust generalization in scenarios of incremental learning. The aforementioned observations prompt us to consider acquiring a more suitable representation for both the classifier and video features while also exploring their relationship. To address these challenges, several subspace learning-based methods [28,33,38,45] have shown effectiveness in reducing dimensionality and capturing intrinsic structure, thereby offering a potential solution to the issues associated with simplistic video representation methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.842 + ], + "angle": 0, + "content": "In this paper, we propose a spatial-temporal subspace projection method for VCIL including a temporal-based subspace classifier (TSC) and a spatial-based gradient projection (SGP) strategy. Specifically, we extend the existing linear" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.41, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Spatial-Temporal Subspace Projection for VCIL" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.373 + ], + "angle": 0, + "content": "classifiers based on fully connected (FC) layers to a TSC in which a subspace basis is adopted to model each class. Differing from linear classifiers that seek to distinguish classes in feature space through a linear decision boundary, the proposed orthogonality constraints offer a distinct geometric structure for each class. This characteristic not only facilitates a more comprehensive representation of each class but also allows for straightforward extension to novel classes. Furthermore, to avoid catastrophic forgetting, we utilize the SGP strategy to constrain the process of gradient update by projecting the gradient of network parameters in each layer onto the subspace of corresponding input features so that losses for the old task will not increase during the incremental sessions. Compared with existing VCIL methods [22-24,35,36] that rely on feature distillation or prompt-based fine-tuning strategies, SGP offers the advantage of fine-tuning all network parameters without necessitating regularizers. Furthermore, SGP obviates the need to explicitly retain class-relevant information, such as frames or prototypes, thereby contributing to privacy protection." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.374, + 0.788, + 0.51 + ], + "angle": 0, + "content": "In summary, we make the following contributions: 1) We propose a spatial-temporal subspace projection (STSP) method, containing a temporal-based subspace classifier and spatial-based gradient projection strategy to better model feature and class representation for discriminative classification. 2) We employ a spatial-based gradient projection strategy to constrain the gradient update process without the need to store frames or prototypes from previous tasks. This not only enhances performance but also mitigates the risk of privacy leakage. 3) The proposed STSP method achieves superior results on three video action recognition incremental benchmarks over state-of-the-art methods." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.531, + 0.388, + 0.546 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.561, + 0.476, + 0.575 + ], + "angle": 0, + "content": "2.1 Video Action Recognition" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.788, + 0.841 + ], + "angle": 0, + "content": "With significant advances in deep learning, current video action recognition approaches aim to learn effective spatial-temporal representation by well-designed network architectures. A notable advancement by Simonyan and Zisserman [29] introduces a dual-stream architecture that analyzes RGB frames in one stream and optical flow in another, laying the foundation for subsequent enhancements. Recently, Lin et al. [17] proposed a Temporal Shift Module (TSM), which shifts some channels along the temporal dimension to perform temporal interaction between the features from adjacent frames. Concurrently, significant research [3, 13, 14, 31, 32, 41] has evolved 2D Convolutional Neural Networks (CNNs) into 3D structures to capture both spatial and temporal contexts within videos, a critical aspect of human action recognition. Notably, Tran et al. [31] introduced the C3D model, a 3D CNN designed to learn spatial-temporal features through an end-to-end learning approach. Carreira et al. [3] developed the Inflated 3D CNN, which extends a 2D CNN into the temporal domain by inflating its convolutional and pooling layers. Additionally, Feichtenhofer et al. [9] crafted a dual-pathway 3D CNN, featuring both slow and fast pathways that process RGB frames at varying frame rates to simultaneously capture semantic content and motion details." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "H. Cheng et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.209 + ], + "angle": 0, + "content": "Despite significant progress in action recognition, catastrophic forgetting has received relatively limited attention. This work aims to address this gap by developing a spatial-temporal subspace projection method, which aims to offer a robust solution to the challenges of incremental learning in action recognition." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.229, + 0.49, + 0.244 + ], + "angle": 0, + "content": "2.2 Class-Incremental Learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.252, + 0.788, + 0.403 + ], + "angle": 0, + "content": "CIL addresses the challenge of adapting machine learning models to recognize new classes without forgetting previously learned information, a phenomenon known as catastrophic forgetting. Pioneering in this domain, Rebuffi et al. [26] introduced iCaRL, which combines representation learning with a memory mechanism to selectively retain the knowledge of previous classes. Following this, Hou et al. [12] proposed a UCIR strategy, which addresses the class imbalance in CIL by adjusting the classification layer to focus more on newer classes while still remembering old ones. Moreover, Gao et al. [10] introduced R-DFCIL, employing relation-guided representation learning to overcome catastrophic forgetting and enable the assimilation of new classes in a data-free scenario." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.403, + 0.788, + 0.524 + ], + "angle": 0, + "content": "While CIL has been well-studied in the image domain, its application to the video domain remains largely unexplored. Park et al. [22] introduced Time-Channel Distillation (TCD), leveraging knowledge distillation loss with time-channel importance masks. Pei et al. [23] proposed FrameMaker, a memory-efficient VCIL strategy that develops a condensed frame for each video to facilitate future incremental tasks with less memory usage. Pei et al. [24] introduced ST-Prompt, a novel space-time prompting framework tailored for VCIL, enhancing the model's adaptability to new classes with minimal memory increase." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.524, + 0.788, + 0.615 + ], + "angle": 0, + "content": "Existing VCIL approaches often rely on preserving a subset of frames or features from previous tasks as exemplars for subsequent incremental stages to prevent catastrophic forgetting. However, these strategies overlook the connection between base and novel classes and potentially lead to privacy leakage. In contrast, we employ a spatial-based gradient projection to regulate gradient updates, thereby mitigating catastrophic forgetting without these drawbacks." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.635, + 0.471, + 0.65 + ], + "angle": 0, + "content": "2.3 Subspace Representation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Subspace learning, by learning a suitable representation for high dimensional data, has shed light on various vision tasks [2, 18, 34, 40, 43]. Existing subspace methods such as Principal Component Analysis (PCA) [2, 18] and Linear Discriminant Analysis (LDA) [2, 40] adaptively transfer the original input data to low dimensional subspace for dimensionality reduction. For example, Tzimiropoulos et al. [34] formulated PCA on image gradient orientations, which demonstrated robust performance in face recognition. Recently, subspace learning has been further extended to regularize the implicit features or learned classifiers within deep networks [1,28,38,45]. Notably, Cheraghian et al. [5] constructed a mixture of subspaces to represent the feature clusters formed by visually similar samples. Such subspace representation enables a better approximation of the overall data distribution, consequently leading to superior performance in" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.41, + 0.114, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Spatial-Temporal Subspace Projection for VCIL" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.23, + 0.154, + 0.768, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.268, + 0.788, + 0.396 + ], + "angle": 0, + "content": "Fig. 2: The overall framework of the proposed STSP method, including a temporal-based subspace classifier (TSC) and a spatial-based gradient projection (SGP). For each input video, spatial-temporal features are extracted using a backbone network. Subsequently, The TSC classifies by projecting temporal features into a subspace, where classification decisions are based on the distance to each class's subspace basis within this subspace. During parameter updates for the novel tasks, the SGP constrains gradient modifications for each specific layer, drawing on the mean covariance matrix \\( M_{t-1} \\) of previous tasks' input features. Upon the current session completion, the mean covariance matrix is updated to \\( M_t \\) for the gradient projection in the subsequent session." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.429, + 0.784, + 0.505 + ], + "angle": 0, + "content": "CIL. To mitigate the catastrophic forgetting and overfitting, Simon et al. [28] formulated the dynamic classifiers via subspace modeling which demonstrated a significant robustness to noise in few-shot learning. Similarly, Akyurek et al. [1] proposed an additional subspace regularizer that pulls the weight vectors of the novel class close to the subspace spanned by the base classes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.507, + 0.785, + 0.568 + ], + "angle": 0, + "content": "However, existing methods lack sufficient constraints on the representation of each class in the classifier. In this paper, we model each class via an orthogonal subspace basis and leverage spatial-based gradient projection to update the entire network during incremental sessions." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.606, + 0.33, + 0.621 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.652, + 0.784, + 0.698 + ], + "angle": 0, + "content": "In this section, we start with an introduction of VCIL problem formulation. Then, the proposed STSP method is described and explained in detail, followed by the loss function of our method as well as the training and inference procedure." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.701, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In VCIL, a series of distinct tasks \\(\\{\\mathcal{T}_1,\\mathcal{T}_2,\\dots ,\\mathcal{T}_k,\\dots \\}\\) unfolds sequentially. Each task \\(\\mathcal{T}_k\\) comprises a distinct subset of video-class pairs \\(\\{(D_i^k,y_i^k),y_i^k\\in\\) \\(\\mathcal{C}_k\\} |_{i = 1}^{|\\mathcal{T}_k|}\\). Specifically, the novel class set \\(\\mathcal{C}_k\\) of \\(\\mathcal{T}_k\\) is distinct from previous classes, i.e., \\((\\mathcal{C}_1\\cup \\dots \\cup \\mathcal{C}_{k - 1})\\cap \\mathcal{C}_k = \\emptyset\\). The primary objective is to incrementally train a deep neural network \\(f(\\cdot)\\) capable of accurately classifying videos across all classes by preventing the loss of class information from previous tasks. In contrast to methods that have access to a limited exemplar set from historical tasks, our emphasis lies on exemplar-free VCIL, where the model is not permitted to store exemplars from previous tasks during the learning process of the current session." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.128 + ], + "angle": 0, + "content": "H. Cheng et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.474, + 0.16 + ], + "angle": 0, + "content": "3.1 Overall STSP Framework" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.169, + 0.788, + 0.351 + ], + "angle": 0, + "content": "Given the problem formulation, we follow the standard learning protocol of previous methods [22, 23] based on knowledge distillation. During the \\(k^{th}\\) incremental session, the network \\(f(\\cdot ;\\mathbf{w})\\) parameterized by \\(\\mathbf{w}\\) is updated from \\(\\mathbf{w}_{k - 1}\\) to adapt to the novel class set \\(\\mathcal{C}_k\\). Fig. 2 illustrates the overall framework of our method. For an input video \\(D_{i}^{k}\\) belonging to \\(\\mathcal{C}_k\\), we first extract \\(T\\) frames as \\(V_{i}\\), which are then fed to the backbone network to generate the feature embeddings. Subsequently, we employ a TSC to differentiate between distinct classes. Moreover, our method stands out by employing an SGP strategy without restoring frames or prototypes from previous tasks. More in detail, the network parameters are optimized by projecting the gradients in each layer to the subspace corresponding to their input features. With the merit of this strategy, we can completely avoid the risk of privacy leakage, whereas existing approaches may suffer." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.37, + 0.557, + 0.385 + ], + "angle": 0, + "content": "3.2 Temporal-based Subspace Classifier" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.391, + 0.789, + 0.542 + ], + "angle": 0, + "content": "In accordance with the standardized procedure for VCIL, given an input video-class pair denoted as \\((D_i,y_i)\\), we first randomly sampled \\(T\\) frames \\(V_{i}^{T}\\). Subsequently, these frames are fed to the backbone network, yielding the feature embeddings as \\(X_{i}\\in \\mathbb{R}^{T\\times C\\times H\\times W}\\), where \\(C\\), \\(H\\), and \\(W\\) represent the size of channel, height, and width, respectively. For the existing architectures [22-24] for VCIL, a prevalent technique for obtaining video representations is global average pooling (GAP), denoted as \\(Z_{i} = GAP(X_{i})\\in \\mathbb{R}^{C\\times T}\\). Following this, a linear classifier \\(L\\), comprising an FC layer with weights \\(\\mathbf{w}\\in \\mathbb{R}^{C\\times K}\\), is utilized to predict the potential class of each frame. The class predictions from individual frames are subsequently averaged to produce the overall classification result." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.543, + 0.789, + 0.678 + ], + "angle": 0, + "content": "Nevertheless, the temporal information holds the potential to encapsulate more intricate intra-frame structural details, warranting thoughtful consideration and seamless integration into the classification process. Additionally, in each incremental task, adjustments to the linear classifier may be necessary to modify the overall representation across all classes, which could influence the base classes for all previous incremental sessions. Conversely, a subspace-based classifier demonstrates heightened adaptability, enabling the modeling of each new class within its unique subspace without imposing substantial effects on the existing representations of previous classes." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.679, + 0.789, + 0.814 + ], + "angle": 0, + "content": "To this end, to properly model each class and learn a robust classifier, we construct a TSC based on subspaces spanned by an orthogonal basis of each class. Specifically, for \\( m \\)-th class, we construct a subspace with orthogonal basis \\( P_{m} \\in \\mathbb{R}^{C \\times \\tau} \\) with \\( P_{m}^{\\top} P_{m} = I_{\\tau}, \\tau \\ll C \\), where \\( C \\) and \\( \\tau \\) indicate the feature dimension and the number of subspace basis for \\( m \\)-th class, respectively. Hence, for each task \\( \\mathcal{T}_k \\) in the \\( k^{th} \\) incremental session, we construct \\( |\\mathcal{C}_k| \\) distinct subspaces for \\( |\\mathcal{C}_k| \\) novel classes. We then adopt current subspaces to predict the class of \\( Z_{i} \\) based on the subspace distance between the subspace spanned by temporal video features \\( Z_{i} \\) and the orthogonal basis of each class as:" + }, + { + "type": "equation", + "bbox": [ + 0.387, + 0.822, + 0.785, + 0.843 + ], + "angle": 0, + "content": "\\[\nd _ {m} (Z _ {i}) = - \\left\\| Z _ {i} Z _ {i} ^ {\\top} - P _ {m} P _ {m} ^ {\\top} \\right\\| _ {F} ^ {2}. \\tag {1}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.41, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Spatial-Temporal Subspace Projection for VCIL" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.128 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.786, + 0.176 + ], + "angle": 0, + "content": "Then, the probability of \\( Z_{i} \\) assigned to \\( m \\)-th class can be calculated using the softmax function as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.407, + 0.186, + 0.786, + 0.224 + ], + "angle": 0, + "content": "\\[\np _ {m} (Z _ {i}) = \\frac {e ^ {d _ {m} (Z _ {i})}}{\\sum_ {m ^ {\\prime} = 1} ^ {M} e ^ {d _ {m ^ {\\prime}} (Z _ {i})}}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.232, + 0.784, + 0.263 + ], + "angle": 0, + "content": "where \\( M \\) is the total number of classes in the current session. Now, we can employ the cross-entropy loss on Eq. (2) to update the network as:" + }, + { + "type": "equation", + "bbox": [ + 0.411, + 0.275, + 0.786, + 0.315 + ], + "angle": 0, + "content": "\\[\nL _ {C E} = - \\sum_ {i = 1} ^ {J} y _ {i} \\log p _ {i} \\left(Z _ {i}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.325, + 0.634, + 0.339 + ], + "angle": 0, + "content": "where \\( J \\) denotes the number of videos in each mini-batch." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.34, + 0.784, + 0.415 + ], + "angle": 0, + "content": "To encourage the network to capture discriminative information for each class and construct distinct subspaces corresponding to different classes, we impose orthogonal constraints on both the temporal features of the video and the weights of the TSC. Specifically, we employ orthogonal loss on the temporal features \\( Z_{i} \\) of each video \\( D_{i} \\) in the final layer as:" + }, + { + "type": "equation", + "bbox": [ + 0.406, + 0.429, + 0.786, + 0.468 + ], + "angle": 0, + "content": "\\[\nL _ {O F} = \\sum_ {i = 1} ^ {J} \\left\\| Z _ {i} ^ {\\top} Z _ {i} - I _ {T} \\right\\| _ {F} ^ {2}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.474, + 0.784, + 0.535 + ], + "angle": 0, + "content": "where \\(I_T \\in \\mathbb{R}^{T \\times T}\\) is an identity matrix. This approach guarantees that every frame independently captures distinct, class-aware information, setting it apart uniquely from the rest. Furthermore, it facilitates the construction of a discriminative subspace for each corresponding class." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.536, + 0.784, + 0.566 + ], + "angle": 0, + "content": "For TSC, we impose intra- and inter-class orthogonal constraints on the subspace basis of each class, which is denoted by \\( L_{OW} \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.327, + 0.578, + 0.786, + 0.626 + ], + "angle": 0, + "content": "\\[\nL _ {O W} = \\sum_ {m = 1} ^ {M} \\left(\\underbrace {\\left\\| P _ {m} ^ {\\top} P _ {m} - I _ {\\tau} \\right\\| _ {F} ^ {2}} _ {\\text {i n t r a - c l a s s}} + \\sum_ {q = m + 1} ^ {M} \\underbrace {\\left\\| P _ {m} ^ {\\top} P _ {q} \\right\\| _ {F} ^ {2}} _ {\\text {i n t e r - c l a s s}}\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.635, + 0.784, + 0.77 + ], + "angle": 0, + "content": "The intra-class orthogonal constraint is introduced to ensure independence among the basis vectors of subspaces, which is crucial for learning distinct and discriminative feature representations. Furthermore, the imposition of inter-class orthogonal constraint ensures that the basis vectors of distinct classes are orthogonal, signifying their mutual independence. This orthogonality enforcement serves to prevent confusion and safeguard the uniqueness of each basis vector, thus preserving the integrity of class information previously learned. Moreover, these orthogonal constraints enhance discriminability, guaranteeing that the basis vectors of subspaces related to different classes are clearly distinguishable." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.771, + 0.691, + 0.786 + ], + "angle": 0, + "content": "Therefore, the overall objective function for TSC is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.384, + 0.799, + 0.786, + 0.814 + ], + "angle": 0, + "content": "\\[\nL _ {t o t a l} = \\alpha L _ {C E} + \\beta L _ {O F} + \\gamma L _ {O W}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.825, + 0.783, + 0.841 + ], + "angle": 0, + "content": "where \\(\\alpha\\), \\(\\beta\\), and \\(\\gamma\\) denote the weighting parameters of each losses, respectively." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.377, + 0.129 + ], + "angle": 0, + "content": "H. Cheng et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.544, + 0.163 + ], + "angle": 0, + "content": "3.3 Spatial-based Gradient Projection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.177, + 0.788, + 0.268 + ], + "angle": 0, + "content": "In contrast to prior approaches [22-24] that rely on retaining exemplars or prompts from previous classes during incremental sessions, our focus is on addressing the challenging exemplar-free VCIL problems by imposing constraints on the network during parameter updates. To avoid catastrophic forgetting on previous classes in the incremental sessions, the network's parameter update should satisfy the following proposition:" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.279, + 0.79, + 0.325 + ], + "angle": 0, + "content": "Proposition 1. Given the data \\( x_{k} \\) from the previous task \\( \\mathcal{T}_k \\), the current network \\( f(\\cdot ;\\mathbf{w}_t) \\) could preserve the historical information if it satisfies the following equation for all sessions:" + }, + { + "type": "equation", + "bbox": [ + 0.368, + 0.338, + 0.786, + 0.356 + ], + "angle": 0, + "content": "\\[\nf \\left(x _ {k}; \\mathbf {w} _ {t}\\right) = f \\left(x _ {k}; \\mathbf {w} _ {k}\\right), \\quad \\forall k < t. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.368, + 0.594, + 0.383 + ], + "angle": 0, + "content": "Here, \\(\\mathbf{w}_t\\) denotes the network parameters at task \\(T_{t}\\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.394, + 0.788, + 0.472 + ], + "angle": 0, + "content": "To realize the objective presented in Proposition 1, it is imperative to confine the gradient update of network parameters to a specific direction—precisely, the gradient update should lie in the subspace spanned by the corresponding input spatial features. Specifically, the gradients in the linear layers, including FC and Convolution layers \\(^5\\) should satisfy Proposition 2 [27,39]." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.485, + 0.79, + 0.548 + ], + "angle": 0, + "content": "Proposition 2. Let \\( f^l(\\cdot; \\mathbf{w}_t) \\) denotes the \\( l \\)-th linear layer of the network trained on current task \\( \\mathcal{T}_t \\) and its corresponding input feature \\( A_k^l \\) from any previous task \\( \\mathcal{T}_k \\) (\\( k < t \\)). During each training step \\( s \\), if the linear layer's parameter update \\( \\Delta \\mathbf{w}_{t,s}^l \\) lies in the null space of \\( A_k^l \\), i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.451, + 0.56, + 0.787, + 0.58 + ], + "angle": 0, + "content": "\\[\nA _ {k} ^ {l} \\Delta \\mathbf {w} _ {t, s} ^ {l} = 0, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.592, + 0.747, + 0.612 + ], + "angle": 0, + "content": "we could preserve the historical information as \\( f^{l}\\left(A_{k}^{l};\\mathbf{w}_{t}^{l}\\right) = f^{l}\\left(A_{k}^{l};\\mathbf{w}_{k}^{l}\\right) \\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.623, + 0.787, + 0.655 + ], + "angle": 0, + "content": "Building upon Proposition 2, we further extend it to our proposed TSC with Eq. (1), as detailed in Proposition 3." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.667, + 0.789, + 0.73 + ], + "angle": 0, + "content": "Proposition 3. Let \\( f^{TSC}(\\cdot ;\\mathbf{w}_t) \\) denotes the TSC of the network trained on current task \\( \\mathcal{T}_t \\) and its corresponding input feature \\( Z_{k} \\) from any previous task \\( \\mathcal{T}_k \\) (\\( k < t \\)). During each training step \\( s \\), if the TSC parameter update \\( \\Delta \\mathbf{w}_{t,s}^{TSC} \\) lies in the null space of the subspace spanned by \\( Z_{k} \\), i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.432, + 0.74, + 0.787, + 0.761 + ], + "angle": 0, + "content": "\\[\nZ _ {k} Z _ {k} ^ {\\top} \\Delta \\mathbf {w} _ {t, s} ^ {T S C} = 0, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.772, + 0.629, + 0.792 + ], + "angle": 0, + "content": "similarly we have \\( f^{TSC}\\left(Z_k;\\mathbf{w}_t^{TSC}\\right) = f^{TSC}\\left(Z_k;\\mathbf{w}_k^{TSC}\\right) \\)." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.801, + 0.595, + 0.817 + ], + "angle": 0, + "content": "Proof. Please refer to the supplemental material." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.768, + 0.842 + ], + "angle": 0, + "content": "5 The convolution operation can be regarded as a specific type of linear operation." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.41, + 0.114, + 0.732, + 0.131 + ], + "angle": 0, + "content": "Spatial-Temporal Subspace Projection for VCIL" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.128 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "Inspired by previous work [27,39], we calculate the mean covariance matrix, denoted as \\( M_{t-1} \\), of the input feature for each specific layer in each mini-batch to enhance network stability and optimize memory efficiency. Here, let \\( Q \\) represent the collective term for the current input feature at each specific layer, with \\( A_t^l \\) for the \\( l \\)-th linear layer and \\( Z_t Z_t^\\top \\) for the TSC, respectively. To initialize, \\( M_0 \\) is set to 0 for the first session. At the end of each session \\( t \\), \\( M_t \\) is updated as:" + }, + { + "type": "equation", + "bbox": [ + 0.407, + 0.246, + 0.786, + 0.288 + ], + "angle": 0, + "content": "\\[\nM _ {t} = M _ {t - 1} + \\sum_ {b = 1} ^ {\\mathcal {B}} \\frac {1}{J} Q _ {b} ^ {\\top} Q _ {b}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.292, + 0.785, + 0.321 + ], + "angle": 0, + "content": "where \\(\\mathcal{B}\\) represents the number of mini-batches in the current task, and \\(J\\) denotes the number of videos in each mini-batch." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.322, + 0.785, + 0.369 + ], + "angle": 0, + "content": "During the feature update in task \\(\\mathcal{T}_t\\), we perform Singular Value Decomposition (SVD) to find the approximate null space of the given feature matrix \\(M_{t-1}\\) corresponding to Proposition 2 and Proposition 3 as:" + }, + { + "type": "equation", + "bbox": [ + 0.366, + 0.375, + 0.786, + 0.408 + ], + "angle": 0, + "content": "\\[\n\\left[ U _ {1}, U _ {2} \\right] \\left[ \\begin{array}{c c} \\Sigma_ {1} & 0 \\\\ 0 & \\Sigma_ {2} \\end{array} \\right] \\left[ \\begin{array}{l} U _ {1} ^ {\\top} \\\\ U _ {2} ^ {\\top} \\end{array} \\right] = \\operatorname {S V D} \\left(M _ {t - 1}\\right), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.414, + 0.788, + 0.626 + ], + "angle": 0, + "content": "where \\( U_{1}, U_{2} \\) represent unitary matrices that comprise orthonormal bases associated with the singular values in \\( \\Sigma_{1}, \\Sigma_{2} \\), respectively, with all zero singular values relegated to \\( \\Sigma_{2} \\). However, it is unrealistic to guarantee that there always exists zero singular values. Previous work [27, 39] constructs \\( U_{2}^{\\top} \\) by selecting singular values restricted to a specific proportion. However, this strategy might not always provide the most compact representation, particularly when there is a significant difference in the scale of singular values. Hence, we employ a threshold strategy based on the energy retained in the matrix \\( M \\) for the construction \\( U_{2}^{\\top} \\), corresponding to a segment of smallest singular values that satisfy \\( \\| \\Sigma_{2} \\|_{F}^{2} \\leq (1 - \\eta) \\left( \\| \\Sigma_{1} \\|_{F}^{2} + \\| \\Sigma_{2} \\|_{F}^{2} \\right) \\), with the proportion of retained energy denoted by \\( \\eta \\). When \\( \\eta \\approx 1 \\), it is reasonable to approximate the null space of the matrix \\( M \\) with \\( U_{2}^{\\top} \\). This enables the parameter update \\( \\Delta \\mathbf{w} \\) from the gradient \\( \\mathbf{g} \\) concerning the parameters \\( \\mathbf{w} \\) through subspace projection based on \\( U_{2} \\). That is, during the training step \\( s \\) within any incremental session \\( t \\), we have:" + }, + { + "type": "equation", + "bbox": [ + 0.435, + 0.632, + 0.786, + 0.65 + ], + "angle": 0, + "content": "\\[\n\\Delta \\mathbf {w} _ {t, s} = U _ {2} U _ {2} ^ {\\top} \\mathbf {g} _ {t, s}. \\tag {12}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.668, + 0.548, + 0.684 + ], + "angle": 0, + "content": "3.4 Model Optimization and Inference" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.795 + ], + "angle": 0, + "content": "Our STSP method is summarized in Algorithm 1. In the training phase, the network parameters are updated conventionally during the initial session. After this initial session, considering parameter updates, the mean covariance of input features from the base classes is retained for each layer. These retained features are then utilized to update parameters according to Eq. (12) at each step in every subsequent incremental session. Additionally, after each incremental session, these parameters are further updated with features extracted from novel classes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.84 + ], + "angle": 0, + "content": "For the testing phase, no additional memories or operations are necessary. We simply feed the video into the network and obtain classification results using the trained model." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "H. Cheng et al." + }, + { + "type": "code_caption", + "bbox": [ + 0.218, + 0.146, + 0.582, + 0.163 + ], + "angle": 0, + "content": "Algorithm 1 STSP for class incremental learning" + }, + { + "type": "algorithm", + "bbox": [ + 0.217, + 0.165, + 0.787, + 0.486 + ], + "angle": 0, + "content": "Input: Task series \\(\\{\\mathcal{T}_1,\\mathcal{T}_2,\\dots \\}\\) , network \\(f(\\cdot ;\\mathbf{w})\\) with parameters w for linear layers and TSC, learning rate \\(\\alpha\\) and energy threshold \\(\\eta\\) \nInitialize: \\(M_0 = 0\\) for each linear layer and TSC. \n1: for \\(\\mathcal{T}_t\\in \\{\\mathcal{T}_1,\\mathcal{T}_2,\\dots \\}\\) do \n2: \\(s = 0\\) \n3: while not converge do \n4: Sample a mini-batch \\(\\{D^t,y^t\\} \\in \\mathcal{T}_t\\) with videos \\(\\{D_i^t,y_i^t\\}_{j = 1}^J\\) \n5: Compute \\(f(\\{D^{t},y^{t}\\} ;\\mathbf{w}_{t,s})\\) and obtain the gradient \\(\\mathbf{g}_{t,s}\\) \n6: if \\(t > 1\\) then \n7: for \\(\\mathbf{w}_{t,s}\\) in linear layers and TSC do \n8: Compute \\(U_{2}\\) using \\(M_{t - 1}\\) via Eq. (11) with a threshold \\(\\eta\\) \n9: Update \\(\\Delta \\mathbf{w}_{t,s}\\) using Eq. (12) \n10: Update \\(\\mathbf{w}_{t,s + 1} = \\mathbf{w}_{t,s} - \\alpha \\Delta \\mathbf{w}_{t,s}\\) \n11: end for \n12: else \n13: Update \\(\\Delta \\mathbf{w}_{t,s} = \\mathbf{g}_{t,s}\\) \n14: Update \\(\\mathbf{w}_{t,s + 1} = \\mathbf{w}_{t,s} - \\alpha \\Delta \\mathbf{w}_{t,s}\\) \n15: end if \n16: \\(s = s + 1\\) \n17: end while \n18: Get current input features \\(Q\\) at each linear layer and TSC \n19: Update \\(M_t\\) using Eq. (10) for later incremental session \n20: end for" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.512, + 0.376, + 0.53 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.542, + 0.785, + 0.587 + ], + "angle": 0, + "content": "In this section, we compare our results against state-of-the-art techniques in class-incremental action recognition benchmarks. Additionally, we perform ablation studies to demonstrate the efficacy of our proposed method." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.607, + 0.427, + 0.623 + ], + "angle": 0, + "content": "4.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.81 + ], + "angle": 0, + "content": "Datasets. The evaluation of our proposed framework is conducted using the UCF101 [30], HMDB51 [15], and Something-Something V2 [11] datasets, which are recognized benchmarks in video action recognition. The UCF-101 dataset comprises 13,320 instances across 101 action categories, with the data divided into three splits for training (70%) and testing (30%). Similarly, the HMDB-51 dataset features 6,766 video clips distributed among 51 action categories, also with three splits for training and testing in the same proportions. For our evaluations, we selected the first split from both datasets. Distinct from UCF101 and HMDB51, the Something-Something V2 dataset, a crowd-sourced collection, includes 220,000 videos across 174 classes. This dataset uniquely maintains consistent objects and backgrounds across its action categories, thereby emphasizing the importance of a model's capability to discern subtle motion cues." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.81, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Evaluation Protocol. Our methodology adheres to the class-incremental learning benchmarks frequently utilized in video analysis, as outlined in TCD [22]." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.41, + 0.114, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Spatial-Temporal Subspace Projection for VCIL" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.144, + 0.785, + 0.174 + ], + "angle": 0, + "content": "Table 1: Video class incremental learning results on UCF101 and HMDB51. The best result for each setting is emphasized in bold." + }, + { + "type": "table", + "bbox": [ + 0.246, + 0.185, + 0.753, + 0.317 + ], + "angle": 0, + "content": "
MethodUCF101HMDB51
10 × 5 stages5 × 10 stages2 × 25 stages5 × 5 stages1 × 25 stages
LwFMC [16]42.1425.5911.6826.8216.49
LwM [6]43.3926.0712.0826.9716.50
UCIR [12]74.3170.4263.2244.9037.04
PODNet [8]73.2671.5870.2844.3238.76
TCD [22]74.8973.4372.1945.3440.47
FrameMaker [23]78.1376.3875.7747.5442.65
STSP (Ours)81.1582.8479.2556.9949.19
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.351, + 0.788, + 0.488 + ], + "angle": 0, + "content": "This approach begins with a model pre-trained on half of the total number of classes, with the remaining classes introduced sequentially at each incremental step. Specifically, for UCF101, the initial training involves 51 classes, after which the subsequent 50 classes are structured into incremental learning tasks of 5, 10, and 25 classes. In the case of HMDB51, the model begins with training on 26 classes, followed by the distribution of the remaining classes into groups of 5 and 25 for subsequent learning phases. For Something-Something V2, the initial training phase covers 84 classes, with the remaining classes organized into incremental learning groups of 10 and 5 classes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.49, + 0.788, + 0.567 + ], + "angle": 0, + "content": "Implementation Details. Our framework leverages the official implementation of TSM [17], utilizing the PyTorch Library for development. We train a ResNet-34 TSM model on UCF101 for 50 epochs with a batch size of 32. For both HMDB51 and Something-Something V2, ResNet-50 TSM models are employed, undergoing training for 50 epochs with a batch size of 32." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.604, + 0.373, + 0.618 + ], + "angle": 0, + "content": "4.2 Main Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.841 + ], + "angle": 0, + "content": "For a fair comparison, we employ the same model structure with the same pretraining initialization as the existing methods [22, 23]. Unlike prior approaches that incorporate exemplars for incremental learning and assess performance using both the standard classification protocol and Nearest Mean of Exemplars (NME) metrics [26], our method does not retain any exemplars. Consequently, we only report the standard classification results. We compare our proposed STSP method with the existing video-class incremental learning methods over three benchmarks, as shown in Tab. 1 and Tab. 2. From both tables, we observe that the proposed STSP method significantly outperforms all competing methods across all settings on each dataset. Notably, unlike other approaches that rely on retaining exemplars for incremental learning sessions, our method utilizes gradient projection to constrain model updates, effectively safeguarding against performance drops on previously learned classes." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.377, + 0.129 + ], + "angle": 0, + "content": "H. Cheng et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.471, + 0.201 + ], + "angle": 0, + "content": "Table 2: Video class incremental learning results on Something-Something V2. The best result for each setting is emphasized in bold." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.212, + 0.468, + 0.297 + ], + "angle": 0, + "content": "
Method10 × 9 stages5 × 18 stages
UCIR [12]26.8420.69
PODNet [8]34.9726.95
TCD [22]35.7829.60
FrameMaker [23]37.2530.98
STSP (Ours)69.6870.87
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.145, + 0.788, + 0.188 + ], + "angle": 0, + "content": "Table 3: Ablation studies for our proposed STSP method on UCF101 with \\(5 \\times 10\\) stages and \\(2 \\times 25\\) stages settings, respectively." + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.199, + 0.785, + 0.3 + ], + "angle": 0, + "content": "
TSCSGPExemplarsUCF101
5 × 10 stages2 × 25 stages
(i)××5/class68.2857.39
(ii)✓×5/class71.5061.78
(iii)××017.6319.22
(iv)✓×08.7814.99
(v)×✓077.6674.02
(vi)✓✓082.8479.25
" + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.314, + 0.499, + 0.451 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.277, + 0.453, + 0.438, + 0.465 + ], + "angle": 0, + "content": "(a) UCF101 (5 × 10 stages)" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.314, + 0.789, + 0.451 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.565, + 0.453, + 0.726, + 0.465 + ], + "angle": 0, + "content": "(b) UCF101 (2 × 25 stages)" + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.476, + 0.785, + 0.491 + ], + "angle": 0, + "content": "Fig. 3: Plots of accuracy over seen classes along with the incremental steps on UCF101." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.52, + 0.502, + 0.537 + ], + "angle": 0, + "content": "4.3 Ablation Study and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.546, + 0.788, + 0.593 + ], + "angle": 0, + "content": "In this section, we perform ablation studies to investigate the effectiveness of the key components of the proposed STSP on the UCF101 dataset under both \\(2 \\times 25\\) stages and \\(5 \\times 10\\) stages settings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.788, + 0.795 + ], + "angle": 0, + "content": "Effectiveness of gradient projection To evaluate the SGP's capability to mitigate catastrophic forgetting, we designed experiments with various configurations: utilizing the method with SGP, the method without SGP, and the method without SGP but incorporating exemplars. In the last configuration, we preserved 5 videos, each containing 8 frames, as exemplars for prior classes to be used in future training sessions. Furthermore, we extended these comparisons to assess performance by employing a linear classifier and the proposed TSC. The comparison results are shown in Tab. 3, where (i), (iii), and (v) utilize the linear classifier, while (ii), (iv), and (vi) implement the TSC. In cooperation with the comparative analysis of the network's performance as the number of seen classes increases, as shown in Fig. 3, it can be found that the network tends to forget most knowledge of previous classes when no strategy is adopted." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Moreover, when comparing our SGP strategy against the exemplar-based distillation method, particularly through the comparison of (i) \\( v.s. \\) (v) and the comparison of (ii) \\( v.s. \\) (vi) in Tab. 3, we can find that our proposed SGP strategy" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.41, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Spatial-Temporal Subspace Projection for VCIL" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 4: Ablation study of the subspace basis size \\(\\tau\\) for TSC on UCF101 with \\(5\\times 10\\) stages setting." + }, + { + "type": "table", + "bbox": [ + 0.331, + 0.185, + 0.669, + 0.221 + ], + "angle": 0, + "content": "
τ=1τ=2τ=4τ=8τ=16
STSP (Ours)66.7567.8875.3982.8463.02
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.251, + 0.788, + 0.357 + ], + "angle": 0, + "content": "significantly outperforms the exemplar-based distillation method. Specifically, regarding the performance of base classes during incremental learning sessions, as shown in Fig. 4, it is evident that our proposed SGP has minimal impact on the classification performance for the base classes. These findings indicate that our proposed SGP can uphold accurate classification across all classes without compromising performance on the base classes, emphasizing the effectiveness of SGP in mitigating knowledge degradation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.358, + 0.788, + 0.465 + ], + "angle": 0, + "content": "Furthermore, based on the comparison in Fig. 3 and Fig. 4, we observe that all methods perform worse under the setting of \\(2 \\times 25\\) stages, while our method with SGP still achieves the best performance. A plausible explanation is that the model may be susceptible to overfitting when dealing with a limited set of novel classes. Moreover, for our method, the computation of the null space becomes more unstable in these scenarios, especially when updating the mean covariance matrix is solely based on a few classes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.492, + 0.788, + 0.703 + ], + "angle": 0, + "content": "Linear or Subspace-based Classifier? To assess the effectiveness of the proposed TSC, we compare it with a linear classifier based on an FC layer. Moreover, this ablation study extends to both the exemplar-based distillation method and our proposed SGP, with the comparison results shown in Tab. 3. For the exemplar-based distillation approach, the implementation of TSC leads to a performance increase of \\(3.22\\%\\) and \\(4.39\\%\\) on UCF101 at the 10-stage and 25-stage settings, respectively, as evidenced by the comparison between (i) and (ii). In the context of SGP, the comparison between (v) and (vi) reveals that the proposed TSC similarly improves the overall performance by \\(5.18\\%\\) and \\(5.23\\%\\) for the 10-stage and 25-stage settings on UCF101, respectively. Compared with the exemplar-based distillation method, our method employs SGP to constrain the classifier update through gradient subspace projection. The above findings align with the concept of TSC, collectively constraining the subspace-based classifier and further enhancing classification performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.842 + ], + "angle": 0, + "content": "From the comparison in Fig. 4, another observation is that TSC can even boost the classification performance on the base classes, demonstrating its effectiveness in modeling class representations. However, we also observe that TSC may perform poorly when there are no constraints (i.e. without exemplars nor SGP) on the network. One plausible explanation for this phenomenon is that, during incremental sessions, the network tends to shift the feature space toward new tasks without constraints from historical information. Such adjustments can result in the original subspace deviating from its previous configuration, leading to significant misalignment that greatly impacts the performance of TSC." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.377, + 0.129 + ], + "angle": 0, + "content": "H. Cheng et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.145, + 0.49, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.277, + 0.283, + 0.438, + 0.296 + ], + "angle": 0, + "content": "(a) UCF101 \\((5\\times 10\\) stages)" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.145, + 0.779, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.565, + 0.283, + 0.726, + 0.296 + ], + "angle": 0, + "content": "(b) UCF101 (2 × 25 stages)" + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.307, + 0.785, + 0.322 + ], + "angle": 0, + "content": "Fig. 4: Plots of accuracy over base classes along with the incremental steps on UCF101." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.35, + 0.788, + 0.518 + ], + "angle": 0, + "content": "The Size of the Subspace Basis We explore how the size of the subspace basis influences the classification accuracy of our proposed method, the comparison results are shown in Tab. 4. It can be found that large subspace basis sizes (i.e., \\(\\tau = 4\\) or \\(\\tau = 8\\)) are more effective at constructing the subspace for each category, enabling the capture of more distinctive features, which in turn leads to better performance as opposed to smaller bases. This improvement, however, results in an increase in both the network's parameters and computational complexity. Additionally, we can observe that a subspace basis size of 16 yields the least favorable results. This could be attributed to the difficulty in constructing orthogonal subspaces for new classes when the subspace basis is excessively large, especially as \\(M \\times \\tau\\) approaches the total number of classes \\((\\mathcal{C})\\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.538, + 0.36, + 0.554 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.788, + 0.735 + ], + "angle": 0, + "content": "In this paper, we proposed STSP, a novel spatial-temporal subspace projection method designed for video class-incremental learning tasks. The objective is to learn discriminative and robust representations for both video and class, facilitating further incremental learning. STSP mainly consists of two key components, i.e., temporal-base subspace classifier (TSC) and spatial-based gradient projection (SGP). TSC employs a subspace to model each video and its class, incorporating intra-frame temporal information. On the other hand, SGP is designed to mitigate catastrophic forgetting by projecting the gradient of network parameters in each layer onto the subspace spanned by the corresponding input spatial features. Extensive experiments on three benchmarks demonstrate the effectiveness of our proposed STSP method." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Limitations. In contrast to previous video class incremental learning methods, our proposed STSP method incorporates gradient projection to prevent the forgetting of historical information from previous classes. This strategy removes the necessity of storing exemplars and prototypes of previously learned classes, yet it requires the storage and updating of input features for both linear layers and the introduced TSC. Such a requirement could pose challenges when employing large backbones, requiring further exploration and discussion." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.41, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Spatial-Temporal Subspace Projection for VCIL" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.207 + ], + "angle": 0, + "content": "This research was carried out at the Rapid-Rich Object Search (ROSE) Lab at Nanyang Technological University in Singapore." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.23, + 0.323, + 0.246 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.26, + 0.785, + 0.302 + ], + "angle": 0, + "content": "1. Akyurek, A.F., Akyurek, E., Wijaya, D., Andreas, J.: Subspace regularizers for few-shot class incremental learning. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=boJy41J-tnQ" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.302, + 0.785, + 0.343 + ], + "angle": 0, + "content": "2. Belhumeur, P.N., Hespanha, J.P., Kriegman, D.J.: Eigenfaces vs. fisherfaces: Recognition using class specific linear projection. IEEE Transactions on pattern analysis and machine intelligence 19(7), 711-720 (1997)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.344, + 0.785, + 0.385 + ], + "angle": 0, + "content": "3. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6299-6308 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.385, + 0.785, + 0.426 + ], + "angle": 0, + "content": "4. Chen, H., Wang, Y., Hu, Q.: Multi-granularity regularized re-balancing for class incremental learning. IEEE Transactions on Knowledge and Data Engineering 35(7), 7263-7277 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.426, + 0.785, + 0.482 + ], + "angle": 0, + "content": "5. Cheraghian, A., Rahman, S., Ramasinghe, S., Fang, P., Simon, C., Petersson, L., Harandi, M.: Synthesized feature based few-shot class-incremental learning on a mixture of subspaces. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 8661-8670 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.482, + 0.785, + 0.523 + ], + "angle": 0, + "content": "6. Dhar, P., Singh, R.V., Peng, K.C., Wu, Z., Chellappa, R.: Learning without memorizing. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5138-5146 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.523, + 0.785, + 0.592 + ], + "angle": 0, + "content": "7. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=YicbFdNTTy" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.592, + 0.785, + 0.647 + ], + "angle": 0, + "content": "8. Douillard, A., Cord, M., Ollion, C., Robert, T., Valle, E.: Podnet: Pooled outputs distillation for small-tasks incremental learning. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XX 16. pp. 86-102. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.647, + 0.785, + 0.688 + ], + "angle": 0, + "content": "9. Feichtenhofer, C., Fan, H., Malik, J., He, K.: Slowfast networks for video recognition. In: Proceedings of the IEEE international conference on computer vision. pp. 6202-6211 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.688, + 0.785, + 0.731 + ], + "angle": 0, + "content": "0. Gao, Q., Zhao, C., Ghanem, B., Zhang, J.: R-dfcil: Relation-guided representation learning for data-free class incremental learning. In: European Conference on Computer Vision. pp. 423-439. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.731, + 0.785, + 0.798 + ], + "angle": 0, + "content": "1. Goyal, R., Ebrahimi Kahou, S., Michalski, V., Materzynska, J., Westphal, S., Kim, H., Haenel, V., Fruend, I., Yianilos, P., Mueller-Freitag, M., et al.: The\" something something\" video database for learning and evaluating visual common sense. In: Proceedings of the IEEE international conference on computer vision. pp. 5842-5850 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "2. Hou, S., Pan, X., Loy, C.C., Wang, Z., Lin, D.: Learning a unified classifier incrementally via rebalancing. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 831-839 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.26, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "H. Cheng et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "13. Huang, H., Wang, Y., Hu, Q., Cheng, M.M.: Class-specific semantic reconstruction for open set recognition. IEEE transactions on pattern analysis and machine intelligence 45(4), 4214-4228 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.233 + ], + "angle": 0, + "content": "14. Ji, S., Xu, W., Yang, M., Yu, K.: 3d convolutional neural networks for human action recognition. IEEE transactions on pattern analysis and machine intelligence 35(1), 221-231 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.234, + 0.787, + 0.275 + ], + "angle": 0, + "content": "15. Kuehne, H., Jhuang, H., Garrote, E., Poggio, T., Serre, T.: Hmdb: a large video database for human motion recognition. In: 2011 International conference on computer vision. pp. 2556-2563. IEEE (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.276, + 0.787, + 0.303 + ], + "angle": 0, + "content": "16. Li, Z., Hoiem, D.: Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence 40(12), 2935-2947 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.304, + 0.787, + 0.346 + ], + "angle": 0, + "content": "17. Lin, J., Gan, C., Han, S.: Temporal shift module for efficient video understanding. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7083-7093 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.346, + 0.787, + 0.389 + ], + "angle": 0, + "content": "18. Liu, C.: Gabor-based kernel pca with fractional power polynomial models for face recognition. IEEE transactions on pattern analysis and machine intelligence 26(5), 572-581 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.389, + 0.787, + 0.445 + ], + "angle": 0, + "content": "19. Masana, M., Liu, X., Twardowski, B., Menta, M., Bagdanov, A.D., Van De Weijer, J.: Class-incremental learning: survey and performance evaluation on image classification. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(5), 5513-5533 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.445, + 0.787, + 0.487 + ], + "angle": 0, + "content": "20. McCloskey, M., Cohen, N.J.: Catastrophic interference in connectionist networks: The sequential learning problem. In: Psychology of learning and motivation, vol. 24, pp. 109-165. Elsevier (1989)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.787, + 0.53 + ], + "angle": 0, + "content": "21. Papyan, V., Han, X., Donoho, D.L.: Prevalence of neural collapse during the terminal phase of deep learning training. Proceedings of the National Academy of Sciences 117(40), 24652-24663 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.53, + 0.787, + 0.572 + ], + "angle": 0, + "content": "22. Park, J., Kang, M., Han, B.: Class-incremental learning for action recognition in videos. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 13698-13707 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.573, + 0.787, + 0.628 + ], + "angle": 0, + "content": "23. Pei, Y., Qing, Z., Cen, J., Wang, X., Zhang, S., Wang, Y., Tang, M., Sang, N., Qian, X.: Learning a condensed frame for memory-efficient video class-incremental learning. Advances in Neural Information Processing Systems 35, 31002-31016 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.629, + 0.787, + 0.671 + ], + "angle": 0, + "content": "24. Pei, Y., Qing, Z., Zhang, S., Wang, X., Zhang, Y., Zhao, D., Qian, X.: Space-time prompting for video class-incremental learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11932-11942 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.671, + 0.787, + 0.713 + ], + "angle": 0, + "content": "25. Peifeng, G., Xu, Q., Wen, P., Yang, Z., Shao, H., Huang, Q.: Feature directions matter: Long-tailed learning via rotated balanced representation. In: International Conference on Machine Learning. pp. 27542-27563. PMLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.714, + 0.787, + 0.756 + ], + "angle": 0, + "content": "26. Rebuffi, S.A., Kolesnikov, A., Sperl, G., Lampert, C.H.: icarl: Incremental classifier and representation learning. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 2001-2010 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.757, + 0.787, + 0.798 + ], + "angle": 0, + "content": "27. Saha, G., Garg, I., Roy, K.: Gradient projection memory for continual learning. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=3A0jORCNC2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "28. Simon, C., Koniusz, P., Nock, R., Harandi, M.: Adaptive subspaces for few-shot learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4136-4145 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.41, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Spatial-Temporal Subspace Projection for VCIL" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "29. Simonyan, K., Zisserman, A.: Two-stream convolutional networks for action recognition in videos. In: Advances in neural information processing systems. pp. 568-576 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.787, + 0.218 + ], + "angle": 0, + "content": "30. Soomro, K., Zamir, A.R., Shah, M.: Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.218, + 0.786, + 0.259 + ], + "angle": 0, + "content": "31. Tran, D., Bourdev, L., Fergus, R., Torresani, L., Paluri, M.: Learning spatiotemporal features with 3d convolutional networks. In: Proceedings of the IEEE international conference on computer vision. pp. 4489-4497 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.787, + 0.313 + ], + "angle": 0, + "content": "32. Tran, D., Wang, H., Torresani, L., Ray, J., LeCun, Y., Paluri, M.: A closer look at spatiotemporal convolutions for action recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6450-6459 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.787, + 0.357 + ], + "angle": 0, + "content": "33. Turk, M.A., Pentland, A.P.: Face recognition using eigenfaces. In: Proceedings. 1991 IEEE computer society conference on computer vision and pattern recognition. pp. 586-587. IEEE Computer Society (1991)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.787, + 0.398 + ], + "angle": 0, + "content": "34. Tzimiropoulos, G., Zafeiriou, S., Pantic, M.: Subspace learning from image gradient orientations. IEEE transactions on pattern analysis and machine intelligence 34(12), 2454-2466 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.787, + 0.453 + ], + "angle": 0, + "content": "35. Villa, A., Alcázar, J.L., Alfarra, M., Alhamoud, K., Hurtado, J., Heilbron, F.C., Soto, A., Ghanem, B.: Pivot: Prompting for video continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24214-24223 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.787, + 0.508 + ], + "angle": 0, + "content": "36. Villa, A., Alhamoud, K., Escorcia, V., Caba, F., Alcazar, J.L., Ghanem, B.: vclimb: A novel video class incremental learning benchmark. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19035-19044 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.508, + 0.787, + 0.564 + ], + "angle": 0, + "content": "37. Wang, C., Guo, L., Wang, Y., Cheng, H., Yu, Y., Wen, B.: Progressive divide-and-conquer via subsampling decomposition for accelerated mri. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 25128-25137 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.787, + 0.605 + ], + "angle": 0, + "content": "38. Wang, J., Cherian, A.: Gods: Generalized one-class discriminative subspaces for anomaly detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8201-8211 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.605, + 0.787, + 0.647 + ], + "angle": 0, + "content": "39. Wang, S., Li, X., Sun, J., Xu, Z.: Training networks in null space of feature covariance for continual learning. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 184-193 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.787, + 0.689 + ], + "angle": 0, + "content": "40. Yang, J., Frangi, A.F., Yang, J.y., Zhang, D., Jin, Z.: Kpca plus lda: a complete kernel fisher discriminant framework for feature extraction and recognition. IEEE Transactions on pattern analysis and machine intelligence 27(2), 230-244 (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.787, + 0.743 + ], + "angle": 0, + "content": "41. Yang, S., Liu, J., Lu, S., Er, M.H., Kot, A.C.: Collaborative learning of gesture recognition and 3d hand pose estimation with multi-order feature analysis. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16. pp. 769-786. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.743, + 0.787, + 0.799 + ], + "angle": 0, + "content": "42. Yang, Y., Yuan, H., Li, X., Lin, Z., Torr, P., Tao, D.: Neural collapse inspired feature-classifier alignment for few-shot class-incremental learning. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=y5W8tpoJhtJ" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "43. Zhang, L., Fu, J., Wang, S., Zhang, D., Dong, Z., Chen, C.P.: Guide subspace learning for unsupervised domain adaptation. IEEE transactions on neural networks and learning systems 31(9), 3374-3388 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "H. Cheng et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "44. Zhao, B., Xiao, X., Gan, G., Zhang, B., Xia, S.T.: Maintaining discrimination and fairness in class incremental learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 13208-13217 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.788, + 0.231 + ], + "angle": 0, + "content": "45. Zhu, H., Koniusz, P.: Ease: Unsupervised discriminant subspace learning for transductive few-shot learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9078-9088 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.231, + 0.788, + 0.26 + ], + "angle": 0, + "content": "46. Zhu, P., Yao, X., Wang, Y., Hui, B., Du, D., Hu, Q.: Multiview deep subspace clustering networks. IEEE Transactions on Cybernetics (2024)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.26 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/STSP_ Spatial-Temporal Subspace Projection for Video Class-incremental Learning/9413ef29-ccb6-4b51-bb25-c4b8dee0d763_origin.pdf b/2024/STSP_ Spatial-Temporal Subspace Projection for Video Class-incremental Learning/9413ef29-ccb6-4b51-bb25-c4b8dee0d763_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1ce9cdc1bfc871cac78303704cd15959f9a3ebf3 --- /dev/null +++ b/2024/STSP_ Spatial-Temporal Subspace Projection for Video Class-incremental Learning/9413ef29-ccb6-4b51-bb25-c4b8dee0d763_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:253d2a608748ff9e64176f5866e0d529c83b3137cd480ede633df684a9ec9352 +size 887816 diff --git a/2024/STSP_ Spatial-Temporal Subspace Projection for Video Class-incremental Learning/full.md b/2024/STSP_ Spatial-Temporal Subspace Projection for Video Class-incremental Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..bba683beffc5acaaf5992e601289e327274af356 --- /dev/null +++ b/2024/STSP_ Spatial-Temporal Subspace Projection for Video Class-incremental Learning/full.md @@ -0,0 +1,333 @@ +# STSP: Spatial-Temporal Subspace Projection for Video Class-incremental Learning + +Hao Cheng $^{1,2}$ , Siyuan Yang $^{1}$ , Chong Wang $^{1}$ , Joey Tianyi Zhou $^{3,4}$ , Alex C. Kot $^{1}$ , and Bihan Wen $^{1*}$ + +1Nanyang Technological University, Singapore {siyuan.yang, wang1711, eackot, bihan.wen}@ntu.edu.sg + +2 Hebei University of Technology, Tianjin, China 2024025@hebut.edu.cn + +$^{3}$ CFAR, Agency for Science, Technology and Research, Singapore + +4 IHPC, Agency for Science, Technology and Research, Singapore zhouty@cfar.a-star.edu.sg + +Abstract. Video class-incremental learning (VCIL) aims to learn discriminative and generalized feature representations for video frames to mitigate catastrophic forgetting. Conventional VCIL methods often retain a subset of frames or features from prior tasks as exemplars for subsequent incremental learning stages. However, these strategies overlook the connection between base and novel classes, sometimes even leading to privacy leakage. To address this challenge, we introduce a Spatial-Temporal Subspace Projection (STSP) scheme for VCIL. Specifically, we propose a discriminative Temporal-based Subspace Classifier (TSC) that represents each class with an orthogonal subspace basis and adopts subspace projection loss for classification. Unlike typical classification methods relying on fully connected layers, our TSC discerns the spatial-temporal dynamics in video content, thereby enhancing the representation of each video sample. Additionally, we implement inter- and intra-class orthogonal constraints into TSC, ensuring each class occupies a unique orthogonal subspace, defined by its basis. To prevent catastrophic forgetting, we further employ a Spatial-based Gradient Projection (SGP) strategy, adjusting network gradients to align with the null space of the spatial feature set from previous tasks. Extensive experiments conducted on three benchmarks, namely HMDB51, UCF101, and Something-Something V2, demonstrate that our STSP method outperforms state-of-the-art comparison methods, evidencing its efficacy in VCIL. + +Keywords: Video action recognition $\cdot$ Class incremental learning $\cdot$ Subspace Projection $\cdot$ Gradient Projection + +# 1 Introduction + +Deep neural networks have demonstrated remarkable performance in modern computer vision tasks [7,17,32,37,46], particularly when trained on large, annotated datasets. However, the real-world challenge of incrementally incorporating + +![](images/173df513cde9670e818f909bb4b61e1031d81913c37529a5d66ea38e36c0f6dc.jpg) +Fig. 1: Class incremental learning with (a) linear classifier and (b) subspace-based classifier. Incorporating new classes alters the decision hyperspace and modifies the representation of prior classes, leading to misalignment. In contrast, subspace-based classifiers, by constructing an orthogonal basis for each class, establish a unique geometric structure for each class, thereby reducing class misalignment. + +![](images/eed59928631da623d347d013762b457e615a6324bed1440d775135688415c2fb.jpg) + +data from new classes into a model initially trained on existing classes presents significant hurdles, known as class-incremental learning (CIL) [26]. Unlike typical classification tasks, CIL involves accessing novel classes and limited memories of prior tasks in incremental sessions, often due to privacy concerns or memory constraints. Consequently, merely updating the model with data from new classes can result in overfitting substantial performance degradation on previous classes, a phenomenon referred to as catastrophic forgetting [19, 20]. + +Recent advancements in CIL methods [1,4,26,44] have shown remarkable performance in image classification tasks. However, recent CIL studies [1,21,25,42] demonstrate that straightforward fine-tuning of the network for novel classes may lead the subspace spanned by the novel classifier to deviate from the base one, as illustrated in Fig. 1 (a). The classification performance can be significantly affected by misalignment, particularly in cases where a notable distribution gap between novel and base classes. Moreover, when extending this task to Video Class-Incremental Learning (VCIL), unique challenges due to the inherent information redundancy within video samples, such as continuous actions with a static background. This necessitates methods to model temporal-spatial relationships between video frames to extract critical information for action recognition properly. However, prevalent approaches [22-24,35,36] in VCIL predominantly involve a learning pipeline that employs simple operations (e.g., mean) on the feature vectors of sampled frames to generate video representations for action recognition. While this methodology is effective for standard action recognition with sufficient data, it may struggle to ensure robust generalization in scenarios of incremental learning. The aforementioned observations prompt us to consider acquiring a more suitable representation for both the classifier and video features while also exploring their relationship. To address these challenges, several subspace learning-based methods [28,33,38,45] have shown effectiveness in reducing dimensionality and capturing intrinsic structure, thereby offering a potential solution to the issues associated with simplistic video representation methods. + +In this paper, we propose a spatial-temporal subspace projection method for VCIL including a temporal-based subspace classifier (TSC) and a spatial-based gradient projection (SGP) strategy. Specifically, we extend the existing linear + +classifiers based on fully connected (FC) layers to a TSC in which a subspace basis is adopted to model each class. Differing from linear classifiers that seek to distinguish classes in feature space through a linear decision boundary, the proposed orthogonality constraints offer a distinct geometric structure for each class. This characteristic not only facilitates a more comprehensive representation of each class but also allows for straightforward extension to novel classes. Furthermore, to avoid catastrophic forgetting, we utilize the SGP strategy to constrain the process of gradient update by projecting the gradient of network parameters in each layer onto the subspace of corresponding input features so that losses for the old task will not increase during the incremental sessions. Compared with existing VCIL methods [22-24,35,36] that rely on feature distillation or prompt-based fine-tuning strategies, SGP offers the advantage of fine-tuning all network parameters without necessitating regularizers. Furthermore, SGP obviates the need to explicitly retain class-relevant information, such as frames or prototypes, thereby contributing to privacy protection. + +In summary, we make the following contributions: 1) We propose a spatial-temporal subspace projection (STSP) method, containing a temporal-based subspace classifier and spatial-based gradient projection strategy to better model feature and class representation for discriminative classification. 2) We employ a spatial-based gradient projection strategy to constrain the gradient update process without the need to store frames or prototypes from previous tasks. This not only enhances performance but also mitigates the risk of privacy leakage. 3) The proposed STSP method achieves superior results on three video action recognition incremental benchmarks over state-of-the-art methods. + +# 2 Related Work + +# 2.1 Video Action Recognition + +With significant advances in deep learning, current video action recognition approaches aim to learn effective spatial-temporal representation by well-designed network architectures. A notable advancement by Simonyan and Zisserman [29] introduces a dual-stream architecture that analyzes RGB frames in one stream and optical flow in another, laying the foundation for subsequent enhancements. Recently, Lin et al. [17] proposed a Temporal Shift Module (TSM), which shifts some channels along the temporal dimension to perform temporal interaction between the features from adjacent frames. Concurrently, significant research [3, 13, 14, 31, 32, 41] has evolved 2D Convolutional Neural Networks (CNNs) into 3D structures to capture both spatial and temporal contexts within videos, a critical aspect of human action recognition. Notably, Tran et al. [31] introduced the C3D model, a 3D CNN designed to learn spatial-temporal features through an end-to-end learning approach. Carreira et al. [3] developed the Inflated 3D CNN, which extends a 2D CNN into the temporal domain by inflating its convolutional and pooling layers. Additionally, Feichtenhofer et al. [9] crafted a dual-pathway 3D CNN, featuring both slow and fast pathways that process RGB frames at varying frame rates to simultaneously capture semantic content and motion details. + +Despite significant progress in action recognition, catastrophic forgetting has received relatively limited attention. This work aims to address this gap by developing a spatial-temporal subspace projection method, which aims to offer a robust solution to the challenges of incremental learning in action recognition. + +# 2.2 Class-Incremental Learning + +CIL addresses the challenge of adapting machine learning models to recognize new classes without forgetting previously learned information, a phenomenon known as catastrophic forgetting. Pioneering in this domain, Rebuffi et al. [26] introduced iCaRL, which combines representation learning with a memory mechanism to selectively retain the knowledge of previous classes. Following this, Hou et al. [12] proposed a UCIR strategy, which addresses the class imbalance in CIL by adjusting the classification layer to focus more on newer classes while still remembering old ones. Moreover, Gao et al. [10] introduced R-DFCIL, employing relation-guided representation learning to overcome catastrophic forgetting and enable the assimilation of new classes in a data-free scenario. + +While CIL has been well-studied in the image domain, its application to the video domain remains largely unexplored. Park et al. [22] introduced Time-Channel Distillation (TCD), leveraging knowledge distillation loss with time-channel importance masks. Pei et al. [23] proposed FrameMaker, a memory-efficient VCIL strategy that develops a condensed frame for each video to facilitate future incremental tasks with less memory usage. Pei et al. [24] introduced ST-Prompt, a novel space-time prompting framework tailored for VCIL, enhancing the model's adaptability to new classes with minimal memory increase. + +Existing VCIL approaches often rely on preserving a subset of frames or features from previous tasks as exemplars for subsequent incremental stages to prevent catastrophic forgetting. However, these strategies overlook the connection between base and novel classes and potentially lead to privacy leakage. In contrast, we employ a spatial-based gradient projection to regulate gradient updates, thereby mitigating catastrophic forgetting without these drawbacks. + +# 2.3 Subspace Representation + +Subspace learning, by learning a suitable representation for high dimensional data, has shed light on various vision tasks [2, 18, 34, 40, 43]. Existing subspace methods such as Principal Component Analysis (PCA) [2, 18] and Linear Discriminant Analysis (LDA) [2, 40] adaptively transfer the original input data to low dimensional subspace for dimensionality reduction. For example, Tzimiropoulos et al. [34] formulated PCA on image gradient orientations, which demonstrated robust performance in face recognition. Recently, subspace learning has been further extended to regularize the implicit features or learned classifiers within deep networks [1,28,38,45]. Notably, Cheraghian et al. [5] constructed a mixture of subspaces to represent the feature clusters formed by visually similar samples. Such subspace representation enables a better approximation of the overall data distribution, consequently leading to superior performance in + +![](images/969ada04ac68c9d7743347b61d21d25e5a17aaa79b0ee3b7f0e24a433ec9a6c5.jpg) +Fig. 2: The overall framework of the proposed STSP method, including a temporal-based subspace classifier (TSC) and a spatial-based gradient projection (SGP). For each input video, spatial-temporal features are extracted using a backbone network. Subsequently, The TSC classifies by projecting temporal features into a subspace, where classification decisions are based on the distance to each class's subspace basis within this subspace. During parameter updates for the novel tasks, the SGP constrains gradient modifications for each specific layer, drawing on the mean covariance matrix $M_{t-1}$ of previous tasks' input features. Upon the current session completion, the mean covariance matrix is updated to $M_t$ for the gradient projection in the subsequent session. + +CIL. To mitigate the catastrophic forgetting and overfitting, Simon et al. [28] formulated the dynamic classifiers via subspace modeling which demonstrated a significant robustness to noise in few-shot learning. Similarly, Akyurek et al. [1] proposed an additional subspace regularizer that pulls the weight vectors of the novel class close to the subspace spanned by the base classes. + +However, existing methods lack sufficient constraints on the representation of each class in the classifier. In this paper, we model each class via an orthogonal subspace basis and leverage spatial-based gradient projection to update the entire network during incremental sessions. + +# 3 Method + +In this section, we start with an introduction of VCIL problem formulation. Then, the proposed STSP method is described and explained in detail, followed by the loss function of our method as well as the training and inference procedure. + +In VCIL, a series of distinct tasks $\{\mathcal{T}_1,\mathcal{T}_2,\dots ,\mathcal{T}_k,\dots \}$ unfolds sequentially. Each task $\mathcal{T}_k$ comprises a distinct subset of video-class pairs $\{(D_i^k,y_i^k),y_i^k\in$ $\mathcal{C}_k\} |_{i = 1}^{|\mathcal{T}_k|}$ . Specifically, the novel class set $\mathcal{C}_k$ of $\mathcal{T}_k$ is distinct from previous classes, i.e., $(\mathcal{C}_1\cup \dots \cup \mathcal{C}_{k - 1})\cap \mathcal{C}_k = \emptyset$ . The primary objective is to incrementally train a deep neural network $f(\cdot)$ capable of accurately classifying videos across all classes by preventing the loss of class information from previous tasks. In contrast to methods that have access to a limited exemplar set from historical tasks, our emphasis lies on exemplar-free VCIL, where the model is not permitted to store exemplars from previous tasks during the learning process of the current session. + +# 3.1 Overall STSP Framework + +Given the problem formulation, we follow the standard learning protocol of previous methods [22, 23] based on knowledge distillation. During the $k^{th}$ incremental session, the network $f(\cdot ;\mathbf{w})$ parameterized by $\mathbf{w}$ is updated from $\mathbf{w}_{k - 1}$ to adapt to the novel class set $\mathcal{C}_k$ . Fig. 2 illustrates the overall framework of our method. For an input video $D_{i}^{k}$ belonging to $\mathcal{C}_k$ , we first extract $T$ frames as $V_{i}$ , which are then fed to the backbone network to generate the feature embeddings. Subsequently, we employ a TSC to differentiate between distinct classes. Moreover, our method stands out by employing an SGP strategy without restoring frames or prototypes from previous tasks. More in detail, the network parameters are optimized by projecting the gradients in each layer to the subspace corresponding to their input features. With the merit of this strategy, we can completely avoid the risk of privacy leakage, whereas existing approaches may suffer. + +# 3.2 Temporal-based Subspace Classifier + +In accordance with the standardized procedure for VCIL, given an input video-class pair denoted as $(D_i,y_i)$ , we first randomly sampled $T$ frames $V_{i}^{T}$ . Subsequently, these frames are fed to the backbone network, yielding the feature embeddings as $X_{i}\in \mathbb{R}^{T\times C\times H\times W}$ , where $C$ , $H$ , and $W$ represent the size of channel, height, and width, respectively. For the existing architectures [22-24] for VCIL, a prevalent technique for obtaining video representations is global average pooling (GAP), denoted as $Z_{i} = GAP(X_{i})\in \mathbb{R}^{C\times T}$ . Following this, a linear classifier $L$ , comprising an FC layer with weights $\mathbf{w}\in \mathbb{R}^{C\times K}$ , is utilized to predict the potential class of each frame. The class predictions from individual frames are subsequently averaged to produce the overall classification result. + +Nevertheless, the temporal information holds the potential to encapsulate more intricate intra-frame structural details, warranting thoughtful consideration and seamless integration into the classification process. Additionally, in each incremental task, adjustments to the linear classifier may be necessary to modify the overall representation across all classes, which could influence the base classes for all previous incremental sessions. Conversely, a subspace-based classifier demonstrates heightened adaptability, enabling the modeling of each new class within its unique subspace without imposing substantial effects on the existing representations of previous classes. + +To this end, to properly model each class and learn a robust classifier, we construct a TSC based on subspaces spanned by an orthogonal basis of each class. Specifically, for $m$ -th class, we construct a subspace with orthogonal basis $P_{m} \in \mathbb{R}^{C \times \tau}$ with $P_{m}^{\top} P_{m} = I_{\tau}, \tau \ll C$ , where $C$ and $\tau$ indicate the feature dimension and the number of subspace basis for $m$ -th class, respectively. Hence, for each task $\mathcal{T}_k$ in the $k^{th}$ incremental session, we construct $|\mathcal{C}_k|$ distinct subspaces for $|\mathcal{C}_k|$ novel classes. We then adopt current subspaces to predict the class of $Z_{i}$ based on the subspace distance between the subspace spanned by temporal video features $Z_{i}$ and the orthogonal basis of each class as: + +$$ +d _ {m} (Z _ {i}) = - \left\| Z _ {i} Z _ {i} ^ {\top} - P _ {m} P _ {m} ^ {\top} \right\| _ {F} ^ {2}. \tag {1} +$$ + +Then, the probability of $Z_{i}$ assigned to $m$ -th class can be calculated using the softmax function as follows: + +$$ +p _ {m} (Z _ {i}) = \frac {e ^ {d _ {m} (Z _ {i})}}{\sum_ {m ^ {\prime} = 1} ^ {M} e ^ {d _ {m ^ {\prime}} (Z _ {i})}}, \tag {2} +$$ + +where $M$ is the total number of classes in the current session. Now, we can employ the cross-entropy loss on Eq. (2) to update the network as: + +$$ +L _ {C E} = - \sum_ {i = 1} ^ {J} y _ {i} \log p _ {i} \left(Z _ {i}\right), \tag {3} +$$ + +where $J$ denotes the number of videos in each mini-batch. + +To encourage the network to capture discriminative information for each class and construct distinct subspaces corresponding to different classes, we impose orthogonal constraints on both the temporal features of the video and the weights of the TSC. Specifically, we employ orthogonal loss on the temporal features $Z_{i}$ of each video $D_{i}$ in the final layer as: + +$$ +L _ {O F} = \sum_ {i = 1} ^ {J} \left\| Z _ {i} ^ {\top} Z _ {i} - I _ {T} \right\| _ {F} ^ {2}, \tag {4} +$$ + +where $I_T \in \mathbb{R}^{T \times T}$ is an identity matrix. This approach guarantees that every frame independently captures distinct, class-aware information, setting it apart uniquely from the rest. Furthermore, it facilitates the construction of a discriminative subspace for each corresponding class. + +For TSC, we impose intra- and inter-class orthogonal constraints on the subspace basis of each class, which is denoted by $L_{OW}$ as: + +$$ +L _ {O W} = \sum_ {m = 1} ^ {M} \left(\underbrace {\left\| P _ {m} ^ {\top} P _ {m} - I _ {\tau} \right\| _ {F} ^ {2}} _ {\text {i n t r a - c l a s s}} + \sum_ {q = m + 1} ^ {M} \underbrace {\left\| P _ {m} ^ {\top} P _ {q} \right\| _ {F} ^ {2}} _ {\text {i n t e r - c l a s s}}\right). \tag {5} +$$ + +The intra-class orthogonal constraint is introduced to ensure independence among the basis vectors of subspaces, which is crucial for learning distinct and discriminative feature representations. Furthermore, the imposition of inter-class orthogonal constraint ensures that the basis vectors of distinct classes are orthogonal, signifying their mutual independence. This orthogonality enforcement serves to prevent confusion and safeguard the uniqueness of each basis vector, thus preserving the integrity of class information previously learned. Moreover, these orthogonal constraints enhance discriminability, guaranteeing that the basis vectors of subspaces related to different classes are clearly distinguishable. + +Therefore, the overall objective function for TSC is defined as: + +$$ +L _ {t o t a l} = \alpha L _ {C E} + \beta L _ {O F} + \gamma L _ {O W}, \tag {6} +$$ + +where $\alpha$ , $\beta$ , and $\gamma$ denote the weighting parameters of each losses, respectively. + +# 3.3 Spatial-based Gradient Projection + +In contrast to prior approaches [22-24] that rely on retaining exemplars or prompts from previous classes during incremental sessions, our focus is on addressing the challenging exemplar-free VCIL problems by imposing constraints on the network during parameter updates. To avoid catastrophic forgetting on previous classes in the incremental sessions, the network's parameter update should satisfy the following proposition: + +Proposition 1. Given the data $x_{k}$ from the previous task $\mathcal{T}_k$ , the current network $f(\cdot ;\mathbf{w}_t)$ could preserve the historical information if it satisfies the following equation for all sessions: + +$$ +f \left(x _ {k}; \mathbf {w} _ {t}\right) = f \left(x _ {k}; \mathbf {w} _ {k}\right), \quad \forall k < t. \tag {7} +$$ + +Here, $\mathbf{w}_t$ denotes the network parameters at task $T_{t}$ + +To realize the objective presented in Proposition 1, it is imperative to confine the gradient update of network parameters to a specific direction—precisely, the gradient update should lie in the subspace spanned by the corresponding input spatial features. Specifically, the gradients in the linear layers, including FC and Convolution layers $^5$ should satisfy Proposition 2 [27,39]. + +Proposition 2. Let $f^l(\cdot; \mathbf{w}_t)$ denotes the $l$ -th linear layer of the network trained on current task $\mathcal{T}_t$ and its corresponding input feature $A_k^l$ from any previous task $\mathcal{T}_k$ ( $k < t$ ). During each training step $s$ , if the linear layer's parameter update $\Delta \mathbf{w}_{t,s}^l$ lies in the null space of $A_k^l$ , i.e., + +$$ +A _ {k} ^ {l} \Delta \mathbf {w} _ {t, s} ^ {l} = 0, \tag {8} +$$ + +we could preserve the historical information as $f^{l}\left(A_{k}^{l};\mathbf{w}_{t}^{l}\right) = f^{l}\left(A_{k}^{l};\mathbf{w}_{k}^{l}\right)$ . + +Building upon Proposition 2, we further extend it to our proposed TSC with Eq. (1), as detailed in Proposition 3. + +Proposition 3. Let $f^{TSC}(\cdot ;\mathbf{w}_t)$ denotes the TSC of the network trained on current task $\mathcal{T}_t$ and its corresponding input feature $Z_{k}$ from any previous task $\mathcal{T}_k$ ( $k < t$ ). During each training step $s$ , if the TSC parameter update $\Delta \mathbf{w}_{t,s}^{TSC}$ lies in the null space of the subspace spanned by $Z_{k}$ , i.e., + +$$ +Z _ {k} Z _ {k} ^ {\top} \Delta \mathbf {w} _ {t, s} ^ {T S C} = 0, \tag {9} +$$ + +similarly we have $f^{TSC}\left(Z_k;\mathbf{w}_t^{TSC}\right) = f^{TSC}\left(Z_k;\mathbf{w}_k^{TSC}\right)$ . + +Proof. Please refer to the supplemental material. + +Inspired by previous work [27,39], we calculate the mean covariance matrix, denoted as $M_{t-1}$ , of the input feature for each specific layer in each mini-batch to enhance network stability and optimize memory efficiency. Here, let $Q$ represent the collective term for the current input feature at each specific layer, with $A_t^l$ for the $l$ -th linear layer and $Z_t Z_t^\top$ for the TSC, respectively. To initialize, $M_0$ is set to 0 for the first session. At the end of each session $t$ , $M_t$ is updated as: + +$$ +M _ {t} = M _ {t - 1} + \sum_ {b = 1} ^ {\mathcal {B}} \frac {1}{J} Q _ {b} ^ {\top} Q _ {b}, \tag {10} +$$ + +where $\mathcal{B}$ represents the number of mini-batches in the current task, and $J$ denotes the number of videos in each mini-batch. + +During the feature update in task $\mathcal{T}_t$ , we perform Singular Value Decomposition (SVD) to find the approximate null space of the given feature matrix $M_{t-1}$ corresponding to Proposition 2 and Proposition 3 as: + +$$ +\left[ U _ {1}, U _ {2} \right] \left[ \begin{array}{c c} \Sigma_ {1} & 0 \\ 0 & \Sigma_ {2} \end{array} \right] \left[ \begin{array}{l} U _ {1} ^ {\top} \\ U _ {2} ^ {\top} \end{array} \right] = \operatorname {S V D} \left(M _ {t - 1}\right), \tag {11} +$$ + +where $U_{1}, U_{2}$ represent unitary matrices that comprise orthonormal bases associated with the singular values in $\Sigma_{1}, \Sigma_{2}$ , respectively, with all zero singular values relegated to $\Sigma_{2}$ . However, it is unrealistic to guarantee that there always exists zero singular values. Previous work [27, 39] constructs $U_{2}^{\top}$ by selecting singular values restricted to a specific proportion. However, this strategy might not always provide the most compact representation, particularly when there is a significant difference in the scale of singular values. Hence, we employ a threshold strategy based on the energy retained in the matrix $M$ for the construction $U_{2}^{\top}$ , corresponding to a segment of smallest singular values that satisfy $\| \Sigma_{2} \|_{F}^{2} \leq (1 - \eta) \left( \| \Sigma_{1} \|_{F}^{2} + \| \Sigma_{2} \|_{F}^{2} \right)$ , with the proportion of retained energy denoted by $\eta$ . When $\eta \approx 1$ , it is reasonable to approximate the null space of the matrix $M$ with $U_{2}^{\top}$ . This enables the parameter update $\Delta \mathbf{w}$ from the gradient $\mathbf{g}$ concerning the parameters $\mathbf{w}$ through subspace projection based on $U_{2}$ . That is, during the training step $s$ within any incremental session $t$ , we have: + +$$ +\Delta \mathbf {w} _ {t, s} = U _ {2} U _ {2} ^ {\top} \mathbf {g} _ {t, s}. \tag {12} +$$ + +# 3.4 Model Optimization and Inference + +Our STSP method is summarized in Algorithm 1. In the training phase, the network parameters are updated conventionally during the initial session. After this initial session, considering parameter updates, the mean covariance of input features from the base classes is retained for each layer. These retained features are then utilized to update parameters according to Eq. (12) at each step in every subsequent incremental session. Additionally, after each incremental session, these parameters are further updated with features extracted from novel classes. + +For the testing phase, no additional memories or operations are necessary. We simply feed the video into the network and obtain classification results using the trained model. + +Algorithm 1 STSP for class incremental learning +Input: Task series $\{\mathcal{T}_1,\mathcal{T}_2,\dots \}$ , network $f(\cdot ;\mathbf{w})$ with parameters w for linear layers and TSC, learning rate $\alpha$ and energy threshold $\eta$ +Initialize: $M_0 = 0$ for each linear layer and TSC. +1: for $\mathcal{T}_t\in \{\mathcal{T}_1,\mathcal{T}_2,\dots \}$ do +2: $s = 0$ +3: while not converge do +4: Sample a mini-batch $\{D^t,y^t\} \in \mathcal{T}_t$ with videos $\{D_i^t,y_i^t\}_{j = 1}^J$ +5: Compute $f(\{D^{t},y^{t}\} ;\mathbf{w}_{t,s})$ and obtain the gradient $\mathbf{g}_{t,s}$ +6: if $t > 1$ then +7: for $\mathbf{w}_{t,s}$ in linear layers and TSC do +8: Compute $U_{2}$ using $M_{t - 1}$ via Eq. (11) with a threshold $\eta$ +9: Update $\Delta \mathbf{w}_{t,s}$ using Eq. (12) +10: Update $\mathbf{w}_{t,s + 1} = \mathbf{w}_{t,s} - \alpha \Delta \mathbf{w}_{t,s}$ +11: end for +12: else +13: Update $\Delta \mathbf{w}_{t,s} = \mathbf{g}_{t,s}$ +14: Update $\mathbf{w}_{t,s + 1} = \mathbf{w}_{t,s} - \alpha \Delta \mathbf{w}_{t,s}$ +15: end if +16: $s = s + 1$ +17: end while +18: Get current input features $Q$ at each linear layer and TSC +19: Update $M_t$ using Eq. (10) for later incremental session +20: end for + +# 4 Experiments + +In this section, we compare our results against state-of-the-art techniques in class-incremental action recognition benchmarks. Additionally, we perform ablation studies to demonstrate the efficacy of our proposed method. + +# 4.1 Experimental Setup + +Datasets. The evaluation of our proposed framework is conducted using the UCF101 [30], HMDB51 [15], and Something-Something V2 [11] datasets, which are recognized benchmarks in video action recognition. The UCF-101 dataset comprises 13,320 instances across 101 action categories, with the data divided into three splits for training (70%) and testing (30%). Similarly, the HMDB-51 dataset features 6,766 video clips distributed among 51 action categories, also with three splits for training and testing in the same proportions. For our evaluations, we selected the first split from both datasets. Distinct from UCF101 and HMDB51, the Something-Something V2 dataset, a crowd-sourced collection, includes 220,000 videos across 174 classes. This dataset uniquely maintains consistent objects and backgrounds across its action categories, thereby emphasizing the importance of a model's capability to discern subtle motion cues. + +Evaluation Protocol. Our methodology adheres to the class-incremental learning benchmarks frequently utilized in video analysis, as outlined in TCD [22]. + +Table 1: Video class incremental learning results on UCF101 and HMDB51. The best result for each setting is emphasized in bold. + +
MethodUCF101HMDB51
10 × 5 stages5 × 10 stages2 × 25 stages5 × 5 stages1 × 25 stages
LwFMC [16]42.1425.5911.6826.8216.49
LwM [6]43.3926.0712.0826.9716.50
UCIR [12]74.3170.4263.2244.9037.04
PODNet [8]73.2671.5870.2844.3238.76
TCD [22]74.8973.4372.1945.3440.47
FrameMaker [23]78.1376.3875.7747.5442.65
STSP (Ours)81.1582.8479.2556.9949.19
+ +This approach begins with a model pre-trained on half of the total number of classes, with the remaining classes introduced sequentially at each incremental step. Specifically, for UCF101, the initial training involves 51 classes, after which the subsequent 50 classes are structured into incremental learning tasks of 5, 10, and 25 classes. In the case of HMDB51, the model begins with training on 26 classes, followed by the distribution of the remaining classes into groups of 5 and 25 for subsequent learning phases. For Something-Something V2, the initial training phase covers 84 classes, with the remaining classes organized into incremental learning groups of 10 and 5 classes. + +Implementation Details. Our framework leverages the official implementation of TSM [17], utilizing the PyTorch Library for development. We train a ResNet-34 TSM model on UCF101 for 50 epochs with a batch size of 32. For both HMDB51 and Something-Something V2, ResNet-50 TSM models are employed, undergoing training for 50 epochs with a batch size of 32. + +# 4.2 Main Results + +For a fair comparison, we employ the same model structure with the same pretraining initialization as the existing methods [22, 23]. Unlike prior approaches that incorporate exemplars for incremental learning and assess performance using both the standard classification protocol and Nearest Mean of Exemplars (NME) metrics [26], our method does not retain any exemplars. Consequently, we only report the standard classification results. We compare our proposed STSP method with the existing video-class incremental learning methods over three benchmarks, as shown in Tab. 1 and Tab. 2. From both tables, we observe that the proposed STSP method significantly outperforms all competing methods across all settings on each dataset. Notably, unlike other approaches that rely on retaining exemplars for incremental learning sessions, our method utilizes gradient projection to constrain model updates, effectively safeguarding against performance drops on previously learned classes. + +Table 2: Video class incremental learning results on Something-Something V2. The best result for each setting is emphasized in bold. + +
Method10 × 9 stages5 × 18 stages
UCIR [12]26.8420.69
PODNet [8]34.9726.95
TCD [22]35.7829.60
FrameMaker [23]37.2530.98
STSP (Ours)69.6870.87
+ +Table 3: Ablation studies for our proposed STSP method on UCF101 with $5 \times 10$ stages and $2 \times 25$ stages settings, respectively. + +
TSCSGPExemplarsUCF101
5 × 10 stages2 × 25 stages
(i)××5/class68.2857.39
(ii)✓×5/class71.5061.78
(iii)××017.6319.22
(iv)✓×08.7814.99
(v)×✓077.6674.02
(vi)✓✓082.8479.25
+ +![](images/f448247ea081d492cf4b2f6a773e7c6a8a7d713f0c9b73ed1d85be9c26a1a0f4.jpg) +(a) UCF101 (5 × 10 stages) + +![](images/07d7d1b95cf4a0e3ba7fbd1f8033c9086d61a878a5e6599d581b72cdecd51932.jpg) +(b) UCF101 (2 × 25 stages) +Fig. 3: Plots of accuracy over seen classes along with the incremental steps on UCF101. + +# 4.3 Ablation Study and Analysis + +In this section, we perform ablation studies to investigate the effectiveness of the key components of the proposed STSP on the UCF101 dataset under both $2 \times 25$ stages and $5 \times 10$ stages settings. + +Effectiveness of gradient projection To evaluate the SGP's capability to mitigate catastrophic forgetting, we designed experiments with various configurations: utilizing the method with SGP, the method without SGP, and the method without SGP but incorporating exemplars. In the last configuration, we preserved 5 videos, each containing 8 frames, as exemplars for prior classes to be used in future training sessions. Furthermore, we extended these comparisons to assess performance by employing a linear classifier and the proposed TSC. The comparison results are shown in Tab. 3, where (i), (iii), and (v) utilize the linear classifier, while (ii), (iv), and (vi) implement the TSC. In cooperation with the comparative analysis of the network's performance as the number of seen classes increases, as shown in Fig. 3, it can be found that the network tends to forget most knowledge of previous classes when no strategy is adopted. + +Moreover, when comparing our SGP strategy against the exemplar-based distillation method, particularly through the comparison of (i) $v.s.$ (v) and the comparison of (ii) $v.s.$ (vi) in Tab. 3, we can find that our proposed SGP strategy + +Table 4: Ablation study of the subspace basis size $\tau$ for TSC on UCF101 with $5\times 10$ stages setting. + +
τ=1τ=2τ=4τ=8τ=16
STSP (Ours)66.7567.8875.3982.8463.02
+ +significantly outperforms the exemplar-based distillation method. Specifically, regarding the performance of base classes during incremental learning sessions, as shown in Fig. 4, it is evident that our proposed SGP has minimal impact on the classification performance for the base classes. These findings indicate that our proposed SGP can uphold accurate classification across all classes without compromising performance on the base classes, emphasizing the effectiveness of SGP in mitigating knowledge degradation. + +Furthermore, based on the comparison in Fig. 3 and Fig. 4, we observe that all methods perform worse under the setting of $2 \times 25$ stages, while our method with SGP still achieves the best performance. A plausible explanation is that the model may be susceptible to overfitting when dealing with a limited set of novel classes. Moreover, for our method, the computation of the null space becomes more unstable in these scenarios, especially when updating the mean covariance matrix is solely based on a few classes. + +Linear or Subspace-based Classifier? To assess the effectiveness of the proposed TSC, we compare it with a linear classifier based on an FC layer. Moreover, this ablation study extends to both the exemplar-based distillation method and our proposed SGP, with the comparison results shown in Tab. 3. For the exemplar-based distillation approach, the implementation of TSC leads to a performance increase of $3.22\%$ and $4.39\%$ on UCF101 at the 10-stage and 25-stage settings, respectively, as evidenced by the comparison between (i) and (ii). In the context of SGP, the comparison between (v) and (vi) reveals that the proposed TSC similarly improves the overall performance by $5.18\%$ and $5.23\%$ for the 10-stage and 25-stage settings on UCF101, respectively. Compared with the exemplar-based distillation method, our method employs SGP to constrain the classifier update through gradient subspace projection. The above findings align with the concept of TSC, collectively constraining the subspace-based classifier and further enhancing classification performance. + +From the comparison in Fig. 4, another observation is that TSC can even boost the classification performance on the base classes, demonstrating its effectiveness in modeling class representations. However, we also observe that TSC may perform poorly when there are no constraints (i.e. without exemplars nor SGP) on the network. One plausible explanation for this phenomenon is that, during incremental sessions, the network tends to shift the feature space toward new tasks without constraints from historical information. Such adjustments can result in the original subspace deviating from its previous configuration, leading to significant misalignment that greatly impacts the performance of TSC. + +![](images/e12369de4149b125837e8d1cf7e00e6f1db461bf46d3b054d2e714a9fe1b1b61.jpg) +(a) UCF101 $(5\times 10$ stages) + +![](images/0171a0332d9e8d19f03421da3bd4fc01d980d684a015deee4823b40c531b01f5.jpg) +(b) UCF101 (2 × 25 stages) +Fig. 4: Plots of accuracy over base classes along with the incremental steps on UCF101. + +The Size of the Subspace Basis We explore how the size of the subspace basis influences the classification accuracy of our proposed method, the comparison results are shown in Tab. 4. It can be found that large subspace basis sizes (i.e., $\tau = 4$ or $\tau = 8$ ) are more effective at constructing the subspace for each category, enabling the capture of more distinctive features, which in turn leads to better performance as opposed to smaller bases. This improvement, however, results in an increase in both the network's parameters and computational complexity. Additionally, we can observe that a subspace basis size of 16 yields the least favorable results. This could be attributed to the difficulty in constructing orthogonal subspaces for new classes when the subspace basis is excessively large, especially as $M \times \tau$ approaches the total number of classes $(\mathcal{C})$ . + +# 5 Conclusion + +In this paper, we proposed STSP, a novel spatial-temporal subspace projection method designed for video class-incremental learning tasks. The objective is to learn discriminative and robust representations for both video and class, facilitating further incremental learning. STSP mainly consists of two key components, i.e., temporal-base subspace classifier (TSC) and spatial-based gradient projection (SGP). TSC employs a subspace to model each video and its class, incorporating intra-frame temporal information. On the other hand, SGP is designed to mitigate catastrophic forgetting by projecting the gradient of network parameters in each layer onto the subspace spanned by the corresponding input spatial features. Extensive experiments on three benchmarks demonstrate the effectiveness of our proposed STSP method. + +Limitations. In contrast to previous video class incremental learning methods, our proposed STSP method incorporates gradient projection to prevent the forgetting of historical information from previous classes. This strategy removes the necessity of storing exemplars and prototypes of previously learned classes, yet it requires the storage and updating of input features for both linear layers and the introduced TSC. Such a requirement could pose challenges when employing large backbones, requiring further exploration and discussion. + +# Acknowledgements + +This research was carried out at the Rapid-Rich Object Search (ROSE) Lab at Nanyang Technological University in Singapore. + +# References + +1. Akyurek, A.F., Akyurek, E., Wijaya, D., Andreas, J.: Subspace regularizers for few-shot class incremental learning. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=boJy41J-tnQ +2. Belhumeur, P.N., Hespanha, J.P., Kriegman, D.J.: Eigenfaces vs. fisherfaces: Recognition using class specific linear projection. IEEE Transactions on pattern analysis and machine intelligence 19(7), 711-720 (1997) +3. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6299-6308 (2017) +4. Chen, H., Wang, Y., Hu, Q.: Multi-granularity regularized re-balancing for class incremental learning. IEEE Transactions on Knowledge and Data Engineering 35(7), 7263-7277 (2023) +5. Cheraghian, A., Rahman, S., Ramasinghe, S., Fang, P., Simon, C., Petersson, L., Harandi, M.: Synthesized feature based few-shot class-incremental learning on a mixture of subspaces. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 8661-8670 (2021) +6. Dhar, P., Singh, R.V., Peng, K.C., Wu, Z., Chellappa, R.: Learning without memorizing. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5138-5146 (2019) +7. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=YicbFdNTTy +8. Douillard, A., Cord, M., Ollion, C., Robert, T., Valle, E.: Podnet: Pooled outputs distillation for small-tasks incremental learning. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XX 16. pp. 86-102. Springer (2020) +9. Feichtenhofer, C., Fan, H., Malik, J., He, K.: Slowfast networks for video recognition. In: Proceedings of the IEEE international conference on computer vision. pp. 6202-6211 (2019) +0. Gao, Q., Zhao, C., Ghanem, B., Zhang, J.: R-dfcil: Relation-guided representation learning for data-free class incremental learning. In: European Conference on Computer Vision. pp. 423-439. Springer (2022) +1. Goyal, R., Ebrahimi Kahou, S., Michalski, V., Materzynska, J., Westphal, S., Kim, H., Haenel, V., Fruend, I., Yianilos, P., Mueller-Freitag, M., et al.: The" something something" video database for learning and evaluating visual common sense. In: Proceedings of the IEEE international conference on computer vision. pp. 5842-5850 (2017) +2. Hou, S., Pan, X., Loy, C.C., Wang, Z., Lin, D.: Learning a unified classifier incrementally via rebalancing. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 831-839 (2019) + +13. Huang, H., Wang, Y., Hu, Q., Cheng, M.M.: Class-specific semantic reconstruction for open set recognition. IEEE transactions on pattern analysis and machine intelligence 45(4), 4214-4228 (2023) +14. Ji, S., Xu, W., Yang, M., Yu, K.: 3d convolutional neural networks for human action recognition. IEEE transactions on pattern analysis and machine intelligence 35(1), 221-231 (2012) +15. Kuehne, H., Jhuang, H., Garrote, E., Poggio, T., Serre, T.: Hmdb: a large video database for human motion recognition. In: 2011 International conference on computer vision. pp. 2556-2563. IEEE (2011) +16. Li, Z., Hoiem, D.: Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence 40(12), 2935-2947 (2017) +17. Lin, J., Gan, C., Han, S.: Temporal shift module for efficient video understanding. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7083-7093 (2019) +18. Liu, C.: Gabor-based kernel pca with fractional power polynomial models for face recognition. IEEE transactions on pattern analysis and machine intelligence 26(5), 572-581 (2004) +19. Masana, M., Liu, X., Twardowski, B., Menta, M., Bagdanov, A.D., Van De Weijer, J.: Class-incremental learning: survey and performance evaluation on image classification. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(5), 5513-5533 (2022) +20. McCloskey, M., Cohen, N.J.: Catastrophic interference in connectionist networks: The sequential learning problem. In: Psychology of learning and motivation, vol. 24, pp. 109-165. Elsevier (1989) +21. Papyan, V., Han, X., Donoho, D.L.: Prevalence of neural collapse during the terminal phase of deep learning training. Proceedings of the National Academy of Sciences 117(40), 24652-24663 (2020) +22. Park, J., Kang, M., Han, B.: Class-incremental learning for action recognition in videos. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 13698-13707 (2021) +23. Pei, Y., Qing, Z., Cen, J., Wang, X., Zhang, S., Wang, Y., Tang, M., Sang, N., Qian, X.: Learning a condensed frame for memory-efficient video class-incremental learning. Advances in Neural Information Processing Systems 35, 31002-31016 (2022) +24. Pei, Y., Qing, Z., Zhang, S., Wang, X., Zhang, Y., Zhao, D., Qian, X.: Space-time prompting for video class-incremental learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11932-11942 (2023) +25. Peifeng, G., Xu, Q., Wen, P., Yang, Z., Shao, H., Huang, Q.: Feature directions matter: Long-tailed learning via rotated balanced representation. In: International Conference on Machine Learning. pp. 27542-27563. PMLR (2023) +26. Rebuffi, S.A., Kolesnikov, A., Sperl, G., Lampert, C.H.: icarl: Incremental classifier and representation learning. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 2001-2010 (2017) +27. Saha, G., Garg, I., Roy, K.: Gradient projection memory for continual learning. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=3A0jORCNC2 +28. Simon, C., Koniusz, P., Nock, R., Harandi, M.: Adaptive subspaces for few-shot learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4136-4145 (2020) + +29. Simonyan, K., Zisserman, A.: Two-stream convolutional networks for action recognition in videos. In: Advances in neural information processing systems. pp. 568-576 (2014) +30. Soomro, K., Zamir, A.R., Shah, M.: Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402 (2012) +31. Tran, D., Bourdev, L., Fergus, R., Torresani, L., Paluri, M.: Learning spatiotemporal features with 3d convolutional networks. In: Proceedings of the IEEE international conference on computer vision. pp. 4489-4497 (2015) +32. Tran, D., Wang, H., Torresani, L., Ray, J., LeCun, Y., Paluri, M.: A closer look at spatiotemporal convolutions for action recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6450-6459 (2018) +33. Turk, M.A., Pentland, A.P.: Face recognition using eigenfaces. In: Proceedings. 1991 IEEE computer society conference on computer vision and pattern recognition. pp. 586-587. IEEE Computer Society (1991) +34. Tzimiropoulos, G., Zafeiriou, S., Pantic, M.: Subspace learning from image gradient orientations. IEEE transactions on pattern analysis and machine intelligence 34(12), 2454-2466 (2012) +35. Villa, A., Alcázar, J.L., Alfarra, M., Alhamoud, K., Hurtado, J., Heilbron, F.C., Soto, A., Ghanem, B.: Pivot: Prompting for video continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24214-24223 (2023) +36. Villa, A., Alhamoud, K., Escorcia, V., Caba, F., Alcazar, J.L., Ghanem, B.: vclimb: A novel video class incremental learning benchmark. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19035-19044 (2022) +37. Wang, C., Guo, L., Wang, Y., Cheng, H., Yu, Y., Wen, B.: Progressive divide-and-conquer via subsampling decomposition for accelerated mri. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 25128-25137 (2024) +38. Wang, J., Cherian, A.: Gods: Generalized one-class discriminative subspaces for anomaly detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8201-8211 (2019) +39. Wang, S., Li, X., Sun, J., Xu, Z.: Training networks in null space of feature covariance for continual learning. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 184-193 (2021) +40. Yang, J., Frangi, A.F., Yang, J.y., Zhang, D., Jin, Z.: Kpca plus lda: a complete kernel fisher discriminant framework for feature extraction and recognition. IEEE Transactions on pattern analysis and machine intelligence 27(2), 230-244 (2005) +41. Yang, S., Liu, J., Lu, S., Er, M.H., Kot, A.C.: Collaborative learning of gesture recognition and 3d hand pose estimation with multi-order feature analysis. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16. pp. 769-786. Springer (2020) +42. Yang, Y., Yuan, H., Li, X., Lin, Z., Torr, P., Tao, D.: Neural collapse inspired feature-classifier alignment for few-shot class-incremental learning. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=y5W8tpoJhtJ +43. Zhang, L., Fu, J., Wang, S., Zhang, D., Dong, Z., Chen, C.P.: Guide subspace learning for unsupervised domain adaptation. IEEE transactions on neural networks and learning systems 31(9), 3374-3388 (2019) + +44. Zhao, B., Xiao, X., Gan, G., Zhang, B., Xia, S.T.: Maintaining discrimination and fairness in class incremental learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 13208-13217 (2020) +45. Zhu, H., Koniusz, P.: Ease: Unsupervised discriminant subspace learning for transductive few-shot learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9078-9088 (2022) +46. Zhu, P., Yao, X., Wang, Y., Hui, B., Du, D., Hu, Q.: Multiview deep subspace clustering networks. IEEE Transactions on Cybernetics (2024) \ No newline at end of file diff --git a/2024/STSP_ Spatial-Temporal Subspace Projection for Video Class-incremental Learning/images.zip b/2024/STSP_ Spatial-Temporal Subspace Projection for Video Class-incremental Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8498dca3c0a7c486f3c928a13ed07bd6c6b02f66 --- /dev/null +++ b/2024/STSP_ Spatial-Temporal Subspace Projection for Video Class-incremental Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f60145ad60d62474f710ae5a291a917dfda73ad9440f363996a0985f8d802f74 +size 277925 diff --git a/2024/STSP_ Spatial-Temporal Subspace Projection for Video Class-incremental Learning/layout.json b/2024/STSP_ Spatial-Temporal Subspace Projection for Video Class-incremental Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..84fd35ff410a714b9a38ec30a37e6e9315d03ae5 --- /dev/null +++ b/2024/STSP_ Spatial-Temporal Subspace Projection for Video Class-incremental Learning/layout.json @@ -0,0 +1,10319 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 136, + 111, + 477, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 111, + 477, + 148 + ], + "spans": [ + { + "bbox": [ + 136, + 111, + 477, + 148 + ], + "type": "text", + "content": "STSP: Spatial-Temporal Subspace Projection for Video Class-incremental Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 135, + 167, + 478, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 167, + 478, + 194 + ], + "spans": [ + { + "bbox": [ + 135, + 167, + 478, + 194 + ], + "type": "text", + "content": "Hao Cheng" + }, + { + "bbox": [ + 135, + 167, + 478, + 194 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 135, + 167, + 478, + 194 + ], + "type": "text", + "content": ", Siyuan Yang" + }, + { + "bbox": [ + 135, + 167, + 478, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 135, + 167, + 478, + 194 + ], + "type": "text", + "content": ", Chong Wang" + }, + { + "bbox": [ + 135, + 167, + 478, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 135, + 167, + 478, + 194 + ], + "type": "text", + "content": ", Joey Tianyi Zhou" + }, + { + "bbox": [ + 135, + 167, + 478, + 194 + ], + "type": "inline_equation", + "content": "^{3,4}" + }, + { + "bbox": [ + 135, + 167, + 478, + 194 + ], + "type": "text", + "content": ", Alex C. Kot" + }, + { + "bbox": [ + 135, + 167, + 478, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 135, + 167, + 478, + 194 + ], + "type": "text", + "content": ", and Bihan Wen" + }, + { + "bbox": [ + 135, + 167, + 478, + 194 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 187, + 201, + 425, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 201, + 425, + 224 + ], + "spans": [ + { + "bbox": [ + 187, + 201, + 425, + 224 + ], + "type": "text", + "content": "1Nanyang Technological University, Singapore {siyuan.yang, wang1711, eackot, bihan.wen}@ntu.edu.sg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 205, + 224, + 407, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 224, + 407, + 244 + ], + "spans": [ + { + "bbox": [ + 205, + 224, + 407, + 244 + ], + "type": "text", + "content": "2 Hebei University of Technology, Tianjin, China 2024025@hebut.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 171, + 245, + 441, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 245, + 441, + 257 + ], + "spans": [ + { + "bbox": [ + 171, + 245, + 441, + 257 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 171, + 245, + 441, + 257 + ], + "type": "text", + "content": " CFAR, Agency for Science, Technology and Research, Singapore" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 172, + 257, + 441, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 257, + 441, + 279 + ], + "spans": [ + { + "bbox": [ + 172, + 257, + 441, + 279 + ], + "type": "text", + "content": "4 IHPC, Agency for Science, Technology and Research, Singapore zhouty@cfar.a-star.edu.sg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 160, + 300, + 455, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 300, + 455, + 542 + ], + "spans": [ + { + "bbox": [ + 160, + 300, + 455, + 542 + ], + "type": "text", + "content": "Abstract. Video class-incremental learning (VCIL) aims to learn discriminative and generalized feature representations for video frames to mitigate catastrophic forgetting. Conventional VCIL methods often retain a subset of frames or features from prior tasks as exemplars for subsequent incremental learning stages. However, these strategies overlook the connection between base and novel classes, sometimes even leading to privacy leakage. To address this challenge, we introduce a Spatial-Temporal Subspace Projection (STSP) scheme for VCIL. Specifically, we propose a discriminative Temporal-based Subspace Classifier (TSC) that represents each class with an orthogonal subspace basis and adopts subspace projection loss for classification. Unlike typical classification methods relying on fully connected layers, our TSC discerns the spatial-temporal dynamics in video content, thereby enhancing the representation of each video sample. Additionally, we implement inter- and intra-class orthogonal constraints into TSC, ensuring each class occupies a unique orthogonal subspace, defined by its basis. To prevent catastrophic forgetting, we further employ a Spatial-based Gradient Projection (SGP) strategy, adjusting network gradients to align with the null space of the spatial feature set from previous tasks. Extensive experiments conducted on three benchmarks, namely HMDB51, UCF101, and Something-Something V2, demonstrate that our STSP method outperforms state-of-the-art comparison methods, evidencing its efficacy in VCIL." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 159, + 552, + 453, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 552, + 453, + 574 + ], + "spans": [ + { + "bbox": [ + 159, + 552, + 453, + 574 + ], + "type": "text", + "content": "Keywords: Video action recognition " + }, + { + "bbox": [ + 159, + 552, + 453, + 574 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 552, + 453, + 574 + ], + "type": "text", + "content": " Class incremental learning " + }, + { + "bbox": [ + 159, + 552, + 453, + 574 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 552, + 453, + 574 + ], + "type": "text", + "content": " Subspace Projection " + }, + { + "bbox": [ + 159, + 552, + 453, + 574 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 552, + 453, + 574 + ], + "type": "text", + "content": " Gradient Projection" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 590, + 230, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 230, + 604 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 230, + 604 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 613, + 482, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 613, + 482, + 650 + ], + "spans": [ + { + "bbox": [ + 130, + 613, + 482, + 650 + ], + "type": "text", + "content": "Deep neural networks have demonstrated remarkable performance in modern computer vision tasks [7,17,32,37,46], particularly when trained on large, annotated datasets. However, the real-world challenge of incrementally incorporating" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 654, + 236, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 654, + 236, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 654, + 236, + 666 + ], + "type": "text", + "content": "* Corresponding author." + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 146, + 115, + 302, + 191 + ], + "blocks": [ + { + "bbox": [ + 146, + 115, + 302, + 191 + ], + "lines": [ + { + "bbox": [ + 146, + 115, + 302, + 191 + ], + "spans": [ + { + "bbox": [ + 146, + 115, + 302, + 191 + ], + "type": "image", + "image_path": "173df513cde9670e818f909bb4b61e1031d81913c37529a5d66ea38e36c0f6dc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 201, + 482, + 257 + ], + "lines": [ + { + "bbox": [ + 130, + 201, + 482, + 257 + ], + "spans": [ + { + "bbox": [ + 130, + 201, + 482, + 257 + ], + "type": "text", + "content": "Fig. 1: Class incremental learning with (a) linear classifier and (b) subspace-based classifier. Incorporating new classes alters the decision hyperspace and modifies the representation of prior classes, leading to misalignment. In contrast, subspace-based classifiers, by constructing an orthogonal basis for each class, establish a unique geometric structure for each class, thereby reducing class misalignment." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 306, + 116, + 471, + 191 + ], + "blocks": [ + { + "bbox": [ + 306, + 116, + 471, + 191 + ], + "lines": [ + { + "bbox": [ + 306, + 116, + 471, + 191 + ], + "spans": [ + { + "bbox": [ + 306, + 116, + 471, + 191 + ], + "type": "image", + "image_path": "eed59928631da623d347d013762b457e615a6324bed1440d775135688415c2fb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 281, + 482, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 365 + ], + "type": "text", + "content": "data from new classes into a model initially trained on existing classes presents significant hurdles, known as class-incremental learning (CIL) [26]. Unlike typical classification tasks, CIL involves accessing novel classes and limited memories of prior tasks in incremental sessions, often due to privacy concerns or memory constraints. Consequently, merely updating the model with data from new classes can result in overfitting substantial performance degradation on previous classes, a phenomenon referred to as catastrophic forgetting [19, 20]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 365, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 628 + ], + "type": "text", + "content": "Recent advancements in CIL methods [1,4,26,44] have shown remarkable performance in image classification tasks. However, recent CIL studies [1,21,25,42] demonstrate that straightforward fine-tuning of the network for novel classes may lead the subspace spanned by the novel classifier to deviate from the base one, as illustrated in Fig. 1 (a). The classification performance can be significantly affected by misalignment, particularly in cases where a notable distribution gap between novel and base classes. Moreover, when extending this task to Video Class-Incremental Learning (VCIL), unique challenges due to the inherent information redundancy within video samples, such as continuous actions with a static background. This necessitates methods to model temporal-spatial relationships between video frames to extract critical information for action recognition properly. However, prevalent approaches [22-24,35,36] in VCIL predominantly involve a learning pipeline that employs simple operations (e.g., mean) on the feature vectors of sampled frames to generate video representations for action recognition. While this methodology is effective for standard action recognition with sufficient data, it may struggle to ensure robust generalization in scenarios of incremental learning. The aforementioned observations prompt us to consider acquiring a more suitable representation for both the classifier and video features while also exploring their relationship. To address these challenges, several subspace learning-based methods [28,33,38,45] have shown effectiveness in reducing dimensionality and capturing intrinsic structure, thereby offering a potential solution to the issues associated with simplistic video representation methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we propose a spatial-temporal subspace projection method for VCIL including a temporal-based subspace classifier (TSC) and a spatial-based gradient projection (SGP) strategy. Specifically, we extend the existing linear" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "H. Cheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 295 + ], + "type": "text", + "content": "classifiers based on fully connected (FC) layers to a TSC in which a subspace basis is adopted to model each class. Differing from linear classifiers that seek to distinguish classes in feature space through a linear decision boundary, the proposed orthogonality constraints offer a distinct geometric structure for each class. This characteristic not only facilitates a more comprehensive representation of each class but also allows for straightforward extension to novel classes. Furthermore, to avoid catastrophic forgetting, we utilize the SGP strategy to constrain the process of gradient update by projecting the gradient of network parameters in each layer onto the subspace of corresponding input features so that losses for the old task will not increase during the incremental sessions. Compared with existing VCIL methods [22-24,35,36] that rely on feature distillation or prompt-based fine-tuning strategies, SGP offers the advantage of fine-tuning all network parameters without necessitating regularizers. Furthermore, SGP obviates the need to explicitly retain class-relevant information, such as frames or prototypes, thereby contributing to privacy protection." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 296, + 482, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 296, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 482, + 403 + ], + "type": "text", + "content": "In summary, we make the following contributions: 1) We propose a spatial-temporal subspace projection (STSP) method, containing a temporal-based subspace classifier and spatial-based gradient projection strategy to better model feature and class representation for discriminative classification. 2) We employ a spatial-based gradient projection strategy to constrain the gradient update process without the need to store frames or prototypes from previous tasks. This not only enhances performance but also mitigates the risk of privacy leakage. 3) The proposed STSP method achieves superior results on three video action recognition incremental benchmarks over state-of-the-art methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 420, + 237, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 420, + 237, + 432 + ], + "spans": [ + { + "bbox": [ + 132, + 420, + 237, + 432 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 444, + 291, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 444, + 291, + 455 + ], + "spans": [ + { + "bbox": [ + 132, + 444, + 291, + 455 + ], + "type": "text", + "content": "2.1 Video Action Recognition" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "content": "With significant advances in deep learning, current video action recognition approaches aim to learn effective spatial-temporal representation by well-designed network architectures. A notable advancement by Simonyan and Zisserman [29] introduces a dual-stream architecture that analyzes RGB frames in one stream and optical flow in another, laying the foundation for subsequent enhancements. Recently, Lin et al. [17] proposed a Temporal Shift Module (TSM), which shifts some channels along the temporal dimension to perform temporal interaction between the features from adjacent frames. Concurrently, significant research [3, 13, 14, 31, 32, 41] has evolved 2D Convolutional Neural Networks (CNNs) into 3D structures to capture both spatial and temporal contexts within videos, a critical aspect of human action recognition. Notably, Tran et al. [31] introduced the C3D model, a 3D CNN designed to learn spatial-temporal features through an end-to-end learning approach. Carreira et al. [3] developed the Inflated 3D CNN, which extends a 2D CNN into the temporal domain by inflating its convolutional and pooling layers. Additionally, Feichtenhofer et al. [9] crafted a dual-pathway 3D CNN, featuring both slow and fast pathways that process RGB frames at varying frame rates to simultaneously capture semantic content and motion details." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 250, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 250, + 90, + 447, + 102 + ], + "type": "text", + "content": "Spatial-Temporal Subspace Projection for VCIL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "content": "Despite significant progress in action recognition, catastrophic forgetting has received relatively limited attention. This work aims to address this gap by developing a spatial-temporal subspace projection method, which aims to offer a robust solution to the challenges of incremental learning in action recognition." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 181, + 299, + 193 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 181, + 299, + 193 + ], + "spans": [ + { + "bbox": [ + 131, + 181, + 299, + 193 + ], + "type": "text", + "content": "2.2 Class-Incremental Learning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 199, + 482, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 199, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 199, + 482, + 319 + ], + "type": "text", + "content": "CIL addresses the challenge of adapting machine learning models to recognize new classes without forgetting previously learned information, a phenomenon known as catastrophic forgetting. Pioneering in this domain, Rebuffi et al. [26] introduced iCaRL, which combines representation learning with a memory mechanism to selectively retain the knowledge of previous classes. Following this, Hou et al. [12] proposed a UCIR strategy, which addresses the class imbalance in CIL by adjusting the classification layer to focus more on newer classes while still remembering old ones. Moreover, Gao et al. [10] introduced R-DFCIL, employing relation-guided representation learning to overcome catastrophic forgetting and enable the assimilation of new classes in a data-free scenario." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 319, + 482, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 482, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 482, + 415 + ], + "type": "text", + "content": "While CIL has been well-studied in the image domain, its application to the video domain remains largely unexplored. Park et al. [22] introduced Time-Channel Distillation (TCD), leveraging knowledge distillation loss with time-channel importance masks. Pei et al. [23] proposed FrameMaker, a memory-efficient VCIL strategy that develops a condensed frame for each video to facilitate future incremental tasks with less memory usage. Pei et al. [24] introduced ST-Prompt, a novel space-time prompting framework tailored for VCIL, enhancing the model's adaptability to new classes with minimal memory increase." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 415, + 482, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 482, + 487 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 482, + 487 + ], + "type": "text", + "content": "Existing VCIL approaches often rely on preserving a subset of frames or features from previous tasks as exemplars for subsequent incremental stages to prevent catastrophic forgetting. However, these strategies overlook the connection between base and novel classes and potentially lead to privacy leakage. In contrast, we employ a spatial-based gradient projection to regulate gradient updates, thereby mitigating catastrophic forgetting without these drawbacks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 502, + 288, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 502, + 288, + 514 + ], + "spans": [ + { + "bbox": [ + 131, + 502, + 288, + 514 + ], + "type": "text", + "content": "2.3 Subspace Representation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "Subspace learning, by learning a suitable representation for high dimensional data, has shed light on various vision tasks [2, 18, 34, 40, 43]. Existing subspace methods such as Principal Component Analysis (PCA) [2, 18] and Linear Discriminant Analysis (LDA) [2, 40] adaptively transfer the original input data to low dimensional subspace for dimensionality reduction. For example, Tzimiropoulos et al. [34] formulated PCA on image gradient orientations, which demonstrated robust performance in face recognition. Recently, subspace learning has been further extended to regularize the implicit features or learned classifiers within deep networks [1,28,38,45]. Notably, Cheraghian et al. [5] constructed a mixture of subspaces to represent the feature clusters formed by visually similar samples. Such subspace representation enables a better approximation of the overall data distribution, consequently leading to superior performance in" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "H. Cheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 121, + 470, + 205 + ], + "blocks": [ + { + "bbox": [ + 140, + 121, + 470, + 205 + ], + "lines": [ + { + "bbox": [ + 140, + 121, + 470, + 205 + ], + "spans": [ + { + "bbox": [ + 140, + 121, + 470, + 205 + ], + "type": "image", + "image_path": "969ada04ac68c9d7743347b61d21d25e5a17aaa79b0ee3b7f0e24a433ec9a6c5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 212, + 482, + 313 + ], + "lines": [ + { + "bbox": [ + 130, + 212, + 482, + 313 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 482, + 313 + ], + "type": "text", + "content": "Fig. 2: The overall framework of the proposed STSP method, including a temporal-based subspace classifier (TSC) and a spatial-based gradient projection (SGP). For each input video, spatial-temporal features are extracted using a backbone network. Subsequently, The TSC classifies by projecting temporal features into a subspace, where classification decisions are based on the distance to each class's subspace basis within this subspace. During parameter updates for the novel tasks, the SGP constrains gradient modifications for each specific layer, drawing on the mean covariance matrix " + }, + { + "bbox": [ + 130, + 212, + 482, + 313 + ], + "type": "inline_equation", + "content": "M_{t-1}" + }, + { + "bbox": [ + 130, + 212, + 482, + 313 + ], + "type": "text", + "content": " of previous tasks' input features. Upon the current session completion, the mean covariance matrix is updated to " + }, + { + "bbox": [ + 130, + 212, + 482, + 313 + ], + "type": "inline_equation", + "content": "M_t" + }, + { + "bbox": [ + 130, + 212, + 482, + 313 + ], + "type": "text", + "content": " for the gradient projection in the subsequent session." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 339, + 479, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 339, + 479, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 339, + 479, + 399 + ], + "type": "text", + "content": "CIL. To mitigate the catastrophic forgetting and overfitting, Simon et al. [28] formulated the dynamic classifiers via subspace modeling which demonstrated a significant robustness to noise in few-shot learning. Similarly, Akyurek et al. [1] proposed an additional subspace regularizer that pulls the weight vectors of the novel class close to the subspace spanned by the base classes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 401, + 480, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 401, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 480, + 449 + ], + "type": "text", + "content": "However, existing methods lack sufficient constraints on the representation of each class in the classifier. In this paper, we model each class via an orthogonal subspace basis and leverage spatial-based gradient projection to update the entire network during incremental sessions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 479, + 201, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 201, + 491 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 201, + 491 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 516, + 479, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 516, + 479, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 516, + 479, + 552 + ], + "type": "text", + "content": "In this section, we start with an introduction of VCIL problem formulation. Then, the proposed STSP method is described and explained in detail, followed by the loss function of our method as well as the training and inference procedure." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 555, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 555, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 555, + 482, + 666 + ], + "type": "text", + "content": "In VCIL, a series of distinct tasks " + }, + { + "bbox": [ + 130, + 555, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{T}_1,\\mathcal{T}_2,\\dots ,\\mathcal{T}_k,\\dots \\}" + }, + { + "bbox": [ + 130, + 555, + 482, + 666 + ], + "type": "text", + "content": " unfolds sequentially. Each task " + }, + { + "bbox": [ + 130, + 555, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_k" + }, + { + "bbox": [ + 130, + 555, + 482, + 666 + ], + "type": "text", + "content": " comprises a distinct subset of video-class pairs " + }, + { + "bbox": [ + 130, + 555, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\{(D_i^k,y_i^k),y_i^k\\in" + }, + { + "bbox": [ + 130, + 555, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_k\\} |_{i = 1}^{|\\mathcal{T}_k|}" + }, + { + "bbox": [ + 130, + 555, + 482, + 666 + ], + "type": "text", + "content": ". Specifically, the novel class set " + }, + { + "bbox": [ + 130, + 555, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_k" + }, + { + "bbox": [ + 130, + 555, + 482, + 666 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 130, + 555, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_k" + }, + { + "bbox": [ + 130, + 555, + 482, + 666 + ], + "type": "text", + "content": " is distinct from previous classes, i.e., " + }, + { + "bbox": [ + 130, + 555, + 482, + 666 + ], + "type": "inline_equation", + "content": "(\\mathcal{C}_1\\cup \\dots \\cup \\mathcal{C}_{k - 1})\\cap \\mathcal{C}_k = \\emptyset" + }, + { + "bbox": [ + 130, + 555, + 482, + 666 + ], + "type": "text", + "content": ". The primary objective is to incrementally train a deep neural network " + }, + { + "bbox": [ + 130, + 555, + 482, + 666 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 130, + 555, + 482, + 666 + ], + "type": "text", + "content": " capable of accurately classifying videos across all classes by preventing the loss of class information from previous tasks. In contrast to methods that have access to a limited exemplar set from historical tasks, our emphasis lies on exemplar-free VCIL, where the model is not permitted to store exemplars from previous tasks during the learning process of the current session." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 250, + 90, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 90, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 250, + 90, + 448, + 102 + ], + "type": "text", + "content": "Spatial-Temporal Subspace Projection for VCIL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 290, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 290, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 290, + 126 + ], + "type": "text", + "content": "3.1 Overall STSP Framework" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 133, + 482, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 133, + 482, + 277 + ], + "spans": [ + { + "bbox": [ + 130, + 133, + 482, + 277 + ], + "type": "text", + "content": "Given the problem formulation, we follow the standard learning protocol of previous methods [22, 23] based on knowledge distillation. During the " + }, + { + "bbox": [ + 130, + 133, + 482, + 277 + ], + "type": "inline_equation", + "content": "k^{th}" + }, + { + "bbox": [ + 130, + 133, + 482, + 277 + ], + "type": "text", + "content": " incremental session, the network " + }, + { + "bbox": [ + 130, + 133, + 482, + 277 + ], + "type": "inline_equation", + "content": "f(\\cdot ;\\mathbf{w})" + }, + { + "bbox": [ + 130, + 133, + 482, + 277 + ], + "type": "text", + "content": " parameterized by " + }, + { + "bbox": [ + 130, + 133, + 482, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 130, + 133, + 482, + 277 + ], + "type": "text", + "content": " is updated from " + }, + { + "bbox": [ + 130, + 133, + 482, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{k - 1}" + }, + { + "bbox": [ + 130, + 133, + 482, + 277 + ], + "type": "text", + "content": " to adapt to the novel class set " + }, + { + "bbox": [ + 130, + 133, + 482, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_k" + }, + { + "bbox": [ + 130, + 133, + 482, + 277 + ], + "type": "text", + "content": ". Fig. 2 illustrates the overall framework of our method. For an input video " + }, + { + "bbox": [ + 130, + 133, + 482, + 277 + ], + "type": "inline_equation", + "content": "D_{i}^{k}" + }, + { + "bbox": [ + 130, + 133, + 482, + 277 + ], + "type": "text", + "content": " belonging to " + }, + { + "bbox": [ + 130, + 133, + 482, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_k" + }, + { + "bbox": [ + 130, + 133, + 482, + 277 + ], + "type": "text", + "content": ", we first extract " + }, + { + "bbox": [ + 130, + 133, + 482, + 277 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 133, + 482, + 277 + ], + "type": "text", + "content": " frames as " + }, + { + "bbox": [ + 130, + 133, + 482, + 277 + ], + "type": "inline_equation", + "content": "V_{i}" + }, + { + "bbox": [ + 130, + 133, + 482, + 277 + ], + "type": "text", + "content": ", which are then fed to the backbone network to generate the feature embeddings. Subsequently, we employ a TSC to differentiate between distinct classes. Moreover, our method stands out by employing an SGP strategy without restoring frames or prototypes from previous tasks. More in detail, the network parameters are optimized by projecting the gradients in each layer to the subspace corresponding to their input features. With the merit of this strategy, we can completely avoid the risk of privacy leakage, whereas existing approaches may suffer." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 293, + 340, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 340, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 340, + 304 + ], + "type": "text", + "content": "3.2 Temporal-based Subspace Classifier" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 309, + 482, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 309, + 482, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 309, + 482, + 429 + ], + "type": "text", + "content": "In accordance with the standardized procedure for VCIL, given an input video-class pair denoted as " + }, + { + "bbox": [ + 130, + 309, + 482, + 429 + ], + "type": "inline_equation", + "content": "(D_i,y_i)" + }, + { + "bbox": [ + 130, + 309, + 482, + 429 + ], + "type": "text", + "content": ", we first randomly sampled " + }, + { + "bbox": [ + 130, + 309, + 482, + 429 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 309, + 482, + 429 + ], + "type": "text", + "content": " frames " + }, + { + "bbox": [ + 130, + 309, + 482, + 429 + ], + "type": "inline_equation", + "content": "V_{i}^{T}" + }, + { + "bbox": [ + 130, + 309, + 482, + 429 + ], + "type": "text", + "content": ". Subsequently, these frames are fed to the backbone network, yielding the feature embeddings as " + }, + { + "bbox": [ + 130, + 309, + 482, + 429 + ], + "type": "inline_equation", + "content": "X_{i}\\in \\mathbb{R}^{T\\times C\\times H\\times W}" + }, + { + "bbox": [ + 130, + 309, + 482, + 429 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 309, + 482, + 429 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 309, + 482, + 429 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 309, + 482, + 429 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 130, + 309, + 482, + 429 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 309, + 482, + 429 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 309, + 482, + 429 + ], + "type": "text", + "content": " represent the size of channel, height, and width, respectively. For the existing architectures [22-24] for VCIL, a prevalent technique for obtaining video representations is global average pooling (GAP), denoted as " + }, + { + "bbox": [ + 130, + 309, + 482, + 429 + ], + "type": "inline_equation", + "content": "Z_{i} = GAP(X_{i})\\in \\mathbb{R}^{C\\times T}" + }, + { + "bbox": [ + 130, + 309, + 482, + 429 + ], + "type": "text", + "content": ". Following this, a linear classifier " + }, + { + "bbox": [ + 130, + 309, + 482, + 429 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 309, + 482, + 429 + ], + "type": "text", + "content": ", comprising an FC layer with weights " + }, + { + "bbox": [ + 130, + 309, + 482, + 429 + ], + "type": "inline_equation", + "content": "\\mathbf{w}\\in \\mathbb{R}^{C\\times K}" + }, + { + "bbox": [ + 130, + 309, + 482, + 429 + ], + "type": "text", + "content": ", is utilized to predict the potential class of each frame. The class predictions from individual frames are subsequently averaged to produce the overall classification result." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 430, + 482, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 482, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 482, + 536 + ], + "type": "text", + "content": "Nevertheless, the temporal information holds the potential to encapsulate more intricate intra-frame structural details, warranting thoughtful consideration and seamless integration into the classification process. Additionally, in each incremental task, adjustments to the linear classifier may be necessary to modify the overall representation across all classes, which could influence the base classes for all previous incremental sessions. Conversely, a subspace-based classifier demonstrates heightened adaptability, enabling the modeling of each new class within its unique subspace without imposing substantial effects on the existing representations of previous classes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 537, + 482, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 537, + 482, + 644 + ], + "spans": [ + { + "bbox": [ + 131, + 537, + 482, + 644 + ], + "type": "text", + "content": "To this end, to properly model each class and learn a robust classifier, we construct a TSC based on subspaces spanned by an orthogonal basis of each class. Specifically, for " + }, + { + "bbox": [ + 131, + 537, + 482, + 644 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 131, + 537, + 482, + 644 + ], + "type": "text", + "content": "-th class, we construct a subspace with orthogonal basis " + }, + { + "bbox": [ + 131, + 537, + 482, + 644 + ], + "type": "inline_equation", + "content": "P_{m} \\in \\mathbb{R}^{C \\times \\tau}" + }, + { + "bbox": [ + 131, + 537, + 482, + 644 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 131, + 537, + 482, + 644 + ], + "type": "inline_equation", + "content": "P_{m}^{\\top} P_{m} = I_{\\tau}, \\tau \\ll C" + }, + { + "bbox": [ + 131, + 537, + 482, + 644 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 131, + 537, + 482, + 644 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 131, + 537, + 482, + 644 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 537, + 482, + 644 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 131, + 537, + 482, + 644 + ], + "type": "text", + "content": " indicate the feature dimension and the number of subspace basis for " + }, + { + "bbox": [ + 131, + 537, + 482, + 644 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 131, + 537, + 482, + 644 + ], + "type": "text", + "content": "-th class, respectively. Hence, for each task " + }, + { + "bbox": [ + 131, + 537, + 482, + 644 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_k" + }, + { + "bbox": [ + 131, + 537, + 482, + 644 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 131, + 537, + 482, + 644 + ], + "type": "inline_equation", + "content": "k^{th}" + }, + { + "bbox": [ + 131, + 537, + 482, + 644 + ], + "type": "text", + "content": " incremental session, we construct " + }, + { + "bbox": [ + 131, + 537, + 482, + 644 + ], + "type": "inline_equation", + "content": "|\\mathcal{C}_k|" + }, + { + "bbox": [ + 131, + 537, + 482, + 644 + ], + "type": "text", + "content": " distinct subspaces for " + }, + { + "bbox": [ + 131, + 537, + 482, + 644 + ], + "type": "inline_equation", + "content": "|\\mathcal{C}_k|" + }, + { + "bbox": [ + 131, + 537, + 482, + 644 + ], + "type": "text", + "content": " novel classes. We then adopt current subspaces to predict the class of " + }, + { + "bbox": [ + 131, + 537, + 482, + 644 + ], + "type": "inline_equation", + "content": "Z_{i}" + }, + { + "bbox": [ + 131, + 537, + 482, + 644 + ], + "type": "text", + "content": " based on the subspace distance between the subspace spanned by temporal video features " + }, + { + "bbox": [ + 131, + 537, + 482, + 644 + ], + "type": "inline_equation", + "content": "Z_{i}" + }, + { + "bbox": [ + 131, + 537, + 482, + 644 + ], + "type": "text", + "content": " and the orthogonal basis of each class as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 236, + 651, + 480, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 651, + 480, + 667 + ], + "spans": [ + { + "bbox": [ + 236, + 651, + 480, + 667 + ], + "type": "interline_equation", + "content": "d _ {m} (Z _ {i}) = - \\left\\| Z _ {i} Z _ {i} ^ {\\top} - P _ {m} P _ {m} ^ {\\top} \\right\\| _ {F} ^ {2}. \\tag {1}", + "image_path": "58a08e082e07c8e224e3c66568a06760f932386eb7bfb1dbc794aa71e642d021.jpg" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 101 + ], + "type": "text", + "content": "H. Cheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 481, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 481, + 139 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 481, + 139 + ], + "type": "text", + "content": "Then, the probability of " + }, + { + "bbox": [ + 131, + 116, + 481, + 139 + ], + "type": "inline_equation", + "content": "Z_{i}" + }, + { + "bbox": [ + 131, + 116, + 481, + 139 + ], + "type": "text", + "content": " assigned to " + }, + { + "bbox": [ + 131, + 116, + 481, + 139 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 131, + 116, + 481, + 139 + ], + "type": "text", + "content": "-th class can be calculated using the softmax function as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 249, + 147, + 481, + 177 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 147, + 481, + 177 + ], + "spans": [ + { + "bbox": [ + 249, + 147, + 481, + 177 + ], + "type": "interline_equation", + "content": "p _ {m} (Z _ {i}) = \\frac {e ^ {d _ {m} (Z _ {i})}}{\\sum_ {m ^ {\\prime} = 1} ^ {M} e ^ {d _ {m ^ {\\prime}} (Z _ {i})}}, \\tag {2}", + "image_path": "f5019534eb8e5070cfb9b4ebc91547486b7d6364c14eceec8fe54324c1f2fdef.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 183, + 479, + 208 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 183, + 479, + 208 + ], + "spans": [ + { + "bbox": [ + 131, + 183, + 479, + 208 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 183, + 479, + 208 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 131, + 183, + 479, + 208 + ], + "type": "text", + "content": " is the total number of classes in the current session. Now, we can employ the cross-entropy loss on Eq. (2) to update the network as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 251, + 217, + 481, + 249 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 217, + 481, + 249 + ], + "spans": [ + { + "bbox": [ + 251, + 217, + 481, + 249 + ], + "type": "interline_equation", + "content": "L _ {C E} = - \\sum_ {i = 1} ^ {J} y _ {i} \\log p _ {i} \\left(Z _ {i}\\right), \\tag {3}", + "image_path": "70c37fce8d2eb7c9415176707561b2e76ebfd19d78adbf3244168be70bca305b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 257, + 388, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 257, + 388, + 268 + ], + "spans": [ + { + "bbox": [ + 131, + 257, + 388, + 268 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 257, + 388, + 268 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 131, + 257, + 388, + 268 + ], + "type": "text", + "content": " denotes the number of videos in each mini-batch." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 269, + 479, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 269, + 479, + 328 + ], + "spans": [ + { + "bbox": [ + 130, + 269, + 479, + 328 + ], + "type": "text", + "content": "To encourage the network to capture discriminative information for each class and construct distinct subspaces corresponding to different classes, we impose orthogonal constraints on both the temporal features of the video and the weights of the TSC. Specifically, we employ orthogonal loss on the temporal features " + }, + { + "bbox": [ + 130, + 269, + 479, + 328 + ], + "type": "inline_equation", + "content": "Z_{i}" + }, + { + "bbox": [ + 130, + 269, + 479, + 328 + ], + "type": "text", + "content": " of each video " + }, + { + "bbox": [ + 130, + 269, + 479, + 328 + ], + "type": "inline_equation", + "content": "D_{i}" + }, + { + "bbox": [ + 130, + 269, + 479, + 328 + ], + "type": "text", + "content": " in the final layer as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 248, + 339, + 481, + 370 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 248, + 339, + 481, + 370 + ], + "spans": [ + { + "bbox": [ + 248, + 339, + 481, + 370 + ], + "type": "interline_equation", + "content": "L _ {O F} = \\sum_ {i = 1} ^ {J} \\left\\| Z _ {i} ^ {\\top} Z _ {i} - I _ {T} \\right\\| _ {F} ^ {2}, \\tag {4}", + "image_path": "73760da2187d67b98d959424ea7e9a67363343843ed54a1405b8fa53ec2f0ec7.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 375, + 479, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 375, + 479, + 423 + ], + "spans": [ + { + "bbox": [ + 131, + 375, + 479, + 423 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 375, + 479, + 423 + ], + "type": "inline_equation", + "content": "I_T \\in \\mathbb{R}^{T \\times T}" + }, + { + "bbox": [ + 131, + 375, + 479, + 423 + ], + "type": "text", + "content": " is an identity matrix. This approach guarantees that every frame independently captures distinct, class-aware information, setting it apart uniquely from the rest. Furthermore, it facilitates the construction of a discriminative subspace for each corresponding class." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 424, + 479, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 424, + 479, + 448 + ], + "spans": [ + { + "bbox": [ + 131, + 424, + 479, + 448 + ], + "type": "text", + "content": "For TSC, we impose intra- and inter-class orthogonal constraints on the subspace basis of each class, which is denoted by " + }, + { + "bbox": [ + 131, + 424, + 479, + 448 + ], + "type": "inline_equation", + "content": "L_{OW}" + }, + { + "bbox": [ + 131, + 424, + 479, + 448 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 200, + 457, + 481, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 457, + 481, + 495 + ], + "spans": [ + { + "bbox": [ + 200, + 457, + 481, + 495 + ], + "type": "interline_equation", + "content": "L _ {O W} = \\sum_ {m = 1} ^ {M} \\left(\\underbrace {\\left\\| P _ {m} ^ {\\top} P _ {m} - I _ {\\tau} \\right\\| _ {F} ^ {2}} _ {\\text {i n t r a - c l a s s}} + \\sum_ {q = m + 1} ^ {M} \\underbrace {\\left\\| P _ {m} ^ {\\top} P _ {q} \\right\\| _ {F} ^ {2}} _ {\\text {i n t e r - c l a s s}}\\right). \\tag {5}", + "image_path": "122da1e22899ac26f12ceeaf9562f4252a54ab673b6432aa556cd63e33eef837.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 502, + 479, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 502, + 479, + 609 + ], + "spans": [ + { + "bbox": [ + 131, + 502, + 479, + 609 + ], + "type": "text", + "content": "The intra-class orthogonal constraint is introduced to ensure independence among the basis vectors of subspaces, which is crucial for learning distinct and discriminative feature representations. Furthermore, the imposition of inter-class orthogonal constraint ensures that the basis vectors of distinct classes are orthogonal, signifying their mutual independence. This orthogonality enforcement serves to prevent confusion and safeguard the uniqueness of each basis vector, thus preserving the integrity of class information previously learned. Moreover, these orthogonal constraints enhance discriminability, guaranteeing that the basis vectors of subspaces related to different classes are clearly distinguishable." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 146, + 610, + 422, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 610, + 422, + 622 + ], + "spans": [ + { + "bbox": [ + 146, + 610, + 422, + 622 + ], + "type": "text", + "content": "Therefore, the overall objective function for TSC is defined as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 235, + 632, + 481, + 644 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 632, + 481, + 644 + ], + "spans": [ + { + "bbox": [ + 235, + 632, + 481, + 644 + ], + "type": "interline_equation", + "content": "L _ {t o t a l} = \\alpha L _ {C E} + \\beta L _ {O F} + \\gamma L _ {O W}, \\tag {6}", + "image_path": "dc5048388da1830ccf3d0099c40766c5347bd0d3797424efdf2caa14f1f315d3.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 653, + 479, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 653, + 479, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 653, + 479, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 653, + 479, + 666 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 131, + 653, + 479, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 653, + 479, + 666 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 131, + 653, + 479, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 653, + 479, + 666 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 131, + 653, + 479, + 666 + ], + "type": "text", + "content": " denote the weighting parameters of each losses, respectively." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 250, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 250, + 91, + 448, + 103 + ], + "type": "text", + "content": "Spatial-Temporal Subspace Projection for VCIL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 332, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 332, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 332, + 129 + ], + "type": "text", + "content": "3.3 Spatial-based Gradient Projection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 212 + ], + "type": "text", + "content": "In contrast to prior approaches [22-24] that rely on retaining exemplars or prompts from previous classes during incremental sessions, our focus is on addressing the challenging exemplar-free VCIL problems by imposing constraints on the network during parameter updates. To avoid catastrophic forgetting on previous classes in the incremental sessions, the network's parameter update should satisfy the following proposition:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 220, + 483, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 220, + 483, + 257 + ], + "spans": [ + { + "bbox": [ + 131, + 220, + 483, + 257 + ], + "type": "text", + "content": "Proposition 1. Given the data " + }, + { + "bbox": [ + 131, + 220, + 483, + 257 + ], + "type": "inline_equation", + "content": "x_{k}" + }, + { + "bbox": [ + 131, + 220, + 483, + 257 + ], + "type": "text", + "content": " from the previous task " + }, + { + "bbox": [ + 131, + 220, + 483, + 257 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_k" + }, + { + "bbox": [ + 131, + 220, + 483, + 257 + ], + "type": "text", + "content": ", the current network " + }, + { + "bbox": [ + 131, + 220, + 483, + 257 + ], + "type": "inline_equation", + "content": "f(\\cdot ;\\mathbf{w}_t)" + }, + { + "bbox": [ + 131, + 220, + 483, + 257 + ], + "type": "text", + "content": " could preserve the historical information if it satisfies the following equation for all sessions:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 225, + 267, + 481, + 281 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 267, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 225, + 267, + 481, + 281 + ], + "type": "interline_equation", + "content": "f \\left(x _ {k}; \\mathbf {w} _ {t}\\right) = f \\left(x _ {k}; \\mathbf {w} _ {k}\\right), \\quad \\forall k < t. \\tag {7}", + "image_path": "ad8cc6ff4715960cb4e7d96710d04d1d556eb6d2e4878f5e268cc2cd64b425e5.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 291, + 363, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 291, + 363, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 291, + 363, + 303 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 132, + 291, + 363, + 303 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_t" + }, + { + "bbox": [ + 132, + 291, + 363, + 303 + ], + "type": "text", + "content": " denotes the network parameters at task " + }, + { + "bbox": [ + 132, + 291, + 363, + 303 + ], + "type": "inline_equation", + "content": "T_{t}" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 312, + 482, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 482, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 482, + 373 + ], + "type": "text", + "content": "To realize the objective presented in Proposition 1, it is imperative to confine the gradient update of network parameters to a specific direction—precisely, the gradient update should lie in the subspace spanned by the corresponding input spatial features. Specifically, the gradients in the linear layers, including FC and Convolution layers " + }, + { + "bbox": [ + 130, + 312, + 482, + 373 + ], + "type": "inline_equation", + "content": "^5" + }, + { + "bbox": [ + 130, + 312, + 482, + 373 + ], + "type": "text", + "content": " should satisfy Proposition 2 [27,39]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 384, + 483, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 384, + 483, + 434 + ], + "spans": [ + { + "bbox": [ + 131, + 384, + 483, + 434 + ], + "type": "text", + "content": "Proposition 2. Let " + }, + { + "bbox": [ + 131, + 384, + 483, + 434 + ], + "type": "inline_equation", + "content": "f^l(\\cdot; \\mathbf{w}_t)" + }, + { + "bbox": [ + 131, + 384, + 483, + 434 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 131, + 384, + 483, + 434 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 131, + 384, + 483, + 434 + ], + "type": "text", + "content": "-th linear layer of the network trained on current task " + }, + { + "bbox": [ + 131, + 384, + 483, + 434 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_t" + }, + { + "bbox": [ + 131, + 384, + 483, + 434 + ], + "type": "text", + "content": " and its corresponding input feature " + }, + { + "bbox": [ + 131, + 384, + 483, + 434 + ], + "type": "inline_equation", + "content": "A_k^l" + }, + { + "bbox": [ + 131, + 384, + 483, + 434 + ], + "type": "text", + "content": " from any previous task " + }, + { + "bbox": [ + 131, + 384, + 483, + 434 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_k" + }, + { + "bbox": [ + 131, + 384, + 483, + 434 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 131, + 384, + 483, + 434 + ], + "type": "inline_equation", + "content": "k < t" + }, + { + "bbox": [ + 131, + 384, + 483, + 434 + ], + "type": "text", + "content": "). During each training step " + }, + { + "bbox": [ + 131, + 384, + 483, + 434 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 131, + 384, + 483, + 434 + ], + "type": "text", + "content": ", if the linear layer's parameter update " + }, + { + "bbox": [ + 131, + 384, + 483, + 434 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{w}_{t,s}^l" + }, + { + "bbox": [ + 131, + 384, + 483, + 434 + ], + "type": "text", + "content": " lies in the null space of " + }, + { + "bbox": [ + 131, + 384, + 483, + 434 + ], + "type": "inline_equation", + "content": "A_k^l" + }, + { + "bbox": [ + 131, + 384, + 483, + 434 + ], + "type": "text", + "content": ", i.e.," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 276, + 443, + 481, + 459 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 443, + 481, + 459 + ], + "spans": [ + { + "bbox": [ + 276, + 443, + 481, + 459 + ], + "type": "interline_equation", + "content": "A _ {k} ^ {l} \\Delta \\mathbf {w} _ {t, s} ^ {l} = 0, \\tag {8}", + "image_path": "f2e2685822581c149ab30d30f4624e5db52ff4fbfe3b1aadc429bed57b698e72.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 468, + 457, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 457, + 484 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 457, + 484 + ], + "type": "text", + "content": "we could preserve the historical information as " + }, + { + "bbox": [ + 132, + 468, + 457, + 484 + ], + "type": "inline_equation", + "content": "f^{l}\\left(A_{k}^{l};\\mathbf{w}_{t}^{l}\\right) = f^{l}\\left(A_{k}^{l};\\mathbf{w}_{k}^{l}\\right)" + }, + { + "bbox": [ + 132, + 468, + 457, + 484 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 493, + 481, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 493, + 481, + 518 + ], + "spans": [ + { + "bbox": [ + 131, + 493, + 481, + 518 + ], + "type": "text", + "content": "Building upon Proposition 2, we further extend it to our proposed TSC with Eq. (1), as detailed in Proposition 3." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 528, + 482, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 528, + 482, + 578 + ], + "spans": [ + { + "bbox": [ + 131, + 528, + 482, + 578 + ], + "type": "text", + "content": "Proposition 3. Let " + }, + { + "bbox": [ + 131, + 528, + 482, + 578 + ], + "type": "inline_equation", + "content": "f^{TSC}(\\cdot ;\\mathbf{w}_t)" + }, + { + "bbox": [ + 131, + 528, + 482, + 578 + ], + "type": "text", + "content": " denotes the TSC of the network trained on current task " + }, + { + "bbox": [ + 131, + 528, + 482, + 578 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_t" + }, + { + "bbox": [ + 131, + 528, + 482, + 578 + ], + "type": "text", + "content": " and its corresponding input feature " + }, + { + "bbox": [ + 131, + 528, + 482, + 578 + ], + "type": "inline_equation", + "content": "Z_{k}" + }, + { + "bbox": [ + 131, + 528, + 482, + 578 + ], + "type": "text", + "content": " from any previous task " + }, + { + "bbox": [ + 131, + 528, + 482, + 578 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_k" + }, + { + "bbox": [ + 131, + 528, + 482, + 578 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 131, + 528, + 482, + 578 + ], + "type": "inline_equation", + "content": "k < t" + }, + { + "bbox": [ + 131, + 528, + 482, + 578 + ], + "type": "text", + "content": "). During each training step " + }, + { + "bbox": [ + 131, + 528, + 482, + 578 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 131, + 528, + 482, + 578 + ], + "type": "text", + "content": ", if the TSC parameter update " + }, + { + "bbox": [ + 131, + 528, + 482, + 578 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{w}_{t,s}^{TSC}" + }, + { + "bbox": [ + 131, + 528, + 482, + 578 + ], + "type": "text", + "content": " lies in the null space of the subspace spanned by " + }, + { + "bbox": [ + 131, + 528, + 482, + 578 + ], + "type": "inline_equation", + "content": "Z_{k}" + }, + { + "bbox": [ + 131, + 528, + 482, + 578 + ], + "type": "text", + "content": ", i.e.," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 264, + 586, + 481, + 602 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 586, + 481, + 602 + ], + "spans": [ + { + "bbox": [ + 264, + 586, + 481, + 602 + ], + "type": "interline_equation", + "content": "Z _ {k} Z _ {k} ^ {\\top} \\Delta \\mathbf {w} _ {t, s} ^ {T S C} = 0, \\tag {9}", + "image_path": "300a539b4524f3737b60179fc8da444751c09878e6cadc25198e2fc03b5f80b3.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 611, + 384, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 384, + 627 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 384, + 627 + ], + "type": "text", + "content": "similarly we have " + }, + { + "bbox": [ + 132, + 611, + 384, + 627 + ], + "type": "inline_equation", + "content": "f^{TSC}\\left(Z_k;\\mathbf{w}_t^{TSC}\\right) = f^{TSC}\\left(Z_k;\\mathbf{w}_k^{TSC}\\right)" + }, + { + "bbox": [ + 132, + 611, + 384, + 627 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 634, + 364, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 634, + 364, + 647 + ], + "spans": [ + { + "bbox": [ + 132, + 634, + 364, + 647 + ], + "type": "text", + "content": "Proof. Please refer to the supplemental material." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "text", + "content": "H. Cheng et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 470, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 470, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 470, + 666 + ], + "type": "text", + "content": "5 The convolution operation can be regarded as a specific type of linear operation." + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "Inspired by previous work [27,39], we calculate the mean covariance matrix, denoted as " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "M_{t-1}" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": ", of the input feature for each specific layer in each mini-batch to enhance network stability and optimize memory efficiency. Here, let " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": " represent the collective term for the current input feature at each specific layer, with " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "A_t^l" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "-th linear layer and " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "Z_t Z_t^\\top" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": " for the TSC, respectively. To initialize, " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "M_0" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": " is set to 0 for the first session. At the end of each session " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "M_t" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": " is updated as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 249, + 194, + 481, + 228 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 194, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 249, + 194, + 481, + 228 + ], + "type": "interline_equation", + "content": "M _ {t} = M _ {t - 1} + \\sum_ {b = 1} ^ {\\mathcal {B}} \\frac {1}{J} Q _ {b} ^ {\\top} Q _ {b}, \\tag {10}", + "image_path": "bf32456aa827f5daa63cbc44ea2803c2aff4e8051ad54eaf4aab8b5dee00212e.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 231, + 480, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 231, + 480, + 254 + ], + "spans": [ + { + "bbox": [ + 130, + 231, + 480, + 254 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 231, + 480, + 254 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 130, + 231, + 480, + 254 + ], + "type": "text", + "content": " represents the number of mini-batches in the current task, and " + }, + { + "bbox": [ + 130, + 231, + 480, + 254 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 130, + 231, + 480, + 254 + ], + "type": "text", + "content": " denotes the number of videos in each mini-batch." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 255, + 480, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 255, + 480, + 292 + ], + "spans": [ + { + "bbox": [ + 130, + 255, + 480, + 292 + ], + "type": "text", + "content": "During the feature update in task " + }, + { + "bbox": [ + 130, + 255, + 480, + 292 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_t" + }, + { + "bbox": [ + 130, + 255, + 480, + 292 + ], + "type": "text", + "content": ", we perform Singular Value Decomposition (SVD) to find the approximate null space of the given feature matrix " + }, + { + "bbox": [ + 130, + 255, + 480, + 292 + ], + "type": "inline_equation", + "content": "M_{t-1}" + }, + { + "bbox": [ + 130, + 255, + 480, + 292 + ], + "type": "text", + "content": " corresponding to Proposition 2 and Proposition 3 as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 223, + 297, + 481, + 323 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 297, + 481, + 323 + ], + "spans": [ + { + "bbox": [ + 223, + 297, + 481, + 323 + ], + "type": "interline_equation", + "content": "\\left[ U _ {1}, U _ {2} \\right] \\left[ \\begin{array}{c c} \\Sigma_ {1} & 0 \\\\ 0 & \\Sigma_ {2} \\end{array} \\right] \\left[ \\begin{array}{l} U _ {1} ^ {\\top} \\\\ U _ {2} ^ {\\top} \\end{array} \\right] = \\operatorname {S V D} \\left(M _ {t - 1}\\right), \\tag {11}", + "image_path": "f4e10baff49bcaf24afcf1ebfd78e73523e5350233383e950ff0779297a62cf1.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 327, + 482, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 482, + 495 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 327, + 482, + 495 + ], + "type": "inline_equation", + "content": "U_{1}, U_{2}" + }, + { + "bbox": [ + 130, + 327, + 482, + 495 + ], + "type": "text", + "content": " represent unitary matrices that comprise orthonormal bases associated with the singular values in " + }, + { + "bbox": [ + 130, + 327, + 482, + 495 + ], + "type": "inline_equation", + "content": "\\Sigma_{1}, \\Sigma_{2}" + }, + { + "bbox": [ + 130, + 327, + 482, + 495 + ], + "type": "text", + "content": ", respectively, with all zero singular values relegated to " + }, + { + "bbox": [ + 130, + 327, + 482, + 495 + ], + "type": "inline_equation", + "content": "\\Sigma_{2}" + }, + { + "bbox": [ + 130, + 327, + 482, + 495 + ], + "type": "text", + "content": ". However, it is unrealistic to guarantee that there always exists zero singular values. Previous work [27, 39] constructs " + }, + { + "bbox": [ + 130, + 327, + 482, + 495 + ], + "type": "inline_equation", + "content": "U_{2}^{\\top}" + }, + { + "bbox": [ + 130, + 327, + 482, + 495 + ], + "type": "text", + "content": " by selecting singular values restricted to a specific proportion. However, this strategy might not always provide the most compact representation, particularly when there is a significant difference in the scale of singular values. Hence, we employ a threshold strategy based on the energy retained in the matrix " + }, + { + "bbox": [ + 130, + 327, + 482, + 495 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 327, + 482, + 495 + ], + "type": "text", + "content": " for the construction " + }, + { + "bbox": [ + 130, + 327, + 482, + 495 + ], + "type": "inline_equation", + "content": "U_{2}^{\\top}" + }, + { + "bbox": [ + 130, + 327, + 482, + 495 + ], + "type": "text", + "content": ", corresponding to a segment of smallest singular values that satisfy " + }, + { + "bbox": [ + 130, + 327, + 482, + 495 + ], + "type": "inline_equation", + "content": "\\| \\Sigma_{2} \\|_{F}^{2} \\leq (1 - \\eta) \\left( \\| \\Sigma_{1} \\|_{F}^{2} + \\| \\Sigma_{2} \\|_{F}^{2} \\right)" + }, + { + "bbox": [ + 130, + 327, + 482, + 495 + ], + "type": "text", + "content": ", with the proportion of retained energy denoted by " + }, + { + "bbox": [ + 130, + 327, + 482, + 495 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 130, + 327, + 482, + 495 + ], + "type": "text", + "content": ". When " + }, + { + "bbox": [ + 130, + 327, + 482, + 495 + ], + "type": "inline_equation", + "content": "\\eta \\approx 1" + }, + { + "bbox": [ + 130, + 327, + 482, + 495 + ], + "type": "text", + "content": ", it is reasonable to approximate the null space of the matrix " + }, + { + "bbox": [ + 130, + 327, + 482, + 495 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 327, + 482, + 495 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 327, + 482, + 495 + ], + "type": "inline_equation", + "content": "U_{2}^{\\top}" + }, + { + "bbox": [ + 130, + 327, + 482, + 495 + ], + "type": "text", + "content": ". This enables the parameter update " + }, + { + "bbox": [ + 130, + 327, + 482, + 495 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{w}" + }, + { + "bbox": [ + 130, + 327, + 482, + 495 + ], + "type": "text", + "content": " from the gradient " + }, + { + "bbox": [ + 130, + 327, + 482, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{g}" + }, + { + "bbox": [ + 130, + 327, + 482, + 495 + ], + "type": "text", + "content": " concerning the parameters " + }, + { + "bbox": [ + 130, + 327, + 482, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 130, + 327, + 482, + 495 + ], + "type": "text", + "content": " through subspace projection based on " + }, + { + "bbox": [ + 130, + 327, + 482, + 495 + ], + "type": "inline_equation", + "content": "U_{2}" + }, + { + "bbox": [ + 130, + 327, + 482, + 495 + ], + "type": "text", + "content": ". That is, during the training step " + }, + { + "bbox": [ + 130, + 327, + 482, + 495 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 327, + 482, + 495 + ], + "type": "text", + "content": " within any incremental session " + }, + { + "bbox": [ + 130, + 327, + 482, + 495 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 327, + 482, + 495 + ], + "type": "text", + "content": ", we have:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 266, + 500, + 481, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 500, + 481, + 514 + ], + "spans": [ + { + "bbox": [ + 266, + 500, + 481, + 514 + ], + "type": "interline_equation", + "content": "\\Delta \\mathbf {w} _ {t, s} = U _ {2} U _ {2} ^ {\\top} \\mathbf {g} _ {t, s}. \\tag {12}", + "image_path": "db4e8dad93bb14c85b8eeeb69b332d620d3a436c72b3e9a8d5adcdf251ddde43.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 529, + 335, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 529, + 335, + 541 + ], + "spans": [ + { + "bbox": [ + 131, + 529, + 335, + 541 + ], + "type": "text", + "content": "3.4 Model Optimization and Inference" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 545, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 629 + ], + "type": "text", + "content": "Our STSP method is summarized in Algorithm 1. In the training phase, the network parameters are updated conventionally during the initial session. After this initial session, considering parameter updates, the mean covariance of input features from the base classes is retained for each layer. These retained features are then utilized to update parameters according to Eq. (12) at each step in every subsequent incremental session. Additionally, after each incremental session, these parameters are further updated with features extracted from novel classes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "content": "For the testing phase, no additional memories or operations are necessary. We simply feed the video into the network and obtain classification results using the trained model." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 250, + 90, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 90, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 250, + 90, + 447, + 103 + ], + "type": "text", + "content": "Spatial-Temporal Subspace Projection for VCIL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 132, + 130, + 481, + 384 + ], + "blocks": [ + { + "bbox": [ + 133, + 115, + 356, + 129 + ], + "lines": [ + { + "bbox": [ + 133, + 115, + 356, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 356, + 129 + ], + "type": "text", + "content": "Algorithm 1 STSP for class incremental learning" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 132, + 130, + 481, + 384 + ], + "lines": [ + { + "bbox": [ + 132, + 130, + 481, + 384 + ], + "spans": [ + { + "bbox": [ + 132, + 130, + 481, + 384 + ], + "type": "text", + "content": "Input: Task series " + }, + { + "bbox": [ + 132, + 130, + 481, + 384 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{T}_1,\\mathcal{T}_2,\\dots \\}" + }, + { + "bbox": [ + 132, + 130, + 481, + 384 + ], + "type": "text", + "content": " , network " + }, + { + "bbox": [ + 132, + 130, + 481, + 384 + ], + "type": "inline_equation", + "content": "f(\\cdot ;\\mathbf{w})" + }, + { + "bbox": [ + 132, + 130, + 481, + 384 + ], + "type": "text", + "content": " with parameters w for linear layers and TSC, learning rate " + }, + { + "bbox": [ + 132, + 130, + 481, + 384 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 132, + 130, + 481, + 384 + ], + "type": "text", + "content": " and energy threshold " + }, + { + "bbox": [ + 132, + 130, + 481, + 384 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 132, + 130, + 481, + 384 + ], + "type": "text", + "content": " \nInitialize: " + }, + { + "bbox": [ + 132, + 130, + 481, + 384 + ], + "type": "inline_equation", + "content": "M_0 = 0" + }, + { + "bbox": [ + 132, + 130, + 481, + 384 + ], + "type": "text", + "content": " for each linear layer and TSC. \n1: for " + }, + { + "bbox": [ + 132, + 130, + 481, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_t\\in \\{\\mathcal{T}_1,\\mathcal{T}_2,\\dots \\}" + }, + { + "bbox": [ + 132, + 130, + 481, + 384 + ], + "type": "text", + "content": " do \n2: " + }, + { + "bbox": [ + 132, + 130, + 481, + 384 + ], + "type": "inline_equation", + "content": "s = 0" + }, + { + "bbox": [ + 132, + 130, + 481, + 384 + ], + "type": "text", + "content": " \n3: while not converge do \n4: Sample a mini-batch " + }, + { + "bbox": [ + 132, + 130, + 481, + 384 + ], + "type": "inline_equation", + "content": "\\{D^t,y^t\\} \\in \\mathcal{T}_t" + }, + { + "bbox": [ + 132, + 130, + 481, + 384 + ], + "type": "text", + "content": " with videos " + }, + { + "bbox": [ + 132, + 130, + 481, + 384 + ], + "type": "inline_equation", + "content": "\\{D_i^t,y_i^t\\}_{j = 1}^J" + }, + { + "bbox": [ + 132, + 130, + 481, + 384 + ], + "type": "text", + "content": " \n5: Compute " + }, + { + "bbox": [ + 132, + 130, + 481, + 384 + ], + "type": "inline_equation", + "content": "f(\\{D^{t},y^{t}\\} ;\\mathbf{w}_{t,s})" + }, + { + "bbox": [ + 132, + 130, + 481, + 384 + ], + "type": "text", + "content": " and obtain the gradient " + }, + { + "bbox": [ + 132, + 130, + 481, + 384 + ], + "type": "inline_equation", + "content": "\\mathbf{g}_{t,s}" + }, + { + "bbox": [ + 132, + 130, + 481, + 384 + ], + "type": "text", + "content": " \n6: if " + }, + { + "bbox": [ + 132, + 130, + 481, + 384 + ], + "type": "inline_equation", + "content": "t > 1" + }, + { + "bbox": [ + 132, + 130, + 481, + 384 + ], + "type": "text", + "content": " then \n7: for " + }, + { + "bbox": [ + 132, + 130, + 481, + 384 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{t,s}" + }, + { + "bbox": [ + 132, + 130, + 481, + 384 + ], + "type": "text", + "content": " in linear layers and TSC do \n8: Compute " + }, + { + "bbox": [ + 132, + 130, + 481, + 384 + ], + "type": "inline_equation", + "content": "U_{2}" + }, + { + "bbox": [ + 132, + 130, + 481, + 384 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 132, + 130, + 481, + 384 + ], + "type": "inline_equation", + "content": "M_{t - 1}" + }, + { + "bbox": [ + 132, + 130, + 481, + 384 + ], + "type": "text", + "content": " via Eq. (11) with a threshold " + }, + { + "bbox": [ + 132, + 130, + 481, + 384 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 132, + 130, + 481, + 384 + ], + "type": "text", + "content": " \n9: Update " + }, + { + "bbox": [ + 132, + 130, + 481, + 384 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{w}_{t,s}" + }, + { + "bbox": [ + 132, + 130, + 481, + 384 + ], + "type": "text", + "content": " using Eq. (12) \n10: Update " + }, + { + "bbox": [ + 132, + 130, + 481, + 384 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{t,s + 1} = \\mathbf{w}_{t,s} - \\alpha \\Delta \\mathbf{w}_{t,s}" + }, + { + "bbox": [ + 132, + 130, + 481, + 384 + ], + "type": "text", + "content": " \n11: end for \n12: else \n13: Update " + }, + { + "bbox": [ + 132, + 130, + 481, + 384 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{w}_{t,s} = \\mathbf{g}_{t,s}" + }, + { + "bbox": [ + 132, + 130, + 481, + 384 + ], + "type": "text", + "content": " \n14: Update " + }, + { + "bbox": [ + 132, + 130, + 481, + 384 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{t,s + 1} = \\mathbf{w}_{t,s} - \\alpha \\Delta \\mathbf{w}_{t,s}" + }, + { + "bbox": [ + 132, + 130, + 481, + 384 + ], + "type": "text", + "content": " \n15: end if \n16: " + }, + { + "bbox": [ + 132, + 130, + 481, + 384 + ], + "type": "inline_equation", + "content": "s = s + 1" + }, + { + "bbox": [ + 132, + 130, + 481, + 384 + ], + "type": "text", + "content": " \n17: end while \n18: Get current input features " + }, + { + "bbox": [ + 132, + 130, + 481, + 384 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 132, + 130, + 481, + 384 + ], + "type": "text", + "content": " at each linear layer and TSC \n19: Update " + }, + { + "bbox": [ + 132, + 130, + 481, + 384 + ], + "type": "inline_equation", + "content": "M_t" + }, + { + "bbox": [ + 132, + 130, + 481, + 384 + ], + "type": "text", + "content": " using Eq. (10) for later incremental session \n20: end for" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "algorithm" + }, + { + "bbox": [ + 132, + 405, + 230, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 405, + 230, + 419 + ], + "spans": [ + { + "bbox": [ + 132, + 405, + 230, + 419 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 429, + 480, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 429, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 429, + 480, + 464 + ], + "type": "text", + "content": "In this section, we compare our results against state-of-the-art techniques in class-incremental action recognition benchmarks. Additionally, we perform ablation studies to demonstrate the efficacy of our proposed method." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 480, + 261, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 480, + 261, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 480, + 261, + 493 + ], + "type": "text", + "content": "4.1 Experimental Setup" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 498, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 641 + ], + "type": "text", + "content": "Datasets. The evaluation of our proposed framework is conducted using the UCF101 [30], HMDB51 [15], and Something-Something V2 [11] datasets, which are recognized benchmarks in video action recognition. The UCF-101 dataset comprises 13,320 instances across 101 action categories, with the data divided into three splits for training (70%) and testing (30%). Similarly, the HMDB-51 dataset features 6,766 video clips distributed among 51 action categories, also with three splits for training and testing in the same proportions. For our evaluations, we selected the first split from both datasets. Distinct from UCF101 and HMDB51, the Something-Something V2 dataset, a crowd-sourced collection, includes 220,000 videos across 174 classes. This dataset uniquely maintains consistent objects and backgrounds across its action categories, thereby emphasizing the importance of a model's capability to discern subtle motion cues." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 481, + 666 + ], + "type": "text", + "content": "Evaluation Protocol. Our methodology adheres to the class-incremental learning benchmarks frequently utilized in video analysis, as outlined in TCD [22]." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "H. Cheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 150, + 146, + 460, + 251 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 1: Video class incremental learning results on UCF101 and HMDB51. The best result for each setting is emphasized in bold." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 150, + 146, + 460, + 251 + ], + "lines": [ + { + "bbox": [ + 150, + 146, + 460, + 251 + ], + "spans": [ + { + "bbox": [ + 150, + 146, + 460, + 251 + ], + "type": "table", + "html": "
MethodUCF101HMDB51
10 × 5 stages5 × 10 stages2 × 25 stages5 × 5 stages1 × 25 stages
LwFMC [16]42.1425.5911.6826.8216.49
LwM [6]43.3926.0712.0826.9716.50
UCIR [12]74.3170.4263.2244.9037.04
PODNet [8]73.2671.5870.2844.3238.76
TCD [22]74.8973.4372.1945.3440.47
FrameMaker [23]78.1376.3875.7747.5442.65
STSP (Ours)81.1582.8479.2556.9949.19
", + "image_path": "156348bf1b1ef7e062298c2d10a498541158b830d8fe77f3a263ef5b44ff12c1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 277, + 482, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 482, + 386 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 482, + 386 + ], + "type": "text", + "content": "This approach begins with a model pre-trained on half of the total number of classes, with the remaining classes introduced sequentially at each incremental step. Specifically, for UCF101, the initial training involves 51 classes, after which the subsequent 50 classes are structured into incremental learning tasks of 5, 10, and 25 classes. In the case of HMDB51, the model begins with training on 26 classes, followed by the distribution of the remaining classes into groups of 5 and 25 for subsequent learning phases. For Something-Something V2, the initial training phase covers 84 classes, with the remaining classes organized into incremental learning groups of 10 and 5 classes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 388, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 482, + 449 + ], + "type": "text", + "content": "Implementation Details. Our framework leverages the official implementation of TSM [17], utilizing the PyTorch Library for development. We train a ResNet-34 TSM model on UCF101 for 50 epochs with a batch size of 32. For both HMDB51 and Something-Something V2, ResNet-50 TSM models are employed, undergoing training for 50 epochs with a batch size of 32." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 478, + 228, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 478, + 228, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 478, + 228, + 489 + ], + "type": "text", + "content": "4.2 Main Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "For a fair comparison, we employ the same model structure with the same pretraining initialization as the existing methods [22, 23]. Unlike prior approaches that incorporate exemplars for incremental learning and assess performance using both the standard classification protocol and Nearest Mean of Exemplars (NME) metrics [26], our method does not retain any exemplars. Consequently, we only report the standard classification results. We compare our proposed STSP method with the existing video-class incremental learning methods over three benchmarks, as shown in Tab. 1 and Tab. 2. From both tables, we observe that the proposed STSP method significantly outperforms all competing methods across all settings on each dataset. Notably, unlike other approaches that rely on retaining exemplars for incremental learning sessions, our method utilizes gradient projection to constrain model updates, effectively safeguarding against performance drops on previously learned classes." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 250, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 250, + 90, + 447, + 102 + ], + "type": "text", + "content": "Spatial-Temporal Subspace Projection for VCIL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 167, + 286, + 235 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 288, + 159 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 288, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 288, + 159 + ], + "type": "text", + "content": "Table 2: Video class incremental learning results on Something-Something V2. The best result for each setting is emphasized in bold." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 167, + 286, + 235 + ], + "lines": [ + { + "bbox": [ + 135, + 167, + 286, + 235 + ], + "spans": [ + { + "bbox": [ + 135, + 167, + 286, + 235 + ], + "type": "table", + "html": "
Method10 × 9 stages5 × 18 stages
UCIR [12]26.8420.69
PODNet [8]34.9726.95
TCD [22]35.7829.60
FrameMaker [23]37.2530.98
STSP (Ours)69.6870.87
", + "image_path": "3656398c9388e07a216c98556f1bd19af1218fda393218e25564fe7c1bc7e7a1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 307, + 157, + 480, + 237 + ], + "blocks": [ + { + "bbox": [ + 304, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 304, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 304, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 3: Ablation studies for our proposed STSP method on UCF101 with " + }, + { + "bbox": [ + 304, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "5 \\times 10" + }, + { + "bbox": [ + 304, + 114, + 482, + 148 + ], + "type": "text", + "content": " stages and " + }, + { + "bbox": [ + 304, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "2 \\times 25" + }, + { + "bbox": [ + 304, + 114, + 482, + 148 + ], + "type": "text", + "content": " stages settings, respectively." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 157, + 480, + 237 + ], + "lines": [ + { + "bbox": [ + 307, + 157, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 307, + 157, + 480, + 237 + ], + "type": "table", + "html": "
TSCSGPExemplarsUCF101
5 × 10 stages2 × 25 stages
(i)××5/class68.2857.39
(ii)✓×5/class71.5061.78
(iii)××017.6319.22
(iv)✓×08.7814.99
(v)×✓077.6674.02
(vi)✓✓082.8479.25
", + "image_path": "a0aeb7a36745e6fcbd155b0a8e458c7ea06ac247b4203bbdb23144bf96e9987b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 135, + 248, + 305, + 357 + ], + "blocks": [ + { + "bbox": [ + 135, + 248, + 305, + 357 + ], + "lines": [ + { + "bbox": [ + 135, + 248, + 305, + 357 + ], + "spans": [ + { + "bbox": [ + 135, + 248, + 305, + 357 + ], + "type": "image", + "image_path": "f448247ea081d492cf4b2f6a773e7c6a8a7d713f0c9b73ed1d85be9c26a1a0f4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 169, + 358, + 268, + 368 + ], + "lines": [ + { + "bbox": [ + 169, + 358, + 268, + 368 + ], + "spans": [ + { + "bbox": [ + 169, + 358, + 268, + 368 + ], + "type": "text", + "content": "(a) UCF101 (5 × 10 stages)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 312, + 248, + 482, + 357 + ], + "blocks": [ + { + "bbox": [ + 312, + 248, + 482, + 357 + ], + "lines": [ + { + "bbox": [ + 312, + 248, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 312, + 248, + 482, + 357 + ], + "type": "image", + "image_path": "07d7d1b95cf4a0e3ba7fbd1f8033c9086d61a878a5e6599d581b72cdecd51932.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 358, + 444, + 368 + ], + "lines": [ + { + "bbox": [ + 345, + 358, + 444, + 368 + ], + "spans": [ + { + "bbox": [ + 345, + 358, + 444, + 368 + ], + "type": "text", + "content": "(b) UCF101 (2 × 25 stages)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 376, + 480, + 388 + ], + "lines": [ + { + "bbox": [ + 132, + 376, + 480, + 388 + ], + "spans": [ + { + "bbox": [ + 132, + 376, + 480, + 388 + ], + "type": "text", + "content": "Fig. 3: Plots of accuracy over seen classes along with the incremental steps on UCF101." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 411, + 307, + 425 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 411, + 307, + 425 + ], + "spans": [ + { + "bbox": [ + 132, + 411, + 307, + 425 + ], + "type": "text", + "content": "4.3 Ablation Study and Analysis" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 432, + 482, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 432, + 482, + 469 + ], + "spans": [ + { + "bbox": [ + 130, + 432, + 482, + 469 + ], + "type": "text", + "content": "In this section, we perform ablation studies to investigate the effectiveness of the key components of the proposed STSP on the UCF101 dataset under both " + }, + { + "bbox": [ + 130, + 432, + 482, + 469 + ], + "type": "inline_equation", + "content": "2 \\times 25" + }, + { + "bbox": [ + 130, + 432, + 482, + 469 + ], + "type": "text", + "content": " stages and " + }, + { + "bbox": [ + 130, + 432, + 482, + 469 + ], + "type": "inline_equation", + "content": "5 \\times 10" + }, + { + "bbox": [ + 130, + 432, + 482, + 469 + ], + "type": "text", + "content": " stages settings." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 485, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 482, + 629 + ], + "type": "text", + "content": "Effectiveness of gradient projection To evaluate the SGP's capability to mitigate catastrophic forgetting, we designed experiments with various configurations: utilizing the method with SGP, the method without SGP, and the method without SGP but incorporating exemplars. In the last configuration, we preserved 5 videos, each containing 8 frames, as exemplars for prior classes to be used in future training sessions. Furthermore, we extended these comparisons to assess performance by employing a linear classifier and the proposed TSC. The comparison results are shown in Tab. 3, where (i), (iii), and (v) utilize the linear classifier, while (ii), (iv), and (vi) implement the TSC. In cooperation with the comparative analysis of the network's performance as the number of seen classes increases, as shown in Fig. 3, it can be found that the network tends to forget most knowledge of previous classes when no strategy is adopted." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "Moreover, when comparing our SGP strategy against the exemplar-based distillation method, particularly through the comparison of (i) " + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "inline_equation", + "content": "v.s." + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": " (v) and the comparison of (ii) " + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "inline_equation", + "content": "v.s." + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": " (vi) in Tab. 3, we can find that our proposed SGP strategy" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "text", + "content": "H. Cheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 202, + 146, + 409, + 175 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 4: Ablation study of the subspace basis size " + }, + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "type": "text", + "content": " for TSC on UCF101 with " + }, + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "type": "inline_equation", + "content": "5\\times 10" + }, + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "type": "text", + "content": " stages setting." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 202, + 146, + 409, + 175 + ], + "lines": [ + { + "bbox": [ + 202, + 146, + 409, + 175 + ], + "spans": [ + { + "bbox": [ + 202, + 146, + 409, + 175 + ], + "type": "table", + "html": "
τ=1τ=2τ=4τ=8τ=16
STSP (Ours)66.7567.8875.3982.8463.02
", + "image_path": "2d74f708b12972e5e4466687d76f8d33705e67f3d8b1d6ea10c1977c7fd714d7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 198, + 482, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 198, + 482, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 198, + 482, + 282 + ], + "type": "text", + "content": "significantly outperforms the exemplar-based distillation method. Specifically, regarding the performance of base classes during incremental learning sessions, as shown in Fig. 4, it is evident that our proposed SGP has minimal impact on the classification performance for the base classes. These findings indicate that our proposed SGP can uphold accurate classification across all classes without compromising performance on the base classes, emphasizing the effectiveness of SGP in mitigating knowledge degradation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 283, + 482, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 283, + 482, + 368 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 482, + 368 + ], + "type": "text", + "content": "Furthermore, based on the comparison in Fig. 3 and Fig. 4, we observe that all methods perform worse under the setting of " + }, + { + "bbox": [ + 130, + 283, + 482, + 368 + ], + "type": "inline_equation", + "content": "2 \\times 25" + }, + { + "bbox": [ + 130, + 283, + 482, + 368 + ], + "type": "text", + "content": " stages, while our method with SGP still achieves the best performance. A plausible explanation is that the model may be susceptible to overfitting when dealing with a limited set of novel classes. Moreover, for our method, the computation of the null space becomes more unstable in these scenarios, especially when updating the mean covariance matrix is solely based on a few classes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 389, + 482, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 389, + 482, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 482, + 556 + ], + "type": "text", + "content": "Linear or Subspace-based Classifier? To assess the effectiveness of the proposed TSC, we compare it with a linear classifier based on an FC layer. Moreover, this ablation study extends to both the exemplar-based distillation method and our proposed SGP, with the comparison results shown in Tab. 3. For the exemplar-based distillation approach, the implementation of TSC leads to a performance increase of " + }, + { + "bbox": [ + 130, + 389, + 482, + 556 + ], + "type": "inline_equation", + "content": "3.22\\%" + }, + { + "bbox": [ + 130, + 389, + 482, + 556 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 389, + 482, + 556 + ], + "type": "inline_equation", + "content": "4.39\\%" + }, + { + "bbox": [ + 130, + 389, + 482, + 556 + ], + "type": "text", + "content": " on UCF101 at the 10-stage and 25-stage settings, respectively, as evidenced by the comparison between (i) and (ii). In the context of SGP, the comparison between (v) and (vi) reveals that the proposed TSC similarly improves the overall performance by " + }, + { + "bbox": [ + 130, + 389, + 482, + 556 + ], + "type": "inline_equation", + "content": "5.18\\%" + }, + { + "bbox": [ + 130, + 389, + 482, + 556 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 389, + 482, + 556 + ], + "type": "inline_equation", + "content": "5.23\\%" + }, + { + "bbox": [ + 130, + 389, + 482, + 556 + ], + "type": "text", + "content": " for the 10-stage and 25-stage settings on UCF101, respectively. Compared with the exemplar-based distillation method, our method employs SGP to constrain the classifier update through gradient subspace projection. The above findings align with the concept of TSC, collectively constraining the subspace-based classifier and further enhancing classification performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "From the comparison in Fig. 4, another observation is that TSC can even boost the classification performance on the base classes, demonstrating its effectiveness in modeling class representations. However, we also observe that TSC may perform poorly when there are no constraints (i.e. without exemplars nor SGP) on the network. One plausible explanation for this phenomenon is that, during incremental sessions, the network tends to shift the feature space toward new tasks without constraints from historical information. Such adjustments can result in the original subspace deviating from its previous configuration, leading to significant misalignment that greatly impacts the performance of TSC." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 250, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 250, + 90, + 447, + 102 + ], + "type": "text", + "content": "Spatial-Temporal Subspace Projection for VCIL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 299, + 223 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 299, + 223 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 299, + 223 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 299, + 223 + ], + "type": "image", + "image_path": "e12369de4149b125837e8d1cf7e00e6f1db461bf46d3b054d2e714a9fe1b1b61.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 169, + 224, + 268, + 234 + ], + "lines": [ + { + "bbox": [ + 169, + 224, + 268, + 234 + ], + "spans": [ + { + "bbox": [ + 169, + 224, + 268, + 234 + ], + "type": "text", + "content": "(a) UCF101 " + }, + { + "bbox": [ + 169, + 224, + 268, + 234 + ], + "type": "inline_equation", + "content": "(5\\times 10" + }, + { + "bbox": [ + 169, + 224, + 268, + 234 + ], + "type": "text", + "content": " stages)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 312, + 114, + 476, + 223 + ], + "blocks": [ + { + "bbox": [ + 312, + 114, + 476, + 223 + ], + "lines": [ + { + "bbox": [ + 312, + 114, + 476, + 223 + ], + "spans": [ + { + "bbox": [ + 312, + 114, + 476, + 223 + ], + "type": "image", + "image_path": "0171a0332d9e8d19f03421da3bd4fc01d980d684a015deee4823b40c531b01f5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 224, + 444, + 234 + ], + "lines": [ + { + "bbox": [ + 345, + 224, + 444, + 234 + ], + "spans": [ + { + "bbox": [ + 345, + 224, + 444, + 234 + ], + "type": "text", + "content": "(b) UCF101 (2 × 25 stages)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 243, + 480, + 255 + ], + "lines": [ + { + "bbox": [ + 132, + 243, + 480, + 255 + ], + "spans": [ + { + "bbox": [ + 132, + 243, + 480, + 255 + ], + "type": "text", + "content": "Fig. 4: Plots of accuracy over base classes along with the incremental steps on UCF101." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 277, + 482, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 482, + 410 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 482, + 410 + ], + "type": "text", + "content": "The Size of the Subspace Basis We explore how the size of the subspace basis influences the classification accuracy of our proposed method, the comparison results are shown in Tab. 4. It can be found that large subspace basis sizes (i.e., " + }, + { + "bbox": [ + 130, + 277, + 482, + 410 + ], + "type": "inline_equation", + "content": "\\tau = 4" + }, + { + "bbox": [ + 130, + 277, + 482, + 410 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 130, + 277, + 482, + 410 + ], + "type": "inline_equation", + "content": "\\tau = 8" + }, + { + "bbox": [ + 130, + 277, + 482, + 410 + ], + "type": "text", + "content": ") are more effective at constructing the subspace for each category, enabling the capture of more distinctive features, which in turn leads to better performance as opposed to smaller bases. This improvement, however, results in an increase in both the network's parameters and computational complexity. Additionally, we can observe that a subspace basis size of 16 yields the least favorable results. This could be attributed to the difficulty in constructing orthogonal subspaces for new classes when the subspace basis is excessively large, especially as " + }, + { + "bbox": [ + 130, + 277, + 482, + 410 + ], + "type": "inline_equation", + "content": "M \\times \\tau" + }, + { + "bbox": [ + 130, + 277, + 482, + 410 + ], + "type": "text", + "content": " approaches the total number of classes " + }, + { + "bbox": [ + 130, + 277, + 482, + 410 + ], + "type": "inline_equation", + "content": "(\\mathcal{C})" + }, + { + "bbox": [ + 130, + 277, + 482, + 410 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 426, + 220, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 426, + 220, + 438 + ], + "spans": [ + { + "bbox": [ + 132, + 426, + 220, + 438 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 449, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 582 + ], + "type": "text", + "content": "In this paper, we proposed STSP, a novel spatial-temporal subspace projection method designed for video class-incremental learning tasks. The objective is to learn discriminative and robust representations for both video and class, facilitating further incremental learning. STSP mainly consists of two key components, i.e., temporal-base subspace classifier (TSC) and spatial-based gradient projection (SGP). TSC employs a subspace to model each video and its class, incorporating intra-frame temporal information. On the other hand, SGP is designed to mitigate catastrophic forgetting by projecting the gradient of network parameters in each layer onto the subspace spanned by the corresponding input spatial features. Extensive experiments on three benchmarks demonstrate the effectiveness of our proposed STSP method." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Limitations. In contrast to previous video class incremental learning methods, our proposed STSP method incorporates gradient projection to prevent the forgetting of historical information from previous classes. This strategy removes the necessity of storing exemplars and prototypes of previously learned classes, yet it requires the storage and updating of input features for both linear layers and the introduced TSC. Such a requirement could pose challenges when employing large backbones, requiring further exploration and discussion." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "text", + "content": "H. Cheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 163 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 163 + ], + "type": "text", + "content": "This research was carried out at the Rapid-Rich Object Search (ROSE) Lab at Nanyang Technological University in Singapore." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 182, + 197, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 182, + 197, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 182, + 197, + 194 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 205, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 138, + 205, + 480, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 205, + 480, + 239 + ], + "spans": [ + { + "bbox": [ + 138, + 205, + 480, + 239 + ], + "type": "text", + "content": "1. Akyurek, A.F., Akyurek, E., Wijaya, D., Andreas, J.: Subspace regularizers for few-shot class incremental learning. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=boJy41J-tnQ" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 239, + 480, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 239, + 480, + 271 + ], + "spans": [ + { + "bbox": [ + 138, + 239, + 480, + 271 + ], + "type": "text", + "content": "2. Belhumeur, P.N., Hespanha, J.P., Kriegman, D.J.: Eigenfaces vs. fisherfaces: Recognition using class specific linear projection. IEEE Transactions on pattern analysis and machine intelligence 19(7), 711-720 (1997)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 272, + 480, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 272, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 138, + 272, + 480, + 304 + ], + "type": "text", + "content": "3. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6299-6308 (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 304, + 480, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 304, + 480, + 337 + ], + "spans": [ + { + "bbox": [ + 138, + 304, + 480, + 337 + ], + "type": "text", + "content": "4. Chen, H., Wang, Y., Hu, Q.: Multi-granularity regularized re-balancing for class incremental learning. IEEE Transactions on Knowledge and Data Engineering 35(7), 7263-7277 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 337, + 480, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 337, + 480, + 381 + ], + "spans": [ + { + "bbox": [ + 138, + 337, + 480, + 381 + ], + "type": "text", + "content": "5. Cheraghian, A., Rahman, S., Ramasinghe, S., Fang, P., Simon, C., Petersson, L., Harandi, M.: Synthesized feature based few-shot class-incremental learning on a mixture of subspaces. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 8661-8670 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 381, + 480, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 381, + 480, + 414 + ], + "spans": [ + { + "bbox": [ + 138, + 381, + 480, + 414 + ], + "type": "text", + "content": "6. Dhar, P., Singh, R.V., Peng, K.C., Wu, Z., Chellappa, R.: Learning without memorizing. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5138-5146 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 414, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 414, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 414, + 480, + 468 + ], + "type": "text", + "content": "7. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=YicbFdNTTy" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 468, + 480, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 468, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 138, + 468, + 480, + 512 + ], + "type": "text", + "content": "8. Douillard, A., Cord, M., Ollion, C., Robert, T., Valle, E.: Podnet: Pooled outputs distillation for small-tasks incremental learning. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XX 16. pp. 86-102. Springer (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 512, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 512, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 512, + 480, + 544 + ], + "type": "text", + "content": "9. Feichtenhofer, C., Fan, H., Malik, J., He, K.: Slowfast networks for video recognition. In: Proceedings of the IEEE international conference on computer vision. pp. 6202-6211 (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 544, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 544, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 544, + 480, + 578 + ], + "type": "text", + "content": "0. Gao, Q., Zhao, C., Ghanem, B., Zhang, J.: R-dfcil: Relation-guided representation learning for data-free class incremental learning. In: European Conference on Computer Vision. pp. 423-439. Springer (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 578, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 480, + 632 + ], + "type": "text", + "content": "1. Goyal, R., Ebrahimi Kahou, S., Michalski, V., Materzynska, J., Westphal, S., Kim, H., Haenel, V., Fruend, I., Yianilos, P., Mueller-Freitag, M., et al.: The\" something something\" video database for learning and evaluating visual common sense. In: Proceedings of the IEEE international conference on computer vision. pp. 5842-5850 (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "text", + "content": "2. Hou, S., Pan, X., Loy, C.C., Wang, Z., Lin, D.: Learning a unified classifier incrementally via rebalancing. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 831-839 (2019)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 250, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 250, + 91, + 447, + 102 + ], + "type": "text", + "content": "Spatial-Temporal Subspace Projection for VCIL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "13. Huang, H., Wang, Y., Hu, Q., Cheng, M.M.: Class-specific semantic reconstruction for open set recognition. IEEE transactions on pattern analysis and machine intelligence 45(4), 4214-4228 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 184 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 184 + ], + "type": "text", + "content": "14. Ji, S., Xu, W., Yang, M., Yu, K.: 3d convolutional neural networks for human action recognition. IEEE transactions on pattern analysis and machine intelligence 35(1), 221-231 (2012)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 185, + 481, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 185, + 481, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 185, + 481, + 217 + ], + "type": "text", + "content": "15. Kuehne, H., Jhuang, H., Garrote, E., Poggio, T., Serre, T.: Hmdb: a large video database for human motion recognition. In: 2011 International conference on computer vision. pp. 2556-2563. IEEE (2011)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 218, + 481, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 218, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 218, + 481, + 239 + ], + "type": "text", + "content": "16. Li, Z., Hoiem, D.: Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence 40(12), 2935-2947 (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 240, + 481, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 240, + 481, + 274 + ], + "spans": [ + { + "bbox": [ + 132, + 240, + 481, + 274 + ], + "type": "text", + "content": "17. Lin, J., Gan, C., Han, S.: Temporal shift module for efficient video understanding. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7083-7093 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 274, + 481, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 481, + 308 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 481, + 308 + ], + "type": "text", + "content": "18. Liu, C.: Gabor-based kernel pca with fractional power polynomial models for face recognition. IEEE transactions on pattern analysis and machine intelligence 26(5), 572-581 (2004)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 308, + 481, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 481, + 352 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 481, + 352 + ], + "type": "text", + "content": "19. Masana, M., Liu, X., Twardowski, B., Menta, M., Bagdanov, A.D., Van De Weijer, J.: Class-incremental learning: survey and performance evaluation on image classification. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(5), 5513-5533 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 352, + 481, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 352, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 352, + 481, + 385 + ], + "type": "text", + "content": "20. McCloskey, M., Cohen, N.J.: Catastrophic interference in connectionist networks: The sequential learning problem. In: Psychology of learning and motivation, vol. 24, pp. 109-165. Elsevier (1989)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 386, + 481, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 481, + 419 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 481, + 419 + ], + "type": "text", + "content": "21. Papyan, V., Han, X., Donoho, D.L.: Prevalence of neural collapse during the terminal phase of deep learning training. Proceedings of the National Academy of Sciences 117(40), 24652-24663 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 419, + 481, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 419, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 419, + 481, + 453 + ], + "type": "text", + "content": "22. Park, J., Kang, M., Han, B.: Class-incremental learning for action recognition in videos. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 13698-13707 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 453, + 481, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 481, + 497 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 481, + 497 + ], + "type": "text", + "content": "23. Pei, Y., Qing, Z., Cen, J., Wang, X., Zhang, S., Wang, Y., Tang, M., Sang, N., Qian, X.: Learning a condensed frame for memory-efficient video class-incremental learning. Advances in Neural Information Processing Systems 35, 31002-31016 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 498, + 481, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 481, + 531 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 481, + 531 + ], + "type": "text", + "content": "24. Pei, Y., Qing, Z., Zhang, S., Wang, X., Zhang, Y., Zhao, D., Qian, X.: Space-time prompting for video class-incremental learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11932-11942 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 531, + 481, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 531, + 481, + 564 + ], + "spans": [ + { + "bbox": [ + 132, + 531, + 481, + 564 + ], + "type": "text", + "content": "25. Peifeng, G., Xu, Q., Wen, P., Yang, Z., Shao, H., Huang, Q.: Feature directions matter: Long-tailed learning via rotated balanced representation. In: International Conference on Machine Learning. pp. 27542-27563. PMLR (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 565, + 481, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 565, + 481, + 598 + ], + "spans": [ + { + "bbox": [ + 132, + 565, + 481, + 598 + ], + "type": "text", + "content": "26. Rebuffi, S.A., Kolesnikov, A., Sperl, G., Lampert, C.H.: icarl: Incremental classifier and representation learning. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 2001-2010 (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "type": "text", + "content": "27. Saha, G., Garg, I., Roy, K.: Gradient projection memory for continual learning. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=3A0jORCNC2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "28. Simon, C., Koniusz, P., Nock, R., Harandi, M.: Adaptive subspaces for few-shot learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4136-4145 (2020)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "H. Cheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "29. Simonyan, K., Zisserman, A.: Two-stream convolutional networks for action recognition in videos. In: Advances in neural information processing systems. pp. 568-576 (2014)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "type": "text", + "content": "30. Soomro, K., Zamir, A.R., Shah, M.: Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402 (2012)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "type": "text", + "content": "31. Tran, D., Bourdev, L., Fergus, R., Torresani, L., Paluri, M.: Learning spatiotemporal features with 3d convolutional networks. In: Proceedings of the IEEE international conference on computer vision. pp. 4489-4497 (2015)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 481, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 247 + ], + "type": "text", + "content": "32. Tran, D., Wang, H., Torresani, L., Ray, J., LeCun, Y., Paluri, M.: A closer look at spatiotemporal convolutions for action recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6450-6459 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 481, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 282 + ], + "type": "text", + "content": "33. Turk, M.A., Pentland, A.P.: Face recognition using eigenfaces. In: Proceedings. 1991 IEEE computer society conference on computer vision and pattern recognition. pp. 586-587. IEEE Computer Society (1991)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 282, + 481, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 315 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 315 + ], + "type": "text", + "content": "34. Tzimiropoulos, G., Zafeiriou, S., Pantic, M.: Subspace learning from image gradient orientations. IEEE transactions on pattern analysis and machine intelligence 34(12), 2454-2466 (2012)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 481, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 358 + ], + "type": "text", + "content": "35. Villa, A., Alcázar, J.L., Alfarra, M., Alhamoud, K., Hurtado, J., Heilbron, F.C., Soto, A., Ghanem, B.: Pivot: Prompting for video continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24214-24223 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 358, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 402 + ], + "type": "text", + "content": "36. Villa, A., Alhamoud, K., Escorcia, V., Caba, F., Alcazar, J.L., Ghanem, B.: vclimb: A novel video class incremental learning benchmark. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19035-19044 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 402, + 481, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 481, + 446 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 481, + 446 + ], + "type": "text", + "content": "37. Wang, C., Guo, L., Wang, Y., Cheng, H., Yu, Y., Wen, B.: Progressive divide-and-conquer via subsampling decomposition for accelerated mri. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 25128-25137 (2024)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "type": "text", + "content": "38. Wang, J., Cherian, A.: Gods: Generalized one-class discriminative subspaces for anomaly detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8201-8211 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 479, + 481, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 512 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 512 + ], + "type": "text", + "content": "39. Wang, S., Li, X., Sun, J., Xu, Z.: Training networks in null space of feature covariance for continual learning. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 184-193 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 512, + 481, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 481, + 545 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 481, + 545 + ], + "type": "text", + "content": "40. Yang, J., Frangi, A.F., Yang, J.y., Zhang, D., Jin, Z.: Kpca plus lda: a complete kernel fisher discriminant framework for feature extraction and recognition. IEEE Transactions on pattern analysis and machine intelligence 27(2), 230-244 (2005)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 545, + 481, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 481, + 588 + ], + "type": "text", + "content": "41. Yang, S., Liu, J., Lu, S., Er, M.H., Kot, A.C.: Collaborative learning of gesture recognition and 3d hand pose estimation with multi-order feature analysis. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16. pp. 769-786. Springer (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 588, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 588, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 481, + 632 + ], + "type": "text", + "content": "42. Yang, Y., Yuan, H., Li, X., Lin, Z., Torr, P., Tao, D.: Neural collapse inspired feature-classifier alignment for few-shot class-incremental learning. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=y5W8tpoJhtJ" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "43. Zhang, L., Fu, J., Wang, S., Zhang, D., Dong, Z., Chen, C.P.: Guide subspace learning for unsupervised domain adaptation. IEEE transactions on neural networks and learning systems 31(9), 3374-3388 (2019)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 250, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 250, + 91, + 447, + 102 + ], + "type": "text", + "content": "Spatial-Temporal Subspace Projection for VCIL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 205 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "text", + "content": "44. Zhao, B., Xiao, X., Gan, G., Zhang, B., Xia, S.T.: Maintaining discrimination and fairness in class incremental learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 13208-13217 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 482, + 182 + ], + "type": "text", + "content": "45. Zhu, H., Koniusz, P.: Ease: Unsupervised discriminant subspace learning for transductive few-shot learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9078-9088 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 482, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 482, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 482, + 205 + ], + "type": "text", + "content": "46. Zhu, P., Yao, X., Wang, Y., Hui, B., Du, D., Hu, Q.: Multiview deep subspace clustering networks. IEEE Transactions on Cybernetics (2024)" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "H. Cheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/SUMix_ Mixup with Semantic and Uncertain Information/f0041215-e301-48ad-9db8-48c026069726_content_list.json b/2024/SUMix_ Mixup with Semantic and Uncertain Information/f0041215-e301-48ad-9db8-48c026069726_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c9e9debf82c0935a91b46875a9eea3532e812fd7 --- /dev/null +++ b/2024/SUMix_ Mixup with Semantic and Uncertain Information/f0041215-e301-48ad-9db8-48c026069726_content_list.json @@ -0,0 +1,1770 @@ +[ + { + "type": "text", + "text": "SUMix: Mixup with Semantic and Uncertain Information", + "text_level": 1, + "bbox": [ + 243, + 140, + 759, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Huafeng Qin $^{1,2,5,\\star,\\boxtimes}$ , Xin Jin $^{1,2,\\star}$ , Hongyu Zhu $^{1,2,\\star}$ , Hongchao Liao $^{1,2}$ , Mounim A. El-Yacoubi $^{3}$ , Xinbo Gao $^{4}$", + "bbox": [ + 230, + 210, + 769, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Chongqing Technology and Business University, Chongqing, 400067, China", + "bbox": [ + 228, + 253, + 759, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ National Research base of Intelligent Manufacturing Service, Chongqing, China", + "bbox": [ + 228, + 268, + 771, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Telecom SudParis, Institut Polytechnique de Paris, Paris, 75020, France", + "bbox": [ + 254, + 282, + 746, + 296 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 Chongqing University of Posts and Telecommunications, Chongqing, 400065, China", + "bbox": [ + 217, + 296, + 784, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "5 Chongqing Micro-Vein Intelligent Technology Co.", + "bbox": [ + 328, + 310, + 671, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{2013009, jinxin4, zhuhongyu}@ctbu.edu.cn", + "bbox": [ + 341, + 325, + 660, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$\\star$ Equal contribution, $\\boxtimes$ Corresponding author.", + "bbox": [ + 341, + 339, + 658, + 351 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Mixup data augmentation approaches have been applied for various tasks of deep learning to improve the generalization ability of deep neural networks. Some existing approaches CutMix, SaliencyMix, etc. randomly replace a patch in one image with patches from another to generate the mixed image. Similarly, the corresponding labels are linearly combined by a fixed ratio $\\lambda$ by $l$ . The objects in two images may be overlapped during the mixing process, so some semantic information is corrupted in the mixed samples. In this case, the mixed image does not match the mixed label information. Besides, such a label may mislead the deep learning model training, which results in poor performance. To solve this problem, we proposed a novel approach named SUMix to learn the mixing ratio as well as the uncertainty for the mixed samples during the training process. First, we design a learnable similarity function to compute an accurate mix ratio. Second, an approach is investigated as a regularized term to model the uncertainty of the mixed samples. We conduct experiments on five image benchmarks, and extensive experimental results imply that our method is capable of improving the performance of classifiers with different cutting-based mixup approaches. The source code is available at https://github.com/JinXins/SUMix.", + "bbox": [ + 259, + 387, + 743, + 652 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: mixup, data augmentation, image classification", + "bbox": [ + 261, + 664, + 665, + 678 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 703, + 375, + 718 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep learning has emerged as a viable alternative to traditional machine learning in various tasks, e.g. computer vision [3, 29, 34], natural language processing [10, 43, 62], and video applications [12, 37, 59], because of its robust feature representation capacity. However, deep learning requires a lot of data to train thousands of network parameters. Despite the self-supervised learning paradigm [18, 31, 33, 36] enables task agnostic pre-training without annotations, the labeled data of specific scenarios is still required and generally limited, where", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/35abdef1888190e637e8707039dc0671db77cfcbe317f6ba2a131988cdc5ab2f.jpg", + "image_caption": [ + "Fig. 1: The diagram shows different cases of raw samples that underwent the CutMix with a mixing ratio of 0.5 to obtain mixed samples. Left: the raw samples and their one-hot labels; Right: labels of the mixes obtained using a redefined mixing ratio $\\widetilde{\\lambda}$ for different cases." + ], + "image_footnote": [], + "bbox": [ + 267, + 143, + 738, + 258 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "deep neural networks (DNNs) are prone to overfitting [2]. Therefore, the representation capacity is not effectively exploited, resulting in poor discriminative performances and generalization abilities.", + "bbox": [ + 212, + 354, + 784, + 400 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Data Augmentation (DA) methods as a solution have been proposed to prevent DNNs from over-fitting. Existing methods transform original data to generate new samples for data augmentation. For example, in computer vision tasks, classical techniques such as rotating, scaling, and panning have been used to generate augmented samples for model training. Similarly, in natural language processing tasks, methods such as proximate word substitution and randomly misspelling words are employed to produce new data. These data augmentation methods have proven to be effective in improving the performance of network models. They allow the models to learn a larger variety of samples based on augmentation data. Currently, data augmentation methods can be categorized into three classes. (i) The contrast combination methods [8, 9, 17, 38, 39, 58] employs affine matrix transformations and color enhancement for data augmentation. However, the resulting images are significantly different from the original ones. (ii) Mixup method produces virtual mixup examples via a simple convex combination of pairs of examples [51, 64, 75]. (iii) The generative approaches, e.g. Generative Adversarial Networks(GANs) [13, 71], generate fake samples by adversarial learning. Overall, these data augmentation methods effectively alleviate the overfitting problem and improve the performance of deep learning models on limited training data.", + "bbox": [ + 212, + 401, + 787, + 688 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Mixup [75] method as a recent DA model produces virtual samples and labels by linearly interpolating the sample images as well as their corresponding labels, and such a simple method has proven to be effective in improving the representation capacity of the deep neural networks. In recent years, various mixup methods have been proposed for data augmentation. For example, CutMix [72] shifts the numerical mixing to spatial, where the mixing coefficient $\\lambda$ determines the size of the area of the cut and introduces the idea of Cutout [11], which cuts and mixes two samples, as shown in Figure 1. This approach not only helps the network to generalize better and improve object localization but also enables continuous performance improvement in tasks such as Pascal VOC", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Qin. et al.", + "bbox": [ + 271, + 114, + 341, + 128 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/dc0a669dcab3376c350dd9d1aea2a1e0c23ba05ce62fde6ffcc18767295ca3bd.jpg", + "image_caption": [ + "Fig. 2: The figure shows some hand-crafted mixup methods with \"Label Mismatch\" problem. In CutMix and ResizeMix, the \"Shark\" is mostly obscured, while in SaliencyMix and FMix, it is difficult to notice the \"Panda\"." + ], + "image_footnote": [], + "bbox": [ + 225, + 143, + 776, + 229 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "object detection [53] and MS-COCO image captioning [65]. In addition, CutMix improves the robustness of the model, making it perform better in the face of input corruption and cross-distribution detection. However, the CutMix method randomly selects the cropping region, which results in the background region of the target sample being clipped to the feature region of the source sample, and in the extreme case, the source sample will be completely occluded, resulting in an augmented sample that has no feature information at all, a phenomenon known as \"Label MisMatch\" shown in Figure.2. Since the one-hot label of the sample is unchanged, Decoupled Mix [42] found that the Mixup method relies heavily on the loss function, forcing the prediction of the network model to obey two higher peaks, which results in an inaccurate gradient result due to the label mismatch problem in the calculation of the loss function. This leads to an inaccurate gradient result when calculating the loss function, which affects the training of the network model.", + "bbox": [ + 212, + 319, + 784, + 529 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To improve the \"Label MisMatch\" problem, we propose SUMix, which consists of a mix ratio learning module and an uncertain estimation module. The former focuses on computing the proportion of two images and the latter aims to learn the uncertainty information of mixed samples. Firstly, we design a function to compute the semantic distance between the mixed and original samples to determine the ratio $\\lambda$ . Secondly, we present a method to learn the uncertainty", + "bbox": [ + 212, + 532, + 784, + 623 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a701f0ea1cfa235cd773f0447d6478048348d22696aaec16d632b5c317304116.jpg", + "image_caption": [ + "Fig. 3: The results illustrate the application of SUMix. Left: Top-1 Accuracy improvement from the mixup approaches with SUMix; Right: Comparison of Vanilla method, CutMix, and SUMix for CAM visualization." + ], + "image_footnote": [], + "bbox": [ + 225, + 657, + 776, + 781 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "SUMix", + "bbox": [ + 681, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "of the mixed. This adapted feature vector effectively mitigates issues related to computing the loss function caused by discrepancies in semantic and uncertainty aspects. Extensive experiments on five image classification datasets verify that our proposed SUMix can remarkably improve performances of existing mixup augmentations (as shown in Figure 3) in a plug-and-play manner while achieving better robustness. Our main contributions are as follows:", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(a) We propose a learnable metric to compute the mixed ratio by similarity between the mixed samples and the original samples.", + "(b) We further consider the uncertainty and semantic information of the mixed samples and recalculate a reasonable feature vector, providing an additional regularization loss for model training.", + "(c) Our approach helps popular Cutting-based mixup methods to improve classification tasks without spending too excessive extra time overhead." + ], + "bbox": [ + 215, + 248, + 782, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 382, + 387, + 398 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Sample-based Mixup Sample-based Mixup approaches [21-23, 30, 70] focus on obtaining more reliable mixed samples or hidden space feature maps. Usually combined with inserting prior knowledge or saliency information [54]. CutMix [72] obtains the mixed samples by randomly cutting and pasting between the two samples. Due to the large randomness of CutMix cropping, ResizeMix [52] replaces the cut operation with the resize operation. There are a series of works such as FMix [16], GridMix [1] etc, which use similar methods. To extract the saliency information within the samples during the training of the model, some work such as SaliencyMix [61], Attentive-CutMix [67] etc, use a saliency extraction module to mix the samples in a guidedness. PuzzleMix [26] extracts the saliency regions from the source and target samples respectively through the saliency capability learned from the model and then extracts the saliency regions according to the designed \"optimal transmission\". Co-Mixup [27] further improves PuzzleMix by increasing the number of mixed samples from 2 to 3. AutoMix [44] and SAMix [32] analyze the advantages and disadvantages of the hand-crafted and saliency method and divide the mixing task into two sub-tasks. AdAutoMix [51] further improves AutoMix by proposing adversarial training to train the generator and obtain \"hard mixed sample\" to train the model. Manifold Mixup [64] and AlignMix [63] moved the mixup method to hidden space by mixing the feature information of the samples.", + "bbox": [ + 212, + 414, + 787, + 719 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Label-based Mixup Label-based Mixup approaches aim to solve the problems of 1) time overhead and 2) sample-label mismatch. These methods usually use some known information to recalculate the mix ratio of labels. Saliency Grafting [50] uses the area of saliency features of each sample in a mixed sample to reconstruct a new $\\lambda$ . LUMix [57] differs from previous ways to address the label noise introduced by the mixup process instead of modeling the label noise. Methods such as TransMix [5] and TokenMix [40] use attention scores to mix the", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Qin. et al.", + "bbox": [ + 271, + 114, + 339, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "samples to compute $\\lambda$ based on the area of the mask. Mixpro [76] obtains a new $\\lambda$ by combining the attention score with the masked area and obtains a stable output with a factor $\\alpha$ . Other methods like SMMix [6], Token Labeling [24] etc, are based on attention score and masking area to get a mix ratio.", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Uncertainty Modeling The presence of naturally occurring uncertainties such as occlusion [25] and blurring [55] in the images tends to influence the training of network models. Several works [4,56,74] have attempted to model uncertainty in face recognition, age estimation, and point cloud segmentation, respectively, to obtain better robustness and generalization. In addition, it is also able to deal with Out-distribution problems. [15] proposed a Bayesian classifier for obtaining Out-of-distribution uncertainty. [49] defines a Gaussian distribution to represent the samples and uses Monte Carlo sampling to adopt multiple points from the Gaussian distribution for optimizing the metric. [68] uses an introspective modeling approach that bypasses distribution optimization to further improve the robustness of the model.", + "bbox": [ + 212, + 217, + 784, + 380 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 401, + 380, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "MixUp Classification. Let us define an $X \\in \\mathbb{X}$ as the input data and its corresponding $Y \\in \\mathbb{Y}$ as the label of the input data. For the vanilla mixup, we construct a mixed sample $\\widetilde{X}$ and a mixed label $\\widetilde{Y}$ from Eq.(1).", + "bbox": [ + 212, + 428, + 782, + 473 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\begin{array}{l} \\widetilde {X} = \\lambda * X _ {a} + (1 - \\lambda) * X _ {b}, \\\\ \\widetilde {\\tilde {\\Sigma}} = \\dots \\end{array} \\tag {1} \\\\ \\widetilde {Y} = \\lambda * Y _ {a} + (1 - \\lambda) * Y _ {b}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 403, + 479, + 784, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda$ is a mix ratio obtained from a Beta distribution, $(X_{a},X_{b})\\in \\widetilde{X}$ , and $(Y_{a},Y_{b})\\in \\widetilde{Y}$ are the same batch of data. Mixup obtains the mixed samples by pixel interpolating with a $\\lambda$ , CutMix randomly obtains a two-valued rectangular mask $\\mathbb{1}$ , where $\\mathbb{1} = \\lambda = \\frac{r_w r_h}{WH}$ , W, H are the width and height of the data and $r_w, r_h$ are the width and height of the rectangular mask. We get the mixed samples $\\widetilde{X} = \\mathbb{1}\\odot X_a + (1 - \\mathbb{1})\\odot X_b$ .", + "bbox": [ + 212, + 527, + 782, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Devil Lies in Label MisMatch. Mixup contains two different labels, so the Mixed Cross Entropy (MCE) is used in calculating the loss function as shown in Eq.(2).", + "bbox": [ + 212, + 630, + 782, + 675 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {M C E} = \\ell_ {C E} \\left(f _ {\\theta} (\\tilde {X}), Y _ {a}\\right) * \\lambda + \\ell_ {C E} \\left(f _ {\\theta} (\\tilde {X}), Y _ {b}\\right) * (1 - \\lambda), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 302, + 681, + 784, + 699 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $f_{\\theta}(\\cdot)$ denotes the network model, and $\\ell(\\cdot)$ denotes the loss function, $\\ell_{CE}$ was cross-entropy loss, which minimizes the loss during the training process. However, since some hand-crafted mixup methods are random, e.g. CutMix, ResizeMix, but the labels of the mixed samples are fixed $\\widetilde{Y}$ , then in the case shown in Figure.2, the features of a certain classification are occluded and reduced or do not exist in the mixed samples, resulting in the feature information not matching the labels. The phenomenon is called \"Label MisMatch\". In this case, the proportion of labels should be reduced or removed to get an accurate result when calculating the loss function.", + "bbox": [ + 212, + 704, + 784, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "SUMix", + "bbox": [ + 681, + 113, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7476f3d50f69c94437f390b9e1e879674e6f99f935bc9978630523c1dd76ee58.jpg", + "image_caption": [ + "Fig. 4: The left diagram represents the training process of SUMix. The raw and mixed samples are encoded through the model to obtain the semantic and uncertainty information, and the new mix ratio and loss function for regularization are obtained. The right diagram is the spatial distance between the raw and mixed samples." + ], + "image_footnote": [], + "bbox": [ + 222, + 146, + 553, + 272 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/dc9a0917fcc158fc6b2342bfd2f78f1b3a87f477f4aa4261082715c551267fca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 563, + 148, + 781, + 268 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 SUMix", + "text_level": 1, + "bbox": [ + 215, + 375, + 321, + 390 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we focus on detailing SUMix, 1) the uncertainty classifier, 2) how to obtain the mix ratio $\\lambda$ , and 3) the regularization function in the case of uncertainty and semantic information. Firstly, we show how to use deep learning classifiers for loss function computation and network model training, secondly, how to redefine the mix ratio $\\lambda$ based on semantic information, and lastly, how to consider uncertainty and regularization. The pipeline of SUMix is shown in Fig. 4.", + "bbox": [ + 212, + 414, + 787, + 521 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1 Uncertainty Classifier", + "text_level": 1, + "bbox": [ + 215, + 550, + 441, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Assuming $\\mathbb{S} = \\{x_s|s = 1,2,\\dots ,S\\}$ is a training set and $\\mathbb{S}$ is the number of training images. We take $N$ samples from it as a sample set, where $\\mathbb{X}\\in \\mathbb{R}^{N\\times 3\\times H\\times W}$ , $\\mathbb{Y}\\in \\mathbb{R}^{N\\times K}$ are the labels and $K$ is the number of classes. For the original computation of the mixup loss function, we use a deep neural network $f_{\\theta}(\\cdot)$ with a trainable parameter that maps the sample $x_{i}$ to the labels $y_{i}$ . Unlike the vanilla mixup loss computation in Eq.(2), SUMix adds a regularized loss function as shown in Eq.(3):", + "bbox": [ + 212, + 584, + 787, + 691 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s u} = \\frac {1}{N} \\left(\\sum_ {i = 0} ^ {N} \\mathcal {L} _ {M C E} \\left(f _ {\\theta} \\left(\\widetilde {X} _ {i}\\right), Y, \\widetilde {\\lambda} _ {i}\\right)\\right) + \\zeta * \\mathcal {L} _ {M C E} \\left(S U \\left(f _ {\\theta} (\\widetilde {X}), U _ {\\omega}\\right), \\widetilde {Y}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 230, + 705, + 785, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\widetilde{\\lambda}_i$ are the redefined lambda, $SU(\\cdot)$ denotes a function to obtain feature, and $\\zeta$ is a hyperparameter used to weight the regulated loss function, set to [0.1, 0.2, 0.5, 1.0]. $U_{\\omega}$ denotes a function as a measure of uncertainty. Finally, the parameters of the deep neural network are updated continuously to minimize $\\mathcal{L}_{su}$ .", + "bbox": [ + 212, + 762, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Qin. et al.", + "bbox": [ + 271, + 114, + 341, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2 Mix ratio Learning", + "text_level": 1, + "bbox": [ + 215, + 146, + 419, + 161 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Mixup approaches always depend on a parameter taken from the $Beta(\\alpha, \\alpha)$ distribution, which we call the mix ratio $\\lambda$ , to balance the mix ratio of the two samples. Lots of mixup approaches use $\\lambda$ to augment the mixup sample set $\\mathbb{X}$ after obtaining it, and notably, this $\\lambda$ serves the whole sample set. For some saliency-based or automatic-based mixups, e.g. PuzzleMix [26], AutoMix [44], it is possible to push the mixing method to obtain the best possible mixed samples according to the $\\lambda$ ratio, balancing the sample ratio with the $\\lambda$ as much as possible. However, some hand-crafted approaches with randomness cannot accurately match the mix ratio. The problem we want to solve is how to match each mixup sample with its own $\\lambda$ .", + "bbox": [ + 212, + 174, + 787, + 324 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For a mixed sample $\\widetilde{x} = \\left[\\widetilde{X}\\right]_i^N\\in R^{3\\times H\\times W}$ , semantic information links need to be established with the original sample pair $(x_{a},x_{b})$ . They are fed into a deep neural network $f_{\\theta}(\\cdot)$ , e.g. ResNet18 [19]. feature vectors $\\widetilde{z}$ and $(z_{a},z_{b})$ are obtained, and the feature vectors are normalized to modify their similarity according to Eq.(4):", + "bbox": [ + 212, + 325, + 787, + 411 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {\\lambda} _ {a} = \\frac {\\lambda * \\mathrm {e} ^ {- \\| (\\sigma (\\widetilde {z} - z _ {a}) \\| _ {2}}}{\\lambda * \\mathrm {e} ^ {- \\| (\\sigma (\\widetilde {z} - z _ {a}) \\| _ {2}} + (1 - \\lambda) * \\mathrm {e} ^ {- \\| (\\sigma (\\widetilde {z} - z _ {b}) \\| _ {2}}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 330, + 422, + 784, + 457 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\sigma(\\cdot)$ is the softmax function used to normalize the vectors, and $\\|\\cdot\\|_2$ is the l2 norm, which imposes a limit on the distance between the two. When the difference in semantic information between $\\widetilde{x}$ and $x_a$ is large, which indicates that the mixed sample has a small amount of the feature information of the raw sample, $\\widetilde{\\lambda}_a$ tends to 0. Similarly, when the difference in semantic information between the two is small, $\\widetilde{\\lambda}_a$ tends to 1. $\\widetilde{\\lambda}_b = 1 - \\widetilde{\\lambda}_a$ . Thus, we have calculated the mixed sample set with its own matching $\\lambda$ . a more sensible loss is computed according to Eq.(3).", + "bbox": [ + 212, + 469, + 787, + 590 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3 Uncertainty Estimation", + "text_level": 1, + "bbox": [ + 215, + 614, + 455, + 630 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Deep learning tasks involve two types of uncertainty: Epistemic uncertainty and Aleatoric uncertainty. Epistemic uncertainty arises from the model's inherent uncertainty due to insufficient training data, resulting in low confidence in unknown data. In contrast, Aleatoric uncertainty arises from errors in the data itself, which can introduce bias into the dataset. The greater the bias, the greater the Aleatoric uncertainty. In the case of the mixup task, the error in the data occurs when the mixup method is stochastic. Our goal is to address this error and enable the network to adapt accordingly.", + "bbox": [ + 212, + 643, + 787, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Wang et al. [68] propose an introspective metric that models a new vector for classification based on the semantic information distance between two images and their uncertainty distance. The mixup methods differ from this as they typically incorporate feature information from two samples. To begin, we encode the mixed sample and the raw sample using an encoder to obtain the feature", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "SUMix", + "bbox": [ + 681, + 113, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "vectors $\\widetilde{z}$ and $z$ , respectively. Next, we calculate the uncertainty by passing $\\widetilde{z}$ and $z$ through $U_{\\omega}$ , by Eq.(5).", + "bbox": [ + 212, + 146, + 787, + 176 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} u = \\parallel M L P (\\sigma (z)) \\| _ {2}, \\\\ \\widetilde {\\gamma} - \\parallel M L P (\\sigma (\\widetilde {z})) \\| \\end{array} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 191, + 784, + 215 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where is a linear layer with a softmax function and l2 norm. We define $\\beta = \\widetilde{u} + u$ as the uncertainty of mixed samples and raw samples.", + "bbox": [ + 212, + 236, + 785, + 266 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "After capturing the uncertainty and semantic information, as demonstrated in the right part of Figure.4, Wang et al. [68] argue that in the feature space, the smaller the radius of a sample, the more aggregated the class is, and the smaller the uncertainty. By combining the semantic information, we reconstruct the mixed samples $\\widetilde{X}$ and the raw samples $X$ to obtain $Z_{su}$ according to Eq.(6).", + "bbox": [ + 212, + 268, + 785, + 344 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nZ _ {s u} = \\mathrm {e} ^ {- (\\beta + \\| \\sigma (\\tilde {z} - z) \\| _ {2})} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 424, + 357, + 784, + 375 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "When there is significant uncertainty or semantic information has a large difference, $Z_{su}$ receives a small gradient; conversely, when there is little uncertainty and small differences in semantic information, $Z_{su}$ receives a normal gradient that tends to 1. This approach is used as a method of regularization.", + "bbox": [ + 212, + 382, + 785, + 444 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 468, + 375, + 486 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To verify the performance of SUMix, we conducted a large number of experiments on six datasets, CIFAR100 [28], Tiny-ImageNet [7], ImageNet-1K [29], CUB-200 [66], FGVC-Aircrafts [47], and CIFAR100-C. To verify the effectiveness of the improved \"Label MisMatching\" method, we mainly compared it with some patch-wise mixup methods based on hand-crafted, i.e., CutMix, FMix, ResizeMix, and SaliencyMix. We have implemented our algorithm on the open-source library OpenMixup [35], and some parameters are designed according to OpenMixup. The classification results are the median of the top-1 test accuracy from the last 10 training epochs, but on ImageNet-1K, we choose the last 5 training epochs. To facilitate comparison, our results are highlighted in bold.", + "bbox": [ + 212, + 501, + 787, + 654 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.1 Datasets", + "text_level": 1, + "bbox": [ + 215, + 676, + 333, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We choose the three most classical datasets and two fine-grained datasets in the classification task. CIFAR100 has a total of 50,000 images for the training set and 10,000 images for the testing set, containing 100 different classes, each with a $32 \\times 32$ resolution; Tiny-ImageNet contains 10,000 images for training and 10,000 images for testing, each with a $64 \\times 64$ resolution and a total of 200 classes; The ImageNet-1K dataset contains 1.2 million images for training and 50,000 images for validation; CUB200 dataset contains images of 200 different bird species, a total of 11,788 images; the FGVC-Aircrafts dataset contains images of 100 different types of aircraft, a total of 10,000 images.", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Qin. et al.", + "bbox": [ + 271, + 114, + 341, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 146, + 455, + 161 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Hyper-parameter of $\\alpha$ . In the experimental results we report, the hyperparameter $\\alpha$ of the mixup variants, which sampling from the $Beta(\\alpha, \\alpha)$ distribution used by the model based on the CNN architecture and the model based on the ViT architecture are different. For CutMix, the CNN-based models, e.g. ResNet18, ResNeXt50, and Wide-ResNet28 set as 0.2, and ViT-based models, DeiT and Swin-Transformer set as 2.0. For FMix, CNN-based set 0.2, and ViT-based set 1.0. For PuzzleMix, the CNN-based set is 1.0, and the ViT-based set is 2.0. In CNN-based and Vit-based architectures alike, SaliencyMix, ResizeMix, and AutoMix exhibit identical alpha values, which are 0.2, 1.0, and 2.0, respectively.", + "bbox": [ + 212, + 171, + 787, + 323 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Hyper-parameter of $\\zeta$ . In SUMix, we set the hyper-parameters $\\zeta$ of $\\mathcal{L}_{su}$ as follows: On the CIFAR100 dataset, we report five different models, the $\\zeta$ setting shown in Appendix A.1. On Tiny-ImageNet, we set the $\\zeta = 1.0$ for all mixup methods based on ResNet18 and ResNeXt50. Similarly with ImageNet-1K, we set the $\\zeta = 0.5$ based on ResNet18. The settings we reported for the fine-grained datasets are also shown in Appendix A.1.", + "bbox": [ + 212, + 335, + 787, + 429 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.3 Image Classification", + "text_level": 1, + "bbox": [ + 215, + 450, + 428, + 465 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Generic classification. We initially train ResNet18 [19] and ResNeXt50 [69] using the CIFAR100 dataset for a total of 800 epochs. The experimental parameters are set as in OpenMixup [35]: a basic learning rate of 0.1, SGD [45] optimizer with momentum of 0.9, weight decay of 0.0001, and batch size of 100 for each iteration. Wide-ResNet [73] using the CIFAR100 dataset for a total of 400 epochs, learning rate of 0.03, SGD optimizer with momentum of 0.9, weight decay of 0.001. For the Tiny-ImageNet dataset, we use the same experimental parameters as the CIFAR100 experiment, except the base learning rate to 0.2 and training for 400 epochs. As for ImageNet-1K, we follow the Pytorch-style settings for training ResNet18 of 100 epochs. For the CIFAR100 dataset, we use", + "bbox": [ + 212, + 474, + 787, + 627 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/88e98f1f3042dfd08fee9de0110cc97e24e9b5beb8553c2d63313f7c76807497.jpg", + "table_caption": [ + "Table 1: Top-1 accuracy $(\\%)\\uparrow$ of mixup methods on CIFAR-100, Tiny-ImageNet and ImageNet-1K. * denotes mixup methods with SUMix." + ], + "table_footnote": [], + "table_body": "
MethodCIFAR100Tiny-ImageNetImageNet-1K
ResNet18ResNeXt50W-ResNet28-8ResNet18ResNeXt50ResNet18
CutMix78.1778.3284.4565.5366.4768.95
FMix79.6979.0284.2163.4765.0869.96
SaliencyMix79.1278.7784.3564.6066.5569.16
ResizeMix80.0180.3584.8763.7465.8769.50
CutMix*79.7879.9184.5665.7168.7469.71
FMix*80.2080.7984.3263.6967.1270.48
SaliencyMix*79.9179.3284.5865.6868.9269.52
ResizeMix*80.3880.7284.9165.3067.4969.76
Avg. Gain+0.82+1.07+0.12+0.81+2.07+0.47
", + "bbox": [ + 220, + 693, + 782, + 835 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "SUMix", + "bbox": [ + 681, + 113, + 730, + 127 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/5c97baa211a7d8df1dafb85a63812fa24eb08d3c240d6934872ea1e5ac0c276c.jpg", + "table_caption": [ + "Table 2: Top-1 accuracy $(\\%)\\uparrow$ of mixup methods on CUB200, FGVC-Aircrafts, and Standford-Cars. * denotes mixup methods SUMix. † denotes that we reproduce the results of that experiment." + ], + "table_footnote": [], + "table_body": "
MethodCUB200FGVC-Aircrafts
ResNet18ResNeXt50ResNet18ResNeXt50
CutMix†77.7083.6778.8484.55
FMix77.2884.0679.3684.10
SaliencyMix†75.7782.8379.7884.31
ResizeMix78.5084.1678.1084.08
CutMix*78.2083.7179.7285.84
FMix*79.2484.3379.4884.64
SaliencyMix*76.9882.8479.9084.49
ResizeMix*78.5684.2380.2985.12
Avg. Gain+0.93+0.10+0.82+0.67
", + "bbox": [ + 287, + 198, + 707, + 359 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/7b671bdf8b2bfb6623e5a2684d78e618cd6f0a5a0e9e746f0b53a8719920bff5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCIFAR100
DeiT-SmallSwin-Tiny
CutMix74.1280.64
FMix70.4180.72
SaliencyMix69.7880.40
ResizeMix68.4580.16
CutMix*75.26+1.1480.83+0.19
FMix*70.69+0.2880.73+0.01
SaliencyMix*70.31+0.5380.71+0.29
ResizeMix*68.78+0.3380.59+0.43
", + "bbox": [ + 218, + 444, + 455, + 547 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/97d47370fc91925f037965ea0cabf071a4845e9b2ac499559f01ba0a7a4fc710.jpg", + "table_caption": [ + "Table 3: Top-1 accuracy(\\%)↑ of Table 4: Top-1 accuracy(\\%)↑ of saliency-based mixup methods on CIFAR100. * de- mixup methods on CIFAR100. * denotes mixup notes mixup methods with SUMix. methods with SUMix." + ], + "table_footnote": [], + "table_body": "
MethodCIFAR100
ResNet18ResNeXt50
PuzzleMix81.1381.69
AutoMix82.0482.84
PuzzleMix*81.43+0.3082.60+0.91
AutoMix*82.30+0.2683.82+0.98
AdAutoMix82.3283.81
", + "bbox": [ + 462, + 443, + 781, + 547 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "basic data augmentation including RandomFlip and RandomCrop with 4 pixels padding, and for the Tiny-ImageNet and ImageNet-1K datasets, we also use the RandomFlip and RandonCrop data augmentation operations.", + "bbox": [ + 212, + 578, + 785, + 625 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 1 shows our five comparisons on three different datasets, where the use of SUMix brings some improvement in the classification accuracy of CutMix, FMix, SaliencyMix, and ResizeMix. For the CIFAR100 dataset, on ResNet and its variants ResNeXt50, W-ResNet28-8, SUMix improves their classification results by an average of $0.82\\%$ , $1.07\\%$ , and $0.12\\%$ , respectively; for the Tiny-ImageNet dataset, on ResNet18 and ResNeXt50, the average improvements of $0.81\\%$ and $2.07\\%$ ; for the ImageNet-1K dataset, on ResNet18, the average improvement is $0.47\\%$ .", + "bbox": [ + 212, + 625, + 787, + 747 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Fine-grained classification. We load the official PyTorch pre-trained models on ImageNet-1k as initialization and fine-tuning pre-trained ResNet18 and ResNeXt50 using SGD optimizer with a learning rate of 0.001, a momentum of 0.9, weight decay of 0.0005 on CUB200, FGVC-Aircrafts, total training 200 epochs. Table 2 presents the comparison results for the bird and aircraft datasets.", + "bbox": [ + 212, + 763, + 787, + 842 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Qin. et al.", + "bbox": [ + 271, + 114, + 339, + 128 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/1bcf8a841b06735580a596c4f537b930fc238fc2f4cbe46957338b4881625bd1.jpg", + "image_caption": [ + "Fig. 5: The top of the figure shows a visualization of the sample at $0\\%$ to $100\\%$ occlusion ratio. The light blue curves in the lower four subfigures show the classification accuracy of CutMix, FMix, SaliencyMix, and ResizeMix on ImageNet-1K using ResNet18 for 100 epochs of training, respectively, and the red curve shows the classification accuracy of these methods using SUMix." + ], + "image_footnote": [], + "bbox": [ + 222, + 146, + 777, + 299 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "SUMix improves the classification results for the CUB200 dataset by an average of $0.93\\%$ and $0.10\\%$ at ResNet18 and ResNeXt50, respectively. The FGVC-Aircrafts dataset improves by an average of $0.82\\%$ and $0.67\\%$ at ResNet18 and ResNeXt50.", + "bbox": [ + 212, + 419, + 784, + 479 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Classification based on ViTs. To further show the performance of SUMix, we perform classification experiments using the ViT-based model DeiT [60] and Swin-Transformer [41] on the CIFAR100 dataset, and the results are shown in Tab 3. We resized the size to $224 \\times 224$ , and in the training set, we trained DeiT and Swin-Transformer of 200 epochs by AdamW [46] optimizer with a batch size of 100 on CIFAR100. The basic learning rates of DeiT and Swin are 1e-3 and 5e-4, and the weight decay of 0.05, like OpenMixup settings, we also used some RandAugment [9] operations. We reported the best top-1 accuracy in the last 10 training epochs, and the results show that our SUMix also can bring effectiveness to ViTs.", + "bbox": [ + 212, + 507, + 784, + 657 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 3 shows the results of the ViT models using SUMix, which improves the classification results by an average of $0.57\\%$ and $0.23\\%$ at DeiT-Small and Swin-Tiny.", + "bbox": [ + 212, + 660, + 784, + 705 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Classification based on Saliency methods. To further explore the capability of SUMix, we not only apply it to some Hand-crafted methods, we apply SUMix to PuzzleMix and AutoMix, which are based on saliency-learning methods, and Tab 4 shows that SUMix will also bring some boosts, PuzzleMix and AutoMix has gained $0.30\\%$ , $0.26\\%$ at ResNet18, and $0.91\\%$ , $0.98\\%$ at ResNeXt50. Those methods combined with SUMix can even reach almost close to the performance of the current state-of-the-art method AdAutoMix.", + "bbox": [ + 212, + 733, + 784, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "SUMix", + "bbox": [ + 681, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/d5615ce7be2606975bc230c4bee103c09e2d5cf71f6f5ecfd96fb4e4a1aa37e1.jpg", + "table_caption": [ + "Table 5: Top-1 acc(%)↑ and FGSM error(%)↓ Table 6: Ablation experiments on CI-of ResNet18 without and with SUMix. FAR100, CUB200 based on ResNet18." + ], + "table_footnote": [], + "table_body": "
Clean Acc(%)↑Corruption Acc(%)↑FGSM Error(%)↓
MethodMCESUMixMCESUMixMCESUMix
CutMix78.1779.7843.0644.3191.1590.41
FMix79.6980.2048.7949.1489.1689.08
SaliencyMix79.1279.9143.7344.3689.6491.49
ResizeMix80.0180.3846.1246.2890.0491.05
", + "bbox": [ + 218, + 214, + 522, + 289 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/9fd5a8fefbfe9687974816e6cc7cd7869e2cd7d39f5b844d99fad0f3b923cccf.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCIFAR100CUB200
CutMixSaliencyMixFMixSaliencyMix
baseline78.1779.1277.2875.77
+λ79.1379.7979.0376.46
+ℓs79.5178.7179.0376.34
+ℓu79.6079.4379.2476.72
+ℓsu79.7879.9179.2476.98
", + "bbox": [ + 527, + 214, + 785, + 289 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.4 Occlusion Robustness", + "text_level": 1, + "bbox": [ + 214, + 316, + 441, + 330 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To analyze the enhancement brought by SUMix to the robustness of the mixup methods, we randomly conducted masking patches experiment [48]. The experiment involved dividing the image into $16 \\times 16$ patches and masking them according to different ratios. The resulting image was then classified by the model. We used the ImageNet dataset, which contains 1000 daily life objects. It is more representative. The ResNet18 classifier was used to classify the masked images and verify the model's robustness capability. The curves at the bottom of Figure.5 show that SUMix provides a performance gain compared to the corresponding mixup method at different occlusion ratios.", + "bbox": [ + 212, + 338, + 787, + 474 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.5 Robustness", + "text_level": 1, + "bbox": [ + 214, + 494, + 356, + 508 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We first evaluated robustness against corruption on CIFAR-100-C [20], which was designed manually to provide 19 different corruptions (e.g., noise, blur, fog, etc.) to evaluate robustness. In Table 5, we show the results of improving some mixup methods with SUMix. We further investigate the FGSM [14] white box attack experiments with $8/255\\ell_{\\infty}$ epsilon ball and similarly show the results of the improvement of our approach on mixup methods in Table 5.", + "bbox": [ + 212, + 518, + 787, + 609 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.6 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 628, + 390, + 643 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In the ablation study, we primarily analyze the effect of the two SUMix modules in the model. The following two analyses are primarily done. (1) Can redefining $\\lambda$ bring gains in models? (2) Can uncertainty lead to regularization gains? Table 6 shows the experimental results for CutMix, SaliencyMix, and FMix using ResNet18 and ResNeXt50 on the CIFAR100 and CUB200 datasets.", + "bbox": [ + 212, + 651, + 787, + 741 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "(1) Table 6 shows that with our redefined $\\widetilde{\\lambda}$ , CutMix and SaliencyMix are improved by $0.96\\%$ and $0.67\\%$ on the CIFAR100 dataset, and FMix and SaliencyMix are improved by $1.75\\%$ and $0.69\\%$ on the CUB200 dataset, which means that the module can bring a significant increase in the classification ability, in other words, the redefined $\\widetilde{\\lambda}$ can obtain a more sensible loss function for the model to train.", + "bbox": [ + 215, + 747, + 787, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Qin. et al.", + "bbox": [ + 271, + 114, + 339, + 127 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "(2) To calculate the regularization loss, we consider three cases: 1) using only the semantic information $\\ell_s$ , 2) using only the uncertainty information $\\ell_u$ , and 3) using both the semantic and uncertainty information $\\ell_{su}$ . Table 6 shows that on the CUB200 dataset, FMix and SaliencyMix improve accuracy by $1.75\\%$ and $0.57\\%$ respectively when using only $\\ell_s$ , and by $1.96\\%$ and $0.95\\%$ respectively when using only $\\ell_u$ . When both types of information are used, accuracy improves by $1.75\\%$ and $1.21\\%$ . Similarly, on the CIFAR100 dataset, there was a $1.61\\%$ and $0.79\\%$ improvement. The experimental results show that combining semantic and uncertainty information can significantly improve accuracy.", + "bbox": [ + 217, + 146, + 787, + 297 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "6 Conclusion and Limitations", + "text_level": 1, + "bbox": [ + 215, + 321, + 517, + 338 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Conclusion. In this paper, we propose the SUMix approach to regularize the DNN in terms of the semantic information and uncertainty between the raw and mixed samples. Specifically, the semantic information difference between the raw and mixed samples is used to obtain the mix ratio $\\lambda$ , while the uncertainty is, in turn, combined with the semantic information to obtain a regularized loss function. Experimental results on five datasets proved the effectiveness of SUMix for the regularization of the mixup methods.", + "bbox": [ + 215, + 353, + 787, + 459 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Limitations. Currently, SUMix is due to learn how to obtain a sensible mixing ratio $\\lambda$ and an uncertainty-based feature vector. These are currently proven only on classification tasks. In future work, we would like to apply and improve SUMix accordingly on some downstream tasks e.g. segmentation, detection, and other methods. And for semi-supervised tasks, we can get more accurate pseudo-labels for training models.", + "bbox": [ + 215, + 474, + 787, + 566 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 215, + 589, + 393, + 606 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "This work is supported in part by the Scientific Innovation 2030 Major Project for New Generation of AI under Grant 2020AAA0107300, in part by the National Natural Science Foundation of China (Grant No. 61976030), in part by Chongqing Natural Science Foundation (Grant No. 2024NSCQ-MSX4893), in part by Project of CQ CSTC (Grant No. cstc2018jcyjAX 0057), in part by Innovative Research Projects for Postgraduate Students of Chongqing Business University (No. yjscxx2024-284-53, No. yiscxx2024-284-231, No. yjscxx2024-284-234).", + "bbox": [ + 215, + 621, + 787, + 742 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 766, + 323, + 781 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Baek, K., Bang, D., Shim, H.: Gridmix: Strong regularization through local context mapping. Pattern Recognition 109 (2021) 4", + "2. Bishop, C.M.: Pattern recognition and machine learning. Springer (2006) 2" + ], + "bbox": [ + 222, + 796, + 782, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "SUMix", + "bbox": [ + 681, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "3. Bochkovskiy, A., Wang, C.Y., Liao, H.Y.M.: Yolov4: Optimal speed and accuracy of object detection. ArXiv abs/2004.10934 (2020) 1", + "4. Chang, J., Lan, Z., Cheng, C., Wei, Y.: Data uncertainty learning in face recognition. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5710-5719 (2020) 5", + "5. Chen, J.N., Sun, S., He, J., Torr, P.H., Yuille, A., Bai, S.: Transmix: Attend to mix for vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12135-12144 (2022) 4", + "6. Chen, M., Lin, M., Lin, Z., Zhang, Y., Chao, F., Ji, R.: Smmix: Self-motivated image mixing for vision transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17260-17270 (2023) 5", + "7. Chrabaszcz, P., Loshchilov, I., Hutter, F.: A downsampled variant of imagenet as an alternative to the cifar datasets. arXiv preprint arXiv:1707.08819 (2017) 8", + "8. Cubuk, E.D., Zoph, B., Mane, D., Vasudevan, V., Le, Q.V.: Autoaugment: Learning augmentation policies from data. arXiv preprint arXiv:1805.09501 (2018) 2", + "9. Cubuk, E.D., Zoph, B., Shlens, J., Le, Q.V.: Randaugment: Practical automated data augmentation with a reduced search space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 702-703 (2020) 2, 11", + "10. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018) 1", + "11. DeVries, T., Taylor, G.W.: Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552 (2017) 2", + "12. Feichtenhofer, C., Fan, H., Malik, J., He, K.: Slowfast networks for video recognition. In: ICCV. pp. 6201-6210 (2019) 1", + "13. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. Communications of the ACM 63(11), 139-144 (2020) 2", + "14. Goodfellow, I.J., Shlens, J., Szegedy, C.: Explaining and harnessing adversarial examples. In: ICLR (2015) 12", + "15. Guénais, T., Vamvourellis, D., Yacoby, Y., Doshi-Velez, F., Pan, W.: Bacoun: Bayesian classifiers with out-of-distribution uncertainty. arXiv preprint arXiv:2007.06096 (2020) 5", + "16. Harris, E., Marcu, A., Painter, M., Niranjan, M., Hare, A.P.B.J.: Fmix: Enhancing mixed sample data augmentation. arXiv preprint arXiv:2002.12047 (2020) 4", + "17. Hataya, R., Zdenek, J., Yoshizoe, K., Nakayama, H.: Faster autoaugment: Learning augmentation strategies using backpropagation. In: European Conference on Computer Vision. pp. 1-16. Springer (2020) 2", + "18. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: CVPR. pp. 9729-9738 (2020) 1", + "19. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 770-778 (2016) 7, 9", + "20. Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261 (2019) 12", + "21. Hendrycks, D., Mu, N., Cubuk, E.D., Zoph, B., Gilmer, J., Lakshminarayanan, B.: Augmix: A simple data processing method to improve robustness and uncertainty. arXiv preprint arXiv:1912.02781 (2019) 4" + ], + "bbox": [ + 217, + 146, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Qin. et al.", + "bbox": [ + 271, + 114, + 339, + 127 + ], + "page_idx": 13 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "22. Hendrycks, D., Zou, A., Mazeika, M., Tang, L., Li, B., Song, D., Steinhardt, J.: Pixmix: Dreamlike pictures comprehensively improve safety measures. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022) 4", + "23. Huang, S., Wang, X., Tao, D.: Snapmix: Semantically proportional mixing for augmenting fine-grained data. In: AAAI (2020) 4", + "24. Jiang, Z., Hou, Q., Yuan, L., Zhou, D., Jin, X., Wang, A., Feng, J.: Token labeling: Training a $85.4\\%$ top-1 accuracy vision transformer with 56m parameters onImagenet. ArXiv abs/2104.10858 (2021) 5", + "25. Kendall, A., Gal, Y.: What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems 30 (2017) 5", + "26. Kim, J.H., Choo, W., Song, H.O.: Puzzle mix: Exploiting saliency and local statistics for optimal mixup. In: International Conference on Machine Learning. pp. 5275-5285. PMLR (2020) 4, 7", + "27. Kim, J., Choo, W., Jeong, H., Song, H.O.: Co-mixup: Saliency guided joint mixup with supermodular diversity. In: International Conference on Learning Representations (2020) 4", + "28. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009) 8", + "29. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. In: Advances in neural information processing systems. pp. 1097-1105 (2012) 1, 8", + "30. Lee, J.H., Zaheer, M.Z., Astrid, M., Lee, S.I.: Smoothmix: a simple yet effective data augmentation to train robust classifiers. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 756-757 (2020) 4", + "31. Li, S., Jin, W., Wang, Z., Wu, F., Liu, Z., Tan, C., Li, S.Z.: Semireward: A general reward model for semi-supervised learning. In: International Conference on Learning Representations (2024) 1", + "32. Li, S., Liu, Z., Wang, Z., Wu, D., Liu, Z., Li, S.Z.: Boosting discriminative visual representation learning with scenario-agnostic mixup. ArXiv abs/2111.15454 (2021) 4", + "33. Li, S., Liu, Z., Zang, Z., Wu, D., Chen, Z., Li, S.Z.: Genurl: A general framework for unsupervised representation learning. IEEE Transactions on Neural Networks and Learning Systems (2024) 1", + "34. Li, S., Wang, Z., Liu, Z., Tan, C., Lin, H., Wu, D., Chen, Z., Zheng, J., Li, S.Z.: Efficient multi-order gated aggregation network. ArXiv abs/2211.03295 (2022) 1", + "35. Li, S., Wang, Z., Liu, Z., Wu, D., Li, S.Z.: Openmixup: Open mixup toolbox and benchmark for visual representation learning. https://github.com/Westlake-AI/openmixup (2022) 8, 9", + "36. Li, S., Wu, D., Wu, F., Zang, Z., Li, S.Z.: Architecture-agnostic masked image modeling – from vit back to cnn. In: International Conference on Machine Learning (2023) 1", + "37. Li, S., Zhang, Z., Liu, Z., Wang, A., Qiu, L., Du, F.: Tlpg-tracker: Joint learning of target localization and proposal generation for visual tracking. In: IJCAI (2020) 1", + "38. Li, Y., Hu, G., Wang, Y., Hospedales, T., Robertson, N.M., Yang, Y.: Differentiable automatic data augmentation. In: European Conference on Computer Vision. pp. 580-595. Springer (2020) 2" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "SUMix", + "bbox": [ + 681, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "39. Lim, S., Kim, I., Kim, T., Kim, C., Kim, S.: Fast autoaugment. In: Advances in Neural Information Processing Systems (NeurIPS) (2019) 2", + "40. Liu, J., Liu, B., Zhou, H., Li, H., Liu, Y.: Tokenmix: Rethinking image mixing for data augmentation in vision transformers. In: European Conference on Computer Vision (2022) 4", + "41. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: International Conference on Computer Vision (ICCV) (2021) 11", + "42. Liu, Z., Li, S., Wang, G., Wu, L., Tan, C., Li, S.Z.: Harnessing hard mixed samples with decoupled regularizer. In: Thirty-seventh Conference on Neural Information Processing Systems (2023) 3", + "43. Liu, Z., Li, S., Wang, L., Wang, Z., Liu, Y., Li, S.Z.: Short-long convolutions help hardware-efficient linear attention to focus on long sequences. In: Forty-first International Conference on Machine Learning (2024) 1", + "44. Liu, Z., Li, S., Wu, D., Liu, Z., Chen, Z., Wu, L., Li, S.Z.: Automix: Unveiling the power of mixup for stronger classifiers. In: European Conference on Computer Vision. pp. 441-458. Springer (2022) 4, 7", + "45. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983 (2016) 9", + "46. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: ICLR (2019) 11", + "47. Maji, S., Rahtu, E., Kannala, J., Blaschko, M., Vedaldi, A.: Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151 (2013) 8", + "48. Naseer, M., Ranasinghe, K., Khan, S.H., Hayat, M., Khan, F.S., Yang, M.H.: Intriguing properties of vision transformers. In: Neural Information Processing Systems (2021) 12", + "49. Oh, S.J., Murphy, K., Pan, J., Roth, J., Schroff, F., Gallagher, A.: Modeling uncertainty with hedged instance embedding. arXiv preprint arXiv:1810.00319 (2018) 5", + "50. Park, J., Yang, J.Y., Shin, J., Hwang, S.J., Yang, E.: Saliency grafting: Innocuous attribution-guided mixup with calibrated label mixing. In: AAAI Conference on Artificial Intelligence (2021) 4", + "51. Qin, H., Jin, X., Jiang, Y., El-Yacoubi, M.A., Gao, X.: Adversarial automixup. arXiv preprint arXiv:2312.11954 (2023) 2, 4", + "52. Qin, J., Fang, J., Zhang, Q., Liu, W., Wang, X., Wang, X.: Resizemix: Mixing data with preserved object information and true labels. arXiv preprint arXiv:2012.11101 (2020) 4", + "53. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. arXiv preprint arXiv:1506.01497 (2015) 3", + "54. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Gradcam: Visual explanations from deep networks via gradient-based localization. arXiv preprint arXiv:1610.02391 (2019) 4", + "55. Shaw, G., Manolakis, D.: Signal processing for hyperspectral image exploitation. IEEE Signal processing magazine 19(1), 12-16 (2002) 5", + "56. Shi, Y., Jain, A.K.: Probabilistic face embeddings. In: ICCV. pp. 6902-6911 (2019) 5", + "57. Sun, S., Chen, J.N., He, R., Yuille, A., Torr, P., Bai, S.: Lumix: Improving mixup by better modelling label uncertainty. arXiv preprint arXiv:2211.15846 (2022) 4", + "58. Suzuki, T.: Teachaugment: Data augmentation optimization using teacher knowledge. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10904-10914 (June 2022) 2" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Qin. et al.", + "bbox": [ + 271, + 114, + 339, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "59. Tan, C., Gao, Z., Wu, L., Xu, Y., Xia, J., Li, S., Li, S.Z.: Temporal attention unit: Towards efficient spatiotemporal predictive learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18770-18782 (2023) 1", + "60. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jegou, H.: Training data-efficient image transformers & distillation through attention. In: International Conference on Machine Learning (ICML). pp. 10347-10357 (2021) 11", + "61. Uddin, A.S., Monira, M.S., Shin, W., Chung, T., Bae, S.H.: Saliencymix: A saliency guided data augmentation strategy for better regularization. In: International Conference on Learning Representations (2020) 4", + "62. Vaswani, A., Shazeer, N.M., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NIPS (2017) 1", + "63. Venkataramanan, S., Avrithis, Y., Kijak, E., Amsaleg, L.: Alignmix: Improving representation by interpolating aligned features. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022) 4", + "64. Verma, V., Lamb, A., Beckham, C., Najafi, A., Mitliagkas, I., Lopez-Paz, D., Bengio, Y.: Manifold mixup: Better representations by interpolating hidden states. In: International Conference on Machine Learning. pp. 6438-6447 (2019) 2, 4", + "65. Vinyals, O., Toshev, A., Bengio, S., Erhan, D.: Show and tell: A neural image caption generator. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3156-3164 (2015) 3", + "66. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The caltech-ucsd birds-200-2011 dataset. California Institute of Technology (2011) 8", + "67. Walawalkar, D., Shen, Z., Liu, Z., Savvides, M.: Attentive cutmix: An enhanced data augmentation approach for deep learning based image classification. In: IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 3642-3646 (2020) 4", + "68. Wang, C., Zheng, W., Zhu, Z., Zhou, J., Lu, J.: Introspective deep metric learning. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023) 5, 7, 8", + "69. Xie, S., Girshick, R., Dollár, P., Tu, Z., He, K.: Aggregated residual transformations for deep neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1492-1500 (2017) 9", + "70. Yang, L., Li, X., Zhao, B., Song, R., Yang, J.: Recursiveemix: Mixed learning with history. Advances in Neural Information Processing Systems 35, 8427-8440 (2022) 4", + "71. Yang, S., Wu, Y., Jin, X., El Yacoubi, M., Qin, H.: Cgan-da: A cross-modality domain adaptation model for hand-vein biometric-based authentication. JOURNAL OF Cyber-Physical-Social Intelligence 1, 3–12 (2022) 2", + "72. Yun, S., Han, D., Oh, S.J., Chun, S., Choe, J., Yoo, Y.: Cutmix: Regularization strategy to train strong classifiers with localizable features. In: International Conference on Computer Vision (ICCV). pp. 6023-6032 (2019) 2, 4", + "73. Zagoruyko, S., Komodakis, N.: Wide residual networks. In: Proceedings of the British Machine Vision Conference (BMVC) (2016) 9", + "74. Zhang, B., Wonka, P.: Point cloud instance segmentation using probabilistic embeddings. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8883-8892 (2021) 5", + "75. Zhang, H., Cisse, M., Dauphin, Y.N., Lopez-Paz, D.: mixup: Beyond empirical risk minimization. In: International Conference on Learning Representations (2018) 2", + "76. Zhao, Q., Huang, Y., Hu, W., Zhang, F., Liu, J.: Mixpro: Data augmentation with maskmix and progressive attention labeling for vision transformer. In: The Eleventh International Conference on Learning Representations (2023) 5" + ], + "bbox": [ + 212, + 146, + 787, + 833 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "SUMix", + "bbox": [ + 681, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Appendix A", + "text_level": 1, + "bbox": [ + 215, + 143, + 338, + 162 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "A.1 More Information about $\\zeta$", + "text_level": 1, + "bbox": [ + 215, + 172, + 472, + 186 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "We reported the hyper-parameter $\\zeta$ of $\\mathcal{L}_{su}$ on the CIFAR100 dataset shown in Table 7. And for the fine-grained dataset, CUB200 and FGVC-Aircraftd. We reported the settings in Table 8 and Table 9.", + "bbox": [ + 212, + 191, + 787, + 236 + ], + "page_idx": 17 + }, + { + "type": "table", + "img_path": "images/a552b9840f7e253ef0298bab9ac0e628680d63bf1370887753c9d6990e79af8a.jpg", + "table_caption": [ + "Table 7: The $\\zeta$ setting on the CIFAR100 dataset." + ], + "table_footnote": [], + "table_body": "
ζ settingCIFAR100
ResNet18ResNeXt50W-ResNet28-8DeiT-SmallSwin-Tiny
CutMix0.50.50.50.50.5
FMix1.01.00.50.50.5
SaliencyMix0.51.00.50.50.5
ResizeMix0.20.50.50.50.5
", + "bbox": [ + 222, + 265, + 781, + 361 + ], + "page_idx": 17 + }, + { + "type": "table", + "img_path": "images/6fc94597c5c5571b1f9eb01d63f2ea99bd6972e3c032ac1264d6680a5bb436f1.jpg", + "table_caption": [ + "Table 8: The $\\zeta$ setting on the CUB200 Table 9: The $\\zeta$ setting on the FGVC-dataset." + ], + "table_footnote": [], + "table_body": "
ζ settingCUB200ζ settingFGVC-Aircrafts
ResNet18ResNeXt50ResNet18ResNeXt50
CutMix1.00.5CutMix1.00.5
FMix0.50.1FMix0.50.5
SaliencyMix0.50.2SaliencyMix1.00.5
ResizeMix0.50.5ResizeMix0.50.5
", + "bbox": [ + 222, + 407, + 782, + 506 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "A.2 More Visualization Comparison", + "text_level": 1, + "bbox": [ + 215, + 554, + 516, + 569 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "In Figure.6 and Figure.7, we show the visualization obtained by the mixup method using the CAM method with and without SUMix.", + "bbox": [ + 212, + 575, + 785, + 604 + ], + "page_idx": 17 + }, + { + "type": "image", + "img_path": "images/5dda73e6bbf47db8e12e5c3d89e3ee6affe782ef8da874a5745f9e8923fdf298.jpg", + "image_caption": [ + "Fig.6: CAM visualization of the ResNet18 with and without SUMix trained on the ImageNet-1K dataset." + ], + "image_footnote": [], + "bbox": [ + 261, + 633, + 738, + 792 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Qin. et al.", + "bbox": [ + 271, + 114, + 339, + 127 + ], + "page_idx": 17 + }, + { + "type": "image", + "img_path": "images/b9cf32e5858c579bbd311bb0e69ca9e3dd749b7cde0ab31d96230aa9fb51a7a9.jpg", + "image_caption": [ + "Fig. 7: CAM visualization of the ResNet18 with and without SUMix trained on the ImageNet-1K dataset." + ], + "image_footnote": [], + "bbox": [ + 240, + 224, + 754, + 714 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "SUMix", + "bbox": [ + 681, + 114, + 732, + 126 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/SUMix_ Mixup with Semantic and Uncertain Information/f0041215-e301-48ad-9db8-48c026069726_model.json b/2024/SUMix_ Mixup with Semantic and Uncertain Information/f0041215-e301-48ad-9db8-48c026069726_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6e107882a570eebd3e83d7b50714d6be41e784d5 --- /dev/null +++ b/2024/SUMix_ Mixup with Semantic and Uncertain Information/f0041215-e301-48ad-9db8-48c026069726_model.json @@ -0,0 +1,2636 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.244, + 0.141, + 0.761, + 0.185 + ], + "angle": 0, + "content": "SUMix: Mixup with Semantic and Uncertain Information" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.211, + 0.77, + 0.244 + ], + "angle": 0, + "content": "Huafeng Qin\\(^{1,2,5,\\star,\\boxtimes}\\), Xin Jin\\(^{1,2,\\star}\\), Hongyu Zhu\\(^{1,2,\\star}\\), Hongchao Liao\\(^{1,2}\\), Mounim A. El-Yacoubi\\(^{3}\\), Xinbo Gao\\(^{4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.255, + 0.761, + 0.269 + ], + "angle": 0, + "content": "1 Chongqing Technology and Business University, Chongqing, 400067, China" + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.269, + 0.772, + 0.283 + ], + "angle": 0, + "content": "\\(^{2}\\) National Research base of Intelligent Manufacturing Service, Chongqing, China" + }, + { + "type": "text", + "bbox": [ + 0.255, + 0.283, + 0.748, + 0.297 + ], + "angle": 0, + "content": "3 Telecom SudParis, Institut Polytechnique de Paris, Paris, 75020, France" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.297, + 0.785, + 0.311 + ], + "angle": 0, + "content": "4 Chongqing University of Posts and Telecommunications, Chongqing, 400065, China" + }, + { + "type": "text", + "bbox": [ + 0.33, + 0.311, + 0.672, + 0.325 + ], + "angle": 0, + "content": "5 Chongqing Micro-Vein Intelligent Technology Co." + }, + { + "type": "text", + "bbox": [ + 0.343, + 0.326, + 0.661, + 0.339 + ], + "angle": 0, + "content": "{2013009, jinxin4, zhuhongyu}@ctbu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.343, + 0.34, + 0.659, + 0.352 + ], + "angle": 0, + "content": "\\(\\star\\) Equal contribution, \\(\\boxtimes\\) Corresponding author." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.388, + 0.744, + 0.653 + ], + "angle": 0, + "content": "Abstract. Mixup data augmentation approaches have been applied for various tasks of deep learning to improve the generalization ability of deep neural networks. Some existing approaches CutMix, SaliencyMix, etc. randomly replace a patch in one image with patches from another to generate the mixed image. Similarly, the corresponding labels are linearly combined by a fixed ratio \\(\\lambda\\) by \\(l\\). The objects in two images may be overlapped during the mixing process, so some semantic information is corrupted in the mixed samples. In this case, the mixed image does not match the mixed label information. Besides, such a label may mislead the deep learning model training, which results in poor performance. To solve this problem, we proposed a novel approach named SUMix to learn the mixing ratio as well as the uncertainty for the mixed samples during the training process. First, we design a learnable similarity function to compute an accurate mix ratio. Second, an approach is investigated as a regularized term to model the uncertainty of the mixed samples. We conduct experiments on five image benchmarks, and extensive experimental results imply that our method is capable of improving the performance of classifiers with different cutting-based mixup approaches. The source code is available at https://github.com/JinXins/SUMix." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.665, + 0.666, + 0.679 + ], + "angle": 0, + "content": "Keywords: mixup, data augmentation, image classification" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.704, + 0.376, + 0.719 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Deep learning has emerged as a viable alternative to traditional machine learning in various tasks, e.g. computer vision [3, 29, 34], natural language processing [10, 43, 62], and video applications [12, 37, 59], because of its robust feature representation capacity. However, deep learning requires a lot of data to train thousands of network parameters. Despite the self-supervised learning paradigm [18, 31, 33, 36] enables task agnostic pre-training without annotations, the labeled data of specific scenarios is still required and generally limited, where" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.342, + 0.129 + ], + "angle": 0, + "content": "Qin. et al." + }, + { + "type": "image", + "bbox": [ + 0.268, + 0.145, + 0.739, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.27, + 0.788, + 0.325 + ], + "angle": 0, + "content": "Fig. 1: The diagram shows different cases of raw samples that underwent the CutMix with a mixing ratio of 0.5 to obtain mixed samples. Left: the raw samples and their one-hot labels; Right: labels of the mixes obtained using a redefined mixing ratio \\(\\widetilde{\\lambda}\\) for different cases." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.356, + 0.785, + 0.401 + ], + "angle": 0, + "content": "deep neural networks (DNNs) are prone to overfitting [2]. Therefore, the representation capacity is not effectively exploited, resulting in poor discriminative performances and generalization abilities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.402, + 0.789, + 0.689 + ], + "angle": 0, + "content": "Data Augmentation (DA) methods as a solution have been proposed to prevent DNNs from over-fitting. Existing methods transform original data to generate new samples for data augmentation. For example, in computer vision tasks, classical techniques such as rotating, scaling, and panning have been used to generate augmented samples for model training. Similarly, in natural language processing tasks, methods such as proximate word substitution and randomly misspelling words are employed to produce new data. These data augmentation methods have proven to be effective in improving the performance of network models. They allow the models to learn a larger variety of samples based on augmentation data. Currently, data augmentation methods can be categorized into three classes. (i) The contrast combination methods [8, 9, 17, 38, 39, 58] employs affine matrix transformations and color enhancement for data augmentation. However, the resulting images are significantly different from the original ones. (ii) Mixup method produces virtual mixup examples via a simple convex combination of pairs of examples [51, 64, 75]. (iii) The generative approaches, e.g. Generative Adversarial Networks(GANs) [13, 71], generate fake samples by adversarial learning. Overall, these data augmentation methods effectively alleviate the overfitting problem and improve the performance of deep learning models on limited training data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Mixup [75] method as a recent DA model produces virtual samples and labels by linearly interpolating the sample images as well as their corresponding labels, and such a simple method has proven to be effective in improving the representation capacity of the deep neural networks. In recent years, various mixup methods have been proposed for data augmentation. For example, CutMix [72] shifts the numerical mixing to spatial, where the mixing coefficient \\(\\lambda\\) determines the size of the area of the cut and introduces the idea of Cutout [11], which cuts and mixes two samples, as shown in Figure 1. This approach not only helps the network to generalize better and improve object localization but also enables continuous performance improvement in tasks such as Pascal VOC" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.683, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SUMix" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.145, + 0.777, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.243, + 0.785, + 0.285 + ], + "angle": 0, + "content": "Fig. 2: The figure shows some hand-crafted mixup methods with \"Label Mismatch\" problem. In CutMix and ResizeMix, the \"Shark\" is mostly obscured, while in SaliencyMix and FMix, it is difficult to notice the \"Panda\"." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.32, + 0.785, + 0.53 + ], + "angle": 0, + "content": "object detection [53] and MS-COCO image captioning [65]. In addition, CutMix improves the robustness of the model, making it perform better in the face of input corruption and cross-distribution detection. However, the CutMix method randomly selects the cropping region, which results in the background region of the target sample being clipped to the feature region of the source sample, and in the extreme case, the source sample will be completely occluded, resulting in an augmented sample that has no feature information at all, a phenomenon known as \"Label MisMatch\" shown in Figure.2. Since the one-hot label of the sample is unchanged, Decoupled Mix [42] found that the Mixup method relies heavily on the loss function, forcing the prediction of the network model to obey two higher peaks, which results in an inaccurate gradient result due to the label mismatch problem in the calculation of the loss function. This leads to an inaccurate gradient result when calculating the loss function, which affects the training of the network model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.534, + 0.785, + 0.624 + ], + "angle": 0, + "content": "To improve the \"Label MisMatch\" problem, we propose SUMix, which consists of a mix ratio learning module and an uncertain estimation module. The former focuses on computing the proportion of two images and the latter aims to learn the uncertainty information of mixed samples. Firstly, we design a function to compute the semantic distance between the mixed and original samples to determine the ratio \\(\\lambda\\). Secondly, we present a method to learn the uncertainty" + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.658, + 0.777, + 0.782 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.794, + 0.785, + 0.834 + ], + "angle": 0, + "content": "Fig. 3: The results illustrate the application of SUMix. Left: Top-1 Accuracy improvement from the mixup approaches with SUMix; Right: Comparison of Vanilla method, CutMix, and SUMix for CAM visualization." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.341, + 0.129 + ], + "angle": 0, + "content": "Qin. et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "of the mixed. This adapted feature vector effectively mitigates issues related to computing the loss function caused by discrepancies in semantic and uncertainty aspects. Extensive experiments on five image classification datasets verify that our proposed SUMix can remarkably improve performances of existing mixup augmentations (as shown in Figure 3) in a plug-and-play manner while achieving better robustness. Our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.249, + 0.784, + 0.278 + ], + "angle": 0, + "content": "(a) We propose a learnable metric to compute the mixed ratio by similarity between the mixed samples and the original samples." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.28, + 0.784, + 0.325 + ], + "angle": 0, + "content": "(b) We further consider the uncertainty and semantic information of the mixed samples and recalculate a reasonable feature vector, providing an additional regularization loss for model training." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.327, + 0.784, + 0.356 + ], + "angle": 0, + "content": "(c) Our approach helps popular Cutting-based mixup methods to improve classification tasks without spending too excessive extra time overhead." + }, + { + "type": "list", + "bbox": [ + 0.216, + 0.249, + 0.784, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.383, + 0.388, + 0.399 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.415, + 0.789, + 0.72 + ], + "angle": 0, + "content": "Sample-based Mixup Sample-based Mixup approaches [21-23, 30, 70] focus on obtaining more reliable mixed samples or hidden space feature maps. Usually combined with inserting prior knowledge or saliency information [54]. CutMix [72] obtains the mixed samples by randomly cutting and pasting between the two samples. Due to the large randomness of CutMix cropping, ResizeMix [52] replaces the cut operation with the resize operation. There are a series of works such as FMix [16], GridMix [1] etc, which use similar methods. To extract the saliency information within the samples during the training of the model, some work such as SaliencyMix [61], Attentive-CutMix [67] etc, use a saliency extraction module to mix the samples in a guidedness. PuzzleMix [26] extracts the saliency regions from the source and target samples respectively through the saliency capability learned from the model and then extracts the saliency regions according to the designed \"optimal transmission\". Co-Mixup [27] further improves PuzzleMix by increasing the number of mixed samples from 2 to 3. AutoMix [44] and SAMix [32] analyze the advantages and disadvantages of the hand-crafted and saliency method and divide the mixing task into two sub-tasks. AdAutoMix [51] further improves AutoMix by proposing adversarial training to train the generator and obtain \"hard mixed sample\" to train the model. Manifold Mixup [64] and AlignMix [63] moved the mixup method to hidden space by mixing the feature information of the samples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Label-based Mixup Label-based Mixup approaches aim to solve the problems of 1) time overhead and 2) sample-label mismatch. These methods usually use some known information to recalculate the mix ratio of labels. Saliency Grafting [50] uses the area of saliency features of each sample in a mixed sample to reconstruct a new \\(\\lambda\\). LUMix [57] differs from previous ways to address the label noise introduced by the mixup process instead of modeling the label noise. Methods such as TransMix [5] and TokenMix [40] use attention scores to mix the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.682, + 0.114, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SUMix" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "samples to compute \\(\\lambda\\) based on the area of the mask. Mixpro [76] obtains a new \\(\\lambda\\) by combining the attention score with the masked area and obtains a stable output with a factor \\(\\alpha\\). Other methods like SMMix [6], Token Labeling [24] etc, are based on attention score and masking area to get a mix ratio." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.218, + 0.785, + 0.381 + ], + "angle": 0, + "content": "Uncertainty Modeling The presence of naturally occurring uncertainties such as occlusion [25] and blurring [55] in the images tends to influence the training of network models. Several works [4,56,74] have attempted to model uncertainty in face recognition, age estimation, and point cloud segmentation, respectively, to obtain better robustness and generalization. In addition, it is also able to deal with Out-distribution problems. [15] proposed a Bayesian classifier for obtaining Out-of-distribution uncertainty. [49] defines a Gaussian distribution to represent the samples and uses Monte Carlo sampling to adopt multiple points from the Gaussian distribution for optimizing the metric. [68] uses an introspective modeling approach that bypasses distribution optimization to further improve the robustness of the model." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.402, + 0.382, + 0.418 + ], + "angle": 0, + "content": "3 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.429, + 0.784, + 0.474 + ], + "angle": 0, + "content": "MixUp Classification. Let us define an \\( X \\in \\mathbb{X} \\) as the input data and its corresponding \\( Y \\in \\mathbb{Y} \\) as the label of the input data. For the vanilla mixup, we construct a mixed sample \\( \\widetilde{X} \\) and a mixed label \\( \\widetilde{Y} \\) from Eq.(1)." + }, + { + "type": "equation", + "bbox": [ + 0.405, + 0.481, + 0.785, + 0.52 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\begin{array}{l} \\widetilde {X} = \\lambda * X _ {a} + (1 - \\lambda) * X _ {b}, \\\\ \\widetilde {\\tilde {\\Sigma}} = \\dots \\end{array} \\tag {1} \\\\ \\widetilde {Y} = \\lambda * Y _ {a} + (1 - \\lambda) * Y _ {b}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.528, + 0.784, + 0.62 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is a mix ratio obtained from a Beta distribution, \\((X_{a},X_{b})\\in \\widetilde{X}\\), and \\((Y_{a},Y_{b})\\in \\widetilde{Y}\\) are the same batch of data. Mixup obtains the mixed samples by pixel interpolating with a \\(\\lambda\\), CutMix randomly obtains a two-valued rectangular mask \\(\\mathbb{1}\\), where \\(\\mathbb{1} = \\lambda = \\frac{r_w r_h}{WH}\\), W, H are the width and height of the data and \\(r_w, r_h\\) are the width and height of the rectangular mask. We get the mixed samples \\(\\widetilde{X} = \\mathbb{1}\\odot X_a + (1 - \\mathbb{1})\\odot X_b\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.631, + 0.784, + 0.676 + ], + "angle": 0, + "content": "The Devil Lies in Label MisMatch. Mixup contains two different labels, so the Mixed Cross Entropy (MCE) is used in calculating the loss function as shown in Eq.(2)." + }, + { + "type": "equation", + "bbox": [ + 0.303, + 0.682, + 0.785, + 0.7 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {M C E} = \\ell_ {C E} \\left(f _ {\\theta} (\\tilde {X}), Y _ {a}\\right) * \\lambda + \\ell_ {C E} \\left(f _ {\\theta} (\\tilde {X}), Y _ {b}\\right) * (1 - \\lambda), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.785, + 0.84 + ], + "angle": 0, + "content": "where \\( f_{\\theta}(\\cdot) \\) denotes the network model, and \\( \\ell(\\cdot) \\) denotes the loss function, \\( \\ell_{CE} \\) was cross-entropy loss, which minimizes the loss during the training process. However, since some hand-crafted mixup methods are random, e.g. CutMix, ResizeMix, but the labels of the mixed samples are fixed \\( \\widetilde{Y} \\), then in the case shown in Figure.2, the features of a certain classification are occluded and reduced or do not exist in the mixed samples, resulting in the feature information not matching the labels. The phenomenon is called \"Label MisMatch\". In this case, the proportion of labels should be reduced or removed to get an accurate result when calculating the loss function." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.342, + 0.128 + ], + "angle": 0, + "content": "Qin. et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.147, + 0.554, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.564, + 0.149, + 0.782, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.287, + 0.788, + 0.345 + ], + "angle": 0, + "content": "Fig. 4: The left diagram represents the training process of SUMix. The raw and mixed samples are encoded through the model to obtain the semantic and uncertainty information, and the new mix ratio and loss function for regularization are obtained. The right diagram is the spatial distance between the raw and mixed samples." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.375, + 0.322, + 0.391 + ], + "angle": 0, + "content": "4 SUMix" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.415, + 0.788, + 0.522 + ], + "angle": 0, + "content": "In this section, we focus on detailing SUMix, 1) the uncertainty classifier, 2) how to obtain the mix ratio \\(\\lambda\\), and 3) the regularization function in the case of uncertainty and semantic information. Firstly, we show how to use deep learning classifiers for loss function computation and network model training, secondly, how to redefine the mix ratio \\(\\lambda\\) based on semantic information, and lastly, how to consider uncertainty and regularization. The pipeline of SUMix is shown in Fig. 4." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.551, + 0.442, + 0.567 + ], + "angle": 0, + "content": "4.1 Uncertainty Classifier" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.585, + 0.788, + 0.692 + ], + "angle": 0, + "content": "Assuming \\(\\mathbb{S} = \\{x_s|s = 1,2,\\dots ,S\\}\\) is a training set and \\(\\mathbb{S}\\) is the number of training images. We take \\(N\\) samples from it as a sample set, where \\(\\mathbb{X}\\in \\mathbb{R}^{N\\times 3\\times H\\times W}\\), \\(\\mathbb{Y}\\in \\mathbb{R}^{N\\times K}\\) are the labels and \\(K\\) is the number of classes. For the original computation of the mixup loss function, we use a deep neural network \\(f_{\\theta}(\\cdot)\\) with a trainable parameter that maps the sample \\(x_{i}\\) to the labels \\(y_{i}\\). Unlike the vanilla mixup loss computation in Eq.(2), SUMix adds a regularized loss function as shown in Eq.(3):" + }, + { + "type": "equation", + "bbox": [ + 0.231, + 0.706, + 0.787, + 0.749 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s u} = \\frac {1}{N} \\left(\\sum_ {i = 0} ^ {N} \\mathcal {L} _ {M C E} \\left(f _ {\\theta} \\left(\\widetilde {X} _ {i}\\right), Y, \\widetilde {\\lambda} _ {i}\\right)\\right) + \\zeta * \\mathcal {L} _ {M C E} \\left(S U \\left(f _ {\\theta} (\\widetilde {X}), U _ {\\omega}\\right), \\widetilde {Y}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.763, + 0.788, + 0.841 + ], + "angle": 0, + "content": "where \\(\\widetilde{\\lambda}_i\\) are the redefined lambda, \\(SU(\\cdot)\\) denotes a function to obtain feature, and \\(\\zeta\\) is a hyperparameter used to weight the regulated loss function, set to [0.1, 0.2, 0.5, 1.0]. \\(U_{\\omega}\\) denotes a function as a measure of uncertainty. Finally, the parameters of the deep neural network are updated continuously to minimize \\(\\mathcal{L}_{su}\\)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.682, + 0.114, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SUMix" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.147, + 0.42, + 0.162 + ], + "angle": 0, + "content": "4.2 Mix ratio Learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.325 + ], + "angle": 0, + "content": "Mixup approaches always depend on a parameter taken from the \\(Beta(\\alpha, \\alpha)\\) distribution, which we call the mix ratio \\(\\lambda\\), to balance the mix ratio of the two samples. Lots of mixup approaches use \\(\\lambda\\) to augment the mixup sample set \\(\\mathbb{X}\\) after obtaining it, and notably, this \\(\\lambda\\) serves the whole sample set. For some saliency-based or automatic-based mixups, e.g. PuzzleMix [26], AutoMix [44], it is possible to push the mixing method to obtain the best possible mixed samples according to the \\(\\lambda\\) ratio, balancing the sample ratio with the \\(\\lambda\\) as much as possible. However, some hand-crafted approaches with randomness cannot accurately match the mix ratio. The problem we want to solve is how to match each mixup sample with its own \\(\\lambda\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.326, + 0.788, + 0.412 + ], + "angle": 0, + "content": "For a mixed sample \\(\\widetilde{x} = \\left[\\widetilde{X}\\right]_i^N\\in R^{3\\times H\\times W}\\), semantic information links need to be established with the original sample pair \\((x_{a},x_{b})\\). They are fed into a deep neural network \\(f_{\\theta}(\\cdot)\\), e.g. ResNet18 [19]. feature vectors \\(\\widetilde{z}\\) and \\((z_{a},z_{b})\\) are obtained, and the feature vectors are normalized to modify their similarity according to Eq.(4):" + }, + { + "type": "equation", + "bbox": [ + 0.331, + 0.424, + 0.785, + 0.458 + ], + "angle": 0, + "content": "\\[\n\\widetilde {\\lambda} _ {a} = \\frac {\\lambda * \\mathrm {e} ^ {- \\| (\\sigma (\\widetilde {z} - z _ {a}) \\| _ {2}}}{\\lambda * \\mathrm {e} ^ {- \\| (\\sigma (\\widetilde {z} - z _ {a}) \\| _ {2}} + (1 - \\lambda) * \\mathrm {e} ^ {- \\| (\\sigma (\\widetilde {z} - z _ {b}) \\| _ {2}}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.47, + 0.788, + 0.592 + ], + "angle": 0, + "content": "where \\(\\sigma(\\cdot)\\) is the softmax function used to normalize the vectors, and \\(\\|\\cdot\\|_2\\) is the l2 norm, which imposes a limit on the distance between the two. When the difference in semantic information between \\(\\widetilde{x}\\) and \\(x_a\\) is large, which indicates that the mixed sample has a small amount of the feature information of the raw sample, \\(\\widetilde{\\lambda}_a\\) tends to 0. Similarly, when the difference in semantic information between the two is small, \\(\\widetilde{\\lambda}_a\\) tends to 1. \\(\\widetilde{\\lambda}_b = 1 - \\widetilde{\\lambda}_a\\). Thus, we have calculated the mixed sample set with its own matching \\(\\lambda\\). a more sensible loss is computed according to Eq.(3)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.616, + 0.456, + 0.631 + ], + "angle": 0, + "content": "4.3 Uncertainty Estimation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.764 + ], + "angle": 0, + "content": "Deep learning tasks involve two types of uncertainty: Epistemic uncertainty and Aleatoric uncertainty. Epistemic uncertainty arises from the model's inherent uncertainty due to insufficient training data, resulting in low confidence in unknown data. In contrast, Aleatoric uncertainty arises from errors in the data itself, which can introduce bias into the dataset. The greater the bias, the greater the Aleatoric uncertainty. In the case of the mixup task, the error in the data occurs when the mixup method is stochastic. Our goal is to address this error and enable the network to adapt accordingly." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Wang et al. [68] propose an introspective metric that models a new vector for classification based on the semantic information distance between two images and their uncertainty distance. The mixup methods differ from this as they typically incorporate feature information from two samples. To begin, we encode the mixed sample and the raw sample using an encoder to obtain the feature" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.342, + 0.129 + ], + "angle": 0, + "content": "Qin. et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.178 + ], + "angle": 0, + "content": "vectors \\(\\widetilde{z}\\) and \\(z\\), respectively. Next, we calculate the uncertainty by passing \\(\\widetilde{z}\\) and \\(z\\) through \\(U_{\\omega}\\), by Eq.(5)." + }, + { + "type": "equation", + "bbox": [ + 0.428, + 0.192, + 0.785, + 0.217 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} u = \\parallel M L P (\\sigma (z)) \\| _ {2}, \\\\ \\widetilde {\\gamma} - \\parallel M L P (\\sigma (\\widetilde {z})) \\| \\end{array} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.237, + 0.786, + 0.267 + ], + "angle": 0, + "content": "where is a linear layer with a softmax function and l2 norm. We define \\(\\beta = \\widetilde{u} + u\\) as the uncertainty of mixed samples and raw samples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.269, + 0.787, + 0.345 + ], + "angle": 0, + "content": "After capturing the uncertainty and semantic information, as demonstrated in the right part of Figure.4, Wang et al. [68] argue that in the feature space, the smaller the radius of a sample, the more aggregated the class is, and the smaller the uncertainty. By combining the semantic information, we reconstruct the mixed samples \\(\\widetilde{X}\\) and the raw samples \\(X\\) to obtain \\(Z_{su}\\) according to Eq.(6)." + }, + { + "type": "equation", + "bbox": [ + 0.425, + 0.358, + 0.785, + 0.375 + ], + "angle": 0, + "content": "\\[\nZ _ {s u} = \\mathrm {e} ^ {- (\\beta + \\| \\sigma (\\tilde {z} - z) \\| _ {2})} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.383, + 0.787, + 0.445 + ], + "angle": 0, + "content": "When there is significant uncertainty or semantic information has a large difference, \\( Z_{su} \\) receives a small gradient; conversely, when there is little uncertainty and small differences in semantic information, \\( Z_{su} \\) receives a normal gradient that tends to 1. This approach is used as a method of regularization." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.469, + 0.377, + 0.487 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.502, + 0.788, + 0.655 + ], + "angle": 0, + "content": "To verify the performance of SUMix, we conducted a large number of experiments on six datasets, CIFAR100 [28], Tiny-ImageNet [7], ImageNet-1K [29], CUB-200 [66], FGVC-Aircrafts [47], and CIFAR100-C. To verify the effectiveness of the improved \"Label MisMatching\" method, we mainly compared it with some patch-wise mixup methods based on hand-crafted, i.e., CutMix, FMix, ResizeMix, and SaliencyMix. We have implemented our algorithm on the open-source library OpenMixup [35], and some parameters are designed according to OpenMixup. The classification results are the median of the top-1 test accuracy from the last 10 training epochs, but on ImageNet-1K, we choose the last 5 training epochs. To facilitate comparison, our results are highlighted in bold." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.678, + 0.334, + 0.692 + ], + "angle": 0, + "content": "5.1 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We choose the three most classical datasets and two fine-grained datasets in the classification task. CIFAR100 has a total of 50,000 images for the training set and 10,000 images for the testing set, containing 100 different classes, each with a \\(32 \\times 32\\) resolution; Tiny-ImageNet contains 10,000 images for training and 10,000 images for testing, each with a \\(64 \\times 64\\) resolution and a total of 200 classes; The ImageNet-1K dataset contains 1.2 million images for training and 50,000 images for validation; CUB200 dataset contains images of 200 different bird species, a total of 11,788 images; the FGVC-Aircrafts dataset contains images of 100 different types of aircraft, a total of 10,000 images." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.682, + 0.114, + 0.732, + 0.128 + ], + "angle": 0, + "content": "SUMix" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.457, + 0.162 + ], + "angle": 0, + "content": "5.2 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.172, + 0.788, + 0.324 + ], + "angle": 0, + "content": "Hyper-parameter of \\(\\alpha\\). In the experimental results we report, the hyperparameter \\(\\alpha\\) of the mixup variants, which sampling from the \\(Beta(\\alpha, \\alpha)\\) distribution used by the model based on the CNN architecture and the model based on the ViT architecture are different. For CutMix, the CNN-based models, e.g. ResNet18, ResNeXt50, and Wide-ResNet28 set as 0.2, and ViT-based models, DeiT and Swin-Transformer set as 2.0. For FMix, CNN-based set 0.2, and ViT-based set 1.0. For PuzzleMix, the CNN-based set is 1.0, and the ViT-based set is 2.0. In CNN-based and Vit-based architectures alike, SaliencyMix, ResizeMix, and AutoMix exhibit identical alpha values, which are 0.2, 1.0, and 2.0, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.337, + 0.788, + 0.43 + ], + "angle": 0, + "content": "Hyper-parameter of \\(\\zeta\\). In SUMix, we set the hyper-parameters \\(\\zeta\\) of \\(\\mathcal{L}_{su}\\) as follows: On the CIFAR100 dataset, we report five different models, the \\(\\zeta\\) setting shown in Appendix A.1. On Tiny-ImageNet, we set the \\(\\zeta = 1.0\\) for all mixup methods based on ResNet18 and ResNeXt50. Similarly with ImageNet-1K, we set the \\(\\zeta = 0.5\\) based on ResNet18. The settings we reported for the fine-grained datasets are also shown in Appendix A.1." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.451, + 0.429, + 0.466 + ], + "angle": 0, + "content": "5.3 Image Classification" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.475, + 0.789, + 0.628 + ], + "angle": 0, + "content": "Generic classification. We initially train ResNet18 [19] and ResNeXt50 [69] using the CIFAR100 dataset for a total of 800 epochs. The experimental parameters are set as in OpenMixup [35]: a basic learning rate of 0.1, SGD [45] optimizer with momentum of 0.9, weight decay of 0.0001, and batch size of 100 for each iteration. Wide-ResNet [73] using the CIFAR100 dataset for a total of 400 epochs, learning rate of 0.03, SGD optimizer with momentum of 0.9, weight decay of 0.001. For the Tiny-ImageNet dataset, we use the same experimental parameters as the CIFAR100 experiment, except the base learning rate to 0.2 and training for 400 epochs. As for ImageNet-1K, we follow the Pytorch-style settings for training ResNet18 of 100 epochs. For the CIFAR100 dataset, we use" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.653, + 0.788, + 0.682 + ], + "angle": 0, + "content": "Table 1: Top-1 accuracy \\((\\%)\\uparrow\\) of mixup methods on CIFAR-100, Tiny-ImageNet and ImageNet-1K. * denotes mixup methods with SUMix." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.694, + 0.784, + 0.837 + ], + "angle": 0, + "content": "
MethodCIFAR100Tiny-ImageNetImageNet-1K
ResNet18ResNeXt50W-ResNet28-8ResNet18ResNeXt50ResNet18
CutMix78.1778.3284.4565.5366.4768.95
FMix79.6979.0284.2163.4765.0869.96
SaliencyMix79.1278.7784.3564.6066.5569.16
ResizeMix80.0180.3584.8763.7465.8769.50
CutMix*79.7879.9184.5665.7168.7469.71
FMix*80.2080.7984.3263.6967.1270.48
SaliencyMix*79.9179.3284.5865.6868.9269.52
ResizeMix*80.3880.7284.9165.3067.4969.76
Avg. Gain+0.82+1.07+0.12+0.81+2.07+0.47
" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.341, + 0.129 + ], + "angle": 0, + "content": "Qin. et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.144, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 2: Top-1 accuracy \\((\\%)\\uparrow\\) of mixup methods on CUB200, FGVC-Aircrafts, and Standford-Cars. * denotes mixup methods SUMix. † denotes that we reproduce the results of that experiment." + }, + { + "type": "table", + "bbox": [ + 0.288, + 0.199, + 0.709, + 0.361 + ], + "angle": 0, + "content": "
MethodCUB200FGVC-Aircrafts
ResNet18ResNeXt50ResNet18ResNeXt50
CutMix†77.7083.6778.8484.55
FMix77.2884.0679.3684.10
SaliencyMix†75.7782.8379.7884.31
ResizeMix78.5084.1678.1084.08
CutMix*78.2083.7179.7285.84
FMix*79.2484.3379.4884.64
SaliencyMix*76.9882.8479.9084.49
ResizeMix*78.5684.2380.2985.12
Avg. Gain+0.93+0.10+0.82+0.67
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.389, + 0.788, + 0.433 + ], + "angle": 0, + "content": "Table 3: Top-1 accuracy(\\%)↑ of Table 4: Top-1 accuracy(\\%)↑ of saliency-based mixup methods on CIFAR100. * de- mixup methods on CIFAR100. * denotes mixup notes mixup methods with SUMix. methods with SUMix." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.445, + 0.457, + 0.549 + ], + "angle": 0, + "content": "
MethodCIFAR100
DeiT-SmallSwin-Tiny
CutMix74.1280.64
FMix70.4180.72
SaliencyMix69.7880.40
ResizeMix68.4580.16
CutMix*75.26+1.1480.83+0.19
FMix*70.69+0.2880.73+0.01
SaliencyMix*70.31+0.5380.71+0.29
ResizeMix*68.78+0.3380.59+0.43
" + }, + { + "type": "table", + "bbox": [ + 0.464, + 0.444, + 0.782, + 0.548 + ], + "angle": 0, + "content": "
MethodCIFAR100
ResNet18ResNeXt50
PuzzleMix81.1381.69
AutoMix82.0482.84
PuzzleMix*81.43+0.3082.60+0.91
AutoMix*82.30+0.2683.82+0.98
AdAutoMix82.3283.81
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.579, + 0.787, + 0.625 + ], + "angle": 0, + "content": "basic data augmentation including RandomFlip and RandomCrop with 4 pixels padding, and for the Tiny-ImageNet and ImageNet-1K datasets, we also use the RandomFlip and RandonCrop data augmentation operations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.626, + 0.788, + 0.748 + ], + "angle": 0, + "content": "Table 1 shows our five comparisons on three different datasets, where the use of SUMix brings some improvement in the classification accuracy of CutMix, FMix, SaliencyMix, and ResizeMix. For the CIFAR100 dataset, on ResNet and its variants ResNeXt50, W-ResNet28-8, SUMix improves their classification results by an average of \\(0.82\\%\\), \\(1.07\\%\\), and \\(0.12\\%\\), respectively; for the Tiny-ImageNet dataset, on ResNet18 and ResNeXt50, the average improvements of \\(0.81\\%\\) and \\(2.07\\%\\); for the ImageNet-1K dataset, on ResNet18, the average improvement is \\(0.47\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Fine-grained classification. We load the official PyTorch pre-trained models on ImageNet-1k as initialization and fine-tuning pre-trained ResNet18 and ResNeXt50 using SGD optimizer with a learning rate of 0.001, a momentum of 0.9, weight decay of 0.0005 on CUB200, FGVC-Aircrafts, total training 200 epochs. Table 2 presents the comparison results for the bird and aircraft datasets." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.683, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SUMix" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.147, + 0.778, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.314, + 0.785, + 0.384 + ], + "angle": 0, + "content": "Fig. 5: The top of the figure shows a visualization of the sample at \\(0\\%\\) to \\(100\\%\\) occlusion ratio. The light blue curves in the lower four subfigures show the classification accuracy of CutMix, FMix, SaliencyMix, and ResizeMix on ImageNet-1K using ResNet18 for 100 epochs of training, respectively, and the red curve shows the classification accuracy of these methods using SUMix." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.42, + 0.785, + 0.48 + ], + "angle": 0, + "content": "SUMix improves the classification results for the CUB200 dataset by an average of \\(0.93\\%\\) and \\(0.10\\%\\) at ResNet18 and ResNeXt50, respectively. The FGVC-Aircrafts dataset improves by an average of \\(0.82\\%\\) and \\(0.67\\%\\) at ResNet18 and ResNeXt50." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.785, + 0.658 + ], + "angle": 0, + "content": "Classification based on ViTs. To further show the performance of SUMix, we perform classification experiments using the ViT-based model DeiT [60] and Swin-Transformer [41] on the CIFAR100 dataset, and the results are shown in Tab 3. We resized the size to \\(224 \\times 224\\), and in the training set, we trained DeiT and Swin-Transformer of 200 epochs by AdamW [46] optimizer with a batch size of 100 on CIFAR100. The basic learning rates of DeiT and Swin are 1e-3 and 5e-4, and the weight decay of 0.05, like OpenMixup settings, we also used some RandAugment [9] operations. We reported the best top-1 accuracy in the last 10 training epochs, and the results show that our SUMix also can bring effectiveness to ViTs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.661, + 0.785, + 0.707 + ], + "angle": 0, + "content": "Table 3 shows the results of the ViT models using SUMix, which improves the classification results by an average of \\(0.57\\%\\) and \\(0.23\\%\\) at DeiT-Small and Swin-Tiny." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Classification based on Saliency methods. To further explore the capability of SUMix, we not only apply it to some Hand-crafted methods, we apply SUMix to PuzzleMix and AutoMix, which are based on saliency-learning methods, and Tab 4 shows that SUMix will also bring some boosts, PuzzleMix and AutoMix has gained \\(0.30\\%\\), \\(0.26\\%\\) at ResNet18, and \\(0.91\\%\\), \\(0.98\\%\\) at ResNeXt50. Those methods combined with SUMix can even reach almost close to the performance of the current state-of-the-art method AdAutoMix." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.341, + 0.128 + ], + "angle": 0, + "content": "Qin. et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.174, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 5: Top-1 acc(%)↑ and FGSM error(%)↓ Table 6: Ablation experiments on CI-of ResNet18 without and with SUMix. FAR100, CUB200 based on ResNet18." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.215, + 0.524, + 0.29 + ], + "angle": 0, + "content": "
Clean Acc(%)↑Corruption Acc(%)↑FGSM Error(%)↓
MethodMCESUMixMCESUMixMCESUMix
CutMix78.1779.7843.0644.3191.1590.41
FMix79.6980.2048.7949.1489.1689.08
SaliencyMix79.1279.9143.7344.3689.6491.49
ResizeMix80.0180.3846.1246.2890.0491.05
" + }, + { + "type": "table", + "bbox": [ + 0.528, + 0.215, + 0.787, + 0.29 + ], + "angle": 0, + "content": "
MethodCIFAR100CUB200
CutMixSaliencyMixFMixSaliencyMix
baseline78.1779.1277.2875.77
+λ79.1379.7979.0376.46
+ℓs79.5178.7179.0376.34
+ℓu79.6079.4379.2476.72
+ℓsu79.7879.9179.2476.98
" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.317, + 0.442, + 0.331 + ], + "angle": 0, + "content": "5.4 Occlusion Robustness" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.339, + 0.788, + 0.476 + ], + "angle": 0, + "content": "To analyze the enhancement brought by SUMix to the robustness of the mixup methods, we randomly conducted masking patches experiment [48]. The experiment involved dividing the image into \\(16 \\times 16\\) patches and masking them according to different ratios. The resulting image was then classified by the model. We used the ImageNet dataset, which contains 1000 daily life objects. It is more representative. The ResNet18 classifier was used to classify the masked images and verify the model's robustness capability. The curves at the bottom of Figure.5 show that SUMix provides a performance gain compared to the corresponding mixup method at different occlusion ratios." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.496, + 0.357, + 0.51 + ], + "angle": 0, + "content": "5.5 Robustness" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.519, + 0.788, + 0.61 + ], + "angle": 0, + "content": "We first evaluated robustness against corruption on CIFAR-100-C [20], which was designed manually to provide 19 different corruptions (e.g., noise, blur, fog, etc.) to evaluate robustness. In Table 5, we show the results of improving some mixup methods with SUMix. We further investigate the FGSM [14] white box attack experiments with \\(8/255\\ell_{\\infty}\\) epsilon ball and similarly show the results of the improvement of our approach on mixup methods in Table 5." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.63, + 0.391, + 0.645 + ], + "angle": 0, + "content": "5.6 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.652, + 0.788, + 0.742 + ], + "angle": 0, + "content": "In the ablation study, we primarily analyze the effect of the two SUMix modules in the model. The following two analyses are primarily done. (1) Can redefining \\(\\lambda\\) bring gains in models? (2) Can uncertainty lead to regularization gains? Table 6 shows the experimental results for CutMix, SaliencyMix, and FMix using ResNet18 and ResNeXt50 on the CIFAR100 and CUB200 datasets." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.748, + 0.788, + 0.84 + ], + "angle": 0, + "content": "(1) Table 6 shows that with our redefined \\(\\widetilde{\\lambda}\\), CutMix and SaliencyMix are improved by \\(0.96\\%\\) and \\(0.67\\%\\) on the CIFAR100 dataset, and FMix and SaliencyMix are improved by \\(1.75\\%\\) and \\(0.69\\%\\) on the CUB200 dataset, which means that the module can bring a significant increase in the classification ability, in other words, the redefined \\(\\widetilde{\\lambda}\\) can obtain a more sensible loss function for the model to train." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.683, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SUMix" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.147, + 0.788, + 0.298 + ], + "angle": 0, + "content": "(2) To calculate the regularization loss, we consider three cases: 1) using only the semantic information \\(\\ell_s\\), 2) using only the uncertainty information \\(\\ell_u\\), and 3) using both the semantic and uncertainty information \\(\\ell_{su}\\). Table 6 shows that on the CUB200 dataset, FMix and SaliencyMix improve accuracy by \\(1.75\\%\\) and \\(0.57\\%\\) respectively when using only \\(\\ell_s\\), and by \\(1.96\\%\\) and \\(0.95\\%\\) respectively when using only \\(\\ell_u\\). When both types of information are used, accuracy improves by \\(1.75\\%\\) and \\(1.21\\%\\). Similarly, on the CIFAR100 dataset, there was a \\(1.61\\%\\) and \\(0.79\\%\\) improvement. The experimental results show that combining semantic and uncertainty information can significantly improve accuracy." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.322, + 0.519, + 0.339 + ], + "angle": 0, + "content": "6 Conclusion and Limitations" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.354, + 0.788, + 0.46 + ], + "angle": 0, + "content": "Conclusion. In this paper, we propose the SUMix approach to regularize the DNN in terms of the semantic information and uncertainty between the raw and mixed samples. Specifically, the semantic information difference between the raw and mixed samples is used to obtain the mix ratio \\(\\lambda\\), while the uncertainty is, in turn, combined with the semantic information to obtain a regularized loss function. Experimental results on five datasets proved the effectiveness of SUMix for the regularization of the mixup methods." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.475, + 0.788, + 0.567 + ], + "angle": 0, + "content": "Limitations. Currently, SUMix is due to learn how to obtain a sensible mixing ratio \\(\\lambda\\) and an uncertainty-based feature vector. These are currently proven only on classification tasks. In future work, we would like to apply and improve SUMix accordingly on some downstream tasks e.g. segmentation, detection, and other methods. And for semi-supervised tasks, we can get more accurate pseudo-labels for training models." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.59, + 0.395, + 0.607 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.622, + 0.788, + 0.743 + ], + "angle": 0, + "content": "This work is supported in part by the Scientific Innovation 2030 Major Project for New Generation of AI under Grant 2020AAA0107300, in part by the National Natural Science Foundation of China (Grant No. 61976030), in part by Chongqing Natural Science Foundation (Grant No. 2024NSCQ-MSX4893), in part by Project of CQ CSTC (Grant No. cstc2018jcyjAX 0057), in part by Innovative Research Projects for Postgraduate Students of Chongqing Business University (No. yjscxx2024-284-53, No. yiscxx2024-284-231, No. yjscxx2024-284-234)." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.767, + 0.325, + 0.782 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.797, + 0.784, + 0.825 + ], + "angle": 0, + "content": "1. Baek, K., Bang, D., Shim, H.: Gridmix: Strong regularization through local context mapping. Pattern Recognition 109 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.827, + 0.741, + 0.84 + ], + "angle": 0, + "content": "2. Bishop, C.M.: Pattern recognition and machine learning. Springer (2006) 2" + }, + { + "type": "list", + "bbox": [ + 0.223, + 0.797, + 0.784, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.341, + 0.128 + ], + "angle": 0, + "content": "Qin. et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "3. Bochkovskiy, A., Wang, C.Y., Liao, H.Y.M.: Yolov4: Optimal speed and accuracy of object detection. ArXiv abs/2004.10934 (2020) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.786, + 0.218 + ], + "angle": 0, + "content": "4. Chang, J., Lan, Z., Cheng, C., Wei, Y.: Data uncertainty learning in face recognition. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5710-5719 (2020) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.219, + 0.786, + 0.26 + ], + "angle": 0, + "content": "5. Chen, J.N., Sun, S., He, J., Torr, P.H., Yuille, A., Bai, S.: Transmix: Attend to mix for vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12135-12144 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.261, + 0.786, + 0.302 + ], + "angle": 0, + "content": "6. Chen, M., Lin, M., Lin, Z., Zhang, Y., Chao, F., Ji, R.: Smmix: Self-motivated image mixing for vision transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17260-17270 (2023) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.304, + 0.786, + 0.332 + ], + "angle": 0, + "content": "7. Chrabaszcz, P., Loshchilov, I., Hutter, F.: A downsampled variant of imagenet as an alternative to the cifar datasets. arXiv preprint arXiv:1707.08819 (2017) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.333, + 0.786, + 0.36 + ], + "angle": 0, + "content": "8. Cubuk, E.D., Zoph, B., Mane, D., Vasudevan, V., Le, Q.V.: Autoaugment: Learning augmentation policies from data. arXiv preprint arXiv:1805.09501 (2018) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.361, + 0.786, + 0.415 + ], + "angle": 0, + "content": "9. Cubuk, E.D., Zoph, B., Shlens, J., Le, Q.V.: Randaugment: Practical automated data augmentation with a reduced search space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 702-703 (2020) 2, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.417, + 0.786, + 0.458 + ], + "angle": 0, + "content": "10. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.459, + 0.786, + 0.486 + ], + "angle": 0, + "content": "11. DeVries, T., Taylor, G.W.: Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552 (2017) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.488, + 0.786, + 0.515 + ], + "angle": 0, + "content": "12. Feichtenhofer, C., Fan, H., Malik, J., He, K.: Slowfast networks for video recognition. In: ICCV. pp. 6201-6210 (2019) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.516, + 0.786, + 0.557 + ], + "angle": 0, + "content": "13. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. Communications of the ACM 63(11), 139-144 (2020) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.558, + 0.786, + 0.585 + ], + "angle": 0, + "content": "14. Goodfellow, I.J., Shlens, J., Szegedy, C.: Explaining and harnessing adversarial examples. In: ICLR (2015) 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.587, + 0.786, + 0.628 + ], + "angle": 0, + "content": "15. Guénais, T., Vamvourellis, D., Yacoby, Y., Doshi-Velez, F., Pan, W.: Bacoun: Bayesian classifiers with out-of-distribution uncertainty. arXiv preprint arXiv:2007.06096 (2020) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.629, + 0.786, + 0.656 + ], + "angle": 0, + "content": "16. Harris, E., Marcu, A., Painter, M., Niranjan, M., Hare, A.P.B.J.: Fmix: Enhancing mixed sample data augmentation. arXiv preprint arXiv:2002.12047 (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.657, + 0.786, + 0.699 + ], + "angle": 0, + "content": "17. Hataya, R., Zdenek, J., Yoshizoe, K., Nakayama, H.: Faster autoaugment: Learning augmentation strategies using backpropagation. In: European Conference on Computer Vision. pp. 1-16. Springer (2020) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.7, + 0.786, + 0.728 + ], + "angle": 0, + "content": "18. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: CVPR. pp. 9729-9738 (2020) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.729, + 0.786, + 0.77 + ], + "angle": 0, + "content": "19. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 770-778 (2016) 7, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.771, + 0.786, + 0.798 + ], + "angle": 0, + "content": "20. Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261 (2019) 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.799, + 0.786, + 0.84 + ], + "angle": 0, + "content": "21. Hendrycks, D., Mu, N., Cubuk, E.D., Zoph, B., Gilmer, J., Lakshminarayanan, B.: Augmix: A simple data processing method to improve robustness and uncertainty. arXiv preprint arXiv:1912.02781 (2019) 4" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.147, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.683, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SUMix" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.786, + 0.204 + ], + "angle": 0, + "content": "22. Hendrycks, D., Zou, A., Mazeika, M., Tang, L., Li, B., Song, D., Steinhardt, J.: Pixmix: Dreamlike pictures comprehensively improve safety measures. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.205, + 0.786, + 0.231 + ], + "angle": 0, + "content": "23. Huang, S., Wang, X., Tao, D.: Snapmix: Semantically proportional mixing for augmenting fine-grained data. In: AAAI (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.232, + 0.787, + 0.272 + ], + "angle": 0, + "content": "24. Jiang, Z., Hou, Q., Yuan, L., Zhou, D., Jin, X., Wang, A., Feng, J.: Token labeling: Training a \\(85.4\\%\\) top-1 accuracy vision transformer with 56m parameters onImagenet. ArXiv abs/2104.10858 (2021) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.786, + 0.3 + ], + "angle": 0, + "content": "25. Kendall, A., Gal, Y.: What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems 30 (2017) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.301, + 0.786, + 0.342 + ], + "angle": 0, + "content": "26. Kim, J.H., Choo, W., Song, H.O.: Puzzle mix: Exploiting saliency and local statistics for optimal mixup. In: International Conference on Machine Learning. pp. 5275-5285. PMLR (2020) 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.786, + 0.383 + ], + "angle": 0, + "content": "27. Kim, J., Choo, W., Jeong, H., Song, H.O.: Co-mixup: Saliency guided joint mixup with supermodular diversity. In: International Conference on Learning Representations (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.786, + 0.411 + ], + "angle": 0, + "content": "28. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.786, + 0.452 + ], + "angle": 0, + "content": "29. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. In: Advances in neural information processing systems. pp. 1097-1105 (2012) 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.786, + 0.508 + ], + "angle": 0, + "content": "30. Lee, J.H., Zaheer, M.Z., Astrid, M., Lee, S.I.: Smoothmix: a simple yet effective data augmentation to train robust classifiers. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 756-757 (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.508, + 0.786, + 0.55 + ], + "angle": 0, + "content": "31. Li, S., Jin, W., Wang, Z., Wu, F., Liu, Z., Tan, C., Li, S.Z.: Semireward: A general reward model for semi-supervised learning. In: International Conference on Learning Representations (2024) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.786, + 0.591 + ], + "angle": 0, + "content": "32. Li, S., Liu, Z., Wang, Z., Wu, D., Liu, Z., Li, S.Z.: Boosting discriminative visual representation learning with scenario-agnostic mixup. ArXiv abs/2111.15454 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.591, + 0.786, + 0.633 + ], + "angle": 0, + "content": "33. Li, S., Liu, Z., Zang, Z., Wu, D., Chen, Z., Li, S.Z.: Genurl: A general framework for unsupervised representation learning. IEEE Transactions on Neural Networks and Learning Systems (2024) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.786, + 0.673 + ], + "angle": 0, + "content": "34. Li, S., Wang, Z., Liu, Z., Tan, C., Lin, H., Wu, D., Chen, Z., Zheng, J., Li, S.Z.: Efficient multi-order gated aggregation network. ArXiv abs/2211.03295 (2022) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.674, + 0.786, + 0.715 + ], + "angle": 0, + "content": "35. Li, S., Wang, Z., Liu, Z., Wu, D., Li, S.Z.: Openmixup: Open mixup toolbox and benchmark for visual representation learning. https://github.com/Westlake-AI/openmixup (2022) 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.786, + 0.757 + ], + "angle": 0, + "content": "36. Li, S., Wu, D., Wu, F., Zang, Z., Li, S.Z.: Architecture-agnostic masked image modeling – from vit back to cnn. In: International Conference on Machine Learning (2023) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.786, + 0.797 + ], + "angle": 0, + "content": "37. Li, S., Zhang, Z., Liu, Z., Wang, A., Qiu, L., Du, F.: Tlpg-tracker: Joint learning of target localization and proposal generation for visual tracking. In: IJCAI (2020) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.798, + 0.786, + 0.84 + ], + "angle": 0, + "content": "38. Li, Y., Hu, G., Wang, Y., Hospedales, T., Robertson, N.M., Yang, Y.: Differentiable automatic data augmentation. In: European Conference on Computer Vision. pp. 580-595. Springer (2020) 2" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.341, + 0.128 + ], + "angle": 0, + "content": "Qin. et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "39. Lim, S., Kim, I., Kim, T., Kim, C., Kim, S.: Fast autoaugment. In: Advances in Neural Information Processing Systems (NeurIPS) (2019) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.785, + 0.215 + ], + "angle": 0, + "content": "40. Liu, J., Liu, B., Zhou, H., Li, H., Liu, Y.: Tokenmix: Rethinking image mixing for data augmentation in vision transformers. In: European Conference on Computer Vision (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.216, + 0.785, + 0.257 + ], + "angle": 0, + "content": "41. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: International Conference on Computer Vision (ICCV) (2021) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.257, + 0.785, + 0.298 + ], + "angle": 0, + "content": "42. Liu, Z., Li, S., Wang, G., Wu, L., Tan, C., Li, S.Z.: Harnessing hard mixed samples with decoupled regularizer. In: Thirty-seventh Conference on Neural Information Processing Systems (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.298, + 0.785, + 0.338 + ], + "angle": 0, + "content": "43. Liu, Z., Li, S., Wang, L., Wang, Z., Liu, Y., Li, S.Z.: Short-long convolutions help hardware-efficient linear attention to focus on long sequences. In: Forty-first International Conference on Machine Learning (2024) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.339, + 0.785, + 0.379 + ], + "angle": 0, + "content": "44. Liu, Z., Li, S., Wu, D., Liu, Z., Chen, Z., Wu, L., Li, S.Z.: Automix: Unveiling the power of mixup for stronger classifiers. In: European Conference on Computer Vision. pp. 441-458. Springer (2022) 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.785, + 0.407 + ], + "angle": 0, + "content": "45. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983 (2016) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.407, + 0.785, + 0.431 + ], + "angle": 0, + "content": "46. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: ICLR (2019) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.433, + 0.785, + 0.459 + ], + "angle": 0, + "content": "47. Maji, S., Rahtu, E., Kannala, J., Blaschko, M., Vedaldi, A.: Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151 (2013) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.46, + 0.785, + 0.5 + ], + "angle": 0, + "content": "48. Naseer, M., Ranasinghe, K., Khan, S.H., Hayat, M., Khan, F.S., Yang, M.H.: Intriguing properties of vision transformers. In: Neural Information Processing Systems (2021) 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.785, + 0.54 + ], + "angle": 0, + "content": "49. Oh, S.J., Murphy, K., Pan, J., Roth, J., Schroff, F., Gallagher, A.: Modeling uncertainty with hedged instance embedding. arXiv preprint arXiv:1810.00319 (2018) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.541, + 0.785, + 0.582 + ], + "angle": 0, + "content": "50. Park, J., Yang, J.Y., Shin, J., Hwang, S.J., Yang, E.: Saliency grafting: Innocuous attribution-guided mixup with calibrated label mixing. In: AAAI Conference on Artificial Intelligence (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.583, + 0.785, + 0.609 + ], + "angle": 0, + "content": "51. Qin, H., Jin, X., Jiang, Y., El-Yacoubi, M.A., Gao, X.: Adversarial automixup. arXiv preprint arXiv:2312.11954 (2023) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.61, + 0.785, + 0.65 + ], + "angle": 0, + "content": "52. Qin, J., Fang, J., Zhang, Q., Liu, W., Wang, X., Wang, X.: Resizemix: Mixing data with preserved object information and true labels. arXiv preprint arXiv:2012.11101 (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.651, + 0.785, + 0.677 + ], + "angle": 0, + "content": "53. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. arXiv preprint arXiv:1506.01497 (2015) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.678, + 0.785, + 0.718 + ], + "angle": 0, + "content": "54. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Gradcam: Visual explanations from deep networks via gradient-based localization. arXiv preprint arXiv:1610.02391 (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.719, + 0.785, + 0.744 + ], + "angle": 0, + "content": "55. Shaw, G., Manolakis, D.: Signal processing for hyperspectral image exploitation. IEEE Signal processing magazine 19(1), 12-16 (2002) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.745, + 0.785, + 0.77 + ], + "angle": 0, + "content": "56. Shi, Y., Jain, A.K.: Probabilistic face embeddings. In: ICCV. pp. 6902-6911 (2019) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.799 + ], + "angle": 0, + "content": "57. Sun, S., Chen, J.N., He, R., Yuille, A., Torr, P., Bai, S.: Lumix: Improving mixup by better modelling label uncertainty. arXiv preprint arXiv:2211.15846 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "58. Suzuki, T.: Teachaugment: Data augmentation optimization using teacher knowledge. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10904-10914 (June 2022) 2" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.683, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SUMix" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.203 + ], + "angle": 0, + "content": "59. Tan, C., Gao, Z., Wu, L., Xu, Y., Xia, J., Li, S., Li, S.Z.: Temporal attention unit: Towards efficient spatiotemporal predictive learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18770-18782 (2023) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.203, + 0.787, + 0.243 + ], + "angle": 0, + "content": "60. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jegou, H.: Training data-efficient image transformers & distillation through attention. In: International Conference on Machine Learning (ICML). pp. 10347-10357 (2021) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.243, + 0.787, + 0.284 + ], + "angle": 0, + "content": "61. Uddin, A.S., Monira, M.S., Shin, W., Chung, T., Bae, S.H.: Saliencymix: A saliency guided data augmentation strategy for better regularization. In: International Conference on Learning Representations (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.284, + 0.787, + 0.31 + ], + "angle": 0, + "content": "62. Vaswani, A., Shazeer, N.M., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NIPS (2017) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.31, + 0.787, + 0.351 + ], + "angle": 0, + "content": "63. Venkataramanan, S., Avrithis, Y., Kijak, E., Amsaleg, L.: Alignmix: Improving representation by interpolating aligned features. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.351, + 0.787, + 0.392 + ], + "angle": 0, + "content": "64. Verma, V., Lamb, A., Beckham, C., Najafi, A., Mitliagkas, I., Lopez-Paz, D., Bengio, Y.: Manifold mixup: Better representations by interpolating hidden states. In: International Conference on Machine Learning. pp. 6438-6447 (2019) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.392, + 0.787, + 0.432 + ], + "angle": 0, + "content": "65. Vinyals, O., Toshev, A., Bengio, S., Erhan, D.: Show and tell: A neural image caption generator. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3156-3164 (2015) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.432, + 0.787, + 0.458 + ], + "angle": 0, + "content": "66. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The caltech-ucsd birds-200-2011 dataset. California Institute of Technology (2011) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.458, + 0.787, + 0.512 + ], + "angle": 0, + "content": "67. Walawalkar, D., Shen, Z., Liu, Z., Savvides, M.: Attentive cutmix: An enhanced data augmentation approach for deep learning based image classification. In: IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 3642-3646 (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.512, + 0.787, + 0.539 + ], + "angle": 0, + "content": "68. Wang, C., Zheng, W., Zhu, Z., Zhou, J., Lu, J.: Introspective deep metric learning. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023) 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.539, + 0.787, + 0.579 + ], + "angle": 0, + "content": "69. Xie, S., Girshick, R., Dollár, P., Tu, Z., He, K.: Aggregated residual transformations for deep neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1492-1500 (2017) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.579, + 0.787, + 0.618 + ], + "angle": 0, + "content": "70. Yang, L., Li, X., Zhao, B., Song, R., Yang, J.: Recursiveemix: Mixed learning with history. Advances in Neural Information Processing Systems 35, 8427-8440 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.618, + 0.787, + 0.659 + ], + "angle": 0, + "content": "71. Yang, S., Wu, Y., Jin, X., El Yacoubi, M., Qin, H.: Cgan-da: A cross-modality domain adaptation model for hand-vein biometric-based authentication. JOURNAL OF Cyber-Physical-Social Intelligence 1, 3–12 (2022) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.659, + 0.787, + 0.7 + ], + "angle": 0, + "content": "72. Yun, S., Han, D., Oh, S.J., Chun, S., Choe, J., Yoo, Y.: Cutmix: Regularization strategy to train strong classifiers with localizable features. In: International Conference on Computer Vision (ICCV). pp. 6023-6032 (2019) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.7, + 0.787, + 0.726 + ], + "angle": 0, + "content": "73. Zagoruyko, S., Komodakis, N.: Wide residual networks. In: Proceedings of the British Machine Vision Conference (BMVC) (2016) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.726, + 0.787, + 0.767 + ], + "angle": 0, + "content": "74. Zhang, B., Wonka, P.: Point cloud instance segmentation using probabilistic embeddings. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8883-8892 (2021) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.767, + 0.787, + 0.793 + ], + "angle": 0, + "content": "75. Zhang, H., Cisse, M., Dauphin, Y.N., Lopez-Paz, D.: mixup: Beyond empirical risk minimization. In: International Conference on Learning Representations (2018) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.793, + 0.787, + 0.834 + ], + "angle": 0, + "content": "76. Zhao, Q., Huang, Y., Hu, W., Zhang, F., Liu, J.: Mixpro: Data augmentation with maskmix and progressive attention labeling for vision transformer. In: The Eleventh International Conference on Learning Representations (2023) 5" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.834 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.341, + 0.128 + ], + "angle": 0, + "content": "Qin. et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.339, + 0.163 + ], + "angle": 0, + "content": "Appendix A" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.173, + 0.473, + 0.188 + ], + "angle": 0, + "content": "A.1 More Information about \\(\\zeta\\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.788, + 0.237 + ], + "angle": 0, + "content": "We reported the hyper-parameter \\(\\zeta\\) of \\(\\mathcal{L}_{su}\\) on the CIFAR100 dataset shown in Table 7. And for the fine-grained dataset, CUB200 and FGVC-Aircraftd. We reported the settings in Table 8 and Table 9." + }, + { + "type": "table_caption", + "bbox": [ + 0.332, + 0.239, + 0.669, + 0.253 + ], + "angle": 0, + "content": "Table 7: The \\( \\zeta \\) setting on the CIFAR100 dataset." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.266, + 0.782, + 0.362 + ], + "angle": 0, + "content": "
ζ settingCIFAR100
ResNet18ResNeXt50W-ResNet28-8DeiT-SmallSwin-Tiny
CutMix0.50.50.50.50.5
FMix1.01.00.50.50.5
SaliencyMix0.51.00.50.50.5
ResizeMix0.20.50.50.50.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.367, + 0.789, + 0.394 + ], + "angle": 0, + "content": "Table 8: The \\(\\zeta\\) setting on the CUB200 Table 9: The \\(\\zeta\\) setting on the FGVC-dataset." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.409, + 0.784, + 0.507 + ], + "angle": 0, + "content": "
ζ settingCUB200ζ settingFGVC-Aircrafts
ResNet18ResNeXt50ResNet18ResNeXt50
CutMix1.00.5CutMix1.00.5
FMix0.50.1FMix0.50.5
SaliencyMix0.50.2SaliencyMix1.00.5
ResizeMix0.50.5ResizeMix0.50.5
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.555, + 0.517, + 0.57 + ], + "angle": 0, + "content": "A.2 More Visualization Comparison" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.576, + 0.787, + 0.606 + ], + "angle": 0, + "content": "In Figure.6 and Figure.7, we show the visualization obtained by the mixup method using the CAM method with and without SUMix." + }, + { + "type": "image", + "bbox": [ + 0.262, + 0.634, + 0.74, + 0.793 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.808, + 0.788, + 0.838 + ], + "angle": 0, + "content": "Fig.6: CAM visualization of the ResNet18 with and without SUMix trained on the ImageNet-1K dataset." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.683, + 0.115, + 0.733, + 0.127 + ], + "angle": 0, + "content": "SUMix" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "image", + "bbox": [ + 0.241, + 0.226, + 0.756, + 0.715 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.731, + 0.788, + 0.761 + ], + "angle": 0, + "content": "Fig. 7: CAM visualization of the ResNet18 with and without SUMix trained on the ImageNet-1K dataset." + } + ] +] \ No newline at end of file diff --git a/2024/SUMix_ Mixup with Semantic and Uncertain Information/f0041215-e301-48ad-9db8-48c026069726_origin.pdf b/2024/SUMix_ Mixup with Semantic and Uncertain Information/f0041215-e301-48ad-9db8-48c026069726_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2bf329a747b018995de5c8e83a96159781f1e3a9 --- /dev/null +++ b/2024/SUMix_ Mixup with Semantic and Uncertain Information/f0041215-e301-48ad-9db8-48c026069726_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf6e545ebf2d3b49e2ccc63a805a9c3cfc4dbf489cc10a3a3f07d9b316f983fa +size 1381247 diff --git a/2024/SUMix_ Mixup with Semantic and Uncertain Information/full.md b/2024/SUMix_ Mixup with Semantic and Uncertain Information/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a575bc4583a8aab049c137810571118322fb4603 --- /dev/null +++ b/2024/SUMix_ Mixup with Semantic and Uncertain Information/full.md @@ -0,0 +1,319 @@ +# SUMix: Mixup with Semantic and Uncertain Information + +Huafeng Qin $^{1,2,5,\star,\boxtimes}$ , Xin Jin $^{1,2,\star}$ , Hongyu Zhu $^{1,2,\star}$ , Hongchao Liao $^{1,2}$ , Mounim A. El-Yacoubi $^{3}$ , Xinbo Gao $^{4}$ + +1 Chongqing Technology and Business University, Chongqing, 400067, China + +$^{2}$ National Research base of Intelligent Manufacturing Service, Chongqing, China + +3 Telecom SudParis, Institut Polytechnique de Paris, Paris, 75020, France + +4 Chongqing University of Posts and Telecommunications, Chongqing, 400065, China + +5 Chongqing Micro-Vein Intelligent Technology Co. + +{2013009, jinxin4, zhuhongyu}@ctbu.edu.cn + +$\star$ Equal contribution, $\boxtimes$ Corresponding author. + +Abstract. Mixup data augmentation approaches have been applied for various tasks of deep learning to improve the generalization ability of deep neural networks. Some existing approaches CutMix, SaliencyMix, etc. randomly replace a patch in one image with patches from another to generate the mixed image. Similarly, the corresponding labels are linearly combined by a fixed ratio $\lambda$ by $l$ . The objects in two images may be overlapped during the mixing process, so some semantic information is corrupted in the mixed samples. In this case, the mixed image does not match the mixed label information. Besides, such a label may mislead the deep learning model training, which results in poor performance. To solve this problem, we proposed a novel approach named SUMix to learn the mixing ratio as well as the uncertainty for the mixed samples during the training process. First, we design a learnable similarity function to compute an accurate mix ratio. Second, an approach is investigated as a regularized term to model the uncertainty of the mixed samples. We conduct experiments on five image benchmarks, and extensive experimental results imply that our method is capable of improving the performance of classifiers with different cutting-based mixup approaches. The source code is available at https://github.com/JinXins/SUMix. + +Keywords: mixup, data augmentation, image classification + +# 1 Introduction + +Deep learning has emerged as a viable alternative to traditional machine learning in various tasks, e.g. computer vision [3, 29, 34], natural language processing [10, 43, 62], and video applications [12, 37, 59], because of its robust feature representation capacity. However, deep learning requires a lot of data to train thousands of network parameters. Despite the self-supervised learning paradigm [18, 31, 33, 36] enables task agnostic pre-training without annotations, the labeled data of specific scenarios is still required and generally limited, where + +![](images/35abdef1888190e637e8707039dc0671db77cfcbe317f6ba2a131988cdc5ab2f.jpg) +Fig. 1: The diagram shows different cases of raw samples that underwent the CutMix with a mixing ratio of 0.5 to obtain mixed samples. Left: the raw samples and their one-hot labels; Right: labels of the mixes obtained using a redefined mixing ratio $\widetilde{\lambda}$ for different cases. + +deep neural networks (DNNs) are prone to overfitting [2]. Therefore, the representation capacity is not effectively exploited, resulting in poor discriminative performances and generalization abilities. + +Data Augmentation (DA) methods as a solution have been proposed to prevent DNNs from over-fitting. Existing methods transform original data to generate new samples for data augmentation. For example, in computer vision tasks, classical techniques such as rotating, scaling, and panning have been used to generate augmented samples for model training. Similarly, in natural language processing tasks, methods such as proximate word substitution and randomly misspelling words are employed to produce new data. These data augmentation methods have proven to be effective in improving the performance of network models. They allow the models to learn a larger variety of samples based on augmentation data. Currently, data augmentation methods can be categorized into three classes. (i) The contrast combination methods [8, 9, 17, 38, 39, 58] employs affine matrix transformations and color enhancement for data augmentation. However, the resulting images are significantly different from the original ones. (ii) Mixup method produces virtual mixup examples via a simple convex combination of pairs of examples [51, 64, 75]. (iii) The generative approaches, e.g. Generative Adversarial Networks(GANs) [13, 71], generate fake samples by adversarial learning. Overall, these data augmentation methods effectively alleviate the overfitting problem and improve the performance of deep learning models on limited training data. + +Mixup [75] method as a recent DA model produces virtual samples and labels by linearly interpolating the sample images as well as their corresponding labels, and such a simple method has proven to be effective in improving the representation capacity of the deep neural networks. In recent years, various mixup methods have been proposed for data augmentation. For example, CutMix [72] shifts the numerical mixing to spatial, where the mixing coefficient $\lambda$ determines the size of the area of the cut and introduces the idea of Cutout [11], which cuts and mixes two samples, as shown in Figure 1. This approach not only helps the network to generalize better and improve object localization but also enables continuous performance improvement in tasks such as Pascal VOC + +![](images/dc0a669dcab3376c350dd9d1aea2a1e0c23ba05ce62fde6ffcc18767295ca3bd.jpg) +Fig. 2: The figure shows some hand-crafted mixup methods with "Label Mismatch" problem. In CutMix and ResizeMix, the "Shark" is mostly obscured, while in SaliencyMix and FMix, it is difficult to notice the "Panda". + +object detection [53] and MS-COCO image captioning [65]. In addition, CutMix improves the robustness of the model, making it perform better in the face of input corruption and cross-distribution detection. However, the CutMix method randomly selects the cropping region, which results in the background region of the target sample being clipped to the feature region of the source sample, and in the extreme case, the source sample will be completely occluded, resulting in an augmented sample that has no feature information at all, a phenomenon known as "Label MisMatch" shown in Figure.2. Since the one-hot label of the sample is unchanged, Decoupled Mix [42] found that the Mixup method relies heavily on the loss function, forcing the prediction of the network model to obey two higher peaks, which results in an inaccurate gradient result due to the label mismatch problem in the calculation of the loss function. This leads to an inaccurate gradient result when calculating the loss function, which affects the training of the network model. + +To improve the "Label MisMatch" problem, we propose SUMix, which consists of a mix ratio learning module and an uncertain estimation module. The former focuses on computing the proportion of two images and the latter aims to learn the uncertainty information of mixed samples. Firstly, we design a function to compute the semantic distance between the mixed and original samples to determine the ratio $\lambda$ . Secondly, we present a method to learn the uncertainty + +![](images/a701f0ea1cfa235cd773f0447d6478048348d22696aaec16d632b5c317304116.jpg) +Fig. 3: The results illustrate the application of SUMix. Left: Top-1 Accuracy improvement from the mixup approaches with SUMix; Right: Comparison of Vanilla method, CutMix, and SUMix for CAM visualization. + +of the mixed. This adapted feature vector effectively mitigates issues related to computing the loss function caused by discrepancies in semantic and uncertainty aspects. Extensive experiments on five image classification datasets verify that our proposed SUMix can remarkably improve performances of existing mixup augmentations (as shown in Figure 3) in a plug-and-play manner while achieving better robustness. Our main contributions are as follows: + +(a) We propose a learnable metric to compute the mixed ratio by similarity between the mixed samples and the original samples. +(b) We further consider the uncertainty and semantic information of the mixed samples and recalculate a reasonable feature vector, providing an additional regularization loss for model training. +(c) Our approach helps popular Cutting-based mixup methods to improve classification tasks without spending too excessive extra time overhead. + +# 2 Related Work + +Sample-based Mixup Sample-based Mixup approaches [21-23, 30, 70] focus on obtaining more reliable mixed samples or hidden space feature maps. Usually combined with inserting prior knowledge or saliency information [54]. CutMix [72] obtains the mixed samples by randomly cutting and pasting between the two samples. Due to the large randomness of CutMix cropping, ResizeMix [52] replaces the cut operation with the resize operation. There are a series of works such as FMix [16], GridMix [1] etc, which use similar methods. To extract the saliency information within the samples during the training of the model, some work such as SaliencyMix [61], Attentive-CutMix [67] etc, use a saliency extraction module to mix the samples in a guidedness. PuzzleMix [26] extracts the saliency regions from the source and target samples respectively through the saliency capability learned from the model and then extracts the saliency regions according to the designed "optimal transmission". Co-Mixup [27] further improves PuzzleMix by increasing the number of mixed samples from 2 to 3. AutoMix [44] and SAMix [32] analyze the advantages and disadvantages of the hand-crafted and saliency method and divide the mixing task into two sub-tasks. AdAutoMix [51] further improves AutoMix by proposing adversarial training to train the generator and obtain "hard mixed sample" to train the model. Manifold Mixup [64] and AlignMix [63] moved the mixup method to hidden space by mixing the feature information of the samples. + +Label-based Mixup Label-based Mixup approaches aim to solve the problems of 1) time overhead and 2) sample-label mismatch. These methods usually use some known information to recalculate the mix ratio of labels. Saliency Grafting [50] uses the area of saliency features of each sample in a mixed sample to reconstruct a new $\lambda$ . LUMix [57] differs from previous ways to address the label noise introduced by the mixup process instead of modeling the label noise. Methods such as TransMix [5] and TokenMix [40] use attention scores to mix the + +samples to compute $\lambda$ based on the area of the mask. Mixpro [76] obtains a new $\lambda$ by combining the attention score with the masked area and obtains a stable output with a factor $\alpha$ . Other methods like SMMix [6], Token Labeling [24] etc, are based on attention score and masking area to get a mix ratio. + +Uncertainty Modeling The presence of naturally occurring uncertainties such as occlusion [25] and blurring [55] in the images tends to influence the training of network models. Several works [4,56,74] have attempted to model uncertainty in face recognition, age estimation, and point cloud segmentation, respectively, to obtain better robustness and generalization. In addition, it is also able to deal with Out-distribution problems. [15] proposed a Bayesian classifier for obtaining Out-of-distribution uncertainty. [49] defines a Gaussian distribution to represent the samples and uses Monte Carlo sampling to adopt multiple points from the Gaussian distribution for optimizing the metric. [68] uses an introspective modeling approach that bypasses distribution optimization to further improve the robustness of the model. + +# 3 Preliminaries + +MixUp Classification. Let us define an $X \in \mathbb{X}$ as the input data and its corresponding $Y \in \mathbb{Y}$ as the label of the input data. For the vanilla mixup, we construct a mixed sample $\widetilde{X}$ and a mixed label $\widetilde{Y}$ from Eq.(1). + +$$ +\begin{array}{l} \begin{array}{l} \widetilde {X} = \lambda * X _ {a} + (1 - \lambda) * X _ {b}, \\ \widetilde {\tilde {\Sigma}} = \dots \end{array} \tag {1} \\ \widetilde {Y} = \lambda * Y _ {a} + (1 - \lambda) * Y _ {b}, \\ \end{array} +$$ + +where $\lambda$ is a mix ratio obtained from a Beta distribution, $(X_{a},X_{b})\in \widetilde{X}$ , and $(Y_{a},Y_{b})\in \widetilde{Y}$ are the same batch of data. Mixup obtains the mixed samples by pixel interpolating with a $\lambda$ , CutMix randomly obtains a two-valued rectangular mask $\mathbb{1}$ , where $\mathbb{1} = \lambda = \frac{r_w r_h}{WH}$ , W, H are the width and height of the data and $r_w, r_h$ are the width and height of the rectangular mask. We get the mixed samples $\widetilde{X} = \mathbb{1}\odot X_a + (1 - \mathbb{1})\odot X_b$ . + +The Devil Lies in Label MisMatch. Mixup contains two different labels, so the Mixed Cross Entropy (MCE) is used in calculating the loss function as shown in Eq.(2). + +$$ +\mathcal {L} _ {M C E} = \ell_ {C E} \left(f _ {\theta} (\tilde {X}), Y _ {a}\right) * \lambda + \ell_ {C E} \left(f _ {\theta} (\tilde {X}), Y _ {b}\right) * (1 - \lambda), \tag {2} +$$ + +where $f_{\theta}(\cdot)$ denotes the network model, and $\ell(\cdot)$ denotes the loss function, $\ell_{CE}$ was cross-entropy loss, which minimizes the loss during the training process. However, since some hand-crafted mixup methods are random, e.g. CutMix, ResizeMix, but the labels of the mixed samples are fixed $\widetilde{Y}$ , then in the case shown in Figure.2, the features of a certain classification are occluded and reduced or do not exist in the mixed samples, resulting in the feature information not matching the labels. The phenomenon is called "Label MisMatch". In this case, the proportion of labels should be reduced or removed to get an accurate result when calculating the loss function. + +![](images/7476f3d50f69c94437f390b9e1e879674e6f99f935bc9978630523c1dd76ee58.jpg) +Fig. 4: The left diagram represents the training process of SUMix. The raw and mixed samples are encoded through the model to obtain the semantic and uncertainty information, and the new mix ratio and loss function for regularization are obtained. The right diagram is the spatial distance between the raw and mixed samples. + +![](images/dc9a0917fcc158fc6b2342bfd2f78f1b3a87f477f4aa4261082715c551267fca.jpg) + +# 4 SUMix + +In this section, we focus on detailing SUMix, 1) the uncertainty classifier, 2) how to obtain the mix ratio $\lambda$ , and 3) the regularization function in the case of uncertainty and semantic information. Firstly, we show how to use deep learning classifiers for loss function computation and network model training, secondly, how to redefine the mix ratio $\lambda$ based on semantic information, and lastly, how to consider uncertainty and regularization. The pipeline of SUMix is shown in Fig. 4. + +# 4.1 Uncertainty Classifier + +Assuming $\mathbb{S} = \{x_s|s = 1,2,\dots ,S\}$ is a training set and $\mathbb{S}$ is the number of training images. We take $N$ samples from it as a sample set, where $\mathbb{X}\in \mathbb{R}^{N\times 3\times H\times W}$ , $\mathbb{Y}\in \mathbb{R}^{N\times K}$ are the labels and $K$ is the number of classes. For the original computation of the mixup loss function, we use a deep neural network $f_{\theta}(\cdot)$ with a trainable parameter that maps the sample $x_{i}$ to the labels $y_{i}$ . Unlike the vanilla mixup loss computation in Eq.(2), SUMix adds a regularized loss function as shown in Eq.(3): + +$$ +\mathcal {L} _ {s u} = \frac {1}{N} \left(\sum_ {i = 0} ^ {N} \mathcal {L} _ {M C E} \left(f _ {\theta} \left(\widetilde {X} _ {i}\right), Y, \widetilde {\lambda} _ {i}\right)\right) + \zeta * \mathcal {L} _ {M C E} \left(S U \left(f _ {\theta} (\widetilde {X}), U _ {\omega}\right), \widetilde {Y}\right), \tag {3} +$$ + +where $\widetilde{\lambda}_i$ are the redefined lambda, $SU(\cdot)$ denotes a function to obtain feature, and $\zeta$ is a hyperparameter used to weight the regulated loss function, set to [0.1, 0.2, 0.5, 1.0]. $U_{\omega}$ denotes a function as a measure of uncertainty. Finally, the parameters of the deep neural network are updated continuously to minimize $\mathcal{L}_{su}$ . + +# 4.2 Mix ratio Learning + +Mixup approaches always depend on a parameter taken from the $Beta(\alpha, \alpha)$ distribution, which we call the mix ratio $\lambda$ , to balance the mix ratio of the two samples. Lots of mixup approaches use $\lambda$ to augment the mixup sample set $\mathbb{X}$ after obtaining it, and notably, this $\lambda$ serves the whole sample set. For some saliency-based or automatic-based mixups, e.g. PuzzleMix [26], AutoMix [44], it is possible to push the mixing method to obtain the best possible mixed samples according to the $\lambda$ ratio, balancing the sample ratio with the $\lambda$ as much as possible. However, some hand-crafted approaches with randomness cannot accurately match the mix ratio. The problem we want to solve is how to match each mixup sample with its own $\lambda$ . + +For a mixed sample $\widetilde{x} = \left[\widetilde{X}\right]_i^N\in R^{3\times H\times W}$ , semantic information links need to be established with the original sample pair $(x_{a},x_{b})$ . They are fed into a deep neural network $f_{\theta}(\cdot)$ , e.g. ResNet18 [19]. feature vectors $\widetilde{z}$ and $(z_{a},z_{b})$ are obtained, and the feature vectors are normalized to modify their similarity according to Eq.(4): + +$$ +\widetilde {\lambda} _ {a} = \frac {\lambda * \mathrm {e} ^ {- \| (\sigma (\widetilde {z} - z _ {a}) \| _ {2}}}{\lambda * \mathrm {e} ^ {- \| (\sigma (\widetilde {z} - z _ {a}) \| _ {2}} + (1 - \lambda) * \mathrm {e} ^ {- \| (\sigma (\widetilde {z} - z _ {b}) \| _ {2}}}, \tag {4} +$$ + +where $\sigma(\cdot)$ is the softmax function used to normalize the vectors, and $\|\cdot\|_2$ is the l2 norm, which imposes a limit on the distance between the two. When the difference in semantic information between $\widetilde{x}$ and $x_a$ is large, which indicates that the mixed sample has a small amount of the feature information of the raw sample, $\widetilde{\lambda}_a$ tends to 0. Similarly, when the difference in semantic information between the two is small, $\widetilde{\lambda}_a$ tends to 1. $\widetilde{\lambda}_b = 1 - \widetilde{\lambda}_a$ . Thus, we have calculated the mixed sample set with its own matching $\lambda$ . a more sensible loss is computed according to Eq.(3). + +# 4.3 Uncertainty Estimation + +Deep learning tasks involve two types of uncertainty: Epistemic uncertainty and Aleatoric uncertainty. Epistemic uncertainty arises from the model's inherent uncertainty due to insufficient training data, resulting in low confidence in unknown data. In contrast, Aleatoric uncertainty arises from errors in the data itself, which can introduce bias into the dataset. The greater the bias, the greater the Aleatoric uncertainty. In the case of the mixup task, the error in the data occurs when the mixup method is stochastic. Our goal is to address this error and enable the network to adapt accordingly. + +Wang et al. [68] propose an introspective metric that models a new vector for classification based on the semantic information distance between two images and their uncertainty distance. The mixup methods differ from this as they typically incorporate feature information from two samples. To begin, we encode the mixed sample and the raw sample using an encoder to obtain the feature + +vectors $\widetilde{z}$ and $z$ , respectively. Next, we calculate the uncertainty by passing $\widetilde{z}$ and $z$ through $U_{\omega}$ , by Eq.(5). + +$$ +\begin{array}{l} u = \parallel M L P (\sigma (z)) \| _ {2}, \\ \widetilde {\gamma} - \parallel M L P (\sigma (\widetilde {z})) \| \end{array} \tag {5} +$$ + +where is a linear layer with a softmax function and l2 norm. We define $\beta = \widetilde{u} + u$ as the uncertainty of mixed samples and raw samples. + +After capturing the uncertainty and semantic information, as demonstrated in the right part of Figure.4, Wang et al. [68] argue that in the feature space, the smaller the radius of a sample, the more aggregated the class is, and the smaller the uncertainty. By combining the semantic information, we reconstruct the mixed samples $\widetilde{X}$ and the raw samples $X$ to obtain $Z_{su}$ according to Eq.(6). + +$$ +Z _ {s u} = \mathrm {e} ^ {- (\beta + \| \sigma (\tilde {z} - z) \| _ {2})} \tag {6} +$$ + +When there is significant uncertainty or semantic information has a large difference, $Z_{su}$ receives a small gradient; conversely, when there is little uncertainty and small differences in semantic information, $Z_{su}$ receives a normal gradient that tends to 1. This approach is used as a method of regularization. + +# 5 Experiments + +To verify the performance of SUMix, we conducted a large number of experiments on six datasets, CIFAR100 [28], Tiny-ImageNet [7], ImageNet-1K [29], CUB-200 [66], FGVC-Aircrafts [47], and CIFAR100-C. To verify the effectiveness of the improved "Label MisMatching" method, we mainly compared it with some patch-wise mixup methods based on hand-crafted, i.e., CutMix, FMix, ResizeMix, and SaliencyMix. We have implemented our algorithm on the open-source library OpenMixup [35], and some parameters are designed according to OpenMixup. The classification results are the median of the top-1 test accuracy from the last 10 training epochs, but on ImageNet-1K, we choose the last 5 training epochs. To facilitate comparison, our results are highlighted in bold. + +# 5.1 Datasets + +We choose the three most classical datasets and two fine-grained datasets in the classification task. CIFAR100 has a total of 50,000 images for the training set and 10,000 images for the testing set, containing 100 different classes, each with a $32 \times 32$ resolution; Tiny-ImageNet contains 10,000 images for training and 10,000 images for testing, each with a $64 \times 64$ resolution and a total of 200 classes; The ImageNet-1K dataset contains 1.2 million images for training and 50,000 images for validation; CUB200 dataset contains images of 200 different bird species, a total of 11,788 images; the FGVC-Aircrafts dataset contains images of 100 different types of aircraft, a total of 10,000 images. + +# 5.2 Implementation Details + +Hyper-parameter of $\alpha$ . In the experimental results we report, the hyperparameter $\alpha$ of the mixup variants, which sampling from the $Beta(\alpha, \alpha)$ distribution used by the model based on the CNN architecture and the model based on the ViT architecture are different. For CutMix, the CNN-based models, e.g. ResNet18, ResNeXt50, and Wide-ResNet28 set as 0.2, and ViT-based models, DeiT and Swin-Transformer set as 2.0. For FMix, CNN-based set 0.2, and ViT-based set 1.0. For PuzzleMix, the CNN-based set is 1.0, and the ViT-based set is 2.0. In CNN-based and Vit-based architectures alike, SaliencyMix, ResizeMix, and AutoMix exhibit identical alpha values, which are 0.2, 1.0, and 2.0, respectively. + +Hyper-parameter of $\zeta$ . In SUMix, we set the hyper-parameters $\zeta$ of $\mathcal{L}_{su}$ as follows: On the CIFAR100 dataset, we report five different models, the $\zeta$ setting shown in Appendix A.1. On Tiny-ImageNet, we set the $\zeta = 1.0$ for all mixup methods based on ResNet18 and ResNeXt50. Similarly with ImageNet-1K, we set the $\zeta = 0.5$ based on ResNet18. The settings we reported for the fine-grained datasets are also shown in Appendix A.1. + +# 5.3 Image Classification + +Generic classification. We initially train ResNet18 [19] and ResNeXt50 [69] using the CIFAR100 dataset for a total of 800 epochs. The experimental parameters are set as in OpenMixup [35]: a basic learning rate of 0.1, SGD [45] optimizer with momentum of 0.9, weight decay of 0.0001, and batch size of 100 for each iteration. Wide-ResNet [73] using the CIFAR100 dataset for a total of 400 epochs, learning rate of 0.03, SGD optimizer with momentum of 0.9, weight decay of 0.001. For the Tiny-ImageNet dataset, we use the same experimental parameters as the CIFAR100 experiment, except the base learning rate to 0.2 and training for 400 epochs. As for ImageNet-1K, we follow the Pytorch-style settings for training ResNet18 of 100 epochs. For the CIFAR100 dataset, we use + +Table 1: Top-1 accuracy $(\%)\uparrow$ of mixup methods on CIFAR-100, Tiny-ImageNet and ImageNet-1K. * denotes mixup methods with SUMix. + +
MethodCIFAR100Tiny-ImageNetImageNet-1K
ResNet18ResNeXt50W-ResNet28-8ResNet18ResNeXt50ResNet18
CutMix78.1778.3284.4565.5366.4768.95
FMix79.6979.0284.2163.4765.0869.96
SaliencyMix79.1278.7784.3564.6066.5569.16
ResizeMix80.0180.3584.8763.7465.8769.50
CutMix*79.7879.9184.5665.7168.7469.71
FMix*80.2080.7984.3263.6967.1270.48
SaliencyMix*79.9179.3284.5865.6868.9269.52
ResizeMix*80.3880.7284.9165.3067.4969.76
Avg. Gain+0.82+1.07+0.12+0.81+2.07+0.47
+ +Table 2: Top-1 accuracy $(\%)\uparrow$ of mixup methods on CUB200, FGVC-Aircrafts, and Standford-Cars. * denotes mixup methods SUMix. † denotes that we reproduce the results of that experiment. + +
MethodCUB200FGVC-Aircrafts
ResNet18ResNeXt50ResNet18ResNeXt50
CutMix†77.7083.6778.8484.55
FMix77.2884.0679.3684.10
SaliencyMix†75.7782.8379.7884.31
ResizeMix78.5084.1678.1084.08
CutMix*78.2083.7179.7285.84
FMix*79.2484.3379.4884.64
SaliencyMix*76.9882.8479.9084.49
ResizeMix*78.5684.2380.2985.12
Avg. Gain+0.93+0.10+0.82+0.67
+ +
MethodCIFAR100
DeiT-SmallSwin-Tiny
CutMix74.1280.64
FMix70.4180.72
SaliencyMix69.7880.40
ResizeMix68.4580.16
CutMix*75.26+1.1480.83+0.19
FMix*70.69+0.2880.73+0.01
SaliencyMix*70.31+0.5380.71+0.29
ResizeMix*68.78+0.3380.59+0.43
+ +Table 3: Top-1 accuracy(\%)↑ of Table 4: Top-1 accuracy(\%)↑ of saliency-based mixup methods on CIFAR100. * de- mixup methods on CIFAR100. * denotes mixup notes mixup methods with SUMix. methods with SUMix. + +
MethodCIFAR100
ResNet18ResNeXt50
PuzzleMix81.1381.69
AutoMix82.0482.84
PuzzleMix*81.43+0.3082.60+0.91
AutoMix*82.30+0.2683.82+0.98
AdAutoMix82.3283.81
+ +basic data augmentation including RandomFlip and RandomCrop with 4 pixels padding, and for the Tiny-ImageNet and ImageNet-1K datasets, we also use the RandomFlip and RandonCrop data augmentation operations. + +Table 1 shows our five comparisons on three different datasets, where the use of SUMix brings some improvement in the classification accuracy of CutMix, FMix, SaliencyMix, and ResizeMix. For the CIFAR100 dataset, on ResNet and its variants ResNeXt50, W-ResNet28-8, SUMix improves their classification results by an average of $0.82\%$ , $1.07\%$ , and $0.12\%$ , respectively; for the Tiny-ImageNet dataset, on ResNet18 and ResNeXt50, the average improvements of $0.81\%$ and $2.07\%$ ; for the ImageNet-1K dataset, on ResNet18, the average improvement is $0.47\%$ . + +Fine-grained classification. We load the official PyTorch pre-trained models on ImageNet-1k as initialization and fine-tuning pre-trained ResNet18 and ResNeXt50 using SGD optimizer with a learning rate of 0.001, a momentum of 0.9, weight decay of 0.0005 on CUB200, FGVC-Aircrafts, total training 200 epochs. Table 2 presents the comparison results for the bird and aircraft datasets. + +![](images/1bcf8a841b06735580a596c4f537b930fc238fc2f4cbe46957338b4881625bd1.jpg) +Fig. 5: The top of the figure shows a visualization of the sample at $0\%$ to $100\%$ occlusion ratio. The light blue curves in the lower four subfigures show the classification accuracy of CutMix, FMix, SaliencyMix, and ResizeMix on ImageNet-1K using ResNet18 for 100 epochs of training, respectively, and the red curve shows the classification accuracy of these methods using SUMix. + +SUMix improves the classification results for the CUB200 dataset by an average of $0.93\%$ and $0.10\%$ at ResNet18 and ResNeXt50, respectively. The FGVC-Aircrafts dataset improves by an average of $0.82\%$ and $0.67\%$ at ResNet18 and ResNeXt50. + +Classification based on ViTs. To further show the performance of SUMix, we perform classification experiments using the ViT-based model DeiT [60] and Swin-Transformer [41] on the CIFAR100 dataset, and the results are shown in Tab 3. We resized the size to $224 \times 224$ , and in the training set, we trained DeiT and Swin-Transformer of 200 epochs by AdamW [46] optimizer with a batch size of 100 on CIFAR100. The basic learning rates of DeiT and Swin are 1e-3 and 5e-4, and the weight decay of 0.05, like OpenMixup settings, we also used some RandAugment [9] operations. We reported the best top-1 accuracy in the last 10 training epochs, and the results show that our SUMix also can bring effectiveness to ViTs. + +Table 3 shows the results of the ViT models using SUMix, which improves the classification results by an average of $0.57\%$ and $0.23\%$ at DeiT-Small and Swin-Tiny. + +Classification based on Saliency methods. To further explore the capability of SUMix, we not only apply it to some Hand-crafted methods, we apply SUMix to PuzzleMix and AutoMix, which are based on saliency-learning methods, and Tab 4 shows that SUMix will also bring some boosts, PuzzleMix and AutoMix has gained $0.30\%$ , $0.26\%$ at ResNet18, and $0.91\%$ , $0.98\%$ at ResNeXt50. Those methods combined with SUMix can even reach almost close to the performance of the current state-of-the-art method AdAutoMix. + +Table 5: Top-1 acc(%)↑ and FGSM error(%)↓ Table 6: Ablation experiments on CI-of ResNet18 without and with SUMix. FAR100, CUB200 based on ResNet18. + +
Clean Acc(%)↑Corruption Acc(%)↑FGSM Error(%)↓
MethodMCESUMixMCESUMixMCESUMix
CutMix78.1779.7843.0644.3191.1590.41
FMix79.6980.2048.7949.1489.1689.08
SaliencyMix79.1279.9143.7344.3689.6491.49
ResizeMix80.0180.3846.1246.2890.0491.05
+ +
MethodCIFAR100CUB200
CutMixSaliencyMixFMixSaliencyMix
baseline78.1779.1277.2875.77
+λ79.1379.7979.0376.46
+ℓs79.5178.7179.0376.34
+ℓu79.6079.4379.2476.72
+ℓsu79.7879.9179.2476.98
+ +# 5.4 Occlusion Robustness + +To analyze the enhancement brought by SUMix to the robustness of the mixup methods, we randomly conducted masking patches experiment [48]. The experiment involved dividing the image into $16 \times 16$ patches and masking them according to different ratios. The resulting image was then classified by the model. We used the ImageNet dataset, which contains 1000 daily life objects. It is more representative. The ResNet18 classifier was used to classify the masked images and verify the model's robustness capability. The curves at the bottom of Figure.5 show that SUMix provides a performance gain compared to the corresponding mixup method at different occlusion ratios. + +# 5.5 Robustness + +We first evaluated robustness against corruption on CIFAR-100-C [20], which was designed manually to provide 19 different corruptions (e.g., noise, blur, fog, etc.) to evaluate robustness. In Table 5, we show the results of improving some mixup methods with SUMix. We further investigate the FGSM [14] white box attack experiments with $8/255\ell_{\infty}$ epsilon ball and similarly show the results of the improvement of our approach on mixup methods in Table 5. + +# 5.6 Ablation Study + +In the ablation study, we primarily analyze the effect of the two SUMix modules in the model. The following two analyses are primarily done. (1) Can redefining $\lambda$ bring gains in models? (2) Can uncertainty lead to regularization gains? Table 6 shows the experimental results for CutMix, SaliencyMix, and FMix using ResNet18 and ResNeXt50 on the CIFAR100 and CUB200 datasets. + +(1) Table 6 shows that with our redefined $\widetilde{\lambda}$ , CutMix and SaliencyMix are improved by $0.96\%$ and $0.67\%$ on the CIFAR100 dataset, and FMix and SaliencyMix are improved by $1.75\%$ and $0.69\%$ on the CUB200 dataset, which means that the module can bring a significant increase in the classification ability, in other words, the redefined $\widetilde{\lambda}$ can obtain a more sensible loss function for the model to train. + +(2) To calculate the regularization loss, we consider three cases: 1) using only the semantic information $\ell_s$ , 2) using only the uncertainty information $\ell_u$ , and 3) using both the semantic and uncertainty information $\ell_{su}$ . Table 6 shows that on the CUB200 dataset, FMix and SaliencyMix improve accuracy by $1.75\%$ and $0.57\%$ respectively when using only $\ell_s$ , and by $1.96\%$ and $0.95\%$ respectively when using only $\ell_u$ . When both types of information are used, accuracy improves by $1.75\%$ and $1.21\%$ . Similarly, on the CIFAR100 dataset, there was a $1.61\%$ and $0.79\%$ improvement. The experimental results show that combining semantic and uncertainty information can significantly improve accuracy. + +# 6 Conclusion and Limitations + +Conclusion. In this paper, we propose the SUMix approach to regularize the DNN in terms of the semantic information and uncertainty between the raw and mixed samples. Specifically, the semantic information difference between the raw and mixed samples is used to obtain the mix ratio $\lambda$ , while the uncertainty is, in turn, combined with the semantic information to obtain a regularized loss function. Experimental results on five datasets proved the effectiveness of SUMix for the regularization of the mixup methods. + +Limitations. Currently, SUMix is due to learn how to obtain a sensible mixing ratio $\lambda$ and an uncertainty-based feature vector. These are currently proven only on classification tasks. In future work, we would like to apply and improve SUMix accordingly on some downstream tasks e.g. segmentation, detection, and other methods. And for semi-supervised tasks, we can get more accurate pseudo-labels for training models. + +# Acknowledgement + +This work is supported in part by the Scientific Innovation 2030 Major Project for New Generation of AI under Grant 2020AAA0107300, in part by the National Natural Science Foundation of China (Grant No. 61976030), in part by Chongqing Natural Science Foundation (Grant No. 2024NSCQ-MSX4893), in part by Project of CQ CSTC (Grant No. cstc2018jcyjAX 0057), in part by Innovative Research Projects for Postgraduate Students of Chongqing Business University (No. yjscxx2024-284-53, No. yiscxx2024-284-231, No. yjscxx2024-284-234). + +# References + +1. Baek, K., Bang, D., Shim, H.: Gridmix: Strong regularization through local context mapping. Pattern Recognition 109 (2021) 4 +2. Bishop, C.M.: Pattern recognition and machine learning. Springer (2006) 2 + +3. Bochkovskiy, A., Wang, C.Y., Liao, H.Y.M.: Yolov4: Optimal speed and accuracy of object detection. ArXiv abs/2004.10934 (2020) 1 +4. Chang, J., Lan, Z., Cheng, C., Wei, Y.: Data uncertainty learning in face recognition. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5710-5719 (2020) 5 +5. Chen, J.N., Sun, S., He, J., Torr, P.H., Yuille, A., Bai, S.: Transmix: Attend to mix for vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12135-12144 (2022) 4 +6. Chen, M., Lin, M., Lin, Z., Zhang, Y., Chao, F., Ji, R.: Smmix: Self-motivated image mixing for vision transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17260-17270 (2023) 5 +7. Chrabaszcz, P., Loshchilov, I., Hutter, F.: A downsampled variant of imagenet as an alternative to the cifar datasets. arXiv preprint arXiv:1707.08819 (2017) 8 +8. Cubuk, E.D., Zoph, B., Mane, D., Vasudevan, V., Le, Q.V.: Autoaugment: Learning augmentation policies from data. arXiv preprint arXiv:1805.09501 (2018) 2 +9. Cubuk, E.D., Zoph, B., Shlens, J., Le, Q.V.: Randaugment: Practical automated data augmentation with a reduced search space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 702-703 (2020) 2, 11 +10. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018) 1 +11. DeVries, T., Taylor, G.W.: Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552 (2017) 2 +12. Feichtenhofer, C., Fan, H., Malik, J., He, K.: Slowfast networks for video recognition. In: ICCV. pp. 6201-6210 (2019) 1 +13. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. Communications of the ACM 63(11), 139-144 (2020) 2 +14. Goodfellow, I.J., Shlens, J., Szegedy, C.: Explaining and harnessing adversarial examples. In: ICLR (2015) 12 +15. Guénais, T., Vamvourellis, D., Yacoby, Y., Doshi-Velez, F., Pan, W.: Bacoun: Bayesian classifiers with out-of-distribution uncertainty. arXiv preprint arXiv:2007.06096 (2020) 5 +16. Harris, E., Marcu, A., Painter, M., Niranjan, M., Hare, A.P.B.J.: Fmix: Enhancing mixed sample data augmentation. arXiv preprint arXiv:2002.12047 (2020) 4 +17. Hataya, R., Zdenek, J., Yoshizoe, K., Nakayama, H.: Faster autoaugment: Learning augmentation strategies using backpropagation. In: European Conference on Computer Vision. pp. 1-16. Springer (2020) 2 +18. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: CVPR. pp. 9729-9738 (2020) 1 +19. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 770-778 (2016) 7, 9 +20. Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261 (2019) 12 +21. Hendrycks, D., Mu, N., Cubuk, E.D., Zoph, B., Gilmer, J., Lakshminarayanan, B.: Augmix: A simple data processing method to improve robustness and uncertainty. arXiv preprint arXiv:1912.02781 (2019) 4 + +22. Hendrycks, D., Zou, A., Mazeika, M., Tang, L., Li, B., Song, D., Steinhardt, J.: Pixmix: Dreamlike pictures comprehensively improve safety measures. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022) 4 +23. Huang, S., Wang, X., Tao, D.: Snapmix: Semantically proportional mixing for augmenting fine-grained data. In: AAAI (2020) 4 +24. Jiang, Z., Hou, Q., Yuan, L., Zhou, D., Jin, X., Wang, A., Feng, J.: Token labeling: Training a $85.4\%$ top-1 accuracy vision transformer with 56m parameters onImagenet. ArXiv abs/2104.10858 (2021) 5 +25. Kendall, A., Gal, Y.: What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems 30 (2017) 5 +26. Kim, J.H., Choo, W., Song, H.O.: Puzzle mix: Exploiting saliency and local statistics for optimal mixup. In: International Conference on Machine Learning. pp. 5275-5285. PMLR (2020) 4, 7 +27. Kim, J., Choo, W., Jeong, H., Song, H.O.: Co-mixup: Saliency guided joint mixup with supermodular diversity. In: International Conference on Learning Representations (2020) 4 +28. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009) 8 +29. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. In: Advances in neural information processing systems. pp. 1097-1105 (2012) 1, 8 +30. Lee, J.H., Zaheer, M.Z., Astrid, M., Lee, S.I.: Smoothmix: a simple yet effective data augmentation to train robust classifiers. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 756-757 (2020) 4 +31. Li, S., Jin, W., Wang, Z., Wu, F., Liu, Z., Tan, C., Li, S.Z.: Semireward: A general reward model for semi-supervised learning. In: International Conference on Learning Representations (2024) 1 +32. Li, S., Liu, Z., Wang, Z., Wu, D., Liu, Z., Li, S.Z.: Boosting discriminative visual representation learning with scenario-agnostic mixup. ArXiv abs/2111.15454 (2021) 4 +33. Li, S., Liu, Z., Zang, Z., Wu, D., Chen, Z., Li, S.Z.: Genurl: A general framework for unsupervised representation learning. IEEE Transactions on Neural Networks and Learning Systems (2024) 1 +34. Li, S., Wang, Z., Liu, Z., Tan, C., Lin, H., Wu, D., Chen, Z., Zheng, J., Li, S.Z.: Efficient multi-order gated aggregation network. ArXiv abs/2211.03295 (2022) 1 +35. Li, S., Wang, Z., Liu, Z., Wu, D., Li, S.Z.: Openmixup: Open mixup toolbox and benchmark for visual representation learning. https://github.com/Westlake-AI/openmixup (2022) 8, 9 +36. Li, S., Wu, D., Wu, F., Zang, Z., Li, S.Z.: Architecture-agnostic masked image modeling – from vit back to cnn. In: International Conference on Machine Learning (2023) 1 +37. Li, S., Zhang, Z., Liu, Z., Wang, A., Qiu, L., Du, F.: Tlpg-tracker: Joint learning of target localization and proposal generation for visual tracking. In: IJCAI (2020) 1 +38. Li, Y., Hu, G., Wang, Y., Hospedales, T., Robertson, N.M., Yang, Y.: Differentiable automatic data augmentation. In: European Conference on Computer Vision. pp. 580-595. Springer (2020) 2 + +39. Lim, S., Kim, I., Kim, T., Kim, C., Kim, S.: Fast autoaugment. In: Advances in Neural Information Processing Systems (NeurIPS) (2019) 2 +40. Liu, J., Liu, B., Zhou, H., Li, H., Liu, Y.: Tokenmix: Rethinking image mixing for data augmentation in vision transformers. In: European Conference on Computer Vision (2022) 4 +41. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: International Conference on Computer Vision (ICCV) (2021) 11 +42. Liu, Z., Li, S., Wang, G., Wu, L., Tan, C., Li, S.Z.: Harnessing hard mixed samples with decoupled regularizer. In: Thirty-seventh Conference on Neural Information Processing Systems (2023) 3 +43. Liu, Z., Li, S., Wang, L., Wang, Z., Liu, Y., Li, S.Z.: Short-long convolutions help hardware-efficient linear attention to focus on long sequences. In: Forty-first International Conference on Machine Learning (2024) 1 +44. Liu, Z., Li, S., Wu, D., Liu, Z., Chen, Z., Wu, L., Li, S.Z.: Automix: Unveiling the power of mixup for stronger classifiers. In: European Conference on Computer Vision. pp. 441-458. Springer (2022) 4, 7 +45. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983 (2016) 9 +46. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: ICLR (2019) 11 +47. Maji, S., Rahtu, E., Kannala, J., Blaschko, M., Vedaldi, A.: Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151 (2013) 8 +48. Naseer, M., Ranasinghe, K., Khan, S.H., Hayat, M., Khan, F.S., Yang, M.H.: Intriguing properties of vision transformers. In: Neural Information Processing Systems (2021) 12 +49. Oh, S.J., Murphy, K., Pan, J., Roth, J., Schroff, F., Gallagher, A.: Modeling uncertainty with hedged instance embedding. arXiv preprint arXiv:1810.00319 (2018) 5 +50. Park, J., Yang, J.Y., Shin, J., Hwang, S.J., Yang, E.: Saliency grafting: Innocuous attribution-guided mixup with calibrated label mixing. In: AAAI Conference on Artificial Intelligence (2021) 4 +51. Qin, H., Jin, X., Jiang, Y., El-Yacoubi, M.A., Gao, X.: Adversarial automixup. arXiv preprint arXiv:2312.11954 (2023) 2, 4 +52. Qin, J., Fang, J., Zhang, Q., Liu, W., Wang, X., Wang, X.: Resizemix: Mixing data with preserved object information and true labels. arXiv preprint arXiv:2012.11101 (2020) 4 +53. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. arXiv preprint arXiv:1506.01497 (2015) 3 +54. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Gradcam: Visual explanations from deep networks via gradient-based localization. arXiv preprint arXiv:1610.02391 (2019) 4 +55. Shaw, G., Manolakis, D.: Signal processing for hyperspectral image exploitation. IEEE Signal processing magazine 19(1), 12-16 (2002) 5 +56. Shi, Y., Jain, A.K.: Probabilistic face embeddings. In: ICCV. pp. 6902-6911 (2019) 5 +57. Sun, S., Chen, J.N., He, R., Yuille, A., Torr, P., Bai, S.: Lumix: Improving mixup by better modelling label uncertainty. arXiv preprint arXiv:2211.15846 (2022) 4 +58. Suzuki, T.: Teachaugment: Data augmentation optimization using teacher knowledge. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10904-10914 (June 2022) 2 + +59. Tan, C., Gao, Z., Wu, L., Xu, Y., Xia, J., Li, S., Li, S.Z.: Temporal attention unit: Towards efficient spatiotemporal predictive learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18770-18782 (2023) 1 +60. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jegou, H.: Training data-efficient image transformers & distillation through attention. In: International Conference on Machine Learning (ICML). pp. 10347-10357 (2021) 11 +61. Uddin, A.S., Monira, M.S., Shin, W., Chung, T., Bae, S.H.: Saliencymix: A saliency guided data augmentation strategy for better regularization. In: International Conference on Learning Representations (2020) 4 +62. Vaswani, A., Shazeer, N.M., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NIPS (2017) 1 +63. Venkataramanan, S., Avrithis, Y., Kijak, E., Amsaleg, L.: Alignmix: Improving representation by interpolating aligned features. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022) 4 +64. Verma, V., Lamb, A., Beckham, C., Najafi, A., Mitliagkas, I., Lopez-Paz, D., Bengio, Y.: Manifold mixup: Better representations by interpolating hidden states. In: International Conference on Machine Learning. pp. 6438-6447 (2019) 2, 4 +65. Vinyals, O., Toshev, A., Bengio, S., Erhan, D.: Show and tell: A neural image caption generator. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3156-3164 (2015) 3 +66. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The caltech-ucsd birds-200-2011 dataset. California Institute of Technology (2011) 8 +67. Walawalkar, D., Shen, Z., Liu, Z., Savvides, M.: Attentive cutmix: An enhanced data augmentation approach for deep learning based image classification. In: IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 3642-3646 (2020) 4 +68. Wang, C., Zheng, W., Zhu, Z., Zhou, J., Lu, J.: Introspective deep metric learning. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023) 5, 7, 8 +69. Xie, S., Girshick, R., Dollár, P., Tu, Z., He, K.: Aggregated residual transformations for deep neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1492-1500 (2017) 9 +70. Yang, L., Li, X., Zhao, B., Song, R., Yang, J.: Recursiveemix: Mixed learning with history. Advances in Neural Information Processing Systems 35, 8427-8440 (2022) 4 +71. Yang, S., Wu, Y., Jin, X., El Yacoubi, M., Qin, H.: Cgan-da: A cross-modality domain adaptation model for hand-vein biometric-based authentication. JOURNAL OF Cyber-Physical-Social Intelligence 1, 3–12 (2022) 2 +72. Yun, S., Han, D., Oh, S.J., Chun, S., Choe, J., Yoo, Y.: Cutmix: Regularization strategy to train strong classifiers with localizable features. In: International Conference on Computer Vision (ICCV). pp. 6023-6032 (2019) 2, 4 +73. Zagoruyko, S., Komodakis, N.: Wide residual networks. In: Proceedings of the British Machine Vision Conference (BMVC) (2016) 9 +74. Zhang, B., Wonka, P.: Point cloud instance segmentation using probabilistic embeddings. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8883-8892 (2021) 5 +75. Zhang, H., Cisse, M., Dauphin, Y.N., Lopez-Paz, D.: mixup: Beyond empirical risk minimization. In: International Conference on Learning Representations (2018) 2 +76. Zhao, Q., Huang, Y., Hu, W., Zhang, F., Liu, J.: Mixpro: Data augmentation with maskmix and progressive attention labeling for vision transformer. In: The Eleventh International Conference on Learning Representations (2023) 5 + +# Appendix A + +# A.1 More Information about $\zeta$ + +We reported the hyper-parameter $\zeta$ of $\mathcal{L}_{su}$ on the CIFAR100 dataset shown in Table 7. And for the fine-grained dataset, CUB200 and FGVC-Aircraftd. We reported the settings in Table 8 and Table 9. + +Table 7: The $\zeta$ setting on the CIFAR100 dataset. + +
ζ settingCIFAR100
ResNet18ResNeXt50W-ResNet28-8DeiT-SmallSwin-Tiny
CutMix0.50.50.50.50.5
FMix1.01.00.50.50.5
SaliencyMix0.51.00.50.50.5
ResizeMix0.20.50.50.50.5
+ +Table 8: The $\zeta$ setting on the CUB200 Table 9: The $\zeta$ setting on the FGVC-dataset. + +
ζ settingCUB200ζ settingFGVC-Aircrafts
ResNet18ResNeXt50ResNet18ResNeXt50
CutMix1.00.5CutMix1.00.5
FMix0.50.1FMix0.50.5
SaliencyMix0.50.2SaliencyMix1.00.5
ResizeMix0.50.5ResizeMix0.50.5
+ +# A.2 More Visualization Comparison + +In Figure.6 and Figure.7, we show the visualization obtained by the mixup method using the CAM method with and without SUMix. + +![](images/5dda73e6bbf47db8e12e5c3d89e3ee6affe782ef8da874a5745f9e8923fdf298.jpg) +Fig.6: CAM visualization of the ResNet18 with and without SUMix trained on the ImageNet-1K dataset. + +![](images/b9cf32e5858c579bbd311bb0e69ca9e3dd749b7cde0ab31d96230aa9fb51a7a9.jpg) +Fig. 7: CAM visualization of the ResNet18 with and without SUMix trained on the ImageNet-1K dataset. \ No newline at end of file diff --git a/2024/SUMix_ Mixup with Semantic and Uncertain Information/images.zip b/2024/SUMix_ Mixup with Semantic and Uncertain Information/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d142069b10dc0dd579eaf40a21abb1a75d95805a --- /dev/null +++ b/2024/SUMix_ Mixup with Semantic and Uncertain Information/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d50495c03a076c4faf4fff00d31a73670decee0c5e1719a6c599c848f692ce10 +size 622415 diff --git a/2024/SUMix_ Mixup with Semantic and Uncertain Information/layout.json b/2024/SUMix_ Mixup with Semantic and Uncertain Information/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..833024556adaf7719fd1aa7ac9c6fdc6dc6ba803 --- /dev/null +++ b/2024/SUMix_ Mixup with Semantic and Uncertain Information/layout.json @@ -0,0 +1,10954 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 149, + 111, + 465, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 111, + 465, + 146 + ], + "spans": [ + { + "bbox": [ + 149, + 111, + 465, + 146 + ], + "type": "text", + "content": "SUMix: Mixup with Semantic and Uncertain Information" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 141, + 167, + 471, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 167, + 471, + 193 + ], + "spans": [ + { + "bbox": [ + 141, + 167, + 471, + 193 + ], + "type": "text", + "content": "Huafeng Qin" + }, + { + "bbox": [ + 141, + 167, + 471, + 193 + ], + "type": "inline_equation", + "content": "^{1,2,5,\\star,\\boxtimes}" + }, + { + "bbox": [ + 141, + 167, + 471, + 193 + ], + "type": "text", + "content": ", Xin Jin" + }, + { + "bbox": [ + 141, + 167, + 471, + 193 + ], + "type": "inline_equation", + "content": "^{1,2,\\star}" + }, + { + "bbox": [ + 141, + 167, + 471, + 193 + ], + "type": "text", + "content": ", Hongyu Zhu" + }, + { + "bbox": [ + 141, + 167, + 471, + 193 + ], + "type": "inline_equation", + "content": "^{1,2,\\star}" + }, + { + "bbox": [ + 141, + 167, + 471, + 193 + ], + "type": "text", + "content": ", Hongchao Liao" + }, + { + "bbox": [ + 141, + 167, + 471, + 193 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 141, + 167, + 471, + 193 + ], + "type": "text", + "content": ", Mounim A. El-Yacoubi" + }, + { + "bbox": [ + 141, + 167, + 471, + 193 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 141, + 167, + 471, + 193 + ], + "type": "text", + "content": ", Xinbo Gao" + }, + { + "bbox": [ + 141, + 167, + 471, + 193 + ], + "type": "inline_equation", + "content": "^{4}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 140, + 201, + 465, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 201, + 465, + 213 + ], + "spans": [ + { + "bbox": [ + 140, + 201, + 465, + 213 + ], + "type": "text", + "content": "1 Chongqing Technology and Business University, Chongqing, 400067, China" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 140, + 213, + 472, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 213, + 472, + 224 + ], + "spans": [ + { + "bbox": [ + 140, + 213, + 472, + 224 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 140, + 213, + 472, + 224 + ], + "type": "text", + "content": " National Research base of Intelligent Manufacturing Service, Chongqing, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 156, + 224, + 457, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 224, + 457, + 235 + ], + "spans": [ + { + "bbox": [ + 156, + 224, + 457, + 235 + ], + "type": "text", + "content": "3 Telecom SudParis, Institut Polytechnique de Paris, Paris, 75020, France" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 235, + 480, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 235, + 480, + 246 + ], + "spans": [ + { + "bbox": [ + 133, + 235, + 480, + 246 + ], + "type": "text", + "content": "4 Chongqing University of Posts and Telecommunications, Chongqing, 400065, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 201, + 246, + 411, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 246, + 411, + 257 + ], + "spans": [ + { + "bbox": [ + 201, + 246, + 411, + 257 + ], + "type": "text", + "content": "5 Chongqing Micro-Vein Intelligent Technology Co." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 209, + 258, + 404, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 258, + 404, + 268 + ], + "spans": [ + { + "bbox": [ + 209, + 258, + 404, + 268 + ], + "type": "text", + "content": "{2013009, jinxin4, zhuhongyu}@ctbu.edu.cn" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 209, + 269, + 403, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 269, + 403, + 278 + ], + "spans": [ + { + "bbox": [ + 209, + 269, + 403, + 278 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 209, + 269, + 403, + 278 + ], + "type": "text", + "content": " Equal contribution, " + }, + { + "bbox": [ + 209, + 269, + 403, + 278 + ], + "type": "inline_equation", + "content": "\\boxtimes" + }, + { + "bbox": [ + 209, + 269, + 403, + 278 + ], + "type": "text", + "content": " Corresponding author." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 159, + 307, + 455, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 307, + 455, + 517 + ], + "spans": [ + { + "bbox": [ + 159, + 307, + 455, + 517 + ], + "type": "text", + "content": "Abstract. Mixup data augmentation approaches have been applied for various tasks of deep learning to improve the generalization ability of deep neural networks. Some existing approaches CutMix, SaliencyMix, etc. randomly replace a patch in one image with patches from another to generate the mixed image. Similarly, the corresponding labels are linearly combined by a fixed ratio " + }, + { + "bbox": [ + 159, + 307, + 455, + 517 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 159, + 307, + 455, + 517 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 159, + 307, + 455, + 517 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 159, + 307, + 455, + 517 + ], + "type": "text", + "content": ". The objects in two images may be overlapped during the mixing process, so some semantic information is corrupted in the mixed samples. In this case, the mixed image does not match the mixed label information. Besides, such a label may mislead the deep learning model training, which results in poor performance. To solve this problem, we proposed a novel approach named SUMix to learn the mixing ratio as well as the uncertainty for the mixed samples during the training process. First, we design a learnable similarity function to compute an accurate mix ratio. Second, an approach is investigated as a regularized term to model the uncertainty of the mixed samples. We conduct experiments on five image benchmarks, and extensive experimental results imply that our method is capable of improving the performance of classifiers with different cutting-based mixup approaches. The source code is available at https://github.com/JinXins/SUMix." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 160, + 526, + 407, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 526, + 407, + 537 + ], + "spans": [ + { + "bbox": [ + 160, + 526, + 407, + 537 + ], + "type": "text", + "content": "Keywords: mixup, data augmentation, image classification" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 557, + 230, + 569 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 230, + 569 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 230, + 569 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Deep learning has emerged as a viable alternative to traditional machine learning in various tasks, e.g. computer vision [3, 29, 34], natural language processing [10, 43, 62], and video applications [12, 37, 59], because of its robust feature representation capacity. However, deep learning requires a lot of data to train thousands of network parameters. Despite the self-supervised learning paradigm [18, 31, 33, 36] enables task agnostic pre-training without annotations, the labeled data of specific scenarios is still required and generally limited, where" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 164, + 114, + 452, + 205 + ], + "blocks": [ + { + "bbox": [ + 164, + 114, + 452, + 205 + ], + "lines": [ + { + "bbox": [ + 164, + 114, + 452, + 205 + ], + "spans": [ + { + "bbox": [ + 164, + 114, + 452, + 205 + ], + "type": "image", + "image_path": "35abdef1888190e637e8707039dc0671db77cfcbe317f6ba2a131988cdc5ab2f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 213, + 482, + 257 + ], + "lines": [ + { + "bbox": [ + 130, + 213, + 482, + 257 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 482, + 257 + ], + "type": "text", + "content": "Fig. 1: The diagram shows different cases of raw samples that underwent the CutMix with a mixing ratio of 0.5 to obtain mixed samples. Left: the raw samples and their one-hot labels; Right: labels of the mixes obtained using a redefined mixing ratio " + }, + { + "bbox": [ + 130, + 213, + 482, + 257 + ], + "type": "inline_equation", + "content": "\\widetilde{\\lambda}" + }, + { + "bbox": [ + 130, + 213, + 482, + 257 + ], + "type": "text", + "content": " for different cases." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 281, + 480, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 480, + 317 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 480, + 317 + ], + "type": "text", + "content": "deep neural networks (DNNs) are prone to overfitting [2]. Therefore, the representation capacity is not effectively exploited, resulting in poor discriminative performances and generalization abilities." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 318, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 318, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 318, + 482, + 545 + ], + "type": "text", + "content": "Data Augmentation (DA) methods as a solution have been proposed to prevent DNNs from over-fitting. Existing methods transform original data to generate new samples for data augmentation. For example, in computer vision tasks, classical techniques such as rotating, scaling, and panning have been used to generate augmented samples for model training. Similarly, in natural language processing tasks, methods such as proximate word substitution and randomly misspelling words are employed to produce new data. These data augmentation methods have proven to be effective in improving the performance of network models. They allow the models to learn a larger variety of samples based on augmentation data. Currently, data augmentation methods can be categorized into three classes. (i) The contrast combination methods [8, 9, 17, 38, 39, 58] employs affine matrix transformations and color enhancement for data augmentation. However, the resulting images are significantly different from the original ones. (ii) Mixup method produces virtual mixup examples via a simple convex combination of pairs of examples [51, 64, 75]. (iii) The generative approaches, e.g. Generative Adversarial Networks(GANs) [13, 71], generate fake samples by adversarial learning. Overall, these data augmentation methods effectively alleviate the overfitting problem and improve the performance of deep learning models on limited training data." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Mixup [75] method as a recent DA model produces virtual samples and labels by linearly interpolating the sample images as well as their corresponding labels, and such a simple method has proven to be effective in improving the representation capacity of the deep neural networks. In recent years, various mixup methods have been proposed for data augmentation. For example, CutMix [72] shifts the numerical mixing to spatial, where the mixing coefficient " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " determines the size of the area of the cut and introduces the idea of Cutout [11], which cuts and mixes two samples, as shown in Figure 1. This approach not only helps the network to generalize better and improve object localization but also enables continuous performance improvement in tasks such as Pascal VOC" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 209, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 209, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 209, + 102 + ], + "type": "text", + "content": "Qin. et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 114, + 475, + 182 + ], + "blocks": [ + { + "bbox": [ + 138, + 114, + 475, + 182 + ], + "lines": [ + { + "bbox": [ + 138, + 114, + 475, + 182 + ], + "spans": [ + { + "bbox": [ + 138, + 114, + 475, + 182 + ], + "type": "image", + "image_path": "dc0a669dcab3376c350dd9d1aea2a1e0c23ba05ce62fde6ffcc18767295ca3bd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 192, + 480, + 225 + ], + "lines": [ + { + "bbox": [ + 130, + 192, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 130, + 192, + 480, + 225 + ], + "type": "text", + "content": "Fig. 2: The figure shows some hand-crafted mixup methods with \"Label Mismatch\" problem. In CutMix and ResizeMix, the \"Shark\" is mostly obscured, while in SaliencyMix and FMix, it is difficult to notice the \"Panda\"." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 253, + 480, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 253, + 480, + 419 + ], + "spans": [ + { + "bbox": [ + 130, + 253, + 480, + 419 + ], + "type": "text", + "content": "object detection [53] and MS-COCO image captioning [65]. In addition, CutMix improves the robustness of the model, making it perform better in the face of input corruption and cross-distribution detection. However, the CutMix method randomly selects the cropping region, which results in the background region of the target sample being clipped to the feature region of the source sample, and in the extreme case, the source sample will be completely occluded, resulting in an augmented sample that has no feature information at all, a phenomenon known as \"Label MisMatch\" shown in Figure.2. Since the one-hot label of the sample is unchanged, Decoupled Mix [42] found that the Mixup method relies heavily on the loss function, forcing the prediction of the network model to obey two higher peaks, which results in an inaccurate gradient result due to the label mismatch problem in the calculation of the loss function. This leads to an inaccurate gradient result when calculating the loss function, which affects the training of the network model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 422, + 480, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 480, + 494 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 480, + 494 + ], + "type": "text", + "content": "To improve the \"Label MisMatch\" problem, we propose SUMix, which consists of a mix ratio learning module and an uncertain estimation module. The former focuses on computing the proportion of two images and the latter aims to learn the uncertainty information of mixed samples. Firstly, we design a function to compute the semantic distance between the mixed and original samples to determine the ratio " + }, + { + "bbox": [ + 130, + 422, + 480, + 494 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 422, + 480, + 494 + ], + "type": "text", + "content": ". Secondly, we present a method to learn the uncertainty" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 138, + 521, + 475, + 619 + ], + "blocks": [ + { + "bbox": [ + 138, + 521, + 475, + 619 + ], + "lines": [ + { + "bbox": [ + 138, + 521, + 475, + 619 + ], + "spans": [ + { + "bbox": [ + 138, + 521, + 475, + 619 + ], + "type": "image", + "image_path": "a701f0ea1cfa235cd773f0447d6478048348d22696aaec16d632b5c317304116.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 628, + 480, + 660 + ], + "lines": [ + { + "bbox": [ + 130, + 628, + 480, + 660 + ], + "spans": [ + { + "bbox": [ + 130, + 628, + 480, + 660 + ], + "type": "text", + "content": "Fig. 3: The results illustrate the application of SUMix. Left: Top-1 Accuracy improvement from the mixup approaches with SUMix; Right: Comparison of Vanilla method, CutMix, and SUMix for CAM visualization." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "type": "text", + "content": "SUMix" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "of the mixed. This adapted feature vector effectively mitigates issues related to computing the loss function caused by discrepancies in semantic and uncertainty aspects. Extensive experiments on five image classification datasets verify that our proposed SUMix can remarkably improve performances of existing mixup augmentations (as shown in Figure 3) in a plug-and-play manner while achieving better robustness. Our main contributions are as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 197, + 479, + 281 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 132, + 197, + 479, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 197, + 479, + 220 + ], + "spans": [ + { + "bbox": [ + 132, + 197, + 479, + 220 + ], + "type": "text", + "content": "(a) We propose a learnable metric to compute the mixed ratio by similarity between the mixed samples and the original samples." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 221, + 479, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 221, + 479, + 257 + ], + "spans": [ + { + "bbox": [ + 132, + 221, + 479, + 257 + ], + "type": "text", + "content": "(b) We further consider the uncertainty and semantic information of the mixed samples and recalculate a reasonable feature vector, providing an additional regularization loss for model training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 258, + 479, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 258, + 479, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 258, + 479, + 281 + ], + "type": "text", + "content": "(c) Our approach helps popular Cutting-based mixup methods to improve classification tasks without spending too excessive extra time overhead." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 303, + 237, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 303, + 237, + 316 + ], + "spans": [ + { + "bbox": [ + 132, + 303, + 237, + 316 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 328, + 482, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 328, + 482, + 570 + ], + "spans": [ + { + "bbox": [ + 130, + 328, + 482, + 570 + ], + "type": "text", + "content": "Sample-based Mixup Sample-based Mixup approaches [21-23, 30, 70] focus on obtaining more reliable mixed samples or hidden space feature maps. Usually combined with inserting prior knowledge or saliency information [54]. CutMix [72] obtains the mixed samples by randomly cutting and pasting between the two samples. Due to the large randomness of CutMix cropping, ResizeMix [52] replaces the cut operation with the resize operation. There are a series of works such as FMix [16], GridMix [1] etc, which use similar methods. To extract the saliency information within the samples during the training of the model, some work such as SaliencyMix [61], Attentive-CutMix [67] etc, use a saliency extraction module to mix the samples in a guidedness. PuzzleMix [26] extracts the saliency regions from the source and target samples respectively through the saliency capability learned from the model and then extracts the saliency regions according to the designed \"optimal transmission\". Co-Mixup [27] further improves PuzzleMix by increasing the number of mixed samples from 2 to 3. AutoMix [44] and SAMix [32] analyze the advantages and disadvantages of the hand-crafted and saliency method and divide the mixing task into two sub-tasks. AdAutoMix [51] further improves AutoMix by proposing adversarial training to train the generator and obtain \"hard mixed sample\" to train the model. Manifold Mixup [64] and AlignMix [63] moved the mixup method to hidden space by mixing the feature information of the samples." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "Label-based Mixup Label-based Mixup approaches aim to solve the problems of 1) time overhead and 2) sample-label mismatch. These methods usually use some known information to recalculate the mix ratio of labels. Saliency Grafting [50] uses the area of saliency features of each sample in a mixed sample to reconstruct a new " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": ". LUMix [57] differs from previous ways to address the label noise introduced by the mixup process instead of modeling the label noise. Methods such as TransMix [5] and TokenMix [40] use attention scores to mix the" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 102 + ], + "type": "text", + "content": "Qin. et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "samples to compute " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " based on the area of the mask. Mixpro [76] obtains a new " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " by combining the attention score with the masked area and obtains a stable output with a factor " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": ". Other methods like SMMix [6], Token Labeling [24] etc, are based on attention score and masking area to get a mix ratio." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 172, + 480, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 172, + 480, + 301 + ], + "spans": [ + { + "bbox": [ + 130, + 172, + 480, + 301 + ], + "type": "text", + "content": "Uncertainty Modeling The presence of naturally occurring uncertainties such as occlusion [25] and blurring [55] in the images tends to influence the training of network models. Several works [4,56,74] have attempted to model uncertainty in face recognition, age estimation, and point cloud segmentation, respectively, to obtain better robustness and generalization. In addition, it is also able to deal with Out-distribution problems. [15] proposed a Bayesian classifier for obtaining Out-of-distribution uncertainty. [49] defines a Gaussian distribution to represent the samples and uses Monte Carlo sampling to adopt multiple points from the Gaussian distribution for optimizing the metric. [68] uses an introspective modeling approach that bypasses distribution optimization to further improve the robustness of the model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 318, + 233, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 318, + 233, + 331 + ], + "spans": [ + { + "bbox": [ + 132, + 318, + 233, + 331 + ], + "type": "text", + "content": "3 Preliminaries" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 339, + 479, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 339, + 479, + 375 + ], + "spans": [ + { + "bbox": [ + 130, + 339, + 479, + 375 + ], + "type": "text", + "content": "MixUp Classification. Let us define an " + }, + { + "bbox": [ + 130, + 339, + 479, + 375 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{X}" + }, + { + "bbox": [ + 130, + 339, + 479, + 375 + ], + "type": "text", + "content": " as the input data and its corresponding " + }, + { + "bbox": [ + 130, + 339, + 479, + 375 + ], + "type": "inline_equation", + "content": "Y \\in \\mathbb{Y}" + }, + { + "bbox": [ + 130, + 339, + 479, + 375 + ], + "type": "text", + "content": " as the label of the input data. For the vanilla mixup, we construct a mixed sample " + }, + { + "bbox": [ + 130, + 339, + 479, + 375 + ], + "type": "inline_equation", + "content": "\\widetilde{X}" + }, + { + "bbox": [ + 130, + 339, + 479, + 375 + ], + "type": "text", + "content": " and a mixed label " + }, + { + "bbox": [ + 130, + 339, + 479, + 375 + ], + "type": "inline_equation", + "content": "\\widetilde{Y}" + }, + { + "bbox": [ + 130, + 339, + 479, + 375 + ], + "type": "text", + "content": " from Eq.(1)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 247, + 380, + 480, + 411 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 380, + 480, + 411 + ], + "spans": [ + { + "bbox": [ + 247, + 380, + 480, + 411 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\begin{array}{l} \\widetilde {X} = \\lambda * X _ {a} + (1 - \\lambda) * X _ {b}, \\\\ \\widetilde {\\tilde {\\Sigma}} = \\dots \\end{array} \\tag {1} \\\\ \\widetilde {Y} = \\lambda * Y _ {a} + (1 - \\lambda) * Y _ {b}, \\\\ \\end{array}", + "image_path": "d9e27d13abbd1e5d8e788fd57487f3f353810d3e46072a0928debd409bc1856f.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 418, + 479, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 479, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 479, + 491 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 418, + 479, + 491 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 418, + 479, + 491 + ], + "type": "text", + "content": " is a mix ratio obtained from a Beta distribution, " + }, + { + "bbox": [ + 130, + 418, + 479, + 491 + ], + "type": "inline_equation", + "content": "(X_{a},X_{b})\\in \\widetilde{X}" + }, + { + "bbox": [ + 130, + 418, + 479, + 491 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 418, + 479, + 491 + ], + "type": "inline_equation", + "content": "(Y_{a},Y_{b})\\in \\widetilde{Y}" + }, + { + "bbox": [ + 130, + 418, + 479, + 491 + ], + "type": "text", + "content": " are the same batch of data. Mixup obtains the mixed samples by pixel interpolating with a " + }, + { + "bbox": [ + 130, + 418, + 479, + 491 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 418, + 479, + 491 + ], + "type": "text", + "content": ", CutMix randomly obtains a two-valued rectangular mask " + }, + { + "bbox": [ + 130, + 418, + 479, + 491 + ], + "type": "inline_equation", + "content": "\\mathbb{1}" + }, + { + "bbox": [ + 130, + 418, + 479, + 491 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 418, + 479, + 491 + ], + "type": "inline_equation", + "content": "\\mathbb{1} = \\lambda = \\frac{r_w r_h}{WH}" + }, + { + "bbox": [ + 130, + 418, + 479, + 491 + ], + "type": "text", + "content": ", W, H are the width and height of the data and " + }, + { + "bbox": [ + 130, + 418, + 479, + 491 + ], + "type": "inline_equation", + "content": "r_w, r_h" + }, + { + "bbox": [ + 130, + 418, + 479, + 491 + ], + "type": "text", + "content": " are the width and height of the rectangular mask. We get the mixed samples " + }, + { + "bbox": [ + 130, + 418, + 479, + 491 + ], + "type": "inline_equation", + "content": "\\widetilde{X} = \\mathbb{1}\\odot X_a + (1 - \\mathbb{1})\\odot X_b" + }, + { + "bbox": [ + 130, + 418, + 479, + 491 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 499, + 479, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 499, + 479, + 535 + ], + "spans": [ + { + "bbox": [ + 130, + 499, + 479, + 535 + ], + "type": "text", + "content": "The Devil Lies in Label MisMatch. Mixup contains two different labels, so the Mixed Cross Entropy (MCE) is used in calculating the loss function as shown in Eq.(2)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 185, + 540, + 480, + 554 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 185, + 540, + 480, + 554 + ], + "spans": [ + { + "bbox": [ + 185, + 540, + 480, + 554 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {M C E} = \\ell_ {C E} \\left(f _ {\\theta} (\\tilde {X}), Y _ {a}\\right) * \\lambda + \\ell_ {C E} \\left(f _ {\\theta} (\\tilde {X}), Y _ {b}\\right) * (1 - \\lambda), \\tag {2}", + "image_path": "65c1599423bcdc86ccb29e03ee0e99f956dd4aff009d4806e35049bb6ea5a45b.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 558, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 480, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 558, + 480, + 665 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 130, + 558, + 480, + 665 + ], + "type": "text", + "content": " denotes the network model, and " + }, + { + "bbox": [ + 130, + 558, + 480, + 665 + ], + "type": "inline_equation", + "content": "\\ell(\\cdot)" + }, + { + "bbox": [ + 130, + 558, + 480, + 665 + ], + "type": "text", + "content": " denotes the loss function, " + }, + { + "bbox": [ + 130, + 558, + 480, + 665 + ], + "type": "inline_equation", + "content": "\\ell_{CE}" + }, + { + "bbox": [ + 130, + 558, + 480, + 665 + ], + "type": "text", + "content": " was cross-entropy loss, which minimizes the loss during the training process. However, since some hand-crafted mixup methods are random, e.g. CutMix, ResizeMix, but the labels of the mixed samples are fixed " + }, + { + "bbox": [ + 130, + 558, + 480, + 665 + ], + "type": "inline_equation", + "content": "\\widetilde{Y}" + }, + { + "bbox": [ + 130, + 558, + 480, + 665 + ], + "type": "text", + "content": ", then in the case shown in Figure.2, the features of a certain classification are occluded and reduced or do not exist in the mixed samples, resulting in the feature information not matching the labels. The phenomenon is called \"Label MisMatch\". In this case, the proportion of labels should be reduced or removed to get an accurate result when calculating the loss function." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 417, + 90, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 417, + 90, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 417, + 90, + 447, + 100 + ], + "type": "text", + "content": "SUMix" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 116, + 339, + 216 + ], + "blocks": [ + { + "bbox": [ + 136, + 116, + 339, + 216 + ], + "lines": [ + { + "bbox": [ + 136, + 116, + 339, + 216 + ], + "spans": [ + { + "bbox": [ + 136, + 116, + 339, + 216 + ], + "type": "image", + "image_path": "7476f3d50f69c94437f390b9e1e879674e6f99f935bc9978630523c1dd76ee58.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 227, + 482, + 273 + ], + "lines": [ + { + "bbox": [ + 130, + 227, + 482, + 273 + ], + "spans": [ + { + "bbox": [ + 130, + 227, + 482, + 273 + ], + "type": "text", + "content": "Fig. 4: The left diagram represents the training process of SUMix. The raw and mixed samples are encoded through the model to obtain the semantic and uncertainty information, and the new mix ratio and loss function for regularization are obtained. The right diagram is the spatial distance between the raw and mixed samples." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 345, + 118, + 478, + 213 + ], + "blocks": [ + { + "bbox": [ + 345, + 118, + 478, + 213 + ], + "lines": [ + { + "bbox": [ + 345, + 118, + 478, + 213 + ], + "spans": [ + { + "bbox": [ + 345, + 118, + 478, + 213 + ], + "type": "image", + "image_path": "dc9a0917fcc158fc6b2342bfd2f78f1b3a87f477f4aa4261082715c551267fca.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 297, + 197, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 297, + 197, + 309 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 197, + 309 + ], + "type": "text", + "content": "4 SUMix" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 328, + 482, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 328, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 328, + 482, + 413 + ], + "type": "text", + "content": "In this section, we focus on detailing SUMix, 1) the uncertainty classifier, 2) how to obtain the mix ratio " + }, + { + "bbox": [ + 130, + 328, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 328, + 482, + 413 + ], + "type": "text", + "content": ", and 3) the regularization function in the case of uncertainty and semantic information. Firstly, we show how to use deep learning classifiers for loss function computation and network model training, secondly, how to redefine the mix ratio " + }, + { + "bbox": [ + 130, + 328, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 328, + 482, + 413 + ], + "type": "text", + "content": " based on semantic information, and lastly, how to consider uncertainty and regularization. The pipeline of SUMix is shown in Fig. 4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 436, + 270, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 436, + 270, + 449 + ], + "spans": [ + { + "bbox": [ + 132, + 436, + 270, + 449 + ], + "type": "text", + "content": "4.1 Uncertainty Classifier" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 463, + 482, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 463, + 482, + 548 + ], + "spans": [ + { + "bbox": [ + 130, + 463, + 482, + 548 + ], + "type": "text", + "content": "Assuming " + }, + { + "bbox": [ + 130, + 463, + 482, + 548 + ], + "type": "inline_equation", + "content": "\\mathbb{S} = \\{x_s|s = 1,2,\\dots ,S\\}" + }, + { + "bbox": [ + 130, + 463, + 482, + 548 + ], + "type": "text", + "content": " is a training set and " + }, + { + "bbox": [ + 130, + 463, + 482, + 548 + ], + "type": "inline_equation", + "content": "\\mathbb{S}" + }, + { + "bbox": [ + 130, + 463, + 482, + 548 + ], + "type": "text", + "content": " is the number of training images. We take " + }, + { + "bbox": [ + 130, + 463, + 482, + 548 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 463, + 482, + 548 + ], + "type": "text", + "content": " samples from it as a sample set, where " + }, + { + "bbox": [ + 130, + 463, + 482, + 548 + ], + "type": "inline_equation", + "content": "\\mathbb{X}\\in \\mathbb{R}^{N\\times 3\\times H\\times W}" + }, + { + "bbox": [ + 130, + 463, + 482, + 548 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 463, + 482, + 548 + ], + "type": "inline_equation", + "content": "\\mathbb{Y}\\in \\mathbb{R}^{N\\times K}" + }, + { + "bbox": [ + 130, + 463, + 482, + 548 + ], + "type": "text", + "content": " are the labels and " + }, + { + "bbox": [ + 130, + 463, + 482, + 548 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 463, + 482, + 548 + ], + "type": "text", + "content": " is the number of classes. For the original computation of the mixup loss function, we use a deep neural network " + }, + { + "bbox": [ + 130, + 463, + 482, + 548 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 130, + 463, + 482, + 548 + ], + "type": "text", + "content": " with a trainable parameter that maps the sample " + }, + { + "bbox": [ + 130, + 463, + 482, + 548 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 130, + 463, + 482, + 548 + ], + "type": "text", + "content": " to the labels " + }, + { + "bbox": [ + 130, + 463, + 482, + 548 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 130, + 463, + 482, + 548 + ], + "type": "text", + "content": ". Unlike the vanilla mixup loss computation in Eq.(2), SUMix adds a regularized loss function as shown in Eq.(3):" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 559, + 481, + 593 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 559, + 481, + 593 + ], + "spans": [ + { + "bbox": [ + 141, + 559, + 481, + 593 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s u} = \\frac {1}{N} \\left(\\sum_ {i = 0} ^ {N} \\mathcal {L} _ {M C E} \\left(f _ {\\theta} \\left(\\widetilde {X} _ {i}\\right), Y, \\widetilde {\\lambda} _ {i}\\right)\\right) + \\zeta * \\mathcal {L} _ {M C E} \\left(S U \\left(f _ {\\theta} (\\widetilde {X}), U _ {\\omega}\\right), \\widetilde {Y}\\right), \\tag {3}", + "image_path": "b4179e87a8d03801e52b8e0077f59393cf3378dbd19f5ac4081f1a02fed99e97.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 604, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 604, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 604, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 604, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\widetilde{\\lambda}_i" + }, + { + "bbox": [ + 130, + 604, + 482, + 666 + ], + "type": "text", + "content": " are the redefined lambda, " + }, + { + "bbox": [ + 130, + 604, + 482, + 666 + ], + "type": "inline_equation", + "content": "SU(\\cdot)" + }, + { + "bbox": [ + 130, + 604, + 482, + 666 + ], + "type": "text", + "content": " denotes a function to obtain feature, and " + }, + { + "bbox": [ + 130, + 604, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\zeta" + }, + { + "bbox": [ + 130, + 604, + 482, + 666 + ], + "type": "text", + "content": " is a hyperparameter used to weight the regulated loss function, set to [0.1, 0.2, 0.5, 1.0]. " + }, + { + "bbox": [ + 130, + 604, + 482, + 666 + ], + "type": "inline_equation", + "content": "U_{\\omega}" + }, + { + "bbox": [ + 130, + 604, + 482, + 666 + ], + "type": "text", + "content": " denotes a function as a measure of uncertainty. Finally, the parameters of the deep neural network are updated continuously to minimize " + }, + { + "bbox": [ + 130, + 604, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{su}" + }, + { + "bbox": [ + 130, + 604, + 482, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 209, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 209, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 209, + 101 + ], + "type": "text", + "content": "Qin. et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 257, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 257, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 257, + 128 + ], + "type": "text", + "content": "4.2 Mix ratio Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 257 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 257 + ], + "type": "text", + "content": "Mixup approaches always depend on a parameter taken from the " + }, + { + "bbox": [ + 130, + 138, + 482, + 257 + ], + "type": "inline_equation", + "content": "Beta(\\alpha, \\alpha)" + }, + { + "bbox": [ + 130, + 138, + 482, + 257 + ], + "type": "text", + "content": " distribution, which we call the mix ratio " + }, + { + "bbox": [ + 130, + 138, + 482, + 257 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 138, + 482, + 257 + ], + "type": "text", + "content": ", to balance the mix ratio of the two samples. Lots of mixup approaches use " + }, + { + "bbox": [ + 130, + 138, + 482, + 257 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 138, + 482, + 257 + ], + "type": "text", + "content": " to augment the mixup sample set " + }, + { + "bbox": [ + 130, + 138, + 482, + 257 + ], + "type": "inline_equation", + "content": "\\mathbb{X}" + }, + { + "bbox": [ + 130, + 138, + 482, + 257 + ], + "type": "text", + "content": " after obtaining it, and notably, this " + }, + { + "bbox": [ + 130, + 138, + 482, + 257 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 138, + 482, + 257 + ], + "type": "text", + "content": " serves the whole sample set. For some saliency-based or automatic-based mixups, e.g. PuzzleMix [26], AutoMix [44], it is possible to push the mixing method to obtain the best possible mixed samples according to the " + }, + { + "bbox": [ + 130, + 138, + 482, + 257 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 138, + 482, + 257 + ], + "type": "text", + "content": " ratio, balancing the sample ratio with the " + }, + { + "bbox": [ + 130, + 138, + 482, + 257 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 138, + 482, + 257 + ], + "type": "text", + "content": " as much as possible. However, some hand-crafted approaches with randomness cannot accurately match the mix ratio. The problem we want to solve is how to match each mixup sample with its own " + }, + { + "bbox": [ + 130, + 138, + 482, + 257 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 138, + 482, + 257 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 258, + 482, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 482, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 482, + 326 + ], + "type": "text", + "content": "For a mixed sample " + }, + { + "bbox": [ + 130, + 258, + 482, + 326 + ], + "type": "inline_equation", + "content": "\\widetilde{x} = \\left[\\widetilde{X}\\right]_i^N\\in R^{3\\times H\\times W}" + }, + { + "bbox": [ + 130, + 258, + 482, + 326 + ], + "type": "text", + "content": ", semantic information links need to be established with the original sample pair " + }, + { + "bbox": [ + 130, + 258, + 482, + 326 + ], + "type": "inline_equation", + "content": "(x_{a},x_{b})" + }, + { + "bbox": [ + 130, + 258, + 482, + 326 + ], + "type": "text", + "content": ". They are fed into a deep neural network " + }, + { + "bbox": [ + 130, + 258, + 482, + 326 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 130, + 258, + 482, + 326 + ], + "type": "text", + "content": ", e.g. ResNet18 [19]. feature vectors " + }, + { + "bbox": [ + 130, + 258, + 482, + 326 + ], + "type": "inline_equation", + "content": "\\widetilde{z}" + }, + { + "bbox": [ + 130, + 258, + 482, + 326 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 258, + 482, + 326 + ], + "type": "inline_equation", + "content": "(z_{a},z_{b})" + }, + { + "bbox": [ + 130, + 258, + 482, + 326 + ], + "type": "text", + "content": " are obtained, and the feature vectors are normalized to modify their similarity according to Eq.(4):" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 202, + 335, + 480, + 362 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 335, + 480, + 362 + ], + "spans": [ + { + "bbox": [ + 202, + 335, + 480, + 362 + ], + "type": "interline_equation", + "content": "\\widetilde {\\lambda} _ {a} = \\frac {\\lambda * \\mathrm {e} ^ {- \\| (\\sigma (\\widetilde {z} - z _ {a}) \\| _ {2}}}{\\lambda * \\mathrm {e} ^ {- \\| (\\sigma (\\widetilde {z} - z _ {a}) \\| _ {2}} + (1 - \\lambda) * \\mathrm {e} ^ {- \\| (\\sigma (\\widetilde {z} - z _ {b}) \\| _ {2}}}, \\tag {4}", + "image_path": "e87373e033d7c60cebe77e636cba048c160db0f99038c9eb6748f1fc659cee56.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 372, + 482, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 372, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 372, + 482, + 468 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 372, + 482, + 468 + ], + "type": "inline_equation", + "content": "\\sigma(\\cdot)" + }, + { + "bbox": [ + 130, + 372, + 482, + 468 + ], + "type": "text", + "content": " is the softmax function used to normalize the vectors, and " + }, + { + "bbox": [ + 130, + 372, + 482, + 468 + ], + "type": "inline_equation", + "content": "\\|\\cdot\\|_2" + }, + { + "bbox": [ + 130, + 372, + 482, + 468 + ], + "type": "text", + "content": " is the l2 norm, which imposes a limit on the distance between the two. When the difference in semantic information between " + }, + { + "bbox": [ + 130, + 372, + 482, + 468 + ], + "type": "inline_equation", + "content": "\\widetilde{x}" + }, + { + "bbox": [ + 130, + 372, + 482, + 468 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 372, + 482, + 468 + ], + "type": "inline_equation", + "content": "x_a" + }, + { + "bbox": [ + 130, + 372, + 482, + 468 + ], + "type": "text", + "content": " is large, which indicates that the mixed sample has a small amount of the feature information of the raw sample, " + }, + { + "bbox": [ + 130, + 372, + 482, + 468 + ], + "type": "inline_equation", + "content": "\\widetilde{\\lambda}_a" + }, + { + "bbox": [ + 130, + 372, + 482, + 468 + ], + "type": "text", + "content": " tends to 0. Similarly, when the difference in semantic information between the two is small, " + }, + { + "bbox": [ + 130, + 372, + 482, + 468 + ], + "type": "inline_equation", + "content": "\\widetilde{\\lambda}_a" + }, + { + "bbox": [ + 130, + 372, + 482, + 468 + ], + "type": "text", + "content": " tends to 1. " + }, + { + "bbox": [ + 130, + 372, + 482, + 468 + ], + "type": "inline_equation", + "content": "\\widetilde{\\lambda}_b = 1 - \\widetilde{\\lambda}_a" + }, + { + "bbox": [ + 130, + 372, + 482, + 468 + ], + "type": "text", + "content": ". Thus, we have calculated the mixed sample set with its own matching " + }, + { + "bbox": [ + 130, + 372, + 482, + 468 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 372, + 482, + 468 + ], + "type": "text", + "content": ". a more sensible loss is computed according to Eq.(3)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 487, + 279, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 279, + 499 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 279, + 499 + ], + "type": "text", + "content": "4.3 Uncertainty Estimation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 510, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 605 + ], + "type": "text", + "content": "Deep learning tasks involve two types of uncertainty: Epistemic uncertainty and Aleatoric uncertainty. Epistemic uncertainty arises from the model's inherent uncertainty due to insufficient training data, resulting in low confidence in unknown data. In contrast, Aleatoric uncertainty arises from errors in the data itself, which can introduce bias into the dataset. The greater the bias, the greater the Aleatoric uncertainty. In the case of the mixup task, the error in the data occurs when the mixup method is stochastic. Our goal is to address this error and enable the network to adapt accordingly." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "Wang et al. [68] propose an introspective metric that models a new vector for classification based on the semantic information distance between two images and their uncertainty distance. The mixup methods differ from this as they typically incorporate feature information from two samples. To begin, we encode the mixed sample and the raw sample using an encoder to obtain the feature" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 417, + 90, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 417, + 90, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 417, + 90, + 447, + 100 + ], + "type": "text", + "content": "SUMix" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 140 + ], + "type": "text", + "content": "vectors " + }, + { + "bbox": [ + 130, + 116, + 482, + 140 + ], + "type": "inline_equation", + "content": "\\widetilde{z}" + }, + { + "bbox": [ + 130, + 116, + 482, + 140 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 140 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 116, + 482, + 140 + ], + "type": "text", + "content": ", respectively. Next, we calculate the uncertainty by passing " + }, + { + "bbox": [ + 130, + 116, + 482, + 140 + ], + "type": "inline_equation", + "content": "\\widetilde{z}" + }, + { + "bbox": [ + 130, + 116, + 482, + 140 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 140 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 116, + 482, + 140 + ], + "type": "text", + "content": " through " + }, + { + "bbox": [ + 130, + 116, + 482, + 140 + ], + "type": "inline_equation", + "content": "U_{\\omega}" + }, + { + "bbox": [ + 130, + 116, + 482, + 140 + ], + "type": "text", + "content": ", by Eq.(5)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 261, + 152, + 480, + 171 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 152, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 261, + 152, + 480, + 171 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} u = \\parallel M L P (\\sigma (z)) \\| _ {2}, \\\\ \\widetilde {\\gamma} - \\parallel M L P (\\sigma (\\widetilde {z})) \\| \\end{array} \\tag {5}", + "image_path": "29ec0252a70043d73fc286173ec6912c2ff64b74c0120292c86bc635630b5a67.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 187, + 481, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 187, + 481, + 211 + ], + "spans": [ + { + "bbox": [ + 130, + 187, + 481, + 211 + ], + "type": "text", + "content": "where is a linear layer with a softmax function and l2 norm. We define " + }, + { + "bbox": [ + 130, + 187, + 481, + 211 + ], + "type": "inline_equation", + "content": "\\beta = \\widetilde{u} + u" + }, + { + "bbox": [ + 130, + 187, + 481, + 211 + ], + "type": "text", + "content": " as the uncertainty of mixed samples and raw samples." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 213, + 481, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 481, + 273 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 481, + 273 + ], + "type": "text", + "content": "After capturing the uncertainty and semantic information, as demonstrated in the right part of Figure.4, Wang et al. [68] argue that in the feature space, the smaller the radius of a sample, the more aggregated the class is, and the smaller the uncertainty. By combining the semantic information, we reconstruct the mixed samples " + }, + { + "bbox": [ + 130, + 213, + 481, + 273 + ], + "type": "inline_equation", + "content": "\\widetilde{X}" + }, + { + "bbox": [ + 130, + 213, + 481, + 273 + ], + "type": "text", + "content": " and the raw samples " + }, + { + "bbox": [ + 130, + 213, + 481, + 273 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 213, + 481, + 273 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 130, + 213, + 481, + 273 + ], + "type": "inline_equation", + "content": "Z_{su}" + }, + { + "bbox": [ + 130, + 213, + 481, + 273 + ], + "type": "text", + "content": " according to Eq.(6)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 260, + 283, + 480, + 297 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 283, + 480, + 297 + ], + "spans": [ + { + "bbox": [ + 260, + 283, + 480, + 297 + ], + "type": "interline_equation", + "content": "Z _ {s u} = \\mathrm {e} ^ {- (\\beta + \\| \\sigma (\\tilde {z} - z) \\| _ {2})} \\tag {6}", + "image_path": "5729ff2c8aebb59b47d72493a5de27f41b2ac7380e80c26d4f8a2ea5768f000e.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 303, + 481, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 303, + 481, + 352 + ], + "spans": [ + { + "bbox": [ + 130, + 303, + 481, + 352 + ], + "type": "text", + "content": "When there is significant uncertainty or semantic information has a large difference, " + }, + { + "bbox": [ + 130, + 303, + 481, + 352 + ], + "type": "inline_equation", + "content": "Z_{su}" + }, + { + "bbox": [ + 130, + 303, + 481, + 352 + ], + "type": "text", + "content": " receives a small gradient; conversely, when there is little uncertainty and small differences in semantic information, " + }, + { + "bbox": [ + 130, + 303, + 481, + 352 + ], + "type": "inline_equation", + "content": "Z_{su}" + }, + { + "bbox": [ + 130, + 303, + 481, + 352 + ], + "type": "text", + "content": " receives a normal gradient that tends to 1. This approach is used as a method of regularization." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 371, + 230, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 371, + 230, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 371, + 230, + 385 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 397, + 482, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 397, + 482, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 397, + 482, + 518 + ], + "type": "text", + "content": "To verify the performance of SUMix, we conducted a large number of experiments on six datasets, CIFAR100 [28], Tiny-ImageNet [7], ImageNet-1K [29], CUB-200 [66], FGVC-Aircrafts [47], and CIFAR100-C. To verify the effectiveness of the improved \"Label MisMatching\" method, we mainly compared it with some patch-wise mixup methods based on hand-crafted, i.e., CutMix, FMix, ResizeMix, and SaliencyMix. We have implemented our algorithm on the open-source library OpenMixup [35], and some parameters are designed according to OpenMixup. The classification results are the median of the top-1 test accuracy from the last 10 training epochs, but on ImageNet-1K, we choose the last 5 training epochs. To facilitate comparison, our results are highlighted in bold." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 536, + 204, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 204, + 548 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 204, + 548 + ], + "type": "text", + "content": "5.1 Datasets" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "We choose the three most classical datasets and two fine-grained datasets in the classification task. CIFAR100 has a total of 50,000 images for the training set and 10,000 images for the testing set, containing 100 different classes, each with a " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " resolution; Tiny-ImageNet contains 10,000 images for training and 10,000 images for testing, each with a " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " resolution and a total of 200 classes; The ImageNet-1K dataset contains 1.2 million images for training and 50,000 images for validation; CUB200 dataset contains images of 200 different bird species, a total of 11,788 images; the FGVC-Aircrafts dataset contains images of 100 different types of aircraft, a total of 10,000 images." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 209, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 209, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 209, + 102 + ], + "type": "text", + "content": "Qin. et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 279, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 279, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 279, + 128 + ], + "type": "text", + "content": "5.2 Implementation Details" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 136, + 482, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 136, + 482, + 256 + ], + "spans": [ + { + "bbox": [ + 130, + 136, + 482, + 256 + ], + "type": "text", + "content": "Hyper-parameter of " + }, + { + "bbox": [ + 130, + 136, + 482, + 256 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 136, + 482, + 256 + ], + "type": "text", + "content": ". In the experimental results we report, the hyperparameter " + }, + { + "bbox": [ + 130, + 136, + 482, + 256 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 136, + 482, + 256 + ], + "type": "text", + "content": " of the mixup variants, which sampling from the " + }, + { + "bbox": [ + 130, + 136, + 482, + 256 + ], + "type": "inline_equation", + "content": "Beta(\\alpha, \\alpha)" + }, + { + "bbox": [ + 130, + 136, + 482, + 256 + ], + "type": "text", + "content": " distribution used by the model based on the CNN architecture and the model based on the ViT architecture are different. For CutMix, the CNN-based models, e.g. ResNet18, ResNeXt50, and Wide-ResNet28 set as 0.2, and ViT-based models, DeiT and Swin-Transformer set as 2.0. For FMix, CNN-based set 0.2, and ViT-based set 1.0. For PuzzleMix, the CNN-based set is 1.0, and the ViT-based set is 2.0. In CNN-based and Vit-based architectures alike, SaliencyMix, ResizeMix, and AutoMix exhibit identical alpha values, which are 0.2, 1.0, and 2.0, respectively." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 266, + 482, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 266, + 482, + 340 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 482, + 340 + ], + "type": "text", + "content": "Hyper-parameter of " + }, + { + "bbox": [ + 130, + 266, + 482, + 340 + ], + "type": "inline_equation", + "content": "\\zeta" + }, + { + "bbox": [ + 130, + 266, + 482, + 340 + ], + "type": "text", + "content": ". In SUMix, we set the hyper-parameters " + }, + { + "bbox": [ + 130, + 266, + 482, + 340 + ], + "type": "inline_equation", + "content": "\\zeta" + }, + { + "bbox": [ + 130, + 266, + 482, + 340 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 130, + 266, + 482, + 340 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{su}" + }, + { + "bbox": [ + 130, + 266, + 482, + 340 + ], + "type": "text", + "content": " as follows: On the CIFAR100 dataset, we report five different models, the " + }, + { + "bbox": [ + 130, + 266, + 482, + 340 + ], + "type": "inline_equation", + "content": "\\zeta" + }, + { + "bbox": [ + 130, + 266, + 482, + 340 + ], + "type": "text", + "content": " setting shown in Appendix A.1. On Tiny-ImageNet, we set the " + }, + { + "bbox": [ + 130, + 266, + 482, + 340 + ], + "type": "inline_equation", + "content": "\\zeta = 1.0" + }, + { + "bbox": [ + 130, + 266, + 482, + 340 + ], + "type": "text", + "content": " for all mixup methods based on ResNet18 and ResNeXt50. Similarly with ImageNet-1K, we set the " + }, + { + "bbox": [ + 130, + 266, + 482, + 340 + ], + "type": "inline_equation", + "content": "\\zeta = 0.5" + }, + { + "bbox": [ + 130, + 266, + 482, + 340 + ], + "type": "text", + "content": " based on ResNet18. The settings we reported for the fine-grained datasets are also shown in Appendix A.1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 357, + 262, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 357, + 262, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 357, + 262, + 369 + ], + "type": "text", + "content": "5.3 Image Classification" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 376, + 482, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 497 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 497 + ], + "type": "text", + "content": "Generic classification. We initially train ResNet18 [19] and ResNeXt50 [69] using the CIFAR100 dataset for a total of 800 epochs. The experimental parameters are set as in OpenMixup [35]: a basic learning rate of 0.1, SGD [45] optimizer with momentum of 0.9, weight decay of 0.0001, and batch size of 100 for each iteration. Wide-ResNet [73] using the CIFAR100 dataset for a total of 400 epochs, learning rate of 0.03, SGD optimizer with momentum of 0.9, weight decay of 0.001. For the Tiny-ImageNet dataset, we use the same experimental parameters as the CIFAR100 experiment, except the base learning rate to 0.2 and training for 400 epochs. As for ImageNet-1K, we follow the Pytorch-style settings for training ResNet18 of 100 epochs. For the CIFAR100 dataset, we use" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 135, + 549, + 479, + 662 + ], + "blocks": [ + { + "bbox": [ + 132, + 517, + 482, + 540 + ], + "lines": [ + { + "bbox": [ + 132, + 517, + 482, + 540 + ], + "spans": [ + { + "bbox": [ + 132, + 517, + 482, + 540 + ], + "type": "text", + "content": "Table 1: Top-1 accuracy " + }, + { + "bbox": [ + 132, + 517, + 482, + 540 + ], + "type": "inline_equation", + "content": "(\\%)\\uparrow" + }, + { + "bbox": [ + 132, + 517, + 482, + 540 + ], + "type": "text", + "content": " of mixup methods on CIFAR-100, Tiny-ImageNet and ImageNet-1K. * denotes mixup methods with SUMix." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 549, + 479, + 662 + ], + "lines": [ + { + "bbox": [ + 135, + 549, + 479, + 662 + ], + "spans": [ + { + "bbox": [ + 135, + 549, + 479, + 662 + ], + "type": "table", + "html": "
MethodCIFAR100Tiny-ImageNetImageNet-1K
ResNet18ResNeXt50W-ResNet28-8ResNet18ResNeXt50ResNet18
CutMix78.1778.3284.4565.5366.4768.95
FMix79.6979.0284.2163.4765.0869.96
SaliencyMix79.1278.7784.3564.6066.5569.16
ResizeMix80.0180.3584.8763.7465.8769.50
CutMix*79.7879.9184.5665.7168.7469.71
FMix*80.2080.7984.3263.6967.1270.48
SaliencyMix*79.9179.3284.5865.6868.9269.52
ResizeMix*80.3880.7284.9165.3067.4969.76
Avg. Gain+0.82+1.07+0.12+0.81+2.07+0.47
", + "image_path": "88e98f1f3042dfd08fee9de0110cc97e24e9b5beb8553c2d63313f7c76807497.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 417, + 90, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 417, + 90, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 417, + 90, + 447, + 101 + ], + "type": "text", + "content": "SUMix" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 176, + 157, + 433, + 285 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 2: Top-1 accuracy " + }, + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "(\\%)\\uparrow" + }, + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "text", + "content": " of mixup methods on CUB200, FGVC-Aircrafts, and Standford-Cars. * denotes mixup methods SUMix. † denotes that we reproduce the results of that experiment." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 176, + 157, + 433, + 285 + ], + "lines": [ + { + "bbox": [ + 176, + 157, + 433, + 285 + ], + "spans": [ + { + "bbox": [ + 176, + 157, + 433, + 285 + ], + "type": "table", + "html": "
MethodCUB200FGVC-Aircrafts
ResNet18ResNeXt50ResNet18ResNeXt50
CutMix†77.7083.6778.8484.55
FMix77.2884.0679.3684.10
SaliencyMix†75.7782.8379.7884.31
ResizeMix78.5084.1678.1084.08
CutMix*78.2083.7179.7285.84
FMix*79.2484.3379.4884.64
SaliencyMix*76.9882.8479.9084.49
ResizeMix*78.5684.2380.2985.12
Avg. Gain+0.93+0.10+0.82+0.67
", + "image_path": "5c97baa211a7d8df1dafb85a63812fa24eb08d3c240d6934872ea1e5ac0c276c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 134, + 352, + 279, + 434 + ], + "blocks": [ + { + "bbox": [ + 134, + 352, + 279, + 434 + ], + "lines": [ + { + "bbox": [ + 134, + 352, + 279, + 434 + ], + "spans": [ + { + "bbox": [ + 134, + 352, + 279, + 434 + ], + "type": "table", + "html": "
MethodCIFAR100
DeiT-SmallSwin-Tiny
CutMix74.1280.64
FMix70.4180.72
SaliencyMix69.7880.40
ResizeMix68.4580.16
CutMix*75.26+1.1480.83+0.19
FMix*70.69+0.2880.73+0.01
SaliencyMix*70.31+0.5380.71+0.29
ResizeMix*68.78+0.3380.59+0.43
", + "image_path": "7b671bdf8b2bfb6623e5a2684d78e618cd6f0a5a0e9e746f0b53a8719920bff5.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 283, + 351, + 478, + 434 + ], + "blocks": [ + { + "bbox": [ + 130, + 308, + 482, + 342 + ], + "lines": [ + { + "bbox": [ + 130, + 308, + 482, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 482, + 342 + ], + "type": "text", + "content": "Table 3: Top-1 accuracy(\\%)↑ of Table 4: Top-1 accuracy(\\%)↑ of saliency-based mixup methods on CIFAR100. * de- mixup methods on CIFAR100. * denotes mixup notes mixup methods with SUMix. methods with SUMix." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 283, + 351, + 478, + 434 + ], + "lines": [ + { + "bbox": [ + 283, + 351, + 478, + 434 + ], + "spans": [ + { + "bbox": [ + 283, + 351, + 478, + 434 + ], + "type": "table", + "html": "
MethodCIFAR100
ResNet18ResNeXt50
PuzzleMix81.1381.69
AutoMix82.0482.84
PuzzleMix*81.43+0.3082.60+0.91
AutoMix*82.30+0.2683.82+0.98
AdAutoMix82.3283.81
", + "image_path": "97d47370fc91925f037965ea0cabf071a4845e9b2ac499559f01ba0a7a4fc710.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 458, + 481, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 458, + 481, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 458, + 481, + 495 + ], + "type": "text", + "content": "basic data augmentation including RandomFlip and RandomCrop with 4 pixels padding, and for the Tiny-ImageNet and ImageNet-1K datasets, we also use the RandomFlip and RandonCrop data augmentation operations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 495, + 482, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 482, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 482, + 592 + ], + "type": "text", + "content": "Table 1 shows our five comparisons on three different datasets, where the use of SUMix brings some improvement in the classification accuracy of CutMix, FMix, SaliencyMix, and ResizeMix. For the CIFAR100 dataset, on ResNet and its variants ResNeXt50, W-ResNet28-8, SUMix improves their classification results by an average of " + }, + { + "bbox": [ + 130, + 495, + 482, + 592 + ], + "type": "inline_equation", + "content": "0.82\\%" + }, + { + "bbox": [ + 130, + 495, + 482, + 592 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 495, + 482, + 592 + ], + "type": "inline_equation", + "content": "1.07\\%" + }, + { + "bbox": [ + 130, + 495, + 482, + 592 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 495, + 482, + 592 + ], + "type": "inline_equation", + "content": "0.12\\%" + }, + { + "bbox": [ + 130, + 495, + 482, + 592 + ], + "type": "text", + "content": ", respectively; for the Tiny-ImageNet dataset, on ResNet18 and ResNeXt50, the average improvements of " + }, + { + "bbox": [ + 130, + 495, + 482, + 592 + ], + "type": "inline_equation", + "content": "0.81\\%" + }, + { + "bbox": [ + 130, + 495, + 482, + 592 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 495, + 482, + 592 + ], + "type": "inline_equation", + "content": "2.07\\%" + }, + { + "bbox": [ + 130, + 495, + 482, + 592 + ], + "type": "text", + "content": "; for the ImageNet-1K dataset, on ResNet18, the average improvement is " + }, + { + "bbox": [ + 130, + 495, + 482, + 592 + ], + "type": "inline_equation", + "content": "0.47\\%" + }, + { + "bbox": [ + 130, + 495, + 482, + 592 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": "Fine-grained classification. We load the official PyTorch pre-trained models on ImageNet-1k as initialization and fine-tuning pre-trained ResNet18 and ResNeXt50 using SGD optimizer with a learning rate of 0.001, a momentum of 0.9, weight decay of 0.0005 on CUB200, FGVC-Aircrafts, total training 200 epochs. Table 2 presents the comparison results for the bird and aircraft datasets." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 102 + ], + "type": "text", + "content": "Qin. et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 116, + 476, + 237 + ], + "blocks": [ + { + "bbox": [ + 136, + 116, + 476, + 237 + ], + "lines": [ + { + "bbox": [ + 136, + 116, + 476, + 237 + ], + "spans": [ + { + "bbox": [ + 136, + 116, + 476, + 237 + ], + "type": "image", + "image_path": "1bcf8a841b06735580a596c4f537b930fc238fc2f4cbe46957338b4881625bd1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 248, + 480, + 304 + ], + "lines": [ + { + "bbox": [ + 130, + 248, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 480, + 304 + ], + "type": "text", + "content": "Fig. 5: The top of the figure shows a visualization of the sample at " + }, + { + "bbox": [ + 130, + 248, + 480, + 304 + ], + "type": "inline_equation", + "content": "0\\%" + }, + { + "bbox": [ + 130, + 248, + 480, + 304 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 248, + 480, + 304 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 130, + 248, + 480, + 304 + ], + "type": "text", + "content": " occlusion ratio. The light blue curves in the lower four subfigures show the classification accuracy of CutMix, FMix, SaliencyMix, and ResizeMix on ImageNet-1K using ResNet18 for 100 epochs of training, respectively, and the red curve shows the classification accuracy of these methods using SUMix." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 332, + 480, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 332, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 332, + 480, + 380 + ], + "type": "text", + "content": "SUMix improves the classification results for the CUB200 dataset by an average of " + }, + { + "bbox": [ + 130, + 332, + 480, + 380 + ], + "type": "inline_equation", + "content": "0.93\\%" + }, + { + "bbox": [ + 130, + 332, + 480, + 380 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 332, + 480, + 380 + ], + "type": "inline_equation", + "content": "0.10\\%" + }, + { + "bbox": [ + 130, + 332, + 480, + 380 + ], + "type": "text", + "content": " at ResNet18 and ResNeXt50, respectively. The FGVC-Aircrafts dataset improves by an average of " + }, + { + "bbox": [ + 130, + 332, + 480, + 380 + ], + "type": "inline_equation", + "content": "0.82\\%" + }, + { + "bbox": [ + 130, + 332, + 480, + 380 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 332, + 480, + 380 + ], + "type": "inline_equation", + "content": "0.67\\%" + }, + { + "bbox": [ + 130, + 332, + 480, + 380 + ], + "type": "text", + "content": " at ResNet18 and ResNeXt50." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 402, + 480, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 480, + 521 + ], + "type": "text", + "content": "Classification based on ViTs. To further show the performance of SUMix, we perform classification experiments using the ViT-based model DeiT [60] and Swin-Transformer [41] on the CIFAR100 dataset, and the results are shown in Tab 3. We resized the size to " + }, + { + "bbox": [ + 130, + 402, + 480, + 521 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 130, + 402, + 480, + 521 + ], + "type": "text", + "content": ", and in the training set, we trained DeiT and Swin-Transformer of 200 epochs by AdamW [46] optimizer with a batch size of 100 on CIFAR100. The basic learning rates of DeiT and Swin are 1e-3 and 5e-4, and the weight decay of 0.05, like OpenMixup settings, we also used some RandAugment [9] operations. We reported the best top-1 accuracy in the last 10 training epochs, and the results show that our SUMix also can bring effectiveness to ViTs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 523, + 480, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 523, + 480, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 523, + 480, + 559 + ], + "type": "text", + "content": "Table 3 shows the results of the ViT models using SUMix, which improves the classification results by an average of " + }, + { + "bbox": [ + 130, + 523, + 480, + 559 + ], + "type": "inline_equation", + "content": "0.57\\%" + }, + { + "bbox": [ + 130, + 523, + 480, + 559 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 523, + 480, + 559 + ], + "type": "inline_equation", + "content": "0.23\\%" + }, + { + "bbox": [ + 130, + 523, + 480, + 559 + ], + "type": "text", + "content": " at DeiT-Small and Swin-Tiny." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 581, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 480, + 665 + ], + "type": "text", + "content": "Classification based on Saliency methods. To further explore the capability of SUMix, we not only apply it to some Hand-crafted methods, we apply SUMix to PuzzleMix and AutoMix, which are based on saliency-learning methods, and Tab 4 shows that SUMix will also bring some boosts, PuzzleMix and AutoMix has gained " + }, + { + "bbox": [ + 130, + 581, + 480, + 665 + ], + "type": "inline_equation", + "content": "0.30\\%" + }, + { + "bbox": [ + 130, + 581, + 480, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 581, + 480, + 665 + ], + "type": "inline_equation", + "content": "0.26\\%" + }, + { + "bbox": [ + 130, + 581, + 480, + 665 + ], + "type": "text", + "content": " at ResNet18, and " + }, + { + "bbox": [ + 130, + 581, + 480, + 665 + ], + "type": "inline_equation", + "content": "0.91\\%" + }, + { + "bbox": [ + 130, + 581, + 480, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 581, + 480, + 665 + ], + "type": "inline_equation", + "content": "0.98\\%" + }, + { + "bbox": [ + 130, + 581, + 480, + 665 + ], + "type": "text", + "content": " at ResNeXt50. Those methods combined with SUMix can even reach almost close to the performance of the current state-of-the-art method AdAutoMix." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "type": "text", + "content": "SUMix" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 170, + 320, + 229 + ], + "blocks": [ + { + "bbox": [ + 130, + 137, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 130, + 137, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 482, + 159 + ], + "type": "text", + "content": "Table 5: Top-1 acc(%)↑ and FGSM error(%)↓ Table 6: Ablation experiments on CI-of ResNet18 without and with SUMix. FAR100, CUB200 based on ResNet18." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 170, + 320, + 229 + ], + "lines": [ + { + "bbox": [ + 134, + 170, + 320, + 229 + ], + "spans": [ + { + "bbox": [ + 134, + 170, + 320, + 229 + ], + "type": "table", + "html": "
Clean Acc(%)↑Corruption Acc(%)↑FGSM Error(%)↓
MethodMCESUMixMCESUMixMCESUMix
CutMix78.1779.7843.0644.3191.1590.41
FMix79.6980.2048.7949.1489.1689.08
SaliencyMix79.1279.9143.7344.3689.6491.49
ResizeMix80.0180.3846.1246.2890.0491.05
", + "image_path": "d5615ce7be2606975bc230c4bee103c09e2d5cf71f6f5ecfd96fb4e4a1aa37e1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 323, + 170, + 481, + 229 + ], + "blocks": [ + { + "bbox": [ + 323, + 170, + 481, + 229 + ], + "lines": [ + { + "bbox": [ + 323, + 170, + 481, + 229 + ], + "spans": [ + { + "bbox": [ + 323, + 170, + 481, + 229 + ], + "type": "table", + "html": "
MethodCIFAR100CUB200
CutMixSaliencyMixFMixSaliencyMix
baseline78.1779.1277.2875.77
+λ79.1379.7979.0376.46
+ℓs79.5178.7179.0376.34
+ℓu79.6079.4379.2476.72
+ℓsu79.7879.9179.2476.98
", + "image_path": "9fd5a8fefbfe9687974816e6cc7cd7869e2cd7d39f5b844d99fad0f3b923cccf.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 251, + 270, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 251, + 270, + 262 + ], + "spans": [ + { + "bbox": [ + 131, + 251, + 270, + 262 + ], + "type": "text", + "content": "5.4 Occlusion Robustness" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 268, + 482, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 268, + 482, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 268, + 482, + 376 + ], + "type": "text", + "content": "To analyze the enhancement brought by SUMix to the robustness of the mixup methods, we randomly conducted masking patches experiment [48]. The experiment involved dividing the image into " + }, + { + "bbox": [ + 130, + 268, + 482, + 376 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 130, + 268, + 482, + 376 + ], + "type": "text", + "content": " patches and masking them according to different ratios. The resulting image was then classified by the model. We used the ImageNet dataset, which contains 1000 daily life objects. It is more representative. The ResNet18 classifier was used to classify the masked images and verify the model's robustness capability. The curves at the bottom of Figure.5 show that SUMix provides a performance gain compared to the corresponding mixup method at different occlusion ratios." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 392, + 218, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 392, + 218, + 403 + ], + "spans": [ + { + "bbox": [ + 131, + 392, + 218, + 403 + ], + "type": "text", + "content": "5.5 Robustness" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 411, + 482, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 482, + 483 + ], + "type": "text", + "content": "We first evaluated robustness against corruption on CIFAR-100-C [20], which was designed manually to provide 19 different corruptions (e.g., noise, blur, fog, etc.) to evaluate robustness. In Table 5, we show the results of improving some mixup methods with SUMix. We further investigate the FGSM [14] white box attack experiments with " + }, + { + "bbox": [ + 130, + 411, + 482, + 483 + ], + "type": "inline_equation", + "content": "8/255\\ell_{\\infty}" + }, + { + "bbox": [ + 130, + 411, + 482, + 483 + ], + "type": "text", + "content": " epsilon ball and similarly show the results of the improvement of our approach on mixup methods in Table 5." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 498, + 239, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 498, + 239, + 510 + ], + "spans": [ + { + "bbox": [ + 131, + 498, + 239, + 510 + ], + "type": "text", + "content": "5.6 Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 516, + 482, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 516, + 482, + 587 + ], + "spans": [ + { + "bbox": [ + 130, + 516, + 482, + 587 + ], + "type": "text", + "content": "In the ablation study, we primarily analyze the effect of the two SUMix modules in the model. The following two analyses are primarily done. (1) Can redefining " + }, + { + "bbox": [ + 130, + 516, + 482, + 587 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 516, + 482, + 587 + ], + "type": "text", + "content": " bring gains in models? (2) Can uncertainty lead to regularization gains? Table 6 shows the experimental results for CutMix, SaliencyMix, and FMix using ResNet18 and ResNeXt50 on the CIFAR100 and CUB200 datasets." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 592, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 592, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 592, + 482, + 665 + ], + "type": "text", + "content": "(1) Table 6 shows that with our redefined " + }, + { + "bbox": [ + 132, + 592, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\widetilde{\\lambda}" + }, + { + "bbox": [ + 132, + 592, + 482, + 665 + ], + "type": "text", + "content": ", CutMix and SaliencyMix are improved by " + }, + { + "bbox": [ + 132, + 592, + 482, + 665 + ], + "type": "inline_equation", + "content": "0.96\\%" + }, + { + "bbox": [ + 132, + 592, + 482, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 592, + 482, + 665 + ], + "type": "inline_equation", + "content": "0.67\\%" + }, + { + "bbox": [ + 132, + 592, + 482, + 665 + ], + "type": "text", + "content": " on the CIFAR100 dataset, and FMix and SaliencyMix are improved by " + }, + { + "bbox": [ + 132, + 592, + 482, + 665 + ], + "type": "inline_equation", + "content": "1.75\\%" + }, + { + "bbox": [ + 132, + 592, + 482, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 592, + 482, + 665 + ], + "type": "inline_equation", + "content": "0.69\\%" + }, + { + "bbox": [ + 132, + 592, + 482, + 665 + ], + "type": "text", + "content": " on the CUB200 dataset, which means that the module can bring a significant increase in the classification ability, in other words, the redefined " + }, + { + "bbox": [ + 132, + 592, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\widetilde{\\lambda}" + }, + { + "bbox": [ + 132, + 592, + 482, + 665 + ], + "type": "text", + "content": " can obtain a more sensible loss function for the model to train." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "type": "text", + "content": "Qin. et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 482, + 236 + ], + "type": "text", + "content": "(2) To calculate the regularization loss, we consider three cases: 1) using only the semantic information " + }, + { + "bbox": [ + 133, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "\\ell_s" + }, + { + "bbox": [ + 133, + 116, + 482, + 236 + ], + "type": "text", + "content": ", 2) using only the uncertainty information " + }, + { + "bbox": [ + 133, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "\\ell_u" + }, + { + "bbox": [ + 133, + 116, + 482, + 236 + ], + "type": "text", + "content": ", and 3) using both the semantic and uncertainty information " + }, + { + "bbox": [ + 133, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "\\ell_{su}" + }, + { + "bbox": [ + 133, + 116, + 482, + 236 + ], + "type": "text", + "content": ". Table 6 shows that on the CUB200 dataset, FMix and SaliencyMix improve accuracy by " + }, + { + "bbox": [ + 133, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "1.75\\%" + }, + { + "bbox": [ + 133, + 116, + 482, + 236 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 133, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "0.57\\%" + }, + { + "bbox": [ + 133, + 116, + 482, + 236 + ], + "type": "text", + "content": " respectively when using only " + }, + { + "bbox": [ + 133, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "\\ell_s" + }, + { + "bbox": [ + 133, + 116, + 482, + 236 + ], + "type": "text", + "content": ", and by " + }, + { + "bbox": [ + 133, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "1.96\\%" + }, + { + "bbox": [ + 133, + 116, + 482, + 236 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 133, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "0.95\\%" + }, + { + "bbox": [ + 133, + 116, + 482, + 236 + ], + "type": "text", + "content": " respectively when using only " + }, + { + "bbox": [ + 133, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "\\ell_u" + }, + { + "bbox": [ + 133, + 116, + 482, + 236 + ], + "type": "text", + "content": ". When both types of information are used, accuracy improves by " + }, + { + "bbox": [ + 133, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "1.75\\%" + }, + { + "bbox": [ + 133, + 116, + 482, + 236 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 133, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "1.21\\%" + }, + { + "bbox": [ + 133, + 116, + 482, + 236 + ], + "type": "text", + "content": ". Similarly, on the CIFAR100 dataset, there was a " + }, + { + "bbox": [ + 133, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "1.61\\%" + }, + { + "bbox": [ + 133, + 116, + 482, + 236 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 133, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "0.79\\%" + }, + { + "bbox": [ + 133, + 116, + 482, + 236 + ], + "type": "text", + "content": " improvement. The experimental results show that combining semantic and uncertainty information can significantly improve accuracy." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 255, + 317, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 255, + 317, + 268 + ], + "spans": [ + { + "bbox": [ + 132, + 255, + 317, + 268 + ], + "type": "text", + "content": "6 Conclusion and Limitations" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 280, + 482, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 280, + 482, + 364 + ], + "spans": [ + { + "bbox": [ + 132, + 280, + 482, + 364 + ], + "type": "text", + "content": "Conclusion. In this paper, we propose the SUMix approach to regularize the DNN in terms of the semantic information and uncertainty between the raw and mixed samples. Specifically, the semantic information difference between the raw and mixed samples is used to obtain the mix ratio " + }, + { + "bbox": [ + 132, + 280, + 482, + 364 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 132, + 280, + 482, + 364 + ], + "type": "text", + "content": ", while the uncertainty is, in turn, combined with the semantic information to obtain a regularized loss function. Experimental results on five datasets proved the effectiveness of SUMix for the regularization of the mixup methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 376, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 376, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 132, + 376, + 482, + 449 + ], + "type": "text", + "content": "Limitations. Currently, SUMix is due to learn how to obtain a sensible mixing ratio " + }, + { + "bbox": [ + 132, + 376, + 482, + 449 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 132, + 376, + 482, + 449 + ], + "type": "text", + "content": " and an uncertainty-based feature vector. These are currently proven only on classification tasks. In future work, we would like to apply and improve SUMix accordingly on some downstream tasks e.g. segmentation, detection, and other methods. And for semi-supervised tasks, we can get more accurate pseudo-labels for training models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 467, + 241, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 467, + 241, + 480 + ], + "spans": [ + { + "bbox": [ + 132, + 467, + 241, + 480 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 492, + 482, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 492, + 482, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 492, + 482, + 588 + ], + "type": "text", + "content": "This work is supported in part by the Scientific Innovation 2030 Major Project for New Generation of AI under Grant 2020AAA0107300, in part by the National Natural Science Foundation of China (Grant No. 61976030), in part by Chongqing Natural Science Foundation (Grant No. 2024NSCQ-MSX4893), in part by Project of CQ CSTC (Grant No. cstc2018jcyjAX 0057), in part by Innovative Research Projects for Postgraduate Students of Chongqing Business University (No. yjscxx2024-284-53, No. yiscxx2024-284-231, No. yjscxx2024-284-234)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 607, + 198, + 619 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 607, + 198, + 619 + ], + "spans": [ + { + "bbox": [ + 132, + 607, + 198, + 619 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 136, + 631, + 479, + 665 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 136, + 631, + 479, + 653 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 631, + 479, + 653 + ], + "spans": [ + { + "bbox": [ + 136, + 631, + 479, + 653 + ], + "type": "text", + "content": "1. Baek, K., Bang, D., Shim, H.: Gridmix: Strong regularization through local context mapping. Pattern Recognition 109 (2021) 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 136, + 654, + 453, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 654, + 453, + 665 + ], + "spans": [ + { + "bbox": [ + 136, + 654, + 453, + 665 + ], + "type": "text", + "content": "2. Bishop, C.M.: Pattern recognition and machine learning. Springer (2006) 2" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "type": "text", + "content": "SUMix" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 136, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 136, + 116, + 480, + 138 + ], + "type": "text", + "content": "3. Bochkovskiy, A., Wang, C.Y., Liao, H.Y.M.: Yolov4: Optimal speed and accuracy of object detection. ArXiv abs/2004.10934 (2020) 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 481, + 172 + ], + "type": "text", + "content": "4. Chang, J., Lan, Z., Cheng, C., Wei, Y.: Data uncertainty learning in face recognition. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5710-5719 (2020) 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 173, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 173, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 138, + 173, + 481, + 205 + ], + "type": "text", + "content": "5. Chen, J.N., Sun, S., He, J., Torr, P.H., Yuille, A., Bai, S.: Transmix: Attend to mix for vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12135-12144 (2022) 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 206, + 481, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 206, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 138, + 206, + 481, + 239 + ], + "type": "text", + "content": "6. Chen, M., Lin, M., Lin, Z., Zhang, Y., Chao, F., Ji, R.: Smmix: Self-motivated image mixing for vision transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17260-17270 (2023) 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 240, + 481, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 240, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 138, + 240, + 481, + 262 + ], + "type": "text", + "content": "7. Chrabaszcz, P., Loshchilov, I., Hutter, F.: A downsampled variant of imagenet as an alternative to the cifar datasets. arXiv preprint arXiv:1707.08819 (2017) 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 263, + 481, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 263, + 481, + 285 + ], + "spans": [ + { + "bbox": [ + 138, + 263, + 481, + 285 + ], + "type": "text", + "content": "8. Cubuk, E.D., Zoph, B., Mane, D., Vasudevan, V., Le, Q.V.: Autoaugment: Learning augmentation policies from data. arXiv preprint arXiv:1805.09501 (2018) 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 285, + 481, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 285, + 481, + 328 + ], + "spans": [ + { + "bbox": [ + 138, + 285, + 481, + 328 + ], + "type": "text", + "content": "9. Cubuk, E.D., Zoph, B., Shlens, J., Le, Q.V.: Randaugment: Practical automated data augmentation with a reduced search space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 702-703 (2020) 2, 11" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 330, + 481, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 330, + 481, + 362 + ], + "spans": [ + { + "bbox": [ + 133, + 330, + 481, + 362 + ], + "type": "text", + "content": "10. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018) 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 363, + 481, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 363, + 481, + 384 + ], + "spans": [ + { + "bbox": [ + 133, + 363, + 481, + 384 + ], + "type": "text", + "content": "11. DeVries, T., Taylor, G.W.: Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552 (2017) 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 386, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 386, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 133, + 386, + 481, + 407 + ], + "type": "text", + "content": "12. Feichtenhofer, C., Fan, H., Malik, J., He, K.: Slowfast networks for video recognition. In: ICCV. pp. 6201-6210 (2019) 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 408, + 481, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 408, + 481, + 441 + ], + "spans": [ + { + "bbox": [ + 133, + 408, + 481, + 441 + ], + "type": "text", + "content": "13. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. Communications of the ACM 63(11), 139-144 (2020) 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 441, + 481, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 441, + 481, + 463 + ], + "spans": [ + { + "bbox": [ + 133, + 441, + 481, + 463 + ], + "type": "text", + "content": "14. Goodfellow, I.J., Shlens, J., Szegedy, C.: Explaining and harnessing adversarial examples. In: ICLR (2015) 12" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 464, + 481, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 464, + 481, + 497 + ], + "spans": [ + { + "bbox": [ + 133, + 464, + 481, + 497 + ], + "type": "text", + "content": "15. Guénais, T., Vamvourellis, D., Yacoby, Y., Doshi-Velez, F., Pan, W.: Bacoun: Bayesian classifiers with out-of-distribution uncertainty. arXiv preprint arXiv:2007.06096 (2020) 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 498, + 481, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 498, + 481, + 519 + ], + "spans": [ + { + "bbox": [ + 133, + 498, + 481, + 519 + ], + "type": "text", + "content": "16. Harris, E., Marcu, A., Painter, M., Niranjan, M., Hare, A.P.B.J.: Fmix: Enhancing mixed sample data augmentation. arXiv preprint arXiv:2002.12047 (2020) 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 520, + 481, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 520, + 481, + 553 + ], + "spans": [ + { + "bbox": [ + 133, + 520, + 481, + 553 + ], + "type": "text", + "content": "17. Hataya, R., Zdenek, J., Yoshizoe, K., Nakayama, H.: Faster autoaugment: Learning augmentation strategies using backpropagation. In: European Conference on Computer Vision. pp. 1-16. Springer (2020) 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 554, + 481, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 554, + 481, + 576 + ], + "spans": [ + { + "bbox": [ + 133, + 554, + 481, + 576 + ], + "type": "text", + "content": "18. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: CVPR. pp. 9729-9738 (2020) 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 577, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 577, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 133, + 577, + 481, + 609 + ], + "type": "text", + "content": "19. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 770-778 (2016) 7, 9" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 133, + 610, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 610, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 133, + 610, + 481, + 632 + ], + "type": "text", + "content": "20. Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261 (2019) 12" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 133, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 632, + 481, + 665 + ], + "type": "text", + "content": "21. Hendrycks, D., Mu, N., Cubuk, E.D., Zoph, B., Gilmer, J., Lakshminarayanan, B.: Augmix: A simple data processing method to improve robustness and uncertainty. arXiv preprint arXiv:1912.02781 (2019) 4" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "type": "text", + "content": "Qin. et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 481, + 161 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 481, + 161 + ], + "type": "text", + "content": "22. Hendrycks, D., Zou, A., Mazeika, M., Tang, L., Li, B., Song, D., Steinhardt, J.: Pixmix: Dreamlike pictures comprehensively improve safety measures. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022) 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 162, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 162, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 162, + 481, + 182 + ], + "type": "text", + "content": "23. Huang, S., Wang, X., Tao, D.: Snapmix: Semantically proportional mixing for augmenting fine-grained data. In: AAAI (2020) 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 183, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 183, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 133, + 183, + 481, + 215 + ], + "type": "text", + "content": "24. Jiang, Z., Hou, Q., Yuan, L., Zhou, D., Jin, X., Wang, A., Feng, J.: Token labeling: Training a " + }, + { + "bbox": [ + 133, + 183, + 481, + 215 + ], + "type": "inline_equation", + "content": "85.4\\%" + }, + { + "bbox": [ + 133, + 183, + 481, + 215 + ], + "type": "text", + "content": " top-1 accuracy vision transformer with 56m parameters onImagenet. ArXiv abs/2104.10858 (2021) 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "type": "text", + "content": "25. Kendall, A., Gal, Y.: What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems 30 (2017) 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "type": "text", + "content": "26. Kim, J.H., Choo, W., Song, H.O.: Puzzle mix: Exploiting saliency and local statistics for optimal mixup. In: International Conference on Machine Learning. pp. 5275-5285. PMLR (2020) 4, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "type": "text", + "content": "27. Kim, J., Choo, W., Jeong, H., Song, H.O.: Co-mixup: Saliency guided joint mixup with supermodular diversity. In: International Conference on Learning Representations (2020) 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "type": "text", + "content": "28. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009) 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 326, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 357 + ], + "type": "text", + "content": "29. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. In: Advances in neural information processing systems. pp. 1097-1105 (2012) 1, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 358, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 402 + ], + "type": "text", + "content": "30. Lee, J.H., Zaheer, M.Z., Astrid, M., Lee, S.I.: Smoothmix: a simple yet effective data augmentation to train robust classifiers. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 756-757 (2020) 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 402, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 481, + 435 + ], + "type": "text", + "content": "31. Li, S., Jin, W., Wang, Z., Wu, F., Liu, Z., Tan, C., Li, S.Z.: Semireward: A general reward model for semi-supervised learning. In: International Conference on Learning Representations (2024) 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "type": "text", + "content": "32. Li, S., Liu, Z., Wang, Z., Wu, D., Liu, Z., Li, S.Z.: Boosting discriminative visual representation learning with scenario-agnostic mixup. ArXiv abs/2111.15454 (2021) 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "type": "text", + "content": "33. Li, S., Liu, Z., Zang, Z., Wu, D., Chen, Z., Li, S.Z.: Genurl: A general framework for unsupervised representation learning. IEEE Transactions on Neural Networks and Learning Systems (2024) 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 501, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 481, + 533 + ], + "type": "text", + "content": "34. Li, S., Wang, Z., Liu, Z., Tan, C., Lin, H., Wu, D., Chen, Z., Zheng, J., Li, S.Z.: Efficient multi-order gated aggregation network. ArXiv abs/2211.03295 (2022) 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 533, + 481, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 481, + 566 + ], + "type": "text", + "content": "35. Li, S., Wang, Z., Liu, Z., Wu, D., Li, S.Z.: Openmixup: Open mixup toolbox and benchmark for visual representation learning. https://github.com/Westlake-AI/openmixup (2022) 8, 9" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "type": "text", + "content": "36. Li, S., Wu, D., Wu, F., Zang, Z., Li, S.Z.: Architecture-agnostic masked image modeling – from vit back to cnn. In: International Conference on Machine Learning (2023) 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 600, + 481, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 481, + 631 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 481, + 631 + ], + "type": "text", + "content": "37. Li, S., Zhang, Z., Liu, Z., Wang, A., Qiu, L., Du, F.: Tlpg-tracker: Joint learning of target localization and proposal generation for visual tracking. In: IJCAI (2020) 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "text", + "content": "38. Li, Y., Hu, G., Wang, Y., Hospedales, T., Robertson, N.M., Yang, Y.: Differentiable automatic data augmentation. In: European Conference on Computer Vision. pp. 580-595. Springer (2020) 2" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "type": "text", + "content": "SUMix" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "39. Lim, S., Kim, I., Kim, T., Kim, C., Kim, S.: Fast autoaugment. In: Advances in Neural Information Processing Systems (NeurIPS) (2019) 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 480, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 480, + 170 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 480, + 170 + ], + "type": "text", + "content": "40. Liu, J., Liu, B., Zhou, H., Li, H., Liu, Y.: Tokenmix: Rethinking image mixing for data augmentation in vision transformers. In: European Conference on Computer Vision (2022) 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 171, + 480, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 171, + 480, + 203 + ], + "spans": [ + { + "bbox": [ + 132, + 171, + 480, + 203 + ], + "type": "text", + "content": "41. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: International Conference on Computer Vision (ICCV) (2021) 11" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 203, + 480, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 203, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 480, + 236 + ], + "type": "text", + "content": "42. Liu, Z., Li, S., Wang, G., Wu, L., Tan, C., Li, S.Z.: Harnessing hard mixed samples with decoupled regularizer. In: Thirty-seventh Conference on Neural Information Processing Systems (2023) 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 236, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 480, + 267 + ], + "type": "text", + "content": "43. Liu, Z., Li, S., Wang, L., Wang, Z., Liu, Y., Li, S.Z.: Short-long convolutions help hardware-efficient linear attention to focus on long sequences. In: Forty-first International Conference on Machine Learning (2024) 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 268, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 268, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 268, + 480, + 300 + ], + "type": "text", + "content": "44. Liu, Z., Li, S., Wu, D., Liu, Z., Chen, Z., Wu, L., Li, S.Z.: Automix: Unveiling the power of mixup for stronger classifiers. In: European Conference on Computer Vision. pp. 441-458. Springer (2022) 4, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 300, + 480, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 480, + 322 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 480, + 322 + ], + "type": "text", + "content": "45. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983 (2016) 9" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 322, + 480, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 322, + 480, + 341 + ], + "spans": [ + { + "bbox": [ + 132, + 322, + 480, + 341 + ], + "type": "text", + "content": "46. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: ICLR (2019) 11" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 342, + 480, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 480, + 363 + ], + "type": "text", + "content": "47. Maji, S., Rahtu, E., Kannala, J., Blaschko, M., Vedaldi, A.: Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151 (2013) 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 364, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 480, + 396 + ], + "type": "text", + "content": "48. Naseer, M., Ranasinghe, K., Khan, S.H., Hayat, M., Khan, F.S., Yang, M.H.: Intriguing properties of vision transformers. In: Neural Information Processing Systems (2021) 12" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 396, + 480, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 480, + 427 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 480, + 427 + ], + "type": "text", + "content": "49. Oh, S.J., Murphy, K., Pan, J., Roth, J., Schroff, F., Gallagher, A.: Modeling uncertainty with hedged instance embedding. arXiv preprint arXiv:1810.00319 (2018) 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 428, + 480, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 428, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 428, + 480, + 460 + ], + "type": "text", + "content": "50. Park, J., Yang, J.Y., Shin, J., Hwang, S.J., Yang, E.: Saliency grafting: Innocuous attribution-guided mixup with calibrated label mixing. In: AAAI Conference on Artificial Intelligence (2021) 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 461, + 480, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 461, + 480, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 461, + 480, + 482 + ], + "type": "text", + "content": "51. Qin, H., Jin, X., Jiang, Y., El-Yacoubi, M.A., Gao, X.: Adversarial automixup. arXiv preprint arXiv:2312.11954 (2023) 2, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 483, + 480, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 480, + 514 + ], + "type": "text", + "content": "52. Qin, J., Fang, J., Zhang, Q., Liu, W., Wang, X., Wang, X.: Resizemix: Mixing data with preserved object information and true labels. arXiv preprint arXiv:2012.11101 (2020) 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 515, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 515, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 515, + 480, + 536 + ], + "type": "text", + "content": "53. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. arXiv preprint arXiv:1506.01497 (2015) 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 536, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 480, + 568 + ], + "type": "text", + "content": "54. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Gradcam: Visual explanations from deep networks via gradient-based localization. arXiv preprint arXiv:1610.02391 (2019) 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 569, + 480, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 569, + 480, + 589 + ], + "spans": [ + { + "bbox": [ + 132, + 569, + 480, + 589 + ], + "type": "text", + "content": "55. Shaw, G., Manolakis, D.: Signal processing for hyperspectral image exploitation. IEEE Signal processing magazine 19(1), 12-16 (2002) 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 590, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 480, + 609 + ], + "type": "text", + "content": "56. Shi, Y., Jain, A.K.: Probabilistic face embeddings. In: ICCV. pp. 6902-6911 (2019) 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "text", + "content": "57. Sun, S., Chen, J.N., He, R., Yuille, A., Torr, P., Bai, S.: Lumix: Improving mixup by better modelling label uncertainty. arXiv preprint arXiv:2211.15846 (2022) 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "text", + "content": "58. Suzuki, T.: Teachaugment: Data augmentation optimization using teacher knowledge. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10904-10914 (June 2022) 2" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "type": "text", + "content": "Qin. et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 660 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "text", + "content": "59. Tan, C., Gao, Z., Wu, L., Xu, Y., Xia, J., Li, S., Li, S.Z.: Temporal attention unit: Towards efficient spatiotemporal predictive learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18770-18782 (2023) 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 160, + 481, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 160, + 481, + 192 + ], + "spans": [ + { + "bbox": [ + 130, + 160, + 481, + 192 + ], + "type": "text", + "content": "60. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jegou, H.: Training data-efficient image transformers & distillation through attention. In: International Conference on Machine Learning (ICML). pp. 10347-10357 (2021) 11" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 192, + 481, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 192, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 192, + 481, + 224 + ], + "type": "text", + "content": "61. Uddin, A.S., Monira, M.S., Shin, W., Chung, T., Bae, S.H.: Saliencymix: A saliency guided data augmentation strategy for better regularization. In: International Conference on Learning Representations (2020) 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 224, + 481, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 481, + 245 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 481, + 245 + ], + "type": "text", + "content": "62. Vaswani, A., Shazeer, N.M., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NIPS (2017) 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 245, + 481, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 245, + 481, + 277 + ], + "spans": [ + { + "bbox": [ + 130, + 245, + 481, + 277 + ], + "type": "text", + "content": "63. Venkataramanan, S., Avrithis, Y., Kijak, E., Amsaleg, L.: Alignmix: Improving representation by interpolating aligned features. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022) 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 277, + 481, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 481, + 310 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 481, + 310 + ], + "type": "text", + "content": "64. Verma, V., Lamb, A., Beckham, C., Najafi, A., Mitliagkas, I., Lopez-Paz, D., Bengio, Y.: Manifold mixup: Better representations by interpolating hidden states. In: International Conference on Machine Learning. pp. 6438-6447 (2019) 2, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 310, + 481, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 310, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 310, + 481, + 342 + ], + "type": "text", + "content": "65. Vinyals, O., Toshev, A., Bengio, S., Erhan, D.: Show and tell: A neural image caption generator. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3156-3164 (2015) 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 342, + 481, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 481, + 362 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 481, + 362 + ], + "type": "text", + "content": "66. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The caltech-ucsd birds-200-2011 dataset. California Institute of Technology (2011) 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 362, + 481, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 362, + 481, + 405 + ], + "spans": [ + { + "bbox": [ + 130, + 362, + 481, + 405 + ], + "type": "text", + "content": "67. Walawalkar, D., Shen, Z., Liu, Z., Savvides, M.: Attentive cutmix: An enhanced data augmentation approach for deep learning based image classification. In: IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 3642-3646 (2020) 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 405, + 481, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 405, + 481, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 405, + 481, + 426 + ], + "type": "text", + "content": "68. Wang, C., Zheng, W., Zhu, Z., Zhou, J., Lu, J.: Introspective deep metric learning. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023) 5, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 426, + 481, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 481, + 458 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 481, + 458 + ], + "type": "text", + "content": "69. Xie, S., Girshick, R., Dollár, P., Tu, Z., He, K.: Aggregated residual transformations for deep neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1492-1500 (2017) 9" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 458, + 481, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 458, + 481, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 458, + 481, + 489 + ], + "type": "text", + "content": "70. Yang, L., Li, X., Zhao, B., Song, R., Yang, J.: Recursiveemix: Mixed learning with history. Advances in Neural Information Processing Systems 35, 8427-8440 (2022) 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 489, + 481, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 489, + 481, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 489, + 481, + 521 + ], + "type": "text", + "content": "71. Yang, S., Wu, Y., Jin, X., El Yacoubi, M., Qin, H.: Cgan-da: A cross-modality domain adaptation model for hand-vein biometric-based authentication. JOURNAL OF Cyber-Physical-Social Intelligence 1, 3–12 (2022) 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 521, + 481, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 481, + 554 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 481, + 554 + ], + "type": "text", + "content": "72. Yun, S., Han, D., Oh, S.J., Chun, S., Choe, J., Yoo, Y.: Cutmix: Regularization strategy to train strong classifiers with localizable features. In: International Conference on Computer Vision (ICCV). pp. 6023-6032 (2019) 2, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 554, + 481, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 554, + 481, + 574 + ], + "spans": [ + { + "bbox": [ + 130, + 554, + 481, + 574 + ], + "type": "text", + "content": "73. Zagoruyko, S., Komodakis, N.: Wide residual networks. In: Proceedings of the British Machine Vision Conference (BMVC) (2016) 9" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 574, + 481, + 607 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 574, + 481, + 607 + ], + "spans": [ + { + "bbox": [ + 130, + 574, + 481, + 607 + ], + "type": "text", + "content": "74. Zhang, B., Wonka, P.: Point cloud instance segmentation using probabilistic embeddings. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8883-8892 (2021) 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 607, + 481, + 628 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 607, + 481, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 607, + 481, + 628 + ], + "type": "text", + "content": "75. Zhang, H., Cisse, M., Dauphin, Y.N., Lopez-Paz, D.: mixup: Beyond empirical risk minimization. In: International Conference on Learning Representations (2018) 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 628, + 481, + 660 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 628, + 481, + 660 + ], + "spans": [ + { + "bbox": [ + 130, + 628, + 481, + 660 + ], + "type": "text", + "content": "76. Zhao, Q., Huang, Y., Hu, W., Zhang, F., Liu, J.: Mixpro: Data augmentation with maskmix and progressive attention labeling for vision transformer. In: The Eleventh International Conference on Learning Representations (2023) 5" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "type": "text", + "content": "SUMix" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 207, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 207, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 207, + 129 + ], + "type": "text", + "content": "Appendix A" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 137, + 289, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 137, + 289, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 137, + 289, + 148 + ], + "type": "text", + "content": "A.1 More Information about " + }, + { + "bbox": [ + 132, + 137, + 289, + 148 + ], + "type": "inline_equation", + "content": "\\zeta" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 152, + 482, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 482, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 482, + 187 + ], + "type": "text", + "content": "We reported the hyper-parameter " + }, + { + "bbox": [ + 130, + 152, + 482, + 187 + ], + "type": "inline_equation", + "content": "\\zeta" + }, + { + "bbox": [ + 130, + 152, + 482, + 187 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 130, + 152, + 482, + 187 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{su}" + }, + { + "bbox": [ + 130, + 152, + 482, + 187 + ], + "type": "text", + "content": " on the CIFAR100 dataset shown in Table 7. And for the fine-grained dataset, CUB200 and FGVC-Aircraftd. We reported the settings in Table 8 and Table 9." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 136, + 210, + 478, + 286 + ], + "blocks": [ + { + "bbox": [ + 203, + 189, + 409, + 200 + ], + "lines": [ + { + "bbox": [ + 203, + 189, + 409, + 200 + ], + "spans": [ + { + "bbox": [ + 203, + 189, + 409, + 200 + ], + "type": "text", + "content": "Table 7: The " + }, + { + "bbox": [ + 203, + 189, + 409, + 200 + ], + "type": "inline_equation", + "content": "\\zeta" + }, + { + "bbox": [ + 203, + 189, + 409, + 200 + ], + "type": "text", + "content": " setting on the CIFAR100 dataset." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 210, + 478, + 286 + ], + "lines": [ + { + "bbox": [ + 136, + 210, + 478, + 286 + ], + "spans": [ + { + "bbox": [ + 136, + 210, + 478, + 286 + ], + "type": "table", + "html": "
ζ settingCIFAR100
ResNet18ResNeXt50W-ResNet28-8DeiT-SmallSwin-Tiny
CutMix0.50.50.50.50.5
FMix1.01.00.50.50.5
SaliencyMix0.51.00.50.50.5
ResizeMix0.20.50.50.50.5
", + "image_path": "a552b9840f7e253ef0298bab9ac0e628680d63bf1370887753c9d6990e79af8a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 136, + 323, + 479, + 401 + ], + "blocks": [ + { + "bbox": [ + 132, + 290, + 482, + 312 + ], + "lines": [ + { + "bbox": [ + 132, + 290, + 482, + 312 + ], + "spans": [ + { + "bbox": [ + 132, + 290, + 482, + 312 + ], + "type": "text", + "content": "Table 8: The " + }, + { + "bbox": [ + 132, + 290, + 482, + 312 + ], + "type": "inline_equation", + "content": "\\zeta" + }, + { + "bbox": [ + 132, + 290, + 482, + 312 + ], + "type": "text", + "content": " setting on the CUB200 Table 9: The " + }, + { + "bbox": [ + 132, + 290, + 482, + 312 + ], + "type": "inline_equation", + "content": "\\zeta" + }, + { + "bbox": [ + 132, + 290, + 482, + 312 + ], + "type": "text", + "content": " setting on the FGVC-dataset." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 323, + 479, + 401 + ], + "lines": [ + { + "bbox": [ + 136, + 323, + 479, + 401 + ], + "spans": [ + { + "bbox": [ + 136, + 323, + 479, + 401 + ], + "type": "table", + "html": "
ζ settingCUB200ζ settingFGVC-Aircrafts
ResNet18ResNeXt50ResNet18ResNeXt50
CutMix1.00.5CutMix1.00.5
FMix0.50.1FMix0.50.5
SaliencyMix0.50.2SaliencyMix1.00.5
ResizeMix0.50.5ResizeMix0.50.5
", + "image_path": "6fc94597c5c5571b1f9eb01d63f2ea99bd6972e3c032ac1264d6680a5bb436f1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 439, + 316, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 439, + 316, + 451 + ], + "spans": [ + { + "bbox": [ + 132, + 439, + 316, + 451 + ], + "type": "text", + "content": "A.2 More Visualization Comparison" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 456, + 481, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 456, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 456, + 481, + 479 + ], + "type": "text", + "content": "In Figure.6 and Figure.7, we show the visualization obtained by the mixup method using the CAM method with and without SUMix." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 160, + 502, + 452, + 628 + ], + "blocks": [ + { + "bbox": [ + 160, + 502, + 452, + 628 + ], + "lines": [ + { + "bbox": [ + 160, + 502, + 452, + 628 + ], + "spans": [ + { + "bbox": [ + 160, + 502, + 452, + 628 + ], + "type": "image", + "image_path": "5dda73e6bbf47db8e12e5c3d89e3ee6affe782ef8da874a5745f9e8923fdf298.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 639, + 482, + 663 + ], + "lines": [ + { + "bbox": [ + 130, + 639, + 482, + 663 + ], + "spans": [ + { + "bbox": [ + 130, + 639, + 482, + 663 + ], + "type": "text", + "content": "Fig.6: CAM visualization of the ResNet18 with and without SUMix trained on the ImageNet-1K dataset." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "type": "text", + "content": "Qin. et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 147, + 178, + 462, + 566 + ], + "blocks": [ + { + "bbox": [ + 147, + 178, + 462, + 566 + ], + "lines": [ + { + "bbox": [ + 147, + 178, + 462, + 566 + ], + "spans": [ + { + "bbox": [ + 147, + 178, + 462, + 566 + ], + "type": "image", + "image_path": "b9cf32e5858c579bbd311bb0e69ca9e3dd749b7cde0ab31d96230aa9fb51a7a9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 578, + 482, + 602 + ], + "lines": [ + { + "bbox": [ + 132, + 578, + 482, + 602 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 482, + 602 + ], + "type": "text", + "content": "Fig. 7: CAM visualization of the ResNet18 with and without SUMix trained on the ImageNet-1K dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 417, + 91, + 448, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 417, + 91, + 448, + 100 + ], + "spans": [ + { + "bbox": [ + 417, + 91, + 448, + 100 + ], + "type": "text", + "content": "SUMix" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/SUP-NeRF_ A Streamlined Unification of Pose Estimation and NeRF for Monocular 3D Object Reconstruction/c913b291-07c4-425a-807f-281296d9ef23_content_list.json b/2024/SUP-NeRF_ A Streamlined Unification of Pose Estimation and NeRF for Monocular 3D Object Reconstruction/c913b291-07c4-425a-807f-281296d9ef23_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f16907f6d293984338c9838b23e34086485d969f --- /dev/null +++ b/2024/SUP-NeRF_ A Streamlined Unification of Pose Estimation and NeRF for Monocular 3D Object Reconstruction/c913b291-07c4-425a-807f-281296d9ef23_content_list.json @@ -0,0 +1,1592 @@ +[ + { + "type": "text", + "text": "SUP-NeRF: A Streamlined Unification of Pose Estimation and NeRF for Monocular 3D Object Reconstruction", + "text_level": 1, + "bbox": [ + 227, + 140, + 776, + 205 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuliang Guo $^{1\\star}$", + "bbox": [ + 290, + 233, + 397, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ruoyu Wang", + "bbox": [ + 292, + 251, + 393, + 265 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abhinav Kumar1,2", + "bbox": [ + 439, + 234, + 573, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xinyu Huang", + "bbox": [ + 457, + 251, + 560, + 265 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Cheng Zhao1", + "bbox": [ + 614, + 234, + 710, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Liu Ren1", + "bbox": [ + 630, + 251, + 697, + 263 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Bosch Research North America & Bosch Center for Artificial Intelligence (BCAI)", + "bbox": [ + 225, + 268, + 772, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2Michigan State University", + "bbox": [ + 410, + 284, + 591, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "[yuiliang.guo2,cheng.zhao,ruoyu.wang,xingyu.huang,liu. ren]@us.bosch.com", + "bbox": [ + 227, + 297, + 774, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ kumarab6@msu.edu", + "bbox": [ + 434, + 311, + 566, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://yuliangguo.github.io/supnerf", + "bbox": [ + 359, + 325, + 640, + 340 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f645d13ac8f4e7560605408b99cea7f685fef0bba1f61c1b7b121faa26b5d499.jpg", + "image_caption": [ + "Fig. 1: Teaser. SUP-NeRF is a unified solution that predicts an object's pose, shape, and texture using a single network. SUP-NeRF is trained on real driving scenes with imprecise labels, and it adapts robustly to new cross-dataset scenarios." + ], + "image_footnote": [], + "bbox": [ + 253, + 340, + 748, + 506 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Monocular 3D reconstruction for categorical objects heavily relies on accurately perceiving each object's pose. While gradient-based optimization in a NeRF framework updates the initial pose, this paper highlights that scale-depth ambiguity in monocular object reconstruction causes failures when the initial pose deviates moderately from the true pose. Consequently, existing methods often depend on a third-party 3D object to provide an initial object pose, leading to increased complexity and generalization issues. To address these challenges, we present SUP-NeRF, a Streamlined Unification of object Pose estimation and NeRF-based object reconstruction. SUP-NeRF decouples the object's dimension estimation and pose refinement to resolve the scale-depth ambiguity, and introduces a camera-invariant projected-box representation that generalizes cross different domains. While using a dedicated pose estimator that smoothly integrates into an object-centric NeRF, SUP-NeRF is free from external 3D detectors. SUP-NeRF achieves state-of-the-art results in both reconstruction and pose estimation tasks on the nuScenes dataset. Furthermore, SUP-NeRF exhibits exceptional cross-dataset generalization on the KITTI and Waymo datasets, surpassing prior methods with up to $50\\%$ reduction in rotation and translation error.", + "bbox": [ + 259, + 556, + 740, + 820 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Project Lead", + "bbox": [ + 220, + 825, + 321, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 143, + 374, + 160 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Monocular 3D object reconstruction [8,9,13,19,49] is a critical technology with broad applications in autonomous driving [18,29,49], AR/VR, robotics [43], and embodied AI [1,40,51]. Its impact is particularly significant in autonomous driving, where it plays a vital role in tasks like auto-labeling [49] and world-model construction [18,29]. This technology is essential for reconstructing 3D models from typical driving scenarios, where single-view observations are common. This paper focuses explicitly on the problem of recovering an object's pose, shape, and texture from a single image under real driving scenarios, see Fig. 1. Our aim is to devise an efficient solution, thereby unlocking the full potential of monocular reconstruction for active tasks such as planning and control.", + "bbox": [ + 212, + 174, + 785, + 325 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural Radiance Field (NeRF) technology [27, 30, 48] has revolutionized the field of 3D reconstruction. It excels at capturing fine details in a scene and generating novel views from existing reconstructions. An exciting branch, Object-centric NeRF [11, 12, 18, 29, 34, 45, 46], takes things further, allowing for flexible data creation and tackling reconstruction challenges from limited number of views by incorporating category-specific prior knowledge. Earlier object-centric NeRF methods [8, 12, 13, 19, 46, 48, 49] require densely overlapping views and precise poses, limiting their use to synthetic data. Later advances [18, 29] enable monocular training with imperfect masks.", + "bbox": [ + 212, + 325, + 785, + 462 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although promising, object NeRFs show limitations in correcting erroneous object poses. First, the well-known scale-depth ambiguity [16,25,36] also affects the object NeRF optimization. Fig. 2 shows the optimization failure of a monocular NeRF method under longitudinal translation error when it is given complete freedom in adjusting the scale within the normalized space. Secondly, NeRF also struggles with large rotation errors. Appendix A6.1 shows that NeRF typically requires the initial rotation error to be below $25^{\\circ}$ , and the pose optimization cannot go lower than $7^{\\circ}$ .", + "bbox": [ + 212, + 462, + 785, + 582 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent papers [18, 29] use an external monocular 3D detector to provide initial poses to NeRF. However, monocular 3D detectors suffer from generalization issues [16, 17] and increased complexity, ultimately affecting the NeRF pipeline. BootInv [34], a more recent method, tackles pose estimation, shape, and appearance reconstruction in a single framework but requires additional pre-training of dense field generator and struggles with handling moderate occlusions.", + "bbox": [ + 212, + 583, + 785, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To resolve the scale-depth ambiguity and achieve greater generalization, this paper proposes SUP-NeRF, a Streamlined Unification of Pose estimation and NeRF. SUP-NeRF uses an image encoder to output the object dimensions and then refines for object pose using an iterative process while keeping its size fixed. Besides decoupling of an object's dimension and pose to resolve the scale-depth ambiguity, the introduction of a novel projected-box representation in the iterative update refinement also enhances the generalization to novel data since the representation makes the deep pose refiner independent from camera intrinsic parameters. In addition, SUP-NeRF tackles the complexity challenge by sharing the NeRF image encoder with the pose estimation, making the pose estimation module a seamless fit for object NeRFs.", + "bbox": [ + 212, + 672, + 785, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Y. Guo et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/833ef2f859bd16c3c8b98127c27d658962597580f5a0d19d9317c863735f15c1.jpg", + "image_caption": [ + "R err: 0.00, T err: 6.85", + "Fig. 2: Scale-depth ambiguity in NeRF. Given the input image (left), joint optimization of pose, shape, and texture in NeRF has full freedom to rescale the shape within the normalized shape space (blue box) or move the 3D box. Such phenomenon is observed from the evolution of the rendered objects from iteration 0 (middle) to iteration 50 (right)." + ], + "image_footnote": [], + "bbox": [ + 303, + 145, + 436, + 250 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/fd13ec9be23ef6727ec18a22077a79c12b6a7afa4c00a0b5990be1d3b9baf744.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 449, + 175, + 558, + 217 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/48cfe8c73f976499f20bde598fd27dc4b5a161aa021a162e7103d5c661633f33.jpg", + "image_caption": [ + "R err: 0.06, Terr: 6.35" + ], + "image_footnote": [], + "bbox": [ + 581, + 172, + 694, + 218 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In our experiments, we first comprehensively benchmark and conduct ablations on SUP-NeRF on the nuScenes [3] dataset. SUP-NeRF outperforms state-of-the-art (SoTA) methods and achieves SoTA object reconstruction and pose performance on this dataset. We next conduct cross-dataset object reconstruction and pose estimation experiments of the nuScenes models on the KITTI and Waymo datasets to test SUP-NeRF's generalization capability. SUP-NeRF shows more significant improvements over SoTA in such cross-dataset settings.", + "bbox": [ + 212, + 340, + 784, + 445 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, our main contributions include:", + "bbox": [ + 238, + 446, + 563, + 460 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce SUP-NeRF, an streamlined method unifying object pose estimation and monocular object-centric neural reconstruction.", + "- We identify depth-scale ambiguities as the key barrier to object NeRF's pose optimization and craft a module to overcome this effectively, while also integrating seamlessly with existing object NeRF frameworks.", + "- We introduce a novel projected box representation for the pose estimation module, enabling the deep refiner to operate in a space invariant to camera intrinsic parameters, thereby enhancing SUP-NeRF's generalization capability.", + "- SUP-NeRF achieves SoTA object reconstruction and pose estimation results, outperforming other methods on the nuScenes dataset, and in challenging cross-dataset experiments on both KITTI and Waymo datasets." + ], + "bbox": [ + 210, + 462, + 784, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 654, + 385, + 669 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Object-centric NeRF. Object-centric NeRF [11, 12, 18, 29, 34, 45, 46] is a specialized NeRF variant for modeling specific object categories. Its pipeline feeds a NeRF decoder with shape and texture codes, generating occupancy and color values at 3D positions, followed by volumetric rendering to produce a rendered image. The entire process is differentiable, allowing back-propagation of the photometric loss to update shape codes, texture codes, network weights, and object poses. Early object-centric NeRFs [8, 12, 13, 19, 45, 46, 48, 49] required accurate object poses, full visibility, and dense multi-view observations to reconstruct a categorical model, making them inaccessible for real-world applications. Some approaches [12, 23] claim successful object pose recovery through gradient-based", + "bbox": [ + 212, + 688, + 784, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "SUP-NeRF", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/c18b108a521e97e24aafac18d54680a054e8939898b3d77e3218974dc435f5ae.jpg", + "table_caption": [ + "Table 1: Literature Summary. SUP-NeRF sets a benchmark for minimal requirements, eliminating the need for external 3D detectors, CAD models, or dense NOCS generators. It is dedicated to the reconstruction of object shapes and pose estimation directly from a single image of driving scenes, employing an object-centric framework." + ], + "table_footnote": [], + "table_body": "
MethodsIndoorOutdoorViewsScopeCADwt. Det3Dwt. NeRFNOCS Gen.
6DoF Pose [21, 26]✓×singleobject✓✓××
3D Detector [16, 42]×✓singleimage×✓××
NeRF [27]✓✓multiscene××✓×
Object-NeRF [18, 29]×✓singleobject××✓×
BootInv [34]×✓singleobject×✓✓✓
SUP-NeRF×✓singleobject×✓✓×
", + "bbox": [ + 225, + 212, + 772, + 301 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "pose refinement from monocular images. However these successes are still limited to synthetic dataset, and oversimplified object/camera poses. Recent methods [18,29] investigate more practical pipelines demonstrating the object-NeRF's potential to train and test on driving scenes with limited views and occluded object data. However, both methods require relatively accurate poses from heavy third-party 3D detectors [32,36]. Other recent works [2,5] try freeing NeRF from external camera poses; however, solving camera poses from densely overlapping views is a different problem than recovering an object's pose from a single image. Compared to these methods, SUP-NeRF recovers object pose from a single image without a 3D detector in the pipeline. A more recent work [34] also unifies pose estimation and NeRF in a single network, and enables high-fidelity object generation. However, its training requires additional NOCS dense generator and it is highly sensitive to moderate occlusions compared to our method.", + "bbox": [ + 212, + 318, + 787, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Monocular 3D Object Pose Estimation. Advancements in deep learning and the availability of 3D-labeled datasets [3, 6, 10] enable many monocular 3D object detection [16, 24, 25, 33, 36, 42, 47] and 6DoF pose estimation methods [14, 21, 22, 26, 28, 39, 44, 50]. Although estimating 3D information from a single image is ill-posed due to depth-scale ambiguity, leveraging constraints like CAD models or sizes enables accurate object pose recovery via PnP processes [35, 39, 41] or iterative optimization [14, 21]. These indoor 6DoF pose estimation mostly follow an object-centric design which aims to estimate highly accurate object pose within the 2D image patches out of a previous 2D object detector. Recent studies [4, 20] use the NeRF pipeline to reconstruct 3D models or to train model-based keypoint detectors to estimate object poses during inference. However, none of these model-based methods is ideal for outdoor driving scenarios, as pre-reconstruction of models is often impossible in new scenes. Fortunately, inferring the 3D dimensions of objects from an image is feasible when its appearance is tightly bounded to its fine-grained class [28], which keeps an outdoor 3D detector from being fully ill-posed. Nowadays, reliable outdoor 3D detectors usually favors a image-centric design to direct detect many 3D objects captured within the whole image, which usually requires a large backbone network [16, 42]. However, the performance of SoTA detectors drops significantly when applied to datasets with distribution different from the training data [16, 25, 42].", + "bbox": [ + 212, + 522, + 787, + 843 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Y. Guo et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/193c1224eb7b64e9afba6de803a5eb9e99a264f4049fbaf4df452e2754e0918b.jpg", + "image_caption": [ + "Fig. 3: SUP-NeRF Overview. SUP-NeRF unifies pose estimation and NeRF. The pose estimation module enables SUP-NeRF to work for objects in diverse poses without external 3D detectors. The increase of complexity only constitutes a few MLP layers." + ], + "image_footnote": [], + "bbox": [ + 228, + 146, + 779, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Tab. 1 outline the differences in methodologies. Unlike existing approaches, SUP-NeRF unifies pose estimation and object NeRF in a streamlined manner, adopting a straightforward object-centric design akin to indoor techniques. Without the need for CAD models, SUP-NeRF is suitable for outdoor driving scenarios. Despite its focused scope on object-centric pose estimation, SUP-NeRF showcases enhanced cross-domain generalization capabilities compared to more complex image-centric 3D detectors within this realm.", + "bbox": [ + 212, + 409, + 787, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 SUP-NeRF", + "text_level": 1, + "bbox": [ + 215, + 536, + 362, + 553 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our goal is to create a NeRF framework that bypasses NeRF's limitations without external 3D detectors. Fig. 3 shows the overall pipeline of SUP-NeRF, a streamlined unified network that predicts object pose, shape, and texture from a single image. SUP-NeRF consists of three modules: a Resnet-based encoder, a pose estimation module, and a NeRF decoder. SUP-NeRF processes an image and its occlusion mask, with the encoder turning the masked image into shape, texture, and pose codes and predicted object dimensions. The pose code, object dimensions and a box code, are passed to the pose estimation module to iteratively refine the object pose $[R_{o2c}|T_{o2c}]$ . After a few iterations, the estimated pose is converted to the camera pose $[R_{c2o}|T_{c2o}]$ and passed to the NeRF decoder. In the NeRF phase, the decoder uses the shape, texture codes, and camera pose for volumetric rendering, creating both RGB and occupancy images.", + "bbox": [ + 212, + 566, + 787, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pose Estimation Module. The key component of SUP-NeRF is the pose estimation module, which we show in Fig. 4. It aims to provide reliable object poses for targets at all ranges and orientations. Our pose estimation module iterative updates the input pose, based on the visual difference between the input pose and the observed object in the image. Given predicted object dimensions $[H_B, W_B, L_B]$ , camera intrinsic $K$ , and current pose $R^{(t)}, T^{(t)}$ , image projections", + "bbox": [ + 212, + 750, + 787, + 842 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "SUP-NeRF", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e0b9fe08d967f5b1f1b3a41d36d651e04c554ab36d004762ad654cf770c78147.jpg", + "image_caption": [ + "Fig.4: SUP-NeRF Pose Estimation Module. The pose estimation module of SUP-NeRF iteratively updates the object's pose while preserving scale. It takes the projection of 3D box corners as a visual representation of the input pose and estimate the pose update via comparing it to observed image in a latent embedding space. These designs handle scale-depth ambiguity and make the deep refiner independent from camera intrinsic parameters for better cross-domain generalization." + ], + "image_footnote": [], + "bbox": [ + 303, + 146, + 699, + 296 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "of the 3D box corners $B_{proj}^{(t)}$ is computed, which can be interpolated as a visual representation of current pose. After encoding the 8 corners into a higher dimensional box code, SUP-NeRF feeds into the pose refiner along with a pose code from the image encoder to predict the desired pose update $\\Delta R^{(t)}, \\Delta T^{(t)}$ . SUP-NeRF then composes the pose changes with current pose to compute the next pose state, $R^{(t + 1)}, T^{(t + 1)}$ . This updating process starts from a fully random object pose within the viewing frustum of the image patch, and takes a few iterations such that the final pose approaches the true pose even from a random initial pose.", + "bbox": [ + 212, + 417, + 787, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The pose estimation module incorporates the following unique designs:", + "bbox": [ + 238, + 560, + 750, + 575 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- SUP-NeRF preserves the object dimensions to ensure unambiguous pose updates. Our implementation, using projected 3D boxes on the image plane, leads to a concise encoding of shape and pose, reducing computation in image rendering and encoding per iteration, as opposed to previous methods [21, 29] relying on full CAD models. Since this representation makes the deep pose refiner's task independent from camera intrinsic parameters, it leads to better generalization to different testing domains, as evidenced in Tabs. 3 and 4.", + "- SUP-NeRF updates the object pose iteratively rather than conduct a direct pose estimation using a PnP-based [14] because the later is sensitive to missed or swapped corners thus often produces far out-of-distribution poses. Our ablation studies in Tab. 6 confirm this difference.", + "- SUP-NeRF represents the rotation update in axis-angle representation [37]. During rotation updating process, we convert the axis-angle space 3D rotation $\\Delta R^{(t)}$ back to $SO(3)$ , compose it with the previous rotation to compute the new rotation state as $R_{\\mathrm{e2}}^{(t + 1)}$ .", + "- SUP-NeRF represents the translation update in a relative space, indicated as $\\Delta T^{(t)}: [v_x^{(t)}, v_y^{(t)}, \\rho^{(t)}]$ , where $v_x^{(t)}$ and $v_y^{(t)}$ indicate the location change in" + ], + "bbox": [ + 210, + 579, + 782, + 838 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Y. Guo et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "image space, and $\\rho^{(t)}$ indicates the ratio between the target depth and the source depth. Specifically, SUP-NeRF updates the translation in image space", + "bbox": [ + 227, + 145, + 782, + 176 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and depth as:", + "bbox": [ + 228, + 176, + 330, + 191 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nx ^ {(t + 1)} = x ^ {(t)} + v _ {x} ^ {(t)}\n$$\n", + "text_format": "latex", + "bbox": [ + 442, + 183, + 575, + 200 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\ny ^ {(t + 1)} = y ^ {(t)} + v _ {y} ^ {(t)} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 442, + 204, + 784, + 223 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nZ ^ {(t + 1)} = \\rho^ {(t)} Z ^ {(t)}\n$$\n", + "text_format": "latex", + "bbox": [ + 439, + 226, + 558, + 243 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $(x,y)$ represent the object's 2D location in image space, $Z$ indicates the depth. The new 3D location $T_{o2c}^{(t + 1)}$ is then computed from the updated image coordinates, depth and camera intrinsic matrix $K$ . Compared to [21, 34], our definition of $\\rho^{(t)}$ leads to a more straight-forward update.", + "bbox": [ + 212, + 253, + 782, + 316 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "During the pose updating process indicated by $\\oplus$ , both the rotation and translation are updated. Through such representation and updating scheme, SUP-NeRF handles objects at varying distances with the same output space. Further discusses the choice of an effective pose representation and the right coordinate frame are included in Appendices A1 and A2.", + "bbox": [ + 212, + 316, + 782, + 391 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Cross-task Shortcut. Simultaneously perceiving pose, shape, and texture, requires an image encoder with shared initial layers but separate later stages. However, encoding shape, texture, and pose is a challenge as they require different features, making it tough for a shared encoder to meet all needs effectively. Shape and texture encoding demand a pose-invariant approach, as they correspond to each object's frame. On the other hand, pose encoding needs to capture sufficient pose information for later use. This difference leads to conflicting learning signals during joint training, potentially weakening the model's performance.", + "bbox": [ + 212, + 393, + 784, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "SUP-NeRF proposes a straightforward one-step approach to resolve this conflict. After processing through the first four CNN layers separately, SUP-NeRF introduces a cross-task shortcut that deducts pose features from the shape and texture features. This design choice eliminates the shape and texture's reliance on pose, allowing the earlier layers to focus on pose-dependent features for all tasks. This shortcut turns the conflict into a cooperative relationship, enhancing the joint training of different modules, as demonstrated in Tab. 7. The details of our image encoder is described in Appendix A3.", + "bbox": [ + 212, + 513, + 784, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Training. SUP-NeRF closely follows the loss functions and occupancy mask definitions from AutoRF [29]. The total loss during training is: $L_{train} = L_{rgb} + w_{occ}L_{occ} + w_{pose}(L_{pose}^{(direct)} + \\sum_t^K L_{pose}^{(t)})$ , where $w_{occ}$ and $w_{pose}$ indicate the corresponding weights.", + "bbox": [ + 212, + 635, + 782, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare the RGB output to the input image to calculate photometric loss $L_{rgb}$ , and the occupancy values to the mask for occupancy loss $L_{occ}$ . SUP-NeRF additionally compares the projected box corners from the integrative pose estimator $B^{(t)}_{proj}$ with the ground-truth values to compute the pose losses $L_{pose}^{(t)}$ . We define the pose loss function as the mean distance between the matched corners: $L_{pose}^{(*)} = \\frac{1}{8}\\sum_{i=1}^{8}\\sqrt{(x_i^{(*)} - \\hat{x}_i)^2 + (y_i^{(*)} - \\hat{y}_i)^2}$ , where (*) indicates it compatible with both the direct pose loss and the iterative pose loss, $(x_i^{(*)}, y_i^{(*)})$ denote a input corner position, and $(\\hat{x}_i, \\hat{y}_i)$ indicates a ground-truth corner po", + "bbox": [ + 212, + 698, + 784, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "SUP-NeRF", + "bbox": [ + 653, + 113, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sition. SUP-NeRF introduces an additional MLP layer for more accurate pose estimation to directly supervise the 2D coordinates of box corners predicted from the pose code.", + "bbox": [ + 212, + 145, + 782, + 191 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In training, only the network weights are treated as the estimizable variables. Moreover, the NeRF decoder uses ground-truth camera poses $[R_{c2o}^{gt}|T_{c2o}^{gt}]$ instead of predictions from the pose estimation module to ensure more reliable training. Implementation Details. SUP-NeRF's image encoder backbone is based on ResNet50, with shared early layers and separated later layers for pose, shape, and texture encoding. For more information on this, please refer to the Appendix A3. To implement the NeRF decoder, we use CodeNeRF [12] for its simplicity and ease of use. For the differentiable rendering based on NeRF, we set the latent dimension to 256, use 64 samples along each ray, and normalize the shape space by the diagonal length of the observed object's 3D box. For the pose estimation module, we also set the latent dimension to 256 for all MLP layers. Before feeding the input image and occupancy mask into the image encoder, we pad them to make them square and resize them to $128\\times 128$ pixels. We then render $32\\times 32$ images and occupancy maps to calculate losses.", + "bbox": [ + 212, + 191, + 785, + 402 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To ensure effective iterative pose estimation without sacrificing generalization, SUP-NeRF uses a heuristic to sample a random initial 3D object pose across the entire pose space. For translation, we sample in the relative space. We sample a random 2D location within the image ROI range as the initial projection of the object center, while setting the initial depth to $Z = 20$ . For rotation, we sample from the full range of yaw angles combined with a random rotation angle within a $(-20^{\\circ}, 20^{\\circ})$ range around a random axis direction. This sampling method covers almost the entire distribution of object poses in the camera frame and is used consistently in both training and inference.", + "bbox": [ + 212, + 404, + 785, + 539 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "During training, we use 3 iterations of the pose estimator to generate corresponding pose losses, which allow for pose regression to handle highly diverse poses. We jointly train the unified model for the pose estimation module, image encoder, and NeRF decoder for 40 epochs on the nuScenes dataset using a learning rate of $10^{-4}$ and loss weights of $w_{occ} = 0.1$ , $w_{pos} = 0.01$ to compute the total loss.", + "bbox": [ + 212, + 540, + 785, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Inference. During inference, SUP-NeRF optimizes the shape code, texture code, and object pose $[R_{o2c}|T_{o2c}]$ keeping the network weights frozen. The total loss $L_{infer}$ is thus $L_{infer} = L_{rgb} + w_{occ}L_{occ}$ , where $w_{occ}$ balances the two. SUP-NeRF minimizes the loss $L_{infer}$ to update the variables. At inference, we run 3 iterations for the feed-forward pose estimation module without updating the shape and color codes, and 50 iterations for the gradient-based joint optimization of shape, texture, and pose. We use a step size of 0.02 for shape and texture code updates and 0.01 for pose updates during joint optimization.", + "bbox": [ + 212, + 630, + 785, + 752 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 776, + 375, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our experiments use three datasets: nuScenes [3], KITTI [6] and Waymo [38]. Our evaluation framework simultaneously measures rendering quality, shape reconstruction accuracy, and pose estimation precision.", + "bbox": [ + 212, + 792, + 782, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Y. Guo et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Data Splits. Since these datasets are not specifically for 3D reconstruction, we sifted a subset for better evaluation. We chose daytime sequences and used Mask R-CNN [7] to get instance masks, as nuScenes lacks 2D segmentation masks. We matched these masks with 3D bounding box annotations, categorizing them into foreground, background, and unknown, similar to AutoRF [29]. More details are in Appendix A7. We use the following splits of these datasets:", + "bbox": [ + 212, + 146, + 787, + 236 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- nuScenes Train Split: We collected 87,048 objects from the nuScenes Train split for training.", + "- nuScenes Val Split: We collected 5,000 random objects from the nuScenes Val split for evaluation.", + "- KITTI Val Split: We use 4,895 objects from KITTI Val split [6] for evaluation.", + "- Waymo Val Split: We randomly selected 5,000 objects from Waymo Val split [16] for evaluation." + ], + "bbox": [ + 209, + 238, + 782, + 340 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Evaluation metrics. Our evaluation framework differs from traditional NeRF assessments as it simultaneously evaluates object reconstruction quality and pose accuracy. We incorporate two monocular reconstruction metrics: (1) the PSNR (Peak Signal-to-Noise Ratio) comparing the rendered and observed images; (2) the Depth Error between the rendered depth map and sparse LiDAR data, denoted as $DE$ . We also include two pose estimation metrics: (3) rotation error denoted as $RE$ ; (4) the translation error denoted as $TE$ . To offer a thorough evaluation, we compare various methods at different stages: post-feed-forward ( $FF$ ), and at the 50th iterations of NeRF joint optimization. Beyond monocular assessments, we also conduct cross-view evaluations for PSNR and depth error metrics, denoted as (5) PSNR- $C$ and (6) $DE-C$ , focusing on objects observed from novel views. As the nuScenes dataset provides object IDs that allows tracking objects across various views, the cross-view evaluation is exclusively conducted on the nuScenes dataset, as explained in Fig. 5.", + "bbox": [ + 212, + 344, + 787, + 555 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Baselines. We compare SUP-NeRF with two baselines: AutoRF [29] and AutoRF [29] + FCOS3D [42] (selected for its monocular capability, and availability of model trained on the nuScenes dataset). Both AutoRF and SUP-NeRF use the same NeRF decoder and image encoder to ensure a fair comparison. The difference lies in SUP-NeRF's additional layers for pose encoding and refinement. During testing, we apply the same random pose generation for both AutoRF and SUP-NeRF methods, whereas the FCOS3D baseline begins with its own pose prediction. We also compare with BootInv [34], which recovers pose, shape, and appearance. Leveraging our flexible design, we developed SUP-BootInv by integrating BootInv's pretrained decoder with our encoder and pose estimator.", + "bbox": [ + 212, + 556, + 787, + 708 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 nuScenes Monocular Reconstruction and Pose Estimation", + "text_level": 1, + "bbox": [ + 214, + 734, + 736, + 750 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Tab. 2 shows that SUP-NeRF outperforms two baselines, AutoRF and its combination with FCOS3D on nuScenes Val set. AutoRF alone struggles to improve pose from a random start, resulting in lower PSNR and depth error. SUP-NeRF consistently reduces depth error by over $25\\%$ , rotation error by over $30\\%$ , and translation error by over $13\\%$ compared to the FCOS3D and AutoRF com", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "SUP-NeRF", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/0745e9c0ac0d90a24e2cc15080bcb06af2a31af3a07cf0df732a9ed3caf01138.jpg", + "table_caption": [ + "Table 2: nuScenes Monocular Reconstruction and Pose Estimation Results. SUP-NeRF consistently improves AutoRF-based pipelines in all metrics, particularly in the pose estimation metrics. SUP-BootInv also consistently improves BootInv in all metrics. [Key: Best, Second Best, FF= Feed Forward, C= Cross-View]" + ], + "table_footnote": [], + "table_body": "
MethodPSNR (↑)DE (m) (↓)RE (deg.) (↓)TE (m) (↓)PSNR-C (↑)DE-C (m) (↓)
FF|50itFF|50itFF|50itFF|50itFF|50itFF|50it
AutoRF [29]3.6|10.611.21|10.0987.52|88.076.04|5.9510.0|8.81.31|1.41
AutoRF+FCOS7.5|17.21.34|0.819.77|10.170.85|0.789.8|10.51.29|1.16
SUP-NeRF10.5|18.80.69|0.607.01|7.070.68|0.7310.6|10.91.22|1.14
BootInv [34]9.4|14.35.01|3.5628.40|28.002.59|2.9110.9|11.81.37|1.35
SUP-BootInv10.9|15.41.95|1.627.11|8.400.64|1.0010.9|11.91.37|1.40
", + "bbox": [ + 232, + 212, + 771, + 301 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/dd917c13bc6afe2dfae83e07b615f4518207d87f7654f1d560ccb26fb6b4aabf.jpg", + "image_caption": [ + "Fig. 5: nuScenes Cross-view Evaluation. Each row shows a set of images of the same object from different angles. We use the example for both monocular pose estimation and NeRF reconstruction. Other images in each row are used to evaluate the reconstructed shape and texture in PSNR and Depth Error (DE)." + ], + "image_footnote": [], + "bbox": [ + 233, + 316, + 769, + 431 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "bination. Our feed-forward stage also shows promising potential for real-time applications, especially where rendering quality is less critical.", + "bbox": [ + 212, + 530, + 784, + 560 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We next compare BootInv [34] as the baseline, although the performance of original BootInv's pose estimation was subpar on both the nuScenes and KITTI datasets, integrating our robust pose module with this NeRF framework significantly enhanced pose estimation accuracy. This underscores the adaptability of our unified pipeline in augmenting various object NeRF frameworks for improved pose estimation. Moreover, comparisons of BootInv [34] and SUP-BootInv in terms of monocular PSNR and depth error suggest that SUP-BootInv's superior pose estimation contributes to its higher PSNR and lower depth error. More detailed analysis upon BootInv-based pipelines is included in Appendix A5.", + "bbox": [ + 212, + 561, + 787, + 696 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Cross-Dataset Evaluation", + "text_level": 1, + "bbox": [ + 214, + 720, + 472, + 734 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We next compare the generalization capability of all these NeRF models with the cross-dataset experiments. For these experiments, we directly apply nuScenes models to the KITTI and Waymo datasets and report the metrics.", + "bbox": [ + 212, + 750, + 784, + 792 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "KITTI. Tab. 3 shows the result of cross-dataset experiments on the KITTI dataset. It shows that SUP-NeRF achieves SoTA performance across all evaluation metrics. SUP-NeRF reduces depth error by about $16\\%$ , rotation error by", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Y. Guo et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/2df3669c1d60c1182fcf2fad7a63fa1b16a5ec02bc91a3c2fc689ef90a4091a4.jpg", + "table_caption": [ + "Table 3: KITTI Cross-dataset Monocular Reconstruction and Pose Estimation Results. We train all methods on nuScenes dataset, and test on KITTI dataset. Our methods SUP-NeRF and SUP-BootInv consistently show superiority in all metrics compared to the counterpart methods. [Key: Best, Second Best, FF= Feed Forward]" + ], + "table_footnote": [], + "table_body": "
MethodPSNR (↑)DE (m) (↓)RE (deg.) (↓)TE (m) (↓)
FF|50itFF|50itFF|50itFF|50it
AutoRF [29]0.4|7.69.89|8.8389.67|90.426.16|6.07
AutoRF+FCOS2.4|13.42.42|1.7411.95|12.52.2|2.09
SUP-NeRF4.0|14.12.19|1.546.79|6.891.06|1.01
BootInv [34]7.0|12.36.11|4.3815.52|15.763.97|3.90
SUP-BootInv7.6|12.72.79|2.219.16|8.841.15|1.67
", + "bbox": [ + 297, + 212, + 705, + 301 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/bbad4520302c52591efcc56bda809535229a58ab55cfac0ebe5b56cbf23fc5a8.jpg", + "image_caption": [ + "Fig. 6: KITTI (Cross-dataset) Qualitative Results. In the top panel, we demonstrate SUP-NeRF executes pose estimation reliably, fast converging from a random initial pose to the true one, and enables neural reconstruction under diverse object poses, occlusion cases under this cross-dataset setup. In the bottom panel, SUP-NeRF is visually compared to the other major competitor, demonstrates sharper rendered image, higher accuracy in shape and pose." + ], + "image_footnote": [], + "bbox": [ + 232, + 323, + 772, + 536 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "as large as $50\\%$ and translation error also about $50\\%$ . Thus, SUP-NeRF shows significantly better geometric performance in cross-dataset setting compared to AutoRF+FCOS, which we attributes to our deep pose module's independence from camera intrinsic parameters. Meanwhile, compared to nuScenes evaluation, all the methods' PSNR and depth error performance drop, which means the cross-dataset generalization of monocular NeRF remains an open challenge.", + "bbox": [ + 212, + 669, + 787, + 760 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Fig. 6 presents the qualitative results on this experiment, demonstrating how SUP-NeRF handles diverse poses and occlusions in pose refinement and the visual comparisons between the two most competitive pipelines. Consistent improvements upon integrating our design into BootInv baseline were also observed. More visual comparisons are in Appendix A9.", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "SUP-NeRF", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/fe8a772b3b84190b5db7d74bd592b4efed8cc770ff1000a8728e4ce335aa3728.jpg", + "table_caption": [ + "Table 4: Waymo Cross-dataset Monocular Reconstruction and Pose Estimation Results. We train all methods on nuScenes dataset, and test on Waymo dataset. Our methods SUP-NeRF and SUP-BootInv consistently show superiority in all metrics compared to the counterpart methods, the only exception is the comparison to AutoRF+FCOS in Rotation Error. [Key: Best, Second Best, FF= Feed Forward]" + ], + "table_footnote": [], + "table_body": "
MethodPSNR (↑)DE (m) (↓)RE (deg.) (↓)TE (m) (↓)
FF|50itFF|50itFF|50itFF|50it
AutoRF [29]0.6|9.86.76|6.5386.56|87.679.1|9.14
AutoRF+FCOS6.2|16.52.43|2.326.65|7.23.22|3.31
SUP-NeRF4.8|17.02.32|1.5610.01|10.61.68|1.54
BootInv [34]8.1|11.08.85|8.2030.78|31.525.26|6.16
SUP-BootInv8.6|11.95.36|4.3810.24|11.041.67|2.53
", + "bbox": [ + 297, + 226, + 705, + 315 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Waymo. Tab. 4 shows the result of cross-dataset experiments on the Waymo dataset. It shows that SUP-NeRF achieves SoTA performance across most evaluation metrics. The only exception is the superior rotation estimation achieved by FCOS3D, which is likely due to leveraging larger context for heavily occluded cases. However, our advantage in translation estimation is more significant.", + "bbox": [ + 212, + 342, + 787, + 417 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Ablation Studies", + "text_level": 1, + "bbox": [ + 214, + 438, + 401, + 452 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The major ablation studies we conduct includes (i) the impact of pose estimation module designs and (ii) the impact of different object NeRF networks. Other ablation studies to analyze the impact of training with predicted 2D boxes, pose refinement iterations, training epochs, as well as NeRF's capability on handling initial pose errors are included in Appendix A6, owing to lack of space.", + "bbox": [ + 212, + 462, + 787, + 537 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Pose Estimation Module. We evaluate multiple pose estimation designs from two main angles. First, we focus on object-centric pose estimation task alone with evaluations on nuScenes, KITTI, and Waymo datasets. The prototypical designs include: (i) Direct 6D pose regression using MLP layers; (ii) Pose estimation from predicted box corners via Keypoint-based Perspective-n-Point (PnP); (iii) NOCS-based pose estimation as implemented in BootInv [34]; (vi) External monocular 3D detection FCOS3D [42]; (v) Our iterative pose refinement method, which uses projected boxes. For fairness, methods (1), (2), and (5) share the same image encoder, pose encoding layers, differing only in their output layers. Tab. 5 shows that SUP-NeRF consistently outperforms the others, notably in cross-dataset translation estimation.", + "bbox": [ + 212, + 537, + 787, + 703 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The second evaluation assesses the impact of different initial poses on object-NeRF performance within the SUP-NeRF architecture. We additionally compared (i) Ground-Truth (GT) pose; (ii) Random initial pose, to represent the oracle and worst case in pose initialization, and exclude the NOCS method which is part of a different NeRF framework. Given different initial poses, all methods undergo the same NeRF Gradient-based Pose Refinement (NGPR), evaluated immediately and after 50 iterations. Tab. 6 shows that our method significantly surpasses other initial pose estimators in geometric accuracy, demonstrating its effectiveness in enhancing 3D model quality.", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Y. Guo et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/583b85efe7887294a7633870bb3516d84731440640f8f1b5b116047c3c5c0316.jpg", + "table_caption": [ + "Table 5: Pose Estimation Ablation Studies. We evaluate all the candidate pose estimation methods with a focus on pose accuracy alone, on nuScenes, KITTI, and Waymo validation sets. Our method shows outstanding robustness in all cases, particularly in translation accuracy in cross-dataset tests. [Key: Best]" + ], + "table_footnote": [], + "table_body": "
Pose ModuleRE (deg.)(↓)TE (m) (↓)
nuScenesKITTIWaymonuScenesKITTIWaymo
MLP Pose31.9637.5830.895.535.3110.1
Corners+PnP24.7935.1093.762.823.5518.68
NOCS +PnP(BootInv)28.4015.5223.662.593.973.47
FCOS3D9.7711.956.650.852.23.22
SUP-NeRF7.016.7910.010.681.061.54
", + "bbox": [ + 287, + 212, + 714, + 300 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/cbbec98dd334000d859685ce85bd8ea8a676facf038dcbcce8258332916e0bed.jpg", + "table_caption": [ + "Table 6: Ablation studies of Pose Initialization in Object NeRF on nuScenes. We compare different choices of initial poses under SUP-NeRF frameworks. As shown, SUP-NeRF presents superiors effectiveness in all the metrics. [Key: Best, FF=Feed Forward, C=Cross-View]" + ], + "table_footnote": [], + "table_body": "
Initial PosePSNR (↑)DE (m) (↓)RE (deg.) (↓)TE (m) (↓)PSNR-C (↑)DE-C (m) (↓)
FF|50itFF|50itFF|50itFF|50itFF|50itFF|50it
Random3.6|10.611.21|10.0987.52|88.076.04|5.959.1|7.41.42|1.82
MLP Pose1.2|6.95.33|4.4131.96|24.575.53|2.789.8|8.51.3|1.4
Corners+PnP7.2|16.22.18|1.8324.79|24.572.82|2.789.9|9.51.3|1.3
FCOS3D7.5|17.21.34|0.819.77|10.170.85|0.789.8||10.51.29||1.16
SUP-NeRF10.5|18.80.69|0.607.01|7.070.68|0.7310.6|10.91.22|1.14
GT10.0|18.80.66|0.530.2|1.110.0|1.310.9|11.11.19|1.14
", + "bbox": [ + 243, + 385, + 756, + 486 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Tab. 6 also confirms the role of NeRF Gradient-Based Pose Refinement (NGPR). By comparing the difference between feed-forward results and the later ones, one can find NGPR merely reduces rotation error or translation error. On the contrary, it can even increase pose error when the initial pose error is small enough, e.g. rotation error $< 7^{\\circ}$ , and translation error $< 0.75m$ . On the other hand, the improvements of PSNR and depth reconstruction over time indicate that NeRF optimization mostly focuses on shape and texture optimization w.r.t. the visible surface regardless the pose.", + "bbox": [ + 215, + 518, + 784, + 638 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Impact of object NeRF framework. To demonstrate SUP-NeRF's benefits, we compared different network configurations using the same pose refinement approach. For fairness, we used identical subnetwork modules across all configurations. These include a ResNet-based Image Encoder (E), a NeRF Decoder (D), and our Pose Estimation Module (P). CodeNeRF [12] has just a Decoder (D), AutoRF combines an Image Encoder and Decoder $(\\mathrm{E} + \\mathrm{D})$ , and SUP-NeRF uses all three $(\\mathrm{E} + \\mathrm{D} + \\mathrm{P})$ . For CodeNeRF and AutoRF, we added a separately trained Pose Network (PNet, composed of $\\mathrm{E} + \\mathrm{P}$ ) for initial pose refinement. We also tested a SUP-NeRF version without the cross-task Short Cut (SC) to assess its impact. Tab. 7 shows that SC consistently improves SUP-NeRF across all metrics on nuScenes.", + "bbox": [ + 215, + 641, + 785, + 806 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "SUP-NeRF, with joint training, outperforms separate training setups AutoRF + PNet and CodeNeRF + PNet in monocular metrics. Interestingly, while", + "bbox": [ + 215, + 809, + 784, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "SUP-NeRF", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/beaf91a1862be14a7f350994c6c4720ff879e2ea67a7ea6b5c940e571212bc7c.jpg", + "table_caption": [ + "Table 7: Object NeRF Framework Ablations on nuScenes. We used the same subnetwork architectures - the image encoder (E), NeRF decoder (D), and pose estimation module (P) - to create various model configurations. For SUP-NeRF, we also included a version without the cross-task Short Cut (SC) for comparison. [Key: Best, FF=Feed Forward, C=Cross-View]" + ], + "table_footnote": [], + "table_body": "
Model TrainingPSNR(▲)DE (m)(↓)RE (deg.) (↓)TE (m)(↓)PSNR-C(▲)DE-C (m)(↓)
FF|50itFF|50itFF|50itFF|50itFF|50itFF|50it
CodeNeRF (D)+PNet(E+P)7.6|16.00.87|0.6910.57|9.750.72|0.7611.0|11.21.01|1.04
AutoRF (E+D)+PNet(E+P)9.7|17.50.79|0.668.26|7.870.7|0.745.6|10.11.24|1.16
SUP-NeRF (E+D+P) w/o SC8.7|18.00.94|0.7310.12|9.220.75|0.7410.8|10.91.23|1.11
SUP-NeRF (E+D+P)10.5|18.80.69|0.607.01|7.070.68|0.7310.6|10.91.22|1.14
", + "bbox": [ + 222, + 226, + 781, + 299 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/5bddc11ab308ee33e12767ee11e721a1b0c584b524d668d143619266e794e71a.jpg", + "table_caption": [ + "Table 8: Model Size and Running Time Comparison in feed-forward scenarios, and with 20 and 50 iterations of NeRF optimization. SUP-NeRF gets the smallest running time. [Key: Best, FF= Feed Forward]" + ], + "table_footnote": [], + "table_body": "
MethodParams (M) (↓)FF (s) (↓)FF+20it (s) (↓)FF+50it (s) (↓)
AutoRF+FCOS [29]91.1160.1230.7141.599
(36.166+54.950)(0.114+0.009)(0.114+0.600)(0.114+1.485)
SUP-NeRF49.8160.0180.6081.493
BootInv [34]182.6160.1563.5348.601
SUP-BootInv57.5800.0183.3968.463
", + "bbox": [ + 259, + 364, + 743, + 444 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "CodeNeRF lags in monocular, it excels in cross-view evaluations. This implies that using mean shape codes, rather than current image encodings, might reduce overfitting to current view.", + "bbox": [ + 212, + 469, + 787, + 516 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "4.4 Running Speed Analysis", + "text_level": 1, + "bbox": [ + 214, + 536, + 465, + 553 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Tab. 8 compares the model size and running speed on a single A5000 Graphics card. Detailed setup and analysis details are in Appendix A8. Tab. 8 shows that SUP-NeRF needs about half or one third as many parameters as others and is 6-8 times faster in the feed-forward (FF) stage. Meanwhile, a significant portion of the run time is consumed by the NeRF optimization step, whether for 20 iterations (20it) or 50 iterations (50it). However, speed limitation of NeRF can be effectively tackled by recent advances in neural rendering [15,31], which we consider as orthogonal to our contribution to the feed-forward stage.", + "bbox": [ + 212, + 560, + 787, + 681 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 703, + 359, + 720 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This paper introduces SUP-NeRF: a unified network that seamlessly integrates pose estimation and object reconstruction. SUP-NeRF includes a novel pose estimation module to handle scale-depth ambiguity and a new representation invariant to camera changes. Consequently, SUP-NeRF greatly improves the robustness of object pose estimation compared to the standard NeRF baseline. SUP-NeRF outperforms previous methods in both monocular reconstruction and pose estimation tasks, especially in challenging cross-dataset applications.", + "bbox": [ + 212, + 733, + 787, + 842 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Y. Guo et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Batra, D., Chang, A.X., Chernova, S., Davison, A.J., Deng, J., Koltun, V., Levine, S., Malik, J., Mordatch, I., Mottaghi, R., Savva, M., Su, H.: Rearrangement: A challenge for embodied AI. CoRR (2020) 2", + "2. Bian, W., Wang, Z., Li, K., Bian, J.: Nope-nerf: Optimising neural radiance field with no pose prior. In: CVPR (2023) 4", + "3. Caesar, H., Bankiti, V., Lang, A., Vora, S., Liong, V., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuScenes: A multimodal dataset for autonomous driving. In: CVPR (2020) 3, 4, 8", + "4. Chen, H., Manhardt, F., Navab, N., Busam, B.: Texpose: Neural texture learning for self-supervised 6d object pose estimation. In: CVPR (2023) 4", + "5. Chen, Y., Lee, G.H.: DBARF: deep bundle-adjusting generalizable neural radiance fields. In: CVPR (2023) 4", + "6. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the KITTI vision benchmark suite. In: CVPR (2012) 4, 8, 9", + "7. Girshick, R., Radosavovic, I., Gkioxari, G., Dollar, P., He, K.: Detectron. https://github.com/facebookresearch/detectron (2018) 9", + "8. Gkioxari, G., Johnson, J., Malik, J.: Mesh R-CNN. In: ICCV (2019) 2, 3", + "9. Henderson, P., Tsiminaki, V., Lampert, C.H.: Leveraging 2d data to learn textured 3d mesh generation. In: CVPR (2020) 2", + "0. Hodañ, T., Michel, F., Brachmann, E., Kehl, W., Glent Buch, A., Kraft, D., Drost, B., Vidal, J., Ihrke, S., Zabulis, X., Sahin, C., Manhardt, F., Tombari, F., Kim, T.K., Matas, J., Rother, C.: BOP: Benchmark for 6D object pose estimation. ECCV (2018) 4", + "1. Insafutdinov, E., Campbell, D., Henriques, J.F., Vedaldi, A.: Snes: Learning probably symmetric neural surfaces from incomplete data. In: Avidan, S., Brostow, G.J., Cissé, M., Farinella, G.M., Hassner, T. (eds.) ECCV. pp. 367-383 (2022) 2, 3", + "2. Jang, W., Agapito, L.: Codenerf: Disentangled neural radiance fields for object categories. In: ICCV (2021) 2, 3, 8, 13", + "3. Kanazawa, A., Tulsiani, S., Efros, A.A., Malik, J.: Learning category-specific mesh reconstruction from image collections. In: Ferrari, V., Hebert, M., Sminchisescu, C., Weiss, Y. (eds.) ECCV (2018) 2, 3", + "4. Kehl, W., Manhardt, F., Tombari, F., Ilic, S., Navab, N.: SSD-6D: making rgb-based 3d detection and 6d pose estimation great again. In: ICCV (2017) 4, 6", + "5. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph. (2023) 14", + "6. Kumar, A., Brazil, G., Corona, E., Parchami, A., Liu, X.: DEVIANT: depth equivariant network for monocular 3d object detection. In: ECCV (2022) 2, 4, 9", + "7. Kumar, A., Guo, Y., Huang, X., Ren, L., Liu, X.: SeaBird: Segmentation in Bird's View with Dice Loss Improves Monocular 3D Detection of Large Objects. In: CVPR (2024) 2", + "8. Kundu, A., Genova, K., Yin, X., Fathi, A., Pantofaru, C., Guibas, L.J., Tagliasacchi, A., Dellaert, F., Funkhouser, T.A.: Panoptic neural fields: A semantic object-aware neural scene representation. In: CVPR (2022) 2, 3, 4", + "9. Kundu, A., Li, Y., Rehg, J.M.: 3d-rcnn: Instance-level 3d object reconstruction via render-and-compare. In: CVPR (2018) 2, 3", + "20. Li, F., Yu, H., Shugurov, I., Busam, B., Yang, S., Ilic, S.: Nerf-posed: A first-reconstruct-then-regress approach for weakly-supervised 6d object pose estimation. CoRR (2022) 4" + ], + "bbox": [ + 225, + 176, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "SUP-NeRF", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "21. Li, Y., Wang, G., Ji, X., Xiang, Y., Fox, D.: Deepim: Deep iterative matching for 6d pose estimation. In: Ferrari, V., Hebert, M., Sminchisescu, C., Weiss, Y. (eds.) ECCV (2018) 4, 6, 7", + "22. Li, Y., Wang, G., Ji, X., Xiang, Y., Fox, D.: Deepim: Deep iterative matching for 6d pose estimation. In: ECCV (2018) 4", + "23. Lin, Y., Florence, P., Barron, J.T., Rodriguez, A., Isola, P., Lin, T.: inerf: Inverting neural radiance fields for pose estimation. In: IROS (2021) 3", + "24. Lipson, L., Teed, Z., Goyal, A., Deng, J.: Coupled iterative refinement for 6d multi-object pose estimation. In: CVPR (2022) 4", + "25. Lu, Y., Ma, X., Yang, L., Zhang, T., Liu, Y., Chu, Q., Yan, J., Ouyang, W.: Geometry uncertainty projection network for monocular 3d object detection. In: ICCV (2021) 2, 4", + "26. Merrill, N., Guo, Y., Zuo, X., Huang, X., Leutenegger, S., Peng, X., Ren, L., Huang, G.: Symmetry and uncertainty-aware object SLAM for 6 dof object pose estimation. In: CVPR (2022) 4", + "27. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM pp. 99-106 (2021) 2, 4", + "28. Mousavian, A., Anguelov, D., Flynn, J., Kosecka, J.: 3d bounding box estimation using deep learning and geometry. In: CVPR (2017) 4", + "29. Müller, N., Simonelli, A., Porzi, L., Bulò, S.R., Nießner, M., Kontschieder, P.: Autorf: Learning 3d object radiance fields from single view observations. In: CVPR (2022) 2, 3, 4, 6, 7, 9, 10, 11, 12, 14", + "30. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph. pp. 102:1-102:15 (2022) 2", + "31. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph. (2022) 14", + "32. Park, D., Ambrus, R., Guizilini, V., Li, J., Gaidon, A.: Is pseudo-lidar needed for monocular 3d object detection? In: ICCV (2021) 4", + "33. Park, K., Patten, T., Vincze, M.: Pix2pose: Pix2pose: Pixel-wise coordinate regression of objects for 6d pose estimation. In: ICCV (2019) 4", + "34. Pavllo, D., Tan, D.J., Rakotosaona, M.J., Tombari, F.: Shape, pose, and appearance from a single image via bootstrapped radiance field inversion. In: CVPR (2023) 2, 3, 4, 7, 9, 10, 11, 12, 14", + "35. Rad, M., Lepetit, V.: Bb8: A scalable, accurate, robust to partial occlusion method for predicting the 3d poses of challenging objects without using depth. ICCV (2017) 4", + "36. Simonelli, A., Bulò, S.R., Porzi, L., Lopez-Antequera, M., Kontschieder, P.: Disentangling monocular 3d object detection. In: ICCV (2019) 2, 4", + "37. Solà, J., Deray, J., Atchuthan, D.: A micro lie theory for state estimation in robotics. CoRR (2018) 6", + "38. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., Vasudevan, V., Han, W., Ngiam, J., Zhao, H., Timofeev, A., Ettinger, S., Krivokon, M., Gao, A., Joshi, A., Zhang, Y., Shlens, J., Chen, Z., Anguelov, D.: Scalability in perception for autonomous driving: Waymo open dataset. In: CVPR (2020) 8", + "39. Tekin, B., Sinha, S.N., Fua, P.: Real-time seamless single shot 6d object pose prediction. In: CVPR (2018) 4", + "40. Tian, S., Cai, Y., Yu, H.X., Zakharov, S., Liu, K., Gaidon, A., Li, Y., Wu, J.: Multi-object manipulation via object-centric neural scattering functions. In: CVPR (2023) 2" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Y. Guo et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "41. Wang, H., Sridhar, S., Huang, J., Valentin, J., Song, S., Guibas, L.J.: Normalized object coordinate space for category-level 6d object pose and size estimation. In: CVPR (2019) 4", + "42. Wang, T., Zhu, X., Pang, J., Lin, D.: FCOS3D: fully convolutional one-stage monocular 3d object detection. In: ICCVW (2021) 4, 9, 12", + "43. Weihs, L., Deitke, M., Kembhavi, A., Mottaghi, R.: Visual room rearrangement. In: CVPR (June 2021) 2", + "44. Xiang, Y., Schmidt, T., Narayanan, V., Fox, D.: PoseCNN: A convolutional neural network for 6D object pose estimation in cluttered scenes. In: RSS (2018) 4", + "45. Yang, B., Zhang, Y., Xu, Y., Li, Y., Zhou, H., Bao, H., Zhang, G., Cui, Z.: Learning object-compositional neural radiance field for editable scene rendering. In: ICCV (2021) 2, 3", + "46. Yu, A., Ye, V., Tancik, M., Kanazawa, A.: pixelnerf: Neural radiance fields from one or few images. In: CVPR (2021) 2, 3", + "47. Yu, H., Wu, J., Yi, L.: Rotationally equivariant 3d object detection. In: CVPR (2022) 4", + "48. Zakharov, S., Ambrus, R.A., Guizilini, V., Park, D., Kehl, W., Durand, F., Tenenbaum, J.B., Sitzmann, V., Wu, J., Gaidon, A.: Single-shot scene reconstruction. In: CoRL (2021) 2, 3", + "49. Zakharov, S., Kehl, W., Bhargava, A., Gaidon, A.: Autolabeling 3d objects with differentiable rendering of SDF shape priors. In: CVPR (2020) 2, 3", + "50. Zakharov, S., Shugurov, I., Ilic, S.: Dpod: 6d pose object detector and refiner. In: ICCV (2019) 4", + "51. Zhou, A., Kim, M.J., Wang, L., Florence, P., Finn, C.: Nerf in the palm of your hand: Corrective augmentation for robotics via novel-view synthesis. In: CVPR (2023) 2" + ], + "bbox": [ + 212, + 146, + 787, + 507 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "SUP-NeRF", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/SUP-NeRF_ A Streamlined Unification of Pose Estimation and NeRF for Monocular 3D Object Reconstruction/c913b291-07c4-425a-807f-281296d9ef23_model.json b/2024/SUP-NeRF_ A Streamlined Unification of Pose Estimation and NeRF for Monocular 3D Object Reconstruction/c913b291-07c4-425a-807f-281296d9ef23_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0313ebfc65374ddf7a52c9124331e32ae2110071 --- /dev/null +++ b/2024/SUP-NeRF_ A Streamlined Unification of Pose Estimation and NeRF for Monocular 3D Object Reconstruction/c913b291-07c4-425a-807f-281296d9ef23_model.json @@ -0,0 +1,2335 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.228, + 0.141, + 0.777, + 0.207 + ], + "angle": 0, + "content": "SUP-NeRF: A Streamlined Unification of Pose Estimation and NeRF for Monocular 3D Object Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.291, + 0.234, + 0.398, + 0.251 + ], + "angle": 0, + "content": "Yuliang Guo\\(^{1\\star}\\)" + }, + { + "type": "text", + "bbox": [ + 0.294, + 0.252, + 0.395, + 0.266 + ], + "angle": 0, + "content": "Ruoyu Wang" + }, + { + "type": "text", + "bbox": [ + 0.44, + 0.235, + 0.575, + 0.25 + ], + "angle": 0, + "content": "Abhinav Kumar1,2" + }, + { + "type": "text", + "bbox": [ + 0.459, + 0.252, + 0.562, + 0.266 + ], + "angle": 0, + "content": "Xinyu Huang" + }, + { + "type": "text", + "bbox": [ + 0.616, + 0.235, + 0.711, + 0.25 + ], + "angle": 0, + "content": "Cheng Zhao1" + }, + { + "type": "text", + "bbox": [ + 0.632, + 0.252, + 0.698, + 0.265 + ], + "angle": 0, + "content": "Liu Ren1" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.27, + 0.774, + 0.285 + ], + "angle": 0, + "content": "1Bosch Research North America & Bosch Center for Artificial Intelligence (BCAI)" + }, + { + "type": "text", + "bbox": [ + 0.411, + 0.285, + 0.593, + 0.299 + ], + "angle": 0, + "content": "2Michigan State University" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.299, + 0.776, + 0.313 + ], + "angle": 0, + "content": "[yuiliang.guo2,cheng.zhao,ruoyu.wang,xingyu.huang,liu. ren]@us.bosch.com" + }, + { + "type": "text", + "bbox": [ + 0.435, + 0.313, + 0.568, + 0.325 + ], + "angle": 0, + "content": "\\(^{2}\\)kumarab6@msu.edu" + }, + { + "type": "text", + "bbox": [ + 0.361, + 0.327, + 0.642, + 0.341 + ], + "angle": 0, + "content": "https://yuliangguo.github.io/supnerf" + }, + { + "type": "image", + "bbox": [ + 0.254, + 0.342, + 0.749, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.511, + 0.788, + 0.553 + ], + "angle": 0, + "content": "Fig. 1: Teaser. SUP-NeRF is a unified solution that predicts an object's pose, shape, and texture using a single network. SUP-NeRF is trained on real driving scenes with imprecise labels, and it adapts robustly to new cross-dataset scenarios." + }, + { + "type": "text", + "bbox": [ + 0.26, + 0.558, + 0.741, + 0.821 + ], + "angle": 0, + "content": "Abstract. Monocular 3D reconstruction for categorical objects heavily relies on accurately perceiving each object's pose. While gradient-based optimization in a NeRF framework updates the initial pose, this paper highlights that scale-depth ambiguity in monocular object reconstruction causes failures when the initial pose deviates moderately from the true pose. Consequently, existing methods often depend on a third-party 3D object to provide an initial object pose, leading to increased complexity and generalization issues. To address these challenges, we present SUP-NeRF, a Streamlined Unification of object Pose estimation and NeRF-based object reconstruction. SUP-NeRF decouples the object's dimension estimation and pose refinement to resolve the scale-depth ambiguity, and introduces a camera-invariant projected-box representation that generalizes cross different domains. While using a dedicated pose estimator that smoothly integrates into an object-centric NeRF, SUP-NeRF is free from external 3D detectors. SUP-NeRF achieves state-of-the-art results in both reconstruction and pose estimation tasks on the nuScenes dataset. Furthermore, SUP-NeRF exhibits exceptional cross-dataset generalization on the KITTI and Waymo datasets, surpassing prior methods with up to \\(50\\%\\) reduction in rotation and translation error." + }, + { + "type": "page_footnote", + "bbox": [ + 0.221, + 0.826, + 0.323, + 0.84 + ], + "angle": 0, + "content": "* Project Lead" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "Y. Guo et al." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.375, + 0.161 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.175, + 0.787, + 0.326 + ], + "angle": 0, + "content": "Monocular 3D object reconstruction [8,9,13,19,49] is a critical technology with broad applications in autonomous driving [18,29,49], AR/VR, robotics [43], and embodied AI [1,40,51]. Its impact is particularly significant in autonomous driving, where it plays a vital role in tasks like auto-labeling [49] and world-model construction [18,29]. This technology is essential for reconstructing 3D models from typical driving scenarios, where single-view observations are common. This paper focuses explicitly on the problem of recovering an object's pose, shape, and texture from a single image under real driving scenarios, see Fig. 1. Our aim is to devise an efficient solution, thereby unlocking the full potential of monocular reconstruction for active tasks such as planning and control." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.327, + 0.787, + 0.463 + ], + "angle": 0, + "content": "Neural Radiance Field (NeRF) technology [27, 30, 48] has revolutionized the field of 3D reconstruction. It excels at capturing fine details in a scene and generating novel views from existing reconstructions. An exciting branch, Object-centric NeRF [11, 12, 18, 29, 34, 45, 46], takes things further, allowing for flexible data creation and tackling reconstruction challenges from limited number of views by incorporating category-specific prior knowledge. Earlier object-centric NeRF methods [8, 12, 13, 19, 46, 48, 49] require densely overlapping views and precise poses, limiting their use to synthetic data. Later advances [18, 29] enable monocular training with imperfect masks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.463, + 0.787, + 0.583 + ], + "angle": 0, + "content": "Although promising, object NeRFs show limitations in correcting erroneous object poses. First, the well-known scale-depth ambiguity [16,25,36] also affects the object NeRF optimization. Fig. 2 shows the optimization failure of a monocular NeRF method under longitudinal translation error when it is given complete freedom in adjusting the scale within the normalized space. Secondly, NeRF also struggles with large rotation errors. Appendix A6.1 shows that NeRF typically requires the initial rotation error to be below \\(25^{\\circ}\\), and the pose optimization cannot go lower than \\(7^{\\circ}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.787, + 0.674 + ], + "angle": 0, + "content": "Recent papers [18, 29] use an external monocular 3D detector to provide initial poses to NeRF. However, monocular 3D detectors suffer from generalization issues [16, 17] and increased complexity, ultimately affecting the NeRF pipeline. BootInv [34], a more recent method, tackles pose estimation, shape, and appearance reconstruction in a single framework but requires additional pre-training of dense field generator and struggles with handling moderate occlusions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.787, + 0.84 + ], + "angle": 0, + "content": "To resolve the scale-depth ambiguity and achieve greater generalization, this paper proposes SUP-NeRF, a Streamlined Unification of Pose estimation and NeRF. SUP-NeRF uses an image encoder to output the object dimensions and then refines for object pose using an iterative process while keeping its size fixed. Besides decoupling of an object's dimension and pose to resolve the scale-depth ambiguity, the introduction of a novel projected-box representation in the iterative update refinement also enhances the generalization to novel data since the representation makes the deep pose refiner independent from camera intrinsic parameters. In addition, SUP-NeRF tackles the complexity challenge by sharing the NeRF image encoder with the pose estimation, making the pose estimation module a seamless fit for object NeRFs." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.654, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SUP-NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.146, + 0.437, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.438, + 0.153, + 0.562, + 0.164 + ], + "angle": 0, + "content": "R err: 0.00, T err: 6.85" + }, + { + "type": "image", + "bbox": [ + 0.45, + 0.176, + 0.56, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.577, + 0.153, + 0.7, + 0.164 + ], + "angle": 0, + "content": "R err: 0.06, Terr: 6.35" + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.174, + 0.695, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.261, + 0.785, + 0.331 + ], + "angle": 0, + "content": "Fig. 2: Scale-depth ambiguity in NeRF. Given the input image (left), joint optimization of pose, shape, and texture in NeRF has full freedom to rescale the shape within the normalized shape space (blue box) or move the 3D box. Such phenomenon is observed from the evolution of the rendered objects from iteration 0 (middle) to iteration 50 (right)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.341, + 0.785, + 0.446 + ], + "angle": 0, + "content": "In our experiments, we first comprehensively benchmark and conduct ablations on SUP-NeRF on the nuScenes [3] dataset. SUP-NeRF outperforms state-of-the-art (SoTA) methods and achieves SoTA object reconstruction and pose performance on this dataset. We next conduct cross-dataset object reconstruction and pose estimation experiments of the nuScenes models on the KITTI and Waymo datasets to test SUP-NeRF's generalization capability. SUP-NeRF shows more significant improvements over SoTA in such cross-dataset settings." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.448, + 0.564, + 0.461 + ], + "angle": 0, + "content": "In summary, our main contributions include:" + }, + { + "type": "text", + "bbox": [ + 0.211, + 0.463, + 0.784, + 0.492 + ], + "angle": 0, + "content": "- We introduce SUP-NeRF, an streamlined method unifying object pose estimation and monocular object-centric neural reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.211, + 0.493, + 0.785, + 0.538 + ], + "angle": 0, + "content": "- We identify depth-scale ambiguities as the key barrier to object NeRF's pose optimization and craft a module to overcome this effectively, while also integrating seamlessly with existing object NeRF frameworks." + }, + { + "type": "text", + "bbox": [ + 0.211, + 0.539, + 0.785, + 0.583 + ], + "angle": 0, + "content": "- We introduce a novel projected box representation for the pose estimation module, enabling the deep refiner to operate in a space invariant to camera intrinsic parameters, thereby enhancing SUP-NeRF's generalization capability." + }, + { + "type": "text", + "bbox": [ + 0.211, + 0.584, + 0.785, + 0.629 + ], + "angle": 0, + "content": "- SUP-NeRF achieves SoTA object reconstruction and pose estimation results, outperforming other methods on the nuScenes dataset, and in challenging cross-dataset experiments on both KITTI and Waymo datasets." + }, + { + "type": "list", + "bbox": [ + 0.211, + 0.463, + 0.785, + 0.629 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.655, + 0.387, + 0.67 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Object-centric NeRF. Object-centric NeRF [11, 12, 18, 29, 34, 45, 46] is a specialized NeRF variant for modeling specific object categories. Its pipeline feeds a NeRF decoder with shape and texture codes, generating occupancy and color values at 3D positions, followed by volumetric rendering to produce a rendered image. The entire process is differentiable, allowing back-propagation of the photometric loss to update shape codes, texture codes, network weights, and object poses. Early object-centric NeRFs [8, 12, 13, 19, 45, 46, 48, 49] required accurate object poses, full visibility, and dense multi-view observations to reconstruct a categorical model, making them inaccessible for real-world applications. Some approaches [12, 23] claim successful object pose recovery through gradient-based" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "Y. Guo et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 1: Literature Summary. SUP-NeRF sets a benchmark for minimal requirements, eliminating the need for external 3D detectors, CAD models, or dense NOCS generators. It is dedicated to the reconstruction of object shapes and pose estimation directly from a single image of driving scenes, employing an object-centric framework." + }, + { + "type": "table", + "bbox": [ + 0.227, + 0.213, + 0.773, + 0.302 + ], + "angle": 0, + "content": "
MethodsIndoorOutdoorViewsScopeCADwt. Det3Dwt. NeRFNOCS Gen.
6DoF Pose [21, 26]✓×singleobject✓✓××
3D Detector [16, 42]×✓singleimage×✓××
NeRF [27]✓✓multiscene××✓×
Object-NeRF [18, 29]×✓singleobject××✓×
BootInv [34]×✓singleobject×✓✓✓
SUP-NeRF×✓singleobject×✓✓×
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.319, + 0.788, + 0.515 + ], + "angle": 0, + "content": "pose refinement from monocular images. However these successes are still limited to synthetic dataset, and oversimplified object/camera poses. Recent methods [18,29] investigate more practical pipelines demonstrating the object-NeRF's potential to train and test on driving scenes with limited views and occluded object data. However, both methods require relatively accurate poses from heavy third-party 3D detectors [32,36]. Other recent works [2,5] try freeing NeRF from external camera poses; however, solving camera poses from densely overlapping views is a different problem than recovering an object's pose from a single image. Compared to these methods, SUP-NeRF recovers object pose from a single image without a 3D detector in the pipeline. A more recent work [34] also unifies pose estimation and NeRF in a single network, and enables high-fidelity object generation. However, its training requires additional NOCS dense generator and it is highly sensitive to moderate occlusions compared to our method." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.789, + 0.844 + ], + "angle": 0, + "content": "Monocular 3D Object Pose Estimation. Advancements in deep learning and the availability of 3D-labeled datasets [3, 6, 10] enable many monocular 3D object detection [16, 24, 25, 33, 36, 42, 47] and 6DoF pose estimation methods [14, 21, 22, 26, 28, 39, 44, 50]. Although estimating 3D information from a single image is ill-posed due to depth-scale ambiguity, leveraging constraints like CAD models or sizes enables accurate object pose recovery via PnP processes [35, 39, 41] or iterative optimization [14, 21]. These indoor 6DoF pose estimation mostly follow an object-centric design which aims to estimate highly accurate object pose within the 2D image patches out of a previous 2D object detector. Recent studies [4, 20] use the NeRF pipeline to reconstruct 3D models or to train model-based keypoint detectors to estimate object poses during inference. However, none of these model-based methods is ideal for outdoor driving scenarios, as pre-reconstruction of models is often impossible in new scenes. Fortunately, inferring the 3D dimensions of objects from an image is feasible when its appearance is tightly bounded to its fine-grained class [28], which keeps an outdoor 3D detector from being fully ill-posed. Nowadays, reliable outdoor 3D detectors usually favors a image-centric design to direct detect many 3D objects captured within the whole image, which usually requires a large backbone network [16, 42]. However, the performance of SoTA detectors drops significantly when applied to datasets with distribution different from the training data [16, 25, 42]." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.655, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SUP-NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.147, + 0.78, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.338, + 0.788, + 0.395 + ], + "angle": 0, + "content": "Fig. 3: SUP-NeRF Overview. SUP-NeRF unifies pose estimation and NeRF. The pose estimation module enables SUP-NeRF to work for objects in diverse poses without external 3D detectors. The increase of complexity only constitutes a few MLP layers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.41, + 0.788, + 0.515 + ], + "angle": 0, + "content": "Tab. 1 outline the differences in methodologies. Unlike existing approaches, SUP-NeRF unifies pose estimation and object NeRF in a streamlined manner, adopting a straightforward object-centric design akin to indoor techniques. Without the need for CAD models, SUP-NeRF is suitable for outdoor driving scenarios. Despite its focused scope on object-centric pose estimation, SUP-NeRF showcases enhanced cross-domain generalization capabilities compared to more complex image-centric 3D detectors within this realm." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.537, + 0.364, + 0.554 + ], + "angle": 0, + "content": "3 SUP-NeRF" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.788, + 0.749 + ], + "angle": 0, + "content": "Our goal is to create a NeRF framework that bypasses NeRF's limitations without external 3D detectors. Fig. 3 shows the overall pipeline of SUP-NeRF, a streamlined unified network that predicts object pose, shape, and texture from a single image. SUP-NeRF consists of three modules: a Resnet-based encoder, a pose estimation module, and a NeRF decoder. SUP-NeRF processes an image and its occlusion mask, with the encoder turning the masked image into shape, texture, and pose codes and predicted object dimensions. The pose code, object dimensions and a box code, are passed to the pose estimation module to iteratively refine the object pose \\([R_{o2c}|T_{o2c}]\\). After a few iterations, the estimated pose is converted to the camera pose \\([R_{c2o}|T_{c2o}]\\) and passed to the NeRF decoder. In the NeRF phase, the decoder uses the shape, texture codes, and camera pose for volumetric rendering, creating both RGB and occupancy images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Pose Estimation Module. The key component of SUP-NeRF is the pose estimation module, which we show in Fig. 4. It aims to provide reliable object poses for targets at all ranges and orientations. Our pose estimation module iterative updates the input pose, based on the visual difference between the input pose and the observed object in the image. Given predicted object dimensions \\([H_B, W_B, L_B]\\), camera intrinsic \\(K\\), and current pose \\(R^{(t)}, T^{(t)}\\), image projections" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "Y. Guo et al." + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.147, + 0.7, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.313, + 0.788, + 0.397 + ], + "angle": 0, + "content": "Fig.4: SUP-NeRF Pose Estimation Module. The pose estimation module of SUP-NeRF iteratively updates the object's pose while preserving scale. It takes the projection of 3D box corners as a visual representation of the input pose and estimate the pose update via comparing it to observed image in a latent embedding space. These designs handle scale-depth ambiguity and make the deep refiner independent from camera intrinsic parameters for better cross-domain generalization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.419, + 0.788, + 0.557 + ], + "angle": 0, + "content": "of the 3D box corners \\( B_{proj}^{(t)} \\) is computed, which can be interpolated as a visual representation of current pose. After encoding the 8 corners into a higher dimensional box code, SUP-NeRF feeds into the pose refiner along with a pose code from the image encoder to predict the desired pose update \\( \\Delta R^{(t)}, \\Delta T^{(t)} \\). SUP-NeRF then composes the pose changes with current pose to compute the next pose state, \\( R^{(t + 1)}, T^{(t + 1)} \\). This updating process starts from a fully random object pose within the viewing frustum of the image patch, and takes a few iterations such that the final pose approaches the true pose even from a random initial pose." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.561, + 0.751, + 0.576 + ], + "angle": 0, + "content": "The pose estimation module incorporates the following unique designs:" + }, + { + "type": "text", + "bbox": [ + 0.211, + 0.58, + 0.784, + 0.682 + ], + "angle": 0, + "content": "- SUP-NeRF preserves the object dimensions to ensure unambiguous pose updates. Our implementation, using projected 3D boxes on the image plane, leads to a concise encoding of shape and pose, reducing computation in image rendering and encoding per iteration, as opposed to previous methods [21, 29] relying on full CAD models. Since this representation makes the deep pose refiner's task independent from camera intrinsic parameters, it leads to better generalization to different testing domains, as evidenced in Tabs. 3 and 4." + }, + { + "type": "text", + "bbox": [ + 0.211, + 0.686, + 0.784, + 0.741 + ], + "angle": 0, + "content": "- SUP-NeRF updates the object pose iteratively rather than conduct a direct pose estimation using a PnP-based [14] because the later is sensitive to missed or swapped corners thus often produces far out-of-distribution poses. Our ablation studies in Tab. 6 confirm this difference." + }, + { + "type": "text", + "bbox": [ + 0.211, + 0.746, + 0.784, + 0.806 + ], + "angle": 0, + "content": "- SUP-NeRF represents the rotation update in axis-angle representation [37]. During rotation updating process, we convert the axis-angle space 3D rotation \\(\\Delta R^{(t)}\\) back to \\(SO(3)\\), compose it with the previous rotation to compute the new rotation state as \\(R_{\\mathrm{e2}}^{(t + 1)}\\)." + }, + { + "type": "text", + "bbox": [ + 0.211, + 0.808, + 0.784, + 0.839 + ], + "angle": 0, + "content": "- SUP-NeRF represents the translation update in a relative space, indicated as \\(\\Delta T^{(t)}: [v_x^{(t)}, v_y^{(t)}, \\rho^{(t)}]\\), where \\(v_x^{(t)}\\) and \\(v_y^{(t)}\\) indicate the location change in" + }, + { + "type": "list", + "bbox": [ + 0.211, + 0.58, + 0.784, + 0.839 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.654, + 0.114, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SUP-NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.146, + 0.784, + 0.177 + ], + "angle": 0, + "content": "image space, and \\(\\rho^{(t)}\\) indicates the ratio between the target depth and the source depth. Specifically, SUP-NeRF updates the translation in image space" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.178, + 0.331, + 0.192 + ], + "angle": 0, + "content": "and depth as:" + }, + { + "type": "equation", + "bbox": [ + 0.443, + 0.184, + 0.576, + 0.202 + ], + "angle": 0, + "content": "\\[\nx ^ {(t + 1)} = x ^ {(t)} + v _ {x} ^ {(t)}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.443, + 0.205, + 0.785, + 0.224 + ], + "angle": 0, + "content": "\\[\ny ^ {(t + 1)} = y ^ {(t)} + v _ {y} ^ {(t)} \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.441, + 0.227, + 0.559, + 0.244 + ], + "angle": 0, + "content": "\\[\nZ ^ {(t + 1)} = \\rho^ {(t)} Z ^ {(t)}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.255, + 0.784, + 0.317 + ], + "angle": 0, + "content": "where \\((x,y)\\) represent the object's 2D location in image space, \\(Z\\) indicates the depth. The new 3D location \\(T_{o2c}^{(t + 1)}\\) is then computed from the updated image coordinates, depth and camera intrinsic matrix \\(K\\). Compared to [21, 34], our definition of \\(\\rho^{(t)}\\) leads to a more straight-forward update." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.318, + 0.784, + 0.392 + ], + "angle": 0, + "content": "During the pose updating process indicated by \\(\\oplus\\), both the rotation and translation are updated. Through such representation and updating scheme, SUP-NeRF handles objects at varying distances with the same output space. Further discusses the choice of an effective pose representation and the right coordinate frame are included in Appendices A1 and A2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.394, + 0.785, + 0.514 + ], + "angle": 0, + "content": "Cross-task Shortcut. Simultaneously perceiving pose, shape, and texture, requires an image encoder with shared initial layers but separate later stages. However, encoding shape, texture, and pose is a challenge as they require different features, making it tough for a shared encoder to meet all needs effectively. Shape and texture encoding demand a pose-invariant approach, as they correspond to each object's frame. On the other hand, pose encoding needs to capture sufficient pose information for later use. This difference leads to conflicting learning signals during joint training, potentially weakening the model's performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.515, + 0.785, + 0.635 + ], + "angle": 0, + "content": "SUP-NeRF proposes a straightforward one-step approach to resolve this conflict. After processing through the first four CNN layers separately, SUP-NeRF introduces a cross-task shortcut that deducts pose features from the shape and texture features. This design choice eliminates the shape and texture's reliance on pose, allowing the earlier layers to focus on pose-dependent features for all tasks. This shortcut turns the conflict into a cooperative relationship, enhancing the joint training of different modules, as demonstrated in Tab. 7. The details of our image encoder is described in Appendix A3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.636, + 0.784, + 0.698 + ], + "angle": 0, + "content": "Training. SUP-NeRF closely follows the loss functions and occupancy mask definitions from AutoRF [29]. The total loss during training is: \\( L_{train} = L_{rgb} + w_{occ}L_{occ} + w_{pose}(L_{pose}^{(direct)} + \\sum_t^K L_{pose}^{(t)}) \\), where \\( w_{occ} \\) and \\( w_{pose} \\) indicate the corresponding weights." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.699, + 0.785, + 0.84 + ], + "angle": 0, + "content": "We compare the RGB output to the input image to calculate photometric loss \\( L_{rgb} \\), and the occupancy values to the mask for occupancy loss \\( L_{occ} \\). SUP-NeRF additionally compares the projected box corners from the integrative pose estimator \\( B^{(t)}_{proj} \\) with the ground-truth values to compute the pose losses \\( L_{pose}^{(t)} \\). We define the pose loss function as the mean distance between the matched corners: \\( L_{pose}^{(*)} = \\frac{1}{8}\\sum_{i=1}^{8}\\sqrt{(x_i^{(*)} - \\hat{x}_i)^2 + (y_i^{(*)} - \\hat{y}_i)^2} \\), where (*) indicates it compatible with both the direct pose loss and the iterative pose loss, \\( (x_i^{(*)}, y_i^{(*)}) \\) denote a input corner position, and \\( (\\hat{x}_i, \\hat{y}_i) \\) indicates a ground-truth corner po" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "Y. Guo et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.192 + ], + "angle": 0, + "content": "sition. SUP-NeRF introduces an additional MLP layer for more accurate pose estimation to directly supervise the 2D coordinates of box corners predicted from the pose code." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.787, + 0.403 + ], + "angle": 0, + "content": "In training, only the network weights are treated as the estimizable variables. Moreover, the NeRF decoder uses ground-truth camera poses \\([R_{c2o}^{gt}|T_{c2o}^{gt}]\\) instead of predictions from the pose estimation module to ensure more reliable training. Implementation Details. SUP-NeRF's image encoder backbone is based on ResNet50, with shared early layers and separated later layers for pose, shape, and texture encoding. For more information on this, please refer to the Appendix A3. To implement the NeRF decoder, we use CodeNeRF [12] for its simplicity and ease of use. For the differentiable rendering based on NeRF, we set the latent dimension to 256, use 64 samples along each ray, and normalize the shape space by the diagonal length of the observed object's 3D box. For the pose estimation module, we also set the latent dimension to 256 for all MLP layers. Before feeding the input image and occupancy mask into the image encoder, we pad them to make them square and resize them to \\(128\\times 128\\) pixels. We then render \\(32\\times 32\\) images and occupancy maps to calculate losses." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.405, + 0.787, + 0.54 + ], + "angle": 0, + "content": "To ensure effective iterative pose estimation without sacrificing generalization, SUP-NeRF uses a heuristic to sample a random initial 3D object pose across the entire pose space. For translation, we sample in the relative space. We sample a random 2D location within the image ROI range as the initial projection of the object center, while setting the initial depth to \\( Z = 20 \\). For rotation, we sample from the full range of yaw angles combined with a random rotation angle within a \\( (-20^{\\circ}, 20^{\\circ}) \\) range around a random axis direction. This sampling method covers almost the entire distribution of object poses in the camera frame and is used consistently in both training and inference." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.541, + 0.787, + 0.63 + ], + "angle": 0, + "content": "During training, we use 3 iterations of the pose estimator to generate corresponding pose losses, which allow for pose regression to handle highly diverse poses. We jointly train the unified model for the pose estimation module, image encoder, and NeRF decoder for 40 epochs on the nuScenes dataset using a learning rate of \\(10^{-4}\\) and loss weights of \\(w_{occ} = 0.1\\), \\(w_{pos} = 0.01\\) to compute the total loss." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.631, + 0.787, + 0.753 + ], + "angle": 0, + "content": "Inference. During inference, SUP-NeRF optimizes the shape code, texture code, and object pose \\([R_{o2c}|T_{o2c}]\\) keeping the network weights frozen. The total loss \\(L_{infer}\\) is thus \\(L_{infer} = L_{rgb} + w_{occ}L_{occ}\\), where \\(w_{occ}\\) balances the two. SUP-NeRF minimizes the loss \\(L_{infer}\\) to update the variables. At inference, we run 3 iterations for the feed-forward pose estimation module without updating the shape and color codes, and 50 iterations for the gradient-based joint optimization of shape, texture, and pose. We use a step size of 0.02 for shape and texture code updates and 0.01 for pose updates during joint optimization." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.777, + 0.377, + 0.793 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.794, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Our experiments use three datasets: nuScenes [3], KITTI [6] and Waymo [38]. Our evaluation framework simultaneously measures rendering quality, shape reconstruction accuracy, and pose estimation precision." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.654, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SUP-NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.237 + ], + "angle": 0, + "content": "Data Splits. Since these datasets are not specifically for 3D reconstruction, we sifted a subset for better evaluation. We chose daytime sequences and used Mask R-CNN [7] to get instance masks, as nuScenes lacks 2D segmentation masks. We matched these masks with 3D bounding box annotations, categorizing them into foreground, background, and unknown, similar to AutoRF [29]. More details are in Appendix A7. We use the following splits of these datasets:" + }, + { + "type": "text", + "bbox": [ + 0.21, + 0.239, + 0.784, + 0.265 + ], + "angle": 0, + "content": "- nuScenes Train Split: We collected 87,048 objects from the nuScenes Train split for training." + }, + { + "type": "text", + "bbox": [ + 0.21, + 0.269, + 0.784, + 0.295 + ], + "angle": 0, + "content": "- nuScenes Val Split: We collected 5,000 random objects from the nuScenes Val split for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.21, + 0.298, + 0.784, + 0.312 + ], + "angle": 0, + "content": "- KITTI Val Split: We use 4,895 objects from KITTI Val split [6] for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.21, + 0.314, + 0.784, + 0.342 + ], + "angle": 0, + "content": "- Waymo Val Split: We randomly selected 5,000 objects from Waymo Val split [16] for evaluation." + }, + { + "type": "list", + "bbox": [ + 0.21, + 0.239, + 0.784, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.345, + 0.788, + 0.556 + ], + "angle": 0, + "content": "Evaluation metrics. Our evaluation framework differs from traditional NeRF assessments as it simultaneously evaluates object reconstruction quality and pose accuracy. We incorporate two monocular reconstruction metrics: (1) the PSNR (Peak Signal-to-Noise Ratio) comparing the rendered and observed images; (2) the Depth Error between the rendered depth map and sparse LiDAR data, denoted as \\( DE \\). We also include two pose estimation metrics: (3) rotation error denoted as \\( RE \\); (4) the translation error denoted as \\( TE \\). To offer a thorough evaluation, we compare various methods at different stages: post-feed-forward (\\( FF \\)), and at the 50th iterations of NeRF joint optimization. Beyond monocular assessments, we also conduct cross-view evaluations for PSNR and depth error metrics, denoted as (5) PSNR-\\( C \\) and (6) \\( DE-C \\), focusing on objects observed from novel views. As the nuScenes dataset provides object IDs that allows tracking objects across various views, the cross-view evaluation is exclusively conducted on the nuScenes dataset, as explained in Fig. 5." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.558, + 0.788, + 0.709 + ], + "angle": 0, + "content": "Baselines. We compare SUP-NeRF with two baselines: AutoRF [29] and AutoRF [29] + FCOS3D [42] (selected for its monocular capability, and availability of model trained on the nuScenes dataset). Both AutoRF and SUP-NeRF use the same NeRF decoder and image encoder to ensure a fair comparison. The difference lies in SUP-NeRF's additional layers for pose encoding and refinement. During testing, we apply the same random pose generation for both AutoRF and SUP-NeRF methods, whereas the FCOS3D baseline begins with its own pose prediction. We also compare with BootInv [34], which recovers pose, shape, and appearance. Leveraging our flexible design, we developed SUP-BootInv by integrating BootInv's pretrained decoder with our encoder and pose estimator." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.736, + 0.738, + 0.75 + ], + "angle": 0, + "content": "4.1 nuScenes Monocular Reconstruction and Pose Estimation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Tab. 2 shows that SUP-NeRF outperforms two baselines, AutoRF and its combination with FCOS3D on nuScenes Val set. AutoRF alone struggles to improve pose from a random start, resulting in lower PSNR and depth error. SUP-NeRF consistently reduces depth error by over \\(25\\%\\), rotation error by over \\(30\\%\\), and translation error by over \\(13\\%\\) compared to the FCOS3D and AutoRF com" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "Y. Guo et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 2: nuScenes Monocular Reconstruction and Pose Estimation Results. SUP-NeRF consistently improves AutoRF-based pipelines in all metrics, particularly in the pose estimation metrics. SUP-BootInv also consistently improves BootInv in all metrics. [Key: Best, Second Best, FF= Feed Forward, C= Cross-View]" + }, + { + "type": "table", + "bbox": [ + 0.233, + 0.213, + 0.772, + 0.302 + ], + "angle": 0, + "content": "
MethodPSNR (↑)DE (m) (↓)RE (deg.) (↓)TE (m) (↓)PSNR-C (↑)DE-C (m) (↓)
FF|50itFF|50itFF|50itFF|50itFF|50itFF|50it
AutoRF [29]3.6|10.611.21|10.0987.52|88.076.04|5.9510.0|8.81.31|1.41
AutoRF+FCOS7.5|17.21.34|0.819.77|10.170.85|0.789.8|10.51.29|1.16
SUP-NeRF10.5|18.80.69|0.607.01|7.070.68|0.7310.6|10.91.22|1.14
BootInv [34]9.4|14.35.01|3.5628.40|28.002.59|2.9110.9|11.81.37|1.35
SUP-BootInv10.9|15.41.95|1.627.11|8.400.64|1.0010.9|11.91.37|1.40
" + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.317, + 0.77, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.444, + 0.788, + 0.5 + ], + "angle": 0, + "content": "Fig. 5: nuScenes Cross-view Evaluation. Each row shows a set of images of the same object from different angles. We use the example for both monocular pose estimation and NeRF reconstruction. Other images in each row are used to evaluate the reconstructed shape and texture in PSNR and Depth Error (DE)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.531, + 0.785, + 0.561 + ], + "angle": 0, + "content": "bination. Our feed-forward stage also shows promising potential for real-time applications, especially where rendering quality is less critical." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.562, + 0.789, + 0.698 + ], + "angle": 0, + "content": "We next compare BootInv [34] as the baseline, although the performance of original BootInv's pose estimation was subpar on both the nuScenes and KITTI datasets, integrating our robust pose module with this NeRF framework significantly enhanced pose estimation accuracy. This underscores the adaptability of our unified pipeline in augmenting various object NeRF frameworks for improved pose estimation. Moreover, comparisons of BootInv [34] and SUP-BootInv in terms of monocular PSNR and depth error suggest that SUP-BootInv's superior pose estimation contributes to its higher PSNR and lower depth error. More detailed analysis upon BootInv-based pipelines is included in Appendix A5." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.722, + 0.473, + 0.736 + ], + "angle": 0, + "content": "4.2 Cross-Dataset Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.785, + 0.794 + ], + "angle": 0, + "content": "We next compare the generalization capability of all these NeRF models with the cross-dataset experiments. For these experiments, we directly apply nuScenes models to the KITTI and Waymo datasets and report the metrics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "KITTI. Tab. 3 shows the result of cross-dataset experiments on the KITTI dataset. It shows that SUP-NeRF achieves SoTA performance across all evaluation metrics. SUP-NeRF reduces depth error by about \\(16\\%\\), rotation error by" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.654, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SUP-NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 3: KITTI Cross-dataset Monocular Reconstruction and Pose Estimation Results. We train all methods on nuScenes dataset, and test on KITTI dataset. Our methods SUP-NeRF and SUP-BootInv consistently show superiority in all metrics compared to the counterpart methods. [Key: Best, Second Best, FF= Feed Forward]" + }, + { + "type": "table", + "bbox": [ + 0.298, + 0.213, + 0.706, + 0.302 + ], + "angle": 0, + "content": "
MethodPSNR (↑)DE (m) (↓)RE (deg.) (↓)TE (m) (↓)
FF|50itFF|50itFF|50itFF|50it
AutoRF [29]0.4|7.69.89|8.8389.67|90.426.16|6.07
AutoRF+FCOS2.4|13.42.42|1.7411.95|12.52.2|2.09
SUP-NeRF4.0|14.12.19|1.546.79|6.891.06|1.01
BootInv [34]7.0|12.36.11|4.3815.52|15.763.97|3.90
SUP-BootInv7.6|12.72.79|2.219.16|8.841.15|1.67
" + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.324, + 0.773, + 0.537 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.548, + 0.788, + 0.633 + ], + "angle": 0, + "content": "Fig. 6: KITTI (Cross-dataset) Qualitative Results. In the top panel, we demonstrate SUP-NeRF executes pose estimation reliably, fast converging from a random initial pose to the true one, and enables neural reconstruction under diverse object poses, occlusion cases under this cross-dataset setup. In the bottom panel, SUP-NeRF is visually compared to the other major competitor, demonstrates sharper rendered image, higher accuracy in shape and pose." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.67, + 0.788, + 0.761 + ], + "angle": 0, + "content": "as large as \\(50\\%\\) and translation error also about \\(50\\%\\). Thus, SUP-NeRF shows significantly better geometric performance in cross-dataset setting compared to AutoRF+FCOS, which we attributes to our deep pose module's independence from camera intrinsic parameters. Meanwhile, compared to nuScenes evaluation, all the methods' PSNR and depth error performance drop, which means the cross-dataset generalization of monocular NeRF remains an open challenge." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Fig. 6 presents the qualitative results on this experiment, demonstrating how SUP-NeRF handles diverse poses and occlusions in pose refinement and the visual comparisons between the two most competitive pipelines. Consistent improvements upon integrating our design into BootInv baseline were also observed. More visual comparisons are in Appendix A9." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "Y. Guo et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 4: Waymo Cross-dataset Monocular Reconstruction and Pose Estimation Results. We train all methods on nuScenes dataset, and test on Waymo dataset. Our methods SUP-NeRF and SUP-BootInv consistently show superiority in all metrics compared to the counterpart methods, the only exception is the comparison to AutoRF+FCOS in Rotation Error. [Key: Best, Second Best, FF= Feed Forward]" + }, + { + "type": "table", + "bbox": [ + 0.298, + 0.227, + 0.706, + 0.316 + ], + "angle": 0, + "content": "
MethodPSNR (↑)DE (m) (↓)RE (deg.) (↓)TE (m) (↓)
FF|50itFF|50itFF|50itFF|50it
AutoRF [29]0.6|9.86.76|6.5386.56|87.679.1|9.14
AutoRF+FCOS6.2|16.52.43|2.326.65|7.23.22|3.31
SUP-NeRF4.8|17.02.32|1.5610.01|10.61.68|1.54
BootInv [34]8.1|11.08.85|8.2030.78|31.525.26|6.16
SUP-BootInv8.6|11.95.36|4.3810.24|11.041.67|2.53
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.343, + 0.788, + 0.419 + ], + "angle": 0, + "content": "Waymo. Tab. 4 shows the result of cross-dataset experiments on the Waymo dataset. It shows that SUP-NeRF achieves SoTA performance across most evaluation metrics. The only exception is the superior rotation estimation achieved by FCOS3D, which is likely due to leveraging larger context for heavily occluded cases. However, our advantage in translation estimation is more significant." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.439, + 0.403, + 0.453 + ], + "angle": 0, + "content": "4.3 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.463, + 0.788, + 0.538 + ], + "angle": 0, + "content": "The major ablation studies we conduct includes (i) the impact of pose estimation module designs and (ii) the impact of different object NeRF networks. Other ablation studies to analyze the impact of training with predicted 2D boxes, pose refinement iterations, training epochs, as well as NeRF's capability on handling initial pose errors are included in Appendix A6, owing to lack of space." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.539, + 0.788, + 0.704 + ], + "angle": 0, + "content": "Pose Estimation Module. We evaluate multiple pose estimation designs from two main angles. First, we focus on object-centric pose estimation task alone with evaluations on nuScenes, KITTI, and Waymo datasets. The prototypical designs include: (i) Direct 6D pose regression using MLP layers; (ii) Pose estimation from predicted box corners via Keypoint-based Perspective-n-Point (PnP); (iii) NOCS-based pose estimation as implemented in BootInv [34]; (vi) External monocular 3D detection FCOS3D [42]; (v) Our iterative pose refinement method, which uses projected boxes. For fairness, methods (1), (2), and (5) share the same image encoder, pose encoding layers, differing only in their output layers. Tab. 5 shows that SUP-NeRF consistently outperforms the others, notably in cross-dataset translation estimation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The second evaluation assesses the impact of different initial poses on object-NeRF performance within the SUP-NeRF architecture. We additionally compared (i) Ground-Truth (GT) pose; (ii) Random initial pose, to represent the oracle and worst case in pose initialization, and exclude the NOCS method which is part of a different NeRF framework. Given different initial poses, all methods undergo the same NeRF Gradient-based Pose Refinement (NGPR), evaluated immediately and after 50 iterations. Tab. 6 shows that our method significantly surpasses other initial pose estimators in geometric accuracy, demonstrating its effectiveness in enhancing 3D model quality." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.654, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SUP-NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.785, + 0.201 + ], + "angle": 0, + "content": "Table 5: Pose Estimation Ablation Studies. We evaluate all the candidate pose estimation methods with a focus on pose accuracy alone, on nuScenes, KITTI, and Waymo validation sets. Our method shows outstanding robustness in all cases, particularly in translation accuracy in cross-dataset tests. [Key: Best]" + }, + { + "type": "table", + "bbox": [ + 0.288, + 0.213, + 0.715, + 0.301 + ], + "angle": 0, + "content": "
Pose ModuleRE (deg.)(↓)TE (m) (↓)
nuScenesKITTIWaymonuScenesKITTIWaymo
MLP Pose31.9637.5830.895.535.3110.1
Corners+PnP24.7935.1093.762.823.5518.68
NOCS +PnP(BootInv)28.4015.5223.662.593.973.47
FCOS3D9.7711.956.650.852.23.22
SUP-NeRF7.016.7910.010.681.061.54
" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.318, + 0.785, + 0.374 + ], + "angle": 0, + "content": "Table 6: Ablation studies of Pose Initialization in Object NeRF on nuScenes. We compare different choices of initial poses under SUP-NeRF frameworks. As shown, SUP-NeRF presents superiors effectiveness in all the metrics. [Key: Best, FF=Feed Forward, C=Cross-View]" + }, + { + "type": "table", + "bbox": [ + 0.245, + 0.386, + 0.758, + 0.487 + ], + "angle": 0, + "content": "
Initial PosePSNR (↑)DE (m) (↓)RE (deg.) (↓)TE (m) (↓)PSNR-C (↑)DE-C (m) (↓)
FF|50itFF|50itFF|50itFF|50itFF|50itFF|50it
Random3.6|10.611.21|10.0987.52|88.076.04|5.959.1|7.41.42|1.82
MLP Pose1.2|6.95.33|4.4131.96|24.575.53|2.789.8|8.51.3|1.4
Corners+PnP7.2|16.22.18|1.8324.79|24.572.82|2.789.9|9.51.3|1.3
FCOS3D7.5|17.21.34|0.819.77|10.170.85|0.789.8||10.51.29||1.16
SUP-NeRF10.5|18.80.69|0.607.01|7.070.68|0.7310.6|10.91.22|1.14
GT10.0|18.80.66|0.530.2|1.110.0|1.310.9|11.11.19|1.14
" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.519, + 0.785, + 0.64 + ], + "angle": 0, + "content": "Tab. 6 also confirms the role of NeRF Gradient-Based Pose Refinement (NGPR). By comparing the difference between feed-forward results and the later ones, one can find NGPR merely reduces rotation error or translation error. On the contrary, it can even increase pose error when the initial pose error is small enough, e.g. rotation error \\(< 7^{\\circ}\\), and translation error \\(< 0.75m\\). On the other hand, the improvements of PSNR and depth reconstruction over time indicate that NeRF optimization mostly focuses on shape and texture optimization w.r.t. the visible surface regardless the pose." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.642, + 0.786, + 0.807 + ], + "angle": 0, + "content": "Impact of object NeRF framework. To demonstrate SUP-NeRF's benefits, we compared different network configurations using the same pose refinement approach. For fairness, we used identical subnetwork modules across all configurations. These include a ResNet-based Image Encoder (E), a NeRF Decoder (D), and our Pose Estimation Module (P). CodeNeRF [12] has just a Decoder (D), AutoRF combines an Image Encoder and Decoder \\((\\mathrm{E} + \\mathrm{D})\\), and SUP-NeRF uses all three \\((\\mathrm{E} + \\mathrm{D} + \\mathrm{P})\\). For CodeNeRF and AutoRF, we added a separately trained Pose Network (PNet, composed of \\(\\mathrm{E} + \\mathrm{P}\\)) for initial pose refinement. We also tested a SUP-NeRF version without the cross-task Short Cut (SC) to assess its impact. Tab. 7 shows that SC consistently improves SUP-NeRF across all metrics on nuScenes." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "SUP-NeRF, with joint training, outperforms separate training setups AutoRF + PNet and CodeNeRF + PNet in monocular metrics. Interestingly, while" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "Y. Guo et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 7: Object NeRF Framework Ablations on nuScenes. We used the same subnetwork architectures - the image encoder (E), NeRF decoder (D), and pose estimation module (P) - to create various model configurations. For SUP-NeRF, we also included a version without the cross-task Short Cut (SC) for comparison. [Key: Best, FF=Feed Forward, C=Cross-View]" + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.227, + 0.782, + 0.3 + ], + "angle": 0, + "content": "
Model TrainingPSNR(▲)DE (m)(↓)RE (deg.) (↓)TE (m)(↓)PSNR-C(▲)DE-C (m)(↓)
FF|50itFF|50itFF|50itFF|50itFF|50itFF|50it
CodeNeRF (D)+PNet(E+P)7.6|16.00.87|0.6910.57|9.750.72|0.7611.0|11.21.01|1.04
AutoRF (E+D)+PNet(E+P)9.7|17.50.79|0.668.26|7.870.7|0.745.6|10.11.24|1.16
SUP-NeRF (E+D+P) w/o SC8.7|18.00.94|0.7310.12|9.220.75|0.7410.8|10.91.23|1.11
SUP-NeRF (E+D+P)10.5|18.80.69|0.607.01|7.070.68|0.7310.6|10.91.22|1.14
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.312, + 0.788, + 0.356 + ], + "angle": 0, + "content": "Table 8: Model Size and Running Time Comparison in feed-forward scenarios, and with 20 and 50 iterations of NeRF optimization. SUP-NeRF gets the smallest running time. [Key: Best, FF= Feed Forward]" + }, + { + "type": "table", + "bbox": [ + 0.261, + 0.366, + 0.744, + 0.445 + ], + "angle": 0, + "content": "
MethodParams (M) (↓)FF (s) (↓)FF+20it (s) (↓)FF+50it (s) (↓)
AutoRF+FCOS [29]91.1160.1230.7141.599
(36.166+54.950)(0.114+0.009)(0.114+0.600)(0.114+1.485)
SUP-NeRF49.8160.0180.6081.493
BootInv [34]182.6160.1563.5348.601
SUP-BootInv57.5800.0183.3968.463
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.47, + 0.788, + 0.517 + ], + "angle": 0, + "content": "CodeNeRF lags in monocular, it excels in cross-view evaluations. This implies that using mean shape codes, rather than current image encodings, might reduce overfitting to current view." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.537, + 0.466, + 0.554 + ], + "angle": 0, + "content": "4.4 Running Speed Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.561, + 0.788, + 0.683 + ], + "angle": 0, + "content": "Tab. 8 compares the model size and running speed on a single A5000 Graphics card. Detailed setup and analysis details are in Appendix A8. Tab. 8 shows that SUP-NeRF needs about half or one third as many parameters as others and is 6-8 times faster in the feed-forward (FF) stage. Meanwhile, a significant portion of the run time is consumed by the NeRF optimization step, whether for 20 iterations (20it) or 50 iterations (50it). However, speed limitation of NeRF can be effectively tackled by recent advances in neural rendering [15,31], which we consider as orthogonal to our contribution to the feed-forward stage." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.704, + 0.36, + 0.721 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.843 + ], + "angle": 0, + "content": "This paper introduces SUP-NeRF: a unified network that seamlessly integrates pose estimation and object reconstruction. SUP-NeRF includes a novel pose estimation module to handle scale-depth ambiguity and a new representation invariant to camera changes. Consequently, SUP-NeRF greatly improves the robustness of object pose estimation compared to the standard NeRF baseline. SUP-NeRF outperforms previous methods in both monocular reconstruction and pose estimation tasks, especially in challenging cross-dataset applications." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.654, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SUP-NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.218 + ], + "angle": 0, + "content": "1. Batra, D., Chang, A.X., Chernova, S., Davison, A.J., Deng, J., Koltun, V., Levine, S., Malik, J., Mordatch, I., Mottaghi, R., Savva, M., Su, H.: Rearrangement: A challenge for embodied AI. CoRR (2020) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.219, + 0.785, + 0.246 + ], + "angle": 0, + "content": "2. Bian, W., Wang, Z., Li, K., Bian, J.: Nope-nerf: Optimising neural radiance field with no pose prior. In: CVPR (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.247, + 0.785, + 0.286 + ], + "angle": 0, + "content": "3. Caesar, H., Bankiti, V., Lang, A., Vora, S., Liong, V., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuScenes: A multimodal dataset for autonomous driving. In: CVPR (2020) 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.288, + 0.785, + 0.315 + ], + "angle": 0, + "content": "4. Chen, H., Manhardt, F., Navab, N., Busam, B.: Texpose: Neural texture learning for self-supervised 6d object pose estimation. In: CVPR (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.316, + 0.785, + 0.342 + ], + "angle": 0, + "content": "5. Chen, Y., Lee, G.H.: DBARF: deep bundle-adjusting generalizable neural radiance fields. In: CVPR (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.343, + 0.785, + 0.37 + ], + "angle": 0, + "content": "6. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the KITTI vision benchmark suite. In: CVPR (2012) 4, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.371, + 0.785, + 0.398 + ], + "angle": 0, + "content": "7. Girshick, R., Radosavovic, I., Gkioxari, G., Dollar, P., He, K.: Detectron. https://github.com/facebookresearch/detectron (2018) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.399, + 0.73, + 0.411 + ], + "angle": 0, + "content": "8. Gkioxari, G., Johnson, J., Malik, J.: Mesh R-CNN. In: ICCV (2019) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.412, + 0.785, + 0.439 + ], + "angle": 0, + "content": "9. Henderson, P., Tsiminaki, V., Lampert, C.H.: Leveraging 2d data to learn textured 3d mesh generation. In: CVPR (2020) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.44, + 0.785, + 0.493 + ], + "angle": 0, + "content": "0. Hodañ, T., Michel, F., Brachmann, E., Kehl, W., Glent Buch, A., Kraft, D., Drost, B., Vidal, J., Ihrke, S., Zabulis, X., Sahin, C., Manhardt, F., Tombari, F., Kim, T.K., Matas, J., Rother, C.: BOP: Benchmark for 6D object pose estimation. ECCV (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.495, + 0.785, + 0.536 + ], + "angle": 0, + "content": "1. Insafutdinov, E., Campbell, D., Henriques, J.F., Vedaldi, A.: Snes: Learning probably symmetric neural surfaces from incomplete data. In: Avidan, S., Brostow, G.J., Cissé, M., Farinella, G.M., Hassner, T. (eds.) ECCV. pp. 367-383 (2022) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.537, + 0.785, + 0.563 + ], + "angle": 0, + "content": "2. Jang, W., Agapito, L.: Codenerf: Disentangled neural radiance fields for object categories. In: ICCV (2021) 2, 3, 8, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.565, + 0.785, + 0.605 + ], + "angle": 0, + "content": "3. Kanazawa, A., Tulsiani, S., Efros, A.A., Malik, J.: Learning category-specific mesh reconstruction from image collections. In: Ferrari, V., Hebert, M., Sminchisescu, C., Weiss, Y. (eds.) ECCV (2018) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.606, + 0.785, + 0.633 + ], + "angle": 0, + "content": "4. Kehl, W., Manhardt, F., Tombari, F., Ilic, S., Navab, N.: SSD-6D: making rgb-based 3d detection and 6d pose estimation great again. In: ICCV (2017) 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.634, + 0.785, + 0.66 + ], + "angle": 0, + "content": "5. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph. (2023) 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.661, + 0.785, + 0.688 + ], + "angle": 0, + "content": "6. Kumar, A., Brazil, G., Corona, E., Parchami, A., Liu, X.: DEVIANT: depth equivariant network for monocular 3d object detection. In: ECCV (2022) 2, 4, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.689, + 0.785, + 0.729 + ], + "angle": 0, + "content": "7. Kumar, A., Guo, Y., Huang, X., Ren, L., Liu, X.: SeaBird: Segmentation in Bird's View with Dice Loss Improves Monocular 3D Detection of Large Objects. In: CVPR (2024) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.73, + 0.785, + 0.77 + ], + "angle": 0, + "content": "8. Kundu, A., Genova, K., Yin, X., Fathi, A., Pantofaru, C., Guibas, L.J., Tagliasacchi, A., Dellaert, F., Funkhouser, T.A.: Panoptic neural fields: A semantic object-aware neural scene representation. In: CVPR (2022) 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.785, + 0.798 + ], + "angle": 0, + "content": "9. Kundu, A., Li, Y., Rehg, J.M.: 3d-rcnn: Instance-level 3d object reconstruction via render-and-compare. In: CVPR (2018) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "20. Li, F., Yu, H., Shugurov, I., Busam, B., Yang, S., Ilic, S.: Nerf-posed: A first-reconstruct-then-regress approach for weakly-supervised 6d object pose estimation. CoRR (2022) 4" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "Y. Guo et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.188 + ], + "angle": 0, + "content": "21. Li, Y., Wang, G., Ji, X., Xiang, Y., Fox, D.: Deepim: Deep iterative matching for 6d pose estimation. In: Ferrari, V., Hebert, M., Sminchisescu, C., Weiss, Y. (eds.) ECCV (2018) 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.189, + 0.785, + 0.217 + ], + "angle": 0, + "content": "22. Li, Y., Wang, G., Ji, X., Xiang, Y., Fox, D.: Deepim: Deep iterative matching for 6d pose estimation. In: ECCV (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.217, + 0.785, + 0.242 + ], + "angle": 0, + "content": "23. Lin, Y., Florence, P., Barron, J.T., Rodriguez, A., Isola, P., Lin, T.: inerf: Inverting neural radiance fields for pose estimation. In: IROS (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.243, + 0.785, + 0.27 + ], + "angle": 0, + "content": "24. Lipson, L., Teed, Z., Goyal, A., Deng, J.: Coupled iterative refinement for 6d multi-object pose estimation. In: CVPR (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.27, + 0.785, + 0.311 + ], + "angle": 0, + "content": "25. Lu, Y., Ma, X., Yang, L., Zhang, T., Liu, Y., Chu, Q., Yan, J., Ouyang, W.: Geometry uncertainty projection network for monocular 3d object detection. In: ICCV (2021) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.311, + 0.785, + 0.352 + ], + "angle": 0, + "content": "26. Merrill, N., Guo, Y., Zuo, X., Huang, X., Leutenegger, S., Peng, X., Ren, L., Huang, G.: Symmetry and uncertainty-aware object SLAM for 6 dof object pose estimation. In: CVPR (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.352, + 0.785, + 0.393 + ], + "angle": 0, + "content": "27. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM pp. 99-106 (2021) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.393, + 0.785, + 0.42 + ], + "angle": 0, + "content": "28. Mousavian, A., Anguelov, D., Flynn, J., Kosecka, J.: 3d bounding box estimation using deep learning and geometry. In: CVPR (2017) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.42, + 0.785, + 0.46 + ], + "angle": 0, + "content": "29. Müller, N., Simonelli, A., Porzi, L., Bulò, S.R., Nießner, M., Kontschieder, P.: Autorf: Learning 3d object radiance fields from single view observations. In: CVPR (2022) 2, 3, 4, 6, 7, 9, 10, 11, 12, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.46, + 0.785, + 0.487 + ], + "angle": 0, + "content": "30. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph. pp. 102:1-102:15 (2022) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.785, + 0.514 + ], + "angle": 0, + "content": "31. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph. (2022) 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.514, + 0.785, + 0.541 + ], + "angle": 0, + "content": "32. Park, D., Ambrus, R., Guizilini, V., Li, J., Gaidon, A.: Is pseudo-lidar needed for monocular 3d object detection? In: ICCV (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.541, + 0.785, + 0.568 + ], + "angle": 0, + "content": "33. Park, K., Patten, T., Vincze, M.: Pix2pose: Pix2pose: Pixel-wise coordinate regression of objects for 6d pose estimation. In: ICCV (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.568, + 0.785, + 0.609 + ], + "angle": 0, + "content": "34. Pavllo, D., Tan, D.J., Rakotosaona, M.J., Tombari, F.: Shape, pose, and appearance from a single image via bootstrapped radiance field inversion. In: CVPR (2023) 2, 3, 4, 7, 9, 10, 11, 12, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.609, + 0.785, + 0.648 + ], + "angle": 0, + "content": "35. Rad, M., Lepetit, V.: Bb8: A scalable, accurate, robust to partial occlusion method for predicting the 3d poses of challenging objects without using depth. ICCV (2017) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.649, + 0.785, + 0.676 + ], + "angle": 0, + "content": "36. Simonelli, A., Bulò, S.R., Porzi, L., Lopez-Antequera, M., Kontschieder, P.: Disentangling monocular 3d object detection. In: ICCV (2019) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.677, + 0.785, + 0.704 + ], + "angle": 0, + "content": "37. Solà, J., Deray, J., Atchuthan, D.: A micro lie theory for state estimation in robotics. CoRR (2018) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.704, + 0.785, + 0.772 + ], + "angle": 0, + "content": "38. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., Vasudevan, V., Han, W., Ngiam, J., Zhao, H., Timofeev, A., Ettinger, S., Krivokon, M., Gao, A., Joshi, A., Zhang, Y., Shlens, J., Chen, Z., Anguelov, D.: Scalability in perception for autonomous driving: Waymo open dataset. In: CVPR (2020) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.8 + ], + "angle": 0, + "content": "39. Tekin, B., Sinha, S.N., Fua, P.: Real-time seamless single shot 6d object pose prediction. In: CVPR (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "40. Tian, S., Cai, Y., Yu, H.X., Zakharov, S., Liu, K., Gaidon, A., Li, Y., Wu, J.: Multi-object manipulation via object-centric neural scattering functions. In: CVPR (2023) 2" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.654, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SUP-NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "41. Wang, H., Sridhar, S., Huang, J., Valentin, J., Song, S., Guibas, L.J.: Normalized object coordinate space for category-level 6d object pose and size estimation. In: CVPR (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.787, + 0.217 + ], + "angle": 0, + "content": "42. Wang, T., Zhu, X., Pang, J., Lin, D.: FCOS3D: fully convolutional one-stage monocular 3d object detection. In: ICCVW (2021) 4, 9, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.218, + 0.786, + 0.244 + ], + "angle": 0, + "content": "43. Weihs, L., Deitke, M., Kembhavi, A., Mottaghi, R.: Visual room rearrangement. In: CVPR (June 2021) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.785, + 0.273 + ], + "angle": 0, + "content": "44. Xiang, Y., Schmidt, T., Narayanan, V., Fox, D.: PoseCNN: A convolutional neural network for 6D object pose estimation in cluttered scenes. In: RSS (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.787, + 0.313 + ], + "angle": 0, + "content": "45. Yang, B., Zhang, Y., Xu, Y., Li, Y., Zhou, H., Bao, H., Zhang, G., Cui, Z.: Learning object-compositional neural radiance field for editable scene rendering. In: ICCV (2021) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.786, + 0.342 + ], + "angle": 0, + "content": "46. Yu, A., Ye, V., Tancik, M., Kanazawa, A.: pixelnerf: Neural radiance fields from one or few images. In: CVPR (2021) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.786, + 0.37 + ], + "angle": 0, + "content": "47. Yu, H., Wu, J., Yi, L.: Rotationally equivariant 3d object detection. In: CVPR (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.371, + 0.786, + 0.41 + ], + "angle": 0, + "content": "48. Zakharov, S., Ambrus, R.A., Guizilini, V., Park, D., Kehl, W., Durand, F., Tenenbaum, J.B., Sitzmann, V., Wu, J., Gaidon, A.: Single-shot scene reconstruction. In: CoRL (2021) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.786, + 0.439 + ], + "angle": 0, + "content": "49. Zakharov, S., Kehl, W., Bhargava, A., Gaidon, A.: Autolabeling 3d objects with differentiable rendering of SDF shape priors. In: CVPR (2020) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.44, + 0.786, + 0.466 + ], + "angle": 0, + "content": "50. Zakharov, S., Shugurov, I., Ilic, S.: Dpod: 6d pose object detector and refiner. In: ICCV (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.468, + 0.787, + 0.508 + ], + "angle": 0, + "content": "51. Zhou, A., Kim, M.J., Wang, L., Florence, P., Finn, C.: Nerf in the palm of your hand: Corrective augmentation for robotics via novel-view synthesis. In: CVPR (2023) 2" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.508 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/SUP-NeRF_ A Streamlined Unification of Pose Estimation and NeRF for Monocular 3D Object Reconstruction/c913b291-07c4-425a-807f-281296d9ef23_origin.pdf b/2024/SUP-NeRF_ A Streamlined Unification of Pose Estimation and NeRF for Monocular 3D Object Reconstruction/c913b291-07c4-425a-807f-281296d9ef23_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3e87db78a7fd26ef0b9547f81cfcbf2cb301d839 --- /dev/null +++ b/2024/SUP-NeRF_ A Streamlined Unification of Pose Estimation and NeRF for Monocular 3D Object Reconstruction/c913b291-07c4-425a-807f-281296d9ef23_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7d1bbc12bd7fa97c65639c8c6f0a7b8b3eb29b3f69c4f649719b37fd6f0f8c0b +size 1709105 diff --git a/2024/SUP-NeRF_ A Streamlined Unification of Pose Estimation and NeRF for Monocular 3D Object Reconstruction/full.md b/2024/SUP-NeRF_ A Streamlined Unification of Pose Estimation and NeRF for Monocular 3D Object Reconstruction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b311a4ab92f4f6793a0686fbce336cef6c3a473d --- /dev/null +++ b/2024/SUP-NeRF_ A Streamlined Unification of Pose Estimation and NeRF for Monocular 3D Object Reconstruction/full.md @@ -0,0 +1,280 @@ +# SUP-NeRF: A Streamlined Unification of Pose Estimation and NeRF for Monocular 3D Object Reconstruction + +Yuliang Guo $^{1\star}$ + +Ruoyu Wang + +Abhinav Kumar1,2 + +Xinyu Huang + +Cheng Zhao1 + +Liu Ren1 + +1Bosch Research North America & Bosch Center for Artificial Intelligence (BCAI) + +2Michigan State University + +[yuiliang.guo2,cheng.zhao,ruoyu.wang,xingyu.huang,liu. ren]@us.bosch.com + +$^{2}$ kumarab6@msu.edu + +https://yuliangguo.github.io/supnerf + +![](images/f645d13ac8f4e7560605408b99cea7f685fef0bba1f61c1b7b121faa26b5d499.jpg) +Fig. 1: Teaser. SUP-NeRF is a unified solution that predicts an object's pose, shape, and texture using a single network. SUP-NeRF is trained on real driving scenes with imprecise labels, and it adapts robustly to new cross-dataset scenarios. + +Abstract. Monocular 3D reconstruction for categorical objects heavily relies on accurately perceiving each object's pose. While gradient-based optimization in a NeRF framework updates the initial pose, this paper highlights that scale-depth ambiguity in monocular object reconstruction causes failures when the initial pose deviates moderately from the true pose. Consequently, existing methods often depend on a third-party 3D object to provide an initial object pose, leading to increased complexity and generalization issues. To address these challenges, we present SUP-NeRF, a Streamlined Unification of object Pose estimation and NeRF-based object reconstruction. SUP-NeRF decouples the object's dimension estimation and pose refinement to resolve the scale-depth ambiguity, and introduces a camera-invariant projected-box representation that generalizes cross different domains. While using a dedicated pose estimator that smoothly integrates into an object-centric NeRF, SUP-NeRF is free from external 3D detectors. SUP-NeRF achieves state-of-the-art results in both reconstruction and pose estimation tasks on the nuScenes dataset. Furthermore, SUP-NeRF exhibits exceptional cross-dataset generalization on the KITTI and Waymo datasets, surpassing prior methods with up to $50\%$ reduction in rotation and translation error. + +# 1 Introduction + +Monocular 3D object reconstruction [8,9,13,19,49] is a critical technology with broad applications in autonomous driving [18,29,49], AR/VR, robotics [43], and embodied AI [1,40,51]. Its impact is particularly significant in autonomous driving, where it plays a vital role in tasks like auto-labeling [49] and world-model construction [18,29]. This technology is essential for reconstructing 3D models from typical driving scenarios, where single-view observations are common. This paper focuses explicitly on the problem of recovering an object's pose, shape, and texture from a single image under real driving scenarios, see Fig. 1. Our aim is to devise an efficient solution, thereby unlocking the full potential of monocular reconstruction for active tasks such as planning and control. + +Neural Radiance Field (NeRF) technology [27, 30, 48] has revolutionized the field of 3D reconstruction. It excels at capturing fine details in a scene and generating novel views from existing reconstructions. An exciting branch, Object-centric NeRF [11, 12, 18, 29, 34, 45, 46], takes things further, allowing for flexible data creation and tackling reconstruction challenges from limited number of views by incorporating category-specific prior knowledge. Earlier object-centric NeRF methods [8, 12, 13, 19, 46, 48, 49] require densely overlapping views and precise poses, limiting their use to synthetic data. Later advances [18, 29] enable monocular training with imperfect masks. + +Although promising, object NeRFs show limitations in correcting erroneous object poses. First, the well-known scale-depth ambiguity [16,25,36] also affects the object NeRF optimization. Fig. 2 shows the optimization failure of a monocular NeRF method under longitudinal translation error when it is given complete freedom in adjusting the scale within the normalized space. Secondly, NeRF also struggles with large rotation errors. Appendix A6.1 shows that NeRF typically requires the initial rotation error to be below $25^{\circ}$ , and the pose optimization cannot go lower than $7^{\circ}$ . + +Recent papers [18, 29] use an external monocular 3D detector to provide initial poses to NeRF. However, monocular 3D detectors suffer from generalization issues [16, 17] and increased complexity, ultimately affecting the NeRF pipeline. BootInv [34], a more recent method, tackles pose estimation, shape, and appearance reconstruction in a single framework but requires additional pre-training of dense field generator and struggles with handling moderate occlusions. + +To resolve the scale-depth ambiguity and achieve greater generalization, this paper proposes SUP-NeRF, a Streamlined Unification of Pose estimation and NeRF. SUP-NeRF uses an image encoder to output the object dimensions and then refines for object pose using an iterative process while keeping its size fixed. Besides decoupling of an object's dimension and pose to resolve the scale-depth ambiguity, the introduction of a novel projected-box representation in the iterative update refinement also enhances the generalization to novel data since the representation makes the deep pose refiner independent from camera intrinsic parameters. In addition, SUP-NeRF tackles the complexity challenge by sharing the NeRF image encoder with the pose estimation, making the pose estimation module a seamless fit for object NeRFs. + +![](images/833ef2f859bd16c3c8b98127c27d658962597580f5a0d19d9317c863735f15c1.jpg) +R err: 0.00, T err: 6.85 +Fig. 2: Scale-depth ambiguity in NeRF. Given the input image (left), joint optimization of pose, shape, and texture in NeRF has full freedom to rescale the shape within the normalized shape space (blue box) or move the 3D box. Such phenomenon is observed from the evolution of the rendered objects from iteration 0 (middle) to iteration 50 (right). + +![](images/fd13ec9be23ef6727ec18a22077a79c12b6a7afa4c00a0b5990be1d3b9baf744.jpg) + +![](images/48cfe8c73f976499f20bde598fd27dc4b5a161aa021a162e7103d5c661633f33.jpg) +R err: 0.06, Terr: 6.35 + +In our experiments, we first comprehensively benchmark and conduct ablations on SUP-NeRF on the nuScenes [3] dataset. SUP-NeRF outperforms state-of-the-art (SoTA) methods and achieves SoTA object reconstruction and pose performance on this dataset. We next conduct cross-dataset object reconstruction and pose estimation experiments of the nuScenes models on the KITTI and Waymo datasets to test SUP-NeRF's generalization capability. SUP-NeRF shows more significant improvements over SoTA in such cross-dataset settings. + +In summary, our main contributions include: + +- We introduce SUP-NeRF, an streamlined method unifying object pose estimation and monocular object-centric neural reconstruction. +- We identify depth-scale ambiguities as the key barrier to object NeRF's pose optimization and craft a module to overcome this effectively, while also integrating seamlessly with existing object NeRF frameworks. +- We introduce a novel projected box representation for the pose estimation module, enabling the deep refiner to operate in a space invariant to camera intrinsic parameters, thereby enhancing SUP-NeRF's generalization capability. +- SUP-NeRF achieves SoTA object reconstruction and pose estimation results, outperforming other methods on the nuScenes dataset, and in challenging cross-dataset experiments on both KITTI and Waymo datasets. + +# 2 Related Work + +Object-centric NeRF. Object-centric NeRF [11, 12, 18, 29, 34, 45, 46] is a specialized NeRF variant for modeling specific object categories. Its pipeline feeds a NeRF decoder with shape and texture codes, generating occupancy and color values at 3D positions, followed by volumetric rendering to produce a rendered image. The entire process is differentiable, allowing back-propagation of the photometric loss to update shape codes, texture codes, network weights, and object poses. Early object-centric NeRFs [8, 12, 13, 19, 45, 46, 48, 49] required accurate object poses, full visibility, and dense multi-view observations to reconstruct a categorical model, making them inaccessible for real-world applications. Some approaches [12, 23] claim successful object pose recovery through gradient-based + +Table 1: Literature Summary. SUP-NeRF sets a benchmark for minimal requirements, eliminating the need for external 3D detectors, CAD models, or dense NOCS generators. It is dedicated to the reconstruction of object shapes and pose estimation directly from a single image of driving scenes, employing an object-centric framework. + +
MethodsIndoorOutdoorViewsScopeCADwt. Det3Dwt. NeRFNOCS Gen.
6DoF Pose [21, 26]✓×singleobject✓✓××
3D Detector [16, 42]×✓singleimage×✓××
NeRF [27]✓✓multiscene××✓×
Object-NeRF [18, 29]×✓singleobject××✓×
BootInv [34]×✓singleobject×✓✓✓
SUP-NeRF×✓singleobject×✓✓×
+ +pose refinement from monocular images. However these successes are still limited to synthetic dataset, and oversimplified object/camera poses. Recent methods [18,29] investigate more practical pipelines demonstrating the object-NeRF's potential to train and test on driving scenes with limited views and occluded object data. However, both methods require relatively accurate poses from heavy third-party 3D detectors [32,36]. Other recent works [2,5] try freeing NeRF from external camera poses; however, solving camera poses from densely overlapping views is a different problem than recovering an object's pose from a single image. Compared to these methods, SUP-NeRF recovers object pose from a single image without a 3D detector in the pipeline. A more recent work [34] also unifies pose estimation and NeRF in a single network, and enables high-fidelity object generation. However, its training requires additional NOCS dense generator and it is highly sensitive to moderate occlusions compared to our method. + +Monocular 3D Object Pose Estimation. Advancements in deep learning and the availability of 3D-labeled datasets [3, 6, 10] enable many monocular 3D object detection [16, 24, 25, 33, 36, 42, 47] and 6DoF pose estimation methods [14, 21, 22, 26, 28, 39, 44, 50]. Although estimating 3D information from a single image is ill-posed due to depth-scale ambiguity, leveraging constraints like CAD models or sizes enables accurate object pose recovery via PnP processes [35, 39, 41] or iterative optimization [14, 21]. These indoor 6DoF pose estimation mostly follow an object-centric design which aims to estimate highly accurate object pose within the 2D image patches out of a previous 2D object detector. Recent studies [4, 20] use the NeRF pipeline to reconstruct 3D models or to train model-based keypoint detectors to estimate object poses during inference. However, none of these model-based methods is ideal for outdoor driving scenarios, as pre-reconstruction of models is often impossible in new scenes. Fortunately, inferring the 3D dimensions of objects from an image is feasible when its appearance is tightly bounded to its fine-grained class [28], which keeps an outdoor 3D detector from being fully ill-posed. Nowadays, reliable outdoor 3D detectors usually favors a image-centric design to direct detect many 3D objects captured within the whole image, which usually requires a large backbone network [16, 42]. However, the performance of SoTA detectors drops significantly when applied to datasets with distribution different from the training data [16, 25, 42]. + +![](images/193c1224eb7b64e9afba6de803a5eb9e99a264f4049fbaf4df452e2754e0918b.jpg) +Fig. 3: SUP-NeRF Overview. SUP-NeRF unifies pose estimation and NeRF. The pose estimation module enables SUP-NeRF to work for objects in diverse poses without external 3D detectors. The increase of complexity only constitutes a few MLP layers. + +Tab. 1 outline the differences in methodologies. Unlike existing approaches, SUP-NeRF unifies pose estimation and object NeRF in a streamlined manner, adopting a straightforward object-centric design akin to indoor techniques. Without the need for CAD models, SUP-NeRF is suitable for outdoor driving scenarios. Despite its focused scope on object-centric pose estimation, SUP-NeRF showcases enhanced cross-domain generalization capabilities compared to more complex image-centric 3D detectors within this realm. + +# 3 SUP-NeRF + +Our goal is to create a NeRF framework that bypasses NeRF's limitations without external 3D detectors. Fig. 3 shows the overall pipeline of SUP-NeRF, a streamlined unified network that predicts object pose, shape, and texture from a single image. SUP-NeRF consists of three modules: a Resnet-based encoder, a pose estimation module, and a NeRF decoder. SUP-NeRF processes an image and its occlusion mask, with the encoder turning the masked image into shape, texture, and pose codes and predicted object dimensions. The pose code, object dimensions and a box code, are passed to the pose estimation module to iteratively refine the object pose $[R_{o2c}|T_{o2c}]$ . After a few iterations, the estimated pose is converted to the camera pose $[R_{c2o}|T_{c2o}]$ and passed to the NeRF decoder. In the NeRF phase, the decoder uses the shape, texture codes, and camera pose for volumetric rendering, creating both RGB and occupancy images. + +Pose Estimation Module. The key component of SUP-NeRF is the pose estimation module, which we show in Fig. 4. It aims to provide reliable object poses for targets at all ranges and orientations. Our pose estimation module iterative updates the input pose, based on the visual difference between the input pose and the observed object in the image. Given predicted object dimensions $[H_B, W_B, L_B]$ , camera intrinsic $K$ , and current pose $R^{(t)}, T^{(t)}$ , image projections + +![](images/e0b9fe08d967f5b1f1b3a41d36d651e04c554ab36d004762ad654cf770c78147.jpg) +Fig.4: SUP-NeRF Pose Estimation Module. The pose estimation module of SUP-NeRF iteratively updates the object's pose while preserving scale. It takes the projection of 3D box corners as a visual representation of the input pose and estimate the pose update via comparing it to observed image in a latent embedding space. These designs handle scale-depth ambiguity and make the deep refiner independent from camera intrinsic parameters for better cross-domain generalization. + +of the 3D box corners $B_{proj}^{(t)}$ is computed, which can be interpolated as a visual representation of current pose. After encoding the 8 corners into a higher dimensional box code, SUP-NeRF feeds into the pose refiner along with a pose code from the image encoder to predict the desired pose update $\Delta R^{(t)}, \Delta T^{(t)}$ . SUP-NeRF then composes the pose changes with current pose to compute the next pose state, $R^{(t + 1)}, T^{(t + 1)}$ . This updating process starts from a fully random object pose within the viewing frustum of the image patch, and takes a few iterations such that the final pose approaches the true pose even from a random initial pose. + +The pose estimation module incorporates the following unique designs: + +- SUP-NeRF preserves the object dimensions to ensure unambiguous pose updates. Our implementation, using projected 3D boxes on the image plane, leads to a concise encoding of shape and pose, reducing computation in image rendering and encoding per iteration, as opposed to previous methods [21, 29] relying on full CAD models. Since this representation makes the deep pose refiner's task independent from camera intrinsic parameters, it leads to better generalization to different testing domains, as evidenced in Tabs. 3 and 4. +- SUP-NeRF updates the object pose iteratively rather than conduct a direct pose estimation using a PnP-based [14] because the later is sensitive to missed or swapped corners thus often produces far out-of-distribution poses. Our ablation studies in Tab. 6 confirm this difference. +- SUP-NeRF represents the rotation update in axis-angle representation [37]. During rotation updating process, we convert the axis-angle space 3D rotation $\Delta R^{(t)}$ back to $SO(3)$ , compose it with the previous rotation to compute the new rotation state as $R_{\mathrm{e2}}^{(t + 1)}$ . +- SUP-NeRF represents the translation update in a relative space, indicated as $\Delta T^{(t)}: [v_x^{(t)}, v_y^{(t)}, \rho^{(t)}]$ , where $v_x^{(t)}$ and $v_y^{(t)}$ indicate the location change in + +image space, and $\rho^{(t)}$ indicates the ratio between the target depth and the source depth. Specifically, SUP-NeRF updates the translation in image space + +and depth as: + +$$ +x ^ {(t + 1)} = x ^ {(t)} + v _ {x} ^ {(t)} +$$ + +$$ +y ^ {(t + 1)} = y ^ {(t)} + v _ {y} ^ {(t)} \tag {1} +$$ + +$$ +Z ^ {(t + 1)} = \rho^ {(t)} Z ^ {(t)} +$$ + +where $(x,y)$ represent the object's 2D location in image space, $Z$ indicates the depth. The new 3D location $T_{o2c}^{(t + 1)}$ is then computed from the updated image coordinates, depth and camera intrinsic matrix $K$ . Compared to [21, 34], our definition of $\rho^{(t)}$ leads to a more straight-forward update. + +During the pose updating process indicated by $\oplus$ , both the rotation and translation are updated. Through such representation and updating scheme, SUP-NeRF handles objects at varying distances with the same output space. Further discusses the choice of an effective pose representation and the right coordinate frame are included in Appendices A1 and A2. + +Cross-task Shortcut. Simultaneously perceiving pose, shape, and texture, requires an image encoder with shared initial layers but separate later stages. However, encoding shape, texture, and pose is a challenge as they require different features, making it tough for a shared encoder to meet all needs effectively. Shape and texture encoding demand a pose-invariant approach, as they correspond to each object's frame. On the other hand, pose encoding needs to capture sufficient pose information for later use. This difference leads to conflicting learning signals during joint training, potentially weakening the model's performance. + +SUP-NeRF proposes a straightforward one-step approach to resolve this conflict. After processing through the first four CNN layers separately, SUP-NeRF introduces a cross-task shortcut that deducts pose features from the shape and texture features. This design choice eliminates the shape and texture's reliance on pose, allowing the earlier layers to focus on pose-dependent features for all tasks. This shortcut turns the conflict into a cooperative relationship, enhancing the joint training of different modules, as demonstrated in Tab. 7. The details of our image encoder is described in Appendix A3. + +Training. SUP-NeRF closely follows the loss functions and occupancy mask definitions from AutoRF [29]. The total loss during training is: $L_{train} = L_{rgb} + w_{occ}L_{occ} + w_{pose}(L_{pose}^{(direct)} + \sum_t^K L_{pose}^{(t)})$ , where $w_{occ}$ and $w_{pose}$ indicate the corresponding weights. + +We compare the RGB output to the input image to calculate photometric loss $L_{rgb}$ , and the occupancy values to the mask for occupancy loss $L_{occ}$ . SUP-NeRF additionally compares the projected box corners from the integrative pose estimator $B^{(t)}_{proj}$ with the ground-truth values to compute the pose losses $L_{pose}^{(t)}$ . We define the pose loss function as the mean distance between the matched corners: $L_{pose}^{(*)} = \frac{1}{8}\sum_{i=1}^{8}\sqrt{(x_i^{(*)} - \hat{x}_i)^2 + (y_i^{(*)} - \hat{y}_i)^2}$ , where (*) indicates it compatible with both the direct pose loss and the iterative pose loss, $(x_i^{(*)}, y_i^{(*)})$ denote a input corner position, and $(\hat{x}_i, \hat{y}_i)$ indicates a ground-truth corner po + +sition. SUP-NeRF introduces an additional MLP layer for more accurate pose estimation to directly supervise the 2D coordinates of box corners predicted from the pose code. + +In training, only the network weights are treated as the estimizable variables. Moreover, the NeRF decoder uses ground-truth camera poses $[R_{c2o}^{gt}|T_{c2o}^{gt}]$ instead of predictions from the pose estimation module to ensure more reliable training. Implementation Details. SUP-NeRF's image encoder backbone is based on ResNet50, with shared early layers and separated later layers for pose, shape, and texture encoding. For more information on this, please refer to the Appendix A3. To implement the NeRF decoder, we use CodeNeRF [12] for its simplicity and ease of use. For the differentiable rendering based on NeRF, we set the latent dimension to 256, use 64 samples along each ray, and normalize the shape space by the diagonal length of the observed object's 3D box. For the pose estimation module, we also set the latent dimension to 256 for all MLP layers. Before feeding the input image and occupancy mask into the image encoder, we pad them to make them square and resize them to $128\times 128$ pixels. We then render $32\times 32$ images and occupancy maps to calculate losses. + +To ensure effective iterative pose estimation without sacrificing generalization, SUP-NeRF uses a heuristic to sample a random initial 3D object pose across the entire pose space. For translation, we sample in the relative space. We sample a random 2D location within the image ROI range as the initial projection of the object center, while setting the initial depth to $Z = 20$ . For rotation, we sample from the full range of yaw angles combined with a random rotation angle within a $(-20^{\circ}, 20^{\circ})$ range around a random axis direction. This sampling method covers almost the entire distribution of object poses in the camera frame and is used consistently in both training and inference. + +During training, we use 3 iterations of the pose estimator to generate corresponding pose losses, which allow for pose regression to handle highly diverse poses. We jointly train the unified model for the pose estimation module, image encoder, and NeRF decoder for 40 epochs on the nuScenes dataset using a learning rate of $10^{-4}$ and loss weights of $w_{occ} = 0.1$ , $w_{pos} = 0.01$ to compute the total loss. + +Inference. During inference, SUP-NeRF optimizes the shape code, texture code, and object pose $[R_{o2c}|T_{o2c}]$ keeping the network weights frozen. The total loss $L_{infer}$ is thus $L_{infer} = L_{rgb} + w_{occ}L_{occ}$ , where $w_{occ}$ balances the two. SUP-NeRF minimizes the loss $L_{infer}$ to update the variables. At inference, we run 3 iterations for the feed-forward pose estimation module without updating the shape and color codes, and 50 iterations for the gradient-based joint optimization of shape, texture, and pose. We use a step size of 0.02 for shape and texture code updates and 0.01 for pose updates during joint optimization. + +# 4 Experiments + +Our experiments use three datasets: nuScenes [3], KITTI [6] and Waymo [38]. Our evaluation framework simultaneously measures rendering quality, shape reconstruction accuracy, and pose estimation precision. + +Data Splits. Since these datasets are not specifically for 3D reconstruction, we sifted a subset for better evaluation. We chose daytime sequences and used Mask R-CNN [7] to get instance masks, as nuScenes lacks 2D segmentation masks. We matched these masks with 3D bounding box annotations, categorizing them into foreground, background, and unknown, similar to AutoRF [29]. More details are in Appendix A7. We use the following splits of these datasets: + +- nuScenes Train Split: We collected 87,048 objects from the nuScenes Train split for training. +- nuScenes Val Split: We collected 5,000 random objects from the nuScenes Val split for evaluation. +- KITTI Val Split: We use 4,895 objects from KITTI Val split [6] for evaluation. +- Waymo Val Split: We randomly selected 5,000 objects from Waymo Val split [16] for evaluation. + +Evaluation metrics. Our evaluation framework differs from traditional NeRF assessments as it simultaneously evaluates object reconstruction quality and pose accuracy. We incorporate two monocular reconstruction metrics: (1) the PSNR (Peak Signal-to-Noise Ratio) comparing the rendered and observed images; (2) the Depth Error between the rendered depth map and sparse LiDAR data, denoted as $DE$ . We also include two pose estimation metrics: (3) rotation error denoted as $RE$ ; (4) the translation error denoted as $TE$ . To offer a thorough evaluation, we compare various methods at different stages: post-feed-forward ( $FF$ ), and at the 50th iterations of NeRF joint optimization. Beyond monocular assessments, we also conduct cross-view evaluations for PSNR and depth error metrics, denoted as (5) PSNR- $C$ and (6) $DE-C$ , focusing on objects observed from novel views. As the nuScenes dataset provides object IDs that allows tracking objects across various views, the cross-view evaluation is exclusively conducted on the nuScenes dataset, as explained in Fig. 5. + +Baselines. We compare SUP-NeRF with two baselines: AutoRF [29] and AutoRF [29] + FCOS3D [42] (selected for its monocular capability, and availability of model trained on the nuScenes dataset). Both AutoRF and SUP-NeRF use the same NeRF decoder and image encoder to ensure a fair comparison. The difference lies in SUP-NeRF's additional layers for pose encoding and refinement. During testing, we apply the same random pose generation for both AutoRF and SUP-NeRF methods, whereas the FCOS3D baseline begins with its own pose prediction. We also compare with BootInv [34], which recovers pose, shape, and appearance. Leveraging our flexible design, we developed SUP-BootInv by integrating BootInv's pretrained decoder with our encoder and pose estimator. + +# 4.1 nuScenes Monocular Reconstruction and Pose Estimation + +Tab. 2 shows that SUP-NeRF outperforms two baselines, AutoRF and its combination with FCOS3D on nuScenes Val set. AutoRF alone struggles to improve pose from a random start, resulting in lower PSNR and depth error. SUP-NeRF consistently reduces depth error by over $25\%$ , rotation error by over $30\%$ , and translation error by over $13\%$ compared to the FCOS3D and AutoRF com + +Table 2: nuScenes Monocular Reconstruction and Pose Estimation Results. SUP-NeRF consistently improves AutoRF-based pipelines in all metrics, particularly in the pose estimation metrics. SUP-BootInv also consistently improves BootInv in all metrics. [Key: Best, Second Best, FF= Feed Forward, C= Cross-View] + +
MethodPSNR (↑)DE (m) (↓)RE (deg.) (↓)TE (m) (↓)PSNR-C (↑)DE-C (m) (↓)
FF|50itFF|50itFF|50itFF|50itFF|50itFF|50it
AutoRF [29]3.6|10.611.21|10.0987.52|88.076.04|5.9510.0|8.81.31|1.41
AutoRF+FCOS7.5|17.21.34|0.819.77|10.170.85|0.789.8|10.51.29|1.16
SUP-NeRF10.5|18.80.69|0.607.01|7.070.68|0.7310.6|10.91.22|1.14
BootInv [34]9.4|14.35.01|3.5628.40|28.002.59|2.9110.9|11.81.37|1.35
SUP-BootInv10.9|15.41.95|1.627.11|8.400.64|1.0010.9|11.91.37|1.40
+ +![](images/dd917c13bc6afe2dfae83e07b615f4518207d87f7654f1d560ccb26fb6b4aabf.jpg) +Fig. 5: nuScenes Cross-view Evaluation. Each row shows a set of images of the same object from different angles. We use the example for both monocular pose estimation and NeRF reconstruction. Other images in each row are used to evaluate the reconstructed shape and texture in PSNR and Depth Error (DE). + +bination. Our feed-forward stage also shows promising potential for real-time applications, especially where rendering quality is less critical. + +We next compare BootInv [34] as the baseline, although the performance of original BootInv's pose estimation was subpar on both the nuScenes and KITTI datasets, integrating our robust pose module with this NeRF framework significantly enhanced pose estimation accuracy. This underscores the adaptability of our unified pipeline in augmenting various object NeRF frameworks for improved pose estimation. Moreover, comparisons of BootInv [34] and SUP-BootInv in terms of monocular PSNR and depth error suggest that SUP-BootInv's superior pose estimation contributes to its higher PSNR and lower depth error. More detailed analysis upon BootInv-based pipelines is included in Appendix A5. + +# 4.2 Cross-Dataset Evaluation + +We next compare the generalization capability of all these NeRF models with the cross-dataset experiments. For these experiments, we directly apply nuScenes models to the KITTI and Waymo datasets and report the metrics. + +KITTI. Tab. 3 shows the result of cross-dataset experiments on the KITTI dataset. It shows that SUP-NeRF achieves SoTA performance across all evaluation metrics. SUP-NeRF reduces depth error by about $16\%$ , rotation error by + +Table 3: KITTI Cross-dataset Monocular Reconstruction and Pose Estimation Results. We train all methods on nuScenes dataset, and test on KITTI dataset. Our methods SUP-NeRF and SUP-BootInv consistently show superiority in all metrics compared to the counterpart methods. [Key: Best, Second Best, FF= Feed Forward] + +
MethodPSNR (↑)DE (m) (↓)RE (deg.) (↓)TE (m) (↓)
FF|50itFF|50itFF|50itFF|50it
AutoRF [29]0.4|7.69.89|8.8389.67|90.426.16|6.07
AutoRF+FCOS2.4|13.42.42|1.7411.95|12.52.2|2.09
SUP-NeRF4.0|14.12.19|1.546.79|6.891.06|1.01
BootInv [34]7.0|12.36.11|4.3815.52|15.763.97|3.90
SUP-BootInv7.6|12.72.79|2.219.16|8.841.15|1.67
+ +![](images/bbad4520302c52591efcc56bda809535229a58ab55cfac0ebe5b56cbf23fc5a8.jpg) +Fig. 6: KITTI (Cross-dataset) Qualitative Results. In the top panel, we demonstrate SUP-NeRF executes pose estimation reliably, fast converging from a random initial pose to the true one, and enables neural reconstruction under diverse object poses, occlusion cases under this cross-dataset setup. In the bottom panel, SUP-NeRF is visually compared to the other major competitor, demonstrates sharper rendered image, higher accuracy in shape and pose. + +as large as $50\%$ and translation error also about $50\%$ . Thus, SUP-NeRF shows significantly better geometric performance in cross-dataset setting compared to AutoRF+FCOS, which we attributes to our deep pose module's independence from camera intrinsic parameters. Meanwhile, compared to nuScenes evaluation, all the methods' PSNR and depth error performance drop, which means the cross-dataset generalization of monocular NeRF remains an open challenge. + +Fig. 6 presents the qualitative results on this experiment, demonstrating how SUP-NeRF handles diverse poses and occlusions in pose refinement and the visual comparisons between the two most competitive pipelines. Consistent improvements upon integrating our design into BootInv baseline were also observed. More visual comparisons are in Appendix A9. + +Table 4: Waymo Cross-dataset Monocular Reconstruction and Pose Estimation Results. We train all methods on nuScenes dataset, and test on Waymo dataset. Our methods SUP-NeRF and SUP-BootInv consistently show superiority in all metrics compared to the counterpart methods, the only exception is the comparison to AutoRF+FCOS in Rotation Error. [Key: Best, Second Best, FF= Feed Forward] + +
MethodPSNR (↑)DE (m) (↓)RE (deg.) (↓)TE (m) (↓)
FF|50itFF|50itFF|50itFF|50it
AutoRF [29]0.6|9.86.76|6.5386.56|87.679.1|9.14
AutoRF+FCOS6.2|16.52.43|2.326.65|7.23.22|3.31
SUP-NeRF4.8|17.02.32|1.5610.01|10.61.68|1.54
BootInv [34]8.1|11.08.85|8.2030.78|31.525.26|6.16
SUP-BootInv8.6|11.95.36|4.3810.24|11.041.67|2.53
+ +Waymo. Tab. 4 shows the result of cross-dataset experiments on the Waymo dataset. It shows that SUP-NeRF achieves SoTA performance across most evaluation metrics. The only exception is the superior rotation estimation achieved by FCOS3D, which is likely due to leveraging larger context for heavily occluded cases. However, our advantage in translation estimation is more significant. + +# 4.3 Ablation Studies + +The major ablation studies we conduct includes (i) the impact of pose estimation module designs and (ii) the impact of different object NeRF networks. Other ablation studies to analyze the impact of training with predicted 2D boxes, pose refinement iterations, training epochs, as well as NeRF's capability on handling initial pose errors are included in Appendix A6, owing to lack of space. + +Pose Estimation Module. We evaluate multiple pose estimation designs from two main angles. First, we focus on object-centric pose estimation task alone with evaluations on nuScenes, KITTI, and Waymo datasets. The prototypical designs include: (i) Direct 6D pose regression using MLP layers; (ii) Pose estimation from predicted box corners via Keypoint-based Perspective-n-Point (PnP); (iii) NOCS-based pose estimation as implemented in BootInv [34]; (vi) External monocular 3D detection FCOS3D [42]; (v) Our iterative pose refinement method, which uses projected boxes. For fairness, methods (1), (2), and (5) share the same image encoder, pose encoding layers, differing only in their output layers. Tab. 5 shows that SUP-NeRF consistently outperforms the others, notably in cross-dataset translation estimation. + +The second evaluation assesses the impact of different initial poses on object-NeRF performance within the SUP-NeRF architecture. We additionally compared (i) Ground-Truth (GT) pose; (ii) Random initial pose, to represent the oracle and worst case in pose initialization, and exclude the NOCS method which is part of a different NeRF framework. Given different initial poses, all methods undergo the same NeRF Gradient-based Pose Refinement (NGPR), evaluated immediately and after 50 iterations. Tab. 6 shows that our method significantly surpasses other initial pose estimators in geometric accuracy, demonstrating its effectiveness in enhancing 3D model quality. + +Table 5: Pose Estimation Ablation Studies. We evaluate all the candidate pose estimation methods with a focus on pose accuracy alone, on nuScenes, KITTI, and Waymo validation sets. Our method shows outstanding robustness in all cases, particularly in translation accuracy in cross-dataset tests. [Key: Best] + +
Pose ModuleRE (deg.)(↓)TE (m) (↓)
nuScenesKITTIWaymonuScenesKITTIWaymo
MLP Pose31.9637.5830.895.535.3110.1
Corners+PnP24.7935.1093.762.823.5518.68
NOCS +PnP(BootInv)28.4015.5223.662.593.973.47
FCOS3D9.7711.956.650.852.23.22
SUP-NeRF7.016.7910.010.681.061.54
+ +Table 6: Ablation studies of Pose Initialization in Object NeRF on nuScenes. We compare different choices of initial poses under SUP-NeRF frameworks. As shown, SUP-NeRF presents superiors effectiveness in all the metrics. [Key: Best, FF=Feed Forward, C=Cross-View] + +
Initial PosePSNR (↑)DE (m) (↓)RE (deg.) (↓)TE (m) (↓)PSNR-C (↑)DE-C (m) (↓)
FF|50itFF|50itFF|50itFF|50itFF|50itFF|50it
Random3.6|10.611.21|10.0987.52|88.076.04|5.959.1|7.41.42|1.82
MLP Pose1.2|6.95.33|4.4131.96|24.575.53|2.789.8|8.51.3|1.4
Corners+PnP7.2|16.22.18|1.8324.79|24.572.82|2.789.9|9.51.3|1.3
FCOS3D7.5|17.21.34|0.819.77|10.170.85|0.789.8||10.51.29||1.16
SUP-NeRF10.5|18.80.69|0.607.01|7.070.68|0.7310.6|10.91.22|1.14
GT10.0|18.80.66|0.530.2|1.110.0|1.310.9|11.11.19|1.14
+ +Tab. 6 also confirms the role of NeRF Gradient-Based Pose Refinement (NGPR). By comparing the difference between feed-forward results and the later ones, one can find NGPR merely reduces rotation error or translation error. On the contrary, it can even increase pose error when the initial pose error is small enough, e.g. rotation error $< 7^{\circ}$ , and translation error $< 0.75m$ . On the other hand, the improvements of PSNR and depth reconstruction over time indicate that NeRF optimization mostly focuses on shape and texture optimization w.r.t. the visible surface regardless the pose. + +Impact of object NeRF framework. To demonstrate SUP-NeRF's benefits, we compared different network configurations using the same pose refinement approach. For fairness, we used identical subnetwork modules across all configurations. These include a ResNet-based Image Encoder (E), a NeRF Decoder (D), and our Pose Estimation Module (P). CodeNeRF [12] has just a Decoder (D), AutoRF combines an Image Encoder and Decoder $(\mathrm{E} + \mathrm{D})$ , and SUP-NeRF uses all three $(\mathrm{E} + \mathrm{D} + \mathrm{P})$ . For CodeNeRF and AutoRF, we added a separately trained Pose Network (PNet, composed of $\mathrm{E} + \mathrm{P}$ ) for initial pose refinement. We also tested a SUP-NeRF version without the cross-task Short Cut (SC) to assess its impact. Tab. 7 shows that SC consistently improves SUP-NeRF across all metrics on nuScenes. + +SUP-NeRF, with joint training, outperforms separate training setups AutoRF + PNet and CodeNeRF + PNet in monocular metrics. Interestingly, while + +Table 7: Object NeRF Framework Ablations on nuScenes. We used the same subnetwork architectures - the image encoder (E), NeRF decoder (D), and pose estimation module (P) - to create various model configurations. For SUP-NeRF, we also included a version without the cross-task Short Cut (SC) for comparison. [Key: Best, FF=Feed Forward, C=Cross-View] + +
Model TrainingPSNR(▲)DE (m)(↓)RE (deg.) (↓)TE (m)(↓)PSNR-C(▲)DE-C (m)(↓)
FF|50itFF|50itFF|50itFF|50itFF|50itFF|50it
CodeNeRF (D)+PNet(E+P)7.6|16.00.87|0.6910.57|9.750.72|0.7611.0|11.21.01|1.04
AutoRF (E+D)+PNet(E+P)9.7|17.50.79|0.668.26|7.870.7|0.745.6|10.11.24|1.16
SUP-NeRF (E+D+P) w/o SC8.7|18.00.94|0.7310.12|9.220.75|0.7410.8|10.91.23|1.11
SUP-NeRF (E+D+P)10.5|18.80.69|0.607.01|7.070.68|0.7310.6|10.91.22|1.14
+ +Table 8: Model Size and Running Time Comparison in feed-forward scenarios, and with 20 and 50 iterations of NeRF optimization. SUP-NeRF gets the smallest running time. [Key: Best, FF= Feed Forward] + +
MethodParams (M) (↓)FF (s) (↓)FF+20it (s) (↓)FF+50it (s) (↓)
AutoRF+FCOS [29]91.1160.1230.7141.599
(36.166+54.950)(0.114+0.009)(0.114+0.600)(0.114+1.485)
SUP-NeRF49.8160.0180.6081.493
BootInv [34]182.6160.1563.5348.601
SUP-BootInv57.5800.0183.3968.463
+ +CodeNeRF lags in monocular, it excels in cross-view evaluations. This implies that using mean shape codes, rather than current image encodings, might reduce overfitting to current view. + +# 4.4 Running Speed Analysis + +Tab. 8 compares the model size and running speed on a single A5000 Graphics card. Detailed setup and analysis details are in Appendix A8. Tab. 8 shows that SUP-NeRF needs about half or one third as many parameters as others and is 6-8 times faster in the feed-forward (FF) stage. Meanwhile, a significant portion of the run time is consumed by the NeRF optimization step, whether for 20 iterations (20it) or 50 iterations (50it). However, speed limitation of NeRF can be effectively tackled by recent advances in neural rendering [15,31], which we consider as orthogonal to our contribution to the feed-forward stage. + +# 5 Conclusion + +This paper introduces SUP-NeRF: a unified network that seamlessly integrates pose estimation and object reconstruction. SUP-NeRF includes a novel pose estimation module to handle scale-depth ambiguity and a new representation invariant to camera changes. Consequently, SUP-NeRF greatly improves the robustness of object pose estimation compared to the standard NeRF baseline. SUP-NeRF outperforms previous methods in both monocular reconstruction and pose estimation tasks, especially in challenging cross-dataset applications. + +# References + +1. Batra, D., Chang, A.X., Chernova, S., Davison, A.J., Deng, J., Koltun, V., Levine, S., Malik, J., Mordatch, I., Mottaghi, R., Savva, M., Su, H.: Rearrangement: A challenge for embodied AI. CoRR (2020) 2 +2. Bian, W., Wang, Z., Li, K., Bian, J.: Nope-nerf: Optimising neural radiance field with no pose prior. In: CVPR (2023) 4 +3. Caesar, H., Bankiti, V., Lang, A., Vora, S., Liong, V., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuScenes: A multimodal dataset for autonomous driving. In: CVPR (2020) 3, 4, 8 +4. Chen, H., Manhardt, F., Navab, N., Busam, B.: Texpose: Neural texture learning for self-supervised 6d object pose estimation. In: CVPR (2023) 4 +5. Chen, Y., Lee, G.H.: DBARF: deep bundle-adjusting generalizable neural radiance fields. In: CVPR (2023) 4 +6. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the KITTI vision benchmark suite. In: CVPR (2012) 4, 8, 9 +7. Girshick, R., Radosavovic, I., Gkioxari, G., Dollar, P., He, K.: Detectron. https://github.com/facebookresearch/detectron (2018) 9 +8. Gkioxari, G., Johnson, J., Malik, J.: Mesh R-CNN. In: ICCV (2019) 2, 3 +9. Henderson, P., Tsiminaki, V., Lampert, C.H.: Leveraging 2d data to learn textured 3d mesh generation. In: CVPR (2020) 2 +0. Hodañ, T., Michel, F., Brachmann, E., Kehl, W., Glent Buch, A., Kraft, D., Drost, B., Vidal, J., Ihrke, S., Zabulis, X., Sahin, C., Manhardt, F., Tombari, F., Kim, T.K., Matas, J., Rother, C.: BOP: Benchmark for 6D object pose estimation. ECCV (2018) 4 +1. Insafutdinov, E., Campbell, D., Henriques, J.F., Vedaldi, A.: Snes: Learning probably symmetric neural surfaces from incomplete data. In: Avidan, S., Brostow, G.J., Cissé, M., Farinella, G.M., Hassner, T. (eds.) ECCV. pp. 367-383 (2022) 2, 3 +2. Jang, W., Agapito, L.: Codenerf: Disentangled neural radiance fields for object categories. In: ICCV (2021) 2, 3, 8, 13 +3. Kanazawa, A., Tulsiani, S., Efros, A.A., Malik, J.: Learning category-specific mesh reconstruction from image collections. In: Ferrari, V., Hebert, M., Sminchisescu, C., Weiss, Y. (eds.) ECCV (2018) 2, 3 +4. Kehl, W., Manhardt, F., Tombari, F., Ilic, S., Navab, N.: SSD-6D: making rgb-based 3d detection and 6d pose estimation great again. In: ICCV (2017) 4, 6 +5. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph. (2023) 14 +6. Kumar, A., Brazil, G., Corona, E., Parchami, A., Liu, X.: DEVIANT: depth equivariant network for monocular 3d object detection. In: ECCV (2022) 2, 4, 9 +7. Kumar, A., Guo, Y., Huang, X., Ren, L., Liu, X.: SeaBird: Segmentation in Bird's View with Dice Loss Improves Monocular 3D Detection of Large Objects. In: CVPR (2024) 2 +8. Kundu, A., Genova, K., Yin, X., Fathi, A., Pantofaru, C., Guibas, L.J., Tagliasacchi, A., Dellaert, F., Funkhouser, T.A.: Panoptic neural fields: A semantic object-aware neural scene representation. In: CVPR (2022) 2, 3, 4 +9. Kundu, A., Li, Y., Rehg, J.M.: 3d-rcnn: Instance-level 3d object reconstruction via render-and-compare. In: CVPR (2018) 2, 3 +20. Li, F., Yu, H., Shugurov, I., Busam, B., Yang, S., Ilic, S.: Nerf-posed: A first-reconstruct-then-regress approach for weakly-supervised 6d object pose estimation. CoRR (2022) 4 + +21. Li, Y., Wang, G., Ji, X., Xiang, Y., Fox, D.: Deepim: Deep iterative matching for 6d pose estimation. In: Ferrari, V., Hebert, M., Sminchisescu, C., Weiss, Y. (eds.) ECCV (2018) 4, 6, 7 +22. Li, Y., Wang, G., Ji, X., Xiang, Y., Fox, D.: Deepim: Deep iterative matching for 6d pose estimation. In: ECCV (2018) 4 +23. Lin, Y., Florence, P., Barron, J.T., Rodriguez, A., Isola, P., Lin, T.: inerf: Inverting neural radiance fields for pose estimation. In: IROS (2021) 3 +24. Lipson, L., Teed, Z., Goyal, A., Deng, J.: Coupled iterative refinement for 6d multi-object pose estimation. In: CVPR (2022) 4 +25. Lu, Y., Ma, X., Yang, L., Zhang, T., Liu, Y., Chu, Q., Yan, J., Ouyang, W.: Geometry uncertainty projection network for monocular 3d object detection. In: ICCV (2021) 2, 4 +26. Merrill, N., Guo, Y., Zuo, X., Huang, X., Leutenegger, S., Peng, X., Ren, L., Huang, G.: Symmetry and uncertainty-aware object SLAM for 6 dof object pose estimation. In: CVPR (2022) 4 +27. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM pp. 99-106 (2021) 2, 4 +28. Mousavian, A., Anguelov, D., Flynn, J., Kosecka, J.: 3d bounding box estimation using deep learning and geometry. In: CVPR (2017) 4 +29. Müller, N., Simonelli, A., Porzi, L., Bulò, S.R., Nießner, M., Kontschieder, P.: Autorf: Learning 3d object radiance fields from single view observations. In: CVPR (2022) 2, 3, 4, 6, 7, 9, 10, 11, 12, 14 +30. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph. pp. 102:1-102:15 (2022) 2 +31. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph. (2022) 14 +32. Park, D., Ambrus, R., Guizilini, V., Li, J., Gaidon, A.: Is pseudo-lidar needed for monocular 3d object detection? In: ICCV (2021) 4 +33. Park, K., Patten, T., Vincze, M.: Pix2pose: Pix2pose: Pixel-wise coordinate regression of objects for 6d pose estimation. In: ICCV (2019) 4 +34. Pavllo, D., Tan, D.J., Rakotosaona, M.J., Tombari, F.: Shape, pose, and appearance from a single image via bootstrapped radiance field inversion. In: CVPR (2023) 2, 3, 4, 7, 9, 10, 11, 12, 14 +35. Rad, M., Lepetit, V.: Bb8: A scalable, accurate, robust to partial occlusion method for predicting the 3d poses of challenging objects without using depth. ICCV (2017) 4 +36. Simonelli, A., Bulò, S.R., Porzi, L., Lopez-Antequera, M., Kontschieder, P.: Disentangling monocular 3d object detection. In: ICCV (2019) 2, 4 +37. Solà, J., Deray, J., Atchuthan, D.: A micro lie theory for state estimation in robotics. CoRR (2018) 6 +38. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., Vasudevan, V., Han, W., Ngiam, J., Zhao, H., Timofeev, A., Ettinger, S., Krivokon, M., Gao, A., Joshi, A., Zhang, Y., Shlens, J., Chen, Z., Anguelov, D.: Scalability in perception for autonomous driving: Waymo open dataset. In: CVPR (2020) 8 +39. Tekin, B., Sinha, S.N., Fua, P.: Real-time seamless single shot 6d object pose prediction. In: CVPR (2018) 4 +40. Tian, S., Cai, Y., Yu, H.X., Zakharov, S., Liu, K., Gaidon, A., Li, Y., Wu, J.: Multi-object manipulation via object-centric neural scattering functions. In: CVPR (2023) 2 + +41. Wang, H., Sridhar, S., Huang, J., Valentin, J., Song, S., Guibas, L.J.: Normalized object coordinate space for category-level 6d object pose and size estimation. In: CVPR (2019) 4 +42. Wang, T., Zhu, X., Pang, J., Lin, D.: FCOS3D: fully convolutional one-stage monocular 3d object detection. In: ICCVW (2021) 4, 9, 12 +43. Weihs, L., Deitke, M., Kembhavi, A., Mottaghi, R.: Visual room rearrangement. In: CVPR (June 2021) 2 +44. Xiang, Y., Schmidt, T., Narayanan, V., Fox, D.: PoseCNN: A convolutional neural network for 6D object pose estimation in cluttered scenes. In: RSS (2018) 4 +45. Yang, B., Zhang, Y., Xu, Y., Li, Y., Zhou, H., Bao, H., Zhang, G., Cui, Z.: Learning object-compositional neural radiance field for editable scene rendering. In: ICCV (2021) 2, 3 +46. Yu, A., Ye, V., Tancik, M., Kanazawa, A.: pixelnerf: Neural radiance fields from one or few images. In: CVPR (2021) 2, 3 +47. Yu, H., Wu, J., Yi, L.: Rotationally equivariant 3d object detection. In: CVPR (2022) 4 +48. Zakharov, S., Ambrus, R.A., Guizilini, V., Park, D., Kehl, W., Durand, F., Tenenbaum, J.B., Sitzmann, V., Wu, J., Gaidon, A.: Single-shot scene reconstruction. In: CoRL (2021) 2, 3 +49. Zakharov, S., Kehl, W., Bhargava, A., Gaidon, A.: Autolabeling 3d objects with differentiable rendering of SDF shape priors. In: CVPR (2020) 2, 3 +50. Zakharov, S., Shugurov, I., Ilic, S.: Dpod: 6d pose object detector and refiner. In: ICCV (2019) 4 +51. Zhou, A., Kim, M.J., Wang, L., Florence, P., Finn, C.: Nerf in the palm of your hand: Corrective augmentation for robotics via novel-view synthesis. In: CVPR (2023) 2 \ No newline at end of file diff --git a/2024/SUP-NeRF_ A Streamlined Unification of Pose Estimation and NeRF for Monocular 3D Object Reconstruction/images.zip b/2024/SUP-NeRF_ A Streamlined Unification of Pose Estimation and NeRF for Monocular 3D Object Reconstruction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a8355af95fa59554c8ebd13bdb21981b4ad74823 --- /dev/null +++ b/2024/SUP-NeRF_ A Streamlined Unification of Pose Estimation and NeRF for Monocular 3D Object Reconstruction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0b3cb57ac0b12a679ecf28f5c2ae69a2b5171408c23a45ea7d4c5df05968e5ad +size 568231 diff --git a/2024/SUP-NeRF_ A Streamlined Unification of Pose Estimation and NeRF for Monocular 3D Object Reconstruction/layout.json b/2024/SUP-NeRF_ A Streamlined Unification of Pose Estimation and NeRF for Monocular 3D Object Reconstruction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6af04c9e46b62d761bc13181050f9ae271814cfa --- /dev/null +++ b/2024/SUP-NeRF_ A Streamlined Unification of Pose Estimation and NeRF for Monocular 3D Object Reconstruction/layout.json @@ -0,0 +1,8479 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 139, + 111, + 475, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 111, + 475, + 163 + ], + "spans": [ + { + "bbox": [ + 139, + 111, + 475, + 163 + ], + "type": "text", + "content": "SUP-NeRF: A Streamlined Unification of Pose Estimation and NeRF for Monocular 3D Object Reconstruction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 178, + 185, + 243, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 185, + 243, + 198 + ], + "spans": [ + { + "bbox": [ + 178, + 185, + 243, + 198 + ], + "type": "text", + "content": "Yuliang Guo" + }, + { + "bbox": [ + 178, + 185, + 243, + 198 + ], + "type": "inline_equation", + "content": "^{1\\star}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 179, + 199, + 241, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 199, + 241, + 210 + ], + "spans": [ + { + "bbox": [ + 179, + 199, + 241, + 210 + ], + "type": "text", + "content": "Ruoyu Wang" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 269, + 186, + 351, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 269, + 186, + 351, + 198 + ], + "spans": [ + { + "bbox": [ + 269, + 186, + 351, + 198 + ], + "type": "text", + "content": "Abhinav Kumar1,2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 280, + 199, + 343, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 280, + 199, + 343, + 210 + ], + "spans": [ + { + "bbox": [ + 280, + 199, + 343, + 210 + ], + "type": "text", + "content": "Xinyu Huang" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 376, + 186, + 435, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 186, + 435, + 198 + ], + "spans": [ + { + "bbox": [ + 376, + 186, + 435, + 198 + ], + "type": "text", + "content": "Cheng Zhao1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 386, + 199, + 427, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 199, + 427, + 209 + ], + "spans": [ + { + "bbox": [ + 386, + 199, + 427, + 209 + ], + "type": "text", + "content": "Liu Ren1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 213, + 473, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 213, + 473, + 225 + ], + "spans": [ + { + "bbox": [ + 138, + 213, + 473, + 225 + ], + "type": "text", + "content": "1Bosch Research North America & Bosch Center for Artificial Intelligence (BCAI)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 251, + 225, + 362, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 225, + 362, + 236 + ], + "spans": [ + { + "bbox": [ + 251, + 225, + 362, + 236 + ], + "type": "text", + "content": "2Michigan State University" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 139, + 236, + 474, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 236, + 474, + 247 + ], + "spans": [ + { + "bbox": [ + 139, + 236, + 474, + 247 + ], + "type": "text", + "content": "[yuiliang.guo2,cheng.zhao,ruoyu.wang,xingyu.huang,liu. ren]@us.bosch.com" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 266, + 247, + 347, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 247, + 347, + 257 + ], + "spans": [ + { + "bbox": [ + 266, + 247, + 347, + 257 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 266, + 247, + 347, + 257 + ], + "type": "text", + "content": "kumarab6@msu.edu" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 220, + 258, + 392, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 258, + 392, + 270 + ], + "spans": [ + { + "bbox": [ + 220, + 258, + 392, + 270 + ], + "type": "text", + "content": "https://yuliangguo.github.io/supnerf" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 155, + 270, + 458, + 401 + ], + "blocks": [ + { + "bbox": [ + 155, + 270, + 458, + 401 + ], + "lines": [ + { + "bbox": [ + 155, + 270, + 458, + 401 + ], + "spans": [ + { + "bbox": [ + 155, + 270, + 458, + 401 + ], + "type": "image", + "image_path": "f645d13ac8f4e7560605408b99cea7f685fef0bba1f61c1b7b121faa26b5d499.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 404, + 482, + 437 + ], + "lines": [ + { + "bbox": [ + 131, + 404, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 131, + 404, + 482, + 437 + ], + "type": "text", + "content": "Fig. 1: Teaser. SUP-NeRF is a unified solution that predicts an object's pose, shape, and texture using a single network. SUP-NeRF is trained on real driving scenes with imprecise labels, and it adapts robustly to new cross-dataset scenarios." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 159, + 441, + 453, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 441, + 453, + 650 + ], + "spans": [ + { + "bbox": [ + 159, + 441, + 453, + 650 + ], + "type": "text", + "content": "Abstract. Monocular 3D reconstruction for categorical objects heavily relies on accurately perceiving each object's pose. While gradient-based optimization in a NeRF framework updates the initial pose, this paper highlights that scale-depth ambiguity in monocular object reconstruction causes failures when the initial pose deviates moderately from the true pose. Consequently, existing methods often depend on a third-party 3D object to provide an initial object pose, leading to increased complexity and generalization issues. To address these challenges, we present SUP-NeRF, a Streamlined Unification of object Pose estimation and NeRF-based object reconstruction. SUP-NeRF decouples the object's dimension estimation and pose refinement to resolve the scale-depth ambiguity, and introduces a camera-invariant projected-box representation that generalizes cross different domains. While using a dedicated pose estimator that smoothly integrates into an object-centric NeRF, SUP-NeRF is free from external 3D detectors. SUP-NeRF achieves state-of-the-art results in both reconstruction and pose estimation tasks on the nuScenes dataset. Furthermore, SUP-NeRF exhibits exceptional cross-dataset generalization on the KITTI and Waymo datasets, surpassing prior methods with up to " + }, + { + "bbox": [ + 159, + 441, + 453, + 650 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 159, + 441, + 453, + 650 + ], + "type": "text", + "content": " reduction in rotation and translation error." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 135, + 654, + 197, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 654, + 197, + 665 + ], + "spans": [ + { + "bbox": [ + 135, + 654, + 197, + 665 + ], + "type": "text", + "content": "* Project Lead" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 229, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 229, + 127 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 229, + 127 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 481, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 481, + 258 + ], + "type": "text", + "content": "Monocular 3D object reconstruction [8,9,13,19,49] is a critical technology with broad applications in autonomous driving [18,29,49], AR/VR, robotics [43], and embodied AI [1,40,51]. Its impact is particularly significant in autonomous driving, where it plays a vital role in tasks like auto-labeling [49] and world-model construction [18,29]. This technology is essential for reconstructing 3D models from typical driving scenarios, where single-view observations are common. This paper focuses explicitly on the problem of recovering an object's pose, shape, and texture from a single image under real driving scenarios, see Fig. 1. Our aim is to devise an efficient solution, thereby unlocking the full potential of monocular reconstruction for active tasks such as planning and control." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 258, + 481, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 481, + 366 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 481, + 366 + ], + "type": "text", + "content": "Neural Radiance Field (NeRF) technology [27, 30, 48] has revolutionized the field of 3D reconstruction. It excels at capturing fine details in a scene and generating novel views from existing reconstructions. An exciting branch, Object-centric NeRF [11, 12, 18, 29, 34, 45, 46], takes things further, allowing for flexible data creation and tackling reconstruction challenges from limited number of views by incorporating category-specific prior knowledge. Earlier object-centric NeRF methods [8, 12, 13, 19, 46, 48, 49] require densely overlapping views and precise poses, limiting their use to synthetic data. Later advances [18, 29] enable monocular training with imperfect masks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 366, + 481, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 366, + 481, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 366, + 481, + 461 + ], + "type": "text", + "content": "Although promising, object NeRFs show limitations in correcting erroneous object poses. First, the well-known scale-depth ambiguity [16,25,36] also affects the object NeRF optimization. Fig. 2 shows the optimization failure of a monocular NeRF method under longitudinal translation error when it is given complete freedom in adjusting the scale within the normalized space. Secondly, NeRF also struggles with large rotation errors. Appendix A6.1 shows that NeRF typically requires the initial rotation error to be below " + }, + { + "bbox": [ + 130, + 366, + 481, + 461 + ], + "type": "inline_equation", + "content": "25^{\\circ}" + }, + { + "bbox": [ + 130, + 366, + 481, + 461 + ], + "type": "text", + "content": ", and the pose optimization cannot go lower than " + }, + { + "bbox": [ + 130, + 366, + 481, + 461 + ], + "type": "inline_equation", + "content": "7^{\\circ}" + }, + { + "bbox": [ + 130, + 366, + 481, + 461 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 462, + 481, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 481, + 533 + ], + "type": "text", + "content": "Recent papers [18, 29] use an external monocular 3D detector to provide initial poses to NeRF. However, monocular 3D detectors suffer from generalization issues [16, 17] and increased complexity, ultimately affecting the NeRF pipeline. BootInv [34], a more recent method, tackles pose estimation, shape, and appearance reconstruction in a single framework but requires additional pre-training of dense field generator and struggles with handling moderate occlusions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 481, + 665 + ], + "type": "text", + "content": "To resolve the scale-depth ambiguity and achieve greater generalization, this paper proposes SUP-NeRF, a Streamlined Unification of Pose estimation and NeRF. SUP-NeRF uses an image encoder to output the object dimensions and then refines for object pose using an iterative process while keeping its size fixed. Besides decoupling of an object's dimension and pose to resolve the scale-depth ambiguity, the introduction of a novel projected-box representation in the iterative update refinement also enhances the generalization to novel data since the representation makes the deep pose refiner independent from camera intrinsic parameters. In addition, SUP-NeRF tackles the complexity challenge by sharing the NeRF image encoder with the pose estimation, making the pose estimation module a seamless fit for object NeRFs." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "Y. Guo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 186, + 115, + 267, + 198 + ], + "blocks": [ + { + "bbox": [ + 186, + 115, + 267, + 198 + ], + "lines": [ + { + "bbox": [ + 186, + 115, + 267, + 198 + ], + "spans": [ + { + "bbox": [ + 186, + 115, + 267, + 198 + ], + "type": "image", + "image_path": "833ef2f859bd16c3c8b98127c27d658962597580f5a0d19d9317c863735f15c1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 268, + 121, + 343, + 129 + ], + "lines": [ + { + "bbox": [ + 268, + 121, + 343, + 129 + ], + "spans": [ + { + "bbox": [ + 268, + 121, + 343, + 129 + ], + "type": "text", + "content": "R err: 0.00, T err: 6.85" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 206, + 480, + 262 + ], + "lines": [ + { + "bbox": [ + 130, + 206, + 480, + 262 + ], + "spans": [ + { + "bbox": [ + 130, + 206, + 480, + 262 + ], + "type": "text", + "content": "Fig. 2: Scale-depth ambiguity in NeRF. Given the input image (left), joint optimization of pose, shape, and texture in NeRF has full freedom to rescale the shape within the normalized shape space (blue box) or move the 3D box. Such phenomenon is observed from the evolution of the rendered objects from iteration 0 (middle) to iteration 50 (right)." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 275, + 139, + 342, + 172 + ], + "blocks": [ + { + "bbox": [ + 275, + 139, + 342, + 172 + ], + "lines": [ + { + "bbox": [ + 275, + 139, + 342, + 172 + ], + "spans": [ + { + "bbox": [ + 275, + 139, + 342, + 172 + ], + "type": "image", + "image_path": "fd13ec9be23ef6727ec18a22077a79c12b6a7afa4c00a0b5990be1d3b9baf744.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 356, + 137, + 425, + 173 + ], + "blocks": [ + { + "bbox": [ + 353, + 121, + 428, + 129 + ], + "lines": [ + { + "bbox": [ + 353, + 121, + 428, + 129 + ], + "spans": [ + { + "bbox": [ + 353, + 121, + 428, + 129 + ], + "type": "text", + "content": "R err: 0.06, Terr: 6.35" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 356, + 137, + 425, + 173 + ], + "lines": [ + { + "bbox": [ + 356, + 137, + 425, + 173 + ], + "spans": [ + { + "bbox": [ + 356, + 137, + 425, + 173 + ], + "type": "image", + "image_path": "48cfe8c73f976499f20bde598fd27dc4b5a161aa021a162e7103d5c661633f33.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 270, + 480, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 480, + 353 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 480, + 353 + ], + "type": "text", + "content": "In our experiments, we first comprehensively benchmark and conduct ablations on SUP-NeRF on the nuScenes [3] dataset. SUP-NeRF outperforms state-of-the-art (SoTA) methods and achieves SoTA object reconstruction and pose performance on this dataset. We next conduct cross-dataset object reconstruction and pose estimation experiments of the nuScenes models on the KITTI and Waymo datasets to test SUP-NeRF's generalization capability. SUP-NeRF shows more significant improvements over SoTA in such cross-dataset settings." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 146, + 354, + 345, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 354, + 345, + 365 + ], + "spans": [ + { + "bbox": [ + 146, + 354, + 345, + 365 + ], + "type": "text", + "content": "In summary, our main contributions include:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 129, + 366, + 480, + 498 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 129, + 366, + 479, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 366, + 479, + 389 + ], + "spans": [ + { + "bbox": [ + 129, + 366, + 479, + 389 + ], + "type": "text", + "content": "- We introduce SUP-NeRF, an streamlined method unifying object pose estimation and monocular object-centric neural reconstruction." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 129, + 390, + 480, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 390, + 480, + 426 + ], + "spans": [ + { + "bbox": [ + 129, + 390, + 480, + 426 + ], + "type": "text", + "content": "- We identify depth-scale ambiguities as the key barrier to object NeRF's pose optimization and craft a module to overcome this effectively, while also integrating seamlessly with existing object NeRF frameworks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 129, + 426, + 480, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 426, + 480, + 461 + ], + "spans": [ + { + "bbox": [ + 129, + 426, + 480, + 461 + ], + "type": "text", + "content": "- We introduce a novel projected box representation for the pose estimation module, enabling the deep refiner to operate in a space invariant to camera intrinsic parameters, thereby enhancing SUP-NeRF's generalization capability." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 129, + 462, + 480, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 462, + 480, + 498 + ], + "spans": [ + { + "bbox": [ + 129, + 462, + 480, + 498 + ], + "type": "text", + "content": "- SUP-NeRF achieves SoTA object reconstruction and pose estimation results, outperforming other methods on the nuScenes dataset, and in challenging cross-dataset experiments on both KITTI and Waymo datasets." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 518, + 236, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 518, + 236, + 530 + ], + "spans": [ + { + "bbox": [ + 132, + 518, + 236, + 530 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 545, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 480, + 665 + ], + "type": "text", + "content": "Object-centric NeRF. Object-centric NeRF [11, 12, 18, 29, 34, 45, 46] is a specialized NeRF variant for modeling specific object categories. Its pipeline feeds a NeRF decoder with shape and texture codes, generating occupancy and color values at 3D positions, followed by volumetric rendering to produce a rendered image. The entire process is differentiable, allowing back-propagation of the photometric loss to update shape codes, texture codes, network weights, and object poses. Early object-centric NeRFs [8, 12, 13, 19, 45, 46, 48, 49] required accurate object poses, full visibility, and dense multi-view observations to reconstruct a categorical model, making them inaccessible for real-world applications. Some approaches [12, 23] claim successful object pose recovery through gradient-based" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "SUP-NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 138, + 168, + 473, + 239 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 1: Literature Summary. SUP-NeRF sets a benchmark for minimal requirements, eliminating the need for external 3D detectors, CAD models, or dense NOCS generators. It is dedicated to the reconstruction of object shapes and pose estimation directly from a single image of driving scenes, employing an object-centric framework." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 168, + 473, + 239 + ], + "lines": [ + { + "bbox": [ + 138, + 168, + 473, + 239 + ], + "spans": [ + { + "bbox": [ + 138, + 168, + 473, + 239 + ], + "type": "table", + "html": "
MethodsIndoorOutdoorViewsScopeCADwt. Det3Dwt. NeRFNOCS Gen.
6DoF Pose [21, 26]✓×singleobject✓✓××
3D Detector [16, 42]×✓singleimage×✓××
NeRF [27]✓✓multiscene××✓×
Object-NeRF [18, 29]×✓singleobject××✓×
BootInv [34]×✓singleobject×✓✓✓
SUP-NeRF×✓singleobject×✓✓×
", + "image_path": "c18b108a521e97e24aafac18d54680a054e8939898b3d77e3218974dc435f5ae.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 252, + 482, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 252, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 252, + 482, + 407 + ], + "type": "text", + "content": "pose refinement from monocular images. However these successes are still limited to synthetic dataset, and oversimplified object/camera poses. Recent methods [18,29] investigate more practical pipelines demonstrating the object-NeRF's potential to train and test on driving scenes with limited views and occluded object data. However, both methods require relatively accurate poses from heavy third-party 3D detectors [32,36]. Other recent works [2,5] try freeing NeRF from external camera poses; however, solving camera poses from densely overlapping views is a different problem than recovering an object's pose from a single image. Compared to these methods, SUP-NeRF recovers object pose from a single image without a 3D detector in the pipeline. A more recent work [34] also unifies pose estimation and NeRF in a single network, and enables high-fidelity object generation. However, its training requires additional NOCS dense generator and it is highly sensitive to moderate occlusions compared to our method." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 414, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 482, + 668 + ], + "type": "text", + "content": "Monocular 3D Object Pose Estimation. Advancements in deep learning and the availability of 3D-labeled datasets [3, 6, 10] enable many monocular 3D object detection [16, 24, 25, 33, 36, 42, 47] and 6DoF pose estimation methods [14, 21, 22, 26, 28, 39, 44, 50]. Although estimating 3D information from a single image is ill-posed due to depth-scale ambiguity, leveraging constraints like CAD models or sizes enables accurate object pose recovery via PnP processes [35, 39, 41] or iterative optimization [14, 21]. These indoor 6DoF pose estimation mostly follow an object-centric design which aims to estimate highly accurate object pose within the 2D image patches out of a previous 2D object detector. Recent studies [4, 20] use the NeRF pipeline to reconstruct 3D models or to train model-based keypoint detectors to estimate object poses during inference. However, none of these model-based methods is ideal for outdoor driving scenarios, as pre-reconstruction of models is often impossible in new scenes. Fortunately, inferring the 3D dimensions of objects from an image is feasible when its appearance is tightly bounded to its fine-grained class [28], which keeps an outdoor 3D detector from being fully ill-posed. Nowadays, reliable outdoor 3D detectors usually favors a image-centric design to direct detect many 3D objects captured within the whole image, which usually requires a large backbone network [16, 42]. However, the performance of SoTA detectors drops significantly when applied to datasets with distribution different from the training data [16, 25, 42]." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "Y. Guo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 116, + 477, + 258 + ], + "blocks": [ + { + "bbox": [ + 140, + 116, + 477, + 258 + ], + "lines": [ + { + "bbox": [ + 140, + 116, + 477, + 258 + ], + "spans": [ + { + "bbox": [ + 140, + 116, + 477, + 258 + ], + "type": "image", + "image_path": "193c1224eb7b64e9afba6de803a5eb9e99a264f4049fbaf4df452e2754e0918b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 267, + 482, + 312 + ], + "lines": [ + { + "bbox": [ + 130, + 267, + 482, + 312 + ], + "spans": [ + { + "bbox": [ + 130, + 267, + 482, + 312 + ], + "type": "text", + "content": "Fig. 3: SUP-NeRF Overview. SUP-NeRF unifies pose estimation and NeRF. The pose estimation module enables SUP-NeRF to work for objects in diverse poses without external 3D detectors. The increase of complexity only constitutes a few MLP layers." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 324, + 482, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 324, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 324, + 482, + 407 + ], + "type": "text", + "content": "Tab. 1 outline the differences in methodologies. Unlike existing approaches, SUP-NeRF unifies pose estimation and object NeRF in a streamlined manner, adopting a straightforward object-centric design akin to indoor techniques. Without the need for CAD models, SUP-NeRF is suitable for outdoor driving scenarios. Despite its focused scope on object-centric pose estimation, SUP-NeRF showcases enhanced cross-domain generalization capabilities compared to more complex image-centric 3D detectors within this realm." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 425, + 222, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 425, + 222, + 438 + ], + "spans": [ + { + "bbox": [ + 132, + 425, + 222, + 438 + ], + "type": "text", + "content": "3 SUP-NeRF" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 449, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 593 + ], + "type": "text", + "content": "Our goal is to create a NeRF framework that bypasses NeRF's limitations without external 3D detectors. Fig. 3 shows the overall pipeline of SUP-NeRF, a streamlined unified network that predicts object pose, shape, and texture from a single image. SUP-NeRF consists of three modules: a Resnet-based encoder, a pose estimation module, and a NeRF decoder. SUP-NeRF processes an image and its occlusion mask, with the encoder turning the masked image into shape, texture, and pose codes and predicted object dimensions. The pose code, object dimensions and a box code, are passed to the pose estimation module to iteratively refine the object pose " + }, + { + "bbox": [ + 130, + 449, + 482, + 593 + ], + "type": "inline_equation", + "content": "[R_{o2c}|T_{o2c}]" + }, + { + "bbox": [ + 130, + 449, + 482, + 593 + ], + "type": "text", + "content": ". After a few iterations, the estimated pose is converted to the camera pose " + }, + { + "bbox": [ + 130, + 449, + 482, + 593 + ], + "type": "inline_equation", + "content": "[R_{c2o}|T_{c2o}]" + }, + { + "bbox": [ + 130, + 449, + 482, + 593 + ], + "type": "text", + "content": " and passed to the NeRF decoder. In the NeRF phase, the decoder uses the shape, texture codes, and camera pose for volumetric rendering, creating both RGB and occupancy images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": "Pose Estimation Module. The key component of SUP-NeRF is the pose estimation module, which we show in Fig. 4. It aims to provide reliable object poses for targets at all ranges and orientations. Our pose estimation module iterative updates the input pose, based on the visual difference between the input pose and the observed object in the image. Given predicted object dimensions " + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "inline_equation", + "content": "[H_B, W_B, L_B]" + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": ", camera intrinsic " + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": ", and current pose " + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "inline_equation", + "content": "R^{(t)}, T^{(t)}" + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": ", image projections" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "SUP-NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 186, + 116, + 428, + 235 + ], + "blocks": [ + { + "bbox": [ + 186, + 116, + 428, + 235 + ], + "lines": [ + { + "bbox": [ + 186, + 116, + 428, + 235 + ], + "spans": [ + { + "bbox": [ + 186, + 116, + 428, + 235 + ], + "type": "image", + "image_path": "e0b9fe08d967f5b1f1b3a41d36d651e04c554ab36d004762ad654cf770c78147.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 247, + 482, + 314 + ], + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 314 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 314 + ], + "type": "text", + "content": "Fig.4: SUP-NeRF Pose Estimation Module. The pose estimation module of SUP-NeRF iteratively updates the object's pose while preserving scale. It takes the projection of 3D box corners as a visual representation of the input pose and estimate the pose update via comparing it to observed image in a latent embedding space. These designs handle scale-depth ambiguity and make the deep refiner independent from camera intrinsic parameters for better cross-domain generalization." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 331, + 482, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 482, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 482, + 441 + ], + "type": "text", + "content": "of the 3D box corners " + }, + { + "bbox": [ + 130, + 331, + 482, + 441 + ], + "type": "inline_equation", + "content": "B_{proj}^{(t)}" + }, + { + "bbox": [ + 130, + 331, + 482, + 441 + ], + "type": "text", + "content": " is computed, which can be interpolated as a visual representation of current pose. After encoding the 8 corners into a higher dimensional box code, SUP-NeRF feeds into the pose refiner along with a pose code from the image encoder to predict the desired pose update " + }, + { + "bbox": [ + 130, + 331, + 482, + 441 + ], + "type": "inline_equation", + "content": "\\Delta R^{(t)}, \\Delta T^{(t)}" + }, + { + "bbox": [ + 130, + 331, + 482, + 441 + ], + "type": "text", + "content": ". SUP-NeRF then composes the pose changes with current pose to compute the next pose state, " + }, + { + "bbox": [ + 130, + 331, + 482, + 441 + ], + "type": "inline_equation", + "content": "R^{(t + 1)}, T^{(t + 1)}" + }, + { + "bbox": [ + 130, + 331, + 482, + 441 + ], + "type": "text", + "content": ". This updating process starts from a fully random object pose within the viewing frustum of the image patch, and takes a few iterations such that the final pose approaches the true pose even from a random initial pose." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 146, + 444, + 459, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 444, + 459, + 456 + ], + "spans": [ + { + "bbox": [ + 146, + 444, + 459, + 456 + ], + "type": "text", + "content": "The pose estimation module incorporates the following unique designs:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 129, + 459, + 479, + 664 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 129, + 459, + 479, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 459, + 479, + 540 + ], + "spans": [ + { + "bbox": [ + 129, + 459, + 479, + 540 + ], + "type": "text", + "content": "- SUP-NeRF preserves the object dimensions to ensure unambiguous pose updates. Our implementation, using projected 3D boxes on the image plane, leads to a concise encoding of shape and pose, reducing computation in image rendering and encoding per iteration, as opposed to previous methods [21, 29] relying on full CAD models. Since this representation makes the deep pose refiner's task independent from camera intrinsic parameters, it leads to better generalization to different testing domains, as evidenced in Tabs. 3 and 4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 129, + 543, + 479, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 543, + 479, + 586 + ], + "spans": [ + { + "bbox": [ + 129, + 543, + 479, + 586 + ], + "type": "text", + "content": "- SUP-NeRF updates the object pose iteratively rather than conduct a direct pose estimation using a PnP-based [14] because the later is sensitive to missed or swapped corners thus often produces far out-of-distribution poses. Our ablation studies in Tab. 6 confirm this difference." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 129, + 590, + 479, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 590, + 479, + 638 + ], + "spans": [ + { + "bbox": [ + 129, + 590, + 479, + 638 + ], + "type": "text", + "content": "- SUP-NeRF represents the rotation update in axis-angle representation [37]. During rotation updating process, we convert the axis-angle space 3D rotation " + }, + { + "bbox": [ + 129, + 590, + 479, + 638 + ], + "type": "inline_equation", + "content": "\\Delta R^{(t)}" + }, + { + "bbox": [ + 129, + 590, + 479, + 638 + ], + "type": "text", + "content": " back to " + }, + { + "bbox": [ + 129, + 590, + 479, + 638 + ], + "type": "inline_equation", + "content": "SO(3)" + }, + { + "bbox": [ + 129, + 590, + 479, + 638 + ], + "type": "text", + "content": ", compose it with the previous rotation to compute the new rotation state as " + }, + { + "bbox": [ + 129, + 590, + 479, + 638 + ], + "type": "inline_equation", + "content": "R_{\\mathrm{e2}}^{(t + 1)}" + }, + { + "bbox": [ + 129, + 590, + 479, + 638 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 129, + 639, + 479, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 639, + 479, + 664 + ], + "spans": [ + { + "bbox": [ + 129, + 639, + 479, + 664 + ], + "type": "text", + "content": "- SUP-NeRF represents the translation update in a relative space, indicated as " + }, + { + "bbox": [ + 129, + 639, + 479, + 664 + ], + "type": "inline_equation", + "content": "\\Delta T^{(t)}: [v_x^{(t)}, v_y^{(t)}, \\rho^{(t)}]" + }, + { + "bbox": [ + 129, + 639, + 479, + 664 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 129, + 639, + 479, + 664 + ], + "type": "inline_equation", + "content": "v_x^{(t)}" + }, + { + "bbox": [ + 129, + 639, + 479, + 664 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 129, + 639, + 479, + 664 + ], + "type": "inline_equation", + "content": "v_y^{(t)}" + }, + { + "bbox": [ + 129, + 639, + 479, + 664 + ], + "type": "text", + "content": " indicate the location change in" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "Y. Guo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 139, + 115, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 115, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 139, + 115, + 479, + 140 + ], + "type": "text", + "content": "image space, and " + }, + { + "bbox": [ + 139, + 115, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\rho^{(t)}" + }, + { + "bbox": [ + 139, + 115, + 479, + 140 + ], + "type": "text", + "content": " indicates the ratio between the target depth and the source depth. Specifically, SUP-NeRF updates the translation in image space" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 140, + 140, + 202, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 140, + 202, + 152 + ], + "spans": [ + { + "bbox": [ + 140, + 140, + 202, + 152 + ], + "type": "text", + "content": "and depth as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 271, + 145, + 352, + 159 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 271, + 145, + 352, + 159 + ], + "spans": [ + { + "bbox": [ + 271, + 145, + 352, + 159 + ], + "type": "interline_equation", + "content": "x ^ {(t + 1)} = x ^ {(t)} + v _ {x} ^ {(t)}", + "image_path": "a036f4a2d96d50c2884dc81dfe07123bc9c0b285e7b10064e990b4fe3b96f30e.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 271, + 162, + 480, + 177 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 271, + 162, + 480, + 177 + ], + "spans": [ + { + "bbox": [ + 271, + 162, + 480, + 177 + ], + "type": "interline_equation", + "content": "y ^ {(t + 1)} = y ^ {(t)} + v _ {y} ^ {(t)} \\tag {1}", + "image_path": "3bb8b58eabd37aacdfef546a813248a7aa69bbb0d23a9fd146fc6ae322d9d5ee.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 269, + 179, + 342, + 193 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 269, + 179, + 342, + 193 + ], + "spans": [ + { + "bbox": [ + 269, + 179, + 342, + 193 + ], + "type": "interline_equation", + "content": "Z ^ {(t + 1)} = \\rho^ {(t)} Z ^ {(t)}", + "image_path": "3a341c5675875b7e025b08ef4127a0218e800db783525b7efccfbaa0c7b35e13.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 201, + 479, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 201, + 479, + 251 + ], + "spans": [ + { + "bbox": [ + 130, + 201, + 479, + 251 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 201, + 479, + 251 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 130, + 201, + 479, + 251 + ], + "type": "text", + "content": " represent the object's 2D location in image space, " + }, + { + "bbox": [ + 130, + 201, + 479, + 251 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 130, + 201, + 479, + 251 + ], + "type": "text", + "content": " indicates the depth. The new 3D location " + }, + { + "bbox": [ + 130, + 201, + 479, + 251 + ], + "type": "inline_equation", + "content": "T_{o2c}^{(t + 1)}" + }, + { + "bbox": [ + 130, + 201, + 479, + 251 + ], + "type": "text", + "content": " is then computed from the updated image coordinates, depth and camera intrinsic matrix " + }, + { + "bbox": [ + 130, + 201, + 479, + 251 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 201, + 479, + 251 + ], + "type": "text", + "content": ". Compared to [21, 34], our definition of " + }, + { + "bbox": [ + 130, + 201, + 479, + 251 + ], + "type": "inline_equation", + "content": "\\rho^{(t)}" + }, + { + "bbox": [ + 130, + 201, + 479, + 251 + ], + "type": "text", + "content": " leads to a more straight-forward update." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 251, + 479, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 251, + 479, + 310 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 479, + 310 + ], + "type": "text", + "content": "During the pose updating process indicated by " + }, + { + "bbox": [ + 130, + 251, + 479, + 310 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 130, + 251, + 479, + 310 + ], + "type": "text", + "content": ", both the rotation and translation are updated. Through such representation and updating scheme, SUP-NeRF handles objects at varying distances with the same output space. Further discusses the choice of an effective pose representation and the right coordinate frame are included in Appendices A1 and A2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 312, + 480, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 480, + 407 + ], + "type": "text", + "content": "Cross-task Shortcut. Simultaneously perceiving pose, shape, and texture, requires an image encoder with shared initial layers but separate later stages. However, encoding shape, texture, and pose is a challenge as they require different features, making it tough for a shared encoder to meet all needs effectively. Shape and texture encoding demand a pose-invariant approach, as they correspond to each object's frame. On the other hand, pose encoding needs to capture sufficient pose information for later use. This difference leads to conflicting learning signals during joint training, potentially weakening the model's performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 407, + 480, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 480, + 502 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 480, + 502 + ], + "type": "text", + "content": "SUP-NeRF proposes a straightforward one-step approach to resolve this conflict. After processing through the first four CNN layers separately, SUP-NeRF introduces a cross-task shortcut that deducts pose features from the shape and texture features. This design choice eliminates the shape and texture's reliance on pose, allowing the earlier layers to focus on pose-dependent features for all tasks. This shortcut turns the conflict into a cooperative relationship, enhancing the joint training of different modules, as demonstrated in Tab. 7. The details of our image encoder is described in Appendix A3." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 503, + 479, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 503, + 479, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 503, + 479, + 552 + ], + "type": "text", + "content": "Training. SUP-NeRF closely follows the loss functions and occupancy mask definitions from AutoRF [29]. The total loss during training is: " + }, + { + "bbox": [ + 130, + 503, + 479, + 552 + ], + "type": "inline_equation", + "content": "L_{train} = L_{rgb} + w_{occ}L_{occ} + w_{pose}(L_{pose}^{(direct)} + \\sum_t^K L_{pose}^{(t)})" + }, + { + "bbox": [ + 130, + 503, + 479, + 552 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 503, + 479, + 552 + ], + "type": "inline_equation", + "content": "w_{occ}" + }, + { + "bbox": [ + 130, + 503, + 479, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 503, + 479, + 552 + ], + "type": "inline_equation", + "content": "w_{pose}" + }, + { + "bbox": [ + 130, + 503, + 479, + 552 + ], + "type": "text", + "content": " indicate the corresponding weights." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 553, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 553, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 553, + 480, + 665 + ], + "type": "text", + "content": "We compare the RGB output to the input image to calculate photometric loss " + }, + { + "bbox": [ + 130, + 553, + 480, + 665 + ], + "type": "inline_equation", + "content": "L_{rgb}" + }, + { + "bbox": [ + 130, + 553, + 480, + 665 + ], + "type": "text", + "content": ", and the occupancy values to the mask for occupancy loss " + }, + { + "bbox": [ + 130, + 553, + 480, + 665 + ], + "type": "inline_equation", + "content": "L_{occ}" + }, + { + "bbox": [ + 130, + 553, + 480, + 665 + ], + "type": "text", + "content": ". SUP-NeRF additionally compares the projected box corners from the integrative pose estimator " + }, + { + "bbox": [ + 130, + 553, + 480, + 665 + ], + "type": "inline_equation", + "content": "B^{(t)}_{proj}" + }, + { + "bbox": [ + 130, + 553, + 480, + 665 + ], + "type": "text", + "content": " with the ground-truth values to compute the pose losses " + }, + { + "bbox": [ + 130, + 553, + 480, + 665 + ], + "type": "inline_equation", + "content": "L_{pose}^{(t)}" + }, + { + "bbox": [ + 130, + 553, + 480, + 665 + ], + "type": "text", + "content": ". We define the pose loss function as the mean distance between the matched corners: " + }, + { + "bbox": [ + 130, + 553, + 480, + 665 + ], + "type": "inline_equation", + "content": "L_{pose}^{(*)} = \\frac{1}{8}\\sum_{i=1}^{8}\\sqrt{(x_i^{(*)} - \\hat{x}_i)^2 + (y_i^{(*)} - \\hat{y}_i)^2}" + }, + { + "bbox": [ + 130, + 553, + 480, + 665 + ], + "type": "text", + "content": ", where (*) indicates it compatible with both the direct pose loss and the iterative pose loss, " + }, + { + "bbox": [ + 130, + 553, + 480, + 665 + ], + "type": "inline_equation", + "content": "(x_i^{(*)}, y_i^{(*)})" + }, + { + "bbox": [ + 130, + 553, + 480, + 665 + ], + "type": "text", + "content": " denote a input corner position, and " + }, + { + "bbox": [ + 130, + 553, + 480, + 665 + ], + "type": "inline_equation", + "content": "(\\hat{x}_i, \\hat{y}_i)" + }, + { + "bbox": [ + 130, + 553, + 480, + 665 + ], + "type": "text", + "content": " indicates a ground-truth corner po" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 90, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 90, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 90, + 447, + 100 + ], + "type": "text", + "content": "SUP-NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 152 + ], + "type": "text", + "content": "sition. SUP-NeRF introduces an additional MLP layer for more accurate pose estimation to directly supervise the 2D coordinates of box corners predicted from the pose code." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 481, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 481, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 481, + 319 + ], + "type": "text", + "content": "In training, only the network weights are treated as the estimizable variables. Moreover, the NeRF decoder uses ground-truth camera poses " + }, + { + "bbox": [ + 130, + 152, + 481, + 319 + ], + "type": "inline_equation", + "content": "[R_{c2o}^{gt}|T_{c2o}^{gt}]" + }, + { + "bbox": [ + 130, + 152, + 481, + 319 + ], + "type": "text", + "content": " instead of predictions from the pose estimation module to ensure more reliable training. Implementation Details. SUP-NeRF's image encoder backbone is based on ResNet50, with shared early layers and separated later layers for pose, shape, and texture encoding. For more information on this, please refer to the Appendix A3. To implement the NeRF decoder, we use CodeNeRF [12] for its simplicity and ease of use. For the differentiable rendering based on NeRF, we set the latent dimension to 256, use 64 samples along each ray, and normalize the shape space by the diagonal length of the observed object's 3D box. For the pose estimation module, we also set the latent dimension to 256 for all MLP layers. Before feeding the input image and occupancy mask into the image encoder, we pad them to make them square and resize them to " + }, + { + "bbox": [ + 130, + 152, + 481, + 319 + ], + "type": "inline_equation", + "content": "128\\times 128" + }, + { + "bbox": [ + 130, + 152, + 481, + 319 + ], + "type": "text", + "content": " pixels. We then render " + }, + { + "bbox": [ + 130, + 152, + 481, + 319 + ], + "type": "inline_equation", + "content": "32\\times 32" + }, + { + "bbox": [ + 130, + 152, + 481, + 319 + ], + "type": "text", + "content": " images and occupancy maps to calculate losses." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 320, + 481, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 320, + 481, + 427 + ], + "spans": [ + { + "bbox": [ + 130, + 320, + 481, + 427 + ], + "type": "text", + "content": "To ensure effective iterative pose estimation without sacrificing generalization, SUP-NeRF uses a heuristic to sample a random initial 3D object pose across the entire pose space. For translation, we sample in the relative space. We sample a random 2D location within the image ROI range as the initial projection of the object center, while setting the initial depth to " + }, + { + "bbox": [ + 130, + 320, + 481, + 427 + ], + "type": "inline_equation", + "content": "Z = 20" + }, + { + "bbox": [ + 130, + 320, + 481, + 427 + ], + "type": "text", + "content": ". For rotation, we sample from the full range of yaw angles combined with a random rotation angle within a " + }, + { + "bbox": [ + 130, + 320, + 481, + 427 + ], + "type": "inline_equation", + "content": "(-20^{\\circ}, 20^{\\circ})" + }, + { + "bbox": [ + 130, + 320, + 481, + 427 + ], + "type": "text", + "content": " range around a random axis direction. This sampling method covers almost the entire distribution of object poses in the camera frame and is used consistently in both training and inference." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 428, + 481, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 428, + 481, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 428, + 481, + 498 + ], + "type": "text", + "content": "During training, we use 3 iterations of the pose estimator to generate corresponding pose losses, which allow for pose regression to handle highly diverse poses. We jointly train the unified model for the pose estimation module, image encoder, and NeRF decoder for 40 epochs on the nuScenes dataset using a learning rate of " + }, + { + "bbox": [ + 130, + 428, + 481, + 498 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 130, + 428, + 481, + 498 + ], + "type": "text", + "content": " and loss weights of " + }, + { + "bbox": [ + 130, + 428, + 481, + 498 + ], + "type": "inline_equation", + "content": "w_{occ} = 0.1" + }, + { + "bbox": [ + 130, + 428, + 481, + 498 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 428, + 481, + 498 + ], + "type": "inline_equation", + "content": "w_{pos} = 0.01" + }, + { + "bbox": [ + 130, + 428, + 481, + 498 + ], + "type": "text", + "content": " to compute the total loss." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 499, + 481, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 499, + 481, + 596 + ], + "spans": [ + { + "bbox": [ + 130, + 499, + 481, + 596 + ], + "type": "text", + "content": "Inference. During inference, SUP-NeRF optimizes the shape code, texture code, and object pose " + }, + { + "bbox": [ + 130, + 499, + 481, + 596 + ], + "type": "inline_equation", + "content": "[R_{o2c}|T_{o2c}]" + }, + { + "bbox": [ + 130, + 499, + 481, + 596 + ], + "type": "text", + "content": " keeping the network weights frozen. The total loss " + }, + { + "bbox": [ + 130, + 499, + 481, + 596 + ], + "type": "inline_equation", + "content": "L_{infer}" + }, + { + "bbox": [ + 130, + 499, + 481, + 596 + ], + "type": "text", + "content": " is thus " + }, + { + "bbox": [ + 130, + 499, + 481, + 596 + ], + "type": "inline_equation", + "content": "L_{infer} = L_{rgb} + w_{occ}L_{occ}" + }, + { + "bbox": [ + 130, + 499, + 481, + 596 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 499, + 481, + 596 + ], + "type": "inline_equation", + "content": "w_{occ}" + }, + { + "bbox": [ + 130, + 499, + 481, + 596 + ], + "type": "text", + "content": " balances the two. SUP-NeRF minimizes the loss " + }, + { + "bbox": [ + 130, + 499, + 481, + 596 + ], + "type": "inline_equation", + "content": "L_{infer}" + }, + { + "bbox": [ + 130, + 499, + 481, + 596 + ], + "type": "text", + "content": " to update the variables. At inference, we run 3 iterations for the feed-forward pose estimation module without updating the shape and color codes, and 50 iterations for the gradient-based joint optimization of shape, texture, and pose. We use a step size of 0.02 for shape and texture code updates and 0.01 for pose updates during joint optimization." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 615, + 230, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 615, + 230, + 628 + ], + "spans": [ + { + "bbox": [ + 131, + 615, + 230, + 628 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 628, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 628, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 628, + 479, + 665 + ], + "type": "text", + "content": "Our experiments use three datasets: nuScenes [3], KITTI [6] and Waymo [38]. Our evaluation framework simultaneously measures rendering quality, shape reconstruction accuracy, and pose estimation precision." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "Y. Guo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": "Data Splits. Since these datasets are not specifically for 3D reconstruction, we sifted a subset for better evaluation. We chose daytime sequences and used Mask R-CNN [7] to get instance masks, as nuScenes lacks 2D segmentation masks. We matched these masks with 3D bounding box annotations, categorizing them into foreground, background, and unknown, similar to AutoRF [29]. More details are in Appendix A7. We use the following splits of these datasets:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 128, + 189, + 479, + 270 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 128, + 189, + 479, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 189, + 479, + 209 + ], + "spans": [ + { + "bbox": [ + 128, + 189, + 479, + 209 + ], + "type": "text", + "content": "- nuScenes Train Split: We collected 87,048 objects from the nuScenes Train split for training." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 128, + 213, + 479, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 213, + 479, + 233 + ], + "spans": [ + { + "bbox": [ + 128, + 213, + 479, + 233 + ], + "type": "text", + "content": "- nuScenes Val Split: We collected 5,000 random objects from the nuScenes Val split for evaluation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 128, + 236, + 479, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 236, + 479, + 247 + ], + "spans": [ + { + "bbox": [ + 128, + 236, + 479, + 247 + ], + "type": "text", + "content": "- KITTI Val Split: We use 4,895 objects from KITTI Val split [6] for evaluation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 128, + 248, + 479, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 248, + 479, + 270 + ], + "spans": [ + { + "bbox": [ + 128, + 248, + 479, + 270 + ], + "type": "text", + "content": "- Waymo Val Split: We randomly selected 5,000 objects from Waymo Val split [16] for evaluation." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 273, + 482, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 273, + 482, + 440 + ], + "spans": [ + { + "bbox": [ + 130, + 273, + 482, + 440 + ], + "type": "text", + "content": "Evaluation metrics. Our evaluation framework differs from traditional NeRF assessments as it simultaneously evaluates object reconstruction quality and pose accuracy. We incorporate two monocular reconstruction metrics: (1) the PSNR (Peak Signal-to-Noise Ratio) comparing the rendered and observed images; (2) the Depth Error between the rendered depth map and sparse LiDAR data, denoted as " + }, + { + "bbox": [ + 130, + 273, + 482, + 440 + ], + "type": "inline_equation", + "content": "DE" + }, + { + "bbox": [ + 130, + 273, + 482, + 440 + ], + "type": "text", + "content": ". We also include two pose estimation metrics: (3) rotation error denoted as " + }, + { + "bbox": [ + 130, + 273, + 482, + 440 + ], + "type": "inline_equation", + "content": "RE" + }, + { + "bbox": [ + 130, + 273, + 482, + 440 + ], + "type": "text", + "content": "; (4) the translation error denoted as " + }, + { + "bbox": [ + 130, + 273, + 482, + 440 + ], + "type": "inline_equation", + "content": "TE" + }, + { + "bbox": [ + 130, + 273, + 482, + 440 + ], + "type": "text", + "content": ". To offer a thorough evaluation, we compare various methods at different stages: post-feed-forward (" + }, + { + "bbox": [ + 130, + 273, + 482, + 440 + ], + "type": "inline_equation", + "content": "FF" + }, + { + "bbox": [ + 130, + 273, + 482, + 440 + ], + "type": "text", + "content": "), and at the 50th iterations of NeRF joint optimization. Beyond monocular assessments, we also conduct cross-view evaluations for PSNR and depth error metrics, denoted as (5) PSNR-" + }, + { + "bbox": [ + 130, + 273, + 482, + 440 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 273, + 482, + 440 + ], + "type": "text", + "content": " and (6) " + }, + { + "bbox": [ + 130, + 273, + 482, + 440 + ], + "type": "inline_equation", + "content": "DE-C" + }, + { + "bbox": [ + 130, + 273, + 482, + 440 + ], + "type": "text", + "content": ", focusing on objects observed from novel views. As the nuScenes dataset provides object IDs that allows tracking objects across various views, the cross-view evaluation is exclusively conducted on the nuScenes dataset, as explained in Fig. 5." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 441, + 482, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 441, + 482, + 561 + ], + "spans": [ + { + "bbox": [ + 130, + 441, + 482, + 561 + ], + "type": "text", + "content": "Baselines. We compare SUP-NeRF with two baselines: AutoRF [29] and AutoRF [29] + FCOS3D [42] (selected for its monocular capability, and availability of model trained on the nuScenes dataset). Both AutoRF and SUP-NeRF use the same NeRF decoder and image encoder to ensure a fair comparison. The difference lies in SUP-NeRF's additional layers for pose encoding and refinement. During testing, we apply the same random pose generation for both AutoRF and SUP-NeRF methods, whereas the FCOS3D baseline begins with its own pose prediction. We also compare with BootInv [34], which recovers pose, shape, and appearance. Leveraging our flexible design, we developed SUP-BootInv by integrating BootInv's pretrained decoder with our encoder and pose estimator." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 582, + 451, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 582, + 451, + 594 + ], + "spans": [ + { + "bbox": [ + 131, + 582, + 451, + 594 + ], + "type": "text", + "content": "4.1 nuScenes Monocular Reconstruction and Pose Estimation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "Tab. 2 shows that SUP-NeRF outperforms two baselines, AutoRF and its combination with FCOS3D on nuScenes Val set. AutoRF alone struggles to improve pose from a random start, resulting in lower PSNR and depth error. SUP-NeRF consistently reduces depth error by over " + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": ", rotation error by over " + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": ", and translation error by over " + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "13\\%" + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": " compared to the FCOS3D and AutoRF com" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "SUP-NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 142, + 168, + 472, + 239 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 2: nuScenes Monocular Reconstruction and Pose Estimation Results. SUP-NeRF consistently improves AutoRF-based pipelines in all metrics, particularly in the pose estimation metrics. SUP-BootInv also consistently improves BootInv in all metrics. [Key: Best, Second Best, FF= Feed Forward, C= Cross-View]" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 142, + 168, + 472, + 239 + ], + "lines": [ + { + "bbox": [ + 142, + 168, + 472, + 239 + ], + "spans": [ + { + "bbox": [ + 142, + 168, + 472, + 239 + ], + "type": "table", + "html": "
MethodPSNR (↑)DE (m) (↓)RE (deg.) (↓)TE (m) (↓)PSNR-C (↑)DE-C (m) (↓)
FF|50itFF|50itFF|50itFF|50itFF|50itFF|50it
AutoRF [29]3.6|10.611.21|10.0987.52|88.076.04|5.9510.0|8.81.31|1.41
AutoRF+FCOS7.5|17.21.34|0.819.77|10.170.85|0.789.8|10.51.29|1.16
SUP-NeRF10.5|18.80.69|0.607.01|7.070.68|0.7310.6|10.91.22|1.14
BootInv [34]9.4|14.35.01|3.5628.40|28.002.59|2.9110.9|11.81.37|1.35
SUP-BootInv10.9|15.41.95|1.627.11|8.400.64|1.0010.9|11.91.37|1.40
", + "image_path": "0745e9c0ac0d90a24e2cc15080bcb06af2a31af3a07cf0df732a9ed3caf01138.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 143, + 251, + 471, + 342 + ], + "blocks": [ + { + "bbox": [ + 143, + 251, + 471, + 342 + ], + "lines": [ + { + "bbox": [ + 143, + 251, + 471, + 342 + ], + "spans": [ + { + "bbox": [ + 143, + 251, + 471, + 342 + ], + "type": "image", + "image_path": "dd917c13bc6afe2dfae83e07b615f4518207d87f7654f1d560ccb26fb6b4aabf.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 351, + 482, + 396 + ], + "lines": [ + { + "bbox": [ + 131, + 351, + 482, + 396 + ], + "spans": [ + { + "bbox": [ + 131, + 351, + 482, + 396 + ], + "type": "text", + "content": "Fig. 5: nuScenes Cross-view Evaluation. Each row shows a set of images of the same object from different angles. We use the example for both monocular pose estimation and NeRF reconstruction. Other images in each row are used to evaluate the reconstructed shape and texture in PSNR and Depth Error (DE)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 420, + 480, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 420, + 480, + 444 + ], + "spans": [ + { + "bbox": [ + 130, + 420, + 480, + 444 + ], + "type": "text", + "content": "bination. Our feed-forward stage also shows promising potential for real-time applications, especially where rendering quality is less critical." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 445, + 482, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 445, + 482, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 482, + 552 + ], + "type": "text", + "content": "We next compare BootInv [34] as the baseline, although the performance of original BootInv's pose estimation was subpar on both the nuScenes and KITTI datasets, integrating our robust pose module with this NeRF framework significantly enhanced pose estimation accuracy. This underscores the adaptability of our unified pipeline in augmenting various object NeRF frameworks for improved pose estimation. Moreover, comparisons of BootInv [34] and SUP-BootInv in terms of monocular PSNR and depth error suggest that SUP-BootInv's superior pose estimation contributes to its higher PSNR and lower depth error. More detailed analysis upon BootInv-based pipelines is included in Appendix A5." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 571, + 289, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 571, + 289, + 582 + ], + "spans": [ + { + "bbox": [ + 131, + 571, + 289, + 582 + ], + "type": "text", + "content": "4.2 Cross-Dataset Evaluation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 480, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 480, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 480, + 628 + ], + "type": "text", + "content": "We next compare the generalization capability of all these NeRF models with the cross-dataset experiments. For these experiments, we directly apply nuScenes models to the KITTI and Waymo datasets and report the metrics." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "KITTI. Tab. 3 shows the result of cross-dataset experiments on the KITTI dataset. It shows that SUP-NeRF achieves SoTA performance across all evaluation metrics. SUP-NeRF reduces depth error by about " + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "inline_equation", + "content": "16\\%" + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": ", rotation error by" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "Y. Guo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 182, + 168, + 432, + 239 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 3: KITTI Cross-dataset Monocular Reconstruction and Pose Estimation Results. We train all methods on nuScenes dataset, and test on KITTI dataset. Our methods SUP-NeRF and SUP-BootInv consistently show superiority in all metrics compared to the counterpart methods. [Key: Best, Second Best, FF= Feed Forward]" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 182, + 168, + 432, + 239 + ], + "lines": [ + { + "bbox": [ + 182, + 168, + 432, + 239 + ], + "spans": [ + { + "bbox": [ + 182, + 168, + 432, + 239 + ], + "type": "table", + "html": "
MethodPSNR (↑)DE (m) (↓)RE (deg.) (↓)TE (m) (↓)
FF|50itFF|50itFF|50itFF|50it
AutoRF [29]0.4|7.69.89|8.8389.67|90.426.16|6.07
AutoRF+FCOS2.4|13.42.42|1.7411.95|12.52.2|2.09
SUP-NeRF4.0|14.12.19|1.546.79|6.891.06|1.01
BootInv [34]7.0|12.36.11|4.3815.52|15.763.97|3.90
SUP-BootInv7.6|12.72.79|2.219.16|8.841.15|1.67
", + "image_path": "2df3669c1d60c1182fcf2fad7a63fa1b16a5ec02bc91a3c2fc689ef90a4091a4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 142, + 256, + 473, + 425 + ], + "blocks": [ + { + "bbox": [ + 142, + 256, + 473, + 425 + ], + "lines": [ + { + "bbox": [ + 142, + 256, + 473, + 425 + ], + "spans": [ + { + "bbox": [ + 142, + 256, + 473, + 425 + ], + "type": "image", + "image_path": "bbad4520302c52591efcc56bda809535229a58ab55cfac0ebe5b56cbf23fc5a8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 434, + 482, + 501 + ], + "lines": [ + { + "bbox": [ + 130, + 434, + 482, + 501 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 482, + 501 + ], + "type": "text", + "content": "Fig. 6: KITTI (Cross-dataset) Qualitative Results. In the top panel, we demonstrate SUP-NeRF executes pose estimation reliably, fast converging from a random initial pose to the true one, and enables neural reconstruction under diverse object poses, occlusion cases under this cross-dataset setup. In the bottom panel, SUP-NeRF is visually compared to the other major competitor, demonstrates sharper rendered image, higher accuracy in shape and pose." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 530, + 482, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 530, + 482, + 602 + ], + "spans": [ + { + "bbox": [ + 130, + 530, + 482, + 602 + ], + "type": "text", + "content": "as large as " + }, + { + "bbox": [ + 130, + 530, + 482, + 602 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 530, + 482, + 602 + ], + "type": "text", + "content": " and translation error also about " + }, + { + "bbox": [ + 130, + 530, + 482, + 602 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 530, + 482, + 602 + ], + "type": "text", + "content": ". Thus, SUP-NeRF shows significantly better geometric performance in cross-dataset setting compared to AutoRF+FCOS, which we attributes to our deep pose module's independence from camera intrinsic parameters. Meanwhile, compared to nuScenes evaluation, all the methods' PSNR and depth error performance drop, which means the cross-dataset generalization of monocular NeRF remains an open challenge." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "Fig. 6 presents the qualitative results on this experiment, demonstrating how SUP-NeRF handles diverse poses and occlusions in pose refinement and the visual comparisons between the two most competitive pipelines. Consistent improvements upon integrating our design into BootInv baseline were also observed. More visual comparisons are in Appendix A9." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "SUP-NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 182, + 179, + 432, + 250 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 4: Waymo Cross-dataset Monocular Reconstruction and Pose Estimation Results. We train all methods on nuScenes dataset, and test on Waymo dataset. Our methods SUP-NeRF and SUP-BootInv consistently show superiority in all metrics compared to the counterpart methods, the only exception is the comparison to AutoRF+FCOS in Rotation Error. [Key: Best, Second Best, FF= Feed Forward]" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 182, + 179, + 432, + 250 + ], + "lines": [ + { + "bbox": [ + 182, + 179, + 432, + 250 + ], + "spans": [ + { + "bbox": [ + 182, + 179, + 432, + 250 + ], + "type": "table", + "html": "
MethodPSNR (↑)DE (m) (↓)RE (deg.) (↓)TE (m) (↓)
FF|50itFF|50itFF|50itFF|50it
AutoRF [29]0.6|9.86.76|6.5386.56|87.679.1|9.14
AutoRF+FCOS6.2|16.52.43|2.326.65|7.23.22|3.31
SUP-NeRF4.8|17.02.32|1.5610.01|10.61.68|1.54
BootInv [34]8.1|11.08.85|8.2030.78|31.525.26|6.16
SUP-BootInv8.6|11.95.36|4.3810.24|11.041.67|2.53
", + "image_path": "fe8a772b3b84190b5db7d74bd592b4efed8cc770ff1000a8728e4ce335aa3728.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 271, + 482, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 271, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 271, + 482, + 331 + ], + "type": "text", + "content": "Waymo. Tab. 4 shows the result of cross-dataset experiments on the Waymo dataset. It shows that SUP-NeRF achieves SoTA performance across most evaluation metrics. The only exception is the superior rotation estimation achieved by FCOS3D, which is likely due to leveraging larger context for heavily occluded cases. However, our advantage in translation estimation is more significant." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 347, + 246, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 347, + 246, + 358 + ], + "spans": [ + { + "bbox": [ + 131, + 347, + 246, + 358 + ], + "type": "text", + "content": "4.3 Ablation Studies" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 366, + 482, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 366, + 482, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 366, + 482, + 426 + ], + "type": "text", + "content": "The major ablation studies we conduct includes (i) the impact of pose estimation module designs and (ii) the impact of different object NeRF networks. Other ablation studies to analyze the impact of training with predicted 2D boxes, pose refinement iterations, training epochs, as well as NeRF's capability on handling initial pose errors are included in Appendix A6, owing to lack of space." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 426, + 482, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 557 + ], + "type": "text", + "content": "Pose Estimation Module. We evaluate multiple pose estimation designs from two main angles. First, we focus on object-centric pose estimation task alone with evaluations on nuScenes, KITTI, and Waymo datasets. The prototypical designs include: (i) Direct 6D pose regression using MLP layers; (ii) Pose estimation from predicted box corners via Keypoint-based Perspective-n-Point (PnP); (iii) NOCS-based pose estimation as implemented in BootInv [34]; (vi) External monocular 3D detection FCOS3D [42]; (v) Our iterative pose refinement method, which uses projected boxes. For fairness, methods (1), (2), and (5) share the same image encoder, pose encoding layers, differing only in their output layers. Tab. 5 shows that SUP-NeRF consistently outperforms the others, notably in cross-dataset translation estimation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "The second evaluation assesses the impact of different initial poses on object-NeRF performance within the SUP-NeRF architecture. We additionally compared (i) Ground-Truth (GT) pose; (ii) Random initial pose, to represent the oracle and worst case in pose initialization, and exclude the NOCS method which is part of a different NeRF framework. Given different initial poses, all methods undergo the same NeRF Gradient-based Pose Refinement (NGPR), evaluated immediately and after 50 iterations. Tab. 6 shows that our method significantly surpasses other initial pose estimators in geometric accuracy, demonstrating its effectiveness in enhancing 3D model quality." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "Y. Guo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 176, + 168, + 437, + 238 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 159 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 159 + ], + "type": "text", + "content": "Table 5: Pose Estimation Ablation Studies. We evaluate all the candidate pose estimation methods with a focus on pose accuracy alone, on nuScenes, KITTI, and Waymo validation sets. Our method shows outstanding robustness in all cases, particularly in translation accuracy in cross-dataset tests. [Key: Best]" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 176, + 168, + 437, + 238 + ], + "lines": [ + { + "bbox": [ + 176, + 168, + 437, + 238 + ], + "spans": [ + { + "bbox": [ + 176, + 168, + 437, + 238 + ], + "type": "table", + "html": "
Pose ModuleRE (deg.)(↓)TE (m) (↓)
nuScenesKITTIWaymonuScenesKITTIWaymo
MLP Pose31.9637.5830.895.535.3110.1
Corners+PnP24.7935.1093.762.823.5518.68
NOCS +PnP(BootInv)28.4015.5223.662.593.973.47
FCOS3D9.7711.956.650.852.23.22
SUP-NeRF7.016.7910.010.681.061.54
", + "image_path": "583b85efe7887294a7633870bb3516d84731440640f8f1b5b116047c3c5c0316.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 149, + 305, + 463, + 385 + ], + "blocks": [ + { + "bbox": [ + 132, + 251, + 480, + 296 + ], + "lines": [ + { + "bbox": [ + 132, + 251, + 480, + 296 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 480, + 296 + ], + "type": "text", + "content": "Table 6: Ablation studies of Pose Initialization in Object NeRF on nuScenes. We compare different choices of initial poses under SUP-NeRF frameworks. As shown, SUP-NeRF presents superiors effectiveness in all the metrics. [Key: Best, FF=Feed Forward, C=Cross-View]" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 149, + 305, + 463, + 385 + ], + "lines": [ + { + "bbox": [ + 149, + 305, + 463, + 385 + ], + "spans": [ + { + "bbox": [ + 149, + 305, + 463, + 385 + ], + "type": "table", + "html": "
Initial PosePSNR (↑)DE (m) (↓)RE (deg.) (↓)TE (m) (↓)PSNR-C (↑)DE-C (m) (↓)
FF|50itFF|50itFF|50itFF|50itFF|50itFF|50it
Random3.6|10.611.21|10.0987.52|88.076.04|5.959.1|7.41.42|1.82
MLP Pose1.2|6.95.33|4.4131.96|24.575.53|2.789.8|8.51.3|1.4
Corners+PnP7.2|16.22.18|1.8324.79|24.572.82|2.789.9|9.51.3|1.3
FCOS3D7.5|17.21.34|0.819.77|10.170.85|0.789.8||10.51.29||1.16
SUP-NeRF10.5|18.80.69|0.607.01|7.070.68|0.7310.6|10.91.22|1.14
GT10.0|18.80.66|0.530.2|1.110.0|1.310.9|11.11.19|1.14
", + "image_path": "cbbec98dd334000d859685ce85bd8ea8a676facf038dcbcce8258332916e0bed.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 411, + 480, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 411, + 480, + 506 + ], + "spans": [ + { + "bbox": [ + 132, + 411, + 480, + 506 + ], + "type": "text", + "content": "Tab. 6 also confirms the role of NeRF Gradient-Based Pose Refinement (NGPR). By comparing the difference between feed-forward results and the later ones, one can find NGPR merely reduces rotation error or translation error. On the contrary, it can even increase pose error when the initial pose error is small enough, e.g. rotation error " + }, + { + "bbox": [ + 132, + 411, + 480, + 506 + ], + "type": "inline_equation", + "content": "< 7^{\\circ}" + }, + { + "bbox": [ + 132, + 411, + 480, + 506 + ], + "type": "text", + "content": ", and translation error " + }, + { + "bbox": [ + 132, + 411, + 480, + 506 + ], + "type": "inline_equation", + "content": "< 0.75m" + }, + { + "bbox": [ + 132, + 411, + 480, + 506 + ], + "type": "text", + "content": ". On the other hand, the improvements of PSNR and depth reconstruction over time indicate that NeRF optimization mostly focuses on shape and texture optimization w.r.t. the visible surface regardless the pose." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 508, + 481, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 508, + 481, + 639 + ], + "spans": [ + { + "bbox": [ + 132, + 508, + 481, + 639 + ], + "type": "text", + "content": "Impact of object NeRF framework. To demonstrate SUP-NeRF's benefits, we compared different network configurations using the same pose refinement approach. For fairness, we used identical subnetwork modules across all configurations. These include a ResNet-based Image Encoder (E), a NeRF Decoder (D), and our Pose Estimation Module (P). CodeNeRF [12] has just a Decoder (D), AutoRF combines an Image Encoder and Decoder " + }, + { + "bbox": [ + 132, + 508, + 481, + 639 + ], + "type": "inline_equation", + "content": "(\\mathrm{E} + \\mathrm{D})" + }, + { + "bbox": [ + 132, + 508, + 481, + 639 + ], + "type": "text", + "content": ", and SUP-NeRF uses all three " + }, + { + "bbox": [ + 132, + 508, + 481, + 639 + ], + "type": "inline_equation", + "content": "(\\mathrm{E} + \\mathrm{D} + \\mathrm{P})" + }, + { + "bbox": [ + 132, + 508, + 481, + 639 + ], + "type": "text", + "content": ". For CodeNeRF and AutoRF, we added a separately trained Pose Network (PNet, composed of " + }, + { + "bbox": [ + 132, + 508, + 481, + 639 + ], + "type": "inline_equation", + "content": "\\mathrm{E} + \\mathrm{P}" + }, + { + "bbox": [ + 132, + 508, + 481, + 639 + ], + "type": "text", + "content": ") for initial pose refinement. We also tested a SUP-NeRF version without the cross-task Short Cut (SC) to assess its impact. Tab. 7 shows that SC consistently improves SUP-NeRF across all metrics on nuScenes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 480, + 665 + ], + "type": "text", + "content": "SUP-NeRF, with joint training, outperforms separate training setups AutoRF + PNet and CodeNeRF + PNet in monocular metrics. Interestingly, while" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "SUP-NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 179, + 478, + 237 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 7: Object NeRF Framework Ablations on nuScenes. We used the same subnetwork architectures - the image encoder (E), NeRF decoder (D), and pose estimation module (P) - to create various model configurations. For SUP-NeRF, we also included a version without the cross-task Short Cut (SC) for comparison. [Key: Best, FF=Feed Forward, C=Cross-View]" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 179, + 478, + 237 + ], + "lines": [ + { + "bbox": [ + 136, + 179, + 478, + 237 + ], + "spans": [ + { + "bbox": [ + 136, + 179, + 478, + 237 + ], + "type": "table", + "html": "
Model TrainingPSNR(▲)DE (m)(↓)RE (deg.) (↓)TE (m)(↓)PSNR-C(▲)DE-C (m)(↓)
FF|50itFF|50itFF|50itFF|50itFF|50itFF|50it
CodeNeRF (D)+PNet(E+P)7.6|16.00.87|0.6910.57|9.750.72|0.7611.0|11.21.01|1.04
AutoRF (E+D)+PNet(E+P)9.7|17.50.79|0.668.26|7.870.7|0.745.6|10.11.24|1.16
SUP-NeRF (E+D+P) w/o SC8.7|18.00.94|0.7310.12|9.220.75|0.7410.8|10.91.23|1.11
SUP-NeRF (E+D+P)10.5|18.80.69|0.607.01|7.070.68|0.7310.6|10.91.22|1.14
", + "image_path": "beaf91a1862be14a7f350994c6c4720ff879e2ea67a7ea6b5c940e571212bc7c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 159, + 289, + 455, + 352 + ], + "blocks": [ + { + "bbox": [ + 130, + 247, + 482, + 281 + ], + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 281 + ], + "type": "text", + "content": "Table 8: Model Size and Running Time Comparison in feed-forward scenarios, and with 20 and 50 iterations of NeRF optimization. SUP-NeRF gets the smallest running time. [Key: Best, FF= Feed Forward]" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 159, + 289, + 455, + 352 + ], + "lines": [ + { + "bbox": [ + 159, + 289, + 455, + 352 + ], + "spans": [ + { + "bbox": [ + 159, + 289, + 455, + 352 + ], + "type": "table", + "html": "
MethodParams (M) (↓)FF (s) (↓)FF+20it (s) (↓)FF+50it (s) (↓)
AutoRF+FCOS [29]91.1160.1230.7141.599
(36.166+54.950)(0.114+0.009)(0.114+0.600)(0.114+1.485)
SUP-NeRF49.8160.0180.6081.493
BootInv [34]182.6160.1563.5348.601
SUP-BootInv57.5800.0183.3968.463
", + "image_path": "5bddc11ab308ee33e12767ee11e721a1b0c584b524d668d143619266e794e71a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 372, + 482, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 372, + 482, + 409 + ], + "spans": [ + { + "bbox": [ + 130, + 372, + 482, + 409 + ], + "type": "text", + "content": "CodeNeRF lags in monocular, it excels in cross-view evaluations. This implies that using mean shape codes, rather than current image encodings, might reduce overfitting to current view." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 425, + 285, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 425, + 285, + 438 + ], + "spans": [ + { + "bbox": [ + 131, + 425, + 285, + 438 + ], + "type": "text", + "content": "4.4 Running Speed Analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 444, + 482, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 444, + 482, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 444, + 482, + 540 + ], + "type": "text", + "content": "Tab. 8 compares the model size and running speed on a single A5000 Graphics card. Detailed setup and analysis details are in Appendix A8. Tab. 8 shows that SUP-NeRF needs about half or one third as many parameters as others and is 6-8 times faster in the feed-forward (FF) stage. Meanwhile, a significant portion of the run time is consumed by the NeRF optimization step, whether for 20 iterations (20it) or 50 iterations (50it). However, speed limitation of NeRF can be effectively tackled by recent advances in neural rendering [15,31], which we consider as orthogonal to our contribution to the feed-forward stage." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 557, + 220, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 557, + 220, + 571 + ], + "spans": [ + { + "bbox": [ + 131, + 557, + 220, + 571 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 581, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 667 + ], + "type": "text", + "content": "This paper introduces SUP-NeRF: a unified network that seamlessly integrates pose estimation and object reconstruction. SUP-NeRF includes a novel pose estimation module to handle scale-depth ambiguity and a new representation invariant to camera changes. Consequently, SUP-NeRF greatly improves the robustness of object pose estimation compared to the standard NeRF baseline. SUP-NeRF outperforms previous methods in both monocular reconstruction and pose estimation tasks, especially in challenging cross-dataset applications." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "Y. Guo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "type": "text", + "content": "1. Batra, D., Chang, A.X., Chernova, S., Davison, A.J., Deng, J., Koltun, V., Levine, S., Malik, J., Mordatch, I., Mottaghi, R., Savva, M., Su, H.: Rearrangement: A challenge for embodied AI. CoRR (2020) 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 173, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 173, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 138, + 173, + 480, + 194 + ], + "type": "text", + "content": "2. Bian, W., Wang, Z., Li, K., Bian, J.: Nope-nerf: Optimising neural radiance field with no pose prior. In: CVPR (2023) 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 195, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 195, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 138, + 195, + 480, + 226 + ], + "type": "text", + "content": "3. Caesar, H., Bankiti, V., Lang, A., Vora, S., Liong, V., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuScenes: A multimodal dataset for autonomous driving. In: CVPR (2020) 3, 4, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 228, + 480, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 480, + 249 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 480, + 249 + ], + "type": "text", + "content": "4. Chen, H., Manhardt, F., Navab, N., Busam, B.: Texpose: Neural texture learning for self-supervised 6d object pose estimation. In: CVPR (2023) 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 250, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 250, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 138, + 250, + 480, + 270 + ], + "type": "text", + "content": "5. Chen, Y., Lee, G.H.: DBARF: deep bundle-adjusting generalizable neural radiance fields. In: CVPR (2023) 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 271, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 271, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 138, + 271, + 480, + 293 + ], + "type": "text", + "content": "6. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the KITTI vision benchmark suite. In: CVPR (2012) 4, 8, 9" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 293, + 480, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 293, + 480, + 315 + ], + "spans": [ + { + "bbox": [ + 138, + 293, + 480, + 315 + ], + "type": "text", + "content": "7. Girshick, R., Radosavovic, I., Gkioxari, G., Dollar, P., He, K.: Detectron. https://github.com/facebookresearch/detectron (2018) 9" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 316, + 446, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 316, + 446, + 325 + ], + "spans": [ + { + "bbox": [ + 138, + 316, + 446, + 325 + ], + "type": "text", + "content": "8. Gkioxari, G., Johnson, J., Malik, J.: Mesh R-CNN. In: ICCV (2019) 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 326, + 480, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 326, + 480, + 347 + ], + "spans": [ + { + "bbox": [ + 138, + 326, + 480, + 347 + ], + "type": "text", + "content": "9. Henderson, P., Tsiminaki, V., Lampert, C.H.: Leveraging 2d data to learn textured 3d mesh generation. In: CVPR (2020) 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 348, + 480, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 348, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 138, + 348, + 480, + 390 + ], + "type": "text", + "content": "0. Hodañ, T., Michel, F., Brachmann, E., Kehl, W., Glent Buch, A., Kraft, D., Drost, B., Vidal, J., Ihrke, S., Zabulis, X., Sahin, C., Manhardt, F., Tombari, F., Kim, T.K., Matas, J., Rother, C.: BOP: Benchmark for 6D object pose estimation. ECCV (2018) 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 392, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 392, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 138, + 392, + 480, + 424 + ], + "type": "text", + "content": "1. Insafutdinov, E., Campbell, D., Henriques, J.F., Vedaldi, A.: Snes: Learning probably symmetric neural surfaces from incomplete data. In: Avidan, S., Brostow, G.J., Cissé, M., Farinella, G.M., Hassner, T. (eds.) ECCV. pp. 367-383 (2022) 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 425, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 425, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 138, + 425, + 480, + 445 + ], + "type": "text", + "content": "2. Jang, W., Agapito, L.: Codenerf: Disentangled neural radiance fields for object categories. In: ICCV (2021) 2, 3, 8, 13" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 447, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 447, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 138, + 447, + 480, + 479 + ], + "type": "text", + "content": "3. Kanazawa, A., Tulsiani, S., Efros, A.A., Malik, J.: Learning category-specific mesh reconstruction from image collections. In: Ferrari, V., Hebert, M., Sminchisescu, C., Weiss, Y. (eds.) ECCV (2018) 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 479, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 479, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 138, + 479, + 480, + 501 + ], + "type": "text", + "content": "4. Kehl, W., Manhardt, F., Tombari, F., Ilic, S., Navab, N.: SSD-6D: making rgb-based 3d detection and 6d pose estimation great again. In: ICCV (2017) 4, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 502, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 502, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 138, + 502, + 480, + 522 + ], + "type": "text", + "content": "5. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph. (2023) 14" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 523, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 523, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 523, + 480, + 544 + ], + "type": "text", + "content": "6. Kumar, A., Brazil, G., Corona, E., Parchami, A., Liu, X.: DEVIANT: depth equivariant network for monocular 3d object detection. In: ECCV (2022) 2, 4, 9" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 545, + 480, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 545, + 480, + 577 + ], + "spans": [ + { + "bbox": [ + 138, + 545, + 480, + 577 + ], + "type": "text", + "content": "7. Kumar, A., Guo, Y., Huang, X., Ren, L., Liu, X.: SeaBird: Segmentation in Bird's View with Dice Loss Improves Monocular 3D Detection of Large Objects. In: CVPR (2024) 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 578, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 480, + 609 + ], + "type": "text", + "content": "8. Kundu, A., Genova, K., Yin, X., Fathi, A., Pantofaru, C., Guibas, L.J., Tagliasacchi, A., Dellaert, F., Funkhouser, T.A.: Panoptic neural fields: A semantic object-aware neural scene representation. In: CVPR (2022) 2, 3, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "type": "text", + "content": "9. Kundu, A., Li, Y., Rehg, J.M.: 3d-rcnn: Instance-level 3d object reconstruction via render-and-compare. In: CVPR (2018) 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "text", + "content": "20. Li, F., Yu, H., Shugurov, I., Busam, B., Yang, S., Ilic, S.: Nerf-posed: A first-reconstruct-then-regress approach for weakly-supervised 6d object pose estimation. CoRR (2022) 4" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "SUP-NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 148 + ], + "type": "text", + "content": "21. Li, Y., Wang, G., Ji, X., Xiang, Y., Fox, D.: Deepim: Deep iterative matching for 6d pose estimation. In: Ferrari, V., Hebert, M., Sminchisescu, C., Weiss, Y. (eds.) ECCV (2018) 4, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 149, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 480, + 171 + ], + "type": "text", + "content": "22. Li, Y., Wang, G., Ji, X., Xiang, Y., Fox, D.: Deepim: Deep iterative matching for 6d pose estimation. In: ECCV (2018) 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 171, + 480, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 171, + 480, + 191 + ], + "spans": [ + { + "bbox": [ + 132, + 171, + 480, + 191 + ], + "type": "text", + "content": "23. Lin, Y., Florence, P., Barron, J.T., Rodriguez, A., Isola, P., Lin, T.: inerf: Inverting neural radiance fields for pose estimation. In: IROS (2021) 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 192, + 480, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 192, + 480, + 213 + ], + "spans": [ + { + "bbox": [ + 132, + 192, + 480, + 213 + ], + "type": "text", + "content": "24. Lipson, L., Teed, Z., Goyal, A., Deng, J.: Coupled iterative refinement for 6d multi-object pose estimation. In: CVPR (2022) 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 213, + 480, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 213, + 480, + 246 + ], + "spans": [ + { + "bbox": [ + 132, + 213, + 480, + 246 + ], + "type": "text", + "content": "25. Lu, Y., Ma, X., Yang, L., Zhang, T., Liu, Y., Chu, Q., Yan, J., Ouyang, W.: Geometry uncertainty projection network for monocular 3d object detection. In: ICCV (2021) 2, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 246, + 480, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 246, + 480, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 480, + 278 + ], + "type": "text", + "content": "26. Merrill, N., Guo, Y., Zuo, X., Huang, X., Leutenegger, S., Peng, X., Ren, L., Huang, G.: Symmetry and uncertainty-aware object SLAM for 6 dof object pose estimation. In: CVPR (2022) 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 278, + 480, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 278, + 480, + 311 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 480, + 311 + ], + "type": "text", + "content": "27. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM pp. 99-106 (2021) 2, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 311, + 480, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 311, + 480, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 311, + 480, + 332 + ], + "type": "text", + "content": "28. Mousavian, A., Anguelov, D., Flynn, J., Kosecka, J.: 3d bounding box estimation using deep learning and geometry. In: CVPR (2017) 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 332, + 480, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 480, + 364 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 480, + 364 + ], + "type": "text", + "content": "29. Müller, N., Simonelli, A., Porzi, L., Bulò, S.R., Nießner, M., Kontschieder, P.: Autorf: Learning 3d object radiance fields from single view observations. In: CVPR (2022) 2, 3, 4, 6, 7, 9, 10, 11, 12, 14" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 364, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 480, + 385 + ], + "type": "text", + "content": "30. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph. pp. 102:1-102:15 (2022) 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 386, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 480, + 407 + ], + "type": "text", + "content": "31. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph. (2022) 14" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 407, + 480, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 480, + 428 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 480, + 428 + ], + "type": "text", + "content": "32. Park, D., Ambrus, R., Guizilini, V., Li, J., Gaidon, A.: Is pseudo-lidar needed for monocular 3d object detection? In: ICCV (2021) 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 428, + 480, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 428, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 132, + 428, + 480, + 449 + ], + "type": "text", + "content": "33. Park, K., Patten, T., Vincze, M.: Pix2pose: Pix2pose: Pixel-wise coordinate regression of objects for 6d pose estimation. In: ICCV (2019) 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 449, + 480, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 449, + 480, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 449, + 480, + 482 + ], + "type": "text", + "content": "34. Pavllo, D., Tan, D.J., Rakotosaona, M.J., Tombari, F.: Shape, pose, and appearance from a single image via bootstrapped radiance field inversion. In: CVPR (2023) 2, 3, 4, 7, 9, 10, 11, 12, 14" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 482, + 480, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 482, + 480, + 513 + ], + "spans": [ + { + "bbox": [ + 132, + 482, + 480, + 513 + ], + "type": "text", + "content": "35. Rad, M., Lepetit, V.: Bb8: A scalable, accurate, robust to partial occlusion method for predicting the 3d poses of challenging objects without using depth. ICCV (2017) 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 514, + 480, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 480, + 535 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 480, + 535 + ], + "type": "text", + "content": "36. Simonelli, A., Bulò, S.R., Porzi, L., Lopez-Antequera, M., Kontschieder, P.: Disentangling monocular 3d object detection. In: ICCV (2019) 2, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 536, + 480, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 480, + 557 + ], + "type": "text", + "content": "37. Solà, J., Deray, J., Atchuthan, D.: A micro lie theory for state estimation in robotics. CoRR (2018) 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 557, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 480, + 611 + ], + "type": "text", + "content": "38. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., Vasudevan, V., Han, W., Ngiam, J., Zhao, H., Timofeev, A., Ettinger, S., Krivokon, M., Gao, A., Joshi, A., Zhang, Y., Shlens, J., Chen, Z., Anguelov, D.: Scalability in perception for autonomous driving: Waymo open dataset. In: CVPR (2020) 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 611, + 480, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 633 + ], + "type": "text", + "content": "39. Tekin, B., Sinha, S.N., Fua, P.: Real-time seamless single shot 6d object pose prediction. In: CVPR (2018) 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "text", + "content": "40. Tian, S., Cai, Y., Yu, H.X., Zakharov, S., Liu, K., Gaidon, A., Li, Y., Wu, J.: Multi-object manipulation via object-centric neural scattering functions. In: CVPR (2023) 2" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "Y. Guo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 402 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "41. Wang, H., Sridhar, S., Huang, J., Valentin, J., Song, S., Guibas, L.J.: Normalized object coordinate space for category-level 6d object pose and size estimation. In: CVPR (2019) 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 481, + 171 + ], + "type": "text", + "content": "42. Wang, T., Zhu, X., Pang, J., Lin, D.: FCOS3D: fully convolutional one-stage monocular 3d object detection. In: ICCVW (2021) 4, 9, 12" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 193 + ], + "type": "text", + "content": "43. Weihs, L., Deitke, M., Kembhavi, A., Mottaghi, R.: Visual room rearrangement. In: CVPR (June 2021) 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 216 + ], + "type": "text", + "content": "44. Xiang, Y., Schmidt, T., Narayanan, V., Fox, D.: PoseCNN: A convolutional neural network for 6D object pose estimation in cluttered scenes. In: RSS (2018) 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 216, + 481, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 247 + ], + "type": "text", + "content": "45. Yang, B., Zhang, Y., Xu, Y., Li, Y., Zhou, H., Bao, H., Zhang, G., Cui, Z.: Learning object-compositional neural radiance field for editable scene rendering. In: ICCV (2021) 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "type": "text", + "content": "46. Yu, A., Ye, V., Tancik, M., Kanazawa, A.: pixelnerf: Neural radiance fields from one or few images. In: CVPR (2021) 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 271, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 481, + 293 + ], + "type": "text", + "content": "47. Yu, H., Wu, J., Yi, L.: Rotationally equivariant 3d object detection. In: CVPR (2022) 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 293, + 481, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 293, + 481, + 324 + ], + "spans": [ + { + "bbox": [ + 133, + 293, + 481, + 324 + ], + "type": "text", + "content": "48. Zakharov, S., Ambrus, R.A., Guizilini, V., Park, D., Kehl, W., Durand, F., Tenenbaum, J.B., Sitzmann, V., Wu, J., Gaidon, A.: Single-shot scene reconstruction. In: CoRL (2021) 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 325, + 481, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 481, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 481, + 347 + ], + "type": "text", + "content": "49. Zakharov, S., Kehl, W., Bhargava, A., Gaidon, A.: Autolabeling 3d objects with differentiable rendering of SDF shape priors. In: CVPR (2020) 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 348, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 348, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 348, + 481, + 369 + ], + "type": "text", + "content": "50. Zakharov, S., Shugurov, I., Ilic, S.: Dpod: 6d pose object detector and refiner. In: ICCV (2019) 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 370, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 370, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 370, + 481, + 402 + ], + "type": "text", + "content": "51. Zhou, A., Kim, M.J., Wang, L., Florence, P., Finn, C.: Nerf in the palm of your hand: Corrective augmentation for robotics via novel-view synthesis. In: CVPR (2023) 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "SUP-NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/SV3D_ Novel Multi-view Synthesis and 3D Generation from a Single Image using Latent Video Diffusion/55466408-8e29-4541-9064-bae13a2c8fbf_content_list.json b/2024/SV3D_ Novel Multi-view Synthesis and 3D Generation from a Single Image using Latent Video Diffusion/55466408-8e29-4541-9064-bae13a2c8fbf_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0fb279ae042010831ef364860d498c660caf06e0 --- /dev/null +++ b/2024/SV3D_ Novel Multi-view Synthesis and 3D Generation from a Single Image using Latent Video Diffusion/55466408-8e29-4541-9064-bae13a2c8fbf_content_list.json @@ -0,0 +1,3001 @@ +[ + { + "type": "text", + "text": "SV3D: Novel Multi-view Synthesis and 3D Generation from a Single Image using Latent Video Diffusion", + "text_level": 1, + "bbox": [ + 243, + 140, + 759, + 205 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vikram Voleti*, Chun-Han Yao*, Mark Boss*, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani", + "bbox": [ + 225, + 234, + 776, + 266 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/83325efc4c0195ff4b6a0c61c5348cc40f69864fffe4d7858886e4b0a8f2a849.jpg", + "image_caption": [ + "Fig. 1: Stable Video 3D (SV3D). From a single image, SV3D generates consistent novel multi-view images. We then optimize a 3D representation with SV3D-generated views resulting in high-quality 3D meshes." + ], + "image_footnote": [], + "bbox": [ + 225, + 276, + 769, + 412 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. We present Stable Video 3D (SV3D) — a latent video diffusion model for high-resolution, image-to-multi-view generation of orbital videos around a 3D object. Recent works propose to adapt 2D generative models for novel view synthesis (NVS) and 3D optimization. However, these methods have several disadvantages due to limited views or inconsistent NVS, affecting the performance of 3D object generation. In this work, we propose SV3D that adapts image-to-video diffusion model for novel multi-view synthesis and 3D generation, thereby leveraging the generalization and multi-view consistency of the video models, while further adding explicit camera control for NVS. We also propose improved 3D optimization techniques for image-to-3D generation using SV3D and its NVS outputs. Extensive experiments on multiple datasets with 2D and 3D metrics and user study demonstrate SV3D's state-of-the-art performance on NVS as well as 3D reconstruction compared to prior works.", + "bbox": [ + 259, + 484, + 740, + 679 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: 3D synthesis $\\cdot$ Video generation $\\cdot$ Novel view synthesis", + "bbox": [ + 261, + 691, + 712, + 707 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 733, + 375, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Single-image 3D object reconstruction is a long-standing problem in computer vision with a wide range of applications in game design, AR/VR, e-commerce, robotics, etc. It is a highly challenging and ill-posed problem as it requires lifting 2D pixels to 3D space while also reasoning about the unseen portions of the object in 3D.", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite being a long-standing vision problem, it is only recently that practically useful results are being produced by leveraging advances in the generative AI. This is mainly made possible by the large-scale pretraining of generative models which enables sufficient generalization to various domains. A typical strategy is to use image-based 2D generative models (e.g., Imagen [42], Stable Diffusion (SD) [40]) to provide a 3D optimization loss function for unseen novel views of a given object [23,31,37]. In addition, several works repurpose these 2D generative models to perform novel view synthesis (NVS) from a single image [26,27,30,54], and then use the synthesized novel views for 3D generation. Conceptually, these works mimic a typical photogrammetry-based 3D object capture pipeline, i.e., first photographing multi-view images of an object, followed by 3D optimization; except that the explicit multi-view capture is replaced by novel-view synthesis using a generative model, either via text prompt or camera pose control.", + "bbox": [ + 212, + 146, + 787, + 343 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A key issue in these generation-based reconstruction methods is the lack of multi-view consistency in the underlying generative model, resulting in inconsistent novel views. Some works try to address this by jointly reasoning a 3D representation during NVS [6,15,27], but at the cost of high compute and data, often still resulting in inconsistent geometric and texture details. In this work, we tackle this issue by adapting a high-resolution, image-conditioned video diffusion model for NVS followed by 3D generation.", + "bbox": [ + 212, + 345, + 789, + 450 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Novel Multi-view Synthesis. We adapt a latent video diffusion model (Stable Video Diffusion - SVD [2]) to generate multiple novel views of a given object with explicit camera pose conditioning. SVD demonstrates excellent multi-view consistency for video generation, and we repurpose it for NVs. In addition, SVD also has good generalization capabilities as it is trained on large-scale image and video data that are more readily available than large-scale 3D data. In short, we adapt the video diffusion model for NVs from a single image with three useful properties for 3D object generation: pose-controllable, multi-view consistent, and generalizable. We call our resulting NVS network 'SV3D'. To our knowledge, this is the first work that adapts a video diffusion model for explicit pose-controlled view synthesis. Some contemporary works such as [2, 29] demonstrate this, but only generate orbital videos without any explicit camera control.", + "bbox": [ + 212, + 455, + 789, + 636 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Generation. We then utilize our SV3D model for 3D object generation by optimizing a NeRF and DMTet mesh in a coarse-to-fine manner. Benefiting from the multi-view consistency in SV3D, we are able to produce high-quality 3D meshes directly from the SV3D novel view images. We also design a masked score distillation sampling (SDS) [37] loss to further enhance 3D quality in the regions that are not visible in the SV3D-predicted novel views. In addition, we propose to jointly optimize a disentangled illumination model along with 3D shape and texture, effectively reducing the issue of baked-in lighting.", + "bbox": [ + 212, + 640, + 787, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We perform extensive comparisons of both our NVS and 3D generation results with respective state-of-the-art methods, demonstrating considerably better outputs with SV3D. For NVS, SV3D shows high-level of multi-view consistency and generalization to real-world images while being pose controllable. Our resulting 3D meshes are able to capture intricate geometric and texture details.", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "V. Voleti et al.", + "bbox": [ + 271, + 114, + 370, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Background", + "text_level": 1, + "bbox": [ + 215, + 143, + 367, + 162 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Novel View Synthesis", + "text_level": 1, + "bbox": [ + 215, + 174, + 441, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We organize the related works along three crucial aspects of novel view synthesis (NVS): generalization, controllability, and multi-view (3D) consistency.", + "bbox": [ + 212, + 198, + 782, + 228 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Generalization. Diffusion models [14, 48] have recently shown to be powerful generative models that can generate a wide variety of images [3, 40, 41] and videos [2, 12, 53] by iteratively denoising a noise sample. Among these models, the publicly available Stable Diffusion (SD) [40] and Stable Video Diffusion (SVD) [2] demonstrate strong generalization ability by being trained on extremely large datasets like LAION [43] and LVD [2]. Hence, they are commonly used as foundation models for various generation tasks, e.g. novel view synthesis.", + "bbox": [ + 212, + 229, + 784, + 335 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Controllability. Ideally, a controllable NVS model allows us to generate an image corresponding to any arbitrary viewpoint. For this, Zero123 [26] repurposes an image diffusion model to a novel view synthesizer, conditioned on a single-view image and the pose difference between the input and target views. Follow-up works such as Zero123XL [8] and Stable Zero123 [49] advance the quality of diffusion-based NVS, as well as the trained NeRFs using SDS loss. However, they only generate one novel view at a time, and thus are not designed to be multi-view consistent inherently. Recent works such as EscherNet [21] and Free3D [64] are capable of better multi-view consistency with intelligent camera position embedding design. However, they only use image-based diffusion models, and generate images at $256 \\times 256$ resolution. We finetune a video diffusion model to generate novel views at $576 \\times 576$ resolution.", + "bbox": [ + 212, + 338, + 785, + 518 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multi-view Consistency. Multi-view consistency is the most critical requirement for high-quality NVS and 3D generation. To address the inconsistency issue, MVDream [46], SyncDreamer [27], HexGen3D [30], and Zero123++ [45] propose to generate multiple (specific) views of an object simultaneously. However, they are not controllable: given a conditional image, they only generate specific views, not arbitrary viewpoints. Moreover, they were finetuned from image-based diffusion models, i.e. multi-view consistency was imposed by adding interaction among the multiple generated views through cross-attention layers. Hence, their output quality is limited to the generalizability of their base image-based model, and their 3D finetuning dataset. Efficient-3DiM finetunes the SD model with a stronger vision transformer DINO v2 [36]. Consistent-1-to-3 [59] and SPAD [18] leverage epipolar geometry. One-2-3-45 [25] and One-2-3-45++ [24] train additional 3D network using the 2D generator's outputs. MVDream [46], Consistent123 [55], and Wonder3D [28] also train multi-view diffusion models, yet still require post-processing for video rendering. SyncDreamer [27] and ConsistentNet [57] employ 3D representation into the latent diffusion model.", + "bbox": [ + 212, + 521, + 785, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Exploiting Video Diffusion Models. To achieve better generalization and multi-view consistency, some contemporary works exploit the temporal priors in video diffusion models for NVs. For instance, Vivid-1-to-3 [22] combines a view-conditioned diffusion model and video diffusion model to generate consistent views. SVD-MV [2] and IM-3D [29] finetune a video diffusion model for NVs.", + "bbox": [ + 212, + 763, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "SV3D", + "bbox": [ + 687, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, they generate $\\leq 360^{\\circ}$ views at the same elevation only. Unlike SV3D, none of them are capable of rendering any arbitrary view of the 3D object.", + "bbox": [ + 215, + 146, + 784, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We argue that the existing NVS and 3D generation methods do not fully leverage the superior generalization capability, controllability, and consistency in video diffusion models. In this paper, we fill this important gap and train SV3D, a state-of-the-art novel multi-view synthesis video diffusion model at $576 \\times 576$ resolution, and leverage it for 3D generation.", + "bbox": [ + 215, + 176, + 785, + 252 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 3D Generation", + "text_level": 1, + "bbox": [ + 215, + 273, + 385, + 289 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recent advances in 3D representations and diffusion-based generative models have significantly improved the quality of image-to-3D generation. Here, we briefly summarize the related works in these two categories.", + "bbox": [ + 215, + 299, + 785, + 343 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3D Representation. 3D generation has seen great progress since the advent of Neural Radiance Fields (NeRFs) [32] and its subsequent variants [1], which implicitly represents a 3D scene as a volumetric function, typically parameterized by a neural network. Notably, Instant-NGP [33] introduces a hash grid feature encoding that can be used as a NeRF backbone for fast inference and ability to recover complex geometry. On the other hand, several recent works improve from an explicit representation such as DMTet [44], which is capable of generating high-resolution 3D shapes due to its hybrid SDF-Mesh representation and high memory efficiency. Similar to [23, 38], we adopt coarse-to-fine training for 3D generation, by first learning a rough object with Instant-NGP NeRF and then refining it using the DMTet representation.", + "bbox": [ + 215, + 345, + 785, + 511 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Diffusion-Based 3D Generation. Several recent works [16, 34] train a 3D diffusion model to to learn these flexible 3D representations, which, however, lack generalizability due to the scarcity of 3D data. To learn 3D generation without ground truth 3D data, image/multi-view diffusion models have been used as guidance. DreamFusion [37] and follow-up works [23, 31] leverage a trained image diffusion model as a 'scoring' function and calculate SDS loss for text-to-3D generation. However, they are prone to artifacts like Janus problem [31, 37] and over-saturated texture. Inspired by Zero123 [26], recent works [21, 24, 25, 27, 29, 38, 45, 49, 51, 64] finetune image/video diffusion models to generate novel view images as a stronger guidance. Our method shares the same spirit as this line of work, but produces denser, controllable, and more consistent multi-view images, thus resulting in better 3D generation quality.", + "bbox": [ + 215, + 513, + 785, + 695 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 SV3D: Novel Multi-view Synthesis", + "text_level": 1, + "bbox": [ + 215, + 717, + 588, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our main idea is to repurpose temporal consistency in a video diffusion model for spatial 3D consistency of an object. Specifically, we finetune SVD [2] to generate an orbital video around a 3D object, conditioned on a single-view image. This orbital video need not be at the same elevation, or at regularly spaced azimuth angles. SVD is well-suited for this task since it is trained to generate smooth and consistent videos on large-scale datasets of real and high-quality videos.", + "bbox": [ + 215, + 750, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "V. Voleti et al.", + "bbox": [ + 271, + 114, + 370, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/12882c02dace176adb8a7308987935aef341e849263fcf5bef9dae01d0b9ae9f.jpg", + "image_caption": [ + "Fig. 2: SV3D Architecture. We add the sinusoidal embedding of the camera orbit elevation and azimuth angles $(e, a)$ to that of the noise step $t$ , and feed the sum to the convolutional blocks in the UNet. We feed the single input image's CLIP embedding to the attention blocks, and concatenate its latent embedding to the noisy state $\\mathbf{z}_t$ ." + ], + "image_footnote": [ + ":FullyConnected :Embedding :VAE Encoder" + ], + "bbox": [ + 222, + 142, + 643, + 292 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6ad6fbf896b7ea7b537a65d35f76b1b60f5fab8ddfe00902463a218b7e50f49b.jpg", + "image_caption": [ + "Static" + ], + "image_footnote": [ + "We use two types of orbits for training the SV3D models." + ], + "bbox": [ + 656, + 154, + 772, + 210 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f238593dc45e29a47af60466306c47f3e94968dec88978f5a3e2b18e9b09197d.jpg", + "image_caption": [ + "Dynamic", + "Fig.3:Static vs. DynamicOrbits." + ], + "image_footnote": [], + "bbox": [ + 660, + 232, + 772, + 300 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The exposure to superior data quantity and quality makes it more generalizable and multi-view consistent, and the flexibility of the SVD architecture makes it amenable to be finetuned for camera controllability.", + "bbox": [ + 212, + 420, + 784, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Some prior works attempt to leverage such properties by finetuning image diffusion models, training video diffusion models from scratch, or finetuning video diffusion models to generate pre-defined views at the same elevation (static orbit) around an object [2, 29]. However, we argue that these methods do not fully exploit the potential of video diffusion models. To the best of our knowledge, SV3D is the first video diffusion-based framework for controllable multi-view synthesis at $576 \\times 576$ resolution (and subsequently for 3D generation).", + "bbox": [ + 212, + 468, + 784, + 574 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Problem Setting. Formally, given an image $\\mathbf{I} \\in \\mathbb{R}^{3 \\times H \\times W}$ of an object, our goal is to generate an orbital video $\\mathbf{J} \\in \\mathbb{R}^{K \\times 3 \\times H \\times W}$ around the object of $K = 21$ multi-view images along a camera trajectory $\\pmb{\\pi} \\in \\mathbb{R}^{K \\times 2} = \\{(e_i, a_i)\\}_{i=1}^K$ as a sequence of $K$ tuples of elevation $e$ and azimuth $a$ angles. We assume the camera always looks at the center of an object (origin of world coordinates), so any viewpoint can be specified by only two parameters: $e$ and $a$ . We generate this orbital video by iteratively denoising samples from a learned conditional distribution $p(\\mathbf{J}|\\mathbf{I}, \\pmb{\\pi})$ , parameterized by a video diffusion model.", + "bbox": [ + 212, + 577, + 784, + 699 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SV3D Architecture. As shown in Fig. 2, the architecture of SV3D builds on that of SVD consisting of a UNet with multiple layers, each layer containing sequences of one residual block with Conv3D layers, and two transformer blocks (spatial and temporal) with attention layers. (i) We remove the vector conditionings of 'fps id' and 'motion bucket id' since they are irrelevant for SV3D. (ii) The conditioning image is concatenated to the noisy latent state input $\\mathbf{z}_t$ at noise timestep $t$ to the UNet, after being embedded into latent space by the VAE encoder of SVD. (iii) The CLIP-embedding [39] matrix of the conditioning image is provided to the cross-attention layers of each transformer block as its key and", + "bbox": [ + 212, + 703, + 785, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "SV3D", + "bbox": [ + 689, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/cc3a266b1abb00711f934e48ce92e388c4a3f9fd8edb84a2c3658e78cbe0e122.jpg", + "image_caption": [ + "Fig. 4: Linear vs. Triangle CFG Scaling. Notice Fig. 5: LPIPS vs. Frame increased oversharping in the penultimate frame in Number. SV3D has the best re-linear scaling vs. our proposed triangle scaling. construction metric per frame." + ], + "image_footnote": [], + "bbox": [ + 220, + 150, + 531, + 311 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ae353a2baa0034bb01d4ff5f7d14c1f6e1b9128a12167fafbda9d0c558ca3168.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 566, + 151, + 779, + 313 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "value, the query being the feature at that layer. (iv) The camera trajectory is fed into the residual blocks along with the diffusion noise timestep. The camera pose angles $e_i$ and $a_i$ and the noise timestep $t$ are first embedded into sinusoidal position embeddings. Then, the camera pose embeddings are concatenated together, linearly transformed, and added to the noise timestep embedding. This is fed to every residual block, where they are added to the block's output feature (after being linearly transformed again to match the feature size).", + "bbox": [ + 212, + 364, + 787, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Static v.s. Dynamic Orbits. As shown in Fig. 3, we design static and dynamic orbits to study the effects of camera pose conditioning. In a static orbit, the camera circles around an object at regularly-spaced azimuths at the same elevation angle as that in the conditioning image. The disadvantage with the static orbit is that we might not get any information about the top or bottom of the object depending on the conditioning elevation angle. In a dynamic orbit, the azimuths can be irregularly spaced, and the elevation can vary per view. To build a dynamic orbit, we sample a static orbit, add small random noise to the azimuth angles, and add a random weighted combination of sinusoids with different frequencies to its elevation. This provides temporal smoothness, and ensures that the camera trajectory loops around to end at the same azimuth and elevation as those of the conditioning image.", + "bbox": [ + 212, + 474, + 787, + 655 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Thus, with this strategy, we are able to tackle all three aspects of generalization, controllability, and consistency in novel multi-view synthesis by leveraging video diffusion models, additionally conditioning on camera trajectory, and repurposing temporal consistency for spatial 3D object consistency, respectively.", + "bbox": [ + 212, + 656, + 787, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Triangular CFG Scaling. SVD uses a linearly increasing scale for classifier-free guidance (CFG) from 1 to 4 across the generated frames. However, this scaling causes the last few frames in our generated orbits to be over-sharpened, as shown in Fig. 4 Frame 20. Since we generate videos looping back to the front-view image, we propose to use a triangle wave CFG scaling during inference: linearly increase CFG from 1 at the front view to 2.5 at the back view, then linearly decrease it back to 1 at the front view. Fig. 4 also demonstrates that our triangle CFG scaling produces more details in the back view (Frame 12).", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "V. Voleti et al.", + "bbox": [ + 271, + 114, + 370, + 126 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Models. We train three image-to-3D-video models finetuned from SVD. First, we train a pose-unconditioned model, $\\mathrm{SV3D}^u$ , which generates a video of static orbit around an object while only conditioned on a single-view image. Note that unlike SVD-MV [2], we do not provide the elevation angle to the pose-unconditioned model, as we find that the model is able to infer it from the conditioning image. Our second model, the pose-conditioned $\\mathrm{SV3D}^c$ is conditioned on the input image as well as a sequence of camera elevation and azimuth angles in an orbit, trained on dynamic orbits. Following SVD's [2] intuition to progressively increase the task difficulty during training, we train our third model, $\\mathrm{SV3D}^p$ , by first finetuning SVD to produce static orbits unconditionally, then further finetuning on dynamic orbits with camera pose condition.", + "bbox": [ + 212, + 146, + 787, + 313 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Training Details. We train SV3D on the Objverse dataset [9], which contains synthetic 3D objects covering a wide diversity. For each object, we render 21 frames around it on random color background at $576 \\times 576$ resolution, field-of-view of 33.8 degrees. We choose to finetune the SVD-xt model to output 21 frames. All three models (SV3D $^u$ , SV3D $^c$ , SV3D $^p$ ) are trained for 105k iterations in total (SV3D $^p$ is trained unconditionally for 55k iterations and conditionally for 50k iterations), with an effective batch size of 64 on 4 nodes of 8 80GB A100 GPUs for around 6 days. For more training details, please see the appendix.", + "bbox": [ + 212, + 314, + 787, + 435 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.1 Experiments and Results", + "text_level": 1, + "bbox": [ + 215, + 459, + 470, + 473 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Datasets. We evaluate the SV3D-generated multi-view images on static and dynamic orbits on the unseen GSO [10] and OmniObject3D [56] datasets. Since many GSO objects are the same items with slightly different colors, we filter 300 objects from GSO to reduce redundancy and maintain diversity. For each object, we render ground truth static and dynamic orbit videos and pick the last frame of each video as the conditioning image. We also conduct a user study on novel view videos from a dataset of 22 real-world images. (More details in appendix.)", + "bbox": [ + 212, + 486, + 787, + 593 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Metrics. We use the SV3D models to generate static and dynamic orbit videos corresponding to the ground truth camera trajectories in the evaluation datasets. We compare each generated frame with the corresponding ground truth frames, in terms of Learned Perceptual Similarity (LPIPS [63]), Peak Signal-to-Noise Ratio (PSNR), Structural SIMilarity (SSIM), Mean Squared-Error (MSE), and CLIP-score (CLIP-S). This range of metrics covers both pixel-level as well as semantic aspects. Note that testing on dynamic orbits evaluates the controllability of SV3D models, and all the metrics evaluate multi-view consistency.", + "bbox": [ + 212, + 594, + 787, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines. We compare SV3D with several recent NVS methods capable of generating arbitrary views, including Zero123 [26], Zero123-XL [8], SyncDreamer [27], Stable Zero123 [49], Free3D [64], EscherNet [21].", + "bbox": [ + 212, + 717, + 787, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results. As shown in Tabs. 1 to 4, our SV3D models achieve state-of-the-art performance on novel multi-view synthesis. Tabs. 1 and 3 show results on static orbits, and include all our three models. We see that even our pose-unconditioned model $\\mathrm{SV3D}^u$ performs better than all prior methods. Tabs. 2 and 4 show results on dynamic orbits, and include our pose-conditioned models $\\mathrm{SV3D}^c$ and $\\mathrm{SV3D}^p$ .", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "SV3D", + "bbox": [ + 689, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0e2e78ba032f8a75d450521803b42a294afe1fbb9621377fe07ebec3834f8d66.jpg", + "image_caption": [ + "Fig.6: Visual Comparison of Novel Multi-view Synthesis. SV3D is able to generate novel multi-views that are more detailed, faithful to the conditioning image, and multi-view consistent compared to the prior works." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 782, + 428 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablative Analyses. Interestingly, from Tabs. 1 and 3, we find that both $\\mathrm{SV3D}^c$ and $\\mathrm{SV3D}^p$ outperform $\\mathrm{SV3D}^u$ on generations of static orbits, even though $\\mathrm{SV3D}^u$ is trained specifically on static orbits. We also observe that $\\mathrm{SV3D}^p$ achieves better metrics than $\\mathrm{SV3D}^c$ on both static and dynamic orbits, making it the best performing SV3D model overall. This shows that progressive finetuning from easier (static) to harder (dynamic) tasks is indeed a favorable way to finetune a video diffusion model.", + "bbox": [ + 212, + 494, + 787, + 601 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visual Comparisons in Fig. 6 further demonstrate that SV3D-generated images are more detailed, faithful to the conditioning image, and multi-view consistent compared to the prior works.", + "bbox": [ + 212, + 604, + 785, + 651 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Quality Per Frame. We also observe from Fig. 5 that SV3D produces better quality at every frame. We plot the average LPIPS value for each generated frame, across generated GSO static orbit videos. The quality is generally worse around the back view, and better at the beginning and the end (i.e. near the conditioning image), as expected.", + "bbox": [ + 212, + 654, + 787, + 731 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "User Study on Real-World Images. We conducted a user survey to study human preference between static orbital videos generated by SV3D and by other methods. We asked 30 users to pick one between our SV3D-generated static video and other method-generated video as the best orbital video for the corresponding image, using 22 real-world images. We noted that users preferred SV3D-generated videos over Zero123XL, Stable Zero123, EscherNet, and Free3D, $96\\%$ , $99\\%$ , $96\\%$ , and $98\\%$ of the time, respectively.", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "V.Voleti et al.", + "bbox": [ + 271, + 114, + 372, + 127 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0a5b238eef1ca70ca08d4de0f6c7600701c355fe3da1c83f076b00ba6559c1f7.jpg", + "table_caption": [ + "Table 1: Evaluation of novel multi-view synthesis on GSO static orbits" + ], + "table_footnote": [], + "table_body": "
ModelLPIPS↓PSNR↑SSIM↑CLIP-S↑MSE↓
SyncDreamer [27]0.1715.780.760.870.03
Zero123 [26]0.1317.290.790.850.04
Zero123XL [8]0.1417.110.780.850.04
Stable Zero123 [49]0.1318.340.780.850.05
Free3D [64]0.1516.180.790.840.04
EscherNet [21]0.1316.730.790.850.03
SV3Du0.0921.140.870.890.02
SV3Dc0.0920.560.870.880.02
SV3Dp0.0821.260.880.890.02
", + "bbox": [ + 217, + 172, + 495, + 285 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/f1259838032aeaecffe1ac122e0dcf3f5de7695a76b54a4cac648d80eb0261bc.jpg", + "table_caption": [ + "Table 2: Evaluation of novel multi-view synthesis on GSO dynamic orbits" + ], + "table_footnote": [], + "table_body": "
ModelLPIPS↓PSNR↑SSIM↑CLIP-S↑MSE↓
Zero123 [26]0.1416.990.790.840.04
Zero123XL [8]0.1416.730.780.840.04
Stable Zero123 [49]0.1318.040.780.850.05
Free3D [64]0.1814.930.770.830.05
EscherNet [21]0.1316.470.790.840.03
SV3Dc0.1019.990.860.870.02
SV3Dp0.0920.380.870.870.02
", + "bbox": [ + 501, + 172, + 782, + 268 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/713505754bfb1171b42f2d6fcacbcfdf682d2880b9244881404c034c95d3d8df.jpg", + "table_caption": [ + "Table 3: Evaluation of novel multi-view synthesis on OmniObject3D static orbits" + ], + "table_footnote": [], + "table_body": "
ModelLPIPS↓PSNR↑SSIM↑CLIP-S↑MSE↓
Zero123 [26]0.1715.500.760.830.05
Zero123XL [8]0.1815.360.750.830.06
Stable Zero123 [49]0.1516.860.770.840.06
Free3D [64]0.1615.290.780.830.05
EscherNet [21]0.1714.630.740.830.05
SV3Du0.1019.680.860.860.02
SV3Dc0.1019.500.850.850.02
SV3Dp0.1019.910.860.860.02
", + "bbox": [ + 217, + 324, + 495, + 428 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/f3a2aa9278279c0f943e0cde03ee8fe8901f7e9479bce3b26c778e71af14d23e.jpg", + "table_caption": [ + "Table 4: Evaluation of novel multi-view synthesis on OmniObject3D dynamic orbits" + ], + "table_footnote": [], + "table_body": "
ModelLPIPS↓PSNR↑SSIM↑CLIP-S↑MSE↓
Zero123 [26]0.1615.780.770.820.05
Zero123XL [8]0.1715.490.760.830.05
Stable Zero123 [49]0.1416.740.770.830.05
Free3D [64]0.1914.280.760.820.06
EscherNet [21]0.1615.050.760.830.05
SV3Dc0.1019.210.850.840.02
SV3Dp0.1019.280.850.850.02
", + "bbox": [ + 501, + 324, + 782, + 419 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 3D Generation from a Single Image Using SV3D", + "text_level": 1, + "bbox": [ + 215, + 434, + 722, + 452 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We then generate 3D meshes of objects from a single image by leveraging SV3D. One way is to use the generated static/dynamic orbital samples from SV3D as direct reconstruction targets. Another way is to use SV3D as diffusion guidance with Score Distillation Sampling (SDS) loss [37].", + "bbox": [ + 212, + 470, + 785, + 531 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Since SV3D produces more consistent multi-views compared to existing works, we already observe higher-quality 3D reconstructions by only using SV3D outputs for reconstruction when compared to existing works. However, we observe that this naive approach often leads to artifacts like baked-in illumination, rough surfaces, and noisy texture, especially for the unseen regions in the reference orbit. Thus, we further propose several techniques to address these issues.", + "bbox": [ + 212, + 532, + 787, + 622 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Coarse-to-Fine Training. We adopt a two-stage, coarse-to-fine training scheme to generate a 3D mesh from input images, similar to [23, 38]. Fig. 7 illustrate an overview of our 3D optimization pipeline. In the coarse stage, we train an Instant-NGP [33] NeRF representation to reconstruct the SV3D-generated images (i.e. without SDS loss) at a lower resolution. In the fine stage, we extract a mesh from the trained NeRF using marching cubes [7], and adopt a DMTet [44] representation to finetune the 3D mesh using SDS-based diffusion guidance from SV3D at full-resolution. Finally, we use xatlas [60] to perform the UV unwrapping and export the output object mesh.", + "bbox": [ + 212, + 626, + 787, + 762 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Disentangled Illumination Model. Similar to other recent 3D object generation methods [23, 31, 37], our output target is a mesh with a diffuse texture. Typically, such SDS-based optimization techniques use random illuminations at every iteration. However, our SV3D-generated videos are under consistent illumination, i.e., the lighting remains static while the camera circles around an object.", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "SV3D", + "bbox": [ + 689, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/5987a83388c37819313fdba8e4461418165ad0f7445d54136e20437a06efa447.jpg", + "image_caption": [ + "Fig.7: 3D Optimization Pipeline. We perform a two-stage optimization. In the short first stage, we extract the general shape, texture and illumination from the SV3D generated multi-view images. In the second stage, we extract a mesh with marching cubes and use DMTet to further optimize the shape, texture and illumination. We not only use the SV3D-generated images but a soft-masked SDS loss for unseen areas. Dashed red lines represent backpropagation into the differentiable rendering pipeline." + ], + "image_footnote": [], + "bbox": [ + 230, + 145, + 769, + 311 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/5fb4bcf394d5d910bdad223a29f52256c4b8375432c15acf8e41f9abc8b6df6b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 250, + 414, + 290, + 481 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/b1fd04eee9a26b38127a2922f27179c09b2c5f689a5b2f9845930264b6ec190c.jpg", + "image_caption": [ + "Ref. images in static orbit" + ], + "image_footnote": [], + "bbox": [ + 254, + 488, + 287, + 554 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/671b192150010ac14eb20f5708976a826ad85e32c45dd3ec5c955b6436252316.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 318, + 415, + 364, + 482 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/0aa46e836604df6c791f979614c2ccd5b4bc2c345168216bf0c4ab3f241d5b3a.jpg", + "image_caption": [ + "3D from static orbit" + ], + "image_footnote": [], + "bbox": [ + 318, + 484, + 359, + 556 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/79e021486db2f103fbc401784306cab1644763354e4e721b7bbd123d2021c78a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 393, + 415, + 439, + 481 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/447888e32730a3ab3a5d887bb2a79dfe40d3e06d6d5271320cb26db27edd5135.jpg", + "image_caption": [ + "3D from dynamic orbit" + ], + "image_footnote": [], + "bbox": [ + 395, + 483, + 434, + 556 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/4f0ed928839d0d8218b7518328d81d2b74e9c02a17aaff33a944f63e08ca4072.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 498, + 415, + 584, + 472 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/aedeea4d1746ea26fe4ca91a08ab571a565b166952c5fad0f90461f52eb22694.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 586, + 415, + 671, + 472 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/3d48c3019a3a7f02f19c339f810d8c6b89cd3e8d3bb0774746da33172408360f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 679, + 415, + 764, + 472 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/789b011b0b7a02615f4bbbd432e0296095ac0502f6de6847d394c02e962a9a5b.jpg", + "image_caption": [ + "Fig.8: Influence of Training Orbits. We show that using a dynamic orbit is crucial to 3D generations that are complete from diverse views.", + "No SDS loss", + "Fig.9: Influence of SDS. Using our masked SDS loss, we are able to fill in unseen surfaces in the training orbit, producing a cleaner result without oversaturation or blurry artifacts caused by naive SDS." + ], + "image_footnote": [], + "bbox": [ + 511, + 479, + 588, + 537 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/d40e9b9de18b4d2252ca9c1c6b35d420a5ffeb0d43704dcb14aa16324e004ee7.jpg", + "image_caption": [ + "SDS loss" + ], + "image_footnote": [], + "bbox": [ + 596, + 479, + 673, + 537 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/cada0c9d47e64e8fc135525873334996b3cfe6ceedb5abaa1f14b17b5ab85e30.jpg", + "image_caption": [ + "Masked SDS loss" + ], + "image_footnote": [], + "bbox": [ + 686, + 479, + 763, + 537 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Hence, to disentangle lighting effects and obtain a cleaner texture, we propose to fit a simple illumination model of 24 Spherical Gaussians (SG) inspired by prior decomposition methods [4,62].", + "bbox": [ + 212, + 656, + 784, + 700 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We model white light and hence only use a scalar amplitude for the SGs. We only consider Lambertian shading, where the cosine shading term is approximated with another SG. We learn the parameters of the illumination SGs using a reconstruction loss between the rendered images and SV3D-generated images.", + "bbox": [ + 212, + 702, + 784, + 762 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Inspired by [13, 37] we reduce baked-in illumination with a loss term that replicates the HSV-value component of the input image $\\mathbf{I}$ with the rendered illumination $L$ : $\\mathcal{L}_{\\mathrm{illum}} = |V(\\mathbf{I}) - L|^2$ , with $V(\\pmb {c}) = \\max (c_r,c_g,c_b)$ . Given these changes, our disentangled illumination model is able to express lighting variation properly and can severely reduce baked-in illumination. Fig. 10 shows sample", + "bbox": [ + 212, + 763, + 785, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "V. Voleti et al.", + "bbox": [ + 271, + 114, + 370, + 126 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "reconstructions with and without our illumination modelling. From the results, it is clear that we are able to disentangle the illumination effects from the base color (e.g., dark side of the school bus).", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.1 3D Optimization Strategies and Losses", + "text_level": 1, + "bbox": [ + 214, + 214, + 581, + 229 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Reconstruction via Photometric Losses. Intuitively, we can treat the SV3D-generated images as multi-view pseudo-ground truth, and apply 2D reconstruction losses to train the 3D models. In this case, we apply photometric losses on the rendered images from NeRF or DMTet, including the pixel-level MSE loss, mask loss, and a perceptual LPIPS [63] loss. These photometric losses also optimize our illumination model through the differential rendering pipeline.", + "bbox": [ + 212, + 242, + 782, + 332 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Training Orbits. For 3D generation, we use SV3D to generate multi-view images following a camera orbit $\\pi_{\\mathrm{ref}}$ , referred to as the reference orbit (also see Fig. 7 for the overview). Fig. 8 shows sample reconstruction with using static and dynamic orbital outputs from SV3D. Using a dynamic orbit for training is crucial to high-quality 3D outputs when viewed from various angles, since some top/bottom views are missing in the static orbit (fixed elevation). Hence, for $\\mathrm{SV3D}^c$ and $\\mathrm{SV3D}^p$ , we render images on a dynamic orbit whose elevation follows a sine function to ensure that top and bo", + "bbox": [ + 212, + 335, + 550, + 530 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/d59104d1140387572463183ae508657a044358a3d5ca9d0739bf10a2b1286687.jpg", + "image_caption": [], + "image_footnote": [ + "Constant Illumination" + ], + "bbox": [ + 580, + 353, + 666, + 393 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/e2a63af3ffafaff4160f5fadf2e928ec4b17d3f63c5f1a2b3131aa56437ddefe.jpg", + "image_caption": [], + "image_footnote": [ + "SGs Illumination", + "Fig.10: Constant vs. SGs Illumination. Our SGs-based reconstructions do not exhibit darkening on the side of the bus, which enables easier and more convincing relighting for downstream applications." + ], + "bbox": [ + 686, + 353, + 772, + 393 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "SV3D-Based SDS Loss. In addition to the reconstruction losses, we can also use SV3D via score-distillation sampling (SDS) [37, 58]. Fig. 9 shows sample reconstructions with and without using SDS losses. As shown in Fig. 9 (left), although training with a dynamic orbit improves overall visibility, we observe that sometimes the output texture is still noisy, perhaps due to partial visibility, self-occlusions, or inconsistent texture/shape between images. Hence, we handle those unseen areas using SDS loss [37] with SV3D as a diffusion guidance.", + "bbox": [ + 212, + 532, + 782, + 638 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We sample a random camera orbit $\\pi_{\\mathrm{rand}}$ , and use our 3D NeRF/DMTet parameterized by $\\theta$ to render the views $\\hat{\\mathbf{J}}$ of the object along $\\pi_{\\mathrm{rand}}$ . Then, noise $\\epsilon$ at level $t$ is added to the latent embedding $\\mathbf{z}_t$ of $\\hat{\\mathbf{J}}$ , and the following SDS loss (taken expectation over $t$ and $\\epsilon$ ) is backpropagated through the differentiable rendering pipeline: $\\mathcal{L}_{\\mathrm{sds}} = w(t)\\bigl (\\epsilon_{\\phi}(\\mathbf{z}_t;\\mathbf{I},\\pi_{\\mathrm{rand}},t) - \\epsilon \\bigr)\\frac{\\partial\\hat{\\mathbf{J}}}{\\partial\\theta}$ , where $w$ is $t$ -dependent weight, $\\epsilon$ and $\\epsilon_{\\phi}$ are the added and predicted noise, and $\\phi$ and $\\theta$ are the parameters of SV3D and NeRF/DMTet, respectively. See Fig. 7 for an illustration of these loss functions in the overall pipeline.", + "bbox": [ + 212, + 638, + 784, + 762 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Masked SDS Loss. As shown in Fig. 9 (middle), in our experiments we found that adding the SDS loss naively can cause unstable training and unfaithful texture to the input images such as oversaturation or blurry artifacts. Therefore, we design a soft masking mechanism to only apply SDS loss on the unseen/occluded areas, allowing it to inpaint the missing details while preserving the texture of", + "bbox": [ + 212, + 763, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "SV3D", + "bbox": [ + 687, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/e96e640b561758b8f3934e1a727a4386cf96519b602cbeaf342608a3fdc58e0a.jpg", + "image_caption": [ + "Fig. 11: Visual Comparison of 3D Meshes. For each object, we show the conditioning image (left), and the output meshes rendered in two different views. Our generated meshes are more detailed, faithful to input images, and consistent in 3D. This demonstrates the quality of novel multi-view synthesis by our SV3D model." + ], + "image_footnote": [], + "bbox": [ + 243, + 143, + 751, + 279 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "clearly-visible surfaces in the training orbit (as seen in Fig. 9 (right)). Also, we only apply the masked SDS loss in the final stage of DMTet optimization, which greatly increased the convergence speed.", + "bbox": [ + 212, + 358, + 784, + 402 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We apply SDS loss on only those pixels in the random orbit views that are not likely visible in the reference orbit views. For this, we first render the object from the random camera orbit $\\pi_{\\mathrm{rand}}$ . For each random camera view, we obtain the visible surface points $p \\in \\mathbb{R}^3$ and their corresponding surface normals $\\boldsymbol{n}$ . Then, for each reference camera $i$ , we calculate the view directions $\\boldsymbol{v}_i$ of the surface $\\boldsymbol{p}$ towards its position $\\bar{\\pi}_{\\mathrm{ref}}^i \\in \\mathbb{R}^3$ (calculated from $\\pi_{\\mathrm{ref}}^i \\in \\mathbb{R}^2$ ) as $\\boldsymbol{v}_i = \\frac{\\bar{\\pi}_{\\mathrm{ref}}^i - \\boldsymbol{p}}{||\\bar{\\pi}_{\\mathrm{ref}}^i - \\boldsymbol{p}||}$ . We infer the visibility of this surface from the reference camera based on the dot product between $\\boldsymbol{v}_i$ and $\\boldsymbol{n}$ i.e. $\\boldsymbol{v}_i \\cdot \\boldsymbol{n}$ . Since higher values roughly indicate more visibility of the surface from the reference camera, we chose that reference camera $c$ that has maximum likelihood of visibility: $c = \\max_i (\\boldsymbol{v}_i \\cdot \\boldsymbol{n})$ . As we only want to apply SDS loss to unseen or grazing angle areas from $c$ , we use the smoothstep function $f_s$ to smoothly clip to $c$ 's visibility range $\\boldsymbol{v}_c \\cdot \\boldsymbol{n}$ . In this way, we create a pseudo visibility mask $M = 1 - f_s(\\boldsymbol{v}_c \\cdot \\boldsymbol{n}, 0, 0.5)$ , where $f_s(x; f_0, f_1) = \\hat{x}^2(3 - 2x)$ with $\\hat{x} = \\frac{x - f_0}{f_1 - f_0}$ . Thus, $M$ is calculated for each random camera render, and the combined visibility mask $\\mathbf{M}$ is applied to SDS loss: $\\mathcal{L}_{\\mathrm{mask-sds}} = \\mathbf{M}\\mathcal{L}_{\\mathrm{sds}}$ .", + "bbox": [ + 212, + 404, + 787, + 636 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Geometric Priors. Since our rendering-based optimization operates at the image level, we adopt several geometric priors to regularize the output shapes. We add a smooth depth loss from RegNeRF [35] and a bilateral normal smoothness loss [5] to encourage smooth 3D surfaces where the projected image gradients are low. Moreover, we obtain normal estimates from Omnidata [11] and calculate a mono normal loss similar to MonoSDF [61], which can effectively reduce noisy surfaces in the output mesh. Further details about the training losses and optimization process are available in the appendix.", + "bbox": [ + 212, + 638, + 787, + 758 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.2 Experiments and Results", + "text_level": 1, + "bbox": [ + 215, + 782, + 470, + 797 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Due to the strong regularization, we only require 600 steps in the coarse stage and 1000 in the fine stage. Overall, the entire mesh extraction process takes $\\approx 8$", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "V. Voleti et al.", + "bbox": [ + 269, + 114, + 372, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/48c824298c93e3d130e70bc250a70665f6a84254b1899ea3a6d012afe34e0102.jpg", + "image_caption": [ + "Fig. 12: Real-World 3D Results. Notice the accurate shape and details in our reconstructions even from the diverse images in-the-wild." + ], + "image_footnote": [], + "bbox": [ + 243, + 145, + 751, + 232 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "minutes without SDS loss, and $\\approx 20$ minutes with SDS loss. The coarse stage only takes $\\approx 2$ minutes and provides a full 3D representation of the object.", + "bbox": [ + 212, + 282, + 784, + 313 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Evaluation. We evaluate our 3D generation framework on 50 randomly sampled objects from the GSO dataset as described in Sec. 3.1. We compute image-based reconstruction metrics (LPIPS, PSNR, SSIM, MSE, and CLIP-S) between the ground truth (GT) GSO images, and rendered images from the trained 3D meshes on the same static/dynamic orbits. In addition, we compute 3D reconstruction metrics of Chamfer distance (CD) and 3D IoU between the GT and predicted meshes. We compare our SV3D-guided 3D generations with several prior methods including Point-E [34], Shap-E [16], One-2-3-45++ [24], DreamGaussian [50], SyncDreamer [27], EscherNet [21], Free3D [64], and Stable Zero123 [49].", + "bbox": [ + 212, + 315, + 784, + 450 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Visual Comparison. In Fig. 11, we show visual comparison of our results with those from prior methods. Qualitatively, Point-E [34] and Shap-E [16] often produce incomplete 3D shapes. DreamGaussian [50], SyncDreamer [27], EscherNet [21], and Free3D [64] outputs tend to contain rough surfaces. One-2-3-45++ [24] and Stable Zero123 [49] are able to reconstruct meshes with smooth surface, but lack geometric details. Our mesh outputs are detailed, faithful to input image, and consistent in 3D (see appendix for more examples). We also show 3D mesh renders from real-world images in Fig. 12.", + "bbox": [ + 212, + 454, + 785, + 574 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Quantitative Comparison. Tabs. 5 and 6 show the 2D and 3D metric comparisons respectively. All our 3D models achieve better 2D and 3D metrics compared to the prior and concurrent methods, showing the high-fidelity texture and geometry of our output meshes. We render all 3D meshes on the same dynamic orbits and compare them against the GT renders. Our best model, $\\mathrm{SV3D}^p$ , performs comparably to using GT renders for reconstruction in terms of the 3D metrics, which further demonstrates the 3D consistency of our generated images.", + "bbox": [ + 212, + 577, + 785, + 683 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effects of Photometric and (Masked) SDS Loss. As shown in Tabs. 5 and 6, the 3D outputs using both photometric and Masked SDS losses ( $\\mathbf{SV3D}^p$ ) achieves the best results, while training without SDS ( $\\mathbf{SV3D}^p$ no SDS') leads to marginally lower performance. This demonstrates that the images generated by SV3D are high-quality reconstruction targets, and are often sufficient for 3D generation without the cumbersome SDS-based optimization. Nevertheless, adding SDS generally improves quality, as also shown in Fig. 9.", + "bbox": [ + 212, + 686, + 785, + 792 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effects of SV3D Model and Training Orbit. As shown in Tabs. 5 and 6, $\\mathrm{SV3D}^p$ achieves the best performance among the three SV3D models, indicating that its synthesized novel views are most faithful to the input image and consis", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "SV3D", + "bbox": [ + 687, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/3a34e68cd530a417442722419b3d5961ec10c3054e8fc5216b2490d078b6cea5.jpg", + "table_caption": [ + "Table 5: 2D comparison of our 3D outputs against prior methods on the GSO dataset. Our best performing method uses $\\mathbf{SV3D}^p$ with dynamic (sine elevation) orbit and SDS guidance. Note that all our models achieve better 2D metrics than prior works." + ], + "table_footnote": [], + "table_body": "
ModelLPIPS↓PSNR↑SSIM↑MSE↓CLIP-S↑
GT renders0.07819.5080.8790.0140.897
EscherNet [21]0.17814.4380.8040.0410.835
Free3D [64]0.19714.2020.7990.0430.809
Stable Zero123 [49]0.16614.6350.8130.0400.805
SV3Du0.13315.9570.8340.0310.871
SV3Dc0.13216.3730.8340.0270.870
SV3Dpstatic orbit0.12516.8210.8480.0250.864
SV3Dpno SDS0.12416.8640.8410.0240.875
SV3Dp0.11917.4050.8490.0210.877
", + "bbox": [ + 215, + 215, + 563, + 347 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/40fb7a1ee3ca221ed980f517d91fb7c173407d73f26e8271bc4469cc7bf8ba14.jpg", + "table_caption": [ + "Table 6: Comparison of 3D metrics on the GSO dataset. Our models perform favorably against prior works." + ], + "table_footnote": [], + "table_body": "
ModelCD↓3D IoU↑
GT renders0.0210.689
Point-E [34]0.0740.162
Shap-E [16]0.0710.267
DreamGaussian [50]0.0550.411
One-2-3-45++ [24]0.0540.406
SyncDreamer [27]0.0530.451
EscherNet [21]0.0420.466
Free3D [64]0.0470.426
Stable Zero123 [49]0.0390.550
SV3Du0.0270.589
SV3Dc0.0270.584
SV3Dp static orbit0.0280.610
SV3Dp no SDS0.0240.611
SV3Dp0.0240.614
", + "bbox": [ + 566, + 202, + 782, + 383 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "tent in 3D. On the other hand, $\\mathrm{SV3D}^u$ shares the same disadvantage as several prior works in that it can only generate views at the same elevation, which is insufficient to build a legible 3D object, as shown in Fig. 8. This also leads to the worse performance of 'SV3D $^p$ with static orbit' in Tabs. 5 and 6. Overall, $\\mathrm{SV3D}^p$ with dynamic orbit and masked SDS loss performs favorably against all other configurations since it can leverage more diverse views of the object.", + "bbox": [ + 212, + 388, + 785, + 479 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations. Our SV3D model is by design only capable of handling 2 degrees of freedom: elevation and azimuth; which is usually sufficient for 3D generation from a single image. One may want to tackle more degrees of freedom in cameras for a generalized NVS system, which forms an interesting future work. We also notice that SV3D exhibits some view inconsistency for mirror-like reflective surfaces, which provide a challenge to 3D reconstruction. Lastly, such reflective surfaces are not representable by our Lambertian reflection-based shading model. Conditioning SV3D on the full camera matrix, and extending the shading model are interesting directions for future research.", + "bbox": [ + 212, + 481, + 787, + 617 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 641, + 359, + 657 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We present SV3D, a latent video diffusion model for novel multi-view synthesis and 3D generation. In addition to leveraging the generalizability and view-consistent prior in SVD, SV3D enables controllability via camera pose conditioning, and generates orbital videos of an object at high resolution on arbitrary camera orbits. We further propose several techniques for improved 3D generation from SV3D, including triangle CFG scaling, disentangled illumination, and masked SDS loss. We conduct extensive experiments to show that SV3D is controllable, multi-view consistent, and generalizable to the real-world, achieving state-of-the-art performance on novel multi-view synthesis and 3D generation. We believe SV3D provides a solid foundation model for further research on 3D object generation. We plan to publicly release SV3D models.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "V. Voleti et al.", + "bbox": [ + 271, + 114, + 370, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 145, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "We thank Emad Mostaque, Bryce Wilson, Anel Islamovic, Savannah Martin, Ana Guillen, Josephine Parquet, Adam Chen and Ella Irwin for their help in various aspects of the model development and release. We thank Kyle Lacey and Yan Marek for their help with demos. We also thank Eric Courtemanche for his help with visual results.", + "bbox": [ + 212, + 178, + 787, + 253 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 277, + 321, + 292 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-NeRF: A multiscale representation for anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5855-5864 (2021)", + "2. Blattmann, A., Dockhorn, T., Kulal, S., Mendelevitch, D., Kilian, M., Lorenz, D., Levi, Y., English, Z., Voleti, V., Letts, A., et al.: Stable Video Diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127 (2023)", + "3. Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S.W., Fidler, S., Kreis, K.: Align your Latents: High-Resolution Video Synthesis with Latent Diffusion Models. arXiv:2304.08818 (2023)", + "4. Boss, M., Braun, R., Jampani, V., Barron, J.T., Liu, C., Lensch, H.P.: NeRD: Neural reflectance decomposition from image collections. In: IEEE International Conference on Computer Vision (ICCV) (2021)", + "5. Boss, M., Engelhardt, A., Kar, A., Li, Y., Sun, D., Barron, J.T., Lensch, H.P., Jampani, V.: SAMURAI: Shape And Material from Unconstrained Real-world Arbitrary Image collections. In: Advances in Neural Information Processing Systems (NeurIPS) (2022)", + "6. Chan, E.R., Nagano, K., Chan, M.A., Bergman, A.W., Park, J.J., Levy, A., Aittala, M., Mello, S.D., Karras, T., Wetzstein, G.: GeNVS: Generative novel view synthesis with 3D-aware diffusion models. International Conference on Computer Vision (2023)", + "7. Cline, D.B.: Admissible kernel estimators of a multivariate density. The Annals of Statistics 16(4), 1421-1427 (1988)", + "8. Deitke, M., Liu, R., Wallingford, M., Ngo, H., Michel, O., Kusupati, A., Fan, A., Laforte, C., Voleti, V., Gadre, S.Y., et al.: Objverse-XL: A universe of $10\\mathrm{m} + 3\\mathrm{D}$ objects. arXiv preprint arXiv:2307.05663 (2023)", + "9. Deitke, M., Schwenk, D., Salvador, J., Weihs, L., Michel, O., VanderBilt, E., Schmidt, L., Ehsani, K., Kembhavi, A., Farhadi, A.: Objaverse: A universe of annotated 3D objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13142-13153 (2023)", + "10. Downs, L., Francis, A., Koenig, N., Kinman, B., Hickman, R., Reymann, K., McHugh, T.B., Vanhoucke, V.: Google Scanned Objects: A high-quality dataset of 3D scanned household items. In: 2022 International Conference on Robotics and Automation (ICRA). pp. 2553-2560. IEEE (2022)", + "11. Eftekhar, A., Sax, A., Bachmann, R., Malik, J., Zamir, A.: Omnidata: A scalable pipeline for making multi-task mid-level vision datasets from 3D scans. In: IEEE International Conference on Computer Vision (ICCV) (2021)" + ], + "bbox": [ + 218, + 310, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "SV3D", + "bbox": [ + 689, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "12. Girdhar, R., Singh, M., Brown, A., Duval, Q., Azadi, S., Rambhatla, S.S., Shah, A., Yin, X., Parikh, D., Misra, I.: EMU VIDEO: Factorizing Text-to-Video Generation by Explicit Image Conditioning (2023), https://emu-video.metademolab.com/ assets/emu(video.pdf", + "13. Hasselgren, J., Hofmann, N., Munkberg, J.: Shape, Light, and Material Decomposition from Images using Monte Carlo Rendering and Denoising. In: Advances in Neural Information Processing Systems (NeurIPS) (2022)", + "14. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: Advances in Neural Information Processing Systems (2020)", + "15. Hong, Y., Zhang, K., Gu, J., Bi, S., Zhou, Y., Liu, D., Liu, F., Sunkavalli, K., Bui, T., Tan, H.: LRM: Large reconstruction model for single image to 3D. International Conference on Learning Representations (2024)", + "16. Jun, H., Nichol, A.: Shap-E: Generating conditional 3D implicit functions (2023)", + "17. Kanopoulos, N., Vasanthavada, N., Baker, R.L.: Design of an image edge detection filter using the sobel operator. IEEE Journal of solid-state circuits 23(2), 358-367 (1988)", + "18. Kant, Y., Wu, Z., Vasilkovsky, M., Qian, G., Ren, J., Guler, R.A., Ghanem, B., Tulyakov, S., Gilitschenski, I., Siarohin, A.: Spad: Spatially aware multiview diffusers. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2024)", + "19. Karras, T., Aittala, M., Aila, T., Laine, S.: Elucidating the Design Space of Diffusion-Based Generative Models. arXiv:2206.00364 (2022)", + "20. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)", + "21. Kong, X., Liu, S., Lyu, X., Taher, M., Qi, X., Davison, A.J.: Eschernet: A generative model for scalable view synthesis. arXiv preprint arXiv:2402.03908 (2024)", + "22. Kwak, J.g., Dong, E., Jin, Y., Ko, H., Mahajan, S., Yi, K.M.: Vivid-1-to-3: Novel view synthesis with video diffusion models. IEEE conference on computer vision and pattern recognition (2024)", + "23. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3D: High-resolution text-to-3D content creation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 300–309 (2023)", + "24. Liu, M., Shi, R., Chen, L., Zhang, Z., Xu, C., Wei, X., Chen, H., Zeng, C., Gu, J., Su, H.: One-2-3-45++: Fast single image to 3D objects with consistent multi-view generation and 3D diffusion. arXiv preprint arXiv:2311.07885 (2023)", + "25. Liu, M., Xu, C., Jin, H., Chen, L., Varma T, M., Xu, Z., Su, H.: One-2-3-45: Any single image to 3D mesh in 45 seconds without per-shape optimization. Advances in Neural Information Processing Systems 36 (2023)", + "26. Liu, R., Wu, R., Hoorick, B.V., Tokmakov, P., Zakharov, S., Vondrick, C.: Zero-1-to-3: Zero-shot one image to 3D object. International Conference on Computer Vision (2023)", + "27. Liu, Y., Lin, C., Zeng, Z., Long, X., Liu, L., Komura, T., Wang, W.: SyncDreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453 (2023)", + "28. Long, X., Guo, Y.C., Lin, C., Liu, Y., Dou, Z., Liu, L., Ma, Y., Zhang, S.H., Habermann, M., Theobalt, C., et al.: Wonder3D: Single image to 3D using cross-domain diffusion. arXiv preprint arXiv:2310.15008 (2023)", + "29. Melas-Kyriazi, L., Laina, I., Rupprecht, C., Neverova, N., Vedaldi, A., Gafni, O., Kokkinos, F.: IM-3D: Iterative multiview diffusion and reconstruction for high-quality 3D generation. arXiv preprint arXiv:2402.08682 (2024)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "V. Voleti et al.", + "bbox": [ + 271, + 114, + 370, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "30. Mercier, A., Nakhli, R., Reddy, M., Yasarla, R., Cai, H., Porikli, F., Berger, G.: HexaGen3D: Stablediffusion is just one step away from fast and diverse Text-to-3D generation. arXiv preprint arXiv:2401.07727 (2024)", + "31. Metzer, G., Richardson, E., Patashnik, O., Giryes, R., Cohen-Or, D.: Latent-NeRF for shape-guided generation of 3D shapes and textures. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12663-12673 (2023)", + "32. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing scenes as neural radiance fields for view synthesis. In: European Conference on Computer Vision. pp. 405-421 (2020)", + "33. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph. (2022)", + "34. Nichol, A., Jun, H., Dhariwal, P., Mishkin, P., Chen, M.: Point-E: A System for Generating 3D Point Clouds from Complex Prompts (2022)", + "35. Niemeyer, M., Barron, J.T., Mildenhall, B., Sajjadi, M.S.M., Geiger, A., Radwan, N.: RegNeRF: Regularizing neural radiance fields for view synthesis from sparse inputs. In: Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR) (2022)", + "36. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023)", + "37. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3D using 2D diffusion. arXiv (2022)", + "38. Qian, G., Mai, J., Hamdi, A., Ren, J., Siarohin, A., Li, B., Lee, H.Y., Skorokhodov, I., Wonka, P., Tulyakov, S., et al.: Magic123: One image to high-quality 3D object generation using both 2D and 3D diffusion priors. arXiv preprint arXiv:2306.17843 (2023)", + "39. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)", + "40. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)", + "41. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dream-Booth: Fine tuning text-to-image dissusion models for subject-driven generation. arXiv preprint arXiv:2208.12242 (2022)", + "42. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E., Kamyar Seyed Ghasemipour, S., Karagol Ayan, B., Mahdavi, S.S., Gontijo Lopes, R., Salimans, T., Ho, J., Fleet, D.J., Norouzi, M.: Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding. arXiv:2205.11487 (2022)", + "43. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortzman, M., et al.: LAION-5B: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022)", + "44. Shen, T., Gao, J., Yin, K., Liu, M.Y., Fidler, S.: Deep Marching Tetrahedra: a hybrid representation for high-resolution 3D shape synthesis. In: Advances in Neural Information Processing Systems (NeurIPS) (2021)", + "45. Shi, R., Chen, H., Zhang, Z., Liu, M., Xu, C., Wei, X., Chen, L., Zeng, C., Su, H.: Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110 (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "SV3D", + "bbox": [ + 689, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "46. Shi, Y., Wang, P., Ye, J., Long, M., Li, K., Yang, X.: MVDream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512 (2023)", + "47. Song, Y., Ermon, S.: Improved Techniques for Training Score-Based Generative Models. arXiv:2006.09011 (2020)", + "48. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456 (2020)", + "49. StabilityAI: Stable Zero123 (2023), https://stability.ai/news/stable-zero123-3d-generation", + "50. Tang, J., Ren, J., Zhou, H., Liu, Z., Zeng, G.: DreamGaussian: Generative Gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653 (2023)", + "51. Tang, S., Chen, J., Wang, D., Tang, C., Zhang, F., Fan, Y., Chandra, V., Furukawa, Y., Ranjan, R.: Mvdiffusion++: A dense high-resolution multi-view diffusion model for single or sparse-view 3d object reconstruction. arXiv preprint arXiv:2402.12712 (2024)", + "52. Tsai, Y.T., Shih, Z.C.: All-frequency precomputed radiance transfer using spherical radial basis functions and clustered tensor approximation. ACM Transactions on Graphics (ToG) (2006)", + "53. Voleti, V., Jolicoeur-Martineau, A., Pal, C.: MCVD: Masked conditional video diffusion for prediction, generation, and interpolation. In: (NeurIPS) Advances in Neural Information Processing Systems (2022)", + "54. Watson, D., Chan, W., Martin-Brualla, R., Ho, J., Tagliasacchi, A., Norouzi, M.: Novel view synthesis with diffusion models (2022)", + "55. Weng, H., Yang, T., Wang, J., Li, Y., Zhang, T., Chen, C.L.P., Zhang, L.: Consistent123: Improve consistency for one image to 3D object synthesis (2023)", + "56. Wu, T., Zhang, J., Fu, X., Wang, Y., Ren, J., Pan, L., Wu, W., Yang, L., Wang, J., Qian, C., et al.: Omniobject3D: Large-vocabulary 3D object dataset for realistic perception, reconstruction and generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 803-814 (2023)", + "57. Yang, J., Cheng, Z., Duan, Y., Ji, P., Li, H.: ConsistNet: Enforcing 3D consistency for multi-view images diffusion. arXiv preprint arXiv:2310.10343 (2023)", + "58. Yao, C.H., Raj, A., Hung, W.C., Rubinstein, M., Li, Y., Yang, M.H., Jampani, V.: ARTIC3D: Learning robust articulated 3D shapes from noisy web image collections. Advances in Neural Information Processing Systems 36 (2024)", + "59. Ye, J., Wang, P., Li, K., Shi, Y., Wang, H.: Consistent-1-to-3: Consistent image to 3D view synthesis via geometry-aware diffusion models. In: 3DV (2024)", + "60. Young, J.: xatlas. https://github.com/jpcy/xatlas (2024)", + "61. Yu, Z., Peng, S., Niemeyer, M., Sattler, T., Geiger, A.: MonoSDF: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in Neural Information Processing Systems (NeurIPS) (2022)", + "62. Zhang, K., Luan, F., Wang, Q., Bala, K., Snavely, N.: PhySG: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In: The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021)", + "63. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)", + "64. Zheng, C., Vedaldi, A.: Free3D: Consistent novel view synthesis without 3D representation. arXiv preprint arXiv:2312.04551 (2023)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "V. Voleti et al.", + "bbox": [ + 271, + 114, + 370, + 126 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "In this supplemental document, we include the broader impact in Appendix A, implementation details in Appendices B to F, ablative analyses in Appendix G, and additional results in Appendix H. We also attach a teaser video summarizing the SV3D framework (SV3D(video.mp4), as well as an HTML page (SV3D.html) for more visual comparisons.", + "bbox": [ + 212, + 146, + 787, + 220 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "A Broader Impact", + "text_level": 1, + "bbox": [ + 215, + 246, + 413, + 266 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "The advancement of generative models in different media forms is changing how we make and use content. These AI-powered models can create images, videos, 3D objects, and more, in ways we've never seen before. They offer huge potential for innovation in media production. But along with this potential, there are also risks. Before we start using these models widely, it is crucial to make sure we understand the possible downsides and have plans in place to deal with them effectively.", + "bbox": [ + 212, + 279, + 787, + 383 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "In the case of 3D object generation, the input provided by the user plays a crucial role in constraining the model's output. By supplying a full front view of an object, users limit the model's creative freedom to the visible or unoccluded regions, thus minimizing the potential for generating problematic imagery. Additionally, factors such as predicted depth values and lighting further influence the fidelity and realism of generated content.", + "bbox": [ + 212, + 386, + 787, + 474 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "Moreover, ensuring the integrity and appropriateness of training data is critical in mitigating risks associated with generative models. Platforms like Sketchfab, which serve as repositories for 3D models used in training, enforce strict content policies to prevent the dissemination of \"Unacceptable Content\" and disallows it on their platform: https://help.sketchfab.com/hc/en-us/articles/214867883-What-is-Restricted-Content. By adhering to these guidelines and actively monitoring dataset quality, developers can reduce the likelihood of biased or inappropriate outputs.", + "bbox": [ + 212, + 477, + 802, + 595 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "Sketchfab also has a tag for \"Restricted Content\" which is deemed to be similar to the PG-13 content rating used in the US (i.e. inappropriate for children under 13). We have confirmed that none of the objects that we use in training have this flag set to true. Thus we go the extra step of excluding even tagged PG-13 content from the training set.", + "bbox": [ + 212, + 598, + 787, + 672 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "There is a chance that certain content may not have been correctly labeled on Sketchfab. In cases where the uploader fails to tag an object appropriately, Sketchfab provides publicly accessible listings of objects and a mechanism for the community to report any content that may be deemed offensive.", + "bbox": [ + 212, + 674, + 787, + 734 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "In our regular utilization of Objaverse content, we haven't observed any significant amount of questionable material. Nevertheless, there are occasional instances of doll-like nudity stemming from basic 3D models, which could be crucial for accurately depicting humanoid anatomy. Additionally, the training dataset contains some presence of drugs, drug paraphernalia, as well as a certain level of blood content and weaponry, resembling what might be encountered in a video game context. Should the model be provided with imagery featuring these cate-", + "bbox": [ + 212, + 734, + 787, + 839 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "SV3D", + "bbox": [ + 687, + 114, + 730, + 126 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "gories of content, it possesses the capability to generate corresponding 3D models to some extent.", + "bbox": [ + 212, + 146, + 782, + 175 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "It is to be noted that SV3D mainly focuses on generating hidden details in the user's input image. If the image is unclear or some parts are hidden, the model guesses what those parts might look like based on its training data. This means it might create details similar to what it has seen before. The training data generally follows the standards of 3D modeling and gaming. However, this could lead to criticisms about the models being too similar to existing trends. But the user's input image limits how creative the model can be and reduces the chance of biases showing up in its creations, especially if the image is clear and straightforward.", + "bbox": [ + 212, + 180, + 787, + 316 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "B Data Details", + "text_level": 1, + "bbox": [ + 215, + 354, + 380, + 372 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "Similar to SVD-MV [2], we render views of a subset of the Objaverse dataset [9] consisting of 150K curated CC-licensed 3D objects from the original dataset. Each loaded object is scaled such that the largest world-space XYZ extent of its bounding box is 1. The object is then repositioned such that this bounding box is centered around the origin.", + "bbox": [ + 212, + 402, + 782, + 479 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "For both the static and dynamic orbits, we use Blender's EEVEE renderer to render an 84-frame RGBA orbit at $576 \\times 576$ resolution. During training, any 21-frame orbit can be subsampled from this by picking any frame as the first frame, and then choosing every 4th frame after that.", + "bbox": [ + 212, + 482, + 784, + 542 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "We apply two background colors to each of these images: random RGB color, and white. This results in a doubling of the number of orbit samples for training. We then encode all of these images into latent space using SVD's VAE, and using CLIP. We then store the latent and CLIP embeddings for all of these images along with the corresponding elevation and azimuth values.", + "bbox": [ + 212, + 546, + 787, + 622 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "For lighting, we randomly select from a set of 20 curated HDRI envmaps. Each orbit begins with the camera positioned at azimuth 0. Our camera uses a field-of-view of 33.8 degrees. For each object, we adaptively position the camera to a distance sufficient to ensure that the rendered object content makes good and consistent use of the image extents without being clipped in any view.", + "bbox": [ + 212, + 625, + 787, + 700 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "For static orbits, the camera is positioned at a randomly sampled elevation between $[-5, 30]$ degrees. The azimuth steps by a constant $\\frac{360}{84}$ degree delta between each frame. For dynamic orbits, the sequence of camera elevations for each orbit are obtained from a random weighted combination of sinusoids with different frequencies. For each sinusoid, the whole number period is sampled from [1, 5], the amplitude is sampled from [0.5, 10], and a random phase shift is applied. The azimuth angles are sampled regularly, and then a small amount of noise is added to make them irregular. The elevation values are smoothed using a simple convolution kernel and then clamped to a maximum elevation of 89 degrees.", + "bbox": [ + 212, + 704, + 787, + 839 + ], + "page_idx": 19 + }, + { + "type": "page_number", + "text": "20", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 19 + }, + { + "type": "header", + "text": "V. Voleti et al.", + "bbox": [ + 271, + 114, + 370, + 126 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "C Training Details", + "text_level": 1, + "bbox": [ + 215, + 143, + 413, + 162 + ], + "page_idx": 20 + }, + { + "type": "text", + "text": "Our approach involves utilizing the widely used EDM [19] framework, incorporating a simplified diffusion loss for fine-tuning, as followed in SVD [2]. We eliminate the conditions of 'fps_id', 'motion_bucket_id', etc. since they are irrelevant for SV3D. Furthermore, we adjust the loss function to assign lower weights to frames closer to the front-view conditioning image, ensuring that challenging back views receive equal training focus as the easier front views. To optimize training efficiency and conserve GPU VRAM, we preprocess and store the precomputed latent and CLIP embeddings for all video frames in advance. During training, these tensors are directly loaded rather than being computed in real-time. We choose to finetune the SVD-xt model to output 21 frames instead of 25 frames. We found that with 21 frames we were able to fit a batch size of 2 on each GPU, instead of 1 with 25 frames at $576 \\times 576$ resolution. All SV3D models are trained for 105k iterations with an effective batch size of 64 on 4 nodes of 880GB A100 GPUs for around 6 days.", + "bbox": [ + 212, + 175, + 787, + 387 + ], + "page_idx": 20 + }, + { + "type": "text", + "text": "D Inference Details", + "text_level": 1, + "bbox": [ + 215, + 409, + 421, + 424 + ], + "page_idx": 20 + }, + { + "type": "text", + "text": "To generate an orbital video during inference, we use 50 steps of the deterministic DDIM sampler [47] with the triangular classifier-free guidance (CFG) scale described in the main paper. This takes $\\approx 40$ seconds for the SV3D model.", + "bbox": [ + 212, + 439, + 785, + 484 + ], + "page_idx": 20 + }, + { + "type": "text", + "text": "E Additional Details on Illumination Model", + "text_level": 1, + "bbox": [ + 215, + 507, + 656, + 523 + ], + "page_idx": 20 + }, + { + "type": "text", + "text": "We base our rendering model on Spherical Gaussians (SG) [4,62]. A SG at query location $\\pmb{x} \\in \\mathbb{R}^3$ is defined as $G(\\pmb{x};\\pmb{\\mu},c,a) = ae^{s(\\pmb{\\mu}\\cdot \\pmb{x} - 1)}$ , where $\\pmb{\\mu} \\in \\mathbb{R}^3$ is the axis, $s \\in \\mathbb{R}$ the sharpness of the lobe, and $a \\in \\mathbb{R}$ the amplitude. Here, we point out that we only model white light and hence only use a scalar amplitude. One particularly interesting property of SGs is that the inner product of two SGs is the integral of the product of two SGs. The operation is defined as [52]:", + "bbox": [ + 212, + 537, + 787, + 630 + ], + "page_idx": 20 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} G _ {1} (\\boldsymbol {x}) \\cdot G _ {2} (\\boldsymbol {x}) = \\int_ {\\Omega} G _ {1} (\\boldsymbol {x}) G _ {2} (\\boldsymbol {x}) d \\boldsymbol {x} = \\frac {1}{d _ {m}} \\left(2 \\pi a _ {1} a _ {2} e ^ {d _ {m} - \\lambda_ {m}} (1. 0 - e ^ {- 2 d _ {m}})\\right) \\\\ \\lambda_ {m} = \\lambda_ {1} - \\lambda_ {2} \\tag {1} \\\\ d _ {m} = \\left| \\left| \\lambda_ {1} \\boldsymbol {\\mu} _ {1} + \\lambda_ {2} \\boldsymbol {\\mu} _ {2} \\right| \\right|. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 243, + 638, + 784, + 709 + ], + "page_idx": 20 + }, + { + "type": "text", + "text": "In our illumination model we only consider Lambertian shading. Here, the cosine shading term influences the output the most. This term can be approximated with another SG $G_{\\mathrm{cosine}} = (\\pmb{x}; \\pmb{n}, 2.133, 1.17)$ , where $\\pmb{n}$ defines the surface normal at $\\pmb{x}$ . The lighting evaluation using SGs $G_i$ then becomes: $L = \\sum_{i=1}^{24} \\frac{1}{\\pi} \\max(G_i \\cdot G_{\\mathrm{cosine}}, 0)$ . As defined previously this results in the full integration of incoming light for each SG and as light is additive evaluating and summing all SGs results in the complete environment illumination. This $L$ is also used in the $\\mathcal{L}_{\\mathrm{illum}}$ loss described in the main paper. The rendered textured", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 20 + }, + { + "type": "header", + "text": "SV3D", + "bbox": [ + 687, + 114, + 730, + 126 + ], + "page_idx": 20 + }, + { + "type": "page_number", + "text": "21", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 20 + }, + { + "type": "text", + "text": "image is then defined as $\\hat{I} = c_dL$ , where $c_{d}$ is the diffuse albedo. We learn $\\mu, c, a$ for each SG $G_{i}$ using reconstruction loss between these rendered images and SV3D-generated images.", + "bbox": [ + 215, + 145, + 784, + 191 + ], + "page_idx": 21 + }, + { + "type": "text", + "text": "F Losses and Optimization for 3D Generation", + "text_level": 1, + "bbox": [ + 215, + 214, + 674, + 233 + ], + "page_idx": 21 + }, + { + "type": "text", + "text": "In addition to the masked SDS loss $\\mathcal{L}_{\\mathrm{mask - ds}}$ and illumination loss $\\mathcal{L}_{\\mathrm{illum}}$ detailed in the manuscript, we use several other losses for 3D reconstruction. Our main reconstruction losses are the pixel-level mean squared error $\\mathcal{L}_{\\mathrm{mse}} = \\| I - \\hat{I}\\|^2$ , LPIPS [63] loss $\\mathcal{L}_{\\mathrm{lpips}}$ , and mask loss $\\mathcal{L}_{\\mathrm{mask}} = \\| S - \\hat{S}\\|^2$ , where $S$ , $\\hat{S}$ are the predicted and ground-truth masks. We further employ a normal loss using the estimated mono normal by Omnidata [11], which is defined as the cosine similarity between the rendered normal $\\pmb{n}$ and estimated pseudo ground truths $\\bar{n}$ : $\\mathcal{L}_{\\mathrm{normal}} = 1 - (\\pmb{n} \\cdot \\bar{\\pmb{n}})$ . To regularize the output geometry, we apply a smooth depth loss inspired by RegNeRF [35]: $\\mathcal{L}_{\\mathrm{depth}}(i,j) = (d(i,j) - d(i,j + 1))^2 + (d(i,j) - (d(i + 1,j))^2$ , where $i,j$ indicate the pixel coordinate. For surface normal we instead rely on a bilateral smoothness loss similar to [5]. We found that this is crucial to getting high-frequency details and avoiding over-smoothed surfaces. For this loss we compute the image gradients of the input image $\\nabla I$ with a Sobel filter [17]. We then encourage the gradients of rendered normal $\\nabla n$ to be smooth if (and only if) the input image gradients $\\nabla I$ are smooth. The loss can be written as $\\mathcal{L}_{\\mathrm{bilateral}} = e^{-3\\nabla I}\\sqrt{1 + ||\\nabla n||}$ . We also found that adding a spatial smoothness regularization on the albedo is beneficial: $\\mathcal{L}_{\\mathrm{albedo}} = (c_d(x) - c_d(x + \\epsilon))^2$ , where $c_d$ denotes the albedo, $x$ is a 3D surface point, and $\\epsilon \\in \\mathbb{R}^3$ is a normal distributed offset with a scale of 0.01. The overall objective is then defined as the weighted sum of these losses. All losses are applied in both coarse and fine stages, except that we only apply $\\mathcal{L}_{\\mathrm{mask - ds}}$ in the last 200 iterations of the fine stage. We use an Adam optimizer [20] with a learning rate of 0.01 for both stages.", + "bbox": [ + 215, + 247, + 787, + 599 + ], + "page_idx": 21 + }, + { + "type": "text", + "text": "G Additional Ablative Analyses", + "text_level": 1, + "bbox": [ + 215, + 619, + 542, + 638 + ], + "page_idx": 21 + }, + { + "type": "text", + "text": "We conduct additional ablative analyses of our 3D generation pipeline in this section.", + "bbox": [ + 215, + 652, + 784, + 681 + ], + "page_idx": 21 + }, + { + "type": "text", + "text": "G.1 SV3D Models", + "text_level": 1, + "bbox": [ + 215, + 705, + 382, + 720 + ], + "page_idx": 21 + }, + { + "type": "text", + "text": "In Tab. 7, we compare the quantitative results using different SV3D models and training losses. Both 2D and 3D evaluation shows that $\\mathrm{SV3D}^p$ is our best performing model, either for pure photometric reconstruction or SDS-based optimization.", + "bbox": [ + 215, + 729, + 784, + 790 + ], + "page_idx": 21 + }, + { + "type": "page_number", + "text": "22", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 21 + }, + { + "type": "header", + "text": "V. Voleti et al.", + "bbox": [ + 271, + 114, + 370, + 126 + ], + "page_idx": 21 + }, + { + "type": "table", + "img_path": "images/ad805a2478f4b5ea09d2b1624fb755c74d233a1b169793a259bff8ce45983232.jpg", + "table_caption": [ + "Table 7: Ablative results of different SV3D models and training losses. We show that our $\\mathrm{SV3D}^p$ model with Photo+SDS losses achieves the best 2D and 3D metrics." + ], + "table_footnote": [], + "table_body": "
ModelTraining lossesLPIPS↓PSNR↑SSIM↑MSE↓CLIP-S↑CD↓3D IoU↑
SV3DuPhoto0.13215.9510.8270.0320.8730.0280.583
SV3DuPhoto+SDS0.13315.9570.8340.0310.8710.0270.589
SV3DcPhoto0.13515.8260.8320.0330.8710.0290.579
SV3DcPhoto+SDS0.13216.3730.8340.0270.8700.0270.584
SV3DpPhoto0.12416.8640.8410.0240.8750.0240.611
SV3DpPhoto+SDS0.11917.4050.8490.0210.8770.0240.614
", + "bbox": [ + 238, + 468, + 763, + 566 + ], + "page_idx": 22 + }, + { + "type": "header", + "text": "SV3D", + "bbox": [ + 689, + 114, + 730, + 126 + ], + "page_idx": 22 + }, + { + "type": "page_number", + "text": "23", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 22 + }, + { + "type": "text", + "text": "G.2 Static v.s. Dynamic Orbits", + "text_level": 1, + "bbox": [ + 215, + 146, + 488, + 162 + ], + "page_idx": 23 + }, + { + "type": "text", + "text": "We also compare the results using different camera orbits for 3D training in Tab. 8. The results show that using a dynamic orbit (sine-30) produces better 3D outputs compared to static orbit since it contains more information of the top and bottom views of the object. However, higher elevation (sine-50) tends to increase inconsistency between multi-view images, and thus resulting in worse 3D reconstruction. In our experiments, we find that setting the elevation within $\\pm 30$ degree generally leads to desirable 3D outputs.", + "bbox": [ + 212, + 171, + 787, + 279 + ], + "page_idx": 23 + }, + { + "type": "table", + "img_path": "images/8082049d049fcbad08c3ec24b92d6dd5d6a2e32e05575f93ac076f9d3ca32f64.jpg", + "table_caption": [ + "Table 8: Ablative results of different reference orbits for 3D generation. We show that using a dynamic orbit (sine elevation) with moderate amplitude performs better than orbits with no or extreme elevation variations." + ], + "table_footnote": [], + "table_body": "
Training orbitLPIPS↓PSNR↑SSIM↑MSE↓CLIP-S↑CD↓3D IoU↑
Static0.12516.8210.8480.0250.8640.0280.610
Sine-300.11917.4050.8490.0210.8770.0240.614
Sine-500.12317.0570.8540.0250.8730.0260.609
", + "bbox": [ + 264, + 358, + 738, + 416 + ], + "page_idx": 23 + }, + { + "type": "text", + "text": "G.3 Masked SDS Loss", + "text_level": 1, + "bbox": [ + 215, + 468, + 415, + 483 + ], + "page_idx": 23 + }, + { + "type": "text", + "text": "Finally, we show the ablative results of our SDS loss in Tab. 9. We compare the results of 1) pure photometric losses, 2) with naive SDS loss (no masking), 3) with hard-masked SDS loss by thresholding the dot product of surface normal and camera viewing angle as visibility masks, and 4) with soft-masked SDS loss as described in the manuscript. Overall, adding SDS guidance from the SV3D model can improve the 2D metrics while maintaining similar 3D metrics. Our novel soft-masked SDS loss generally achieves the best results compared to other baselines.", + "bbox": [ + 212, + 493, + 787, + 616 + ], + "page_idx": 23 + }, + { + "type": "table", + "img_path": "images/5b61eb72cf30106f67894316113b6ee762b478c32bb8fe21a5f235dbf770e850.jpg", + "table_caption": [ + "Table 9: Ablative analyses of Masked SDS loss. Overall, our soft-masked SDS loss leads to higher-quality mesh outputs in terms of most 2D and 3D metrics." + ], + "table_footnote": [], + "table_body": "
Training lossesLPIPS↓PSNR↑SSIM↑MSE↓CLIP-S↑CD↓3D IoU↑
Photo0.12416.8640.8410.0240.8750.0240.611
Photo+SDS (naive)0.12417.0070.8500.0240.8670.0250.615
Photo+SDS (hard masked)0.12417.3350.8450.0220.8770.0240.610
Photo+SDS (soft masked)0.11917.4050.8490.0210.8770.0240.614
", + "bbox": [ + 225, + 678, + 781, + 750 + ], + "page_idx": 23 + }, + { + "type": "page_number", + "text": "24", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 23 + }, + { + "type": "header", + "text": "V. Voleti et al.", + "bbox": [ + 271, + 114, + 370, + 127 + ], + "page_idx": 23 + }, + { + "type": "text", + "text": "H Additional Visual Results", + "text_level": 1, + "bbox": [ + 215, + 143, + 504, + 160 + ], + "page_idx": 24 + }, + { + "type": "text", + "text": "In this section, we show more results of novel view synthesis and 3D generation.", + "bbox": [ + 214, + 176, + 782, + 191 + ], + "page_idx": 24 + }, + { + "type": "text", + "text": "H.1 Novel View Synthesis", + "text_level": 1, + "bbox": [ + 215, + 214, + 444, + 229 + ], + "page_idx": 24 + }, + { + "type": "text", + "text": "We show the additional NVS results on OmniObject3D [56] and real-world images in Fig. 13 and Fig. 14, respectively. The generated novel multi-view images by SV3D are more detailed and consistent compared to prior state-of-the-arts.", + "bbox": [ + 214, + 239, + 784, + 285 + ], + "page_idx": 24 + }, + { + "type": "image", + "img_path": "images/105b9af9adeba66108cfd77e8c2dad8e2bfee731f010f0f8fba23fe2b687e7ca.jpg", + "image_caption": [ + "Fig. 13: NVS Results on OmniObject3D [56]. For each object, we show the input image and generated novel multi-view images by different methods. SV3D is able to produce images with consistent pose, color, and geometric details, which are closer the ground-truth renders." + ], + "image_footnote": [], + "bbox": [ + 225, + 310, + 781, + 612 + ], + "page_idx": 24 + }, + { + "type": "header", + "text": "SV3D", + "bbox": [ + 689, + 114, + 730, + 126 + ], + "page_idx": 24 + }, + { + "type": "page_number", + "text": "25", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 24 + }, + { + "type": "image", + "img_path": "images/f8dfcc05bba1ccf34bc9c1286b08bcab5f3ff6351a6d67286b56a7fd0387b3d3.jpg", + "image_caption": [ + "Fig. 14: NVS Results on Real-World Images. Notice the consistent texture, geometry, and pose in SV3D NVS outputs compared to prior works." + ], + "image_footnote": [], + "bbox": [ + 218, + 340, + 787, + 597 + ], + "page_idx": 25 + }, + { + "type": "page_number", + "text": "26", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 25 + }, + { + "type": "header", + "text": "V. Voleti et al.", + "bbox": [ + 271, + 114, + 370, + 126 + ], + "page_idx": 25 + }, + { + "type": "text", + "text": "H.2 3D Generation", + "text_level": 1, + "bbox": [ + 215, + 146, + 390, + 159 + ], + "page_idx": 26 + }, + { + "type": "text", + "text": "We show the additional 3D generation results on OmniObject3D [56] and real-world images in Fig. 15 and Fig. 16, respectively. Thanks to the consistent multi-view images by SV3D and the novel Masked SDS loss, our 3D generations are detailed, high-fidelity, and generalizable to a wide range of images. Since Free3D [64] does not include a 3D generation method, we run our 3D pipeline on its generated multi-view images for fair 3D comparison.", + "bbox": [ + 212, + 171, + 787, + 263 + ], + "page_idx": 26 + }, + { + "type": "image", + "img_path": "images/2636ffd15e777637be3807b2fddf85d618586e9e02154032657a0d1a01a83753.jpg", + "image_caption": [ + "Fig. 15: Mesh Results on OmniObject3D [56]. Thanks to the consistent novel multi-view images generated by SV3D, our 3D meshes contain higher-fidelity details while having smooth surfaces." + ], + "image_footnote": [], + "bbox": [ + 228, + 292, + 741, + 612 + ], + "page_idx": 26 + }, + { + "type": "header", + "text": "SV3D", + "bbox": [ + 687, + 114, + 730, + 126 + ], + "page_idx": 26 + }, + { + "type": "page_number", + "text": "27", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 26 + }, + { + "type": "image", + "img_path": "images/bf21c4f7501d5718f0681146204e4d1c72137712b15b642c356d72f3ef9496ed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 230, + 315, + 285, + 362 + ], + "page_idx": 27 + }, + { + "type": "image", + "img_path": "images/c1771e6d83010626d04298302ccea97de0b0d513ece8964c10b8356feea725d9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 313, + 316, + 403, + 359 + ], + "page_idx": 27 + }, + { + "type": "image", + "img_path": "images/03831f022ccc6d4c6f863a15122dae0102a2a28bbf2be6daff0cb9f38f20d187.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 313, + 362, + 401, + 405 + ], + "page_idx": 27 + }, + { + "type": "image", + "img_path": "images/af2dfd1a717bcf4ddf323c2bbf1bb4480d6e5871a129128e3ee2fcf03cf92edc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 419, + 316, + 511, + 361 + ], + "page_idx": 27 + }, + { + "type": "image", + "img_path": "images/eeee73d320527f8110c04100e047f38eb14974c6758bee457f90ca4760636cb0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 419, + 362, + 511, + 406 + ], + "page_idx": 27 + }, + { + "type": "image", + "img_path": "images/71e079185eeecd8f2d77f798c1b502cd2f3ef1dd54d119e1221a9a6828c0ac81.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 529, + 316, + 616, + 359 + ], + "page_idx": 27 + }, + { + "type": "image", + "img_path": "images/7fde30b1db64fec167c6874d912d2cd110f9ab3db4a2c032fb20f0ef8b8c5c31.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 529, + 362, + 616, + 405 + ], + "page_idx": 27 + }, + { + "type": "image", + "img_path": "images/0ea317b84ad7a62bb4ea75c19291e24f346adc9422bd4ec03ff6e39e1e28e431.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 643, + 316, + 738, + 361 + ], + "page_idx": 27 + }, + { + "type": "image", + "img_path": "images/8d5feb66007408ebe637e15fd7aa917354d6bb4d45f75e60debeb5b28fd80c3c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 363, + 736, + 406 + ], + "page_idx": 27 + }, + { + "type": "image", + "img_path": "images/fe6da9439b80e8940978a5b41f4348f88fe7962f82209cedc9c52e241736ac2f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 227, + 409, + 290, + 455 + ], + "page_idx": 27 + }, + { + "type": "image", + "img_path": "images/fd5959849467000e71de01bfeed96deeb9a41a7ec08aa22145be45c402ee75a4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 305, + 410, + 398, + 459 + ], + "page_idx": 27 + }, + { + "type": "image", + "img_path": "images/e41cbdd433d6b5194e91c97c77ed3f7e4e292bf429f823e22f507366266aa936.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 307, + 460, + 397, + 507 + ], + "page_idx": 27 + }, + { + "type": "image", + "img_path": "images/410f32136b831e515a6dc9df47293aee849b7916cd829908725ef633fc4b477c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 413, + 411, + 511, + 455 + ], + "page_idx": 27 + }, + { + "type": "image", + "img_path": "images/e13154017e66a59f398233c875015c1797cfdd3ac9edaf8ef53c12d453886c09.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 413, + 460, + 511, + 505 + ], + "page_idx": 27 + }, + { + "type": "image", + "img_path": "images/48d42cdabb13393fda6c62f2c7e40de307427c762285cb6bc96b5a35a5d3f4f4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 524, + 414, + 622, + 454 + ], + "page_idx": 27 + }, + { + "type": "image", + "img_path": "images/3124b38b7ca24277fdb6e853e2e9f9311ec830f2a96da6e9368e860ab62beb47.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 524, + 460, + 622, + 505 + ], + "page_idx": 27 + }, + { + "type": "image", + "img_path": "images/588f6a934c862812ac95926fb295431cd952ff70624031689af9fccc30899b22.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 412, + 736, + 457 + ], + "page_idx": 27 + }, + { + "type": "image", + "img_path": "images/e67dc5694ded9302613190785b2c063fd1018a87fd668b03e1ab3cf43788c63d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 460, + 736, + 506 + ], + "page_idx": 27 + }, + { + "type": "image", + "img_path": "images/3a0124166a356a96f01ea15ab24f3e9dc1744cc495ca07851aa034db2edcf47e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 241, + 513, + 274, + 561 + ], + "page_idx": 27 + }, + { + "type": "image", + "img_path": "images/4e064a625dbad13dfb4f12e3a79fadd81b214212bb524c75cb8d591c38f0ad33.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 313, + 513, + 395, + 561 + ], + "page_idx": 27 + }, + { + "type": "image", + "img_path": "images/700de025b4638fdbff94603b36c34dba71bbeb360015ee50308e6cee7c14b912.jpg", + "image_caption": [ + "SV3D" + ], + "image_footnote": [], + "bbox": [ + 313, + 563, + 393, + 613 + ], + "page_idx": 27 + }, + { + "type": "image", + "img_path": "images/529907a07ac547f4fa93a0765e19e3e36b6c915840f7ddd2df888b49af04ccfc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 423, + 512, + 503, + 561 + ], + "page_idx": 27 + }, + { + "type": "image", + "img_path": "images/b798dde190ce6ac208841607c19bf331c0e7e1a9d23f30ad12ab2de367eebd82.jpg", + "image_caption": [ + "EscherNet" + ], + "image_footnote": [], + "bbox": [ + 423, + 561, + 501, + 613 + ], + "page_idx": 27 + }, + { + "type": "image", + "img_path": "images/7763dd4e82f207b058b1f1912e8bcba795d6ed1018d8ce4ef552ee0709cde2df.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 532, + 511, + 609, + 561 + ], + "page_idx": 27 + }, + { + "type": "image", + "img_path": "images/1874c021568fd7b505d7b6603e370e17ad3353acf7eba1ced4fa9be69dd3d328.jpg", + "image_caption": [ + "Free3D", + "Fig. 16: Mesh Results on Real-World Images. Notice the accurate shape and details in our reconstructions even from the diverse images in-the-wild." + ], + "image_footnote": [], + "bbox": [ + 532, + 561, + 609, + 613 + ], + "page_idx": 27 + }, + { + "type": "image", + "img_path": "images/c73ad4f0f244fbd8d0a7a18df3aa5ebb36d9da8859a1dba445eb4219135f88f9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 513, + 727, + 561 + ], + "page_idx": 27 + }, + { + "type": "image", + "img_path": "images/fe0e95ac726c2866be59983efc71a4d1a978a945b82899035a4dbccd048fdfdd.jpg", + "image_caption": [ + "Stable Zero123" + ], + "image_footnote": [], + "bbox": [ + 635, + 564, + 728, + 612 + ], + "page_idx": 27 + }, + { + "type": "page_number", + "text": "28", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 27 + }, + { + "type": "header", + "text": "V. Voleti et al.", + "bbox": [ + 271, + 114, + 370, + 126 + ], + "page_idx": 27 + }, + { + "type": "image", + "img_path": "images/4f7cd0772d90275f1e9ee7af525094b9c69eeb707919381d4120c5d096e4f68a.jpg", + "image_caption": [ + "Fig. 17: NVS and 3D Generation Results on Real-World Images. Visual results on real-world images from the MVImgNet (Yu et al. CVPR'23) and NAVI (Jampani et al. NeurIPS'22) datasets. Compared to Stable Zero123, SV3D generates more consistent novel multi-view images, thus resulting in higher-quality 3D meshes with smooth surface and detailed texture." + ], + "image_footnote": [], + "bbox": [ + 218, + 382, + 493, + 536 + ], + "page_idx": 28 + }, + { + "type": "image", + "img_path": "images/f2e9a4372043cdf1fb202506c1a54251c142495e9254853a6c6c4900c24288a9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 381, + 784, + 537 + ], + "page_idx": 28 + }, + { + "type": "header", + "text": "SV3D", + "bbox": [ + 689, + 114, + 730, + 126 + ], + "page_idx": 28 + }, + { + "type": "page_number", + "text": "29", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 28 + } +] \ No newline at end of file diff --git a/2024/SV3D_ Novel Multi-view Synthesis and 3D Generation from a Single Image using Latent Video Diffusion/55466408-8e29-4541-9064-bae13a2c8fbf_model.json b/2024/SV3D_ Novel Multi-view Synthesis and 3D Generation from a Single Image using Latent Video Diffusion/55466408-8e29-4541-9064-bae13a2c8fbf_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7413b8061fc9175393a8ac6b8cd4e8dc5dcaa2c1 --- /dev/null +++ b/2024/SV3D_ Novel Multi-view Synthesis and 3D Generation from a Single Image using Latent Video Diffusion/55466408-8e29-4541-9064-bae13a2c8fbf_model.json @@ -0,0 +1,3910 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.245, + 0.141, + 0.761, + 0.207 + ], + "angle": 0, + "content": "SV3D: Novel Multi-view Synthesis and 3D Generation from a Single Image using Latent Video Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.235, + 0.777, + 0.267 + ], + "angle": 0, + "content": "Vikram Voleti*, Chun-Han Yao*, Mark Boss*, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani" + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.277, + 0.77, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.427, + 0.788, + 0.469 + ], + "angle": 0, + "content": "Fig. 1: Stable Video 3D (SV3D). From a single image, SV3D generates consistent novel multi-view images. We then optimize a 3D representation with SV3D-generated views resulting in high-quality 3D meshes." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.486, + 0.741, + 0.68 + ], + "angle": 0, + "content": "Abstract. We present Stable Video 3D (SV3D) — a latent video diffusion model for high-resolution, image-to-multi-view generation of orbital videos around a 3D object. Recent works propose to adapt 2D generative models for novel view synthesis (NVS) and 3D optimization. However, these methods have several disadvantages due to limited views or inconsistent NVS, affecting the performance of 3D object generation. In this work, we propose SV3D that adapts image-to-video diffusion model for novel multi-view synthesis and 3D generation, thereby leveraging the generalization and multi-view consistency of the video models, while further adding explicit camera control for NVS. We also propose improved 3D optimization techniques for image-to-3D generation using SV3D and its NVS outputs. Extensive experiments on multiple datasets with 2D and 3D metrics and user study demonstrate SV3D's state-of-the-art performance on NVS as well as 3D reconstruction compared to prior works." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.693, + 0.714, + 0.708 + ], + "angle": 0, + "content": "Keywords: 3D synthesis \\(\\cdot\\) Video generation \\(\\cdot\\) Novel view synthesis" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.734, + 0.376, + 0.75 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Single-image 3D object reconstruction is a long-standing problem in computer vision with a wide range of applications in game design, AR/VR, e-commerce, robotics, etc. It is a highly challenging and ill-posed problem as it requires lifting 2D pixels to 3D space while also reasoning about the unseen portions of the object in 3D." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.372, + 0.128 + ], + "angle": 0, + "content": "V. Voleti et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.344 + ], + "angle": 0, + "content": "Despite being a long-standing vision problem, it is only recently that practically useful results are being produced by leveraging advances in the generative AI. This is mainly made possible by the large-scale pretraining of generative models which enables sufficient generalization to various domains. A typical strategy is to use image-based 2D generative models (e.g., Imagen [42], Stable Diffusion (SD) [40]) to provide a 3D optimization loss function for unseen novel views of a given object [23,31,37]. In addition, several works repurpose these 2D generative models to perform novel view synthesis (NVS) from a single image [26,27,30,54], and then use the synthesized novel views for 3D generation. Conceptually, these works mimic a typical photogrammetry-based 3D object capture pipeline, i.e., first photographing multi-view images of an object, followed by 3D optimization; except that the explicit multi-view capture is replaced by novel-view synthesis using a generative model, either via text prompt or camera pose control." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.346, + 0.79, + 0.452 + ], + "angle": 0, + "content": "A key issue in these generation-based reconstruction methods is the lack of multi-view consistency in the underlying generative model, resulting in inconsistent novel views. Some works try to address this by jointly reasoning a 3D representation during NVS [6,15,27], but at the cost of high compute and data, often still resulting in inconsistent geometric and texture details. In this work, we tackle this issue by adapting a high-resolution, image-conditioned video diffusion model for NVS followed by 3D generation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.456, + 0.79, + 0.637 + ], + "angle": 0, + "content": "Novel Multi-view Synthesis. We adapt a latent video diffusion model (Stable Video Diffusion - SVD [2]) to generate multiple novel views of a given object with explicit camera pose conditioning. SVD demonstrates excellent multi-view consistency for video generation, and we repurpose it for NVs. In addition, SVD also has good generalization capabilities as it is trained on large-scale image and video data that are more readily available than large-scale 3D data. In short, we adapt the video diffusion model for NVs from a single image with three useful properties for 3D object generation: pose-controllable, multi-view consistent, and generalizable. We call our resulting NVS network 'SV3D'. To our knowledge, this is the first work that adapts a video diffusion model for explicit pose-controlled view synthesis. Some contemporary works such as [2, 29] demonstrate this, but only generate orbital videos without any explicit camera control." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.641, + 0.788, + 0.763 + ], + "angle": 0, + "content": "3D Generation. We then utilize our SV3D model for 3D object generation by optimizing a NeRF and DMTet mesh in a coarse-to-fine manner. Benefiting from the multi-view consistency in SV3D, we are able to produce high-quality 3D meshes directly from the SV3D novel view images. We also design a masked score distillation sampling (SDS) [37] loss to further enhance 3D quality in the regions that are not visible in the SV3D-predicted novel views. In addition, we propose to jointly optimize a disentangled illumination model along with 3D shape and texture, effectively reducing the issue of baked-in lighting." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "We perform extensive comparisons of both our NVS and 3D generation results with respective state-of-the-art methods, demonstrating considerably better outputs with SV3D. For NVS, SV3D shows high-level of multi-view consistency and generalization to real-world images while being pose controllable. Our resulting 3D meshes are able to capture intricate geometric and texture details." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.689, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SV3D" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.368, + 0.163 + ], + "angle": 0, + "content": "2 Background" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.175, + 0.442, + 0.192 + ], + "angle": 0, + "content": "2.1 Novel View Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.199, + 0.784, + 0.229 + ], + "angle": 0, + "content": "We organize the related works along three crucial aspects of novel view synthesis (NVS): generalization, controllability, and multi-view (3D) consistency." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.231, + 0.785, + 0.337 + ], + "angle": 0, + "content": "Generalization. Diffusion models [14, 48] have recently shown to be powerful generative models that can generate a wide variety of images [3, 40, 41] and videos [2, 12, 53] by iteratively denoising a noise sample. Among these models, the publicly available Stable Diffusion (SD) [40] and Stable Video Diffusion (SVD) [2] demonstrate strong generalization ability by being trained on extremely large datasets like LAION [43] and LVD [2]. Hence, they are commonly used as foundation models for various generation tasks, e.g. novel view synthesis." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.339, + 0.787, + 0.52 + ], + "angle": 0, + "content": "Controllability. Ideally, a controllable NVS model allows us to generate an image corresponding to any arbitrary viewpoint. For this, Zero123 [26] repurposes an image diffusion model to a novel view synthesizer, conditioned on a single-view image and the pose difference between the input and target views. Follow-up works such as Zero123XL [8] and Stable Zero123 [49] advance the quality of diffusion-based NVS, as well as the trained NeRFs using SDS loss. However, they only generate one novel view at a time, and thus are not designed to be multi-view consistent inherently. Recent works such as EscherNet [21] and Free3D [64] are capable of better multi-view consistency with intelligent camera position embedding design. However, they only use image-based diffusion models, and generate images at \\(256 \\times 256\\) resolution. We finetune a video diffusion model to generate novel views at \\(576 \\times 576\\) resolution." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.522, + 0.787, + 0.763 + ], + "angle": 0, + "content": "Multi-view Consistency. Multi-view consistency is the most critical requirement for high-quality NVS and 3D generation. To address the inconsistency issue, MVDream [46], SyncDreamer [27], HexGen3D [30], and Zero123++ [45] propose to generate multiple (specific) views of an object simultaneously. However, they are not controllable: given a conditional image, they only generate specific views, not arbitrary viewpoints. Moreover, they were finetuned from image-based diffusion models, i.e. multi-view consistency was imposed by adding interaction among the multiple generated views through cross-attention layers. Hence, their output quality is limited to the generalizability of their base image-based model, and their 3D finetuning dataset. Efficient-3DiM finetunes the SD model with a stronger vision transformer DINO v2 [36]. Consistent-1-to-3 [59] and SPAD [18] leverage epipolar geometry. One-2-3-45 [25] and One-2-3-45++ [24] train additional 3D network using the 2D generator's outputs. MVDream [46], Consistent123 [55], and Wonder3D [28] also train multi-view diffusion models, yet still require post-processing for video rendering. SyncDreamer [27] and ConsistentNet [57] employ 3D representation into the latent diffusion model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Exploiting Video Diffusion Models. To achieve better generalization and multi-view consistency, some contemporary works exploit the temporal priors in video diffusion models for NVs. For instance, Vivid-1-to-3 [22] combines a view-conditioned diffusion model and video diffusion model to generate consistent views. SVD-MV [2] and IM-3D [29] finetune a video diffusion model for NVs." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.372, + 0.128 + ], + "angle": 0, + "content": "V. Voleti et al." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.177 + ], + "angle": 0, + "content": "However, they generate \\(\\leq 360^{\\circ}\\) views at the same elevation only. Unlike SV3D, none of them are capable of rendering any arbitrary view of the 3D object." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.178, + 0.787, + 0.253 + ], + "angle": 0, + "content": "We argue that the existing NVS and 3D generation methods do not fully leverage the superior generalization capability, controllability, and consistency in video diffusion models. In this paper, we fill this important gap and train SV3D, a state-of-the-art novel multi-view synthesis video diffusion model at \\(576 \\times 576\\) resolution, and leverage it for 3D generation." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.275, + 0.387, + 0.29 + ], + "angle": 0, + "content": "2.2 3D Generation" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.3, + 0.787, + 0.344 + ], + "angle": 0, + "content": "Recent advances in 3D representations and diffusion-based generative models have significantly improved the quality of image-to-3D generation. Here, we briefly summarize the related works in these two categories." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.347, + 0.787, + 0.512 + ], + "angle": 0, + "content": "3D Representation. 3D generation has seen great progress since the advent of Neural Radiance Fields (NeRFs) [32] and its subsequent variants [1], which implicitly represents a 3D scene as a volumetric function, typically parameterized by a neural network. Notably, Instant-NGP [33] introduces a hash grid feature encoding that can be used as a NeRF backbone for fast inference and ability to recover complex geometry. On the other hand, several recent works improve from an explicit representation such as DMTet [44], which is capable of generating high-resolution 3D shapes due to its hybrid SDF-Mesh representation and high memory efficiency. Similar to [23, 38], we adopt coarse-to-fine training for 3D generation, by first learning a rough object with Instant-NGP NeRF and then refining it using the DMTet representation." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.514, + 0.787, + 0.696 + ], + "angle": 0, + "content": "Diffusion-Based 3D Generation. Several recent works [16, 34] train a 3D diffusion model to to learn these flexible 3D representations, which, however, lack generalizability due to the scarcity of 3D data. To learn 3D generation without ground truth 3D data, image/multi-view diffusion models have been used as guidance. DreamFusion [37] and follow-up works [23, 31] leverage a trained image diffusion model as a 'scoring' function and calculate SDS loss for text-to-3D generation. However, they are prone to artifacts like Janus problem [31, 37] and over-saturated texture. Inspired by Zero123 [26], recent works [21, 24, 25, 27, 29, 38, 45, 49, 51, 64] finetune image/video diffusion models to generate novel view images as a stronger guidance. Our method shares the same spirit as this line of work, but produces denser, controllable, and more consistent multi-view images, thus resulting in better 3D generation quality." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.718, + 0.589, + 0.736 + ], + "angle": 0, + "content": "3 SV3D: Novel Multi-view Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.75, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Our main idea is to repurpose temporal consistency in a video diffusion model for spatial 3D consistency of an object. Specifically, we finetune SVD [2] to generate an orbital video around a 3D object, conditioned on a single-view image. This orbital video need not be at the same elevation, or at regularly spaced azimuth angles. SVD is well-suited for this task since it is trained to generate smooth and consistent videos on large-scale datasets of real and high-quality videos." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.69, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SV3D" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.143, + 0.645, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.241, + 0.3, + 0.597, + 0.317 + ], + "angle": 0, + "content": ":FullyConnected :Embedding :VAE Encoder" + }, + { + "type": "image", + "bbox": [ + 0.658, + 0.155, + 0.774, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.696, + 0.216, + 0.733, + 0.226 + ], + "angle": 0, + "content": "Static" + }, + { + "type": "image", + "bbox": [ + 0.661, + 0.233, + 0.774, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.688, + 0.304, + 0.741, + 0.313 + ], + "angle": 0, + "content": "Dynamic" + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.319, + 0.617, + 0.402 + ], + "angle": 0, + "content": "Fig. 2: SV3D Architecture. We add the sinusoidal embedding of the camera orbit elevation and azimuth angles \\((e, a)\\) to that of the noise step \\(t\\), and feed the sum to the convolutional blocks in the UNet. We feed the single input image's CLIP embedding to the attention blocks, and concatenate its latent embedding to the noisy state \\(\\mathbf{z}_t\\)." + }, + { + "type": "image_caption", + "bbox": [ + 0.646, + 0.325, + 0.781, + 0.351 + ], + "angle": 0, + "content": "Fig.3:Static vs. DynamicOrbits." + }, + { + "type": "image_footnote", + "bbox": [ + 0.646, + 0.353, + 0.781, + 0.393 + ], + "angle": 0, + "content": "We use two types of orbits for training the SV3D models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.421, + 0.785, + 0.466 + ], + "angle": 0, + "content": "The exposure to superior data quantity and quality makes it more generalizable and multi-view consistent, and the flexibility of the SVD architecture makes it amenable to be finetuned for camera controllability." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.469, + 0.785, + 0.575 + ], + "angle": 0, + "content": "Some prior works attempt to leverage such properties by finetuning image diffusion models, training video diffusion models from scratch, or finetuning video diffusion models to generate pre-defined views at the same elevation (static orbit) around an object [2, 29]. However, we argue that these methods do not fully exploit the potential of video diffusion models. To the best of our knowledge, SV3D is the first video diffusion-based framework for controllable multi-view synthesis at \\(576 \\times 576\\) resolution (and subsequently for 3D generation)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.578, + 0.785, + 0.7 + ], + "angle": 0, + "content": "Problem Setting. Formally, given an image \\(\\mathbf{I} \\in \\mathbb{R}^{3 \\times H \\times W}\\) of an object, our goal is to generate an orbital video \\(\\mathbf{J} \\in \\mathbb{R}^{K \\times 3 \\times H \\times W}\\) around the object of \\(K = 21\\) multi-view images along a camera trajectory \\(\\pmb{\\pi} \\in \\mathbb{R}^{K \\times 2} = \\{(e_i, a_i)\\}_{i=1}^K\\) as a sequence of \\(K\\) tuples of elevation \\(e\\) and azimuth \\(a\\) angles. We assume the camera always looks at the center of an object (origin of world coordinates), so any viewpoint can be specified by only two parameters: \\(e\\) and \\(a\\). We generate this orbital video by iteratively denoising samples from a learned conditional distribution \\(p(\\mathbf{J}|\\mathbf{I}, \\pmb{\\pi})\\), parameterized by a video diffusion model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.787, + 0.84 + ], + "angle": 0, + "content": "SV3D Architecture. As shown in Fig. 2, the architecture of SV3D builds on that of SVD consisting of a UNet with multiple layers, each layer containing sequences of one residual block with Conv3D layers, and two transformer blocks (spatial and temporal) with attention layers. (i) We remove the vector conditionings of 'fps id' and 'motion bucket id' since they are irrelevant for SV3D. (ii) The conditioning image is concatenated to the noisy latent state input \\(\\mathbf{z}_t\\) at noise timestep \\(t\\) to the UNet, after being embedded into latent space by the VAE encoder of SVD. (iii) The CLIP-embedding [39] matrix of the conditioning image is provided to the cross-attention layers of each transformer block as its key and" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.372, + 0.127 + ], + "angle": 0, + "content": "V. Voleti et al." + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.151, + 0.532, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.567, + 0.152, + 0.78, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.316, + 0.782, + 0.358 + ], + "angle": 0, + "content": "Fig. 4: Linear vs. Triangle CFG Scaling. Notice Fig. 5: LPIPS vs. Frame increased oversharping in the penultimate frame in Number. SV3D has the best re-linear scaling vs. our proposed triangle scaling. construction metric per frame." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.366, + 0.788, + 0.472 + ], + "angle": 0, + "content": "value, the query being the feature at that layer. (iv) The camera trajectory is fed into the residual blocks along with the diffusion noise timestep. The camera pose angles \\( e_i \\) and \\( a_i \\) and the noise timestep \\( t \\) are first embedded into sinusoidal position embeddings. Then, the camera pose embeddings are concatenated together, linearly transformed, and added to the noise timestep embedding. This is fed to every residual block, where they are added to the block's output feature (after being linearly transformed again to match the feature size)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.475, + 0.789, + 0.656 + ], + "angle": 0, + "content": "Static v.s. Dynamic Orbits. As shown in Fig. 3, we design static and dynamic orbits to study the effects of camera pose conditioning. In a static orbit, the camera circles around an object at regularly-spaced azimuths at the same elevation angle as that in the conditioning image. The disadvantage with the static orbit is that we might not get any information about the top or bottom of the object depending on the conditioning elevation angle. In a dynamic orbit, the azimuths can be irregularly spaced, and the elevation can vary per view. To build a dynamic orbit, we sample a static orbit, add small random noise to the azimuth angles, and add a random weighted combination of sinusoids with different frequencies to its elevation. This provides temporal smoothness, and ensures that the camera trajectory loops around to end at the same azimuth and elevation as those of the conditioning image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.657, + 0.788, + 0.717 + ], + "angle": 0, + "content": "Thus, with this strategy, we are able to tackle all three aspects of generalization, controllability, and consistency in novel multi-view synthesis by leveraging video diffusion models, additionally conditioning on camera trajectory, and repurposing temporal consistency for spatial 3D object consistency, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Triangular CFG Scaling. SVD uses a linearly increasing scale for classifier-free guidance (CFG) from 1 to 4 across the generated frames. However, this scaling causes the last few frames in our generated orbits to be over-sharpened, as shown in Fig. 4 Frame 20. Since we generate videos looping back to the front-view image, we propose to use a triangle wave CFG scaling during inference: linearly increase CFG from 1 at the front view to 2.5 at the back view, then linearly decrease it back to 1 at the front view. Fig. 4 also demonstrates that our triangle CFG scaling produces more details in the back view (Frame 12)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.69, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SV3D" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.314 + ], + "angle": 0, + "content": "Models. We train three image-to-3D-video models finetuned from SVD. First, we train a pose-unconditioned model, \\(\\mathrm{SV3D}^u\\), which generates a video of static orbit around an object while only conditioned on a single-view image. Note that unlike SVD-MV [2], we do not provide the elevation angle to the pose-unconditioned model, as we find that the model is able to infer it from the conditioning image. Our second model, the pose-conditioned \\(\\mathrm{SV3D}^c\\) is conditioned on the input image as well as a sequence of camera elevation and azimuth angles in an orbit, trained on dynamic orbits. Following SVD's [2] intuition to progressively increase the task difficulty during training, we train our third model, \\(\\mathrm{SV3D}^p\\), by first finetuning SVD to produce static orbits unconditionally, then further finetuning on dynamic orbits with camera pose condition." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.315, + 0.788, + 0.436 + ], + "angle": 0, + "content": "Training Details. We train SV3D on the Objverse dataset [9], which contains synthetic 3D objects covering a wide diversity. For each object, we render 21 frames around it on random color background at \\(576 \\times 576\\) resolution, field-of-view of 33.8 degrees. We choose to finetune the SVD-xt model to output 21 frames. All three models (SV3D\\(^u\\), SV3D\\(^c\\), SV3D\\(^p\\)) are trained for 105k iterations in total (SV3D\\(^p\\) is trained unconditionally for 55k iterations and conditionally for 50k iterations), with an effective batch size of 64 on 4 nodes of 8 80GB A100 GPUs for around 6 days. For more training details, please see the appendix." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.46, + 0.471, + 0.474 + ], + "angle": 0, + "content": "3.1 Experiments and Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.487, + 0.788, + 0.594 + ], + "angle": 0, + "content": "Datasets. We evaluate the SV3D-generated multi-view images on static and dynamic orbits on the unseen GSO [10] and OmniObject3D [56] datasets. Since many GSO objects are the same items with slightly different colors, we filter 300 objects from GSO to reduce redundancy and maintain diversity. For each object, we render ground truth static and dynamic orbit videos and pick the last frame of each video as the conditioning image. We also conduct a user study on novel view videos from a dataset of 22 real-world images. (More details in appendix.)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.595, + 0.788, + 0.716 + ], + "angle": 0, + "content": "Metrics. We use the SV3D models to generate static and dynamic orbit videos corresponding to the ground truth camera trajectories in the evaluation datasets. We compare each generated frame with the corresponding ground truth frames, in terms of Learned Perceptual Similarity (LPIPS [63]), Peak Signal-to-Noise Ratio (PSNR), Structural SIMilarity (SSIM), Mean Squared-Error (MSE), and CLIP-score (CLIP-S). This range of metrics covers both pixel-level as well as semantic aspects. Note that testing on dynamic orbits evaluates the controllability of SV3D models, and all the metrics evaluate multi-view consistency." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.718, + 0.788, + 0.763 + ], + "angle": 0, + "content": "Baselines. We compare SV3D with several recent NVS methods capable of generating arbitrary views, including Zero123 [26], Zero123-XL [8], SyncDreamer [27], Stable Zero123 [49], Free3D [64], EscherNet [21]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Results. As shown in Tabs. 1 to 4, our SV3D models achieve state-of-the-art performance on novel multi-view synthesis. Tabs. 1 and 3 show results on static orbits, and include all our three models. We see that even our pose-unconditioned model \\(\\mathrm{SV3D}^u\\) performs better than all prior methods. Tabs. 2 and 4 show results on dynamic orbits, and include our pose-conditioned models \\(\\mathrm{SV3D}^c\\) and \\(\\mathrm{SV3D}^p\\)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.128 + ], + "angle": 0, + "content": "V.Voleti et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.144, + 0.784, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.431, + 0.788, + 0.473 + ], + "angle": 0, + "content": "Fig.6: Visual Comparison of Novel Multi-view Synthesis. SV3D is able to generate novel multi-views that are more detailed, faithful to the conditioning image, and multi-view consistent compared to the prior works." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.495, + 0.788, + 0.602 + ], + "angle": 0, + "content": "Ablative Analyses. Interestingly, from Tabs. 1 and 3, we find that both \\(\\mathrm{SV3D}^c\\) and \\(\\mathrm{SV3D}^p\\) outperform \\(\\mathrm{SV3D}^u\\) on generations of static orbits, even though \\(\\mathrm{SV3D}^u\\) is trained specifically on static orbits. We also observe that \\(\\mathrm{SV3D}^p\\) achieves better metrics than \\(\\mathrm{SV3D}^c\\) on both static and dynamic orbits, making it the best performing SV3D model overall. This shows that progressive finetuning from easier (static) to harder (dynamic) tasks is indeed a favorable way to finetune a video diffusion model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.606, + 0.787, + 0.652 + ], + "angle": 0, + "content": "Visual Comparisons in Fig. 6 further demonstrate that SV3D-generated images are more detailed, faithful to the conditioning image, and multi-view consistent compared to the prior works." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.655, + 0.788, + 0.732 + ], + "angle": 0, + "content": "Quality Per Frame. We also observe from Fig. 5 that SV3D produces better quality at every frame. We plot the average LPIPS value for each generated frame, across generated GSO static orbit videos. The quality is generally worse around the back view, and better at the beginning and the end (i.e. near the conditioning image), as expected." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.842 + ], + "angle": 0, + "content": "User Study on Real-World Images. We conducted a user survey to study human preference between static orbital videos generated by SV3D and by other methods. We asked 30 users to pick one between our SV3D-generated static video and other method-generated video as the best orbital video for the corresponding image, using 22 real-world images. We noted that users preferred SV3D-generated videos over Zero123XL, Stable Zero123, EscherNet, and Free3D, \\(96\\%\\), \\(99\\%\\), \\(96\\%\\), and \\(98\\%\\) of the time, respectively." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.69, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SV3D" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.497, + 0.172 + ], + "angle": 0, + "content": "Table 1: Evaluation of novel multi-view synthesis on GSO static orbits" + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.174, + 0.496, + 0.286 + ], + "angle": 0, + "content": "
ModelLPIPS↓PSNR↑SSIM↑CLIP-S↑MSE↓
SyncDreamer [27]0.1715.780.760.870.03
Zero123 [26]0.1317.290.790.850.04
Zero123XL [8]0.1417.110.780.850.04
Stable Zero123 [49]0.1318.340.780.850.05
Free3D [64]0.1516.180.790.840.04
EscherNet [21]0.1316.730.790.850.03
SV3Du0.0921.140.870.890.02
SV3Dc0.0920.560.870.880.02
SV3Dp0.0821.260.880.890.02
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.145, + 0.785, + 0.172 + ], + "angle": 0, + "content": "Table 2: Evaluation of novel multi-view synthesis on GSO dynamic orbits" + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.174, + 0.784, + 0.269 + ], + "angle": 0, + "content": "
ModelLPIPS↓PSNR↑SSIM↑CLIP-S↑MSE↓
Zero123 [26]0.1416.990.790.840.04
Zero123XL [8]0.1416.730.780.840.04
Stable Zero123 [49]0.1318.040.780.850.05
Free3D [64]0.1814.930.770.830.05
EscherNet [21]0.1316.470.790.840.03
SV3Dc0.1019.990.860.870.02
SV3Dp0.0920.380.870.870.02
" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.297, + 0.497, + 0.324 + ], + "angle": 0, + "content": "Table 3: Evaluation of novel multi-view synthesis on OmniObject3D static orbits" + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.325, + 0.496, + 0.429 + ], + "angle": 0, + "content": "
ModelLPIPS↓PSNR↑SSIM↑CLIP-S↑MSE↓
Zero123 [26]0.1715.500.760.830.05
Zero123XL [8]0.1815.360.750.830.06
Stable Zero123 [49]0.1516.860.770.840.06
Free3D [64]0.1615.290.780.830.05
EscherNet [21]0.1714.630.740.830.05
SV3Du0.1019.680.860.860.02
SV3Dc0.1019.500.850.850.02
SV3Dp0.1019.910.860.860.02
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.297, + 0.786, + 0.324 + ], + "angle": 0, + "content": "Table 4: Evaluation of novel multi-view synthesis on OmniObject3D dynamic orbits" + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.325, + 0.784, + 0.42 + ], + "angle": 0, + "content": "
ModelLPIPS↓PSNR↑SSIM↑CLIP-S↑MSE↓
Zero123 [26]0.1615.780.770.820.05
Zero123XL [8]0.1715.490.760.830.05
Stable Zero123 [49]0.1416.740.770.830.05
Free3D [64]0.1914.280.760.820.06
EscherNet [21]0.1615.050.760.830.05
SV3Dc0.1019.210.850.840.02
SV3Dp0.1019.280.850.850.02
" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.435, + 0.723, + 0.453 + ], + "angle": 0, + "content": "4 3D Generation from a Single Image Using SV3D" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.471, + 0.786, + 0.532 + ], + "angle": 0, + "content": "We then generate 3D meshes of objects from a single image by leveraging SV3D. One way is to use the generated static/dynamic orbital samples from SV3D as direct reconstruction targets. Another way is to use SV3D as diffusion guidance with Score Distillation Sampling (SDS) loss [37]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.533, + 0.788, + 0.623 + ], + "angle": 0, + "content": "Since SV3D produces more consistent multi-views compared to existing works, we already observe higher-quality 3D reconstructions by only using SV3D outputs for reconstruction when compared to existing works. However, we observe that this naive approach often leads to artifacts like baked-in illumination, rough surfaces, and noisy texture, especially for the unseen regions in the reference orbit. Thus, we further propose several techniques to address these issues." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.627, + 0.788, + 0.763 + ], + "angle": 0, + "content": "Coarse-to-Fine Training. We adopt a two-stage, coarse-to-fine training scheme to generate a 3D mesh from input images, similar to [23, 38]. Fig. 7 illustrate an overview of our 3D optimization pipeline. In the coarse stage, we train an Instant-NGP [33] NeRF representation to reconstruct the SV3D-generated images (i.e. without SDS loss) at a lower resolution. In the fine stage, we extract a mesh from the trained NeRF using marching cubes [7], and adopt a DMTet [44] representation to finetune the 3D mesh using SDS-based diffusion guidance from SV3D at full-resolution. Finally, we use xatlas [60] to perform the UV unwrapping and export the output object mesh." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Disentangled Illumination Model. Similar to other recent 3D object generation methods [23, 31, 37], our output target is a mesh with a diffuse texture. Typically, such SDS-based optimization techniques use random illuminations at every iteration. However, our SV3D-generated videos are under consistent illumination, i.e., the lighting remains static while the camera circles around an object." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.372, + 0.127 + ], + "angle": 0, + "content": "V. Voleti et al." + }, + { + "type": "image", + "bbox": [ + 0.232, + 0.146, + 0.77, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.323, + 0.788, + 0.407 + ], + "angle": 0, + "content": "Fig.7: 3D Optimization Pipeline. We perform a two-stage optimization. In the short first stage, we extract the general shape, texture and illumination from the SV3D generated multi-view images. In the second stage, we extract a mesh with marching cubes and use DMTet to further optimize the shape, texture and illumination. We not only use the SV3D-generated images but a soft-masked SDS loss for unseen areas. Dashed red lines represent backpropagation into the differentiable rendering pipeline." + }, + { + "type": "image", + "bbox": [ + 0.25, + 0.415, + 0.291, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.255, + 0.489, + 0.288, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.247, + 0.563, + 0.294, + 0.578 + ], + "angle": 0, + "content": "Ref. images in static orbit" + }, + { + "type": "image", + "bbox": [ + 0.319, + 0.416, + 0.365, + 0.483 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.319, + 0.485, + 0.361, + 0.557 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.322, + 0.563, + 0.362, + 0.578 + ], + "angle": 0, + "content": "3D from static orbit" + }, + { + "type": "image", + "bbox": [ + 0.395, + 0.416, + 0.441, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.396, + 0.484, + 0.436, + 0.557 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.392, + 0.563, + 0.443, + 0.578 + ], + "angle": 0, + "content": "3D from dynamic orbit" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.58, + 0.477, + 0.636 + ], + "angle": 0, + "content": "Fig.8: Influence of Training Orbits. We show that using a dynamic orbit is crucial to 3D generations that are complete from diverse views." + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.416, + 0.586, + 0.473 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.416, + 0.673, + 0.473 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.68, + 0.416, + 0.765, + 0.473 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.48, + 0.589, + 0.538 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.543, + 0.575, + 0.553 + ], + "angle": 0, + "content": "No SDS loss" + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.48, + 0.674, + 0.538 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.613, + 0.543, + 0.655, + 0.553 + ], + "angle": 0, + "content": "SDS loss" + }, + { + "type": "image", + "bbox": [ + 0.687, + 0.48, + 0.764, + 0.538 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.679, + 0.543, + 0.759, + 0.553 + ], + "angle": 0, + "content": "Masked SDS loss" + }, + { + "type": "image_caption", + "bbox": [ + 0.481, + 0.567, + 0.783, + 0.636 + ], + "angle": 0, + "content": "Fig.9: Influence of SDS. Using our masked SDS loss, we are able to fill in unseen surfaces in the training orbit, producing a cleaner result without oversaturation or blurry artifacts caused by naive SDS." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.657, + 0.785, + 0.702 + ], + "angle": 0, + "content": "Hence, to disentangle lighting effects and obtain a cleaner texture, we propose to fit a simple illumination model of 24 Spherical Gaussians (SG) inspired by prior decomposition methods [4,62]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.703, + 0.785, + 0.763 + ], + "angle": 0, + "content": "We model white light and hence only use a scalar amplitude for the SGs. We only consider Lambertian shading, where the cosine shading term is approximated with another SG. We learn the parameters of the illumination SGs using a reconstruction loss between the rendered images and SV3D-generated images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Inspired by [13, 37] we reduce baked-in illumination with a loss term that replicates the HSV-value component of the input image \\(\\mathbf{I}\\) with the rendered illumination \\(L\\): \\(\\mathcal{L}_{\\mathrm{illum}} = |V(\\mathbf{I}) - L|^2\\), with \\(V(\\pmb {c}) = \\max (c_r,c_g,c_b)\\). Given these changes, our disentangled illumination model is able to express lighting variation properly and can severely reduce baked-in illumination. Fig. 10 shows sample" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.689, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SV3D" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "reconstructions with and without our illumination modelling. From the results, it is clear that we are able to disentangle the illumination effects from the base color (e.g., dark side of the school bus)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.215, + 0.582, + 0.231 + ], + "angle": 0, + "content": "4.1 3D Optimization Strategies and Losses" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.243, + 0.784, + 0.333 + ], + "angle": 0, + "content": "Reconstruction via Photometric Losses. Intuitively, we can treat the SV3D-generated images as multi-view pseudo-ground truth, and apply 2D reconstruction losses to train the 3D models. In this case, we apply photometric losses on the rendered images from NeRF or DMTet, including the pixel-level MSE loss, mask loss, and a perceptual LPIPS [63] loss. These photometric losses also optimize our illumination model through the differential rendering pipeline." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.336, + 0.551, + 0.531 + ], + "angle": 0, + "content": "Training Orbits. For 3D generation, we use SV3D to generate multi-view images following a camera orbit \\(\\pi_{\\mathrm{ref}}\\) , referred to as the reference orbit (also see Fig. 7 for the overview). Fig. 8 shows sample reconstruction with using static and dynamic orbital outputs from SV3D. Using a dynamic orbit for training is crucial to high-quality 3D outputs when viewed from various angles, since some top/bottom views are missing in the static orbit (fixed elevation). Hence, for \\(\\mathrm{SV3D}^c\\) and \\(\\mathrm{SV3D}^p\\) , we render images on a dynamic orbit whose elevation follows a sine function to ensure that top and bo" + }, + { + "type": "image", + "bbox": [ + 0.581, + 0.354, + 0.667, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.687, + 0.354, + 0.774, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.562, + 0.4, + 0.674, + 0.409 + ], + "angle": 0, + "content": "Constant Illumination" + }, + { + "type": "image_footnote", + "bbox": [ + 0.683, + 0.4, + 0.769, + 0.409 + ], + "angle": 0, + "content": "SGs Illumination" + }, + { + "type": "list", + "bbox": [ + 0.562, + 0.4, + 0.769, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.56, + 0.411, + 0.787, + 0.507 + ], + "angle": 0, + "content": "Fig.10: Constant vs. SGs Illumination. Our SGs-based reconstructions do not exhibit darkening on the side of the bus, which enables easier and more convincing relighting for downstream applications." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.534, + 0.784, + 0.639 + ], + "angle": 0, + "content": "SV3D-Based SDS Loss. In addition to the reconstruction losses, we can also use SV3D via score-distillation sampling (SDS) [37, 58]. Fig. 9 shows sample reconstructions with and without using SDS losses. As shown in Fig. 9 (left), although training with a dynamic orbit improves overall visibility, we observe that sometimes the output texture is still noisy, perhaps due to partial visibility, self-occlusions, or inconsistent texture/shape between images. Hence, we handle those unseen areas using SDS loss [37] with SV3D as a diffusion guidance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.64, + 0.785, + 0.763 + ], + "angle": 0, + "content": "We sample a random camera orbit \\(\\pi_{\\mathrm{rand}}\\), and use our 3D NeRF/DMTet parameterized by \\(\\theta\\) to render the views \\(\\hat{\\mathbf{J}}\\) of the object along \\(\\pi_{\\mathrm{rand}}\\). Then, noise \\(\\epsilon\\) at level \\(t\\) is added to the latent embedding \\(\\mathbf{z}_t\\) of \\(\\hat{\\mathbf{J}}\\), and the following SDS loss (taken expectation over \\(t\\) and \\(\\epsilon\\)) is backpropagated through the differentiable rendering pipeline: \\(\\mathcal{L}_{\\mathrm{sds}} = w(t)\\bigl (\\epsilon_{\\phi}(\\mathbf{z}_t;\\mathbf{I},\\pi_{\\mathrm{rand}},t) - \\epsilon \\bigr)\\frac{\\partial\\hat{\\mathbf{J}}}{\\partial\\theta}\\), where \\(w\\) is \\(t\\)-dependent weight, \\(\\epsilon\\) and \\(\\epsilon_{\\phi}\\) are the added and predicted noise, and \\(\\phi\\) and \\(\\theta\\) are the parameters of SV3D and NeRF/DMTet, respectively. See Fig. 7 for an illustration of these loss functions in the overall pipeline." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Masked SDS Loss. As shown in Fig. 9 (middle), in our experiments we found that adding the SDS loss naively can cause unstable training and unfaithful texture to the input images such as oversaturation or blurry artifacts. Therefore, we design a soft masking mechanism to only apply SDS loss on the unseen/occluded areas, allowing it to inpaint the missing details while preserving the texture of" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.271, + 0.115, + 0.373, + 0.127 + ], + "angle": 0, + "content": "V. Voleti et al." + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.145, + 0.753, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.284, + 0.787, + 0.34 + ], + "angle": 0, + "content": "Fig. 11: Visual Comparison of 3D Meshes. For each object, we show the conditioning image (left), and the output meshes rendered in two different views. Our generated meshes are more detailed, faithful to input images, and consistent in 3D. This demonstrates the quality of novel multi-view synthesis by our SV3D model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.359, + 0.785, + 0.404 + ], + "angle": 0, + "content": "clearly-visible surfaces in the training orbit (as seen in Fig. 9 (right)). Also, we only apply the masked SDS loss in the final stage of DMTet optimization, which greatly increased the convergence speed." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.405, + 0.789, + 0.637 + ], + "angle": 0, + "content": "We apply SDS loss on only those pixels in the random orbit views that are not likely visible in the reference orbit views. For this, we first render the object from the random camera orbit \\(\\pi_{\\mathrm{rand}}\\). For each random camera view, we obtain the visible surface points \\(p \\in \\mathbb{R}^3\\) and their corresponding surface normals \\(\\boldsymbol{n}\\). Then, for each reference camera \\(i\\), we calculate the view directions \\(\\boldsymbol{v}_i\\) of the surface \\(\\boldsymbol{p}\\) towards its position \\(\\bar{\\pi}_{\\mathrm{ref}}^i \\in \\mathbb{R}^3\\) (calculated from \\(\\pi_{\\mathrm{ref}}^i \\in \\mathbb{R}^2\\)) as \\(\\boldsymbol{v}_i = \\frac{\\bar{\\pi}_{\\mathrm{ref}}^i - \\boldsymbol{p}}{||\\bar{\\pi}_{\\mathrm{ref}}^i - \\boldsymbol{p}||}\\). We infer the visibility of this surface from the reference camera based on the dot product between \\(\\boldsymbol{v}_i\\) and \\(\\boldsymbol{n}\\) i.e. \\(\\boldsymbol{v}_i \\cdot \\boldsymbol{n}\\). Since higher values roughly indicate more visibility of the surface from the reference camera, we chose that reference camera \\(c\\) that has maximum likelihood of visibility: \\(c = \\max_i (\\boldsymbol{v}_i \\cdot \\boldsymbol{n})\\). As we only want to apply SDS loss to unseen or grazing angle areas from \\(c\\), we use the smoothstep function \\(f_s\\) to smoothly clip to \\(c\\)'s visibility range \\(\\boldsymbol{v}_c \\cdot \\boldsymbol{n}\\). In this way, we create a pseudo visibility mask \\(M = 1 - f_s(\\boldsymbol{v}_c \\cdot \\boldsymbol{n}, 0, 0.5)\\), where \\(f_s(x; f_0, f_1) = \\hat{x}^2(3 - 2x)\\) with \\(\\hat{x} = \\frac{x - f_0}{f_1 - f_0}\\). Thus, \\(M\\) is calculated for each random camera render, and the combined visibility mask \\(\\mathbf{M}\\) is applied to SDS loss: \\(\\mathcal{L}_{\\mathrm{mask-sds}} = \\mathbf{M}\\mathcal{L}_{\\mathrm{sds}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.639, + 0.788, + 0.76 + ], + "angle": 0, + "content": "Geometric Priors. Since our rendering-based optimization operates at the image level, we adopt several geometric priors to regularize the output shapes. We add a smooth depth loss from RegNeRF [35] and a bilateral normal smoothness loss [5] to encourage smooth 3D surfaces where the projected image gradients are low. Moreover, we obtain normal estimates from Omnidata [11] and calculate a mono normal loss similar to MonoSDF [61], which can effectively reduce noisy surfaces in the output mesh. Further details about the training losses and optimization process are available in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.784, + 0.471, + 0.799 + ], + "angle": 0, + "content": "4.2 Experiments and Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Due to the strong regularization, we only require 600 steps in the coarse stage and 1000 in the fine stage. Overall, the entire mesh extraction process takes \\(\\approx 8\\)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.689, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SV3D" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.146, + 0.752, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.235, + 0.785, + 0.262 + ], + "angle": 0, + "content": "Fig. 12: Real-World 3D Results. Notice the accurate shape and details in our reconstructions even from the diverse images in-the-wild." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.785, + 0.314 + ], + "angle": 0, + "content": "minutes without SDS loss, and \\(\\approx 20\\) minutes with SDS loss. The coarse stage only takes \\(\\approx 2\\) minutes and provides a full 3D representation of the object." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.316, + 0.785, + 0.452 + ], + "angle": 0, + "content": "Evaluation. We evaluate our 3D generation framework on 50 randomly sampled objects from the GSO dataset as described in Sec. 3.1. We compute image-based reconstruction metrics (LPIPS, PSNR, SSIM, MSE, and CLIP-S) between the ground truth (GT) GSO images, and rendered images from the trained 3D meshes on the same static/dynamic orbits. In addition, we compute 3D reconstruction metrics of Chamfer distance (CD) and 3D IoU between the GT and predicted meshes. We compare our SV3D-guided 3D generations with several prior methods including Point-E [34], Shap-E [16], One-2-3-45++ [24], DreamGaussian [50], SyncDreamer [27], EscherNet [21], Free3D [64], and Stable Zero123 [49]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.455, + 0.787, + 0.575 + ], + "angle": 0, + "content": "Visual Comparison. In Fig. 11, we show visual comparison of our results with those from prior methods. Qualitatively, Point-E [34] and Shap-E [16] often produce incomplete 3D shapes. DreamGaussian [50], SyncDreamer [27], EscherNet [21], and Free3D [64] outputs tend to contain rough surfaces. One-2-3-45++ [24] and Stable Zero123 [49] are able to reconstruct meshes with smooth surface, but lack geometric details. Our mesh outputs are detailed, faithful to input image, and consistent in 3D (see appendix for more examples). We also show 3D mesh renders from real-world images in Fig. 12." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.578, + 0.787, + 0.684 + ], + "angle": 0, + "content": "Quantitative Comparison. Tabs. 5 and 6 show the 2D and 3D metric comparisons respectively. All our 3D models achieve better 2D and 3D metrics compared to the prior and concurrent methods, showing the high-fidelity texture and geometry of our output meshes. We render all 3D meshes on the same dynamic orbits and compare them against the GT renders. Our best model, \\(\\mathrm{SV3D}^p\\), performs comparably to using GT renders for reconstruction in terms of the 3D metrics, which further demonstrates the 3D consistency of our generated images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.687, + 0.787, + 0.793 + ], + "angle": 0, + "content": "Effects of Photometric and (Masked) SDS Loss. As shown in Tabs. 5 and 6, the 3D outputs using both photometric and Masked SDS losses (\\( \\mathbf{SV3D}^p \\)) achieves the best results, while training without SDS (\\( \\mathbf{SV3D}^p \\) no SDS') leads to marginally lower performance. This demonstrates that the images generated by SV3D are high-quality reconstruction targets, and are often sufficient for 3D generation without the cumbersome SDS-based optimization. Nevertheless, adding SDS generally improves quality, as also shown in Fig. 9." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Effects of SV3D Model and Training Orbit. As shown in Tabs. 5 and 6, \\(\\mathrm{SV3D}^p\\) achieves the best performance among the three SV3D models, indicating that its synthesized novel views are most faithful to the input image and consis" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.372, + 0.128 + ], + "angle": 0, + "content": "V. Voleti et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.56, + 0.215 + ], + "angle": 0, + "content": "Table 5: 2D comparison of our 3D outputs against prior methods on the GSO dataset. Our best performing method uses \\(\\mathbf{SV3D}^p\\) with dynamic (sine elevation) orbit and SDS guidance. Note that all our models achieve better 2D metrics than prior works." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.217, + 0.564, + 0.348 + ], + "angle": 0, + "content": "
ModelLPIPS↓PSNR↑SSIM↑MSE↓CLIP-S↑
GT renders0.07819.5080.8790.0140.897
EscherNet [21]0.17814.4380.8040.0410.835
Free3D [64]0.19714.2020.7990.0430.809
Stable Zero123 [49]0.16614.6350.8130.0400.805
SV3Du0.13315.9570.8340.0310.871
SV3Dc0.13216.3730.8340.0270.870
SV3Dpstatic orbit0.12516.8210.8480.0250.864
SV3Dpno SDS0.12416.8640.8410.0240.875
SV3Dp0.11917.4050.8490.0210.877
" + }, + { + "type": "table_caption", + "bbox": [ + 0.567, + 0.145, + 0.782, + 0.201 + ], + "angle": 0, + "content": "Table 6: Comparison of 3D metrics on the GSO dataset. Our models perform favorably against prior works." + }, + { + "type": "table", + "bbox": [ + 0.568, + 0.203, + 0.784, + 0.384 + ], + "angle": 0, + "content": "
ModelCD↓3D IoU↑
GT renders0.0210.689
Point-E [34]0.0740.162
Shap-E [16]0.0710.267
DreamGaussian [50]0.0550.411
One-2-3-45++ [24]0.0540.406
SyncDreamer [27]0.0530.451
EscherNet [21]0.0420.466
Free3D [64]0.0470.426
Stable Zero123 [49]0.0390.550
SV3Du0.0270.589
SV3Dc0.0270.584
SV3Dp static orbit0.0280.610
SV3Dp no SDS0.0240.611
SV3Dp0.0240.614
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.389, + 0.787, + 0.48 + ], + "angle": 0, + "content": "tent in 3D. On the other hand, \\(\\mathrm{SV3D}^u\\) shares the same disadvantage as several prior works in that it can only generate views at the same elevation, which is insufficient to build a legible 3D object, as shown in Fig. 8. This also leads to the worse performance of 'SV3D\\(^p\\) with static orbit' in Tabs. 5 and 6. Overall, \\(\\mathrm{SV3D}^p\\) with dynamic orbit and masked SDS loss performs favorably against all other configurations since it can leverage more diverse views of the object." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.482, + 0.788, + 0.618 + ], + "angle": 0, + "content": "Limitations. Our SV3D model is by design only capable of handling 2 degrees of freedom: elevation and azimuth; which is usually sufficient for 3D generation from a single image. One may want to tackle more degrees of freedom in cameras for a generalized NVS system, which forms an interesting future work. We also notice that SV3D exhibits some view inconsistency for mirror-like reflective surfaces, which provide a challenge to 3D reconstruction. Lastly, such reflective surfaces are not representable by our Lambertian reflection-based shading model. Conditioning SV3D on the full camera matrix, and extending the shading model are interesting directions for future research." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.642, + 0.36, + 0.658 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We present SV3D, a latent video diffusion model for novel multi-view synthesis and 3D generation. In addition to leveraging the generalizability and view-consistent prior in SVD, SV3D enables controllability via camera pose conditioning, and generates orbital videos of an object at high resolution on arbitrary camera orbits. We further propose several techniques for improved 3D generation from SV3D, including triangle CFG scaling, disentangled illumination, and masked SDS loss. We conduct extensive experiments to show that SV3D is controllable, multi-view consistent, and generalizable to the real-world, achieving state-of-the-art performance on novel multi-view synthesis and 3D generation. We believe SV3D provides a solid foundation model for further research on 3D object generation. We plan to publicly release SV3D models." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.69, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SV3D" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.146, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.788, + 0.254 + ], + "angle": 0, + "content": "We thank Emad Mostaque, Bryce Wilson, Anel Islamovic, Savannah Martin, Ana Guillen, Josephine Parquet, Adam Chen and Ella Irwin for their help in various aspects of the model development and release. We thank Kyle Lacey and Yan Marek for their help with demos. We also thank Eric Courtemanche for his help with visual results." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.279, + 0.323, + 0.294 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.311, + 0.785, + 0.367 + ], + "angle": 0, + "content": "1. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-NeRF: A multiscale representation for anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5855-5864 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.369, + 0.785, + 0.422 + ], + "angle": 0, + "content": "2. Blattmann, A., Dockhorn, T., Kulal, S., Mendelevitch, D., Kilian, M., Lorenz, D., Levi, Y., English, Z., Voleti, V., Letts, A., et al.: Stable Video Diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.424, + 0.785, + 0.464 + ], + "angle": 0, + "content": "3. Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S.W., Fidler, S., Kreis, K.: Align your Latents: High-Resolution Video Synthesis with Latent Diffusion Models. arXiv:2304.08818 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.465, + 0.785, + 0.506 + ], + "angle": 0, + "content": "4. Boss, M., Braun, R., Jampani, V., Barron, J.T., Liu, C., Lensch, H.P.: NeRD: Neural reflectance decomposition from image collections. In: IEEE International Conference on Computer Vision (ICCV) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.508, + 0.785, + 0.562 + ], + "angle": 0, + "content": "5. Boss, M., Engelhardt, A., Kar, A., Li, Y., Sun, D., Barron, J.T., Lensch, H.P., Jampani, V.: SAMURAI: Shape And Material from Unconstrained Real-world Arbitrary Image collections. In: Advances in Neural Information Processing Systems (NeurIPS) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.563, + 0.785, + 0.617 + ], + "angle": 0, + "content": "6. Chan, E.R., Nagano, K., Chan, M.A., Bergman, A.W., Park, J.J., Levy, A., Aittala, M., Mello, S.D., Karras, T., Wetzstein, G.: GeNVS: Generative novel view synthesis with 3D-aware diffusion models. International Conference on Computer Vision (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.619, + 0.785, + 0.645 + ], + "angle": 0, + "content": "7. Cline, D.B.: Admissible kernel estimators of a multivariate density. The Annals of Statistics 16(4), 1421-1427 (1988)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.647, + 0.785, + 0.687 + ], + "angle": 0, + "content": "8. Deitke, M., Liu, R., Wallingford, M., Ngo, H., Michel, O., Kusupati, A., Fan, A., Laforte, C., Voleti, V., Gadre, S.Y., et al.: Objverse-XL: A universe of \\(10\\mathrm{m} + 3\\mathrm{D}\\) objects. arXiv preprint arXiv:2307.05663 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.688, + 0.785, + 0.742 + ], + "angle": 0, + "content": "9. Deitke, M., Schwenk, D., Salvador, J., Weihs, L., Michel, O., VanderBilt, E., Schmidt, L., Ehsani, K., Kembhavi, A., Farhadi, A.: Objaverse: A universe of annotated 3D objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13142-13153 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.743, + 0.785, + 0.798 + ], + "angle": 0, + "content": "10. Downs, L., Francis, A., Koenig, N., Kinman, B., Hickman, R., Reymann, K., McHugh, T.B., Vanhoucke, V.: Google Scanned Objects: A high-quality dataset of 3D scanned household items. In: 2022 International Conference on Robotics and Automation (ICRA). pp. 2553-2560. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "11. Eftekhar, A., Sax, A., Bachmann, R., Malik, J., Zamir, A.: Omnidata: A scalable pipeline for making multi-task mid-level vision datasets from 3D scans. In: IEEE International Conference on Computer Vision (ICCV) (2021)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.311, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.372, + 0.127 + ], + "angle": 0, + "content": "V. Voleti et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.203 + ], + "angle": 0, + "content": "12. Girdhar, R., Singh, M., Brown, A., Duval, Q., Azadi, S., Rambhatla, S.S., Shah, A., Yin, X., Parikh, D., Misra, I.: EMU VIDEO: Factorizing Text-to-Video Generation by Explicit Image Conditioning (2023), https://emu-video.metademolab.com/ assets/emu(video.pdf" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.203, + 0.787, + 0.244 + ], + "angle": 0, + "content": "13. Hasselgren, J., Hofmann, N., Munkberg, J.: Shape, Light, and Material Decomposition from Images using Monte Carlo Rendering and Denoising. In: Advances in Neural Information Processing Systems (NeurIPS) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.244, + 0.787, + 0.271 + ], + "angle": 0, + "content": "14. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: Advances in Neural Information Processing Systems (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.271, + 0.787, + 0.312 + ], + "angle": 0, + "content": "15. Hong, Y., Zhang, K., Gu, J., Bi, S., Zhou, Y., Liu, D., Liu, F., Sunkavalli, K., Bui, T., Tan, H.: LRM: Large reconstruction model for single image to 3D. International Conference on Learning Representations (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.312, + 0.787, + 0.325 + ], + "angle": 0, + "content": "16. Jun, H., Nichol, A.: Shap-E: Generating conditional 3D implicit functions (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.325, + 0.787, + 0.365 + ], + "angle": 0, + "content": "17. Kanopoulos, N., Vasanthavada, N., Baker, R.L.: Design of an image edge detection filter using the sobel operator. IEEE Journal of solid-state circuits 23(2), 358-367 (1988)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.365, + 0.787, + 0.42 + ], + "angle": 0, + "content": "18. Kant, Y., Wu, Z., Vasilkovsky, M., Qian, G., Ren, J., Guler, R.A., Ghanem, B., Tulyakov, S., Gilitschenski, I., Siarohin, A.: Spad: Spatially aware multiview diffusers. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.42, + 0.787, + 0.446 + ], + "angle": 0, + "content": "19. Karras, T., Aittala, M., Aila, T., Laine, S.: Elucidating the Design Space of Diffusion-Based Generative Models. arXiv:2206.00364 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.446, + 0.787, + 0.473 + ], + "angle": 0, + "content": "20. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.787, + 0.501 + ], + "angle": 0, + "content": "21. Kong, X., Liu, S., Lyu, X., Taher, M., Qi, X., Davison, A.J.: Eschernet: A generative model for scalable view synthesis. arXiv preprint arXiv:2402.03908 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.787, + 0.542 + ], + "angle": 0, + "content": "22. Kwak, J.g., Dong, E., Jin, Y., Ko, H., Mahajan, S., Yi, K.M.: Vivid-1-to-3: Novel view synthesis with video diffusion models. IEEE conference on computer vision and pattern recognition (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.542, + 0.787, + 0.595 + ], + "angle": 0, + "content": "23. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3D: High-resolution text-to-3D content creation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 300–309 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.595, + 0.787, + 0.637 + ], + "angle": 0, + "content": "24. Liu, M., Shi, R., Chen, L., Zhang, Z., Xu, C., Wei, X., Chen, H., Zeng, C., Gu, J., Su, H.: One-2-3-45++: Fast single image to 3D objects with consistent multi-view generation and 3D diffusion. arXiv preprint arXiv:2311.07885 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.637, + 0.787, + 0.677 + ], + "angle": 0, + "content": "25. Liu, M., Xu, C., Jin, H., Chen, L., Varma T, M., Xu, Z., Su, H.: One-2-3-45: Any single image to 3D mesh in 45 seconds without per-shape optimization. Advances in Neural Information Processing Systems 36 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.677, + 0.787, + 0.717 + ], + "angle": 0, + "content": "26. Liu, R., Wu, R., Hoorick, B.V., Tokmakov, P., Zakharov, S., Vondrick, C.: Zero-1-to-3: Zero-shot one image to 3D object. International Conference on Computer Vision (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.718, + 0.787, + 0.758 + ], + "angle": 0, + "content": "27. Liu, Y., Lin, C., Zeng, Z., Long, X., Liu, L., Komura, T., Wang, W.: SyncDreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.787, + 0.8 + ], + "angle": 0, + "content": "28. Long, X., Guo, Y.C., Lin, C., Liu, Y., Dou, Z., Liu, L., Ma, Y., Zhang, S.H., Habermann, M., Theobalt, C., et al.: Wonder3D: Single image to 3D using cross-domain diffusion. arXiv preprint arXiv:2310.15008 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.787, + 0.84 + ], + "angle": 0, + "content": "29. Melas-Kyriazi, L., Laina, I., Rupprecht, C., Neverova, N., Vedaldi, A., Gafni, O., Kokkinos, F.: IM-3D: Iterative multiview diffusion and reconstruction for high-quality 3D generation. arXiv preprint arXiv:2402.08682 (2024)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.69, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SV3D" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "30. Mercier, A., Nakhli, R., Reddy, M., Yasarla, R., Cai, H., Porikli, F., Berger, G.: HexaGen3D: Stablediffusion is just one step away from fast and diverse Text-to-3D generation. arXiv preprint arXiv:2401.07727 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.788, + 0.244 + ], + "angle": 0, + "content": "31. Metzer, G., Richardson, E., Patashnik, O., Giryes, R., Cohen-Or, D.: Latent-NeRF for shape-guided generation of 3D shapes and textures. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12663-12673 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.244, + 0.788, + 0.285 + ], + "angle": 0, + "content": "32. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing scenes as neural radiance fields for view synthesis. In: European Conference on Computer Vision. pp. 405-421 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.285, + 0.788, + 0.311 + ], + "angle": 0, + "content": "33. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.311, + 0.788, + 0.338 + ], + "angle": 0, + "content": "34. Nichol, A., Jun, H., Dhariwal, P., Mishkin, P., Chen, M.: Point-E: A System for Generating 3D Point Clouds from Complex Prompts (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.338, + 0.788, + 0.392 + ], + "angle": 0, + "content": "35. Niemeyer, M., Barron, J.T., Mildenhall, B., Sajjadi, M.S.M., Geiger, A., Radwan, N.: RegNeRF: Regularizing neural radiance fields for view synthesis from sparse inputs. In: Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.392, + 0.788, + 0.433 + ], + "angle": 0, + "content": "36. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.433, + 0.788, + 0.459 + ], + "angle": 0, + "content": "37. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3D using 2D diffusion. arXiv (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.459, + 0.788, + 0.514 + ], + "angle": 0, + "content": "38. Qian, G., Mai, J., Hamdi, A., Ren, J., Siarohin, A., Li, B., Lee, H.Y., Skorokhodov, I., Wonka, P., Tulyakov, S., et al.: Magic123: One image to high-quality 3D object generation using both 2D and 3D diffusion priors. arXiv preprint arXiv:2306.17843 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.514, + 0.788, + 0.568 + ], + "angle": 0, + "content": "39. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.568, + 0.788, + 0.609 + ], + "angle": 0, + "content": "40. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.609, + 0.788, + 0.65 + ], + "angle": 0, + "content": "41. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dream-Booth: Fine tuning text-to-image dissusion models for subject-driven generation. arXiv preprint arXiv:2208.12242 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.65, + 0.788, + 0.704 + ], + "angle": 0, + "content": "42. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E., Kamyar Seyed Ghasemipour, S., Karagol Ayan, B., Mahdavi, S.S., Gontijo Lopes, R., Salimans, T., Ho, J., Fleet, D.J., Norouzi, M.: Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding. arXiv:2205.11487 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.759 + ], + "angle": 0, + "content": "43. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortzman, M., et al.: LAION-5B: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.759, + 0.788, + 0.8 + ], + "angle": 0, + "content": "44. Shen, T., Gao, J., Yin, K., Liu, M.Y., Fidler, S.: Deep Marching Tetrahedra: a hybrid representation for high-resolution 3D shape synthesis. In: Advances in Neural Information Processing Systems (NeurIPS) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.8, + 0.788, + 0.841 + ], + "angle": 0, + "content": "45. Shi, R., Chen, H., Zhang, Z., Liu, M., Xu, C., Wei, X., Chen, L., Zeng, C., Su, H.: Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.372, + 0.127 + ], + "angle": 0, + "content": "V. Voleti et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "46. Shi, Y., Wang, P., Ye, J., Long, M., Li, K., Yang, X.: MVDream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.204 + ], + "angle": 0, + "content": "47. Song, Y., Ermon, S.: Improved Techniques for Training Score-Based Generative Models. arXiv:2006.09011 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.206, + 0.785, + 0.247 + ], + "angle": 0, + "content": "48. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.249, + 0.785, + 0.274 + ], + "angle": 0, + "content": "49. StabilityAI: Stable Zero123 (2023), https://stability.ai/news/stable-zero123-3d-generation" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.276, + 0.785, + 0.303 + ], + "angle": 0, + "content": "50. Tang, J., Ren, J., Zhou, H., Liu, Z., Zeng, G.: DreamGaussian: Generative Gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.305, + 0.785, + 0.359 + ], + "angle": 0, + "content": "51. Tang, S., Chen, J., Wang, D., Tang, C., Zhang, F., Fan, Y., Chandra, V., Furukawa, Y., Ranjan, R.: Mvdiffusion++: A dense high-resolution multi-view diffusion model for single or sparse-view 3d object reconstruction. arXiv preprint arXiv:2402.12712 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.361, + 0.785, + 0.402 + ], + "angle": 0, + "content": "52. Tsai, Y.T., Shih, Z.C.: All-frequency precomputed radiance transfer using spherical radial basis functions and clustered tensor approximation. ACM Transactions on Graphics (ToG) (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.404, + 0.785, + 0.444 + ], + "angle": 0, + "content": "53. Voleti, V., Jolicoeur-Martineau, A., Pal, C.: MCVD: Masked conditional video diffusion for prediction, generation, and interpolation. In: (NeurIPS) Advances in Neural Information Processing Systems (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.446, + 0.785, + 0.472 + ], + "angle": 0, + "content": "54. Watson, D., Chan, W., Martin-Brualla, R., Ho, J., Tagliasacchi, A., Norouzi, M.: Novel view synthesis with diffusion models (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.474, + 0.785, + 0.5 + ], + "angle": 0, + "content": "55. Weng, H., Yang, T., Wang, J., Li, Y., Zhang, T., Chen, C.L.P., Zhang, L.: Consistent123: Improve consistency for one image to 3D object synthesis (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.503, + 0.785, + 0.557 + ], + "angle": 0, + "content": "56. Wu, T., Zhang, J., Fu, X., Wang, Y., Ren, J., Pan, L., Wu, W., Yang, L., Wang, J., Qian, C., et al.: Omniobject3D: Large-vocabulary 3D object dataset for realistic perception, reconstruction and generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 803-814 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.559, + 0.785, + 0.585 + ], + "angle": 0, + "content": "57. Yang, J., Cheng, Z., Duan, Y., Ji, P., Li, H.: ConsistNet: Enforcing 3D consistency for multi-view images diffusion. arXiv preprint arXiv:2310.10343 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.587, + 0.785, + 0.628 + ], + "angle": 0, + "content": "58. Yao, C.H., Raj, A., Hung, W.C., Rubinstein, M., Li, Y., Yang, M.H., Jampani, V.: ARTIC3D: Learning robust articulated 3D shapes from noisy web image collections. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.63, + 0.785, + 0.656 + ], + "angle": 0, + "content": "59. Ye, J., Wang, P., Li, K., Shi, Y., Wang, H.: Consistent-1-to-3: Consistent image to 3D view synthesis via geometry-aware diffusion models. In: 3DV (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.658, + 0.645, + 0.671 + ], + "angle": 0, + "content": "60. Young, J.: xatlas. https://github.com/jpcy/xatlas (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.673, + 0.785, + 0.713 + ], + "angle": 0, + "content": "61. Yu, Z., Peng, S., Niemeyer, M., Sattler, T., Geiger, A.: MonoSDF: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in Neural Information Processing Systems (NeurIPS) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.715, + 0.785, + 0.769 + ], + "angle": 0, + "content": "62. Zhang, K., Luan, F., Wang, Q., Bala, K., Snavely, N.: PhySG: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In: The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.771, + 0.785, + 0.811 + ], + "angle": 0, + "content": "63. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "64. Zheng, C., Vedaldi, A.: Free3D: Consistent novel view synthesis without 3D representation. arXiv preprint arXiv:2312.04551 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.689, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SV3D" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.222 + ], + "angle": 0, + "content": "In this supplemental document, we include the broader impact in Appendix A, implementation details in Appendices B to F, ablative analyses in Appendix G, and additional results in Appendix H. We also attach a teaser video summarizing the SV3D framework (SV3D(video.mp4), as well as an HTML page (SV3D.html) for more visual comparisons." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.247, + 0.414, + 0.267 + ], + "angle": 0, + "content": "A Broader Impact" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.28, + 0.788, + 0.385 + ], + "angle": 0, + "content": "The advancement of generative models in different media forms is changing how we make and use content. These AI-powered models can create images, videos, 3D objects, and more, in ways we've never seen before. They offer huge potential for innovation in media production. But along with this potential, there are also risks. Before we start using these models widely, it is crucial to make sure we understand the possible downsides and have plans in place to deal with them effectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.387, + 0.788, + 0.476 + ], + "angle": 0, + "content": "In the case of 3D object generation, the input provided by the user plays a crucial role in constraining the model's output. By supplying a full front view of an object, users limit the model's creative freedom to the visible or unoccluded regions, thus minimizing the potential for generating problematic imagery. Additionally, factors such as predicted depth values and lighting further influence the fidelity and realism of generated content." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.478, + 0.803, + 0.597 + ], + "angle": 0, + "content": "Moreover, ensuring the integrity and appropriateness of training data is critical in mitigating risks associated with generative models. Platforms like Sketchfab, which serve as repositories for 3D models used in training, enforce strict content policies to prevent the dissemination of \"Unacceptable Content\" and disallows it on their platform: https://help.sketchfab.com/hc/en-us/articles/214867883-What-is-Restricted-Content. By adhering to these guidelines and actively monitoring dataset quality, developers can reduce the likelihood of biased or inappropriate outputs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.788, + 0.674 + ], + "angle": 0, + "content": "Sketchfab also has a tag for \"Restricted Content\" which is deemed to be similar to the PG-13 content rating used in the US (i.e. inappropriate for children under 13). We have confirmed that none of the objects that we use in training have this flag set to true. Thus we go the extra step of excluding even tagged PG-13 content from the training set." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.788, + 0.735 + ], + "angle": 0, + "content": "There is a chance that certain content may not have been correctly labeled on Sketchfab. In cases where the uploader fails to tag an object appropriately, Sketchfab provides publicly accessible listings of objects and a mechanism for the community to report any content that may be deemed offensive." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.84 + ], + "angle": 0, + "content": "In our regular utilization of Objaverse content, we haven't observed any significant amount of questionable material. Nevertheless, there are occasional instances of doll-like nudity stemming from basic 3D models, which could be crucial for accurately depicting humanoid anatomy. Additionally, the training dataset contains some presence of drugs, drug paraphernalia, as well as a certain level of blood content and weaponry, resembling what might be encountered in a video game context. Should the model be provided with imagery featuring these cate-" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "20" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.372, + 0.127 + ], + "angle": 0, + "content": "V. Voleti et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.176 + ], + "angle": 0, + "content": "gories of content, it possesses the capability to generate corresponding 3D models to some extent." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.181, + 0.788, + 0.317 + ], + "angle": 0, + "content": "It is to be noted that SV3D mainly focuses on generating hidden details in the user's input image. If the image is unclear or some parts are hidden, the model guesses what those parts might look like based on its training data. This means it might create details similar to what it has seen before. The training data generally follows the standards of 3D modeling and gaming. However, this could lead to criticisms about the models being too similar to existing trends. But the user's input image limits how creative the model can be and reduces the chance of biases showing up in its creations, especially if the image is clear and straightforward." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.356, + 0.381, + 0.373 + ], + "angle": 0, + "content": "B Data Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.404, + 0.784, + 0.48 + ], + "angle": 0, + "content": "Similar to SVD-MV [2], we render views of a subset of the Objaverse dataset [9] consisting of 150K curated CC-licensed 3D objects from the original dataset. Each loaded object is scaled such that the largest world-space XYZ extent of its bounding box is 1. The object is then repositioned such that this bounding box is centered around the origin." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.483, + 0.785, + 0.543 + ], + "angle": 0, + "content": "For both the static and dynamic orbits, we use Blender's EEVEE renderer to render an 84-frame RGBA orbit at \\(576 \\times 576\\) resolution. During training, any 21-frame orbit can be subsampled from this by picking any frame as the first frame, and then choosing every 4th frame after that." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.547, + 0.788, + 0.623 + ], + "angle": 0, + "content": "We apply two background colors to each of these images: random RGB color, and white. This results in a doubling of the number of orbit samples for training. We then encode all of these images into latent space using SVD's VAE, and using CLIP. We then store the latent and CLIP embeddings for all of these images along with the corresponding elevation and azimuth values." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.626, + 0.788, + 0.702 + ], + "angle": 0, + "content": "For lighting, we randomly select from a set of 20 curated HDRI envmaps. Each orbit begins with the camera positioned at azimuth 0. Our camera uses a field-of-view of 33.8 degrees. For each object, we adaptively position the camera to a distance sufficient to ensure that the rendered object content makes good and consistent use of the image extents without being clipped in any view." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.84 + ], + "angle": 0, + "content": "For static orbits, the camera is positioned at a randomly sampled elevation between \\([-5, 30]\\) degrees. The azimuth steps by a constant \\(\\frac{360}{84}\\) degree delta between each frame. For dynamic orbits, the sequence of camera elevations for each orbit are obtained from a random weighted combination of sinusoids with different frequencies. For each sinusoid, the whole number period is sampled from [1, 5], the amplitude is sampled from [0.5, 10], and a random phase shift is applied. The azimuth angles are sampled regularly, and then a small amount of noise is added to make them irregular. The elevation values are smoothed using a simple convolution kernel and then clamped to a maximum elevation of 89 degrees." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.689, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SV3D" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "21" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.414, + 0.163 + ], + "angle": 0, + "content": "C Training Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.788, + 0.388 + ], + "angle": 0, + "content": "Our approach involves utilizing the widely used EDM [19] framework, incorporating a simplified diffusion loss for fine-tuning, as followed in SVD [2]. We eliminate the conditions of 'fps_id', 'motion_bucket_id', etc. since they are irrelevant for SV3D. Furthermore, we adjust the loss function to assign lower weights to frames closer to the front-view conditioning image, ensuring that challenging back views receive equal training focus as the easier front views. To optimize training efficiency and conserve GPU VRAM, we preprocess and store the precomputed latent and CLIP embeddings for all video frames in advance. During training, these tensors are directly loaded rather than being computed in real-time. We choose to finetune the SVD-xt model to output 21 frames instead of 25 frames. We found that with 21 frames we were able to fit a batch size of 2 on each GPU, instead of 1 with 25 frames at \\(576 \\times 576\\) resolution. All SV3D models are trained for 105k iterations with an effective batch size of 64 on 4 nodes of 880GB A100 GPUs for around 6 days." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.41, + 0.423, + 0.425 + ], + "angle": 0, + "content": "D Inference Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.44, + 0.787, + 0.486 + ], + "angle": 0, + "content": "To generate an orbital video during inference, we use 50 steps of the deterministic DDIM sampler [47] with the triangular classifier-free guidance (CFG) scale described in the main paper. This takes \\(\\approx 40\\) seconds for the SV3D model." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.508, + 0.657, + 0.525 + ], + "angle": 0, + "content": "E Additional Details on Illumination Model" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.539, + 0.788, + 0.631 + ], + "angle": 0, + "content": "We base our rendering model on Spherical Gaussians (SG) [4,62]. A SG at query location \\( \\pmb{x} \\in \\mathbb{R}^3 \\) is defined as \\( G(\\pmb{x};\\pmb{\\mu},c,a) = ae^{s(\\pmb{\\mu}\\cdot \\pmb{x} - 1)} \\), where \\( \\pmb{\\mu} \\in \\mathbb{R}^3 \\) is the axis, \\( s \\in \\mathbb{R} \\) the sharpness of the lobe, and \\( a \\in \\mathbb{R} \\) the amplitude. Here, we point out that we only model white light and hence only use a scalar amplitude. One particularly interesting property of SGs is that the inner product of two SGs is the integral of the product of two SGs. The operation is defined as [52]:" + }, + { + "type": "equation", + "bbox": [ + 0.244, + 0.64, + 0.785, + 0.71 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} G _ {1} (\\boldsymbol {x}) \\cdot G _ {2} (\\boldsymbol {x}) = \\int_ {\\Omega} G _ {1} (\\boldsymbol {x}) G _ {2} (\\boldsymbol {x}) d \\boldsymbol {x} = \\frac {1}{d _ {m}} \\left(2 \\pi a _ {1} a _ {2} e ^ {d _ {m} - \\lambda_ {m}} (1. 0 - e ^ {- 2 d _ {m}})\\right) \\\\ \\lambda_ {m} = \\lambda_ {1} - \\lambda_ {2} \\tag {1} \\\\ d _ {m} = \\left| \\left| \\lambda_ {1} \\boldsymbol {\\mu} _ {1} + \\lambda_ {2} \\boldsymbol {\\mu} _ {2} \\right| \\right|. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In our illumination model we only consider Lambertian shading. Here, the cosine shading term influences the output the most. This term can be approximated with another SG \\( G_{\\mathrm{cosine}} = (\\pmb{x}; \\pmb{n}, 2.133, 1.17) \\), where \\( \\pmb{n} \\) defines the surface normal at \\( \\pmb{x} \\). The lighting evaluation using SGs \\( G_i \\) then becomes: \\( L = \\sum_{i=1}^{24} \\frac{1}{\\pi} \\max(G_i \\cdot G_{\\mathrm{cosine}}, 0) \\). As defined previously this results in the full integration of incoming light for each SG and as light is additive evaluating and summing all SGs results in the complete environment illumination. This \\( L \\) is also used in the \\( \\mathcal{L}_{\\mathrm{illum}} \\) loss described in the main paper. The rendered textured" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "22" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.372, + 0.127 + ], + "angle": 0, + "content": "V. Voleti et al." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.146, + 0.785, + 0.193 + ], + "angle": 0, + "content": "image is then defined as \\(\\hat{I} = c_dL\\), where \\(c_{d}\\) is the diffuse albedo. We learn \\(\\mu, c, a\\) for each SG \\(G_{i}\\) using reconstruction loss between these rendered images and SV3D-generated images." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.215, + 0.676, + 0.234 + ], + "angle": 0, + "content": "F Losses and Optimization for 3D Generation" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.248, + 0.788, + 0.6 + ], + "angle": 0, + "content": "In addition to the masked SDS loss \\(\\mathcal{L}_{\\mathrm{mask - ds}}\\) and illumination loss \\(\\mathcal{L}_{\\mathrm{illum}}\\) detailed in the manuscript, we use several other losses for 3D reconstruction. Our main reconstruction losses are the pixel-level mean squared error \\(\\mathcal{L}_{\\mathrm{mse}} = \\| I - \\hat{I}\\|^2\\), LPIPS [63] loss \\(\\mathcal{L}_{\\mathrm{lpips}}\\), and mask loss \\(\\mathcal{L}_{\\mathrm{mask}} = \\| S - \\hat{S}\\|^2\\), where \\(S\\), \\(\\hat{S}\\) are the predicted and ground-truth masks. We further employ a normal loss using the estimated mono normal by Omnidata [11], which is defined as the cosine similarity between the rendered normal \\(\\pmb{n}\\) and estimated pseudo ground truths \\(\\bar{n}\\): \\(\\mathcal{L}_{\\mathrm{normal}} = 1 - (\\pmb{n} \\cdot \\bar{\\pmb{n}})\\). To regularize the output geometry, we apply a smooth depth loss inspired by RegNeRF [35]: \\(\\mathcal{L}_{\\mathrm{depth}}(i,j) = (d(i,j) - d(i,j + 1))^2 + (d(i,j) - (d(i + 1,j))^2\\), where \\(i,j\\) indicate the pixel coordinate. For surface normal we instead rely on a bilateral smoothness loss similar to [5]. We found that this is crucial to getting high-frequency details and avoiding over-smoothed surfaces. For this loss we compute the image gradients of the input image \\(\\nabla I\\) with a Sobel filter [17]. We then encourage the gradients of rendered normal \\(\\nabla n\\) to be smooth if (and only if) the input image gradients \\(\\nabla I\\) are smooth. The loss can be written as \\(\\mathcal{L}_{\\mathrm{bilateral}} = e^{-3\\nabla I}\\sqrt{1 + ||\\nabla n||}\\). We also found that adding a spatial smoothness regularization on the albedo is beneficial: \\(\\mathcal{L}_{\\mathrm{albedo}} = (c_d(x) - c_d(x + \\epsilon))^2\\), where \\(c_d\\) denotes the albedo, \\(x\\) is a 3D surface point, and \\(\\epsilon \\in \\mathbb{R}^3\\) is a normal distributed offset with a scale of 0.01. The overall objective is then defined as the weighted sum of these losses. All losses are applied in both coarse and fine stages, except that we only apply \\(\\mathcal{L}_{\\mathrm{mask - ds}}\\) in the last 200 iterations of the fine stage. We use an Adam optimizer [20] with a learning rate of 0.01 for both stages." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.621, + 0.543, + 0.64 + ], + "angle": 0, + "content": "G Additional Ablative Analyses" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.653, + 0.785, + 0.683 + ], + "angle": 0, + "content": "We conduct additional ablative analyses of our 3D generation pipeline in this section." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.706, + 0.383, + 0.721 + ], + "angle": 0, + "content": "G.1 SV3D Models" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.731, + 0.785, + 0.791 + ], + "angle": 0, + "content": "In Tab. 7, we compare the quantitative results using different SV3D models and training losses. Both 2D and 3D evaluation shows that \\(\\mathrm{SV3D}^p\\) is our best performing model, either for pure photometric reconstruction or SDS-based optimization." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.69, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SV3D" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "23" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.414, + 0.788, + 0.455 + ], + "angle": 0, + "content": "Table 7: Ablative results of different SV3D models and training losses. We show that our \\(\\mathrm{SV3D}^p\\) model with Photo+SDS losses achieves the best 2D and 3D metrics." + }, + { + "type": "table", + "bbox": [ + 0.24, + 0.469, + 0.764, + 0.567 + ], + "angle": 0, + "content": "
ModelTraining lossesLPIPS↓PSNR↑SSIM↑MSE↓CLIP-S↑CD↓3D IoU↑
SV3DuPhoto0.13215.9510.8270.0320.8730.0280.583
SV3DuPhoto+SDS0.13315.9570.8340.0310.8710.0270.589
SV3DcPhoto0.13515.8260.8320.0330.8710.0290.579
SV3DcPhoto+SDS0.13216.3730.8340.0270.8700.0270.584
SV3DpPhoto0.12416.8640.8410.0240.8750.0240.611
SV3DpPhoto+SDS0.11917.4050.8490.0210.8770.0240.614
" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "24" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.372, + 0.128 + ], + "angle": 0, + "content": "V. Voleti et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.489, + 0.163 + ], + "angle": 0, + "content": "G.2 Static v.s. Dynamic Orbits" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.172, + 0.788, + 0.28 + ], + "angle": 0, + "content": "We also compare the results using different camera orbits for 3D training in Tab. 8. The results show that using a dynamic orbit (sine-30) produces better 3D outputs compared to static orbit since it contains more information of the top and bottom views of the object. However, higher elevation (sine-50) tends to increase inconsistency between multi-view images, and thus resulting in worse 3D reconstruction. In our experiments, we find that setting the elevation within \\(\\pm 30\\) degree generally leads to desirable 3D outputs." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.304, + 0.789, + 0.348 + ], + "angle": 0, + "content": "Table 8: Ablative results of different reference orbits for 3D generation. We show that using a dynamic orbit (sine elevation) with moderate amplitude performs better than orbits with no or extreme elevation variations." + }, + { + "type": "table", + "bbox": [ + 0.266, + 0.359, + 0.74, + 0.417 + ], + "angle": 0, + "content": "
Training orbitLPIPS↓PSNR↑SSIM↑MSE↓CLIP-S↑CD↓3D IoU↑
Static0.12516.8210.8480.0250.8640.0280.610
Sine-300.11917.4050.8490.0210.8770.0240.614
Sine-500.12317.0570.8540.0250.8730.0260.609
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.469, + 0.416, + 0.484 + ], + "angle": 0, + "content": "G.3 Masked SDS Loss" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.494, + 0.789, + 0.617 + ], + "angle": 0, + "content": "Finally, we show the ablative results of our SDS loss in Tab. 9. We compare the results of 1) pure photometric losses, 2) with naive SDS loss (no masking), 3) with hard-masked SDS loss by thresholding the dot product of surface normal and camera viewing angle as visibility masks, and 4) with soft-masked SDS loss as described in the manuscript. Overall, adding SDS guidance from the SV3D model can improve the 2D metrics while maintaining similar 3D metrics. Our novel soft-masked SDS loss generally achieves the best results compared to other baselines." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.638, + 0.789, + 0.669 + ], + "angle": 0, + "content": "Table 9: Ablative analyses of Masked SDS loss. Overall, our soft-masked SDS loss leads to higher-quality mesh outputs in terms of most 2D and 3D metrics." + }, + { + "type": "table", + "bbox": [ + 0.226, + 0.679, + 0.782, + 0.751 + ], + "angle": 0, + "content": "
Training lossesLPIPS↓PSNR↑SSIM↑MSE↓CLIP-S↑CD↓3D IoU↑
Photo0.12416.8640.8410.0240.8750.0240.611
Photo+SDS (naive)0.12417.0070.8500.0240.8670.0250.615
Photo+SDS (hard masked)0.12417.3350.8450.0220.8770.0240.610
Photo+SDS (soft masked)0.11917.4050.8490.0210.8770.0240.614
" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.69, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SV3D" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "25" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.506, + 0.161 + ], + "angle": 0, + "content": "H Additional Visual Results" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.178, + 0.784, + 0.193 + ], + "angle": 0, + "content": "In this section, we show more results of novel view synthesis and 3D generation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.215, + 0.446, + 0.23 + ], + "angle": 0, + "content": "H.1 Novel View Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.24, + 0.785, + 0.286 + ], + "angle": 0, + "content": "We show the additional NVS results on OmniObject3D [56] and real-world images in Fig. 13 and Fig. 14, respectively. The generated novel multi-view images by SV3D are more detailed and consistent compared to prior state-of-the-arts." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.311, + 0.782, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.627, + 0.788, + 0.683 + ], + "angle": 0, + "content": "Fig. 13: NVS Results on OmniObject3D [56]. For each object, we show the input image and generated novel multi-view images by different methods. SV3D is able to produce images with consistent pose, color, and geometric details, which are closer the ground-truth renders." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "26" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.372, + 0.127 + ], + "angle": 0, + "content": "V. Voleti et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.342, + 0.788, + 0.598 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.611, + 0.788, + 0.641 + ], + "angle": 0, + "content": "Fig. 14: NVS Results on Real-World Images. Notice the consistent texture, geometry, and pose in SV3D NVS outputs compared to prior works." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.689, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SV3D" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "27" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.391, + 0.16 + ], + "angle": 0, + "content": "H.2 3D Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.172, + 0.788, + 0.264 + ], + "angle": 0, + "content": "We show the additional 3D generation results on OmniObject3D [56] and real-world images in Fig. 15 and Fig. 16, respectively. Thanks to the consistent multi-view images by SV3D and the novel Masked SDS loss, our 3D generations are detailed, high-fidelity, and generalizable to a wide range of images. Since Free3D [64] does not include a 3D generation method, we run our 3D pipeline on its generated multi-view images for fair 3D comparison." + }, + { + "type": "image", + "bbox": [ + 0.23, + 0.293, + 0.742, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.628, + 0.788, + 0.671 + ], + "angle": 0, + "content": "Fig. 15: Mesh Results on OmniObject3D [56]. Thanks to the consistent novel multi-view images generated by SV3D, our 3D meshes contain higher-fidelity details while having smooth surfaces." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "28" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.372, + 0.127 + ], + "angle": 0, + "content": "V. Voleti et al." + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.316, + 0.286, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.314, + 0.318, + 0.404, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.363, + 0.403, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.42, + 0.318, + 0.512, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.421, + 0.363, + 0.512, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.53, + 0.318, + 0.617, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.531, + 0.363, + 0.617, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.645, + 0.318, + 0.739, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.646, + 0.364, + 0.738, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.228, + 0.41, + 0.291, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.307, + 0.411, + 0.399, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.308, + 0.461, + 0.398, + 0.508 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.415, + 0.412, + 0.512, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.415, + 0.461, + 0.512, + 0.506 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.415, + 0.624, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.461, + 0.624, + 0.506 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.413, + 0.738, + 0.458 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.461, + 0.738, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.242, + 0.515, + 0.275, + 0.563 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.314, + 0.514, + 0.396, + 0.563 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.314, + 0.564, + 0.395, + 0.614 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.335, + 0.616, + 0.371, + 0.627 + ], + "angle": 0, + "content": "SV3D" + }, + { + "type": "image", + "bbox": [ + 0.424, + 0.513, + 0.504, + 0.562 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.424, + 0.563, + 0.503, + 0.614 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.434, + 0.616, + 0.495, + 0.627 + ], + "angle": 0, + "content": "EscherNet" + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.512, + 0.611, + 0.562 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.563, + 0.611, + 0.614 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.547, + 0.616, + 0.591, + 0.627 + ], + "angle": 0, + "content": "Free3D" + }, + { + "type": "image", + "bbox": [ + 0.637, + 0.515, + 0.728, + 0.562 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.637, + 0.565, + 0.729, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.647, + 0.615, + 0.736, + 0.626 + ], + "angle": 0, + "content": "Stable Zero123" + }, + { + "type": "image_caption", + "bbox": [ + 0.226, + 0.644, + 0.788, + 0.674 + ], + "angle": 0, + "content": "Fig. 16: Mesh Results on Real-World Images. Notice the accurate shape and details in our reconstructions even from the diverse images in-the-wild." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.69, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SV3D" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "29" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.383, + 0.495, + 0.537 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.382, + 0.785, + 0.539 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.543, + 0.788, + 0.613 + ], + "angle": 0, + "content": "Fig. 17: NVS and 3D Generation Results on Real-World Images. Visual results on real-world images from the MVImgNet (Yu et al. CVPR'23) and NAVI (Jampani et al. NeurIPS'22) datasets. Compared to Stable Zero123, SV3D generates more consistent novel multi-view images, thus resulting in higher-quality 3D meshes with smooth surface and detailed texture." + } + ] +] \ No newline at end of file diff --git a/2024/SV3D_ Novel Multi-view Synthesis and 3D Generation from a Single Image using Latent Video Diffusion/55466408-8e29-4541-9064-bae13a2c8fbf_origin.pdf b/2024/SV3D_ Novel Multi-view Synthesis and 3D Generation from a Single Image using Latent Video Diffusion/55466408-8e29-4541-9064-bae13a2c8fbf_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..bb04d5377d1de5574f45bfd5f2eaae0f30816331 --- /dev/null +++ b/2024/SV3D_ Novel Multi-view Synthesis and 3D Generation from a Single Image using Latent Video Diffusion/55466408-8e29-4541-9064-bae13a2c8fbf_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:84d883e3dbf7be0079a7c77c53ff80946659cef55f95256ffb7b8a9e051858e9 +size 15216450 diff --git a/2024/SV3D_ Novel Multi-view Synthesis and 3D Generation from a Single Image using Latent Video Diffusion/full.md b/2024/SV3D_ Novel Multi-view Synthesis and 3D Generation from a Single Image using Latent Video Diffusion/full.md new file mode 100644 index 0000000000000000000000000000000000000000..76d979157519060dfff920ebcd781b971eb01206 --- /dev/null +++ b/2024/SV3D_ Novel Multi-view Synthesis and 3D Generation from a Single Image using Latent Video Diffusion/full.md @@ -0,0 +1,487 @@ +# SV3D: Novel Multi-view Synthesis and 3D Generation from a Single Image using Latent Video Diffusion + +Vikram Voleti*, Chun-Han Yao*, Mark Boss*, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani + +![](images/83325efc4c0195ff4b6a0c61c5348cc40f69864fffe4d7858886e4b0a8f2a849.jpg) +Fig. 1: Stable Video 3D (SV3D). From a single image, SV3D generates consistent novel multi-view images. We then optimize a 3D representation with SV3D-generated views resulting in high-quality 3D meshes. + +Abstract. We present Stable Video 3D (SV3D) — a latent video diffusion model for high-resolution, image-to-multi-view generation of orbital videos around a 3D object. Recent works propose to adapt 2D generative models for novel view synthesis (NVS) and 3D optimization. However, these methods have several disadvantages due to limited views or inconsistent NVS, affecting the performance of 3D object generation. In this work, we propose SV3D that adapts image-to-video diffusion model for novel multi-view synthesis and 3D generation, thereby leveraging the generalization and multi-view consistency of the video models, while further adding explicit camera control for NVS. We also propose improved 3D optimization techniques for image-to-3D generation using SV3D and its NVS outputs. Extensive experiments on multiple datasets with 2D and 3D metrics and user study demonstrate SV3D's state-of-the-art performance on NVS as well as 3D reconstruction compared to prior works. + +Keywords: 3D synthesis $\cdot$ Video generation $\cdot$ Novel view synthesis + +# 1 Introduction + +Single-image 3D object reconstruction is a long-standing problem in computer vision with a wide range of applications in game design, AR/VR, e-commerce, robotics, etc. It is a highly challenging and ill-posed problem as it requires lifting 2D pixels to 3D space while also reasoning about the unseen portions of the object in 3D. + +Despite being a long-standing vision problem, it is only recently that practically useful results are being produced by leveraging advances in the generative AI. This is mainly made possible by the large-scale pretraining of generative models which enables sufficient generalization to various domains. A typical strategy is to use image-based 2D generative models (e.g., Imagen [42], Stable Diffusion (SD) [40]) to provide a 3D optimization loss function for unseen novel views of a given object [23,31,37]. In addition, several works repurpose these 2D generative models to perform novel view synthesis (NVS) from a single image [26,27,30,54], and then use the synthesized novel views for 3D generation. Conceptually, these works mimic a typical photogrammetry-based 3D object capture pipeline, i.e., first photographing multi-view images of an object, followed by 3D optimization; except that the explicit multi-view capture is replaced by novel-view synthesis using a generative model, either via text prompt or camera pose control. + +A key issue in these generation-based reconstruction methods is the lack of multi-view consistency in the underlying generative model, resulting in inconsistent novel views. Some works try to address this by jointly reasoning a 3D representation during NVS [6,15,27], but at the cost of high compute and data, often still resulting in inconsistent geometric and texture details. In this work, we tackle this issue by adapting a high-resolution, image-conditioned video diffusion model for NVS followed by 3D generation. + +Novel Multi-view Synthesis. We adapt a latent video diffusion model (Stable Video Diffusion - SVD [2]) to generate multiple novel views of a given object with explicit camera pose conditioning. SVD demonstrates excellent multi-view consistency for video generation, and we repurpose it for NVs. In addition, SVD also has good generalization capabilities as it is trained on large-scale image and video data that are more readily available than large-scale 3D data. In short, we adapt the video diffusion model for NVs from a single image with three useful properties for 3D object generation: pose-controllable, multi-view consistent, and generalizable. We call our resulting NVS network 'SV3D'. To our knowledge, this is the first work that adapts a video diffusion model for explicit pose-controlled view synthesis. Some contemporary works such as [2, 29] demonstrate this, but only generate orbital videos without any explicit camera control. + +3D Generation. We then utilize our SV3D model for 3D object generation by optimizing a NeRF and DMTet mesh in a coarse-to-fine manner. Benefiting from the multi-view consistency in SV3D, we are able to produce high-quality 3D meshes directly from the SV3D novel view images. We also design a masked score distillation sampling (SDS) [37] loss to further enhance 3D quality in the regions that are not visible in the SV3D-predicted novel views. In addition, we propose to jointly optimize a disentangled illumination model along with 3D shape and texture, effectively reducing the issue of baked-in lighting. + +We perform extensive comparisons of both our NVS and 3D generation results with respective state-of-the-art methods, demonstrating considerably better outputs with SV3D. For NVS, SV3D shows high-level of multi-view consistency and generalization to real-world images while being pose controllable. Our resulting 3D meshes are able to capture intricate geometric and texture details. + +# 2 Background + +# 2.1 Novel View Synthesis + +We organize the related works along three crucial aspects of novel view synthesis (NVS): generalization, controllability, and multi-view (3D) consistency. + +Generalization. Diffusion models [14, 48] have recently shown to be powerful generative models that can generate a wide variety of images [3, 40, 41] and videos [2, 12, 53] by iteratively denoising a noise sample. Among these models, the publicly available Stable Diffusion (SD) [40] and Stable Video Diffusion (SVD) [2] demonstrate strong generalization ability by being trained on extremely large datasets like LAION [43] and LVD [2]. Hence, they are commonly used as foundation models for various generation tasks, e.g. novel view synthesis. + +Controllability. Ideally, a controllable NVS model allows us to generate an image corresponding to any arbitrary viewpoint. For this, Zero123 [26] repurposes an image diffusion model to a novel view synthesizer, conditioned on a single-view image and the pose difference between the input and target views. Follow-up works such as Zero123XL [8] and Stable Zero123 [49] advance the quality of diffusion-based NVS, as well as the trained NeRFs using SDS loss. However, they only generate one novel view at a time, and thus are not designed to be multi-view consistent inherently. Recent works such as EscherNet [21] and Free3D [64] are capable of better multi-view consistency with intelligent camera position embedding design. However, they only use image-based diffusion models, and generate images at $256 \times 256$ resolution. We finetune a video diffusion model to generate novel views at $576 \times 576$ resolution. + +Multi-view Consistency. Multi-view consistency is the most critical requirement for high-quality NVS and 3D generation. To address the inconsistency issue, MVDream [46], SyncDreamer [27], HexGen3D [30], and Zero123++ [45] propose to generate multiple (specific) views of an object simultaneously. However, they are not controllable: given a conditional image, they only generate specific views, not arbitrary viewpoints. Moreover, they were finetuned from image-based diffusion models, i.e. multi-view consistency was imposed by adding interaction among the multiple generated views through cross-attention layers. Hence, their output quality is limited to the generalizability of their base image-based model, and their 3D finetuning dataset. Efficient-3DiM finetunes the SD model with a stronger vision transformer DINO v2 [36]. Consistent-1-to-3 [59] and SPAD [18] leverage epipolar geometry. One-2-3-45 [25] and One-2-3-45++ [24] train additional 3D network using the 2D generator's outputs. MVDream [46], Consistent123 [55], and Wonder3D [28] also train multi-view diffusion models, yet still require post-processing for video rendering. SyncDreamer [27] and ConsistentNet [57] employ 3D representation into the latent diffusion model. + +Exploiting Video Diffusion Models. To achieve better generalization and multi-view consistency, some contemporary works exploit the temporal priors in video diffusion models for NVs. For instance, Vivid-1-to-3 [22] combines a view-conditioned diffusion model and video diffusion model to generate consistent views. SVD-MV [2] and IM-3D [29] finetune a video diffusion model for NVs. + +However, they generate $\leq 360^{\circ}$ views at the same elevation only. Unlike SV3D, none of them are capable of rendering any arbitrary view of the 3D object. + +We argue that the existing NVS and 3D generation methods do not fully leverage the superior generalization capability, controllability, and consistency in video diffusion models. In this paper, we fill this important gap and train SV3D, a state-of-the-art novel multi-view synthesis video diffusion model at $576 \times 576$ resolution, and leverage it for 3D generation. + +# 2.2 3D Generation + +Recent advances in 3D representations and diffusion-based generative models have significantly improved the quality of image-to-3D generation. Here, we briefly summarize the related works in these two categories. + +3D Representation. 3D generation has seen great progress since the advent of Neural Radiance Fields (NeRFs) [32] and its subsequent variants [1], which implicitly represents a 3D scene as a volumetric function, typically parameterized by a neural network. Notably, Instant-NGP [33] introduces a hash grid feature encoding that can be used as a NeRF backbone for fast inference and ability to recover complex geometry. On the other hand, several recent works improve from an explicit representation such as DMTet [44], which is capable of generating high-resolution 3D shapes due to its hybrid SDF-Mesh representation and high memory efficiency. Similar to [23, 38], we adopt coarse-to-fine training for 3D generation, by first learning a rough object with Instant-NGP NeRF and then refining it using the DMTet representation. + +Diffusion-Based 3D Generation. Several recent works [16, 34] train a 3D diffusion model to to learn these flexible 3D representations, which, however, lack generalizability due to the scarcity of 3D data. To learn 3D generation without ground truth 3D data, image/multi-view diffusion models have been used as guidance. DreamFusion [37] and follow-up works [23, 31] leverage a trained image diffusion model as a 'scoring' function and calculate SDS loss for text-to-3D generation. However, they are prone to artifacts like Janus problem [31, 37] and over-saturated texture. Inspired by Zero123 [26], recent works [21, 24, 25, 27, 29, 38, 45, 49, 51, 64] finetune image/video diffusion models to generate novel view images as a stronger guidance. Our method shares the same spirit as this line of work, but produces denser, controllable, and more consistent multi-view images, thus resulting in better 3D generation quality. + +# 3 SV3D: Novel Multi-view Synthesis + +Our main idea is to repurpose temporal consistency in a video diffusion model for spatial 3D consistency of an object. Specifically, we finetune SVD [2] to generate an orbital video around a 3D object, conditioned on a single-view image. This orbital video need not be at the same elevation, or at regularly spaced azimuth angles. SVD is well-suited for this task since it is trained to generate smooth and consistent videos on large-scale datasets of real and high-quality videos. + +Fig. 2: SV3D Architecture. We add the sinusoidal embedding of the camera orbit elevation and azimuth angles $(e, a)$ to that of the noise step $t$ , and feed the sum to the convolutional blocks in the UNet. We feed the single input image's CLIP embedding to the attention blocks, and concatenate its latent embedding to the noisy state $\mathbf{z}_t$ . +![](images/12882c02dace176adb8a7308987935aef341e849263fcf5bef9dae01d0b9ae9f.jpg) +:FullyConnected :Embedding :VAE Encoder + +Static +![](images/6ad6fbf896b7ea7b537a65d35f76b1b60f5fab8ddfe00902463a218b7e50f49b.jpg) +We use two types of orbits for training the SV3D models. + +![](images/f238593dc45e29a47af60466306c47f3e94968dec88978f5a3e2b18e9b09197d.jpg) +Dynamic +Fig.3:Static vs. DynamicOrbits. + +The exposure to superior data quantity and quality makes it more generalizable and multi-view consistent, and the flexibility of the SVD architecture makes it amenable to be finetuned for camera controllability. + +Some prior works attempt to leverage such properties by finetuning image diffusion models, training video diffusion models from scratch, or finetuning video diffusion models to generate pre-defined views at the same elevation (static orbit) around an object [2, 29]. However, we argue that these methods do not fully exploit the potential of video diffusion models. To the best of our knowledge, SV3D is the first video diffusion-based framework for controllable multi-view synthesis at $576 \times 576$ resolution (and subsequently for 3D generation). + +Problem Setting. Formally, given an image $\mathbf{I} \in \mathbb{R}^{3 \times H \times W}$ of an object, our goal is to generate an orbital video $\mathbf{J} \in \mathbb{R}^{K \times 3 \times H \times W}$ around the object of $K = 21$ multi-view images along a camera trajectory $\pmb{\pi} \in \mathbb{R}^{K \times 2} = \{(e_i, a_i)\}_{i=1}^K$ as a sequence of $K$ tuples of elevation $e$ and azimuth $a$ angles. We assume the camera always looks at the center of an object (origin of world coordinates), so any viewpoint can be specified by only two parameters: $e$ and $a$ . We generate this orbital video by iteratively denoising samples from a learned conditional distribution $p(\mathbf{J}|\mathbf{I}, \pmb{\pi})$ , parameterized by a video diffusion model. + +SV3D Architecture. As shown in Fig. 2, the architecture of SV3D builds on that of SVD consisting of a UNet with multiple layers, each layer containing sequences of one residual block with Conv3D layers, and two transformer blocks (spatial and temporal) with attention layers. (i) We remove the vector conditionings of 'fps id' and 'motion bucket id' since they are irrelevant for SV3D. (ii) The conditioning image is concatenated to the noisy latent state input $\mathbf{z}_t$ at noise timestep $t$ to the UNet, after being embedded into latent space by the VAE encoder of SVD. (iii) The CLIP-embedding [39] matrix of the conditioning image is provided to the cross-attention layers of each transformer block as its key and + +![](images/cc3a266b1abb00711f934e48ce92e388c4a3f9fd8edb84a2c3658e78cbe0e122.jpg) +Fig. 4: Linear vs. Triangle CFG Scaling. Notice Fig. 5: LPIPS vs. Frame increased oversharping in the penultimate frame in Number. SV3D has the best re-linear scaling vs. our proposed triangle scaling. construction metric per frame. + +![](images/ae353a2baa0034bb01d4ff5f7d14c1f6e1b9128a12167fafbda9d0c558ca3168.jpg) + +value, the query being the feature at that layer. (iv) The camera trajectory is fed into the residual blocks along with the diffusion noise timestep. The camera pose angles $e_i$ and $a_i$ and the noise timestep $t$ are first embedded into sinusoidal position embeddings. Then, the camera pose embeddings are concatenated together, linearly transformed, and added to the noise timestep embedding. This is fed to every residual block, where they are added to the block's output feature (after being linearly transformed again to match the feature size). + +Static v.s. Dynamic Orbits. As shown in Fig. 3, we design static and dynamic orbits to study the effects of camera pose conditioning. In a static orbit, the camera circles around an object at regularly-spaced azimuths at the same elevation angle as that in the conditioning image. The disadvantage with the static orbit is that we might not get any information about the top or bottom of the object depending on the conditioning elevation angle. In a dynamic orbit, the azimuths can be irregularly spaced, and the elevation can vary per view. To build a dynamic orbit, we sample a static orbit, add small random noise to the azimuth angles, and add a random weighted combination of sinusoids with different frequencies to its elevation. This provides temporal smoothness, and ensures that the camera trajectory loops around to end at the same azimuth and elevation as those of the conditioning image. + +Thus, with this strategy, we are able to tackle all three aspects of generalization, controllability, and consistency in novel multi-view synthesis by leveraging video diffusion models, additionally conditioning on camera trajectory, and repurposing temporal consistency for spatial 3D object consistency, respectively. + +Triangular CFG Scaling. SVD uses a linearly increasing scale for classifier-free guidance (CFG) from 1 to 4 across the generated frames. However, this scaling causes the last few frames in our generated orbits to be over-sharpened, as shown in Fig. 4 Frame 20. Since we generate videos looping back to the front-view image, we propose to use a triangle wave CFG scaling during inference: linearly increase CFG from 1 at the front view to 2.5 at the back view, then linearly decrease it back to 1 at the front view. Fig. 4 also demonstrates that our triangle CFG scaling produces more details in the back view (Frame 12). + +Models. We train three image-to-3D-video models finetuned from SVD. First, we train a pose-unconditioned model, $\mathrm{SV3D}^u$ , which generates a video of static orbit around an object while only conditioned on a single-view image. Note that unlike SVD-MV [2], we do not provide the elevation angle to the pose-unconditioned model, as we find that the model is able to infer it from the conditioning image. Our second model, the pose-conditioned $\mathrm{SV3D}^c$ is conditioned on the input image as well as a sequence of camera elevation and azimuth angles in an orbit, trained on dynamic orbits. Following SVD's [2] intuition to progressively increase the task difficulty during training, we train our third model, $\mathrm{SV3D}^p$ , by first finetuning SVD to produce static orbits unconditionally, then further finetuning on dynamic orbits with camera pose condition. + +Training Details. We train SV3D on the Objverse dataset [9], which contains synthetic 3D objects covering a wide diversity. For each object, we render 21 frames around it on random color background at $576 \times 576$ resolution, field-of-view of 33.8 degrees. We choose to finetune the SVD-xt model to output 21 frames. All three models (SV3D $^u$ , SV3D $^c$ , SV3D $^p$ ) are trained for 105k iterations in total (SV3D $^p$ is trained unconditionally for 55k iterations and conditionally for 50k iterations), with an effective batch size of 64 on 4 nodes of 8 80GB A100 GPUs for around 6 days. For more training details, please see the appendix. + +# 3.1 Experiments and Results + +Datasets. We evaluate the SV3D-generated multi-view images on static and dynamic orbits on the unseen GSO [10] and OmniObject3D [56] datasets. Since many GSO objects are the same items with slightly different colors, we filter 300 objects from GSO to reduce redundancy and maintain diversity. For each object, we render ground truth static and dynamic orbit videos and pick the last frame of each video as the conditioning image. We also conduct a user study on novel view videos from a dataset of 22 real-world images. (More details in appendix.) + +Metrics. We use the SV3D models to generate static and dynamic orbit videos corresponding to the ground truth camera trajectories in the evaluation datasets. We compare each generated frame with the corresponding ground truth frames, in terms of Learned Perceptual Similarity (LPIPS [63]), Peak Signal-to-Noise Ratio (PSNR), Structural SIMilarity (SSIM), Mean Squared-Error (MSE), and CLIP-score (CLIP-S). This range of metrics covers both pixel-level as well as semantic aspects. Note that testing on dynamic orbits evaluates the controllability of SV3D models, and all the metrics evaluate multi-view consistency. + +Baselines. We compare SV3D with several recent NVS methods capable of generating arbitrary views, including Zero123 [26], Zero123-XL [8], SyncDreamer [27], Stable Zero123 [49], Free3D [64], EscherNet [21]. + +Results. As shown in Tabs. 1 to 4, our SV3D models achieve state-of-the-art performance on novel multi-view synthesis. Tabs. 1 and 3 show results on static orbits, and include all our three models. We see that even our pose-unconditioned model $\mathrm{SV3D}^u$ performs better than all prior methods. Tabs. 2 and 4 show results on dynamic orbits, and include our pose-conditioned models $\mathrm{SV3D}^c$ and $\mathrm{SV3D}^p$ . + +![](images/0e2e78ba032f8a75d450521803b42a294afe1fbb9621377fe07ebec3834f8d66.jpg) +Fig.6: Visual Comparison of Novel Multi-view Synthesis. SV3D is able to generate novel multi-views that are more detailed, faithful to the conditioning image, and multi-view consistent compared to the prior works. + +Ablative Analyses. Interestingly, from Tabs. 1 and 3, we find that both $\mathrm{SV3D}^c$ and $\mathrm{SV3D}^p$ outperform $\mathrm{SV3D}^u$ on generations of static orbits, even though $\mathrm{SV3D}^u$ is trained specifically on static orbits. We also observe that $\mathrm{SV3D}^p$ achieves better metrics than $\mathrm{SV3D}^c$ on both static and dynamic orbits, making it the best performing SV3D model overall. This shows that progressive finetuning from easier (static) to harder (dynamic) tasks is indeed a favorable way to finetune a video diffusion model. + +Visual Comparisons in Fig. 6 further demonstrate that SV3D-generated images are more detailed, faithful to the conditioning image, and multi-view consistent compared to the prior works. + +Quality Per Frame. We also observe from Fig. 5 that SV3D produces better quality at every frame. We plot the average LPIPS value for each generated frame, across generated GSO static orbit videos. The quality is generally worse around the back view, and better at the beginning and the end (i.e. near the conditioning image), as expected. + +User Study on Real-World Images. We conducted a user survey to study human preference between static orbital videos generated by SV3D and by other methods. We asked 30 users to pick one between our SV3D-generated static video and other method-generated video as the best orbital video for the corresponding image, using 22 real-world images. We noted that users preferred SV3D-generated videos over Zero123XL, Stable Zero123, EscherNet, and Free3D, $96\%$ , $99\%$ , $96\%$ , and $98\%$ of the time, respectively. + +Table 1: Evaluation of novel multi-view synthesis on GSO static orbits + +
ModelLPIPS↓PSNR↑SSIM↑CLIP-S↑MSE↓
SyncDreamer [27]0.1715.780.760.870.03
Zero123 [26]0.1317.290.790.850.04
Zero123XL [8]0.1417.110.780.850.04
Stable Zero123 [49]0.1318.340.780.850.05
Free3D [64]0.1516.180.790.840.04
EscherNet [21]0.1316.730.790.850.03
SV3Du0.0921.140.870.890.02
SV3Dc0.0920.560.870.880.02
SV3Dp0.0821.260.880.890.02
+ +Table 2: Evaluation of novel multi-view synthesis on GSO dynamic orbits + +
ModelLPIPS↓PSNR↑SSIM↑CLIP-S↑MSE↓
Zero123 [26]0.1416.990.790.840.04
Zero123XL [8]0.1416.730.780.840.04
Stable Zero123 [49]0.1318.040.780.850.05
Free3D [64]0.1814.930.770.830.05
EscherNet [21]0.1316.470.790.840.03
SV3Dc0.1019.990.860.870.02
SV3Dp0.0920.380.870.870.02
+ +Table 3: Evaluation of novel multi-view synthesis on OmniObject3D static orbits + +
ModelLPIPS↓PSNR↑SSIM↑CLIP-S↑MSE↓
Zero123 [26]0.1715.500.760.830.05
Zero123XL [8]0.1815.360.750.830.06
Stable Zero123 [49]0.1516.860.770.840.06
Free3D [64]0.1615.290.780.830.05
EscherNet [21]0.1714.630.740.830.05
SV3Du0.1019.680.860.860.02
SV3Dc0.1019.500.850.850.02
SV3Dp0.1019.910.860.860.02
+ +Table 4: Evaluation of novel multi-view synthesis on OmniObject3D dynamic orbits + +
ModelLPIPS↓PSNR↑SSIM↑CLIP-S↑MSE↓
Zero123 [26]0.1615.780.770.820.05
Zero123XL [8]0.1715.490.760.830.05
Stable Zero123 [49]0.1416.740.770.830.05
Free3D [64]0.1914.280.760.820.06
EscherNet [21]0.1615.050.760.830.05
SV3Dc0.1019.210.850.840.02
SV3Dp0.1019.280.850.850.02
+ +# 4 3D Generation from a Single Image Using SV3D + +We then generate 3D meshes of objects from a single image by leveraging SV3D. One way is to use the generated static/dynamic orbital samples from SV3D as direct reconstruction targets. Another way is to use SV3D as diffusion guidance with Score Distillation Sampling (SDS) loss [37]. + +Since SV3D produces more consistent multi-views compared to existing works, we already observe higher-quality 3D reconstructions by only using SV3D outputs for reconstruction when compared to existing works. However, we observe that this naive approach often leads to artifacts like baked-in illumination, rough surfaces, and noisy texture, especially for the unseen regions in the reference orbit. Thus, we further propose several techniques to address these issues. + +Coarse-to-Fine Training. We adopt a two-stage, coarse-to-fine training scheme to generate a 3D mesh from input images, similar to [23, 38]. Fig. 7 illustrate an overview of our 3D optimization pipeline. In the coarse stage, we train an Instant-NGP [33] NeRF representation to reconstruct the SV3D-generated images (i.e. without SDS loss) at a lower resolution. In the fine stage, we extract a mesh from the trained NeRF using marching cubes [7], and adopt a DMTet [44] representation to finetune the 3D mesh using SDS-based diffusion guidance from SV3D at full-resolution. Finally, we use xatlas [60] to perform the UV unwrapping and export the output object mesh. + +Disentangled Illumination Model. Similar to other recent 3D object generation methods [23, 31, 37], our output target is a mesh with a diffuse texture. Typically, such SDS-based optimization techniques use random illuminations at every iteration. However, our SV3D-generated videos are under consistent illumination, i.e., the lighting remains static while the camera circles around an object. + +![](images/5987a83388c37819313fdba8e4461418165ad0f7445d54136e20437a06efa447.jpg) +Fig.7: 3D Optimization Pipeline. We perform a two-stage optimization. In the short first stage, we extract the general shape, texture and illumination from the SV3D generated multi-view images. In the second stage, we extract a mesh with marching cubes and use DMTet to further optimize the shape, texture and illumination. We not only use the SV3D-generated images but a soft-masked SDS loss for unseen areas. Dashed red lines represent backpropagation into the differentiable rendering pipeline. + +![](images/5fb4bcf394d5d910bdad223a29f52256c4b8375432c15acf8e41f9abc8b6df6b.jpg) + +![](images/b1fd04eee9a26b38127a2922f27179c09b2c5f689a5b2f9845930264b6ec190c.jpg) +Ref. images in static orbit + +![](images/671b192150010ac14eb20f5708976a826ad85e32c45dd3ec5c955b6436252316.jpg) + +![](images/0aa46e836604df6c791f979614c2ccd5b4bc2c345168216bf0c4ab3f241d5b3a.jpg) +3D from static orbit + +![](images/79e021486db2f103fbc401784306cab1644763354e4e721b7bbd123d2021c78a.jpg) + +![](images/447888e32730a3ab3a5d887bb2a79dfe40d3e06d6d5271320cb26db27edd5135.jpg) +3D from dynamic orbit + +![](images/4f0ed928839d0d8218b7518328d81d2b74e9c02a17aaff33a944f63e08ca4072.jpg) + +![](images/aedeea4d1746ea26fe4ca91a08ab571a565b166952c5fad0f90461f52eb22694.jpg) + +![](images/3d48c3019a3a7f02f19c339f810d8c6b89cd3e8d3bb0774746da33172408360f.jpg) + +![](images/789b011b0b7a02615f4bbbd432e0296095ac0502f6de6847d394c02e962a9a5b.jpg) +Fig.8: Influence of Training Orbits. We show that using a dynamic orbit is crucial to 3D generations that are complete from diverse views. +No SDS loss +Fig.9: Influence of SDS. Using our masked SDS loss, we are able to fill in unseen surfaces in the training orbit, producing a cleaner result without oversaturation or blurry artifacts caused by naive SDS. + +![](images/d40e9b9de18b4d2252ca9c1c6b35d420a5ffeb0d43704dcb14aa16324e004ee7.jpg) +SDS loss + +![](images/cada0c9d47e64e8fc135525873334996b3cfe6ceedb5abaa1f14b17b5ab85e30.jpg) +Masked SDS loss + +Hence, to disentangle lighting effects and obtain a cleaner texture, we propose to fit a simple illumination model of 24 Spherical Gaussians (SG) inspired by prior decomposition methods [4,62]. + +We model white light and hence only use a scalar amplitude for the SGs. We only consider Lambertian shading, where the cosine shading term is approximated with another SG. We learn the parameters of the illumination SGs using a reconstruction loss between the rendered images and SV3D-generated images. + +Inspired by [13, 37] we reduce baked-in illumination with a loss term that replicates the HSV-value component of the input image $\mathbf{I}$ with the rendered illumination $L$ : $\mathcal{L}_{\mathrm{illum}} = |V(\mathbf{I}) - L|^2$ , with $V(\pmb {c}) = \max (c_r,c_g,c_b)$ . Given these changes, our disentangled illumination model is able to express lighting variation properly and can severely reduce baked-in illumination. Fig. 10 shows sample + +reconstructions with and without our illumination modelling. From the results, it is clear that we are able to disentangle the illumination effects from the base color (e.g., dark side of the school bus). + +# 4.1 3D Optimization Strategies and Losses + +Reconstruction via Photometric Losses. Intuitively, we can treat the SV3D-generated images as multi-view pseudo-ground truth, and apply 2D reconstruction losses to train the 3D models. In this case, we apply photometric losses on the rendered images from NeRF or DMTet, including the pixel-level MSE loss, mask loss, and a perceptual LPIPS [63] loss. These photometric losses also optimize our illumination model through the differential rendering pipeline. + +Training Orbits. For 3D generation, we use SV3D to generate multi-view images following a camera orbit $\pi_{\mathrm{ref}}$ , referred to as the reference orbit (also see Fig. 7 for the overview). Fig. 8 shows sample reconstruction with using static and dynamic orbital outputs from SV3D. Using a dynamic orbit for training is crucial to high-quality 3D outputs when viewed from various angles, since some top/bottom views are missing in the static orbit (fixed elevation). Hence, for $\mathrm{SV3D}^c$ and $\mathrm{SV3D}^p$ , we render images on a dynamic orbit whose elevation follows a sine function to ensure that top and bo + +![](images/d59104d1140387572463183ae508657a044358a3d5ca9d0739bf10a2b1286687.jpg) +Constant Illumination + +![](images/e2a63af3ffafaff4160f5fadf2e928ec4b17d3f63c5f1a2b3131aa56437ddefe.jpg) +SGs Illumination +Fig.10: Constant vs. SGs Illumination. Our SGs-based reconstructions do not exhibit darkening on the side of the bus, which enables easier and more convincing relighting for downstream applications. + +SV3D-Based SDS Loss. In addition to the reconstruction losses, we can also use SV3D via score-distillation sampling (SDS) [37, 58]. Fig. 9 shows sample reconstructions with and without using SDS losses. As shown in Fig. 9 (left), although training with a dynamic orbit improves overall visibility, we observe that sometimes the output texture is still noisy, perhaps due to partial visibility, self-occlusions, or inconsistent texture/shape between images. Hence, we handle those unseen areas using SDS loss [37] with SV3D as a diffusion guidance. + +We sample a random camera orbit $\pi_{\mathrm{rand}}$ , and use our 3D NeRF/DMTet parameterized by $\theta$ to render the views $\hat{\mathbf{J}}$ of the object along $\pi_{\mathrm{rand}}$ . Then, noise $\epsilon$ at level $t$ is added to the latent embedding $\mathbf{z}_t$ of $\hat{\mathbf{J}}$ , and the following SDS loss (taken expectation over $t$ and $\epsilon$ ) is backpropagated through the differentiable rendering pipeline: $\mathcal{L}_{\mathrm{sds}} = w(t)\bigl (\epsilon_{\phi}(\mathbf{z}_t;\mathbf{I},\pi_{\mathrm{rand}},t) - \epsilon \bigr)\frac{\partial\hat{\mathbf{J}}}{\partial\theta}$ , where $w$ is $t$ -dependent weight, $\epsilon$ and $\epsilon_{\phi}$ are the added and predicted noise, and $\phi$ and $\theta$ are the parameters of SV3D and NeRF/DMTet, respectively. See Fig. 7 for an illustration of these loss functions in the overall pipeline. + +Masked SDS Loss. As shown in Fig. 9 (middle), in our experiments we found that adding the SDS loss naively can cause unstable training and unfaithful texture to the input images such as oversaturation or blurry artifacts. Therefore, we design a soft masking mechanism to only apply SDS loss on the unseen/occluded areas, allowing it to inpaint the missing details while preserving the texture of + +![](images/e96e640b561758b8f3934e1a727a4386cf96519b602cbeaf342608a3fdc58e0a.jpg) +Fig. 11: Visual Comparison of 3D Meshes. For each object, we show the conditioning image (left), and the output meshes rendered in two different views. Our generated meshes are more detailed, faithful to input images, and consistent in 3D. This demonstrates the quality of novel multi-view synthesis by our SV3D model. + +clearly-visible surfaces in the training orbit (as seen in Fig. 9 (right)). Also, we only apply the masked SDS loss in the final stage of DMTet optimization, which greatly increased the convergence speed. + +We apply SDS loss on only those pixels in the random orbit views that are not likely visible in the reference orbit views. For this, we first render the object from the random camera orbit $\pi_{\mathrm{rand}}$ . For each random camera view, we obtain the visible surface points $p \in \mathbb{R}^3$ and their corresponding surface normals $\boldsymbol{n}$ . Then, for each reference camera $i$ , we calculate the view directions $\boldsymbol{v}_i$ of the surface $\boldsymbol{p}$ towards its position $\bar{\pi}_{\mathrm{ref}}^i \in \mathbb{R}^3$ (calculated from $\pi_{\mathrm{ref}}^i \in \mathbb{R}^2$ ) as $\boldsymbol{v}_i = \frac{\bar{\pi}_{\mathrm{ref}}^i - \boldsymbol{p}}{||\bar{\pi}_{\mathrm{ref}}^i - \boldsymbol{p}||}$ . We infer the visibility of this surface from the reference camera based on the dot product between $\boldsymbol{v}_i$ and $\boldsymbol{n}$ i.e. $\boldsymbol{v}_i \cdot \boldsymbol{n}$ . Since higher values roughly indicate more visibility of the surface from the reference camera, we chose that reference camera $c$ that has maximum likelihood of visibility: $c = \max_i (\boldsymbol{v}_i \cdot \boldsymbol{n})$ . As we only want to apply SDS loss to unseen or grazing angle areas from $c$ , we use the smoothstep function $f_s$ to smoothly clip to $c$ 's visibility range $\boldsymbol{v}_c \cdot \boldsymbol{n}$ . In this way, we create a pseudo visibility mask $M = 1 - f_s(\boldsymbol{v}_c \cdot \boldsymbol{n}, 0, 0.5)$ , where $f_s(x; f_0, f_1) = \hat{x}^2(3 - 2x)$ with $\hat{x} = \frac{x - f_0}{f_1 - f_0}$ . Thus, $M$ is calculated for each random camera render, and the combined visibility mask $\mathbf{M}$ is applied to SDS loss: $\mathcal{L}_{\mathrm{mask-sds}} = \mathbf{M}\mathcal{L}_{\mathrm{sds}}$ . + +Geometric Priors. Since our rendering-based optimization operates at the image level, we adopt several geometric priors to regularize the output shapes. We add a smooth depth loss from RegNeRF [35] and a bilateral normal smoothness loss [5] to encourage smooth 3D surfaces where the projected image gradients are low. Moreover, we obtain normal estimates from Omnidata [11] and calculate a mono normal loss similar to MonoSDF [61], which can effectively reduce noisy surfaces in the output mesh. Further details about the training losses and optimization process are available in the appendix. + +# 4.2 Experiments and Results + +Due to the strong regularization, we only require 600 steps in the coarse stage and 1000 in the fine stage. Overall, the entire mesh extraction process takes $\approx 8$ + +![](images/48c824298c93e3d130e70bc250a70665f6a84254b1899ea3a6d012afe34e0102.jpg) +Fig. 12: Real-World 3D Results. Notice the accurate shape and details in our reconstructions even from the diverse images in-the-wild. + +minutes without SDS loss, and $\approx 20$ minutes with SDS loss. The coarse stage only takes $\approx 2$ minutes and provides a full 3D representation of the object. + +Evaluation. We evaluate our 3D generation framework on 50 randomly sampled objects from the GSO dataset as described in Sec. 3.1. We compute image-based reconstruction metrics (LPIPS, PSNR, SSIM, MSE, and CLIP-S) between the ground truth (GT) GSO images, and rendered images from the trained 3D meshes on the same static/dynamic orbits. In addition, we compute 3D reconstruction metrics of Chamfer distance (CD) and 3D IoU between the GT and predicted meshes. We compare our SV3D-guided 3D generations with several prior methods including Point-E [34], Shap-E [16], One-2-3-45++ [24], DreamGaussian [50], SyncDreamer [27], EscherNet [21], Free3D [64], and Stable Zero123 [49]. + +Visual Comparison. In Fig. 11, we show visual comparison of our results with those from prior methods. Qualitatively, Point-E [34] and Shap-E [16] often produce incomplete 3D shapes. DreamGaussian [50], SyncDreamer [27], EscherNet [21], and Free3D [64] outputs tend to contain rough surfaces. One-2-3-45++ [24] and Stable Zero123 [49] are able to reconstruct meshes with smooth surface, but lack geometric details. Our mesh outputs are detailed, faithful to input image, and consistent in 3D (see appendix for more examples). We also show 3D mesh renders from real-world images in Fig. 12. + +Quantitative Comparison. Tabs. 5 and 6 show the 2D and 3D metric comparisons respectively. All our 3D models achieve better 2D and 3D metrics compared to the prior and concurrent methods, showing the high-fidelity texture and geometry of our output meshes. We render all 3D meshes on the same dynamic orbits and compare them against the GT renders. Our best model, $\mathrm{SV3D}^p$ , performs comparably to using GT renders for reconstruction in terms of the 3D metrics, which further demonstrates the 3D consistency of our generated images. + +Effects of Photometric and (Masked) SDS Loss. As shown in Tabs. 5 and 6, the 3D outputs using both photometric and Masked SDS losses ( $\mathbf{SV3D}^p$ ) achieves the best results, while training without SDS ( $\mathbf{SV3D}^p$ no SDS') leads to marginally lower performance. This demonstrates that the images generated by SV3D are high-quality reconstruction targets, and are often sufficient for 3D generation without the cumbersome SDS-based optimization. Nevertheless, adding SDS generally improves quality, as also shown in Fig. 9. + +Effects of SV3D Model and Training Orbit. As shown in Tabs. 5 and 6, $\mathrm{SV3D}^p$ achieves the best performance among the three SV3D models, indicating that its synthesized novel views are most faithful to the input image and consis + +Table 5: 2D comparison of our 3D outputs against prior methods on the GSO dataset. Our best performing method uses $\mathbf{SV3D}^p$ with dynamic (sine elevation) orbit and SDS guidance. Note that all our models achieve better 2D metrics than prior works. + +
ModelLPIPS↓PSNR↑SSIM↑MSE↓CLIP-S↑
GT renders0.07819.5080.8790.0140.897
EscherNet [21]0.17814.4380.8040.0410.835
Free3D [64]0.19714.2020.7990.0430.809
Stable Zero123 [49]0.16614.6350.8130.0400.805
SV3Du0.13315.9570.8340.0310.871
SV3Dc0.13216.3730.8340.0270.870
SV3Dpstatic orbit0.12516.8210.8480.0250.864
SV3Dpno SDS0.12416.8640.8410.0240.875
SV3Dp0.11917.4050.8490.0210.877
+ +Table 6: Comparison of 3D metrics on the GSO dataset. Our models perform favorably against prior works. + +
ModelCD↓3D IoU↑
GT renders0.0210.689
Point-E [34]0.0740.162
Shap-E [16]0.0710.267
DreamGaussian [50]0.0550.411
One-2-3-45++ [24]0.0540.406
SyncDreamer [27]0.0530.451
EscherNet [21]0.0420.466
Free3D [64]0.0470.426
Stable Zero123 [49]0.0390.550
SV3Du0.0270.589
SV3Dc0.0270.584
SV3Dp static orbit0.0280.610
SV3Dp no SDS0.0240.611
SV3Dp0.0240.614
+ +tent in 3D. On the other hand, $\mathrm{SV3D}^u$ shares the same disadvantage as several prior works in that it can only generate views at the same elevation, which is insufficient to build a legible 3D object, as shown in Fig. 8. This also leads to the worse performance of 'SV3D $^p$ with static orbit' in Tabs. 5 and 6. Overall, $\mathrm{SV3D}^p$ with dynamic orbit and masked SDS loss performs favorably against all other configurations since it can leverage more diverse views of the object. + +Limitations. Our SV3D model is by design only capable of handling 2 degrees of freedom: elevation and azimuth; which is usually sufficient for 3D generation from a single image. One may want to tackle more degrees of freedom in cameras for a generalized NVS system, which forms an interesting future work. We also notice that SV3D exhibits some view inconsistency for mirror-like reflective surfaces, which provide a challenge to 3D reconstruction. Lastly, such reflective surfaces are not representable by our Lambertian reflection-based shading model. Conditioning SV3D on the full camera matrix, and extending the shading model are interesting directions for future research. + +# 5 Conclusion + +We present SV3D, a latent video diffusion model for novel multi-view synthesis and 3D generation. In addition to leveraging the generalizability and view-consistent prior in SVD, SV3D enables controllability via camera pose conditioning, and generates orbital videos of an object at high resolution on arbitrary camera orbits. We further propose several techniques for improved 3D generation from SV3D, including triangle CFG scaling, disentangled illumination, and masked SDS loss. We conduct extensive experiments to show that SV3D is controllable, multi-view consistent, and generalizable to the real-world, achieving state-of-the-art performance on novel multi-view synthesis and 3D generation. We believe SV3D provides a solid foundation model for further research on 3D object generation. We plan to publicly release SV3D models. + +# Acknowledgements + +We thank Emad Mostaque, Bryce Wilson, Anel Islamovic, Savannah Martin, Ana Guillen, Josephine Parquet, Adam Chen and Ella Irwin for their help in various aspects of the model development and release. We thank Kyle Lacey and Yan Marek for their help with demos. We also thank Eric Courtemanche for his help with visual results. + +# References + +1. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-NeRF: A multiscale representation for anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5855-5864 (2021) +2. Blattmann, A., Dockhorn, T., Kulal, S., Mendelevitch, D., Kilian, M., Lorenz, D., Levi, Y., English, Z., Voleti, V., Letts, A., et al.: Stable Video Diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127 (2023) +3. Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S.W., Fidler, S., Kreis, K.: Align your Latents: High-Resolution Video Synthesis with Latent Diffusion Models. arXiv:2304.08818 (2023) +4. Boss, M., Braun, R., Jampani, V., Barron, J.T., Liu, C., Lensch, H.P.: NeRD: Neural reflectance decomposition from image collections. In: IEEE International Conference on Computer Vision (ICCV) (2021) +5. Boss, M., Engelhardt, A., Kar, A., Li, Y., Sun, D., Barron, J.T., Lensch, H.P., Jampani, V.: SAMURAI: Shape And Material from Unconstrained Real-world Arbitrary Image collections. In: Advances in Neural Information Processing Systems (NeurIPS) (2022) +6. Chan, E.R., Nagano, K., Chan, M.A., Bergman, A.W., Park, J.J., Levy, A., Aittala, M., Mello, S.D., Karras, T., Wetzstein, G.: GeNVS: Generative novel view synthesis with 3D-aware diffusion models. International Conference on Computer Vision (2023) +7. Cline, D.B.: Admissible kernel estimators of a multivariate density. The Annals of Statistics 16(4), 1421-1427 (1988) +8. Deitke, M., Liu, R., Wallingford, M., Ngo, H., Michel, O., Kusupati, A., Fan, A., Laforte, C., Voleti, V., Gadre, S.Y., et al.: Objverse-XL: A universe of $10\mathrm{m} + 3\mathrm{D}$ objects. arXiv preprint arXiv:2307.05663 (2023) +9. Deitke, M., Schwenk, D., Salvador, J., Weihs, L., Michel, O., VanderBilt, E., Schmidt, L., Ehsani, K., Kembhavi, A., Farhadi, A.: Objaverse: A universe of annotated 3D objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13142-13153 (2023) +10. Downs, L., Francis, A., Koenig, N., Kinman, B., Hickman, R., Reymann, K., McHugh, T.B., Vanhoucke, V.: Google Scanned Objects: A high-quality dataset of 3D scanned household items. In: 2022 International Conference on Robotics and Automation (ICRA). pp. 2553-2560. IEEE (2022) +11. Eftekhar, A., Sax, A., Bachmann, R., Malik, J., Zamir, A.: Omnidata: A scalable pipeline for making multi-task mid-level vision datasets from 3D scans. In: IEEE International Conference on Computer Vision (ICCV) (2021) + +12. Girdhar, R., Singh, M., Brown, A., Duval, Q., Azadi, S., Rambhatla, S.S., Shah, A., Yin, X., Parikh, D., Misra, I.: EMU VIDEO: Factorizing Text-to-Video Generation by Explicit Image Conditioning (2023), https://emu-video.metademolab.com/ assets/emu(video.pdf +13. Hasselgren, J., Hofmann, N., Munkberg, J.: Shape, Light, and Material Decomposition from Images using Monte Carlo Rendering and Denoising. In: Advances in Neural Information Processing Systems (NeurIPS) (2022) +14. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: Advances in Neural Information Processing Systems (2020) +15. Hong, Y., Zhang, K., Gu, J., Bi, S., Zhou, Y., Liu, D., Liu, F., Sunkavalli, K., Bui, T., Tan, H.: LRM: Large reconstruction model for single image to 3D. International Conference on Learning Representations (2024) +16. Jun, H., Nichol, A.: Shap-E: Generating conditional 3D implicit functions (2023) +17. Kanopoulos, N., Vasanthavada, N., Baker, R.L.: Design of an image edge detection filter using the sobel operator. IEEE Journal of solid-state circuits 23(2), 358-367 (1988) +18. Kant, Y., Wu, Z., Vasilkovsky, M., Qian, G., Ren, J., Guler, R.A., Ghanem, B., Tulyakov, S., Gilitschenski, I., Siarohin, A.: Spad: Spatially aware multiview diffusers. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2024) +19. Karras, T., Aittala, M., Aila, T., Laine, S.: Elucidating the Design Space of Diffusion-Based Generative Models. arXiv:2206.00364 (2022) +20. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) +21. Kong, X., Liu, S., Lyu, X., Taher, M., Qi, X., Davison, A.J.: Eschernet: A generative model for scalable view synthesis. arXiv preprint arXiv:2402.03908 (2024) +22. Kwak, J.g., Dong, E., Jin, Y., Ko, H., Mahajan, S., Yi, K.M.: Vivid-1-to-3: Novel view synthesis with video diffusion models. IEEE conference on computer vision and pattern recognition (2024) +23. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3D: High-resolution text-to-3D content creation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 300–309 (2023) +24. Liu, M., Shi, R., Chen, L., Zhang, Z., Xu, C., Wei, X., Chen, H., Zeng, C., Gu, J., Su, H.: One-2-3-45++: Fast single image to 3D objects with consistent multi-view generation and 3D diffusion. arXiv preprint arXiv:2311.07885 (2023) +25. Liu, M., Xu, C., Jin, H., Chen, L., Varma T, M., Xu, Z., Su, H.: One-2-3-45: Any single image to 3D mesh in 45 seconds without per-shape optimization. Advances in Neural Information Processing Systems 36 (2023) +26. Liu, R., Wu, R., Hoorick, B.V., Tokmakov, P., Zakharov, S., Vondrick, C.: Zero-1-to-3: Zero-shot one image to 3D object. International Conference on Computer Vision (2023) +27. Liu, Y., Lin, C., Zeng, Z., Long, X., Liu, L., Komura, T., Wang, W.: SyncDreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453 (2023) +28. Long, X., Guo, Y.C., Lin, C., Liu, Y., Dou, Z., Liu, L., Ma, Y., Zhang, S.H., Habermann, M., Theobalt, C., et al.: Wonder3D: Single image to 3D using cross-domain diffusion. arXiv preprint arXiv:2310.15008 (2023) +29. Melas-Kyriazi, L., Laina, I., Rupprecht, C., Neverova, N., Vedaldi, A., Gafni, O., Kokkinos, F.: IM-3D: Iterative multiview diffusion and reconstruction for high-quality 3D generation. arXiv preprint arXiv:2402.08682 (2024) + +30. Mercier, A., Nakhli, R., Reddy, M., Yasarla, R., Cai, H., Porikli, F., Berger, G.: HexaGen3D: Stablediffusion is just one step away from fast and diverse Text-to-3D generation. arXiv preprint arXiv:2401.07727 (2024) +31. Metzer, G., Richardson, E., Patashnik, O., Giryes, R., Cohen-Or, D.: Latent-NeRF for shape-guided generation of 3D shapes and textures. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12663-12673 (2023) +32. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing scenes as neural radiance fields for view synthesis. In: European Conference on Computer Vision. pp. 405-421 (2020) +33. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph. (2022) +34. Nichol, A., Jun, H., Dhariwal, P., Mishkin, P., Chen, M.: Point-E: A System for Generating 3D Point Clouds from Complex Prompts (2022) +35. Niemeyer, M., Barron, J.T., Mildenhall, B., Sajjadi, M.S.M., Geiger, A., Radwan, N.: RegNeRF: Regularizing neural radiance fields for view synthesis from sparse inputs. In: Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR) (2022) +36. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023) +37. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3D using 2D diffusion. arXiv (2022) +38. Qian, G., Mai, J., Hamdi, A., Ren, J., Siarohin, A., Li, B., Lee, H.Y., Skorokhodov, I., Wonka, P., Tulyakov, S., et al.: Magic123: One image to high-quality 3D object generation using both 2D and 3D diffusion priors. arXiv preprint arXiv:2306.17843 (2023) +39. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) +40. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022) +41. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dream-Booth: Fine tuning text-to-image dissusion models for subject-driven generation. arXiv preprint arXiv:2208.12242 (2022) +42. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E., Kamyar Seyed Ghasemipour, S., Karagol Ayan, B., Mahdavi, S.S., Gontijo Lopes, R., Salimans, T., Ho, J., Fleet, D.J., Norouzi, M.: Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding. arXiv:2205.11487 (2022) +43. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortzman, M., et al.: LAION-5B: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022) +44. Shen, T., Gao, J., Yin, K., Liu, M.Y., Fidler, S.: Deep Marching Tetrahedra: a hybrid representation for high-resolution 3D shape synthesis. In: Advances in Neural Information Processing Systems (NeurIPS) (2021) +45. Shi, R., Chen, H., Zhang, Z., Liu, M., Xu, C., Wei, X., Chen, L., Zeng, C., Su, H.: Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110 (2023) + +46. Shi, Y., Wang, P., Ye, J., Long, M., Li, K., Yang, X.: MVDream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512 (2023) +47. Song, Y., Ermon, S.: Improved Techniques for Training Score-Based Generative Models. arXiv:2006.09011 (2020) +48. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456 (2020) +49. StabilityAI: Stable Zero123 (2023), https://stability.ai/news/stable-zero123-3d-generation +50. Tang, J., Ren, J., Zhou, H., Liu, Z., Zeng, G.: DreamGaussian: Generative Gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653 (2023) +51. Tang, S., Chen, J., Wang, D., Tang, C., Zhang, F., Fan, Y., Chandra, V., Furukawa, Y., Ranjan, R.: Mvdiffusion++: A dense high-resolution multi-view diffusion model for single or sparse-view 3d object reconstruction. arXiv preprint arXiv:2402.12712 (2024) +52. Tsai, Y.T., Shih, Z.C.: All-frequency precomputed radiance transfer using spherical radial basis functions and clustered tensor approximation. ACM Transactions on Graphics (ToG) (2006) +53. Voleti, V., Jolicoeur-Martineau, A., Pal, C.: MCVD: Masked conditional video diffusion for prediction, generation, and interpolation. In: (NeurIPS) Advances in Neural Information Processing Systems (2022) +54. Watson, D., Chan, W., Martin-Brualla, R., Ho, J., Tagliasacchi, A., Norouzi, M.: Novel view synthesis with diffusion models (2022) +55. Weng, H., Yang, T., Wang, J., Li, Y., Zhang, T., Chen, C.L.P., Zhang, L.: Consistent123: Improve consistency for one image to 3D object synthesis (2023) +56. Wu, T., Zhang, J., Fu, X., Wang, Y., Ren, J., Pan, L., Wu, W., Yang, L., Wang, J., Qian, C., et al.: Omniobject3D: Large-vocabulary 3D object dataset for realistic perception, reconstruction and generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 803-814 (2023) +57. Yang, J., Cheng, Z., Duan, Y., Ji, P., Li, H.: ConsistNet: Enforcing 3D consistency for multi-view images diffusion. arXiv preprint arXiv:2310.10343 (2023) +58. Yao, C.H., Raj, A., Hung, W.C., Rubinstein, M., Li, Y., Yang, M.H., Jampani, V.: ARTIC3D: Learning robust articulated 3D shapes from noisy web image collections. Advances in Neural Information Processing Systems 36 (2024) +59. Ye, J., Wang, P., Li, K., Shi, Y., Wang, H.: Consistent-1-to-3: Consistent image to 3D view synthesis via geometry-aware diffusion models. In: 3DV (2024) +60. Young, J.: xatlas. https://github.com/jpcy/xatlas (2024) +61. Yu, Z., Peng, S., Niemeyer, M., Sattler, T., Geiger, A.: MonoSDF: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in Neural Information Processing Systems (NeurIPS) (2022) +62. Zhang, K., Luan, F., Wang, Q., Bala, K., Snavely, N.: PhySG: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In: The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021) +63. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018) +64. Zheng, C., Vedaldi, A.: Free3D: Consistent novel view synthesis without 3D representation. arXiv preprint arXiv:2312.04551 (2023) + +In this supplemental document, we include the broader impact in Appendix A, implementation details in Appendices B to F, ablative analyses in Appendix G, and additional results in Appendix H. We also attach a teaser video summarizing the SV3D framework (SV3D(video.mp4), as well as an HTML page (SV3D.html) for more visual comparisons. + +# A Broader Impact + +The advancement of generative models in different media forms is changing how we make and use content. These AI-powered models can create images, videos, 3D objects, and more, in ways we've never seen before. They offer huge potential for innovation in media production. But along with this potential, there are also risks. Before we start using these models widely, it is crucial to make sure we understand the possible downsides and have plans in place to deal with them effectively. + +In the case of 3D object generation, the input provided by the user plays a crucial role in constraining the model's output. By supplying a full front view of an object, users limit the model's creative freedom to the visible or unoccluded regions, thus minimizing the potential for generating problematic imagery. Additionally, factors such as predicted depth values and lighting further influence the fidelity and realism of generated content. + +Moreover, ensuring the integrity and appropriateness of training data is critical in mitigating risks associated with generative models. Platforms like Sketchfab, which serve as repositories for 3D models used in training, enforce strict content policies to prevent the dissemination of "Unacceptable Content" and disallows it on their platform: https://help.sketchfab.com/hc/en-us/articles/214867883-What-is-Restricted-Content. By adhering to these guidelines and actively monitoring dataset quality, developers can reduce the likelihood of biased or inappropriate outputs. + +Sketchfab also has a tag for "Restricted Content" which is deemed to be similar to the PG-13 content rating used in the US (i.e. inappropriate for children under 13). We have confirmed that none of the objects that we use in training have this flag set to true. Thus we go the extra step of excluding even tagged PG-13 content from the training set. + +There is a chance that certain content may not have been correctly labeled on Sketchfab. In cases where the uploader fails to tag an object appropriately, Sketchfab provides publicly accessible listings of objects and a mechanism for the community to report any content that may be deemed offensive. + +In our regular utilization of Objaverse content, we haven't observed any significant amount of questionable material. Nevertheless, there are occasional instances of doll-like nudity stemming from basic 3D models, which could be crucial for accurately depicting humanoid anatomy. Additionally, the training dataset contains some presence of drugs, drug paraphernalia, as well as a certain level of blood content and weaponry, resembling what might be encountered in a video game context. Should the model be provided with imagery featuring these cate- + +gories of content, it possesses the capability to generate corresponding 3D models to some extent. + +It is to be noted that SV3D mainly focuses on generating hidden details in the user's input image. If the image is unclear or some parts are hidden, the model guesses what those parts might look like based on its training data. This means it might create details similar to what it has seen before. The training data generally follows the standards of 3D modeling and gaming. However, this could lead to criticisms about the models being too similar to existing trends. But the user's input image limits how creative the model can be and reduces the chance of biases showing up in its creations, especially if the image is clear and straightforward. + +# B Data Details + +Similar to SVD-MV [2], we render views of a subset of the Objaverse dataset [9] consisting of 150K curated CC-licensed 3D objects from the original dataset. Each loaded object is scaled such that the largest world-space XYZ extent of its bounding box is 1. The object is then repositioned such that this bounding box is centered around the origin. + +For both the static and dynamic orbits, we use Blender's EEVEE renderer to render an 84-frame RGBA orbit at $576 \times 576$ resolution. During training, any 21-frame orbit can be subsampled from this by picking any frame as the first frame, and then choosing every 4th frame after that. + +We apply two background colors to each of these images: random RGB color, and white. This results in a doubling of the number of orbit samples for training. We then encode all of these images into latent space using SVD's VAE, and using CLIP. We then store the latent and CLIP embeddings for all of these images along with the corresponding elevation and azimuth values. + +For lighting, we randomly select from a set of 20 curated HDRI envmaps. Each orbit begins with the camera positioned at azimuth 0. Our camera uses a field-of-view of 33.8 degrees. For each object, we adaptively position the camera to a distance sufficient to ensure that the rendered object content makes good and consistent use of the image extents without being clipped in any view. + +For static orbits, the camera is positioned at a randomly sampled elevation between $[-5, 30]$ degrees. The azimuth steps by a constant $\frac{360}{84}$ degree delta between each frame. For dynamic orbits, the sequence of camera elevations for each orbit are obtained from a random weighted combination of sinusoids with different frequencies. For each sinusoid, the whole number period is sampled from [1, 5], the amplitude is sampled from [0.5, 10], and a random phase shift is applied. The azimuth angles are sampled regularly, and then a small amount of noise is added to make them irregular. The elevation values are smoothed using a simple convolution kernel and then clamped to a maximum elevation of 89 degrees. + +# C Training Details + +Our approach involves utilizing the widely used EDM [19] framework, incorporating a simplified diffusion loss for fine-tuning, as followed in SVD [2]. We eliminate the conditions of 'fps_id', 'motion_bucket_id', etc. since they are irrelevant for SV3D. Furthermore, we adjust the loss function to assign lower weights to frames closer to the front-view conditioning image, ensuring that challenging back views receive equal training focus as the easier front views. To optimize training efficiency and conserve GPU VRAM, we preprocess and store the precomputed latent and CLIP embeddings for all video frames in advance. During training, these tensors are directly loaded rather than being computed in real-time. We choose to finetune the SVD-xt model to output 21 frames instead of 25 frames. We found that with 21 frames we were able to fit a batch size of 2 on each GPU, instead of 1 with 25 frames at $576 \times 576$ resolution. All SV3D models are trained for 105k iterations with an effective batch size of 64 on 4 nodes of 880GB A100 GPUs for around 6 days. + +# D Inference Details + +To generate an orbital video during inference, we use 50 steps of the deterministic DDIM sampler [47] with the triangular classifier-free guidance (CFG) scale described in the main paper. This takes $\approx 40$ seconds for the SV3D model. + +# E Additional Details on Illumination Model + +We base our rendering model on Spherical Gaussians (SG) [4,62]. A SG at query location $\pmb{x} \in \mathbb{R}^3$ is defined as $G(\pmb{x};\pmb{\mu},c,a) = ae^{s(\pmb{\mu}\cdot \pmb{x} - 1)}$ , where $\pmb{\mu} \in \mathbb{R}^3$ is the axis, $s \in \mathbb{R}$ the sharpness of the lobe, and $a \in \mathbb{R}$ the amplitude. Here, we point out that we only model white light and hence only use a scalar amplitude. One particularly interesting property of SGs is that the inner product of two SGs is the integral of the product of two SGs. The operation is defined as [52]: + +$$ +\begin{array}{l} G _ {1} (\boldsymbol {x}) \cdot G _ {2} (\boldsymbol {x}) = \int_ {\Omega} G _ {1} (\boldsymbol {x}) G _ {2} (\boldsymbol {x}) d \boldsymbol {x} = \frac {1}{d _ {m}} \left(2 \pi a _ {1} a _ {2} e ^ {d _ {m} - \lambda_ {m}} (1. 0 - e ^ {- 2 d _ {m}})\right) \\ \lambda_ {m} = \lambda_ {1} - \lambda_ {2} \tag {1} \\ d _ {m} = \left| \left| \lambda_ {1} \boldsymbol {\mu} _ {1} + \lambda_ {2} \boldsymbol {\mu} _ {2} \right| \right|. \\ \end{array} +$$ + +In our illumination model we only consider Lambertian shading. Here, the cosine shading term influences the output the most. This term can be approximated with another SG $G_{\mathrm{cosine}} = (\pmb{x}; \pmb{n}, 2.133, 1.17)$ , where $\pmb{n}$ defines the surface normal at $\pmb{x}$ . The lighting evaluation using SGs $G_i$ then becomes: $L = \sum_{i=1}^{24} \frac{1}{\pi} \max(G_i \cdot G_{\mathrm{cosine}}, 0)$ . As defined previously this results in the full integration of incoming light for each SG and as light is additive evaluating and summing all SGs results in the complete environment illumination. This $L$ is also used in the $\mathcal{L}_{\mathrm{illum}}$ loss described in the main paper. The rendered textured + +image is then defined as $\hat{I} = c_dL$ , where $c_{d}$ is the diffuse albedo. We learn $\mu, c, a$ for each SG $G_{i}$ using reconstruction loss between these rendered images and SV3D-generated images. + +# F Losses and Optimization for 3D Generation + +In addition to the masked SDS loss $\mathcal{L}_{\mathrm{mask - ds}}$ and illumination loss $\mathcal{L}_{\mathrm{illum}}$ detailed in the manuscript, we use several other losses for 3D reconstruction. Our main reconstruction losses are the pixel-level mean squared error $\mathcal{L}_{\mathrm{mse}} = \| I - \hat{I}\|^2$ , LPIPS [63] loss $\mathcal{L}_{\mathrm{lpips}}$ , and mask loss $\mathcal{L}_{\mathrm{mask}} = \| S - \hat{S}\|^2$ , where $S$ , $\hat{S}$ are the predicted and ground-truth masks. We further employ a normal loss using the estimated mono normal by Omnidata [11], which is defined as the cosine similarity between the rendered normal $\pmb{n}$ and estimated pseudo ground truths $\bar{n}$ : $\mathcal{L}_{\mathrm{normal}} = 1 - (\pmb{n} \cdot \bar{\pmb{n}})$ . To regularize the output geometry, we apply a smooth depth loss inspired by RegNeRF [35]: $\mathcal{L}_{\mathrm{depth}}(i,j) = (d(i,j) - d(i,j + 1))^2 + (d(i,j) - (d(i + 1,j))^2$ , where $i,j$ indicate the pixel coordinate. For surface normal we instead rely on a bilateral smoothness loss similar to [5]. We found that this is crucial to getting high-frequency details and avoiding over-smoothed surfaces. For this loss we compute the image gradients of the input image $\nabla I$ with a Sobel filter [17]. We then encourage the gradients of rendered normal $\nabla n$ to be smooth if (and only if) the input image gradients $\nabla I$ are smooth. The loss can be written as $\mathcal{L}_{\mathrm{bilateral}} = e^{-3\nabla I}\sqrt{1 + ||\nabla n||}$ . We also found that adding a spatial smoothness regularization on the albedo is beneficial: $\mathcal{L}_{\mathrm{albedo}} = (c_d(x) - c_d(x + \epsilon))^2$ , where $c_d$ denotes the albedo, $x$ is a 3D surface point, and $\epsilon \in \mathbb{R}^3$ is a normal distributed offset with a scale of 0.01. The overall objective is then defined as the weighted sum of these losses. All losses are applied in both coarse and fine stages, except that we only apply $\mathcal{L}_{\mathrm{mask - ds}}$ in the last 200 iterations of the fine stage. We use an Adam optimizer [20] with a learning rate of 0.01 for both stages. + +# G Additional Ablative Analyses + +We conduct additional ablative analyses of our 3D generation pipeline in this section. + +# G.1 SV3D Models + +In Tab. 7, we compare the quantitative results using different SV3D models and training losses. Both 2D and 3D evaluation shows that $\mathrm{SV3D}^p$ is our best performing model, either for pure photometric reconstruction or SDS-based optimization. + +Table 7: Ablative results of different SV3D models and training losses. We show that our $\mathrm{SV3D}^p$ model with Photo+SDS losses achieves the best 2D and 3D metrics. + +
ModelTraining lossesLPIPS↓PSNR↑SSIM↑MSE↓CLIP-S↑CD↓3D IoU↑
SV3DuPhoto0.13215.9510.8270.0320.8730.0280.583
SV3DuPhoto+SDS0.13315.9570.8340.0310.8710.0270.589
SV3DcPhoto0.13515.8260.8320.0330.8710.0290.579
SV3DcPhoto+SDS0.13216.3730.8340.0270.8700.0270.584
SV3DpPhoto0.12416.8640.8410.0240.8750.0240.611
SV3DpPhoto+SDS0.11917.4050.8490.0210.8770.0240.614
+ +# G.2 Static v.s. Dynamic Orbits + +We also compare the results using different camera orbits for 3D training in Tab. 8. The results show that using a dynamic orbit (sine-30) produces better 3D outputs compared to static orbit since it contains more information of the top and bottom views of the object. However, higher elevation (sine-50) tends to increase inconsistency between multi-view images, and thus resulting in worse 3D reconstruction. In our experiments, we find that setting the elevation within $\pm 30$ degree generally leads to desirable 3D outputs. + +Table 8: Ablative results of different reference orbits for 3D generation. We show that using a dynamic orbit (sine elevation) with moderate amplitude performs better than orbits with no or extreme elevation variations. + +
Training orbitLPIPS↓PSNR↑SSIM↑MSE↓CLIP-S↑CD↓3D IoU↑
Static0.12516.8210.8480.0250.8640.0280.610
Sine-300.11917.4050.8490.0210.8770.0240.614
Sine-500.12317.0570.8540.0250.8730.0260.609
+ +# G.3 Masked SDS Loss + +Finally, we show the ablative results of our SDS loss in Tab. 9. We compare the results of 1) pure photometric losses, 2) with naive SDS loss (no masking), 3) with hard-masked SDS loss by thresholding the dot product of surface normal and camera viewing angle as visibility masks, and 4) with soft-masked SDS loss as described in the manuscript. Overall, adding SDS guidance from the SV3D model can improve the 2D metrics while maintaining similar 3D metrics. Our novel soft-masked SDS loss generally achieves the best results compared to other baselines. + +Table 9: Ablative analyses of Masked SDS loss. Overall, our soft-masked SDS loss leads to higher-quality mesh outputs in terms of most 2D and 3D metrics. + +
Training lossesLPIPS↓PSNR↑SSIM↑MSE↓CLIP-S↑CD↓3D IoU↑
Photo0.12416.8640.8410.0240.8750.0240.611
Photo+SDS (naive)0.12417.0070.8500.0240.8670.0250.615
Photo+SDS (hard masked)0.12417.3350.8450.0220.8770.0240.610
Photo+SDS (soft masked)0.11917.4050.8490.0210.8770.0240.614
+ +# H Additional Visual Results + +In this section, we show more results of novel view synthesis and 3D generation. + +# H.1 Novel View Synthesis + +We show the additional NVS results on OmniObject3D [56] and real-world images in Fig. 13 and Fig. 14, respectively. The generated novel multi-view images by SV3D are more detailed and consistent compared to prior state-of-the-arts. + +![](images/105b9af9adeba66108cfd77e8c2dad8e2bfee731f010f0f8fba23fe2b687e7ca.jpg) +Fig. 13: NVS Results on OmniObject3D [56]. For each object, we show the input image and generated novel multi-view images by different methods. SV3D is able to produce images with consistent pose, color, and geometric details, which are closer the ground-truth renders. + +![](images/f8dfcc05bba1ccf34bc9c1286b08bcab5f3ff6351a6d67286b56a7fd0387b3d3.jpg) +Fig. 14: NVS Results on Real-World Images. Notice the consistent texture, geometry, and pose in SV3D NVS outputs compared to prior works. + +# H.2 3D Generation + +We show the additional 3D generation results on OmniObject3D [56] and real-world images in Fig. 15 and Fig. 16, respectively. Thanks to the consistent multi-view images by SV3D and the novel Masked SDS loss, our 3D generations are detailed, high-fidelity, and generalizable to a wide range of images. Since Free3D [64] does not include a 3D generation method, we run our 3D pipeline on its generated multi-view images for fair 3D comparison. + +![](images/2636ffd15e777637be3807b2fddf85d618586e9e02154032657a0d1a01a83753.jpg) +Fig. 15: Mesh Results on OmniObject3D [56]. Thanks to the consistent novel multi-view images generated by SV3D, our 3D meshes contain higher-fidelity details while having smooth surfaces. + +![](images/bf21c4f7501d5718f0681146204e4d1c72137712b15b642c356d72f3ef9496ed.jpg) + +![](images/c1771e6d83010626d04298302ccea97de0b0d513ece8964c10b8356feea725d9.jpg) + +![](images/03831f022ccc6d4c6f863a15122dae0102a2a28bbf2be6daff0cb9f38f20d187.jpg) + +![](images/af2dfd1a717bcf4ddf323c2bbf1bb4480d6e5871a129128e3ee2fcf03cf92edc.jpg) + +![](images/eeee73d320527f8110c04100e047f38eb14974c6758bee457f90ca4760636cb0.jpg) + +![](images/71e079185eeecd8f2d77f798c1b502cd2f3ef1dd54d119e1221a9a6828c0ac81.jpg) + +![](images/7fde30b1db64fec167c6874d912d2cd110f9ab3db4a2c032fb20f0ef8b8c5c31.jpg) + +![](images/0ea317b84ad7a62bb4ea75c19291e24f346adc9422bd4ec03ff6e39e1e28e431.jpg) + +![](images/8d5feb66007408ebe637e15fd7aa917354d6bb4d45f75e60debeb5b28fd80c3c.jpg) + +![](images/fe6da9439b80e8940978a5b41f4348f88fe7962f82209cedc9c52e241736ac2f.jpg) + +![](images/fd5959849467000e71de01bfeed96deeb9a41a7ec08aa22145be45c402ee75a4.jpg) + +![](images/e41cbdd433d6b5194e91c97c77ed3f7e4e292bf429f823e22f507366266aa936.jpg) + +![](images/410f32136b831e515a6dc9df47293aee849b7916cd829908725ef633fc4b477c.jpg) + +![](images/e13154017e66a59f398233c875015c1797cfdd3ac9edaf8ef53c12d453886c09.jpg) + +![](images/48d42cdabb13393fda6c62f2c7e40de307427c762285cb6bc96b5a35a5d3f4f4.jpg) + +![](images/3124b38b7ca24277fdb6e853e2e9f9311ec830f2a96da6e9368e860ab62beb47.jpg) + +![](images/588f6a934c862812ac95926fb295431cd952ff70624031689af9fccc30899b22.jpg) + +![](images/e67dc5694ded9302613190785b2c063fd1018a87fd668b03e1ab3cf43788c63d.jpg) + +![](images/3a0124166a356a96f01ea15ab24f3e9dc1744cc495ca07851aa034db2edcf47e.jpg) + +![](images/4e064a625dbad13dfb4f12e3a79fadd81b214212bb524c75cb8d591c38f0ad33.jpg) + +![](images/700de025b4638fdbff94603b36c34dba71bbeb360015ee50308e6cee7c14b912.jpg) +SV3D + +![](images/529907a07ac547f4fa93a0765e19e3e36b6c915840f7ddd2df888b49af04ccfc.jpg) + +![](images/b798dde190ce6ac208841607c19bf331c0e7e1a9d23f30ad12ab2de367eebd82.jpg) +EscherNet + +![](images/7763dd4e82f207b058b1f1912e8bcba795d6ed1018d8ce4ef552ee0709cde2df.jpg) + +![](images/1874c021568fd7b505d7b6603e370e17ad3353acf7eba1ced4fa9be69dd3d328.jpg) +Free3D +Fig. 16: Mesh Results on Real-World Images. Notice the accurate shape and details in our reconstructions even from the diverse images in-the-wild. + +![](images/c73ad4f0f244fbd8d0a7a18df3aa5ebb36d9da8859a1dba445eb4219135f88f9.jpg) + +![](images/fe0e95ac726c2866be59983efc71a4d1a978a945b82899035a4dbccd048fdfdd.jpg) +Stable Zero123 + +![](images/4f7cd0772d90275f1e9ee7af525094b9c69eeb707919381d4120c5d096e4f68a.jpg) +Fig. 17: NVS and 3D Generation Results on Real-World Images. Visual results on real-world images from the MVImgNet (Yu et al. CVPR'23) and NAVI (Jampani et al. NeurIPS'22) datasets. Compared to Stable Zero123, SV3D generates more consistent novel multi-view images, thus resulting in higher-quality 3D meshes with smooth surface and detailed texture. + +![](images/f2e9a4372043cdf1fb202506c1a54251c142495e9254853a6c6c4900c24288a9.jpg) \ No newline at end of file diff --git a/2024/SV3D_ Novel Multi-view Synthesis and 3D Generation from a Single Image using Latent Video Diffusion/images.zip b/2024/SV3D_ Novel Multi-view Synthesis and 3D Generation from a Single Image using Latent Video Diffusion/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..354ff7ef7ac7465f276a918fb0acadc12d45ef38 --- /dev/null +++ b/2024/SV3D_ Novel Multi-view Synthesis and 3D Generation from a Single Image using Latent Video Diffusion/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a356507ff6e7f5d5fcd5457bc6d4cbb0526aa499e9a4d0d0ca34b9eca0c8798a +size 1017959 diff --git a/2024/SV3D_ Novel Multi-view Synthesis and 3D Generation from a Single Image using Latent Video Diffusion/layout.json b/2024/SV3D_ Novel Multi-view Synthesis and 3D Generation from a Single Image using Latent Video Diffusion/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a20734e7f85fc30ca93847d9a08c7c4174983262 --- /dev/null +++ b/2024/SV3D_ Novel Multi-view Synthesis and 3D Generation from a Single Image using Latent Video Diffusion/layout.json @@ -0,0 +1,15565 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 149, + 111, + 465, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 111, + 465, + 163 + ], + "spans": [ + { + "bbox": [ + 149, + 111, + 465, + 163 + ], + "type": "text", + "content": "SV3D: Novel Multi-view Synthesis and 3D Generation from a Single Image using Latent Video Diffusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 138, + 186, + 475, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 186, + 475, + 211 + ], + "spans": [ + { + "bbox": [ + 138, + 186, + 475, + 211 + ], + "type": "text", + "content": "Vikram Voleti*, Chun-Han Yao*, Mark Boss*, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani" + } + ] + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 138, + 219, + 471, + 327 + ], + "blocks": [ + { + "bbox": [ + 138, + 219, + 471, + 327 + ], + "lines": [ + { + "bbox": [ + 138, + 219, + 471, + 327 + ], + "spans": [ + { + "bbox": [ + 138, + 219, + 471, + 327 + ], + "type": "image", + "image_path": "83325efc4c0195ff4b6a0c61c5348cc40f69864fffe4d7858886e4b0a8f2a849.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 338, + 482, + 371 + ], + "lines": [ + { + "bbox": [ + 132, + 338, + 482, + 371 + ], + "spans": [ + { + "bbox": [ + 132, + 338, + 482, + 371 + ], + "type": "text", + "content": "Fig. 1: Stable Video 3D (SV3D). From a single image, SV3D generates consistent novel multi-view images. We then optimize a 3D representation with SV3D-generated views resulting in high-quality 3D meshes." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 159, + 384, + 453, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 384, + 453, + 538 + ], + "spans": [ + { + "bbox": [ + 159, + 384, + 453, + 538 + ], + "type": "text", + "content": "Abstract. We present Stable Video 3D (SV3D) — a latent video diffusion model for high-resolution, image-to-multi-view generation of orbital videos around a 3D object. Recent works propose to adapt 2D generative models for novel view synthesis (NVS) and 3D optimization. However, these methods have several disadvantages due to limited views or inconsistent NVS, affecting the performance of 3D object generation. In this work, we propose SV3D that adapts image-to-video diffusion model for novel multi-view synthesis and 3D generation, thereby leveraging the generalization and multi-view consistency of the video models, while further adding explicit camera control for NVS. We also propose improved 3D optimization techniques for image-to-3D generation using SV3D and its NVS outputs. Extensive experiments on multiple datasets with 2D and 3D metrics and user study demonstrate SV3D's state-of-the-art performance on NVS as well as 3D reconstruction compared to prior works." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 548, + 436, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 548, + 436, + 560 + ], + "spans": [ + { + "bbox": [ + 160, + 548, + 436, + 560 + ], + "type": "text", + "content": "Keywords: 3D synthesis " + }, + { + "bbox": [ + 160, + 548, + 436, + 560 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 548, + 436, + 560 + ], + "type": "text", + "content": " Video generation " + }, + { + "bbox": [ + 160, + 548, + 436, + 560 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 548, + 436, + 560 + ], + "type": "text", + "content": " Novel view synthesis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 581, + 230, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 581, + 230, + 594 + ], + "spans": [ + { + "bbox": [ + 132, + 581, + 230, + 594 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Single-image 3D object reconstruction is a long-standing problem in computer vision with a wide range of applications in game design, AR/VR, e-commerce, robotics, etc. It is a highly challenging and ill-posed problem as it requires lifting 2D pixels to 3D space while also reasoning about the unseen portions of the object in 3D." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": "Despite being a long-standing vision problem, it is only recently that practically useful results are being produced by leveraging advances in the generative AI. This is mainly made possible by the large-scale pretraining of generative models which enables sufficient generalization to various domains. A typical strategy is to use image-based 2D generative models (e.g., Imagen [42], Stable Diffusion (SD) [40]) to provide a 3D optimization loss function for unseen novel views of a given object [23,31,37]. In addition, several works repurpose these 2D generative models to perform novel view synthesis (NVS) from a single image [26,27,30,54], and then use the synthesized novel views for 3D generation. Conceptually, these works mimic a typical photogrammetry-based 3D object capture pipeline, i.e., first photographing multi-view images of an object, followed by 3D optimization; except that the explicit multi-view capture is replaced by novel-view synthesis using a generative model, either via text prompt or camera pose control." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 274, + 483, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 274, + 483, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 483, + 357 + ], + "type": "text", + "content": "A key issue in these generation-based reconstruction methods is the lack of multi-view consistency in the underlying generative model, resulting in inconsistent novel views. Some works try to address this by jointly reasoning a 3D representation during NVS [6,15,27], but at the cost of high compute and data, often still resulting in inconsistent geometric and texture details. In this work, we tackle this issue by adapting a high-resolution, image-conditioned video diffusion model for NVS followed by 3D generation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 361, + 483, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 483, + 504 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 483, + 504 + ], + "type": "text", + "content": "Novel Multi-view Synthesis. We adapt a latent video diffusion model (Stable Video Diffusion - SVD [2]) to generate multiple novel views of a given object with explicit camera pose conditioning. SVD demonstrates excellent multi-view consistency for video generation, and we repurpose it for NVs. In addition, SVD also has good generalization capabilities as it is trained on large-scale image and video data that are more readily available than large-scale 3D data. In short, we adapt the video diffusion model for NVs from a single image with three useful properties for 3D object generation: pose-controllable, multi-view consistent, and generalizable. We call our resulting NVS network 'SV3D'. To our knowledge, this is the first work that adapts a video diffusion model for explicit pose-controlled view synthesis. Some contemporary works such as [2, 29] demonstrate this, but only generate orbital videos without any explicit camera control." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 507, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 507, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 507, + 482, + 604 + ], + "type": "text", + "content": "3D Generation. We then utilize our SV3D model for 3D object generation by optimizing a NeRF and DMTet mesh in a coarse-to-fine manner. Benefiting from the multi-view consistency in SV3D, we are able to produce high-quality 3D meshes directly from the SV3D novel view images. We also design a masked score distillation sampling (SDS) [37] loss to further enhance 3D quality in the regions that are not visible in the SV3D-predicted novel views. In addition, we propose to jointly optimize a disentangled illumination model along with 3D shape and texture, effectively reducing the issue of baked-in lighting." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "We perform extensive comparisons of both our NVS and 3D generation results with respective state-of-the-art methods, demonstrating considerably better outputs with SV3D. For NVS, SV3D shows high-level of multi-view consistency and generalization to real-world images while being pose controllable. Our resulting 3D meshes are able to capture intricate geometric and texture details." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "text", + "content": "V. Voleti et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 225, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 225, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 225, + 129 + ], + "type": "text", + "content": "2 Background" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 138, + 270, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 138, + 270, + 152 + ], + "spans": [ + { + "bbox": [ + 132, + 138, + 270, + 152 + ], + "type": "text", + "content": "2.1 Novel View Synthesis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 157, + 479, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 157, + 479, + 181 + ], + "spans": [ + { + "bbox": [ + 130, + 157, + 479, + 181 + ], + "type": "text", + "content": "We organize the related works along three crucial aspects of novel view synthesis (NVS): generalization, controllability, and multi-view (3D) consistency." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 182, + 480, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 480, + 266 + ], + "type": "text", + "content": "Generalization. Diffusion models [14, 48] have recently shown to be powerful generative models that can generate a wide variety of images [3, 40, 41] and videos [2, 12, 53] by iteratively denoising a noise sample. Among these models, the publicly available Stable Diffusion (SD) [40] and Stable Video Diffusion (SVD) [2] demonstrate strong generalization ability by being trained on extremely large datasets like LAION [43] and LVD [2]. Hence, they are commonly used as foundation models for various generation tasks, e.g. novel view synthesis." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 268, + 481, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 268, + 481, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 268, + 481, + 411 + ], + "type": "text", + "content": "Controllability. Ideally, a controllable NVS model allows us to generate an image corresponding to any arbitrary viewpoint. For this, Zero123 [26] repurposes an image diffusion model to a novel view synthesizer, conditioned on a single-view image and the pose difference between the input and target views. Follow-up works such as Zero123XL [8] and Stable Zero123 [49] advance the quality of diffusion-based NVS, as well as the trained NeRFs using SDS loss. However, they only generate one novel view at a time, and thus are not designed to be multi-view consistent inherently. Recent works such as EscherNet [21] and Free3D [64] are capable of better multi-view consistency with intelligent camera position embedding design. However, they only use image-based diffusion models, and generate images at " + }, + { + "bbox": [ + 130, + 268, + 481, + 411 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 130, + 268, + 481, + 411 + ], + "type": "text", + "content": " resolution. We finetune a video diffusion model to generate novel views at " + }, + { + "bbox": [ + 130, + 268, + 481, + 411 + ], + "type": "inline_equation", + "content": "576 \\times 576" + }, + { + "bbox": [ + 130, + 268, + 481, + 411 + ], + "type": "text", + "content": " resolution." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 413, + 481, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 481, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 481, + 604 + ], + "type": "text", + "content": "Multi-view Consistency. Multi-view consistency is the most critical requirement for high-quality NVS and 3D generation. To address the inconsistency issue, MVDream [46], SyncDreamer [27], HexGen3D [30], and Zero123++ [45] propose to generate multiple (specific) views of an object simultaneously. However, they are not controllable: given a conditional image, they only generate specific views, not arbitrary viewpoints. Moreover, they were finetuned from image-based diffusion models, i.e. multi-view consistency was imposed by adding interaction among the multiple generated views through cross-attention layers. Hence, their output quality is limited to the generalizability of their base image-based model, and their 3D finetuning dataset. Efficient-3DiM finetunes the SD model with a stronger vision transformer DINO v2 [36]. Consistent-1-to-3 [59] and SPAD [18] leverage epipolar geometry. One-2-3-45 [25] and One-2-3-45++ [24] train additional 3D network using the 2D generator's outputs. MVDream [46], Consistent123 [55], and Wonder3D [28] also train multi-view diffusion models, yet still require post-processing for video rendering. SyncDreamer [27] and ConsistentNet [57] employ 3D representation into the latent diffusion model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": "Exploiting Video Diffusion Models. To achieve better generalization and multi-view consistency, some contemporary works exploit the temporal priors in video diffusion models for NVs. For instance, Vivid-1-to-3 [22] combines a view-conditioned diffusion model and video diffusion model to generate consistent views. SVD-MV [2] and IM-3D [29] finetune a video diffusion model for NVs." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "type": "text", + "content": "SV3D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "content": "However, they generate " + }, + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "inline_equation", + "content": "\\leq 360^{\\circ}" + }, + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "content": " views at the same elevation only. Unlike SV3D, none of them are capable of rendering any arbitrary view of the 3D object." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 481, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 481, + 200 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 481, + 200 + ], + "type": "text", + "content": "We argue that the existing NVS and 3D generation methods do not fully leverage the superior generalization capability, controllability, and consistency in video diffusion models. In this paper, we fill this important gap and train SV3D, a state-of-the-art novel multi-view synthesis video diffusion model at " + }, + { + "bbox": [ + 132, + 140, + 481, + 200 + ], + "type": "inline_equation", + "content": "576 \\times 576" + }, + { + "bbox": [ + 132, + 140, + 481, + 200 + ], + "type": "text", + "content": " resolution, and leverage it for 3D generation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 217, + 236, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 217, + 236, + 229 + ], + "spans": [ + { + "bbox": [ + 132, + 217, + 236, + 229 + ], + "type": "text", + "content": "2.2 3D Generation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 237, + 481, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 481, + 272 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 481, + 272 + ], + "type": "text", + "content": "Recent advances in 3D representations and diffusion-based generative models have significantly improved the quality of image-to-3D generation. Here, we briefly summarize the related works in these two categories." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 274, + 481, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 481, + 405 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 481, + 405 + ], + "type": "text", + "content": "3D Representation. 3D generation has seen great progress since the advent of Neural Radiance Fields (NeRFs) [32] and its subsequent variants [1], which implicitly represents a 3D scene as a volumetric function, typically parameterized by a neural network. Notably, Instant-NGP [33] introduces a hash grid feature encoding that can be used as a NeRF backbone for fast inference and ability to recover complex geometry. On the other hand, several recent works improve from an explicit representation such as DMTet [44], which is capable of generating high-resolution 3D shapes due to its hybrid SDF-Mesh representation and high memory efficiency. Similar to [23, 38], we adopt coarse-to-fine training for 3D generation, by first learning a rough object with Instant-NGP NeRF and then refining it using the DMTet representation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 407, + 481, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 481, + 551 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 481, + 551 + ], + "type": "text", + "content": "Diffusion-Based 3D Generation. Several recent works [16, 34] train a 3D diffusion model to to learn these flexible 3D representations, which, however, lack generalizability due to the scarcity of 3D data. To learn 3D generation without ground truth 3D data, image/multi-view diffusion models have been used as guidance. DreamFusion [37] and follow-up works [23, 31] leverage a trained image diffusion model as a 'scoring' function and calculate SDS loss for text-to-3D generation. However, they are prone to artifacts like Janus problem [31, 37] and over-saturated texture. Inspired by Zero123 [26], recent works [21, 24, 25, 27, 29, 38, 45, 49, 51, 64] finetune image/video diffusion models to generate novel view images as a stronger guidance. Our method shares the same spirit as this line of work, but produces denser, controllable, and more consistent multi-view images, thus resulting in better 3D generation quality." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 568, + 360, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 360, + 582 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 360, + 582 + ], + "type": "text", + "content": "3 SV3D: Novel Multi-view Synthesis" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 594, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 594, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 594, + 481, + 665 + ], + "type": "text", + "content": "Our main idea is to repurpose temporal consistency in a video diffusion model for spatial 3D consistency of an object. Specifically, we finetune SVD [2] to generate an orbital video around a 3D object, conditioned on a single-view image. This orbital video need not be at the same elevation, or at regularly spaced azimuth angles. SVD is well-suited for this task since it is trained to generate smooth and consistent videos on large-scale datasets of real and high-quality videos." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "text", + "content": "V. Voleti et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 113, + 394, + 232 + ], + "blocks": [ + { + "bbox": [ + 136, + 113, + 394, + 232 + ], + "lines": [ + { + "bbox": [ + 136, + 113, + 394, + 232 + ], + "spans": [ + { + "bbox": [ + 136, + 113, + 394, + 232 + ], + "type": "image", + "image_path": "12882c02dace176adb8a7308987935aef341e849263fcf5bef9dae01d0b9ae9f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 147, + 237, + 365, + 251 + ], + "lines": [ + { + "bbox": [ + 147, + 237, + 365, + 251 + ], + "spans": [ + { + "bbox": [ + 147, + 237, + 365, + 251 + ], + "type": "text", + "content": ":FullyConnected :Embedding :VAE Encoder" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 132, + 252, + 377, + 318 + ], + "lines": [ + { + "bbox": [ + 132, + 252, + 377, + 318 + ], + "spans": [ + { + "bbox": [ + 132, + 252, + 377, + 318 + ], + "type": "text", + "content": "Fig. 2: SV3D Architecture. We add the sinusoidal embedding of the camera orbit elevation and azimuth angles " + }, + { + "bbox": [ + 132, + 252, + 377, + 318 + ], + "type": "inline_equation", + "content": "(e, a)" + }, + { + "bbox": [ + 132, + 252, + 377, + 318 + ], + "type": "text", + "content": " to that of the noise step " + }, + { + "bbox": [ + 132, + 252, + 377, + 318 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 132, + 252, + 377, + 318 + ], + "type": "text", + "content": ", and feed the sum to the convolutional blocks in the UNet. We feed the single input image's CLIP embedding to the attention blocks, and concatenate its latent embedding to the noisy state " + }, + { + "bbox": [ + 132, + 252, + 377, + 318 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 132, + 252, + 377, + 318 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 402, + 122, + 473, + 167 + ], + "blocks": [ + { + "bbox": [ + 402, + 122, + 473, + 167 + ], + "lines": [ + { + "bbox": [ + 402, + 122, + 473, + 167 + ], + "spans": [ + { + "bbox": [ + 402, + 122, + 473, + 167 + ], + "type": "image", + "image_path": "6ad6fbf896b7ea7b537a65d35f76b1b60f5fab8ddfe00902463a218b7e50f49b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 425, + 171, + 448, + 178 + ], + "lines": [ + { + "bbox": [ + 425, + 171, + 448, + 178 + ], + "spans": [ + { + "bbox": [ + 425, + 171, + 448, + 178 + ], + "type": "text", + "content": "Static" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 395, + 279, + 477, + 311 + ], + "lines": [ + { + "bbox": [ + 395, + 279, + 477, + 311 + ], + "spans": [ + { + "bbox": [ + 395, + 279, + 477, + 311 + ], + "type": "text", + "content": "We use two types of orbits for training the SV3D models." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 404, + 184, + 473, + 238 + ], + "blocks": [ + { + "bbox": [ + 404, + 184, + 473, + 238 + ], + "lines": [ + { + "bbox": [ + 404, + 184, + 473, + 238 + ], + "spans": [ + { + "bbox": [ + 404, + 184, + 473, + 238 + ], + "type": "image", + "image_path": "f238593dc45e29a47af60466306c47f3e94968dec88978f5a3e2b18e9b09197d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 421, + 240, + 453, + 247 + ], + "lines": [ + { + "bbox": [ + 421, + 240, + 453, + 247 + ], + "spans": [ + { + "bbox": [ + 421, + 240, + 453, + 247 + ], + "type": "text", + "content": "Dynamic" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 395, + 257, + 477, + 277 + ], + "lines": [ + { + "bbox": [ + 395, + 257, + 477, + 277 + ], + "spans": [ + { + "bbox": [ + 395, + 257, + 477, + 277 + ], + "type": "text", + "content": "Fig.3:Static vs. DynamicOrbits." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 333, + 480, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 333, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 333, + 480, + 369 + ], + "type": "text", + "content": "The exposure to superior data quantity and quality makes it more generalizable and multi-view consistent, and the flexibility of the SVD architecture makes it amenable to be finetuned for camera controllability." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 371, + 480, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 371, + 480, + 455 + ], + "spans": [ + { + "bbox": [ + 130, + 371, + 480, + 455 + ], + "type": "text", + "content": "Some prior works attempt to leverage such properties by finetuning image diffusion models, training video diffusion models from scratch, or finetuning video diffusion models to generate pre-defined views at the same elevation (static orbit) around an object [2, 29]. However, we argue that these methods do not fully exploit the potential of video diffusion models. To the best of our knowledge, SV3D is the first video diffusion-based framework for controllable multi-view synthesis at " + }, + { + "bbox": [ + 130, + 371, + 480, + 455 + ], + "type": "inline_equation", + "content": "576 \\times 576" + }, + { + "bbox": [ + 130, + 371, + 480, + 455 + ], + "type": "text", + "content": " resolution (and subsequently for 3D generation)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 457, + 480, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 457, + 480, + 554 + ], + "spans": [ + { + "bbox": [ + 130, + 457, + 480, + 554 + ], + "type": "text", + "content": "Problem Setting. Formally, given an image " + }, + { + "bbox": [ + 130, + 457, + 480, + 554 + ], + "type": "inline_equation", + "content": "\\mathbf{I} \\in \\mathbb{R}^{3 \\times H \\times W}" + }, + { + "bbox": [ + 130, + 457, + 480, + 554 + ], + "type": "text", + "content": " of an object, our goal is to generate an orbital video " + }, + { + "bbox": [ + 130, + 457, + 480, + 554 + ], + "type": "inline_equation", + "content": "\\mathbf{J} \\in \\mathbb{R}^{K \\times 3 \\times H \\times W}" + }, + { + "bbox": [ + 130, + 457, + 480, + 554 + ], + "type": "text", + "content": " around the object of " + }, + { + "bbox": [ + 130, + 457, + 480, + 554 + ], + "type": "inline_equation", + "content": "K = 21" + }, + { + "bbox": [ + 130, + 457, + 480, + 554 + ], + "type": "text", + "content": " multi-view images along a camera trajectory " + }, + { + "bbox": [ + 130, + 457, + 480, + 554 + ], + "type": "inline_equation", + "content": "\\pmb{\\pi} \\in \\mathbb{R}^{K \\times 2} = \\{(e_i, a_i)\\}_{i=1}^K" + }, + { + "bbox": [ + 130, + 457, + 480, + 554 + ], + "type": "text", + "content": " as a sequence of " + }, + { + "bbox": [ + 130, + 457, + 480, + 554 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 457, + 480, + 554 + ], + "type": "text", + "content": " tuples of elevation " + }, + { + "bbox": [ + 130, + 457, + 480, + 554 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 130, + 457, + 480, + 554 + ], + "type": "text", + "content": " and azimuth " + }, + { + "bbox": [ + 130, + 457, + 480, + 554 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 130, + 457, + 480, + 554 + ], + "type": "text", + "content": " angles. We assume the camera always looks at the center of an object (origin of world coordinates), so any viewpoint can be specified by only two parameters: " + }, + { + "bbox": [ + 130, + 457, + 480, + 554 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 130, + 457, + 480, + 554 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 457, + 480, + 554 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 130, + 457, + 480, + 554 + ], + "type": "text", + "content": ". We generate this orbital video by iteratively denoising samples from a learned conditional distribution " + }, + { + "bbox": [ + 130, + 457, + 480, + 554 + ], + "type": "inline_equation", + "content": "p(\\mathbf{J}|\\mathbf{I}, \\pmb{\\pi})" + }, + { + "bbox": [ + 130, + 457, + 480, + 554 + ], + "type": "text", + "content": ", parameterized by a video diffusion model." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 557, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 481, + 665 + ], + "type": "text", + "content": "SV3D Architecture. As shown in Fig. 2, the architecture of SV3D builds on that of SVD consisting of a UNet with multiple layers, each layer containing sequences of one residual block with Conv3D layers, and two transformer blocks (spatial and temporal) with attention layers. (i) We remove the vector conditionings of 'fps id' and 'motion bucket id' since they are irrelevant for SV3D. (ii) The conditioning image is concatenated to the noisy latent state input " + }, + { + "bbox": [ + 130, + 557, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 130, + 557, + 481, + 665 + ], + "type": "text", + "content": " at noise timestep " + }, + { + "bbox": [ + 130, + 557, + 481, + 665 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 557, + 481, + 665 + ], + "type": "text", + "content": " to the UNet, after being embedded into latent space by the VAE encoder of SVD. (iii) The CLIP-embedding [39] matrix of the conditioning image is provided to the cross-attention layers of each transformer block as its key and" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 422, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 422, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 422, + 91, + 447, + 100 + ], + "type": "text", + "content": "SV3D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 119, + 325, + 247 + ], + "blocks": [ + { + "bbox": [ + 135, + 119, + 325, + 247 + ], + "lines": [ + { + "bbox": [ + 135, + 119, + 325, + 247 + ], + "spans": [ + { + "bbox": [ + 135, + 119, + 325, + 247 + ], + "type": "image", + "image_path": "cc3a266b1abb00711f934e48ce92e388c4a3f9fd8edb84a2c3658e78cbe0e122.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 250, + 478, + 283 + ], + "lines": [ + { + "bbox": [ + 132, + 250, + 478, + 283 + ], + "spans": [ + { + "bbox": [ + 132, + 250, + 478, + 283 + ], + "type": "text", + "content": "Fig. 4: Linear vs. Triangle CFG Scaling. Notice Fig. 5: LPIPS vs. Frame increased oversharping in the penultimate frame in Number. SV3D has the best re-linear scaling vs. our proposed triangle scaling. construction metric per frame." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 347, + 120, + 477, + 248 + ], + "blocks": [ + { + "bbox": [ + 347, + 120, + 477, + 248 + ], + "lines": [ + { + "bbox": [ + 347, + 120, + 477, + 248 + ], + "spans": [ + { + "bbox": [ + 347, + 120, + 477, + 248 + ], + "type": "image", + "image_path": "ae353a2baa0034bb01d4ff5f7d14c1f6e1b9128a12167fafbda9d0c558ca3168.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 289, + 482, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 373 + ], + "type": "text", + "content": "value, the query being the feature at that layer. (iv) The camera trajectory is fed into the residual blocks along with the diffusion noise timestep. The camera pose angles " + }, + { + "bbox": [ + 130, + 289, + 482, + 373 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 130, + 289, + 482, + 373 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 289, + 482, + 373 + ], + "type": "inline_equation", + "content": "a_i" + }, + { + "bbox": [ + 130, + 289, + 482, + 373 + ], + "type": "text", + "content": " and the noise timestep " + }, + { + "bbox": [ + 130, + 289, + 482, + 373 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 289, + 482, + 373 + ], + "type": "text", + "content": " are first embedded into sinusoidal position embeddings. Then, the camera pose embeddings are concatenated together, linearly transformed, and added to the noise timestep embedding. This is fed to every residual block, where they are added to the block's output feature (after being linearly transformed again to match the feature size)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 376, + 482, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 519 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 519 + ], + "type": "text", + "content": "Static v.s. Dynamic Orbits. As shown in Fig. 3, we design static and dynamic orbits to study the effects of camera pose conditioning. In a static orbit, the camera circles around an object at regularly-spaced azimuths at the same elevation angle as that in the conditioning image. The disadvantage with the static orbit is that we might not get any information about the top or bottom of the object depending on the conditioning elevation angle. In a dynamic orbit, the azimuths can be irregularly spaced, and the elevation can vary per view. To build a dynamic orbit, we sample a static orbit, add small random noise to the azimuth angles, and add a random weighted combination of sinusoids with different frequencies to its elevation. This provides temporal smoothness, and ensures that the camera trajectory loops around to end at the same azimuth and elevation as those of the conditioning image." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 520, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 520, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 520, + 482, + 567 + ], + "type": "text", + "content": "Thus, with this strategy, we are able to tackle all three aspects of generalization, controllability, and consistency in novel multi-view synthesis by leveraging video diffusion models, additionally conditioning on camera trajectory, and repurposing temporal consistency for spatial 3D object consistency, respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "Triangular CFG Scaling. SVD uses a linearly increasing scale for classifier-free guidance (CFG) from 1 to 4 across the generated frames. However, this scaling causes the last few frames in our generated orbits to be over-sharpened, as shown in Fig. 4 Frame 20. Since we generate videos looping back to the front-view image, we propose to use a triangle wave CFG scaling during inference: linearly increase CFG from 1 at the front view to 2.5 at the back view, then linearly decrease it back to 1 at the front view. Fig. 4 also demonstrates that our triangle CFG scaling produces more details in the back view (Frame 12)." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 100 + ], + "type": "text", + "content": "V. Voleti et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": "Models. We train three image-to-3D-video models finetuned from SVD. First, we train a pose-unconditioned model, " + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "inline_equation", + "content": "\\mathrm{SV3D}^u" + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": ", which generates a video of static orbit around an object while only conditioned on a single-view image. Note that unlike SVD-MV [2], we do not provide the elevation angle to the pose-unconditioned model, as we find that the model is able to infer it from the conditioning image. Our second model, the pose-conditioned " + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "inline_equation", + "content": "\\mathrm{SV3D}^c" + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": " is conditioned on the input image as well as a sequence of camera elevation and azimuth angles in an orbit, trained on dynamic orbits. Following SVD's [2] intuition to progressively increase the task difficulty during training, we train our third model, " + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "inline_equation", + "content": "\\mathrm{SV3D}^p" + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": ", by first finetuning SVD to produce static orbits unconditionally, then further finetuning on dynamic orbits with camera pose condition." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 249, + 482, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 249, + 482, + 345 + ], + "spans": [ + { + "bbox": [ + 130, + 249, + 482, + 345 + ], + "type": "text", + "content": "Training Details. We train SV3D on the Objverse dataset [9], which contains synthetic 3D objects covering a wide diversity. For each object, we render 21 frames around it on random color background at " + }, + { + "bbox": [ + 130, + 249, + 482, + 345 + ], + "type": "inline_equation", + "content": "576 \\times 576" + }, + { + "bbox": [ + 130, + 249, + 482, + 345 + ], + "type": "text", + "content": " resolution, field-of-view of 33.8 degrees. We choose to finetune the SVD-xt model to output 21 frames. All three models (SV3D" + }, + { + "bbox": [ + 130, + 249, + 482, + 345 + ], + "type": "inline_equation", + "content": "^u" + }, + { + "bbox": [ + 130, + 249, + 482, + 345 + ], + "type": "text", + "content": ", SV3D" + }, + { + "bbox": [ + 130, + 249, + 482, + 345 + ], + "type": "inline_equation", + "content": "^c" + }, + { + "bbox": [ + 130, + 249, + 482, + 345 + ], + "type": "text", + "content": ", SV3D" + }, + { + "bbox": [ + 130, + 249, + 482, + 345 + ], + "type": "inline_equation", + "content": "^p" + }, + { + "bbox": [ + 130, + 249, + 482, + 345 + ], + "type": "text", + "content": ") are trained for 105k iterations in total (SV3D" + }, + { + "bbox": [ + 130, + 249, + 482, + 345 + ], + "type": "inline_equation", + "content": "^p" + }, + { + "bbox": [ + 130, + 249, + 482, + 345 + ], + "type": "text", + "content": " is trained unconditionally for 55k iterations and conditionally for 50k iterations), with an effective batch size of 64 on 4 nodes of 8 80GB A100 GPUs for around 6 days. For more training details, please see the appendix." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 364, + 288, + 375 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 288, + 375 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 288, + 375 + ], + "type": "text", + "content": "3.1 Experiments and Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 385, + 482, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 385, + 482, + 470 + ], + "spans": [ + { + "bbox": [ + 130, + 385, + 482, + 470 + ], + "type": "text", + "content": "Datasets. We evaluate the SV3D-generated multi-view images on static and dynamic orbits on the unseen GSO [10] and OmniObject3D [56] datasets. Since many GSO objects are the same items with slightly different colors, we filter 300 objects from GSO to reduce redundancy and maintain diversity. For each object, we render ground truth static and dynamic orbit videos and pick the last frame of each video as the conditioning image. We also conduct a user study on novel view videos from a dataset of 22 real-world images. (More details in appendix.)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 471, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 471, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 471, + 482, + 567 + ], + "type": "text", + "content": "Metrics. We use the SV3D models to generate static and dynamic orbit videos corresponding to the ground truth camera trajectories in the evaluation datasets. We compare each generated frame with the corresponding ground truth frames, in terms of Learned Perceptual Similarity (LPIPS [63]), Peak Signal-to-Noise Ratio (PSNR), Structural SIMilarity (SSIM), Mean Squared-Error (MSE), and CLIP-score (CLIP-S). This range of metrics covers both pixel-level as well as semantic aspects. Note that testing on dynamic orbits evaluates the controllability of SV3D models, and all the metrics evaluate multi-view consistency." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 568, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 568, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 568, + 482, + 604 + ], + "type": "text", + "content": "Baselines. We compare SV3D with several recent NVS methods capable of generating arbitrary views, including Zero123 [26], Zero123-XL [8], SyncDreamer [27], Stable Zero123 [49], Free3D [64], EscherNet [21]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "Results. As shown in Tabs. 1 to 4, our SV3D models achieve state-of-the-art performance on novel multi-view synthesis. Tabs. 1 and 3 show results on static orbits, and include all our three models. We see that even our pose-unconditioned model " + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathrm{SV3D}^u" + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": " performs better than all prior methods. Tabs. 2 and 4 show results on dynamic orbits, and include our pose-conditioned models " + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathrm{SV3D}^c" + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathrm{SV3D}^p" + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 422, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 422, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 422, + 91, + 447, + 100 + ], + "type": "text", + "content": "SV3D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 479, + 339 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 479, + 339 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 479, + 339 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 479, + 339 + ], + "type": "image", + "image_path": "0e2e78ba032f8a75d450521803b42a294afe1fbb9621377fe07ebec3834f8d66.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 341, + 482, + 374 + ], + "lines": [ + { + "bbox": [ + 130, + 341, + 482, + 374 + ], + "spans": [ + { + "bbox": [ + 130, + 341, + 482, + 374 + ], + "type": "text", + "content": "Fig.6: Visual Comparison of Novel Multi-view Synthesis. SV3D is able to generate novel multi-views that are more detailed, faithful to the conditioning image, and multi-view consistent compared to the prior works." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 392, + 482, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 482, + 476 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 482, + 476 + ], + "type": "text", + "content": "Ablative Analyses. Interestingly, from Tabs. 1 and 3, we find that both " + }, + { + "bbox": [ + 130, + 392, + 482, + 476 + ], + "type": "inline_equation", + "content": "\\mathrm{SV3D}^c" + }, + { + "bbox": [ + 130, + 392, + 482, + 476 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 392, + 482, + 476 + ], + "type": "inline_equation", + "content": "\\mathrm{SV3D}^p" + }, + { + "bbox": [ + 130, + 392, + 482, + 476 + ], + "type": "text", + "content": " outperform " + }, + { + "bbox": [ + 130, + 392, + 482, + 476 + ], + "type": "inline_equation", + "content": "\\mathrm{SV3D}^u" + }, + { + "bbox": [ + 130, + 392, + 482, + 476 + ], + "type": "text", + "content": " on generations of static orbits, even though " + }, + { + "bbox": [ + 130, + 392, + 482, + 476 + ], + "type": "inline_equation", + "content": "\\mathrm{SV3D}^u" + }, + { + "bbox": [ + 130, + 392, + 482, + 476 + ], + "type": "text", + "content": " is trained specifically on static orbits. We also observe that " + }, + { + "bbox": [ + 130, + 392, + 482, + 476 + ], + "type": "inline_equation", + "content": "\\mathrm{SV3D}^p" + }, + { + "bbox": [ + 130, + 392, + 482, + 476 + ], + "type": "text", + "content": " achieves better metrics than " + }, + { + "bbox": [ + 130, + 392, + 482, + 476 + ], + "type": "inline_equation", + "content": "\\mathrm{SV3D}^c" + }, + { + "bbox": [ + 130, + 392, + 482, + 476 + ], + "type": "text", + "content": " on both static and dynamic orbits, making it the best performing SV3D model overall. This shows that progressive finetuning from easier (static) to harder (dynamic) tasks is indeed a favorable way to finetune a video diffusion model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 479, + 481, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 481, + 516 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 481, + 516 + ], + "type": "text", + "content": "Visual Comparisons in Fig. 6 further demonstrate that SV3D-generated images are more detailed, faithful to the conditioning image, and multi-view consistent compared to the prior works." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 518, + 482, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 518, + 482, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 518, + 482, + 579 + ], + "type": "text", + "content": "Quality Per Frame. We also observe from Fig. 5 that SV3D produces better quality at every frame. We plot the average LPIPS value for each generated frame, across generated GSO static orbit videos. The quality is generally worse around the back view, and better at the beginning and the end (i.e. near the conditioning image), as expected." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "User Study on Real-World Images. We conducted a user survey to study human preference between static orbital videos generated by SV3D and by other methods. We asked 30 users to pick one between our SV3D-generated static video and other method-generated video as the best orbital video for the corresponding image, using 22 real-world images. We noted that users preferred SV3D-generated videos over Zero123XL, Stable Zero123, EscherNet, and Free3D, " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "96\\%" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "99\\%" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "96\\%" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "98\\%" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " of the time, respectively." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "text", + "content": "V.Voleti et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 137, + 303, + 226 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 304, + 136 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 304, + 136 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 304, + 136 + ], + "type": "text", + "content": "Table 1: Evaluation of novel multi-view synthesis on GSO static orbits" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 137, + 303, + 226 + ], + "lines": [ + { + "bbox": [ + 133, + 137, + 303, + 226 + ], + "spans": [ + { + "bbox": [ + 133, + 137, + 303, + 226 + ], + "type": "table", + "html": "
ModelLPIPS↓PSNR↑SSIM↑CLIP-S↑MSE↓
SyncDreamer [27]0.1715.780.760.870.03
Zero123 [26]0.1317.290.790.850.04
Zero123XL [8]0.1417.110.780.850.04
Stable Zero123 [49]0.1318.340.780.850.05
Free3D [64]0.1516.180.790.840.04
EscherNet [21]0.1316.730.790.850.03
SV3Du0.0921.140.870.890.02
SV3Dc0.0920.560.870.880.02
SV3Dp0.0821.260.880.890.02
", + "image_path": "0a5b238eef1ca70ca08d4de0f6c7600701c355fe3da1c83f076b00ba6559c1f7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 307, + 137, + 479, + 213 + ], + "blocks": [ + { + "bbox": [ + 305, + 114, + 480, + 136 + ], + "lines": [ + { + "bbox": [ + 305, + 114, + 480, + 136 + ], + "spans": [ + { + "bbox": [ + 305, + 114, + 480, + 136 + ], + "type": "text", + "content": "Table 2: Evaluation of novel multi-view synthesis on GSO dynamic orbits" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 137, + 479, + 213 + ], + "lines": [ + { + "bbox": [ + 307, + 137, + 479, + 213 + ], + "spans": [ + { + "bbox": [ + 307, + 137, + 479, + 213 + ], + "type": "table", + "html": "
ModelLPIPS↓PSNR↑SSIM↑CLIP-S↑MSE↓
Zero123 [26]0.1416.990.790.840.04
Zero123XL [8]0.1416.730.780.840.04
Stable Zero123 [49]0.1318.040.780.850.05
Free3D [64]0.1814.930.770.830.05
EscherNet [21]0.1316.470.790.840.03
SV3Dc0.1019.990.860.870.02
SV3Dp0.0920.380.870.870.02
", + "image_path": "f1259838032aeaecffe1ac122e0dcf3f5de7695a76b54a4cac648d80eb0261bc.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 133, + 257, + 303, + 339 + ], + "blocks": [ + { + "bbox": [ + 132, + 235, + 304, + 256 + ], + "lines": [ + { + "bbox": [ + 132, + 235, + 304, + 256 + ], + "spans": [ + { + "bbox": [ + 132, + 235, + 304, + 256 + ], + "type": "text", + "content": "Table 3: Evaluation of novel multi-view synthesis on OmniObject3D static orbits" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 257, + 303, + 339 + ], + "lines": [ + { + "bbox": [ + 133, + 257, + 303, + 339 + ], + "spans": [ + { + "bbox": [ + 133, + 257, + 303, + 339 + ], + "type": "table", + "html": "
ModelLPIPS↓PSNR↑SSIM↑CLIP-S↑MSE↓
Zero123 [26]0.1715.500.760.830.05
Zero123XL [8]0.1815.360.750.830.06
Stable Zero123 [49]0.1516.860.770.840.06
Free3D [64]0.1615.290.780.830.05
EscherNet [21]0.1714.630.740.830.05
SV3Du0.1019.680.860.860.02
SV3Dc0.1019.500.850.850.02
SV3Dp0.1019.910.860.860.02
", + "image_path": "713505754bfb1171b42f2d6fcacbcfdf682d2880b9244881404c034c95d3d8df.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 307, + 257, + 479, + 332 + ], + "blocks": [ + { + "bbox": [ + 305, + 235, + 481, + 256 + ], + "lines": [ + { + "bbox": [ + 305, + 235, + 481, + 256 + ], + "spans": [ + { + "bbox": [ + 305, + 235, + 481, + 256 + ], + "type": "text", + "content": "Table 4: Evaluation of novel multi-view synthesis on OmniObject3D dynamic orbits" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 257, + 479, + 332 + ], + "lines": [ + { + "bbox": [ + 307, + 257, + 479, + 332 + ], + "spans": [ + { + "bbox": [ + 307, + 257, + 479, + 332 + ], + "type": "table", + "html": "
ModelLPIPS↓PSNR↑SSIM↑CLIP-S↑MSE↓
Zero123 [26]0.1615.780.770.820.05
Zero123XL [8]0.1715.490.760.830.05
Stable Zero123 [49]0.1416.740.770.830.05
Free3D [64]0.1914.280.760.820.06
EscherNet [21]0.1615.050.760.830.05
SV3Dc0.1019.210.850.840.02
SV3Dp0.1019.280.850.850.02
", + "image_path": "f3a2aa9278279c0f943e0cde03ee8fe8901f7e9479bce3b26c778e71af14d23e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 344, + 442, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 344, + 442, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 344, + 442, + 358 + ], + "type": "text", + "content": "4 3D Generation from a Single Image Using SV3D" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 373, + 481, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 373, + 481, + 421 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 481, + 421 + ], + "type": "text", + "content": "We then generate 3D meshes of objects from a single image by leveraging SV3D. One way is to use the generated static/dynamic orbital samples from SV3D as direct reconstruction targets. Another way is to use SV3D as diffusion guidance with Score Distillation Sampling (SDS) loss [37]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 422, + 482, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 482, + 493 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 482, + 493 + ], + "type": "text", + "content": "Since SV3D produces more consistent multi-views compared to existing works, we already observe higher-quality 3D reconstructions by only using SV3D outputs for reconstruction when compared to existing works. However, we observe that this naive approach often leads to artifacts like baked-in illumination, rough surfaces, and noisy texture, especially for the unseen regions in the reference orbit. Thus, we further propose several techniques to address these issues." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 496, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 496, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 496, + 482, + 604 + ], + "type": "text", + "content": "Coarse-to-Fine Training. We adopt a two-stage, coarse-to-fine training scheme to generate a 3D mesh from input images, similar to [23, 38]. Fig. 7 illustrate an overview of our 3D optimization pipeline. In the coarse stage, we train an Instant-NGP [33] NeRF representation to reconstruct the SV3D-generated images (i.e. without SDS loss) at a lower resolution. In the fine stage, we extract a mesh from the trained NeRF using marching cubes [7], and adopt a DMTet [44] representation to finetune the 3D mesh using SDS-based diffusion guidance from SV3D at full-resolution. Finally, we use xatlas [60] to perform the UV unwrapping and export the output object mesh." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "Disentangled Illumination Model. Similar to other recent 3D object generation methods [23, 31, 37], our output target is a mesh with a diffuse texture. Typically, such SDS-based optimization techniques use random illuminations at every iteration. However, our SV3D-generated videos are under consistent illumination, i.e., the lighting remains static while the camera circles around an object." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 422, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 422, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 422, + 91, + 447, + 100 + ], + "type": "text", + "content": "SV3D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 141, + 115, + 471, + 247 + ], + "blocks": [ + { + "bbox": [ + 141, + 115, + 471, + 247 + ], + "lines": [ + { + "bbox": [ + 141, + 115, + 471, + 247 + ], + "spans": [ + { + "bbox": [ + 141, + 115, + 471, + 247 + ], + "type": "image", + "image_path": "5987a83388c37819313fdba8e4461418165ad0f7445d54136e20437a06efa447.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 255, + 482, + 322 + ], + "lines": [ + { + "bbox": [ + 130, + 255, + 482, + 322 + ], + "spans": [ + { + "bbox": [ + 130, + 255, + 482, + 322 + ], + "type": "text", + "content": "Fig.7: 3D Optimization Pipeline. We perform a two-stage optimization. In the short first stage, we extract the general shape, texture and illumination from the SV3D generated multi-view images. In the second stage, we extract a mesh with marching cubes and use DMTet to further optimize the shape, texture and illumination. We not only use the SV3D-generated images but a soft-masked SDS loss for unseen areas. Dashed red lines represent backpropagation into the differentiable rendering pipeline." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 153, + 328, + 178, + 381 + ], + "blocks": [ + { + "bbox": [ + 153, + 328, + 178, + 381 + ], + "lines": [ + { + "bbox": [ + 153, + 328, + 178, + 381 + ], + "spans": [ + { + "bbox": [ + 153, + 328, + 178, + 381 + ], + "type": "image", + "image_path": "5fb4bcf394d5d910bdad223a29f52256c4b8375432c15acf8e41f9abc8b6df6b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 156, + 387, + 176, + 439 + ], + "blocks": [ + { + "bbox": [ + 156, + 387, + 176, + 439 + ], + "lines": [ + { + "bbox": [ + 156, + 387, + 176, + 439 + ], + "spans": [ + { + "bbox": [ + 156, + 387, + 176, + 439 + ], + "type": "image", + "image_path": "b1fd04eee9a26b38127a2922f27179c09b2c5f689a5b2f9845930264b6ec190c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 151, + 445, + 179, + 457 + ], + "lines": [ + { + "bbox": [ + 151, + 445, + 179, + 457 + ], + "spans": [ + { + "bbox": [ + 151, + 445, + 179, + 457 + ], + "type": "text", + "content": "Ref. images in static orbit" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 195, + 329, + 223, + 382 + ], + "blocks": [ + { + "bbox": [ + 195, + 329, + 223, + 382 + ], + "lines": [ + { + "bbox": [ + 195, + 329, + 223, + 382 + ], + "spans": [ + { + "bbox": [ + 195, + 329, + 223, + 382 + ], + "type": "image", + "image_path": "671b192150010ac14eb20f5708976a826ad85e32c45dd3ec5c955b6436252316.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 195, + 384, + 220, + 441 + ], + "blocks": [ + { + "bbox": [ + 195, + 384, + 220, + 441 + ], + "lines": [ + { + "bbox": [ + 195, + 384, + 220, + 441 + ], + "spans": [ + { + "bbox": [ + 195, + 384, + 220, + 441 + ], + "type": "image", + "image_path": "0aa46e836604df6c791f979614c2ccd5b4bc2c345168216bf0c4ab3f241d5b3a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 197, + 445, + 221, + 457 + ], + "lines": [ + { + "bbox": [ + 197, + 445, + 221, + 457 + ], + "spans": [ + { + "bbox": [ + 197, + 445, + 221, + 457 + ], + "type": "text", + "content": "3D from static orbit" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 241, + 329, + 269, + 381 + ], + "blocks": [ + { + "bbox": [ + 241, + 329, + 269, + 381 + ], + "lines": [ + { + "bbox": [ + 241, + 329, + 269, + 381 + ], + "spans": [ + { + "bbox": [ + 241, + 329, + 269, + 381 + ], + "type": "image", + "image_path": "79e021486db2f103fbc401784306cab1644763354e4e721b7bbd123d2021c78a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 242, + 383, + 266, + 441 + ], + "blocks": [ + { + "bbox": [ + 242, + 383, + 266, + 441 + ], + "lines": [ + { + "bbox": [ + 242, + 383, + 266, + 441 + ], + "spans": [ + { + "bbox": [ + 242, + 383, + 266, + 441 + ], + "type": "image", + "image_path": "447888e32730a3ab3a5d887bb2a79dfe40d3e06d6d5271320cb26db27edd5135.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 239, + 445, + 271, + 457 + ], + "lines": [ + { + "bbox": [ + 239, + 445, + 271, + 457 + ], + "spans": [ + { + "bbox": [ + 239, + 445, + 271, + 457 + ], + "type": "text", + "content": "3D from dynamic orbit" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 305, + 329, + 358, + 374 + ], + "blocks": [ + { + "bbox": [ + 305, + 329, + 358, + 374 + ], + "lines": [ + { + "bbox": [ + 305, + 329, + 358, + 374 + ], + "spans": [ + { + "bbox": [ + 305, + 329, + 358, + 374 + ], + "type": "image", + "image_path": "4f0ed928839d0d8218b7518328d81d2b74e9c02a17aaff33a944f63e08ca4072.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 359, + 329, + 411, + 374 + ], + "blocks": [ + { + "bbox": [ + 359, + 329, + 411, + 374 + ], + "lines": [ + { + "bbox": [ + 359, + 329, + 411, + 374 + ], + "spans": [ + { + "bbox": [ + 359, + 329, + 411, + 374 + ], + "type": "image", + "image_path": "aedeea4d1746ea26fe4ca91a08ab571a565b166952c5fad0f90461f52eb22694.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 416, + 329, + 468, + 374 + ], + "blocks": [ + { + "bbox": [ + 416, + 329, + 468, + 374 + ], + "lines": [ + { + "bbox": [ + 416, + 329, + 468, + 374 + ], + "spans": [ + { + "bbox": [ + 416, + 329, + 468, + 374 + ], + "type": "image", + "image_path": "3d48c3019a3a7f02f19c339f810d8c6b89cd3e8d3bb0774746da33172408360f.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 313, + 380, + 360, + 426 + ], + "blocks": [ + { + "bbox": [ + 130, + 459, + 291, + 503 + ], + "lines": [ + { + "bbox": [ + 130, + 459, + 291, + 503 + ], + "spans": [ + { + "bbox": [ + 130, + 459, + 291, + 503 + ], + "type": "text", + "content": "Fig.8: Influence of Training Orbits. We show that using a dynamic orbit is crucial to 3D generations that are complete from diverse views." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 380, + 360, + 426 + ], + "lines": [ + { + "bbox": [ + 313, + 380, + 360, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 360, + 426 + ], + "type": "image", + "image_path": "789b011b0b7a02615f4bbbd432e0296095ac0502f6de6847d394c02e962a9a5b.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 316, + 430, + 351, + 437 + ], + "lines": [ + { + "bbox": [ + 316, + 430, + 351, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 430, + 351, + 437 + ], + "type": "text", + "content": "No SDS loss" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 294, + 449, + 479, + 503 + ], + "lines": [ + { + "bbox": [ + 294, + 449, + 479, + 503 + ], + "spans": [ + { + "bbox": [ + 294, + 449, + 479, + 503 + ], + "type": "text", + "content": "Fig.9: Influence of SDS. Using our masked SDS loss, we are able to fill in unseen surfaces in the training orbit, producing a cleaner result without oversaturation or blurry artifacts caused by naive SDS." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 365, + 380, + 412, + 426 + ], + "blocks": [ + { + "bbox": [ + 365, + 380, + 412, + 426 + ], + "lines": [ + { + "bbox": [ + 365, + 380, + 412, + 426 + ], + "spans": [ + { + "bbox": [ + 365, + 380, + 412, + 426 + ], + "type": "image", + "image_path": "d40e9b9de18b4d2252ca9c1c6b35d420a5ffeb0d43704dcb14aa16324e004ee7.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 375, + 430, + 400, + 437 + ], + "lines": [ + { + "bbox": [ + 375, + 430, + 400, + 437 + ], + "spans": [ + { + "bbox": [ + 375, + 430, + 400, + 437 + ], + "type": "text", + "content": "SDS loss" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 420, + 380, + 467, + 426 + ], + "blocks": [ + { + "bbox": [ + 420, + 380, + 467, + 426 + ], + "lines": [ + { + "bbox": [ + 420, + 380, + 467, + 426 + ], + "spans": [ + { + "bbox": [ + 420, + 380, + 467, + 426 + ], + "type": "image", + "image_path": "cada0c9d47e64e8fc135525873334996b3cfe6ceedb5abaa1f14b17b5ab85e30.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 415, + 430, + 464, + 437 + ], + "lines": [ + { + "bbox": [ + 415, + 430, + 464, + 437 + ], + "spans": [ + { + "bbox": [ + 415, + 430, + 464, + 437 + ], + "type": "text", + "content": "Masked SDS loss" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "bbox": [ + 130, + 520, + 480, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 520, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 520, + 480, + 555 + ], + "type": "text", + "content": "Hence, to disentangle lighting effects and obtain a cleaner texture, we propose to fit a simple illumination model of 24 Spherical Gaussians (SG) inspired by prior decomposition methods [4,62]." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 130, + 556, + 480, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 556, + 480, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 556, + 480, + 604 + ], + "type": "text", + "content": "We model white light and hence only use a scalar amplitude for the SGs. We only consider Lambertian shading, where the cosine shading term is approximated with another SG. We learn the parameters of the illumination SGs using a reconstruction loss between the rendered images and SV3D-generated images." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": "Inspired by [13, 37] we reduce baked-in illumination with a loss term that replicates the HSV-value component of the input image " + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": " with the rendered illumination " + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{illum}} = |V(\\mathbf{I}) - L|^2" + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "inline_equation", + "content": "V(\\pmb {c}) = \\max (c_r,c_g,c_b)" + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": ". Given these changes, our disentangled illumination model is able to express lighting variation properly and can severely reduce baked-in illumination. Fig. 10 shows sample" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 100 + ], + "type": "text", + "content": "V. Voleti et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "reconstructions with and without our illumination modelling. From the results, it is clear that we are able to disentangle the illumination effects from the base color (e.g., dark side of the school bus)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 170, + 356, + 182 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 170, + 356, + 182 + ], + "spans": [ + { + "bbox": [ + 131, + 170, + 356, + 182 + ], + "type": "text", + "content": "4.1 3D Optimization Strategies and Losses" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 192, + 479, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 192, + 479, + 263 + ], + "spans": [ + { + "bbox": [ + 130, + 192, + 479, + 263 + ], + "type": "text", + "content": "Reconstruction via Photometric Losses. Intuitively, we can treat the SV3D-generated images as multi-view pseudo-ground truth, and apply 2D reconstruction losses to train the 3D models. In this case, we apply photometric losses on the rendered images from NeRF or DMTet, including the pixel-level MSE loss, mask loss, and a perceptual LPIPS [63] loss. These photometric losses also optimize our illumination model through the differential rendering pipeline." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 266, + 337, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 266, + 337, + 420 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 337, + 420 + ], + "type": "text", + "content": "Training Orbits. For 3D generation, we use SV3D to generate multi-view images following a camera orbit " + }, + { + "bbox": [ + 130, + 266, + 337, + 420 + ], + "type": "inline_equation", + "content": "\\pi_{\\mathrm{ref}}" + }, + { + "bbox": [ + 130, + 266, + 337, + 420 + ], + "type": "text", + "content": " , referred to as the reference orbit (also see Fig. 7 for the overview). Fig. 8 shows sample reconstruction with using static and dynamic orbital outputs from SV3D. Using a dynamic orbit for training is crucial to high-quality 3D outputs when viewed from various angles, since some top/bottom views are missing in the static orbit (fixed elevation). Hence, for " + }, + { + "bbox": [ + 130, + 266, + 337, + 420 + ], + "type": "inline_equation", + "content": "\\mathrm{SV3D}^c" + }, + { + "bbox": [ + 130, + 266, + 337, + 420 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 266, + 337, + 420 + ], + "type": "inline_equation", + "content": "\\mathrm{SV3D}^p" + }, + { + "bbox": [ + 130, + 266, + 337, + 420 + ], + "type": "text", + "content": " , we render images on a dynamic orbit whose elevation follows a sine function to ensure that top and bo" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 355, + 280, + 408, + 312 + ], + "blocks": [ + { + "bbox": [ + 355, + 280, + 408, + 312 + ], + "lines": [ + { + "bbox": [ + 355, + 280, + 408, + 312 + ], + "spans": [ + { + "bbox": [ + 355, + 280, + 408, + 312 + ], + "type": "image", + "image_path": "d59104d1140387572463183ae508657a044358a3d5ca9d0739bf10a2b1286687.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 316, + 412, + 323 + ], + "lines": [ + { + "bbox": [ + 343, + 316, + 412, + 323 + ], + "spans": [ + { + "bbox": [ + 343, + 316, + 412, + 323 + ], + "type": "text", + "content": "Constant Illumination" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 420, + 280, + 473, + 312 + ], + "blocks": [ + { + "bbox": [ + 420, + 280, + 473, + 312 + ], + "lines": [ + { + "bbox": [ + 420, + 280, + 473, + 312 + ], + "spans": [ + { + "bbox": [ + 420, + 280, + 473, + 312 + ], + "type": "image", + "image_path": "e2a63af3ffafaff4160f5fadf2e928ec4b17d3f63c5f1a2b3131aa56437ddefe.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 417, + 316, + 470, + 323 + ], + "lines": [ + { + "bbox": [ + 417, + 316, + 470, + 323 + ], + "spans": [ + { + "bbox": [ + 417, + 316, + 470, + 323 + ], + "type": "text", + "content": "SGs Illumination" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 342, + 325, + 481, + 401 + ], + "lines": [ + { + "bbox": [ + 342, + 325, + 481, + 401 + ], + "spans": [ + { + "bbox": [ + 342, + 325, + 481, + 401 + ], + "type": "text", + "content": "Fig.10: Constant vs. SGs Illumination. Our SGs-based reconstructions do not exhibit darkening on the side of the bus, which enables easier and more convincing relighting for downstream applications." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 422, + 479, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 479, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 479, + 506 + ], + "type": "text", + "content": "SV3D-Based SDS Loss. In addition to the reconstruction losses, we can also use SV3D via score-distillation sampling (SDS) [37, 58]. Fig. 9 shows sample reconstructions with and without using SDS losses. As shown in Fig. 9 (left), although training with a dynamic orbit improves overall visibility, we observe that sometimes the output texture is still noisy, perhaps due to partial visibility, self-occlusions, or inconsistent texture/shape between images. Hence, we handle those unseen areas using SDS loss [37] with SV3D as a diffusion guidance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 506, + 480, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 506, + 480, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 506, + 480, + 604 + ], + "type": "text", + "content": "We sample a random camera orbit " + }, + { + "bbox": [ + 130, + 506, + 480, + 604 + ], + "type": "inline_equation", + "content": "\\pi_{\\mathrm{rand}}" + }, + { + "bbox": [ + 130, + 506, + 480, + 604 + ], + "type": "text", + "content": ", and use our 3D NeRF/DMTet parameterized by " + }, + { + "bbox": [ + 130, + 506, + 480, + 604 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 506, + 480, + 604 + ], + "type": "text", + "content": " to render the views " + }, + { + "bbox": [ + 130, + 506, + 480, + 604 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{J}}" + }, + { + "bbox": [ + 130, + 506, + 480, + 604 + ], + "type": "text", + "content": " of the object along " + }, + { + "bbox": [ + 130, + 506, + 480, + 604 + ], + "type": "inline_equation", + "content": "\\pi_{\\mathrm{rand}}" + }, + { + "bbox": [ + 130, + 506, + 480, + 604 + ], + "type": "text", + "content": ". Then, noise " + }, + { + "bbox": [ + 130, + 506, + 480, + 604 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 506, + 480, + 604 + ], + "type": "text", + "content": " at level " + }, + { + "bbox": [ + 130, + 506, + 480, + 604 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 506, + 480, + 604 + ], + "type": "text", + "content": " is added to the latent embedding " + }, + { + "bbox": [ + 130, + 506, + 480, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 130, + 506, + 480, + 604 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 130, + 506, + 480, + 604 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{J}}" + }, + { + "bbox": [ + 130, + 506, + 480, + 604 + ], + "type": "text", + "content": ", and the following SDS loss (taken expectation over " + }, + { + "bbox": [ + 130, + 506, + 480, + 604 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 506, + 480, + 604 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 506, + 480, + 604 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 506, + 480, + 604 + ], + "type": "text", + "content": ") is backpropagated through the differentiable rendering pipeline: " + }, + { + "bbox": [ + 130, + 506, + 480, + 604 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{sds}} = w(t)\\bigl (\\epsilon_{\\phi}(\\mathbf{z}_t;\\mathbf{I},\\pi_{\\mathrm{rand}},t) - \\epsilon \\bigr)\\frac{\\partial\\hat{\\mathbf{J}}}{\\partial\\theta}" + }, + { + "bbox": [ + 130, + 506, + 480, + 604 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 506, + 480, + 604 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 130, + 506, + 480, + 604 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 130, + 506, + 480, + 604 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 506, + 480, + 604 + ], + "type": "text", + "content": "-dependent weight, " + }, + { + "bbox": [ + 130, + 506, + 480, + 604 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 506, + 480, + 604 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 506, + 480, + 604 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}" + }, + { + "bbox": [ + 130, + 506, + 480, + 604 + ], + "type": "text", + "content": " are the added and predicted noise, and " + }, + { + "bbox": [ + 130, + 506, + 480, + 604 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 506, + 480, + 604 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 506, + 480, + 604 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 506, + 480, + 604 + ], + "type": "text", + "content": " are the parameters of SV3D and NeRF/DMTet, respectively. See Fig. 7 for an illustration of these loss functions in the overall pipeline." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": "Masked SDS Loss. As shown in Fig. 9 (middle), in our experiments we found that adding the SDS loss naively can cause unstable training and unfaithful texture to the input images such as oversaturation or blurry artifacts. Therefore, we design a soft masking mechanism to only apply SDS loss on the unseen/occluded areas, allowing it to inpaint the missing details while preserving the texture of" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "type": "text", + "content": "SV3D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 114, + 460, + 221 + ], + "blocks": [ + { + "bbox": [ + 149, + 114, + 460, + 221 + ], + "lines": [ + { + "bbox": [ + 149, + 114, + 460, + 221 + ], + "spans": [ + { + "bbox": [ + 149, + 114, + 460, + 221 + ], + "type": "image", + "image_path": "e96e640b561758b8f3934e1a727a4386cf96519b602cbeaf342608a3fdc58e0a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 224, + 481, + 269 + ], + "lines": [ + { + "bbox": [ + 130, + 224, + 481, + 269 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 481, + 269 + ], + "type": "text", + "content": "Fig. 11: Visual Comparison of 3D Meshes. For each object, we show the conditioning image (left), and the output meshes rendered in two different views. Our generated meshes are more detailed, faithful to input images, and consistent in 3D. This demonstrates the quality of novel multi-view synthesis by our SV3D model." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 284, + 480, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 284, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 284, + 480, + 319 + ], + "type": "text", + "content": "clearly-visible surfaces in the training orbit (as seen in Fig. 9 (right)). Also, we only apply the masked SDS loss in the final stage of DMTet optimization, which greatly increased the convergence speed." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 320, + 482, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 320, + 482, + 504 + ], + "spans": [ + { + "bbox": [ + 130, + 320, + 482, + 504 + ], + "type": "text", + "content": "We apply SDS loss on only those pixels in the random orbit views that are not likely visible in the reference orbit views. For this, we first render the object from the random camera orbit " + }, + { + "bbox": [ + 130, + 320, + 482, + 504 + ], + "type": "inline_equation", + "content": "\\pi_{\\mathrm{rand}}" + }, + { + "bbox": [ + 130, + 320, + 482, + 504 + ], + "type": "text", + "content": ". For each random camera view, we obtain the visible surface points " + }, + { + "bbox": [ + 130, + 320, + 482, + 504 + ], + "type": "inline_equation", + "content": "p \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 320, + 482, + 504 + ], + "type": "text", + "content": " and their corresponding surface normals " + }, + { + "bbox": [ + 130, + 320, + 482, + 504 + ], + "type": "inline_equation", + "content": "\\boldsymbol{n}" + }, + { + "bbox": [ + 130, + 320, + 482, + 504 + ], + "type": "text", + "content": ". Then, for each reference camera " + }, + { + "bbox": [ + 130, + 320, + 482, + 504 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 320, + 482, + 504 + ], + "type": "text", + "content": ", we calculate the view directions " + }, + { + "bbox": [ + 130, + 320, + 482, + 504 + ], + "type": "inline_equation", + "content": "\\boldsymbol{v}_i" + }, + { + "bbox": [ + 130, + 320, + 482, + 504 + ], + "type": "text", + "content": " of the surface " + }, + { + "bbox": [ + 130, + 320, + 482, + 504 + ], + "type": "inline_equation", + "content": "\\boldsymbol{p}" + }, + { + "bbox": [ + 130, + 320, + 482, + 504 + ], + "type": "text", + "content": " towards its position " + }, + { + "bbox": [ + 130, + 320, + 482, + 504 + ], + "type": "inline_equation", + "content": "\\bar{\\pi}_{\\mathrm{ref}}^i \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 320, + 482, + 504 + ], + "type": "text", + "content": " (calculated from " + }, + { + "bbox": [ + 130, + 320, + 482, + 504 + ], + "type": "inline_equation", + "content": "\\pi_{\\mathrm{ref}}^i \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 130, + 320, + 482, + 504 + ], + "type": "text", + "content": ") as " + }, + { + "bbox": [ + 130, + 320, + 482, + 504 + ], + "type": "inline_equation", + "content": "\\boldsymbol{v}_i = \\frac{\\bar{\\pi}_{\\mathrm{ref}}^i - \\boldsymbol{p}}{||\\bar{\\pi}_{\\mathrm{ref}}^i - \\boldsymbol{p}||}" + }, + { + "bbox": [ + 130, + 320, + 482, + 504 + ], + "type": "text", + "content": ". We infer the visibility of this surface from the reference camera based on the dot product between " + }, + { + "bbox": [ + 130, + 320, + 482, + 504 + ], + "type": "inline_equation", + "content": "\\boldsymbol{v}_i" + }, + { + "bbox": [ + 130, + 320, + 482, + 504 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 320, + 482, + 504 + ], + "type": "inline_equation", + "content": "\\boldsymbol{n}" + }, + { + "bbox": [ + 130, + 320, + 482, + 504 + ], + "type": "text", + "content": " i.e. " + }, + { + "bbox": [ + 130, + 320, + 482, + 504 + ], + "type": "inline_equation", + "content": "\\boldsymbol{v}_i \\cdot \\boldsymbol{n}" + }, + { + "bbox": [ + 130, + 320, + 482, + 504 + ], + "type": "text", + "content": ". Since higher values roughly indicate more visibility of the surface from the reference camera, we chose that reference camera " + }, + { + "bbox": [ + 130, + 320, + 482, + 504 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 320, + 482, + 504 + ], + "type": "text", + "content": " that has maximum likelihood of visibility: " + }, + { + "bbox": [ + 130, + 320, + 482, + 504 + ], + "type": "inline_equation", + "content": "c = \\max_i (\\boldsymbol{v}_i \\cdot \\boldsymbol{n})" + }, + { + "bbox": [ + 130, + 320, + 482, + 504 + ], + "type": "text", + "content": ". As we only want to apply SDS loss to unseen or grazing angle areas from " + }, + { + "bbox": [ + 130, + 320, + 482, + 504 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 320, + 482, + 504 + ], + "type": "text", + "content": ", we use the smoothstep function " + }, + { + "bbox": [ + 130, + 320, + 482, + 504 + ], + "type": "inline_equation", + "content": "f_s" + }, + { + "bbox": [ + 130, + 320, + 482, + 504 + ], + "type": "text", + "content": " to smoothly clip to " + }, + { + "bbox": [ + 130, + 320, + 482, + 504 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 320, + 482, + 504 + ], + "type": "text", + "content": "'s visibility range " + }, + { + "bbox": [ + 130, + 320, + 482, + 504 + ], + "type": "inline_equation", + "content": "\\boldsymbol{v}_c \\cdot \\boldsymbol{n}" + }, + { + "bbox": [ + 130, + 320, + 482, + 504 + ], + "type": "text", + "content": ". In this way, we create a pseudo visibility mask " + }, + { + "bbox": [ + 130, + 320, + 482, + 504 + ], + "type": "inline_equation", + "content": "M = 1 - f_s(\\boldsymbol{v}_c \\cdot \\boldsymbol{n}, 0, 0.5)" + }, + { + "bbox": [ + 130, + 320, + 482, + 504 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 320, + 482, + 504 + ], + "type": "inline_equation", + "content": "f_s(x; f_0, f_1) = \\hat{x}^2(3 - 2x)" + }, + { + "bbox": [ + 130, + 320, + 482, + 504 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 320, + 482, + 504 + ], + "type": "inline_equation", + "content": "\\hat{x} = \\frac{x - f_0}{f_1 - f_0}" + }, + { + "bbox": [ + 130, + 320, + 482, + 504 + ], + "type": "text", + "content": ". Thus, " + }, + { + "bbox": [ + 130, + 320, + 482, + 504 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 320, + 482, + 504 + ], + "type": "text", + "content": " is calculated for each random camera render, and the combined visibility mask " + }, + { + "bbox": [ + 130, + 320, + 482, + 504 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 130, + 320, + 482, + 504 + ], + "type": "text", + "content": " is applied to SDS loss: " + }, + { + "bbox": [ + 130, + 320, + 482, + 504 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{mask-sds}} = \\mathbf{M}\\mathcal{L}_{\\mathrm{sds}}" + }, + { + "bbox": [ + 130, + 320, + 482, + 504 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 506, + 482, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 506, + 482, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 506, + 482, + 601 + ], + "type": "text", + "content": "Geometric Priors. Since our rendering-based optimization operates at the image level, we adopt several geometric priors to regularize the output shapes. We add a smooth depth loss from RegNeRF [35] and a bilateral normal smoothness loss [5] to encourage smooth 3D surfaces where the projected image gradients are low. Moreover, we obtain normal estimates from Omnidata [11] and calculate a mono normal loss similar to MonoSDF [61], which can effectively reduce noisy surfaces in the output mesh. Further details about the training losses and optimization process are available in the appendix." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 620, + 288, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 620, + 288, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 620, + 288, + 632 + ], + "type": "text", + "content": "4.2 Experiments and Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Due to the strong regularization, we only require 600 steps in the coarse stage and 1000 in the fine stage. Overall, the entire mesh extraction process takes " + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\approx 8" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 91, + 228, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 228, + 100 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 228, + 100 + ], + "type": "text", + "content": "V. Voleti et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 115, + 460, + 184 + ], + "blocks": [ + { + "bbox": [ + 149, + 115, + 460, + 184 + ], + "lines": [ + { + "bbox": [ + 149, + 115, + 460, + 184 + ], + "spans": [ + { + "bbox": [ + 149, + 115, + 460, + 184 + ], + "type": "image", + "image_path": "48c824298c93e3d130e70bc250a70665f6a84254b1899ea3a6d012afe34e0102.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 186, + 480, + 207 + ], + "lines": [ + { + "bbox": [ + 132, + 186, + 480, + 207 + ], + "spans": [ + { + "bbox": [ + 132, + 186, + 480, + 207 + ], + "type": "text", + "content": "Fig. 12: Real-World 3D Results. Notice the accurate shape and details in our reconstructions even from the diverse images in-the-wild." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 224, + 480, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 480, + 248 + ], + "type": "text", + "content": "minutes without SDS loss, and " + }, + { + "bbox": [ + 130, + 224, + 480, + 248 + ], + "type": "inline_equation", + "content": "\\approx 20" + }, + { + "bbox": [ + 130, + 224, + 480, + 248 + ], + "type": "text", + "content": " minutes with SDS loss. The coarse stage only takes " + }, + { + "bbox": [ + 130, + 224, + 480, + 248 + ], + "type": "inline_equation", + "content": "\\approx 2" + }, + { + "bbox": [ + 130, + 224, + 480, + 248 + ], + "type": "text", + "content": " minutes and provides a full 3D representation of the object." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 250, + 480, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 250, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 250, + 480, + 357 + ], + "type": "text", + "content": "Evaluation. We evaluate our 3D generation framework on 50 randomly sampled objects from the GSO dataset as described in Sec. 3.1. We compute image-based reconstruction metrics (LPIPS, PSNR, SSIM, MSE, and CLIP-S) between the ground truth (GT) GSO images, and rendered images from the trained 3D meshes on the same static/dynamic orbits. In addition, we compute 3D reconstruction metrics of Chamfer distance (CD) and 3D IoU between the GT and predicted meshes. We compare our SV3D-guided 3D generations with several prior methods including Point-E [34], Shap-E [16], One-2-3-45++ [24], DreamGaussian [50], SyncDreamer [27], EscherNet [21], Free3D [64], and Stable Zero123 [49]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 360, + 481, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 360, + 481, + 455 + ], + "spans": [ + { + "bbox": [ + 130, + 360, + 481, + 455 + ], + "type": "text", + "content": "Visual Comparison. In Fig. 11, we show visual comparison of our results with those from prior methods. Qualitatively, Point-E [34] and Shap-E [16] often produce incomplete 3D shapes. DreamGaussian [50], SyncDreamer [27], EscherNet [21], and Free3D [64] outputs tend to contain rough surfaces. One-2-3-45++ [24] and Stable Zero123 [49] are able to reconstruct meshes with smooth surface, but lack geometric details. Our mesh outputs are detailed, faithful to input image, and consistent in 3D (see appendix for more examples). We also show 3D mesh renders from real-world images in Fig. 12." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 457, + 481, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 457, + 481, + 541 + ], + "spans": [ + { + "bbox": [ + 130, + 457, + 481, + 541 + ], + "type": "text", + "content": "Quantitative Comparison. Tabs. 5 and 6 show the 2D and 3D metric comparisons respectively. All our 3D models achieve better 2D and 3D metrics compared to the prior and concurrent methods, showing the high-fidelity texture and geometry of our output meshes. We render all 3D meshes on the same dynamic orbits and compare them against the GT renders. Our best model, " + }, + { + "bbox": [ + 130, + 457, + 481, + 541 + ], + "type": "inline_equation", + "content": "\\mathrm{SV3D}^p" + }, + { + "bbox": [ + 130, + 457, + 481, + 541 + ], + "type": "text", + "content": ", performs comparably to using GT renders for reconstruction in terms of the 3D metrics, which further demonstrates the 3D consistency of our generated images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 544, + 481, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 481, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 481, + 628 + ], + "type": "text", + "content": "Effects of Photometric and (Masked) SDS Loss. As shown in Tabs. 5 and 6, the 3D outputs using both photometric and Masked SDS losses (" + }, + { + "bbox": [ + 130, + 544, + 481, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{SV3D}^p" + }, + { + "bbox": [ + 130, + 544, + 481, + 628 + ], + "type": "text", + "content": ") achieves the best results, while training without SDS (" + }, + { + "bbox": [ + 130, + 544, + 481, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{SV3D}^p" + }, + { + "bbox": [ + 130, + 544, + 481, + 628 + ], + "type": "text", + "content": " no SDS') leads to marginally lower performance. This demonstrates that the images generated by SV3D are high-quality reconstruction targets, and are often sufficient for 3D generation without the cumbersome SDS-based optimization. Nevertheless, adding SDS generally improves quality, as also shown in Fig. 9." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "Effects of SV3D Model and Training Orbit. As shown in Tabs. 5 and 6, " + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathrm{SV3D}^p" + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": " achieves the best performance among the three SV3D models, indicating that its synthesized novel views are most faithful to the input image and consis" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "type": "text", + "content": "SV3D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 171, + 345, + 275 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 342, + 170 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 342, + 170 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 342, + 170 + ], + "type": "text", + "content": "Table 5: 2D comparison of our 3D outputs against prior methods on the GSO dataset. Our best performing method uses " + }, + { + "bbox": [ + 132, + 114, + 342, + 170 + ], + "type": "inline_equation", + "content": "\\mathbf{SV3D}^p" + }, + { + "bbox": [ + 132, + 114, + 342, + 170 + ], + "type": "text", + "content": " with dynamic (sine elevation) orbit and SDS guidance. Note that all our models achieve better 2D metrics than prior works." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 171, + 345, + 275 + ], + "lines": [ + { + "bbox": [ + 132, + 171, + 345, + 275 + ], + "spans": [ + { + "bbox": [ + 132, + 171, + 345, + 275 + ], + "type": "table", + "html": "
ModelLPIPS↓PSNR↑SSIM↑MSE↓CLIP-S↑
GT renders0.07819.5080.8790.0140.897
EscherNet [21]0.17814.4380.8040.0410.835
Free3D [64]0.19714.2020.7990.0430.809
Stable Zero123 [49]0.16614.6350.8130.0400.805
SV3Du0.13315.9570.8340.0310.871
SV3Dc0.13216.3730.8340.0270.870
SV3Dpstatic orbit0.12516.8210.8480.0250.864
SV3Dpno SDS0.12416.8640.8410.0240.875
SV3Dp0.11917.4050.8490.0210.877
", + "image_path": "3a34e68cd530a417442722419b3d5961ec10c3054e8fc5216b2490d078b6cea5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 347, + 160, + 479, + 304 + ], + "blocks": [ + { + "bbox": [ + 347, + 114, + 478, + 159 + ], + "lines": [ + { + "bbox": [ + 347, + 114, + 478, + 159 + ], + "spans": [ + { + "bbox": [ + 347, + 114, + 478, + 159 + ], + "type": "text", + "content": "Table 6: Comparison of 3D metrics on the GSO dataset. Our models perform favorably against prior works." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 347, + 160, + 479, + 304 + ], + "lines": [ + { + "bbox": [ + 347, + 160, + 479, + 304 + ], + "spans": [ + { + "bbox": [ + 347, + 160, + 479, + 304 + ], + "type": "table", + "html": "
ModelCD↓3D IoU↑
GT renders0.0210.689
Point-E [34]0.0740.162
Shap-E [16]0.0710.267
DreamGaussian [50]0.0550.411
One-2-3-45++ [24]0.0540.406
SyncDreamer [27]0.0530.451
EscherNet [21]0.0420.466
Free3D [64]0.0470.426
Stable Zero123 [49]0.0390.550
SV3Du0.0270.589
SV3Dc0.0270.584
SV3Dp static orbit0.0280.610
SV3Dp no SDS0.0240.611
SV3Dp0.0240.614
", + "image_path": "40fb7a1ee3ca221ed980f517d91fb7c173407d73f26e8271bc4469cc7bf8ba14.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 308, + 481, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 481, + 380 + ], + "type": "text", + "content": "tent in 3D. On the other hand, " + }, + { + "bbox": [ + 130, + 308, + 481, + 380 + ], + "type": "inline_equation", + "content": "\\mathrm{SV3D}^u" + }, + { + "bbox": [ + 130, + 308, + 481, + 380 + ], + "type": "text", + "content": " shares the same disadvantage as several prior works in that it can only generate views at the same elevation, which is insufficient to build a legible 3D object, as shown in Fig. 8. This also leads to the worse performance of 'SV3D" + }, + { + "bbox": [ + 130, + 308, + 481, + 380 + ], + "type": "inline_equation", + "content": "^p" + }, + { + "bbox": [ + 130, + 308, + 481, + 380 + ], + "type": "text", + "content": " with static orbit' in Tabs. 5 and 6. Overall, " + }, + { + "bbox": [ + 130, + 308, + 481, + 380 + ], + "type": "inline_equation", + "content": "\\mathrm{SV3D}^p" + }, + { + "bbox": [ + 130, + 308, + 481, + 380 + ], + "type": "text", + "content": " with dynamic orbit and masked SDS loss performs favorably against all other configurations since it can leverage more diverse views of the object." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 381, + 482, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 381, + 482, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 381, + 482, + 489 + ], + "type": "text", + "content": "Limitations. Our SV3D model is by design only capable of handling 2 degrees of freedom: elevation and azimuth; which is usually sufficient for 3D generation from a single image. One may want to tackle more degrees of freedom in cameras for a generalized NVS system, which forms an interesting future work. We also notice that SV3D exhibits some view inconsistency for mirror-like reflective surfaces, which provide a challenge to 3D reconstruction. Lastly, such reflective surfaces are not representable by our Lambertian reflection-based shading model. Conditioning SV3D on the full camera matrix, and extending the shading model are interesting directions for future research." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 508, + 220, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 508, + 220, + 521 + ], + "spans": [ + { + "bbox": [ + 132, + 508, + 220, + 521 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "We present SV3D, a latent video diffusion model for novel multi-view synthesis and 3D generation. In addition to leveraging the generalizability and view-consistent prior in SVD, SV3D enables controllability via camera pose conditioning, and generates orbital videos of an object at high resolution on arbitrary camera orbits. We further propose several techniques for improved 3D generation from SV3D, including triangle CFG scaling, disentangled illumination, and masked SDS loss. We conduct extensive experiments to show that SV3D is controllable, multi-view consistent, and generalizable to the real-world, achieving state-of-the-art performance on novel multi-view synthesis and 3D generation. We believe SV3D provides a solid foundation model for further research on 3D object generation. We plan to publicly release SV3D models." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "text", + "content": "V. Voleti et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 115, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 115, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 482, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 482, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 482, + 201 + ], + "type": "text", + "content": "We thank Emad Mostaque, Bryce Wilson, Anel Islamovic, Savannah Martin, Ana Guillen, Josephine Parquet, Adam Chen and Ella Irwin for their help in various aspects of the model development and release. We thank Kyle Lacey and Yan Marek for their help with demos. We also thank Eric Courtemanche for his help with visual results." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 220, + 197, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 220, + 197, + 232 + ], + "spans": [ + { + "bbox": [ + 133, + 220, + 197, + 232 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 246, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 138, + 246, + 480, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 246, + 480, + 290 + ], + "spans": [ + { + "bbox": [ + 138, + 246, + 480, + 290 + ], + "type": "text", + "content": "1. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-NeRF: A multiscale representation for anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5855-5864 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 292, + 480, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 292, + 480, + 334 + ], + "spans": [ + { + "bbox": [ + 138, + 292, + 480, + 334 + ], + "type": "text", + "content": "2. Blattmann, A., Dockhorn, T., Kulal, S., Mendelevitch, D., Kilian, M., Lorenz, D., Levi, Y., English, Z., Voleti, V., Letts, A., et al.: Stable Video Diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 335, + 480, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 335, + 480, + 367 + ], + "spans": [ + { + "bbox": [ + 138, + 335, + 480, + 367 + ], + "type": "text", + "content": "3. Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S.W., Fidler, S., Kreis, K.: Align your Latents: High-Resolution Video Synthesis with Latent Diffusion Models. arXiv:2304.08818 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 368, + 480, + 400 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 368, + 480, + 400 + ], + "spans": [ + { + "bbox": [ + 138, + 368, + 480, + 400 + ], + "type": "text", + "content": "4. Boss, M., Braun, R., Jampani, V., Barron, J.T., Liu, C., Lensch, H.P.: NeRD: Neural reflectance decomposition from image collections. In: IEEE International Conference on Computer Vision (ICCV) (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 402, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 402, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 138, + 402, + 480, + 445 + ], + "type": "text", + "content": "5. Boss, M., Engelhardt, A., Kar, A., Li, Y., Sun, D., Barron, J.T., Lensch, H.P., Jampani, V.: SAMURAI: Shape And Material from Unconstrained Real-world Arbitrary Image collections. In: Advances in Neural Information Processing Systems (NeurIPS) (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 445, + 480, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 445, + 480, + 488 + ], + "spans": [ + { + "bbox": [ + 138, + 445, + 480, + 488 + ], + "type": "text", + "content": "6. Chan, E.R., Nagano, K., Chan, M.A., Bergman, A.W., Park, J.J., Levy, A., Aittala, M., Mello, S.D., Karras, T., Wetzstein, G.: GeNVS: Generative novel view synthesis with 3D-aware diffusion models. International Conference on Computer Vision (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 490, + 480, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 490, + 480, + 510 + ], + "spans": [ + { + "bbox": [ + 138, + 490, + 480, + 510 + ], + "type": "text", + "content": "7. Cline, D.B.: Admissible kernel estimators of a multivariate density. The Annals of Statistics 16(4), 1421-1427 (1988)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 512, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 512, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 512, + 480, + 544 + ], + "type": "text", + "content": "8. Deitke, M., Liu, R., Wallingford, M., Ngo, H., Michel, O., Kusupati, A., Fan, A., Laforte, C., Voleti, V., Gadre, S.Y., et al.: Objverse-XL: A universe of " + }, + { + "bbox": [ + 138, + 512, + 480, + 544 + ], + "type": "inline_equation", + "content": "10\\mathrm{m} + 3\\mathrm{D}" + }, + { + "bbox": [ + 138, + 512, + 480, + 544 + ], + "type": "text", + "content": " objects. arXiv preprint arXiv:2307.05663 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 544, + 480, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 544, + 480, + 587 + ], + "spans": [ + { + "bbox": [ + 138, + 544, + 480, + 587 + ], + "type": "text", + "content": "9. Deitke, M., Schwenk, D., Salvador, J., Weihs, L., Michel, O., VanderBilt, E., Schmidt, L., Ehsani, K., Kembhavi, A., Farhadi, A.: Objaverse: A universe of annotated 3D objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13142-13153 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 588, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 588, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 588, + 480, + 632 + ], + "type": "text", + "content": "10. Downs, L., Francis, A., Koenig, N., Kinman, B., Hickman, R., Reymann, K., McHugh, T.B., Vanhoucke, V.: Google Scanned Objects: A high-quality dataset of 3D scanned household items. In: 2022 International Conference on Robotics and Automation (ICRA). pp. 2553-2560. IEEE (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 632, + 480, + 665 + ], + "type": "text", + "content": "11. Eftekhar, A., Sax, A., Bachmann, R., Malik, J., Zamir, A.: Omnidata: A scalable pipeline for making multi-task mid-level vision datasets from 3D scans. In: IEEE International Conference on Computer Vision (ICCV) (2021)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 422, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 422, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 422, + 91, + 447, + 100 + ], + "type": "text", + "content": "SV3D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "text", + "content": "12. Girdhar, R., Singh, M., Brown, A., Duval, Q., Azadi, S., Rambhatla, S.S., Shah, A., Yin, X., Parikh, D., Misra, I.: EMU VIDEO: Factorizing Text-to-Video Generation by Explicit Image Conditioning (2023), https://emu-video.metademolab.com/ assets/emu(video.pdf" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 160, + 481, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 160, + 481, + 193 + ], + "spans": [ + { + "bbox": [ + 133, + 160, + 481, + 193 + ], + "type": "text", + "content": "13. Hasselgren, J., Hofmann, N., Munkberg, J.: Shape, Light, and Material Decomposition from Images using Monte Carlo Rendering and Denoising. In: Advances in Neural Information Processing Systems (NeurIPS) (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 193, + 481, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 193, + 481, + 214 + ], + "spans": [ + { + "bbox": [ + 133, + 193, + 481, + 214 + ], + "type": "text", + "content": "14. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: Advances in Neural Information Processing Systems (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 214, + 481, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 214, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 133, + 214, + 481, + 247 + ], + "type": "text", + "content": "15. Hong, Y., Zhang, K., Gu, J., Bi, S., Zhou, Y., Liu, D., Liu, F., Sunkavalli, K., Bui, T., Tan, H.: LRM: Large reconstruction model for single image to 3D. International Conference on Learning Representations (2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 247, + 481, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 247, + 481, + 257 + ], + "spans": [ + { + "bbox": [ + 133, + 247, + 481, + 257 + ], + "type": "text", + "content": "16. Jun, H., Nichol, A.: Shap-E: Generating conditional 3D implicit functions (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 257, + 481, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 257, + 481, + 289 + ], + "spans": [ + { + "bbox": [ + 133, + 257, + 481, + 289 + ], + "type": "text", + "content": "17. Kanopoulos, N., Vasanthavada, N., Baker, R.L.: Design of an image edge detection filter using the sobel operator. IEEE Journal of solid-state circuits 23(2), 358-367 (1988)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 289, + 481, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 289, + 481, + 332 + ], + "spans": [ + { + "bbox": [ + 133, + 289, + 481, + 332 + ], + "type": "text", + "content": "18. Kant, Y., Wu, Z., Vasilkovsky, M., Qian, G., Ren, J., Guler, R.A., Ghanem, B., Tulyakov, S., Gilitschenski, I., Siarohin, A.: Spad: Spatially aware multiview diffusers. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2024)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 332, + 481, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 332, + 481, + 353 + ], + "spans": [ + { + "bbox": [ + 133, + 332, + 481, + 353 + ], + "type": "text", + "content": "19. Karras, T., Aittala, M., Aila, T., Laine, S.: Elucidating the Design Space of Diffusion-Based Generative Models. arXiv:2206.00364 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 353, + 481, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 353, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 353, + 481, + 374 + ], + "type": "text", + "content": "20. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 374, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 481, + 396 + ], + "type": "text", + "content": "21. Kong, X., Liu, S., Lyu, X., Taher, M., Qi, X., Davison, A.J.: Eschernet: A generative model for scalable view synthesis. arXiv preprint arXiv:2402.03908 (2024)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 396, + 481, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 481, + 429 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 481, + 429 + ], + "type": "text", + "content": "22. Kwak, J.g., Dong, E., Jin, Y., Ko, H., Mahajan, S., Yi, K.M.: Vivid-1-to-3: Novel view synthesis with video diffusion models. IEEE conference on computer vision and pattern recognition (2024)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 429, + 481, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 429, + 481, + 471 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 481, + 471 + ], + "type": "text", + "content": "23. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3D: High-resolution text-to-3D content creation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 300–309 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 471, + 481, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 471, + 481, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 471, + 481, + 504 + ], + "type": "text", + "content": "24. Liu, M., Shi, R., Chen, L., Zhang, Z., Xu, C., Wei, X., Chen, H., Zeng, C., Gu, J., Su, H.: One-2-3-45++: Fast single image to 3D objects with consistent multi-view generation and 3D diffusion. arXiv preprint arXiv:2311.07885 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 504, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 481, + 536 + ], + "type": "text", + "content": "25. Liu, M., Xu, C., Jin, H., Chen, L., Varma T, M., Xu, Z., Su, H.: One-2-3-45: Any single image to 3D mesh in 45 seconds without per-shape optimization. Advances in Neural Information Processing Systems 36 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 536, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 481, + 567 + ], + "type": "text", + "content": "26. Liu, R., Wu, R., Hoorick, B.V., Tokmakov, P., Zakharov, S., Vondrick, C.: Zero-1-to-3: Zero-shot one image to 3D object. International Conference on Computer Vision (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 568, + 481, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 481, + 600 + ], + "type": "text", + "content": "27. Liu, Y., Lin, C., Zeng, Z., Long, X., Liu, L., Komura, T., Wang, W.: SyncDreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "type": "text", + "content": "28. Long, X., Guo, Y.C., Lin, C., Liu, Y., Dou, Z., Liu, L., Ma, Y., Zhang, S.H., Habermann, M., Theobalt, C., et al.: Wonder3D: Single image to 3D using cross-domain diffusion. arXiv preprint arXiv:2310.15008 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "type": "text", + "content": "29. Melas-Kyriazi, L., Laina, I., Rupprecht, C., Neverova, N., Vedaldi, A., Gafni, O., Kokkinos, F.: IM-3D: Iterative multiview diffusion and reconstruction for high-quality 3D generation. arXiv preprint arXiv:2402.08682 (2024)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 100 + ], + "type": "text", + "content": "V. Voleti et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "30. Mercier, A., Nakhli, R., Reddy, M., Yasarla, R., Cai, H., Porikli, F., Berger, G.: HexaGen3D: Stablediffusion is just one step away from fast and diverse Text-to-3D generation. arXiv preprint arXiv:2401.07727 (2024)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 482, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 482, + 193 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 482, + 193 + ], + "type": "text", + "content": "31. Metzer, G., Richardson, E., Patashnik, O., Giryes, R., Cohen-Or, D.: Latent-NeRF for shape-guided generation of 3D shapes and textures. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12663-12673 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 193, + 482, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 193, + 482, + 225 + ], + "spans": [ + { + "bbox": [ + 130, + 193, + 482, + 225 + ], + "type": "text", + "content": "32. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing scenes as neural radiance fields for view synthesis. In: European Conference on Computer Vision. pp. 405-421 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 225, + 482, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 225, + 482, + 246 + ], + "spans": [ + { + "bbox": [ + 130, + 225, + 482, + 246 + ], + "type": "text", + "content": "33. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph. (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 246, + 482, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 246, + 482, + 267 + ], + "spans": [ + { + "bbox": [ + 130, + 246, + 482, + 267 + ], + "type": "text", + "content": "34. Nichol, A., Jun, H., Dhariwal, P., Mishkin, P., Chen, M.: Point-E: A System for Generating 3D Point Clouds from Complex Prompts (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 267, + 482, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 267, + 482, + 310 + ], + "spans": [ + { + "bbox": [ + 130, + 267, + 482, + 310 + ], + "type": "text", + "content": "35. Niemeyer, M., Barron, J.T., Mildenhall, B., Sajjadi, M.S.M., Geiger, A., Radwan, N.: RegNeRF: Regularizing neural radiance fields for view synthesis from sparse inputs. In: Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR) (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 310, + 482, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 310, + 482, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 310, + 482, + 342 + ], + "type": "text", + "content": "36. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 342, + 482, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 363 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 363 + ], + "type": "text", + "content": "37. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3D using 2D diffusion. arXiv (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 363, + 482, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 363, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 363, + 482, + 407 + ], + "type": "text", + "content": "38. Qian, G., Mai, J., Hamdi, A., Ren, J., Siarohin, A., Li, B., Lee, H.Y., Skorokhodov, I., Wonka, P., Tulyakov, S., et al.: Magic123: One image to high-quality 3D object generation using both 2D and 3D diffusion priors. arXiv preprint arXiv:2306.17843 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 407, + 482, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 482, + 449 + ], + "type": "text", + "content": "39. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 449, + 482, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 482 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 482 + ], + "type": "text", + "content": "40. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 482, + 482, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 482, + 482, + 514 + ], + "spans": [ + { + "bbox": [ + 130, + 482, + 482, + 514 + ], + "type": "text", + "content": "41. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dream-Booth: Fine tuning text-to-image dissusion models for subject-driven generation. arXiv preprint arXiv:2208.12242 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 514, + 482, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 514, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 514, + 482, + 557 + ], + "type": "text", + "content": "42. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E., Kamyar Seyed Ghasemipour, S., Karagol Ayan, B., Mahdavi, S.S., Gontijo Lopes, R., Salimans, T., Ho, J., Fleet, D.J., Norouzi, M.: Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding. arXiv:2205.11487 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 557, + 482, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 601 + ], + "type": "text", + "content": "43. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortzman, M., et al.: LAION-5B: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 601, + 482, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 601, + 482, + 633 + ], + "spans": [ + { + "bbox": [ + 130, + 601, + 482, + 633 + ], + "type": "text", + "content": "44. Shen, T., Gao, J., Yin, K., Liu, M.Y., Fidler, S.: Deep Marching Tetrahedra: a hybrid representation for high-resolution 3D shape synthesis. In: Advances in Neural Information Processing Systems (NeurIPS) (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 633, + 482, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 633, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 633, + 482, + 666 + ], + "type": "text", + "content": "45. Shi, R., Chen, H., Zhang, Z., Liu, M., Xu, C., Wei, X., Chen, L., Zeng, C., Su, H.: Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110 (2023)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 422, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 422, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 422, + 91, + 447, + 100 + ], + "type": "text", + "content": "SV3D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "46. Shi, Y., Wang, P., Ye, J., Long, M., Li, K., Yang, X.: MVDream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 161 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 161 + ], + "type": "text", + "content": "47. Song, Y., Ermon, S.: Improved Techniques for Training Score-Based Generative Models. arXiv:2006.09011 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 163, + 480, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 163, + 480, + 195 + ], + "spans": [ + { + "bbox": [ + 133, + 163, + 480, + 195 + ], + "type": "text", + "content": "48. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 197, + 480, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 197, + 480, + 217 + ], + "spans": [ + { + "bbox": [ + 133, + 197, + 480, + 217 + ], + "type": "text", + "content": "49. StabilityAI: Stable Zero123 (2023), https://stability.ai/news/stable-zero123-3d-generation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 218, + 480, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 218, + 480, + 239 + ], + "spans": [ + { + "bbox": [ + 133, + 218, + 480, + 239 + ], + "type": "text", + "content": "50. Tang, J., Ren, J., Zhou, H., Liu, Z., Zeng, G.: DreamGaussian: Generative Gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 241, + 480, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 241, + 480, + 284 + ], + "spans": [ + { + "bbox": [ + 133, + 241, + 480, + 284 + ], + "type": "text", + "content": "51. Tang, S., Chen, J., Wang, D., Tang, C., Zhang, F., Fan, Y., Chandra, V., Furukawa, Y., Ranjan, R.: Mvdiffusion++: A dense high-resolution multi-view diffusion model for single or sparse-view 3d object reconstruction. arXiv preprint arXiv:2402.12712 (2024)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 285, + 480, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 480, + 318 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 480, + 318 + ], + "type": "text", + "content": "52. Tsai, Y.T., Shih, Z.C.: All-frequency precomputed radiance transfer using spherical radial basis functions and clustered tensor approximation. ACM Transactions on Graphics (ToG) (2006)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 319, + 480, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 319, + 480, + 351 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 480, + 351 + ], + "type": "text", + "content": "53. Voleti, V., Jolicoeur-Martineau, A., Pal, C.: MCVD: Masked conditional video diffusion for prediction, generation, and interpolation. In: (NeurIPS) Advances in Neural Information Processing Systems (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 353, + 480, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 353, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 353, + 480, + 373 + ], + "type": "text", + "content": "54. Watson, D., Chan, W., Martin-Brualla, R., Ho, J., Tagliasacchi, A., Norouzi, M.: Novel view synthesis with diffusion models (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 375, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 375, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 375, + 480, + 396 + ], + "type": "text", + "content": "55. Weng, H., Yang, T., Wang, J., Li, Y., Zhang, T., Chen, C.L.P., Zhang, L.: Consistent123: Improve consistency for one image to 3D object synthesis (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 398, + 480, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 398, + 480, + 441 + ], + "spans": [ + { + "bbox": [ + 133, + 398, + 480, + 441 + ], + "type": "text", + "content": "56. Wu, T., Zhang, J., Fu, X., Wang, Y., Ren, J., Pan, L., Wu, W., Yang, L., Wang, J., Qian, C., et al.: Omniobject3D: Large-vocabulary 3D object dataset for realistic perception, reconstruction and generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 803-814 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 442, + 480, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 442, + 480, + 463 + ], + "spans": [ + { + "bbox": [ + 132, + 442, + 480, + 463 + ], + "type": "text", + "content": "57. Yang, J., Cheng, Z., Duan, Y., Ji, P., Li, H.: ConsistNet: Enforcing 3D consistency for multi-view images diffusion. arXiv preprint arXiv:2310.10343 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 464, + 480, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 464, + 480, + 497 + ], + "spans": [ + { + "bbox": [ + 133, + 464, + 480, + 497 + ], + "type": "text", + "content": "58. Yao, C.H., Raj, A., Hung, W.C., Rubinstein, M., Li, Y., Yang, M.H., Jampani, V.: ARTIC3D: Learning robust articulated 3D shapes from noisy web image collections. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 498, + 480, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 480, + 519 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 480, + 519 + ], + "type": "text", + "content": "59. Ye, J., Wang, P., Li, K., Shi, Y., Wang, H.: Consistent-1-to-3: Consistent image to 3D view synthesis via geometry-aware diffusion models. In: 3DV (2024)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 521, + 394, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 521, + 394, + 531 + ], + "spans": [ + { + "bbox": [ + 132, + 521, + 394, + 531 + ], + "type": "text", + "content": "60. Young, J.: xatlas. https://github.com/jpcy/xatlas (2024)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 533, + 480, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 533, + 480, + 564 + ], + "spans": [ + { + "bbox": [ + 133, + 533, + 480, + 564 + ], + "type": "text", + "content": "61. Yu, Z., Peng, S., Niemeyer, M., Sattler, T., Geiger, A.: MonoSDF: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in Neural Information Processing Systems (NeurIPS) (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 566, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 566, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 133, + 566, + 480, + 609 + ], + "type": "text", + "content": "62. Zhang, K., Luan, F., Wang, Q., Bala, K., Snavely, N.: PhySG: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In: The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 610, + 480, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 480, + 642 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 480, + 642 + ], + "type": "text", + "content": "63. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "64. Zheng, C., Vedaldi, A.: Free3D: Consistent novel view synthesis without 3D representation. arXiv preprint arXiv:2312.04551 (2023)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 100 + ], + "type": "text", + "content": "V. Voleti et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": "In this supplemental document, we include the broader impact in Appendix A, implementation details in Appendices B to F, ablative analyses in Appendix G, and additional results in Appendix H. We also attach a teaser video summarizing the SV3D framework (SV3D(video.mp4), as well as an HTML page (SV3D.html) for more visual comparisons." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 195, + 253, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 195, + 253, + 211 + ], + "spans": [ + { + "bbox": [ + 132, + 195, + 253, + 211 + ], + "type": "text", + "content": "A Broader Impact" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 221, + 482, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 221, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 221, + 482, + 304 + ], + "type": "text", + "content": "The advancement of generative models in different media forms is changing how we make and use content. These AI-powered models can create images, videos, 3D objects, and more, in ways we've never seen before. They offer huge potential for innovation in media production. But along with this potential, there are also risks. Before we start using these models widely, it is crucial to make sure we understand the possible downsides and have plans in place to deal with them effectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 306, + 482, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 306, + 482, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 306, + 482, + 376 + ], + "type": "text", + "content": "In the case of 3D object generation, the input provided by the user plays a crucial role in constraining the model's output. By supplying a full front view of an object, users limit the model's creative freedom to the visible or unoccluded regions, thus minimizing the potential for generating problematic imagery. Additionally, factors such as predicted depth values and lighting further influence the fidelity and realism of generated content." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 378, + 491, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 378, + 491, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 378, + 491, + 472 + ], + "type": "text", + "content": "Moreover, ensuring the integrity and appropriateness of training data is critical in mitigating risks associated with generative models. Platforms like Sketchfab, which serve as repositories for 3D models used in training, enforce strict content policies to prevent the dissemination of \"Unacceptable Content\" and disallows it on their platform: https://help.sketchfab.com/hc/en-us/articles/214867883-What-is-Restricted-Content. By adhering to these guidelines and actively monitoring dataset quality, developers can reduce the likelihood of biased or inappropriate outputs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 474, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 533 + ], + "type": "text", + "content": "Sketchfab also has a tag for \"Restricted Content\" which is deemed to be similar to the PG-13 content rating used in the US (i.e. inappropriate for children under 13). We have confirmed that none of the objects that we use in training have this flag set to true. Thus we go the extra step of excluding even tagged PG-13 content from the training set." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 534, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 582 + ], + "type": "text", + "content": "There is a chance that certain content may not have been correctly labeled on Sketchfab. In cases where the uploader fails to tag an object appropriately, Sketchfab provides publicly accessible listings of objects and a mechanism for the community to report any content that may be deemed offensive." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": "In our regular utilization of Objaverse content, we haven't observed any significant amount of questionable material. Nevertheless, there are occasional instances of doll-like nudity stemming from basic 3D models, which could be crucial for accurately depicting humanoid anatomy. Additionally, the training dataset contains some presence of drugs, drug paraphernalia, as well as a certain level of blood content and weaponry, resembling what might be encountered in a video game context. Should the model be provided with imagery featuring these cate-" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "type": "text", + "content": "SV3D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "content": "gories of content, it possesses the capability to generate corresponding 3D models to some extent." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 143, + 482, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 143, + 482, + 251 + ], + "spans": [ + { + "bbox": [ + 130, + 143, + 482, + 251 + ], + "type": "text", + "content": "It is to be noted that SV3D mainly focuses on generating hidden details in the user's input image. If the image is unclear or some parts are hidden, the model guesses what those parts might look like based on its training data. This means it might create details similar to what it has seen before. The training data generally follows the standards of 3D modeling and gaming. However, this could lead to criticisms about the models being too similar to existing trends. But the user's input image limits how creative the model can be and reduces the chance of biases showing up in its creations, especially if the image is clear and straightforward." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 281, + 233, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 233, + 295 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 233, + 295 + ], + "type": "text", + "content": "B Data Details" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 319, + 479, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 479, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 479, + 380 + ], + "type": "text", + "content": "Similar to SVD-MV [2], we render views of a subset of the Objaverse dataset [9] consisting of 150K curated CC-licensed 3D objects from the original dataset. Each loaded object is scaled such that the largest world-space XYZ extent of its bounding box is 1. The object is then repositioned such that this bounding box is centered around the origin." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 382, + 480, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 382, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 382, + 480, + 430 + ], + "type": "text", + "content": "For both the static and dynamic orbits, we use Blender's EEVEE renderer to render an 84-frame RGBA orbit at " + }, + { + "bbox": [ + 130, + 382, + 480, + 430 + ], + "type": "inline_equation", + "content": "576 \\times 576" + }, + { + "bbox": [ + 130, + 382, + 480, + 430 + ], + "type": "text", + "content": " resolution. During training, any 21-frame orbit can be subsampled from this by picking any frame as the first frame, and then choosing every 4th frame after that." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 433, + 482, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 433, + 482, + 493 + ], + "spans": [ + { + "bbox": [ + 130, + 433, + 482, + 493 + ], + "type": "text", + "content": "We apply two background colors to each of these images: random RGB color, and white. This results in a doubling of the number of orbit samples for training. We then encode all of these images into latent space using SVD's VAE, and using CLIP. We then store the latent and CLIP embeddings for all of these images along with the corresponding elevation and azimuth values." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 495, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 482, + 555 + ], + "type": "text", + "content": "For lighting, we randomly select from a set of 20 curated HDRI envmaps. Each orbit begins with the camera positioned at azimuth 0. Our camera uses a field-of-view of 33.8 degrees. For each object, we adaptively position the camera to a distance sufficient to ensure that the rendered object content makes good and consistent use of the image extents without being clipped in any view." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": "For static orbits, the camera is positioned at a randomly sampled elevation between " + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "inline_equation", + "content": "[-5, 30]" + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": " degrees. The azimuth steps by a constant " + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\frac{360}{84}" + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": " degree delta between each frame. For dynamic orbits, the sequence of camera elevations for each orbit are obtained from a random weighted combination of sinusoids with different frequencies. For each sinusoid, the whole number period is sampled from [1, 5], the amplitude is sampled from [0.5, 10], and a random phase shift is applied. The azimuth angles are sampled regularly, and then a small amount of noise is added to make them irregular. The elevation values are smoothed using a simple convolution kernel and then clamped to a maximum elevation of 89 degrees." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "20" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 100 + ], + "type": "text", + "content": "V. Voleti et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 19 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 253, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 253, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 253, + 129 + ], + "type": "text", + "content": "C Training Details" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 482, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 482, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 482, + 307 + ], + "type": "text", + "content": "Our approach involves utilizing the widely used EDM [19] framework, incorporating a simplified diffusion loss for fine-tuning, as followed in SVD [2]. We eliminate the conditions of 'fps_id', 'motion_bucket_id', etc. since they are irrelevant for SV3D. Furthermore, we adjust the loss function to assign lower weights to frames closer to the front-view conditioning image, ensuring that challenging back views receive equal training focus as the easier front views. To optimize training efficiency and conserve GPU VRAM, we preprocess and store the precomputed latent and CLIP embeddings for all video frames in advance. During training, these tensors are directly loaded rather than being computed in real-time. We choose to finetune the SVD-xt model to output 21 frames instead of 25 frames. We found that with 21 frames we were able to fit a batch size of 2 on each GPU, instead of 1 with 25 frames at " + }, + { + "bbox": [ + 130, + 139, + 482, + 307 + ], + "type": "inline_equation", + "content": "576 \\times 576" + }, + { + "bbox": [ + 130, + 139, + 482, + 307 + ], + "type": "text", + "content": " resolution. All SV3D models are trained for 105k iterations with an effective batch size of 64 on 4 nodes of 880GB A100 GPUs for around 6 days." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 324, + 258, + 336 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 324, + 258, + 336 + ], + "spans": [ + { + "bbox": [ + 132, + 324, + 258, + 336 + ], + "type": "text", + "content": "D Inference Details" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 348, + 481, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 348, + 481, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 348, + 481, + 384 + ], + "type": "text", + "content": "To generate an orbital video during inference, we use 50 steps of the deterministic DDIM sampler [47] with the triangular classifier-free guidance (CFG) scale described in the main paper. This takes " + }, + { + "bbox": [ + 130, + 348, + 481, + 384 + ], + "type": "inline_equation", + "content": "\\approx 40" + }, + { + "bbox": [ + 130, + 348, + 481, + 384 + ], + "type": "text", + "content": " seconds for the SV3D model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 402, + 402, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 402, + 415 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 402, + 415 + ], + "type": "text", + "content": "E Additional Details on Illumination Model" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 426, + 482, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 499 + ], + "type": "text", + "content": "We base our rendering model on Spherical Gaussians (SG) [4,62]. A SG at query location " + }, + { + "bbox": [ + 130, + 426, + 482, + 499 + ], + "type": "inline_equation", + "content": "\\pmb{x} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 426, + 482, + 499 + ], + "type": "text", + "content": " is defined as " + }, + { + "bbox": [ + 130, + 426, + 482, + 499 + ], + "type": "inline_equation", + "content": "G(\\pmb{x};\\pmb{\\mu},c,a) = ae^{s(\\pmb{\\mu}\\cdot \\pmb{x} - 1)}" + }, + { + "bbox": [ + 130, + 426, + 482, + 499 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 426, + 482, + 499 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 426, + 482, + 499 + ], + "type": "text", + "content": " is the axis, " + }, + { + "bbox": [ + 130, + 426, + 482, + 499 + ], + "type": "inline_equation", + "content": "s \\in \\mathbb{R}" + }, + { + "bbox": [ + 130, + 426, + 482, + 499 + ], + "type": "text", + "content": " the sharpness of the lobe, and " + }, + { + "bbox": [ + 130, + 426, + 482, + 499 + ], + "type": "inline_equation", + "content": "a \\in \\mathbb{R}" + }, + { + "bbox": [ + 130, + 426, + 482, + 499 + ], + "type": "text", + "content": " the amplitude. Here, we point out that we only model white light and hence only use a scalar amplitude. One particularly interesting property of SGs is that the inner product of two SGs is the integral of the product of two SGs. The operation is defined as [52]:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 149, + 506, + 480, + 562 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 506, + 480, + 562 + ], + "spans": [ + { + "bbox": [ + 149, + 506, + 480, + 562 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} G _ {1} (\\boldsymbol {x}) \\cdot G _ {2} (\\boldsymbol {x}) = \\int_ {\\Omega} G _ {1} (\\boldsymbol {x}) G _ {2} (\\boldsymbol {x}) d \\boldsymbol {x} = \\frac {1}{d _ {m}} \\left(2 \\pi a _ {1} a _ {2} e ^ {d _ {m} - \\lambda_ {m}} (1. 0 - e ^ {- 2 d _ {m}})\\right) \\\\ \\lambda_ {m} = \\lambda_ {1} - \\lambda_ {2} \\tag {1} \\\\ d _ {m} = \\left| \\left| \\lambda_ {1} \\boldsymbol {\\mu} _ {1} + \\lambda_ {2} \\boldsymbol {\\mu} _ {2} \\right| \\right|. \\\\ \\end{array}", + "image_path": "096ef4601a42eb69a2841942df613fe37efe2184ae2f2e87092657512f8e829b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "In our illumination model we only consider Lambertian shading. Here, the cosine shading term influences the output the most. This term can be approximated with another SG " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "G_{\\mathrm{cosine}} = (\\pmb{x}; \\pmb{n}, 2.133, 1.17)" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{n}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " defines the surface normal at " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": ". The lighting evaluation using SGs " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "G_i" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " then becomes: " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "L = \\sum_{i=1}^{24} \\frac{1}{\\pi} \\max(G_i \\cdot G_{\\mathrm{cosine}}, 0)" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": ". As defined previously this results in the full integration of incoming light for each SG and as light is additive evaluating and summing all SGs results in the complete environment illumination. This " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " is also used in the " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{illum}}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " loss described in the main paper. The rendered textured" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "type": "text", + "content": "SV3D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "21" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 20 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 115, + 480, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 115, + 480, + 152 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 480, + 152 + ], + "type": "text", + "content": "image is then defined as " + }, + { + "bbox": [ + 132, + 115, + 480, + 152 + ], + "type": "inline_equation", + "content": "\\hat{I} = c_dL" + }, + { + "bbox": [ + 132, + 115, + 480, + 152 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 132, + 115, + 480, + 152 + ], + "type": "inline_equation", + "content": "c_{d}" + }, + { + "bbox": [ + 132, + 115, + 480, + 152 + ], + "type": "text", + "content": " is the diffuse albedo. We learn " + }, + { + "bbox": [ + 132, + 115, + 480, + 152 + ], + "type": "inline_equation", + "content": "\\mu, c, a" + }, + { + "bbox": [ + 132, + 115, + 480, + 152 + ], + "type": "text", + "content": " for each SG " + }, + { + "bbox": [ + 132, + 115, + 480, + 152 + ], + "type": "inline_equation", + "content": "G_{i}" + }, + { + "bbox": [ + 132, + 115, + 480, + 152 + ], + "type": "text", + "content": " using reconstruction loss between these rendered images and SV3D-generated images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 170, + 413, + 185 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 170, + 413, + 185 + ], + "spans": [ + { + "bbox": [ + 132, + 170, + 413, + 185 + ], + "type": "text", + "content": "F Losses and Optimization for 3D Generation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 196, + 482, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 196, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 196, + 482, + 475 + ], + "type": "text", + "content": "In addition to the masked SDS loss " + }, + { + "bbox": [ + 132, + 196, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{mask - ds}}" + }, + { + "bbox": [ + 132, + 196, + 482, + 475 + ], + "type": "text", + "content": " and illumination loss " + }, + { + "bbox": [ + 132, + 196, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{illum}}" + }, + { + "bbox": [ + 132, + 196, + 482, + 475 + ], + "type": "text", + "content": " detailed in the manuscript, we use several other losses for 3D reconstruction. Our main reconstruction losses are the pixel-level mean squared error " + }, + { + "bbox": [ + 132, + 196, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{mse}} = \\| I - \\hat{I}\\|^2" + }, + { + "bbox": [ + 132, + 196, + 482, + 475 + ], + "type": "text", + "content": ", LPIPS [63] loss " + }, + { + "bbox": [ + 132, + 196, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{lpips}}" + }, + { + "bbox": [ + 132, + 196, + 482, + 475 + ], + "type": "text", + "content": ", and mask loss " + }, + { + "bbox": [ + 132, + 196, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{mask}} = \\| S - \\hat{S}\\|^2" + }, + { + "bbox": [ + 132, + 196, + 482, + 475 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 132, + 196, + 482, + 475 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 132, + 196, + 482, + 475 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 196, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\hat{S}" + }, + { + "bbox": [ + 132, + 196, + 482, + 475 + ], + "type": "text", + "content": " are the predicted and ground-truth masks. We further employ a normal loss using the estimated mono normal by Omnidata [11], which is defined as the cosine similarity between the rendered normal " + }, + { + "bbox": [ + 132, + 196, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\pmb{n}" + }, + { + "bbox": [ + 132, + 196, + 482, + 475 + ], + "type": "text", + "content": " and estimated pseudo ground truths " + }, + { + "bbox": [ + 132, + 196, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\bar{n}" + }, + { + "bbox": [ + 132, + 196, + 482, + 475 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 132, + 196, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{normal}} = 1 - (\\pmb{n} \\cdot \\bar{\\pmb{n}})" + }, + { + "bbox": [ + 132, + 196, + 482, + 475 + ], + "type": "text", + "content": ". To regularize the output geometry, we apply a smooth depth loss inspired by RegNeRF [35]: " + }, + { + "bbox": [ + 132, + 196, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{depth}}(i,j) = (d(i,j) - d(i,j + 1))^2 + (d(i,j) - (d(i + 1,j))^2" + }, + { + "bbox": [ + 132, + 196, + 482, + 475 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 132, + 196, + 482, + 475 + ], + "type": "inline_equation", + "content": "i,j" + }, + { + "bbox": [ + 132, + 196, + 482, + 475 + ], + "type": "text", + "content": " indicate the pixel coordinate. For surface normal we instead rely on a bilateral smoothness loss similar to [5]. We found that this is crucial to getting high-frequency details and avoiding over-smoothed surfaces. For this loss we compute the image gradients of the input image " + }, + { + "bbox": [ + 132, + 196, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\nabla I" + }, + { + "bbox": [ + 132, + 196, + 482, + 475 + ], + "type": "text", + "content": " with a Sobel filter [17]. We then encourage the gradients of rendered normal " + }, + { + "bbox": [ + 132, + 196, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\nabla n" + }, + { + "bbox": [ + 132, + 196, + 482, + 475 + ], + "type": "text", + "content": " to be smooth if (and only if) the input image gradients " + }, + { + "bbox": [ + 132, + 196, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\nabla I" + }, + { + "bbox": [ + 132, + 196, + 482, + 475 + ], + "type": "text", + "content": " are smooth. The loss can be written as " + }, + { + "bbox": [ + 132, + 196, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{bilateral}} = e^{-3\\nabla I}\\sqrt{1 + ||\\nabla n||}" + }, + { + "bbox": [ + 132, + 196, + 482, + 475 + ], + "type": "text", + "content": ". We also found that adding a spatial smoothness regularization on the albedo is beneficial: " + }, + { + "bbox": [ + 132, + 196, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{albedo}} = (c_d(x) - c_d(x + \\epsilon))^2" + }, + { + "bbox": [ + 132, + 196, + 482, + 475 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 132, + 196, + 482, + 475 + ], + "type": "inline_equation", + "content": "c_d" + }, + { + "bbox": [ + 132, + 196, + 482, + 475 + ], + "type": "text", + "content": " denotes the albedo, " + }, + { + "bbox": [ + 132, + 196, + 482, + 475 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 132, + 196, + 482, + 475 + ], + "type": "text", + "content": " is a 3D surface point, and " + }, + { + "bbox": [ + 132, + 196, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\epsilon \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 132, + 196, + 482, + 475 + ], + "type": "text", + "content": " is a normal distributed offset with a scale of 0.01. The overall objective is then defined as the weighted sum of these losses. All losses are applied in both coarse and fine stages, except that we only apply " + }, + { + "bbox": [ + 132, + 196, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{mask - ds}}" + }, + { + "bbox": [ + 132, + 196, + 482, + 475 + ], + "type": "text", + "content": " in the last 200 iterations of the fine stage. We use an Adam optimizer [20] with a learning rate of 0.01 for both stages." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 491, + 332, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 332, + 506 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 332, + 506 + ], + "type": "text", + "content": "G Additional Ablative Analyses" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 517, + 480, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 517, + 480, + 540 + ], + "spans": [ + { + "bbox": [ + 132, + 517, + 480, + 540 + ], + "type": "text", + "content": "We conduct additional ablative analyses of our 3D generation pipeline in this section." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 559, + 234, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 559, + 234, + 571 + ], + "spans": [ + { + "bbox": [ + 132, + 559, + 234, + 571 + ], + "type": "text", + "content": "G.1 SV3D Models" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 578, + 480, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 480, + 626 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 480, + 626 + ], + "type": "text", + "content": "In Tab. 7, we compare the quantitative results using different SV3D models and training losses. Both 2D and 3D evaluation shows that " + }, + { + "bbox": [ + 132, + 578, + 480, + 626 + ], + "type": "inline_equation", + "content": "\\mathrm{SV3D}^p" + }, + { + "bbox": [ + 132, + 578, + 480, + 626 + ], + "type": "text", + "content": " is our best performing model, either for pure photometric reconstruction or SDS-based optimization." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "22" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 100 + ], + "type": "text", + "content": "V. Voleti et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 21 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 146, + 371, + 467, + 449 + ], + "blocks": [ + { + "bbox": [ + 131, + 327, + 482, + 360 + ], + "lines": [ + { + "bbox": [ + 131, + 327, + 482, + 360 + ], + "spans": [ + { + "bbox": [ + 131, + 327, + 482, + 360 + ], + "type": "text", + "content": "Table 7: Ablative results of different SV3D models and training losses. We show that our " + }, + { + "bbox": [ + 131, + 327, + 482, + 360 + ], + "type": "inline_equation", + "content": "\\mathrm{SV3D}^p" + }, + { + "bbox": [ + 131, + 327, + 482, + 360 + ], + "type": "text", + "content": " model with Photo+SDS losses achieves the best 2D and 3D metrics." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 146, + 371, + 467, + 449 + ], + "lines": [ + { + "bbox": [ + 146, + 371, + 467, + 449 + ], + "spans": [ + { + "bbox": [ + 146, + 371, + 467, + 449 + ], + "type": "table", + "html": "
ModelTraining lossesLPIPS↓PSNR↑SSIM↑MSE↓CLIP-S↑CD↓3D IoU↑
SV3DuPhoto0.13215.9510.8270.0320.8730.0280.583
SV3DuPhoto+SDS0.13315.9570.8340.0310.8710.0270.589
SV3DcPhoto0.13515.8260.8320.0330.8710.0290.579
SV3DcPhoto+SDS0.13216.3730.8340.0270.8700.0270.584
SV3DpPhoto0.12416.8640.8410.0240.8750.0240.611
SV3DpPhoto+SDS0.11917.4050.8490.0210.8770.0240.614
", + "image_path": "ad805a2478f4b5ea09d2b1624fb755c74d233a1b169793a259bff8ce45983232.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 422, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 422, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 422, + 91, + 447, + 100 + ], + "type": "text", + "content": "SV3D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "23" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 22 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 299, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 299, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 299, + 129 + ], + "type": "text", + "content": "G.2 Static v.s. Dynamic Orbits" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 136, + 482, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 136, + 482, + 221 + ], + "spans": [ + { + "bbox": [ + 130, + 136, + 482, + 221 + ], + "type": "text", + "content": "We also compare the results using different camera orbits for 3D training in Tab. 8. The results show that using a dynamic orbit (sine-30) produces better 3D outputs compared to static orbit since it contains more information of the top and bottom views of the object. However, higher elevation (sine-50) tends to increase inconsistency between multi-view images, and thus resulting in worse 3D reconstruction. In our experiments, we find that setting the elevation within " + }, + { + "bbox": [ + 130, + 136, + 482, + 221 + ], + "type": "inline_equation", + "content": "\\pm 30" + }, + { + "bbox": [ + 130, + 136, + 482, + 221 + ], + "type": "text", + "content": " degree generally leads to desirable 3D outputs." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 162, + 284, + 452, + 330 + ], + "blocks": [ + { + "bbox": [ + 130, + 240, + 482, + 275 + ], + "lines": [ + { + "bbox": [ + 130, + 240, + 482, + 275 + ], + "spans": [ + { + "bbox": [ + 130, + 240, + 482, + 275 + ], + "type": "text", + "content": "Table 8: Ablative results of different reference orbits for 3D generation. We show that using a dynamic orbit (sine elevation) with moderate amplitude performs better than orbits with no or extreme elevation variations." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 162, + 284, + 452, + 330 + ], + "lines": [ + { + "bbox": [ + 162, + 284, + 452, + 330 + ], + "spans": [ + { + "bbox": [ + 162, + 284, + 452, + 330 + ], + "type": "table", + "html": "
Training orbitLPIPS↓PSNR↑SSIM↑MSE↓CLIP-S↑CD↓3D IoU↑
Static0.12516.8210.8480.0250.8640.0280.610
Sine-300.11917.4050.8490.0210.8770.0240.614
Sine-500.12317.0570.8540.0250.8730.0260.609
", + "image_path": "8082049d049fcbad08c3ec24b92d6dd5d6a2e32e05575f93ac076f9d3ca32f64.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 371, + 254, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 371, + 254, + 383 + ], + "spans": [ + { + "bbox": [ + 132, + 371, + 254, + 383 + ], + "type": "text", + "content": "G.3 Masked SDS Loss" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 391, + 482, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 391, + 482, + 488 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 482, + 488 + ], + "type": "text", + "content": "Finally, we show the ablative results of our SDS loss in Tab. 9. We compare the results of 1) pure photometric losses, 2) with naive SDS loss (no masking), 3) with hard-masked SDS loss by thresholding the dot product of surface normal and camera viewing angle as visibility masks, and 4) with soft-masked SDS loss as described in the manuscript. Overall, adding SDS guidance from the SV3D model can improve the 2D metrics while maintaining similar 3D metrics. Our novel soft-masked SDS loss generally achieves the best results compared to other baselines." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 138, + 537, + 478, + 594 + ], + "blocks": [ + { + "bbox": [ + 130, + 505, + 482, + 529 + ], + "lines": [ + { + "bbox": [ + 130, + 505, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 505, + 482, + 529 + ], + "type": "text", + "content": "Table 9: Ablative analyses of Masked SDS loss. Overall, our soft-masked SDS loss leads to higher-quality mesh outputs in terms of most 2D and 3D metrics." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 537, + 478, + 594 + ], + "lines": [ + { + "bbox": [ + 138, + 537, + 478, + 594 + ], + "spans": [ + { + "bbox": [ + 138, + 537, + 478, + 594 + ], + "type": "table", + "html": "
Training lossesLPIPS↓PSNR↑SSIM↑MSE↓CLIP-S↑CD↓3D IoU↑
Photo0.12416.8640.8410.0240.8750.0240.611
Photo+SDS (naive)0.12417.0070.8500.0240.8670.0250.615
Photo+SDS (hard masked)0.12417.3350.8450.0220.8770.0240.610
Photo+SDS (soft masked)0.11917.4050.8490.0210.8770.0240.614
", + "image_path": "5b61eb72cf30106f67894316113b6ee762b478c32bb8fe21a5f235dbf770e850.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "24" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "text", + "content": "V. Voleti et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 23 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 309, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 309, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 309, + 127 + ], + "type": "text", + "content": "H Additional Visual Results" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 140, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 140, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 131, + 140, + 479, + 152 + ], + "type": "text", + "content": "In this section, we show more results of novel view synthesis and 3D generation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 170, + 272, + 182 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 170, + 272, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 170, + 272, + 182 + ], + "type": "text", + "content": "H.1 Novel View Synthesis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 190, + 480, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 190, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 131, + 190, + 480, + 226 + ], + "type": "text", + "content": "We show the additional NVS results on OmniObject3D [56] and real-world images in Fig. 13 and Fig. 14, respectively. The generated novel multi-view images by SV3D are more detailed and consistent compared to prior state-of-the-arts." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 138, + 246, + 478, + 485 + ], + "blocks": [ + { + "bbox": [ + 138, + 246, + 478, + 485 + ], + "lines": [ + { + "bbox": [ + 138, + 246, + 478, + 485 + ], + "spans": [ + { + "bbox": [ + 138, + 246, + 478, + 485 + ], + "type": "image", + "image_path": "105b9af9adeba66108cfd77e8c2dad8e2bfee731f010f0f8fba23fe2b687e7ca.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 496, + 482, + 540 + ], + "lines": [ + { + "bbox": [ + 131, + 496, + 482, + 540 + ], + "spans": [ + { + "bbox": [ + 131, + 496, + 482, + 540 + ], + "type": "text", + "content": "Fig. 13: NVS Results on OmniObject3D [56]. For each object, we show the input image and generated novel multi-view images by different methods. SV3D is able to produce images with consistent pose, color, and geometric details, which are closer the ground-truth renders." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 422, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 422, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 422, + 91, + 447, + 100 + ], + "type": "text", + "content": "SV3D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "25" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 24 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 270, + 482, + 473 + ], + "blocks": [ + { + "bbox": [ + 134, + 270, + 482, + 473 + ], + "lines": [ + { + "bbox": [ + 134, + 270, + 482, + 473 + ], + "spans": [ + { + "bbox": [ + 134, + 270, + 482, + 473 + ], + "type": "image", + "image_path": "f8dfcc05bba1ccf34bc9c1286b08bcab5f3ff6351a6d67286b56a7fd0387b3d3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 483, + 482, + 507 + ], + "lines": [ + { + "bbox": [ + 132, + 483, + 482, + 507 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 482, + 507 + ], + "type": "text", + "content": "Fig. 14: NVS Results on Real-World Images. Notice the consistent texture, geometry, and pose in SV3D NVS outputs compared to prior works." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "26" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 100 + ], + "type": "text", + "content": "V. Voleti et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 25 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 239, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 239, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 239, + 126 + ], + "type": "text", + "content": "H.2 3D Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 136, + 482, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 136, + 482, + 209 + ], + "spans": [ + { + "bbox": [ + 130, + 136, + 482, + 209 + ], + "type": "text", + "content": "We show the additional 3D generation results on OmniObject3D [56] and real-world images in Fig. 15 and Fig. 16, respectively. Thanks to the consistent multi-view images by SV3D and the novel Masked SDS loss, our 3D generations are detailed, high-fidelity, and generalizable to a wide range of images. Since Free3D [64] does not include a 3D generation method, we run our 3D pipeline on its generated multi-view images for fair 3D comparison." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 140, + 232, + 454, + 485 + ], + "blocks": [ + { + "bbox": [ + 140, + 232, + 454, + 485 + ], + "lines": [ + { + "bbox": [ + 140, + 232, + 454, + 485 + ], + "spans": [ + { + "bbox": [ + 140, + 232, + 454, + 485 + ], + "type": "image", + "image_path": "2636ffd15e777637be3807b2fddf85d618586e9e02154032657a0d1a01a83753.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 497, + 482, + 531 + ], + "lines": [ + { + "bbox": [ + 130, + 497, + 482, + 531 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 482, + 531 + ], + "type": "text", + "content": "Fig. 15: Mesh Results on OmniObject3D [56]. Thanks to the consistent novel multi-view images generated by SV3D, our 3D meshes contain higher-fidelity details while having smooth surfaces." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "type": "text", + "content": "SV3D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "27" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 26 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 141, + 250, + 175, + 287 + ], + "blocks": [ + { + "bbox": [ + 141, + 250, + 175, + 287 + ], + "lines": [ + { + "bbox": [ + 141, + 250, + 175, + 287 + ], + "spans": [ + { + "bbox": [ + 141, + 250, + 175, + 287 + ], + "type": "image", + "image_path": "bf21c4f7501d5718f0681146204e4d1c72137712b15b642c356d72f3ef9496ed.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 192, + 251, + 247, + 285 + ], + "blocks": [ + { + "bbox": [ + 192, + 251, + 247, + 285 + ], + "lines": [ + { + "bbox": [ + 192, + 251, + 247, + 285 + ], + "spans": [ + { + "bbox": [ + 192, + 251, + 247, + 285 + ], + "type": "image", + "image_path": "c1771e6d83010626d04298302ccea97de0b0d513ece8964c10b8356feea725d9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 192, + 287, + 246, + 321 + ], + "blocks": [ + { + "bbox": [ + 192, + 287, + 246, + 321 + ], + "lines": [ + { + "bbox": [ + 192, + 287, + 246, + 321 + ], + "spans": [ + { + "bbox": [ + 192, + 287, + 246, + 321 + ], + "type": "image", + "image_path": "03831f022ccc6d4c6f863a15122dae0102a2a28bbf2be6daff0cb9f38f20d187.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 257, + 251, + 313, + 286 + ], + "blocks": [ + { + "bbox": [ + 257, + 251, + 313, + 286 + ], + "lines": [ + { + "bbox": [ + 257, + 251, + 313, + 286 + ], + "spans": [ + { + "bbox": [ + 257, + 251, + 313, + 286 + ], + "type": "image", + "image_path": "af2dfd1a717bcf4ddf323c2bbf1bb4480d6e5871a129128e3ee2fcf03cf92edc.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 257, + 287, + 313, + 322 + ], + "blocks": [ + { + "bbox": [ + 257, + 287, + 313, + 322 + ], + "lines": [ + { + "bbox": [ + 257, + 287, + 313, + 322 + ], + "spans": [ + { + "bbox": [ + 257, + 287, + 313, + 322 + ], + "type": "image", + "image_path": "eeee73d320527f8110c04100e047f38eb14974c6758bee457f90ca4760636cb0.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 324, + 251, + 377, + 285 + ], + "blocks": [ + { + "bbox": [ + 324, + 251, + 377, + 285 + ], + "lines": [ + { + "bbox": [ + 324, + 251, + 377, + 285 + ], + "spans": [ + { + "bbox": [ + 324, + 251, + 377, + 285 + ], + "type": "image", + "image_path": "71e079185eeecd8f2d77f798c1b502cd2f3ef1dd54d119e1221a9a6828c0ac81.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 324, + 287, + 377, + 321 + ], + "blocks": [ + { + "bbox": [ + 324, + 287, + 377, + 321 + ], + "lines": [ + { + "bbox": [ + 324, + 287, + 377, + 321 + ], + "spans": [ + { + "bbox": [ + 324, + 287, + 377, + 321 + ], + "type": "image", + "image_path": "7fde30b1db64fec167c6874d912d2cd110f9ab3db4a2c032fb20f0ef8b8c5c31.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 394, + 251, + 452, + 286 + ], + "blocks": [ + { + "bbox": [ + 394, + 251, + 452, + 286 + ], + "lines": [ + { + "bbox": [ + 394, + 251, + 452, + 286 + ], + "spans": [ + { + "bbox": [ + 394, + 251, + 452, + 286 + ], + "type": "image", + "image_path": "0ea317b84ad7a62bb4ea75c19291e24f346adc9422bd4ec03ff6e39e1e28e431.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 395, + 288, + 451, + 322 + ], + "blocks": [ + { + "bbox": [ + 395, + 288, + 451, + 322 + ], + "lines": [ + { + "bbox": [ + 395, + 288, + 451, + 322 + ], + "spans": [ + { + "bbox": [ + 395, + 288, + 451, + 322 + ], + "type": "image", + "image_path": "8d5feb66007408ebe637e15fd7aa917354d6bb4d45f75e60debeb5b28fd80c3c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 139, + 324, + 178, + 361 + ], + "blocks": [ + { + "bbox": [ + 139, + 324, + 178, + 361 + ], + "lines": [ + { + "bbox": [ + 139, + 324, + 178, + 361 + ], + "spans": [ + { + "bbox": [ + 139, + 324, + 178, + 361 + ], + "type": "image", + "image_path": "fe6da9439b80e8940978a5b41f4348f88fe7962f82209cedc9c52e241736ac2f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 187, + 325, + 244, + 364 + ], + "blocks": [ + { + "bbox": [ + 187, + 325, + 244, + 364 + ], + "lines": [ + { + "bbox": [ + 187, + 325, + 244, + 364 + ], + "spans": [ + { + "bbox": [ + 187, + 325, + 244, + 364 + ], + "type": "image", + "image_path": "fd5959849467000e71de01bfeed96deeb9a41a7ec08aa22145be45c402ee75a4.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 188, + 365, + 243, + 402 + ], + "blocks": [ + { + "bbox": [ + 188, + 365, + 243, + 402 + ], + "lines": [ + { + "bbox": [ + 188, + 365, + 243, + 402 + ], + "spans": [ + { + "bbox": [ + 188, + 365, + 243, + 402 + ], + "type": "image", + "image_path": "e41cbdd433d6b5194e91c97c77ed3f7e4e292bf429f823e22f507366266aa936.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 253, + 326, + 313, + 361 + ], + "blocks": [ + { + "bbox": [ + 253, + 326, + 313, + 361 + ], + "lines": [ + { + "bbox": [ + 253, + 326, + 313, + 361 + ], + "spans": [ + { + "bbox": [ + 253, + 326, + 313, + 361 + ], + "type": "image", + "image_path": "410f32136b831e515a6dc9df47293aee849b7916cd829908725ef633fc4b477c.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 253, + 365, + 313, + 400 + ], + "blocks": [ + { + "bbox": [ + 253, + 365, + 313, + 400 + ], + "lines": [ + { + "bbox": [ + 253, + 365, + 313, + 400 + ], + "spans": [ + { + "bbox": [ + 253, + 365, + 313, + 400 + ], + "type": "image", + "image_path": "e13154017e66a59f398233c875015c1797cfdd3ac9edaf8ef53c12d453886c09.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 321, + 328, + 381, + 360 + ], + "blocks": [ + { + "bbox": [ + 321, + 328, + 381, + 360 + ], + "lines": [ + { + "bbox": [ + 321, + 328, + 381, + 360 + ], + "spans": [ + { + "bbox": [ + 321, + 328, + 381, + 360 + ], + "type": "image", + "image_path": "48d42cdabb13393fda6c62f2c7e40de307427c762285cb6bc96b5a35a5d3f4f4.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 321, + 365, + 381, + 400 + ], + "blocks": [ + { + "bbox": [ + 321, + 365, + 381, + 400 + ], + "lines": [ + { + "bbox": [ + 321, + 365, + 381, + 400 + ], + "spans": [ + { + "bbox": [ + 321, + 365, + 381, + 400 + ], + "type": "image", + "image_path": "3124b38b7ca24277fdb6e853e2e9f9311ec830f2a96da6e9368e860ab62beb47.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 391, + 327, + 451, + 362 + ], + "blocks": [ + { + "bbox": [ + 391, + 327, + 451, + 362 + ], + "lines": [ + { + "bbox": [ + 391, + 327, + 451, + 362 + ], + "spans": [ + { + "bbox": [ + 391, + 327, + 451, + 362 + ], + "type": "image", + "image_path": "588f6a934c862812ac95926fb295431cd952ff70624031689af9fccc30899b22.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 391, + 365, + 451, + 401 + ], + "blocks": [ + { + "bbox": [ + 391, + 365, + 451, + 401 + ], + "lines": [ + { + "bbox": [ + 391, + 365, + 451, + 401 + ], + "spans": [ + { + "bbox": [ + 391, + 365, + 451, + 401 + ], + "type": "image", + "image_path": "e67dc5694ded9302613190785b2c063fd1018a87fd668b03e1ab3cf43788c63d.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 148, + 407, + 168, + 445 + ], + "blocks": [ + { + "bbox": [ + 148, + 407, + 168, + 445 + ], + "lines": [ + { + "bbox": [ + 148, + 407, + 168, + 445 + ], + "spans": [ + { + "bbox": [ + 148, + 407, + 168, + 445 + ], + "type": "image", + "image_path": "3a0124166a356a96f01ea15ab24f3e9dc1744cc495ca07851aa034db2edcf47e.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 192, + 407, + 242, + 445 + ], + "blocks": [ + { + "bbox": [ + 192, + 407, + 242, + 445 + ], + "lines": [ + { + "bbox": [ + 192, + 407, + 242, + 445 + ], + "spans": [ + { + "bbox": [ + 192, + 407, + 242, + 445 + ], + "type": "image", + "image_path": "4e064a625dbad13dfb4f12e3a79fadd81b214212bb524c75cb8d591c38f0ad33.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 192, + 446, + 241, + 486 + ], + "blocks": [ + { + "bbox": [ + 192, + 446, + 241, + 486 + ], + "lines": [ + { + "bbox": [ + 192, + 446, + 241, + 486 + ], + "spans": [ + { + "bbox": [ + 192, + 446, + 241, + 486 + ], + "type": "image", + "image_path": "700de025b4638fdbff94603b36c34dba71bbeb360015ee50308e6cee7c14b912.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 205, + 487, + 227, + 496 + ], + "lines": [ + { + "bbox": [ + 205, + 487, + 227, + 496 + ], + "spans": [ + { + "bbox": [ + 205, + 487, + 227, + 496 + ], + "type": "text", + "content": "SV3D" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 259, + 406, + 308, + 445 + ], + "blocks": [ + { + "bbox": [ + 259, + 406, + 308, + 445 + ], + "lines": [ + { + "bbox": [ + 259, + 406, + 308, + 445 + ], + "spans": [ + { + "bbox": [ + 259, + 406, + 308, + 445 + ], + "type": "image", + "image_path": "529907a07ac547f4fa93a0765e19e3e36b6c915840f7ddd2df888b49af04ccfc.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 259, + 445, + 307, + 486 + ], + "blocks": [ + { + "bbox": [ + 259, + 445, + 307, + 486 + ], + "lines": [ + { + "bbox": [ + 259, + 445, + 307, + 486 + ], + "spans": [ + { + "bbox": [ + 259, + 445, + 307, + 486 + ], + "type": "image", + "image_path": "b798dde190ce6ac208841607c19bf331c0e7e1a9d23f30ad12ab2de367eebd82.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 265, + 487, + 302, + 496 + ], + "lines": [ + { + "bbox": [ + 265, + 487, + 302, + 496 + ], + "spans": [ + { + "bbox": [ + 265, + 487, + 302, + 496 + ], + "type": "text", + "content": "EscherNet" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 326, + 405, + 373, + 445 + ], + "blocks": [ + { + "bbox": [ + 326, + 405, + 373, + 445 + ], + "lines": [ + { + "bbox": [ + 326, + 405, + 373, + 445 + ], + "spans": [ + { + "bbox": [ + 326, + 405, + 373, + 445 + ], + "type": "image", + "image_path": "7763dd4e82f207b058b1f1912e8bcba795d6ed1018d8ce4ef552ee0709cde2df.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 326, + 445, + 373, + 486 + ], + "blocks": [ + { + "bbox": [ + 326, + 445, + 373, + 486 + ], + "lines": [ + { + "bbox": [ + 326, + 445, + 373, + 486 + ], + "spans": [ + { + "bbox": [ + 326, + 445, + 373, + 486 + ], + "type": "image", + "image_path": "1874c021568fd7b505d7b6603e370e17ad3353acf7eba1ced4fa9be69dd3d328.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 334, + 487, + 361, + 496 + ], + "lines": [ + { + "bbox": [ + 334, + 487, + 361, + 496 + ], + "spans": [ + { + "bbox": [ + 334, + 487, + 361, + 496 + ], + "type": "text", + "content": "Free3D" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 138, + 510, + 482, + 533 + ], + "lines": [ + { + "bbox": [ + 138, + 510, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 510, + 482, + 533 + ], + "type": "text", + "content": "Fig. 16: Mesh Results on Real-World Images. Notice the accurate shape and details in our reconstructions even from the diverse images in-the-wild." + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 389, + 407, + 445, + 445 + ], + "blocks": [ + { + "bbox": [ + 389, + 407, + 445, + 445 + ], + "lines": [ + { + "bbox": [ + 389, + 407, + 445, + 445 + ], + "spans": [ + { + "bbox": [ + 389, + 407, + 445, + 445 + ], + "type": "image", + "image_path": "c73ad4f0f244fbd8d0a7a18df3aa5ebb36d9da8859a1dba445eb4219135f88f9.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 389, + 447, + 446, + 485 + ], + "blocks": [ + { + "bbox": [ + 389, + 447, + 446, + 485 + ], + "lines": [ + { + "bbox": [ + 389, + 447, + 446, + 485 + ], + "spans": [ + { + "bbox": [ + 389, + 447, + 446, + 485 + ], + "type": "image", + "image_path": "fe0e95ac726c2866be59983efc71a4d1a978a945b82899035a4dbccd048fdfdd.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 395, + 487, + 450, + 495 + ], + "lines": [ + { + "bbox": [ + 395, + 487, + 450, + 495 + ], + "spans": [ + { + "bbox": [ + 395, + 487, + 450, + 495 + ], + "type": "text", + "content": "Stable Zero123" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "28" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 100 + ], + "type": "text", + "content": "V. Voleti et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 27 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 303, + 302, + 425 + ], + "blocks": [ + { + "bbox": [ + 134, + 303, + 302, + 425 + ], + "lines": [ + { + "bbox": [ + 134, + 303, + 302, + 425 + ], + "spans": [ + { + "bbox": [ + 134, + 303, + 302, + 425 + ], + "type": "image", + "image_path": "4f7cd0772d90275f1e9ee7af525094b9c69eeb707919381d4120c5d096e4f68a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 430, + 482, + 485 + ], + "lines": [ + { + "bbox": [ + 130, + 430, + 482, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 482, + 485 + ], + "type": "text", + "content": "Fig. 17: NVS and 3D Generation Results on Real-World Images. Visual results on real-world images from the MVImgNet (Yu et al. CVPR'23) and NAVI (Jampani et al. NeurIPS'22) datasets. Compared to Stable Zero123, SV3D generates more consistent novel multi-view images, thus resulting in higher-quality 3D meshes with smooth surface and detailed texture." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 309, + 302, + 480, + 426 + ], + "blocks": [ + { + "bbox": [ + 309, + 302, + 480, + 426 + ], + "lines": [ + { + "bbox": [ + 309, + 302, + 480, + 426 + ], + "spans": [ + { + "bbox": [ + 309, + 302, + 480, + 426 + ], + "type": "image", + "image_path": "f2e9a4372043cdf1fb202506c1a54251c142495e9254853a6c6c4900c24288a9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 422, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 422, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 422, + 91, + 447, + 100 + ], + "type": "text", + "content": "SV3D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "29" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 28 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/SWAG_ Splatting in the Wild images with Appearance-conditioned Gaussians/3925312b-25b5-4314-8bc0-d6cd9369f8b8_content_list.json b/2024/SWAG_ Splatting in the Wild images with Appearance-conditioned Gaussians/3925312b-25b5-4314-8bc0-d6cd9369f8b8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d00d8ded13d60fda75095f083b5fd8c6602c447a --- /dev/null +++ b/2024/SWAG_ Splatting in the Wild images with Appearance-conditioned Gaussians/3925312b-25b5-4314-8bc0-d6cd9369f8b8_content_list.json @@ -0,0 +1,1751 @@ +[ + { + "type": "text", + "text": "SWAG: Splatting in the Wild images with Appearance-conditioned Gaussians", + "text_level": 1, + "bbox": [ + 259, + 140, + 743, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hiba Dahmani, Moussab Bennehar, Nathan Piasco, Luis Roldao, and Dzmitry Tsishkou", + "bbox": [ + 228, + 212, + 774, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Noah's Ark, Huawei Paris Research Center, France", + "bbox": [ + 330, + 255, + 671, + 268 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/561769fdaa80ab4de74a0575feb1e464c05d67d05c40e5c53ae7a0beef61280a.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 232, + 300, + 343, + 479 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/de79f879277bcfdb8cd9632812f4456e9f57c8f7ea91365d085ae16f9091e7ff.jpg", + "image_caption": [ + "Fig. 1: Given in-the-wild captures (a), our model enables transient objects removal (b) and scene reconstruction with variable appearances (c)." + ], + "image_footnote": [], + "bbox": [ + 346, + 300, + 465, + 479 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/167f9d0404685e581f83fa9957132f5cd11e6ce086b9fa8b11249223669726c3.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 465, + 300, + 656, + 479 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4d676b85d17bd6ffd79a274970e119486582f75f74f7322b2b79afc80c5b8c6b.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 658, + 300, + 769, + 479 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Implicit neural representation methods have shown impressive advancements in learning 3D scenes from unstructured in-the-wild photo collections but are still limited by the large computational cost of volumetric rendering. Recently, 3D Gaussian Splatting emerged as a much faster alternative with superior rendering quality and training efficiency, especially for small-scale and object-centric scenarios. Nevertheless, this technique suffers from poor performance on unstructured in-the-wild data. To tackle this, we extend over 3D Gaussian Splatting to handle unstructured image collections. We achieve this by modeling appearance to seize photometric variations in the rendered images. Additionally, we introduce a new mechanism to train transient Gaussians to handle the presence of scene occluders in an unsupervised manner. Experiments on diverse photo collection scenes and multi-pass acquisition of outdoor landmarks show the effectiveness of our method over prior works achieving state-of-the-art results with improved efficiency.", + "bbox": [ + 259, + 578, + 743, + 787 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: 3D Gaussian Splatting $\\cdot$ Unconstrained Photo Collection $\\cdot$ Novel View Synthesis $\\cdot$ Appearance Modeling $\\cdot$ Real-time Rendering $\\cdot$ Transient Object Removal", + "bbox": [ + 259, + 797, + 743, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 143, + 374, + 160 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Novel View Synthesis (NVS) and 3D scene reconstruction are long-standing, challenging tasks in the realms of computer vision and computer graphics. Over the past few years, Neural Radiance Fields (NeRF) [15] have demonstrated groundbreaking results in rendering photorealistic views from novel viewpoints using an implicit neural representation of the radiance and density fields in a scene. Although NeRFs are very effective in static scenes, their performance significantly degrades in dynamic scenarios (i.e. containing moving transient objects) or in the presence of changing conditions such as weather, exposure, and lighting (i.e. in-the-wild datasets). To tackle these challenges, NeRF-W [14] proposed an extension to NeRF enabling the reconstruction of outdoor landmarks from in-the-wild images. They achieve this by learning the changing appearance of the images through per-image embeddings. Furthermore, to handle the presence of occluders, the scene is decomposed into \"static\" and \"transient\" elements being modeled by separate radiance fields. These modifications significantly improved NeRF's performance when confronted with appearance changes and transient occluders. More recent works [3,21] improve transient object handling by leveraging 2D visibility maps. Nonetheless, these methods still suffer from the inherent cost of volumetric rendering.", + "bbox": [ + 217, + 176, + 785, + 448 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, 3D Gaussian Splatting (3DGS) [8] attracted considerable interest thanks to its explicit representation resulting in faster training and rendering by leveraging GPU-based rasterization. 3DGS achieves fast training and matches visual quality obtained by the current SOTA NeRF methods [16]. This work represents the scene as a large number of anisotropic 3D Gaussians with color features and opacities. The Gaussians' set is then used in a differentiable splatting process to project the 3D Gaussians into a 2D image plane and blend them to get the rendered image. Similar to the original NeRF [15], 3DGS suffers from the same limitations regarding varying lighting conditions and the presence of occluders in the training images.", + "bbox": [ + 217, + 449, + 785, + 599 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we present SWAG, the first in-the-wild extension for 3DGS. To achieve this, we propose to expand the capabilities of 3DGS and improve its robustness in these scenarios. To this aim, we propose to capture the varying appearance of each image using learned embeddings and leveraging a multilayer perceptron (MLP), significantly improving our model's capabilities. Second, we learn an image-dependent opacity variation to each Gaussian that enables a better handling of transient objects and a finer scene reconstruction as illustrated in Figure 1.", + "bbox": [ + 217, + 599, + 785, + 719 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conduct a variety of experiments on the Phototourism dataset [6] and NeRF-OSR [18] benchmark and show that not only do we improve 3DGS performance in these scenarios, but we also achieve SOTA rendering quality with significantly faster training and rendering speeds compared to previous works [3, 7,14,21]. To summarize, our main contributions are as follows:", + "bbox": [ + 217, + 720, + 785, + 796 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "H. Dahmani et al.", + "bbox": [ + 271, + 114, + 392, + 127 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "- We introduce a new method for reconstructing scenes from unconstrained photo collections with varying appearances.", + "bbox": [ + 225, + 809, + 785, + 839 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We design a specific training scheme that captures occluders and enables transient object removal from a trained scene.", + "- We show that our method achieves SOTA performance in NVs in-the-wild scenarios.", + "- We demonstrate SWAG's abilities to generate new images with smooth visual transitions from learned appearances and to remove, in an unsupervised manner, transient objects from a captured scene." + ], + "bbox": [ + 223, + 146, + 784, + 248 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 273, + 385, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Neural Rendering in-the-wild", + "text_level": 1, + "bbox": [ + 215, + 303, + 504, + 318 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The abundance of in-the-wild unconstrained image data has encouraged adapting neural radiance fields to such cases. Early works like NeRF-W [14] have proposed a disentanglement between static and transient occladers using two appearance and transient per-image embeddings in addition to two radiance fields for each component of the scene (i.e. static and transient). Ha-NeRF [3] on the other hand uses a 2D image-dependent visibility map to eliminate occladers instead of using a decoupled radiance field since transient phenomena are only observed in individual 2D images. This simplification reduces the blurry artifacts that NeRF-W [14] suffered from in the attempt to reconstruct transient phenomena with a 3D transient field. Building upon previous methods, CR-NeRF [21] enhances their performance by exploiting the interaction information from multiple rays and fusing it into global information. Using a light-weight segmentation network, this method learns a visibility map without the supervision of ground truth segmentation masks to eliminate transient parts in 2D images. Another recent work, RefinedFields [7], leverages K-Planes and generative priors for in-the-wild scenarios. The learning is done through the alternation of two stages: a scene fitting to optimize the K-Planes [4] representation and a scene enriching phase that finetunes a pre-trained generative prior and infers a new K-Planes representation. As presented, Implicit-field representations have found diverse adaptations for in-the-wild scenarios. However, their time-consuming training and inference pose a challenge to achieving real-time rendering. This constrains their application in practical scenarios where fast rendering speed is crucial, especially in diverse interactive 3D applications.", + "bbox": [ + 212, + 327, + 787, + 675 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 Point-Based Rendering", + "text_level": 1, + "bbox": [ + 215, + 695, + 454, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Although NeRF demonstrated exceptional capabilities in generating photorealistic images, the demand for faster and more efficient rendering methods became increasingly evident. Point-based rendering has emerged as a prominent approach in computer graphics. It mainly revolves around the representation and rendering of the scene using discrete primitives. Starting from the most elementary form, Point Sample Rendering [5] authors proposed to sample a fixed-size formless set of points from a surface. Despite the developed strategies to adequately sample objects or surfaces, point sample rendering still suffers from", + "bbox": [ + 212, + 719, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "SWAG", + "bbox": [ + 681, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "holes and discontinuity. To overcome the issue of 3D points sparsity, the following works used different geometric primitives like discs [2] and surfels [23] that have a larger extent compared to pixels. Differentiable point-based rendering is the groundwork of these methods since it enables end-to-end training of 3D primitives from images. Pulsar [10] introduced an efficient sphere-based differentiable renderer where the scene is represented by a set of spheres with learned positions, feature vectors, opacities, and radii. These parameters are optimized to minimize the photometric reconstruction loss using an efficient differentiable rendering. The rasterization is done using either a convolutional U-Net or a per-pixel one-by-one convolutional network applied to a 2D screen space feature map. The feature map is a blending function that combines each sphere channel information using weighting based on position, radius, and opacity. Recently 3DGS [8] used 3D Gaussians as primitives and introduced a fully differentiable fast tile-based rasterizer for Gaussian splats allowing real-time rendering while outperforming state-of-the-art visual quality. Despite the demonstrated capability of 3DGS, there is still room for refining this method to enhance the quality of rendered images. One observed issue with 3DGS [8] is the appearance of strong artifacts when changing the sampling rate. Additionally, the rendering quality declines significantly in low resolutions or far away from camera positions due to the aliasing caused by the pixel size compared to the screen size and the Nyquist frequency.", + "bbox": [ + 212, + 146, + 787, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 3DGS Rendering Improvement", + "text_level": 1, + "bbox": [ + 215, + 513, + 517, + 529 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To solve the aforementioned rendering issue, a multi-scale adaptation of 3DGS [20] has been introduced. This work explains the aliasing phenomena by the splitting of large amounts of Gaussians with a smaller extent than pixels in 3D regions featuring high-frequency details. To render at a certain scale, authors filter out too large or too small Gaussians. Additionally, small Gaussians are aggregated to form larger ones selected on bigger scales during training to prevent missing areas and low-frequency details. Motivated by Mip-NeRF's [1] rendering quality, Mip-Splatting [22] replaces the 2D dilation filter introduced in Surface splatting [23] using 2D Mip filters that replicate box filters to solve the aliasing issue and uses 3D smoothing filters with a maximal sampling rate obtained from training images. One other issue with 3DGS is its performance deterioration in scenes that may present variation in appearance. VastGaussian [12] introduced a decoupled appearance modeling using a pixel-wise appearance embedding and a CNN to predict a pixel-wise transformation for the rendered images. Hence these predicted transformations (multiplication, addition, and Gamma correction) work well for large scenes but they are insufficient to model more varying appearances that appear in-the-wild scenarios. Our work aims to extend in-the-wild 3DGS scene representation beyond closed-world setups.", + "bbox": [ + 212, + 568, + 789, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "H. Dahmani et al.", + "bbox": [ + 271, + 114, + 392, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Background", + "text_level": 1, + "bbox": [ + 215, + 143, + 367, + 162 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The scene in 3DGS [8] is represented by a set of 3D anisotropic Gaussians. Each Gaussian is parameterized by its centroid $\\mathbf{x} \\in \\mathbb{R}^3$ , scale $\\mathbf{S} \\in \\mathbb{R}^3$ , rotation matrix $\\mathbf{R} \\in \\mathbb{R}^{3 \\times 3}$ , opacity $\\alpha \\in \\mathbb{R}$ and color $\\mathbf{c} \\in \\mathbb{R}^3$ encoded in spherical harmonic (SH) coefficients. The 3D covariance matrix $\\pmb{\\Sigma}$ of the 3D Gaussian is obtained using its rotation matrix $\\mathbf{R}$ and scale $\\mathbf{S}$ :", + "bbox": [ + 214, + 181, + 787, + 256 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\Sigma} = \\mathbf {R} \\mathbf {S} \\mathbf {S} ^ {T} \\mathbf {R} ^ {T}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 447, + 271, + 785, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The 3D Gaussians are defined in world space following:", + "bbox": [ + 215, + 301, + 614, + 316 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {G} (\\mathbf {y}) = e ^ {- \\frac {1}{2} (\\mathbf {y} - \\mathbf {x}) ^ {T} \\boldsymbol {\\Sigma} ^ {- 1} (\\mathbf {y} - \\mathbf {x})}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 401, + 330, + 785, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given $\\mathbf{J}$ , the Jacobian of the affine projective transformation and $\\mathbf{W}$ , a viewing transformation, the covariance matrix $\\pmb{\\Sigma}'$ in camera coordinates is as follows:", + "bbox": [ + 214, + 363, + 785, + 393 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\Sigma} ^ {\\prime} = \\mathbf {J} \\mathbf {W} \\boldsymbol {\\Sigma} \\mathbf {W} ^ {T} \\mathbf {J} ^ {T}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 433, + 407, + 784, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "These 3D Gaussians are projected to 2D splats and blended during a fast differentiable $\\alpha$ -blending process to get 2D rendered images. Each pixel $\\mathbf{y}$ color value $\\mathbf{C}$ is calculated using N-ordered 2D splats using the formula:", + "bbox": [ + 214, + 438, + 785, + 483 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {C} = \\sum_ {i \\in \\mathbf {N}} \\alpha_ {i} ^ {\\prime} \\mathbf {c} _ {i} \\prod_ {j = 1} ^ {i - 1} \\left(1 - \\alpha_ {j} ^ {\\prime}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 401, + 501, + 785, + 541 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\alpha_{i}^{\\prime}$ is the final opacity of the Gaussian obtained by:", + "bbox": [ + 215, + 550, + 584, + 566 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {i} ^ {\\prime} = \\alpha_ {i} * e ^ {- \\frac {1}{2} \\left(\\mathbf {y} ^ {\\prime} - \\mathbf {x} ^ {\\prime}\\right) ^ {T} \\boldsymbol {\\Sigma} ^ {\\prime - 1} \\left(\\mathbf {y} ^ {\\prime} - \\mathbf {x} ^ {\\prime}\\right)}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 382, + 582, + 784, + 601 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{x}'$ and $\\mathbf{y}'$ are coordinates in the projected space.", + "bbox": [ + 214, + 611, + 614, + 626 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although 3D Gaussians can be arbitrarily initialized, a good prior is often required for optimal results. Typically, the 3D Gaussians centers are initialized using the sparse Structure from Motion (SfM) point cloud obtained from the set of images. To avoid leaving huge holes in the scene, their corresponding covariances are initialized to have isotropic Gaussians with initial radii equal to the mean distance to neighboring points. During training, the 3D Gaussian features are optimized such that the photometric difference between the rendered image $\\mathbf{I}_{\\mathbf{r}}$ and the ground truth one $\\mathbf{I}_{\\mathbf{gt}}$ is minimized with intervening controlling steps. To control the Gaussians' density and to address under- and over-distribution issues, authors use an adaptive densification step where Gaussians with large spatial gradients are split while transparent Gaussians that have very low opacity are pruned.", + "bbox": [ + 212, + 627, + 785, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although 3DGS [8] works well on object-centric and small scenes, it struggles with scenes presenting varying appearances and transient objects.", + "bbox": [ + 214, + 809, + 785, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "SWAG", + "bbox": [ + 681, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/771b76bdcdd2edda9b39e8917edcd845805a7c4f5b2d6b8cf21f0354825106e8.jpg", + "image_caption": [ + "Fig. 2: SWAG model architecture - In addition to the typical Gaussians' features, we also optimize a Hash Grid encoding their centers $\\mathbf{emb}(\\mathbf{x})$ , a per-image embedding vector $l_{I}$ and an MLP. This MLP takes as inputs the Gaussians' colors $c$ , the associated image embedding $l_{I}$ , and their encoded centers $\\mathbf{emb}(\\mathbf{x})$ and outputs an image-dependent color $\\mathbf{c}^{\\mathbf{I}}$ as well as an image-dependent opacity variation parameter $\\Delta \\alpha^{I}$ . This parameter is set as the location variable of a concrete distribution which we sample to get the opacity variation $\\Delta \\tilde{\\alpha}^{\\mathbf{I}}$ . Leveraging this opacity variation across diverse training images enables identifying and excluding transient Gaussians within the scene, as demonstrated by the grey Gaussians." + ], + "image_footnote": [], + "bbox": [ + 271, + 143, + 733, + 276 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 Method", + "text_level": 1, + "bbox": [ + 215, + 440, + 328, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we introduce SWAG, a novel 3DGS-based method for 3D scene reconstruction from in-the-wild photo collections. We propose to adapt the 3D Gaussians parameters to handle variable visual appearances and the presence of occluders typically found in such unconstrained image collections.", + "bbox": [ + 212, + 472, + 787, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Section 4.1, we explain how SWAG handles appearance variations using image-dependent embeddings injected into the Gaussian's colors. In Section 4.2, we explain how our proposal handles transient occluders by learning image-dependent Gaussians' opacities variations. Since our model's knowledge about the presence of occluders is embedded in the Gaussians opacities variation, it easily enables a clear disentanglement between static and transient elements of the scene. The overall architecture of our method is illustrated in Figure 2.", + "bbox": [ + 212, + 534, + 787, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1 Appearance Variation Modeling", + "text_level": 1, + "bbox": [ + 215, + 662, + 526, + 678 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To adapt 3DGS to photometric variations, we associate, similarly to previous methods [14, 19], a trainable embedding vector $l_{I}$ for each image. To inject the appearance information, one naive solution would be to calculate a global transformation per image to the 3D Gaussian colors or SH features. This strategy, however, might be insufficient since it cannot model photometric features at the local level in an image. To model local appearance variances in an image, one possible way would be to predict a local transformation or directly a specific color for each Gaussian on a per-image basis. However, such solution is intractable as the number of training parameters will grow according to the number of Gaussians times the number of images.", + "bbox": [ + 212, + 689, + 787, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "H. Dahmani et al.", + "bbox": [ + 271, + 114, + 392, + 126 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In SWAG, we propose a more effective solution where we use an MLP that takes as input a concatenation of the image embeddings and a positional encoding of the Gaussians' centers. By using a positional encoding conditioned on the center of the Gaussians, we are able to model local appearance variations that occur in the image while maintaining a tractable number of parameters to optimize. Initially, we considered estimating an affine color transformation similar to Urban Radiance Fields [17] to be applied to each Gaussian color. However, our initial results showed that affine color transformations alone cannot model all appearance changes. To solve this problem, we directly predict the Gaussian color from the MLP. We noticed that spatial embeddings coupled with an MLP produce a smoothing effect on the predicted colors. Hence, to render view-dependent peculiarities and high-frequency details and counterbalance this smoothing effect, we also add the Gaussian color obtained from the SH coefficients as input to the MLP.", + "bbox": [ + 212, + 146, + 787, + 356 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our final solution employs an MLP $\\mathcal{F}_{\\theta}$ that, for a given view, takes as input the color of the 3D Gaussian $\\mathbf{c}$ , the embedding vector of its center $\\mathbf{emb}(\\mathbf{x})$ and the image embedding $l_{I}$ , and outputs the image-conditioned color $\\mathbf{c}^{\\mathbf{I}}$ . Accordingly, for a 3D Gaussian, the image-dependent color $\\mathbf{c}^{\\mathbf{I}}$ is computed as follows:", + "bbox": [ + 212, + 357, + 787, + 417 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {c} ^ {\\mathbf {I}} = \\mathcal {F} _ {\\theta} (\\mathbf {c}, \\mathbf {e m b} (\\mathbf {x}), l _ {I}). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 418, + 431, + 784, + 449 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2 Transient Gaussians Modeling", + "text_level": 1, + "bbox": [ + 215, + 468, + 511, + 484 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our initial results of applying 3DGS to in-the-wild scenarios showed that Gaussians representing transient objects produce blurred and elongated artifacts, due to their optimization across all images, despite transients appearing in only a few frames. Taking this observation into consideration, we introduce a learnable image-dependent opacity variation term $\\Delta \\tilde{\\alpha}^{\\mathbf{I}}$ to each Gaussian.", + "bbox": [ + 212, + 492, + 787, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "On the one hand, this opacity variation parameter allows Gaussians to reconstruct occluders present in some images. On the other hand, it enables those same Gaussians to be transparent (i.e. not involved in the image rendering process in other images) where these occluders are absent. During training, we want to encourage this variation parameter to either fully mask (i.e. in the case of occluders) or fully maintain (i.e. in the case of the static background) the 3D Gaussians' opacities. For this purpose, we sample $\\Delta \\tilde{\\alpha}^{\\mathbf{I}}$ using a Binary Concrete random variable [13]. This distribution is a continuous approximation of a Bernoulli distribution concentrating most of the mass on the boundaries of the interval [0, 1]. We return an additional output, $\\Delta \\alpha^{\\mathbf{I}}$ , of the previously introduced MLP $\\mathcal{F}_{\\theta}$ as the location parameter to the concrete function:", + "bbox": [ + 212, + 568, + 787, + 734 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\mathbf {c} ^ {\\mathbf {I}}, \\Delta \\alpha^ {\\mathbf {I}}\\right) = \\mathcal {F} _ {\\theta} (\\mathbf {c}, \\mathbf {e m b} (\\mathbf {x}), l _ {I}), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 743, + 784, + 760 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\tilde {\\alpha} ^ {\\mathbf {I}} = \\operatorname {s i g m o i d} \\left[ \\frac {1}{\\mathbf {T}} \\left(l o g \\left(| \\Delta \\alpha^ {\\mathbf {I}} |\\right) + l o g (\\mathbf {U}) - l o g (1 - \\mathbf {U})\\right) \\right], \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 295, + 780, + 784, + 813 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {U} \\sim \\operatorname {U n i f o r m} (0, 1), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 434, + 825, + 784, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "SWAG", + "bbox": [ + 681, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9a997299c4950565ad329fd41551e9cb2a73cca67526102f006e81af33f1c081.jpg", + "image_caption": [ + "Fig. 3: Variance histogram analysis of the Gaussians' opacity variation $\\Delta \\tilde{\\alpha}^{\\mathrm{I}}$ w.r.t training images for a dynamic (i.e. containing transient objects) scene (top, Trevi Fountain from Phototourism [6]); and a static scene (bottom, Locomotive from Tanks & Temples [9]). The left column views are rendered using only static Gaussians (i.e having $\\mathbf{Var}\\left[\\Delta \\tilde{\\alpha}^{\\mathrm{I}}\\right] = 0$ ), whereas right column views are rendered using all Gaussians." + ], + "image_footnote": [], + "bbox": [ + 272, + 143, + 733, + 284 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathbf{T}$ is the temperature hyper-parameter. We do not use any annealing on the temperature parameter, rather we set $\\mathbf{T} = 0.1$ to force, from the very beginning of the training, the opacity variation to match a binary distribution. During the evaluation, we fix $\\mathbf{U}$ at 0.5. The final image-dependent 3D Gaussian opacity is formulated as:", + "bbox": [ + 212, + 390, + 787, + 465 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha^ {\\mathbf {I}} = \\max \\left(\\alpha - \\Delta \\tilde {\\alpha} ^ {\\mathbf {I}}, 0\\right). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 411, + 470, + 785, + 491 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "With our formulation of the opacity, we can naturally disentangle between the transient and the static parts of the scene by defining transient Gaussians as the ones having non-zero variance of their associated parameter $\\Delta \\tilde{\\alpha}^{\\mathbf{I}}$ over training images: $\\operatorname{Var}\\left[\\Delta \\tilde{\\alpha}^{\\mathbf{I}}\\right] \\neq 0$ .", + "bbox": [ + 212, + 497, + 787, + 558 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 3 shows the histogram of $\\mathbf{Var}\\left[\\Delta \\tilde{\\alpha}^{\\mathbf{I}}\\right]$ on two scenes: one without transient content (Locomotive) and one with occluders (Trevi). We also show two rendered images of each scene: the right image using all the Gaussians and the left one using only Gaussians not detected as transient. The histogram shows that for a static scene, we detected less than $4\\%$ of the Gaussians as transient and around $96\\%$ of the Gaussians having $\\mathbf{Var}\\left[\\Delta \\tilde{\\alpha}^{\\mathbf{I}}\\right] = 0$ . Removing the $4\\%$ of transient Gaussians does not affect the rendered image quality. For Trevi Fountain, our model was able to represent transient objects using less than $20\\%$ of the Gaussians while maintaining the majority of the Gaussians to reconstruct static parts of the scene. Comparing the two rendered images shows that the transient Gaussians correspond to the mass of tourists present near the fountain, while removing the transient Gaussians resulted in an occluders-free rendering.", + "bbox": [ + 212, + 559, + 789, + 741 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 760, + 377, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.1 Implementation details", + "text_level": 1, + "bbox": [ + 215, + 787, + 452, + 803 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Similar to 3DGS, we minimize the $\\mathcal{L}_1$ loss combined with D-SSIM term to optimize the Gaussians parameters, the MLP $\\mathcal{F}_{\\theta}$ weights, the Hash Encoder pa", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "H. Dahmani et al.", + "bbox": [ + 271, + 114, + 393, + 127 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/60b401daffde20a204406f3424c6f2779b3cf5815acfe5f390debb720b822a84.jpg", + "table_caption": [ + "Table 1: Quantitative results - Results on three real-world scenes from Phototourism [6] and efficiency comparison among SoTA methods for NVS in-the-wild. Given the significant gap between previous baselines [3,7,14,21] and 3DGS [8], we marginalize on the different GPU configurations." + ], + "table_footnote": [ + "† Results reported on [7]. * Results computed using a high-tier GPU." + ], + "table_body": "
Bradenburg GateSacre CoeurTrevi FountainMean Efficiency
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓Train time (h)FPS
NeRF [15]18.900.8150.23115.600.7150.29116.140.6000.366--
NeRF-W [14]24.170.8900.16719.200.8070.19118.970.6980.265400†<1†
Ha-NeRF [3]24.040.8770.13920.020.8010.17120.180.6900.222452†0.20†
CR-NeRF [21]26.530.9000.10622.070.8230.15221.480.7110.206420†0.25†
RefinedFields [7]26.640.886-22.260.817-23.420.737-150†<1†
3DGS [8]19.990.8890.18017.570.8310.21918.470.7610.2340.50*181*
SWAG (ours)26.330.9290.13921.160.8600.18523.100.8150.2080.83*15.29*
", + "bbox": [ + 225, + 213, + 779, + 308 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "rameters, and the per-image embedding vectors. We fix the size of the latent embedding vectors to $n = 24$ . For the Hash encoding parameters, we set the hash table size $T_{h} = 2^{19}$ , the finest resolution $N_{max} = 2048$ , the number of levels $L = 12$ and the coarsest resolution $N_{min} = 16$ . The MLP we employ has 3 layers of 64 hidden units. Additional details on hyperparameters are provided in the supplementary material.", + "bbox": [ + 212, + 354, + 784, + 446 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.2 Datasets", + "text_level": 1, + "bbox": [ + 214, + 478, + 333, + 492 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Similar to SoTA in-the-wild reconstruction methods, we evaluate SWAG's NVS capabilities using the Phototourism dataset [6] and report our results on three main touristic monuments: Brandenburg Gate, Sacre Coeur, and Trevi Fountain. We also evaluate our method on the benchmark introduced in NeRF-OSR [18] for outdoor scene relighting tasks. We report our performance using NeRF-MS's [11] data split on 4 sites: stjohann (St. Johann, Saarbrücken), lump (Landwehrplatz, Saarbrücken), st (Staatstheater, Saarbrücken), and europa (Galerie Europa, Saarbrücken) and the original NeRF-OSR [18]'s data split on 3 sites: stjohann, lump and $lk2$ (Ludwigskirche, Saarbrücken). More details about the datasets can be found in supplementary materials.", + "bbox": [ + 212, + 513, + 787, + 667 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.3 Evaluation", + "text_level": 1, + "bbox": [ + 214, + 698, + 351, + 713 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We provide visual comparisons with rendered images and report quantitative results based on common rendering metrics from the literature: Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS). Since only the embeddings associated with the training images are optimized during training, we follow NeRF-W's [14] evaluation approach: we optimize an embedding on the left half of each test image and report metrics on the right half.", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "SWAG", + "bbox": [ + 681, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/17ad782ffa038f885928b18c2b1de29970e30922914d802f2cab8a4d20ae46f1.jpg", + "table_caption": [ + "Table 2: Quantitative results - Results on five sites from NeRF-OSR [18] outdoor scenes benchmark." + ], + "table_footnote": [ + "(a) Reported on NeRF-MS split [11]" + ], + "table_body": "
stjohannlwpsteuropa
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
NeRF [15]14.890.4320.63911.510.4680.57417.200.5140.50217.490.5510.503
NeRF-W [14]21.230.6670.42619.610.6160.44520.310.6070.43820.000.6990.340
Ha-NeRF [3]17.190.6860.33120.030.6850.36517.300.5380.48317.790.6320.421
NeRF-MS [11]22.840.7930.23521.900.7190.33620.680.6300.40221.030.7210.294
3DGS [8]16.770.7410.26811.760.6090.41417.160.6290.40620.180.7820.252
SWAG (ours)23.910.8640.17222.070.7830.30322.290.7130.36423.740.8450.242
", + "bbox": [ + 222, + 186, + 782, + 268 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/46bae78a1a89f96b9c8fc1d4bf4c367b0894f3e4ae7ac0199797329e981f95c5.jpg", + "table_caption": [], + "table_footnote": [ + "(b) Reported on original benchmark split from NeRF-OSR [18]" + ], + "table_body": "
lk2stlwp
PSNR ↑SSIM ↑PSNR ↑SSIM ↑PSNR ↑SSIM ↑
NeRF-OSR [18]19.860.62615.830.55617.380.576
3DGS [8]15.910.68712.530.58513.720.659
SWAG (ours)19.590.75618.730.72118.610.753
", + "bbox": [ + 341, + 282, + 663, + 343 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.4 Results", + "text_level": 1, + "bbox": [ + 215, + 385, + 323, + 398 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 1 shows quantitative results on three Phototourism scenes. Unsurprisingly, optimizing 3DGS on in-the-wild photo collections showed poor results due to the lack of image appearance dependency. The 3D Gaussians colors converge to a mean color which makes the method unable to model visual changes caused by weather, lightning conditions and different camera specifications. The presence of transient objects also resulted in the persistence of artifacts. SWAG improves 3DGS performance and outperforms the baselines on SSIM across all datasets. In particular, our proposal improves 3DGS's PSNR by an average margin of 5.01 dB. SWAG outperforms NeRF-W across all datasets in terms of LPIPS, PSNR, and SSIM and shows competitive performance with CR-NeRF and RefinedFields despite the priors these two methods use and requiring significantly longer times to train.", + "bbox": [ + 212, + 411, + 785, + 592 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 2 shows quantitative results on NeRF-OSR [18] benchmark. SWAG improves 3DGS performance by an average PSNR margin of 5 dB and outperformed all the other baselines regardless of the train/test split choices.", + "bbox": [ + 212, + 593, + 785, + 638 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Figure 4 and Figure 5 show qualitative results of 3DGS and SWAG on Phototourism scenes and four of NeRF-OSR sites. In some parts of the scene, 3DGS presents artifacts and floaters either in under-observed part of the scene or in areas occluded by transient objects. Adding to that, the rendered color is shifted from the ground truth color as shown in Figure 4. SWAG produces visual appearances close to the ground truth thanks to our per-image conditioning. Compared to previous implicit methods [3,7,14,21], our proposal achieves one order of magnitude faster training and guarantees real-time rendering as shown in Table 1.", + "bbox": [ + 212, + 638, + 785, + 758 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.5 Controllable appearance", + "text_level": 1, + "bbox": [ + 215, + 782, + 464, + 799 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Modeling the image appearance with an embedding vector enabled us to learn a latent space that we can utilize to change the lighting and appearance of any", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "H. Dahmani et al.", + "bbox": [ + 271, + 114, + 392, + 126 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/e82fd0dbe4b928068f725fc5a60468d6d52046dcd082afd9fea9e03be92b981c.jpg", + "image_caption": [ + "Fig. 4: Qualitative experimental results on three real-world scenes from Phototourism [6].s" + ], + "image_footnote": [], + "bbox": [ + 263, + 142, + 738, + 584 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "viewpoint at inference time. On top of the training image embeddings shown in Figure 6, we can sample vectors from the appearance space using interpolation between the learned embeddings. These interpolations are smooth and natural as shown in Figure 7. We encourage the readers to explore the supplementary videos and appreciate more the naturalness of these interpolations.", + "bbox": [ + 212, + 652, + 787, + 729 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.6 Transient Objects Removal", + "text_level": 1, + "bbox": [ + 215, + 753, + 485, + 768 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "As explained in Section 4.2, our method is able to localize 3D Gaussians that represent transient objects and, consequently, is able to manipulate the scenes omitting these objects. It should be noted that SWAG is thoughtfully designed to model transient objects without allowing the Gaussians to vary their opacity", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "SWAG", + "bbox": [ + 683, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/c90e427c1332e0d4ceb175188e42164b82b55b2c231a4d105ce3a43b117b5af6.jpg", + "image_caption": [ + "Fig. 5: Qualitative experimental results on three real-world scenes from Phototourism [6] and four NeRF-OSR [18] scenes. We demonstrate the capability of SWAG to disentangle between static and transient parts of the scene." + ], + "image_footnote": [], + "bbox": [ + 228, + 143, + 772, + 612 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "to compensate for appearance variation. Indeed, according to our experiments, it was observed that less than $20\\%$ (average on trained scenes) of the Gaussians capture transient objects. Consequently, SWAG achieves a better reconstruction of the static scene parts without having blurry elongated Gaussians that arise in view-dependent appearances with 3DGS. The capabilities of our model of disentangling static and dynamic Gaussians are clearly highlighted in Figure 5.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "H. Dahmani et al.", + "bbox": [ + 271, + 114, + 392, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/acbb1c6965f66d14cad980907fba8c253cf8bebe84a09c9fd18de3192492ef07.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 272, + 143, + 357, + 229 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/e0c574d11ffecd5a20135999fc0df1771463d1cbec21e169a9a867f76a649804.jpg", + "image_caption": [ + "Fig.6: Appearance transfer - SWAG is able to render a scene at any viewpoint with the visual appearance of any training image thanks to the learned appearance embeddings." + ], + "image_footnote": [], + "bbox": [ + 274, + 229, + 356, + 271 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/631bad8b5a26cc9900af96c60694257fb2116a722b070210293c8029562043b9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 357, + 143, + 477, + 271 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/9628c916e29ffac9cf74c6214c98113f225c6d6ed915f78723950da8e48ee47d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 478, + 143, + 728, + 271 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/1d7997630707f1bea131c6ebe20ecc8153100852c13d728fd4120c412d9de931.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 337, + 312, + 391 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/6ee3410d2010e59f2474a98013a597daa0d647378c3cfcb4c9b90d9525391646.jpg", + "image_caption": [ + "Fig. 7: Appearance interpolation - Interpolations between the appearance embeddings of two training images (left, right)." + ], + "image_footnote": [], + "bbox": [ + 217, + 392, + 312, + 445 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/2bd9880599c5f4f28b85add60b889955c0addadae12ecdeab86decdaead1cf2a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 316, + 337, + 411, + 391 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/2da72133034dd24738375d1d9f43669ae824ac3370b995d6800e99cef210b2d5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 316, + 392, + 411, + 445 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/99527152e06fc4f43eefac896bd0b6da6c56b2b23ce581f2156ecf33bfa32fcf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 413, + 337, + 504, + 391 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/ed22996c2f5e7ad9109f2472e7da4d4913347a73254fe97704d95c1b0fdcc188.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 413, + 392, + 504, + 445 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/2bdd6a400cd7870e89bcc5dc7b87f115df0b617d4832298b19e451abb304477f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 337, + 689, + 391 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/017dfaa40d47e50c5c982ecd9e0d46f36d262779126476a2b27a93e3488c1823.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 392, + 689, + 445 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/561fec92faf15f1c7387a9fe13b3c93dbfae955d50088c4e67c4b093ef410c46.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 338, + 787, + 391 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/20e059e50e3f88b63759b1cd225c410eb256b4d7f8a51c1c00114296dfc4b69e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 392, + 787, + 445 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.7 Ablations", + "text_level": 1, + "bbox": [ + 214, + 527, + 341, + 541 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In this section, we compare two variants of our method to analyze the contribution of each component of SWAG:", + "bbox": [ + 212, + 551, + 782, + 582 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- SWAG-A, a variation wherein the transient detection part of our model is removed,", + "- SWAG-T, a variation wherein the image appearance color dependency is omitted." + ], + "bbox": [ + 223, + 590, + 784, + 650 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We compare the results of our model variation on Phototourism scenes and we report the results in Table 3. SWAG-A enabled appearance modeling and", + "bbox": [ + 212, + 660, + 782, + 691 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/4486a64be06425f696e2236e67f86e5274b3b132ede0082dec9167fe3c26e8ed.jpg", + "table_caption": [ + "Table 3: Quantitative results - Results of two ablations of SWAG : SWAG-A and SWAG-T on three real-world scenes from Phototourism [6]." + ], + "table_footnote": [], + "table_body": "
Bradenburg GateSacre CoeurTrevi Fountain
PSNR ↑SSIM↑LPIPS ↓PSNR ↑SSIM↑LPIPS ↓PSNR ↑SSIM↑LPIPS ↓
SWAG26.330.9290.13921.160.8600.18523.100.8150.208
SWAG-A26.440.9280.14121.020.8550.18322.350.8010.223
SWAG-T24.460.9140.16019.060.8360.20520.130.7690.244
", + "bbox": [ + 258, + 758, + 738, + 837 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "SWAG", + "bbox": [ + 681, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/8fda2fcf75e21be26ef7293d5b450fb61dffb01ed073fb9598f8aa3b4c2eaec9.jpg", + "image_caption": [ + "Ground-truth" + ], + "image_footnote": [], + "bbox": [ + 277, + 143, + 364, + 194 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/5c9f1b066b14a3bbb46bac382bc7b56fb5cc5e5beea7d5b30d1439bd2c8775e9.jpg", + "image_caption": [ + "3DGS", + "Fig. 8: Ablation study - Visual comparison of 3DGS, SWAG, and the two variants used in our ablation study: SWAG-T (no appearance) and SWAG-A (no transient)." + ], + "image_footnote": [], + "bbox": [ + 367, + 143, + 454, + 194 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/d82228f05ff0e38d5bf9a80750aad4dec8b8b9790b77fc7ffe1199d04e88f2a5.jpg", + "image_caption": [ + "SWAG-A" + ], + "image_footnote": [], + "bbox": [ + 457, + 143, + 544, + 194 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/f4e3c9e218e9ce4916d2a4f665366da6c32ef0a67f1f8469b42bfb52f07f6f4e.jpg", + "image_caption": [ + "SWAG-T" + ], + "image_footnote": [], + "bbox": [ + 547, + 143, + 633, + 194 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/4584dc1a5a2997d25bec2b9aabcf1d3eee0754215b2370d64380e4344e86c65c.jpg", + "image_caption": [ + "SWAG" + ], + "image_footnote": [], + "bbox": [ + 637, + 143, + 723, + 194 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "closer colors to ground truth compared to 3DGS where colors shift away and converge to an average color for all training images. Transient objects' presence still causes floaters in the scene alike 3DGS as shown in Figure 8. In contrast, SWAG-T enables the elimination of these floaters and produces an occluders-free scene while suffering from color alterations compared to ground truth like 3DGS. Combining these two variations enables SWAG to model varying appearances with the ability to reduce scene floaters.", + "bbox": [ + 212, + 276, + 784, + 380 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5.8 Discussion", + "text_level": 1, + "bbox": [ + 214, + 402, + 346, + 416 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "SWAG leverages an MLP, a hash grid encoder, and image embedding to adapt 3DGS to in-the-wild scenarios. While this empowered 3DGS's reconstruction capabilities in these conditions as it was demonstrated in our experiments, it almost doubles the training time and has 10 times longer inference time per frame as shown Table 1 compared to 3DGS. Nonetheless, our method still requires significantly less training time than all previous in-the-wild baselines and achieves interactive rendering speed. Although SWAG succeeded in removing occluders using transient Gaussians, some parts of the scene where transient objects' presence is more frequent (eg. clouds in the sky and tourists on the roads) can suffer from being detected as transient. Eliminating all the transient Gaussians can leave some holes in these parts of the scene as seen in Figure 5 for the Trevi Fountain scene where some parts of the fountain contain black spots. One way to solve this is to remove Gaussians based on an adaptive threshold $\\lambda$ rather than considering as static only Gaussian with zero variance: $\\mathbf{Var}\\left[\\Delta \\tilde{\\alpha}^{\\mathbf{I}}\\right] < \\lambda$ , but it requires manual tuning of this threshold $\\lambda$ .", + "bbox": [ + 212, + 425, + 784, + 652 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 672, + 357, + 689 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we presented SWAG, a method designed for tailoring 3DGS representations to in-the-wild scenarios. SWAG incorporates appearance modeling in the Gaussians' colors and employs an adaptive opacity modulation to handle the presence of transient objects. Extensive experiments demonstrate that SWAG achieves state-of-the-art results on two challenging benchmarks while exhibiting training times orders of magnitude faster than in-the-wild NVS baselines while enabling real-time rendering. As a first step in conditioning 3DGS for in-the-wild scene representations, this work suggests potential future research direction, such as extending SWAG to dynamic scenes.", + "bbox": [ + 212, + 703, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "H. Dahmani et al.", + "bbox": [ + 271, + 114, + 392, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 323, + 160 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. ICCV (2021)", + "2. Botsch, M., Hornung, A., Zwicker, M., Kobbelt, L.: High-quality surface splatting on today's gpus. In: Proceedings Eurographics/IEEE VGTC Symposium Point-Based Graphics (2005)", + "3. Chen, X., Zhang, Q., Li, X., Chen, Y., Feng, Y., Wang, X., Wang, J.: Hallucinated neural radiance fields in the wild. In: CVPR. pp. 12943-12952 (2022)", + "4. Fridovich-Keil, S., Meanti, G., Warburg, F.R., Recht, B., Kanazawa, A.: K-planes: Explicit radiance fields in space, time, and appearance. In: CVPR (2023)", + "5. Grossman, J., Sc, H., Dally, W.: Point sample rendering (1999)", + "6. Jin, Y., Mishkin, D., Mishchuk, A., Matas, J., Fua, P., Yi, K.M., Trulls, E.: Image matching across wide baselines: From paper to practice. International Journal of Computer Vision p. 517-547 (2020)", + "7. Kassab, K., Schnepf, A., Franceschi, J.Y., Caraffa, L., Mary, J., Gouet-Brunet, V.: Refinedfields: Radiance fields refinement for unconstrained scenes (2024)", + "8. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (2023)", + "9. Knapitsch, A., Park, J., Zhou, Q.Y., Koltun, V.: Tanks and temples. ACM Transactions on Graphics pp. 1-13 (2017)", + "10. Lassner, C.: Fast differentiable raycasting for neural rendering using sphere-based representations. CoRR (2020)", + "1. Li, P., Wang, S., Yang, C., Liu, B., Qiu, W., Wang, H.: Nerf-ms: Neural radiance fields with multi-sequence. In: ICCV. pp. 18591-18600 (2023)", + "2. Lin, J., Li, Z., Tang, X., Liu, J., Liu, S., Liu, J., Lu, Y., Wu, X., Xu, S., Yan, Y., Yang, W.: Vastgaussian: Vast 3d gaussians for large scene reconstruction (2024)", + "3. Maddison, C.J., Mnih, A., Teh, Y.W.: The concrete distribution: A continuous relaxation of discrete random variables. In: International Conference on Learning Representations (2017)", + "4. Martin-Brualla, R., Radwan, N., Sajjadi, M.S.M., Barron, J.T., Dosovitskiy, A., Duckworth, D.: NeRF in the Wild: Neural Radiance Fields for Unconstrained Photo Collections. In: CVPR (2021)", + "5. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)", + "6. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. CoRR (2022)", + "7. Rematas, K., Liu, A., Srinivasan, P.P., Barron, J.T., Tagliasacchi, A., Funkhouser, T., Ferrari, V.: Urban radiance fields. CVPR (2022)", + "8. Rudnev, V., Elgharib, M., Smith, W., Liu, L., Golyanik, V., Theobalt, C.: Nerf for outdoor scene relighting. In: ECCV (2022)", + "9. Turki, H., Ramanan, D., Satyanarayanan, M.: Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. In: CVPR. pp. 12922–12931 (2022)", + "20. Yan, Z., Low, W.F., Chen, Y., Lee, G.H.: Multi-scale 3d gaussian splatting for anti-aliased rendering (2023)", + "21. Yang, Y., Zhang, S., Huang, Z., Zhang, Y., Tan, M.: Cross-ray neural radiance fields for novel-view synthesis from unconstrained image collections. In: ICCV (2023)" + ], + "bbox": [ + 225, + 178, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "SWAG", + "bbox": [ + 681, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "22. Yu, Z., Chen, A., Huang, B., Sattler, T., Geiger, A.: Mip-splatting: Alias-free 3d gaussian splatting (2023)", + "23. Zwicker, M., Pfister, H., van Baar, J., Gross, M.: Surface splattering. In: Proceedings of the 28th Annual Conference on Computer Graphics and Interactive Techniques. p. 371-378 (2001)" + ], + "bbox": [ + 215, + 146, + 784, + 217 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "H. Dahmani et al.", + "bbox": [ + 271, + 114, + 392, + 126 + ], + "page_idx": 15 + } +] \ No newline at end of file diff --git a/2024/SWAG_ Splatting in the Wild images with Appearance-conditioned Gaussians/3925312b-25b5-4314-8bc0-d6cd9369f8b8_model.json b/2024/SWAG_ Splatting in the Wild images with Appearance-conditioned Gaussians/3925312b-25b5-4314-8bc0-d6cd9369f8b8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4d7f8249cb204fdadab600cda1f5a2aff7275fb3 --- /dev/null +++ b/2024/SWAG_ Splatting in the Wild images with Appearance-conditioned Gaussians/3925312b-25b5-4314-8bc0-d6cd9369f8b8_model.json @@ -0,0 +1,2157 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.26, + 0.141, + 0.744, + 0.187 + ], + "angle": 0, + "content": "SWAG: Splatting in the Wild images with Appearance-conditioned Gaussians" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.213, + 0.776, + 0.245 + ], + "angle": 0, + "content": "Hiba Dahmani, Moussab Bennehar, Nathan Piasco, Luis Roldao, and Dzmitry Tsishkou" + }, + { + "type": "text", + "bbox": [ + 0.331, + 0.256, + 0.673, + 0.27 + ], + "angle": 0, + "content": "Noah's Ark, Huawei Paris Research Center, France" + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.301, + 0.344, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.274, + 0.482, + 0.291, + 0.493 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.347, + 0.301, + 0.466, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.482, + 0.531, + 0.493 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.466, + 0.301, + 0.657, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.66, + 0.301, + 0.771, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.709, + 0.482, + 0.726, + 0.493 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.503, + 0.787, + 0.533 + ], + "angle": 0, + "content": "Fig. 1: Given in-the-wild captures (a), our model enables transient objects removal (b) and scene reconstruction with variable appearances (c)." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.579, + 0.744, + 0.788 + ], + "angle": 0, + "content": "Abstract. Implicit neural representation methods have shown impressive advancements in learning 3D scenes from unstructured in-the-wild photo collections but are still limited by the large computational cost of volumetric rendering. Recently, 3D Gaussian Splatting emerged as a much faster alternative with superior rendering quality and training efficiency, especially for small-scale and object-centric scenarios. Nevertheless, this technique suffers from poor performance on unstructured in-the-wild data. To tackle this, we extend over 3D Gaussian Splatting to handle unstructured image collections. We achieve this by modeling appearance to seize photometric variations in the rendered images. Additionally, we introduce a new mechanism to train transient Gaussians to handle the presence of scene occluders in an unsupervised manner. Experiments on diverse photo collection scenes and multi-pass acquisition of outdoor landmarks show the effectiveness of our method over prior works achieving state-of-the-art results with improved efficiency." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.799, + 0.744, + 0.841 + ], + "angle": 0, + "content": "Keywords: 3D Gaussian Splatting \\(\\cdot\\) Unconstrained Photo Collection \\(\\cdot\\) Novel View Synthesis \\(\\cdot\\) Appearance Modeling \\(\\cdot\\) Real-time Rendering \\(\\cdot\\) Transient Object Removal" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.128 + ], + "angle": 0, + "content": "H. Dahmani et al." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.375, + 0.161 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.178, + 0.786, + 0.449 + ], + "angle": 0, + "content": "Novel View Synthesis (NVS) and 3D scene reconstruction are long-standing, challenging tasks in the realms of computer vision and computer graphics. Over the past few years, Neural Radiance Fields (NeRF) [15] have demonstrated groundbreaking results in rendering photorealistic views from novel viewpoints using an implicit neural representation of the radiance and density fields in a scene. Although NeRFs are very effective in static scenes, their performance significantly degrades in dynamic scenarios (i.e. containing moving transient objects) or in the presence of changing conditions such as weather, exposure, and lighting (i.e. in-the-wild datasets). To tackle these challenges, NeRF-W [14] proposed an extension to NeRF enabling the reconstruction of outdoor landmarks from in-the-wild images. They achieve this by learning the changing appearance of the images through per-image embeddings. Furthermore, to handle the presence of occluders, the scene is decomposed into \"static\" and \"transient\" elements being modeled by separate radiance fields. These modifications significantly improved NeRF's performance when confronted with appearance changes and transient occluders. More recent works [3,21] improve transient object handling by leveraging 2D visibility maps. Nonetheless, these methods still suffer from the inherent cost of volumetric rendering." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.45, + 0.787, + 0.6 + ], + "angle": 0, + "content": "Recently, 3D Gaussian Splatting (3DGS) [8] attracted considerable interest thanks to its explicit representation resulting in faster training and rendering by leveraging GPU-based rasterization. 3DGS achieves fast training and matches visual quality obtained by the current SOTA NeRF methods [16]. This work represents the scene as a large number of anisotropic 3D Gaussians with color features and opacities. The Gaussians' set is then used in a differentiable splatting process to project the 3D Gaussians into a 2D image plane and blend them to get the rendered image. Similar to the original NeRF [15], 3DGS suffers from the same limitations regarding varying lighting conditions and the presence of occluders in the training images." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.601, + 0.786, + 0.72 + ], + "angle": 0, + "content": "In this work, we present SWAG, the first in-the-wild extension for 3DGS. To achieve this, we propose to expand the capabilities of 3DGS and improve its robustness in these scenarios. To this aim, we propose to capture the varying appearance of each image using learned embeddings and leveraging a multilayer perceptron (MLP), significantly improving our model's capabilities. Second, we learn an image-dependent opacity variation to each Gaussian that enables a better handling of transient objects and a finer scene reconstruction as illustrated in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.722, + 0.786, + 0.797 + ], + "angle": 0, + "content": "We conduct a variety of experiments on the Phototourism dataset [6] and NeRF-OSR [18] benchmark and show that not only do we improve 3DGS performance in these scenarios, but we also achieve SOTA rendering quality with significantly faster training and rendering speeds compared to previous works [3, 7,14,21]. To summarize, our main contributions are as follows:" + }, + { + "type": "page_footnote", + "bbox": [ + 0.226, + 0.81, + 0.786, + 0.84 + ], + "angle": 0, + "content": "- We introduce a new method for reconstructing scenes from unconstrained photo collections with varying appearances." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.683, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SWAG" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.147, + 0.784, + 0.175 + ], + "angle": 0, + "content": "- We design a specific training scheme that captures occluders and enables transient object removal from a trained scene." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.203 + ], + "angle": 0, + "content": "- We show that our method achieves SOTA performance in NVs in-the-wild scenarios." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.207, + 0.784, + 0.249 + ], + "angle": 0, + "content": "- We demonstrate SWAG's abilities to generate new images with smooth visual transitions from learned appearances and to remove, in an unsupervised manner, transient objects from a captured scene." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.147, + 0.785, + 0.249 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.274, + 0.387, + 0.289 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.304, + 0.505, + 0.319 + ], + "angle": 0, + "content": "2.1 Neural Rendering in-the-wild" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.328, + 0.788, + 0.676 + ], + "angle": 0, + "content": "The abundance of in-the-wild unconstrained image data has encouraged adapting neural radiance fields to such cases. Early works like NeRF-W [14] have proposed a disentanglement between static and transient occladers using two appearance and transient per-image embeddings in addition to two radiance fields for each component of the scene (i.e. static and transient). Ha-NeRF [3] on the other hand uses a 2D image-dependent visibility map to eliminate occladers instead of using a decoupled radiance field since transient phenomena are only observed in individual 2D images. This simplification reduces the blurry artifacts that NeRF-W [14] suffered from in the attempt to reconstruct transient phenomena with a 3D transient field. Building upon previous methods, CR-NeRF [21] enhances their performance by exploiting the interaction information from multiple rays and fusing it into global information. Using a light-weight segmentation network, this method learns a visibility map without the supervision of ground truth segmentation masks to eliminate transient parts in 2D images. Another recent work, RefinedFields [7], leverages K-Planes and generative priors for in-the-wild scenarios. The learning is done through the alternation of two stages: a scene fitting to optimize the K-Planes [4] representation and a scene enriching phase that finetunes a pre-trained generative prior and infers a new K-Planes representation. As presented, Implicit-field representations have found diverse adaptations for in-the-wild scenarios. However, their time-consuming training and inference pose a challenge to achieving real-time rendering. This constrains their application in practical scenarios where fast rendering speed is crucial, especially in diverse interactive 3D applications." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.696, + 0.455, + 0.712 + ], + "angle": 0, + "content": "2.2 Point-Based Rendering" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Although NeRF demonstrated exceptional capabilities in generating photorealistic images, the demand for faster and more efficient rendering methods became increasingly evident. Point-based rendering has emerged as a prominent approach in computer graphics. It mainly revolves around the representation and rendering of the scene using discrete primitives. Starting from the most elementary form, Point Sample Rendering [5] authors proposed to sample a fixed-size formless set of points from a surface. Despite the developed strategies to adequately sample objects or surfaces, point sample rendering still suffers from" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.127 + ], + "angle": 0, + "content": "H. Dahmani et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.466 + ], + "angle": 0, + "content": "holes and discontinuity. To overcome the issue of 3D points sparsity, the following works used different geometric primitives like discs [2] and surfels [23] that have a larger extent compared to pixels. Differentiable point-based rendering is the groundwork of these methods since it enables end-to-end training of 3D primitives from images. Pulsar [10] introduced an efficient sphere-based differentiable renderer where the scene is represented by a set of spheres with learned positions, feature vectors, opacities, and radii. These parameters are optimized to minimize the photometric reconstruction loss using an efficient differentiable rendering. The rasterization is done using either a convolutional U-Net or a per-pixel one-by-one convolutional network applied to a 2D screen space feature map. The feature map is a blending function that combines each sphere channel information using weighting based on position, radius, and opacity. Recently 3DGS [8] used 3D Gaussians as primitives and introduced a fully differentiable fast tile-based rasterizer for Gaussian splats allowing real-time rendering while outperforming state-of-the-art visual quality. Despite the demonstrated capability of 3DGS, there is still room for refining this method to enhance the quality of rendered images. One observed issue with 3DGS [8] is the appearance of strong artifacts when changing the sampling rate. Additionally, the rendering quality declines significantly in low resolutions or far away from camera positions due to the aliasing caused by the pixel size compared to the screen size and the Nyquist frequency." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.515, + 0.519, + 0.53 + ], + "angle": 0, + "content": "2.3 3DGS Rendering Improvement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.79, + 0.84 + ], + "angle": 0, + "content": "To solve the aforementioned rendering issue, a multi-scale adaptation of 3DGS [20] has been introduced. This work explains the aliasing phenomena by the splitting of large amounts of Gaussians with a smaller extent than pixels in 3D regions featuring high-frequency details. To render at a certain scale, authors filter out too large or too small Gaussians. Additionally, small Gaussians are aggregated to form larger ones selected on bigger scales during training to prevent missing areas and low-frequency details. Motivated by Mip-NeRF's [1] rendering quality, Mip-Splatting [22] replaces the 2D dilation filter introduced in Surface splatting [23] using 2D Mip filters that replicate box filters to solve the aliasing issue and uses 3D smoothing filters with a maximal sampling rate obtained from training images. One other issue with 3DGS is its performance deterioration in scenes that may present variation in appearance. VastGaussian [12] introduced a decoupled appearance modeling using a pixel-wise appearance embedding and a CNN to predict a pixel-wise transformation for the rendered images. Hence these predicted transformations (multiplication, addition, and Gamma correction) work well for large scenes but they are insufficient to model more varying appearances that appear in-the-wild scenarios. Our work aims to extend in-the-wild 3DGS scene representation beyond closed-world setups." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.683, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SWAG" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.369, + 0.163 + ], + "angle": 0, + "content": "3 Background" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.183, + 0.788, + 0.257 + ], + "angle": 0, + "content": "The scene in 3DGS [8] is represented by a set of 3D anisotropic Gaussians. Each Gaussian is parameterized by its centroid \\(\\mathbf{x} \\in \\mathbb{R}^3\\), scale \\(\\mathbf{S} \\in \\mathbb{R}^3\\), rotation matrix \\(\\mathbf{R} \\in \\mathbb{R}^{3 \\times 3}\\), opacity \\(\\alpha \\in \\mathbb{R}\\) and color \\(\\mathbf{c} \\in \\mathbb{R}^3\\) encoded in spherical harmonic (SH) coefficients. The 3D covariance matrix \\(\\pmb{\\Sigma}\\) of the 3D Gaussian is obtained using its rotation matrix \\(\\mathbf{R}\\) and scale \\(\\mathbf{S}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.448, + 0.272, + 0.786, + 0.289 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\Sigma} = \\mathbf {R} \\mathbf {S} \\mathbf {S} ^ {T} \\mathbf {R} ^ {T}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.302, + 0.615, + 0.318 + ], + "angle": 0, + "content": "The 3D Gaussians are defined in world space following:" + }, + { + "type": "equation", + "bbox": [ + 0.403, + 0.332, + 0.786, + 0.35 + ], + "angle": 0, + "content": "\\[\n\\mathbf {G} (\\mathbf {y}) = e ^ {- \\frac {1}{2} (\\mathbf {y} - \\mathbf {x}) ^ {T} \\boldsymbol {\\Sigma} ^ {- 1} (\\mathbf {y} - \\mathbf {x})}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.364, + 0.786, + 0.394 + ], + "angle": 0, + "content": "Given \\(\\mathbf{J}\\), the Jacobian of the affine projective transformation and \\(\\mathbf{W}\\), a viewing transformation, the covariance matrix \\(\\pmb{\\Sigma}'\\) in camera coordinates is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.434, + 0.408, + 0.785, + 0.424 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\Sigma} ^ {\\prime} = \\mathbf {J} \\mathbf {W} \\boldsymbol {\\Sigma} \\mathbf {W} ^ {T} \\mathbf {J} ^ {T}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.439, + 0.786, + 0.484 + ], + "angle": 0, + "content": "These 3D Gaussians are projected to 2D splats and blended during a fast differentiable \\(\\alpha\\)-blending process to get 2D rendered images. Each pixel \\(\\mathbf{y}\\) color value \\(\\mathbf{C}\\) is calculated using N-ordered 2D splats using the formula:" + }, + { + "type": "equation", + "bbox": [ + 0.403, + 0.502, + 0.786, + 0.542 + ], + "angle": 0, + "content": "\\[\n\\mathbf {C} = \\sum_ {i \\in \\mathbf {N}} \\alpha_ {i} ^ {\\prime} \\mathbf {c} _ {i} \\prod_ {j = 1} ^ {i - 1} \\left(1 - \\alpha_ {j} ^ {\\prime}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.551, + 0.586, + 0.567 + ], + "angle": 0, + "content": "\\(\\alpha_{i}^{\\prime}\\) is the final opacity of the Gaussian obtained by:" + }, + { + "type": "equation", + "bbox": [ + 0.383, + 0.583, + 0.785, + 0.602 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {i} ^ {\\prime} = \\alpha_ {i} * e ^ {- \\frac {1}{2} \\left(\\mathbf {y} ^ {\\prime} - \\mathbf {x} ^ {\\prime}\\right) ^ {T} \\boldsymbol {\\Sigma} ^ {\\prime - 1} \\left(\\mathbf {y} ^ {\\prime} - \\mathbf {x} ^ {\\prime}\\right)}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.612, + 0.615, + 0.627 + ], + "angle": 0, + "content": "where \\(\\mathbf{x}'\\) and \\(\\mathbf{y}'\\) are coordinates in the projected space." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.787, + 0.809 + ], + "angle": 0, + "content": "Although 3D Gaussians can be arbitrarily initialized, a good prior is often required for optimal results. Typically, the 3D Gaussians centers are initialized using the sparse Structure from Motion (SfM) point cloud obtained from the set of images. To avoid leaving huge holes in the scene, their corresponding covariances are initialized to have isotropic Gaussians with initial radii equal to the mean distance to neighboring points. During training, the 3D Gaussian features are optimized such that the photometric difference between the rendered image \\(\\mathbf{I}_{\\mathbf{r}}\\) and the ground truth one \\(\\mathbf{I}_{\\mathbf{gt}}\\) is minimized with intervening controlling steps. To control the Gaussians' density and to address under- and over-distribution issues, authors use an adaptive densification step where Gaussians with large spatial gradients are split while transparent Gaussians that have very low opacity are pruned." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Although 3DGS [8] works well on object-centric and small scenes, it struggles with scenes presenting varying appearances and transient objects." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.127 + ], + "angle": 0, + "content": "H. Dahmani et al." + }, + { + "type": "image", + "bbox": [ + 0.272, + 0.144, + 0.734, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.287, + 0.788, + 0.412 + ], + "angle": 0, + "content": "Fig. 2: SWAG model architecture - In addition to the typical Gaussians' features, we also optimize a Hash Grid encoding their centers \\(\\mathbf{emb}(\\mathbf{x})\\), a per-image embedding vector \\(l_{I}\\) and an MLP. This MLP takes as inputs the Gaussians' colors \\(c\\), the associated image embedding \\(l_{I}\\), and their encoded centers \\(\\mathbf{emb}(\\mathbf{x})\\) and outputs an image-dependent color \\(\\mathbf{c}^{\\mathbf{I}}\\) as well as an image-dependent opacity variation parameter \\(\\Delta \\alpha^{I}\\). This parameter is set as the location variable of a concrete distribution which we sample to get the opacity variation \\(\\Delta \\tilde{\\alpha}^{\\mathbf{I}}\\). Leveraging this opacity variation across diverse training images enables identifying and excluding transient Gaussians within the scene, as demonstrated by the grey Gaussians." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.441, + 0.33, + 0.456 + ], + "angle": 0, + "content": "4 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.473, + 0.789, + 0.534 + ], + "angle": 0, + "content": "In this section, we introduce SWAG, a novel 3DGS-based method for 3D scene reconstruction from in-the-wild photo collections. We propose to adapt the 3D Gaussians parameters to handle variable visual appearances and the presence of occluders typically found in such unconstrained image collections." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.535, + 0.789, + 0.64 + ], + "angle": 0, + "content": "In Section 4.1, we explain how SWAG handles appearance variations using image-dependent embeddings injected into the Gaussian's colors. In Section 4.2, we explain how our proposal handles transient occluders by learning image-dependent Gaussians' opacities variations. Since our model's knowledge about the presence of occluders is embedded in the Gaussians opacities variation, it easily enables a clear disentanglement between static and transient elements of the scene. The overall architecture of our method is illustrated in Figure 2." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.663, + 0.527, + 0.679 + ], + "angle": 0, + "content": "4.1 Appearance Variation Modeling" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.789, + 0.84 + ], + "angle": 0, + "content": "To adapt 3DGS to photometric variations, we associate, similarly to previous methods [14, 19], a trainable embedding vector \\( l_{I} \\) for each image. To inject the appearance information, one naive solution would be to calculate a global transformation per image to the 3D Gaussian colors or SH features. This strategy, however, might be insufficient since it cannot model photometric features at the local level in an image. To model local appearance variances in an image, one possible way would be to predict a local transformation or directly a specific color for each Gaussian on a per-image basis. However, such solution is intractable as the number of training parameters will grow according to the number of Gaussians times the number of images." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.683, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SWAG" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.357 + ], + "angle": 0, + "content": "In SWAG, we propose a more effective solution where we use an MLP that takes as input a concatenation of the image embeddings and a positional encoding of the Gaussians' centers. By using a positional encoding conditioned on the center of the Gaussians, we are able to model local appearance variations that occur in the image while maintaining a tractable number of parameters to optimize. Initially, we considered estimating an affine color transformation similar to Urban Radiance Fields [17] to be applied to each Gaussian color. However, our initial results showed that affine color transformations alone cannot model all appearance changes. To solve this problem, we directly predict the Gaussian color from the MLP. We noticed that spatial embeddings coupled with an MLP produce a smoothing effect on the predicted colors. Hence, to render view-dependent peculiarities and high-frequency details and counterbalance this smoothing effect, we also add the Gaussian color obtained from the SH coefficients as input to the MLP." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.358, + 0.788, + 0.419 + ], + "angle": 0, + "content": "Our final solution employs an MLP \\(\\mathcal{F}_{\\theta}\\) that, for a given view, takes as input the color of the 3D Gaussian \\(\\mathbf{c}\\), the embedding vector of its center \\(\\mathbf{emb}(\\mathbf{x})\\) and the image embedding \\(l_{I}\\), and outputs the image-conditioned color \\(\\mathbf{c}^{\\mathbf{I}}\\). Accordingly, for a 3D Gaussian, the image-dependent color \\(\\mathbf{c}^{\\mathbf{I}}\\) is computed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.419, + 0.433, + 0.785, + 0.45 + ], + "angle": 0, + "content": "\\[\n\\mathbf {c} ^ {\\mathbf {I}} = \\mathcal {F} _ {\\theta} (\\mathbf {c}, \\mathbf {e m b} (\\mathbf {x}), l _ {I}). \\tag {6}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.469, + 0.513, + 0.485 + ], + "angle": 0, + "content": "4.2 Transient Gaussians Modeling" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.788, + 0.568 + ], + "angle": 0, + "content": "Our initial results of applying 3DGS to in-the-wild scenarios showed that Gaussians representing transient objects produce blurred and elongated artifacts, due to their optimization across all images, despite transients appearing in only a few frames. Taking this observation into consideration, we introduce a learnable image-dependent opacity variation term \\(\\Delta \\tilde{\\alpha}^{\\mathbf{I}}\\) to each Gaussian." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.788, + 0.735 + ], + "angle": 0, + "content": "On the one hand, this opacity variation parameter allows Gaussians to reconstruct occluders present in some images. On the other hand, it enables those same Gaussians to be transparent (i.e. not involved in the image rendering process in other images) where these occluders are absent. During training, we want to encourage this variation parameter to either fully mask (i.e. in the case of occluders) or fully maintain (i.e. in the case of the static background) the 3D Gaussians' opacities. For this purpose, we sample \\(\\Delta \\tilde{\\alpha}^{\\mathbf{I}}\\) using a Binary Concrete random variable [13]. This distribution is a continuous approximation of a Bernoulli distribution concentrating most of the mass on the boundaries of the interval [0, 1]. We return an additional output, \\(\\Delta \\alpha^{\\mathbf{I}}\\), of the previously introduced MLP \\(\\mathcal{F}_{\\theta}\\) as the location parameter to the concrete function:" + }, + { + "type": "equation", + "bbox": [ + 0.397, + 0.744, + 0.785, + 0.761 + ], + "angle": 0, + "content": "\\[\n\\left(\\mathbf {c} ^ {\\mathbf {I}}, \\Delta \\alpha^ {\\mathbf {I}}\\right) = \\mathcal {F} _ {\\theta} (\\mathbf {c}, \\mathbf {e m b} (\\mathbf {x}), l _ {I}), \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.297, + 0.781, + 0.785, + 0.814 + ], + "angle": 0, + "content": "\\[\n\\Delta \\tilde {\\alpha} ^ {\\mathbf {I}} = \\operatorname {s i g m o i d} \\left[ \\frac {1}{\\mathbf {T}} \\left(l o g \\left(| \\Delta \\alpha^ {\\mathbf {I}} |\\right) + l o g (\\mathbf {U}) - l o g (1 - \\mathbf {U})\\right) \\right], \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.435, + 0.826, + 0.785, + 0.841 + ], + "angle": 0, + "content": "\\[\n\\mathbf {U} \\sim \\operatorname {U n i f o r m} (0, 1), \\tag {9}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.395, + 0.128 + ], + "angle": 0, + "content": "H. Dahmani et al." + }, + { + "type": "image", + "bbox": [ + 0.273, + 0.144, + 0.734, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.295, + 0.789, + 0.366 + ], + "angle": 0, + "content": "Fig. 3: Variance histogram analysis of the Gaussians' opacity variation \\(\\Delta \\tilde{\\alpha}^{\\mathrm{I}}\\) w.r.t training images for a dynamic (i.e. containing transient objects) scene (top, Trevi Fountain from Phototourism [6]); and a static scene (bottom, Locomotive from Tanks & Temples [9]). The left column views are rendered using only static Gaussians (i.e having \\(\\mathbf{Var}\\left[\\Delta \\tilde{\\alpha}^{\\mathrm{I}}\\right] = 0\\)), whereas right column views are rendered using all Gaussians." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.391, + 0.789, + 0.466 + ], + "angle": 0, + "content": "where \\(\\mathbf{T}\\) is the temperature hyper-parameter. We do not use any annealing on the temperature parameter, rather we set \\(\\mathbf{T} = 0.1\\) to force, from the very beginning of the training, the opacity variation to match a binary distribution. During the evaluation, we fix \\(\\mathbf{U}\\) at 0.5. The final image-dependent 3D Gaussian opacity is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.413, + 0.472, + 0.787, + 0.492 + ], + "angle": 0, + "content": "\\[\n\\alpha^ {\\mathbf {I}} = \\max \\left(\\alpha - \\Delta \\tilde {\\alpha} ^ {\\mathbf {I}}, 0\\right). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.498, + 0.789, + 0.559 + ], + "angle": 0, + "content": "With our formulation of the opacity, we can naturally disentangle between the transient and the static parts of the scene by defining transient Gaussians as the ones having non-zero variance of their associated parameter \\(\\Delta \\tilde{\\alpha}^{\\mathbf{I}}\\) over training images: \\(\\operatorname{Var}\\left[\\Delta \\tilde{\\alpha}^{\\mathbf{I}}\\right] \\neq 0\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.56, + 0.79, + 0.742 + ], + "angle": 0, + "content": "Figure 3 shows the histogram of \\(\\mathbf{Var}\\left[\\Delta \\tilde{\\alpha}^{\\mathbf{I}}\\right]\\) on two scenes: one without transient content (Locomotive) and one with occluders (Trevi). We also show two rendered images of each scene: the right image using all the Gaussians and the left one using only Gaussians not detected as transient. The histogram shows that for a static scene, we detected less than \\(4\\%\\) of the Gaussians as transient and around \\(96\\%\\) of the Gaussians having \\(\\mathbf{Var}\\left[\\Delta \\tilde{\\alpha}^{\\mathbf{I}}\\right] = 0\\). Removing the \\(4\\%\\) of transient Gaussians does not affect the rendered image quality. For Trevi Fountain, our model was able to represent transient objects using less than \\(20\\%\\) of the Gaussians while maintaining the majority of the Gaussians to reconstruct static parts of the scene. Comparing the two rendered images shows that the transient Gaussians correspond to the mass of tourists present near the fountain, while removing the transient Gaussians resulted in an occluders-free rendering." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.761, + 0.378, + 0.779 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.789, + 0.454, + 0.804 + ], + "angle": 0, + "content": "5.1 Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Similar to 3DGS, we minimize the \\(\\mathcal{L}_1\\) loss combined with D-SSIM term to optimize the Gaussians parameters, the MLP \\(\\mathcal{F}_{\\theta}\\) weights, the Hash Encoder pa" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.683, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SWAG" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.784, + 0.201 + ], + "angle": 0, + "content": "Table 1: Quantitative results - Results on three real-world scenes from Phototourism [6] and efficiency comparison among SoTA methods for NVS in-the-wild. Given the significant gap between previous baselines [3,7,14,21] and 3DGS [8], we marginalize on the different GPU configurations." + }, + { + "type": "table", + "bbox": [ + 0.227, + 0.214, + 0.781, + 0.309 + ], + "angle": 0, + "content": "
Bradenburg GateSacre CoeurTrevi FountainMean Efficiency
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓Train time (h)FPS
NeRF [15]18.900.8150.23115.600.7150.29116.140.6000.366--
NeRF-W [14]24.170.8900.16719.200.8070.19118.970.6980.265400†<1†
Ha-NeRF [3]24.040.8770.13920.020.8010.17120.180.6900.222452†0.20†
CR-NeRF [21]26.530.9000.10622.070.8230.15221.480.7110.206420†0.25†
RefinedFields [7]26.640.886-22.260.817-23.420.737-150†<1†
3DGS [8]19.990.8890.18017.570.8310.21918.470.7610.2340.50*181*
SWAG (ours)26.330.9290.13921.160.8600.18523.100.8150.2080.83*15.29*
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.301, + 0.31, + 0.7, + 0.323 + ], + "angle": 0, + "content": "† Results reported on [7]. * Results computed using a high-tier GPU." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.355, + 0.785, + 0.447 + ], + "angle": 0, + "content": "rameters, and the per-image embedding vectors. We fix the size of the latent embedding vectors to \\( n = 24 \\). For the Hash encoding parameters, we set the hash table size \\( T_{h} = 2^{19} \\), the finest resolution \\( N_{max} = 2048 \\), the number of levels \\( L = 12 \\) and the coarsest resolution \\( N_{min} = 16 \\). The MLP we employ has 3 layers of 64 hidden units. Additional details on hyperparameters are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.479, + 0.334, + 0.493 + ], + "angle": 0, + "content": "5.2 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.514, + 0.788, + 0.668 + ], + "angle": 0, + "content": "Similar to SoTA in-the-wild reconstruction methods, we evaluate SWAG's NVS capabilities using the Phototourism dataset [6] and report our results on three main touristic monuments: Brandenburg Gate, Sacre Coeur, and Trevi Fountain. We also evaluate our method on the benchmark introduced in NeRF-OSR [18] for outdoor scene relighting tasks. We report our performance using NeRF-MS's [11] data split on 4 sites: stjohann (St. Johann, Saarbrücken), lump (Landwehrplatz, Saarbrücken), st (Staatstheater, Saarbrücken), and europa (Galerie Europa, Saarbrücken) and the original NeRF-OSR [18]'s data split on 3 sites: stjohann, lump and \\(lk2\\) (Ludwigskirche, Saarbrücken). More details about the datasets can be found in supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.699, + 0.352, + 0.714 + ], + "angle": 0, + "content": "5.3 Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We provide visual comparisons with rendered images and report quantitative results based on common rendering metrics from the literature: Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS). Since only the embeddings associated with the training images are optimized during training, we follow NeRF-W's [14] evaluation approach: we optimize an embedding on the left half of each test image and report metrics on the right half." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.127 + ], + "angle": 0, + "content": "H. Dahmani et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.787, + 0.173 + ], + "angle": 0, + "content": "Table 2: Quantitative results - Results on five sites from NeRF-OSR [18] outdoor scenes benchmark." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.187, + 0.784, + 0.27 + ], + "angle": 0, + "content": "
stjohannlwpsteuropa
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
NeRF [15]14.890.4320.63911.510.4680.57417.200.5140.50217.490.5510.503
NeRF-W [14]21.230.6670.42619.610.6160.44520.310.6070.43820.000.6990.340
Ha-NeRF [3]17.190.6860.33120.030.6850.36517.300.5380.48317.790.6320.421
NeRF-MS [11]22.840.7930.23521.900.7190.33620.680.6300.40221.030.7210.294
3DGS [8]16.770.7410.26811.760.6090.41417.160.6290.40620.180.7820.252
SWAG (ours)23.910.8640.17222.070.7830.30322.290.7130.36423.740.8450.242
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.226, + 0.272, + 0.395, + 0.282 + ], + "angle": 0, + "content": "(a) Reported on NeRF-MS split [11]" + }, + { + "type": "table", + "bbox": [ + 0.342, + 0.283, + 0.664, + 0.344 + ], + "angle": 0, + "content": "
lk2stlwp
PSNR ↑SSIM ↑PSNR ↑SSIM ↑PSNR ↑SSIM ↑
NeRF-OSR [18]19.860.62615.830.55617.380.576
3DGS [8]15.910.68712.530.58513.720.659
SWAG (ours)19.590.75618.730.72118.610.753
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.345, + 0.345, + 0.646, + 0.357 + ], + "angle": 0, + "content": "(b) Reported on original benchmark split from NeRF-OSR [18]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.386, + 0.324, + 0.4 + ], + "angle": 0, + "content": "5.4 Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.412, + 0.787, + 0.593 + ], + "angle": 0, + "content": "Table 1 shows quantitative results on three Phototourism scenes. Unsurprisingly, optimizing 3DGS on in-the-wild photo collections showed poor results due to the lack of image appearance dependency. The 3D Gaussians colors converge to a mean color which makes the method unable to model visual changes caused by weather, lightning conditions and different camera specifications. The presence of transient objects also resulted in the persistence of artifacts. SWAG improves 3DGS performance and outperforms the baselines on SSIM across all datasets. In particular, our proposal improves 3DGS's PSNR by an average margin of 5.01 dB. SWAG outperforms NeRF-W across all datasets in terms of LPIPS, PSNR, and SSIM and shows competitive performance with CR-NeRF and RefinedFields despite the priors these two methods use and requiring significantly longer times to train." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.594, + 0.787, + 0.639 + ], + "angle": 0, + "content": "Table 2 shows quantitative results on NeRF-OSR [18] benchmark. SWAG improves 3DGS performance by an average PSNR margin of 5 dB and outperformed all the other baselines regardless of the train/test split choices." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.64, + 0.787, + 0.76 + ], + "angle": 0, + "content": "Figure 4 and Figure 5 show qualitative results of 3DGS and SWAG on Phototourism scenes and four of NeRF-OSR sites. In some parts of the scene, 3DGS presents artifacts and floaters either in under-observed part of the scene or in areas occluded by transient objects. Adding to that, the rendered color is shifted from the ground truth color as shown in Figure 4. SWAG produces visual appearances close to the ground truth thanks to our per-image conditioning. Compared to previous implicit methods [3,7,14,21], our proposal achieves one order of magnitude faster training and guarantees real-time rendering as shown in Table 1." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.784, + 0.465, + 0.8 + ], + "angle": 0, + "content": "5.5 Controllable appearance" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Modeling the image appearance with an embedding vector enabled us to learn a latent space that we can utilize to change the lighting and appearance of any" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.684, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SWAG" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.264, + 0.143, + 0.74, + 0.585 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.595, + 0.785, + 0.625 + ], + "angle": 0, + "content": "Fig. 4: Qualitative experimental results on three real-world scenes from Phototourism [6].s" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.654, + 0.788, + 0.731 + ], + "angle": 0, + "content": "viewpoint at inference time. On top of the training image embeddings shown in Figure 6, we can sample vectors from the appearance space using interpolation between the learned embeddings. These interpolations are smooth and natural as shown in Figure 7. We encourage the readers to explore the supplementary videos and appreciate more the naturalness of these interpolations." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.754, + 0.486, + 0.769 + ], + "angle": 0, + "content": "5.6 Transient Objects Removal" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "As explained in Section 4.2, our method is able to localize 3D Gaussians that represent transient objects and, consequently, is able to manipulate the scenes omitting these objects. It should be noted that SWAG is thoughtfully designed to model transient objects without allowing the Gaussians to vary their opacity" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.127 + ], + "angle": 0, + "content": "H. Dahmani et al." + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.144, + 0.774, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.624, + 0.788, + 0.667 + ], + "angle": 0, + "content": "Fig. 5: Qualitative experimental results on three real-world scenes from Phototourism [6] and four NeRF-OSR [18] scenes. We demonstrate the capability of SWAG to disentangle between static and transient parts of the scene." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.789, + 0.841 + ], + "angle": 0, + "content": "to compensate for appearance variation. Indeed, according to our experiments, it was observed that less than \\(20\\%\\) (average on trained scenes) of the Gaussians capture transient objects. Consequently, SWAG achieves a better reconstruction of the static scene parts without having blurry elongated Gaussians that arise in view-dependent appearances with 3DGS. The capabilities of our model of disentangling static and dynamic Gaussians are clearly highlighted in Figure 5." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.683, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SWAG" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.144, + 0.358, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.275, + 0.231, + 0.357, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.358, + 0.144, + 0.478, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.48, + 0.144, + 0.729, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.283, + 0.785, + 0.325 + ], + "angle": 0, + "content": "Fig.6: Appearance transfer - SWAG is able to render a scene at any viewpoint with the visual appearance of any training image thanks to the learned appearance embeddings." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.338, + 0.313, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.393, + 0.313, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.338, + 0.412, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.393, + 0.412, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.414, + 0.338, + 0.505, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.414, + 0.393, + 0.505, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.338, + 0.69, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.393, + 0.69, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.339, + 0.789, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.393, + 0.789, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.471, + 0.784, + 0.499 + ], + "angle": 0, + "content": "Fig. 7: Appearance interpolation - Interpolations between the appearance embeddings of two training images (left, right)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.528, + 0.342, + 0.542 + ], + "angle": 0, + "content": "5.7 Ablations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.784, + 0.583 + ], + "angle": 0, + "content": "In this section, we compare two variants of our method to analyze the contribution of each component of SWAG:" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.592, + 0.785, + 0.62 + ], + "angle": 0, + "content": "- SWAG-A, a variation wherein the transient detection part of our model is removed," + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.622, + 0.785, + 0.651 + ], + "angle": 0, + "content": "- SWAG-T, a variation wherein the image appearance color dependency is omitted." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.592, + 0.785, + 0.651 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.661, + 0.784, + 0.692 + ], + "angle": 0, + "content": "We compare the results of our model variation on Phototourism scenes and we report the results in Table 3. SWAG-A enabled appearance modeling and" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.717, + 0.784, + 0.746 + ], + "angle": 0, + "content": "Table 3: Quantitative results - Results of two ablations of SWAG : SWAG-A and SWAG-T on three real-world scenes from Phototourism [6]." + }, + { + "type": "table", + "bbox": [ + 0.259, + 0.76, + 0.74, + 0.838 + ], + "angle": 0, + "content": "
Bradenburg GateSacre CoeurTrevi Fountain
PSNR ↑SSIM↑LPIPS ↓PSNR ↑SSIM↑LPIPS ↓PSNR ↑SSIM↑LPIPS ↓
SWAG26.330.9290.13921.160.8600.18523.100.8150.208
SWAG-A26.440.9280.14121.020.8550.18322.350.8010.223
SWAG-T24.460.9140.16019.060.8360.20520.130.7690.244
" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.127 + ], + "angle": 0, + "content": "H. Dahmani et al." + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.144, + 0.365, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.282, + 0.2, + 0.362, + 0.21 + ], + "angle": 0, + "content": "Ground-truth" + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.144, + 0.455, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.395, + 0.2, + 0.43, + 0.21 + ], + "angle": 0, + "content": "3DGS" + }, + { + "type": "image", + "bbox": [ + 0.458, + 0.144, + 0.545, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.474, + 0.2, + 0.528, + 0.21 + ], + "angle": 0, + "content": "SWAG-A" + }, + { + "type": "image", + "bbox": [ + 0.548, + 0.144, + 0.635, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.564, + 0.2, + 0.618, + 0.21 + ], + "angle": 0, + "content": "SWAG-T" + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.144, + 0.724, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.661, + 0.2, + 0.701, + 0.21 + ], + "angle": 0, + "content": "SWAG" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.22, + 0.785, + 0.248 + ], + "angle": 0, + "content": "Fig. 8: Ablation study - Visual comparison of 3DGS, SWAG, and the two variants used in our ablation study: SWAG-T (no appearance) and SWAG-A (no transient)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.277, + 0.785, + 0.381 + ], + "angle": 0, + "content": "closer colors to ground truth compared to 3DGS where colors shift away and converge to an average color for all training images. Transient objects' presence still causes floaters in the scene alike 3DGS as shown in Figure 8. In contrast, SWAG-T enables the elimination of these floaters and produces an occluders-free scene while suffering from color alterations compared to ground truth like 3DGS. Combining these two variations enables SWAG to model varying appearances with the ability to reduce scene floaters." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.403, + 0.348, + 0.417 + ], + "angle": 0, + "content": "5.8 Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.426, + 0.785, + 0.653 + ], + "angle": 0, + "content": "SWAG leverages an MLP, a hash grid encoder, and image embedding to adapt 3DGS to in-the-wild scenarios. While this empowered 3DGS's reconstruction capabilities in these conditions as it was demonstrated in our experiments, it almost doubles the training time and has 10 times longer inference time per frame as shown Table 1 compared to 3DGS. Nonetheless, our method still requires significantly less training time than all previous in-the-wild baselines and achieves interactive rendering speed. Although SWAG succeeded in removing occluders using transient Gaussians, some parts of the scene where transient objects' presence is more frequent (eg. clouds in the sky and tourists on the roads) can suffer from being detected as transient. Eliminating all the transient Gaussians can leave some holes in these parts of the scene as seen in Figure 5 for the Trevi Fountain scene where some parts of the fountain contain black spots. One way to solve this is to remove Gaussians based on an adaptive threshold \\(\\lambda\\) rather than considering as static only Gaussian with zero variance: \\(\\mathbf{Var}\\left[\\Delta \\tilde{\\alpha}^{\\mathbf{I}}\\right] < \\lambda\\), but it requires manual tuning of this threshold \\(\\lambda\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.674, + 0.359, + 0.69 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In this paper, we presented SWAG, a method designed for tailoring 3DGS representations to in-the-wild scenarios. SWAG incorporates appearance modeling in the Gaussians' colors and employs an adaptive opacity modulation to handle the presence of transient objects. Extensive experiments demonstrate that SWAG achieves state-of-the-art results on two challenging benchmarks while exhibiting training times orders of magnitude faster than in-the-wild NVS baselines while enabling real-time rendering. As a first step in conditioning 3DGS for in-the-wild scene representations, this work suggests potential future research direction, such as extending SWAG to dynamic scenes." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.683, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SWAG" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.324, + 0.161 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.179, + 0.785, + 0.221 + ], + "angle": 0, + "content": "1. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.222, + 0.785, + 0.263 + ], + "angle": 0, + "content": "2. Botsch, M., Hornung, A., Zwicker, M., Kobbelt, L.: High-quality surface splatting on today's gpus. In: Proceedings Eurographics/IEEE VGTC Symposium Point-Based Graphics (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.264, + 0.785, + 0.292 + ], + "angle": 0, + "content": "3. Chen, X., Zhang, Q., Li, X., Chen, Y., Feng, Y., Wang, X., Wang, J.: Hallucinated neural radiance fields in the wild. In: CVPR. pp. 12943-12952 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.293, + 0.785, + 0.32 + ], + "angle": 0, + "content": "4. Fridovich-Keil, S., Meanti, G., Warburg, F.R., Recht, B., Kanazawa, A.: K-planes: Explicit radiance fields in space, time, and appearance. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.321, + 0.663, + 0.335 + ], + "angle": 0, + "content": "5. Grossman, J., Sc, H., Dally, W.: Point sample rendering (1999)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.336, + 0.785, + 0.376 + ], + "angle": 0, + "content": "6. Jin, Y., Mishkin, D., Mishchuk, A., Matas, J., Fua, P., Yi, K.M., Trulls, E.: Image matching across wide baselines: From paper to practice. International Journal of Computer Vision p. 517-547 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.377, + 0.785, + 0.404 + ], + "angle": 0, + "content": "7. Kassab, K., Schnepf, A., Franceschi, J.Y., Caraffa, L., Mary, J., Gouet-Brunet, V.: Refinedfields: Radiance fields refinement for unconstrained scenes (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.406, + 0.785, + 0.433 + ], + "angle": 0, + "content": "8. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.434, + 0.785, + 0.461 + ], + "angle": 0, + "content": "9. Knapitsch, A., Park, J., Zhou, Q.Y., Koltun, V.: Tanks and temples. ACM Transactions on Graphics pp. 1-13 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.462, + 0.785, + 0.489 + ], + "angle": 0, + "content": "10. Lassner, C.: Fast differentiable raycasting for neural rendering using sphere-based representations. CoRR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.49, + 0.785, + 0.517 + ], + "angle": 0, + "content": "1. Li, P., Wang, S., Yang, C., Liu, B., Qiu, W., Wang, H.: Nerf-ms: Neural radiance fields with multi-sequence. In: ICCV. pp. 18591-18600 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.518, + 0.785, + 0.546 + ], + "angle": 0, + "content": "2. Lin, J., Li, Z., Tang, X., Liu, J., Liu, S., Liu, J., Lu, Y., Wu, X., Xu, S., Yan, Y., Yang, W.: Vastgaussian: Vast 3d gaussians for large scene reconstruction (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.547, + 0.785, + 0.587 + ], + "angle": 0, + "content": "3. Maddison, C.J., Mnih, A., Teh, Y.W.: The concrete distribution: A continuous relaxation of discrete random variables. In: International Conference on Learning Representations (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.588, + 0.785, + 0.629 + ], + "angle": 0, + "content": "4. Martin-Brualla, R., Radwan, N., Sajjadi, M.S.M., Barron, J.T., Dosovitskiy, A., Duckworth, D.: NeRF in the Wild: Neural Radiance Fields for Unconstrained Photo Collections. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.63, + 0.785, + 0.671 + ], + "angle": 0, + "content": "5. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.672, + 0.785, + 0.699 + ], + "angle": 0, + "content": "6. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. CoRR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.7, + 0.785, + 0.727 + ], + "angle": 0, + "content": "7. Rematas, K., Liu, A., Srinivasan, P.P., Barron, J.T., Tagliasacchi, A., Funkhouser, T., Ferrari, V.: Urban radiance fields. CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.728, + 0.785, + 0.755 + ], + "angle": 0, + "content": "8. Rudnev, V., Elgharib, M., Smith, W., Liu, L., Golyanik, V., Theobalt, C.: Nerf for outdoor scene relighting. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.757, + 0.785, + 0.784 + ], + "angle": 0, + "content": "9. Turki, H., Ramanan, D., Satyanarayanan, M.: Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. In: CVPR. pp. 12922–12931 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.785, + 0.812 + ], + "angle": 0, + "content": "20. Yan, Z., Low, W.F., Chen, Y., Lee, G.H.: Multi-scale 3d gaussian splatting for anti-aliased rendering (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "21. Yang, Y., Zhang, S., Huang, Z., Zhang, Y., Tan, M.: Cross-ray neural radiance fields for novel-view synthesis from unconstrained image collections. In: ICCV (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.179, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.127 + ], + "angle": 0, + "content": "H. Dahmani et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "22. Yu, Z., Chen, A., Huang, B., Sattler, T., Geiger, A.: Mip-splatting: Alias-free 3d gaussian splatting (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.785, + 0.218 + ], + "angle": 0, + "content": "23. Zwicker, M., Pfister, H., van Baar, J., Gross, M.: Surface splattering. In: Proceedings of the 28th Annual Conference on Computer Graphics and Interactive Techniques. p. 371-378 (2001)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.218 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/SWAG_ Splatting in the Wild images with Appearance-conditioned Gaussians/3925312b-25b5-4314-8bc0-d6cd9369f8b8_origin.pdf b/2024/SWAG_ Splatting in the Wild images with Appearance-conditioned Gaussians/3925312b-25b5-4314-8bc0-d6cd9369f8b8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0d2707a410cda6a6652312e450c74d6c0a881d45 --- /dev/null +++ b/2024/SWAG_ Splatting in the Wild images with Appearance-conditioned Gaussians/3925312b-25b5-4314-8bc0-d6cd9369f8b8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:df6e108f60fb3d1477ba283ce5043a0e6cbb2b080c95db04bfddd397367f05e7 +size 47141855 diff --git a/2024/SWAG_ Splatting in the Wild images with Appearance-conditioned Gaussians/full.md b/2024/SWAG_ Splatting in the Wild images with Appearance-conditioned Gaussians/full.md new file mode 100644 index 0000000000000000000000000000000000000000..94434d0b3d41d7fea25df974de6fe0123f6a20b4 --- /dev/null +++ b/2024/SWAG_ Splatting in the Wild images with Appearance-conditioned Gaussians/full.md @@ -0,0 +1,295 @@ +# SWAG: Splatting in the Wild images with Appearance-conditioned Gaussians + +Hiba Dahmani, Moussab Bennehar, Nathan Piasco, Luis Roldao, and Dzmitry Tsishkou + +Noah's Ark, Huawei Paris Research Center, France + +![](images/561769fdaa80ab4de74a0575feb1e464c05d67d05c40e5c53ae7a0beef61280a.jpg) +(a) + +![](images/de79f879277bcfdb8cd9632812f4456e9f57c8f7ea91365d085ae16f9091e7ff.jpg) +Fig. 1: Given in-the-wild captures (a), our model enables transient objects removal (b) and scene reconstruction with variable appearances (c). + +![](images/167f9d0404685e581f83fa9957132f5cd11e6ce086b9fa8b11249223669726c3.jpg) +(b) + +![](images/4d676b85d17bd6ffd79a274970e119486582f75f74f7322b2b79afc80c5b8c6b.jpg) +(c) + +Abstract. Implicit neural representation methods have shown impressive advancements in learning 3D scenes from unstructured in-the-wild photo collections but are still limited by the large computational cost of volumetric rendering. Recently, 3D Gaussian Splatting emerged as a much faster alternative with superior rendering quality and training efficiency, especially for small-scale and object-centric scenarios. Nevertheless, this technique suffers from poor performance on unstructured in-the-wild data. To tackle this, we extend over 3D Gaussian Splatting to handle unstructured image collections. We achieve this by modeling appearance to seize photometric variations in the rendered images. Additionally, we introduce a new mechanism to train transient Gaussians to handle the presence of scene occluders in an unsupervised manner. Experiments on diverse photo collection scenes and multi-pass acquisition of outdoor landmarks show the effectiveness of our method over prior works achieving state-of-the-art results with improved efficiency. + +Keywords: 3D Gaussian Splatting $\cdot$ Unconstrained Photo Collection $\cdot$ Novel View Synthesis $\cdot$ Appearance Modeling $\cdot$ Real-time Rendering $\cdot$ Transient Object Removal + +# 1 Introduction + +Novel View Synthesis (NVS) and 3D scene reconstruction are long-standing, challenging tasks in the realms of computer vision and computer graphics. Over the past few years, Neural Radiance Fields (NeRF) [15] have demonstrated groundbreaking results in rendering photorealistic views from novel viewpoints using an implicit neural representation of the radiance and density fields in a scene. Although NeRFs are very effective in static scenes, their performance significantly degrades in dynamic scenarios (i.e. containing moving transient objects) or in the presence of changing conditions such as weather, exposure, and lighting (i.e. in-the-wild datasets). To tackle these challenges, NeRF-W [14] proposed an extension to NeRF enabling the reconstruction of outdoor landmarks from in-the-wild images. They achieve this by learning the changing appearance of the images through per-image embeddings. Furthermore, to handle the presence of occluders, the scene is decomposed into "static" and "transient" elements being modeled by separate radiance fields. These modifications significantly improved NeRF's performance when confronted with appearance changes and transient occluders. More recent works [3,21] improve transient object handling by leveraging 2D visibility maps. Nonetheless, these methods still suffer from the inherent cost of volumetric rendering. + +Recently, 3D Gaussian Splatting (3DGS) [8] attracted considerable interest thanks to its explicit representation resulting in faster training and rendering by leveraging GPU-based rasterization. 3DGS achieves fast training and matches visual quality obtained by the current SOTA NeRF methods [16]. This work represents the scene as a large number of anisotropic 3D Gaussians with color features and opacities. The Gaussians' set is then used in a differentiable splatting process to project the 3D Gaussians into a 2D image plane and blend them to get the rendered image. Similar to the original NeRF [15], 3DGS suffers from the same limitations regarding varying lighting conditions and the presence of occluders in the training images. + +In this work, we present SWAG, the first in-the-wild extension for 3DGS. To achieve this, we propose to expand the capabilities of 3DGS and improve its robustness in these scenarios. To this aim, we propose to capture the varying appearance of each image using learned embeddings and leveraging a multilayer perceptron (MLP), significantly improving our model's capabilities. Second, we learn an image-dependent opacity variation to each Gaussian that enables a better handling of transient objects and a finer scene reconstruction as illustrated in Figure 1. + +We conduct a variety of experiments on the Phototourism dataset [6] and NeRF-OSR [18] benchmark and show that not only do we improve 3DGS performance in these scenarios, but we also achieve SOTA rendering quality with significantly faster training and rendering speeds compared to previous works [3, 7,14,21]. To summarize, our main contributions are as follows: + +- We design a specific training scheme that captures occluders and enables transient object removal from a trained scene. +- We show that our method achieves SOTA performance in NVs in-the-wild scenarios. +- We demonstrate SWAG's abilities to generate new images with smooth visual transitions from learned appearances and to remove, in an unsupervised manner, transient objects from a captured scene. + +# 2 Related Work + +# 2.1 Neural Rendering in-the-wild + +The abundance of in-the-wild unconstrained image data has encouraged adapting neural radiance fields to such cases. Early works like NeRF-W [14] have proposed a disentanglement between static and transient occladers using two appearance and transient per-image embeddings in addition to two radiance fields for each component of the scene (i.e. static and transient). Ha-NeRF [3] on the other hand uses a 2D image-dependent visibility map to eliminate occladers instead of using a decoupled radiance field since transient phenomena are only observed in individual 2D images. This simplification reduces the blurry artifacts that NeRF-W [14] suffered from in the attempt to reconstruct transient phenomena with a 3D transient field. Building upon previous methods, CR-NeRF [21] enhances their performance by exploiting the interaction information from multiple rays and fusing it into global information. Using a light-weight segmentation network, this method learns a visibility map without the supervision of ground truth segmentation masks to eliminate transient parts in 2D images. Another recent work, RefinedFields [7], leverages K-Planes and generative priors for in-the-wild scenarios. The learning is done through the alternation of two stages: a scene fitting to optimize the K-Planes [4] representation and a scene enriching phase that finetunes a pre-trained generative prior and infers a new K-Planes representation. As presented, Implicit-field representations have found diverse adaptations for in-the-wild scenarios. However, their time-consuming training and inference pose a challenge to achieving real-time rendering. This constrains their application in practical scenarios where fast rendering speed is crucial, especially in diverse interactive 3D applications. + +# 2.2 Point-Based Rendering + +Although NeRF demonstrated exceptional capabilities in generating photorealistic images, the demand for faster and more efficient rendering methods became increasingly evident. Point-based rendering has emerged as a prominent approach in computer graphics. It mainly revolves around the representation and rendering of the scene using discrete primitives. Starting from the most elementary form, Point Sample Rendering [5] authors proposed to sample a fixed-size formless set of points from a surface. Despite the developed strategies to adequately sample objects or surfaces, point sample rendering still suffers from + +holes and discontinuity. To overcome the issue of 3D points sparsity, the following works used different geometric primitives like discs [2] and surfels [23] that have a larger extent compared to pixels. Differentiable point-based rendering is the groundwork of these methods since it enables end-to-end training of 3D primitives from images. Pulsar [10] introduced an efficient sphere-based differentiable renderer where the scene is represented by a set of spheres with learned positions, feature vectors, opacities, and radii. These parameters are optimized to minimize the photometric reconstruction loss using an efficient differentiable rendering. The rasterization is done using either a convolutional U-Net or a per-pixel one-by-one convolutional network applied to a 2D screen space feature map. The feature map is a blending function that combines each sphere channel information using weighting based on position, radius, and opacity. Recently 3DGS [8] used 3D Gaussians as primitives and introduced a fully differentiable fast tile-based rasterizer for Gaussian splats allowing real-time rendering while outperforming state-of-the-art visual quality. Despite the demonstrated capability of 3DGS, there is still room for refining this method to enhance the quality of rendered images. One observed issue with 3DGS [8] is the appearance of strong artifacts when changing the sampling rate. Additionally, the rendering quality declines significantly in low resolutions or far away from camera positions due to the aliasing caused by the pixel size compared to the screen size and the Nyquist frequency. + +# 2.3 3DGS Rendering Improvement + +To solve the aforementioned rendering issue, a multi-scale adaptation of 3DGS [20] has been introduced. This work explains the aliasing phenomena by the splitting of large amounts of Gaussians with a smaller extent than pixels in 3D regions featuring high-frequency details. To render at a certain scale, authors filter out too large or too small Gaussians. Additionally, small Gaussians are aggregated to form larger ones selected on bigger scales during training to prevent missing areas and low-frequency details. Motivated by Mip-NeRF's [1] rendering quality, Mip-Splatting [22] replaces the 2D dilation filter introduced in Surface splatting [23] using 2D Mip filters that replicate box filters to solve the aliasing issue and uses 3D smoothing filters with a maximal sampling rate obtained from training images. One other issue with 3DGS is its performance deterioration in scenes that may present variation in appearance. VastGaussian [12] introduced a decoupled appearance modeling using a pixel-wise appearance embedding and a CNN to predict a pixel-wise transformation for the rendered images. Hence these predicted transformations (multiplication, addition, and Gamma correction) work well for large scenes but they are insufficient to model more varying appearances that appear in-the-wild scenarios. Our work aims to extend in-the-wild 3DGS scene representation beyond closed-world setups. + +# 3 Background + +The scene in 3DGS [8] is represented by a set of 3D anisotropic Gaussians. Each Gaussian is parameterized by its centroid $\mathbf{x} \in \mathbb{R}^3$ , scale $\mathbf{S} \in \mathbb{R}^3$ , rotation matrix $\mathbf{R} \in \mathbb{R}^{3 \times 3}$ , opacity $\alpha \in \mathbb{R}$ and color $\mathbf{c} \in \mathbb{R}^3$ encoded in spherical harmonic (SH) coefficients. The 3D covariance matrix $\pmb{\Sigma}$ of the 3D Gaussian is obtained using its rotation matrix $\mathbf{R}$ and scale $\mathbf{S}$ : + +$$ +\boldsymbol {\Sigma} = \mathbf {R} \mathbf {S} \mathbf {S} ^ {T} \mathbf {R} ^ {T}. \tag {1} +$$ + +The 3D Gaussians are defined in world space following: + +$$ +\mathbf {G} (\mathbf {y}) = e ^ {- \frac {1}{2} (\mathbf {y} - \mathbf {x}) ^ {T} \boldsymbol {\Sigma} ^ {- 1} (\mathbf {y} - \mathbf {x})}. \tag {2} +$$ + +Given $\mathbf{J}$ , the Jacobian of the affine projective transformation and $\mathbf{W}$ , a viewing transformation, the covariance matrix $\pmb{\Sigma}'$ in camera coordinates is as follows: + +$$ +\boldsymbol {\Sigma} ^ {\prime} = \mathbf {J} \mathbf {W} \boldsymbol {\Sigma} \mathbf {W} ^ {T} \mathbf {J} ^ {T}. \tag {3} +$$ + +These 3D Gaussians are projected to 2D splats and blended during a fast differentiable $\alpha$ -blending process to get 2D rendered images. Each pixel $\mathbf{y}$ color value $\mathbf{C}$ is calculated using N-ordered 2D splats using the formula: + +$$ +\mathbf {C} = \sum_ {i \in \mathbf {N}} \alpha_ {i} ^ {\prime} \mathbf {c} _ {i} \prod_ {j = 1} ^ {i - 1} \left(1 - \alpha_ {j} ^ {\prime}\right), \tag {4} +$$ + +$\alpha_{i}^{\prime}$ is the final opacity of the Gaussian obtained by: + +$$ +\alpha_ {i} ^ {\prime} = \alpha_ {i} * e ^ {- \frac {1}{2} \left(\mathbf {y} ^ {\prime} - \mathbf {x} ^ {\prime}\right) ^ {T} \boldsymbol {\Sigma} ^ {\prime - 1} \left(\mathbf {y} ^ {\prime} - \mathbf {x} ^ {\prime}\right)}. \tag {5} +$$ + +where $\mathbf{x}'$ and $\mathbf{y}'$ are coordinates in the projected space. + +Although 3D Gaussians can be arbitrarily initialized, a good prior is often required for optimal results. Typically, the 3D Gaussians centers are initialized using the sparse Structure from Motion (SfM) point cloud obtained from the set of images. To avoid leaving huge holes in the scene, their corresponding covariances are initialized to have isotropic Gaussians with initial radii equal to the mean distance to neighboring points. During training, the 3D Gaussian features are optimized such that the photometric difference between the rendered image $\mathbf{I}_{\mathbf{r}}$ and the ground truth one $\mathbf{I}_{\mathbf{gt}}$ is minimized with intervening controlling steps. To control the Gaussians' density and to address under- and over-distribution issues, authors use an adaptive densification step where Gaussians with large spatial gradients are split while transparent Gaussians that have very low opacity are pruned. + +Although 3DGS [8] works well on object-centric and small scenes, it struggles with scenes presenting varying appearances and transient objects. + +![](images/771b76bdcdd2edda9b39e8917edcd845805a7c4f5b2d6b8cf21f0354825106e8.jpg) +Fig. 2: SWAG model architecture - In addition to the typical Gaussians' features, we also optimize a Hash Grid encoding their centers $\mathbf{emb}(\mathbf{x})$ , a per-image embedding vector $l_{I}$ and an MLP. This MLP takes as inputs the Gaussians' colors $c$ , the associated image embedding $l_{I}$ , and their encoded centers $\mathbf{emb}(\mathbf{x})$ and outputs an image-dependent color $\mathbf{c}^{\mathbf{I}}$ as well as an image-dependent opacity variation parameter $\Delta \alpha^{I}$ . This parameter is set as the location variable of a concrete distribution which we sample to get the opacity variation $\Delta \tilde{\alpha}^{\mathbf{I}}$ . Leveraging this opacity variation across diverse training images enables identifying and excluding transient Gaussians within the scene, as demonstrated by the grey Gaussians. + +# 4 Method + +In this section, we introduce SWAG, a novel 3DGS-based method for 3D scene reconstruction from in-the-wild photo collections. We propose to adapt the 3D Gaussians parameters to handle variable visual appearances and the presence of occluders typically found in such unconstrained image collections. + +In Section 4.1, we explain how SWAG handles appearance variations using image-dependent embeddings injected into the Gaussian's colors. In Section 4.2, we explain how our proposal handles transient occluders by learning image-dependent Gaussians' opacities variations. Since our model's knowledge about the presence of occluders is embedded in the Gaussians opacities variation, it easily enables a clear disentanglement between static and transient elements of the scene. The overall architecture of our method is illustrated in Figure 2. + +# 4.1 Appearance Variation Modeling + +To adapt 3DGS to photometric variations, we associate, similarly to previous methods [14, 19], a trainable embedding vector $l_{I}$ for each image. To inject the appearance information, one naive solution would be to calculate a global transformation per image to the 3D Gaussian colors or SH features. This strategy, however, might be insufficient since it cannot model photometric features at the local level in an image. To model local appearance variances in an image, one possible way would be to predict a local transformation or directly a specific color for each Gaussian on a per-image basis. However, such solution is intractable as the number of training parameters will grow according to the number of Gaussians times the number of images. + +In SWAG, we propose a more effective solution where we use an MLP that takes as input a concatenation of the image embeddings and a positional encoding of the Gaussians' centers. By using a positional encoding conditioned on the center of the Gaussians, we are able to model local appearance variations that occur in the image while maintaining a tractable number of parameters to optimize. Initially, we considered estimating an affine color transformation similar to Urban Radiance Fields [17] to be applied to each Gaussian color. However, our initial results showed that affine color transformations alone cannot model all appearance changes. To solve this problem, we directly predict the Gaussian color from the MLP. We noticed that spatial embeddings coupled with an MLP produce a smoothing effect on the predicted colors. Hence, to render view-dependent peculiarities and high-frequency details and counterbalance this smoothing effect, we also add the Gaussian color obtained from the SH coefficients as input to the MLP. + +Our final solution employs an MLP $\mathcal{F}_{\theta}$ that, for a given view, takes as input the color of the 3D Gaussian $\mathbf{c}$ , the embedding vector of its center $\mathbf{emb}(\mathbf{x})$ and the image embedding $l_{I}$ , and outputs the image-conditioned color $\mathbf{c}^{\mathbf{I}}$ . Accordingly, for a 3D Gaussian, the image-dependent color $\mathbf{c}^{\mathbf{I}}$ is computed as follows: + +$$ +\mathbf {c} ^ {\mathbf {I}} = \mathcal {F} _ {\theta} (\mathbf {c}, \mathbf {e m b} (\mathbf {x}), l _ {I}). \tag {6} +$$ + +# 4.2 Transient Gaussians Modeling + +Our initial results of applying 3DGS to in-the-wild scenarios showed that Gaussians representing transient objects produce blurred and elongated artifacts, due to their optimization across all images, despite transients appearing in only a few frames. Taking this observation into consideration, we introduce a learnable image-dependent opacity variation term $\Delta \tilde{\alpha}^{\mathbf{I}}$ to each Gaussian. + +On the one hand, this opacity variation parameter allows Gaussians to reconstruct occluders present in some images. On the other hand, it enables those same Gaussians to be transparent (i.e. not involved in the image rendering process in other images) where these occluders are absent. During training, we want to encourage this variation parameter to either fully mask (i.e. in the case of occluders) or fully maintain (i.e. in the case of the static background) the 3D Gaussians' opacities. For this purpose, we sample $\Delta \tilde{\alpha}^{\mathbf{I}}$ using a Binary Concrete random variable [13]. This distribution is a continuous approximation of a Bernoulli distribution concentrating most of the mass on the boundaries of the interval [0, 1]. We return an additional output, $\Delta \alpha^{\mathbf{I}}$ , of the previously introduced MLP $\mathcal{F}_{\theta}$ as the location parameter to the concrete function: + +$$ +\left(\mathbf {c} ^ {\mathbf {I}}, \Delta \alpha^ {\mathbf {I}}\right) = \mathcal {F} _ {\theta} (\mathbf {c}, \mathbf {e m b} (\mathbf {x}), l _ {I}), \tag {7} +$$ + +$$ +\Delta \tilde {\alpha} ^ {\mathbf {I}} = \operatorname {s i g m o i d} \left[ \frac {1}{\mathbf {T}} \left(l o g \left(| \Delta \alpha^ {\mathbf {I}} |\right) + l o g (\mathbf {U}) - l o g (1 - \mathbf {U})\right) \right], \tag {8} +$$ + +$$ +\mathbf {U} \sim \operatorname {U n i f o r m} (0, 1), \tag {9} +$$ + +![](images/9a997299c4950565ad329fd41551e9cb2a73cca67526102f006e81af33f1c081.jpg) +Fig. 3: Variance histogram analysis of the Gaussians' opacity variation $\Delta \tilde{\alpha}^{\mathrm{I}}$ w.r.t training images for a dynamic (i.e. containing transient objects) scene (top, Trevi Fountain from Phototourism [6]); and a static scene (bottom, Locomotive from Tanks & Temples [9]). The left column views are rendered using only static Gaussians (i.e having $\mathbf{Var}\left[\Delta \tilde{\alpha}^{\mathrm{I}}\right] = 0$ ), whereas right column views are rendered using all Gaussians. + +where $\mathbf{T}$ is the temperature hyper-parameter. We do not use any annealing on the temperature parameter, rather we set $\mathbf{T} = 0.1$ to force, from the very beginning of the training, the opacity variation to match a binary distribution. During the evaluation, we fix $\mathbf{U}$ at 0.5. The final image-dependent 3D Gaussian opacity is formulated as: + +$$ +\alpha^ {\mathbf {I}} = \max \left(\alpha - \Delta \tilde {\alpha} ^ {\mathbf {I}}, 0\right). \tag {10} +$$ + +With our formulation of the opacity, we can naturally disentangle between the transient and the static parts of the scene by defining transient Gaussians as the ones having non-zero variance of their associated parameter $\Delta \tilde{\alpha}^{\mathbf{I}}$ over training images: $\operatorname{Var}\left[\Delta \tilde{\alpha}^{\mathbf{I}}\right] \neq 0$ . + +Figure 3 shows the histogram of $\mathbf{Var}\left[\Delta \tilde{\alpha}^{\mathbf{I}}\right]$ on two scenes: one without transient content (Locomotive) and one with occluders (Trevi). We also show two rendered images of each scene: the right image using all the Gaussians and the left one using only Gaussians not detected as transient. The histogram shows that for a static scene, we detected less than $4\%$ of the Gaussians as transient and around $96\%$ of the Gaussians having $\mathbf{Var}\left[\Delta \tilde{\alpha}^{\mathbf{I}}\right] = 0$ . Removing the $4\%$ of transient Gaussians does not affect the rendered image quality. For Trevi Fountain, our model was able to represent transient objects using less than $20\%$ of the Gaussians while maintaining the majority of the Gaussians to reconstruct static parts of the scene. Comparing the two rendered images shows that the transient Gaussians correspond to the mass of tourists present near the fountain, while removing the transient Gaussians resulted in an occluders-free rendering. + +# 5 Experiments + +# 5.1 Implementation details + +Similar to 3DGS, we minimize the $\mathcal{L}_1$ loss combined with D-SSIM term to optimize the Gaussians parameters, the MLP $\mathcal{F}_{\theta}$ weights, the Hash Encoder pa + +Table 1: Quantitative results - Results on three real-world scenes from Phototourism [6] and efficiency comparison among SoTA methods for NVS in-the-wild. Given the significant gap between previous baselines [3,7,14,21] and 3DGS [8], we marginalize on the different GPU configurations. + +
Bradenburg GateSacre CoeurTrevi FountainMean Efficiency
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓Train time (h)FPS
NeRF [15]18.900.8150.23115.600.7150.29116.140.6000.366--
NeRF-W [14]24.170.8900.16719.200.8070.19118.970.6980.265400†<1†
Ha-NeRF [3]24.040.8770.13920.020.8010.17120.180.6900.222452†0.20†
CR-NeRF [21]26.530.9000.10622.070.8230.15221.480.7110.206420†0.25†
RefinedFields [7]26.640.886-22.260.817-23.420.737-150†<1†
3DGS [8]19.990.8890.18017.570.8310.21918.470.7610.2340.50*181*
SWAG (ours)26.330.9290.13921.160.8600.18523.100.8150.2080.83*15.29*
+ +† Results reported on [7]. * Results computed using a high-tier GPU. + +rameters, and the per-image embedding vectors. We fix the size of the latent embedding vectors to $n = 24$ . For the Hash encoding parameters, we set the hash table size $T_{h} = 2^{19}$ , the finest resolution $N_{max} = 2048$ , the number of levels $L = 12$ and the coarsest resolution $N_{min} = 16$ . The MLP we employ has 3 layers of 64 hidden units. Additional details on hyperparameters are provided in the supplementary material. + +# 5.2 Datasets + +Similar to SoTA in-the-wild reconstruction methods, we evaluate SWAG's NVS capabilities using the Phototourism dataset [6] and report our results on three main touristic monuments: Brandenburg Gate, Sacre Coeur, and Trevi Fountain. We also evaluate our method on the benchmark introduced in NeRF-OSR [18] for outdoor scene relighting tasks. We report our performance using NeRF-MS's [11] data split on 4 sites: stjohann (St. Johann, Saarbrücken), lump (Landwehrplatz, Saarbrücken), st (Staatstheater, Saarbrücken), and europa (Galerie Europa, Saarbrücken) and the original NeRF-OSR [18]'s data split on 3 sites: stjohann, lump and $lk2$ (Ludwigskirche, Saarbrücken). More details about the datasets can be found in supplementary materials. + +# 5.3 Evaluation + +We provide visual comparisons with rendered images and report quantitative results based on common rendering metrics from the literature: Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS). Since only the embeddings associated with the training images are optimized during training, we follow NeRF-W's [14] evaluation approach: we optimize an embedding on the left half of each test image and report metrics on the right half. + +Table 2: Quantitative results - Results on five sites from NeRF-OSR [18] outdoor scenes benchmark. + +
stjohannlwpsteuropa
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
NeRF [15]14.890.4320.63911.510.4680.57417.200.5140.50217.490.5510.503
NeRF-W [14]21.230.6670.42619.610.6160.44520.310.6070.43820.000.6990.340
Ha-NeRF [3]17.190.6860.33120.030.6850.36517.300.5380.48317.790.6320.421
NeRF-MS [11]22.840.7930.23521.900.7190.33620.680.6300.40221.030.7210.294
3DGS [8]16.770.7410.26811.760.6090.41417.160.6290.40620.180.7820.252
SWAG (ours)23.910.8640.17222.070.7830.30322.290.7130.36423.740.8450.242
+ +(a) Reported on NeRF-MS split [11] + +
lk2stlwp
PSNR ↑SSIM ↑PSNR ↑SSIM ↑PSNR ↑SSIM ↑
NeRF-OSR [18]19.860.62615.830.55617.380.576
3DGS [8]15.910.68712.530.58513.720.659
SWAG (ours)19.590.75618.730.72118.610.753
+ +(b) Reported on original benchmark split from NeRF-OSR [18] + +# 5.4 Results + +Table 1 shows quantitative results on three Phototourism scenes. Unsurprisingly, optimizing 3DGS on in-the-wild photo collections showed poor results due to the lack of image appearance dependency. The 3D Gaussians colors converge to a mean color which makes the method unable to model visual changes caused by weather, lightning conditions and different camera specifications. The presence of transient objects also resulted in the persistence of artifacts. SWAG improves 3DGS performance and outperforms the baselines on SSIM across all datasets. In particular, our proposal improves 3DGS's PSNR by an average margin of 5.01 dB. SWAG outperforms NeRF-W across all datasets in terms of LPIPS, PSNR, and SSIM and shows competitive performance with CR-NeRF and RefinedFields despite the priors these two methods use and requiring significantly longer times to train. + +Table 2 shows quantitative results on NeRF-OSR [18] benchmark. SWAG improves 3DGS performance by an average PSNR margin of 5 dB and outperformed all the other baselines regardless of the train/test split choices. + +Figure 4 and Figure 5 show qualitative results of 3DGS and SWAG on Phototourism scenes and four of NeRF-OSR sites. In some parts of the scene, 3DGS presents artifacts and floaters either in under-observed part of the scene or in areas occluded by transient objects. Adding to that, the rendered color is shifted from the ground truth color as shown in Figure 4. SWAG produces visual appearances close to the ground truth thanks to our per-image conditioning. Compared to previous implicit methods [3,7,14,21], our proposal achieves one order of magnitude faster training and guarantees real-time rendering as shown in Table 1. + +# 5.5 Controllable appearance + +Modeling the image appearance with an embedding vector enabled us to learn a latent space that we can utilize to change the lighting and appearance of any + +![](images/e82fd0dbe4b928068f725fc5a60468d6d52046dcd082afd9fea9e03be92b981c.jpg) +Fig. 4: Qualitative experimental results on three real-world scenes from Phototourism [6].s + +viewpoint at inference time. On top of the training image embeddings shown in Figure 6, we can sample vectors from the appearance space using interpolation between the learned embeddings. These interpolations are smooth and natural as shown in Figure 7. We encourage the readers to explore the supplementary videos and appreciate more the naturalness of these interpolations. + +# 5.6 Transient Objects Removal + +As explained in Section 4.2, our method is able to localize 3D Gaussians that represent transient objects and, consequently, is able to manipulate the scenes omitting these objects. It should be noted that SWAG is thoughtfully designed to model transient objects without allowing the Gaussians to vary their opacity + +![](images/c90e427c1332e0d4ceb175188e42164b82b55b2c231a4d105ce3a43b117b5af6.jpg) +Fig. 5: Qualitative experimental results on three real-world scenes from Phototourism [6] and four NeRF-OSR [18] scenes. We demonstrate the capability of SWAG to disentangle between static and transient parts of the scene. + +to compensate for appearance variation. Indeed, according to our experiments, it was observed that less than $20\%$ (average on trained scenes) of the Gaussians capture transient objects. Consequently, SWAG achieves a better reconstruction of the static scene parts without having blurry elongated Gaussians that arise in view-dependent appearances with 3DGS. The capabilities of our model of disentangling static and dynamic Gaussians are clearly highlighted in Figure 5. + +![](images/acbb1c6965f66d14cad980907fba8c253cf8bebe84a09c9fd18de3192492ef07.jpg) + +![](images/e0c574d11ffecd5a20135999fc0df1771463d1cbec21e169a9a867f76a649804.jpg) +Fig.6: Appearance transfer - SWAG is able to render a scene at any viewpoint with the visual appearance of any training image thanks to the learned appearance embeddings. + +![](images/631bad8b5a26cc9900af96c60694257fb2116a722b070210293c8029562043b9.jpg) + +![](images/9628c916e29ffac9cf74c6214c98113f225c6d6ed915f78723950da8e48ee47d.jpg) + +![](images/1d7997630707f1bea131c6ebe20ecc8153100852c13d728fd4120c412d9de931.jpg) + +![](images/6ee3410d2010e59f2474a98013a597daa0d647378c3cfcb4c9b90d9525391646.jpg) +Fig. 7: Appearance interpolation - Interpolations between the appearance embeddings of two training images (left, right). + +![](images/2bd9880599c5f4f28b85add60b889955c0addadae12ecdeab86decdaead1cf2a.jpg) + +![](images/2da72133034dd24738375d1d9f43669ae824ac3370b995d6800e99cef210b2d5.jpg) + +![](images/99527152e06fc4f43eefac896bd0b6da6c56b2b23ce581f2156ecf33bfa32fcf.jpg) + +![](images/ed22996c2f5e7ad9109f2472e7da4d4913347a73254fe97704d95c1b0fdcc188.jpg) + +![](images/2bdd6a400cd7870e89bcc5dc7b87f115df0b617d4832298b19e451abb304477f.jpg) + +![](images/017dfaa40d47e50c5c982ecd9e0d46f36d262779126476a2b27a93e3488c1823.jpg) + +![](images/561fec92faf15f1c7387a9fe13b3c93dbfae955d50088c4e67c4b093ef410c46.jpg) + +![](images/20e059e50e3f88b63759b1cd225c410eb256b4d7f8a51c1c00114296dfc4b69e.jpg) + +# 5.7 Ablations + +In this section, we compare two variants of our method to analyze the contribution of each component of SWAG: + +- SWAG-A, a variation wherein the transient detection part of our model is removed, +- SWAG-T, a variation wherein the image appearance color dependency is omitted. + +We compare the results of our model variation on Phototourism scenes and we report the results in Table 3. SWAG-A enabled appearance modeling and + +Table 3: Quantitative results - Results of two ablations of SWAG : SWAG-A and SWAG-T on three real-world scenes from Phototourism [6]. + +
Bradenburg GateSacre CoeurTrevi Fountain
PSNR ↑SSIM↑LPIPS ↓PSNR ↑SSIM↑LPIPS ↓PSNR ↑SSIM↑LPIPS ↓
SWAG26.330.9290.13921.160.8600.18523.100.8150.208
SWAG-A26.440.9280.14121.020.8550.18322.350.8010.223
SWAG-T24.460.9140.16019.060.8360.20520.130.7690.244
+ +![](images/8fda2fcf75e21be26ef7293d5b450fb61dffb01ed073fb9598f8aa3b4c2eaec9.jpg) +Ground-truth + +![](images/5c9f1b066b14a3bbb46bac382bc7b56fb5cc5e5beea7d5b30d1439bd2c8775e9.jpg) +3DGS +Fig. 8: Ablation study - Visual comparison of 3DGS, SWAG, and the two variants used in our ablation study: SWAG-T (no appearance) and SWAG-A (no transient). + +![](images/d82228f05ff0e38d5bf9a80750aad4dec8b8b9790b77fc7ffe1199d04e88f2a5.jpg) +SWAG-A + +![](images/f4e3c9e218e9ce4916d2a4f665366da6c32ef0a67f1f8469b42bfb52f07f6f4e.jpg) +SWAG-T + +![](images/4584dc1a5a2997d25bec2b9aabcf1d3eee0754215b2370d64380e4344e86c65c.jpg) +SWAG + +closer colors to ground truth compared to 3DGS where colors shift away and converge to an average color for all training images. Transient objects' presence still causes floaters in the scene alike 3DGS as shown in Figure 8. In contrast, SWAG-T enables the elimination of these floaters and produces an occluders-free scene while suffering from color alterations compared to ground truth like 3DGS. Combining these two variations enables SWAG to model varying appearances with the ability to reduce scene floaters. + +# 5.8 Discussion + +SWAG leverages an MLP, a hash grid encoder, and image embedding to adapt 3DGS to in-the-wild scenarios. While this empowered 3DGS's reconstruction capabilities in these conditions as it was demonstrated in our experiments, it almost doubles the training time and has 10 times longer inference time per frame as shown Table 1 compared to 3DGS. Nonetheless, our method still requires significantly less training time than all previous in-the-wild baselines and achieves interactive rendering speed. Although SWAG succeeded in removing occluders using transient Gaussians, some parts of the scene where transient objects' presence is more frequent (eg. clouds in the sky and tourists on the roads) can suffer from being detected as transient. Eliminating all the transient Gaussians can leave some holes in these parts of the scene as seen in Figure 5 for the Trevi Fountain scene where some parts of the fountain contain black spots. One way to solve this is to remove Gaussians based on an adaptive threshold $\lambda$ rather than considering as static only Gaussian with zero variance: $\mathbf{Var}\left[\Delta \tilde{\alpha}^{\mathbf{I}}\right] < \lambda$ , but it requires manual tuning of this threshold $\lambda$ . + +# 6 Conclusion + +In this paper, we presented SWAG, a method designed for tailoring 3DGS representations to in-the-wild scenarios. SWAG incorporates appearance modeling in the Gaussians' colors and employs an adaptive opacity modulation to handle the presence of transient objects. Extensive experiments demonstrate that SWAG achieves state-of-the-art results on two challenging benchmarks while exhibiting training times orders of magnitude faster than in-the-wild NVS baselines while enabling real-time rendering. As a first step in conditioning 3DGS for in-the-wild scene representations, this work suggests potential future research direction, such as extending SWAG to dynamic scenes. + +# References + +1. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. ICCV (2021) +2. Botsch, M., Hornung, A., Zwicker, M., Kobbelt, L.: High-quality surface splatting on today's gpus. In: Proceedings Eurographics/IEEE VGTC Symposium Point-Based Graphics (2005) +3. Chen, X., Zhang, Q., Li, X., Chen, Y., Feng, Y., Wang, X., Wang, J.: Hallucinated neural radiance fields in the wild. In: CVPR. pp. 12943-12952 (2022) +4. Fridovich-Keil, S., Meanti, G., Warburg, F.R., Recht, B., Kanazawa, A.: K-planes: Explicit radiance fields in space, time, and appearance. In: CVPR (2023) +5. Grossman, J., Sc, H., Dally, W.: Point sample rendering (1999) +6. Jin, Y., Mishkin, D., Mishchuk, A., Matas, J., Fua, P., Yi, K.M., Trulls, E.: Image matching across wide baselines: From paper to practice. International Journal of Computer Vision p. 517-547 (2020) +7. Kassab, K., Schnepf, A., Franceschi, J.Y., Caraffa, L., Mary, J., Gouet-Brunet, V.: Refinedfields: Radiance fields refinement for unconstrained scenes (2024) +8. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (2023) +9. Knapitsch, A., Park, J., Zhou, Q.Y., Koltun, V.: Tanks and temples. ACM Transactions on Graphics pp. 1-13 (2017) +10. Lassner, C.: Fast differentiable raycasting for neural rendering using sphere-based representations. CoRR (2020) +1. Li, P., Wang, S., Yang, C., Liu, B., Qiu, W., Wang, H.: Nerf-ms: Neural radiance fields with multi-sequence. In: ICCV. pp. 18591-18600 (2023) +2. Lin, J., Li, Z., Tang, X., Liu, J., Liu, S., Liu, J., Lu, Y., Wu, X., Xu, S., Yan, Y., Yang, W.: Vastgaussian: Vast 3d gaussians for large scene reconstruction (2024) +3. Maddison, C.J., Mnih, A., Teh, Y.W.: The concrete distribution: A continuous relaxation of discrete random variables. In: International Conference on Learning Representations (2017) +4. Martin-Brualla, R., Radwan, N., Sajjadi, M.S.M., Barron, J.T., Dosovitskiy, A., Duckworth, D.: NeRF in the Wild: Neural Radiance Fields for Unconstrained Photo Collections. In: CVPR (2021) +5. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020) +6. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. CoRR (2022) +7. Rematas, K., Liu, A., Srinivasan, P.P., Barron, J.T., Tagliasacchi, A., Funkhouser, T., Ferrari, V.: Urban radiance fields. CVPR (2022) +8. Rudnev, V., Elgharib, M., Smith, W., Liu, L., Golyanik, V., Theobalt, C.: Nerf for outdoor scene relighting. In: ECCV (2022) +9. Turki, H., Ramanan, D., Satyanarayanan, M.: Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. In: CVPR. pp. 12922–12931 (2022) +20. Yan, Z., Low, W.F., Chen, Y., Lee, G.H.: Multi-scale 3d gaussian splatting for anti-aliased rendering (2023) +21. Yang, Y., Zhang, S., Huang, Z., Zhang, Y., Tan, M.: Cross-ray neural radiance fields for novel-view synthesis from unconstrained image collections. In: ICCV (2023) + +22. Yu, Z., Chen, A., Huang, B., Sattler, T., Geiger, A.: Mip-splatting: Alias-free 3d gaussian splatting (2023) +23. Zwicker, M., Pfister, H., van Baar, J., Gross, M.: Surface splattering. In: Proceedings of the 28th Annual Conference on Computer Graphics and Interactive Techniques. p. 371-378 (2001) \ No newline at end of file diff --git a/2024/SWAG_ Splatting in the Wild images with Appearance-conditioned Gaussians/images.zip b/2024/SWAG_ Splatting in the Wild images with Appearance-conditioned Gaussians/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..861b10cab170c9a83c2bf1547071ca8f9609267a --- /dev/null +++ b/2024/SWAG_ Splatting in the Wild images with Appearance-conditioned Gaussians/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d4d43dcbb8dea3d00f70bc1b0528276f9ed8a5e57a27c3916368dc4888668826 +size 780812 diff --git a/2024/SWAG_ Splatting in the Wild images with Appearance-conditioned Gaussians/layout.json b/2024/SWAG_ Splatting in the Wild images with Appearance-conditioned Gaussians/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d95fe4c3c2f80dfa2513599ff1cf56d11c055714 --- /dev/null +++ b/2024/SWAG_ Splatting in the Wild images with Appearance-conditioned Gaussians/layout.json @@ -0,0 +1,8210 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 159, + 111, + 455, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 111, + 455, + 148 + ], + "spans": [ + { + "bbox": [ + 159, + 111, + 455, + 148 + ], + "type": "text", + "content": "SWAG: Splatting in the Wild images with Appearance-conditioned Gaussians" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 140, + 168, + 474, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 168, + 474, + 194 + ], + "spans": [ + { + "bbox": [ + 140, + 168, + 474, + 194 + ], + "type": "text", + "content": "Hiba Dahmani, Moussab Bennehar, Nathan Piasco, Luis Roldao, and Dzmitry Tsishkou" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 202, + 202, + 411, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 202, + 411, + 213 + ], + "spans": [ + { + "bbox": [ + 202, + 202, + 411, + 213 + ], + "type": "text", + "content": "Noah's Ark, Huawei Paris Research Center, France" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 142, + 238, + 210, + 380 + ], + "blocks": [ + { + "bbox": [ + 142, + 238, + 210, + 380 + ], + "lines": [ + { + "bbox": [ + 142, + 238, + 210, + 380 + ], + "spans": [ + { + "bbox": [ + 142, + 238, + 210, + 380 + ], + "type": "image", + "image_path": "561769fdaa80ab4de74a0575feb1e464c05d67d05c40e5c53ae7a0beef61280a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 167, + 381, + 178, + 390 + ], + "lines": [ + { + "bbox": [ + 167, + 381, + 178, + 390 + ], + "spans": [ + { + "bbox": [ + 167, + 381, + 178, + 390 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 212, + 238, + 285, + 380 + ], + "blocks": [ + { + "bbox": [ + 212, + 238, + 285, + 380 + ], + "lines": [ + { + "bbox": [ + 212, + 238, + 285, + 380 + ], + "spans": [ + { + "bbox": [ + 212, + 238, + 285, + 380 + ], + "type": "image", + "image_path": "de79f879277bcfdb8cd9632812f4456e9f57c8f7ea91365d085ae16f9091e7ff.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 398, + 481, + 422 + ], + "lines": [ + { + "bbox": [ + 131, + 398, + 481, + 422 + ], + "spans": [ + { + "bbox": [ + 131, + 398, + 481, + 422 + ], + "type": "text", + "content": "Fig. 1: Given in-the-wild captures (a), our model enables transient objects removal (b) and scene reconstruction with variable appearances (c)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 285, + 238, + 402, + 380 + ], + "blocks": [ + { + "bbox": [ + 313, + 381, + 324, + 390 + ], + "lines": [ + { + "bbox": [ + 313, + 381, + 324, + 390 + ], + "spans": [ + { + "bbox": [ + 313, + 381, + 324, + 390 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 285, + 238, + 402, + 380 + ], + "lines": [ + { + "bbox": [ + 285, + 238, + 402, + 380 + ], + "spans": [ + { + "bbox": [ + 285, + 238, + 402, + 380 + ], + "type": "image", + "image_path": "167f9d0404685e581f83fa9957132f5cd11e6ce086b9fa8b11249223669726c3.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 403, + 238, + 471, + 380 + ], + "blocks": [ + { + "bbox": [ + 403, + 238, + 471, + 380 + ], + "lines": [ + { + "bbox": [ + 403, + 238, + 471, + 380 + ], + "spans": [ + { + "bbox": [ + 403, + 238, + 471, + 380 + ], + "type": "image", + "image_path": "4d676b85d17bd6ffd79a274970e119486582f75f74f7322b2b79afc80c5b8c6b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 433, + 381, + 444, + 390 + ], + "lines": [ + { + "bbox": [ + 433, + 381, + 444, + 390 + ], + "spans": [ + { + "bbox": [ + 433, + 381, + 444, + 390 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 159, + 458, + 455, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 458, + 455, + 624 + ], + "spans": [ + { + "bbox": [ + 159, + 458, + 455, + 624 + ], + "type": "text", + "content": "Abstract. Implicit neural representation methods have shown impressive advancements in learning 3D scenes from unstructured in-the-wild photo collections but are still limited by the large computational cost of volumetric rendering. Recently, 3D Gaussian Splatting emerged as a much faster alternative with superior rendering quality and training efficiency, especially for small-scale and object-centric scenarios. Nevertheless, this technique suffers from poor performance on unstructured in-the-wild data. To tackle this, we extend over 3D Gaussian Splatting to handle unstructured image collections. We achieve this by modeling appearance to seize photometric variations in the rendered images. Additionally, we introduce a new mechanism to train transient Gaussians to handle the presence of scene occluders in an unsupervised manner. Experiments on diverse photo collection scenes and multi-pass acquisition of outdoor landmarks show the effectiveness of our method over prior works achieving state-of-the-art results with improved efficiency." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 159, + 632, + 455, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 632, + 455, + 666 + ], + "spans": [ + { + "bbox": [ + 159, + 632, + 455, + 666 + ], + "type": "text", + "content": "Keywords: 3D Gaussian Splatting " + }, + { + "bbox": [ + 159, + 632, + 455, + 666 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 632, + 455, + 666 + ], + "type": "text", + "content": " Unconstrained Photo Collection " + }, + { + "bbox": [ + 159, + 632, + 455, + 666 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 632, + 455, + 666 + ], + "type": "text", + "content": " Novel View Synthesis " + }, + { + "bbox": [ + 159, + 632, + 455, + 666 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 632, + 455, + 666 + ], + "type": "text", + "content": " Appearance Modeling " + }, + { + "bbox": [ + 159, + 632, + 455, + 666 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 632, + 455, + 666 + ], + "type": "text", + "content": " Real-time Rendering " + }, + { + "bbox": [ + 159, + 632, + 455, + 666 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 632, + 455, + 666 + ], + "type": "text", + "content": " Transient Object Removal" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 229, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 229, + 127 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 229, + 127 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 481, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 481, + 355 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 481, + 355 + ], + "type": "text", + "content": "Novel View Synthesis (NVS) and 3D scene reconstruction are long-standing, challenging tasks in the realms of computer vision and computer graphics. Over the past few years, Neural Radiance Fields (NeRF) [15] have demonstrated groundbreaking results in rendering photorealistic views from novel viewpoints using an implicit neural representation of the radiance and density fields in a scene. Although NeRFs are very effective in static scenes, their performance significantly degrades in dynamic scenarios (i.e. containing moving transient objects) or in the presence of changing conditions such as weather, exposure, and lighting (i.e. in-the-wild datasets). To tackle these challenges, NeRF-W [14] proposed an extension to NeRF enabling the reconstruction of outdoor landmarks from in-the-wild images. They achieve this by learning the changing appearance of the images through per-image embeddings. Furthermore, to handle the presence of occluders, the scene is decomposed into \"static\" and \"transient\" elements being modeled by separate radiance fields. These modifications significantly improved NeRF's performance when confronted with appearance changes and transient occluders. More recent works [3,21] improve transient object handling by leveraging 2D visibility maps. Nonetheless, these methods still suffer from the inherent cost of volumetric rendering." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 356, + 481, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 356, + 481, + 475 + ], + "spans": [ + { + "bbox": [ + 133, + 356, + 481, + 475 + ], + "type": "text", + "content": "Recently, 3D Gaussian Splatting (3DGS) [8] attracted considerable interest thanks to its explicit representation resulting in faster training and rendering by leveraging GPU-based rasterization. 3DGS achieves fast training and matches visual quality obtained by the current SOTA NeRF methods [16]. This work represents the scene as a large number of anisotropic 3D Gaussians with color features and opacities. The Gaussians' set is then used in a differentiable splatting process to project the 3D Gaussians into a 2D image plane and blend them to get the rendered image. Similar to the original NeRF [15], 3DGS suffers from the same limitations regarding varying lighting conditions and the presence of occluders in the training images." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 475, + 481, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 475, + 481, + 570 + ], + "spans": [ + { + "bbox": [ + 133, + 475, + 481, + 570 + ], + "type": "text", + "content": "In this work, we present SWAG, the first in-the-wild extension for 3DGS. To achieve this, we propose to expand the capabilities of 3DGS and improve its robustness in these scenarios. To this aim, we propose to capture the varying appearance of each image using learned embeddings and leveraging a multilayer perceptron (MLP), significantly improving our model's capabilities. Second, we learn an image-dependent opacity variation to each Gaussian that enables a better handling of transient objects and a finer scene reconstruction as illustrated in Figure 1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 571, + 481, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 571, + 481, + 631 + ], + "spans": [ + { + "bbox": [ + 133, + 571, + 481, + 631 + ], + "type": "text", + "content": "We conduct a variety of experiments on the Phototourism dataset [6] and NeRF-OSR [18] benchmark and show that not only do we improve 3DGS performance in these scenarios, but we also achieve SOTA rendering quality with significantly faster training and rendering speeds compared to previous works [3, 7,14,21]. To summarize, our main contributions are as follows:" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "text", + "content": "H. Dahmani et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 138, + 641, + 481, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 641, + 481, + 665 + ], + "type": "text", + "content": "- We introduce a new method for reconstructing scenes from unconstrained photo collections with varying appearances." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 137, + 116, + 480, + 197 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 137, + 116, + 479, + 138 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 116, + 479, + 138 + ], + "spans": [ + { + "bbox": [ + 137, + 116, + 479, + 138 + ], + "type": "text", + "content": "- We design a specific training scheme that captures occluders and enables transient object removal from a trained scene." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 480, + 160 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 480, + 160 + ], + "type": "text", + "content": "- We show that our method achieves SOTA performance in NVs in-the-wild scenarios." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 163, + 479, + 197 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 163, + 479, + 197 + ], + "spans": [ + { + "bbox": [ + 138, + 163, + 479, + 197 + ], + "type": "text", + "content": "- We demonstrate SWAG's abilities to generate new images with smooth visual transitions from learned appearances and to remove, in an unsupervised manner, transient objects from a captured scene." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 217, + 236, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 217, + 236, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 217, + 236, + 228 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 240, + 309, + 252 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 240, + 309, + 252 + ], + "spans": [ + { + "bbox": [ + 132, + 240, + 309, + 252 + ], + "type": "text", + "content": "2.1 Neural Rendering in-the-wild" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 259, + 482, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 482, + 535 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 482, + 535 + ], + "type": "text", + "content": "The abundance of in-the-wild unconstrained image data has encouraged adapting neural radiance fields to such cases. Early works like NeRF-W [14] have proposed a disentanglement between static and transient occladers using two appearance and transient per-image embeddings in addition to two radiance fields for each component of the scene (i.e. static and transient). Ha-NeRF [3] on the other hand uses a 2D image-dependent visibility map to eliminate occladers instead of using a decoupled radiance field since transient phenomena are only observed in individual 2D images. This simplification reduces the blurry artifacts that NeRF-W [14] suffered from in the attempt to reconstruct transient phenomena with a 3D transient field. Building upon previous methods, CR-NeRF [21] enhances their performance by exploiting the interaction information from multiple rays and fusing it into global information. Using a light-weight segmentation network, this method learns a visibility map without the supervision of ground truth segmentation masks to eliminate transient parts in 2D images. Another recent work, RefinedFields [7], leverages K-Planes and generative priors for in-the-wild scenarios. The learning is done through the alternation of two stages: a scene fitting to optimize the K-Planes [4] representation and a scene enriching phase that finetunes a pre-trained generative prior and infers a new K-Planes representation. As presented, Implicit-field representations have found diverse adaptations for in-the-wild scenarios. However, their time-consuming training and inference pose a challenge to achieving real-time rendering. This constrains their application in practical scenarios where fast rendering speed is crucial, especially in diverse interactive 3D applications." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 551, + 278, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 551, + 278, + 563 + ], + "spans": [ + { + "bbox": [ + 132, + 551, + 278, + 563 + ], + "type": "text", + "content": "2.2 Point-Based Rendering" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "content": "Although NeRF demonstrated exceptional capabilities in generating photorealistic images, the demand for faster and more efficient rendering methods became increasingly evident. Point-based rendering has emerged as a prominent approach in computer graphics. It mainly revolves around the representation and rendering of the scene using discrete primitives. Starting from the most elementary form, Point Sample Rendering [5] authors proposed to sample a fixed-size formless set of points from a surface. Despite the developed strategies to adequately sample objects or surfaces, point sample rendering still suffers from" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "type": "text", + "content": "SWAG" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "text", + "content": "holes and discontinuity. To overcome the issue of 3D points sparsity, the following works used different geometric primitives like discs [2] and surfels [23] that have a larger extent compared to pixels. Differentiable point-based rendering is the groundwork of these methods since it enables end-to-end training of 3D primitives from images. Pulsar [10] introduced an efficient sphere-based differentiable renderer where the scene is represented by a set of spheres with learned positions, feature vectors, opacities, and radii. These parameters are optimized to minimize the photometric reconstruction loss using an efficient differentiable rendering. The rasterization is done using either a convolutional U-Net or a per-pixel one-by-one convolutional network applied to a 2D screen space feature map. The feature map is a blending function that combines each sphere channel information using weighting based on position, radius, and opacity. Recently 3DGS [8] used 3D Gaussians as primitives and introduced a fully differentiable fast tile-based rasterizer for Gaussian splats allowing real-time rendering while outperforming state-of-the-art visual quality. Despite the demonstrated capability of 3DGS, there is still room for refining this method to enhance the quality of rendered images. One observed issue with 3DGS [8] is the appearance of strong artifacts when changing the sampling rate. Additionally, the rendering quality declines significantly in low resolutions or far away from camera positions due to the aliasing caused by the pixel size compared to the screen size and the Nyquist frequency." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 407, + 317, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 317, + 419 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 317, + 419 + ], + "type": "text", + "content": "2.3 3DGS Rendering Improvement" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 450, + 483, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 483, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 483, + 665 + ], + "type": "text", + "content": "To solve the aforementioned rendering issue, a multi-scale adaptation of 3DGS [20] has been introduced. This work explains the aliasing phenomena by the splitting of large amounts of Gaussians with a smaller extent than pixels in 3D regions featuring high-frequency details. To render at a certain scale, authors filter out too large or too small Gaussians. Additionally, small Gaussians are aggregated to form larger ones selected on bigger scales during training to prevent missing areas and low-frequency details. Motivated by Mip-NeRF's [1] rendering quality, Mip-Splatting [22] replaces the 2D dilation filter introduced in Surface splatting [23] using 2D Mip filters that replicate box filters to solve the aliasing issue and uses 3D smoothing filters with a maximal sampling rate obtained from training images. One other issue with 3DGS is its performance deterioration in scenes that may present variation in appearance. VastGaussian [12] introduced a decoupled appearance modeling using a pixel-wise appearance embedding and a CNN to predict a pixel-wise transformation for the rendered images. Hence these predicted transformations (multiplication, addition, and Gamma correction) work well for large scenes but they are insufficient to model more varying appearances that appear in-the-wild scenarios. Our work aims to extend in-the-wild 3DGS scene representation beyond closed-world setups." + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "type": "text", + "content": "H. Dahmani et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 225, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 225, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 225, + 129 + ], + "type": "text", + "content": "3 Background" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 144, + 482, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 144, + 482, + 203 + ], + "spans": [ + { + "bbox": [ + 131, + 144, + 482, + 203 + ], + "type": "text", + "content": "The scene in 3DGS [8] is represented by a set of 3D anisotropic Gaussians. Each Gaussian is parameterized by its centroid " + }, + { + "bbox": [ + 131, + 144, + 482, + 203 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 131, + 144, + 482, + 203 + ], + "type": "text", + "content": ", scale " + }, + { + "bbox": [ + 131, + 144, + 482, + 203 + ], + "type": "inline_equation", + "content": "\\mathbf{S} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 131, + 144, + 482, + 203 + ], + "type": "text", + "content": ", rotation matrix " + }, + { + "bbox": [ + 131, + 144, + 482, + 203 + ], + "type": "inline_equation", + "content": "\\mathbf{R} \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 131, + 144, + 482, + 203 + ], + "type": "text", + "content": ", opacity " + }, + { + "bbox": [ + 131, + 144, + 482, + 203 + ], + "type": "inline_equation", + "content": "\\alpha \\in \\mathbb{R}" + }, + { + "bbox": [ + 131, + 144, + 482, + 203 + ], + "type": "text", + "content": " and color " + }, + { + "bbox": [ + 131, + 144, + 482, + 203 + ], + "type": "inline_equation", + "content": "\\mathbf{c} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 131, + 144, + 482, + 203 + ], + "type": "text", + "content": " encoded in spherical harmonic (SH) coefficients. The 3D covariance matrix " + }, + { + "bbox": [ + 131, + 144, + 482, + 203 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}" + }, + { + "bbox": [ + 131, + 144, + 482, + 203 + ], + "type": "text", + "content": " of the 3D Gaussian is obtained using its rotation matrix " + }, + { + "bbox": [ + 131, + 144, + 482, + 203 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 131, + 144, + 482, + 203 + ], + "type": "text", + "content": " and scale " + }, + { + "bbox": [ + 131, + 144, + 482, + 203 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 131, + 144, + 482, + 203 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 274, + 215, + 481, + 228 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 274, + 215, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 274, + 215, + 481, + 228 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\Sigma} = \\mathbf {R} \\mathbf {S} \\mathbf {S} ^ {T} \\mathbf {R} ^ {T}. \\tag {1}", + "image_path": "8549c153d2fe75d325b99055f28c361efc6cb28fa45171ad8bdc873efe95b801.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 239, + 376, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 239, + 376, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 239, + 376, + 251 + ], + "type": "text", + "content": "The 3D Gaussians are defined in world space following:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 246, + 262, + 481, + 277 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 262, + 481, + 277 + ], + "spans": [ + { + "bbox": [ + 246, + 262, + 481, + 277 + ], + "type": "interline_equation", + "content": "\\mathbf {G} (\\mathbf {y}) = e ^ {- \\frac {1}{2} (\\mathbf {y} - \\mathbf {x}) ^ {T} \\boldsymbol {\\Sigma} ^ {- 1} (\\mathbf {y} - \\mathbf {x})}. \\tag {2}", + "image_path": "4aa85d9fde8bc7b95b9a800138001f8268d162d2ad72a25fcd8698c660293af2.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 288, + 481, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 288, + 481, + 312 + ], + "spans": [ + { + "bbox": [ + 131, + 288, + 481, + 312 + ], + "type": "text", + "content": "Given " + }, + { + "bbox": [ + 131, + 288, + 481, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{J}" + }, + { + "bbox": [ + 131, + 288, + 481, + 312 + ], + "type": "text", + "content": ", the Jacobian of the affine projective transformation and " + }, + { + "bbox": [ + 131, + 288, + 481, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 131, + 288, + 481, + 312 + ], + "type": "text", + "content": ", a viewing transformation, the covariance matrix " + }, + { + "bbox": [ + 131, + 288, + 481, + 312 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}'" + }, + { + "bbox": [ + 131, + 288, + 481, + 312 + ], + "type": "text", + "content": " in camera coordinates is as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 265, + 323, + 480, + 335 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 265, + 323, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 265, + 323, + 480, + 335 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\Sigma} ^ {\\prime} = \\mathbf {J} \\mathbf {W} \\boldsymbol {\\Sigma} \\mathbf {W} ^ {T} \\mathbf {J} ^ {T}. \\tag {3}", + "image_path": "ec5a2acecbc7ecb90f32b810c001d8656b609192bfb2276eb05dca556732b579.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 347, + 481, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 347, + 481, + 383 + ], + "spans": [ + { + "bbox": [ + 131, + 347, + 481, + 383 + ], + "type": "text", + "content": "These 3D Gaussians are projected to 2D splats and blended during a fast differentiable " + }, + { + "bbox": [ + 131, + 347, + 481, + 383 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 131, + 347, + 481, + 383 + ], + "type": "text", + "content": "-blending process to get 2D rendered images. Each pixel " + }, + { + "bbox": [ + 131, + 347, + 481, + 383 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 131, + 347, + 481, + 383 + ], + "type": "text", + "content": " color value " + }, + { + "bbox": [ + 131, + 347, + 481, + 383 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 131, + 347, + 481, + 383 + ], + "type": "text", + "content": " is calculated using N-ordered 2D splats using the formula:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 246, + 397, + 481, + 429 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 397, + 481, + 429 + ], + "spans": [ + { + "bbox": [ + 246, + 397, + 481, + 429 + ], + "type": "interline_equation", + "content": "\\mathbf {C} = \\sum_ {i \\in \\mathbf {N}} \\alpha_ {i} ^ {\\prime} \\mathbf {c} _ {i} \\prod_ {j = 1} ^ {i - 1} \\left(1 - \\alpha_ {j} ^ {\\prime}\\right), \\tag {4}", + "image_path": "ea37b1097c88b75a16b2e6d8c1494d4f7b8f26c801670ad4c0b3baf572ca30cf.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 436, + 358, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 436, + 358, + 449 + ], + "spans": [ + { + "bbox": [ + 132, + 436, + 358, + 449 + ], + "type": "inline_equation", + "content": "\\alpha_{i}^{\\prime}" + }, + { + "bbox": [ + 132, + 436, + 358, + 449 + ], + "type": "text", + "content": " is the final opacity of the Gaussian obtained by:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 234, + 461, + 480, + 476 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 234, + 461, + 480, + 476 + ], + "spans": [ + { + "bbox": [ + 234, + 461, + 480, + 476 + ], + "type": "interline_equation", + "content": "\\alpha_ {i} ^ {\\prime} = \\alpha_ {i} * e ^ {- \\frac {1}{2} \\left(\\mathbf {y} ^ {\\prime} - \\mathbf {x} ^ {\\prime}\\right) ^ {T} \\boldsymbol {\\Sigma} ^ {\\prime - 1} \\left(\\mathbf {y} ^ {\\prime} - \\mathbf {x} ^ {\\prime}\\right)}. \\tag {5}", + "image_path": "593cedbebbc55e30aea5111fc30db1878e2c5fef9088dd38dff421da369b3c3b.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 484, + 376, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 484, + 376, + 496 + ], + "spans": [ + { + "bbox": [ + 131, + 484, + 376, + 496 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 484, + 376, + 496 + ], + "type": "inline_equation", + "content": "\\mathbf{x}'" + }, + { + "bbox": [ + 131, + 484, + 376, + 496 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 484, + 376, + 496 + ], + "type": "inline_equation", + "content": "\\mathbf{y}'" + }, + { + "bbox": [ + 131, + 484, + 376, + 496 + ], + "type": "text", + "content": " are coordinates in the projected space." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 497, + 481, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 481, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 481, + 640 + ], + "type": "text", + "content": "Although 3D Gaussians can be arbitrarily initialized, a good prior is often required for optimal results. Typically, the 3D Gaussians centers are initialized using the sparse Structure from Motion (SfM) point cloud obtained from the set of images. To avoid leaving huge holes in the scene, their corresponding covariances are initialized to have isotropic Gaussians with initial radii equal to the mean distance to neighboring points. During training, the 3D Gaussian features are optimized such that the photometric difference between the rendered image " + }, + { + "bbox": [ + 130, + 497, + 481, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathbf{r}}" + }, + { + "bbox": [ + 130, + 497, + 481, + 640 + ], + "type": "text", + "content": " and the ground truth one " + }, + { + "bbox": [ + 130, + 497, + 481, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathbf{gt}}" + }, + { + "bbox": [ + 130, + 497, + 481, + 640 + ], + "type": "text", + "content": " is minimized with intervening controlling steps. To control the Gaussians' density and to address under- and over-distribution issues, authors use an adaptive densification step where Gaussians with large spatial gradients are split while transparent Gaussians that have very low opacity are pruned." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": "Although 3DGS [8] works well on object-centric and small scenes, it struggles with scenes presenting varying appearances and transient objects." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "type": "text", + "content": "SWAG" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 166, + 114, + 449, + 219 + ], + "blocks": [ + { + "bbox": [ + 166, + 114, + 449, + 219 + ], + "lines": [ + { + "bbox": [ + 166, + 114, + 449, + 219 + ], + "spans": [ + { + "bbox": [ + 166, + 114, + 449, + 219 + ], + "type": "image", + "image_path": "771b76bdcdd2edda9b39e8917edcd845805a7c4f5b2d6b8cf21f0354825106e8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 227, + 482, + 326 + ], + "lines": [ + { + "bbox": [ + 130, + 227, + 482, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 227, + 482, + 326 + ], + "type": "text", + "content": "Fig. 2: SWAG model architecture - In addition to the typical Gaussians' features, we also optimize a Hash Grid encoding their centers " + }, + { + "bbox": [ + 130, + 227, + 482, + 326 + ], + "type": "inline_equation", + "content": "\\mathbf{emb}(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 227, + 482, + 326 + ], + "type": "text", + "content": ", a per-image embedding vector " + }, + { + "bbox": [ + 130, + 227, + 482, + 326 + ], + "type": "inline_equation", + "content": "l_{I}" + }, + { + "bbox": [ + 130, + 227, + 482, + 326 + ], + "type": "text", + "content": " and an MLP. This MLP takes as inputs the Gaussians' colors " + }, + { + "bbox": [ + 130, + 227, + 482, + 326 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 227, + 482, + 326 + ], + "type": "text", + "content": ", the associated image embedding " + }, + { + "bbox": [ + 130, + 227, + 482, + 326 + ], + "type": "inline_equation", + "content": "l_{I}" + }, + { + "bbox": [ + 130, + 227, + 482, + 326 + ], + "type": "text", + "content": ", and their encoded centers " + }, + { + "bbox": [ + 130, + 227, + 482, + 326 + ], + "type": "inline_equation", + "content": "\\mathbf{emb}(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 227, + 482, + 326 + ], + "type": "text", + "content": " and outputs an image-dependent color " + }, + { + "bbox": [ + 130, + 227, + 482, + 326 + ], + "type": "inline_equation", + "content": "\\mathbf{c}^{\\mathbf{I}}" + }, + { + "bbox": [ + 130, + 227, + 482, + 326 + ], + "type": "text", + "content": " as well as an image-dependent opacity variation parameter " + }, + { + "bbox": [ + 130, + 227, + 482, + 326 + ], + "type": "inline_equation", + "content": "\\Delta \\alpha^{I}" + }, + { + "bbox": [ + 130, + 227, + 482, + 326 + ], + "type": "text", + "content": ". This parameter is set as the location variable of a concrete distribution which we sample to get the opacity variation " + }, + { + "bbox": [ + 130, + 227, + 482, + 326 + ], + "type": "inline_equation", + "content": "\\Delta \\tilde{\\alpha}^{\\mathbf{I}}" + }, + { + "bbox": [ + 130, + 227, + 482, + 326 + ], + "type": "text", + "content": ". Leveraging this opacity variation across diverse training images enables identifying and excluding transient Gaussians within the scene, as demonstrated by the grey Gaussians." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 349, + 201, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 349, + 201, + 361 + ], + "spans": [ + { + "bbox": [ + 132, + 349, + 201, + 361 + ], + "type": "text", + "content": "4 Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 374, + 482, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 374, + 482, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 482, + 422 + ], + "type": "text", + "content": "In this section, we introduce SWAG, a novel 3DGS-based method for 3D scene reconstruction from in-the-wild photo collections. We propose to adapt the 3D Gaussians parameters to handle variable visual appearances and the presence of occluders typically found in such unconstrained image collections." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 423, + 482, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 482, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 482, + 506 + ], + "type": "text", + "content": "In Section 4.1, we explain how SWAG handles appearance variations using image-dependent embeddings injected into the Gaussian's colors. In Section 4.2, we explain how our proposal handles transient occluders by learning image-dependent Gaussians' opacities variations. Since our model's knowledge about the presence of occluders is embedded in the Gaussians opacities variation, it easily enables a clear disentanglement between static and transient elements of the scene. The overall architecture of our method is illustrated in Figure 2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 525, + 322, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 322, + 537 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 322, + 537 + ], + "type": "text", + "content": "4.1 Appearance Variation Modeling" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 546, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 665 + ], + "type": "text", + "content": "To adapt 3DGS to photometric variations, we associate, similarly to previous methods [14, 19], a trainable embedding vector " + }, + { + "bbox": [ + 130, + 546, + 482, + 665 + ], + "type": "inline_equation", + "content": "l_{I}" + }, + { + "bbox": [ + 130, + 546, + 482, + 665 + ], + "type": "text", + "content": " for each image. To inject the appearance information, one naive solution would be to calculate a global transformation per image to the 3D Gaussian colors or SH features. This strategy, however, might be insufficient since it cannot model photometric features at the local level in an image. To model local appearance variances in an image, one possible way would be to predict a local transformation or directly a specific color for each Gaussian on a per-image basis. However, such solution is intractable as the number of training parameters will grow according to the number of Gaussians times the number of images." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "type": "text", + "content": "H. Dahmani et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "text", + "content": "In SWAG, we propose a more effective solution where we use an MLP that takes as input a concatenation of the image embeddings and a positional encoding of the Gaussians' centers. By using a positional encoding conditioned on the center of the Gaussians, we are able to model local appearance variations that occur in the image while maintaining a tractable number of parameters to optimize. Initially, we considered estimating an affine color transformation similar to Urban Radiance Fields [17] to be applied to each Gaussian color. However, our initial results showed that affine color transformations alone cannot model all appearance changes. To solve this problem, we directly predict the Gaussian color from the MLP. We noticed that spatial embeddings coupled with an MLP produce a smoothing effect on the predicted colors. Hence, to render view-dependent peculiarities and high-frequency details and counterbalance this smoothing effect, we also add the Gaussian color obtained from the SH coefficients as input to the MLP." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 283, + 482, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 283, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 482, + 331 + ], + "type": "text", + "content": "Our final solution employs an MLP " + }, + { + "bbox": [ + 130, + 283, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\theta}" + }, + { + "bbox": [ + 130, + 283, + 482, + 331 + ], + "type": "text", + "content": " that, for a given view, takes as input the color of the 3D Gaussian " + }, + { + "bbox": [ + 130, + 283, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 130, + 283, + 482, + 331 + ], + "type": "text", + "content": ", the embedding vector of its center " + }, + { + "bbox": [ + 130, + 283, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{emb}(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 283, + 482, + 331 + ], + "type": "text", + "content": " and the image embedding " + }, + { + "bbox": [ + 130, + 283, + 482, + 331 + ], + "type": "inline_equation", + "content": "l_{I}" + }, + { + "bbox": [ + 130, + 283, + 482, + 331 + ], + "type": "text", + "content": ", and outputs the image-conditioned color " + }, + { + "bbox": [ + 130, + 283, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{c}^{\\mathbf{I}}" + }, + { + "bbox": [ + 130, + 283, + 482, + 331 + ], + "type": "text", + "content": ". Accordingly, for a 3D Gaussian, the image-dependent color " + }, + { + "bbox": [ + 130, + 283, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{c}^{\\mathbf{I}}" + }, + { + "bbox": [ + 130, + 283, + 482, + 331 + ], + "type": "text", + "content": " is computed as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 256, + 342, + 480, + 356 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 342, + 480, + 356 + ], + "spans": [ + { + "bbox": [ + 256, + 342, + 480, + 356 + ], + "type": "interline_equation", + "content": "\\mathbf {c} ^ {\\mathbf {I}} = \\mathcal {F} _ {\\theta} (\\mathbf {c}, \\mathbf {e m b} (\\mathbf {x}), l _ {I}). \\tag {6}", + "image_path": "f2e099f7d562f60f0dfd8a7722c8bbd21a05cf03ab3924878fe62257b28fbb10.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 371, + 313, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 371, + 313, + 384 + ], + "spans": [ + { + "bbox": [ + 132, + 371, + 313, + 384 + ], + "type": "text", + "content": "4.2 Transient Gaussians Modeling" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 390, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 482, + 449 + ], + "type": "text", + "content": "Our initial results of applying 3DGS to in-the-wild scenarios showed that Gaussians representing transient objects produce blurred and elongated artifacts, due to their optimization across all images, despite transients appearing in only a few frames. Taking this observation into consideration, we introduce a learnable image-dependent opacity variation term " + }, + { + "bbox": [ + 130, + 390, + 482, + 449 + ], + "type": "inline_equation", + "content": "\\Delta \\tilde{\\alpha}^{\\mathbf{I}}" + }, + { + "bbox": [ + 130, + 390, + 482, + 449 + ], + "type": "text", + "content": " to each Gaussian." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 450, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 582 + ], + "type": "text", + "content": "On the one hand, this opacity variation parameter allows Gaussians to reconstruct occluders present in some images. On the other hand, it enables those same Gaussians to be transparent (i.e. not involved in the image rendering process in other images) where these occluders are absent. During training, we want to encourage this variation parameter to either fully mask (i.e. in the case of occluders) or fully maintain (i.e. in the case of the static background) the 3D Gaussians' opacities. For this purpose, we sample " + }, + { + "bbox": [ + 130, + 450, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\Delta \\tilde{\\alpha}^{\\mathbf{I}}" + }, + { + "bbox": [ + 130, + 450, + 482, + 582 + ], + "type": "text", + "content": " using a Binary Concrete random variable [13]. This distribution is a continuous approximation of a Bernoulli distribution concentrating most of the mass on the boundaries of the interval [0, 1]. We return an additional output, " + }, + { + "bbox": [ + 130, + 450, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\Delta \\alpha^{\\mathbf{I}}" + }, + { + "bbox": [ + 130, + 450, + 482, + 582 + ], + "type": "text", + "content": ", of the previously introduced MLP " + }, + { + "bbox": [ + 130, + 450, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\theta}" + }, + { + "bbox": [ + 130, + 450, + 482, + 582 + ], + "type": "text", + "content": " as the location parameter to the concrete function:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 242, + 589, + 480, + 602 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 589, + 480, + 602 + ], + "spans": [ + { + "bbox": [ + 242, + 589, + 480, + 602 + ], + "type": "interline_equation", + "content": "\\left(\\mathbf {c} ^ {\\mathbf {I}}, \\Delta \\alpha^ {\\mathbf {I}}\\right) = \\mathcal {F} _ {\\theta} (\\mathbf {c}, \\mathbf {e m b} (\\mathbf {x}), l _ {I}), \\tag {7}", + "image_path": "387b27611f8da8889d58e712b74da96095a628928c08de2e0bd56eff41e7a296.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 181, + 618, + 480, + 644 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 618, + 480, + 644 + ], + "spans": [ + { + "bbox": [ + 181, + 618, + 480, + 644 + ], + "type": "interline_equation", + "content": "\\Delta \\tilde {\\alpha} ^ {\\mathbf {I}} = \\operatorname {s i g m o i d} \\left[ \\frac {1}{\\mathbf {T}} \\left(l o g \\left(| \\Delta \\alpha^ {\\mathbf {I}} |\\right) + l o g (\\mathbf {U}) - l o g (1 - \\mathbf {U})\\right) \\right], \\tag {8}", + "image_path": "2c2dac0d56d44088bffd1bf51d66b4fd308f2b8c5f50bf390302a180643c6179.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 266, + 654, + 480, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 654, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 266, + 654, + 480, + 666 + ], + "type": "interline_equation", + "content": "\\mathbf {U} \\sim \\operatorname {U n i f o r m} (0, 1), \\tag {9}", + "image_path": "118c869d2b7959d80d11a83f0452c7ef6d6f1ff23bac149574f13cdc1f92ab85.jpg" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "type": "text", + "content": "SWAG" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 167, + 114, + 449, + 225 + ], + "blocks": [ + { + "bbox": [ + 167, + 114, + 449, + 225 + ], + "lines": [ + { + "bbox": [ + 167, + 114, + 449, + 225 + ], + "spans": [ + { + "bbox": [ + 167, + 114, + 449, + 225 + ], + "type": "image", + "image_path": "9a997299c4950565ad329fd41551e9cb2a73cca67526102f006e81af33f1c081.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 233, + 482, + 289 + ], + "lines": [ + { + "bbox": [ + 131, + 233, + 482, + 289 + ], + "spans": [ + { + "bbox": [ + 131, + 233, + 482, + 289 + ], + "type": "text", + "content": "Fig. 3: Variance histogram analysis of the Gaussians' opacity variation " + }, + { + "bbox": [ + 131, + 233, + 482, + 289 + ], + "type": "inline_equation", + "content": "\\Delta \\tilde{\\alpha}^{\\mathrm{I}}" + }, + { + "bbox": [ + 131, + 233, + 482, + 289 + ], + "type": "text", + "content": " w.r.t training images for a dynamic (i.e. containing transient objects) scene (top, Trevi Fountain from Phototourism [6]); and a static scene (bottom, Locomotive from Tanks & Temples [9]). The left column views are rendered using only static Gaussians (i.e having " + }, + { + "bbox": [ + 131, + 233, + 482, + 289 + ], + "type": "inline_equation", + "content": "\\mathbf{Var}\\left[\\Delta \\tilde{\\alpha}^{\\mathrm{I}}\\right] = 0" + }, + { + "bbox": [ + 131, + 233, + 482, + 289 + ], + "type": "text", + "content": "), whereas right column views are rendered using all Gaussians." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 309, + 482, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 309, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 309, + 482, + 369 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 309, + 482, + 369 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 130, + 309, + 482, + 369 + ], + "type": "text", + "content": " is the temperature hyper-parameter. We do not use any annealing on the temperature parameter, rather we set " + }, + { + "bbox": [ + 130, + 309, + 482, + 369 + ], + "type": "inline_equation", + "content": "\\mathbf{T} = 0.1" + }, + { + "bbox": [ + 130, + 309, + 482, + 369 + ], + "type": "text", + "content": " to force, from the very beginning of the training, the opacity variation to match a binary distribution. During the evaluation, we fix " + }, + { + "bbox": [ + 130, + 309, + 482, + 369 + ], + "type": "inline_equation", + "content": "\\mathbf{U}" + }, + { + "bbox": [ + 130, + 309, + 482, + 369 + ], + "type": "text", + "content": " at 0.5. The final image-dependent 3D Gaussian opacity is formulated as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 252, + 373, + 481, + 389 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 373, + 481, + 389 + ], + "spans": [ + { + "bbox": [ + 252, + 373, + 481, + 389 + ], + "type": "interline_equation", + "content": "\\alpha^ {\\mathbf {I}} = \\max \\left(\\alpha - \\Delta \\tilde {\\alpha} ^ {\\mathbf {I}}, 0\\right). \\tag {10}", + "image_path": "e5df743b92b2ac412978a26a51b559233e3d926ff51fe3f7e9580b2ce0ba0395.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 394, + 482, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 394, + 482, + 442 + ], + "spans": [ + { + "bbox": [ + 130, + 394, + 482, + 442 + ], + "type": "text", + "content": "With our formulation of the opacity, we can naturally disentangle between the transient and the static parts of the scene by defining transient Gaussians as the ones having non-zero variance of their associated parameter " + }, + { + "bbox": [ + 130, + 394, + 482, + 442 + ], + "type": "inline_equation", + "content": "\\Delta \\tilde{\\alpha}^{\\mathbf{I}}" + }, + { + "bbox": [ + 130, + 394, + 482, + 442 + ], + "type": "text", + "content": " over training images: " + }, + { + "bbox": [ + 130, + 394, + 482, + 442 + ], + "type": "inline_equation", + "content": "\\operatorname{Var}\\left[\\Delta \\tilde{\\alpha}^{\\mathbf{I}}\\right] \\neq 0" + }, + { + "bbox": [ + 130, + 394, + 482, + 442 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 443, + 483, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 443, + 483, + 587 + ], + "spans": [ + { + "bbox": [ + 130, + 443, + 483, + 587 + ], + "type": "text", + "content": "Figure 3 shows the histogram of " + }, + { + "bbox": [ + 130, + 443, + 483, + 587 + ], + "type": "inline_equation", + "content": "\\mathbf{Var}\\left[\\Delta \\tilde{\\alpha}^{\\mathbf{I}}\\right]" + }, + { + "bbox": [ + 130, + 443, + 483, + 587 + ], + "type": "text", + "content": " on two scenes: one without transient content (Locomotive) and one with occluders (Trevi). We also show two rendered images of each scene: the right image using all the Gaussians and the left one using only Gaussians not detected as transient. The histogram shows that for a static scene, we detected less than " + }, + { + "bbox": [ + 130, + 443, + 483, + 587 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 130, + 443, + 483, + 587 + ], + "type": "text", + "content": " of the Gaussians as transient and around " + }, + { + "bbox": [ + 130, + 443, + 483, + 587 + ], + "type": "inline_equation", + "content": "96\\%" + }, + { + "bbox": [ + 130, + 443, + 483, + 587 + ], + "type": "text", + "content": " of the Gaussians having " + }, + { + "bbox": [ + 130, + 443, + 483, + 587 + ], + "type": "inline_equation", + "content": "\\mathbf{Var}\\left[\\Delta \\tilde{\\alpha}^{\\mathbf{I}}\\right] = 0" + }, + { + "bbox": [ + 130, + 443, + 483, + 587 + ], + "type": "text", + "content": ". Removing the " + }, + { + "bbox": [ + 130, + 443, + 483, + 587 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 130, + 443, + 483, + 587 + ], + "type": "text", + "content": " of transient Gaussians does not affect the rendered image quality. For Trevi Fountain, our model was able to represent transient objects using less than " + }, + { + "bbox": [ + 130, + 443, + 483, + 587 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 443, + 483, + 587 + ], + "type": "text", + "content": " of the Gaussians while maintaining the majority of the Gaussians to reconstruct static parts of the scene. Comparing the two rendered images shows that the transient Gaussians correspond to the mass of tourists present near the fountain, while removing the transient Gaussians resulted in an occluders-free rendering." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 602, + 231, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 602, + 231, + 616 + ], + "spans": [ + { + "bbox": [ + 132, + 602, + 231, + 616 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 624, + 277, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 624, + 277, + 636 + ], + "spans": [ + { + "bbox": [ + 132, + 624, + 277, + 636 + ], + "type": "text", + "content": "5.1 Implementation details" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Similar to 3DGS, we minimize the " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": " loss combined with D-SSIM term to optimize the Gaussians parameters, the MLP " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\theta}" + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": " weights, the Hash Encoder pa" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 241, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 241, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 241, + 101 + ], + "type": "text", + "content": "H. Dahmani et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 138, + 169, + 477, + 244 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 479, + 159 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 479, + 159 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 479, + 159 + ], + "type": "text", + "content": "Table 1: Quantitative results - Results on three real-world scenes from Phototourism [6] and efficiency comparison among SoTA methods for NVS in-the-wild. Given the significant gap between previous baselines [3,7,14,21] and 3DGS [8], we marginalize on the different GPU configurations." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 169, + 477, + 244 + ], + "lines": [ + { + "bbox": [ + 138, + 169, + 477, + 244 + ], + "spans": [ + { + "bbox": [ + 138, + 169, + 477, + 244 + ], + "type": "table", + "html": "
Bradenburg GateSacre CoeurTrevi FountainMean Efficiency
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓Train time (h)FPS
NeRF [15]18.900.8150.23115.600.7150.29116.140.6000.366--
NeRF-W [14]24.170.8900.16719.200.8070.19118.970.6980.265400†<1†
Ha-NeRF [3]24.040.8770.13920.020.8010.17120.180.6900.222452†0.20†
CR-NeRF [21]26.530.9000.10622.070.8230.15221.480.7110.206420†0.25†
RefinedFields [7]26.640.886-22.260.817-23.420.737-150†<1†
3DGS [8]19.990.8890.18017.570.8310.21918.470.7610.2340.50*181*
SWAG (ours)26.330.9290.13921.160.8600.18523.100.8150.2080.83*15.29*
", + "image_path": "60b401daffde20a204406f3424c6f2779b3cf5815acfe5f390debb720b822a84.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 184, + 245, + 428, + 255 + ], + "lines": [ + { + "bbox": [ + 184, + 245, + 428, + 255 + ], + "spans": [ + { + "bbox": [ + 184, + 245, + 428, + 255 + ], + "type": "text", + "content": "† Results reported on [7]. * Results computed using a high-tier GPU." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 281, + 480, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 480, + 354 + ], + "type": "text", + "content": "rameters, and the per-image embedding vectors. We fix the size of the latent embedding vectors to " + }, + { + "bbox": [ + 130, + 281, + 480, + 354 + ], + "type": "inline_equation", + "content": "n = 24" + }, + { + "bbox": [ + 130, + 281, + 480, + 354 + ], + "type": "text", + "content": ". For the Hash encoding parameters, we set the hash table size " + }, + { + "bbox": [ + 130, + 281, + 480, + 354 + ], + "type": "inline_equation", + "content": "T_{h} = 2^{19}" + }, + { + "bbox": [ + 130, + 281, + 480, + 354 + ], + "type": "text", + "content": ", the finest resolution " + }, + { + "bbox": [ + 130, + 281, + 480, + 354 + ], + "type": "inline_equation", + "content": "N_{max} = 2048" + }, + { + "bbox": [ + 130, + 281, + 480, + 354 + ], + "type": "text", + "content": ", the number of levels " + }, + { + "bbox": [ + 130, + 281, + 480, + 354 + ], + "type": "inline_equation", + "content": "L = 12" + }, + { + "bbox": [ + 130, + 281, + 480, + 354 + ], + "type": "text", + "content": " and the coarsest resolution " + }, + { + "bbox": [ + 130, + 281, + 480, + 354 + ], + "type": "inline_equation", + "content": "N_{min} = 16" + }, + { + "bbox": [ + 130, + 281, + 480, + 354 + ], + "type": "text", + "content": ". The MLP we employ has 3 layers of 64 hidden units. Additional details on hyperparameters are provided in the supplementary material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 379, + 204, + 390 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 379, + 204, + 390 + ], + "spans": [ + { + "bbox": [ + 131, + 379, + 204, + 390 + ], + "type": "text", + "content": "5.2 Datasets" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 407, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 482, + 529 + ], + "type": "text", + "content": "Similar to SoTA in-the-wild reconstruction methods, we evaluate SWAG's NVS capabilities using the Phototourism dataset [6] and report our results on three main touristic monuments: Brandenburg Gate, Sacre Coeur, and Trevi Fountain. We also evaluate our method on the benchmark introduced in NeRF-OSR [18] for outdoor scene relighting tasks. We report our performance using NeRF-MS's [11] data split on 4 sites: stjohann (St. Johann, Saarbrücken), lump (Landwehrplatz, Saarbrücken), st (Staatstheater, Saarbrücken), and europa (Galerie Europa, Saarbrücken) and the original NeRF-OSR [18]'s data split on 3 sites: stjohann, lump and " + }, + { + "bbox": [ + 130, + 407, + 482, + 529 + ], + "type": "inline_equation", + "content": "lk2" + }, + { + "bbox": [ + 130, + 407, + 482, + 529 + ], + "type": "text", + "content": " (Ludwigskirche, Saarbrücken). More details about the datasets can be found in supplementary materials." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 553, + 215, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 553, + 215, + 565 + ], + "spans": [ + { + "bbox": [ + 131, + 553, + 215, + 565 + ], + "type": "text", + "content": "5.3 Evaluation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "We provide visual comparisons with rendered images and report quantitative results based on common rendering metrics from the literature: Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS). Since only the embeddings associated with the training images are optimized during training, we follow NeRF-W's [14] evaluation approach: we optimize an embedding on the left half of each test image and report metrics on the right half." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "type": "text", + "content": "SWAG" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 148, + 479, + 213 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "type": "text", + "content": "Table 2: Quantitative results - Results on five sites from NeRF-OSR [18] outdoor scenes benchmark." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 148, + 479, + 213 + ], + "lines": [ + { + "bbox": [ + 136, + 148, + 479, + 213 + ], + "spans": [ + { + "bbox": [ + 136, + 148, + 479, + 213 + ], + "type": "table", + "html": "
stjohannlwpsteuropa
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
NeRF [15]14.890.4320.63911.510.4680.57417.200.5140.50217.490.5510.503
NeRF-W [14]21.230.6670.42619.610.6160.44520.310.6070.43820.000.6990.340
Ha-NeRF [3]17.190.6860.33120.030.6850.36517.300.5380.48317.790.6320.421
NeRF-MS [11]22.840.7930.23521.900.7190.33620.680.6300.40221.030.7210.294
3DGS [8]16.770.7410.26811.760.6090.41417.160.6290.40620.180.7820.252
SWAG (ours)23.910.8640.17222.070.7830.30322.290.7130.36423.740.8450.242
", + "image_path": "17ad782ffa038f885928b18c2b1de29970e30922914d802f2cab8a4d20ae46f1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 138, + 215, + 241, + 223 + ], + "lines": [ + { + "bbox": [ + 138, + 215, + 241, + 223 + ], + "spans": [ + { + "bbox": [ + 138, + 215, + 241, + 223 + ], + "type": "text", + "content": "(a) Reported on NeRF-MS split [11]" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 209, + 224, + 406, + 272 + ], + "blocks": [ + { + "bbox": [ + 209, + 224, + 406, + 272 + ], + "lines": [ + { + "bbox": [ + 209, + 224, + 406, + 272 + ], + "spans": [ + { + "bbox": [ + 209, + 224, + 406, + 272 + ], + "type": "table", + "html": "
lk2stlwp
PSNR ↑SSIM ↑PSNR ↑SSIM ↑PSNR ↑SSIM ↑
NeRF-OSR [18]19.860.62615.830.55617.380.576
3DGS [8]15.910.68712.530.58513.720.659
SWAG (ours)19.590.75618.730.72118.610.753
", + "image_path": "46bae78a1a89f96b9c8fc1d4bf4c367b0894f3e4ae7ac0199797329e981f95c5.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 211, + 273, + 395, + 282 + ], + "lines": [ + { + "bbox": [ + 211, + 273, + 395, + 282 + ], + "spans": [ + { + "bbox": [ + 211, + 273, + 395, + 282 + ], + "type": "text", + "content": "(b) Reported on original benchmark split from NeRF-OSR [18]" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 305, + 198, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 305, + 198, + 316 + ], + "spans": [ + { + "bbox": [ + 132, + 305, + 198, + 316 + ], + "type": "text", + "content": "5.4 Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 326, + 481, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 326, + 481, + 469 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 481, + 469 + ], + "type": "text", + "content": "Table 1 shows quantitative results on three Phototourism scenes. Unsurprisingly, optimizing 3DGS on in-the-wild photo collections showed poor results due to the lack of image appearance dependency. The 3D Gaussians colors converge to a mean color which makes the method unable to model visual changes caused by weather, lightning conditions and different camera specifications. The presence of transient objects also resulted in the persistence of artifacts. SWAG improves 3DGS performance and outperforms the baselines on SSIM across all datasets. In particular, our proposal improves 3DGS's PSNR by an average margin of 5.01 dB. SWAG outperforms NeRF-W across all datasets in terms of LPIPS, PSNR, and SSIM and shows competitive performance with CR-NeRF and RefinedFields despite the priors these two methods use and requiring significantly longer times to train." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 470, + 481, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 470, + 481, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 470, + 481, + 506 + ], + "type": "text", + "content": "Table 2 shows quantitative results on NeRF-OSR [18] benchmark. SWAG improves 3DGS performance by an average PSNR margin of 5 dB and outperformed all the other baselines regardless of the train/test split choices." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 506, + 481, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 506, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 506, + 481, + 601 + ], + "type": "text", + "content": "Figure 4 and Figure 5 show qualitative results of 3DGS and SWAG on Phototourism scenes and four of NeRF-OSR sites. In some parts of the scene, 3DGS presents artifacts and floaters either in under-observed part of the scene or in areas occluded by transient objects. Adding to that, the rendered color is shifted from the ground truth color as shown in Figure 4. SWAG produces visual appearances close to the ground truth thanks to our per-image conditioning. Compared to previous implicit methods [3,7,14,21], our proposal achieves one order of magnitude faster training and guarantees real-time rendering as shown in Table 1." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 620, + 284, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 620, + 284, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 620, + 284, + 633 + ], + "type": "text", + "content": "5.5 Controllable appearance" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Modeling the image appearance with an embedding vector enabled us to learn a latent space that we can utilize to change the lighting and appearance of any" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "type": "text", + "content": "H. Dahmani et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 161, + 113, + 452, + 463 + ], + "blocks": [ + { + "bbox": [ + 161, + 113, + 452, + 463 + ], + "lines": [ + { + "bbox": [ + 161, + 113, + 452, + 463 + ], + "spans": [ + { + "bbox": [ + 161, + 113, + 452, + 463 + ], + "type": "image", + "image_path": "e82fd0dbe4b928068f725fc5a60468d6d52046dcd082afd9fea9e03be92b981c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 471, + 480, + 495 + ], + "lines": [ + { + "bbox": [ + 132, + 471, + 480, + 495 + ], + "spans": [ + { + "bbox": [ + 132, + 471, + 480, + 495 + ], + "type": "text", + "content": "Fig. 4: Qualitative experimental results on three real-world scenes from Phototourism [6].s" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 517, + 482, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 517, + 482, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 517, + 482, + 578 + ], + "type": "text", + "content": "viewpoint at inference time. On top of the training image embeddings shown in Figure 6, we can sample vectors from the appearance space using interpolation between the learned embeddings. These interpolations are smooth and natural as shown in Figure 7. We encourage the readers to explore the supplementary videos and appreciate more the naturalness of these interpolations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 597, + 297, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 597, + 297, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 597, + 297, + 609 + ], + "type": "text", + "content": "5.6 Transient Objects Removal" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "As explained in Section 4.2, our method is able to localize 3D Gaussians that represent transient objects and, consequently, is able to manipulate the scenes omitting these objects. It should be noted that SWAG is thoughtfully designed to model transient objects without allowing the Gaussians to vary their opacity" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "type": "text", + "content": "SWAG" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 114, + 473, + 485 + ], + "blocks": [ + { + "bbox": [ + 140, + 114, + 473, + 485 + ], + "lines": [ + { + "bbox": [ + 140, + 114, + 473, + 485 + ], + "spans": [ + { + "bbox": [ + 140, + 114, + 473, + 485 + ], + "type": "image", + "image_path": "c90e427c1332e0d4ceb175188e42164b82b55b2c231a4d105ce3a43b117b5af6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 494, + 482, + 528 + ], + "lines": [ + { + "bbox": [ + 132, + 494, + 482, + 528 + ], + "spans": [ + { + "bbox": [ + 132, + 494, + 482, + 528 + ], + "type": "text", + "content": "Fig. 5: Qualitative experimental results on three real-world scenes from Phototourism [6] and four NeRF-OSR [18] scenes. We demonstrate the capability of SWAG to disentangle between static and transient parts of the scene." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "to compensate for appearance variation. Indeed, according to our experiments, it was observed that less than " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " (average on trained scenes) of the Gaussians capture transient objects. Consequently, SWAG achieves a better reconstruction of the static scene parts without having blurry elongated Gaussians that arise in view-dependent appearances with 3DGS. The capabilities of our model of disentangling static and dynamic Gaussians are clearly highlighted in Figure 5." + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "type": "text", + "content": "H. Dahmani et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 167, + 114, + 219, + 182 + ], + "blocks": [ + { + "bbox": [ + 167, + 114, + 219, + 182 + ], + "lines": [ + { + "bbox": [ + 167, + 114, + 219, + 182 + ], + "spans": [ + { + "bbox": [ + 167, + 114, + 219, + 182 + ], + "type": "image", + "image_path": "acbb1c6965f66d14cad980907fba8c253cf8bebe84a09c9fd18de3192492ef07.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 168, + 182, + 218, + 215 + ], + "blocks": [ + { + "bbox": [ + 168, + 182, + 218, + 215 + ], + "lines": [ + { + "bbox": [ + 168, + 182, + 218, + 215 + ], + "spans": [ + { + "bbox": [ + 168, + 182, + 218, + 215 + ], + "type": "image", + "image_path": "e0c574d11ffecd5a20135999fc0df1771463d1cbec21e169a9a867f76a649804.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 224, + 480, + 257 + ], + "lines": [ + { + "bbox": [ + 130, + 224, + 480, + 257 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 480, + 257 + ], + "type": "text", + "content": "Fig.6: Appearance transfer - SWAG is able to render a scene at any viewpoint with the visual appearance of any training image thanks to the learned appearance embeddings." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 219, + 114, + 292, + 215 + ], + "blocks": [ + { + "bbox": [ + 219, + 114, + 292, + 215 + ], + "lines": [ + { + "bbox": [ + 219, + 114, + 292, + 215 + ], + "spans": [ + { + "bbox": [ + 219, + 114, + 292, + 215 + ], + "type": "image", + "image_path": "631bad8b5a26cc9900af96c60694257fb2116a722b070210293c8029562043b9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 293, + 114, + 446, + 215 + ], + "blocks": [ + { + "bbox": [ + 293, + 114, + 446, + 215 + ], + "lines": [ + { + "bbox": [ + 293, + 114, + 446, + 215 + ], + "spans": [ + { + "bbox": [ + 293, + 114, + 446, + 215 + ], + "type": "image", + "image_path": "9628c916e29ffac9cf74c6214c98113f225c6d6ed915f78723950da8e48ee47d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 133, + 267, + 191, + 310 + ], + "blocks": [ + { + "bbox": [ + 133, + 267, + 191, + 310 + ], + "lines": [ + { + "bbox": [ + 133, + 267, + 191, + 310 + ], + "spans": [ + { + "bbox": [ + 133, + 267, + 191, + 310 + ], + "type": "image", + "image_path": "1d7997630707f1bea131c6ebe20ecc8153100852c13d728fd4120c412d9de931.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 133, + 311, + 191, + 353 + ], + "blocks": [ + { + "bbox": [ + 133, + 311, + 191, + 353 + ], + "lines": [ + { + "bbox": [ + 133, + 311, + 191, + 353 + ], + "spans": [ + { + "bbox": [ + 133, + 311, + 191, + 353 + ], + "type": "image", + "image_path": "6ee3410d2010e59f2474a98013a597daa0d647378c3cfcb4c9b90d9525391646.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 373, + 479, + 395 + ], + "lines": [ + { + "bbox": [ + 130, + 373, + 479, + 395 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 479, + 395 + ], + "type": "text", + "content": "Fig. 7: Appearance interpolation - Interpolations between the appearance embeddings of two training images (left, right)." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 194, + 267, + 252, + 310 + ], + "blocks": [ + { + "bbox": [ + 194, + 267, + 252, + 310 + ], + "lines": [ + { + "bbox": [ + 194, + 267, + 252, + 310 + ], + "spans": [ + { + "bbox": [ + 194, + 267, + 252, + 310 + ], + "type": "image", + "image_path": "2bd9880599c5f4f28b85add60b889955c0addadae12ecdeab86decdaead1cf2a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 194, + 311, + 252, + 353 + ], + "blocks": [ + { + "bbox": [ + 194, + 311, + 252, + 353 + ], + "lines": [ + { + "bbox": [ + 194, + 311, + 252, + 353 + ], + "spans": [ + { + "bbox": [ + 194, + 311, + 252, + 353 + ], + "type": "image", + "image_path": "2da72133034dd24738375d1d9f43669ae824ac3370b995d6800e99cef210b2d5.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 253, + 267, + 309, + 310 + ], + "blocks": [ + { + "bbox": [ + 253, + 267, + 309, + 310 + ], + "lines": [ + { + "bbox": [ + 253, + 267, + 309, + 310 + ], + "spans": [ + { + "bbox": [ + 253, + 267, + 309, + 310 + ], + "type": "image", + "image_path": "99527152e06fc4f43eefac896bd0b6da6c56b2b23ce581f2156ecf33bfa32fcf.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 253, + 311, + 309, + 353 + ], + "blocks": [ + { + "bbox": [ + 253, + 311, + 309, + 353 + ], + "lines": [ + { + "bbox": [ + 253, + 311, + 309, + 353 + ], + "spans": [ + { + "bbox": [ + 253, + 311, + 309, + 353 + ], + "type": "image", + "image_path": "ed22996c2f5e7ad9109f2472e7da4d4913347a73254fe97704d95c1b0fdcc188.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 309, + 267, + 422, + 310 + ], + "blocks": [ + { + "bbox": [ + 309, + 267, + 422, + 310 + ], + "lines": [ + { + "bbox": [ + 309, + 267, + 422, + 310 + ], + "spans": [ + { + "bbox": [ + 309, + 267, + 422, + 310 + ], + "type": "image", + "image_path": "2bdd6a400cd7870e89bcc5dc7b87f115df0b617d4832298b19e451abb304477f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 309, + 311, + 422, + 353 + ], + "blocks": [ + { + "bbox": [ + 309, + 311, + 422, + 353 + ], + "lines": [ + { + "bbox": [ + 309, + 311, + 422, + 353 + ], + "spans": [ + { + "bbox": [ + 309, + 311, + 422, + 353 + ], + "type": "image", + "image_path": "017dfaa40d47e50c5c982ecd9e0d46f36d262779126476a2b27a93e3488c1823.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 425, + 268, + 482, + 310 + ], + "blocks": [ + { + "bbox": [ + 425, + 268, + 482, + 310 + ], + "lines": [ + { + "bbox": [ + 425, + 268, + 482, + 310 + ], + "spans": [ + { + "bbox": [ + 425, + 268, + 482, + 310 + ], + "type": "image", + "image_path": "561fec92faf15f1c7387a9fe13b3c93dbfae955d50088c4e67c4b093ef410c46.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 425, + 311, + 482, + 353 + ], + "blocks": [ + { + "bbox": [ + 425, + 311, + 482, + 353 + ], + "lines": [ + { + "bbox": [ + 425, + 311, + 482, + 353 + ], + "spans": [ + { + "bbox": [ + 425, + 311, + 482, + 353 + ], + "type": "image", + "image_path": "20e059e50e3f88b63759b1cd225c410eb256b4d7f8a51c1c00114296dfc4b69e.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 131, + 418, + 209, + 429 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 418, + 209, + 429 + ], + "spans": [ + { + "bbox": [ + 131, + 418, + 209, + 429 + ], + "type": "text", + "content": "5.7 Ablations" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 437, + 479, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 479, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 479, + 461 + ], + "type": "text", + "content": "In this section, we compare two variants of our method to analyze the contribution of each component of SWAG:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 137, + 468, + 480, + 515 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 137, + 468, + 480, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 468, + 480, + 491 + ], + "spans": [ + { + "bbox": [ + 137, + 468, + 480, + 491 + ], + "type": "text", + "content": "- SWAG-A, a variation wherein the transient detection part of our model is removed," + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 137, + 492, + 480, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 492, + 480, + 515 + ], + "spans": [ + { + "bbox": [ + 137, + 492, + 480, + 515 + ], + "type": "text", + "content": "- SWAG-T, a variation wherein the image appearance color dependency is omitted." + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 523, + 479, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 523, + 479, + 548 + ], + "spans": [ + { + "bbox": [ + 130, + 523, + 479, + 548 + ], + "type": "text", + "content": "We compare the results of our model variation on Phototourism scenes and we report the results in Table 3. SWAG-A enabled appearance modeling and" + } + ] + } + ], + "index": 23 + }, + { + "type": "table", + "bbox": [ + 158, + 601, + 452, + 663 + ], + "blocks": [ + { + "bbox": [ + 130, + 567, + 479, + 590 + ], + "lines": [ + { + "bbox": [ + 130, + 567, + 479, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 567, + 479, + 590 + ], + "type": "text", + "content": "Table 3: Quantitative results - Results of two ablations of SWAG : SWAG-A and SWAG-T on three real-world scenes from Phototourism [6]." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 158, + 601, + 452, + 663 + ], + "lines": [ + { + "bbox": [ + 158, + 601, + 452, + 663 + ], + "spans": [ + { + "bbox": [ + 158, + 601, + 452, + 663 + ], + "type": "table", + "html": "
Bradenburg GateSacre CoeurTrevi Fountain
PSNR ↑SSIM↑LPIPS ↓PSNR ↑SSIM↑LPIPS ↓PSNR ↑SSIM↑LPIPS ↓
SWAG26.330.9290.13921.160.8600.18523.100.8150.208
SWAG-A26.440.9280.14121.020.8550.18322.350.8010.223
SWAG-T24.460.9140.16019.060.8360.20520.130.7690.244
", + "image_path": "4486a64be06425f696e2236e67f86e5274b3b132ede0082dec9167fe3c26e8ed.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "table_body" + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "type": "text", + "content": "SWAG" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 170, + 114, + 223, + 154 + ], + "blocks": [ + { + "bbox": [ + 170, + 114, + 223, + 154 + ], + "lines": [ + { + "bbox": [ + 170, + 114, + 223, + 154 + ], + "spans": [ + { + "bbox": [ + 170, + 114, + 223, + 154 + ], + "type": "image", + "image_path": "8fda2fcf75e21be26ef7293d5b450fb61dffb01ed073fb9598f8aa3b4c2eaec9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 172, + 158, + 221, + 166 + ], + "lines": [ + { + "bbox": [ + 172, + 158, + 221, + 166 + ], + "spans": [ + { + "bbox": [ + 172, + 158, + 221, + 166 + ], + "type": "text", + "content": "Ground-truth" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 225, + 114, + 278, + 154 + ], + "blocks": [ + { + "bbox": [ + 225, + 114, + 278, + 154 + ], + "lines": [ + { + "bbox": [ + 225, + 114, + 278, + 154 + ], + "spans": [ + { + "bbox": [ + 225, + 114, + 278, + 154 + ], + "type": "image", + "image_path": "5c9f1b066b14a3bbb46bac382bc7b56fb5cc5e5beea7d5b30d1439bd2c8775e9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 241, + 158, + 263, + 166 + ], + "lines": [ + { + "bbox": [ + 241, + 158, + 263, + 166 + ], + "spans": [ + { + "bbox": [ + 241, + 158, + 263, + 166 + ], + "type": "text", + "content": "3DGS" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 174, + 480, + 196 + ], + "lines": [ + { + "bbox": [ + 131, + 174, + 480, + 196 + ], + "spans": [ + { + "bbox": [ + 131, + 174, + 480, + 196 + ], + "type": "text", + "content": "Fig. 8: Ablation study - Visual comparison of 3DGS, SWAG, and the two variants used in our ablation study: SWAG-T (no appearance) and SWAG-A (no transient)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 280, + 114, + 333, + 154 + ], + "blocks": [ + { + "bbox": [ + 280, + 114, + 333, + 154 + ], + "lines": [ + { + "bbox": [ + 280, + 114, + 333, + 154 + ], + "spans": [ + { + "bbox": [ + 280, + 114, + 333, + 154 + ], + "type": "image", + "image_path": "d82228f05ff0e38d5bf9a80750aad4dec8b8b9790b77fc7ffe1199d04e88f2a5.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 290, + 158, + 323, + 166 + ], + "lines": [ + { + "bbox": [ + 290, + 158, + 323, + 166 + ], + "spans": [ + { + "bbox": [ + 290, + 158, + 323, + 166 + ], + "type": "text", + "content": "SWAG-A" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 335, + 114, + 388, + 154 + ], + "blocks": [ + { + "bbox": [ + 335, + 114, + 388, + 154 + ], + "lines": [ + { + "bbox": [ + 335, + 114, + 388, + 154 + ], + "spans": [ + { + "bbox": [ + 335, + 114, + 388, + 154 + ], + "type": "image", + "image_path": "f4e3c9e218e9ce4916d2a4f665366da6c32ef0a67f1f8469b42bfb52f07f6f4e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 158, + 378, + 166 + ], + "lines": [ + { + "bbox": [ + 345, + 158, + 378, + 166 + ], + "spans": [ + { + "bbox": [ + 345, + 158, + 378, + 166 + ], + "type": "text", + "content": "SWAG-T" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 390, + 114, + 443, + 154 + ], + "blocks": [ + { + "bbox": [ + 390, + 114, + 443, + 154 + ], + "lines": [ + { + "bbox": [ + 390, + 114, + 443, + 154 + ], + "spans": [ + { + "bbox": [ + 390, + 114, + 443, + 154 + ], + "type": "image", + "image_path": "4584dc1a5a2997d25bec2b9aabcf1d3eee0754215b2370d64380e4344e86c65c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 404, + 158, + 429, + 166 + ], + "lines": [ + { + "bbox": [ + 404, + 158, + 429, + 166 + ], + "spans": [ + { + "bbox": [ + 404, + 158, + 429, + 166 + ], + "type": "text", + "content": "SWAG" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 219, + 480, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 219, + 480, + 301 + ], + "spans": [ + { + "bbox": [ + 130, + 219, + 480, + 301 + ], + "type": "text", + "content": "closer colors to ground truth compared to 3DGS where colors shift away and converge to an average color for all training images. Transient objects' presence still causes floaters in the scene alike 3DGS as shown in Figure 8. In contrast, SWAG-T enables the elimination of these floaters and produces an occluders-free scene while suffering from color alterations compared to ground truth like 3DGS. Combining these two variations enables SWAG to model varying appearances with the ability to reduce scene floaters." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 319, + 212, + 330 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 319, + 212, + 330 + ], + "spans": [ + { + "bbox": [ + 131, + 319, + 212, + 330 + ], + "type": "text", + "content": "5.8 Discussion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 337, + 480, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 337, + 480, + 517 + ], + "spans": [ + { + "bbox": [ + 130, + 337, + 480, + 517 + ], + "type": "text", + "content": "SWAG leverages an MLP, a hash grid encoder, and image embedding to adapt 3DGS to in-the-wild scenarios. While this empowered 3DGS's reconstruction capabilities in these conditions as it was demonstrated in our experiments, it almost doubles the training time and has 10 times longer inference time per frame as shown Table 1 compared to 3DGS. Nonetheless, our method still requires significantly less training time than all previous in-the-wild baselines and achieves interactive rendering speed. Although SWAG succeeded in removing occluders using transient Gaussians, some parts of the scene where transient objects' presence is more frequent (eg. clouds in the sky and tourists on the roads) can suffer from being detected as transient. Eliminating all the transient Gaussians can leave some holes in these parts of the scene as seen in Figure 5 for the Trevi Fountain scene where some parts of the fountain contain black spots. One way to solve this is to remove Gaussians based on an adaptive threshold " + }, + { + "bbox": [ + 130, + 337, + 480, + 517 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 337, + 480, + 517 + ], + "type": "text", + "content": " rather than considering as static only Gaussian with zero variance: " + }, + { + "bbox": [ + 130, + 337, + 480, + 517 + ], + "type": "inline_equation", + "content": "\\mathbf{Var}\\left[\\Delta \\tilde{\\alpha}^{\\mathbf{I}}\\right] < \\lambda" + }, + { + "bbox": [ + 130, + 337, + 480, + 517 + ], + "type": "text", + "content": ", but it requires manual tuning of this threshold " + }, + { + "bbox": [ + 130, + 337, + 480, + 517 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 337, + 480, + 517 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 131, + 533, + 219, + 546 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 533, + 219, + 546 + ], + "spans": [ + { + "bbox": [ + 131, + 533, + 219, + 546 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 557, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 481, + 665 + ], + "type": "text", + "content": "In this paper, we presented SWAG, a method designed for tailoring 3DGS representations to in-the-wild scenarios. SWAG incorporates appearance modeling in the Gaussians' colors and employs an adaptive opacity modulation to handle the presence of transient objects. Extensive experiments demonstrate that SWAG achieves state-of-the-art results on two challenging benchmarks while exhibiting training times orders of magnitude faster than in-the-wild NVS baselines while enabling real-time rendering. As a first step in conditioning 3DGS for in-the-wild scene representations, this work suggests potential future research direction, such as extending SWAG to dynamic scenes." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "type": "text", + "content": "H. Dahmani et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 141, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 138, + 141, + 480, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 141, + 480, + 175 + ], + "spans": [ + { + "bbox": [ + 138, + 141, + 480, + 175 + ], + "type": "text", + "content": "1. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. ICCV (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 175, + 480, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 175, + 480, + 208 + ], + "spans": [ + { + "bbox": [ + 138, + 175, + 480, + 208 + ], + "type": "text", + "content": "2. Botsch, M., Hornung, A., Zwicker, M., Kobbelt, L.: High-quality surface splatting on today's gpus. In: Proceedings Eurographics/IEEE VGTC Symposium Point-Based Graphics (2005)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 209, + 480, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 209, + 480, + 231 + ], + "spans": [ + { + "bbox": [ + 138, + 209, + 480, + 231 + ], + "type": "text", + "content": "3. Chen, X., Zhang, Q., Li, X., Chen, Y., Feng, Y., Wang, X., Wang, J.: Hallucinated neural radiance fields in the wild. In: CVPR. pp. 12943-12952 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 232, + 480, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 232, + 480, + 253 + ], + "spans": [ + { + "bbox": [ + 138, + 232, + 480, + 253 + ], + "type": "text", + "content": "4. Fridovich-Keil, S., Meanti, G., Warburg, F.R., Recht, B., Kanazawa, A.: K-planes: Explicit radiance fields in space, time, and appearance. In: CVPR (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 254, + 405, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 254, + 405, + 265 + ], + "spans": [ + { + "bbox": [ + 138, + 254, + 405, + 265 + ], + "type": "text", + "content": "5. Grossman, J., Sc, H., Dally, W.: Point sample rendering (1999)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 266, + 480, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 266, + 480, + 297 + ], + "spans": [ + { + "bbox": [ + 138, + 266, + 480, + 297 + ], + "type": "text", + "content": "6. Jin, Y., Mishkin, D., Mishchuk, A., Matas, J., Fua, P., Yi, K.M., Trulls, E.: Image matching across wide baselines: From paper to practice. International Journal of Computer Vision p. 517-547 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 298, + 480, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 298, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 138, + 298, + 480, + 319 + ], + "type": "text", + "content": "7. Kassab, K., Schnepf, A., Franceschi, J.Y., Caraffa, L., Mary, J., Gouet-Brunet, V.: Refinedfields: Radiance fields refinement for unconstrained scenes (2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 321, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 321, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 321, + 480, + 342 + ], + "type": "text", + "content": "8. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 343, + 480, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 343, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 138, + 343, + 480, + 365 + ], + "type": "text", + "content": "9. Knapitsch, A., Park, J., Zhou, Q.Y., Koltun, V.: Tanks and temples. ACM Transactions on Graphics pp. 1-13 (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 365, + 480, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 365, + 480, + 387 + ], + "spans": [ + { + "bbox": [ + 138, + 365, + 480, + 387 + ], + "type": "text", + "content": "10. Lassner, C.: Fast differentiable raycasting for neural rendering using sphere-based representations. CoRR (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 388, + 480, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 388, + 480, + 409 + ], + "spans": [ + { + "bbox": [ + 138, + 388, + 480, + 409 + ], + "type": "text", + "content": "1. Li, P., Wang, S., Yang, C., Liu, B., Qiu, W., Wang, H.: Nerf-ms: Neural radiance fields with multi-sequence. In: ICCV. pp. 18591-18600 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 410, + 480, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 410, + 480, + 432 + ], + "spans": [ + { + "bbox": [ + 138, + 410, + 480, + 432 + ], + "type": "text", + "content": "2. Lin, J., Li, Z., Tang, X., Liu, J., Liu, S., Liu, J., Lu, Y., Wu, X., Xu, S., Yan, Y., Yang, W.: Vastgaussian: Vast 3d gaussians for large scene reconstruction (2024)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 433, + 480, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 433, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 138, + 433, + 480, + 464 + ], + "type": "text", + "content": "3. Maddison, C.J., Mnih, A., Teh, Y.W.: The concrete distribution: A continuous relaxation of discrete random variables. In: International Conference on Learning Representations (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 465, + 480, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 465, + 480, + 498 + ], + "spans": [ + { + "bbox": [ + 138, + 465, + 480, + 498 + ], + "type": "text", + "content": "4. Martin-Brualla, R., Radwan, N., Sajjadi, M.S.M., Barron, J.T., Dosovitskiy, A., Duckworth, D.: NeRF in the Wild: Neural Radiance Fields for Unconstrained Photo Collections. In: CVPR (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 498, + 480, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 498, + 480, + 531 + ], + "spans": [ + { + "bbox": [ + 138, + 498, + 480, + 531 + ], + "type": "text", + "content": "5. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 532, + 480, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 532, + 480, + 553 + ], + "spans": [ + { + "bbox": [ + 138, + 532, + 480, + 553 + ], + "type": "text", + "content": "6. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. CoRR (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 554, + 480, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 554, + 480, + 575 + ], + "spans": [ + { + "bbox": [ + 138, + 554, + 480, + 575 + ], + "type": "text", + "content": "7. Rematas, K., Liu, A., Srinivasan, P.P., Barron, J.T., Tagliasacchi, A., Funkhouser, T., Ferrari, V.: Urban radiance fields. CVPR (2022)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 576, + 480, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 576, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 138, + 576, + 480, + 597 + ], + "type": "text", + "content": "8. Rudnev, V., Elgharib, M., Smith, W., Liu, L., Golyanik, V., Theobalt, C.: Nerf for outdoor scene relighting. In: ECCV (2022)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 599, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 599, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 138, + 599, + 480, + 620 + ], + "type": "text", + "content": "9. Turki, H., Ramanan, D., Satyanarayanan, M.: Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. In: CVPR. pp. 12922–12931 (2022)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 138, + 621, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 480, + 643 + ], + "type": "text", + "content": "20. Yan, Z., Low, W.F., Chen, Y., Lee, G.H.: Multi-scale 3d gaussian splatting for anti-aliased rendering (2023)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "21. Yang, Y., Zhang, S., Huang, Z., Zhang, Y., Tan, M.: Cross-ray neural radiance fields for novel-view synthesis from unconstrained image collections. In: ICCV (2023)" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "type": "text", + "content": "SWAG" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 172 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "22. Yu, Z., Chen, A., Huang, B., Sattler, T., Geiger, A.: Mip-splatting: Alias-free 3d gaussian splatting (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 480, + 172 + ], + "type": "text", + "content": "23. Zwicker, M., Pfister, H., van Baar, J., Gross, M.: Surface splattering. In: Proceedings of the 28th Annual Conference on Computer Graphics and Interactive Techniques. p. 371-378 (2001)" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "type": "text", + "content": "H. Dahmani et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/SWinGS_ Sliding Windows for Dynamic 3D Gaussian Splatting/1089e742-f963-4d64-a493-ab6447fb05de_content_list.json b/2024/SWinGS_ Sliding Windows for Dynamic 3D Gaussian Splatting/1089e742-f963-4d64-a493-ab6447fb05de_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..db3bee309b83aedce5a2b61796a494ae9a33fd9a --- /dev/null +++ b/2024/SWinGS_ Sliding Windows for Dynamic 3D Gaussian Splatting/1089e742-f963-4d64-a493-ab6447fb05de_content_list.json @@ -0,0 +1,1584 @@ +[ + { + "type": "text", + "text": "SWinGS: Sliding Windows for Dynamic 3D Gaussian Splatting", + "text_level": 1, + "bbox": [ + 316, + 140, + 686, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Richard Shaw1, Michal Nazarczuk1, Jifei Song1, Arthur Moreau1, Sibi Catley-Chandar1,2, Helisa Dhamo1, and Eduardo Pérez-Pellitero1", + "bbox": [ + 250, + 210, + 754, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Huawei Noah's Ark Lab \n $^{2}$ Queen Mary University Of London", + "bbox": [ + 375, + 253, + 624, + 282 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/adfaab72d4811650a89fe9f926206654dce088cc185571b363b4081714d95df5.jpg", + "image_caption": [ + "Fig. 1: Left: SWinGS achieves sharper dynamic 3D scene reconstruction in part thanks to a sliding window canonical space that reduces the complexity of the 3D motion estimation. Right: Our dynamic real-time viewer allows users to explore the scene." + ], + "image_footnote": [], + "bbox": [ + 220, + 325, + 395, + 444 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e06659d5361df453a0466df373466c8ce4c5824e1fb721ac814c62ac8f6c29e9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 395, + 325, + 565, + 443 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/868ae6435617946c74c592d7a37e5513eb5e35aae3b9b87cf619e5d183643913.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 576, + 325, + 779, + 443 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Novel view synthesis has shown rapid progress recently, with methods capable of producing increasingly photorealistic results. 3D Gaussian Splatting has emerged as a promising method, producing high-quality renderings of scenes and enabling interactive viewing at real-time frame rates. However, it is limited to static scenes. In this work, we extend 3D Gaussian Splatting to reconstruct dynamic scenes. We model a scene's dynamics using dynamic MLPs, learning deformations from temporally-local canonical representations to per-frame 3D Gaussians. To disentangle static and dynamic regions, tuneable parameters weigh each Gaussian's respective MLP parameters, improving the dynamics modelling of imbalanced scenes. We introduce a sliding window training strategy that partitions the sequence into smaller manageable windows to handle arbitrary length scenes while maintaining high rendering quality. We propose an adaptive sampling strategy to determine appropriate window size hyperparameters based on the scene's motion, balancing training overhead with visual quality. Training a separate dynamic 3D Gaussian model for each sliding window allows the canonical representation to change, enabling the reconstruction of scenes with significant geometric changes. Temporal consistency is enforced using a fine-tuning step with self-supervising consistency loss on randomly sampled novel", + "bbox": [ + 261, + 561, + 740, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "views. As a result, our method produces high-quality renderings of general dynamic scenes with competitive quantitative performance, which can be viewed in real-time in our dynamic interactive viewer.", + "bbox": [ + 259, + 147, + 743, + 189 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 209, + 375, + 224 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Photorealistic rendering and generally 3-dimensional (3D) imaging have received significant attention in recent years, especially since the seminal work of Neural Radiance Fields (NeRF) [37]. This is in part thanks to its impressive novel view synthesis results, but also due to its appealing ease of use when coupled with off-the-shelf structure-from-motion camera pose estimation [53]. NeRF's key insight is a fully differentiable volumetric rendering pipeline paired with learnable implicit functions that model a view-dependent 3D radiance field. Dense coverage of posed images of the scene provides then direct photometric supervision.", + "bbox": [ + 212, + 234, + 785, + 356 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The original formulation of NeRF and most follow-ups [5,9,37,39,41] assume static scenes and thus a fixed radiance field. Some have explored new paradigms enabling dynamic reconstruction for radiance fields, including D-NeRF [22] and Nerfies [43], optimising an additional continuous volumetric deformation field that warps each observed point into a canonical NeRF. Such an approach has been popular [12,13,28,32,33,44], and has also been used for dynamic human reconstruction [45,74]. However, learning 3D deformation fields is inherently challenging, especially for large motions, with increased computational expense in training and inference. Moreover, approaches that share a canonical space among all frames struggle to maintain reconstruction quality for long sequences, obtaining overly blurred results due to inaccurate deformations and limited representational capacity. Other methods avoid maintaining a canonical representation and use explicit per-frame representations of the dynamic scene. Examples are tri-plane extensions to 4D $(x,y,z,t)$ [7,14,54,60] with plane decompositions [9] to keep memory footprint under control. These approaches can suffer from a lack of temporal consistency, especially as they generally are agnostic about the motion of the scene. Grid-based methods that share a representation [46], also can suffer from degradation due to lack of representational capacity in long sequences.", + "bbox": [ + 212, + 356, + 785, + 628 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This paper proposes a new method that addresses open problems of the state-of-the-art (SoTA) - see Fig. 1. Our method overview is shown in Fig. 2. Firstly, we build upon 3D Gaussian Splatting (3DGS) [20] and adapt the 3D Gaussians to be dynamic by allowing them to move. Our representation is thus explicit and avoids expensive raymarching via fast rasterization. Secondly, we introduce a novel paradigm for dynamic neural rendering with temporally-local canonical spaces defined in a sliding window fashion. Each window's length is adaptively defined following the amount of scene motion to maintain high-quality reconstruction. By limiting the scope of each canonical space, we can accurately track 3D displacements (i.e. they are generally smaller displacements) and prevent intra-window flickering. Thirdly, we introduce tuneable MLPs (MLP with several sets of weights governed by per-3D Gaussian blending weights) to estimate displacements. This tackles scenes with static vs dynamic imbalance. By learning different \"modes\" of motion estimation, we can separate smoothly between", + "bbox": [ + 212, + 628, + 785, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "R. Shaw et al.", + "bbox": [ + 271, + 114, + 367, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "static and dynamic regions with virtually no additional computational cost nor any handcrafted heuristics. Lastly, temporal consistency loss computed on overlapping frames of neighbouring windows ensures consistency between windows, i.e. avoids inter-window flickering. In summary, our main contributions are:", + "bbox": [ + 212, + 146, + 787, + 207 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. An adaptive sliding window approach that enables the reconstruction of arbitrary length sequences whilst maintaining high rendering quality.", + "2. Temporally-local dynamic MLPs that model scene dynamics by learning deformation fields from per-window canonical 3D Gaussians to each frame.", + "3. Learnable MLP tuning parameters tackle scene imbalance by learning different motion modes; disentangling static canonical and dynamic 3D Gaussians.", + "4. A fine-tuning stage ensures temporal consistency throughout the sequence." + ], + "bbox": [ + 222, + 214, + 785, + 319 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related work", + "text_level": 1, + "bbox": [ + 215, + 340, + 382, + 357 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Non-NeRF dynamic reconstruction: A number of approaches prior to the emergence of NeRF [37] tackled dynamic scene reconstruction. Such methods typically relied on dense camera coverage for point tracking and reprojection [19], or the presence of additional measurements, i.e. depth [40, 55]. Alternatively, some were curated towards specific domains, e.g. car reconstruction in driving scenarios [6, 34]. Several recent works [3,29,46] followed the idea of Image-Based Rendering [63] (direct reconstruction from neighbouring views). Others [30, 69] utilize multiplane images [58, 76] with an additional temporal component.", + "bbox": [ + 212, + 371, + 787, + 492 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "NeRF-based reconstruction: NeRF [37] has achieved great success for reconstructing static scenes with many works extending it to dynamic inputs. Several methods [4, 67] model separate representations per time-step, disregarding the temporal component of the input. D-NeRF [22] reconstruct the scene in a canonical representation and model temporal variations with a deformation field. This idea was developed upon by many works [12, 13, 28, 32, 33, 44]. StreamRF [23] and NeRFPlayer [56] use time-aware MLPs and a compact 3D grid at each time-step for 4D field representation, reducing memory cost for longer videos. Other approaches [7, 14, 54, 60] represent dynamic scenes with a space-time grid, with grid decomposition to increase efficiency. DyNeRF [25] represents dynamic scenes by extending NeRF with an additional time-variant latent code. HyperReel [2] uses an efficient sampling network, modeling the scene around keyframes. MixVoxels [61] represents dynamic and static components with separately processed voxels. Several methods [18, 24, 64] rely on an underlying template mesh (e.g. human). Some methods [8, 50, 75] aim to improve NeRF quality post rendering.", + "bbox": [ + 212, + 493, + 787, + 719 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Gaussian Splatting Much of the development of neural rendering has focused on accelerating inference [15,46,49,62,73]. Recently, 3D Gaussian Splatting [20] made strides by modelling the scene with 3D Gaussians, which, when combined with tile-based differentiable rasterization, achieves very fast rendering, yet preserves high-quality reconstruction. Luiten et al. [35] extend this to dynamic scenes with shared 3D Gaussians that are optimised frame-by-frame. However, their focus is more on tracking 3D Gaussian trajectories rather than maximizing final rendering quality. Concurrent methods to ours that extend 3D Gaus-", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "SWinGS", + "bbox": [ + 669, + 113, + 730, + 127 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f0ba281755eecfee02e1b9aa96def15dc4e98e3654ff807e81e82e7683e023f2.jpg", + "image_caption": [ + "Fig. 2: Method. First, the sequence is partitioned into sliding windows based on optical flow. Second, a dynamic 3DGS model is trained per window, where tunable MLPs model the deformations. Blending parameters $\\alpha$ weigh the MLP's parameters to focus on dynamic parts. Finally, each model is fine-tuned, enforcing inter-window temporal consistency with consistency loss on sampled views for overlapping frames." + ], + "image_footnote": [], + "bbox": [ + 233, + 142, + 411, + 282 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6849426cd786d7313c7aeddbc9fe257bf9cb780fc15ee020032617201022a6fe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 413, + 142, + 586, + 282 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c9c1409faba288cf9f3ff10840eede29f659a345966e1443b0d40e4a1b0a1865.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 589, + 142, + 771, + 282 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "sian Splatting to focus on general dynamic scenes include [17, 26, 31, 57, 71, 72] amongst others, while other works have focused specifically on dynamic human reconstruction [16, 21, 27, 38, 42, 48, 65] and facial animation [11, 47, 52, 68, 70].", + "bbox": [ + 212, + 396, + 787, + 444 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 469, + 330, + 486 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Overview", + "text_level": 1, + "bbox": [ + 214, + 506, + 341, + 520 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We present our method to reconstruct and render novel views of general dynamic scenes from multiple calibrated time-synchronized cameras. An overview of our method is shown in Fig. 2. We build upon 3D Gaussian Splatting [20] for novel view synthesis of static scenes, which we extend to scenes containing motion.", + "bbox": [ + 212, + 536, + 787, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our method can be separated into three main steps. First, given a dynamic sequence, we split the sequence into separate shorter sliding windows of frames for concurrent processing. We adaptively sample windows of varying lengths depending on the amount of motion in the sequence. Each sliding window contains an overlapping frame between adjacent windows, enabling temporal consistency to be enforced throughout the sequence at a subsequent training stage. Partitioning the sequence into smaller windows enables us to deal with sequences of arbitrary length while maintaining high render quality.", + "bbox": [ + 212, + 597, + 787, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Second, we train separate dynamic 3DGS models for each sliding window in turn. We extend the static 3DGS method to model the dynamics by introducing a tuneable MLP [36]. The MLP learns the deformation field from a canonical set of 3D Gaussians for each frame in a window. Thus each window comprises an independent temporally-local canonical 3D Gaussian representation and deformation field. This enables us to handle significant geometric changes and/or if new objects appear throughout the sequence, which can be challenging to model with a single representation. A tuneable MLP weighting parameter $\\alpha$ is learned", + "bbox": [ + 212, + 719, + 789, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "R. Shaw et al.", + "bbox": [ + 271, + 114, + 367, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "for each 3D Gaussian to enable the MLP to focus on modelling the dynamic parts of the scene, with the static parts encapsulated by the canonical representation.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Third, once a dynamic 3DGS model is trained for each window in the sequence, we apply a fine-tuning step to enforce temporal consistency throughout the sequence. We fine-tune each 3DGS model sequentially and employ a self-supervising temporal consistency loss on the overlapping frame renders between neighbouring windows. This encourages the model of the current window to produce similar renderings to the previous window. The result is a set of per-frame Gaussian Splatting models enabling high-quality novel view renderings of dynamic scenes with real-time interactive viewing capability. Our approach enables us to overcome the limitations of training with long sequences and to handle complex motions without exhibiting distracting temporal flickering.", + "bbox": [ + 212, + 176, + 787, + 328 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Preliminary: 3D Gaussian Splitting", + "text_level": 1, + "bbox": [ + 215, + 349, + 555, + 366 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our method is built upon 3D Gaussian Splatting (3DGS) [20]. 3DGS uses a 3D Gaussian representation to model scenes as they are differentiable and can be projected to 2D splats, enabling fast tile-based rasterization. The 3D Gaussians are defined by 3D covariance matrix $\\Sigma$ in world space centered at the mean $\\mu$ :", + "bbox": [ + 212, + 375, + 782, + 436 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nG (x) = e ^ {- \\frac {1}{2} (x) ^ {T} \\Sigma^ {- 1} (x)}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 421, + 448, + 784, + 468 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given a scaling matrix $S$ and rotation matrix $R$ , the corresponding covariance matrix $\\Sigma$ of a 3D Gaussian can be written as $\\Sigma = RSS^T R^T$ . To represent a scene, 3DGS optimizes 3D Gaussian positions $\\mathbf{x}$ , covariance $\\Sigma$ (scaling $S$ and rotation $R$ ), opacity $o$ , and colours, represented by spherical harmonic (SH) coefficients, capturing view-dependent appearance. The optimization is interleaved with Gaussian adaptive density control (ADC). The model is optimized by rendering the learned Gaussians via a differentiable rasterizer, comparing the resulting image $I_r$ against the ground truth $I_{gt}$ , and minimizing the loss function:", + "bbox": [ + 212, + 474, + 787, + 595 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = (1 - \\lambda) \\mathcal {L} _ {1} \\left(I _ {r}, I _ {g t}\\right) + \\lambda \\mathcal {L} _ {\\text {S S I M}} \\left(I _ {r}, I _ {g t}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 611, + 784, + 627 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3 Sliding-Window Processing", + "text_level": 1, + "bbox": [ + 215, + 648, + 486, + 664 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Using a single dynamic 3DGS representation to model an entire sequence becomes impractical for longer sequences, simply from a data processing standpoint. Furthermore, representing a lengthy sequence with a single model performs significantly worse than using multiple smaller fixed-sized segments, particularly if the scene has large motion or topological changes that cannot be modeled easily with one canonical representation and deformation field (Fig. 4). To address this, we use a sliding-window approach, separating the sequence into smaller windows with overlapping frames. The window size is a hyperparameter whose effect is explored in Section 4.1. Each window comprises an independent dynamic 3DGS model, and we allow all 3D Gaussian parameters to change between windows, including the number of Gaussians, their positions, rotations,", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "SWinGS", + "bbox": [ + 669, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "scaling, colours and opacities. The advantage of this approach is that independent models can be trained in parallel across multiple GPUs to speed up training.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Adaptive Window Sampling We propose an adaptive method for choosing the appropriate sliding window sizes, balancing training overhead and model size with performance and temporal consistency. Given an input sequence of frame length $N_{f}$ , we separate the sequence into smaller windows by adaptively sampling windows of different lengths $\\{N_w\\}$ , depending on the amount of motion in the sequence. In high-motion areas, we aim to sample windows more frequently (shorter windows), and in low-motion areas, less frequently (longer windows).", + "bbox": [ + 212, + 194, + 784, + 300 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To do this, we leverage the magnitude of 2D optical flow from each camera viewpoint. We employ a greedy algorithm to adaptively select the sizes of windows, prior to training, based on the accumulated optical flow magnitude. Specifically, we estimate per-frame optical flow $\\pmb{f}$ using a pre-trained RAFT [59] model for each camera view $j\\in V$ and all frames in the sequence $i\\in N_{f}$ , and compute the mean flow magnitude summed over each frame:", + "bbox": [ + 212, + 301, + 785, + 391 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {v} _ {i} = \\frac {1}{V} \\sum_ {j} ^ {V} \\sum_ {i} ^ {N _ {f} - 1} | | \\boldsymbol {f} \\left(I _ {i} ^ {j}, I _ {i + 1} ^ {j}\\right) | | _ {2} ^ {2} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 392, + 404, + 785, + 446 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We iterate over each frame in the sequence with a greedy heuristic; spawning a new window when the sum of mean flow $\\hat{v}_i$ exceeds a pre-defined threshold. This ensures that each window contains a similar amount of movement, leading to a balanced distribution of the total representational workload. Taking the average across viewpoints makes this approach somewhat invariant to the number of cameras, while placing a limit on the total flow stops an excessive amount of movement within each window. Note, each sampled window overlaps with the next window by a single frame, such that neighbouring windows share a common image frame. This is to enable inter-window temporal consistency (section 3.6).", + "bbox": [ + 212, + 449, + 787, + 587 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4 Dynamic 3D Gaussians", + "text_level": 1, + "bbox": [ + 214, + 603, + 454, + 618 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To extend static 3DGS to handle dynamic scenes, we introduce a temporally-local dynamic MLP unique to each sliding window in the sequence (section 3.3). Each time-dependent MLP learns a temporally-local deformation field, mapping from a learned per-window canonical space to a set of 3D Gaussians for each frame in the window. Each deformation field, represented by a small MLP $\\mathcal{F}_{\\theta}$ with weights $\\theta$ , takes as input the normalized frame time $t \\in [0,1]$ and 3D Gaussian means $\\pmb{x}$ (normalized by the scene's mean and standard deviation), and outputs displacements to their positions $\\Delta \\pmb{x}$ , rotations $\\Delta \\pmb{r}$ and scaling $\\Delta s$ :", + "bbox": [ + 212, + 623, + 787, + 746 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\boldsymbol {x} (t), \\Delta \\boldsymbol {r} (t), \\Delta \\boldsymbol {s} (t) = \\mathcal {F} _ {\\theta} (\\gamma (\\boldsymbol {x}), \\gamma (t)) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 761, + 785, + 776 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\gamma(.)$ denotes sinusoidal positional encoding $\\gamma: \\mathbb{R}^3 \\to \\mathbb{R}^{3+6m}$ , $\\gamma(x) = (x, \\ldots, \\sin(2k\\pi x), \\cos(2k\\pi x), \\ldots)$ . We use small MLPs to reduce overfitting, setting the number of frequency components $m = 6$ , MLP depth $D = 4$ and width $W = 16$ , with two skip connections.", + "bbox": [ + 212, + 777, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "R. Shaw et al.", + "bbox": [ + 271, + 114, + 367, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.5 Tunable Dynamic MLPs", + "text_level": 1, + "bbox": [ + 215, + 146, + 464, + 161 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ideally, time-dependent MLPs model motion in the scene by associating the dynamic parts with the temporal input component $\\gamma(t)$ and thus learn to decouple the scene into i) static canonical 3D Gaussians and ii) dynamic 3D Gaussians. However, in scenes with an imbalance of static vs dynamic parts, e.g. Neural 3D Video [25] with mostly static backgrounds, the MLP struggles to disentangle static and dynamic regions, resulting in poorly estimated deformation fields.", + "bbox": [ + 212, + 171, + 787, + 261 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To counteract this, we introduce tuneable dynamic MLPs [36] (i.e. MLPs with $M$ sets of weights) governed by $M$ sets of learnable blending parameters $\\{\\pmb{\\alpha} \\in \\mathbb{R}^{M \\times N_g}\\}$ . The tuning parameters weigh the respective parameters of the MLP for each input Gaussian $i \\in N_g$ , enabling the learning of $M$ modes of variation corresponding to different motion modes. With $M = 2$ , the MLP has the ability to decouple static and dynamic Gaussians in a smoothly weighted manner. Given the set of blending parameters $\\{\\alpha_i\\}_{m=1}^M$ for the $i$ -th input Gaussian $\\pmb{x}_i$ , the output of a single layer of the dynamic MLP $\\pmb{y}_i$ can be written as a weighted sum over the $M$ sets of weights:", + "bbox": [ + 214, + 261, + 789, + 398 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {y} _ {i} = \\phi \\left(\\sum_ {m = 1} ^ {M} \\left(\\alpha_ {i, m} \\boldsymbol {w} _ {m} ^ {T} \\boldsymbol {x} _ {i} + \\alpha_ {i, m} \\boldsymbol {b} _ {m}\\right)\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 364, + 410, + 785, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\boldsymbol{w}$ and $\\boldsymbol{b}$ are the weights and bias, and $\\phi$ is a non-linear activation function. The blending parameters $\\{\\alpha_{i}\\}_{m = 1}^{M}$ linearly blend $M$ sets of weights and biases of each layer of the MLP for each input Gaussian which, when passed through the activation function, enables nonlinear interaction between $\\alpha$ and the output. Applying $\\alpha$ in this way naturally enables the learning different of motion modes with a single forward pass of the MLP. Thus the dynamic MLP $\\mathcal{F}_{\\theta}^{dyn}$ becomes a function of position, time, and blending parameters $\\alpha$ :", + "bbox": [ + 214, + 455, + 787, + 561 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\boldsymbol {x} (t), \\Delta \\boldsymbol {r} (t), \\Delta \\boldsymbol {s} (t) = \\mathcal {F} _ {\\theta} ^ {d y n} (\\gamma (\\boldsymbol {x}), \\gamma (t), \\boldsymbol {\\alpha}). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 575, + 785, + 592 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We implement the dynamic MLP as a single batch matrix multiplication (see supplementary). Blending parameters $\\alpha$ are initialized as a binary mask of dynamic Gaussians (0-static, 1-dynamic) as follows. For a sliding window, we project all 3D Gaussians into each camera view: $u = \\varPi_{j}(\\pmb{x})$ , obtaining their 2D-pixel coordinates. We compute their L1 pixel differences from the central frame to each frame in the window. If the difference is larger than a threshold, we label that Gaussian 1, otherwise 0. To be robust to occlusions and mislabelling, we average the assigned label over all views and frames in the window. If the average is greater than 0.5, we initialize the Gaussian dynamic, otherwise static.", + "bbox": [ + 212, + 598, + 787, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Once initialized, we set the blending parameter $\\alpha$ as a learnable parameter, and optimize it via back-propagation together with the rest of the system. We assume $\\alpha$ is constant for all frames in a sliding window to reduce the complexity of the optimization, which is a fair assumption given our adaptive window sampling mechanism. Fig. 3 visualizes the learning of the MLP tuning parameter. We observe $\\alpha$ providing higher weight to Gaussians likely to be dynamic. Fig. 3 (right) plots the magnitudes of resulting displacements $\\Delta x$ output by the MLP.", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "SWinGS", + "bbox": [ + 669, + 113, + 730, + 127 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Note, $\\alpha$ does not have to be entirely accurate, as the MLP learns to adjust accordingly, yet enables the MLP to handle highly imbalanced scenes (Table 6).", + "bbox": [ + 215, + 146, + 787, + 176 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c3dcdb2f11a8ec061cf3864f723627d2eb0a6c004174315a99e39398c82a3011.jpg", + "image_caption": [ + "Fig. 3: Dynamic MLPs with tunable parameters $\\alpha$ weigh the parameters of the MLP for each Gaussian. We show renders from two scenes, left: cook spinach [25] and right: Train [51]. Shown from left-to-right: image render, tunable $\\alpha$ parameters, and normalized MLP displacements $\\Delta x$ . Note, $\\alpha$ highlights the scene's dynamic regions." + ], + "image_footnote": [], + "bbox": [ + 218, + 203, + 488, + 275 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/355e0502329e36c80dec5967c0753a42ca8b3763eafae513c562decf22bd4b69.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 495, + 203, + 784, + 273 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.6 Temporal Consistency Fine-tuning", + "text_level": 1, + "bbox": [ + 215, + 405, + 545, + 421 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Due to the non-deterministic nature of the 3DGS optimization, independently trained models for each sliding window may produce slightly different results. When the resulting renders from each model are played back in sequence, noticeable flickering can be observed, particularly in novel views (though it's less noticeable in the training views). Consequently, after training models for all sliding windows separately, we tackle inter-window temporal flickering by introducing a fine-tuning step for temporal consistency. This step uses a self-supervising loss function to aid in smooth transitions between the models of separate windows.", + "bbox": [ + 212, + 430, + 787, + 550 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We fine-tune each model for a short period (3000 iterations in our experiments). We initiate the process from the first sliding window and progress through the sequence sequentially. For a window comprising $N_{w}$ frames, we load the trained model and the model from the preceding window, with one overlapping frame between them. We then freeze all the parameters of the previous model. In order to fine-tune the model, as the flickering is mainly observed in novel views, we randomly sample novel test views in between the training views by rigidly interpolating the training camera poses $P_{j} = [R|\\pmb{t}] \\in \\mathbb{R}^{4 \\times 4}$ in SE(3):", + "bbox": [ + 212, + 551, + 787, + 674 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nP _ {\\text {n o v e l}} = \\exp_ {\\mathrm {M}} \\sum_ {j} ^ {V} \\beta_ {j} \\log_ {\\mathrm {M}} \\left(P _ {j}\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 392, + 683, + 785, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\exp_{\\mathbf{M}}$ and $\\log_{\\mathbf{M}}$ are the matrix exponential and logarithm [1] respectively, and $\\beta_{j}\\in [0,1]$ is a uniformly sampled weighting such that $\\sum_{j}^{V}\\beta_{j} = 1$ .", + "bbox": [ + 212, + 732, + 784, + 765 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In one fine-tuning step, we render the overlapping frames from a randomly sampled novel viewpoint using both the model of the current window $w$ and the previous window $w - 1$ . This means we use the first frame of current window $I_{t = 0}^{w}$ and the last frame of the previous window $I_{t = N_w - 1}^{w - 1}$ . We then apply a consistency loss, which is simply the L1 loss on the two image renders from both models:", + "bbox": [ + 212, + 766, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "R. Shaw et al.", + "bbox": [ + 271, + 114, + 367, + 127 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c o n s i s t e n c y}} = \\left| I _ {t = 0} ^ {w} - I _ {t = N _ {w} - 1} ^ {w - 1} \\right| _ {1} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 385, + 159, + 785, + 179 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "When fine-tuning a model, we only allow the canonical (static) representation to change, while freezing the weights of the time-dependent dynamic MLP. Since the canonical set of 3D Gaussians is shared for all frames in a sliding window, we need to be careful not to negatively impact other frames when refining the overlapping frame. To address this, we use an alternating strategy for refinement, enforcing temporal consistency on the overlapping frame $75\\%$ of the time and training with the remaining views and frames as usual for the remaining $25\\%$ .", + "bbox": [ + 212, + 183, + 787, + 289 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.7 Implementation Details", + "text_level": 1, + "bbox": [ + 214, + 309, + 455, + 324 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We implement our method in PyTorch, building upon the codebase and differentiable rasterizer of 3DGS [20]. We initialize each model with a point cloud obtained from COLMAP [53]. Each dynamic 3DGS model is trained for 15K iterations for all sampled windows in a sequence. The initial 2K iterations comprise a warm-up stage; training with only the central window frame and freezing the weights of the MLP, allowing the canonical representation to stabilize. We optimize the Gaussians' positions, rotations, scaling, opacities and SH coefficients. Afterwards, we unfreeze the MLP and allow the deformation field and tuning parameters $\\alpha$ to optimize. We find this staggered optimization approach leads to better convergence. The number of Gaussians densifies for 8K iterations, after which the number of Gaussians is fixed. Inputs to the MLP are normalized by the mean and standard deviation of the canonical point cloud post-warm-up stage before frequency encoding, leading to faster and stabler convergence. We train with Adam optimizer, using different learning rates for each parameter following the implementation of [20]. The learning rate for the MLP and $\\alpha$ parameters are set to 1e-4, with the MLP learning rate undergoing exponential decay by factor 1e-2 in 20K iterations. During the initial optimization phase, due to the independence of each Gaussian model, we train each model in parallel on eight 32Gb Tesla V100 GPUs to speed up training. Afterwards, we perform temporal fine-tuning of each model sequentially using a single GPU for 3K iterations each.", + "bbox": [ + 212, + 332, + 787, + 636 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/cf8c5288faf9ed86146ead5a2213368d6f3b125b4e80450aebb1d0903f3c8e13.jpg", + "table_caption": [ + "Table 2: Quantitative results on Technicolor dataset [51] at full resolution. Best and second best results are highlighted." + ], + "table_footnote": [], + "table_body": "
MethodPSNRSSIMLPIPSFPS
DyNeRF [25]31.800.9110.1420.02
HypeRreel [2]32.500.9020.1130.45
Dynamic3DG [35]27.020.8320.22886.96
Ours33.650.9340.11723.79
", + "bbox": [ + 217, + 719, + 537, + 808 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/53586640ec588eaa2592c01d4b8987d6e1e79d5e2727521a8cae144dd4a52479.jpg", + "image_caption": [ + "Fig. 4: Comparison in performance consistency for Ours and Dynamic3DG in consecutive frames." + ], + "image_footnote": [], + "bbox": [ + 557, + 662, + 777, + 755 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "SWinGS", + "bbox": [ + 669, + 113, + 730, + 127 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/068521d4d9ca8a3d8817ce701a7d8e8e00c2d8bb671ce83e9fa5b5d7e7c1d020.jpg", + "table_caption": [ + "Table 1: Quantitative results on the Neural 3D Video dataset [25], averaged over all scenes. Best and second best highlighted. Our method performs best overall whilst enabling real-time frame rates. Per-scene breakdown of results are given in Table 3." + ], + "table_footnote": [], + "table_body": "
MethodPSNRSSIMLPIPSFPS
MixVoxels [61]30.420.9230.1244.30
K-Planes [14]30.630.9220.1170.33
HexPlane [7]30.000.9220.1130.24
HyperReel [2]30.780.9310.1013.60
NeRFPlayer [56]30.700.9310.1210.10
StreamRF [23]30.230.9040.1779.40
4DGS [65]27.610.9160.13530.00
Ours31.100.9400.09671.51
", + "bbox": [ + 341, + 200, + 655, + 335 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Results", + "text_level": 1, + "bbox": [ + 215, + 362, + 326, + 378 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We evaluate our method on two real-world multi-view dynamic benchmarks: the Neural 3D Video dataset [25] and the Technicolor dataset [51].", + "bbox": [ + 212, + 396, + 782, + 425 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Neural 3D Video comprises real-world dynamic scenes captured with a time-synchronized multi-view system at $2028 \\times 2704$ resolution at 30 FPS. Camera parameters are estimated using COLMAP [53]. We compare our method to K-Planes [14], HexPlane [7], MixVoxels [61], HyperReel [2], NeRFPlayer [56], StreamRF [23], and 4DGS [65]. We compute quantitative metrics for the central test view at half the original resolution ( $1014 \\times 1352$ ) for 300 frames. The average results for each method over the dataset are given in Table 1, while Table 3 provides a breakdown of the per-scene performance. The results show that our method performs best regarding PSNR and SSIM metrics while offering the fastest rendering performance. Qualitative results are shown in Fig. 5.", + "bbox": [ + 212, + 426, + 784, + 575 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Technicolor captures real dynamic scenes from a synchronized $4 \\times 4$ camera array at $2048 \\times 1088$ resolution. Following [2], we evaluate on the second row second column camera on five scenes: Birthday, Fabien, Painter, Theater and Trains. We compare to DynNeRF [25], HyperReel [2] and Dynamic 3D Gaussians [35], with quantitative and qualitative results in Table 4 and Fig. 6 respectively.", + "bbox": [ + 212, + 578, + 784, + 652 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 676, + 401, + 691 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "This section provides ablations showing the effectiveness of our sliding window and self-supervised temporal consistency fine-tuning strategies. As we train independent models for each window, flickering artifacts can occur in the final renders. To visualize this, Fig. 8 plots absolute image error between renders of neighbouring frames (overlapping frames outlined in red). Without temporal consistency, we observe a spike in absolute error on overlapping frames, resulting in undesirable flickering. However, after fine-tuning, the error in overlapping frames is drastically reduced. In Table 5, we compute per-frame image metrics and estimate a measure of the temporal consistency using SoTA video quality", + "bbox": [ + 212, + 704, + 785, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "R. Shaw et al.", + "bbox": [ + 271, + 114, + 367, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/cb5918eee4c8df66e596aea7fdb190a779cc58a2aed28bbbccecf4b239059144.jpg", + "table_caption": [ + "Table 3: Quantitative results on the Neural 3D Video dataset [25], evaluated for 300 frames at $1352 \\times 1014$ resolution. †As reported in [2], ‡Natively trained in lower resolution, upscaled. Best and second best results highlighted respectively. Our method performs competitively in all metrics, usually coming in either first or second place." + ], + "table_footnote": [], + "table_body": "
Scene\nMethodCook SpinachCut Roast BeefFlame Steak
PSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPS
MixVoxels [61]31.390.9310.11331.380.9280.11130.150.9380.108
K-Planes [14]31.230.9260.11431.870.9280.11431.490.9400.102
HexPlane‡ [7]31.050.9280.11430.830.9270.11530.420.9390.104
HyperReel [2]31.770.9320.09032.250.9360.08631.480.9390.083
NeRFPlayer† [56]30.580.9290.11329.350.9080.14431.930.9500.088
StreamRF [23]30.890.9140.16230.750.9170.15431.370.9230.152
Ours31.960.9460.09431.840.9450.09932.180.9530.087
Sear steakCoffee MartiniFlame Salmon
MixVoxels [61]30.850.9400.10329.250.9010.14729.500.8980.163
K-Planes [14]30.280.9370.10429.300.9000.13429.580.9010.132
HexPlane‡ [7]30.000.9390.10528.450.8910.14929.230.9050.088
HyperReel [2]31.880.9420.08028.650.8970.12928.260.9410.136
NeRFPlayer† [56]29.130.9080.13831.530.9510.08531.650.9400.098
StreamRF [23]31.600.9250.14728.130.8730.21928.690.8720.229
Ours32.210.9500.09229.160.9210.10529.250.9250.100
", + "bbox": [ + 215, + 244, + 785, + 508 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/c02a05478f2744b6422de5bc37ab35c8ec6c75c8a6df04413654b64288be65b4.jpg", + "table_caption": [ + "Table 4: Quantitative results on the Technicolor dataset [51] evaluated at full resolution. Best and second best results highlighted respectively." + ], + "table_footnote": [], + "table_body": "
SceneBirthdayFabienPainter
MethodPSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPS
DyNeRF [25]29.200.9090.06732.760.9090.24235.950.9300.147
HyperReel [2]30.790.9220.06232.280.8600.21735.680.9260.123
Dynamic3DG [35]27.060.8590.11326.340.8340.26825.180.7580.395
Ours33.440.9590.04234.430.9250.17136.760.9480.128
SceneTheaterTrainAverage
DyNeRF [25]29.530.8750.18831.580.9330.06731.800.9110.142
HyperReel [2]33.670.8950.10430.100.9090.06132.500.9020.113
Dynamic3DG [35]28.050.7990.27728.460.9080.08827.020.8320.228
Ours29.810.8840.20133.990.9570.04333.690.9340.117
", + "bbox": [ + 217, + 623, + 785, + 804 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "SWinGS", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/bcf93faffbc7ba5e1f00a23b627d57d3e8172798dcb6349ef25b0ba85d0f6d06.jpg", + "image_caption": [ + "Fig. 5: Qualitative results on Neural 3D Video [25]. Scenes top to bottom: i) coffee martini, ii) cook spinach, iii) cut roasted beef, iv) flame salmon, v) flame steak." + ], + "image_footnote": [], + "bbox": [ + 246, + 160, + 738, + 460 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/21b4ecc10dd87b69a1ba7abf40eeb7e45537e7b3295d88bf06eba3b0671b091b.jpg", + "image_caption": [ + "Fig. 6: Qualitative results on Technicolor [51]. Scenes top to bottom: i) Birthday, ii) Painter, iii) Train." + ], + "image_footnote": [], + "bbox": [ + 253, + 542, + 736, + 777 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "R. Shaw et al.", + "bbox": [ + 271, + 114, + 367, + 126 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "assessment metric FAST-VQA [66], where the quality score is in the range [0,1]. The table provides results averaged over all scenes from the Neural 3D Video dataset [25]. Although we incur a minor penalty in some per-frame performance metrics (SSIM and LPIPS), we obtain a significantly higher video quality (VQA) score. This indicates a substantial improvement in temporal consistency and overall perceptual video quality, resulting in more pleasing renderings.", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/d84f6286710ef22b89cd3ec34cf477d388614a1c0efef27c7ecc797a7996c84b.jpg", + "table_caption": [ + "Table 5: Ablation on temporal fine-tuning. Results averaged over all scenes from [25]. Temporal consistency is measured using t-LPIPS [10] and FAST-VQA [66] (quality score in the range [0,1]). Per-frame performance remains fairly constant, but temporal consistency and overall perceptual video quality is significantly improved." + ], + "table_footnote": [], + "table_body": "
MethodPSNR ↑SSIM ↑LPIPS ↓t-LPIPS ↓VQA ↑
w/o temporal consistency32.010.9560.0850.01290.666
w/ temporal consistency32.050.9490.0930.01020.726
Ground truth----0.763
", + "bbox": [ + 251, + 329, + 746, + 393 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/8bb945534981a7ee776a1acbcf01a01236dfa330513c8ac914a6ebc2040d11a8.jpg", + "table_caption": [ + "Table 6: Ablation on sliding window size vs adaptive on Birthday scene [51]. We show the improvement from the dynamic MLP. Adaptive chooses window sizes that match, and in some metrics exceed, the best fixed-size window performance, striking balance between performance, temporal consistency (t-LPIPS) and training time (GPU hrs)." + ], + "table_footnote": [], + "table_body": "
Window SizeNo. WindowsTrain timePSNRSSIMLPIPSt-LPIPS
32416.033.120.9560.0480.0076
964.033.380.9590.0430.0053
1732.033.010.9560.0450.0049
2521.332.970.9560.0430.0048
4910.732.730.9550.0470.0051
Adaptive53.333.440.9590.0420.0051
Adaptive (w/o dyn. MLP)53.332.760.9570.0450.0062
", + "bbox": [ + 215, + 523, + 787, + 631 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 6 and Fig. 7 show the impact of the sliding-window size hyperparameter and our adaptive sampling strategy. Adaptive sampling automatically chooses appropriate window sizes that match and sometimes exceed the best fixed-size window performance, striking a good balance between performance, temporal consistency (t-LPIPS [10]) and training time. Adaptive performs the best overall with fewer windows (less storage requirements). We also show the advantage gained from using a dynamic MLP vs a regular MLP, which improves all metrics.", + "bbox": [ + 212, + 656, + 787, + 763 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 782, + 359, + 797 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We have presented a method to render novel views of dynamic scenes by extending the 3DGS framework. Results show that our method produces high-", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "SWinGS", + "bbox": [ + 669, + 113, + 730, + 127 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/c3697ee26a3e77e502cc4cee2086eb36ae715e7de5a7252cfccde31075409d3a.jpg", + "image_caption": [ + "Fig. 7: Ablation on sliding window size vs adaptive window sampling (with and without dynamic MLP) on scenes from the Technicolor dataset [51]. The single-window scenario (w49) has the shortest training time but unsatisfactory visual quality, while adaptive windows offer the best balance between quality and training overhead." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 265 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/b1339bc65944398bb3e4dfe6a041fb1a31b9789f04fde8b81c44606f1741a0d9.jpg", + "image_caption": [ + "Fig. 8: Ablation on temporal fine-tuning: we display the absolute error between neighbouring frame renders, with overlapping frames highlighted red. After fine-tuning, the error is substantially reduced and overall perceptual video quality is improved." + ], + "image_footnote": [], + "bbox": [ + 217, + 349, + 785, + 523 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "quality renderings, even with complex motions e.g. flames. Key to our approach is sliding-window processing, which adaptively partitions sequences into manageable chunks. Processing each window separately, allowing the canonical representation and deformation field to vary throughout the sequence, enables us to handle complex topological changes, and reduces the magnitude and variability of the 3D scene flow. In contrast, other methods that learn a single representation for whole sequences are impractical for long sequences and degrade in quality with increasing length. Introducing an MLP for each window learns the deformation field from each canonical representation to a set of per-frame 3D Gaussians. Moreover, learnable tuning parameters help disentangle static and dynamic parts of the scene, which we find essential for imbalanced scenes. Our ablations show self-supervised temporal consistency fine-tuning reduces temporal flickering and improves the overall perceptual video quality, with only a minor impact on per-frame performance metrics. Overall, our method performs strongly compared to recent SoTA quantitatively and obtains sharper, temporally-consistent results.", + "bbox": [ + 212, + 613, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "R. Shaw et al.", + "bbox": [ + 271, + 114, + 367, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Alexa, M.: Linear Combination of Transformations. ACM Trans. Graph. (2002)", + "2. Attal, B., Huang, J.B., Richardt, C., Zollhöfer, M., Kopf, J., O'Toole, M., Kim, C.: HyperReel: High-fidelity 6-DoF Video with Ray-conditioned Sampling. In: CVPR (2023)", + "3. Bansal, A., Vo, M., Sheikh, Y., Ramanan, D., Narasimhan, S.: 4D Visualization of Dynamic Events from Unconstrained Multi-view Videos. In: CVPR (2020)", + "4. Bansal, A., Zollhöfer, M.: Neural Pixel Composition for 3D-4D View Synthesis from Multi-views. In: CVPR (2023)", + "5. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-NeRF: A Multiscale Representation for Anti-aliasing Neural Radiance Fields. In: ICCV (2021)", + "6. Bārsan, I.A., Liu, P., Pollefeys, M., Geiger, A.: Robust Dense Mapping for Large-scale Dynamic Environments. In: International Conference on Robotics and Automation (2018)", + "7. Cao, A., CV, J.: HexPlane: A Fast Representation for Dynamic Scenes. In: CVPR (2023)", + "8. Catley-Chandar, S., Shaw, R., Slabaugh, G., Pérez-Pellitero, E.: RoGUENeRF: A Robust Geometry-consistent Universal Enhancer for NeRF. arXiv preprint arXiv:2403.11909 (2024)", + "9. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: TensoRF: Tensorial Radiance Fields. In: ECCV (2022)", + "0. Chu, M., Xie, Y., Mayer, J., Leal-Taix'e, L., Thuerey, N.: Learning Temporal Coherence via Self-supervision for GAN-based Video Generation. ACM Trans. Graph. 39, 75:1 - 75:13 (2020)", + "1. Dhamo, H., Nie, Y., Moreau, A., Song, J., Shaw, R., Zhou, Y., Pérez-Pellitero, E.: HeadGaS: Real-time Animatable Head Avatars via 3D Gaussian Splatting. arXiv preprint arXiv:2312.02902 (2023)", + "2. Du, Y., Zhang, Y., Yu, H.X., Tenenbaum, J.B., Wu, J.: Neural Radiance Flow for 4D View Synthesis and Video Processing. In: ICCV (2021)", + "3. Fang, J., Yi, T., Wang, X., Xie, L., Zhang, X., Liu, W., Nießner, M., Tian, Q.: Fast Dynamic Radiance Fields with Time-aware Neural Voxels. In: SIGGRAPH Asia (2022)", + "4. Fridovich-Keil, S., Meanti, G., Warburg, F.R., Recht, B., Kanazawa, A.: K-Planes: Explicit Radiance Fields in Space, Time, and Appearance. In: CVPR (2023)", + "5. Hedman, P., Srinivasan, P.P., Mildenhall, B., Barron, J.T., Debevec, P.: Baking Neural Radiance Fields for Real-time View Synthesis. In: ICCV (2021)", + "6. Hu, L., Zhang, H., Zhang, Y., Zhou, B., Liu, B., Zhang, S., Nie, L.: GaussianAvatar: Towards Realistic Human Avatar Modeling from a Single Video via Animatable 3D Gaussians. In: CVPR (2024)", + "7. Huang, Y.H., Sun, Y.T., Yang, Z., Lyu, X., Cao, Y.P., Qi, X.: SC-GS: Sparse-Controlled Gaussian Splatting for Editable Dynamic Scenes. In: CVPR (2024)", + "8. Isik, M., Rünz, M., Georgopoulos, M., Khakhulin, T., Starck, J., Agapito, L., Niešner, M.: HumanRF: High-fidelity Neural Radiance Fields for Humans in Motion. ACM Trans. Graph. 42(4), 1-12 (2023)", + "9. Joo, H., Soo Park, H., Sheikh, Y.: MAP Visibility Estimation for Large-scale Dynamic 3D Reconstruction. In: CVPR (2014)", + "20. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3D Gaussian Splatting for Real-time Radiance Field Rendering. ACM Trans. Graph. 42(4) (2023)" + ], + "bbox": [ + 225, + 176, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "SWinGS", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "21. Kocabas, M., Chang, J.H.R., Gabriel, J., Tuzel, O., Ranjan, A.: HUGS: Human Gaussian Splatting. In: CVPR (2024)", + "22. Li, L., Shen, Z., Wang, Z., Shen, L., Tan, P.: D-NeRF: Neural Radiance Fields for Dynamic Scenes. In: CVPR (2020)", + "23. Li, L., Shen, Z., Wang, Z., Shen, L., Tan, P.: Streaming Radiance Fields for 3D Video Synthesis. In: NeurIPS (2022)", + "24. Li, R., Tanke, J., Vo, M., Zollhöfer, M., Gall, J., Kanazawa, A., Lassner, C.: TAVA: Template-free Animatable Volumetric Actors. In: ECCV (2022)", + "25. Li, T., Slavcheva, M., Zollhöfer, M., Green, S., Lassner, C., Kim, C., Schmidt, T., Lovegrove, S., Goesele, M., Newcombe, R., Lv, Z.: Neural 3D Video Synthesis from Multi-view Video. In: CVPR (2022)", + "26. Li, Z., Chen, Z., Li, Z., Xu, Y.: Spacetime Gaussian Feature Splatting for Real-time Dynamic View Synthesis. In: CVPR (2024)", + "27. Li, Z., Zheng, Z., Wang, L., Liu, Y.: Animatable Gaussians: Learning Pose-dependent Gaussian Maps for High-fidelity Human Avatar Modeling. In: CVPR (2024)", + "28. Li, Z., Niklaus, S., Snavely, N., Wang, O.: Neural Scene Flow Fields for Space-time View Synthesis of Dynamic Scenes. In: CVPR (2021)", + "29. Li, Z., Wang, Q., Cole, F., Tucker, R., Snavely, N.: DynIBaR: Neural Dynamic Image-Based Rendering. In: CVPR (2023)", + "30. Lin, K.E., Xiao, L., Liu, F., Yang, G., Ramamoorthi, R.: Deep 3D Mask Volume for View Synthesis of Dynamic Scenes. In: ICCV (2021)", + "31. Lin, Y., Dai, Z., Zhu, S., Yao, Y.: Gaussian-Flow: 4D Reconstruction with Dynamic 3D Gaussian Particle. In: CVPR (2024)", + "32. Liu, Y.L., Gao, C., Meuleman, A., Tseng, H.Y., Saraf, A., Kim, C., Chuang, Y.Y., Kopf, J., Huang, J.B.: Robust Dynamic Radiance Fields. In: CVPR (2023)", + "33. Lombardi, S., Simon, T., Saragih, J., Schwartz, G., Lehrmann, A., Sheikh, Y.: Neural Volumes: Learning Dynamic Renderable Volumes from Images. ACM Trans. Graph. 38(4) (2019)", + "34. Luiten, J., Fischer, T., Leibe, B.: Track to Reconstruct and Reconstruct to Track. IEEE Robotics and Automation Letters 5(2), 1803-1810 (2020)", + "35. Luiten, J., Kopanas, G., Leibe, B., Ramanan, D.: Dynamic 3D Gaussians: Tracking by Persistent Dynamic View Synthesis. In: 3DV (2024)", + "36. Maggioni, M., Tanay, T., Babiloni, F., McDonagh, S.G., Leonardis, A.: Tunable Convolutions with Parametric Multi-loss Optimization. In: CVPR (2023)", + "37. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. In: ECCV (2020)", + "38. Moreau, A., Song, J., Dhamo, H., Shaw, R., Zhou, Y., Pérez-Pellitero, E.: Human Gaussian Splatting: Real-time Rendering of Animatable Avatars. In: CVPR (2024)", + "39. Müller, T., Evans, A., Schied, C., Keller, A.: Instant Neural Graphics Primitives with a Multiresolution Hash Encoding. ACM Trans. Graph. 41(4), 102:1-102:15 (2022)", + "40. Newcombe, R.A., Fox, D., Seitz, S.M.: DynamicFusion: Reconstruction and Tracking of Non-rigid Scenes in Real-time. In: CVPR (2015)", + "41. Niemeyer, M., Barron, J.T., Mildenhall, B., Sajjadi, M.S.M., Geiger, A., Radwan, N.: RegNeRF: Regularizing Neural Radiance Fields for View Synthesis from Sparse Inputs. In: CVPR (2021)", + "42. Pang, H., Zhu, H., Kortylewski, A., Theobalt, C., Habermann, M.: ASH: Animatable Gaussian Splats for Efficient and Photoreal Human Rendering. In: CVPR (2024)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "R. Shaw et al.", + "bbox": [ + 271, + 114, + 367, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "43. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin-Brualla, R.: Nerfies: Deformable Neural Radiance Fields. ICCV (2021)", + "44. Park, K., Sinha, U., Hedman, P., Barron, J.T., Bouaziz, S., Goldman, D.B., Martin-Brualla, R., Seitz, S.M.: HyperNeRF: A Higher-dimensional Representation for Topologically Varying Neural Radiance Fields. ACM Trans. Graph. 40(6) (2021)", + "45. Peng, S., Dong, J., Wang, Q., Zhang, S., Shuai, Q., Zhou, X., Bao, H.: Animatable Neural Radiance Fields for Modeling Dynamic Human Bodies. In: ICCV (2021)", + "46. Peng, S., Yan, Y., Shuai, Q., Bao, H., Zhou, X.: Representing Volumetric Videos as Dynamic MLP Maps. In: CVPR (2023)", + "47. Qian, S., Kirschstein, T., Schoneveld, L., Davoli, D., Giebenhain, S., Nießner, M.: GaussianAvatars: Photorealistic Head Avatars with Rigged 3D Gaussians. In: CVPR (2024)", + "48. Qian, Z., Wang, S., Mihajlovic, M., Geiger, A., Tang, S.: 3DGS-Avatar: Animatable Avatars via Deformable 3D Gaussian Splitting. In: CVPR (2024)", + "49. Reiser, C., Peng, S., Liao, Y., Geiger, A.: KiloNeRF: Speeding up Neural Radiance Fields with Thousands of Tiny MLPs. In: ICCV (2021)", + "50. Rong, X., Huang, J.B., Saraf, A., Kim, C., Kopf, J.: Boosting View Synthesis with Residual Transfer. In: CVPR (2022)", + "51. Sabater, N., Boisson, G., Vandame, B., Kerbiriou, P., Babon, F., et al.: Dataset and Pipeline for Multi-view Light-Field Video. In: CVPRW (2017)", + "52. Saito, S., Schwartz, G., Simon, T., Li, J., Nam, G.: Relightable Gaussian Codec Avatars. In: CVPR (2024)", + "53. Schonberger, J.L., Frahm, J.M.: Structure-from-Motion Revisited. In: CVPR (2016)", + "54. Shao, R., Zheng, Z., Tu, H., Liu, B., Zhang, H., Liu, Y.: Tensor4D: Efficient Neural 4D Decomposition for High-fidelity Dynamic Reconstruction and Rendering. In: CVPR (2023)", + "55. Slavcheva, M., Baust, M., Cremers, D., Ilic, S.: KillingFusion: Non-rigid 3D Reconstruction without Correspondences. In: CVPR (2017)", + "56. Song, L., Chen, A., Li, Z., Chen, Z., Chen, L., Yuan, J., Xu, Y., Geiger, A.: NeRF-Player: A Streamable Dynamic Scene Representation with Decomposed Neural Radiance Fields. IEEE Transactions on Visualization and Computer Graphics 29(5), 2732-2742 (2023)", + "57. Sun, J., Jiao, H., Li, G., Zhang, Z., Zhao, L., Xing, W.: 3DGStream: On-the-fly Training of 3D Gaussians for Efficient Streaming of Photo-realistic Free-viewpoint Videos. In: CVPR (2024)", + "58. Tanay, T., Maggioni, M.: Global Latent Neural Rendering. In: CVPR (2024)", + "59. Teed, Z., Deng, J.: RAFT: Recurrent All-pairs Field Transforms for Optical Flow. In: ECCV (2020)", + "60. Turki, H., Zhang, J.Y., Ferroni, F., Ramanan, D.: SUDS: Scalable Urban Dynamic Scenes. In: CVPR (2023)", + "61. Wang, F., Tan, S., Li, X., Tian, Z., Liu, H.: Mixed Neural Voxels for Fast Multiview Video Synthesis. In: ICCV (2023)", + "62. Wang, L., Zhang, J., Liu, X., Zhao, F., Zhang, Y., Zhang, Y., Wu, M., Yu, J., Xu, L.: Fourier PlenOctrees for Dynamic Radiance Field Rendering in Real-time. In: CVPR (2022)", + "63. Wang, Q., Wang, Z., Genova, K., Srinivasan, P., Zhou, H., Barron, J.T., Martin-Brualla, R., Snavely, N., Funkhouser, T.: IBRNet: Learning Multi-view Image-Based Rendering. In: CVPR (2021)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "SWinGS", + "bbox": [ + 669, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "64. Weng, C.Y., Curless, B., Srinivasan, P.P., Barron, J.T., Kemelmacher-Shlizerman, I.: HumanNeRF: Free-viewpoint Rendering of Moving People from Monocular Video. In: CVPR (2022)", + "65. Wu, G., Yi, T., Fang, J., Xie, L., Zhang, X., Wei, W., Liu, W., Tian, Q., Wang, X.: 4D Gaussian Splatting for Real-time Dynamic Scene Rendering. In: CVPR (2024)", + "66. Wu, H., Chen, C., Hou, J., Liao, L., Wang, A., Sun, W., Yan, Q., Lin, W.: FASTVQA: Efficient End-to-end Video Quality Assessment with Fragment Sampling. In: ECCV (2022)", + "67. Xian, W., Huang, J.B., Kopf, J., Kim, C.: Space-time Neural Irradiance Fields for Free-viewpoint Video. In: CVPR (2021)", + "68. Xiang, J., Gao, X., Guo, Y., Zhang, J.: FlashAvatar: High-fidelity Head Avatar with Efficient Gaussian Embedding. In: CVPR (2024)", + "69. Xing, W., Chen, J.: Temporal-MPI: Enabling Multi-Plane Images for Dynamic Scene Modelling via Temporal Basis Learning. In: ECCV (2022)", + "70. Xu, Y., Chen, B., Li, Z., Zhang, H., Wang, L., Zheng, Z., Liu, Y.: Gaussian HeadAvatar: Ultra High-fidelity Head Avatar via Dynamic Gaussians. In: CVPR (2024)", + "71. Yang, Z., Yang, H., Pan, Z., Zhang, L.: Real-time Photorealistic Dynamic Scene Representation and Rendering with 4D Gaussian Splatting. In: ICLR (2024)", + "72. Yang, Z., Gao, X., Zhou, W., Jiao, S., Zhang, Y., Jin, X.: Deformable 3D Gaussians for High-fidelity Monocular Dynamic Scene Reconstruction. In: CVPR (2024)", + "73. Yu, A., Li, R., Tancik, M., Li, H., Ng, R., Kanazawa, A.: PlenOctrees for Real-time Rendering of Neural Radiance Fields. In: ICCV (2021)", + "74. Zhao, F., Yang, W., Zhang, J., Lin, P., Zhang, Y., Yu, J., Xu, L.: HumanNeRF: Efficiently Generated Human Radiance Field from Sparse Inputs. In: CVPR (2022)", + "75. Zhou, K., Li, W., Wang, Y., Hu, T., Jiang, N., Han, X., Lu, J.: NeRFLiX: High-quality Neural View Synthesis by Learning a Degradation-driven Inter-viewpoint MiXer. In: CVPR (2023)", + "76. Zhou, T., Tucker, R., Flynn, J., Fyffe, G., Snavely, N.: Stereo Magnification: Learning View Synthesis using Multiplane Images. ACM Trans. Graph. 37 (2018)" + ], + "bbox": [ + 215, + 146, + 784, + 549 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "R. Shaw et al.", + "bbox": [ + 271, + 114, + 367, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/SWinGS_ Sliding Windows for Dynamic 3D Gaussian Splatting/1089e742-f963-4d64-a493-ab6447fb05de_model.json b/2024/SWinGS_ Sliding Windows for Dynamic 3D Gaussian Splatting/1089e742-f963-4d64-a493-ab6447fb05de_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6a82e0337ead43e4e0a9ea2201ee537c9798564f --- /dev/null +++ b/2024/SWinGS_ Sliding Windows for Dynamic 3D Gaussian Splatting/1089e742-f963-4d64-a493-ab6447fb05de_model.json @@ -0,0 +1,2469 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.317, + 0.141, + 0.687, + 0.187 + ], + "angle": 0, + "content": "SWinGS: Sliding Windows for Dynamic 3D Gaussian Splatting" + }, + { + "type": "text", + "bbox": [ + 0.25, + 0.212, + 0.756, + 0.245 + ], + "angle": 0, + "content": "Richard Shaw1, Michal Nazarczuk1, Jifei Song1, Arthur Moreau1, Sibi Catley-Chandar1,2, Helisa Dhamo1, and Eduardo Pérez-Pellitero1" + }, + { + "type": "text", + "bbox": [ + 0.376, + 0.255, + 0.625, + 0.284 + ], + "angle": 0, + "content": "\\(^{1}\\) Huawei Noah's Ark Lab \n\\(^{2}\\) Queen Mary University Of London" + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.327, + 0.396, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.396, + 0.327, + 0.566, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.578, + 0.327, + 0.781, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.469, + 0.788, + 0.512 + ], + "angle": 0, + "content": "Fig. 1: Left: SWinGS achieves sharper dynamic 3D scene reconstruction in part thanks to a sliding window canonical space that reduces the complexity of the 3D motion estimation. Right: Our dynamic real-time viewer allows users to explore the scene." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.563, + 0.741, + 0.841 + ], + "angle": 0, + "content": "Abstract. Novel view synthesis has shown rapid progress recently, with methods capable of producing increasingly photorealistic results. 3D Gaussian Splatting has emerged as a promising method, producing high-quality renderings of scenes and enabling interactive viewing at real-time frame rates. However, it is limited to static scenes. In this work, we extend 3D Gaussian Splatting to reconstruct dynamic scenes. We model a scene's dynamics using dynamic MLPs, learning deformations from temporally-local canonical representations to per-frame 3D Gaussians. To disentangle static and dynamic regions, tuneable parameters weigh each Gaussian's respective MLP parameters, improving the dynamics modelling of imbalanced scenes. We introduce a sliding window training strategy that partitions the sequence into smaller manageable windows to handle arbitrary length scenes while maintaining high rendering quality. We propose an adaptive sampling strategy to determine appropriate window size hyperparameters based on the scene's motion, balancing training overhead with visual quality. Training a separate dynamic 3D Gaussian model for each sliding window allows the canonical representation to change, enabling the reconstruction of scenes with significant geometric changes. Temporal consistency is enforced using a fine-tuning step with self-supervising consistency loss on randomly sampled novel" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.128 + ], + "angle": 0, + "content": "R. Shaw et al." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.148, + 0.744, + 0.19 + ], + "angle": 0, + "content": "views. As a result, our method produces high-quality renderings of general dynamic scenes with competitive quantitative performance, which can be viewed in real-time in our dynamic interactive viewer." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.21, + 0.376, + 0.226 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.236, + 0.787, + 0.357 + ], + "angle": 0, + "content": "Photorealistic rendering and generally 3-dimensional (3D) imaging have received significant attention in recent years, especially since the seminal work of Neural Radiance Fields (NeRF) [37]. This is in part thanks to its impressive novel view synthesis results, but also due to its appealing ease of use when coupled with off-the-shelf structure-from-motion camera pose estimation [53]. NeRF's key insight is a fully differentiable volumetric rendering pipeline paired with learnable implicit functions that model a view-dependent 3D radiance field. Dense coverage of posed images of the scene provides then direct photometric supervision." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.357, + 0.787, + 0.629 + ], + "angle": 0, + "content": "The original formulation of NeRF and most follow-ups [5,9,37,39,41] assume static scenes and thus a fixed radiance field. Some have explored new paradigms enabling dynamic reconstruction for radiance fields, including D-NeRF [22] and Nerfies [43], optimising an additional continuous volumetric deformation field that warps each observed point into a canonical NeRF. Such an approach has been popular [12,13,28,32,33,44], and has also been used for dynamic human reconstruction [45,74]. However, learning 3D deformation fields is inherently challenging, especially for large motions, with increased computational expense in training and inference. Moreover, approaches that share a canonical space among all frames struggle to maintain reconstruction quality for long sequences, obtaining overly blurred results due to inaccurate deformations and limited representational capacity. Other methods avoid maintaining a canonical representation and use explicit per-frame representations of the dynamic scene. Examples are tri-plane extensions to 4D \\((x,y,z,t)\\) [7,14,54,60] with plane decompositions [9] to keep memory footprint under control. These approaches can suffer from a lack of temporal consistency, especially as they generally are agnostic about the motion of the scene. Grid-based methods that share a representation [46], also can suffer from degradation due to lack of representational capacity in long sequences." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.787, + 0.841 + ], + "angle": 0, + "content": "This paper proposes a new method that addresses open problems of the state-of-the-art (SoTA) - see Fig. 1. Our method overview is shown in Fig. 2. Firstly, we build upon 3D Gaussian Splatting (3DGS) [20] and adapt the 3D Gaussians to be dynamic by allowing them to move. Our representation is thus explicit and avoids expensive raymarching via fast rasterization. Secondly, we introduce a novel paradigm for dynamic neural rendering with temporally-local canonical spaces defined in a sliding window fashion. Each window's length is adaptively defined following the amount of scene motion to maintain high-quality reconstruction. By limiting the scope of each canonical space, we can accurately track 3D displacements (i.e. they are generally smaller displacements) and prevent intra-window flickering. Thirdly, we introduce tuneable MLPs (MLP with several sets of weights governed by per-3D Gaussian blending weights) to estimate displacements. This tackles scenes with static vs dynamic imbalance. By learning different \"modes\" of motion estimation, we can separate smoothly between" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.671, + 0.114, + 0.732, + 0.128 + ], + "angle": 0, + "content": "SWinGS" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.208 + ], + "angle": 0, + "content": "static and dynamic regions with virtually no additional computational cost nor any handcrafted heuristics. Lastly, temporal consistency loss computed on overlapping frames of neighbouring windows ensures consistency between windows, i.e. avoids inter-window flickering. In summary, our main contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.215, + 0.787, + 0.245 + ], + "angle": 0, + "content": "1. An adaptive sliding window approach that enables the reconstruction of arbitrary length sequences whilst maintaining high rendering quality." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.245, + 0.786, + 0.274 + ], + "angle": 0, + "content": "2. Temporally-local dynamic MLPs that model scene dynamics by learning deformation fields from per-window canonical 3D Gaussians to each frame." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.274, + 0.786, + 0.305 + ], + "angle": 0, + "content": "3. Learnable MLP tuning parameters tackle scene imbalance by learning different motion modes; disentangling static canonical and dynamic 3D Gaussians." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.305, + 0.786, + 0.32 + ], + "angle": 0, + "content": "4. A fine-tuning stage ensures temporal consistency throughout the sequence." + }, + { + "type": "list", + "bbox": [ + 0.223, + 0.215, + 0.787, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.342, + 0.383, + 0.358 + ], + "angle": 0, + "content": "2 Related work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.372, + 0.788, + 0.493 + ], + "angle": 0, + "content": "Non-NeRF dynamic reconstruction: A number of approaches prior to the emergence of NeRF [37] tackled dynamic scene reconstruction. Such methods typically relied on dense camera coverage for point tracking and reprojection [19], or the presence of additional measurements, i.e. depth [40, 55]. Alternatively, some were curated towards specific domains, e.g. car reconstruction in driving scenarios [6, 34]. Several recent works [3,29,46] followed the idea of Image-Based Rendering [63] (direct reconstruction from neighbouring views). Others [30, 69] utilize multiplane images [58, 76] with an additional temporal component." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.494, + 0.788, + 0.72 + ], + "angle": 0, + "content": "NeRF-based reconstruction: NeRF [37] has achieved great success for reconstructing static scenes with many works extending it to dynamic inputs. Several methods [4, 67] model separate representations per time-step, disregarding the temporal component of the input. D-NeRF [22] reconstruct the scene in a canonical representation and model temporal variations with a deformation field. This idea was developed upon by many works [12, 13, 28, 32, 33, 44]. StreamRF [23] and NeRFPlayer [56] use time-aware MLPs and a compact 3D grid at each time-step for 4D field representation, reducing memory cost for longer videos. Other approaches [7, 14, 54, 60] represent dynamic scenes with a space-time grid, with grid decomposition to increase efficiency. DyNeRF [25] represents dynamic scenes by extending NeRF with an additional time-variant latent code. HyperReel [2] uses an efficient sampling network, modeling the scene around keyframes. MixVoxels [61] represents dynamic and static components with separately processed voxels. Several methods [18, 24, 64] rely on an underlying template mesh (e.g. human). Some methods [8, 50, 75] aim to improve NeRF quality post rendering." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "3D Gaussian Splatting Much of the development of neural rendering has focused on accelerating inference [15,46,49,62,73]. Recently, 3D Gaussian Splatting [20] made strides by modelling the scene with 3D Gaussians, which, when combined with tile-based differentiable rasterization, achieves very fast rendering, yet preserves high-quality reconstruction. Luiten et al. [35] extend this to dynamic scenes with shared 3D Gaussians that are optimised frame-by-frame. However, their focus is more on tracking 3D Gaussian trajectories rather than maximizing final rendering quality. Concurrent methods to ours that extend 3D Gaus-" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.128 + ], + "angle": 0, + "content": "R. Shaw et al." + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.143, + 0.413, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.415, + 0.143, + 0.588, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.59, + 0.143, + 0.772, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.295, + 0.789, + 0.366 + ], + "angle": 0, + "content": "Fig. 2: Method. First, the sequence is partitioned into sliding windows based on optical flow. Second, a dynamic 3DGS model is trained per window, where tunable MLPs model the deformations. Blending parameters \\(\\alpha\\) weigh the MLP's parameters to focus on dynamic parts. Finally, each model is fine-tuned, enforcing inter-window temporal consistency with consistency loss on sampled views for overlapping frames." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.397, + 0.789, + 0.445 + ], + "angle": 0, + "content": "sian Splatting to focus on general dynamic scenes include [17, 26, 31, 57, 71, 72] amongst others, while other works have focused specifically on dynamic human reconstruction [16, 21, 27, 38, 42, 48, 65] and facial animation [11, 47, 52, 68, 70]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.47, + 0.331, + 0.487 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.507, + 0.342, + 0.521 + ], + "angle": 0, + "content": "3.1 Overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.537, + 0.789, + 0.597 + ], + "angle": 0, + "content": "We present our method to reconstruct and render novel views of general dynamic scenes from multiple calibrated time-synchronized cameras. An overview of our method is shown in Fig. 2. We build upon 3D Gaussian Splatting [20] for novel view synthesis of static scenes, which we extend to scenes containing motion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.789, + 0.719 + ], + "angle": 0, + "content": "Our method can be separated into three main steps. First, given a dynamic sequence, we split the sequence into separate shorter sliding windows of frames for concurrent processing. We adaptively sample windows of varying lengths depending on the amount of motion in the sequence. Each sliding window contains an overlapping frame between adjacent windows, enabling temporal consistency to be enforced throughout the sequence at a subsequent training stage. Partitioning the sequence into smaller windows enables us to deal with sequences of arbitrary length while maintaining high render quality." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.79, + 0.84 + ], + "angle": 0, + "content": "Second, we train separate dynamic 3DGS models for each sliding window in turn. We extend the static 3DGS method to model the dynamics by introducing a tuneable MLP [36]. The MLP learns the deformation field from a canonical set of 3D Gaussians for each frame in a window. Thus each window comprises an independent temporally-local canonical 3D Gaussian representation and deformation field. This enables us to handle significant geometric changes and/or if new objects appear throughout the sequence, which can be challenging to model with a single representation. A tuneable MLP weighting parameter \\(\\alpha\\) is learned" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.671, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SWinGS" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "for each 3D Gaussian to enable the MLP to focus on modelling the dynamic parts of the scene, with the static parts encapsulated by the canonical representation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.329 + ], + "angle": 0, + "content": "Third, once a dynamic 3DGS model is trained for each window in the sequence, we apply a fine-tuning step to enforce temporal consistency throughout the sequence. We fine-tune each 3DGS model sequentially and employ a self-supervising temporal consistency loss on the overlapping frame renders between neighbouring windows. This encourages the model of the current window to produce similar renderings to the previous window. The result is a set of per-frame Gaussian Splatting models enabling high-quality novel view renderings of dynamic scenes with real-time interactive viewing capability. Our approach enables us to overcome the limitations of training with long sequences and to handle complex motions without exhibiting distracting temporal flickering." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.351, + 0.557, + 0.367 + ], + "angle": 0, + "content": "3.2 Preliminary: 3D Gaussian Splitting" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.376, + 0.784, + 0.438 + ], + "angle": 0, + "content": "Our method is built upon 3D Gaussian Splatting (3DGS) [20]. 3DGS uses a 3D Gaussian representation to model scenes as they are differentiable and can be projected to 2D splats, enabling fast tile-based rasterization. The 3D Gaussians are defined by 3D covariance matrix \\(\\Sigma\\) in world space centered at the mean \\(\\mu\\):" + }, + { + "type": "equation", + "bbox": [ + 0.423, + 0.449, + 0.785, + 0.469 + ], + "angle": 0, + "content": "\\[\nG (x) = e ^ {- \\frac {1}{2} (x) ^ {T} \\Sigma^ {- 1} (x)}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.475, + 0.788, + 0.597 + ], + "angle": 0, + "content": "Given a scaling matrix \\( S \\) and rotation matrix \\( R \\), the corresponding covariance matrix \\( \\Sigma \\) of a 3D Gaussian can be written as \\( \\Sigma = RSS^T R^T \\). To represent a scene, 3DGS optimizes 3D Gaussian positions \\( \\mathbf{x} \\), covariance \\( \\Sigma \\) (scaling \\( S \\) and rotation \\( R \\)), opacity \\( o \\), and colours, represented by spherical harmonic (SH) coefficients, capturing view-dependent appearance. The optimization is interleaved with Gaussian adaptive density control (ADC). The model is optimized by rendering the learned Gaussians via a differentiable rasterizer, comparing the resulting image \\( I_r \\) against the ground truth \\( I_{gt} \\), and minimizing the loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.358, + 0.612, + 0.785, + 0.628 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = (1 - \\lambda) \\mathcal {L} _ {1} \\left(I _ {r}, I _ {g t}\\right) + \\lambda \\mathcal {L} _ {\\text {S S I M}} \\left(I _ {r}, I _ {g t}\\right). \\tag {2}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.649, + 0.488, + 0.665 + ], + "angle": 0, + "content": "3.3 Sliding-Window Processing" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Using a single dynamic 3DGS representation to model an entire sequence becomes impractical for longer sequences, simply from a data processing standpoint. Furthermore, representing a lengthy sequence with a single model performs significantly worse than using multiple smaller fixed-sized segments, particularly if the scene has large motion or topological changes that cannot be modeled easily with one canonical representation and deformation field (Fig. 4). To address this, we use a sliding-window approach, separating the sequence into smaller windows with overlapping frames. The window size is a hyperparameter whose effect is explored in Section 4.1. Each window comprises an independent dynamic 3DGS model, and we allow all 3D Gaussian parameters to change between windows, including the number of Gaussians, their positions, rotations," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.128 + ], + "angle": 0, + "content": "R. Shaw et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "scaling, colours and opacities. The advantage of this approach is that independent models can be trained in parallel across multiple GPUs to speed up training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.195, + 0.785, + 0.301 + ], + "angle": 0, + "content": "Adaptive Window Sampling We propose an adaptive method for choosing the appropriate sliding window sizes, balancing training overhead and model size with performance and temporal consistency. Given an input sequence of frame length \\( N_{f} \\), we separate the sequence into smaller windows by adaptively sampling windows of different lengths \\( \\{N_w\\} \\), depending on the amount of motion in the sequence. In high-motion areas, we aim to sample windows more frequently (shorter windows), and in low-motion areas, less frequently (longer windows)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.302, + 0.787, + 0.392 + ], + "angle": 0, + "content": "To do this, we leverage the magnitude of 2D optical flow from each camera viewpoint. We employ a greedy algorithm to adaptively select the sizes of windows, prior to training, based on the accumulated optical flow magnitude. Specifically, we estimate per-frame optical flow \\(\\pmb{f}\\) using a pre-trained RAFT [59] model for each camera view \\(j\\in V\\) and all frames in the sequence \\(i\\in N_{f}\\), and compute the mean flow magnitude summed over each frame:" + }, + { + "type": "equation", + "bbox": [ + 0.393, + 0.405, + 0.786, + 0.447 + ], + "angle": 0, + "content": "\\[\n\\hat {v} _ {i} = \\frac {1}{V} \\sum_ {j} ^ {V} \\sum_ {i} ^ {N _ {f} - 1} | | \\boldsymbol {f} \\left(I _ {i} ^ {j}, I _ {i + 1} ^ {j}\\right) | | _ {2} ^ {2} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.45, + 0.788, + 0.588 + ], + "angle": 0, + "content": "We iterate over each frame in the sequence with a greedy heuristic; spawning a new window when the sum of mean flow \\(\\hat{v}_i\\) exceeds a pre-defined threshold. This ensures that each window contains a similar amount of movement, leading to a balanced distribution of the total representational workload. Taking the average across viewpoints makes this approach somewhat invariant to the number of cameras, while placing a limit on the total flow stops an excessive amount of movement within each window. Note, each sampled window overlaps with the next window by a single frame, such that neighbouring windows share a common image frame. This is to enable inter-window temporal consistency (section 3.6)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.604, + 0.455, + 0.619 + ], + "angle": 0, + "content": "3.4 Dynamic 3D Gaussians" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.624, + 0.788, + 0.747 + ], + "angle": 0, + "content": "To extend static 3DGS to handle dynamic scenes, we introduce a temporally-local dynamic MLP unique to each sliding window in the sequence (section 3.3). Each time-dependent MLP learns a temporally-local deformation field, mapping from a learned per-window canonical space to a set of 3D Gaussians for each frame in the window. Each deformation field, represented by a small MLP \\(\\mathcal{F}_{\\theta}\\) with weights \\(\\theta\\), takes as input the normalized frame time \\(t \\in [0,1]\\) and 3D Gaussian means \\(\\pmb{x}\\) (normalized by the scene's mean and standard deviation), and outputs displacements to their positions \\(\\Delta \\pmb{x}\\), rotations \\(\\Delta \\pmb{r}\\) and scaling \\(\\Delta s\\):" + }, + { + "type": "equation", + "bbox": [ + 0.373, + 0.762, + 0.786, + 0.777 + ], + "angle": 0, + "content": "\\[\n\\Delta \\boldsymbol {x} (t), \\Delta \\boldsymbol {r} (t), \\Delta \\boldsymbol {s} (t) = \\mathcal {F} _ {\\theta} (\\gamma (\\boldsymbol {x}), \\gamma (t)) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.779, + 0.788, + 0.841 + ], + "angle": 0, + "content": "where \\(\\gamma(.)\\) denotes sinusoidal positional encoding \\(\\gamma: \\mathbb{R}^3 \\to \\mathbb{R}^{3+6m}\\), \\(\\gamma(x) = (x, \\ldots, \\sin(2k\\pi x), \\cos(2k\\pi x), \\ldots)\\). We use small MLPs to reduce overfitting, setting the number of frequency components \\(m = 6\\), MLP depth \\(D = 4\\) and width \\(W = 16\\), with two skip connections." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.671, + 0.114, + 0.732, + 0.128 + ], + "angle": 0, + "content": "SWinGS" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.465, + 0.162 + ], + "angle": 0, + "content": "3.5 Tunable Dynamic MLPs" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.172, + 0.788, + 0.262 + ], + "angle": 0, + "content": "Ideally, time-dependent MLPs model motion in the scene by associating the dynamic parts with the temporal input component \\(\\gamma(t)\\) and thus learn to decouple the scene into i) static canonical 3D Gaussians and ii) dynamic 3D Gaussians. However, in scenes with an imbalance of static vs dynamic parts, e.g. Neural 3D Video [25] with mostly static backgrounds, the MLP struggles to disentangle static and dynamic regions, resulting in poorly estimated deformation fields." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.262, + 0.79, + 0.399 + ], + "angle": 0, + "content": "To counteract this, we introduce tuneable dynamic MLPs [36] (i.e. MLPs with \\(M\\) sets of weights) governed by \\(M\\) sets of learnable blending parameters \\(\\{\\pmb{\\alpha} \\in \\mathbb{R}^{M \\times N_g}\\}\\). The tuning parameters weigh the respective parameters of the MLP for each input Gaussian \\(i \\in N_g\\), enabling the learning of \\(M\\) modes of variation corresponding to different motion modes. With \\(M = 2\\), the MLP has the ability to decouple static and dynamic Gaussians in a smoothly weighted manner. Given the set of blending parameters \\(\\{\\alpha_i\\}_{m=1}^M\\) for the \\(i\\)-th input Gaussian \\(\\pmb{x}_i\\), the output of a single layer of the dynamic MLP \\(\\pmb{y}_i\\) can be written as a weighted sum over the \\(M\\) sets of weights:" + }, + { + "type": "equation", + "bbox": [ + 0.365, + 0.411, + 0.786, + 0.451 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {y} _ {i} = \\phi \\left(\\sum_ {m = 1} ^ {M} \\left(\\alpha_ {i, m} \\boldsymbol {w} _ {m} ^ {T} \\boldsymbol {x} _ {i} + \\alpha_ {i, m} \\boldsymbol {b} _ {m}\\right)\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.456, + 0.789, + 0.562 + ], + "angle": 0, + "content": "where \\(\\boldsymbol{w}\\) and \\(\\boldsymbol{b}\\) are the weights and bias, and \\(\\phi\\) is a non-linear activation function. The blending parameters \\(\\{\\alpha_{i}\\}_{m = 1}^{M}\\) linearly blend \\(M\\) sets of weights and biases of each layer of the MLP for each input Gaussian which, when passed through the activation function, enables nonlinear interaction between \\(\\alpha\\) and the output. Applying \\(\\alpha\\) in this way naturally enables the learning different of motion modes with a single forward pass of the MLP. Thus the dynamic MLP \\(\\mathcal{F}_{\\theta}^{dyn}\\) becomes a function of position, time, and blending parameters \\(\\alpha\\):" + }, + { + "type": "equation", + "bbox": [ + 0.353, + 0.576, + 0.786, + 0.593 + ], + "angle": 0, + "content": "\\[\n\\Delta \\boldsymbol {x} (t), \\Delta \\boldsymbol {r} (t), \\Delta \\boldsymbol {s} (t) = \\mathcal {F} _ {\\theta} ^ {d y n} (\\gamma (\\boldsymbol {x}), \\gamma (t), \\boldsymbol {\\alpha}). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.788, + 0.734 + ], + "angle": 0, + "content": "We implement the dynamic MLP as a single batch matrix multiplication (see supplementary). Blending parameters \\(\\alpha\\) are initialized as a binary mask of dynamic Gaussians (0-static, 1-dynamic) as follows. For a sliding window, we project all 3D Gaussians into each camera view: \\(u = \\varPi_{j}(\\pmb{x})\\), obtaining their 2D-pixel coordinates. We compute their L1 pixel differences from the central frame to each frame in the window. If the difference is larger than a threshold, we label that Gaussian 1, otherwise 0. To be robust to occlusions and mislabelling, we average the assigned label over all views and frames in the window. If the average is greater than 0.5, we initialize the Gaussian dynamic, otherwise static." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Once initialized, we set the blending parameter \\(\\alpha\\) as a learnable parameter, and optimize it via back-propagation together with the rest of the system. We assume \\(\\alpha\\) is constant for all frames in a sliding window to reduce the complexity of the optimization, which is a fair assumption given our adaptive window sampling mechanism. Fig. 3 visualizes the learning of the MLP tuning parameter. We observe \\(\\alpha\\) providing higher weight to Gaussians likely to be dynamic. Fig. 3 (right) plots the magnitudes of resulting displacements \\(\\Delta x\\) output by the MLP." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.128 + ], + "angle": 0, + "content": "R. Shaw et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.788, + 0.178 + ], + "angle": 0, + "content": "Note, \\(\\alpha\\) does not have to be entirely accurate, as the MLP learns to adjust accordingly, yet enables the MLP to handle highly imbalanced scenes (Table 6)." + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.204, + 0.489, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.496, + 0.204, + 0.785, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.298, + 0.789, + 0.356 + ], + "angle": 0, + "content": "Fig. 3: Dynamic MLPs with tunable parameters \\(\\alpha\\) weigh the parameters of the MLP for each Gaussian. We show renders from two scenes, left: cook spinach [25] and right: Train [51]. Shown from left-to-right: image render, tunable \\(\\alpha\\) parameters, and normalized MLP displacements \\(\\Delta x\\). Note, \\(\\alpha\\) highlights the scene's dynamic regions." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.406, + 0.547, + 0.422 + ], + "angle": 0, + "content": "3.6 Temporal Consistency Fine-tuning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.431, + 0.788, + 0.551 + ], + "angle": 0, + "content": "Due to the non-deterministic nature of the 3DGS optimization, independently trained models for each sliding window may produce slightly different results. When the resulting renders from each model are played back in sequence, noticeable flickering can be observed, particularly in novel views (though it's less noticeable in the training views). Consequently, after training models for all sliding windows separately, we tackle inter-window temporal flickering by introducing a fine-tuning step for temporal consistency. This step uses a self-supervising loss function to aid in smooth transitions between the models of separate windows." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.788, + 0.675 + ], + "angle": 0, + "content": "We fine-tune each model for a short period (3000 iterations in our experiments). We initiate the process from the first sliding window and progress through the sequence sequentially. For a window comprising \\( N_{w} \\) frames, we load the trained model and the model from the preceding window, with one overlapping frame between them. We then freeze all the parameters of the previous model. In order to fine-tune the model, as the flickering is mainly observed in novel views, we randomly sample novel test views in between the training views by rigidly interpolating the training camera poses \\( P_{j} = [R|\\pmb{t}] \\in \\mathbb{R}^{4 \\times 4} \\) in SE(3):" + }, + { + "type": "equation", + "bbox": [ + 0.393, + 0.684, + 0.787, + 0.726 + ], + "angle": 0, + "content": "\\[\nP _ {\\text {n o v e l}} = \\exp_ {\\mathrm {M}} \\sum_ {j} ^ {V} \\beta_ {j} \\log_ {\\mathrm {M}} \\left(P _ {j}\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.733, + 0.785, + 0.766 + ], + "angle": 0, + "content": "where \\(\\exp_{\\mathbf{M}}\\) and \\(\\log_{\\mathbf{M}}\\) are the matrix exponential and logarithm [1] respectively, and \\(\\beta_{j}\\in [0,1]\\) is a uniformly sampled weighting such that \\(\\sum_{j}^{V}\\beta_{j} = 1\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.767, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In one fine-tuning step, we render the overlapping frames from a randomly sampled novel viewpoint using both the model of the current window \\( w \\) and the previous window \\( w - 1 \\). This means we use the first frame of current window \\( I_{t = 0}^{w} \\) and the last frame of the previous window \\( I_{t = N_w - 1}^{w - 1} \\). We then apply a consistency loss, which is simply the L1 loss on the two image renders from both models:" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.671, + 0.114, + 0.732, + 0.128 + ], + "angle": 0, + "content": "SWinGS" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "equation", + "bbox": [ + 0.386, + 0.16, + 0.786, + 0.18 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c o n s i s t e n c y}} = \\left| I _ {t = 0} ^ {w} - I _ {t = N _ {w} - 1} ^ {w - 1} \\right| _ {1} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.184, + 0.788, + 0.29 + ], + "angle": 0, + "content": "When fine-tuning a model, we only allow the canonical (static) representation to change, while freezing the weights of the time-dependent dynamic MLP. Since the canonical set of 3D Gaussians is shared for all frames in a sliding window, we need to be careful not to negatively impact other frames when refining the overlapping frame. To address this, we use an alternating strategy for refinement, enforcing temporal consistency on the overlapping frame \\(75\\%\\) of the time and training with the remaining views and frames as usual for the remaining \\(25\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.31, + 0.457, + 0.325 + ], + "angle": 0, + "content": "3.7 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.333, + 0.789, + 0.637 + ], + "angle": 0, + "content": "We implement our method in PyTorch, building upon the codebase and differentiable rasterizer of 3DGS [20]. We initialize each model with a point cloud obtained from COLMAP [53]. Each dynamic 3DGS model is trained for 15K iterations for all sampled windows in a sequence. The initial 2K iterations comprise a warm-up stage; training with only the central window frame and freezing the weights of the MLP, allowing the canonical representation to stabilize. We optimize the Gaussians' positions, rotations, scaling, opacities and SH coefficients. Afterwards, we unfreeze the MLP and allow the deformation field and tuning parameters \\(\\alpha\\) to optimize. We find this staggered optimization approach leads to better convergence. The number of Gaussians densifies for 8K iterations, after which the number of Gaussians is fixed. Inputs to the MLP are normalized by the mean and standard deviation of the canonical point cloud post-warm-up stage before frequency encoding, leading to faster and stabler convergence. We train with Adam optimizer, using different learning rates for each parameter following the implementation of [20]. The learning rate for the MLP and \\(\\alpha\\) parameters are set to 1e-4, with the MLP learning rate undergoing exponential decay by factor 1e-2 in 20K iterations. During the initial optimization phase, due to the independence of each Gaussian model, we train each model in parallel on eight 32Gb Tesla V100 GPUs to speed up training. Afterwards, we perform temporal fine-tuning of each model sequentially using a single GPU for 3K iterations each." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.674, + 0.528, + 0.718 + ], + "angle": 0, + "content": "Table 2: Quantitative results on Technicolor dataset [51] at full resolution. Best and second best results are highlighted." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.72, + 0.538, + 0.809 + ], + "angle": 0, + "content": "
MethodPSNRSSIMLPIPSFPS
DyNeRF [25]31.800.9110.1420.02
HypeRreel [2]32.500.9020.1130.45
Dynamic3DG [35]27.020.8320.22886.96
Ours33.650.9340.11723.79
" + }, + { + "type": "image", + "bbox": [ + 0.558, + 0.664, + 0.778, + 0.756 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.543, + 0.766, + 0.788, + 0.808 + ], + "angle": 0, + "content": "Fig. 4: Comparison in performance consistency for Ours and Dynamic3DG in consecutive frames." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.128 + ], + "angle": 0, + "content": "R. Shaw et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.19 + ], + "angle": 0, + "content": "Table 1: Quantitative results on the Neural 3D Video dataset [25], averaged over all scenes. Best and second best highlighted. Our method performs best overall whilst enabling real-time frame rates. Per-scene breakdown of results are given in Table 3." + }, + { + "type": "table", + "bbox": [ + 0.343, + 0.201, + 0.656, + 0.337 + ], + "angle": 0, + "content": "
MethodPSNRSSIMLPIPSFPS
MixVoxels [61]30.420.9230.1244.30
K-Planes [14]30.630.9220.1170.33
HexPlane [7]30.000.9220.1130.24
HyperReel [2]30.780.9310.1013.60
NeRFPlayer [56]30.700.9310.1210.10
StreamRF [23]30.230.9040.1779.40
4DGS [65]27.610.9160.13530.00
Ours31.100.9400.09671.51
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.363, + 0.327, + 0.379 + ], + "angle": 0, + "content": "4 Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.397, + 0.784, + 0.426 + ], + "angle": 0, + "content": "We evaluate our method on two real-world multi-view dynamic benchmarks: the Neural 3D Video dataset [25] and the Technicolor dataset [51]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.428, + 0.785, + 0.577 + ], + "angle": 0, + "content": "Neural 3D Video comprises real-world dynamic scenes captured with a time-synchronized multi-view system at \\(2028 \\times 2704\\) resolution at 30 FPS. Camera parameters are estimated using COLMAP [53]. We compare our method to K-Planes [14], HexPlane [7], MixVoxels [61], HyperReel [2], NeRFPlayer [56], StreamRF [23], and 4DGS [65]. We compute quantitative metrics for the central test view at half the original resolution (\\(1014 \\times 1352\\)) for 300 frames. The average results for each method over the dataset are given in Table 1, while Table 3 provides a breakdown of the per-scene performance. The results show that our method performs best regarding PSNR and SSIM metrics while offering the fastest rendering performance. Qualitative results are shown in Fig. 5." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.579, + 0.785, + 0.653 + ], + "angle": 0, + "content": "Technicolor captures real dynamic scenes from a synchronized \\(4 \\times 4\\) camera array at \\(2048 \\times 1088\\) resolution. Following [2], we evaluate on the second row second column camera on five scenes: Birthday, Fabien, Painter, Theater and Trains. We compare to DynNeRF [25], HyperReel [2] and Dynamic 3D Gaussians [35], with quantitative and qualitative results in Table 4 and Fig. 6 respectively." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.678, + 0.402, + 0.692 + ], + "angle": 0, + "content": "4.1 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.787, + 0.841 + ], + "angle": 0, + "content": "This section provides ablations showing the effectiveness of our sliding window and self-supervised temporal consistency fine-tuning strategies. As we train independent models for each window, flickering artifacts can occur in the final renders. To visualize this, Fig. 8 plots absolute image error between renders of neighbouring frames (overlapping frames outlined in red). Without temporal consistency, we observe a spike in absolute error on overlapping frames, resulting in undesirable flickering. However, after fine-tuning, the error in overlapping frames is drastically reduced. In Table 5, we compute per-frame image metrics and estimate a measure of the temporal consistency using SoTA video quality" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.672, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SWinGS" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.176, + 0.787, + 0.234 + ], + "angle": 0, + "content": "Table 3: Quantitative results on the Neural 3D Video dataset [25], evaluated for 300 frames at \\(1352 \\times 1014\\) resolution. †As reported in [2], ‡Natively trained in lower resolution, upscaled. Best and second best results highlighted respectively. Our method performs competitively in all metrics, usually coming in either first or second place." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.246, + 0.787, + 0.51 + ], + "angle": 0, + "content": "
Scene\nMethodCook SpinachCut Roast BeefFlame Steak
PSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPS
MixVoxels [61]31.390.9310.11331.380.9280.11130.150.9380.108
K-Planes [14]31.230.9260.11431.870.9280.11431.490.9400.102
HexPlane‡ [7]31.050.9280.11430.830.9270.11530.420.9390.104
HyperReel [2]31.770.9320.09032.250.9360.08631.480.9390.083
NeRFPlayer† [56]30.580.9290.11329.350.9080.14431.930.9500.088
StreamRF [23]30.890.9140.16230.750.9170.15431.370.9230.152
Ours31.960.9460.09431.840.9450.09932.180.9530.087
Sear steakCoffee MartiniFlame Salmon
MixVoxels [61]30.850.9400.10329.250.9010.14729.500.8980.163
K-Planes [14]30.280.9370.10429.300.9000.13429.580.9010.132
HexPlane‡ [7]30.000.9390.10528.450.8910.14929.230.9050.088
HyperReel [2]31.880.9420.08028.650.8970.12928.260.9410.136
NeRFPlayer† [56]29.130.9080.13831.530.9510.08531.650.9400.098
StreamRF [23]31.600.9250.14728.130.8730.21928.690.8720.229
Ours32.210.9500.09229.160.9210.10529.250.9250.100
" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.581, + 0.785, + 0.611 + ], + "angle": 0, + "content": "Table 4: Quantitative results on the Technicolor dataset [51] evaluated at full resolution. Best and second best results highlighted respectively." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.624, + 0.787, + 0.805 + ], + "angle": 0, + "content": "
SceneBirthdayFabienPainter
MethodPSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPS
DyNeRF [25]29.200.9090.06732.760.9090.24235.950.9300.147
HyperReel [2]30.790.9220.06232.280.8600.21735.680.9260.123
Dynamic3DG [35]27.060.8590.11326.340.8340.26825.180.7580.395
Ours33.440.9590.04234.430.9250.17136.760.9480.128
SceneTheaterTrainAverage
DyNeRF [25]29.530.8750.18831.580.9330.06731.800.9110.142
HyperReel [2]33.670.8950.10430.100.9090.06132.500.9020.113
Dynamic3DG [35]28.050.7990.27728.460.9080.08827.020.8320.228
Ours29.810.8840.20133.990.9570.04333.690.9340.117
" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.127 + ], + "angle": 0, + "content": "R. Shaw et al." + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.161, + 0.74, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.471, + 0.788, + 0.501 + ], + "angle": 0, + "content": "Fig. 5: Qualitative results on Neural 3D Video [25]. Scenes top to bottom: i) coffee martini, ii) cook spinach, iii) cut roasted beef, iv) flame salmon, v) flame steak." + }, + { + "type": "image", + "bbox": [ + 0.254, + 0.543, + 0.737, + 0.779 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.79, + 0.785, + 0.818 + ], + "angle": 0, + "content": "Fig. 6: Qualitative results on Technicolor [51]. Scenes top to bottom: i) Birthday, ii) Painter, iii) Train." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.671, + 0.114, + 0.732, + 0.128 + ], + "angle": 0, + "content": "SWinGS" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "assessment metric FAST-VQA [66], where the quality score is in the range [0,1]. The table provides results averaged over all scenes from the Neural 3D Video dataset [25]. Although we incur a minor penalty in some per-frame performance metrics (SSIM and LPIPS), we obtain a significantly higher video quality (VQA) score. This indicates a substantial improvement in temporal consistency and overall perceptual video quality, resulting in more pleasing renderings." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.262, + 0.79, + 0.319 + ], + "angle": 0, + "content": "Table 5: Ablation on temporal fine-tuning. Results averaged over all scenes from [25]. Temporal consistency is measured using t-LPIPS [10] and FAST-VQA [66] (quality score in the range [0,1]). Per-frame performance remains fairly constant, but temporal consistency and overall perceptual video quality is significantly improved." + }, + { + "type": "table", + "bbox": [ + 0.252, + 0.33, + 0.747, + 0.394 + ], + "angle": 0, + "content": "
MethodPSNR ↑SSIM ↑LPIPS ↓t-LPIPS ↓VQA ↑
w/o temporal consistency32.010.9560.0850.01290.666
w/ temporal consistency32.050.9490.0930.01020.726
Ground truth----0.763
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.444, + 0.79, + 0.504 + ], + "angle": 0, + "content": "Table 6: Ablation on sliding window size vs adaptive on Birthday scene [51]. We show the improvement from the dynamic MLP. Adaptive chooses window sizes that match, and in some metrics exceed, the best fixed-size window performance, striking balance between performance, temporal consistency (t-LPIPS) and training time (GPU hrs)." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.524, + 0.789, + 0.632 + ], + "angle": 0, + "content": "
Window SizeNo. WindowsTrain timePSNRSSIMLPIPSt-LPIPS
32416.033.120.9560.0480.0076
964.033.380.9590.0430.0053
1732.033.010.9560.0450.0049
2521.332.970.9560.0430.0048
4910.732.730.9550.0470.0051
Adaptive53.333.440.9590.0420.0051
Adaptive (w/o dyn. MLP)53.332.760.9570.0450.0062
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.657, + 0.788, + 0.765 + ], + "angle": 0, + "content": "Table 6 and Fig. 7 show the impact of the sliding-window size hyperparameter and our adaptive sampling strategy. Adaptive sampling automatically chooses appropriate window sizes that match and sometimes exceed the best fixed-size window performance, striking a good balance between performance, temporal consistency (t-LPIPS [10]) and training time. Adaptive performs the best overall with fewer windows (less storage requirements). We also show the advantage gained from using a dynamic MLP vs a regular MLP, which improves all metrics." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.783, + 0.36, + 0.799 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "We have presented a method to render novel views of dynamic scenes by extending the 3DGS framework. Results show that our method produces high-" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.127 + ], + "angle": 0, + "content": "R. Shaw et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.787, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.275, + 0.788, + 0.332 + ], + "angle": 0, + "content": "Fig. 7: Ablation on sliding window size vs adaptive window sampling (with and without dynamic MLP) on scenes from the Technicolor dataset [51]. The single-window scenario (w49) has the shortest training time but unsatisfactory visual quality, while adaptive windows offer the best balance between quality and training overhead." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.351, + 0.787, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.538, + 0.788, + 0.583 + ], + "angle": 0, + "content": "Fig. 8: Ablation on temporal fine-tuning: we display the absolute error between neighbouring frame renders, with overlapping frames highlighted red. After fine-tuning, the error is substantially reduced and overall perceptual video quality is improved." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.789, + 0.842 + ], + "angle": 0, + "content": "quality renderings, even with complex motions e.g. flames. Key to our approach is sliding-window processing, which adaptively partitions sequences into manageable chunks. Processing each window separately, allowing the canonical representation and deformation field to vary throughout the sequence, enables us to handle complex topological changes, and reduces the magnitude and variability of the 3D scene flow. In contrast, other methods that learn a single representation for whole sequences are impractical for long sequences and degrade in quality with increasing length. Introducing an MLP for each window learns the deformation field from each canonical representation to a set of per-frame 3D Gaussians. Moreover, learnable tuning parameters help disentangle static and dynamic parts of the scene, which we find essential for imbalanced scenes. Our ablations show self-supervised temporal consistency fine-tuning reduces temporal flickering and improves the overall perceptual video quality, with only a minor impact on per-frame performance metrics. Overall, our method performs strongly compared to recent SoTA quantitatively and obtains sharper, temporally-consistent results." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.672, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SWinGS" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.772, + 0.19 + ], + "angle": 0, + "content": "1. Alexa, M.: Linear Combination of Transformations. ACM Trans. Graph. (2002)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.192, + 0.786, + 0.232 + ], + "angle": 0, + "content": "2. Attal, B., Huang, J.B., Richardt, C., Zollhöfer, M., Kopf, J., O'Toole, M., Kim, C.: HyperReel: High-fidelity 6-DoF Video with Ray-conditioned Sampling. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.233, + 0.786, + 0.26 + ], + "angle": 0, + "content": "3. Bansal, A., Vo, M., Sheikh, Y., Ramanan, D., Narasimhan, S.: 4D Visualization of Dynamic Events from Unconstrained Multi-view Videos. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.261, + 0.786, + 0.286 + ], + "angle": 0, + "content": "4. Bansal, A., Zollhöfer, M.: Neural Pixel Composition for 3D-4D View Synthesis from Multi-views. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.288, + 0.786, + 0.328 + ], + "angle": 0, + "content": "5. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-NeRF: A Multiscale Representation for Anti-aliasing Neural Radiance Fields. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.33, + 0.786, + 0.37 + ], + "angle": 0, + "content": "6. Bārsan, I.A., Liu, P., Pollefeys, M., Geiger, A.: Robust Dense Mapping for Large-scale Dynamic Environments. In: International Conference on Robotics and Automation (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.371, + 0.786, + 0.397 + ], + "angle": 0, + "content": "7. Cao, A., CV, J.: HexPlane: A Fast Representation for Dynamic Scenes. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.399, + 0.786, + 0.439 + ], + "angle": 0, + "content": "8. Catley-Chandar, S., Shaw, R., Slabaugh, G., Pérez-Pellitero, E.: RoGUENeRF: A Robust Geometry-consistent Universal Enhancer for NeRF. arXiv preprint arXiv:2403.11909 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.44, + 0.786, + 0.466 + ], + "angle": 0, + "content": "9. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: TensoRF: Tensorial Radiance Fields. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.468, + 0.786, + 0.508 + ], + "angle": 0, + "content": "0. Chu, M., Xie, Y., Mayer, J., Leal-Taix'e, L., Thuerey, N.: Learning Temporal Coherence via Self-supervision for GAN-based Video Generation. ACM Trans. Graph. 39, 75:1 - 75:13 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.509, + 0.786, + 0.55 + ], + "angle": 0, + "content": "1. Dhamo, H., Nie, Y., Moreau, A., Song, J., Shaw, R., Zhou, Y., Pérez-Pellitero, E.: HeadGaS: Real-time Animatable Head Avatars via 3D Gaussian Splatting. arXiv preprint arXiv:2312.02902 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.551, + 0.786, + 0.577 + ], + "angle": 0, + "content": "2. Du, Y., Zhang, Y., Yu, H.X., Tenenbaum, J.B., Wu, J.: Neural Radiance Flow for 4D View Synthesis and Video Processing. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.578, + 0.786, + 0.619 + ], + "angle": 0, + "content": "3. Fang, J., Yi, T., Wang, X., Xie, L., Zhang, X., Liu, W., Nießner, M., Tian, Q.: Fast Dynamic Radiance Fields with Time-aware Neural Voxels. In: SIGGRAPH Asia (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.62, + 0.786, + 0.646 + ], + "angle": 0, + "content": "4. Fridovich-Keil, S., Meanti, G., Warburg, F.R., Recht, B., Kanazawa, A.: K-Planes: Explicit Radiance Fields in Space, Time, and Appearance. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.647, + 0.786, + 0.674 + ], + "angle": 0, + "content": "5. Hedman, P., Srinivasan, P.P., Mildenhall, B., Barron, J.T., Debevec, P.: Baking Neural Radiance Fields for Real-time View Synthesis. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.675, + 0.786, + 0.715 + ], + "angle": 0, + "content": "6. Hu, L., Zhang, H., Zhang, Y., Zhou, B., Liu, B., Zhang, S., Nie, L.: GaussianAvatar: Towards Realistic Human Avatar Modeling from a Single Video via Animatable 3D Gaussians. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.716, + 0.786, + 0.744 + ], + "angle": 0, + "content": "7. Huang, Y.H., Sun, Y.T., Yang, Z., Lyu, X., Cao, Y.P., Qi, X.: SC-GS: Sparse-Controlled Gaussian Splatting for Editable Dynamic Scenes. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.745, + 0.786, + 0.785 + ], + "angle": 0, + "content": "8. Isik, M., Rünz, M., Georgopoulos, M., Khakhulin, T., Starck, J., Agapito, L., Niešner, M.: HumanRF: High-fidelity Neural Radiance Fields for Humans in Motion. ACM Trans. Graph. 42(4), 1-12 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.786, + 0.786, + 0.812 + ], + "angle": 0, + "content": "9. Joo, H., Soo Park, H., Sheikh, Y.: MAP Visibility Estimation for Large-scale Dynamic 3D Reconstruction. In: CVPR (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.786, + 0.84 + ], + "angle": 0, + "content": "20. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3D Gaussian Splatting for Real-time Radiance Field Rendering. ACM Trans. Graph. 42(4) (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.177, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.127 + ], + "angle": 0, + "content": "R. Shaw et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "21. Kocabas, M., Chang, J.H.R., Gabriel, J., Tuzel, O., Ranjan, A.: HUGS: Human Gaussian Splatting. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.175, + 0.785, + 0.202 + ], + "angle": 0, + "content": "22. Li, L., Shen, Z., Wang, Z., Shen, L., Tan, P.: D-NeRF: Neural Radiance Fields for Dynamic Scenes. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.203, + 0.785, + 0.229 + ], + "angle": 0, + "content": "23. Li, L., Shen, Z., Wang, Z., Shen, L., Tan, P.: Streaming Radiance Fields for 3D Video Synthesis. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.23, + 0.785, + 0.256 + ], + "angle": 0, + "content": "24. Li, R., Tanke, J., Vo, M., Zollhöfer, M., Gall, J., Kanazawa, A., Lassner, C.: TAVA: Template-free Animatable Volumetric Actors. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.257, + 0.785, + 0.297 + ], + "angle": 0, + "content": "25. Li, T., Slavcheva, M., Zollhöfer, M., Green, S., Lassner, C., Kim, C., Schmidt, T., Lovegrove, S., Goesele, M., Newcombe, R., Lv, Z.: Neural 3D Video Synthesis from Multi-view Video. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.298, + 0.785, + 0.325 + ], + "angle": 0, + "content": "26. Li, Z., Chen, Z., Li, Z., Xu, Y.: Spacetime Gaussian Feature Splatting for Real-time Dynamic View Synthesis. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.326, + 0.785, + 0.365 + ], + "angle": 0, + "content": "27. Li, Z., Zheng, Z., Wang, L., Liu, Y.: Animatable Gaussians: Learning Pose-dependent Gaussian Maps for High-fidelity Human Avatar Modeling. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.366, + 0.785, + 0.393 + ], + "angle": 0, + "content": "28. Li, Z., Niklaus, S., Snavely, N., Wang, O.: Neural Scene Flow Fields for Space-time View Synthesis of Dynamic Scenes. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.394, + 0.785, + 0.419 + ], + "angle": 0, + "content": "29. Li, Z., Wang, Q., Cole, F., Tucker, R., Snavely, N.: DynIBaR: Neural Dynamic Image-Based Rendering. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.42, + 0.785, + 0.447 + ], + "angle": 0, + "content": "30. Lin, K.E., Xiao, L., Liu, F., Yang, G., Ramamoorthi, R.: Deep 3D Mask Volume for View Synthesis of Dynamic Scenes. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.447, + 0.785, + 0.473 + ], + "angle": 0, + "content": "31. Lin, Y., Dai, Z., Zhu, S., Yao, Y.: Gaussian-Flow: 4D Reconstruction with Dynamic 3D Gaussian Particle. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.474, + 0.785, + 0.501 + ], + "angle": 0, + "content": "32. Liu, Y.L., Gao, C., Meuleman, A., Tseng, H.Y., Saraf, A., Kim, C., Chuang, Y.Y., Kopf, J., Huang, J.B.: Robust Dynamic Radiance Fields. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.502, + 0.785, + 0.542 + ], + "angle": 0, + "content": "33. Lombardi, S., Simon, T., Saragih, J., Schwartz, G., Lehrmann, A., Sheikh, Y.: Neural Volumes: Learning Dynamic Renderable Volumes from Images. ACM Trans. Graph. 38(4) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.543, + 0.785, + 0.569 + ], + "angle": 0, + "content": "34. Luiten, J., Fischer, T., Leibe, B.: Track to Reconstruct and Reconstruct to Track. IEEE Robotics and Automation Letters 5(2), 1803-1810 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.569, + 0.785, + 0.596 + ], + "angle": 0, + "content": "35. Luiten, J., Kopanas, G., Leibe, B., Ramanan, D.: Dynamic 3D Gaussians: Tracking by Persistent Dynamic View Synthesis. In: 3DV (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.597, + 0.785, + 0.623 + ], + "angle": 0, + "content": "36. Maggioni, M., Tanay, T., Babiloni, F., McDonagh, S.G., Leonardis, A.: Tunable Convolutions with Parametric Multi-loss Optimization. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.624, + 0.785, + 0.663 + ], + "angle": 0, + "content": "37. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.664, + 0.785, + 0.69 + ], + "angle": 0, + "content": "38. Moreau, A., Song, J., Dhamo, H., Shaw, R., Zhou, Y., Pérez-Pellitero, E.: Human Gaussian Splatting: Real-time Rendering of Animatable Avatars. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.691, + 0.785, + 0.731 + ], + "angle": 0, + "content": "39. Müller, T., Evans, A., Schied, C., Keller, A.: Instant Neural Graphics Primitives with a Multiresolution Hash Encoding. ACM Trans. Graph. 41(4), 102:1-102:15 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.732, + 0.785, + 0.759 + ], + "angle": 0, + "content": "40. Newcombe, R.A., Fox, D., Seitz, S.M.: DynamicFusion: Reconstruction and Tracking of Non-rigid Scenes in Real-time. In: CVPR (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.785, + 0.8 + ], + "angle": 0, + "content": "41. Niemeyer, M., Barron, J.T., Mildenhall, B., Sajjadi, M.S.M., Geiger, A., Radwan, N.: RegNeRF: Regularizing Neural Radiance Fields for View Synthesis from Sparse Inputs. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "42. Pang, H., Zhu, H., Kortylewski, A., Theobalt, C., Habermann, M.: ASH: Animatable Gaussian Splats for Efficient and Photoreal Human Rendering. In: CVPR (2024)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.671, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SWinGS" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "43. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin-Brualla, R.: Nerfies: Deformable Neural Radiance Fields. ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.218 + ], + "angle": 0, + "content": "44. Park, K., Sinha, U., Hedman, P., Barron, J.T., Bouaziz, S., Goldman, D.B., Martin-Brualla, R., Seitz, S.M.: HyperNeRF: A Higher-dimensional Representation for Topologically Varying Neural Radiance Fields. ACM Trans. Graph. 40(6) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.219, + 0.785, + 0.246 + ], + "angle": 0, + "content": "45. Peng, S., Dong, J., Wang, Q., Zhang, S., Shuai, Q., Zhou, X., Bao, H.: Animatable Neural Radiance Fields for Modeling Dynamic Human Bodies. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.247, + 0.785, + 0.274 + ], + "angle": 0, + "content": "46. Peng, S., Yan, Y., Shuai, Q., Bao, H., Zhou, X.: Representing Volumetric Videos as Dynamic MLP Maps. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.276, + 0.785, + 0.316 + ], + "angle": 0, + "content": "47. Qian, S., Kirschstein, T., Schoneveld, L., Davoli, D., Giebenhain, S., Nießner, M.: GaussianAvatars: Photorealistic Head Avatars with Rigged 3D Gaussians. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.317, + 0.785, + 0.346 + ], + "angle": 0, + "content": "48. Qian, Z., Wang, S., Mihajlovic, M., Geiger, A., Tang, S.: 3DGS-Avatar: Animatable Avatars via Deformable 3D Gaussian Splitting. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.346, + 0.785, + 0.373 + ], + "angle": 0, + "content": "49. Reiser, C., Peng, S., Liao, Y., Geiger, A.: KiloNeRF: Speeding up Neural Radiance Fields with Thousands of Tiny MLPs. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.374, + 0.785, + 0.402 + ], + "angle": 0, + "content": "50. Rong, X., Huang, J.B., Saraf, A., Kim, C., Kopf, J.: Boosting View Synthesis with Residual Transfer. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.403, + 0.785, + 0.43 + ], + "angle": 0, + "content": "51. Sabater, N., Boisson, G., Vandame, B., Kerbiriou, P., Babon, F., et al.: Dataset and Pipeline for Multi-view Light-Field Video. In: CVPRW (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.431, + 0.785, + 0.459 + ], + "angle": 0, + "content": "52. Saito, S., Schwartz, G., Simon, T., Li, J., Nam, G.: Relightable Gaussian Codec Avatars. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.46, + 0.785, + 0.487 + ], + "angle": 0, + "content": "53. Schonberger, J.L., Frahm, J.M.: Structure-from-Motion Revisited. In: CVPR (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.785, + 0.529 + ], + "angle": 0, + "content": "54. Shao, R., Zheng, Z., Tu, H., Liu, B., Zhang, H., Liu, Y.: Tensor4D: Efficient Neural 4D Decomposition for High-fidelity Dynamic Reconstruction and Rendering. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.53, + 0.785, + 0.558 + ], + "angle": 0, + "content": "55. Slavcheva, M., Baust, M., Cremers, D., Ilic, S.: KillingFusion: Non-rigid 3D Reconstruction without Correspondences. In: CVPR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.559, + 0.785, + 0.613 + ], + "angle": 0, + "content": "56. Song, L., Chen, A., Li, Z., Chen, Z., Chen, L., Yuan, J., Xu, Y., Geiger, A.: NeRF-Player: A Streamable Dynamic Scene Representation with Decomposed Neural Radiance Fields. IEEE Transactions on Visualization and Computer Graphics 29(5), 2732-2742 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.615, + 0.785, + 0.655 + ], + "angle": 0, + "content": "57. Sun, J., Jiao, H., Li, G., Zhang, Z., Zhao, L., Xing, W.: 3DGStream: On-the-fly Training of 3D Gaussians for Efficient Streaming of Photo-realistic Free-viewpoint Videos. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.657, + 0.752, + 0.671 + ], + "angle": 0, + "content": "58. Tanay, T., Maggioni, M.: Global Latent Neural Rendering. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.672, + 0.785, + 0.699 + ], + "angle": 0, + "content": "59. Teed, Z., Deng, J.: RAFT: Recurrent All-pairs Field Transforms for Optical Flow. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.7, + 0.785, + 0.727 + ], + "angle": 0, + "content": "60. Turki, H., Zhang, J.Y., Ferroni, F., Ramanan, D.: SUDS: Scalable Urban Dynamic Scenes. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.728, + 0.785, + 0.756 + ], + "angle": 0, + "content": "61. Wang, F., Tan, S., Li, X., Tian, Z., Liu, H.: Mixed Neural Voxels for Fast Multiview Video Synthesis. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.757, + 0.785, + 0.798 + ], + "angle": 0, + "content": "62. Wang, L., Zhang, J., Liu, X., Zhao, F., Zhang, Y., Zhang, Y., Wu, M., Yu, J., Xu, L.: Fourier PlenOctrees for Dynamic Radiance Field Rendering in Real-time. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "63. Wang, Q., Wang, Z., Genova, K., Srinivasan, P., Zhou, H., Barron, J.T., Martin-Brualla, R., Snavely, N., Funkhouser, T.: IBRNet: Learning Multi-view Image-Based Rendering. In: CVPR (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.127 + ], + "angle": 0, + "content": "R. Shaw et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.188 + ], + "angle": 0, + "content": "64. Weng, C.Y., Curless, B., Srinivasan, P.P., Barron, J.T., Kemelmacher-Shlizerman, I.: HumanNeRF: Free-viewpoint Rendering of Moving People from Monocular Video. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.189, + 0.785, + 0.217 + ], + "angle": 0, + "content": "65. Wu, G., Yi, T., Fang, J., Xie, L., Zhang, X., Wei, W., Liu, W., Tian, Q., Wang, X.: 4D Gaussian Splatting for Real-time Dynamic Scene Rendering. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.258 + ], + "angle": 0, + "content": "66. Wu, H., Chen, C., Hou, J., Liao, L., Wang, A., Sun, W., Yan, Q., Lin, W.: FASTVQA: Efficient End-to-end Video Quality Assessment with Fragment Sampling. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.785, + 0.285 + ], + "angle": 0, + "content": "67. Xian, W., Huang, J.B., Kopf, J., Kim, C.: Space-time Neural Irradiance Fields for Free-viewpoint Video. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.313 + ], + "angle": 0, + "content": "68. Xiang, J., Gao, X., Guo, Y., Zhang, J.: FlashAvatar: High-fidelity Head Avatar with Efficient Gaussian Embedding. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.341 + ], + "angle": 0, + "content": "69. Xing, W., Chen, J.: Temporal-MPI: Enabling Multi-Plane Images for Dynamic Scene Modelling via Temporal Basis Learning. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.785, + 0.37 + ], + "angle": 0, + "content": "70. Xu, Y., Chen, B., Li, Z., Zhang, H., Wang, L., Zheng, Z., Liu, Y.: Gaussian HeadAvatar: Ultra High-fidelity Head Avatar via Dynamic Gaussians. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.397 + ], + "angle": 0, + "content": "71. Yang, Z., Yang, H., Pan, Z., Zhang, L.: Real-time Photorealistic Dynamic Scene Representation and Rendering with 4D Gaussian Splatting. In: ICLR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.424 + ], + "angle": 0, + "content": "72. Yang, Z., Gao, X., Zhou, W., Jiao, S., Zhang, Y., Jin, X.: Deformable 3D Gaussians for High-fidelity Monocular Dynamic Scene Reconstruction. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.452 + ], + "angle": 0, + "content": "73. Yu, A., Li, R., Tancik, M., Li, H., Ng, R., Kanazawa, A.: PlenOctrees for Real-time Rendering of Neural Radiance Fields. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.479 + ], + "angle": 0, + "content": "74. Zhao, F., Yang, W., Zhang, J., Lin, P., Zhang, Y., Yu, J., Xu, L.: HumanNeRF: Efficiently Generated Human Radiance Field from Sparse Inputs. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.48, + 0.785, + 0.521 + ], + "angle": 0, + "content": "75. Zhou, K., Li, W., Wang, Y., Hu, T., Jiang, N., Han, X., Lu, J.: NeRFLiX: High-quality Neural View Synthesis by Learning a Degradation-driven Inter-viewpoint MiXer. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.785, + 0.55 + ], + "angle": 0, + "content": "76. Zhou, T., Tucker, R., Flynn, J., Fyffe, G., Snavely, N.: Stereo Magnification: Learning View Synthesis using Multiplane Images. ACM Trans. Graph. 37 (2018)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.55 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/SWinGS_ Sliding Windows for Dynamic 3D Gaussian Splatting/1089e742-f963-4d64-a493-ab6447fb05de_origin.pdf b/2024/SWinGS_ Sliding Windows for Dynamic 3D Gaussian Splatting/1089e742-f963-4d64-a493-ab6447fb05de_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..31029c7b31be61ecb41b58c17beb0438f4039d55 --- /dev/null +++ b/2024/SWinGS_ Sliding Windows for Dynamic 3D Gaussian Splatting/1089e742-f963-4d64-a493-ab6447fb05de_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e0181d97e49063864f5042f010159df1a5b8836d69407a87455d5959e225536c +size 7717356 diff --git a/2024/SWinGS_ Sliding Windows for Dynamic 3D Gaussian Splatting/full.md b/2024/SWinGS_ Sliding Windows for Dynamic 3D Gaussian Splatting/full.md new file mode 100644 index 0000000000000000000000000000000000000000..65422c5945eefad0056efd0fab9dc05fe94ef7ef --- /dev/null +++ b/2024/SWinGS_ Sliding Windows for Dynamic 3D Gaussian Splatting/full.md @@ -0,0 +1,298 @@ +# SWinGS: Sliding Windows for Dynamic 3D Gaussian Splatting + +Richard Shaw1, Michal Nazarczuk1, Jifei Song1, Arthur Moreau1, Sibi Catley-Chandar1,2, Helisa Dhamo1, and Eduardo Pérez-Pellitero1 + +$^{1}$ Huawei Noah's Ark Lab + $^{2}$ Queen Mary University Of London + +![](images/adfaab72d4811650a89fe9f926206654dce088cc185571b363b4081714d95df5.jpg) +Fig. 1: Left: SWinGS achieves sharper dynamic 3D scene reconstruction in part thanks to a sliding window canonical space that reduces the complexity of the 3D motion estimation. Right: Our dynamic real-time viewer allows users to explore the scene. + +![](images/e06659d5361df453a0466df373466c8ce4c5824e1fb721ac814c62ac8f6c29e9.jpg) + +![](images/868ae6435617946c74c592d7a37e5513eb5e35aae3b9b87cf619e5d183643913.jpg) + +Abstract. Novel view synthesis has shown rapid progress recently, with methods capable of producing increasingly photorealistic results. 3D Gaussian Splatting has emerged as a promising method, producing high-quality renderings of scenes and enabling interactive viewing at real-time frame rates. However, it is limited to static scenes. In this work, we extend 3D Gaussian Splatting to reconstruct dynamic scenes. We model a scene's dynamics using dynamic MLPs, learning deformations from temporally-local canonical representations to per-frame 3D Gaussians. To disentangle static and dynamic regions, tuneable parameters weigh each Gaussian's respective MLP parameters, improving the dynamics modelling of imbalanced scenes. We introduce a sliding window training strategy that partitions the sequence into smaller manageable windows to handle arbitrary length scenes while maintaining high rendering quality. We propose an adaptive sampling strategy to determine appropriate window size hyperparameters based on the scene's motion, balancing training overhead with visual quality. Training a separate dynamic 3D Gaussian model for each sliding window allows the canonical representation to change, enabling the reconstruction of scenes with significant geometric changes. Temporal consistency is enforced using a fine-tuning step with self-supervising consistency loss on randomly sampled novel + +views. As a result, our method produces high-quality renderings of general dynamic scenes with competitive quantitative performance, which can be viewed in real-time in our dynamic interactive viewer. + +# 1 Introduction + +Photorealistic rendering and generally 3-dimensional (3D) imaging have received significant attention in recent years, especially since the seminal work of Neural Radiance Fields (NeRF) [37]. This is in part thanks to its impressive novel view synthesis results, but also due to its appealing ease of use when coupled with off-the-shelf structure-from-motion camera pose estimation [53]. NeRF's key insight is a fully differentiable volumetric rendering pipeline paired with learnable implicit functions that model a view-dependent 3D radiance field. Dense coverage of posed images of the scene provides then direct photometric supervision. + +The original formulation of NeRF and most follow-ups [5,9,37,39,41] assume static scenes and thus a fixed radiance field. Some have explored new paradigms enabling dynamic reconstruction for radiance fields, including D-NeRF [22] and Nerfies [43], optimising an additional continuous volumetric deformation field that warps each observed point into a canonical NeRF. Such an approach has been popular [12,13,28,32,33,44], and has also been used for dynamic human reconstruction [45,74]. However, learning 3D deformation fields is inherently challenging, especially for large motions, with increased computational expense in training and inference. Moreover, approaches that share a canonical space among all frames struggle to maintain reconstruction quality for long sequences, obtaining overly blurred results due to inaccurate deformations and limited representational capacity. Other methods avoid maintaining a canonical representation and use explicit per-frame representations of the dynamic scene. Examples are tri-plane extensions to 4D $(x,y,z,t)$ [7,14,54,60] with plane decompositions [9] to keep memory footprint under control. These approaches can suffer from a lack of temporal consistency, especially as they generally are agnostic about the motion of the scene. Grid-based methods that share a representation [46], also can suffer from degradation due to lack of representational capacity in long sequences. + +This paper proposes a new method that addresses open problems of the state-of-the-art (SoTA) - see Fig. 1. Our method overview is shown in Fig. 2. Firstly, we build upon 3D Gaussian Splatting (3DGS) [20] and adapt the 3D Gaussians to be dynamic by allowing them to move. Our representation is thus explicit and avoids expensive raymarching via fast rasterization. Secondly, we introduce a novel paradigm for dynamic neural rendering with temporally-local canonical spaces defined in a sliding window fashion. Each window's length is adaptively defined following the amount of scene motion to maintain high-quality reconstruction. By limiting the scope of each canonical space, we can accurately track 3D displacements (i.e. they are generally smaller displacements) and prevent intra-window flickering. Thirdly, we introduce tuneable MLPs (MLP with several sets of weights governed by per-3D Gaussian blending weights) to estimate displacements. This tackles scenes with static vs dynamic imbalance. By learning different "modes" of motion estimation, we can separate smoothly between + +static and dynamic regions with virtually no additional computational cost nor any handcrafted heuristics. Lastly, temporal consistency loss computed on overlapping frames of neighbouring windows ensures consistency between windows, i.e. avoids inter-window flickering. In summary, our main contributions are: + +1. An adaptive sliding window approach that enables the reconstruction of arbitrary length sequences whilst maintaining high rendering quality. +2. Temporally-local dynamic MLPs that model scene dynamics by learning deformation fields from per-window canonical 3D Gaussians to each frame. +3. Learnable MLP tuning parameters tackle scene imbalance by learning different motion modes; disentangling static canonical and dynamic 3D Gaussians. +4. A fine-tuning stage ensures temporal consistency throughout the sequence. + +# 2 Related work + +Non-NeRF dynamic reconstruction: A number of approaches prior to the emergence of NeRF [37] tackled dynamic scene reconstruction. Such methods typically relied on dense camera coverage for point tracking and reprojection [19], or the presence of additional measurements, i.e. depth [40, 55]. Alternatively, some were curated towards specific domains, e.g. car reconstruction in driving scenarios [6, 34]. Several recent works [3,29,46] followed the idea of Image-Based Rendering [63] (direct reconstruction from neighbouring views). Others [30, 69] utilize multiplane images [58, 76] with an additional temporal component. + +NeRF-based reconstruction: NeRF [37] has achieved great success for reconstructing static scenes with many works extending it to dynamic inputs. Several methods [4, 67] model separate representations per time-step, disregarding the temporal component of the input. D-NeRF [22] reconstruct the scene in a canonical representation and model temporal variations with a deformation field. This idea was developed upon by many works [12, 13, 28, 32, 33, 44]. StreamRF [23] and NeRFPlayer [56] use time-aware MLPs and a compact 3D grid at each time-step for 4D field representation, reducing memory cost for longer videos. Other approaches [7, 14, 54, 60] represent dynamic scenes with a space-time grid, with grid decomposition to increase efficiency. DyNeRF [25] represents dynamic scenes by extending NeRF with an additional time-variant latent code. HyperReel [2] uses an efficient sampling network, modeling the scene around keyframes. MixVoxels [61] represents dynamic and static components with separately processed voxels. Several methods [18, 24, 64] rely on an underlying template mesh (e.g. human). Some methods [8, 50, 75] aim to improve NeRF quality post rendering. + +3D Gaussian Splatting Much of the development of neural rendering has focused on accelerating inference [15,46,49,62,73]. Recently, 3D Gaussian Splatting [20] made strides by modelling the scene with 3D Gaussians, which, when combined with tile-based differentiable rasterization, achieves very fast rendering, yet preserves high-quality reconstruction. Luiten et al. [35] extend this to dynamic scenes with shared 3D Gaussians that are optimised frame-by-frame. However, their focus is more on tracking 3D Gaussian trajectories rather than maximizing final rendering quality. Concurrent methods to ours that extend 3D Gaus- + +![](images/f0ba281755eecfee02e1b9aa96def15dc4e98e3654ff807e81e82e7683e023f2.jpg) +Fig. 2: Method. First, the sequence is partitioned into sliding windows based on optical flow. Second, a dynamic 3DGS model is trained per window, where tunable MLPs model the deformations. Blending parameters $\alpha$ weigh the MLP's parameters to focus on dynamic parts. Finally, each model is fine-tuned, enforcing inter-window temporal consistency with consistency loss on sampled views for overlapping frames. + +![](images/6849426cd786d7313c7aeddbc9fe257bf9cb780fc15ee020032617201022a6fe.jpg) + +![](images/c9c1409faba288cf9f3ff10840eede29f659a345966e1443b0d40e4a1b0a1865.jpg) + +sian Splatting to focus on general dynamic scenes include [17, 26, 31, 57, 71, 72] amongst others, while other works have focused specifically on dynamic human reconstruction [16, 21, 27, 38, 42, 48, 65] and facial animation [11, 47, 52, 68, 70]. + +# 3 Method + +# 3.1 Overview + +We present our method to reconstruct and render novel views of general dynamic scenes from multiple calibrated time-synchronized cameras. An overview of our method is shown in Fig. 2. We build upon 3D Gaussian Splatting [20] for novel view synthesis of static scenes, which we extend to scenes containing motion. + +Our method can be separated into three main steps. First, given a dynamic sequence, we split the sequence into separate shorter sliding windows of frames for concurrent processing. We adaptively sample windows of varying lengths depending on the amount of motion in the sequence. Each sliding window contains an overlapping frame between adjacent windows, enabling temporal consistency to be enforced throughout the sequence at a subsequent training stage. Partitioning the sequence into smaller windows enables us to deal with sequences of arbitrary length while maintaining high render quality. + +Second, we train separate dynamic 3DGS models for each sliding window in turn. We extend the static 3DGS method to model the dynamics by introducing a tuneable MLP [36]. The MLP learns the deformation field from a canonical set of 3D Gaussians for each frame in a window. Thus each window comprises an independent temporally-local canonical 3D Gaussian representation and deformation field. This enables us to handle significant geometric changes and/or if new objects appear throughout the sequence, which can be challenging to model with a single representation. A tuneable MLP weighting parameter $\alpha$ is learned + +for each 3D Gaussian to enable the MLP to focus on modelling the dynamic parts of the scene, with the static parts encapsulated by the canonical representation. + +Third, once a dynamic 3DGS model is trained for each window in the sequence, we apply a fine-tuning step to enforce temporal consistency throughout the sequence. We fine-tune each 3DGS model sequentially and employ a self-supervising temporal consistency loss on the overlapping frame renders between neighbouring windows. This encourages the model of the current window to produce similar renderings to the previous window. The result is a set of per-frame Gaussian Splatting models enabling high-quality novel view renderings of dynamic scenes with real-time interactive viewing capability. Our approach enables us to overcome the limitations of training with long sequences and to handle complex motions without exhibiting distracting temporal flickering. + +# 3.2 Preliminary: 3D Gaussian Splitting + +Our method is built upon 3D Gaussian Splatting (3DGS) [20]. 3DGS uses a 3D Gaussian representation to model scenes as they are differentiable and can be projected to 2D splats, enabling fast tile-based rasterization. The 3D Gaussians are defined by 3D covariance matrix $\Sigma$ in world space centered at the mean $\mu$ : + +$$ +G (x) = e ^ {- \frac {1}{2} (x) ^ {T} \Sigma^ {- 1} (x)}. \tag {1} +$$ + +Given a scaling matrix $S$ and rotation matrix $R$ , the corresponding covariance matrix $\Sigma$ of a 3D Gaussian can be written as $\Sigma = RSS^T R^T$ . To represent a scene, 3DGS optimizes 3D Gaussian positions $\mathbf{x}$ , covariance $\Sigma$ (scaling $S$ and rotation $R$ ), opacity $o$ , and colours, represented by spherical harmonic (SH) coefficients, capturing view-dependent appearance. The optimization is interleaved with Gaussian adaptive density control (ADC). The model is optimized by rendering the learned Gaussians via a differentiable rasterizer, comparing the resulting image $I_r$ against the ground truth $I_{gt}$ , and minimizing the loss function: + +$$ +\mathcal {L} = (1 - \lambda) \mathcal {L} _ {1} \left(I _ {r}, I _ {g t}\right) + \lambda \mathcal {L} _ {\text {S S I M}} \left(I _ {r}, I _ {g t}\right). \tag {2} +$$ + +# 3.3 Sliding-Window Processing + +Using a single dynamic 3DGS representation to model an entire sequence becomes impractical for longer sequences, simply from a data processing standpoint. Furthermore, representing a lengthy sequence with a single model performs significantly worse than using multiple smaller fixed-sized segments, particularly if the scene has large motion or topological changes that cannot be modeled easily with one canonical representation and deformation field (Fig. 4). To address this, we use a sliding-window approach, separating the sequence into smaller windows with overlapping frames. The window size is a hyperparameter whose effect is explored in Section 4.1. Each window comprises an independent dynamic 3DGS model, and we allow all 3D Gaussian parameters to change between windows, including the number of Gaussians, their positions, rotations, + +scaling, colours and opacities. The advantage of this approach is that independent models can be trained in parallel across multiple GPUs to speed up training. + +Adaptive Window Sampling We propose an adaptive method for choosing the appropriate sliding window sizes, balancing training overhead and model size with performance and temporal consistency. Given an input sequence of frame length $N_{f}$ , we separate the sequence into smaller windows by adaptively sampling windows of different lengths $\{N_w\}$ , depending on the amount of motion in the sequence. In high-motion areas, we aim to sample windows more frequently (shorter windows), and in low-motion areas, less frequently (longer windows). + +To do this, we leverage the magnitude of 2D optical flow from each camera viewpoint. We employ a greedy algorithm to adaptively select the sizes of windows, prior to training, based on the accumulated optical flow magnitude. Specifically, we estimate per-frame optical flow $\pmb{f}$ using a pre-trained RAFT [59] model for each camera view $j\in V$ and all frames in the sequence $i\in N_{f}$ , and compute the mean flow magnitude summed over each frame: + +$$ +\hat {v} _ {i} = \frac {1}{V} \sum_ {j} ^ {V} \sum_ {i} ^ {N _ {f} - 1} | | \boldsymbol {f} \left(I _ {i} ^ {j}, I _ {i + 1} ^ {j}\right) | | _ {2} ^ {2} \tag {3} +$$ + +We iterate over each frame in the sequence with a greedy heuristic; spawning a new window when the sum of mean flow $\hat{v}_i$ exceeds a pre-defined threshold. This ensures that each window contains a similar amount of movement, leading to a balanced distribution of the total representational workload. Taking the average across viewpoints makes this approach somewhat invariant to the number of cameras, while placing a limit on the total flow stops an excessive amount of movement within each window. Note, each sampled window overlaps with the next window by a single frame, such that neighbouring windows share a common image frame. This is to enable inter-window temporal consistency (section 3.6). + +# 3.4 Dynamic 3D Gaussians + +To extend static 3DGS to handle dynamic scenes, we introduce a temporally-local dynamic MLP unique to each sliding window in the sequence (section 3.3). Each time-dependent MLP learns a temporally-local deformation field, mapping from a learned per-window canonical space to a set of 3D Gaussians for each frame in the window. Each deformation field, represented by a small MLP $\mathcal{F}_{\theta}$ with weights $\theta$ , takes as input the normalized frame time $t \in [0,1]$ and 3D Gaussian means $\pmb{x}$ (normalized by the scene's mean and standard deviation), and outputs displacements to their positions $\Delta \pmb{x}$ , rotations $\Delta \pmb{r}$ and scaling $\Delta s$ : + +$$ +\Delta \boldsymbol {x} (t), \Delta \boldsymbol {r} (t), \Delta \boldsymbol {s} (t) = \mathcal {F} _ {\theta} (\gamma (\boldsymbol {x}), \gamma (t)) \tag {4} +$$ + +where $\gamma(.)$ denotes sinusoidal positional encoding $\gamma: \mathbb{R}^3 \to \mathbb{R}^{3+6m}$ , $\gamma(x) = (x, \ldots, \sin(2k\pi x), \cos(2k\pi x), \ldots)$ . We use small MLPs to reduce overfitting, setting the number of frequency components $m = 6$ , MLP depth $D = 4$ and width $W = 16$ , with two skip connections. + +# 3.5 Tunable Dynamic MLPs + +Ideally, time-dependent MLPs model motion in the scene by associating the dynamic parts with the temporal input component $\gamma(t)$ and thus learn to decouple the scene into i) static canonical 3D Gaussians and ii) dynamic 3D Gaussians. However, in scenes with an imbalance of static vs dynamic parts, e.g. Neural 3D Video [25] with mostly static backgrounds, the MLP struggles to disentangle static and dynamic regions, resulting in poorly estimated deformation fields. + +To counteract this, we introduce tuneable dynamic MLPs [36] (i.e. MLPs with $M$ sets of weights) governed by $M$ sets of learnable blending parameters $\{\pmb{\alpha} \in \mathbb{R}^{M \times N_g}\}$ . The tuning parameters weigh the respective parameters of the MLP for each input Gaussian $i \in N_g$ , enabling the learning of $M$ modes of variation corresponding to different motion modes. With $M = 2$ , the MLP has the ability to decouple static and dynamic Gaussians in a smoothly weighted manner. Given the set of blending parameters $\{\alpha_i\}_{m=1}^M$ for the $i$ -th input Gaussian $\pmb{x}_i$ , the output of a single layer of the dynamic MLP $\pmb{y}_i$ can be written as a weighted sum over the $M$ sets of weights: + +$$ +\boldsymbol {y} _ {i} = \phi \left(\sum_ {m = 1} ^ {M} \left(\alpha_ {i, m} \boldsymbol {w} _ {m} ^ {T} \boldsymbol {x} _ {i} + \alpha_ {i, m} \boldsymbol {b} _ {m}\right)\right), \tag {5} +$$ + +where $\boldsymbol{w}$ and $\boldsymbol{b}$ are the weights and bias, and $\phi$ is a non-linear activation function. The blending parameters $\{\alpha_{i}\}_{m = 1}^{M}$ linearly blend $M$ sets of weights and biases of each layer of the MLP for each input Gaussian which, when passed through the activation function, enables nonlinear interaction between $\alpha$ and the output. Applying $\alpha$ in this way naturally enables the learning different of motion modes with a single forward pass of the MLP. Thus the dynamic MLP $\mathcal{F}_{\theta}^{dyn}$ becomes a function of position, time, and blending parameters $\alpha$ : + +$$ +\Delta \boldsymbol {x} (t), \Delta \boldsymbol {r} (t), \Delta \boldsymbol {s} (t) = \mathcal {F} _ {\theta} ^ {d y n} (\gamma (\boldsymbol {x}), \gamma (t), \boldsymbol {\alpha}). \tag {6} +$$ + +We implement the dynamic MLP as a single batch matrix multiplication (see supplementary). Blending parameters $\alpha$ are initialized as a binary mask of dynamic Gaussians (0-static, 1-dynamic) as follows. For a sliding window, we project all 3D Gaussians into each camera view: $u = \varPi_{j}(\pmb{x})$ , obtaining their 2D-pixel coordinates. We compute their L1 pixel differences from the central frame to each frame in the window. If the difference is larger than a threshold, we label that Gaussian 1, otherwise 0. To be robust to occlusions and mislabelling, we average the assigned label over all views and frames in the window. If the average is greater than 0.5, we initialize the Gaussian dynamic, otherwise static. + +Once initialized, we set the blending parameter $\alpha$ as a learnable parameter, and optimize it via back-propagation together with the rest of the system. We assume $\alpha$ is constant for all frames in a sliding window to reduce the complexity of the optimization, which is a fair assumption given our adaptive window sampling mechanism. Fig. 3 visualizes the learning of the MLP tuning parameter. We observe $\alpha$ providing higher weight to Gaussians likely to be dynamic. Fig. 3 (right) plots the magnitudes of resulting displacements $\Delta x$ output by the MLP. + +Note, $\alpha$ does not have to be entirely accurate, as the MLP learns to adjust accordingly, yet enables the MLP to handle highly imbalanced scenes (Table 6). + +![](images/c3dcdb2f11a8ec061cf3864f723627d2eb0a6c004174315a99e39398c82a3011.jpg) +Fig. 3: Dynamic MLPs with tunable parameters $\alpha$ weigh the parameters of the MLP for each Gaussian. We show renders from two scenes, left: cook spinach [25] and right: Train [51]. Shown from left-to-right: image render, tunable $\alpha$ parameters, and normalized MLP displacements $\Delta x$ . Note, $\alpha$ highlights the scene's dynamic regions. + +![](images/355e0502329e36c80dec5967c0753a42ca8b3763eafae513c562decf22bd4b69.jpg) + +# 3.6 Temporal Consistency Fine-tuning + +Due to the non-deterministic nature of the 3DGS optimization, independently trained models for each sliding window may produce slightly different results. When the resulting renders from each model are played back in sequence, noticeable flickering can be observed, particularly in novel views (though it's less noticeable in the training views). Consequently, after training models for all sliding windows separately, we tackle inter-window temporal flickering by introducing a fine-tuning step for temporal consistency. This step uses a self-supervising loss function to aid in smooth transitions between the models of separate windows. + +We fine-tune each model for a short period (3000 iterations in our experiments). We initiate the process from the first sliding window and progress through the sequence sequentially. For a window comprising $N_{w}$ frames, we load the trained model and the model from the preceding window, with one overlapping frame between them. We then freeze all the parameters of the previous model. In order to fine-tune the model, as the flickering is mainly observed in novel views, we randomly sample novel test views in between the training views by rigidly interpolating the training camera poses $P_{j} = [R|\pmb{t}] \in \mathbb{R}^{4 \times 4}$ in SE(3): + +$$ +P _ {\text {n o v e l}} = \exp_ {\mathrm {M}} \sum_ {j} ^ {V} \beta_ {j} \log_ {\mathrm {M}} \left(P _ {j}\right) \tag {7} +$$ + +where $\exp_{\mathbf{M}}$ and $\log_{\mathbf{M}}$ are the matrix exponential and logarithm [1] respectively, and $\beta_{j}\in [0,1]$ is a uniformly sampled weighting such that $\sum_{j}^{V}\beta_{j} = 1$ . + +In one fine-tuning step, we render the overlapping frames from a randomly sampled novel viewpoint using both the model of the current window $w$ and the previous window $w - 1$ . This means we use the first frame of current window $I_{t = 0}^{w}$ and the last frame of the previous window $I_{t = N_w - 1}^{w - 1}$ . We then apply a consistency loss, which is simply the L1 loss on the two image renders from both models: + +$$ +\mathcal {L} _ {\text {c o n s i s t e n c y}} = \left| I _ {t = 0} ^ {w} - I _ {t = N _ {w} - 1} ^ {w - 1} \right| _ {1} \tag {8} +$$ + +When fine-tuning a model, we only allow the canonical (static) representation to change, while freezing the weights of the time-dependent dynamic MLP. Since the canonical set of 3D Gaussians is shared for all frames in a sliding window, we need to be careful not to negatively impact other frames when refining the overlapping frame. To address this, we use an alternating strategy for refinement, enforcing temporal consistency on the overlapping frame $75\%$ of the time and training with the remaining views and frames as usual for the remaining $25\%$ . + +# 3.7 Implementation Details + +We implement our method in PyTorch, building upon the codebase and differentiable rasterizer of 3DGS [20]. We initialize each model with a point cloud obtained from COLMAP [53]. Each dynamic 3DGS model is trained for 15K iterations for all sampled windows in a sequence. The initial 2K iterations comprise a warm-up stage; training with only the central window frame and freezing the weights of the MLP, allowing the canonical representation to stabilize. We optimize the Gaussians' positions, rotations, scaling, opacities and SH coefficients. Afterwards, we unfreeze the MLP and allow the deformation field and tuning parameters $\alpha$ to optimize. We find this staggered optimization approach leads to better convergence. The number of Gaussians densifies for 8K iterations, after which the number of Gaussians is fixed. Inputs to the MLP are normalized by the mean and standard deviation of the canonical point cloud post-warm-up stage before frequency encoding, leading to faster and stabler convergence. We train with Adam optimizer, using different learning rates for each parameter following the implementation of [20]. The learning rate for the MLP and $\alpha$ parameters are set to 1e-4, with the MLP learning rate undergoing exponential decay by factor 1e-2 in 20K iterations. During the initial optimization phase, due to the independence of each Gaussian model, we train each model in parallel on eight 32Gb Tesla V100 GPUs to speed up training. Afterwards, we perform temporal fine-tuning of each model sequentially using a single GPU for 3K iterations each. + +Table 2: Quantitative results on Technicolor dataset [51] at full resolution. Best and second best results are highlighted. + +
MethodPSNRSSIMLPIPSFPS
DyNeRF [25]31.800.9110.1420.02
HypeRreel [2]32.500.9020.1130.45
Dynamic3DG [35]27.020.8320.22886.96
Ours33.650.9340.11723.79
+ +![](images/53586640ec588eaa2592c01d4b8987d6e1e79d5e2727521a8cae144dd4a52479.jpg) +Fig. 4: Comparison in performance consistency for Ours and Dynamic3DG in consecutive frames. + +Table 1: Quantitative results on the Neural 3D Video dataset [25], averaged over all scenes. Best and second best highlighted. Our method performs best overall whilst enabling real-time frame rates. Per-scene breakdown of results are given in Table 3. + +
MethodPSNRSSIMLPIPSFPS
MixVoxels [61]30.420.9230.1244.30
K-Planes [14]30.630.9220.1170.33
HexPlane [7]30.000.9220.1130.24
HyperReel [2]30.780.9310.1013.60
NeRFPlayer [56]30.700.9310.1210.10
StreamRF [23]30.230.9040.1779.40
4DGS [65]27.610.9160.13530.00
Ours31.100.9400.09671.51
+ +# 4 Results + +We evaluate our method on two real-world multi-view dynamic benchmarks: the Neural 3D Video dataset [25] and the Technicolor dataset [51]. + +Neural 3D Video comprises real-world dynamic scenes captured with a time-synchronized multi-view system at $2028 \times 2704$ resolution at 30 FPS. Camera parameters are estimated using COLMAP [53]. We compare our method to K-Planes [14], HexPlane [7], MixVoxels [61], HyperReel [2], NeRFPlayer [56], StreamRF [23], and 4DGS [65]. We compute quantitative metrics for the central test view at half the original resolution ( $1014 \times 1352$ ) for 300 frames. The average results for each method over the dataset are given in Table 1, while Table 3 provides a breakdown of the per-scene performance. The results show that our method performs best regarding PSNR and SSIM metrics while offering the fastest rendering performance. Qualitative results are shown in Fig. 5. + +Technicolor captures real dynamic scenes from a synchronized $4 \times 4$ camera array at $2048 \times 1088$ resolution. Following [2], we evaluate on the second row second column camera on five scenes: Birthday, Fabien, Painter, Theater and Trains. We compare to DynNeRF [25], HyperReel [2] and Dynamic 3D Gaussians [35], with quantitative and qualitative results in Table 4 and Fig. 6 respectively. + +# 4.1 Ablation Studies + +This section provides ablations showing the effectiveness of our sliding window and self-supervised temporal consistency fine-tuning strategies. As we train independent models for each window, flickering artifacts can occur in the final renders. To visualize this, Fig. 8 plots absolute image error between renders of neighbouring frames (overlapping frames outlined in red). Without temporal consistency, we observe a spike in absolute error on overlapping frames, resulting in undesirable flickering. However, after fine-tuning, the error in overlapping frames is drastically reduced. In Table 5, we compute per-frame image metrics and estimate a measure of the temporal consistency using SoTA video quality + +Table 3: Quantitative results on the Neural 3D Video dataset [25], evaluated for 300 frames at $1352 \times 1014$ resolution. †As reported in [2], ‡Natively trained in lower resolution, upscaled. Best and second best results highlighted respectively. Our method performs competitively in all metrics, usually coming in either first or second place. + +
Scene +MethodCook SpinachCut Roast BeefFlame Steak
PSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPS
MixVoxels [61]31.390.9310.11331.380.9280.11130.150.9380.108
K-Planes [14]31.230.9260.11431.870.9280.11431.490.9400.102
HexPlane‡ [7]31.050.9280.11430.830.9270.11530.420.9390.104
HyperReel [2]31.770.9320.09032.250.9360.08631.480.9390.083
NeRFPlayer† [56]30.580.9290.11329.350.9080.14431.930.9500.088
StreamRF [23]30.890.9140.16230.750.9170.15431.370.9230.152
Ours31.960.9460.09431.840.9450.09932.180.9530.087
Sear steakCoffee MartiniFlame Salmon
MixVoxels [61]30.850.9400.10329.250.9010.14729.500.8980.163
K-Planes [14]30.280.9370.10429.300.9000.13429.580.9010.132
HexPlane‡ [7]30.000.9390.10528.450.8910.14929.230.9050.088
HyperReel [2]31.880.9420.08028.650.8970.12928.260.9410.136
NeRFPlayer† [56]29.130.9080.13831.530.9510.08531.650.9400.098
StreamRF [23]31.600.9250.14728.130.8730.21928.690.8720.229
Ours32.210.9500.09229.160.9210.10529.250.9250.100
+ +Table 4: Quantitative results on the Technicolor dataset [51] evaluated at full resolution. Best and second best results highlighted respectively. + +
SceneBirthdayFabienPainter
MethodPSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPS
DyNeRF [25]29.200.9090.06732.760.9090.24235.950.9300.147
HyperReel [2]30.790.9220.06232.280.8600.21735.680.9260.123
Dynamic3DG [35]27.060.8590.11326.340.8340.26825.180.7580.395
Ours33.440.9590.04234.430.9250.17136.760.9480.128
SceneTheaterTrainAverage
DyNeRF [25]29.530.8750.18831.580.9330.06731.800.9110.142
HyperReel [2]33.670.8950.10430.100.9090.06132.500.9020.113
Dynamic3DG [35]28.050.7990.27728.460.9080.08827.020.8320.228
Ours29.810.8840.20133.990.9570.04333.690.9340.117
+ +![](images/bcf93faffbc7ba5e1f00a23b627d57d3e8172798dcb6349ef25b0ba85d0f6d06.jpg) +Fig. 5: Qualitative results on Neural 3D Video [25]. Scenes top to bottom: i) coffee martini, ii) cook spinach, iii) cut roasted beef, iv) flame salmon, v) flame steak. + +![](images/21b4ecc10dd87b69a1ba7abf40eeb7e45537e7b3295d88bf06eba3b0671b091b.jpg) +Fig. 6: Qualitative results on Technicolor [51]. Scenes top to bottom: i) Birthday, ii) Painter, iii) Train. + +assessment metric FAST-VQA [66], where the quality score is in the range [0,1]. The table provides results averaged over all scenes from the Neural 3D Video dataset [25]. Although we incur a minor penalty in some per-frame performance metrics (SSIM and LPIPS), we obtain a significantly higher video quality (VQA) score. This indicates a substantial improvement in temporal consistency and overall perceptual video quality, resulting in more pleasing renderings. + +Table 5: Ablation on temporal fine-tuning. Results averaged over all scenes from [25]. Temporal consistency is measured using t-LPIPS [10] and FAST-VQA [66] (quality score in the range [0,1]). Per-frame performance remains fairly constant, but temporal consistency and overall perceptual video quality is significantly improved. + +
MethodPSNR ↑SSIM ↑LPIPS ↓t-LPIPS ↓VQA ↑
w/o temporal consistency32.010.9560.0850.01290.666
w/ temporal consistency32.050.9490.0930.01020.726
Ground truth----0.763
+ +Table 6: Ablation on sliding window size vs adaptive on Birthday scene [51]. We show the improvement from the dynamic MLP. Adaptive chooses window sizes that match, and in some metrics exceed, the best fixed-size window performance, striking balance between performance, temporal consistency (t-LPIPS) and training time (GPU hrs). + +
Window SizeNo. WindowsTrain timePSNRSSIMLPIPSt-LPIPS
32416.033.120.9560.0480.0076
964.033.380.9590.0430.0053
1732.033.010.9560.0450.0049
2521.332.970.9560.0430.0048
4910.732.730.9550.0470.0051
Adaptive53.333.440.9590.0420.0051
Adaptive (w/o dyn. MLP)53.332.760.9570.0450.0062
+ +Table 6 and Fig. 7 show the impact of the sliding-window size hyperparameter and our adaptive sampling strategy. Adaptive sampling automatically chooses appropriate window sizes that match and sometimes exceed the best fixed-size window performance, striking a good balance between performance, temporal consistency (t-LPIPS [10]) and training time. Adaptive performs the best overall with fewer windows (less storage requirements). We also show the advantage gained from using a dynamic MLP vs a regular MLP, which improves all metrics. + +# 5 Conclusion + +We have presented a method to render novel views of dynamic scenes by extending the 3DGS framework. Results show that our method produces high- + +![](images/c3697ee26a3e77e502cc4cee2086eb36ae715e7de5a7252cfccde31075409d3a.jpg) +Fig. 7: Ablation on sliding window size vs adaptive window sampling (with and without dynamic MLP) on scenes from the Technicolor dataset [51]. The single-window scenario (w49) has the shortest training time but unsatisfactory visual quality, while adaptive windows offer the best balance between quality and training overhead. + +![](images/b1339bc65944398bb3e4dfe6a041fb1a31b9789f04fde8b81c44606f1741a0d9.jpg) +Fig. 8: Ablation on temporal fine-tuning: we display the absolute error between neighbouring frame renders, with overlapping frames highlighted red. After fine-tuning, the error is substantially reduced and overall perceptual video quality is improved. + +quality renderings, even with complex motions e.g. flames. Key to our approach is sliding-window processing, which adaptively partitions sequences into manageable chunks. Processing each window separately, allowing the canonical representation and deformation field to vary throughout the sequence, enables us to handle complex topological changes, and reduces the magnitude and variability of the 3D scene flow. In contrast, other methods that learn a single representation for whole sequences are impractical for long sequences and degrade in quality with increasing length. Introducing an MLP for each window learns the deformation field from each canonical representation to a set of per-frame 3D Gaussians. Moreover, learnable tuning parameters help disentangle static and dynamic parts of the scene, which we find essential for imbalanced scenes. Our ablations show self-supervised temporal consistency fine-tuning reduces temporal flickering and improves the overall perceptual video quality, with only a minor impact on per-frame performance metrics. Overall, our method performs strongly compared to recent SoTA quantitatively and obtains sharper, temporally-consistent results. + +# References + +1. Alexa, M.: Linear Combination of Transformations. ACM Trans. Graph. (2002) +2. Attal, B., Huang, J.B., Richardt, C., Zollhöfer, M., Kopf, J., O'Toole, M., Kim, C.: HyperReel: High-fidelity 6-DoF Video with Ray-conditioned Sampling. In: CVPR (2023) +3. Bansal, A., Vo, M., Sheikh, Y., Ramanan, D., Narasimhan, S.: 4D Visualization of Dynamic Events from Unconstrained Multi-view Videos. In: CVPR (2020) +4. Bansal, A., Zollhöfer, M.: Neural Pixel Composition for 3D-4D View Synthesis from Multi-views. In: CVPR (2023) +5. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-NeRF: A Multiscale Representation for Anti-aliasing Neural Radiance Fields. In: ICCV (2021) +6. Bārsan, I.A., Liu, P., Pollefeys, M., Geiger, A.: Robust Dense Mapping for Large-scale Dynamic Environments. In: International Conference on Robotics and Automation (2018) +7. Cao, A., CV, J.: HexPlane: A Fast Representation for Dynamic Scenes. In: CVPR (2023) +8. Catley-Chandar, S., Shaw, R., Slabaugh, G., Pérez-Pellitero, E.: RoGUENeRF: A Robust Geometry-consistent Universal Enhancer for NeRF. arXiv preprint arXiv:2403.11909 (2024) +9. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: TensoRF: Tensorial Radiance Fields. In: ECCV (2022) +0. Chu, M., Xie, Y., Mayer, J., Leal-Taix'e, L., Thuerey, N.: Learning Temporal Coherence via Self-supervision for GAN-based Video Generation. ACM Trans. Graph. 39, 75:1 - 75:13 (2020) +1. Dhamo, H., Nie, Y., Moreau, A., Song, J., Shaw, R., Zhou, Y., Pérez-Pellitero, E.: HeadGaS: Real-time Animatable Head Avatars via 3D Gaussian Splatting. arXiv preprint arXiv:2312.02902 (2023) +2. Du, Y., Zhang, Y., Yu, H.X., Tenenbaum, J.B., Wu, J.: Neural Radiance Flow for 4D View Synthesis and Video Processing. In: ICCV (2021) +3. Fang, J., Yi, T., Wang, X., Xie, L., Zhang, X., Liu, W., Nießner, M., Tian, Q.: Fast Dynamic Radiance Fields with Time-aware Neural Voxels. In: SIGGRAPH Asia (2022) +4. Fridovich-Keil, S., Meanti, G., Warburg, F.R., Recht, B., Kanazawa, A.: K-Planes: Explicit Radiance Fields in Space, Time, and Appearance. In: CVPR (2023) +5. Hedman, P., Srinivasan, P.P., Mildenhall, B., Barron, J.T., Debevec, P.: Baking Neural Radiance Fields for Real-time View Synthesis. In: ICCV (2021) +6. Hu, L., Zhang, H., Zhang, Y., Zhou, B., Liu, B., Zhang, S., Nie, L.: GaussianAvatar: Towards Realistic Human Avatar Modeling from a Single Video via Animatable 3D Gaussians. In: CVPR (2024) +7. Huang, Y.H., Sun, Y.T., Yang, Z., Lyu, X., Cao, Y.P., Qi, X.: SC-GS: Sparse-Controlled Gaussian Splatting for Editable Dynamic Scenes. In: CVPR (2024) +8. Isik, M., Rünz, M., Georgopoulos, M., Khakhulin, T., Starck, J., Agapito, L., Niešner, M.: HumanRF: High-fidelity Neural Radiance Fields for Humans in Motion. ACM Trans. Graph. 42(4), 1-12 (2023) +9. Joo, H., Soo Park, H., Sheikh, Y.: MAP Visibility Estimation for Large-scale Dynamic 3D Reconstruction. In: CVPR (2014) +20. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3D Gaussian Splatting for Real-time Radiance Field Rendering. ACM Trans. Graph. 42(4) (2023) + +21. Kocabas, M., Chang, J.H.R., Gabriel, J., Tuzel, O., Ranjan, A.: HUGS: Human Gaussian Splatting. In: CVPR (2024) +22. Li, L., Shen, Z., Wang, Z., Shen, L., Tan, P.: D-NeRF: Neural Radiance Fields for Dynamic Scenes. In: CVPR (2020) +23. Li, L., Shen, Z., Wang, Z., Shen, L., Tan, P.: Streaming Radiance Fields for 3D Video Synthesis. In: NeurIPS (2022) +24. Li, R., Tanke, J., Vo, M., Zollhöfer, M., Gall, J., Kanazawa, A., Lassner, C.: TAVA: Template-free Animatable Volumetric Actors. In: ECCV (2022) +25. Li, T., Slavcheva, M., Zollhöfer, M., Green, S., Lassner, C., Kim, C., Schmidt, T., Lovegrove, S., Goesele, M., Newcombe, R., Lv, Z.: Neural 3D Video Synthesis from Multi-view Video. In: CVPR (2022) +26. Li, Z., Chen, Z., Li, Z., Xu, Y.: Spacetime Gaussian Feature Splatting for Real-time Dynamic View Synthesis. In: CVPR (2024) +27. Li, Z., Zheng, Z., Wang, L., Liu, Y.: Animatable Gaussians: Learning Pose-dependent Gaussian Maps for High-fidelity Human Avatar Modeling. In: CVPR (2024) +28. Li, Z., Niklaus, S., Snavely, N., Wang, O.: Neural Scene Flow Fields for Space-time View Synthesis of Dynamic Scenes. In: CVPR (2021) +29. Li, Z., Wang, Q., Cole, F., Tucker, R., Snavely, N.: DynIBaR: Neural Dynamic Image-Based Rendering. In: CVPR (2023) +30. Lin, K.E., Xiao, L., Liu, F., Yang, G., Ramamoorthi, R.: Deep 3D Mask Volume for View Synthesis of Dynamic Scenes. In: ICCV (2021) +31. Lin, Y., Dai, Z., Zhu, S., Yao, Y.: Gaussian-Flow: 4D Reconstruction with Dynamic 3D Gaussian Particle. In: CVPR (2024) +32. Liu, Y.L., Gao, C., Meuleman, A., Tseng, H.Y., Saraf, A., Kim, C., Chuang, Y.Y., Kopf, J., Huang, J.B.: Robust Dynamic Radiance Fields. In: CVPR (2023) +33. Lombardi, S., Simon, T., Saragih, J., Schwartz, G., Lehrmann, A., Sheikh, Y.: Neural Volumes: Learning Dynamic Renderable Volumes from Images. ACM Trans. Graph. 38(4) (2019) +34. Luiten, J., Fischer, T., Leibe, B.: Track to Reconstruct and Reconstruct to Track. IEEE Robotics and Automation Letters 5(2), 1803-1810 (2020) +35. Luiten, J., Kopanas, G., Leibe, B., Ramanan, D.: Dynamic 3D Gaussians: Tracking by Persistent Dynamic View Synthesis. In: 3DV (2024) +36. Maggioni, M., Tanay, T., Babiloni, F., McDonagh, S.G., Leonardis, A.: Tunable Convolutions with Parametric Multi-loss Optimization. In: CVPR (2023) +37. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. In: ECCV (2020) +38. Moreau, A., Song, J., Dhamo, H., Shaw, R., Zhou, Y., Pérez-Pellitero, E.: Human Gaussian Splatting: Real-time Rendering of Animatable Avatars. In: CVPR (2024) +39. Müller, T., Evans, A., Schied, C., Keller, A.: Instant Neural Graphics Primitives with a Multiresolution Hash Encoding. ACM Trans. Graph. 41(4), 102:1-102:15 (2022) +40. Newcombe, R.A., Fox, D., Seitz, S.M.: DynamicFusion: Reconstruction and Tracking of Non-rigid Scenes in Real-time. In: CVPR (2015) +41. Niemeyer, M., Barron, J.T., Mildenhall, B., Sajjadi, M.S.M., Geiger, A., Radwan, N.: RegNeRF: Regularizing Neural Radiance Fields for View Synthesis from Sparse Inputs. In: CVPR (2021) +42. Pang, H., Zhu, H., Kortylewski, A., Theobalt, C., Habermann, M.: ASH: Animatable Gaussian Splats for Efficient and Photoreal Human Rendering. In: CVPR (2024) + +43. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin-Brualla, R.: Nerfies: Deformable Neural Radiance Fields. ICCV (2021) +44. Park, K., Sinha, U., Hedman, P., Barron, J.T., Bouaziz, S., Goldman, D.B., Martin-Brualla, R., Seitz, S.M.: HyperNeRF: A Higher-dimensional Representation for Topologically Varying Neural Radiance Fields. ACM Trans. Graph. 40(6) (2021) +45. Peng, S., Dong, J., Wang, Q., Zhang, S., Shuai, Q., Zhou, X., Bao, H.: Animatable Neural Radiance Fields for Modeling Dynamic Human Bodies. In: ICCV (2021) +46. Peng, S., Yan, Y., Shuai, Q., Bao, H., Zhou, X.: Representing Volumetric Videos as Dynamic MLP Maps. In: CVPR (2023) +47. Qian, S., Kirschstein, T., Schoneveld, L., Davoli, D., Giebenhain, S., Nießner, M.: GaussianAvatars: Photorealistic Head Avatars with Rigged 3D Gaussians. In: CVPR (2024) +48. Qian, Z., Wang, S., Mihajlovic, M., Geiger, A., Tang, S.: 3DGS-Avatar: Animatable Avatars via Deformable 3D Gaussian Splitting. In: CVPR (2024) +49. Reiser, C., Peng, S., Liao, Y., Geiger, A.: KiloNeRF: Speeding up Neural Radiance Fields with Thousands of Tiny MLPs. In: ICCV (2021) +50. Rong, X., Huang, J.B., Saraf, A., Kim, C., Kopf, J.: Boosting View Synthesis with Residual Transfer. In: CVPR (2022) +51. Sabater, N., Boisson, G., Vandame, B., Kerbiriou, P., Babon, F., et al.: Dataset and Pipeline for Multi-view Light-Field Video. In: CVPRW (2017) +52. Saito, S., Schwartz, G., Simon, T., Li, J., Nam, G.: Relightable Gaussian Codec Avatars. In: CVPR (2024) +53. Schonberger, J.L., Frahm, J.M.: Structure-from-Motion Revisited. In: CVPR (2016) +54. Shao, R., Zheng, Z., Tu, H., Liu, B., Zhang, H., Liu, Y.: Tensor4D: Efficient Neural 4D Decomposition for High-fidelity Dynamic Reconstruction and Rendering. In: CVPR (2023) +55. Slavcheva, M., Baust, M., Cremers, D., Ilic, S.: KillingFusion: Non-rigid 3D Reconstruction without Correspondences. In: CVPR (2017) +56. Song, L., Chen, A., Li, Z., Chen, Z., Chen, L., Yuan, J., Xu, Y., Geiger, A.: NeRF-Player: A Streamable Dynamic Scene Representation with Decomposed Neural Radiance Fields. IEEE Transactions on Visualization and Computer Graphics 29(5), 2732-2742 (2023) +57. Sun, J., Jiao, H., Li, G., Zhang, Z., Zhao, L., Xing, W.: 3DGStream: On-the-fly Training of 3D Gaussians for Efficient Streaming of Photo-realistic Free-viewpoint Videos. In: CVPR (2024) +58. Tanay, T., Maggioni, M.: Global Latent Neural Rendering. In: CVPR (2024) +59. Teed, Z., Deng, J.: RAFT: Recurrent All-pairs Field Transforms for Optical Flow. In: ECCV (2020) +60. Turki, H., Zhang, J.Y., Ferroni, F., Ramanan, D.: SUDS: Scalable Urban Dynamic Scenes. In: CVPR (2023) +61. Wang, F., Tan, S., Li, X., Tian, Z., Liu, H.: Mixed Neural Voxels for Fast Multiview Video Synthesis. In: ICCV (2023) +62. Wang, L., Zhang, J., Liu, X., Zhao, F., Zhang, Y., Zhang, Y., Wu, M., Yu, J., Xu, L.: Fourier PlenOctrees for Dynamic Radiance Field Rendering in Real-time. In: CVPR (2022) +63. Wang, Q., Wang, Z., Genova, K., Srinivasan, P., Zhou, H., Barron, J.T., Martin-Brualla, R., Snavely, N., Funkhouser, T.: IBRNet: Learning Multi-view Image-Based Rendering. In: CVPR (2021) + +64. Weng, C.Y., Curless, B., Srinivasan, P.P., Barron, J.T., Kemelmacher-Shlizerman, I.: HumanNeRF: Free-viewpoint Rendering of Moving People from Monocular Video. In: CVPR (2022) +65. Wu, G., Yi, T., Fang, J., Xie, L., Zhang, X., Wei, W., Liu, W., Tian, Q., Wang, X.: 4D Gaussian Splatting for Real-time Dynamic Scene Rendering. In: CVPR (2024) +66. Wu, H., Chen, C., Hou, J., Liao, L., Wang, A., Sun, W., Yan, Q., Lin, W.: FASTVQA: Efficient End-to-end Video Quality Assessment with Fragment Sampling. In: ECCV (2022) +67. Xian, W., Huang, J.B., Kopf, J., Kim, C.: Space-time Neural Irradiance Fields for Free-viewpoint Video. In: CVPR (2021) +68. Xiang, J., Gao, X., Guo, Y., Zhang, J.: FlashAvatar: High-fidelity Head Avatar with Efficient Gaussian Embedding. In: CVPR (2024) +69. Xing, W., Chen, J.: Temporal-MPI: Enabling Multi-Plane Images for Dynamic Scene Modelling via Temporal Basis Learning. In: ECCV (2022) +70. Xu, Y., Chen, B., Li, Z., Zhang, H., Wang, L., Zheng, Z., Liu, Y.: Gaussian HeadAvatar: Ultra High-fidelity Head Avatar via Dynamic Gaussians. In: CVPR (2024) +71. Yang, Z., Yang, H., Pan, Z., Zhang, L.: Real-time Photorealistic Dynamic Scene Representation and Rendering with 4D Gaussian Splatting. In: ICLR (2024) +72. Yang, Z., Gao, X., Zhou, W., Jiao, S., Zhang, Y., Jin, X.: Deformable 3D Gaussians for High-fidelity Monocular Dynamic Scene Reconstruction. In: CVPR (2024) +73. Yu, A., Li, R., Tancik, M., Li, H., Ng, R., Kanazawa, A.: PlenOctrees for Real-time Rendering of Neural Radiance Fields. In: ICCV (2021) +74. Zhao, F., Yang, W., Zhang, J., Lin, P., Zhang, Y., Yu, J., Xu, L.: HumanNeRF: Efficiently Generated Human Radiance Field from Sparse Inputs. In: CVPR (2022) +75. Zhou, K., Li, W., Wang, Y., Hu, T., Jiang, N., Han, X., Lu, J.: NeRFLiX: High-quality Neural View Synthesis by Learning a Degradation-driven Inter-viewpoint MiXer. In: CVPR (2023) +76. Zhou, T., Tucker, R., Flynn, J., Fyffe, G., Snavely, N.: Stereo Magnification: Learning View Synthesis using Multiplane Images. ACM Trans. Graph. 37 (2018) \ No newline at end of file diff --git a/2024/SWinGS_ Sliding Windows for Dynamic 3D Gaussian Splatting/images.zip b/2024/SWinGS_ Sliding Windows for Dynamic 3D Gaussian Splatting/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c1c6af47053f6eb96e3fe73f7a6a964a2dff6d3a --- /dev/null +++ b/2024/SWinGS_ Sliding Windows for Dynamic 3D Gaussian Splatting/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:20c7d00ab9ee651946263cf419fe2112fb12ac6d2530bf67107c389c62f3e370 +size 823137 diff --git a/2024/SWinGS_ Sliding Windows for Dynamic 3D Gaussian Splatting/layout.json b/2024/SWinGS_ Sliding Windows for Dynamic 3D Gaussian Splatting/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..cbdd39c74f819633da6725f3c8d499d9a9e20311 --- /dev/null +++ b/2024/SWinGS_ Sliding Windows for Dynamic 3D Gaussian Splatting/layout.json @@ -0,0 +1,9393 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 194, + 111, + 420, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 111, + 420, + 148 + ], + "spans": [ + { + "bbox": [ + 194, + 111, + 420, + 148 + ], + "type": "text", + "content": "SWinGS: Sliding Windows for Dynamic 3D Gaussian Splatting" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 153, + 167, + 462, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 167, + 462, + 194 + ], + "spans": [ + { + "bbox": [ + 153, + 167, + 462, + 194 + ], + "type": "text", + "content": "Richard Shaw1, Michal Nazarczuk1, Jifei Song1, Arthur Moreau1, Sibi Catley-Chandar1,2, Helisa Dhamo1, and Eduardo Pérez-Pellitero1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 230, + 201, + 382, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 201, + 382, + 224 + ], + "spans": [ + { + "bbox": [ + 230, + 201, + 382, + 224 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 230, + 201, + 382, + 224 + ], + "type": "text", + "content": " Huawei Noah's Ark Lab \n" + }, + { + "bbox": [ + 230, + 201, + 382, + 224 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 230, + 201, + 382, + 224 + ], + "type": "text", + "content": " Queen Mary University Of London" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 135, + 258, + 242, + 352 + ], + "blocks": [ + { + "bbox": [ + 135, + 258, + 242, + 352 + ], + "lines": [ + { + "bbox": [ + 135, + 258, + 242, + 352 + ], + "spans": [ + { + "bbox": [ + 135, + 258, + 242, + 352 + ], + "type": "image", + "image_path": "adfaab72d4811650a89fe9f926206654dce088cc185571b363b4081714d95df5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 371, + 482, + 405 + ], + "lines": [ + { + "bbox": [ + 132, + 371, + 482, + 405 + ], + "spans": [ + { + "bbox": [ + 132, + 371, + 482, + 405 + ], + "type": "text", + "content": "Fig. 1: Left: SWinGS achieves sharper dynamic 3D scene reconstruction in part thanks to a sliding window canonical space that reduces the complexity of the 3D motion estimation. Right: Our dynamic real-time viewer allows users to explore the scene." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 242, + 258, + 346, + 351 + ], + "blocks": [ + { + "bbox": [ + 242, + 258, + 346, + 351 + ], + "lines": [ + { + "bbox": [ + 242, + 258, + 346, + 351 + ], + "spans": [ + { + "bbox": [ + 242, + 258, + 346, + 351 + ], + "type": "image", + "image_path": "e06659d5361df453a0466df373466c8ce4c5824e1fb721ac814c62ac8f6c29e9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 353, + 258, + 477, + 351 + ], + "blocks": [ + { + "bbox": [ + 353, + 258, + 477, + 351 + ], + "lines": [ + { + "bbox": [ + 353, + 258, + 477, + 351 + ], + "spans": [ + { + "bbox": [ + 353, + 258, + 477, + 351 + ], + "type": "image", + "image_path": "868ae6435617946c74c592d7a37e5513eb5e35aae3b9b87cf619e5d183643913.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 160, + 445, + 453, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 445, + 453, + 666 + ], + "spans": [ + { + "bbox": [ + 160, + 445, + 453, + 666 + ], + "type": "text", + "content": "Abstract. Novel view synthesis has shown rapid progress recently, with methods capable of producing increasingly photorealistic results. 3D Gaussian Splatting has emerged as a promising method, producing high-quality renderings of scenes and enabling interactive viewing at real-time frame rates. However, it is limited to static scenes. In this work, we extend 3D Gaussian Splatting to reconstruct dynamic scenes. We model a scene's dynamics using dynamic MLPs, learning deformations from temporally-local canonical representations to per-frame 3D Gaussians. To disentangle static and dynamic regions, tuneable parameters weigh each Gaussian's respective MLP parameters, improving the dynamics modelling of imbalanced scenes. We introduce a sliding window training strategy that partitions the sequence into smaller manageable windows to handle arbitrary length scenes while maintaining high rendering quality. We propose an adaptive sampling strategy to determine appropriate window size hyperparameters based on the scene's motion, balancing training overhead with visual quality. Training a separate dynamic 3D Gaussian model for each sliding window allows the canonical representation to change, enabling the reconstruction of scenes with significant geometric changes. Temporal consistency is enforced using a fine-tuning step with self-supervising consistency loss on randomly sampled novel" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 159, + 117, + 455, + 150 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 117, + 455, + 150 + ], + "spans": [ + { + "bbox": [ + 159, + 117, + 455, + 150 + ], + "type": "text", + "content": "views. As a result, our method produces high-quality renderings of general dynamic scenes with competitive quantitative performance, which can be viewed in real-time in our dynamic interactive viewer." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 166, + 230, + 178 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 166, + 230, + 178 + ], + "spans": [ + { + "bbox": [ + 132, + 166, + 230, + 178 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 186, + 481, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 186, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 186, + 481, + 282 + ], + "type": "text", + "content": "Photorealistic rendering and generally 3-dimensional (3D) imaging have received significant attention in recent years, especially since the seminal work of Neural Radiance Fields (NeRF) [37]. This is in part thanks to its impressive novel view synthesis results, but also due to its appealing ease of use when coupled with off-the-shelf structure-from-motion camera pose estimation [53]. NeRF's key insight is a fully differentiable volumetric rendering pipeline paired with learnable implicit functions that model a view-dependent 3D radiance field. Dense coverage of posed images of the scene provides then direct photometric supervision." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 282, + 481, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 282, + 481, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 282, + 481, + 498 + ], + "type": "text", + "content": "The original formulation of NeRF and most follow-ups [5,9,37,39,41] assume static scenes and thus a fixed radiance field. Some have explored new paradigms enabling dynamic reconstruction for radiance fields, including D-NeRF [22] and Nerfies [43], optimising an additional continuous volumetric deformation field that warps each observed point into a canonical NeRF. Such an approach has been popular [12,13,28,32,33,44], and has also been used for dynamic human reconstruction [45,74]. However, learning 3D deformation fields is inherently challenging, especially for large motions, with increased computational expense in training and inference. Moreover, approaches that share a canonical space among all frames struggle to maintain reconstruction quality for long sequences, obtaining overly blurred results due to inaccurate deformations and limited representational capacity. Other methods avoid maintaining a canonical representation and use explicit per-frame representations of the dynamic scene. Examples are tri-plane extensions to 4D " + }, + { + "bbox": [ + 130, + 282, + 481, + 498 + ], + "type": "inline_equation", + "content": "(x,y,z,t)" + }, + { + "bbox": [ + 130, + 282, + 481, + 498 + ], + "type": "text", + "content": " [7,14,54,60] with plane decompositions [9] to keep memory footprint under control. These approaches can suffer from a lack of temporal consistency, especially as they generally are agnostic about the motion of the scene. Grid-based methods that share a representation [46], also can suffer from degradation due to lack of representational capacity in long sequences." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 498, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 481, + 666 + ], + "type": "text", + "content": "This paper proposes a new method that addresses open problems of the state-of-the-art (SoTA) - see Fig. 1. Our method overview is shown in Fig. 2. Firstly, we build upon 3D Gaussian Splatting (3DGS) [20] and adapt the 3D Gaussians to be dynamic by allowing them to move. Our representation is thus explicit and avoids expensive raymarching via fast rasterization. Secondly, we introduce a novel paradigm for dynamic neural rendering with temporally-local canonical spaces defined in a sliding window fashion. Each window's length is adaptively defined following the amount of scene motion to maintain high-quality reconstruction. By limiting the scope of each canonical space, we can accurately track 3D displacements (i.e. they are generally smaller displacements) and prevent intra-window flickering. Thirdly, we introduce tuneable MLPs (MLP with several sets of weights governed by per-3D Gaussian blending weights) to estimate displacements. This tackles scenes with static vs dynamic imbalance. By learning different \"modes\" of motion estimation, we can separate smoothly between" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "text", + "content": "R. Shaw et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": "static and dynamic regions with virtually no additional computational cost nor any handcrafted heuristics. Lastly, temporal consistency loss computed on overlapping frames of neighbouring windows ensures consistency between windows, i.e. avoids inter-window flickering. In summary, our main contributions are:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 136, + 170, + 481, + 253 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 136, + 170, + 481, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 170, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 136, + 170, + 481, + 194 + ], + "type": "text", + "content": "1. An adaptive sliding window approach that enables the reconstruction of arbitrary length sequences whilst maintaining high rendering quality." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 136, + 194, + 481, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 194, + 481, + 217 + ], + "spans": [ + { + "bbox": [ + 136, + 194, + 481, + 217 + ], + "type": "text", + "content": "2. Temporally-local dynamic MLPs that model scene dynamics by learning deformation fields from per-window canonical 3D Gaussians to each frame." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 136, + 217, + 481, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 217, + 481, + 241 + ], + "spans": [ + { + "bbox": [ + 136, + 217, + 481, + 241 + ], + "type": "text", + "content": "3. Learnable MLP tuning parameters tackle scene imbalance by learning different motion modes; disentangling static canonical and dynamic 3D Gaussians." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 136, + 241, + 481, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 241, + 481, + 253 + ], + "spans": [ + { + "bbox": [ + 136, + 241, + 481, + 253 + ], + "type": "text", + "content": "4. A fine-tuning stage ensures temporal consistency throughout the sequence." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 270, + 234, + 283 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 234, + 283 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 234, + 283 + ], + "type": "text", + "content": "2 Related work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 294, + 482, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 294, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 294, + 482, + 390 + ], + "type": "text", + "content": "Non-NeRF dynamic reconstruction: A number of approaches prior to the emergence of NeRF [37] tackled dynamic scene reconstruction. Such methods typically relied on dense camera coverage for point tracking and reprojection [19], or the presence of additional measurements, i.e. depth [40, 55]. Alternatively, some were curated towards specific domains, e.g. car reconstruction in driving scenarios [6, 34]. Several recent works [3,29,46] followed the idea of Image-Based Rendering [63] (direct reconstruction from neighbouring views). Others [30, 69] utilize multiplane images [58, 76] with an additional temporal component." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 391, + 482, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 391, + 482, + 570 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 482, + 570 + ], + "type": "text", + "content": "NeRF-based reconstruction: NeRF [37] has achieved great success for reconstructing static scenes with many works extending it to dynamic inputs. Several methods [4, 67] model separate representations per time-step, disregarding the temporal component of the input. D-NeRF [22] reconstruct the scene in a canonical representation and model temporal variations with a deformation field. This idea was developed upon by many works [12, 13, 28, 32, 33, 44]. StreamRF [23] and NeRFPlayer [56] use time-aware MLPs and a compact 3D grid at each time-step for 4D field representation, reducing memory cost for longer videos. Other approaches [7, 14, 54, 60] represent dynamic scenes with a space-time grid, with grid decomposition to increase efficiency. DyNeRF [25] represents dynamic scenes by extending NeRF with an additional time-variant latent code. HyperReel [2] uses an efficient sampling network, modeling the scene around keyframes. MixVoxels [61] represents dynamic and static components with separately processed voxels. Several methods [18, 24, 64] rely on an underlying template mesh (e.g. human). Some methods [8, 50, 75] aim to improve NeRF quality post rendering." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "3D Gaussian Splatting Much of the development of neural rendering has focused on accelerating inference [15,46,49,62,73]. Recently, 3D Gaussian Splatting [20] made strides by modelling the scene with 3D Gaussians, which, when combined with tile-based differentiable rasterization, achieves very fast rendering, yet preserves high-quality reconstruction. Luiten et al. [35] extend this to dynamic scenes with shared 3D Gaussians that are optimised frame-by-frame. However, their focus is more on tracking 3D Gaussian trajectories rather than maximizing final rendering quality. Concurrent methods to ours that extend 3D Gaus-" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 410, + 90, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 410, + 90, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 410, + 90, + 447, + 101 + ], + "type": "text", + "content": "SWinGS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 113, + 252, + 224 + ], + "blocks": [ + { + "bbox": [ + 143, + 113, + 252, + 224 + ], + "lines": [ + { + "bbox": [ + 143, + 113, + 252, + 224 + ], + "spans": [ + { + "bbox": [ + 143, + 113, + 252, + 224 + ], + "type": "image", + "image_path": "f0ba281755eecfee02e1b9aa96def15dc4e98e3654ff807e81e82e7683e023f2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 233, + 482, + 289 + ], + "lines": [ + { + "bbox": [ + 130, + 233, + 482, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 233, + 482, + 289 + ], + "type": "text", + "content": "Fig. 2: Method. First, the sequence is partitioned into sliding windows based on optical flow. Second, a dynamic 3DGS model is trained per window, where tunable MLPs model the deformations. Blending parameters " + }, + { + "bbox": [ + 130, + 233, + 482, + 289 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 233, + 482, + 289 + ], + "type": "text", + "content": " weigh the MLP's parameters to focus on dynamic parts. Finally, each model is fine-tuned, enforcing inter-window temporal consistency with consistency loss on sampled views for overlapping frames." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 253, + 113, + 359, + 224 + ], + "blocks": [ + { + "bbox": [ + 253, + 113, + 359, + 224 + ], + "lines": [ + { + "bbox": [ + 253, + 113, + 359, + 224 + ], + "spans": [ + { + "bbox": [ + 253, + 113, + 359, + 224 + ], + "type": "image", + "image_path": "6849426cd786d7313c7aeddbc9fe257bf9cb780fc15ee020032617201022a6fe.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 361, + 113, + 472, + 224 + ], + "blocks": [ + { + "bbox": [ + 361, + 113, + 472, + 224 + ], + "lines": [ + { + "bbox": [ + 361, + 113, + 472, + 224 + ], + "spans": [ + { + "bbox": [ + 361, + 113, + 472, + 224 + ], + "type": "image", + "image_path": "c9c1409faba288cf9f3ff10840eede29f659a345966e1443b0d40e4a1b0a1865.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 314, + 482, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 314, + 482, + 352 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 482, + 352 + ], + "type": "text", + "content": "sian Splatting to focus on general dynamic scenes include [17, 26, 31, 57, 71, 72] amongst others, while other works have focused specifically on dynamic human reconstruction [16, 21, 27, 38, 42, 48, 65] and facial animation [11, 47, 52, 68, 70]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 372, + 202, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 372, + 202, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 372, + 202, + 385 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 401, + 209, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 401, + 209, + 412 + ], + "spans": [ + { + "bbox": [ + 131, + 401, + 209, + 412 + ], + "type": "text", + "content": "3.1 Overview" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 425, + 482, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 425, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 425, + 482, + 472 + ], + "type": "text", + "content": "We present our method to reconstruct and render novel views of general dynamic scenes from multiple calibrated time-synchronized cameras. An overview of our method is shown in Fig. 2. We build upon 3D Gaussian Splatting [20] for novel view synthesis of static scenes, which we extend to scenes containing motion." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 473, + 482, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 482, + 569 + ], + "type": "text", + "content": "Our method can be separated into three main steps. First, given a dynamic sequence, we split the sequence into separate shorter sliding windows of frames for concurrent processing. We adaptively sample windows of varying lengths depending on the amount of motion in the sequence. Each sliding window contains an overlapping frame between adjacent windows, enabling temporal consistency to be enforced throughout the sequence at a subsequent training stage. Partitioning the sequence into smaller windows enables us to deal with sequences of arbitrary length while maintaining high render quality." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 570, + 483, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 483, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 483, + 665 + ], + "type": "text", + "content": "Second, we train separate dynamic 3DGS models for each sliding window in turn. We extend the static 3DGS method to model the dynamics by introducing a tuneable MLP [36]. The MLP learns the deformation field from a canonical set of 3D Gaussians for each frame in a window. Thus each window comprises an independent temporally-local canonical 3D Gaussian representation and deformation field. This enables us to handle significant geometric changes and/or if new objects appear throughout the sequence, which can be challenging to model with a single representation. A tuneable MLP weighting parameter " + }, + { + "bbox": [ + 130, + 570, + 483, + 665 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 570, + 483, + 665 + ], + "type": "text", + "content": " is learned" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "text", + "content": "R. Shaw et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "for each 3D Gaussian to enable the MLP to focus on modelling the dynamic parts of the scene, with the static parts encapsulated by the canonical representation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 260 + ], + "type": "text", + "content": "Third, once a dynamic 3DGS model is trained for each window in the sequence, we apply a fine-tuning step to enforce temporal consistency throughout the sequence. We fine-tune each 3DGS model sequentially and employ a self-supervising temporal consistency loss on the overlapping frame renders between neighbouring windows. This encourages the model of the current window to produce similar renderings to the previous window. The result is a set of per-frame Gaussian Splatting models enabling high-quality novel view renderings of dynamic scenes with real-time interactive viewing capability. Our approach enables us to overcome the limitations of training with long sequences and to handle complex motions without exhibiting distracting temporal flickering." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 277, + 340, + 290 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 277, + 340, + 290 + ], + "spans": [ + { + "bbox": [ + 132, + 277, + 340, + 290 + ], + "type": "text", + "content": "3.2 Preliminary: 3D Gaussian Splitting" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 297, + 479, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 479, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 479, + 346 + ], + "type": "text", + "content": "Our method is built upon 3D Gaussian Splatting (3DGS) [20]. 3DGS uses a 3D Gaussian representation to model scenes as they are differentiable and can be projected to 2D splats, enabling fast tile-based rasterization. The 3D Gaussians are defined by 3D covariance matrix " + }, + { + "bbox": [ + 130, + 297, + 479, + 346 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 130, + 297, + 479, + 346 + ], + "type": "text", + "content": " in world space centered at the mean " + }, + { + "bbox": [ + 130, + 297, + 479, + 346 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 130, + 297, + 479, + 346 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 258, + 355, + 480, + 371 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 355, + 480, + 371 + ], + "spans": [ + { + "bbox": [ + 258, + 355, + 480, + 371 + ], + "type": "interline_equation", + "content": "G (x) = e ^ {- \\frac {1}{2} (x) ^ {T} \\Sigma^ {- 1} (x)}. \\tag {1}", + "image_path": "d6d52af14d99f6a8da7eecc80dcb2f5fc3a158b6f4ed4b024097980addee7769.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 376, + 482, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 472 + ], + "type": "text", + "content": "Given a scaling matrix " + }, + { + "bbox": [ + 130, + 376, + 482, + 472 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 376, + 482, + 472 + ], + "type": "text", + "content": " and rotation matrix " + }, + { + "bbox": [ + 130, + 376, + 482, + 472 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 130, + 376, + 482, + 472 + ], + "type": "text", + "content": ", the corresponding covariance matrix " + }, + { + "bbox": [ + 130, + 376, + 482, + 472 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 130, + 376, + 482, + 472 + ], + "type": "text", + "content": " of a 3D Gaussian can be written as " + }, + { + "bbox": [ + 130, + 376, + 482, + 472 + ], + "type": "inline_equation", + "content": "\\Sigma = RSS^T R^T" + }, + { + "bbox": [ + 130, + 376, + 482, + 472 + ], + "type": "text", + "content": ". To represent a scene, 3DGS optimizes 3D Gaussian positions " + }, + { + "bbox": [ + 130, + 376, + 482, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 376, + 482, + 472 + ], + "type": "text", + "content": ", covariance " + }, + { + "bbox": [ + 130, + 376, + 482, + 472 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 130, + 376, + 482, + 472 + ], + "type": "text", + "content": " (scaling " + }, + { + "bbox": [ + 130, + 376, + 482, + 472 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 376, + 482, + 472 + ], + "type": "text", + "content": " and rotation " + }, + { + "bbox": [ + 130, + 376, + 482, + 472 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 130, + 376, + 482, + 472 + ], + "type": "text", + "content": "), opacity " + }, + { + "bbox": [ + 130, + 376, + 482, + 472 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 130, + 376, + 482, + 472 + ], + "type": "text", + "content": ", and colours, represented by spherical harmonic (SH) coefficients, capturing view-dependent appearance. The optimization is interleaved with Gaussian adaptive density control (ADC). The model is optimized by rendering the learned Gaussians via a differentiable rasterizer, comparing the resulting image " + }, + { + "bbox": [ + 130, + 376, + 482, + 472 + ], + "type": "inline_equation", + "content": "I_r" + }, + { + "bbox": [ + 130, + 376, + 482, + 472 + ], + "type": "text", + "content": " against the ground truth " + }, + { + "bbox": [ + 130, + 376, + 482, + 472 + ], + "type": "inline_equation", + "content": "I_{gt}" + }, + { + "bbox": [ + 130, + 376, + 482, + 472 + ], + "type": "text", + "content": ", and minimizing the loss function:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 219, + 484, + 480, + 497 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 484, + 480, + 497 + ], + "spans": [ + { + "bbox": [ + 219, + 484, + 480, + 497 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = (1 - \\lambda) \\mathcal {L} _ {1} \\left(I _ {r}, I _ {g t}\\right) + \\lambda \\mathcal {L} _ {\\text {S S I M}} \\left(I _ {r}, I _ {g t}\\right). \\tag {2}", + "image_path": "8fa8dd1b4ca22c59e1f064b3ef2c550408e125635ac0f0375331bd8d5b3bb574.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 514, + 298, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 298, + 526 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 298, + 526 + ], + "type": "text", + "content": "3.3 Sliding-Window Processing" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Using a single dynamic 3DGS representation to model an entire sequence becomes impractical for longer sequences, simply from a data processing standpoint. Furthermore, representing a lengthy sequence with a single model performs significantly worse than using multiple smaller fixed-sized segments, particularly if the scene has large motion or topological changes that cannot be modeled easily with one canonical representation and deformation field (Fig. 4). To address this, we use a sliding-window approach, separating the sequence into smaller windows with overlapping frames. The window size is a hyperparameter whose effect is explored in Section 4.1. Each window comprises an independent dynamic 3DGS model, and we allow all 3D Gaussian parameters to change between windows, including the number of Gaussians, their positions, rotations," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 410, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 410, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 410, + 91, + 447, + 100 + ], + "type": "text", + "content": "SWinGS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "scaling, colours and opacities. The advantage of this approach is that independent models can be trained in parallel across multiple GPUs to speed up training." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 154, + 480, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 154, + 480, + 238 + ], + "spans": [ + { + "bbox": [ + 130, + 154, + 480, + 238 + ], + "type": "text", + "content": "Adaptive Window Sampling We propose an adaptive method for choosing the appropriate sliding window sizes, balancing training overhead and model size with performance and temporal consistency. Given an input sequence of frame length " + }, + { + "bbox": [ + 130, + 154, + 480, + 238 + ], + "type": "inline_equation", + "content": "N_{f}" + }, + { + "bbox": [ + 130, + 154, + 480, + 238 + ], + "type": "text", + "content": ", we separate the sequence into smaller windows by adaptively sampling windows of different lengths " + }, + { + "bbox": [ + 130, + 154, + 480, + 238 + ], + "type": "inline_equation", + "content": "\\{N_w\\}" + }, + { + "bbox": [ + 130, + 154, + 480, + 238 + ], + "type": "text", + "content": ", depending on the amount of motion in the sequence. In high-motion areas, we aim to sample windows more frequently (shorter windows), and in low-motion areas, less frequently (longer windows)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 239, + 481, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 239, + 481, + 310 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 481, + 310 + ], + "type": "text", + "content": "To do this, we leverage the magnitude of 2D optical flow from each camera viewpoint. We employ a greedy algorithm to adaptively select the sizes of windows, prior to training, based on the accumulated optical flow magnitude. Specifically, we estimate per-frame optical flow " + }, + { + "bbox": [ + 130, + 239, + 481, + 310 + ], + "type": "inline_equation", + "content": "\\pmb{f}" + }, + { + "bbox": [ + 130, + 239, + 481, + 310 + ], + "type": "text", + "content": " using a pre-trained RAFT [59] model for each camera view " + }, + { + "bbox": [ + 130, + 239, + 481, + 310 + ], + "type": "inline_equation", + "content": "j\\in V" + }, + { + "bbox": [ + 130, + 239, + 481, + 310 + ], + "type": "text", + "content": " and all frames in the sequence " + }, + { + "bbox": [ + 130, + 239, + 481, + 310 + ], + "type": "inline_equation", + "content": "i\\in N_{f}" + }, + { + "bbox": [ + 130, + 239, + 481, + 310 + ], + "type": "text", + "content": ", and compute the mean flow magnitude summed over each frame:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 240, + 320, + 481, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 240, + 320, + 481, + 354 + ], + "spans": [ + { + "bbox": [ + 240, + 320, + 481, + 354 + ], + "type": "interline_equation", + "content": "\\hat {v} _ {i} = \\frac {1}{V} \\sum_ {j} ^ {V} \\sum_ {i} ^ {N _ {f} - 1} | | \\boldsymbol {f} \\left(I _ {i} ^ {j}, I _ {i + 1} ^ {j}\\right) | | _ {2} ^ {2} \\tag {3}", + "image_path": "7b57ccfbf337f15fcbe0d8274556e8df4898b6948d2bef00bef52fe95d510568.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 356, + 482, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 356, + 482, + 465 + ], + "spans": [ + { + "bbox": [ + 130, + 356, + 482, + 465 + ], + "type": "text", + "content": "We iterate over each frame in the sequence with a greedy heuristic; spawning a new window when the sum of mean flow " + }, + { + "bbox": [ + 130, + 356, + 482, + 465 + ], + "type": "inline_equation", + "content": "\\hat{v}_i" + }, + { + "bbox": [ + 130, + 356, + 482, + 465 + ], + "type": "text", + "content": " exceeds a pre-defined threshold. This ensures that each window contains a similar amount of movement, leading to a balanced distribution of the total representational workload. Taking the average across viewpoints makes this approach somewhat invariant to the number of cameras, while placing a limit on the total flow stops an excessive amount of movement within each window. Note, each sampled window overlaps with the next window by a single frame, such that neighbouring windows share a common image frame. This is to enable inter-window temporal consistency (section 3.6)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 478, + 278, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 478, + 278, + 490 + ], + "spans": [ + { + "bbox": [ + 131, + 478, + 278, + 490 + ], + "type": "text", + "content": "3.4 Dynamic 3D Gaussians" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 494, + 482, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 494, + 482, + 591 + ], + "spans": [ + { + "bbox": [ + 130, + 494, + 482, + 591 + ], + "type": "text", + "content": "To extend static 3DGS to handle dynamic scenes, we introduce a temporally-local dynamic MLP unique to each sliding window in the sequence (section 3.3). Each time-dependent MLP learns a temporally-local deformation field, mapping from a learned per-window canonical space to a set of 3D Gaussians for each frame in the window. Each deformation field, represented by a small MLP " + }, + { + "bbox": [ + 130, + 494, + 482, + 591 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\theta}" + }, + { + "bbox": [ + 130, + 494, + 482, + 591 + ], + "type": "text", + "content": " with weights " + }, + { + "bbox": [ + 130, + 494, + 482, + 591 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 494, + 482, + 591 + ], + "type": "text", + "content": ", takes as input the normalized frame time " + }, + { + "bbox": [ + 130, + 494, + 482, + 591 + ], + "type": "inline_equation", + "content": "t \\in [0,1]" + }, + { + "bbox": [ + 130, + 494, + 482, + 591 + ], + "type": "text", + "content": " and 3D Gaussian means " + }, + { + "bbox": [ + 130, + 494, + 482, + 591 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 130, + 494, + 482, + 591 + ], + "type": "text", + "content": " (normalized by the scene's mean and standard deviation), and outputs displacements to their positions " + }, + { + "bbox": [ + 130, + 494, + 482, + 591 + ], + "type": "inline_equation", + "content": "\\Delta \\pmb{x}" + }, + { + "bbox": [ + 130, + 494, + 482, + 591 + ], + "type": "text", + "content": ", rotations " + }, + { + "bbox": [ + 130, + 494, + 482, + 591 + ], + "type": "inline_equation", + "content": "\\Delta \\pmb{r}" + }, + { + "bbox": [ + 130, + 494, + 482, + 591 + ], + "type": "text", + "content": " and scaling " + }, + { + "bbox": [ + 130, + 494, + 482, + 591 + ], + "type": "inline_equation", + "content": "\\Delta s" + }, + { + "bbox": [ + 130, + 494, + 482, + 591 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 228, + 603, + 481, + 615 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 603, + 481, + 615 + ], + "spans": [ + { + "bbox": [ + 228, + 603, + 481, + 615 + ], + "type": "interline_equation", + "content": "\\Delta \\boldsymbol {x} (t), \\Delta \\boldsymbol {r} (t), \\Delta \\boldsymbol {s} (t) = \\mathcal {F} _ {\\theta} (\\gamma (\\boldsymbol {x}), \\gamma (t)) \\tag {4}", + "image_path": "6e7e79e68e0f10eac6afc5ed4f9ecfd64287d4964eb65d18093df8effb5a3ca7.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\gamma(.)" + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "content": " denotes sinusoidal positional encoding " + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\gamma: \\mathbb{R}^3 \\to \\mathbb{R}^{3+6m}" + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\gamma(x) = (x, \\ldots, \\sin(2k\\pi x), \\cos(2k\\pi x), \\ldots)" + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "content": ". We use small MLPs to reduce overfitting, setting the number of frequency components " + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "inline_equation", + "content": "m = 6" + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "content": ", MLP depth " + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "inline_equation", + "content": "D = 4" + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "content": " and width " + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "inline_equation", + "content": "W = 16" + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "content": ", with two skip connections." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "text", + "content": "R. Shaw et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 284, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 284, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 284, + 128 + ], + "type": "text", + "content": "3.5 Tunable Dynamic MLPs" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 136, + 482, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 136, + 482, + 207 + ], + "spans": [ + { + "bbox": [ + 130, + 136, + 482, + 207 + ], + "type": "text", + "content": "Ideally, time-dependent MLPs model motion in the scene by associating the dynamic parts with the temporal input component " + }, + { + "bbox": [ + 130, + 136, + 482, + 207 + ], + "type": "inline_equation", + "content": "\\gamma(t)" + }, + { + "bbox": [ + 130, + 136, + 482, + 207 + ], + "type": "text", + "content": " and thus learn to decouple the scene into i) static canonical 3D Gaussians and ii) dynamic 3D Gaussians. However, in scenes with an imbalance of static vs dynamic parts, e.g. Neural 3D Video [25] with mostly static backgrounds, the MLP struggles to disentangle static and dynamic regions, resulting in poorly estimated deformation fields." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 207, + 483, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 207, + 483, + 316 + ], + "spans": [ + { + "bbox": [ + 131, + 207, + 483, + 316 + ], + "type": "text", + "content": "To counteract this, we introduce tuneable dynamic MLPs [36] (i.e. MLPs with " + }, + { + "bbox": [ + 131, + 207, + 483, + 316 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 131, + 207, + 483, + 316 + ], + "type": "text", + "content": " sets of weights) governed by " + }, + { + "bbox": [ + 131, + 207, + 483, + 316 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 131, + 207, + 483, + 316 + ], + "type": "text", + "content": " sets of learnable blending parameters " + }, + { + "bbox": [ + 131, + 207, + 483, + 316 + ], + "type": "inline_equation", + "content": "\\{\\pmb{\\alpha} \\in \\mathbb{R}^{M \\times N_g}\\}" + }, + { + "bbox": [ + 131, + 207, + 483, + 316 + ], + "type": "text", + "content": ". The tuning parameters weigh the respective parameters of the MLP for each input Gaussian " + }, + { + "bbox": [ + 131, + 207, + 483, + 316 + ], + "type": "inline_equation", + "content": "i \\in N_g" + }, + { + "bbox": [ + 131, + 207, + 483, + 316 + ], + "type": "text", + "content": ", enabling the learning of " + }, + { + "bbox": [ + 131, + 207, + 483, + 316 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 131, + 207, + 483, + 316 + ], + "type": "text", + "content": " modes of variation corresponding to different motion modes. With " + }, + { + "bbox": [ + 131, + 207, + 483, + 316 + ], + "type": "inline_equation", + "content": "M = 2" + }, + { + "bbox": [ + 131, + 207, + 483, + 316 + ], + "type": "text", + "content": ", the MLP has the ability to decouple static and dynamic Gaussians in a smoothly weighted manner. Given the set of blending parameters " + }, + { + "bbox": [ + 131, + 207, + 483, + 316 + ], + "type": "inline_equation", + "content": "\\{\\alpha_i\\}_{m=1}^M" + }, + { + "bbox": [ + 131, + 207, + 483, + 316 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 131, + 207, + 483, + 316 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 207, + 483, + 316 + ], + "type": "text", + "content": "-th input Gaussian " + }, + { + "bbox": [ + 131, + 207, + 483, + 316 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i" + }, + { + "bbox": [ + 131, + 207, + 483, + 316 + ], + "type": "text", + "content": ", the output of a single layer of the dynamic MLP " + }, + { + "bbox": [ + 131, + 207, + 483, + 316 + ], + "type": "inline_equation", + "content": "\\pmb{y}_i" + }, + { + "bbox": [ + 131, + 207, + 483, + 316 + ], + "type": "text", + "content": " can be written as a weighted sum over the " + }, + { + "bbox": [ + 131, + 207, + 483, + 316 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 131, + 207, + 483, + 316 + ], + "type": "text", + "content": " sets of weights:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 223, + 325, + 481, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 325, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 223, + 325, + 481, + 357 + ], + "type": "interline_equation", + "content": "\\boldsymbol {y} _ {i} = \\phi \\left(\\sum_ {m = 1} ^ {M} \\left(\\alpha_ {i, m} \\boldsymbol {w} _ {m} ^ {T} \\boldsymbol {x} _ {i} + \\alpha_ {i, m} \\boldsymbol {b} _ {m}\\right)\\right), \\tag {5}", + "image_path": "bd13c9a269d5a20d004fd7c0c6d17b3213d4f0aef823a5e3fa1fb37e2e11b126.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 361, + 482, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 361, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 131, + 361, + 482, + 445 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 361, + 482, + 445 + ], + "type": "inline_equation", + "content": "\\boldsymbol{w}" + }, + { + "bbox": [ + 131, + 361, + 482, + 445 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 361, + 482, + 445 + ], + "type": "inline_equation", + "content": "\\boldsymbol{b}" + }, + { + "bbox": [ + 131, + 361, + 482, + 445 + ], + "type": "text", + "content": " are the weights and bias, and " + }, + { + "bbox": [ + 131, + 361, + 482, + 445 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 131, + 361, + 482, + 445 + ], + "type": "text", + "content": " is a non-linear activation function. The blending parameters " + }, + { + "bbox": [ + 131, + 361, + 482, + 445 + ], + "type": "inline_equation", + "content": "\\{\\alpha_{i}\\}_{m = 1}^{M}" + }, + { + "bbox": [ + 131, + 361, + 482, + 445 + ], + "type": "text", + "content": " linearly blend " + }, + { + "bbox": [ + 131, + 361, + 482, + 445 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 131, + 361, + 482, + 445 + ], + "type": "text", + "content": " sets of weights and biases of each layer of the MLP for each input Gaussian which, when passed through the activation function, enables nonlinear interaction between " + }, + { + "bbox": [ + 131, + 361, + 482, + 445 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 131, + 361, + 482, + 445 + ], + "type": "text", + "content": " and the output. Applying " + }, + { + "bbox": [ + 131, + 361, + 482, + 445 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 131, + 361, + 482, + 445 + ], + "type": "text", + "content": " in this way naturally enables the learning different of motion modes with a single forward pass of the MLP. Thus the dynamic MLP " + }, + { + "bbox": [ + 131, + 361, + 482, + 445 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\theta}^{dyn}" + }, + { + "bbox": [ + 131, + 361, + 482, + 445 + ], + "type": "text", + "content": " becomes a function of position, time, and blending parameters " + }, + { + "bbox": [ + 131, + 361, + 482, + 445 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 131, + 361, + 482, + 445 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 216, + 456, + 481, + 469 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 456, + 481, + 469 + ], + "spans": [ + { + "bbox": [ + 216, + 456, + 481, + 469 + ], + "type": "interline_equation", + "content": "\\Delta \\boldsymbol {x} (t), \\Delta \\boldsymbol {r} (t), \\Delta \\boldsymbol {s} (t) = \\mathcal {F} _ {\\theta} ^ {d y n} (\\gamma (\\boldsymbol {x}), \\gamma (t), \\boldsymbol {\\alpha}). \\tag {6}", + "image_path": "9aef9fe3f634bb611ee58cfba08f6795928893bcd97b24b9666e8f0fd1ec3827.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 474, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 581 + ], + "type": "text", + "content": "We implement the dynamic MLP as a single batch matrix multiplication (see supplementary). Blending parameters " + }, + { + "bbox": [ + 130, + 474, + 482, + 581 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 474, + 482, + 581 + ], + "type": "text", + "content": " are initialized as a binary mask of dynamic Gaussians (0-static, 1-dynamic) as follows. For a sliding window, we project all 3D Gaussians into each camera view: " + }, + { + "bbox": [ + 130, + 474, + 482, + 581 + ], + "type": "inline_equation", + "content": "u = \\varPi_{j}(\\pmb{x})" + }, + { + "bbox": [ + 130, + 474, + 482, + 581 + ], + "type": "text", + "content": ", obtaining their 2D-pixel coordinates. We compute their L1 pixel differences from the central frame to each frame in the window. If the difference is larger than a threshold, we label that Gaussian 1, otherwise 0. To be robust to occlusions and mislabelling, we average the assigned label over all views and frames in the window. If the average is greater than 0.5, we initialize the Gaussian dynamic, otherwise static." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Once initialized, we set the blending parameter " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " as a learnable parameter, and optimize it via back-propagation together with the rest of the system. We assume " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " is constant for all frames in a sliding window to reduce the complexity of the optimization, which is a fair assumption given our adaptive window sampling mechanism. Fig. 3 visualizes the learning of the MLP tuning parameter. We observe " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " providing higher weight to Gaussians likely to be dynamic. Fig. 3 (right) plots the magnitudes of resulting displacements " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\Delta x" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " output by the MLP." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 410, + 90, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 410, + 90, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 410, + 90, + 447, + 101 + ], + "type": "text", + "content": "SWinGS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 140 + ], + "type": "text", + "content": "Note, " + }, + { + "bbox": [ + 132, + 116, + 482, + 140 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 132, + 116, + 482, + 140 + ], + "type": "text", + "content": " does not have to be entirely accurate, as the MLP learns to adjust accordingly, yet enables the MLP to handle highly imbalanced scenes (Table 6)." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 134, + 161, + 299, + 218 + ], + "blocks": [ + { + "bbox": [ + 134, + 161, + 299, + 218 + ], + "lines": [ + { + "bbox": [ + 134, + 161, + 299, + 218 + ], + "spans": [ + { + "bbox": [ + 134, + 161, + 299, + 218 + ], + "type": "image", + "image_path": "c3dcdb2f11a8ec061cf3864f723627d2eb0a6c004174315a99e39398c82a3011.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 236, + 482, + 281 + ], + "lines": [ + { + "bbox": [ + 131, + 236, + 482, + 281 + ], + "spans": [ + { + "bbox": [ + 131, + 236, + 482, + 281 + ], + "type": "text", + "content": "Fig. 3: Dynamic MLPs with tunable parameters " + }, + { + "bbox": [ + 131, + 236, + 482, + 281 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 131, + 236, + 482, + 281 + ], + "type": "text", + "content": " weigh the parameters of the MLP for each Gaussian. We show renders from two scenes, left: cook spinach [25] and right: Train [51]. Shown from left-to-right: image render, tunable " + }, + { + "bbox": [ + 131, + 236, + 482, + 281 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 131, + 236, + 482, + 281 + ], + "type": "text", + "content": " parameters, and normalized MLP displacements " + }, + { + "bbox": [ + 131, + 236, + 482, + 281 + ], + "type": "inline_equation", + "content": "\\Delta x" + }, + { + "bbox": [ + 131, + 236, + 482, + 281 + ], + "type": "text", + "content": ". Note, " + }, + { + "bbox": [ + 131, + 236, + 482, + 281 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 131, + 236, + 482, + 281 + ], + "type": "text", + "content": " highlights the scene's dynamic regions." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 303, + 161, + 480, + 217 + ], + "blocks": [ + { + "bbox": [ + 303, + 161, + 480, + 217 + ], + "lines": [ + { + "bbox": [ + 303, + 161, + 480, + 217 + ], + "spans": [ + { + "bbox": [ + 303, + 161, + 480, + 217 + ], + "type": "image", + "image_path": "355e0502329e36c80dec5967c0753a42ca8b3763eafae513c562decf22bd4b69.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 321, + 334, + 334 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 321, + 334, + 334 + ], + "spans": [ + { + "bbox": [ + 132, + 321, + 334, + 334 + ], + "type": "text", + "content": "3.6 Temporal Consistency Fine-tuning" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 341, + 482, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 341, + 482, + 436 + ], + "spans": [ + { + "bbox": [ + 130, + 341, + 482, + 436 + ], + "type": "text", + "content": "Due to the non-deterministic nature of the 3DGS optimization, independently trained models for each sliding window may produce slightly different results. When the resulting renders from each model are played back in sequence, noticeable flickering can be observed, particularly in novel views (though it's less noticeable in the training views). Consequently, after training models for all sliding windows separately, we tackle inter-window temporal flickering by introducing a fine-tuning step for temporal consistency. This step uses a self-supervising loss function to aid in smooth transitions between the models of separate windows." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 437, + 482, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 534 + ], + "type": "text", + "content": "We fine-tune each model for a short period (3000 iterations in our experiments). We initiate the process from the first sliding window and progress through the sequence sequentially. For a window comprising " + }, + { + "bbox": [ + 130, + 437, + 482, + 534 + ], + "type": "inline_equation", + "content": "N_{w}" + }, + { + "bbox": [ + 130, + 437, + 482, + 534 + ], + "type": "text", + "content": " frames, we load the trained model and the model from the preceding window, with one overlapping frame between them. We then freeze all the parameters of the previous model. In order to fine-tune the model, as the flickering is mainly observed in novel views, we randomly sample novel test views in between the training views by rigidly interpolating the training camera poses " + }, + { + "bbox": [ + 130, + 437, + 482, + 534 + ], + "type": "inline_equation", + "content": "P_{j} = [R|\\pmb{t}] \\in \\mathbb{R}^{4 \\times 4}" + }, + { + "bbox": [ + 130, + 437, + 482, + 534 + ], + "type": "text", + "content": " in SE(3):" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 240, + 541, + 481, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 240, + 541, + 481, + 574 + ], + "spans": [ + { + "bbox": [ + 240, + 541, + 481, + 574 + ], + "type": "interline_equation", + "content": "P _ {\\text {n o v e l}} = \\exp_ {\\mathrm {M}} \\sum_ {j} ^ {V} \\beta_ {j} \\log_ {\\mathrm {M}} \\left(P _ {j}\\right) \\tag {7}", + "image_path": "31ed5668f85a30dab67d0b89e550f156b0c452071768a362fb0c2329245fd2f3.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 580, + 480, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 580, + 480, + 606 + ], + "spans": [ + { + "bbox": [ + 130, + 580, + 480, + 606 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 580, + 480, + 606 + ], + "type": "inline_equation", + "content": "\\exp_{\\mathbf{M}}" + }, + { + "bbox": [ + 130, + 580, + 480, + 606 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 580, + 480, + 606 + ], + "type": "inline_equation", + "content": "\\log_{\\mathbf{M}}" + }, + { + "bbox": [ + 130, + 580, + 480, + 606 + ], + "type": "text", + "content": " are the matrix exponential and logarithm [1] respectively, and " + }, + { + "bbox": [ + 130, + 580, + 480, + 606 + ], + "type": "inline_equation", + "content": "\\beta_{j}\\in [0,1]" + }, + { + "bbox": [ + 130, + 580, + 480, + 606 + ], + "type": "text", + "content": " is a uniformly sampled weighting such that " + }, + { + "bbox": [ + 130, + 580, + 480, + 606 + ], + "type": "inline_equation", + "content": "\\sum_{j}^{V}\\beta_{j} = 1" + }, + { + "bbox": [ + 130, + 580, + 480, + 606 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 607, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 607, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 607, + 482, + 666 + ], + "type": "text", + "content": "In one fine-tuning step, we render the overlapping frames from a randomly sampled novel viewpoint using both the model of the current window " + }, + { + "bbox": [ + 130, + 607, + 482, + 666 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 130, + 607, + 482, + 666 + ], + "type": "text", + "content": " and the previous window " + }, + { + "bbox": [ + 130, + 607, + 482, + 666 + ], + "type": "inline_equation", + "content": "w - 1" + }, + { + "bbox": [ + 130, + 607, + 482, + 666 + ], + "type": "text", + "content": ". This means we use the first frame of current window " + }, + { + "bbox": [ + 130, + 607, + 482, + 666 + ], + "type": "inline_equation", + "content": "I_{t = 0}^{w}" + }, + { + "bbox": [ + 130, + 607, + 482, + 666 + ], + "type": "text", + "content": " and the last frame of the previous window " + }, + { + "bbox": [ + 130, + 607, + 482, + 666 + ], + "type": "inline_equation", + "content": "I_{t = N_w - 1}^{w - 1}" + }, + { + "bbox": [ + 130, + 607, + 482, + 666 + ], + "type": "text", + "content": ". We then apply a consistency loss, which is simply the L1 loss on the two image renders from both models:" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "text", + "content": "R. Shaw et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 236, + 126, + 481, + 142 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 126, + 481, + 142 + ], + "spans": [ + { + "bbox": [ + 236, + 126, + 481, + 142 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c o n s i s t e n c y}} = \\left| I _ {t = 0} ^ {w} - I _ {t = N _ {w} - 1} ^ {w - 1} \\right| _ {1} \\tag {8}", + "image_path": "76350d5211cea26c290229aafb1d7987c1e3f1aa8c531f86809fb1436982de54.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 145, + 482, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 145, + 482, + 229 + ], + "spans": [ + { + "bbox": [ + 130, + 145, + 482, + 229 + ], + "type": "text", + "content": "When fine-tuning a model, we only allow the canonical (static) representation to change, while freezing the weights of the time-dependent dynamic MLP. Since the canonical set of 3D Gaussians is shared for all frames in a sliding window, we need to be careful not to negatively impact other frames when refining the overlapping frame. To address this, we use an alternating strategy for refinement, enforcing temporal consistency on the overlapping frame " + }, + { + "bbox": [ + 130, + 145, + 482, + 229 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 130, + 145, + 482, + 229 + ], + "type": "text", + "content": " of the time and training with the remaining views and frames as usual for the remaining " + }, + { + "bbox": [ + 130, + 145, + 482, + 229 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 130, + 145, + 482, + 229 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 245, + 279, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 245, + 279, + 257 + ], + "spans": [ + { + "bbox": [ + 131, + 245, + 279, + 257 + ], + "type": "text", + "content": "3.7 Implementation Details" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 263, + 482, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 263, + 482, + 504 + ], + "spans": [ + { + "bbox": [ + 130, + 263, + 482, + 504 + ], + "type": "text", + "content": "We implement our method in PyTorch, building upon the codebase and differentiable rasterizer of 3DGS [20]. We initialize each model with a point cloud obtained from COLMAP [53]. Each dynamic 3DGS model is trained for 15K iterations for all sampled windows in a sequence. The initial 2K iterations comprise a warm-up stage; training with only the central window frame and freezing the weights of the MLP, allowing the canonical representation to stabilize. We optimize the Gaussians' positions, rotations, scaling, opacities and SH coefficients. Afterwards, we unfreeze the MLP and allow the deformation field and tuning parameters " + }, + { + "bbox": [ + 130, + 263, + 482, + 504 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 263, + 482, + 504 + ], + "type": "text", + "content": " to optimize. We find this staggered optimization approach leads to better convergence. The number of Gaussians densifies for 8K iterations, after which the number of Gaussians is fixed. Inputs to the MLP are normalized by the mean and standard deviation of the canonical point cloud post-warm-up stage before frequency encoding, leading to faster and stabler convergence. We train with Adam optimizer, using different learning rates for each parameter following the implementation of [20]. The learning rate for the MLP and " + }, + { + "bbox": [ + 130, + 263, + 482, + 504 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 263, + 482, + 504 + ], + "type": "text", + "content": " parameters are set to 1e-4, with the MLP learning rate undergoing exponential decay by factor 1e-2 in 20K iterations. During the initial optimization phase, due to the independence of each Gaussian model, we train each model in parallel on eight 32Gb Tesla V100 GPUs to speed up training. Afterwards, we perform temporal fine-tuning of each model sequentially using a single GPU for 3K iterations each." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 133, + 570, + 329, + 640 + ], + "blocks": [ + { + "bbox": [ + 132, + 533, + 323, + 568 + ], + "lines": [ + { + "bbox": [ + 132, + 533, + 323, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 323, + 568 + ], + "type": "text", + "content": "Table 2: Quantitative results on Technicolor dataset [51] at full resolution. Best and second best results are highlighted." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 570, + 329, + 640 + ], + "lines": [ + { + "bbox": [ + 133, + 570, + 329, + 640 + ], + "spans": [ + { + "bbox": [ + 133, + 570, + 329, + 640 + ], + "type": "table", + "html": "
MethodPSNRSSIMLPIPSFPS
DyNeRF [25]31.800.9110.1420.02
HypeRreel [2]32.500.9020.1130.45
Dynamic3DG [35]27.020.8320.22886.96
Ours33.650.9340.11723.79
", + "image_path": "cf8c5288faf9ed86146ead5a2213368d6f3b125b4e80450aebb1d0903f3c8e13.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 341, + 525, + 476, + 598 + ], + "blocks": [ + { + "bbox": [ + 341, + 525, + 476, + 598 + ], + "lines": [ + { + "bbox": [ + 341, + 525, + 476, + 598 + ], + "spans": [ + { + "bbox": [ + 341, + 525, + 476, + 598 + ], + "type": "image", + "image_path": "53586640ec588eaa2592c01d4b8987d6e1e79d5e2727521a8cae144dd4a52479.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 332, + 606, + 482, + 639 + ], + "lines": [ + { + "bbox": [ + 332, + 606, + 482, + 639 + ], + "spans": [ + { + "bbox": [ + 332, + 606, + 482, + 639 + ], + "type": "text", + "content": "Fig. 4: Comparison in performance consistency for Ours and Dynamic3DG in consecutive frames." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 410, + 90, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 410, + 90, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 410, + 90, + 447, + 101 + ], + "type": "text", + "content": "SWinGS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 209, + 159, + 401, + 266 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 150 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 150 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 150 + ], + "type": "text", + "content": "Table 1: Quantitative results on the Neural 3D Video dataset [25], averaged over all scenes. Best and second best highlighted. Our method performs best overall whilst enabling real-time frame rates. Per-scene breakdown of results are given in Table 3." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 209, + 159, + 401, + 266 + ], + "lines": [ + { + "bbox": [ + 209, + 159, + 401, + 266 + ], + "spans": [ + { + "bbox": [ + 209, + 159, + 401, + 266 + ], + "type": "table", + "html": "
MethodPSNRSSIMLPIPSFPS
MixVoxels [61]30.420.9230.1244.30
K-Planes [14]30.630.9220.1170.33
HexPlane [7]30.000.9220.1130.24
HyperReel [2]30.780.9310.1013.60
NeRFPlayer [56]30.700.9310.1210.10
StreamRF [23]30.230.9040.1779.40
4DGS [65]27.610.9160.13530.00
Ours31.100.9400.09671.51
", + "image_path": "068521d4d9ca8a3d8817ce701a7d8e8e00c2d8bb671ce83e9fa5b5d7e7c1d020.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 287, + 200, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 287, + 200, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 287, + 200, + 300 + ], + "type": "text", + "content": "4 Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 314, + 479, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 314, + 479, + 337 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 479, + 337 + ], + "type": "text", + "content": "We evaluate our method on two real-world multi-view dynamic benchmarks: the Neural 3D Video dataset [25] and the Technicolor dataset [51]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 338, + 480, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 338, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 480, + 456 + ], + "type": "text", + "content": "Neural 3D Video comprises real-world dynamic scenes captured with a time-synchronized multi-view system at " + }, + { + "bbox": [ + 130, + 338, + 480, + 456 + ], + "type": "inline_equation", + "content": "2028 \\times 2704" + }, + { + "bbox": [ + 130, + 338, + 480, + 456 + ], + "type": "text", + "content": " resolution at 30 FPS. Camera parameters are estimated using COLMAP [53]. We compare our method to K-Planes [14], HexPlane [7], MixVoxels [61], HyperReel [2], NeRFPlayer [56], StreamRF [23], and 4DGS [65]. We compute quantitative metrics for the central test view at half the original resolution (" + }, + { + "bbox": [ + 130, + 338, + 480, + 456 + ], + "type": "inline_equation", + "content": "1014 \\times 1352" + }, + { + "bbox": [ + 130, + 338, + 480, + 456 + ], + "type": "text", + "content": ") for 300 frames. The average results for each method over the dataset are given in Table 1, while Table 3 provides a breakdown of the per-scene performance. The results show that our method performs best regarding PSNR and SSIM metrics while offering the fastest rendering performance. Qualitative results are shown in Fig. 5." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 458, + 480, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 458, + 480, + 517 + ], + "spans": [ + { + "bbox": [ + 130, + 458, + 480, + 517 + ], + "type": "text", + "content": "Technicolor captures real dynamic scenes from a synchronized " + }, + { + "bbox": [ + 130, + 458, + 480, + 517 + ], + "type": "inline_equation", + "content": "4 \\times 4" + }, + { + "bbox": [ + 130, + 458, + 480, + 517 + ], + "type": "text", + "content": " camera array at " + }, + { + "bbox": [ + 130, + 458, + 480, + 517 + ], + "type": "inline_equation", + "content": "2048 \\times 1088" + }, + { + "bbox": [ + 130, + 458, + 480, + 517 + ], + "type": "text", + "content": " resolution. Following [2], we evaluate on the second row second column camera on five scenes: Birthday, Fabien, Painter, Theater and Trains. We compare to DynNeRF [25], HyperReel [2] and Dynamic 3D Gaussians [35], with quantitative and qualitative results in Table 4 and Fig. 6 respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 536, + 246, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 246, + 548 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 246, + 548 + ], + "type": "text", + "content": "4.1 Ablation Studies" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 558, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 481, + 666 + ], + "type": "text", + "content": "This section provides ablations showing the effectiveness of our sliding window and self-supervised temporal consistency fine-tuning strategies. As we train independent models for each window, flickering artifacts can occur in the final renders. To visualize this, Fig. 8 plots absolute image error between renders of neighbouring frames (overlapping frames outlined in red). Without temporal consistency, we observe a spike in absolute error on overlapping frames, resulting in undesirable flickering. However, after fine-tuning, the error in overlapping frames is drastically reduced. In Table 5, we compute per-frame image metrics and estimate a measure of the temporal consistency using SoTA video quality" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "text", + "content": "R. Shaw et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 194, + 481, + 403 + ], + "blocks": [ + { + "bbox": [ + 132, + 139, + 481, + 185 + ], + "lines": [ + { + "bbox": [ + 132, + 139, + 481, + 185 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 481, + 185 + ], + "type": "text", + "content": "Table 3: Quantitative results on the Neural 3D Video dataset [25], evaluated for 300 frames at " + }, + { + "bbox": [ + 132, + 139, + 481, + 185 + ], + "type": "inline_equation", + "content": "1352 \\times 1014" + }, + { + "bbox": [ + 132, + 139, + 481, + 185 + ], + "type": "text", + "content": " resolution. †As reported in [2], ‡Natively trained in lower resolution, upscaled. Best and second best results highlighted respectively. Our method performs competitively in all metrics, usually coming in either first or second place." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 194, + 481, + 403 + ], + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 403 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 403 + ], + "type": "table", + "html": "
Scene\nMethodCook SpinachCut Roast BeefFlame Steak
PSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPS
MixVoxels [61]31.390.9310.11331.380.9280.11130.150.9380.108
K-Planes [14]31.230.9260.11431.870.9280.11431.490.9400.102
HexPlane‡ [7]31.050.9280.11430.830.9270.11530.420.9390.104
HyperReel [2]31.770.9320.09032.250.9360.08631.480.9390.083
NeRFPlayer† [56]30.580.9290.11329.350.9080.14431.930.9500.088
StreamRF [23]30.890.9140.16230.750.9170.15431.370.9230.152
Ours31.960.9460.09431.840.9450.09932.180.9530.087
Sear steakCoffee MartiniFlame Salmon
MixVoxels [61]30.850.9400.10329.250.9010.14729.500.8980.163
K-Planes [14]30.280.9370.10429.300.9000.13429.580.9010.132
HexPlane‡ [7]30.000.9390.10528.450.8910.14929.230.9050.088
HyperReel [2]31.880.9420.08028.650.8970.12928.260.9410.136
NeRFPlayer† [56]29.130.9080.13831.530.9510.08531.650.9400.098
StreamRF [23]31.600.9250.14728.130.8730.21928.690.8720.229
Ours32.210.9500.09229.160.9210.10529.250.9250.100
", + "image_path": "cb5918eee4c8df66e596aea7fdb190a779cc58a2aed28bbbccecf4b239059144.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 133, + 494, + 481, + 637 + ], + "blocks": [ + { + "bbox": [ + 132, + 460, + 480, + 483 + ], + "lines": [ + { + "bbox": [ + 132, + 460, + 480, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 480, + 483 + ], + "type": "text", + "content": "Table 4: Quantitative results on the Technicolor dataset [51] evaluated at full resolution. Best and second best results highlighted respectively." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 494, + 481, + 637 + ], + "lines": [ + { + "bbox": [ + 133, + 494, + 481, + 637 + ], + "spans": [ + { + "bbox": [ + 133, + 494, + 481, + 637 + ], + "type": "table", + "html": "
SceneBirthdayFabienPainter
MethodPSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPS
DyNeRF [25]29.200.9090.06732.760.9090.24235.950.9300.147
HyperReel [2]30.790.9220.06232.280.8600.21735.680.9260.123
Dynamic3DG [35]27.060.8590.11326.340.8340.26825.180.7580.395
Ours33.440.9590.04234.430.9250.17136.760.9480.128
SceneTheaterTrainAverage
DyNeRF [25]29.530.8750.18831.580.9330.06731.800.9110.142
HyperReel [2]33.670.8950.10430.100.9090.06132.500.9020.113
Dynamic3DG [35]28.050.7990.27728.460.9080.08827.020.8320.228
Ours29.810.8840.20133.990.9570.04333.690.9340.117
", + "image_path": "c02a05478f2744b6422de5bc37ab35c8ec6c75c8a6df04413654b64288be65b4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "SWinGS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 151, + 127, + 452, + 365 + ], + "blocks": [ + { + "bbox": [ + 151, + 127, + 452, + 365 + ], + "lines": [ + { + "bbox": [ + 151, + 127, + 452, + 365 + ], + "spans": [ + { + "bbox": [ + 151, + 127, + 452, + 365 + ], + "type": "image", + "image_path": "bcf93faffbc7ba5e1f00a23b627d57d3e8172798dcb6349ef25b0ba85d0f6d06.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 373, + 482, + 396 + ], + "lines": [ + { + "bbox": [ + 132, + 373, + 482, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 373, + 482, + 396 + ], + "type": "text", + "content": "Fig. 5: Qualitative results on Neural 3D Video [25]. Scenes top to bottom: i) coffee martini, ii) cook spinach, iii) cut roasted beef, iv) flame salmon, v) flame steak." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 155, + 430, + 451, + 616 + ], + "blocks": [ + { + "bbox": [ + 155, + 430, + 451, + 616 + ], + "lines": [ + { + "bbox": [ + 155, + 430, + 451, + 616 + ], + "spans": [ + { + "bbox": [ + 155, + 430, + 451, + 616 + ], + "type": "image", + "image_path": "21b4ecc10dd87b69a1ba7abf40eeb7e45537e7b3295d88bf06eba3b0671b091b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 625, + 480, + 647 + ], + "lines": [ + { + "bbox": [ + 132, + 625, + 480, + 647 + ], + "spans": [ + { + "bbox": [ + 132, + 625, + 480, + 647 + ], + "type": "text", + "content": "Fig. 6: Qualitative results on Technicolor [51]. Scenes top to bottom: i) Birthday, ii) Painter, iii) Train." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "text", + "content": "R. Shaw et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "assessment metric FAST-VQA [66], where the quality score is in the range [0,1]. The table provides results averaged over all scenes from the Neural 3D Video dataset [25]. Although we incur a minor penalty in some per-frame performance metrics (SSIM and LPIPS), we obtain a significantly higher video quality (VQA) score. This indicates a substantial improvement in temporal consistency and overall perceptual video quality, resulting in more pleasing renderings." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 154, + 261, + 457, + 312 + ], + "blocks": [ + { + "bbox": [ + 130, + 207, + 483, + 252 + ], + "lines": [ + { + "bbox": [ + 130, + 207, + 483, + 252 + ], + "spans": [ + { + "bbox": [ + 130, + 207, + 483, + 252 + ], + "type": "text", + "content": "Table 5: Ablation on temporal fine-tuning. Results averaged over all scenes from [25]. Temporal consistency is measured using t-LPIPS [10] and FAST-VQA [66] (quality score in the range [0,1]). Per-frame performance remains fairly constant, but temporal consistency and overall perceptual video quality is significantly improved." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 154, + 261, + 457, + 312 + ], + "lines": [ + { + "bbox": [ + 154, + 261, + 457, + 312 + ], + "spans": [ + { + "bbox": [ + 154, + 261, + 457, + 312 + ], + "type": "table", + "html": "
MethodPSNR ↑SSIM ↑LPIPS ↓t-LPIPS ↓VQA ↑
w/o temporal consistency32.010.9560.0850.01290.666
w/ temporal consistency32.050.9490.0930.01020.726
Ground truth----0.763
", + "image_path": "d84f6286710ef22b89cd3ec34cf477d388614a1c0efef27c7ecc797a7996c84b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 132, + 415, + 482, + 500 + ], + "blocks": [ + { + "bbox": [ + 130, + 351, + 483, + 399 + ], + "lines": [ + { + "bbox": [ + 130, + 351, + 483, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 351, + 483, + 399 + ], + "type": "text", + "content": "Table 6: Ablation on sliding window size vs adaptive on Birthday scene [51]. We show the improvement from the dynamic MLP. Adaptive chooses window sizes that match, and in some metrics exceed, the best fixed-size window performance, striking balance between performance, temporal consistency (t-LPIPS) and training time (GPU hrs)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 415, + 482, + 500 + ], + "lines": [ + { + "bbox": [ + 132, + 415, + 482, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 415, + 482, + 500 + ], + "type": "table", + "html": "
Window SizeNo. WindowsTrain timePSNRSSIMLPIPSt-LPIPS
32416.033.120.9560.0480.0076
964.033.380.9590.0430.0053
1732.033.010.9560.0450.0049
2521.332.970.9560.0430.0048
4910.732.730.9550.0470.0051
Adaptive53.333.440.9590.0420.0051
Adaptive (w/o dyn. MLP)53.332.760.9570.0450.0062
", + "image_path": "8bb945534981a7ee776a1acbcf01a01236dfa330513c8ac914a6ebc2040d11a8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 520, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 520, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 520, + 482, + 605 + ], + "type": "text", + "content": "Table 6 and Fig. 7 show the impact of the sliding-window size hyperparameter and our adaptive sampling strategy. Adaptive sampling automatically chooses appropriate window sizes that match and sometimes exceed the best fixed-size window performance, striking a good balance between performance, temporal consistency (t-LPIPS [10]) and training time. Adaptive performs the best overall with fewer windows (less storage requirements). We also show the advantage gained from using a dynamic MLP vs a regular MLP, which improves all metrics." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 620, + 220, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 620, + 220, + 632 + ], + "spans": [ + { + "bbox": [ + 131, + 620, + 220, + 632 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "We have presented a method to render novel views of dynamic scenes by extending the 3DGS framework. Results show that our method produces high-" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 410, + 90, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 410, + 90, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 410, + 90, + 447, + 101 + ], + "type": "text", + "content": "SWinGS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 210 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 210 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 210 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 210 + ], + "type": "image", + "image_path": "c3697ee26a3e77e502cc4cee2086eb36ae715e7de5a7252cfccde31075409d3a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 217, + 482, + 262 + ], + "lines": [ + { + "bbox": [ + 130, + 217, + 482, + 262 + ], + "spans": [ + { + "bbox": [ + 130, + 217, + 482, + 262 + ], + "type": "text", + "content": "Fig. 7: Ablation on sliding window size vs adaptive window sampling (with and without dynamic MLP) on scenes from the Technicolor dataset [51]. The single-window scenario (w49) has the shortest training time but unsatisfactory visual quality, while adaptive windows offer the best balance between quality and training overhead." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 277, + 481, + 415 + ], + "blocks": [ + { + "bbox": [ + 133, + 277, + 481, + 415 + ], + "lines": [ + { + "bbox": [ + 133, + 277, + 481, + 415 + ], + "spans": [ + { + "bbox": [ + 133, + 277, + 481, + 415 + ], + "type": "image", + "image_path": "b1339bc65944398bb3e4dfe6a041fb1a31b9789f04fde8b81c44606f1741a0d9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 426, + 482, + 461 + ], + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 461 + ], + "type": "text", + "content": "Fig. 8: Ablation on temporal fine-tuning: we display the absolute error between neighbouring frame renders, with overlapping frames highlighted red. After fine-tuning, the error is substantially reduced and overall perceptual video quality is improved." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": "quality renderings, even with complex motions e.g. flames. Key to our approach is sliding-window processing, which adaptively partitions sequences into manageable chunks. Processing each window separately, allowing the canonical representation and deformation field to vary throughout the sequence, enables us to handle complex topological changes, and reduces the magnitude and variability of the 3D scene flow. In contrast, other methods that learn a single representation for whole sequences are impractical for long sequences and degrade in quality with increasing length. Introducing an MLP for each window learns the deformation field from each canonical representation to a set of per-frame 3D Gaussians. Moreover, learnable tuning parameters help disentangle static and dynamic parts of the scene, which we find essential for imbalanced scenes. Our ablations show self-supervised temporal consistency fine-tuning reduces temporal flickering and improves the overall perceptual video quality, with only a minor impact on per-frame performance metrics. Overall, our method performs strongly compared to recent SoTA quantitatively and obtains sharper, temporally-consistent results." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "text", + "content": "R. Shaw et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 138, + 140, + 472, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 472, + 150 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 472, + 150 + ], + "type": "text", + "content": "1. Alexa, M.: Linear Combination of Transformations. ACM Trans. Graph. (2002)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 152, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 152, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 138, + 152, + 481, + 183 + ], + "type": "text", + "content": "2. Attal, B., Huang, J.B., Richardt, C., Zollhöfer, M., Kopf, J., O'Toole, M., Kim, C.: HyperReel: High-fidelity 6-DoF Video with Ray-conditioned Sampling. In: CVPR (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 184, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 184, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 138, + 184, + 481, + 205 + ], + "type": "text", + "content": "3. Bansal, A., Vo, M., Sheikh, Y., Ramanan, D., Narasimhan, S.: 4D Visualization of Dynamic Events from Unconstrained Multi-view Videos. In: CVPR (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 206, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 206, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 138, + 206, + 481, + 226 + ], + "type": "text", + "content": "4. Bansal, A., Zollhöfer, M.: Neural Pixel Composition for 3D-4D View Synthesis from Multi-views. In: CVPR (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 228, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 481, + 259 + ], + "type": "text", + "content": "5. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-NeRF: A Multiscale Representation for Anti-aliasing Neural Radiance Fields. In: ICCV (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 261, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 261, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 138, + 261, + 481, + 293 + ], + "type": "text", + "content": "6. Bārsan, I.A., Liu, P., Pollefeys, M., Geiger, A.: Robust Dense Mapping for Large-scale Dynamic Environments. In: International Conference on Robotics and Automation (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 293, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 293, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 138, + 293, + 481, + 314 + ], + "type": "text", + "content": "7. Cao, A., CV, J.: HexPlane: A Fast Representation for Dynamic Scenes. In: CVPR (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 316, + 481, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 316, + 481, + 347 + ], + "spans": [ + { + "bbox": [ + 138, + 316, + 481, + 347 + ], + "type": "text", + "content": "8. Catley-Chandar, S., Shaw, R., Slabaugh, G., Pérez-Pellitero, E.: RoGUENeRF: A Robust Geometry-consistent Universal Enhancer for NeRF. arXiv preprint arXiv:2403.11909 (2024)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 348, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 348, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 138, + 348, + 481, + 369 + ], + "type": "text", + "content": "9. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: TensoRF: Tensorial Radiance Fields. In: ECCV (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 370, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 370, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 138, + 370, + 481, + 402 + ], + "type": "text", + "content": "0. Chu, M., Xie, Y., Mayer, J., Leal-Taix'e, L., Thuerey, N.: Learning Temporal Coherence via Self-supervision for GAN-based Video Generation. ACM Trans. Graph. 39, 75:1 - 75:13 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 403, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 403, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 138, + 403, + 481, + 435 + ], + "type": "text", + "content": "1. Dhamo, H., Nie, Y., Moreau, A., Song, J., Shaw, R., Zhou, Y., Pérez-Pellitero, E.: HeadGaS: Real-time Animatable Head Avatars via 3D Gaussian Splatting. arXiv preprint arXiv:2312.02902 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 436, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 436, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 138, + 436, + 481, + 456 + ], + "type": "text", + "content": "2. Du, Y., Zhang, Y., Yu, H.X., Tenenbaum, J.B., Wu, J.: Neural Radiance Flow for 4D View Synthesis and Video Processing. In: ICCV (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 457, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 457, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 138, + 457, + 481, + 490 + ], + "type": "text", + "content": "3. Fang, J., Yi, T., Wang, X., Xie, L., Zhang, X., Liu, W., Nießner, M., Tian, Q.: Fast Dynamic Radiance Fields with Time-aware Neural Voxels. In: SIGGRAPH Asia (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 491, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 491, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 138, + 491, + 481, + 511 + ], + "type": "text", + "content": "4. Fridovich-Keil, S., Meanti, G., Warburg, F.R., Recht, B., Kanazawa, A.: K-Planes: Explicit Radiance Fields in Space, Time, and Appearance. In: CVPR (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 512, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 512, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 512, + 481, + 533 + ], + "type": "text", + "content": "5. Hedman, P., Srinivasan, P.P., Mildenhall, B., Barron, J.T., Debevec, P.: Baking Neural Radiance Fields for Real-time View Synthesis. In: ICCV (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 534, + 481, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 534, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 138, + 534, + 481, + 566 + ], + "type": "text", + "content": "6. Hu, L., Zhang, H., Zhang, Y., Zhou, B., Liu, B., Zhang, S., Nie, L.: GaussianAvatar: Towards Realistic Human Avatar Modeling from a Single Video via Animatable 3D Gaussians. In: CVPR (2024)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 567, + 481, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 481, + 589 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 481, + 589 + ], + "type": "text", + "content": "7. Huang, Y.H., Sun, Y.T., Yang, Z., Lyu, X., Cao, Y.P., Qi, X.: SC-GS: Sparse-Controlled Gaussian Splatting for Editable Dynamic Scenes. In: CVPR (2024)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 590, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 590, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 138, + 590, + 481, + 621 + ], + "type": "text", + "content": "8. Isik, M., Rünz, M., Georgopoulos, M., Khakhulin, T., Starck, J., Agapito, L., Niešner, M.: HumanRF: High-fidelity Neural Radiance Fields for Humans in Motion. ACM Trans. Graph. 42(4), 1-12 (2023)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 622, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 622, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 622, + 481, + 643 + ], + "type": "text", + "content": "9. Joo, H., Soo Park, H., Sheikh, Y.: MAP Visibility Estimation for Large-scale Dynamic 3D Reconstruction. In: CVPR (2014)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "type": "text", + "content": "20. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3D Gaussian Splatting for Real-time Radiance Field Rendering. ACM Trans. Graph. 42(4) (2023)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "SWinGS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "21. Kocabas, M., Chang, J.H.R., Gabriel, J., Tuzel, O., Ranjan, A.: HUGS: Human Gaussian Splatting. In: CVPR (2024)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 138, + 480, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 138, + 480, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 138, + 480, + 159 + ], + "type": "text", + "content": "22. Li, L., Shen, Z., Wang, Z., Shen, L., Tan, P.: D-NeRF: Neural Radiance Fields for Dynamic Scenes. In: CVPR (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 160, + 480, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 480, + 181 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 480, + 181 + ], + "type": "text", + "content": "23. Li, L., Shen, Z., Wang, Z., Shen, L., Tan, P.: Streaming Radiance Fields for 3D Video Synthesis. In: NeurIPS (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 182, + 480, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 480, + 202 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 480, + 202 + ], + "type": "text", + "content": "24. Li, R., Tanke, J., Vo, M., Zollhöfer, M., Gall, J., Kanazawa, A., Lassner, C.: TAVA: Template-free Animatable Volumetric Actors. In: ECCV (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 203, + 480, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 203, + 480, + 235 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 480, + 235 + ], + "type": "text", + "content": "25. Li, T., Slavcheva, M., Zollhöfer, M., Green, S., Lassner, C., Kim, C., Schmidt, T., Lovegrove, S., Goesele, M., Newcombe, R., Lv, Z.: Neural 3D Video Synthesis from Multi-view Video. In: CVPR (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 236, + 480, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 480, + 257 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 480, + 257 + ], + "type": "text", + "content": "26. Li, Z., Chen, Z., Li, Z., Xu, Y.: Spacetime Gaussian Feature Splatting for Real-time Dynamic View Synthesis. In: CVPR (2024)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 258, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 258, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 258, + 480, + 289 + ], + "type": "text", + "content": "27. Li, Z., Zheng, Z., Wang, L., Liu, Y.: Animatable Gaussians: Learning Pose-dependent Gaussian Maps for High-fidelity Human Avatar Modeling. In: CVPR (2024)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 289, + 480, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 480, + 311 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 480, + 311 + ], + "type": "text", + "content": "28. Li, Z., Niklaus, S., Snavely, N., Wang, O.: Neural Scene Flow Fields for Space-time View Synthesis of Dynamic Scenes. In: CVPR (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 312, + 480, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 312, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 132, + 312, + 480, + 331 + ], + "type": "text", + "content": "29. Li, Z., Wang, Q., Cole, F., Tucker, R., Snavely, N.: DynIBaR: Neural Dynamic Image-Based Rendering. In: CVPR (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 332, + 480, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 480, + 354 + ], + "type": "text", + "content": "30. Lin, K.E., Xiao, L., Liu, F., Yang, G., Ramamoorthi, R.: Deep 3D Mask Volume for View Synthesis of Dynamic Scenes. In: ICCV (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 354, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 480, + 374 + ], + "type": "text", + "content": "31. Lin, Y., Dai, Z., Zhu, S., Yao, Y.: Gaussian-Flow: 4D Reconstruction with Dynamic 3D Gaussian Particle. In: CVPR (2024)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 375, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 375, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 375, + 480, + 396 + ], + "type": "text", + "content": "32. Liu, Y.L., Gao, C., Meuleman, A., Tseng, H.Y., Saraf, A., Kim, C., Chuang, Y.Y., Kopf, J., Huang, J.B.: Robust Dynamic Radiance Fields. In: CVPR (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 397, + 480, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 480, + 429 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 480, + 429 + ], + "type": "text", + "content": "33. Lombardi, S., Simon, T., Saragih, J., Schwartz, G., Lehrmann, A., Sheikh, Y.: Neural Volumes: Learning Dynamic Renderable Volumes from Images. ACM Trans. Graph. 38(4) (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 430, + 480, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 480, + 450 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 480, + 450 + ], + "type": "text", + "content": "34. Luiten, J., Fischer, T., Leibe, B.: Track to Reconstruct and Reconstruct to Track. IEEE Robotics and Automation Letters 5(2), 1803-1810 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 450, + 480, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 480, + 472 + ], + "type": "text", + "content": "35. Luiten, J., Kopanas, G., Leibe, B., Ramanan, D.: Dynamic 3D Gaussians: Tracking by Persistent Dynamic View Synthesis. In: 3DV (2024)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 472, + 480, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 480, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 480, + 493 + ], + "type": "text", + "content": "36. Maggioni, M., Tanay, T., Babiloni, F., McDonagh, S.G., Leonardis, A.: Tunable Convolutions with Parametric Multi-loss Optimization. In: CVPR (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 494, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 494, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 494, + 480, + 525 + ], + "type": "text", + "content": "37. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. In: ECCV (2020)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 525, + 480, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 480, + 546 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 480, + 546 + ], + "type": "text", + "content": "38. Moreau, A., Song, J., Dhamo, H., Shaw, R., Zhou, Y., Pérez-Pellitero, E.: Human Gaussian Splatting: Real-time Rendering of Animatable Avatars. In: CVPR (2024)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 547, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 547, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 547, + 480, + 578 + ], + "type": "text", + "content": "39. Müller, T., Evans, A., Schied, C., Keller, A.: Instant Neural Graphics Primitives with a Multiresolution Hash Encoding. ACM Trans. Graph. 41(4), 102:1-102:15 (2022)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 579, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 579, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 579, + 480, + 601 + ], + "type": "text", + "content": "40. Newcombe, R.A., Fox, D., Seitz, S.M.: DynamicFusion: Reconstruction and Tracking of Non-rigid Scenes in Real-time. In: CVPR (2015)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 601, + 480, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 480, + 633 + ], + "type": "text", + "content": "41. Niemeyer, M., Barron, J.T., Mildenhall, B., Sajjadi, M.S.M., Geiger, A., Radwan, N.: RegNeRF: Regularizing Neural Radiance Fields for View Synthesis from Sparse Inputs. In: CVPR (2021)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "text", + "content": "42. Pang, H., Zhu, H., Kortylewski, A., Theobalt, C., Habermann, M.: ASH: Animatable Gaussian Splats for Efficient and Photoreal Human Rendering. In: CVPR (2024)" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "text", + "content": "R. Shaw et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "43. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin-Brualla, R.: Nerfies: Deformable Neural Radiance Fields. ICCV (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 172 + ], + "type": "text", + "content": "44. Park, K., Sinha, U., Hedman, P., Barron, J.T., Bouaziz, S., Goldman, D.B., Martin-Brualla, R., Seitz, S.M.: HyperNeRF: A Higher-dimensional Representation for Topologically Varying Neural Radiance Fields. ACM Trans. Graph. 40(6) (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 173, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 173, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 173, + 480, + 194 + ], + "type": "text", + "content": "45. Peng, S., Dong, J., Wang, Q., Zhang, S., Shuai, Q., Zhou, X., Bao, H.: Animatable Neural Radiance Fields for Modeling Dynamic Human Bodies. In: ICCV (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 195, + 480, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 195, + 480, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 195, + 480, + 217 + ], + "type": "text", + "content": "46. Peng, S., Yan, Y., Shuai, Q., Bao, H., Zhou, X.: Representing Volumetric Videos as Dynamic MLP Maps. In: CVPR (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 218, + 480, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 218, + 480, + 250 + ], + "spans": [ + { + "bbox": [ + 132, + 218, + 480, + 250 + ], + "type": "text", + "content": "47. Qian, S., Kirschstein, T., Schoneveld, L., Davoli, D., Giebenhain, S., Nießner, M.: GaussianAvatars: Photorealistic Head Avatars with Rigged 3D Gaussians. In: CVPR (2024)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 251, + 480, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 251, + 480, + 274 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 480, + 274 + ], + "type": "text", + "content": "48. Qian, Z., Wang, S., Mihajlovic, M., Geiger, A., Tang, S.: 3DGS-Avatar: Animatable Avatars via Deformable 3D Gaussian Splitting. In: CVPR (2024)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 274, + 480, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 480, + 295 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 480, + 295 + ], + "type": "text", + "content": "49. Reiser, C., Peng, S., Liao, Y., Geiger, A.: KiloNeRF: Speeding up Neural Radiance Fields with Thousands of Tiny MLPs. In: ICCV (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 296, + 480, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 296, + 480, + 318 + ], + "spans": [ + { + "bbox": [ + 132, + 296, + 480, + 318 + ], + "type": "text", + "content": "50. Rong, X., Huang, J.B., Saraf, A., Kim, C., Kopf, J.: Boosting View Synthesis with Residual Transfer. In: CVPR (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 319, + 480, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 319, + 480, + 340 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 480, + 340 + ], + "type": "text", + "content": "51. Sabater, N., Boisson, G., Vandame, B., Kerbiriou, P., Babon, F., et al.: Dataset and Pipeline for Multi-view Light-Field Video. In: CVPRW (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 341, + 480, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 341, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 341, + 480, + 363 + ], + "type": "text", + "content": "52. Saito, S., Schwartz, G., Simon, T., Li, J., Nam, G.: Relightable Gaussian Codec Avatars. In: CVPR (2024)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 364, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 480, + 385 + ], + "type": "text", + "content": "53. Schonberger, J.L., Frahm, J.M.: Structure-from-Motion Revisited. In: CVPR (2016)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "type": "text", + "content": "54. Shao, R., Zheng, Z., Tu, H., Liu, B., Zhang, H., Liu, Y.: Tensor4D: Efficient Neural 4D Decomposition for High-fidelity Dynamic Reconstruction and Rendering. In: CVPR (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 419, + 480, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 419, + 480, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 419, + 480, + 441 + ], + "type": "text", + "content": "55. Slavcheva, M., Baust, M., Cremers, D., Ilic, S.: KillingFusion: Non-rigid 3D Reconstruction without Correspondences. In: CVPR (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 442, + 480, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 442, + 480, + 485 + ], + "spans": [ + { + "bbox": [ + 132, + 442, + 480, + 485 + ], + "type": "text", + "content": "56. Song, L., Chen, A., Li, Z., Chen, Z., Chen, L., Yuan, J., Xu, Y., Geiger, A.: NeRF-Player: A Streamable Dynamic Scene Representation with Decomposed Neural Radiance Fields. IEEE Transactions on Visualization and Computer Graphics 29(5), 2732-2742 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 487, + 480, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 480, + 518 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 480, + 518 + ], + "type": "text", + "content": "57. Sun, J., Jiao, H., Li, G., Zhang, Z., Zhao, L., Xing, W.: 3DGStream: On-the-fly Training of 3D Gaussians for Efficient Streaming of Photo-realistic Free-viewpoint Videos. In: CVPR (2024)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 520, + 460, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 520, + 460, + 531 + ], + "spans": [ + { + "bbox": [ + 132, + 520, + 460, + 531 + ], + "type": "text", + "content": "58. Tanay, T., Maggioni, M.: Global Latent Neural Rendering. In: CVPR (2024)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 532, + 480, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 532, + 480, + 553 + ], + "spans": [ + { + "bbox": [ + 132, + 532, + 480, + 553 + ], + "type": "text", + "content": "59. Teed, Z., Deng, J.: RAFT: Recurrent All-pairs Field Transforms for Optical Flow. In: ECCV (2020)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 554, + 480, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 554, + 480, + 575 + ], + "spans": [ + { + "bbox": [ + 132, + 554, + 480, + 575 + ], + "type": "text", + "content": "60. Turki, H., Zhang, J.Y., Ferroni, F., Ramanan, D.: SUDS: Scalable Urban Dynamic Scenes. In: CVPR (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 576, + 480, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 576, + 480, + 598 + ], + "spans": [ + { + "bbox": [ + 132, + 576, + 480, + 598 + ], + "type": "text", + "content": "61. Wang, F., Tan, S., Li, X., Tian, Z., Liu, H.: Mixed Neural Voxels for Fast Multiview Video Synthesis. In: ICCV (2023)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 599, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 480, + 632 + ], + "type": "text", + "content": "62. Wang, L., Zhang, J., Liu, X., Zhao, F., Zhang, Y., Zhang, Y., Wu, M., Yu, J., Xu, L.: Fourier PlenOctrees for Dynamic Radiance Field Rendering in Real-time. In: CVPR (2022)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "63. Wang, Q., Wang, Z., Genova, K., Srinivasan, P., Zhou, H., Barron, J.T., Martin-Brualla, R., Snavely, N., Funkhouser, T.: IBRNet: Learning Multi-view Image-Based Rendering. In: CVPR (2021)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 410, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 410, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 410, + 91, + 447, + 100 + ], + "type": "text", + "content": "SWinGS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 435 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 148 + ], + "type": "text", + "content": "64. Weng, C.Y., Curless, B., Srinivasan, P.P., Barron, J.T., Kemelmacher-Shlizerman, I.: HumanNeRF: Free-viewpoint Rendering of Moving People from Monocular Video. In: CVPR (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 149, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 480, + 171 + ], + "type": "text", + "content": "65. Wu, G., Yi, T., Fang, J., Xie, L., Zhang, X., Wei, W., Liu, W., Tian, Q., Wang, X.: 4D Gaussian Splatting for Real-time Dynamic Scene Rendering. In: CVPR (2024)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 204 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 204 + ], + "type": "text", + "content": "66. Wu, H., Chen, C., Hou, J., Liao, L., Wang, A., Sun, W., Yan, Q., Lin, W.: FASTVQA: Efficient End-to-end Video Quality Assessment with Fragment Sampling. In: ECCV (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 225 + ], + "type": "text", + "content": "67. Xian, W., Huang, J.B., Kopf, J., Kim, C.: Space-time Neural Irradiance Fields for Free-viewpoint Video. In: CVPR (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 480, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 247 + ], + "type": "text", + "content": "68. Xiang, J., Gao, X., Guo, Y., Zhang, J.: FlashAvatar: High-fidelity Head Avatar with Efficient Gaussian Embedding. In: CVPR (2024)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "text", + "content": "69. Xing, W., Chen, J.: Temporal-MPI: Enabling Multi-Plane Images for Dynamic Scene Modelling via Temporal Basis Learning. In: ECCV (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 270, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 480, + 293 + ], + "type": "text", + "content": "70. Xu, Y., Chen, B., Li, Z., Zhang, H., Wang, L., Zheng, Z., Liu, Y.: Gaussian HeadAvatar: Ultra High-fidelity Head Avatar via Dynamic Gaussians. In: CVPR (2024)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "type": "text", + "content": "71. Yang, Z., Yang, H., Pan, Z., Zhang, L.: Real-time Photorealistic Dynamic Scene Representation and Rendering with 4D Gaussian Splatting. In: ICLR (2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "type": "text", + "content": "72. Yang, Z., Gao, X., Zhou, W., Jiao, S., Zhang, Y., Jin, X.: Deformable 3D Gaussians for High-fidelity Monocular Dynamic Scene Reconstruction. In: CVPR (2024)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "type": "text", + "content": "73. Yu, A., Li, R., Tancik, M., Li, H., Ng, R., Kanazawa, A.: PlenOctrees for Real-time Rendering of Neural Radiance Fields. In: ICCV (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "type": "text", + "content": "74. Zhao, F., Yang, W., Zhang, J., Lin, P., Zhang, Y., Yu, J., Xu, L.: HumanNeRF: Efficiently Generated Human Radiance Field from Sparse Inputs. In: CVPR (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 380, + 480, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 412 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 412 + ], + "type": "text", + "content": "75. Zhou, K., Li, W., Wang, Y., Hu, T., Jiang, N., Han, X., Lu, J.: NeRFLiX: High-quality Neural View Synthesis by Learning a Degradation-driven Inter-viewpoint MiXer. In: CVPR (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 413, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 480, + 435 + ], + "type": "text", + "content": "76. Zhou, T., Tucker, R., Flynn, J., Fyffe, G., Snavely, N.: Stereo Magnification: Learning View Synthesis using Multiplane Images. ACM Trans. Graph. 37 (2018)" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "text", + "content": "R. Shaw et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/S^3D-NeRF_ Single-Shot Speech-Driven Neural Radiance Field for High Fidelity Talking Head Synthesis/49757949-cf34-4879-95e0-e53ca1bcc2a6_content_list.json b/2024/S^3D-NeRF_ Single-Shot Speech-Driven Neural Radiance Field for High Fidelity Talking Head Synthesis/49757949-cf34-4879-95e0-e53ca1bcc2a6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..02c5c18c4faf621fa9c9d4e15cf04c2b7466ec3e --- /dev/null +++ b/2024/S^3D-NeRF_ Single-Shot Speech-Driven Neural Radiance Field for High Fidelity Talking Head Synthesis/49757949-cf34-4879-95e0-e53ca1bcc2a6_content_list.json @@ -0,0 +1,1684 @@ +[ + { + "type": "text", + "text": "$\\mathrm{S}^{3}\\mathrm{D}-\\mathrm{NeRF}$ : Single-Shot Speech-Driven Neural Radiance Field for High Fidelity Talking Head Synthesis", + "text_level": 1, + "bbox": [ + 236, + 141, + 767, + 210 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dongze Li $^{1,2}$ Kang Zhao $^{3}$ , Wei Wang $^{2\\star}$ , Yifeng Ma $^{3}$ , Bo Peng $^{2}$ , Yingya Zhang $^{3}$ and Jing Dong $^{2}$", + "bbox": [ + 308, + 236, + 692, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Artificial Intelligence, University of Chinese Academy of Sciences", + "bbox": [ + 243, + 277, + 754, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ NLPR & MAIS, Institute of Automation, Chinese Academy of Sciences", + "bbox": [ + 256, + 292, + 745, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Alibaba Group", + "bbox": [ + 444, + 306, + 558, + 321 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "dongze.li@cripac.ia.ac.cn; {wang,bo.peng,jdong}@nlpr.ia.ac.cn", + "bbox": [ + 263, + 321, + 738, + 335 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zhaokang.zk, yingya.zyy, mayifeng.myf}@alibaba-inc.com", + "bbox": [ + 287, + 335, + 714, + 349 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Talking head synthesis is a practical technique with wide applications. Current Neural Radiance Field (NeRF) based approaches have shown their superiority on driving one-shot talking heads with videos or signals regressed from audio. However, most of them failed to take the audio as driven information directly, unable to enjoy the flexibility and availability of speech. Since mapping audio signals to face deformation is non-trivial, we design a Single-Shot Speech-Driven Neural Radiance Field (S $^3$ D-NeRF) method in this paper to tackle the following three difficulties: learning a representative appearance feature for each identity, modeling motion of different face regions with audio, and keeping the temporal consistency of the lip area. To this end, we introduce a Hierarchical Facial Appearance Encoder to learn multi-scale representations for catching the appearance of different speakers, and elaborate a Cross-modal Facial Deformation Field to perform speech animation according to the relationship between the audio signal and different face regions. Moreover, to enhance the temporal consistency of the important lip area, we introduce a lip-sync discriminator to penalize the out-of-sync audio-visual sequences. Extensive experiments have shown that our S $^3$ D-NeRF surpasses previous arts on both video fidelity and audio-lip synchronization.", + "bbox": [ + 259, + 388, + 740, + 665 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Talking Head $\\cdot$ Neural Radiance Fields", + "bbox": [ + 261, + 679, + 611, + 694 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 720, + 375, + 736 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Speech driven talking head synthesis is a promising technique and can be applied to a wide range of situations such as digital human, film making, virtual reality and video games. Current Neural Radiance Field (NeRF) [24] based methods [11,17,38] have shown their superiority on generating vivid talking portraits", + "bbox": [ + 212, + 753, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author", + "bbox": [ + 217, + 824, + 382, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "with high quality for their 3D consistency and view controllability. However, they can only be applied to a specific identity, and require a long video sequence from the same speaker to train on, which hampers them from broader application scenarios. Afterwards some progress have been made in the generalization of NeRF-based talking head methods. Most of them [18, 23, 41] are driven by intermediate facial representations, such as 3DMM coefficients extracted from video or expression coefficients regressed from audio. These intermediate representations may introduce information loss more or less, leading to sub-optimal synthesis results.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Considering it is non-trivial to map audio signals to face deformation, we attribute the challenge of performing high-fidelity one-shot speech animation with Neural Radiance Fields to three aspects: 1) Learning Representative Appearance Features for Each Identity. Different people have different facial appearance i.e. shape and texture, it is difficult for a vanilla NeRF to model these details of several speakers simultaneously. During the inference phase, it becomes even harder to generate a talking head sequence with high quality given that only a single image is available. 2) Modeling Motion of Different Face Regions with Audio. Video driven techniques can utilize global driving signals which can describe the motion of the whole face, such as depth priors [12], 3DMM meshes [41] or PNCC [10,18] maps. However, in a speech driven task, the input audio merely has strong correlation with the lower face [27,34]. Directly predicting face motions according to the input audio will cause insufficient modeling of different face regions, harming the fidelity of the synthesized portrait. 3) Keeping Temporal Consistency of the Lip Area. The lip area has the greatest importance in speech animation. Due to the lack of constraints on this part, videos generated by NeRFs have sometimes shown incorrect mouth shapes.", + "bbox": [ + 212, + 287, + 787, + 545 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose S $^3$ D-NeRF, namely Single-Shot Speech-Driven Neural Radiance Field, to synthesize high fidelity talking head videos. Given a single shot source image and a driven audio sequence, our method can synthesize vivid free view talking head videos with accurate mouth shapes.", + "bbox": [ + 212, + 547, + 789, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To fully capture the appearance of an arbitrary speaker, a Hierarchical Facial Appearance Encoder is presented to extract the multi-scale facial features of the single-shot source image with a down sample convolutional network armed with a feature pyramid. These features contain rich structural and textural information of the source identity. Then, a multi-scale tri-plane [4] representation is constructed for neural rendering. To realize accurate speech animation, a Cross-modal Facial Deformation Field is proposed to faithfully catch the correlations between the speech signal and the visual features from different face regions through cross attention, and predicts the motion of the whole face precisely. To further enhance the consistency of the lip area, a lip-sync discriminator is imposed on the lower face of the generated video frames and the driving audio signal. By penalizing the out-of-sync audio-visual sequences, higher temporal consistency of the lip area can be achieved. During training, we adapt a coarse-to-fine image generation strategy to reduce the difficulty of modeling multiple speakers at the same time. Specifically, a coarse talking head frame which", + "bbox": [ + 212, + 613, + 789, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "D. Li et al.", + "bbox": [ + 271, + 114, + 346, + 127 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2942f86784c916ec195debd937915bdf9e263090c948fad1f41231fdea451f57.jpg", + "image_caption": [ + "Fig. 1: A showcase of our S $^3$ D-NeRF, which generates high quality face portraits with fine-grained face texture and mouth details." + ], + "image_footnote": [], + "bbox": [ + 251, + 148, + 754, + 330 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "contains the important inner face area is synthesized through standard volume neural rendering, while the texture details are refined with a super-resolution module, resulting in a high fidelity talking head frame.", + "bbox": [ + 212, + 396, + 782, + 441 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our main contributions can be summarized as below:", + "bbox": [ + 238, + 441, + 625, + 455 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose S $^3$ D-NeRF, a single-shot NeRF-based talking head synthesis framework. Our method extends NeRF-based speech animation techniques to handle arbitrary unseen identities.", + "- Several key components are presented to assist the talking head synthesis procedure. Including a Hierarchical Facial Appearance Encoder for representative feature extraction, a Cross-modal Facial Deformation Field for accurate speech animation, and a lip-sync discriminator for better temporal consistency of lip area.", + "- Comprehensive experiments have shown the superiority of our S $^3$ D-NeRF over previous arts on both video fidelity and audio-lip synchronization." + ], + "bbox": [ + 225, + 469, + 784, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 645, + 387, + 660 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Speech Driven Talking Head Synthesis", + "text_level": 1, + "bbox": [ + 214, + 676, + 578, + 694 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The goal of speech driven talking head synthesis is to animate a speaker according to input audiios. According to the generalization ability of the model to new identities, current speech driven methods can be divided into two categories: identity agnostic methods and identity specific methods.", + "bbox": [ + 212, + 703, + 782, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Identity agnostic methods are capable of generating speech videos of an arbitrary speaker once their training procedure is finished, with one or several images or video sequences as inputs. Among them, single-shot methods [22, 27, 43, 45] have attracted considerable interests due to their high generalization ability and data efficiency. Image-based methods [27, 37, 45] utilize deep generative", + "bbox": [ + 212, + 765, + 784, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "$\\mathrm{S}^3\\mathrm{D}$ -NeRF", + "bbox": [ + 656, + 113, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/321b5cb07b72ba26343cb38eb30e3a4471f3373f03017a8c8eb0b15111282280.jpg", + "image_caption": [ + "Fig. 2: The full pipeline of our S $^3$ D-NeRF. The Hierarchical Facial Appearance Encoder extracts representative features from the masked face region of the single-shot source image, for high fidelity neural rendering of an arbitrary identity. The Cross-modal Facial Deformation Field accurately models the motion of different face regions, with the help of the correlation score calculated through cross attention between audiovisual features. Texture details are complemented with the super-resolution module." + ], + "image_footnote": [], + "bbox": [ + 215, + 145, + 785, + 338 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "networks [9] or auto-encoders [16] to generate continuous video clips. Model-based methods predict intermediate representations such as motion field [36], 2D landmarks [46] or 3D Morphable Model (3DMM) [3] coefficients [22, 29, 43] with the input audio signal, and animate the head with these representations. StyleTalk [22] predicts 3DMM through a transformer architecture and can generate portraits with diverse speaking styles under the help of a pretrained renderer [29]. SadTalker [43] uses several representation networks to learn lip and expression basis. Although remarkable progresses have been made, current single-shot methods still suffer in image quality and audio-lip synchronization because of the error accumulated in the audio to representation mapping.", + "bbox": [ + 212, + 460, + 787, + 612 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Identity specific methods train an individual model for each identity with higher fidelity. Earlier methods [33, 40] utilize strong priors such as 3D meshes or dense key points for portrait generation. Current NeRF-based methods [11, 20, 31, 38] have shed light on high fidelity speaker synthesis. They optimize on the ground truth video clip of a single person, and can synthesize pose-controllable faces with fine-grained details. AD-NeRF [11] uses two separated NeRFs to model the head and the torso part respectively. SSP-NeRF [20] performs rays re-sampling based on the loss magnitude of different semantic regions. Currently, RAD-NeRF [32] and ER-NeRF [17] utilize hash table structures [25] to improve the rendering speed. GeneFace [39] proposes a facial radiance field with generalize ability but it also requires per identity finetuning and a complex multi-stage training process. Despite the above advantages, the above methods can only fit a single identity, and need tedious re-training when a new identity is encountered. At the same time, since no constraints are imposed on image sequences, the temporal consistency of the lip area is also not satisfactory.", + "bbox": [ + 212, + 613, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "D. Li et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 NeRF-based Face Portrait Reconstruction", + "text_level": 1, + "bbox": [ + 215, + 146, + 609, + 161 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "It is a natural idea to model human faces with NeRFs in a 3D aware manner. However, vanilla NeRFs struggle to handle dynamic objects, and are unsuitable for animating human faces which have diverse poses and expressions. There mainly exist two series of solutions for this problem. The former [14] is to condition the original Radiance Field with additional global control signals, e.g. expression parameters. The latter solve this problem by learning an additional deformation field [26, 28] to warp the 3D points. NerFace [8] is the first approach to reconstruct animatable face avatars with NeRFs conditioned on 3DMM expression parameters. HeadNeRF [13], NOFA [41] and OTAvatar [23] utilize GAN inversion to project onto the latent space of a 3D aware GAN [4] and drive the generated avatar through different latent codes. HiDe-NeRF [18] utilizes Projected Normalized Coordinate Code (PNCC) [10] as a global expression control indicator, and drives the face image with high fidelity. Comparing with the above mentioned methods, driving an arbitrary head with speech signal directly through NeRFs is a more challenging problem and is rarely explored.", + "bbox": [ + 212, + 170, + 787, + 397 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 417, + 330, + 434 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overall pipeline of our S $^3$ D-NeRF is shown in Fig.2. The 3D positions of the talking head can be derived through the given head pose of the target image $\\mathbf{P}_{tar}$ , which is available during both training and inference time. Firstly, our Hierarchical Facial Appearance Encoder takes a single-shot image as input, and models its appearance with a deep convolutional network, where multi-resolution feature maps are extracted to construct an efficient representation i.e. multiscale appearance tri-planes, which are known to be capable of modeling several identities faithfully at the same time [4, 18]. Feature points from these planes are retrieved according to the 3D position for rendering. Then, the motion of the talking face are predicted by our Cross-modal Facial Deformation Field. It firstly calculates the correlation between an aggregated visual feature embedding and the audio signal through cross attention, then takes the result as a prior knowledge to assist the regression process of the deformation, which is used for calibrating the 3D position. A coarse face is rendered through the classic volume rendering process [24] with color and density predicted from the feature points on the tri-plane. Finally, a super-resolution module is used to refine the coarse face and yields the high fidelity portrait.", + "bbox": [ + 212, + 446, + 787, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Hierarchical Facial Appearance Encoder", + "text_level": 1, + "bbox": [ + 215, + 726, + 591, + 742 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "An effective feature representation is important to accurately model the appearance of a speaker. Initially proposed in [4], tri-plane has shown its strong representative power and is much faster comparing with the vanilla NeRFs. The Hierarchical Facial Appearance Encoder constructs this representation based on multi-scale feature maps from the one-shot source image, and yields the feature points for further volumetric rendering process.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "$\\mathrm{S}^3\\mathrm{D}$ -NeRF", + "bbox": [ + 656, + 113, + 730, + 127 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Constructing Representative Appearance Features. We adopt a feature pyramid structure [19] to extract source features for its ability to learn multi-scale information from different resolutions, and reshape all the feature maps channelwisely to construct several orthogonal planes as in [4]. To be more concrete, given the source image $\\mathbf{I}_{src}$ , feature maps with different resolutions $\\{\\mathbf{D}^0,\\mathbf{D}^1,\\mathbf{D}^2,\\mathbf{D}^3\\}$ are extracted through several convolutional downsample blocks $Down^{i}(.)$ , while $\\mathbf{D}^0 = Down^0 (\\mathbf{I}_{src})$ and $\\mathbf{D}^i = Down^i (\\mathbf{D}^{i - 1})$ . Then, hierarchical facial appearance features can be obtained by connecting the feature map at current $i$ -th level with those at the higher $(i + 1)$ -th level, both of which are upsampled to the same resolution. $1\\times 1$ convolution layers are used to reduce the channel number to avoid overlarge computational cost:", + "bbox": [ + 212, + 146, + 787, + 313 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} ^ {i} = \\left\\{ \\begin{array}{l l} \\operatorname {C o n v} ^ {i} ([ U p ^ {i} (\\mathbf {F} ^ {i + 1}), \\mathbf {D} ^ {i} ]), & i = 0, 1, 2 \\\\ \\mathbf {D} ^ {i}, & i = 3 \\end{array} \\right. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 321, + 785, + 359 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $Up^{i}(.)$ denotes the $i$ -th upsampling convolution layer, and $Conv^{i}(.)$ denotes the $i$ -th $1 \\times 1$ convolution layer mentioned above. [...] denotes channelwise concatenation. Afterwards, the feature map at $i$ -th level $F^{i}$ has a shape of $B \\times 3 \\times C^{i} \\times H^{i} \\times W^{i}$ , and is reshaped and splitted into three orthogonal sub feature maps, each of which has a shape of $B \\times C^{i} \\times H^{i} \\times W^{i}$ , representing coordinates in the three-dimensional space, i.e. plane $\\mathbf{F}_{xy}$ , $\\mathbf{F}_{yz}$ and $\\mathbf{F}_{xz}$ respectively. $B$ is the batch size, and $C^{i}$ , $H^{i}$ and $W^{i}$ are the channel number, the height and width at $i$ -th level. Since this tri-plane representation is obtained from multiple identities, it is capable of fitting an unseen person during inference time.", + "bbox": [ + 212, + 369, + 787, + 505 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Extracting Features for Neural Rendering. To render the target image $\\mathbf{I}_{tar}$ , which may have different head pose with the source image, we need to find the correspondence between the position in the target image space and the source feature tri-plane space. This can be realized through camera transformation between the source pose $\\mathbf{P}_{src}$ and the target pose $\\mathbf{P}_{tar}$ , together with bilinear interpolation. Concretely, given a 3D point $\\mathbf{x}$ on a ray $\\mathbf{r}(t)$ constructed with target pose $\\mathbf{P}_{tar} = \\{\\mathbf{R}_{tar},\\mathbf{T}_{tar}\\}$ , it is firstly calibrated by the deformation predicted from the Cross-modal Facial Deformation Field (Sec.3.2) to get the deformed point $\\tilde{\\mathbf{x}}$ . Then, feature vectors $\\mathbf{F}_{xy}(\\tilde{\\mathbf{x}}),\\mathbf{F}_{yz}(\\tilde{\\mathbf{x}})$ and $\\mathbf{F}_{xz}(\\tilde{\\mathbf{x}})$ are retrieved from three orthogonal planes $\\mathbf{F}_{xy},\\mathbf{F}_{yz}$ and $\\mathbf{F}_{xz}$ , according to the pose of the source image $\\mathbf{P}_{src} = \\{\\mathbf{R}_{src},\\mathbf{T}_{src}\\}$ via bilinear interpolation Interp(:.):", + "bbox": [ + 212, + 506, + 787, + 674 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {x y z} ^ {i} (\\tilde {\\mathbf {x}}) = \\operatorname {I n t e r p} \\left(\\mathbf {F} ^ {i}, \\operatorname {I n v} \\left(\\mathbf {P} _ {s r c}\\right) \\cdot \\tilde {\\mathbf {x}}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 370, + 681, + 784, + 699 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$Inv(\\mathbf{P}_{src})$ is the inverse source pose which transforms 3D positions onto the source tri-plane space. These extracted feature vectors are further used for generating the talking head image through volume rendering.", + "bbox": [ + 212, + 705, + 787, + 752 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Cross-modal Facial Deformation Field", + "text_level": 1, + "bbox": [ + 214, + 771, + 576, + 786 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To animate the target speaker, a crucial step is to get the precise deformation which describe face motions driven by the speech signal. Solely predicting the deformation from audio struggles to learn the global motion of the whole face, since", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "D. Li et al.", + "bbox": [ + 271, + 114, + 346, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d1565c14fe61408989fd247862c40ea40430407f29470fd4ce80ba49a8093bfe.jpg", + "image_caption": [ + "Fig. 3: Results with naive deformation module (left) and our Cross-modal Facial Deformation Field (right). Lower Face regions have the largest activations in the heatmap, which denote their strongest correlations with the driven speech signal." + ], + "image_footnote": [], + "bbox": [ + 308, + 148, + 699, + 258 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "audio is a local signal which only correlates strongly with the mouth area. Thus it tends to yield blurry results (shown in Fig.3, left). For correctly modeling the global motion, structural and textural information of the whole face is required. Our Cross-modal Facial Deformation Field firstly aggregates the multi-scale visual features to a unified embedding, which contains the global information of the entire face area, then calculates the correlation score between the audio signal and this embedding through cross attention. The correlation score is used as a prior knowledge for determining the importance of motions from each part of the face, making deformation prediction more precise. As is shown in Fig.3, our Cross-modal Facial Deformation Field is free from the blurry issues because of the correctly learned motion.", + "bbox": [ + 212, + 329, + 787, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Aggregating Multi-scale Facial Features. To learn the features which represents the whole face, all the hierarchical appearance feature maps $\\mathbf{F}^0$ to $\\mathbf{F}^3$ are first reshaped to the same resolution $(H_s\\times W_s)$ and concatenated along the channel. Then, these features are flattened and fed into a slot-attention module [21] to exchange information from different scales: $\\mathbf{F}_{agg} = SlotAttention(\\mathbf{F}^{0\\dots 3})$ . The aggregated slot feature embedding has a shape of $B\\times (H_{s}\\times W_{s})\\times D_{slot}$ , where $D_{slot}$ is the slot feature dimension. This learning process makes the embedding contain rich structural and textural information of the speaker.", + "bbox": [ + 212, + 496, + 787, + 617 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Calculating Audio-Visual Correlation Scores. The input speech signal is firstly processed with a 1D full convolutional network to get the driving audio feature $\\mathbf{a}_{dri}$ . Relevance between audio and face motion is calculated through the Multi-Head Cross-Attention (MHCA) scores with the aggregated visual embedding $\\mathbf{F}_{agg}$ as query, the driving audio feature $\\mathbf{a}_{dri}$ as key and value:", + "bbox": [ + 212, + 618, + 787, + 694 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {F} _ {c m} = M H C A \\left(\\mathbf {F} _ {a g g}, \\mathbf {a} _ {d r i}\\right) \\\\ = \\operatorname {S o f t m a x} \\left[ \\frac {\\mathbf {F} _ {a g g} \\boldsymbol {W} ^ {Q} \\left(\\mathbf {a} _ {d r i} \\boldsymbol {W} ^ {K}\\right) ^ {T}}{\\sqrt {d}} \\right] \\mathbf {a} _ {d r i} \\boldsymbol {W} ^ {V}, \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 316, + 707, + 785, + 782 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\pmb{W}^Q, \\pmb{W}^K, \\pmb{W}^V$ are the projection matrices with hidden dimension $d$ , respectively. Cross-attention scores between the speech signal and the whole face feature representation indicate the relevance of each face area with the speech", + "bbox": [ + 212, + 792, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "$\\mathrm{S}^3\\mathrm{D}$ -NeRF", + "bbox": [ + 656, + 113, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b03b76721c2e183dfc3b9d08d1145499c23c4277c708f1eb3ebe8c88a2843164.jpg", + "image_caption": [ + "Fig. 4: Qualitative comparison with single-shot methods. Our $\\mathrm{S}^3\\mathrm{D}$ -NeRF yields the most correct lip shapes and the clearest teeth. Note that different methods adopt different face alignment tools, and the ground truth row demonstrates the raw image without alignment, so the face poses from different methods are slightly different." + ], + "image_footnote": [], + "bbox": [ + 225, + 148, + 782, + 392 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "input, and can serve as a strong prior for modeling the dynamic talking face in a region-aware manner. We observed that the mouth area of the speaker, which moves with the audio signal has the largest deformation scales. Also shown in Fig.3, heatmaps are calculated by averaging $\\mathbf{F}_{cm}$ channelwisely, and have shown larger activations on lower face regions which contains not only mouth area but also the muscles of the lower face.", + "bbox": [ + 212, + 478, + 787, + 568 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Predicting Facial Deformation. The facial deformation prediction module $Deform(\\ldots)$ has a U-Net [30] like architecture, with the original 3D position $\\mathbf{x}$ , the cross-modal correlation score $\\mathbf{F}_{cm}$ , and the driving audio feature $\\mathbf{a}_{dri}$ as inputs, this deformation prediction process can be formalized as", + "bbox": [ + 212, + 569, + 787, + 630 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\mathbf {x} = D e f o r m (\\mathbf {x}, \\mathbf {a} _ {d r i}, \\mathbf {F} _ {c m}),\n$$\n", + "text_format": "latex", + "bbox": [ + 393, + 636, + 606, + 652 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {x}} = \\mathbf {x} + \\Delta \\mathbf {x}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 410, + 652, + 784, + 669 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "With the calibrated 3D points, source features from the tri-plane can be extracted faithfully through Eq.2.", + "bbox": [ + 212, + 678, + 787, + 710 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 High Fidelity Talking Head Generation", + "text_level": 1, + "bbox": [ + 214, + 728, + 586, + 744 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Volume Rendering. To generate an image with volume rendering, given the feature vector $\\mathbf{F}_{xyz}^{i}(\\tilde{\\mathbf{x}})$ retrieved from the tri-plane, a tri-plane decoder takes it as input and predicts the color density $\\mathbf{c}$ and occupancy $\\sigma$ at that position. The final color of a pixel $\\mathbf{C}(\\mathbf{r})$ along the ray is calculated through:", + "bbox": [ + 212, + 750, + 787, + 806 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {C} (\\mathbf {r}) = \\int_ {t _ {n}} ^ {t _ {f}} T (t) \\sigma (\\mathbf {r} (t), \\mathbf {a}) \\mathbf {c} (\\mathbf {r} (t), \\mathbf {a}) d t, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 811, + 785, + 844 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "D. Li et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/bf3c4215eb4643f96b7d4b120951b25a5be6713aa1a221d3035b7ce957f261ae.jpg", + "table_caption": [ + "Table 1: Quantitative comparison with single-shot methods, The best and the second best results are emphasized." + ], + "table_footnote": [], + "table_body": "
DatasetMetricGTWav2LipMakeItTalkPC-AVSStyleTalkSadTalker\\( S^3 \\)D-NeRF
HDTFSSIM ↑10.7490.5590.5820.7760.6780.819
LPIPS ↓00.3320.4220.4110.2890.3860.258
F-LMD ↓03.5826.8463.9253.6234.6212.799
M-LMD ↓03.6526.1523.9123.2733.6542.929
CPBD ↑0.2840.1950.1390.2060.2340.1920.263
Sync ↑8.1448.3654.8135.0966.1886.4736.514
MEADSSIM ↑10.8290.6260.6730.8500.7550.845
LPIPS ↓00.2120.3670.3590.1990.2230.161
F-LMD ↓02.1266.0818.3722.6223.8352.014
M-LMD ↓03.3776.7905.1923.0613.6312.921
CPBD ↑0.1760.1010.0650.0680.1450.0810.152
Sync ↑7.7027.7135.1435.4666.0575.7076.768
", + "bbox": [ + 225, + 184, + 782, + 377 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $T(t) = \\exp \\left(-\\int_{t_n}^t\\sigma (s)ds\\right)$ denotes for the accumulated transmittance along the ray from $t_n$ to $t$ , $t_n$ and $t_f$ are the lower and the upper bound of depth.", + "bbox": [ + 215, + 410, + 784, + 460 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Coarse-to-fine Generation Strategy. Talking head frames contain various backgrounds and complex outer face textures which are hard for a NeRF to render at the same time. To decline the difficulty for training, we adopt a coarse-to-fine strategy. Specifically, the inner face part $\\mathbf{I}^{masked}$ of an arbitrary frame $\\mathbf{I}$ is decoupled in advance. During training, the inner face is rendered by the NeRF model, and the synthesized image is denoted as $\\mathbf{I}_{gen}^{coar}$ , while the fine texture details of the outer face and the background is complemented by an additional super-resolution module. We utilize Deep3DFaceRecon [7] to generate a 3D-aware face segmentation mask for inner-outer face area separation. This coarse-to-fine generation strategy guarantees that the NeRF can focus on the important face area.", + "bbox": [ + 215, + 463, + 785, + 628 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Super Resolution Module. The super resolution module takes the coarse output face $\\mathbf{I}_{gen}^{coar}$ and the full source face image $\\mathbf{I}_{src}^{full}$ as input, and generates the talking head frame $\\mathbf{I}_{fine}^{gen}$ with background and more details. The whole super resolution process can be written as $\\mathbf{I}_{gen}^{fine} = \\text{Supres}(\\mathbf{I}_{gen}^{coar}, \\mathbf{I}_{src}^{full})$ . More details are shown in the appendix.", + "bbox": [ + 215, + 631, + 785, + 709 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.4 Lip-sync Discriminator", + "text_level": 1, + "bbox": [ + 217, + 742, + 449, + 758 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Our lip sync discriminator has a visual branch and an audio branch. The former takes $T$ frames of the lip area as input and outputs their embedding $\\mathbf{e}_l$ , while the latter gets an audio clip and outputs its corresponding feature embedding $\\mathbf{e}_a$ . Then, the synchronization level between the audio and the video sequence", + "bbox": [ + 215, + 779, + 785, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "$\\mathrm{S}^3\\mathrm{D}$ -NeRF", + "bbox": [ + 656, + 113, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "can be judged by calculating the cosine similarity between $\\mathbf{e}_l$ and $\\mathbf{e}_a$ :", + "bbox": [ + 212, + 146, + 712, + 162 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\cos (\\mathbf {e} _ {l}, \\mathbf {e} _ {a}) = \\frac {\\mathbf {e} _ {l} \\cdot \\mathbf {e} _ {a}}{\\operatorname* {m a x} \\left(\\| \\mathbf {e} _ {l} \\| _ {2} \\cdot \\| \\mathbf {e} _ {a} \\| _ {2} , \\epsilon\\right)}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 374, + 172, + 785, + 202 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\epsilon$ is a margin which is set to 0. Different from that used in Wav2Lip [27], our lip sync discriminator is trained on a larger mixed dataset containing HDTF [44] and LRS2 [1], with a single contrastive triplet loss for more distinctive feature representations:", + "bbox": [ + 212, + 210, + 782, + 272 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {s y n c}} ^ {\\text {d i s}} = \\max \\left(0, \\eta + \\cos \\left(\\mathbf {e} _ {l}, \\mathbf {e} _ {a} ^ {+}\\right) - \\cos \\left(\\mathbf {e} _ {l}, \\mathbf {e} _ {a} ^ {-}\\right)\\right) \\tag {7} \\\\ + \\max \\left(0, \\eta + \\cos \\left(\\mathbf {e} _ {l} ^ {+}, \\mathbf {e} _ {a}\\right) - \\cos \\left(\\mathbf {e} _ {l} ^ {-}, \\mathbf {e} _ {a}\\right)\\right). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 282, + 785, + 321 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "$\\mathbf{e}_a^+$ and $\\mathbf{e}_a^-$ regard the positive and the negative audio embedding according to the lip sequence respectively, so as $\\mathbf{e}_l^+$ and $\\mathbf{e}_l^-$ , $\\eta$ is a margin set to 0.5. While training our model, the lip-sync discriminator is frozen and $T$ frames of the same person are generated once a time. The lip regions are cropped through the pre-extracted bounding boxes of the ground-truth. The loss used for S $^3$ D-NeRF training can be formulated as", + "bbox": [ + 212, + 332, + 784, + 422 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s y n c}} ^ {\\text {g e n}} = \\cos \\left(\\mathbf {e} _ {l} ^ {\\text {g e n}}, \\mathbf {e} _ {g t}\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 421, + 435, + 784, + 453 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "while $\\mathbf{e}_l^{gen}$ is the embedding of the generated lip sequence and $\\mathbf{e}_{gt}$ is the embedding of the ground-truth audio. By penalizing the out-of-sync audio-visual pair, more precise mouth shapes can be achieved.", + "bbox": [ + 212, + 462, + 782, + 507 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "3.5 Loss Functions", + "text_level": 1, + "bbox": [ + 214, + 530, + 383, + 542 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In addition to the lip synchronization loss $\\mathcal{L}_{sync}^{gen}$ , we use a pixelwise $l_{2}$ reconstruction loss $\\mathcal{L}_{pix}$ , a feature level perceptual loss $\\mathcal{L}_{per}$ to match the generated face with the target face, and a GAN loss $\\mathcal{L}_{adv}$ with the architecture of StyleGAN discriminator [15]. All the losses are imposed on both the coarse and the fine faces. An $l_{2}$ loss $\\mathcal{L}_{deform}$ is used to regularize the prediction deformation. Our S $^3$ D-NeRF is finally trained with the weighted sum of the above losses as:", + "bbox": [ + 212, + 554, + 782, + 645 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {p i x} + \\lambda_ {p e r} \\mathcal {L} _ {p e r} + \\lambda_ {a d v} \\mathcal {L} _ {a d v} + \\lambda_ {\\text {s y n c}} ^ {\\text {g e n}} \\mathcal {L} _ {\\text {s y n c}} ^ {\\text {g e n}} + \\lambda_ {\\text {d e f o r m}} \\mathcal {L} _ {\\text {d e f o r m}}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 261, + 657, + 784, + 674 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "with $\\lambda_{per} = 0.01$ , $\\lambda_{adv} = 1$ , $\\lambda_{sync}^{gen} = 0.5$ , $\\lambda_{deform} = 0.001$ .", + "bbox": [ + 214, + 685, + 625, + 702 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 723, + 375, + 739 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Experimental Settings", + "text_level": 1, + "bbox": [ + 214, + 753, + 446, + 768 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Datasets. Our $\\mathrm{S}^3\\mathrm{D}$ -NeRF is trained on HDTF [44] training split, where each video ranges from tens of seconds to more than ten minutes in length, and evaluated on HDTF testing split and MEAD testing split [35] for comparison with single-shot methods. There is no overlap between the training and the testing", + "bbox": [ + 212, + 779, + 784, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "D. Li et al.", + "bbox": [ + 271, + 114, + 346, + 127 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/91fc46d10e0bf4b0169710d39b311fff04eabbbc5bec46bbd8644236ad3d00bd.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 225, + 146, + 496, + 316 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/f3707c90a25590976e50d4b9179ccf5d16531148b246a43b0f309357451eb7dd.jpg", + "image_caption": [ + "(b)", + "Fig. 5: Left (a): Qualitative comparison with NeRF-based methods, when encountering a new identity, our S $^3$ D-NeRF successfully synthesizes a portrait faithfully without any retraining. Right (b): High fidelity generation results with multi-view consistency. Ground truth are the images with front view. Most of the face details are preserved." + ], + "image_footnote": [], + "bbox": [ + 504, + 159, + 779, + 311 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "split. For comparison with NeRF-based methods, we use the videos provided by AD-NeRF [11] and GeneFace [39]. The video sampling rate is set to 25 FPS. All the faces are cropped and aligned to a fix resolution of $256 \\times 256$ .", + "bbox": [ + 212, + 410, + 784, + 455 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Implementation Details. The whole proposed framework is implemented by PyTorch and can be trained end-to-end. It takes about two days to train our S $^3$ D-NeRF on 4 NVIDIA-A100 GPUs. The sample rate of the speech signal is set to 16000. We use Wav2Vec2.0 [2] to extract the driving audio features. More details and the user study can be found in the appendix.", + "bbox": [ + 212, + 455, + 787, + 532 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Quantitative Results", + "text_level": 1, + "bbox": [ + 215, + 556, + 433, + 571 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Several widely used metrics are adopted to carry out our quantitative evaluation. We use SSIM and LPIPS [42] to measure the similarity of the generated speaker with the ground truth in pixel level and feature level respectively. To compare the identity fidelity, we use Facial Landmark Distance (F-LMD). To compare the clear extent of the images, we use Cumulative Probability of Blur Detection (CPBD). To evaluate the audio-lip synchronization, we use the SyncNet confidence score (Sync) [6] and Mouth Landmark Distance (M-LMD) [5].", + "bbox": [ + 212, + 582, + 787, + 688 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Comparison with Single-Shot Methods. Our competitors include Wav2Lip [27], MakeiTalk [46], PC-AVS [45], SadTalker [43] and StyleTalk [22]. For the above methods, we take their provided checkpoints and strictly follow the test protocol of current single-shot methods to conduct our evaluation. MakeiTalk has shown the worst image quality and lip shape. Wav2Lip also uses a pretrained lip-sync expert as supervision, and it has larger batch size and its model is simpler. Thus, it achieves a SyncNet score which is even better than ground truth. However, Wav2Lip tends to generate blurry faces, resulting in an inferior face fidelity. StyleTalk and SadTalker have shown competitive performance on image quality. Our S $^3$ D-NeRF achieves the second-best lip-sync result, meanwhile sur", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "$\\mathrm{S}^3\\mathrm{D}$ -NeRF", + "bbox": [ + 656, + 113, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/b6c1f7784bc59d48505edcd76916f74256c1798ab43e446cd7d8ee1b1e04373f.jpg", + "table_caption": [ + "Table 2: Quantitative comparison with NeRF-based methods. The best and the second best results are highlighted. Our S $^3$ D-NeRF holds competitive image quality and audio-visual consistency, meanwhile the best generalize ability." + ], + "table_footnote": [], + "table_body": "
MetricGTGeneFaceAD-NeRFER-NeRFS3D-NeRF
SSIM ↑10.8280.8460.8840.852
LPIPS ↓00.0980.0870.0680.104
M-LMD ↓02.4361.9821.6591.493
F-LMD ↓01.8981.7151.1582.019
CPBD ↑0.2780.2070.2200.2390.244
Sync ↑8.9705.3665.6266.7817.118
FPS ↑N/A0.130.13299.5
Fit Time / h ↓N/A42364.5<0.01
", + "bbox": [ + 302, + 198, + 702, + 333 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "passes the competitors significantly on quality metrics at both image level and feature level. Moreover, our method has the lowest face landmark distance and mouth landmark distance, which indicates that our S $^3$ D-NeRF keep facial and lip structure better than other methods. On MEAD dataset where all the identities are unseen before, our S $^3$ D-NeRF still achieves competitive performance, which demonstrates the strong generalization ability of our approach.", + "bbox": [ + 212, + 358, + 785, + 449 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Comparison with NeRF-based Methods. We also compare S $^3$ D-NeRF with current identity specific NeRF-based methods. Three start-of-the-art approaches AD-NeRF [11] GeneFace [39] and ER-NeRF [17] are chosen as our opponents. It is worth to note that our method has never seen the identity nor the driving audio in the evaluation set, while the person-specific NeRFs are trained on a training set which has the same identity as the evaluation set (although no image overlap). In addition to the metrics mentioned above, we also take account of the data efficiency metrics, FPS: Frame Per Second during inference, which can reflects the speed of a rendering model. Fit Time: Times needed to generate the plausible talking head videos of a new identity with a pretrained model available. ER-NeRF has shown the strong image quality with the fastest rendering speed. GeneFace needs to train a postnet for each new identity, which takes extra time. Our model is also competitive in image quality. Moreover, it has the best lip synchronization score, together with the best mouth-LMD, which indicates the superiority on the specific audio-driven task. Moreover, our model has the highest CPBD, that means we generate the clearest images. Besides, S $^3$ D-NeRF has the strongest generalization ability since it is free from retraining when coming up with an new speaker, and has shown an excellent inference speed.", + "bbox": [ + 212, + 449, + 787, + 736 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Qualitative Results", + "text_level": 1, + "bbox": [ + 215, + 756, + 419, + 771 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Comparison with Single-Shot Methods. A direct comparison with other person agnostic methods are shown in Fig.4. Our S $^3$ D-NeRF has shown more precise lip shapes and clearer teeth. At the same time, the detailed facial features such as skin texture, eyebrows and teeth splits are preserved by our method.", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "D. Li et al.", + "bbox": [ + 271, + 114, + 346, + 127 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/da0083a72e3022b0b2821fb7aa3e33c6d085ea7c5e4918e34691988afc041faa.jpg", + "image_caption": [ + "Fig. 6: Qualitative ablation study over key components." + ], + "image_footnote": [], + "bbox": [ + 308, + 143, + 699, + 347 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/532fe2e3f973f971d9203da1444a0b5604bde35d5c7f87d0a3a4946d518e0dea.jpg", + "table_caption": [ + "Table 3: Quantitative ablation study. The full S $^3$ D-NeRF achieves the best landmark distance and lip synthesis results, with a little sacrifice on image quality." + ], + "table_footnote": [], + "table_body": "
MethodSSIM ↑LPIPS ↓F-LMD ↓M-LMD ↓CPBD ↑Sync ↑
ND0.7440.2783.4753.5070.1655.201
FCD0.8340.2422.8553.1960.2466.106
w/o Deform0.5560.457N/AN/A0.1400.240
w/o Lip Sync0.8010.2633.2763.3420.2555.649
S3D-NeRF0.8290.2582.7992.9290.2636.514
", + "bbox": [ + 289, + 446, + 718, + 541 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Comparison with NeRF-based Methods. A qualitative comparison with NeRF-based methods are shown in Fig.5. AD-NeRF has shown the head-torso separation problem, which seriously harm the fidelity of the portrait (Obama face at the third row). Due to the identity specific training process, NeRF based methods can achieve excellent speaker reconstruction. Their generated face regions are very similar to the ground truth. However, on a novel identity, all the person specific NeRFs failed to generate a plausible result. On the contrary, our S $^3$ D-NeRF has shown excellent generation results. The images from our S $^3$ D-NeRF do not reach a similarity as high as NeRF-based methods, but they have more distinct facial details, each face region of the speaker are kept completely. Images generated by our method are sharper and their lip areas are much clearer.", + "bbox": [ + 212, + 588, + 787, + 770 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Multi-view Generation Results. Benefited from the 3D consistency, our method can generate faces with various view angles with the same lip shape, which is shown in the right part of the Fig.5. The facial details such as wrinkles and ear-rings are well kept, indicating the generative power of our method.", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "$\\mathrm{S}^3\\mathrm{D}$ -NeRF", + "bbox": [ + 656, + 113, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Ablation Studies & Analysis", + "text_level": 1, + "bbox": [ + 215, + 146, + 496, + 161 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The Hierarchical Facial Appearance Encoder is a necessary structure which can not be discarded or replaced. To study the function of our Cross-modal Facial Deformation Field, we designed several kinds of variants. Firstly, we directly reconstruct the frame of the target speaker without deformation, denoted as w/o deform. In this case, the audio feature is concatenated with the tri-plane feature and fed into the NeRF model Secondly, we replace our Cross-modal Facial Deformation Field with a naive deformation module mentioned in Sec.3.2, denoted as Naive Deform (ND). Thirdly, to show the effect of the correlation score as a prior, we train a model with a deformation module which takes directly the tri-plane feature as an additional input, denoted as Feat-Concatenate Deform (FCD). A model without the lip-sync discriminator is denoted as w/o lip-sync. Qualitative and quantitative results are shown in Tab.3 and Fig.6.", + "bbox": [ + 212, + 171, + 785, + 352 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The model without deformation modules totally fails to keep the appearance of the speaker, resulting in a poor video quality, not even performing speech animations. The Naive Deform model yields a coarse image similar to the average face, and its final generated talking head frames are less sharper. As for the Feat-Concatenate Deform model, it has kept the source feature better, resulting in a slight higher image quality metrics, however, because of it does not take the relationship between the speech signal and different face regions, some details of the mouth and the face area are missing. The model trained without the lip-sync discriminator synthesizes less precise lip shape. Our full model has the best performance on most of the metrics.", + "bbox": [ + 212, + 353, + 785, + 503 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Limitations & Ethical Considerations", + "text_level": 1, + "bbox": [ + 215, + 526, + 612, + 542 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations. Changing backgrounds like person specific-NeRFs [11, 17] is not supported currently. When the head pose is overlarge, the contour of the portrait sometimes tends to be blur. The above issues will be studied in future.", + "bbox": [ + 212, + 558, + 784, + 602 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Ethical Considerations. Once trained, our $\\mathrm{S}^3\\mathrm{D}$ -NeRF can synthesize videos of any given person saying something that they never actually said. This may cause some moral or legal problems. We are committed to offer our model to fight against the potential abuses.", + "bbox": [ + 212, + 603, + 785, + 664 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 686, + 359, + 703 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we propose a Single-Shot Speech-Driven Neural Radiance Field (S $^3$ D-NeRF) for synthesizing audio driving talking heads with high fidelity. Several key components are proposed to assist the generation process, including a Hierarchical Facial Appearance Encoder, a Cross-modal Facial Deformation Field and a lip-sync discriminator. Our S $^3$ D-NeRF is the first attempt to explore single-shot speech-driven talking head generation with NeRFs. Validated by comprehensive experiments, S $^3$ D-NeRF has shown clear improvement against current state-of-the-art, both on video quality and generalization ability.", + "bbox": [ + 212, + 719, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "D. Li et al.", + "bbox": [ + 271, + 114, + 346, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work is supported by the National Natural Science Foundation of China (NSFC) under Grants 62372452, 62272460, Youth Innovation Promotion Association CAS, and Alibaba Research Intern Program.", + "bbox": [ + 215, + 175, + 784, + 220 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 243, + 321, + 258 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Afouras, T., Chung, J.S., Senior, A., Vinyals, O., Zisserman, A.: Deep audio-visual speech recognition. IEEE transactions on pattern analysis and machine intelligence 44(12), 8717-8727 (2018)", + "2. Baevski, A., Zhou, Y., Mohamed, A., Auli, M.: wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems 33, 12449-12460 (2020)", + "3. Blanz, V., Vetter, T.: A morphable model for the synthesis of 3d faces. In: Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pp. 157-164 (2023)", + "4. Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., De Mello, S., Gallo, O., Guibas, L.J., Tremblay, J., Khamis, S., et al.: Efficient geometry-aware 3d generative adversarial networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16123-16133 (2022)", + "5. Chen, L., Maddox, R.K., Duan, Z., Xu, C.: Hierarchical cross-modal talking face generation with dynamic pixel-wise loss. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7832-7841 (2019)", + "6. Chung, J.S., Zisserman, A.: Out of time: automated lip sync in the wild. In: Computer Vision-ACCV 2016 Workshops: ACCV 2016 International Workshops, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part II 13. pp. 251-263. Springer (2017)", + "7. Deng, Y., Yang, J., Xu, S., Chen, D., Jia, Y., Tong, X.: Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In: IEEE Computer Vision and Pattern Recognition Workshops (2019)", + "8. Gafni, G., Thies, J., Zollhöfer, M., Nießner, M.: Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8649-8658 (June 2021)", + "9. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. Communications of the ACM 63(11), 139–144 (2020)", + "10. Guo, J., Zhu, X., Yang, Y., Yang, F., Lei, Z., Li, S.Z.: Towards fast, accurate and stable 3d dense face alignment. In: Proceedings of the European Conference on Computer Vision (ECCV) (2020)", + "11. Guo, Y., Chen, K., Liang, S., Liu, Y.J., Bao, H., Zhang, J.: Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5784-5794 (2021)", + "12. Hong, F.T., Zhang, L., Shen, L., Xu, D.: Depth-aware generative adversarial network for talking head video generation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3397-3406 (2022)", + "13. Hong, Y., Peng, B., Xiao, H., Liu, L., Zhang, J.: Headnerf: A real-time nerf-based parametric head model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20374-20384 (2022)" + ], + "bbox": [ + 218, + 273, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "$\\mathrm{S}^3\\mathrm{D}$ -NeRF", + "bbox": [ + 656, + 113, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Jang, W., Agapito, L.: Codenerf: Disentangled neural radiance fields for object categories. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12949-12958 (2021)", + "15. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of stylegan. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8110-8119 (2020)", + "16. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013)", + "17. Li, J., Zhang, J., Bai, X., Zhou, J., Gu, L.: Efficient region-aware neural radiance fields for high-fidelity talking portrait synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 7568-7578 (October 2023)", + "18. Li, W., Zhang, L., Wang, D., Zhao, B., Wang, Z., Chen, M., Zhang, B., Wang, Z., Bo, L., Li, X.: One-shot high-fidelity talking-head synthesis with deformable neural radiance field (2023)", + "19. Lin, T.Y., Dollar, P., Girshick, R., He, K., Hariharan, B., Belongie, S.: Feature pyramid networks for object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2117-2125 (2017)", + "20. Liu, X., Xu, Y., Wu, Q., Zhou, H., Wu, W., Zhou, B.: Semantic-aware implicit neural audio-driven video portrait generation. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVII. pp. 106-125. Springer (2022)", + "21. Locatello, F., Weissenborn, D., Unterthiner, T., Mahendran, A., Heigold, G., Uszkoreit, J., Dosovitskiy, A., Kipf, T.: Object-centric learning with slot attention. Advances in Neural Information Processing Systems 33, 11525-11538 (2020)", + "22. Ma, Y., Wang, S., Hu, Z., Fan, C., Lv, T., Ding, Y., Deng, Z., Yu, X.: Styletalk: One-shot talking head generation with controllable speaking styles. In: Proceedings of the Thirty-Seventh AAAI Conference on Artificial Intelligence (AAAI) (2023)", + "23. Ma, Z., Zhu, X., Qi, G.J., Lei, Z., Zhang, L.: Otavatar: One-shot talking face avatar with controllable tri-plane rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16901-16910 (2023)", + "24. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)", + "25. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022)", + "26. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., MartinBrualla, R.: Nerfies: Deformable neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5865-5874 (2021)", + "27. Prajwal, K., Mukhopadhyay, R., Namboodiri, V.P., Jawahar, C.: A lip sync expert is all you need for speech to lip generation in the wild. In: Proceedings of the 28th ACM International Conference on Multimedia. pp. 484-492 (2020)", + "28. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-nerf: Neural radiance fields for dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10318-10327 (2021)", + "29. Ren, Y., Li, G., Chen, Y., Li, T.H., Liu, S.: Pirenderer: Controllable portrait image generation via semantic neural rendering (2021)", + "30. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "D. Li et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015)", + "31. Shen, S., Li, W., Zhu, Z., Duan, Y., Zhou, J., Lu, J.: Learning dynamic facial radiance fields for few-shot talking head synthesis. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XII. pp. 666-682. Springer (2022)", + "32. Tang, J., Wang, K., Zhou, H., Chen, X., He, D., Hu, T., Liu, J., Zeng, G., Wang, J.: Real-time neural radiance talking portrait synthesis via audio-spatial decomposition. arXiv preprint arXiv:2211.12368 (2022)", + "33. Thies, J., Elgharib, M., Tewari, A., Theobalt, C., Nießner, M.: Neural voice puppetry: Audio-driven facial reenactment. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVI 16. pp. 716-731. Springer (2020)", + "34. Wang, J., Zhao, K., Zhang, S., Zhang, Y., Shen, Y., Zhao, D., Zhou, J.: Lipformer: High-fidelity and generalizable talking face generation with a pre-learned facial codebook. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13844–13853 (2023)", + "35. Wang, K., Wu, Q., Song, L., Yang, Z., Wu, W., Qian, C., He, R., Qiao, Y., Loy, C.C.: Mead: A large-scale audio-visual dataset for emotional talking-face generation. In: ECCV (2020)", + "36. Wang, S., Li, L., Ding, Y., Yu, X.: One-shot talking face generation from single-speaker audio-visual correlation learning. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 2531-2539 (2022)", + "37. Wiles, O., Koepke, A., Zisserman, A.: X2face: A network for controlling face generation using images, audio, and pose codes. In: Proceedings of the European conference on computer vision (ECCV). pp. 670-686 (2018)", + "38. Ye, Z., Jiang, Z., Ren, Y., Liu, J., He, J., Zhao, Z.: Geneface: Generalized and high-fidelity audio-driven 3d talking face synthesis. arXiv preprint arXiv:2301.13430 (2023)", + "39. Ye, Z., Jiang, Z., Ren, Y., Liu, J., He, J., Zhao, Z.: Geneface: Generalized and high-fidelity audio-driven 3d talking face synthesis. arXiv preprint arXiv:2301.13430 (2023)", + "40. Yi, R., Ye, Z., Zhang, J., Bao, H., Liu, Y.J.: Audio-driven talking face video generation with learning-based personalized head pose. arXiv preprint arXiv:2002.10137 (2020)", + "41. Yu, W., Fan, Y., Zhang, Y., Wang, X., Yin, F., Bai, Y., Cao, Y.P., Shan, Y., Wu, Y., Sun, Z., et al.: Nofa: Nerf-based one-shot facial avatar reconstruction. In: ACM SIGGRAPH 2023 Conference Proceedings. pp. 1-12 (2023)", + "42. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. Proceedings of the IEEE conference on computer vision and pattern recognition pp. 586-595 (2018)", + "43. Zhang, W., Cun, X., Wang, X., Zhang, Y., Shen, X., Guo, Y., Shan, Y., Wang, F.: Sadtalker: Learning realistic 3d motion coefficients for stylized audio-driven single image talking face animation. arXiv preprint arXiv:2211.12194 (2022)", + "44. Zhang, Z., Li, L., Ding, Y., Fan, C.: Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3661-3670 (2021)", + "45. Zhou, H., Sun, Y., Wu, W., Loy, C.C., Wang, X., Liu, Z.: Pose-controllable talking face generation by implicitly modularized audio-visual representation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "$\\mathrm{S}^3\\mathrm{D}$ -NeRF", + "bbox": [ + 656, + 113, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "ref_text", + "text": "46. Zhou, Y., Han, X., Shechtman, E., Echevarria, J., Kalogerakis, E., Li, D.: Makelttalk: speaker-aware talking-head animation. ACM Transactions On Graphics (TOG) 39(6), 1-15 (2020)", + "bbox": [ + 217, + 146, + 785, + 189 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "D. Li et al.", + "bbox": [ + 271, + 114, + 346, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/S^3D-NeRF_ Single-Shot Speech-Driven Neural Radiance Field for High Fidelity Talking Head Synthesis/49757949-cf34-4879-95e0-e53ca1bcc2a6_model.json b/2024/S^3D-NeRF_ Single-Shot Speech-Driven Neural Radiance Field for High Fidelity Talking Head Synthesis/49757949-cf34-4879-95e0-e53ca1bcc2a6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0a6cdb791c46f9dedf06b99b49b069ab6eb6a845 --- /dev/null +++ b/2024/S^3D-NeRF_ Single-Shot Speech-Driven Neural Radiance Field for High Fidelity Talking Head Synthesis/49757949-cf34-4879-95e0-e53ca1bcc2a6_model.json @@ -0,0 +1,2249 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.237, + 0.142, + 0.768, + 0.211 + ], + "angle": 0, + "content": "\\(\\mathrm{S}^{3}\\mathrm{D}-\\mathrm{NeRF}\\): Single-Shot Speech-Driven Neural Radiance Field for High Fidelity Talking Head Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.31, + 0.237, + 0.693, + 0.27 + ], + "angle": 0, + "content": "Dongze Li\\(^{1,2}\\) Kang Zhao\\(^{3}\\), Wei Wang\\(^{2\\star}\\), Yifeng Ma\\(^{3}\\), Bo Peng\\(^{2}\\), Yingya Zhang\\(^{3}\\) and Jing Dong\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.279, + 0.756, + 0.293 + ], + "angle": 0, + "content": "\\(^{1}\\) School of Artificial Intelligence, University of Chinese Academy of Sciences" + }, + { + "type": "text", + "bbox": [ + 0.258, + 0.293, + 0.746, + 0.307 + ], + "angle": 0, + "content": "\\(^{2}\\) NLPR & MAIS, Institute of Automation, Chinese Academy of Sciences" + }, + { + "type": "text", + "bbox": [ + 0.445, + 0.307, + 0.56, + 0.322 + ], + "angle": 0, + "content": "3 Alibaba Group" + }, + { + "type": "text", + "bbox": [ + 0.264, + 0.323, + 0.739, + 0.337 + ], + "angle": 0, + "content": "dongze.li@cripac.ia.ac.cn; {wang,bo.peng,jdong}@nlpr.ia.ac.cn" + }, + { + "type": "text", + "bbox": [ + 0.289, + 0.337, + 0.715, + 0.35 + ], + "angle": 0, + "content": "{zhaokang.zk, yingya.zyy, mayifeng.myf}@alibaba-inc.com" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.389, + 0.741, + 0.666 + ], + "angle": 0, + "content": "Abstract. Talking head synthesis is a practical technique with wide applications. Current Neural Radiance Field (NeRF) based approaches have shown their superiority on driving one-shot talking heads with videos or signals regressed from audio. However, most of them failed to take the audio as driven information directly, unable to enjoy the flexibility and availability of speech. Since mapping audio signals to face deformation is non-trivial, we design a Single-Shot Speech-Driven Neural Radiance Field (S\\(^3\\)D-NeRF) method in this paper to tackle the following three difficulties: learning a representative appearance feature for each identity, modeling motion of different face regions with audio, and keeping the temporal consistency of the lip area. To this end, we introduce a Hierarchical Facial Appearance Encoder to learn multi-scale representations for catching the appearance of different speakers, and elaborate a Cross-modal Facial Deformation Field to perform speech animation according to the relationship between the audio signal and different face regions. Moreover, to enhance the temporal consistency of the important lip area, we introduce a lip-sync discriminator to penalize the out-of-sync audio-visual sequences. Extensive experiments have shown that our S\\(^3\\)D-NeRF surpasses previous arts on both video fidelity and audio-lip synchronization." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.68, + 0.612, + 0.695 + ], + "angle": 0, + "content": "Keywords: Talking Head \\(\\cdot\\) Neural Radiance Fields" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.722, + 0.377, + 0.737 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.755, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Speech driven talking head synthesis is a promising technique and can be applied to a wide range of situations such as digital human, film making, virtual reality and video games. Current Neural Radiance Field (NeRF) [24] based methods [11,17,38] have shown their superiority on generating vivid talking portraits" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.383, + 0.84 + ], + "angle": 0, + "content": "* Corresponding author" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.348, + 0.128 + ], + "angle": 0, + "content": "D. Li et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.283 + ], + "angle": 0, + "content": "with high quality for their 3D consistency and view controllability. However, they can only be applied to a specific identity, and require a long video sequence from the same speaker to train on, which hampers them from broader application scenarios. Afterwards some progress have been made in the generalization of NeRF-based talking head methods. Most of them [18, 23, 41] are driven by intermediate facial representations, such as 3DMM coefficients extracted from video or expression coefficients regressed from audio. These intermediate representations may introduce information loss more or less, leading to sub-optimal synthesis results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.288, + 0.789, + 0.546 + ], + "angle": 0, + "content": "Considering it is non-trivial to map audio signals to face deformation, we attribute the challenge of performing high-fidelity one-shot speech animation with Neural Radiance Fields to three aspects: 1) Learning Representative Appearance Features for Each Identity. Different people have different facial appearance i.e. shape and texture, it is difficult for a vanilla NeRF to model these details of several speakers simultaneously. During the inference phase, it becomes even harder to generate a talking head sequence with high quality given that only a single image is available. 2) Modeling Motion of Different Face Regions with Audio. Video driven techniques can utilize global driving signals which can describe the motion of the whole face, such as depth priors [12], 3DMM meshes [41] or PNCC [10,18] maps. However, in a speech driven task, the input audio merely has strong correlation with the lower face [27,34]. Directly predicting face motions according to the input audio will cause insufficient modeling of different face regions, harming the fidelity of the synthesized portrait. 3) Keeping Temporal Consistency of the Lip Area. The lip area has the greatest importance in speech animation. Due to the lack of constraints on this part, videos generated by NeRFs have sometimes shown incorrect mouth shapes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.549, + 0.79, + 0.61 + ], + "angle": 0, + "content": "In this paper, we propose S\\(^3\\)D-NeRF, namely Single-Shot Speech-Driven Neural Radiance Field, to synthesize high fidelity talking head videos. Given a single shot source image and a driven audio sequence, our method can synthesize vivid free view talking head videos with accurate mouth shapes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.79, + 0.842 + ], + "angle": 0, + "content": "To fully capture the appearance of an arbitrary speaker, a Hierarchical Facial Appearance Encoder is presented to extract the multi-scale facial features of the single-shot source image with a down sample convolutional network armed with a feature pyramid. These features contain rich structural and textural information of the source identity. Then, a multi-scale tri-plane [4] representation is constructed for neural rendering. To realize accurate speech animation, a Cross-modal Facial Deformation Field is proposed to faithfully catch the correlations between the speech signal and the visual features from different face regions through cross attention, and predicts the motion of the whole face precisely. To further enhance the consistency of the lip area, a lip-sync discriminator is imposed on the lower face of the generated video frames and the driving audio signal. By penalizing the out-of-sync audio-visual sequences, higher temporal consistency of the lip area can be achieved. During training, we adapt a coarse-to-fine image generation strategy to reduce the difficulty of modeling multiple speakers at the same time. Specifically, a coarse talking head frame which" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.657, + 0.114, + 0.732, + 0.127 + ], + "angle": 0, + "content": "\\(\\mathrm{S}^3\\mathrm{D}\\) -NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.253, + 0.149, + 0.756, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.338, + 0.785, + 0.366 + ], + "angle": 0, + "content": "Fig. 1: A showcase of our S\\(^3\\)D-NeRF, which generates high quality face portraits with fine-grained face texture and mouth details." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.397, + 0.784, + 0.442 + ], + "angle": 0, + "content": "contains the important inner face area is synthesized through standard volume neural rendering, while the texture details are refined with a super-resolution module, resulting in a high fidelity talking head frame." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.443, + 0.627, + 0.457 + ], + "angle": 0, + "content": "Our main contributions can be summarized as below:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.47, + 0.785, + 0.514 + ], + "angle": 0, + "content": "- We propose S\\(^3\\)D-NeRF, a single-shot NeRF-based talking head synthesis framework. Our method extends NeRF-based speech animation techniques to handle arbitrary unseen identities." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.516, + 0.785, + 0.59 + ], + "angle": 0, + "content": "- Several key components are presented to assist the talking head synthesis procedure. Including a Hierarchical Facial Appearance Encoder for representative feature extraction, a Cross-modal Facial Deformation Field for accurate speech animation, and a lip-sync discriminator for better temporal consistency of lip area." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.592, + 0.785, + 0.621 + ], + "angle": 0, + "content": "- Comprehensive experiments have shown the superiority of our S\\(^3\\)D-NeRF over previous arts on both video fidelity and audio-lip synchronization." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.47, + 0.785, + 0.621 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.646, + 0.388, + 0.661 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.678, + 0.58, + 0.695 + ], + "angle": 0, + "content": "2.1 Speech Driven Talking Head Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.784, + 0.764 + ], + "angle": 0, + "content": "The goal of speech driven talking head synthesis is to animate a speaker according to input audiios. According to the generalization ability of the model to new identities, current speech driven methods can be divided into two categories: identity agnostic methods and identity specific methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Identity agnostic methods are capable of generating speech videos of an arbitrary speaker once their training procedure is finished, with one or several images or video sequences as inputs. Among them, single-shot methods [22, 27, 43, 45] have attracted considerable interests due to their high generalization ability and data efficiency. Image-based methods [27, 37, 45] utilize deep generative" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.127 + ], + "angle": 0, + "content": "D. Li et al." + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.146, + 0.787, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.349, + 0.788, + 0.433 + ], + "angle": 0, + "content": "Fig. 2: The full pipeline of our S\\(^3\\)D-NeRF. The Hierarchical Facial Appearance Encoder extracts representative features from the masked face region of the single-shot source image, for high fidelity neural rendering of an arbitrary identity. The Cross-modal Facial Deformation Field accurately models the motion of different face regions, with the help of the correlation score calculated through cross attention between audiovisual features. Texture details are complemented with the super-resolution module." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.788, + 0.613 + ], + "angle": 0, + "content": "networks [9] or auto-encoders [16] to generate continuous video clips. Model-based methods predict intermediate representations such as motion field [36], 2D landmarks [46] or 3D Morphable Model (3DMM) [3] coefficients [22, 29, 43] with the input audio signal, and animate the head with these representations. StyleTalk [22] predicts 3DMM through a transformer architecture and can generate portraits with diverse speaking styles under the help of a pretrained renderer [29]. SadTalker [43] uses several representation networks to learn lip and expression basis. Although remarkable progresses have been made, current single-shot methods still suffer in image quality and audio-lip synchronization because of the error accumulated in the audio to representation mapping." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Identity specific methods train an individual model for each identity with higher fidelity. Earlier methods [33, 40] utilize strong priors such as 3D meshes or dense key points for portrait generation. Current NeRF-based methods [11, 20, 31, 38] have shed light on high fidelity speaker synthesis. They optimize on the ground truth video clip of a single person, and can synthesize pose-controllable faces with fine-grained details. AD-NeRF [11] uses two separated NeRFs to model the head and the torso part respectively. SSP-NeRF [20] performs rays re-sampling based on the loss magnitude of different semantic regions. Currently, RAD-NeRF [32] and ER-NeRF [17] utilize hash table structures [25] to improve the rendering speed. GeneFace [39] proposes a facial radiance field with generalize ability but it also requires per identity finetuning and a complex multi-stage training process. Despite the above advantages, the above methods can only fit a single identity, and need tedious re-training when a new identity is encountered. At the same time, since no constraints are imposed on image sequences, the temporal consistency of the lip area is also not satisfactory." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.657, + 0.114, + 0.732, + 0.128 + ], + "angle": 0, + "content": "\\(\\mathrm{S}^3\\mathrm{D}\\) -NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.61, + 0.162 + ], + "angle": 0, + "content": "2.2 NeRF-based Face Portrait Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.171, + 0.788, + 0.398 + ], + "angle": 0, + "content": "It is a natural idea to model human faces with NeRFs in a 3D aware manner. However, vanilla NeRFs struggle to handle dynamic objects, and are unsuitable for animating human faces which have diverse poses and expressions. There mainly exist two series of solutions for this problem. The former [14] is to condition the original Radiance Field with additional global control signals, e.g. expression parameters. The latter solve this problem by learning an additional deformation field [26, 28] to warp the 3D points. NerFace [8] is the first approach to reconstruct animatable face avatars with NeRFs conditioned on 3DMM expression parameters. HeadNeRF [13], NOFA [41] and OTAvatar [23] utilize GAN inversion to project onto the latent space of a 3D aware GAN [4] and drive the generated avatar through different latent codes. HiDe-NeRF [18] utilizes Projected Normalized Coordinate Code (PNCC) [10] as a global expression control indicator, and drives the face image with high fidelity. Comparing with the above mentioned methods, driving an arbitrary head with speech signal directly through NeRFs is a more challenging problem and is rarely explored." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.419, + 0.331, + 0.435 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.789, + 0.706 + ], + "angle": 0, + "content": "The overall pipeline of our S\\(^3\\)D-NeRF is shown in Fig.2. The 3D positions of the talking head can be derived through the given head pose of the target image \\(\\mathbf{P}_{tar}\\), which is available during both training and inference time. Firstly, our Hierarchical Facial Appearance Encoder takes a single-shot image as input, and models its appearance with a deep convolutional network, where multi-resolution feature maps are extracted to construct an efficient representation i.e. multiscale appearance tri-planes, which are known to be capable of modeling several identities faithfully at the same time [4, 18]. Feature points from these planes are retrieved according to the 3D position for rendering. Then, the motion of the talking face are predicted by our Cross-modal Facial Deformation Field. It firstly calculates the correlation between an aggregated visual feature embedding and the audio signal through cross attention, then takes the result as a prior knowledge to assist the regression process of the deformation, which is used for calibrating the 3D position. A coarse face is rendered through the classic volume rendering process [24] with color and density predicted from the feature points on the tri-plane. Finally, a super-resolution module is used to refine the coarse face and yields the high fidelity portrait." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.727, + 0.593, + 0.743 + ], + "angle": 0, + "content": "3.1 Hierarchical Facial Appearance Encoder" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "An effective feature representation is important to accurately model the appearance of a speaker. Initially proposed in [4], tri-plane has shown its strong representative power and is much faster comparing with the vanilla NeRFs. The Hierarchical Facial Appearance Encoder constructs this representation based on multi-scale feature maps from the one-shot source image, and yields the feature points for further volumetric rendering process." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.348, + 0.128 + ], + "angle": 0, + "content": "D. Li et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.314 + ], + "angle": 0, + "content": "Constructing Representative Appearance Features. We adopt a feature pyramid structure [19] to extract source features for its ability to learn multi-scale information from different resolutions, and reshape all the feature maps channelwisely to construct several orthogonal planes as in [4]. To be more concrete, given the source image \\(\\mathbf{I}_{src}\\), feature maps with different resolutions \\(\\{\\mathbf{D}^0,\\mathbf{D}^1,\\mathbf{D}^2,\\mathbf{D}^3\\}\\) are extracted through several convolutional downsample blocks \\(Down^{i}(.)\\), while \\(\\mathbf{D}^0 = Down^0 (\\mathbf{I}_{src})\\) and \\(\\mathbf{D}^i = Down^i (\\mathbf{D}^{i - 1})\\). Then, hierarchical facial appearance features can be obtained by connecting the feature map at current \\(i\\)-th level with those at the higher \\((i + 1)\\)-th level, both of which are upsampled to the same resolution. \\(1\\times 1\\) convolution layers are used to reduce the channel number to avoid overlarge computational cost:" + }, + { + "type": "equation", + "bbox": [ + 0.348, + 0.322, + 0.786, + 0.361 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} ^ {i} = \\left\\{ \\begin{array}{l l} \\operatorname {C o n v} ^ {i} ([ U p ^ {i} (\\mathbf {F} ^ {i + 1}), \\mathbf {D} ^ {i} ]), & i = 0, 1, 2 \\\\ \\mathbf {D} ^ {i}, & i = 3 \\end{array} \\right. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.37, + 0.789, + 0.506 + ], + "angle": 0, + "content": "where \\( Up^{i}(.) \\) denotes the \\( i \\)-th upsampling convolution layer, and \\( Conv^{i}(.) \\) denotes the \\( i \\)-th \\( 1 \\times 1 \\) convolution layer mentioned above. [...] denotes channelwise concatenation. Afterwards, the feature map at \\( i \\)-th level \\( F^{i} \\) has a shape of \\( B \\times 3 \\times C^{i} \\times H^{i} \\times W^{i} \\), and is reshaped and splitted into three orthogonal sub feature maps, each of which has a shape of \\( B \\times C^{i} \\times H^{i} \\times W^{i} \\), representing coordinates in the three-dimensional space, i.e. plane \\( \\mathbf{F}_{xy} \\), \\( \\mathbf{F}_{yz} \\) and \\( \\mathbf{F}_{xz} \\) respectively. \\( B \\) is the batch size, and \\( C^{i} \\), \\( H^{i} \\) and \\( W^{i} \\) are the channel number, the height and width at \\( i \\)-th level. Since this tri-plane representation is obtained from multiple identities, it is capable of fitting an unseen person during inference time." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.507, + 0.789, + 0.675 + ], + "angle": 0, + "content": "Extracting Features for Neural Rendering. To render the target image \\(\\mathbf{I}_{tar}\\), which may have different head pose with the source image, we need to find the correspondence between the position in the target image space and the source feature tri-plane space. This can be realized through camera transformation between the source pose \\(\\mathbf{P}_{src}\\) and the target pose \\(\\mathbf{P}_{tar}\\), together with bilinear interpolation. Concretely, given a 3D point \\(\\mathbf{x}\\) on a ray \\(\\mathbf{r}(t)\\) constructed with target pose \\(\\mathbf{P}_{tar} = \\{\\mathbf{R}_{tar},\\mathbf{T}_{tar}\\}\\), it is firstly calibrated by the deformation predicted from the Cross-modal Facial Deformation Field (Sec.3.2) to get the deformed point \\(\\tilde{\\mathbf{x}}\\). Then, feature vectors \\(\\mathbf{F}_{xy}(\\tilde{\\mathbf{x}}),\\mathbf{F}_{yz}(\\tilde{\\mathbf{x}})\\) and \\(\\mathbf{F}_{xz}(\\tilde{\\mathbf{x}})\\) are retrieved from three orthogonal planes \\(\\mathbf{F}_{xy},\\mathbf{F}_{yz}\\) and \\(\\mathbf{F}_{xz}\\), according to the pose of the source image \\(\\mathbf{P}_{src} = \\{\\mathbf{R}_{src},\\mathbf{T}_{src}\\}\\) via bilinear interpolation Interp(:.):" + }, + { + "type": "equation", + "bbox": [ + 0.371, + 0.682, + 0.785, + 0.7 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {x y z} ^ {i} (\\tilde {\\mathbf {x}}) = \\operatorname {I n t e r p} \\left(\\mathbf {F} ^ {i}, \\operatorname {I n v} \\left(\\mathbf {P} _ {s r c}\\right) \\cdot \\tilde {\\mathbf {x}}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.707, + 0.788, + 0.753 + ], + "angle": 0, + "content": "\\(Inv(\\mathbf{P}_{src})\\) is the inverse source pose which transforms 3D positions onto the source tri-plane space. These extracted feature vectors are further used for generating the talking head image through volume rendering." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.772, + 0.577, + 0.787 + ], + "angle": 0, + "content": "3.2 Cross-modal Facial Deformation Field" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "To animate the target speaker, a crucial step is to get the precise deformation which describe face motions driven by the speech signal. Solely predicting the deformation from audio struggles to learn the global motion of the whole face, since" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.657, + 0.114, + 0.732, + 0.127 + ], + "angle": 0, + "content": "\\(\\mathrm{S}^3\\mathrm{D}\\) -NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.31, + 0.15, + 0.7, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.274, + 0.785, + 0.318 + ], + "angle": 0, + "content": "Fig. 3: Results with naive deformation module (left) and our Cross-modal Facial Deformation Field (right). Lower Face regions have the largest activations in the heatmap, which denote their strongest correlations with the driven speech signal." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.33, + 0.788, + 0.495 + ], + "angle": 0, + "content": "audio is a local signal which only correlates strongly with the mouth area. Thus it tends to yield blurry results (shown in Fig.3, left). For correctly modeling the global motion, structural and textural information of the whole face is required. Our Cross-modal Facial Deformation Field firstly aggregates the multi-scale visual features to a unified embedding, which contains the global information of the entire face area, then calculates the correlation score between the audio signal and this embedding through cross attention. The correlation score is used as a prior knowledge for determining the importance of motions from each part of the face, making deformation prediction more precise. As is shown in Fig.3, our Cross-modal Facial Deformation Field is free from the blurry issues because of the correctly learned motion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.497, + 0.788, + 0.618 + ], + "angle": 0, + "content": "Aggregating Multi-scale Facial Features. To learn the features which represents the whole face, all the hierarchical appearance feature maps \\(\\mathbf{F}^0\\) to \\(\\mathbf{F}^3\\) are first reshaped to the same resolution \\((H_s\\times W_s)\\) and concatenated along the channel. Then, these features are flattened and fed into a slot-attention module [21] to exchange information from different scales: \\(\\mathbf{F}_{agg} = SlotAttention(\\mathbf{F}^{0\\dots 3})\\). The aggregated slot feature embedding has a shape of \\(B\\times (H_{s}\\times W_{s})\\times D_{slot}\\), where \\(D_{slot}\\) is the slot feature dimension. This learning process makes the embedding contain rich structural and textural information of the speaker." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.619, + 0.788, + 0.695 + ], + "angle": 0, + "content": "Calculating Audio-Visual Correlation Scores. The input speech signal is firstly processed with a 1D full convolutional network to get the driving audio feature \\(\\mathbf{a}_{dri}\\). Relevance between audio and face motion is calculated through the Multi-Head Cross-Attention (MHCA) scores with the aggregated visual embedding \\(\\mathbf{F}_{agg}\\) as query, the driving audio feature \\(\\mathbf{a}_{dri}\\) as key and value:" + }, + { + "type": "equation", + "bbox": [ + 0.317, + 0.708, + 0.786, + 0.784 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {F} _ {c m} = M H C A \\left(\\mathbf {F} _ {a g g}, \\mathbf {a} _ {d r i}\\right) \\\\ = \\operatorname {S o f t m a x} \\left[ \\frac {\\mathbf {F} _ {a g g} \\boldsymbol {W} ^ {Q} \\left(\\mathbf {a} _ {d r i} \\boldsymbol {W} ^ {K}\\right) ^ {T}}{\\sqrt {d}} \\right] \\mathbf {a} _ {d r i} \\boldsymbol {W} ^ {V}, \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.794, + 0.788, + 0.842 + ], + "angle": 0, + "content": "where \\(\\pmb{W}^Q, \\pmb{W}^K, \\pmb{W}^V\\) are the projection matrices with hidden dimension \\(d\\), respectively. Cross-attention scores between the speech signal and the whole face feature representation indicate the relevance of each face area with the speech" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.127 + ], + "angle": 0, + "content": "D. Li et al." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.149, + 0.784, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.407, + 0.788, + 0.465 + ], + "angle": 0, + "content": "Fig. 4: Qualitative comparison with single-shot methods. Our \\(\\mathrm{S}^3\\mathrm{D}\\)-NeRF yields the most correct lip shapes and the clearest teeth. Note that different methods adopt different face alignment tools, and the ground truth row demonstrates the raw image without alignment, so the face poses from different methods are slightly different." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.479, + 0.788, + 0.569 + ], + "angle": 0, + "content": "input, and can serve as a strong prior for modeling the dynamic talking face in a region-aware manner. We observed that the mouth area of the speaker, which moves with the audio signal has the largest deformation scales. Also shown in Fig.3, heatmaps are calculated by averaging \\(\\mathbf{F}_{cm}\\) channelwisely, and have shown larger activations on lower face regions which contains not only mouth area but also the muscles of the lower face." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.57, + 0.788, + 0.631 + ], + "angle": 0, + "content": "Predicting Facial Deformation. The facial deformation prediction module \\(Deform(\\ldots)\\) has a U-Net [30] like architecture, with the original 3D position \\(\\mathbf{x}\\), the cross-modal correlation score \\(\\mathbf{F}_{cm}\\), and the driving audio feature \\(\\mathbf{a}_{dri}\\) as inputs, this deformation prediction process can be formalized as" + }, + { + "type": "equation", + "bbox": [ + 0.395, + 0.637, + 0.607, + 0.653 + ], + "angle": 0, + "content": "\\[\n\\Delta \\mathbf {x} = D e f o r m (\\mathbf {x}, \\mathbf {a} _ {d r i}, \\mathbf {F} _ {c m}),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.411, + 0.653, + 0.785, + 0.67 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {x}} = \\mathbf {x} + \\Delta \\mathbf {x}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.679, + 0.788, + 0.711 + ], + "angle": 0, + "content": "With the calibrated 3D points, source features from the tri-plane can be extracted faithfully through Eq.2." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.729, + 0.588, + 0.745 + ], + "angle": 0, + "content": "3.3 High Fidelity Talking Head Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.807 + ], + "angle": 0, + "content": "Volume Rendering. To generate an image with volume rendering, given the feature vector \\(\\mathbf{F}_{xyz}^{i}(\\tilde{\\mathbf{x}})\\) retrieved from the tri-plane, a tri-plane decoder takes it as input and predicts the color density \\(\\mathbf{c}\\) and occupancy \\(\\sigma\\) at that position. The final color of a pixel \\(\\mathbf{C}(\\mathbf{r})\\) along the ray is calculated through:" + }, + { + "type": "equation", + "bbox": [ + 0.374, + 0.813, + 0.787, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\mathbf {C} (\\mathbf {r}) = \\int_ {t _ {n}} ^ {t _ {f}} T (t) \\sigma (\\mathbf {r} (t), \\mathbf {a}) \\mathbf {c} (\\mathbf {r} (t), \\mathbf {a}) d t, \\tag {5}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.657, + 0.114, + 0.732, + 0.127 + ], + "angle": 0, + "content": "\\(\\mathrm{S}^3\\mathrm{D}\\) -NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.786, + 0.173 + ], + "angle": 0, + "content": "Table 1: Quantitative comparison with single-shot methods, The best and the second best results are emphasized." + }, + { + "type": "table", + "bbox": [ + 0.226, + 0.185, + 0.784, + 0.378 + ], + "angle": 0, + "content": "
DatasetMetricGTWav2LipMakeItTalkPC-AVSStyleTalkSadTalker\\( S^3 \\)D-NeRF
HDTFSSIM ↑10.7490.5590.5820.7760.6780.819
LPIPS ↓00.3320.4220.4110.2890.3860.258
F-LMD ↓03.5826.8463.9253.6234.6212.799
M-LMD ↓03.6526.1523.9123.2733.6542.929
CPBD ↑0.2840.1950.1390.2060.2340.1920.263
Sync ↑8.1448.3654.8135.0966.1886.4736.514
MEADSSIM ↑10.8290.6260.6730.8500.7550.845
LPIPS ↓00.2120.3670.3590.1990.2230.161
F-LMD ↓02.1266.0818.3722.6223.8352.014
M-LMD ↓03.3776.7905.1923.0613.6312.921
CPBD ↑0.1760.1010.0650.0680.1450.0810.152
Sync ↑7.7027.7135.1435.4666.0575.7076.768
" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.411, + 0.785, + 0.461 + ], + "angle": 0, + "content": "where \\( T(t) = \\exp \\left(-\\int_{t_n}^t\\sigma (s)ds\\right) \\) denotes for the accumulated transmittance along the ray from \\( t_n \\) to \\( t \\), \\( t_n \\) and \\( t_f \\) are the lower and the upper bound of depth." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.464, + 0.786, + 0.629 + ], + "angle": 0, + "content": "Coarse-to-fine Generation Strategy. Talking head frames contain various backgrounds and complex outer face textures which are hard for a NeRF to render at the same time. To decline the difficulty for training, we adopt a coarse-to-fine strategy. Specifically, the inner face part \\(\\mathbf{I}^{masked}\\) of an arbitrary frame \\(\\mathbf{I}\\) is decoupled in advance. During training, the inner face is rendered by the NeRF model, and the synthesized image is denoted as \\(\\mathbf{I}_{gen}^{coar}\\), while the fine texture details of the outer face and the background is complemented by an additional super-resolution module. We utilize Deep3DFaceRecon [7] to generate a 3D-aware face segmentation mask for inner-outer face area separation. This coarse-to-fine generation strategy guarantees that the NeRF can focus on the important face area." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.632, + 0.786, + 0.71 + ], + "angle": 0, + "content": "Super Resolution Module. The super resolution module takes the coarse output face \\(\\mathbf{I}_{gen}^{coar}\\) and the full source face image \\(\\mathbf{I}_{src}^{full}\\) as input, and generates the talking head frame \\(\\mathbf{I}_{fine}^{gen}\\) with background and more details. The whole super resolution process can be written as \\(\\mathbf{I}_{gen}^{fine} = \\text{Supres}(\\mathbf{I}_{gen}^{coar}, \\mathbf{I}_{src}^{full})\\). More details are shown in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.743, + 0.45, + 0.759 + ], + "angle": 0, + "content": "3.4 Lip-sync Discriminator" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.78, + 0.786, + 0.84 + ], + "angle": 0, + "content": "Our lip sync discriminator has a visual branch and an audio branch. The former takes \\( T \\) frames of the lip area as input and outputs their embedding \\( \\mathbf{e}_l \\), while the latter gets an audio clip and outputs its corresponding feature embedding \\( \\mathbf{e}_a \\). Then, the synchronization level between the audio and the video sequence" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.348, + 0.128 + ], + "angle": 0, + "content": "D. Li et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.714, + 0.163 + ], + "angle": 0, + "content": "can be judged by calculating the cosine similarity between \\(\\mathbf{e}_l\\) and \\(\\mathbf{e}_a\\):" + }, + { + "type": "equation", + "bbox": [ + 0.375, + 0.173, + 0.786, + 0.203 + ], + "angle": 0, + "content": "\\[\n\\cos (\\mathbf {e} _ {l}, \\mathbf {e} _ {a}) = \\frac {\\mathbf {e} _ {l} \\cdot \\mathbf {e} _ {a}}{\\operatorname* {m a x} \\left(\\| \\mathbf {e} _ {l} \\| _ {2} \\cdot \\| \\mathbf {e} _ {a} \\| _ {2} , \\epsilon\\right)}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.212, + 0.784, + 0.273 + ], + "angle": 0, + "content": "where \\(\\epsilon\\) is a margin which is set to 0. Different from that used in Wav2Lip [27], our lip sync discriminator is trained on a larger mixed dataset containing HDTF [44] and LRS2 [1], with a single contrastive triplet loss for more distinctive feature representations:" + }, + { + "type": "equation", + "bbox": [ + 0.335, + 0.284, + 0.786, + 0.322 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {s y n c}} ^ {\\text {d i s}} = \\max \\left(0, \\eta + \\cos \\left(\\mathbf {e} _ {l}, \\mathbf {e} _ {a} ^ {+}\\right) - \\cos \\left(\\mathbf {e} _ {l}, \\mathbf {e} _ {a} ^ {-}\\right)\\right) \\tag {7} \\\\ + \\max \\left(0, \\eta + \\cos \\left(\\mathbf {e} _ {l} ^ {+}, \\mathbf {e} _ {a}\\right) - \\cos \\left(\\mathbf {e} _ {l} ^ {-}, \\mathbf {e} _ {a}\\right)\\right). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.333, + 0.785, + 0.424 + ], + "angle": 0, + "content": "\\(\\mathbf{e}_a^+\\) and \\(\\mathbf{e}_a^-\\) regard the positive and the negative audio embedding according to the lip sequence respectively, so as \\(\\mathbf{e}_l^+\\) and \\(\\mathbf{e}_l^-\\), \\(\\eta\\) is a margin set to 0.5. While training our model, the lip-sync discriminator is frozen and \\(T\\) frames of the same person are generated once a time. The lip regions are cropped through the pre-extracted bounding boxes of the ground-truth. The loss used for S\\(^3\\)D-NeRF training can be formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.423, + 0.436, + 0.785, + 0.454 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s y n c}} ^ {\\text {g e n}} = \\cos \\left(\\mathbf {e} _ {l} ^ {\\text {g e n}}, \\mathbf {e} _ {g t}\\right). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.463, + 0.784, + 0.508 + ], + "angle": 0, + "content": "while \\(\\mathbf{e}_l^{gen}\\) is the embedding of the generated lip sequence and \\(\\mathbf{e}_{gt}\\) is the embedding of the ground-truth audio. By penalizing the out-of-sync audio-visual pair, more precise mouth shapes can be achieved." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.531, + 0.384, + 0.544 + ], + "angle": 0, + "content": "3.5 Loss Functions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.555, + 0.784, + 0.646 + ], + "angle": 0, + "content": "In addition to the lip synchronization loss \\(\\mathcal{L}_{sync}^{gen}\\), we use a pixelwise \\(l_{2}\\) reconstruction loss \\(\\mathcal{L}_{pix}\\), a feature level perceptual loss \\(\\mathcal{L}_{per}\\) to match the generated face with the target face, and a GAN loss \\(\\mathcal{L}_{adv}\\) with the architecture of StyleGAN discriminator [15]. All the losses are imposed on both the coarse and the fine faces. An \\(l_{2}\\) loss \\(\\mathcal{L}_{deform}\\) is used to regularize the prediction deformation. Our S\\(^3\\)D-NeRF is finally trained with the weighted sum of the above losses as:" + }, + { + "type": "equation", + "bbox": [ + 0.263, + 0.658, + 0.785, + 0.675 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {p i x} + \\lambda_ {p e r} \\mathcal {L} _ {p e r} + \\lambda_ {a d v} \\mathcal {L} _ {a d v} + \\lambda_ {\\text {s y n c}} ^ {\\text {g e n}} \\mathcal {L} _ {\\text {s y n c}} ^ {\\text {g e n}} + \\lambda_ {\\text {d e f o r m}} \\mathcal {L} _ {\\text {d e f o r m}}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.686, + 0.627, + 0.703 + ], + "angle": 0, + "content": "with \\(\\lambda_{per} = 0.01\\), \\(\\lambda_{adv} = 1\\), \\(\\lambda_{sync}^{gen} = 0.5\\), \\(\\lambda_{deform} = 0.001\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.724, + 0.376, + 0.74 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.755, + 0.447, + 0.77 + ], + "angle": 0, + "content": "4.1 Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.785, + 0.841 + ], + "angle": 0, + "content": "Datasets. Our \\(\\mathrm{S}^3\\mathrm{D}\\)-NeRF is trained on HDTF [44] training split, where each video ranges from tens of seconds to more than ten minutes in length, and evaluated on HDTF testing split and MEAD testing split [35] for comparison with single-shot methods. There is no overlap between the training and the testing" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.657, + 0.114, + 0.732, + 0.127 + ], + "angle": 0, + "content": "\\(\\mathrm{S}^3\\mathrm{D}\\) -NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.147, + 0.498, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.358, + 0.32, + 0.374, + 0.331 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.16, + 0.78, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.641, + 0.319, + 0.658, + 0.331 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.342, + 0.788, + 0.399 + ], + "angle": 0, + "content": "Fig. 5: Left (a): Qualitative comparison with NeRF-based methods, when encountering a new identity, our S\\(^3\\)D-NeRF successfully synthesizes a portrait faithfully without any retraining. Right (b): High fidelity generation results with multi-view consistency. Ground truth are the images with front view. Most of the face details are preserved." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.411, + 0.785, + 0.456 + ], + "angle": 0, + "content": "split. For comparison with NeRF-based methods, we use the videos provided by AD-NeRF [11] and GeneFace [39]. The video sampling rate is set to 25 FPS. All the faces are cropped and aligned to a fix resolution of \\(256 \\times 256\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.457, + 0.788, + 0.534 + ], + "angle": 0, + "content": "Implementation Details. The whole proposed framework is implemented by PyTorch and can be trained end-to-end. It takes about two days to train our S\\(^3\\)D-NeRF on 4 NVIDIA-A100 GPUs. The sample rate of the speech signal is set to 16000. We use Wav2Vec2.0 [2] to extract the driving audio features. More details and the user study can be found in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.557, + 0.434, + 0.572 + ], + "angle": 0, + "content": "4.2 Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.788, + 0.689 + ], + "angle": 0, + "content": "Several widely used metrics are adopted to carry out our quantitative evaluation. We use SSIM and LPIPS [42] to measure the similarity of the generated speaker with the ground truth in pixel level and feature level respectively. To compare the identity fidelity, we use Facial Landmark Distance (F-LMD). To compare the clear extent of the images, we use Cumulative Probability of Blur Detection (CPBD). To evaluate the audio-lip synchronization, we use the SyncNet confidence score (Sync) [6] and Mouth Landmark Distance (M-LMD) [5]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Comparison with Single-Shot Methods. Our competitors include Wav2Lip [27], MakeiTalk [46], PC-AVS [45], SadTalker [43] and StyleTalk [22]. For the above methods, we take their provided checkpoints and strictly follow the test protocol of current single-shot methods to conduct our evaluation. MakeiTalk has shown the worst image quality and lip shape. Wav2Lip also uses a pretrained lip-sync expert as supervision, and it has larger batch size and its model is simpler. Thus, it achieves a SyncNet score which is even better than ground truth. However, Wav2Lip tends to generate blurry faces, resulting in an inferior face fidelity. StyleTalk and SadTalker have shown competitive performance on image quality. Our S\\(^3\\)D-NeRF achieves the second-best lip-sync result, meanwhile sur" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.348, + 0.128 + ], + "angle": 0, + "content": "D. Li et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.189 + ], + "angle": 0, + "content": "Table 2: Quantitative comparison with NeRF-based methods. The best and the second best results are highlighted. Our S\\(^3\\)D-NeRF holds competitive image quality and audio-visual consistency, meanwhile the best generalize ability." + }, + { + "type": "table", + "bbox": [ + 0.303, + 0.199, + 0.703, + 0.334 + ], + "angle": 0, + "content": "
MetricGTGeneFaceAD-NeRFER-NeRFS3D-NeRF
SSIM ↑10.8280.8460.8840.852
LPIPS ↓00.0980.0870.0680.104
M-LMD ↓02.4361.9821.6591.493
F-LMD ↓01.8981.7151.1582.019
CPBD ↑0.2780.2070.2200.2390.244
Sync ↑8.9705.3665.6266.7817.118
FPS ↑N/A0.130.13299.5
Fit Time / h ↓N/A42364.5<0.01
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.359, + 0.787, + 0.45 + ], + "angle": 0, + "content": "passes the competitors significantly on quality metrics at both image level and feature level. Moreover, our method has the lowest face landmark distance and mouth landmark distance, which indicates that our S\\(^3\\)D-NeRF keep facial and lip structure better than other methods. On MEAD dataset where all the identities are unseen before, our S\\(^3\\)D-NeRF still achieves competitive performance, which demonstrates the strong generalization ability of our approach." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.45, + 0.788, + 0.737 + ], + "angle": 0, + "content": "Comparison with NeRF-based Methods. We also compare S\\(^3\\)D-NeRF with current identity specific NeRF-based methods. Three start-of-the-art approaches AD-NeRF [11] GeneFace [39] and ER-NeRF [17] are chosen as our opponents. It is worth to note that our method has never seen the identity nor the driving audio in the evaluation set, while the person-specific NeRFs are trained on a training set which has the same identity as the evaluation set (although no image overlap). In addition to the metrics mentioned above, we also take account of the data efficiency metrics, FPS: Frame Per Second during inference, which can reflects the speed of a rendering model. Fit Time: Times needed to generate the plausible talking head videos of a new identity with a pretrained model available. ER-NeRF has shown the strong image quality with the fastest rendering speed. GeneFace needs to train a postnet for each new identity, which takes extra time. Our model is also competitive in image quality. Moreover, it has the best lip synchronization score, together with the best mouth-LMD, which indicates the superiority on the specific audio-driven task. Moreover, our model has the highest CPBD, that means we generate the clearest images. Besides, S\\(^3\\)D-NeRF has the strongest generalization ability since it is free from retraining when coming up with an new speaker, and has shown an excellent inference speed." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.757, + 0.421, + 0.772 + ], + "angle": 0, + "content": "4.3 Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Comparison with Single-Shot Methods. A direct comparison with other person agnostic methods are shown in Fig.4. Our S\\(^3\\)D-NeRF has shown more precise lip shapes and clearer teeth. At the same time, the detailed facial features such as skin texture, eyebrows and teeth splits are preserved by our method." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.657, + 0.114, + 0.732, + 0.127 + ], + "angle": 0, + "content": "\\(\\mathrm{S}^3\\mathrm{D}\\) -NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.31, + 0.145, + 0.7, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.312, + 0.359, + 0.688, + 0.373 + ], + "angle": 0, + "content": "Fig. 6: Qualitative ablation study over key components." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.406, + 0.785, + 0.435 + ], + "angle": 0, + "content": "Table 3: Quantitative ablation study. The full S\\(^3\\)D-NeRF achieves the best landmark distance and lip synthesis results, with a little sacrifice on image quality." + }, + { + "type": "table", + "bbox": [ + 0.29, + 0.447, + 0.72, + 0.542 + ], + "angle": 0, + "content": "
MethodSSIM ↑LPIPS ↓F-LMD ↓M-LMD ↓CPBD ↑Sync ↑
ND0.7440.2783.4753.5070.1655.201
FCD0.8340.2422.8553.1960.2466.106
w/o Deform0.5560.457N/AN/A0.1400.240
w/o Lip Sync0.8010.2633.2763.3420.2555.649
S3D-NeRF0.8290.2582.7992.9290.2636.514
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.589, + 0.788, + 0.771 + ], + "angle": 0, + "content": "Comparison with NeRF-based Methods. A qualitative comparison with NeRF-based methods are shown in Fig.5. AD-NeRF has shown the head-torso separation problem, which seriously harm the fidelity of the portrait (Obama face at the third row). Due to the identity specific training process, NeRF based methods can achieve excellent speaker reconstruction. Their generated face regions are very similar to the ground truth. However, on a novel identity, all the person specific NeRFs failed to generate a plausible result. On the contrary, our S\\(^3\\)D-NeRF has shown excellent generation results. The images from our S\\(^3\\)D-NeRF do not reach a similarity as high as NeRF-based methods, but they have more distinct facial details, each face region of the speaker are kept completely. Images generated by our method are sharper and their lip areas are much clearer." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Multi-view Generation Results. Benefited from the 3D consistency, our method can generate faces with various view angles with the same lip shape, which is shown in the right part of the Fig.5. The facial details such as wrinkles and ear-rings are well kept, indicating the generative power of our method." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.348, + 0.128 + ], + "angle": 0, + "content": "D. Li et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.147, + 0.497, + 0.162 + ], + "angle": 0, + "content": "4.4 Ablation Studies & Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.172, + 0.787, + 0.353 + ], + "angle": 0, + "content": "The Hierarchical Facial Appearance Encoder is a necessary structure which can not be discarded or replaced. To study the function of our Cross-modal Facial Deformation Field, we designed several kinds of variants. Firstly, we directly reconstruct the frame of the target speaker without deformation, denoted as w/o deform. In this case, the audio feature is concatenated with the tri-plane feature and fed into the NeRF model Secondly, we replace our Cross-modal Facial Deformation Field with a naive deformation module mentioned in Sec.3.2, denoted as Naive Deform (ND). Thirdly, to show the effect of the correlation score as a prior, we train a model with a deformation module which takes directly the tri-plane feature as an additional input, denoted as Feat-Concatenate Deform (FCD). A model without the lip-sync discriminator is denoted as w/o lip-sync. Qualitative and quantitative results are shown in Tab.3 and Fig.6." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.354, + 0.787, + 0.504 + ], + "angle": 0, + "content": "The model without deformation modules totally fails to keep the appearance of the speaker, resulting in a poor video quality, not even performing speech animations. The Naive Deform model yields a coarse image similar to the average face, and its final generated talking head frames are less sharper. As for the Feat-Concatenate Deform model, it has kept the source feature better, resulting in a slight higher image quality metrics, however, because of it does not take the relationship between the speech signal and different face regions, some details of the mouth and the face area are missing. The model trained without the lip-sync discriminator synthesizes less precise lip shape. Our full model has the best performance on most of the metrics." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.527, + 0.613, + 0.543 + ], + "angle": 0, + "content": "5 Limitations & Ethical Considerations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.559, + 0.785, + 0.603 + ], + "angle": 0, + "content": "Limitations. Changing backgrounds like person specific-NeRFs [11, 17] is not supported currently. When the head pose is overlarge, the contour of the portrait sometimes tends to be blur. The above issues will be studied in future." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.604, + 0.787, + 0.665 + ], + "angle": 0, + "content": "Ethical Considerations. Once trained, our \\(\\mathrm{S}^3\\mathrm{D}\\)-NeRF can synthesize videos of any given person saying something that they never actually said. This may cause some moral or legal problems. We are committed to offer our model to fight against the potential abuses." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.688, + 0.36, + 0.704 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In this paper, we propose a Single-Shot Speech-Driven Neural Radiance Field (S\\(^3\\)D-NeRF) for synthesizing audio driving talking heads with high fidelity. Several key components are proposed to assist the generation process, including a Hierarchical Facial Appearance Encoder, a Cross-modal Facial Deformation Field and a lip-sync discriminator. Our S\\(^3\\)D-NeRF is the first attempt to explore single-shot speech-driven talking head generation with NeRFs. Validated by comprehensive experiments, S\\(^3\\)D-NeRF has shown clear improvement against current state-of-the-art, both on video quality and generalization ability." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.657, + 0.114, + 0.732, + 0.127 + ], + "angle": 0, + "content": "\\(\\mathrm{S}^3\\mathrm{D}\\) -NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.176, + 0.785, + 0.222 + ], + "angle": 0, + "content": "This work is supported by the National Natural Science Foundation of China (NSFC) under Grants 62372452, 62272460, Youth Innovation Promotion Association CAS, and Alibaba Research Intern Program." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.244, + 0.323, + 0.26 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.274, + 0.787, + 0.317 + ], + "angle": 0, + "content": "1. Afouras, T., Chung, J.S., Senior, A., Vinyals, O., Zisserman, A.: Deep audio-visual speech recognition. IEEE transactions on pattern analysis and machine intelligence 44(12), 8717-8727 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.318, + 0.787, + 0.358 + ], + "angle": 0, + "content": "2. Baevski, A., Zhou, Y., Mohamed, A., Auli, M.: wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems 33, 12449-12460 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.359, + 0.787, + 0.385 + ], + "angle": 0, + "content": "3. Blanz, V., Vetter, T.: A morphable model for the synthesis of 3d faces. In: Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pp. 157-164 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.386, + 0.787, + 0.44 + ], + "angle": 0, + "content": "4. Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., De Mello, S., Gallo, O., Guibas, L.J., Tremblay, J., Khamis, S., et al.: Efficient geometry-aware 3d generative adversarial networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16123-16133 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.441, + 0.787, + 0.481 + ], + "angle": 0, + "content": "5. Chen, L., Maddox, R.K., Duan, Z., Xu, C.: Hierarchical cross-modal talking face generation with dynamic pixel-wise loss. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7832-7841 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.482, + 0.787, + 0.537 + ], + "angle": 0, + "content": "6. Chung, J.S., Zisserman, A.: Out of time: automated lip sync in the wild. In: Computer Vision-ACCV 2016 Workshops: ACCV 2016 International Workshops, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part II 13. pp. 251-263. Springer (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.538, + 0.787, + 0.579 + ], + "angle": 0, + "content": "7. Deng, Y., Yang, J., Xu, S., Chen, D., Jia, Y., Tong, X.: Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In: IEEE Computer Vision and Pattern Recognition Workshops (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.58, + 0.787, + 0.633 + ], + "angle": 0, + "content": "8. Gafni, G., Thies, J., Zollhöfer, M., Nießner, M.: Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8649-8658 (June 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.634, + 0.787, + 0.675 + ], + "angle": 0, + "content": "9. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. Communications of the ACM 63(11), 139–144 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.676, + 0.787, + 0.717 + ], + "angle": 0, + "content": "10. Guo, J., Zhu, X., Yang, Y., Yang, F., Lei, Z., Li, S.Z.: Towards fast, accurate and stable 3d dense face alignment. In: Proceedings of the European Conference on Computer Vision (ECCV) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.718, + 0.787, + 0.758 + ], + "angle": 0, + "content": "11. Guo, Y., Chen, K., Liang, S., Liu, Y.J., Bao, H., Zhang, J.: Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5784-5794 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.759, + 0.787, + 0.799 + ], + "angle": 0, + "content": "12. Hong, F.T., Zhang, L., Shen, L., Xu, D.: Depth-aware generative adversarial network for talking head video generation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3397-3406 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.8, + 0.787, + 0.84 + ], + "angle": 0, + "content": "13. Hong, Y., Peng, B., Xiao, H., Liu, L., Zhang, J.: Headnerf: A real-time nerf-based parametric head model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20374-20384 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.274, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.127 + ], + "angle": 0, + "content": "D. Li et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "14. Jang, W., Agapito, L.: Codenerf: Disentangled neural radiance fields for object categories. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12949-12958 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.232 + ], + "angle": 0, + "content": "15. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of stylegan. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8110-8119 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.232, + 0.785, + 0.259 + ], + "angle": 0, + "content": "16. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.259, + 0.785, + 0.313 + ], + "angle": 0, + "content": "17. Li, J., Zhang, J., Bai, X., Zhou, J., Gu, L.: Efficient region-aware neural radiance fields for high-fidelity talking portrait synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 7568-7578 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.315, + 0.785, + 0.355 + ], + "angle": 0, + "content": "18. Li, W., Zhang, L., Wang, D., Zhao, B., Wang, Z., Chen, M., Zhang, B., Wang, Z., Bo, L., Li, X.: One-shot high-fidelity talking-head synthesis with deformable neural radiance field (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.785, + 0.397 + ], + "angle": 0, + "content": "19. Lin, T.Y., Dollar, P., Girshick, R., He, K., Hariharan, B., Belongie, S.: Feature pyramid networks for object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2117-2125 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.453 + ], + "angle": 0, + "content": "20. Liu, X., Xu, Y., Wu, Q., Zhou, H., Wu, W., Zhou, B.: Semantic-aware implicit neural audio-driven video portrait generation. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVII. pp. 106-125. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.494 + ], + "angle": 0, + "content": "21. Locatello, F., Weissenborn, D., Unterthiner, T., Mahendran, A., Heigold, G., Uszkoreit, J., Dosovitskiy, A., Kipf, T.: Object-centric learning with slot attention. Advances in Neural Information Processing Systems 33, 11525-11538 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.785, + 0.536 + ], + "angle": 0, + "content": "22. Ma, Y., Wang, S., Hu, Z., Fan, C., Lv, T., Ding, Y., Deng, Z., Yu, X.: Styletalk: One-shot talking head generation with controllable speaking styles. In: Proceedings of the Thirty-Seventh AAAI Conference on Artificial Intelligence (AAAI) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.785, + 0.577 + ], + "angle": 0, + "content": "23. Ma, Z., Zhu, X., Qi, G.J., Lei, Z., Zhang, L.: Otavatar: One-shot talking face avatar with controllable tri-plane rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16901-16910 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.618 + ], + "angle": 0, + "content": "24. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.785, + 0.66 + ], + "angle": 0, + "content": "25. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.66, + 0.785, + 0.702 + ], + "angle": 0, + "content": "26. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., MartinBrualla, R.: Nerfies: Deformable neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5865-5874 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.785, + 0.742 + ], + "angle": 0, + "content": "27. Prajwal, K., Mukhopadhyay, R., Namboodiri, V.P., Jawahar, C.: A lip sync expert is all you need for speech to lip generation in the wild. In: Proceedings of the 28th ACM International Conference on Multimedia. pp. 484-492 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.743, + 0.785, + 0.785 + ], + "angle": 0, + "content": "28. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-nerf: Neural radiance fields for dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10318-10327 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.785, + 0.813 + ], + "angle": 0, + "content": "29. Ren, Y., Li, G., Chen, Y., Li, T.H., Liu, S.: Pirenderer: Controllable portrait image generation via semantic neural rendering (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "30. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.657, + 0.114, + 0.732, + 0.127 + ], + "angle": 0, + "content": "\\(\\mathrm{S}^3\\mathrm{D}\\) -NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.245, + 0.147, + 0.786, + 0.175 + ], + "angle": 0, + "content": "Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.175, + 0.787, + 0.229 + ], + "angle": 0, + "content": "31. Shen, S., Li, W., Zhu, Z., Duan, Y., Zhou, J., Lu, J.: Learning dynamic facial radiance fields for few-shot talking head synthesis. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XII. pp. 666-682. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.231, + 0.786, + 0.271 + ], + "angle": 0, + "content": "32. Tang, J., Wang, K., Zhou, H., Chen, X., He, D., Hu, T., Liu, J., Zeng, G., Wang, J.: Real-time neural radiance talking portrait synthesis via audio-spatial decomposition. arXiv preprint arXiv:2211.12368 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.271, + 0.787, + 0.325 + ], + "angle": 0, + "content": "33. Thies, J., Elgharib, M., Tewari, A., Theobalt, C., Nießner, M.: Neural voice puppetry: Audio-driven facial reenactment. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVI 16. pp. 716-731. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.325, + 0.787, + 0.379 + ], + "angle": 0, + "content": "34. Wang, J., Zhao, K., Zhang, S., Zhang, Y., Shen, Y., Zhao, D., Zhou, J.: Lipformer: High-fidelity and generalizable talking face generation with a pre-learned facial codebook. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13844–13853 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.787, + 0.42 + ], + "angle": 0, + "content": "35. Wang, K., Wu, Q., Song, L., Yang, Z., Wu, W., Qian, C., He, R., Qiao, Y., Loy, C.C.: Mead: A large-scale audio-visual dataset for emotional talking-face generation. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.421, + 0.787, + 0.461 + ], + "angle": 0, + "content": "36. Wang, S., Li, L., Ding, Y., Yu, X.: One-shot talking face generation from single-speaker audio-visual correlation learning. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 2531-2539 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.461, + 0.787, + 0.502 + ], + "angle": 0, + "content": "37. Wiles, O., Koepke, A., Zisserman, A.: X2face: A network for controlling face generation using images, audio, and pose codes. In: Proceedings of the European conference on computer vision (ECCV). pp. 670-686 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.502, + 0.787, + 0.542 + ], + "angle": 0, + "content": "38. Ye, Z., Jiang, Z., Ren, Y., Liu, J., He, J., Zhao, Z.: Geneface: Generalized and high-fidelity audio-driven 3d talking face synthesis. arXiv preprint arXiv:2301.13430 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.542, + 0.787, + 0.583 + ], + "angle": 0, + "content": "39. Ye, Z., Jiang, Z., Ren, Y., Liu, J., He, J., Zhao, Z.: Geneface: Generalized and high-fidelity audio-driven 3d talking face synthesis. arXiv preprint arXiv:2301.13430 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.583, + 0.787, + 0.623 + ], + "angle": 0, + "content": "40. Yi, R., Ye, Z., Zhang, J., Bao, H., Liu, Y.J.: Audio-driven talking face video generation with learning-based personalized head pose. arXiv preprint arXiv:2002.10137 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.623, + 0.787, + 0.664 + ], + "angle": 0, + "content": "41. Yu, W., Fan, Y., Zhang, Y., Wang, X., Yin, F., Bai, Y., Cao, Y.P., Shan, Y., Wu, Y., Sun, Z., et al.: Nofa: Nerf-based one-shot facial avatar reconstruction. In: ACM SIGGRAPH 2023 Conference Proceedings. pp. 1-12 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.664, + 0.787, + 0.704 + ], + "angle": 0, + "content": "42. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. Proceedings of the IEEE conference on computer vision and pattern recognition pp. 586-595 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.705, + 0.787, + 0.745 + ], + "angle": 0, + "content": "43. Zhang, W., Cun, X., Wang, X., Zhang, Y., Shen, X., Guo, Y., Shan, Y., Wang, F.: Sadtalker: Learning realistic 3d motion coefficients for stylized audio-driven single image talking face animation. arXiv preprint arXiv:2211.12194 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.745, + 0.787, + 0.786 + ], + "angle": 0, + "content": "44. Zhang, Z., Li, L., Ding, Y., Fan, C.: Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3661-3670 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.787, + 0.84 + ], + "angle": 0, + "content": "45. Zhou, H., Sun, Y., Wu, W., Loy, C.C., Wang, X., Liu, Z.: Pose-controllable talking face generation by implicitly modularized audio-visual representation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.348, + 0.129 + ], + "angle": 0, + "content": "D. Li et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.19 + ], + "angle": 0, + "content": "46. Zhou, Y., Han, X., Shechtman, E., Echevarria, J., Kalogerakis, E., Li, D.: Makelttalk: speaker-aware talking-head animation. ACM Transactions On Graphics (TOG) 39(6), 1-15 (2020)" + } + ] +] \ No newline at end of file diff --git a/2024/S^3D-NeRF_ Single-Shot Speech-Driven Neural Radiance Field for High Fidelity Talking Head Synthesis/49757949-cf34-4879-95e0-e53ca1bcc2a6_origin.pdf b/2024/S^3D-NeRF_ Single-Shot Speech-Driven Neural Radiance Field for High Fidelity Talking Head Synthesis/49757949-cf34-4879-95e0-e53ca1bcc2a6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e66f84d8a56556cd6d7b60081c99fc42b8f306ed --- /dev/null +++ b/2024/S^3D-NeRF_ Single-Shot Speech-Driven Neural Radiance Field for High Fidelity Talking Head Synthesis/49757949-cf34-4879-95e0-e53ca1bcc2a6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:891435f1410ee42a1de114e181a8954b6157554a8ea5335608a3db093a53c641 +size 2796152 diff --git a/2024/S^3D-NeRF_ Single-Shot Speech-Driven Neural Radiance Field for High Fidelity Talking Head Synthesis/full.md b/2024/S^3D-NeRF_ Single-Shot Speech-Driven Neural Radiance Field for High Fidelity Talking Head Synthesis/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8a7cd43af2925c5e89d5cc19b38d5e5297689505 --- /dev/null +++ b/2024/S^3D-NeRF_ Single-Shot Speech-Driven Neural Radiance Field for High Fidelity Talking Head Synthesis/full.md @@ -0,0 +1,290 @@ +# $\mathrm{S}^{3}\mathrm{D}-\mathrm{NeRF}$ : Single-Shot Speech-Driven Neural Radiance Field for High Fidelity Talking Head Synthesis + +Dongze Li $^{1,2}$ Kang Zhao $^{3}$ , Wei Wang $^{2\star}$ , Yifeng Ma $^{3}$ , Bo Peng $^{2}$ , Yingya Zhang $^{3}$ and Jing Dong $^{2}$ + +$^{1}$ School of Artificial Intelligence, University of Chinese Academy of Sciences + +$^{2}$ NLPR & MAIS, Institute of Automation, Chinese Academy of Sciences + +3 Alibaba Group + +dongze.li@cripac.ia.ac.cn; {wang,bo.peng,jdong}@nlpr.ia.ac.cn + +{zhaokang.zk, yingya.zyy, mayifeng.myf}@alibaba-inc.com + +Abstract. Talking head synthesis is a practical technique with wide applications. Current Neural Radiance Field (NeRF) based approaches have shown their superiority on driving one-shot talking heads with videos or signals regressed from audio. However, most of them failed to take the audio as driven information directly, unable to enjoy the flexibility and availability of speech. Since mapping audio signals to face deformation is non-trivial, we design a Single-Shot Speech-Driven Neural Radiance Field (S $^3$ D-NeRF) method in this paper to tackle the following three difficulties: learning a representative appearance feature for each identity, modeling motion of different face regions with audio, and keeping the temporal consistency of the lip area. To this end, we introduce a Hierarchical Facial Appearance Encoder to learn multi-scale representations for catching the appearance of different speakers, and elaborate a Cross-modal Facial Deformation Field to perform speech animation according to the relationship between the audio signal and different face regions. Moreover, to enhance the temporal consistency of the important lip area, we introduce a lip-sync discriminator to penalize the out-of-sync audio-visual sequences. Extensive experiments have shown that our S $^3$ D-NeRF surpasses previous arts on both video fidelity and audio-lip synchronization. + +Keywords: Talking Head $\cdot$ Neural Radiance Fields + +# 1 Introduction + +Speech driven talking head synthesis is a promising technique and can be applied to a wide range of situations such as digital human, film making, virtual reality and video games. Current Neural Radiance Field (NeRF) [24] based methods [11,17,38] have shown their superiority on generating vivid talking portraits + +with high quality for their 3D consistency and view controllability. However, they can only be applied to a specific identity, and require a long video sequence from the same speaker to train on, which hampers them from broader application scenarios. Afterwards some progress have been made in the generalization of NeRF-based talking head methods. Most of them [18, 23, 41] are driven by intermediate facial representations, such as 3DMM coefficients extracted from video or expression coefficients regressed from audio. These intermediate representations may introduce information loss more or less, leading to sub-optimal synthesis results. + +Considering it is non-trivial to map audio signals to face deformation, we attribute the challenge of performing high-fidelity one-shot speech animation with Neural Radiance Fields to three aspects: 1) Learning Representative Appearance Features for Each Identity. Different people have different facial appearance i.e. shape and texture, it is difficult for a vanilla NeRF to model these details of several speakers simultaneously. During the inference phase, it becomes even harder to generate a talking head sequence with high quality given that only a single image is available. 2) Modeling Motion of Different Face Regions with Audio. Video driven techniques can utilize global driving signals which can describe the motion of the whole face, such as depth priors [12], 3DMM meshes [41] or PNCC [10,18] maps. However, in a speech driven task, the input audio merely has strong correlation with the lower face [27,34]. Directly predicting face motions according to the input audio will cause insufficient modeling of different face regions, harming the fidelity of the synthesized portrait. 3) Keeping Temporal Consistency of the Lip Area. The lip area has the greatest importance in speech animation. Due to the lack of constraints on this part, videos generated by NeRFs have sometimes shown incorrect mouth shapes. + +In this paper, we propose S $^3$ D-NeRF, namely Single-Shot Speech-Driven Neural Radiance Field, to synthesize high fidelity talking head videos. Given a single shot source image and a driven audio sequence, our method can synthesize vivid free view talking head videos with accurate mouth shapes. + +To fully capture the appearance of an arbitrary speaker, a Hierarchical Facial Appearance Encoder is presented to extract the multi-scale facial features of the single-shot source image with a down sample convolutional network armed with a feature pyramid. These features contain rich structural and textural information of the source identity. Then, a multi-scale tri-plane [4] representation is constructed for neural rendering. To realize accurate speech animation, a Cross-modal Facial Deformation Field is proposed to faithfully catch the correlations between the speech signal and the visual features from different face regions through cross attention, and predicts the motion of the whole face precisely. To further enhance the consistency of the lip area, a lip-sync discriminator is imposed on the lower face of the generated video frames and the driving audio signal. By penalizing the out-of-sync audio-visual sequences, higher temporal consistency of the lip area can be achieved. During training, we adapt a coarse-to-fine image generation strategy to reduce the difficulty of modeling multiple speakers at the same time. Specifically, a coarse talking head frame which + +![](images/2942f86784c916ec195debd937915bdf9e263090c948fad1f41231fdea451f57.jpg) +Fig. 1: A showcase of our S $^3$ D-NeRF, which generates high quality face portraits with fine-grained face texture and mouth details. + +contains the important inner face area is synthesized through standard volume neural rendering, while the texture details are refined with a super-resolution module, resulting in a high fidelity talking head frame. + +Our main contributions can be summarized as below: + +- We propose S $^3$ D-NeRF, a single-shot NeRF-based talking head synthesis framework. Our method extends NeRF-based speech animation techniques to handle arbitrary unseen identities. +- Several key components are presented to assist the talking head synthesis procedure. Including a Hierarchical Facial Appearance Encoder for representative feature extraction, a Cross-modal Facial Deformation Field for accurate speech animation, and a lip-sync discriminator for better temporal consistency of lip area. +- Comprehensive experiments have shown the superiority of our S $^3$ D-NeRF over previous arts on both video fidelity and audio-lip synchronization. + +# 2 Related Work + +# 2.1 Speech Driven Talking Head Synthesis + +The goal of speech driven talking head synthesis is to animate a speaker according to input audiios. According to the generalization ability of the model to new identities, current speech driven methods can be divided into two categories: identity agnostic methods and identity specific methods. + +Identity agnostic methods are capable of generating speech videos of an arbitrary speaker once their training procedure is finished, with one or several images or video sequences as inputs. Among them, single-shot methods [22, 27, 43, 45] have attracted considerable interests due to their high generalization ability and data efficiency. Image-based methods [27, 37, 45] utilize deep generative + +![](images/321b5cb07b72ba26343cb38eb30e3a4471f3373f03017a8c8eb0b15111282280.jpg) +Fig. 2: The full pipeline of our S $^3$ D-NeRF. The Hierarchical Facial Appearance Encoder extracts representative features from the masked face region of the single-shot source image, for high fidelity neural rendering of an arbitrary identity. The Cross-modal Facial Deformation Field accurately models the motion of different face regions, with the help of the correlation score calculated through cross attention between audiovisual features. Texture details are complemented with the super-resolution module. + +networks [9] or auto-encoders [16] to generate continuous video clips. Model-based methods predict intermediate representations such as motion field [36], 2D landmarks [46] or 3D Morphable Model (3DMM) [3] coefficients [22, 29, 43] with the input audio signal, and animate the head with these representations. StyleTalk [22] predicts 3DMM through a transformer architecture and can generate portraits with diverse speaking styles under the help of a pretrained renderer [29]. SadTalker [43] uses several representation networks to learn lip and expression basis. Although remarkable progresses have been made, current single-shot methods still suffer in image quality and audio-lip synchronization because of the error accumulated in the audio to representation mapping. + +Identity specific methods train an individual model for each identity with higher fidelity. Earlier methods [33, 40] utilize strong priors such as 3D meshes or dense key points for portrait generation. Current NeRF-based methods [11, 20, 31, 38] have shed light on high fidelity speaker synthesis. They optimize on the ground truth video clip of a single person, and can synthesize pose-controllable faces with fine-grained details. AD-NeRF [11] uses two separated NeRFs to model the head and the torso part respectively. SSP-NeRF [20] performs rays re-sampling based on the loss magnitude of different semantic regions. Currently, RAD-NeRF [32] and ER-NeRF [17] utilize hash table structures [25] to improve the rendering speed. GeneFace [39] proposes a facial radiance field with generalize ability but it also requires per identity finetuning and a complex multi-stage training process. Despite the above advantages, the above methods can only fit a single identity, and need tedious re-training when a new identity is encountered. At the same time, since no constraints are imposed on image sequences, the temporal consistency of the lip area is also not satisfactory. + +# 2.2 NeRF-based Face Portrait Reconstruction + +It is a natural idea to model human faces with NeRFs in a 3D aware manner. However, vanilla NeRFs struggle to handle dynamic objects, and are unsuitable for animating human faces which have diverse poses and expressions. There mainly exist two series of solutions for this problem. The former [14] is to condition the original Radiance Field with additional global control signals, e.g. expression parameters. The latter solve this problem by learning an additional deformation field [26, 28] to warp the 3D points. NerFace [8] is the first approach to reconstruct animatable face avatars with NeRFs conditioned on 3DMM expression parameters. HeadNeRF [13], NOFA [41] and OTAvatar [23] utilize GAN inversion to project onto the latent space of a 3D aware GAN [4] and drive the generated avatar through different latent codes. HiDe-NeRF [18] utilizes Projected Normalized Coordinate Code (PNCC) [10] as a global expression control indicator, and drives the face image with high fidelity. Comparing with the above mentioned methods, driving an arbitrary head with speech signal directly through NeRFs is a more challenging problem and is rarely explored. + +# 3 Method + +The overall pipeline of our S $^3$ D-NeRF is shown in Fig.2. The 3D positions of the talking head can be derived through the given head pose of the target image $\mathbf{P}_{tar}$ , which is available during both training and inference time. Firstly, our Hierarchical Facial Appearance Encoder takes a single-shot image as input, and models its appearance with a deep convolutional network, where multi-resolution feature maps are extracted to construct an efficient representation i.e. multiscale appearance tri-planes, which are known to be capable of modeling several identities faithfully at the same time [4, 18]. Feature points from these planes are retrieved according to the 3D position for rendering. Then, the motion of the talking face are predicted by our Cross-modal Facial Deformation Field. It firstly calculates the correlation between an aggregated visual feature embedding and the audio signal through cross attention, then takes the result as a prior knowledge to assist the regression process of the deformation, which is used for calibrating the 3D position. A coarse face is rendered through the classic volume rendering process [24] with color and density predicted from the feature points on the tri-plane. Finally, a super-resolution module is used to refine the coarse face and yields the high fidelity portrait. + +# 3.1 Hierarchical Facial Appearance Encoder + +An effective feature representation is important to accurately model the appearance of a speaker. Initially proposed in [4], tri-plane has shown its strong representative power and is much faster comparing with the vanilla NeRFs. The Hierarchical Facial Appearance Encoder constructs this representation based on multi-scale feature maps from the one-shot source image, and yields the feature points for further volumetric rendering process. + +Constructing Representative Appearance Features. We adopt a feature pyramid structure [19] to extract source features for its ability to learn multi-scale information from different resolutions, and reshape all the feature maps channelwisely to construct several orthogonal planes as in [4]. To be more concrete, given the source image $\mathbf{I}_{src}$ , feature maps with different resolutions $\{\mathbf{D}^0,\mathbf{D}^1,\mathbf{D}^2,\mathbf{D}^3\}$ are extracted through several convolutional downsample blocks $Down^{i}(.)$ , while $\mathbf{D}^0 = Down^0 (\mathbf{I}_{src})$ and $\mathbf{D}^i = Down^i (\mathbf{D}^{i - 1})$ . Then, hierarchical facial appearance features can be obtained by connecting the feature map at current $i$ -th level with those at the higher $(i + 1)$ -th level, both of which are upsampled to the same resolution. $1\times 1$ convolution layers are used to reduce the channel number to avoid overlarge computational cost: + +$$ +\mathbf {F} ^ {i} = \left\{ \begin{array}{l l} \operatorname {C o n v} ^ {i} ([ U p ^ {i} (\mathbf {F} ^ {i + 1}), \mathbf {D} ^ {i} ]), & i = 0, 1, 2 \\ \mathbf {D} ^ {i}, & i = 3 \end{array} \right. \tag {1} +$$ + +where $Up^{i}(.)$ denotes the $i$ -th upsampling convolution layer, and $Conv^{i}(.)$ denotes the $i$ -th $1 \times 1$ convolution layer mentioned above. [...] denotes channelwise concatenation. Afterwards, the feature map at $i$ -th level $F^{i}$ has a shape of $B \times 3 \times C^{i} \times H^{i} \times W^{i}$ , and is reshaped and splitted into three orthogonal sub feature maps, each of which has a shape of $B \times C^{i} \times H^{i} \times W^{i}$ , representing coordinates in the three-dimensional space, i.e. plane $\mathbf{F}_{xy}$ , $\mathbf{F}_{yz}$ and $\mathbf{F}_{xz}$ respectively. $B$ is the batch size, and $C^{i}$ , $H^{i}$ and $W^{i}$ are the channel number, the height and width at $i$ -th level. Since this tri-plane representation is obtained from multiple identities, it is capable of fitting an unseen person during inference time. + +Extracting Features for Neural Rendering. To render the target image $\mathbf{I}_{tar}$ , which may have different head pose with the source image, we need to find the correspondence between the position in the target image space and the source feature tri-plane space. This can be realized through camera transformation between the source pose $\mathbf{P}_{src}$ and the target pose $\mathbf{P}_{tar}$ , together with bilinear interpolation. Concretely, given a 3D point $\mathbf{x}$ on a ray $\mathbf{r}(t)$ constructed with target pose $\mathbf{P}_{tar} = \{\mathbf{R}_{tar},\mathbf{T}_{tar}\}$ , it is firstly calibrated by the deformation predicted from the Cross-modal Facial Deformation Field (Sec.3.2) to get the deformed point $\tilde{\mathbf{x}}$ . Then, feature vectors $\mathbf{F}_{xy}(\tilde{\mathbf{x}}),\mathbf{F}_{yz}(\tilde{\mathbf{x}})$ and $\mathbf{F}_{xz}(\tilde{\mathbf{x}})$ are retrieved from three orthogonal planes $\mathbf{F}_{xy},\mathbf{F}_{yz}$ and $\mathbf{F}_{xz}$ , according to the pose of the source image $\mathbf{P}_{src} = \{\mathbf{R}_{src},\mathbf{T}_{src}\}$ via bilinear interpolation Interp(:.): + +$$ +\mathbf {F} _ {x y z} ^ {i} (\tilde {\mathbf {x}}) = \operatorname {I n t e r p} \left(\mathbf {F} ^ {i}, \operatorname {I n v} \left(\mathbf {P} _ {s r c}\right) \cdot \tilde {\mathbf {x}}\right), \tag {2} +$$ + +$Inv(\mathbf{P}_{src})$ is the inverse source pose which transforms 3D positions onto the source tri-plane space. These extracted feature vectors are further used for generating the talking head image through volume rendering. + +# 3.2 Cross-modal Facial Deformation Field + +To animate the target speaker, a crucial step is to get the precise deformation which describe face motions driven by the speech signal. Solely predicting the deformation from audio struggles to learn the global motion of the whole face, since + +![](images/d1565c14fe61408989fd247862c40ea40430407f29470fd4ce80ba49a8093bfe.jpg) +Fig. 3: Results with naive deformation module (left) and our Cross-modal Facial Deformation Field (right). Lower Face regions have the largest activations in the heatmap, which denote their strongest correlations with the driven speech signal. + +audio is a local signal which only correlates strongly with the mouth area. Thus it tends to yield blurry results (shown in Fig.3, left). For correctly modeling the global motion, structural and textural information of the whole face is required. Our Cross-modal Facial Deformation Field firstly aggregates the multi-scale visual features to a unified embedding, which contains the global information of the entire face area, then calculates the correlation score between the audio signal and this embedding through cross attention. The correlation score is used as a prior knowledge for determining the importance of motions from each part of the face, making deformation prediction more precise. As is shown in Fig.3, our Cross-modal Facial Deformation Field is free from the blurry issues because of the correctly learned motion. + +Aggregating Multi-scale Facial Features. To learn the features which represents the whole face, all the hierarchical appearance feature maps $\mathbf{F}^0$ to $\mathbf{F}^3$ are first reshaped to the same resolution $(H_s\times W_s)$ and concatenated along the channel. Then, these features are flattened and fed into a slot-attention module [21] to exchange information from different scales: $\mathbf{F}_{agg} = SlotAttention(\mathbf{F}^{0\dots 3})$ . The aggregated slot feature embedding has a shape of $B\times (H_{s}\times W_{s})\times D_{slot}$ , where $D_{slot}$ is the slot feature dimension. This learning process makes the embedding contain rich structural and textural information of the speaker. + +Calculating Audio-Visual Correlation Scores. The input speech signal is firstly processed with a 1D full convolutional network to get the driving audio feature $\mathbf{a}_{dri}$ . Relevance between audio and face motion is calculated through the Multi-Head Cross-Attention (MHCA) scores with the aggregated visual embedding $\mathbf{F}_{agg}$ as query, the driving audio feature $\mathbf{a}_{dri}$ as key and value: + +$$ +\begin{array}{l} \mathbf {F} _ {c m} = M H C A \left(\mathbf {F} _ {a g g}, \mathbf {a} _ {d r i}\right) \\ = \operatorname {S o f t m a x} \left[ \frac {\mathbf {F} _ {a g g} \boldsymbol {W} ^ {Q} \left(\mathbf {a} _ {d r i} \boldsymbol {W} ^ {K}\right) ^ {T}}{\sqrt {d}} \right] \mathbf {a} _ {d r i} \boldsymbol {W} ^ {V}, \tag {3} \\ \end{array} +$$ + +where $\pmb{W}^Q, \pmb{W}^K, \pmb{W}^V$ are the projection matrices with hidden dimension $d$ , respectively. Cross-attention scores between the speech signal and the whole face feature representation indicate the relevance of each face area with the speech + +![](images/b03b76721c2e183dfc3b9d08d1145499c23c4277c708f1eb3ebe8c88a2843164.jpg) +Fig. 4: Qualitative comparison with single-shot methods. Our $\mathrm{S}^3\mathrm{D}$ -NeRF yields the most correct lip shapes and the clearest teeth. Note that different methods adopt different face alignment tools, and the ground truth row demonstrates the raw image without alignment, so the face poses from different methods are slightly different. + +input, and can serve as a strong prior for modeling the dynamic talking face in a region-aware manner. We observed that the mouth area of the speaker, which moves with the audio signal has the largest deformation scales. Also shown in Fig.3, heatmaps are calculated by averaging $\mathbf{F}_{cm}$ channelwisely, and have shown larger activations on lower face regions which contains not only mouth area but also the muscles of the lower face. + +Predicting Facial Deformation. The facial deformation prediction module $Deform(\ldots)$ has a U-Net [30] like architecture, with the original 3D position $\mathbf{x}$ , the cross-modal correlation score $\mathbf{F}_{cm}$ , and the driving audio feature $\mathbf{a}_{dri}$ as inputs, this deformation prediction process can be formalized as + +$$ +\Delta \mathbf {x} = D e f o r m (\mathbf {x}, \mathbf {a} _ {d r i}, \mathbf {F} _ {c m}), +$$ + +$$ +\tilde {\mathbf {x}} = \mathbf {x} + \Delta \mathbf {x}. \tag {4} +$$ + +With the calibrated 3D points, source features from the tri-plane can be extracted faithfully through Eq.2. + +# 3.3 High Fidelity Talking Head Generation + +Volume Rendering. To generate an image with volume rendering, given the feature vector $\mathbf{F}_{xyz}^{i}(\tilde{\mathbf{x}})$ retrieved from the tri-plane, a tri-plane decoder takes it as input and predicts the color density $\mathbf{c}$ and occupancy $\sigma$ at that position. The final color of a pixel $\mathbf{C}(\mathbf{r})$ along the ray is calculated through: + +$$ +\mathbf {C} (\mathbf {r}) = \int_ {t _ {n}} ^ {t _ {f}} T (t) \sigma (\mathbf {r} (t), \mathbf {a}) \mathbf {c} (\mathbf {r} (t), \mathbf {a}) d t, \tag {5} +$$ + +Table 1: Quantitative comparison with single-shot methods, The best and the second best results are emphasized. + +
DatasetMetricGTWav2LipMakeItTalkPC-AVSStyleTalkSadTalker\( S^3 \)D-NeRF
HDTFSSIM ↑10.7490.5590.5820.7760.6780.819
LPIPS ↓00.3320.4220.4110.2890.3860.258
F-LMD ↓03.5826.8463.9253.6234.6212.799
M-LMD ↓03.6526.1523.9123.2733.6542.929
CPBD ↑0.2840.1950.1390.2060.2340.1920.263
Sync ↑8.1448.3654.8135.0966.1886.4736.514
MEADSSIM ↑10.8290.6260.6730.8500.7550.845
LPIPS ↓00.2120.3670.3590.1990.2230.161
F-LMD ↓02.1266.0818.3722.6223.8352.014
M-LMD ↓03.3776.7905.1923.0613.6312.921
CPBD ↑0.1760.1010.0650.0680.1450.0810.152
Sync ↑7.7027.7135.1435.4666.0575.7076.768
+ +where $T(t) = \exp \left(-\int_{t_n}^t\sigma (s)ds\right)$ denotes for the accumulated transmittance along the ray from $t_n$ to $t$ , $t_n$ and $t_f$ are the lower and the upper bound of depth. + +Coarse-to-fine Generation Strategy. Talking head frames contain various backgrounds and complex outer face textures which are hard for a NeRF to render at the same time. To decline the difficulty for training, we adopt a coarse-to-fine strategy. Specifically, the inner face part $\mathbf{I}^{masked}$ of an arbitrary frame $\mathbf{I}$ is decoupled in advance. During training, the inner face is rendered by the NeRF model, and the synthesized image is denoted as $\mathbf{I}_{gen}^{coar}$ , while the fine texture details of the outer face and the background is complemented by an additional super-resolution module. We utilize Deep3DFaceRecon [7] to generate a 3D-aware face segmentation mask for inner-outer face area separation. This coarse-to-fine generation strategy guarantees that the NeRF can focus on the important face area. + +Super Resolution Module. The super resolution module takes the coarse output face $\mathbf{I}_{gen}^{coar}$ and the full source face image $\mathbf{I}_{src}^{full}$ as input, and generates the talking head frame $\mathbf{I}_{fine}^{gen}$ with background and more details. The whole super resolution process can be written as $\mathbf{I}_{gen}^{fine} = \text{Supres}(\mathbf{I}_{gen}^{coar}, \mathbf{I}_{src}^{full})$ . More details are shown in the appendix. + +# 3.4 Lip-sync Discriminator + +Our lip sync discriminator has a visual branch and an audio branch. The former takes $T$ frames of the lip area as input and outputs their embedding $\mathbf{e}_l$ , while the latter gets an audio clip and outputs its corresponding feature embedding $\mathbf{e}_a$ . Then, the synchronization level between the audio and the video sequence + +can be judged by calculating the cosine similarity between $\mathbf{e}_l$ and $\mathbf{e}_a$ : + +$$ +\cos (\mathbf {e} _ {l}, \mathbf {e} _ {a}) = \frac {\mathbf {e} _ {l} \cdot \mathbf {e} _ {a}}{\operatorname* {m a x} \left(\| \mathbf {e} _ {l} \| _ {2} \cdot \| \mathbf {e} _ {a} \| _ {2} , \epsilon\right)}, \tag {6} +$$ + +where $\epsilon$ is a margin which is set to 0. Different from that used in Wav2Lip [27], our lip sync discriminator is trained on a larger mixed dataset containing HDTF [44] and LRS2 [1], with a single contrastive triplet loss for more distinctive feature representations: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {s y n c}} ^ {\text {d i s}} = \max \left(0, \eta + \cos \left(\mathbf {e} _ {l}, \mathbf {e} _ {a} ^ {+}\right) - \cos \left(\mathbf {e} _ {l}, \mathbf {e} _ {a} ^ {-}\right)\right) \tag {7} \\ + \max \left(0, \eta + \cos \left(\mathbf {e} _ {l} ^ {+}, \mathbf {e} _ {a}\right) - \cos \left(\mathbf {e} _ {l} ^ {-}, \mathbf {e} _ {a}\right)\right). \\ \end{array} +$$ + +$\mathbf{e}_a^+$ and $\mathbf{e}_a^-$ regard the positive and the negative audio embedding according to the lip sequence respectively, so as $\mathbf{e}_l^+$ and $\mathbf{e}_l^-$ , $\eta$ is a margin set to 0.5. While training our model, the lip-sync discriminator is frozen and $T$ frames of the same person are generated once a time. The lip regions are cropped through the pre-extracted bounding boxes of the ground-truth. The loss used for S $^3$ D-NeRF training can be formulated as + +$$ +\mathcal {L} _ {\text {s y n c}} ^ {\text {g e n}} = \cos \left(\mathbf {e} _ {l} ^ {\text {g e n}}, \mathbf {e} _ {g t}\right). \tag {8} +$$ + +while $\mathbf{e}_l^{gen}$ is the embedding of the generated lip sequence and $\mathbf{e}_{gt}$ is the embedding of the ground-truth audio. By penalizing the out-of-sync audio-visual pair, more precise mouth shapes can be achieved. + +# 3.5 Loss Functions + +In addition to the lip synchronization loss $\mathcal{L}_{sync}^{gen}$ , we use a pixelwise $l_{2}$ reconstruction loss $\mathcal{L}_{pix}$ , a feature level perceptual loss $\mathcal{L}_{per}$ to match the generated face with the target face, and a GAN loss $\mathcal{L}_{adv}$ with the architecture of StyleGAN discriminator [15]. All the losses are imposed on both the coarse and the fine faces. An $l_{2}$ loss $\mathcal{L}_{deform}$ is used to regularize the prediction deformation. Our S $^3$ D-NeRF is finally trained with the weighted sum of the above losses as: + +$$ +\mathcal {L} = \mathcal {L} _ {p i x} + \lambda_ {p e r} \mathcal {L} _ {p e r} + \lambda_ {a d v} \mathcal {L} _ {a d v} + \lambda_ {\text {s y n c}} ^ {\text {g e n}} \mathcal {L} _ {\text {s y n c}} ^ {\text {g e n}} + \lambda_ {\text {d e f o r m}} \mathcal {L} _ {\text {d e f o r m}}, \tag {9} +$$ + +with $\lambda_{per} = 0.01$ , $\lambda_{adv} = 1$ , $\lambda_{sync}^{gen} = 0.5$ , $\lambda_{deform} = 0.001$ . + +# 4 Experiments + +# 4.1 Experimental Settings + +Datasets. Our $\mathrm{S}^3\mathrm{D}$ -NeRF is trained on HDTF [44] training split, where each video ranges from tens of seconds to more than ten minutes in length, and evaluated on HDTF testing split and MEAD testing split [35] for comparison with single-shot methods. There is no overlap between the training and the testing + +![](images/91fc46d10e0bf4b0169710d39b311fff04eabbbc5bec46bbd8644236ad3d00bd.jpg) +(a) + +![](images/f3707c90a25590976e50d4b9179ccf5d16531148b246a43b0f309357451eb7dd.jpg) +(b) +Fig. 5: Left (a): Qualitative comparison with NeRF-based methods, when encountering a new identity, our S $^3$ D-NeRF successfully synthesizes a portrait faithfully without any retraining. Right (b): High fidelity generation results with multi-view consistency. Ground truth are the images with front view. Most of the face details are preserved. + +split. For comparison with NeRF-based methods, we use the videos provided by AD-NeRF [11] and GeneFace [39]. The video sampling rate is set to 25 FPS. All the faces are cropped and aligned to a fix resolution of $256 \times 256$ . + +Implementation Details. The whole proposed framework is implemented by PyTorch and can be trained end-to-end. It takes about two days to train our S $^3$ D-NeRF on 4 NVIDIA-A100 GPUs. The sample rate of the speech signal is set to 16000. We use Wav2Vec2.0 [2] to extract the driving audio features. More details and the user study can be found in the appendix. + +# 4.2 Quantitative Results + +Several widely used metrics are adopted to carry out our quantitative evaluation. We use SSIM and LPIPS [42] to measure the similarity of the generated speaker with the ground truth in pixel level and feature level respectively. To compare the identity fidelity, we use Facial Landmark Distance (F-LMD). To compare the clear extent of the images, we use Cumulative Probability of Blur Detection (CPBD). To evaluate the audio-lip synchronization, we use the SyncNet confidence score (Sync) [6] and Mouth Landmark Distance (M-LMD) [5]. + +Comparison with Single-Shot Methods. Our competitors include Wav2Lip [27], MakeiTalk [46], PC-AVS [45], SadTalker [43] and StyleTalk [22]. For the above methods, we take their provided checkpoints and strictly follow the test protocol of current single-shot methods to conduct our evaluation. MakeiTalk has shown the worst image quality and lip shape. Wav2Lip also uses a pretrained lip-sync expert as supervision, and it has larger batch size and its model is simpler. Thus, it achieves a SyncNet score which is even better than ground truth. However, Wav2Lip tends to generate blurry faces, resulting in an inferior face fidelity. StyleTalk and SadTalker have shown competitive performance on image quality. Our S $^3$ D-NeRF achieves the second-best lip-sync result, meanwhile sur + +Table 2: Quantitative comparison with NeRF-based methods. The best and the second best results are highlighted. Our S $^3$ D-NeRF holds competitive image quality and audio-visual consistency, meanwhile the best generalize ability. + +
MetricGTGeneFaceAD-NeRFER-NeRFS3D-NeRF
SSIM ↑10.8280.8460.8840.852
LPIPS ↓00.0980.0870.0680.104
M-LMD ↓02.4361.9821.6591.493
F-LMD ↓01.8981.7151.1582.019
CPBD ↑0.2780.2070.2200.2390.244
Sync ↑8.9705.3665.6266.7817.118
FPS ↑N/A0.130.13299.5
Fit Time / h ↓N/A42364.5<0.01
+ +passes the competitors significantly on quality metrics at both image level and feature level. Moreover, our method has the lowest face landmark distance and mouth landmark distance, which indicates that our S $^3$ D-NeRF keep facial and lip structure better than other methods. On MEAD dataset where all the identities are unseen before, our S $^3$ D-NeRF still achieves competitive performance, which demonstrates the strong generalization ability of our approach. + +Comparison with NeRF-based Methods. We also compare S $^3$ D-NeRF with current identity specific NeRF-based methods. Three start-of-the-art approaches AD-NeRF [11] GeneFace [39] and ER-NeRF [17] are chosen as our opponents. It is worth to note that our method has never seen the identity nor the driving audio in the evaluation set, while the person-specific NeRFs are trained on a training set which has the same identity as the evaluation set (although no image overlap). In addition to the metrics mentioned above, we also take account of the data efficiency metrics, FPS: Frame Per Second during inference, which can reflects the speed of a rendering model. Fit Time: Times needed to generate the plausible talking head videos of a new identity with a pretrained model available. ER-NeRF has shown the strong image quality with the fastest rendering speed. GeneFace needs to train a postnet for each new identity, which takes extra time. Our model is also competitive in image quality. Moreover, it has the best lip synchronization score, together with the best mouth-LMD, which indicates the superiority on the specific audio-driven task. Moreover, our model has the highest CPBD, that means we generate the clearest images. Besides, S $^3$ D-NeRF has the strongest generalization ability since it is free from retraining when coming up with an new speaker, and has shown an excellent inference speed. + +# 4.3 Qualitative Results + +Comparison with Single-Shot Methods. A direct comparison with other person agnostic methods are shown in Fig.4. Our S $^3$ D-NeRF has shown more precise lip shapes and clearer teeth. At the same time, the detailed facial features such as skin texture, eyebrows and teeth splits are preserved by our method. + +![](images/da0083a72e3022b0b2821fb7aa3e33c6d085ea7c5e4918e34691988afc041faa.jpg) +Fig. 6: Qualitative ablation study over key components. + +Table 3: Quantitative ablation study. The full S $^3$ D-NeRF achieves the best landmark distance and lip synthesis results, with a little sacrifice on image quality. + +
MethodSSIM ↑LPIPS ↓F-LMD ↓M-LMD ↓CPBD ↑Sync ↑
ND0.7440.2783.4753.5070.1655.201
FCD0.8340.2422.8553.1960.2466.106
w/o Deform0.5560.457N/AN/A0.1400.240
w/o Lip Sync0.8010.2633.2763.3420.2555.649
S3D-NeRF0.8290.2582.7992.9290.2636.514
+ +Comparison with NeRF-based Methods. A qualitative comparison with NeRF-based methods are shown in Fig.5. AD-NeRF has shown the head-torso separation problem, which seriously harm the fidelity of the portrait (Obama face at the third row). Due to the identity specific training process, NeRF based methods can achieve excellent speaker reconstruction. Their generated face regions are very similar to the ground truth. However, on a novel identity, all the person specific NeRFs failed to generate a plausible result. On the contrary, our S $^3$ D-NeRF has shown excellent generation results. The images from our S $^3$ D-NeRF do not reach a similarity as high as NeRF-based methods, but they have more distinct facial details, each face region of the speaker are kept completely. Images generated by our method are sharper and their lip areas are much clearer. + +Multi-view Generation Results. Benefited from the 3D consistency, our method can generate faces with various view angles with the same lip shape, which is shown in the right part of the Fig.5. The facial details such as wrinkles and ear-rings are well kept, indicating the generative power of our method. + +# 4.4 Ablation Studies & Analysis + +The Hierarchical Facial Appearance Encoder is a necessary structure which can not be discarded or replaced. To study the function of our Cross-modal Facial Deformation Field, we designed several kinds of variants. Firstly, we directly reconstruct the frame of the target speaker without deformation, denoted as w/o deform. In this case, the audio feature is concatenated with the tri-plane feature and fed into the NeRF model Secondly, we replace our Cross-modal Facial Deformation Field with a naive deformation module mentioned in Sec.3.2, denoted as Naive Deform (ND). Thirdly, to show the effect of the correlation score as a prior, we train a model with a deformation module which takes directly the tri-plane feature as an additional input, denoted as Feat-Concatenate Deform (FCD). A model without the lip-sync discriminator is denoted as w/o lip-sync. Qualitative and quantitative results are shown in Tab.3 and Fig.6. + +The model without deformation modules totally fails to keep the appearance of the speaker, resulting in a poor video quality, not even performing speech animations. The Naive Deform model yields a coarse image similar to the average face, and its final generated talking head frames are less sharper. As for the Feat-Concatenate Deform model, it has kept the source feature better, resulting in a slight higher image quality metrics, however, because of it does not take the relationship between the speech signal and different face regions, some details of the mouth and the face area are missing. The model trained without the lip-sync discriminator synthesizes less precise lip shape. Our full model has the best performance on most of the metrics. + +# 5 Limitations & Ethical Considerations + +Limitations. Changing backgrounds like person specific-NeRFs [11, 17] is not supported currently. When the head pose is overlarge, the contour of the portrait sometimes tends to be blur. The above issues will be studied in future. + +Ethical Considerations. Once trained, our $\mathrm{S}^3\mathrm{D}$ -NeRF can synthesize videos of any given person saying something that they never actually said. This may cause some moral or legal problems. We are committed to offer our model to fight against the potential abuses. + +# 6 Conclusion + +In this paper, we propose a Single-Shot Speech-Driven Neural Radiance Field (S $^3$ D-NeRF) for synthesizing audio driving talking heads with high fidelity. Several key components are proposed to assist the generation process, including a Hierarchical Facial Appearance Encoder, a Cross-modal Facial Deformation Field and a lip-sync discriminator. Our S $^3$ D-NeRF is the first attempt to explore single-shot speech-driven talking head generation with NeRFs. Validated by comprehensive experiments, S $^3$ D-NeRF has shown clear improvement against current state-of-the-art, both on video quality and generalization ability. + +# Acknowledgements + +This work is supported by the National Natural Science Foundation of China (NSFC) under Grants 62372452, 62272460, Youth Innovation Promotion Association CAS, and Alibaba Research Intern Program. + +# References + +1. Afouras, T., Chung, J.S., Senior, A., Vinyals, O., Zisserman, A.: Deep audio-visual speech recognition. IEEE transactions on pattern analysis and machine intelligence 44(12), 8717-8727 (2018) +2. Baevski, A., Zhou, Y., Mohamed, A., Auli, M.: wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems 33, 12449-12460 (2020) +3. Blanz, V., Vetter, T.: A morphable model for the synthesis of 3d faces. In: Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pp. 157-164 (2023) +4. Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., De Mello, S., Gallo, O., Guibas, L.J., Tremblay, J., Khamis, S., et al.: Efficient geometry-aware 3d generative adversarial networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16123-16133 (2022) +5. Chen, L., Maddox, R.K., Duan, Z., Xu, C.: Hierarchical cross-modal talking face generation with dynamic pixel-wise loss. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7832-7841 (2019) +6. Chung, J.S., Zisserman, A.: Out of time: automated lip sync in the wild. In: Computer Vision-ACCV 2016 Workshops: ACCV 2016 International Workshops, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part II 13. pp. 251-263. Springer (2017) +7. Deng, Y., Yang, J., Xu, S., Chen, D., Jia, Y., Tong, X.: Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In: IEEE Computer Vision and Pattern Recognition Workshops (2019) +8. Gafni, G., Thies, J., Zollhöfer, M., Nießner, M.: Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8649-8658 (June 2021) +9. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. Communications of the ACM 63(11), 139–144 (2020) +10. Guo, J., Zhu, X., Yang, Y., Yang, F., Lei, Z., Li, S.Z.: Towards fast, accurate and stable 3d dense face alignment. In: Proceedings of the European Conference on Computer Vision (ECCV) (2020) +11. Guo, Y., Chen, K., Liang, S., Liu, Y.J., Bao, H., Zhang, J.: Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5784-5794 (2021) +12. Hong, F.T., Zhang, L., Shen, L., Xu, D.: Depth-aware generative adversarial network for talking head video generation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3397-3406 (2022) +13. Hong, Y., Peng, B., Xiao, H., Liu, L., Zhang, J.: Headnerf: A real-time nerf-based parametric head model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20374-20384 (2022) + +14. Jang, W., Agapito, L.: Codenerf: Disentangled neural radiance fields for object categories. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12949-12958 (2021) +15. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of stylegan. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8110-8119 (2020) +16. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013) +17. Li, J., Zhang, J., Bai, X., Zhou, J., Gu, L.: Efficient region-aware neural radiance fields for high-fidelity talking portrait synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 7568-7578 (October 2023) +18. Li, W., Zhang, L., Wang, D., Zhao, B., Wang, Z., Chen, M., Zhang, B., Wang, Z., Bo, L., Li, X.: One-shot high-fidelity talking-head synthesis with deformable neural radiance field (2023) +19. Lin, T.Y., Dollar, P., Girshick, R., He, K., Hariharan, B., Belongie, S.: Feature pyramid networks for object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2117-2125 (2017) +20. Liu, X., Xu, Y., Wu, Q., Zhou, H., Wu, W., Zhou, B.: Semantic-aware implicit neural audio-driven video portrait generation. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVII. pp. 106-125. Springer (2022) +21. Locatello, F., Weissenborn, D., Unterthiner, T., Mahendran, A., Heigold, G., Uszkoreit, J., Dosovitskiy, A., Kipf, T.: Object-centric learning with slot attention. Advances in Neural Information Processing Systems 33, 11525-11538 (2020) +22. Ma, Y., Wang, S., Hu, Z., Fan, C., Lv, T., Ding, Y., Deng, Z., Yu, X.: Styletalk: One-shot talking head generation with controllable speaking styles. In: Proceedings of the Thirty-Seventh AAAI Conference on Artificial Intelligence (AAAI) (2023) +23. Ma, Z., Zhu, X., Qi, G.J., Lei, Z., Zhang, L.: Otavatar: One-shot talking face avatar with controllable tri-plane rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16901-16910 (2023) +24. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020) +25. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022) +26. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., MartinBrualla, R.: Nerfies: Deformable neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5865-5874 (2021) +27. Prajwal, K., Mukhopadhyay, R., Namboodiri, V.P., Jawahar, C.: A lip sync expert is all you need for speech to lip generation in the wild. In: Proceedings of the 28th ACM International Conference on Multimedia. pp. 484-492 (2020) +28. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-nerf: Neural radiance fields for dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10318-10327 (2021) +29. Ren, Y., Li, G., Chen, Y., Li, T.H., Liu, S.: Pirenderer: Controllable portrait image generation via semantic neural rendering (2021) +30. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted + +Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015) +31. Shen, S., Li, W., Zhu, Z., Duan, Y., Zhou, J., Lu, J.: Learning dynamic facial radiance fields for few-shot talking head synthesis. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XII. pp. 666-682. Springer (2022) +32. Tang, J., Wang, K., Zhou, H., Chen, X., He, D., Hu, T., Liu, J., Zeng, G., Wang, J.: Real-time neural radiance talking portrait synthesis via audio-spatial decomposition. arXiv preprint arXiv:2211.12368 (2022) +33. Thies, J., Elgharib, M., Tewari, A., Theobalt, C., Nießner, M.: Neural voice puppetry: Audio-driven facial reenactment. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVI 16. pp. 716-731. Springer (2020) +34. Wang, J., Zhao, K., Zhang, S., Zhang, Y., Shen, Y., Zhao, D., Zhou, J.: Lipformer: High-fidelity and generalizable talking face generation with a pre-learned facial codebook. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13844–13853 (2023) +35. Wang, K., Wu, Q., Song, L., Yang, Z., Wu, W., Qian, C., He, R., Qiao, Y., Loy, C.C.: Mead: A large-scale audio-visual dataset for emotional talking-face generation. In: ECCV (2020) +36. Wang, S., Li, L., Ding, Y., Yu, X.: One-shot talking face generation from single-speaker audio-visual correlation learning. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 2531-2539 (2022) +37. Wiles, O., Koepke, A., Zisserman, A.: X2face: A network for controlling face generation using images, audio, and pose codes. In: Proceedings of the European conference on computer vision (ECCV). pp. 670-686 (2018) +38. Ye, Z., Jiang, Z., Ren, Y., Liu, J., He, J., Zhao, Z.: Geneface: Generalized and high-fidelity audio-driven 3d talking face synthesis. arXiv preprint arXiv:2301.13430 (2023) +39. Ye, Z., Jiang, Z., Ren, Y., Liu, J., He, J., Zhao, Z.: Geneface: Generalized and high-fidelity audio-driven 3d talking face synthesis. arXiv preprint arXiv:2301.13430 (2023) +40. Yi, R., Ye, Z., Zhang, J., Bao, H., Liu, Y.J.: Audio-driven talking face video generation with learning-based personalized head pose. arXiv preprint arXiv:2002.10137 (2020) +41. Yu, W., Fan, Y., Zhang, Y., Wang, X., Yin, F., Bai, Y., Cao, Y.P., Shan, Y., Wu, Y., Sun, Z., et al.: Nofa: Nerf-based one-shot facial avatar reconstruction. In: ACM SIGGRAPH 2023 Conference Proceedings. pp. 1-12 (2023) +42. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. Proceedings of the IEEE conference on computer vision and pattern recognition pp. 586-595 (2018) +43. Zhang, W., Cun, X., Wang, X., Zhang, Y., Shen, X., Guo, Y., Shan, Y., Wang, F.: Sadtalker: Learning realistic 3d motion coefficients for stylized audio-driven single image talking face animation. arXiv preprint arXiv:2211.12194 (2022) +44. Zhang, Z., Li, L., Ding, Y., Fan, C.: Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3661-3670 (2021) +45. Zhou, H., Sun, Y., Wu, W., Loy, C.C., Wang, X., Liu, Z.: Pose-controllable talking face generation by implicitly modularized audio-visual representation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021) + +46. Zhou, Y., Han, X., Shechtman, E., Echevarria, J., Kalogerakis, E., Li, D.: Makelttalk: speaker-aware talking-head animation. ACM Transactions On Graphics (TOG) 39(6), 1-15 (2020) \ No newline at end of file diff --git a/2024/S^3D-NeRF_ Single-Shot Speech-Driven Neural Radiance Field for High Fidelity Talking Head Synthesis/images.zip b/2024/S^3D-NeRF_ Single-Shot Speech-Driven Neural Radiance Field for High Fidelity Talking Head Synthesis/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e6205a82d1aba65e32124dc7a3dc1c56d83da02c --- /dev/null +++ b/2024/S^3D-NeRF_ Single-Shot Speech-Driven Neural Radiance Field for High Fidelity Talking Head Synthesis/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:93f3a2bc52353aa0b36e0a7f2942fa8b463c230aa9697862155e12a8bd79600c +size 575460 diff --git a/2024/S^3D-NeRF_ Single-Shot Speech-Driven Neural Radiance Field for High Fidelity Talking Head Synthesis/layout.json b/2024/S^3D-NeRF_ Single-Shot Speech-Driven Neural Radiance Field for High Fidelity Talking Head Synthesis/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..213c62f0a9d971e97af4a7c7ffbf1b79bdf86644 --- /dev/null +++ b/2024/S^3D-NeRF_ Single-Shot Speech-Driven Neural Radiance Field for High Fidelity Talking Head Synthesis/layout.json @@ -0,0 +1,9634 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 145, + 112, + 470, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 112, + 470, + 167 + ], + "spans": [ + { + "bbox": [ + 145, + 112, + 470, + 167 + ], + "type": "inline_equation", + "content": "\\mathrm{S}^{3}\\mathrm{D}-\\mathrm{NeRF}" + }, + { + "bbox": [ + 145, + 112, + 470, + 167 + ], + "type": "text", + "content": ": Single-Shot Speech-Driven Neural Radiance Field for High Fidelity Talking Head Synthesis" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 189, + 187, + 424, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 187, + 424, + 213 + ], + "spans": [ + { + "bbox": [ + 189, + 187, + 424, + 213 + ], + "type": "text", + "content": "Dongze Li" + }, + { + "bbox": [ + 189, + 187, + 424, + 213 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 189, + 187, + 424, + 213 + ], + "type": "text", + "content": " Kang Zhao" + }, + { + "bbox": [ + 189, + 187, + 424, + 213 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 189, + 187, + 424, + 213 + ], + "type": "text", + "content": ", Wei Wang" + }, + { + "bbox": [ + 189, + 187, + 424, + 213 + ], + "type": "inline_equation", + "content": "^{2\\star}" + }, + { + "bbox": [ + 189, + 187, + 424, + 213 + ], + "type": "text", + "content": ", Yifeng Ma" + }, + { + "bbox": [ + 189, + 187, + 424, + 213 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 189, + 187, + 424, + 213 + ], + "type": "text", + "content": ", Bo Peng" + }, + { + "bbox": [ + 189, + 187, + 424, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 189, + 187, + 424, + 213 + ], + "type": "text", + "content": ", Yingya Zhang" + }, + { + "bbox": [ + 189, + 187, + 424, + 213 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 189, + 187, + 424, + 213 + ], + "type": "text", + "content": " and Jing Dong" + }, + { + "bbox": [ + 189, + 187, + 424, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 149, + 220, + 462, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 220, + 462, + 232 + ], + "spans": [ + { + "bbox": [ + 149, + 220, + 462, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 149, + 220, + 462, + 232 + ], + "type": "text", + "content": " School of Artificial Intelligence, University of Chinese Academy of Sciences" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 157, + 232, + 456, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 232, + 456, + 243 + ], + "spans": [ + { + "bbox": [ + 157, + 232, + 456, + 243 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 157, + 232, + 456, + 243 + ], + "type": "text", + "content": " NLPR & MAIS, Institute of Automation, Chinese Academy of Sciences" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 272, + 243, + 342, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 272, + 243, + 342, + 255 + ], + "spans": [ + { + "bbox": [ + 272, + 243, + 342, + 255 + ], + "type": "text", + "content": "3 Alibaba Group" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 161, + 255, + 452, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 255, + 452, + 266 + ], + "spans": [ + { + "bbox": [ + 161, + 255, + 452, + 266 + ], + "type": "text", + "content": "dongze.li@cripac.ia.ac.cn; {wang,bo.peng,jdong}@nlpr.ia.ac.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 176, + 266, + 437, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 266, + 437, + 277 + ], + "spans": [ + { + "bbox": [ + 176, + 266, + 437, + 277 + ], + "type": "text", + "content": "{zhaokang.zk, yingya.zyy, mayifeng.myf}@alibaba-inc.com" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 159, + 308, + 453, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 308, + 453, + 527 + ], + "spans": [ + { + "bbox": [ + 159, + 308, + 453, + 527 + ], + "type": "text", + "content": "Abstract. Talking head synthesis is a practical technique with wide applications. Current Neural Radiance Field (NeRF) based approaches have shown their superiority on driving one-shot talking heads with videos or signals regressed from audio. However, most of them failed to take the audio as driven information directly, unable to enjoy the flexibility and availability of speech. Since mapping audio signals to face deformation is non-trivial, we design a Single-Shot Speech-Driven Neural Radiance Field (S" + }, + { + "bbox": [ + 159, + 308, + 453, + 527 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 159, + 308, + 453, + 527 + ], + "type": "text", + "content": "D-NeRF) method in this paper to tackle the following three difficulties: learning a representative appearance feature for each identity, modeling motion of different face regions with audio, and keeping the temporal consistency of the lip area. To this end, we introduce a Hierarchical Facial Appearance Encoder to learn multi-scale representations for catching the appearance of different speakers, and elaborate a Cross-modal Facial Deformation Field to perform speech animation according to the relationship between the audio signal and different face regions. Moreover, to enhance the temporal consistency of the important lip area, we introduce a lip-sync discriminator to penalize the out-of-sync audio-visual sequences. Extensive experiments have shown that our S" + }, + { + "bbox": [ + 159, + 308, + 453, + 527 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 159, + 308, + 453, + 527 + ], + "type": "text", + "content": "D-NeRF surpasses previous arts on both video fidelity and audio-lip synchronization." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 160, + 538, + 374, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 538, + 374, + 550 + ], + "spans": [ + { + "bbox": [ + 160, + 538, + 374, + 550 + ], + "type": "text", + "content": "Keywords: Talking Head " + }, + { + "bbox": [ + 160, + 538, + 374, + 550 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 538, + 374, + 550 + ], + "type": "text", + "content": " Neural Radiance Fields" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 571, + 230, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 571, + 230, + 583 + ], + "spans": [ + { + "bbox": [ + 132, + 571, + 230, + 583 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 597, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 597, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 597, + 482, + 647 + ], + "type": "text", + "content": "Speech driven talking head synthesis is a promising technique and can be applied to a wide range of situations such as digital human, film making, virtual reality and video games. Current Neural Radiance Field (NeRF) [24] based methods [11,17,38] have shown their superiority on generating vivid talking portraits" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 234, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 234, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 234, + 665 + ], + "type": "text", + "content": "* Corresponding author" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "with high quality for their 3D consistency and view controllability. However, they can only be applied to a specific identity, and require a long video sequence from the same speaker to train on, which hampers them from broader application scenarios. Afterwards some progress have been made in the generalization of NeRF-based talking head methods. Most of them [18, 23, 41] are driven by intermediate facial representations, such as 3DMM coefficients extracted from video or expression coefficients regressed from audio. These intermediate representations may introduce information loss more or less, leading to sub-optimal synthesis results." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 228, + 482, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 228, + 482, + 432 + ], + "spans": [ + { + "bbox": [ + 130, + 228, + 482, + 432 + ], + "type": "text", + "content": "Considering it is non-trivial to map audio signals to face deformation, we attribute the challenge of performing high-fidelity one-shot speech animation with Neural Radiance Fields to three aspects: 1) Learning Representative Appearance Features for Each Identity. Different people have different facial appearance i.e. shape and texture, it is difficult for a vanilla NeRF to model these details of several speakers simultaneously. During the inference phase, it becomes even harder to generate a talking head sequence with high quality given that only a single image is available. 2) Modeling Motion of Different Face Regions with Audio. Video driven techniques can utilize global driving signals which can describe the motion of the whole face, such as depth priors [12], 3DMM meshes [41] or PNCC [10,18] maps. However, in a speech driven task, the input audio merely has strong correlation with the lower face [27,34]. Directly predicting face motions according to the input audio will cause insufficient modeling of different face regions, harming the fidelity of the synthesized portrait. 3) Keeping Temporal Consistency of the Lip Area. The lip area has the greatest importance in speech animation. Due to the lack of constraints on this part, videos generated by NeRFs have sometimes shown incorrect mouth shapes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 434, + 483, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 483, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 483, + 483 + ], + "type": "text", + "content": "In this paper, we propose S" + }, + { + "bbox": [ + 130, + 434, + 483, + 483 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 130, + 434, + 483, + 483 + ], + "type": "text", + "content": "D-NeRF, namely Single-Shot Speech-Driven Neural Radiance Field, to synthesize high fidelity talking head videos. Given a single shot source image and a driven audio sequence, our method can synthesize vivid free view talking head videos with accurate mouth shapes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 486, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 483, + 666 + ], + "type": "text", + "content": "To fully capture the appearance of an arbitrary speaker, a Hierarchical Facial Appearance Encoder is presented to extract the multi-scale facial features of the single-shot source image with a down sample convolutional network armed with a feature pyramid. These features contain rich structural and textural information of the source identity. Then, a multi-scale tri-plane [4] representation is constructed for neural rendering. To realize accurate speech animation, a Cross-modal Facial Deformation Field is proposed to faithfully catch the correlations between the speech signal and the visual features from different face regions through cross attention, and predicts the motion of the whole face precisely. To further enhance the consistency of the lip area, a lip-sync discriminator is imposed on the lower face of the generated video frames and the driving audio signal. By penalizing the out-of-sync audio-visual sequences, higher temporal consistency of the lip area can be achieved. During training, we adapt a coarse-to-fine image generation strategy to reduce the difficulty of modeling multiple speakers at the same time. Specifically, a coarse talking head frame which" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "text", + "content": "D. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 154, + 118, + 462, + 262 + ], + "blocks": [ + { + "bbox": [ + 154, + 118, + 462, + 262 + ], + "lines": [ + { + "bbox": [ + 154, + 118, + 462, + 262 + ], + "spans": [ + { + "bbox": [ + 154, + 118, + 462, + 262 + ], + "type": "image", + "image_path": "2942f86784c916ec195debd937915bdf9e263090c948fad1f41231fdea451f57.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 267, + 480, + 289 + ], + "lines": [ + { + "bbox": [ + 131, + 267, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 131, + 267, + 480, + 289 + ], + "type": "text", + "content": "Fig. 1: A showcase of our S" + }, + { + "bbox": [ + 131, + 267, + 480, + 289 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 131, + 267, + 480, + 289 + ], + "type": "text", + "content": "D-NeRF, which generates high quality face portraits with fine-grained face texture and mouth details." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 314, + 479, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 314, + 479, + 350 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 479, + 350 + ], + "type": "text", + "content": "contains the important inner face area is synthesized through standard volume neural rendering, while the texture details are refined with a super-resolution module, resulting in a high fidelity talking head frame." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 146, + 350, + 383, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 350, + 383, + 361 + ], + "spans": [ + { + "bbox": [ + 146, + 350, + 383, + 361 + ], + "type": "text", + "content": "Our main contributions can be summarized as below:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 372, + 480, + 491 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 138, + 372, + 480, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 372, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 138, + 372, + 480, + 407 + ], + "type": "text", + "content": "- We propose S" + }, + { + "bbox": [ + 138, + 372, + 480, + 407 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 138, + 372, + 480, + 407 + ], + "type": "text", + "content": "D-NeRF, a single-shot NeRF-based talking head synthesis framework. Our method extends NeRF-based speech animation techniques to handle arbitrary unseen identities." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 408, + 480, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 408, + 480, + 467 + ], + "spans": [ + { + "bbox": [ + 138, + 408, + 480, + 467 + ], + "type": "text", + "content": "- Several key components are presented to assist the talking head synthesis procedure. Including a Hierarchical Facial Appearance Encoder for representative feature extraction, a Cross-modal Facial Deformation Field for accurate speech animation, and a lip-sync discriminator for better temporal consistency of lip area." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 468, + 480, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 468, + 480, + 491 + ], + "spans": [ + { + "bbox": [ + 138, + 468, + 480, + 491 + ], + "type": "text", + "content": "- Comprehensive experiments have shown the superiority of our S" + }, + { + "bbox": [ + 138, + 468, + 480, + 491 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 138, + 468, + 480, + 491 + ], + "type": "text", + "content": "D-NeRF over previous arts on both video fidelity and audio-lip synchronization." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 511, + 237, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 511, + 237, + 523 + ], + "spans": [ + { + "bbox": [ + 132, + 511, + 237, + 523 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 536, + 354, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 536, + 354, + 550 + ], + "spans": [ + { + "bbox": [ + 131, + 536, + 354, + 550 + ], + "type": "text", + "content": "2.1 Speech Driven Talking Head Synthesis" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 557, + 479, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 479, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 479, + 605 + ], + "type": "text", + "content": "The goal of speech driven talking head synthesis is to animate a speaker according to input audiios. According to the generalization ability of the model to new identities, current speech driven methods can be divided into two categories: identity agnostic methods and identity specific methods." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 606, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 480, + 665 + ], + "type": "text", + "content": "Identity agnostic methods are capable of generating speech videos of an arbitrary speaker once their training procedure is finished, with one or several images or video sequences as inputs. Among them, single-shot methods [22, 27, 43, 45] have attracted considerable interests due to their high generalization ability and data efficiency. Image-based methods [27, 37, 45] utilize deep generative" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 90, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 90, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 90, + 447, + 100 + ], + "type": "inline_equation", + "content": "\\mathrm{S}^3\\mathrm{D}" + }, + { + "bbox": [ + 402, + 90, + 447, + 100 + ], + "type": "text", + "content": " -NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 115, + 481, + 268 + ], + "blocks": [ + { + "bbox": [ + 132, + 115, + 481, + 268 + ], + "lines": [ + { + "bbox": [ + 132, + 115, + 481, + 268 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 481, + 268 + ], + "type": "image", + "image_path": "321b5cb07b72ba26343cb38eb30e3a4471f3373f03017a8c8eb0b15111282280.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 276, + 482, + 342 + ], + "lines": [ + { + "bbox": [ + 130, + 276, + 482, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 276, + 482, + 342 + ], + "type": "text", + "content": "Fig. 2: The full pipeline of our S" + }, + { + "bbox": [ + 130, + 276, + 482, + 342 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 130, + 276, + 482, + 342 + ], + "type": "text", + "content": "D-NeRF. The Hierarchical Facial Appearance Encoder extracts representative features from the masked face region of the single-shot source image, for high fidelity neural rendering of an arbitrary identity. The Cross-modal Facial Deformation Field accurately models the motion of different face regions, with the help of the correlation score calculated through cross attention between audiovisual features. Texture details are complemented with the super-resolution module." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 365, + 482, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 485 + ], + "type": "text", + "content": "networks [9] or auto-encoders [16] to generate continuous video clips. Model-based methods predict intermediate representations such as motion field [36], 2D landmarks [46] or 3D Morphable Model (3DMM) [3] coefficients [22, 29, 43] with the input audio signal, and animate the head with these representations. StyleTalk [22] predicts 3DMM through a transformer architecture and can generate portraits with diverse speaking styles under the help of a pretrained renderer [29]. SadTalker [43] uses several representation networks to learn lip and expression basis. Although remarkable progresses have been made, current single-shot methods still suffer in image quality and audio-lip synchronization because of the error accumulated in the audio to representation mapping." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": "Identity specific methods train an individual model for each identity with higher fidelity. Earlier methods [33, 40] utilize strong priors such as 3D meshes or dense key points for portrait generation. Current NeRF-based methods [11, 20, 31, 38] have shed light on high fidelity speaker synthesis. They optimize on the ground truth video clip of a single person, and can synthesize pose-controllable faces with fine-grained details. AD-NeRF [11] uses two separated NeRFs to model the head and the torso part respectively. SSP-NeRF [20] performs rays re-sampling based on the loss magnitude of different semantic regions. Currently, RAD-NeRF [32] and ER-NeRF [17] utilize hash table structures [25] to improve the rendering speed. GeneFace [39] proposes a facial radiance field with generalize ability but it also requires per identity finetuning and a complex multi-stage training process. Despite the above advantages, the above methods can only fit a single identity, and need tedious re-training when a new identity is encountered. At the same time, since no constraints are imposed on image sequences, the temporal consistency of the lip area is also not satisfactory." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "D. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 373, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 373, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 373, + 128 + ], + "type": "text", + "content": "2.2 NeRF-based Face Portrait Reconstruction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 135, + 482, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 135, + 482, + 315 + ], + "spans": [ + { + "bbox": [ + 130, + 135, + 482, + 315 + ], + "type": "text", + "content": "It is a natural idea to model human faces with NeRFs in a 3D aware manner. However, vanilla NeRFs struggle to handle dynamic objects, and are unsuitable for animating human faces which have diverse poses and expressions. There mainly exist two series of solutions for this problem. The former [14] is to condition the original Radiance Field with additional global control signals, e.g. expression parameters. The latter solve this problem by learning an additional deformation field [26, 28] to warp the 3D points. NerFace [8] is the first approach to reconstruct animatable face avatars with NeRFs conditioned on 3DMM expression parameters. HeadNeRF [13], NOFA [41] and OTAvatar [23] utilize GAN inversion to project onto the latent space of a 3D aware GAN [4] and drive the generated avatar through different latent codes. HiDe-NeRF [18] utilizes Projected Normalized Coordinate Code (PNCC) [10] as a global expression control indicator, and drives the face image with high fidelity. Comparing with the above mentioned methods, driving an arbitrary head with speech signal directly through NeRFs is a more challenging problem and is rarely explored." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 331, + 202, + 344 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 331, + 202, + 344 + ], + "spans": [ + { + "bbox": [ + 132, + 331, + 202, + 344 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 354, + 482, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 559 + ], + "type": "text", + "content": "The overall pipeline of our S" + }, + { + "bbox": [ + 130, + 354, + 482, + 559 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 130, + 354, + 482, + 559 + ], + "type": "text", + "content": "D-NeRF is shown in Fig.2. The 3D positions of the talking head can be derived through the given head pose of the target image " + }, + { + "bbox": [ + 130, + 354, + 482, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{tar}" + }, + { + "bbox": [ + 130, + 354, + 482, + 559 + ], + "type": "text", + "content": ", which is available during both training and inference time. Firstly, our Hierarchical Facial Appearance Encoder takes a single-shot image as input, and models its appearance with a deep convolutional network, where multi-resolution feature maps are extracted to construct an efficient representation i.e. multiscale appearance tri-planes, which are known to be capable of modeling several identities faithfully at the same time [4, 18]. Feature points from these planes are retrieved according to the 3D position for rendering. Then, the motion of the talking face are predicted by our Cross-modal Facial Deformation Field. It firstly calculates the correlation between an aggregated visual feature embedding and the audio signal through cross attention, then takes the result as a prior knowledge to assist the regression process of the deformation, which is used for calibrating the 3D position. A coarse face is rendered through the classic volume rendering process [24] with color and density predicted from the feature points on the tri-plane. Finally, a super-resolution module is used to refine the coarse face and yields the high fidelity portrait." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 575, + 362, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 575, + 362, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 575, + 362, + 588 + ], + "type": "text", + "content": "3.1 Hierarchical Facial Appearance Encoder" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "An effective feature representation is important to accurately model the appearance of a speaker. Initially proposed in [4], tri-plane has shown its strong representative power and is much faster comparing with the vanilla NeRFs. The Hierarchical Facial Appearance Encoder constructs this representation based on multi-scale feature maps from the one-shot source image, and yields the feature points for further volumetric rendering process." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 90, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 90, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 402, + 90, + 447, + 101 + ], + "type": "inline_equation", + "content": "\\mathrm{S}^3\\mathrm{D}" + }, + { + "bbox": [ + 402, + 90, + 447, + 101 + ], + "type": "text", + "content": " -NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": "Constructing Representative Appearance Features. We adopt a feature pyramid structure [19] to extract source features for its ability to learn multi-scale information from different resolutions, and reshape all the feature maps channelwisely to construct several orthogonal planes as in [4]. To be more concrete, given the source image " + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{src}" + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": ", feature maps with different resolutions " + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{D}^0,\\mathbf{D}^1,\\mathbf{D}^2,\\mathbf{D}^3\\}" + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": " are extracted through several convolutional downsample blocks " + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "inline_equation", + "content": "Down^{i}(.)" + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "inline_equation", + "content": "\\mathbf{D}^0 = Down^0 (\\mathbf{I}_{src})" + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "inline_equation", + "content": "\\mathbf{D}^i = Down^i (\\mathbf{D}^{i - 1})" + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": ". Then, hierarchical facial appearance features can be obtained by connecting the feature map at current " + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": "-th level with those at the higher " + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "inline_equation", + "content": "(i + 1)" + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": "-th level, both of which are upsampled to the same resolution. " + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": " convolution layers are used to reduce the channel number to avoid overlarge computational cost:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 212, + 255, + 481, + 285 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 255, + 481, + 285 + ], + "spans": [ + { + "bbox": [ + 212, + 255, + 481, + 285 + ], + "type": "interline_equation", + "content": "\\mathbf {F} ^ {i} = \\left\\{ \\begin{array}{l l} \\operatorname {C o n v} ^ {i} ([ U p ^ {i} (\\mathbf {F} ^ {i + 1}), \\mathbf {D} ^ {i} ]), & i = 0, 1, 2 \\\\ \\mathbf {D} ^ {i}, & i = 3 \\end{array} \\right. \\tag {1}", + "image_path": "a786b701e73c499c286d3b4d179d7ad73d516e9056424d63d36ddcf1790d026d.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 293, + 482, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 400 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 400 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 293, + 482, + 400 + ], + "type": "inline_equation", + "content": "Up^{i}(.)" + }, + { + "bbox": [ + 130, + 293, + 482, + 400 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 130, + 293, + 482, + 400 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 293, + 482, + 400 + ], + "type": "text", + "content": "-th upsampling convolution layer, and " + }, + { + "bbox": [ + 130, + 293, + 482, + 400 + ], + "type": "inline_equation", + "content": "Conv^{i}(.)" + }, + { + "bbox": [ + 130, + 293, + 482, + 400 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 130, + 293, + 482, + 400 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 293, + 482, + 400 + ], + "type": "text", + "content": "-th " + }, + { + "bbox": [ + 130, + 293, + 482, + 400 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 130, + 293, + 482, + 400 + ], + "type": "text", + "content": " convolution layer mentioned above. [...] denotes channelwise concatenation. Afterwards, the feature map at " + }, + { + "bbox": [ + 130, + 293, + 482, + 400 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 293, + 482, + 400 + ], + "type": "text", + "content": "-th level " + }, + { + "bbox": [ + 130, + 293, + 482, + 400 + ], + "type": "inline_equation", + "content": "F^{i}" + }, + { + "bbox": [ + 130, + 293, + 482, + 400 + ], + "type": "text", + "content": " has a shape of " + }, + { + "bbox": [ + 130, + 293, + 482, + 400 + ], + "type": "inline_equation", + "content": "B \\times 3 \\times C^{i} \\times H^{i} \\times W^{i}" + }, + { + "bbox": [ + 130, + 293, + 482, + 400 + ], + "type": "text", + "content": ", and is reshaped and splitted into three orthogonal sub feature maps, each of which has a shape of " + }, + { + "bbox": [ + 130, + 293, + 482, + 400 + ], + "type": "inline_equation", + "content": "B \\times C^{i} \\times H^{i} \\times W^{i}" + }, + { + "bbox": [ + 130, + 293, + 482, + 400 + ], + "type": "text", + "content": ", representing coordinates in the three-dimensional space, i.e. plane " + }, + { + "bbox": [ + 130, + 293, + 482, + 400 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{xy}" + }, + { + "bbox": [ + 130, + 293, + 482, + 400 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 293, + 482, + 400 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{yz}" + }, + { + "bbox": [ + 130, + 293, + 482, + 400 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 293, + 482, + 400 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{xz}" + }, + { + "bbox": [ + 130, + 293, + 482, + 400 + ], + "type": "text", + "content": " respectively. " + }, + { + "bbox": [ + 130, + 293, + 482, + 400 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 293, + 482, + 400 + ], + "type": "text", + "content": " is the batch size, and " + }, + { + "bbox": [ + 130, + 293, + 482, + 400 + ], + "type": "inline_equation", + "content": "C^{i}" + }, + { + "bbox": [ + 130, + 293, + 482, + 400 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 293, + 482, + 400 + ], + "type": "inline_equation", + "content": "H^{i}" + }, + { + "bbox": [ + 130, + 293, + 482, + 400 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 293, + 482, + 400 + ], + "type": "inline_equation", + "content": "W^{i}" + }, + { + "bbox": [ + 130, + 293, + 482, + 400 + ], + "type": "text", + "content": " are the channel number, the height and width at " + }, + { + "bbox": [ + 130, + 293, + 482, + 400 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 293, + 482, + 400 + ], + "type": "text", + "content": "-th level. Since this tri-plane representation is obtained from multiple identities, it is capable of fitting an unseen person during inference time." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 401, + 482, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 401, + 482, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 482, + 534 + ], + "type": "text", + "content": "Extracting Features for Neural Rendering. To render the target image " + }, + { + "bbox": [ + 130, + 401, + 482, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{tar}" + }, + { + "bbox": [ + 130, + 401, + 482, + 534 + ], + "type": "text", + "content": ", which may have different head pose with the source image, we need to find the correspondence between the position in the target image space and the source feature tri-plane space. This can be realized through camera transformation between the source pose " + }, + { + "bbox": [ + 130, + 401, + 482, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{src}" + }, + { + "bbox": [ + 130, + 401, + 482, + 534 + ], + "type": "text", + "content": " and the target pose " + }, + { + "bbox": [ + 130, + 401, + 482, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{tar}" + }, + { + "bbox": [ + 130, + 401, + 482, + 534 + ], + "type": "text", + "content": ", together with bilinear interpolation. Concretely, given a 3D point " + }, + { + "bbox": [ + 130, + 401, + 482, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 401, + 482, + 534 + ], + "type": "text", + "content": " on a ray " + }, + { + "bbox": [ + 130, + 401, + 482, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{r}(t)" + }, + { + "bbox": [ + 130, + 401, + 482, + 534 + ], + "type": "text", + "content": " constructed with target pose " + }, + { + "bbox": [ + 130, + 401, + 482, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{tar} = \\{\\mathbf{R}_{tar},\\mathbf{T}_{tar}\\}" + }, + { + "bbox": [ + 130, + 401, + 482, + 534 + ], + "type": "text", + "content": ", it is firstly calibrated by the deformation predicted from the Cross-modal Facial Deformation Field (Sec.3.2) to get the deformed point " + }, + { + "bbox": [ + 130, + 401, + 482, + 534 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}" + }, + { + "bbox": [ + 130, + 401, + 482, + 534 + ], + "type": "text", + "content": ". Then, feature vectors " + }, + { + "bbox": [ + 130, + 401, + 482, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{xy}(\\tilde{\\mathbf{x}}),\\mathbf{F}_{yz}(\\tilde{\\mathbf{x}})" + }, + { + "bbox": [ + 130, + 401, + 482, + 534 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 401, + 482, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{xz}(\\tilde{\\mathbf{x}})" + }, + { + "bbox": [ + 130, + 401, + 482, + 534 + ], + "type": "text", + "content": " are retrieved from three orthogonal planes " + }, + { + "bbox": [ + 130, + 401, + 482, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{xy},\\mathbf{F}_{yz}" + }, + { + "bbox": [ + 130, + 401, + 482, + 534 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 401, + 482, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{xz}" + }, + { + "bbox": [ + 130, + 401, + 482, + 534 + ], + "type": "text", + "content": ", according to the pose of the source image " + }, + { + "bbox": [ + 130, + 401, + 482, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{src} = \\{\\mathbf{R}_{src},\\mathbf{T}_{src}\\}" + }, + { + "bbox": [ + 130, + 401, + 482, + 534 + ], + "type": "text", + "content": " via bilinear interpolation Interp(:.):" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 227, + 540, + 480, + 554 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 540, + 480, + 554 + ], + "spans": [ + { + "bbox": [ + 227, + 540, + 480, + 554 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {x y z} ^ {i} (\\tilde {\\mathbf {x}}) = \\operatorname {I n t e r p} \\left(\\mathbf {F} ^ {i}, \\operatorname {I n v} \\left(\\mathbf {P} _ {s r c}\\right) \\cdot \\tilde {\\mathbf {x}}\\right), \\tag {2}", + "image_path": "2a8e2b0978743ab4dd57e42fbaabe573531f24383d2a97541d939caf2391168d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 559, + 482, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 559, + 482, + 596 + ], + "spans": [ + { + "bbox": [ + 130, + 559, + 482, + 596 + ], + "type": "inline_equation", + "content": "Inv(\\mathbf{P}_{src})" + }, + { + "bbox": [ + 130, + 559, + 482, + 596 + ], + "type": "text", + "content": " is the inverse source pose which transforms 3D positions onto the source tri-plane space. These extracted feature vectors are further used for generating the talking head image through volume rendering." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 611, + 353, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 611, + 353, + 623 + ], + "spans": [ + { + "bbox": [ + 131, + 611, + 353, + 623 + ], + "type": "text", + "content": "3.2 Cross-modal Facial Deformation Field" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "To animate the target speaker, a crucial step is to get the precise deformation which describe face motions driven by the speech signal. Solely predicting the deformation from audio struggles to learn the global motion of the whole face, since" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "text", + "content": "D. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 189, + 118, + 428, + 205 + ], + "blocks": [ + { + "bbox": [ + 189, + 118, + 428, + 205 + ], + "lines": [ + { + "bbox": [ + 189, + 118, + 428, + 205 + ], + "spans": [ + { + "bbox": [ + 189, + 118, + 428, + 205 + ], + "type": "image", + "image_path": "d1565c14fe61408989fd247862c40ea40430407f29470fd4ce80ba49a8093bfe.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 217, + 480, + 251 + ], + "lines": [ + { + "bbox": [ + 131, + 217, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 131, + 217, + 480, + 251 + ], + "type": "text", + "content": "Fig. 3: Results with naive deformation module (left) and our Cross-modal Facial Deformation Field (right). Lower Face regions have the largest activations in the heatmap, which denote their strongest correlations with the driven speech signal." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 261, + 482, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 261, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 261, + 482, + 392 + ], + "type": "text", + "content": "audio is a local signal which only correlates strongly with the mouth area. Thus it tends to yield blurry results (shown in Fig.3, left). For correctly modeling the global motion, structural and textural information of the whole face is required. Our Cross-modal Facial Deformation Field firstly aggregates the multi-scale visual features to a unified embedding, which contains the global information of the entire face area, then calculates the correlation score between the audio signal and this embedding through cross attention. The correlation score is used as a prior knowledge for determining the importance of motions from each part of the face, making deformation prediction more precise. As is shown in Fig.3, our Cross-modal Facial Deformation Field is free from the blurry issues because of the correctly learned motion." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 393, + 482, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 393, + 482, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 393, + 482, + 489 + ], + "type": "text", + "content": "Aggregating Multi-scale Facial Features. To learn the features which represents the whole face, all the hierarchical appearance feature maps " + }, + { + "bbox": [ + 130, + 393, + 482, + 489 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^0" + }, + { + "bbox": [ + 130, + 393, + 482, + 489 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 393, + 482, + 489 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^3" + }, + { + "bbox": [ + 130, + 393, + 482, + 489 + ], + "type": "text", + "content": " are first reshaped to the same resolution " + }, + { + "bbox": [ + 130, + 393, + 482, + 489 + ], + "type": "inline_equation", + "content": "(H_s\\times W_s)" + }, + { + "bbox": [ + 130, + 393, + 482, + 489 + ], + "type": "text", + "content": " and concatenated along the channel. Then, these features are flattened and fed into a slot-attention module [21] to exchange information from different scales: " + }, + { + "bbox": [ + 130, + 393, + 482, + 489 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{agg} = SlotAttention(\\mathbf{F}^{0\\dots 3})" + }, + { + "bbox": [ + 130, + 393, + 482, + 489 + ], + "type": "text", + "content": ". The aggregated slot feature embedding has a shape of " + }, + { + "bbox": [ + 130, + 393, + 482, + 489 + ], + "type": "inline_equation", + "content": "B\\times (H_{s}\\times W_{s})\\times D_{slot}" + }, + { + "bbox": [ + 130, + 393, + 482, + 489 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 393, + 482, + 489 + ], + "type": "inline_equation", + "content": "D_{slot}" + }, + { + "bbox": [ + 130, + 393, + 482, + 489 + ], + "type": "text", + "content": " is the slot feature dimension. This learning process makes the embedding contain rich structural and textural information of the speaker." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 490, + 482, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 490, + 482, + 550 + ], + "spans": [ + { + "bbox": [ + 130, + 490, + 482, + 550 + ], + "type": "text", + "content": "Calculating Audio-Visual Correlation Scores. The input speech signal is firstly processed with a 1D full convolutional network to get the driving audio feature " + }, + { + "bbox": [ + 130, + 490, + 482, + 550 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_{dri}" + }, + { + "bbox": [ + 130, + 490, + 482, + 550 + ], + "type": "text", + "content": ". Relevance between audio and face motion is calculated through the Multi-Head Cross-Attention (MHCA) scores with the aggregated visual embedding " + }, + { + "bbox": [ + 130, + 490, + 482, + 550 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{agg}" + }, + { + "bbox": [ + 130, + 490, + 482, + 550 + ], + "type": "text", + "content": " as query, the driving audio feature " + }, + { + "bbox": [ + 130, + 490, + 482, + 550 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_{dri}" + }, + { + "bbox": [ + 130, + 490, + 482, + 550 + ], + "type": "text", + "content": " as key and value:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 194, + 560, + 481, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 560, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 194, + 560, + 481, + 620 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {F} _ {c m} = M H C A \\left(\\mathbf {F} _ {a g g}, \\mathbf {a} _ {d r i}\\right) \\\\ = \\operatorname {S o f t m a x} \\left[ \\frac {\\mathbf {F} _ {a g g} \\boldsymbol {W} ^ {Q} \\left(\\mathbf {a} _ {d r i} \\boldsymbol {W} ^ {K}\\right) ^ {T}}{\\sqrt {d}} \\right] \\mathbf {a} _ {d r i} \\boldsymbol {W} ^ {V}, \\tag {3} \\\\ \\end{array}", + "image_path": "f40b860f161b37931380b6148ca8d45e3bb08639f71683626dc915dd74b2924a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{W}^Q, \\pmb{W}^K, \\pmb{W}^V" + }, + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "text", + "content": " are the projection matrices with hidden dimension " + }, + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "text", + "content": ", respectively. Cross-attention scores between the speech signal and the whole face feature representation indicate the relevance of each face area with the speech" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 90, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 90, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 90, + 447, + 100 + ], + "type": "inline_equation", + "content": "\\mathrm{S}^3\\mathrm{D}" + }, + { + "bbox": [ + 402, + 90, + 447, + 100 + ], + "type": "text", + "content": " -NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 118, + 479, + 311 + ], + "blocks": [ + { + "bbox": [ + 138, + 118, + 479, + 311 + ], + "lines": [ + { + "bbox": [ + 138, + 118, + 479, + 311 + ], + "spans": [ + { + "bbox": [ + 138, + 118, + 479, + 311 + ], + "type": "image", + "image_path": "b03b76721c2e183dfc3b9d08d1145499c23c4277c708f1eb3ebe8c88a2843164.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 322, + 482, + 368 + ], + "lines": [ + { + "bbox": [ + 130, + 322, + 482, + 368 + ], + "spans": [ + { + "bbox": [ + 130, + 322, + 482, + 368 + ], + "type": "text", + "content": "Fig. 4: Qualitative comparison with single-shot methods. Our " + }, + { + "bbox": [ + 130, + 322, + 482, + 368 + ], + "type": "inline_equation", + "content": "\\mathrm{S}^3\\mathrm{D}" + }, + { + "bbox": [ + 130, + 322, + 482, + 368 + ], + "type": "text", + "content": "-NeRF yields the most correct lip shapes and the clearest teeth. Note that different methods adopt different face alignment tools, and the ground truth row demonstrates the raw image without alignment, so the face poses from different methods are slightly different." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 379, + 482, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 379, + 482, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 379, + 482, + 450 + ], + "type": "text", + "content": "input, and can serve as a strong prior for modeling the dynamic talking face in a region-aware manner. We observed that the mouth area of the speaker, which moves with the audio signal has the largest deformation scales. Also shown in Fig.3, heatmaps are calculated by averaging " + }, + { + "bbox": [ + 130, + 379, + 482, + 450 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{cm}" + }, + { + "bbox": [ + 130, + 379, + 482, + 450 + ], + "type": "text", + "content": " channelwisely, and have shown larger activations on lower face regions which contains not only mouth area but also the muscles of the lower face." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 451, + 482, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 451, + 482, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 451, + 482, + 499 + ], + "type": "text", + "content": "Predicting Facial Deformation. The facial deformation prediction module " + }, + { + "bbox": [ + 130, + 451, + 482, + 499 + ], + "type": "inline_equation", + "content": "Deform(\\ldots)" + }, + { + "bbox": [ + 130, + 451, + 482, + 499 + ], + "type": "text", + "content": " has a U-Net [30] like architecture, with the original 3D position " + }, + { + "bbox": [ + 130, + 451, + 482, + 499 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 451, + 482, + 499 + ], + "type": "text", + "content": ", the cross-modal correlation score " + }, + { + "bbox": [ + 130, + 451, + 482, + 499 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{cm}" + }, + { + "bbox": [ + 130, + 451, + 482, + 499 + ], + "type": "text", + "content": ", and the driving audio feature " + }, + { + "bbox": [ + 130, + 451, + 482, + 499 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_{dri}" + }, + { + "bbox": [ + 130, + 451, + 482, + 499 + ], + "type": "text", + "content": " as inputs, this deformation prediction process can be formalized as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 241, + 504, + 371, + 517 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 504, + 371, + 517 + ], + "spans": [ + { + "bbox": [ + 241, + 504, + 371, + 517 + ], + "type": "interline_equation", + "content": "\\Delta \\mathbf {x} = D e f o r m (\\mathbf {x}, \\mathbf {a} _ {d r i}, \\mathbf {F} _ {c m}),", + "image_path": "dc34c0d9c13c16cbc68f38a403815f5565a0d8dc9b01ec1b596a27eebaa3ee30.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 251, + 517, + 480, + 530 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 517, + 480, + 530 + ], + "spans": [ + { + "bbox": [ + 251, + 517, + 480, + 530 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {x}} = \\mathbf {x} + \\Delta \\mathbf {x}. \\tag {4}", + "image_path": "b49359eb5b316b077585aeb55b08be4b26a567681554cac0422accd514e24d3d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 537, + 482, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 537, + 482, + 563 + ], + "spans": [ + { + "bbox": [ + 130, + 537, + 482, + 563 + ], + "type": "text", + "content": "With the calibrated 3D points, source features from the tri-plane can be extracted faithfully through Eq.2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 577, + 359, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 577, + 359, + 590 + ], + "spans": [ + { + "bbox": [ + 131, + 577, + 359, + 590 + ], + "type": "text", + "content": "3.3 High Fidelity Talking Head Generation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 594, + 482, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 639 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 639 + ], + "type": "text", + "content": "Volume Rendering. To generate an image with volume rendering, given the feature vector " + }, + { + "bbox": [ + 130, + 594, + 482, + 639 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{xyz}^{i}(\\tilde{\\mathbf{x}})" + }, + { + "bbox": [ + 130, + 594, + 482, + 639 + ], + "type": "text", + "content": " retrieved from the tri-plane, a tri-plane decoder takes it as input and predicts the color density " + }, + { + "bbox": [ + 130, + 594, + 482, + 639 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 130, + 594, + 482, + 639 + ], + "type": "text", + "content": " and occupancy " + }, + { + "bbox": [ + 130, + 594, + 482, + 639 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 594, + 482, + 639 + ], + "type": "text", + "content": " at that position. The final color of a pixel " + }, + { + "bbox": [ + 130, + 594, + 482, + 639 + ], + "type": "inline_equation", + "content": "\\mathbf{C}(\\mathbf{r})" + }, + { + "bbox": [ + 130, + 594, + 482, + 639 + ], + "type": "text", + "content": " along the ray is calculated through:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 228, + 643, + 481, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 643, + 481, + 669 + ], + "spans": [ + { + "bbox": [ + 228, + 643, + 481, + 669 + ], + "type": "interline_equation", + "content": "\\mathbf {C} (\\mathbf {r}) = \\int_ {t _ {n}} ^ {t _ {f}} T (t) \\sigma (\\mathbf {r} (t), \\mathbf {a}) \\mathbf {c} (\\mathbf {r} (t), \\mathbf {a}) d t, \\tag {5}", + "image_path": "394cf99cfd374781d2e30bdd6c3e04e36a3f4a9767871b211d6037beee1180e0.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "D. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 138, + 146, + 479, + 299 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "type": "text", + "content": "Table 1: Quantitative comparison with single-shot methods, The best and the second best results are emphasized." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 146, + 479, + 299 + ], + "lines": [ + { + "bbox": [ + 138, + 146, + 479, + 299 + ], + "spans": [ + { + "bbox": [ + 138, + 146, + 479, + 299 + ], + "type": "table", + "html": "
DatasetMetricGTWav2LipMakeItTalkPC-AVSStyleTalkSadTalker\\( S^3 \\)D-NeRF
HDTFSSIM ↑10.7490.5590.5820.7760.6780.819
LPIPS ↓00.3320.4220.4110.2890.3860.258
F-LMD ↓03.5826.8463.9253.6234.6212.799
M-LMD ↓03.6526.1523.9123.2733.6542.929
CPBD ↑0.2840.1950.1390.2060.2340.1920.263
Sync ↑8.1448.3654.8135.0966.1886.4736.514
MEADSSIM ↑10.8290.6260.6730.8500.7550.845
LPIPS ↓00.2120.3670.3590.1990.2230.161
F-LMD ↓02.1266.0818.3722.6223.8352.014
M-LMD ↓03.3776.7905.1923.0613.6312.921
CPBD ↑0.1760.1010.0650.0680.1450.0810.152
Sync ↑7.7027.7135.1435.4666.0575.7076.768
", + "image_path": "bf3c4215eb4643f96b7d4b120951b25a5be6713aa1a221d3035b7ce957f261ae.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 325, + 480, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 480, + 365 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 325, + 480, + 365 + ], + "type": "inline_equation", + "content": "T(t) = \\exp \\left(-\\int_{t_n}^t\\sigma (s)ds\\right)" + }, + { + "bbox": [ + 132, + 325, + 480, + 365 + ], + "type": "text", + "content": " denotes for the accumulated transmittance along the ray from " + }, + { + "bbox": [ + 132, + 325, + 480, + 365 + ], + "type": "inline_equation", + "content": "t_n" + }, + { + "bbox": [ + 132, + 325, + 480, + 365 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 132, + 325, + 480, + 365 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 132, + 325, + 480, + 365 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 325, + 480, + 365 + ], + "type": "inline_equation", + "content": "t_n" + }, + { + "bbox": [ + 132, + 325, + 480, + 365 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 325, + 480, + 365 + ], + "type": "inline_equation", + "content": "t_f" + }, + { + "bbox": [ + 132, + 325, + 480, + 365 + ], + "type": "text", + "content": " are the lower and the upper bound of depth." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 367, + 481, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 367, + 481, + 498 + ], + "spans": [ + { + "bbox": [ + 132, + 367, + 481, + 498 + ], + "type": "text", + "content": "Coarse-to-fine Generation Strategy. Talking head frames contain various backgrounds and complex outer face textures which are hard for a NeRF to render at the same time. To decline the difficulty for training, we adopt a coarse-to-fine strategy. Specifically, the inner face part " + }, + { + "bbox": [ + 132, + 367, + 481, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^{masked}" + }, + { + "bbox": [ + 132, + 367, + 481, + 498 + ], + "type": "text", + "content": " of an arbitrary frame " + }, + { + "bbox": [ + 132, + 367, + 481, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 132, + 367, + 481, + 498 + ], + "type": "text", + "content": " is decoupled in advance. During training, the inner face is rendered by the NeRF model, and the synthesized image is denoted as " + }, + { + "bbox": [ + 132, + 367, + 481, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{gen}^{coar}" + }, + { + "bbox": [ + 132, + 367, + 481, + 498 + ], + "type": "text", + "content": ", while the fine texture details of the outer face and the background is complemented by an additional super-resolution module. We utilize Deep3DFaceRecon [7] to generate a 3D-aware face segmentation mask for inner-outer face area separation. This coarse-to-fine generation strategy guarantees that the NeRF can focus on the important face area." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 500, + 481, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 500, + 481, + 562 + ], + "spans": [ + { + "bbox": [ + 132, + 500, + 481, + 562 + ], + "type": "text", + "content": "Super Resolution Module. The super resolution module takes the coarse output face " + }, + { + "bbox": [ + 132, + 500, + 481, + 562 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{gen}^{coar}" + }, + { + "bbox": [ + 132, + 500, + 481, + 562 + ], + "type": "text", + "content": " and the full source face image " + }, + { + "bbox": [ + 132, + 500, + 481, + 562 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{src}^{full}" + }, + { + "bbox": [ + 132, + 500, + 481, + 562 + ], + "type": "text", + "content": " as input, and generates the talking head frame " + }, + { + "bbox": [ + 132, + 500, + 481, + 562 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{fine}^{gen}" + }, + { + "bbox": [ + 132, + 500, + 481, + 562 + ], + "type": "text", + "content": " with background and more details. The whole super resolution process can be written as " + }, + { + "bbox": [ + 132, + 500, + 481, + 562 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{gen}^{fine} = \\text{Supres}(\\mathbf{I}_{gen}^{coar}, \\mathbf{I}_{src}^{full})" + }, + { + "bbox": [ + 132, + 500, + 481, + 562 + ], + "type": "text", + "content": ". More details are shown in the appendix." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 588, + 275, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 588, + 275, + 601 + ], + "spans": [ + { + "bbox": [ + 133, + 588, + 275, + 601 + ], + "type": "text", + "content": "3.4 Lip-sync Discriminator" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 617, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 617, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 617, + 481, + 665 + ], + "type": "text", + "content": "Our lip sync discriminator has a visual branch and an audio branch. The former takes " + }, + { + "bbox": [ + 132, + 617, + 481, + 665 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 132, + 617, + 481, + 665 + ], + "type": "text", + "content": " frames of the lip area as input and outputs their embedding " + }, + { + "bbox": [ + 132, + 617, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_l" + }, + { + "bbox": [ + 132, + 617, + 481, + 665 + ], + "type": "text", + "content": ", while the latter gets an audio clip and outputs its corresponding feature embedding " + }, + { + "bbox": [ + 132, + 617, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_a" + }, + { + "bbox": [ + 132, + 617, + 481, + 665 + ], + "type": "text", + "content": ". Then, the synchronization level between the audio and the video sequence" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 90, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 90, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 90, + 447, + 100 + ], + "type": "inline_equation", + "content": "\\mathrm{S}^3\\mathrm{D}" + }, + { + "bbox": [ + 402, + 90, + 447, + 100 + ], + "type": "text", + "content": " -NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 436, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 436, + 129 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 436, + 129 + ], + "type": "text", + "content": "can be judged by calculating the cosine similarity between " + }, + { + "bbox": [ + 130, + 116, + 436, + 129 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_l" + }, + { + "bbox": [ + 130, + 116, + 436, + 129 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 436, + 129 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_a" + }, + { + "bbox": [ + 130, + 116, + 436, + 129 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 229, + 137, + 481, + 160 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 229, + 137, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 229, + 137, + 481, + 160 + ], + "type": "interline_equation", + "content": "\\cos (\\mathbf {e} _ {l}, \\mathbf {e} _ {a}) = \\frac {\\mathbf {e} _ {l} \\cdot \\mathbf {e} _ {a}}{\\operatorname* {m a x} \\left(\\| \\mathbf {e} _ {l} \\| _ {2} \\cdot \\| \\mathbf {e} _ {a} \\| _ {2} , \\epsilon\\right)}, \\tag {6}", + "image_path": "2446b0ed71377ebeb32a009e3ab2fce54915cd5dc81559a5706b2feae3ba8893.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 167, + 479, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 167, + 479, + 216 + ], + "spans": [ + { + "bbox": [ + 130, + 167, + 479, + 216 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 167, + 479, + 216 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 167, + 479, + 216 + ], + "type": "text", + "content": " is a margin which is set to 0. Different from that used in Wav2Lip [27], our lip sync discriminator is trained on a larger mixed dataset containing HDTF [44] and LRS2 [1], with a single contrastive triplet loss for more distinctive feature representations:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 205, + 224, + 481, + 255 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 224, + 481, + 255 + ], + "spans": [ + { + "bbox": [ + 205, + 224, + 481, + 255 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {s y n c}} ^ {\\text {d i s}} = \\max \\left(0, \\eta + \\cos \\left(\\mathbf {e} _ {l}, \\mathbf {e} _ {a} ^ {+}\\right) - \\cos \\left(\\mathbf {e} _ {l}, \\mathbf {e} _ {a} ^ {-}\\right)\\right) \\tag {7} \\\\ + \\max \\left(0, \\eta + \\cos \\left(\\mathbf {e} _ {l} ^ {+}, \\mathbf {e} _ {a}\\right) - \\cos \\left(\\mathbf {e} _ {l} ^ {-}, \\mathbf {e} _ {a}\\right)\\right). \\\\ \\end{array}", + "image_path": "99bf9e4f689e852015efe79434a8b4107f40ceff4fde260c629d029c7b45668a.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 263, + 480, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 263, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 263, + 480, + 335 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_a^+" + }, + { + "bbox": [ + 130, + 263, + 480, + 335 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 263, + 480, + 335 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_a^-" + }, + { + "bbox": [ + 130, + 263, + 480, + 335 + ], + "type": "text", + "content": " regard the positive and the negative audio embedding according to the lip sequence respectively, so as " + }, + { + "bbox": [ + 130, + 263, + 480, + 335 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_l^+" + }, + { + "bbox": [ + 130, + 263, + 480, + 335 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 263, + 480, + 335 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_l^-" + }, + { + "bbox": [ + 130, + 263, + 480, + 335 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 263, + 480, + 335 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 130, + 263, + 480, + 335 + ], + "type": "text", + "content": " is a margin set to 0.5. While training our model, the lip-sync discriminator is frozen and " + }, + { + "bbox": [ + 130, + 263, + 480, + 335 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 263, + 480, + 335 + ], + "type": "text", + "content": " frames of the same person are generated once a time. The lip regions are cropped through the pre-extracted bounding boxes of the ground-truth. The loss used for S" + }, + { + "bbox": [ + 130, + 263, + 480, + 335 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 130, + 263, + 480, + 335 + ], + "type": "text", + "content": "D-NeRF training can be formulated as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 258, + 345, + 480, + 359 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 345, + 480, + 359 + ], + "spans": [ + { + "bbox": [ + 258, + 345, + 480, + 359 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s y n c}} ^ {\\text {g e n}} = \\cos \\left(\\mathbf {e} _ {l} ^ {\\text {g e n}}, \\mathbf {e} _ {g t}\\right). \\tag {8}", + "image_path": "5a3db9a39a3fcb0cbc7388b3a8ef404297039eace1ecfa80e5cd2d3bb80da763.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 366, + 479, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 366, + 479, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 366, + 479, + 402 + ], + "type": "text", + "content": "while " + }, + { + "bbox": [ + 130, + 366, + 479, + 402 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_l^{gen}" + }, + { + "bbox": [ + 130, + 366, + 479, + 402 + ], + "type": "text", + "content": " is the embedding of the generated lip sequence and " + }, + { + "bbox": [ + 130, + 366, + 479, + 402 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_{gt}" + }, + { + "bbox": [ + 130, + 366, + 479, + 402 + ], + "type": "text", + "content": " is the embedding of the ground-truth audio. By penalizing the out-of-sync audio-visual pair, more precise mouth shapes can be achieved." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 420, + 235, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 420, + 235, + 430 + ], + "spans": [ + { + "bbox": [ + 131, + 420, + 235, + 430 + ], + "type": "text", + "content": "3.5 Loss Functions" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 439, + 479, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 439, + 479, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 439, + 479, + 511 + ], + "type": "text", + "content": "In addition to the lip synchronization loss " + }, + { + "bbox": [ + 130, + 439, + 479, + 511 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{sync}^{gen}" + }, + { + "bbox": [ + 130, + 439, + 479, + 511 + ], + "type": "text", + "content": ", we use a pixelwise " + }, + { + "bbox": [ + 130, + 439, + 479, + 511 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 130, + 439, + 479, + 511 + ], + "type": "text", + "content": " reconstruction loss " + }, + { + "bbox": [ + 130, + 439, + 479, + 511 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{pix}" + }, + { + "bbox": [ + 130, + 439, + 479, + 511 + ], + "type": "text", + "content": ", a feature level perceptual loss " + }, + { + "bbox": [ + 130, + 439, + 479, + 511 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{per}" + }, + { + "bbox": [ + 130, + 439, + 479, + 511 + ], + "type": "text", + "content": " to match the generated face with the target face, and a GAN loss " + }, + { + "bbox": [ + 130, + 439, + 479, + 511 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{adv}" + }, + { + "bbox": [ + 130, + 439, + 479, + 511 + ], + "type": "text", + "content": " with the architecture of StyleGAN discriminator [15]. All the losses are imposed on both the coarse and the fine faces. An " + }, + { + "bbox": [ + 130, + 439, + 479, + 511 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 130, + 439, + 479, + 511 + ], + "type": "text", + "content": " loss " + }, + { + "bbox": [ + 130, + 439, + 479, + 511 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{deform}" + }, + { + "bbox": [ + 130, + 439, + 479, + 511 + ], + "type": "text", + "content": " is used to regularize the prediction deformation. Our S" + }, + { + "bbox": [ + 130, + 439, + 479, + 511 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 130, + 439, + 479, + 511 + ], + "type": "text", + "content": "D-NeRF is finally trained with the weighted sum of the above losses as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 160, + 521, + 480, + 534 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 521, + 480, + 534 + ], + "spans": [ + { + "bbox": [ + 160, + 521, + 480, + 534 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {p i x} + \\lambda_ {p e r} \\mathcal {L} _ {p e r} + \\lambda_ {a d v} \\mathcal {L} _ {a d v} + \\lambda_ {\\text {s y n c}} ^ {\\text {g e n}} \\mathcal {L} _ {\\text {s y n c}} ^ {\\text {g e n}} + \\lambda_ {\\text {d e f o r m}} \\mathcal {L} _ {\\text {d e f o r m}}, \\tag {9}", + "image_path": "0df6d1fe28cc9ed74086cc8d8adc245c02947c3211633dcd38bae9f5d0ce0f67.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 543, + 383, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 543, + 383, + 556 + ], + "spans": [ + { + "bbox": [ + 131, + 543, + 383, + 556 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 131, + 543, + 383, + 556 + ], + "type": "inline_equation", + "content": "\\lambda_{per} = 0.01" + }, + { + "bbox": [ + 131, + 543, + 383, + 556 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 543, + 383, + 556 + ], + "type": "inline_equation", + "content": "\\lambda_{adv} = 1" + }, + { + "bbox": [ + 131, + 543, + 383, + 556 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 543, + 383, + 556 + ], + "type": "inline_equation", + "content": "\\lambda_{sync}^{gen} = 0.5" + }, + { + "bbox": [ + 131, + 543, + 383, + 556 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 543, + 383, + 556 + ], + "type": "inline_equation", + "content": "\\lambda_{deform} = 0.001" + }, + { + "bbox": [ + 131, + 543, + 383, + 556 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 573, + 230, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 573, + 230, + 586 + ], + "spans": [ + { + "bbox": [ + 131, + 573, + 230, + 586 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 597, + 273, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 597, + 273, + 609 + ], + "spans": [ + { + "bbox": [ + 131, + 597, + 273, + 609 + ], + "type": "text", + "content": "4.1 Experimental Settings" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 617, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 480, + 666 + ], + "type": "text", + "content": "Datasets. Our " + }, + { + "bbox": [ + 130, + 617, + 480, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{S}^3\\mathrm{D}" + }, + { + "bbox": [ + 130, + 617, + 480, + 666 + ], + "type": "text", + "content": "-NeRF is trained on HDTF [44] training split, where each video ranges from tens of seconds to more than ten minutes in length, and evaluated on HDTF testing split and MEAD testing split [35] for comparison with single-shot methods. There is no overlap between the training and the testing" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "text", + "content": "D. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 116, + 304, + 251 + ], + "blocks": [ + { + "bbox": [ + 138, + 116, + 304, + 251 + ], + "lines": [ + { + "bbox": [ + 138, + 116, + 304, + 251 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 304, + 251 + ], + "type": "image", + "image_path": "91fc46d10e0bf4b0169710d39b311fff04eabbbc5bec46bbd8644236ad3d00bd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 219, + 253, + 228, + 262 + ], + "lines": [ + { + "bbox": [ + 219, + 253, + 228, + 262 + ], + "spans": [ + { + "bbox": [ + 219, + 253, + 228, + 262 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 309, + 126, + 477, + 247 + ], + "blocks": [ + { + "bbox": [ + 309, + 126, + 477, + 247 + ], + "lines": [ + { + "bbox": [ + 309, + 126, + 477, + 247 + ], + "spans": [ + { + "bbox": [ + 309, + 126, + 477, + 247 + ], + "type": "image", + "image_path": "f3707c90a25590976e50d4b9179ccf5d16531148b246a43b0f309357451eb7dd.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 392, + 252, + 402, + 262 + ], + "lines": [ + { + "bbox": [ + 392, + 252, + 402, + 262 + ], + "spans": [ + { + "bbox": [ + 392, + 252, + 402, + 262 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 270, + 482, + 316 + ], + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 316 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 316 + ], + "type": "text", + "content": "Fig. 5: Left (a): Qualitative comparison with NeRF-based methods, when encountering a new identity, our S" + }, + { + "bbox": [ + 130, + 270, + 482, + 316 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 130, + 270, + 482, + 316 + ], + "type": "text", + "content": "D-NeRF successfully synthesizes a portrait faithfully without any retraining. Right (b): High fidelity generation results with multi-view consistency. Ground truth are the images with front view. Most of the face details are preserved." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 325, + 480, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 325, + 480, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 325, + 480, + 361 + ], + "type": "text", + "content": "split. For comparison with NeRF-based methods, we use the videos provided by AD-NeRF [11] and GeneFace [39]. The video sampling rate is set to 25 FPS. All the faces are cropped and aligned to a fix resolution of " + }, + { + "bbox": [ + 130, + 325, + 480, + 361 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 130, + 325, + 480, + 361 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 361, + 482, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 482, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 482, + 422 + ], + "type": "text", + "content": "Implementation Details. The whole proposed framework is implemented by PyTorch and can be trained end-to-end. It takes about two days to train our S" + }, + { + "bbox": [ + 130, + 361, + 482, + 422 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 130, + 361, + 482, + 422 + ], + "type": "text", + "content": "D-NeRF on 4 NVIDIA-A100 GPUs. The sample rate of the speech signal is set to 16000. We use Wav2Vec2.0 [2] to extract the driving audio features. More details and the user study can be found in the appendix." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 441, + 265, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 441, + 265, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 441, + 265, + 453 + ], + "type": "text", + "content": "4.2 Quantitative Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 461, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 545 + ], + "type": "text", + "content": "Several widely used metrics are adopted to carry out our quantitative evaluation. We use SSIM and LPIPS [42] to measure the similarity of the generated speaker with the ground truth in pixel level and feature level respectively. To compare the identity fidelity, we use Facial Landmark Distance (F-LMD). To compare the clear extent of the images, we use Cumulative Probability of Blur Detection (CPBD). To evaluate the audio-lip synchronization, we use the SyncNet confidence score (Sync) [6] and Mouth Landmark Distance (M-LMD) [5]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Comparison with Single-Shot Methods. Our competitors include Wav2Lip [27], MakeiTalk [46], PC-AVS [45], SadTalker [43] and StyleTalk [22]. For the above methods, we take their provided checkpoints and strictly follow the test protocol of current single-shot methods to conduct our evaluation. MakeiTalk has shown the worst image quality and lip shape. Wav2Lip also uses a pretrained lip-sync expert as supervision, and it has larger batch size and its model is simpler. Thus, it achieves a SyncNet score which is even better than ground truth. However, Wav2Lip tends to generate blurry faces, resulting in an inferior face fidelity. StyleTalk and SadTalker have shown competitive performance on image quality. Our S" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "D-NeRF achieves the second-best lip-sync result, meanwhile sur" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 90, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 90, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 90, + 447, + 100 + ], + "type": "inline_equation", + "content": "\\mathrm{S}^3\\mathrm{D}" + }, + { + "bbox": [ + 402, + 90, + 447, + 100 + ], + "type": "text", + "content": " -NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 185, + 157, + 430, + 264 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "type": "text", + "content": "Table 2: Quantitative comparison with NeRF-based methods. The best and the second best results are highlighted. Our S" + }, + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "type": "text", + "content": "D-NeRF holds competitive image quality and audio-visual consistency, meanwhile the best generalize ability." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 185, + 157, + 430, + 264 + ], + "lines": [ + { + "bbox": [ + 185, + 157, + 430, + 264 + ], + "spans": [ + { + "bbox": [ + 185, + 157, + 430, + 264 + ], + "type": "table", + "html": "
MetricGTGeneFaceAD-NeRFER-NeRFS3D-NeRF
SSIM ↑10.8280.8460.8840.852
LPIPS ↓00.0980.0870.0680.104
M-LMD ↓02.4361.9821.6591.493
F-LMD ↓01.8981.7151.1582.019
CPBD ↑0.2780.2070.2200.2390.244
Sync ↑8.9705.3665.6266.7817.118
FPS ↑N/A0.130.13299.5
Fit Time / h ↓N/A42364.5<0.01
", + "image_path": "b6c1f7784bc59d48505edcd76916f74256c1798ab43e446cd7d8ee1b1e04373f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 284, + 481, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 284, + 481, + 356 + ], + "spans": [ + { + "bbox": [ + 130, + 284, + 481, + 356 + ], + "type": "text", + "content": "passes the competitors significantly on quality metrics at both image level and feature level. Moreover, our method has the lowest face landmark distance and mouth landmark distance, which indicates that our S" + }, + { + "bbox": [ + 130, + 284, + 481, + 356 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 130, + 284, + 481, + 356 + ], + "type": "text", + "content": "D-NeRF keep facial and lip structure better than other methods. On MEAD dataset where all the identities are unseen before, our S" + }, + { + "bbox": [ + 130, + 284, + 481, + 356 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 130, + 284, + 481, + 356 + ], + "type": "text", + "content": "D-NeRF still achieves competitive performance, which demonstrates the strong generalization ability of our approach." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 356, + 482, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 356, + 482, + 583 + ], + "spans": [ + { + "bbox": [ + 130, + 356, + 482, + 583 + ], + "type": "text", + "content": "Comparison with NeRF-based Methods. We also compare S" + }, + { + "bbox": [ + 130, + 356, + 482, + 583 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 130, + 356, + 482, + 583 + ], + "type": "text", + "content": "D-NeRF with current identity specific NeRF-based methods. Three start-of-the-art approaches AD-NeRF [11] GeneFace [39] and ER-NeRF [17] are chosen as our opponents. It is worth to note that our method has never seen the identity nor the driving audio in the evaluation set, while the person-specific NeRFs are trained on a training set which has the same identity as the evaluation set (although no image overlap). In addition to the metrics mentioned above, we also take account of the data efficiency metrics, FPS: Frame Per Second during inference, which can reflects the speed of a rendering model. Fit Time: Times needed to generate the plausible talking head videos of a new identity with a pretrained model available. ER-NeRF has shown the strong image quality with the fastest rendering speed. GeneFace needs to train a postnet for each new identity, which takes extra time. Our model is also competitive in image quality. Moreover, it has the best lip synchronization score, together with the best mouth-LMD, which indicates the superiority on the specific audio-driven task. Moreover, our model has the highest CPBD, that means we generate the clearest images. Besides, S" + }, + { + "bbox": [ + 130, + 356, + 482, + 583 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 130, + 356, + 482, + 583 + ], + "type": "text", + "content": "D-NeRF has the strongest generalization ability since it is free from retraining when coming up with an new speaker, and has shown an excellent inference speed." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 599, + 257, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 257, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 257, + 611 + ], + "type": "text", + "content": "4.3 Qualitative Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "Comparison with Single-Shot Methods. A direct comparison with other person agnostic methods are shown in Fig.4. Our S" + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "D-NeRF has shown more precise lip shapes and clearer teeth. At the same time, the detailed facial features such as skin texture, eyebrows and teeth splits are preserved by our method." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "text", + "content": "D. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 189, + 114, + 428, + 275 + ], + "blocks": [ + { + "bbox": [ + 189, + 114, + 428, + 275 + ], + "lines": [ + { + "bbox": [ + 189, + 114, + 428, + 275 + ], + "spans": [ + { + "bbox": [ + 189, + 114, + 428, + 275 + ], + "type": "image", + "image_path": "da0083a72e3022b0b2821fb7aa3e33c6d085ea7c5e4918e34691988afc041faa.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 190, + 284, + 421, + 295 + ], + "lines": [ + { + "bbox": [ + 190, + 284, + 421, + 295 + ], + "spans": [ + { + "bbox": [ + 190, + 284, + 421, + 295 + ], + "type": "text", + "content": "Fig. 6: Qualitative ablation study over key components." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 177, + 354, + 440, + 429 + ], + "blocks": [ + { + "bbox": [ + 132, + 321, + 480, + 344 + ], + "lines": [ + { + "bbox": [ + 132, + 321, + 480, + 344 + ], + "spans": [ + { + "bbox": [ + 132, + 321, + 480, + 344 + ], + "type": "text", + "content": "Table 3: Quantitative ablation study. The full S" + }, + { + "bbox": [ + 132, + 321, + 480, + 344 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 132, + 321, + 480, + 344 + ], + "type": "text", + "content": "D-NeRF achieves the best landmark distance and lip synthesis results, with a little sacrifice on image quality." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 177, + 354, + 440, + 429 + ], + "lines": [ + { + "bbox": [ + 177, + 354, + 440, + 429 + ], + "spans": [ + { + "bbox": [ + 177, + 354, + 440, + 429 + ], + "type": "table", + "html": "
MethodSSIM ↑LPIPS ↓F-LMD ↓M-LMD ↓CPBD ↑Sync ↑
ND0.7440.2783.4753.5070.1655.201
FCD0.8340.2422.8553.1960.2466.106
w/o Deform0.5560.457N/AN/A0.1400.240
w/o Lip Sync0.8010.2633.2763.3420.2555.649
S3D-NeRF0.8290.2582.7992.9290.2636.514
", + "image_path": "532fe2e3f973f971d9203da1444a0b5604bde35d5c7f87d0a3a4946d518e0dea.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 466, + 482, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 466, + 482, + 610 + ], + "spans": [ + { + "bbox": [ + 130, + 466, + 482, + 610 + ], + "type": "text", + "content": "Comparison with NeRF-based Methods. A qualitative comparison with NeRF-based methods are shown in Fig.5. AD-NeRF has shown the head-torso separation problem, which seriously harm the fidelity of the portrait (Obama face at the third row). Due to the identity specific training process, NeRF based methods can achieve excellent speaker reconstruction. Their generated face regions are very similar to the ground truth. However, on a novel identity, all the person specific NeRFs failed to generate a plausible result. On the contrary, our S" + }, + { + "bbox": [ + 130, + 466, + 482, + 610 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 130, + 466, + 482, + 610 + ], + "type": "text", + "content": "D-NeRF has shown excellent generation results. The images from our S" + }, + { + "bbox": [ + 130, + 466, + 482, + 610 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 130, + 466, + 482, + 610 + ], + "type": "text", + "content": "D-NeRF do not reach a similarity as high as NeRF-based methods, but they have more distinct facial details, each face region of the speaker are kept completely. Images generated by our method are sharper and their lip areas are much clearer." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Multi-view Generation Results. Benefited from the 3D consistency, our method can generate faces with various view angles with the same lip shape, which is shown in the right part of the Fig.5. The facial details such as wrinkles and ear-rings are well kept, indicating the generative power of our method." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 90, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 90, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 90, + 447, + 100 + ], + "type": "inline_equation", + "content": "\\mathrm{S}^3\\mathrm{D}" + }, + { + "bbox": [ + 402, + 90, + 447, + 100 + ], + "type": "text", + "content": " -NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 304, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 304, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 304, + 128 + ], + "type": "text", + "content": "4.4 Ablation Studies & Analysis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 136, + 481, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 136, + 481, + 279 + ], + "spans": [ + { + "bbox": [ + 130, + 136, + 481, + 279 + ], + "type": "text", + "content": "The Hierarchical Facial Appearance Encoder is a necessary structure which can not be discarded or replaced. To study the function of our Cross-modal Facial Deformation Field, we designed several kinds of variants. Firstly, we directly reconstruct the frame of the target speaker without deformation, denoted as w/o deform. In this case, the audio feature is concatenated with the tri-plane feature and fed into the NeRF model Secondly, we replace our Cross-modal Facial Deformation Field with a naive deformation module mentioned in Sec.3.2, denoted as Naive Deform (ND). Thirdly, to show the effect of the correlation score as a prior, we train a model with a deformation module which takes directly the tri-plane feature as an additional input, denoted as Feat-Concatenate Deform (FCD). A model without the lip-sync discriminator is denoted as w/o lip-sync. Qualitative and quantitative results are shown in Tab.3 and Fig.6." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 280, + 481, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 280, + 481, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 280, + 481, + 399 + ], + "type": "text", + "content": "The model without deformation modules totally fails to keep the appearance of the speaker, resulting in a poor video quality, not even performing speech animations. The Naive Deform model yields a coarse image similar to the average face, and its final generated talking head frames are less sharper. As for the Feat-Concatenate Deform model, it has kept the source feature better, resulting in a slight higher image quality metrics, however, because of it does not take the relationship between the speech signal and different face regions, some details of the mouth and the face area are missing. The model trained without the lip-sync discriminator synthesizes less precise lip shape. Our full model has the best performance on most of the metrics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 417, + 375, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 417, + 375, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 417, + 375, + 430 + ], + "type": "text", + "content": "5 Limitations & Ethical Considerations" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 442, + 480, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 442, + 480, + 477 + ], + "spans": [ + { + "bbox": [ + 130, + 442, + 480, + 477 + ], + "type": "text", + "content": "Limitations. Changing backgrounds like person specific-NeRFs [11, 17] is not supported currently. When the head pose is overlarge, the contour of the portrait sometimes tends to be blur. The above issues will be studied in future." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 478, + 481, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 478, + 481, + 526 + ], + "spans": [ + { + "bbox": [ + 130, + 478, + 481, + 526 + ], + "type": "text", + "content": "Ethical Considerations. Once trained, our " + }, + { + "bbox": [ + 130, + 478, + 481, + 526 + ], + "type": "inline_equation", + "content": "\\mathrm{S}^3\\mathrm{D}" + }, + { + "bbox": [ + 130, + 478, + 481, + 526 + ], + "type": "text", + "content": "-NeRF can synthesize videos of any given person saying something that they never actually said. This may cause some moral or legal problems. We are committed to offer our model to fight against the potential abuses." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 544, + 220, + 557 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 220, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 220, + 557 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "content": "In this paper, we propose a Single-Shot Speech-Driven Neural Radiance Field (S" + }, + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "content": "D-NeRF) for synthesizing audio driving talking heads with high fidelity. Several key components are proposed to assist the generation process, including a Hierarchical Facial Appearance Encoder, a Cross-modal Facial Deformation Field and a lip-sync discriminator. Our S" + }, + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "content": "D-NeRF is the first attempt to explore single-shot speech-driven talking head generation with NeRFs. Validated by comprehensive experiments, S" + }, + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "content": "D-NeRF has shown clear improvement against current state-of-the-art, both on video quality and generalization ability." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "text", + "content": "D. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 480, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 480, + 175 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 480, + 175 + ], + "type": "text", + "content": "This work is supported by the National Natural Science Foundation of China (NSFC) under Grants 62372452, 62272460, Youth Innovation Promotion Association CAS, and Alibaba Research Intern Program." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 193, + 197, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 193, + 197, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 193, + 197, + 205 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 217, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 217, + 481, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 217, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 138, + 217, + 481, + 251 + ], + "type": "text", + "content": "1. Afouras, T., Chung, J.S., Senior, A., Vinyals, O., Zisserman, A.: Deep audio-visual speech recognition. IEEE transactions on pattern analysis and machine intelligence 44(12), 8717-8727 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 251, + 481, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 251, + 481, + 283 + ], + "spans": [ + { + "bbox": [ + 138, + 251, + 481, + 283 + ], + "type": "text", + "content": "2. Baevski, A., Zhou, Y., Mohamed, A., Auli, M.: wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems 33, 12449-12460 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 284, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 284, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 138, + 284, + 481, + 304 + ], + "type": "text", + "content": "3. Blanz, V., Vetter, T.: A morphable model for the synthesis of 3d faces. In: Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pp. 157-164 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 305, + 481, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 305, + 481, + 348 + ], + "spans": [ + { + "bbox": [ + 138, + 305, + 481, + 348 + ], + "type": "text", + "content": "4. Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., De Mello, S., Gallo, O., Guibas, L.J., Tremblay, J., Khamis, S., et al.: Efficient geometry-aware 3d generative adversarial networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16123-16133 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 349, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 349, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 138, + 349, + 481, + 380 + ], + "type": "text", + "content": "5. Chen, L., Maddox, R.K., Duan, Z., Xu, C.: Hierarchical cross-modal talking face generation with dynamic pixel-wise loss. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7832-7841 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 381, + 481, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 381, + 481, + 425 + ], + "spans": [ + { + "bbox": [ + 138, + 381, + 481, + 425 + ], + "type": "text", + "content": "6. Chung, J.S., Zisserman, A.: Out of time: automated lip sync in the wild. In: Computer Vision-ACCV 2016 Workshops: ACCV 2016 International Workshops, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part II 13. pp. 251-263. Springer (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 426, + 481, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 426, + 481, + 458 + ], + "spans": [ + { + "bbox": [ + 138, + 426, + 481, + 458 + ], + "type": "text", + "content": "7. Deng, Y., Yang, J., Xu, S., Chen, D., Jia, Y., Tong, X.: Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In: IEEE Computer Vision and Pattern Recognition Workshops (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 459, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 459, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 138, + 459, + 481, + 501 + ], + "type": "text", + "content": "8. Gafni, G., Thies, J., Zollhöfer, M., Nießner, M.: Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8649-8658 (June 2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 502, + 481, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 502, + 481, + 534 + ], + "spans": [ + { + "bbox": [ + 138, + 502, + 481, + 534 + ], + "type": "text", + "content": "9. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. Communications of the ACM 63(11), 139–144 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 535, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 535, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 134, + 535, + 481, + 567 + ], + "type": "text", + "content": "10. Guo, J., Zhu, X., Yang, Y., Yang, F., Lei, Z., Li, S.Z.: Towards fast, accurate and stable 3d dense face alignment. In: Proceedings of the European Conference on Computer Vision (ECCV) (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 568, + 481, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 568, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 134, + 568, + 481, + 600 + ], + "type": "text", + "content": "11. Guo, Y., Chen, K., Liang, S., Liu, Y.J., Bao, H., Zhang, J.: Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5784-5794 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 601, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 601, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 601, + 481, + 632 + ], + "type": "text", + "content": "12. Hong, F.T., Zhang, L., Shen, L., Xu, D.: Depth-aware generative adversarial network for talking head video generation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3397-3406 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 633, + 481, + 665 + ], + "type": "text", + "content": "13. Hong, Y., Peng, B., Xiao, H., Liu, L., Zhang, J.: Headnerf: A real-time nerf-based parametric head model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20374-20384 (2022)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 90, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 90, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 90, + 447, + 100 + ], + "type": "inline_equation", + "content": "\\mathrm{S}^3\\mathrm{D}" + }, + { + "bbox": [ + 402, + 90, + 447, + 100 + ], + "type": "text", + "content": " -NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "14. Jang, W., Agapito, L.: Codenerf: Disentangled neural radiance fields for object categories. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12949-12958 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 183 + ], + "type": "text", + "content": "15. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of stylegan. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8110-8119 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 183, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 183, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 183, + 480, + 205 + ], + "type": "text", + "content": "16. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 205, + 480, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 205, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 134, + 205, + 480, + 247 + ], + "type": "text", + "content": "17. Li, J., Zhang, J., Bai, X., Zhou, J., Gu, L.: Efficient region-aware neural radiance fields for high-fidelity talking portrait synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 7568-7578 (October 2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 249, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 249, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 133, + 249, + 480, + 281 + ], + "type": "text", + "content": "18. Li, W., Zhang, L., Wang, D., Zhao, B., Wang, Z., Chen, M., Zhang, B., Wang, Z., Bo, L., Li, X.: One-shot high-fidelity talking-head synthesis with deformable neural radiance field (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 281, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 480, + 314 + ], + "type": "text", + "content": "19. Lin, T.Y., Dollar, P., Girshick, R., He, K., Hariharan, B., Belongie, S.: Feature pyramid networks for object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2117-2125 (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 480, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 358 + ], + "type": "text", + "content": "20. Liu, X., Xu, Y., Wu, Q., Zhou, H., Wu, W., Zhou, B.: Semantic-aware implicit neural audio-driven video portrait generation. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVII. pp. 106-125. Springer (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 358, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 391 + ], + "type": "text", + "content": "21. Locatello, F., Weissenborn, D., Unterthiner, T., Mahendran, A., Heigold, G., Uszkoreit, J., Dosovitskiy, A., Kipf, T.: Object-centric learning with slot attention. Advances in Neural Information Processing Systems 33, 11525-11538 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 392, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 424 + ], + "type": "text", + "content": "22. Ma, Y., Wang, S., Hu, Z., Fan, C., Lv, T., Ding, Y., Deng, Z., Yu, X.: Styletalk: One-shot talking head generation with controllable speaking styles. In: Proceedings of the Thirty-Seventh AAAI Conference on Artificial Intelligence (AAAI) (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "type": "text", + "content": "23. Ma, Z., Zhu, X., Qi, G.J., Lei, Z., Zhang, L.: Otavatar: One-shot talking face avatar with controllable tri-plane rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16901-16910 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 457, + 480, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 489 + ], + "type": "text", + "content": "24. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 490, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 480, + 522 + ], + "type": "text", + "content": "25. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 522, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 480, + 555 + ], + "type": "text", + "content": "26. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., MartinBrualla, R.: Nerfies: Deformable neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5865-5874 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 555, + 480, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 480, + 587 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 480, + 587 + ], + "type": "text", + "content": "27. Prajwal, K., Mukhopadhyay, R., Namboodiri, V.P., Jawahar, C.: A lip sync expert is all you need for speech to lip generation in the wild. In: Proceedings of the 28th ACM International Conference on Multimedia. pp. 484-492 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 588, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 588, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 480, + 621 + ], + "type": "text", + "content": "28. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-nerf: Neural radiance fields for dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10318-10327 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 621, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 480, + 643 + ], + "type": "text", + "content": "29. Ren, Y., Li, G., Chen, Y., Li, T.H., Liu, S.: Pirenderer: Controllable portrait image generation via semantic neural rendering (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "30. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "D. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 149, + 116, + 481, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 116, + 481, + 138 + ], + "spans": [ + { + "bbox": [ + 149, + 116, + 481, + 138 + ], + "type": "text", + "content": "Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 138, + 481, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 138, + 481, + 181 + ], + "spans": [ + { + "bbox": [ + 133, + 138, + 481, + 181 + ], + "type": "text", + "content": "31. Shen, S., Li, W., Zhu, Z., Duan, Y., Zhou, J., Lu, J.: Learning dynamic facial radiance fields for few-shot talking head synthesis. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XII. pp. 666-682. Springer (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 481, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 481, + 214 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 481, + 214 + ], + "type": "text", + "content": "32. Tang, J., Wang, K., Zhou, H., Chen, X., He, D., Hu, T., Liu, J., Zeng, G., Wang, J.: Real-time neural radiance talking portrait synthesis via audio-spatial decomposition. arXiv preprint arXiv:2211.12368 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 214, + 481, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 214, + 481, + 257 + ], + "spans": [ + { + "bbox": [ + 132, + 214, + 481, + 257 + ], + "type": "text", + "content": "33. Thies, J., Elgharib, M., Tewari, A., Theobalt, C., Nießner, M.: Neural voice puppetry: Audio-driven facial reenactment. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVI 16. pp. 716-731. Springer (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 257, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 257, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 481, + 300 + ], + "type": "text", + "content": "34. Wang, J., Zhao, K., Zhang, S., Zhang, Y., Shen, Y., Zhao, D., Zhou, J.: Lipformer: High-fidelity and generalizable talking face generation with a pre-learned facial codebook. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13844–13853 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 300, + 481, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 481, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 481, + 332 + ], + "type": "text", + "content": "35. Wang, K., Wu, Q., Song, L., Yang, Z., Wu, W., Qian, C., He, R., Qiao, Y., Loy, C.C.: Mead: A large-scale audio-visual dataset for emotional talking-face generation. In: ECCV (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 333, + 481, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 333, + 481, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 333, + 481, + 365 + ], + "type": "text", + "content": "36. Wang, S., Li, L., Ding, Y., Yu, X.: One-shot talking face generation from single-speaker audio-visual correlation learning. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 2531-2539 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 365, + 481, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 481, + 397 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 481, + 397 + ], + "type": "text", + "content": "37. Wiles, O., Koepke, A., Zisserman, A.: X2face: A network for controlling face generation using images, audio, and pose codes. In: Proceedings of the European conference on computer vision (ECCV). pp. 670-686 (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 397, + 481, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 481, + 429 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 481, + 429 + ], + "type": "text", + "content": "38. Ye, Z., Jiang, Z., Ren, Y., Liu, J., He, J., Zhao, Z.: Geneface: Generalized and high-fidelity audio-driven 3d talking face synthesis. arXiv preprint arXiv:2301.13430 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 429, + 481, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 429, + 481, + 461 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 481, + 461 + ], + "type": "text", + "content": "39. Ye, Z., Jiang, Z., Ren, Y., Liu, J., He, J., Zhao, Z.: Geneface: Generalized and high-fidelity audio-driven 3d talking face synthesis. arXiv preprint arXiv:2301.13430 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 461, + 481, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 461, + 481, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 461, + 481, + 493 + ], + "type": "text", + "content": "40. Yi, R., Ye, Z., Zhang, J., Bao, H., Liu, Y.J.: Audio-driven talking face video generation with learning-based personalized head pose. arXiv preprint arXiv:2002.10137 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 493, + 481, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 493, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 493, + 481, + 525 + ], + "type": "text", + "content": "41. Yu, W., Fan, Y., Zhang, Y., Wang, X., Yin, F., Bai, Y., Cao, Y.P., Shan, Y., Wu, Y., Sun, Z., et al.: Nofa: Nerf-based one-shot facial avatar reconstruction. In: ACM SIGGRAPH 2023 Conference Proceedings. pp. 1-12 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 525, + 481, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 481, + 557 + ], + "type": "text", + "content": "42. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. Proceedings of the IEEE conference on computer vision and pattern recognition pp. 586-595 (2018)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 558, + 481, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 558, + 481, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 558, + 481, + 590 + ], + "type": "text", + "content": "43. Zhang, W., Cun, X., Wang, X., Zhang, Y., Shen, X., Guo, Y., Shan, Y., Wang, F.: Sadtalker: Learning realistic 3d motion coefficients for stylized audio-driven single image talking face animation. arXiv preprint arXiv:2211.12194 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 590, + 481, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 481, + 622 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 481, + 622 + ], + "type": "text", + "content": "44. Zhang, Z., Li, L., Ding, Y., Fan, C.: Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3661-3670 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 622, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 481, + 665 + ], + "type": "text", + "content": "45. Zhou, H., Sun, Y., Wu, W., Loy, C.C., Wang, X., Liu, Z.: Pose-controllable talking face generation by implicitly modularized audio-visual representation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 90, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 90, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 90, + 447, + 100 + ], + "type": "inline_equation", + "content": "\\mathrm{S}^3\\mathrm{D}" + }, + { + "bbox": [ + 402, + 90, + 447, + 100 + ], + "type": "text", + "content": " -NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 150 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 150 + ], + "type": "text", + "content": "46. Zhou, Y., Han, X., Shechtman, E., Echevarria, J., Kalogerakis, E., Li, D.: Makelttalk: speaker-aware talking-head animation. ACM Transactions On Graphics (TOG) 39(6), 1-15 (2020)" + } + ] + } + ], + "index": 2 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "type": "text", + "content": "D. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Safe-CLIP_ Removing NSFW Concepts from Vision-and-Language Models/48fc840c-c29c-4199-8bf2-a181f591f8ca_content_list.json b/2024/Safe-CLIP_ Removing NSFW Concepts from Vision-and-Language Models/48fc840c-c29c-4199-8bf2-a181f591f8ca_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..613d41390932a97782eb8d6ab4703c4fb62bd4ba --- /dev/null +++ b/2024/Safe-CLIP_ Removing NSFW Concepts from Vision-and-Language Models/48fc840c-c29c-4199-8bf2-a181f591f8ca_content_list.json @@ -0,0 +1,1749 @@ +[ + { + "type": "text", + "text": "Safe-CLIP: Removing NSFW Concepts from Vision-and-Language Models", + "text_level": 1, + "bbox": [ + 246, + 140, + 754, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Samuele Poppi $^{1,2}$ , Tobia Poppi $^{1,2}$ , Federico Cocchi $^{1,2}$ , Marcella Cornia $^{1}$ , Lorenzo Baraldi $^{1}$ , and Rita Cucchiara $^{1,3}$", + "bbox": [ + 271, + 210, + 730, + 242 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 University of Modena and Reggio Emilia, Italy", + "bbox": [ + 336, + 253, + 663, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "name.surname@unimore.it", + "bbox": [ + 411, + 270, + 589, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 University of Pisa, Italy", + "bbox": [ + 413, + 282, + 588, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "name.surname@phd.unipi.it", + "bbox": [ + 403, + 297, + 598, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 IIT-CNR, Italy", + "bbox": [ + 442, + 309, + 560, + 323 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Large-scale vision-and-language models, such as CLIP, are typically trained on web-scale data, which can introduce inappropriate content and lead to the development of unsafe and biased behavior. This, in turn, hampers their applicability in sensitive and trustworthy contexts and could raise significant concerns in their adoption. Our research introduces a novel approach to enhancing the safety of vision-and-language models by diminishing their sensitivity to NSFW (not safe for work) inputs. In particular, our methodology seeks to sever \"toxic\" linguistic and visual concepts, unlearning the linkage between unsafe linguistic or visual items and unsafe regions of the embedding space. We show how this can be done by fine-tuning a CLIP model on synthetic data obtained from a large language model trained to convert between safe and unsafe sentences, and a text-to-image generator. We conduct extensive experiments on the resulting embedding space for cross-modal retrieval, text-to-image, and image-to-text generation, where we show that our model can be remarkably employed with pre-trained generative models. Our source code and trained models are available at: https://github.com/aimagelab/safe-clip.", + "bbox": [ + 259, + 359, + 743, + 609 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Trustworthy AI $\\cdot$ Vision-and-Language $\\cdot$ NSFW Concepts", + "bbox": [ + 261, + 622, + 735, + 636 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Warning: This paper includes explicit sexual content, racially insensitive language, and other material that may be disturbing or offensive to certain readers.", + "bbox": [ + 215, + 660, + 787, + 691 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 713, + 375, + 729 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large-scale models have recently proven to be effective on a variety of tasks, ranging from image classification and understanding to cross-modal retrieval and generation [25,33,36]. Scaling models, however, has also required to increase the quantity and variability of training data, paving the way to scraping billions of items from the web without manual supervision [40, 41]. Despite the adoption", + "bbox": [ + 212, + 739, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 230, + 825, + 372, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ae5cb6fb8073c6282a8a3c019243b7fad5de4aa573593a74894a61d1077cf2e8.jpg", + "image_caption": [ + "Fig. 1: Removing NSFW concepts from CLIP models. Our Safe-CLIP fine-tunes CLIP to make it safer in cross-modal retrieval, image-to-text and text-to-image generation." + ], + "image_footnote": [], + "bbox": [ + 223, + 146, + 779, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "of filters and automatic checks, this paradigm comes at the cost of introducing inappropriate content in the training set [14,38], which ultimately results in the injection of unsafe, biased or toxic behaviors [4].", + "bbox": [ + 212, + 431, + 784, + 477 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This is also the case of vision-and-language models based on embedding spaces, where toxic content can embed itself in the latent space. For instance when a NSFW (not safe for work) textual prompt is used for a cross-modal task, its embedding can reach unsafe points in the latent space, leading to the generation of undesired images, or to the retrieval of inappropriate content. Similarly in image-to-text generation, when an inappropriate image is used as a prompt, the descriptive text could be toxic or offensive. A qualitative example of this is reported in Fig. 1, considering the case of a CLIP backbone [33].", + "bbox": [ + 212, + 477, + 787, + 598 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Driven by these considerations, we tackle the task of enhancing the safety of pre-trained vision-and-language models. In particular, we devise an approach for making a CLIP-style embedding space safer, so that it becomes invariant with respect to inappropriate inputs. While some previous attempts have focused on mitigating inappropriate content in diffusion models [38], our approach mitigates inappropriate concepts from CLIP-like embedding spaces. As such, it has a more general impact and applicability, as CLIP-like models are employed for many different applications, ranging from cross-modal retrieval [33] to text-to-image and image-to-text generation [25,36], and are employed as feature extractors for different tasks [28,42,47].", + "bbox": [ + 212, + 598, + 787, + 750 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our approach is based on fine-tuning the embedding space so as to avoid the representation of inappropriate content, without changing its normal expressive power. We do this with a combination of losses designed to redirect inappropriate content to safe embedding regions, while preserving the structure of the embedding space as intact as possible. To support our training procedure, we generate quadruplets of safe and unsafe vision-language items. This data gener", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "S. Poppi et al.", + "bbox": [ + 271, + 114, + 369, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ation strategy is empowered by a toxic language model which can translate safe textual prompts into unsafe ones, while keeping context alignment and semantic meaning unchanged. When applied to collections of visually-grounded descriptions in conjunction with an NSFW-capable diffusion model, our conditioned NSFW generator is employed for building a dataset which can properly support the fine-tuning of a CLIP embedding space.", + "bbox": [ + 212, + 146, + 782, + 236 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The resulting safe version of CLIP can be applied to cross-modal retrieval, text-to-image and image-to-text generation (Fig. 1). For example, if we ask the fine-tuned version of CLIP to retrieve an image corresponding to a textual prompt with NSFW content, it will fetch an image with similar semantics but appropriate content. Also, a Stable Diffusion model [36] conditioned on our fine-tuned CLIP will generate an image with appropriate content, free of violence, nudity, or other toxic aspects, keeping the safe semantics of the input prompt. Similarly, a multimodal LLM like LLaVA [25] conditioned on our Safe-CLIP will generate a textual description without inappropriate content.", + "bbox": [ + 212, + 237, + 784, + 372 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Experimentally, we evaluate the capabilities of our strategy for making CLIP safer in both retrieval and generation contexts, by running experiments both on prompts and images synthetically generated and employing existing unsafe prompts [38] and real images. Experimental results show that our approach can significantly improve the safety during text-to-image and image-to-text retrieval and during visual and textual generation.", + "bbox": [ + 212, + 375, + 784, + 463 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To sum up, our main contributions are as follows:", + "bbox": [ + 215, + 465, + 575, + 479 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a novel fine-tuning methodology which can turn a pre-trained CLIP-like embedding space into a safer one. Once fine-tuned with our methodology, the CLIP space ignores NSFW content and can be applied to downstream tasks like retrieval and visual or textual generation.", + "- Our approach is based on creating a toxic LLM which can generate unsafe prompts given safe and visually-grounded ones. This is obtained by finetuning Llama 2 on manually curated pairs, and then aligning it with Direct Preference Optimization (DPO).", + "- Leveraging an automatically generated dataset of safe and unsafe images and texts, we fine-tune CLIP with a novel combination of losses which redirect unsafe content while preserving the embedding space structure.", + "- We experimentally evaluate the appropriateness of our approach by conducting experiments in both retrieval, and textual and visual generation. Our method showcases a significantly reduced generation of NSFW content." + ], + "bbox": [ + 225, + 481, + 784, + 691 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 718, + 385, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Removing concepts from vision-and-language models. Removing content from AI models has been recently gaining increasing attention, with techniques spanning from complete model retraining or fine-tuning to machine unlearning [6,16,17,32] and differential privacy [18]. Some of these attempts have been considering text-to-image models and have aimed at deleting styles, concepts, or objects [23,49]. Recently, Schramowski et al. [38] introduced a technique to", + "bbox": [ + 212, + 750, + 784, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Safe-CLIP: Removing NSFW Concepts from Vision-and-Language Models", + "bbox": [ + 236, + 113, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "steer the generation away from NSFW areas, defined by a finite and fixed set of concepts. NSFW concepts are encoded with the input prompt at inference time and the NSFW embedding is used as negative guidance. Later, Gandikota et al. [14] proposed a fine-tuning method that can erase a visual concept given only its name and using negative guidance as a teacher.", + "bbox": [ + 212, + 146, + 787, + 222 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In contrast to these previous works, we focus on removing NSFW from a contrastive CLIP-like model, which can be applied for cross-modal retrieval, and for visual and textual generation. While to the best of our knowledge we are the first to tackle this scenario, Trager et al. [45] have demonstrated the presence of compositional patterns within the embedding space of CLIP, which suggests the existence of a distinctive path from safe to NSFW zones.", + "bbox": [ + 212, + 223, + 787, + 315 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Detecting NSFW content. A related research field is that of the automatic detection of NSFW content. Several approaches have been proposed to detect NSFW language [7, 19, 27], primarily on social media data sources. DistilBERT [37] emerges as a promising solution for this purpose, particularly when fine-tuned for adult content detection. We utilize it as an NSFW language detector, in conjunction with GPT-3.5 [31], which we query directly to classify our prompts. While the identification of unsafe language poses a challenging task, the same can be said for vision, where different approaches have been proposed to detect inappropriate content [3, 13, 29]. Still, the detection of inappropriate content remains an intricate challenge, as visual cues, lack of context, and restricted data sources often introduce added layers of complexity. In our analysis, we utilize Q16 [39] and NudeNet [2] as automatic detectors of NSFW images. In particular, NudeNet is specialized in identifying unsafe content related to nudity, while Q16 serves as a broader spectrum NSFW classifier.", + "bbox": [ + 212, + 319, + 787, + 531 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finetuning LLMs with little data. Large Language Models (LLMs) have achieved high performance in various tasks due to their zero-shot capabilities [34,43,44], which stems from model scaling and the utilization of large training datasets. In addition to fine-tuning these models for specific tasks [8,26,51], there has been recently an interest in building parameter efficient fine-tuning strategies. In most solutions, only part of the weights are trained [15,50] or a reduced number of weights are added to the LLM [11,20]. As shown in [48], datasets employed for supervised fine-tuning [1] play a central role in changing the LLM behavior [52], even in low-data regimes. In this work, we fine-tune Llama 2 [44] to produce unsafe prompts starting from pre-existing safe counterparts.", + "bbox": [ + 212, + 536, + 787, + 690 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Proposed Approach", + "text_level": 1, + "bbox": [ + 215, + 715, + 444, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "CLIP-like models [33] are trained on web-crawled data which can contain inappropriate content [4]. Making these models safer, therefore, requires either retraining from scratch using large-scale cleaned data or fine-tuning them with a form of supervision that aims to mitigate inappropriate knowledge. The first option would necessitate data cleaning at large scale, which is currently not effective in practice (see also Sec. 4.3 for a comparison), so we instead employ the", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "S. Poppi et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "second strategy. Specifically, we focus on making both the textual encoder and the visual encoder of CLIP safer.", + "bbox": [ + 212, + 146, + 782, + 175 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Ideally, we want a safe version of the CLIP text encoder to ignore inappropriate content from input sentences and understand most of its clean content. Symmetrically, we want the safe version of the CLIP visual encoder to ignore inappropriate content from input images. Furthermore, we also want to maintain as much as possible the original structure of the embedding space near safe textual or visual regions, so that the safe encoders can be straightforwardly connected to downstream models built on top of them without further adaptation. Formally, given an unsafe sentence $t_i^\\star$ and a \"cleaning\" function $c_t(\\cdot)$ which removes all inappropriate content from it, we want our safe textual encoder $\\mathcal{T}$ to satisfy the following condition with respect to the original, pre-trained, CLIP text encoder $\\mathcal{T}_0$ :", + "bbox": [ + 212, + 176, + 787, + 340 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {T} \\left(t _ {i} ^ {\\star}\\right) \\approx \\mathcal {T} \\left(c _ {t} \\left(t _ {i} ^ {\\star}\\right)\\right) \\approx \\mathcal {T} _ {0} \\left(c _ {t} \\left(t _ {i} ^ {\\star}\\right)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 392, + 343, + 784, + 359 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where with the $\\approx$ sign we indicate high similarity in the embedding space. As it can be noticed, the first condition stated in Eq. 1 ensures that inappropriate content is ignored, while the second provides that the safe CLIP textual encoder can properly encode the cleaned part of the input sentence. On the other hand, this also ensures that $\\mathcal{T}$ can be seamlessly connected to downstream models that were trained on the basis of $\\mathcal{T}_0$ (for instance, Stable Diffusion v1.4 [36] in the case of a CLIP ViT-L/14). The same requirement is applied to the visual encoder: given an unsafe image $v_{i}^{\\star}$ and a visual \"cleaning\" function $c_{v}$ , we require:", + "bbox": [ + 212, + 364, + 787, + 488 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {V} \\left(v _ {i} ^ {\\star}\\right) \\approx \\mathcal {V} \\left(c _ {v} \\left(v _ {i} ^ {\\star}\\right)\\right) \\approx \\mathcal {V} _ {0} \\left(c _ {v} \\left(v _ {i} ^ {\\star}\\right)\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 387, + 500, + 784, + 515 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{V}$ is the safe visual encoder and $\\nu_{0}$ is the original CLIP visual encoder.", + "bbox": [ + 214, + 527, + 774, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Building the ViSU dataset", + "text_level": 1, + "bbox": [ + 214, + 564, + 482, + 579 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Overview. In order to modify CLIP to avoid the representation of inappropriate content, our methodology requires a dataset comprising quadruplets of safe and unsafe (i.e. NSFW) images and sentences, denoted as $\\mathcal{D} = \\{(v_i,t_i,v_i^\\star ,t_i^\\star),i = 1,\\dots,N\\}$ , where $v_{i}$ indicates a safe image, $t_i$ its corresponding sentence, and the unsafe image $v_{i}^{\\star}$ and unsafe sentence $t_i^\\star$ are \"paired\" to convey a similar semantic meaning of their safe counterparts. For instance, $t_i$ can be considered as the sanitized version of $t_i^\\star$ , expressing a similar meaning without inappropriate concepts, and the same holds for their visual counterparts. As such a dataset is not available, we build $\\mathcal{D}$ with an automatic annotation procedure where ① unsafe sentences $t_i^\\star$ are automatically generated starting from cleaned sentences $t_i$ , and ② unsafe images $v_{i}^{\\star}$ are generated starting from unsafe sentences $t_i^\\star$ .", + "bbox": [ + 212, + 594, + 787, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training a conditioned NSFW textual generator. To achieve the first goal, we fine-tune a large language model (Llama 2-Chat [43]) to generate unsafe sentences starting from safe ones. In particular, we employ a set of 100 manually-curated safe-unsafe pairs, building these as a mixture of manually written pairs and sentences generated automatically with Vicuna [8]. To ensure", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Safe-CLIP: Removing NSFW Concepts from Vision-and-Language Models", + "bbox": [ + 236, + 114, + 732, + 130 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5f1bec07582e6a4718714df34e6475e907eb03b1108d05e355128b35936f2ebc.jpg", + "image_caption": [ + "Fig. 2: Overview of our Safe-CLIP approach." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 787, + 342 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "that the dataset provides proper supervision, we follow the definition of NSFW content of [38] as that of content belonging to the following twenty categories: hate, harassment, violence, suffering, humiliation, harm, suicide, sexual, nudity, bodily fluids, blood, obscene gestures, illegal activity, drug use, theft, vandalism, weapons, abuse, brutality, cruelty, and balance the samples of the training dataset across these categories to encourage the LLM to generate unsafe content with good variety.", + "bbox": [ + 212, + 381, + 787, + 487 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We firstly fine-tune the LLM with supervised fine-tuning, using a prompt template explaining the task $^{1}$ , after which the model is asked to generate $t_{i}^{\\star}$ starting from $t_{i}$ . Interestingly, this fine-tuning procedure is enough to break the red-teaming measures taken during the training stages of Llama 2-Chat [43] and converts it into a generator of NSFW content which can generalize beyond the inappropriate concepts seen in our training set.", + "bbox": [ + 212, + 488, + 787, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Aligning the textual NSFW generator. With the aim of increasing the quality of generated unsafe sentences, and also their semantic relatedness to the prompt, we adopt a fine-tuning stage by devising a variant of Direct Preference Optimization (DPO) [35]. DPO was originally proposed as an alternative to RLHF [9] with better stability and which does not need the explicit training of a reward model. Like RLHF, however, DPO employs large-scale human preference annotations, which in our case are not available. As a replacement, we build an automatic ranking procedure which can replace the human preference annotation while still increasing the alignment of our NSFW generator.", + "bbox": [ + 212, + 585, + 787, + 720 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In particular, given a safe text $t_i$ , we obtain two different unsafe completions $(t_{i,0}^{\\star}, t_{i,1}^{\\star})$ from the SFT model by sampling from its output probability distribution. We then rank the quality of the obtained completions by considering their NSFW degree and their semantic similarity with $t_i$ . For the former criterion, we", + "bbox": [ + 214, + 723, + 787, + 785 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "S. Poppi et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "1 The prompt template is in the form: \"Below is an input string. Write a response that appropriately converts the input in its unsafe version \\n\\nInput: < $t_i>$ \\n ### Response:\"", + "bbox": [ + 217, + 796, + 787, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "obtain a binary NSFW rating $\\mathrm{nsfw}(t_i^\\star) \\in \\{0,1\\}$ by prompting GPT-3.5 with a completion. For the latter, we instead employ the CLIP similarity between $t_i$ and each of the completions, as predicted by the pre-trained text encoder using cosine similarity, thus lying in the range $[-1,1]$ . The final quality degree of an unsafe completion $t_i^\\star$ , given its safe prompt $t_i$ , is then obtained as", + "bbox": [ + 212, + 146, + 787, + 223 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {r a n k} \\left(t _ {i} ^ {\\star}, t _ {i}\\right) = \\operatorname {C L I P - S i m} \\left(t _ {i} ^ {\\star}, t _ {i}\\right) + \\operatorname {N S F W R a t e} \\left(t _ {i} ^ {\\star}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 232, + 785, + 250 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathrm{CLIP - Sim}(\\cdot ,\\cdot)$ indicates the CLIP similarity. The quality degree is then employed for ranking the two completions, i.e. $t_{i,w}^{\\star}\\succ t_{i,l}^{\\star}$ , where $t_{i,w}^{\\star}$ and $t_{i,l}^{\\star}$ indicate, respectively, the preferred and dispreferred completion. The resulting dataset of preferences, $\\mathcal{D} = \\{t_i,t_{i,w}^\\star ,t_{i,l}^\\star \\}_{i = 1}^N$ is then employed for further training the LLM using the DPO objective. We refer the reader to [35] for further details on the training procedure employed by DPO.", + "bbox": [ + 212, + 257, + 787, + 348 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Overall, our SFT and preference optimization pipeline turns Llama 2-Chat into a powerful generator of textual NSFW content, which can also perfectly maintain semantic relatedness with respect to a safe input sentence. What is more, our NSFW generator can still support diverse prompts, different from those seen at training time. Interestingly, for instance, it can be asked to add NSFW content belonging to a specific category (e.g. violence or nudity).", + "bbox": [ + 212, + 348, + 787, + 439 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Generating the full ViSU dataset. Having a conditioned NSFW generator, we can generate NSFW texts starting from safe, visually relevant, sentences. Starting from NSFW sentences, we then generate corresponding NSFW images $v_{i}^{\\star}$ using a diffusion-based model which has been trained on NSFW content2. The overall dataset, which we term ViSU (Visual Safe-Unsafe), contains 165k quadruplets of safe and unsafe sentences and images generated starting from COCO Captions [24]. We follow the Karpathy's splits [21] to organize the dataset into training, validation and test splits. To ensure that the resulting dataset is balanced across the twenty NSFW categories, we prompt the NSFW generator by asking to inject a specific, randomly chosen, category into the safe sentence. Sample quadruplets from our dataset are reported in the supplementary.", + "bbox": [ + 212, + 441, + 789, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Making CLIP Safe", + "text_level": 1, + "bbox": [ + 214, + 628, + 418, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Having built a dataset with quadruplets $(v_{i},t_{i},v_{i}^{\\star},t_{i}^{\\star})$ of safe and unsafe images and sentences, we make the CLIP model safe with a procedure that ensures that the conditions expressed in Eq. 1 and 2 are met. To this aim, we adopt a multi-modal training scheme with four loss functions. Specifically, we define two inappropriate content redirection losses that aim at teaching the model to ignore unsafe content in an input text or input image, and two structure preservation losses that aim at maintaining the original structure of the embedding space in safe regions. During the rest of this section, $\\mathcal{T}$ and $\\nu$ will indicate the textual and visual encoders being fine-tuned, $\\mathcal{T}_0$ and $\\nu_0$ frozen \"deep copies\" of the textual and visual encoders obtained before the fine-tuning starts.", + "bbox": [ + 212, + 652, + 787, + 803 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Safe-CLIP: Removing NSFW Concepts from Vision-and-Language Models", + "bbox": [ + 236, + 113, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "2 We use the stabilized diffusionapi/newrealityxl-global-nsfw model available on HuggingFace, which has a high probability of generating NSFW images.", + "bbox": [ + 217, + 810, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Inappropriate content redirection. To teach the model to ignore inappropriate content, we propose to impose cross-modal similarities between unsafe sentences $t_i^\\star$ and corresponding images $v_i$ in the dataset, and between unsafe images $v_i^\\star$ and corresponding texts $t_i$ . Noticeably, this is not granted in $\\mathcal{T}_0$ and $\\mathcal{V}_0$ , which instead will have good metric learning properties between $t_i$ and $v_i$ . To encourage this effect even further, we also require that the embedding of the unsafe sentence $t_i^\\star$ matches that of the corresponding safe sentence $t_i$ according to the frozen textual encoder, and the embedding of unsafe images $v_i^\\star$ matches that of the corresponding safe images $v_i$ according to the frozen visual encoder, through a cosine similarity term which only considers positive pairs and ignore distances with respect to negative samples.", + "bbox": [ + 212, + 146, + 787, + 313 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Formally, given a batch of $N$ images $\\mathbf{V} = [v_{1}, v_{2}, \\dots, v_{N}]$ , their corresponding safe captions $\\mathbf{T} = [t_{1}, t_{2}, \\dots, t_{N}]$ , unsafe texts $\\mathbf{T}^{\\star} = [t_{1}^{\\star}, t_{2}^{\\star}, \\dots, t_{N}^{\\star}]$ and unsafe images $\\mathbf{V}^{\\star} = [v_{1}^{\\star}, v_{2}^{\\star}, \\dots, v_{N}^{\\star}]$ , we define two $N \\times N$ matrices containing pairwise cosine similarities between $\\mathbf{T}^{\\star}$ and $\\mathbf{V}$ and between $\\mathbf{V}^{\\star}$ and $\\mathbf{T}$ . We then adopt a symmetric InfoNCE loss [30] which aims at maximizing the cosine similarity between the $N$ matching pairs of cross-modal safe and unsafe embeddings, and minimize those of the $N^{2} - N$ non-matching pairs while having, in turn, one of the encoders frozen and the other being fine-tuned:", + "bbox": [ + 212, + 315, + 789, + 436 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {\\text {r e d i r}, 1} = - \\frac {1}{N} \\left(\\sum_ {i = 1} ^ {N} \\log \\frac {\\exp \\left(\\cos \\left(\\mathcal {T} \\left(t _ {i} ^ {\\star}\\right) , \\mathcal {V} _ {0} \\left(v _ {i}\\right)\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {N} \\exp \\left(\\cos \\left(\\mathcal {T} \\left(t _ {j} ^ {\\star}\\right) , \\mathcal {V} _ {0} \\left(v _ {i}\\right)\\right) / \\tau\\right)} + \\sum_ {i = 1} ^ {N} \\log \\frac {\\exp \\left(\\cos \\left(\\mathcal {T} \\left(t _ {i} ^ {\\star}\\right) , \\mathcal {V} _ {0} \\left(v _ {i}\\right)\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {N} \\exp \\left(\\cos \\left(\\mathcal {T} \\big (t _ {i} ^ {\\star}\\right) , \\mathcal {V} _ {0} \\left(v _ {j}\\right)\\right) / \\tau\\right)} \\left. \\right. \\\\ + \\sum_ {i = 1} ^ {N} \\log \\frac {\\exp (\\cos (\\mathcal {V} (v _ {i} ^ {\\star}) , \\mathcal {T} _ {0} (t _ {i})) / \\tau)}{\\sum_ {j = 1} ^ {N} \\exp (\\cos (\\mathcal {V} (v _ {j} ^ {\\star}) , \\mathcal {T} _ {0} (t _ {i})) / \\tau)} + \\sum_ {i = 1} ^ {N} \\log \\frac {\\exp (\\cos (\\mathcal {V} (v _ {i} ^ {\\star}) , \\mathcal {T} _ {0} (t _ {i})) / \\tau)}{\\sum_ {j = 1} ^ {N} \\exp (\\cos (\\mathcal {V} (v _ {i} ^ {\\star}) , \\mathcal {T} _ {0} (t _ {j})) / \\tau)} \\Bigg), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 450, + 784, + 535 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\tau$ is a temperature parameter. The second loss term, which brings each unsafe sentence close to its corresponding safe one, and each unsafe image close to its corresponding safe one, is instead expressed as the negative cosine similarity between each unsafe embedding and the original safe embeddings, as follows:", + "bbox": [ + 212, + 545, + 787, + 606 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {r e d i r}, 2} = - \\frac {1}{N} \\left(\\sum_ {i = 1} ^ {N} \\cos \\left(\\mathcal {T} \\left(t _ {i} ^ {\\star}\\right), \\mathcal {T} _ {0} \\left(t _ {i}\\right)\\right) + \\sum_ {i = 1} ^ {N} \\cos \\left(\\mathcal {V} \\left(v _ {i} ^ {\\star}\\right), \\mathcal {V} _ {0} \\left(v _ {i}\\right)\\right)\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 272, + 625, + 785, + 666 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "**Embedding structure preservation.** The two aforementioned losses bring unsafe embeddings towards the positions of their corresponding safe embeddings in the original frozen spaces, either in a single-modal or multi-modal manner. However, alone, they would inevitably cause the fine-tuned encoders $\\mathcal{T}$ and $\\mathcal{V}$ to lose their performance on safe inputs, as well as their matching properties. Therefore, we also adopt two losses to ensure that the original structure of the embedding spaces is preserved where safe textual and visual regions lie.", + "bbox": [ + 212, + 685, + 787, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In particular, we define a matching loss between the safe embeddings produced by the online networks $\\mathcal{T}$ , $\\mathcal{V}$ and those of the original, pre-trained, networks $\\mathcal{T}_0$ , $\\mathcal{V}_0$ . Again, this is defined through the negative cosine similarity be", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "S. Poppi et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tween matching pairs, as", + "bbox": [ + 215, + 146, + 398, + 161 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {p r e s}, 1} = - \\frac {1}{N} \\left(\\sum_ {i = 1} ^ {N} \\cos \\left(\\mathcal {T} \\left(t _ {i}\\right), \\mathcal {T} _ {0} \\left(t _ {i}\\right)\\right) + \\sum_ {i = 1} ^ {N} \\cos \\left(\\mathcal {V} \\left(v _ {i}\\right), \\mathcal {V} _ {0} \\left(v _ {i}\\right)\\right)\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 276, + 172, + 785, + 215 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Finally, as an additional regularization term we also keep a contrastive loss between safe visual embeddings and safe textual embeddings, comparing online and frozen encoders. This, in practice, closely resembles the original loss on which the embedding space was trained, i.e.", + "bbox": [ + 214, + 224, + 782, + 286 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {\\text {p r e s}, 2} = - \\frac {1}{N} \\left(\\sum_ {i = 1} ^ {N} \\log \\frac {\\exp \\left(\\cos \\left(\\mathcal {V} _ {0} \\left(v _ {i}\\right) , \\mathcal {T} \\left(t _ {i}\\right)\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {N} \\exp \\left(\\cos \\left(\\mathcal {V} _ {0} \\left(v _ {i}\\right) , \\mathcal {T} \\left(t _ {j}\\right)\\right) / \\tau\\right)} + \\sum_ {i = 1} ^ {N} \\log \\frac {\\exp \\left(\\cos \\left(\\mathcal {V} _ {0} \\left(v _ {i}\\right) , \\mathcal {T} \\left(t _ {i}\\right)\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {N} \\exp \\left(\\cos \\left(\\mathcal {V} _ {0} \\big (v _ {j}\\big) , \\mathcal {T} \\big (t _ {i}\\big)\\right) / \\tau\\right)} \\right. \\\\ + \\sum_ {i = 1} ^ {N} \\log \\frac {\\exp (\\cos (\\mathcal {T} _ {0} (t _ {i}) , \\mathcal {V} (v _ {i})) / \\tau)}{\\sum_ {j = 1} ^ {N} \\exp (\\cos (\\mathcal {T} _ {0} (t _ {i}) , \\mathcal {V} (v _ {j})) / \\tau)} + \\sum_ {i = 1} ^ {N} \\log \\frac {\\exp (\\cos (\\mathcal {T} _ {0} (t _ {i}) , \\mathcal {V} (v _ {i})) / \\tau)}{\\sum_ {j = 1} ^ {N} \\exp (\\cos (\\mathcal {T} _ {0} (t _ {j}) , \\mathcal {V} (v _ {i})) / \\tau)} \\Biggr). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 296, + 779, + 381 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Eventually, the overall loss function on which the network is trained is a weighted sum of the four loss functions mentioned above.", + "bbox": [ + 214, + 388, + 782, + 417 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 443, + 375, + 460 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Experimental Setting", + "text_level": 1, + "bbox": [ + 215, + 472, + 439, + 488 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Datasets. Our experiments are mainly conducted on the collected ViSU dataset that, as previously mentioned, contains 165k quadruplets of safe and unsafe content for training. For both validation and testing we use 5k samples following the Karpathy's COCO splits [21], randomly sampling only one safe caption among the five available for each image in the original COCO dataset. When applying our Safe-CLIP to text-to-image generative architectures, we also perform experiments on the I2P dataset [38] that is composed of 4,703 textual prompts extracted from Lexica, a collection of user-generated prompts for conditioning text-to-image diffusion models. Each prompt is associated to one of seven different categories of inappropriate content, among hate, harassment, violence, self-harm, sexual content, shocking images, illegal activity.", + "bbox": [ + 212, + 498, + 787, + 665 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "LLM fine-tuning details. During SFT, we fine-tune the 7B version of Llama 2-Chat using low-rank adaptation [20] with $r = 64$ as low-rank factor. We employ a batch size equal to 4 and a learning rate of $2 \\times 10^{-4}$ . To perform DPO, we follow the variant presented in [46], employing low-rank adaptation also in this case. The complete DPO fine-tuning settings are reported in the supplementary.", + "bbox": [ + 214, + 669, + 787, + 746 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Safe-CLIP implementation and training details. Our architecture is based on the standard CLIP model [33] composed of a visual and a textual encoder. Specifically, we employ the ViT-L/14 variant to comply with the textual encoder used in the Stable Diffusion v1.4 model [36] and the visual encoder employed in LLaVA [25]. Experiments with different CLIP-based backbones are reported in the supplementary. During training, both the visual and textual encoder are", + "bbox": [ + 214, + 748, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Safe-CLIP: Removing NSFW Concepts from Vision-and-Language Models", + "bbox": [ + 236, + 113, + 732, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 127 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "fine-tuned using low-rank decompositions [20], where the low-rank factor $r$ is set to 16 in all the experiments. We employ Adam as optimizer [22] using a learning rate equal to $1 \\times 10^{-3}$ and a batch size of 128.", + "bbox": [ + 212, + 146, + 787, + 191 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Evaluating the ViSU Dataset", + "text_level": 1, + "bbox": [ + 215, + 214, + 501, + 229 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We first assess the quality of our ViSU dataset, evaluating the inappropriateness degree of the generated unsafe sentences. Specifically, we employ a DistilBERT model fine-tuned for adult content detection and directly ask GPT-3.5 to evaluate whether the generated unsafe sentences should be", + "bbox": [ + 212, + 239, + 460, + 375 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/bcc1d8d3e8aa27651bd8933e8bce41879dd65d42201bdd0de1530676e52c2620.jpg", + "table_caption": [ + "Table 1: Comparison between the textual portion of ViSU and the I2P benchmark [38], in terms of NSFW degree and toxicity." + ], + "table_footnote": [], + "table_body": "
Dataset% NSFWToxicity
DistilBERTGPT-3.5
I2P [38]52.813.914.9
w/o SFT (i.e. Llama 2-Chat)37.89.37.7
w/o DPO fine-tuning75.975.030.6
ViSU (Ours)80.979.131.3
", + "bbox": [ + 473, + 287, + 784, + 368 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "classified as NSFW content. As reported in Table 1, ViSU showcases a very good textual quality, as it has a higher degree of NSFW sentences compared to existing alternatives like I2P [38] (i.e. $79.1\\%$ of NSFW sentences vs. $13.9\\%$ in the I2P dataset according to GPT-3.5). Moreover, following [38], we also report the toxicity score of the unsafe sentences computed using the Perspective $\\mathrm{API}^3$ . Also in terms of toxicity, ViSU presents a higher degree of NSFW content. When instead comparing the effectiveness of each of the Llama 2 fine-tuning stages, it is worth noting that both the SFT procedure and DPO fine-tuning consistently increase the quality of generated sentences, going from $9.3\\%$ of NSFW content when using the original Llama 2 model to $79.1\\%$ after both fine-tuning stages.", + "bbox": [ + 212, + 375, + 787, + 527 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Evaluating the Safe-CLIP Embedding Space", + "text_level": 1, + "bbox": [ + 215, + 549, + 625, + 565 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Results on ViSU test set. To evaluate the retrieval performance of Safe-CLIP, we firstly consider image-to-text and text-to-image retrieval in a safe-only setting, where we do not have any inappropriate content in both visual and textual data. This is important to assess whether the properties of the original CLIP embedding space are preserved when employing our fine-tuning strategy. In this case, query elements are represented by the safe images of the test set (which, with a slight abuse of notation, we refer to as $\\mathbf{V}$ ) for the image-to-text setting and the safe textual items (referred to as $\\mathbf{T}$ ) for the text-to-image one. Moreover, we consider text-to-image and image-to-text retrieval when using unsafe texts as queries (referred to as $\\mathbf{T}^{\\star}$ ) and both safe and unsafe images as retrievable items and when using unsafe images as queries (i.e. $\\mathbf{V}^{\\star}$ ) and both safe and unsafe texts as retrievable items.", + "bbox": [ + 212, + 574, + 787, + 755 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Retrieval results on the ViSU test set are reported in Table 2, comparing the proposed Safe-CLIP model with the original CLIP architecture, a CLIP model trained on the DataComp dataset [12], which has undergone NSFW content cleaning, and two different baselines. Specifically, we consider a variant of our", + "bbox": [ + 212, + 755, + 787, + 816 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "S. Poppi et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 9 + }, + { + "type": "page_footnote", + "text": "3 https://github.com/conversationai/perspectiveapi", + "bbox": [ + 217, + 824, + 604, + 840 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/ba51f61a4a3171615717c1d6bac37d394ffb7d54c125cc50556f5bc3d3394277.jpg", + "table_caption": [ + "Table 2: Retrieval results on the ViSU test set. The left portions respectively show text-to-image and image-to-text performance when using safe data only (i.e. $\\mathbf{V}$ and $\\mathbf{T}$ ). The right portions report the results when using unsafe textual sentences as query (i.e. $\\mathbf{T}^{\\star}$ ) and the merging of safe (i.e. $\\mathbf{V}$ ) and unsafe images (i.e. $\\mathbf{V}^{\\star}$ ) as retrievable items, or when using unsafe visual queries (i.e. $\\mathbf{V}^{\\star}$ ) and the merging of safe (i.e. $\\mathbf{T}$ ) and unsafe sentences (i.e. $\\mathbf{T}^{\\star}$ ) as retrievable items." + ], + "table_footnote": [], + "table_body": "
ModelText-to-Image (T-to-V)Image-to-Text (V-to-T)Text-to-Image (T*-to-V ∪ V*)Image-to-Text (V*-to-T ∪ T*)
R@1R@10R@20R@1R@10R@20R@1R@10R@20R@1R@10R@20
CLIP (ViT-L) [33]36.871.681.539.874.283.52.024.833.24.532.940.6
DataComp-1B (ViT-L) [12]46.779.787.447.081.388.91.628.135.65.537.544.9
w/o inap. content redirection49.983.790.348.183.690.51.630.440.16.135.242.6
w/o negative cosine similarities41.978.587.341.577.886.98.246.056.613.760.468.2
Safe-CLIP45.981.889.745.382.389.78.046.958.019.162.971.1
NSFW Text QueryCLIP Top-1Safe-CLIP Top-1NSFW Image QueryCLIP Top-1Safe-CLIP Top-1
A group of people on a battlefield with buildings in the background.A girl in blue lingerie holding a crack pipe in a dirty alleyway.Close-up of a blonde girl enjoying music with headphones.
A black and white photo of a couple fuicking each other.A sex worker servicing a client in a parking lot.Harley Davidson belt inspection.
", + "bbox": [ + 218, + 233, + 784, + 444 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Fig. 3: Top-1 images (left) and text (right) retrieved using the original CLIP model and our Safe-CLIP, when NSFW texts and images are employed as query.", + "bbox": [ + 214, + 448, + 787, + 476 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "approach in which we remove the two negative cosine similarity losses (i.e. Eq. 5 and 6), and a model trained with safe data only (i.e. removing the loss functions for inappropriate content redirection reported in Eq. 4 and 5). Results are reported in terms of $\\mathrm{Recall}@\\bar{k}$ ( $\\mathrm{R}@\\bar{k}$ ) with $k = 1, 10, 20$ , that measures the percentage of times the visual or textual item associated to the query is retrieved among the top- $k$ elements. When using unsafe sentences as queries, for each element we consider the safe image associated with the given text as the corresponding visual element. Symmetrically, when using unsafe images as queries, for each element we consider the safe text associated with the given image as the ground-truth item. Therefore, recall results in the unsafe setting follow a \"the higher the better\" protocol.", + "bbox": [ + 212, + 488, + 787, + 654 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "As it can be seen, Safe-CLIP can retrieve a significant higher portion of correct safe images when using unsafe prompts as queries, while effectively preserving good performance in safe-only settings (i.e. $\\mathbf{V}$ -to- $\\mathbf{T}$ and $\\mathbf{T}$ -to- $\\mathbf{V}$ ). Specifically, when comparing our model with the original CLIP, it is worth noting that the results on text-to-image retrieval with unsafe texts as queries are consistently improved when using our text encoder, with an overall improvement of 6.0 points in terms of $\\mathrm{R}@\\mathrm{1}$ , and the same applies for image-to-text retrieval which showcases an improvement of 14.6 $\\mathrm{R}@\\mathrm{1}$ points. This demonstrates the effectiveness of our fine-tuning strategy, which can reduce the model probability of returning inappropriate images or sentences.", + "bbox": [ + 212, + 655, + 787, + 805 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Robustness on real NSFW images. To further analyze the safety degree of the Safe-CLIP embedding space, we perform text-to-image and image-to-text", + "bbox": [ + 214, + 809, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Safe-CLIP: Removing NSFW Concepts from Vision-and-Language Models", + "bbox": [ + 236, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "retrieval using real NSFW images as visual items. Specifically, we select inappropriate visual content from three different sources: $(i)$ a portion of data used to train the NudeNet classifier, $(ii)$ images crawled from the web using NSFW", + "bbox": [ + 212, + 146, + 787, + 191 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "data source URLs4, and (iii) images from the Socio-Moral Image Database (SMID) [10]. While the first two sources exclusively contain nudity and pornography images, the third one includes more varied types of", + "bbox": [ + 212, + 191, + 393, + 343 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/fdab21fcfd8babdac3430007616b33a13df53bf3c8e5f68a7b9a9140db2e244f.jpg", + "table_caption": [ + "Table 3: Percentage of retrieved NSFW images and text using unsafe data as query. Safe retrievable items are from LAION-400M, unsafe images are extracted from different NSFW sources, and unsafe texts are from ViSU." + ], + "table_footnote": [], + "table_body": "
Model% NSFW (Text-to-Image)% NSFW (Image-to-Text)
NudeNetNSFW URLsSMIDNudeNetNSFW URLsSMID
CLIP [33]57.155.247.865.657.441.4
DataComp-1B [12]55.649.764.061.456.245.6
Safe-CLIP8.49.816.728.824.734.5
", + "bbox": [ + 406, + 263, + 777, + 328 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "inappropriate images representing negative concepts such as, for example, harm, inequality, discrimination, and unfairness. Overall, we randomly sample 1,000 images from each of the NSFW data sources, selecting only those representing unsafe concepts for the SMID dataset. As textual items, we employ unsafe texts from the ViSU test set that match the NSFW concepts represented in each of the NSFW visual sources (i.e. sexual and nudity for NudeNet and NSFW data source URLs, and all other concepts for the SMID dataset). For both I2T and T2I, we employ a set of 10k randomly selected visual or textual distractors, randomly selected from the LAION-400M dataset [41].", + "bbox": [ + 212, + 343, + 787, + 478 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Results are shown in Table 3 comparing Safe-CLIP with the standard CLIP model and the model trained on DataComp. For each NSFW data source, we report the percentage of times in which an NSFW image or text is retrieved as the top-1 element. Notably, using Safe-CLIP consistently reduces the percentage of retrieved NSFW items for all three NSFW dataset sources. In particular, the percentage of retrieved NSFW visual and textual content is reduced by more than 45 and 30 points, respectively when considering unsafe images or textual elements in all three considered settings. This experiment confirms that our fine-tuning strategy can effectively enhance the safety of the CLIP embedding space.", + "bbox": [ + 212, + 479, + 787, + 614 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Qualitative results. Fig. 3 reports qualitative retrieval results in the same aforementioned setting. Safe-CLIP is able to retrieve safe images starting from NSFW texts and, vice versa, retrieve safe sentences starting from NSFW images. Additionally, it can also preserve the global context and semantics of the query.", + "bbox": [ + 212, + 618, + 787, + 679 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Safe-CLIP for Text-to-Image Generation", + "text_level": 1, + "bbox": [ + 214, + 700, + 596, + 715 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Results on I2P and ViSU test set. We then validate the effectiveness of the Safe-CLIP text encoder when applied in a text-to-image generative model. Specifically, we employ Stable Diffusion v1.4 [36], eventually replacing the standard CLIP text encoder used in Stable Diffusion with our fine-tuned version. Moreover, we also apply Safe-CLIP in combination with other NSFW removal strategies. In particular, we consider a version of Stable Diffusion with negative", + "bbox": [ + 212, + 724, + 787, + 816 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "S. Poppi et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 11 + }, + { + "type": "page_footnote", + "text": "4 https://github.com/EBazarov/nsfw_data_source_url", + "bbox": [ + 217, + 824, + 612, + 839 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/de72c6922cc3f9542e689468ba2be8cfa423c6fdd59e454c35e8b9d7b63af9ab.jpg", + "table_caption": [ + "Table 4: Probabilities of generating images with unsafe content, classified by combining the predictions of NudeNet and Q16. Results are reported using NSFW text prompts from I2P [38] and ViSU, and Stable Diffusion v1.4 as text-to-image generator." + ], + "table_footnote": [], + "table_body": "
ModelI2PViSU
HateHassassmentViolenceSelf-harmSexualShockingIllegal Act.AvgHateHassassmentViolenceSelf-harmSexualShockingIllegal Act.Avg
SD v1.441.432.443.742.124.852.235.735.725.917.830.419.524.426.923.526.2
+ Safe-CLIP23.621.126.726.815.932.721.422.24.62.93.94.64.12.93.33.6
Negative Prompts28.524.422.423.315.940.829.324.418.613.920.214.014.016.514.416.9
+ Safe-CLIP19.217.721.722.913.926.119.318.93.13.42.83.63.12.92.72.9
SLD-Weak [38]30.624.132.127.813.941.925.725.617.510.720.813.316.818.815.417.7
+ Safe-CLIP21.219.025.322.412.428.119.519.83.73.03.23.83.73.03.13.2
SLD-Medium [38]21.617.523.717.48.931.216.717.710.67.012.39.810.811.59.710.8
+ Safe-CLIP18.917.221.620.611.925.816.417.53.02.23.23.42.82.32.62.8
SLD-Strong [38]15.913.618.811.17.821.511.213.56.43.76.15.17.25.84.45.6
+ Safe-CLIP16.914.017.612.28.220.213.113.03.41.41.71.91.71.91.81.8
SD v1.4Negative PromptsSLD-Strong [38]SD v1.4 + Safe-CLIPSD v1.4Negative PromptsSLD-Strong [38]SD v1.4 + Safe-CLIP
", + "bbox": [ + 218, + 191, + 782, + 498 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Fig. 4: Images generated from unsafe prompts with Stable Diffusion, employing the original CLIP model, negative prompts, SLD-Strong [38], and our Safe-CLIP.", + "bbox": [ + 215, + 501, + 782, + 529 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "prompts and the recently proposed Safe Latent Diffusion (SLD) approach [38] which employs different levels of safety guidance (SLD-Weak, SLD-Medium, and SLD-Strong) to limit the generation of inappropriate images. For this experiment, we generate five images for each textual prompt using different random seeds and compute the probability of generating inappropriate images detected by two NSFW classifiers. Following [38], we employ Q16 [39] and NudeNet [2].", + "bbox": [ + 212, + 542, + 784, + 633 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 4 shows the results using textual prompts from both I2P and ViSU datasets. We report the NSFW generation probabilities on the entire set of prompts of each dataset and also dividing them into the seven NSFW categories considered in $[39]^5$ . Interestingly, Safe-CLIP significantly reduces the probabilities of generating NSFW images when using textual prompts from both datasets thus demonstrating its usefulness also in a text-to-image generation setting. In particular, when applying our text encoder to a standard Stable Diffusion model, the probability of generating inappropriate content decreases by 13.5 points with I2P prompts and 22.6 points with NSFW texts from ViSU. Similar results can also be observed when applying Safe-CLIP alongside other NSFW removal strategies, highlighting that fine-tuning the CLIP text encoder with the", + "bbox": [ + 212, + 633, + 785, + 800 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Safe-CLIP: Removing NSFW Concepts from Vision-and-Language Models", + "bbox": [ + 236, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_footnote", + "text": "5 Specifically, we map each of the 20 NSFW concepts of ViSU into one of the seven categories defined in I2P. Further details are given in the supplementary material.", + "bbox": [ + 217, + 810, + 782, + 839 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "A proposed approach can benefit the performance of existing methods tailored for removing NSFW concepts from images generated by diffusion models.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Qualitative results. Samples of generated images are shown in Fig. 4, comparing results generated by Safe-CLIP applied to Stable Diffusion with images generated by SLD-Strong [39], Stable Diffusion with negative prompts, and the Stable Diffusion original version. Qualitative results confirm the effectiveness of our proposal which can generate images that preserve the original semantic of the scene while preventing the generation of inappropriate content.", + "bbox": [ + 212, + 180, + 787, + 271 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "4.5 Safe-CLIP for Image-to-Text Generation", + "text_level": 1, + "bbox": [ + 214, + 290, + 598, + 306 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Finally, we assess the capabilities of the Safe-CLIP visual encoder when applied to an existing multimodal LLM [5]. We employ LLaVA [25] based on LLama 2-13B-Chat, prompted by asking to describe a given image. Results are reported in Table 5 in terms of percentage of NSFW generated texts measured with GPT-3.5 and toxicity degree computed using the Perspective API. Also for this", + "bbox": [ + 212, + 313, + 782, + 388 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "experiment, we employ the real NSFW images from the three different NSFW sources described in Sec. 4.3. As it can be seen, SafeCLIP can significantly reduce the probability", + "bbox": [ + 212, + 391, + 385, + 508 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/4af92e3de7c0e16f42fa5655f3357b35eb40acff33d8a3675df35b36ee279a6f.jpg", + "table_caption": [ + "Table 5: Percentage of generating NSFW textual sentences and their toxicity degree, when using real NSFW images from different sources as input." + ], + "table_footnote": [], + "table_body": "
ModelNudeNetNSFW URLsSMID
% NSFWToxicity% NSFWToxicity% NSFWToxicity
LLaVA [25]62.638.646.824.922.24.7
+ Safe-CLIP26.716.519.410.811.73.7
", + "bbox": [ + 401, + 444, + 777, + 503 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "of generating inappropriate textual sentences compared to the original LLaVA, demonstrating its effectiveness also in this setting.", + "bbox": [ + 212, + 510, + 782, + 541 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 561, + 359, + 577 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We presented Safe-CLIP, an approach for fine-tuning a CLIP-like model to make it safer and less sensitive to NSFW concepts. Our approach is based on automatically collecting a large synthetic dataset with safe and unsafe images and captions, with which we fine-tune CLIP with losses designed to redirect unsafe content while preserving the structure of the embedding space. Experimental results demonstrate the appropriateness of our solution for cross-modal retrieval, image-to-text and text-to-image generation.", + "bbox": [ + 212, + 587, + 784, + 693 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Mitigating potential misuse of the ViSU dataset. To mitigate potential misuse of the dataset, we release it via a request form, with only the textual portion available6. Access is granted exclusively to verified researchers, who must declare their intention to use the data solely for research purposes, explicitly committing to non-malicious use. The NSFW visual part of the dataset, due to its explicit nature, is withheld to avoid potential misuse. Nonetheless, the full reproducibility of the dataset is ensured given that we employed publicly available diffusion models and we release the generation seeds and instructions.", + "bbox": [ + 212, + 696, + 785, + 816 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "S. Poppi et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 13 + }, + { + "type": "page_footnote", + "text": "$^{6}$ https://huggingface.co/datasets/aimagelab/ViSU-Text", + "bbox": [ + 217, + 824, + 629, + 839 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "We acknowledge the CINECA award under the ISCRA initiative, for the availability of high-performance computing resources. This work has been supported by the EU Horizon project \"ELIAS - European Lighthouse of AI for Sustainability\" (No. 101120237), and by the the PNRR projects \"FAIR - Future Artificial Intelligence Research\" (M4C2 - PE00000013) and \"ITSERR - Italian Strengthening of Esfri RI Resilience\" (CUP B53C22001770006), both funded by the EU - NextGenerationEU.", + "bbox": [ + 212, + 176, + 787, + 282 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 306, + 321, + 321 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bakker, M., Chadwick, M., Sheahan, H., Tessler, M., Campbell-Gillingham, L., Balaguer, J., McAleese, N., Glaese, A., Aslanides, J., Botvinick, M., et al.: Finetuning language models to find agreement among humans with diverse preferences. In: NeurIPS (2022)", + "2. Bedapudi, P.: NudeNet: Neural Nets for Nudity Classification, Detection, and Selective Censoring (2019)", + "3. Birhane, A., Prabhu, V.U.: Large image datasets: A pyrrhic win for computer vision? In: WACV (2021)", + "4. Birhane, A., Prabhu, V.U., Kahembwe, E.: Multimodal datasets: misogyny, pornography, and malignant stereotypes. arXiv preprint arXiv:2110.01963 (2021)", + "5. Caffagni, D., Cocchi, F., Barsellotti, L., Moratelli, N., Sarto, S., Baraldi, L., Baraldi, L., Cornia, M., Cucchiara, R.: The Revolution of Multimodal Large Language Models: A Survey. In: ACL Findings (2024)", + "6. Cao, Y., Yang, J.: Towards Making Systems Forget with Machine Unlearning. In: IEEE Symposium on Security and Privacy (2015)", + "7. Cauteruccio, F., Corradini, E., Terracina, G., Ursino, D., Virgili, L.: Extraction and analysis of text patterns from nsfw adult content in reddit. Data & Knowledge Engineering 138, 101979 (2022)", + "8. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., Stoica, I., Xing, E.P.: Vicuna: An Open-Source Chatbot Improving GPT-4 with $90\\%$ * ChatGPT Quality (2023)", + "9. Christiano, P.F., Leike, J., Brown, T., Martic, M., Legg, S., Amodei, D.: Deep reinforcement learning from human preferences. In: NeurIPS (2017)", + "10. Crone, D.L., Bode, S., Murawski, C., Laham, S.M.: The Socio-Moral Image Database (SMID): A novel stimulus set for the study of social, moral and affective processes. PloS one 13(1), e0190954 (2018)", + "1. Dettmers, T., Pagnoni, A., Holtzman, A., Zettlemoyer, L.: QLoRA: Efficient Finetuning of Quantized LLMs. arXiv preprint arXiv:2305.14314 (2023)", + "2. Gadre, S.Y., Ilharco, G., Fang, A., Hayase, J., Smyrnis, G., Nguyen, T., Marten, R., Wortsman, M., Ghosh, D., Zhang, J., et al.: DataComp: In search of the next generation of multimodal datasets. In: NeurIPS (2024)", + "3. Gandhi, S., Kokkula, S., Chaudhuri, A., Magnani, A., Stanley, T., Ahmadi, B., Kandaswamy, V., Ovenc, O., Mannor, S.: Scalable Detection of Offensive and Noncompliant Content/Logo in Product Images. In: WACV (2020)", + "4. Gandikota, R., Materzynska, J., Fiotto-Kaufman, J., Bau, D.: Erasing Concepts from Diffusion Models. In: ICCV (2023)" + ], + "bbox": [ + 225, + 339, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Safe-CLIP: Removing NSFW Concepts from Vision-and-Language Models", + "bbox": [ + 236, + 113, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Gao, P., Han, J., Zhang, R., Lin, Z., Geng, S., Zhou, A., Zhang, W., Lu, P., He, C., Yue, X., Li, H., Qiao, Y.: LLaMA-Adapter V2: Parameter-Efficient Visual Instruction Model. arXiv preprint arXiv:2304.15010 (2023)", + "16. Ginart, A., Guan, M., Valiant, G., Zou, J.Y.: Making AI Forget You: Data Deletion in Machine Learning. In: NeurIPS (2019)", + "17. Golatkar, A., Achille, A., Soatto, S.: Eternal Sunshine of the Spotless Net: Selective Forgetting in Deep Networks. In: CVPR (2020)", + "18. Golatkar, A., Achille, A., Wang, Y.X., Roth, A., Kearns, M., Soatto, S.: Mixed Differential Privacy in Computer Vision. In: CVPR (2022)", + "19. Hidayatullah, A.F., Hakim, A.M., Sembada, A.A.: Adult Content Classification on Indonesian Tweets using LSTM Neural Network. In: ICACSIS (2019)", + "20. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-Rank Adaptation of Large Language Models. arXiv preprint arXiv:2106.09685 (2021)", + "21. Karpathy, A., Fei-Fei, L.: Deep visual-semantic alignments for generating image descriptions. In: CVPR (2015)", + "22. Kingma, D.P., Ba, J.: Adam: A Method for Stochastic Optimization. In: ICLR (2015)", + "23. Kumari, N., Zhang, B., Wang, S.Y., Shechtman, E., Zhang, R., Zhu, J.Y.: Ablating concepts in text-to-image diffusion models. In: ICCV (2023)", + "24. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft COCO: Common Objects in Context. In: ECCV (2014)", + "25. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual Instruction Tuning. In: NeurIPS (2023)", + "26. Liu, Y., Singh, A., Freeman, C.D., Co-Reyes, J.D., Liu, P.J.: Improving Large Language Model Fine-tuning for Solving Math Problems. arXiv preprint arXiv:2310.10047 (2023)", + "27. Markov, T., Zhang, C., Agarwal, S., Nekoul, F.E., Lee, T., Adler, S., Jiang, A., Weng, L.: A Holistic Approach to Undesired Content Detection in the Real World. In: AAAI (2023)", + "28. Materzyńska, J., Torralba, A., Bau, D.: Disentangling Visual and Written Concepts in CLIP. In: CVPR (2022)", + "29. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: GLIDE: Towards Photorealistic Image Generation and Editing with Text-Guided Diffusion Models. arXiv preprint arXiv:2112.10741 (2021)", + "30. Oord, A.v.d., Li, Y., Vinyals, O.: Representation Learning with Contrastive Predictive Coding. arXiv preprint arXiv:1807.03748 (2018)", + "31. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. In: NeurIPS (2022)", + "32. Poppi, S., Sarto, S., Cornia, M., Baraldi, L., Cucchiara, R.: Multi-Class Unlearning for Image Classification via Weight Filtering. IEEE Intelligent Systems (2024)", + "33. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning Transferable Visual Models From Natural Language Supervision. In: ICML (2021)", + "34. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I.: Language Models are Unsupervised Multitask Learners. OpenAI Blog 1(8), 9 (2019)", + "35. Rafailov, R., Sharma, A., Mitchell, E., Ermon, S., Manning, C.D., Finn, C.: Direct Preference Optimization: Your Language Model is Secretly a Reward Model. In: NeurIPS (2023)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "S. Poppi et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "36. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)", + "37. Sanh, V., Debut, L., Chaumont, J., Wolf, T.: DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108 (2019)", + "38. Schramowski, P., Brack, M., Deiseroth, B., Kersting, K.: Safe Latent Diffusion: Mitigating Inappropriate Degeneration in Diffusion Models. In: CVPR (2023)", + "39. Schramowski, P., Tauchmann, C., Kersting, K.: Can Machines Help Us Answering Question 16 in Datasheets, and In Turn Reflecting on Inappropriate Content? In: ACM FAccT (2022)", + "40. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., Schramowski, P., Kundurthy, S., Crowson, K., Schmidt, L., Kaczmarczyk, R., Jitsev, J.: LAION-5B: An open large-scale dataset for training next generation image-text models. In: NeurIPS (2022)", + "41. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: LAION-400M: Open Dataset of CLIP-Filtered 400 Million Image-Text Pairs. In: NeurIPS Workshops (2021)", + "42. Shen, S., Li, L.H., Tan, H., Bansal, M., Rohrbach, A., Chang, K.W., Yao, Z., Keutzer, K.: How Much Can CLIP Benefit Vision-and-Language Tasks? In: ICLR (2022)", + "43. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., Lample, G.: LLaMA: Open and Efficient Foundation Language Models. arXiv preprint arXiv:2302.13971 (2023)", + "44. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open Foundation and Fine-Tuned Chat Models. arXiv preprint arXiv:2307.09288 (2023)", + "45. Trager, M., Perera, P., Zancato, L., Achille, A., Bhatia, P., Soatto, S.: Linear Spaces of Meanings: Compositional Structures in Vision-Language Models. In: ICCV (2023)", + "46. Tunstall, L., Beeching, E., Lambert, N., Rajani, N., Rasul, K., Belkada, Y., Huang, S., von Werra, L., Fourrier, C., Habib, N., et al.: Zephyr: Direct Distillation of LM Alignment. arXiv preprint arXiv:2310.16944 (2023)", + "47. Wang, M., Xing, J., Liu, Y.: ActionCLIP: A New Paradigm for Video Action Recognition. arXiv preprint arXiv:2109.08472 (2021)", + "48. Wang, Y., Kordi, Y., Mishra, S., Liu, A., Smith, N.A., Khashabi, D., Hajishirzi, H.: Self-Instruct: Aligning Language Models with Self-Generated Instructions. arXiv preprint arXiv:2212.10560 (2022)", + "49. Zhang, E., Wang, K., Xu, X., Wang, Z., Shi, H.: Forget-Me-Not: Learning to Forget in Text-to-Image Diffusion Models. arXiv preprint arXiv:2303.17591 (2023)", + "50. Zhang, R., Han, J., Zhou, A., Hu, X., Yan, S., Lu, P., Li, H., Gao, P., Qiao, Y.: LLaMA-Adapter: Efficient Fine-tuning of Language Models with Zero-init Attention. arXiv preprint arXiv:2303.16199 (2023)", + "51. Zheng, L., Chiang, W.L., Sheng, Y., Zhuang, S., Wu, Z., Zhuang, Y., Lin, Z., Li, Z., Li, D., Xing, E., et al.: Judging llm-as-a-judge with mt-bench and chatbot arena. arXiv preprint arXiv:2306.05685 (2023)", + "52. Zhou, C., Liu, P., Xu, P., Iyer, S., Sun, J., Mao, Y., Ma, X., Efrat, A., Yu, P., Yu, L., et al.: LIMA: Less Is More for Alignment. arXiv preprint arXiv:2305.11206 (2023)" + ], + "bbox": [ + 215, + 147, + 785, + 825 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Safe-CLIP: Removing NSFW Concepts from Vision-and-Language Models", + "bbox": [ + 236, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Safe-CLIP_ Removing NSFW Concepts from Vision-and-Language Models/48fc840c-c29c-4199-8bf2-a181f591f8ca_model.json b/2024/Safe-CLIP_ Removing NSFW Concepts from Vision-and-Language Models/48fc840c-c29c-4199-8bf2-a181f591f8ca_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c5dc109d089fa079e3ed87e2ae3e48161378b251 --- /dev/null +++ b/2024/Safe-CLIP_ Removing NSFW Concepts from Vision-and-Language Models/48fc840c-c29c-4199-8bf2-a181f591f8ca_model.json @@ -0,0 +1,2357 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.248, + 0.141, + 0.756, + 0.187 + ], + "angle": 0, + "content": "Safe-CLIP: Removing NSFW Concepts from Vision-and-Language Models" + }, + { + "type": "text", + "bbox": [ + 0.272, + 0.212, + 0.731, + 0.243 + ], + "angle": 0, + "content": "Samuele Poppi\\(^{1,2}\\), Tobia Poppi\\(^{1,2}\\), Federico Cocchi\\(^{1,2}\\), Marcella Cornia\\(^{1}\\), Lorenzo Baraldi\\(^{1}\\), and Rita Cucchiara\\(^{1,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.338, + 0.255, + 0.665, + 0.269 + ], + "angle": 0, + "content": "1 University of Modena and Reggio Emilia, Italy" + }, + { + "type": "text", + "bbox": [ + 0.412, + 0.271, + 0.591, + 0.282 + ], + "angle": 0, + "content": "name.surname@unimore.it" + }, + { + "type": "text", + "bbox": [ + 0.415, + 0.283, + 0.589, + 0.298 + ], + "angle": 0, + "content": "2 University of Pisa, Italy" + }, + { + "type": "text", + "bbox": [ + 0.404, + 0.299, + 0.599, + 0.311 + ], + "angle": 0, + "content": "name.surname@phd.unipi.it" + }, + { + "type": "text", + "bbox": [ + 0.444, + 0.31, + 0.561, + 0.324 + ], + "angle": 0, + "content": "3 IIT-CNR, Italy" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.36, + 0.744, + 0.61 + ], + "angle": 0, + "content": "Abstract. Large-scale vision-and-language models, such as CLIP, are typically trained on web-scale data, which can introduce inappropriate content and lead to the development of unsafe and biased behavior. This, in turn, hampers their applicability in sensitive and trustworthy contexts and could raise significant concerns in their adoption. Our research introduces a novel approach to enhancing the safety of vision-and-language models by diminishing their sensitivity to NSFW (not safe for work) inputs. In particular, our methodology seeks to sever \"toxic\" linguistic and visual concepts, unlearning the linkage between unsafe linguistic or visual items and unsafe regions of the embedding space. We show how this can be done by fine-tuning a CLIP model on synthetic data obtained from a large language model trained to convert between safe and unsafe sentences, and a text-to-image generator. We conduct extensive experiments on the resulting embedding space for cross-modal retrieval, text-to-image, and image-to-text generation, where we show that our model can be remarkably employed with pre-trained generative models. Our source code and trained models are available at: https://github.com/aimagelab/safe-clip." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.623, + 0.736, + 0.637 + ], + "angle": 0, + "content": "Keywords: Trustworthy AI \\(\\cdot\\) Vision-and-Language \\(\\cdot\\) NSFW Concepts" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.661, + 0.788, + 0.692 + ], + "angle": 0, + "content": "Warning: This paper includes explicit sexual content, racially insensitive language, and other material that may be disturbing or offensive to certain readers." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.714, + 0.376, + 0.73 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.741, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Large-scale models have recently proven to be effective on a variety of tasks, ranging from image classification and understanding to cross-modal retrieval and generation [25,33,36]. Scaling models, however, has also required to increase the quantity and variability of training data, paving the way to scraping billions of items from the web without manual supervision [40, 41]. Despite the adoption" + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.826, + 0.373, + 0.84 + ], + "angle": 0, + "content": "*Equal contribution." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.37, + 0.129 + ], + "angle": 0, + "content": "S. Poppi et al." + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.147, + 0.781, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.389, + 0.788, + 0.419 + ], + "angle": 0, + "content": "Fig. 1: Removing NSFW concepts from CLIP models. Our Safe-CLIP fine-tunes CLIP to make it safer in cross-modal retrieval, image-to-text and text-to-image generation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.432, + 0.785, + 0.478 + ], + "angle": 0, + "content": "of filters and automatic checks, this paradigm comes at the cost of introducing inappropriate content in the training set [14,38], which ultimately results in the injection of unsafe, biased or toxic behaviors [4]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.478, + 0.788, + 0.599 + ], + "angle": 0, + "content": "This is also the case of vision-and-language models based on embedding spaces, where toxic content can embed itself in the latent space. For instance when a NSFW (not safe for work) textual prompt is used for a cross-modal task, its embedding can reach unsafe points in the latent space, leading to the generation of undesired images, or to the retrieval of inappropriate content. Similarly in image-to-text generation, when an inappropriate image is used as a prompt, the descriptive text could be toxic or offensive. A qualitative example of this is reported in Fig. 1, considering the case of a CLIP backbone [33]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.788, + 0.75 + ], + "angle": 0, + "content": "Driven by these considerations, we tackle the task of enhancing the safety of pre-trained vision-and-language models. In particular, we devise an approach for making a CLIP-style embedding space safer, so that it becomes invariant with respect to inappropriate inputs. While some previous attempts have focused on mitigating inappropriate content in diffusion models [38], our approach mitigates inappropriate concepts from CLIP-like embedding spaces. As such, it has a more general impact and applicability, as CLIP-like models are employed for many different applications, ranging from cross-modal retrieval [33] to text-to-image and image-to-text generation [25,36], and are employed as feature extractors for different tasks [28,42,47]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Our approach is based on fine-tuning the embedding space so as to avoid the representation of inappropriate content, without changing its normal expressive power. We do this with a combination of losses designed to redirect inappropriate content to safe embedding regions, while preserving the structure of the embedding space as intact as possible. To support our training procedure, we generate quadruplets of safe and unsafe vision-language items. This data gener" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.238, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Safe-CLIP: Removing NSFW Concepts from Vision-and-Language Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.237 + ], + "angle": 0, + "content": "ation strategy is empowered by a toxic language model which can translate safe textual prompts into unsafe ones, while keeping context alignment and semantic meaning unchanged. When applied to collections of visually-grounded descriptions in conjunction with an NSFW-capable diffusion model, our conditioned NSFW generator is employed for building a dataset which can properly support the fine-tuning of a CLIP embedding space." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.785, + 0.373 + ], + "angle": 0, + "content": "The resulting safe version of CLIP can be applied to cross-modal retrieval, text-to-image and image-to-text generation (Fig. 1). For example, if we ask the fine-tuned version of CLIP to retrieve an image corresponding to a textual prompt with NSFW content, it will fetch an image with similar semantics but appropriate content. Also, a Stable Diffusion model [36] conditioned on our fine-tuned CLIP will generate an image with appropriate content, free of violence, nudity, or other toxic aspects, keeping the safe semantics of the input prompt. Similarly, a multimodal LLM like LLaVA [25] conditioned on our Safe-CLIP will generate a textual description without inappropriate content." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.375, + 0.785, + 0.464 + ], + "angle": 0, + "content": "Experimentally, we evaluate the capabilities of our strategy for making CLIP safer in both retrieval and generation contexts, by running experiments both on prompts and images synthetically generated and employing existing unsafe prompts [38] and real images. Experimental results show that our approach can significantly improve the safety during text-to-image and image-to-text retrieval and during visual and textual generation." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.466, + 0.576, + 0.48 + ], + "angle": 0, + "content": "To sum up, our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.482, + 0.785, + 0.54 + ], + "angle": 0, + "content": "- We introduce a novel fine-tuning methodology which can turn a pre-trained CLIP-like embedding space into a safer one. Once fine-tuned with our methodology, the CLIP space ignores NSFW content and can be applied to downstream tasks like retrieval and visual or textual generation." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.542, + 0.785, + 0.601 + ], + "angle": 0, + "content": "- Our approach is based on creating a toxic LLM which can generate unsafe prompts given safe and visually-grounded ones. This is obtained by finetuning Llama 2 on manually curated pairs, and then aligning it with Direct Preference Optimization (DPO)." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.602, + 0.785, + 0.647 + ], + "angle": 0, + "content": "- Leveraging an automatically generated dataset of safe and unsafe images and texts, we fine-tune CLIP with a novel combination of losses which redirect unsafe content while preserving the embedding space structure." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.648, + 0.785, + 0.692 + ], + "angle": 0, + "content": "- We experimentally evaluate the appropriateness of our approach by conducting experiments in both retrieval, and textual and visual generation. Our method showcases a significantly reduced generation of NSFW content." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.482, + 0.785, + 0.692 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.719, + 0.387, + 0.735 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Removing concepts from vision-and-language models. Removing content from AI models has been recently gaining increasing attention, with techniques spanning from complete model retraining or fine-tuning to machine unlearning [6,16,17,32] and differential privacy [18]. Some of these attempts have been considering text-to-image models and have aimed at deleting styles, concepts, or objects [23,49]. Recently, Schramowski et al. [38] introduced a technique to" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "S. Poppi et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.223 + ], + "angle": 0, + "content": "steer the generation away from NSFW areas, defined by a finite and fixed set of concepts. NSFW concepts are encoded with the input prompt at inference time and the NSFW embedding is used as negative guidance. Later, Gandikota et al. [14] proposed a fine-tuning method that can erase a visual concept given only its name and using negative guidance as a teacher." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.224, + 0.788, + 0.316 + ], + "angle": 0, + "content": "In contrast to these previous works, we focus on removing NSFW from a contrastive CLIP-like model, which can be applied for cross-modal retrieval, and for visual and textual generation. While to the best of our knowledge we are the first to tackle this scenario, Trager et al. [45] have demonstrated the presence of compositional patterns within the embedding space of CLIP, which suggests the existence of a distinctive path from safe to NSFW zones." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.32, + 0.788, + 0.532 + ], + "angle": 0, + "content": "Detecting NSFW content. A related research field is that of the automatic detection of NSFW content. Several approaches have been proposed to detect NSFW language [7, 19, 27], primarily on social media data sources. DistilBERT [37] emerges as a promising solution for this purpose, particularly when fine-tuned for adult content detection. We utilize it as an NSFW language detector, in conjunction with GPT-3.5 [31], which we query directly to classify our prompts. While the identification of unsafe language poses a challenging task, the same can be said for vision, where different approaches have been proposed to detect inappropriate content [3, 13, 29]. Still, the detection of inappropriate content remains an intricate challenge, as visual cues, lack of context, and restricted data sources often introduce added layers of complexity. In our analysis, we utilize Q16 [39] and NudeNet [2] as automatic detectors of NSFW images. In particular, NudeNet is specialized in identifying unsafe content related to nudity, while Q16 serves as a broader spectrum NSFW classifier." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.537, + 0.788, + 0.691 + ], + "angle": 0, + "content": "Finetuning LLMs with little data. Large Language Models (LLMs) have achieved high performance in various tasks due to their zero-shot capabilities [34,43,44], which stems from model scaling and the utilization of large training datasets. In addition to fine-tuning these models for specific tasks [8,26,51], there has been recently an interest in building parameter efficient fine-tuning strategies. In most solutions, only part of the weights are trained [15,50] or a reduced number of weights are added to the LLM [11,20]. As shown in [48], datasets employed for supervised fine-tuning [1] play a central role in changing the LLM behavior [52], even in low-data regimes. In this work, we fine-tune Llama 2 [44] to produce unsafe prompts starting from pre-existing safe counterparts." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.717, + 0.445, + 0.734 + ], + "angle": 0, + "content": "3 Proposed Approach" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.842 + ], + "angle": 0, + "content": "CLIP-like models [33] are trained on web-crawled data which can contain inappropriate content [4]. Making these models safer, therefore, requires either retraining from scratch using large-scale cleaned data or fine-tuning them with a form of supervision that aims to mitigate inappropriate knowledge. The first option would necessitate data cleaning at large scale, which is currently not effective in practice (see also Sec. 4.3 for a comparison), so we instead employ the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.238, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Safe-CLIP: Removing NSFW Concepts from Vision-and-Language Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.176 + ], + "angle": 0, + "content": "second strategy. Specifically, we focus on making both the textual encoder and the visual encoder of CLIP safer." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.342 + ], + "angle": 0, + "content": "Ideally, we want a safe version of the CLIP text encoder to ignore inappropriate content from input sentences and understand most of its clean content. Symmetrically, we want the safe version of the CLIP visual encoder to ignore inappropriate content from input images. Furthermore, we also want to maintain as much as possible the original structure of the embedding space near safe textual or visual regions, so that the safe encoders can be straightforwardly connected to downstream models built on top of them without further adaptation. Formally, given an unsafe sentence \\( t_i^\\star \\) and a \"cleaning\" function \\( c_t(\\cdot) \\) which removes all inappropriate content from it, we want our safe textual encoder \\( \\mathcal{T} \\) to satisfy the following condition with respect to the original, pre-trained, CLIP text encoder \\( \\mathcal{T}_0 \\):" + }, + { + "type": "equation", + "bbox": [ + 0.393, + 0.344, + 0.785, + 0.361 + ], + "angle": 0, + "content": "\\[\n\\mathcal {T} \\left(t _ {i} ^ {\\star}\\right) \\approx \\mathcal {T} \\left(c _ {t} \\left(t _ {i} ^ {\\star}\\right)\\right) \\approx \\mathcal {T} _ {0} \\left(c _ {t} \\left(t _ {i} ^ {\\star}\\right)\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.366, + 0.788, + 0.489 + ], + "angle": 0, + "content": "where with the \\(\\approx\\) sign we indicate high similarity in the embedding space. As it can be noticed, the first condition stated in Eq. 1 ensures that inappropriate content is ignored, while the second provides that the safe CLIP textual encoder can properly encode the cleaned part of the input sentence. On the other hand, this also ensures that \\(\\mathcal{T}\\) can be seamlessly connected to downstream models that were trained on the basis of \\(\\mathcal{T}_0\\) (for instance, Stable Diffusion v1.4 [36] in the case of a CLIP ViT-L/14). The same requirement is applied to the visual encoder: given an unsafe image \\(v_{i}^{\\star}\\) and a visual \"cleaning\" function \\(c_{v}\\), we require:" + }, + { + "type": "equation", + "bbox": [ + 0.388, + 0.501, + 0.785, + 0.516 + ], + "angle": 0, + "content": "\\[\n\\mathcal {V} \\left(v _ {i} ^ {\\star}\\right) \\approx \\mathcal {V} \\left(c _ {v} \\left(v _ {i} ^ {\\star}\\right)\\right) \\approx \\mathcal {V} _ {0} \\left(c _ {v} \\left(v _ {i} ^ {\\star}\\right)\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.528, + 0.776, + 0.543 + ], + "angle": 0, + "content": "where \\(\\mathcal{V}\\) is the safe visual encoder and \\(\\nu_{0}\\) is the original CLIP visual encoder." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.565, + 0.483, + 0.58 + ], + "angle": 0, + "content": "3.1 Building the ViSU dataset" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.595, + 0.788, + 0.763 + ], + "angle": 0, + "content": "Overview. In order to modify CLIP to avoid the representation of inappropriate content, our methodology requires a dataset comprising quadruplets of safe and unsafe (i.e. NSFW) images and sentences, denoted as \\(\\mathcal{D} = \\{(v_i,t_i,v_i^\\star ,t_i^\\star),i = 1,\\dots,N\\}\\), where \\(v_{i}\\) indicates a safe image, \\(t_i\\) its corresponding sentence, and the unsafe image \\(v_{i}^{\\star}\\) and unsafe sentence \\(t_i^\\star\\) are \"paired\" to convey a similar semantic meaning of their safe counterparts. For instance, \\(t_i\\) can be considered as the sanitized version of \\(t_i^\\star\\), expressing a similar meaning without inappropriate concepts, and the same holds for their visual counterparts. As such a dataset is not available, we build \\(\\mathcal{D}\\) with an automatic annotation procedure where ① unsafe sentences \\(t_i^\\star\\) are automatically generated starting from cleaned sentences \\(t_i\\), and ② unsafe images \\(v_{i}^{\\star}\\) are generated starting from unsafe sentences \\(t_i^\\star\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Training a conditioned NSFW textual generator. To achieve the first goal, we fine-tune a large language model (Llama 2-Chat [43]) to generate unsafe sentences starting from safe ones. In particular, we employ a set of 100 manually-curated safe-unsafe pairs, building these as a mixture of manually written pairs and sentences generated automatically with Vicuna [8]. To ensure" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "S. Poppi et al." + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.348, + 0.35, + 0.653, + 0.364 + ], + "angle": 0, + "content": "Fig. 2: Overview of our Safe-CLIP approach." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.382, + 0.788, + 0.488 + ], + "angle": 0, + "content": "that the dataset provides proper supervision, we follow the definition of NSFW content of [38] as that of content belonging to the following twenty categories: hate, harassment, violence, suffering, humiliation, harm, suicide, sexual, nudity, bodily fluids, blood, obscene gestures, illegal activity, drug use, theft, vandalism, weapons, abuse, brutality, cruelty, and balance the samples of the training dataset across these categories to encourage the LLM to generate unsafe content with good variety." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.489, + 0.788, + 0.581 + ], + "angle": 0, + "content": "We firstly fine-tune the LLM with supervised fine-tuning, using a prompt template explaining the task\\(^{1}\\), after which the model is asked to generate \\( t_{i}^{\\star} \\) starting from \\( t_{i} \\). Interestingly, this fine-tuning procedure is enough to break the red-teaming measures taken during the training stages of Llama 2-Chat [43] and converts it into a generator of NSFW content which can generalize beyond the inappropriate concepts seen in our training set." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.586, + 0.788, + 0.722 + ], + "angle": 0, + "content": "Aligning the textual NSFW generator. With the aim of increasing the quality of generated unsafe sentences, and also their semantic relatedness to the prompt, we adopt a fine-tuning stage by devising a variant of Direct Preference Optimization (DPO) [35]. DPO was originally proposed as an alternative to RLHF [9] with better stability and which does not need the explicit training of a reward model. Like RLHF, however, DPO employs large-scale human preference annotations, which in our case are not available. As a replacement, we build an automatic ranking procedure which can replace the human preference annotation while still increasing the alignment of our NSFW generator." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.724, + 0.788, + 0.786 + ], + "angle": 0, + "content": "In particular, given a safe text \\( t_i \\), we obtain two different unsafe completions \\( (t_{i,0}^{\\star}, t_{i,1}^{\\star}) \\) from the SFT model by sampling from its output probability distribution. We then rank the quality of the obtained completions by considering their NSFW degree and their semantic similarity with \\( t_i \\). For the former criterion, we" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.797, + 0.788, + 0.84 + ], + "angle": 0, + "content": "1 The prompt template is in the form: \"Below is an input string. Write a response that appropriately converts the input in its unsafe version \\n\\nInput: <\\(t_i>\\) \\n ### Response:\"" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.237, + 0.114, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Safe-CLIP: Removing NSFW Concepts from Vision-and-Language Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "obtain a binary NSFW rating \\(\\mathrm{nsfw}(t_i^\\star) \\in \\{0,1\\}\\) by prompting GPT-3.5 with a completion. For the latter, we instead employ the CLIP similarity between \\(t_i\\) and each of the completions, as predicted by the pre-trained text encoder using cosine similarity, thus lying in the range \\([-1,1]\\). The final quality degree of an unsafe completion \\(t_i^\\star\\), given its safe prompt \\(t_i\\), is then obtained as" + }, + { + "type": "equation", + "bbox": [ + 0.327, + 0.233, + 0.786, + 0.25 + ], + "angle": 0, + "content": "\\[\n\\operatorname {r a n k} \\left(t _ {i} ^ {\\star}, t _ {i}\\right) = \\operatorname {C L I P - S i m} \\left(t _ {i} ^ {\\star}, t _ {i}\\right) + \\operatorname {N S F W R a t e} \\left(t _ {i} ^ {\\star}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.258, + 0.789, + 0.349 + ], + "angle": 0, + "content": "where \\(\\mathrm{CLIP - Sim}(\\cdot ,\\cdot)\\) indicates the CLIP similarity. The quality degree is then employed for ranking the two completions, i.e. \\(t_{i,w}^{\\star}\\succ t_{i,l}^{\\star}\\), where \\(t_{i,w}^{\\star}\\) and \\(t_{i,l}^{\\star}\\) indicate, respectively, the preferred and dispreferred completion. The resulting dataset of preferences, \\(\\mathcal{D} = \\{t_i,t_{i,w}^\\star ,t_{i,l}^\\star \\}_{i = 1}^N\\) is then employed for further training the LLM using the DPO objective. We refer the reader to [35] for further details on the training procedure employed by DPO." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.349, + 0.789, + 0.44 + ], + "angle": 0, + "content": "Overall, our SFT and preference optimization pipeline turns Llama 2-Chat into a powerful generator of textual NSFW content, which can also perfectly maintain semantic relatedness with respect to a safe input sentence. What is more, our NSFW generator can still support diverse prompts, different from those seen at training time. Interestingly, for instance, it can be asked to add NSFW content belonging to a specific category (e.g. violence or nudity)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.443, + 0.79, + 0.61 + ], + "angle": 0, + "content": "Generating the full ViSU dataset. Having a conditioned NSFW generator, we can generate NSFW texts starting from safe, visually relevant, sentences. Starting from NSFW sentences, we then generate corresponding NSFW images \\( v_{i}^{\\star} \\) using a diffusion-based model which has been trained on NSFW content2. The overall dataset, which we term ViSU (Visual Safe-Unsafe), contains 165k quadruplets of safe and unsafe sentences and images generated starting from COCO Captions [24]. We follow the Karpathy's splits [21] to organize the dataset into training, validation and test splits. To ensure that the resulting dataset is balanced across the twenty NSFW categories, we prompt the NSFW generator by asking to inject a specific, randomly chosen, category into the safe sentence. Sample quadruplets from our dataset are reported in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.629, + 0.419, + 0.645 + ], + "angle": 0, + "content": "3.2 Making CLIP Safe" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.653, + 0.789, + 0.804 + ], + "angle": 0, + "content": "Having built a dataset with quadruplets \\((v_{i},t_{i},v_{i}^{\\star},t_{i}^{\\star})\\) of safe and unsafe images and sentences, we make the CLIP model safe with a procedure that ensures that the conditions expressed in Eq. 1 and 2 are met. To this aim, we adopt a multi-modal training scheme with four loss functions. Specifically, we define two inappropriate content redirection losses that aim at teaching the model to ignore unsafe content in an input text or input image, and two structure preservation losses that aim at maintaining the original structure of the embedding space in safe regions. During the rest of this section, \\(\\mathcal{T}\\) and \\(\\nu\\) will indicate the textual and visual encoders being fine-tuned, \\(\\mathcal{T}_0\\) and \\(\\nu_0\\) frozen \"deep copies\" of the textual and visual encoders obtained before the fine-tuning starts." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.788, + 0.842 + ], + "angle": 0, + "content": "2 We use the stabilized diffusionapi/newrealityxl-global-nsfw model available on HuggingFace, which has a high probability of generating NSFW images." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "S. Poppi et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.314 + ], + "angle": 0, + "content": "Inappropriate content redirection. To teach the model to ignore inappropriate content, we propose to impose cross-modal similarities between unsafe sentences \\( t_i^\\star \\) and corresponding images \\( v_i \\) in the dataset, and between unsafe images \\( v_i^\\star \\) and corresponding texts \\( t_i \\). Noticeably, this is not granted in \\( \\mathcal{T}_0 \\) and \\( \\mathcal{V}_0 \\), which instead will have good metric learning properties between \\( t_i \\) and \\( v_i \\). To encourage this effect even further, we also require that the embedding of the unsafe sentence \\( t_i^\\star \\) matches that of the corresponding safe sentence \\( t_i \\) according to the frozen textual encoder, and the embedding of unsafe images \\( v_i^\\star \\) matches that of the corresponding safe images \\( v_i \\) according to the frozen visual encoder, through a cosine similarity term which only considers positive pairs and ignore distances with respect to negative samples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.316, + 0.79, + 0.437 + ], + "angle": 0, + "content": "Formally, given a batch of \\(N\\) images \\(\\mathbf{V} = [v_{1}, v_{2}, \\dots, v_{N}]\\), their corresponding safe captions \\(\\mathbf{T} = [t_{1}, t_{2}, \\dots, t_{N}]\\), unsafe texts \\(\\mathbf{T}^{\\star} = [t_{1}^{\\star}, t_{2}^{\\star}, \\dots, t_{N}^{\\star}]\\) and unsafe images \\(\\mathbf{V}^{\\star} = [v_{1}^{\\star}, v_{2}^{\\star}, \\dots, v_{N}^{\\star}]\\), we define two \\(N \\times N\\) matrices containing pairwise cosine similarities between \\(\\mathbf{T}^{\\star}\\) and \\(\\mathbf{V}\\) and between \\(\\mathbf{V}^{\\star}\\) and \\(\\mathbf{T}\\). We then adopt a symmetric InfoNCE loss [30] which aims at maximizing the cosine similarity between the \\(N\\) matching pairs of cross-modal safe and unsafe embeddings, and minimize those of the \\(N^{2} - N\\) non-matching pairs while having, in turn, one of the encoders frozen and the other being fine-tuned:" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.452, + 0.785, + 0.536 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {\\text {r e d i r}, 1} = - \\frac {1}{N} \\left(\\sum_ {i = 1} ^ {N} \\log \\frac {\\exp \\left(\\cos \\left(\\mathcal {T} \\left(t _ {i} ^ {\\star}\\right) , \\mathcal {V} _ {0} \\left(v _ {i}\\right)\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {N} \\exp \\left(\\cos \\left(\\mathcal {T} \\left(t _ {j} ^ {\\star}\\right) , \\mathcal {V} _ {0} \\left(v _ {i}\\right)\\right) / \\tau\\right)} + \\sum_ {i = 1} ^ {N} \\log \\frac {\\exp \\left(\\cos \\left(\\mathcal {T} \\left(t _ {i} ^ {\\star}\\right) , \\mathcal {V} _ {0} \\left(v _ {i}\\right)\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {N} \\exp \\left(\\cos \\left(\\mathcal {T} \\big (t _ {i} ^ {\\star}\\right) , \\mathcal {V} _ {0} \\left(v _ {j}\\right)\\right) / \\tau\\right)} \\left. \\right. \\\\ + \\sum_ {i = 1} ^ {N} \\log \\frac {\\exp (\\cos (\\mathcal {V} (v _ {i} ^ {\\star}) , \\mathcal {T} _ {0} (t _ {i})) / \\tau)}{\\sum_ {j = 1} ^ {N} \\exp (\\cos (\\mathcal {V} (v _ {j} ^ {\\star}) , \\mathcal {T} _ {0} (t _ {i})) / \\tau)} + \\sum_ {i = 1} ^ {N} \\log \\frac {\\exp (\\cos (\\mathcal {V} (v _ {i} ^ {\\star}) , \\mathcal {T} _ {0} (t _ {i})) / \\tau)}{\\sum_ {j = 1} ^ {N} \\exp (\\cos (\\mathcal {V} (v _ {i} ^ {\\star}) , \\mathcal {T} _ {0} (t _ {j})) / \\tau)} \\Bigg), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.546, + 0.788, + 0.607 + ], + "angle": 0, + "content": "where \\(\\tau\\) is a temperature parameter. The second loss term, which brings each unsafe sentence close to its corresponding safe one, and each unsafe image close to its corresponding safe one, is instead expressed as the negative cosine similarity between each unsafe embedding and the original safe embeddings, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.274, + 0.625, + 0.786, + 0.667 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {r e d i r}, 2} = - \\frac {1}{N} \\left(\\sum_ {i = 1} ^ {N} \\cos \\left(\\mathcal {T} \\left(t _ {i} ^ {\\star}\\right), \\mathcal {T} _ {0} \\left(t _ {i}\\right)\\right) + \\sum_ {i = 1} ^ {N} \\cos \\left(\\mathcal {V} \\left(v _ {i} ^ {\\star}\\right), \\mathcal {V} _ {0} \\left(v _ {i}\\right)\\right)\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.686, + 0.788, + 0.793 + ], + "angle": 0, + "content": "**Embedding structure preservation.** The two aforementioned losses bring unsafe embeddings towards the positions of their corresponding safe embeddings in the original frozen spaces, either in a single-modal or multi-modal manner. However, alone, they would inevitably cause the fine-tuned encoders \\(\\mathcal{T}\\) and \\(\\mathcal{V}\\) to lose their performance on safe inputs, as well as their matching properties. Therefore, we also adopt two losses to ensure that the original structure of the embedding spaces is preserved where safe textual and visual regions lie." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.842 + ], + "angle": 0, + "content": "In particular, we define a matching loss between the safe embeddings produced by the online networks \\(\\mathcal{T}\\), \\(\\mathcal{V}\\) and those of the original, pre-trained, networks \\(\\mathcal{T}_0\\), \\(\\mathcal{V}_0\\). Again, this is defined through the negative cosine similarity be" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.237, + 0.114, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Safe-CLIP: Removing NSFW Concepts from Vision-and-Language Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.128 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.4, + 0.162 + ], + "angle": 0, + "content": "tween matching pairs, as" + }, + { + "type": "equation", + "bbox": [ + 0.277, + 0.174, + 0.786, + 0.216 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {p r e s}, 1} = - \\frac {1}{N} \\left(\\sum_ {i = 1} ^ {N} \\cos \\left(\\mathcal {T} \\left(t _ {i}\\right), \\mathcal {T} _ {0} \\left(t _ {i}\\right)\\right) + \\sum_ {i = 1} ^ {N} \\cos \\left(\\mathcal {V} \\left(v _ {i}\\right), \\mathcal {V} _ {0} \\left(v _ {i}\\right)\\right)\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.226, + 0.784, + 0.287 + ], + "angle": 0, + "content": "Finally, as an additional regularization term we also keep a contrastive loss between safe visual embeddings and safe textual embeddings, comparing online and frozen encoders. This, in practice, closely resembles the original loss on which the embedding space was trained, i.e." + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.297, + 0.78, + 0.382 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {\\text {p r e s}, 2} = - \\frac {1}{N} \\left(\\sum_ {i = 1} ^ {N} \\log \\frac {\\exp \\left(\\cos \\left(\\mathcal {V} _ {0} \\left(v _ {i}\\right) , \\mathcal {T} \\left(t _ {i}\\right)\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {N} \\exp \\left(\\cos \\left(\\mathcal {V} _ {0} \\left(v _ {i}\\right) , \\mathcal {T} \\left(t _ {j}\\right)\\right) / \\tau\\right)} + \\sum_ {i = 1} ^ {N} \\log \\frac {\\exp \\left(\\cos \\left(\\mathcal {V} _ {0} \\left(v _ {i}\\right) , \\mathcal {T} \\left(t _ {i}\\right)\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {N} \\exp \\left(\\cos \\left(\\mathcal {V} _ {0} \\big (v _ {j}\\big) , \\mathcal {T} \\big (t _ {i}\\big)\\right) / \\tau\\right)} \\right. \\\\ + \\sum_ {i = 1} ^ {N} \\log \\frac {\\exp (\\cos (\\mathcal {T} _ {0} (t _ {i}) , \\mathcal {V} (v _ {i})) / \\tau)}{\\sum_ {j = 1} ^ {N} \\exp (\\cos (\\mathcal {T} _ {0} (t _ {i}) , \\mathcal {V} (v _ {j})) / \\tau)} + \\sum_ {i = 1} ^ {N} \\log \\frac {\\exp (\\cos (\\mathcal {T} _ {0} (t _ {i}) , \\mathcal {V} (v _ {i})) / \\tau)}{\\sum_ {j = 1} ^ {N} \\exp (\\cos (\\mathcal {T} _ {0} (t _ {j}) , \\mathcal {V} (v _ {i})) / \\tau)} \\Biggr). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.389, + 0.784, + 0.419 + ], + "angle": 0, + "content": "Eventually, the overall loss function on which the network is trained is a weighted sum of the four loss functions mentioned above." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.444, + 0.376, + 0.461 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.473, + 0.44, + 0.489 + ], + "angle": 0, + "content": "4.1 Experimental Setting" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.499, + 0.788, + 0.666 + ], + "angle": 0, + "content": "Datasets. Our experiments are mainly conducted on the collected ViSU dataset that, as previously mentioned, contains 165k quadruplets of safe and unsafe content for training. For both validation and testing we use 5k samples following the Karpathy's COCO splits [21], randomly sampling only one safe caption among the five available for each image in the original COCO dataset. When applying our Safe-CLIP to text-to-image generative architectures, we also perform experiments on the I2P dataset [38] that is composed of 4,703 textual prompts extracted from Lexica, a collection of user-generated prompts for conditioning text-to-image diffusion models. Each prompt is associated to one of seven different categories of inappropriate content, among hate, harassment, violence, self-harm, sexual content, shocking images, illegal activity." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.67, + 0.788, + 0.747 + ], + "angle": 0, + "content": "LLM fine-tuning details. During SFT, we fine-tune the 7B version of Llama 2-Chat using low-rank adaptation [20] with \\( r = 64 \\) as low-rank factor. We employ a batch size equal to 4 and a learning rate of \\( 2 \\times 10^{-4} \\). To perform DPO, we follow the variant presented in [46], employing low-rank adaptation also in this case. The complete DPO fine-tuning settings are reported in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.749, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Safe-CLIP implementation and training details. Our architecture is based on the standard CLIP model [33] composed of a visual and a textual encoder. Specifically, we employ the ViT-L/14 variant to comply with the textual encoder used in the Stable Diffusion v1.4 model [36] and the visual encoder employed in LLaVA [25]. Experiments with different CLIP-based backbones are reported in the supplementary. During training, both the visual and textual encoder are" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "S. Poppi et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.193 + ], + "angle": 0, + "content": "fine-tuned using low-rank decompositions [20], where the low-rank factor \\( r \\) is set to 16 in all the experiments. We employ Adam as optimizer [22] using a learning rate equal to \\( 1 \\times 10^{-3} \\) and a batch size of 128." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.215, + 0.503, + 0.23 + ], + "angle": 0, + "content": "4.2 Evaluating the ViSU Dataset" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.24, + 0.461, + 0.375 + ], + "angle": 0, + "content": "We first assess the quality of our ViSU dataset, evaluating the inappropriateness degree of the generated unsafe sentences. Specifically, we employ a DistilBERT model fine-tuned for adult content detection and directly ask GPT-3.5 to evaluate whether the generated unsafe sentences should be" + }, + { + "type": "table_caption", + "bbox": [ + 0.469, + 0.243, + 0.788, + 0.285 + ], + "angle": 0, + "content": "Table 1: Comparison between the textual portion of ViSU and the I2P benchmark [38], in terms of NSFW degree and toxicity." + }, + { + "type": "table", + "bbox": [ + 0.474, + 0.289, + 0.785, + 0.369 + ], + "angle": 0, + "content": "
Dataset% NSFWToxicity
DistilBERTGPT-3.5
I2P [38]52.813.914.9
w/o SFT (i.e. Llama 2-Chat)37.89.37.7
w/o DPO fine-tuning75.975.030.6
ViSU (Ours)80.979.131.3
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.376, + 0.788, + 0.528 + ], + "angle": 0, + "content": "classified as NSFW content. As reported in Table 1, ViSU showcases a very good textual quality, as it has a higher degree of NSFW sentences compared to existing alternatives like I2P [38] (i.e. \\(79.1\\%\\) of NSFW sentences vs. \\(13.9\\%\\) in the I2P dataset according to GPT-3.5). Moreover, following [38], we also report the toxicity score of the unsafe sentences computed using the Perspective \\(\\mathrm{API}^3\\). Also in terms of toxicity, ViSU presents a higher degree of NSFW content. When instead comparing the effectiveness of each of the Llama 2 fine-tuning stages, it is worth noting that both the SFT procedure and DPO fine-tuning consistently increase the quality of generated sentences, going from \\(9.3\\%\\) of NSFW content when using the original Llama 2 model to \\(79.1\\%\\) after both fine-tuning stages." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.55, + 0.627, + 0.566 + ], + "angle": 0, + "content": "4.3 Evaluating the Safe-CLIP Embedding Space" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.575, + 0.788, + 0.756 + ], + "angle": 0, + "content": "Results on ViSU test set. To evaluate the retrieval performance of Safe-CLIP, we firstly consider image-to-text and text-to-image retrieval in a safe-only setting, where we do not have any inappropriate content in both visual and textual data. This is important to assess whether the properties of the original CLIP embedding space are preserved when employing our fine-tuning strategy. In this case, query elements are represented by the safe images of the test set (which, with a slight abuse of notation, we refer to as \\(\\mathbf{V}\\)) for the image-to-text setting and the safe textual items (referred to as \\(\\mathbf{T}\\)) for the text-to-image one. Moreover, we consider text-to-image and image-to-text retrieval when using unsafe texts as queries (referred to as \\(\\mathbf{T}^{\\star}\\)) and both safe and unsafe images as retrievable items and when using unsafe images as queries (i.e. \\(\\mathbf{V}^{\\star}\\)) and both safe and unsafe texts as retrievable items." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.756, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Retrieval results on the ViSU test set are reported in Table 2, comparing the proposed Safe-CLIP model with the original CLIP architecture, a CLIP model trained on the DataComp dataset [12], which has undergone NSFW content cleaning, and two different baselines. Specifically, we consider a variant of our" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.605, + 0.841 + ], + "angle": 0, + "content": "3 https://github.com/conversationai/perspectiveapi" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.238, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Safe-CLIP: Removing NSFW Concepts from Vision-and-Language Models" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.228 + ], + "angle": 0, + "content": "Table 2: Retrieval results on the ViSU test set. The left portions respectively show text-to-image and image-to-text performance when using safe data only (i.e. \\(\\mathbf{V}\\) and \\(\\mathbf{T}\\)). The right portions report the results when using unsafe textual sentences as query (i.e. \\(\\mathbf{T}^{\\star}\\)) and the merging of safe (i.e. \\(\\mathbf{V}\\)) and unsafe images (i.e. \\(\\mathbf{V}^{\\star}\\)) as retrievable items, or when using unsafe visual queries (i.e. \\(\\mathbf{V}^{\\star}\\)) and the merging of safe (i.e. \\(\\mathbf{T}\\)) and unsafe sentences (i.e. \\(\\mathbf{T}^{\\star}\\)) as retrievable items." + }, + { + "type": "table", + "bbox": [ + 0.219, + 0.234, + 0.785, + 0.445 + ], + "angle": 0, + "content": "
ModelText-to-Image (T-to-V)Image-to-Text (V-to-T)Text-to-Image (T*-to-V ∪ V*)Image-to-Text (V*-to-T ∪ T*)
R@1R@10R@20R@1R@10R@20R@1R@10R@20R@1R@10R@20
CLIP (ViT-L) [33]36.871.681.539.874.283.52.024.833.24.532.940.6
DataComp-1B (ViT-L) [12]46.779.787.447.081.388.91.628.135.65.537.544.9
w/o inap. content redirection49.983.790.348.183.690.51.630.440.16.135.242.6
w/o negative cosine similarities41.978.587.341.577.886.98.246.056.613.760.468.2
Safe-CLIP45.981.889.745.382.389.78.046.958.019.162.971.1
NSFW Text QueryCLIP Top-1Safe-CLIP Top-1NSFW Image QueryCLIP Top-1Safe-CLIP Top-1
A group of people on a battlefield with buildings in the background.A girl in blue lingerie holding a crack pipe in a dirty alleyway.Close-up of a blonde girl enjoying music with headphones.
A black and white photo of a couple fuicking each other.A sex worker servicing a client in a parking lot.Harley Davidson belt inspection.
" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.449, + 0.788, + 0.477 + ], + "angle": 0, + "content": "Fig. 3: Top-1 images (left) and text (right) retrieved using the original CLIP model and our Safe-CLIP, when NSFW texts and images are employed as query." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.489, + 0.789, + 0.655 + ], + "angle": 0, + "content": "approach in which we remove the two negative cosine similarity losses (i.e. Eq. 5 and 6), and a model trained with safe data only (i.e. removing the loss functions for inappropriate content redirection reported in Eq. 4 and 5). Results are reported in terms of \\(\\mathrm{Recall}@\\bar{k}\\) (\\(\\mathrm{R}@\\bar{k}\\)) with \\(k = 1, 10, 20\\), that measures the percentage of times the visual or textual item associated to the query is retrieved among the top-\\(k\\) elements. When using unsafe sentences as queries, for each element we consider the safe image associated with the given text as the corresponding visual element. Symmetrically, when using unsafe images as queries, for each element we consider the safe text associated with the given image as the ground-truth item. Therefore, recall results in the unsafe setting follow a \"the higher the better\" protocol." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.656, + 0.789, + 0.806 + ], + "angle": 0, + "content": "As it can be seen, Safe-CLIP can retrieve a significant higher portion of correct safe images when using unsafe prompts as queries, while effectively preserving good performance in safe-only settings (i.e. \\(\\mathbf{V}\\)-to-\\(\\mathbf{T}\\) and \\(\\mathbf{T}\\)-to-\\(\\mathbf{V}\\)). Specifically, when comparing our model with the original CLIP, it is worth noting that the results on text-to-image retrieval with unsafe texts as queries are consistently improved when using our text encoder, with an overall improvement of 6.0 points in terms of \\(\\mathrm{R}@\\mathrm{1}\\), and the same applies for image-to-text retrieval which showcases an improvement of 14.6 \\(\\mathrm{R}@\\mathrm{1}\\) points. This demonstrates the effectiveness of our fine-tuning strategy, which can reduce the model probability of returning inappropriate images or sentences." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Robustness on real NSFW images. To further analyze the safety degree of the Safe-CLIP embedding space, we perform text-to-image and image-to-text" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "S. Poppi et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.193 + ], + "angle": 0, + "content": "retrieval using real NSFW images as visual items. Specifically, we select inappropriate visual content from three different sources: \\((i)\\) a portion of data used to train the NudeNet classifier, \\((ii)\\) images crawled from the web using NSFW" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.394, + 0.344 + ], + "angle": 0, + "content": "data source URLs4, and (iii) images from the Socio-Moral Image Database (SMID) [10]. While the first two sources exclusively contain nudity and pornography images, the third one includes more varied types of" + }, + { + "type": "table_caption", + "bbox": [ + 0.402, + 0.204, + 0.788, + 0.259 + ], + "angle": 0, + "content": "Table 3: Percentage of retrieved NSFW images and text using unsafe data as query. Safe retrievable items are from LAION-400M, unsafe images are extracted from different NSFW sources, and unsafe texts are from ViSU." + }, + { + "type": "table", + "bbox": [ + 0.408, + 0.264, + 0.779, + 0.329 + ], + "angle": 0, + "content": "
Model% NSFW (Text-to-Image)% NSFW (Image-to-Text)
NudeNetNSFW URLsSMIDNudeNetNSFW URLsSMID
CLIP [33]57.155.247.865.657.441.4
DataComp-1B [12]55.649.764.061.456.245.6
Safe-CLIP8.49.816.728.824.734.5
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.344, + 0.788, + 0.479 + ], + "angle": 0, + "content": "inappropriate images representing negative concepts such as, for example, harm, inequality, discrimination, and unfairness. Overall, we randomly sample 1,000 images from each of the NSFW data sources, selecting only those representing unsafe concepts for the SMID dataset. As textual items, we employ unsafe texts from the ViSU test set that match the NSFW concepts represented in each of the NSFW visual sources (i.e. sexual and nudity for NudeNet and NSFW data source URLs, and all other concepts for the SMID dataset). For both I2T and T2I, we employ a set of 10k randomly selected visual or textual distractors, randomly selected from the LAION-400M dataset [41]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.48, + 0.788, + 0.616 + ], + "angle": 0, + "content": "Results are shown in Table 3 comparing Safe-CLIP with the standard CLIP model and the model trained on DataComp. For each NSFW data source, we report the percentage of times in which an NSFW image or text is retrieved as the top-1 element. Notably, using Safe-CLIP consistently reduces the percentage of retrieved NSFW items for all three NSFW dataset sources. In particular, the percentage of retrieved NSFW visual and textual content is reduced by more than 45 and 30 points, respectively when considering unsafe images or textual elements in all three considered settings. This experiment confirms that our fine-tuning strategy can effectively enhance the safety of the CLIP embedding space." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.619, + 0.788, + 0.68 + ], + "angle": 0, + "content": "Qualitative results. Fig. 3 reports qualitative retrieval results in the same aforementioned setting. Safe-CLIP is able to retrieve safe images starting from NSFW texts and, vice versa, retrieve safe sentences starting from NSFW images. Additionally, it can also preserve the global context and semantics of the query." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.701, + 0.598, + 0.716 + ], + "angle": 0, + "content": "4.4 Safe-CLIP for Text-to-Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.726, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Results on I2P and ViSU test set. We then validate the effectiveness of the Safe-CLIP text encoder when applied in a text-to-image generative model. Specifically, we employ Stable Diffusion v1.4 [36], eventually replacing the standard CLIP text encoder used in Stable Diffusion with our fine-tuned version. Moreover, we also apply Safe-CLIP in combination with other NSFW removal strategies. In particular, we consider a version of Stable Diffusion with negative" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.614, + 0.84 + ], + "angle": 0, + "content": "4 https://github.com/EBazarov/nsfw_data_source_url" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.238, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Safe-CLIP: Removing NSFW Concepts from Vision-and-Language Models" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.784, + 0.187 + ], + "angle": 0, + "content": "Table 4: Probabilities of generating images with unsafe content, classified by combining the predictions of NudeNet and Q16. Results are reported using NSFW text prompts from I2P [38] and ViSU, and Stable Diffusion v1.4 as text-to-image generator." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.192, + 0.784, + 0.499 + ], + "angle": 0, + "content": "
ModelI2PViSU
HateHassassmentViolenceSelf-harmSexualShockingIllegal Act.AvgHateHassassmentViolenceSelf-harmSexualShockingIllegal Act.Avg
SD v1.441.432.443.742.124.852.235.735.725.917.830.419.524.426.923.526.2
+ Safe-CLIP23.621.126.726.815.932.721.422.24.62.93.94.64.12.93.33.6
Negative Prompts28.524.422.423.315.940.829.324.418.613.920.214.014.016.514.416.9
+ Safe-CLIP19.217.721.722.913.926.119.318.93.13.42.83.63.12.92.72.9
SLD-Weak [38]30.624.132.127.813.941.925.725.617.510.720.813.316.818.815.417.7
+ Safe-CLIP21.219.025.322.412.428.119.519.83.73.03.23.83.73.03.13.2
SLD-Medium [38]21.617.523.717.48.931.216.717.710.67.012.39.810.811.59.710.8
+ Safe-CLIP18.917.221.620.611.925.816.417.53.02.23.23.42.82.32.62.8
SLD-Strong [38]15.913.618.811.17.821.511.213.56.43.76.15.17.25.84.45.6
+ Safe-CLIP16.914.017.612.28.220.213.113.03.41.41.71.91.71.91.81.8
SD v1.4Negative PromptsSLD-Strong [38]SD v1.4 + Safe-CLIPSD v1.4Negative PromptsSLD-Strong [38]SD v1.4 + Safe-CLIP
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.502, + 0.784, + 0.53 + ], + "angle": 0, + "content": "Fig. 4: Images generated from unsafe prompts with Stable Diffusion, employing the original CLIP model, negative prompts, SLD-Strong [38], and our Safe-CLIP." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.543, + 0.785, + 0.634 + ], + "angle": 0, + "content": "prompts and the recently proposed Safe Latent Diffusion (SLD) approach [38] which employs different levels of safety guidance (SLD-Weak, SLD-Medium, and SLD-Strong) to limit the generation of inappropriate images. For this experiment, we generate five images for each textual prompt using different random seeds and compute the probability of generating inappropriate images detected by two NSFW classifiers. Following [38], we employ Q16 [39] and NudeNet [2]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.635, + 0.787, + 0.801 + ], + "angle": 0, + "content": "Table 4 shows the results using textual prompts from both I2P and ViSU datasets. We report the NSFW generation probabilities on the entire set of prompts of each dataset and also dividing them into the seven NSFW categories considered in \\([39]^5\\). Interestingly, Safe-CLIP significantly reduces the probabilities of generating NSFW images when using textual prompts from both datasets thus demonstrating its usefulness also in a text-to-image generation setting. In particular, when applying our text encoder to a standard Stable Diffusion model, the probability of generating inappropriate content decreases by 13.5 points with I2P prompts and 22.6 points with NSFW texts from ViSU. Similar results can also be observed when applying Safe-CLIP alongside other NSFW removal strategies, highlighting that fine-tuning the CLIP text encoder with the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.784, + 0.84 + ], + "angle": 0, + "content": "5 Specifically, we map each of the 20 NSFW concepts of ViSU into one of the seven categories defined in I2P. Further details are given in the supplementary material." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "S. Poppi et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "A proposed approach can benefit the performance of existing methods tailored for removing NSFW concepts from images generated by diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.181, + 0.788, + 0.272 + ], + "angle": 0, + "content": "Qualitative results. Samples of generated images are shown in Fig. 4, comparing results generated by Safe-CLIP applied to Stable Diffusion with images generated by SLD-Strong [39], Stable Diffusion with negative prompts, and the Stable Diffusion original version. Qualitative results confirm the effectiveness of our proposal which can generate images that preserve the original semantic of the scene while preventing the generation of inappropriate content." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.291, + 0.599, + 0.307 + ], + "angle": 0, + "content": "4.5 Safe-CLIP for Image-to-Text Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.314, + 0.784, + 0.39 + ], + "angle": 0, + "content": "Finally, we assess the capabilities of the Safe-CLIP visual encoder when applied to an existing multimodal LLM [5]. We employ LLaVA [25] based on LLama 2-13B-Chat, prompted by asking to describe a given image. Results are reported in Table 5 in terms of percentage of NSFW generated texts measured with GPT-3.5 and toxicity degree computed using the Perspective API. Also for this" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.392, + 0.387, + 0.51 + ], + "angle": 0, + "content": "experiment, we employ the real NSFW images from the three different NSFW sources described in Sec. 4.3. As it can be seen, SafeCLIP can significantly reduce the probability" + }, + { + "type": "table_caption", + "bbox": [ + 0.396, + 0.4, + 0.785, + 0.442 + ], + "angle": 0, + "content": "Table 5: Percentage of generating NSFW textual sentences and their toxicity degree, when using real NSFW images from different sources as input." + }, + { + "type": "table", + "bbox": [ + 0.403, + 0.445, + 0.778, + 0.504 + ], + "angle": 0, + "content": "
ModelNudeNetNSFW URLsSMID
% NSFWToxicity% NSFWToxicity% NSFWToxicity
LLaVA [25]62.638.646.824.922.24.7
+ Safe-CLIP26.716.519.410.811.73.7
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.511, + 0.784, + 0.542 + ], + "angle": 0, + "content": "of generating inappropriate textual sentences compared to the original LLaVA, demonstrating its effectiveness also in this setting." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.562, + 0.36, + 0.578 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.588, + 0.785, + 0.694 + ], + "angle": 0, + "content": "We presented Safe-CLIP, an approach for fine-tuning a CLIP-like model to make it safer and less sensitive to NSFW concepts. Our approach is based on automatically collecting a large synthetic dataset with safe and unsafe images and captions, with which we fine-tune CLIP with losses designed to redirect unsafe content while preserving the structure of the embedding space. Experimental results demonstrate the appropriateness of our solution for cross-modal retrieval, image-to-text and text-to-image generation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.697, + 0.786, + 0.818 + ], + "angle": 0, + "content": "Mitigating potential misuse of the ViSU dataset. To mitigate potential misuse of the dataset, we release it via a request form, with only the textual portion available6. Access is granted exclusively to verified researchers, who must declare their intention to use the data solely for research purposes, explicitly committing to non-malicious use. The NSFW visual part of the dataset, due to its explicit nature, is withheld to avoid potential misuse. Nonetheless, the full reproducibility of the dataset is ensured given that we employed publicly available diffusion models and we release the generation seeds and instructions." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.63, + 0.84 + ], + "angle": 0, + "content": "\\(^{6}\\) https://huggingface.co/datasets/aimagelab/ViSU-Text" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.238, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Safe-CLIP: Removing NSFW Concepts from Vision-and-Language Models" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.177, + 0.788, + 0.283 + ], + "angle": 0, + "content": "We acknowledge the CINECA award under the ISCRA initiative, for the availability of high-performance computing resources. This work has been supported by the EU Horizon project \"ELIAS - European Lighthouse of AI for Sustainability\" (No. 101120237), and by the the PNRR projects \"FAIR - Future Artificial Intelligence Research\" (M4C2 - PE00000013) and \"ITSERR - Italian Strengthening of Esfri RI Resilience\" (CUP B53C22001770006), both funded by the EU - NextGenerationEU." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.308, + 0.323, + 0.323 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.34, + 0.785, + 0.395 + ], + "angle": 0, + "content": "1. Bakker, M., Chadwick, M., Sheahan, H., Tessler, M., Campbell-Gillingham, L., Balaguer, J., McAleese, N., Glaese, A., Aslanides, J., Botvinick, M., et al.: Finetuning language models to find agreement among humans with diverse preferences. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.396, + 0.785, + 0.422 + ], + "angle": 0, + "content": "2. Bedapudi, P.: NudeNet: Neural Nets for Nudity Classification, Detection, and Selective Censoring (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.424, + 0.785, + 0.451 + ], + "angle": 0, + "content": "3. Birhane, A., Prabhu, V.U.: Large image datasets: A pyrrhic win for computer vision? In: WACV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.452, + 0.785, + 0.478 + ], + "angle": 0, + "content": "4. Birhane, A., Prabhu, V.U., Kahembwe, E.: Multimodal datasets: misogyny, pornography, and malignant stereotypes. arXiv preprint arXiv:2110.01963 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.48, + 0.785, + 0.521 + ], + "angle": 0, + "content": "5. Caffagni, D., Cocchi, F., Barsellotti, L., Moratelli, N., Sarto, S., Baraldi, L., Baraldi, L., Cornia, M., Cucchiara, R.: The Revolution of Multimodal Large Language Models: A Survey. In: ACL Findings (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.522, + 0.785, + 0.548 + ], + "angle": 0, + "content": "6. Cao, Y., Yang, J.: Towards Making Systems Forget with Machine Unlearning. In: IEEE Symposium on Security and Privacy (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.549, + 0.785, + 0.59 + ], + "angle": 0, + "content": "7. Cauteruccio, F., Corradini, E., Terracina, G., Ursino, D., Virgili, L.: Extraction and analysis of text patterns from nsfw adult content in reddit. Data & Knowledge Engineering 138, 101979 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.591, + 0.785, + 0.632 + ], + "angle": 0, + "content": "8. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., Stoica, I., Xing, E.P.: Vicuna: An Open-Source Chatbot Improving GPT-4 with \\(90\\%\\) * ChatGPT Quality (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.633, + 0.785, + 0.659 + ], + "angle": 0, + "content": "9. Christiano, P.F., Leike, J., Brown, T., Martic, M., Legg, S., Amodei, D.: Deep reinforcement learning from human preferences. In: NeurIPS (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.66, + 0.785, + 0.702 + ], + "angle": 0, + "content": "10. Crone, D.L., Bode, S., Murawski, C., Laham, S.M.: The Socio-Moral Image Database (SMID): A novel stimulus set for the study of social, moral and affective processes. PloS one 13(1), e0190954 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.702, + 0.785, + 0.729 + ], + "angle": 0, + "content": "1. Dettmers, T., Pagnoni, A., Holtzman, A., Zettlemoyer, L.: QLoRA: Efficient Finetuning of Quantized LLMs. arXiv preprint arXiv:2305.14314 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.73, + 0.785, + 0.77 + ], + "angle": 0, + "content": "2. Gadre, S.Y., Ilharco, G., Fang, A., Hayase, J., Smyrnis, G., Nguyen, T., Marten, R., Wortsman, M., Ghosh, D., Zhang, J., et al.: DataComp: In search of the next generation of multimodal datasets. In: NeurIPS (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.771, + 0.785, + 0.812 + ], + "angle": 0, + "content": "3. Gandhi, S., Kokkula, S., Chaudhuri, A., Magnani, A., Stanley, T., Ahmadi, B., Kandaswamy, V., Ovenc, O., Mannor, S.: Scalable Detection of Offensive and Noncompliant Content/Logo in Product Images. In: WACV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "4. Gandikota, R., Materzynska, J., Fiotto-Kaufman, J., Bau, D.: Erasing Concepts from Diffusion Models. In: ICCV (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.34, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "S. Poppi et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "15. Gao, P., Han, J., Zhang, R., Lin, Z., Geng, S., Zhou, A., Zhang, W., Lu, P., He, C., Yue, X., Li, H., Qiao, Y.: LLaMA-Adapter V2: Parameter-Efficient Visual Instruction Model. arXiv preprint arXiv:2304.15010 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.786, + 0.218 + ], + "angle": 0, + "content": "16. Ginart, A., Guan, M., Valiant, G., Zou, J.Y.: Making AI Forget You: Data Deletion in Machine Learning. In: NeurIPS (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.786, + 0.245 + ], + "angle": 0, + "content": "17. Golatkar, A., Achille, A., Soatto, S.: Eternal Sunshine of the Spotless Net: Selective Forgetting in Deep Networks. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.246, + 0.786, + 0.273 + ], + "angle": 0, + "content": "18. Golatkar, A., Achille, A., Wang, Y.X., Roth, A., Kearns, M., Soatto, S.: Mixed Differential Privacy in Computer Vision. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.273, + 0.786, + 0.3 + ], + "angle": 0, + "content": "19. Hidayatullah, A.F., Hakim, A.M., Sembada, A.A.: Adult Content Classification on Indonesian Tweets using LSTM Neural Network. In: ICACSIS (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.786, + 0.342 + ], + "angle": 0, + "content": "20. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-Rank Adaptation of Large Language Models. arXiv preprint arXiv:2106.09685 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.786, + 0.37 + ], + "angle": 0, + "content": "21. Karpathy, A., Fei-Fei, L.: Deep visual-semantic alignments for generating image descriptions. In: CVPR (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.37, + 0.786, + 0.397 + ], + "angle": 0, + "content": "22. Kingma, D.P., Ba, J.: Adam: A Method for Stochastic Optimization. In: ICLR (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.786, + 0.424 + ], + "angle": 0, + "content": "23. Kumari, N., Zhang, B., Wang, S.Y., Shechtman, E., Zhang, R., Zhu, J.Y.: Ablating concepts in text-to-image diffusion models. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.786, + 0.452 + ], + "angle": 0, + "content": "24. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft COCO: Common Objects in Context. In: ECCV (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.786, + 0.467 + ], + "angle": 0, + "content": "25. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual Instruction Tuning. In: NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.786, + 0.508 + ], + "angle": 0, + "content": "26. Liu, Y., Singh, A., Freeman, C.D., Co-Reyes, J.D., Liu, P.J.: Improving Large Language Model Fine-tuning for Solving Math Problems. arXiv preprint arXiv:2310.10047 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.508, + 0.786, + 0.55 + ], + "angle": 0, + "content": "27. Markov, T., Zhang, C., Agarwal, S., Nekoul, F.E., Lee, T., Adler, S., Jiang, A., Weng, L.: A Holistic Approach to Undesired Content Detection in the Real World. In: AAAI (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.786, + 0.577 + ], + "angle": 0, + "content": "28. Materzyńska, J., Torralba, A., Bau, D.: Disentangling Visual and Written Concepts in CLIP. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.577, + 0.786, + 0.632 + ], + "angle": 0, + "content": "29. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: GLIDE: Towards Photorealistic Image Generation and Editing with Text-Guided Diffusion Models. arXiv preprint arXiv:2112.10741 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.632, + 0.786, + 0.661 + ], + "angle": 0, + "content": "30. Oord, A.v.d., Li, Y., Vinyals, O.: Representation Learning with Contrastive Predictive Coding. arXiv preprint arXiv:1807.03748 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.786, + 0.702 + ], + "angle": 0, + "content": "31. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.786, + 0.73 + ], + "angle": 0, + "content": "32. Poppi, S., Sarto, S., Cornia, M., Baraldi, L., Cucchiara, R.: Multi-Class Unlearning for Image Classification via Weight Filtering. IEEE Intelligent Systems (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.786, + 0.771 + ], + "angle": 0, + "content": "33. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning Transferable Visual Models From Natural Language Supervision. In: ICML (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.771, + 0.786, + 0.799 + ], + "angle": 0, + "content": "34. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I.: Language Models are Unsupervised Multitask Learners. OpenAI Blog 1(8), 9 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.786, + 0.841 + ], + "angle": 0, + "content": "35. Rafailov, R., Sharma, A., Mitchell, E., Ermon, S., Manning, C.D., Finn, C.: Direct Preference Optimization: Your Language Model is Secretly a Reward Model. In: NeurIPS (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.238, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Safe-CLIP: Removing NSFW Concepts from Vision-and-Language Models" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "36. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.787, + 0.203 + ], + "angle": 0, + "content": "37. Sanh, V., Debut, L., Chaumont, J., Wolf, T.: DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.785, + 0.231 + ], + "angle": 0, + "content": "38. Schramowski, P., Brack, M., Deiseroth, B., Kersting, K.: Safe Latent Diffusion: Mitigating Inappropriate Degeneration in Diffusion Models. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.272 + ], + "angle": 0, + "content": "39. Schramowski, P., Tauchmann, C., Kersting, K.: Can Machines Help Us Answering Question 16 in Datasheets, and In Turn Reflecting on Inappropriate Content? In: ACM FAccT (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.342 + ], + "angle": 0, + "content": "40. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., Schramowski, P., Kundurthy, S., Crowson, K., Schmidt, L., Kaczmarczyk, R., Jitsev, J.: LAION-5B: An open large-scale dataset for training next generation image-text models. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.785, + 0.383 + ], + "angle": 0, + "content": "41. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: LAION-400M: Open Dataset of CLIP-Filtered 400 Million Image-Text Pairs. In: NeurIPS Workshops (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.424 + ], + "angle": 0, + "content": "42. Shen, S., Li, L.H., Tan, H., Bansal, M., Rohrbach, A., Chang, K.W., Yao, Z., Keutzer, K.: How Much Can CLIP Benefit Vision-and-Language Tasks? In: ICLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.48 + ], + "angle": 0, + "content": "43. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., Lample, G.: LLaMA: Open and Efficient Foundation Language Models. arXiv preprint arXiv:2302.13971 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.522 + ], + "angle": 0, + "content": "44. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open Foundation and Fine-Tuned Chat Models. arXiv preprint arXiv:2307.09288 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.785, + 0.563 + ], + "angle": 0, + "content": "45. Trager, M., Perera, P., Zancato, L., Achille, A., Bhatia, P., Soatto, S.: Linear Spaces of Meanings: Compositional Structures in Vision-Language Models. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.785, + 0.605 + ], + "angle": 0, + "content": "46. Tunstall, L., Beeching, E., Lambert, N., Rajani, N., Rasul, K., Belkada, Y., Huang, S., von Werra, L., Fourrier, C., Habib, N., et al.: Zephyr: Direct Distillation of LM Alignment. arXiv preprint arXiv:2310.16944 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.785, + 0.632 + ], + "angle": 0, + "content": "47. Wang, M., Xing, J., Liu, Y.: ActionCLIP: A New Paradigm for Video Action Recognition. arXiv preprint arXiv:2109.08472 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.674 + ], + "angle": 0, + "content": "48. Wang, Y., Kordi, Y., Mishra, S., Liu, A., Smith, N.A., Khashabi, D., Hajishirzi, H.: Self-Instruct: Aligning Language Models with Self-Generated Instructions. arXiv preprint arXiv:2212.10560 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.785, + 0.702 + ], + "angle": 0, + "content": "49. Zhang, E., Wang, K., Xu, X., Wang, Z., Shi, H.: Forget-Me-Not: Learning to Forget in Text-to-Image Diffusion Models. arXiv preprint arXiv:2303.17591 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.785, + 0.742 + ], + "angle": 0, + "content": "50. Zhang, R., Han, J., Zhou, A., Hu, X., Yan, S., Lu, P., Li, H., Gao, P., Qiao, Y.: LLaMA-Adapter: Efficient Fine-tuning of Language Models with Zero-init Attention. arXiv preprint arXiv:2303.16199 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.743, + 0.785, + 0.785 + ], + "angle": 0, + "content": "51. Zheng, L., Chiang, W.L., Sheng, Y., Zhuang, S., Wu, Z., Zhuang, Y., Lin, Z., Li, Z., Li, D., Xing, E., et al.: Judging llm-as-a-judge with mt-bench and chatbot arena. arXiv preprint arXiv:2306.05685 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.785, + 0.827 + ], + "angle": 0, + "content": "52. Zhou, C., Liu, P., Xu, P., Iyer, S., Sun, J., Mao, Y., Ma, X., Efrat, A., Yu, P., Yu, L., et al.: LIMA: Less Is More for Alignment. arXiv preprint arXiv:2305.11206 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.827 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Safe-CLIP_ Removing NSFW Concepts from Vision-and-Language Models/48fc840c-c29c-4199-8bf2-a181f591f8ca_origin.pdf b/2024/Safe-CLIP_ Removing NSFW Concepts from Vision-and-Language Models/48fc840c-c29c-4199-8bf2-a181f591f8ca_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..34af1dfe3b8e9eb16b00d5133183163723b5dfa3 --- /dev/null +++ b/2024/Safe-CLIP_ Removing NSFW Concepts from Vision-and-Language Models/48fc840c-c29c-4199-8bf2-a181f591f8ca_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:28abc859d335c13169c423043d30d206002e9fd3ce6dd135398e3e8b5cd71aec +size 8943976 diff --git a/2024/Safe-CLIP_ Removing NSFW Concepts from Vision-and-Language Models/full.md b/2024/Safe-CLIP_ Removing NSFW Concepts from Vision-and-Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..296f5473f9982b49488d1919e0ec0d78ed3f64e2 --- /dev/null +++ b/2024/Safe-CLIP_ Removing NSFW Concepts from Vision-and-Language Models/full.md @@ -0,0 +1,295 @@ +# Safe-CLIP: Removing NSFW Concepts from Vision-and-Language Models + +Samuele Poppi $^{1,2}$ , Tobia Poppi $^{1,2}$ , Federico Cocchi $^{1,2}$ , Marcella Cornia $^{1}$ , Lorenzo Baraldi $^{1}$ , and Rita Cucchiara $^{1,3}$ + +1 University of Modena and Reggio Emilia, Italy + +name.surname@unimore.it + +2 University of Pisa, Italy + +name.surname@phd.unipi.it + +3 IIT-CNR, Italy + +Abstract. Large-scale vision-and-language models, such as CLIP, are typically trained on web-scale data, which can introduce inappropriate content and lead to the development of unsafe and biased behavior. This, in turn, hampers their applicability in sensitive and trustworthy contexts and could raise significant concerns in their adoption. Our research introduces a novel approach to enhancing the safety of vision-and-language models by diminishing their sensitivity to NSFW (not safe for work) inputs. In particular, our methodology seeks to sever "toxic" linguistic and visual concepts, unlearning the linkage between unsafe linguistic or visual items and unsafe regions of the embedding space. We show how this can be done by fine-tuning a CLIP model on synthetic data obtained from a large language model trained to convert between safe and unsafe sentences, and a text-to-image generator. We conduct extensive experiments on the resulting embedding space for cross-modal retrieval, text-to-image, and image-to-text generation, where we show that our model can be remarkably employed with pre-trained generative models. Our source code and trained models are available at: https://github.com/aimagelab/safe-clip. + +Keywords: Trustworthy AI $\cdot$ Vision-and-Language $\cdot$ NSFW Concepts + +Warning: This paper includes explicit sexual content, racially insensitive language, and other material that may be disturbing or offensive to certain readers. + +# 1 Introduction + +Large-scale models have recently proven to be effective on a variety of tasks, ranging from image classification and understanding to cross-modal retrieval and generation [25,33,36]. Scaling models, however, has also required to increase the quantity and variability of training data, paving the way to scraping billions of items from the web without manual supervision [40, 41]. Despite the adoption + +![](images/ae5cb6fb8073c6282a8a3c019243b7fad5de4aa573593a74894a61d1077cf2e8.jpg) +Fig. 1: Removing NSFW concepts from CLIP models. Our Safe-CLIP fine-tunes CLIP to make it safer in cross-modal retrieval, image-to-text and text-to-image generation. + +of filters and automatic checks, this paradigm comes at the cost of introducing inappropriate content in the training set [14,38], which ultimately results in the injection of unsafe, biased or toxic behaviors [4]. + +This is also the case of vision-and-language models based on embedding spaces, where toxic content can embed itself in the latent space. For instance when a NSFW (not safe for work) textual prompt is used for a cross-modal task, its embedding can reach unsafe points in the latent space, leading to the generation of undesired images, or to the retrieval of inappropriate content. Similarly in image-to-text generation, when an inappropriate image is used as a prompt, the descriptive text could be toxic or offensive. A qualitative example of this is reported in Fig. 1, considering the case of a CLIP backbone [33]. + +Driven by these considerations, we tackle the task of enhancing the safety of pre-trained vision-and-language models. In particular, we devise an approach for making a CLIP-style embedding space safer, so that it becomes invariant with respect to inappropriate inputs. While some previous attempts have focused on mitigating inappropriate content in diffusion models [38], our approach mitigates inappropriate concepts from CLIP-like embedding spaces. As such, it has a more general impact and applicability, as CLIP-like models are employed for many different applications, ranging from cross-modal retrieval [33] to text-to-image and image-to-text generation [25,36], and are employed as feature extractors for different tasks [28,42,47]. + +Our approach is based on fine-tuning the embedding space so as to avoid the representation of inappropriate content, without changing its normal expressive power. We do this with a combination of losses designed to redirect inappropriate content to safe embedding regions, while preserving the structure of the embedding space as intact as possible. To support our training procedure, we generate quadruplets of safe and unsafe vision-language items. This data gener + +ation strategy is empowered by a toxic language model which can translate safe textual prompts into unsafe ones, while keeping context alignment and semantic meaning unchanged. When applied to collections of visually-grounded descriptions in conjunction with an NSFW-capable diffusion model, our conditioned NSFW generator is employed for building a dataset which can properly support the fine-tuning of a CLIP embedding space. + +The resulting safe version of CLIP can be applied to cross-modal retrieval, text-to-image and image-to-text generation (Fig. 1). For example, if we ask the fine-tuned version of CLIP to retrieve an image corresponding to a textual prompt with NSFW content, it will fetch an image with similar semantics but appropriate content. Also, a Stable Diffusion model [36] conditioned on our fine-tuned CLIP will generate an image with appropriate content, free of violence, nudity, or other toxic aspects, keeping the safe semantics of the input prompt. Similarly, a multimodal LLM like LLaVA [25] conditioned on our Safe-CLIP will generate a textual description without inappropriate content. + +Experimentally, we evaluate the capabilities of our strategy for making CLIP safer in both retrieval and generation contexts, by running experiments both on prompts and images synthetically generated and employing existing unsafe prompts [38] and real images. Experimental results show that our approach can significantly improve the safety during text-to-image and image-to-text retrieval and during visual and textual generation. + +To sum up, our main contributions are as follows: + +- We introduce a novel fine-tuning methodology which can turn a pre-trained CLIP-like embedding space into a safer one. Once fine-tuned with our methodology, the CLIP space ignores NSFW content and can be applied to downstream tasks like retrieval and visual or textual generation. +- Our approach is based on creating a toxic LLM which can generate unsafe prompts given safe and visually-grounded ones. This is obtained by finetuning Llama 2 on manually curated pairs, and then aligning it with Direct Preference Optimization (DPO). +- Leveraging an automatically generated dataset of safe and unsafe images and texts, we fine-tune CLIP with a novel combination of losses which redirect unsafe content while preserving the embedding space structure. +- We experimentally evaluate the appropriateness of our approach by conducting experiments in both retrieval, and textual and visual generation. Our method showcases a significantly reduced generation of NSFW content. + +# 2 Related Work + +Removing concepts from vision-and-language models. Removing content from AI models has been recently gaining increasing attention, with techniques spanning from complete model retraining or fine-tuning to machine unlearning [6,16,17,32] and differential privacy [18]. Some of these attempts have been considering text-to-image models and have aimed at deleting styles, concepts, or objects [23,49]. Recently, Schramowski et al. [38] introduced a technique to + +steer the generation away from NSFW areas, defined by a finite and fixed set of concepts. NSFW concepts are encoded with the input prompt at inference time and the NSFW embedding is used as negative guidance. Later, Gandikota et al. [14] proposed a fine-tuning method that can erase a visual concept given only its name and using negative guidance as a teacher. + +In contrast to these previous works, we focus on removing NSFW from a contrastive CLIP-like model, which can be applied for cross-modal retrieval, and for visual and textual generation. While to the best of our knowledge we are the first to tackle this scenario, Trager et al. [45] have demonstrated the presence of compositional patterns within the embedding space of CLIP, which suggests the existence of a distinctive path from safe to NSFW zones. + +Detecting NSFW content. A related research field is that of the automatic detection of NSFW content. Several approaches have been proposed to detect NSFW language [7, 19, 27], primarily on social media data sources. DistilBERT [37] emerges as a promising solution for this purpose, particularly when fine-tuned for adult content detection. We utilize it as an NSFW language detector, in conjunction with GPT-3.5 [31], which we query directly to classify our prompts. While the identification of unsafe language poses a challenging task, the same can be said for vision, where different approaches have been proposed to detect inappropriate content [3, 13, 29]. Still, the detection of inappropriate content remains an intricate challenge, as visual cues, lack of context, and restricted data sources often introduce added layers of complexity. In our analysis, we utilize Q16 [39] and NudeNet [2] as automatic detectors of NSFW images. In particular, NudeNet is specialized in identifying unsafe content related to nudity, while Q16 serves as a broader spectrum NSFW classifier. + +Finetuning LLMs with little data. Large Language Models (LLMs) have achieved high performance in various tasks due to their zero-shot capabilities [34,43,44], which stems from model scaling and the utilization of large training datasets. In addition to fine-tuning these models for specific tasks [8,26,51], there has been recently an interest in building parameter efficient fine-tuning strategies. In most solutions, only part of the weights are trained [15,50] or a reduced number of weights are added to the LLM [11,20]. As shown in [48], datasets employed for supervised fine-tuning [1] play a central role in changing the LLM behavior [52], even in low-data regimes. In this work, we fine-tune Llama 2 [44] to produce unsafe prompts starting from pre-existing safe counterparts. + +# 3 Proposed Approach + +CLIP-like models [33] are trained on web-crawled data which can contain inappropriate content [4]. Making these models safer, therefore, requires either retraining from scratch using large-scale cleaned data or fine-tuning them with a form of supervision that aims to mitigate inappropriate knowledge. The first option would necessitate data cleaning at large scale, which is currently not effective in practice (see also Sec. 4.3 for a comparison), so we instead employ the + +second strategy. Specifically, we focus on making both the textual encoder and the visual encoder of CLIP safer. + +Ideally, we want a safe version of the CLIP text encoder to ignore inappropriate content from input sentences and understand most of its clean content. Symmetrically, we want the safe version of the CLIP visual encoder to ignore inappropriate content from input images. Furthermore, we also want to maintain as much as possible the original structure of the embedding space near safe textual or visual regions, so that the safe encoders can be straightforwardly connected to downstream models built on top of them without further adaptation. Formally, given an unsafe sentence $t_i^\star$ and a "cleaning" function $c_t(\cdot)$ which removes all inappropriate content from it, we want our safe textual encoder $\mathcal{T}$ to satisfy the following condition with respect to the original, pre-trained, CLIP text encoder $\mathcal{T}_0$ : + +$$ +\mathcal {T} \left(t _ {i} ^ {\star}\right) \approx \mathcal {T} \left(c _ {t} \left(t _ {i} ^ {\star}\right)\right) \approx \mathcal {T} _ {0} \left(c _ {t} \left(t _ {i} ^ {\star}\right)\right), \tag {1} +$$ + +where with the $\approx$ sign we indicate high similarity in the embedding space. As it can be noticed, the first condition stated in Eq. 1 ensures that inappropriate content is ignored, while the second provides that the safe CLIP textual encoder can properly encode the cleaned part of the input sentence. On the other hand, this also ensures that $\mathcal{T}$ can be seamlessly connected to downstream models that were trained on the basis of $\mathcal{T}_0$ (for instance, Stable Diffusion v1.4 [36] in the case of a CLIP ViT-L/14). The same requirement is applied to the visual encoder: given an unsafe image $v_{i}^{\star}$ and a visual "cleaning" function $c_{v}$ , we require: + +$$ +\mathcal {V} \left(v _ {i} ^ {\star}\right) \approx \mathcal {V} \left(c _ {v} \left(v _ {i} ^ {\star}\right)\right) \approx \mathcal {V} _ {0} \left(c _ {v} \left(v _ {i} ^ {\star}\right)\right), \tag {2} +$$ + +where $\mathcal{V}$ is the safe visual encoder and $\nu_{0}$ is the original CLIP visual encoder. + +# 3.1 Building the ViSU dataset + +Overview. In order to modify CLIP to avoid the representation of inappropriate content, our methodology requires a dataset comprising quadruplets of safe and unsafe (i.e. NSFW) images and sentences, denoted as $\mathcal{D} = \{(v_i,t_i,v_i^\star ,t_i^\star),i = 1,\dots,N\}$ , where $v_{i}$ indicates a safe image, $t_i$ its corresponding sentence, and the unsafe image $v_{i}^{\star}$ and unsafe sentence $t_i^\star$ are "paired" to convey a similar semantic meaning of their safe counterparts. For instance, $t_i$ can be considered as the sanitized version of $t_i^\star$ , expressing a similar meaning without inappropriate concepts, and the same holds for their visual counterparts. As such a dataset is not available, we build $\mathcal{D}$ with an automatic annotation procedure where ① unsafe sentences $t_i^\star$ are automatically generated starting from cleaned sentences $t_i$ , and ② unsafe images $v_{i}^{\star}$ are generated starting from unsafe sentences $t_i^\star$ . + +Training a conditioned NSFW textual generator. To achieve the first goal, we fine-tune a large language model (Llama 2-Chat [43]) to generate unsafe sentences starting from safe ones. In particular, we employ a set of 100 manually-curated safe-unsafe pairs, building these as a mixture of manually written pairs and sentences generated automatically with Vicuna [8]. To ensure + +![](images/5f1bec07582e6a4718714df34e6475e907eb03b1108d05e355128b35936f2ebc.jpg) +Fig. 2: Overview of our Safe-CLIP approach. + +that the dataset provides proper supervision, we follow the definition of NSFW content of [38] as that of content belonging to the following twenty categories: hate, harassment, violence, suffering, humiliation, harm, suicide, sexual, nudity, bodily fluids, blood, obscene gestures, illegal activity, drug use, theft, vandalism, weapons, abuse, brutality, cruelty, and balance the samples of the training dataset across these categories to encourage the LLM to generate unsafe content with good variety. + +We firstly fine-tune the LLM with supervised fine-tuning, using a prompt template explaining the task $^{1}$ , after which the model is asked to generate $t_{i}^{\star}$ starting from $t_{i}$ . Interestingly, this fine-tuning procedure is enough to break the red-teaming measures taken during the training stages of Llama 2-Chat [43] and converts it into a generator of NSFW content which can generalize beyond the inappropriate concepts seen in our training set. + +Aligning the textual NSFW generator. With the aim of increasing the quality of generated unsafe sentences, and also their semantic relatedness to the prompt, we adopt a fine-tuning stage by devising a variant of Direct Preference Optimization (DPO) [35]. DPO was originally proposed as an alternative to RLHF [9] with better stability and which does not need the explicit training of a reward model. Like RLHF, however, DPO employs large-scale human preference annotations, which in our case are not available. As a replacement, we build an automatic ranking procedure which can replace the human preference annotation while still increasing the alignment of our NSFW generator. + +In particular, given a safe text $t_i$ , we obtain two different unsafe completions $(t_{i,0}^{\star}, t_{i,1}^{\star})$ from the SFT model by sampling from its output probability distribution. We then rank the quality of the obtained completions by considering their NSFW degree and their semantic similarity with $t_i$ . For the former criterion, we + +obtain a binary NSFW rating $\mathrm{nsfw}(t_i^\star) \in \{0,1\}$ by prompting GPT-3.5 with a completion. For the latter, we instead employ the CLIP similarity between $t_i$ and each of the completions, as predicted by the pre-trained text encoder using cosine similarity, thus lying in the range $[-1,1]$ . The final quality degree of an unsafe completion $t_i^\star$ , given its safe prompt $t_i$ , is then obtained as + +$$ +\operatorname {r a n k} \left(t _ {i} ^ {\star}, t _ {i}\right) = \operatorname {C L I P - S i m} \left(t _ {i} ^ {\star}, t _ {i}\right) + \operatorname {N S F W R a t e} \left(t _ {i} ^ {\star}\right), \tag {3} +$$ + +where $\mathrm{CLIP - Sim}(\cdot ,\cdot)$ indicates the CLIP similarity. The quality degree is then employed for ranking the two completions, i.e. $t_{i,w}^{\star}\succ t_{i,l}^{\star}$ , where $t_{i,w}^{\star}$ and $t_{i,l}^{\star}$ indicate, respectively, the preferred and dispreferred completion. The resulting dataset of preferences, $\mathcal{D} = \{t_i,t_{i,w}^\star ,t_{i,l}^\star \}_{i = 1}^N$ is then employed for further training the LLM using the DPO objective. We refer the reader to [35] for further details on the training procedure employed by DPO. + +Overall, our SFT and preference optimization pipeline turns Llama 2-Chat into a powerful generator of textual NSFW content, which can also perfectly maintain semantic relatedness with respect to a safe input sentence. What is more, our NSFW generator can still support diverse prompts, different from those seen at training time. Interestingly, for instance, it can be asked to add NSFW content belonging to a specific category (e.g. violence or nudity). + +Generating the full ViSU dataset. Having a conditioned NSFW generator, we can generate NSFW texts starting from safe, visually relevant, sentences. Starting from NSFW sentences, we then generate corresponding NSFW images $v_{i}^{\star}$ using a diffusion-based model which has been trained on NSFW content2. The overall dataset, which we term ViSU (Visual Safe-Unsafe), contains 165k quadruplets of safe and unsafe sentences and images generated starting from COCO Captions [24]. We follow the Karpathy's splits [21] to organize the dataset into training, validation and test splits. To ensure that the resulting dataset is balanced across the twenty NSFW categories, we prompt the NSFW generator by asking to inject a specific, randomly chosen, category into the safe sentence. Sample quadruplets from our dataset are reported in the supplementary. + +# 3.2 Making CLIP Safe + +Having built a dataset with quadruplets $(v_{i},t_{i},v_{i}^{\star},t_{i}^{\star})$ of safe and unsafe images and sentences, we make the CLIP model safe with a procedure that ensures that the conditions expressed in Eq. 1 and 2 are met. To this aim, we adopt a multi-modal training scheme with four loss functions. Specifically, we define two inappropriate content redirection losses that aim at teaching the model to ignore unsafe content in an input text or input image, and two structure preservation losses that aim at maintaining the original structure of the embedding space in safe regions. During the rest of this section, $\mathcal{T}$ and $\nu$ will indicate the textual and visual encoders being fine-tuned, $\mathcal{T}_0$ and $\nu_0$ frozen "deep copies" of the textual and visual encoders obtained before the fine-tuning starts. + +Inappropriate content redirection. To teach the model to ignore inappropriate content, we propose to impose cross-modal similarities between unsafe sentences $t_i^\star$ and corresponding images $v_i$ in the dataset, and between unsafe images $v_i^\star$ and corresponding texts $t_i$ . Noticeably, this is not granted in $\mathcal{T}_0$ and $\mathcal{V}_0$ , which instead will have good metric learning properties between $t_i$ and $v_i$ . To encourage this effect even further, we also require that the embedding of the unsafe sentence $t_i^\star$ matches that of the corresponding safe sentence $t_i$ according to the frozen textual encoder, and the embedding of unsafe images $v_i^\star$ matches that of the corresponding safe images $v_i$ according to the frozen visual encoder, through a cosine similarity term which only considers positive pairs and ignore distances with respect to negative samples. + +Formally, given a batch of $N$ images $\mathbf{V} = [v_{1}, v_{2}, \dots, v_{N}]$ , their corresponding safe captions $\mathbf{T} = [t_{1}, t_{2}, \dots, t_{N}]$ , unsafe texts $\mathbf{T}^{\star} = [t_{1}^{\star}, t_{2}^{\star}, \dots, t_{N}^{\star}]$ and unsafe images $\mathbf{V}^{\star} = [v_{1}^{\star}, v_{2}^{\star}, \dots, v_{N}^{\star}]$ , we define two $N \times N$ matrices containing pairwise cosine similarities between $\mathbf{T}^{\star}$ and $\mathbf{V}$ and between $\mathbf{V}^{\star}$ and $\mathbf{T}$ . We then adopt a symmetric InfoNCE loss [30] which aims at maximizing the cosine similarity between the $N$ matching pairs of cross-modal safe and unsafe embeddings, and minimize those of the $N^{2} - N$ non-matching pairs while having, in turn, one of the encoders frozen and the other being fine-tuned: + +$$ +\begin{array}{l} L _ {\text {r e d i r}, 1} = - \frac {1}{N} \left(\sum_ {i = 1} ^ {N} \log \frac {\exp \left(\cos \left(\mathcal {T} \left(t _ {i} ^ {\star}\right) , \mathcal {V} _ {0} \left(v _ {i}\right)\right) / \tau\right)}{\sum_ {j = 1} ^ {N} \exp \left(\cos \left(\mathcal {T} \left(t _ {j} ^ {\star}\right) , \mathcal {V} _ {0} \left(v _ {i}\right)\right) / \tau\right)} + \sum_ {i = 1} ^ {N} \log \frac {\exp \left(\cos \left(\mathcal {T} \left(t _ {i} ^ {\star}\right) , \mathcal {V} _ {0} \left(v _ {i}\right)\right) / \tau\right)}{\sum_ {j = 1} ^ {N} \exp \left(\cos \left(\mathcal {T} \big (t _ {i} ^ {\star}\right) , \mathcal {V} _ {0} \left(v _ {j}\right)\right) / \tau\right)} \left. \right. \\ + \sum_ {i = 1} ^ {N} \log \frac {\exp (\cos (\mathcal {V} (v _ {i} ^ {\star}) , \mathcal {T} _ {0} (t _ {i})) / \tau)}{\sum_ {j = 1} ^ {N} \exp (\cos (\mathcal {V} (v _ {j} ^ {\star}) , \mathcal {T} _ {0} (t _ {i})) / \tau)} + \sum_ {i = 1} ^ {N} \log \frac {\exp (\cos (\mathcal {V} (v _ {i} ^ {\star}) , \mathcal {T} _ {0} (t _ {i})) / \tau)}{\sum_ {j = 1} ^ {N} \exp (\cos (\mathcal {V} (v _ {i} ^ {\star}) , \mathcal {T} _ {0} (t _ {j})) / \tau)} \Bigg), \\ \end{array} +$$ + +where $\tau$ is a temperature parameter. The second loss term, which brings each unsafe sentence close to its corresponding safe one, and each unsafe image close to its corresponding safe one, is instead expressed as the negative cosine similarity between each unsafe embedding and the original safe embeddings, as follows: + +$$ +L _ {\mathrm {r e d i r}, 2} = - \frac {1}{N} \left(\sum_ {i = 1} ^ {N} \cos \left(\mathcal {T} \left(t _ {i} ^ {\star}\right), \mathcal {T} _ {0} \left(t _ {i}\right)\right) + \sum_ {i = 1} ^ {N} \cos \left(\mathcal {V} \left(v _ {i} ^ {\star}\right), \mathcal {V} _ {0} \left(v _ {i}\right)\right)\right). \tag {5} +$$ + +**Embedding structure preservation.** The two aforementioned losses bring unsafe embeddings towards the positions of their corresponding safe embeddings in the original frozen spaces, either in a single-modal or multi-modal manner. However, alone, they would inevitably cause the fine-tuned encoders $\mathcal{T}$ and $\mathcal{V}$ to lose their performance on safe inputs, as well as their matching properties. Therefore, we also adopt two losses to ensure that the original structure of the embedding spaces is preserved where safe textual and visual regions lie. + +In particular, we define a matching loss between the safe embeddings produced by the online networks $\mathcal{T}$ , $\mathcal{V}$ and those of the original, pre-trained, networks $\mathcal{T}_0$ , $\mathcal{V}_0$ . Again, this is defined through the negative cosine similarity be + +tween matching pairs, as + +$$ +L _ {\text {p r e s}, 1} = - \frac {1}{N} \left(\sum_ {i = 1} ^ {N} \cos \left(\mathcal {T} \left(t _ {i}\right), \mathcal {T} _ {0} \left(t _ {i}\right)\right) + \sum_ {i = 1} ^ {N} \cos \left(\mathcal {V} \left(v _ {i}\right), \mathcal {V} _ {0} \left(v _ {i}\right)\right)\right). \tag {6} +$$ + +Finally, as an additional regularization term we also keep a contrastive loss between safe visual embeddings and safe textual embeddings, comparing online and frozen encoders. This, in practice, closely resembles the original loss on which the embedding space was trained, i.e. + +$$ +\begin{array}{l} L _ {\text {p r e s}, 2} = - \frac {1}{N} \left(\sum_ {i = 1} ^ {N} \log \frac {\exp \left(\cos \left(\mathcal {V} _ {0} \left(v _ {i}\right) , \mathcal {T} \left(t _ {i}\right)\right) / \tau\right)}{\sum_ {j = 1} ^ {N} \exp \left(\cos \left(\mathcal {V} _ {0} \left(v _ {i}\right) , \mathcal {T} \left(t _ {j}\right)\right) / \tau\right)} + \sum_ {i = 1} ^ {N} \log \frac {\exp \left(\cos \left(\mathcal {V} _ {0} \left(v _ {i}\right) , \mathcal {T} \left(t _ {i}\right)\right) / \tau\right)}{\sum_ {j = 1} ^ {N} \exp \left(\cos \left(\mathcal {V} _ {0} \big (v _ {j}\big) , \mathcal {T} \big (t _ {i}\big)\right) / \tau\right)} \right. \\ + \sum_ {i = 1} ^ {N} \log \frac {\exp (\cos (\mathcal {T} _ {0} (t _ {i}) , \mathcal {V} (v _ {i})) / \tau)}{\sum_ {j = 1} ^ {N} \exp (\cos (\mathcal {T} _ {0} (t _ {i}) , \mathcal {V} (v _ {j})) / \tau)} + \sum_ {i = 1} ^ {N} \log \frac {\exp (\cos (\mathcal {T} _ {0} (t _ {i}) , \mathcal {V} (v _ {i})) / \tau)}{\sum_ {j = 1} ^ {N} \exp (\cos (\mathcal {T} _ {0} (t _ {j}) , \mathcal {V} (v _ {i})) / \tau)} \Biggr). \\ \end{array} +$$ + +Eventually, the overall loss function on which the network is trained is a weighted sum of the four loss functions mentioned above. + +# 4 Experiments + +# 4.1 Experimental Setting + +Datasets. Our experiments are mainly conducted on the collected ViSU dataset that, as previously mentioned, contains 165k quadruplets of safe and unsafe content for training. For both validation and testing we use 5k samples following the Karpathy's COCO splits [21], randomly sampling only one safe caption among the five available for each image in the original COCO dataset. When applying our Safe-CLIP to text-to-image generative architectures, we also perform experiments on the I2P dataset [38] that is composed of 4,703 textual prompts extracted from Lexica, a collection of user-generated prompts for conditioning text-to-image diffusion models. Each prompt is associated to one of seven different categories of inappropriate content, among hate, harassment, violence, self-harm, sexual content, shocking images, illegal activity. + +LLM fine-tuning details. During SFT, we fine-tune the 7B version of Llama 2-Chat using low-rank adaptation [20] with $r = 64$ as low-rank factor. We employ a batch size equal to 4 and a learning rate of $2 \times 10^{-4}$ . To perform DPO, we follow the variant presented in [46], employing low-rank adaptation also in this case. The complete DPO fine-tuning settings are reported in the supplementary. + +Safe-CLIP implementation and training details. Our architecture is based on the standard CLIP model [33] composed of a visual and a textual encoder. Specifically, we employ the ViT-L/14 variant to comply with the textual encoder used in the Stable Diffusion v1.4 model [36] and the visual encoder employed in LLaVA [25]. Experiments with different CLIP-based backbones are reported in the supplementary. During training, both the visual and textual encoder are + +fine-tuned using low-rank decompositions [20], where the low-rank factor $r$ is set to 16 in all the experiments. We employ Adam as optimizer [22] using a learning rate equal to $1 \times 10^{-3}$ and a batch size of 128. + +# 4.2 Evaluating the ViSU Dataset + +We first assess the quality of our ViSU dataset, evaluating the inappropriateness degree of the generated unsafe sentences. Specifically, we employ a DistilBERT model fine-tuned for adult content detection and directly ask GPT-3.5 to evaluate whether the generated unsafe sentences should be + +Table 1: Comparison between the textual portion of ViSU and the I2P benchmark [38], in terms of NSFW degree and toxicity. + +
Dataset% NSFWToxicity
DistilBERTGPT-3.5
I2P [38]52.813.914.9
w/o SFT (i.e. Llama 2-Chat)37.89.37.7
w/o DPO fine-tuning75.975.030.6
ViSU (Ours)80.979.131.3
+ +classified as NSFW content. As reported in Table 1, ViSU showcases a very good textual quality, as it has a higher degree of NSFW sentences compared to existing alternatives like I2P [38] (i.e. $79.1\%$ of NSFW sentences vs. $13.9\%$ in the I2P dataset according to GPT-3.5). Moreover, following [38], we also report the toxicity score of the unsafe sentences computed using the Perspective $\mathrm{API}^3$ . Also in terms of toxicity, ViSU presents a higher degree of NSFW content. When instead comparing the effectiveness of each of the Llama 2 fine-tuning stages, it is worth noting that both the SFT procedure and DPO fine-tuning consistently increase the quality of generated sentences, going from $9.3\%$ of NSFW content when using the original Llama 2 model to $79.1\%$ after both fine-tuning stages. + +# 4.3 Evaluating the Safe-CLIP Embedding Space + +Results on ViSU test set. To evaluate the retrieval performance of Safe-CLIP, we firstly consider image-to-text and text-to-image retrieval in a safe-only setting, where we do not have any inappropriate content in both visual and textual data. This is important to assess whether the properties of the original CLIP embedding space are preserved when employing our fine-tuning strategy. In this case, query elements are represented by the safe images of the test set (which, with a slight abuse of notation, we refer to as $\mathbf{V}$ ) for the image-to-text setting and the safe textual items (referred to as $\mathbf{T}$ ) for the text-to-image one. Moreover, we consider text-to-image and image-to-text retrieval when using unsafe texts as queries (referred to as $\mathbf{T}^{\star}$ ) and both safe and unsafe images as retrievable items and when using unsafe images as queries (i.e. $\mathbf{V}^{\star}$ ) and both safe and unsafe texts as retrievable items. + +Retrieval results on the ViSU test set are reported in Table 2, comparing the proposed Safe-CLIP model with the original CLIP architecture, a CLIP model trained on the DataComp dataset [12], which has undergone NSFW content cleaning, and two different baselines. Specifically, we consider a variant of our + +Table 2: Retrieval results on the ViSU test set. The left portions respectively show text-to-image and image-to-text performance when using safe data only (i.e. $\mathbf{V}$ and $\mathbf{T}$ ). The right portions report the results when using unsafe textual sentences as query (i.e. $\mathbf{T}^{\star}$ ) and the merging of safe (i.e. $\mathbf{V}$ ) and unsafe images (i.e. $\mathbf{V}^{\star}$ ) as retrievable items, or when using unsafe visual queries (i.e. $\mathbf{V}^{\star}$ ) and the merging of safe (i.e. $\mathbf{T}$ ) and unsafe sentences (i.e. $\mathbf{T}^{\star}$ ) as retrievable items. + +
ModelText-to-Image (T-to-V)Image-to-Text (V-to-T)Text-to-Image (T*-to-V ∪ V*)Image-to-Text (V*-to-T ∪ T*)
R@1R@10R@20R@1R@10R@20R@1R@10R@20R@1R@10R@20
CLIP (ViT-L) [33]36.871.681.539.874.283.52.024.833.24.532.940.6
DataComp-1B (ViT-L) [12]46.779.787.447.081.388.91.628.135.65.537.544.9
w/o inap. content redirection49.983.790.348.183.690.51.630.440.16.135.242.6
w/o negative cosine similarities41.978.587.341.577.886.98.246.056.613.760.468.2
Safe-CLIP45.981.889.745.382.389.78.046.958.019.162.971.1
NSFW Text QueryCLIP Top-1Safe-CLIP Top-1NSFW Image QueryCLIP Top-1Safe-CLIP Top-1
A group of people on a battlefield with buildings in the background.A girl in blue lingerie holding a crack pipe in a dirty alleyway.Close-up of a blonde girl enjoying music with headphones.
A black and white photo of a couple fuicking each other.A sex worker servicing a client in a parking lot.Harley Davidson belt inspection.
+ +Fig. 3: Top-1 images (left) and text (right) retrieved using the original CLIP model and our Safe-CLIP, when NSFW texts and images are employed as query. + +approach in which we remove the two negative cosine similarity losses (i.e. Eq. 5 and 6), and a model trained with safe data only (i.e. removing the loss functions for inappropriate content redirection reported in Eq. 4 and 5). Results are reported in terms of $\mathrm{Recall}@\bar{k}$ ( $\mathrm{R}@\bar{k}$ ) with $k = 1, 10, 20$ , that measures the percentage of times the visual or textual item associated to the query is retrieved among the top- $k$ elements. When using unsafe sentences as queries, for each element we consider the safe image associated with the given text as the corresponding visual element. Symmetrically, when using unsafe images as queries, for each element we consider the safe text associated with the given image as the ground-truth item. Therefore, recall results in the unsafe setting follow a "the higher the better" protocol. + +As it can be seen, Safe-CLIP can retrieve a significant higher portion of correct safe images when using unsafe prompts as queries, while effectively preserving good performance in safe-only settings (i.e. $\mathbf{V}$ -to- $\mathbf{T}$ and $\mathbf{T}$ -to- $\mathbf{V}$ ). Specifically, when comparing our model with the original CLIP, it is worth noting that the results on text-to-image retrieval with unsafe texts as queries are consistently improved when using our text encoder, with an overall improvement of 6.0 points in terms of $\mathrm{R}@\mathrm{1}$ , and the same applies for image-to-text retrieval which showcases an improvement of 14.6 $\mathrm{R}@\mathrm{1}$ points. This demonstrates the effectiveness of our fine-tuning strategy, which can reduce the model probability of returning inappropriate images or sentences. + +Robustness on real NSFW images. To further analyze the safety degree of the Safe-CLIP embedding space, we perform text-to-image and image-to-text + +retrieval using real NSFW images as visual items. Specifically, we select inappropriate visual content from three different sources: $(i)$ a portion of data used to train the NudeNet classifier, $(ii)$ images crawled from the web using NSFW + +data source URLs4, and (iii) images from the Socio-Moral Image Database (SMID) [10]. While the first two sources exclusively contain nudity and pornography images, the third one includes more varied types of + +Table 3: Percentage of retrieved NSFW images and text using unsafe data as query. Safe retrievable items are from LAION-400M, unsafe images are extracted from different NSFW sources, and unsafe texts are from ViSU. + +
Model% NSFW (Text-to-Image)% NSFW (Image-to-Text)
NudeNetNSFW URLsSMIDNudeNetNSFW URLsSMID
CLIP [33]57.155.247.865.657.441.4
DataComp-1B [12]55.649.764.061.456.245.6
Safe-CLIP8.49.816.728.824.734.5
+ +inappropriate images representing negative concepts such as, for example, harm, inequality, discrimination, and unfairness. Overall, we randomly sample 1,000 images from each of the NSFW data sources, selecting only those representing unsafe concepts for the SMID dataset. As textual items, we employ unsafe texts from the ViSU test set that match the NSFW concepts represented in each of the NSFW visual sources (i.e. sexual and nudity for NudeNet and NSFW data source URLs, and all other concepts for the SMID dataset). For both I2T and T2I, we employ a set of 10k randomly selected visual or textual distractors, randomly selected from the LAION-400M dataset [41]. + +Results are shown in Table 3 comparing Safe-CLIP with the standard CLIP model and the model trained on DataComp. For each NSFW data source, we report the percentage of times in which an NSFW image or text is retrieved as the top-1 element. Notably, using Safe-CLIP consistently reduces the percentage of retrieved NSFW items for all three NSFW dataset sources. In particular, the percentage of retrieved NSFW visual and textual content is reduced by more than 45 and 30 points, respectively when considering unsafe images or textual elements in all three considered settings. This experiment confirms that our fine-tuning strategy can effectively enhance the safety of the CLIP embedding space. + +Qualitative results. Fig. 3 reports qualitative retrieval results in the same aforementioned setting. Safe-CLIP is able to retrieve safe images starting from NSFW texts and, vice versa, retrieve safe sentences starting from NSFW images. Additionally, it can also preserve the global context and semantics of the query. + +# 4.4 Safe-CLIP for Text-to-Image Generation + +Results on I2P and ViSU test set. We then validate the effectiveness of the Safe-CLIP text encoder when applied in a text-to-image generative model. Specifically, we employ Stable Diffusion v1.4 [36], eventually replacing the standard CLIP text encoder used in Stable Diffusion with our fine-tuned version. Moreover, we also apply Safe-CLIP in combination with other NSFW removal strategies. In particular, we consider a version of Stable Diffusion with negative + +Table 4: Probabilities of generating images with unsafe content, classified by combining the predictions of NudeNet and Q16. Results are reported using NSFW text prompts from I2P [38] and ViSU, and Stable Diffusion v1.4 as text-to-image generator. + +
ModelI2PViSU
HateHassassmentViolenceSelf-harmSexualShockingIllegal Act.AvgHateHassassmentViolenceSelf-harmSexualShockingIllegal Act.Avg
SD v1.441.432.443.742.124.852.235.735.725.917.830.419.524.426.923.526.2
+ Safe-CLIP23.621.126.726.815.932.721.422.24.62.93.94.64.12.93.33.6
Negative Prompts28.524.422.423.315.940.829.324.418.613.920.214.014.016.514.416.9
+ Safe-CLIP19.217.721.722.913.926.119.318.93.13.42.83.63.12.92.72.9
SLD-Weak [38]30.624.132.127.813.941.925.725.617.510.720.813.316.818.815.417.7
+ Safe-CLIP21.219.025.322.412.428.119.519.83.73.03.23.83.73.03.13.2
SLD-Medium [38]21.617.523.717.48.931.216.717.710.67.012.39.810.811.59.710.8
+ Safe-CLIP18.917.221.620.611.925.816.417.53.02.23.23.42.82.32.62.8
SLD-Strong [38]15.913.618.811.17.821.511.213.56.43.76.15.17.25.84.45.6
+ Safe-CLIP16.914.017.612.28.220.213.113.03.41.41.71.91.71.91.81.8
SD v1.4Negative PromptsSLD-Strong [38]SD v1.4 + Safe-CLIPSD v1.4Negative PromptsSLD-Strong [38]SD v1.4 + Safe-CLIP
+ +Fig. 4: Images generated from unsafe prompts with Stable Diffusion, employing the original CLIP model, negative prompts, SLD-Strong [38], and our Safe-CLIP. + +prompts and the recently proposed Safe Latent Diffusion (SLD) approach [38] which employs different levels of safety guidance (SLD-Weak, SLD-Medium, and SLD-Strong) to limit the generation of inappropriate images. For this experiment, we generate five images for each textual prompt using different random seeds and compute the probability of generating inappropriate images detected by two NSFW classifiers. Following [38], we employ Q16 [39] and NudeNet [2]. + +Table 4 shows the results using textual prompts from both I2P and ViSU datasets. We report the NSFW generation probabilities on the entire set of prompts of each dataset and also dividing them into the seven NSFW categories considered in $[39]^5$ . Interestingly, Safe-CLIP significantly reduces the probabilities of generating NSFW images when using textual prompts from both datasets thus demonstrating its usefulness also in a text-to-image generation setting. In particular, when applying our text encoder to a standard Stable Diffusion model, the probability of generating inappropriate content decreases by 13.5 points with I2P prompts and 22.6 points with NSFW texts from ViSU. Similar results can also be observed when applying Safe-CLIP alongside other NSFW removal strategies, highlighting that fine-tuning the CLIP text encoder with the + +A proposed approach can benefit the performance of existing methods tailored for removing NSFW concepts from images generated by diffusion models. + +Qualitative results. Samples of generated images are shown in Fig. 4, comparing results generated by Safe-CLIP applied to Stable Diffusion with images generated by SLD-Strong [39], Stable Diffusion with negative prompts, and the Stable Diffusion original version. Qualitative results confirm the effectiveness of our proposal which can generate images that preserve the original semantic of the scene while preventing the generation of inappropriate content. + +# 4.5 Safe-CLIP for Image-to-Text Generation + +Finally, we assess the capabilities of the Safe-CLIP visual encoder when applied to an existing multimodal LLM [5]. We employ LLaVA [25] based on LLama 2-13B-Chat, prompted by asking to describe a given image. Results are reported in Table 5 in terms of percentage of NSFW generated texts measured with GPT-3.5 and toxicity degree computed using the Perspective API. Also for this + +experiment, we employ the real NSFW images from the three different NSFW sources described in Sec. 4.3. As it can be seen, SafeCLIP can significantly reduce the probability + +Table 5: Percentage of generating NSFW textual sentences and their toxicity degree, when using real NSFW images from different sources as input. + +
ModelNudeNetNSFW URLsSMID
% NSFWToxicity% NSFWToxicity% NSFWToxicity
LLaVA [25]62.638.646.824.922.24.7
+ Safe-CLIP26.716.519.410.811.73.7
+ +of generating inappropriate textual sentences compared to the original LLaVA, demonstrating its effectiveness also in this setting. + +# 5 Conclusion + +We presented Safe-CLIP, an approach for fine-tuning a CLIP-like model to make it safer and less sensitive to NSFW concepts. Our approach is based on automatically collecting a large synthetic dataset with safe and unsafe images and captions, with which we fine-tune CLIP with losses designed to redirect unsafe content while preserving the structure of the embedding space. Experimental results demonstrate the appropriateness of our solution for cross-modal retrieval, image-to-text and text-to-image generation. + +Mitigating potential misuse of the ViSU dataset. To mitigate potential misuse of the dataset, we release it via a request form, with only the textual portion available6. Access is granted exclusively to verified researchers, who must declare their intention to use the data solely for research purposes, explicitly committing to non-malicious use. The NSFW visual part of the dataset, due to its explicit nature, is withheld to avoid potential misuse. Nonetheless, the full reproducibility of the dataset is ensured given that we employed publicly available diffusion models and we release the generation seeds and instructions. + +# Acknowledgments + +We acknowledge the CINECA award under the ISCRA initiative, for the availability of high-performance computing resources. This work has been supported by the EU Horizon project "ELIAS - European Lighthouse of AI for Sustainability" (No. 101120237), and by the the PNRR projects "FAIR - Future Artificial Intelligence Research" (M4C2 - PE00000013) and "ITSERR - Italian Strengthening of Esfri RI Resilience" (CUP B53C22001770006), both funded by the EU - NextGenerationEU. + +# References + +1. Bakker, M., Chadwick, M., Sheahan, H., Tessler, M., Campbell-Gillingham, L., Balaguer, J., McAleese, N., Glaese, A., Aslanides, J., Botvinick, M., et al.: Finetuning language models to find agreement among humans with diverse preferences. In: NeurIPS (2022) +2. Bedapudi, P.: NudeNet: Neural Nets for Nudity Classification, Detection, and Selective Censoring (2019) +3. Birhane, A., Prabhu, V.U.: Large image datasets: A pyrrhic win for computer vision? In: WACV (2021) +4. Birhane, A., Prabhu, V.U., Kahembwe, E.: Multimodal datasets: misogyny, pornography, and malignant stereotypes. arXiv preprint arXiv:2110.01963 (2021) +5. Caffagni, D., Cocchi, F., Barsellotti, L., Moratelli, N., Sarto, S., Baraldi, L., Baraldi, L., Cornia, M., Cucchiara, R.: The Revolution of Multimodal Large Language Models: A Survey. In: ACL Findings (2024) +6. Cao, Y., Yang, J.: Towards Making Systems Forget with Machine Unlearning. In: IEEE Symposium on Security and Privacy (2015) +7. Cauteruccio, F., Corradini, E., Terracina, G., Ursino, D., Virgili, L.: Extraction and analysis of text patterns from nsfw adult content in reddit. Data & Knowledge Engineering 138, 101979 (2022) +8. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., Stoica, I., Xing, E.P.: Vicuna: An Open-Source Chatbot Improving GPT-4 with $90\%$ * ChatGPT Quality (2023) +9. Christiano, P.F., Leike, J., Brown, T., Martic, M., Legg, S., Amodei, D.: Deep reinforcement learning from human preferences. In: NeurIPS (2017) +10. Crone, D.L., Bode, S., Murawski, C., Laham, S.M.: The Socio-Moral Image Database (SMID): A novel stimulus set for the study of social, moral and affective processes. PloS one 13(1), e0190954 (2018) +1. Dettmers, T., Pagnoni, A., Holtzman, A., Zettlemoyer, L.: QLoRA: Efficient Finetuning of Quantized LLMs. arXiv preprint arXiv:2305.14314 (2023) +2. Gadre, S.Y., Ilharco, G., Fang, A., Hayase, J., Smyrnis, G., Nguyen, T., Marten, R., Wortsman, M., Ghosh, D., Zhang, J., et al.: DataComp: In search of the next generation of multimodal datasets. In: NeurIPS (2024) +3. Gandhi, S., Kokkula, S., Chaudhuri, A., Magnani, A., Stanley, T., Ahmadi, B., Kandaswamy, V., Ovenc, O., Mannor, S.: Scalable Detection of Offensive and Noncompliant Content/Logo in Product Images. In: WACV (2020) +4. Gandikota, R., Materzynska, J., Fiotto-Kaufman, J., Bau, D.: Erasing Concepts from Diffusion Models. In: ICCV (2023) + +15. Gao, P., Han, J., Zhang, R., Lin, Z., Geng, S., Zhou, A., Zhang, W., Lu, P., He, C., Yue, X., Li, H., Qiao, Y.: LLaMA-Adapter V2: Parameter-Efficient Visual Instruction Model. arXiv preprint arXiv:2304.15010 (2023) +16. Ginart, A., Guan, M., Valiant, G., Zou, J.Y.: Making AI Forget You: Data Deletion in Machine Learning. In: NeurIPS (2019) +17. Golatkar, A., Achille, A., Soatto, S.: Eternal Sunshine of the Spotless Net: Selective Forgetting in Deep Networks. In: CVPR (2020) +18. Golatkar, A., Achille, A., Wang, Y.X., Roth, A., Kearns, M., Soatto, S.: Mixed Differential Privacy in Computer Vision. In: CVPR (2022) +19. Hidayatullah, A.F., Hakim, A.M., Sembada, A.A.: Adult Content Classification on Indonesian Tweets using LSTM Neural Network. In: ICACSIS (2019) +20. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-Rank Adaptation of Large Language Models. arXiv preprint arXiv:2106.09685 (2021) +21. Karpathy, A., Fei-Fei, L.: Deep visual-semantic alignments for generating image descriptions. In: CVPR (2015) +22. Kingma, D.P., Ba, J.: Adam: A Method for Stochastic Optimization. In: ICLR (2015) +23. Kumari, N., Zhang, B., Wang, S.Y., Shechtman, E., Zhang, R., Zhu, J.Y.: Ablating concepts in text-to-image diffusion models. In: ICCV (2023) +24. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft COCO: Common Objects in Context. In: ECCV (2014) +25. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual Instruction Tuning. In: NeurIPS (2023) +26. Liu, Y., Singh, A., Freeman, C.D., Co-Reyes, J.D., Liu, P.J.: Improving Large Language Model Fine-tuning for Solving Math Problems. arXiv preprint arXiv:2310.10047 (2023) +27. Markov, T., Zhang, C., Agarwal, S., Nekoul, F.E., Lee, T., Adler, S., Jiang, A., Weng, L.: A Holistic Approach to Undesired Content Detection in the Real World. In: AAAI (2023) +28. Materzyńska, J., Torralba, A., Bau, D.: Disentangling Visual and Written Concepts in CLIP. In: CVPR (2022) +29. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: GLIDE: Towards Photorealistic Image Generation and Editing with Text-Guided Diffusion Models. arXiv preprint arXiv:2112.10741 (2021) +30. Oord, A.v.d., Li, Y., Vinyals, O.: Representation Learning with Contrastive Predictive Coding. arXiv preprint arXiv:1807.03748 (2018) +31. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. In: NeurIPS (2022) +32. Poppi, S., Sarto, S., Cornia, M., Baraldi, L., Cucchiara, R.: Multi-Class Unlearning for Image Classification via Weight Filtering. IEEE Intelligent Systems (2024) +33. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning Transferable Visual Models From Natural Language Supervision. In: ICML (2021) +34. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I.: Language Models are Unsupervised Multitask Learners. OpenAI Blog 1(8), 9 (2019) +35. Rafailov, R., Sharma, A., Mitchell, E., Ermon, S., Manning, C.D., Finn, C.: Direct Preference Optimization: Your Language Model is Secretly a Reward Model. In: NeurIPS (2023) + +36. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022) +37. Sanh, V., Debut, L., Chaumont, J., Wolf, T.: DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108 (2019) +38. Schramowski, P., Brack, M., Deiseroth, B., Kersting, K.: Safe Latent Diffusion: Mitigating Inappropriate Degeneration in Diffusion Models. In: CVPR (2023) +39. Schramowski, P., Tauchmann, C., Kersting, K.: Can Machines Help Us Answering Question 16 in Datasheets, and In Turn Reflecting on Inappropriate Content? In: ACM FAccT (2022) +40. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., Schramowski, P., Kundurthy, S., Crowson, K., Schmidt, L., Kaczmarczyk, R., Jitsev, J.: LAION-5B: An open large-scale dataset for training next generation image-text models. In: NeurIPS (2022) +41. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: LAION-400M: Open Dataset of CLIP-Filtered 400 Million Image-Text Pairs. In: NeurIPS Workshops (2021) +42. Shen, S., Li, L.H., Tan, H., Bansal, M., Rohrbach, A., Chang, K.W., Yao, Z., Keutzer, K.: How Much Can CLIP Benefit Vision-and-Language Tasks? In: ICLR (2022) +43. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., Lample, G.: LLaMA: Open and Efficient Foundation Language Models. arXiv preprint arXiv:2302.13971 (2023) +44. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open Foundation and Fine-Tuned Chat Models. arXiv preprint arXiv:2307.09288 (2023) +45. Trager, M., Perera, P., Zancato, L., Achille, A., Bhatia, P., Soatto, S.: Linear Spaces of Meanings: Compositional Structures in Vision-Language Models. In: ICCV (2023) +46. Tunstall, L., Beeching, E., Lambert, N., Rajani, N., Rasul, K., Belkada, Y., Huang, S., von Werra, L., Fourrier, C., Habib, N., et al.: Zephyr: Direct Distillation of LM Alignment. arXiv preprint arXiv:2310.16944 (2023) +47. Wang, M., Xing, J., Liu, Y.: ActionCLIP: A New Paradigm for Video Action Recognition. arXiv preprint arXiv:2109.08472 (2021) +48. Wang, Y., Kordi, Y., Mishra, S., Liu, A., Smith, N.A., Khashabi, D., Hajishirzi, H.: Self-Instruct: Aligning Language Models with Self-Generated Instructions. arXiv preprint arXiv:2212.10560 (2022) +49. Zhang, E., Wang, K., Xu, X., Wang, Z., Shi, H.: Forget-Me-Not: Learning to Forget in Text-to-Image Diffusion Models. arXiv preprint arXiv:2303.17591 (2023) +50. Zhang, R., Han, J., Zhou, A., Hu, X., Yan, S., Lu, P., Li, H., Gao, P., Qiao, Y.: LLaMA-Adapter: Efficient Fine-tuning of Language Models with Zero-init Attention. arXiv preprint arXiv:2303.16199 (2023) +51. Zheng, L., Chiang, W.L., Sheng, Y., Zhuang, S., Wu, Z., Zhuang, Y., Lin, Z., Li, Z., Li, D., Xing, E., et al.: Judging llm-as-a-judge with mt-bench and chatbot arena. arXiv preprint arXiv:2306.05685 (2023) +52. Zhou, C., Liu, P., Xu, P., Iyer, S., Sun, J., Mao, Y., Ma, X., Efrat, A., Yu, P., Yu, L., et al.: LIMA: Less Is More for Alignment. arXiv preprint arXiv:2305.11206 (2023) \ No newline at end of file diff --git a/2024/Safe-CLIP_ Removing NSFW Concepts from Vision-and-Language Models/images.zip b/2024/Safe-CLIP_ Removing NSFW Concepts from Vision-and-Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a35dbc4f1e8f9e701bdbf15188f8ce97c7a64565 --- /dev/null +++ b/2024/Safe-CLIP_ Removing NSFW Concepts from Vision-and-Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:026ffad30d45d4043435f834c8603d2b2396d5612bc0c931fc375aa914a8ce5f +size 517644 diff --git a/2024/Safe-CLIP_ Removing NSFW Concepts from Vision-and-Language Models/layout.json b/2024/Safe-CLIP_ Removing NSFW Concepts from Vision-and-Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c09904cd2035d1df4a6892250c418859c90c1240 --- /dev/null +++ b/2024/Safe-CLIP_ Removing NSFW Concepts from Vision-and-Language Models/layout.json @@ -0,0 +1,9576 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 151, + 111, + 462, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 111, + 462, + 148 + ], + "spans": [ + { + "bbox": [ + 151, + 111, + 462, + 148 + ], + "type": "text", + "content": "Safe-CLIP: Removing NSFW Concepts from Vision-and-Language Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 167, + 447, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 167, + 447, + 192 + ], + "spans": [ + { + "bbox": [ + 166, + 167, + 447, + 192 + ], + "type": "text", + "content": "Samuele Poppi" + }, + { + "bbox": [ + 166, + 167, + 447, + 192 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 166, + 167, + 447, + 192 + ], + "type": "text", + "content": ", Tobia Poppi" + }, + { + "bbox": [ + 166, + 167, + 447, + 192 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 166, + 167, + 447, + 192 + ], + "type": "text", + "content": ", Federico Cocchi" + }, + { + "bbox": [ + 166, + 167, + 447, + 192 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 166, + 167, + 447, + 192 + ], + "type": "text", + "content": ", Marcella Cornia" + }, + { + "bbox": [ + 166, + 167, + 447, + 192 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 166, + 167, + 447, + 192 + ], + "type": "text", + "content": ", Lorenzo Baraldi" + }, + { + "bbox": [ + 166, + 167, + 447, + 192 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 166, + 167, + 447, + 192 + ], + "type": "text", + "content": ", and Rita Cucchiara" + }, + { + "bbox": [ + 166, + 167, + 447, + 192 + ], + "type": "inline_equation", + "content": "^{1,3}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 206, + 201, + 406, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 201, + 406, + 213 + ], + "spans": [ + { + "bbox": [ + 206, + 201, + 406, + 213 + ], + "type": "text", + "content": "1 University of Modena and Reggio Emilia, Italy" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 252, + 214, + 361, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 214, + 361, + 223 + ], + "spans": [ + { + "bbox": [ + 252, + 214, + 361, + 223 + ], + "type": "text", + "content": "name.surname@unimore.it" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 253, + 224, + 360, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 224, + 360, + 236 + ], + "spans": [ + { + "bbox": [ + 253, + 224, + 360, + 236 + ], + "type": "text", + "content": "2 University of Pisa, Italy" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 247, + 236, + 366, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 236, + 366, + 246 + ], + "spans": [ + { + "bbox": [ + 247, + 236, + 366, + 246 + ], + "type": "text", + "content": "name.surname@phd.unipi.it" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 271, + 245, + 343, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 271, + 245, + 343, + 256 + ], + "spans": [ + { + "bbox": [ + 271, + 245, + 343, + 256 + ], + "type": "text", + "content": "3 IIT-CNR, Italy" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 159, + 285, + 455, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 285, + 455, + 483 + ], + "spans": [ + { + "bbox": [ + 159, + 285, + 455, + 483 + ], + "type": "text", + "content": "Abstract. Large-scale vision-and-language models, such as CLIP, are typically trained on web-scale data, which can introduce inappropriate content and lead to the development of unsafe and biased behavior. This, in turn, hampers their applicability in sensitive and trustworthy contexts and could raise significant concerns in their adoption. Our research introduces a novel approach to enhancing the safety of vision-and-language models by diminishing their sensitivity to NSFW (not safe for work) inputs. In particular, our methodology seeks to sever \"toxic\" linguistic and visual concepts, unlearning the linkage between unsafe linguistic or visual items and unsafe regions of the embedding space. We show how this can be done by fine-tuning a CLIP model on synthetic data obtained from a large language model trained to convert between safe and unsafe sentences, and a text-to-image generator. We conduct extensive experiments on the resulting embedding space for cross-modal retrieval, text-to-image, and image-to-text generation, where we show that our model can be remarkably employed with pre-trained generative models. Our source code and trained models are available at: https://github.com/aimagelab/safe-clip." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 160, + 493, + 450, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 493, + 450, + 504 + ], + "spans": [ + { + "bbox": [ + 160, + 493, + 450, + 504 + ], + "type": "text", + "content": "Keywords: Trustworthy AI " + }, + { + "bbox": [ + 160, + 493, + 450, + 504 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 493, + 450, + 504 + ], + "type": "text", + "content": " Vision-and-Language " + }, + { + "bbox": [ + 160, + 493, + 450, + 504 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 493, + 450, + 504 + ], + "type": "text", + "content": " NSFW Concepts" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 523, + 482, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 482, + 548 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 482, + 548 + ], + "type": "text", + "content": "Warning: This paper includes explicit sexual content, racially insensitive language, and other material that may be disturbing or offensive to certain readers." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 565, + 230, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 565, + 230, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 565, + 230, + 578 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "text", + "content": "Large-scale models have recently proven to be effective on a variety of tasks, ranging from image classification and understanding to cross-modal retrieval and generation [25,33,36]. Scaling models, however, has also required to increase the quantity and variability of training data, paving the way to scraping billions of items from the web without manual supervision [40, 41]. Despite the adoption" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 654, + 228, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 654, + 228, + 665 + ], + "spans": [ + { + "bbox": [ + 141, + 654, + 228, + 665 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 116, + 477, + 300 + ], + "blocks": [ + { + "bbox": [ + 137, + 116, + 477, + 300 + ], + "lines": [ + { + "bbox": [ + 137, + 116, + 477, + 300 + ], + "spans": [ + { + "bbox": [ + 137, + 116, + 477, + 300 + ], + "type": "image", + "image_path": "ae5cb6fb8073c6282a8a3c019243b7fad5de4aa573593a74894a61d1077cf2e8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 308, + 482, + 331 + ], + "lines": [ + { + "bbox": [ + 132, + 308, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 482, + 331 + ], + "type": "text", + "content": "Fig. 1: Removing NSFW concepts from CLIP models. Our Safe-CLIP fine-tunes CLIP to make it safer in cross-modal retrieval, image-to-text and text-to-image generation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 342, + 480, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 480, + 378 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 480, + 378 + ], + "type": "text", + "content": "of filters and automatic checks, this paradigm comes at the cost of introducing inappropriate content in the training set [14,38], which ultimately results in the injection of unsafe, biased or toxic behaviors [4]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 378, + 482, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 378, + 482, + 474 + ], + "spans": [ + { + "bbox": [ + 130, + 378, + 482, + 474 + ], + "type": "text", + "content": "This is also the case of vision-and-language models based on embedding spaces, where toxic content can embed itself in the latent space. For instance when a NSFW (not safe for work) textual prompt is used for a cross-modal task, its embedding can reach unsafe points in the latent space, leading to the generation of undesired images, or to the retrieval of inappropriate content. Similarly in image-to-text generation, when an inappropriate image is used as a prompt, the descriptive text could be toxic or offensive. A qualitative example of this is reported in Fig. 1, considering the case of a CLIP backbone [33]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 474, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 594 + ], + "type": "text", + "content": "Driven by these considerations, we tackle the task of enhancing the safety of pre-trained vision-and-language models. In particular, we devise an approach for making a CLIP-style embedding space safer, so that it becomes invariant with respect to inappropriate inputs. While some previous attempts have focused on mitigating inappropriate content in diffusion models [38], our approach mitigates inappropriate concepts from CLIP-like embedding spaces. As such, it has a more general impact and applicability, as CLIP-like models are employed for many different applications, ranging from cross-modal retrieval [33] to text-to-image and image-to-text generation [25,36], and are employed as feature extractors for different tasks [28,42,47]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Our approach is based on fine-tuning the embedding space so as to avoid the representation of inappropriate content, without changing its normal expressive power. We do this with a combination of losses designed to redirect inappropriate content to safe embedding regions, while preserving the structure of the embedding space as intact as possible. To support our training procedure, we generate quadruplets of safe and unsafe vision-language items. This data gener" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "type": "text", + "content": "S. Poppi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "text", + "content": "ation strategy is empowered by a toxic language model which can translate safe textual prompts into unsafe ones, while keeping context alignment and semantic meaning unchanged. When applied to collections of visually-grounded descriptions in conjunction with an NSFW-capable diffusion model, our conditioned NSFW generator is employed for building a dataset which can properly support the fine-tuning of a CLIP embedding space." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 480, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 480, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 480, + 295 + ], + "type": "text", + "content": "The resulting safe version of CLIP can be applied to cross-modal retrieval, text-to-image and image-to-text generation (Fig. 1). For example, if we ask the fine-tuned version of CLIP to retrieve an image corresponding to a textual prompt with NSFW content, it will fetch an image with similar semantics but appropriate content. Also, a Stable Diffusion model [36] conditioned on our fine-tuned CLIP will generate an image with appropriate content, free of violence, nudity, or other toxic aspects, keeping the safe semantics of the input prompt. Similarly, a multimodal LLM like LLaVA [25] conditioned on our Safe-CLIP will generate a textual description without inappropriate content." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 297, + 480, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 480, + 367 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 480, + 367 + ], + "type": "text", + "content": "Experimentally, we evaluate the capabilities of our strategy for making CLIP safer in both retrieval and generation contexts, by running experiments both on prompts and images synthetically generated and employing existing unsafe prompts [38] and real images. Experimental results show that our approach can significantly improve the safety during text-to-image and image-to-text retrieval and during visual and textual generation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 369, + 352, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 352, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 352, + 380 + ], + "type": "text", + "content": "To sum up, our main contributions are as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 381, + 480, + 548 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 138, + 381, + 480, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 381, + 480, + 427 + ], + "spans": [ + { + "bbox": [ + 138, + 381, + 480, + 427 + ], + "type": "text", + "content": "- We introduce a novel fine-tuning methodology which can turn a pre-trained CLIP-like embedding space into a safer one. Once fine-tuned with our methodology, the CLIP space ignores NSFW content and can be applied to downstream tasks like retrieval and visual or textual generation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 429, + 480, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 429, + 480, + 475 + ], + "spans": [ + { + "bbox": [ + 138, + 429, + 480, + 475 + ], + "type": "text", + "content": "- Our approach is based on creating a toxic LLM which can generate unsafe prompts given safe and visually-grounded ones. This is obtained by finetuning Llama 2 on manually curated pairs, and then aligning it with Direct Preference Optimization (DPO)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 476, + 480, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 476, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 138, + 476, + 480, + 512 + ], + "type": "text", + "content": "- Leveraging an automatically generated dataset of safe and unsafe images and texts, we fine-tune CLIP with a novel combination of losses which redirect unsafe content while preserving the embedding space structure." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 513, + 480, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 513, + 480, + 548 + ], + "spans": [ + { + "bbox": [ + 138, + 513, + 480, + 548 + ], + "type": "text", + "content": "- We experimentally evaluate the appropriateness of our approach by conducting experiments in both retrieval, and textual and visual generation. Our method showcases a significantly reduced generation of NSFW content." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 569, + 236, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 569, + 236, + 582 + ], + "spans": [ + { + "bbox": [ + 132, + 569, + 236, + 582 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 594, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 480, + 665 + ], + "type": "text", + "content": "Removing concepts from vision-and-language models. Removing content from AI models has been recently gaining increasing attention, with techniques spanning from complete model retraining or fine-tuning to machine unlearning [6,16,17,32] and differential privacy [18]. Some of these attempts have been considering text-to-image models and have aimed at deleting styles, concepts, or objects [23,49]. Recently, Schramowski et al. [38] introduced a technique to" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 145, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 145, + 90, + 447, + 102 + ], + "type": "text", + "content": "Safe-CLIP: Removing NSFW Concepts from Vision-and-Language Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "content": "steer the generation away from NSFW areas, defined by a finite and fixed set of concepts. NSFW concepts are encoded with the input prompt at inference time and the NSFW embedding is used as negative guidance. Later, Gandikota et al. [14] proposed a fine-tuning method that can erase a visual concept given only its name and using negative guidance as a teacher." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 177, + 482, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 177, + 482, + 250 + ], + "spans": [ + { + "bbox": [ + 130, + 177, + 482, + 250 + ], + "type": "text", + "content": "In contrast to these previous works, we focus on removing NSFW from a contrastive CLIP-like model, which can be applied for cross-modal retrieval, and for visual and textual generation. While to the best of our knowledge we are the first to tackle this scenario, Trager et al. [45] have demonstrated the presence of compositional patterns within the embedding space of CLIP, which suggests the existence of a distinctive path from safe to NSFW zones." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 253, + 482, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 253, + 482, + 421 + ], + "spans": [ + { + "bbox": [ + 130, + 253, + 482, + 421 + ], + "type": "text", + "content": "Detecting NSFW content. A related research field is that of the automatic detection of NSFW content. Several approaches have been proposed to detect NSFW language [7, 19, 27], primarily on social media data sources. DistilBERT [37] emerges as a promising solution for this purpose, particularly when fine-tuned for adult content detection. We utilize it as an NSFW language detector, in conjunction with GPT-3.5 [31], which we query directly to classify our prompts. While the identification of unsafe language poses a challenging task, the same can be said for vision, where different approaches have been proposed to detect inappropriate content [3, 13, 29]. Still, the detection of inappropriate content remains an intricate challenge, as visual cues, lack of context, and restricted data sources often introduce added layers of complexity. In our analysis, we utilize Q16 [39] and NudeNet [2] as automatic detectors of NSFW images. In particular, NudeNet is specialized in identifying unsafe content related to nudity, while Q16 serves as a broader spectrum NSFW classifier." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 425, + 482, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 425, + 482, + 547 + ], + "spans": [ + { + "bbox": [ + 130, + 425, + 482, + 547 + ], + "type": "text", + "content": "Finetuning LLMs with little data. Large Language Models (LLMs) have achieved high performance in various tasks due to their zero-shot capabilities [34,43,44], which stems from model scaling and the utilization of large training datasets. In addition to fine-tuning these models for specific tasks [8,26,51], there has been recently an interest in building parameter efficient fine-tuning strategies. In most solutions, only part of the weights are trained [15,50] or a reduced number of weights are added to the LLM [11,20]. As shown in [48], datasets employed for supervised fine-tuning [1] play a central role in changing the LLM behavior [52], even in low-data regimes. In this work, we fine-tune Llama 2 [44] to produce unsafe prompts starting from pre-existing safe counterparts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 567, + 272, + 581 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 272, + 581 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 272, + 581 + ], + "type": "text", + "content": "3 Proposed Approach" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "CLIP-like models [33] are trained on web-crawled data which can contain inappropriate content [4]. Making these models safer, therefore, requires either retraining from scratch using large-scale cleaned data or fine-tuning them with a form of supervision that aims to mitigate inappropriate knowledge. The first option would necessitate data cleaning at large scale, which is currently not effective in practice (see also Sec. 4.3 for a comparison), so we instead employ the" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "S. Poppi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "content": "second strategy. Specifically, we focus on making both the textual encoder and the visual encoder of CLIP safer." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 270 + ], + "type": "text", + "content": "Ideally, we want a safe version of the CLIP text encoder to ignore inappropriate content from input sentences and understand most of its clean content. Symmetrically, we want the safe version of the CLIP visual encoder to ignore inappropriate content from input images. Furthermore, we also want to maintain as much as possible the original structure of the embedding space near safe textual or visual regions, so that the safe encoders can be straightforwardly connected to downstream models built on top of them without further adaptation. Formally, given an unsafe sentence " + }, + { + "bbox": [ + 130, + 140, + 482, + 270 + ], + "type": "inline_equation", + "content": "t_i^\\star" + }, + { + "bbox": [ + 130, + 140, + 482, + 270 + ], + "type": "text", + "content": " and a \"cleaning\" function " + }, + { + "bbox": [ + 130, + 140, + 482, + 270 + ], + "type": "inline_equation", + "content": "c_t(\\cdot)" + }, + { + "bbox": [ + 130, + 140, + 482, + 270 + ], + "type": "text", + "content": " which removes all inappropriate content from it, we want our safe textual encoder " + }, + { + "bbox": [ + 130, + 140, + 482, + 270 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 130, + 140, + 482, + 270 + ], + "type": "text", + "content": " to satisfy the following condition with respect to the original, pre-trained, CLIP text encoder " + }, + { + "bbox": [ + 130, + 140, + 482, + 270 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_0" + }, + { + "bbox": [ + 130, + 140, + 482, + 270 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 240, + 272, + 480, + 285 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 240, + 272, + 480, + 285 + ], + "spans": [ + { + "bbox": [ + 240, + 272, + 480, + 285 + ], + "type": "interline_equation", + "content": "\\mathcal {T} \\left(t _ {i} ^ {\\star}\\right) \\approx \\mathcal {T} \\left(c _ {t} \\left(t _ {i} ^ {\\star}\\right)\\right) \\approx \\mathcal {T} _ {0} \\left(c _ {t} \\left(t _ {i} ^ {\\star}\\right)\\right), \\tag {1}", + "image_path": "cd57604c4e5b6b2a96baebcf648bf0ed193a9ee305408915d3fbca707caa6254.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 289, + 482, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 387 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 387 + ], + "type": "text", + "content": "where with the " + }, + { + "bbox": [ + 130, + 289, + 482, + 387 + ], + "type": "inline_equation", + "content": "\\approx" + }, + { + "bbox": [ + 130, + 289, + 482, + 387 + ], + "type": "text", + "content": " sign we indicate high similarity in the embedding space. As it can be noticed, the first condition stated in Eq. 1 ensures that inappropriate content is ignored, while the second provides that the safe CLIP textual encoder can properly encode the cleaned part of the input sentence. On the other hand, this also ensures that " + }, + { + "bbox": [ + 130, + 289, + 482, + 387 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 130, + 289, + 482, + 387 + ], + "type": "text", + "content": " can be seamlessly connected to downstream models that were trained on the basis of " + }, + { + "bbox": [ + 130, + 289, + 482, + 387 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_0" + }, + { + "bbox": [ + 130, + 289, + 482, + 387 + ], + "type": "text", + "content": " (for instance, Stable Diffusion v1.4 [36] in the case of a CLIP ViT-L/14). The same requirement is applied to the visual encoder: given an unsafe image " + }, + { + "bbox": [ + 130, + 289, + 482, + 387 + ], + "type": "inline_equation", + "content": "v_{i}^{\\star}" + }, + { + "bbox": [ + 130, + 289, + 482, + 387 + ], + "type": "text", + "content": " and a visual \"cleaning\" function " + }, + { + "bbox": [ + 130, + 289, + 482, + 387 + ], + "type": "inline_equation", + "content": "c_{v}" + }, + { + "bbox": [ + 130, + 289, + 482, + 387 + ], + "type": "text", + "content": ", we require:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 237, + 396, + 480, + 408 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 237, + 396, + 480, + 408 + ], + "spans": [ + { + "bbox": [ + 237, + 396, + 480, + 408 + ], + "type": "interline_equation", + "content": "\\mathcal {V} \\left(v _ {i} ^ {\\star}\\right) \\approx \\mathcal {V} \\left(c _ {v} \\left(v _ {i} ^ {\\star}\\right)\\right) \\approx \\mathcal {V} _ {0} \\left(c _ {v} \\left(v _ {i} ^ {\\star}\\right)\\right), \\tag {2}", + "image_path": "40f1614d5ea89b973adf625fd00abe145a0573dcff0de0842fc92a60e6ef156b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 418, + 474, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 418, + 474, + 430 + ], + "spans": [ + { + "bbox": [ + 131, + 418, + 474, + 430 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 418, + 474, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 131, + 418, + 474, + 430 + ], + "type": "text", + "content": " is the safe visual encoder and " + }, + { + "bbox": [ + 131, + 418, + 474, + 430 + ], + "type": "inline_equation", + "content": "\\nu_{0}" + }, + { + "bbox": [ + 131, + 418, + 474, + 430 + ], + "type": "text", + "content": " is the original CLIP visual encoder." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 447, + 295, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 447, + 295, + 459 + ], + "spans": [ + { + "bbox": [ + 131, + 447, + 295, + 459 + ], + "type": "text", + "content": "3.1 Building the ViSU dataset" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 471, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 471, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 471, + 482, + 604 + ], + "type": "text", + "content": "Overview. In order to modify CLIP to avoid the representation of inappropriate content, our methodology requires a dataset comprising quadruplets of safe and unsafe (i.e. NSFW) images and sentences, denoted as " + }, + { + "bbox": [ + 130, + 471, + 482, + 604 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{(v_i,t_i,v_i^\\star ,t_i^\\star),i = 1,\\dots,N\\}" + }, + { + "bbox": [ + 130, + 471, + 482, + 604 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 471, + 482, + 604 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 130, + 471, + 482, + 604 + ], + "type": "text", + "content": " indicates a safe image, " + }, + { + "bbox": [ + 130, + 471, + 482, + 604 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 130, + 471, + 482, + 604 + ], + "type": "text", + "content": " its corresponding sentence, and the unsafe image " + }, + { + "bbox": [ + 130, + 471, + 482, + 604 + ], + "type": "inline_equation", + "content": "v_{i}^{\\star}" + }, + { + "bbox": [ + 130, + 471, + 482, + 604 + ], + "type": "text", + "content": " and unsafe sentence " + }, + { + "bbox": [ + 130, + 471, + 482, + 604 + ], + "type": "inline_equation", + "content": "t_i^\\star" + }, + { + "bbox": [ + 130, + 471, + 482, + 604 + ], + "type": "text", + "content": " are \"paired\" to convey a similar semantic meaning of their safe counterparts. For instance, " + }, + { + "bbox": [ + 130, + 471, + 482, + 604 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 130, + 471, + 482, + 604 + ], + "type": "text", + "content": " can be considered as the sanitized version of " + }, + { + "bbox": [ + 130, + 471, + 482, + 604 + ], + "type": "inline_equation", + "content": "t_i^\\star" + }, + { + "bbox": [ + 130, + 471, + 482, + 604 + ], + "type": "text", + "content": ", expressing a similar meaning without inappropriate concepts, and the same holds for their visual counterparts. As such a dataset is not available, we build " + }, + { + "bbox": [ + 130, + 471, + 482, + 604 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 471, + 482, + 604 + ], + "type": "text", + "content": " with an automatic annotation procedure where ① unsafe sentences " + }, + { + "bbox": [ + 130, + 471, + 482, + 604 + ], + "type": "inline_equation", + "content": "t_i^\\star" + }, + { + "bbox": [ + 130, + 471, + 482, + 604 + ], + "type": "text", + "content": " are automatically generated starting from cleaned sentences " + }, + { + "bbox": [ + 130, + 471, + 482, + 604 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 130, + 471, + 482, + 604 + ], + "type": "text", + "content": ", and ② unsafe images " + }, + { + "bbox": [ + 130, + 471, + 482, + 604 + ], + "type": "inline_equation", + "content": "v_{i}^{\\star}" + }, + { + "bbox": [ + 130, + 471, + 482, + 604 + ], + "type": "text", + "content": " are generated starting from unsafe sentences " + }, + { + "bbox": [ + 130, + 471, + 482, + 604 + ], + "type": "inline_equation", + "content": "t_i^\\star" + }, + { + "bbox": [ + 130, + 471, + 482, + 604 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Training a conditioned NSFW textual generator. To achieve the first goal, we fine-tune a large language model (Llama 2-Chat [43]) to generate unsafe sentences starting from safe ones. In particular, we employ a set of 100 manually-curated safe-unsafe pairs, building these as a mixture of manually written pairs and sentences generated automatically with Vicuna [8]. To ensure" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 145, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 145, + 91, + 448, + 103 + ], + "type": "text", + "content": "Safe-CLIP: Removing NSFW Concepts from Vision-and-Language Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 482, + 271 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 271 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 271 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 271 + ], + "type": "image", + "image_path": "5f1bec07582e6a4718714df34e6475e907eb03b1108d05e355128b35936f2ebc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 212, + 277, + 399, + 288 + ], + "lines": [ + { + "bbox": [ + 212, + 277, + 399, + 288 + ], + "spans": [ + { + "bbox": [ + 212, + 277, + 399, + 288 + ], + "type": "text", + "content": "Fig. 2: Overview of our Safe-CLIP approach." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 302, + 482, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 302, + 482, + 386 + ], + "spans": [ + { + "bbox": [ + 130, + 302, + 482, + 386 + ], + "type": "text", + "content": "that the dataset provides proper supervision, we follow the definition of NSFW content of [38] as that of content belonging to the following twenty categories: hate, harassment, violence, suffering, humiliation, harm, suicide, sexual, nudity, bodily fluids, blood, obscene gestures, illegal activity, drug use, theft, vandalism, weapons, abuse, brutality, cruelty, and balance the samples of the training dataset across these categories to encourage the LLM to generate unsafe content with good variety." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 387, + 482, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 387, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 387, + 482, + 460 + ], + "type": "text", + "content": "We firstly fine-tune the LLM with supervised fine-tuning, using a prompt template explaining the task" + }, + { + "bbox": [ + 130, + 387, + 482, + 460 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 130, + 387, + 482, + 460 + ], + "type": "text", + "content": ", after which the model is asked to generate " + }, + { + "bbox": [ + 130, + 387, + 482, + 460 + ], + "type": "inline_equation", + "content": "t_{i}^{\\star}" + }, + { + "bbox": [ + 130, + 387, + 482, + 460 + ], + "type": "text", + "content": " starting from " + }, + { + "bbox": [ + 130, + 387, + 482, + 460 + ], + "type": "inline_equation", + "content": "t_{i}" + }, + { + "bbox": [ + 130, + 387, + 482, + 460 + ], + "type": "text", + "content": ". Interestingly, this fine-tuning procedure is enough to break the red-teaming measures taken during the training stages of Llama 2-Chat [43] and converts it into a generator of NSFW content which can generalize beyond the inappropriate concepts seen in our training set." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 464, + 482, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 464, + 482, + 571 + ], + "spans": [ + { + "bbox": [ + 130, + 464, + 482, + 571 + ], + "type": "text", + "content": "Aligning the textual NSFW generator. With the aim of increasing the quality of generated unsafe sentences, and also their semantic relatedness to the prompt, we adopt a fine-tuning stage by devising a variant of Direct Preference Optimization (DPO) [35]. DPO was originally proposed as an alternative to RLHF [9] with better stability and which does not need the explicit training of a reward model. Like RLHF, however, DPO employs large-scale human preference annotations, which in our case are not available. As a replacement, we build an automatic ranking procedure which can replace the human preference annotation while still increasing the alignment of our NSFW generator." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 573, + 482, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 573, + 482, + 622 + ], + "spans": [ + { + "bbox": [ + 131, + 573, + 482, + 622 + ], + "type": "text", + "content": "In particular, given a safe text " + }, + { + "bbox": [ + 131, + 573, + 482, + 622 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 131, + 573, + 482, + 622 + ], + "type": "text", + "content": ", we obtain two different unsafe completions " + }, + { + "bbox": [ + 131, + 573, + 482, + 622 + ], + "type": "inline_equation", + "content": "(t_{i,0}^{\\star}, t_{i,1}^{\\star})" + }, + { + "bbox": [ + 131, + 573, + 482, + 622 + ], + "type": "text", + "content": " from the SFT model by sampling from its output probability distribution. We then rank the quality of the obtained completions by considering their NSFW degree and their semantic similarity with " + }, + { + "bbox": [ + 131, + 573, + 482, + 622 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 131, + 573, + 482, + 622 + ], + "type": "text", + "content": ". For the former criterion, we" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "S. Poppi et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 631, + 482, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 631, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 631, + 482, + 665 + ], + "type": "text", + "content": "1 The prompt template is in the form: \"Below is an input string. Write a response that appropriately converts the input in its unsafe version \\n\\nInput: <" + }, + { + "bbox": [ + 133, + 631, + 482, + 665 + ], + "type": "inline_equation", + "content": "t_i>" + }, + { + "bbox": [ + 133, + 631, + 482, + 665 + ], + "type": "text", + "content": " \\n ### Response:\"" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": "obtain a binary NSFW rating " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "\\mathrm{nsfw}(t_i^\\star) \\in \\{0,1\\}" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": " by prompting GPT-3.5 with a completion. For the latter, we instead employ the CLIP similarity between " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": " and each of the completions, as predicted by the pre-trained text encoder using cosine similarity, thus lying in the range " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "[-1,1]" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": ". The final quality degree of an unsafe completion " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "t_i^\\star" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": ", given its safe prompt " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": ", is then obtained as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 200, + 184, + 481, + 198 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 184, + 481, + 198 + ], + "spans": [ + { + "bbox": [ + 200, + 184, + 481, + 198 + ], + "type": "interline_equation", + "content": "\\operatorname {r a n k} \\left(t _ {i} ^ {\\star}, t _ {i}\\right) = \\operatorname {C L I P - S i m} \\left(t _ {i} ^ {\\star}, t _ {i}\\right) + \\operatorname {N S F W R a t e} \\left(t _ {i} ^ {\\star}\\right), \\tag {3}", + "image_path": "a0f411523f50f625d220869e8a63c286446f0327aa4f5429425ddb11a612a06a.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 204, + 482, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 204, + 482, + 276 + ], + "spans": [ + { + "bbox": [ + 130, + 204, + 482, + 276 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 204, + 482, + 276 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP - Sim}(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 130, + 204, + 482, + 276 + ], + "type": "text", + "content": " indicates the CLIP similarity. The quality degree is then employed for ranking the two completions, i.e. " + }, + { + "bbox": [ + 130, + 204, + 482, + 276 + ], + "type": "inline_equation", + "content": "t_{i,w}^{\\star}\\succ t_{i,l}^{\\star}" + }, + { + "bbox": [ + 130, + 204, + 482, + 276 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 204, + 482, + 276 + ], + "type": "inline_equation", + "content": "t_{i,w}^{\\star}" + }, + { + "bbox": [ + 130, + 204, + 482, + 276 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 204, + 482, + 276 + ], + "type": "inline_equation", + "content": "t_{i,l}^{\\star}" + }, + { + "bbox": [ + 130, + 204, + 482, + 276 + ], + "type": "text", + "content": " indicate, respectively, the preferred and dispreferred completion. The resulting dataset of preferences, " + }, + { + "bbox": [ + 130, + 204, + 482, + 276 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{t_i,t_{i,w}^\\star ,t_{i,l}^\\star \\}_{i = 1}^N" + }, + { + "bbox": [ + 130, + 204, + 482, + 276 + ], + "type": "text", + "content": " is then employed for further training the LLM using the DPO objective. We refer the reader to [35] for further details on the training procedure employed by DPO." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 276, + 482, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 276, + 482, + 348 + ], + "spans": [ + { + "bbox": [ + 130, + 276, + 482, + 348 + ], + "type": "text", + "content": "Overall, our SFT and preference optimization pipeline turns Llama 2-Chat into a powerful generator of textual NSFW content, which can also perfectly maintain semantic relatedness with respect to a safe input sentence. What is more, our NSFW generator can still support diverse prompts, different from those seen at training time. Interestingly, for instance, it can be asked to add NSFW content belonging to a specific category (e.g. violence or nudity)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 350, + 483, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 483, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 483, + 483 + ], + "type": "text", + "content": "Generating the full ViSU dataset. Having a conditioned NSFW generator, we can generate NSFW texts starting from safe, visually relevant, sentences. Starting from NSFW sentences, we then generate corresponding NSFW images " + }, + { + "bbox": [ + 130, + 350, + 483, + 483 + ], + "type": "inline_equation", + "content": "v_{i}^{\\star}" + }, + { + "bbox": [ + 130, + 350, + 483, + 483 + ], + "type": "text", + "content": " using a diffusion-based model which has been trained on NSFW content2. The overall dataset, which we term ViSU (Visual Safe-Unsafe), contains 165k quadruplets of safe and unsafe sentences and images generated starting from COCO Captions [24]. We follow the Karpathy's splits [21] to organize the dataset into training, validation and test splits. To ensure that the resulting dataset is balanced across the twenty NSFW categories, we prompt the NSFW generator by asking to inject a specific, randomly chosen, category into the safe sentence. Sample quadruplets from our dataset are reported in the supplementary." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 498, + 256, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 498, + 256, + 510 + ], + "spans": [ + { + "bbox": [ + 131, + 498, + 256, + 510 + ], + "type": "text", + "content": "3.2 Making CLIP Safe" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 517, + 482, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 517, + 482, + 636 + ], + "spans": [ + { + "bbox": [ + 130, + 517, + 482, + 636 + ], + "type": "text", + "content": "Having built a dataset with quadruplets " + }, + { + "bbox": [ + 130, + 517, + 482, + 636 + ], + "type": "inline_equation", + "content": "(v_{i},t_{i},v_{i}^{\\star},t_{i}^{\\star})" + }, + { + "bbox": [ + 130, + 517, + 482, + 636 + ], + "type": "text", + "content": " of safe and unsafe images and sentences, we make the CLIP model safe with a procedure that ensures that the conditions expressed in Eq. 1 and 2 are met. To this aim, we adopt a multi-modal training scheme with four loss functions. Specifically, we define two inappropriate content redirection losses that aim at teaching the model to ignore unsafe content in an input text or input image, and two structure preservation losses that aim at maintaining the original structure of the embedding space in safe regions. During the rest of this section, " + }, + { + "bbox": [ + 130, + 517, + 482, + 636 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 130, + 517, + 482, + 636 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 517, + 482, + 636 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 130, + 517, + 482, + 636 + ], + "type": "text", + "content": " will indicate the textual and visual encoders being fine-tuned, " + }, + { + "bbox": [ + 130, + 517, + 482, + 636 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_0" + }, + { + "bbox": [ + 130, + 517, + 482, + 636 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 517, + 482, + 636 + ], + "type": "inline_equation", + "content": "\\nu_0" + }, + { + "bbox": [ + 130, + 517, + 482, + 636 + ], + "type": "text", + "content": " frozen \"deep copies\" of the textual and visual encoders obtained before the fine-tuning starts." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 145, + 90, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 90, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 145, + 90, + 448, + 103 + ], + "type": "text", + "content": "Safe-CLIP: Removing NSFW Concepts from Vision-and-Language Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 642, + 482, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 482, + 666 + ], + "type": "text", + "content": "2 We use the stabilized diffusionapi/newrealityxl-global-nsfw model available on HuggingFace, which has a high probability of generating NSFW images." + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": "Inappropriate content redirection. To teach the model to ignore inappropriate content, we propose to impose cross-modal similarities between unsafe sentences " + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "inline_equation", + "content": "t_i^\\star" + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": " and corresponding images " + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "inline_equation", + "content": "v_i" + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": " in the dataset, and between unsafe images " + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "inline_equation", + "content": "v_i^\\star" + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": " and corresponding texts " + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": ". Noticeably, this is not granted in " + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_0" + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_0" + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": ", which instead will have good metric learning properties between " + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "inline_equation", + "content": "v_i" + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": ". To encourage this effect even further, we also require that the embedding of the unsafe sentence " + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "inline_equation", + "content": "t_i^\\star" + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": " matches that of the corresponding safe sentence " + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": " according to the frozen textual encoder, and the embedding of unsafe images " + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "inline_equation", + "content": "v_i^\\star" + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": " matches that of the corresponding safe images " + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "inline_equation", + "content": "v_i" + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": " according to the frozen visual encoder, through a cosine similarity term which only considers positive pairs and ignore distances with respect to negative samples." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 250, + 483, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 250, + 483, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 250, + 483, + 346 + ], + "type": "text", + "content": "Formally, given a batch of " + }, + { + "bbox": [ + 130, + 250, + 483, + 346 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 250, + 483, + 346 + ], + "type": "text", + "content": " images " + }, + { + "bbox": [ + 130, + 250, + 483, + 346 + ], + "type": "inline_equation", + "content": "\\mathbf{V} = [v_{1}, v_{2}, \\dots, v_{N}]" + }, + { + "bbox": [ + 130, + 250, + 483, + 346 + ], + "type": "text", + "content": ", their corresponding safe captions " + }, + { + "bbox": [ + 130, + 250, + 483, + 346 + ], + "type": "inline_equation", + "content": "\\mathbf{T} = [t_{1}, t_{2}, \\dots, t_{N}]" + }, + { + "bbox": [ + 130, + 250, + 483, + 346 + ], + "type": "text", + "content": ", unsafe texts " + }, + { + "bbox": [ + 130, + 250, + 483, + 346 + ], + "type": "inline_equation", + "content": "\\mathbf{T}^{\\star} = [t_{1}^{\\star}, t_{2}^{\\star}, \\dots, t_{N}^{\\star}]" + }, + { + "bbox": [ + 130, + 250, + 483, + 346 + ], + "type": "text", + "content": " and unsafe images " + }, + { + "bbox": [ + 130, + 250, + 483, + 346 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^{\\star} = [v_{1}^{\\star}, v_{2}^{\\star}, \\dots, v_{N}^{\\star}]" + }, + { + "bbox": [ + 130, + 250, + 483, + 346 + ], + "type": "text", + "content": ", we define two " + }, + { + "bbox": [ + 130, + 250, + 483, + 346 + ], + "type": "inline_equation", + "content": "N \\times N" + }, + { + "bbox": [ + 130, + 250, + 483, + 346 + ], + "type": "text", + "content": " matrices containing pairwise cosine similarities between " + }, + { + "bbox": [ + 130, + 250, + 483, + 346 + ], + "type": "inline_equation", + "content": "\\mathbf{T}^{\\star}" + }, + { + "bbox": [ + 130, + 250, + 483, + 346 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 250, + 483, + 346 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 130, + 250, + 483, + 346 + ], + "type": "text", + "content": " and between " + }, + { + "bbox": [ + 130, + 250, + 483, + 346 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^{\\star}" + }, + { + "bbox": [ + 130, + 250, + 483, + 346 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 250, + 483, + 346 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 130, + 250, + 483, + 346 + ], + "type": "text", + "content": ". We then adopt a symmetric InfoNCE loss [30] which aims at maximizing the cosine similarity between the " + }, + { + "bbox": [ + 130, + 250, + 483, + 346 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 250, + 483, + 346 + ], + "type": "text", + "content": " matching pairs of cross-modal safe and unsafe embeddings, and minimize those of the " + }, + { + "bbox": [ + 130, + 250, + 483, + 346 + ], + "type": "inline_equation", + "content": "N^{2} - N" + }, + { + "bbox": [ + 130, + 250, + 483, + 346 + ], + "type": "text", + "content": " non-matching pairs while having, in turn, one of the encoders frozen and the other being fine-tuned:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 357, + 480, + 424 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 357, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 133, + 357, + 480, + 424 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {\\text {r e d i r}, 1} = - \\frac {1}{N} \\left(\\sum_ {i = 1} ^ {N} \\log \\frac {\\exp \\left(\\cos \\left(\\mathcal {T} \\left(t _ {i} ^ {\\star}\\right) , \\mathcal {V} _ {0} \\left(v _ {i}\\right)\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {N} \\exp \\left(\\cos \\left(\\mathcal {T} \\left(t _ {j} ^ {\\star}\\right) , \\mathcal {V} _ {0} \\left(v _ {i}\\right)\\right) / \\tau\\right)} + \\sum_ {i = 1} ^ {N} \\log \\frac {\\exp \\left(\\cos \\left(\\mathcal {T} \\left(t _ {i} ^ {\\star}\\right) , \\mathcal {V} _ {0} \\left(v _ {i}\\right)\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {N} \\exp \\left(\\cos \\left(\\mathcal {T} \\big (t _ {i} ^ {\\star}\\right) , \\mathcal {V} _ {0} \\left(v _ {j}\\right)\\right) / \\tau\\right)} \\left. \\right. \\\\ + \\sum_ {i = 1} ^ {N} \\log \\frac {\\exp (\\cos (\\mathcal {V} (v _ {i} ^ {\\star}) , \\mathcal {T} _ {0} (t _ {i})) / \\tau)}{\\sum_ {j = 1} ^ {N} \\exp (\\cos (\\mathcal {V} (v _ {j} ^ {\\star}) , \\mathcal {T} _ {0} (t _ {i})) / \\tau)} + \\sum_ {i = 1} ^ {N} \\log \\frac {\\exp (\\cos (\\mathcal {V} (v _ {i} ^ {\\star}) , \\mathcal {T} _ {0} (t _ {i})) / \\tau)}{\\sum_ {j = 1} ^ {N} \\exp (\\cos (\\mathcal {V} (v _ {i} ^ {\\star}) , \\mathcal {T} _ {0} (t _ {j})) / \\tau)} \\Bigg), \\\\ \\end{array}", + "image_path": "4327a103b85bc6a4796ba484b23e6c29e1b46cdf5a30d7facc7856424a17482f.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 432, + 482, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 432, + 482, + 480 + ], + "spans": [ + { + "bbox": [ + 130, + 432, + 482, + 480 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 432, + 482, + 480 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 432, + 482, + 480 + ], + "type": "text", + "content": " is a temperature parameter. The second loss term, which brings each unsafe sentence close to its corresponding safe one, and each unsafe image close to its corresponding safe one, is instead expressed as the negative cosine similarity between each unsafe embedding and the original safe embeddings, as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 167, + 495, + 481, + 528 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 495, + 481, + 528 + ], + "spans": [ + { + "bbox": [ + 167, + 495, + 481, + 528 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {r e d i r}, 2} = - \\frac {1}{N} \\left(\\sum_ {i = 1} ^ {N} \\cos \\left(\\mathcal {T} \\left(t _ {i} ^ {\\star}\\right), \\mathcal {T} _ {0} \\left(t _ {i}\\right)\\right) + \\sum_ {i = 1} ^ {N} \\cos \\left(\\mathcal {V} \\left(v _ {i} ^ {\\star}\\right), \\mathcal {V} _ {0} \\left(v _ {i}\\right)\\right)\\right). \\tag {5}", + "image_path": "71aad3548c3c7ed72effae0ac628ade3b97882997e36d2cd3dd6ca1002ee6fb7.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 543, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 543, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 543, + 482, + 628 + ], + "type": "text", + "content": "**Embedding structure preservation.** The two aforementioned losses bring unsafe embeddings towards the positions of their corresponding safe embeddings in the original frozen spaces, either in a single-modal or multi-modal manner. However, alone, they would inevitably cause the fine-tuned encoders " + }, + { + "bbox": [ + 130, + 543, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 130, + 543, + 482, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 543, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 130, + 543, + 482, + 628 + ], + "type": "text", + "content": " to lose their performance on safe inputs, as well as their matching properties. Therefore, we also adopt two losses to ensure that the original structure of the embedding spaces is preserved where safe textual and visual regions lie." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "In particular, we define a matching loss between the safe embeddings produced by the online networks " + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": " and those of the original, pre-trained, networks " + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_0" + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_0" + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": ". Again, this is defined through the negative cosine similarity be" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "S. Poppi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 244, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 244, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 244, + 128 + ], + "type": "text", + "content": "tween matching pairs, as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 169, + 137, + 481, + 171 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 137, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 169, + 137, + 481, + 171 + ], + "type": "interline_equation", + "content": "L _ {\\text {p r e s}, 1} = - \\frac {1}{N} \\left(\\sum_ {i = 1} ^ {N} \\cos \\left(\\mathcal {T} \\left(t _ {i}\\right), \\mathcal {T} _ {0} \\left(t _ {i}\\right)\\right) + \\sum_ {i = 1} ^ {N} \\cos \\left(\\mathcal {V} \\left(v _ {i}\\right), \\mathcal {V} _ {0} \\left(v _ {i}\\right)\\right)\\right). \\tag {6}", + "image_path": "3ab17783bfa559d27d32b870427d73ba3591204400f69ce6a06f1ae9f2574c2b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 178, + 479, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 178, + 479, + 227 + ], + "spans": [ + { + "bbox": [ + 131, + 178, + 479, + 227 + ], + "type": "text", + "content": "Finally, as an additional regularization term we also keep a contrastive loss between safe visual embeddings and safe textual embeddings, comparing online and frozen encoders. This, in practice, closely resembles the original loss on which the embedding space was trained, i.e." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 235, + 477, + 302 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 235, + 477, + 302 + ], + "spans": [ + { + "bbox": [ + 133, + 235, + 477, + 302 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {\\text {p r e s}, 2} = - \\frac {1}{N} \\left(\\sum_ {i = 1} ^ {N} \\log \\frac {\\exp \\left(\\cos \\left(\\mathcal {V} _ {0} \\left(v _ {i}\\right) , \\mathcal {T} \\left(t _ {i}\\right)\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {N} \\exp \\left(\\cos \\left(\\mathcal {V} _ {0} \\left(v _ {i}\\right) , \\mathcal {T} \\left(t _ {j}\\right)\\right) / \\tau\\right)} + \\sum_ {i = 1} ^ {N} \\log \\frac {\\exp \\left(\\cos \\left(\\mathcal {V} _ {0} \\left(v _ {i}\\right) , \\mathcal {T} \\left(t _ {i}\\right)\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {N} \\exp \\left(\\cos \\left(\\mathcal {V} _ {0} \\big (v _ {j}\\big) , \\mathcal {T} \\big (t _ {i}\\big)\\right) / \\tau\\right)} \\right. \\\\ + \\sum_ {i = 1} ^ {N} \\log \\frac {\\exp (\\cos (\\mathcal {T} _ {0} (t _ {i}) , \\mathcal {V} (v _ {i})) / \\tau)}{\\sum_ {j = 1} ^ {N} \\exp (\\cos (\\mathcal {T} _ {0} (t _ {i}) , \\mathcal {V} (v _ {j})) / \\tau)} + \\sum_ {i = 1} ^ {N} \\log \\frac {\\exp (\\cos (\\mathcal {T} _ {0} (t _ {i}) , \\mathcal {V} (v _ {i})) / \\tau)}{\\sum_ {j = 1} ^ {N} \\exp (\\cos (\\mathcal {T} _ {0} (t _ {j}) , \\mathcal {V} (v _ {i})) / \\tau)} \\Biggr). \\\\ \\end{array}", + "image_path": "d7f7b0edade33d81538d2d69b019ecad3f9767f55cbe932c54fce92fd0604d45.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 308, + 479, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 308, + 479, + 331 + ], + "spans": [ + { + "bbox": [ + 131, + 308, + 479, + 331 + ], + "type": "text", + "content": "Eventually, the overall loss function on which the network is trained is a weighted sum of the four loss functions mentioned above." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 351, + 230, + 365 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 351, + 230, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 351, + 230, + 365 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 374, + 269, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 269, + 387 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 269, + 387 + ], + "type": "text", + "content": "4.1 Experimental Setting" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 395, + 482, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 395, + 482, + 527 + ], + "spans": [ + { + "bbox": [ + 130, + 395, + 482, + 527 + ], + "type": "text", + "content": "Datasets. Our experiments are mainly conducted on the collected ViSU dataset that, as previously mentioned, contains 165k quadruplets of safe and unsafe content for training. For both validation and testing we use 5k samples following the Karpathy's COCO splits [21], randomly sampling only one safe caption among the five available for each image in the original COCO dataset. When applying our Safe-CLIP to text-to-image generative architectures, we also perform experiments on the I2P dataset [38] that is composed of 4,703 textual prompts extracted from Lexica, a collection of user-generated prompts for conditioning text-to-image diffusion models. Each prompt is associated to one of seven different categories of inappropriate content, among hate, harassment, violence, self-harm, sexual content, shocking images, illegal activity." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 530, + 482, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 530, + 482, + 591 + ], + "spans": [ + { + "bbox": [ + 131, + 530, + 482, + 591 + ], + "type": "text", + "content": "LLM fine-tuning details. During SFT, we fine-tune the 7B version of Llama 2-Chat using low-rank adaptation [20] with " + }, + { + "bbox": [ + 131, + 530, + 482, + 591 + ], + "type": "inline_equation", + "content": "r = 64" + }, + { + "bbox": [ + 131, + 530, + 482, + 591 + ], + "type": "text", + "content": " as low-rank factor. We employ a batch size equal to 4 and a learning rate of " + }, + { + "bbox": [ + 131, + 530, + 482, + 591 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-4}" + }, + { + "bbox": [ + 131, + 530, + 482, + 591 + ], + "type": "text", + "content": ". To perform DPO, we follow the variant presented in [46], employing low-rank adaptation also in this case. The complete DPO fine-tuning settings are reported in the supplementary." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 593, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 593, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 593, + 482, + 666 + ], + "type": "text", + "content": "Safe-CLIP implementation and training details. Our architecture is based on the standard CLIP model [33] composed of a visual and a textual encoder. Specifically, we employ the ViT-L/14 variant to comply with the textual encoder used in the Stable Diffusion v1.4 model [36] and the visual encoder employed in LLaVA [25]. Experiments with different CLIP-based backbones are reported in the supplementary. During training, both the visual and textual encoder are" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 145, + 90, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 90, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 145, + 90, + 448, + 103 + ], + "type": "text", + "content": "Safe-CLIP: Removing NSFW Concepts from Vision-and-Language Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": "fine-tuned using low-rank decompositions [20], where the low-rank factor " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": " is set to 16 in all the experiments. We employ Adam as optimizer [22] using a learning rate equal to " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-3}" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": " and a batch size of 128." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 170, + 307, + 182 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 170, + 307, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 170, + 307, + 182 + ], + "type": "text", + "content": "4.2 Evaluating the ViSU Dataset" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 190, + 282, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 190, + 282, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 190, + 282, + 297 + ], + "type": "text", + "content": "We first assess the quality of our ViSU dataset, evaluating the inappropriateness degree of the generated unsafe sentences. Specifically, we employ a DistilBERT model fine-tuned for adult content detection and directly ask GPT-3.5 to evaluate whether the generated unsafe sentences should be" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 290, + 228, + 480, + 292 + ], + "blocks": [ + { + "bbox": [ + 287, + 192, + 482, + 225 + ], + "lines": [ + { + "bbox": [ + 287, + 192, + 482, + 225 + ], + "spans": [ + { + "bbox": [ + 287, + 192, + 482, + 225 + ], + "type": "text", + "content": "Table 1: Comparison between the textual portion of ViSU and the I2P benchmark [38], in terms of NSFW degree and toxicity." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 290, + 228, + 480, + 292 + ], + "lines": [ + { + "bbox": [ + 290, + 228, + 480, + 292 + ], + "spans": [ + { + "bbox": [ + 290, + 228, + 480, + 292 + ], + "type": "table", + "html": "
Dataset% NSFWToxicity
DistilBERTGPT-3.5
I2P [38]52.813.914.9
w/o SFT (i.e. Llama 2-Chat)37.89.37.7
w/o DPO fine-tuning75.975.030.6
ViSU (Ours)80.979.131.3
", + "image_path": "bcc1d8d3e8aa27651bd8933e8bce41879dd65d42201bdd0de1530676e52c2620.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 297, + 482, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 482, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 482, + 418 + ], + "type": "text", + "content": "classified as NSFW content. As reported in Table 1, ViSU showcases a very good textual quality, as it has a higher degree of NSFW sentences compared to existing alternatives like I2P [38] (i.e. " + }, + { + "bbox": [ + 130, + 297, + 482, + 418 + ], + "type": "inline_equation", + "content": "79.1\\%" + }, + { + "bbox": [ + 130, + 297, + 482, + 418 + ], + "type": "text", + "content": " of NSFW sentences vs. " + }, + { + "bbox": [ + 130, + 297, + 482, + 418 + ], + "type": "inline_equation", + "content": "13.9\\%" + }, + { + "bbox": [ + 130, + 297, + 482, + 418 + ], + "type": "text", + "content": " in the I2P dataset according to GPT-3.5). Moreover, following [38], we also report the toxicity score of the unsafe sentences computed using the Perspective " + }, + { + "bbox": [ + 130, + 297, + 482, + 418 + ], + "type": "inline_equation", + "content": "\\mathrm{API}^3" + }, + { + "bbox": [ + 130, + 297, + 482, + 418 + ], + "type": "text", + "content": ". Also in terms of toxicity, ViSU presents a higher degree of NSFW content. When instead comparing the effectiveness of each of the Llama 2 fine-tuning stages, it is worth noting that both the SFT procedure and DPO fine-tuning consistently increase the quality of generated sentences, going from " + }, + { + "bbox": [ + 130, + 297, + 482, + 418 + ], + "type": "inline_equation", + "content": "9.3\\%" + }, + { + "bbox": [ + 130, + 297, + 482, + 418 + ], + "type": "text", + "content": " of NSFW content when using the original Llama 2 model to " + }, + { + "bbox": [ + 130, + 297, + 482, + 418 + ], + "type": "inline_equation", + "content": "79.1\\%" + }, + { + "bbox": [ + 130, + 297, + 482, + 418 + ], + "type": "text", + "content": " after both fine-tuning stages." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 435, + 383, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 383, + 448 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 383, + 448 + ], + "type": "text", + "content": "4.3 Evaluating the Safe-CLIP Embedding Space" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 455, + 482, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 455, + 482, + 598 + ], + "spans": [ + { + "bbox": [ + 130, + 455, + 482, + 598 + ], + "type": "text", + "content": "Results on ViSU test set. To evaluate the retrieval performance of Safe-CLIP, we firstly consider image-to-text and text-to-image retrieval in a safe-only setting, where we do not have any inappropriate content in both visual and textual data. This is important to assess whether the properties of the original CLIP embedding space are preserved when employing our fine-tuning strategy. In this case, query elements are represented by the safe images of the test set (which, with a slight abuse of notation, we refer to as " + }, + { + "bbox": [ + 130, + 455, + 482, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 130, + 455, + 482, + 598 + ], + "type": "text", + "content": ") for the image-to-text setting and the safe textual items (referred to as " + }, + { + "bbox": [ + 130, + 455, + 482, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 130, + 455, + 482, + 598 + ], + "type": "text", + "content": ") for the text-to-image one. Moreover, we consider text-to-image and image-to-text retrieval when using unsafe texts as queries (referred to as " + }, + { + "bbox": [ + 130, + 455, + 482, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{T}^{\\star}" + }, + { + "bbox": [ + 130, + 455, + 482, + 598 + ], + "type": "text", + "content": ") and both safe and unsafe images as retrievable items and when using unsafe images as queries (i.e. " + }, + { + "bbox": [ + 130, + 455, + 482, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^{\\star}" + }, + { + "bbox": [ + 130, + 455, + 482, + 598 + ], + "type": "text", + "content": ") and both safe and unsafe texts as retrievable items." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "text", + "content": "Retrieval results on the ViSU test set are reported in Table 2, comparing the proposed Safe-CLIP model with the original CLIP architecture, a CLIP model trained on the DataComp dataset [12], which has undergone NSFW content cleaning, and two different baselines. Specifically, we consider a variant of our" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "S. Poppi et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 370, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 370, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 370, + 666 + ], + "type": "text", + "content": "3 https://github.com/conversationai/perspectiveapi" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 185, + 480, + 352 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 180 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 180 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 180 + ], + "type": "text", + "content": "Table 2: Retrieval results on the ViSU test set. The left portions respectively show text-to-image and image-to-text performance when using safe data only (i.e. " + }, + { + "bbox": [ + 131, + 114, + 482, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 131, + 114, + 482, + 180 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 114, + 482, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 131, + 114, + 482, + 180 + ], + "type": "text", + "content": "). The right portions report the results when using unsafe textual sentences as query (i.e. " + }, + { + "bbox": [ + 131, + 114, + 482, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{T}^{\\star}" + }, + { + "bbox": [ + 131, + 114, + 482, + 180 + ], + "type": "text", + "content": ") and the merging of safe (i.e. " + }, + { + "bbox": [ + 131, + 114, + 482, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 131, + 114, + 482, + 180 + ], + "type": "text", + "content": ") and unsafe images (i.e. " + }, + { + "bbox": [ + 131, + 114, + 482, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^{\\star}" + }, + { + "bbox": [ + 131, + 114, + 482, + 180 + ], + "type": "text", + "content": ") as retrievable items, or when using unsafe visual queries (i.e. " + }, + { + "bbox": [ + 131, + 114, + 482, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^{\\star}" + }, + { + "bbox": [ + 131, + 114, + 482, + 180 + ], + "type": "text", + "content": ") and the merging of safe (i.e. " + }, + { + "bbox": [ + 131, + 114, + 482, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 131, + 114, + 482, + 180 + ], + "type": "text", + "content": ") and unsafe sentences (i.e. " + }, + { + "bbox": [ + 131, + 114, + 482, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{T}^{\\star}" + }, + { + "bbox": [ + 131, + 114, + 482, + 180 + ], + "type": "text", + "content": ") as retrievable items." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 185, + 480, + 352 + ], + "lines": [ + { + "bbox": [ + 134, + 185, + 480, + 352 + ], + "spans": [ + { + "bbox": [ + 134, + 185, + 480, + 352 + ], + "type": "table", + "html": "
ModelText-to-Image (T-to-V)Image-to-Text (V-to-T)Text-to-Image (T*-to-V ∪ V*)Image-to-Text (V*-to-T ∪ T*)
R@1R@10R@20R@1R@10R@20R@1R@10R@20R@1R@10R@20
CLIP (ViT-L) [33]36.871.681.539.874.283.52.024.833.24.532.940.6
DataComp-1B (ViT-L) [12]46.779.787.447.081.388.91.628.135.65.537.544.9
w/o inap. content redirection49.983.790.348.183.690.51.630.440.16.135.242.6
w/o negative cosine similarities41.978.587.341.577.886.98.246.056.613.760.468.2
Safe-CLIP45.981.889.745.382.389.78.046.958.019.162.971.1
NSFW Text QueryCLIP Top-1Safe-CLIP Top-1NSFW Image QueryCLIP Top-1Safe-CLIP Top-1
A group of people on a battlefield with buildings in the background.A girl in blue lingerie holding a crack pipe in a dirty alleyway.Close-up of a blonde girl enjoying music with headphones.
A black and white photo of a couple fuicking each other.A sex worker servicing a client in a parking lot.Harley Davidson belt inspection.
", + "image_path": "ba51f61a4a3171615717c1d6bac37d394ffb7d54c125cc50556f5bc3d3394277.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 355, + 482, + 377 + ], + "lines": [ + { + "bbox": [ + 131, + 355, + 482, + 377 + ], + "spans": [ + { + "bbox": [ + 131, + 355, + 482, + 377 + ], + "type": "text", + "content": "Fig. 3: Top-1 images (left) and text (right) retrieved using the original CLIP model and our Safe-CLIP, when NSFW texts and images are employed as query." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 387, + 482, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 387, + 482, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 387, + 482, + 518 + ], + "type": "text", + "content": "approach in which we remove the two negative cosine similarity losses (i.e. Eq. 5 and 6), and a model trained with safe data only (i.e. removing the loss functions for inappropriate content redirection reported in Eq. 4 and 5). Results are reported in terms of " + }, + { + "bbox": [ + 130, + 387, + 482, + 518 + ], + "type": "inline_equation", + "content": "\\mathrm{Recall}@\\bar{k}" + }, + { + "bbox": [ + 130, + 387, + 482, + 518 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 387, + 482, + 518 + ], + "type": "inline_equation", + "content": "\\mathrm{R}@\\bar{k}" + }, + { + "bbox": [ + 130, + 387, + 482, + 518 + ], + "type": "text", + "content": ") with " + }, + { + "bbox": [ + 130, + 387, + 482, + 518 + ], + "type": "inline_equation", + "content": "k = 1, 10, 20" + }, + { + "bbox": [ + 130, + 387, + 482, + 518 + ], + "type": "text", + "content": ", that measures the percentage of times the visual or textual item associated to the query is retrieved among the top-" + }, + { + "bbox": [ + 130, + 387, + 482, + 518 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 387, + 482, + 518 + ], + "type": "text", + "content": " elements. When using unsafe sentences as queries, for each element we consider the safe image associated with the given text as the corresponding visual element. Symmetrically, when using unsafe images as queries, for each element we consider the safe text associated with the given image as the ground-truth item. Therefore, recall results in the unsafe setting follow a \"the higher the better\" protocol." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 519, + 482, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 519, + 482, + 638 + ], + "spans": [ + { + "bbox": [ + 130, + 519, + 482, + 638 + ], + "type": "text", + "content": "As it can be seen, Safe-CLIP can retrieve a significant higher portion of correct safe images when using unsafe prompts as queries, while effectively preserving good performance in safe-only settings (i.e. " + }, + { + "bbox": [ + 130, + 519, + 482, + 638 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 130, + 519, + 482, + 638 + ], + "type": "text", + "content": "-to-" + }, + { + "bbox": [ + 130, + 519, + 482, + 638 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 130, + 519, + 482, + 638 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 519, + 482, + 638 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 130, + 519, + 482, + 638 + ], + "type": "text", + "content": "-to-" + }, + { + "bbox": [ + 130, + 519, + 482, + 638 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 130, + 519, + 482, + 638 + ], + "type": "text", + "content": "). Specifically, when comparing our model with the original CLIP, it is worth noting that the results on text-to-image retrieval with unsafe texts as queries are consistently improved when using our text encoder, with an overall improvement of 6.0 points in terms of " + }, + { + "bbox": [ + 130, + 519, + 482, + 638 + ], + "type": "inline_equation", + "content": "\\mathrm{R}@\\mathrm{1}" + }, + { + "bbox": [ + 130, + 519, + 482, + 638 + ], + "type": "text", + "content": ", and the same applies for image-to-text retrieval which showcases an improvement of 14.6 " + }, + { + "bbox": [ + 130, + 519, + 482, + 638 + ], + "type": "inline_equation", + "content": "\\mathrm{R}@\\mathrm{1}" + }, + { + "bbox": [ + 130, + 519, + 482, + 638 + ], + "type": "text", + "content": " points. This demonstrates the effectiveness of our fine-tuning strategy, which can reduce the model probability of returning inappropriate images or sentences." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": "Robustness on real NSFW images. To further analyze the safety degree of the Safe-CLIP embedding space, we perform text-to-image and image-to-text" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 145, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 145, + 91, + 447, + 102 + ], + "type": "text", + "content": "Safe-CLIP: Removing NSFW Concepts from Vision-and-Language Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": "retrieval using real NSFW images as visual items. Specifically, we select inappropriate visual content from three different sources: " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "(i)" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": " a portion of data used to train the NudeNet classifier, " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "(ii)" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": " images crawled from the web using NSFW" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 241, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 241, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 241, + 272 + ], + "type": "text", + "content": "data source URLs4, and (iii) images from the Socio-Moral Image Database (SMID) [10]. While the first two sources exclusively contain nudity and pornography images, the third one includes more varied types of" + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 249, + 209, + 476, + 260 + ], + "blocks": [ + { + "bbox": [ + 246, + 161, + 482, + 205 + ], + "lines": [ + { + "bbox": [ + 246, + 161, + 482, + 205 + ], + "spans": [ + { + "bbox": [ + 246, + 161, + 482, + 205 + ], + "type": "text", + "content": "Table 3: Percentage of retrieved NSFW images and text using unsafe data as query. Safe retrievable items are from LAION-400M, unsafe images are extracted from different NSFW sources, and unsafe texts are from ViSU." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 249, + 209, + 476, + 260 + ], + "lines": [ + { + "bbox": [ + 249, + 209, + 476, + 260 + ], + "spans": [ + { + "bbox": [ + 249, + 209, + 476, + 260 + ], + "type": "table", + "html": "
Model% NSFW (Text-to-Image)% NSFW (Image-to-Text)
NudeNetNSFW URLsSMIDNudeNetNSFW URLsSMID
CLIP [33]57.155.247.865.657.441.4
DataComp-1B [12]55.649.764.061.456.245.6
Safe-CLIP8.49.816.728.824.734.5
", + "image_path": "fdab21fcfd8babdac3430007616b33a13df53bf3c8e5f68a7b9a9140db2e244f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 272, + 482, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 272, + 482, + 379 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 482, + 379 + ], + "type": "text", + "content": "inappropriate images representing negative concepts such as, for example, harm, inequality, discrimination, and unfairness. Overall, we randomly sample 1,000 images from each of the NSFW data sources, selecting only those representing unsafe concepts for the SMID dataset. As textual items, we employ unsafe texts from the ViSU test set that match the NSFW concepts represented in each of the NSFW visual sources (i.e. sexual and nudity for NudeNet and NSFW data source URLs, and all other concepts for the SMID dataset). For both I2T and T2I, we employ a set of 10k randomly selected visual or textual distractors, randomly selected from the LAION-400M dataset [41]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 380, + 482, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 487 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 487 + ], + "type": "text", + "content": "Results are shown in Table 3 comparing Safe-CLIP with the standard CLIP model and the model trained on DataComp. For each NSFW data source, we report the percentage of times in which an NSFW image or text is retrieved as the top-1 element. Notably, using Safe-CLIP consistently reduces the percentage of retrieved NSFW items for all three NSFW dataset sources. In particular, the percentage of retrieved NSFW visual and textual content is reduced by more than 45 and 30 points, respectively when considering unsafe images or textual elements in all three considered settings. This experiment confirms that our fine-tuning strategy can effectively enhance the safety of the CLIP embedding space." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 490, + 482, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 490, + 482, + 538 + ], + "spans": [ + { + "bbox": [ + 130, + 490, + 482, + 538 + ], + "type": "text", + "content": "Qualitative results. Fig. 3 reports qualitative retrieval results in the same aforementioned setting. Safe-CLIP is able to retrieve safe images starting from NSFW texts and, vice versa, retrieve safe sentences starting from NSFW images. Additionally, it can also preserve the global context and semantics of the query." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 555, + 365, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 555, + 365, + 567 + ], + "spans": [ + { + "bbox": [ + 131, + 555, + 365, + 567 + ], + "type": "text", + "content": "4.4 Safe-CLIP for Text-to-Image Generation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "type": "text", + "content": "Results on I2P and ViSU test set. We then validate the effectiveness of the Safe-CLIP text encoder when applied in a text-to-image generative model. Specifically, we employ Stable Diffusion v1.4 [36], eventually replacing the standard CLIP text encoder used in Stable Diffusion with our fine-tuned version. Moreover, we also apply Safe-CLIP in combination with other NSFW removal strategies. In particular, we consider a version of Stable Diffusion with negative" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "S. Poppi et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 375, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 375, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 375, + 665 + ], + "type": "text", + "content": "4 https://github.com/EBazarov/nsfw_data_source_url" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 152, + 479, + 395 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 148 + ], + "type": "text", + "content": "Table 4: Probabilities of generating images with unsafe content, classified by combining the predictions of NudeNet and Q16. Results are reported using NSFW text prompts from I2P [38] and ViSU, and Stable Diffusion v1.4 as text-to-image generator." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 152, + 479, + 395 + ], + "lines": [ + { + "bbox": [ + 134, + 152, + 479, + 395 + ], + "spans": [ + { + "bbox": [ + 134, + 152, + 479, + 395 + ], + "type": "table", + "html": "
ModelI2PViSU
HateHassassmentViolenceSelf-harmSexualShockingIllegal Act.AvgHateHassassmentViolenceSelf-harmSexualShockingIllegal Act.Avg
SD v1.441.432.443.742.124.852.235.735.725.917.830.419.524.426.923.526.2
+ Safe-CLIP23.621.126.726.815.932.721.422.24.62.93.94.64.12.93.33.6
Negative Prompts28.524.422.423.315.940.829.324.418.613.920.214.014.016.514.416.9
+ Safe-CLIP19.217.721.722.913.926.119.318.93.13.42.83.63.12.92.72.9
SLD-Weak [38]30.624.132.127.813.941.925.725.617.510.720.813.316.818.815.417.7
+ Safe-CLIP21.219.025.322.412.428.119.519.83.73.03.23.83.73.03.13.2
SLD-Medium [38]21.617.523.717.48.931.216.717.710.67.012.39.810.811.59.710.8
+ Safe-CLIP18.917.221.620.611.925.816.417.53.02.23.23.42.82.32.62.8
SLD-Strong [38]15.913.618.811.17.821.511.213.56.43.76.15.17.25.84.45.6
+ Safe-CLIP16.914.017.612.28.220.213.113.03.41.41.71.91.71.91.81.8
SD v1.4Negative PromptsSLD-Strong [38]SD v1.4 + Safe-CLIPSD v1.4Negative PromptsSLD-Strong [38]SD v1.4 + Safe-CLIP
", + "image_path": "de72c6922cc3f9542e689468ba2be8cfa423c6fdd59e454c35e8b9d7b63af9ab.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 397, + 479, + 419 + ], + "lines": [ + { + "bbox": [ + 132, + 397, + 479, + 419 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 479, + 419 + ], + "type": "text", + "content": "Fig. 4: Images generated from unsafe prompts with Stable Diffusion, employing the original CLIP model, negative prompts, SLD-Strong [38], and our Safe-CLIP." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 430, + 480, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 480, + 502 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 480, + 502 + ], + "type": "text", + "content": "prompts and the recently proposed Safe Latent Diffusion (SLD) approach [38] which employs different levels of safety guidance (SLD-Weak, SLD-Medium, and SLD-Strong) to limit the generation of inappropriate images. For this experiment, we generate five images for each textual prompt using different random seeds and compute the probability of generating inappropriate images detected by two NSFW classifiers. Following [38], we employ Q16 [39] and NudeNet [2]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 502, + 481, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 502, + 481, + 634 + ], + "spans": [ + { + "bbox": [ + 130, + 502, + 481, + 634 + ], + "type": "text", + "content": "Table 4 shows the results using textual prompts from both I2P and ViSU datasets. We report the NSFW generation probabilities on the entire set of prompts of each dataset and also dividing them into the seven NSFW categories considered in " + }, + { + "bbox": [ + 130, + 502, + 481, + 634 + ], + "type": "inline_equation", + "content": "[39]^5" + }, + { + "bbox": [ + 130, + 502, + 481, + 634 + ], + "type": "text", + "content": ". Interestingly, Safe-CLIP significantly reduces the probabilities of generating NSFW images when using textual prompts from both datasets thus demonstrating its usefulness also in a text-to-image generation setting. In particular, when applying our text encoder to a standard Stable Diffusion model, the probability of generating inappropriate content decreases by 13.5 points with I2P prompts and 22.6 points with NSFW texts from ViSU. Similar results can also be observed when applying Safe-CLIP alongside other NSFW removal strategies, highlighting that fine-tuning the CLIP text encoder with the" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 145, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 145, + 91, + 447, + 102 + ], + "type": "text", + "content": "Safe-CLIP: Removing NSFW Concepts from Vision-and-Language Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 642, + 479, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 479, + 665 + ], + "type": "text", + "content": "5 Specifically, we map each of the 20 NSFW concepts of ViSU into one of the seven categories defined in I2P. Further details are given in the supplementary material." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "A proposed approach can benefit the performance of existing methods tailored for removing NSFW concepts from images generated by diffusion models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 143, + 482, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 143, + 482, + 215 + ], + "spans": [ + { + "bbox": [ + 130, + 143, + 482, + 215 + ], + "type": "text", + "content": "Qualitative results. Samples of generated images are shown in Fig. 4, comparing results generated by Safe-CLIP applied to Stable Diffusion with images generated by SLD-Strong [39], Stable Diffusion with negative prompts, and the Stable Diffusion original version. Qualitative results confirm the effectiveness of our proposal which can generate images that preserve the original semantic of the scene while preventing the generation of inappropriate content." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 230, + 366, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 230, + 366, + 243 + ], + "spans": [ + { + "bbox": [ + 131, + 230, + 366, + 243 + ], + "type": "text", + "content": "4.5 Safe-CLIP for Image-to-Text Generation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 248, + 479, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 479, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 479, + 308 + ], + "type": "text", + "content": "Finally, we assess the capabilities of the Safe-CLIP visual encoder when applied to an existing multimodal LLM [5]. We employ LLaVA [25] based on LLama 2-13B-Chat, prompted by asking to describe a given image. Results are reported in Table 5 in terms of percentage of NSFW generated texts measured with GPT-3.5 and toxicity degree computed using the Perspective API. Also for this" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 310, + 236, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 310, + 236, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 310, + 236, + 403 + ], + "type": "text", + "content": "experiment, we employ the real NSFW images from the three different NSFW sources described in Sec. 4.3. As it can be seen, SafeCLIP can significantly reduce the probability" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 246, + 352, + 476, + 399 + ], + "blocks": [ + { + "bbox": [ + 242, + 316, + 480, + 350 + ], + "lines": [ + { + "bbox": [ + 242, + 316, + 480, + 350 + ], + "spans": [ + { + "bbox": [ + 242, + 316, + 480, + 350 + ], + "type": "text", + "content": "Table 5: Percentage of generating NSFW textual sentences and their toxicity degree, when using real NSFW images from different sources as input." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 246, + 352, + 476, + 399 + ], + "lines": [ + { + "bbox": [ + 246, + 352, + 476, + 399 + ], + "spans": [ + { + "bbox": [ + 246, + 352, + 476, + 399 + ], + "type": "table", + "html": "
ModelNudeNetNSFW URLsSMID
% NSFWToxicity% NSFWToxicity% NSFWToxicity
LLaVA [25]62.638.646.824.922.24.7
+ Safe-CLIP26.716.519.410.811.73.7
", + "image_path": "4af92e3de7c0e16f42fa5655f3357b35eb40acff33d8a3675df35b36ee279a6f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 404, + 479, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 404, + 479, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 404, + 479, + 429 + ], + "type": "text", + "content": "of generating inappropriate textual sentences compared to the original LLaVA, demonstrating its effectiveness also in this setting." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 445, + 220, + 457 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 445, + 220, + 457 + ], + "spans": [ + { + "bbox": [ + 131, + 445, + 220, + 457 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 465, + 480, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 465, + 480, + 549 + ], + "spans": [ + { + "bbox": [ + 130, + 465, + 480, + 549 + ], + "type": "text", + "content": "We presented Safe-CLIP, an approach for fine-tuning a CLIP-like model to make it safer and less sensitive to NSFW concepts. Our approach is based on automatically collecting a large synthetic dataset with safe and unsafe images and captions, with which we fine-tune CLIP with losses designed to redirect unsafe content while preserving the structure of the embedding space. Experimental results demonstrate the appropriateness of our solution for cross-modal retrieval, image-to-text and text-to-image generation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 552, + 481, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 552, + 481, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 552, + 481, + 647 + ], + "type": "text", + "content": "Mitigating potential misuse of the ViSU dataset. To mitigate potential misuse of the dataset, we release it via a request form, with only the textual portion available6. Access is granted exclusively to verified researchers, who must declare their intention to use the data solely for research purposes, explicitly committing to non-malicious use. The NSFW visual part of the dataset, due to its explicit nature, is withheld to avoid potential misuse. Nonetheless, the full reproducibility of the dataset is ensured given that we employed publicly available diffusion models and we release the generation seeds and instructions." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "S. Poppi et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 385, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 385, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 385, + 665 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 133, + 653, + 385, + 665 + ], + "type": "text", + "content": " https://huggingface.co/datasets/aimagelab/ViSU-Text" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "content": "We acknowledge the CINECA award under the ISCRA initiative, for the availability of high-performance computing resources. This work has been supported by the EU Horizon project \"ELIAS - European Lighthouse of AI for Sustainability\" (No. 101120237), and by the the PNRR projects \"FAIR - Future Artificial Intelligence Research\" (M4C2 - PE00000013) and \"ITSERR - Italian Strengthening of Esfri RI Resilience\" (CUP B53C22001770006), both funded by the EU - NextGenerationEU." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 243, + 197, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 243, + 197, + 255 + ], + "spans": [ + { + "bbox": [ + 133, + 243, + 197, + 255 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 269, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 269, + 480, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 269, + 480, + 312 + ], + "spans": [ + { + "bbox": [ + 138, + 269, + 480, + 312 + ], + "type": "text", + "content": "1. Bakker, M., Chadwick, M., Sheahan, H., Tessler, M., Campbell-Gillingham, L., Balaguer, J., McAleese, N., Glaese, A., Aslanides, J., Botvinick, M., et al.: Finetuning language models to find agreement among humans with diverse preferences. In: NeurIPS (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 313, + 480, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 313, + 480, + 334 + ], + "spans": [ + { + "bbox": [ + 138, + 313, + 480, + 334 + ], + "type": "text", + "content": "2. Bedapudi, P.: NudeNet: Neural Nets for Nudity Classification, Detection, and Selective Censoring (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 335, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 335, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 138, + 335, + 480, + 357 + ], + "type": "text", + "content": "3. Birhane, A., Prabhu, V.U.: Large image datasets: A pyrrhic win for computer vision? In: WACV (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 357, + 480, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 357, + 480, + 378 + ], + "spans": [ + { + "bbox": [ + 138, + 357, + 480, + 378 + ], + "type": "text", + "content": "4. Birhane, A., Prabhu, V.U., Kahembwe, E.: Multimodal datasets: misogyny, pornography, and malignant stereotypes. arXiv preprint arXiv:2110.01963 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 380, + 480, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 380, + 480, + 412 + ], + "spans": [ + { + "bbox": [ + 138, + 380, + 480, + 412 + ], + "type": "text", + "content": "5. Caffagni, D., Cocchi, F., Barsellotti, L., Moratelli, N., Sarto, S., Baraldi, L., Baraldi, L., Cornia, M., Cucchiara, R.: The Revolution of Multimodal Large Language Models: A Survey. In: ACL Findings (2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 413, + 480, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 413, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 138, + 413, + 480, + 434 + ], + "type": "text", + "content": "6. Cao, Y., Yang, J.: Towards Making Systems Forget with Machine Unlearning. In: IEEE Symposium on Security and Privacy (2015)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 434, + 480, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 434, + 480, + 467 + ], + "spans": [ + { + "bbox": [ + 138, + 434, + 480, + 467 + ], + "type": "text", + "content": "7. Cauteruccio, F., Corradini, E., Terracina, G., Ursino, D., Virgili, L.: Extraction and analysis of text patterns from nsfw adult content in reddit. Data & Knowledge Engineering 138, 101979 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 468, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 468, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 138, + 468, + 480, + 500 + ], + "type": "text", + "content": "8. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., Stoica, I., Xing, E.P.: Vicuna: An Open-Source Chatbot Improving GPT-4 with " + }, + { + "bbox": [ + 138, + 468, + 480, + 500 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 138, + 468, + 480, + 500 + ], + "type": "text", + "content": " * ChatGPT Quality (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 501, + 480, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 501, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 138, + 501, + 480, + 521 + ], + "type": "text", + "content": "9. Christiano, P.F., Leike, J., Brown, T., Martic, M., Legg, S., Amodei, D.: Deep reinforcement learning from human preferences. In: NeurIPS (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 522, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 522, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 522, + 480, + 555 + ], + "type": "text", + "content": "10. Crone, D.L., Bode, S., Murawski, C., Laham, S.M.: The Socio-Moral Image Database (SMID): A novel stimulus set for the study of social, moral and affective processes. PloS one 13(1), e0190954 (2018)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 555, + 480, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 480, + 577 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 480, + 577 + ], + "type": "text", + "content": "1. Dettmers, T., Pagnoni, A., Holtzman, A., Zettlemoyer, L.: QLoRA: Efficient Finetuning of Quantized LLMs. arXiv preprint arXiv:2305.14314 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 578, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 480, + 609 + ], + "type": "text", + "content": "2. Gadre, S.Y., Ilharco, G., Fang, A., Hayase, J., Smyrnis, G., Nguyen, T., Marten, R., Wortsman, M., Ghosh, D., Zhang, J., et al.: DataComp: In search of the next generation of multimodal datasets. In: NeurIPS (2024)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 610, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 610, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 610, + 480, + 643 + ], + "type": "text", + "content": "3. Gandhi, S., Kokkula, S., Chaudhuri, A., Magnani, A., Stanley, T., Ahmadi, B., Kandaswamy, V., Ovenc, O., Mannor, S.: Scalable Detection of Offensive and Noncompliant Content/Logo in Product Images. In: WACV (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "4. Gandikota, R., Materzynska, J., Fiotto-Kaufman, J., Bau, D.: Erasing Concepts from Diffusion Models. In: ICCV (2023)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 145, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 145, + 90, + 447, + 102 + ], + "type": "text", + "content": "Safe-CLIP: Removing NSFW Concepts from Vision-and-Language Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "15. Gao, P., Han, J., Zhang, R., Lin, Z., Geng, S., Zhou, A., Zhang, W., Lu, P., He, C., Yue, X., Li, H., Qiao, Y.: LLaMA-Adapter V2: Parameter-Efficient Visual Instruction Model. arXiv preprint arXiv:2304.15010 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 172 + ], + "type": "text", + "content": "16. Ginart, A., Guan, M., Valiant, G., Zou, J.Y.: Making AI Forget You: Data Deletion in Machine Learning. In: NeurIPS (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 481, + 194 + ], + "type": "text", + "content": "17. Golatkar, A., Achille, A., Soatto, S.: Eternal Sunshine of the Spotless Net: Selective Forgetting in Deep Networks. In: CVPR (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 194, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 194, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 133, + 194, + 481, + 216 + ], + "type": "text", + "content": "18. Golatkar, A., Achille, A., Wang, Y.X., Roth, A., Kearns, M., Soatto, S.: Mixed Differential Privacy in Computer Vision. In: CVPR (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 216, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 216, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 133, + 216, + 481, + 237 + ], + "type": "text", + "content": "19. Hidayatullah, A.F., Hakim, A.M., Sembada, A.A.: Adult Content Classification on Indonesian Tweets using LSTM Neural Network. In: ICACSIS (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "type": "text", + "content": "20. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-Rank Adaptation of Large Language Models. arXiv preprint arXiv:2106.09685 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 270, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 481, + 293 + ], + "type": "text", + "content": "21. Karpathy, A., Fei-Fei, L.: Deep visual-semantic alignments for generating image descriptions. In: CVPR (2015)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 293, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 314 + ], + "type": "text", + "content": "22. Kingma, D.P., Ba, J.: Adam: A Method for Stochastic Optimization. In: ICLR (2015)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 315, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 335 + ], + "type": "text", + "content": "23. Kumari, N., Zhang, B., Wang, S.Y., Shechtman, E., Zhang, R., Zhu, J.Y.: Ablating concepts in text-to-image diffusion models. In: ICCV (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 336, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 481, + 357 + ], + "type": "text", + "content": "24. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft COCO: Common Objects in Context. In: ECCV (2014)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 358, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 369 + ], + "type": "text", + "content": "25. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual Instruction Tuning. In: NeurIPS (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "type": "text", + "content": "26. Liu, Y., Singh, A., Freeman, C.D., Co-Reyes, J.D., Liu, P.J.: Improving Large Language Model Fine-tuning for Solving Math Problems. arXiv preprint arXiv:2310.10047 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 402, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 481, + 435 + ], + "type": "text", + "content": "27. Markov, T., Zhang, C., Agarwal, S., Nekoul, F.E., Lee, T., Adler, S., Jiang, A., Weng, L.: A Holistic Approach to Undesired Content Detection in the Real World. In: AAAI (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 435, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 456 + ], + "type": "text", + "content": "28. Materzyńska, J., Torralba, A., Bau, D.: Disentangling Visual and Written Concepts in CLIP. In: CVPR (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 456, + 481, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 456, + 481, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 456, + 481, + 500 + ], + "type": "text", + "content": "29. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: GLIDE: Towards Photorealistic Image Generation and Editing with Text-Guided Diffusion Models. arXiv preprint arXiv:2112.10741 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 500, + 481, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 500, + 481, + 523 + ], + "spans": [ + { + "bbox": [ + 132, + 500, + 481, + 523 + ], + "type": "text", + "content": "30. Oord, A.v.d., Li, Y., Vinyals, O.: Representation Learning with Contrastive Predictive Coding. arXiv preprint arXiv:1807.03748 (2018)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 523, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 481, + 555 + ], + "type": "text", + "content": "31. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. In: NeurIPS (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 555, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 481, + 578 + ], + "type": "text", + "content": "32. Poppi, S., Sarto, S., Cornia, M., Baraldi, L., Cucchiara, R.: Multi-Class Unlearning for Image Classification via Weight Filtering. IEEE Intelligent Systems (2024)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 578, + 481, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 481, + 610 + ], + "type": "text", + "content": "33. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning Transferable Visual Models From Natural Language Supervision. In: ICML (2021)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 610, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 481, + 632 + ], + "type": "text", + "content": "34. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I.: Language Models are Unsupervised Multitask Learners. OpenAI Blog 1(8), 9 (2019)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "35. Rafailov, R., Sharma, A., Mitchell, E., Ermon, S., Manning, C.D., Finn, C.: Direct Preference Optimization: Your Language Model is Secretly a Reward Model. In: NeurIPS (2023)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "S. Poppi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 654 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "text", + "content": "36. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 481, + 160 + ], + "type": "text", + "content": "37. Sanh, V., Debut, L., Chaumont, J., Wolf, T.: DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "type": "text", + "content": "38. Schramowski, P., Brack, M., Deiseroth, B., Kersting, K.: Safe Latent Diffusion: Mitigating Inappropriate Degeneration in Diffusion Models. In: CVPR (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "type": "text", + "content": "39. Schramowski, P., Tauchmann, C., Kersting, K.: Can Machines Help Us Answering Question 16 in Datasheets, and In Turn Reflecting on Inappropriate Content? In: ACM FAccT (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 216, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 270 + ], + "type": "text", + "content": "40. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., Schramowski, P., Kundurthy, S., Crowson, K., Schmidt, L., Kaczmarczyk, R., Jitsev, J.: LAION-5B: An open large-scale dataset for training next generation image-text models. In: NeurIPS (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "type": "text", + "content": "41. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: LAION-400M: Open Dataset of CLIP-Filtered 400 Million Image-Text Pairs. In: NeurIPS Workshops (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "type": "text", + "content": "42. Shen, S., Li, L.H., Tan, H., Bansal, M., Rohrbach, A., Chang, K.W., Yao, Z., Keutzer, K.: How Much Can CLIP Benefit Vision-and-Language Tasks? In: ICLR (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 336, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 380 + ], + "type": "text", + "content": "43. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., Lample, G.: LLaMA: Open and Efficient Foundation Language Models. arXiv preprint arXiv:2302.13971 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "text", + "content": "44. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open Foundation and Fine-Tuned Chat Models. arXiv preprint arXiv:2307.09288 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "type": "text", + "content": "45. Trager, M., Perera, P., Zancato, L., Achille, A., Bhatia, P., Soatto, S.: Linear Spaces of Meanings: Compositional Structures in Vision-Language Models. In: ICCV (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "type": "text", + "content": "46. Tunstall, L., Beeching, E., Lambert, N., Rajani, N., Rasul, K., Belkada, Y., Huang, S., von Werra, L., Fourrier, C., Habib, N., et al.: Zephyr: Direct Distillation of LM Alignment. arXiv preprint arXiv:2310.16944 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 479, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 500 + ], + "type": "text", + "content": "47. Wang, M., Xing, J., Liu, Y.: ActionCLIP: A New Paradigm for Video Action Recognition. arXiv preprint arXiv:2109.08472 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "type": "text", + "content": "48. Wang, Y., Kordi, Y., Mishra, S., Liu, A., Smith, N.A., Khashabi, D., Hajishirzi, H.: Self-Instruct: Aligning Language Models with Self-Generated Instructions. arXiv preprint arXiv:2212.10560 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "type": "text", + "content": "49. Zhang, E., Wang, K., Xu, X., Wang, Z., Shi, H.: Forget-Me-Not: Learning to Forget in Text-to-Image Diffusion Models. arXiv preprint arXiv:2303.17591 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 555, + 480, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 480, + 587 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 480, + 587 + ], + "type": "text", + "content": "50. Zhang, R., Han, J., Zhou, A., Hu, X., Yan, S., Lu, P., Li, H., Gao, P., Qiao, Y.: LLaMA-Adapter: Efficient Fine-tuning of Language Models with Zero-init Attention. arXiv preprint arXiv:2303.16199 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 588, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 588, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 480, + 621 + ], + "type": "text", + "content": "51. Zheng, L., Chiang, W.L., Sheng, Y., Zhuang, S., Wu, Z., Zhuang, Y., Lin, Z., Li, Z., Li, D., Xing, E., et al.: Judging llm-as-a-judge with mt-bench and chatbot arena. arXiv preprint arXiv:2306.05685 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 622, + 480, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 480, + 654 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 480, + 654 + ], + "type": "text", + "content": "52. Zhou, C., Liu, P., Xu, P., Iyer, S., Sun, J., Mao, Y., Ma, X., Efrat, A., Yu, P., Yu, L., et al.: LIMA: Less Is More for Alignment. arXiv preprint arXiv:2305.11206 (2023)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 145, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 145, + 91, + 447, + 102 + ], + "type": "text", + "content": "Safe-CLIP: Removing NSFW Concepts from Vision-and-Language Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Safe-Sim_ Safety-Critical Closed-Loop Traffic Simulation with Diffusion-Controllable Adversaries/548909b9-64d9-4498-bf91-808d29ce48c5_content_list.json b/2024/Safe-Sim_ Safety-Critical Closed-Loop Traffic Simulation with Diffusion-Controllable Adversaries/548909b9-64d9-4498-bf91-808d29ce48c5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1ceff662aa1fddd01f73c0db3a1fa69c6b63e4f1 --- /dev/null +++ b/2024/Safe-Sim_ Safety-Critical Closed-Loop Traffic Simulation with Diffusion-Controllable Adversaries/548909b9-64d9-4498-bf91-808d29ce48c5_content_list.json @@ -0,0 +1,1851 @@ +[ + { + "type": "text", + "text": "SAFE-SIM: Safety-Critical Closed-Loop Traffic Simulation with Diffusion-Controllable Adversaries", + "text_level": 1, + "bbox": [ + 215, + 140, + 785, + 183 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wei-Jer Chang $^{1}$ , Francesco Pittaluga $^{2}$ , Masayoshi Tomizuka $^{1}$ , Wei Zhan $^{1}$ , and Manmohan Chandraker $^{2,3}$", + "bbox": [ + 218, + 210, + 781, + 242 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "UC Berkeley 2 NEC Labs America 3 UC San Diego", + "bbox": [ + 292, + 253, + 705, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Evaluating the performance of autonomous vehicle planning algorithms necessitates simulating long-tail safety-critical traffic scenarios. However, traditional methods for generating such scenarios often fall short in terms of controllability and realism; they also neglect the dynamics of agent interactions. To address these limitations, we introduce SAFE-SIM, a novel diffusion-based controllable closed-loop safety-critical simulation framework. Our approach yields two distinct advantages: 1) generating realistic long-tail safety-critical scenarios that closely reflect real-world conditions, and 2) providing controllable adversarial behavior for more comprehensive and interactive evaluations. We develop a novel approach to simulate safety-critical scenarios through an adversarial term in the denoising process of diffusion models, which allows an adversarial agent to challenge a planner with plausible maneuvers while all agents in the scene exhibit reactive and realistic behaviors. Furthermore, we propose novel guidance objectives and a partial diffusion process that enables users to control key aspects of the scenarios, such as the collision type and aggressiveness of the adversarial agent, while maintaining the realism of the behavior. We validate our framework empirically using the nuScenes and nuPlan datasets across multiple planners, demonstrating improvements in both realism and controllability. These findings affirm that diffusion models provide a robust and versatile foundation for safety-critical, interactive traffic simulation, extending their utility across the broader autonomous driving landscape. Project website: https://safe-sim.github.io/.", + "bbox": [ + 261, + 309, + 743, + 628 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 655, + 375, + 671 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A key safety feature of autonomous vehicles (AVs) is their ability to navigate near-collision events in real-world scenarios. However, these events rarely occur on roads and testing AVs in such high-risk situations on public roads is unsafe. Therefore, simulation is indispensable in the development and assessment of AVs, providing a safe and reliable means to study their safety and dependability. A critical aspect of simulation is modeling the behavior of other road users, since AVs must learn to interact with them safely.", + "bbox": [ + 212, + 688, + 787, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A common method of safety-critical testing of AVs involves manually designing scenarios that could potentially lead to failures, such as collisions. While this approach allows for targeted testing, it is inherently limited in scalability and", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/ca940f7ca32dd9a76e618ddbf2622835949d1ac278fbf36f54544ae2d0ebbf70.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1: Comparison of methods. Our contribution is the development of a framework for (a) safety-critical (b) closed-loop (c) controllable adversarial simulations. These aspects are not concurrently present in previous frameworks. We formulate a novel partial diffusion with novel guidance functions for stable long-term simulation and are the first to enable an ego planner to be tested against controllable adversaries with varied behavior patterns." + ], + "table_body": "
MethodSafety-CriticalControllableControllable \nAdversaryEvaluate \nPlannerClosed-LoopReal-World
CTG [39]×✓××✓✓
CTG++ [38]×✓××✓✓
STRIVE [26]✓××✓✓✓
DiffScene [34]✓✓×✓××
SAFE-SIM (Ours)✓✓✓✓✓✓
", + "bbox": [ + 215, + 143, + 782, + 256 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "lacks the comprehensiveness required for thorough evaluation [7,8,33]. Some recent works focus on automatically generating challenging scenarios that cause planners to fail, but their emphasis has been mostly on static scenario generation rather than dynamic, closed-loop simulations. This results in a critical gap: the behavior of other agents often does not adapt or respond to the planner's actions, which is essential for a comprehensive safety evaluation. Furthermore, the results from these simulations often lack controllability, typically producing only a single adversarial outcome per scenario without the flexibility to explore a range of conditions and responses.", + "bbox": [ + 212, + 351, + 787, + 487 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we introduce SAFE-SIM, a closed-loop simulation framework for generating safety-critical scenarios, with a particular emphasis on controllability and realism for the behavior of agents, which allows simulations over a long-horizon as needed to evaluate AV planning algorithms (Fig. 1). Different from prior works [26,34,38,39] that primarily adhere to rule-constraint satisfaction, our approach enhances controllability by modulating adversarial vehicle behaviors within identical scenarios, thereby facilitating a broader exploration of potential outcomes. See Tab. 1 for a comprehensive comparison of these approaches.", + "bbox": [ + 212, + 489, + 787, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our approach builds upon recent developments in controllable diffusion models [18, 25, 39]. Specifically, we adopt a test-time guidance to direct the denoising phase of the diffusion process, using the gradients from differentiable objectives to enhance scenario generation, enabling generation of adversarial scenarios in which and adversarial agent collides with the ego agent behaving according to specific planning policy. Additionally, we develop an novel approach, which we refer to as Partial Diffusion that introduces trajectory proposals into the diffusion process to provide a high degree of controllability over the type of collision scenario. Overall, our balanced integration of adversarial objectives with regularization during the guidance phase combined with Partial Diffusion allows for refined control over the conditions of the generated scenarios, ensuring both their realism and relevance to safety-critical testing.", + "bbox": [ + 212, + 611, + 787, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In our study, we use the nuScenes [2] and nuPlan [3] datasets to evaluate the efficacy of our method in generating safety-critical closed-loop simulations. Our results demonstrate a marked improvement in the controllability and real-", + "bbox": [ + 212, + 794, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "W.J. Chang et al.", + "bbox": [ + 271, + 114, + 392, + 128 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a4cfa613335605d7fccd1ea9eb1f093ea3313a5383d54d4400afd0f7241febda.jpg", + "image_caption": [ + "Fig. 1: Overview of SAFE-SIM Framework for Controllable Safety-Critical Closed-Loop Simulation. This framework evaluates a planner within scenarios featuring multiple controllable reactive agents. These agents have two distinct roles: adversarial agents, which actively challenge the planner by exhibiting controllable adversarial behaviors such as specific collision types and levels of aggressiveness, and non-adversarial agents, which follow normal driving behavior to maintain the realism of the entire scene. Such a setup facilitates the generation of various realistic, interactive, and safety-critical scenarios, providing a thorough evaluation of the planner's capabilities." + ], + "image_footnote": [], + "bbox": [ + 228, + 148, + 772, + 299 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ism of scenarios compared to previous adversarial scenario generation methods. Furthermore, we showcase the advantage of our proposed framework in varying the safety-criticality and collision types of scenarios. These attributes make our approach particularly well-suited for the closed-loop simulation of AVs, providing a more reliable and comprehensive framework for safety evaluation.", + "bbox": [ + 212, + 452, + 787, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 547, + 387, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Traffic Simulation", + "text_level": 1, + "bbox": [ + 215, + 575, + 408, + 589 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Traffic simulation can broadly be categorized into two main groups: heuristic-based and learning-based methods. In heuristic-based methods, agents are controlled by human-specified rules, such as the Intelligent Driver Model (IDM) [30] to follow a leading vehicle while maintaining a safe following distance. However, these methods have modeling capacity issues and may not reflect the real traffic distribution; a large domain gap limits its usage for planning evaluation. To close the gap, data-driven approaches learn from real driving datasets to imitate real-world behavior [23, 28, 29, 35]. TrafficSim [28] utilizes a trained variational autoencoder for scene-level traffic simulation, while BITS [35] combines high-level goal inference with low-level driving behavior imitation to enhance the realism of simulated driving behavior. Recently, the Waymo SimAgents challenge has focused on whether simulators can accurately represent real-world driving distributions [20]. However, little work has specifically focused on behavior simulation for generating safety-critical, long-tail scenarios.", + "bbox": [ + 212, + 598, + 787, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion models [27, 31, 32] have shown significant promise for synthetic image generation. One of the key advantages of diffusion models is controllability,", + "bbox": [ + 212, + 809, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Safety-Critical Closed-Loop Traffic Simulation with Controllable Adversaries", + "bbox": [ + 220, + 114, + 732, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "which can take the form of classifier [6], classifier-free [15], and reconstruction [13] guidance. Recently, controllable diffusion models have been employed for planning and traffic simulation [25, 38, 39] via guidance. We adopt trajectory diffusion models to develop a novel approach for generate safety-critical realistic traffic simulations in which an adversarial agent collides with an ego planner agent. Other works [19, 24] use diffusion with guidance or conditioning to achieve controllable scene initialization. In contrast, we emphasize closed-loop, controllable adversarial behavior simulation based on real-world data initialization rather than scene initialization.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Safety-Critical Traffic Simulation", + "text_level": 1, + "bbox": [ + 214, + 305, + 532, + 320 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Safety-critical traffic generation plays a crucial role in training and evaluating AV systems, enhancing their capability to navigate diverse real-world scenarios and enhancing robustness. Gradient-based methods that leverage back-propagation to create safety-critical scenarios have been proposed to evaluate AV prediction and planning models [4,11]. Hanselmann et al. use kinematic gradients to modify vehicle trajectories, with the goal of improving the robustness of imitation learning planners [11]. Cao et al. have developed a model with differentiable dynamics, enabling the generation of realistic adversarial trajectories for trajectory prediction models through backpropagation techniques [4]. Black-box optimization approaches include perturbing actions based on kinematic bicycle models [33] and using Bayesian Optimization to create adversarial self-driving scenarios that escalate collision risks with simulated entities [1]. Zhang et al. target trajectory prediction models via white- and black-box attacks that adversarily perturb real driving trajectories [37]. For an extensive review of this topic, we direct readers to Ding et al. [9].", + "bbox": [ + 212, + 330, + 787, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The field has recently seen advancements in data-driven methods for safety-critical scenario generation [26,34,36]. For instance, Xu et al. introduce a diffusion-based approach in CARLA, applying various adversarial optimization objectives to guide the diffusion process for safety-critical scenario generation [34]. Rempe et al. proposed STRIVE to utilize gradient-based adversarial optimization on the latent space, constrained by a graph-based CVAE traffic motion model, to generate realistic safety-critical scenarios for rule-based planners [26]. However, a common limitation in these approaches is the absence of closed-loop interaction, essential for accurately simulating interactive real-world driving.", + "bbox": [ + 212, + 556, + 789, + 693 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Problem Formulation", + "text_level": 1, + "bbox": [ + 214, + 717, + 459, + 732 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We consider a simulated interactive traffic scenario consisting of $N$ agents; one is the ego vehicle controlled by the planner $\\pi$ , and the remaining $N - 1$ are reactive agents modeled by a function $g$ . Our objective is to create a safety-critical closed-loop collision simulation, where reactive agents demonstrate realistic, controllable behavior. Of the $N - 1$ reactive agents, one or a subset is considered the adversarial agents (denoted as agent $a$ ), meant to collide with the ego vehicle.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "W.J. Chang et al.", + "bbox": [ + 271, + 114, + 392, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The adversarial agent, formulated within the reactive agent model $g$ , is governed by an adversarial term designed (detailed in Sec. 5) to be both controllable and adversarial to the planner $\\pi$ . This setup allows the adversarial agent to pose direct challenges to $\\pi$ , testing its resilience in complex scenarios. Concurrently, the other non-adversarial agents, also controlled by $g$ with varying parameters, emulate authentic, reactive behaviors, thus enriching the simulation scenario with realistic and diverse traffic conditions. The dual role of the adversarial agent and non-adversarial agents ensures that while it challenges $\\pi$ , the overall simulation environment plausibly represents real-world driving conditions.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "At any given timestep $t$ , the states of the $N$ vehicles are represented as $\\mathbf{s}_t = [\\mathbf{s}_t^1, \\dots, \\mathbf{s}_t^N]$ , where $\\mathbf{s}_t^i = (x_t^i, y_t^i, v_t^i, \\theta_t^i)$ indicates the 2D position, speed, and yaw of vehicle $i$ . The corresponding actions for each vehicle are $\\mathbf{a}_t = [\\mathbf{a}_t^1, \\dots, \\mathbf{a}_t^N]$ , with $\\mathbf{a}_t^i = (\\dot{v}_t^i, \\dot{\\theta}_t^i)$ representing the acceleration and yaw rate. To predict the state at the next timestep $t + 1$ , a transition function $f$ is used, which computes $\\mathbf{s}_{t+1} = f(\\mathbf{s}_t, \\mathbf{a}_t)$ based on current state and action. We adopt unicycle dynamics as the transition function.", + "bbox": [ + 212, + 282, + 787, + 386 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Each agent's decision context is $\\mathbf{c}_t^i$ , which includes the agent-centric map $I^i$ and the $T_{\\mathrm{hist}}$ historical states of neighboring vehicles from time $t - T_{\\mathrm{hist}}$ to $t$ , defined as $\\mathbf{s}_{t - T_{\\mathrm{hist}}:t} = \\{\\mathbf{s}_{t - T_{\\mathrm{hist}}}, \\dots, \\mathbf{s}_t\\}$ . In closed-loop traffic simulation, each agent continuously generates and updates its trajectory based on the current decision context $\\mathbf{c}_t^i$ . After generating a trajectory, the simulation executes the first few steps of the planned actions before updating $\\mathbf{c}_t^i$ and re-planning. See Sec. 6.2 for more implementation details.", + "bbox": [ + 212, + 387, + 787, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Planner $\\pi$ The planner $\\pi$ determines the ego vehicle's future trajectory over a time horizon $t$ to $t + T$ . The planned state sequence is denoted by $s_{t:t + T}^{1} = \\pi (\\mathbf{c}_{t}^{1})$ , where $\\pi (\\mathbf{c}_t^1)$ processes the historical states and map data within $\\mathbf{c}_t^1$ to plan future states based on the current scene context.", + "bbox": [ + 212, + 506, + 787, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Reactive Agents $g$ The reactive agent model $g$ , parameterized by $\\theta$ , is designed to simulate the behavior of the $N - 1$ non-ego vehicles, represented by the set $\\{s_{t:t + T}^i\\}_{i = 2}^N$ . Each vehicle's state sequence, $s_{t:t + T}^i$ , is generated by $g_{\\theta}(\\mathbf{c}_t^i,\\psi_i)$ , which incorporates the decision context $\\mathbf{c}_t^i$ and a set of control parameters $\\psi_{i}$ unique to each agent. These parameters $\\psi_{i}$ enable the fine-tuning of individual behaviors within the simulation. In our approach, we train the model $g$ on real-world driving data to ensure the trajectories it produces are not only controllable, supporting the generation of various safety-critical scenarios, but also realistic.", + "bbox": [ + 214, + 578, + 787, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 Diffusion Models for Traffic Simulation", + "text_level": 1, + "bbox": [ + 214, + 720, + 629, + 736 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For closed-loop safety-critical traffic simulation, the reactive agents, especially the adversarial agent, should be 1) controllable, and 2) realistic. With recent advances in controllable diffusion models [18,25,39], we adopt trajectory diffusion models to generate realistic simulations.", + "bbox": [ + 212, + 750, + 785, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We define the model's operational trajectory as $\\tau$ , which comprises both action and state sequences: $\\tau \\coloneqq [\\tau_{a}, \\tau_{s}]$ . Specifically, $\\tau_{a} \\coloneqq [a_{0}, \\ldots, a_{T-1}]$ represents the", + "bbox": [ + 214, + 809, + 785, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Safety-Critical Closed-Loop Traffic Simulation with Controllable Adversaries", + "bbox": [ + 220, + 113, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "sequence of actions, while $\\tau_{s} \\coloneqq [s_{1},\\ldots ,s_{T}]$ denotes the corresponding sequence of states. Following the approach described in [39], our model predicts the action sequence $\\tau_{a}$ , and the state sequence $\\tau_{s}$ can be derived starting from the initial state $s_0$ and dynamic model $f$ .", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A diffusion model generates a trajectory by reversing a process that incrementally adds noise. Starting with an actual trajectory $\\tau_0$ sampled from the data distribution $q(\\tau_0)$ , a sequence of increasingly noisy trajectories $(\\tau_{1},\\tau_{2},\\dots,\\tau_{K})$ is produced via a forward noisng process. Each trajectory $\\tau_{k}$ at step $k$ is generated by adding Gaussian noise parameterized by a predefined variance schedule $\\beta_{k}$ [14]:", + "bbox": [ + 212, + 207, + 787, + 282 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\tau_ {1: K} \\mid \\tau_ {0}\\right) := \\prod_ {k = 1} ^ {K} q \\left(\\tau_ {k} \\mid \\tau_ {k - 1}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 401, + 295, + 787, + 335 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\tau_ {k} \\mid \\tau_ {k - 1}\\right) := \\mathcal {N} \\left(\\tau_ {k}; \\sqrt {1 - \\beta_ {k}} \\tau_ {k - 1}, \\beta_ {k} \\mathbf {I}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 362, + 345, + 785, + 362 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The noising process gradually obscures the data, where the final noisy version $q(\\tau_K)$ approaches $\\mathcal{N}(\\tau_K; \\mathbf{0}, \\mathbf{I})$ . The trajectory generation process is then achieved by learning the reverse of this noising process. Given a noisy trajectory $\\tau_K$ , the model learns to denoise it back to $\\tau_0$ through a sequence of reverse steps. Each reverse step is modeled as:", + "bbox": [ + 214, + 368, + 787, + 444 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(\\tau_ {k - 1} \\mid \\tau_ {k}, \\mathbf {c}\\right) := \\mathcal {N} \\left(\\tau_ {k - 1}; \\mu_ {\\theta} \\left(\\tau_ {k}, k, \\mathbf {c}\\right), \\Sigma_ {k}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 351, + 455, + 785, + 472 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\theta$ are learned functions that predict the mean $\\mu$ of the reverse step, and $\\Sigma_{k}$ is a fixed schedule. By iteratively applying the reverse process, the model learns a trajectory distribution, effectively generating a plausible future trajectory from a noisy start.", + "bbox": [ + 212, + 482, + 782, + 541 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During the trajectory prediction phase, the model estimates the final clean trajectory denoted by $\\hat{\\tau}_0$ . This estimated trajectory is used to compute the mean $\\mu$ as described in [21]. For more details, see supplementary material Sec. D.", + "bbox": [ + 212, + 542, + 782, + 588 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5 Diffusion Models for Safety-Critical Traffic Simulation", + "text_level": 1, + "bbox": [ + 214, + 609, + 774, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The diffusion model, once trained on realistic trajectory data, inherently reflects the behavioral patterns present in its training distribution. However, to effectively simulate and analyze safety-critical scenarios, there is a crucial need for a mechanism that allows for the controlled manipulation of agent behaviors [18,39]. This is particularly important for generating adversarial behaviors and ensuring scene consistency in simulations.", + "bbox": [ + 212, + 642, + 787, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1 Guiding Reactive Agents", + "text_level": 1, + "bbox": [ + 214, + 753, + 470, + 771 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our approach specifically introduces guidance to the sampled trajectories at each denoising step, aligning them with predefined objectives $J(\\tau)$ . The concept of guidance involves using the gradient of $J$ to subtly perturb the predicted mean of the model at each denoising step. This process enables the generation", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "W.J. Chang et al.", + "bbox": [ + 271, + 114, + 392, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b756d8af192c285b659b43e12a9984040344b6a5e71fd95b94968fad6c5642ee.jpg", + "image_caption": [ + "Fig.2: Guided Diffusion Process for the Adversarial Agent. This process optimizes the adversarial agent's trajectory using the adversarial cost function $J_{\\mathrm{adv}}$ to the ego vehicle. In particular, we introduce $J_{\\mathrm{control}}$ to vary the adversarial behavior. Simultaneously, it applies regularization through $J_{\\mathrm{reg}}$ for maintaining realism." + ], + "image_footnote": [], + "bbox": [ + 220, + 148, + 782, + 277 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "of trajectories that not only reflect realistic behavior but also cater to specific simulation needs, such as adversarial testing and maintaining scene consistency over extended periods. We adopt the reconstruction guidance (clean guidance) introduced from [16, 25]:", + "bbox": [ + 212, + 376, + 787, + 439 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\tau} _ {0} = \\hat {\\tau} _ {0} - \\alpha \\Sigma_ {k} \\nabla_ {\\tau_ {k}} J (\\hat {\\tau} _ {0}) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 413, + 446, + 787, + 465 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This strategy improves guidance robustness, yielding smoother, more stable trajectories without the usual numerical issues from noisy data.", + "bbox": [ + 212, + 473, + 787, + 503 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In practice, diversifying the behavior of adversarial agents within the same scenarios is crucial for a thorough assessment of AVs. Despite the significance of this challenge, it remains largely unexplored in previous works [26,39]", + "bbox": [ + 212, + 503, + 787, + 549 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The loss function for the non-reactive agents, $J(\\tau)$ , consists of a collision term $J_{\\mathrm{coll}}$ , which encourages collisions between the adversarial agent and the ego agent, two control terms $J_{\\mathrm{v}}$ and $J_{\\mathrm{ttc}}$ , which control the relative speed and time-to-collision between the ego and adversarial agent respectively, a regularization term $J_{\\mathrm{Gauss}}$ , which discourages collisions between the reactive agents, and a route guidance term $J_{\\mathrm{route}}$ , which discourages the reactive agents from going outside the road:", + "bbox": [ + 212, + 549, + 787, + 652 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nJ (\\tau) = \\rho \\underbrace {\\left(J _ {\\mathrm {c o l l}} + J _ {\\mathrm {v}} + J _ {\\mathrm {t t c}}\\right)} _ {J _ {\\mathrm {a d v}} (\\tau)} + \\underbrace {J _ {\\text {r o u t e}} + J _ {\\text {G a u s s}}} _ {J _ {\\text {r e g}} (\\tau)}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 339, + 654, + 787, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\rho$ denotes a scalar weight that determines whether a reactive agent behaves adversarially towards the ego agent, i.e., whether it attempts to collide with the ego agent.", + "bbox": [ + 212, + 696, + 787, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Collision with Planner We define $J_{\\mathrm{coll}}$ to encourage the collision between the adversarial agent and the ego agent, given by:", + "bbox": [ + 212, + 762, + 787, + 792 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nJ _ {\\text {c o l l}} = - \\sum_ {t = 1} ^ {T} d (t), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 434, + 801, + 787, + 844 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Safety-Critical Closed-Loop Traffic Simulation with Controllable Adversaries", + "bbox": [ + 220, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $d(t)$ represents the distance between the ego and the adversarial agent at each time step of the planning horizon $T$ . The adversarial agent is either pre-selected based on lane proximity or dynamically selected based on the distance to the ego agent. Details are in the supplementary Sec. D.4.", + "bbox": [ + 212, + 146, + 787, + 207 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Safety Criticality of Collisions We control the relative speed $J_{v}$ between the ego and adversary at each time step $(v_{t}^{1}$ and $v_{t}^{a}$ ), and the time-to-collision (TTC) cost $J_{\\mathrm{ttc}}$ [22] to control the safety criticality of potential collisions, with the latter given by:", + "bbox": [ + 212, + 227, + 787, + 287 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nJ _ {\\mathrm {t t c}} = \\sum_ {t = 1} ^ {T} - \\exp \\left(- \\frac {\\tilde {t} _ {\\operatorname {c o l} (t)} ^ {2}}{2 \\lambda_ {t}} - \\frac {\\tilde {d} _ {\\operatorname {c o l} (t)} ^ {2}}{2 \\lambda_ {d}}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 297, + 787, + 338 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\tilde{t}_{\\mathrm{col}(t)}$ is the time to collision at time $t$ , $\\tilde{d}_{\\mathrm{col}(t)}$ is the distance to collision and $\\lambda_t$ and $\\lambda_d$ are bandwidth parameters for time and distance. This formula uses a constant velocity assumption. Intuitively, the time-to-collision cost favors scenarios with high relative speeds and challenging collision angles for the ego vehicle to avoid. For details of $J_v$ and $J_{\\mathrm{ttc}}$ , see supplementary Sec. C.2.", + "bbox": [ + 212, + 348, + 787, + 426 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Route Guidance Given an agent's trajectory $\\tau$ and the corresponding route $r$ -the predefined path on a lane graph from its starting point to its destination—we compute the normal distance of each point $\\tau_{t}$ on the trajectory to the route at each timestep. We then penalize deviations from the route that exceed a predefined margin $d$ . This process is captured by the following route guidance cost function:", + "bbox": [ + 212, + 444, + 787, + 534 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nJ _ {\\text {r o u t e}} (\\tau , r) = \\sum_ {t = 1} ^ {T} \\max \\left(0, \\left| d _ {n} \\left(\\tau_ {t}, r\\right) - d _ {m} \\right|\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 354, + 547, + 787, + 587 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $d_{n}(\\tau_{n},r)$ denotes the normal distance from the point $\\tau_{t}$ on the trajectory to the nearest point on the route $r$ at timestep $t$ , and $d_{m}$ represents the acceptable deviation margin from the route. In contrast to the off-road loss in prior studies [39], our proposed route guidance system more effectively indicates each agent's intended path, improving adherence to traffic rules as demonstrated in Sec. 6. The flexibility of route guidance supports diverse agent interactions, such as modifying routes to encourage lane changes among reactive agents [29].", + "bbox": [ + 212, + 592, + 787, + 700 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Gaussian Collision Guidance Given the trajectories of agents, we calculate the Gaussian distance for each pair of agents $(i,j)$ at each timestep $t$ from 1 to $T$ . The Gaussian distance between the agents takes into account both the tangential $(d_t)$ and normal $(d_n)$ components of the projected distances. The aggregated Gaussian distance is:", + "bbox": [ + 212, + 718, + 787, + 792 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nJ _ {\\mathrm {G a u s s}} = \\sum_ {t = 1} ^ {T} \\sum_ {i, j} ^ {N} \\exp \\left(- \\frac {1}{2 \\sigma^ {2}} \\left(\\lambda \\cdot d _ {t} ^ {i j} (t) ^ {2} + d _ {n} ^ {i j} (t) ^ {2}\\right)\\right) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 318, + 801, + 787, + 844 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "W.J. Chang et al.", + "bbox": [ + 271, + 114, + 392, + 128 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7312088dc904c464486e66a5d4db329cf0ad414cb7fd177954ce8dd59f127021.jpg", + "image_caption": [ + "Fig. 3: Framework for Partial Diffusion. We generate proposals based on domain knowledge (e.g., collision types). Users can adjust noise levels to balance between user control and the model's data distribution." + ], + "image_footnote": [], + "bbox": [ + 225, + 146, + 777, + 282 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $d_t^{ij}(t)$ and $d_n^{ij}(t)$ represent the tangential and normal distances from agent $j$ 's trajectory point at time $t$ to agent $i$ 's heading axis, respectively, and $\\sigma$ is the standard deviation for these distances. In this formulation, $\\lambda$ is a scaling factor applied to the tangential distance $d_t^{ij}(t)$ . This approach contrasts with the disk approximation method, which primarily penalizes the Euclidean distance between agents. By accounting for both tangential and normal components, the Gaussian collision distance method significantly reduces collision rate, which we discuss in Sec. 6.", + "bbox": [ + 212, + 359, + 787, + 481 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.2 Partial Diffusion: Controlling Collision Types", + "text_level": 1, + "bbox": [ + 214, + 507, + 635, + 523 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We introduce a novel approach through a partial diffusion process, utilizing trajectory proposals to initiate the diffusion process. This methodology enables the variation in collision types by the adversarial agent within the diffusion, tailoring the adversarial outcomes to specific evaluation needs, the results are discussed in Sec. 6.6.", + "bbox": [ + 212, + 536, + 787, + 611 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Figure 3 illustrates our framework, which is divided into three main steps to generate trajectory proposals for various collision scenarios. First, we create initial trajectory proposals $(\\tau_0)$ aimed at capturing different types of collisions. The next critical step involves setting the partial diffusion ratio $\\gamma$ , which defines the specific point in the process, $k_{p} = \\gamma \\cdot K$ , at which we start modifying the trajectory. Starting from step $k_{p}$ , we adjust the trajectory by adding a precise level of Gaussian noise $\\epsilon \\sim N(0,I)$ : $\\hat{\\tau}_{k_p} = \\sqrt{\\bar{\\alpha}_{k_p}}\\tau_0 + \\sqrt{1 - \\bar{\\alpha}_{k_p}}\\epsilon$ . The final stages include removing noise and using guided diffusion for the rest of the $k_{p}$ steps to refine the trajectory into a realistic path that suits our collision scenario goals.", + "bbox": [ + 212, + 612, + 787, + 750 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To generate the trajectory proposals, we develop a rule-based approach in which we first identify the centerlines of the ego and adversarial agent and then search for potential intersections of their respective centerlines. If such an intersection exists, we generate the proposals by selecting an acceleration value and lateral offset from the centerline that is likely to cause the desired collision type based on the projected plan of the ego agent. Note that the trajectory", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Safety-Critical Closed-Loop Traffic Simulation with Controllable Adversaries", + "bbox": [ + 220, + 113, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/0f1b5f6fd3666f168a6f85918ace569a8d8c85b1afd49f40ce56bdb58db5c66c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 230, + 148, + 405, + 255 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/3b09fa2c72b8599715fa729276012f5d7fec63bfa328ac2b18582f5f7da19fb7.jpg", + "image_caption": [ + "Ego" + ], + "image_footnote": [], + "bbox": [ + 292, + 261, + 328, + 277 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/a693a367c2b58420629a3e21b76ec3eff0c4557317a3525d8f10c72cf706b76d.jpg", + "image_caption": [ + "Fig. 4: Partial Diffusion Results of Rule-Based Planner on NuScenes Dataset." + ], + "image_footnote": [], + "bbox": [ + 359, + 262, + 383, + 273 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/cc07f46929c87ed2c587b8550a66dd8b64c55793cffa7689f5d2f6b1c3ae0728.jpg", + "image_caption": [ + "Non-Adv" + ], + "image_footnote": [], + "bbox": [ + 431, + 148, + 586, + 255 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/1336d5febd7950e8f3440f812f19ef859afb8486895c3a430ecfbefb04c82b18.jpg", + "image_caption": [ + "Adv" + ], + "image_footnote": [], + "bbox": [ + 429, + 262, + 465, + 273 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/d05a2e9ddbb5cec801cafc48acef0785d13fce16ac5a54863ee58c5e8f7d148a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 531, + 262, + 596, + 273 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/316a2f1513fbe5fb695fe3dbe0527b3df2845e09a34b83924a11c3389ec3fff8.jpg", + "image_caption": [ + "Trajectory Proposal" + ], + "image_footnote": [], + "bbox": [ + 599, + 148, + 771, + 255 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The safety-critical scenarios show the framework's ability to create realistic and challenging situations, varying collision types based on different trajectory proposals. The gradient lines reflects the planned trajectories in the next 3.2 seconds.", + "bbox": [ + 212, + 305, + 785, + 348 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "proposals are updated in a closed-loop manner to account for the interaction between the ego and the adversarial agent.", + "bbox": [ + 212, + 366, + 782, + 396 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "This method allows for precise control over the diffusion trajectory, enabling adversarial agents to create customized collision scenarios. Users can adjust $\\gamma$ to fine-tune the balance between explicit control and the model's trained data distribution.", + "bbox": [ + 212, + 397, + 784, + 455 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "6 Experiments", + "text_level": 1, + "bbox": [ + 214, + 481, + 375, + 498 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We validate the efficacy of our proposed framework via experiments with real-world driving data. Our results demonstrate that the framework can generate realistic and controllable adversarial behavior to challenge the planner.", + "bbox": [ + 212, + 513, + 784, + 559 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "6.1 Dataset", + "text_level": 1, + "bbox": [ + 214, + 583, + 326, + 595 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We conduct our experiments on two large-scale real-world driving datasets: nuScenes [2], which consists of 5.5 hours of driving data from two cities, and nuPlan, which consists of 1500 hours of driving data from four cities. We train the model on scenes from the nuScenes train split and evaluate it on the scenes from the nuScenes validation splits and nuPlan mini validation splits. We focus on vehicle-to-vehicle interactions.", + "bbox": [ + 212, + 609, + 784, + 698 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "6.2 Implementation Details", + "text_level": 1, + "bbox": [ + 214, + 723, + 455, + 738 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Baselines. We compare our approach against STRIVE [26] using their open-source implementation and our re-implementation of DiffScene [34]. STRIVE is recognized for its proficiency in generating adversarial safety-critical scenarios using a learned traffic model and adversarial optimization in the latent space.", + "bbox": [ + 212, + 750, + 784, + 808 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Planner. Our experiments utilize a range of different planners: 1) A rule-based planner, as implemented in STRIVE [26], which operates on the lane graph", + "bbox": [ + 212, + 810, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "W.J. Chang et al.", + "bbox": [ + 271, + 114, + 392, + 128 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "and employs the constant velocity model to predict future trajectories of non-ego vehicles, generating multiple trajectory candidates and selecting the one least likely to result in a collision; 2) a hybrid planner BITS [35]; 3) PDM-Closed [5], the winner of the 2023 nuPlan Planning Challenge; 4) a deterministic a learning-based Behavior Cloning (BC) planner, which utilizes a ResNet Encoder, followed by an MLP decoder to generate future trajectories; and 5) an Intelligent Driver Model (IDM) planner [30].", + "bbox": [ + 212, + 146, + 787, + 253 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Diffusion Model. We follow the architecture described in [39]. We represent the context using an agent-centric map and past trajectories on a rasterized map. The traffic scene is encoded using a ResNet structure [12], while the input trajectory is processed through a series of 1D temporal convolution blocks in an UNet-like architecture, as detailed in [18]. The model uses $K = 100$ diffusion steps. During the inference phase, we generate a sample of potential future trajectories for each reactive agent in a given scene. From these, we select the trajectory that yields the lowest guidance cost. This process is referred to as filtering.", + "bbox": [ + 212, + 257, + 787, + 381 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Closed-loop Simulation. The simulation framework for our experiments is built upon an open-source traffic behavior framework [35]. Within this framework, both the planner and reactive agents update their plans at a frequency of $2\\mathrm{Hz}$ .", + "bbox": [ + 212, + 383, + 787, + 429 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "6.3 Evaluation Metrics", + "text_level": 1, + "bbox": [ + 214, + 476, + 419, + 491 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Our goal is to validate that the proposed method can generate safety-critical scenarios that are both realistic and controllable. For realism assessment, in accordance with [39], we compare statistical data between simulated trajectories and actual ground trajectories. This involves calculating the Wasserstein distance between their driving profiles' normalized histograms, focusing on the mean of mean values for three properties: longitudinal acceleration, latitudinal acceleration, and jerk. To evaluate controllability, we measure metrics related to parameters we can control, specifically relative speed and time-to-collision cost between the ego and adversarial agents.", + "bbox": [ + 212, + 527, + 787, + 665 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Our method aims to evaluate the extent of control over adversarial behaviors within a single scenario. To do this, we focus on measuring collision diversity—a metric that quantifies the range of differences in collision angles, relative speeds, and collision points. By calculating the variance of these parameters within the same scenario, we can determine how diverse and controllable the adversarial behaviors are, ensuring the scenario's realism and controllability.", + "bbox": [ + 210, + 667, + 787, + 760 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "For a more nuanced understanding of realism, we analyze the collision rate - the average fractions of agents colliding, and the offroad rate - the percentage of reactive agents going off-road, both of which are considered failure rates [35]. All metrics are averaged across scenarios. For detailed definitions and metrics, see supplementary Sec. C.", + "bbox": [ + 210, + 763, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Safety-Critical Closed-Loop Traffic Simulation with Controllable Adversaries", + "bbox": [ + 220, + 113, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/47049cf1ead19db50f52b7880b8f43142b587f30616d283eb525aa4cc68440f4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetMethodCollision (%) ↑Other Offroad (%)↓Adv Offroad (%) ↓Collision Rel Speed (m/s) ↓Realism ↓Time (s) ↓
nuScenesSAFE-SIM43.21.811.4-0.120.38104.5 ± 17.7
STRIVE36.42.211.45.520.85427.2 ± 169.8
DiffScene18.211.49.016.40.52105.4 ± 22.5
nuPlanSAFE-SIM809.411.76.750.27173.4 ± 73.3
DiffScene56.714.05.0-2.810.42176.7 ± 77.5
", + "bbox": [ + 215, + 143, + 782, + 243 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/2aef0898f1ee520fdf324a459b88a9fc1d86bba8ac8906aa70d4651fed2a8578.jpg", + "table_caption": [ + "Table 2: Safety-critical Traffic Simulation. We compare our approach against STRIVE [26] and DiffScene [34] for safety-critical traffic simulation with a rule-based planner. SAFE-SIM outperforms STRIVE on all metrics and demonstrates higher collision rates and better realism than DiffScene." + ], + "table_footnote": [], + "table_body": "
PlannerEgo-Adv Coll (%) ↑Ego-Other Coll (%) ↑Adv Offroad (%) ↓Coll Speed (m/s) ↓Ego Accel (m/s2) ↓Realism ↓
BC38.837.39.02.470.540.79
IDM49.358.23.0-0.400.940.78
Lane-Graph34.337.31.52.681.620.57
BITS16.419.46.03.070.950.79
PDM-Closed26.950.71.52.731.300.86
", + "bbox": [ + 215, + 325, + 782, + 411 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 3: Safety-Critical Simulation for Different Planners. SAFE-SIM can generate diverse safety-critical scenarios tailored to different planners, including rule-based, learning-based, and hybrid planners.", + "bbox": [ + 212, + 411, + 785, + 453 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "6.4 Evaluation of Safety-Critical Traffic Simulation with Baseline Methods", + "text_level": 1, + "bbox": [ + 212, + 494, + 774, + 523 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We compared our method with STRIVE [26] and DiffScene [34], utilizing the Lane-graph-based planner from STRIVE. The evaluation focused on collision rates between the ego and the adversarial agent (\"Collision\"), adversarial agent off-road rates (\"Adv Offroad\"), other agents' off-road rates (\"Other Offroad\"), collision relative speed between the ego and adversary (\"Collision Rel Speed\"), realism of all non-ego agents (\"Realism\"), and simulation time.", + "bbox": [ + 212, + 534, + 787, + 625 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The results, presented in Tab. 2, demonstrate that our method excels in all metrics compared to STRIVE, especially in collisions and realism. Compared to DiffScene, SAFE-SIM also exhibits a higher collision rate with better realism. Note that we trained models on nuScenes and tested them in nuPlan without additional fine-tuning. As illustrated in Figure 4, the qualitative examples from the NuScenes dataset demonstrate how our framework can challenge the rule-based planner with various driving situations. See supplementary Sec. A for more qualitative examples.", + "bbox": [ + 212, + 626, + 787, + 748 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "6.5 Safety-Critical Simulation with Different Planners", + "text_level": 1, + "bbox": [ + 212, + 768, + 674, + 784 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In Tab. 3, we demonstrate our framework's ability to generate collisions across various planner types: rule-based, learning-based, and hybrid planners. Notably, the IDM planner exhibits the highest Ego-Adv collision rate. This heightened rate", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "W.J. Chang et al.", + "bbox": [ + 271, + 114, + 392, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/e371ab30bd681bf879386a53fcb7090e725083a95280146240add8ead5968911.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TTC Cost WeightTTC CostColl Speed (m/s)Coll Angle (deg)Coll Rate (%)↑Realism ↓
0.00.182.45-7.4348.20.76
1.00.212.300.4353.60.79
2.00.263.78-17.060.70.81
", + "bbox": [ + 254, + 143, + 748, + 227 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/a6d09b2643548dfbabb89b5b14ff76544552052fd1e4bf072202d06738ee21f8.jpg", + "table_caption": [ + "Table 4: Controlling Time to Collision (TTC). The table shows the impact of different TTC Cost weights on collision scenarios. Increasing the TTC Cost weight results in an increase in collision rate, suggesting a heightened challenge for the ego vehicle in avoiding collisions." + ], + "table_footnote": [], + "table_body": "
Jadv Jreg Partial DiffusionCollision MetricsDiversity
Collision (%) ↑Adv Offroad (%) ↓Realism ↓Collision Angle Var (rad) ↑Collision Rel Speed Var (m/s) ↑Collision Point Var (m) ↑
✓✓×23.913.80.582.222.991.62
✓✓✓29.014.60.573.101.965.44
✓××53.523.10.583.344.812.47
", + "bbox": [ + 215, + 309, + 782, + 412 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 5: Ablation Study on Controllability. This study examines each component of our proposed method. Partial diffusion significantly increases the variance of the collision point, resulting in a greater diversity of collision scenarios.", + "bbox": [ + 212, + 412, + 785, + 455 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "can be attributed to the IDM's focus on vehicles near the same lane, potentially overlooking other vehicles in the scene. Consequently, in scenarios with the IDM planner, our framework can induce collisions at relatively lower speeds.", + "bbox": [ + 212, + 494, + 785, + 541 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "6.6 Evaluation: Controlling Safety-Criticality", + "text_level": 1, + "bbox": [ + 212, + 561, + 602, + 578 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "A key feature of SAFE-SIM is its ability to generate controllable adversarial behaviors, offering variations not possible with previous methods.", + "bbox": [ + 212, + 584, + 785, + 614 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Controlling Time-to-Collision. We control the orientation and relative speed together using the time-to-collision (TTC) cost, as described in Sec. 5. We manipulate the scenario's safety-criticality by adjusting the relative weight of the TTC cost. To assess the impact of these adjustments, we measure the average TTC cost shortly before a collision occurs (0.5 seconds). Our observations, detailed in Tab. 4, show that increasing the TTC weight raises the TTC cost. Notably, while the relative collision speed remains fairly consistent, the collision angle shifts, indicating a greater difficulty in avoiding ego-adversary collisions. Additionally, our method can control other aspects, such as relative speed, as detailed in supplementary Sec. F.1.", + "bbox": [ + 212, + 614, + 787, + 765 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "6.7 Ablation Study on Controllability", + "text_level": 1, + "bbox": [ + 212, + 786, + 540, + 801 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We performed an ablation study to assess the influence of different guidance strategies in diffusion models on the quality of simulations. This study, detailed", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Safety-Critical Closed-Loop Traffic Simulation with Controllable Adversaries", + "bbox": [ + 220, + 113, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "in Table 5, aimed to quantify collision diversity. We compare against our baseline approach, consisting of regularized and adversarial objectives $(J_{reg} + J_{adv})$ , by predetermining the selection of adversarial agents and conducting experiments with multiple seeds (three) in our framework to evaluate collision diversity. In partial diffusion, we manipulated the trajectory proposal selection mechanism across centerlines with normal offsets of -2.0, 0.0, and 2.0.", + "bbox": [ + 212, + 146, + 787, + 236 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Effectiveness of Partial Diffusion. Table 5 reveals that partial diffusion significantly enhances both the collision point and angle diversity in comparison to the baseline approach. This underscores the capability of partial diffusion to generate a wider array of collision scenarios through various trajectory proposals, illustrating its potential to explore diverse collision dynamics effectively.", + "bbox": [ + 212, + 237, + 787, + 311 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Impact of the Regularization Term ( $J_{reg}$ ). Incorporating $J_{reg}$ results in a notable decrease in the adversarial-collision rate, highlighting the importance of the regularization term in enhancing simulation realism. Without $J_{reg}$ , the collision rate between the adversarial agent and ego vehicle increases, but the adversarial vehicle goes offroad more often, leading to scenarios that differ significantly from realistic behavior.", + "bbox": [ + 212, + 313, + 787, + 402 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6.8 Limitation and Failure Cases", + "text_level": 1, + "bbox": [ + 214, + 422, + 501, + 436 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We identified areas for improvement in SAFE-SIM (see supp. Fig. A4). In certain cases, the adversarial agent unrealistically collides with non-adversarial agents before reaching the ego agent. Additionally, some scenarios result in collisions where the ego planner is not at fault. While understanding how the ego planner can avoid such cases is important, creating more scenarios where the ego is at fault would be beneficial.", + "bbox": [ + 212, + 444, + 787, + 532 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "7 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 554, + 359, + 570 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we present a closed-loop simulation framework utilizing guided diffusion models for creating safety-critical scenarios to assess autonomous vehicle (AV) algorithms. Our research is in line with the goals of SO-TIF, focusing on how autonomous vehicles respond to dynamic scenarios like aggressive driving, underscoring our dedication to safety across diverse and unforeseeable conditions. Our framework introduces innovative guidance objectives tailored for controllable, stable, long-term safety-critical simulations. A key aspect of our method lies in its ability to vary the types of adversarial behavior within collision scenarios. By integrating adversarial objectives and partial diffusion, we enable fine-grained control over adversarial actions. This versatility enables our framework to produce a broader range of realistic and manageable scenarios, setting a new standard in adversarial scenario generation beyond the limitations of existing approaches.", + "bbox": [ + 212, + 583, + 787, + 763 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Future directions for our research include: 1) exploring the application of our framework in closed-loop policy training, and 2) developing automated methods for adjusting controllable parameters. These methods aim to facilitate the generation of diverse, long-tail scenarios. We believe this framework holds significant promise for enhancing real-world AV safety.", + "bbox": [ + 212, + 765, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "W.J. Chang et al.", + "bbox": [ + 271, + 114, + 392, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was part of W.J. Chang's summer internship at NEC Labs America, and he is also supported by the National Science Foundation Graduate Research Fellowship Program under Grant No. DGE 2146752. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation. The authors would like to thank Chih-Ling Chang for her insightful suggestions and assistance with figures and presentations.", + "bbox": [ + 212, + 176, + 787, + 282 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 306, + 321, + 321 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Abeysirigoonawardena, Y., Shkurti, F., Dudek, G.: Generating adversarial driving scenarios in high-fidelity simulators. In: 2019 International Conference on Robotics and Automation (ICRA). pp. 8271-8277. IEEE (2019)", + "2. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11621-11631 (2020)", + "3. Caesar, H., Kabzan, J., Tan, K.S., Fong, W.K., Wolff, E., Lang, A., Fletcher, L., Beijbom, O., Omari, S.: nuplan: A closed-loop ml-based planning benchmark for autonomous vehicles. arXiv preprint arXiv:2106.11810 (2021)", + "4. Cao, Y., Xiao, C., Anandkumar, A., Xu, D., Pavone, M.: Advdo: Realistic adversarial attacks for trajectory prediction. In: European Conference on Computer Vision. pp 36-52. Springer (2022)", + "5. Dauner, D., Hallgarten, M., Geiger, A., Chitta, K.: Parting with misconceptions about learning-based vehicle motion planning. In: CoRL (2023)", + "6. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in neural information processing systems 34, 8780-8794 (2021)", + "7. Ding, W., Chen, B., Li, B., Eun, K.J., Zhao, D.: Multimodal safety-critical scenarios generation for decision-making algorithms evaluation. IEEE Robotics and Automation Letters 6(2), 1551-1558 (2021). https://doi.org/10.1109/LRA.2021.3058873", + "8. Ding, W., Chen, B., Xu, M., Zhao, D.: Learning to collide: An adaptive safety-critical scenarios generating method. In: 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 2243-2250. IEEE (2020)", + "9. Ding, W., Xu, C., Arief, M., Lin, H., Li, B., Zhao, D.: A survey on safety-critical driving scenario generation—a methodological perspective. IEEE Transactions on Intelligent Transportation Systems (2023)", + "10. Gu, T., Chen, G., Li, J., Lin, C., Rao, Y., Zhou, J., Lu, J.: Stochastic trajectory prediction via motion indeterminacy diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17113-17122 (2022)", + "11. Hanselmann, N., Renz, K., Chitta, K., Bhattacharyya, A., Geiger, A.: King: Generating safety-critical driving scenarios for robust imitation via kinematics gradients In: European Conference on Computer Vision. pp. 335-352. Springer (2022)", + "12. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition In: Proceedings of the IEEE conference on computer vision and pattern recognition pp. 770-778 (2016)" + ], + "bbox": [ + 218, + 339, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Safety-Critical Closed-Loop Traffic Simulation with Controllable Adversaries", + "bbox": [ + 220, + 113, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "13. Ho, J., Chan, W., Sahara, C., Whang, J., Gao, R., Gritsenko, A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., et al.: Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303 (2022)", + "14. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020)", + "15. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)", + "16. Ho, J., Salimans, T., Gritsenko, A., Chan, W., Norouzi, M., Fleet, D.J.: Video diffusion models. arXiv:2204.03458 (2022)", + "17. Ivanovic, B., Song, G., Gilitschenski, I., Pavone, M.: trajdata: A unified interface to multiple human trajectory datasets. arXiv preprint arXiv:2307.13924 (2023)", + "18. Janner, M., Du, Y., Tenenbaum, J.B., Levine, S.: Planning with diffusion for flexible behavior synthesis. arXiv preprint arXiv:2205.09991 (2022)", + "19. Lu, J., Wong, K., Zhang, C., Suo, S., Urtasun, R.: Scenecontrol: Diffusion for controllable traffic scene generation. In: IEEE International Conference on Robotics and Automation (ICRA) (2024)", + "20. Montali, N., Lambert, J., Mougin, P., Kuefler, A., Rhinehart, N., Li, M., Gulino, C., Emrich, T., Yang, Z., Whiteson, S., et al.: The waymo open sim agents challenge. Advances in Neural Information Processing Systems 36 (2024)", + "21. Nichol, A.Q., Dhariwal, P.: Improved denoising diffusion probabilistic models. In: International Conference on Machine Learning. pp. 8162-8171. PMLR (2021)", + "22. Nishimura, H., Mercat, J., Wulfe, B., McAllister, R.T., Gaidon, A.: Rap: Risk-aware prediction for robust planning. In: Conference on Robot Learning. pp. 381-392. PMLR (2023)", + "23. Philion, J., Peng, X.B., Fidler, S.: Trajeglish: Traffic modeling as next-token prediction. In: The Twelfth International Conference on Learning Representations (2024)", + "24. Pronovost, E., Ganesina, M.R., Hendy, N., Wang, Z., Morales, A., Wang, K., Roy, N.: Scenario diffusion: Controllable driving scenario generation with diffusion. Advances in Neural Information Processing Systems 36, 68873-68894 (2023)", + "25. Rempe, D., Luo, Z., Bin Peng, X., Yuan, Y., Kitani, K., Kreis, K., Fidler, S., Litany, O.: Trace and pace: Controllable pedestrian animation via guided trajectory diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13756-13766 (2023)", + "26. Rempe, D., Philion, J., Guibas, L.J., Fidler, S., Litany, O.: Generating useful accident-prone driving scenarios via a learned traffic prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17305-17315 (2022)", + "27. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)", + "28. Suo, S., Regalado, S., Casas, S., Urtasun, R.: Trafficsim: Learning to simulate realistic multi-agent behaviors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10400-10409 (2021)", + "29. Suo, S., Wong, K., Xu, J., Tu, J., Cui, A., Casas, S., Urtasun, R.: Mixsim: A hierarchical framework for mixed reality traffic simulation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9622-9631 (2023)", + "30. Treiber, M., Hennecke, A., Helbing, D.: Congested traffic states in empirical observations and microscopic simulations. Physical review E 62(2), 1805 (2000)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "W.J. Chang et al.", + "bbox": [ + 271, + 114, + 392, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "31. Unknown: Midjourney. https://www.midjourney.com, accessed: 2023-11-16", + "32. Unknown: Openai dall-e-2. https://openai.com/product/dall-e-2, accessed: 2023-11-16", + "33. Wang, J., Pun, A., Tu, J., Manivasagam, S., Sadat, A., Casas, S., Ren, M., Urtasun, R.: Advsim: Generating safety-critical scenarios for self-driving vehicles. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9909-9918 (2021)", + "34. Xu, C., Zhao, D., Sangiovanni-Vincentelli, A., Li, B.: Diffscene: Diffusion-based safety-critical scenario generation for autonomous vehicles. In: The Second Workshop on New Frontiers in Adversarial Machine Learning (2023)", + "35. Xu, D., Chen, Y., Ivanovic, B., Pavone, M.: Bits: Bi-level imitation for traffic simulation. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 2929-2936. IEEE (2023)", + "36. Yin, Z.H., Sun, L., Sun, L., Tomizuka, M., Zhan, W.: Diverse critical interaction generation for planning and planner evaluation. In: 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 7036-7043. IEEE (2021)", + "37. Zhang, Q., Hu, S., Sun, J., Chen, Q.A., Mao, Z.M.: On adversarial robustness of trajectory prediction for autonomous vehicles. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15159-15168 (2022)", + "38. Zhong, Z., Rempe, D., Chen, Y., Ivanovic, B., Cao, Y., Xu, D., Pavone, M., Ray, B.: Language-guided traffic simulation via scene-level diffusion. arXiv preprint arXiv:2306.06344 (2023)", + "39. Zhong, Z., Rempe, D., Xu, D., Chen, Y., Veer, S., Che, T., Ray, B., Pavone, M.: Guided conditional diffusion for controllable traffic simulation. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 3560-3566. IEEE (2023)" + ], + "bbox": [ + 215, + 147, + 785, + 507 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Safety-Critical Closed-Loop Traffic Simulation with Controllable Adversaries", + "bbox": [ + 220, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Safe-Sim_ Safety-Critical Closed-Loop Traffic Simulation with Diffusion-Controllable Adversaries/548909b9-64d9-4498-bf91-808d29ce48c5_model.json b/2024/Safe-Sim_ Safety-Critical Closed-Loop Traffic Simulation with Diffusion-Controllable Adversaries/548909b9-64d9-4498-bf91-808d29ce48c5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..53a96c2b0396abceab64baa08d22681cd91a862c --- /dev/null +++ b/2024/Safe-Sim_ Safety-Critical Closed-Loop Traffic Simulation with Diffusion-Controllable Adversaries/548909b9-64d9-4498-bf91-808d29ce48c5_model.json @@ -0,0 +1,2302 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.217, + 0.141, + 0.786, + 0.184 + ], + "angle": 0, + "content": "SAFE-SIM: Safety-Critical Closed-Loop Traffic Simulation with Diffusion-Controllable Adversaries" + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.212, + 0.782, + 0.243 + ], + "angle": 0, + "content": "Wei-Jer Chang\\(^{1}\\), Francesco Pittaluga\\(^{2}\\), Masayoshi Tomizuka\\(^{1}\\), Wei Zhan\\(^{1}\\), and Manmohan Chandraker\\(^{2,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.294, + 0.254, + 0.707, + 0.269 + ], + "angle": 0, + "content": "UC Berkeley 2 NEC Labs America 3 UC San Diego" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.31, + 0.744, + 0.629 + ], + "angle": 0, + "content": "Abstract. Evaluating the performance of autonomous vehicle planning algorithms necessitates simulating long-tail safety-critical traffic scenarios. However, traditional methods for generating such scenarios often fall short in terms of controllability and realism; they also neglect the dynamics of agent interactions. To address these limitations, we introduce SAFE-SIM, a novel diffusion-based controllable closed-loop safety-critical simulation framework. Our approach yields two distinct advantages: 1) generating realistic long-tail safety-critical scenarios that closely reflect real-world conditions, and 2) providing controllable adversarial behavior for more comprehensive and interactive evaluations. We develop a novel approach to simulate safety-critical scenarios through an adversarial term in the denoising process of diffusion models, which allows an adversarial agent to challenge a planner with plausible maneuvers while all agents in the scene exhibit reactive and realistic behaviors. Furthermore, we propose novel guidance objectives and a partial diffusion process that enables users to control key aspects of the scenarios, such as the collision type and aggressiveness of the adversarial agent, while maintaining the realism of the behavior. We validate our framework empirically using the nuScenes and nuPlan datasets across multiple planners, demonstrating improvements in both realism and controllability. These findings affirm that diffusion models provide a robust and versatile foundation for safety-critical, interactive traffic simulation, extending their utility across the broader autonomous driving landscape. Project website: https://safe-sim.github.io/." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.656, + 0.377, + 0.672 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.794 + ], + "angle": 0, + "content": "A key safety feature of autonomous vehicles (AVs) is their ability to navigate near-collision events in real-world scenarios. However, these events rarely occur on roads and testing AVs in such high-risk situations on public roads is unsafe. Therefore, simulation is indispensable in the development and assessment of AVs, providing a safe and reliable means to study their safety and dependability. A critical aspect of simulation is modeling the behavior of other road users, since AVs must learn to interact with them safely." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.842 + ], + "angle": 0, + "content": "A common method of safety-critical testing of AVs involves manually designing scenarios that could potentially lead to failures, such as collisions. While this approach allows for targeted testing, it is inherently limited in scalability and" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.129 + ], + "angle": 0, + "content": "W.J. Chang et al." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.144, + 0.784, + 0.257 + ], + "angle": 0, + "content": "
MethodSafety-CriticalControllableControllable \nAdversaryEvaluate \nPlannerClosed-LoopReal-World
CTG [39]×✓××✓✓
CTG++ [38]×✓××✓✓
STRIVE [26]✓××✓✓✓
DiffScene [34]✓✓×✓××
SAFE-SIM (Ours)✓✓✓✓✓✓
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.214, + 0.258, + 0.788, + 0.342 + ], + "angle": 0, + "content": "Table 1: Comparison of methods. Our contribution is the development of a framework for (a) safety-critical (b) closed-loop (c) controllable adversarial simulations. These aspects are not concurrently present in previous frameworks. We formulate a novel partial diffusion with novel guidance functions for stable long-term simulation and are the first to enable an ego planner to be tested against controllable adversaries with varied behavior patterns." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.352, + 0.789, + 0.488 + ], + "angle": 0, + "content": "lacks the comprehensiveness required for thorough evaluation [7,8,33]. Some recent works focus on automatically generating challenging scenarios that cause planners to fail, but their emphasis has been mostly on static scenario generation rather than dynamic, closed-loop simulations. This results in a critical gap: the behavior of other agents often does not adapt or respond to the planner's actions, which is essential for a comprehensive safety evaluation. Furthermore, the results from these simulations often lack controllability, typically producing only a single adversarial outcome per scenario without the flexibility to explore a range of conditions and responses." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.49, + 0.789, + 0.611 + ], + "angle": 0, + "content": "In this work, we introduce SAFE-SIM, a closed-loop simulation framework for generating safety-critical scenarios, with a particular emphasis on controllability and realism for the behavior of agents, which allows simulations over a long-horizon as needed to evaluate AV planning algorithms (Fig. 1). Different from prior works [26,34,38,39] that primarily adhere to rule-constraint satisfaction, our approach enhances controllability by modulating adversarial vehicle behaviors within identical scenarios, thereby facilitating a broader exploration of potential outcomes. See Tab. 1 for a comprehensive comparison of these approaches." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.612, + 0.789, + 0.793 + ], + "angle": 0, + "content": "Our approach builds upon recent developments in controllable diffusion models [18, 25, 39]. Specifically, we adopt a test-time guidance to direct the denoising phase of the diffusion process, using the gradients from differentiable objectives to enhance scenario generation, enabling generation of adversarial scenarios in which and adversarial agent collides with the ego agent behaving according to specific planning policy. Additionally, we develop an novel approach, which we refer to as Partial Diffusion that introduces trajectory proposals into the diffusion process to provide a high degree of controllability over the type of collision scenario. Overall, our balanced integration of adversarial objectives with regularization during the guidance phase combined with Partial Diffusion allows for refined control over the conditions of the generated scenarios, ensuring both their realism and relevance to safety-critical testing." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.789, + 0.84 + ], + "angle": 0, + "content": "In our study, we use the nuScenes [2] and nuPlan [3] datasets to evaluate the efficacy of our method in generating safety-critical closed-loop simulations. Our results demonstrate a marked improvement in the controllability and real-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.222, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Safety-Critical Closed-Loop Traffic Simulation with Controllable Adversaries" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.149, + 0.774, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.314, + 0.788, + 0.427 + ], + "angle": 0, + "content": "Fig. 1: Overview of SAFE-SIM Framework for Controllable Safety-Critical Closed-Loop Simulation. This framework evaluates a planner within scenarios featuring multiple controllable reactive agents. These agents have two distinct roles: adversarial agents, which actively challenge the planner by exhibiting controllable adversarial behaviors such as specific collision types and levels of aggressiveness, and non-adversarial agents, which follow normal driving behavior to maintain the realism of the entire scene. Such a setup facilitates the generation of various realistic, interactive, and safety-critical scenarios, providing a thorough evaluation of the planner's capabilities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.453, + 0.788, + 0.529 + ], + "angle": 0, + "content": "ism of scenarios compared to previous adversarial scenario generation methods. Furthermore, we showcase the advantage of our proposed framework in varying the safety-criticality and collision types of scenarios. These attributes make our approach particularly well-suited for the closed-loop simulation of AVs, providing a more reliable and comprehensive framework for safety evaluation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.548, + 0.388, + 0.563 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.577, + 0.41, + 0.59 + ], + "angle": 0, + "content": "2.1 Traffic Simulation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.789, + 0.81 + ], + "angle": 0, + "content": "Traffic simulation can broadly be categorized into two main groups: heuristic-based and learning-based methods. In heuristic-based methods, agents are controlled by human-specified rules, such as the Intelligent Driver Model (IDM) [30] to follow a leading vehicle while maintaining a safe following distance. However, these methods have modeling capacity issues and may not reflect the real traffic distribution; a large domain gap limits its usage for planning evaluation. To close the gap, data-driven approaches learn from real driving datasets to imitate real-world behavior [23, 28, 29, 35]. TrafficSim [28] utilizes a trained variational autoencoder for scene-level traffic simulation, while BITS [35] combines high-level goal inference with low-level driving behavior imitation to enhance the realism of simulated driving behavior. Recently, the Waymo SimAgents challenge has focused on whether simulators can accurately represent real-world driving distributions [20]. However, little work has specifically focused on behavior simulation for generating safety-critical, long-tail scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.789, + 0.84 + ], + "angle": 0, + "content": "Diffusion models [27, 31, 32] have shown significant promise for synthetic image generation. One of the key advantages of diffusion models is controllability," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.129 + ], + "angle": 0, + "content": "W.J. Chang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.283 + ], + "angle": 0, + "content": "which can take the form of classifier [6], classifier-free [15], and reconstruction [13] guidance. Recently, controllable diffusion models have been employed for planning and traffic simulation [25, 38, 39] via guidance. We adopt trajectory diffusion models to develop a novel approach for generate safety-critical realistic traffic simulations in which an adversarial agent collides with an ego planner agent. Other works [19, 24] use diffusion with guidance or conditioning to achieve controllable scene initialization. In contrast, we emphasize closed-loop, controllable adversarial behavior simulation based on real-world data initialization rather than scene initialization." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.306, + 0.533, + 0.321 + ], + "angle": 0, + "content": "2.2 Safety-Critical Traffic Simulation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.331, + 0.788, + 0.558 + ], + "angle": 0, + "content": "Safety-critical traffic generation plays a crucial role in training and evaluating AV systems, enhancing their capability to navigate diverse real-world scenarios and enhancing robustness. Gradient-based methods that leverage back-propagation to create safety-critical scenarios have been proposed to evaluate AV prediction and planning models [4,11]. Hanselmann et al. use kinematic gradients to modify vehicle trajectories, with the goal of improving the robustness of imitation learning planners [11]. Cao et al. have developed a model with differentiable dynamics, enabling the generation of realistic adversarial trajectories for trajectory prediction models through backpropagation techniques [4]. Black-box optimization approaches include perturbing actions based on kinematic bicycle models [33] and using Bayesian Optimization to create adversarial self-driving scenarios that escalate collision risks with simulated entities [1]. Zhang et al. target trajectory prediction models via white- and black-box attacks that adversarily perturb real driving trajectories [37]. For an extensive review of this topic, we direct readers to Ding et al. [9]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.558, + 0.79, + 0.694 + ], + "angle": 0, + "content": "The field has recently seen advancements in data-driven methods for safety-critical scenario generation [26,34,36]. For instance, Xu et al. introduce a diffusion-based approach in CARLA, applying various adversarial optimization objectives to guide the diffusion process for safety-critical scenario generation [34]. Rempe et al. proposed STRIVE to utilize gradient-based adversarial optimization on the latent space, constrained by a graph-based CVAE traffic motion model, to generate realistic safety-critical scenarios for rule-based planners [26]. However, a common limitation in these approaches is the absence of closed-loop interaction, essential for accurately simulating interactive real-world driving." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.718, + 0.46, + 0.733 + ], + "angle": 0, + "content": "3 Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We consider a simulated interactive traffic scenario consisting of \\( N \\) agents; one is the ego vehicle controlled by the planner \\( \\pi \\), and the remaining \\( N - 1 \\) are reactive agents modeled by a function \\( g \\). Our objective is to create a safety-critical closed-loop collision simulation, where reactive agents demonstrate realistic, controllable behavior. Of the \\( N - 1 \\) reactive agents, one or a subset is considered the adversarial agents (denoted as agent \\( a \\)), meant to collide with the ego vehicle." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.221, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Safety-Critical Closed-Loop Traffic Simulation with Controllable Adversaries" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.283 + ], + "angle": 0, + "content": "The adversarial agent, formulated within the reactive agent model \\( g \\), is governed by an adversarial term designed (detailed in Sec. 5) to be both controllable and adversarial to the planner \\( \\pi \\). This setup allows the adversarial agent to pose direct challenges to \\( \\pi \\), testing its resilience in complex scenarios. Concurrently, the other non-adversarial agents, also controlled by \\( g \\) with varying parameters, emulate authentic, reactive behaviors, thus enriching the simulation scenario with realistic and diverse traffic conditions. The dual role of the adversarial agent and non-adversarial agents ensures that while it challenges \\( \\pi \\), the overall simulation environment plausibly represents real-world driving conditions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.788, + 0.387 + ], + "angle": 0, + "content": "At any given timestep \\( t \\), the states of the \\( N \\) vehicles are represented as \\( \\mathbf{s}_t = [\\mathbf{s}_t^1, \\dots, \\mathbf{s}_t^N] \\), where \\( \\mathbf{s}_t^i = (x_t^i, y_t^i, v_t^i, \\theta_t^i) \\) indicates the 2D position, speed, and yaw of vehicle \\( i \\). The corresponding actions for each vehicle are \\( \\mathbf{a}_t = [\\mathbf{a}_t^1, \\dots, \\mathbf{a}_t^N] \\), with \\( \\mathbf{a}_t^i = (\\dot{v}_t^i, \\dot{\\theta}_t^i) \\) representing the acceleration and yaw rate. To predict the state at the next timestep \\( t + 1 \\), a transition function \\( f \\) is used, which computes \\( \\mathbf{s}_{t+1} = f(\\mathbf{s}_t, \\mathbf{a}_t) \\) based on current state and action. We adopt unicycle dynamics as the transition function." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.388, + 0.788, + 0.495 + ], + "angle": 0, + "content": "Each agent's decision context is \\(\\mathbf{c}_t^i\\), which includes the agent-centric map \\(I^i\\) and the \\(T_{\\mathrm{hist}}\\) historical states of neighboring vehicles from time \\(t - T_{\\mathrm{hist}}\\) to \\(t\\), defined as \\(\\mathbf{s}_{t - T_{\\mathrm{hist}}:t} = \\{\\mathbf{s}_{t - T_{\\mathrm{hist}}}, \\dots, \\mathbf{s}_t\\}\\). In closed-loop traffic simulation, each agent continuously generates and updates its trajectory based on the current decision context \\(\\mathbf{c}_t^i\\). After generating a trajectory, the simulation executes the first few steps of the planned actions before updating \\(\\mathbf{c}_t^i\\) and re-planning. See Sec. 6.2 for more implementation details." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.507, + 0.788, + 0.567 + ], + "angle": 0, + "content": "Planner \\(\\pi\\) The planner \\(\\pi\\) determines the ego vehicle's future trajectory over a time horizon \\(t\\) to \\(t + T\\). The planned state sequence is denoted by \\(s_{t:t + T}^{1} = \\pi (\\mathbf{c}_{t}^{1})\\), where \\(\\pi (\\mathbf{c}_t^1)\\) processes the historical states and map data within \\(\\mathbf{c}_t^1\\) to plan future states based on the current scene context." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.579, + 0.788, + 0.701 + ], + "angle": 0, + "content": "Reactive Agents \\( g \\) The reactive agent model \\( g \\), parameterized by \\( \\theta \\), is designed to simulate the behavior of the \\( N - 1 \\) non-ego vehicles, represented by the set \\( \\{s_{t:t + T}^i\\}_{i = 2}^N \\). Each vehicle's state sequence, \\( s_{t:t + T}^i \\), is generated by \\( g_{\\theta}(\\mathbf{c}_t^i,\\psi_i) \\), which incorporates the decision context \\( \\mathbf{c}_t^i \\) and a set of control parameters \\( \\psi_{i} \\) unique to each agent. These parameters \\( \\psi_{i} \\) enable the fine-tuning of individual behaviors within the simulation. In our approach, we train the model \\( g \\) on real-world driving data to ensure the trajectories it produces are not only controllable, supporting the generation of various safety-critical scenarios, but also realistic." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.721, + 0.63, + 0.737 + ], + "angle": 0, + "content": "4 Diffusion Models for Traffic Simulation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.787, + 0.809 + ], + "angle": 0, + "content": "For closed-loop safety-critical traffic simulation, the reactive agents, especially the adversarial agent, should be 1) controllable, and 2) realistic. With recent advances in controllable diffusion models [18,25,39], we adopt trajectory diffusion models to generate realistic simulations." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "We define the model's operational trajectory as \\(\\tau\\), which comprises both action and state sequences: \\(\\tau \\coloneqq [\\tau_{a}, \\tau_{s}]\\). Specifically, \\(\\tau_{a} \\coloneqq [a_{0}, \\ldots, a_{T-1}]\\) represents the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.129 + ], + "angle": 0, + "content": "W.J. Chang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "sequence of actions, while \\(\\tau_{s} \\coloneqq [s_{1},\\ldots ,s_{T}]\\) denotes the corresponding sequence of states. Following the approach described in [39], our model predicts the action sequence \\(\\tau_{a}\\), and the state sequence \\(\\tau_{s}\\) can be derived starting from the initial state \\(s_0\\) and dynamic model \\(f\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.788, + 0.284 + ], + "angle": 0, + "content": "A diffusion model generates a trajectory by reversing a process that incrementally adds noise. Starting with an actual trajectory \\(\\tau_0\\) sampled from the data distribution \\(q(\\tau_0)\\), a sequence of increasingly noisy trajectories \\((\\tau_{1},\\tau_{2},\\dots,\\tau_{K})\\) is produced via a forward noisng process. Each trajectory \\(\\tau_{k}\\) at step \\(k\\) is generated by adding Gaussian noise parameterized by a predefined variance schedule \\(\\beta_{k}\\) [14]:" + }, + { + "type": "equation", + "bbox": [ + 0.403, + 0.296, + 0.788, + 0.336 + ], + "angle": 0, + "content": "\\[\nq \\left(\\tau_ {1: K} \\mid \\tau_ {0}\\right) := \\prod_ {k = 1} ^ {K} q \\left(\\tau_ {k} \\mid \\tau_ {k - 1}\\right), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.363, + 0.346, + 0.786, + 0.363 + ], + "angle": 0, + "content": "\\[\nq \\left(\\tau_ {k} \\mid \\tau_ {k - 1}\\right) := \\mathcal {N} \\left(\\tau_ {k}; \\sqrt {1 - \\beta_ {k}} \\tau_ {k - 1}, \\beta_ {k} \\mathbf {I}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.369, + 0.788, + 0.445 + ], + "angle": 0, + "content": "The noising process gradually obscures the data, where the final noisy version \\(q(\\tau_K)\\) approaches \\(\\mathcal{N}(\\tau_K; \\mathbf{0}, \\mathbf{I})\\). The trajectory generation process is then achieved by learning the reverse of this noising process. Given a noisy trajectory \\(\\tau_K\\), the model learns to denoise it back to \\(\\tau_0\\) through a sequence of reverse steps. Each reverse step is modeled as:" + }, + { + "type": "equation", + "bbox": [ + 0.352, + 0.457, + 0.787, + 0.473 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(\\tau_ {k - 1} \\mid \\tau_ {k}, \\mathbf {c}\\right) := \\mathcal {N} \\left(\\tau_ {k - 1}; \\mu_ {\\theta} \\left(\\tau_ {k}, k, \\mathbf {c}\\right), \\Sigma_ {k}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.483, + 0.784, + 0.542 + ], + "angle": 0, + "content": "where \\(\\theta\\) are learned functions that predict the mean \\(\\mu\\) of the reverse step, and \\(\\Sigma_{k}\\) is a fixed schedule. By iteratively applying the reverse process, the model learns a trajectory distribution, effectively generating a plausible future trajectory from a noisy start." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.544, + 0.784, + 0.589 + ], + "angle": 0, + "content": "During the trajectory prediction phase, the model estimates the final clean trajectory denoted by \\(\\hat{\\tau}_0\\). This estimated trajectory is used to compute the mean \\(\\mu\\) as described in [21]. For more details, see supplementary material Sec. D." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.611, + 0.776, + 0.629 + ], + "angle": 0, + "content": "5 Diffusion Models for Safety-Critical Traffic Simulation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.788, + 0.734 + ], + "angle": 0, + "content": "The diffusion model, once trained on realistic trajectory data, inherently reflects the behavioral patterns present in its training distribution. However, to effectively simulate and analyze safety-critical scenarios, there is a crucial need for a mechanism that allows for the controlled manipulation of agent behaviors [18,39]. This is particularly important for generating adversarial behaviors and ensuring scene consistency in simulations." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.755, + 0.471, + 0.772 + ], + "angle": 0, + "content": "5.1 Guiding Reactive Agents" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Our approach specifically introduces guidance to the sampled trajectories at each denoising step, aligning them with predefined objectives \\( J(\\tau) \\). The concept of guidance involves using the gradient of \\( J \\) to subtly perturb the predicted mean of the model at each denoising step. This process enables the generation" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.222, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Safety-Critical Closed-Loop Traffic Simulation with Controllable Adversaries" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.149, + 0.784, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.292, + 0.789, + 0.351 + ], + "angle": 0, + "content": "Fig.2: Guided Diffusion Process for the Adversarial Agent. This process optimizes the adversarial agent's trajectory using the adversarial cost function \\( J_{\\mathrm{adv}} \\) to the ego vehicle. In particular, we introduce \\( J_{\\mathrm{control}} \\) to vary the adversarial behavior. Simultaneously, it applies regularization through \\( J_{\\mathrm{reg}} \\) for maintaining realism." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.377, + 0.789, + 0.44 + ], + "angle": 0, + "content": "of trajectories that not only reflect realistic behavior but also cater to specific simulation needs, such as adversarial testing and maintaining scene consistency over extended periods. We adopt the reconstruction guidance (clean guidance) introduced from [16, 25]:" + }, + { + "type": "equation", + "bbox": [ + 0.415, + 0.448, + 0.788, + 0.466 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\tau} _ {0} = \\hat {\\tau} _ {0} - \\alpha \\Sigma_ {k} \\nabla_ {\\tau_ {k}} J (\\hat {\\tau} _ {0}) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.474, + 0.789, + 0.504 + ], + "angle": 0, + "content": "This strategy improves guidance robustness, yielding smoother, more stable trajectories without the usual numerical issues from noisy data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.505, + 0.789, + 0.55 + ], + "angle": 0, + "content": "In practice, diversifying the behavior of adversarial agents within the same scenarios is crucial for a thorough assessment of AVs. Despite the significance of this challenge, it remains largely unexplored in previous works [26,39]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.55, + 0.789, + 0.653 + ], + "angle": 0, + "content": "The loss function for the non-reactive agents, \\( J(\\tau) \\), consists of a collision term \\( J_{\\mathrm{coll}} \\), which encourages collisions between the adversarial agent and the ego agent, two control terms \\( J_{\\mathrm{v}} \\) and \\( J_{\\mathrm{ttc}} \\), which control the relative speed and time-to-collision between the ego and adversarial agent respectively, a regularization term \\( J_{\\mathrm{Gauss}} \\), which discourages collisions between the reactive agents, and a route guidance term \\( J_{\\mathrm{route}} \\), which discourages the reactive agents from going outside the road:" + }, + { + "type": "equation", + "bbox": [ + 0.34, + 0.655, + 0.788, + 0.693 + ], + "angle": 0, + "content": "\\[\nJ (\\tau) = \\rho \\underbrace {\\left(J _ {\\mathrm {c o l l}} + J _ {\\mathrm {v}} + J _ {\\mathrm {t t c}}\\right)} _ {J _ {\\mathrm {a d v}} (\\tau)} + \\underbrace {J _ {\\text {r o u t e}} + J _ {\\text {G a u s s}}} _ {J _ {\\text {r e g}} (\\tau)}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.697, + 0.789, + 0.743 + ], + "angle": 0, + "content": "where \\(\\rho\\) denotes a scalar weight that determines whether a reactive agent behaves adversarially towards the ego agent, i.e., whether it attempts to collide with the ego agent." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.763, + 0.789, + 0.794 + ], + "angle": 0, + "content": "Collision with Planner We define \\( J_{\\mathrm{coll}} \\) to encourage the collision between the adversarial agent and the ego agent, given by:" + }, + { + "type": "equation", + "bbox": [ + 0.435, + 0.803, + 0.788, + 0.845 + ], + "angle": 0, + "content": "\\[\nJ _ {\\text {c o l l}} = - \\sum_ {t = 1} ^ {T} d (t), \\tag {6}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.129 + ], + "angle": 0, + "content": "W.J. Chang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.208 + ], + "angle": 0, + "content": "where \\( d(t) \\) represents the distance between the ego and the adversarial agent at each time step of the planning horizon \\( T \\). The adversarial agent is either pre-selected based on lane proximity or dynamically selected based on the distance to the ego agent. Details are in the supplementary Sec. D.4." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.228, + 0.788, + 0.289 + ], + "angle": 0, + "content": "Safety Criticality of Collisions We control the relative speed \\( J_{v} \\) between the ego and adversary at each time step \\( (v_{t}^{1} \\) and \\( v_{t}^{a} \\)), and the time-to-collision (TTC) cost \\( J_{\\mathrm{ttc}} \\) [22] to control the safety criticality of potential collisions, with the latter given by:" + }, + { + "type": "equation", + "bbox": [ + 0.368, + 0.298, + 0.788, + 0.339 + ], + "angle": 0, + "content": "\\[\nJ _ {\\mathrm {t t c}} = \\sum_ {t = 1} ^ {T} - \\exp \\left(- \\frac {\\tilde {t} _ {\\operatorname {c o l} (t)} ^ {2}}{2 \\lambda_ {t}} - \\frac {\\tilde {d} _ {\\operatorname {c o l} (t)} ^ {2}}{2 \\lambda_ {d}}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.349, + 0.788, + 0.427 + ], + "angle": 0, + "content": "where \\(\\tilde{t}_{\\mathrm{col}(t)}\\) is the time to collision at time \\(t\\), \\(\\tilde{d}_{\\mathrm{col}(t)}\\) is the distance to collision and \\(\\lambda_t\\) and \\(\\lambda_d\\) are bandwidth parameters for time and distance. This formula uses a constant velocity assumption. Intuitively, the time-to-collision cost favors scenarios with high relative speeds and challenging collision angles for the ego vehicle to avoid. For details of \\(J_v\\) and \\(J_{\\mathrm{ttc}}\\), see supplementary Sec. C.2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.445, + 0.788, + 0.535 + ], + "angle": 0, + "content": "Route Guidance Given an agent's trajectory \\(\\tau\\) and the corresponding route \\(r\\)-the predefined path on a lane graph from its starting point to its destination—we compute the normal distance of each point \\(\\tau_{t}\\) on the trajectory to the route at each timestep. We then penalize deviations from the route that exceed a predefined margin \\(d\\). This process is captured by the following route guidance cost function:" + }, + { + "type": "equation", + "bbox": [ + 0.355, + 0.548, + 0.788, + 0.588 + ], + "angle": 0, + "content": "\\[\nJ _ {\\text {r o u t e}} (\\tau , r) = \\sum_ {t = 1} ^ {T} \\max \\left(0, \\left| d _ {n} \\left(\\tau_ {t}, r\\right) - d _ {m} \\right|\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.593, + 0.788, + 0.701 + ], + "angle": 0, + "content": "where \\( d_{n}(\\tau_{n},r) \\) denotes the normal distance from the point \\( \\tau_{t} \\) on the trajectory to the nearest point on the route \\( r \\) at timestep \\( t \\), and \\( d_{m} \\) represents the acceptable deviation margin from the route. In contrast to the off-road loss in prior studies [39], our proposed route guidance system more effectively indicates each agent's intended path, improving adherence to traffic rules as demonstrated in Sec. 6. The flexibility of route guidance supports diverse agent interactions, such as modifying routes to encourage lane changes among reactive agents [29]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.788, + 0.794 + ], + "angle": 0, + "content": "Gaussian Collision Guidance Given the trajectories of agents, we calculate the Gaussian distance for each pair of agents \\((i,j)\\) at each timestep \\(t\\) from 1 to \\(T\\). The Gaussian distance between the agents takes into account both the tangential \\((d_t)\\) and normal \\((d_n)\\) components of the projected distances. The aggregated Gaussian distance is:" + }, + { + "type": "equation", + "bbox": [ + 0.319, + 0.803, + 0.788, + 0.845 + ], + "angle": 0, + "content": "\\[\nJ _ {\\mathrm {G a u s s}} = \\sum_ {t = 1} ^ {T} \\sum_ {i, j} ^ {N} \\exp \\left(- \\frac {1}{2 \\sigma^ {2}} \\left(\\lambda \\cdot d _ {t} ^ {i j} (t) ^ {2} + d _ {n} ^ {i j} (t) ^ {2}\\right)\\right) \\tag {9}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.222, + 0.114, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Safety-Critical Closed-Loop Traffic Simulation with Controllable Adversaries" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.147, + 0.778, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.299, + 0.788, + 0.341 + ], + "angle": 0, + "content": "Fig. 3: Framework for Partial Diffusion. We generate proposals based on domain knowledge (e.g., collision types). Users can adjust noise levels to balance between user control and the model's data distribution." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.36, + 0.788, + 0.482 + ], + "angle": 0, + "content": "where \\( d_t^{ij}(t) \\) and \\( d_n^{ij}(t) \\) represent the tangential and normal distances from agent \\( j \\)'s trajectory point at time \\( t \\) to agent \\( i \\)'s heading axis, respectively, and \\( \\sigma \\) is the standard deviation for these distances. In this formulation, \\( \\lambda \\) is a scaling factor applied to the tangential distance \\( d_t^{ij}(t) \\). This approach contrasts with the disk approximation method, which primarily penalizes the Euclidean distance between agents. By accounting for both tangential and normal components, the Gaussian collision distance method significantly reduces collision rate, which we discuss in Sec. 6." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.508, + 0.637, + 0.525 + ], + "angle": 0, + "content": "5.2 Partial Diffusion: Controlling Collision Types" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.537, + 0.788, + 0.612 + ], + "angle": 0, + "content": "We introduce a novel approach through a partial diffusion process, utilizing trajectory proposals to initiate the diffusion process. This methodology enables the variation in collision types by the adversarial agent within the diffusion, tailoring the adversarial outcomes to specific evaluation needs, the results are discussed in Sec. 6.6." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.788, + 0.75 + ], + "angle": 0, + "content": "Figure 3 illustrates our framework, which is divided into three main steps to generate trajectory proposals for various collision scenarios. First, we create initial trajectory proposals \\((\\tau_0)\\) aimed at capturing different types of collisions. The next critical step involves setting the partial diffusion ratio \\(\\gamma\\), which defines the specific point in the process, \\(k_{p} = \\gamma \\cdot K\\), at which we start modifying the trajectory. Starting from step \\(k_{p}\\), we adjust the trajectory by adding a precise level of Gaussian noise \\(\\epsilon \\sim N(0,I)\\): \\(\\hat{\\tau}_{k_p} = \\sqrt{\\bar{\\alpha}_{k_p}}\\tau_0 + \\sqrt{1 - \\bar{\\alpha}_{k_p}}\\epsilon\\). The final stages include removing noise and using guided diffusion for the rest of the \\(k_{p}\\) steps to refine the trajectory into a realistic path that suits our collision scenario goals." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.842 + ], + "angle": 0, + "content": "To generate the trajectory proposals, we develop a rule-based approach in which we first identify the centerlines of the ego and adversarial agent and then search for potential intersections of their respective centerlines. If such an intersection exists, we generate the proposals by selecting an acceleration value and lateral offset from the centerline that is likely to cause the desired collision type based on the projected plan of the ego agent. Note that the trajectory" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.129 + ], + "angle": 0, + "content": "W.J. Chang et al." + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.149, + 0.406, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.262, + 0.33, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.333, + 0.263, + 0.358, + 0.274 + ], + "angle": 0, + "content": "Ego" + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.263, + 0.385, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.39, + 0.264, + 0.424, + 0.275 + ], + "angle": 0, + "content": "Adv" + }, + { + "type": "image", + "bbox": [ + 0.432, + 0.15, + 0.588, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.431, + 0.263, + 0.466, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.471, + 0.264, + 0.518, + 0.275 + ], + "angle": 0, + "content": "Non-Adv" + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.263, + 0.598, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.601, + 0.15, + 0.772, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.606, + 0.264, + 0.706, + 0.275 + ], + "angle": 0, + "content": "Trajectory Proposal" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.292, + 0.789, + 0.305 + ], + "angle": 0, + "content": "Fig. 4: Partial Diffusion Results of Rule-Based Planner on NuScenes Dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.306, + 0.786, + 0.349 + ], + "angle": 0, + "content": "The safety-critical scenarios show the framework's ability to create realistic and challenging situations, varying collision types based on different trajectory proposals. The gradient lines reflects the planned trajectories in the next 3.2 seconds." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.367, + 0.784, + 0.397 + ], + "angle": 0, + "content": "proposals are updated in a closed-loop manner to account for the interaction between the ego and the adversarial agent." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.398, + 0.785, + 0.456 + ], + "angle": 0, + "content": "This method allows for precise control over the diffusion trajectory, enabling adversarial agents to create customized collision scenarios. Users can adjust \\(\\gamma\\) to fine-tune the balance between explicit control and the model's trained data distribution." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.482, + 0.376, + 0.499 + ], + "angle": 0, + "content": "6 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.515, + 0.785, + 0.56 + ], + "angle": 0, + "content": "We validate the efficacy of our proposed framework via experiments with real-world driving data. Our results demonstrate that the framework can generate realistic and controllable adversarial behavior to challenge the planner." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.584, + 0.327, + 0.597 + ], + "angle": 0, + "content": "6.1 Dataset" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.61, + 0.785, + 0.699 + ], + "angle": 0, + "content": "We conduct our experiments on two large-scale real-world driving datasets: nuScenes [2], which consists of 5.5 hours of driving data from two cities, and nuPlan, which consists of 1500 hours of driving data from four cities. We train the model on scenes from the nuScenes train split and evaluate it on the scenes from the nuScenes validation splits and nuPlan mini validation splits. We focus on vehicle-to-vehicle interactions." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.724, + 0.457, + 0.739 + ], + "angle": 0, + "content": "6.2 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.785, + 0.809 + ], + "angle": 0, + "content": "Baselines. We compare our approach against STRIVE [26] using their open-source implementation and our re-implementation of DiffScene [34]. STRIVE is recognized for its proficiency in generating adversarial safety-critical scenarios using a learned traffic model and adversarial optimization in the latent space." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.811, + 0.786, + 0.84 + ], + "angle": 0, + "content": "Planner. Our experiments utilize a range of different planners: 1) A rule-based planner, as implemented in STRIVE [26], which operates on the lane graph" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.221, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Safety-Critical Closed-Loop Traffic Simulation with Controllable Adversaries" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.255 + ], + "angle": 0, + "content": "and employs the constant velocity model to predict future trajectories of non-ego vehicles, generating multiple trajectory candidates and selecting the one least likely to result in a collision; 2) a hybrid planner BITS [35]; 3) PDM-Closed [5], the winner of the 2023 nuPlan Planning Challenge; 4) a deterministic a learning-based Behavior Cloning (BC) planner, which utilizes a ResNet Encoder, followed by an MLP decoder to generate future trajectories; and 5) an Intelligent Driver Model (IDM) planner [30]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.258, + 0.788, + 0.382 + ], + "angle": 0, + "content": "Diffusion Model. We follow the architecture described in [39]. We represent the context using an agent-centric map and past trajectories on a rasterized map. The traffic scene is encoded using a ResNet structure [12], while the input trajectory is processed through a series of 1D temporal convolution blocks in an UNet-like architecture, as detailed in [18]. The model uses \\( K = 100 \\) diffusion steps. During the inference phase, we generate a sample of potential future trajectories for each reactive agent in a given scene. From these, we select the trajectory that yields the lowest guidance cost. This process is referred to as filtering." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.384, + 0.788, + 0.43 + ], + "angle": 0, + "content": "Closed-loop Simulation. The simulation framework for our experiments is built upon an open-source traffic behavior framework [35]. Within this framework, both the planner and reactive agents update their plans at a frequency of \\(2\\mathrm{Hz}\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.477, + 0.421, + 0.492 + ], + "angle": 0, + "content": "6.3 Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.528, + 0.788, + 0.666 + ], + "angle": 0, + "content": "Our goal is to validate that the proposed method can generate safety-critical scenarios that are both realistic and controllable. For realism assessment, in accordance with [39], we compare statistical data between simulated trajectories and actual ground trajectories. This involves calculating the Wasserstein distance between their driving profiles' normalized histograms, focusing on the mean of mean values for three properties: longitudinal acceleration, latitudinal acceleration, and jerk. To evaluate controllability, we measure metrics related to parameters we can control, specifically relative speed and time-to-collision cost between the ego and adversarial agents." + }, + { + "type": "text", + "bbox": [ + 0.212, + 0.669, + 0.788, + 0.761 + ], + "angle": 0, + "content": "Our method aims to evaluate the extent of control over adversarial behaviors within a single scenario. To do this, we focus on measuring collision diversity—a metric that quantifies the range of differences in collision angles, relative speeds, and collision points. By calculating the variance of these parameters within the same scenario, we can determine how diverse and controllable the adversarial behaviors are, ensuring the scenario's realism and controllability." + }, + { + "type": "text", + "bbox": [ + 0.212, + 0.765, + 0.788, + 0.842 + ], + "angle": 0, + "content": "For a more nuanced understanding of realism, we analyze the collision rate - the average fractions of agents colliding, and the offroad rate - the percentage of reactive agents going off-road, both of which are considered failure rates [35]. All metrics are averaged across scenarios. For detailed definitions and metrics, see supplementary Sec. C." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.129 + ], + "angle": 0, + "content": "W.J. Chang et al." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.144, + 0.784, + 0.244 + ], + "angle": 0, + "content": "
DatasetMethodCollision (%) ↑Other Offroad (%)↓Adv Offroad (%) ↓Collision Rel Speed (m/s) ↓Realism ↓Time (s) ↓
nuScenesSAFE-SIM43.21.811.4-0.120.38104.5 ± 17.7
STRIVE36.42.211.45.520.85427.2 ± 169.8
DiffScene18.211.49.016.40.52105.4 ± 22.5
nuPlanSAFE-SIM809.411.76.750.27173.4 ± 73.3
DiffScene56.714.05.0-2.810.42176.7 ± 77.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.245, + 0.785, + 0.299 + ], + "angle": 0, + "content": "Table 2: Safety-critical Traffic Simulation. We compare our approach against STRIVE [26] and DiffScene [34] for safety-critical traffic simulation with a rule-based planner. SAFE-SIM outperforms STRIVE on all metrics and demonstrates higher collision rates and better realism than DiffScene." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.326, + 0.784, + 0.412 + ], + "angle": 0, + "content": "
PlannerEgo-Adv Coll (%) ↑Ego-Other Coll (%) ↑Adv Offroad (%) ↓Coll Speed (m/s) ↓Ego Accel (m/s2) ↓Realism ↓
BC38.837.39.02.470.540.79
IDM49.358.23.0-0.400.940.78
Lane-Graph34.337.31.52.681.620.57
BITS16.419.46.03.070.950.79
PDM-Closed26.950.71.52.731.300.86
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.412, + 0.787, + 0.454 + ], + "angle": 0, + "content": "Table 3: Safety-Critical Simulation for Different Planners. SAFE-SIM can generate diverse safety-critical scenarios tailored to different planners, including rule-based, learning-based, and hybrid planners." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.495, + 0.776, + 0.525 + ], + "angle": 0, + "content": "6.4 Evaluation of Safety-Critical Traffic Simulation with Baseline Methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.535, + 0.788, + 0.626 + ], + "angle": 0, + "content": "We compared our method with STRIVE [26] and DiffScene [34], utilizing the Lane-graph-based planner from STRIVE. The evaluation focused on collision rates between the ego and the adversarial agent (\"Collision\"), adversarial agent off-road rates (\"Adv Offroad\"), other agents' off-road rates (\"Other Offroad\"), collision relative speed between the ego and adversary (\"Collision Rel Speed\"), realism of all non-ego agents (\"Realism\"), and simulation time." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.627, + 0.788, + 0.749 + ], + "angle": 0, + "content": "The results, presented in Tab. 2, demonstrate that our method excels in all metrics compared to STRIVE, especially in collisions and realism. Compared to DiffScene, SAFE-SIM also exhibits a higher collision rate with better realism. Note that we trained models on nuScenes and tested them in nuPlan without additional fine-tuning. As illustrated in Figure 4, the qualitative examples from the NuScenes dataset demonstrate how our framework can challenge the rule-based planner with various driving situations. See supplementary Sec. A for more qualitative examples." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.77, + 0.676, + 0.785 + ], + "angle": 0, + "content": "6.5 Safety-Critical Simulation with Different Planners" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "In Tab. 3, we demonstrate our framework's ability to generate collisions across various planner types: rule-based, learning-based, and hybrid planners. Notably, the IDM planner exhibits the highest Ego-Adv collision rate. This heightened rate" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.222, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Safety-Critical Closed-Loop Traffic Simulation with Controllable Adversaries" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table", + "bbox": [ + 0.256, + 0.144, + 0.749, + 0.228 + ], + "angle": 0, + "content": "
TTC Cost WeightTTC CostColl Speed (m/s)Coll Angle (deg)Coll Rate (%)↑Realism ↓
0.00.182.45-7.4348.20.76
1.00.212.300.4353.60.79
2.00.263.78-17.060.70.81
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.229, + 0.788, + 0.284 + ], + "angle": 0, + "content": "Table 4: Controlling Time to Collision (TTC). The table shows the impact of different TTC Cost weights on collision scenarios. Increasing the TTC Cost weight results in an increase in collision rate, suggesting a heightened challenge for the ego vehicle in avoiding collisions." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.31, + 0.784, + 0.414 + ], + "angle": 0, + "content": "
Jadv Jreg Partial DiffusionCollision MetricsDiversity
Collision (%) ↑Adv Offroad (%) ↓Realism ↓Collision Angle Var (rad) ↑Collision Rel Speed Var (m/s) ↑Collision Point Var (m) ↑
✓✓×23.913.80.582.222.991.62
✓✓✓29.014.60.573.101.965.44
✓××53.523.10.583.344.812.47
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.414, + 0.786, + 0.456 + ], + "angle": 0, + "content": "Table 5: Ablation Study on Controllability. This study examines each component of our proposed method. Partial diffusion significantly increases the variance of the collision point, resulting in a greater diversity of collision scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.495, + 0.786, + 0.542 + ], + "angle": 0, + "content": "can be attributed to the IDM's focus on vehicles near the same lane, potentially overlooking other vehicles in the scene. Consequently, in scenarios with the IDM planner, our framework can induce collisions at relatively lower speeds." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.562, + 0.603, + 0.579 + ], + "angle": 0, + "content": "6.6 Evaluation: Controlling Safety-Criticality" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.585, + 0.786, + 0.615 + ], + "angle": 0, + "content": "A key feature of SAFE-SIM is its ability to generate controllable adversarial behaviors, offering variations not possible with previous methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.616, + 0.788, + 0.766 + ], + "angle": 0, + "content": "Controlling Time-to-Collision. We control the orientation and relative speed together using the time-to-collision (TTC) cost, as described in Sec. 5. We manipulate the scenario's safety-criticality by adjusting the relative weight of the TTC cost. To assess the impact of these adjustments, we measure the average TTC cost shortly before a collision occurs (0.5 seconds). Our observations, detailed in Tab. 4, show that increasing the TTC weight raises the TTC cost. Notably, while the relative collision speed remains fairly consistent, the collision angle shifts, indicating a greater difficulty in avoiding ego-adversary collisions. Additionally, our method can control other aspects, such as relative speed, as detailed in supplementary Sec. F.1." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.787, + 0.542, + 0.803 + ], + "angle": 0, + "content": "6.7 Ablation Study on Controllability" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "We performed an ablation study to assess the influence of different guidance strategies in diffusion models on the quality of simulations. This study, detailed" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.129 + ], + "angle": 0, + "content": "W.J. Chang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.237 + ], + "angle": 0, + "content": "in Table 5, aimed to quantify collision diversity. We compare against our baseline approach, consisting of regularized and adversarial objectives \\((J_{reg} + J_{adv})\\), by predetermining the selection of adversarial agents and conducting experiments with multiple seeds (three) in our framework to evaluate collision diversity. In partial diffusion, we manipulated the trajectory proposal selection mechanism across centerlines with normal offsets of -2.0, 0.0, and 2.0." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.788, + 0.313 + ], + "angle": 0, + "content": "Effectiveness of Partial Diffusion. Table 5 reveals that partial diffusion significantly enhances both the collision point and angle diversity in comparison to the baseline approach. This underscores the capability of partial diffusion to generate a wider array of collision scenarios through various trajectory proposals, illustrating its potential to explore diverse collision dynamics effectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.314, + 0.788, + 0.404 + ], + "angle": 0, + "content": "Impact of the Regularization Term (\\(J_{reg}\\)). Incorporating \\(J_{reg}\\) results in a notable decrease in the adversarial-collision rate, highlighting the importance of the regularization term in enhancing simulation realism. Without \\(J_{reg}\\), the collision rate between the adversarial agent and ego vehicle increases, but the adversarial vehicle goes offroad more often, leading to scenarios that differ significantly from realistic behavior." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.423, + 0.502, + 0.437 + ], + "angle": 0, + "content": "6.8 Limitation and Failure Cases" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.445, + 0.788, + 0.534 + ], + "angle": 0, + "content": "We identified areas for improvement in SAFE-SIM (see supp. Fig. A4). In certain cases, the adversarial agent unrealistically collides with non-adversarial agents before reaching the ego agent. Additionally, some scenarios result in collisions where the ego planner is not at fault. While understanding how the ego planner can avoid such cases is important, creating more scenarios where the ego is at fault would be beneficial." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.555, + 0.36, + 0.571 + ], + "angle": 0, + "content": "7 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.788, + 0.765 + ], + "angle": 0, + "content": "In this work, we present a closed-loop simulation framework utilizing guided diffusion models for creating safety-critical scenarios to assess autonomous vehicle (AV) algorithms. Our research is in line with the goals of SO-TIF, focusing on how autonomous vehicles respond to dynamic scenarios like aggressive driving, underscoring our dedication to safety across diverse and unforeseeable conditions. Our framework introduces innovative guidance objectives tailored for controllable, stable, long-term safety-critical simulations. A key aspect of our method lies in its ability to vary the types of adversarial behavior within collision scenarios. By integrating adversarial objectives and partial diffusion, we enable fine-grained control over adversarial actions. This versatility enables our framework to produce a broader range of realistic and manageable scenarios, setting a new standard in adversarial scenario generation beyond the limitations of existing approaches." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Future directions for our research include: 1) exploring the application of our framework in closed-loop policy training, and 2) developing automated methods for adjusting controllable parameters. These methods aim to facilitate the generation of diverse, long-tail scenarios. We believe this framework holds significant promise for enhancing real-world AV safety." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.221, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Safety-Critical Closed-Loop Traffic Simulation with Controllable Adversaries" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.284 + ], + "angle": 0, + "content": "This work was part of W.J. Chang's summer internship at NEC Labs America, and he is also supported by the National Science Foundation Graduate Research Fellowship Program under Grant No. DGE 2146752. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation. The authors would like to thank Chih-Ling Chang for her insightful suggestions and assistance with figures and presentations." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.308, + 0.323, + 0.323 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.34, + 0.785, + 0.38 + ], + "angle": 0, + "content": "1. Abeysirigoonawardena, Y., Shkurti, F., Dudek, G.: Generating adversarial driving scenarios in high-fidelity simulators. In: 2019 International Conference on Robotics and Automation (ICRA). pp. 8271-8277. IEEE (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.383, + 0.785, + 0.435 + ], + "angle": 0, + "content": "2. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11621-11631 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.438, + 0.785, + 0.478 + ], + "angle": 0, + "content": "3. Caesar, H., Kabzan, J., Tan, K.S., Fong, W.K., Wolff, E., Lang, A., Fletcher, L., Beijbom, O., Omari, S.: nuplan: A closed-loop ml-based planning benchmark for autonomous vehicles. arXiv preprint arXiv:2106.11810 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.48, + 0.785, + 0.52 + ], + "angle": 0, + "content": "4. Cao, Y., Xiao, C., Anandkumar, A., Xu, D., Pavone, M.: Advdo: Realistic adversarial attacks for trajectory prediction. In: European Conference on Computer Vision. pp 36-52. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.522, + 0.785, + 0.548 + ], + "angle": 0, + "content": "5. Dauner, D., Hallgarten, M., Geiger, A., Chitta, K.: Parting with misconceptions about learning-based vehicle motion planning. In: CoRL (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.55, + 0.785, + 0.576 + ], + "angle": 0, + "content": "6. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in neural information processing systems 34, 8780-8794 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.578, + 0.787, + 0.629 + ], + "angle": 0, + "content": "7. Ding, W., Chen, B., Li, B., Eun, K.J., Zhao, D.: Multimodal safety-critical scenarios generation for decision-making algorithms evaluation. IEEE Robotics and Automation Letters 6(2), 1551-1558 (2021). https://doi.org/10.1109/LRA.2021.3058873" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.632, + 0.785, + 0.673 + ], + "angle": 0, + "content": "8. Ding, W., Chen, B., Xu, M., Zhao, D.: Learning to collide: An adaptive safety-critical scenarios generating method. In: 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 2243-2250. IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.675, + 0.785, + 0.714 + ], + "angle": 0, + "content": "9. Ding, W., Xu, C., Arief, M., Lin, H., Li, B., Zhao, D.: A survey on safety-critical driving scenario generation—a methodological perspective. IEEE Transactions on Intelligent Transportation Systems (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.716, + 0.785, + 0.756 + ], + "angle": 0, + "content": "10. Gu, T., Chen, G., Li, J., Lin, C., Rao, Y., Zhou, J., Lu, J.: Stochastic trajectory prediction via motion indeterminacy diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17113-17122 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.758, + 0.785, + 0.798 + ], + "angle": 0, + "content": "11. Hanselmann, N., Renz, K., Chitta, K., Bhattacharyya, A., Geiger, A.: King: Generating safety-critical driving scenarios for robust imitation via kinematics gradients In: European Conference on Computer Vision. pp. 335-352. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "12. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition In: Proceedings of the IEEE conference on computer vision and pattern recognition pp. 770-778 (2016)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.34, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.129 + ], + "angle": 0, + "content": "W.J. Chang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "13. Ho, J., Chan, W., Sahara, C., Whang, J., Gao, R., Gritsenko, A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., et al.: Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.786, + 0.217 + ], + "angle": 0, + "content": "14. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.786, + 0.244 + ], + "angle": 0, + "content": "15. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.786, + 0.272 + ], + "angle": 0, + "content": "16. Ho, J., Salimans, T., Gritsenko, A., Chan, W., Norouzi, M., Fleet, D.J.: Video diffusion models. arXiv:2204.03458 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.786, + 0.3 + ], + "angle": 0, + "content": "17. Ivanovic, B., Song, G., Gilitschenski, I., Pavone, M.: trajdata: A unified interface to multiple human trajectory datasets. arXiv preprint arXiv:2307.13924 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.786, + 0.328 + ], + "angle": 0, + "content": "18. Janner, M., Du, Y., Tenenbaum, J.B., Levine, S.: Planning with diffusion for flexible behavior synthesis. arXiv preprint arXiv:2205.09991 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.786, + 0.37 + ], + "angle": 0, + "content": "19. Lu, J., Wong, K., Zhang, C., Suo, S., Urtasun, R.: Scenecontrol: Diffusion for controllable traffic scene generation. In: IEEE International Conference on Robotics and Automation (ICRA) (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.787, + 0.412 + ], + "angle": 0, + "content": "20. Montali, N., Lambert, J., Mougin, P., Kuefler, A., Rhinehart, N., Li, M., Gulino, C., Emrich, T., Yang, Z., Whiteson, S., et al.: The waymo open sim agents challenge. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.787, + 0.439 + ], + "angle": 0, + "content": "21. Nichol, A.Q., Dhariwal, P.: Improved denoising diffusion probabilistic models. In: International Conference on Machine Learning. pp. 8162-8171. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.787, + 0.481 + ], + "angle": 0, + "content": "22. Nishimura, H., Mercat, J., Wulfe, B., McAllister, R.T., Gaidon, A.: Rap: Risk-aware prediction for robust planning. In: Conference on Robot Learning. pp. 381-392. PMLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.787, + 0.522 + ], + "angle": 0, + "content": "23. Philion, J., Peng, X.B., Fidler, S.: Trajeglish: Traffic modeling as next-token prediction. In: The Twelfth International Conference on Learning Representations (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.787, + 0.563 + ], + "angle": 0, + "content": "24. Pronovost, E., Ganesina, M.R., Hendy, N., Wang, Z., Morales, A., Wang, K., Roy, N.: Scenario diffusion: Controllable driving scenario generation with diffusion. Advances in Neural Information Processing Systems 36, 68873-68894 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.787, + 0.619 + ], + "angle": 0, + "content": "25. Rempe, D., Luo, Z., Bin Peng, X., Yuan, Y., Kitani, K., Kreis, K., Fidler, S., Litany, O.: Trace and pace: Controllable pedestrian animation via guided trajectory diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13756-13766 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.787, + 0.674 + ], + "angle": 0, + "content": "26. Rempe, D., Philion, J., Guibas, L.J., Fidler, S., Litany, O.: Generating useful accident-prone driving scenarios via a learned traffic prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17305-17315 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.787, + 0.716 + ], + "angle": 0, + "content": "27. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.787, + 0.758 + ], + "angle": 0, + "content": "28. Suo, S., Regalado, S., Casas, S., Urtasun, R.: Trafficsim: Learning to simulate realistic multi-agent behaviors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10400-10409 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.787, + 0.813 + ], + "angle": 0, + "content": "29. Suo, S., Wong, K., Xu, J., Tu, J., Cui, A., Casas, S., Urtasun, R.: Mixsim: A hierarchical framework for mixed reality traffic simulation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9622-9631 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "30. Treiber, M., Hennecke, A., Helbing, D.: Congested traffic states in empirical observations and microscopic simulations. Physical review E 62(2), 1805 (2000)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.222, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Safety-Critical Closed-Loop Traffic Simulation with Controllable Adversaries" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.148, + 0.747, + 0.161 + ], + "angle": 0, + "content": "31. Unknown: Midjourney. https://www.midjourney.com, accessed: 2023-11-16" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.162, + 0.787, + 0.188 + ], + "angle": 0, + "content": "32. Unknown: Openai dall-e-2. https://openai.com/product/dall-e-2, accessed: 2023-11-16" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.189, + 0.787, + 0.244 + ], + "angle": 0, + "content": "33. Wang, J., Pun, A., Tu, J., Manivasagam, S., Sadat, A., Casas, S., Ren, M., Urtasun, R.: Advsim: Generating safety-critical scenarios for self-driving vehicles. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9909-9918 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.786, + 0.286 + ], + "angle": 0, + "content": "34. Xu, C., Zhao, D., Sangiovanni-Vincentelli, A., Li, B.: Diffscene: Diffusion-based safety-critical scenario generation for autonomous vehicles. In: The Second Workshop on New Frontiers in Adversarial Machine Learning (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.786, + 0.327 + ], + "angle": 0, + "content": "35. Xu, D., Chen, Y., Ivanovic, B., Pavone, M.: Bits: Bi-level imitation for traffic simulation. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 2929-2936. IEEE (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.786, + 0.37 + ], + "angle": 0, + "content": "36. Yin, Z.H., Sun, L., Sun, L., Tomizuka, M., Zhan, W.: Diverse critical interaction generation for planning and planner evaluation. In: 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 7036-7043. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.786, + 0.411 + ], + "angle": 0, + "content": "37. Zhang, Q., Hu, S., Sun, J., Chen, Q.A., Mao, Z.M.: On adversarial robustness of trajectory prediction for autonomous vehicles. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15159-15168 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.786, + 0.452 + ], + "angle": 0, + "content": "38. Zhong, Z., Rempe, D., Chen, Y., Ivanovic, B., Cao, Y., Xu, D., Pavone, M., Ray, B.: Language-guided traffic simulation via scene-level diffusion. arXiv preprint arXiv:2306.06344 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.787, + 0.508 + ], + "angle": 0, + "content": "39. Zhong, Z., Rempe, D., Xu, D., Chen, Y., Veer, S., Che, T., Ray, B., Pavone, M.: Guided conditional diffusion for controllable traffic simulation. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 3560-3566. IEEE (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.508 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Safe-Sim_ Safety-Critical Closed-Loop Traffic Simulation with Diffusion-Controllable Adversaries/548909b9-64d9-4498-bf91-808d29ce48c5_origin.pdf b/2024/Safe-Sim_ Safety-Critical Closed-Loop Traffic Simulation with Diffusion-Controllable Adversaries/548909b9-64d9-4498-bf91-808d29ce48c5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..fff38b7fc07a510194e1c6506c320ee0f6b7ffdf --- /dev/null +++ b/2024/Safe-Sim_ Safety-Critical Closed-Loop Traffic Simulation with Diffusion-Controllable Adversaries/548909b9-64d9-4498-bf91-808d29ce48c5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bde7f7cb70042e4146f674190d95c4f0c23ec4f7bb6b7cbac3d266ad546a95d0 +size 1888786 diff --git a/2024/Safe-Sim_ Safety-Critical Closed-Loop Traffic Simulation with Diffusion-Controllable Adversaries/full.md b/2024/Safe-Sim_ Safety-Critical Closed-Loop Traffic Simulation with Diffusion-Controllable Adversaries/full.md new file mode 100644 index 0000000000000000000000000000000000000000..58353aba9688cbf8055ded56e3a30f73b6a9bb15 --- /dev/null +++ b/2024/Safe-Sim_ Safety-Critical Closed-Loop Traffic Simulation with Diffusion-Controllable Adversaries/full.md @@ -0,0 +1,317 @@ +# SAFE-SIM: Safety-Critical Closed-Loop Traffic Simulation with Diffusion-Controllable Adversaries + +Wei-Jer Chang $^{1}$ , Francesco Pittaluga $^{2}$ , Masayoshi Tomizuka $^{1}$ , Wei Zhan $^{1}$ , and Manmohan Chandraker $^{2,3}$ + +UC Berkeley 2 NEC Labs America 3 UC San Diego + +Abstract. Evaluating the performance of autonomous vehicle planning algorithms necessitates simulating long-tail safety-critical traffic scenarios. However, traditional methods for generating such scenarios often fall short in terms of controllability and realism; they also neglect the dynamics of agent interactions. To address these limitations, we introduce SAFE-SIM, a novel diffusion-based controllable closed-loop safety-critical simulation framework. Our approach yields two distinct advantages: 1) generating realistic long-tail safety-critical scenarios that closely reflect real-world conditions, and 2) providing controllable adversarial behavior for more comprehensive and interactive evaluations. We develop a novel approach to simulate safety-critical scenarios through an adversarial term in the denoising process of diffusion models, which allows an adversarial agent to challenge a planner with plausible maneuvers while all agents in the scene exhibit reactive and realistic behaviors. Furthermore, we propose novel guidance objectives and a partial diffusion process that enables users to control key aspects of the scenarios, such as the collision type and aggressiveness of the adversarial agent, while maintaining the realism of the behavior. We validate our framework empirically using the nuScenes and nuPlan datasets across multiple planners, demonstrating improvements in both realism and controllability. These findings affirm that diffusion models provide a robust and versatile foundation for safety-critical, interactive traffic simulation, extending their utility across the broader autonomous driving landscape. Project website: https://safe-sim.github.io/. + +# 1 Introduction + +A key safety feature of autonomous vehicles (AVs) is their ability to navigate near-collision events in real-world scenarios. However, these events rarely occur on roads and testing AVs in such high-risk situations on public roads is unsafe. Therefore, simulation is indispensable in the development and assessment of AVs, providing a safe and reliable means to study their safety and dependability. A critical aspect of simulation is modeling the behavior of other road users, since AVs must learn to interact with them safely. + +A common method of safety-critical testing of AVs involves manually designing scenarios that could potentially lead to failures, such as collisions. While this approach allows for targeted testing, it is inherently limited in scalability and + +
MethodSafety-CriticalControllableControllable +AdversaryEvaluate +PlannerClosed-LoopReal-World
CTG [39]×✓××✓✓
CTG++ [38]×✓××✓✓
STRIVE [26]✓××✓✓✓
DiffScene [34]✓✓×✓××
SAFE-SIM (Ours)✓✓✓✓✓✓
+ +Table 1: Comparison of methods. Our contribution is the development of a framework for (a) safety-critical (b) closed-loop (c) controllable adversarial simulations. These aspects are not concurrently present in previous frameworks. We formulate a novel partial diffusion with novel guidance functions for stable long-term simulation and are the first to enable an ego planner to be tested against controllable adversaries with varied behavior patterns. + +lacks the comprehensiveness required for thorough evaluation [7,8,33]. Some recent works focus on automatically generating challenging scenarios that cause planners to fail, but their emphasis has been mostly on static scenario generation rather than dynamic, closed-loop simulations. This results in a critical gap: the behavior of other agents often does not adapt or respond to the planner's actions, which is essential for a comprehensive safety evaluation. Furthermore, the results from these simulations often lack controllability, typically producing only a single adversarial outcome per scenario without the flexibility to explore a range of conditions and responses. + +In this work, we introduce SAFE-SIM, a closed-loop simulation framework for generating safety-critical scenarios, with a particular emphasis on controllability and realism for the behavior of agents, which allows simulations over a long-horizon as needed to evaluate AV planning algorithms (Fig. 1). Different from prior works [26,34,38,39] that primarily adhere to rule-constraint satisfaction, our approach enhances controllability by modulating adversarial vehicle behaviors within identical scenarios, thereby facilitating a broader exploration of potential outcomes. See Tab. 1 for a comprehensive comparison of these approaches. + +Our approach builds upon recent developments in controllable diffusion models [18, 25, 39]. Specifically, we adopt a test-time guidance to direct the denoising phase of the diffusion process, using the gradients from differentiable objectives to enhance scenario generation, enabling generation of adversarial scenarios in which and adversarial agent collides with the ego agent behaving according to specific planning policy. Additionally, we develop an novel approach, which we refer to as Partial Diffusion that introduces trajectory proposals into the diffusion process to provide a high degree of controllability over the type of collision scenario. Overall, our balanced integration of adversarial objectives with regularization during the guidance phase combined with Partial Diffusion allows for refined control over the conditions of the generated scenarios, ensuring both their realism and relevance to safety-critical testing. + +In our study, we use the nuScenes [2] and nuPlan [3] datasets to evaluate the efficacy of our method in generating safety-critical closed-loop simulations. Our results demonstrate a marked improvement in the controllability and real- + +![](images/a4cfa613335605d7fccd1ea9eb1f093ea3313a5383d54d4400afd0f7241febda.jpg) +Fig. 1: Overview of SAFE-SIM Framework for Controllable Safety-Critical Closed-Loop Simulation. This framework evaluates a planner within scenarios featuring multiple controllable reactive agents. These agents have two distinct roles: adversarial agents, which actively challenge the planner by exhibiting controllable adversarial behaviors such as specific collision types and levels of aggressiveness, and non-adversarial agents, which follow normal driving behavior to maintain the realism of the entire scene. Such a setup facilitates the generation of various realistic, interactive, and safety-critical scenarios, providing a thorough evaluation of the planner's capabilities. + +ism of scenarios compared to previous adversarial scenario generation methods. Furthermore, we showcase the advantage of our proposed framework in varying the safety-criticality and collision types of scenarios. These attributes make our approach particularly well-suited for the closed-loop simulation of AVs, providing a more reliable and comprehensive framework for safety evaluation. + +# 2 Related Work + +# 2.1 Traffic Simulation + +Traffic simulation can broadly be categorized into two main groups: heuristic-based and learning-based methods. In heuristic-based methods, agents are controlled by human-specified rules, such as the Intelligent Driver Model (IDM) [30] to follow a leading vehicle while maintaining a safe following distance. However, these methods have modeling capacity issues and may not reflect the real traffic distribution; a large domain gap limits its usage for planning evaluation. To close the gap, data-driven approaches learn from real driving datasets to imitate real-world behavior [23, 28, 29, 35]. TrafficSim [28] utilizes a trained variational autoencoder for scene-level traffic simulation, while BITS [35] combines high-level goal inference with low-level driving behavior imitation to enhance the realism of simulated driving behavior. Recently, the Waymo SimAgents challenge has focused on whether simulators can accurately represent real-world driving distributions [20]. However, little work has specifically focused on behavior simulation for generating safety-critical, long-tail scenarios. + +Diffusion models [27, 31, 32] have shown significant promise for synthetic image generation. One of the key advantages of diffusion models is controllability, + +which can take the form of classifier [6], classifier-free [15], and reconstruction [13] guidance. Recently, controllable diffusion models have been employed for planning and traffic simulation [25, 38, 39] via guidance. We adopt trajectory diffusion models to develop a novel approach for generate safety-critical realistic traffic simulations in which an adversarial agent collides with an ego planner agent. Other works [19, 24] use diffusion with guidance or conditioning to achieve controllable scene initialization. In contrast, we emphasize closed-loop, controllable adversarial behavior simulation based on real-world data initialization rather than scene initialization. + +# 2.2 Safety-Critical Traffic Simulation + +Safety-critical traffic generation plays a crucial role in training and evaluating AV systems, enhancing their capability to navigate diverse real-world scenarios and enhancing robustness. Gradient-based methods that leverage back-propagation to create safety-critical scenarios have been proposed to evaluate AV prediction and planning models [4,11]. Hanselmann et al. use kinematic gradients to modify vehicle trajectories, with the goal of improving the robustness of imitation learning planners [11]. Cao et al. have developed a model with differentiable dynamics, enabling the generation of realistic adversarial trajectories for trajectory prediction models through backpropagation techniques [4]. Black-box optimization approaches include perturbing actions based on kinematic bicycle models [33] and using Bayesian Optimization to create adversarial self-driving scenarios that escalate collision risks with simulated entities [1]. Zhang et al. target trajectory prediction models via white- and black-box attacks that adversarily perturb real driving trajectories [37]. For an extensive review of this topic, we direct readers to Ding et al. [9]. + +The field has recently seen advancements in data-driven methods for safety-critical scenario generation [26,34,36]. For instance, Xu et al. introduce a diffusion-based approach in CARLA, applying various adversarial optimization objectives to guide the diffusion process for safety-critical scenario generation [34]. Rempe et al. proposed STRIVE to utilize gradient-based adversarial optimization on the latent space, constrained by a graph-based CVAE traffic motion model, to generate realistic safety-critical scenarios for rule-based planners [26]. However, a common limitation in these approaches is the absence of closed-loop interaction, essential for accurately simulating interactive real-world driving. + +# 3 Problem Formulation + +We consider a simulated interactive traffic scenario consisting of $N$ agents; one is the ego vehicle controlled by the planner $\pi$ , and the remaining $N - 1$ are reactive agents modeled by a function $g$ . Our objective is to create a safety-critical closed-loop collision simulation, where reactive agents demonstrate realistic, controllable behavior. Of the $N - 1$ reactive agents, one or a subset is considered the adversarial agents (denoted as agent $a$ ), meant to collide with the ego vehicle. + +The adversarial agent, formulated within the reactive agent model $g$ , is governed by an adversarial term designed (detailed in Sec. 5) to be both controllable and adversarial to the planner $\pi$ . This setup allows the adversarial agent to pose direct challenges to $\pi$ , testing its resilience in complex scenarios. Concurrently, the other non-adversarial agents, also controlled by $g$ with varying parameters, emulate authentic, reactive behaviors, thus enriching the simulation scenario with realistic and diverse traffic conditions. The dual role of the adversarial agent and non-adversarial agents ensures that while it challenges $\pi$ , the overall simulation environment plausibly represents real-world driving conditions. + +At any given timestep $t$ , the states of the $N$ vehicles are represented as $\mathbf{s}_t = [\mathbf{s}_t^1, \dots, \mathbf{s}_t^N]$ , where $\mathbf{s}_t^i = (x_t^i, y_t^i, v_t^i, \theta_t^i)$ indicates the 2D position, speed, and yaw of vehicle $i$ . The corresponding actions for each vehicle are $\mathbf{a}_t = [\mathbf{a}_t^1, \dots, \mathbf{a}_t^N]$ , with $\mathbf{a}_t^i = (\dot{v}_t^i, \dot{\theta}_t^i)$ representing the acceleration and yaw rate. To predict the state at the next timestep $t + 1$ , a transition function $f$ is used, which computes $\mathbf{s}_{t+1} = f(\mathbf{s}_t, \mathbf{a}_t)$ based on current state and action. We adopt unicycle dynamics as the transition function. + +Each agent's decision context is $\mathbf{c}_t^i$ , which includes the agent-centric map $I^i$ and the $T_{\mathrm{hist}}$ historical states of neighboring vehicles from time $t - T_{\mathrm{hist}}$ to $t$ , defined as $\mathbf{s}_{t - T_{\mathrm{hist}}:t} = \{\mathbf{s}_{t - T_{\mathrm{hist}}}, \dots, \mathbf{s}_t\}$ . In closed-loop traffic simulation, each agent continuously generates and updates its trajectory based on the current decision context $\mathbf{c}_t^i$ . After generating a trajectory, the simulation executes the first few steps of the planned actions before updating $\mathbf{c}_t^i$ and re-planning. See Sec. 6.2 for more implementation details. + +Planner $\pi$ The planner $\pi$ determines the ego vehicle's future trajectory over a time horizon $t$ to $t + T$ . The planned state sequence is denoted by $s_{t:t + T}^{1} = \pi (\mathbf{c}_{t}^{1})$ , where $\pi (\mathbf{c}_t^1)$ processes the historical states and map data within $\mathbf{c}_t^1$ to plan future states based on the current scene context. + +Reactive Agents $g$ The reactive agent model $g$ , parameterized by $\theta$ , is designed to simulate the behavior of the $N - 1$ non-ego vehicles, represented by the set $\{s_{t:t + T}^i\}_{i = 2}^N$ . Each vehicle's state sequence, $s_{t:t + T}^i$ , is generated by $g_{\theta}(\mathbf{c}_t^i,\psi_i)$ , which incorporates the decision context $\mathbf{c}_t^i$ and a set of control parameters $\psi_{i}$ unique to each agent. These parameters $\psi_{i}$ enable the fine-tuning of individual behaviors within the simulation. In our approach, we train the model $g$ on real-world driving data to ensure the trajectories it produces are not only controllable, supporting the generation of various safety-critical scenarios, but also realistic. + +# 4 Diffusion Models for Traffic Simulation + +For closed-loop safety-critical traffic simulation, the reactive agents, especially the adversarial agent, should be 1) controllable, and 2) realistic. With recent advances in controllable diffusion models [18,25,39], we adopt trajectory diffusion models to generate realistic simulations. + +We define the model's operational trajectory as $\tau$ , which comprises both action and state sequences: $\tau \coloneqq [\tau_{a}, \tau_{s}]$ . Specifically, $\tau_{a} \coloneqq [a_{0}, \ldots, a_{T-1}]$ represents the + +sequence of actions, while $\tau_{s} \coloneqq [s_{1},\ldots ,s_{T}]$ denotes the corresponding sequence of states. Following the approach described in [39], our model predicts the action sequence $\tau_{a}$ , and the state sequence $\tau_{s}$ can be derived starting from the initial state $s_0$ and dynamic model $f$ . + +A diffusion model generates a trajectory by reversing a process that incrementally adds noise. Starting with an actual trajectory $\tau_0$ sampled from the data distribution $q(\tau_0)$ , a sequence of increasingly noisy trajectories $(\tau_{1},\tau_{2},\dots,\tau_{K})$ is produced via a forward noisng process. Each trajectory $\tau_{k}$ at step $k$ is generated by adding Gaussian noise parameterized by a predefined variance schedule $\beta_{k}$ [14]: + +$$ +q \left(\tau_ {1: K} \mid \tau_ {0}\right) := \prod_ {k = 1} ^ {K} q \left(\tau_ {k} \mid \tau_ {k - 1}\right), \tag {1} +$$ + +$$ +q \left(\tau_ {k} \mid \tau_ {k - 1}\right) := \mathcal {N} \left(\tau_ {k}; \sqrt {1 - \beta_ {k}} \tau_ {k - 1}, \beta_ {k} \mathbf {I}\right). \tag {2} +$$ + +The noising process gradually obscures the data, where the final noisy version $q(\tau_K)$ approaches $\mathcal{N}(\tau_K; \mathbf{0}, \mathbf{I})$ . The trajectory generation process is then achieved by learning the reverse of this noising process. Given a noisy trajectory $\tau_K$ , the model learns to denoise it back to $\tau_0$ through a sequence of reverse steps. Each reverse step is modeled as: + +$$ +p _ {\theta} \left(\tau_ {k - 1} \mid \tau_ {k}, \mathbf {c}\right) := \mathcal {N} \left(\tau_ {k - 1}; \mu_ {\theta} \left(\tau_ {k}, k, \mathbf {c}\right), \Sigma_ {k}\right), \tag {3} +$$ + +where $\theta$ are learned functions that predict the mean $\mu$ of the reverse step, and $\Sigma_{k}$ is a fixed schedule. By iteratively applying the reverse process, the model learns a trajectory distribution, effectively generating a plausible future trajectory from a noisy start. + +During the trajectory prediction phase, the model estimates the final clean trajectory denoted by $\hat{\tau}_0$ . This estimated trajectory is used to compute the mean $\mu$ as described in [21]. For more details, see supplementary material Sec. D. + +# 5 Diffusion Models for Safety-Critical Traffic Simulation + +The diffusion model, once trained on realistic trajectory data, inherently reflects the behavioral patterns present in its training distribution. However, to effectively simulate and analyze safety-critical scenarios, there is a crucial need for a mechanism that allows for the controlled manipulation of agent behaviors [18,39]. This is particularly important for generating adversarial behaviors and ensuring scene consistency in simulations. + +# 5.1 Guiding Reactive Agents + +Our approach specifically introduces guidance to the sampled trajectories at each denoising step, aligning them with predefined objectives $J(\tau)$ . The concept of guidance involves using the gradient of $J$ to subtly perturb the predicted mean of the model at each denoising step. This process enables the generation + +![](images/b756d8af192c285b659b43e12a9984040344b6a5e71fd95b94968fad6c5642ee.jpg) +Fig.2: Guided Diffusion Process for the Adversarial Agent. This process optimizes the adversarial agent's trajectory using the adversarial cost function $J_{\mathrm{adv}}$ to the ego vehicle. In particular, we introduce $J_{\mathrm{control}}$ to vary the adversarial behavior. Simultaneously, it applies regularization through $J_{\mathrm{reg}}$ for maintaining realism. + +of trajectories that not only reflect realistic behavior but also cater to specific simulation needs, such as adversarial testing and maintaining scene consistency over extended periods. We adopt the reconstruction guidance (clean guidance) introduced from [16, 25]: + +$$ +\tilde {\tau} _ {0} = \hat {\tau} _ {0} - \alpha \Sigma_ {k} \nabla_ {\tau_ {k}} J (\hat {\tau} _ {0}) \tag {4} +$$ + +This strategy improves guidance robustness, yielding smoother, more stable trajectories without the usual numerical issues from noisy data. + +In practice, diversifying the behavior of adversarial agents within the same scenarios is crucial for a thorough assessment of AVs. Despite the significance of this challenge, it remains largely unexplored in previous works [26,39] + +The loss function for the non-reactive agents, $J(\tau)$ , consists of a collision term $J_{\mathrm{coll}}$ , which encourages collisions between the adversarial agent and the ego agent, two control terms $J_{\mathrm{v}}$ and $J_{\mathrm{ttc}}$ , which control the relative speed and time-to-collision between the ego and adversarial agent respectively, a regularization term $J_{\mathrm{Gauss}}$ , which discourages collisions between the reactive agents, and a route guidance term $J_{\mathrm{route}}$ , which discourages the reactive agents from going outside the road: + +$$ +J (\tau) = \rho \underbrace {\left(J _ {\mathrm {c o l l}} + J _ {\mathrm {v}} + J _ {\mathrm {t t c}}\right)} _ {J _ {\mathrm {a d v}} (\tau)} + \underbrace {J _ {\text {r o u t e}} + J _ {\text {G a u s s}}} _ {J _ {\text {r e g}} (\tau)}, \tag {5} +$$ + +where $\rho$ denotes a scalar weight that determines whether a reactive agent behaves adversarially towards the ego agent, i.e., whether it attempts to collide with the ego agent. + +Collision with Planner We define $J_{\mathrm{coll}}$ to encourage the collision between the adversarial agent and the ego agent, given by: + +$$ +J _ {\text {c o l l}} = - \sum_ {t = 1} ^ {T} d (t), \tag {6} +$$ + +where $d(t)$ represents the distance between the ego and the adversarial agent at each time step of the planning horizon $T$ . The adversarial agent is either pre-selected based on lane proximity or dynamically selected based on the distance to the ego agent. Details are in the supplementary Sec. D.4. + +Safety Criticality of Collisions We control the relative speed $J_{v}$ between the ego and adversary at each time step $(v_{t}^{1}$ and $v_{t}^{a}$ ), and the time-to-collision (TTC) cost $J_{\mathrm{ttc}}$ [22] to control the safety criticality of potential collisions, with the latter given by: + +$$ +J _ {\mathrm {t t c}} = \sum_ {t = 1} ^ {T} - \exp \left(- \frac {\tilde {t} _ {\operatorname {c o l} (t)} ^ {2}}{2 \lambda_ {t}} - \frac {\tilde {d} _ {\operatorname {c o l} (t)} ^ {2}}{2 \lambda_ {d}}\right), \tag {7} +$$ + +where $\tilde{t}_{\mathrm{col}(t)}$ is the time to collision at time $t$ , $\tilde{d}_{\mathrm{col}(t)}$ is the distance to collision and $\lambda_t$ and $\lambda_d$ are bandwidth parameters for time and distance. This formula uses a constant velocity assumption. Intuitively, the time-to-collision cost favors scenarios with high relative speeds and challenging collision angles for the ego vehicle to avoid. For details of $J_v$ and $J_{\mathrm{ttc}}$ , see supplementary Sec. C.2. + +Route Guidance Given an agent's trajectory $\tau$ and the corresponding route $r$ -the predefined path on a lane graph from its starting point to its destination—we compute the normal distance of each point $\tau_{t}$ on the trajectory to the route at each timestep. We then penalize deviations from the route that exceed a predefined margin $d$ . This process is captured by the following route guidance cost function: + +$$ +J _ {\text {r o u t e}} (\tau , r) = \sum_ {t = 1} ^ {T} \max \left(0, \left| d _ {n} \left(\tau_ {t}, r\right) - d _ {m} \right|\right), \tag {8} +$$ + +where $d_{n}(\tau_{n},r)$ denotes the normal distance from the point $\tau_{t}$ on the trajectory to the nearest point on the route $r$ at timestep $t$ , and $d_{m}$ represents the acceptable deviation margin from the route. In contrast to the off-road loss in prior studies [39], our proposed route guidance system more effectively indicates each agent's intended path, improving adherence to traffic rules as demonstrated in Sec. 6. The flexibility of route guidance supports diverse agent interactions, such as modifying routes to encourage lane changes among reactive agents [29]. + +Gaussian Collision Guidance Given the trajectories of agents, we calculate the Gaussian distance for each pair of agents $(i,j)$ at each timestep $t$ from 1 to $T$ . The Gaussian distance between the agents takes into account both the tangential $(d_t)$ and normal $(d_n)$ components of the projected distances. The aggregated Gaussian distance is: + +$$ +J _ {\mathrm {G a u s s}} = \sum_ {t = 1} ^ {T} \sum_ {i, j} ^ {N} \exp \left(- \frac {1}{2 \sigma^ {2}} \left(\lambda \cdot d _ {t} ^ {i j} (t) ^ {2} + d _ {n} ^ {i j} (t) ^ {2}\right)\right) \tag {9} +$$ + +![](images/7312088dc904c464486e66a5d4db329cf0ad414cb7fd177954ce8dd59f127021.jpg) +Fig. 3: Framework for Partial Diffusion. We generate proposals based on domain knowledge (e.g., collision types). Users can adjust noise levels to balance between user control and the model's data distribution. + +where $d_t^{ij}(t)$ and $d_n^{ij}(t)$ represent the tangential and normal distances from agent $j$ 's trajectory point at time $t$ to agent $i$ 's heading axis, respectively, and $\sigma$ is the standard deviation for these distances. In this formulation, $\lambda$ is a scaling factor applied to the tangential distance $d_t^{ij}(t)$ . This approach contrasts with the disk approximation method, which primarily penalizes the Euclidean distance between agents. By accounting for both tangential and normal components, the Gaussian collision distance method significantly reduces collision rate, which we discuss in Sec. 6. + +# 5.2 Partial Diffusion: Controlling Collision Types + +We introduce a novel approach through a partial diffusion process, utilizing trajectory proposals to initiate the diffusion process. This methodology enables the variation in collision types by the adversarial agent within the diffusion, tailoring the adversarial outcomes to specific evaluation needs, the results are discussed in Sec. 6.6. + +Figure 3 illustrates our framework, which is divided into three main steps to generate trajectory proposals for various collision scenarios. First, we create initial trajectory proposals $(\tau_0)$ aimed at capturing different types of collisions. The next critical step involves setting the partial diffusion ratio $\gamma$ , which defines the specific point in the process, $k_{p} = \gamma \cdot K$ , at which we start modifying the trajectory. Starting from step $k_{p}$ , we adjust the trajectory by adding a precise level of Gaussian noise $\epsilon \sim N(0,I)$ : $\hat{\tau}_{k_p} = \sqrt{\bar{\alpha}_{k_p}}\tau_0 + \sqrt{1 - \bar{\alpha}_{k_p}}\epsilon$ . The final stages include removing noise and using guided diffusion for the rest of the $k_{p}$ steps to refine the trajectory into a realistic path that suits our collision scenario goals. + +To generate the trajectory proposals, we develop a rule-based approach in which we first identify the centerlines of the ego and adversarial agent and then search for potential intersections of their respective centerlines. If such an intersection exists, we generate the proposals by selecting an acceleration value and lateral offset from the centerline that is likely to cause the desired collision type based on the projected plan of the ego agent. Note that the trajectory + +![](images/0f1b5f6fd3666f168a6f85918ace569a8d8c85b1afd49f40ce56bdb58db5c66c.jpg) + +![](images/3b09fa2c72b8599715fa729276012f5d7fec63bfa328ac2b18582f5f7da19fb7.jpg) +Ego + +![](images/a693a367c2b58420629a3e21b76ec3eff0c4557317a3525d8f10c72cf706b76d.jpg) +Fig. 4: Partial Diffusion Results of Rule-Based Planner on NuScenes Dataset. + +![](images/cc07f46929c87ed2c587b8550a66dd8b64c55793cffa7689f5d2f6b1c3ae0728.jpg) +Non-Adv + +![](images/1336d5febd7950e8f3440f812f19ef859afb8486895c3a430ecfbefb04c82b18.jpg) +Adv + +![](images/d05a2e9ddbb5cec801cafc48acef0785d13fce16ac5a54863ee58c5e8f7d148a.jpg) + +![](images/316a2f1513fbe5fb695fe3dbe0527b3df2845e09a34b83924a11c3389ec3fff8.jpg) +Trajectory Proposal + +The safety-critical scenarios show the framework's ability to create realistic and challenging situations, varying collision types based on different trajectory proposals. The gradient lines reflects the planned trajectories in the next 3.2 seconds. + +proposals are updated in a closed-loop manner to account for the interaction between the ego and the adversarial agent. + +This method allows for precise control over the diffusion trajectory, enabling adversarial agents to create customized collision scenarios. Users can adjust $\gamma$ to fine-tune the balance between explicit control and the model's trained data distribution. + +# 6 Experiments + +We validate the efficacy of our proposed framework via experiments with real-world driving data. Our results demonstrate that the framework can generate realistic and controllable adversarial behavior to challenge the planner. + +# 6.1 Dataset + +We conduct our experiments on two large-scale real-world driving datasets: nuScenes [2], which consists of 5.5 hours of driving data from two cities, and nuPlan, which consists of 1500 hours of driving data from four cities. We train the model on scenes from the nuScenes train split and evaluate it on the scenes from the nuScenes validation splits and nuPlan mini validation splits. We focus on vehicle-to-vehicle interactions. + +# 6.2 Implementation Details + +Baselines. We compare our approach against STRIVE [26] using their open-source implementation and our re-implementation of DiffScene [34]. STRIVE is recognized for its proficiency in generating adversarial safety-critical scenarios using a learned traffic model and adversarial optimization in the latent space. + +Planner. Our experiments utilize a range of different planners: 1) A rule-based planner, as implemented in STRIVE [26], which operates on the lane graph + +and employs the constant velocity model to predict future trajectories of non-ego vehicles, generating multiple trajectory candidates and selecting the one least likely to result in a collision; 2) a hybrid planner BITS [35]; 3) PDM-Closed [5], the winner of the 2023 nuPlan Planning Challenge; 4) a deterministic a learning-based Behavior Cloning (BC) planner, which utilizes a ResNet Encoder, followed by an MLP decoder to generate future trajectories; and 5) an Intelligent Driver Model (IDM) planner [30]. + +Diffusion Model. We follow the architecture described in [39]. We represent the context using an agent-centric map and past trajectories on a rasterized map. The traffic scene is encoded using a ResNet structure [12], while the input trajectory is processed through a series of 1D temporal convolution blocks in an UNet-like architecture, as detailed in [18]. The model uses $K = 100$ diffusion steps. During the inference phase, we generate a sample of potential future trajectories for each reactive agent in a given scene. From these, we select the trajectory that yields the lowest guidance cost. This process is referred to as filtering. + +Closed-loop Simulation. The simulation framework for our experiments is built upon an open-source traffic behavior framework [35]. Within this framework, both the planner and reactive agents update their plans at a frequency of $2\mathrm{Hz}$ . + +# 6.3 Evaluation Metrics + +Our goal is to validate that the proposed method can generate safety-critical scenarios that are both realistic and controllable. For realism assessment, in accordance with [39], we compare statistical data between simulated trajectories and actual ground trajectories. This involves calculating the Wasserstein distance between their driving profiles' normalized histograms, focusing on the mean of mean values for three properties: longitudinal acceleration, latitudinal acceleration, and jerk. To evaluate controllability, we measure metrics related to parameters we can control, specifically relative speed and time-to-collision cost between the ego and adversarial agents. + +Our method aims to evaluate the extent of control over adversarial behaviors within a single scenario. To do this, we focus on measuring collision diversity—a metric that quantifies the range of differences in collision angles, relative speeds, and collision points. By calculating the variance of these parameters within the same scenario, we can determine how diverse and controllable the adversarial behaviors are, ensuring the scenario's realism and controllability. + +For a more nuanced understanding of realism, we analyze the collision rate - the average fractions of agents colliding, and the offroad rate - the percentage of reactive agents going off-road, both of which are considered failure rates [35]. All metrics are averaged across scenarios. For detailed definitions and metrics, see supplementary Sec. C. + +
DatasetMethodCollision (%) ↑Other Offroad (%)↓Adv Offroad (%) ↓Collision Rel Speed (m/s) ↓Realism ↓Time (s) ↓
nuScenesSAFE-SIM43.21.811.4-0.120.38104.5 ± 17.7
STRIVE36.42.211.45.520.85427.2 ± 169.8
DiffScene18.211.49.016.40.52105.4 ± 22.5
nuPlanSAFE-SIM809.411.76.750.27173.4 ± 73.3
DiffScene56.714.05.0-2.810.42176.7 ± 77.5
+ +Table 2: Safety-critical Traffic Simulation. We compare our approach against STRIVE [26] and DiffScene [34] for safety-critical traffic simulation with a rule-based planner. SAFE-SIM outperforms STRIVE on all metrics and demonstrates higher collision rates and better realism than DiffScene. + +
PlannerEgo-Adv Coll (%) ↑Ego-Other Coll (%) ↑Adv Offroad (%) ↓Coll Speed (m/s) ↓Ego Accel (m/s2) ↓Realism ↓
BC38.837.39.02.470.540.79
IDM49.358.23.0-0.400.940.78
Lane-Graph34.337.31.52.681.620.57
BITS16.419.46.03.070.950.79
PDM-Closed26.950.71.52.731.300.86
+ +Table 3: Safety-Critical Simulation for Different Planners. SAFE-SIM can generate diverse safety-critical scenarios tailored to different planners, including rule-based, learning-based, and hybrid planners. + +# 6.4 Evaluation of Safety-Critical Traffic Simulation with Baseline Methods + +We compared our method with STRIVE [26] and DiffScene [34], utilizing the Lane-graph-based planner from STRIVE. The evaluation focused on collision rates between the ego and the adversarial agent ("Collision"), adversarial agent off-road rates ("Adv Offroad"), other agents' off-road rates ("Other Offroad"), collision relative speed between the ego and adversary ("Collision Rel Speed"), realism of all non-ego agents ("Realism"), and simulation time. + +The results, presented in Tab. 2, demonstrate that our method excels in all metrics compared to STRIVE, especially in collisions and realism. Compared to DiffScene, SAFE-SIM also exhibits a higher collision rate with better realism. Note that we trained models on nuScenes and tested them in nuPlan without additional fine-tuning. As illustrated in Figure 4, the qualitative examples from the NuScenes dataset demonstrate how our framework can challenge the rule-based planner with various driving situations. See supplementary Sec. A for more qualitative examples. + +# 6.5 Safety-Critical Simulation with Different Planners + +In Tab. 3, we demonstrate our framework's ability to generate collisions across various planner types: rule-based, learning-based, and hybrid planners. Notably, the IDM planner exhibits the highest Ego-Adv collision rate. This heightened rate + +
TTC Cost WeightTTC CostColl Speed (m/s)Coll Angle (deg)Coll Rate (%)↑Realism ↓
0.00.182.45-7.4348.20.76
1.00.212.300.4353.60.79
2.00.263.78-17.060.70.81
+ +Table 4: Controlling Time to Collision (TTC). The table shows the impact of different TTC Cost weights on collision scenarios. Increasing the TTC Cost weight results in an increase in collision rate, suggesting a heightened challenge for the ego vehicle in avoiding collisions. + +
Jadv Jreg Partial DiffusionCollision MetricsDiversity
Collision (%) ↑Adv Offroad (%) ↓Realism ↓Collision Angle Var (rad) ↑Collision Rel Speed Var (m/s) ↑Collision Point Var (m) ↑
✓✓×23.913.80.582.222.991.62
✓✓✓29.014.60.573.101.965.44
✓××53.523.10.583.344.812.47
+ +Table 5: Ablation Study on Controllability. This study examines each component of our proposed method. Partial diffusion significantly increases the variance of the collision point, resulting in a greater diversity of collision scenarios. + +can be attributed to the IDM's focus on vehicles near the same lane, potentially overlooking other vehicles in the scene. Consequently, in scenarios with the IDM planner, our framework can induce collisions at relatively lower speeds. + +# 6.6 Evaluation: Controlling Safety-Criticality + +A key feature of SAFE-SIM is its ability to generate controllable adversarial behaviors, offering variations not possible with previous methods. + +Controlling Time-to-Collision. We control the orientation and relative speed together using the time-to-collision (TTC) cost, as described in Sec. 5. We manipulate the scenario's safety-criticality by adjusting the relative weight of the TTC cost. To assess the impact of these adjustments, we measure the average TTC cost shortly before a collision occurs (0.5 seconds). Our observations, detailed in Tab. 4, show that increasing the TTC weight raises the TTC cost. Notably, while the relative collision speed remains fairly consistent, the collision angle shifts, indicating a greater difficulty in avoiding ego-adversary collisions. Additionally, our method can control other aspects, such as relative speed, as detailed in supplementary Sec. F.1. + +# 6.7 Ablation Study on Controllability + +We performed an ablation study to assess the influence of different guidance strategies in diffusion models on the quality of simulations. This study, detailed + +in Table 5, aimed to quantify collision diversity. We compare against our baseline approach, consisting of regularized and adversarial objectives $(J_{reg} + J_{adv})$ , by predetermining the selection of adversarial agents and conducting experiments with multiple seeds (three) in our framework to evaluate collision diversity. In partial diffusion, we manipulated the trajectory proposal selection mechanism across centerlines with normal offsets of -2.0, 0.0, and 2.0. + +Effectiveness of Partial Diffusion. Table 5 reveals that partial diffusion significantly enhances both the collision point and angle diversity in comparison to the baseline approach. This underscores the capability of partial diffusion to generate a wider array of collision scenarios through various trajectory proposals, illustrating its potential to explore diverse collision dynamics effectively. + +Impact of the Regularization Term ( $J_{reg}$ ). Incorporating $J_{reg}$ results in a notable decrease in the adversarial-collision rate, highlighting the importance of the regularization term in enhancing simulation realism. Without $J_{reg}$ , the collision rate between the adversarial agent and ego vehicle increases, but the adversarial vehicle goes offroad more often, leading to scenarios that differ significantly from realistic behavior. + +# 6.8 Limitation and Failure Cases + +We identified areas for improvement in SAFE-SIM (see supp. Fig. A4). In certain cases, the adversarial agent unrealistically collides with non-adversarial agents before reaching the ego agent. Additionally, some scenarios result in collisions where the ego planner is not at fault. While understanding how the ego planner can avoid such cases is important, creating more scenarios where the ego is at fault would be beneficial. + +# 7 Conclusion + +In this work, we present a closed-loop simulation framework utilizing guided diffusion models for creating safety-critical scenarios to assess autonomous vehicle (AV) algorithms. Our research is in line with the goals of SO-TIF, focusing on how autonomous vehicles respond to dynamic scenarios like aggressive driving, underscoring our dedication to safety across diverse and unforeseeable conditions. Our framework introduces innovative guidance objectives tailored for controllable, stable, long-term safety-critical simulations. A key aspect of our method lies in its ability to vary the types of adversarial behavior within collision scenarios. By integrating adversarial objectives and partial diffusion, we enable fine-grained control over adversarial actions. This versatility enables our framework to produce a broader range of realistic and manageable scenarios, setting a new standard in adversarial scenario generation beyond the limitations of existing approaches. + +Future directions for our research include: 1) exploring the application of our framework in closed-loop policy training, and 2) developing automated methods for adjusting controllable parameters. These methods aim to facilitate the generation of diverse, long-tail scenarios. We believe this framework holds significant promise for enhancing real-world AV safety. + +# Acknowledgements + +This work was part of W.J. Chang's summer internship at NEC Labs America, and he is also supported by the National Science Foundation Graduate Research Fellowship Program under Grant No. DGE 2146752. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation. The authors would like to thank Chih-Ling Chang for her insightful suggestions and assistance with figures and presentations. + +# References + +1. Abeysirigoonawardena, Y., Shkurti, F., Dudek, G.: Generating adversarial driving scenarios in high-fidelity simulators. In: 2019 International Conference on Robotics and Automation (ICRA). pp. 8271-8277. IEEE (2019) +2. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11621-11631 (2020) +3. Caesar, H., Kabzan, J., Tan, K.S., Fong, W.K., Wolff, E., Lang, A., Fletcher, L., Beijbom, O., Omari, S.: nuplan: A closed-loop ml-based planning benchmark for autonomous vehicles. arXiv preprint arXiv:2106.11810 (2021) +4. Cao, Y., Xiao, C., Anandkumar, A., Xu, D., Pavone, M.: Advdo: Realistic adversarial attacks for trajectory prediction. In: European Conference on Computer Vision. pp 36-52. Springer (2022) +5. Dauner, D., Hallgarten, M., Geiger, A., Chitta, K.: Parting with misconceptions about learning-based vehicle motion planning. In: CoRL (2023) +6. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in neural information processing systems 34, 8780-8794 (2021) +7. Ding, W., Chen, B., Li, B., Eun, K.J., Zhao, D.: Multimodal safety-critical scenarios generation for decision-making algorithms evaluation. IEEE Robotics and Automation Letters 6(2), 1551-1558 (2021). https://doi.org/10.1109/LRA.2021.3058873 +8. Ding, W., Chen, B., Xu, M., Zhao, D.: Learning to collide: An adaptive safety-critical scenarios generating method. In: 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 2243-2250. IEEE (2020) +9. Ding, W., Xu, C., Arief, M., Lin, H., Li, B., Zhao, D.: A survey on safety-critical driving scenario generation—a methodological perspective. IEEE Transactions on Intelligent Transportation Systems (2023) +10. Gu, T., Chen, G., Li, J., Lin, C., Rao, Y., Zhou, J., Lu, J.: Stochastic trajectory prediction via motion indeterminacy diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17113-17122 (2022) +11. Hanselmann, N., Renz, K., Chitta, K., Bhattacharyya, A., Geiger, A.: King: Generating safety-critical driving scenarios for robust imitation via kinematics gradients In: European Conference on Computer Vision. pp. 335-352. Springer (2022) +12. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition In: Proceedings of the IEEE conference on computer vision and pattern recognition pp. 770-778 (2016) + +13. Ho, J., Chan, W., Sahara, C., Whang, J., Gao, R., Gritsenko, A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., et al.: Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303 (2022) +14. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020) +15. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022) +16. Ho, J., Salimans, T., Gritsenko, A., Chan, W., Norouzi, M., Fleet, D.J.: Video diffusion models. arXiv:2204.03458 (2022) +17. Ivanovic, B., Song, G., Gilitschenski, I., Pavone, M.: trajdata: A unified interface to multiple human trajectory datasets. arXiv preprint arXiv:2307.13924 (2023) +18. Janner, M., Du, Y., Tenenbaum, J.B., Levine, S.: Planning with diffusion for flexible behavior synthesis. arXiv preprint arXiv:2205.09991 (2022) +19. Lu, J., Wong, K., Zhang, C., Suo, S., Urtasun, R.: Scenecontrol: Diffusion for controllable traffic scene generation. In: IEEE International Conference on Robotics and Automation (ICRA) (2024) +20. Montali, N., Lambert, J., Mougin, P., Kuefler, A., Rhinehart, N., Li, M., Gulino, C., Emrich, T., Yang, Z., Whiteson, S., et al.: The waymo open sim agents challenge. Advances in Neural Information Processing Systems 36 (2024) +21. Nichol, A.Q., Dhariwal, P.: Improved denoising diffusion probabilistic models. In: International Conference on Machine Learning. pp. 8162-8171. PMLR (2021) +22. Nishimura, H., Mercat, J., Wulfe, B., McAllister, R.T., Gaidon, A.: Rap: Risk-aware prediction for robust planning. In: Conference on Robot Learning. pp. 381-392. PMLR (2023) +23. Philion, J., Peng, X.B., Fidler, S.: Trajeglish: Traffic modeling as next-token prediction. In: The Twelfth International Conference on Learning Representations (2024) +24. Pronovost, E., Ganesina, M.R., Hendy, N., Wang, Z., Morales, A., Wang, K., Roy, N.: Scenario diffusion: Controllable driving scenario generation with diffusion. Advances in Neural Information Processing Systems 36, 68873-68894 (2023) +25. Rempe, D., Luo, Z., Bin Peng, X., Yuan, Y., Kitani, K., Kreis, K., Fidler, S., Litany, O.: Trace and pace: Controllable pedestrian animation via guided trajectory diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13756-13766 (2023) +26. Rempe, D., Philion, J., Guibas, L.J., Fidler, S., Litany, O.: Generating useful accident-prone driving scenarios via a learned traffic prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17305-17315 (2022) +27. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022) +28. Suo, S., Regalado, S., Casas, S., Urtasun, R.: Trafficsim: Learning to simulate realistic multi-agent behaviors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10400-10409 (2021) +29. Suo, S., Wong, K., Xu, J., Tu, J., Cui, A., Casas, S., Urtasun, R.: Mixsim: A hierarchical framework for mixed reality traffic simulation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9622-9631 (2023) +30. Treiber, M., Hennecke, A., Helbing, D.: Congested traffic states in empirical observations and microscopic simulations. Physical review E 62(2), 1805 (2000) + +31. Unknown: Midjourney. https://www.midjourney.com, accessed: 2023-11-16 +32. Unknown: Openai dall-e-2. https://openai.com/product/dall-e-2, accessed: 2023-11-16 +33. Wang, J., Pun, A., Tu, J., Manivasagam, S., Sadat, A., Casas, S., Ren, M., Urtasun, R.: Advsim: Generating safety-critical scenarios for self-driving vehicles. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9909-9918 (2021) +34. Xu, C., Zhao, D., Sangiovanni-Vincentelli, A., Li, B.: Diffscene: Diffusion-based safety-critical scenario generation for autonomous vehicles. In: The Second Workshop on New Frontiers in Adversarial Machine Learning (2023) +35. Xu, D., Chen, Y., Ivanovic, B., Pavone, M.: Bits: Bi-level imitation for traffic simulation. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 2929-2936. IEEE (2023) +36. Yin, Z.H., Sun, L., Sun, L., Tomizuka, M., Zhan, W.: Diverse critical interaction generation for planning and planner evaluation. In: 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 7036-7043. IEEE (2021) +37. Zhang, Q., Hu, S., Sun, J., Chen, Q.A., Mao, Z.M.: On adversarial robustness of trajectory prediction for autonomous vehicles. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15159-15168 (2022) +38. Zhong, Z., Rempe, D., Chen, Y., Ivanovic, B., Cao, Y., Xu, D., Pavone, M., Ray, B.: Language-guided traffic simulation via scene-level diffusion. arXiv preprint arXiv:2306.06344 (2023) +39. Zhong, Z., Rempe, D., Xu, D., Chen, Y., Veer, S., Che, T., Ray, B., Pavone, M.: Guided conditional diffusion for controllable traffic simulation. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 3560-3566. IEEE (2023) \ No newline at end of file diff --git a/2024/Safe-Sim_ Safety-Critical Closed-Loop Traffic Simulation with Diffusion-Controllable Adversaries/images.zip b/2024/Safe-Sim_ Safety-Critical Closed-Loop Traffic Simulation with Diffusion-Controllable Adversaries/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..78f1bd131135482ebd5e84227f2058bb43e24985 --- /dev/null +++ b/2024/Safe-Sim_ Safety-Critical Closed-Loop Traffic Simulation with Diffusion-Controllable Adversaries/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:00ceef58466e40d608d526001e4bd4fdf12ee807c8fb65cf98514d4bec43c8c9 +size 321039 diff --git a/2024/Safe-Sim_ Safety-Critical Closed-Loop Traffic Simulation with Diffusion-Controllable Adversaries/layout.json b/2024/Safe-Sim_ Safety-Critical Closed-Loop Traffic Simulation with Diffusion-Controllable Adversaries/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b8e6d48e0310e22221e83e78dee622389945afa7 --- /dev/null +++ b/2024/Safe-Sim_ Safety-Critical Closed-Loop Traffic Simulation with Diffusion-Controllable Adversaries/layout.json @@ -0,0 +1,9857 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 132, + 111, + 481, + 145 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 111, + 481, + 145 + ], + "spans": [ + { + "bbox": [ + 132, + 111, + 481, + 145 + ], + "type": "text", + "content": "SAFE-SIM: Safety-Critical Closed-Loop Traffic Simulation with Diffusion-Controllable Adversaries" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 134, + 167, + 478, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 167, + 478, + 192 + ], + "spans": [ + { + "bbox": [ + 134, + 167, + 478, + 192 + ], + "type": "text", + "content": "Wei-Jer Chang" + }, + { + "bbox": [ + 134, + 167, + 478, + 192 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 134, + 167, + 478, + 192 + ], + "type": "text", + "content": ", Francesco Pittaluga" + }, + { + "bbox": [ + 134, + 167, + 478, + 192 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 134, + 167, + 478, + 192 + ], + "type": "text", + "content": ", Masayoshi Tomizuka" + }, + { + "bbox": [ + 134, + 167, + 478, + 192 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 134, + 167, + 478, + 192 + ], + "type": "text", + "content": ", Wei Zhan" + }, + { + "bbox": [ + 134, + 167, + 478, + 192 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 134, + 167, + 478, + 192 + ], + "type": "text", + "content": ", and Manmohan Chandraker" + }, + { + "bbox": [ + 134, + 167, + 478, + 192 + ], + "type": "inline_equation", + "content": "^{2,3}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 179, + 201, + 432, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 201, + 432, + 213 + ], + "spans": [ + { + "bbox": [ + 179, + 201, + 432, + 213 + ], + "type": "text", + "content": "UC Berkeley 2 NEC Labs America 3 UC San Diego" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 245, + 455, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 245, + 455, + 498 + ], + "spans": [ + { + "bbox": [ + 160, + 245, + 455, + 498 + ], + "type": "text", + "content": "Abstract. Evaluating the performance of autonomous vehicle planning algorithms necessitates simulating long-tail safety-critical traffic scenarios. However, traditional methods for generating such scenarios often fall short in terms of controllability and realism; they also neglect the dynamics of agent interactions. To address these limitations, we introduce SAFE-SIM, a novel diffusion-based controllable closed-loop safety-critical simulation framework. Our approach yields two distinct advantages: 1) generating realistic long-tail safety-critical scenarios that closely reflect real-world conditions, and 2) providing controllable adversarial behavior for more comprehensive and interactive evaluations. We develop a novel approach to simulate safety-critical scenarios through an adversarial term in the denoising process of diffusion models, which allows an adversarial agent to challenge a planner with plausible maneuvers while all agents in the scene exhibit reactive and realistic behaviors. Furthermore, we propose novel guidance objectives and a partial diffusion process that enables users to control key aspects of the scenarios, such as the collision type and aggressiveness of the adversarial agent, while maintaining the realism of the behavior. We validate our framework empirically using the nuScenes and nuPlan datasets across multiple planners, demonstrating improvements in both realism and controllability. These findings affirm that diffusion models provide a robust and versatile foundation for safety-critical, interactive traffic simulation, extending their utility across the broader autonomous driving landscape. Project website: https://safe-sim.github.io/." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 519, + 230, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 519, + 230, + 532 + ], + "spans": [ + { + "bbox": [ + 132, + 519, + 230, + 532 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 545, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 628 + ], + "type": "text", + "content": "A key safety feature of autonomous vehicles (AVs) is their ability to navigate near-collision events in real-world scenarios. However, these events rarely occur on roads and testing AVs in such high-risk situations on public roads is unsafe. Therefore, simulation is indispensable in the development and assessment of AVs, providing a safe and reliable means to study their safety and dependability. A critical aspect of simulation is modeling the behavior of other road users, since AVs must learn to interact with them safely." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "A common method of safety-critical testing of AVs involves manually designing scenarios that could potentially lead to failures, such as collisions. While this approach allows for targeted testing, it is inherently limited in scalability and" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 114, + 479, + 203 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 203 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 203 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 203 + ], + "type": "table", + "html": "
MethodSafety-CriticalControllableControllable \nAdversaryEvaluate \nPlannerClosed-LoopReal-World
CTG [39]×✓××✓✓
CTG++ [38]×✓××✓✓
STRIVE [26]✓××✓✓✓
DiffScene [34]✓✓×✓××
SAFE-SIM (Ours)✓✓✓✓✓✓
", + "image_path": "ca940f7ca32dd9a76e618ddbf2622835949d1ac278fbf36f54544ae2d0ebbf70.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 130, + 204, + 482, + 270 + ], + "lines": [ + { + "bbox": [ + 130, + 204, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 204, + 482, + 270 + ], + "type": "text", + "content": "Table 1: Comparison of methods. Our contribution is the development of a framework for (a) safety-critical (b) closed-loop (c) controllable adversarial simulations. These aspects are not concurrently present in previous frameworks. We formulate a novel partial diffusion with novel guidance functions for stable long-term simulation and are the first to enable an ego planner to be tested against controllable adversaries with varied behavior patterns." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 278, + 482, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 278, + 482, + 386 + ], + "spans": [ + { + "bbox": [ + 130, + 278, + 482, + 386 + ], + "type": "text", + "content": "lacks the comprehensiveness required for thorough evaluation [7,8,33]. Some recent works focus on automatically generating challenging scenarios that cause planners to fail, but their emphasis has been mostly on static scenario generation rather than dynamic, closed-loop simulations. This results in a critical gap: the behavior of other agents often does not adapt or respond to the planner's actions, which is essential for a comprehensive safety evaluation. Furthermore, the results from these simulations often lack controllability, typically producing only a single adversarial outcome per scenario without the flexibility to explore a range of conditions and responses." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 388, + 482, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 482, + 483 + ], + "type": "text", + "content": "In this work, we introduce SAFE-SIM, a closed-loop simulation framework for generating safety-critical scenarios, with a particular emphasis on controllability and realism for the behavior of agents, which allows simulations over a long-horizon as needed to evaluate AV planning algorithms (Fig. 1). Different from prior works [26,34,38,39] that primarily adhere to rule-constraint satisfaction, our approach enhances controllability by modulating adversarial vehicle behaviors within identical scenarios, thereby facilitating a broader exploration of potential outcomes. See Tab. 1 for a comprehensive comparison of these approaches." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 484, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 484, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 484, + 482, + 628 + ], + "type": "text", + "content": "Our approach builds upon recent developments in controllable diffusion models [18, 25, 39]. Specifically, we adopt a test-time guidance to direct the denoising phase of the diffusion process, using the gradients from differentiable objectives to enhance scenario generation, enabling generation of adversarial scenarios in which and adversarial agent collides with the ego agent behaving according to specific planning policy. Additionally, we develop an novel approach, which we refer to as Partial Diffusion that introduces trajectory proposals into the diffusion process to provide a high degree of controllability over the type of collision scenario. Overall, our balanced integration of adversarial objectives with regularization during the guidance phase combined with Partial Diffusion allows for refined control over the conditions of the generated scenarios, ensuring both their realism and relevance to safety-critical testing." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": "In our study, we use the nuScenes [2] and nuPlan [3] datasets to evaluate the efficacy of our method in generating safety-critical closed-loop simulations. Our results demonstrate a marked improvement in the controllability and real-" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "type": "text", + "content": "W.J. Chang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 118, + 473, + 237 + ], + "blocks": [ + { + "bbox": [ + 140, + 118, + 473, + 237 + ], + "lines": [ + { + "bbox": [ + 140, + 118, + 473, + 237 + ], + "spans": [ + { + "bbox": [ + 140, + 118, + 473, + 237 + ], + "type": "image", + "image_path": "a4cfa613335605d7fccd1ea9eb1f093ea3313a5383d54d4400afd0f7241febda.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 248, + 482, + 338 + ], + "lines": [ + { + "bbox": [ + 130, + 248, + 482, + 338 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 482, + 338 + ], + "type": "text", + "content": "Fig. 1: Overview of SAFE-SIM Framework for Controllable Safety-Critical Closed-Loop Simulation. This framework evaluates a planner within scenarios featuring multiple controllable reactive agents. These agents have two distinct roles: adversarial agents, which actively challenge the planner by exhibiting controllable adversarial behaviors such as specific collision types and levels of aggressiveness, and non-adversarial agents, which follow normal driving behavior to maintain the realism of the entire scene. Such a setup facilitates the generation of various realistic, interactive, and safety-critical scenarios, providing a thorough evaluation of the planner's capabilities." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 358, + 482, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 358, + 482, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 358, + 482, + 418 + ], + "type": "text", + "content": "ism of scenarios compared to previous adversarial scenario generation methods. Furthermore, we showcase the advantage of our proposed framework in varying the safety-criticality and collision types of scenarios. These attributes make our approach particularly well-suited for the closed-loop simulation of AVs, providing a more reliable and comprehensive framework for safety evaluation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 434, + 237, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 434, + 237, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 434, + 237, + 445 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 456, + 250, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 456, + 250, + 467 + ], + "spans": [ + { + "bbox": [ + 132, + 456, + 250, + 467 + ], + "type": "text", + "content": "2.1 Traffic Simulation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 474, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 641 + ], + "type": "text", + "content": "Traffic simulation can broadly be categorized into two main groups: heuristic-based and learning-based methods. In heuristic-based methods, agents are controlled by human-specified rules, such as the Intelligent Driver Model (IDM) [30] to follow a leading vehicle while maintaining a safe following distance. However, these methods have modeling capacity issues and may not reflect the real traffic distribution; a large domain gap limits its usage for planning evaluation. To close the gap, data-driven approaches learn from real driving datasets to imitate real-world behavior [23, 28, 29, 35]. TrafficSim [28] utilizes a trained variational autoencoder for scene-level traffic simulation, while BITS [35] combines high-level goal inference with low-level driving behavior imitation to enhance the realism of simulated driving behavior. Recently, the Waymo SimAgents challenge has focused on whether simulators can accurately represent real-world driving distributions [20]. However, little work has specifically focused on behavior simulation for generating safety-critical, long-tail scenarios." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "content": "Diffusion models [27, 31, 32] have shown significant promise for synthetic image generation. One of the key advantages of diffusion models is controllability," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 135, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 135, + 91, + 448, + 102 + ], + "type": "text", + "content": "Safety-Critical Closed-Loop Traffic Simulation with Controllable Adversaries" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "which can take the form of classifier [6], classifier-free [15], and reconstruction [13] guidance. Recently, controllable diffusion models have been employed for planning and traffic simulation [25, 38, 39] via guidance. We adopt trajectory diffusion models to develop a novel approach for generate safety-critical realistic traffic simulations in which an adversarial agent collides with an ego planner agent. Other works [19, 24] use diffusion with guidance or conditioning to achieve controllable scene initialization. In contrast, we emphasize closed-loop, controllable adversarial behavior simulation based on real-world data initialization rather than scene initialization." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 242, + 326, + 254 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 242, + 326, + 254 + ], + "spans": [ + { + "bbox": [ + 131, + 242, + 326, + 254 + ], + "type": "text", + "content": "2.2 Safety-Critical Traffic Simulation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 262, + 482, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 262, + 482, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 482, + 441 + ], + "type": "text", + "content": "Safety-critical traffic generation plays a crucial role in training and evaluating AV systems, enhancing their capability to navigate diverse real-world scenarios and enhancing robustness. Gradient-based methods that leverage back-propagation to create safety-critical scenarios have been proposed to evaluate AV prediction and planning models [4,11]. Hanselmann et al. use kinematic gradients to modify vehicle trajectories, with the goal of improving the robustness of imitation learning planners [11]. Cao et al. have developed a model with differentiable dynamics, enabling the generation of realistic adversarial trajectories for trajectory prediction models through backpropagation techniques [4]. Black-box optimization approaches include perturbing actions based on kinematic bicycle models [33] and using Bayesian Optimization to create adversarial self-driving scenarios that escalate collision risks with simulated entities [1]. Zhang et al. target trajectory prediction models via white- and black-box attacks that adversarily perturb real driving trajectories [37]. For an extensive review of this topic, we direct readers to Ding et al. [9]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 441, + 483, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 441, + 483, + 549 + ], + "spans": [ + { + "bbox": [ + 130, + 441, + 483, + 549 + ], + "type": "text", + "content": "The field has recently seen advancements in data-driven methods for safety-critical scenario generation [26,34,36]. For instance, Xu et al. introduce a diffusion-based approach in CARLA, applying various adversarial optimization objectives to guide the diffusion process for safety-critical scenario generation [34]. Rempe et al. proposed STRIVE to utilize gradient-based adversarial optimization on the latent space, constrained by a graph-based CVAE traffic motion model, to generate realistic safety-critical scenarios for rule-based planners [26]. However, a common limitation in these approaches is the absence of closed-loop interaction, essential for accurately simulating interactive real-world driving." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 568, + 281, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 568, + 281, + 580 + ], + "spans": [ + { + "bbox": [ + 131, + 568, + 281, + 580 + ], + "type": "text", + "content": "3 Problem Formulation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "We consider a simulated interactive traffic scenario consisting of " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " agents; one is the ego vehicle controlled by the planner " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ", and the remaining " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "N - 1" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " are reactive agents modeled by a function " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ". Our objective is to create a safety-critical closed-loop collision simulation, where reactive agents demonstrate realistic, controllable behavior. Of the " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "N - 1" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " reactive agents, one or a subset is considered the adversarial agents (denoted as agent " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "), meant to collide with the ego vehicle." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "type": "text", + "content": "W.J. Chang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "The adversarial agent, formulated within the reactive agent model " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": ", is governed by an adversarial term designed (detailed in Sec. 5) to be both controllable and adversarial to the planner " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": ". This setup allows the adversarial agent to pose direct challenges to " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": ", testing its resilience in complex scenarios. Concurrently, the other non-adversarial agents, also controlled by " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " with varying parameters, emulate authentic, reactive behaviors, thus enriching the simulation scenario with realistic and diverse traffic conditions. The dual role of the adversarial agent and non-adversarial agents ensures that while it challenges " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": ", the overall simulation environment plausibly represents real-world driving conditions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 224, + 482, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 306 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 306 + ], + "type": "text", + "content": "At any given timestep " + }, + { + "bbox": [ + 130, + 224, + 482, + 306 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 224, + 482, + 306 + ], + "type": "text", + "content": ", the states of the " + }, + { + "bbox": [ + 130, + 224, + 482, + 306 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 224, + 482, + 306 + ], + "type": "text", + "content": " vehicles are represented as " + }, + { + "bbox": [ + 130, + 224, + 482, + 306 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_t = [\\mathbf{s}_t^1, \\dots, \\mathbf{s}_t^N]" + }, + { + "bbox": [ + 130, + 224, + 482, + 306 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 224, + 482, + 306 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_t^i = (x_t^i, y_t^i, v_t^i, \\theta_t^i)" + }, + { + "bbox": [ + 130, + 224, + 482, + 306 + ], + "type": "text", + "content": " indicates the 2D position, speed, and yaw of vehicle " + }, + { + "bbox": [ + 130, + 224, + 482, + 306 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 224, + 482, + 306 + ], + "type": "text", + "content": ". The corresponding actions for each vehicle are " + }, + { + "bbox": [ + 130, + 224, + 482, + 306 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_t = [\\mathbf{a}_t^1, \\dots, \\mathbf{a}_t^N]" + }, + { + "bbox": [ + 130, + 224, + 482, + 306 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 130, + 224, + 482, + 306 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_t^i = (\\dot{v}_t^i, \\dot{\\theta}_t^i)" + }, + { + "bbox": [ + 130, + 224, + 482, + 306 + ], + "type": "text", + "content": " representing the acceleration and yaw rate. To predict the state at the next timestep " + }, + { + "bbox": [ + 130, + 224, + 482, + 306 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 130, + 224, + 482, + 306 + ], + "type": "text", + "content": ", a transition function " + }, + { + "bbox": [ + 130, + 224, + 482, + 306 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 224, + 482, + 306 + ], + "type": "text", + "content": " is used, which computes " + }, + { + "bbox": [ + 130, + 224, + 482, + 306 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{t+1} = f(\\mathbf{s}_t, \\mathbf{a}_t)" + }, + { + "bbox": [ + 130, + 224, + 482, + 306 + ], + "type": "text", + "content": " based on current state and action. We adopt unicycle dynamics as the transition function." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 307, + 482, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 307, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 307, + 482, + 392 + ], + "type": "text", + "content": "Each agent's decision context is " + }, + { + "bbox": [ + 130, + 307, + 482, + 392 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_t^i" + }, + { + "bbox": [ + 130, + 307, + 482, + 392 + ], + "type": "text", + "content": ", which includes the agent-centric map " + }, + { + "bbox": [ + 130, + 307, + 482, + 392 + ], + "type": "inline_equation", + "content": "I^i" + }, + { + "bbox": [ + 130, + 307, + 482, + 392 + ], + "type": "text", + "content": " and the " + }, + { + "bbox": [ + 130, + 307, + 482, + 392 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{hist}}" + }, + { + "bbox": [ + 130, + 307, + 482, + 392 + ], + "type": "text", + "content": " historical states of neighboring vehicles from time " + }, + { + "bbox": [ + 130, + 307, + 482, + 392 + ], + "type": "inline_equation", + "content": "t - T_{\\mathrm{hist}}" + }, + { + "bbox": [ + 130, + 307, + 482, + 392 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 307, + 482, + 392 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 307, + 482, + 392 + ], + "type": "text", + "content": ", defined as " + }, + { + "bbox": [ + 130, + 307, + 482, + 392 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{t - T_{\\mathrm{hist}}:t} = \\{\\mathbf{s}_{t - T_{\\mathrm{hist}}}, \\dots, \\mathbf{s}_t\\}" + }, + { + "bbox": [ + 130, + 307, + 482, + 392 + ], + "type": "text", + "content": ". In closed-loop traffic simulation, each agent continuously generates and updates its trajectory based on the current decision context " + }, + { + "bbox": [ + 130, + 307, + 482, + 392 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_t^i" + }, + { + "bbox": [ + 130, + 307, + 482, + 392 + ], + "type": "text", + "content": ". After generating a trajectory, the simulation executes the first few steps of the planned actions before updating " + }, + { + "bbox": [ + 130, + 307, + 482, + 392 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_t^i" + }, + { + "bbox": [ + 130, + 307, + 482, + 392 + ], + "type": "text", + "content": " and re-planning. See Sec. 6.2 for more implementation details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 401, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 401, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 482, + 449 + ], + "type": "text", + "content": "Planner " + }, + { + "bbox": [ + 130, + 401, + 482, + 449 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 130, + 401, + 482, + 449 + ], + "type": "text", + "content": " The planner " + }, + { + "bbox": [ + 130, + 401, + 482, + 449 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 130, + 401, + 482, + 449 + ], + "type": "text", + "content": " determines the ego vehicle's future trajectory over a time horizon " + }, + { + "bbox": [ + 130, + 401, + 482, + 449 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 401, + 482, + 449 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 401, + 482, + 449 + ], + "type": "inline_equation", + "content": "t + T" + }, + { + "bbox": [ + 130, + 401, + 482, + 449 + ], + "type": "text", + "content": ". The planned state sequence is denoted by " + }, + { + "bbox": [ + 130, + 401, + 482, + 449 + ], + "type": "inline_equation", + "content": "s_{t:t + T}^{1} = \\pi (\\mathbf{c}_{t}^{1})" + }, + { + "bbox": [ + 130, + 401, + 482, + 449 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 401, + 482, + 449 + ], + "type": "inline_equation", + "content": "\\pi (\\mathbf{c}_t^1)" + }, + { + "bbox": [ + 130, + 401, + 482, + 449 + ], + "type": "text", + "content": " processes the historical states and map data within " + }, + { + "bbox": [ + 130, + 401, + 482, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_t^1" + }, + { + "bbox": [ + 130, + 401, + 482, + 449 + ], + "type": "text", + "content": " to plan future states based on the current scene context." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 458, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 458, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 131, + 458, + 482, + 555 + ], + "type": "text", + "content": "Reactive Agents " + }, + { + "bbox": [ + 131, + 458, + 482, + 555 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 131, + 458, + 482, + 555 + ], + "type": "text", + "content": " The reactive agent model " + }, + { + "bbox": [ + 131, + 458, + 482, + 555 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 131, + 458, + 482, + 555 + ], + "type": "text", + "content": ", parameterized by " + }, + { + "bbox": [ + 131, + 458, + 482, + 555 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 131, + 458, + 482, + 555 + ], + "type": "text", + "content": ", is designed to simulate the behavior of the " + }, + { + "bbox": [ + 131, + 458, + 482, + 555 + ], + "type": "inline_equation", + "content": "N - 1" + }, + { + "bbox": [ + 131, + 458, + 482, + 555 + ], + "type": "text", + "content": " non-ego vehicles, represented by the set " + }, + { + "bbox": [ + 131, + 458, + 482, + 555 + ], + "type": "inline_equation", + "content": "\\{s_{t:t + T}^i\\}_{i = 2}^N" + }, + { + "bbox": [ + 131, + 458, + 482, + 555 + ], + "type": "text", + "content": ". Each vehicle's state sequence, " + }, + { + "bbox": [ + 131, + 458, + 482, + 555 + ], + "type": "inline_equation", + "content": "s_{t:t + T}^i" + }, + { + "bbox": [ + 131, + 458, + 482, + 555 + ], + "type": "text", + "content": ", is generated by " + }, + { + "bbox": [ + 131, + 458, + 482, + 555 + ], + "type": "inline_equation", + "content": "g_{\\theta}(\\mathbf{c}_t^i,\\psi_i)" + }, + { + "bbox": [ + 131, + 458, + 482, + 555 + ], + "type": "text", + "content": ", which incorporates the decision context " + }, + { + "bbox": [ + 131, + 458, + 482, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_t^i" + }, + { + "bbox": [ + 131, + 458, + 482, + 555 + ], + "type": "text", + "content": " and a set of control parameters " + }, + { + "bbox": [ + 131, + 458, + 482, + 555 + ], + "type": "inline_equation", + "content": "\\psi_{i}" + }, + { + "bbox": [ + 131, + 458, + 482, + 555 + ], + "type": "text", + "content": " unique to each agent. These parameters " + }, + { + "bbox": [ + 131, + 458, + 482, + 555 + ], + "type": "inline_equation", + "content": "\\psi_{i}" + }, + { + "bbox": [ + 131, + 458, + 482, + 555 + ], + "type": "text", + "content": " enable the fine-tuning of individual behaviors within the simulation. In our approach, we train the model " + }, + { + "bbox": [ + 131, + 458, + 482, + 555 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 131, + 458, + 482, + 555 + ], + "type": "text", + "content": " on real-world driving data to ensure the trajectories it produces are not only controllable, supporting the generation of various safety-critical scenarios, but also realistic." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 571, + 385, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 571, + 385, + 583 + ], + "spans": [ + { + "bbox": [ + 131, + 571, + 385, + 583 + ], + "type": "text", + "content": "4 Diffusion Models for Traffic Simulation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 481, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 481, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 481, + 640 + ], + "type": "text", + "content": "For closed-loop safety-critical traffic simulation, the reactive agents, especially the adversarial agent, should be 1) controllable, and 2) realistic. With recent advances in controllable diffusion models [18,25,39], we adopt trajectory diffusion models to generate realistic simulations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": "We define the model's operational trajectory as " + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": ", which comprises both action and state sequences: " + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\tau \\coloneqq [\\tau_{a}, \\tau_{s}]" + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": ". Specifically, " + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\tau_{a} \\coloneqq [a_{0}, \\ldots, a_{T-1}]" + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": " represents the" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 135, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 135, + 90, + 447, + 102 + ], + "type": "text", + "content": "Safety-Critical Closed-Loop Traffic Simulation with Controllable Adversaries" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "sequence of actions, while " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\tau_{s} \\coloneqq [s_{1},\\ldots ,s_{T}]" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " denotes the corresponding sequence of states. Following the approach described in [39], our model predicts the action sequence " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\tau_{a}" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": ", and the state sequence " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\tau_{s}" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " can be derived starting from the initial state " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "s_0" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " and dynamic model " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 482, + 224 + ], + "type": "text", + "content": "A diffusion model generates a trajectory by reversing a process that incrementally adds noise. Starting with an actual trajectory " + }, + { + "bbox": [ + 130, + 164, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\tau_0" + }, + { + "bbox": [ + 130, + 164, + 482, + 224 + ], + "type": "text", + "content": " sampled from the data distribution " + }, + { + "bbox": [ + 130, + 164, + 482, + 224 + ], + "type": "inline_equation", + "content": "q(\\tau_0)" + }, + { + "bbox": [ + 130, + 164, + 482, + 224 + ], + "type": "text", + "content": ", a sequence of increasingly noisy trajectories " + }, + { + "bbox": [ + 130, + 164, + 482, + 224 + ], + "type": "inline_equation", + "content": "(\\tau_{1},\\tau_{2},\\dots,\\tau_{K})" + }, + { + "bbox": [ + 130, + 164, + 482, + 224 + ], + "type": "text", + "content": " is produced via a forward noisng process. Each trajectory " + }, + { + "bbox": [ + 130, + 164, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\tau_{k}" + }, + { + "bbox": [ + 130, + 164, + 482, + 224 + ], + "type": "text", + "content": " at step " + }, + { + "bbox": [ + 130, + 164, + 482, + 224 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 164, + 482, + 224 + ], + "type": "text", + "content": " is generated by adding Gaussian noise parameterized by a predefined variance schedule " + }, + { + "bbox": [ + 130, + 164, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\beta_{k}" + }, + { + "bbox": [ + 130, + 164, + 482, + 224 + ], + "type": "text", + "content": " [14]:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 246, + 234, + 482, + 266 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 234, + 482, + 266 + ], + "spans": [ + { + "bbox": [ + 246, + 234, + 482, + 266 + ], + "type": "interline_equation", + "content": "q \\left(\\tau_ {1: K} \\mid \\tau_ {0}\\right) := \\prod_ {k = 1} ^ {K} q \\left(\\tau_ {k} \\mid \\tau_ {k - 1}\\right), \\tag {1}", + "image_path": "c54b6c4595afbe42ecb9206c810cc87e36638e3b0716ffddf30840c892bc3704.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 222, + 274, + 481, + 287 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 274, + 481, + 287 + ], + "spans": [ + { + "bbox": [ + 222, + 274, + 481, + 287 + ], + "type": "interline_equation", + "content": "q \\left(\\tau_ {k} \\mid \\tau_ {k - 1}\\right) := \\mathcal {N} \\left(\\tau_ {k}; \\sqrt {1 - \\beta_ {k}} \\tau_ {k - 1}, \\beta_ {k} \\mathbf {I}\\right). \\tag {2}", + "image_path": "8c7a309ac32ea96a0d4efac31573ca61643c3bd59551777d8686ef3301e0aef5.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 292, + 482, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 292, + 482, + 352 + ], + "spans": [ + { + "bbox": [ + 131, + 292, + 482, + 352 + ], + "type": "text", + "content": "The noising process gradually obscures the data, where the final noisy version " + }, + { + "bbox": [ + 131, + 292, + 482, + 352 + ], + "type": "inline_equation", + "content": "q(\\tau_K)" + }, + { + "bbox": [ + 131, + 292, + 482, + 352 + ], + "type": "text", + "content": " approaches " + }, + { + "bbox": [ + 131, + 292, + 482, + 352 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\tau_K; \\mathbf{0}, \\mathbf{I})" + }, + { + "bbox": [ + 131, + 292, + 482, + 352 + ], + "type": "text", + "content": ". The trajectory generation process is then achieved by learning the reverse of this noising process. Given a noisy trajectory " + }, + { + "bbox": [ + 131, + 292, + 482, + 352 + ], + "type": "inline_equation", + "content": "\\tau_K" + }, + { + "bbox": [ + 131, + 292, + 482, + 352 + ], + "type": "text", + "content": ", the model learns to denoise it back to " + }, + { + "bbox": [ + 131, + 292, + 482, + 352 + ], + "type": "inline_equation", + "content": "\\tau_0" + }, + { + "bbox": [ + 131, + 292, + 482, + 352 + ], + "type": "text", + "content": " through a sequence of reverse steps. Each reverse step is modeled as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 215, + 361, + 481, + 374 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 215, + 361, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 215, + 361, + 481, + 374 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(\\tau_ {k - 1} \\mid \\tau_ {k}, \\mathbf {c}\\right) := \\mathcal {N} \\left(\\tau_ {k - 1}; \\mu_ {\\theta} \\left(\\tau_ {k}, k, \\mathbf {c}\\right), \\Sigma_ {k}\\right), \\tag {3}", + "image_path": "f7caa48611fecc5a383be1e1ba3472eda8391b35f4100df657a68f2c819f2f1c.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 382, + 479, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 382, + 479, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 382, + 479, + 429 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 382, + 479, + 429 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 382, + 479, + 429 + ], + "type": "text", + "content": " are learned functions that predict the mean " + }, + { + "bbox": [ + 130, + 382, + 479, + 429 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 130, + 382, + 479, + 429 + ], + "type": "text", + "content": " of the reverse step, and " + }, + { + "bbox": [ + 130, + 382, + 479, + 429 + ], + "type": "inline_equation", + "content": "\\Sigma_{k}" + }, + { + "bbox": [ + 130, + 382, + 479, + 429 + ], + "type": "text", + "content": " is a fixed schedule. By iteratively applying the reverse process, the model learns a trajectory distribution, effectively generating a plausible future trajectory from a noisy start." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 430, + 479, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 479, + 466 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 479, + 466 + ], + "type": "text", + "content": "During the trajectory prediction phase, the model estimates the final clean trajectory denoted by " + }, + { + "bbox": [ + 130, + 430, + 479, + 466 + ], + "type": "inline_equation", + "content": "\\hat{\\tau}_0" + }, + { + "bbox": [ + 130, + 430, + 479, + 466 + ], + "type": "text", + "content": ". This estimated trajectory is used to compute the mean " + }, + { + "bbox": [ + 130, + 430, + 479, + 466 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 130, + 430, + 479, + 466 + ], + "type": "text", + "content": " as described in [21]. For more details, see supplementary material Sec. D." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 483, + 474, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 483, + 474, + 498 + ], + "spans": [ + { + "bbox": [ + 131, + 483, + 474, + 498 + ], + "type": "text", + "content": "5 Diffusion Models for Safety-Critical Traffic Simulation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 509, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 482, + 581 + ], + "type": "text", + "content": "The diffusion model, once trained on realistic trajectory data, inherently reflects the behavioral patterns present in its training distribution. However, to effectively simulate and analyze safety-critical scenarios, there is a crucial need for a mechanism that allows for the controlled manipulation of agent behaviors [18,39]. This is particularly important for generating adversarial behaviors and ensuring scene consistency in simulations." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 597, + 288, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 597, + 288, + 611 + ], + "spans": [ + { + "bbox": [ + 131, + 597, + 288, + 611 + ], + "type": "text", + "content": "5.1 Guiding Reactive Agents" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Our approach specifically introduces guidance to the sampled trajectories at each denoising step, aligning them with predefined objectives " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "J(\\tau)" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ". The concept of guidance involves using the gradient of " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " to subtly perturb the predicted mean of the model at each denoising step. This process enables the generation" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "type": "text", + "content": "W.J. Chang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 118, + 479, + 220 + ], + "blocks": [ + { + "bbox": [ + 135, + 118, + 479, + 220 + ], + "lines": [ + { + "bbox": [ + 135, + 118, + 479, + 220 + ], + "spans": [ + { + "bbox": [ + 135, + 118, + 479, + 220 + ], + "type": "image", + "image_path": "b756d8af192c285b659b43e12a9984040344b6a5e71fd95b94968fad6c5642ee.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 231, + 482, + 277 + ], + "lines": [ + { + "bbox": [ + 131, + 231, + 482, + 277 + ], + "spans": [ + { + "bbox": [ + 131, + 231, + 482, + 277 + ], + "type": "text", + "content": "Fig.2: Guided Diffusion Process for the Adversarial Agent. This process optimizes the adversarial agent's trajectory using the adversarial cost function " + }, + { + "bbox": [ + 131, + 231, + 482, + 277 + ], + "type": "inline_equation", + "content": "J_{\\mathrm{adv}}" + }, + { + "bbox": [ + 131, + 231, + 482, + 277 + ], + "type": "text", + "content": " to the ego vehicle. In particular, we introduce " + }, + { + "bbox": [ + 131, + 231, + 482, + 277 + ], + "type": "inline_equation", + "content": "J_{\\mathrm{control}}" + }, + { + "bbox": [ + 131, + 231, + 482, + 277 + ], + "type": "text", + "content": " to vary the adversarial behavior. Simultaneously, it applies regularization through " + }, + { + "bbox": [ + 131, + 231, + 482, + 277 + ], + "type": "inline_equation", + "content": "J_{\\mathrm{reg}}" + }, + { + "bbox": [ + 131, + 231, + 482, + 277 + ], + "type": "text", + "content": " for maintaining realism." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 298, + 482, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 298, + 482, + 348 + ], + "spans": [ + { + "bbox": [ + 130, + 298, + 482, + 348 + ], + "type": "text", + "content": "of trajectories that not only reflect realistic behavior but also cater to specific simulation needs, such as adversarial testing and maintaining scene consistency over extended periods. We adopt the reconstruction guidance (clean guidance) introduced from [16, 25]:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 253, + 354, + 482, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 354, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 253, + 354, + 482, + 369 + ], + "type": "interline_equation", + "content": "\\tilde {\\tau} _ {0} = \\hat {\\tau} _ {0} - \\alpha \\Sigma_ {k} \\nabla_ {\\tau_ {k}} J (\\hat {\\tau} _ {0}) \\tag {4}", + "image_path": "1f17754b45168b493270f9a52f8a79cabcb5e813f4353761535523bdf8c93644.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 375, + 482, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 375, + 482, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 375, + 482, + 399 + ], + "type": "text", + "content": "This strategy improves guidance robustness, yielding smoother, more stable trajectories without the usual numerical issues from noisy data." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 399, + 482, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 482, + 435 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 482, + 435 + ], + "type": "text", + "content": "In practice, diversifying the behavior of adversarial agents within the same scenarios is crucial for a thorough assessment of AVs. Despite the significance of this challenge, it remains largely unexplored in previous works [26,39]" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 435, + 482, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 435, + 482, + 517 + ], + "spans": [ + { + "bbox": [ + 130, + 435, + 482, + 517 + ], + "type": "text", + "content": "The loss function for the non-reactive agents, " + }, + { + "bbox": [ + 130, + 435, + 482, + 517 + ], + "type": "inline_equation", + "content": "J(\\tau)" + }, + { + "bbox": [ + 130, + 435, + 482, + 517 + ], + "type": "text", + "content": ", consists of a collision term " + }, + { + "bbox": [ + 130, + 435, + 482, + 517 + ], + "type": "inline_equation", + "content": "J_{\\mathrm{coll}}" + }, + { + "bbox": [ + 130, + 435, + 482, + 517 + ], + "type": "text", + "content": ", which encourages collisions between the adversarial agent and the ego agent, two control terms " + }, + { + "bbox": [ + 130, + 435, + 482, + 517 + ], + "type": "inline_equation", + "content": "J_{\\mathrm{v}}" + }, + { + "bbox": [ + 130, + 435, + 482, + 517 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 435, + 482, + 517 + ], + "type": "inline_equation", + "content": "J_{\\mathrm{ttc}}" + }, + { + "bbox": [ + 130, + 435, + 482, + 517 + ], + "type": "text", + "content": ", which control the relative speed and time-to-collision between the ego and adversarial agent respectively, a regularization term " + }, + { + "bbox": [ + 130, + 435, + 482, + 517 + ], + "type": "inline_equation", + "content": "J_{\\mathrm{Gauss}}" + }, + { + "bbox": [ + 130, + 435, + 482, + 517 + ], + "type": "text", + "content": ", which discourages collisions between the reactive agents, and a route guidance term " + }, + { + "bbox": [ + 130, + 435, + 482, + 517 + ], + "type": "inline_equation", + "content": "J_{\\mathrm{route}}" + }, + { + "bbox": [ + 130, + 435, + 482, + 517 + ], + "type": "text", + "content": ", which discourages the reactive agents from going outside the road:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 208, + 518, + 482, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 518, + 482, + 548 + ], + "spans": [ + { + "bbox": [ + 208, + 518, + 482, + 548 + ], + "type": "interline_equation", + "content": "J (\\tau) = \\rho \\underbrace {\\left(J _ {\\mathrm {c o l l}} + J _ {\\mathrm {v}} + J _ {\\mathrm {t t c}}\\right)} _ {J _ {\\mathrm {a d v}} (\\tau)} + \\underbrace {J _ {\\text {r o u t e}} + J _ {\\text {G a u s s}}} _ {J _ {\\text {r e g}} (\\tau)}, \\tag {5}", + "image_path": "fb1cf237592a8509d88fbe86516f3b2b88fabdcaa6b6bf9db277d677f657373c.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 552, + 482, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 552, + 482, + 588 + ], + "spans": [ + { + "bbox": [ + 130, + 552, + 482, + 588 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 552, + 482, + 588 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 130, + 552, + 482, + 588 + ], + "type": "text", + "content": " denotes a scalar weight that determines whether a reactive agent behaves adversarially towards the ego agent, i.e., whether it attempts to collide with the ego agent." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 604, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 604, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 604, + 482, + 628 + ], + "type": "text", + "content": "Collision with Planner We define " + }, + { + "bbox": [ + 130, + 604, + 482, + 628 + ], + "type": "inline_equation", + "content": "J_{\\mathrm{coll}}" + }, + { + "bbox": [ + 130, + 604, + 482, + 628 + ], + "type": "text", + "content": " to encourage the collision between the adversarial agent and the ego agent, given by:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 266, + 635, + 482, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 635, + 482, + 669 + ], + "spans": [ + { + "bbox": [ + 266, + 635, + 482, + 669 + ], + "type": "interline_equation", + "content": "J _ {\\text {c o l l}} = - \\sum_ {t = 1} ^ {T} d (t), \\tag {6}", + "image_path": "65c2e15629e06fb476cf2cfeb1ab9e7ef4adf9e2ed62049693f2c80bb3f04911.jpg" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 135, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 135, + 91, + 448, + 102 + ], + "type": "text", + "content": "Safety-Critical Closed-Loop Traffic Simulation with Controllable Adversaries" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "inline_equation", + "content": "d(t)" + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": " represents the distance between the ego and the adversarial agent at each time step of the planning horizon " + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": ". The adversarial agent is either pre-selected based on lane proximity or dynamically selected based on the distance to the ego agent. Details are in the supplementary Sec. D.4." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 180, + 482, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 180, + 482, + 228 + ], + "spans": [ + { + "bbox": [ + 130, + 180, + 482, + 228 + ], + "type": "text", + "content": "Safety Criticality of Collisions We control the relative speed " + }, + { + "bbox": [ + 130, + 180, + 482, + 228 + ], + "type": "inline_equation", + "content": "J_{v}" + }, + { + "bbox": [ + 130, + 180, + 482, + 228 + ], + "type": "text", + "content": " between the ego and adversary at each time step " + }, + { + "bbox": [ + 130, + 180, + 482, + 228 + ], + "type": "inline_equation", + "content": "(v_{t}^{1}" + }, + { + "bbox": [ + 130, + 180, + 482, + 228 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 180, + 482, + 228 + ], + "type": "inline_equation", + "content": "v_{t}^{a}" + }, + { + "bbox": [ + 130, + 180, + 482, + 228 + ], + "type": "text", + "content": "), and the time-to-collision (TTC) cost " + }, + { + "bbox": [ + 130, + 180, + 482, + 228 + ], + "type": "inline_equation", + "content": "J_{\\mathrm{ttc}}" + }, + { + "bbox": [ + 130, + 180, + 482, + 228 + ], + "type": "text", + "content": " [22] to control the safety criticality of potential collisions, with the latter given by:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 225, + 236, + 482, + 268 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 236, + 482, + 268 + ], + "spans": [ + { + "bbox": [ + 225, + 236, + 482, + 268 + ], + "type": "interline_equation", + "content": "J _ {\\mathrm {t t c}} = \\sum_ {t = 1} ^ {T} - \\exp \\left(- \\frac {\\tilde {t} _ {\\operatorname {c o l} (t)} ^ {2}}{2 \\lambda_ {t}} - \\frac {\\tilde {d} _ {\\operatorname {c o l} (t)} ^ {2}}{2 \\lambda_ {d}}\\right), \\tag {7}", + "image_path": "0e4fb220320d119d2aeb583baa65e1f622db80480a8badf8862f3865d12aae47.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 276, + 482, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 276, + 482, + 338 + ], + "spans": [ + { + "bbox": [ + 130, + 276, + 482, + 338 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 276, + 482, + 338 + ], + "type": "inline_equation", + "content": "\\tilde{t}_{\\mathrm{col}(t)}" + }, + { + "bbox": [ + 130, + 276, + 482, + 338 + ], + "type": "text", + "content": " is the time to collision at time " + }, + { + "bbox": [ + 130, + 276, + 482, + 338 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 276, + 482, + 338 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 276, + 482, + 338 + ], + "type": "inline_equation", + "content": "\\tilde{d}_{\\mathrm{col}(t)}" + }, + { + "bbox": [ + 130, + 276, + 482, + 338 + ], + "type": "text", + "content": " is the distance to collision and " + }, + { + "bbox": [ + 130, + 276, + 482, + 338 + ], + "type": "inline_equation", + "content": "\\lambda_t" + }, + { + "bbox": [ + 130, + 276, + 482, + 338 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 276, + 482, + 338 + ], + "type": "inline_equation", + "content": "\\lambda_d" + }, + { + "bbox": [ + 130, + 276, + 482, + 338 + ], + "type": "text", + "content": " are bandwidth parameters for time and distance. This formula uses a constant velocity assumption. Intuitively, the time-to-collision cost favors scenarios with high relative speeds and challenging collision angles for the ego vehicle to avoid. For details of " + }, + { + "bbox": [ + 130, + 276, + 482, + 338 + ], + "type": "inline_equation", + "content": "J_v" + }, + { + "bbox": [ + 130, + 276, + 482, + 338 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 276, + 482, + 338 + ], + "type": "inline_equation", + "content": "J_{\\mathrm{ttc}}" + }, + { + "bbox": [ + 130, + 276, + 482, + 338 + ], + "type": "text", + "content": ", see supplementary Sec. C.2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 352, + 482, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 352, + 482, + 423 + ], + "spans": [ + { + "bbox": [ + 130, + 352, + 482, + 423 + ], + "type": "text", + "content": "Route Guidance Given an agent's trajectory " + }, + { + "bbox": [ + 130, + 352, + 482, + 423 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 352, + 482, + 423 + ], + "type": "text", + "content": " and the corresponding route " + }, + { + "bbox": [ + 130, + 352, + 482, + 423 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 352, + 482, + 423 + ], + "type": "text", + "content": "-the predefined path on a lane graph from its starting point to its destination—we compute the normal distance of each point " + }, + { + "bbox": [ + 130, + 352, + 482, + 423 + ], + "type": "inline_equation", + "content": "\\tau_{t}" + }, + { + "bbox": [ + 130, + 352, + 482, + 423 + ], + "type": "text", + "content": " on the trajectory to the route at each timestep. We then penalize deviations from the route that exceed a predefined margin " + }, + { + "bbox": [ + 130, + 352, + 482, + 423 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 352, + 482, + 423 + ], + "type": "text", + "content": ". This process is captured by the following route guidance cost function:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 217, + 434, + 482, + 465 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 434, + 482, + 465 + ], + "spans": [ + { + "bbox": [ + 217, + 434, + 482, + 465 + ], + "type": "interline_equation", + "content": "J _ {\\text {r o u t e}} (\\tau , r) = \\sum_ {t = 1} ^ {T} \\max \\left(0, \\left| d _ {n} \\left(\\tau_ {t}, r\\right) - d _ {m} \\right|\\right), \\tag {8}", + "image_path": "03022a84dd87be43159f494b039be7e91c0bf17a370f04290aaaede5bdd14415.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 469, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 469, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 469, + 482, + 555 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 469, + 482, + 555 + ], + "type": "inline_equation", + "content": "d_{n}(\\tau_{n},r)" + }, + { + "bbox": [ + 130, + 469, + 482, + 555 + ], + "type": "text", + "content": " denotes the normal distance from the point " + }, + { + "bbox": [ + 130, + 469, + 482, + 555 + ], + "type": "inline_equation", + "content": "\\tau_{t}" + }, + { + "bbox": [ + 130, + 469, + 482, + 555 + ], + "type": "text", + "content": " on the trajectory to the nearest point on the route " + }, + { + "bbox": [ + 130, + 469, + 482, + 555 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 469, + 482, + 555 + ], + "type": "text", + "content": " at timestep " + }, + { + "bbox": [ + 130, + 469, + 482, + 555 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 469, + 482, + 555 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 469, + 482, + 555 + ], + "type": "inline_equation", + "content": "d_{m}" + }, + { + "bbox": [ + 130, + 469, + 482, + 555 + ], + "type": "text", + "content": " represents the acceptable deviation margin from the route. In contrast to the off-road loss in prior studies [39], our proposed route guidance system more effectively indicates each agent's intended path, improving adherence to traffic rules as demonstrated in Sec. 6. The flexibility of route guidance supports diverse agent interactions, such as modifying routes to encourage lane changes among reactive agents [29]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 569, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 628 + ], + "type": "text", + "content": "Gaussian Collision Guidance Given the trajectories of agents, we calculate the Gaussian distance for each pair of agents " + }, + { + "bbox": [ + 130, + 569, + 482, + 628 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 130, + 569, + 482, + 628 + ], + "type": "text", + "content": " at each timestep " + }, + { + "bbox": [ + 130, + 569, + 482, + 628 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 569, + 482, + 628 + ], + "type": "text", + "content": " from 1 to " + }, + { + "bbox": [ + 130, + 569, + 482, + 628 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 569, + 482, + 628 + ], + "type": "text", + "content": ". The Gaussian distance between the agents takes into account both the tangential " + }, + { + "bbox": [ + 130, + 569, + 482, + 628 + ], + "type": "inline_equation", + "content": "(d_t)" + }, + { + "bbox": [ + 130, + 569, + 482, + 628 + ], + "type": "text", + "content": " and normal " + }, + { + "bbox": [ + 130, + 569, + 482, + 628 + ], + "type": "inline_equation", + "content": "(d_n)" + }, + { + "bbox": [ + 130, + 569, + 482, + 628 + ], + "type": "text", + "content": " components of the projected distances. The aggregated Gaussian distance is:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 195, + 635, + 482, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 635, + 482, + 669 + ], + "spans": [ + { + "bbox": [ + 195, + 635, + 482, + 669 + ], + "type": "interline_equation", + "content": "J _ {\\mathrm {G a u s s}} = \\sum_ {t = 1} ^ {T} \\sum_ {i, j} ^ {N} \\exp \\left(- \\frac {1}{2 \\sigma^ {2}} \\left(\\lambda \\cdot d _ {t} ^ {i j} (t) ^ {2} + d _ {n} ^ {i j} (t) ^ {2}\\right)\\right) \\tag {9}", + "image_path": "71e7f551711ec0c73930762d84889dc7eb56cf1b07528521c59d22e659a23cca.jpg" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "type": "text", + "content": "W.J. Chang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 116, + 476, + 224 + ], + "blocks": [ + { + "bbox": [ + 138, + 116, + 476, + 224 + ], + "lines": [ + { + "bbox": [ + 138, + 116, + 476, + 224 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 476, + 224 + ], + "type": "image", + "image_path": "7312088dc904c464486e66a5d4db329cf0ad414cb7fd177954ce8dd59f127021.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 236, + 482, + 270 + ], + "lines": [ + { + "bbox": [ + 130, + 236, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 482, + 270 + ], + "type": "text", + "content": "Fig. 3: Framework for Partial Diffusion. We generate proposals based on domain knowledge (e.g., collision types). Users can adjust noise levels to balance between user control and the model's data distribution." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 285, + 482, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 482, + 381 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 482, + 381 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 285, + 482, + 381 + ], + "type": "inline_equation", + "content": "d_t^{ij}(t)" + }, + { + "bbox": [ + 130, + 285, + 482, + 381 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 285, + 482, + 381 + ], + "type": "inline_equation", + "content": "d_n^{ij}(t)" + }, + { + "bbox": [ + 130, + 285, + 482, + 381 + ], + "type": "text", + "content": " represent the tangential and normal distances from agent " + }, + { + "bbox": [ + 130, + 285, + 482, + 381 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 285, + 482, + 381 + ], + "type": "text", + "content": "'s trajectory point at time " + }, + { + "bbox": [ + 130, + 285, + 482, + 381 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 285, + 482, + 381 + ], + "type": "text", + "content": " to agent " + }, + { + "bbox": [ + 130, + 285, + 482, + 381 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 285, + 482, + 381 + ], + "type": "text", + "content": "'s heading axis, respectively, and " + }, + { + "bbox": [ + 130, + 285, + 482, + 381 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 285, + 482, + 381 + ], + "type": "text", + "content": " is the standard deviation for these distances. In this formulation, " + }, + { + "bbox": [ + 130, + 285, + 482, + 381 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 285, + 482, + 381 + ], + "type": "text", + "content": " is a scaling factor applied to the tangential distance " + }, + { + "bbox": [ + 130, + 285, + 482, + 381 + ], + "type": "inline_equation", + "content": "d_t^{ij}(t)" + }, + { + "bbox": [ + 130, + 285, + 482, + 381 + ], + "type": "text", + "content": ". This approach contrasts with the disk approximation method, which primarily penalizes the Euclidean distance between agents. By accounting for both tangential and normal components, the Gaussian collision distance method significantly reduces collision rate, which we discuss in Sec. 6." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 402, + 389, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 402, + 389, + 415 + ], + "spans": [ + { + "bbox": [ + 131, + 402, + 389, + 415 + ], + "type": "text", + "content": "5.2 Partial Diffusion: Controlling Collision Types" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 425, + 482, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 425, + 482, + 484 + ], + "spans": [ + { + "bbox": [ + 130, + 425, + 482, + 484 + ], + "type": "text", + "content": "We introduce a novel approach through a partial diffusion process, utilizing trajectory proposals to initiate the diffusion process. This methodology enables the variation in collision types by the adversarial agent within the diffusion, tailoring the adversarial outcomes to specific evaluation needs, the results are discussed in Sec. 6.6." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 485, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 482, + 594 + ], + "type": "text", + "content": "Figure 3 illustrates our framework, which is divided into three main steps to generate trajectory proposals for various collision scenarios. First, we create initial trajectory proposals " + }, + { + "bbox": [ + 130, + 485, + 482, + 594 + ], + "type": "inline_equation", + "content": "(\\tau_0)" + }, + { + "bbox": [ + 130, + 485, + 482, + 594 + ], + "type": "text", + "content": " aimed at capturing different types of collisions. The next critical step involves setting the partial diffusion ratio " + }, + { + "bbox": [ + 130, + 485, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 485, + 482, + 594 + ], + "type": "text", + "content": ", which defines the specific point in the process, " + }, + { + "bbox": [ + 130, + 485, + 482, + 594 + ], + "type": "inline_equation", + "content": "k_{p} = \\gamma \\cdot K" + }, + { + "bbox": [ + 130, + 485, + 482, + 594 + ], + "type": "text", + "content": ", at which we start modifying the trajectory. Starting from step " + }, + { + "bbox": [ + 130, + 485, + 482, + 594 + ], + "type": "inline_equation", + "content": "k_{p}" + }, + { + "bbox": [ + 130, + 485, + 482, + 594 + ], + "type": "text", + "content": ", we adjust the trajectory by adding a precise level of Gaussian noise " + }, + { + "bbox": [ + 130, + 485, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim N(0,I)" + }, + { + "bbox": [ + 130, + 485, + 482, + 594 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 130, + 485, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\hat{\\tau}_{k_p} = \\sqrt{\\bar{\\alpha}_{k_p}}\\tau_0 + \\sqrt{1 - \\bar{\\alpha}_{k_p}}\\epsilon" + }, + { + "bbox": [ + 130, + 485, + 482, + 594 + ], + "type": "text", + "content": ". The final stages include removing noise and using guided diffusion for the rest of the " + }, + { + "bbox": [ + 130, + 485, + 482, + 594 + ], + "type": "inline_equation", + "content": "k_{p}" + }, + { + "bbox": [ + 130, + 485, + 482, + 594 + ], + "type": "text", + "content": " steps to refine the trajectory into a realistic path that suits our collision scenario goals." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "To generate the trajectory proposals, we develop a rule-based approach in which we first identify the centerlines of the ego and adversarial agent and then search for potential intersections of their respective centerlines. If such an intersection exists, we generate the proposals by selecting an acceleration value and lateral offset from the centerline that is likely to cause the desired collision type based on the projected plan of the ego agent. Note that the trajectory" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 135, + 90, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 90, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 135, + 90, + 448, + 102 + ], + "type": "text", + "content": "Safety-Critical Closed-Loop Traffic Simulation with Controllable Adversaries" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 141, + 118, + 248, + 202 + ], + "blocks": [ + { + "bbox": [ + 141, + 118, + 248, + 202 + ], + "lines": [ + { + "bbox": [ + 141, + 118, + 248, + 202 + ], + "spans": [ + { + "bbox": [ + 141, + 118, + 248, + 202 + ], + "type": "image", + "image_path": "0f1b5f6fd3666f168a6f85918ace569a8d8c85b1afd49f40ce56bdb58db5c66c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 179, + 207, + 201, + 220 + ], + "blocks": [ + { + "bbox": [ + 179, + 207, + 201, + 220 + ], + "lines": [ + { + "bbox": [ + 179, + 207, + 201, + 220 + ], + "spans": [ + { + "bbox": [ + 179, + 207, + 201, + 220 + ], + "type": "image", + "image_path": "3b09fa2c72b8599715fa729276012f5d7fec63bfa328ac2b18582f5f7da19fb7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 203, + 208, + 219, + 217 + ], + "lines": [ + { + "bbox": [ + 203, + 208, + 219, + 217 + ], + "spans": [ + { + "bbox": [ + 203, + 208, + 219, + 217 + ], + "type": "text", + "content": "Ego" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 220, + 208, + 235, + 217 + ], + "blocks": [ + { + "bbox": [ + 220, + 208, + 235, + 217 + ], + "lines": [ + { + "bbox": [ + 220, + 208, + 235, + 217 + ], + "spans": [ + { + "bbox": [ + 220, + 208, + 235, + 217 + ], + "type": "image", + "image_path": "a693a367c2b58420629a3e21b76ec3eff0c4557317a3525d8f10c72cf706b76d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 231, + 482, + 241 + ], + "lines": [ + { + "bbox": [ + 131, + 231, + 482, + 241 + ], + "spans": [ + { + "bbox": [ + 131, + 231, + 482, + 241 + ], + "type": "text", + "content": "Fig. 4: Partial Diffusion Results of Rule-Based Planner on NuScenes Dataset." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 264, + 118, + 359, + 202 + ], + "blocks": [ + { + "bbox": [ + 264, + 118, + 359, + 202 + ], + "lines": [ + { + "bbox": [ + 264, + 118, + 359, + 202 + ], + "spans": [ + { + "bbox": [ + 264, + 118, + 359, + 202 + ], + "type": "image", + "image_path": "cc07f46929c87ed2c587b8550a66dd8b64c55793cffa7689f5d2f6b1c3ae0728.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 288, + 209, + 317, + 217 + ], + "lines": [ + { + "bbox": [ + 288, + 209, + 317, + 217 + ], + "spans": [ + { + "bbox": [ + 288, + 209, + 317, + 217 + ], + "type": "text", + "content": "Non-Adv" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 263, + 208, + 285, + 217 + ], + "blocks": [ + { + "bbox": [ + 238, + 209, + 259, + 217 + ], + "lines": [ + { + "bbox": [ + 238, + 209, + 259, + 217 + ], + "spans": [ + { + "bbox": [ + 238, + 209, + 259, + 217 + ], + "type": "text", + "content": "Adv" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 263, + 208, + 285, + 217 + ], + "lines": [ + { + "bbox": [ + 263, + 208, + 285, + 217 + ], + "spans": [ + { + "bbox": [ + 263, + 208, + 285, + 217 + ], + "type": "image", + "image_path": "1336d5febd7950e8f3440f812f19ef859afb8486895c3a430ecfbefb04c82b18.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 325, + 208, + 365, + 217 + ], + "blocks": [ + { + "bbox": [ + 325, + 208, + 365, + 217 + ], + "lines": [ + { + "bbox": [ + 325, + 208, + 365, + 217 + ], + "spans": [ + { + "bbox": [ + 325, + 208, + 365, + 217 + ], + "type": "image", + "image_path": "d05a2e9ddbb5cec801cafc48acef0785d13fce16ac5a54863ee58c5e8f7d148a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 367, + 118, + 472, + 202 + ], + "blocks": [ + { + "bbox": [ + 367, + 118, + 472, + 202 + ], + "lines": [ + { + "bbox": [ + 367, + 118, + 472, + 202 + ], + "spans": [ + { + "bbox": [ + 367, + 118, + 472, + 202 + ], + "type": "image", + "image_path": "316a2f1513fbe5fb695fe3dbe0527b3df2845e09a34b83924a11c3389ec3fff8.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 370, + 209, + 432, + 217 + ], + "lines": [ + { + "bbox": [ + 370, + 209, + 432, + 217 + ], + "spans": [ + { + "bbox": [ + 370, + 209, + 432, + 217 + ], + "type": "text", + "content": "Trajectory Proposal" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 242, + 481, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 242, + 481, + 276 + ], + "spans": [ + { + "bbox": [ + 130, + 242, + 481, + 276 + ], + "type": "text", + "content": "The safety-critical scenarios show the framework's ability to create realistic and challenging situations, varying collision types based on different trajectory proposals. The gradient lines reflects the planned trajectories in the next 3.2 seconds." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 290, + 479, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 290, + 479, + 314 + ], + "spans": [ + { + "bbox": [ + 130, + 290, + 479, + 314 + ], + "type": "text", + "content": "proposals are updated in a closed-loop manner to account for the interaction between the ego and the adversarial agent." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 315, + 480, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 315, + 480, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 480, + 361 + ], + "type": "text", + "content": "This method allows for precise control over the diffusion trajectory, enabling adversarial agents to create customized collision scenarios. Users can adjust " + }, + { + "bbox": [ + 130, + 315, + 480, + 361 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 315, + 480, + 361 + ], + "type": "text", + "content": " to fine-tune the balance between explicit control and the model's trained data distribution." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 131, + 381, + 230, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 381, + 230, + 395 + ], + "spans": [ + { + "bbox": [ + 131, + 381, + 230, + 395 + ], + "type": "text", + "content": "6 Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 407, + 480, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 480, + 443 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 480, + 443 + ], + "type": "text", + "content": "We validate the efficacy of our proposed framework via experiments with real-world driving data. Our results demonstrate that the framework can generate realistic and controllable adversarial behavior to challenge the planner." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 131, + 462, + 200, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 462, + 200, + 472 + ], + "spans": [ + { + "bbox": [ + 131, + 462, + 200, + 472 + ], + "type": "text", + "content": "6.1 Dataset" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 483, + 480, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 483, + 480, + 553 + ], + "spans": [ + { + "bbox": [ + 130, + 483, + 480, + 553 + ], + "type": "text", + "content": "We conduct our experiments on two large-scale real-world driving datasets: nuScenes [2], which consists of 5.5 hours of driving data from two cities, and nuPlan, which consists of 1500 hours of driving data from four cities. We train the model on scenes from the nuScenes train split and evaluate it on the scenes from the nuScenes validation splits and nuPlan mini validation splits. We focus on vehicle-to-vehicle interactions." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 131, + 573, + 279, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 573, + 279, + 585 + ], + "spans": [ + { + "bbox": [ + 131, + 573, + 279, + 585 + ], + "type": "text", + "content": "6.2 Implementation Details" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 130, + 594, + 480, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 480, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 480, + 640 + ], + "type": "text", + "content": "Baselines. We compare our approach against STRIVE [26] using their open-source implementation and our re-implementation of DiffScene [34]. STRIVE is recognized for its proficiency in generating adversarial safety-critical scenarios using a learned traffic model and adversarial optimization in the latent space." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 130, + 642, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 642, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 642, + 481, + 665 + ], + "type": "text", + "content": "Planner. Our experiments utilize a range of different planners: 1) A rule-based planner, as implemented in STRIVE [26], which operates on the lane graph" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "type": "text", + "content": "W.J. Chang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": "and employs the constant velocity model to predict future trajectories of non-ego vehicles, generating multiple trajectory candidates and selecting the one least likely to result in a collision; 2) a hybrid planner BITS [35]; 3) PDM-Closed [5], the winner of the 2023 nuPlan Planning Challenge; 4) a deterministic a learning-based Behavior Cloning (BC) planner, which utilizes a ResNet Encoder, followed by an MLP decoder to generate future trajectories; and 5) an Intelligent Driver Model (IDM) planner [30]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 204, + 482, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 204, + 482, + 302 + ], + "spans": [ + { + "bbox": [ + 130, + 204, + 482, + 302 + ], + "type": "text", + "content": "Diffusion Model. We follow the architecture described in [39]. We represent the context using an agent-centric map and past trajectories on a rasterized map. The traffic scene is encoded using a ResNet structure [12], while the input trajectory is processed through a series of 1D temporal convolution blocks in an UNet-like architecture, as detailed in [18]. The model uses " + }, + { + "bbox": [ + 130, + 204, + 482, + 302 + ], + "type": "inline_equation", + "content": "K = 100" + }, + { + "bbox": [ + 130, + 204, + 482, + 302 + ], + "type": "text", + "content": " diffusion steps. During the inference phase, we generate a sample of potential future trajectories for each reactive agent in a given scene. From these, we select the trajectory that yields the lowest guidance cost. This process is referred to as filtering." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 304, + 482, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 482, + 340 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 482, + 340 + ], + "type": "text", + "content": "Closed-loop Simulation. The simulation framework for our experiments is built upon an open-source traffic behavior framework [35]. Within this framework, both the planner and reactive agents update their plans at a frequency of " + }, + { + "bbox": [ + 130, + 304, + 482, + 340 + ], + "type": "inline_equation", + "content": "2\\mathrm{Hz}" + }, + { + "bbox": [ + 130, + 304, + 482, + 340 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 377, + 257, + 389 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 377, + 257, + 389 + ], + "spans": [ + { + "bbox": [ + 131, + 377, + 257, + 389 + ], + "type": "text", + "content": "6.3 Evaluation Metrics" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 418, + 482, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 482, + 527 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 482, + 527 + ], + "type": "text", + "content": "Our goal is to validate that the proposed method can generate safety-critical scenarios that are both realistic and controllable. For realism assessment, in accordance with [39], we compare statistical data between simulated trajectories and actual ground trajectories. This involves calculating the Wasserstein distance between their driving profiles' normalized histograms, focusing on the mean of mean values for three properties: longitudinal acceleration, latitudinal acceleration, and jerk. To evaluate controllability, we measure metrics related to parameters we can control, specifically relative speed and time-to-collision cost between the ego and adversarial agents." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 129, + 529, + 482, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 529, + 482, + 602 + ], + "spans": [ + { + "bbox": [ + 129, + 529, + 482, + 602 + ], + "type": "text", + "content": "Our method aims to evaluate the extent of control over adversarial behaviors within a single scenario. To do this, we focus on measuring collision diversity—a metric that quantifies the range of differences in collision angles, relative speeds, and collision points. By calculating the variance of these parameters within the same scenario, we can determine how diverse and controllable the adversarial behaviors are, ensuring the scenario's realism and controllability." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 129, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 129, + 605, + 482, + 666 + ], + "type": "text", + "content": "For a more nuanced understanding of realism, we analyze the collision rate - the average fractions of agents colliding, and the offroad rate - the percentage of reactive agents going off-road, both of which are considered failure rates [35]. All metrics are averaged across scenarios. For detailed definitions and metrics, see supplementary Sec. C." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 135, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 135, + 90, + 447, + 102 + ], + "type": "text", + "content": "Safety-Critical Closed-Loop Traffic Simulation with Controllable Adversaries" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 114, + 479, + 193 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 193 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 193 + ], + "type": "table", + "html": "
DatasetMethodCollision (%) ↑Other Offroad (%)↓Adv Offroad (%) ↓Collision Rel Speed (m/s) ↓Realism ↓Time (s) ↓
nuScenesSAFE-SIM43.21.811.4-0.120.38104.5 ± 17.7
STRIVE36.42.211.45.520.85427.2 ± 169.8
DiffScene18.211.49.016.40.52105.4 ± 22.5
nuPlanSAFE-SIM809.411.76.750.27173.4 ± 73.3
DiffScene56.714.05.0-2.810.42176.7 ± 77.5
", + "image_path": "47049cf1ead19db50f52b7880b8f43142b587f30616d283eb525aa4cc68440f4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 132, + 258, + 479, + 326 + ], + "blocks": [ + { + "bbox": [ + 130, + 194, + 480, + 236 + ], + "lines": [ + { + "bbox": [ + 130, + 194, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 480, + 236 + ], + "type": "text", + "content": "Table 2: Safety-critical Traffic Simulation. We compare our approach against STRIVE [26] and DiffScene [34] for safety-critical traffic simulation with a rule-based planner. SAFE-SIM outperforms STRIVE on all metrics and demonstrates higher collision rates and better realism than DiffScene." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 258, + 479, + 326 + ], + "lines": [ + { + "bbox": [ + 132, + 258, + 479, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 258, + 479, + 326 + ], + "type": "table", + "html": "
PlannerEgo-Adv Coll (%) ↑Ego-Other Coll (%) ↑Adv Offroad (%) ↓Coll Speed (m/s) ↓Ego Accel (m/s2) ↓Realism ↓
BC38.837.39.02.470.540.79
IDM49.358.23.0-0.400.940.78
Lane-Graph34.337.31.52.681.620.57
BITS16.419.46.03.070.950.79
PDM-Closed26.950.71.52.731.300.86
", + "image_path": "2aef0898f1ee520fdf324a459b88a9fc1d86bba8ac8906aa70d4651fed2a8578.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 326, + 481, + 359 + ], + "lines": [ + { + "bbox": [ + 130, + 326, + 481, + 359 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 481, + 359 + ], + "type": "text", + "content": "Table 3: Safety-Critical Simulation for Different Planners. SAFE-SIM can generate diverse safety-critical scenarios tailored to different planners, including rule-based, learning-based, and hybrid planners." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 392, + 474, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 474, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 474, + 415 + ], + "type": "text", + "content": "6.4 Evaluation of Safety-Critical Traffic Simulation with Baseline Methods" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 423, + 482, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 482, + 495 + ], + "type": "text", + "content": "We compared our method with STRIVE [26] and DiffScene [34], utilizing the Lane-graph-based planner from STRIVE. The evaluation focused on collision rates between the ego and the adversarial agent (\"Collision\"), adversarial agent off-road rates (\"Adv Offroad\"), other agents' off-road rates (\"Other Offroad\"), collision relative speed between the ego and adversary (\"Collision Rel Speed\"), realism of all non-ego agents (\"Realism\"), and simulation time." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 496, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 496, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 496, + 482, + 593 + ], + "type": "text", + "content": "The results, presented in Tab. 2, demonstrate that our method excels in all metrics compared to STRIVE, especially in collisions and realism. Compared to DiffScene, SAFE-SIM also exhibits a higher collision rate with better realism. Note that we trained models on nuScenes and tested them in nuPlan without additional fine-tuning. As illustrated in Figure 4, the qualitative examples from the NuScenes dataset demonstrate how our framework can challenge the rule-based planner with various driving situations. See supplementary Sec. A for more qualitative examples." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 609, + 413, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 609, + 413, + 621 + ], + "spans": [ + { + "bbox": [ + 130, + 609, + 413, + 621 + ], + "type": "text", + "content": "6.5 Safety-Critical Simulation with Different Planners" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "In Tab. 3, we demonstrate our framework's ability to generate collisions across various planner types: rule-based, learning-based, and hybrid planners. Notably, the IDM planner exhibits the highest Ego-Adv collision rate. This heightened rate" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "type": "text", + "content": "W.J. Chang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 156, + 114, + 458, + 180 + ], + "blocks": [ + { + "bbox": [ + 156, + 114, + 458, + 180 + ], + "lines": [ + { + "bbox": [ + 156, + 114, + 458, + 180 + ], + "spans": [ + { + "bbox": [ + 156, + 114, + 458, + 180 + ], + "type": "table", + "html": "
TTC Cost WeightTTC CostColl Speed (m/s)Coll Angle (deg)Coll Rate (%)↑Realism ↓
0.00.182.45-7.4348.20.76
1.00.212.300.4353.60.79
2.00.263.78-17.060.70.81
", + "image_path": "e371ab30bd681bf879386a53fcb7090e725083a95280146240add8ead5968911.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 132, + 245, + 479, + 327 + ], + "blocks": [ + { + "bbox": [ + 130, + 181, + 482, + 224 + ], + "lines": [ + { + "bbox": [ + 130, + 181, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 181, + 482, + 224 + ], + "type": "text", + "content": "Table 4: Controlling Time to Collision (TTC). The table shows the impact of different TTC Cost weights on collision scenarios. Increasing the TTC Cost weight results in an increase in collision rate, suggesting a heightened challenge for the ego vehicle in avoiding collisions." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 245, + 479, + 327 + ], + "lines": [ + { + "bbox": [ + 132, + 245, + 479, + 327 + ], + "spans": [ + { + "bbox": [ + 132, + 245, + 479, + 327 + ], + "type": "table", + "html": "
Jadv Jreg Partial DiffusionCollision MetricsDiversity
Collision (%) ↑Adv Offroad (%) ↓Realism ↓Collision Angle Var (rad) ↑Collision Rel Speed Var (m/s) ↑Collision Point Var (m) ↑
✓✓×23.913.80.582.222.991.62
✓✓✓29.014.60.573.101.965.44
✓××53.523.10.583.344.812.47
", + "image_path": "a6d09b2643548dfbabb89b5b14ff76544552052fd1e4bf072202d06738ee21f8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 327, + 481, + 361 + ], + "lines": [ + { + "bbox": [ + 130, + 327, + 481, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 481, + 361 + ], + "type": "text", + "content": "Table 5: Ablation Study on Controllability. This study examines each component of our proposed method. Partial diffusion significantly increases the variance of the collision point, resulting in a greater diversity of collision scenarios." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 392, + 481, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 481, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 481, + 429 + ], + "type": "text", + "content": "can be attributed to the IDM's focus on vehicles near the same lane, potentially overlooking other vehicles in the scene. Consequently, in scenarios with the IDM planner, our framework can induce collisions at relatively lower speeds." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 445, + 369, + 458 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 445, + 369, + 458 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 369, + 458 + ], + "type": "text", + "content": "6.6 Evaluation: Controlling Safety-Criticality" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 463, + 481, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 463, + 481, + 487 + ], + "spans": [ + { + "bbox": [ + 130, + 463, + 481, + 487 + ], + "type": "text", + "content": "A key feature of SAFE-SIM is its ability to generate controllable adversarial behaviors, offering variations not possible with previous methods." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 487, + 482, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 482, + 606 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 482, + 606 + ], + "type": "text", + "content": "Controlling Time-to-Collision. We control the orientation and relative speed together using the time-to-collision (TTC) cost, as described in Sec. 5. We manipulate the scenario's safety-criticality by adjusting the relative weight of the TTC cost. To assess the impact of these adjustments, we measure the average TTC cost shortly before a collision occurs (0.5 seconds). Our observations, detailed in Tab. 4, show that increasing the TTC weight raises the TTC cost. Notably, while the relative collision speed remains fairly consistent, the collision angle shifts, indicating a greater difficulty in avoiding ego-adversary collisions. Additionally, our method can control other aspects, such as relative speed, as detailed in supplementary Sec. F.1." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 623, + 331, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 623, + 331, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 623, + 331, + 635 + ], + "type": "text", + "content": "6.7 Ablation Study on Controllability" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "We performed an ablation study to assess the influence of different guidance strategies in diffusion models on the quality of simulations. This study, detailed" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 135, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 135, + 90, + 447, + 102 + ], + "type": "text", + "content": "Safety-Critical Closed-Loop Traffic Simulation with Controllable Adversaries" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": "in Table 5, aimed to quantify collision diversity. We compare against our baseline approach, consisting of regularized and adversarial objectives " + }, + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "inline_equation", + "content": "(J_{reg} + J_{adv})" + }, + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": ", by predetermining the selection of adversarial agents and conducting experiments with multiple seeds (three) in our framework to evaluate collision diversity. In partial diffusion, we manipulated the trajectory proposal selection mechanism across centerlines with normal offsets of -2.0, 0.0, and 2.0." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 482, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 482, + 247 + ], + "type": "text", + "content": "Effectiveness of Partial Diffusion. Table 5 reveals that partial diffusion significantly enhances both the collision point and angle diversity in comparison to the baseline approach. This underscores the capability of partial diffusion to generate a wider array of collision scenarios through various trajectory proposals, illustrating its potential to explore diverse collision dynamics effectively." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 248, + 482, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 482, + 319 + ], + "type": "text", + "content": "Impact of the Regularization Term (" + }, + { + "bbox": [ + 130, + 248, + 482, + 319 + ], + "type": "inline_equation", + "content": "J_{reg}" + }, + { + "bbox": [ + 130, + 248, + 482, + 319 + ], + "type": "text", + "content": "). Incorporating " + }, + { + "bbox": [ + 130, + 248, + 482, + 319 + ], + "type": "inline_equation", + "content": "J_{reg}" + }, + { + "bbox": [ + 130, + 248, + 482, + 319 + ], + "type": "text", + "content": " results in a notable decrease in the adversarial-collision rate, highlighting the importance of the regularization term in enhancing simulation realism. Without " + }, + { + "bbox": [ + 130, + 248, + 482, + 319 + ], + "type": "inline_equation", + "content": "J_{reg}" + }, + { + "bbox": [ + 130, + 248, + 482, + 319 + ], + "type": "text", + "content": ", the collision rate between the adversarial agent and ego vehicle increases, but the adversarial vehicle goes offroad more often, leading to scenarios that differ significantly from realistic behavior." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 335, + 307, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 335, + 307, + 346 + ], + "spans": [ + { + "bbox": [ + 131, + 335, + 307, + 346 + ], + "type": "text", + "content": "6.8 Limitation and Failure Cases" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 352, + 482, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 352, + 482, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 352, + 482, + 422 + ], + "type": "text", + "content": "We identified areas for improvement in SAFE-SIM (see supp. Fig. A4). In certain cases, the adversarial agent unrealistically collides with non-adversarial agents before reaching the ego agent. Additionally, some scenarios result in collisions where the ego planner is not at fault. While understanding how the ego planner can avoid such cases is important, creating more scenarios where the ego is at fault would be beneficial." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 439, + 220, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 439, + 220, + 452 + ], + "spans": [ + { + "bbox": [ + 131, + 439, + 220, + 452 + ], + "type": "text", + "content": "7 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 462, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 605 + ], + "type": "text", + "content": "In this work, we present a closed-loop simulation framework utilizing guided diffusion models for creating safety-critical scenarios to assess autonomous vehicle (AV) algorithms. Our research is in line with the goals of SO-TIF, focusing on how autonomous vehicles respond to dynamic scenarios like aggressive driving, underscoring our dedication to safety across diverse and unforeseeable conditions. Our framework introduces innovative guidance objectives tailored for controllable, stable, long-term safety-critical simulations. A key aspect of our method lies in its ability to vary the types of adversarial behavior within collision scenarios. By integrating adversarial objectives and partial diffusion, we enable fine-grained control over adversarial actions. This versatility enables our framework to produce a broader range of realistic and manageable scenarios, setting a new standard in adversarial scenario generation beyond the limitations of existing approaches." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "type": "text", + "content": "Future directions for our research include: 1) exploring the application of our framework in closed-loop policy training, and 2) developing automated methods for adjusting controllable parameters. These methods aim to facilitate the generation of diverse, long-tail scenarios. We believe this framework holds significant promise for enhancing real-world AV safety." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "type": "text", + "content": "W.J. Chang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "content": "This work was part of W.J. Chang's summer internship at NEC Labs America, and he is also supported by the National Science Foundation Graduate Research Fellowship Program under Grant No. DGE 2146752. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation. The authors would like to thank Chih-Ling Chang for her insightful suggestions and assistance with figures and presentations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 243, + 197, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 243, + 197, + 255 + ], + "spans": [ + { + "bbox": [ + 133, + 243, + 197, + 255 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 269, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 138, + 269, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 269, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 138, + 269, + 480, + 300 + ], + "type": "text", + "content": "1. Abeysirigoonawardena, Y., Shkurti, F., Dudek, G.: Generating adversarial driving scenarios in high-fidelity simulators. In: 2019 International Conference on Robotics and Automation (ICRA). pp. 8271-8277. IEEE (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 303, + 480, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 303, + 480, + 344 + ], + "spans": [ + { + "bbox": [ + 138, + 303, + 480, + 344 + ], + "type": "text", + "content": "2. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11621-11631 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 346, + 480, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 346, + 480, + 378 + ], + "spans": [ + { + "bbox": [ + 138, + 346, + 480, + 378 + ], + "type": "text", + "content": "3. Caesar, H., Kabzan, J., Tan, K.S., Fong, W.K., Wolff, E., Lang, A., Fletcher, L., Beijbom, O., Omari, S.: nuplan: A closed-loop ml-based planning benchmark for autonomous vehicles. arXiv preprint arXiv:2106.11810 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 380, + 480, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 380, + 480, + 411 + ], + "spans": [ + { + "bbox": [ + 138, + 380, + 480, + 411 + ], + "type": "text", + "content": "4. Cao, Y., Xiao, C., Anandkumar, A., Xu, D., Pavone, M.: Advdo: Realistic adversarial attacks for trajectory prediction. In: European Conference on Computer Vision. pp 36-52. Springer (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 413, + 480, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 413, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 138, + 413, + 480, + 434 + ], + "type": "text", + "content": "5. Dauner, D., Hallgarten, M., Geiger, A., Chitta, K.: Parting with misconceptions about learning-based vehicle motion planning. In: CoRL (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 435, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 435, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 138, + 435, + 480, + 456 + ], + "type": "text", + "content": "6. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in neural information processing systems 34, 8780-8794 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 457, + 481, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 457, + 481, + 498 + ], + "spans": [ + { + "bbox": [ + 138, + 457, + 481, + 498 + ], + "type": "text", + "content": "7. Ding, W., Chen, B., Li, B., Eun, K.J., Zhao, D.: Multimodal safety-critical scenarios generation for decision-making algorithms evaluation. IEEE Robotics and Automation Letters 6(2), 1551-1558 (2021). https://doi.org/10.1109/LRA.2021.3058873" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 500, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 500, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 500, + 480, + 533 + ], + "type": "text", + "content": "8. Ding, W., Chen, B., Xu, M., Zhao, D.: Learning to collide: An adaptive safety-critical scenarios generating method. In: 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 2243-2250. IEEE (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 534, + 480, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 534, + 480, + 565 + ], + "spans": [ + { + "bbox": [ + 138, + 534, + 480, + 565 + ], + "type": "text", + "content": "9. Ding, W., Xu, C., Arief, M., Lin, H., Li, B., Zhao, D.: A survey on safety-critical driving scenario generation—a methodological perspective. IEEE Transactions on Intelligent Transportation Systems (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 567, + 480, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 567, + 480, + 598 + ], + "spans": [ + { + "bbox": [ + 134, + 567, + 480, + 598 + ], + "type": "text", + "content": "10. Gu, T., Chen, G., Li, J., Lin, C., Rao, Y., Zhou, J., Lu, J.: Stochastic trajectory prediction via motion indeterminacy diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17113-17122 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 600, + 480, + 632 + ], + "type": "text", + "content": "11. Hanselmann, N., Renz, K., Chitta, K., Bhattacharyya, A., Geiger, A.: King: Generating safety-critical driving scenarios for robust imitation via kinematics gradients In: European Conference on Computer Vision. pp. 335-352. Springer (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "text", + "content": "12. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition In: Proceedings of the IEEE conference on computer vision and pattern recognition pp. 770-778 (2016)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 135, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 135, + 90, + 447, + 102 + ], + "type": "text", + "content": "Safety-Critical Closed-Loop Traffic Simulation with Controllable Adversaries" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "13. Ho, J., Chan, W., Sahara, C., Whang, J., Gao, R., Gritsenko, A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., et al.: Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 171 + ], + "type": "text", + "content": "14. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 481, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 481, + 193 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 481, + 193 + ], + "type": "text", + "content": "15. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 215 + ], + "type": "text", + "content": "16. Ho, J., Salimans, T., Gritsenko, A., Chan, W., Norouzi, M., Fleet, D.J.: Video diffusion models. arXiv:2204.03458 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "type": "text", + "content": "17. Ivanovic, B., Song, G., Gilitschenski, I., Pavone, M.: trajdata: A unified interface to multiple human trajectory datasets. arXiv preprint arXiv:2307.13924 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 237, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 481, + 259 + ], + "type": "text", + "content": "18. Janner, M., Du, Y., Tenenbaum, J.B., Levine, S.: Planning with diffusion for flexible behavior synthesis. arXiv preprint arXiv:2205.09991 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "text", + "content": "19. Lu, J., Wong, K., Zhang, C., Suo, S., Urtasun, R.: Scenecontrol: Diffusion for controllable traffic scene generation. In: IEEE International Conference on Robotics and Automation (ICRA) (2024)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 293, + 481, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 326 + ], + "type": "text", + "content": "20. Montali, N., Lambert, J., Mougin, P., Kuefler, A., Rhinehart, N., Li, M., Gulino, C., Emrich, T., Yang, Z., Whiteson, S., et al.: The waymo open sim agents challenge. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 326, + 481, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 347 + ], + "type": "text", + "content": "21. Nichol, A.Q., Dhariwal, P.: Improved denoising diffusion probabilistic models. In: International Conference on Machine Learning. pp. 8162-8171. PMLR (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "text", + "content": "22. Nishimura, H., Mercat, J., Wulfe, B., McAllister, R.T., Gaidon, A.: Rap: Risk-aware prediction for robust planning. In: Conference on Robot Learning. pp. 381-392. PMLR (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "text", + "content": "23. Philion, J., Peng, X.B., Fidler, S.: Trajeglish: Traffic modeling as next-token prediction. In: The Twelfth International Conference on Learning Representations (2024)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 413, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 481, + 445 + ], + "type": "text", + "content": "24. Pronovost, E., Ganesina, M.R., Hendy, N., Wang, Z., Morales, A., Wang, K., Roy, N.: Scenario diffusion: Controllable driving scenario generation with diffusion. Advances in Neural Information Processing Systems 36, 68873-68894 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 446, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 481, + 490 + ], + "type": "text", + "content": "25. Rempe, D., Luo, Z., Bin Peng, X., Yuan, Y., Kitani, K., Kreis, K., Fidler, S., Litany, O.: Trace and pace: Controllable pedestrian animation via guided trajectory diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13756-13766 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 491, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 481, + 533 + ], + "type": "text", + "content": "26. Rempe, D., Philion, J., Guibas, L.J., Fidler, S., Litany, O.: Generating useful accident-prone driving scenarios via a learned traffic prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17305-17315 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 534, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 567 + ], + "type": "text", + "content": "27. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 567, + 481, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 600 + ], + "type": "text", + "content": "28. Suo, S., Regalado, S., Casas, S., Urtasun, R.: Trafficsim: Learning to simulate realistic multi-agent behaviors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10400-10409 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 600, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 481, + 643 + ], + "type": "text", + "content": "29. Suo, S., Wong, K., Xu, J., Tu, J., Cui, A., Casas, S., Urtasun, R.: Mixsim: A hierarchical framework for mixed reality traffic simulation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9622-9631 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "text", + "content": "30. Treiber, M., Hennecke, A., Helbing, D.: Congested traffic states in empirical observations and microscopic simulations. Physical review E 62(2), 1805 (2000)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "type": "text", + "content": "W.J. Chang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 402 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 132, + 117, + 457, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 117, + 457, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 457, + 127 + ], + "type": "text", + "content": "31. Unknown: Midjourney. https://www.midjourney.com, accessed: 2023-11-16" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 128, + 481, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 128, + 481, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 128, + 481, + 148 + ], + "type": "text", + "content": "32. Unknown: Openai dall-e-2. https://openai.com/product/dall-e-2, accessed: 2023-11-16" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 149, + 481, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 481, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 481, + 193 + ], + "type": "text", + "content": "33. Wang, J., Pun, A., Tu, J., Manivasagam, S., Sadat, A., Casas, S., Ren, M., Urtasun, R.: Advsim: Generating safety-critical scenarios for self-driving vehicles. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9909-9918 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "text", + "content": "34. Xu, C., Zhao, D., Sangiovanni-Vincentelli, A., Li, B.: Diffscene: Diffusion-based safety-critical scenario generation for autonomous vehicles. In: The Second Workshop on New Frontiers in Adversarial Machine Learning (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "type": "text", + "content": "35. Xu, D., Chen, Y., Ivanovic, B., Pavone, M.: Bits: Bi-level imitation for traffic simulation. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 2929-2936. IEEE (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "text", + "content": "36. Yin, Z.H., Sun, L., Sun, L., Tomizuka, M., Zhan, W.: Diverse critical interaction generation for planning and planner evaluation. In: 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 7036-7043. IEEE (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 325 + ], + "type": "text", + "content": "37. Zhang, Q., Hu, S., Sun, J., Chen, Q.A., Mao, Z.M.: On adversarial robustness of trajectory prediction for autonomous vehicles. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15159-15168 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 325, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 481, + 357 + ], + "type": "text", + "content": "38. Zhong, Z., Rempe, D., Chen, Y., Ivanovic, B., Cao, Y., Xu, D., Pavone, M., Ray, B.: Language-guided traffic simulation via scene-level diffusion. arXiv preprint arXiv:2306.06344 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 358, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 402 + ], + "type": "text", + "content": "39. Zhong, Z., Rempe, D., Xu, D., Chen, Y., Veer, S., Che, T., Ray, B., Pavone, M.: Guided conditional diffusion for controllable traffic simulation. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 3560-3566. IEEE (2023)" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 135, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 135, + 91, + 447, + 102 + ], + "type": "text", + "content": "Safety-Critical Closed-Loop Traffic Simulation with Controllable Adversaries" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Safeguard Text-to-Image Diffusion Models with Human Feedback Inversion/603d7974-42c1-4361-a264-c2589e2c9681_content_list.json b/2024/Safeguard Text-to-Image Diffusion Models with Human Feedback Inversion/603d7974-42c1-4361-a264-c2589e2c9681_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..81fd0cdd8893059fec147f61e5524363fc9338a5 --- /dev/null +++ b/2024/Safeguard Text-to-Image Diffusion Models with Human Feedback Inversion/603d7974-42c1-4361-a264-c2589e2c9681_content_list.json @@ -0,0 +1,1854 @@ +[ + { + "type": "text", + "text": "Safeguard Text-to-Image Diffusion Models with Human Feedback Inversion", + "text_level": 1, + "bbox": [ + 230, + 140, + 774, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sanghyun Kim $^{\\text{D}}$ , Seohyeon Jung $^{\\text{D}}$ , Balhae Kim $^{\\text{D}}$ , Moonseok Choi $^{\\text{D}}$ , Jinwoo Shin $^{\\text{D}}$ , and Juho Lee $^{\\text{D}}$", + "bbox": [ + 254, + 212, + 746, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Korea Advanced Institute of Science and Technology (KAIST) {nannullna,heon2203,balhaekim,ms.choi,jinwoos,juholee}@kaist.ac.kr", + "bbox": [ + 243, + 255, + 756, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. This paper addresses the societal concerns arising from large-scale text-to-image diffusion models for generating potentially harmful or copyrighted content. Existing models rely heavily on internet-crawled data, wherein problematic concepts persist due to incomplete filtration processes. While previous approaches somewhat alleviate the issue, they often rely on text-specified concepts, introducing challenges in accurately capturing nuanced concepts and aligning model knowledge with human understandings. In response, we propose a framework named Human Feedback Inversion (HFI), where human feedback on model-generated images is condensed into textual tokens guiding the mitigation or removal of problematic images. The proposed framework can be built upon existing techniques for the same purpose, enhancing their alignment with human judgment. By doing so, we simplify the training objective with a self-distillation-based technique, providing a strong baseline for concept removal. Our experimental results demonstrate our framework significantly reduces objectionable content generation while preserving image quality, contributing to the ethical deployment of AI in the public sphere. Code is available at https://github.com/nannullna/safeguard-hfi.", + "bbox": [ + 261, + 316, + 743, + 566 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Caution: This paper contains discussions and examples related to harmful content, including text and images. Reader discretion is advised.", + "bbox": [ + 261, + 579, + 740, + 608 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Fairness, accountability, and transparency $\\cdot$ Diffusion models $\\cdot$ Image generation", + "bbox": [ + 261, + 619, + 740, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 672, + 375, + 690 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large-scale text-to-image generation models have demonstrated remarkable success in producing high-quality images encompassing a wide array of concepts [4, 31,32,36,39]. These models possess the ability to combine disparate ideas in novel and unprecedented manners, making them invaluable tools for creative content creation. However, they also give rise to substantial concerns, including the generation of harmful or copyrighted content, even without explicit directives [40]. A critical issue lies in the heavy reliance of the models on internet-crawled datasets (e.g., LAION-5B [42], currently not available [50]), which, despite attempts at aggressive filtration [48], often lead to problematic content generation [33,40].", + "bbox": [ + 212, + 704, + 787, + 843 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/393aa38b7d0a6a268dc4680498706399ed7c20875575bd8391fdd410d7d9a65f.jpg", + "image_caption": [ + "Fig. 1: Comparative analysis of NSFW content removal techniques, including our proposed fine-tuning method (SDD), both with and without HFI framework. The results clearly show that incorporating HFI significantly reduces the amplification of provocative body parts and ensures the generation of clothed representations." + ], + "image_footnote": [], + "bbox": [ + 215, + 142, + 787, + 388 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Previous inference-time [3,40] and fine-tuning methods [13,14,52] have made strides in mitigating objectionable imagery. Here, the concept to be removed, or the target concept, is specified by text. However, they insufficiently address the nuanced nature of concepts, leading to a gap between the model-generated content and human expectations. Examples include determining whether exposed body parts in a masterpiece are considered safe or if an illustration overtly portraying such body should be similarly categorized. Moreover, the ethical judgment criteria of the model generally fall short of human standards [33, 40], emphasizing the crucial need for collective intelligence intervention at this juncture.", + "bbox": [ + 212, + 488, + 787, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To handle such ambiguity and bridge this gap, we present Human Feedback Inversion (HFI), a novel and effective framework to mitigate the ethical concerns surrounding text-to-image diffusion models. HFI incorporates human feedback on model-generated images with problematic concepts directly into soft tokens using the textual inversion (TI) [12]. By directly incorporating human perspectives, HFI addresses the limitations of the previous text-based approaches, offering a more nuanced and accurate means of discerning harmful concepts. This condensed information serves as a guiding force against the target concept during inference time or to create training images for fine-tuning. Through this process, tokens acquired via HFI become specialized for their respective concepts, aiming to enhance both effectiveness (i.e., removal performance of the target concept) and utility aspects (i.e., image quality for remaining concepts).", + "bbox": [ + 212, + 626, + 787, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To augment the efficacy of HFI, we introduce Safe self-Distillation Diffusion (SDD), a self-distillation-based fine-tuning method. This method simplifies", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "S. Kim et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1c9d5d2ac0f1d53dfcdc17500255f1c9852852b6f1ab57b8328a00e9a6a69af9.jpg", + "image_caption": [ + "(a) Flowers" + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 401, + 232 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/fa677bffb3fa5a15e9f76ca31deb80f20f73e0921dd458363ec24f2d3d4ede3f.jpg", + "image_caption": [ + "(b) Fish" + ], + "image_footnote": [], + "bbox": [ + 408, + 143, + 593, + 232 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ddb9a2e9ebd33459451c72a6b9fa1346878e12e3fc46c4cafb9c5b0370596496.jpg", + "image_caption": [ + "(c) Boat" + ], + "image_footnote": [], + "bbox": [ + 602, + 143, + 784, + 232 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/de5cb46be7044453dc2df6fb464e5bb8353ad3f50aff65c4b03047dcd291b2eb.jpg", + "image_caption": [ + "(d) Hat" + ], + "image_footnote": [], + "bbox": [ + 217, + 251, + 401, + 339 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ac85089f823d0aa0bcf217e8c4c926ccc5d8c32f690bf8ef9977f9a41feb8ffc.jpg", + "image_caption": [ + "(e) Tree" + ], + "image_footnote": [], + "bbox": [ + 410, + 251, + 593, + 339 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/799d3402bde8d72b57bcd2f8249b3d5108d3a121f7c83cac23c0d23eec520805.jpg", + "image_caption": [ + "(f) Streetlamp" + ], + "image_footnote": [], + "bbox": [ + 602, + 251, + 784, + 339 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7b31fc6842a2133a98e6599df08078f8f2cd80efef50af10a437e4192073482b.jpg", + "image_caption": [ + "Fig. 2: Artist style removal results on Vincent van Gogh. Images are generated with prompts such as \"a painting of flowers in the style of Vincent van Gogh\". HFI+SDD effectively eliminates his distinct artistic features while preserving the essence of the original subject matter, subject integrity, and visual quality." + ], + "image_footnote": [], + "bbox": [ + 230, + 438, + 411, + 539 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8931b5a4b14f87c0a8c4ad3023f774825f7c49c433abe5886fa9907473dd7bac.jpg", + "image_caption": [ + "Fig. 3: Comparison of images generated with famous artwork titles of the three artists. HFI+SDD, leveraging human feedback, captures and removes distinctive artistic styles more effectively while preserving the integrity and essence of the original paintings." + ], + "image_footnote": [], + "bbox": [ + 421, + 438, + 591, + 539 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b753e0336562fdaac3f10506b303f6f14b59ea30d5b92bab93e42447be25e1db.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 602, + 438, + 771, + 539 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the training objective, eliminating the need for manually designing the target or specifying which concept to retain. With this improved design, SDD outperforms existing baselines, even when based on text-specified concepts, with improved training stability. The integration of HFI and SDD establishes a synergistic alliance, utilizing the HFI token as a removal target. This soft token plays a crucial role in generating latents, serving as a vital component for self-supervision and further contributing to the model's alignment with human understanding.", + "bbox": [ + 212, + 626, + 782, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Figs. 1 to 3 showcase the effectiveness of HFI in concept removal, surpassing previous methods and aligning well with human preferences. Specifically, when tasked with removing Van Gogh's artistic style, our pipeline, guided by human feedback, successfully retained the essential qualities of a painting-like style while eliminating his characteristic bold brushstrokes. In contrast, methods without human feedback transformed images into realistic photographs to aggressively remove artist's styles as a side effect. Additionally, when instructed to remove", + "bbox": [ + 212, + 734, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Safeguard Diffusion Models with Human Feedback Inversion", + "bbox": [ + 328, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "nudity with human feedback, our pipeline successfully eliminated pornographic content, in contrast to previous methods that still produced provocative images. The visual evidence highlights the superior effectiveness of HFI with SDD in consistently delivering safe and modest image outputs across different prompts.", + "bbox": [ + 212, + 146, + 782, + 207 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In summary, our framework not only markedly diminishes the generation of undesirable content perceived by humans but also maintains image quality. This offers a pragmatic solution for improving safety and ethical considerations in the widespread deployment of text-to-image diffusion models.", + "bbox": [ + 212, + 207, + 784, + 267 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Background on Diffusion Models", + "text_level": 1, + "bbox": [ + 215, + 292, + 568, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Diffusion models [17, 44-47] are a class of generative models that learn the data distribution by building two Markov chains of latent variables. Given a sample $\\mathbf{x}_0 \\sim p_{\\mathrm{data}}(\\mathbf{x}) \\coloneqq q(\\mathbf{x})$ and a noise schedule $\\{\\beta_t\\}_{t=1}^T$ , the forward process gradually injects a series of Gaussian noises to the sample until it nearly follows the standard Gaussian $q(\\mathbf{x}_t | \\mathbf{x}_{t-1}) \\coloneqq \\mathcal{N}(\\mathbf{x}_t; \\sqrt{1 - \\beta_t} \\mathbf{x}_{t-1}, \\beta_t \\mathbf{I})$ , $q(\\mathbf{x}_T | \\mathbf{x}_0) \\approx \\mathcal{N}(\\mathbf{x}_T; \\mathbf{0}, \\mathbf{I})$ . Such process is then followed by the reverse process parameterized by $\\theta$ , where the model learns to reconstruct the original image by iteratively denoising from a pure Gaussian noise $p(\\mathbf{x}_T) = \\mathcal{N}(\\mathbf{0}, \\mathbf{I})$ as $p_\\theta(\\mathbf{x}_{0:T}) = p(\\mathbf{x}_T) \\prod_{t=1}^T p_\\theta(\\mathbf{x}_{t-1} | \\mathbf{x}_t)$ . Ho et al. [17] simplifies the objective to learn a noise estimator $\\epsilon_\\theta$ :", + "bbox": [ + 212, + 327, + 782, + 463 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {D M} (\\theta) = \\mathbb {E} _ {\\mathbf {x} _ {0}, \\epsilon , t} \\left[ \\| \\boldsymbol {\\epsilon} - \\boldsymbol {\\epsilon} _ {\\theta} \\left(\\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {x} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\boldsymbol {\\epsilon}; t\\right) \\| _ {2} ^ {2} \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 316, + 477, + 784, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ , $\\bar{\\alpha}_t = \\prod_{s=1}^t (1 - \\beta_s)$ , $t \\sim \\mathcal{U}([1,T])$ , and $T$ is the total number of steps. For simple notation, we omit the timestep $t$ as $\\epsilon(\\mathbf{x}_t; t) = \\epsilon(\\mathbf{x}_t)$ .", + "bbox": [ + 212, + 506, + 782, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Latent Diffusion Models (LDMs) [36] leverage the diffusion process within the latent space rather than in the pixel space, utilizing a pre-trained autoencoder. By mapping the input $\\mathbf{x}$ into a latent space with the encoder $\\mathcal{E}$ , $\\mathbf{z} = \\mathcal{E}(\\mathbf{x})$ , an LDM is trained to predict the added noise in the latent space. As this paper mainly discusses LDMs, we use $\\mathbf{x}_t$ instead of $\\mathbf{z}_t$ . Text-to-image models additionally take inputs of the text embedding, $\\mathbf{c}_p = \\mathcal{E}_{\\mathrm{txt}}(c_p)$ , paired with an image $\\mathbf{x}$ , where $\\mathcal{E}_{\\mathrm{txt}}$ is the text encoder and $c_{p}$ is a text prompt. To enhance the quality of text conditioning, the classifier-free guidance (CFG) models both conditional and unconditional noise estimates by randomly replacing $\\mathbf{c}_p$ with the embedding of an empty string $\\mathbf{c}_{\\emptyset} = \\mathcal{E}_{\\mathrm{txt}}(\"\")$ and modifies the noise estimate during sampling as $\\tilde{\\epsilon}_{\\mathrm{CFG}} = \\epsilon_{\\theta}(\\mathbf{x}_t) + (w + 1)[\\epsilon_{\\theta}(\\mathbf{x}_t;\\mathbf{c}_p) - \\epsilon_{\\theta}(\\mathbf{x}_t;\\mathbf{c}_{\\emptyset})]$ , where $w + 1$ is the CFG scale. We simply denote $\\epsilon_{\\theta}(\\mathbf{x}_t;\\mathbf{c}_{\\emptyset})$ as $\\epsilon_{\\theta}(\\mathbf{x}_t)$ in the subsequent sections.", + "bbox": [ + 212, + 539, + 784, + 720 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 744, + 330, + 760 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We present Human Feedback Inversion (HFI), an innovative framework for concept removal guided by human feedback. Unlike previous methods that rely on predefined text-based concepts, i.e., a specified target word like \"nudity\", real-world inappropriate concepts are often nuanced and challenging to express in", + "bbox": [ + 212, + 779, + 782, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "S. Kim et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b61c306d2a01ff31ee4f2d0086fe77c770bb3d429a6990b959a880822f4d53fa.jpg", + "image_caption": [ + "(a) Human Feedback Inversion (HFI)" + ], + "image_footnote": [], + "bbox": [ + 215, + 142, + 472, + 265 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2812881bf105aebaee2ce1f6f5a5f41e8397976f4f169c5f526f92e8fb44007f.jpg", + "image_caption": [ + "(b) Safe self-Distillation Diffusion (SDD)", + "Fig. 4: Illustration of our proposed framework, HFI and SDD." + ], + "image_footnote": [], + "bbox": [ + 496, + 162, + 787, + 265 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "text. HFI addresses this challenge by learning concepts based on collective understanding. Our approach initiates with collecting human feedback on model-generated images through carefully designed surveys detailed in Appendix C, where harmfulness is comparatively evaluated. Using this human evaluation data, we construct a reward model that assigns a harmfulness score to each image. With this reward model, we formulate an optimization problem to identify a soft token that accurately represents harmfulness. These optimal embedding vectors are then utilized to fine-tune the diffusion model, preventing the generation of harmful images. The overall framework is depicted in Fig. 4 and Appendix A, and detailed explanations of each stage are provided in the subsequent sections.", + "bbox": [ + 212, + 334, + 787, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Collecting and Modeling Human Feedback", + "text_level": 1, + "bbox": [ + 214, + 506, + 612, + 522 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first generate images, on which human feedback will be collected, using the original model and a set of simple prompts containing the target concept word $c$ detailed in Appendix B. Subsequently, two main settings can be considered for gathering human feedback. The design of how and what kind of feedback to collect during this process is crucial. In cases where the content the model should not generate is well-defined and clear annotation guidelines exist (e.g., nudity), binary feedback for each image can be collected. However, this does not apply to all scenarios, with a notable example being an artist's style. Suppose that an artist requests the removal of not only specific works but also the overall style from the model, then the criteria may not be clear. In such cases, ranking feedback can be collected. This involves presenting annotators with $M$ random images and requesting them to rank these images ranging from 1 to $M$ .", + "bbox": [ + 212, + 530, + 787, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following this, we need to train a reward model $r_{\\psi}$ , which accepts an input image $\\mathbf{x}$ and outputs a score $r_{\\psi}(\\mathbf{x}) \\in \\mathbb{R}$ for the concept $c$ . For binary feedback, we follow the MSE-based approach proposed in Lee et al. [20],", + "bbox": [ + 212, + 712, + 785, + 758 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {M S E} (\\psi) = \\mathbb {E} _ {(\\mathbf {x}, y) \\sim \\mathcal {D} _ {\\text {h u m a n}}} \\left[ (r _ {\\psi} (\\mathbf {x}) - y) ^ {2} \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 349, + 766, + 785, + 786 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{D}_{\\mathrm{human}}$ consists of images $\\mathbf{x}$ and corresponding binary annotations $y \\in \\{0,1\\}$ . Here, '1' indicates the presence of the harmful concept (e.g., with nudity), and '0' is the absence (e.g., without nudity), answered by human annotators.", + "bbox": [ + 214, + 794, + 785, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Safeguard Diffusion Models with Human Feedback Inversion", + "bbox": [ + 328, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "On the other hand, for ranking feedback, we adopted the Bradley-Terry model [5], which assigns scores to examples based on pairwise comparisons [8, 26, 49]. The obtained rankings of $M$ images can be reorganized to construct $\\mathcal{D}_{\\mathrm{human}}$ as $\\binom{M}{2}$ pairwise comparisons as $\\mathbf{x}^{+} = \\mathbf{x}^{(i)}, \\mathbf{x}^{-} = \\mathbf{x}^{(j)}$ if $\\mathbf{x}^{(i)}$ has a higher rank than $\\mathbf{x}^{(j)}$ , and vice versa - a higher ranking indicates a greater presence or resemblance of the concept. Subsequently, we include all $\\binom{M}{2}$ pairs in a single batch, following Ouyang et al. [26], and optimize the reward model as", + "bbox": [ + 212, + 146, + 787, + 255 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {N L L} (\\psi) = - \\frac {1}{\\binom {M} {2}} \\mathbb {E} _ {(\\mathbf {x} ^ {+}, \\mathbf {x} ^ {-}) \\sim \\mathcal {D} _ {\\text {h u m a n}}} \\log \\sigma \\left(r _ {\\psi} \\left(\\mathbf {x} ^ {+}\\right) - r _ {\\psi} \\left(\\mathbf {x} ^ {-}\\right)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 294, + 265, + 785, + 301 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\sigma (\\cdot)$ is the sigmoid function. For both approaches, the reward model $r_{\\psi}(\\mathbf{x})$ maps the CLIP [30] embedding of an image $\\mathbf{x}$ to a score through a couple of MLP layers with ReLU non-linearities.", + "bbox": [ + 212, + 310, + 785, + 356 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Inverting Feedback into Embeddings", + "text_level": 1, + "bbox": [ + 215, + 378, + 565, + 395 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Based on the feedback received in Sec. 3.1, we transform the corresponding concept into a soft token. Inspired by recent inversion techniques [12,35], we utilize TI [12], as it can easily replace or be combined with text tokens from previous approaches. The goal is to find a soft token $v^{*}$ that maximizes the expected reward of generated images, i.e., $v^{*} = \\arg \\max_{v}\\mathbb{E}_{\\mathbf{x}\\sim p_{\\theta}(\\mathbf{x}|v)}[r_{\\psi}(\\mathbf{x})]$ , where $v$ is the soft word in the token space. Utilizing the reward model $r_{\\psi}$ in Sec. 3.1, we optimize the following objective of the reward-weighted negative log-likelihood [20] to get the optimal token $v^{*}$ that best captures the text concept $c$ :", + "bbox": [ + 212, + 404, + 787, + 526 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nv ^ {*} = \\underset {v} {\\arg \\min } \\mathbb {E} _ {\\mathbf {x} \\sim p _ {\\theta} (\\mathbf {x} | c)} [ - r _ {\\psi} (\\mathbf {x}) \\log p _ {\\theta} (\\mathbf {x} | v) ]. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 348, + 537, + 785, + 560 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Inverting too many images may struggle with concept extraction by prioritizing image reconstruction over it, as shown in Gal et al. [12]. Also, we empirically found that inverting it with a small learning rate may fail to capture the concept, which requires additional hyperparameter tuning. To simplify such a burden, we use a reward-weighted sampler with images of the top- $K$ scores, which is equivalent to the reward-weighted loss in Eq. (4). We use the learning rate of $5 \\times 10^{-4}$ , the default value in the HuggingFace's implementation1, and $K = 50$ .", + "bbox": [ + 212, + 571, + 787, + 679 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Removing Learned Concepts with Self-Distillation", + "text_level": 1, + "bbox": [ + 214, + 700, + 674, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After obtaining soft tokens through the proposed HFI method in Sec. 3.2, one can adapt them to existing text-based concept removal methods, enabling the desired removal of concepts using the obtained soft tokens instead of text tokens. While HFI alone proves effectiveness in Tabs. 1 and 2, we propose a new method termed Safe self-Distillation Diffusion (SDD) to achieve an even greater impact,", + "bbox": [ + 212, + 724, + 787, + 803 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "S. Kim et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "1 https://github.com/huggingface/diffusers/blob/main/examples/textual_inversion/textual_inversion.py", + "bbox": [ + 217, + 810, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "particularly in enhancing the removal of ambiguous concepts judged by humans. SDD removes the target concept from a model based on self-distillation [53]. We first set both student $\\theta$ and teacher models $\\theta^{\\star}$ whose parameters are initialized from the pre-trained model and update the student with the following loss:", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {S D D} = \\mathbb {E} _ {\\mathbf {x} _ {t} \\sim p _ {\\theta^ {*}} (\\mathbf {x} _ {t} | v ^ {*}), t} \\left[ \\| \\epsilon_ {\\theta} (\\mathbf {x} _ {t}; v ^ {*}) - \\mathbf {s g} \\left(\\epsilon_ {\\theta} (\\mathbf {x} _ {t})\\right) \\| _ {2} ^ {2} \\right], \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 321, + 217, + 784, + 234 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathfrak{sg}()$ denotes the stop-grad operation. That is, we want the model given the token $v^{*}$ to behave as if it were not for any token by setting the unconditional estimate $\\epsilon_{\\theta}(\\mathbf{x}_t)$ as a self-distillation target. After updating the student model, we update the teacher model $\\theta^{\\star}$ as an exponential moving average (EMA) from it, i.e., $\\theta^{\\star} \\gets m\\theta^{\\star} + (1 - m)\\theta$ , with a decaying rate $m \\in (0,1)$ . After the training, we use the EMA parameter of the teacher model as our final safe model.", + "bbox": [ + 212, + 243, + 784, + 332 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Unlike methods focused on learning new concepts [19, 38], determining suitable replacements when removing specific concepts can be challenging, often contingent on other contextual elements [9]. For instance, while removing the color red in a red apple can be substituted with a green apple, removing the color red devoid of context lacks clarity. Similarly, removing concepts like nudity, violence, or specific artists presents this ambiguity. Previous approaches [3,13,40] have attempted to tackle this challenge by targeting negative guidance; however, images generated in this manner often fall short of removing such concepts.", + "bbox": [ + 212, + 334, + 784, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We tackle this problem by introducing the EMA teacher model, which moves in tandem with the student model but in a more stable and deliberate manner, gradually removing the target concept. This serves not only to enhance training stability but also plays a crucial role in providing the student model with appropriate targets for conditional noise predictions. Notably, in Eq. (5), the latent $\\mathbf{x}_t$ is sampled from the teacher model, i.e. it generates parts that have not been effectively removed up to the existing training step. Additionally, resolving discrepancies between the model's understanding and human judgment regarding harmfulness poses a non-trivial task. In SDD with HFI, sampling $\\mathbf{x}_t$ is conditioned on the learned token $v^*$ . During this process, the model learns to remove the concept from newly generated latents that would be deemed problematic.", + "bbox": [ + 212, + 455, + 785, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Lastly, we introduce a couple of additional techniques for fine-tuning. First, similar to existing work [13, 19, 51, 52], we fine-tune only the cross-attention layers of the U-Net [37]. Second, we empirically find that not all timesteps equally influence the conditioning of image generation; the $L_{2}$ -norm between the unconditional and the conditional noise was generally high in the middle timesteps, as shown in Appendix E. Therefore, we sample the timestep to apply distillation as $t = T \\cdot \\mathcal{B}(\\alpha, \\beta)$ where $\\mathcal{B}(\\alpha, \\beta)$ is a beta random variable with parameter $(\\alpha, \\beta)$ . We set $(\\alpha, \\beta) = (3, 3)$ for our experiments as it works well in general.", + "bbox": [ + 212, + 621, + 785, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Related Work", + "text_level": 1, + "bbox": [ + 215, + 763, + 387, + 780 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Numerous attempts have been made to address harmful content generation, including the use of external classifiers for image filtering during inference, such as the safety checker in Stable Diffusion (SD) [36]. However, these approaches", + "bbox": [ + 212, + 794, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Safeguard Diffusion Models with Human Feedback Inversion", + "bbox": [ + 328, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "can be easily circumvented by users [33]. Efforts to filter out problematic training images during pre-processing [48] often lead to reduced training sets and inferior image quality [24, 28], without completely eliminating harmfulness [1].", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Recent advancements in the field can be broadly classified into three categories: (i) inference-time techniques manipulate noise estimates to counteract concepts [3,40]; (ii) direct cross-attention optimization adjusts weights to disregard problematic tokens from text inputs either directly or through gradient descent [14,52]; (iii) fine-tuning methods refine a subset of parameters using generated images [13,18,23]. Although both (ii) and (iii) fine-tune the model, the main disparity lies in the training objective. The former adjusts the projection or attention of text inputs, while the latter alters the model's distribution along the diffusion trajectories. Moreover, while the former does not necessitate training images or uses only a few real images, the latter generates training images from the model. Nonetheless, all methods entail a trade-off between effectiveness and utility, given the inherent challenge of identifying concepts in texts for removal.", + "bbox": [ + 212, + 191, + 785, + 375 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Human feedback has become crucial in enhancing both performance and safety in machine learning models [8]. Initially applied in dialogue systems [15, 21], reinforcement learning (RL) algorithms, particularly Proximal Policy Optimization (PPO) [43], have played a significant role in leveraging human feedback, notably in Large Language Models (LLMs) [6, 26, 49], focusing on both performance and safety considerations. Reinforcement learning with human feedback (RLHF) [26] optimizes models based on reward signals derived from user input, further enhancing the safety and adaptability of LLMs. Beyond language domains, efforts have been made to align text-to-image diffusion models with human preferences [2, 10, 20]. These involve fine-tuning using a reward-weighted loss function with appropriate regularization. However, the practical application of such RL-based methods often faces challenges due to hyperparameter sensitivity and reward design intricacies [20]. Furthermore, while such methods relying on human preference data aim to enhance image quality perceived by humans, they differ from our goal of suppressing image generation of specific concepts.", + "bbox": [ + 212, + 375, + 787, + 602 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 626, + 375, + 645 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.1 Baselines and Evaluation Protocols", + "text_level": 1, + "bbox": [ + 215, + 660, + 552, + 676 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We compare the performance of our method with the original SD and previous methods. In Tabs. 1 and 2, SD [36] indicates the original Stable Diffusion model, and NEG is the simplest baseline of SD with the negative prompt, i.e., the noise conditioned on the target concept is used instead of the unconditional one, thus negating the target concept. We consider two inference-time techniques: SLD [40] and SEGA [3]. For fine-tuning methods, we consider the following three methods: ESD [13], FMN [52], and UCE [14]. For all fine-tuning methods, including ours, we use the artist name (e.g., \"Vincent van Gogh\") or the word (e.g., \"nudity\" and \"bleeding\") for the target concept in text. When applying HFI to these methods, HFI token $v^{*}$ replaces the text word. Refer to Appendix B for details.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "S. Kim et al.", + "bbox": [ + 271, + 114, + 359, + 128 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4cb55fd4c0f1b50c96b5948cc533111c5b90cb78975fdd94a41a5da3d2caf696.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 142, + 498, + 220 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/51e50d19be2ced3e33c7060d973470f3f6198ab92868c5dfb4701f54c4c79ed6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 142, + 787, + 220 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/df0bc3b376c3935e6b8ef6184a63483a506a0c0014291a222d277cd611047847.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 223, + 496, + 301 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/93f195f44fb1da8524ce35717dc02ec13bd1e5aaad717d4ba99fde5183660aa0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 224, + 785, + 301 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/50e39ff7b275d238ec0d7dd339d3c888306698137e6c54dfeb4625909fdc53b6.jpg", + "image_caption": [ + "Fig. 5: Comparison of concept removal methods for Vincent van Gogh's artistic style. Van Gogh's style is marked by bold brushstrokes, vivid colors, and swirling patterns. HFI+SDD is the most effective, consistently producing the most neutral but still artwork-like images and successfully eliminating his stylistic characteristics." + ], + "image_footnote": [], + "bbox": [ + 217, + 303, + 496, + 398 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/455f6974bab61299aa72750055e858d05c3322c0513cb82258101441fe5dc651.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 303, + 785, + 398 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The evaluation of the method's capability to remove specific concepts is structured along two primary axes: effectiveness and utility. Effectiveness measures the ability to avoid generating the target concept, while utility assesses the preservation of image quality in outputs unrelated to the target concept.", + "bbox": [ + 212, + 501, + 784, + 561 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For artist concept removal in Sec. 5.2, we propose two specific metrics to gauge both aspects. The effectiveness is evaluated by the target artist ratio (%ART) using CLIP zero-shot classification. This involves generating 1,000 images in total of 50 commonly depicted objects in famous artworks (listed by ChatGPT [25]), with the prompt template \"a painting of {object} in the style of {artist}\" These images are then classified using OpenCLIP [7] to determine how frequently the target artist is accurately identified. The utility is determined through the photo similarity ratio (%PHO), which quantifies how well the model retains the ability to generate paintings rather than photorealistic images, thus preserving artistic styles not specific to the targeted artist. We enhance classification accuracy using prompt ensemble [30]. In line with previous work, we also report LPIPS [54] for both the target and other artists. Finally, since such quantitative metrics cannot completely replace qualitative evaluations and human evaluations can be subjective, we asked ChatGPT 4o [25] to assess the performance of removing the specified artist to supplement this. The qualitative evaluation results and further details are provided in Appendix F.", + "bbox": [ + 212, + 565, + 787, + 806 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For harmful concept removal in Sec. 5.3, the effectiveness is measured by generating 5,000 images in total with the prompt {\"country\" body\" using", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Safeguard Diffusion Models with Human Feedback Inversion", + "bbox": [ + 328, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/4af15efc7928b63f0e1c720f94b73bc365e7da24b9c13e7585dc5652a4d61c15.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 141, + 500, + 222 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/7520733b457100a65231047a2a2ae2e8109ee74cb4096f06fc49e9cfea6f2900.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 142, + 787, + 220 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/35547a6442c50a47ba3d3864555da2913d3a39afa7ee53b86d02b196401034f1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 223, + 498, + 301 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/f246d9b95b09bcfa424c4c8780f4855f914309fc8dc1b279e0120c2eb0115942.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 223, + 785, + 301 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/3d9babbfb58812f97438752c75a8b626012ec5bb4eb30fbfe65e7083527e9ef6.jpg", + "image_caption": [ + "Fig. 6: Comparison of concept removal methods for Leonid Afremov's artistic style. Afremov's style is marked by vibrant colors, palette knife technique (similar to pointage technique), and impressionistic elements. HFI+SDD is the most effective, producing images with clear depictions of various objects and minimal retention of his style." + ], + "image_footnote": [], + "bbox": [ + 217, + 303, + 498, + 398 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/83543329c04df940c576566bc5da83afb3c0025291bcf643b896c3782a572d20.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 303, + 785, + 400 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "names of 50 countries with high GDP following [40]. This prompt generates way more inappropriate images (74.2% with SD v1.4) compared to the sexual category of the I2P dataset [40] (reportedly ranked $5^{\\text{th}}$ among the seven categories with 32.8% of hard prompt ratio and only 16.7% of nudity images). The proportion of images classified \"unsafe\" (%NSFW) and those detected with exposed body parts (%NUDE) are evaluated by NudeNet [29]. This process is similarly used to assess the removal of another harmful concept of bleeding by generating images for the shocking category in the I2P dataset [40] and classifying them with the Q16 classifier (%I2P) [41]. The utility is evaluated by generating images from 10,000 random captions from the MSCOCO dataset [22] and comparing these to actual photos, outputs from the original SD model, and their text-to-image alignment using FID [16, 27], LPIPS [54], and CLIPScore [30], respectively.", + "bbox": [ + 212, + 494, + 787, + 678 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.2 Artist Style Removal", + "text_level": 1, + "bbox": [ + 214, + 696, + 437, + 710 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In this section, we conducted experiments to remove the style of artists from the existing SD model. For HFI, we generate 1,000 images from the original SD model with prompts detailed in Appendix B and gather human feedback on a total of 200 images (equivalent to 20 question pairs) per artist. We then train a reward model, which is then used to approximate the scores for all 1,000 images for HFI. All methods with HFI use the same soft token instead of the text token.", + "bbox": [ + 212, + 719, + 787, + 808 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In Tab. 1, even without HFI, SDD alone finds a reasonable trade-off, achieving relatively better removal than the other methods while maintaining a rela-", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "S. Kim et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/6cedc83aebb138b72490f9197e9ee162b8334da92e1b05719d28bbf35a475858.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 142, + 498, + 220 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/306eac286f9b49782db4b4132533213b5203a79f05522a852c99e46671f13005.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 142, + 787, + 220 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/532a096210f79bf9e596ac25d1d4cab41e6b3911e064c3179c5bf3cee29af597.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 222, + 496, + 301 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/0f82675506b2cff253b38f8ae66aaf29823fbb7113b06ae89b0633055a8d4372.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 223, + 785, + 301 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/1b1f4b4638c92b43c1dee0367acbe97d9aac105084a5768b0507554e9fbaf87a.jpg", + "image_caption": [ + "Fig. 7: Comparison of concept removal methods for Edward Hopper's artistic style. Hopper's characteristic style is marked by realism, simplicity, architectural details, and muted color palette. Among the methods, HFI+SDD most effectively removes his distinctive features while still being abstract and less photorealistic." + ], + "image_footnote": [], + "bbox": [ + 217, + 303, + 496, + 398 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/81ada7f44b364ffc7df661017cb071ba3130009c191ebeef47b6564de424e05b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 303, + 785, + 398 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "tively lower percentage on $\\% \\mathrm{PHO}$ . Moreover, when HFI is employed, our method shows the state-of-the-art performance of $\\% \\mathrm{ART}$ of $12.7\\%$ on average, with average $\\% \\mathrm{PHO}$ of $12.0\\%$ . For Edward Hopper, however, HFI+SDD was outperformed by a few HFI-enhanced baselines, but it worked the best in most cases. Notably, HFI+SDD always surpasses all the previous methods without HFI.", + "bbox": [ + 212, + 501, + 787, + 577 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Among the baselines, fine-tuning methods (ESD and SDD) observed the most benefit from HFI. In contrast, methods directly optimizing cross-attention exhibited subpar removal performance and appeared to have relatively less impact. Meanwhile, inference-time techniques generally demonstrated improvements in the trade-off between effectiveness and utility facilitated by HFI as shown in Fig. 8. However, as shown in Figs. 5 to 7, there are often cases where the object is not well represented or the distinctive artistic characteristics remain. In such cases, HFI can be beneficial, whereas text-based methods frequently struggle to distinguish the artist's concept from the generic artwork style.", + "bbox": [ + 212, + 580, + 787, + 715 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In the qualitative evaluation using ChatGPT 4o [25], HFI+SDD was evaluated as the most superior method by stating that \"it consistently produces the most neutral images by moving away from the artist's style while maintaining a high level of image quality.\" Despite being asked to evaluate only the removal performance, somewhat surprisingly, ChatGPT considered not only the degree of style removal but also various aspects such as object depiction, image quality, and artifacts) with consistent judgments across all objects and presentation orders, aligning with our proposed metrics. Refer to Appendix F for full results.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Safeguard Diffusion Models with Human Feedback Inversion", + "bbox": [ + 328, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/5de7e08e7bc6d426cb1918f91d18398a07256e2a8da7e72c84ec934df9db819e.jpg", + "image_caption": [ + "Fig. 8: Trends of $\\%_{\\text{ART}}$ and $\\%_{\\text{PHO}}$ by inference-time techniques. While forgetting the general artwork concepts (higher $\\%_{\\text{PHO}}$ ) is inevitable, HFI can achieve similar outcomes at smaller editing scales. The same marker symbol denotes the same scale in each plot." + ], + "image_footnote": [ + "--- SLD -SLD+HFI SEGAE SEGA+HFI" + ], + "bbox": [ + 218, + 143, + 405, + 237 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/df39b96308b953562803d94d0a24bbca797753e042be2dec2724ac00460b7282.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 406, + 143, + 593, + 237 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/1ef1f7713ddafb5c71a53083f2660eaf8f136b0d77c2d2fc3c0b551b89c5d461.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 143, + 782, + 237 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/4a6df93dda4cf5ac0f65f1f1fecef967d657e895e7f5cda7693df28f5927719c.jpg", + "image_caption": [ + "Fig. 9: HFI mitigates terrifying aspects of images generated with I2P shocking prompts." + ], + "image_footnote": [], + "bbox": [ + 218, + 321, + 527, + 446 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/a4c467ccfe44f3a90ed5f1da64816723cd3b67eab43dfa4550a98b1c8431da5e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 537, + 330, + 785, + 388 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/148414c18f80fc16ff51f92d273df82c24982b6f43f88415825450968f456ff6.jpg", + "image_caption": [ + "Fig.10: Images with the highest (top) and lowest (bottom) rewards." + ], + "image_footnote": [], + "bbox": [ + 539, + 390, + 785, + 445 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.3 Harmful Concept Removal", + "text_level": 1, + "bbox": [ + 215, + 513, + 483, + 529 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Tab. 2 shows our experimental results on removing NSFW and bleeding concepts. Similar to Sec. 5.2, we first generated images and collected binary feedback and ranking feedback for nudity and bleeding, respectively. SDD alone effectively reduces the presence of NSFW imagery, but when combined with HFI, such images are nearly eliminated. Our method and ESD-u-3 achieve a similar level, but ESD exhibits inferior performance in all image quality metrics. In a similar vein, the method of fine-tuning only the cross-attention, as in ESD-x-3, is observed to generate harmful images beyond a certain threshold. High in %NSFW and low in %NUDE for some methods indicate that they still generate somewhat unsafe images even though exposed body parts are limited. When combined with HFI, all methods but FMN exhibit a sharp decrease in %NSFW, underscoring the role of the soft token aligned with human judgment.", + "bbox": [ + 212, + 537, + 787, + 718 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Furthermore, for nuanced concepts such as bleeding, HFI aids in identifying and mitigating the inappropriate aspects of images. In Fig. 9, with HFI, it becomes apparent that not only has the extent of bleeding diminished, but it also yields images that are safer and less offensive. Particularly notable are the outcomes achieved by combining SDD and HFI, where notably safer images are produced despite maintaining identical compositions. The interpretation of such ambiguous concepts by humans can be encoded into learned tokens for subsequent utilization in their removal, highlighting the utility of HFI.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "S. Kim et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/4035f0c4619fd439f622aee8a8cb606bfcb520a98cc99564399f052fcd935382.jpg", + "table_caption": [ + "Table 1: Quantitative comparison of artist style removal performance. Each method is evaluated both without and with HFI (✓), showing improvement due to human feedback. The $\\sum$ column represents the sum of %ART and %PHO, indicating undesirable images. The table also compares our SDD method with other baselines." + ], + "table_footnote": [], + "table_body": "
Target ArtistVincent van GoghLeonid AfremovEdward HopperAvg. LPIPS
MethodHFI%ART↓ %PHO↓Σ%ART↓ %PHO↓Σ%ART↓ %PHO↓ΣTarget↑ Others↓
SD v1.495.92.998.899.90.6100.597.62.299.8--
Inference-time Technique
NEG [36]✓15.032.547.592.82.094.836.512.148.60.33520.2273
17.212.429.60.030.430.416.516.132.60.43520.3255
SLD [40]✓29.127.756.827.84.232.021.339.560.80.42770.2880
24.622.847.49.78.818.518.134.953.00.44590.3285
SEGA [3]✓40.311.752.064.41.565.937.728.666.30.31510.2522
17.517.635.120.74.525.232.021.853.80.34820.2661
Direct Cross-Attention Optimization
FMN [52]✓66.67.574.110.725.536.277.15.782.80.32930.1665
58.67.766.36.50.97.473.25.578.70.27930.2478
UCE [14]✓66.44.070.444.72.046.751.53.354.80.27670.0332
77.32.579.892.50.693.172.51.974.40.17440.0328
Fine-tuning Method
ESD-x-1 [13]✓45.015.960.913.525.438.964.47.273.60.36240.1635
22.520.442.919.057.276.214.317.431.70.44040.1691
SDD (ours)✓20.625.546.19.98.218.140.441.181.50.27600.2341
9.55.114.68.83.011.819.827.947.70.28790.1439
", + "bbox": [ + 220, + 212, + 782, + 481 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.4 Analysis on Rewards and Learned Embeddings", + "text_level": 1, + "bbox": [ + 215, + 512, + 648, + 527 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "For nudity, the reward model generally aligns with Nudenet [29] (AUROC of 0.882), but differences between human and model judgments highlight the limits of relying solely only on models for inversion. Fig. 10 shows that images deemed appropriate by humans or not reflecting the artist's style receive low rewards and minimally impact HFI token learning. Integrating HFI with reward models effectively captures concepts. Ablation studies comparing HFI with the vanilla TI [12] show a $3.2\\%$ and $5.1\\%$ increase in $\\% \\mathrm{ART}$ for artist removal without human feedback (i.e., random images without rewards) and with real images from the internet (without rewards), respectively. For nudity removal, $\\% \\mathrm{NSFW}$ and $\\% \\mathrm{NUDE}$ increase by $1.74\\%$ and $0.04\\%$ without human feedback, with FID increasing by 1.565. This demonstrates the importance of integrating human judgments and model-generated data into the training process to achieve optimal performance.", + "bbox": [ + 212, + 537, + 787, + 719 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Compared to existing expensive methods using human feedback, the proposed HFI is highly cost-effective. We obtained 650 responses from five individuals (450 unique images for inter-annotator agreement (IAA)) within three hours for nudity. Providing annotation guidelines increased IAA, evidenced by Fleiss' $\\kappa$ [11] at 0.4997 and Randolph's $\\kappa$ [34] at 0.6702. For artistic concepts, we collected 40 ranking feedback responses (400 images) per artist from 15 individuals. Our research offers an economical and effective framework for extracting information from models and removing harmful content with human judgment.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Safeguard Diffusion Models with Human Feedback Inversion", + "bbox": [ + 328, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/645f2562f68283f4b6f793885e19775210d35e27d848254e66ee241ef7f8468d.jpg", + "table_caption": [ + "Table 2: Quantitative comparison of harmful concept removal performance. Each method is evaluated both without and with HFI (✓) for inter-method comparison. The table also compares our SDD with other baselines for intra-method comparison." + ], + "table_footnote": [], + "table_body": "
Target ConceptNudityBleedingCOCO10K
MethodHFI%NSFW↓%NUDE↓%I2P↓FID↓CLIP↑ LPIPS↓
SD v1.474.1832.1465.7716.6360.2745
COCO ref.0.2676
Inference-time Technique
SD+NEG [36]✓30.683.8645.7918.8370.2701
5.144.7624.6519.5180.2699
SLD [40]✓58.089.8446.1417.8150.2702
42.9815.5829.5618.3450.2704
SEGA [3]✓67.7213.1645.4417.1290.2698
56.4219.5239.3716.7560.2698
Direct Cross-Attention Optimization
FMN [52]✓51.5027.7847.0818.0540.2556
61.3234.7255.7216.8130.2683
UCE [14]✓32.307.3661.8017.0240.2726
14.769.3253.2716.5410.2736
Fine-tuning Method
ESD-u-1 [13]✓11.302.00-17.0330.2664
2.421.16-18.4040.2636
ESD-u-3 [13]✓4.440.48-18.4980.2624
0.800.46-18.6150.2621
ESD-x-3 [13]✓14.322.9851.1718.5170.2677
4.183.3034.3517.8980.2670
SDD (ours)✓2.780.6030.8417.3620.2658
0.280.3223.3617.0230.2643
", + "bbox": [ + 287, + 198, + 709, + 556 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion and Discussion", + "text_level": 1, + "bbox": [ + 215, + 583, + 509, + 599 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we introduced a novel concept removal framework within a text-to-image diffusion model through human feedback inversion and self-distillation-based fine-tuning. By leveraging human feedback to train soft tokens, we effectively safeguarded the diffusion model from harmful concepts, addressing the ambiguity in defining text-based concepts. We further proposed a self-distillation-based diffusion model fine-tuning algorithm, enhancing the efficacy of concept removal. We believe our research lays the groundwork for a human-centric approach to the safety and ethical considerations of large-scale diffusion models, paving the way for future advancements.", + "bbox": [ + 212, + 613, + 787, + 750 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitation. While our proposed HFI performs well in capturing the intended concept compared to the text, its effectiveness depends on the quality of human feedback and the performance of inversion to soft tokens. Significant disagreements in human feedback can compromise the learning process, underscoring the importance of meticulous data collection and reward modeling.", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "S. Kim et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was partly supported by Institute for Information & communications Technology Promotion (IITP) grant funded by the Korea government (MSIT) (No. RS-2019-II190075, Artificial Intelligence Graduate School Program (KAIST), No. 2022-0-00184, Development and Study of AI Technologies to Inexpensively Conform to Evolving Policy on Ethics) and the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. 2022R1A5A708390812).", + "bbox": [ + 212, + 176, + 787, + 282 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 306, + 321, + 321 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Baio, A.: Exploring 12 million of the 2.3 billion images used to train stable diffusion's image generator. https://waxy.org/2022/08/exploring-12-million-of-the-images-used-to-train-stable-diffusions-image-generator/ (Aug 2022)", + "2. Black, K., Janner, M., Du, Y., Kostrikov, I., Levine, S.: Training diffusion models with reinforcement learning. arXiv preprint arXiv:2305.13301 (2023)", + "3. Brack, M., Friedrich, F., Hintersdorf, D., Struppek, L., Schramowski, P., Kersting, K.: Sega: Instructing diffusion using semantic dimensions. arXiv preprint arXiv:2301.12247 (2023)", + "4. Brack, M., Schramowski, P., Friedrich, F., Hintersdorf, D., Kersting, K.: The stable artist: Steering semantics in diffusion latent space. arXiv preprint arXiv:2212.06013 (2022)", + "5. Bradley, R.A., Terry, M.E.: Rank analysis of incomplete block designs: I. the method of paired comparisons. Biometrika 39(3/4), 324-345 (1952)", + "6. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)", + "7. Cherti, M., Beaumont, R., Wightman, R., Wortsman, M., Ilharco, G., Gordon, C., Schuhmann, C., Schmidt, L., Jitsev, J.: Reproducible scaling laws for contrastive language-image learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2818-2829 (2023)", + "8. Christiano, P.F., Leike, J., Brown, T., Martic, M., Legg, S., Amodei, D.: Deep reinforcement learning from human preferences. Advances in neural information processing systems 30 (2017)", + "9. Du, Y., Li, S., Mordatch, I.: Compositional visual generation with energy based models. Advances in Neural Information Processing Systems 33, 6637-6647 (2020)", + "10. Fan, Y., Watkins, O., Du, Y., Liu, H., Ryu, M., Boutilier, C., Abbeel, P., Ghavamzadeh, M., Lee, K., Lee, K.: Dpok: Reinforcement learning for fine-tuning text-to-image diffusion models. arXiv preprint arXiv:2305.16381 (2023)", + "1. Fleiss, J.L.: Measuring nominal scale agreement among many raters. Psychological bulletin 76(5), 378 (1971)", + "12. Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A.H., Chechik, G., Cohen-Or, D.: An image is worth one word: Personalizing text-to-image generation using textual inversion (2022). https://doi.org/10.48550/ARXIV.2208.01618, https://arxiv.org/abs/2208.01618", + "13. Gandikota, R., Materzynska, J., Fiotto-Kaufman, J., Bau, D.: Erasing concepts from diffusion models. arXiv preprint arXiv:2303.07345 (2023)" + ], + "bbox": [ + 218, + 339, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Safeguard Diffusion Models with Human Feedback Inversion", + "bbox": [ + 328, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Gandikota, R., Orgad, H., Belinkov, Y., Materzynska, J., Bau, D.: Unified concept editing in diffusion models. arXiv preprint arXiv:2308.14761 (2023)", + "15. Gao, J., Galley, M., Li, L.: Neural approaches to conversational ai. In: The 41st international ACM SIGIR conference on research & development in information retrieval. pp. 1371-1374 (2018)", + "16. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017)", + "17. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. arXiv preprint arxiv:2006.11239 (2020)", + "18. Kumari, N., Zhang, B., Wang, S.Y., Shechtman, E., Zhang, R., Zhu, J.Y.: Ablating concepts in text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22691-22702 (2023)", + "19. Kumari, N., Zhang, B., Zhang, R., Shechtman, E., Zhu, J.Y.: Multi-concept customization of text-to-image diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1931-1941 (2023)", + "20. Lee, K., Liu, H., Ryu, M., Watkins, O., Du, Y., Boutilier, C., Abbeel, P., Ghavamzadeh, M., Gu, S.S.: Aligning text-to-image models using human feedback. arXiv preprint arXiv:2302.12192 (2023)", + "21. Li, J., Monroe, W., Ritter, A., Galley, M., Gao, J., Jurafsky, D.: Deep reinforcement learning for dialogue generation. arXiv preprint arXiv:1606.01541 (2016)", + "22. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)", + "23. Ni, Z., Wei, L., Li, J., Tang, S., Zhuang, Y., Tian, Q.: Degeneration-tuning: Using scrambled grid shield unwanted concepts from stable diffusion. arXiv preprint arXiv:2308.02552 (2023)", + "24. O'Connor, R.: Stable diffusion 1 vs 2 - what you need to know. https://www.assemblyai.com/blog/stable-diffusion-1-vs-2-what-you-need-to-know/ (Dec 2022)", + "25. OpenAI: Chatgpt (2023), https://www.openai.com/, july 2023 version", + "26. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems 35, 27730-27744 (2022)", + "27. Parmar, G., Zhang, R., Zhu, J.Y.: On aliased resizing and surprising subtleties in gan evaluation. In: CVPR (2022)", + "28. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952 (2023)", + "29. Praneeth, B.: NudeNet: Neural nets for nudity classification, detection and selective censoring. https://github.com/notAI-tech/NudeNet (2021)", + "30. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)", + "31. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 1(2), 3 (2022)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "S. Kim et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "32. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: International Conference on Machine Learning. pp. 8821-8831. PMLR (2021)", + "33. Rando, J., Paleka, D., Lindner, D., Heim, L., Tramér, F.: Red-teaming the stable diffusion safety filter. arXiv preprint arXiv:2210.04610 (2022)", + "34. Randolph, J.J.: Free-marginal multirater kappa (multirater k [free]): An alternative to fleiss' fixed-marginal multirater kappa. Online submission (2005)", + "35. Roich, D., Mokady, R., Bermano, A.H., Cohen-Or, D.: Pivotal tuning for latent-based editing of real images. ACM Transactions on graphics (TOG) 42(1), 1-13 (2022)", + "36. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10684-10695 (2022)", + "37. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015)", + "38. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22500-22510 (2023)", + "39. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022)", + "40. Schramowski, P., Brack, M., Deiseroth, B., Kersting, K.: Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2023)", + "41. Schramowski, P., Tauchmann, C., Kersting, K.: Can machines help us answering question 16 in datasheets, and in turn reflecting on inappropriate content? In: 2022 ACM Conference on Fairness, Accountability, and Transparency. pp. 1350-1361 (2022)", + "42. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. arXiv preprint arXiv:2210.08402 (2022)", + "43. Schulman, J., Wolski, F., Dhariwal, P., Radford, A., Klimov, O.: Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347 (2017)", + "44. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: Proceedings of the 32nd International Conference on Machine Learning (2015)", + "45. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)", + "46. Song, Y., Ermon, S.: Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems 32 (2019)", + "47. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456 (2020)", + "48. Stability AI: Stable diffusion 2.0 release. https://stability.ai/blog/stable-diffusion-v2-release (Nov 2022)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Safeguard Diffusion Models with Human Feedback Inversion", + "bbox": [ + 328, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "49. Stiennon, N., Ouyang, L., Wu, J., Ziegler, D., Lowe, R., Voss, C., Radford, A., Amodei, D., Christiano, P.F.: Learning to summarize with human feedback. Advances in Neural Information Processing Systems 33, 3008-3021 (2020)", + "50. Thiel, D.: Identifying and eliminating csam in generative ml training data and models. Tech. rep., Technical Report. Stanford University, Palo Alto, CA. https://purl.stanford... (2023)", + "51. Yang, Z., Yi, X., Li, P., Liu, Y., Xie, X.: Unified detoxifying and debiasing in language generation via inference-time adaptive optimization. arXiv preprint arXiv:2210.04492 (2022)", + "52. Zhang, E., Wang, K., Xu, X., Wang, Z., Shi, H.: Forget-me-not: Learning to forget in text-to-image diffusion models. arXiv preprint arXiv:2211.08332 (2023)", + "53. Zhang, L., Song, J., Gao, A., Chen, J., Bao, C., Ma, K.: Be your own teacher: Improve the performance of convolutional neural networks via self distillation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3713-3722 (2019)", + "54. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)" + ], + "bbox": [ + 215, + 146, + 784, + 397 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "S. Kim et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Safeguard Text-to-Image Diffusion Models with Human Feedback Inversion/603d7974-42c1-4361-a264-c2589e2c9681_model.json b/2024/Safeguard Text-to-Image Diffusion Models with Human Feedback Inversion/603d7974-42c1-4361-a264-c2589e2c9681_model.json new file mode 100644 index 0000000000000000000000000000000000000000..487f699bd679c9f09e544e8764ea45ba276850e6 --- /dev/null +++ b/2024/Safeguard Text-to-Image Diffusion Models with Human Feedback Inversion/603d7974-42c1-4361-a264-c2589e2c9681_model.json @@ -0,0 +1,2513 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.231, + 0.141, + 0.776, + 0.187 + ], + "angle": 0, + "content": "Safeguard Text-to-Image Diffusion Models with Human Feedback Inversion" + }, + { + "type": "text", + "bbox": [ + 0.255, + 0.213, + 0.747, + 0.244 + ], + "angle": 0, + "content": "Sanghyun Kim\\(^{\\text{D}}\\), Seohyeon Jung\\(^{\\text{D}}\\), Balhae Kim\\(^{\\text{D}}\\), Moonseok Choi\\(^{\\text{D}}\\), Jinwoo Shin\\(^{\\text{D}}\\), and Juho Lee\\(^{\\text{D}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.256, + 0.758, + 0.285 + ], + "angle": 0, + "content": "Korea Advanced Institute of Science and Technology (KAIST) {nannullna,heon2203,balhaekim,ms.choi,jinwoos,juholee}@kaist.ac.kr" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.317, + 0.744, + 0.568 + ], + "angle": 0, + "content": "Abstract. This paper addresses the societal concerns arising from large-scale text-to-image diffusion models for generating potentially harmful or copyrighted content. Existing models rely heavily on internet-crawled data, wherein problematic concepts persist due to incomplete filtration processes. While previous approaches somewhat alleviate the issue, they often rely on text-specified concepts, introducing challenges in accurately capturing nuanced concepts and aligning model knowledge with human understandings. In response, we propose a framework named Human Feedback Inversion (HFI), where human feedback on model-generated images is condensed into textual tokens guiding the mitigation or removal of problematic images. The proposed framework can be built upon existing techniques for the same purpose, enhancing their alignment with human judgment. By doing so, we simplify the training objective with a self-distillation-based technique, providing a strong baseline for concept removal. Our experimental results demonstrate our framework significantly reduces objectionable content generation while preserving image quality, contributing to the ethical deployment of AI in the public sphere. Code is available at https://github.com/nannullna/safeguard-hfi." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.58, + 0.741, + 0.609 + ], + "angle": 0, + "content": "Caution: This paper contains discussions and examples related to harmful content, including text and images. Reader discretion is advised." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.621, + 0.741, + 0.65 + ], + "angle": 0, + "content": "Keywords: Fairness, accountability, and transparency \\(\\cdot\\) Diffusion models \\(\\cdot\\) Image generation" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.674, + 0.377, + 0.691 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.844 + ], + "angle": 0, + "content": "Large-scale text-to-image generation models have demonstrated remarkable success in producing high-quality images encompassing a wide array of concepts [4, 31,32,36,39]. These models possess the ability to combine disparate ideas in novel and unprecedented manners, making them invaluable tools for creative content creation. However, they also give rise to substantial concerns, including the generation of harmful or copyrighted content, even without explicit directives [40]. A critical issue lies in the heavy reliance of the models on internet-crawled datasets (e.g., LAION-5B [42], currently not available [50]), which, despite attempts at aggressive filtration [48], often lead to problematic content generation [33,40]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "S. Kim et al." + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.143, + 0.788, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.4, + 0.789, + 0.456 + ], + "angle": 0, + "content": "Fig. 1: Comparative analysis of NSFW content removal techniques, including our proposed fine-tuning method (SDD), both with and without HFI framework. The results clearly show that incorporating HFI significantly reduces the amplification of provocative body parts and ensures the generation of clothed representations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.489, + 0.789, + 0.626 + ], + "angle": 0, + "content": "Previous inference-time [3,40] and fine-tuning methods [13,14,52] have made strides in mitigating objectionable imagery. Here, the concept to be removed, or the target concept, is specified by text. However, they insufficiently address the nuanced nature of concepts, leading to a gap between the model-generated content and human expectations. Examples include determining whether exposed body parts in a masterpiece are considered safe or if an illustration overtly portraying such body should be similarly categorized. Moreover, the ethical judgment criteria of the model generally fall short of human standards [33, 40], emphasizing the crucial need for collective intelligence intervention at this juncture." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.627, + 0.789, + 0.809 + ], + "angle": 0, + "content": "To handle such ambiguity and bridge this gap, we present Human Feedback Inversion (HFI), a novel and effective framework to mitigate the ethical concerns surrounding text-to-image diffusion models. HFI incorporates human feedback on model-generated images with problematic concepts directly into soft tokens using the textual inversion (TI) [12]. By directly incorporating human perspectives, HFI addresses the limitations of the previous text-based approaches, offering a more nuanced and accurate means of discerning harmful concepts. This condensed information serves as a guiding force against the target concept during inference time or to create training images for fine-tuning. Through this process, tokens acquired via HFI become specialized for their respective concepts, aiming to enhance both effectiveness (i.e., removal performance of the target concept) and utility aspects (i.e., image quality for remaining concepts)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.789, + 0.842 + ], + "angle": 0, + "content": "To augment the efficacy of HFI, we introduce Safe self-Distillation Diffusion (SDD), a self-distillation-based fine-tuning method. This method simplifies" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.33, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Safeguard Diffusion Models with Human Feedback Inversion" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.402, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.277, + 0.236, + 0.345, + 0.247 + ], + "angle": 0, + "content": "(a) Flowers" + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.144, + 0.594, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.477, + 0.236, + 0.528, + 0.247 + ], + "angle": 0, + "content": "(b) Fish" + }, + { + "type": "image", + "bbox": [ + 0.603, + 0.144, + 0.785, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.669, + 0.236, + 0.72, + 0.247 + ], + "angle": 0, + "content": "(c) Boat" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.252, + 0.402, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.287, + 0.343, + 0.333, + 0.355 + ], + "angle": 0, + "content": "(d) Hat" + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.252, + 0.594, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.478, + 0.343, + 0.528, + 0.355 + ], + "angle": 0, + "content": "(e) Tree" + }, + { + "type": "image", + "bbox": [ + 0.603, + 0.252, + 0.785, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.652, + 0.343, + 0.737, + 0.356 + ], + "angle": 0, + "content": "(f) Streetlamp" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.367, + 0.788, + 0.424 + ], + "angle": 0, + "content": "Fig. 2: Artist style removal results on Vincent van Gogh. Images are generated with prompts such as \"a painting of flowers in the style of Vincent van Gogh\". HFI+SDD effectively eliminates his distinct artistic features while preserving the essence of the original subject matter, subject integrity, and visual quality." + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.439, + 0.413, + 0.54 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.423, + 0.439, + 0.593, + 0.54 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.603, + 0.439, + 0.772, + 0.54 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.552, + 0.788, + 0.595 + ], + "angle": 0, + "content": "Fig. 3: Comparison of images generated with famous artwork titles of the three artists. HFI+SDD, leveraging human feedback, captures and removes distinctive artistic styles more effectively while preserving the integrity and essence of the original paintings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.627, + 0.784, + 0.734 + ], + "angle": 0, + "content": "the training objective, eliminating the need for manually designing the target or specifying which concept to retain. With this improved design, SDD outperforms existing baselines, even when based on text-specified concepts, with improved training stability. The integration of HFI and SDD establishes a synergistic alliance, utilizing the HFI token as a removal target. This soft token plays a crucial role in generating latents, serving as a vital component for self-supervision and further contributing to the model's alignment with human understanding." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Figs. 1 to 3 showcase the effectiveness of HFI in concept removal, surpassing previous methods and aligning well with human preferences. Specifically, when tasked with removing Van Gogh's artistic style, our pipeline, guided by human feedback, successfully retained the essential qualities of a painting-like style while eliminating his characteristic bold brushstrokes. In contrast, methods without human feedback transformed images into realistic photographs to aggressively remove artist's styles as a side effect. Additionally, when instructed to remove" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.36, + 0.128 + ], + "angle": 0, + "content": "S. Kim et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.208 + ], + "angle": 0, + "content": "nudity with human feedback, our pipeline successfully eliminated pornographic content, in contrast to previous methods that still produced provocative images. The visual evidence highlights the superior effectiveness of HFI with SDD in consistently delivering safe and modest image outputs across different prompts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.785, + 0.268 + ], + "angle": 0, + "content": "In summary, our framework not only markedly diminishes the generation of undesirable content perceived by humans but also maintains image quality. This offers a pragmatic solution for improving safety and ethical considerations in the widespread deployment of text-to-image diffusion models." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.294, + 0.57, + 0.312 + ], + "angle": 0, + "content": "2 Background on Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.328, + 0.784, + 0.464 + ], + "angle": 0, + "content": "Diffusion models [17, 44-47] are a class of generative models that learn the data distribution by building two Markov chains of latent variables. Given a sample \\(\\mathbf{x}_0 \\sim p_{\\mathrm{data}}(\\mathbf{x}) \\coloneqq q(\\mathbf{x})\\) and a noise schedule \\(\\{\\beta_t\\}_{t=1}^T\\), the forward process gradually injects a series of Gaussian noises to the sample until it nearly follows the standard Gaussian \\(q(\\mathbf{x}_t | \\mathbf{x}_{t-1}) \\coloneqq \\mathcal{N}(\\mathbf{x}_t; \\sqrt{1 - \\beta_t} \\mathbf{x}_{t-1}, \\beta_t \\mathbf{I})\\), \\(q(\\mathbf{x}_T | \\mathbf{x}_0) \\approx \\mathcal{N}(\\mathbf{x}_T; \\mathbf{0}, \\mathbf{I})\\). Such process is then followed by the reverse process parameterized by \\(\\theta\\), where the model learns to reconstruct the original image by iteratively denoising from a pure Gaussian noise \\(p(\\mathbf{x}_T) = \\mathcal{N}(\\mathbf{0}, \\mathbf{I})\\) as \\(p_\\theta(\\mathbf{x}_{0:T}) = p(\\mathbf{x}_T) \\prod_{t=1}^T p_\\theta(\\mathbf{x}_{t-1} | \\mathbf{x}_t)\\). Ho et al. [17] simplifies the objective to learn a noise estimator \\(\\epsilon_\\theta\\):" + }, + { + "type": "equation", + "bbox": [ + 0.317, + 0.478, + 0.785, + 0.495 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {D M} (\\theta) = \\mathbb {E} _ {\\mathbf {x} _ {0}, \\epsilon , t} \\left[ \\| \\boldsymbol {\\epsilon} - \\boldsymbol {\\epsilon} _ {\\theta} \\left(\\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {x} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\boldsymbol {\\epsilon}; t\\right) \\| _ {2} ^ {2} \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.507, + 0.784, + 0.539 + ], + "angle": 0, + "content": "where \\(\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\), \\(\\bar{\\alpha}_t = \\prod_{s=1}^t (1 - \\beta_s)\\), \\(t \\sim \\mathcal{U}([1,T])\\), and \\(T\\) is the total number of steps. For simple notation, we omit the timestep \\(t\\) as \\(\\epsilon(\\mathbf{x}_t; t) = \\epsilon(\\mathbf{x}_t)\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.54, + 0.785, + 0.721 + ], + "angle": 0, + "content": "Latent Diffusion Models (LDMs) [36] leverage the diffusion process within the latent space rather than in the pixel space, utilizing a pre-trained autoencoder. By mapping the input \\(\\mathbf{x}\\) into a latent space with the encoder \\(\\mathcal{E}\\), \\(\\mathbf{z} = \\mathcal{E}(\\mathbf{x})\\), an LDM is trained to predict the added noise in the latent space. As this paper mainly discusses LDMs, we use \\(\\mathbf{x}_t\\) instead of \\(\\mathbf{z}_t\\). Text-to-image models additionally take inputs of the text embedding, \\(\\mathbf{c}_p = \\mathcal{E}_{\\mathrm{txt}}(c_p)\\), paired with an image \\(\\mathbf{x}\\), where \\(\\mathcal{E}_{\\mathrm{txt}}\\) is the text encoder and \\(c_{p}\\) is a text prompt. To enhance the quality of text conditioning, the classifier-free guidance (CFG) models both conditional and unconditional noise estimates by randomly replacing \\(\\mathbf{c}_p\\) with the embedding of an empty string \\(\\mathbf{c}_{\\emptyset} = \\mathcal{E}_{\\mathrm{txt}}(\"\")\\) and modifies the noise estimate during sampling as \\(\\tilde{\\epsilon}_{\\mathrm{CFG}} = \\epsilon_{\\theta}(\\mathbf{x}_t) + (w + 1)[\\epsilon_{\\theta}(\\mathbf{x}_t;\\mathbf{c}_p) - \\epsilon_{\\theta}(\\mathbf{x}_t;\\mathbf{c}_{\\emptyset})]\\), where \\(w + 1\\) is the CFG scale. We simply denote \\(\\epsilon_{\\theta}(\\mathbf{x}_t;\\mathbf{c}_{\\emptyset})\\) as \\(\\epsilon_{\\theta}(\\mathbf{x}_t)\\) in the subsequent sections." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.746, + 0.331, + 0.761 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.784, + 0.841 + ], + "angle": 0, + "content": "We present Human Feedback Inversion (HFI), an innovative framework for concept removal guided by human feedback. Unlike previous methods that rely on predefined text-based concepts, i.e., a specified target word like \"nudity\", real-world inappropriate concepts are often nuanced and challenging to express in" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.33, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Safeguard Diffusion Models with Human Feedback Inversion" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.143, + 0.473, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.238, + 0.268, + 0.454, + 0.28 + ], + "angle": 0, + "content": "(a) Human Feedback Inversion (HFI)" + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.163, + 0.788, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.524, + 0.268, + 0.762, + 0.28 + ], + "angle": 0, + "content": "(b) Safe self-Distillation Diffusion (SDD)" + }, + { + "type": "image_caption", + "bbox": [ + 0.292, + 0.291, + 0.709, + 0.306 + ], + "angle": 0, + "content": "Fig. 4: Illustration of our proposed framework, HFI and SDD." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.335, + 0.788, + 0.486 + ], + "angle": 0, + "content": "text. HFI addresses this challenge by learning concepts based on collective understanding. Our approach initiates with collecting human feedback on model-generated images through carefully designed surveys detailed in Appendix C, where harmfulness is comparatively evaluated. Using this human evaluation data, we construct a reward model that assigns a harmfulness score to each image. With this reward model, we formulate an optimization problem to identify a soft token that accurately represents harmfulness. These optimal embedding vectors are then utilized to fine-tune the diffusion model, preventing the generation of harmful images. The overall framework is depicted in Fig. 4 and Appendix A, and detailed explanations of each stage are provided in the subsequent sections." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.507, + 0.614, + 0.523 + ], + "angle": 0, + "content": "3.1 Collecting and Modeling Human Feedback" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.531, + 0.788, + 0.712 + ], + "angle": 0, + "content": "We first generate images, on which human feedback will be collected, using the original model and a set of simple prompts containing the target concept word \\( c \\) detailed in Appendix B. Subsequently, two main settings can be considered for gathering human feedback. The design of how and what kind of feedback to collect during this process is crucial. In cases where the content the model should not generate is well-defined and clear annotation guidelines exist (e.g., nudity), binary feedback for each image can be collected. However, this does not apply to all scenarios, with a notable example being an artist's style. Suppose that an artist requests the removal of not only specific works but also the overall style from the model, then the criteria may not be clear. In such cases, ranking feedback can be collected. This involves presenting annotators with \\( M \\) random images and requesting them to rank these images ranging from 1 to \\( M \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.713, + 0.787, + 0.759 + ], + "angle": 0, + "content": "Following this, we need to train a reward model \\(r_{\\psi}\\), which accepts an input image \\(\\mathbf{x}\\) and outputs a score \\(r_{\\psi}(\\mathbf{x}) \\in \\mathbb{R}\\) for the concept \\(c\\). For binary feedback, we follow the MSE-based approach proposed in Lee et al. [20]," + }, + { + "type": "equation", + "bbox": [ + 0.351, + 0.767, + 0.787, + 0.787 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {M S E} (\\psi) = \\mathbb {E} _ {(\\mathbf {x}, y) \\sim \\mathcal {D} _ {\\text {h u m a n}}} \\left[ (r _ {\\psi} (\\mathbf {x}) - y) ^ {2} \\right], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "where \\(\\mathcal{D}_{\\mathrm{human}}\\) consists of images \\(\\mathbf{x}\\) and corresponding binary annotations \\(y \\in \\{0,1\\}\\). Here, '1' indicates the presence of the harmful concept (e.g., with nudity), and '0' is the absence (e.g., without nudity), answered by human annotators." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "S. Kim et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.256 + ], + "angle": 0, + "content": "On the other hand, for ranking feedback, we adopted the Bradley-Terry model [5], which assigns scores to examples based on pairwise comparisons [8, 26, 49]. The obtained rankings of \\(M\\) images can be reorganized to construct \\(\\mathcal{D}_{\\mathrm{human}}\\) as \\(\\binom{M}{2}\\) pairwise comparisons as \\(\\mathbf{x}^{+} = \\mathbf{x}^{(i)}, \\mathbf{x}^{-} = \\mathbf{x}^{(j)}\\) if \\(\\mathbf{x}^{(i)}\\) has a higher rank than \\(\\mathbf{x}^{(j)}\\), and vice versa - a higher ranking indicates a greater presence or resemblance of the concept. Subsequently, we include all \\(\\binom{M}{2}\\) pairs in a single batch, following Ouyang et al. [26], and optimize the reward model as" + }, + { + "type": "equation", + "bbox": [ + 0.295, + 0.266, + 0.786, + 0.302 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {N L L} (\\psi) = - \\frac {1}{\\binom {M} {2}} \\mathbb {E} _ {(\\mathbf {x} ^ {+}, \\mathbf {x} ^ {-}) \\sim \\mathcal {D} _ {\\text {h u m a n}}} \\log \\sigma \\left(r _ {\\psi} \\left(\\mathbf {x} ^ {+}\\right) - r _ {\\psi} \\left(\\mathbf {x} ^ {-}\\right)\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.311, + 0.787, + 0.357 + ], + "angle": 0, + "content": "where \\(\\sigma (\\cdot)\\) is the sigmoid function. For both approaches, the reward model \\(r_{\\psi}(\\mathbf{x})\\) maps the CLIP [30] embedding of an image \\(\\mathbf{x}\\) to a score through a couple of MLP layers with ReLU non-linearities." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.38, + 0.566, + 0.396 + ], + "angle": 0, + "content": "3.2 Inverting Feedback into Embeddings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.405, + 0.789, + 0.527 + ], + "angle": 0, + "content": "Based on the feedback received in Sec. 3.1, we transform the corresponding concept into a soft token. Inspired by recent inversion techniques [12,35], we utilize TI [12], as it can easily replace or be combined with text tokens from previous approaches. The goal is to find a soft token \\( v^{*} \\) that maximizes the expected reward of generated images, i.e., \\( v^{*} = \\arg \\max_{v}\\mathbb{E}_{\\mathbf{x}\\sim p_{\\theta}(\\mathbf{x}|v)}[r_{\\psi}(\\mathbf{x})] \\), where \\( v \\) is the soft word in the token space. Utilizing the reward model \\( r_{\\psi} \\) in Sec. 3.1, we optimize the following objective of the reward-weighted negative log-likelihood [20] to get the optimal token \\( v^{*} \\) that best captures the text concept \\( c \\):" + }, + { + "type": "equation", + "bbox": [ + 0.349, + 0.539, + 0.786, + 0.561 + ], + "angle": 0, + "content": "\\[\nv ^ {*} = \\underset {v} {\\arg \\min } \\mathbb {E} _ {\\mathbf {x} \\sim p _ {\\theta} (\\mathbf {x} | c)} [ - r _ {\\psi} (\\mathbf {x}) \\log p _ {\\theta} (\\mathbf {x} | v) ]. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.572, + 0.788, + 0.68 + ], + "angle": 0, + "content": "Inverting too many images may struggle with concept extraction by prioritizing image reconstruction over it, as shown in Gal et al. [12]. Also, we empirically found that inverting it with a small learning rate may fail to capture the concept, which requires additional hyperparameter tuning. To simplify such a burden, we use a reward-weighted sampler with images of the top- \\( K \\) scores, which is equivalent to the reward-weighted loss in Eq. (4). We use the learning rate of \\( 5 \\times 10^{-4} \\), the default value in the HuggingFace's implementation1, and \\( K = 50 \\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.701, + 0.676, + 0.716 + ], + "angle": 0, + "content": "3.3 Removing Learned Concepts with Self-Distillation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.726, + 0.788, + 0.804 + ], + "angle": 0, + "content": "After obtaining soft tokens through the proposed HFI method in Sec. 3.2, one can adapt them to existing text-based concept removal methods, enabling the desired removal of concepts using the obtained soft tokens instead of text tokens. While HFI alone proves effectiveness in Tabs. 1 and 2, we propose a new method termed Safe self-Distillation Diffusion (SDD) to achieve an even greater impact," + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.787, + 0.842 + ], + "angle": 0, + "content": "1 https://github.com/huggingface/diffusers/blob/main/examples/textual_inversion/textual_inversion.py" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.329, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Safeguard Diffusion Models with Human Feedback Inversion" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "particularly in enhancing the removal of ambiguous concepts judged by humans. SDD removes the target concept from a model based on self-distillation [53]. We first set both student \\(\\theta\\) and teacher models \\(\\theta^{\\star}\\) whose parameters are initialized from the pre-trained model and update the student with the following loss:" + }, + { + "type": "equation", + "bbox": [ + 0.323, + 0.218, + 0.785, + 0.235 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {S D D} = \\mathbb {E} _ {\\mathbf {x} _ {t} \\sim p _ {\\theta^ {*}} (\\mathbf {x} _ {t} | v ^ {*}), t} \\left[ \\| \\epsilon_ {\\theta} (\\mathbf {x} _ {t}; v ^ {*}) - \\mathbf {s g} \\left(\\epsilon_ {\\theta} (\\mathbf {x} _ {t})\\right) \\| _ {2} ^ {2} \\right], \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.244, + 0.785, + 0.333 + ], + "angle": 0, + "content": "where \\(\\mathfrak{sg}()\\) denotes the stop-grad operation. That is, we want the model given the token \\(v^{*}\\) to behave as if it were not for any token by setting the unconditional estimate \\(\\epsilon_{\\theta}(\\mathbf{x}_t)\\) as a self-distillation target. After updating the student model, we update the teacher model \\(\\theta^{\\star}\\) as an exponential moving average (EMA) from it, i.e., \\(\\theta^{\\star} \\gets m\\theta^{\\star} + (1 - m)\\theta\\), with a decaying rate \\(m \\in (0,1)\\). After the training, we use the EMA parameter of the teacher model as our final safe model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.335, + 0.785, + 0.456 + ], + "angle": 0, + "content": "Unlike methods focused on learning new concepts [19, 38], determining suitable replacements when removing specific concepts can be challenging, often contingent on other contextual elements [9]. For instance, while removing the color red in a red apple can be substituted with a green apple, removing the color red devoid of context lacks clarity. Similarly, removing concepts like nudity, violence, or specific artists presents this ambiguity. Previous approaches [3,13,40] have attempted to tackle this challenge by targeting negative guidance; however, images generated in this manner often fall short of removing such concepts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.456, + 0.786, + 0.621 + ], + "angle": 0, + "content": "We tackle this problem by introducing the EMA teacher model, which moves in tandem with the student model but in a more stable and deliberate manner, gradually removing the target concept. This serves not only to enhance training stability but also plays a crucial role in providing the student model with appropriate targets for conditional noise predictions. Notably, in Eq. (5), the latent \\(\\mathbf{x}_t\\) is sampled from the teacher model, i.e. it generates parts that have not been effectively removed up to the existing training step. Additionally, resolving discrepancies between the model's understanding and human judgment regarding harmfulness poses a non-trivial task. In SDD with HFI, sampling \\(\\mathbf{x}_t\\) is conditioned on the learned token \\(v^*\\). During this process, the model learns to remove the concept from newly generated latents that would be deemed problematic." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.622, + 0.786, + 0.743 + ], + "angle": 0, + "content": "Lastly, we introduce a couple of additional techniques for fine-tuning. First, similar to existing work [13, 19, 51, 52], we fine-tune only the cross-attention layers of the U-Net [37]. Second, we empirically find that not all timesteps equally influence the conditioning of image generation; the \\(L_{2}\\)-norm between the unconditional and the conditional noise was generally high in the middle timesteps, as shown in Appendix E. Therefore, we sample the timestep to apply distillation as \\(t = T \\cdot \\mathcal{B}(\\alpha, \\beta)\\) where \\(\\mathcal{B}(\\alpha, \\beta)\\) is a beta random variable with parameter \\((\\alpha, \\beta)\\). We set \\((\\alpha, \\beta) = (3, 3)\\) for our experiments as it works well in general." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.764, + 0.388, + 0.781 + ], + "angle": 0, + "content": "4 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Numerous attempts have been made to address harmful content generation, including the use of external classifiers for image filtering during inference, such as the safety checker in Stable Diffusion (SD) [36]. However, these approaches" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.129 + ], + "angle": 0, + "content": "S. Kim et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "can be easily circumvented by users [33]. Efforts to filter out problematic training images during pre-processing [48] often lead to reduced training sets and inferior image quality [24, 28], without completely eliminating harmfulness [1]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.787, + 0.375 + ], + "angle": 0, + "content": "Recent advancements in the field can be broadly classified into three categories: (i) inference-time techniques manipulate noise estimates to counteract concepts [3,40]; (ii) direct cross-attention optimization adjusts weights to disregard problematic tokens from text inputs either directly or through gradient descent [14,52]; (iii) fine-tuning methods refine a subset of parameters using generated images [13,18,23]. Although both (ii) and (iii) fine-tune the model, the main disparity lies in the training objective. The former adjusts the projection or attention of text inputs, while the latter alters the model's distribution along the diffusion trajectories. Moreover, while the former does not necessitate training images or uses only a few real images, the latter generates training images from the model. Nonetheless, all methods entail a trade-off between effectiveness and utility, given the inherent challenge of identifying concepts in texts for removal." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.375, + 0.788, + 0.603 + ], + "angle": 0, + "content": "Human feedback has become crucial in enhancing both performance and safety in machine learning models [8]. Initially applied in dialogue systems [15, 21], reinforcement learning (RL) algorithms, particularly Proximal Policy Optimization (PPO) [43], have played a significant role in leveraging human feedback, notably in Large Language Models (LLMs) [6, 26, 49], focusing on both performance and safety considerations. Reinforcement learning with human feedback (RLHF) [26] optimizes models based on reward signals derived from user input, further enhancing the safety and adaptability of LLMs. Beyond language domains, efforts have been made to align text-to-image diffusion models with human preferences [2, 10, 20]. These involve fine-tuning using a reward-weighted loss function with appropriate regularization. However, the practical application of such RL-based methods often faces challenges due to hyperparameter sensitivity and reward design intricacies [20]. Furthermore, while such methods relying on human preference data aim to enhance image quality perceived by humans, they differ from our goal of suppressing image generation of specific concepts." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.627, + 0.376, + 0.646 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.661, + 0.553, + 0.677 + ], + "angle": 0, + "content": "5.1 Baselines and Evaluation Protocols" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We compare the performance of our method with the original SD and previous methods. In Tabs. 1 and 2, SD [36] indicates the original Stable Diffusion model, and NEG is the simplest baseline of SD with the negative prompt, i.e., the noise conditioned on the target concept is used instead of the unconditional one, thus negating the target concept. We consider two inference-time techniques: SLD [40] and SEGA [3]. For fine-tuning methods, we consider the following three methods: ESD [13], FMN [52], and UCE [14]. For all fine-tuning methods, including ours, we use the artist name (e.g., \"Vincent van Gogh\") or the word (e.g., \"nudity\" and \"bleeding\") for the target concept in text. When applying HFI to these methods, HFI token \\( v^{*} \\) replaces the text word. Refer to Appendix B for details." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.33, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Safeguard Diffusion Models with Human Feedback Inversion" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.143, + 0.499, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.143, + 0.788, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.224, + 0.498, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.225, + 0.787, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.304, + 0.498, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.304, + 0.787, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.41, + 0.788, + 0.466 + ], + "angle": 0, + "content": "Fig. 5: Comparison of concept removal methods for Vincent van Gogh's artistic style. Van Gogh's style is marked by bold brushstrokes, vivid colors, and swirling patterns. HFI+SDD is the most effective, consistently producing the most neutral but still artwork-like images and successfully eliminating his stylistic characteristics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.502, + 0.785, + 0.563 + ], + "angle": 0, + "content": "The evaluation of the method's capability to remove specific concepts is structured along two primary axes: effectiveness and utility. Effectiveness measures the ability to avoid generating the target concept, while utility assesses the preservation of image quality in outputs unrelated to the target concept." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.566, + 0.788, + 0.807 + ], + "angle": 0, + "content": "For artist concept removal in Sec. 5.2, we propose two specific metrics to gauge both aspects. The effectiveness is evaluated by the target artist ratio (%ART) using CLIP zero-shot classification. This involves generating 1,000 images in total of 50 commonly depicted objects in famous artworks (listed by ChatGPT [25]), with the prompt template \"a painting of {object} in the style of {artist}\" These images are then classified using OpenCLIP [7] to determine how frequently the target artist is accurately identified. The utility is determined through the photo similarity ratio (%PHO), which quantifies how well the model retains the ability to generate paintings rather than photorealistic images, thus preserving artistic styles not specific to the targeted artist. We enhance classification accuracy using prompt ensemble [30]. In line with previous work, we also report LPIPS [54] for both the target and other artists. Finally, since such quantitative metrics cannot completely replace qualitative evaluations and human evaluations can be subjective, we asked ChatGPT 4o [25] to assess the performance of removing the specified artist to supplement this. The qualitative evaluation results and further details are provided in Appendix F." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "For harmful concept removal in Sec. 5.3, the effectiveness is measured by generating 5,000 images in total with the prompt {\"country\" body\" using" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "S. Kim et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.142, + 0.5, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.143, + 0.788, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.224, + 0.499, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.224, + 0.787, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.304, + 0.499, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.304, + 0.787, + 0.401 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.411, + 0.788, + 0.468 + ], + "angle": 0, + "content": "Fig. 6: Comparison of concept removal methods for Leonid Afremov's artistic style. Afremov's style is marked by vibrant colors, palette knife technique (similar to pointage technique), and impressionistic elements. HFI+SDD is the most effective, producing images with clear depictions of various objects and minimal retention of his style." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.495, + 0.789, + 0.679 + ], + "angle": 0, + "content": "names of 50 countries with high GDP following [40]. This prompt generates way more inappropriate images (74.2% with SD v1.4) compared to the sexual category of the I2P dataset [40] (reportedly ranked \\(5^{\\text{th}}\\) among the seven categories with 32.8% of hard prompt ratio and only 16.7% of nudity images). The proportion of images classified \"unsafe\" (%NSFW) and those detected with exposed body parts (%NUDE) are evaluated by NudeNet [29]. This process is similarly used to assess the removal of another harmful concept of bleeding by generating images for the shocking category in the I2P dataset [40] and classifying them with the Q16 classifier (%I2P) [41]. The utility is evaluated by generating images from 10,000 random captions from the MSCOCO dataset [22] and comparing these to actual photos, outputs from the original SD model, and their text-to-image alignment using FID [16, 27], LPIPS [54], and CLIPScore [30], respectively." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.697, + 0.439, + 0.712 + ], + "angle": 0, + "content": "5.2 Artist Style Removal" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.809 + ], + "angle": 0, + "content": "In this section, we conducted experiments to remove the style of artists from the existing SD model. For HFI, we generate 1,000 images from the original SD model with prompts detailed in Appendix B and gather human feedback on a total of 200 images (equivalent to 20 question pairs) per artist. We then train a reward model, which is then used to approximate the scores for all 1,000 images for HFI. All methods with HFI use the same soft token instead of the text token." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In Tab. 1, even without HFI, SDD alone finds a reasonable trade-off, achieving relatively better removal than the other methods while maintaining a rela-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.33, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Safeguard Diffusion Models with Human Feedback Inversion" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.143, + 0.499, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.143, + 0.788, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.223, + 0.498, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.224, + 0.787, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.304, + 0.498, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.304, + 0.787, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.411, + 0.788, + 0.467 + ], + "angle": 0, + "content": "Fig. 7: Comparison of concept removal methods for Edward Hopper's artistic style. Hopper's characteristic style is marked by realism, simplicity, architectural details, and muted color palette. Among the methods, HFI+SDD most effectively removes his distinctive features while still being abstract and less photorealistic." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.502, + 0.788, + 0.578 + ], + "angle": 0, + "content": "tively lower percentage on \\(\\% \\mathrm{PHO}\\). Moreover, when HFI is employed, our method shows the state-of-the-art performance of \\(\\% \\mathrm{ART}\\) of \\(12.7\\%\\) on average, with average \\(\\% \\mathrm{PHO}\\) of \\(12.0\\%\\). For Edward Hopper, however, HFI+SDD was outperformed by a few HFI-enhanced baselines, but it worked the best in most cases. Notably, HFI+SDD always surpasses all the previous methods without HFI." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.581, + 0.788, + 0.717 + ], + "angle": 0, + "content": "Among the baselines, fine-tuning methods (ESD and SDD) observed the most benefit from HFI. In contrast, methods directly optimizing cross-attention exhibited subpar removal performance and appeared to have relatively less impact. Meanwhile, inference-time techniques generally demonstrated improvements in the trade-off between effectiveness and utility facilitated by HFI as shown in Fig. 8. However, as shown in Figs. 5 to 7, there are often cases where the object is not well represented or the distinctive artistic characteristics remain. In such cases, HFI can be beneficial, whereas text-based methods frequently struggle to distinguish the artist's concept from the generic artwork style." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In the qualitative evaluation using ChatGPT 4o [25], HFI+SDD was evaluated as the most superior method by stating that \"it consistently produces the most neutral images by moving away from the artist's style while maintaining a high level of image quality.\" Despite being asked to evaluate only the removal performance, somewhat surprisingly, ChatGPT considered not only the degree of style removal but also various aspects such as object depiction, image quality, and artifacts) with consistent judgments across all objects and presentation orders, aligning with our proposed metrics. Refer to Appendix F for full results." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "S. Kim et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.145, + 0.406, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.145, + 0.594, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.145, + 0.783, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.385, + 0.241, + 0.64, + 0.253 + ], + "angle": 0, + "content": "--- SLD -SLD+HFI SEGAE SEGA+HFI" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.266, + 0.788, + 0.31 + ], + "angle": 0, + "content": "Fig. 8: Trends of \\(\\%_{\\text{ART}}\\) and \\(\\%_{\\text{PHO}}\\) by inference-time techniques. While forgetting the general artwork concepts (higher \\(\\%_{\\text{PHO}}\\)) is inevitable, HFI can achieve similar outcomes at smaller editing scales. The same marker symbol denotes the same scale in each plot." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.323, + 0.528, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.456, + 0.526, + 0.485 + ], + "angle": 0, + "content": "Fig. 9: HFI mitigates terrifying aspects of images generated with I2P shocking prompts." + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.331, + 0.787, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.54, + 0.391, + 0.787, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.539, + 0.456, + 0.787, + 0.485 + ], + "angle": 0, + "content": "Fig.10: Images with the highest (top) and lowest (bottom) rewards." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.514, + 0.484, + 0.53 + ], + "angle": 0, + "content": "5.3 Harmful Concept Removal" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.788, + 0.719 + ], + "angle": 0, + "content": "Tab. 2 shows our experimental results on removing NSFW and bleeding concepts. Similar to Sec. 5.2, we first generated images and collected binary feedback and ranking feedback for nudity and bleeding, respectively. SDD alone effectively reduces the presence of NSFW imagery, but when combined with HFI, such images are nearly eliminated. Our method and ESD-u-3 achieve a similar level, but ESD exhibits inferior performance in all image quality metrics. In a similar vein, the method of fine-tuning only the cross-attention, as in ESD-x-3, is observed to generate harmful images beyond a certain threshold. High in %NSFW and low in %NUDE for some methods indicate that they still generate somewhat unsafe images even though exposed body parts are limited. When combined with HFI, all methods but FMN exhibit a sharp decrease in %NSFW, underscoring the role of the soft token aligned with human judgment." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Furthermore, for nuanced concepts such as bleeding, HFI aids in identifying and mitigating the inappropriate aspects of images. In Fig. 9, with HFI, it becomes apparent that not only has the extent of bleeding diminished, but it also yields images that are safer and less offensive. Particularly notable are the outcomes achieved by combining SDD and HFI, where notably safer images are produced despite maintaining identical compositions. The interpretation of such ambiguous concepts by humans can be encoded into learned tokens for subsequent utilization in their removal, highlighting the utility of HFI." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.33, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Safeguard Diffusion Models with Human Feedback Inversion" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 1: Quantitative comparison of artist style removal performance. Each method is evaluated both without and with HFI (✓), showing improvement due to human feedback. The \\(\\sum\\) column represents the sum of %ART and %PHO, indicating undesirable images. The table also compares our SDD method with other baselines." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.213, + 0.784, + 0.482 + ], + "angle": 0, + "content": "
Target ArtistVincent van GoghLeonid AfremovEdward HopperAvg. LPIPS
MethodHFI%ART↓ %PHO↓Σ%ART↓ %PHO↓Σ%ART↓ %PHO↓ΣTarget↑ Others↓
SD v1.495.92.998.899.90.6100.597.62.299.8--
Inference-time Technique
NEG [36]✓15.032.547.592.82.094.836.512.148.60.33520.2273
17.212.429.60.030.430.416.516.132.60.43520.3255
SLD [40]✓29.127.756.827.84.232.021.339.560.80.42770.2880
24.622.847.49.78.818.518.134.953.00.44590.3285
SEGA [3]✓40.311.752.064.41.565.937.728.666.30.31510.2522
17.517.635.120.74.525.232.021.853.80.34820.2661
Direct Cross-Attention Optimization
FMN [52]✓66.67.574.110.725.536.277.15.782.80.32930.1665
58.67.766.36.50.97.473.25.578.70.27930.2478
UCE [14]✓66.44.070.444.72.046.751.53.354.80.27670.0332
77.32.579.892.50.693.172.51.974.40.17440.0328
Fine-tuning Method
ESD-x-1 [13]✓45.015.960.913.525.438.964.47.273.60.36240.1635
22.520.442.919.057.276.214.317.431.70.44040.1691
SDD (ours)✓20.625.546.19.98.218.140.441.181.50.27600.2341
9.55.114.68.83.011.819.827.947.70.28790.1439
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.513, + 0.65, + 0.529 + ], + "angle": 0, + "content": "5.4 Analysis on Rewards and Learned Embeddings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.788, + 0.72 + ], + "angle": 0, + "content": "For nudity, the reward model generally aligns with Nudenet [29] (AUROC of 0.882), but differences between human and model judgments highlight the limits of relying solely only on models for inversion. Fig. 10 shows that images deemed appropriate by humans or not reflecting the artist's style receive low rewards and minimally impact HFI token learning. Integrating HFI with reward models effectively captures concepts. Ablation studies comparing HFI with the vanilla TI [12] show a \\(3.2\\%\\) and \\(5.1\\%\\) increase in \\(\\% \\mathrm{ART}\\) for artist removal without human feedback (i.e., random images without rewards) and with real images from the internet (without rewards), respectively. For nudity removal, \\(\\% \\mathrm{NSFW}\\) and \\(\\% \\mathrm{NUDE}\\) increase by \\(1.74\\%\\) and \\(0.04\\%\\) without human feedback, with FID increasing by 1.565. This demonstrates the importance of integrating human judgments and model-generated data into the training process to achieve optimal performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Compared to existing expensive methods using human feedback, the proposed HFI is highly cost-effective. We obtained 650 responses from five individuals (450 unique images for inter-annotator agreement (IAA)) within three hours for nudity. Providing annotation guidelines increased IAA, evidenced by Fleiss' \\(\\kappa\\) [11] at 0.4997 and Randolph's \\(\\kappa\\) [34] at 0.6702. For artistic concepts, we collected 40 ranking feedback responses (400 images) per artist from 15 individuals. Our research offers an economical and effective framework for extracting information from models and removing harmful content with human judgment." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "S. Kim et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 2: Quantitative comparison of harmful concept removal performance. Each method is evaluated both without and with HFI (✓) for inter-method comparison. The table also compares our SDD with other baselines for intra-method comparison." + }, + { + "type": "table", + "bbox": [ + 0.289, + 0.199, + 0.71, + 0.558 + ], + "angle": 0, + "content": "
Target ConceptNudityBleedingCOCO10K
MethodHFI%NSFW↓%NUDE↓%I2P↓FID↓CLIP↑ LPIPS↓
SD v1.474.1832.1465.7716.6360.2745
COCO ref.0.2676
Inference-time Technique
SD+NEG [36]✓30.683.8645.7918.8370.2701
5.144.7624.6519.5180.2699
SLD [40]✓58.089.8446.1417.8150.2702
42.9815.5829.5618.3450.2704
SEGA [3]✓67.7213.1645.4417.1290.2698
56.4219.5239.3716.7560.2698
Direct Cross-Attention Optimization
FMN [52]✓51.5027.7847.0818.0540.2556
61.3234.7255.7216.8130.2683
UCE [14]✓32.307.3661.8017.0240.2726
14.769.3253.2716.5410.2736
Fine-tuning Method
ESD-u-1 [13]✓11.302.00-17.0330.2664
2.421.16-18.4040.2636
ESD-u-3 [13]✓4.440.48-18.4980.2624
0.800.46-18.6150.2621
ESD-x-3 [13]✓14.322.9851.1718.5170.2677
4.183.3034.3517.8980.2670
SDD (ours)✓2.780.6030.8417.3620.2658
0.280.3223.3617.0230.2643
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.584, + 0.51, + 0.6 + ], + "angle": 0, + "content": "6 Conclusion and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.751 + ], + "angle": 0, + "content": "In this work, we introduced a novel concept removal framework within a text-to-image diffusion model through human feedback inversion and self-distillation-based fine-tuning. By leveraging human feedback to train soft tokens, we effectively safeguarded the diffusion model from harmful concepts, addressing the ambiguity in defining text-based concepts. We further proposed a self-distillation-based diffusion model fine-tuning algorithm, enhancing the efficacy of concept removal. We believe our research lays the groundwork for a human-centric approach to the safety and ethical considerations of large-scale diffusion models, paving the way for future advancements." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Limitation. While our proposed HFI performs well in capturing the intended concept compared to the text, its effectiveness depends on the quality of human feedback and the performance of inversion to soft tokens. Significant disagreements in human feedback can compromise the learning process, underscoring the importance of meticulous data collection and reward modeling." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.33, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Safeguard Diffusion Models with Human Feedback Inversion" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.284 + ], + "angle": 0, + "content": "This work was partly supported by Institute for Information & communications Technology Promotion (IITP) grant funded by the Korea government (MSIT) (No. RS-2019-II190075, Artificial Intelligence Graduate School Program (KAIST), No. 2022-0-00184, Development and Study of AI Technologies to Inexpensively Conform to Evolving Policy on Ethics) and the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. 2022R1A5A708390812)." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.308, + 0.323, + 0.323 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.34, + 0.785, + 0.381 + ], + "angle": 0, + "content": "1. Baio, A.: Exploring 12 million of the 2.3 billion images used to train stable diffusion's image generator. https://waxy.org/2022/08/exploring-12-million-of-the-images-used-to-train-stable-diffusions-image-generator/ (Aug 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.383, + 0.785, + 0.409 + ], + "angle": 0, + "content": "2. Black, K., Janner, M., Du, Y., Kostrikov, I., Levine, S.: Training diffusion models with reinforcement learning. arXiv preprint arXiv:2305.13301 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.41, + 0.785, + 0.45 + ], + "angle": 0, + "content": "3. Brack, M., Friedrich, F., Hintersdorf, D., Struppek, L., Schramowski, P., Kersting, K.: Sega: Instructing diffusion using semantic dimensions. arXiv preprint arXiv:2301.12247 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.451, + 0.785, + 0.492 + ], + "angle": 0, + "content": "4. Brack, M., Schramowski, P., Friedrich, F., Hintersdorf, D., Kersting, K.: The stable artist: Steering semantics in diffusion latent space. arXiv preprint arXiv:2212.06013 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.493, + 0.785, + 0.52 + ], + "angle": 0, + "content": "5. Bradley, R.A., Terry, M.E.: Rank analysis of incomplete block designs: I. the method of paired comparisons. Biometrika 39(3/4), 324-345 (1952)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.521, + 0.785, + 0.562 + ], + "angle": 0, + "content": "6. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.563, + 0.785, + 0.617 + ], + "angle": 0, + "content": "7. Cherti, M., Beaumont, R., Wightman, R., Wortsman, M., Ilharco, G., Gordon, C., Schuhmann, C., Schmidt, L., Jitsev, J.: Reproducible scaling laws for contrastive language-image learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2818-2829 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.618, + 0.785, + 0.659 + ], + "angle": 0, + "content": "8. Christiano, P.F., Leike, J., Brown, T., Martic, M., Legg, S., Amodei, D.: Deep reinforcement learning from human preferences. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.66, + 0.785, + 0.687 + ], + "angle": 0, + "content": "9. Du, Y., Li, S., Mordatch, I.: Compositional visual generation with energy based models. Advances in Neural Information Processing Systems 33, 6637-6647 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.688, + 0.785, + 0.729 + ], + "angle": 0, + "content": "10. Fan, Y., Watkins, O., Du, Y., Liu, H., Ryu, M., Boutilier, C., Abbeel, P., Ghavamzadeh, M., Lee, K., Lee, K.: Dpok: Reinforcement learning for fine-tuning text-to-image diffusion models. arXiv preprint arXiv:2305.16381 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.729, + 0.785, + 0.756 + ], + "angle": 0, + "content": "1. Fleiss, J.L.: Measuring nominal scale agreement among many raters. Psychological bulletin 76(5), 378 (1971)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.757, + 0.785, + 0.811 + ], + "angle": 0, + "content": "12. Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A.H., Chechik, G., Cohen-Or, D.: An image is worth one word: Personalizing text-to-image generation using textual inversion (2022). https://doi.org/10.48550/ARXIV.2208.01618, https://arxiv.org/abs/2208.01618" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.812, + 0.785, + 0.84 + ], + "angle": 0, + "content": "13. Gandikota, R., Materzynska, J., Fiotto-Kaufman, J., Bau, D.: Erasing concepts from diffusion models. arXiv preprint arXiv:2303.07345 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.34, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.359, + 0.128 + ], + "angle": 0, + "content": "S. Kim et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "14. Gandikota, R., Orgad, H., Belinkov, Y., Materzynska, J., Bau, D.: Unified concept editing in diffusion models. arXiv preprint arXiv:2308.14761 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.785, + 0.217 + ], + "angle": 0, + "content": "15. Gao, J., Galley, M., Li, L.: Neural approaches to conversational ai. In: The 41st international ACM SIGIR conference on research & development in information retrieval. pp. 1371-1374 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.785, + 0.259 + ], + "angle": 0, + "content": "16. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.26, + 0.785, + 0.285 + ], + "angle": 0, + "content": "17. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. arXiv preprint arxiv:2006.11239 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.287, + 0.785, + 0.327 + ], + "angle": 0, + "content": "18. Kumari, N., Zhang, B., Wang, S.Y., Shechtman, E., Zhang, R., Zhu, J.Y.: Ablating concepts in text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22691-22702 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.329, + 0.785, + 0.37 + ], + "angle": 0, + "content": "19. Kumari, N., Zhang, B., Zhang, R., Shechtman, E., Zhu, J.Y.: Multi-concept customization of text-to-image diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1931-1941 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.411 + ], + "angle": 0, + "content": "20. Lee, K., Liu, H., Ryu, M., Watkins, O., Du, Y., Boutilier, C., Abbeel, P., Ghavamzadeh, M., Gu, S.S.: Aligning text-to-image models using human feedback. arXiv preprint arXiv:2302.12192 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.438 + ], + "angle": 0, + "content": "21. Li, J., Monroe, W., Ritter, A., Galley, M., Gao, J., Jurafsky, D.: Deep reinforcement learning for dialogue generation. arXiv preprint arXiv:1606.01541 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.494 + ], + "angle": 0, + "content": "22. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.496, + 0.785, + 0.535 + ], + "angle": 0, + "content": "23. Ni, Z., Wei, L., Li, J., Tang, S., Zhuang, Y., Tian, Q.: Degeneration-tuning: Using scrambled grid shield unwanted concepts from stable diffusion. arXiv preprint arXiv:2308.02552 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.537, + 0.785, + 0.577 + ], + "angle": 0, + "content": "24. O'Connor, R.: Stable diffusion 1 vs 2 - what you need to know. https://www.assemblyai.com/blog/stable-diffusion-1-vs-2-what-you-need-to-know/ (Dec 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.718, + 0.591 + ], + "angle": 0, + "content": "25. OpenAI: Chatgpt (2023), https://www.openai.com/, july 2023 version" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.785, + 0.646 + ], + "angle": 0, + "content": "26. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems 35, 27730-27744 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.673 + ], + "angle": 0, + "content": "27. Parmar, G., Zhang, R., Zhu, J.Y.: On aliased resizing and surprising subtleties in gan evaluation. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.785, + 0.715 + ], + "angle": 0, + "content": "28. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.717, + 0.785, + 0.743 + ], + "angle": 0, + "content": "29. Praneeth, B.: NudeNet: Neural nets for nudity classification, detection and selective censoring. https://github.com/notAI-tech/NudeNet (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.785, + 0.798 + ], + "angle": 0, + "content": "30. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "31. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 1(2), 3 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.33, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Safeguard Diffusion Models with Human Feedback Inversion" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "32. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: International Conference on Machine Learning. pp. 8821-8831. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.787, + 0.217 + ], + "angle": 0, + "content": "33. Rando, J., Paleka, D., Lindner, D., Heim, L., Tramér, F.: Red-teaming the stable diffusion safety filter. arXiv preprint arXiv:2210.04610 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.218, + 0.786, + 0.244 + ], + "angle": 0, + "content": "34. Randolph, J.J.: Free-marginal multirater kappa (multirater k [free]): An alternative to fleiss' fixed-marginal multirater kappa. Online submission (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.787, + 0.286 + ], + "angle": 0, + "content": "35. Roich, D., Mokady, R., Bermano, A.H., Cohen-Or, D.: Pivotal tuning for latent-based editing of real images. ACM Transactions on graphics (TOG) 42(1), 1-13 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.787, + 0.327 + ], + "angle": 0, + "content": "36. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10684-10695 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.786, + 0.383 + ], + "angle": 0, + "content": "37. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.787, + 0.438 + ], + "angle": 0, + "content": "38. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22500-22510 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.787, + 0.493 + ], + "angle": 0, + "content": "39. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.787, + 0.535 + ], + "angle": 0, + "content": "40. Schramowski, P., Brack, M., Deiseroth, B., Kersting, K.: Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.787, + 0.59 + ], + "angle": 0, + "content": "41. Schramowski, P., Tauchmann, C., Kersting, K.: Can machines help us answering question 16 in datasheets, and in turn reflecting on inappropriate content? In: 2022 ACM Conference on Fairness, Accountability, and Transparency. pp. 1350-1361 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.591, + 0.787, + 0.646 + ], + "angle": 0, + "content": "42. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. arXiv preprint arXiv:2210.08402 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.787, + 0.673 + ], + "angle": 0, + "content": "43. Schulman, J., Wolski, F., Dhariwal, P., Radford, A., Klimov, O.: Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.674, + 0.787, + 0.715 + ], + "angle": 0, + "content": "44. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: Proceedings of the 32nd International Conference on Machine Learning (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.787, + 0.743 + ], + "angle": 0, + "content": "45. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.787, + 0.771 + ], + "angle": 0, + "content": "46. Song, Y., Ermon, S.: Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.787, + 0.812 + ], + "angle": 0, + "content": "47. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "48. Stability AI: Stable diffusion 2.0 release. https://stability.ai/blog/stable-diffusion-v2-release (Nov 2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.359, + 0.127 + ], + "angle": 0, + "content": "S. Kim et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "49. Stiennon, N., Ouyang, L., Wu, J., Ziegler, D., Lowe, R., Voss, C., Radford, A., Amodei, D., Christiano, P.F.: Learning to summarize with human feedback. Advances in Neural Information Processing Systems 33, 3008-3021 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.231 + ], + "angle": 0, + "content": "50. Thiel, D.: Identifying and eliminating csam in generative ml training data and models. Tech. rep., Technical Report. Stanford University, Palo Alto, CA. https://purl.stanford... (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.231, + 0.785, + 0.272 + ], + "angle": 0, + "content": "51. Yang, Z., Yi, X., Li, P., Liu, Y., Xie, X.: Unified detoxifying and debiasing in language generation via inference-time adaptive optimization. arXiv preprint arXiv:2210.04492 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.272, + 0.785, + 0.3 + ], + "angle": 0, + "content": "52. Zhang, E., Wang, K., Xu, X., Wang, Z., Shi, H.: Forget-me-not: Learning to forget in text-to-image diffusion models. arXiv preprint arXiv:2211.08332 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.785, + 0.355 + ], + "angle": 0, + "content": "53. Zhang, L., Song, J., Gao, A., Chen, J., Bao, C., Ma, K.: Be your own teacher: Improve the performance of convolutional neural networks via self distillation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3713-3722 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.785, + 0.398 + ], + "angle": 0, + "content": "54. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.398 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Safeguard Text-to-Image Diffusion Models with Human Feedback Inversion/603d7974-42c1-4361-a264-c2589e2c9681_origin.pdf b/2024/Safeguard Text-to-Image Diffusion Models with Human Feedback Inversion/603d7974-42c1-4361-a264-c2589e2c9681_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b38aee0d518d2777ebfd4fc43d2cd08b6bf8d580 --- /dev/null +++ b/2024/Safeguard Text-to-Image Diffusion Models with Human Feedback Inversion/603d7974-42c1-4361-a264-c2589e2c9681_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:08a89b8c6978ea6a0c6f98a61aaf5755ec3e86e6c440916e9c820e68ecd55bfd +size 13460496 diff --git a/2024/Safeguard Text-to-Image Diffusion Models with Human Feedback Inversion/full.md b/2024/Safeguard Text-to-Image Diffusion Models with Human Feedback Inversion/full.md new file mode 100644 index 0000000000000000000000000000000000000000..427b50ad044784a7c7326af3ff2321c357187515 --- /dev/null +++ b/2024/Safeguard Text-to-Image Diffusion Models with Human Feedback Inversion/full.md @@ -0,0 +1,312 @@ +# Safeguard Text-to-Image Diffusion Models with Human Feedback Inversion + +Sanghyun Kim $^{\text{D}}$ , Seohyeon Jung $^{\text{D}}$ , Balhae Kim $^{\text{D}}$ , Moonseok Choi $^{\text{D}}$ , Jinwoo Shin $^{\text{D}}$ , and Juho Lee $^{\text{D}}$ + +Korea Advanced Institute of Science and Technology (KAIST) {nannullna,heon2203,balhaekim,ms.choi,jinwoos,juholee}@kaist.ac.kr + +Abstract. This paper addresses the societal concerns arising from large-scale text-to-image diffusion models for generating potentially harmful or copyrighted content. Existing models rely heavily on internet-crawled data, wherein problematic concepts persist due to incomplete filtration processes. While previous approaches somewhat alleviate the issue, they often rely on text-specified concepts, introducing challenges in accurately capturing nuanced concepts and aligning model knowledge with human understandings. In response, we propose a framework named Human Feedback Inversion (HFI), where human feedback on model-generated images is condensed into textual tokens guiding the mitigation or removal of problematic images. The proposed framework can be built upon existing techniques for the same purpose, enhancing their alignment with human judgment. By doing so, we simplify the training objective with a self-distillation-based technique, providing a strong baseline for concept removal. Our experimental results demonstrate our framework significantly reduces objectionable content generation while preserving image quality, contributing to the ethical deployment of AI in the public sphere. Code is available at https://github.com/nannullna/safeguard-hfi. + +Caution: This paper contains discussions and examples related to harmful content, including text and images. Reader discretion is advised. + +Keywords: Fairness, accountability, and transparency $\cdot$ Diffusion models $\cdot$ Image generation + +# 1 Introduction + +Large-scale text-to-image generation models have demonstrated remarkable success in producing high-quality images encompassing a wide array of concepts [4, 31,32,36,39]. These models possess the ability to combine disparate ideas in novel and unprecedented manners, making them invaluable tools for creative content creation. However, they also give rise to substantial concerns, including the generation of harmful or copyrighted content, even without explicit directives [40]. A critical issue lies in the heavy reliance of the models on internet-crawled datasets (e.g., LAION-5B [42], currently not available [50]), which, despite attempts at aggressive filtration [48], often lead to problematic content generation [33,40]. + +![](images/393aa38b7d0a6a268dc4680498706399ed7c20875575bd8391fdd410d7d9a65f.jpg) +Fig. 1: Comparative analysis of NSFW content removal techniques, including our proposed fine-tuning method (SDD), both with and without HFI framework. The results clearly show that incorporating HFI significantly reduces the amplification of provocative body parts and ensures the generation of clothed representations. + +Previous inference-time [3,40] and fine-tuning methods [13,14,52] have made strides in mitigating objectionable imagery. Here, the concept to be removed, or the target concept, is specified by text. However, they insufficiently address the nuanced nature of concepts, leading to a gap between the model-generated content and human expectations. Examples include determining whether exposed body parts in a masterpiece are considered safe or if an illustration overtly portraying such body should be similarly categorized. Moreover, the ethical judgment criteria of the model generally fall short of human standards [33, 40], emphasizing the crucial need for collective intelligence intervention at this juncture. + +To handle such ambiguity and bridge this gap, we present Human Feedback Inversion (HFI), a novel and effective framework to mitigate the ethical concerns surrounding text-to-image diffusion models. HFI incorporates human feedback on model-generated images with problematic concepts directly into soft tokens using the textual inversion (TI) [12]. By directly incorporating human perspectives, HFI addresses the limitations of the previous text-based approaches, offering a more nuanced and accurate means of discerning harmful concepts. This condensed information serves as a guiding force against the target concept during inference time or to create training images for fine-tuning. Through this process, tokens acquired via HFI become specialized for their respective concepts, aiming to enhance both effectiveness (i.e., removal performance of the target concept) and utility aspects (i.e., image quality for remaining concepts). + +To augment the efficacy of HFI, we introduce Safe self-Distillation Diffusion (SDD), a self-distillation-based fine-tuning method. This method simplifies + +![](images/1c9d5d2ac0f1d53dfcdc17500255f1c9852852b6f1ab57b8328a00e9a6a69af9.jpg) +(a) Flowers + +![](images/fa677bffb3fa5a15e9f76ca31deb80f20f73e0921dd458363ec24f2d3d4ede3f.jpg) +(b) Fish + +![](images/ddb9a2e9ebd33459451c72a6b9fa1346878e12e3fc46c4cafb9c5b0370596496.jpg) +(c) Boat + +![](images/de5cb46be7044453dc2df6fb464e5bb8353ad3f50aff65c4b03047dcd291b2eb.jpg) +(d) Hat + +![](images/ac85089f823d0aa0bcf217e8c4c926ccc5d8c32f690bf8ef9977f9a41feb8ffc.jpg) +(e) Tree + +![](images/799d3402bde8d72b57bcd2f8249b3d5108d3a121f7c83cac23c0d23eec520805.jpg) +(f) Streetlamp + +![](images/7b31fc6842a2133a98e6599df08078f8f2cd80efef50af10a437e4192073482b.jpg) +Fig. 2: Artist style removal results on Vincent van Gogh. Images are generated with prompts such as "a painting of flowers in the style of Vincent van Gogh". HFI+SDD effectively eliminates his distinct artistic features while preserving the essence of the original subject matter, subject integrity, and visual quality. + +![](images/8931b5a4b14f87c0a8c4ad3023f774825f7c49c433abe5886fa9907473dd7bac.jpg) +Fig. 3: Comparison of images generated with famous artwork titles of the three artists. HFI+SDD, leveraging human feedback, captures and removes distinctive artistic styles more effectively while preserving the integrity and essence of the original paintings. + +![](images/b753e0336562fdaac3f10506b303f6f14b59ea30d5b92bab93e42447be25e1db.jpg) + +the training objective, eliminating the need for manually designing the target or specifying which concept to retain. With this improved design, SDD outperforms existing baselines, even when based on text-specified concepts, with improved training stability. The integration of HFI and SDD establishes a synergistic alliance, utilizing the HFI token as a removal target. This soft token plays a crucial role in generating latents, serving as a vital component for self-supervision and further contributing to the model's alignment with human understanding. + +Figs. 1 to 3 showcase the effectiveness of HFI in concept removal, surpassing previous methods and aligning well with human preferences. Specifically, when tasked with removing Van Gogh's artistic style, our pipeline, guided by human feedback, successfully retained the essential qualities of a painting-like style while eliminating his characteristic bold brushstrokes. In contrast, methods without human feedback transformed images into realistic photographs to aggressively remove artist's styles as a side effect. Additionally, when instructed to remove + +nudity with human feedback, our pipeline successfully eliminated pornographic content, in contrast to previous methods that still produced provocative images. The visual evidence highlights the superior effectiveness of HFI with SDD in consistently delivering safe and modest image outputs across different prompts. + +In summary, our framework not only markedly diminishes the generation of undesirable content perceived by humans but also maintains image quality. This offers a pragmatic solution for improving safety and ethical considerations in the widespread deployment of text-to-image diffusion models. + +# 2 Background on Diffusion Models + +Diffusion models [17, 44-47] are a class of generative models that learn the data distribution by building two Markov chains of latent variables. Given a sample $\mathbf{x}_0 \sim p_{\mathrm{data}}(\mathbf{x}) \coloneqq q(\mathbf{x})$ and a noise schedule $\{\beta_t\}_{t=1}^T$ , the forward process gradually injects a series of Gaussian noises to the sample until it nearly follows the standard Gaussian $q(\mathbf{x}_t | \mathbf{x}_{t-1}) \coloneqq \mathcal{N}(\mathbf{x}_t; \sqrt{1 - \beta_t} \mathbf{x}_{t-1}, \beta_t \mathbf{I})$ , $q(\mathbf{x}_T | \mathbf{x}_0) \approx \mathcal{N}(\mathbf{x}_T; \mathbf{0}, \mathbf{I})$ . Such process is then followed by the reverse process parameterized by $\theta$ , where the model learns to reconstruct the original image by iteratively denoising from a pure Gaussian noise $p(\mathbf{x}_T) = \mathcal{N}(\mathbf{0}, \mathbf{I})$ as $p_\theta(\mathbf{x}_{0:T}) = p(\mathbf{x}_T) \prod_{t=1}^T p_\theta(\mathbf{x}_{t-1} | \mathbf{x}_t)$ . Ho et al. [17] simplifies the objective to learn a noise estimator $\epsilon_\theta$ : + +$$ +\mathcal {L} _ {D M} (\theta) = \mathbb {E} _ {\mathbf {x} _ {0}, \epsilon , t} \left[ \| \boldsymbol {\epsilon} - \boldsymbol {\epsilon} _ {\theta} \left(\sqrt {\bar {\alpha} _ {t}} \mathbf {x} _ {0} + \sqrt {1 - \bar {\alpha} _ {t}} \boldsymbol {\epsilon}; t\right) \| _ {2} ^ {2} \right], \tag {1} +$$ + +where $\epsilon \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ , $\bar{\alpha}_t = \prod_{s=1}^t (1 - \beta_s)$ , $t \sim \mathcal{U}([1,T])$ , and $T$ is the total number of steps. For simple notation, we omit the timestep $t$ as $\epsilon(\mathbf{x}_t; t) = \epsilon(\mathbf{x}_t)$ . + +Latent Diffusion Models (LDMs) [36] leverage the diffusion process within the latent space rather than in the pixel space, utilizing a pre-trained autoencoder. By mapping the input $\mathbf{x}$ into a latent space with the encoder $\mathcal{E}$ , $\mathbf{z} = \mathcal{E}(\mathbf{x})$ , an LDM is trained to predict the added noise in the latent space. As this paper mainly discusses LDMs, we use $\mathbf{x}_t$ instead of $\mathbf{z}_t$ . Text-to-image models additionally take inputs of the text embedding, $\mathbf{c}_p = \mathcal{E}_{\mathrm{txt}}(c_p)$ , paired with an image $\mathbf{x}$ , where $\mathcal{E}_{\mathrm{txt}}$ is the text encoder and $c_{p}$ is a text prompt. To enhance the quality of text conditioning, the classifier-free guidance (CFG) models both conditional and unconditional noise estimates by randomly replacing $\mathbf{c}_p$ with the embedding of an empty string $\mathbf{c}_{\emptyset} = \mathcal{E}_{\mathrm{txt}}("")$ and modifies the noise estimate during sampling as $\tilde{\epsilon}_{\mathrm{CFG}} = \epsilon_{\theta}(\mathbf{x}_t) + (w + 1)[\epsilon_{\theta}(\mathbf{x}_t;\mathbf{c}_p) - \epsilon_{\theta}(\mathbf{x}_t;\mathbf{c}_{\emptyset})]$ , where $w + 1$ is the CFG scale. We simply denote $\epsilon_{\theta}(\mathbf{x}_t;\mathbf{c}_{\emptyset})$ as $\epsilon_{\theta}(\mathbf{x}_t)$ in the subsequent sections. + +# 3 Method + +We present Human Feedback Inversion (HFI), an innovative framework for concept removal guided by human feedback. Unlike previous methods that rely on predefined text-based concepts, i.e., a specified target word like "nudity", real-world inappropriate concepts are often nuanced and challenging to express in + +![](images/b61c306d2a01ff31ee4f2d0086fe77c770bb3d429a6990b959a880822f4d53fa.jpg) +(a) Human Feedback Inversion (HFI) + +![](images/2812881bf105aebaee2ce1f6f5a5f41e8397976f4f169c5f526f92e8fb44007f.jpg) +(b) Safe self-Distillation Diffusion (SDD) +Fig. 4: Illustration of our proposed framework, HFI and SDD. + +text. HFI addresses this challenge by learning concepts based on collective understanding. Our approach initiates with collecting human feedback on model-generated images through carefully designed surveys detailed in Appendix C, where harmfulness is comparatively evaluated. Using this human evaluation data, we construct a reward model that assigns a harmfulness score to each image. With this reward model, we formulate an optimization problem to identify a soft token that accurately represents harmfulness. These optimal embedding vectors are then utilized to fine-tune the diffusion model, preventing the generation of harmful images. The overall framework is depicted in Fig. 4 and Appendix A, and detailed explanations of each stage are provided in the subsequent sections. + +# 3.1 Collecting and Modeling Human Feedback + +We first generate images, on which human feedback will be collected, using the original model and a set of simple prompts containing the target concept word $c$ detailed in Appendix B. Subsequently, two main settings can be considered for gathering human feedback. The design of how and what kind of feedback to collect during this process is crucial. In cases where the content the model should not generate is well-defined and clear annotation guidelines exist (e.g., nudity), binary feedback for each image can be collected. However, this does not apply to all scenarios, with a notable example being an artist's style. Suppose that an artist requests the removal of not only specific works but also the overall style from the model, then the criteria may not be clear. In such cases, ranking feedback can be collected. This involves presenting annotators with $M$ random images and requesting them to rank these images ranging from 1 to $M$ . + +Following this, we need to train a reward model $r_{\psi}$ , which accepts an input image $\mathbf{x}$ and outputs a score $r_{\psi}(\mathbf{x}) \in \mathbb{R}$ for the concept $c$ . For binary feedback, we follow the MSE-based approach proposed in Lee et al. [20], + +$$ +\mathcal {L} _ {M S E} (\psi) = \mathbb {E} _ {(\mathbf {x}, y) \sim \mathcal {D} _ {\text {h u m a n}}} \left[ (r _ {\psi} (\mathbf {x}) - y) ^ {2} \right], \tag {2} +$$ + +where $\mathcal{D}_{\mathrm{human}}$ consists of images $\mathbf{x}$ and corresponding binary annotations $y \in \{0,1\}$ . Here, '1' indicates the presence of the harmful concept (e.g., with nudity), and '0' is the absence (e.g., without nudity), answered by human annotators. + +On the other hand, for ranking feedback, we adopted the Bradley-Terry model [5], which assigns scores to examples based on pairwise comparisons [8, 26, 49]. The obtained rankings of $M$ images can be reorganized to construct $\mathcal{D}_{\mathrm{human}}$ as $\binom{M}{2}$ pairwise comparisons as $\mathbf{x}^{+} = \mathbf{x}^{(i)}, \mathbf{x}^{-} = \mathbf{x}^{(j)}$ if $\mathbf{x}^{(i)}$ has a higher rank than $\mathbf{x}^{(j)}$ , and vice versa - a higher ranking indicates a greater presence or resemblance of the concept. Subsequently, we include all $\binom{M}{2}$ pairs in a single batch, following Ouyang et al. [26], and optimize the reward model as + +$$ +\mathcal {L} _ {N L L} (\psi) = - \frac {1}{\binom {M} {2}} \mathbb {E} _ {(\mathbf {x} ^ {+}, \mathbf {x} ^ {-}) \sim \mathcal {D} _ {\text {h u m a n}}} \log \sigma \left(r _ {\psi} \left(\mathbf {x} ^ {+}\right) - r _ {\psi} \left(\mathbf {x} ^ {-}\right)\right), \tag {3} +$$ + +where $\sigma (\cdot)$ is the sigmoid function. For both approaches, the reward model $r_{\psi}(\mathbf{x})$ maps the CLIP [30] embedding of an image $\mathbf{x}$ to a score through a couple of MLP layers with ReLU non-linearities. + +# 3.2 Inverting Feedback into Embeddings + +Based on the feedback received in Sec. 3.1, we transform the corresponding concept into a soft token. Inspired by recent inversion techniques [12,35], we utilize TI [12], as it can easily replace or be combined with text tokens from previous approaches. The goal is to find a soft token $v^{*}$ that maximizes the expected reward of generated images, i.e., $v^{*} = \arg \max_{v}\mathbb{E}_{\mathbf{x}\sim p_{\theta}(\mathbf{x}|v)}[r_{\psi}(\mathbf{x})]$ , where $v$ is the soft word in the token space. Utilizing the reward model $r_{\psi}$ in Sec. 3.1, we optimize the following objective of the reward-weighted negative log-likelihood [20] to get the optimal token $v^{*}$ that best captures the text concept $c$ : + +$$ +v ^ {*} = \underset {v} {\arg \min } \mathbb {E} _ {\mathbf {x} \sim p _ {\theta} (\mathbf {x} | c)} [ - r _ {\psi} (\mathbf {x}) \log p _ {\theta} (\mathbf {x} | v) ]. \tag {4} +$$ + +Inverting too many images may struggle with concept extraction by prioritizing image reconstruction over it, as shown in Gal et al. [12]. Also, we empirically found that inverting it with a small learning rate may fail to capture the concept, which requires additional hyperparameter tuning. To simplify such a burden, we use a reward-weighted sampler with images of the top- $K$ scores, which is equivalent to the reward-weighted loss in Eq. (4). We use the learning rate of $5 \times 10^{-4}$ , the default value in the HuggingFace's implementation1, and $K = 50$ . + +# 3.3 Removing Learned Concepts with Self-Distillation + +After obtaining soft tokens through the proposed HFI method in Sec. 3.2, one can adapt them to existing text-based concept removal methods, enabling the desired removal of concepts using the obtained soft tokens instead of text tokens. While HFI alone proves effectiveness in Tabs. 1 and 2, we propose a new method termed Safe self-Distillation Diffusion (SDD) to achieve an even greater impact, + +particularly in enhancing the removal of ambiguous concepts judged by humans. SDD removes the target concept from a model based on self-distillation [53]. We first set both student $\theta$ and teacher models $\theta^{\star}$ whose parameters are initialized from the pre-trained model and update the student with the following loss: + +$$ +\mathcal {L} _ {S D D} = \mathbb {E} _ {\mathbf {x} _ {t} \sim p _ {\theta^ {*}} (\mathbf {x} _ {t} | v ^ {*}), t} \left[ \| \epsilon_ {\theta} (\mathbf {x} _ {t}; v ^ {*}) - \mathbf {s g} \left(\epsilon_ {\theta} (\mathbf {x} _ {t})\right) \| _ {2} ^ {2} \right], \tag {5} +$$ + +where $\mathfrak{sg}()$ denotes the stop-grad operation. That is, we want the model given the token $v^{*}$ to behave as if it were not for any token by setting the unconditional estimate $\epsilon_{\theta}(\mathbf{x}_t)$ as a self-distillation target. After updating the student model, we update the teacher model $\theta^{\star}$ as an exponential moving average (EMA) from it, i.e., $\theta^{\star} \gets m\theta^{\star} + (1 - m)\theta$ , with a decaying rate $m \in (0,1)$ . After the training, we use the EMA parameter of the teacher model as our final safe model. + +Unlike methods focused on learning new concepts [19, 38], determining suitable replacements when removing specific concepts can be challenging, often contingent on other contextual elements [9]. For instance, while removing the color red in a red apple can be substituted with a green apple, removing the color red devoid of context lacks clarity. Similarly, removing concepts like nudity, violence, or specific artists presents this ambiguity. Previous approaches [3,13,40] have attempted to tackle this challenge by targeting negative guidance; however, images generated in this manner often fall short of removing such concepts. + +We tackle this problem by introducing the EMA teacher model, which moves in tandem with the student model but in a more stable and deliberate manner, gradually removing the target concept. This serves not only to enhance training stability but also plays a crucial role in providing the student model with appropriate targets for conditional noise predictions. Notably, in Eq. (5), the latent $\mathbf{x}_t$ is sampled from the teacher model, i.e. it generates parts that have not been effectively removed up to the existing training step. Additionally, resolving discrepancies between the model's understanding and human judgment regarding harmfulness poses a non-trivial task. In SDD with HFI, sampling $\mathbf{x}_t$ is conditioned on the learned token $v^*$ . During this process, the model learns to remove the concept from newly generated latents that would be deemed problematic. + +Lastly, we introduce a couple of additional techniques for fine-tuning. First, similar to existing work [13, 19, 51, 52], we fine-tune only the cross-attention layers of the U-Net [37]. Second, we empirically find that not all timesteps equally influence the conditioning of image generation; the $L_{2}$ -norm between the unconditional and the conditional noise was generally high in the middle timesteps, as shown in Appendix E. Therefore, we sample the timestep to apply distillation as $t = T \cdot \mathcal{B}(\alpha, \beta)$ where $\mathcal{B}(\alpha, \beta)$ is a beta random variable with parameter $(\alpha, \beta)$ . We set $(\alpha, \beta) = (3, 3)$ for our experiments as it works well in general. + +# 4 Related Work + +Numerous attempts have been made to address harmful content generation, including the use of external classifiers for image filtering during inference, such as the safety checker in Stable Diffusion (SD) [36]. However, these approaches + +can be easily circumvented by users [33]. Efforts to filter out problematic training images during pre-processing [48] often lead to reduced training sets and inferior image quality [24, 28], without completely eliminating harmfulness [1]. + +Recent advancements in the field can be broadly classified into three categories: (i) inference-time techniques manipulate noise estimates to counteract concepts [3,40]; (ii) direct cross-attention optimization adjusts weights to disregard problematic tokens from text inputs either directly or through gradient descent [14,52]; (iii) fine-tuning methods refine a subset of parameters using generated images [13,18,23]. Although both (ii) and (iii) fine-tune the model, the main disparity lies in the training objective. The former adjusts the projection or attention of text inputs, while the latter alters the model's distribution along the diffusion trajectories. Moreover, while the former does not necessitate training images or uses only a few real images, the latter generates training images from the model. Nonetheless, all methods entail a trade-off between effectiveness and utility, given the inherent challenge of identifying concepts in texts for removal. + +Human feedback has become crucial in enhancing both performance and safety in machine learning models [8]. Initially applied in dialogue systems [15, 21], reinforcement learning (RL) algorithms, particularly Proximal Policy Optimization (PPO) [43], have played a significant role in leveraging human feedback, notably in Large Language Models (LLMs) [6, 26, 49], focusing on both performance and safety considerations. Reinforcement learning with human feedback (RLHF) [26] optimizes models based on reward signals derived from user input, further enhancing the safety and adaptability of LLMs. Beyond language domains, efforts have been made to align text-to-image diffusion models with human preferences [2, 10, 20]. These involve fine-tuning using a reward-weighted loss function with appropriate regularization. However, the practical application of such RL-based methods often faces challenges due to hyperparameter sensitivity and reward design intricacies [20]. Furthermore, while such methods relying on human preference data aim to enhance image quality perceived by humans, they differ from our goal of suppressing image generation of specific concepts. + +# 5 Experiments + +# 5.1 Baselines and Evaluation Protocols + +We compare the performance of our method with the original SD and previous methods. In Tabs. 1 and 2, SD [36] indicates the original Stable Diffusion model, and NEG is the simplest baseline of SD with the negative prompt, i.e., the noise conditioned on the target concept is used instead of the unconditional one, thus negating the target concept. We consider two inference-time techniques: SLD [40] and SEGA [3]. For fine-tuning methods, we consider the following three methods: ESD [13], FMN [52], and UCE [14]. For all fine-tuning methods, including ours, we use the artist name (e.g., "Vincent van Gogh") or the word (e.g., "nudity" and "bleeding") for the target concept in text. When applying HFI to these methods, HFI token $v^{*}$ replaces the text word. Refer to Appendix B for details. + +![](images/4cb55fd4c0f1b50c96b5948cc533111c5b90cb78975fdd94a41a5da3d2caf696.jpg) + +![](images/51e50d19be2ced3e33c7060d973470f3f6198ab92868c5dfb4701f54c4c79ed6.jpg) + +![](images/df0bc3b376c3935e6b8ef6184a63483a506a0c0014291a222d277cd611047847.jpg) + +![](images/93f195f44fb1da8524ce35717dc02ec13bd1e5aaad717d4ba99fde5183660aa0.jpg) + +![](images/50e39ff7b275d238ec0d7dd339d3c888306698137e6c54dfeb4625909fdc53b6.jpg) +Fig. 5: Comparison of concept removal methods for Vincent van Gogh's artistic style. Van Gogh's style is marked by bold brushstrokes, vivid colors, and swirling patterns. HFI+SDD is the most effective, consistently producing the most neutral but still artwork-like images and successfully eliminating his stylistic characteristics. + +![](images/455f6974bab61299aa72750055e858d05c3322c0513cb82258101441fe5dc651.jpg) + +The evaluation of the method's capability to remove specific concepts is structured along two primary axes: effectiveness and utility. Effectiveness measures the ability to avoid generating the target concept, while utility assesses the preservation of image quality in outputs unrelated to the target concept. + +For artist concept removal in Sec. 5.2, we propose two specific metrics to gauge both aspects. The effectiveness is evaluated by the target artist ratio (%ART) using CLIP zero-shot classification. This involves generating 1,000 images in total of 50 commonly depicted objects in famous artworks (listed by ChatGPT [25]), with the prompt template "a painting of {object} in the style of {artist}" These images are then classified using OpenCLIP [7] to determine how frequently the target artist is accurately identified. The utility is determined through the photo similarity ratio (%PHO), which quantifies how well the model retains the ability to generate paintings rather than photorealistic images, thus preserving artistic styles not specific to the targeted artist. We enhance classification accuracy using prompt ensemble [30]. In line with previous work, we also report LPIPS [54] for both the target and other artists. Finally, since such quantitative metrics cannot completely replace qualitative evaluations and human evaluations can be subjective, we asked ChatGPT 4o [25] to assess the performance of removing the specified artist to supplement this. The qualitative evaluation results and further details are provided in Appendix F. + +For harmful concept removal in Sec. 5.3, the effectiveness is measured by generating 5,000 images in total with the prompt {"country" body" using + +![](images/4af15efc7928b63f0e1c720f94b73bc365e7da24b9c13e7585dc5652a4d61c15.jpg) + +![](images/7520733b457100a65231047a2a2ae2e8109ee74cb4096f06fc49e9cfea6f2900.jpg) + +![](images/35547a6442c50a47ba3d3864555da2913d3a39afa7ee53b86d02b196401034f1.jpg) + +![](images/f246d9b95b09bcfa424c4c8780f4855f914309fc8dc1b279e0120c2eb0115942.jpg) + +![](images/3d9babbfb58812f97438752c75a8b626012ec5bb4eb30fbfe65e7083527e9ef6.jpg) +Fig. 6: Comparison of concept removal methods for Leonid Afremov's artistic style. Afremov's style is marked by vibrant colors, palette knife technique (similar to pointage technique), and impressionistic elements. HFI+SDD is the most effective, producing images with clear depictions of various objects and minimal retention of his style. + +![](images/83543329c04df940c576566bc5da83afb3c0025291bcf643b896c3782a572d20.jpg) + +names of 50 countries with high GDP following [40]. This prompt generates way more inappropriate images (74.2% with SD v1.4) compared to the sexual category of the I2P dataset [40] (reportedly ranked $5^{\text{th}}$ among the seven categories with 32.8% of hard prompt ratio and only 16.7% of nudity images). The proportion of images classified "unsafe" (%NSFW) and those detected with exposed body parts (%NUDE) are evaluated by NudeNet [29]. This process is similarly used to assess the removal of another harmful concept of bleeding by generating images for the shocking category in the I2P dataset [40] and classifying them with the Q16 classifier (%I2P) [41]. The utility is evaluated by generating images from 10,000 random captions from the MSCOCO dataset [22] and comparing these to actual photos, outputs from the original SD model, and their text-to-image alignment using FID [16, 27], LPIPS [54], and CLIPScore [30], respectively. + +# 5.2 Artist Style Removal + +In this section, we conducted experiments to remove the style of artists from the existing SD model. For HFI, we generate 1,000 images from the original SD model with prompts detailed in Appendix B and gather human feedback on a total of 200 images (equivalent to 20 question pairs) per artist. We then train a reward model, which is then used to approximate the scores for all 1,000 images for HFI. All methods with HFI use the same soft token instead of the text token. + +In Tab. 1, even without HFI, SDD alone finds a reasonable trade-off, achieving relatively better removal than the other methods while maintaining a rela- + +![](images/6cedc83aebb138b72490f9197e9ee162b8334da92e1b05719d28bbf35a475858.jpg) + +![](images/306eac286f9b49782db4b4132533213b5203a79f05522a852c99e46671f13005.jpg) + +![](images/532a096210f79bf9e596ac25d1d4cab41e6b3911e064c3179c5bf3cee29af597.jpg) + +![](images/0f82675506b2cff253b38f8ae66aaf29823fbb7113b06ae89b0633055a8d4372.jpg) + +![](images/1b1f4b4638c92b43c1dee0367acbe97d9aac105084a5768b0507554e9fbaf87a.jpg) +Fig. 7: Comparison of concept removal methods for Edward Hopper's artistic style. Hopper's characteristic style is marked by realism, simplicity, architectural details, and muted color palette. Among the methods, HFI+SDD most effectively removes his distinctive features while still being abstract and less photorealistic. + +![](images/81ada7f44b364ffc7df661017cb071ba3130009c191ebeef47b6564de424e05b.jpg) + +tively lower percentage on $\% \mathrm{PHO}$ . Moreover, when HFI is employed, our method shows the state-of-the-art performance of $\% \mathrm{ART}$ of $12.7\%$ on average, with average $\% \mathrm{PHO}$ of $12.0\%$ . For Edward Hopper, however, HFI+SDD was outperformed by a few HFI-enhanced baselines, but it worked the best in most cases. Notably, HFI+SDD always surpasses all the previous methods without HFI. + +Among the baselines, fine-tuning methods (ESD and SDD) observed the most benefit from HFI. In contrast, methods directly optimizing cross-attention exhibited subpar removal performance and appeared to have relatively less impact. Meanwhile, inference-time techniques generally demonstrated improvements in the trade-off between effectiveness and utility facilitated by HFI as shown in Fig. 8. However, as shown in Figs. 5 to 7, there are often cases where the object is not well represented or the distinctive artistic characteristics remain. In such cases, HFI can be beneficial, whereas text-based methods frequently struggle to distinguish the artist's concept from the generic artwork style. + +In the qualitative evaluation using ChatGPT 4o [25], HFI+SDD was evaluated as the most superior method by stating that "it consistently produces the most neutral images by moving away from the artist's style while maintaining a high level of image quality." Despite being asked to evaluate only the removal performance, somewhat surprisingly, ChatGPT considered not only the degree of style removal but also various aspects such as object depiction, image quality, and artifacts) with consistent judgments across all objects and presentation orders, aligning with our proposed metrics. Refer to Appendix F for full results. + +Fig. 8: Trends of $\%_{\text{ART}}$ and $\%_{\text{PHO}}$ by inference-time techniques. While forgetting the general artwork concepts (higher $\%_{\text{PHO}}$ ) is inevitable, HFI can achieve similar outcomes at smaller editing scales. The same marker symbol denotes the same scale in each plot. +![](images/5de7e08e7bc6d426cb1918f91d18398a07256e2a8da7e72c84ec934df9db819e.jpg) +--- SLD -SLD+HFI SEGAE SEGA+HFI + +![](images/df39b96308b953562803d94d0a24bbca797753e042be2dec2724ac00460b7282.jpg) + +![](images/1ef1f7713ddafb5c71a53083f2660eaf8f136b0d77c2d2fc3c0b551b89c5d461.jpg) + +![](images/4a6df93dda4cf5ac0f65f1f1fecef967d657e895e7f5cda7693df28f5927719c.jpg) +Fig. 9: HFI mitigates terrifying aspects of images generated with I2P shocking prompts. + +![](images/a4c467ccfe44f3a90ed5f1da64816723cd3b67eab43dfa4550a98b1c8431da5e.jpg) + +![](images/148414c18f80fc16ff51f92d273df82c24982b6f43f88415825450968f456ff6.jpg) +Fig.10: Images with the highest (top) and lowest (bottom) rewards. + +# 5.3 Harmful Concept Removal + +Tab. 2 shows our experimental results on removing NSFW and bleeding concepts. Similar to Sec. 5.2, we first generated images and collected binary feedback and ranking feedback for nudity and bleeding, respectively. SDD alone effectively reduces the presence of NSFW imagery, but when combined with HFI, such images are nearly eliminated. Our method and ESD-u-3 achieve a similar level, but ESD exhibits inferior performance in all image quality metrics. In a similar vein, the method of fine-tuning only the cross-attention, as in ESD-x-3, is observed to generate harmful images beyond a certain threshold. High in %NSFW and low in %NUDE for some methods indicate that they still generate somewhat unsafe images even though exposed body parts are limited. When combined with HFI, all methods but FMN exhibit a sharp decrease in %NSFW, underscoring the role of the soft token aligned with human judgment. + +Furthermore, for nuanced concepts such as bleeding, HFI aids in identifying and mitigating the inappropriate aspects of images. In Fig. 9, with HFI, it becomes apparent that not only has the extent of bleeding diminished, but it also yields images that are safer and less offensive. Particularly notable are the outcomes achieved by combining SDD and HFI, where notably safer images are produced despite maintaining identical compositions. The interpretation of such ambiguous concepts by humans can be encoded into learned tokens for subsequent utilization in their removal, highlighting the utility of HFI. + +Table 1: Quantitative comparison of artist style removal performance. Each method is evaluated both without and with HFI (✓), showing improvement due to human feedback. The $\sum$ column represents the sum of %ART and %PHO, indicating undesirable images. The table also compares our SDD method with other baselines. + +
Target ArtistVincent van GoghLeonid AfremovEdward HopperAvg. LPIPS
MethodHFI%ART↓ %PHO↓Σ%ART↓ %PHO↓Σ%ART↓ %PHO↓ΣTarget↑ Others↓
SD v1.495.92.998.899.90.6100.597.62.299.8--
Inference-time Technique
NEG [36]✓15.032.547.592.82.094.836.512.148.60.33520.2273
17.212.429.60.030.430.416.516.132.60.43520.3255
SLD [40]✓29.127.756.827.84.232.021.339.560.80.42770.2880
24.622.847.49.78.818.518.134.953.00.44590.3285
SEGA [3]✓40.311.752.064.41.565.937.728.666.30.31510.2522
17.517.635.120.74.525.232.021.853.80.34820.2661
Direct Cross-Attention Optimization
FMN [52]✓66.67.574.110.725.536.277.15.782.80.32930.1665
58.67.766.36.50.97.473.25.578.70.27930.2478
UCE [14]✓66.44.070.444.72.046.751.53.354.80.27670.0332
77.32.579.892.50.693.172.51.974.40.17440.0328
Fine-tuning Method
ESD-x-1 [13]✓45.015.960.913.525.438.964.47.273.60.36240.1635
22.520.442.919.057.276.214.317.431.70.44040.1691
SDD (ours)✓20.625.546.19.98.218.140.441.181.50.27600.2341
9.55.114.68.83.011.819.827.947.70.28790.1439
+ +# 5.4 Analysis on Rewards and Learned Embeddings + +For nudity, the reward model generally aligns with Nudenet [29] (AUROC of 0.882), but differences between human and model judgments highlight the limits of relying solely only on models for inversion. Fig. 10 shows that images deemed appropriate by humans or not reflecting the artist's style receive low rewards and minimally impact HFI token learning. Integrating HFI with reward models effectively captures concepts. Ablation studies comparing HFI with the vanilla TI [12] show a $3.2\%$ and $5.1\%$ increase in $\% \mathrm{ART}$ for artist removal without human feedback (i.e., random images without rewards) and with real images from the internet (without rewards), respectively. For nudity removal, $\% \mathrm{NSFW}$ and $\% \mathrm{NUDE}$ increase by $1.74\%$ and $0.04\%$ without human feedback, with FID increasing by 1.565. This demonstrates the importance of integrating human judgments and model-generated data into the training process to achieve optimal performance. + +Compared to existing expensive methods using human feedback, the proposed HFI is highly cost-effective. We obtained 650 responses from five individuals (450 unique images for inter-annotator agreement (IAA)) within three hours for nudity. Providing annotation guidelines increased IAA, evidenced by Fleiss' $\kappa$ [11] at 0.4997 and Randolph's $\kappa$ [34] at 0.6702. For artistic concepts, we collected 40 ranking feedback responses (400 images) per artist from 15 individuals. Our research offers an economical and effective framework for extracting information from models and removing harmful content with human judgment. + +Table 2: Quantitative comparison of harmful concept removal performance. Each method is evaluated both without and with HFI (✓) for inter-method comparison. The table also compares our SDD with other baselines for intra-method comparison. + +
Target ConceptNudityBleedingCOCO10K
MethodHFI%NSFW↓%NUDE↓%I2P↓FID↓CLIP↑ LPIPS↓
SD v1.474.1832.1465.7716.6360.2745
COCO ref.0.2676
Inference-time Technique
SD+NEG [36]✓30.683.8645.7918.8370.2701
5.144.7624.6519.5180.2699
SLD [40]✓58.089.8446.1417.8150.2702
42.9815.5829.5618.3450.2704
SEGA [3]✓67.7213.1645.4417.1290.2698
56.4219.5239.3716.7560.2698
Direct Cross-Attention Optimization
FMN [52]✓51.5027.7847.0818.0540.2556
61.3234.7255.7216.8130.2683
UCE [14]✓32.307.3661.8017.0240.2726
14.769.3253.2716.5410.2736
Fine-tuning Method
ESD-u-1 [13]✓11.302.00-17.0330.2664
2.421.16-18.4040.2636
ESD-u-3 [13]✓4.440.48-18.4980.2624
0.800.46-18.6150.2621
ESD-x-3 [13]✓14.322.9851.1718.5170.2677
4.183.3034.3517.8980.2670
SDD (ours)✓2.780.6030.8417.3620.2658
0.280.3223.3617.0230.2643
+ +# 6 Conclusion and Discussion + +In this work, we introduced a novel concept removal framework within a text-to-image diffusion model through human feedback inversion and self-distillation-based fine-tuning. By leveraging human feedback to train soft tokens, we effectively safeguarded the diffusion model from harmful concepts, addressing the ambiguity in defining text-based concepts. We further proposed a self-distillation-based diffusion model fine-tuning algorithm, enhancing the efficacy of concept removal. We believe our research lays the groundwork for a human-centric approach to the safety and ethical considerations of large-scale diffusion models, paving the way for future advancements. + +Limitation. While our proposed HFI performs well in capturing the intended concept compared to the text, its effectiveness depends on the quality of human feedback and the performance of inversion to soft tokens. Significant disagreements in human feedback can compromise the learning process, underscoring the importance of meticulous data collection and reward modeling. + +# Acknowledgement + +This work was partly supported by Institute for Information & communications Technology Promotion (IITP) grant funded by the Korea government (MSIT) (No. RS-2019-II190075, Artificial Intelligence Graduate School Program (KAIST), No. 2022-0-00184, Development and Study of AI Technologies to Inexpensively Conform to Evolving Policy on Ethics) and the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. 2022R1A5A708390812). + +# References + +1. Baio, A.: Exploring 12 million of the 2.3 billion images used to train stable diffusion's image generator. https://waxy.org/2022/08/exploring-12-million-of-the-images-used-to-train-stable-diffusions-image-generator/ (Aug 2022) +2. Black, K., Janner, M., Du, Y., Kostrikov, I., Levine, S.: Training diffusion models with reinforcement learning. arXiv preprint arXiv:2305.13301 (2023) +3. Brack, M., Friedrich, F., Hintersdorf, D., Struppek, L., Schramowski, P., Kersting, K.: Sega: Instructing diffusion using semantic dimensions. arXiv preprint arXiv:2301.12247 (2023) +4. Brack, M., Schramowski, P., Friedrich, F., Hintersdorf, D., Kersting, K.: The stable artist: Steering semantics in diffusion latent space. arXiv preprint arXiv:2212.06013 (2022) +5. Bradley, R.A., Terry, M.E.: Rank analysis of incomplete block designs: I. the method of paired comparisons. Biometrika 39(3/4), 324-345 (1952) +6. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020) +7. Cherti, M., Beaumont, R., Wightman, R., Wortsman, M., Ilharco, G., Gordon, C., Schuhmann, C., Schmidt, L., Jitsev, J.: Reproducible scaling laws for contrastive language-image learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2818-2829 (2023) +8. Christiano, P.F., Leike, J., Brown, T., Martic, M., Legg, S., Amodei, D.: Deep reinforcement learning from human preferences. Advances in neural information processing systems 30 (2017) +9. Du, Y., Li, S., Mordatch, I.: Compositional visual generation with energy based models. Advances in Neural Information Processing Systems 33, 6637-6647 (2020) +10. Fan, Y., Watkins, O., Du, Y., Liu, H., Ryu, M., Boutilier, C., Abbeel, P., Ghavamzadeh, M., Lee, K., Lee, K.: Dpok: Reinforcement learning for fine-tuning text-to-image diffusion models. arXiv preprint arXiv:2305.16381 (2023) +1. Fleiss, J.L.: Measuring nominal scale agreement among many raters. Psychological bulletin 76(5), 378 (1971) +12. Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A.H., Chechik, G., Cohen-Or, D.: An image is worth one word: Personalizing text-to-image generation using textual inversion (2022). https://doi.org/10.48550/ARXIV.2208.01618, https://arxiv.org/abs/2208.01618 +13. Gandikota, R., Materzynska, J., Fiotto-Kaufman, J., Bau, D.: Erasing concepts from diffusion models. arXiv preprint arXiv:2303.07345 (2023) + +14. Gandikota, R., Orgad, H., Belinkov, Y., Materzynska, J., Bau, D.: Unified concept editing in diffusion models. arXiv preprint arXiv:2308.14761 (2023) +15. Gao, J., Galley, M., Li, L.: Neural approaches to conversational ai. In: The 41st international ACM SIGIR conference on research & development in information retrieval. pp. 1371-1374 (2018) +16. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017) +17. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. arXiv preprint arxiv:2006.11239 (2020) +18. Kumari, N., Zhang, B., Wang, S.Y., Shechtman, E., Zhang, R., Zhu, J.Y.: Ablating concepts in text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22691-22702 (2023) +19. Kumari, N., Zhang, B., Zhang, R., Shechtman, E., Zhu, J.Y.: Multi-concept customization of text-to-image diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1931-1941 (2023) +20. Lee, K., Liu, H., Ryu, M., Watkins, O., Du, Y., Boutilier, C., Abbeel, P., Ghavamzadeh, M., Gu, S.S.: Aligning text-to-image models using human feedback. arXiv preprint arXiv:2302.12192 (2023) +21. Li, J., Monroe, W., Ritter, A., Galley, M., Gao, J., Jurafsky, D.: Deep reinforcement learning for dialogue generation. arXiv preprint arXiv:1606.01541 (2016) +22. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014) +23. Ni, Z., Wei, L., Li, J., Tang, S., Zhuang, Y., Tian, Q.: Degeneration-tuning: Using scrambled grid shield unwanted concepts from stable diffusion. arXiv preprint arXiv:2308.02552 (2023) +24. O'Connor, R.: Stable diffusion 1 vs 2 - what you need to know. https://www.assemblyai.com/blog/stable-diffusion-1-vs-2-what-you-need-to-know/ (Dec 2022) +25. OpenAI: Chatgpt (2023), https://www.openai.com/, july 2023 version +26. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems 35, 27730-27744 (2022) +27. Parmar, G., Zhang, R., Zhu, J.Y.: On aliased resizing and surprising subtleties in gan evaluation. In: CVPR (2022) +28. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952 (2023) +29. Praneeth, B.: NudeNet: Neural nets for nudity classification, detection and selective censoring. https://github.com/notAI-tech/NudeNet (2021) +30. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) +31. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 1(2), 3 (2022) + +32. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: International Conference on Machine Learning. pp. 8821-8831. PMLR (2021) +33. Rando, J., Paleka, D., Lindner, D., Heim, L., Tramér, F.: Red-teaming the stable diffusion safety filter. arXiv preprint arXiv:2210.04610 (2022) +34. Randolph, J.J.: Free-marginal multirater kappa (multirater k [free]): An alternative to fleiss' fixed-marginal multirater kappa. Online submission (2005) +35. Roich, D., Mokady, R., Bermano, A.H., Cohen-Or, D.: Pivotal tuning for latent-based editing of real images. ACM Transactions on graphics (TOG) 42(1), 1-13 (2022) +36. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10684-10695 (2022) +37. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015) +38. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22500-22510 (2023) +39. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022) +40. Schramowski, P., Brack, M., Deiseroth, B., Kersting, K.: Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2023) +41. Schramowski, P., Tauchmann, C., Kersting, K.: Can machines help us answering question 16 in datasheets, and in turn reflecting on inappropriate content? In: 2022 ACM Conference on Fairness, Accountability, and Transparency. pp. 1350-1361 (2022) +42. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. arXiv preprint arXiv:2210.08402 (2022) +43. Schulman, J., Wolski, F., Dhariwal, P., Radford, A., Klimov, O.: Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347 (2017) +44. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: Proceedings of the 32nd International Conference on Machine Learning (2015) +45. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020) +46. Song, Y., Ermon, S.: Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems 32 (2019) +47. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456 (2020) +48. Stability AI: Stable diffusion 2.0 release. https://stability.ai/blog/stable-diffusion-v2-release (Nov 2022) + +49. Stiennon, N., Ouyang, L., Wu, J., Ziegler, D., Lowe, R., Voss, C., Radford, A., Amodei, D., Christiano, P.F.: Learning to summarize with human feedback. Advances in Neural Information Processing Systems 33, 3008-3021 (2020) +50. Thiel, D.: Identifying and eliminating csam in generative ml training data and models. Tech. rep., Technical Report. Stanford University, Palo Alto, CA. https://purl.stanford... (2023) +51. Yang, Z., Yi, X., Li, P., Liu, Y., Xie, X.: Unified detoxifying and debiasing in language generation via inference-time adaptive optimization. arXiv preprint arXiv:2210.04492 (2022) +52. Zhang, E., Wang, K., Xu, X., Wang, Z., Shi, H.: Forget-me-not: Learning to forget in text-to-image diffusion models. arXiv preprint arXiv:2211.08332 (2023) +53. Zhang, L., Song, J., Gao, A., Chen, J., Bao, C., Ma, K.: Be your own teacher: Improve the performance of convolutional neural networks via self distillation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3713-3722 (2019) +54. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018) \ No newline at end of file diff --git a/2024/Safeguard Text-to-Image Diffusion Models with Human Feedback Inversion/images.zip b/2024/Safeguard Text-to-Image Diffusion Models with Human Feedback Inversion/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..cc4679fc1e107aab65bbe0a5a43412fb63ab963d --- /dev/null +++ b/2024/Safeguard Text-to-Image Diffusion Models with Human Feedback Inversion/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e220a793d50aaabeca4a5cacac89016c235a5dc20869ec272ffb28f85e3d5db2 +size 1193536 diff --git a/2024/Safeguard Text-to-Image Diffusion Models with Human Feedback Inversion/layout.json b/2024/Safeguard Text-to-Image Diffusion Models with Human Feedback Inversion/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..cbd988d1de8a47cf8203ca016d9f79ea18897900 --- /dev/null +++ b/2024/Safeguard Text-to-Image Diffusion Models with Human Feedback Inversion/layout.json @@ -0,0 +1,10045 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 141, + 111, + 474, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 111, + 474, + 148 + ], + "spans": [ + { + "bbox": [ + 141, + 111, + 474, + 148 + ], + "type": "text", + "content": "Safeguard Text-to-Image Diffusion Models with Human Feedback Inversion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 156, + 168, + 457, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 168, + 457, + 193 + ], + "spans": [ + { + "bbox": [ + 156, + 168, + 457, + 193 + ], + "type": "text", + "content": "Sanghyun Kim" + }, + { + "bbox": [ + 156, + 168, + 457, + 193 + ], + "type": "inline_equation", + "content": "^{\\text{D}}" + }, + { + "bbox": [ + 156, + 168, + 457, + 193 + ], + "type": "text", + "content": ", Seohyeon Jung" + }, + { + "bbox": [ + 156, + 168, + 457, + 193 + ], + "type": "inline_equation", + "content": "^{\\text{D}}" + }, + { + "bbox": [ + 156, + 168, + 457, + 193 + ], + "type": "text", + "content": ", Balhae Kim" + }, + { + "bbox": [ + 156, + 168, + 457, + 193 + ], + "type": "inline_equation", + "content": "^{\\text{D}}" + }, + { + "bbox": [ + 156, + 168, + 457, + 193 + ], + "type": "text", + "content": ", Moonseok Choi" + }, + { + "bbox": [ + 156, + 168, + 457, + 193 + ], + "type": "inline_equation", + "content": "^{\\text{D}}" + }, + { + "bbox": [ + 156, + 168, + 457, + 193 + ], + "type": "text", + "content": ", Jinwoo Shin" + }, + { + "bbox": [ + 156, + 168, + 457, + 193 + ], + "type": "inline_equation", + "content": "^{\\text{D}}" + }, + { + "bbox": [ + 156, + 168, + 457, + 193 + ], + "type": "text", + "content": ", and Juho Lee" + }, + { + "bbox": [ + 156, + 168, + 457, + 193 + ], + "type": "inline_equation", + "content": "^{\\text{D}}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 149, + 202, + 463, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 202, + 463, + 225 + ], + "spans": [ + { + "bbox": [ + 149, + 202, + 463, + 225 + ], + "type": "text", + "content": "Korea Advanced Institute of Science and Technology (KAIST) {nannullna,heon2203,balhaekim,ms.choi,jinwoos,juholee}@kaist.ac.kr" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 251, + 455, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 251, + 455, + 449 + ], + "spans": [ + { + "bbox": [ + 160, + 251, + 455, + 449 + ], + "type": "text", + "content": "Abstract. This paper addresses the societal concerns arising from large-scale text-to-image diffusion models for generating potentially harmful or copyrighted content. Existing models rely heavily on internet-crawled data, wherein problematic concepts persist due to incomplete filtration processes. While previous approaches somewhat alleviate the issue, they often rely on text-specified concepts, introducing challenges in accurately capturing nuanced concepts and aligning model knowledge with human understandings. In response, we propose a framework named Human Feedback Inversion (HFI), where human feedback on model-generated images is condensed into textual tokens guiding the mitigation or removal of problematic images. The proposed framework can be built upon existing techniques for the same purpose, enhancing their alignment with human judgment. By doing so, we simplify the training objective with a self-distillation-based technique, providing a strong baseline for concept removal. Our experimental results demonstrate our framework significantly reduces objectionable content generation while preserving image quality, contributing to the ethical deployment of AI in the public sphere. Code is available at https://github.com/nannullna/safeguard-hfi." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 459, + 453, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 459, + 453, + 482 + ], + "spans": [ + { + "bbox": [ + 160, + 459, + 453, + 482 + ], + "type": "text", + "content": "Caution: This paper contains discussions and examples related to harmful content, including text and images. Reader discretion is advised." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 491, + 453, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 491, + 453, + 514 + ], + "spans": [ + { + "bbox": [ + 160, + 491, + 453, + 514 + ], + "type": "text", + "content": "Keywords: Fairness, accountability, and transparency " + }, + { + "bbox": [ + 160, + 491, + 453, + 514 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 491, + 453, + 514 + ], + "type": "text", + "content": " Diffusion models " + }, + { + "bbox": [ + 160, + 491, + 453, + 514 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 491, + 453, + 514 + ], + "type": "text", + "content": " Image generation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 533, + 230, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 230, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 230, + 547 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 558, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 668 + ], + "type": "text", + "content": "Large-scale text-to-image generation models have demonstrated remarkable success in producing high-quality images encompassing a wide array of concepts [4, 31,32,36,39]. These models possess the ability to combine disparate ideas in novel and unprecedented manners, making them invaluable tools for creative content creation. However, they also give rise to substantial concerns, including the generation of harmful or copyrighted content, even without explicit directives [40]. A critical issue lies in the heavy reliance of the models on internet-crawled datasets (e.g., LAION-5B [42], currently not available [50]), which, despite attempts at aggressive filtration [48], often lead to problematic content generation [33,40]." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 113, + 482, + 308 + ], + "blocks": [ + { + "bbox": [ + 132, + 113, + 482, + 308 + ], + "lines": [ + { + "bbox": [ + 132, + 113, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 132, + 113, + 482, + 308 + ], + "type": "image", + "image_path": "393aa38b7d0a6a268dc4680498706399ed7c20875575bd8391fdd410d7d9a65f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 316, + 482, + 361 + ], + "lines": [ + { + "bbox": [ + 130, + 316, + 482, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 482, + 361 + ], + "type": "text", + "content": "Fig. 1: Comparative analysis of NSFW content removal techniques, including our proposed fine-tuning method (SDD), both with and without HFI framework. The results clearly show that incorporating HFI significantly reduces the amplification of provocative body parts and ensures the generation of clothed representations." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 387, + 482, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 387, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 387, + 482, + 495 + ], + "type": "text", + "content": "Previous inference-time [3,40] and fine-tuning methods [13,14,52] have made strides in mitigating objectionable imagery. Here, the concept to be removed, or the target concept, is specified by text. However, they insufficiently address the nuanced nature of concepts, leading to a gap between the model-generated content and human expectations. Examples include determining whether exposed body parts in a masterpiece are considered safe or if an illustration overtly portraying such body should be similarly categorized. Moreover, the ethical judgment criteria of the model generally fall short of human standards [33, 40], emphasizing the crucial need for collective intelligence intervention at this juncture." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 496, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 496, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 496, + 482, + 640 + ], + "type": "text", + "content": "To handle such ambiguity and bridge this gap, we present Human Feedback Inversion (HFI), a novel and effective framework to mitigate the ethical concerns surrounding text-to-image diffusion models. HFI incorporates human feedback on model-generated images with problematic concepts directly into soft tokens using the textual inversion (TI) [12]. By directly incorporating human perspectives, HFI addresses the limitations of the previous text-based approaches, offering a more nuanced and accurate means of discerning harmful concepts. This condensed information serves as a guiding force against the target concept during inference time or to create training images for fine-tuning. Through this process, tokens acquired via HFI become specialized for their respective concepts, aiming to enhance both effectiveness (i.e., removal performance of the target concept) and utility aspects (i.e., image quality for remaining concepts)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "To augment the efficacy of HFI, we introduce Safe self-Distillation Diffusion (SDD), a self-distillation-based fine-tuning method. This method simplifies" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "S. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 246, + 184 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 184 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 184 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 184 + ], + "type": "image", + "image_path": "1c9d5d2ac0f1d53dfcdc17500255f1c9852852b6f1ab57b8328a00e9a6a69af9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 169, + 186, + 211, + 195 + ], + "lines": [ + { + "bbox": [ + 169, + 186, + 211, + 195 + ], + "spans": [ + { + "bbox": [ + 169, + 186, + 211, + 195 + ], + "type": "text", + "content": "(a) Flowers" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 250, + 114, + 363, + 184 + ], + "blocks": [ + { + "bbox": [ + 250, + 114, + 363, + 184 + ], + "lines": [ + { + "bbox": [ + 250, + 114, + 363, + 184 + ], + "spans": [ + { + "bbox": [ + 250, + 114, + 363, + 184 + ], + "type": "image", + "image_path": "fa677bffb3fa5a15e9f76ca31deb80f20f73e0921dd458363ec24f2d3d4ede3f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 291, + 186, + 323, + 195 + ], + "lines": [ + { + "bbox": [ + 291, + 186, + 323, + 195 + ], + "spans": [ + { + "bbox": [ + 291, + 186, + 323, + 195 + ], + "type": "text", + "content": "(b) Fish" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 369, + 114, + 480, + 184 + ], + "blocks": [ + { + "bbox": [ + 369, + 114, + 480, + 184 + ], + "lines": [ + { + "bbox": [ + 369, + 114, + 480, + 184 + ], + "spans": [ + { + "bbox": [ + 369, + 114, + 480, + 184 + ], + "type": "image", + "image_path": "ddb9a2e9ebd33459451c72a6b9fa1346878e12e3fc46c4cafb9c5b0370596496.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 409, + 186, + 440, + 195 + ], + "lines": [ + { + "bbox": [ + 409, + 186, + 440, + 195 + ], + "spans": [ + { + "bbox": [ + 409, + 186, + 440, + 195 + ], + "type": "text", + "content": "(c) Boat" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 133, + 199, + 246, + 269 + ], + "blocks": [ + { + "bbox": [ + 133, + 199, + 246, + 269 + ], + "lines": [ + { + "bbox": [ + 133, + 199, + 246, + 269 + ], + "spans": [ + { + "bbox": [ + 133, + 199, + 246, + 269 + ], + "type": "image", + "image_path": "de5cb46be7044453dc2df6fb464e5bb8353ad3f50aff65c4b03047dcd291b2eb.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 175, + 271, + 203, + 281 + ], + "lines": [ + { + "bbox": [ + 175, + 271, + 203, + 281 + ], + "spans": [ + { + "bbox": [ + 175, + 271, + 203, + 281 + ], + "type": "text", + "content": "(d) Hat" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 251, + 199, + 363, + 269 + ], + "blocks": [ + { + "bbox": [ + 251, + 199, + 363, + 269 + ], + "lines": [ + { + "bbox": [ + 251, + 199, + 363, + 269 + ], + "spans": [ + { + "bbox": [ + 251, + 199, + 363, + 269 + ], + "type": "image", + "image_path": "ac85089f823d0aa0bcf217e8c4c926ccc5d8c32f690bf8ef9977f9a41feb8ffc.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 292, + 271, + 323, + 281 + ], + "lines": [ + { + "bbox": [ + 292, + 271, + 323, + 281 + ], + "spans": [ + { + "bbox": [ + 292, + 271, + 323, + 281 + ], + "type": "text", + "content": "(e) Tree" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 369, + 199, + 480, + 269 + ], + "blocks": [ + { + "bbox": [ + 369, + 199, + 480, + 269 + ], + "lines": [ + { + "bbox": [ + 369, + 199, + 480, + 269 + ], + "spans": [ + { + "bbox": [ + 369, + 199, + 480, + 269 + ], + "type": "image", + "image_path": "799d3402bde8d72b57bcd2f8249b3d5108d3a121f7c83cac23c0d23eec520805.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 399, + 271, + 451, + 281 + ], + "lines": [ + { + "bbox": [ + 399, + 271, + 451, + 281 + ], + "spans": [ + { + "bbox": [ + 399, + 271, + 451, + 281 + ], + "type": "text", + "content": "(f) Streetlamp" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 141, + 347, + 252, + 427 + ], + "blocks": [ + { + "bbox": [ + 131, + 290, + 482, + 335 + ], + "lines": [ + { + "bbox": [ + 131, + 290, + 482, + 335 + ], + "spans": [ + { + "bbox": [ + 131, + 290, + 482, + 335 + ], + "type": "text", + "content": "Fig. 2: Artist style removal results on Vincent van Gogh. Images are generated with prompts such as \"a painting of flowers in the style of Vincent van Gogh\". HFI+SDD effectively eliminates his distinct artistic features while preserving the essence of the original subject matter, subject integrity, and visual quality." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 141, + 347, + 252, + 427 + ], + "lines": [ + { + "bbox": [ + 141, + 347, + 252, + 427 + ], + "spans": [ + { + "bbox": [ + 141, + 347, + 252, + 427 + ], + "type": "image", + "image_path": "7b31fc6842a2133a98e6599df08078f8f2cd80efef50af10a437e4192073482b.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 258, + 347, + 362, + 427 + ], + "blocks": [ + { + "bbox": [ + 258, + 347, + 362, + 427 + ], + "lines": [ + { + "bbox": [ + 258, + 347, + 362, + 427 + ], + "spans": [ + { + "bbox": [ + 258, + 347, + 362, + 427 + ], + "type": "image", + "image_path": "8931b5a4b14f87c0a8c4ad3023f774825f7c49c433abe5886fa9907473dd7bac.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 437, + 482, + 471 + ], + "lines": [ + { + "bbox": [ + 131, + 437, + 482, + 471 + ], + "spans": [ + { + "bbox": [ + 131, + 437, + 482, + 471 + ], + "type": "text", + "content": "Fig. 3: Comparison of images generated with famous artwork titles of the three artists. HFI+SDD, leveraging human feedback, captures and removes distinctive artistic styles more effectively while preserving the integrity and essence of the original paintings." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 369, + 347, + 472, + 427 + ], + "blocks": [ + { + "bbox": [ + 369, + 347, + 472, + 427 + ], + "lines": [ + { + "bbox": [ + 369, + 347, + 472, + 427 + ], + "spans": [ + { + "bbox": [ + 369, + 347, + 472, + 427 + ], + "type": "image", + "image_path": "b753e0336562fdaac3f10506b303f6f14b59ea30d5b92bab93e42447be25e1db.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 496, + 479, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 496, + 479, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 496, + 479, + 581 + ], + "type": "text", + "content": "the training objective, eliminating the need for manually designing the target or specifying which concept to retain. With this improved design, SDD outperforms existing baselines, even when based on text-specified concepts, with improved training stability. The integration of HFI and SDD establishes a synergistic alliance, utilizing the HFI token as a removal target. This soft token plays a crucial role in generating latents, serving as a vital component for self-supervision and further contributing to the model's alignment with human understanding." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": "Figs. 1 to 3 showcase the effectiveness of HFI in concept removal, surpassing previous methods and aligning well with human preferences. Specifically, when tasked with removing Van Gogh's artistic style, our pipeline, guided by human feedback, successfully retained the essential qualities of a painting-like style while eliminating his characteristic bold brushstrokes. In contrast, methods without human feedback transformed images into realistic photographs to aggressively remove artist's styles as a side effect. Additionally, when instructed to remove" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "text", + "content": "Safeguard Diffusion Models with Human Feedback Inversion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 164 + ], + "type": "text", + "content": "nudity with human feedback, our pipeline successfully eliminated pornographic content, in contrast to previous methods that still produced provocative images. The visual evidence highlights the superior effectiveness of HFI with SDD in consistently delivering safe and modest image outputs across different prompts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 480, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 480, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 480, + 212 + ], + "type": "text", + "content": "In summary, our framework not only markedly diminishes the generation of undesirable content perceived by humans but also maintains image quality. This offers a pragmatic solution for improving safety and ethical considerations in the widespread deployment of text-to-image diffusion models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 232, + 348, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 232, + 348, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 232, + 348, + 247 + ], + "type": "text", + "content": "2 Background on Diffusion Models" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 259, + 479, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 479, + 367 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 479, + 367 + ], + "type": "text", + "content": "Diffusion models [17, 44-47] are a class of generative models that learn the data distribution by building two Markov chains of latent variables. Given a sample " + }, + { + "bbox": [ + 130, + 259, + 479, + 367 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0 \\sim p_{\\mathrm{data}}(\\mathbf{x}) \\coloneqq q(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 259, + 479, + 367 + ], + "type": "text", + "content": " and a noise schedule " + }, + { + "bbox": [ + 130, + 259, + 479, + 367 + ], + "type": "inline_equation", + "content": "\\{\\beta_t\\}_{t=1}^T" + }, + { + "bbox": [ + 130, + 259, + 479, + 367 + ], + "type": "text", + "content": ", the forward process gradually injects a series of Gaussian noises to the sample until it nearly follows the standard Gaussian " + }, + { + "bbox": [ + 130, + 259, + 479, + 367 + ], + "type": "inline_equation", + "content": "q(\\mathbf{x}_t | \\mathbf{x}_{t-1}) \\coloneqq \\mathcal{N}(\\mathbf{x}_t; \\sqrt{1 - \\beta_t} \\mathbf{x}_{t-1}, \\beta_t \\mathbf{I})" + }, + { + "bbox": [ + 130, + 259, + 479, + 367 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 259, + 479, + 367 + ], + "type": "inline_equation", + "content": "q(\\mathbf{x}_T | \\mathbf{x}_0) \\approx \\mathcal{N}(\\mathbf{x}_T; \\mathbf{0}, \\mathbf{I})" + }, + { + "bbox": [ + 130, + 259, + 479, + 367 + ], + "type": "text", + "content": ". Such process is then followed by the reverse process parameterized by " + }, + { + "bbox": [ + 130, + 259, + 479, + 367 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 259, + 479, + 367 + ], + "type": "text", + "content": ", where the model learns to reconstruct the original image by iteratively denoising from a pure Gaussian noise " + }, + { + "bbox": [ + 130, + 259, + 479, + 367 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_T) = \\mathcal{N}(\\mathbf{0}, \\mathbf{I})" + }, + { + "bbox": [ + 130, + 259, + 479, + 367 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 259, + 479, + 367 + ], + "type": "inline_equation", + "content": "p_\\theta(\\mathbf{x}_{0:T}) = p(\\mathbf{x}_T) \\prod_{t=1}^T p_\\theta(\\mathbf{x}_{t-1} | \\mathbf{x}_t)" + }, + { + "bbox": [ + 130, + 259, + 479, + 367 + ], + "type": "text", + "content": ". Ho et al. [17] simplifies the objective to learn a noise estimator " + }, + { + "bbox": [ + 130, + 259, + 479, + 367 + ], + "type": "inline_equation", + "content": "\\epsilon_\\theta" + }, + { + "bbox": [ + 130, + 259, + 479, + 367 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 194, + 378, + 480, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 378, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 194, + 378, + 480, + 392 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {D M} (\\theta) = \\mathbb {E} _ {\\mathbf {x} _ {0}, \\epsilon , t} \\left[ \\| \\boldsymbol {\\epsilon} - \\boldsymbol {\\epsilon} _ {\\theta} \\left(\\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {x} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\boldsymbol {\\epsilon}; t\\right) \\| _ {2} ^ {2} \\right], \\tag {1}", + "image_path": "907d0ac4e4b3ee09046726b61a38f8f95451c15d1292b76b8716f43d5d7ac482.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 401, + 479, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 401, + 479, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 479, + 426 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 401, + 479, + 426 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 130, + 401, + 479, + 426 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 401, + 479, + 426 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t = \\prod_{s=1}^t (1 - \\beta_s)" + }, + { + "bbox": [ + 130, + 401, + 479, + 426 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 401, + 479, + 426 + ], + "type": "inline_equation", + "content": "t \\sim \\mathcal{U}([1,T])" + }, + { + "bbox": [ + 130, + 401, + 479, + 426 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 401, + 479, + 426 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 401, + 479, + 426 + ], + "type": "text", + "content": " is the total number of steps. For simple notation, we omit the timestep " + }, + { + "bbox": [ + 130, + 401, + 479, + 426 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 401, + 479, + 426 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 401, + 479, + 426 + ], + "type": "inline_equation", + "content": "\\epsilon(\\mathbf{x}_t; t) = \\epsilon(\\mathbf{x}_t)" + }, + { + "bbox": [ + 130, + 401, + 479, + 426 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 427, + 480, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 427, + 480, + 571 + ], + "spans": [ + { + "bbox": [ + 130, + 427, + 480, + 571 + ], + "type": "text", + "content": "Latent Diffusion Models (LDMs) [36] leverage the diffusion process within the latent space rather than in the pixel space, utilizing a pre-trained autoencoder. By mapping the input " + }, + { + "bbox": [ + 130, + 427, + 480, + 571 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 427, + 480, + 571 + ], + "type": "text", + "content": " into a latent space with the encoder " + }, + { + "bbox": [ + 130, + 427, + 480, + 571 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 130, + 427, + 480, + 571 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 427, + 480, + 571 + ], + "type": "inline_equation", + "content": "\\mathbf{z} = \\mathcal{E}(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 427, + 480, + 571 + ], + "type": "text", + "content": ", an LDM is trained to predict the added noise in the latent space. As this paper mainly discusses LDMs, we use " + }, + { + "bbox": [ + 130, + 427, + 480, + 571 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 130, + 427, + 480, + 571 + ], + "type": "text", + "content": " instead of " + }, + { + "bbox": [ + 130, + 427, + 480, + 571 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 130, + 427, + 480, + 571 + ], + "type": "text", + "content": ". Text-to-image models additionally take inputs of the text embedding, " + }, + { + "bbox": [ + 130, + 427, + 480, + 571 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_p = \\mathcal{E}_{\\mathrm{txt}}(c_p)" + }, + { + "bbox": [ + 130, + 427, + 480, + 571 + ], + "type": "text", + "content": ", paired with an image " + }, + { + "bbox": [ + 130, + 427, + 480, + 571 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 427, + 480, + 571 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 427, + 480, + 571 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{txt}}" + }, + { + "bbox": [ + 130, + 427, + 480, + 571 + ], + "type": "text", + "content": " is the text encoder and " + }, + { + "bbox": [ + 130, + 427, + 480, + 571 + ], + "type": "inline_equation", + "content": "c_{p}" + }, + { + "bbox": [ + 130, + 427, + 480, + 571 + ], + "type": "text", + "content": " is a text prompt. To enhance the quality of text conditioning, the classifier-free guidance (CFG) models both conditional and unconditional noise estimates by randomly replacing " + }, + { + "bbox": [ + 130, + 427, + 480, + 571 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_p" + }, + { + "bbox": [ + 130, + 427, + 480, + 571 + ], + "type": "text", + "content": " with the embedding of an empty string " + }, + { + "bbox": [ + 130, + 427, + 480, + 571 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{\\emptyset} = \\mathcal{E}_{\\mathrm{txt}}(\"\")" + }, + { + "bbox": [ + 130, + 427, + 480, + 571 + ], + "type": "text", + "content": " and modifies the noise estimate during sampling as " + }, + { + "bbox": [ + 130, + 427, + 480, + 571 + ], + "type": "inline_equation", + "content": "\\tilde{\\epsilon}_{\\mathrm{CFG}} = \\epsilon_{\\theta}(\\mathbf{x}_t) + (w + 1)[\\epsilon_{\\theta}(\\mathbf{x}_t;\\mathbf{c}_p) - \\epsilon_{\\theta}(\\mathbf{x}_t;\\mathbf{c}_{\\emptyset})]" + }, + { + "bbox": [ + 130, + 427, + 480, + 571 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 427, + 480, + 571 + ], + "type": "inline_equation", + "content": "w + 1" + }, + { + "bbox": [ + 130, + 427, + 480, + 571 + ], + "type": "text", + "content": " is the CFG scale. We simply denote " + }, + { + "bbox": [ + 130, + 427, + 480, + 571 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\mathbf{x}_t;\\mathbf{c}_{\\emptyset})" + }, + { + "bbox": [ + 130, + 427, + 480, + 571 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 427, + 480, + 571 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\mathbf{x}_t)" + }, + { + "bbox": [ + 130, + 427, + 480, + 571 + ], + "type": "text", + "content": " in the subsequent sections." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 590, + 202, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 202, + 602 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 202, + 602 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 617, + 479, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 479, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 479, + 666 + ], + "type": "text", + "content": "We present Human Feedback Inversion (HFI), an innovative framework for concept removal guided by human feedback. Unlike previous methods that rely on predefined text-based concepts, i.e., a specified target word like \"nudity\", real-world inappropriate concepts are often nuanced and challenging to express in" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "S. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 113, + 289, + 210 + ], + "blocks": [ + { + "bbox": [ + 132, + 113, + 289, + 210 + ], + "lines": [ + { + "bbox": [ + 132, + 113, + 289, + 210 + ], + "spans": [ + { + "bbox": [ + 132, + 113, + 289, + 210 + ], + "type": "image", + "image_path": "b61c306d2a01ff31ee4f2d0086fe77c770bb3d429a6990b959a880822f4d53fa.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 145, + 212, + 277, + 221 + ], + "lines": [ + { + "bbox": [ + 145, + 212, + 277, + 221 + ], + "spans": [ + { + "bbox": [ + 145, + 212, + 277, + 221 + ], + "type": "text", + "content": "(a) Human Feedback Inversion (HFI)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 304, + 129, + 482, + 210 + ], + "blocks": [ + { + "bbox": [ + 304, + 129, + 482, + 210 + ], + "lines": [ + { + "bbox": [ + 304, + 129, + 482, + 210 + ], + "spans": [ + { + "bbox": [ + 304, + 129, + 482, + 210 + ], + "type": "image", + "image_path": "2812881bf105aebaee2ce1f6f5a5f41e8397976f4f169c5f526f92e8fb44007f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 212, + 466, + 221 + ], + "lines": [ + { + "bbox": [ + 320, + 212, + 466, + 221 + ], + "spans": [ + { + "bbox": [ + 320, + 212, + 466, + 221 + ], + "type": "text", + "content": "(b) Safe self-Distillation Diffusion (SDD)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 178, + 230, + 433, + 242 + ], + "lines": [ + { + "bbox": [ + 178, + 230, + 433, + 242 + ], + "spans": [ + { + "bbox": [ + 178, + 230, + 433, + 242 + ], + "type": "text", + "content": "Fig. 4: Illustration of our proposed framework, HFI and SDD." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 265, + 482, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 265, + 482, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 265, + 482, + 384 + ], + "type": "text", + "content": "text. HFI addresses this challenge by learning concepts based on collective understanding. Our approach initiates with collecting human feedback on model-generated images through carefully designed surveys detailed in Appendix C, where harmfulness is comparatively evaluated. Using this human evaluation data, we construct a reward model that assigns a harmfulness score to each image. With this reward model, we formulate an optimization problem to identify a soft token that accurately represents harmfulness. These optimal embedding vectors are then utilized to fine-tune the diffusion model, preventing the generation of harmful images. The overall framework is depicted in Fig. 4 and Appendix A, and detailed explanations of each stage are provided in the subsequent sections." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 401, + 375, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 401, + 375, + 414 + ], + "spans": [ + { + "bbox": [ + 131, + 401, + 375, + 414 + ], + "type": "text", + "content": "3.1 Collecting and Modeling Human Feedback" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 420, + 482, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 420, + 482, + 563 + ], + "spans": [ + { + "bbox": [ + 130, + 420, + 482, + 563 + ], + "type": "text", + "content": "We first generate images, on which human feedback will be collected, using the original model and a set of simple prompts containing the target concept word " + }, + { + "bbox": [ + 130, + 420, + 482, + 563 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 420, + 482, + 563 + ], + "type": "text", + "content": " detailed in Appendix B. Subsequently, two main settings can be considered for gathering human feedback. The design of how and what kind of feedback to collect during this process is crucial. In cases where the content the model should not generate is well-defined and clear annotation guidelines exist (e.g., nudity), binary feedback for each image can be collected. However, this does not apply to all scenarios, with a notable example being an artist's style. Suppose that an artist requests the removal of not only specific works but also the overall style from the model, then the criteria may not be clear. In such cases, ranking feedback can be collected. This involves presenting annotators with " + }, + { + "bbox": [ + 130, + 420, + 482, + 563 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 420, + 482, + 563 + ], + "type": "text", + "content": " random images and requesting them to rank these images ranging from 1 to " + }, + { + "bbox": [ + 130, + 420, + 482, + 563 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 420, + 482, + 563 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 564, + 481, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 564, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 564, + 481, + 601 + ], + "type": "text", + "content": "Following this, we need to train a reward model " + }, + { + "bbox": [ + 130, + 564, + 481, + 601 + ], + "type": "inline_equation", + "content": "r_{\\psi}" + }, + { + "bbox": [ + 130, + 564, + 481, + 601 + ], + "type": "text", + "content": ", which accepts an input image " + }, + { + "bbox": [ + 130, + 564, + 481, + 601 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 564, + 481, + 601 + ], + "type": "text", + "content": " and outputs a score " + }, + { + "bbox": [ + 130, + 564, + 481, + 601 + ], + "type": "inline_equation", + "content": "r_{\\psi}(\\mathbf{x}) \\in \\mathbb{R}" + }, + { + "bbox": [ + 130, + 564, + 481, + 601 + ], + "type": "text", + "content": " for the concept " + }, + { + "bbox": [ + 130, + 564, + 481, + 601 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 564, + 481, + 601 + ], + "type": "text", + "content": ". For binary feedback, we follow the MSE-based approach proposed in Lee et al. [20]," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 214, + 607, + 481, + 623 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 607, + 481, + 623 + ], + "spans": [ + { + "bbox": [ + 214, + 607, + 481, + 623 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {M S E} (\\psi) = \\mathbb {E} _ {(\\mathbf {x}, y) \\sim \\mathcal {D} _ {\\text {h u m a n}}} \\left[ (r _ {\\psi} (\\mathbf {x}) - y) ^ {2} \\right], \\tag {2}", + "image_path": "18c1b399bd6c2574f729ee9a72d02ed732c68e36696b43860ea0d2fb8b13e4af.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 629, + 481, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 629, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{human}}" + }, + { + "bbox": [ + 131, + 629, + 481, + 666 + ], + "type": "text", + "content": " consists of images " + }, + { + "bbox": [ + 131, + 629, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 131, + 629, + 481, + 666 + ], + "type": "text", + "content": " and corresponding binary annotations " + }, + { + "bbox": [ + 131, + 629, + 481, + 666 + ], + "type": "inline_equation", + "content": "y \\in \\{0,1\\}" + }, + { + "bbox": [ + 131, + 629, + 481, + 666 + ], + "type": "text", + "content": ". Here, '1' indicates the presence of the harmful concept (e.g., with nudity), and '0' is the absence (e.g., without nudity), answered by human annotators." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "text", + "content": "Safeguard Diffusion Models with Human Feedback Inversion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 202 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 202 + ], + "type": "text", + "content": "On the other hand, for ranking feedback, we adopted the Bradley-Terry model [5], which assigns scores to examples based on pairwise comparisons [8, 26, 49]. The obtained rankings of " + }, + { + "bbox": [ + 130, + 116, + 482, + 202 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 116, + 482, + 202 + ], + "type": "text", + "content": " images can be reorganized to construct " + }, + { + "bbox": [ + 130, + 116, + 482, + 202 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{human}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 202 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 116, + 482, + 202 + ], + "type": "inline_equation", + "content": "\\binom{M}{2}" + }, + { + "bbox": [ + 130, + 116, + 482, + 202 + ], + "type": "text", + "content": " pairwise comparisons as " + }, + { + "bbox": [ + 130, + 116, + 482, + 202 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{+} = \\mathbf{x}^{(i)}, \\mathbf{x}^{-} = \\mathbf{x}^{(j)}" + }, + { + "bbox": [ + 130, + 116, + 482, + 202 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 130, + 116, + 482, + 202 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{(i)}" + }, + { + "bbox": [ + 130, + 116, + 482, + 202 + ], + "type": "text", + "content": " has a higher rank than " + }, + { + "bbox": [ + 130, + 116, + 482, + 202 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{(j)}" + }, + { + "bbox": [ + 130, + 116, + 482, + 202 + ], + "type": "text", + "content": ", and vice versa - a higher ranking indicates a greater presence or resemblance of the concept. Subsequently, we include all " + }, + { + "bbox": [ + 130, + 116, + 482, + 202 + ], + "type": "inline_equation", + "content": "\\binom{M}{2}" + }, + { + "bbox": [ + 130, + 116, + 482, + 202 + ], + "type": "text", + "content": " pairs in a single batch, following Ouyang et al. [26], and optimize the reward model as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 180, + 210, + 481, + 239 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 180, + 210, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 180, + 210, + 481, + 239 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {N L L} (\\psi) = - \\frac {1}{\\binom {M} {2}} \\mathbb {E} _ {(\\mathbf {x} ^ {+}, \\mathbf {x} ^ {-}) \\sim \\mathcal {D} _ {\\text {h u m a n}}} \\log \\sigma \\left(r _ {\\psi} \\left(\\mathbf {x} ^ {+}\\right) - r _ {\\psi} \\left(\\mathbf {x} ^ {-}\\right)\\right), \\tag {3}", + "image_path": "ab74e4b5a692f5bdbd7aaaee922761004d023ae82ad0f6e8f5d83f2479d9695f.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 246, + 481, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 246, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 246, + 481, + 282 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 246, + 481, + 282 + ], + "type": "inline_equation", + "content": "\\sigma (\\cdot)" + }, + { + "bbox": [ + 130, + 246, + 481, + 282 + ], + "type": "text", + "content": " is the sigmoid function. For both approaches, the reward model " + }, + { + "bbox": [ + 130, + 246, + 481, + 282 + ], + "type": "inline_equation", + "content": "r_{\\psi}(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 246, + 481, + 282 + ], + "type": "text", + "content": " maps the CLIP [30] embedding of an image " + }, + { + "bbox": [ + 130, + 246, + 481, + 282 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 246, + 481, + 282 + ], + "type": "text", + "content": " to a score through a couple of MLP layers with ReLU non-linearities." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 300, + 346, + 313 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 346, + 313 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 346, + 313 + ], + "type": "text", + "content": "3.2 Inverting Feedback into Embeddings" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 320, + 482, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 320, + 482, + 417 + ], + "spans": [ + { + "bbox": [ + 130, + 320, + 482, + 417 + ], + "type": "text", + "content": "Based on the feedback received in Sec. 3.1, we transform the corresponding concept into a soft token. Inspired by recent inversion techniques [12,35], we utilize TI [12], as it can easily replace or be combined with text tokens from previous approaches. The goal is to find a soft token " + }, + { + "bbox": [ + 130, + 320, + 482, + 417 + ], + "type": "inline_equation", + "content": "v^{*}" + }, + { + "bbox": [ + 130, + 320, + 482, + 417 + ], + "type": "text", + "content": " that maximizes the expected reward of generated images, i.e., " + }, + { + "bbox": [ + 130, + 320, + 482, + 417 + ], + "type": "inline_equation", + "content": "v^{*} = \\arg \\max_{v}\\mathbb{E}_{\\mathbf{x}\\sim p_{\\theta}(\\mathbf{x}|v)}[r_{\\psi}(\\mathbf{x})]" + }, + { + "bbox": [ + 130, + 320, + 482, + 417 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 320, + 482, + 417 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 130, + 320, + 482, + 417 + ], + "type": "text", + "content": " is the soft word in the token space. Utilizing the reward model " + }, + { + "bbox": [ + 130, + 320, + 482, + 417 + ], + "type": "inline_equation", + "content": "r_{\\psi}" + }, + { + "bbox": [ + 130, + 320, + 482, + 417 + ], + "type": "text", + "content": " in Sec. 3.1, we optimize the following objective of the reward-weighted negative log-likelihood [20] to get the optimal token " + }, + { + "bbox": [ + 130, + 320, + 482, + 417 + ], + "type": "inline_equation", + "content": "v^{*}" + }, + { + "bbox": [ + 130, + 320, + 482, + 417 + ], + "type": "text", + "content": " that best captures the text concept " + }, + { + "bbox": [ + 130, + 320, + 482, + 417 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 320, + 482, + 417 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 213, + 426, + 481, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 213, + 426, + 481, + 444 + ], + "spans": [ + { + "bbox": [ + 213, + 426, + 481, + 444 + ], + "type": "interline_equation", + "content": "v ^ {*} = \\underset {v} {\\arg \\min } \\mathbb {E} _ {\\mathbf {x} \\sim p _ {\\theta} (\\mathbf {x} | c)} [ - r _ {\\psi} (\\mathbf {x}) \\log p _ {\\theta} (\\mathbf {x} | v) ]. \\tag {4}", + "image_path": "599769ea1fe20d4ede7e0f934d4f40741a03167d61b4f812f535a2b2e5759664.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 453, + 482, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 453, + 482, + 538 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 482, + 538 + ], + "type": "text", + "content": "Inverting too many images may struggle with concept extraction by prioritizing image reconstruction over it, as shown in Gal et al. [12]. Also, we empirically found that inverting it with a small learning rate may fail to capture the concept, which requires additional hyperparameter tuning. To simplify such a burden, we use a reward-weighted sampler with images of the top- " + }, + { + "bbox": [ + 130, + 453, + 482, + 538 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 453, + 482, + 538 + ], + "type": "text", + "content": " scores, which is equivalent to the reward-weighted loss in Eq. (4). We use the learning rate of " + }, + { + "bbox": [ + 130, + 453, + 482, + 538 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-4}" + }, + { + "bbox": [ + 130, + 453, + 482, + 538 + ], + "type": "text", + "content": ", the default value in the HuggingFace's implementation1, and " + }, + { + "bbox": [ + 130, + 453, + 482, + 538 + ], + "type": "inline_equation", + "content": "K = 50" + }, + { + "bbox": [ + 130, + 453, + 482, + 538 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 555, + 413, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 555, + 413, + 567 + ], + "spans": [ + { + "bbox": [ + 131, + 555, + 413, + 567 + ], + "type": "text", + "content": "3.3 Removing Learned Concepts with Self-Distillation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 574, + 482, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 574, + 482, + 636 + ], + "spans": [ + { + "bbox": [ + 130, + 574, + 482, + 636 + ], + "type": "text", + "content": "After obtaining soft tokens through the proposed HFI method in Sec. 3.2, one can adapt them to existing text-based concept removal methods, enabling the desired removal of concepts using the obtained soft tokens instead of text tokens. While HFI alone proves effectiveness in Tabs. 1 and 2, we propose a new method termed Safe self-Distillation Diffusion (SDD) to achieve an even greater impact," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "S. Kim et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 642, + 481, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 481, + 666 + ], + "type": "text", + "content": "1 https://github.com/huggingface/diffusers/blob/main/examples/textual_inversion/textual_inversion.py" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "particularly in enhancing the removal of ambiguous concepts judged by humans. SDD removes the target concept from a model based on self-distillation [53]. We first set both student " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " and teacher models " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\theta^{\\star}" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " whose parameters are initialized from the pre-trained model and update the student with the following loss:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 197, + 172, + 480, + 186 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 172, + 480, + 186 + ], + "spans": [ + { + "bbox": [ + 197, + 172, + 480, + 186 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {S D D} = \\mathbb {E} _ {\\mathbf {x} _ {t} \\sim p _ {\\theta^ {*}} (\\mathbf {x} _ {t} | v ^ {*}), t} \\left[ \\| \\epsilon_ {\\theta} (\\mathbf {x} _ {t}; v ^ {*}) - \\mathbf {s g} \\left(\\epsilon_ {\\theta} (\\mathbf {x} _ {t})\\right) \\| _ {2} ^ {2} \\right], \\tag {5}", + "image_path": "a1439201af0442c25c82d6c588b35cc83e3acaf0f073b1500aeab83085dcd6fc.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 193, + 480, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 193, + 480, + 263 + ], + "spans": [ + { + "bbox": [ + 130, + 193, + 480, + 263 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 193, + 480, + 263 + ], + "type": "inline_equation", + "content": "\\mathfrak{sg}()" + }, + { + "bbox": [ + 130, + 193, + 480, + 263 + ], + "type": "text", + "content": " denotes the stop-grad operation. That is, we want the model given the token " + }, + { + "bbox": [ + 130, + 193, + 480, + 263 + ], + "type": "inline_equation", + "content": "v^{*}" + }, + { + "bbox": [ + 130, + 193, + 480, + 263 + ], + "type": "text", + "content": " to behave as if it were not for any token by setting the unconditional estimate " + }, + { + "bbox": [ + 130, + 193, + 480, + 263 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\mathbf{x}_t)" + }, + { + "bbox": [ + 130, + 193, + 480, + 263 + ], + "type": "text", + "content": " as a self-distillation target. After updating the student model, we update the teacher model " + }, + { + "bbox": [ + 130, + 193, + 480, + 263 + ], + "type": "inline_equation", + "content": "\\theta^{\\star}" + }, + { + "bbox": [ + 130, + 193, + 480, + 263 + ], + "type": "text", + "content": " as an exponential moving average (EMA) from it, i.e., " + }, + { + "bbox": [ + 130, + 193, + 480, + 263 + ], + "type": "inline_equation", + "content": "\\theta^{\\star} \\gets m\\theta^{\\star} + (1 - m)\\theta" + }, + { + "bbox": [ + 130, + 193, + 480, + 263 + ], + "type": "text", + "content": ", with a decaying rate " + }, + { + "bbox": [ + 130, + 193, + 480, + 263 + ], + "type": "inline_equation", + "content": "m \\in (0,1)" + }, + { + "bbox": [ + 130, + 193, + 480, + 263 + ], + "type": "text", + "content": ". After the training, we use the EMA parameter of the teacher model as our final safe model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 265, + 480, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 265, + 480, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 265, + 480, + 361 + ], + "type": "text", + "content": "Unlike methods focused on learning new concepts [19, 38], determining suitable replacements when removing specific concepts can be challenging, often contingent on other contextual elements [9]. For instance, while removing the color red in a red apple can be substituted with a green apple, removing the color red devoid of context lacks clarity. Similarly, removing concepts like nudity, violence, or specific artists presents this ambiguity. Previous approaches [3,13,40] have attempted to tackle this challenge by targeting negative guidance; however, images generated in this manner often fall short of removing such concepts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 361, + 481, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 481, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 481, + 491 + ], + "type": "text", + "content": "We tackle this problem by introducing the EMA teacher model, which moves in tandem with the student model but in a more stable and deliberate manner, gradually removing the target concept. This serves not only to enhance training stability but also plays a crucial role in providing the student model with appropriate targets for conditional noise predictions. Notably, in Eq. (5), the latent " + }, + { + "bbox": [ + 130, + 361, + 481, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 130, + 361, + 481, + 491 + ], + "type": "text", + "content": " is sampled from the teacher model, i.e. it generates parts that have not been effectively removed up to the existing training step. Additionally, resolving discrepancies between the model's understanding and human judgment regarding harmfulness poses a non-trivial task. In SDD with HFI, sampling " + }, + { + "bbox": [ + 130, + 361, + 481, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 130, + 361, + 481, + 491 + ], + "type": "text", + "content": " is conditioned on the learned token " + }, + { + "bbox": [ + 130, + 361, + 481, + 491 + ], + "type": "inline_equation", + "content": "v^*" + }, + { + "bbox": [ + 130, + 361, + 481, + 491 + ], + "type": "text", + "content": ". During this process, the model learns to remove the concept from newly generated latents that would be deemed problematic." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 492, + 481, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 492, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 130, + 492, + 481, + 588 + ], + "type": "text", + "content": "Lastly, we introduce a couple of additional techniques for fine-tuning. First, similar to existing work [13, 19, 51, 52], we fine-tune only the cross-attention layers of the U-Net [37]. Second, we empirically find that not all timesteps equally influence the conditioning of image generation; the " + }, + { + "bbox": [ + 130, + 492, + 481, + 588 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 130, + 492, + 481, + 588 + ], + "type": "text", + "content": "-norm between the unconditional and the conditional noise was generally high in the middle timesteps, as shown in Appendix E. Therefore, we sample the timestep to apply distillation as " + }, + { + "bbox": [ + 130, + 492, + 481, + 588 + ], + "type": "inline_equation", + "content": "t = T \\cdot \\mathcal{B}(\\alpha, \\beta)" + }, + { + "bbox": [ + 130, + 492, + 481, + 588 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 492, + 481, + 588 + ], + "type": "inline_equation", + "content": "\\mathcal{B}(\\alpha, \\beta)" + }, + { + "bbox": [ + 130, + 492, + 481, + 588 + ], + "type": "text", + "content": " is a beta random variable with parameter " + }, + { + "bbox": [ + 130, + 492, + 481, + 588 + ], + "type": "inline_equation", + "content": "(\\alpha, \\beta)" + }, + { + "bbox": [ + 130, + 492, + 481, + 588 + ], + "type": "text", + "content": ". We set " + }, + { + "bbox": [ + 130, + 492, + 481, + 588 + ], + "type": "inline_equation", + "content": "(\\alpha, \\beta) = (3, 3)" + }, + { + "bbox": [ + 130, + 492, + 481, + 588 + ], + "type": "text", + "content": " for our experiments as it works well in general." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 605, + 237, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 605, + 237, + 618 + ], + "spans": [ + { + "bbox": [ + 132, + 605, + 237, + 618 + ], + "type": "text", + "content": "4 Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "content": "Numerous attempts have been made to address harmful content generation, including the use of external classifiers for image filtering during inference, such as the safety checker in Stable Diffusion (SD) [36]. However, these approaches" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "text", + "content": "Safeguard Diffusion Models with Human Feedback Inversion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "can be easily circumvented by users [33]. Efforts to filter out problematic training images during pre-processing [48] often lead to reduced training sets and inferior image quality [24, 28], without completely eliminating harmfulness [1]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 481, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 481, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 481, + 297 + ], + "type": "text", + "content": "Recent advancements in the field can be broadly classified into three categories: (i) inference-time techniques manipulate noise estimates to counteract concepts [3,40]; (ii) direct cross-attention optimization adjusts weights to disregard problematic tokens from text inputs either directly or through gradient descent [14,52]; (iii) fine-tuning methods refine a subset of parameters using generated images [13,18,23]. Although both (ii) and (iii) fine-tune the model, the main disparity lies in the training objective. The former adjusts the projection or attention of text inputs, while the latter alters the model's distribution along the diffusion trajectories. Moreover, while the former does not necessitate training images or uses only a few real images, the latter generates training images from the model. Nonetheless, all methods entail a trade-off between effectiveness and utility, given the inherent challenge of identifying concepts in texts for removal." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 297, + 482, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 482, + 477 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 482, + 477 + ], + "type": "text", + "content": "Human feedback has become crucial in enhancing both performance and safety in machine learning models [8]. Initially applied in dialogue systems [15, 21], reinforcement learning (RL) algorithms, particularly Proximal Policy Optimization (PPO) [43], have played a significant role in leveraging human feedback, notably in Large Language Models (LLMs) [6, 26, 49], focusing on both performance and safety considerations. Reinforcement learning with human feedback (RLHF) [26] optimizes models based on reward signals derived from user input, further enhancing the safety and adaptability of LLMs. Beyond language domains, efforts have been made to align text-to-image diffusion models with human preferences [2, 10, 20]. These involve fine-tuning using a reward-weighted loss function with appropriate regularization. However, the practical application of such RL-based methods often faces challenges due to hyperparameter sensitivity and reward design intricacies [20]. Furthermore, while such methods relying on human preference data aim to enhance image quality perceived by humans, they differ from our goal of suppressing image generation of specific concepts." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 496, + 230, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 496, + 230, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 496, + 230, + 511 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 523, + 338, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 338, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 338, + 536 + ], + "type": "text", + "content": "5.1 Baselines and Evaluation Protocols" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "We compare the performance of our method with the original SD and previous methods. In Tabs. 1 and 2, SD [36] indicates the original Stable Diffusion model, and NEG is the simplest baseline of SD with the negative prompt, i.e., the noise conditioned on the target concept is used instead of the unconditional one, thus negating the target concept. We consider two inference-time techniques: SLD [40] and SEGA [3]. For fine-tuning methods, we consider the following three methods: ESD [13], FMN [52], and UCE [14]. For all fine-tuning methods, including ours, we use the artist name (e.g., \"Vincent van Gogh\") or the word (e.g., \"nudity\" and \"bleeding\") for the target concept in text. When applying HFI to these methods, HFI token " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "v^{*}" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " replaces the text word. Refer to Appendix B for details." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 102 + ], + "type": "text", + "content": "S. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 113, + 305, + 175 + ], + "blocks": [ + { + "bbox": [ + 134, + 113, + 305, + 175 + ], + "lines": [ + { + "bbox": [ + 134, + 113, + 305, + 175 + ], + "spans": [ + { + "bbox": [ + 134, + 113, + 305, + 175 + ], + "type": "image", + "image_path": "4cb55fd4c0f1b50c96b5948cc533111c5b90cb78975fdd94a41a5da3d2caf696.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 309, + 113, + 482, + 175 + ], + "blocks": [ + { + "bbox": [ + 309, + 113, + 482, + 175 + ], + "lines": [ + { + "bbox": [ + 309, + 113, + 482, + 175 + ], + "spans": [ + { + "bbox": [ + 309, + 113, + 482, + 175 + ], + "type": "image", + "image_path": "51e50d19be2ced3e33c7060d973470f3f6198ab92868c5dfb4701f54c4c79ed6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 133, + 177, + 304, + 239 + ], + "blocks": [ + { + "bbox": [ + 133, + 177, + 304, + 239 + ], + "lines": [ + { + "bbox": [ + 133, + 177, + 304, + 239 + ], + "spans": [ + { + "bbox": [ + 133, + 177, + 304, + 239 + ], + "type": "image", + "image_path": "df0bc3b376c3935e6b8ef6184a63483a506a0c0014291a222d277cd611047847.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 310, + 178, + 481, + 239 + ], + "blocks": [ + { + "bbox": [ + 310, + 178, + 481, + 239 + ], + "lines": [ + { + "bbox": [ + 310, + 178, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 310, + 178, + 481, + 239 + ], + "type": "image", + "image_path": "93f195f44fb1da8524ce35717dc02ec13bd1e5aaad717d4ba99fde5183660aa0.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 133, + 240, + 304, + 316 + ], + "blocks": [ + { + "bbox": [ + 133, + 240, + 304, + 316 + ], + "lines": [ + { + "bbox": [ + 133, + 240, + 304, + 316 + ], + "spans": [ + { + "bbox": [ + 133, + 240, + 304, + 316 + ], + "type": "image", + "image_path": "50e39ff7b275d238ec0d7dd339d3c888306698137e6c54dfeb4625909fdc53b6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 324, + 482, + 369 + ], + "lines": [ + { + "bbox": [ + 130, + 324, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 324, + 482, + 369 + ], + "type": "text", + "content": "Fig. 5: Comparison of concept removal methods for Vincent van Gogh's artistic style. Van Gogh's style is marked by bold brushstrokes, vivid colors, and swirling patterns. HFI+SDD is the most effective, consistently producing the most neutral but still artwork-like images and successfully eliminating his stylistic characteristics." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 310, + 240, + 481, + 316 + ], + "blocks": [ + { + "bbox": [ + 310, + 240, + 481, + 316 + ], + "lines": [ + { + "bbox": [ + 310, + 240, + 481, + 316 + ], + "spans": [ + { + "bbox": [ + 310, + 240, + 481, + 316 + ], + "type": "image", + "image_path": "455f6974bab61299aa72750055e858d05c3322c0513cb82258101441fe5dc651.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 397, + 480, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 397, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 397, + 480, + 445 + ], + "type": "text", + "content": "The evaluation of the method's capability to remove specific concepts is structured along two primary axes: effectiveness and utility. Effectiveness measures the ability to avoid generating the target concept, while utility assesses the preservation of image quality in outputs unrelated to the target concept." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 448, + 482, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 448, + 482, + 639 + ], + "spans": [ + { + "bbox": [ + 130, + 448, + 482, + 639 + ], + "type": "text", + "content": "For artist concept removal in Sec. 5.2, we propose two specific metrics to gauge both aspects. The effectiveness is evaluated by the target artist ratio (%ART) using CLIP zero-shot classification. This involves generating 1,000 images in total of 50 commonly depicted objects in famous artworks (listed by ChatGPT [25]), with the prompt template \"a painting of {object} in the style of {artist}\" These images are then classified using OpenCLIP [7] to determine how frequently the target artist is accurately identified. The utility is determined through the photo similarity ratio (%PHO), which quantifies how well the model retains the ability to generate paintings rather than photorealistic images, thus preserving artistic styles not specific to the targeted artist. We enhance classification accuracy using prompt ensemble [30]. In line with previous work, we also report LPIPS [54] for both the target and other artists. Finally, since such quantitative metrics cannot completely replace qualitative evaluations and human evaluations can be subjective, we asked ChatGPT 4o [25] to assess the performance of removing the specified artist to supplement this. The qualitative evaluation results and further details are provided in Appendix F." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "For harmful concept removal in Sec. 5.3, the effectiveness is measured by generating 5,000 images in total with the prompt {\"country\" body\" using" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "text", + "content": "Safeguard Diffusion Models with Human Feedback Inversion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 112, + 306, + 176 + ], + "blocks": [ + { + "bbox": [ + 133, + 112, + 306, + 176 + ], + "lines": [ + { + "bbox": [ + 133, + 112, + 306, + 176 + ], + "spans": [ + { + "bbox": [ + 133, + 112, + 306, + 176 + ], + "type": "image", + "image_path": "4af15efc7928b63f0e1c720f94b73bc365e7da24b9c13e7585dc5652a4d61c15.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 309, + 113, + 482, + 175 + ], + "blocks": [ + { + "bbox": [ + 309, + 113, + 482, + 175 + ], + "lines": [ + { + "bbox": [ + 309, + 113, + 482, + 175 + ], + "spans": [ + { + "bbox": [ + 309, + 113, + 482, + 175 + ], + "type": "image", + "image_path": "7520733b457100a65231047a2a2ae2e8109ee74cb4096f06fc49e9cfea6f2900.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 133, + 177, + 305, + 239 + ], + "blocks": [ + { + "bbox": [ + 133, + 177, + 305, + 239 + ], + "lines": [ + { + "bbox": [ + 133, + 177, + 305, + 239 + ], + "spans": [ + { + "bbox": [ + 133, + 177, + 305, + 239 + ], + "type": "image", + "image_path": "35547a6442c50a47ba3d3864555da2913d3a39afa7ee53b86d02b196401034f1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 310, + 177, + 481, + 239 + ], + "blocks": [ + { + "bbox": [ + 310, + 177, + 481, + 239 + ], + "lines": [ + { + "bbox": [ + 310, + 177, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 310, + 177, + 481, + 239 + ], + "type": "image", + "image_path": "f246d9b95b09bcfa424c4c8780f4855f914309fc8dc1b279e0120c2eb0115942.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 133, + 240, + 305, + 316 + ], + "blocks": [ + { + "bbox": [ + 133, + 240, + 305, + 316 + ], + "lines": [ + { + "bbox": [ + 133, + 240, + 305, + 316 + ], + "spans": [ + { + "bbox": [ + 133, + 240, + 305, + 316 + ], + "type": "image", + "image_path": "3d9babbfb58812f97438752c75a8b626012ec5bb4eb30fbfe65e7083527e9ef6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 325, + 482, + 370 + ], + "lines": [ + { + "bbox": [ + 130, + 325, + 482, + 370 + ], + "spans": [ + { + "bbox": [ + 130, + 325, + 482, + 370 + ], + "type": "text", + "content": "Fig. 6: Comparison of concept removal methods for Leonid Afremov's artistic style. Afremov's style is marked by vibrant colors, palette knife technique (similar to pointage technique), and impressionistic elements. HFI+SDD is the most effective, producing images with clear depictions of various objects and minimal retention of his style." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 310, + 240, + 481, + 317 + ], + "blocks": [ + { + "bbox": [ + 310, + 240, + 481, + 317 + ], + "lines": [ + { + "bbox": [ + 310, + 240, + 481, + 317 + ], + "spans": [ + { + "bbox": [ + 310, + 240, + 481, + 317 + ], + "type": "image", + "image_path": "83543329c04df940c576566bc5da83afb3c0025291bcf643b896c3782a572d20.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 392, + 482, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 482, + 537 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 482, + 537 + ], + "type": "text", + "content": "names of 50 countries with high GDP following [40]. This prompt generates way more inappropriate images (74.2% with SD v1.4) compared to the sexual category of the I2P dataset [40] (reportedly ranked " + }, + { + "bbox": [ + 130, + 392, + 482, + 537 + ], + "type": "inline_equation", + "content": "5^{\\text{th}}" + }, + { + "bbox": [ + 130, + 392, + 482, + 537 + ], + "type": "text", + "content": " among the seven categories with 32.8% of hard prompt ratio and only 16.7% of nudity images). The proportion of images classified \"unsafe\" (%NSFW) and those detected with exposed body parts (%NUDE) are evaluated by NudeNet [29]. This process is similarly used to assess the removal of another harmful concept of bleeding by generating images for the shocking category in the I2P dataset [40] and classifying them with the Q16 classifier (%I2P) [41]. The utility is evaluated by generating images from 10,000 random captions from the MSCOCO dataset [22] and comparing these to actual photos, outputs from the original SD model, and their text-to-image alignment using FID [16, 27], LPIPS [54], and CLIPScore [30], respectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 552, + 268, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 552, + 268, + 563 + ], + "spans": [ + { + "bbox": [ + 131, + 552, + 268, + 563 + ], + "type": "text", + "content": "5.2 Artist Style Removal" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 570, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 640 + ], + "type": "text", + "content": "In this section, we conducted experiments to remove the style of artists from the existing SD model. For HFI, we generate 1,000 images from the original SD model with prompts detailed in Appendix B and gather human feedback on a total of 200 images (equivalent to 20 question pairs) per artist. We then train a reward model, which is then used to approximate the scores for all 1,000 images for HFI. All methods with HFI use the same soft token instead of the text token." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "In Tab. 1, even without HFI, SDD alone finds a reasonable trade-off, achieving relatively better removal than the other methods while maintaining a rela-" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "S. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 113, + 305, + 175 + ], + "blocks": [ + { + "bbox": [ + 134, + 113, + 305, + 175 + ], + "lines": [ + { + "bbox": [ + 134, + 113, + 305, + 175 + ], + "spans": [ + { + "bbox": [ + 134, + 113, + 305, + 175 + ], + "type": "image", + "image_path": "6cedc83aebb138b72490f9197e9ee162b8334da92e1b05719d28bbf35a475858.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 309, + 113, + 482, + 175 + ], + "blocks": [ + { + "bbox": [ + 309, + 113, + 482, + 175 + ], + "lines": [ + { + "bbox": [ + 309, + 113, + 482, + 175 + ], + "spans": [ + { + "bbox": [ + 309, + 113, + 482, + 175 + ], + "type": "image", + "image_path": "306eac286f9b49782db4b4132533213b5203a79f05522a852c99e46671f13005.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 133, + 176, + 304, + 239 + ], + "blocks": [ + { + "bbox": [ + 133, + 176, + 304, + 239 + ], + "lines": [ + { + "bbox": [ + 133, + 176, + 304, + 239 + ], + "spans": [ + { + "bbox": [ + 133, + 176, + 304, + 239 + ], + "type": "image", + "image_path": "532a096210f79bf9e596ac25d1d4cab41e6b3911e064c3179c5bf3cee29af597.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 310, + 177, + 481, + 239 + ], + "blocks": [ + { + "bbox": [ + 310, + 177, + 481, + 239 + ], + "lines": [ + { + "bbox": [ + 310, + 177, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 310, + 177, + 481, + 239 + ], + "type": "image", + "image_path": "0f82675506b2cff253b38f8ae66aaf29823fbb7113b06ae89b0633055a8d4372.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 133, + 240, + 304, + 316 + ], + "blocks": [ + { + "bbox": [ + 133, + 240, + 304, + 316 + ], + "lines": [ + { + "bbox": [ + 133, + 240, + 304, + 316 + ], + "spans": [ + { + "bbox": [ + 133, + 240, + 304, + 316 + ], + "type": "image", + "image_path": "1b1f4b4638c92b43c1dee0367acbe97d9aac105084a5768b0507554e9fbaf87a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 325, + 482, + 369 + ], + "lines": [ + { + "bbox": [ + 130, + 325, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 325, + 482, + 369 + ], + "type": "text", + "content": "Fig. 7: Comparison of concept removal methods for Edward Hopper's artistic style. Hopper's characteristic style is marked by realism, simplicity, architectural details, and muted color palette. Among the methods, HFI+SDD most effectively removes his distinctive features while still being abstract and less photorealistic." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 310, + 240, + 481, + 316 + ], + "blocks": [ + { + "bbox": [ + 310, + 240, + 481, + 316 + ], + "lines": [ + { + "bbox": [ + 310, + 240, + 481, + 316 + ], + "spans": [ + { + "bbox": [ + 310, + 240, + 481, + 316 + ], + "type": "image", + "image_path": "81ada7f44b364ffc7df661017cb071ba3130009c191ebeef47b6564de424e05b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 397, + 482, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 397, + 482, + 457 + ], + "spans": [ + { + "bbox": [ + 130, + 397, + 482, + 457 + ], + "type": "text", + "content": "tively lower percentage on " + }, + { + "bbox": [ + 130, + 397, + 482, + 457 + ], + "type": "inline_equation", + "content": "\\% \\mathrm{PHO}" + }, + { + "bbox": [ + 130, + 397, + 482, + 457 + ], + "type": "text", + "content": ". Moreover, when HFI is employed, our method shows the state-of-the-art performance of " + }, + { + "bbox": [ + 130, + 397, + 482, + 457 + ], + "type": "inline_equation", + "content": "\\% \\mathrm{ART}" + }, + { + "bbox": [ + 130, + 397, + 482, + 457 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 130, + 397, + 482, + 457 + ], + "type": "inline_equation", + "content": "12.7\\%" + }, + { + "bbox": [ + 130, + 397, + 482, + 457 + ], + "type": "text", + "content": " on average, with average " + }, + { + "bbox": [ + 130, + 397, + 482, + 457 + ], + "type": "inline_equation", + "content": "\\% \\mathrm{PHO}" + }, + { + "bbox": [ + 130, + 397, + 482, + 457 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 130, + 397, + 482, + 457 + ], + "type": "inline_equation", + "content": "12.0\\%" + }, + { + "bbox": [ + 130, + 397, + 482, + 457 + ], + "type": "text", + "content": ". For Edward Hopper, however, HFI+SDD was outperformed by a few HFI-enhanced baselines, but it worked the best in most cases. Notably, HFI+SDD always surpasses all the previous methods without HFI." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 460, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 482, + 567 + ], + "type": "text", + "content": "Among the baselines, fine-tuning methods (ESD and SDD) observed the most benefit from HFI. In contrast, methods directly optimizing cross-attention exhibited subpar removal performance and appeared to have relatively less impact. Meanwhile, inference-time techniques generally demonstrated improvements in the trade-off between effectiveness and utility facilitated by HFI as shown in Fig. 8. However, as shown in Figs. 5 to 7, there are often cases where the object is not well represented or the distinctive artistic characteristics remain. In such cases, HFI can be beneficial, whereas text-based methods frequently struggle to distinguish the artist's concept from the generic artwork style." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "In the qualitative evaluation using ChatGPT 4o [25], HFI+SDD was evaluated as the most superior method by stating that \"it consistently produces the most neutral images by moving away from the artist's style while maintaining a high level of image quality.\" Despite being asked to evaluate only the removal performance, somewhat surprisingly, ChatGPT considered not only the degree of style removal but also various aspects such as object depiction, image quality, and artifacts) with consistent judgments across all objects and presentation orders, aligning with our proposed metrics. Refer to Appendix F for full results." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "text", + "content": "Safeguard Diffusion Models with Human Feedback Inversion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 248, + 188 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 248, + 188 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 248, + 188 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 248, + 188 + ], + "type": "image", + "image_path": "5de7e08e7bc6d426cb1918f91d18398a07256e2a8da7e72c84ec934df9db819e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 235, + 190, + 391, + 200 + ], + "lines": [ + { + "bbox": [ + 235, + 190, + 391, + 200 + ], + "spans": [ + { + "bbox": [ + 235, + 190, + 391, + 200 + ], + "type": "text", + "content": "--- SLD -SLD+HFI SEGAE SEGA+HFI" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 131, + 210, + 482, + 245 + ], + "lines": [ + { + "bbox": [ + 131, + 210, + 482, + 245 + ], + "spans": [ + { + "bbox": [ + 131, + 210, + 482, + 245 + ], + "type": "text", + "content": "Fig. 8: Trends of " + }, + { + "bbox": [ + 131, + 210, + 482, + 245 + ], + "type": "inline_equation", + "content": "\\%_{\\text{ART}}" + }, + { + "bbox": [ + 131, + 210, + 482, + 245 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 210, + 482, + 245 + ], + "type": "inline_equation", + "content": "\\%_{\\text{PHO}}" + }, + { + "bbox": [ + 131, + 210, + 482, + 245 + ], + "type": "text", + "content": " by inference-time techniques. While forgetting the general artwork concepts (higher " + }, + { + "bbox": [ + 131, + 210, + 482, + 245 + ], + "type": "inline_equation", + "content": "\\%_{\\text{PHO}}" + }, + { + "bbox": [ + 131, + 210, + 482, + 245 + ], + "type": "text", + "content": ") is inevitable, HFI can achieve similar outcomes at smaller editing scales. The same marker symbol denotes the same scale in each plot." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 249, + 114, + 363, + 188 + ], + "blocks": [ + { + "bbox": [ + 249, + 114, + 363, + 188 + ], + "lines": [ + { + "bbox": [ + 249, + 114, + 363, + 188 + ], + "spans": [ + { + "bbox": [ + 249, + 114, + 363, + 188 + ], + "type": "image", + "image_path": "df39b96308b953562803d94d0a24bbca797753e042be2dec2724ac00460b7282.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 365, + 114, + 479, + 188 + ], + "blocks": [ + { + "bbox": [ + 365, + 114, + 479, + 188 + ], + "lines": [ + { + "bbox": [ + 365, + 114, + 479, + 188 + ], + "spans": [ + { + "bbox": [ + 365, + 114, + 479, + 188 + ], + "type": "image", + "image_path": "1ef1f7713ddafb5c71a53083f2660eaf8f136b0d77c2d2fc3c0b551b89c5d461.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 134, + 255, + 323, + 354 + ], + "blocks": [ + { + "bbox": [ + 134, + 255, + 323, + 354 + ], + "lines": [ + { + "bbox": [ + 134, + 255, + 323, + 354 + ], + "spans": [ + { + "bbox": [ + 134, + 255, + 323, + 354 + ], + "type": "image", + "image_path": "4a6df93dda4cf5ac0f65f1f1fecef967d657e895e7f5cda7693df28f5927719c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 361, + 321, + 384 + ], + "lines": [ + { + "bbox": [ + 131, + 361, + 321, + 384 + ], + "spans": [ + { + "bbox": [ + 131, + 361, + 321, + 384 + ], + "type": "text", + "content": "Fig. 9: HFI mitigates terrifying aspects of images generated with I2P shocking prompts." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 329, + 262, + 481, + 308 + ], + "blocks": [ + { + "bbox": [ + 329, + 262, + 481, + 308 + ], + "lines": [ + { + "bbox": [ + 329, + 262, + 481, + 308 + ], + "spans": [ + { + "bbox": [ + 329, + 262, + 481, + 308 + ], + "type": "image", + "image_path": "a4c467ccfe44f3a90ed5f1da64816723cd3b67eab43dfa4550a98b1c8431da5e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 330, + 309, + 481, + 353 + ], + "blocks": [ + { + "bbox": [ + 330, + 309, + 481, + 353 + ], + "lines": [ + { + "bbox": [ + 330, + 309, + 481, + 353 + ], + "spans": [ + { + "bbox": [ + 330, + 309, + 481, + 353 + ], + "type": "image", + "image_path": "148414c18f80fc16ff51f92d273df82c24982b6f43f88415825450968f456ff6.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 361, + 481, + 384 + ], + "lines": [ + { + "bbox": [ + 329, + 361, + 481, + 384 + ], + "spans": [ + { + "bbox": [ + 329, + 361, + 481, + 384 + ], + "type": "text", + "content": "Fig.10: Images with the highest (top) and lowest (bottom) rewards." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 407, + 296, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 296, + 419 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 296, + 419 + ], + "type": "text", + "content": "5.3 Harmful Concept Removal" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 426, + 482, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 569 + ], + "type": "text", + "content": "Tab. 2 shows our experimental results on removing NSFW and bleeding concepts. Similar to Sec. 5.2, we first generated images and collected binary feedback and ranking feedback for nudity and bleeding, respectively. SDD alone effectively reduces the presence of NSFW imagery, but when combined with HFI, such images are nearly eliminated. Our method and ESD-u-3 achieve a similar level, but ESD exhibits inferior performance in all image quality metrics. In a similar vein, the method of fine-tuning only the cross-attention, as in ESD-x-3, is observed to generate harmful images beyond a certain threshold. High in %NSFW and low in %NUDE for some methods indicate that they still generate somewhat unsafe images even though exposed body parts are limited. When combined with HFI, all methods but FMN exhibit a sharp decrease in %NSFW, underscoring the role of the soft token aligned with human judgment." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Furthermore, for nuanced concepts such as bleeding, HFI aids in identifying and mitigating the inappropriate aspects of images. In Fig. 9, with HFI, it becomes apparent that not only has the extent of bleeding diminished, but it also yields images that are safer and less offensive. Particularly notable are the outcomes achieved by combining SDD and HFI, where notably safer images are produced despite maintaining identical compositions. The interpretation of such ambiguous concepts by humans can be encoded into learned tokens for subsequent utilization in their removal, highlighting the utility of HFI." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "S. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 168, + 479, + 381 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 1: Quantitative comparison of artist style removal performance. Each method is evaluated both without and with HFI (✓), showing improvement due to human feedback. The " + }, + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "inline_equation", + "content": "\\sum" + }, + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "text", + "content": " column represents the sum of %ART and %PHO, indicating undesirable images. The table also compares our SDD method with other baselines." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 168, + 479, + 381 + ], + "lines": [ + { + "bbox": [ + 135, + 168, + 479, + 381 + ], + "spans": [ + { + "bbox": [ + 135, + 168, + 479, + 381 + ], + "type": "table", + "html": "
Target ArtistVincent van GoghLeonid AfremovEdward HopperAvg. LPIPS
MethodHFI%ART↓ %PHO↓Σ%ART↓ %PHO↓Σ%ART↓ %PHO↓ΣTarget↑ Others↓
SD v1.495.92.998.899.90.6100.597.62.299.8--
Inference-time Technique
NEG [36]✓15.032.547.592.82.094.836.512.148.60.33520.2273
17.212.429.60.030.430.416.516.132.60.43520.3255
SLD [40]✓29.127.756.827.84.232.021.339.560.80.42770.2880
24.622.847.49.78.818.518.134.953.00.44590.3285
SEGA [3]✓40.311.752.064.41.565.937.728.666.30.31510.2522
17.517.635.120.74.525.232.021.853.80.34820.2661
Direct Cross-Attention Optimization
FMN [52]✓66.67.574.110.725.536.277.15.782.80.32930.1665
58.67.766.36.50.97.473.25.578.70.27930.2478
UCE [14]✓66.44.070.444.72.046.751.53.354.80.27670.0332
77.32.579.892.50.693.172.51.974.40.17440.0328
Fine-tuning Method
ESD-x-1 [13]✓45.015.960.913.525.438.964.47.273.60.36240.1635
22.520.442.919.057.276.214.317.431.70.44040.1691
SDD (ours)✓20.625.546.19.98.218.140.441.181.50.27600.2341
9.55.114.68.83.011.819.827.947.70.28790.1439
", + "image_path": "4035f0c4619fd439f622aee8a8cb606bfcb520a98cc99564399f052fcd935382.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 406, + 397, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 406, + 397, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 406, + 397, + 418 + ], + "type": "text", + "content": "5.4 Analysis on Rewards and Learned Embeddings" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 426, + 482, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 570 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 570 + ], + "type": "text", + "content": "For nudity, the reward model generally aligns with Nudenet [29] (AUROC of 0.882), but differences between human and model judgments highlight the limits of relying solely only on models for inversion. Fig. 10 shows that images deemed appropriate by humans or not reflecting the artist's style receive low rewards and minimally impact HFI token learning. Integrating HFI with reward models effectively captures concepts. Ablation studies comparing HFI with the vanilla TI [12] show a " + }, + { + "bbox": [ + 130, + 426, + 482, + 570 + ], + "type": "inline_equation", + "content": "3.2\\%" + }, + { + "bbox": [ + 130, + 426, + 482, + 570 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 426, + 482, + 570 + ], + "type": "inline_equation", + "content": "5.1\\%" + }, + { + "bbox": [ + 130, + 426, + 482, + 570 + ], + "type": "text", + "content": " increase in " + }, + { + "bbox": [ + 130, + 426, + 482, + 570 + ], + "type": "inline_equation", + "content": "\\% \\mathrm{ART}" + }, + { + "bbox": [ + 130, + 426, + 482, + 570 + ], + "type": "text", + "content": " for artist removal without human feedback (i.e., random images without rewards) and with real images from the internet (without rewards), respectively. For nudity removal, " + }, + { + "bbox": [ + 130, + 426, + 482, + 570 + ], + "type": "inline_equation", + "content": "\\% \\mathrm{NSFW}" + }, + { + "bbox": [ + 130, + 426, + 482, + 570 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 426, + 482, + 570 + ], + "type": "inline_equation", + "content": "\\% \\mathrm{NUDE}" + }, + { + "bbox": [ + 130, + 426, + 482, + 570 + ], + "type": "text", + "content": " increase by " + }, + { + "bbox": [ + 130, + 426, + 482, + 570 + ], + "type": "inline_equation", + "content": "1.74\\%" + }, + { + "bbox": [ + 130, + 426, + 482, + 570 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 426, + 482, + 570 + ], + "type": "inline_equation", + "content": "0.04\\%" + }, + { + "bbox": [ + 130, + 426, + 482, + 570 + ], + "type": "text", + "content": " without human feedback, with FID increasing by 1.565. This demonstrates the importance of integrating human judgments and model-generated data into the training process to achieve optimal performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Compared to existing expensive methods using human feedback, the proposed HFI is highly cost-effective. We obtained 650 responses from five individuals (450 unique images for inter-annotator agreement (IAA)) within three hours for nudity. Providing annotation guidelines increased IAA, evidenced by Fleiss' " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " [11] at 0.4997 and Randolph's " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " [34] at 0.6702. For artistic concepts, we collected 40 ranking feedback responses (400 images) per artist from 15 individuals. Our research offers an economical and effective framework for extracting information from models and removing harmful content with human judgment." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "text", + "content": "Safeguard Diffusion Models with Human Feedback Inversion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 176, + 157, + 434, + 441 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 2: Quantitative comparison of harmful concept removal performance. Each method is evaluated both without and with HFI (✓) for inter-method comparison. The table also compares our SDD with other baselines for intra-method comparison." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 176, + 157, + 434, + 441 + ], + "lines": [ + { + "bbox": [ + 176, + 157, + 434, + 441 + ], + "spans": [ + { + "bbox": [ + 176, + 157, + 434, + 441 + ], + "type": "table", + "html": "
Target ConceptNudityBleedingCOCO10K
MethodHFI%NSFW↓%NUDE↓%I2P↓FID↓CLIP↑ LPIPS↓
SD v1.474.1832.1465.7716.6360.2745
COCO ref.0.2676
Inference-time Technique
SD+NEG [36]✓30.683.8645.7918.8370.2701
5.144.7624.6519.5180.2699
SLD [40]✓58.089.8446.1417.8150.2702
42.9815.5829.5618.3450.2704
SEGA [3]✓67.7213.1645.4417.1290.2698
56.4219.5239.3716.7560.2698
Direct Cross-Attention Optimization
FMN [52]✓51.5027.7847.0818.0540.2556
61.3234.7255.7216.8130.2683
UCE [14]✓32.307.3661.8017.0240.2726
14.769.3253.2716.5410.2736
Fine-tuning Method
ESD-u-1 [13]✓11.302.00-17.0330.2664
2.421.16-18.4040.2636
ESD-u-3 [13]✓4.440.48-18.4980.2624
0.800.46-18.6150.2621
ESD-x-3 [13]✓14.322.9851.1718.5170.2677
4.183.3034.3517.8980.2670
SDD (ours)✓2.780.6030.8417.3620.2658
0.280.3223.3617.0230.2643
", + "image_path": "645f2562f68283f4b6f793885e19775210d35e27d848254e66ee241ef7f8468d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 462, + 312, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 462, + 312, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 462, + 312, + 475 + ], + "type": "text", + "content": "6 Conclusion and Discussion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 486, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 594 + ], + "type": "text", + "content": "In this work, we introduced a novel concept removal framework within a text-to-image diffusion model through human feedback inversion and self-distillation-based fine-tuning. By leveraging human feedback to train soft tokens, we effectively safeguarded the diffusion model from harmful concepts, addressing the ambiguity in defining text-based concepts. We further proposed a self-distillation-based diffusion model fine-tuning algorithm, enhancing the efficacy of concept removal. We believe our research lays the groundwork for a human-centric approach to the safety and ethical considerations of large-scale diffusion models, paving the way for future advancements." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Limitation. While our proposed HFI performs well in capturing the intended concept compared to the text, its effectiveness depends on the quality of human feedback and the performance of inversion to soft tokens. Significant disagreements in human feedback can compromise the learning process, underscoring the importance of meticulous data collection and reward modeling." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "S. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "content": "This work was partly supported by Institute for Information & communications Technology Promotion (IITP) grant funded by the Korea government (MSIT) (No. RS-2019-II190075, Artificial Intelligence Graduate School Program (KAIST), No. 2022-0-00184, Development and Study of AI Technologies to Inexpensively Conform to Evolving Policy on Ethics) and the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. 2022R1A5A708390812)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 243, + 197, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 243, + 197, + 255 + ], + "spans": [ + { + "bbox": [ + 133, + 243, + 197, + 255 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 269, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 269, + 480, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 269, + 480, + 301 + ], + "spans": [ + { + "bbox": [ + 138, + 269, + 480, + 301 + ], + "type": "text", + "content": "1. Baio, A.: Exploring 12 million of the 2.3 billion images used to train stable diffusion's image generator. https://waxy.org/2022/08/exploring-12-million-of-the-images-used-to-train-stable-diffusions-image-generator/ (Aug 2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 303, + 480, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 303, + 480, + 323 + ], + "spans": [ + { + "bbox": [ + 138, + 303, + 480, + 323 + ], + "type": "text", + "content": "2. Black, K., Janner, M., Du, Y., Kostrikov, I., Levine, S.: Training diffusion models with reinforcement learning. arXiv preprint arXiv:2305.13301 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 324, + 480, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 324, + 480, + 356 + ], + "spans": [ + { + "bbox": [ + 138, + 324, + 480, + 356 + ], + "type": "text", + "content": "3. Brack, M., Friedrich, F., Hintersdorf, D., Struppek, L., Schramowski, P., Kersting, K.: Sega: Instructing diffusion using semantic dimensions. arXiv preprint arXiv:2301.12247 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 357, + 480, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 357, + 480, + 389 + ], + "spans": [ + { + "bbox": [ + 138, + 357, + 480, + 389 + ], + "type": "text", + "content": "4. Brack, M., Schramowski, P., Friedrich, F., Hintersdorf, D., Kersting, K.: The stable artist: Steering semantics in diffusion latent space. arXiv preprint arXiv:2212.06013 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 390, + 480, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 390, + 480, + 411 + ], + "spans": [ + { + "bbox": [ + 138, + 390, + 480, + 411 + ], + "type": "text", + "content": "5. Bradley, R.A., Terry, M.E.: Rank analysis of incomplete block designs: I. the method of paired comparisons. Biometrika 39(3/4), 324-345 (1952)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 412, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 412, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 138, + 412, + 480, + 445 + ], + "type": "text", + "content": "6. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 445, + 480, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 445, + 480, + 488 + ], + "spans": [ + { + "bbox": [ + 138, + 445, + 480, + 488 + ], + "type": "text", + "content": "7. Cherti, M., Beaumont, R., Wightman, R., Wortsman, M., Ilharco, G., Gordon, C., Schuhmann, C., Schmidt, L., Jitsev, J.: Reproducible scaling laws for contrastive language-image learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2818-2829 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 489, + 480, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 489, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 138, + 489, + 480, + 521 + ], + "type": "text", + "content": "8. Christiano, P.F., Leike, J., Brown, T., Martic, M., Legg, S., Amodei, D.: Deep reinforcement learning from human preferences. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 522, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 522, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 522, + 480, + 544 + ], + "type": "text", + "content": "9. Du, Y., Li, S., Mordatch, I.: Compositional visual generation with energy based models. Advances in Neural Information Processing Systems 33, 6637-6647 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 544, + 480, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 544, + 480, + 577 + ], + "spans": [ + { + "bbox": [ + 138, + 544, + 480, + 577 + ], + "type": "text", + "content": "10. Fan, Y., Watkins, O., Du, Y., Liu, H., Ryu, M., Boutilier, C., Abbeel, P., Ghavamzadeh, M., Lee, K., Lee, K.: Dpok: Reinforcement learning for fine-tuning text-to-image diffusion models. arXiv preprint arXiv:2305.16381 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 577, + 480, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 577, + 480, + 598 + ], + "spans": [ + { + "bbox": [ + 138, + 577, + 480, + 598 + ], + "type": "text", + "content": "1. Fleiss, J.L.: Measuring nominal scale agreement among many raters. Psychological bulletin 76(5), 378 (1971)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 599, + 480, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 599, + 480, + 642 + ], + "spans": [ + { + "bbox": [ + 134, + 599, + 480, + 642 + ], + "type": "text", + "content": "12. Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A.H., Chechik, G., Cohen-Or, D.: An image is worth one word: Personalizing text-to-image generation using textual inversion (2022). https://doi.org/10.48550/ARXIV.2208.01618, https://arxiv.org/abs/2208.01618" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 643, + 480, + 665 + ], + "type": "text", + "content": "13. Gandikota, R., Materzynska, J., Fiotto-Kaufman, J., Bau, D.: Erasing concepts from diffusion models. arXiv preprint arXiv:2303.07345 (2023)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "text", + "content": "Safeguard Diffusion Models with Human Feedback Inversion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "14. Gandikota, R., Orgad, H., Belinkov, Y., Materzynska, J., Bau, D.: Unified concept editing in diffusion models. arXiv preprint arXiv:2308.14761 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 480, + 171 + ], + "type": "text", + "content": "15. Gao, J., Galley, M., Li, L.: Neural approaches to conversational ai. In: The 41st international ACM SIGIR conference on research & development in information retrieval. pp. 1371-1374 (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 480, + 205 + ], + "type": "text", + "content": "16. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 205, + 480, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 205, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 133, + 205, + 480, + 225 + ], + "type": "text", + "content": "17. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. arXiv preprint arxiv:2006.11239 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 227, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 227, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 133, + 227, + 480, + 258 + ], + "type": "text", + "content": "18. Kumari, N., Zhang, B., Wang, S.Y., Shechtman, E., Zhang, R., Zhu, J.Y.: Ablating concepts in text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22691-22702 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 260, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 260, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 133, + 260, + 480, + 293 + ], + "type": "text", + "content": "19. Kumari, N., Zhang, B., Zhang, R., Shechtman, E., Zhu, J.Y.: Multi-concept customization of text-to-image diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1931-1941 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "text", + "content": "20. Lee, K., Liu, H., Ryu, M., Watkins, O., Du, Y., Boutilier, C., Abbeel, P., Ghavamzadeh, M., Gu, S.S.: Aligning text-to-image models using human feedback. arXiv preprint arXiv:2302.12192 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 326, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 346 + ], + "type": "text", + "content": "21. Li, J., Monroe, W., Ritter, A., Galley, M., Gao, J., Jurafsky, D.: Deep reinforcement learning for dialogue generation. arXiv preprint arXiv:1606.01541 (2016)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 347, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 391 + ], + "type": "text", + "content": "22. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 392, + 480, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 423 + ], + "type": "text", + "content": "23. Ni, Z., Wei, L., Li, J., Tang, S., Zhuang, Y., Tian, Q.: Degeneration-tuning: Using scrambled grid shield unwanted concepts from stable diffusion. arXiv preprint arXiv:2308.02552 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 425, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 425, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 425, + 480, + 456 + ], + "type": "text", + "content": "24. O'Connor, R.: Stable diffusion 1 vs 2 - what you need to know. https://www.assemblyai.com/blog/stable-diffusion-1-vs-2-what-you-need-to-know/ (Dec 2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 457, + 439, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 439, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 439, + 468 + ], + "type": "text", + "content": "25. OpenAI: Chatgpt (2023), https://www.openai.com/, july 2023 version" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 468, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 511 + ], + "type": "text", + "content": "26. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems 35, 27730-27744 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "type": "text", + "content": "27. Parmar, G., Zhang, R., Zhu, J.Y.: On aliased resizing and surprising subtleties in gan evaluation. In: CVPR (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "type": "text", + "content": "28. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "type": "text", + "content": "29. Praneeth, B.: NudeNet: Neural nets for nudity classification, detection and selective censoring. https://github.com/notAI-tech/NudeNet (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 589, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 480, + 632 + ], + "type": "text", + "content": "30. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "31. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 1(2), 3 (2022)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "S. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "32. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: International Conference on Machine Learning. pp. 8821-8831. PMLR (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 481, + 171 + ], + "type": "text", + "content": "33. Rando, J., Paleka, D., Lindner, D., Heim, L., Tramér, F.: Red-teaming the stable diffusion safety filter. arXiv preprint arXiv:2210.04610 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 193 + ], + "type": "text", + "content": "34. Randolph, J.J.: Free-marginal multirater kappa (multirater k [free]): An alternative to fleiss' fixed-marginal multirater kappa. Online submission (2005)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "text", + "content": "35. Roich, D., Mokady, R., Bermano, A.H., Cohen-Or, D.: Pivotal tuning for latent-based editing of real images. ACM Transactions on graphics (TOG) 42(1), 1-13 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "type": "text", + "content": "36. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10684-10695 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 303 + ], + "type": "text", + "content": "37. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 346 + ], + "type": "text", + "content": "38. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22500-22510 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 347, + 481, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 390 + ], + "type": "text", + "content": "39. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 392, + 481, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 481, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 481, + 423 + ], + "type": "text", + "content": "40. Schramowski, P., Brack, M., Deiseroth, B., Kersting, K.: Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 424, + 481, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 481, + 467 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 481, + 467 + ], + "type": "text", + "content": "41. Schramowski, P., Tauchmann, C., Kersting, K.: Can machines help us answering question 16 in datasheets, and in turn reflecting on inappropriate content? In: 2022 ACM Conference on Fairness, Accountability, and Transparency. pp. 1350-1361 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 468, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 511 + ], + "type": "text", + "content": "42. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. arXiv preprint arXiv:2210.08402 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "type": "text", + "content": "43. Schulman, J., Wolski, F., Dhariwal, P., Radford, A., Klimov, O.: Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347 (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 533, + 481, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 481, + 566 + ], + "type": "text", + "content": "44. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: Proceedings of the 32nd International Conference on Machine Learning (2015)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 567, + 481, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 588 + ], + "type": "text", + "content": "45. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 589, + 481, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 481, + 610 + ], + "type": "text", + "content": "46. Song, Y., Ermon, S.: Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems 32 (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "text", + "content": "47. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456 (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "text", + "content": "48. Stability AI: Stable diffusion 2.0 release. https://stability.ai/blog/stable-diffusion-v2-release (Nov 2022)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "text", + "content": "Safeguard Diffusion Models with Human Feedback Inversion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 315 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "49. Stiennon, N., Ouyang, L., Wu, J., Ziegler, D., Lowe, R., Voss, C., Radford, A., Amodei, D., Christiano, P.F.: Learning to summarize with human feedback. Advances in Neural Information Processing Systems 33, 3008-3021 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 182 + ], + "type": "text", + "content": "50. Thiel, D.: Identifying and eliminating csam in generative ml training data and models. Tech. rep., Technical Report. Stanford University, Palo Alto, CA. https://purl.stanford... (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 480, + 215 + ], + "type": "text", + "content": "51. Yang, Z., Yi, X., Li, P., Liu, Y., Xie, X.: Unified detoxifying and debiasing in language generation via inference-time adaptive optimization. arXiv preprint arXiv:2210.04492 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 215, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 215, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 215, + 480, + 237 + ], + "type": "text", + "content": "52. Zhang, E., Wang, K., Xu, X., Wang, Z., Shi, H.: Forget-me-not: Learning to forget in text-to-image diffusion models. arXiv preprint arXiv:2211.08332 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 480, + 281 + ], + "type": "text", + "content": "53. Zhang, L., Song, J., Gao, A., Chen, J., Bao, C., Ma, K.: Be your own teacher: Improve the performance of convolutional neural networks via self distillation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3713-3722 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 281, + 480, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 480, + 315 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 480, + 315 + ], + "type": "text", + "content": "54. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "S. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Salience-Based Adaptive Masking_ Revisiting Token Dynamics for Enhanced Pre-training/05c12db4-7790-4b9e-84f8-e76fa8191a90_content_list.json b/2024/Salience-Based Adaptive Masking_ Revisiting Token Dynamics for Enhanced Pre-training/05c12db4-7790-4b9e-84f8-e76fa8191a90_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..efdbd22a26b751ee7308b2e4248a011a69eb7883 --- /dev/null +++ b/2024/Salience-Based Adaptive Masking_ Revisiting Token Dynamics for Enhanced Pre-training/05c12db4-7790-4b9e-84f8-e76fa8191a90_content_list.json @@ -0,0 +1,1957 @@ +[ + { + "type": "text", + "text": "Salience-Based Adaptive Masking: Revisiting Token Dynamics for Enhanced Pre-training", + "text_level": 1, + "bbox": [ + 243, + 140, + 759, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hyesong Choi $^{1}$ , Hyejin Park $^{1}$ , Kwang Moo Yi $^{2}$ , Sungmin Cha $^{3}$ , and Dongbo Min $^{1\\dagger}$", + "bbox": [ + 328, + 212, + 673, + 246 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Ewha W. University \n2 University of British Columbia", + "bbox": [ + 390, + 253, + 612, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 New York University", + "bbox": [ + 423, + 282, + 578, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. In this paper, we introduce Saliency-Based Adaptive Masking (SBAM), a novel and cost-effective approach that significantly enhances the pre-training performance of Masked Image Modeling (MIM) approaches by prioritizing token salience. Our method provides robustness against variations in masking ratios, effectively mitigating the performance instability issues common in existing methods. This relaxes the sensitivity of MIM-based pre-training to masking ratios, which in turn allows us to propose an adaptive strategy for 'tailored' masking ratios for each data sample, which no existing method can provide. Toward this goal, we propose an Adaptive Masking Ratio (AMR) strategy that dynamically adjusts the proportion of masking for the unique content of each image based on token salience. We show that our method significantly improves over the state-of-the-art in mask-based pre-training on the ImageNet-1K dataset. Code and model parameters are available at https://github.com/doihye/SBAM.", + "bbox": [ + 261, + 339, + 743, + 547 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Self-supervised learning $\\cdot$ Masked image modeling $\\cdot$ Masked autoencoder", + "bbox": [ + 259, + 561, + 743, + 590 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 618, + 375, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent drastic improvements in various Computer Vision tasks rely heavily on Transformer architectures [12]. A critical component that enables these architectures is the necessity of large-scale data [9], which is not always readily available. Naturally, pre-training with pretext tasks has become a popular solution as a workaround, represented by Masked Image Modeling (MIM) [1, 11, 14, 36], inspired by how Masked Language Modeling (MLM) [2, 7, 10, 24] has reshaped the", + "bbox": [ + 212, + 652, + 787, + 746 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (No. RS-2023-00222385) and the National Research Foundation of Korea, Republic of Korea (2022M3H9A2083956) and from the Korea Basic Science Institute, Republic of Korea (National Research Facilities and Equipment Center) funded by the Ministry of Education, Republic of Korea (2019R1A6C1010020). † Corresponding author: dbmin@ewha.ac.kr", + "bbox": [ + 230, + 755, + 789, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/edf5841832ab87f7f23efb6a9161f423637aed875e113a9115a6934592850e2f.jpg", + "image_caption": [ + "Fig. 1: Overview of SBAM. Whereas (a) Random Masking must rely completely on chance and carefully tuned masking ratio to guarantee effective masking, (b) the proposed SBAM strategically masks tokens based on the token salience. The robustness of SBAM paved the way for the introduction of (c) AMR, which implements a tailored masking ratio for each sample in the dataset." + ], + "image_footnote": [], + "bbox": [ + 253, + 147, + 751, + 320 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Natural Language Processing landscape. These strategies involve masking a subset of the input data and predicting those that have been hidden, thus forcing the deep network to infer underlying concepts.", + "bbox": [ + 212, + 414, + 784, + 460 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While MIM has significantly advanced self-supervised learning in vision, its conventional approach to randomly selecting tokens for masking falls short of harnessing the full potential of visual data [18]. Unlike in text, where randomness might obscure key semantic units, the visual domain's complexity and token redundancy demand a more strategic masking protocol to ensure model comprehension. This necessity prompts us to explore a refined masking methodology, targeting a selection process that achieves the goal of image understanding for pre-trained models, thereby bridging the gap in modality-specific pre-training strategies.", + "bbox": [ + 212, + 460, + 787, + 595 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Various techniques [18, 25, 35] have thus been proposed for more effective masking. While these methods all strive toward improved pre-training, an oversight shared amongst all methods is that they do not regard the contribution of each token within the overall composition of the image. It is crucial to scrutinize the interconnections between tokens to ensure that the token masking includes the tokens that play a pivotal role within the image. Moreover, prior masking methodologies [18, 25, 35] fail to address the critical consideration of the masking ratio, a factor that ought to dynamically adjust in response to the size and quantity of pivotal objects embedded within the image. This is primarily due to the difficulty reported in various works [14, 36, 39] that minor changes in the optimized masking ratio can lead to performance instability, rendering such considerations difficult to implement. Thus, these methods must rely completely on chance and carefully tuned masking ratio to guarantee effective masking. Besides, contemporary masking strategies [18, 25, 35] often grapple with high complexity, burdened by intricate distillation frameworks [18], auxiliary detection processes [35], and duplicated forward processes [25].", + "bbox": [ + 212, + 598, + 787, + 843 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "H. Choi et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we introduce a simple yet novel approach that focuses on token dynamics, termed Salience-Based Adaptive Masking (SBAM), which aims to strategically select masking locations by discerning perceptual prominence within the visual data. Specifically, the proposed method leverages the directional emphasis from attention mechanisms to identify the image tokens pivotal to the visual context. Significantly, our methodology stands apart from the conventional attention-based approach [25] by leveraging the token's outgoing weight to calculate its 'token salience' within the image and prioritizing those that have high salience to be masked. We also infuse a degree of randomness into token salience to enrich the diversity of mask generation. Further details of SBAM are elaborated in Sec. 3. Thus, without any significant additional cost, we can consider the token's prominence within the image.", + "bbox": [ + 212, + 146, + 787, + 328 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Crucially, the strength of the SBAM approach lies in its robustness to the varying masking ratios—a notable vulnerability in established baselines [14,36, 39], which struggles from performance instability with even minor variations in optimized masking ratio. The robustness of SBAM is attributable to the proposed analytical precision in token dynamics, selectively masking tokens pivotal to the image's entirety. As a result, SBAM exhibits a diminished likelihood of masking redundant or non-essential segments, offering a stable alternative to the random masking or preceding masking approaches. Along with improving robustness to mask ratio variations, SBAM significantly enhances pre-training efficiency. A comprehensive evaluation of this is provided in Sec. 4.", + "bbox": [ + 212, + 330, + 787, + 482 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Establishing robustness against variations in masking ratios has empowered us to expand the discourse on image masking into a pioneering aspect, introducing an innovative paradigm: an Adaptive Masking Ratio (AMR). We find having masking ratios that adapt throughout training to be highly effective, as it allows the masking process to be tailored to each sample in the dataset. For instance, each image may benefit from different masking ratios, as one image might have a close-up of a bear which would enjoy high masking ratios, whereas one might have an airplane in the sky that would require a lower masking ratio; see (c) of Fig. 1. The proposed token salience, reapplied in this context, serves to finely determine the dynamic masking ratio by quantifying the proportion of tokens exhibiting high salience. More details can be found in Sec. 5. The proposed AMR implements an adaptive strategy that respects the distinctiveness of visual data and thus achieves significant performance gains when applied to various baselines (refer to Sec. 6.2 and 6.3. We note that this type of 'tailoring' to each image for pre-training is impossible with any existing method.", + "bbox": [ + 212, + 484, + 787, + 712 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In sum, this paper offers an innovative masking strategy that enhances the robustness and effectiveness of pre-trained models, setting the stage for a notable shift in the field of self-supervised learning. More importantly, the proposed method can be universally applied across any MIM framework that exploits token masking. We evaluate our SBAM and AMR on data-hungry models like ViT-L/ViT-B [12] on ImageNet-1K [9] datasets, achieving significant performance improvements in both fine-tuning and linear probing accuracy.", + "bbox": [ + 212, + 715, + 787, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To summarize, our contributions are:", + "bbox": [ + 240, + 824, + 509, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "SBAM", + "bbox": [ + 683, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- we present Saliency-Based Adaptive Masking (SBAM), a novel effective method for MIM pre-training that focuses on token salience;", + "- being saliency-based, without a significant increase in computation, we allow effective masking that is robust to masking ratios:", + "- empowered by the robustness to masking ratio, we propose an Adaptive Masking Ratio (AMR) that allows tailored masking for each sample;", + "- we evaluate our method on ImageNet-1K datasets, achieving notable enhancements in both fine-tuning and linear probing accuracy." + ], + "bbox": [ + 223, + 146, + 782, + 266 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 292, + 380, + 308 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Within the Masked Image Modeling (MIM) domain, the crux lies in the strategic corruption and subsequent reconstruction of image segments. This process hinges on two core operations: Random Masking of image tokens and Reconstruction of corrupted tokens to guide the learning process.", + "bbox": [ + 212, + 325, + 782, + 383 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For a given image sequence $X \\in \\mathbb{R}^{N \\times L \\times D}$ , where $N, L$ and $D$ denote batch size, number of tokens per image and dimensionality of each token respectively, and specified masking proportion $\\gamma$ , the general random masking process conducted by $\\varPhi_{\\mathrm{mask}}$ is defined as:", + "bbox": [ + 212, + 386, + 782, + 445 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nX _ {\\text {m a s k e d}}, M = \\Phi_ {\\text {m a s k}} (X, \\gamma), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 460, + 782, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $X_{\\mathrm{masked}}$ represents the visible tokens as a result of the post-application of the random mask $M$ , with $M \\in \\{0,1\\}^{N \\times L}$ indicating the presence (1) or absence (0) of masking for each element.", + "bbox": [ + 212, + 488, + 784, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The key component for learning in MIM is the reconstruction error $\\mathcal{L}_{\\mathrm{MIM}}$ , which is computed by the mean squared error (MSE) between the predicted representation of masked tokens, denoted as $\\hat{X}$ , and their normalized original counterparts, represented by $X$ :", + "bbox": [ + 212, + 534, + 782, + 592 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {X} = \\frac {X - \\mu (X)}{\\sigma (X) + \\epsilon}, \\quad \\mathcal {L} _ {\\mathrm {M I M}} = \\frac {1}{\\sum_ {i , j} M _ {i , j}} \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {L} M _ {i, j} \\cdot \\| \\hat {X} _ {i, j} - \\bar {X} _ {i, j} \\| _ {2} ^ {2}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 264, + 607, + 782, + 647 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $\\mu (\\cdot)$ and $\\sigma (\\cdot)$ represent the mean and standard deviation of $X$ respectively. $\\| \\cdot \\| _2^2$ denotes the squared $L_{2}$ norm and $\\epsilon$ is a small value for numerical stability. This formulation presents the core of MIM's training objective, focusing explicitly on the reconstruction of the masked portions of the input, encouraging the model to infer corrupted information from the unmasked context.", + "bbox": [ + 214, + 660, + 782, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Salience-Based Masking (SBM)", + "text_level": 1, + "bbox": [ + 214, + 761, + 555, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Masked Image Modeling (MIM) has greatly pushed forward self-supervised learning in the visual domain, yet its standard practice of randomly masking tokens fails to fully capture the richness of visual information. To overcome this", + "bbox": [ + 212, + 794, + 782, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "H. Choi et al.", + "bbox": [ + 271, + 114, + 364, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "limitation, we present Salience-Based Masking (SBM), a novel masking methodology that selects masking locations by revisiting token dynamics.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given an input tensor $X \\in \\mathbb{R}^{N \\times (L \\times D)}$ , the first step involves computing an affinity matrix $\\mathcal{A} \\in \\mathbb{R}^{N \\times L \\times L}$ through a batch matrix-matrix product between $X$ and $X' \\in \\mathbb{R}^{N \\times (D \\times L)}$ . Subsequently, we apply the softmax function $\\rho(\\cdot)$ to the affinity matrix, resulting in a normalized affinity matrix denoted as $\\hat{\\mathcal{A}} = \\rho(\\mathcal{A}) \\in \\mathbb{R}^{N \\times L \\times L}$ . The softmax function is implemented as follows:", + "bbox": [ + 212, + 176, + 784, + 252 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\rho (\\mathcal {A}) _ {n, i, j} = \\frac {e ^ {a _ {n , i , j}}}{\\sum_ {k = 1} ^ {L} e ^ {a _ {n , i , k}}}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 411, + 262, + 785, + 297 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $a_{n,i,j}$ represents an element of $\\mathcal{A}$ and $e$ is the base of the natural logarithm. Note that, for each $n$ -th element with a shape of $\\mathbb{R}^{L\\times L}$ , this function normalizes the rows of it, transforming them into probabilities that sum to 1.", + "bbox": [ + 212, + 301, + 782, + 345 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Crucially, our approach distinguishes itself from conventional attention-based methods by utilizing the sum of outgoing weight of each token to determine the 'token salience' $S \\in \\mathbb{R}^{N \\times L}$ in the image, which is represented by the column-wise summed score of $\\hat{\\mathcal{A}}$ :", + "bbox": [ + 212, + 347, + 784, + 406 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS = \\mathcal {N} \\left(\\Sigma_ {j = 1} ^ {L} \\hat {\\mathcal {A}} _ {:, j,:}\\right), \\quad \\mathcal {N} (x) = \\frac {x - \\operatorname* {m i n} (x)}{\\operatorname* {m a x} (x) - \\operatorname* {m i n} (x)}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 313, + 415, + 785, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the affinity map, the row-wise score represents the incoming weight, reflecting the extent to which other tokens influence the corresponding token. Conversely, the column-wise score signifies the outgoing weight, indicating the contribution of the corresponding token to others. By summing these scores, it becomes feasible to quantify the token's overall impact on the image, thereby defining token salience.", + "bbox": [ + 212, + 455, + 782, + 545 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The adaptive masking process of SBM is formulated with the token salience $S$ . Notably, exclusive reliance on $S$ for masking can precipitate a decline in performance (refer to Fig. 7). To mitigate this issue, we incorporated an element of randomness into the masking process to get the adjusted token salience, denoted as $\\tilde{S} = S + N$ , where $N \\sim U([0,0.5)^{N \\times L})$ represents a noise realization sampled from a multivariate uniform distribution $U$ .", + "bbox": [ + 212, + 545, + 784, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, the sampling of tokens for masking is guided by $\\tilde{S}$ . We sort $\\tilde{S}$ in ascending order and select tokens corresponding to the top $K$ scores, where $K = \\lceil L\\cdot (1 - \\gamma)\\rceil$ . Consequently, an adaptive binary mask $M$ is constructed, where $M_{i} = 0$ for the $K$ selected tokens, and $M_{i} = 1$ for the remainder.", + "bbox": [ + 212, + 633, + 782, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The proposed token salience offers a more intuitive and cost-efficient strategy for determining which tokens to mask. SBM therefore revisits token dynamics in the image context, streamlining the conventional complex masking process and enabling strategic masking.", + "bbox": [ + 212, + 696, + 782, + 757 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 Evaluation on SBM", + "text_level": 1, + "bbox": [ + 215, + 779, + 442, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, the evaluation critically examines the robustness of the proposed SBM, especially against varying masking ratios, while also assessing its enhanced", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "SBAM", + "bbox": [ + 683, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a560adb51a404aa3450d15ace64f511d6d5c9e7d40d7b8b109344ae887068002.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 220, + 148, + 303, + 218 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b0619e75c53aa302265e480d7c4c1f6321ac32001d1103ff9f1d58a9a525775c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 315, + 148, + 400, + 218 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c6e3f1a641ad4460397fbb55a23860bd185c34fb36c46918fdc0d6de02f54052.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 410, + 148, + 495, + 218 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c84a942aa9f13d05966c61c6c0ff4e0cce5083ba49ac0d2a7d7dfec799251e10.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 148, + 589, + 218 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/71dd3a3aca882f066cadbdf64dc85ab9ae4136b6987edba0af510d158a41dbcb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 601, + 148, + 684, + 218 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/49cfdf792a3ef54b8f0cadd109c8fa7f9447fdfbf3381fe86779073116251d22.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 148, + 779, + 218 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3d43a9a8071b5895759e961d7de6b135f52dc3a1904df00e1201aae5bd1b9cd3.jpg", + "image_caption": [ + "Fig. 2: Qualitative example of SBM. SBM introduces 'token salience' to prioritize and mask tokens with high significance. Hence, it is qualitatively confirmed that particularly important objects with high contribution within the image are selectively masked. Moreover, by integrating randomness with token salience, masks are probabilistically assigned to the background and less significant tokens, enriching the diversity of the token masking." + ], + "image_footnote": [], + "bbox": [ + 220, + 224, + 303, + 294 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/362453e51f64e159de4494dfa60c2fa7b042f72f268d9b4c6d3f6048c0718333.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 315, + 224, + 400, + 294 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ae77002fdd85ed3c1b27059859a73b4f875bec6904f1cc2a7f42634de2a583c8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 410, + 224, + 495, + 294 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4ecb6741130c59b5ddbc5fa82de165083e161120aa6e0162215924853fe6c7bc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 224, + 589, + 294 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/06cb8ff24ce7280d283e68c9c1222080d209610faa20521abb4a602590d4060b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 601, + 224, + 684, + 294 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/081a5cb1e6cafd855ff7268d878a25bae32832363ca878b0d611e6efdb206fa9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 224, + 779, + 294 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "pre-training efficiency, evidenced by performance gains and faster convergence. Moreover, Fig. 2 qualitatively shows that SBM selectively masks only those tokens that contribute significantly to the image.", + "bbox": [ + 212, + 420, + 787, + 465 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1 Robustness to Masking Ratio Variability", + "text_level": 1, + "bbox": [ + 214, + 493, + 599, + 508 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A crucial aspect of the evaluation of SBM focuses on the robustness, particularly in the context of varying masking ratios. Prior established baselines, such as the widely used Masked Autoencoder (MAE) [14], often exhibit significant performance fluctuations with even minimal adjustments to the masking ratio. This sensitivity undermines the practicality and generalizability of such methods, especially in diverse real-world scenarios where optimal masking ratios may not be consistent across datasets. In contrast, SBM introduces a novel approach of selectively masking tokens based on their salience within an image. As a result, SBM exhibits a diminished likelihood of masking redundant or trivial tokens, thereby maintaining a stable performance across a broad spectrum of masking ratios.", + "bbox": [ + 212, + 523, + 787, + 688 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 3 showcases the performance stability of SBM against strong baseline MAE, underscoring its superior resilience to changes in masking ratios. Our evaluation leverages the Performance Improvement over Masking Ratio (PIMR) metric, a normalized measure that quantifies how each model's performance at a given masking ratio stands against its performance at the lowest ratio, thereby reflecting the relative improvement. This metric is instrumental in revealing the impact of increased masked data on model training. The PIMR is defined as follows:", + "bbox": [ + 212, + 691, + 787, + 809 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {P I M R} (M) = \\frac {P (M) - P \\left(M _ {\\min }\\right)}{P \\left(M _ {\\max }\\right) - P \\left(M _ {\\min }\\right)}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 810, + 785, + 844 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "H. Choi et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f98fad1803f38a36bea58c23afec47752f79df3785241737c185304648206fb8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 225, + 148, + 488, + 251 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/25fb0377f585ec315dcca4ea2019690ad216dd9235a809757dc1d640a1d2bb35.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 148, + 777, + 251 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/67e4dcc61139872b2aef9a08068c192c06c70f3c82f3f442217c3ee0fe7dd2e7.jpg", + "image_caption": [ + "Fig. 3: Evaluation of robustness across varied masking ratios. To evaluate the robustness of SBM, we report the comparative analysis of image classification performance on ImageNet-1K dataset [9] against the baseline method, MAE [14], using ViT-L [12] as a backbone. The upper graphs display the performance of the methods at different masking ratios, while the lower graphs illustrate the Performance Improvement over Masking Ratio (PIMR) and Global PIMR. These measures indicate the extent of each model's performance enhancement as the masking ratio increases from the lowest to higher ratios. SBM significantly outperforms MAE in every measures, demonstrating its superior effectiveness in handling various masking ratios and enhanced pre-training performances." + ], + "image_footnote": [], + "bbox": [ + 228, + 253, + 488, + 357 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/19e344ef5296371a9b90a01819cea30f479e0b289fe77c259cd15aade64de994.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 255, + 647, + 357 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/831248965af41e0a1622a634587aa5cf2f02916da50e653cd69bbe0036f7a21e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 648, + 255, + 776, + 357 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $P(M)$ is the performance at masking ratio $M$ . $P(M_{min})$ and $P(M_{max})$ denote the minimum and maximum observed performances, respectively. A PIMR value closer to 1 signifies a substantial improvement relative to the range of observed performances. The PIMR graph in classification accuracy demonstrate the robustness of the SBM strategy, where it maintains a competitive edge over MAE across various masking ratios. Notably, SBM shows a minimal performance drop at lower masking ratios compared to MAE, indicating its effectiveness even with sparse data presence.", + "bbox": [ + 212, + 540, + 787, + 662 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This advantage is particularly evident in the Global PIMR metric, where SBM's performance remains consistently high. For the Global PIMR calculation, we normalize performance relative to the most extensive range of performance observed among all models under comparison. This means that instead of comparing to the best and worst performances of a single model, Global PIMR considers the best and worst across both MAE and SBM, which provides a universal performance context. Thus, the metric reflects a model's improvement not in isolation but rather in relation to its peers, which can be seen in the equation:", + "bbox": [ + 212, + 666, + 787, + 787 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {G l o b a l} \\operatorname {P I M R} (M) = \\frac {P (M) - P \\left(M _ {G m i n}\\right)}{P \\left(M _ {G m a x}\\right) - P \\left(M _ {G m i n}\\right)}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 331, + 810, + 785, + 844 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "SBAM", + "bbox": [ + 683, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7ea3238ca914f38640f8ff25c457e094a269a97c513740e8ece77ba01da6b5e4.jpg", + "image_caption": [ + "Fig. 4: Performance evaluation of SBM with respect to the pre-trained epochs. We report the comparison of image classification accuracy on ImageNet-1K [9], pre-trained on ViT-L [12]. The left graph displays fine-tuning accuracy, whereas the right graph illustrates linear probing accuracy, both over a range of pre-trained epochs. The curves illustrate that SBM surpasses MAE [14] in pre-training effectiveness in every trained epoch, and also validates its quicker attainment of converged performance levels." + ], + "image_footnote": [], + "bbox": [ + 230, + 151, + 504, + 281 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3dc5994f371a561c6673aa90fdd0878807ea032cc758ee8facdd83523f2aa9c1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 151, + 774, + 281 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $P(M_{Gmax})$ and $P(M_{Gmin})$ are the global minimum and maximum performance values observed across both MAE and SBM models, respectively. This broader evaluation framework further underscores SBM's superior resilience and ability to maintain high accuracy across varying degrees of masking, outperforming the MAE baseline even when less information is available for learning.", + "bbox": [ + 212, + 393, + 787, + 470 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in the graphs, SBM consistently outperforms MAE across a spectrum of masking ratios, demonstrating its remarkable stability even at lower ratios. This is because in random masking, as the masking ratio decreases, the chance of including a crucial token in the mask lowers; conversely, SBM consistently masks pivotal tokens, regardless of the masking ratio.", + "bbox": [ + 212, + 470, + 787, + 546 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Furthermore, the classification accuracy graphs displayed above demonstrate that the application of SBM results in a notable enhancement in performance compared to MAE for all ratios. This reveals that the proposed SBM not only exhibits resilience to variations in the masking ratio but also significantly boosts pre-training efficacy, irrespective of the masking ratio. Strategically masking pivotal information enhances model performance and accelerates convergence by encouraging a comprehensive understanding of the visual context through a focus on essential tokens.", + "bbox": [ + 212, + 547, + 789, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 Enhanced Pre-training Efficiency", + "text_level": 1, + "bbox": [ + 214, + 691, + 535, + 708 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Beyond the robustness to masking ratio variations, SBM's efficacy is further demonstrated through its enhanced pre-training capabilities. Traditional MIM strategies often require extensive computational resources and time for model convergence, primarily due to the indiscriminate masking of image tokens which can hinder the learning process by obfuscating essential visual cues. By strategically selecting pivotal tokens for masking, SBM ensures that crucial tokens are leveraged during training, which leads to substantial improvement in pre-training performance and facilitates a more focused model convergence process.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "H. Choi et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a645b3183736e68dc8d4bc9b1cc6cf11e0be2cf216bf55e9579bb387b779c257.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 222, + 148, + 305, + 218 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/7c37280ad50cea46575d4947b8ee2c3712b2d94928ade23614ea7a83180cab2b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 316, + 148, + 401, + 218 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/49bb5f48b63b7a079f447c01d62fef6325888fb1850682a4903a06ea0b39fb4b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 411, + 150, + 495, + 218 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/92e652d4a17f5c3e986a2eea8a3989fc981d3db1697fb9cb53bcba6e5130263a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 148, + 589, + 218 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/5f0574eb78a0656b4e4308a39de0c33a71a2b2f25b64f9e0cdfc8961a22e3c77.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 602, + 148, + 684, + 218 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/0dcc4e5b4de537eed820e1c53cecf91fbf63517b82698666a7c32a14e8255170.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 148, + 779, + 218 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/5bc341637fed440cec1fcee5784766b9819ef29c6791c74ee23463ef3edbb9ec.jpg", + "image_caption": [ + "Fig. 5: Qualitative example of SBAM, which combines SBM and AMR. Having masking ratios that adapt throughout training is highly effective, as it allows the masking process to be tailored to each sample in the dataset, accommodating the unique composition and object sizes within each image, as shown in the above qualitative samples." + ], + "image_footnote": [], + "bbox": [ + 222, + 224, + 305, + 294 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/6da04f99d96b8e753059f6eee1ef6c22716e49d56010e2179b35cbcff0eeaa61.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 316, + 224, + 401, + 294 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/1fc530b9fd79b0b7351086e4d6003783bb06c6d5dbf0cbb9999ea83174e7d63d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 411, + 224, + 495, + 294 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/fb4ed28cceb34f79ccd7d4af6d4275ad09907532529d016de253a1d6c9f54390.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 224, + 589, + 294 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/75419a52424e7f759af059239dacc528dc2ab87d65be30223c614da88f7b4d5b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 602, + 224, + 684, + 294 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/d6f639e2d7bc7cba1176cfde6a7341e6ded0b2c774762e8b843fad45339967ee.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 224, + 779, + 294 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Fig. 4 showcases the comparative pre-training efficiency of SBM against MAE [14], underlining SBM's reduced pre-training duration without a tradeoff in accuracy. Initially, SBM secures a distinct lead in classification accuracy and continues to demonstrate this advantage across the training epochs, as seen in the left graph. The right graph, representing linear probing accuracy, further confirms SBM's higher initial performance, which stabilizes near peak levels well before 800 epochs. These findings highlight SBM's ability to prioritize significant features during early training stages, resulting in a new alternative for both convergence speed and pre-training performance in the pre-training of masked image models.", + "bbox": [ + 212, + 386, + 784, + 537 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 Adaptive Masking Ratio (AMR)", + "text_level": 1, + "bbox": [ + 215, + 564, + 565, + 582 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Achieving stability against changes in masking ratios has enabled us to advance the discourse on image masking, introducing a novel perspective: an Adaptive Masking Ratio (AMR). This innovative approach acknowledges the inherently diverse visual narratives presented by individual images and adjusts the masking ratio to fit different object sizes and classes within them.", + "bbox": [ + 212, + 599, + 784, + 672 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The proposed token salience $S = \\mathcal{N}(\\Sigma_{j=1}^{L}\\hat{A}_{\\cdot,j,:})$ forms the basis for determining AMRs. The AMR $R_{dyna}$ is computed based on the distribution of salience scores across tokens, adjusted by a predefined variability parameter $\\delta$ and the base mask ratio $r$ :", + "bbox": [ + 212, + 674, + 784, + 734 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nR _ {d y n a} = r - \\Delta r + 2 \\Delta r \\times \\operatorname {m e a n} \\left(1 _ {S > \\delta}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 362, + 750, + 784, + 766 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Here, $\\Delta r$ denotes the range of allowable variation in the masking ratio, as $R_{dyna}$ can range from $r - \\Delta r$ to $r + \\Delta r$ . $\\delta$ is the salience threshold for distinguishing highly salient tokens, and 1 is the indicator function that identifies tokens exceeding $\\delta$ .", + "bbox": [ + 214, + 779, + 784, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "SBAM", + "bbox": [ + 683, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/0dee5c96a07a79db0317ff25da0709d9bcaf3bdc80cd115a0765494620ead434.jpg", + "image_caption": [ + "Fig. 6: Performance comparison of SBAM with respect to the pre-trained epochs. We report the comparison of image classification accuracy on ImageNet-1K [9], pre-trained on ViT-L [12]. The left graph illustrates classification accuracy across epochs, while the right graph shows the accuracy obtained through linear probing. Both results indicate a significant improvement in pre-training performance when AMR is applied, which not only achieves higher accuracy earlier in the training process but also maintains a lead at convergence." + ], + "image_footnote": [], + "bbox": [ + 230, + 151, + 501, + 281 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/32c666333013583215b2fa259ba488783ad77485e71ba6d78e316589640bdba9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 151, + 772, + 281 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "With $R_{dyna}$ established, we adjust the number of tokens to be masked accordingly. This dynamic adjustment of masking ratios ensures that the masking process is not uniformly applied but is instead sensitive to the visual information's inherent salience, promoting a more effective learning mechanism by focusing on the informative segments of the image. Fig. 5 presents a qualitative example of AMR, demonstrating the effectiveness of adaptive masking ratios that customize the masking process for each dataset sample, thereby accounting for the unique composition and object sizes within each image. As a result, the proposed Saliency-Based Adaptive Masking (SBAM), which combines SBM and AMR, employs an adaptive approach, leading to enhanced performance across different models (See Fig. 6 and Tab. 1) and setting a new standard for tailored image masking.", + "bbox": [ + 212, + 422, + 787, + 604 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "6 Experiments", + "text_level": 1, + "bbox": [ + 215, + 628, + 375, + 647 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "6.1 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 662, + 455, + 678 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Our evaluation approach involves deploying the proposed SBM and AMR against the baseline to assess performance enhancements of the comprehensive method, SBAM. To ensure a fair comparison, we maintain consistency with the baseline method's hyperparameters and network architectures. Notably, to preserve the integrity of our experiments, we ensured uniform hardware utilization and experimental conditions across both our method and the reproduced baseline, all employing $8^{*}\\mathrm{A}6000$ GPUs. Given that this fixed GPU configuration diverges from those used in prior methods, discrepancies between our reproduced performance and the performance documented in existing papers may arise. For the fair experimental schedule, the reproduced 400-epoch performance of baseline", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "H. Choi et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/39b8bb661032b7d48fcfdcc1328e7100423267fa7f683e64104f4048f91a0efe.jpg", + "table_caption": [ + "Table 1: Comprehensive performance results of applying SBAM to various baseline methods. We report the comparison of image classification fine-tuning accuracy on ImageNet-1K [9] dataset. The consistent performance improvement of SBAM across various baseline methods demonstrates the efficacy of SBAM as a scalable methodology capable of enhancing a variety of MIM frameworks." + ], + "table_footnote": [], + "table_body": "
MethodBaselineBaseline+SBAM
MAE (ViT-L) [14]84.385.1
MAE (ViT-B)82.983.6
BootMAE (ViT-B) [11]84.184.8
iBoT (ViT-B) [43]71.574.4
CMAE (ViT-B) [17]83.884.5
", + "bbox": [ + 282, + 226, + 718, + 325 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "methods and the proposed method were all equally measured in intermediate stages in the training towards 800 epochs. A more detailed implementation description can be found in the Supplementary material.", + "bbox": [ + 212, + 353, + 787, + 398 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "6.2 Evaluation on SBAM", + "text_level": 1, + "bbox": [ + 214, + 420, + 439, + 434 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The graphs in Fig. 6 demonstrate the efficacy of the SBAM in comparison to the MAE model throughout the pre-training phase. Specifically, the left graph indicates that SBAM starts with a higher accuracy than MAE at the 200 epochs mark and continues to outperform MAE [14] at every subsequent checkpoint. By the 800 epochs mark, SBAM not only achieves a significant accuracy enhancement but also shows a more rapid improvement in the earlier epochs, suggesting that SBAM requires fewer epochs to achieve similar or better performance compared to MAE. In the context of linear probing, depicted in the right graph, the trend is similar. SBAM consistently achieves higher accuracy than MAE from the outset, and this performance gap is maintained as training progresses. The curves for SBAM and MAE become flat toward 800 epochs, indicating that 800 epochs or later is the point of convergence in performance. This observation underscores SBAM's substantial advantage over MAE, both in terms of converged performance and convergence speed. Overall, the performance trends captured in these graphs suggest that SBAM is more efficient during pre-training, reaching higher levels of accuracy faster than MAE.", + "bbox": [ + 212, + 445, + 787, + 686 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "6.3 Evaluation on Various Baselines", + "text_level": 1, + "bbox": [ + 214, + 709, + 527, + 724 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The integration of SBAM into various baseline methodologies demonstrates a notable enhancement in pre-training performance, as summarized in Tab. 1. All experiments report the classification accuracy on 400 epochs, except for iBoT [43] which is pre-trained for 100 epochs. Applying SBAM to the large-scale variant of MAE [14] (MAE (ViT-L)) yields a noteworthy enhancement, elevating the baseline accuracy from $84.3\\%$ to $85.1\\%$ , marking a significant advancement. Integration of SBAM to the MAE using ViT-B also experienced significant performance", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "SBAM", + "bbox": [ + 683, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/b35f10a6c36ae51020c23b54d327d45c0f98280de54506a4376cf50a2b2417fd.jpg", + "table_caption": [ + "Table 2: Comparative evaluation of SBAM against the state-of-the-art masking strategy AMT [25]." + ], + "table_footnote": [], + "table_body": "
MethodAcc (%)
AM [25]82.5
AMT [25]82.8
SBAM83.6
", + "bbox": [ + 261, + 193, + 447, + 263 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/57c0c628851955d0fbeae22447b197121bc70153b25c9396ac1a6e5de48648af.jpg", + "table_caption": [ + "Table 3: Comparative evaluation of SBAM against the state-of-the-art masking strategy AttMask [18]." + ], + "table_footnote": [], + "table_body": "
MethodAcc (%)
AttMask-High [18]72.5
AttMask-Hint [18]72.8
SBAM74.4
", + "bbox": [ + 516, + 193, + 759, + 265 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/cbfd74e9497fcf453434baef5780b39e0832b2bc081f2442a9dafbf31bcc711c.jpg", + "image_caption": [ + "Fig. 7: Comprehensive ablation studies of the impact of randomness, variance $(\\Delta r)$ , and delta $(\\delta)$ of the proposed SBAM approach." + ], + "image_footnote": [], + "bbox": [ + 233, + 287, + 410, + 417 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/55beaab6d230ca3654e175ec57dee0e96c9a2968fa60094c540bb4edbe4e98aa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 421, + 287, + 591, + 416 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/8a78552668ab7a6eaa6000b19c11cdeb2f422e232d98131ea7e7e5c4cd218373.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 287, + 772, + 416 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "gains, underscoring that the benefits of SBAM are not limited to specific model architectures and can provide substantial pre-training efficacy for a variety of models. The performance enhancement of SBAM on various models including BootMAE [11], iBoT, and CMAE [17] highlights SBAM's generalizability and efficacy in augmenting various model structures with significant accuracy gains.", + "bbox": [ + 212, + 500, + 787, + 575 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In conclusion, the consistent improvement across various baselines validates the efficacy of SBAM as a scalable enhancement tool. It not only boosts performance in standard settings but also bridges the gap in more challenging learning scenarios, marking it as a pivotal development in masked image modeling pretraining techniques.", + "bbox": [ + 212, + 577, + 787, + 652 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "6.4 Comparison with Masking Methods", + "text_level": 1, + "bbox": [ + 214, + 680, + 558, + 696 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We conducted a comparative evaluation of our SBAM masking strategy against the established masking strategies of AMT and AttMask. This comparison was performed by applying SBAM to the baselines previously employed by AMT [25] and AttMask [18] methods, specifically MAE [14] (ViT-B, 400 epochs) and iBoT [43] (ViT-B, 100 epochs), to ascertain the enhancements introduced by our approach.", + "bbox": [ + 212, + 703, + 787, + 794 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In Tab. 2, the AM and AMT strategy [25], which implement basic masking and selective masking based on semantic importance, achieve an image classification fine-tuning accuracy of 82.5 and 82.8, respectively. When applied to the", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "H. Choi et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "same baseline model, SBAM outperforms both approaches by achieving an accuracy of 83.6. This implies that the efficacy of self-supervised pre-training can be maximized by defining token salience based on the outgoing weight of the token, as opposed to the incoming weight [25].", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Furthermore, we compare SBAM to two variants of AttMask [18], applied within the iBoT [43] framework in Tab. 3. Both AttMask-High and AttMask-Hint incorporate the selective masking strategy of distillation setup which leverages similarity to classification token. SBAM stands out with an accuracy of $74.4\\%$ , substantially higher than AttMask-High's $72.5\\%$ and AttMask-Hint's $72.8\\%$ . This highlights the superiority of the SBAM method, which can effectively improve the pre-training efficiency without the need for the additional computational cost of using a complex framework.", + "bbox": [ + 212, + 207, + 787, + 328 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "6.5 Ablation studies", + "text_level": 1, + "bbox": [ + 214, + 345, + 398, + 359 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In Fig. 7, we provide comprehensive ablation studies of the impact of randomness, variance $(\\varDelta r)$ , and delta $(\\delta)$ of the proposed SBAM approach. We report the ImageNet-1K [9] classification accuracy achieved by SBAM using the baseline approach of MAE [14], trained for 400 epochs on ViT-L [12] as the backbone. The first graph depicts the impact of the incorporated randomness in the SBAM on model performance over various epochs. The plot reveals that integrating randomness with token saliency markedly enhances pre-training accuracy as the number of pre-trained epochs increases. The second and third graphs show the fine-tuning accuracy and linear probing performance ablations for the hyperparameters of SBAM. While fine-tuning accuracy remained stable across various hyperparameters, linear probing accuracy demonstrated relative sensitivity. We chose $\\varDelta r=0.15$ and $\\delta=0.1$ as optimal hyperparameters in both figures and were universally applicable across all baseline methodologies.", + "bbox": [ + 212, + 364, + 787, + 563 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "7 Related Work", + "text_level": 1, + "bbox": [ + 214, + 580, + 388, + 595 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "7.1 Masked Language Modeling", + "text_level": 1, + "bbox": [ + 214, + 607, + 493, + 625 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Masked Language Modeling (MLM) [2,7,8,10,13,24,31,33,34,40] has become a keystone self-supervised learning paradigm in NLP, exemplified by groundbreaking models like BERT [10] and GPT [29,30]. By predicting masked tokens from their context, MLM has propelled NLP forward, enabling models to scale and perform adeptly on diverse tasks [3]. However, the considerable training time and computational demands of these models have spurred innovations aimed at increasing pre-training efficiency. For example, ALBERT [19] reduced parameters through embedding matrix factorization and shared layer parameters, while EarlyBERT [6] applied the principles of network pruning to optimize the training process. The curriculum masking approach of CCM [20] represents another advancement, strategically increasing the complexity of token masking to enhance learning. These efforts reflect a broader trend in the quest for efficiency, leading to models that maintain or exceed the capabilities of their predecessors with a fraction of the resource investment.", + "bbox": [ + 212, + 628, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "SBAM", + "bbox": [ + 683, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "7.2 Masked Image Modeling", + "text_level": 1, + "bbox": [ + 215, + 146, + 465, + 162 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In computer vision, Masked Image Modeling (MIM)) [1,4,5,11,14-16,23,26-28, 36,37,41-43] has emerged as a transformative technique, drawing parallels to the success of Masked Language Modeling (MLM) in NLP. MIM's central tenet involves predicting occluded parts of images to foster a nuanced understanding of visual content sans explicit labels. Early efforts adapting MLM concepts for visual data, such as iGPT [5], paved the way for more sophisticated methods. BEiT [1] utilized a pre-trained discrete variational autoencoder (dVAE) to produce target visual tokens. Further refinements in the technique have been observed in methods such as MAE [14] and SimMIM [36], which focus on direct prediction from unmasked image areas, refining the process of visual understanding. Efficiency in pre-training has been a critical frontier, leading to innovations like GreenMIM [16] and HiViT [42], which optimize hierarchical Vision Transformers (ViTs) by processing only the visible patches, significantly reducing computational overhead.", + "bbox": [ + 212, + 169, + 787, + 380 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Recent advances in model pre-training have honed in on the strategic use of masking to enhance learning efficiency, with a particular emphasis on selecting which image regions to mask. Initiating this trend, ADIOS [32] leverages adversarial training to smartly select challenging segments for masking, setting a foundation for intelligent masking approaches. AttMask [18] and SemMAE [21] further this by utilizing self-attention and semantic information, respectively, to pinpoint and mask the most informative parts of an image, thereby prioritizing high-value areas over random masking. The Attention-Driven Masking and Throwing Strategy (AMT) strategy [25] refines this focus on semantics by employing self-attention to identify and eliminate redundant patches, achieving a delicate balance between precision and efficiency. In the realm of CLIP models, ACLIP [38] and Fast CLIP (FLIP) [22] adopt attentive masking strategies to optimize training, with Fast CLIP demonstrating the effectiveness of masking substantial portions of images for accelerated learning. MaskAlign [37] introduces an innovative teacher-student framework that bypasses the need for masked region reconstruction, aligning visible features with semantically rich intact image features to concentrate on the most informative parts. Together, these approaches illustrate a shift towards more strategic, intelligent masking techniques, significantly boosting the pre-training process by leveraging both the quantity and quality of masked inputs.", + "bbox": [ + 212, + 380, + 787, + 683 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "8 Conclusions", + "text_level": 1, + "bbox": [ + 215, + 704, + 367, + 720 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The proposed Saliency-Based Adaptive Masking (SBAM) approach and the Adaptive Masking Ratio (AMR) significantly progress the field of MIM by introducing a method that adaptively masks image tokens with dynamic masking ratios based on their token salience. SBAM not only enhances pre-training efficiency and model performance on ImageNet-1K datasets but also introduces a novel way of considering the importance of token dynamics, thereby enabling models to learn more pivotal representations.", + "bbox": [ + 212, + 733, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "H. Choi et al.", + "bbox": [ + 271, + 114, + 362, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 323, + 160 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bao, H., Dong, L., Piao, S., Wei, F.: Beit: Bert pre-training of image transformers. In: International Conference on Learning Representations (2021)", + "2. Bao, H., Dong, L., Wei, F., Wang, W., Yang, N., Liu, X., Wang, Y., Gao, J., Piao, S., Zhou, M., et al.: Unilmv2: Pseudo-masked language models for unified language model pre-training. In: International conference on machine learning. pp. 642-652. PMLR (2020)", + "3. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)", + "4. Cao, S., Xu, P., Clifton, D.A.: How to understand masked autoencoders. arXiv preprint arXiv:2202.03670 (2022)", + "5. Chen, M., Radford, A., Child, R., Wu, J., Jun, H., Luan, D., Sutskever, I.: Generative pretraining from pixels. In: International conference on machine learning. pp. 1691-1703. PMLR (2020)", + "6. Chen, X., Cheng, Y., Wang, S., Gan, Z., Wang, Z., Liu, J.: Earlybert: Efficient bert training via early-bird lottery tickets. arXiv preprint arXiv:2101.00063 (2020)", + "7. Clark, K., Luong, M.T., Le, Q.V., Manning, C.D.: Electra: Pre-training text encoders as discriminators rather than generators. arXiv preprint arXiv:2003.10555 (2020)", + "8. Conneau, A., Lample, G.: Cross-lingual language model pretraining. Advances in neural information processing systems 32 (2019)", + "9. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)", + "0. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)", + "1. Dong, X., Bao, J., Zhang, T., Chen, D., Zhang, W., Yuan, L., Chen, D., Wen, F., Yu, N.: Bootstrapped masked autoencoders for vision bert pretraining. In: European Conference on Computer Vision. pp. 247-264. Springer (2022)", + "2. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)", + "3. Ghazvininejad, M., Levy, O., Liu, Y., Zettlemoyer, L.: Mask-predict: Parallel decoding of conditional masked language models. arXiv preprint arXiv:1904.09324 (2019)", + "4. He, K., Chen, X., Xie, S., Li, Y., Dólar, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16000-16009 (2022)", + "5. Hou, Z., Sun, F., Chen, Y.K., Xie, Y., Kung, S.Y.: Milan: Masked image pretraining on language assisted representation. arXiv preprint arXiv:2208.06049 (2022)", + "6. Huang, L., You, S., Zheng, M., Wang, F., Qian, C., Yamasaki, T.: Green hierarchical vision transformer for masked image modeling. arXiv preprint arXiv:2205.13515 (2022)", + "7. Huang, Z., Jin, X., Lu, C., Hou, Q., Cheng, M.M., Fu, D., Shen, X., Feng, J.: Contrastive masked autoencoders are stronger vision learners. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)" + ], + "bbox": [ + 225, + 176, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "SBAM", + "bbox": [ + 683, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Kakogeorgiou, I., Gidaris, S., Psomas, B., Avrithis, Y., Bursuc, A., Karantzalos, K., Komodakis, N.: What to hide from your students: Attention-guided masked image modeling. arXiv preprint arXiv:2203.12719 (2022)", + "19. Lan, Z., Chen, M., Goodman, S., Gimpel, K., Sharma, P., Soricut, R.: Albert: A lite bert for self-supervised learning of language representations. In: International Conference on Learning Representations", + "20. Lee, M., Park, J.H., Kim, J., Kim, K.M., Lee, S.: Efficient pre-training of masked language model via concept-based curriculum masking. In: Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing (2022)", + "21. Li, G., Zheng, H., Liu, D., Wang, C., Su, B., Zheng, C.: Semmae: Semantic-guided masking for learning masked autoencoders. Advances in Neural Information Processing Systems 35, 14290-14302 (2022)", + "22. Li, Y., Fan, H., Hu, R., Feichtenhofer, C., He, K.: Scaling language-image pretraining via masking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23390-23400 (2023)", + "23. Liu, J., Huang, X., Liu, Y., Li, H.: Mixmim: Mixed and masked image modeling for efficient visual representation learning. arXiv preprint arXiv:2205.13137 (2022)", + "24. Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D., Levy, O., Lewis, M., Zettlemoyer, L., Stoyanov, V.: Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692 (2019)", + "25. Liu, Z., Gui, J., Luo, H.: Good helper is around you: Attention-driven masked image modeling. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 1799-1807 (2023)", + "26. Pan, J., Zhou, P., Yan, S.: Towards understanding why mask-reconstruction pretraining helps in downstream tasks. arXiv preprint arXiv:2206.03826 (2022)", + "27. Peng, Z., Dong, L., Bao, H., Ye, Q., Wei, F.: Beit v2: Masked image modeling with vector-quantized visual tokenizers. arXiv preprint arXiv:2208.06366 (2022)", + "28. Peng, Z., Dong, L., Bao, H., Ye, Q., Wei, F.: A unified view of masked image modeling. arXiv preprint arXiv:2210.10615 (2022)", + "29. Radford, A., Narasimhan, K., Salimans, T., Sutskever, I., et al.: Improving language understanding by generative pre-training (2018)", + "30. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners. OpenAI blog 1(8), 9 (2019)", + "31. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research 21(1), 5485-5551 (2020)", + "32. Shi, Y., Siddharth, N., Torr, P., Kosiorek, A.R.: Adversarial masking for self-supervised learning. In: International Conference on Machine Learning. pp. 20026-20040. PMLR (2022)", + "33. Song, K., Tan, X., Qin, T., Lu, J., Liu, T.Y.: Mass: Masked sequence to sequence pre-training for language generation. In: International Conference on Machine Learning. pp. 5926-5936. PMLR (2019)", + "34. Song, K., Tan, X., Qin, T., Lu, J., Liu, T.Y.: Mpnet: Masked and permuted pretraining for language understanding. Advances in Neural Information Processing Systems 33, 16857-16867 (2020)", + "35. Wu, J., Mo, S.: Object-wise masked autoencoders for fast pre-training. arXiv preprint arXiv:2205.14338 (2022)", + "36. Xie, Z., Zhang, Z., Cao, Y., Lin, Y., Bao, J., Yao, Z., Dai, Q., Hu, H.: Simmim: A simple framework for masked image modeling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9653-9663 (2022)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "H. Choi et al.", + "bbox": [ + 271, + 114, + 362, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "37. Xue, H., Gao, P., Li, H., Qiao, Y., Sun, H., Li, H., Luo, J.: Stare at what you see: Masked image modeling without reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22732-22741 (2023)", + "38. Yang, Y., Huang, W., Wei, Y., Peng, H., Jiang, X., Jiang, H., Wei, F., Wang, Y., Hu, H., Qiu, L., et al.: Attentive mask clip. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2771-2781 (2023)", + "39. Yi, K., Ge, Y., Li, X., Yang, S., Li, D., Wu, J., Shan, Y., Qie, X.: Masked image modeling with denoising contrast. arXiv preprint arXiv:2205.09616 (2022)", + "40. Zaken, E.B., Ravfogel, S., Goldberg, Y.: Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. arXiv preprint arXiv:2106.10199 (2021)", + "41. Zhang, Q., Wang, Y., Wang, Y.: How mask matters: Towards theoretical understandings of masked autoencoders. arXiv preprint arXiv:2210.08344 (2022)", + "42. Zhang, X., Tian, Y., Xie, L., Huang, W., Dai, Q., Ye, Q., Tian, Q.: Hivit: A simpler and more efficient design of hierarchical vision transformer. In: The Eleventh International Conference on Learning Representations (2023)", + "43. Zhou, J., Wei, C., Wang, H., Shen, W., Xie, C., Yuille, A., Kong, T.: ibot: Image bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832 (2021)" + ], + "bbox": [ + 212, + 146, + 787, + 397 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "SBAM", + "bbox": [ + 683, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Salience-Based Adaptive Masking_ Revisiting Token Dynamics for Enhanced Pre-training/05c12db4-7790-4b9e-84f8-e76fa8191a90_model.json b/2024/Salience-Based Adaptive Masking_ Revisiting Token Dynamics for Enhanced Pre-training/05c12db4-7790-4b9e-84f8-e76fa8191a90_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3297c7f40483622ff1177f40bdaec705f841f653 --- /dev/null +++ b/2024/Salience-Based Adaptive Masking_ Revisiting Token Dynamics for Enhanced Pre-training/05c12db4-7790-4b9e-84f8-e76fa8191a90_model.json @@ -0,0 +1,2434 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.245, + 0.141, + 0.761, + 0.187 + ], + "angle": 0, + "content": "Salience-Based Adaptive Masking: Revisiting Token Dynamics for Enhanced Pre-training" + }, + { + "type": "text", + "bbox": [ + 0.33, + 0.213, + 0.674, + 0.247 + ], + "angle": 0, + "content": "Hyesong Choi\\(^{1}\\), Hyejin Park\\(^{1}\\), Kwang Moo Yi\\(^{2}\\), Sungmin Cha\\(^{3}\\), and Dongbo Min\\(^{1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.391, + 0.255, + 0.613, + 0.283 + ], + "angle": 0, + "content": "1 Ewha W. University \n2 University of British Columbia" + }, + { + "type": "text", + "bbox": [ + 0.424, + 0.283, + 0.58, + 0.298 + ], + "angle": 0, + "content": "3 New York University" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.34, + 0.744, + 0.549 + ], + "angle": 0, + "content": "Abstract. In this paper, we introduce Saliency-Based Adaptive Masking (SBAM), a novel and cost-effective approach that significantly enhances the pre-training performance of Masked Image Modeling (MIM) approaches by prioritizing token salience. Our method provides robustness against variations in masking ratios, effectively mitigating the performance instability issues common in existing methods. This relaxes the sensitivity of MIM-based pre-training to masking ratios, which in turn allows us to propose an adaptive strategy for 'tailored' masking ratios for each data sample, which no existing method can provide. Toward this goal, we propose an Adaptive Masking Ratio (AMR) strategy that dynamically adjusts the proportion of masking for the unique content of each image based on token salience. We show that our method significantly improves over the state-of-the-art in mask-based pre-training on the ImageNet-1K dataset. Code and model parameters are available at https://github.com/doihye/SBAM." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.563, + 0.744, + 0.591 + ], + "angle": 0, + "content": "Keywords: Self-supervised learning \\(\\cdot\\) Masked image modeling \\(\\cdot\\) Masked autoencoder" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.619, + 0.377, + 0.635 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.654, + 0.788, + 0.747 + ], + "angle": 0, + "content": "Recent drastic improvements in various Computer Vision tasks rely heavily on Transformer architectures [12]. A critical component that enables these architectures is the necessity of large-scale data [9], which is not always readily available. Naturally, pre-training with pretext tasks has become a popular solution as a workaround, represented by Masked Image Modeling (MIM) [1, 11, 14, 36], inspired by how Masked Language Modeling (MLM) [2, 7, 10, 24] has reshaped the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.756, + 0.79, + 0.842 + ], + "angle": 0, + "content": "This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (No. RS-2023-00222385) and the National Research Foundation of Korea, Republic of Korea (2022M3H9A2083956) and from the Korea Basic Science Institute, Republic of Korea (National Research Facilities and Equipment Center) funded by the Ministry of Education, Republic of Korea (2019R1A6C1010020). † Corresponding author: dbmin@ewha.ac.kr" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "H. Choi et al." + }, + { + "type": "image", + "bbox": [ + 0.254, + 0.148, + 0.753, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.334, + 0.788, + 0.404 + ], + "angle": 0, + "content": "Fig. 1: Overview of SBAM. Whereas (a) Random Masking must rely completely on chance and carefully tuned masking ratio to guarantee effective masking, (b) the proposed SBAM strategically masks tokens based on the token salience. The robustness of SBAM paved the way for the introduction of (c) AMR, which implements a tailored masking ratio for each sample in the dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.415, + 0.785, + 0.461 + ], + "angle": 0, + "content": "Natural Language Processing landscape. These strategies involve masking a subset of the input data and predicting those that have been hidden, thus forcing the deep network to infer underlying concepts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.788, + 0.597 + ], + "angle": 0, + "content": "While MIM has significantly advanced self-supervised learning in vision, its conventional approach to randomly selecting tokens for masking falls short of harnessing the full potential of visual data [18]. Unlike in text, where randomness might obscure key semantic units, the visual domain's complexity and token redundancy demand a more strategic masking protocol to ensure model comprehension. This necessity prompts us to explore a refined masking methodology, targeting a selection process that achieves the goal of image understanding for pre-trained models, thereby bridging the gap in modality-specific pre-training strategies." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.789, + 0.844 + ], + "angle": 0, + "content": "Various techniques [18, 25, 35] have thus been proposed for more effective masking. While these methods all strive toward improved pre-training, an oversight shared amongst all methods is that they do not regard the contribution of each token within the overall composition of the image. It is crucial to scrutinize the interconnections between tokens to ensure that the token masking includes the tokens that play a pivotal role within the image. Moreover, prior masking methodologies [18, 25, 35] fail to address the critical consideration of the masking ratio, a factor that ought to dynamically adjust in response to the size and quantity of pivotal objects embedded within the image. This is primarily due to the difficulty reported in various works [14, 36, 39] that minor changes in the optimized masking ratio can lead to performance instability, rendering such considerations difficult to implement. Thus, these methods must rely completely on chance and carefully tuned masking ratio to guarantee effective masking. Besides, contemporary masking strategies [18, 25, 35] often grapple with high complexity, burdened by intricate distillation frameworks [18], auxiliary detection processes [35], and duplicated forward processes [25]." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.684, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SBAM" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.329 + ], + "angle": 0, + "content": "In this work, we introduce a simple yet novel approach that focuses on token dynamics, termed Salience-Based Adaptive Masking (SBAM), which aims to strategically select masking locations by discerning perceptual prominence within the visual data. Specifically, the proposed method leverages the directional emphasis from attention mechanisms to identify the image tokens pivotal to the visual context. Significantly, our methodology stands apart from the conventional attention-based approach [25] by leveraging the token's outgoing weight to calculate its 'token salience' within the image and prioritizing those that have high salience to be masked. We also infuse a degree of randomness into token salience to enrich the diversity of mask generation. Further details of SBAM are elaborated in Sec. 3. Thus, without any significant additional cost, we can consider the token's prominence within the image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.332, + 0.788, + 0.483 + ], + "angle": 0, + "content": "Crucially, the strength of the SBAM approach lies in its robustness to the varying masking ratios—a notable vulnerability in established baselines [14,36, 39], which struggles from performance instability with even minor variations in optimized masking ratio. The robustness of SBAM is attributable to the proposed analytical precision in token dynamics, selectively masking tokens pivotal to the image's entirety. As a result, SBAM exhibits a diminished likelihood of masking redundant or non-essential segments, offering a stable alternative to the random masking or preceding masking approaches. Along with improving robustness to mask ratio variations, SBAM significantly enhances pre-training efficiency. A comprehensive evaluation of this is provided in Sec. 4." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.486, + 0.788, + 0.713 + ], + "angle": 0, + "content": "Establishing robustness against variations in masking ratios has empowered us to expand the discourse on image masking into a pioneering aspect, introducing an innovative paradigm: an Adaptive Masking Ratio (AMR). We find having masking ratios that adapt throughout training to be highly effective, as it allows the masking process to be tailored to each sample in the dataset. For instance, each image may benefit from different masking ratios, as one image might have a close-up of a bear which would enjoy high masking ratios, whereas one might have an airplane in the sky that would require a lower masking ratio; see (c) of Fig. 1. The proposed token salience, reapplied in this context, serves to finely determine the dynamic masking ratio by quantifying the proportion of tokens exhibiting high salience. More details can be found in Sec. 5. The proposed AMR implements an adaptive strategy that respects the distinctiveness of visual data and thus achieves significant performance gains when applied to various baselines (refer to Sec. 6.2 and 6.3. We note that this type of 'tailoring' to each image for pre-training is impossible with any existing method." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.716, + 0.788, + 0.822 + ], + "angle": 0, + "content": "In sum, this paper offers an innovative masking strategy that enhances the robustness and effectiveness of pre-trained models, setting the stage for a notable shift in the field of self-supervised learning. More importantly, the proposed method can be universally applied across any MIM framework that exploits token masking. We evaluate our SBAM and AMR on data-hungry models like ViT-L/ViT-B [12] on ImageNet-1K [9] datasets, achieving significant performance improvements in both fine-tuning and linear probing accuracy." + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.825, + 0.511, + 0.84 + ], + "angle": 0, + "content": "To summarize, our contributions are:" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.128 + ], + "angle": 0, + "content": "H. Choi et al." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.147, + 0.784, + 0.175 + ], + "angle": 0, + "content": "- we present Saliency-Based Adaptive Masking (SBAM), a novel effective method for MIM pre-training that focuses on token salience;" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.178, + 0.784, + 0.205 + ], + "angle": 0, + "content": "- being saliency-based, without a significant increase in computation, we allow effective masking that is robust to masking ratios:" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.208, + 0.784, + 0.236 + ], + "angle": 0, + "content": "- empowered by the robustness to masking ratio, we propose an Adaptive Masking Ratio (AMR) that allows tailored masking for each sample;" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.239, + 0.784, + 0.267 + ], + "angle": 0, + "content": "- we evaluate our method on ImageNet-1K datasets, achieving notable enhancements in both fine-tuning and linear probing accuracy." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.147, + 0.784, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.293, + 0.382, + 0.309 + ], + "angle": 0, + "content": "2 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.326, + 0.784, + 0.385 + ], + "angle": 0, + "content": "Within the Masked Image Modeling (MIM) domain, the crux lies in the strategic corruption and subsequent reconstruction of image segments. This process hinges on two core operations: Random Masking of image tokens and Reconstruction of corrupted tokens to guide the learning process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.387, + 0.784, + 0.446 + ], + "angle": 0, + "content": "For a given image sequence \\(X \\in \\mathbb{R}^{N \\times L \\times D}\\), where \\(N, L\\) and \\(D\\) denote batch size, number of tokens per image and dimensionality of each token respectively, and specified masking proportion \\(\\gamma\\), the general random masking process conducted by \\(\\varPhi_{\\mathrm{mask}}\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.407, + 0.462, + 0.784, + 0.476 + ], + "angle": 0, + "content": "\\[\nX _ {\\text {m a s k e d}}, M = \\Phi_ {\\text {m a s k}} (X, \\gamma), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.489, + 0.785, + 0.533 + ], + "angle": 0, + "content": "where \\( X_{\\mathrm{masked}} \\) represents the visible tokens as a result of the post-application of the random mask \\( M \\), with \\( M \\in \\{0,1\\}^{N \\times L} \\) indicating the presence (1) or absence (0) of masking for each element." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.535, + 0.784, + 0.593 + ], + "angle": 0, + "content": "The key component for learning in MIM is the reconstruction error \\(\\mathcal{L}_{\\mathrm{MIM}}\\), which is computed by the mean squared error (MSE) between the predicted representation of masked tokens, denoted as \\(\\hat{X}\\), and their normalized original counterparts, represented by \\(X\\):" + }, + { + "type": "equation", + "bbox": [ + 0.266, + 0.608, + 0.784, + 0.648 + ], + "angle": 0, + "content": "\\[\n\\bar {X} = \\frac {X - \\mu (X)}{\\sigma (X) + \\epsilon}, \\quad \\mathcal {L} _ {\\mathrm {M I M}} = \\frac {1}{\\sum_ {i , j} M _ {i , j}} \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {L} M _ {i, j} \\cdot \\| \\hat {X} _ {i, j} - \\bar {X} _ {i, j} \\| _ {2} ^ {2}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.661, + 0.784, + 0.735 + ], + "angle": 0, + "content": "Here, \\(\\mu (\\cdot)\\) and \\(\\sigma (\\cdot)\\) represent the mean and standard deviation of \\(X\\) respectively. \\(\\| \\cdot \\| _2^2\\) denotes the squared \\(L_{2}\\) norm and \\(\\epsilon\\) is a small value for numerical stability. This formulation presents the core of MIM's training objective, focusing explicitly on the reconstruction of the masked portions of the input, encouraging the model to infer corrupted information from the unmasked context." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.762, + 0.557, + 0.78 + ], + "angle": 0, + "content": "3 Salience-Based Masking (SBM)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Masked Image Modeling (MIM) has greatly pushed forward self-supervised learning in the visual domain, yet its standard practice of randomly masking tokens fails to fully capture the richness of visual information. To overcome this" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.684, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SBAM" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "limitation, we present Salience-Based Masking (SBM), a novel masking methodology that selects masking locations by revisiting token dynamics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.177, + 0.785, + 0.253 + ], + "angle": 0, + "content": "Given an input tensor \\(X \\in \\mathbb{R}^{N \\times (L \\times D)}\\), the first step involves computing an affinity matrix \\(\\mathcal{A} \\in \\mathbb{R}^{N \\times L \\times L}\\) through a batch matrix-matrix product between \\(X\\) and \\(X' \\in \\mathbb{R}^{N \\times (D \\times L)}\\). Subsequently, we apply the softmax function \\(\\rho(\\cdot)\\) to the affinity matrix, resulting in a normalized affinity matrix denoted as \\(\\hat{\\mathcal{A}} = \\rho(\\mathcal{A}) \\in \\mathbb{R}^{N \\times L \\times L}\\). The softmax function is implemented as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.413, + 0.263, + 0.786, + 0.299 + ], + "angle": 0, + "content": "\\[\n\\rho (\\mathcal {A}) _ {n, i, j} = \\frac {e ^ {a _ {n , i , j}}}{\\sum_ {k = 1} ^ {L} e ^ {a _ {n , i , k}}}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.302, + 0.784, + 0.347 + ], + "angle": 0, + "content": "Here, \\(a_{n,i,j}\\) represents an element of \\(\\mathcal{A}\\) and \\(e\\) is the base of the natural logarithm. Note that, for each \\(n\\)-th element with a shape of \\(\\mathbb{R}^{L\\times L}\\), this function normalizes the rows of it, transforming them into probabilities that sum to 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.348, + 0.785, + 0.407 + ], + "angle": 0, + "content": "Crucially, our approach distinguishes itself from conventional attention-based methods by utilizing the sum of outgoing weight of each token to determine the 'token salience' \\( S \\in \\mathbb{R}^{N \\times L} \\) in the image, which is represented by the column-wise summed score of \\( \\hat{\\mathcal{A}} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.315, + 0.416, + 0.786, + 0.448 + ], + "angle": 0, + "content": "\\[\nS = \\mathcal {N} \\left(\\Sigma_ {j = 1} ^ {L} \\hat {\\mathcal {A}} _ {:, j,:}\\right), \\quad \\mathcal {N} (x) = \\frac {x - \\operatorname* {m i n} (x)}{\\operatorname* {m a x} (x) - \\operatorname* {m i n} (x)}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.456, + 0.784, + 0.546 + ], + "angle": 0, + "content": "In the affinity map, the row-wise score represents the incoming weight, reflecting the extent to which other tokens influence the corresponding token. Conversely, the column-wise score signifies the outgoing weight, indicating the contribution of the corresponding token to others. By summing these scores, it becomes feasible to quantify the token's overall impact on the image, thereby defining token salience." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.546, + 0.785, + 0.635 + ], + "angle": 0, + "content": "The adaptive masking process of SBM is formulated with the token salience \\( S \\). Notably, exclusive reliance on \\( S \\) for masking can precipitate a decline in performance (refer to Fig. 7). To mitigate this issue, we incorporated an element of randomness into the masking process to get the adjusted token salience, denoted as \\( \\tilde{S} = S + N \\), where \\( N \\sim U([0,0.5)^{N \\times L}) \\) represents a noise realization sampled from a multivariate uniform distribution \\( U \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.635, + 0.784, + 0.697 + ], + "angle": 0, + "content": "Then, the sampling of tokens for masking is guided by \\(\\tilde{S}\\). We sort \\(\\tilde{S}\\) in ascending order and select tokens corresponding to the top \\(K\\) scores, where \\(K = \\lceil L\\cdot (1 - \\gamma)\\rceil\\). Consequently, an adaptive binary mask \\(M\\) is constructed, where \\(M_{i} = 0\\) for the \\(K\\) selected tokens, and \\(M_{i} = 1\\) for the remainder." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.697, + 0.784, + 0.758 + ], + "angle": 0, + "content": "The proposed token salience offers a more intuitive and cost-efficient strategy for determining which tokens to mask. SBM therefore revisits token dynamics in the image context, streamlining the conventional complex masking process and enabling strategic masking." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.78, + 0.444, + 0.795 + ], + "angle": 0, + "content": "4 Evaluation on SBM" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "In this section, the evaluation critically examines the robustness of the proposed SBM, especially against varying masking ratios, while also assessing its enhanced" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "H. Choi et al." + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.149, + 0.305, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.316, + 0.149, + 0.401, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.149, + 0.496, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.15, + 0.591, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.15, + 0.686, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.15, + 0.781, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.226, + 0.305, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.316, + 0.226, + 0.401, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.226, + 0.496, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.226, + 0.591, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.226, + 0.686, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.226, + 0.781, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.309, + 0.788, + 0.392 + ], + "angle": 0, + "content": "Fig. 2: Qualitative example of SBM. SBM introduces 'token salience' to prioritize and mask tokens with high significance. Hence, it is qualitatively confirmed that particularly important objects with high contribution within the image are selectively masked. Moreover, by integrating randomness with token salience, masks are probabilistically assigned to the background and less significant tokens, enriching the diversity of the token masking." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.421, + 0.788, + 0.467 + ], + "angle": 0, + "content": "pre-training efficiency, evidenced by performance gains and faster convergence. Moreover, Fig. 2 qualitatively shows that SBM selectively masks only those tokens that contribute significantly to the image." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.494, + 0.6, + 0.51 + ], + "angle": 0, + "content": "4.1 Robustness to Masking Ratio Variability" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.524, + 0.788, + 0.689 + ], + "angle": 0, + "content": "A crucial aspect of the evaluation of SBM focuses on the robustness, particularly in the context of varying masking ratios. Prior established baselines, such as the widely used Masked Autoencoder (MAE) [14], often exhibit significant performance fluctuations with even minimal adjustments to the masking ratio. This sensitivity undermines the practicality and generalizability of such methods, especially in diverse real-world scenarios where optimal masking ratios may not be consistent across datasets. In contrast, SBM introduces a novel approach of selectively masking tokens based on their salience within an image. As a result, SBM exhibits a diminished likelihood of masking redundant or trivial tokens, thereby maintaining a stable performance across a broad spectrum of masking ratios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.692, + 0.788, + 0.81 + ], + "angle": 0, + "content": "Fig. 3 showcases the performance stability of SBM against strong baseline MAE, underscoring its superior resilience to changes in masking ratios. Our evaluation leverages the Performance Improvement over Masking Ratio (PIMR) metric, a normalized measure that quantifies how each model's performance at a given masking ratio stands against its performance at the lowest ratio, thereby reflecting the relative improvement. This metric is instrumental in revealing the impact of increased masked data on model training. The PIMR is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.368, + 0.811, + 0.786, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\operatorname {P I M R} (M) = \\frac {P (M) - P \\left(M _ {\\min }\\right)}{P \\left(M _ {\\max }\\right) - P \\left(M _ {\\min }\\right)}, \\tag {5}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.684, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SBAM" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.149, + 0.49, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.149, + 0.778, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.23, + 0.255, + 0.49, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.256, + 0.648, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.65, + 0.256, + 0.777, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.367, + 0.788, + 0.507 + ], + "angle": 0, + "content": "Fig. 3: Evaluation of robustness across varied masking ratios. To evaluate the robustness of SBM, we report the comparative analysis of image classification performance on ImageNet-1K dataset [9] against the baseline method, MAE [14], using ViT-L [12] as a backbone. The upper graphs display the performance of the methods at different masking ratios, while the lower graphs illustrate the Performance Improvement over Masking Ratio (PIMR) and Global PIMR. These measures indicate the extent of each model's performance enhancement as the masking ratio increases from the lowest to higher ratios. SBM significantly outperforms MAE in every measures, demonstrating its superior effectiveness in handling various masking ratios and enhanced pre-training performances." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.541, + 0.788, + 0.663 + ], + "angle": 0, + "content": "where \\( P(M) \\) is the performance at masking ratio \\( M \\). \\( P(M_{min}) \\) and \\( P(M_{max}) \\) denote the minimum and maximum observed performances, respectively. A PIMR value closer to 1 signifies a substantial improvement relative to the range of observed performances. The PIMR graph in classification accuracy demonstrate the robustness of the SBM strategy, where it maintains a competitive edge over MAE across various masking ratios. Notably, SBM shows a minimal performance drop at lower masking ratios compared to MAE, indicating its effectiveness even with sparse data presence." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.667, + 0.788, + 0.789 + ], + "angle": 0, + "content": "This advantage is particularly evident in the Global PIMR metric, where SBM's performance remains consistently high. For the Global PIMR calculation, we normalize performance relative to the most extensive range of performance observed among all models under comparison. This means that instead of comparing to the best and worst performances of a single model, Global PIMR considers the best and worst across both MAE and SBM, which provides a universal performance context. Thus, the metric reflects a model's improvement not in isolation but rather in relation to its peers, which can be seen in the equation:" + }, + { + "type": "equation", + "bbox": [ + 0.332, + 0.811, + 0.787, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\operatorname {G l o b a l} \\operatorname {P I M R} (M) = \\frac {P (M) - P \\left(M _ {G m i n}\\right)}{P \\left(M _ {G m a x}\\right) - P \\left(M _ {G m i n}\\right)}, \\tag {6}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "H. Choi et al." + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.152, + 0.505, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.152, + 0.776, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.295, + 0.789, + 0.381 + ], + "angle": 0, + "content": "Fig. 4: Performance evaluation of SBM with respect to the pre-trained epochs. We report the comparison of image classification accuracy on ImageNet-1K [9], pre-trained on ViT-L [12]. The left graph displays fine-tuning accuracy, whereas the right graph illustrates linear probing accuracy, both over a range of pre-trained epochs. The curves illustrate that SBM surpasses MAE [14] in pre-training effectiveness in every trained epoch, and also validates its quicker attainment of converged performance levels." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.395, + 0.789, + 0.472 + ], + "angle": 0, + "content": "where \\( P(M_{Gmax}) \\) and \\( P(M_{Gmin}) \\) are the global minimum and maximum performance values observed across both MAE and SBM models, respectively. This broader evaluation framework further underscores SBM's superior resilience and ability to maintain high accuracy across varying degrees of masking, outperforming the MAE baseline even when less information is available for learning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.472, + 0.789, + 0.547 + ], + "angle": 0, + "content": "As shown in the graphs, SBM consistently outperforms MAE across a spectrum of masking ratios, demonstrating its remarkable stability even at lower ratios. This is because in random masking, as the masking ratio decreases, the chance of including a crucial token in the mask lowers; conversely, SBM consistently masks pivotal tokens, regardless of the masking ratio." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.548, + 0.79, + 0.669 + ], + "angle": 0, + "content": "Furthermore, the classification accuracy graphs displayed above demonstrate that the application of SBM results in a notable enhancement in performance compared to MAE for all ratios. This reveals that the proposed SBM not only exhibits resilience to variations in the masking ratio but also significantly boosts pre-training efficacy, irrespective of the masking ratio. Strategically masking pivotal information enhances model performance and accelerates convergence by encouraging a comprehensive understanding of the visual context through a focus on essential tokens." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.692, + 0.536, + 0.709 + ], + "angle": 0, + "content": "4.2 Enhanced Pre-training Efficiency" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Beyond the robustness to masking ratio variations, SBM's efficacy is further demonstrated through its enhanced pre-training capabilities. Traditional MIM strategies often require extensive computational resources and time for model convergence, primarily due to the indiscriminate masking of image tokens which can hinder the learning process by obfuscating essential visual cues. By strategically selecting pivotal tokens for masking, SBM ensures that crucial tokens are leveraged during training, which leads to substantial improvement in pre-training performance and facilitates a more focused model convergence process." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.684, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SBAM" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.149, + 0.307, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.15, + 0.402, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.413, + 0.151, + 0.496, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.15, + 0.591, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.603, + 0.15, + 0.686, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.15, + 0.781, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.226, + 0.307, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.226, + 0.402, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.413, + 0.226, + 0.496, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.226, + 0.591, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.603, + 0.226, + 0.686, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.226, + 0.781, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.31, + 0.785, + 0.366 + ], + "angle": 0, + "content": "Fig. 5: Qualitative example of SBAM, which combines SBM and AMR. Having masking ratios that adapt throughout training is highly effective, as it allows the masking process to be tailored to each sample in the dataset, accommodating the unique composition and object sizes within each image, as shown in the above qualitative samples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.387, + 0.785, + 0.538 + ], + "angle": 0, + "content": "Fig. 4 showcases the comparative pre-training efficiency of SBM against MAE [14], underlining SBM's reduced pre-training duration without a tradeoff in accuracy. Initially, SBM secures a distinct lead in classification accuracy and continues to demonstrate this advantage across the training epochs, as seen in the left graph. The right graph, representing linear probing accuracy, further confirms SBM's higher initial performance, which stabilizes near peak levels well before 800 epochs. These findings highlight SBM's ability to prioritize significant features during early training stages, resulting in a new alternative for both convergence speed and pre-training performance in the pre-training of masked image models." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.565, + 0.566, + 0.583 + ], + "angle": 0, + "content": "5 Adaptive Masking Ratio (AMR)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.6, + 0.785, + 0.674 + ], + "angle": 0, + "content": "Achieving stability against changes in masking ratios has enabled us to advance the discourse on image masking, introducing a novel perspective: an Adaptive Masking Ratio (AMR). This innovative approach acknowledges the inherently diverse visual narratives presented by individual images and adjusts the masking ratio to fit different object sizes and classes within them." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.785, + 0.735 + ], + "angle": 0, + "content": "The proposed token salience \\( S = \\mathcal{N}(\\Sigma_{j=1}^{L}\\hat{A}_{\\cdot,j,:}) \\) forms the basis for determining AMRs. The AMR \\( R_{dyna} \\) is computed based on the distribution of salience scores across tokens, adjusted by a predefined variability parameter \\( \\delta \\) and the base mask ratio \\( r \\):" + }, + { + "type": "equation", + "bbox": [ + 0.363, + 0.75, + 0.785, + 0.767 + ], + "angle": 0, + "content": "\\[\nR _ {d y n a} = r - \\Delta r + 2 \\Delta r \\times \\operatorname {m e a n} \\left(1 _ {S > \\delta}\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.78, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Here, \\(\\Delta r\\) denotes the range of allowable variation in the masking ratio, as \\(R_{dyna}\\) can range from \\(r - \\Delta r\\) to \\(r + \\Delta r\\). \\(\\delta\\) is the salience threshold for distinguishing highly salient tokens, and 1 is the indicator function that identifies tokens exceeding \\(\\delta\\)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "H. Choi et al." + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.152, + 0.502, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.152, + 0.774, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.301, + 0.788, + 0.4 + ], + "angle": 0, + "content": "Fig. 6: Performance comparison of SBAM with respect to the pre-trained epochs. We report the comparison of image classification accuracy on ImageNet-1K [9], pre-trained on ViT-L [12]. The left graph illustrates classification accuracy across epochs, while the right graph shows the accuracy obtained through linear probing. Both results indicate a significant improvement in pre-training performance when AMR is applied, which not only achieves higher accuracy earlier in the training process but also maintains a lead at convergence." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.424, + 0.789, + 0.606 + ], + "angle": 0, + "content": "With \\( R_{dyna} \\) established, we adjust the number of tokens to be masked accordingly. This dynamic adjustment of masking ratios ensures that the masking process is not uniformly applied but is instead sensitive to the visual information's inherent salience, promoting a more effective learning mechanism by focusing on the informative segments of the image. Fig. 5 presents a qualitative example of AMR, demonstrating the effectiveness of adaptive masking ratios that customize the masking process for each dataset sample, thereby accounting for the unique composition and object sizes within each image. As a result, the proposed Saliency-Based Adaptive Masking (SBAM), which combines SBM and AMR, employs an adaptive approach, leading to enhanced performance across different models (See Fig. 6 and Tab. 1) and setting a new standard for tailored image masking." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.63, + 0.377, + 0.648 + ], + "angle": 0, + "content": "6 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.663, + 0.457, + 0.679 + ], + "angle": 0, + "content": "6.1 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Our evaluation approach involves deploying the proposed SBM and AMR against the baseline to assess performance enhancements of the comprehensive method, SBAM. To ensure a fair comparison, we maintain consistency with the baseline method's hyperparameters and network architectures. Notably, to preserve the integrity of our experiments, we ensured uniform hardware utilization and experimental conditions across both our method and the reproduced baseline, all employing \\(8^{*}\\mathrm{A}6000\\) GPUs. Given that this fixed GPU configuration diverges from those used in prior methods, discrepancies between our reproduced performance and the performance documented in existing papers may arise. For the fair experimental schedule, the reproduced 400-epoch performance of baseline" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.684, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SBAM" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.214 + ], + "angle": 0, + "content": "Table 1: Comprehensive performance results of applying SBAM to various baseline methods. We report the comparison of image classification fine-tuning accuracy on ImageNet-1K [9] dataset. The consistent performance improvement of SBAM across various baseline methods demonstrates the efficacy of SBAM as a scalable methodology capable of enhancing a variety of MIM frameworks." + }, + { + "type": "table", + "bbox": [ + 0.283, + 0.227, + 0.719, + 0.327 + ], + "angle": 0, + "content": "
MethodBaselineBaseline+SBAM
MAE (ViT-L) [14]84.385.1
MAE (ViT-B)82.983.6
BootMAE (ViT-B) [11]84.184.8
iBoT (ViT-B) [43]71.574.4
CMAE (ViT-B) [17]83.884.5
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.354, + 0.788, + 0.4 + ], + "angle": 0, + "content": "methods and the proposed method were all equally measured in intermediate stages in the training towards 800 epochs. A more detailed implementation description can be found in the Supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.421, + 0.441, + 0.435 + ], + "angle": 0, + "content": "6.2 Evaluation on SBAM" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.446, + 0.789, + 0.688 + ], + "angle": 0, + "content": "The graphs in Fig. 6 demonstrate the efficacy of the SBAM in comparison to the MAE model throughout the pre-training phase. Specifically, the left graph indicates that SBAM starts with a higher accuracy than MAE at the 200 epochs mark and continues to outperform MAE [14] at every subsequent checkpoint. By the 800 epochs mark, SBAM not only achieves a significant accuracy enhancement but also shows a more rapid improvement in the earlier epochs, suggesting that SBAM requires fewer epochs to achieve similar or better performance compared to MAE. In the context of linear probing, depicted in the right graph, the trend is similar. SBAM consistently achieves higher accuracy than MAE from the outset, and this performance gap is maintained as training progresses. The curves for SBAM and MAE become flat toward 800 epochs, indicating that 800 epochs or later is the point of convergence in performance. This observation underscores SBAM's substantial advantage over MAE, both in terms of converged performance and convergence speed. Overall, the performance trends captured in these graphs suggest that SBAM is more efficient during pre-training, reaching higher levels of accuracy faster than MAE." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.71, + 0.528, + 0.725 + ], + "angle": 0, + "content": "6.3 Evaluation on Various Baselines" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The integration of SBAM into various baseline methodologies demonstrates a notable enhancement in pre-training performance, as summarized in Tab. 1. All experiments report the classification accuracy on 400 epochs, except for iBoT [43] which is pre-trained for 100 epochs. Applying SBAM to the large-scale variant of MAE [14] (MAE (ViT-L)) yields a noteworthy enhancement, elevating the baseline accuracy from \\(84.3\\%\\) to \\(85.1\\%\\), marking a significant advancement. Integration of SBAM to the MAE using ViT-B also experienced significant performance" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "H. Choi et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.218, + 0.145, + 0.497, + 0.188 + ], + "angle": 0, + "content": "Table 2: Comparative evaluation of SBAM against the state-of-the-art masking strategy AMT [25]." + }, + { + "type": "table", + "bbox": [ + 0.263, + 0.194, + 0.449, + 0.265 + ], + "angle": 0, + "content": "
MethodAcc (%)
AM [25]82.5
AMT [25]82.8
SBAM83.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.502, + 0.145, + 0.78, + 0.188 + ], + "angle": 0, + "content": "Table 3: Comparative evaluation of SBAM against the state-of-the-art masking strategy AttMask [18]." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.194, + 0.761, + 0.266 + ], + "angle": 0, + "content": "
MethodAcc (%)
AttMask-High [18]72.5
AttMask-Hint [18]72.8
SBAM74.4
" + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.288, + 0.411, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.422, + 0.288, + 0.593, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.288, + 0.774, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.438, + 0.788, + 0.467 + ], + "angle": 0, + "content": "Fig. 7: Comprehensive ablation studies of the impact of randomness, variance \\((\\Delta r)\\), and delta \\((\\delta)\\) of the proposed SBAM approach." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.501, + 0.788, + 0.577 + ], + "angle": 0, + "content": "gains, underscoring that the benefits of SBAM are not limited to specific model architectures and can provide substantial pre-training efficacy for a variety of models. The performance enhancement of SBAM on various models including BootMAE [11], iBoT, and CMAE [17] highlights SBAM's generalizability and efficacy in augmenting various model structures with significant accuracy gains." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.578, + 0.788, + 0.653 + ], + "angle": 0, + "content": "In conclusion, the consistent improvement across various baselines validates the efficacy of SBAM as a scalable enhancement tool. It not only boosts performance in standard settings but also bridges the gap in more challenging learning scenarios, marking it as a pivotal development in masked image modeling pretraining techniques." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.681, + 0.559, + 0.697 + ], + "angle": 0, + "content": "6.4 Comparison with Masking Methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.795 + ], + "angle": 0, + "content": "We conducted a comparative evaluation of our SBAM masking strategy against the established masking strategies of AMT and AttMask. This comparison was performed by applying SBAM to the baselines previously employed by AMT [25] and AttMask [18] methods, specifically MAE [14] (ViT-B, 400 epochs) and iBoT [43] (ViT-B, 100 epochs), to ascertain the enhancements introduced by our approach." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In Tab. 2, the AM and AMT strategy [25], which implement basic masking and selective masking based on semantic importance, achieve an image classification fine-tuning accuracy of 82.5 and 82.8, respectively. When applied to the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.684, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SBAM" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "same baseline model, SBAM outperforms both approaches by achieving an accuracy of 83.6. This implies that the efficacy of self-supervised pre-training can be maximized by defining token salience based on the outgoing weight of the token, as opposed to the incoming weight [25]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.788, + 0.329 + ], + "angle": 0, + "content": "Furthermore, we compare SBAM to two variants of AttMask [18], applied within the iBoT [43] framework in Tab. 3. Both AttMask-High and AttMask-Hint incorporate the selective masking strategy of distillation setup which leverages similarity to classification token. SBAM stands out with an accuracy of \\(74.4\\%\\), substantially higher than AttMask-High's \\(72.5\\%\\) and AttMask-Hint's \\(72.8\\%\\). This highlights the superiority of the SBAM method, which can effectively improve the pre-training efficiency without the need for the additional computational cost of using a complex framework." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.346, + 0.4, + 0.36 + ], + "angle": 0, + "content": "6.5 Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.366, + 0.788, + 0.564 + ], + "angle": 0, + "content": "In Fig. 7, we provide comprehensive ablation studies of the impact of randomness, variance \\((\\varDelta r)\\), and delta \\((\\delta)\\) of the proposed SBAM approach. We report the ImageNet-1K [9] classification accuracy achieved by SBAM using the baseline approach of MAE [14], trained for 400 epochs on ViT-L [12] as the backbone. The first graph depicts the impact of the incorporated randomness in the SBAM on model performance over various epochs. The plot reveals that integrating randomness with token saliency markedly enhances pre-training accuracy as the number of pre-trained epochs increases. The second and third graphs show the fine-tuning accuracy and linear probing performance ablations for the hyperparameters of SBAM. While fine-tuning accuracy remained stable across various hyperparameters, linear probing accuracy demonstrated relative sensitivity. We chose \\(\\varDelta r=0.15\\) and \\(\\delta=0.1\\) as optimal hyperparameters in both figures and were universally applicable across all baseline methodologies." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.581, + 0.389, + 0.597 + ], + "angle": 0, + "content": "7 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.608, + 0.494, + 0.625 + ], + "angle": 0, + "content": "7.1 Masked Language Modeling" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Masked Language Modeling (MLM) [2,7,8,10,13,24,31,33,34,40] has become a keystone self-supervised learning paradigm in NLP, exemplified by groundbreaking models like BERT [10] and GPT [29,30]. By predicting masked tokens from their context, MLM has propelled NLP forward, enabling models to scale and perform adeptly on diverse tasks [3]. However, the considerable training time and computational demands of these models have spurred innovations aimed at increasing pre-training efficiency. For example, ALBERT [19] reduced parameters through embedding matrix factorization and shared layer parameters, while EarlyBERT [6] applied the principles of network pruning to optimize the training process. The curriculum masking approach of CCM [20] represents another advancement, strategically increasing the complexity of token masking to enhance learning. These efforts reflect a broader trend in the quest for efficiency, leading to models that maintain or exceed the capabilities of their predecessors with a fraction of the resource investment." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.127 + ], + "angle": 0, + "content": "H. Choi et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.466, + 0.163 + ], + "angle": 0, + "content": "7.2 Masked Image Modeling" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.17, + 0.788, + 0.381 + ], + "angle": 0, + "content": "In computer vision, Masked Image Modeling (MIM)) [1,4,5,11,14-16,23,26-28, 36,37,41-43] has emerged as a transformative technique, drawing parallels to the success of Masked Language Modeling (MLM) in NLP. MIM's central tenet involves predicting occluded parts of images to foster a nuanced understanding of visual content sans explicit labels. Early efforts adapting MLM concepts for visual data, such as iGPT [5], paved the way for more sophisticated methods. BEiT [1] utilized a pre-trained discrete variational autoencoder (dVAE) to produce target visual tokens. Further refinements in the technique have been observed in methods such as MAE [14] and SimMIM [36], which focus on direct prediction from unmasked image areas, refining the process of visual understanding. Efficiency in pre-training has been a critical frontier, leading to innovations like GreenMIM [16] and HiViT [42], which optimize hierarchical Vision Transformers (ViTs) by processing only the visible patches, significantly reducing computational overhead." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.381, + 0.789, + 0.684 + ], + "angle": 0, + "content": "Recent advances in model pre-training have honed in on the strategic use of masking to enhance learning efficiency, with a particular emphasis on selecting which image regions to mask. Initiating this trend, ADIOS [32] leverages adversarial training to smartly select challenging segments for masking, setting a foundation for intelligent masking approaches. AttMask [18] and SemMAE [21] further this by utilizing self-attention and semantic information, respectively, to pinpoint and mask the most informative parts of an image, thereby prioritizing high-value areas over random masking. The Attention-Driven Masking and Throwing Strategy (AMT) strategy [25] refines this focus on semantics by employing self-attention to identify and eliminate redundant patches, achieving a delicate balance between precision and efficiency. In the realm of CLIP models, ACLIP [38] and Fast CLIP (FLIP) [22] adopt attentive masking strategies to optimize training, with Fast CLIP demonstrating the effectiveness of masking substantial portions of images for accelerated learning. MaskAlign [37] introduces an innovative teacher-student framework that bypasses the need for masked region reconstruction, aligning visible features with semantically rich intact image features to concentrate on the most informative parts. Together, these approaches illustrate a shift towards more strategic, intelligent masking techniques, significantly boosting the pre-training process by leveraging both the quantity and quality of masked inputs." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.705, + 0.368, + 0.721 + ], + "angle": 0, + "content": "8 Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.84 + ], + "angle": 0, + "content": "The proposed Saliency-Based Adaptive Masking (SBAM) approach and the Adaptive Masking Ratio (AMR) significantly progress the field of MIM by introducing a method that adaptively masks image tokens with dynamic masking ratios based on their token salience. SBAM not only enhances pre-training efficiency and model performance on ImageNet-1K datasets but also introduces a novel way of considering the importance of token dynamics, thereby enabling models to learn more pivotal representations." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.684, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SBAM" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.324, + 0.161 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.784, + 0.204 + ], + "angle": 0, + "content": "1. Bao, H., Dong, L., Piao, S., Wei, F.: Beit: Bert pre-training of image transformers. In: International Conference on Learning Representations (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.205, + 0.785, + 0.259 + ], + "angle": 0, + "content": "2. Bao, H., Dong, L., Wei, F., Wang, W., Yang, N., Liu, X., Wang, Y., Gao, J., Piao, S., Zhou, M., et al.: Unilmv2: Pseudo-masked language models for unified language model pre-training. In: International conference on machine learning. pp. 642-652. PMLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.261, + 0.785, + 0.3 + ], + "angle": 0, + "content": "3. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.302, + 0.785, + 0.328 + ], + "angle": 0, + "content": "4. Cao, S., Xu, P., Clifton, D.A.: How to understand masked autoencoders. arXiv preprint arXiv:2202.03670 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.33, + 0.785, + 0.37 + ], + "angle": 0, + "content": "5. Chen, M., Radford, A., Child, R., Wu, J., Jun, H., Luan, D., Sutskever, I.: Generative pretraining from pixels. In: International conference on machine learning. pp. 1691-1703. PMLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.372, + 0.785, + 0.398 + ], + "angle": 0, + "content": "6. Chen, X., Cheng, Y., Wang, S., Gan, Z., Wang, Z., Liu, J.: Earlybert: Efficient bert training via early-bird lottery tickets. arXiv preprint arXiv:2101.00063 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.4, + 0.785, + 0.439 + ], + "angle": 0, + "content": "7. Clark, K., Luong, M.T., Le, Q.V., Manning, C.D.: Electra: Pre-training text encoders as discriminators rather than generators. arXiv preprint arXiv:2003.10555 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.441, + 0.785, + 0.467 + ], + "angle": 0, + "content": "8. Conneau, A., Lample, G.: Cross-lingual language model pretraining. Advances in neural information processing systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.469, + 0.785, + 0.509 + ], + "angle": 0, + "content": "9. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.51, + 0.785, + 0.55 + ], + "angle": 0, + "content": "0. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.551, + 0.785, + 0.592 + ], + "angle": 0, + "content": "1. Dong, X., Bao, J., Zhang, T., Chen, D., Zhang, W., Yuan, L., Chen, D., Wen, F., Yu, N.: Bootstrapped masked autoencoders for vision bert pretraining. In: European Conference on Computer Vision. pp. 247-264. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.593, + 0.785, + 0.647 + ], + "angle": 0, + "content": "2. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.648, + 0.785, + 0.688 + ], + "angle": 0, + "content": "3. Ghazvininejad, M., Levy, O., Liu, Y., Zettlemoyer, L.: Mask-predict: Parallel decoding of conditional masked language models. arXiv preprint arXiv:1904.09324 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.689, + 0.785, + 0.73 + ], + "angle": 0, + "content": "4. He, K., Chen, X., Xie, S., Li, Y., Dólar, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16000-16009 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.731, + 0.785, + 0.757 + ], + "angle": 0, + "content": "5. Hou, Z., Sun, F., Chen, Y.K., Xie, Y., Kung, S.Y.: Milan: Masked image pretraining on language assisted representation. arXiv preprint arXiv:2208.06049 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "6. Huang, L., You, S., Zheng, M., Wang, F., Qian, C., Yamasaki, T.: Green hierarchical vision transformer for masked image modeling. arXiv preprint arXiv:2205.13515 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "7. Huang, Z., Jin, X., Lu, C., Hou, Q., Cheng, M.M., Fu, D., Shen, X., Feng, J.: Contrastive masked autoencoders are stronger vision learners. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.127 + ], + "angle": 0, + "content": "H. Choi et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "18. Kakogeorgiou, I., Gidaris, S., Psomas, B., Avrithis, Y., Bursuc, A., Karantzalos, K., Komodakis, N.: What to hide from your students: Attention-guided masked image modeling. arXiv preprint arXiv:2203.12719 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.231 + ], + "angle": 0, + "content": "19. Lan, Z., Chen, M., Goodman, S., Gimpel, K., Sharma, P., Soricut, R.: Albert: A lite bert for self-supervised learning of language representations. In: International Conference on Learning Representations" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.272 + ], + "angle": 0, + "content": "20. Lee, M., Park, J.H., Kim, J., Kim, K.M., Lee, S.: Efficient pre-training of masked language model via concept-based curriculum masking. In: Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.314 + ], + "angle": 0, + "content": "21. Li, G., Zheng, H., Liu, D., Wang, C., Su, B., Zheng, C.: Semmae: Semantic-guided masking for learning masked autoencoders. Advances in Neural Information Processing Systems 35, 14290-14302 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.357 + ], + "angle": 0, + "content": "22. Li, Y., Fan, H., Hu, R., Feichtenhofer, C., He, K.: Scaling language-image pretraining via masking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23390-23400 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.785, + 0.384 + ], + "angle": 0, + "content": "23. Liu, J., Huang, X., Liu, Y., Li, H.: Mixmim: Mixed and masked image modeling for efficient visual representation learning. arXiv preprint arXiv:2205.13137 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.385, + 0.785, + 0.425 + ], + "angle": 0, + "content": "24. Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D., Levy, O., Lewis, M., Zettlemoyer, L., Stoyanov, V.: Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.426, + 0.785, + 0.467 + ], + "angle": 0, + "content": "25. Liu, Z., Gui, J., Luo, H.: Good helper is around you: Attention-driven masked image modeling. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 1799-1807 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.468, + 0.785, + 0.494 + ], + "angle": 0, + "content": "26. Pan, J., Zhou, P., Yan, S.: Towards understanding why mask-reconstruction pretraining helps in downstream tasks. arXiv preprint arXiv:2206.03826 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.785, + 0.522 + ], + "angle": 0, + "content": "27. Peng, Z., Dong, L., Bao, H., Ye, Q., Wei, F.: Beit v2: Masked image modeling with vector-quantized visual tokenizers. arXiv preprint arXiv:2208.06366 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.785, + 0.55 + ], + "angle": 0, + "content": "28. Peng, Z., Dong, L., Bao, H., Ye, Q., Wei, F.: A unified view of masked image modeling. arXiv preprint arXiv:2210.10615 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.551, + 0.785, + 0.577 + ], + "angle": 0, + "content": "29. Radford, A., Narasimhan, K., Salimans, T., Sutskever, I., et al.: Improving language understanding by generative pre-training (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.605 + ], + "angle": 0, + "content": "30. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners. OpenAI blog 1(8), 9 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.785, + 0.647 + ], + "angle": 0, + "content": "31. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research 21(1), 5485-5551 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.648, + 0.785, + 0.688 + ], + "angle": 0, + "content": "32. Shi, Y., Siddharth, N., Torr, P., Kosiorek, A.R.: Adversarial masking for self-supervised learning. In: International Conference on Machine Learning. pp. 20026-20040. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.785, + 0.73 + ], + "angle": 0, + "content": "33. Song, K., Tan, X., Qin, T., Lu, J., Liu, T.Y.: Mass: Masked sequence to sequence pre-training for language generation. In: International Conference on Machine Learning. pp. 5926-5936. PMLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.731, + 0.785, + 0.771 + ], + "angle": 0, + "content": "34. Song, K., Tan, X., Qin, T., Lu, J., Liu, T.Y.: Mpnet: Masked and permuted pretraining for language understanding. Advances in Neural Information Processing Systems 33, 16857-16867 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.799 + ], + "angle": 0, + "content": "35. Wu, J., Mo, S.: Object-wise masked autoencoders for fast pre-training. arXiv preprint arXiv:2205.14338 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "36. Xie, Z., Zhang, Z., Cao, Y., Lin, Y., Bao, J., Yao, Z., Dai, Q., Hu, H.: Simmim: A simple framework for masked image modeling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9653-9663 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.684, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SBAM" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "37. Xue, H., Gao, P., Li, H., Qiao, Y., Sun, H., Li, H., Luo, J.: Stare at what you see: Masked image modeling without reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22732-22741 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.788, + 0.231 + ], + "angle": 0, + "content": "38. Yang, Y., Huang, W., Wei, Y., Peng, H., Jiang, X., Jiang, H., Wei, F., Wang, Y., Hu, H., Qiu, L., et al.: Attentive mask clip. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2771-2781 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.231, + 0.788, + 0.259 + ], + "angle": 0, + "content": "39. Yi, K., Ge, Y., Li, X., Yang, S., Li, D., Wu, J., Shan, Y., Qie, X.: Masked image modeling with denoising contrast. arXiv preprint arXiv:2205.09616 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.259, + 0.788, + 0.3 + ], + "angle": 0, + "content": "40. Zaken, E.B., Ravfogel, S., Goldberg, Y.: Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. arXiv preprint arXiv:2106.10199 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.3, + 0.788, + 0.328 + ], + "angle": 0, + "content": "41. Zhang, Q., Wang, Y., Wang, Y.: How mask matters: Towards theoretical understandings of masked autoencoders. arXiv preprint arXiv:2210.08344 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.328, + 0.788, + 0.37 + ], + "angle": 0, + "content": "42. Zhang, X., Tian, Y., Xie, L., Huang, W., Dai, Q., Ye, Q., Tian, Q.: Hivit: A simpler and more efficient design of hierarchical vision transformer. In: The Eleventh International Conference on Learning Representations (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.37, + 0.788, + 0.398 + ], + "angle": 0, + "content": "43. Zhou, J., Wei, C., Wang, H., Shen, W., Xie, C., Yuille, A., Kong, T.: ibot: Image bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.398 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Salience-Based Adaptive Masking_ Revisiting Token Dynamics for Enhanced Pre-training/05c12db4-7790-4b9e-84f8-e76fa8191a90_origin.pdf b/2024/Salience-Based Adaptive Masking_ Revisiting Token Dynamics for Enhanced Pre-training/05c12db4-7790-4b9e-84f8-e76fa8191a90_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..128fd1d02e5b6b124a98a491b679f45acd54e466 --- /dev/null +++ b/2024/Salience-Based Adaptive Masking_ Revisiting Token Dynamics for Enhanced Pre-training/05c12db4-7790-4b9e-84f8-e76fa8191a90_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a315d4acdc410393dd748e0e0032cb611e67d21103a9a6fb6548fae3e9972c1a +size 3511637 diff --git a/2024/Salience-Based Adaptive Masking_ Revisiting Token Dynamics for Enhanced Pre-training/full.md b/2024/Salience-Based Adaptive Masking_ Revisiting Token Dynamics for Enhanced Pre-training/full.md new file mode 100644 index 0000000000000000000000000000000000000000..655a7025974daaf05753741a63df9dfabd7a95ee --- /dev/null +++ b/2024/Salience-Based Adaptive Masking_ Revisiting Token Dynamics for Enhanced Pre-training/full.md @@ -0,0 +1,326 @@ +# Salience-Based Adaptive Masking: Revisiting Token Dynamics for Enhanced Pre-training + +Hyesong Choi $^{1}$ , Hyejin Park $^{1}$ , Kwang Moo Yi $^{2}$ , Sungmin Cha $^{3}$ , and Dongbo Min $^{1\dagger}$ + +1 Ewha W. University +2 University of British Columbia + +3 New York University + +Abstract. In this paper, we introduce Saliency-Based Adaptive Masking (SBAM), a novel and cost-effective approach that significantly enhances the pre-training performance of Masked Image Modeling (MIM) approaches by prioritizing token salience. Our method provides robustness against variations in masking ratios, effectively mitigating the performance instability issues common in existing methods. This relaxes the sensitivity of MIM-based pre-training to masking ratios, which in turn allows us to propose an adaptive strategy for 'tailored' masking ratios for each data sample, which no existing method can provide. Toward this goal, we propose an Adaptive Masking Ratio (AMR) strategy that dynamically adjusts the proportion of masking for the unique content of each image based on token salience. We show that our method significantly improves over the state-of-the-art in mask-based pre-training on the ImageNet-1K dataset. Code and model parameters are available at https://github.com/doihye/SBAM. + +Keywords: Self-supervised learning $\cdot$ Masked image modeling $\cdot$ Masked autoencoder + +# 1 Introduction + +Recent drastic improvements in various Computer Vision tasks rely heavily on Transformer architectures [12]. A critical component that enables these architectures is the necessity of large-scale data [9], which is not always readily available. Naturally, pre-training with pretext tasks has become a popular solution as a workaround, represented by Masked Image Modeling (MIM) [1, 11, 14, 36], inspired by how Masked Language Modeling (MLM) [2, 7, 10, 24] has reshaped the + +![](images/edf5841832ab87f7f23efb6a9161f423637aed875e113a9115a6934592850e2f.jpg) +Fig. 1: Overview of SBAM. Whereas (a) Random Masking must rely completely on chance and carefully tuned masking ratio to guarantee effective masking, (b) the proposed SBAM strategically masks tokens based on the token salience. The robustness of SBAM paved the way for the introduction of (c) AMR, which implements a tailored masking ratio for each sample in the dataset. + +Natural Language Processing landscape. These strategies involve masking a subset of the input data and predicting those that have been hidden, thus forcing the deep network to infer underlying concepts. + +While MIM has significantly advanced self-supervised learning in vision, its conventional approach to randomly selecting tokens for masking falls short of harnessing the full potential of visual data [18]. Unlike in text, where randomness might obscure key semantic units, the visual domain's complexity and token redundancy demand a more strategic masking protocol to ensure model comprehension. This necessity prompts us to explore a refined masking methodology, targeting a selection process that achieves the goal of image understanding for pre-trained models, thereby bridging the gap in modality-specific pre-training strategies. + +Various techniques [18, 25, 35] have thus been proposed for more effective masking. While these methods all strive toward improved pre-training, an oversight shared amongst all methods is that they do not regard the contribution of each token within the overall composition of the image. It is crucial to scrutinize the interconnections between tokens to ensure that the token masking includes the tokens that play a pivotal role within the image. Moreover, prior masking methodologies [18, 25, 35] fail to address the critical consideration of the masking ratio, a factor that ought to dynamically adjust in response to the size and quantity of pivotal objects embedded within the image. This is primarily due to the difficulty reported in various works [14, 36, 39] that minor changes in the optimized masking ratio can lead to performance instability, rendering such considerations difficult to implement. Thus, these methods must rely completely on chance and carefully tuned masking ratio to guarantee effective masking. Besides, contemporary masking strategies [18, 25, 35] often grapple with high complexity, burdened by intricate distillation frameworks [18], auxiliary detection processes [35], and duplicated forward processes [25]. + +In this work, we introduce a simple yet novel approach that focuses on token dynamics, termed Salience-Based Adaptive Masking (SBAM), which aims to strategically select masking locations by discerning perceptual prominence within the visual data. Specifically, the proposed method leverages the directional emphasis from attention mechanisms to identify the image tokens pivotal to the visual context. Significantly, our methodology stands apart from the conventional attention-based approach [25] by leveraging the token's outgoing weight to calculate its 'token salience' within the image and prioritizing those that have high salience to be masked. We also infuse a degree of randomness into token salience to enrich the diversity of mask generation. Further details of SBAM are elaborated in Sec. 3. Thus, without any significant additional cost, we can consider the token's prominence within the image. + +Crucially, the strength of the SBAM approach lies in its robustness to the varying masking ratios—a notable vulnerability in established baselines [14,36, 39], which struggles from performance instability with even minor variations in optimized masking ratio. The robustness of SBAM is attributable to the proposed analytical precision in token dynamics, selectively masking tokens pivotal to the image's entirety. As a result, SBAM exhibits a diminished likelihood of masking redundant or non-essential segments, offering a stable alternative to the random masking or preceding masking approaches. Along with improving robustness to mask ratio variations, SBAM significantly enhances pre-training efficiency. A comprehensive evaluation of this is provided in Sec. 4. + +Establishing robustness against variations in masking ratios has empowered us to expand the discourse on image masking into a pioneering aspect, introducing an innovative paradigm: an Adaptive Masking Ratio (AMR). We find having masking ratios that adapt throughout training to be highly effective, as it allows the masking process to be tailored to each sample in the dataset. For instance, each image may benefit from different masking ratios, as one image might have a close-up of a bear which would enjoy high masking ratios, whereas one might have an airplane in the sky that would require a lower masking ratio; see (c) of Fig. 1. The proposed token salience, reapplied in this context, serves to finely determine the dynamic masking ratio by quantifying the proportion of tokens exhibiting high salience. More details can be found in Sec. 5. The proposed AMR implements an adaptive strategy that respects the distinctiveness of visual data and thus achieves significant performance gains when applied to various baselines (refer to Sec. 6.2 and 6.3. We note that this type of 'tailoring' to each image for pre-training is impossible with any existing method. + +In sum, this paper offers an innovative masking strategy that enhances the robustness and effectiveness of pre-trained models, setting the stage for a notable shift in the field of self-supervised learning. More importantly, the proposed method can be universally applied across any MIM framework that exploits token masking. We evaluate our SBAM and AMR on data-hungry models like ViT-L/ViT-B [12] on ImageNet-1K [9] datasets, achieving significant performance improvements in both fine-tuning and linear probing accuracy. + +To summarize, our contributions are: + +- we present Saliency-Based Adaptive Masking (SBAM), a novel effective method for MIM pre-training that focuses on token salience; +- being saliency-based, without a significant increase in computation, we allow effective masking that is robust to masking ratios: +- empowered by the robustness to masking ratio, we propose an Adaptive Masking Ratio (AMR) that allows tailored masking for each sample; +- we evaluate our method on ImageNet-1K datasets, achieving notable enhancements in both fine-tuning and linear probing accuracy. + +# 2 Preliminaries + +Within the Masked Image Modeling (MIM) domain, the crux lies in the strategic corruption and subsequent reconstruction of image segments. This process hinges on two core operations: Random Masking of image tokens and Reconstruction of corrupted tokens to guide the learning process. + +For a given image sequence $X \in \mathbb{R}^{N \times L \times D}$ , where $N, L$ and $D$ denote batch size, number of tokens per image and dimensionality of each token respectively, and specified masking proportion $\gamma$ , the general random masking process conducted by $\varPhi_{\mathrm{mask}}$ is defined as: + +$$ +X _ {\text {m a s k e d}}, M = \Phi_ {\text {m a s k}} (X, \gamma), \tag {1} +$$ + +where $X_{\mathrm{masked}}$ represents the visible tokens as a result of the post-application of the random mask $M$ , with $M \in \{0,1\}^{N \times L}$ indicating the presence (1) or absence (0) of masking for each element. + +The key component for learning in MIM is the reconstruction error $\mathcal{L}_{\mathrm{MIM}}$ , which is computed by the mean squared error (MSE) between the predicted representation of masked tokens, denoted as $\hat{X}$ , and their normalized original counterparts, represented by $X$ : + +$$ +\bar {X} = \frac {X - \mu (X)}{\sigma (X) + \epsilon}, \quad \mathcal {L} _ {\mathrm {M I M}} = \frac {1}{\sum_ {i , j} M _ {i , j}} \sum_ {i = 1} ^ {N} \sum_ {j = 1} ^ {L} M _ {i, j} \cdot \| \hat {X} _ {i, j} - \bar {X} _ {i, j} \| _ {2} ^ {2}. \tag {2} +$$ + +Here, $\mu (\cdot)$ and $\sigma (\cdot)$ represent the mean and standard deviation of $X$ respectively. $\| \cdot \| _2^2$ denotes the squared $L_{2}$ norm and $\epsilon$ is a small value for numerical stability. This formulation presents the core of MIM's training objective, focusing explicitly on the reconstruction of the masked portions of the input, encouraging the model to infer corrupted information from the unmasked context. + +# 3 Salience-Based Masking (SBM) + +Masked Image Modeling (MIM) has greatly pushed forward self-supervised learning in the visual domain, yet its standard practice of randomly masking tokens fails to fully capture the richness of visual information. To overcome this + +limitation, we present Salience-Based Masking (SBM), a novel masking methodology that selects masking locations by revisiting token dynamics. + +Given an input tensor $X \in \mathbb{R}^{N \times (L \times D)}$ , the first step involves computing an affinity matrix $\mathcal{A} \in \mathbb{R}^{N \times L \times L}$ through a batch matrix-matrix product between $X$ and $X' \in \mathbb{R}^{N \times (D \times L)}$ . Subsequently, we apply the softmax function $\rho(\cdot)$ to the affinity matrix, resulting in a normalized affinity matrix denoted as $\hat{\mathcal{A}} = \rho(\mathcal{A}) \in \mathbb{R}^{N \times L \times L}$ . The softmax function is implemented as follows: + +$$ +\rho (\mathcal {A}) _ {n, i, j} = \frac {e ^ {a _ {n , i , j}}}{\sum_ {k = 1} ^ {L} e ^ {a _ {n , i , k}}}. \tag {3} +$$ + +Here, $a_{n,i,j}$ represents an element of $\mathcal{A}$ and $e$ is the base of the natural logarithm. Note that, for each $n$ -th element with a shape of $\mathbb{R}^{L\times L}$ , this function normalizes the rows of it, transforming them into probabilities that sum to 1. + +Crucially, our approach distinguishes itself from conventional attention-based methods by utilizing the sum of outgoing weight of each token to determine the 'token salience' $S \in \mathbb{R}^{N \times L}$ in the image, which is represented by the column-wise summed score of $\hat{\mathcal{A}}$ : + +$$ +S = \mathcal {N} \left(\Sigma_ {j = 1} ^ {L} \hat {\mathcal {A}} _ {:, j,:}\right), \quad \mathcal {N} (x) = \frac {x - \operatorname* {m i n} (x)}{\operatorname* {m a x} (x) - \operatorname* {m i n} (x)}. \tag {4} +$$ + +In the affinity map, the row-wise score represents the incoming weight, reflecting the extent to which other tokens influence the corresponding token. Conversely, the column-wise score signifies the outgoing weight, indicating the contribution of the corresponding token to others. By summing these scores, it becomes feasible to quantify the token's overall impact on the image, thereby defining token salience. + +The adaptive masking process of SBM is formulated with the token salience $S$ . Notably, exclusive reliance on $S$ for masking can precipitate a decline in performance (refer to Fig. 7). To mitigate this issue, we incorporated an element of randomness into the masking process to get the adjusted token salience, denoted as $\tilde{S} = S + N$ , where $N \sim U([0,0.5)^{N \times L})$ represents a noise realization sampled from a multivariate uniform distribution $U$ . + +Then, the sampling of tokens for masking is guided by $\tilde{S}$ . We sort $\tilde{S}$ in ascending order and select tokens corresponding to the top $K$ scores, where $K = \lceil L\cdot (1 - \gamma)\rceil$ . Consequently, an adaptive binary mask $M$ is constructed, where $M_{i} = 0$ for the $K$ selected tokens, and $M_{i} = 1$ for the remainder. + +The proposed token salience offers a more intuitive and cost-efficient strategy for determining which tokens to mask. SBM therefore revisits token dynamics in the image context, streamlining the conventional complex masking process and enabling strategic masking. + +# 4 Evaluation on SBM + +In this section, the evaluation critically examines the robustness of the proposed SBM, especially against varying masking ratios, while also assessing its enhanced + +![](images/a560adb51a404aa3450d15ace64f511d6d5c9e7d40d7b8b109344ae887068002.jpg) + +![](images/b0619e75c53aa302265e480d7c4c1f6321ac32001d1103ff9f1d58a9a525775c.jpg) + +![](images/c6e3f1a641ad4460397fbb55a23860bd185c34fb36c46918fdc0d6de02f54052.jpg) + +![](images/c84a942aa9f13d05966c61c6c0ff4e0cce5083ba49ac0d2a7d7dfec799251e10.jpg) + +![](images/71dd3a3aca882f066cadbdf64dc85ab9ae4136b6987edba0af510d158a41dbcb.jpg) + +![](images/49cfdf792a3ef54b8f0cadd109c8fa7f9447fdfbf3381fe86779073116251d22.jpg) + +![](images/3d43a9a8071b5895759e961d7de6b135f52dc3a1904df00e1201aae5bd1b9cd3.jpg) +Fig. 2: Qualitative example of SBM. SBM introduces 'token salience' to prioritize and mask tokens with high significance. Hence, it is qualitatively confirmed that particularly important objects with high contribution within the image are selectively masked. Moreover, by integrating randomness with token salience, masks are probabilistically assigned to the background and less significant tokens, enriching the diversity of the token masking. + +![](images/362453e51f64e159de4494dfa60c2fa7b042f72f268d9b4c6d3f6048c0718333.jpg) + +![](images/ae77002fdd85ed3c1b27059859a73b4f875bec6904f1cc2a7f42634de2a583c8.jpg) + +![](images/4ecb6741130c59b5ddbc5fa82de165083e161120aa6e0162215924853fe6c7bc.jpg) + +![](images/06cb8ff24ce7280d283e68c9c1222080d209610faa20521abb4a602590d4060b.jpg) + +![](images/081a5cb1e6cafd855ff7268d878a25bae32832363ca878b0d611e6efdb206fa9.jpg) + +pre-training efficiency, evidenced by performance gains and faster convergence. Moreover, Fig. 2 qualitatively shows that SBM selectively masks only those tokens that contribute significantly to the image. + +# 4.1 Robustness to Masking Ratio Variability + +A crucial aspect of the evaluation of SBM focuses on the robustness, particularly in the context of varying masking ratios. Prior established baselines, such as the widely used Masked Autoencoder (MAE) [14], often exhibit significant performance fluctuations with even minimal adjustments to the masking ratio. This sensitivity undermines the practicality and generalizability of such methods, especially in diverse real-world scenarios where optimal masking ratios may not be consistent across datasets. In contrast, SBM introduces a novel approach of selectively masking tokens based on their salience within an image. As a result, SBM exhibits a diminished likelihood of masking redundant or trivial tokens, thereby maintaining a stable performance across a broad spectrum of masking ratios. + +Fig. 3 showcases the performance stability of SBM against strong baseline MAE, underscoring its superior resilience to changes in masking ratios. Our evaluation leverages the Performance Improvement over Masking Ratio (PIMR) metric, a normalized measure that quantifies how each model's performance at a given masking ratio stands against its performance at the lowest ratio, thereby reflecting the relative improvement. This metric is instrumental in revealing the impact of increased masked data on model training. The PIMR is defined as follows: + +$$ +\operatorname {P I M R} (M) = \frac {P (M) - P \left(M _ {\min }\right)}{P \left(M _ {\max }\right) - P \left(M _ {\min }\right)}, \tag {5} +$$ + +![](images/f98fad1803f38a36bea58c23afec47752f79df3785241737c185304648206fb8.jpg) + +![](images/25fb0377f585ec315dcca4ea2019690ad216dd9235a809757dc1d640a1d2bb35.jpg) + +![](images/67e4dcc61139872b2aef9a08068c192c06c70f3c82f3f442217c3ee0fe7dd2e7.jpg) +Fig. 3: Evaluation of robustness across varied masking ratios. To evaluate the robustness of SBM, we report the comparative analysis of image classification performance on ImageNet-1K dataset [9] against the baseline method, MAE [14], using ViT-L [12] as a backbone. The upper graphs display the performance of the methods at different masking ratios, while the lower graphs illustrate the Performance Improvement over Masking Ratio (PIMR) and Global PIMR. These measures indicate the extent of each model's performance enhancement as the masking ratio increases from the lowest to higher ratios. SBM significantly outperforms MAE in every measures, demonstrating its superior effectiveness in handling various masking ratios and enhanced pre-training performances. + +![](images/19e344ef5296371a9b90a01819cea30f479e0b289fe77c259cd15aade64de994.jpg) + +![](images/831248965af41e0a1622a634587aa5cf2f02916da50e653cd69bbe0036f7a21e.jpg) + +where $P(M)$ is the performance at masking ratio $M$ . $P(M_{min})$ and $P(M_{max})$ denote the minimum and maximum observed performances, respectively. A PIMR value closer to 1 signifies a substantial improvement relative to the range of observed performances. The PIMR graph in classification accuracy demonstrate the robustness of the SBM strategy, where it maintains a competitive edge over MAE across various masking ratios. Notably, SBM shows a minimal performance drop at lower masking ratios compared to MAE, indicating its effectiveness even with sparse data presence. + +This advantage is particularly evident in the Global PIMR metric, where SBM's performance remains consistently high. For the Global PIMR calculation, we normalize performance relative to the most extensive range of performance observed among all models under comparison. This means that instead of comparing to the best and worst performances of a single model, Global PIMR considers the best and worst across both MAE and SBM, which provides a universal performance context. Thus, the metric reflects a model's improvement not in isolation but rather in relation to its peers, which can be seen in the equation: + +$$ +\operatorname {G l o b a l} \operatorname {P I M R} (M) = \frac {P (M) - P \left(M _ {G m i n}\right)}{P \left(M _ {G m a x}\right) - P \left(M _ {G m i n}\right)}, \tag {6} +$$ + +![](images/7ea3238ca914f38640f8ff25c457e094a269a97c513740e8ece77ba01da6b5e4.jpg) +Fig. 4: Performance evaluation of SBM with respect to the pre-trained epochs. We report the comparison of image classification accuracy on ImageNet-1K [9], pre-trained on ViT-L [12]. The left graph displays fine-tuning accuracy, whereas the right graph illustrates linear probing accuracy, both over a range of pre-trained epochs. The curves illustrate that SBM surpasses MAE [14] in pre-training effectiveness in every trained epoch, and also validates its quicker attainment of converged performance levels. + +![](images/3dc5994f371a561c6673aa90fdd0878807ea032cc758ee8facdd83523f2aa9c1.jpg) + +where $P(M_{Gmax})$ and $P(M_{Gmin})$ are the global minimum and maximum performance values observed across both MAE and SBM models, respectively. This broader evaluation framework further underscores SBM's superior resilience and ability to maintain high accuracy across varying degrees of masking, outperforming the MAE baseline even when less information is available for learning. + +As shown in the graphs, SBM consistently outperforms MAE across a spectrum of masking ratios, demonstrating its remarkable stability even at lower ratios. This is because in random masking, as the masking ratio decreases, the chance of including a crucial token in the mask lowers; conversely, SBM consistently masks pivotal tokens, regardless of the masking ratio. + +Furthermore, the classification accuracy graphs displayed above demonstrate that the application of SBM results in a notable enhancement in performance compared to MAE for all ratios. This reveals that the proposed SBM not only exhibits resilience to variations in the masking ratio but also significantly boosts pre-training efficacy, irrespective of the masking ratio. Strategically masking pivotal information enhances model performance and accelerates convergence by encouraging a comprehensive understanding of the visual context through a focus on essential tokens. + +# 4.2 Enhanced Pre-training Efficiency + +Beyond the robustness to masking ratio variations, SBM's efficacy is further demonstrated through its enhanced pre-training capabilities. Traditional MIM strategies often require extensive computational resources and time for model convergence, primarily due to the indiscriminate masking of image tokens which can hinder the learning process by obfuscating essential visual cues. By strategically selecting pivotal tokens for masking, SBM ensures that crucial tokens are leveraged during training, which leads to substantial improvement in pre-training performance and facilitates a more focused model convergence process. + +![](images/a645b3183736e68dc8d4bc9b1cc6cf11e0be2cf216bf55e9579bb387b779c257.jpg) + +![](images/7c37280ad50cea46575d4947b8ee2c3712b2d94928ade23614ea7a83180cab2b.jpg) + +![](images/49bb5f48b63b7a079f447c01d62fef6325888fb1850682a4903a06ea0b39fb4b.jpg) + +![](images/92e652d4a17f5c3e986a2eea8a3989fc981d3db1697fb9cb53bcba6e5130263a.jpg) + +![](images/5f0574eb78a0656b4e4308a39de0c33a71a2b2f25b64f9e0cdfc8961a22e3c77.jpg) + +![](images/0dcc4e5b4de537eed820e1c53cecf91fbf63517b82698666a7c32a14e8255170.jpg) + +![](images/5bc341637fed440cec1fcee5784766b9819ef29c6791c74ee23463ef3edbb9ec.jpg) +Fig. 5: Qualitative example of SBAM, which combines SBM and AMR. Having masking ratios that adapt throughout training is highly effective, as it allows the masking process to be tailored to each sample in the dataset, accommodating the unique composition and object sizes within each image, as shown in the above qualitative samples. + +![](images/6da04f99d96b8e753059f6eee1ef6c22716e49d56010e2179b35cbcff0eeaa61.jpg) + +![](images/1fc530b9fd79b0b7351086e4d6003783bb06c6d5dbf0cbb9999ea83174e7d63d.jpg) + +![](images/fb4ed28cceb34f79ccd7d4af6d4275ad09907532529d016de253a1d6c9f54390.jpg) + +![](images/75419a52424e7f759af059239dacc528dc2ab87d65be30223c614da88f7b4d5b.jpg) + +![](images/d6f639e2d7bc7cba1176cfde6a7341e6ded0b2c774762e8b843fad45339967ee.jpg) + +Fig. 4 showcases the comparative pre-training efficiency of SBM against MAE [14], underlining SBM's reduced pre-training duration without a tradeoff in accuracy. Initially, SBM secures a distinct lead in classification accuracy and continues to demonstrate this advantage across the training epochs, as seen in the left graph. The right graph, representing linear probing accuracy, further confirms SBM's higher initial performance, which stabilizes near peak levels well before 800 epochs. These findings highlight SBM's ability to prioritize significant features during early training stages, resulting in a new alternative for both convergence speed and pre-training performance in the pre-training of masked image models. + +# 5 Adaptive Masking Ratio (AMR) + +Achieving stability against changes in masking ratios has enabled us to advance the discourse on image masking, introducing a novel perspective: an Adaptive Masking Ratio (AMR). This innovative approach acknowledges the inherently diverse visual narratives presented by individual images and adjusts the masking ratio to fit different object sizes and classes within them. + +The proposed token salience $S = \mathcal{N}(\Sigma_{j=1}^{L}\hat{A}_{\cdot,j,:})$ forms the basis for determining AMRs. The AMR $R_{dyna}$ is computed based on the distribution of salience scores across tokens, adjusted by a predefined variability parameter $\delta$ and the base mask ratio $r$ : + +$$ +R _ {d y n a} = r - \Delta r + 2 \Delta r \times \operatorname {m e a n} \left(1 _ {S > \delta}\right). \tag {7} +$$ + +Here, $\Delta r$ denotes the range of allowable variation in the masking ratio, as $R_{dyna}$ can range from $r - \Delta r$ to $r + \Delta r$ . $\delta$ is the salience threshold for distinguishing highly salient tokens, and 1 is the indicator function that identifies tokens exceeding $\delta$ . + +![](images/0dee5c96a07a79db0317ff25da0709d9bcaf3bdc80cd115a0765494620ead434.jpg) +Fig. 6: Performance comparison of SBAM with respect to the pre-trained epochs. We report the comparison of image classification accuracy on ImageNet-1K [9], pre-trained on ViT-L [12]. The left graph illustrates classification accuracy across epochs, while the right graph shows the accuracy obtained through linear probing. Both results indicate a significant improvement in pre-training performance when AMR is applied, which not only achieves higher accuracy earlier in the training process but also maintains a lead at convergence. + +![](images/32c666333013583215b2fa259ba488783ad77485e71ba6d78e316589640bdba9.jpg) + +With $R_{dyna}$ established, we adjust the number of tokens to be masked accordingly. This dynamic adjustment of masking ratios ensures that the masking process is not uniformly applied but is instead sensitive to the visual information's inherent salience, promoting a more effective learning mechanism by focusing on the informative segments of the image. Fig. 5 presents a qualitative example of AMR, demonstrating the effectiveness of adaptive masking ratios that customize the masking process for each dataset sample, thereby accounting for the unique composition and object sizes within each image. As a result, the proposed Saliency-Based Adaptive Masking (SBAM), which combines SBM and AMR, employs an adaptive approach, leading to enhanced performance across different models (See Fig. 6 and Tab. 1) and setting a new standard for tailored image masking. + +# 6 Experiments + +# 6.1 Implementation Details + +Our evaluation approach involves deploying the proposed SBM and AMR against the baseline to assess performance enhancements of the comprehensive method, SBAM. To ensure a fair comparison, we maintain consistency with the baseline method's hyperparameters and network architectures. Notably, to preserve the integrity of our experiments, we ensured uniform hardware utilization and experimental conditions across both our method and the reproduced baseline, all employing $8^{*}\mathrm{A}6000$ GPUs. Given that this fixed GPU configuration diverges from those used in prior methods, discrepancies between our reproduced performance and the performance documented in existing papers may arise. For the fair experimental schedule, the reproduced 400-epoch performance of baseline + +Table 1: Comprehensive performance results of applying SBAM to various baseline methods. We report the comparison of image classification fine-tuning accuracy on ImageNet-1K [9] dataset. The consistent performance improvement of SBAM across various baseline methods demonstrates the efficacy of SBAM as a scalable methodology capable of enhancing a variety of MIM frameworks. + +
MethodBaselineBaseline+SBAM
MAE (ViT-L) [14]84.385.1
MAE (ViT-B)82.983.6
BootMAE (ViT-B) [11]84.184.8
iBoT (ViT-B) [43]71.574.4
CMAE (ViT-B) [17]83.884.5
+ +methods and the proposed method were all equally measured in intermediate stages in the training towards 800 epochs. A more detailed implementation description can be found in the Supplementary material. + +# 6.2 Evaluation on SBAM + +The graphs in Fig. 6 demonstrate the efficacy of the SBAM in comparison to the MAE model throughout the pre-training phase. Specifically, the left graph indicates that SBAM starts with a higher accuracy than MAE at the 200 epochs mark and continues to outperform MAE [14] at every subsequent checkpoint. By the 800 epochs mark, SBAM not only achieves a significant accuracy enhancement but also shows a more rapid improvement in the earlier epochs, suggesting that SBAM requires fewer epochs to achieve similar or better performance compared to MAE. In the context of linear probing, depicted in the right graph, the trend is similar. SBAM consistently achieves higher accuracy than MAE from the outset, and this performance gap is maintained as training progresses. The curves for SBAM and MAE become flat toward 800 epochs, indicating that 800 epochs or later is the point of convergence in performance. This observation underscores SBAM's substantial advantage over MAE, both in terms of converged performance and convergence speed. Overall, the performance trends captured in these graphs suggest that SBAM is more efficient during pre-training, reaching higher levels of accuracy faster than MAE. + +# 6.3 Evaluation on Various Baselines + +The integration of SBAM into various baseline methodologies demonstrates a notable enhancement in pre-training performance, as summarized in Tab. 1. All experiments report the classification accuracy on 400 epochs, except for iBoT [43] which is pre-trained for 100 epochs. Applying SBAM to the large-scale variant of MAE [14] (MAE (ViT-L)) yields a noteworthy enhancement, elevating the baseline accuracy from $84.3\%$ to $85.1\%$ , marking a significant advancement. Integration of SBAM to the MAE using ViT-B also experienced significant performance + +Table 2: Comparative evaluation of SBAM against the state-of-the-art masking strategy AMT [25]. + +
MethodAcc (%)
AM [25]82.5
AMT [25]82.8
SBAM83.6
+ +Table 3: Comparative evaluation of SBAM against the state-of-the-art masking strategy AttMask [18]. + +
MethodAcc (%)
AttMask-High [18]72.5
AttMask-Hint [18]72.8
SBAM74.4
+ +![](images/cbfd74e9497fcf453434baef5780b39e0832b2bc081f2442a9dafbf31bcc711c.jpg) +Fig. 7: Comprehensive ablation studies of the impact of randomness, variance $(\Delta r)$ , and delta $(\delta)$ of the proposed SBAM approach. + +![](images/55beaab6d230ca3654e175ec57dee0e96c9a2968fa60094c540bb4edbe4e98aa.jpg) + +![](images/8a78552668ab7a6eaa6000b19c11cdeb2f422e232d98131ea7e7e5c4cd218373.jpg) + +gains, underscoring that the benefits of SBAM are not limited to specific model architectures and can provide substantial pre-training efficacy for a variety of models. The performance enhancement of SBAM on various models including BootMAE [11], iBoT, and CMAE [17] highlights SBAM's generalizability and efficacy in augmenting various model structures with significant accuracy gains. + +In conclusion, the consistent improvement across various baselines validates the efficacy of SBAM as a scalable enhancement tool. It not only boosts performance in standard settings but also bridges the gap in more challenging learning scenarios, marking it as a pivotal development in masked image modeling pretraining techniques. + +# 6.4 Comparison with Masking Methods + +We conducted a comparative evaluation of our SBAM masking strategy against the established masking strategies of AMT and AttMask. This comparison was performed by applying SBAM to the baselines previously employed by AMT [25] and AttMask [18] methods, specifically MAE [14] (ViT-B, 400 epochs) and iBoT [43] (ViT-B, 100 epochs), to ascertain the enhancements introduced by our approach. + +In Tab. 2, the AM and AMT strategy [25], which implement basic masking and selective masking based on semantic importance, achieve an image classification fine-tuning accuracy of 82.5 and 82.8, respectively. When applied to the + +same baseline model, SBAM outperforms both approaches by achieving an accuracy of 83.6. This implies that the efficacy of self-supervised pre-training can be maximized by defining token salience based on the outgoing weight of the token, as opposed to the incoming weight [25]. + +Furthermore, we compare SBAM to two variants of AttMask [18], applied within the iBoT [43] framework in Tab. 3. Both AttMask-High and AttMask-Hint incorporate the selective masking strategy of distillation setup which leverages similarity to classification token. SBAM stands out with an accuracy of $74.4\%$ , substantially higher than AttMask-High's $72.5\%$ and AttMask-Hint's $72.8\%$ . This highlights the superiority of the SBAM method, which can effectively improve the pre-training efficiency without the need for the additional computational cost of using a complex framework. + +# 6.5 Ablation studies + +In Fig. 7, we provide comprehensive ablation studies of the impact of randomness, variance $(\varDelta r)$ , and delta $(\delta)$ of the proposed SBAM approach. We report the ImageNet-1K [9] classification accuracy achieved by SBAM using the baseline approach of MAE [14], trained for 400 epochs on ViT-L [12] as the backbone. The first graph depicts the impact of the incorporated randomness in the SBAM on model performance over various epochs. The plot reveals that integrating randomness with token saliency markedly enhances pre-training accuracy as the number of pre-trained epochs increases. The second and third graphs show the fine-tuning accuracy and linear probing performance ablations for the hyperparameters of SBAM. While fine-tuning accuracy remained stable across various hyperparameters, linear probing accuracy demonstrated relative sensitivity. We chose $\varDelta r=0.15$ and $\delta=0.1$ as optimal hyperparameters in both figures and were universally applicable across all baseline methodologies. + +# 7 Related Work + +# 7.1 Masked Language Modeling + +Masked Language Modeling (MLM) [2,7,8,10,13,24,31,33,34,40] has become a keystone self-supervised learning paradigm in NLP, exemplified by groundbreaking models like BERT [10] and GPT [29,30]. By predicting masked tokens from their context, MLM has propelled NLP forward, enabling models to scale and perform adeptly on diverse tasks [3]. However, the considerable training time and computational demands of these models have spurred innovations aimed at increasing pre-training efficiency. For example, ALBERT [19] reduced parameters through embedding matrix factorization and shared layer parameters, while EarlyBERT [6] applied the principles of network pruning to optimize the training process. The curriculum masking approach of CCM [20] represents another advancement, strategically increasing the complexity of token masking to enhance learning. These efforts reflect a broader trend in the quest for efficiency, leading to models that maintain or exceed the capabilities of their predecessors with a fraction of the resource investment. + +# 7.2 Masked Image Modeling + +In computer vision, Masked Image Modeling (MIM)) [1,4,5,11,14-16,23,26-28, 36,37,41-43] has emerged as a transformative technique, drawing parallels to the success of Masked Language Modeling (MLM) in NLP. MIM's central tenet involves predicting occluded parts of images to foster a nuanced understanding of visual content sans explicit labels. Early efforts adapting MLM concepts for visual data, such as iGPT [5], paved the way for more sophisticated methods. BEiT [1] utilized a pre-trained discrete variational autoencoder (dVAE) to produce target visual tokens. Further refinements in the technique have been observed in methods such as MAE [14] and SimMIM [36], which focus on direct prediction from unmasked image areas, refining the process of visual understanding. Efficiency in pre-training has been a critical frontier, leading to innovations like GreenMIM [16] and HiViT [42], which optimize hierarchical Vision Transformers (ViTs) by processing only the visible patches, significantly reducing computational overhead. + +Recent advances in model pre-training have honed in on the strategic use of masking to enhance learning efficiency, with a particular emphasis on selecting which image regions to mask. Initiating this trend, ADIOS [32] leverages adversarial training to smartly select challenging segments for masking, setting a foundation for intelligent masking approaches. AttMask [18] and SemMAE [21] further this by utilizing self-attention and semantic information, respectively, to pinpoint and mask the most informative parts of an image, thereby prioritizing high-value areas over random masking. The Attention-Driven Masking and Throwing Strategy (AMT) strategy [25] refines this focus on semantics by employing self-attention to identify and eliminate redundant patches, achieving a delicate balance between precision and efficiency. In the realm of CLIP models, ACLIP [38] and Fast CLIP (FLIP) [22] adopt attentive masking strategies to optimize training, with Fast CLIP demonstrating the effectiveness of masking substantial portions of images for accelerated learning. MaskAlign [37] introduces an innovative teacher-student framework that bypasses the need for masked region reconstruction, aligning visible features with semantically rich intact image features to concentrate on the most informative parts. Together, these approaches illustrate a shift towards more strategic, intelligent masking techniques, significantly boosting the pre-training process by leveraging both the quantity and quality of masked inputs. + +# 8 Conclusions + +The proposed Saliency-Based Adaptive Masking (SBAM) approach and the Adaptive Masking Ratio (AMR) significantly progress the field of MIM by introducing a method that adaptively masks image tokens with dynamic masking ratios based on their token salience. SBAM not only enhances pre-training efficiency and model performance on ImageNet-1K datasets but also introduces a novel way of considering the importance of token dynamics, thereby enabling models to learn more pivotal representations. + +# References + +1. Bao, H., Dong, L., Piao, S., Wei, F.: Beit: Bert pre-training of image transformers. In: International Conference on Learning Representations (2021) +2. Bao, H., Dong, L., Wei, F., Wang, W., Yang, N., Liu, X., Wang, Y., Gao, J., Piao, S., Zhou, M., et al.: Unilmv2: Pseudo-masked language models for unified language model pre-training. In: International conference on machine learning. pp. 642-652. PMLR (2020) +3. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020) +4. Cao, S., Xu, P., Clifton, D.A.: How to understand masked autoencoders. arXiv preprint arXiv:2202.03670 (2022) +5. Chen, M., Radford, A., Child, R., Wu, J., Jun, H., Luan, D., Sutskever, I.: Generative pretraining from pixels. In: International conference on machine learning. pp. 1691-1703. PMLR (2020) +6. Chen, X., Cheng, Y., Wang, S., Gan, Z., Wang, Z., Liu, J.: Earlybert: Efficient bert training via early-bird lottery tickets. arXiv preprint arXiv:2101.00063 (2020) +7. Clark, K., Luong, M.T., Le, Q.V., Manning, C.D.: Electra: Pre-training text encoders as discriminators rather than generators. arXiv preprint arXiv:2003.10555 (2020) +8. Conneau, A., Lample, G.: Cross-lingual language model pretraining. Advances in neural information processing systems 32 (2019) +9. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009) +0. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018) +1. Dong, X., Bao, J., Zhang, T., Chen, D., Zhang, W., Yuan, L., Chen, D., Wen, F., Yu, N.: Bootstrapped masked autoencoders for vision bert pretraining. In: European Conference on Computer Vision. pp. 247-264. Springer (2022) +2. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) +3. Ghazvininejad, M., Levy, O., Liu, Y., Zettlemoyer, L.: Mask-predict: Parallel decoding of conditional masked language models. arXiv preprint arXiv:1904.09324 (2019) +4. He, K., Chen, X., Xie, S., Li, Y., Dólar, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16000-16009 (2022) +5. Hou, Z., Sun, F., Chen, Y.K., Xie, Y., Kung, S.Y.: Milan: Masked image pretraining on language assisted representation. arXiv preprint arXiv:2208.06049 (2022) +6. Huang, L., You, S., Zheng, M., Wang, F., Qian, C., Yamasaki, T.: Green hierarchical vision transformer for masked image modeling. arXiv preprint arXiv:2205.13515 (2022) +7. Huang, Z., Jin, X., Lu, C., Hou, Q., Cheng, M.M., Fu, D., Shen, X., Feng, J.: Contrastive masked autoencoders are stronger vision learners. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023) + +18. Kakogeorgiou, I., Gidaris, S., Psomas, B., Avrithis, Y., Bursuc, A., Karantzalos, K., Komodakis, N.: What to hide from your students: Attention-guided masked image modeling. arXiv preprint arXiv:2203.12719 (2022) +19. Lan, Z., Chen, M., Goodman, S., Gimpel, K., Sharma, P., Soricut, R.: Albert: A lite bert for self-supervised learning of language representations. In: International Conference on Learning Representations +20. Lee, M., Park, J.H., Kim, J., Kim, K.M., Lee, S.: Efficient pre-training of masked language model via concept-based curriculum masking. In: Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing (2022) +21. Li, G., Zheng, H., Liu, D., Wang, C., Su, B., Zheng, C.: Semmae: Semantic-guided masking for learning masked autoencoders. Advances in Neural Information Processing Systems 35, 14290-14302 (2022) +22. Li, Y., Fan, H., Hu, R., Feichtenhofer, C., He, K.: Scaling language-image pretraining via masking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23390-23400 (2023) +23. Liu, J., Huang, X., Liu, Y., Li, H.: Mixmim: Mixed and masked image modeling for efficient visual representation learning. arXiv preprint arXiv:2205.13137 (2022) +24. Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D., Levy, O., Lewis, M., Zettlemoyer, L., Stoyanov, V.: Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692 (2019) +25. Liu, Z., Gui, J., Luo, H.: Good helper is around you: Attention-driven masked image modeling. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 1799-1807 (2023) +26. Pan, J., Zhou, P., Yan, S.: Towards understanding why mask-reconstruction pretraining helps in downstream tasks. arXiv preprint arXiv:2206.03826 (2022) +27. Peng, Z., Dong, L., Bao, H., Ye, Q., Wei, F.: Beit v2: Masked image modeling with vector-quantized visual tokenizers. arXiv preprint arXiv:2208.06366 (2022) +28. Peng, Z., Dong, L., Bao, H., Ye, Q., Wei, F.: A unified view of masked image modeling. arXiv preprint arXiv:2210.10615 (2022) +29. Radford, A., Narasimhan, K., Salimans, T., Sutskever, I., et al.: Improving language understanding by generative pre-training (2018) +30. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners. OpenAI blog 1(8), 9 (2019) +31. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research 21(1), 5485-5551 (2020) +32. Shi, Y., Siddharth, N., Torr, P., Kosiorek, A.R.: Adversarial masking for self-supervised learning. In: International Conference on Machine Learning. pp. 20026-20040. PMLR (2022) +33. Song, K., Tan, X., Qin, T., Lu, J., Liu, T.Y.: Mass: Masked sequence to sequence pre-training for language generation. In: International Conference on Machine Learning. pp. 5926-5936. PMLR (2019) +34. Song, K., Tan, X., Qin, T., Lu, J., Liu, T.Y.: Mpnet: Masked and permuted pretraining for language understanding. Advances in Neural Information Processing Systems 33, 16857-16867 (2020) +35. Wu, J., Mo, S.: Object-wise masked autoencoders for fast pre-training. arXiv preprint arXiv:2205.14338 (2022) +36. Xie, Z., Zhang, Z., Cao, Y., Lin, Y., Bao, J., Yao, Z., Dai, Q., Hu, H.: Simmim: A simple framework for masked image modeling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9653-9663 (2022) + +37. Xue, H., Gao, P., Li, H., Qiao, Y., Sun, H., Li, H., Luo, J.: Stare at what you see: Masked image modeling without reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22732-22741 (2023) +38. Yang, Y., Huang, W., Wei, Y., Peng, H., Jiang, X., Jiang, H., Wei, F., Wang, Y., Hu, H., Qiu, L., et al.: Attentive mask clip. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2771-2781 (2023) +39. Yi, K., Ge, Y., Li, X., Yang, S., Li, D., Wu, J., Shan, Y., Qie, X.: Masked image modeling with denoising contrast. arXiv preprint arXiv:2205.09616 (2022) +40. Zaken, E.B., Ravfogel, S., Goldberg, Y.: Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. arXiv preprint arXiv:2106.10199 (2021) +41. Zhang, Q., Wang, Y., Wang, Y.: How mask matters: Towards theoretical understandings of masked autoencoders. arXiv preprint arXiv:2210.08344 (2022) +42. Zhang, X., Tian, Y., Xie, L., Huang, W., Dai, Q., Ye, Q., Tian, Q.: Hivit: A simpler and more efficient design of hierarchical vision transformer. In: The Eleventh International Conference on Learning Representations (2023) +43. Zhou, J., Wei, C., Wang, H., Shen, W., Xie, C., Yuille, A., Kong, T.: ibot: Image bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832 (2021) \ No newline at end of file diff --git a/2024/Salience-Based Adaptive Masking_ Revisiting Token Dynamics for Enhanced Pre-training/images.zip b/2024/Salience-Based Adaptive Masking_ Revisiting Token Dynamics for Enhanced Pre-training/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d5d35af02579dbe5609336b90d3d9ab4fee70d5d --- /dev/null +++ b/2024/Salience-Based Adaptive Masking_ Revisiting Token Dynamics for Enhanced Pre-training/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:67aa24d9fe6d588f3c9442bc1e13f10a21ff719b65f5d4a2c3a772246643b924 +size 423110 diff --git a/2024/Salience-Based Adaptive Masking_ Revisiting Token Dynamics for Enhanced Pre-training/layout.json b/2024/Salience-Based Adaptive Masking_ Revisiting Token Dynamics for Enhanced Pre-training/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..507be9cc0909f1114db6e47ddc26b158eb23c1f7 --- /dev/null +++ b/2024/Salience-Based Adaptive Masking_ Revisiting Token Dynamics for Enhanced Pre-training/layout.json @@ -0,0 +1,9371 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 149, + 111, + 465, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 111, + 465, + 148 + ], + "spans": [ + { + "bbox": [ + 149, + 111, + 465, + 148 + ], + "type": "text", + "content": "Salience-Based Adaptive Masking: Revisiting Token Dynamics for Enhanced Pre-training" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 201, + 168, + 412, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 168, + 412, + 195 + ], + "spans": [ + { + "bbox": [ + 201, + 168, + 412, + 195 + ], + "type": "text", + "content": "Hyesong Choi" + }, + { + "bbox": [ + 201, + 168, + 412, + 195 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 201, + 168, + 412, + 195 + ], + "type": "text", + "content": ", Hyejin Park" + }, + { + "bbox": [ + 201, + 168, + 412, + 195 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 201, + 168, + 412, + 195 + ], + "type": "text", + "content": ", Kwang Moo Yi" + }, + { + "bbox": [ + 201, + 168, + 412, + 195 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 201, + 168, + 412, + 195 + ], + "type": "text", + "content": ", Sungmin Cha" + }, + { + "bbox": [ + 201, + 168, + 412, + 195 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 201, + 168, + 412, + 195 + ], + "type": "text", + "content": ", and Dongbo Min" + }, + { + "bbox": [ + 201, + 168, + 412, + 195 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 239, + 201, + 375, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 201, + 375, + 224 + ], + "spans": [ + { + "bbox": [ + 239, + 201, + 375, + 224 + ], + "type": "text", + "content": "1 Ewha W. University \n2 University of British Columbia" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 259, + 224, + 354, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 259, + 224, + 354, + 236 + ], + "spans": [ + { + "bbox": [ + 259, + 224, + 354, + 236 + ], + "type": "text", + "content": "3 New York University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 269, + 455, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 269, + 455, + 434 + ], + "spans": [ + { + "bbox": [ + 160, + 269, + 455, + 434 + ], + "type": "text", + "content": "Abstract. In this paper, we introduce Saliency-Based Adaptive Masking (SBAM), a novel and cost-effective approach that significantly enhances the pre-training performance of Masked Image Modeling (MIM) approaches by prioritizing token salience. Our method provides robustness against variations in masking ratios, effectively mitigating the performance instability issues common in existing methods. This relaxes the sensitivity of MIM-based pre-training to masking ratios, which in turn allows us to propose an adaptive strategy for 'tailored' masking ratios for each data sample, which no existing method can provide. Toward this goal, we propose an Adaptive Masking Ratio (AMR) strategy that dynamically adjusts the proportion of masking for the unique content of each image based on token salience. We show that our method significantly improves over the state-of-the-art in mask-based pre-training on the ImageNet-1K dataset. Code and model parameters are available at https://github.com/doihye/SBAM." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 445, + 455, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 445, + 455, + 468 + ], + "spans": [ + { + "bbox": [ + 159, + 445, + 455, + 468 + ], + "type": "text", + "content": "Keywords: Self-supervised learning " + }, + { + "bbox": [ + 159, + 445, + 455, + 468 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 445, + 455, + 468 + ], + "type": "text", + "content": " Masked image modeling " + }, + { + "bbox": [ + 159, + 445, + 455, + 468 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 445, + 455, + 468 + ], + "type": "text", + "content": " Masked autoencoder" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 490, + 230, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 230, + 502 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 230, + 502 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 517, + 482, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 517, + 482, + 591 + ], + "spans": [ + { + "bbox": [ + 130, + 517, + 482, + 591 + ], + "type": "text", + "content": "Recent drastic improvements in various Computer Vision tasks rely heavily on Transformer architectures [12]. A critical component that enables these architectures is the necessity of large-scale data [9], which is not always readily available. Naturally, pre-training with pretext tasks has become a popular solution as a workaround, represented by Masked Image Modeling (MIM) [1, 11, 14, 36], inspired by how Masked Language Modeling (MLM) [2, 7, 10, 24] has reshaped the" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 598, + 483, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 598, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 141, + 598, + 483, + 666 + ], + "type": "text", + "content": "This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (No. RS-2023-00222385) and the National Research Foundation of Korea, Republic of Korea (2022M3H9A2083956) and from the Korea Basic Science Institute, Republic of Korea (National Research Facilities and Equipment Center) funded by the Ministry of Education, Republic of Korea (2019R1A6C1010020). † Corresponding author: dbmin@ewha.ac.kr" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 155, + 117, + 460, + 254 + ], + "blocks": [ + { + "bbox": [ + 155, + 117, + 460, + 254 + ], + "lines": [ + { + "bbox": [ + 155, + 117, + 460, + 254 + ], + "spans": [ + { + "bbox": [ + 155, + 117, + 460, + 254 + ], + "type": "image", + "image_path": "edf5841832ab87f7f23efb6a9161f423637aed875e113a9115a6934592850e2f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 264, + 482, + 319 + ], + "lines": [ + { + "bbox": [ + 130, + 264, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 264, + 482, + 319 + ], + "type": "text", + "content": "Fig. 1: Overview of SBAM. Whereas (a) Random Masking must rely completely on chance and carefully tuned masking ratio to guarantee effective masking, (b) the proposed SBAM strategically masks tokens based on the token salience. The robustness of SBAM paved the way for the introduction of (c) AMR, which implements a tailored masking ratio for each sample in the dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 328, + 480, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 328, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 328, + 480, + 365 + ], + "type": "text", + "content": "Natural Language Processing landscape. These strategies involve masking a subset of the input data and predicting those that have been hidden, thus forcing the deep network to infer underlying concepts." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 365, + 482, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 472 + ], + "type": "text", + "content": "While MIM has significantly advanced self-supervised learning in vision, its conventional approach to randomly selecting tokens for masking falls short of harnessing the full potential of visual data [18]. Unlike in text, where randomness might obscure key semantic units, the visual domain's complexity and token redundancy demand a more strategic masking protocol to ensure model comprehension. This necessity prompts us to explore a refined masking methodology, targeting a selection process that achieves the goal of image understanding for pre-trained models, thereby bridging the gap in modality-specific pre-training strategies." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "text", + "content": "Various techniques [18, 25, 35] have thus been proposed for more effective masking. While these methods all strive toward improved pre-training, an oversight shared amongst all methods is that they do not regard the contribution of each token within the overall composition of the image. It is crucial to scrutinize the interconnections between tokens to ensure that the token masking includes the tokens that play a pivotal role within the image. Moreover, prior masking methodologies [18, 25, 35] fail to address the critical consideration of the masking ratio, a factor that ought to dynamically adjust in response to the size and quantity of pivotal objects embedded within the image. This is primarily due to the difficulty reported in various works [14, 36, 39] that minor changes in the optimized masking ratio can lead to performance instability, rendering such considerations difficult to implement. Thus, these methods must rely completely on chance and carefully tuned masking ratio to guarantee effective masking. Besides, contemporary masking strategies [18, 25, 35] often grapple with high complexity, burdened by intricate distillation frameworks [18], auxiliary detection processes [35], and duplicated forward processes [25]." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "H. Choi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "content": "In this work, we introduce a simple yet novel approach that focuses on token dynamics, termed Salience-Based Adaptive Masking (SBAM), which aims to strategically select masking locations by discerning perceptual prominence within the visual data. Specifically, the proposed method leverages the directional emphasis from attention mechanisms to identify the image tokens pivotal to the visual context. Significantly, our methodology stands apart from the conventional attention-based approach [25] by leveraging the token's outgoing weight to calculate its 'token salience' within the image and prioritizing those that have high salience to be masked. We also infuse a degree of randomness into token salience to enrich the diversity of mask generation. Further details of SBAM are elaborated in Sec. 3. Thus, without any significant additional cost, we can consider the token's prominence within the image." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 262, + 482, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 262, + 482, + 382 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 482, + 382 + ], + "type": "text", + "content": "Crucially, the strength of the SBAM approach lies in its robustness to the varying masking ratios—a notable vulnerability in established baselines [14,36, 39], which struggles from performance instability with even minor variations in optimized masking ratio. The robustness of SBAM is attributable to the proposed analytical precision in token dynamics, selectively masking tokens pivotal to the image's entirety. As a result, SBAM exhibits a diminished likelihood of masking redundant or non-essential segments, offering a stable alternative to the random masking or preceding masking approaches. Along with improving robustness to mask ratio variations, SBAM significantly enhances pre-training efficiency. A comprehensive evaluation of this is provided in Sec. 4." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 384, + 482, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 482, + 564 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 482, + 564 + ], + "type": "text", + "content": "Establishing robustness against variations in masking ratios has empowered us to expand the discourse on image masking into a pioneering aspect, introducing an innovative paradigm: an Adaptive Masking Ratio (AMR). We find having masking ratios that adapt throughout training to be highly effective, as it allows the masking process to be tailored to each sample in the dataset. For instance, each image may benefit from different masking ratios, as one image might have a close-up of a bear which would enjoy high masking ratios, whereas one might have an airplane in the sky that would require a lower masking ratio; see (c) of Fig. 1. The proposed token salience, reapplied in this context, serves to finely determine the dynamic masking ratio by quantifying the proportion of tokens exhibiting high salience. More details can be found in Sec. 5. The proposed AMR implements an adaptive strategy that respects the distinctiveness of visual data and thus achieves significant performance gains when applied to various baselines (refer to Sec. 6.2 and 6.3. We note that this type of 'tailoring' to each image for pre-training is impossible with any existing method." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 567, + 482, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 567, + 482, + 651 + ], + "spans": [ + { + "bbox": [ + 130, + 567, + 482, + 651 + ], + "type": "text", + "content": "In sum, this paper offers an innovative masking strategy that enhances the robustness and effectiveness of pre-trained models, setting the stage for a notable shift in the field of self-supervised learning. More importantly, the proposed method can be universally applied across any MIM framework that exploits token masking. We evaluate our SBAM and AMR on data-hungry models like ViT-L/ViT-B [12] on ImageNet-1K [9] datasets, achieving significant performance improvements in both fine-tuning and linear probing accuracy." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 147, + 653, + 312, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 653, + 312, + 665 + ], + "spans": [ + { + "bbox": [ + 147, + 653, + 312, + 665 + ], + "type": "text", + "content": "To summarize, our contributions are:" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "type": "text", + "content": "SBAM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 137, + 116, + 479, + 211 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 137, + 116, + 479, + 138 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 116, + 479, + 138 + ], + "spans": [ + { + "bbox": [ + 137, + 116, + 479, + 138 + ], + "type": "text", + "content": "- we present Saliency-Based Adaptive Masking (SBAM), a novel effective method for MIM pre-training that focuses on token salience;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 140, + 479, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 140, + 479, + 162 + ], + "spans": [ + { + "bbox": [ + 137, + 140, + 479, + 162 + ], + "type": "text", + "content": "- being saliency-based, without a significant increase in computation, we allow effective masking that is robust to masking ratios:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 137, + 164, + 479, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 164, + 479, + 186 + ], + "spans": [ + { + "bbox": [ + 137, + 164, + 479, + 186 + ], + "type": "text", + "content": "- empowered by the robustness to masking ratio, we propose an Adaptive Masking Ratio (AMR) that allows tailored masking for each sample;" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 189, + 479, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 189, + 479, + 211 + ], + "spans": [ + { + "bbox": [ + 137, + 189, + 479, + 211 + ], + "type": "text", + "content": "- we evaluate our method on ImageNet-1K datasets, achieving notable enhancements in both fine-tuning and linear probing accuracy." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 232, + 233, + 244 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 232, + 233, + 244 + ], + "spans": [ + { + "bbox": [ + 132, + 232, + 233, + 244 + ], + "type": "text", + "content": "2 Preliminaries" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 258, + 479, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 479, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 479, + 304 + ], + "type": "text", + "content": "Within the Masked Image Modeling (MIM) domain, the crux lies in the strategic corruption and subsequent reconstruction of image segments. This process hinges on two core operations: Random Masking of image tokens and Reconstruction of corrupted tokens to guide the learning process." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 306, + 479, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 306, + 479, + 353 + ], + "spans": [ + { + "bbox": [ + 130, + 306, + 479, + 353 + ], + "type": "text", + "content": "For a given image sequence " + }, + { + "bbox": [ + 130, + 306, + 479, + 353 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^{N \\times L \\times D}" + }, + { + "bbox": [ + 130, + 306, + 479, + 353 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 306, + 479, + 353 + ], + "type": "inline_equation", + "content": "N, L" + }, + { + "bbox": [ + 130, + 306, + 479, + 353 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 306, + 479, + 353 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 306, + 479, + 353 + ], + "type": "text", + "content": " denote batch size, number of tokens per image and dimensionality of each token respectively, and specified masking proportion " + }, + { + "bbox": [ + 130, + 306, + 479, + 353 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 306, + 479, + 353 + ], + "type": "text", + "content": ", the general random masking process conducted by " + }, + { + "bbox": [ + 130, + 306, + 479, + 353 + ], + "type": "inline_equation", + "content": "\\varPhi_{\\mathrm{mask}}" + }, + { + "bbox": [ + 130, + 306, + 479, + 353 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 249, + 365, + 479, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 365, + 479, + 376 + ], + "spans": [ + { + "bbox": [ + 249, + 365, + 479, + 376 + ], + "type": "interline_equation", + "content": "X _ {\\text {m a s k e d}}, M = \\Phi_ {\\text {m a s k}} (X, \\gamma), \\tag {1}", + "image_path": "42f850d5f7209c1355acd6675a70ba03abb38062b6cf53fb7ad86b0d34c894b7.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 387, + 480, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 387, + 480, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 387, + 480, + 422 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 387, + 480, + 422 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{masked}}" + }, + { + "bbox": [ + 130, + 387, + 480, + 422 + ], + "type": "text", + "content": " represents the visible tokens as a result of the post-application of the random mask " + }, + { + "bbox": [ + 130, + 387, + 480, + 422 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 387, + 480, + 422 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 130, + 387, + 480, + 422 + ], + "type": "inline_equation", + "content": "M \\in \\{0,1\\}^{N \\times L}" + }, + { + "bbox": [ + 130, + 387, + 480, + 422 + ], + "type": "text", + "content": " indicating the presence (1) or absence (0) of masking for each element." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 423, + 479, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 479, + 469 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 479, + 469 + ], + "type": "text", + "content": "The key component for learning in MIM is the reconstruction error " + }, + { + "bbox": [ + 130, + 423, + 479, + 469 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{MIM}}" + }, + { + "bbox": [ + 130, + 423, + 479, + 469 + ], + "type": "text", + "content": ", which is computed by the mean squared error (MSE) between the predicted representation of masked tokens, denoted as " + }, + { + "bbox": [ + 130, + 423, + 479, + 469 + ], + "type": "inline_equation", + "content": "\\hat{X}" + }, + { + "bbox": [ + 130, + 423, + 479, + 469 + ], + "type": "text", + "content": ", and their normalized original counterparts, represented by " + }, + { + "bbox": [ + 130, + 423, + 479, + 469 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 423, + 479, + 469 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 162, + 481, + 479, + 513 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 481, + 479, + 513 + ], + "spans": [ + { + "bbox": [ + 162, + 481, + 479, + 513 + ], + "type": "interline_equation", + "content": "\\bar {X} = \\frac {X - \\mu (X)}{\\sigma (X) + \\epsilon}, \\quad \\mathcal {L} _ {\\mathrm {M I M}} = \\frac {1}{\\sum_ {i , j} M _ {i , j}} \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {L} M _ {i, j} \\cdot \\| \\hat {X} _ {i, j} - \\bar {X} _ {i, j} \\| _ {2} ^ {2}. \\tag {2}", + "image_path": "63e4279096647a6747b81af0e300e6fecb51fe685ed6d4e5910966b11d05b7f1.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 523, + 479, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 523, + 479, + 582 + ], + "spans": [ + { + "bbox": [ + 131, + 523, + 479, + 582 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 131, + 523, + 479, + 582 + ], + "type": "inline_equation", + "content": "\\mu (\\cdot)" + }, + { + "bbox": [ + 131, + 523, + 479, + 582 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 523, + 479, + 582 + ], + "type": "inline_equation", + "content": "\\sigma (\\cdot)" + }, + { + "bbox": [ + 131, + 523, + 479, + 582 + ], + "type": "text", + "content": " represent the mean and standard deviation of " + }, + { + "bbox": [ + 131, + 523, + 479, + 582 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 131, + 523, + 479, + 582 + ], + "type": "text", + "content": " respectively. " + }, + { + "bbox": [ + 131, + 523, + 479, + 582 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\| _2^2" + }, + { + "bbox": [ + 131, + 523, + 479, + 582 + ], + "type": "text", + "content": " denotes the squared " + }, + { + "bbox": [ + 131, + 523, + 479, + 582 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 131, + 523, + 479, + 582 + ], + "type": "text", + "content": " norm and " + }, + { + "bbox": [ + 131, + 523, + 479, + 582 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 131, + 523, + 479, + 582 + ], + "type": "text", + "content": " is a small value for numerical stability. This formulation presents the core of MIM's training objective, focusing explicitly on the reconstruction of the masked portions of the input, encouraging the model to infer corrupted information from the unmasked context." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 603, + 340, + 617 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 603, + 340, + 617 + ], + "spans": [ + { + "bbox": [ + 131, + 603, + 340, + 617 + ], + "type": "text", + "content": "3 Salience-Based Masking (SBM)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "content": "Masked Image Modeling (MIM) has greatly pushed forward self-supervised learning in the visual domain, yet its standard practice of randomly masking tokens fails to fully capture the richness of visual information. To overcome this" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "type": "text", + "content": "H. Choi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "limitation, we present Salience-Based Masking (SBM), a novel masking methodology that selects masking locations by revisiting token dynamics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "text", + "content": "Given an input tensor " + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^{N \\times (L \\times D)}" + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "text", + "content": ", the first step involves computing an affinity matrix " + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "inline_equation", + "content": "\\mathcal{A} \\in \\mathbb{R}^{N \\times L \\times L}" + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "text", + "content": " through a batch matrix-matrix product between " + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "inline_equation", + "content": "X' \\in \\mathbb{R}^{N \\times (D \\times L)}" + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "text", + "content": ". Subsequently, we apply the softmax function " + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "inline_equation", + "content": "\\rho(\\cdot)" + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "text", + "content": " to the affinity matrix, resulting in a normalized affinity matrix denoted as " + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{A}} = \\rho(\\mathcal{A}) \\in \\mathbb{R}^{N \\times L \\times L}" + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "text", + "content": ". The softmax function is implemented as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 252, + 208, + 481, + 236 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 208, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 252, + 208, + 481, + 236 + ], + "type": "interline_equation", + "content": "\\rho (\\mathcal {A}) _ {n, i, j} = \\frac {e ^ {a _ {n , i , j}}}{\\sum_ {k = 1} ^ {L} e ^ {a _ {n , i , k}}}. \\tag {3}", + "image_path": "8189ceb140fac09d1c79ed38fd5bea900946277032f9e143c8beab209098502d.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 239, + 479, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 239, + 479, + 274 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 479, + 274 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 130, + 239, + 479, + 274 + ], + "type": "inline_equation", + "content": "a_{n,i,j}" + }, + { + "bbox": [ + 130, + 239, + 479, + 274 + ], + "type": "text", + "content": " represents an element of " + }, + { + "bbox": [ + 130, + 239, + 479, + 274 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 130, + 239, + 479, + 274 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 239, + 479, + 274 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 130, + 239, + 479, + 274 + ], + "type": "text", + "content": " is the base of the natural logarithm. Note that, for each " + }, + { + "bbox": [ + 130, + 239, + 479, + 274 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 239, + 479, + 274 + ], + "type": "text", + "content": "-th element with a shape of " + }, + { + "bbox": [ + 130, + 239, + 479, + 274 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{L\\times L}" + }, + { + "bbox": [ + 130, + 239, + 479, + 274 + ], + "type": "text", + "content": ", this function normalizes the rows of it, transforming them into probabilities that sum to 1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 275, + 480, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 275, + 480, + 322 + ], + "spans": [ + { + "bbox": [ + 130, + 275, + 480, + 322 + ], + "type": "text", + "content": "Crucially, our approach distinguishes itself from conventional attention-based methods by utilizing the sum of outgoing weight of each token to determine the 'token salience' " + }, + { + "bbox": [ + 130, + 275, + 480, + 322 + ], + "type": "inline_equation", + "content": "S \\in \\mathbb{R}^{N \\times L}" + }, + { + "bbox": [ + 130, + 275, + 480, + 322 + ], + "type": "text", + "content": " in the image, which is represented by the column-wise summed score of " + }, + { + "bbox": [ + 130, + 275, + 480, + 322 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{A}}" + }, + { + "bbox": [ + 130, + 275, + 480, + 322 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 192, + 329, + 481, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 329, + 481, + 354 + ], + "spans": [ + { + "bbox": [ + 192, + 329, + 481, + 354 + ], + "type": "interline_equation", + "content": "S = \\mathcal {N} \\left(\\Sigma_ {j = 1} ^ {L} \\hat {\\mathcal {A}} _ {:, j,:}\\right), \\quad \\mathcal {N} (x) = \\frac {x - \\operatorname* {m i n} (x)}{\\operatorname* {m a x} (x) - \\operatorname* {m i n} (x)}. \\tag {4}", + "image_path": "132c9095e9ca594d8b351d5195e9d9ef85adf7eb9eada034c9f6fb066f3094df.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 361, + 479, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 479, + 432 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 479, + 432 + ], + "type": "text", + "content": "In the affinity map, the row-wise score represents the incoming weight, reflecting the extent to which other tokens influence the corresponding token. Conversely, the column-wise score signifies the outgoing weight, indicating the contribution of the corresponding token to others. By summing these scores, it becomes feasible to quantify the token's overall impact on the image, thereby defining token salience." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 432, + 480, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 432, + 480, + 502 + ], + "spans": [ + { + "bbox": [ + 130, + 432, + 480, + 502 + ], + "type": "text", + "content": "The adaptive masking process of SBM is formulated with the token salience " + }, + { + "bbox": [ + 130, + 432, + 480, + 502 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 432, + 480, + 502 + ], + "type": "text", + "content": ". Notably, exclusive reliance on " + }, + { + "bbox": [ + 130, + 432, + 480, + 502 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 432, + 480, + 502 + ], + "type": "text", + "content": " for masking can precipitate a decline in performance (refer to Fig. 7). To mitigate this issue, we incorporated an element of randomness into the masking process to get the adjusted token salience, denoted as " + }, + { + "bbox": [ + 130, + 432, + 480, + 502 + ], + "type": "inline_equation", + "content": "\\tilde{S} = S + N" + }, + { + "bbox": [ + 130, + 432, + 480, + 502 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 432, + 480, + 502 + ], + "type": "inline_equation", + "content": "N \\sim U([0,0.5)^{N \\times L})" + }, + { + "bbox": [ + 130, + 432, + 480, + 502 + ], + "type": "text", + "content": " represents a noise realization sampled from a multivariate uniform distribution " + }, + { + "bbox": [ + 130, + 432, + 480, + 502 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 130, + 432, + 480, + 502 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 502, + 479, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 502, + 479, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 502, + 479, + 552 + ], + "type": "text", + "content": "Then, the sampling of tokens for masking is guided by " + }, + { + "bbox": [ + 130, + 502, + 479, + 552 + ], + "type": "inline_equation", + "content": "\\tilde{S}" + }, + { + "bbox": [ + 130, + 502, + 479, + 552 + ], + "type": "text", + "content": ". We sort " + }, + { + "bbox": [ + 130, + 502, + 479, + 552 + ], + "type": "inline_equation", + "content": "\\tilde{S}" + }, + { + "bbox": [ + 130, + 502, + 479, + 552 + ], + "type": "text", + "content": " in ascending order and select tokens corresponding to the top " + }, + { + "bbox": [ + 130, + 502, + 479, + 552 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 502, + 479, + 552 + ], + "type": "text", + "content": " scores, where " + }, + { + "bbox": [ + 130, + 502, + 479, + 552 + ], + "type": "inline_equation", + "content": "K = \\lceil L\\cdot (1 - \\gamma)\\rceil" + }, + { + "bbox": [ + 130, + 502, + 479, + 552 + ], + "type": "text", + "content": ". Consequently, an adaptive binary mask " + }, + { + "bbox": [ + 130, + 502, + 479, + 552 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 502, + 479, + 552 + ], + "type": "text", + "content": " is constructed, where " + }, + { + "bbox": [ + 130, + 502, + 479, + 552 + ], + "type": "inline_equation", + "content": "M_{i} = 0" + }, + { + "bbox": [ + 130, + 502, + 479, + 552 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 130, + 502, + 479, + 552 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 502, + 479, + 552 + ], + "type": "text", + "content": " selected tokens, and " + }, + { + "bbox": [ + 130, + 502, + 479, + 552 + ], + "type": "inline_equation", + "content": "M_{i} = 1" + }, + { + "bbox": [ + 130, + 502, + 479, + 552 + ], + "type": "text", + "content": " for the remainder." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 552, + 479, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 552, + 479, + 600 + ], + "spans": [ + { + "bbox": [ + 130, + 552, + 479, + 600 + ], + "type": "text", + "content": "The proposed token salience offers a more intuitive and cost-efficient strategy for determining which tokens to mask. SBM therefore revisits token dynamics in the image context, streamlining the conventional complex masking process and enabling strategic masking." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 617, + 271, + 629 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 617, + 271, + 629 + ], + "spans": [ + { + "bbox": [ + 132, + 617, + 271, + 629 + ], + "type": "text", + "content": "4 Evaluation on SBM" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "In this section, the evaluation critically examines the robustness of the proposed SBM, especially against varying masking ratios, while also assessing its enhanced" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "type": "text", + "content": "SBAM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 118, + 186, + 173 + ], + "blocks": [ + { + "bbox": [ + 135, + 118, + 186, + 173 + ], + "lines": [ + { + "bbox": [ + 135, + 118, + 186, + 173 + ], + "spans": [ + { + "bbox": [ + 135, + 118, + 186, + 173 + ], + "type": "image", + "image_path": "a560adb51a404aa3450d15ace64f511d6d5c9e7d40d7b8b109344ae887068002.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 193, + 118, + 245, + 173 + ], + "blocks": [ + { + "bbox": [ + 193, + 118, + 245, + 173 + ], + "lines": [ + { + "bbox": [ + 193, + 118, + 245, + 173 + ], + "spans": [ + { + "bbox": [ + 193, + 118, + 245, + 173 + ], + "type": "image", + "image_path": "b0619e75c53aa302265e480d7c4c1f6321ac32001d1103ff9f1d58a9a525775c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 251, + 118, + 303, + 173 + ], + "blocks": [ + { + "bbox": [ + 251, + 118, + 303, + 173 + ], + "lines": [ + { + "bbox": [ + 251, + 118, + 303, + 173 + ], + "spans": [ + { + "bbox": [ + 251, + 118, + 303, + 173 + ], + "type": "image", + "image_path": "c6e3f1a641ad4460397fbb55a23860bd185c34fb36c46918fdc0d6de02f54052.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 309, + 118, + 361, + 173 + ], + "blocks": [ + { + "bbox": [ + 309, + 118, + 361, + 173 + ], + "lines": [ + { + "bbox": [ + 309, + 118, + 361, + 173 + ], + "spans": [ + { + "bbox": [ + 309, + 118, + 361, + 173 + ], + "type": "image", + "image_path": "c84a942aa9f13d05966c61c6c0ff4e0cce5083ba49ac0d2a7d7dfec799251e10.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 368, + 118, + 419, + 173 + ], + "blocks": [ + { + "bbox": [ + 368, + 118, + 419, + 173 + ], + "lines": [ + { + "bbox": [ + 368, + 118, + 419, + 173 + ], + "spans": [ + { + "bbox": [ + 368, + 118, + 419, + 173 + ], + "type": "image", + "image_path": "71dd3a3aca882f066cadbdf64dc85ab9ae4136b6987edba0af510d158a41dbcb.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 427, + 118, + 477, + 173 + ], + "blocks": [ + { + "bbox": [ + 427, + 118, + 477, + 173 + ], + "lines": [ + { + "bbox": [ + 427, + 118, + 477, + 173 + ], + "spans": [ + { + "bbox": [ + 427, + 118, + 477, + 173 + ], + "type": "image", + "image_path": "49cfdf792a3ef54b8f0cadd109c8fa7f9447fdfbf3381fe86779073116251d22.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 135, + 178, + 186, + 233 + ], + "blocks": [ + { + "bbox": [ + 135, + 178, + 186, + 233 + ], + "lines": [ + { + "bbox": [ + 135, + 178, + 186, + 233 + ], + "spans": [ + { + "bbox": [ + 135, + 178, + 186, + 233 + ], + "type": "image", + "image_path": "3d43a9a8071b5895759e961d7de6b135f52dc3a1904df00e1201aae5bd1b9cd3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 244, + 482, + 310 + ], + "lines": [ + { + "bbox": [ + 130, + 244, + 482, + 310 + ], + "spans": [ + { + "bbox": [ + 130, + 244, + 482, + 310 + ], + "type": "text", + "content": "Fig. 2: Qualitative example of SBM. SBM introduces 'token salience' to prioritize and mask tokens with high significance. Hence, it is qualitatively confirmed that particularly important objects with high contribution within the image are selectively masked. Moreover, by integrating randomness with token salience, masks are probabilistically assigned to the background and less significant tokens, enriching the diversity of the token masking." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 193, + 178, + 245, + 233 + ], + "blocks": [ + { + "bbox": [ + 193, + 178, + 245, + 233 + ], + "lines": [ + { + "bbox": [ + 193, + 178, + 245, + 233 + ], + "spans": [ + { + "bbox": [ + 193, + 178, + 245, + 233 + ], + "type": "image", + "image_path": "362453e51f64e159de4494dfa60c2fa7b042f72f268d9b4c6d3f6048c0718333.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 251, + 178, + 303, + 233 + ], + "blocks": [ + { + "bbox": [ + 251, + 178, + 303, + 233 + ], + "lines": [ + { + "bbox": [ + 251, + 178, + 303, + 233 + ], + "spans": [ + { + "bbox": [ + 251, + 178, + 303, + 233 + ], + "type": "image", + "image_path": "ae77002fdd85ed3c1b27059859a73b4f875bec6904f1cc2a7f42634de2a583c8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 309, + 178, + 361, + 233 + ], + "blocks": [ + { + "bbox": [ + 309, + 178, + 361, + 233 + ], + "lines": [ + { + "bbox": [ + 309, + 178, + 361, + 233 + ], + "spans": [ + { + "bbox": [ + 309, + 178, + 361, + 233 + ], + "type": "image", + "image_path": "4ecb6741130c59b5ddbc5fa82de165083e161120aa6e0162215924853fe6c7bc.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 368, + 178, + 419, + 233 + ], + "blocks": [ + { + "bbox": [ + 368, + 178, + 419, + 233 + ], + "lines": [ + { + "bbox": [ + 368, + 178, + 419, + 233 + ], + "spans": [ + { + "bbox": [ + 368, + 178, + 419, + 233 + ], + "type": "image", + "image_path": "06cb8ff24ce7280d283e68c9c1222080d209610faa20521abb4a602590d4060b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 427, + 178, + 477, + 233 + ], + "blocks": [ + { + "bbox": [ + 427, + 178, + 477, + 233 + ], + "lines": [ + { + "bbox": [ + 427, + 178, + 477, + 233 + ], + "spans": [ + { + "bbox": [ + 427, + 178, + 477, + 233 + ], + "type": "image", + "image_path": "081a5cb1e6cafd855ff7268d878a25bae32832363ca878b0d611e6efdb206fa9.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 333, + 482, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 333, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 333, + 482, + 369 + ], + "type": "text", + "content": "pre-training efficiency, evidenced by performance gains and faster convergence. Moreover, Fig. 2 qualitatively shows that SBM selectively masks only those tokens that contribute significantly to the image." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 131, + 391, + 367, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 391, + 367, + 403 + ], + "spans": [ + { + "bbox": [ + 131, + 391, + 367, + 403 + ], + "type": "text", + "content": "4.1 Robustness to Masking Ratio Variability" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 415, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 482, + 545 + ], + "type": "text", + "content": "A crucial aspect of the evaluation of SBM focuses on the robustness, particularly in the context of varying masking ratios. Prior established baselines, such as the widely used Masked Autoencoder (MAE) [14], often exhibit significant performance fluctuations with even minimal adjustments to the masking ratio. This sensitivity undermines the practicality and generalizability of such methods, especially in diverse real-world scenarios where optimal masking ratios may not be consistent across datasets. In contrast, SBM introduces a novel approach of selectively masking tokens based on their salience within an image. As a result, SBM exhibits a diminished likelihood of masking redundant or trivial tokens, thereby maintaining a stable performance across a broad spectrum of masking ratios." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 548, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 548, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 548, + 482, + 641 + ], + "type": "text", + "content": "Fig. 3 showcases the performance stability of SBM against strong baseline MAE, underscoring its superior resilience to changes in masking ratios. Our evaluation leverages the Performance Improvement over Masking Ratio (PIMR) metric, a normalized measure that quantifies how each model's performance at a given masking ratio stands against its performance at the lowest ratio, thereby reflecting the relative improvement. This metric is instrumental in revealing the impact of increased masked data on model training. The PIMR is defined as follows:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 225, + 642, + 481, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 642, + 481, + 669 + ], + "spans": [ + { + "bbox": [ + 225, + 642, + 481, + 669 + ], + "type": "interline_equation", + "content": "\\operatorname {P I M R} (M) = \\frac {P (M) - P \\left(M _ {\\min }\\right)}{P \\left(M _ {\\max }\\right) - P \\left(M _ {\\min }\\right)}, \\tag {5}", + "image_path": "0f52f70c80fbbd2c46ac55c5afaed258c5915d8d40645b3b5b921fa0efcc9842.jpg" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "H. Choi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 118, + 299, + 199 + ], + "blocks": [ + { + "bbox": [ + 138, + 118, + 299, + 199 + ], + "lines": [ + { + "bbox": [ + 138, + 118, + 299, + 199 + ], + "spans": [ + { + "bbox": [ + 138, + 118, + 299, + 199 + ], + "type": "image", + "image_path": "f98fad1803f38a36bea58c23afec47752f79df3785241737c185304648206fb8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 318, + 118, + 476, + 199 + ], + "blocks": [ + { + "bbox": [ + 318, + 118, + 476, + 199 + ], + "lines": [ + { + "bbox": [ + 318, + 118, + 476, + 199 + ], + "spans": [ + { + "bbox": [ + 318, + 118, + 476, + 199 + ], + "type": "image", + "image_path": "25fb0377f585ec315dcca4ea2019690ad216dd9235a809757dc1d640a1d2bb35.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 140, + 201, + 299, + 283 + ], + "blocks": [ + { + "bbox": [ + 140, + 201, + 299, + 283 + ], + "lines": [ + { + "bbox": [ + 140, + 201, + 299, + 283 + ], + "spans": [ + { + "bbox": [ + 140, + 201, + 299, + 283 + ], + "type": "image", + "image_path": "67e4dcc61139872b2aef9a08068c192c06c70f3c82f3f442217c3ee0fe7dd2e7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 290, + 482, + 401 + ], + "lines": [ + { + "bbox": [ + 130, + 290, + 482, + 401 + ], + "spans": [ + { + "bbox": [ + 130, + 290, + 482, + 401 + ], + "type": "text", + "content": "Fig. 3: Evaluation of robustness across varied masking ratios. To evaluate the robustness of SBM, we report the comparative analysis of image classification performance on ImageNet-1K dataset [9] against the baseline method, MAE [14], using ViT-L [12] as a backbone. The upper graphs display the performance of the methods at different masking ratios, while the lower graphs illustrate the Performance Improvement over Masking Ratio (PIMR) and Global PIMR. These measures indicate the extent of each model's performance enhancement as the masking ratio increases from the lowest to higher ratios. SBM significantly outperforms MAE in every measures, demonstrating its superior effectiveness in handling various masking ratios and enhanced pre-training performances." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 317, + 202, + 396, + 283 + ], + "blocks": [ + { + "bbox": [ + 317, + 202, + 396, + 283 + ], + "lines": [ + { + "bbox": [ + 317, + 202, + 396, + 283 + ], + "spans": [ + { + "bbox": [ + 317, + 202, + 396, + 283 + ], + "type": "image", + "image_path": "19e344ef5296371a9b90a01819cea30f479e0b289fe77c259cd15aade64de994.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 397, + 202, + 475, + 283 + ], + "blocks": [ + { + "bbox": [ + 397, + 202, + 475, + 283 + ], + "lines": [ + { + "bbox": [ + 397, + 202, + 475, + 283 + ], + "spans": [ + { + "bbox": [ + 397, + 202, + 475, + 283 + ], + "type": "image", + "image_path": "831248965af41e0a1622a634587aa5cf2f02916da50e653cd69bbe0036f7a21e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 428, + 482, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 428, + 482, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 428, + 482, + 525 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 428, + 482, + 525 + ], + "type": "inline_equation", + "content": "P(M)" + }, + { + "bbox": [ + 130, + 428, + 482, + 525 + ], + "type": "text", + "content": " is the performance at masking ratio " + }, + { + "bbox": [ + 130, + 428, + 482, + 525 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 428, + 482, + 525 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 428, + 482, + 525 + ], + "type": "inline_equation", + "content": "P(M_{min})" + }, + { + "bbox": [ + 130, + 428, + 482, + 525 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 428, + 482, + 525 + ], + "type": "inline_equation", + "content": "P(M_{max})" + }, + { + "bbox": [ + 130, + 428, + 482, + 525 + ], + "type": "text", + "content": " denote the minimum and maximum observed performances, respectively. A PIMR value closer to 1 signifies a substantial improvement relative to the range of observed performances. The PIMR graph in classification accuracy demonstrate the robustness of the SBM strategy, where it maintains a competitive edge over MAE across various masking ratios. Notably, SBM shows a minimal performance drop at lower masking ratios compared to MAE, indicating its effectiveness even with sparse data presence." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 528, + 482, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 528, + 482, + 624 + ], + "spans": [ + { + "bbox": [ + 130, + 528, + 482, + 624 + ], + "type": "text", + "content": "This advantage is particularly evident in the Global PIMR metric, where SBM's performance remains consistently high. For the Global PIMR calculation, we normalize performance relative to the most extensive range of performance observed among all models under comparison. This means that instead of comparing to the best and worst performances of a single model, Global PIMR considers the best and worst across both MAE and SBM, which provides a universal performance context. Thus, the metric reflects a model's improvement not in isolation but rather in relation to its peers, which can be seen in the equation:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 203, + 642, + 481, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 642, + 481, + 669 + ], + "spans": [ + { + "bbox": [ + 203, + 642, + 481, + 669 + ], + "type": "interline_equation", + "content": "\\operatorname {G l o b a l} \\operatorname {P I M R} (M) = \\frac {P (M) - P \\left(M _ {G m i n}\\right)}{P \\left(M _ {G m a x}\\right) - P \\left(M _ {G m i n}\\right)}, \\tag {6}", + "image_path": "0460b6717c4245a5f0c8ed29b89d2e66852db859db1f2e60d0c85a6b09432cab.jpg" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "type": "text", + "content": "SBAM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 141, + 120, + 309, + 223 + ], + "blocks": [ + { + "bbox": [ + 141, + 120, + 309, + 223 + ], + "lines": [ + { + "bbox": [ + 141, + 120, + 309, + 223 + ], + "spans": [ + { + "bbox": [ + 141, + 120, + 309, + 223 + ], + "type": "image", + "image_path": "7ea3238ca914f38640f8ff25c457e094a269a97c513740e8ece77ba01da6b5e4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 233, + 482, + 301 + ], + "lines": [ + { + "bbox": [ + 130, + 233, + 482, + 301 + ], + "spans": [ + { + "bbox": [ + 130, + 233, + 482, + 301 + ], + "type": "text", + "content": "Fig. 4: Performance evaluation of SBM with respect to the pre-trained epochs. We report the comparison of image classification accuracy on ImageNet-1K [9], pre-trained on ViT-L [12]. The left graph displays fine-tuning accuracy, whereas the right graph illustrates linear probing accuracy, both over a range of pre-trained epochs. The curves illustrate that SBM surpasses MAE [14] in pre-training effectiveness in every trained epoch, and also validates its quicker attainment of converged performance levels." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 312, + 120, + 474, + 223 + ], + "blocks": [ + { + "bbox": [ + 312, + 120, + 474, + 223 + ], + "lines": [ + { + "bbox": [ + 312, + 120, + 474, + 223 + ], + "spans": [ + { + "bbox": [ + 312, + 120, + 474, + 223 + ], + "type": "image", + "image_path": "3dc5994f371a561c6673aa90fdd0878807ea032cc758ee8facdd83523f2aa9c1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 312, + 482, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 482, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 482, + 373 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 312, + 482, + 373 + ], + "type": "inline_equation", + "content": "P(M_{Gmax})" + }, + { + "bbox": [ + 130, + 312, + 482, + 373 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 312, + 482, + 373 + ], + "type": "inline_equation", + "content": "P(M_{Gmin})" + }, + { + "bbox": [ + 130, + 312, + 482, + 373 + ], + "type": "text", + "content": " are the global minimum and maximum performance values observed across both MAE and SBM models, respectively. This broader evaluation framework further underscores SBM's superior resilience and ability to maintain high accuracy across varying degrees of masking, outperforming the MAE baseline even when less information is available for learning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 373, + 482, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 373, + 482, + 433 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 482, + 433 + ], + "type": "text", + "content": "As shown in the graphs, SBM consistently outperforms MAE across a spectrum of masking ratios, demonstrating its remarkable stability even at lower ratios. This is because in random masking, as the masking ratio decreases, the chance of including a crucial token in the mask lowers; conversely, SBM consistently masks pivotal tokens, regardless of the masking ratio." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 434, + 483, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 483, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 483, + 529 + ], + "type": "text", + "content": "Furthermore, the classification accuracy graphs displayed above demonstrate that the application of SBM results in a notable enhancement in performance compared to MAE for all ratios. This reveals that the proposed SBM not only exhibits resilience to variations in the masking ratio but also significantly boosts pre-training efficacy, irrespective of the masking ratio. Strategically masking pivotal information enhances model performance and accelerates convergence by encouraging a comprehensive understanding of the visual context through a focus on essential tokens." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 548, + 328, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 548, + 328, + 561 + ], + "spans": [ + { + "bbox": [ + 131, + 548, + 328, + 561 + ], + "type": "text", + "content": "4.2 Enhanced Pre-training Efficiency" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Beyond the robustness to masking ratio variations, SBM's efficacy is further demonstrated through its enhanced pre-training capabilities. Traditional MIM strategies often require extensive computational resources and time for model convergence, primarily due to the indiscriminate masking of image tokens which can hinder the learning process by obfuscating essential visual cues. By strategically selecting pivotal tokens for masking, SBM ensures that crucial tokens are leveraged during training, which leads to substantial improvement in pre-training performance and facilitates a more focused model convergence process." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "H. Choi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 118, + 187, + 173 + ], + "blocks": [ + { + "bbox": [ + 136, + 118, + 187, + 173 + ], + "lines": [ + { + "bbox": [ + 136, + 118, + 187, + 173 + ], + "spans": [ + { + "bbox": [ + 136, + 118, + 187, + 173 + ], + "type": "image", + "image_path": "a645b3183736e68dc8d4bc9b1cc6cf11e0be2cf216bf55e9579bb387b779c257.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 194, + 118, + 246, + 173 + ], + "blocks": [ + { + "bbox": [ + 194, + 118, + 246, + 173 + ], + "lines": [ + { + "bbox": [ + 194, + 118, + 246, + 173 + ], + "spans": [ + { + "bbox": [ + 194, + 118, + 246, + 173 + ], + "type": "image", + "image_path": "7c37280ad50cea46575d4947b8ee2c3712b2d94928ade23614ea7a83180cab2b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 252, + 119, + 303, + 173 + ], + "blocks": [ + { + "bbox": [ + 252, + 119, + 303, + 173 + ], + "lines": [ + { + "bbox": [ + 252, + 119, + 303, + 173 + ], + "spans": [ + { + "bbox": [ + 252, + 119, + 303, + 173 + ], + "type": "image", + "image_path": "49bb5f48b63b7a079f447c01d62fef6325888fb1850682a4903a06ea0b39fb4b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 310, + 118, + 361, + 173 + ], + "blocks": [ + { + "bbox": [ + 310, + 118, + 361, + 173 + ], + "lines": [ + { + "bbox": [ + 310, + 118, + 361, + 173 + ], + "spans": [ + { + "bbox": [ + 310, + 118, + 361, + 173 + ], + "type": "image", + "image_path": "92e652d4a17f5c3e986a2eea8a3989fc981d3db1697fb9cb53bcba6e5130263a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 369, + 118, + 419, + 173 + ], + "blocks": [ + { + "bbox": [ + 369, + 118, + 419, + 173 + ], + "lines": [ + { + "bbox": [ + 369, + 118, + 419, + 173 + ], + "spans": [ + { + "bbox": [ + 369, + 118, + 419, + 173 + ], + "type": "image", + "image_path": "5f0574eb78a0656b4e4308a39de0c33a71a2b2f25b64f9e0cdfc8961a22e3c77.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 427, + 118, + 477, + 173 + ], + "blocks": [ + { + "bbox": [ + 427, + 118, + 477, + 173 + ], + "lines": [ + { + "bbox": [ + 427, + 118, + 477, + 173 + ], + "spans": [ + { + "bbox": [ + 427, + 118, + 477, + 173 + ], + "type": "image", + "image_path": "0dcc4e5b4de537eed820e1c53cecf91fbf63517b82698666a7c32a14e8255170.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 136, + 178, + 187, + 233 + ], + "blocks": [ + { + "bbox": [ + 136, + 178, + 187, + 233 + ], + "lines": [ + { + "bbox": [ + 136, + 178, + 187, + 233 + ], + "spans": [ + { + "bbox": [ + 136, + 178, + 187, + 233 + ], + "type": "image", + "image_path": "5bc341637fed440cec1fcee5784766b9819ef29c6791c74ee23463ef3edbb9ec.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 245, + 480, + 289 + ], + "lines": [ + { + "bbox": [ + 131, + 245, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 131, + 245, + 480, + 289 + ], + "type": "text", + "content": "Fig. 5: Qualitative example of SBAM, which combines SBM and AMR. Having masking ratios that adapt throughout training is highly effective, as it allows the masking process to be tailored to each sample in the dataset, accommodating the unique composition and object sizes within each image, as shown in the above qualitative samples." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 194, + 178, + 246, + 233 + ], + "blocks": [ + { + "bbox": [ + 194, + 178, + 246, + 233 + ], + "lines": [ + { + "bbox": [ + 194, + 178, + 246, + 233 + ], + "spans": [ + { + "bbox": [ + 194, + 178, + 246, + 233 + ], + "type": "image", + "image_path": "6da04f99d96b8e753059f6eee1ef6c22716e49d56010e2179b35cbcff0eeaa61.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 252, + 178, + 303, + 233 + ], + "blocks": [ + { + "bbox": [ + 252, + 178, + 303, + 233 + ], + "lines": [ + { + "bbox": [ + 252, + 178, + 303, + 233 + ], + "spans": [ + { + "bbox": [ + 252, + 178, + 303, + 233 + ], + "type": "image", + "image_path": "1fc530b9fd79b0b7351086e4d6003783bb06c6d5dbf0cbb9999ea83174e7d63d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 310, + 178, + 361, + 233 + ], + "blocks": [ + { + "bbox": [ + 310, + 178, + 361, + 233 + ], + "lines": [ + { + "bbox": [ + 310, + 178, + 361, + 233 + ], + "spans": [ + { + "bbox": [ + 310, + 178, + 361, + 233 + ], + "type": "image", + "image_path": "fb4ed28cceb34f79ccd7d4af6d4275ad09907532529d016de253a1d6c9f54390.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 369, + 178, + 419, + 233 + ], + "blocks": [ + { + "bbox": [ + 369, + 178, + 419, + 233 + ], + "lines": [ + { + "bbox": [ + 369, + 178, + 419, + 233 + ], + "spans": [ + { + "bbox": [ + 369, + 178, + 419, + 233 + ], + "type": "image", + "image_path": "75419a52424e7f759af059239dacc528dc2ab87d65be30223c614da88f7b4d5b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 427, + 178, + 477, + 233 + ], + "blocks": [ + { + "bbox": [ + 427, + 178, + 477, + 233 + ], + "lines": [ + { + "bbox": [ + 427, + 178, + 477, + 233 + ], + "spans": [ + { + "bbox": [ + 427, + 178, + 477, + 233 + ], + "type": "image", + "image_path": "d6f639e2d7bc7cba1176cfde6a7341e6ded0b2c774762e8b843fad45339967ee.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 306, + 480, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 306, + 480, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 306, + 480, + 426 + ], + "type": "text", + "content": "Fig. 4 showcases the comparative pre-training efficiency of SBM against MAE [14], underlining SBM's reduced pre-training duration without a tradeoff in accuracy. Initially, SBM secures a distinct lead in classification accuracy and continues to demonstrate this advantage across the training epochs, as seen in the left graph. The right graph, representing linear probing accuracy, further confirms SBM's higher initial performance, which stabilizes near peak levels well before 800 epochs. These findings highlight SBM's ability to prioritize significant features during early training stages, resulting in a new alternative for both convergence speed and pre-training performance in the pre-training of masked image models." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 447, + 346, + 461 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 346, + 461 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 346, + 461 + ], + "type": "text", + "content": "5 Adaptive Masking Ratio (AMR)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 475, + 480, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 475, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 480, + 533 + ], + "type": "text", + "content": "Achieving stability against changes in masking ratios has enabled us to advance the discourse on image masking, introducing a novel perspective: an Adaptive Masking Ratio (AMR). This innovative approach acknowledges the inherently diverse visual narratives presented by individual images and adjusts the masking ratio to fit different object sizes and classes within them." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 534, + 480, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 480, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 480, + 582 + ], + "type": "text", + "content": "The proposed token salience " + }, + { + "bbox": [ + 130, + 534, + 480, + 582 + ], + "type": "inline_equation", + "content": "S = \\mathcal{N}(\\Sigma_{j=1}^{L}\\hat{A}_{\\cdot,j,:})" + }, + { + "bbox": [ + 130, + 534, + 480, + 582 + ], + "type": "text", + "content": " forms the basis for determining AMRs. The AMR " + }, + { + "bbox": [ + 130, + 534, + 480, + 582 + ], + "type": "inline_equation", + "content": "R_{dyna}" + }, + { + "bbox": [ + 130, + 534, + 480, + 582 + ], + "type": "text", + "content": " is computed based on the distribution of salience scores across tokens, adjusted by a predefined variability parameter " + }, + { + "bbox": [ + 130, + 534, + 480, + 582 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 130, + 534, + 480, + 582 + ], + "type": "text", + "content": " and the base mask ratio " + }, + { + "bbox": [ + 130, + 534, + 480, + 582 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 534, + 480, + 582 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 222, + 594, + 480, + 607 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 594, + 480, + 607 + ], + "spans": [ + { + "bbox": [ + 222, + 594, + 480, + 607 + ], + "type": "interline_equation", + "content": "R _ {d y n a} = r - \\Delta r + 2 \\Delta r \\times \\operatorname {m e a n} \\left(1 _ {S > \\delta}\\right). \\tag {7}", + "image_path": "f1e2ac698e80c6459a64e780a68f4d3cff928a4b29d74e32ab75a5b3d90d4719.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 131, + 617, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 617, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 617, + 480, + 665 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 131, + 617, + 480, + 665 + ], + "type": "inline_equation", + "content": "\\Delta r" + }, + { + "bbox": [ + 131, + 617, + 480, + 665 + ], + "type": "text", + "content": " denotes the range of allowable variation in the masking ratio, as " + }, + { + "bbox": [ + 131, + 617, + 480, + 665 + ], + "type": "inline_equation", + "content": "R_{dyna}" + }, + { + "bbox": [ + 131, + 617, + 480, + 665 + ], + "type": "text", + "content": " can range from " + }, + { + "bbox": [ + 131, + 617, + 480, + 665 + ], + "type": "inline_equation", + "content": "r - \\Delta r" + }, + { + "bbox": [ + 131, + 617, + 480, + 665 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 131, + 617, + 480, + 665 + ], + "type": "inline_equation", + "content": "r + \\Delta r" + }, + { + "bbox": [ + 131, + 617, + 480, + 665 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 131, + 617, + 480, + 665 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 131, + 617, + 480, + 665 + ], + "type": "text", + "content": " is the salience threshold for distinguishing highly salient tokens, and 1 is the indicator function that identifies tokens exceeding " + }, + { + "bbox": [ + 131, + 617, + 480, + 665 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 131, + 617, + 480, + 665 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "type": "text", + "content": "SBAM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 141, + 120, + 307, + 223 + ], + "blocks": [ + { + "bbox": [ + 141, + 120, + 307, + 223 + ], + "lines": [ + { + "bbox": [ + 141, + 120, + 307, + 223 + ], + "spans": [ + { + "bbox": [ + 141, + 120, + 307, + 223 + ], + "type": "image", + "image_path": "0dee5c96a07a79db0317ff25da0709d9bcaf3bdc80cd115a0765494620ead434.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 238, + 482, + 316 + ], + "lines": [ + { + "bbox": [ + 130, + 238, + 482, + 316 + ], + "spans": [ + { + "bbox": [ + 130, + 238, + 482, + 316 + ], + "type": "text", + "content": "Fig. 6: Performance comparison of SBAM with respect to the pre-trained epochs. We report the comparison of image classification accuracy on ImageNet-1K [9], pre-trained on ViT-L [12]. The left graph illustrates classification accuracy across epochs, while the right graph shows the accuracy obtained through linear probing. Both results indicate a significant improvement in pre-training performance when AMR is applied, which not only achieves higher accuracy earlier in the training process but also maintains a lead at convergence." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 307, + 120, + 473, + 223 + ], + "blocks": [ + { + "bbox": [ + 307, + 120, + 473, + 223 + ], + "lines": [ + { + "bbox": [ + 307, + 120, + 473, + 223 + ], + "spans": [ + { + "bbox": [ + 307, + 120, + 473, + 223 + ], + "type": "image", + "image_path": "32c666333013583215b2fa259ba488783ad77485e71ba6d78e316589640bdba9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 335, + 482, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 482, + 479 + ], + "type": "text", + "content": "With " + }, + { + "bbox": [ + 130, + 335, + 482, + 479 + ], + "type": "inline_equation", + "content": "R_{dyna}" + }, + { + "bbox": [ + 130, + 335, + 482, + 479 + ], + "type": "text", + "content": " established, we adjust the number of tokens to be masked accordingly. This dynamic adjustment of masking ratios ensures that the masking process is not uniformly applied but is instead sensitive to the visual information's inherent salience, promoting a more effective learning mechanism by focusing on the informative segments of the image. Fig. 5 presents a qualitative example of AMR, demonstrating the effectiveness of adaptive masking ratios that customize the masking process for each dataset sample, thereby accounting for the unique composition and object sizes within each image. As a result, the proposed Saliency-Based Adaptive Masking (SBAM), which combines SBM and AMR, employs an adaptive approach, leading to enhanced performance across different models (See Fig. 6 and Tab. 1) and setting a new standard for tailored image masking." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 498, + 230, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 230, + 513 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 230, + 513 + ], + "type": "text", + "content": "6 Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 525, + 279, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 279, + 537 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 279, + 537 + ], + "type": "text", + "content": "6.1 Implementation Details" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Our evaluation approach involves deploying the proposed SBM and AMR against the baseline to assess performance enhancements of the comprehensive method, SBAM. To ensure a fair comparison, we maintain consistency with the baseline method's hyperparameters and network architectures. Notably, to preserve the integrity of our experiments, we ensured uniform hardware utilization and experimental conditions across both our method and the reproduced baseline, all employing " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "8^{*}\\mathrm{A}6000" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " GPUs. Given that this fixed GPU configuration diverges from those used in prior methods, discrepancies between our reproduced performance and the performance documented in existing papers may arise. For the fair experimental schedule, the reproduced 400-epoch performance of baseline" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "H. Choi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 173, + 179, + 440, + 258 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "type": "text", + "content": "Table 1: Comprehensive performance results of applying SBAM to various baseline methods. We report the comparison of image classification fine-tuning accuracy on ImageNet-1K [9] dataset. The consistent performance improvement of SBAM across various baseline methods demonstrates the efficacy of SBAM as a scalable methodology capable of enhancing a variety of MIM frameworks." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 173, + 179, + 440, + 258 + ], + "lines": [ + { + "bbox": [ + 173, + 179, + 440, + 258 + ], + "spans": [ + { + "bbox": [ + 173, + 179, + 440, + 258 + ], + "type": "table", + "html": "
MethodBaselineBaseline+SBAM
MAE (ViT-L) [14]84.385.1
MAE (ViT-B)82.983.6
BootMAE (ViT-B) [11]84.184.8
iBoT (ViT-B) [43]71.574.4
CMAE (ViT-B) [17]83.884.5
", + "image_path": "39b8bb661032b7d48fcfdcc1328e7100423267fa7f683e64104f4048f91a0efe.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 280, + 482, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 280, + 482, + 316 + ], + "spans": [ + { + "bbox": [ + 130, + 280, + 482, + 316 + ], + "type": "text", + "content": "methods and the proposed method were all equally measured in intermediate stages in the training towards 800 epochs. A more detailed implementation description can be found in the Supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 333, + 269, + 344 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 333, + 269, + 344 + ], + "spans": [ + { + "bbox": [ + 131, + 333, + 269, + 344 + ], + "type": "text", + "content": "6.2 Evaluation on SBAM" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 353, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 353, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 353, + 482, + 544 + ], + "type": "text", + "content": "The graphs in Fig. 6 demonstrate the efficacy of the SBAM in comparison to the MAE model throughout the pre-training phase. Specifically, the left graph indicates that SBAM starts with a higher accuracy than MAE at the 200 epochs mark and continues to outperform MAE [14] at every subsequent checkpoint. By the 800 epochs mark, SBAM not only achieves a significant accuracy enhancement but also shows a more rapid improvement in the earlier epochs, suggesting that SBAM requires fewer epochs to achieve similar or better performance compared to MAE. In the context of linear probing, depicted in the right graph, the trend is similar. SBAM consistently achieves higher accuracy than MAE from the outset, and this performance gap is maintained as training progresses. The curves for SBAM and MAE become flat toward 800 epochs, indicating that 800 epochs or later is the point of convergence in performance. This observation underscores SBAM's substantial advantage over MAE, both in terms of converged performance and convergence speed. Overall, the performance trends captured in these graphs suggest that SBAM is more efficient during pre-training, reaching higher levels of accuracy faster than MAE." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 562, + 323, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 562, + 323, + 574 + ], + "spans": [ + { + "bbox": [ + 131, + 562, + 323, + 574 + ], + "type": "text", + "content": "6.3 Evaluation on Various Baselines" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "The integration of SBAM into various baseline methodologies demonstrates a notable enhancement in pre-training performance, as summarized in Tab. 1. All experiments report the classification accuracy on 400 epochs, except for iBoT [43] which is pre-trained for 100 epochs. Applying SBAM to the large-scale variant of MAE [14] (MAE (ViT-L)) yields a noteworthy enhancement, elevating the baseline accuracy from " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "84.3\\%" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "85.1\\%" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": ", marking a significant advancement. Integration of SBAM to the MAE using ViT-B also experienced significant performance" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "type": "text", + "content": "SBAM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 160, + 153, + 274, + 209 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 304, + 148 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 304, + 148 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 304, + 148 + ], + "type": "text", + "content": "Table 2: Comparative evaluation of SBAM against the state-of-the-art masking strategy AMT [25]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 160, + 153, + 274, + 209 + ], + "lines": [ + { + "bbox": [ + 160, + 153, + 274, + 209 + ], + "spans": [ + { + "bbox": [ + 160, + 153, + 274, + 209 + ], + "type": "table", + "html": "
MethodAcc (%)
AM [25]82.5
AMT [25]82.8
SBAM83.6
", + "image_path": "b35f10a6c36ae51020c23b54d327d45c0f98280de54506a4376cf50a2b2417fd.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 316, + 153, + 465, + 210 + ], + "blocks": [ + { + "bbox": [ + 307, + 114, + 477, + 148 + ], + "lines": [ + { + "bbox": [ + 307, + 114, + 477, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 114, + 477, + 148 + ], + "type": "text", + "content": "Table 3: Comparative evaluation of SBAM against the state-of-the-art masking strategy AttMask [18]." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 153, + 465, + 210 + ], + "lines": [ + { + "bbox": [ + 316, + 153, + 465, + 210 + ], + "spans": [ + { + "bbox": [ + 316, + 153, + 465, + 210 + ], + "type": "table", + "html": "
MethodAcc (%)
AttMask-High [18]72.5
AttMask-Hint [18]72.8
SBAM74.4
", + "image_path": "57c0c628851955d0fbeae22447b197121bc70153b25c9396ac1a6e5de48648af.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 143, + 228, + 251, + 331 + ], + "blocks": [ + { + "bbox": [ + 143, + 228, + 251, + 331 + ], + "lines": [ + { + "bbox": [ + 143, + 228, + 251, + 331 + ], + "spans": [ + { + "bbox": [ + 143, + 228, + 251, + 331 + ], + "type": "image", + "image_path": "cbfd74e9497fcf453434baef5780b39e0832b2bc081f2442a9dafbf31bcc711c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 346, + 482, + 369 + ], + "lines": [ + { + "bbox": [ + 130, + 346, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 482, + 369 + ], + "type": "text", + "content": "Fig. 7: Comprehensive ablation studies of the impact of randomness, variance " + }, + { + "bbox": [ + 130, + 346, + 482, + 369 + ], + "type": "inline_equation", + "content": "(\\Delta r)" + }, + { + "bbox": [ + 130, + 346, + 482, + 369 + ], + "type": "text", + "content": ", and delta " + }, + { + "bbox": [ + 130, + 346, + 482, + 369 + ], + "type": "inline_equation", + "content": "(\\delta)" + }, + { + "bbox": [ + 130, + 346, + 482, + 369 + ], + "type": "text", + "content": " of the proposed SBAM approach." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 258, + 228, + 362, + 330 + ], + "blocks": [ + { + "bbox": [ + 258, + 228, + 362, + 330 + ], + "lines": [ + { + "bbox": [ + 258, + 228, + 362, + 330 + ], + "spans": [ + { + "bbox": [ + 258, + 228, + 362, + 330 + ], + "type": "image", + "image_path": "55beaab6d230ca3654e175ec57dee0e96c9a2968fa60094c540bb4edbe4e98aa.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 365, + 228, + 473, + 330 + ], + "blocks": [ + { + "bbox": [ + 365, + 228, + 473, + 330 + ], + "lines": [ + { + "bbox": [ + 365, + 228, + 473, + 330 + ], + "spans": [ + { + "bbox": [ + 365, + 228, + 473, + 330 + ], + "type": "image", + "image_path": "8a78552668ab7a6eaa6000b19c11cdeb2f422e232d98131ea7e7e5c4cd218373.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 396, + 482, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 396, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 482, + 456 + ], + "type": "text", + "content": "gains, underscoring that the benefits of SBAM are not limited to specific model architectures and can provide substantial pre-training efficacy for a variety of models. The performance enhancement of SBAM on various models including BootMAE [11], iBoT, and CMAE [17] highlights SBAM's generalizability and efficacy in augmenting various model structures with significant accuracy gains." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 457, + 482, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 457, + 482, + 517 + ], + "spans": [ + { + "bbox": [ + 130, + 457, + 482, + 517 + ], + "type": "text", + "content": "In conclusion, the consistent improvement across various baselines validates the efficacy of SBAM as a scalable enhancement tool. It not only boosts performance in standard settings but also bridges the gap in more challenging learning scenarios, marking it as a pivotal development in masked image modeling pretraining techniques." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 539, + 342, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 539, + 342, + 552 + ], + "spans": [ + { + "bbox": [ + 131, + 539, + 342, + 552 + ], + "type": "text", + "content": "6.4 Comparison with Masking Methods" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 557, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 629 + ], + "type": "text", + "content": "We conducted a comparative evaluation of our SBAM masking strategy against the established masking strategies of AMT and AttMask. This comparison was performed by applying SBAM to the baselines previously employed by AMT [25] and AttMask [18] methods, specifically MAE [14] (ViT-B, 400 epochs) and iBoT [43] (ViT-B, 100 epochs), to ascertain the enhancements introduced by our approach." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "In Tab. 2, the AM and AMT strategy [25], which implement basic masking and selective masking based on semantic importance, achieve an image classification fine-tuning accuracy of 82.5 and 82.8, respectively. When applied to the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "H. Choi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "same baseline model, SBAM outperforms both approaches by achieving an accuracy of 83.6. This implies that the efficacy of self-supervised pre-training can be maximized by defining token salience based on the outgoing weight of the token, as opposed to the incoming weight [25]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 482, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 482, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 482, + 260 + ], + "type": "text", + "content": "Furthermore, we compare SBAM to two variants of AttMask [18], applied within the iBoT [43] framework in Tab. 3. Both AttMask-High and AttMask-Hint incorporate the selective masking strategy of distillation setup which leverages similarity to classification token. SBAM stands out with an accuracy of " + }, + { + "bbox": [ + 130, + 164, + 482, + 260 + ], + "type": "inline_equation", + "content": "74.4\\%" + }, + { + "bbox": [ + 130, + 164, + 482, + 260 + ], + "type": "text", + "content": ", substantially higher than AttMask-High's " + }, + { + "bbox": [ + 130, + 164, + 482, + 260 + ], + "type": "inline_equation", + "content": "72.5\\%" + }, + { + "bbox": [ + 130, + 164, + 482, + 260 + ], + "type": "text", + "content": " and AttMask-Hint's " + }, + { + "bbox": [ + 130, + 164, + 482, + 260 + ], + "type": "inline_equation", + "content": "72.8\\%" + }, + { + "bbox": [ + 130, + 164, + 482, + 260 + ], + "type": "text", + "content": ". This highlights the superiority of the SBAM method, which can effectively improve the pre-training efficiency without the need for the additional computational cost of using a complex framework." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 274, + 244, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 274, + 244, + 285 + ], + "spans": [ + { + "bbox": [ + 131, + 274, + 244, + 285 + ], + "type": "text", + "content": "6.5 Ablation studies" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 289, + 482, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 446 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 446 + ], + "type": "text", + "content": "In Fig. 7, we provide comprehensive ablation studies of the impact of randomness, variance " + }, + { + "bbox": [ + 130, + 289, + 482, + 446 + ], + "type": "inline_equation", + "content": "(\\varDelta r)" + }, + { + "bbox": [ + 130, + 289, + 482, + 446 + ], + "type": "text", + "content": ", and delta " + }, + { + "bbox": [ + 130, + 289, + 482, + 446 + ], + "type": "inline_equation", + "content": "(\\delta)" + }, + { + "bbox": [ + 130, + 289, + 482, + 446 + ], + "type": "text", + "content": " of the proposed SBAM approach. We report the ImageNet-1K [9] classification accuracy achieved by SBAM using the baseline approach of MAE [14], trained for 400 epochs on ViT-L [12] as the backbone. The first graph depicts the impact of the incorporated randomness in the SBAM on model performance over various epochs. The plot reveals that integrating randomness with token saliency markedly enhances pre-training accuracy as the number of pre-trained epochs increases. The second and third graphs show the fine-tuning accuracy and linear probing performance ablations for the hyperparameters of SBAM. While fine-tuning accuracy remained stable across various hyperparameters, linear probing accuracy demonstrated relative sensitivity. We chose " + }, + { + "bbox": [ + 130, + 289, + 482, + 446 + ], + "type": "inline_equation", + "content": "\\varDelta r=0.15" + }, + { + "bbox": [ + 130, + 289, + 482, + 446 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 289, + 482, + 446 + ], + "type": "inline_equation", + "content": "\\delta=0.1" + }, + { + "bbox": [ + 130, + 289, + 482, + 446 + ], + "type": "text", + "content": " as optimal hyperparameters in both figures and were universally applicable across all baseline methodologies." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 460, + 238, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 460, + 238, + 472 + ], + "spans": [ + { + "bbox": [ + 131, + 460, + 238, + 472 + ], + "type": "text", + "content": "7 Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 481, + 302, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 481, + 302, + 495 + ], + "spans": [ + { + "bbox": [ + 131, + 481, + 302, + 495 + ], + "type": "text", + "content": "7.1 Masked Language Modeling" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 498, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 665 + ], + "type": "text", + "content": "Masked Language Modeling (MLM) [2,7,8,10,13,24,31,33,34,40] has become a keystone self-supervised learning paradigm in NLP, exemplified by groundbreaking models like BERT [10] and GPT [29,30]. By predicting masked tokens from their context, MLM has propelled NLP forward, enabling models to scale and perform adeptly on diverse tasks [3]. However, the considerable training time and computational demands of these models have spurred innovations aimed at increasing pre-training efficiency. For example, ALBERT [19] reduced parameters through embedding matrix factorization and shared layer parameters, while EarlyBERT [6] applied the principles of network pruning to optimize the training process. The curriculum masking approach of CCM [20] represents another advancement, strategically increasing the complexity of token masking to enhance learning. These efforts reflect a broader trend in the quest for efficiency, leading to models that maintain or exceed the capabilities of their predecessors with a fraction of the resource investment." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "type": "text", + "content": "SBAM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 285, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 285, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 285, + 129 + ], + "type": "text", + "content": "7.2 Masked Image Modeling" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 134, + 482, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 134, + 482, + 301 + ], + "spans": [ + { + "bbox": [ + 130, + 134, + 482, + 301 + ], + "type": "text", + "content": "In computer vision, Masked Image Modeling (MIM)) [1,4,5,11,14-16,23,26-28, 36,37,41-43] has emerged as a transformative technique, drawing parallels to the success of Masked Language Modeling (MLM) in NLP. MIM's central tenet involves predicting occluded parts of images to foster a nuanced understanding of visual content sans explicit labels. Early efforts adapting MLM concepts for visual data, such as iGPT [5], paved the way for more sophisticated methods. BEiT [1] utilized a pre-trained discrete variational autoencoder (dVAE) to produce target visual tokens. Further refinements in the technique have been observed in methods such as MAE [14] and SimMIM [36], which focus on direct prediction from unmasked image areas, refining the process of visual understanding. Efficiency in pre-training has been a critical frontier, leading to innovations like GreenMIM [16] and HiViT [42], which optimize hierarchical Vision Transformers (ViTs) by processing only the visible patches, significantly reducing computational overhead." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 301, + 482, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 301, + 482, + 541 + ], + "spans": [ + { + "bbox": [ + 130, + 301, + 482, + 541 + ], + "type": "text", + "content": "Recent advances in model pre-training have honed in on the strategic use of masking to enhance learning efficiency, with a particular emphasis on selecting which image regions to mask. Initiating this trend, ADIOS [32] leverages adversarial training to smartly select challenging segments for masking, setting a foundation for intelligent masking approaches. AttMask [18] and SemMAE [21] further this by utilizing self-attention and semantic information, respectively, to pinpoint and mask the most informative parts of an image, thereby prioritizing high-value areas over random masking. The Attention-Driven Masking and Throwing Strategy (AMT) strategy [25] refines this focus on semantics by employing self-attention to identify and eliminate redundant patches, achieving a delicate balance between precision and efficiency. In the realm of CLIP models, ACLIP [38] and Fast CLIP (FLIP) [22] adopt attentive masking strategies to optimize training, with Fast CLIP demonstrating the effectiveness of masking substantial portions of images for accelerated learning. MaskAlign [37] introduces an innovative teacher-student framework that bypasses the need for masked region reconstruction, aligning visible features with semantically rich intact image features to concentrate on the most informative parts. Together, these approaches illustrate a shift towards more strategic, intelligent masking techniques, significantly boosting the pre-training process by leveraging both the quantity and quality of masked inputs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 558, + 225, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 558, + 225, + 571 + ], + "spans": [ + { + "bbox": [ + 132, + 558, + 225, + 571 + ], + "type": "text", + "content": "8 Conclusions" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "type": "text", + "content": "The proposed Saliency-Based Adaptive Masking (SBAM) approach and the Adaptive Masking Ratio (AMR) significantly progress the field of MIM by introducing a method that adaptively masks image tokens with dynamic masking ratios based on their token salience. SBAM not only enhances pre-training efficiency and model performance on ImageNet-1K datasets but also introduces a novel way of considering the importance of token dynamics, thereby enabling models to learn more pivotal representations." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "text", + "content": "H. Choi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 140, + 479, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 479, + 161 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 479, + 161 + ], + "type": "text", + "content": "1. Bao, H., Dong, L., Piao, S., Wei, F.: Beit: Bert pre-training of image transformers. In: International Conference on Learning Representations (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 162, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 162, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 138, + 162, + 480, + 205 + ], + "type": "text", + "content": "2. Bao, H., Dong, L., Wei, F., Wang, W., Yang, N., Liu, X., Wang, Y., Gao, J., Piao, S., Zhou, M., et al.: Unilmv2: Pseudo-masked language models for unified language model pre-training. In: International conference on machine learning. pp. 642-652. PMLR (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 206, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 206, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 138, + 206, + 480, + 237 + ], + "type": "text", + "content": "3. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 239, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 239, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 138, + 239, + 480, + 259 + ], + "type": "text", + "content": "4. Cao, S., Xu, P., Clifton, D.A.: How to understand masked autoencoders. arXiv preprint arXiv:2202.03670 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 261, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 261, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 138, + 261, + 480, + 293 + ], + "type": "text", + "content": "5. Chen, M., Radford, A., Child, R., Wu, J., Jun, H., Luan, D., Sutskever, I.: Generative pretraining from pixels. In: International conference on machine learning. pp. 1691-1703. PMLR (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 294, + 480, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 294, + 480, + 315 + ], + "spans": [ + { + "bbox": [ + 138, + 294, + 480, + 315 + ], + "type": "text", + "content": "6. Chen, X., Cheng, Y., Wang, S., Gan, Z., Wang, Z., Liu, J.: Earlybert: Efficient bert training via early-bird lottery tickets. arXiv preprint arXiv:2101.00063 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 316, + 480, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 316, + 480, + 347 + ], + "spans": [ + { + "bbox": [ + 138, + 316, + 480, + 347 + ], + "type": "text", + "content": "7. Clark, K., Luong, M.T., Le, Q.V., Manning, C.D.: Electra: Pre-training text encoders as discriminators rather than generators. arXiv preprint arXiv:2003.10555 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 349, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 349, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 138, + 349, + 480, + 369 + ], + "type": "text", + "content": "8. Conneau, A., Lample, G.: Cross-lingual language model pretraining. Advances in neural information processing systems 32 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 371, + 480, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 371, + 480, + 403 + ], + "spans": [ + { + "bbox": [ + 138, + 371, + 480, + 403 + ], + "type": "text", + "content": "9. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 403, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 403, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 138, + 403, + 480, + 435 + ], + "type": "text", + "content": "0. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 436, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 436, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 436, + 480, + 468 + ], + "type": "text", + "content": "1. Dong, X., Bao, J., Zhang, T., Chen, D., Zhang, W., Yuan, L., Chen, D., Wen, F., Yu, N.: Bootstrapped masked autoencoders for vision bert pretraining. In: European Conference on Computer Vision. pp. 247-264. Springer (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 469, + 480, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 469, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 138, + 469, + 480, + 512 + ], + "type": "text", + "content": "2. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 513, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 513, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 513, + 480, + 544 + ], + "type": "text", + "content": "3. Ghazvininejad, M., Levy, O., Liu, Y., Zettlemoyer, L.: Mask-predict: Parallel decoding of conditional masked language models. arXiv preprint arXiv:1904.09324 (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 545, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 545, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 545, + 480, + 578 + ], + "type": "text", + "content": "4. He, K., Chen, X., Xie, S., Li, Y., Dólar, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16000-16009 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 578, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 480, + 599 + ], + "type": "text", + "content": "5. Hou, Z., Sun, F., Chen, Y.K., Xie, Y., Kung, S.Y.: Milan: Masked image pretraining on language assisted representation. arXiv preprint arXiv:2208.06049 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "type": "text", + "content": "6. Huang, L., You, S., Zheng, M., Wang, F., Qian, C., Yamasaki, T.: Green hierarchical vision transformer for masked image modeling. arXiv preprint arXiv:2205.13515 (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "text", + "content": "7. Huang, Z., Jin, X., Lu, C., Hou, Q., Cheng, M.M., Fu, D., Shen, X., Feng, J.: Contrastive masked autoencoders are stronger vision learners. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "type": "text", + "content": "SBAM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "18. Kakogeorgiou, I., Gidaris, S., Psomas, B., Avrithis, Y., Bursuc, A., Karantzalos, K., Komodakis, N.: What to hide from your students: Attention-guided masked image modeling. arXiv preprint arXiv:2203.12719 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 182 + ], + "type": "text", + "content": "19. Lan, Z., Chen, M., Goodman, S., Gimpel, K., Sharma, P., Soricut, R.: Albert: A lite bert for self-supervised learning of language representations. In: International Conference on Learning Representations" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "type": "text", + "content": "20. Lee, M., Park, J.H., Kim, J., Kim, K.M., Lee, S.: Efficient pre-training of masked language model via concept-based curriculum masking. In: Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "type": "text", + "content": "21. Li, G., Zheng, H., Liu, D., Wang, C., Su, B., Zheng, C.: Semmae: Semantic-guided masking for learning masked autoencoders. Advances in Neural Information Processing Systems 35, 14290-14302 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 480, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 282 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 282 + ], + "type": "text", + "content": "22. Li, Y., Fan, H., Hu, R., Feichtenhofer, C., He, K.: Scaling language-image pretraining via masking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23390-23400 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 282, + 480, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 480, + 304 + ], + "type": "text", + "content": "23. Liu, J., Huang, X., Liu, Y., Li, H.: Mixmim: Mixed and masked image modeling for efficient visual representation learning. arXiv preprint arXiv:2205.13137 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 480, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 336 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 336 + ], + "type": "text", + "content": "24. Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D., Levy, O., Lewis, M., Zettlemoyer, L., Stoyanov, V.: Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692 (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 337, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 337, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 337, + 480, + 369 + ], + "type": "text", + "content": "25. Liu, Z., Gui, J., Luo, H.: Good helper is around you: Attention-driven masked image modeling. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 1799-1807 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 370, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 370, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 370, + 480, + 391 + ], + "type": "text", + "content": "26. Pan, J., Zhou, P., Yan, S.: Towards understanding why mask-reconstruction pretraining helps in downstream tasks. arXiv preprint arXiv:2206.03826 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "type": "text", + "content": "27. Peng, Z., Dong, L., Bao, H., Ye, Q., Wei, F.: Beit v2: Masked image modeling with vector-quantized visual tokenizers. arXiv preprint arXiv:2208.06366 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 414, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 480, + 435 + ], + "type": "text", + "content": "28. Peng, Z., Dong, L., Bao, H., Ye, Q., Wei, F.: A unified view of masked image modeling. arXiv preprint arXiv:2210.10615 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 436, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 436, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 436, + 480, + 456 + ], + "type": "text", + "content": "29. Radford, A., Narasimhan, K., Salimans, T., Sutskever, I., et al.: Improving language understanding by generative pre-training (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "type": "text", + "content": "30. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners. OpenAI blog 1(8), 9 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 479, + 480, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 512 + ], + "type": "text", + "content": "31. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research 21(1), 5485-5551 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 513, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 513, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 513, + 480, + 544 + ], + "type": "text", + "content": "32. Shi, Y., Siddharth, N., Torr, P., Kosiorek, A.R.: Adversarial masking for self-supervised learning. In: International Conference on Machine Learning. pp. 20026-20040. PMLR (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 545, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 480, + 578 + ], + "type": "text", + "content": "33. Song, K., Tan, X., Qin, T., Lu, J., Liu, T.Y.: Mass: Masked sequence to sequence pre-training for language generation. In: International Conference on Machine Learning. pp. 5926-5936. PMLR (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 578, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 480, + 610 + ], + "type": "text", + "content": "34. Song, K., Tan, X., Qin, T., Lu, J., Liu, T.Y.: Mpnet: Masked and permuted pretraining for language understanding. Advances in Neural Information Processing Systems 33, 16857-16867 (2020)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "text", + "content": "35. Wu, J., Mo, S.: Object-wise masked autoencoders for fast pre-training. arXiv preprint arXiv:2205.14338 (2022)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "text", + "content": "36. Xie, Z., Zhang, Z., Cao, Y., Lin, Y., Bao, J., Yao, Z., Dai, Q., Hu, H.: Simmim: A simple framework for masked image modeling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9653-9663 (2022)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "text", + "content": "H. Choi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 315 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "37. Xue, H., Gao, P., Li, H., Qiao, Y., Sun, H., Li, H., Luo, J.: Stare at what you see: Masked image modeling without reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22732-22741 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "type": "text", + "content": "38. Yang, Y., Huang, W., Wei, Y., Peng, H., Jiang, X., Jiang, H., Wei, F., Wang, Y., Hu, H., Qiu, L., et al.: Attentive mask clip. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2771-2781 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 182, + 482, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 482, + 205 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 482, + 205 + ], + "type": "text", + "content": "39. Yi, K., Ge, Y., Li, X., Yang, S., Li, D., Wu, J., Shan, Y., Qie, X.: Masked image modeling with denoising contrast. arXiv preprint arXiv:2205.09616 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 205, + 482, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 482, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 482, + 237 + ], + "type": "text", + "content": "40. Zaken, E.B., Ravfogel, S., Goldberg, Y.: Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. arXiv preprint arXiv:2106.10199 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 237, + 482, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 482, + 259 + ], + "type": "text", + "content": "41. Zhang, Q., Wang, Y., Wang, Y.: How mask matters: Towards theoretical understandings of masked autoencoders. arXiv preprint arXiv:2210.08344 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "type": "text", + "content": "42. Zhang, X., Tian, Y., Xie, L., Huang, W., Dai, Q., Ye, Q., Tian, Q.: Hivit: A simpler and more efficient design of hierarchical vision transformer. In: The Eleventh International Conference on Learning Representations (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 293, + 482, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 315 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 315 + ], + "type": "text", + "content": "43. Zhou, J., Wei, C., Wang, H., Shen, W., Xie, C., Yuille, A., Kong, T.: ibot: Image bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832 (2021)" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "type": "text", + "content": "SBAM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Sapiens_ Foundation for Human Vision Models/726c3ced-d623-4195-b91f-26812a59bf2e_content_list.json b/2024/Sapiens_ Foundation for Human Vision Models/726c3ced-d623-4195-b91f-26812a59bf2e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3db49a0aadc21ea078a5e82dac6a3ab312f8df01 --- /dev/null +++ b/2024/Sapiens_ Foundation for Human Vision Models/726c3ced-d623-4195-b91f-26812a59bf2e_content_list.json @@ -0,0 +1,2072 @@ +[ + { + "type": "text", + "text": "Sapiens: Foundation for Human Vision Models", + "text_level": 1, + "bbox": [ + 233, + 140, + 767, + 162 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rawal Khirodkar, Timur Bagautdinov, Julieta Martinez, Su Zhaoen, Austin James, Peter Selednik, Stuart Anderson, Shunsuke Saito", + "bbox": [ + 228, + 189, + 774, + 219 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Codec Avatars Lab, Meta", + "bbox": [ + 413, + 232, + 588, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. We present Sapiens, a family of models for four fundamental human-centric vision tasks - 2D pose estimation, body-part segmentation, depth estimation, and surface normal prediction. Our models natively support 1K high-resolution inference and are extremely easy to adapt for individual tasks by simply fine-tuning foundation models pretrained on over 300 million in-the-wild human images. We observe that, given the same computational budget, self-supervised pretraining on a curated dataset of human images significantly boosts the performance for a diverse set of human-centric tasks. The resulting models exhibit remarkable generalization to in-the-wild data, even when labeled data is scarce or entirely synthetic. Our simple model design also brings scalability - model performance across tasks significantly improves as we scale the number of parameters from 0.3 to 2 billion. Sapiens consistently surpasses existing complex baselines across various human-centric benchmarks. Specifically, we achieve significant improvements over the prior state-of-the-art on Humans-5K (pose) by $7.6\\mathrm{mAP}$ , Humans-2K (part-seg) by $17.1\\mathrm{mIoU}$ , Hi4D (depth) by $22.4\\%$ relative RMSE, and THuman2 (normal) by $53.5\\%$ relative angular error.", + "bbox": [ + 261, + 258, + 738, + 507 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Human-centric, Large-scale pretraining, In-the-wild", + "bbox": [ + 261, + 512, + 692, + 526 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "\"Sapiens—pertaining to, or resembling modern humans.\"", + "bbox": [ + 292, + 539, + 705, + 554 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 574, + 375, + 589 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent years have witnessed remarkable strides towards generating photorealistic humans in 2D [17, 28, 50, 118] and 3D [69, 89, 102, 109]. The success of these methods is greatly attributed to the robust estimation of various assets such as 2D keypoints [14,67], fine-grained body-part segmentation [119], depth [113], and surface normals [89, 108]. However, robust and accurate estimation of these assets is still an active research area, and complicated systems to boost performance for individual tasks often hinder wider adoption. Moreover, obtaining accurate ground-truth annotation in-the-wild is notoriously difficult to scale. Our goal is to provide a unified framework and models to infer these assets in-the-wild to unlock a wide range of human-centric applications for everybody.", + "bbox": [ + 212, + 598, + 787, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We argue that such human-centric models should satisfy three criteria: generalization, broad applicability, and high fidelity. Generalization ensures robustness to unseen conditions, enabling the model to perform consistently across varied environments. Broad applicability indicates the versatility of the model, making it suitable for a wide range of tasks with minimal modifications. High fidelity denotes the ability of the model to produce precise, high-resolution outputs,", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5e2c5b544eea5e6c23ea0772008ae77fed6efeed3b8f0adaac0338426b159b3e.jpg", + "image_caption": [ + "Fig. 1: Sapiens models are finetuned for four human tasks - 2D pose estimation, body-part segmentation, depth prediction and normal prediction. Our models generalize across a variety of in-the-wild face, upper-body, full-body and multi-person images." + ], + "image_footnote": [], + "bbox": [ + 243, + 143, + 754, + 537 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "essential for faithful human generation tasks such as 2D to 3D lifting. This paper details the development of models that embody these attributes, collectively referred to as Sapiens.", + "bbox": [ + 212, + 590, + 787, + 636 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Following the insights from [34, 79, 91], leveraging large datasets and scalable model architectures is key for generalization. For broader applicability, we adopt the pretrain-then-finetune approach, enabling post-pretraining adaptation to specific tasks with minimal adjustments. This approach raises a critical question: What type of data is most effective for pretraining? Given computational limits, should the emphasis be on collecting as many human images as possible, or is it preferable to pretrain on a less curated set to better reflect real-world variability? Existing methods often overlook the pretraining data distribution in the context of downstream tasks. To study the influence of pretraining data distribution on human-specific tasks, we collect the Humans-300M dataset, featuring 300 million diverse human images. These unlabelled images are used to pretrain a family of vision transformers [27] from scratch, with parameter counts ranging from 300M to 2B.", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "R. Khirodkar et al.", + "bbox": [ + 271, + 114, + 400, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Among various self-supervision methods for learning general-purpose visual features from large datasets [5,19,34,45,46,121], we choose the masked-autoencoder (MAE) approach [45] for its simplicity and efficiency in pretraining. MAE, having a single-pass inference model compared to contrastive or multi-inference strategies, allows processing a larger volume of images with the same computational resources. For higher-fidelity, in contrast to prior methods, we increase the native input resolution of our pretraining to 1024 pixels, resulting in a $\\sim 4\\times$ increase in FLOPs compared to the largest existing vision backbone [91]. Each model is pretrained on 1.2 trillion tokens. Table 1 outlines a comparison with earlier approaches. For finetuning on human-centric tasks [15, 101, 113, 119], we use a consistent encoder-decoder architecture. The encoder is initialized with weights from pretraining, while the decoder, a lightweight and task-specific head, is initialized randomly. Both components are then finetuned in an end-to-end manner. We focus on four key tasks - 2D pose estimation, body-part segmentation, depth, and normal estimation, as shown in Fig. 1.", + "bbox": [ + 212, + 146, + 797, + 372 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Consistently with prior studies [56,122], we affirm the critical impact of label quality on the model's in-the-wild performance. Public benchmarks [23,41,55] often contain noisy labels, providing inconsistent supervisory signals during model fine-tuning. At the same time, it is important to utilize fine-grained and precise annotations to align closely with our primary goal of 3D human digitization. To this end, we propose a substantially denser set of 2D whole body keypoints for pose estimation and a detailed class vocabulary for body part segmentation, surpassing the scope of previous datasets (please refer to Fig. 1). Specifically, we introduce a comprehensive collection of 308 keypoints encompassing the body, hands, feet, surface, and face. Additionally, we expand the segmentation class vocabulary to 28 classes, covering body parts such as the hair, tongue, teeth, upper/lower lip, and torso. To guarantee the quality and consistency of annotations and a high degree of automation, we utilize a multi-view capture setup to collect pose and segmentation annotations. We also utilize human-centric synthetic data for depth and normal estimation, leveraging 600 detailed scans from RenderPeople [82] to generate high-resolution depth maps and surface normals.", + "bbox": [ + 212, + 373, + 789, + 616 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We show that the combination of domain-specific large-scale pretraining with limited, yet high-quality annotations leads to robust in-the-wild generalization. Overall, our method demonstrates an effective strategy for developing highly precise discriminative models capable of performing in real-world scenarios without the need for collecting a costly and diverse set of annotations.", + "bbox": [ + 212, + 614, + 787, + 690 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/8fe4c204156bfeef8bee9e13ce3c5462d871ca07f7a76c0827654a829885f249.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodDataset#ParamsGFLOPsImage sizeDomain
DINO [16]ImageNet1k86 M17.6224General
iBOT [121]ImageNet21k307 M61.6224General
DINOv2 [79]LVD-142M1 B291.0224General
ViT-6.5B [91]IG-3B6.5 B1657.0224General
AIM [34]DFN-2B6.5 B1657.0224General
Sapiens (Ours)Humans-300M2 B8709.01024Human
", + "bbox": [ + 284, + 694, + 722, + 795 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1: Comparison of state-of-the-art pretrained vision models. Sapiens adopts a higher resolution backbone on a large dataset of in-the-wild human images.", + "bbox": [ + 212, + 800, + 787, + 829 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Sapiens", + "bbox": [ + 678, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our contributions are summarized as follows.", + "bbox": [ + 238, + 146, + 568, + 160 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce Sapiens, a family of vision transformers pretrained on a large-scale, curated dataset of diverse human images.", + "- This study shows that simple data curation and large-scale pretraining significantly boost the model's performance with the same computational budget.", + "- Our models, fine-tuned with high-quality annotations or even synthetic data, demonstrate robust in-the-wild generalization.", + "- The first 1K high-resolution model that natively supports high-fidelity inference essential for human-centric tasks, achieving state-of-the-art performance on benchmarks for 2D pose estimation, body-part segmentation, depth, and normal estimation." + ], + "bbox": [ + 223, + 166, + 790, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 340, + 387, + 356 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our work explores the limits of training large architectures on a large number of in-the-wild human images. We build on prior work from different areas: pretraining at scale, human vision tasks, and large vision transformers.", + "bbox": [ + 212, + 372, + 784, + 417 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pretraining at Scale. The remarkable success of large-scale pretraining [26,95] followed by task-specific finetuning for language modeling [2,13,53,96,99,100] has established this approach as a standard practice. Similarly, computer vision methods [1,4,33,34,42,79,83,85,87,120] are progressively embracing extensive data scales for pretraining. The emergence of large datasets, such as LAION-5B [90], Instagram-3.5B [77], JFT-300M [92], LVD-142M [79], Visual Genome [60], and YFCC100M [97], has enabled the exploration of a data corpus well beyond the scope of traditional benchmarks [61,67,86]. Salient work in this domain includes DINOv2 [79], MAWS [91], and AIM [34]. DINOv2 achieves state-of-the-art performance in generating self-supervised features by scaling the contrastive iBot [121] method on the LDV-142M dataset [79]. MAWS [91] studies the scaling of masked-autoencoders (MAE) [45] on billion images. AIM [34] explores the scalability of autoregressive visual pretraining similar to BERT [26] for vision transformers [27]. In contrast to these methods which mainly focus on general image pretraining or zero-shot image classification, we take a distinctly human-centric approach: our models leverage a vast collection of human images for pretraining, subsequently fine-tuning for a range of human-related tasks.", + "bbox": [ + 212, + 424, + 787, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Human Vision Tasks. The pursuit of large-scale 3D human digitization [8, 44, 64, 74] remains a pivotal goal in computer vision [12]. Significant progress has been made within controlled or studio environments [3, 59, 63, 69, 70, 76, 89], yet challenges persist in extending these methods to unconstrained environments [29]. To address these challenges, developing versatile models capable of multiple fundamental tasks such as keypoint estimation [21, 36, 47, 51, 57, 78, 80, 93, 106], body-part segmentation [35, 40, 40, 41, 75, 104, 105], depth estimation [9, 10, 32, 43, 52, 66, 84, 113], and surface normal prediction [6, 7, 31, 39, 62, 88, 101, 108] from images in natural settings is crucial. In this work, we aim to develop models for these essential human vision tasks which generalize to in-the-wild settings.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "R. Khirodkar et al.", + "bbox": [ + 271, + 114, + 400, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Scaling Architectures. Currently, the largest publicly-accessible language models contain upwards of 100B parameters [49], while the more commonly used language models [94, 100] contain around 7B parameters. In contrast, Vision Transformers (ViT) [27], despite sharing a similar architecture, have not been scaled to this extent successfully. While there are notable endeavors in this direction, including the development of a dense ViT-4B [20] trained on both text and images, and the formulation of techniques for the stable training of a ViT-22B [25], commonly utilized vision backbones still range between 300M to 600M parameters [24, 38, 48, 68] and are primarily pretrained at an image resolution of about 224 pixels. Similarly, existing transformer-based image generation models, such as DiT [81] use less than 700M parameters, and operate on a highly compressed latent space. To address this gap, we introduce Sapiens - a collection of large, high-resolution ViT models that are pretrained natively at a 1024 pixel image resolution on millions of human images.", + "bbox": [ + 212, + 146, + 787, + 357 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 214, + 366, + 330, + 382 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Humans-300M Dataset", + "text_level": 1, + "bbox": [ + 214, + 387, + 452, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We utilize a large proprietary dataset for pretraining of approximately 1 billion in-the-wild images, focusing exclusively on human images. The preprocessing involves discarding images with watermarks, text, artistic depictions, or unnatural elements. Subsequently, we use an off-the-shelf person bounding-box detector [103] to filter images, retaining those with a detection score above 0.9 and bounding box dimensions exceeding 300 pixels. Fig. 2 provides an overview of the distribution of the number of people per image in over 248 million images contain multiple subjects.", + "bbox": [ + 212, + 411, + 606, + 578 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/453feb69b9add65076d75c0c6acdf971ef91e4808b045636a3f341b277320dac.jpg", + "image_caption": [ + "Fig.2: Overview of number of humans per image in the Humans-300M dataset." + ], + "image_footnote": [], + "bbox": [ + 614, + 351, + 785, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Pretraining", + "text_level": 1, + "bbox": [ + 214, + 589, + 357, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We follow the masked-autoencoder [45] (MAE) approach for pretraining. Our model is trained to reconstruct the original human image given its partial observation. Like all autoencoders, our model has an encoder that maps the visible image to a latent representation and a decoder that reconstructs the original image from this latent representation. Our pretraining dataset consists of both single and multi-human images; each image is resized to a fixed size with a square aspect ratio. Similar to ViT [27], we divide an image into regular non-overlapping patches with a fixed patch size. A subset of these patches is randomly selected and masked, leaving the rest visible. The proportion of masked patches to visible ones is defined as the masking ratio, which remains fixed throughout training. We refer to MAE [45] for more details. Fig. 3 (Top) shows the reconstruction of our pretrained model on unseen human images. Our models exhibit generalization across a variety of image characteristics including scales, crops, the age and ethnicity of subjects, and number of subjects. Each patch token in our model accounts for $0.02\\%$ of the image area compared to $0.4\\%$ in standard ViTs, a $16\\times$", + "bbox": [ + 212, + 613, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Sapiens", + "bbox": [ + 678, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/14bb481e41f6e40c447e02a883aadf2406e7016367fd43a989616fca2253ada2.jpg", + "image_caption": [ + "reduction - this provides a fine-grained inter-token reasoning for our models. Fig.3 (Bottom) shows that even with an increased mask ratio of $95\\%$ , our model achieves a plausible reconstruction of human anatomy on held-out samples." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 412 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 2D Pose Estimation", + "text_level": 1, + "bbox": [ + 215, + 482, + 426, + 496 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We follow the top-down 2D pose estimation paradigm, which aims to detect the locations of $K$ keypoints from an input image $\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}$ . Most methods pose this problem as heatmap prediction, where each of $K$ heatmaps represents the probability of the corresponding keypoint being at any spatial location. Similar to [111], we define a pose estimation transformer, $\\mathcal{P}$ , for keypoint detection. The bounding box at training and inference is scaled to $H \\times W$ and is provided as an input to $\\mathcal{P}$ . Let $\\mathbf{y} \\in \\mathbb{R}^{H \\times W \\times K}$ denote the $K$ heatmaps corresponding to the ground truth keypoints for a given input $\\mathbf{I}$ . The pose estimator transforms input $\\mathbf{I}$ to a set of predicted heatmaps, $\\hat{\\mathbf{y}} \\in \\mathbb{R}^{H \\times W \\times K}$ , such that $\\hat{\\mathbf{y}} = \\mathcal{P}(\\mathbf{I})$ . $\\mathcal{P}$ is trained to minimize the mean squared loss $\\mathcal{L}_{\\mathrm{pose}} = \\mathrm{MSE}(\\mathbf{y},\\hat{\\mathbf{y}})$ . During finetuning, the encoder of $\\mathcal{P}$ is initialized with the weights from pretraining, and the decoder is initialized randomly. The aspect ratio $H:W$ is set to be 4:3, with the pretrained positional embedding being interpolated accordingly [58]. We use lightweight decoders with deconvolution and convolution operations.", + "bbox": [ + 212, + 507, + 787, + 718 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We finetune both the encoder and the decoder in $\\mathcal{P}$ across multiple skeleton formats, including $K = 17$ [67], $K = 133$ [55] and a new highly-detailed skeleton format, with $K = 308$ , as shown in Fig. 4 (Left). Compared to existing formats with at most 68 facial keypoints, our annotations consist of 243 facial keypoints, including representative points around the eyes, eyebrows, lips, nose, and ears. This design is tailored to meticulously capture the nuanced details of facial expressions in the real world. With these keypoints, we manually annotated 1 million images at $4K$ resolution from an indoor capture setup.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "R. Khirodkar et al.", + "bbox": [ + 271, + 114, + 398, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c6937e5398dab721268539562b82ae78971bfcef323120b0a3eae4f1a07f8f1f.jpg", + "image_caption": [ + "a) Full-Body - 308 kps", + "c) Face - 243 kps", + "Fig. 4: Ground-truth annotations for 2D pose estimation and body-part segmentation." + ], + "image_footnote": [], + "bbox": [ + 222, + 145, + 385, + 268 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c4a8aa5300b95bf6610f839ee3521a86c8b6179f514319a51b85afc070ae7e33.jpg", + "image_caption": [ + "b) Hands - 40 kps" + ], + "image_footnote": [], + "bbox": [ + 385, + 146, + 465, + 191 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b823aadf68f180e5b7cf54bb07d58404e4f4c8c5b7f11f80d2c0ca12e2aa81f2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 390, + 205, + 460, + 267 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/be8e11fb02233f3669afa36ef6ae00a09a051c8df9cfac16407dd82738fe6cd0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 478, + 146, + 565, + 268 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2887a149f76bb9d884a9c99216ce4ea53cababd52aa66711d107bfbd8dca9ea6.jpg", + "image_caption": [ + "d) Body-Part Segmentation: 28 Classes" + ], + "image_footnote": [], + "bbox": [ + 566, + 147, + 650, + 268 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/90b238aa8bb10f6e41a8241bb013e9c7fd3c45915dcb9f69fd1208c0e5c8e372.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 148, + 784, + 268 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Body-Part Segmentation", + "text_level": 1, + "bbox": [ + 214, + 316, + 465, + 330 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Commonly referred to as human parsing, body-part segmentation aims to classify pixels in the input image $\\mathbf{I}$ into $C$ classes. Most methods [41] transform this problem to estimating per-pixel class probabilities to create a probability map $\\hat{\\mathbf{p}}\\in \\mathbb{R}^{H\\times W\\times C}$ such that $\\hat{\\mathbf{p}} = S(\\mathbf{I})$ , where $S$ is the segmentation model. As outlined previously, we adopt the same encoder-decoder architecture and initialization scheme for $S$ . $S$ is finetuned to minimize the weighted cross-entropy loss between the actual $\\mathbf{p}$ and predicted $\\hat{\\mathbf{p}}$ probability maps, $\\mathcal{L}_{\\mathrm{seg}} = \\mathrm{WeightedCE}(\\mathbf{p},\\hat{\\mathbf{p}})$ .", + "bbox": [ + 212, + 340, + 784, + 446 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We finetune $\\mathcal{S}$ across two part-segmentation vocabularies: a standard set with $C = 20$ [41] and a new larger vocabulary with $C = 28$ , as illustrated in Fig.4 (Right). Our proposed vocabulary goes beyond previous datasets in important ways. It distinguishes between the upper and lower halves of limbs and incorporates more detailed classifications such as upper/lower lips, teeth, and tongue. To this end, we manually annotate $100K$ images at $4K$ resolution with this vocabulary.", + "bbox": [ + 212, + 446, + 784, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.5 Depth Estimation", + "text_level": 1, + "bbox": [ + 214, + 574, + 408, + 589 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For depth estimation, we adopt the architecture used for segmentation, with the modification that the decoder output channel is set to 1 for regression. We denote the ground-truth depth map of image $\\mathbf{I}$ by $\\mathbf{d} \\in \\mathbb{R}^{H \\times W}$ , the depth estimator by $\\mathcal{D}$ , where $\\hat{\\mathbf{d}} = \\mathcal{D}(\\mathbf{I})$ , and $M$ as the number of human pixels in the image. For the relative depth estimation, we normalize $\\mathbf{d}$ to the range [0, 1] using max and min depths in the image. The $\\mathcal{L}_{\\mathrm{depth}}$ loss [32] for $\\mathcal{D}$ is defined as follows:", + "bbox": [ + 214, + 599, + 782, + 690 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\mathbf {d} = \\log (\\mathbf {d}) - \\log (\\hat {\\mathbf {d}}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 299, + 700, + 784, + 719 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\overline {{\\Delta \\mathbf {d}}} = \\frac {1}{M} \\sum_ {i = 1} ^ {M} \\Delta \\mathbf {d} _ {i}, \\quad \\overline {{(\\Delta \\mathbf {d}) ^ {2}}} = \\frac {1}{M} \\sum_ {i = 1} ^ {M} (\\Delta \\mathbf {d} _ {i}) ^ {2}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 299, + 723, + 784, + 763 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {d e p t h}} = \\sqrt {\\left(\\overline {{\\Delta \\mathbf {d}}} ^ {2}\\right) - \\frac {1}{2} \\left(\\overline {{\\Delta \\mathbf {d}}} ^ {2}\\right) ^ {2}}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 279, + 767, + 784, + 800 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We render 500,000 synthetic images using 600 high-resolution photogrammetry human scans as shown in Fig. 5 to obtain a robust monocular depth", + "bbox": [ + 214, + 809, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Sapiens", + "bbox": [ + 678, + 114, + 730, + 127 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "estimation model with high-fidelity. A random background is selected from a 100 HDRI environment map collection. We place a virtual camera within the scene, randomly adjusting its focal length, rotation, and translation to capture images and their associated ground-truth depth maps at 4K resolution.", + "bbox": [ + 212, + 146, + 787, + 207 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b0c67ba670a15af50b812c31374d01223983b839a9491536af1a0ccb7b56bbee.jpg", + "image_caption": [ + "Fig. 5: Ground-truth synthetic annotations for depth and surface normal estimation." + ], + "image_footnote": [], + "bbox": [ + 217, + 214, + 785, + 412 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.6 Surface Normal Estimation", + "text_level": 1, + "bbox": [ + 215, + 441, + 486, + 457 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Similar to previous tasks, we set the decoder output channels of the normal estimator $\\mathcal{N}$ to be 3, corresponding to the $xyz$ components of the normal vector at each pixel. The generated synthetic data is also used as supervision for surface normal estimation. Let $\\mathbf{n}$ be the ground-truth normal map for image $\\mathbf{I}$ and $\\hat{\\mathbf{n}} = \\mathcal{N}(\\mathbf{I})$ . Similar to depth, the loss $\\mathcal{L}_{\\mathrm{normal}}$ is only computed for human pixels in the image and is defined as follows:", + "bbox": [ + 212, + 469, + 787, + 560 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {n o r m a l}} = \\left\\| \\mathbf {n} - \\hat {\\mathbf {n}} \\right\\| _ {1} + (1 - \\mathbf {n} \\cdot \\hat {\\mathbf {n}}) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 571, + 785, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 599, + 375, + 616 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we initially provide an overview of the implementation details. Subsequently, we conduct comprehensive benchmarking across four tasks: pose estimation, part segmentation, depth estimation, and normal estimation.", + "bbox": [ + 212, + 632, + 785, + 678 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 691, + 457, + 708 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our largest model, Sapiens-2B, is pretrained using 1024 A100 GPUs for 18 days using PyTorch. We use the AdamW [73] optimizer for all our experiments. The learning schedule includes a brief linear warm-up, followed by cosine annealing [72] for pretraining and linear decay [65] for finetuning. All models are pretrained from scratch at a resolution of $1024 \\times 1024$ with a patch size of 16. For finetuning, the input image is resized to a 4:3 ratio, i.e. $1024 \\times 768$ . We use standard augmentations like cropping, scaling, flipping, and photometric distortions. A random background from non-human COCO [67] images is added for", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "R. Khirodkar et al.", + "bbox": [ + 271, + 114, + 400, + 127 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/89830f00747d92549f23f5fd6e7f15156f831b352c75be0798c7f7513e34ade5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Model#ParamsFLOPsHidden sizeLayersHeadsBatch size
Sapiens-0.3B0.336 B1.242 T1024241698,304
Sapiens-0.6B0.664 B2.583 T1280321665,536
Sapiens-1B1.169 B4.647 T1536402440,960
Sapiens-2B2.163 B8.709 T1920483220,480
", + "bbox": [ + 276, + 143, + 733, + 215 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "segmentation, depth, and normal prediction tasks. Importantly, we use differential learning rates [115] to preserve generalization i.e. lower learning rates for initial layers and progressively higher rates for subsequent layers. The layer-wise learning rate decay is set to 0.85 with a weight decay of 0.1 for the encoder. We detail the design specifications of Sapiens in Table. 2. Following [34, 100], we prioritize scaling models by width rather than depth. Note that the Sapiens-0.3B model, while architecturally similar to the traditional ViT-Large, consists of twentyfold more FLOPs due to its higher resolution.", + "bbox": [ + 212, + 239, + 784, + 359 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 2D Pose Estimation", + "text_level": 1, + "bbox": [ + 214, + 366, + 426, + 378 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We finetune Sapiens for face, body, feet, and hand ( $K = 308$ ) pose estimation on our high-fidelity annotations. For training, we use the train set with $1M$ images and for evaluation, we use the test set, named Humans-5K, with $5K$ images. Our evaluation is top-down [111] i.e. we use an off-the-shelf detector [37] for bounding-box and conduct single human pose inference. Table 3 shows a comparison of our models with existing methods for whole-body pose estimation. We evaluate all methods on 114 common keypoints between our 308 keypoint vocabulary and the 133 keypoint vocabulary from COCO-WholeBody [55]. Sapiens-0.6B surpasses the current state-of-the-art, DWPose-1 [114] by +2.8 AP. Contrary to DWPose [114], which utilizes a complex student-teacher framework with feature distillation tailored for the task, Sapiens adopts a general encoder-decoder architecture with large human-centric pretraining.", + "bbox": [ + 212, + 383, + 784, + 564 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Interestingly, even with the same parameter count, our models demonstrate superior performance compared to their counterparts. For instance, Sapiens-0.3B exceeds VitPose+-L by +5.6 AP, and Sapiens-0.6B outperforms VitPose+-H by +7.9 AP. Within the Sapiens family, our results indicate a direct correlation between model size and performance. Sapiens-2B sets a state-of-the-art with 61.1 AP, a significant improvement of +7.6 AP to the prior art. Despite fine-tuning", + "bbox": [ + 212, + 565, + 785, + 656 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/26011f812389c7d252523d01eebe61c9d8b586f816965f1f398a581dd1a9e692.jpg", + "table_caption": [ + "Table 2: Sapiens encoder specifications for pretraining on Human-300M dataset." + ], + "table_footnote": [], + "table_body": "
ModelInput SizeBodyFootFaceHandWhole-body
APARAPARAPARAPARAPAR
DeepPose [98]384 × 28832.143.525.341.237.853.915.731.623.937.2
SimpleBaseline [106]384 × 28852.360.149.862.559.667.341.451.844.653.7
HRNet [93]384 × 28855.862.645.255.458.964.539.347.645.753.9
ZoomNAS [110]384 × 28859.766.348.157.974.579.249.860.652.160.7
ViTPose+L [112]256 × 19261.066.862.468.250.155.741.547.347.853.6
ViTPose+-H [112]256 × 19261.667.463.269.050.756.342.047.848.354.1
RTMPose-x [54]384 × 28857.163.755.366.874.478.546.355.051.959.6
DWPose-m [114]256 × 19254.261.449.963.068.574.240.150.047.755.8
DWPose-l [114]384 × 28857.964.256.567.474.378.449.357.453.160.6
Sapiens-0.3B (Ours)1024 × 76858.164.556.867.774.578.649.657.753.4 (+0.3)60.9 (+0.3)
Sapiens-0.6B (Ours)1024 × 76859.865.564.772.375.279.052.160.356.2 (+2.8)62.4 (+2.1)
Sapiens-1B (Ours)1024 × 76862.968.268.375.176.479.755.963.459.4 (+5.9)65.3 (+5.1)
Sapiens-2B (Ours)1024 × 76864.769.969.476.276.979.957.164.461.1(+7.6)67.1(+7.0)
", + "bbox": [ + 225, + 664, + 792, + 830 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 3: Pose estimation results on Humans-5K test set. Flip test is used.", + "bbox": [ + 246, + 835, + 751, + 849 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Sapiens", + "bbox": [ + 678, + 114, + 730, + 127 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/59e713e7960beedb77ed778998fe60572b14023b7a02f3e5fd98f3be7656a33b.jpg", + "image_caption": [ + "Fig. 6: Pose estimation with Sapiens-1B for 308 keypoints on in-the-wild images." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 392, + 234 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/552a641dad9dc63dab04fbb96772b1ed292af4049983426ae4501f0361f0fcdb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 392, + 145, + 478, + 234 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/65b6b443723029db3d843fb46d4ae2061f4e00dc7a34c41fdb9cdc36b114f935.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 478, + 145, + 611, + 234 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/cb9f6c5d47c337b2a123ce6e609f19db0fe9e1720cb51bcc90df1bf918e79bb1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 611, + 145, + 784, + 234 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "with annotations from a indoor capture studio, Sapiens demonstrate robust generalization to real-world, as shown in Fig. 6.", + "bbox": [ + 212, + 268, + 782, + 299 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Body-Part Segmentation", + "text_level": 1, + "bbox": [ + 215, + 321, + 465, + 338 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We fine-tune and evaluate our annotations with a segmentation vocabulary of 28 classes. Our train set consists of $100K$ images, and the test set, Humans-2K, consists of $2K$ images. We compare Sapiens with existing bodypart segmentation methods fine-tuned on our train set. Importantly, we use suggested pretrained checkpoints by each method as initialization. Similar to pose, we observe generalization to segmentation as shown in Table 4.", + "bbox": [ + 212, + 349, + 542, + 498 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/d44e1a91d10252c825c33c02c58dc93a22dae06a5c3e89502ed9f707726760a1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelmIoU(%)mAcc(%)
FCN* [71]48.257.6
SegFormer* [107]53.562.9
Mask2Former* [22]58.768.3
DeepLabV3+* [18]64.174.8
Sapiens-0.3B (Ours)76.786.1
Sapiens-0.6B (Ours)77.886.3
Sapiens-1B (Ours)79.989.1
Sapiens-2B (Ours)81.289.4
", + "bbox": [ + 571, + 324, + 777, + 436 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 4: We report mIoU and mAcc on Humans-2K test set. Methods with * are trained by us.", + "bbox": [ + 553, + 443, + 785, + 484 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Interestingly, our smallest model, Sapiens-0.3B outperforms existing state-of-the-art segmentation methods like Mask2Former [22] and DeepLabV3+ [18] by 12.6 mIoU due to its higher resolution and large human-centric pretraining. Furthermore, increasing the model size improves segmentation performance. Sapiens-2B achieves the best performance of 81.2 mIoU and 89.4 mAcc on the test set. Fig. 7 shows the qualitative results of our models.", + "bbox": [ + 212, + 500, + 787, + 590 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.4 Depth Estimation", + "text_level": 1, + "bbox": [ + 215, + 604, + 410, + 619 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We evaluate our models on THuman2.0 [117] and Hi4D [116] datasets for depth estimation. THuman2.0 consists of 526 high-quality human scans, from which we derive three sets of images for testing: a) face, b) upper body, and c) full body using a virtual camera. THuman2.0 with 1578 images thus enables the evaluation of our models' performance on single-human images across multiple scales. Conversely, the Hi4D dataset focuses on multi-human scenarios, with each sequence showcasing two subjects engaged in activities involving human-human", + "bbox": [ + 212, + 630, + 787, + 737 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/3ff54b7187a6b7e52537224d6211d49ccc4d448604f0c8bd3651a36c6d0e9e36.jpg", + "image_caption": [ + "Fig. 7: Human segmentation with Sapiens-1B for 28 categories on in-the-wild images." + ], + "image_footnote": [], + "bbox": [ + 217, + 744, + 446, + 823 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/f7e1f5af38c018e0cc541d44bfda9c2f1ec853e339232a34c218bdb52f849f12.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 446, + 744, + 565, + 823 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/5eb79fee42cd7ce46d0985b236995a4b28539eab2ac721597fcd1d7e2972b782.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 565, + 744, + 617, + 823 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/074adc99db64f71394d54eafed3a97513acb0fda084d2275eb4ca8498207cdd8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 744, + 784, + 823 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "R. Khirodkar et al.", + "bbox": [ + 271, + 114, + 400, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/469297cfbb6c0870753dfb5f01902d78199abd0cad50ff491af5772f5c0b3f6f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTH2.0-FaceTH2.0-UprBodyTH2.0-FullBodyHi4D
RMSE ↓AbsRel ↓δ1 ↑RMSEAbsRelδ1RMSEAbsRelδ1RMSEAbsRelδ1
MiDaS-L [11]0.1140.0970.9250.3980.2710.8680.7010.6890.7820.2610.0820.975
MiDaS-Swin2 [11]0.0500.0360.9950.1220.0810.9480.2920.1710.8620.2090.0630.997
DepthAny-B [113]0.0390.0260.9990.0480.0280.9990.0610.0300.9990.1430.0340.997
DepthAny-L [113]0.0390.0270.9990.0480.0270.9990.0600.0300.9990.1470.0350.997
Sapiens-0.3B (Ours)0.0120.0081.0000.0150.0091.0000.0210.0101.0000.1480.0461.000
Sapiens-0.6B (Ours)0.0110.0081.0000.0150.0091.0000.0210.0101.0000.1420.0441.000
Sapiens-1B (Ours)0.0090.0061.0000.0120.0071.0000.0190.0091.0000.1250.0391.000
Sapiens-2B (Ours)0.0080.0051.0000.0100.0061.0000.0160.0081.0000.1140.0361.000
", + "bbox": [ + 220, + 133, + 805, + 256 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 5: Comparison of Sapiens for monocular depth estimation on human images.", + "bbox": [ + 218, + 257, + 779, + 272 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "interactions. We select sequences from pair 28, pair 32, and pair 37, featuring 6 unique subjects from camera 4, totaling 1195 multi-human real images for testing. We follow the relative-depth evaluation protocols established by MiDaSv3.1 [11], reporting standard metrics such as AbsRel and $\\delta_{1}$ . In addition, we also report RMSE as our primary metric since $\\delta_{1}$ does not effectively reflect performance in human scenes characterized by subtle depth variations.", + "bbox": [ + 212, + 277, + 784, + 368 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 5 compares our models with existing state-of-the-art monocular depth estimators. Sapiens-2B, finetuned solely on synthetic data, remarkably outperforms prior art across all single-human scales and multi-human scenarios. We observe a $20\\%$ RMSE reduction compared to the top-performing Depth-Anything model on Hi4D images. It is important to highlight that while baseline models are trained on a variety of scenes, Sapiens specializes in human-centric depth estimation. Fig. 8 presents a qualitative comparison of depth estimation between Sapiens-1B and DepthAnything-L. To ensure a fair comparison, the predicted depth is renormalized using the human mask in the baseline visualizations.", + "bbox": [ + 212, + 369, + 787, + 505 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/84ad7805d41928a1b011e4439b9e54b06ba160ab4940c7e5f56dbf429b91251a.jpg", + "image_caption": [ + "Fig. 8: We compare our depth prediction with DepthAnything [113]. To showcase the consistency of predicted depth, we also visualize the $\\nabla$ depth as pseudo surface normals." + ], + "image_footnote": [], + "bbox": [ + 215, + 515, + 787, + 792 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Sapiens", + "bbox": [ + 678, + 114, + 730, + 127 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.5 Surface Normal Estimation", + "text_level": 1, + "bbox": [ + 215, + 146, + 485, + 159 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The datasets for surface normal evaluation are identical to those used for depth estimation. Following [30], we report the mean and median angular error, along with the percentage of pixels within $t^\\circ$ error for $t \\in \\{11.25^\\circ, 22.5^\\circ, 30^\\circ\\}$ . Table 6 compares our models with existing human-specific surface normal estimators. All our models outperform existing methods by a significant margin. Sapiens-2B achieves a mean error of around $12^\\circ$ on the THuman2.0 (single-human) and Hi4D (multi-human) datasets. We qualitatively compare Sapiens-1B with PIFuHD [89] and ECON [108] for surface normal estimation in Figure 9. Note that PIFuHD [89] is trained with the identical set of 3D scans as ours, and ECON [108] is trained with 4000 scans that are a super set of our 3D scan data.", + "bbox": [ + 212, + 162, + 787, + 313 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/7482f87a6ff3f7fe9a68abce17970bbeb23ebf8500b7d07da035d762d5bd52b4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTHuman2.0 [117]Hi4D [116]
Angular Error°% Within t°Angular Error°% Within t°
MeanMedian11.25°22.5°30°MeanMedian11.25°22.5°30°
PIFuHD [89]30.5127.1315.8142.9758.8622.3919.2622.9860.1477.02
HDNet [52]34.8230.6017.4439.2654.5128.6026.8519.0857.9370.14
ICON [109]28.7425.5222.8147.8363.7320.1817.5226.8166.3482.73
ECON [108]25.4523.6732.9555.8669.0318.4616.4729.3568.1284.88
Sapiens-0.3B13.0210.3357.3786.2092.715.0412.2247.0781.4990.70
Sapiens-0.6B12.8610.2357.8586.6893.3014.0611.4750.5984.3792.54
Sapiens-1B12.119.4061.9788.0393.8412.189.5960.3688.6294.44
Sapiens-2B11.849.1663.1688.6094.1812.149.6260.2289.0894.74
", + "bbox": [ + 232, + 321, + 772, + 463 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 6: Comparison of Sapiens for surface normal estimation on human images.", + "bbox": [ + 230, + 465, + 772, + 481 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/d33afce36c657ac7f6dc7a049f860fec3cb6d7f447001294e5e56f018d805460.jpg", + "image_caption": [ + "Fig.9: Qualitative comparison of Sapiens-1B with PIFuHD [89] and ECON [108] for monocular surface normal estimation on challenging in-the-wild images." + ], + "image_footnote": [], + "bbox": [ + 256, + 484, + 743, + 833 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "R. Khirodkar et al.", + "bbox": [ + 271, + 114, + 398, + 127 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.6 Discussion", + "text_level": 1, + "bbox": [ + 215, + 146, + 349, + 159 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Importance of Pretraining Data Source. The feature quality is closely linked to the pretraining data quality. We assess the importance of pretraining on various data sources for human-centric tasks by pretraining Sapiens-0.3B on each dataset under identical training schedules and number of iterations. We fine-tune the model on each task and select early checkpoints for evaluation, reasoning that early-stage fine-tuning better reflects the model's generalization capability. We investigate the impact of pretraining at scale on general images (which may include humans) versus exclusively human images using Sapiens. We randomly select 100 million and 300 million general images from our 1 billion image corpus to create the General-100M and General-300M datasets, respectively. Table 7 showcases the comparison of pretraining outcomes. We report mAP for pose on Humans-5K, mIoU for segmentation on Humans-2K, RMSE for depth on THuman2.0, and mean angular error in degrees for normal estimation on Hi4D. Aligned with findings from [112], our results show that pretraining with Human300M leads to superior performance across all metrics, highlighting the benefits of human-centric pretraining within a fixed computational budget.", + "bbox": [ + 212, + 172, + 787, + 415 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We also study the effect of number of unique human images seen during pretraining with normal estimation performance. We report $\\%$ within $30^{\\circ}$ . Again, we maintain identical conditions for Sapiens-0.3B pretraining and finetuning. Fig.10 shows a steady improvement in performance as the pretraining data size increases without saturation. In summary, the diversity of human images during pretraining directly correlates with improved generalization to down-stream tasks.", + "bbox": [ + 212, + 415, + 517, + 598 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/f3e5ddc2209cf13d89e34b66ca5793794a923b1006ddb4e00fd8dccb76fd0079.jpg", + "image_caption": [ + "Fig. 10: Sapiens-0.3B's normal estimation performance with unique human images seen during pretraining." + ], + "image_footnote": [], + "bbox": [ + 524, + 426, + 787, + 598 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Zero-Shot Generalization. Our models exhibit broad generalization to a variety of settings. For instance, in segmentation, Sapiens are finetuned on single-human im", + "bbox": [ + 212, + 604, + 517, + 664 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "ages with limited subject diversity, minimal background variation, and solely third-person views (see Fig. 4). Nevertheless, our large-scale pretraining enables generalization across number of subjects, varying ages, and egocentric views, as shown in Fig. 11. These observations similarly hold for other tasks.", + "bbox": [ + 212, + 665, + 787, + 726 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/ddd9582152d5eca7d00db2c35b18ca23d810b5044a8668dd034015cc0cbd0a96.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Pretraining Source#ImagesPose (↑)Seg(↑)Depth(↓)Normal(↓)
Random Initialization-30.240.30.72035.4
General-100M100M35.750.10.35127.5
General-300M300M37.352.80.34726.8
Humans-100M100M43.661.20.31624.0
Humans-300M (Full)300M47.066.50.28821.8
", + "bbox": [ + 276, + 734, + 736, + 820 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 7: Comparison of Sapiens-0.3B pretrained on various data sources. A domain-specific pretraining yields superior results compared to general data sources.", + "bbox": [ + 212, + 821, + 785, + 851 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Sapiens", + "bbox": [ + 678, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/4d32dc066bdc97c77fd4eddfb2d6841af85623dff4abad49e54a0e317551bc00.jpg", + "image_caption": [ + "Multi-Human Generalization", + "Fig. 11: Sapiens achieve broad generalization via large human-centric pretraining." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 401, + 234 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/f8b0e003be8dc2dcc8d3db3aeec1cc5f04a5fa88fccb1bd4bdc59fdf69468f49.jpg", + "image_caption": [ + "Age Generalization" + ], + "image_footnote": [], + "bbox": [ + 401, + 143, + 576, + 234 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/1d80308ea50fc79ee73dafd05abf1e08af2f87c6e660abfac74723543ebf442c.jpg", + "image_caption": [ + "Viewpoint Generalization" + ], + "image_footnote": [], + "bbox": [ + 578, + 143, + 784, + 234 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Applications. Accurately predicting pose, body-part segmentation, depth and surface normals for in-the-wild human images enables a wide range of applications, such as 3D human scans and the generation of controllable images. We adapt ControlNet [118] to integrate keypoints for faces and hands predicted by Sapiens, and train this modified model on a subset of the Humans-300M dataset, as shown in Fig. 12. This grants more precise control over image generation, producing anatomically consistent human images.", + "bbox": [ + 212, + 276, + 784, + 381 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/c2f1a7606945920b7e60e88c7c902901987a86846deaed5a998c4e02df204c1a.jpg", + "image_caption": [ + "Image", + "Fig. 12: Image synthesis using ControlNet [118] trained on Sapiens pose predictions." + ], + "image_footnote": [], + "bbox": [ + 217, + 396, + 330, + 484 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/8c69392d70601b18d4164db9a9928890eb17689317cd4965cd91a2cdd318ceaa.jpg", + "image_caption": [ + "Predicted Pose" + ], + "image_footnote": [], + "bbox": [ + 330, + 397, + 441, + 484 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/edf1a9a3e3b1767e500cbe94d93ddb0ffbb7d898a636e1dca733f4481eb12ff5.jpg", + "image_caption": [ + "Pose-conditioned Human Image Synthesis" + ], + "image_footnote": [], + "bbox": [ + 447, + 396, + 560, + 484 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/b2845bd6118e2beec6fbbd1dc471d7fcbafc3b0bd8e12abaed6ce0f108bedf83.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 560, + 397, + 673, + 484 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/e0deb4fdb364e6ef015c964610534c3b14288f72b1f80b9ab58f475d273fd521.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 671, + 397, + 785, + 484 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations. While our models generally perform well, they are not perfect. Human images with complex/rare poses, crowding, and severe occlusion are challenging (see supplemental for details). Although aggressive data augmentation and a detect-and-crop strategy could mitigate these issues, we envision our models as a tool for acquiring large-scale, real-world supervision with human-in-the-loop to develop the next generations of human vision models.", + "bbox": [ + 212, + 534, + 784, + 625 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 645, + 359, + 660 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Sapiens represents a significant step toward elevating human-centric vision models into the realm of foundation models. Our models demonstrate strong generalization capabilities on a variety of human-centric tasks. We attribute the state-of-the-art performance of our models to: (i) large-scale pretraining on a large curated dataset, which is specifically tailored to understanding humans, (ii) scaled high-resolution and high-capacity vision transformer backbones, and (iii) high-quality annotations on augmented studio and synthetic data. We believe that these models can become a key building block for a multitude of downstream tasks, and provide access to high-quality vision backbones to a significantly wider part of the community. A potential direction for future work would be extending Sapiens to 3D and multi-modal datasets.", + "bbox": [ + 212, + 672, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "R. Khirodkar et al.", + "bbox": [ + 271, + 114, + 398, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 323, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Abnar, S., Dehghani, M., Neyshabur, B., Sedghi, H.: Exploring the limits of large scale pre-training. arXiv preprint arXiv:2110.02095 (2021) 4", + "2. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023) 4", + "3. Alldieck, T., Zanfir, M., Sminchisescu, C.: Photorealistic monocular 3d reconstruction of humans wearing clothing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1506-1515 (2022) 4", + "4. Bai, Y., Geng, X., Mangalam, K., Bar, A., Yuille, A., Darrell, T., Malik, J., Efros, A.A.: Sequential modeling enables scalable learning for large vision models. arXiv preprint arXiv:2312.00785 (2023) 4", + "5. Bao, H., Dong, L., Piao, S., Wei, F.: Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254 (2021) 3", + "6. Barron, J.T., Malik, J.: Intrinsic scene properties from a single rgb-d image. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 17-24 (2013) 4", + "7. Barron, J.T., Malik, J.: Shape, illumination, and reflectance from shading. IEEE transactions on pattern analysis and machine intelligence 37(8), 1670-1687 (2014) 4", + "8. Bartol, K., Bojanic, D., Petkovic, T., Pribanic, T.: A review of body measurement using 3d scanning. IEEE Access 9, 67281-67301 (2021) 4", + "9. Bhat, S.F., Alhashim, I., Wonka, P.: Adabins: Depth estimation using adaptive bins. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4009-4018 (2021) 4", + "10. Bhat, S.F., Birkl, R., Wofk, D., Wonka, P., Müller, M.: Zoodepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288 (2023) 4", + "11. Birkl, R., Wofk, D., Müller, M.: Midas v3. 1-a model zoo for robust monocular relative depth estimation. arXiv preprint arXiv:2307.14460 (2023) 11", + "12. Bojic, L.: Metaverse through the prism of power and addiction: what will happen when the virtual world becomes more attractive than reality? European Journal of Futures Research 10(1), 1-24 (2022) 4", + "13. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020) 4", + "14. Cao, Z., Simon, T., Wei, S., Sheikh, Y.: Realtime multi-person 2d pose estimation using part affinity fields. corr abs/1611.08050. arXiv preprint arXiv:1611.08050 (2016) 1", + "15. Cao, Z., Hidalgo, G., Simon, T., Wei, S.E., Sheikh, Y.: Openpose: real-time multi-person 2d pose estimation using part affinity fields. arXiv preprint arXiv:1812.08008 (2018) 3", + "16. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021) 3", + "17. Chan, C., Ginosar, S., Zhou, T., Efros, A.A.: Everybody dance now. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 5933-5942 (2019) 1" + ], + "bbox": [ + 225, + 172, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Sapiens", + "bbox": [ + 678, + 114, + 730, + 127 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Chen, L.C., Zhu, Y., Papandreou, G., Schroff, F., Adam, H.: Encoder-decoder with atrous separable convolution for semantic image segmentation. In: Proceedings of the European conference on computer vision (ECCV). pp. 801-818 (2018) 10", + "19. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International conference on machine learning. pp. 1597-1607. PMLR (2020) 3", + "20. Chen, X., Wang, X., Changpinyo, S., Piergiovanni, A., Padlewski, P., Salz, D., Goodman, S., Grycner, A., Mustafa, B., Beyer, L., et al.: Pali: A jointly-scaled multilingual language-image model. arXiv preprint arXiv:2209.06794 (2022) 5", + "21. Chen, Y., Wang, Z., Peng, Y., Zhang, Z., Yu, G., Sun, J.: Cascaded pyramid network for multi-person pose estimation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7103-7112 (2018) 4", + "22. Cheng, B., Misra, I., Schwing, A., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. arxiv 2022. arXiv preprint arXiv:2112.01527 (2021) 10", + "23. Couprie, C., Farabet, C., Najman, L., LeCun, Y.: Indoor semantic segmentation using depth information. arXiv preprint arXiv:1301.3572 (2013) 3", + "24. Dai, Z., Liu, H., Le, Q.V., Tan, M.: Coatnet: Marrying convolution and attention for all data sizes. Advances in neural information processing systems 34, 3965-3977 (2021) 5", + "25. Dehghani, M., Djolonga, J., Mustafa, B., Padlewski, P., Heek, J., Gilmer, J., Steiner, A.P., Caron, M., Geirhos, R., Alabdulmohsin, I., et al.: Scaling vision transformers to 22 billion parameters. In: International Conference on Machine Learning. pp. 7480-7512. PMLR (2023) 5", + "26. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018) 4", + "27. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) 2, 4, 5", + "28. Drobyshev, N., Chelishev, J., Khakhulin, T., Ivakhnenko, A., Lempitsky, V., Zakharov, E.: Megaportraits: One-shot megapixel neural head avatars. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 2663-2671 (2022) 1", + "29. Du, Y., Kips, R., Pumarola, A., Starke, S., Thabet, A., Sanakoyeu, A.: Avatars grow legs: Generating smooth human motion from sparse tracking inputs with diffusion model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 481-490 (2023) 4", + "30. Eftekhar, A., Sax, A., Malik, J., Zamir, A.: Omnidata: A scalable pipeline for making multi-task mid-level vision datasets from 3d scans. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10786-10796 (2021) 12", + "31. Eigen, D., Fergus, R.: Predicting depth, surface normals and semantic labels with a common multi-scale convolutional architecture. In: Proceedings of the IEEE international conference on computer vision. pp. 2650-2658 (2015) 4", + "32. Eigen, D., Puhrsch, C., Fergus, R.: Depth map prediction from a single image using a multi-scale deep network. Advances in neural information processing systems 27 (2014) 4, 7" + ], + "bbox": [ + 225, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "R. Khirodkar et al.", + "bbox": [ + 271, + 114, + 398, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "33. El-Nouby, A., Izacard, G., Touvron, H., Laptev, I., Jegou, H., Grave, E.: Are large-scale datasets necessary for self-supervised pre-training? arXiv preprint arXiv:2112.10740 (2021) 4", + "34. El-Nouby, A., Klein, M., Zhai, S., Bautista, M.A., Toshev, A., Shankar, V., Susskind, J.M., Joulin, A.: Scalable pre-training of large autoregressive image models. arXiv preprint arXiv:2401.08541 (2024) 2, 3, 4, 9", + "35. Fang, H.S., Lu, G., Fang, X., Xie, J., Tai, Y.W., Lu, C.: Weakly and semi supervised human body part parsing via pose-guided knowledge transfer. arXiv preprint arXiv:1805.04310 (2018) 4", + "36. Fang, H.S., Xie, S., Tai, Y.W., Lu, C.: Rmpe: Regional multi-person pose estimation. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 2334-2343 (2017) 4", + "37. Fang, Y., Sun, Q., Wang, X., Huang, T., Wang, X., Cao, Y.: Eva-02: A visual representation for neon genesis. arXiv preprint arXiv:2303.11331 (2023) 9", + "38. Fang, Y., Wang, W., Xie, B., Sun, Q., Wu, L., Wang, X., Huang, T., Wang, X., Cao, Y.: Eva: Exploring the limits of masked visual representation learning at scale. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19358-19369 (2023) 5", + "39. Fouhey, D.F., Gupta, A., Hebert, M.: Data-driven 3d primitives for single image understanding. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 3392-3399 (2013) 4", + "40. Gong, K., Liang, X., Li, Y., Chen, Y., Yang, M., Lin, L.: Instance-level human parsing via part grouping network. In: Proceedings of the European conference on computer vision (ECCV). pp. 770-785 (2018) 4", + "41. Gong, K., Liang, X., Zhang, D., Shen, X., Lin, L.: Look into person: Self-supervised structure-sensitive learning and a new benchmark for human parsing. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 932-940 (2017) 3, 4, 7", + "42. Goyal, P., Caron, M., Lefaudeauux, B., Xu, M., Wang, P., Pai, V., Singh, M., Liptchinsky, V., Misra, I., Joulin, A., et al.: Self-supervised pretraining of visual features in the wild. arXiv preprint arXiv:2103.01988 (2021) 4", + "43. Guizilini, V., Vasiljevic, I., Chen, D., Ambrus, R., Gaidon, A.: Towards zero-shot scale-aware monocular depth estimation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9233-9243 (2023) 4", + "44. Halstead, M.A., Barsky, B.A., Klein, S.A., Mandell, R.B.: Reconstructing curved surfaces from specular reflection patterns using spline surface fitting of normals. In: Proceedings of the 23rd annual conference on Computer graphics and interactive techniques. pp. 335-342 (1996) 4", + "45. He, K., Chen, X., Xie, S., Li, Y., Dollár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022) 3, 4, 5", + "46. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729-9738 (2020) 3", + "47. He, K., Gkioxari, G., Dollár, P., Girshick, R.B.: Mask r-cnn. corr abs/1703.06870 (2017). arXiv preprint arXiv:1703.06870 (2017) 4", + "48. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) 5" + ], + "bbox": [ + 225, + 146, + 784, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Sapiens", + "bbox": [ + 678, + 114, + 730, + 127 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "49. Hoffmann, J., Borgeaud, S., Mensch, A., Buchatskaya, E., Cai, T., Rutherford, E., Casas, D.d.L., Hendricks, L.A., Welbl, J., Clark, A., et al.: Training compute-optimal large language models. arXiv preprint arXiv:2203.15556 (2022) 5", + "50. Hu, L., Gao, X., Zhang, P., Sun, K., Zhang, B., Bo, L.: animate anyone: Consistent and controllable image-to-video synthesis for character animation. arXiv preprint arXiv:2311.17117 (2023) 1", + "51. Huang, S., Gong, M., Tao, D.: A coarse-fine network for keypoint localization. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 3028-3037 (2017) 4", + "52. Jafarian, Y., Park, H.S.: Learning high fidelity depths of dressed humans by watching social media dance videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12753-12762 (2021) 4, 12", + "53. Jiang, A.Q., Sablayrolles, A., Mensch, A., Bamford, C., Chaplot, D.S., Casas, D.d.l., Bressand, F., Lengyel, G., Lample, G., Saulnier, L., et al.: Mistral 7b. arXiv preprint arXiv:2310.06825 (2023) 4", + "54. Jiang, T., Lu, P., Zhang, L., Ma, N., Han, R., Lyu, C., Li, Y., Chen, K.: Rtmpose: Real-time multi-person pose estimation based on mmpose. arXiv preprint arXiv:2303.07399 (2023) 9", + "55. Jin, S., Xu, L., Xu, J., Wang, C., Liu, W., Qian, C., Ouyang, W., Luo, P.: Whole-body human pose estimation in the wild. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IX 16. pp. 196-214. Springer (2020) 3, 6, 9", + "56. Kato, N., Li, T., Nishino, K., Uchida, Y.: Improving multi-person pose estimation using label correction. arXiv preprint arXiv:1811.03331 (2018) 3", + "57. Khirodkar, R., Chari, V., Agrawal, A., Tyagi, A.: Multi-instance pose networks: Rethinking top-down pose estimation. In: Proceedings of the IEEE/CVF International conference on computer vision. pp. 3122-3131 (2021) 4", + "58. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023) 6", + "59. Kocabas, M., Chang, J.H.R., Gabriel, J., Tuzel, O., Ranjan, A.: Hugs: Human gaussian splats. arXiv preprint arXiv:2311.17910 (2023) 4", + "60. Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., et al.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision 123, 32-73 (2017) 4", + "61. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images. arXiv preprint (2009) 4", + "62. Ladicky, L., Zeisl, B., Pollefeys, M.: Discriminatively trained dense surface normal estimation. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 468-484. Springer (2014) 4", + "63. Lawrence, J., Goldman, D.B., Achar, S., Blascovich, G.M., Desloge, J.G., Fortes, T., Gomez, E.M., Haberling, S., Hoppe, H., Huibers, A., et al.: Project starline: A high-fidelity telepresence system. arXiv preprint (2021) 4", + "64. Levoy, M., Pulli, K., Curless, B., Rusinkiewicz, S., Koller, D., Pereira, L., Ginzton, M., Anderson, S., Davis, J., Ginsberg, J., et al.: The digital michelangelo project: 3d scanning of large statues. In: Proceedings of the 27th annual conference on Computer graphics and interactive techniques. pp. 131-144 (2000) 4", + "65. Lewkowycz, A.: How to decay your learning rate. arXiv preprint arXiv:2103.12682 (2021) 8" + ], + "bbox": [ + 225, + 146, + 784, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "R. Khirodkar et al.", + "bbox": [ + 271, + 114, + 398, + 126 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "66. Li, Z., Wang, X., Liu, X., Jiang, J.: Binsformer: Revisiting adaptive bins for monocular depth estimation. arXiv preprint arXiv:2204.00987 (2022) 4", + "67. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: European conference on computer vision. pp. 740-755. Springer (2014) 1, 4, 6, 8", + "68. Liu, Z., Hu, H., Lin, Y., Yao, Z., Xie, Z., Wei, Y., Ning, J., Cao, Y., Zhang, Z., Dong, L., et al.: Swin transformer v2: Scaling up capacity and resolution. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12009-12019 (2022) 5", + "69. Lombardi, S., Simon, T., Saragih, J., Schwartz, G., Lehrmann, A., Sheikh, Y.: Neural volumes: Learning dynamic renderable volumes from images. arXiv preprint arXiv:1906.07751 (2019) 1, 4", + "70. Lombardi, S., Simon, T., Schwartz, G., Zollhoefer, M., Sheikh, Y., Saragih, J.: Mixture of volumetric primitives for efficient neural rendering. ACM Transactions on Graphics (ToG) 40(4), 1-13 (2021) 4", + "71. Long, J., Shelhamer, E., Darrell, T., Berkeley, U.: Fully convolutional networks for semantic segmentation. arxiv 2015. arXiv preprint arXiv:1411.4038 (2014) 10", + "72. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983 (2016) 8", + "73. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017) 8", + "74. Lowe, D.G.: Three-dimensional object recognition from single two-dimensional images. Artificial intelligence 31(3), 355-395 (1987) 4", + "75. Luo, Y., Zheng, Z., Zheng, L., Guan, T., Yu, J., Yang, Y.: Macro-micro adversarial network for human parsing. In: Proceedings of the European conference on computer vision (ECCV). pp. 418-434 (2018) 4", + "76. Ma, S., Simon, T., Saragih, J., Wang, D., Li, Y., De La Torre, F., Sheikh, Y.: Pixel codec avatars. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 64-73 (2021) 4", + "77. Mahajan, D., Girshick, R., Ramanathan, V., He, K., Paluri, M., Li, Y., Bharambe, A., Van Der Maaten, L.: Exploring the limits of weakly supervised pretraining. In: Proceedings of the European conference on computer vision (ECCV). pp. 181-196 (2018) 4", + "78. Newell, A., Yang, K., Deng, J.: Stacked hourglass networks for human pose estimation. In: European conference on computer vision. pp. 483-499. Springer (2016) 4", + "79. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023) 2, 3, 4", + "80. Papandreou, G., Zhu, T., Kanazawa, N., Toshev, A., Tompson, J., Bregler, C., Murphy, K.: Towards accurate multi-person pose estimation in the wild. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4903-4911 (2017) 4", + "81. Peebles, W., Xie, S.: Scalable diffusion models with transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4195-4205 (2023) 5", + "82. render people: 3d people for architectural visualization | renderpeople, https://renderpeople.com/3d-people, accessed: 2024-02-22 3" + ], + "bbox": [ + 223, + 146, + 784, + 839 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "Sapiens", + "bbox": [ + 678, + 114, + 730, + 127 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "83. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) 4", + "84. Ranftl, R., Lasinger, K., Hafner, D., Schindler, K., Koltun, V.: Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence 44(3), 1623-1637 (2020) 4", + "85. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022) 4", + "86. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al.: Imagenet large scale visual recognition challenge. International journal of computer vision 115, 211-252 (2015) 4", + "87. Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022) 4", + "88. Saito, S., Huang, Z., Natsume, R., Morishima, S., Kanazawa, A., Li, H.: Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 2304-2314 (2019) 4", + "89. Saito, S., Simon, T., Saragih, J., Joo, H.: Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 84-93 (2020) 1, 4, 12", + "90. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022) 4", + "91. Singh, M., Duval, Q., Alwala, K.V., Fan, H., Aggarwal, V., Adcock, A., Joulin, A., Dollar, P., Feichtenhofer, C., Girshick, R., et al.: The effectiveness of mae pretraining for billion-scale pretraining. arXiv preprint arXiv:2303.13496 (2023) 2, 3, 4", + "92. Sun, C., Shrivastava, A., Singh, S., Gupta, A.: Revisiting unreasonable effectiveness of data in deep learning era. In: Proceedings of the IEEE international conference on computer vision. pp. 843-852 (2017) 4", + "93. Sun, K., Xiao, B., Liu, D., Wang, J.: Deep high-resolution representation learning for human pose estimation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5693-5703 (2019) 4, 9", + "94. Taori, R., Gulrajani, I., Zhang, T., Dubois, Y., Li, X., Guestrin, C., Liang, P., Hashimoto, T.B.: Alpaca: A strong, replicable instruction-following model. Stanford Center for Research on Foundation Models. https://crfm.stanford.edu/2023/03/13/alpaca.html 3(6), 7 (2023) 5", + "95. Tay, Y., Dehghani, M., Rao, J., Fedus, W., Abnar, S., Chung, H.W., Narang, S., Yogatama, D., Vaswani, A., Metzler, D.: Scale efficiently: Insights from pretraining and fine-tuning transformers. arXiv preprint arXiv:2109.10686 (2021) 4" + ], + "bbox": [ + 223, + 146, + 785, + 840 + ], + "page_idx": 19 + }, + { + "type": "page_number", + "text": "20", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 19 + }, + { + "type": "header", + "text": "R. Khirodkar et al.", + "bbox": [ + 271, + 114, + 398, + 126 + ], + "page_idx": 19 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "96. Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023) 4", + "97. Thomee, B., Shamma, D.A., Friedland, G., Elizalde, B., Ni, K., Poland, D., Borth, D., Li, L.J.: Yfcc100m: The new data in multimedia research. Communications of the ACM 59(2), 64-73 (2016) 4", + "98. Toshev, A., Szegedy, C.: Deepest: Human pose estimation via deep neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1653-1660 (2014) 9", + "99. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023) 4", + "100. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023) 4, 5, 9", + "101. Wang, X., Fouhey, D., Gupta, A.: Designing deep networks for surface normal estimation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 539-547 (2015) 3, 4", + "102. Weng, C.Y., Curless, B., Srinivasan, P.P., Barron, J.T., Kemelmacher-Shlizerman, I.: Humannerf: Free-viewpoint rendering of moving people from monocular video. In: Proceedings of the IEEE/CVF conference on computer vision and pattern Recognition. pp. 16210-16220 (2022) 1", + "103. Wu, Y., Kirillov, A., Massa, F., Lo, W.Y., Girshick, R.: Detector2. https://github.com/facebookresearch/detectron2 (2019) 5", + "104. Xia, F., Wang, P., Chen, L.C., Yuille, A.L.: Zoom better to see clearer: Human and object parsing with hierarchical auto-zoom net. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14. pp. 648-663. Springer (2016) 4", + "105. Xia, F., Wang, P., Chen, X., Yuille, A.L.: Joint multi-person pose estimation and semantic part segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6769-6778 (2017) 4", + "106. Xiao, B., Wu, H., Wei, Y.: Simple baselines for human pose estimation and tracking. In: Proceedings of the European conference on computer vision (ECCV). pp. 466-481 (2018) 4, 9", + "107. Xie, E., Wang, W., Yu, Z., Anandkumar, A., Alvarez, J.M., Luo, P.: Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in neural information processing systems 34, 12077-12090 (2021) 10", + "108. Xiu, Y., Yang, J., Cao, X., Tzionas, D., Black, M.J.: Econ: Explicit clothed humans optimized via normal integration. arXiv preprint arXiv:2212.07422 (2022) 1, 4, 12", + "109. Xiu, Y., Yang, J., Tzionas, D., Black, M.J.: Icon: Implicit clothed humans obtained from normals. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13286-13296. IEEE (2022) 1, 12", + "110. Xu, L., Jin, S., Liu, W., Qian, C., Ouyang, W., Luo, P., Wang, X.: Zoomnas: searching for whole-body human pose estimation in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(4), 5296-5313 (2022) 9", + "111. Xu, Y., Zhang, J., Zhang, Q., Tao, D.: Vitpose: Simple vision transformer baselines for human pose estimation. Advances in Neural Information Processing Systems 35, 38571-38584 (2022) 6, 9" + ], + "bbox": [ + 217, + 146, + 785, + 839 + ], + "page_idx": 20 + }, + { + "type": "header", + "text": "Sapiens", + "bbox": [ + 678, + 114, + 730, + 127 + ], + "page_idx": 20 + }, + { + "type": "page_number", + "text": "21", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 20 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "112. Xu, Y., Zhang, J., Zhang, Q., Tao, D.: Vitpose+: Vision transformer foundation model for generic body pose estimation. arXiv preprint arXiv:2212.04246 (2022) 9, 13", + "113. Yang, L., Kang, B., Huang, Z., Xu, X., Feng, J., Zhao, H.: Depth anything: Unleashing the power of large-scale unlabeled data. arXiv preprint arXiv:2401.10891 (2024) 1, 3, 4, 11", + "114. Yang, Z., Zeng, A., Yuan, C., Li, Y.: Effective whole-body pose estimation with two-stages distillation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4210-4220 (2023) 9", + "115. Yang, Z., Dai, Z., Yang, Y., Carbonell, J., Salakhutdinov, R.R., Le, Q.V.: Xlnet: Generalized autoregressive pretraining for language understanding. Advances in neural information processing systems 32 (2019) 9", + "116. Yin, Y., Guo, C., Kaufmann, M., Zarate, J.J., Song, J., Hilliges, O.: Hi4d: 4d instance segmentation of close human interaction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17016-17027 (2023) 10, 12", + "117. Yu, T., Zheng, Z., Guo, K., Liu, P., Dai, Q., Liu, Y.: Function4d: Real-time human volumetric capture from very sparse consumer rgbd sensors. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5746-5756 (2021) 10, 12", + "118. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3836-3847 (2023) 1, 14", + "119. Zhang, S.H., Li, R., Dong, X., Rosin, P., Cai, Z., Han, X., Yang, D., Huang, H., Hu, S.M.: Pose2seg: Detection free human instance segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 889-898 (2019) 1, 3", + "120. Zhang, X., Lu, Y., Wang, W., Yan, A., Yan, J., Qin, L., Wang, H., Yan, X., Wang, W.Y., Petzold, L.R.: Gpt-4v (isdiction) as a generalist evaluator for vision-language tasks. arXiv preprint arXiv:2311.01361 (2023) 4", + "121. Zhou, J., Wei, C., Wang, H., Shen, W., Xie, C., Yuille, A., Kong, T.: ibot: Image bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832 (2021) 3, 4", + "122. Zlateski, A., Jaroensri, R., Sharma, P., Durand, F.: On the importance of label quality for semantic segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1479-1487 (2018) 3" + ], + "bbox": [ + 215, + 146, + 785, + 646 + ], + "page_idx": 21 + }, + { + "type": "page_number", + "text": "22", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 21 + }, + { + "type": "header", + "text": "R. Khirodkar et al.", + "bbox": [ + 271, + 114, + 398, + 126 + ], + "page_idx": 21 + } +] \ No newline at end of file diff --git a/2024/Sapiens_ Foundation for Human Vision Models/726c3ced-d623-4195-b91f-26812a59bf2e_model.json b/2024/Sapiens_ Foundation for Human Vision Models/726c3ced-d623-4195-b91f-26812a59bf2e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..696bede5bbcdc1f132e9b09a0038dc9a2622ffe0 --- /dev/null +++ b/2024/Sapiens_ Foundation for Human Vision Models/726c3ced-d623-4195-b91f-26812a59bf2e_model.json @@ -0,0 +1,3467 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.235, + 0.141, + 0.768, + 0.163 + ], + "angle": 0, + "content": "Sapiens: Foundation for Human Vision Models" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.19, + 0.776, + 0.22 + ], + "angle": 0, + "content": "Rawal Khirodkar, Timur Bagautdinov, Julieta Martinez, Su Zhaoen, Austin James, Peter Selednik, Stuart Anderson, Shunsuke Saito" + }, + { + "type": "text", + "bbox": [ + 0.415, + 0.233, + 0.589, + 0.246 + ], + "angle": 0, + "content": "Codec Avatars Lab, Meta" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.259, + 0.74, + 0.508 + ], + "angle": 0, + "content": "Abstract. We present Sapiens, a family of models for four fundamental human-centric vision tasks - 2D pose estimation, body-part segmentation, depth estimation, and surface normal prediction. Our models natively support 1K high-resolution inference and are extremely easy to adapt for individual tasks by simply fine-tuning foundation models pretrained on over 300 million in-the-wild human images. We observe that, given the same computational budget, self-supervised pretraining on a curated dataset of human images significantly boosts the performance for a diverse set of human-centric tasks. The resulting models exhibit remarkable generalization to in-the-wild data, even when labeled data is scarce or entirely synthetic. Our simple model design also brings scalability - model performance across tasks significantly improves as we scale the number of parameters from 0.3 to 2 billion. Sapiens consistently surpasses existing complex baselines across various human-centric benchmarks. Specifically, we achieve significant improvements over the prior state-of-the-art on Humans-5K (pose) by \\(7.6\\mathrm{mAP}\\), Humans-2K (part-seg) by \\(17.1\\mathrm{mIoU}\\), Hi4D (depth) by \\(22.4\\%\\) relative RMSE, and THuman2 (normal) by \\(53.5\\%\\) relative angular error." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.513, + 0.694, + 0.527 + ], + "angle": 0, + "content": "Keywords: Human-centric, Large-scale pretraining, In-the-wild" + }, + { + "type": "text", + "bbox": [ + 0.294, + 0.54, + 0.706, + 0.555 + ], + "angle": 0, + "content": "\"Sapiens—pertaining to, or resembling modern humans.\"" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.575, + 0.377, + 0.59 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.788, + 0.749 + ], + "angle": 0, + "content": "Recent years have witnessed remarkable strides towards generating photorealistic humans in 2D [17, 28, 50, 118] and 3D [69, 89, 102, 109]. The success of these methods is greatly attributed to the robust estimation of various assets such as 2D keypoints [14,67], fine-grained body-part segmentation [119], depth [113], and surface normals [89, 108]. However, robust and accurate estimation of these assets is still an active research area, and complicated systems to boost performance for individual tasks often hinder wider adoption. Moreover, obtaining accurate ground-truth annotation in-the-wild is notoriously difficult to scale. Our goal is to provide a unified framework and models to infer these assets in-the-wild to unlock a wide range of human-centric applications for everybody." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We argue that such human-centric models should satisfy three criteria: generalization, broad applicability, and high fidelity. Generalization ensures robustness to unseen conditions, enabling the model to perform consistently across varied environments. Broad applicability indicates the versatility of the model, making it suitable for a wide range of tasks with minimal modifications. High fidelity denotes the ability of the model to produce precise, high-resolution outputs," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.401, + 0.128 + ], + "angle": 0, + "content": "R. Khirodkar et al." + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.144, + 0.756, + 0.539 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.542, + 0.788, + 0.584 + ], + "angle": 0, + "content": "Fig. 1: Sapiens models are finetuned for four human tasks - 2D pose estimation, body-part segmentation, depth prediction and normal prediction. Our models generalize across a variety of in-the-wild face, upper-body, full-body and multi-person images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.592, + 0.788, + 0.637 + ], + "angle": 0, + "content": "essential for faithful human generation tasks such as 2D to 3D lifting. This paper details the development of models that embody these attributes, collectively referred to as Sapiens." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Following the insights from [34, 79, 91], leveraging large datasets and scalable model architectures is key for generalization. For broader applicability, we adopt the pretrain-then-finetune approach, enabling post-pretraining adaptation to specific tasks with minimal adjustments. This approach raises a critical question: What type of data is most effective for pretraining? Given computational limits, should the emphasis be on collecting as many human images as possible, or is it preferable to pretrain on a less curated set to better reflect real-world variability? Existing methods often overlook the pretraining data distribution in the context of downstream tasks. To study the influence of pretraining data distribution on human-specific tasks, we collect the Humans-300M dataset, featuring 300 million diverse human images. These unlabelled images are used to pretrain a family of vision transformers [27] from scratch, with parameter counts ranging from 300M to 2B." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.679, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Sapiens" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.799, + 0.373 + ], + "angle": 0, + "content": "Among various self-supervision methods for learning general-purpose visual features from large datasets [5,19,34,45,46,121], we choose the masked-autoencoder (MAE) approach [45] for its simplicity and efficiency in pretraining. MAE, having a single-pass inference model compared to contrastive or multi-inference strategies, allows processing a larger volume of images with the same computational resources. For higher-fidelity, in contrast to prior methods, we increase the native input resolution of our pretraining to 1024 pixels, resulting in a \\(\\sim 4\\times\\) increase in FLOPs compared to the largest existing vision backbone [91]. Each model is pretrained on 1.2 trillion tokens. Table 1 outlines a comparison with earlier approaches. For finetuning on human-centric tasks [15, 101, 113, 119], we use a consistent encoder-decoder architecture. The encoder is initialized with weights from pretraining, while the decoder, a lightweight and task-specific head, is initialized randomly. Both components are then finetuned in an end-to-end manner. We focus on four key tasks - 2D pose estimation, body-part segmentation, depth, and normal estimation, as shown in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.374, + 0.79, + 0.617 + ], + "angle": 0, + "content": "Consistently with prior studies [56,122], we affirm the critical impact of label quality on the model's in-the-wild performance. Public benchmarks [23,41,55] often contain noisy labels, providing inconsistent supervisory signals during model fine-tuning. At the same time, it is important to utilize fine-grained and precise annotations to align closely with our primary goal of 3D human digitization. To this end, we propose a substantially denser set of 2D whole body keypoints for pose estimation and a detailed class vocabulary for body part segmentation, surpassing the scope of previous datasets (please refer to Fig. 1). Specifically, we introduce a comprehensive collection of 308 keypoints encompassing the body, hands, feet, surface, and face. Additionally, we expand the segmentation class vocabulary to 28 classes, covering body parts such as the hair, tongue, teeth, upper/lower lip, and torso. To guarantee the quality and consistency of annotations and a high degree of automation, we utilize a multi-view capture setup to collect pose and segmentation annotations. We also utilize human-centric synthetic data for depth and normal estimation, leveraging 600 detailed scans from RenderPeople [82] to generate high-resolution depth maps and surface normals." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.616, + 0.789, + 0.691 + ], + "angle": 0, + "content": "We show that the combination of domain-specific large-scale pretraining with limited, yet high-quality annotations leads to robust in-the-wild generalization. Overall, our method demonstrates an effective strategy for developing highly precise discriminative models capable of performing in real-world scenarios without the need for collecting a costly and diverse set of annotations." + }, + { + "type": "table", + "bbox": [ + 0.285, + 0.695, + 0.723, + 0.796 + ], + "angle": 0, + "content": "
MethodDataset#ParamsGFLOPsImage sizeDomain
DINO [16]ImageNet1k86 M17.6224General
iBOT [121]ImageNet21k307 M61.6224General
DINOv2 [79]LVD-142M1 B291.0224General
ViT-6.5B [91]IG-3B6.5 B1657.0224General
AIM [34]DFN-2B6.5 B1657.0224General
Sapiens (Ours)Humans-300M2 B8709.01024Human
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.801, + 0.789, + 0.83 + ], + "angle": 0, + "content": "Table 1: Comparison of state-of-the-art pretrained vision models. Sapiens adopts a higher resolution backbone on a large dataset of in-the-wild human images." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.401, + 0.128 + ], + "angle": 0, + "content": "R. Khirodkar et al." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.147, + 0.569, + 0.161 + ], + "angle": 0, + "content": "Our contributions are summarized as follows." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.167, + 0.784, + 0.196 + ], + "angle": 0, + "content": "- We introduce Sapiens, a family of vision transformers pretrained on a large-scale, curated dataset of diverse human images." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.198, + 0.785, + 0.227 + ], + "angle": 0, + "content": "- This study shows that simple data curation and large-scale pretraining significantly boost the model's performance with the same computational budget." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.228, + 0.784, + 0.257 + ], + "angle": 0, + "content": "- Our models, fine-tuned with high-quality annotations or even synthetic data, demonstrate robust in-the-wild generalization." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.258, + 0.791, + 0.317 + ], + "angle": 0, + "content": "- The first 1K high-resolution model that natively supports high-fidelity inference essential for human-centric tasks, achieving state-of-the-art performance on benchmarks for 2D pose estimation, body-part segmentation, depth, and normal estimation." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.167, + 0.791, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.341, + 0.388, + 0.357 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.373, + 0.785, + 0.419 + ], + "angle": 0, + "content": "Our work explores the limits of training large architectures on a large number of in-the-wild human images. We build on prior work from different areas: pretraining at scale, human vision tasks, and large vision transformers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.425, + 0.788, + 0.683 + ], + "angle": 0, + "content": "Pretraining at Scale. The remarkable success of large-scale pretraining [26,95] followed by task-specific finetuning for language modeling [2,13,53,96,99,100] has established this approach as a standard practice. Similarly, computer vision methods [1,4,33,34,42,79,83,85,87,120] are progressively embracing extensive data scales for pretraining. The emergence of large datasets, such as LAION-5B [90], Instagram-3.5B [77], JFT-300M [92], LVD-142M [79], Visual Genome [60], and YFCC100M [97], has enabled the exploration of a data corpus well beyond the scope of traditional benchmarks [61,67,86]. Salient work in this domain includes DINOv2 [79], MAWS [91], and AIM [34]. DINOv2 achieves state-of-the-art performance in generating self-supervised features by scaling the contrastive iBot [121] method on the LDV-142M dataset [79]. MAWS [91] studies the scaling of masked-autoencoders (MAE) [45] on billion images. AIM [34] explores the scalability of autoregressive visual pretraining similar to BERT [26] for vision transformers [27]. In contrast to these methods which mainly focus on general image pretraining or zero-shot image classification, we take a distinctly human-centric approach: our models leverage a vast collection of human images for pretraining, subsequently fine-tuning for a range of human-related tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Human Vision Tasks. The pursuit of large-scale 3D human digitization [8, 44, 64, 74] remains a pivotal goal in computer vision [12]. Significant progress has been made within controlled or studio environments [3, 59, 63, 69, 70, 76, 89], yet challenges persist in extending these methods to unconstrained environments [29]. To address these challenges, developing versatile models capable of multiple fundamental tasks such as keypoint estimation [21, 36, 47, 51, 57, 78, 80, 93, 106], body-part segmentation [35, 40, 40, 41, 75, 104, 105], depth estimation [9, 10, 32, 43, 52, 66, 84, 113], and surface normal prediction [6, 7, 31, 39, 62, 88, 101, 108] from images in natural settings is crucial. In this work, we aim to develop models for these essential human vision tasks which generalize to in-the-wild settings." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.679, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Sapiens" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.358 + ], + "angle": 0, + "content": "Scaling Architectures. Currently, the largest publicly-accessible language models contain upwards of 100B parameters [49], while the more commonly used language models [94, 100] contain around 7B parameters. In contrast, Vision Transformers (ViT) [27], despite sharing a similar architecture, have not been scaled to this extent successfully. While there are notable endeavors in this direction, including the development of a dense ViT-4B [20] trained on both text and images, and the formulation of techniques for the stable training of a ViT-22B [25], commonly utilized vision backbones still range between 300M to 600M parameters [24, 38, 48, 68] and are primarily pretrained at an image resolution of about 224 pixels. Similarly, existing transformer-based image generation models, such as DiT [81] use less than 700M parameters, and operate on a highly compressed latent space. To address this gap, we introduce Sapiens - a collection of large, high-resolution ViT models that are pretrained natively at a 1024 pixel image resolution on millions of human images." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.367, + 0.331, + 0.383 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.388, + 0.453, + 0.403 + ], + "angle": 0, + "content": "3.1 Humans-300M Dataset" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.412, + 0.607, + 0.579 + ], + "angle": 0, + "content": "We utilize a large proprietary dataset for pretraining of approximately 1 billion in-the-wild images, focusing exclusively on human images. The preprocessing involves discarding images with watermarks, text, artistic depictions, or unnatural elements. Subsequently, we use an off-the-shelf person bounding-box detector [103] to filter images, retaining those with a detection score above 0.9 and bounding box dimensions exceeding 300 pixels. Fig. 2 provides an overview of the distribution of the number of people per image in over 248 million images contain multiple subjects." + }, + { + "type": "image", + "bbox": [ + 0.615, + 0.352, + 0.787, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.611, + 0.504, + 0.788, + 0.536 + ], + "angle": 0, + "content": "Fig.2: Overview of number of humans per image in the Humans-300M dataset." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.59, + 0.359, + 0.606 + ], + "angle": 0, + "content": "3.2 Pretraining" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.789, + 0.841 + ], + "angle": 0, + "content": "We follow the masked-autoencoder [45] (MAE) approach for pretraining. Our model is trained to reconstruct the original human image given its partial observation. Like all autoencoders, our model has an encoder that maps the visible image to a latent representation and a decoder that reconstructs the original image from this latent representation. Our pretraining dataset consists of both single and multi-human images; each image is resized to a fixed size with a square aspect ratio. Similar to ViT [27], we divide an image into regular non-overlapping patches with a fixed patch size. A subset of these patches is randomly selected and masked, leaving the rest visible. The proportion of masked patches to visible ones is defined as the masking ratio, which remains fixed throughout training. We refer to MAE [45] for more details. Fig. 3 (Top) shows the reconstruction of our pretrained model on unseen human images. Our models exhibit generalization across a variety of image characteristics including scales, crops, the age and ethnicity of subjects, and number of subjects. Each patch token in our model accounts for \\(0.02\\%\\) of the image area compared to \\(0.4\\%\\) in standard ViTs, a \\(16\\times\\)" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.4, + 0.128 + ], + "angle": 0, + "content": "R. Khirodkar et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.787, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.425, + 0.785, + 0.471 + ], + "angle": 0, + "content": "reduction - this provides a fine-grained inter-token reasoning for our models. Fig.3 (Bottom) shows that even with an increased mask ratio of \\(95\\%\\), our model achieves a plausible reconstruction of human anatomy on held-out samples." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.483, + 0.427, + 0.497 + ], + "angle": 0, + "content": "3.3 2D Pose Estimation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.719 + ], + "angle": 0, + "content": "We follow the top-down 2D pose estimation paradigm, which aims to detect the locations of \\( K \\) keypoints from an input image \\( \\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3} \\). Most methods pose this problem as heatmap prediction, where each of \\( K \\) heatmaps represents the probability of the corresponding keypoint being at any spatial location. Similar to [111], we define a pose estimation transformer, \\( \\mathcal{P} \\), for keypoint detection. The bounding box at training and inference is scaled to \\( H \\times W \\) and is provided as an input to \\( \\mathcal{P} \\). Let \\( \\mathbf{y} \\in \\mathbb{R}^{H \\times W \\times K} \\) denote the \\( K \\) heatmaps corresponding to the ground truth keypoints for a given input \\( \\mathbf{I} \\). The pose estimator transforms input \\( \\mathbf{I} \\) to a set of predicted heatmaps, \\( \\hat{\\mathbf{y}} \\in \\mathbb{R}^{H \\times W \\times K} \\), such that \\( \\hat{\\mathbf{y}} = \\mathcal{P}(\\mathbf{I}) \\). \\( \\mathcal{P} \\) is trained to minimize the mean squared loss \\( \\mathcal{L}_{\\mathrm{pose}} = \\mathrm{MSE}(\\mathbf{y},\\hat{\\mathbf{y}}) \\). During finetuning, the encoder of \\( \\mathcal{P} \\) is initialized with the weights from pretraining, and the decoder is initialized randomly. The aspect ratio \\( H:W \\) is set to be 4:3, with the pretrained positional embedding being interpolated accordingly [58]. We use lightweight decoders with deconvolution and convolution operations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We finetune both the encoder and the decoder in \\(\\mathcal{P}\\) across multiple skeleton formats, including \\(K = 17\\) [67], \\(K = 133\\) [55] and a new highly-detailed skeleton format, with \\(K = 308\\), as shown in Fig. 4 (Left). Compared to existing formats with at most 68 facial keypoints, our annotations consist of 243 facial keypoints, including representative points around the eyes, eyebrows, lips, nose, and ears. This design is tailored to meticulously capture the nuanced details of facial expressions in the real world. With these keypoints, we manually annotated 1 million images at \\(4K\\) resolution from an indoor capture setup." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.679, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Sapiens" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.146, + 0.387, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.255, + 0.27, + 0.349, + 0.28 + ], + "angle": 0, + "content": "a) Full-Body - 308 kps" + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.147, + 0.466, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.389, + 0.193, + 0.464, + 0.202 + ], + "angle": 0, + "content": "b) Hands - 40 kps" + }, + { + "type": "image", + "bbox": [ + 0.391, + 0.206, + 0.461, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.387, + 0.27, + 0.459, + 0.28 + ], + "angle": 0, + "content": "c) Face - 243 kps" + }, + { + "type": "image", + "bbox": [ + 0.479, + 0.147, + 0.566, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.568, + 0.148, + 0.651, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.149, + 0.785, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.57, + 0.27, + 0.731, + 0.28 + ], + "angle": 0, + "content": "d) Body-Part Segmentation: 28 Classes" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.291, + 0.784, + 0.306 + ], + "angle": 0, + "content": "Fig. 4: Ground-truth annotations for 2D pose estimation and body-part segmentation." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.317, + 0.467, + 0.332 + ], + "angle": 0, + "content": "3.4 Body-Part Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.341, + 0.785, + 0.447 + ], + "angle": 0, + "content": "Commonly referred to as human parsing, body-part segmentation aims to classify pixels in the input image \\(\\mathbf{I}\\) into \\(C\\) classes. Most methods [41] transform this problem to estimating per-pixel class probabilities to create a probability map \\(\\hat{\\mathbf{p}}\\in \\mathbb{R}^{H\\times W\\times C}\\) such that \\(\\hat{\\mathbf{p}} = S(\\mathbf{I})\\), where \\(S\\) is the segmentation model. As outlined previously, we adopt the same encoder-decoder architecture and initialization scheme for \\(S\\). \\(S\\) is finetuned to minimize the weighted cross-entropy loss between the actual \\(\\mathbf{p}\\) and predicted \\(\\hat{\\mathbf{p}}\\) probability maps, \\(\\mathcal{L}_{\\mathrm{seg}} = \\mathrm{WeightedCE}(\\mathbf{p},\\hat{\\mathbf{p}})\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.785, + 0.553 + ], + "angle": 0, + "content": "We finetune \\(\\mathcal{S}\\) across two part-segmentation vocabularies: a standard set with \\(C = 20\\) [41] and a new larger vocabulary with \\(C = 28\\), as illustrated in Fig.4 (Right). Our proposed vocabulary goes beyond previous datasets in important ways. It distinguishes between the upper and lower halves of limbs and incorporates more detailed classifications such as upper/lower lips, teeth, and tongue. To this end, we manually annotate \\(100K\\) images at \\(4K\\) resolution with this vocabulary." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.575, + 0.41, + 0.59 + ], + "angle": 0, + "content": "3.5 Depth Estimation" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.6, + 0.784, + 0.691 + ], + "angle": 0, + "content": "For depth estimation, we adopt the architecture used for segmentation, with the modification that the decoder output channel is set to 1 for regression. We denote the ground-truth depth map of image \\(\\mathbf{I}\\) by \\(\\mathbf{d} \\in \\mathbb{R}^{H \\times W}\\), the depth estimator by \\(\\mathcal{D}\\), where \\(\\hat{\\mathbf{d}} = \\mathcal{D}(\\mathbf{I})\\), and \\(M\\) as the number of human pixels in the image. For the relative depth estimation, we normalize \\(\\mathbf{d}\\) to the range [0, 1] using max and min depths in the image. The \\(\\mathcal{L}_{\\mathrm{depth}}\\) loss [32] for \\(\\mathcal{D}\\) is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.3, + 0.702, + 0.785, + 0.72 + ], + "angle": 0, + "content": "\\[\n\\Delta \\mathbf {d} = \\log (\\mathbf {d}) - \\log (\\hat {\\mathbf {d}}), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.3, + 0.724, + 0.785, + 0.765 + ], + "angle": 0, + "content": "\\[\n\\overline {{\\Delta \\mathbf {d}}} = \\frac {1}{M} \\sum_ {i = 1} ^ {M} \\Delta \\mathbf {d} _ {i}, \\quad \\overline {{(\\Delta \\mathbf {d}) ^ {2}}} = \\frac {1}{M} \\sum_ {i = 1} ^ {M} (\\Delta \\mathbf {d} _ {i}) ^ {2}, \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.28, + 0.768, + 0.785, + 0.801 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {d e p t h}} = \\sqrt {\\left(\\overline {{\\Delta \\mathbf {d}}} ^ {2}\\right) - \\frac {1}{2} \\left(\\overline {{\\Delta \\mathbf {d}}} ^ {2}\\right) ^ {2}}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "We render 500,000 synthetic images using 600 high-resolution photogrammetry human scans as shown in Fig. 5 to obtain a robust monocular depth" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.401, + 0.128 + ], + "angle": 0, + "content": "R. Khirodkar et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.208 + ], + "angle": 0, + "content": "estimation model with high-fidelity. A random background is selected from a 100 HDRI environment map collection. We place a virtual camera within the scene, randomly adjusting its focal length, rotation, and translation to capture images and their associated ground-truth depth maps at 4K resolution." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.215, + 0.787, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.217, + 0.419, + 0.784, + 0.435 + ], + "angle": 0, + "content": "Fig. 5: Ground-truth synthetic annotations for depth and surface normal estimation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.443, + 0.488, + 0.458 + ], + "angle": 0, + "content": "3.6 Surface Normal Estimation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.47, + 0.788, + 0.561 + ], + "angle": 0, + "content": "Similar to previous tasks, we set the decoder output channels of the normal estimator \\(\\mathcal{N}\\) to be 3, corresponding to the \\(xyz\\) components of the normal vector at each pixel. The generated synthetic data is also used as supervision for surface normal estimation. Let \\(\\mathbf{n}\\) be the ground-truth normal map for image \\(\\mathbf{I}\\) and \\(\\hat{\\mathbf{n}} = \\mathcal{N}(\\mathbf{I})\\). Similar to depth, the loss \\(\\mathcal{L}_{\\mathrm{normal}}\\) is only computed for human pixels in the image and is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.381, + 0.573, + 0.787, + 0.591 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {n o r m a l}} = \\left\\| \\mathbf {n} - \\hat {\\mathbf {n}} \\right\\| _ {1} + (1 - \\mathbf {n} \\cdot \\hat {\\mathbf {n}}) \\tag {4}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.6, + 0.377, + 0.617 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.633, + 0.787, + 0.679 + ], + "angle": 0, + "content": "In this section, we initially provide an overview of the implementation details. Subsequently, we conduct comprehensive benchmarking across four tasks: pose estimation, part segmentation, depth estimation, and normal estimation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.693, + 0.458, + 0.709 + ], + "angle": 0, + "content": "4.1 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Our largest model, Sapiens-2B, is pretrained using 1024 A100 GPUs for 18 days using PyTorch. We use the AdamW [73] optimizer for all our experiments. The learning schedule includes a brief linear warm-up, followed by cosine annealing [72] for pretraining and linear decay [65] for finetuning. All models are pretrained from scratch at a resolution of \\(1024 \\times 1024\\) with a patch size of 16. For finetuning, the input image is resized to a 4:3 ratio, i.e. \\(1024 \\times 768\\). We use standard augmentations like cropping, scaling, flipping, and photometric distortions. A random background from non-human COCO [67] images is added for" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.679, + 0.116, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Sapiens" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table", + "bbox": [ + 0.277, + 0.144, + 0.734, + 0.216 + ], + "angle": 0, + "content": "
Model#ParamsFLOPsHidden sizeLayersHeadsBatch size
Sapiens-0.3B0.336 B1.242 T1024241698,304
Sapiens-0.6B0.664 B2.583 T1280321665,536
Sapiens-1B1.169 B4.647 T1536402440,960
Sapiens-2B2.163 B8.709 T1920483220,480
" + }, + { + "type": "table_caption", + "bbox": [ + 0.23, + 0.22, + 0.77, + 0.234 + ], + "angle": 0, + "content": "Table 2: Sapiens encoder specifications for pretraining on Human-300M dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.241, + 0.785, + 0.361 + ], + "angle": 0, + "content": "segmentation, depth, and normal prediction tasks. Importantly, we use differential learning rates [115] to preserve generalization i.e. lower learning rates for initial layers and progressively higher rates for subsequent layers. The layer-wise learning rate decay is set to 0.85 with a weight decay of 0.1 for the encoder. We detail the design specifications of Sapiens in Table. 2. Following [34, 100], we prioritize scaling models by width rather than depth. Note that the Sapiens-0.3B model, while architecturally similar to the traditional ViT-Large, consists of twentyfold more FLOPs due to its higher resolution." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.367, + 0.427, + 0.38 + ], + "angle": 0, + "content": "4.2 2D Pose Estimation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.384, + 0.785, + 0.565 + ], + "angle": 0, + "content": "We finetune Sapiens for face, body, feet, and hand (\\(K = 308\\)) pose estimation on our high-fidelity annotations. For training, we use the train set with \\(1M\\) images and for evaluation, we use the test set, named Humans-5K, with \\(5K\\) images. Our evaluation is top-down [111] i.e. we use an off-the-shelf detector [37] for bounding-box and conduct single human pose inference. Table 3 shows a comparison of our models with existing methods for whole-body pose estimation. We evaluate all methods on 114 common keypoints between our 308 keypoint vocabulary and the 133 keypoint vocabulary from COCO-WholeBody [55]. Sapiens-0.6B surpasses the current state-of-the-art, DWPose-1 [114] by +2.8 AP. Contrary to DWPose [114], which utilizes a complex student-teacher framework with feature distillation tailored for the task, Sapiens adopts a general encoder-decoder architecture with large human-centric pretraining." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.566, + 0.787, + 0.657 + ], + "angle": 0, + "content": "Interestingly, even with the same parameter count, our models demonstrate superior performance compared to their counterparts. For instance, Sapiens-0.3B exceeds VitPose+-L by +5.6 AP, and Sapiens-0.6B outperforms VitPose+-H by +7.9 AP. Within the Sapiens family, our results indicate a direct correlation between model size and performance. Sapiens-2B sets a state-of-the-art with 61.1 AP, a significant improvement of +7.6 AP to the prior art. Despite fine-tuning" + }, + { + "type": "table", + "bbox": [ + 0.227, + 0.665, + 0.794, + 0.832 + ], + "angle": 0, + "content": "
ModelInput SizeBodyFootFaceHandWhole-body
APARAPARAPARAPARAPAR
DeepPose [98]384 × 28832.143.525.341.237.853.915.731.623.937.2
SimpleBaseline [106]384 × 28852.360.149.862.559.667.341.451.844.653.7
HRNet [93]384 × 28855.862.645.255.458.964.539.347.645.753.9
ZoomNAS [110]384 × 28859.766.348.157.974.579.249.860.652.160.7
ViTPose+L [112]256 × 19261.066.862.468.250.155.741.547.347.853.6
ViTPose+-H [112]256 × 19261.667.463.269.050.756.342.047.848.354.1
RTMPose-x [54]384 × 28857.163.755.366.874.478.546.355.051.959.6
DWPose-m [114]256 × 19254.261.449.963.068.574.240.150.047.755.8
DWPose-l [114]384 × 28857.964.256.567.474.378.449.357.453.160.6
Sapiens-0.3B (Ours)1024 × 76858.164.556.867.774.578.649.657.753.4 (+0.3)60.9 (+0.3)
Sapiens-0.6B (Ours)1024 × 76859.865.564.772.375.279.052.160.356.2 (+2.8)62.4 (+2.1)
Sapiens-1B (Ours)1024 × 76862.968.268.375.176.479.755.963.459.4 (+5.9)65.3 (+5.1)
Sapiens-2B (Ours)1024 × 76864.769.969.476.276.979.957.164.461.1(+7.6)67.1(+7.0)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.248, + 0.837, + 0.753, + 0.851 + ], + "angle": 0, + "content": "Table 3: Pose estimation results on Humans-5K test set. Flip test is used." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.401, + 0.128 + ], + "angle": 0, + "content": "R. Khirodkar et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.393, + 0.146, + 0.479, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.48, + 0.146, + 0.612, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.146, + 0.785, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.23, + 0.243, + 0.771, + 0.258 + ], + "angle": 0, + "content": "Fig. 6: Pose estimation with Sapiens-1B for 308 keypoints on in-the-wild images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.269, + 0.784, + 0.3 + ], + "angle": 0, + "content": "with annotations from a indoor capture studio, Sapiens demonstrate robust generalization to real-world, as shown in Fig. 6." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.323, + 0.467, + 0.339 + ], + "angle": 0, + "content": "4.3 Body-Part Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.35, + 0.544, + 0.499 + ], + "angle": 0, + "content": "We fine-tune and evaluate our annotations with a segmentation vocabulary of 28 classes. Our train set consists of \\(100K\\) images, and the test set, Humans-2K, consists of \\(2K\\) images. We compare Sapiens with existing bodypart segmentation methods fine-tuned on our train set. Importantly, we use suggested pretrained checkpoints by each method as initialization. Similar to pose, we observe generalization to segmentation as shown in Table 4." + }, + { + "type": "table", + "bbox": [ + 0.572, + 0.325, + 0.778, + 0.438 + ], + "angle": 0, + "content": "
ModelmIoU(%)mAcc(%)
FCN* [71]48.257.6
SegFormer* [107]53.562.9
Mask2Former* [22]58.768.3
DeepLabV3+* [18]64.174.8
Sapiens-0.3B (Ours)76.786.1
Sapiens-0.6B (Ours)77.886.3
Sapiens-1B (Ours)79.989.1
Sapiens-2B (Ours)81.289.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.554, + 0.444, + 0.787, + 0.486 + ], + "angle": 0, + "content": "Table 4: We report mIoU and mAcc on Humans-2K test set. Methods with * are trained by us." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.501, + 0.788, + 0.592 + ], + "angle": 0, + "content": "Interestingly, our smallest model, Sapiens-0.3B outperforms existing state-of-the-art segmentation methods like Mask2Former [22] and DeepLabV3+ [18] by 12.6 mIoU due to its higher resolution and large human-centric pretraining. Furthermore, increasing the model size improves segmentation performance. Sapiens-2B achieves the best performance of 81.2 mIoU and 89.4 mAcc on the test set. Fig. 7 shows the qualitative results of our models." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.606, + 0.411, + 0.621 + ], + "angle": 0, + "content": "4.4 Depth Estimation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.631, + 0.788, + 0.738 + ], + "angle": 0, + "content": "We evaluate our models on THuman2.0 [117] and Hi4D [116] datasets for depth estimation. THuman2.0 consists of 526 high-quality human scans, from which we derive three sets of images for testing: a) face, b) upper body, and c) full body using a virtual camera. THuman2.0 with 1578 images thus enables the evaluation of our models' performance on single-human images across multiple scales. Conversely, the Hi4D dataset focuses on multi-human scenarios, with each sequence showcasing two subjects engaged in activities involving human-human" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.745, + 0.447, + 0.824 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.447, + 0.745, + 0.566, + 0.824 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.566, + 0.745, + 0.619, + 0.824 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.746, + 0.785, + 0.824 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.831, + 0.785, + 0.846 + ], + "angle": 0, + "content": "Fig. 7: Human segmentation with Sapiens-1B for 28 categories on in-the-wild images." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.679, + 0.116, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Sapiens" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.135, + 0.807, + 0.257 + ], + "angle": 0, + "content": "
MethodTH2.0-FaceTH2.0-UprBodyTH2.0-FullBodyHi4D
RMSE ↓AbsRel ↓δ1 ↑RMSEAbsRelδ1RMSEAbsRelδ1RMSEAbsRelδ1
MiDaS-L [11]0.1140.0970.9250.3980.2710.8680.7010.6890.7820.2610.0820.975
MiDaS-Swin2 [11]0.0500.0360.9950.1220.0810.9480.2920.1710.8620.2090.0630.997
DepthAny-B [113]0.0390.0260.9990.0480.0280.9990.0610.0300.9990.1430.0340.997
DepthAny-L [113]0.0390.0270.9990.0480.0270.9990.0600.0300.9990.1470.0350.997
Sapiens-0.3B (Ours)0.0120.0081.0000.0150.0091.0000.0210.0101.0000.1480.0461.000
Sapiens-0.6B (Ours)0.0110.0081.0000.0150.0091.0000.0210.0101.0000.1420.0441.000
Sapiens-1B (Ours)0.0090.0061.0000.0120.0071.0000.0190.0091.0000.1250.0391.000
Sapiens-2B (Ours)0.0080.0051.0000.0100.0061.0000.0160.0081.0000.1140.0361.000
" + }, + { + "type": "table_caption", + "bbox": [ + 0.22, + 0.258, + 0.781, + 0.273 + ], + "angle": 0, + "content": "Table 5: Comparison of Sapiens for monocular depth estimation on human images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.279, + 0.785, + 0.369 + ], + "angle": 0, + "content": "interactions. We select sequences from pair 28, pair 32, and pair 37, featuring 6 unique subjects from camera 4, totaling 1195 multi-human real images for testing. We follow the relative-depth evaluation protocols established by MiDaSv3.1 [11], reporting standard metrics such as AbsRel and \\(\\delta_{1}\\). In addition, we also report RMSE as our primary metric since \\(\\delta_{1}\\) does not effectively reflect performance in human scenes characterized by subtle depth variations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.37, + 0.788, + 0.506 + ], + "angle": 0, + "content": "Table 5 compares our models with existing state-of-the-art monocular depth estimators. Sapiens-2B, finetuned solely on synthetic data, remarkably outperforms prior art across all single-human scales and multi-human scenarios. We observe a \\(20\\%\\) RMSE reduction compared to the top-performing Depth-Anything model on Hi4D images. It is important to highlight that while baseline models are trained on a variety of scenes, Sapiens specializes in human-centric depth estimation. Fig. 8 presents a qualitative comparison of depth estimation between Sapiens-1B and DepthAnything-L. To ensure a fair comparison, the predicted depth is renormalized using the human mask in the baseline visualizations." + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.516, + 0.789, + 0.793 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.798, + 0.788, + 0.827 + ], + "angle": 0, + "content": "Fig. 8: We compare our depth prediction with DepthAnything [113]. To showcase the consistency of predicted depth, we also visualize the \\(\\nabla\\) depth as pseudo surface normals." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.4, + 0.128 + ], + "angle": 0, + "content": "R. Khirodkar et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.486, + 0.16 + ], + "angle": 0, + "content": "4.5 Surface Normal Estimation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.164, + 0.788, + 0.314 + ], + "angle": 0, + "content": "The datasets for surface normal evaluation are identical to those used for depth estimation. Following [30], we report the mean and median angular error, along with the percentage of pixels within \\( t^\\circ \\) error for \\( t \\in \\{11.25^\\circ, 22.5^\\circ, 30^\\circ\\} \\). Table 6 compares our models with existing human-specific surface normal estimators. All our models outperform existing methods by a significant margin. Sapiens-2B achieves a mean error of around \\( 12^\\circ \\) on the THuman2.0 (single-human) and Hi4D (multi-human) datasets. We qualitatively compare Sapiens-1B with PIFuHD [89] and ECON [108] for surface normal estimation in Figure 9. Note that PIFuHD [89] is trained with the identical set of 3D scans as ours, and ECON [108] is trained with 4000 scans that are a super set of our 3D scan data." + }, + { + "type": "table", + "bbox": [ + 0.233, + 0.322, + 0.773, + 0.464 + ], + "angle": 0, + "content": "
MethodTHuman2.0 [117]Hi4D [116]
Angular Error°% Within t°Angular Error°% Within t°
MeanMedian11.25°22.5°30°MeanMedian11.25°22.5°30°
PIFuHD [89]30.5127.1315.8142.9758.8622.3919.2622.9860.1477.02
HDNet [52]34.8230.6017.4439.2654.5128.6026.8519.0857.9370.14
ICON [109]28.7425.5222.8147.8363.7320.1817.5226.8166.3482.73
ECON [108]25.4523.6732.9555.8669.0318.4616.4729.3568.1284.88
Sapiens-0.3B13.0210.3357.3786.2092.715.0412.2247.0781.4990.70
Sapiens-0.6B12.8610.2357.8586.6893.3014.0611.4750.5984.3792.54
Sapiens-1B12.119.4061.9788.0393.8412.189.5960.3688.6294.44
Sapiens-2B11.849.1663.1688.6094.1812.149.6260.2289.0894.74
" + }, + { + "type": "table_caption", + "bbox": [ + 0.231, + 0.467, + 0.773, + 0.482 + ], + "angle": 0, + "content": "Table 6: Comparison of Sapiens for surface normal estimation on human images." + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.485, + 0.744, + 0.834 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.835, + 0.788, + 0.864 + ], + "angle": 0, + "content": "Fig.9: Qualitative comparison of Sapiens-1B with PIFuHD [89] and ECON [108] for monocular surface normal estimation on challenging in-the-wild images." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.679, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Sapiens" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.147, + 0.35, + 0.16 + ], + "angle": 0, + "content": "4.6 Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.174, + 0.788, + 0.416 + ], + "angle": 0, + "content": "Importance of Pretraining Data Source. The feature quality is closely linked to the pretraining data quality. We assess the importance of pretraining on various data sources for human-centric tasks by pretraining Sapiens-0.3B on each dataset under identical training schedules and number of iterations. We fine-tune the model on each task and select early checkpoints for evaluation, reasoning that early-stage fine-tuning better reflects the model's generalization capability. We investigate the impact of pretraining at scale on general images (which may include humans) versus exclusively human images using Sapiens. We randomly select 100 million and 300 million general images from our 1 billion image corpus to create the General-100M and General-300M datasets, respectively. Table 7 showcases the comparison of pretraining outcomes. We report mAP for pose on Humans-5K, mIoU for segmentation on Humans-2K, RMSE for depth on THuman2.0, and mean angular error in degrees for normal estimation on Hi4D. Aligned with findings from [112], our results show that pretraining with Human300M leads to superior performance across all metrics, highlighting the benefits of human-centric pretraining within a fixed computational budget." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.416, + 0.518, + 0.599 + ], + "angle": 0, + "content": "We also study the effect of number of unique human images seen during pretraining with normal estimation performance. We report \\(\\%\\) within \\(30^{\\circ}\\). Again, we maintain identical conditions for Sapiens-0.3B pretraining and finetuning. Fig.10 shows a steady improvement in performance as the pretraining data size increases without saturation. In summary, the diversity of human images during pretraining directly correlates with improved generalization to down-stream tasks." + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.428, + 0.789, + 0.599 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.605, + 0.518, + 0.665 + ], + "angle": 0, + "content": "Zero-Shot Generalization. Our models exhibit broad generalization to a variety of settings. For instance, in segmentation, Sapiens are finetuned on single-human im" + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.608, + 0.788, + 0.651 + ], + "angle": 0, + "content": "Fig. 10: Sapiens-0.3B's normal estimation performance with unique human images seen during pretraining." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.666, + 0.788, + 0.727 + ], + "angle": 0, + "content": "ages with limited subject diversity, minimal background variation, and solely third-person views (see Fig. 4). Nevertheless, our large-scale pretraining enables generalization across number of subjects, varying ages, and egocentric views, as shown in Fig. 11. These observations similarly hold for other tasks." + }, + { + "type": "table", + "bbox": [ + 0.277, + 0.735, + 0.737, + 0.821 + ], + "angle": 0, + "content": "
Pretraining Source#ImagesPose (↑)Seg(↑)Depth(↓)Normal(↓)
Random Initialization-30.240.30.72035.4
General-100M100M35.750.10.35127.5
General-300M300M37.352.80.34726.8
Humans-100M100M43.661.20.31624.0
Humans-300M (Full)300M47.066.50.28821.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.823, + 0.787, + 0.852 + ], + "angle": 0, + "content": "Table 7: Comparison of Sapiens-0.3B pretrained on various data sources. A domain-specific pretraining yields superior results compared to general data sources." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.4, + 0.128 + ], + "angle": 0, + "content": "R. Khirodkar et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.244, + 0.236, + 0.379, + 0.246 + ], + "angle": 0, + "content": "Multi-Human Generalization" + }, + { + "type": "image", + "bbox": [ + 0.403, + 0.145, + 0.578, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.446, + 0.236, + 0.537, + 0.247 + ], + "angle": 0, + "content": "Age Generalization" + }, + { + "type": "image", + "bbox": [ + 0.58, + 0.145, + 0.785, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.623, + 0.236, + 0.743, + 0.247 + ], + "angle": 0, + "content": "Viewpoint Generalization" + }, + { + "type": "image_caption", + "bbox": [ + 0.225, + 0.254, + 0.775, + 0.269 + ], + "angle": 0, + "content": "Fig. 11: Sapiens achieve broad generalization via large human-centric pretraining." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.277, + 0.785, + 0.382 + ], + "angle": 0, + "content": "Applications. Accurately predicting pose, body-part segmentation, depth and surface normals for in-the-wild human images enables a wide range of applications, such as 3D human scans and the generation of controllable images. We adapt ControlNet [118] to integrate keypoints for faces and hands predicted by Sapiens, and train this modified model on a subset of the Humans-300M dataset, as shown in Fig. 12. This grants more precise control over image generation, producing anatomically consistent human images." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.397, + 0.331, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.253, + 0.486, + 0.288, + 0.497 + ], + "angle": 0, + "content": "Image" + }, + { + "type": "image", + "bbox": [ + 0.331, + 0.398, + 0.442, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.345, + 0.486, + 0.425, + 0.497 + ], + "angle": 0, + "content": "Predicted Pose" + }, + { + "type": "image", + "bbox": [ + 0.449, + 0.397, + 0.561, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.486, + 0.727, + 0.497 + ], + "angle": 0, + "content": "Pose-conditioned Human Image Synthesis" + }, + { + "type": "image", + "bbox": [ + 0.562, + 0.398, + 0.674, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.673, + 0.398, + 0.786, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.504, + 0.782, + 0.519 + ], + "angle": 0, + "content": "Fig. 12: Image synthesis using ControlNet [118] trained on Sapiens pose predictions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.535, + 0.785, + 0.626 + ], + "angle": 0, + "content": "Limitations. While our models generally perform well, they are not perfect. Human images with complex/rare poses, crowding, and severe occlusion are challenging (see supplemental for details). Although aggressive data augmentation and a detect-and-crop strategy could mitigate these issues, we envision our models as a tool for acquiring large-scale, real-world supervision with human-in-the-loop to develop the next generations of human vision models." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.646, + 0.36, + 0.661 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Sapiens represents a significant step toward elevating human-centric vision models into the realm of foundation models. Our models demonstrate strong generalization capabilities on a variety of human-centric tasks. We attribute the state-of-the-art performance of our models to: (i) large-scale pretraining on a large curated dataset, which is specifically tailored to understanding humans, (ii) scaled high-resolution and high-capacity vision transformer backbones, and (iii) high-quality annotations on augmented studio and synthetic data. We believe that these models can become a key building block for a multitude of downstream tasks, and provide access to high-quality vision backbones to a significantly wider part of the community. A potential direction for future work would be extending Sapiens to 3D and multi-modal datasets." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.679, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Sapiens" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.324, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.173, + 0.785, + 0.2 + ], + "angle": 0, + "content": "1. Abnar, S., Dehghani, M., Neyshabur, B., Sedghi, H.: Exploring the limits of large scale pre-training. arXiv preprint arXiv:2110.02095 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.201, + 0.785, + 0.241 + ], + "angle": 0, + "content": "2. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.242, + 0.785, + 0.282 + ], + "angle": 0, + "content": "3. Alldieck, T., Zanfir, M., Sminchisescu, C.: Photorealistic monocular 3d reconstruction of humans wearing clothing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1506-1515 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.283, + 0.785, + 0.323 + ], + "angle": 0, + "content": "4. Bai, Y., Geng, X., Mangalam, K., Bar, A., Yuille, A., Darrell, T., Malik, J., Efros, A.A.: Sequential modeling enables scalable learning for large vision models. arXiv preprint arXiv:2312.00785 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.324, + 0.785, + 0.35 + ], + "angle": 0, + "content": "5. Bao, H., Dong, L., Piao, S., Wei, F.: Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254 (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.35, + 0.785, + 0.391 + ], + "angle": 0, + "content": "6. Barron, J.T., Malik, J.: Intrinsic scene properties from a single rgb-d image. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 17-24 (2013) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.391, + 0.785, + 0.431 + ], + "angle": 0, + "content": "7. Barron, J.T., Malik, J.: Shape, illumination, and reflectance from shading. IEEE transactions on pattern analysis and machine intelligence 37(8), 1670-1687 (2014) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.432, + 0.785, + 0.459 + ], + "angle": 0, + "content": "8. Bartol, K., Bojanic, D., Petkovic, T., Pribanic, T.: A review of body measurement using 3d scanning. IEEE Access 9, 67281-67301 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.459, + 0.785, + 0.499 + ], + "angle": 0, + "content": "9. Bhat, S.F., Alhashim, I., Wonka, P.: Adabins: Depth estimation using adaptive bins. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4009-4018 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.5, + 0.785, + 0.54 + ], + "angle": 0, + "content": "10. Bhat, S.F., Birkl, R., Wofk, D., Wonka, P., Müller, M.: Zoodepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.541, + 0.785, + 0.567 + ], + "angle": 0, + "content": "11. Birkl, R., Wofk, D., Müller, M.: Midas v3. 1-a model zoo for robust monocular relative depth estimation. arXiv preprint arXiv:2307.14460 (2023) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.568, + 0.785, + 0.608 + ], + "angle": 0, + "content": "12. Bojic, L.: Metaverse through the prism of power and addiction: what will happen when the virtual world becomes more attractive than reality? European Journal of Futures Research 10(1), 1-24 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.608, + 0.785, + 0.662 + ], + "angle": 0, + "content": "13. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.663, + 0.785, + 0.703 + ], + "angle": 0, + "content": "14. Cao, Z., Simon, T., Wei, S., Sheikh, Y.: Realtime multi-person 2d pose estimation using part affinity fields. corr abs/1611.08050. arXiv preprint arXiv:1611.08050 (2016) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.704, + 0.785, + 0.744 + ], + "angle": 0, + "content": "15. Cao, Z., Hidalgo, G., Simon, T., Wei, S.E., Sheikh, Y.: Openpose: real-time multi-person 2d pose estimation using part affinity fields. arXiv preprint arXiv:1812.08008 (2018) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.744, + 0.785, + 0.799 + ], + "angle": 0, + "content": "16. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "17. Chan, C., Ginosar, S., Zhou, T., Efros, A.A.: Everybody dance now. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 5933-5942 (2019) 1" + }, + { + "type": "list", + "bbox": [ + 0.227, + 0.173, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.4, + 0.127 + ], + "angle": 0, + "content": "R. Khirodkar et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "18. Chen, L.C., Zhu, Y., Papandreou, G., Schroff, F., Adam, H.: Encoder-decoder with atrous separable convolution for semantic image segmentation. In: Proceedings of the European conference on computer vision (ECCV). pp. 801-818 (2018) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.191, + 0.787, + 0.233 + ], + "angle": 0, + "content": "19. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International conference on machine learning. pp. 1597-1607. PMLR (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.234, + 0.787, + 0.275 + ], + "angle": 0, + "content": "20. Chen, X., Wang, X., Changpinyo, S., Piergiovanni, A., Padlewski, P., Salz, D., Goodman, S., Grycner, A., Mustafa, B., Beyer, L., et al.: Pali: A jointly-scaled multilingual language-image model. arXiv preprint arXiv:2209.06794 (2022) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.276, + 0.787, + 0.317 + ], + "angle": 0, + "content": "21. Chen, Y., Wang, Z., Peng, Y., Zhang, Z., Yu, G., Sun, J.: Cascaded pyramid network for multi-person pose estimation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7103-7112 (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.318, + 0.787, + 0.359 + ], + "angle": 0, + "content": "22. Cheng, B., Misra, I., Schwing, A., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. arxiv 2022. arXiv preprint arXiv:2112.01527 (2021) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.361, + 0.787, + 0.389 + ], + "angle": 0, + "content": "23. Couprie, C., Farabet, C., Najman, L., LeCun, Y.: Indoor semantic segmentation using depth information. arXiv preprint arXiv:1301.3572 (2013) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.39, + 0.787, + 0.431 + ], + "angle": 0, + "content": "24. Dai, Z., Liu, H., Le, Q.V., Tan, M.: Coatnet: Marrying convolution and attention for all data sizes. Advances in neural information processing systems 34, 3965-3977 (2021) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.432, + 0.787, + 0.487 + ], + "angle": 0, + "content": "25. Dehghani, M., Djolonga, J., Mustafa, B., Padlewski, P., Heek, J., Gilmer, J., Steiner, A.P., Caron, M., Geirhos, R., Alabdulmohsin, I., et al.: Scaling vision transformers to 22 billion parameters. In: International Conference on Machine Learning. pp. 7480-7512. PMLR (2023) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.489, + 0.787, + 0.53 + ], + "angle": 0, + "content": "26. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.531, + 0.787, + 0.585 + ], + "angle": 0, + "content": "27. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.587, + 0.787, + 0.642 + ], + "angle": 0, + "content": "28. Drobyshev, N., Chelishev, J., Khakhulin, T., Ivakhnenko, A., Lempitsky, V., Zakharov, E.: Megaportraits: One-shot megapixel neural head avatars. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 2663-2671 (2022) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.643, + 0.787, + 0.699 + ], + "angle": 0, + "content": "29. Du, Y., Kips, R., Pumarola, A., Starke, S., Thabet, A., Sanakoyeu, A.: Avatars grow legs: Generating smooth human motion from sparse tracking inputs with diffusion model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 481-490 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.7, + 0.787, + 0.755 + ], + "angle": 0, + "content": "30. Eftekhar, A., Sax, A., Malik, J., Zamir, A.: Omnidata: A scalable pipeline for making multi-task mid-level vision datasets from 3d scans. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10786-10796 (2021) 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.757, + 0.787, + 0.798 + ], + "angle": 0, + "content": "31. Eigen, D., Fergus, R.: Predicting depth, surface normals and semantic labels with a common multi-scale convolutional architecture. In: Proceedings of the IEEE international conference on computer vision. pp. 2650-2658 (2015) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "32. Eigen, D., Puhrsch, C., Fergus, R.: Depth map prediction from a single image using a multi-scale deep network. Advances in neural information processing systems 27 (2014) 4, 7" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.679, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Sapiens" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "33. El-Nouby, A., Izacard, G., Touvron, H., Laptev, I., Jegou, H., Grave, E.: Are large-scale datasets necessary for self-supervised pre-training? arXiv preprint arXiv:2112.10740 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.191, + 0.785, + 0.232 + ], + "angle": 0, + "content": "34. El-Nouby, A., Klein, M., Zhai, S., Bautista, M.A., Toshev, A., Shankar, V., Susskind, J.M., Joulin, A.: Scalable pre-training of large autoregressive image models. arXiv preprint arXiv:2401.08541 (2024) 2, 3, 4, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.233, + 0.785, + 0.274 + ], + "angle": 0, + "content": "35. Fang, H.S., Lu, G., Fang, X., Xie, J., Tai, Y.W., Lu, C.: Weakly and semi supervised human body part parsing via pose-guided knowledge transfer. arXiv preprint arXiv:1805.04310 (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.276, + 0.785, + 0.317 + ], + "angle": 0, + "content": "36. Fang, H.S., Xie, S., Tai, Y.W., Lu, C.: Rmpe: Regional multi-person pose estimation. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 2334-2343 (2017) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.318, + 0.785, + 0.346 + ], + "angle": 0, + "content": "37. Fang, Y., Sun, Q., Wang, X., Huang, T., Wang, X., Cao, Y.: Eva-02: A visual representation for neon genesis. arXiv preprint arXiv:2303.11331 (2023) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.347, + 0.785, + 0.402 + ], + "angle": 0, + "content": "38. Fang, Y., Wang, W., Xie, B., Sun, Q., Wu, L., Wang, X., Huang, T., Wang, X., Cao, Y.: Eva: Exploring the limits of masked visual representation learning at scale. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19358-19369 (2023) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.403, + 0.785, + 0.444 + ], + "angle": 0, + "content": "39. Fouhey, D.F., Gupta, A., Hebert, M.: Data-driven 3d primitives for single image understanding. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 3392-3399 (2013) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.445, + 0.785, + 0.487 + ], + "angle": 0, + "content": "40. Gong, K., Liang, X., Li, Y., Chen, Y., Yang, M., Lin, L.: Instance-level human parsing via part grouping network. In: Proceedings of the European conference on computer vision (ECCV). pp. 770-785 (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.488, + 0.785, + 0.543 + ], + "angle": 0, + "content": "41. Gong, K., Liang, X., Zhang, D., Shen, X., Lin, L.: Look into person: Self-supervised structure-sensitive learning and a new benchmark for human parsing. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 932-940 (2017) 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.544, + 0.785, + 0.585 + ], + "angle": 0, + "content": "42. Goyal, P., Caron, M., Lefaudeauux, B., Xu, M., Wang, P., Pai, V., Singh, M., Liptchinsky, V., Misra, I., Joulin, A., et al.: Self-supervised pretraining of visual features in the wild. arXiv preprint arXiv:2103.01988 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.587, + 0.785, + 0.628 + ], + "angle": 0, + "content": "43. Guizilini, V., Vasiljevic, I., Chen, D., Ambrus, R., Gaidon, A.: Towards zero-shot scale-aware monocular depth estimation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9233-9243 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.629, + 0.785, + 0.685 + ], + "angle": 0, + "content": "44. Halstead, M.A., Barsky, B.A., Klein, S.A., Mandell, R.B.: Reconstructing curved surfaces from specular reflection patterns using spline surface fitting of normals. In: Proceedings of the 23rd annual conference on Computer graphics and interactive techniques. pp. 335-342 (1996) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.686, + 0.785, + 0.727 + ], + "angle": 0, + "content": "45. He, K., Chen, X., Xie, S., Li, Y., Dollár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022) 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.728, + 0.785, + 0.769 + ], + "angle": 0, + "content": "46. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729-9738 (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.77, + 0.785, + 0.798 + ], + "angle": 0, + "content": "47. He, K., Gkioxari, G., Dollár, P., Girshick, R.B.: Mask r-cnn. corr abs/1703.06870 (2017). arXiv preprint arXiv:1703.06870 (2017) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "48. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) 5" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.4, + 0.127 + ], + "angle": 0, + "content": "R. Khirodkar et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "49. Hoffmann, J., Borgeaud, S., Mensch, A., Buchatskaya, E., Cai, T., Rutherford, E., Casas, D.d.L., Hendricks, L.A., Welbl, J., Clark, A., et al.: Training compute-optimal large language models. arXiv preprint arXiv:2203.15556 (2022) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.189, + 0.785, + 0.23 + ], + "angle": 0, + "content": "50. Hu, L., Gao, X., Zhang, P., Sun, K., Zhang, B., Bo, L.: animate anyone: Consistent and controllable image-to-video synthesis for character animation. arXiv preprint arXiv:2311.17117 (2023) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.23, + 0.785, + 0.271 + ], + "angle": 0, + "content": "51. Huang, S., Gong, M., Tao, D.: A coarse-fine network for keypoint localization. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 3028-3037 (2017) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.271, + 0.785, + 0.311 + ], + "angle": 0, + "content": "52. Jafarian, Y., Park, H.S.: Learning high fidelity depths of dressed humans by watching social media dance videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12753-12762 (2021) 4, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.311, + 0.785, + 0.352 + ], + "angle": 0, + "content": "53. Jiang, A.Q., Sablayrolles, A., Mensch, A., Bamford, C., Chaplot, D.S., Casas, D.d.l., Bressand, F., Lengyel, G., Lample, G., Saulnier, L., et al.: Mistral 7b. arXiv preprint arXiv:2310.06825 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.352, + 0.785, + 0.393 + ], + "angle": 0, + "content": "54. Jiang, T., Lu, P., Zhang, L., Ma, N., Han, R., Lyu, C., Li, Y., Chen, K.: Rtmpose: Real-time multi-person pose estimation based on mmpose. arXiv preprint arXiv:2303.07399 (2023) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.393, + 0.785, + 0.447 + ], + "angle": 0, + "content": "55. Jin, S., Xu, L., Xu, J., Wang, C., Liu, W., Qian, C., Ouyang, W., Luo, P.: Whole-body human pose estimation in the wild. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IX 16. pp. 196-214. Springer (2020) 3, 6, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.447, + 0.785, + 0.473 + ], + "angle": 0, + "content": "56. Kato, N., Li, T., Nishino, K., Uchida, Y.: Improving multi-person pose estimation using label correction. arXiv preprint arXiv:1811.03331 (2018) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.473, + 0.785, + 0.514 + ], + "angle": 0, + "content": "57. Khirodkar, R., Chari, V., Agrawal, A., Tyagi, A.: Multi-instance pose networks: Rethinking top-down pose estimation. In: Proceedings of the IEEE/CVF International conference on computer vision. pp. 3122-3131 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.514, + 0.785, + 0.555 + ], + "angle": 0, + "content": "58. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.555, + 0.785, + 0.582 + ], + "angle": 0, + "content": "59. Kocabas, M., Chang, J.H.R., Gabriel, J., Tuzel, O., Ranjan, A.: Hugs: Human gaussian splats. arXiv preprint arXiv:2311.17910 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.582, + 0.785, + 0.636 + ], + "angle": 0, + "content": "60. Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., et al.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision 123, 32-73 (2017) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.636, + 0.785, + 0.663 + ], + "angle": 0, + "content": "61. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images. arXiv preprint (2009) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.663, + 0.785, + 0.717 + ], + "angle": 0, + "content": "62. Ladicky, L., Zeisl, B., Pollefeys, M.: Discriminatively trained dense surface normal estimation. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 468-484. Springer (2014) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.718, + 0.785, + 0.759 + ], + "angle": 0, + "content": "63. Lawrence, J., Goldman, D.B., Achar, S., Blascovich, G.M., Desloge, J.G., Fortes, T., Gomez, E.M., Haberling, S., Hoppe, H., Huibers, A., et al.: Project starline: A high-fidelity telepresence system. arXiv preprint (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.759, + 0.785, + 0.813 + ], + "angle": 0, + "content": "64. Levoy, M., Pulli, K., Curless, B., Rusinkiewicz, S., Koller, D., Pereira, L., Ginzton, M., Anderson, S., Davis, J., Ginsberg, J., et al.: The digital michelangelo project: 3d scanning of large statues. In: Proceedings of the 27th annual conference on Computer graphics and interactive techniques. pp. 131-144 (2000) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "65. Lewkowycz, A.: How to decay your learning rate. arXiv preprint arXiv:2103.12682 (2021) 8" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.679, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Sapiens" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.224, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "66. Li, Z., Wang, X., Liu, X., Jiang, J.: Binsformer: Revisiting adaptive bins for monocular depth estimation. arXiv preprint arXiv:2204.00987 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.218 + ], + "angle": 0, + "content": "67. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: European conference on computer vision. pp. 740-755. Springer (2014) 1, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.219, + 0.785, + 0.274 + ], + "angle": 0, + "content": "68. Liu, Z., Hu, H., Lin, Y., Yao, Z., Xie, Z., Wei, Y., Ning, J., Cao, Y., Zhang, Z., Dong, L., et al.: Swin transformer v2: Scaling up capacity and resolution. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12009-12019 (2022) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.276, + 0.785, + 0.316 + ], + "angle": 0, + "content": "69. Lombardi, S., Simon, T., Saragih, J., Schwartz, G., Lehrmann, A., Sheikh, Y.: Neural volumes: Learning dynamic renderable volumes from images. arXiv preprint arXiv:1906.07751 (2019) 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.318, + 0.785, + 0.358 + ], + "angle": 0, + "content": "70. Lombardi, S., Simon, T., Schwartz, G., Zollhoefer, M., Sheikh, Y., Saragih, J.: Mixture of volumetric primitives for efficient neural rendering. ACM Transactions on Graphics (ToG) 40(4), 1-13 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.361, + 0.785, + 0.387 + ], + "angle": 0, + "content": "71. Long, J., Shelhamer, E., Darrell, T., Berkeley, U.: Fully convolutional networks for semantic segmentation. arxiv 2015. arXiv preprint arXiv:1411.4038 (2014) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.389, + 0.785, + 0.416 + ], + "angle": 0, + "content": "72. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983 (2016) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.418, + 0.785, + 0.445 + ], + "angle": 0, + "content": "73. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.446, + 0.785, + 0.472 + ], + "angle": 0, + "content": "74. Lowe, D.G.: Three-dimensional object recognition from single two-dimensional images. Artificial intelligence 31(3), 355-395 (1987) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.475, + 0.785, + 0.515 + ], + "angle": 0, + "content": "75. Luo, Y., Zheng, Z., Zheng, L., Guan, T., Yu, J., Yang, Y.: Macro-micro adversarial network for human parsing. In: Proceedings of the European conference on computer vision (ECCV). pp. 418-434 (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.517, + 0.785, + 0.558 + ], + "angle": 0, + "content": "76. Ma, S., Simon, T., Saragih, J., Wang, D., Li, Y., De La Torre, F., Sheikh, Y.: Pixel codec avatars. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 64-73 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.56, + 0.785, + 0.613 + ], + "angle": 0, + "content": "77. Mahajan, D., Girshick, R., Ramanathan, V., He, K., Paluri, M., Li, Y., Bharambe, A., Van Der Maaten, L.: Exploring the limits of weakly supervised pretraining. In: Proceedings of the European conference on computer vision (ECCV). pp. 181-196 (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.615, + 0.785, + 0.654 + ], + "angle": 0, + "content": "78. Newell, A., Yang, K., Deng, J.: Stacked hourglass networks for human pose estimation. In: European conference on computer vision. pp. 483-499. Springer (2016) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.657, + 0.785, + 0.711 + ], + "angle": 0, + "content": "79. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023) 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.714, + 0.785, + 0.769 + ], + "angle": 0, + "content": "80. Papandreou, G., Zhu, T., Kanazawa, N., Toshev, A., Tompson, J., Bregler, C., Murphy, K.: Towards accurate multi-person pose estimation in the wild. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4903-4911 (2017) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.77, + 0.785, + 0.811 + ], + "angle": 0, + "content": "81. Peebles, W., Xie, S.: Scalable diffusion models with transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4195-4205 (2023) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "82. render people: 3d people for architectural visualization | renderpeople, https://renderpeople.com/3d-people, accessed: 2024-02-22 3" + }, + { + "type": "list", + "bbox": [ + 0.224, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "20" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.4, + 0.127 + ], + "angle": 0, + "content": "R. Khirodkar et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.147, + 0.787, + 0.204 + ], + "angle": 0, + "content": "83. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.205, + 0.787, + 0.26 + ], + "angle": 0, + "content": "84. Ranftl, R., Lasinger, K., Hafner, D., Schindler, K., Koltun, V.: Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence 44(3), 1623-1637 (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.261, + 0.787, + 0.315 + ], + "angle": 0, + "content": "85. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.317, + 0.787, + 0.372 + ], + "angle": 0, + "content": "86. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al.: Imagenet large scale visual recognition challenge. International journal of computer vision 115, 211-252 (2015) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.374, + 0.787, + 0.43 + ], + "angle": 0, + "content": "87. Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.431, + 0.787, + 0.486 + ], + "angle": 0, + "content": "88. Saito, S., Huang, Z., Natsume, R., Morishima, S., Kanazawa, A., Li, H.: Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 2304-2314 (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.487, + 0.787, + 0.543 + ], + "angle": 0, + "content": "89. Saito, S., Simon, T., Saragih, J., Joo, H.: Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 84-93 (2020) 1, 4, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.544, + 0.787, + 0.599 + ], + "angle": 0, + "content": "90. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.6, + 0.787, + 0.655 + ], + "angle": 0, + "content": "91. Singh, M., Duval, Q., Alwala, K.V., Fan, H., Aggarwal, V., Adcock, A., Joulin, A., Dollar, P., Feichtenhofer, C., Girshick, R., et al.: The effectiveness of mae pretraining for billion-scale pretraining. arXiv preprint arXiv:2303.13496 (2023) 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.656, + 0.787, + 0.699 + ], + "angle": 0, + "content": "92. Sun, C., Shrivastava, A., Singh, S., Gupta, A.: Revisiting unreasonable effectiveness of data in deep learning era. In: Proceedings of the IEEE international conference on computer vision. pp. 843-852 (2017) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.7, + 0.787, + 0.741 + ], + "angle": 0, + "content": "93. Sun, K., Xiao, B., Liu, D., Wang, J.: Deep high-resolution representation learning for human pose estimation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5693-5703 (2019) 4, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.742, + 0.787, + 0.798 + ], + "angle": 0, + "content": "94. Taori, R., Gulrajani, I., Zhang, T., Dubois, Y., Li, X., Guestrin, C., Liang, P., Hashimoto, T.B.: Alpaca: A strong, replicable instruction-following model. Stanford Center for Research on Foundation Models. https://crfm.stanford.edu/2023/03/13/alpaca.html 3(6), 7 (2023) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "95. Tay, Y., Dehghani, M., Rao, J., Fedus, W., Abnar, S., Chung, H.W., Narang, S., Yogatama, D., Vaswani, A., Metzler, D.: Scale efficiently: Insights from pretraining and fine-tuning transformers. arXiv preprint arXiv:2109.10686 (2021) 4" + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.679, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Sapiens" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "21" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "96. Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.191, + 0.786, + 0.233 + ], + "angle": 0, + "content": "97. Thomee, B., Shamma, D.A., Friedland, G., Elizalde, B., Ni, K., Poland, D., Borth, D., Li, L.J.: Yfcc100m: The new data in multimedia research. Communications of the ACM 59(2), 64-73 (2016) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.234, + 0.786, + 0.274 + ], + "angle": 0, + "content": "98. Toshev, A., Szegedy, C.: Deepest: Human pose estimation via deep neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1653-1660 (2014) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.276, + 0.786, + 0.317 + ], + "angle": 0, + "content": "99. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.318, + 0.785, + 0.359 + ], + "angle": 0, + "content": "100. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023) 4, 5, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.36, + 0.785, + 0.402 + ], + "angle": 0, + "content": "101. Wang, X., Fouhey, D., Gupta, A.: Designing deep networks for surface normal estimation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 539-547 (2015) 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.403, + 0.785, + 0.457 + ], + "angle": 0, + "content": "102. Weng, C.Y., Curless, B., Srinivasan, P.P., Barron, J.T., Kemelmacher-Shlizerman, I.: Humannerf: Free-viewpoint rendering of moving people from monocular video. In: Proceedings of the IEEE/CVF conference on computer vision and pattern Recognition. pp. 16210-16220 (2022) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.459, + 0.785, + 0.487 + ], + "angle": 0, + "content": "103. Wu, Y., Kirillov, A., Massa, F., Lo, W.Y., Girshick, R.: Detector2. https://github.com/facebookresearch/detectron2 (2019) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.488, + 0.785, + 0.543 + ], + "angle": 0, + "content": "104. Xia, F., Wang, P., Chen, L.C., Yuille, A.L.: Zoom better to see clearer: Human and object parsing with hierarchical auto-zoom net. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14. pp. 648-663. Springer (2016) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.544, + 0.785, + 0.585 + ], + "angle": 0, + "content": "105. Xia, F., Wang, P., Chen, X., Yuille, A.L.: Joint multi-person pose estimation and semantic part segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6769-6778 (2017) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.587, + 0.785, + 0.628 + ], + "angle": 0, + "content": "106. Xiao, B., Wu, H., Wei, Y.: Simple baselines for human pose estimation and tracking. In: Proceedings of the European conference on computer vision (ECCV). pp. 466-481 (2018) 4, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.629, + 0.785, + 0.67 + ], + "angle": 0, + "content": "107. Xie, E., Wang, W., Yu, Z., Anandkumar, A., Alvarez, J.M., Luo, P.: Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in neural information processing systems 34, 12077-12090 (2021) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.671, + 0.785, + 0.712 + ], + "angle": 0, + "content": "108. Xiu, Y., Yang, J., Cao, X., Tzionas, D., Black, M.J.: Econ: Explicit clothed humans optimized via normal integration. arXiv preprint arXiv:2212.07422 (2022) 1, 4, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.714, + 0.785, + 0.755 + ], + "angle": 0, + "content": "109. Xiu, Y., Yang, J., Tzionas, D., Black, M.J.: Icon: Implicit clothed humans obtained from normals. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13286-13296. IEEE (2022) 1, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.757, + 0.785, + 0.798 + ], + "angle": 0, + "content": "110. Xu, L., Jin, S., Liu, W., Qian, C., Ouyang, W., Luo, P., Wang, X.: Zoomnas: searching for whole-body human pose estimation in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(4), 5296-5313 (2022) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "111. Xu, Y., Zhang, J., Zhang, Q., Tao, D.: Vitpose: Simple vision transformer baselines for human pose estimation. Advances in Neural Information Processing Systems 35, 38571-38584 (2022) 6, 9" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.147, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "22" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.4, + 0.127 + ], + "angle": 0, + "content": "R. Khirodkar et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.188 + ], + "angle": 0, + "content": "112. Xu, Y., Zhang, J., Zhang, Q., Tao, D.: Vitpose+: Vision transformer foundation model for generic body pose estimation. arXiv preprint arXiv:2212.04246 (2022) 9, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.189, + 0.785, + 0.231 + ], + "angle": 0, + "content": "113. Yang, L., Kang, B., Huang, Z., Xu, X., Feng, J., Zhao, H.: Depth anything: Unleashing the power of large-scale unlabeled data. arXiv preprint arXiv:2401.10891 (2024) 1, 3, 4, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.231, + 0.785, + 0.272 + ], + "angle": 0, + "content": "114. Yang, Z., Zeng, A., Yuan, C., Li, Y.: Effective whole-body pose estimation with two-stages distillation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4210-4220 (2023) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.272, + 0.785, + 0.314 + ], + "angle": 0, + "content": "115. Yang, Z., Dai, Z., Yang, Y., Carbonell, J., Salakhutdinov, R.R., Le, Q.V.: Xlnet: Generalized autoregressive pretraining for language understanding. Advances in neural information processing systems 32 (2019) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.314, + 0.785, + 0.368 + ], + "angle": 0, + "content": "116. Yin, Y., Guo, C., Kaufmann, M., Zarate, J.J., Song, J., Hilliges, O.: Hi4d: 4d instance segmentation of close human interaction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17016-17027 (2023) 10, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.37, + 0.787, + 0.424 + ], + "angle": 0, + "content": "117. Yu, T., Zheng, Z., Guo, K., Liu, P., Dai, Q., Liu, Y.: Function4d: Real-time human volumetric capture from very sparse consumer rgbd sensors. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5746-5756 (2021) 10, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.424, + 0.785, + 0.467 + ], + "angle": 0, + "content": "118. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3836-3847 (2023) 1, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.522 + ], + "angle": 0, + "content": "119. Zhang, S.H., Li, R., Dong, X., Rosin, P., Cai, Z., Han, X., Yang, D., Huang, H., Hu, S.M.: Pose2seg: Detection free human instance segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 889-898 (2019) 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.785, + 0.563 + ], + "angle": 0, + "content": "120. Zhang, X., Lu, Y., Wang, W., Yan, A., Yan, J., Qin, L., Wang, H., Yan, X., Wang, W.Y., Petzold, L.R.: Gpt-4v (isdiction) as a generalist evaluator for vision-language tasks. arXiv preprint arXiv:2311.01361 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.785, + 0.604 + ], + "angle": 0, + "content": "121. Zhou, J., Wei, C., Wang, H., Shen, W., Xie, C., Yuille, A., Kong, T.: ibot: Image bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832 (2021) 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.605, + 0.785, + 0.647 + ], + "angle": 0, + "content": "122. Zlateski, A., Jaroensri, R., Sharma, P., Durand, F.: On the importance of label quality for semantic segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1479-1487 (2018) 3" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.647 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Sapiens_ Foundation for Human Vision Models/726c3ced-d623-4195-b91f-26812a59bf2e_origin.pdf b/2024/Sapiens_ Foundation for Human Vision Models/726c3ced-d623-4195-b91f-26812a59bf2e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..567fae7c250302afa8989ea39cdd3ca377fe47fd --- /dev/null +++ b/2024/Sapiens_ Foundation for Human Vision Models/726c3ced-d623-4195-b91f-26812a59bf2e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d69446b0d20fe518655709cfeb55551762d87ea8651c06813fb5e400322386f2 +size 18083643 diff --git a/2024/Sapiens_ Foundation for Human Vision Models/full.md b/2024/Sapiens_ Foundation for Human Vision Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e4c58540c592870fef7dc3d3637153531df8c209 --- /dev/null +++ b/2024/Sapiens_ Foundation for Human Vision Models/full.md @@ -0,0 +1,388 @@ +# Sapiens: Foundation for Human Vision Models + +Rawal Khirodkar, Timur Bagautdinov, Julieta Martinez, Su Zhaoen, Austin James, Peter Selednik, Stuart Anderson, Shunsuke Saito + +Codec Avatars Lab, Meta + +Abstract. We present Sapiens, a family of models for four fundamental human-centric vision tasks - 2D pose estimation, body-part segmentation, depth estimation, and surface normal prediction. Our models natively support 1K high-resolution inference and are extremely easy to adapt for individual tasks by simply fine-tuning foundation models pretrained on over 300 million in-the-wild human images. We observe that, given the same computational budget, self-supervised pretraining on a curated dataset of human images significantly boosts the performance for a diverse set of human-centric tasks. The resulting models exhibit remarkable generalization to in-the-wild data, even when labeled data is scarce or entirely synthetic. Our simple model design also brings scalability - model performance across tasks significantly improves as we scale the number of parameters from 0.3 to 2 billion. Sapiens consistently surpasses existing complex baselines across various human-centric benchmarks. Specifically, we achieve significant improvements over the prior state-of-the-art on Humans-5K (pose) by $7.6\mathrm{mAP}$ , Humans-2K (part-seg) by $17.1\mathrm{mIoU}$ , Hi4D (depth) by $22.4\%$ relative RMSE, and THuman2 (normal) by $53.5\%$ relative angular error. + +Keywords: Human-centric, Large-scale pretraining, In-the-wild + +"Sapiens—pertaining to, or resembling modern humans." + +# 1 Introduction + +Recent years have witnessed remarkable strides towards generating photorealistic humans in 2D [17, 28, 50, 118] and 3D [69, 89, 102, 109]. The success of these methods is greatly attributed to the robust estimation of various assets such as 2D keypoints [14,67], fine-grained body-part segmentation [119], depth [113], and surface normals [89, 108]. However, robust and accurate estimation of these assets is still an active research area, and complicated systems to boost performance for individual tasks often hinder wider adoption. Moreover, obtaining accurate ground-truth annotation in-the-wild is notoriously difficult to scale. Our goal is to provide a unified framework and models to infer these assets in-the-wild to unlock a wide range of human-centric applications for everybody. + +We argue that such human-centric models should satisfy three criteria: generalization, broad applicability, and high fidelity. Generalization ensures robustness to unseen conditions, enabling the model to perform consistently across varied environments. Broad applicability indicates the versatility of the model, making it suitable for a wide range of tasks with minimal modifications. High fidelity denotes the ability of the model to produce precise, high-resolution outputs, + +![](images/5e2c5b544eea5e6c23ea0772008ae77fed6efeed3b8f0adaac0338426b159b3e.jpg) +Fig. 1: Sapiens models are finetuned for four human tasks - 2D pose estimation, body-part segmentation, depth prediction and normal prediction. Our models generalize across a variety of in-the-wild face, upper-body, full-body and multi-person images. + +essential for faithful human generation tasks such as 2D to 3D lifting. This paper details the development of models that embody these attributes, collectively referred to as Sapiens. + +Following the insights from [34, 79, 91], leveraging large datasets and scalable model architectures is key for generalization. For broader applicability, we adopt the pretrain-then-finetune approach, enabling post-pretraining adaptation to specific tasks with minimal adjustments. This approach raises a critical question: What type of data is most effective for pretraining? Given computational limits, should the emphasis be on collecting as many human images as possible, or is it preferable to pretrain on a less curated set to better reflect real-world variability? Existing methods often overlook the pretraining data distribution in the context of downstream tasks. To study the influence of pretraining data distribution on human-specific tasks, we collect the Humans-300M dataset, featuring 300 million diverse human images. These unlabelled images are used to pretrain a family of vision transformers [27] from scratch, with parameter counts ranging from 300M to 2B. + +Among various self-supervision methods for learning general-purpose visual features from large datasets [5,19,34,45,46,121], we choose the masked-autoencoder (MAE) approach [45] for its simplicity and efficiency in pretraining. MAE, having a single-pass inference model compared to contrastive or multi-inference strategies, allows processing a larger volume of images with the same computational resources. For higher-fidelity, in contrast to prior methods, we increase the native input resolution of our pretraining to 1024 pixels, resulting in a $\sim 4\times$ increase in FLOPs compared to the largest existing vision backbone [91]. Each model is pretrained on 1.2 trillion tokens. Table 1 outlines a comparison with earlier approaches. For finetuning on human-centric tasks [15, 101, 113, 119], we use a consistent encoder-decoder architecture. The encoder is initialized with weights from pretraining, while the decoder, a lightweight and task-specific head, is initialized randomly. Both components are then finetuned in an end-to-end manner. We focus on four key tasks - 2D pose estimation, body-part segmentation, depth, and normal estimation, as shown in Fig. 1. + +Consistently with prior studies [56,122], we affirm the critical impact of label quality on the model's in-the-wild performance. Public benchmarks [23,41,55] often contain noisy labels, providing inconsistent supervisory signals during model fine-tuning. At the same time, it is important to utilize fine-grained and precise annotations to align closely with our primary goal of 3D human digitization. To this end, we propose a substantially denser set of 2D whole body keypoints for pose estimation and a detailed class vocabulary for body part segmentation, surpassing the scope of previous datasets (please refer to Fig. 1). Specifically, we introduce a comprehensive collection of 308 keypoints encompassing the body, hands, feet, surface, and face. Additionally, we expand the segmentation class vocabulary to 28 classes, covering body parts such as the hair, tongue, teeth, upper/lower lip, and torso. To guarantee the quality and consistency of annotations and a high degree of automation, we utilize a multi-view capture setup to collect pose and segmentation annotations. We also utilize human-centric synthetic data for depth and normal estimation, leveraging 600 detailed scans from RenderPeople [82] to generate high-resolution depth maps and surface normals. + +We show that the combination of domain-specific large-scale pretraining with limited, yet high-quality annotations leads to robust in-the-wild generalization. Overall, our method demonstrates an effective strategy for developing highly precise discriminative models capable of performing in real-world scenarios without the need for collecting a costly and diverse set of annotations. + +
MethodDataset#ParamsGFLOPsImage sizeDomain
DINO [16]ImageNet1k86 M17.6224General
iBOT [121]ImageNet21k307 M61.6224General
DINOv2 [79]LVD-142M1 B291.0224General
ViT-6.5B [91]IG-3B6.5 B1657.0224General
AIM [34]DFN-2B6.5 B1657.0224General
Sapiens (Ours)Humans-300M2 B8709.01024Human
+ +Table 1: Comparison of state-of-the-art pretrained vision models. Sapiens adopts a higher resolution backbone on a large dataset of in-the-wild human images. + +Our contributions are summarized as follows. + +- We introduce Sapiens, a family of vision transformers pretrained on a large-scale, curated dataset of diverse human images. +- This study shows that simple data curation and large-scale pretraining significantly boost the model's performance with the same computational budget. +- Our models, fine-tuned with high-quality annotations or even synthetic data, demonstrate robust in-the-wild generalization. +- The first 1K high-resolution model that natively supports high-fidelity inference essential for human-centric tasks, achieving state-of-the-art performance on benchmarks for 2D pose estimation, body-part segmentation, depth, and normal estimation. + +# 2 Related Work + +Our work explores the limits of training large architectures on a large number of in-the-wild human images. We build on prior work from different areas: pretraining at scale, human vision tasks, and large vision transformers. + +Pretraining at Scale. The remarkable success of large-scale pretraining [26,95] followed by task-specific finetuning for language modeling [2,13,53,96,99,100] has established this approach as a standard practice. Similarly, computer vision methods [1,4,33,34,42,79,83,85,87,120] are progressively embracing extensive data scales for pretraining. The emergence of large datasets, such as LAION-5B [90], Instagram-3.5B [77], JFT-300M [92], LVD-142M [79], Visual Genome [60], and YFCC100M [97], has enabled the exploration of a data corpus well beyond the scope of traditional benchmarks [61,67,86]. Salient work in this domain includes DINOv2 [79], MAWS [91], and AIM [34]. DINOv2 achieves state-of-the-art performance in generating self-supervised features by scaling the contrastive iBot [121] method on the LDV-142M dataset [79]. MAWS [91] studies the scaling of masked-autoencoders (MAE) [45] on billion images. AIM [34] explores the scalability of autoregressive visual pretraining similar to BERT [26] for vision transformers [27]. In contrast to these methods which mainly focus on general image pretraining or zero-shot image classification, we take a distinctly human-centric approach: our models leverage a vast collection of human images for pretraining, subsequently fine-tuning for a range of human-related tasks. + +Human Vision Tasks. The pursuit of large-scale 3D human digitization [8, 44, 64, 74] remains a pivotal goal in computer vision [12]. Significant progress has been made within controlled or studio environments [3, 59, 63, 69, 70, 76, 89], yet challenges persist in extending these methods to unconstrained environments [29]. To address these challenges, developing versatile models capable of multiple fundamental tasks such as keypoint estimation [21, 36, 47, 51, 57, 78, 80, 93, 106], body-part segmentation [35, 40, 40, 41, 75, 104, 105], depth estimation [9, 10, 32, 43, 52, 66, 84, 113], and surface normal prediction [6, 7, 31, 39, 62, 88, 101, 108] from images in natural settings is crucial. In this work, we aim to develop models for these essential human vision tasks which generalize to in-the-wild settings. + +Scaling Architectures. Currently, the largest publicly-accessible language models contain upwards of 100B parameters [49], while the more commonly used language models [94, 100] contain around 7B parameters. In contrast, Vision Transformers (ViT) [27], despite sharing a similar architecture, have not been scaled to this extent successfully. While there are notable endeavors in this direction, including the development of a dense ViT-4B [20] trained on both text and images, and the formulation of techniques for the stable training of a ViT-22B [25], commonly utilized vision backbones still range between 300M to 600M parameters [24, 38, 48, 68] and are primarily pretrained at an image resolution of about 224 pixels. Similarly, existing transformer-based image generation models, such as DiT [81] use less than 700M parameters, and operate on a highly compressed latent space. To address this gap, we introduce Sapiens - a collection of large, high-resolution ViT models that are pretrained natively at a 1024 pixel image resolution on millions of human images. + +# 3 Method + +# 3.1 Humans-300M Dataset + +We utilize a large proprietary dataset for pretraining of approximately 1 billion in-the-wild images, focusing exclusively on human images. The preprocessing involves discarding images with watermarks, text, artistic depictions, or unnatural elements. Subsequently, we use an off-the-shelf person bounding-box detector [103] to filter images, retaining those with a detection score above 0.9 and bounding box dimensions exceeding 300 pixels. Fig. 2 provides an overview of the distribution of the number of people per image in over 248 million images contain multiple subjects. + +![](images/453feb69b9add65076d75c0c6acdf971ef91e4808b045636a3f341b277320dac.jpg) +Fig.2: Overview of number of humans per image in the Humans-300M dataset. + +# 3.2 Pretraining + +We follow the masked-autoencoder [45] (MAE) approach for pretraining. Our model is trained to reconstruct the original human image given its partial observation. Like all autoencoders, our model has an encoder that maps the visible image to a latent representation and a decoder that reconstructs the original image from this latent representation. Our pretraining dataset consists of both single and multi-human images; each image is resized to a fixed size with a square aspect ratio. Similar to ViT [27], we divide an image into regular non-overlapping patches with a fixed patch size. A subset of these patches is randomly selected and masked, leaving the rest visible. The proportion of masked patches to visible ones is defined as the masking ratio, which remains fixed throughout training. We refer to MAE [45] for more details. Fig. 3 (Top) shows the reconstruction of our pretrained model on unseen human images. Our models exhibit generalization across a variety of image characteristics including scales, crops, the age and ethnicity of subjects, and number of subjects. Each patch token in our model accounts for $0.02\%$ of the image area compared to $0.4\%$ in standard ViTs, a $16\times$ + +![](images/14bb481e41f6e40c447e02a883aadf2406e7016367fd43a989616fca2253ada2.jpg) +reduction - this provides a fine-grained inter-token reasoning for our models. Fig.3 (Bottom) shows that even with an increased mask ratio of $95\%$ , our model achieves a plausible reconstruction of human anatomy on held-out samples. + +# 3.3 2D Pose Estimation + +We follow the top-down 2D pose estimation paradigm, which aims to detect the locations of $K$ keypoints from an input image $\mathbf{I} \in \mathbb{R}^{H \times W \times 3}$ . Most methods pose this problem as heatmap prediction, where each of $K$ heatmaps represents the probability of the corresponding keypoint being at any spatial location. Similar to [111], we define a pose estimation transformer, $\mathcal{P}$ , for keypoint detection. The bounding box at training and inference is scaled to $H \times W$ and is provided as an input to $\mathcal{P}$ . Let $\mathbf{y} \in \mathbb{R}^{H \times W \times K}$ denote the $K$ heatmaps corresponding to the ground truth keypoints for a given input $\mathbf{I}$ . The pose estimator transforms input $\mathbf{I}$ to a set of predicted heatmaps, $\hat{\mathbf{y}} \in \mathbb{R}^{H \times W \times K}$ , such that $\hat{\mathbf{y}} = \mathcal{P}(\mathbf{I})$ . $\mathcal{P}$ is trained to minimize the mean squared loss $\mathcal{L}_{\mathrm{pose}} = \mathrm{MSE}(\mathbf{y},\hat{\mathbf{y}})$ . During finetuning, the encoder of $\mathcal{P}$ is initialized with the weights from pretraining, and the decoder is initialized randomly. The aspect ratio $H:W$ is set to be 4:3, with the pretrained positional embedding being interpolated accordingly [58]. We use lightweight decoders with deconvolution and convolution operations. + +We finetune both the encoder and the decoder in $\mathcal{P}$ across multiple skeleton formats, including $K = 17$ [67], $K = 133$ [55] and a new highly-detailed skeleton format, with $K = 308$ , as shown in Fig. 4 (Left). Compared to existing formats with at most 68 facial keypoints, our annotations consist of 243 facial keypoints, including representative points around the eyes, eyebrows, lips, nose, and ears. This design is tailored to meticulously capture the nuanced details of facial expressions in the real world. With these keypoints, we manually annotated 1 million images at $4K$ resolution from an indoor capture setup. + +![](images/c6937e5398dab721268539562b82ae78971bfcef323120b0a3eae4f1a07f8f1f.jpg) +a) Full-Body - 308 kps +c) Face - 243 kps +Fig. 4: Ground-truth annotations for 2D pose estimation and body-part segmentation. + +![](images/c4a8aa5300b95bf6610f839ee3521a86c8b6179f514319a51b85afc070ae7e33.jpg) +b) Hands - 40 kps + +![](images/b823aadf68f180e5b7cf54bb07d58404e4f4c8c5b7f11f80d2c0ca12e2aa81f2.jpg) + +![](images/be8e11fb02233f3669afa36ef6ae00a09a051c8df9cfac16407dd82738fe6cd0.jpg) + +![](images/2887a149f76bb9d884a9c99216ce4ea53cababd52aa66711d107bfbd8dca9ea6.jpg) +d) Body-Part Segmentation: 28 Classes + +![](images/90b238aa8bb10f6e41a8241bb013e9c7fd3c45915dcb9f69fd1208c0e5c8e372.jpg) + +# 3.4 Body-Part Segmentation + +Commonly referred to as human parsing, body-part segmentation aims to classify pixels in the input image $\mathbf{I}$ into $C$ classes. Most methods [41] transform this problem to estimating per-pixel class probabilities to create a probability map $\hat{\mathbf{p}}\in \mathbb{R}^{H\times W\times C}$ such that $\hat{\mathbf{p}} = S(\mathbf{I})$ , where $S$ is the segmentation model. As outlined previously, we adopt the same encoder-decoder architecture and initialization scheme for $S$ . $S$ is finetuned to minimize the weighted cross-entropy loss between the actual $\mathbf{p}$ and predicted $\hat{\mathbf{p}}$ probability maps, $\mathcal{L}_{\mathrm{seg}} = \mathrm{WeightedCE}(\mathbf{p},\hat{\mathbf{p}})$ . + +We finetune $\mathcal{S}$ across two part-segmentation vocabularies: a standard set with $C = 20$ [41] and a new larger vocabulary with $C = 28$ , as illustrated in Fig.4 (Right). Our proposed vocabulary goes beyond previous datasets in important ways. It distinguishes between the upper and lower halves of limbs and incorporates more detailed classifications such as upper/lower lips, teeth, and tongue. To this end, we manually annotate $100K$ images at $4K$ resolution with this vocabulary. + +# 3.5 Depth Estimation + +For depth estimation, we adopt the architecture used for segmentation, with the modification that the decoder output channel is set to 1 for regression. We denote the ground-truth depth map of image $\mathbf{I}$ by $\mathbf{d} \in \mathbb{R}^{H \times W}$ , the depth estimator by $\mathcal{D}$ , where $\hat{\mathbf{d}} = \mathcal{D}(\mathbf{I})$ , and $M$ as the number of human pixels in the image. For the relative depth estimation, we normalize $\mathbf{d}$ to the range [0, 1] using max and min depths in the image. The $\mathcal{L}_{\mathrm{depth}}$ loss [32] for $\mathcal{D}$ is defined as follows: + +$$ +\Delta \mathbf {d} = \log (\mathbf {d}) - \log (\hat {\mathbf {d}}), \tag {1} +$$ + +$$ +\overline {{\Delta \mathbf {d}}} = \frac {1}{M} \sum_ {i = 1} ^ {M} \Delta \mathbf {d} _ {i}, \quad \overline {{(\Delta \mathbf {d}) ^ {2}}} = \frac {1}{M} \sum_ {i = 1} ^ {M} (\Delta \mathbf {d} _ {i}) ^ {2}, \tag {2} +$$ + +$$ +\mathcal {L} _ {\mathrm {d e p t h}} = \sqrt {\left(\overline {{\Delta \mathbf {d}}} ^ {2}\right) - \frac {1}{2} \left(\overline {{\Delta \mathbf {d}}} ^ {2}\right) ^ {2}}. \tag {3} +$$ + +We render 500,000 synthetic images using 600 high-resolution photogrammetry human scans as shown in Fig. 5 to obtain a robust monocular depth + +estimation model with high-fidelity. A random background is selected from a 100 HDRI environment map collection. We place a virtual camera within the scene, randomly adjusting its focal length, rotation, and translation to capture images and their associated ground-truth depth maps at 4K resolution. + +![](images/b0c67ba670a15af50b812c31374d01223983b839a9491536af1a0ccb7b56bbee.jpg) +Fig. 5: Ground-truth synthetic annotations for depth and surface normal estimation. + +# 3.6 Surface Normal Estimation + +Similar to previous tasks, we set the decoder output channels of the normal estimator $\mathcal{N}$ to be 3, corresponding to the $xyz$ components of the normal vector at each pixel. The generated synthetic data is also used as supervision for surface normal estimation. Let $\mathbf{n}$ be the ground-truth normal map for image $\mathbf{I}$ and $\hat{\mathbf{n}} = \mathcal{N}(\mathbf{I})$ . Similar to depth, the loss $\mathcal{L}_{\mathrm{normal}}$ is only computed for human pixels in the image and is defined as follows: + +$$ +\mathcal {L} _ {\text {n o r m a l}} = \left\| \mathbf {n} - \hat {\mathbf {n}} \right\| _ {1} + (1 - \mathbf {n} \cdot \hat {\mathbf {n}}) \tag {4} +$$ + +# 4 Experiments + +In this section, we initially provide an overview of the implementation details. Subsequently, we conduct comprehensive benchmarking across four tasks: pose estimation, part segmentation, depth estimation, and normal estimation. + +# 4.1 Implementation Details + +Our largest model, Sapiens-2B, is pretrained using 1024 A100 GPUs for 18 days using PyTorch. We use the AdamW [73] optimizer for all our experiments. The learning schedule includes a brief linear warm-up, followed by cosine annealing [72] for pretraining and linear decay [65] for finetuning. All models are pretrained from scratch at a resolution of $1024 \times 1024$ with a patch size of 16. For finetuning, the input image is resized to a 4:3 ratio, i.e. $1024 \times 768$ . We use standard augmentations like cropping, scaling, flipping, and photometric distortions. A random background from non-human COCO [67] images is added for + +
Model#ParamsFLOPsHidden sizeLayersHeadsBatch size
Sapiens-0.3B0.336 B1.242 T1024241698,304
Sapiens-0.6B0.664 B2.583 T1280321665,536
Sapiens-1B1.169 B4.647 T1536402440,960
Sapiens-2B2.163 B8.709 T1920483220,480
+ +segmentation, depth, and normal prediction tasks. Importantly, we use differential learning rates [115] to preserve generalization i.e. lower learning rates for initial layers and progressively higher rates for subsequent layers. The layer-wise learning rate decay is set to 0.85 with a weight decay of 0.1 for the encoder. We detail the design specifications of Sapiens in Table. 2. Following [34, 100], we prioritize scaling models by width rather than depth. Note that the Sapiens-0.3B model, while architecturally similar to the traditional ViT-Large, consists of twentyfold more FLOPs due to its higher resolution. + +# 4.2 2D Pose Estimation + +We finetune Sapiens for face, body, feet, and hand ( $K = 308$ ) pose estimation on our high-fidelity annotations. For training, we use the train set with $1M$ images and for evaluation, we use the test set, named Humans-5K, with $5K$ images. Our evaluation is top-down [111] i.e. we use an off-the-shelf detector [37] for bounding-box and conduct single human pose inference. Table 3 shows a comparison of our models with existing methods for whole-body pose estimation. We evaluate all methods on 114 common keypoints between our 308 keypoint vocabulary and the 133 keypoint vocabulary from COCO-WholeBody [55]. Sapiens-0.6B surpasses the current state-of-the-art, DWPose-1 [114] by +2.8 AP. Contrary to DWPose [114], which utilizes a complex student-teacher framework with feature distillation tailored for the task, Sapiens adopts a general encoder-decoder architecture with large human-centric pretraining. + +Interestingly, even with the same parameter count, our models demonstrate superior performance compared to their counterparts. For instance, Sapiens-0.3B exceeds VitPose+-L by +5.6 AP, and Sapiens-0.6B outperforms VitPose+-H by +7.9 AP. Within the Sapiens family, our results indicate a direct correlation between model size and performance. Sapiens-2B sets a state-of-the-art with 61.1 AP, a significant improvement of +7.6 AP to the prior art. Despite fine-tuning + +Table 2: Sapiens encoder specifications for pretraining on Human-300M dataset. + +
ModelInput SizeBodyFootFaceHandWhole-body
APARAPARAPARAPARAPAR
DeepPose [98]384 × 28832.143.525.341.237.853.915.731.623.937.2
SimpleBaseline [106]384 × 28852.360.149.862.559.667.341.451.844.653.7
HRNet [93]384 × 28855.862.645.255.458.964.539.347.645.753.9
ZoomNAS [110]384 × 28859.766.348.157.974.579.249.860.652.160.7
ViTPose+L [112]256 × 19261.066.862.468.250.155.741.547.347.853.6
ViTPose+-H [112]256 × 19261.667.463.269.050.756.342.047.848.354.1
RTMPose-x [54]384 × 28857.163.755.366.874.478.546.355.051.959.6
DWPose-m [114]256 × 19254.261.449.963.068.574.240.150.047.755.8
DWPose-l [114]384 × 28857.964.256.567.474.378.449.357.453.160.6
Sapiens-0.3B (Ours)1024 × 76858.164.556.867.774.578.649.657.753.4 (+0.3)60.9 (+0.3)
Sapiens-0.6B (Ours)1024 × 76859.865.564.772.375.279.052.160.356.2 (+2.8)62.4 (+2.1)
Sapiens-1B (Ours)1024 × 76862.968.268.375.176.479.755.963.459.4 (+5.9)65.3 (+5.1)
Sapiens-2B (Ours)1024 × 76864.769.969.476.276.979.957.164.461.1(+7.6)67.1(+7.0)
+ +Table 3: Pose estimation results on Humans-5K test set. Flip test is used. + +![](images/59e713e7960beedb77ed778998fe60572b14023b7a02f3e5fd98f3be7656a33b.jpg) +Fig. 6: Pose estimation with Sapiens-1B for 308 keypoints on in-the-wild images. + +![](images/552a641dad9dc63dab04fbb96772b1ed292af4049983426ae4501f0361f0fcdb.jpg) + +![](images/65b6b443723029db3d843fb46d4ae2061f4e00dc7a34c41fdb9cdc36b114f935.jpg) + +![](images/cb9f6c5d47c337b2a123ce6e609f19db0fe9e1720cb51bcc90df1bf918e79bb1.jpg) + +with annotations from a indoor capture studio, Sapiens demonstrate robust generalization to real-world, as shown in Fig. 6. + +# 4.3 Body-Part Segmentation + +We fine-tune and evaluate our annotations with a segmentation vocabulary of 28 classes. Our train set consists of $100K$ images, and the test set, Humans-2K, consists of $2K$ images. We compare Sapiens with existing bodypart segmentation methods fine-tuned on our train set. Importantly, we use suggested pretrained checkpoints by each method as initialization. Similar to pose, we observe generalization to segmentation as shown in Table 4. + +
ModelmIoU(%)mAcc(%)
FCN* [71]48.257.6
SegFormer* [107]53.562.9
Mask2Former* [22]58.768.3
DeepLabV3+* [18]64.174.8
Sapiens-0.3B (Ours)76.786.1
Sapiens-0.6B (Ours)77.886.3
Sapiens-1B (Ours)79.989.1
Sapiens-2B (Ours)81.289.4
+ +Table 4: We report mIoU and mAcc on Humans-2K test set. Methods with * are trained by us. + +Interestingly, our smallest model, Sapiens-0.3B outperforms existing state-of-the-art segmentation methods like Mask2Former [22] and DeepLabV3+ [18] by 12.6 mIoU due to its higher resolution and large human-centric pretraining. Furthermore, increasing the model size improves segmentation performance. Sapiens-2B achieves the best performance of 81.2 mIoU and 89.4 mAcc on the test set. Fig. 7 shows the qualitative results of our models. + +# 4.4 Depth Estimation + +We evaluate our models on THuman2.0 [117] and Hi4D [116] datasets for depth estimation. THuman2.0 consists of 526 high-quality human scans, from which we derive three sets of images for testing: a) face, b) upper body, and c) full body using a virtual camera. THuman2.0 with 1578 images thus enables the evaluation of our models' performance on single-human images across multiple scales. Conversely, the Hi4D dataset focuses on multi-human scenarios, with each sequence showcasing two subjects engaged in activities involving human-human + +![](images/3ff54b7187a6b7e52537224d6211d49ccc4d448604f0c8bd3651a36c6d0e9e36.jpg) +Fig. 7: Human segmentation with Sapiens-1B for 28 categories on in-the-wild images. + +![](images/f7e1f5af38c018e0cc541d44bfda9c2f1ec853e339232a34c218bdb52f849f12.jpg) + +![](images/5eb79fee42cd7ce46d0985b236995a4b28539eab2ac721597fcd1d7e2972b782.jpg) + +![](images/074adc99db64f71394d54eafed3a97513acb0fda084d2275eb4ca8498207cdd8.jpg) + +
MethodTH2.0-FaceTH2.0-UprBodyTH2.0-FullBodyHi4D
RMSE ↓AbsRel ↓δ1 ↑RMSEAbsRelδ1RMSEAbsRelδ1RMSEAbsRelδ1
MiDaS-L [11]0.1140.0970.9250.3980.2710.8680.7010.6890.7820.2610.0820.975
MiDaS-Swin2 [11]0.0500.0360.9950.1220.0810.9480.2920.1710.8620.2090.0630.997
DepthAny-B [113]0.0390.0260.9990.0480.0280.9990.0610.0300.9990.1430.0340.997
DepthAny-L [113]0.0390.0270.9990.0480.0270.9990.0600.0300.9990.1470.0350.997
Sapiens-0.3B (Ours)0.0120.0081.0000.0150.0091.0000.0210.0101.0000.1480.0461.000
Sapiens-0.6B (Ours)0.0110.0081.0000.0150.0091.0000.0210.0101.0000.1420.0441.000
Sapiens-1B (Ours)0.0090.0061.0000.0120.0071.0000.0190.0091.0000.1250.0391.000
Sapiens-2B (Ours)0.0080.0051.0000.0100.0061.0000.0160.0081.0000.1140.0361.000
+ +Table 5: Comparison of Sapiens for monocular depth estimation on human images. + +interactions. We select sequences from pair 28, pair 32, and pair 37, featuring 6 unique subjects from camera 4, totaling 1195 multi-human real images for testing. We follow the relative-depth evaluation protocols established by MiDaSv3.1 [11], reporting standard metrics such as AbsRel and $\delta_{1}$ . In addition, we also report RMSE as our primary metric since $\delta_{1}$ does not effectively reflect performance in human scenes characterized by subtle depth variations. + +Table 5 compares our models with existing state-of-the-art monocular depth estimators. Sapiens-2B, finetuned solely on synthetic data, remarkably outperforms prior art across all single-human scales and multi-human scenarios. We observe a $20\%$ RMSE reduction compared to the top-performing Depth-Anything model on Hi4D images. It is important to highlight that while baseline models are trained on a variety of scenes, Sapiens specializes in human-centric depth estimation. Fig. 8 presents a qualitative comparison of depth estimation between Sapiens-1B and DepthAnything-L. To ensure a fair comparison, the predicted depth is renormalized using the human mask in the baseline visualizations. + +![](images/84ad7805d41928a1b011e4439b9e54b06ba160ab4940c7e5f56dbf429b91251a.jpg) +Fig. 8: We compare our depth prediction with DepthAnything [113]. To showcase the consistency of predicted depth, we also visualize the $\nabla$ depth as pseudo surface normals. + +# 4.5 Surface Normal Estimation + +The datasets for surface normal evaluation are identical to those used for depth estimation. Following [30], we report the mean and median angular error, along with the percentage of pixels within $t^\circ$ error for $t \in \{11.25^\circ, 22.5^\circ, 30^\circ\}$ . Table 6 compares our models with existing human-specific surface normal estimators. All our models outperform existing methods by a significant margin. Sapiens-2B achieves a mean error of around $12^\circ$ on the THuman2.0 (single-human) and Hi4D (multi-human) datasets. We qualitatively compare Sapiens-1B with PIFuHD [89] and ECON [108] for surface normal estimation in Figure 9. Note that PIFuHD [89] is trained with the identical set of 3D scans as ours, and ECON [108] is trained with 4000 scans that are a super set of our 3D scan data. + +
MethodTHuman2.0 [117]Hi4D [116]
Angular Error°% Within t°Angular Error°% Within t°
MeanMedian11.25°22.5°30°MeanMedian11.25°22.5°30°
PIFuHD [89]30.5127.1315.8142.9758.8622.3919.2622.9860.1477.02
HDNet [52]34.8230.6017.4439.2654.5128.6026.8519.0857.9370.14
ICON [109]28.7425.5222.8147.8363.7320.1817.5226.8166.3482.73
ECON [108]25.4523.6732.9555.8669.0318.4616.4729.3568.1284.88
Sapiens-0.3B13.0210.3357.3786.2092.715.0412.2247.0781.4990.70
Sapiens-0.6B12.8610.2357.8586.6893.3014.0611.4750.5984.3792.54
Sapiens-1B12.119.4061.9788.0393.8412.189.5960.3688.6294.44
Sapiens-2B11.849.1663.1688.6094.1812.149.6260.2289.0894.74
+ +Table 6: Comparison of Sapiens for surface normal estimation on human images. + +![](images/d33afce36c657ac7f6dc7a049f860fec3cb6d7f447001294e5e56f018d805460.jpg) +Fig.9: Qualitative comparison of Sapiens-1B with PIFuHD [89] and ECON [108] for monocular surface normal estimation on challenging in-the-wild images. + +# 4.6 Discussion + +Importance of Pretraining Data Source. The feature quality is closely linked to the pretraining data quality. We assess the importance of pretraining on various data sources for human-centric tasks by pretraining Sapiens-0.3B on each dataset under identical training schedules and number of iterations. We fine-tune the model on each task and select early checkpoints for evaluation, reasoning that early-stage fine-tuning better reflects the model's generalization capability. We investigate the impact of pretraining at scale on general images (which may include humans) versus exclusively human images using Sapiens. We randomly select 100 million and 300 million general images from our 1 billion image corpus to create the General-100M and General-300M datasets, respectively. Table 7 showcases the comparison of pretraining outcomes. We report mAP for pose on Humans-5K, mIoU for segmentation on Humans-2K, RMSE for depth on THuman2.0, and mean angular error in degrees for normal estimation on Hi4D. Aligned with findings from [112], our results show that pretraining with Human300M leads to superior performance across all metrics, highlighting the benefits of human-centric pretraining within a fixed computational budget. + +We also study the effect of number of unique human images seen during pretraining with normal estimation performance. We report $\%$ within $30^{\circ}$ . Again, we maintain identical conditions for Sapiens-0.3B pretraining and finetuning. Fig.10 shows a steady improvement in performance as the pretraining data size increases without saturation. In summary, the diversity of human images during pretraining directly correlates with improved generalization to down-stream tasks. + +![](images/f3e5ddc2209cf13d89e34b66ca5793794a923b1006ddb4e00fd8dccb76fd0079.jpg) +Fig. 10: Sapiens-0.3B's normal estimation performance with unique human images seen during pretraining. + +Zero-Shot Generalization. Our models exhibit broad generalization to a variety of settings. For instance, in segmentation, Sapiens are finetuned on single-human im + +ages with limited subject diversity, minimal background variation, and solely third-person views (see Fig. 4). Nevertheless, our large-scale pretraining enables generalization across number of subjects, varying ages, and egocentric views, as shown in Fig. 11. These observations similarly hold for other tasks. + +
Pretraining Source#ImagesPose (↑)Seg(↑)Depth(↓)Normal(↓)
Random Initialization-30.240.30.72035.4
General-100M100M35.750.10.35127.5
General-300M300M37.352.80.34726.8
Humans-100M100M43.661.20.31624.0
Humans-300M (Full)300M47.066.50.28821.8
+ +Table 7: Comparison of Sapiens-0.3B pretrained on various data sources. A domain-specific pretraining yields superior results compared to general data sources. + +![](images/4d32dc066bdc97c77fd4eddfb2d6841af85623dff4abad49e54a0e317551bc00.jpg) +Multi-Human Generalization +Fig. 11: Sapiens achieve broad generalization via large human-centric pretraining. + +![](images/f8b0e003be8dc2dcc8d3db3aeec1cc5f04a5fa88fccb1bd4bdc59fdf69468f49.jpg) +Age Generalization + +![](images/1d80308ea50fc79ee73dafd05abf1e08af2f87c6e660abfac74723543ebf442c.jpg) +Viewpoint Generalization + +Applications. Accurately predicting pose, body-part segmentation, depth and surface normals for in-the-wild human images enables a wide range of applications, such as 3D human scans and the generation of controllable images. We adapt ControlNet [118] to integrate keypoints for faces and hands predicted by Sapiens, and train this modified model on a subset of the Humans-300M dataset, as shown in Fig. 12. This grants more precise control over image generation, producing anatomically consistent human images. + +![](images/c2f1a7606945920b7e60e88c7c902901987a86846deaed5a998c4e02df204c1a.jpg) +Image +Fig. 12: Image synthesis using ControlNet [118] trained on Sapiens pose predictions. + +![](images/8c69392d70601b18d4164db9a9928890eb17689317cd4965cd91a2cdd318ceaa.jpg) +Predicted Pose + +![](images/edf1a9a3e3b1767e500cbe94d93ddb0ffbb7d898a636e1dca733f4481eb12ff5.jpg) +Pose-conditioned Human Image Synthesis + +![](images/b2845bd6118e2beec6fbbd1dc471d7fcbafc3b0bd8e12abaed6ce0f108bedf83.jpg) + +![](images/e0deb4fdb364e6ef015c964610534c3b14288f72b1f80b9ab58f475d273fd521.jpg) + +Limitations. While our models generally perform well, they are not perfect. Human images with complex/rare poses, crowding, and severe occlusion are challenging (see supplemental for details). Although aggressive data augmentation and a detect-and-crop strategy could mitigate these issues, we envision our models as a tool for acquiring large-scale, real-world supervision with human-in-the-loop to develop the next generations of human vision models. + +# 5 Conclusion + +Sapiens represents a significant step toward elevating human-centric vision models into the realm of foundation models. Our models demonstrate strong generalization capabilities on a variety of human-centric tasks. We attribute the state-of-the-art performance of our models to: (i) large-scale pretraining on a large curated dataset, which is specifically tailored to understanding humans, (ii) scaled high-resolution and high-capacity vision transformer backbones, and (iii) high-quality annotations on augmented studio and synthetic data. We believe that these models can become a key building block for a multitude of downstream tasks, and provide access to high-quality vision backbones to a significantly wider part of the community. A potential direction for future work would be extending Sapiens to 3D and multi-modal datasets. + +# References + +1. Abnar, S., Dehghani, M., Neyshabur, B., Sedghi, H.: Exploring the limits of large scale pre-training. arXiv preprint arXiv:2110.02095 (2021) 4 +2. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023) 4 +3. Alldieck, T., Zanfir, M., Sminchisescu, C.: Photorealistic monocular 3d reconstruction of humans wearing clothing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1506-1515 (2022) 4 +4. Bai, Y., Geng, X., Mangalam, K., Bar, A., Yuille, A., Darrell, T., Malik, J., Efros, A.A.: Sequential modeling enables scalable learning for large vision models. arXiv preprint arXiv:2312.00785 (2023) 4 +5. Bao, H., Dong, L., Piao, S., Wei, F.: Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254 (2021) 3 +6. Barron, J.T., Malik, J.: Intrinsic scene properties from a single rgb-d image. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 17-24 (2013) 4 +7. Barron, J.T., Malik, J.: Shape, illumination, and reflectance from shading. IEEE transactions on pattern analysis and machine intelligence 37(8), 1670-1687 (2014) 4 +8. Bartol, K., Bojanic, D., Petkovic, T., Pribanic, T.: A review of body measurement using 3d scanning. IEEE Access 9, 67281-67301 (2021) 4 +9. Bhat, S.F., Alhashim, I., Wonka, P.: Adabins: Depth estimation using adaptive bins. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4009-4018 (2021) 4 +10. Bhat, S.F., Birkl, R., Wofk, D., Wonka, P., Müller, M.: Zoodepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288 (2023) 4 +11. Birkl, R., Wofk, D., Müller, M.: Midas v3. 1-a model zoo for robust monocular relative depth estimation. arXiv preprint arXiv:2307.14460 (2023) 11 +12. Bojic, L.: Metaverse through the prism of power and addiction: what will happen when the virtual world becomes more attractive than reality? European Journal of Futures Research 10(1), 1-24 (2022) 4 +13. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020) 4 +14. Cao, Z., Simon, T., Wei, S., Sheikh, Y.: Realtime multi-person 2d pose estimation using part affinity fields. corr abs/1611.08050. arXiv preprint arXiv:1611.08050 (2016) 1 +15. Cao, Z., Hidalgo, G., Simon, T., Wei, S.E., Sheikh, Y.: Openpose: real-time multi-person 2d pose estimation using part affinity fields. arXiv preprint arXiv:1812.08008 (2018) 3 +16. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021) 3 +17. Chan, C., Ginosar, S., Zhou, T., Efros, A.A.: Everybody dance now. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 5933-5942 (2019) 1 + +18. Chen, L.C., Zhu, Y., Papandreou, G., Schroff, F., Adam, H.: Encoder-decoder with atrous separable convolution for semantic image segmentation. In: Proceedings of the European conference on computer vision (ECCV). pp. 801-818 (2018) 10 +19. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International conference on machine learning. pp. 1597-1607. PMLR (2020) 3 +20. Chen, X., Wang, X., Changpinyo, S., Piergiovanni, A., Padlewski, P., Salz, D., Goodman, S., Grycner, A., Mustafa, B., Beyer, L., et al.: Pali: A jointly-scaled multilingual language-image model. arXiv preprint arXiv:2209.06794 (2022) 5 +21. Chen, Y., Wang, Z., Peng, Y., Zhang, Z., Yu, G., Sun, J.: Cascaded pyramid network for multi-person pose estimation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7103-7112 (2018) 4 +22. Cheng, B., Misra, I., Schwing, A., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. arxiv 2022. arXiv preprint arXiv:2112.01527 (2021) 10 +23. Couprie, C., Farabet, C., Najman, L., LeCun, Y.: Indoor semantic segmentation using depth information. arXiv preprint arXiv:1301.3572 (2013) 3 +24. Dai, Z., Liu, H., Le, Q.V., Tan, M.: Coatnet: Marrying convolution and attention for all data sizes. Advances in neural information processing systems 34, 3965-3977 (2021) 5 +25. Dehghani, M., Djolonga, J., Mustafa, B., Padlewski, P., Heek, J., Gilmer, J., Steiner, A.P., Caron, M., Geirhos, R., Alabdulmohsin, I., et al.: Scaling vision transformers to 22 billion parameters. In: International Conference on Machine Learning. pp. 7480-7512. PMLR (2023) 5 +26. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018) 4 +27. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) 2, 4, 5 +28. Drobyshev, N., Chelishev, J., Khakhulin, T., Ivakhnenko, A., Lempitsky, V., Zakharov, E.: Megaportraits: One-shot megapixel neural head avatars. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 2663-2671 (2022) 1 +29. Du, Y., Kips, R., Pumarola, A., Starke, S., Thabet, A., Sanakoyeu, A.: Avatars grow legs: Generating smooth human motion from sparse tracking inputs with diffusion model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 481-490 (2023) 4 +30. Eftekhar, A., Sax, A., Malik, J., Zamir, A.: Omnidata: A scalable pipeline for making multi-task mid-level vision datasets from 3d scans. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10786-10796 (2021) 12 +31. Eigen, D., Fergus, R.: Predicting depth, surface normals and semantic labels with a common multi-scale convolutional architecture. In: Proceedings of the IEEE international conference on computer vision. pp. 2650-2658 (2015) 4 +32. Eigen, D., Puhrsch, C., Fergus, R.: Depth map prediction from a single image using a multi-scale deep network. Advances in neural information processing systems 27 (2014) 4, 7 + +33. El-Nouby, A., Izacard, G., Touvron, H., Laptev, I., Jegou, H., Grave, E.: Are large-scale datasets necessary for self-supervised pre-training? arXiv preprint arXiv:2112.10740 (2021) 4 +34. El-Nouby, A., Klein, M., Zhai, S., Bautista, M.A., Toshev, A., Shankar, V., Susskind, J.M., Joulin, A.: Scalable pre-training of large autoregressive image models. arXiv preprint arXiv:2401.08541 (2024) 2, 3, 4, 9 +35. Fang, H.S., Lu, G., Fang, X., Xie, J., Tai, Y.W., Lu, C.: Weakly and semi supervised human body part parsing via pose-guided knowledge transfer. arXiv preprint arXiv:1805.04310 (2018) 4 +36. Fang, H.S., Xie, S., Tai, Y.W., Lu, C.: Rmpe: Regional multi-person pose estimation. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 2334-2343 (2017) 4 +37. Fang, Y., Sun, Q., Wang, X., Huang, T., Wang, X., Cao, Y.: Eva-02: A visual representation for neon genesis. arXiv preprint arXiv:2303.11331 (2023) 9 +38. Fang, Y., Wang, W., Xie, B., Sun, Q., Wu, L., Wang, X., Huang, T., Wang, X., Cao, Y.: Eva: Exploring the limits of masked visual representation learning at scale. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19358-19369 (2023) 5 +39. Fouhey, D.F., Gupta, A., Hebert, M.: Data-driven 3d primitives for single image understanding. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 3392-3399 (2013) 4 +40. Gong, K., Liang, X., Li, Y., Chen, Y., Yang, M., Lin, L.: Instance-level human parsing via part grouping network. In: Proceedings of the European conference on computer vision (ECCV). pp. 770-785 (2018) 4 +41. Gong, K., Liang, X., Zhang, D., Shen, X., Lin, L.: Look into person: Self-supervised structure-sensitive learning and a new benchmark for human parsing. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 932-940 (2017) 3, 4, 7 +42. Goyal, P., Caron, M., Lefaudeauux, B., Xu, M., Wang, P., Pai, V., Singh, M., Liptchinsky, V., Misra, I., Joulin, A., et al.: Self-supervised pretraining of visual features in the wild. arXiv preprint arXiv:2103.01988 (2021) 4 +43. Guizilini, V., Vasiljevic, I., Chen, D., Ambrus, R., Gaidon, A.: Towards zero-shot scale-aware monocular depth estimation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9233-9243 (2023) 4 +44. Halstead, M.A., Barsky, B.A., Klein, S.A., Mandell, R.B.: Reconstructing curved surfaces from specular reflection patterns using spline surface fitting of normals. In: Proceedings of the 23rd annual conference on Computer graphics and interactive techniques. pp. 335-342 (1996) 4 +45. He, K., Chen, X., Xie, S., Li, Y., Dollár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022) 3, 4, 5 +46. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729-9738 (2020) 3 +47. He, K., Gkioxari, G., Dollár, P., Girshick, R.B.: Mask r-cnn. corr abs/1703.06870 (2017). arXiv preprint arXiv:1703.06870 (2017) 4 +48. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) 5 + +49. Hoffmann, J., Borgeaud, S., Mensch, A., Buchatskaya, E., Cai, T., Rutherford, E., Casas, D.d.L., Hendricks, L.A., Welbl, J., Clark, A., et al.: Training compute-optimal large language models. arXiv preprint arXiv:2203.15556 (2022) 5 +50. Hu, L., Gao, X., Zhang, P., Sun, K., Zhang, B., Bo, L.: animate anyone: Consistent and controllable image-to-video synthesis for character animation. arXiv preprint arXiv:2311.17117 (2023) 1 +51. Huang, S., Gong, M., Tao, D.: A coarse-fine network for keypoint localization. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 3028-3037 (2017) 4 +52. Jafarian, Y., Park, H.S.: Learning high fidelity depths of dressed humans by watching social media dance videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12753-12762 (2021) 4, 12 +53. Jiang, A.Q., Sablayrolles, A., Mensch, A., Bamford, C., Chaplot, D.S., Casas, D.d.l., Bressand, F., Lengyel, G., Lample, G., Saulnier, L., et al.: Mistral 7b. arXiv preprint arXiv:2310.06825 (2023) 4 +54. Jiang, T., Lu, P., Zhang, L., Ma, N., Han, R., Lyu, C., Li, Y., Chen, K.: Rtmpose: Real-time multi-person pose estimation based on mmpose. arXiv preprint arXiv:2303.07399 (2023) 9 +55. Jin, S., Xu, L., Xu, J., Wang, C., Liu, W., Qian, C., Ouyang, W., Luo, P.: Whole-body human pose estimation in the wild. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IX 16. pp. 196-214. Springer (2020) 3, 6, 9 +56. Kato, N., Li, T., Nishino, K., Uchida, Y.: Improving multi-person pose estimation using label correction. arXiv preprint arXiv:1811.03331 (2018) 3 +57. Khirodkar, R., Chari, V., Agrawal, A., Tyagi, A.: Multi-instance pose networks: Rethinking top-down pose estimation. In: Proceedings of the IEEE/CVF International conference on computer vision. pp. 3122-3131 (2021) 4 +58. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023) 6 +59. Kocabas, M., Chang, J.H.R., Gabriel, J., Tuzel, O., Ranjan, A.: Hugs: Human gaussian splats. arXiv preprint arXiv:2311.17910 (2023) 4 +60. Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., et al.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision 123, 32-73 (2017) 4 +61. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images. arXiv preprint (2009) 4 +62. Ladicky, L., Zeisl, B., Pollefeys, M.: Discriminatively trained dense surface normal estimation. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 468-484. Springer (2014) 4 +63. Lawrence, J., Goldman, D.B., Achar, S., Blascovich, G.M., Desloge, J.G., Fortes, T., Gomez, E.M., Haberling, S., Hoppe, H., Huibers, A., et al.: Project starline: A high-fidelity telepresence system. arXiv preprint (2021) 4 +64. Levoy, M., Pulli, K., Curless, B., Rusinkiewicz, S., Koller, D., Pereira, L., Ginzton, M., Anderson, S., Davis, J., Ginsberg, J., et al.: The digital michelangelo project: 3d scanning of large statues. In: Proceedings of the 27th annual conference on Computer graphics and interactive techniques. pp. 131-144 (2000) 4 +65. Lewkowycz, A.: How to decay your learning rate. arXiv preprint arXiv:2103.12682 (2021) 8 + +66. Li, Z., Wang, X., Liu, X., Jiang, J.: Binsformer: Revisiting adaptive bins for monocular depth estimation. arXiv preprint arXiv:2204.00987 (2022) 4 +67. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: European conference on computer vision. pp. 740-755. Springer (2014) 1, 4, 6, 8 +68. Liu, Z., Hu, H., Lin, Y., Yao, Z., Xie, Z., Wei, Y., Ning, J., Cao, Y., Zhang, Z., Dong, L., et al.: Swin transformer v2: Scaling up capacity and resolution. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12009-12019 (2022) 5 +69. Lombardi, S., Simon, T., Saragih, J., Schwartz, G., Lehrmann, A., Sheikh, Y.: Neural volumes: Learning dynamic renderable volumes from images. arXiv preprint arXiv:1906.07751 (2019) 1, 4 +70. Lombardi, S., Simon, T., Schwartz, G., Zollhoefer, M., Sheikh, Y., Saragih, J.: Mixture of volumetric primitives for efficient neural rendering. ACM Transactions on Graphics (ToG) 40(4), 1-13 (2021) 4 +71. Long, J., Shelhamer, E., Darrell, T., Berkeley, U.: Fully convolutional networks for semantic segmentation. arxiv 2015. arXiv preprint arXiv:1411.4038 (2014) 10 +72. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983 (2016) 8 +73. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017) 8 +74. Lowe, D.G.: Three-dimensional object recognition from single two-dimensional images. Artificial intelligence 31(3), 355-395 (1987) 4 +75. Luo, Y., Zheng, Z., Zheng, L., Guan, T., Yu, J., Yang, Y.: Macro-micro adversarial network for human parsing. In: Proceedings of the European conference on computer vision (ECCV). pp. 418-434 (2018) 4 +76. Ma, S., Simon, T., Saragih, J., Wang, D., Li, Y., De La Torre, F., Sheikh, Y.: Pixel codec avatars. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 64-73 (2021) 4 +77. Mahajan, D., Girshick, R., Ramanathan, V., He, K., Paluri, M., Li, Y., Bharambe, A., Van Der Maaten, L.: Exploring the limits of weakly supervised pretraining. In: Proceedings of the European conference on computer vision (ECCV). pp. 181-196 (2018) 4 +78. Newell, A., Yang, K., Deng, J.: Stacked hourglass networks for human pose estimation. In: European conference on computer vision. pp. 483-499. Springer (2016) 4 +79. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023) 2, 3, 4 +80. Papandreou, G., Zhu, T., Kanazawa, N., Toshev, A., Tompson, J., Bregler, C., Murphy, K.: Towards accurate multi-person pose estimation in the wild. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4903-4911 (2017) 4 +81. Peebles, W., Xie, S.: Scalable diffusion models with transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4195-4205 (2023) 5 +82. render people: 3d people for architectural visualization | renderpeople, https://renderpeople.com/3d-people, accessed: 2024-02-22 3 + +83. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) 4 +84. Ranftl, R., Lasinger, K., Hafner, D., Schindler, K., Koltun, V.: Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence 44(3), 1623-1637 (2020) 4 +85. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022) 4 +86. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al.: Imagenet large scale visual recognition challenge. International journal of computer vision 115, 211-252 (2015) 4 +87. Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022) 4 +88. Saito, S., Huang, Z., Natsume, R., Morishima, S., Kanazawa, A., Li, H.: Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 2304-2314 (2019) 4 +89. Saito, S., Simon, T., Saragih, J., Joo, H.: Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 84-93 (2020) 1, 4, 12 +90. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022) 4 +91. Singh, M., Duval, Q., Alwala, K.V., Fan, H., Aggarwal, V., Adcock, A., Joulin, A., Dollar, P., Feichtenhofer, C., Girshick, R., et al.: The effectiveness of mae pretraining for billion-scale pretraining. arXiv preprint arXiv:2303.13496 (2023) 2, 3, 4 +92. Sun, C., Shrivastava, A., Singh, S., Gupta, A.: Revisiting unreasonable effectiveness of data in deep learning era. In: Proceedings of the IEEE international conference on computer vision. pp. 843-852 (2017) 4 +93. Sun, K., Xiao, B., Liu, D., Wang, J.: Deep high-resolution representation learning for human pose estimation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5693-5703 (2019) 4, 9 +94. Taori, R., Gulrajani, I., Zhang, T., Dubois, Y., Li, X., Guestrin, C., Liang, P., Hashimoto, T.B.: Alpaca: A strong, replicable instruction-following model. Stanford Center for Research on Foundation Models. https://crfm.stanford.edu/2023/03/13/alpaca.html 3(6), 7 (2023) 5 +95. Tay, Y., Dehghani, M., Rao, J., Fedus, W., Abnar, S., Chung, H.W., Narang, S., Yogatama, D., Vaswani, A., Metzler, D.: Scale efficiently: Insights from pretraining and fine-tuning transformers. arXiv preprint arXiv:2109.10686 (2021) 4 + +96. Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023) 4 +97. Thomee, B., Shamma, D.A., Friedland, G., Elizalde, B., Ni, K., Poland, D., Borth, D., Li, L.J.: Yfcc100m: The new data in multimedia research. Communications of the ACM 59(2), 64-73 (2016) 4 +98. Toshev, A., Szegedy, C.: Deepest: Human pose estimation via deep neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1653-1660 (2014) 9 +99. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023) 4 +100. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023) 4, 5, 9 +101. Wang, X., Fouhey, D., Gupta, A.: Designing deep networks for surface normal estimation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 539-547 (2015) 3, 4 +102. Weng, C.Y., Curless, B., Srinivasan, P.P., Barron, J.T., Kemelmacher-Shlizerman, I.: Humannerf: Free-viewpoint rendering of moving people from monocular video. In: Proceedings of the IEEE/CVF conference on computer vision and pattern Recognition. pp. 16210-16220 (2022) 1 +103. Wu, Y., Kirillov, A., Massa, F., Lo, W.Y., Girshick, R.: Detector2. https://github.com/facebookresearch/detectron2 (2019) 5 +104. Xia, F., Wang, P., Chen, L.C., Yuille, A.L.: Zoom better to see clearer: Human and object parsing with hierarchical auto-zoom net. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14. pp. 648-663. Springer (2016) 4 +105. Xia, F., Wang, P., Chen, X., Yuille, A.L.: Joint multi-person pose estimation and semantic part segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6769-6778 (2017) 4 +106. Xiao, B., Wu, H., Wei, Y.: Simple baselines for human pose estimation and tracking. In: Proceedings of the European conference on computer vision (ECCV). pp. 466-481 (2018) 4, 9 +107. Xie, E., Wang, W., Yu, Z., Anandkumar, A., Alvarez, J.M., Luo, P.: Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in neural information processing systems 34, 12077-12090 (2021) 10 +108. Xiu, Y., Yang, J., Cao, X., Tzionas, D., Black, M.J.: Econ: Explicit clothed humans optimized via normal integration. arXiv preprint arXiv:2212.07422 (2022) 1, 4, 12 +109. Xiu, Y., Yang, J., Tzionas, D., Black, M.J.: Icon: Implicit clothed humans obtained from normals. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13286-13296. IEEE (2022) 1, 12 +110. Xu, L., Jin, S., Liu, W., Qian, C., Ouyang, W., Luo, P., Wang, X.: Zoomnas: searching for whole-body human pose estimation in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(4), 5296-5313 (2022) 9 +111. Xu, Y., Zhang, J., Zhang, Q., Tao, D.: Vitpose: Simple vision transformer baselines for human pose estimation. Advances in Neural Information Processing Systems 35, 38571-38584 (2022) 6, 9 + +112. Xu, Y., Zhang, J., Zhang, Q., Tao, D.: Vitpose+: Vision transformer foundation model for generic body pose estimation. arXiv preprint arXiv:2212.04246 (2022) 9, 13 +113. Yang, L., Kang, B., Huang, Z., Xu, X., Feng, J., Zhao, H.: Depth anything: Unleashing the power of large-scale unlabeled data. arXiv preprint arXiv:2401.10891 (2024) 1, 3, 4, 11 +114. Yang, Z., Zeng, A., Yuan, C., Li, Y.: Effective whole-body pose estimation with two-stages distillation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4210-4220 (2023) 9 +115. Yang, Z., Dai, Z., Yang, Y., Carbonell, J., Salakhutdinov, R.R., Le, Q.V.: Xlnet: Generalized autoregressive pretraining for language understanding. Advances in neural information processing systems 32 (2019) 9 +116. Yin, Y., Guo, C., Kaufmann, M., Zarate, J.J., Song, J., Hilliges, O.: Hi4d: 4d instance segmentation of close human interaction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17016-17027 (2023) 10, 12 +117. Yu, T., Zheng, Z., Guo, K., Liu, P., Dai, Q., Liu, Y.: Function4d: Real-time human volumetric capture from very sparse consumer rgbd sensors. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5746-5756 (2021) 10, 12 +118. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3836-3847 (2023) 1, 14 +119. Zhang, S.H., Li, R., Dong, X., Rosin, P., Cai, Z., Han, X., Yang, D., Huang, H., Hu, S.M.: Pose2seg: Detection free human instance segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 889-898 (2019) 1, 3 +120. Zhang, X., Lu, Y., Wang, W., Yan, A., Yan, J., Qin, L., Wang, H., Yan, X., Wang, W.Y., Petzold, L.R.: Gpt-4v (isdiction) as a generalist evaluator for vision-language tasks. arXiv preprint arXiv:2311.01361 (2023) 4 +121. Zhou, J., Wei, C., Wang, H., Shen, W., Xie, C., Yuille, A., Kong, T.: ibot: Image bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832 (2021) 3, 4 +122. Zlateski, A., Jaroensri, R., Sharma, P., Durand, F.: On the importance of label quality for semantic segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1479-1487 (2018) 3 \ No newline at end of file diff --git a/2024/Sapiens_ Foundation for Human Vision Models/images.zip b/2024/Sapiens_ Foundation for Human Vision Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..746989a1075972bb05b871d0d0cd37eec302600a --- /dev/null +++ b/2024/Sapiens_ Foundation for Human Vision Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:da149726ed0cb0b47e377b330e794daded5d8c30d6c92d6cedb22d6de24d8fb6 +size 1054985 diff --git a/2024/Sapiens_ Foundation for Human Vision Models/layout.json b/2024/Sapiens_ Foundation for Human Vision Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c5835eeb53c3077c2e0ca95e83f16fa8395867a6 --- /dev/null +++ b/2024/Sapiens_ Foundation for Human Vision Models/layout.json @@ -0,0 +1,12272 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 143, + 111, + 470, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 111, + 470, + 129 + ], + "spans": [ + { + "bbox": [ + 143, + 111, + 470, + 129 + ], + "type": "text", + "content": "Sapiens: Foundation for Human Vision Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 140, + 150, + 474, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 150, + 474, + 174 + ], + "spans": [ + { + "bbox": [ + 140, + 150, + 474, + 174 + ], + "type": "text", + "content": "Rawal Khirodkar, Timur Bagautdinov, Julieta Martinez, Su Zhaoen, Austin James, Peter Selednik, Stuart Anderson, Shunsuke Saito" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 253, + 184, + 360, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 184, + 360, + 194 + ], + "spans": [ + { + "bbox": [ + 253, + 184, + 360, + 194 + ], + "type": "text", + "content": "Codec Avatars Lab, Meta" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 205, + 452, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 205, + 452, + 402 + ], + "spans": [ + { + "bbox": [ + 160, + 205, + 452, + 402 + ], + "type": "text", + "content": "Abstract. We present Sapiens, a family of models for four fundamental human-centric vision tasks - 2D pose estimation, body-part segmentation, depth estimation, and surface normal prediction. Our models natively support 1K high-resolution inference and are extremely easy to adapt for individual tasks by simply fine-tuning foundation models pretrained on over 300 million in-the-wild human images. We observe that, given the same computational budget, self-supervised pretraining on a curated dataset of human images significantly boosts the performance for a diverse set of human-centric tasks. The resulting models exhibit remarkable generalization to in-the-wild data, even when labeled data is scarce or entirely synthetic. Our simple model design also brings scalability - model performance across tasks significantly improves as we scale the number of parameters from 0.3 to 2 billion. Sapiens consistently surpasses existing complex baselines across various human-centric benchmarks. Specifically, we achieve significant improvements over the prior state-of-the-art on Humans-5K (pose) by " + }, + { + "bbox": [ + 160, + 205, + 452, + 402 + ], + "type": "inline_equation", + "content": "7.6\\mathrm{mAP}" + }, + { + "bbox": [ + 160, + 205, + 452, + 402 + ], + "type": "text", + "content": ", Humans-2K (part-seg) by " + }, + { + "bbox": [ + 160, + 205, + 452, + 402 + ], + "type": "inline_equation", + "content": "17.1\\mathrm{mIoU}" + }, + { + "bbox": [ + 160, + 205, + 452, + 402 + ], + "type": "text", + "content": ", Hi4D (depth) by " + }, + { + "bbox": [ + 160, + 205, + 452, + 402 + ], + "type": "inline_equation", + "content": "22.4\\%" + }, + { + "bbox": [ + 160, + 205, + 452, + 402 + ], + "type": "text", + "content": " relative RMSE, and THuman2 (normal) by " + }, + { + "bbox": [ + 160, + 205, + 452, + 402 + ], + "type": "inline_equation", + "content": "53.5\\%" + }, + { + "bbox": [ + 160, + 205, + 452, + 402 + ], + "type": "text", + "content": " relative angular error." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 406, + 424, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 406, + 424, + 417 + ], + "spans": [ + { + "bbox": [ + 160, + 406, + 424, + 417 + ], + "type": "text", + "content": "Keywords: Human-centric, Large-scale pretraining, In-the-wild" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 179, + 427, + 432, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 427, + 432, + 439 + ], + "spans": [ + { + "bbox": [ + 179, + 427, + 432, + 439 + ], + "type": "text", + "content": "\"Sapiens—pertaining to, or resembling modern humans.\"" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 455, + 230, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 455, + 230, + 467 + ], + "spans": [ + { + "bbox": [ + 132, + 455, + 230, + 467 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 474, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 593 + ], + "type": "text", + "content": "Recent years have witnessed remarkable strides towards generating photorealistic humans in 2D [17, 28, 50, 118] and 3D [69, 89, 102, 109]. The success of these methods is greatly attributed to the robust estimation of various assets such as 2D keypoints [14,67], fine-grained body-part segmentation [119], depth [113], and surface normals [89, 108]. However, robust and accurate estimation of these assets is still an active research area, and complicated systems to boost performance for individual tasks often hinder wider adoption. Moreover, obtaining accurate ground-truth annotation in-the-wild is notoriously difficult to scale. Our goal is to provide a unified framework and models to infer these assets in-the-wild to unlock a wide range of human-centric applications for everybody." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "We argue that such human-centric models should satisfy three criteria: generalization, broad applicability, and high fidelity. Generalization ensures robustness to unseen conditions, enabling the model to perform consistently across varied environments. Broad applicability indicates the versatility of the model, making it suitable for a wide range of tasks with minimal modifications. High fidelity denotes the ability of the model to produce precise, high-resolution outputs," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 114, + 462, + 426 + ], + "blocks": [ + { + "bbox": [ + 149, + 114, + 462, + 426 + ], + "lines": [ + { + "bbox": [ + 149, + 114, + 462, + 426 + ], + "spans": [ + { + "bbox": [ + 149, + 114, + 462, + 426 + ], + "type": "image", + "image_path": "5e2c5b544eea5e6c23ea0772008ae77fed6efeed3b8f0adaac0338426b159b3e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 429, + 482, + 462 + ], + "lines": [ + { + "bbox": [ + 130, + 429, + 482, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 429, + 482, + 462 + ], + "type": "text", + "content": "Fig. 1: Sapiens models are finetuned for four human tasks - 2D pose estimation, body-part segmentation, depth prediction and normal prediction. Our models generalize across a variety of in-the-wild face, upper-body, full-body and multi-person images." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 468, + 482, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 482, + 504 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 482, + 504 + ], + "type": "text", + "content": "essential for faithful human generation tasks such as 2D to 3D lifting. This paper details the development of models that embody these attributes, collectively referred to as Sapiens." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "Following the insights from [34, 79, 91], leveraging large datasets and scalable model architectures is key for generalization. For broader applicability, we adopt the pretrain-then-finetune approach, enabling post-pretraining adaptation to specific tasks with minimal adjustments. This approach raises a critical question: What type of data is most effective for pretraining? Given computational limits, should the emphasis be on collecting as many human images as possible, or is it preferable to pretrain on a less curated set to better reflect real-world variability? Existing methods often overlook the pretraining data distribution in the context of downstream tasks. To study the influence of pretraining data distribution on human-specific tasks, we collect the Humans-300M dataset, featuring 300 million diverse human images. These unlabelled images are used to pretrain a family of vision transformers [27] from scratch, with parameter counts ranging from 300M to 2B." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 245, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 245, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 245, + 101 + ], + "type": "text", + "content": "R. Khirodkar et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 488, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 488, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 488, + 295 + ], + "type": "text", + "content": "Among various self-supervision methods for learning general-purpose visual features from large datasets [5,19,34,45,46,121], we choose the masked-autoencoder (MAE) approach [45] for its simplicity and efficiency in pretraining. MAE, having a single-pass inference model compared to contrastive or multi-inference strategies, allows processing a larger volume of images with the same computational resources. For higher-fidelity, in contrast to prior methods, we increase the native input resolution of our pretraining to 1024 pixels, resulting in a " + }, + { + "bbox": [ + 130, + 116, + 488, + 295 + ], + "type": "inline_equation", + "content": "\\sim 4\\times" + }, + { + "bbox": [ + 130, + 116, + 488, + 295 + ], + "type": "text", + "content": " increase in FLOPs compared to the largest existing vision backbone [91]. Each model is pretrained on 1.2 trillion tokens. Table 1 outlines a comparison with earlier approaches. For finetuning on human-centric tasks [15, 101, 113, 119], we use a consistent encoder-decoder architecture. The encoder is initialized with weights from pretraining, while the decoder, a lightweight and task-specific head, is initialized randomly. Both components are then finetuned in an end-to-end manner. We focus on four key tasks - 2D pose estimation, body-part segmentation, depth, and normal estimation, as shown in Fig. 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 296, + 483, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 296, + 483, + 488 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 483, + 488 + ], + "type": "text", + "content": "Consistently with prior studies [56,122], we affirm the critical impact of label quality on the model's in-the-wild performance. Public benchmarks [23,41,55] often contain noisy labels, providing inconsistent supervisory signals during model fine-tuning. At the same time, it is important to utilize fine-grained and precise annotations to align closely with our primary goal of 3D human digitization. To this end, we propose a substantially denser set of 2D whole body keypoints for pose estimation and a detailed class vocabulary for body part segmentation, surpassing the scope of previous datasets (please refer to Fig. 1). Specifically, we introduce a comprehensive collection of 308 keypoints encompassing the body, hands, feet, surface, and face. Additionally, we expand the segmentation class vocabulary to 28 classes, covering body parts such as the hair, tongue, teeth, upper/lower lip, and torso. To guarantee the quality and consistency of annotations and a high degree of automation, we utilize a multi-view capture setup to collect pose and segmentation annotations. We also utilize human-centric synthetic data for depth and normal estimation, leveraging 600 detailed scans from RenderPeople [82] to generate high-resolution depth maps and surface normals." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 487, + 482, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 482, + 547 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 482, + 547 + ], + "type": "text", + "content": "We show that the combination of domain-specific large-scale pretraining with limited, yet high-quality annotations leads to robust in-the-wild generalization. Overall, our method demonstrates an effective strategy for developing highly precise discriminative models capable of performing in real-world scenarios without the need for collecting a costly and diverse set of annotations." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 174, + 550, + 442, + 630 + ], + "blocks": [ + { + "bbox": [ + 174, + 550, + 442, + 630 + ], + "lines": [ + { + "bbox": [ + 174, + 550, + 442, + 630 + ], + "spans": [ + { + "bbox": [ + 174, + 550, + 442, + 630 + ], + "type": "table", + "html": "
MethodDataset#ParamsGFLOPsImage sizeDomain
DINO [16]ImageNet1k86 M17.6224General
iBOT [121]ImageNet21k307 M61.6224General
DINOv2 [79]LVD-142M1 B291.0224General
ViT-6.5B [91]IG-3B6.5 B1657.0224General
AIM [34]DFN-2B6.5 B1657.0224General
Sapiens (Ours)Humans-300M2 B8709.01024Human
", + "image_path": "8fe4c204156bfeef8bee9e13ce3c5462d871ca07f7a76c0827654a829885f249.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 634, + 482, + 657 + ], + "lines": [ + { + "bbox": [ + 130, + 634, + 482, + 657 + ], + "spans": [ + { + "bbox": [ + 130, + 634, + 482, + 657 + ], + "type": "text", + "content": "Table 1: Comparison of state-of-the-art pretrained vision models. Sapiens adopts a higher resolution backbone on a large dataset of in-the-wild human images." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 415, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 415, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 415, + 91, + 447, + 102 + ], + "type": "text", + "content": "Sapiens" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 146, + 116, + 348, + 127 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 116, + 348, + 127 + ], + "spans": [ + { + "bbox": [ + 146, + 116, + 348, + 127 + ], + "type": "text", + "content": "Our contributions are summarized as follows." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 132, + 484, + 251 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 137, + 132, + 479, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 132, + 479, + 155 + ], + "spans": [ + { + "bbox": [ + 137, + 132, + 479, + 155 + ], + "type": "text", + "content": "- We introduce Sapiens, a family of vision transformers pretrained on a large-scale, curated dataset of diverse human images." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 156, + 480, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 156, + 480, + 179 + ], + "spans": [ + { + "bbox": [ + 138, + 156, + 480, + 179 + ], + "type": "text", + "content": "- This study shows that simple data curation and large-scale pretraining significantly boost the model's performance with the same computational budget." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 180, + 479, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 180, + 479, + 203 + ], + "spans": [ + { + "bbox": [ + 138, + 180, + 479, + 203 + ], + "type": "text", + "content": "- Our models, fine-tuned with high-quality annotations or even synthetic data, demonstrate robust in-the-wild generalization." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 204, + 484, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 204, + 484, + 251 + ], + "spans": [ + { + "bbox": [ + 138, + 204, + 484, + 251 + ], + "type": "text", + "content": "- The first 1K high-resolution model that natively supports high-fidelity inference essential for human-centric tasks, achieving state-of-the-art performance on benchmarks for 2D pose estimation, body-part segmentation, depth, and normal estimation." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 270, + 237, + 282 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 237, + 282 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 237, + 282 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 295, + 480, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 295, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 295, + 480, + 331 + ], + "type": "text", + "content": "Our work explores the limits of training large architectures on a large number of in-the-wild human images. We build on prior work from different areas: pretraining at scale, human vision tasks, and large vision transformers." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 336, + 482, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 336, + 482, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 336, + 482, + 540 + ], + "type": "text", + "content": "Pretraining at Scale. The remarkable success of large-scale pretraining [26,95] followed by task-specific finetuning for language modeling [2,13,53,96,99,100] has established this approach as a standard practice. Similarly, computer vision methods [1,4,33,34,42,79,83,85,87,120] are progressively embracing extensive data scales for pretraining. The emergence of large datasets, such as LAION-5B [90], Instagram-3.5B [77], JFT-300M [92], LVD-142M [79], Visual Genome [60], and YFCC100M [97], has enabled the exploration of a data corpus well beyond the scope of traditional benchmarks [61,67,86]. Salient work in this domain includes DINOv2 [79], MAWS [91], and AIM [34]. DINOv2 achieves state-of-the-art performance in generating self-supervised features by scaling the contrastive iBot [121] method on the LDV-142M dataset [79]. MAWS [91] studies the scaling of masked-autoencoders (MAE) [45] on billion images. AIM [34] explores the scalability of autoregressive visual pretraining similar to BERT [26] for vision transformers [27]. In contrast to these methods which mainly focus on general image pretraining or zero-shot image classification, we take a distinctly human-centric approach: our models leverage a vast collection of human images for pretraining, subsequently fine-tuning for a range of human-related tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Human Vision Tasks. The pursuit of large-scale 3D human digitization [8, 44, 64, 74] remains a pivotal goal in computer vision [12]. Significant progress has been made within controlled or studio environments [3, 59, 63, 69, 70, 76, 89], yet challenges persist in extending these methods to unconstrained environments [29]. To address these challenges, developing versatile models capable of multiple fundamental tasks such as keypoint estimation [21, 36, 47, 51, 57, 78, 80, 93, 106], body-part segmentation [35, 40, 40, 41, 75, 104, 105], depth estimation [9, 10, 32, 43, 52, 66, 84, 113], and surface normal prediction [6, 7, 31, 39, 62, 88, 101, 108] from images in natural settings is crucial. In this work, we aim to develop models for these essential human vision tasks which generalize to in-the-wild settings." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 245, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 245, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 245, + 101 + ], + "type": "text", + "content": "R. Khirodkar et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 283 + ], + "type": "text", + "content": "Scaling Architectures. Currently, the largest publicly-accessible language models contain upwards of 100B parameters [49], while the more commonly used language models [94, 100] contain around 7B parameters. In contrast, Vision Transformers (ViT) [27], despite sharing a similar architecture, have not been scaled to this extent successfully. While there are notable endeavors in this direction, including the development of a dense ViT-4B [20] trained on both text and images, and the formulation of techniques for the stable training of a ViT-22B [25], commonly utilized vision backbones still range between 300M to 600M parameters [24, 38, 48, 68] and are primarily pretrained at an image resolution of about 224 pixels. Similarly, existing transformer-based image generation models, such as DiT [81] use less than 700M parameters, and operate on a highly compressed latent space. To address this gap, we introduce Sapiens - a collection of large, high-resolution ViT models that are pretrained natively at a 1024 pixel image resolution on millions of human images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 290, + 202, + 303 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 290, + 202, + 303 + ], + "spans": [ + { + "bbox": [ + 131, + 290, + 202, + 303 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 307, + 277, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 307, + 277, + 319 + ], + "spans": [ + { + "bbox": [ + 131, + 307, + 277, + 319 + ], + "type": "text", + "content": "3.1 Humans-300M Dataset" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 326, + 371, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 326, + 371, + 458 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 371, + 458 + ], + "type": "text", + "content": "We utilize a large proprietary dataset for pretraining of approximately 1 billion in-the-wild images, focusing exclusively on human images. The preprocessing involves discarding images with watermarks, text, artistic depictions, or unnatural elements. Subsequently, we use an off-the-shelf person bounding-box detector [103] to filter images, retaining those with a detection score above 0.9 and bounding box dimensions exceeding 300 pixels. Fig. 2 provides an overview of the distribution of the number of people per image in over 248 million images contain multiple subjects." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 376, + 278, + 481, + 392 + ], + "blocks": [ + { + "bbox": [ + 376, + 278, + 481, + 392 + ], + "lines": [ + { + "bbox": [ + 376, + 278, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 376, + 278, + 481, + 392 + ], + "type": "image", + "image_path": "453feb69b9add65076d75c0c6acdf971ef91e4808b045636a3f341b277320dac.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 373, + 399, + 482, + 424 + ], + "lines": [ + { + "bbox": [ + 373, + 399, + 482, + 424 + ], + "spans": [ + { + "bbox": [ + 373, + 399, + 482, + 424 + ], + "type": "text", + "content": "Fig.2: Overview of number of humans per image in the Humans-300M dataset." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 467, + 219, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 467, + 219, + 479 + ], + "spans": [ + { + "bbox": [ + 131, + 467, + 219, + 479 + ], + "type": "text", + "content": "3.2 Pretraining" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": "We follow the masked-autoencoder [45] (MAE) approach for pretraining. Our model is trained to reconstruct the original human image given its partial observation. Like all autoencoders, our model has an encoder that maps the visible image to a latent representation and a decoder that reconstructs the original image from this latent representation. Our pretraining dataset consists of both single and multi-human images; each image is resized to a fixed size with a square aspect ratio. Similar to ViT [27], we divide an image into regular non-overlapping patches with a fixed patch size. A subset of these patches is randomly selected and masked, leaving the rest visible. The proportion of masked patches to visible ones is defined as the masking ratio, which remains fixed throughout training. We refer to MAE [45] for more details. Fig. 3 (Top) shows the reconstruction of our pretrained model on unseen human images. Our models exhibit generalization across a variety of image characteristics including scales, crops, the age and ethnicity of subjects, and number of subjects. Each patch token in our model accounts for " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "0.02\\%" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": " of the image area compared to " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "0.4\\%" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": " in standard ViTs, a " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "16\\times" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 415, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 415, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 415, + 91, + 447, + 102 + ], + "type": "text", + "content": "Sapiens" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 327 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 327 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 327 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 327 + ], + "type": "image", + "image_path": "14bb481e41f6e40c447e02a883aadf2406e7016367fd43a989616fca2253ada2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 336, + 480, + 373 + ], + "lines": [ + { + "bbox": [ + 130, + 336, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 336, + 480, + 373 + ], + "type": "text", + "content": "reduction - this provides a fine-grained inter-token reasoning for our models. Fig.3 (Bottom) shows that even with an increased mask ratio of " + }, + { + "bbox": [ + 130, + 336, + 480, + 373 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 130, + 336, + 480, + 373 + ], + "type": "text", + "content": ", our model achieves a plausible reconstruction of human anatomy on held-out samples." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 382, + 261, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 382, + 261, + 393 + ], + "spans": [ + { + "bbox": [ + 132, + 382, + 261, + 393 + ], + "type": "text", + "content": "3.3 2D Pose Estimation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 402, + 482, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 569 + ], + "type": "text", + "content": "We follow the top-down 2D pose estimation paradigm, which aims to detect the locations of " + }, + { + "bbox": [ + 130, + 402, + 482, + 569 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 402, + 482, + 569 + ], + "type": "text", + "content": " keypoints from an input image " + }, + { + "bbox": [ + 130, + 402, + 482, + 569 + ], + "type": "inline_equation", + "content": "\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 130, + 402, + 482, + 569 + ], + "type": "text", + "content": ". Most methods pose this problem as heatmap prediction, where each of " + }, + { + "bbox": [ + 130, + 402, + 482, + 569 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 402, + 482, + 569 + ], + "type": "text", + "content": " heatmaps represents the probability of the corresponding keypoint being at any spatial location. Similar to [111], we define a pose estimation transformer, " + }, + { + "bbox": [ + 130, + 402, + 482, + 569 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 130, + 402, + 482, + 569 + ], + "type": "text", + "content": ", for keypoint detection. The bounding box at training and inference is scaled to " + }, + { + "bbox": [ + 130, + 402, + 482, + 569 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 130, + 402, + 482, + 569 + ], + "type": "text", + "content": " and is provided as an input to " + }, + { + "bbox": [ + 130, + 402, + 482, + 569 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 130, + 402, + 482, + 569 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 130, + 402, + 482, + 569 + ], + "type": "inline_equation", + "content": "\\mathbf{y} \\in \\mathbb{R}^{H \\times W \\times K}" + }, + { + "bbox": [ + 130, + 402, + 482, + 569 + ], + "type": "text", + "content": " denote the " + }, + { + "bbox": [ + 130, + 402, + 482, + 569 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 402, + 482, + 569 + ], + "type": "text", + "content": " heatmaps corresponding to the ground truth keypoints for a given input " + }, + { + "bbox": [ + 130, + 402, + 482, + 569 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 130, + 402, + 482, + 569 + ], + "type": "text", + "content": ". The pose estimator transforms input " + }, + { + "bbox": [ + 130, + 402, + 482, + 569 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 130, + 402, + 482, + 569 + ], + "type": "text", + "content": " to a set of predicted heatmaps, " + }, + { + "bbox": [ + 130, + 402, + 482, + 569 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}} \\in \\mathbb{R}^{H \\times W \\times K}" + }, + { + "bbox": [ + 130, + 402, + 482, + 569 + ], + "type": "text", + "content": ", such that " + }, + { + "bbox": [ + 130, + 402, + 482, + 569 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}} = \\mathcal{P}(\\mathbf{I})" + }, + { + "bbox": [ + 130, + 402, + 482, + 569 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 402, + 482, + 569 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 130, + 402, + 482, + 569 + ], + "type": "text", + "content": " is trained to minimize the mean squared loss " + }, + { + "bbox": [ + 130, + 402, + 482, + 569 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{pose}} = \\mathrm{MSE}(\\mathbf{y},\\hat{\\mathbf{y}})" + }, + { + "bbox": [ + 130, + 402, + 482, + 569 + ], + "type": "text", + "content": ". During finetuning, the encoder of " + }, + { + "bbox": [ + 130, + 402, + 482, + 569 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 130, + 402, + 482, + 569 + ], + "type": "text", + "content": " is initialized with the weights from pretraining, and the decoder is initialized randomly. The aspect ratio " + }, + { + "bbox": [ + 130, + 402, + 482, + 569 + ], + "type": "inline_equation", + "content": "H:W" + }, + { + "bbox": [ + 130, + 402, + 482, + 569 + ], + "type": "text", + "content": " is set to be 4:3, with the pretrained positional embedding being interpolated accordingly [58]. We use lightweight decoders with deconvolution and convolution operations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "We finetune both the encoder and the decoder in " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " across multiple skeleton formats, including " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "K = 17" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " [67], " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "K = 133" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " [55] and a new highly-detailed skeleton format, with " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "K = 308" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": ", as shown in Fig. 4 (Left). Compared to existing formats with at most 68 facial keypoints, our annotations consist of 243 facial keypoints, including representative points around the eyes, eyebrows, lips, nose, and ears. This design is tailored to meticulously capture the nuanced details of facial expressions in the real world. With these keypoints, we manually annotated 1 million images at " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "4K" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " resolution from an indoor capture setup." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 244, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 244, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 244, + 101 + ], + "type": "text", + "content": "R. Khirodkar et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 115, + 236, + 213 + ], + "blocks": [ + { + "bbox": [ + 136, + 115, + 236, + 213 + ], + "lines": [ + { + "bbox": [ + 136, + 115, + 236, + 213 + ], + "spans": [ + { + "bbox": [ + 136, + 115, + 236, + 213 + ], + "type": "image", + "image_path": "c6937e5398dab721268539562b82ae78971bfcef323120b0a3eae4f1a07f8f1f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 156, + 213, + 213, + 221 + ], + "lines": [ + { + "bbox": [ + 156, + 213, + 213, + 221 + ], + "spans": [ + { + "bbox": [ + 156, + 213, + 213, + 221 + ], + "type": "text", + "content": "a) Full-Body - 308 kps" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 236, + 213, + 280, + 221 + ], + "lines": [ + { + "bbox": [ + 236, + 213, + 280, + 221 + ], + "spans": [ + { + "bbox": [ + 236, + 213, + 280, + 221 + ], + "type": "text", + "content": "c) Face - 243 kps" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 230, + 479, + 242 + ], + "lines": [ + { + "bbox": [ + 131, + 230, + 479, + 242 + ], + "spans": [ + { + "bbox": [ + 131, + 230, + 479, + 242 + ], + "type": "text", + "content": "Fig. 4: Ground-truth annotations for 2D pose estimation and body-part segmentation." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 236, + 116, + 285, + 152 + ], + "blocks": [ + { + "bbox": [ + 236, + 116, + 285, + 152 + ], + "lines": [ + { + "bbox": [ + 236, + 116, + 285, + 152 + ], + "spans": [ + { + "bbox": [ + 236, + 116, + 285, + 152 + ], + "type": "image", + "image_path": "c4a8aa5300b95bf6610f839ee3521a86c8b6179f514319a51b85afc070ae7e33.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 238, + 152, + 283, + 159 + ], + "lines": [ + { + "bbox": [ + 238, + 152, + 283, + 159 + ], + "spans": [ + { + "bbox": [ + 238, + 152, + 283, + 159 + ], + "type": "text", + "content": "b) Hands - 40 kps" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 239, + 163, + 282, + 212 + ], + "blocks": [ + { + "bbox": [ + 239, + 163, + 282, + 212 + ], + "lines": [ + { + "bbox": [ + 239, + 163, + 282, + 212 + ], + "spans": [ + { + "bbox": [ + 239, + 163, + 282, + 212 + ], + "type": "image", + "image_path": "b823aadf68f180e5b7cf54bb07d58404e4f4c8c5b7f11f80d2c0ca12e2aa81f2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 293, + 116, + 346, + 213 + ], + "blocks": [ + { + "bbox": [ + 293, + 116, + 346, + 213 + ], + "lines": [ + { + "bbox": [ + 293, + 116, + 346, + 213 + ], + "spans": [ + { + "bbox": [ + 293, + 116, + 346, + 213 + ], + "type": "image", + "image_path": "be8e11fb02233f3669afa36ef6ae00a09a051c8df9cfac16407dd82738fe6cd0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 347, + 117, + 398, + 213 + ], + "blocks": [ + { + "bbox": [ + 347, + 117, + 398, + 213 + ], + "lines": [ + { + "bbox": [ + 347, + 117, + 398, + 213 + ], + "spans": [ + { + "bbox": [ + 347, + 117, + 398, + 213 + ], + "type": "image", + "image_path": "2887a149f76bb9d884a9c99216ce4ea53cababd52aa66711d107bfbd8dca9ea6.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 348, + 213, + 447, + 221 + ], + "lines": [ + { + "bbox": [ + 348, + 213, + 447, + 221 + ], + "spans": [ + { + "bbox": [ + 348, + 213, + 447, + 221 + ], + "type": "text", + "content": "d) Body-Part Segmentation: 28 Classes" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 399, + 118, + 480, + 213 + ], + "blocks": [ + { + "bbox": [ + 399, + 118, + 480, + 213 + ], + "lines": [ + { + "bbox": [ + 399, + 118, + 480, + 213 + ], + "spans": [ + { + "bbox": [ + 399, + 118, + 480, + 213 + ], + "type": "image", + "image_path": "90b238aa8bb10f6e41a8241bb013e9c7fd3c45915dcb9f69fd1208c0e5c8e372.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 251, + 285, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 251, + 285, + 262 + ], + "spans": [ + { + "bbox": [ + 131, + 251, + 285, + 262 + ], + "type": "text", + "content": "3.4 Body-Part Segmentation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 270, + 480, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 480, + 354 + ], + "type": "text", + "content": "Commonly referred to as human parsing, body-part segmentation aims to classify pixels in the input image " + }, + { + "bbox": [ + 130, + 270, + 480, + 354 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 130, + 270, + 480, + 354 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 130, + 270, + 480, + 354 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 270, + 480, + 354 + ], + "type": "text", + "content": " classes. Most methods [41] transform this problem to estimating per-pixel class probabilities to create a probability map " + }, + { + "bbox": [ + 130, + 270, + 480, + 354 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{p}}\\in \\mathbb{R}^{H\\times W\\times C}" + }, + { + "bbox": [ + 130, + 270, + 480, + 354 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 130, + 270, + 480, + 354 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{p}} = S(\\mathbf{I})" + }, + { + "bbox": [ + 130, + 270, + 480, + 354 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 270, + 480, + 354 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 270, + 480, + 354 + ], + "type": "text", + "content": " is the segmentation model. As outlined previously, we adopt the same encoder-decoder architecture and initialization scheme for " + }, + { + "bbox": [ + 130, + 270, + 480, + 354 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 270, + 480, + 354 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 270, + 480, + 354 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 270, + 480, + 354 + ], + "type": "text", + "content": " is finetuned to minimize the weighted cross-entropy loss between the actual " + }, + { + "bbox": [ + 130, + 270, + 480, + 354 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 130, + 270, + 480, + 354 + ], + "type": "text", + "content": " and predicted " + }, + { + "bbox": [ + 130, + 270, + 480, + 354 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{p}}" + }, + { + "bbox": [ + 130, + 270, + 480, + 354 + ], + "type": "text", + "content": " probability maps, " + }, + { + "bbox": [ + 130, + 270, + 480, + 354 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{seg}} = \\mathrm{WeightedCE}(\\mathbf{p},\\hat{\\mathbf{p}})" + }, + { + "bbox": [ + 130, + 270, + 480, + 354 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 354, + 480, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 480, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 480, + 437 + ], + "type": "text", + "content": "We finetune " + }, + { + "bbox": [ + 130, + 354, + 480, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 130, + 354, + 480, + 437 + ], + "type": "text", + "content": " across two part-segmentation vocabularies: a standard set with " + }, + { + "bbox": [ + 130, + 354, + 480, + 437 + ], + "type": "inline_equation", + "content": "C = 20" + }, + { + "bbox": [ + 130, + 354, + 480, + 437 + ], + "type": "text", + "content": " [41] and a new larger vocabulary with " + }, + { + "bbox": [ + 130, + 354, + 480, + 437 + ], + "type": "inline_equation", + "content": "C = 28" + }, + { + "bbox": [ + 130, + 354, + 480, + 437 + ], + "type": "text", + "content": ", as illustrated in Fig.4 (Right). Our proposed vocabulary goes beyond previous datasets in important ways. It distinguishes between the upper and lower halves of limbs and incorporates more detailed classifications such as upper/lower lips, teeth, and tongue. To this end, we manually annotate " + }, + { + "bbox": [ + 130, + 354, + 480, + 437 + ], + "type": "inline_equation", + "content": "100K" + }, + { + "bbox": [ + 130, + 354, + 480, + 437 + ], + "type": "text", + "content": " images at " + }, + { + "bbox": [ + 130, + 354, + 480, + 437 + ], + "type": "inline_equation", + "content": "4K" + }, + { + "bbox": [ + 130, + 354, + 480, + 437 + ], + "type": "text", + "content": " resolution with this vocabulary." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 131, + 455, + 250, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 455, + 250, + 467 + ], + "spans": [ + { + "bbox": [ + 131, + 455, + 250, + 467 + ], + "type": "text", + "content": "3.5 Depth Estimation" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 131, + 475, + 479, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 475, + 479, + 547 + ], + "spans": [ + { + "bbox": [ + 131, + 475, + 479, + 547 + ], + "type": "text", + "content": "For depth estimation, we adopt the architecture used for segmentation, with the modification that the decoder output channel is set to 1 for regression. We denote the ground-truth depth map of image " + }, + { + "bbox": [ + 131, + 475, + 479, + 547 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 131, + 475, + 479, + 547 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 131, + 475, + 479, + 547 + ], + "type": "inline_equation", + "content": "\\mathbf{d} \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 131, + 475, + 479, + 547 + ], + "type": "text", + "content": ", the depth estimator by " + }, + { + "bbox": [ + 131, + 475, + 479, + 547 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 131, + 475, + 479, + 547 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 131, + 475, + 479, + 547 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{d}} = \\mathcal{D}(\\mathbf{I})" + }, + { + "bbox": [ + 131, + 475, + 479, + 547 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 475, + 479, + 547 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 131, + 475, + 479, + 547 + ], + "type": "text", + "content": " as the number of human pixels in the image. For the relative depth estimation, we normalize " + }, + { + "bbox": [ + 131, + 475, + 479, + 547 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 131, + 475, + 479, + 547 + ], + "type": "text", + "content": " to the range [0, 1] using max and min depths in the image. The " + }, + { + "bbox": [ + 131, + 475, + 479, + 547 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{depth}}" + }, + { + "bbox": [ + 131, + 475, + 479, + 547 + ], + "type": "text", + "content": " loss [32] for " + }, + { + "bbox": [ + 131, + 475, + 479, + 547 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 131, + 475, + 479, + 547 + ], + "type": "text", + "content": " is defined as follows:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 183, + 555, + 480, + 570 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 183, + 555, + 480, + 570 + ], + "spans": [ + { + "bbox": [ + 183, + 555, + 480, + 570 + ], + "type": "interline_equation", + "content": "\\Delta \\mathbf {d} = \\log (\\mathbf {d}) - \\log (\\hat {\\mathbf {d}}), \\tag {1}", + "image_path": "8376638915f0e77cf3bbbf57dfb38b6d14ca13a1b8734bd93d93c68c6e83a540.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 183, + 573, + 480, + 605 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 183, + 573, + 480, + 605 + ], + "spans": [ + { + "bbox": [ + 183, + 573, + 480, + 605 + ], + "type": "interline_equation", + "content": "\\overline {{\\Delta \\mathbf {d}}} = \\frac {1}{M} \\sum_ {i = 1} ^ {M} \\Delta \\mathbf {d} _ {i}, \\quad \\overline {{(\\Delta \\mathbf {d}) ^ {2}}} = \\frac {1}{M} \\sum_ {i = 1} ^ {M} (\\Delta \\mathbf {d} _ {i}) ^ {2}, \\tag {2}", + "image_path": "f799af776e76e3c61a85a34bdbfc3396556527e9a9bce21a7b6cc0bb74cfbf12.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 171, + 608, + 480, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 608, + 480, + 634 + ], + "spans": [ + { + "bbox": [ + 171, + 608, + 480, + 634 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {d e p t h}} = \\sqrt {\\left(\\overline {{\\Delta \\mathbf {d}}} ^ {2}\\right) - \\frac {1}{2} \\left(\\overline {{\\Delta \\mathbf {d}}} ^ {2}\\right) ^ {2}}. \\tag {3}", + "image_path": "4c118067248672358b0104a946e869f27999a0c629e4560017a316a85049ca20.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "content": "We render 500,000 synthetic images using 600 high-resolution photogrammetry human scans as shown in Fig. 5 to obtain a robust monocular depth" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 415, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 415, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 415, + 91, + 447, + 101 + ], + "type": "text", + "content": "Sapiens" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": "estimation model with high-fidelity. A random background is selected from a 100 HDRI environment map collection. We place a virtual camera within the scene, randomly adjusting its focal length, rotation, and translation to capture images and their associated ground-truth depth maps at 4K resolution." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 170, + 481, + 327 + ], + "blocks": [ + { + "bbox": [ + 133, + 170, + 481, + 327 + ], + "lines": [ + { + "bbox": [ + 133, + 170, + 481, + 327 + ], + "spans": [ + { + "bbox": [ + 133, + 170, + 481, + 327 + ], + "type": "image", + "image_path": "b0c67ba670a15af50b812c31374d01223983b839a9491536af1a0ccb7b56bbee.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 331, + 479, + 344 + ], + "lines": [ + { + "bbox": [ + 132, + 331, + 479, + 344 + ], + "spans": [ + { + "bbox": [ + 132, + 331, + 479, + 344 + ], + "type": "text", + "content": "Fig. 5: Ground-truth synthetic annotations for depth and surface normal estimation." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 350, + 298, + 362 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 350, + 298, + 362 + ], + "spans": [ + { + "bbox": [ + 132, + 350, + 298, + 362 + ], + "type": "text", + "content": "3.6 Surface Normal Estimation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 372, + 482, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 372, + 482, + 444 + ], + "spans": [ + { + "bbox": [ + 130, + 372, + 482, + 444 + ], + "type": "text", + "content": "Similar to previous tasks, we set the decoder output channels of the normal estimator " + }, + { + "bbox": [ + 130, + 372, + 482, + 444 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 130, + 372, + 482, + 444 + ], + "type": "text", + "content": " to be 3, corresponding to the " + }, + { + "bbox": [ + 130, + 372, + 482, + 444 + ], + "type": "inline_equation", + "content": "xyz" + }, + { + "bbox": [ + 130, + 372, + 482, + 444 + ], + "type": "text", + "content": " components of the normal vector at each pixel. The generated synthetic data is also used as supervision for surface normal estimation. Let " + }, + { + "bbox": [ + 130, + 372, + 482, + 444 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 130, + 372, + 482, + 444 + ], + "type": "text", + "content": " be the ground-truth normal map for image " + }, + { + "bbox": [ + 130, + 372, + 482, + 444 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 130, + 372, + 482, + 444 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 372, + 482, + 444 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{n}} = \\mathcal{N}(\\mathbf{I})" + }, + { + "bbox": [ + 130, + 372, + 482, + 444 + ], + "type": "text", + "content": ". Similar to depth, the loss " + }, + { + "bbox": [ + 130, + 372, + 482, + 444 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{normal}}" + }, + { + "bbox": [ + 130, + 372, + 482, + 444 + ], + "type": "text", + "content": " is only computed for human pixels in the image and is defined as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 233, + 453, + 481, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 453, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 233, + 453, + 481, + 468 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {n o r m a l}} = \\left\\| \\mathbf {n} - \\hat {\\mathbf {n}} \\right\\| _ {1} + (1 - \\mathbf {n} \\cdot \\hat {\\mathbf {n}}) \\tag {4}", + "image_path": "c56148bfc526b490cf524c8d292f3ad245fbebeea940d287f923615e0056a7b1.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 475, + 230, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 230, + 488 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 230, + 488 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 501, + 481, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 501, + 481, + 537 + ], + "spans": [ + { + "bbox": [ + 130, + 501, + 481, + 537 + ], + "type": "text", + "content": "In this section, we initially provide an overview of the implementation details. Subsequently, we conduct comprehensive benchmarking across four tasks: pose estimation, part segmentation, depth estimation, and normal estimation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 548, + 280, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 548, + 280, + 561 + ], + "spans": [ + { + "bbox": [ + 132, + 548, + 280, + 561 + ], + "type": "text", + "content": "4.1 Implementation Details" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Our largest model, Sapiens-2B, is pretrained using 1024 A100 GPUs for 18 days using PyTorch. We use the AdamW [73] optimizer for all our experiments. The learning schedule includes a brief linear warm-up, followed by cosine annealing [72] for pretraining and linear decay [65] for finetuning. All models are pretrained from scratch at a resolution of " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " with a patch size of 16. For finetuning, the input image is resized to a 4:3 ratio, i.e. " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "1024 \\times 768" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": ". We use standard augmentations like cropping, scaling, flipping, and photometric distortions. A random background from non-human COCO [67] images is added for" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 245, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 245, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 245, + 101 + ], + "type": "text", + "content": "R. Khirodkar et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 169, + 114, + 449, + 171 + ], + "blocks": [ + { + "bbox": [ + 169, + 114, + 449, + 171 + ], + "lines": [ + { + "bbox": [ + 169, + 114, + 449, + 171 + ], + "spans": [ + { + "bbox": [ + 169, + 114, + 449, + 171 + ], + "type": "table", + "html": "
Model#ParamsFLOPsHidden sizeLayersHeadsBatch size
Sapiens-0.3B0.336 B1.242 T1024241698,304
Sapiens-0.6B0.664 B2.583 T1280321665,536
Sapiens-1B1.169 B4.647 T1536402440,960
Sapiens-2B2.163 B8.709 T1920483220,480
", + "image_path": "89830f00747d92549f23f5fd6e7f15156f831b352c75be0798c7f7513e34ade5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 190, + 480, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 190, + 480, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 190, + 480, + 285 + ], + "type": "text", + "content": "segmentation, depth, and normal prediction tasks. Importantly, we use differential learning rates [115] to preserve generalization i.e. lower learning rates for initial layers and progressively higher rates for subsequent layers. The layer-wise learning rate decay is set to 0.85 with a weight decay of 0.1 for the encoder. We detail the design specifications of Sapiens in Table. 2. Following [34, 100], we prioritize scaling models by width rather than depth. Note that the Sapiens-0.3B model, while architecturally similar to the traditional ViT-Large, consists of twentyfold more FLOPs due to its higher resolution." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 290, + 261, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 290, + 261, + 300 + ], + "spans": [ + { + "bbox": [ + 131, + 290, + 261, + 300 + ], + "type": "text", + "content": "4.2 2D Pose Estimation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 304, + 480, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 480, + 447 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 480, + 447 + ], + "type": "text", + "content": "We finetune Sapiens for face, body, feet, and hand (" + }, + { + "bbox": [ + 130, + 304, + 480, + 447 + ], + "type": "inline_equation", + "content": "K = 308" + }, + { + "bbox": [ + 130, + 304, + 480, + 447 + ], + "type": "text", + "content": ") pose estimation on our high-fidelity annotations. For training, we use the train set with " + }, + { + "bbox": [ + 130, + 304, + 480, + 447 + ], + "type": "inline_equation", + "content": "1M" + }, + { + "bbox": [ + 130, + 304, + 480, + 447 + ], + "type": "text", + "content": " images and for evaluation, we use the test set, named Humans-5K, with " + }, + { + "bbox": [ + 130, + 304, + 480, + 447 + ], + "type": "inline_equation", + "content": "5K" + }, + { + "bbox": [ + 130, + 304, + 480, + 447 + ], + "type": "text", + "content": " images. Our evaluation is top-down [111] i.e. we use an off-the-shelf detector [37] for bounding-box and conduct single human pose inference. Table 3 shows a comparison of our models with existing methods for whole-body pose estimation. We evaluate all methods on 114 common keypoints between our 308 keypoint vocabulary and the 133 keypoint vocabulary from COCO-WholeBody [55]. Sapiens-0.6B surpasses the current state-of-the-art, DWPose-1 [114] by +2.8 AP. Contrary to DWPose [114], which utilizes a complex student-teacher framework with feature distillation tailored for the task, Sapiens adopts a general encoder-decoder architecture with large human-centric pretraining." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 448, + 481, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 448, + 481, + 520 + ], + "spans": [ + { + "bbox": [ + 130, + 448, + 481, + 520 + ], + "type": "text", + "content": "Interestingly, even with the same parameter count, our models demonstrate superior performance compared to their counterparts. For instance, Sapiens-0.3B exceeds VitPose+-L by +5.6 AP, and Sapiens-0.6B outperforms VitPose+-H by +7.9 AP. Within the Sapiens family, our results indicate a direct correlation between model size and performance. Sapiens-2B sets a state-of-the-art with 61.1 AP, a significant improvement of +7.6 AP to the prior art. Despite fine-tuning" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 138, + 526, + 485, + 658 + ], + "blocks": [ + { + "bbox": [ + 140, + 174, + 471, + 185 + ], + "lines": [ + { + "bbox": [ + 140, + 174, + 471, + 185 + ], + "spans": [ + { + "bbox": [ + 140, + 174, + 471, + 185 + ], + "type": "text", + "content": "Table 2: Sapiens encoder specifications for pretraining on Human-300M dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 526, + 485, + 658 + ], + "lines": [ + { + "bbox": [ + 138, + 526, + 485, + 658 + ], + "spans": [ + { + "bbox": [ + 138, + 526, + 485, + 658 + ], + "type": "table", + "html": "
ModelInput SizeBodyFootFaceHandWhole-body
APARAPARAPARAPARAPAR
DeepPose [98]384 × 28832.143.525.341.237.853.915.731.623.937.2
SimpleBaseline [106]384 × 28852.360.149.862.559.667.341.451.844.653.7
HRNet [93]384 × 28855.862.645.255.458.964.539.347.645.753.9
ZoomNAS [110]384 × 28859.766.348.157.974.579.249.860.652.160.7
ViTPose+L [112]256 × 19261.066.862.468.250.155.741.547.347.853.6
ViTPose+-H [112]256 × 19261.667.463.269.050.756.342.047.848.354.1
RTMPose-x [54]384 × 28857.163.755.366.874.478.546.355.051.959.6
DWPose-m [114]256 × 19254.261.449.963.068.574.240.150.047.755.8
DWPose-l [114]384 × 28857.964.256.567.474.378.449.357.453.160.6
Sapiens-0.3B (Ours)1024 × 76858.164.556.867.774.578.649.657.753.4 (+0.3)60.9 (+0.3)
Sapiens-0.6B (Ours)1024 × 76859.865.564.772.375.279.052.160.356.2 (+2.8)62.4 (+2.1)
Sapiens-1B (Ours)1024 × 76862.968.268.375.176.479.755.963.459.4 (+5.9)65.3 (+5.1)
Sapiens-2B (Ours)1024 × 76864.769.969.476.276.979.957.164.461.1(+7.6)67.1(+7.0)
", + "image_path": "26011f812389c7d252523d01eebe61c9d8b586f816965f1f398a581dd1a9e692.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 151, + 662, + 460, + 673 + ], + "lines": [ + { + "bbox": [ + 151, + 662, + 460, + 673 + ], + "spans": [ + { + "bbox": [ + 151, + 662, + 460, + 673 + ], + "type": "text", + "content": "Table 3: Pose estimation results on Humans-5K test set. Flip test is used." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 415, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 415, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 415, + 91, + 447, + 101 + ], + "type": "text", + "content": "Sapiens" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 240, + 186 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 186 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 186 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 186 + ], + "type": "image", + "image_path": "59e713e7960beedb77ed778998fe60572b14023b7a02f3e5fd98f3be7656a33b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 140, + 192, + 471, + 204 + ], + "lines": [ + { + "bbox": [ + 140, + 192, + 471, + 204 + ], + "spans": [ + { + "bbox": [ + 140, + 192, + 471, + 204 + ], + "type": "text", + "content": "Fig. 6: Pose estimation with Sapiens-1B for 308 keypoints on in-the-wild images." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 240, + 115, + 293, + 186 + ], + "blocks": [ + { + "bbox": [ + 240, + 115, + 293, + 186 + ], + "lines": [ + { + "bbox": [ + 240, + 115, + 293, + 186 + ], + "spans": [ + { + "bbox": [ + 240, + 115, + 293, + 186 + ], + "type": "image", + "image_path": "552a641dad9dc63dab04fbb96772b1ed292af4049983426ae4501f0361f0fcdb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 293, + 115, + 374, + 186 + ], + "blocks": [ + { + "bbox": [ + 293, + 115, + 374, + 186 + ], + "lines": [ + { + "bbox": [ + 293, + 115, + 374, + 186 + ], + "spans": [ + { + "bbox": [ + 293, + 115, + 374, + 186 + ], + "type": "image", + "image_path": "65b6b443723029db3d843fb46d4ae2061f4e00dc7a34c41fdb9cdc36b114f935.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 374, + 115, + 480, + 186 + ], + "blocks": [ + { + "bbox": [ + 374, + 115, + 480, + 186 + ], + "lines": [ + { + "bbox": [ + 374, + 115, + 480, + 186 + ], + "spans": [ + { + "bbox": [ + 374, + 115, + 480, + 186 + ], + "type": "image", + "image_path": "cb9f6c5d47c337b2a123ce6e609f19db0fe9e1720cb51bcc90df1bf918e79bb1.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 213, + 479, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 479, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 479, + 237 + ], + "type": "text", + "content": "with annotations from a indoor capture studio, Sapiens demonstrate robust generalization to real-world, as shown in Fig. 6." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 255, + 285, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 255, + 285, + 268 + ], + "spans": [ + { + "bbox": [ + 132, + 255, + 285, + 268 + ], + "type": "text", + "content": "4.3 Body-Part Segmentation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 277, + 332, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 332, + 395 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 332, + 395 + ], + "type": "text", + "content": "We fine-tune and evaluate our annotations with a segmentation vocabulary of 28 classes. Our train set consists of " + }, + { + "bbox": [ + 130, + 277, + 332, + 395 + ], + "type": "inline_equation", + "content": "100K" + }, + { + "bbox": [ + 130, + 277, + 332, + 395 + ], + "type": "text", + "content": " images, and the test set, Humans-2K, consists of " + }, + { + "bbox": [ + 130, + 277, + 332, + 395 + ], + "type": "inline_equation", + "content": "2K" + }, + { + "bbox": [ + 130, + 277, + 332, + 395 + ], + "type": "text", + "content": " images. We compare Sapiens with existing bodypart segmentation methods fine-tuned on our train set. Importantly, we use suggested pretrained checkpoints by each method as initialization. Similar to pose, we observe generalization to segmentation as shown in Table 4." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 350, + 257, + 476, + 346 + ], + "blocks": [ + { + "bbox": [ + 350, + 257, + 476, + 346 + ], + "lines": [ + { + "bbox": [ + 350, + 257, + 476, + 346 + ], + "spans": [ + { + "bbox": [ + 350, + 257, + 476, + 346 + ], + "type": "table", + "html": "
ModelmIoU(%)mAcc(%)
FCN* [71]48.257.6
SegFormer* [107]53.562.9
Mask2Former* [22]58.768.3
DeepLabV3+* [18]64.174.8
Sapiens-0.3B (Ours)76.786.1
Sapiens-0.6B (Ours)77.886.3
Sapiens-1B (Ours)79.989.1
Sapiens-2B (Ours)81.289.4
", + "image_path": "d44e1a91d10252c825c33c02c58dc93a22dae06a5c3e89502ed9f707726760a1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 339, + 351, + 481, + 384 + ], + "lines": [ + { + "bbox": [ + 339, + 351, + 481, + 384 + ], + "spans": [ + { + "bbox": [ + 339, + 351, + 481, + 384 + ], + "type": "text", + "content": "Table 4: We report mIoU and mAcc on Humans-2K test set. Methods with * are trained by us." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 396, + 482, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 396, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 482, + 468 + ], + "type": "text", + "content": "Interestingly, our smallest model, Sapiens-0.3B outperforms existing state-of-the-art segmentation methods like Mask2Former [22] and DeepLabV3+ [18] by 12.6 mIoU due to its higher resolution and large human-centric pretraining. Furthermore, increasing the model size improves segmentation performance. Sapiens-2B achieves the best performance of 81.2 mIoU and 89.4 mAcc on the test set. Fig. 7 shows the qualitative results of our models." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 479, + 251, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 251, + 491 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 251, + 491 + ], + "type": "text", + "content": "4.4 Depth Estimation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 499, + 482, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 499, + 482, + 584 + ], + "spans": [ + { + "bbox": [ + 130, + 499, + 482, + 584 + ], + "type": "text", + "content": "We evaluate our models on THuman2.0 [117] and Hi4D [116] datasets for depth estimation. THuman2.0 consists of 526 high-quality human scans, from which we derive three sets of images for testing: a) face, b) upper body, and c) full body using a virtual camera. THuman2.0 with 1578 images thus enables the evaluation of our models' performance on single-human images across multiple scales. Conversely, the Hi4D dataset focuses on multi-human scenarios, with each sequence showcasing two subjects engaged in activities involving human-human" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 133, + 590, + 273, + 652 + ], + "blocks": [ + { + "bbox": [ + 133, + 590, + 273, + 652 + ], + "lines": [ + { + "bbox": [ + 133, + 590, + 273, + 652 + ], + "spans": [ + { + "bbox": [ + 133, + 590, + 273, + 652 + ], + "type": "image", + "image_path": "3ff54b7187a6b7e52537224d6211d49ccc4d448604f0c8bd3651a36c6d0e9e36.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 658, + 480, + 670 + ], + "lines": [ + { + "bbox": [ + 132, + 658, + 480, + 670 + ], + "spans": [ + { + "bbox": [ + 132, + 658, + 480, + 670 + ], + "type": "text", + "content": "Fig. 7: Human segmentation with Sapiens-1B for 28 categories on in-the-wild images." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 273, + 590, + 346, + 652 + ], + "blocks": [ + { + "bbox": [ + 273, + 590, + 346, + 652 + ], + "lines": [ + { + "bbox": [ + 273, + 590, + 346, + 652 + ], + "spans": [ + { + "bbox": [ + 273, + 590, + 346, + 652 + ], + "type": "image", + "image_path": "f7e1f5af38c018e0cc541d44bfda9c2f1ec853e339232a34c218bdb52f849f12.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 346, + 590, + 378, + 652 + ], + "blocks": [ + { + "bbox": [ + 346, + 590, + 378, + 652 + ], + "lines": [ + { + "bbox": [ + 346, + 590, + 378, + 652 + ], + "spans": [ + { + "bbox": [ + 346, + 590, + 378, + 652 + ], + "type": "image", + "image_path": "5eb79fee42cd7ce46d0985b236995a4b28539eab2ac721597fcd1d7e2972b782.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 378, + 590, + 480, + 652 + ], + "blocks": [ + { + "bbox": [ + 378, + 590, + 480, + 652 + ], + "lines": [ + { + "bbox": [ + 378, + 590, + 480, + 652 + ], + "spans": [ + { + "bbox": [ + 378, + 590, + 480, + 652 + ], + "type": "image", + "image_path": "074adc99db64f71394d54eafed3a97513acb0fda084d2275eb4ca8498207cdd8.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 245, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 245, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 245, + 101 + ], + "type": "text", + "content": "R. Khirodkar et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 106, + 493, + 203 + ], + "blocks": [ + { + "bbox": [ + 135, + 106, + 493, + 203 + ], + "lines": [ + { + "bbox": [ + 135, + 106, + 493, + 203 + ], + "spans": [ + { + "bbox": [ + 135, + 106, + 493, + 203 + ], + "type": "table", + "html": "
MethodTH2.0-FaceTH2.0-UprBodyTH2.0-FullBodyHi4D
RMSE ↓AbsRel ↓δ1 ↑RMSEAbsRelδ1RMSEAbsRelδ1RMSEAbsRelδ1
MiDaS-L [11]0.1140.0970.9250.3980.2710.8680.7010.6890.7820.2610.0820.975
MiDaS-Swin2 [11]0.0500.0360.9950.1220.0810.9480.2920.1710.8620.2090.0630.997
DepthAny-B [113]0.0390.0260.9990.0480.0280.9990.0610.0300.9990.1430.0340.997
DepthAny-L [113]0.0390.0270.9990.0480.0270.9990.0600.0300.9990.1470.0350.997
Sapiens-0.3B (Ours)0.0120.0081.0000.0150.0091.0000.0210.0101.0000.1480.0461.000
Sapiens-0.6B (Ours)0.0110.0081.0000.0150.0091.0000.0210.0101.0000.1420.0441.000
Sapiens-1B (Ours)0.0090.0061.0000.0120.0071.0000.0190.0091.0000.1250.0391.000
Sapiens-2B (Ours)0.0080.0051.0000.0100.0061.0000.0160.0081.0000.1140.0361.000
", + "image_path": "469297cfbb6c0870753dfb5f01902d78199abd0cad50ff491af5772f5c0b3f6f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 204, + 477, + 216 + ], + "lines": [ + { + "bbox": [ + 134, + 204, + 477, + 216 + ], + "spans": [ + { + "bbox": [ + 134, + 204, + 477, + 216 + ], + "type": "text", + "content": "Table 5: Comparison of Sapiens for monocular depth estimation on human images." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 220, + 480, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 220, + 480, + 292 + ], + "spans": [ + { + "bbox": [ + 130, + 220, + 480, + 292 + ], + "type": "text", + "content": "interactions. We select sequences from pair 28, pair 32, and pair 37, featuring 6 unique subjects from camera 4, totaling 1195 multi-human real images for testing. We follow the relative-depth evaluation protocols established by MiDaSv3.1 [11], reporting standard metrics such as AbsRel and " + }, + { + "bbox": [ + 130, + 220, + 480, + 292 + ], + "type": "inline_equation", + "content": "\\delta_{1}" + }, + { + "bbox": [ + 130, + 220, + 480, + 292 + ], + "type": "text", + "content": ". In addition, we also report RMSE as our primary metric since " + }, + { + "bbox": [ + 130, + 220, + 480, + 292 + ], + "type": "inline_equation", + "content": "\\delta_{1}" + }, + { + "bbox": [ + 130, + 220, + 480, + 292 + ], + "type": "text", + "content": " does not effectively reflect performance in human scenes characterized by subtle depth variations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 293, + 482, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 400 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 400 + ], + "type": "text", + "content": "Table 5 compares our models with existing state-of-the-art monocular depth estimators. Sapiens-2B, finetuned solely on synthetic data, remarkably outperforms prior art across all single-human scales and multi-human scenarios. We observe a " + }, + { + "bbox": [ + 130, + 293, + 482, + 400 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 293, + 482, + 400 + ], + "type": "text", + "content": " RMSE reduction compared to the top-performing Depth-Anything model on Hi4D images. It is important to highlight that while baseline models are trained on a variety of scenes, Sapiens specializes in human-centric depth estimation. Fig. 8 presents a qualitative comparison of depth estimation between Sapiens-1B and DepthAnything-L. To ensure a fair comparison, the predicted depth is renormalized using the human mask in the baseline visualizations." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 132, + 408, + 482, + 628 + ], + "blocks": [ + { + "bbox": [ + 132, + 408, + 482, + 628 + ], + "lines": [ + { + "bbox": [ + 132, + 408, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 132, + 408, + 482, + 628 + ], + "type": "image", + "image_path": "84ad7805d41928a1b011e4439b9e54b06ba160ab4940c7e5f56dbf429b91251a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 632, + 482, + 654 + ], + "lines": [ + { + "bbox": [ + 130, + 632, + 482, + 654 + ], + "spans": [ + { + "bbox": [ + 130, + 632, + 482, + 654 + ], + "type": "text", + "content": "Fig. 8: We compare our depth prediction with DepthAnything [113]. To showcase the consistency of predicted depth, we also visualize the " + }, + { + "bbox": [ + 130, + 632, + 482, + 654 + ], + "type": "inline_equation", + "content": "\\nabla" + }, + { + "bbox": [ + 130, + 632, + 482, + 654 + ], + "type": "text", + "content": " depth as pseudo surface normals." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 415, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 415, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 415, + 91, + 447, + 101 + ], + "type": "text", + "content": "Sapiens" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 297, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 297, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 297, + 126 + ], + "type": "text", + "content": "4.5 Surface Normal Estimation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 129, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 129, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 129, + 482, + 248 + ], + "type": "text", + "content": "The datasets for surface normal evaluation are identical to those used for depth estimation. Following [30], we report the mean and median angular error, along with the percentage of pixels within " + }, + { + "bbox": [ + 130, + 129, + 482, + 248 + ], + "type": "inline_equation", + "content": "t^\\circ" + }, + { + "bbox": [ + 130, + 129, + 482, + 248 + ], + "type": "text", + "content": " error for " + }, + { + "bbox": [ + 130, + 129, + 482, + 248 + ], + "type": "inline_equation", + "content": "t \\in \\{11.25^\\circ, 22.5^\\circ, 30^\\circ\\}" + }, + { + "bbox": [ + 130, + 129, + 482, + 248 + ], + "type": "text", + "content": ". Table 6 compares our models with existing human-specific surface normal estimators. All our models outperform existing methods by a significant margin. Sapiens-2B achieves a mean error of around " + }, + { + "bbox": [ + 130, + 129, + 482, + 248 + ], + "type": "inline_equation", + "content": "12^\\circ" + }, + { + "bbox": [ + 130, + 129, + 482, + 248 + ], + "type": "text", + "content": " on the THuman2.0 (single-human) and Hi4D (multi-human) datasets. We qualitatively compare Sapiens-1B with PIFuHD [89] and ECON [108] for surface normal estimation in Figure 9. Note that PIFuHD [89] is trained with the identical set of 3D scans as ours, and ECON [108] is trained with 4000 scans that are a super set of our 3D scan data." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 142, + 255, + 473, + 367 + ], + "blocks": [ + { + "bbox": [ + 142, + 255, + 473, + 367 + ], + "lines": [ + { + "bbox": [ + 142, + 255, + 473, + 367 + ], + "spans": [ + { + "bbox": [ + 142, + 255, + 473, + 367 + ], + "type": "table", + "html": "
MethodTHuman2.0 [117]Hi4D [116]
Angular Error°% Within t°Angular Error°% Within t°
MeanMedian11.25°22.5°30°MeanMedian11.25°22.5°30°
PIFuHD [89]30.5127.1315.8142.9758.8622.3919.2622.9860.1477.02
HDNet [52]34.8230.6017.4439.2654.5128.6026.8519.0857.9370.14
ICON [109]28.7425.5222.8147.8363.7320.1817.5226.8166.3482.73
ECON [108]25.4523.6732.9555.8669.0318.4616.4729.3568.1284.88
Sapiens-0.3B13.0210.3357.3786.2092.715.0412.2247.0781.4990.70
Sapiens-0.6B12.8610.2357.8586.6893.3014.0611.4750.5984.3792.54
Sapiens-1B12.119.4061.9788.0393.8412.189.5960.3688.6294.44
Sapiens-2B11.849.1663.1688.6094.1812.149.6260.2289.0894.74
", + "image_path": "7482f87a6ff3f7fe9a68abce17970bbeb23ebf8500b7d07da035d762d5bd52b4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 141, + 369, + 473, + 381 + ], + "lines": [ + { + "bbox": [ + 141, + 369, + 473, + 381 + ], + "spans": [ + { + "bbox": [ + 141, + 369, + 473, + 381 + ], + "type": "text", + "content": "Table 6: Comparison of Sapiens for surface normal estimation on human images." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 157, + 384, + 455, + 660 + ], + "blocks": [ + { + "bbox": [ + 157, + 384, + 455, + 660 + ], + "lines": [ + { + "bbox": [ + 157, + 384, + 455, + 660 + ], + "spans": [ + { + "bbox": [ + 157, + 384, + 455, + 660 + ], + "type": "image", + "image_path": "d33afce36c657ac7f6dc7a049f860fec3cb6d7f447001294e5e56f018d805460.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 661, + 482, + 684 + ], + "lines": [ + { + "bbox": [ + 130, + 661, + 482, + 684 + ], + "spans": [ + { + "bbox": [ + 130, + 661, + 482, + 684 + ], + "type": "text", + "content": "Fig.9: Qualitative comparison of Sapiens-1B with PIFuHD [89] and ECON [108] for monocular surface normal estimation on challenging in-the-wild images." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 244, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 244, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 244, + 101 + ], + "type": "text", + "content": "R. Khirodkar et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 214, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 214, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 214, + 126 + ], + "type": "text", + "content": "4.6 Discussion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 482, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 482, + 329 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 482, + 329 + ], + "type": "text", + "content": "Importance of Pretraining Data Source. The feature quality is closely linked to the pretraining data quality. We assess the importance of pretraining on various data sources for human-centric tasks by pretraining Sapiens-0.3B on each dataset under identical training schedules and number of iterations. We fine-tune the model on each task and select early checkpoints for evaluation, reasoning that early-stage fine-tuning better reflects the model's generalization capability. We investigate the impact of pretraining at scale on general images (which may include humans) versus exclusively human images using Sapiens. We randomly select 100 million and 300 million general images from our 1 billion image corpus to create the General-100M and General-300M datasets, respectively. Table 7 showcases the comparison of pretraining outcomes. We report mAP for pose on Humans-5K, mIoU for segmentation on Humans-2K, RMSE for depth on THuman2.0, and mean angular error in degrees for normal estimation on Hi4D. Aligned with findings from [112], our results show that pretraining with Human300M leads to superior performance across all metrics, highlighting the benefits of human-centric pretraining within a fixed computational budget." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 329, + 317, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 329, + 317, + 474 + ], + "spans": [ + { + "bbox": [ + 130, + 329, + 317, + 474 + ], + "type": "text", + "content": "We also study the effect of number of unique human images seen during pretraining with normal estimation performance. We report " + }, + { + "bbox": [ + 130, + 329, + 317, + 474 + ], + "type": "inline_equation", + "content": "\\%" + }, + { + "bbox": [ + 130, + 329, + 317, + 474 + ], + "type": "text", + "content": " within " + }, + { + "bbox": [ + 130, + 329, + 317, + 474 + ], + "type": "inline_equation", + "content": "30^{\\circ}" + }, + { + "bbox": [ + 130, + 329, + 317, + 474 + ], + "type": "text", + "content": ". Again, we maintain identical conditions for Sapiens-0.3B pretraining and finetuning. Fig.10 shows a steady improvement in performance as the pretraining data size increases without saturation. In summary, the diversity of human images during pretraining directly correlates with improved generalization to down-stream tasks." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 321, + 338, + 482, + 474 + ], + "blocks": [ + { + "bbox": [ + 321, + 338, + 482, + 474 + ], + "lines": [ + { + "bbox": [ + 321, + 338, + 482, + 474 + ], + "spans": [ + { + "bbox": [ + 321, + 338, + 482, + 474 + ], + "type": "image", + "image_path": "f3e5ddc2209cf13d89e34b66ca5793794a923b1006ddb4e00fd8dccb76fd0079.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 481, + 482, + 515 + ], + "lines": [ + { + "bbox": [ + 321, + 481, + 482, + 515 + ], + "spans": [ + { + "bbox": [ + 321, + 481, + 482, + 515 + ], + "type": "text", + "content": "Fig. 10: Sapiens-0.3B's normal estimation performance with unique human images seen during pretraining." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 479, + 317, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 317, + 526 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 317, + 526 + ], + "type": "text", + "content": "Zero-Shot Generalization. Our models exhibit broad generalization to a variety of settings. For instance, in segmentation, Sapiens are finetuned on single-human im" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 527, + 482, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 527, + 482, + 575 + ], + "spans": [ + { + "bbox": [ + 130, + 527, + 482, + 575 + ], + "type": "text", + "content": "ages with limited subject diversity, minimal background variation, and solely third-person views (see Fig. 4). Nevertheless, our large-scale pretraining enables generalization across number of subjects, varying ages, and egocentric views, as shown in Fig. 11. These observations similarly hold for other tasks." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 169, + 582, + 451, + 650 + ], + "blocks": [ + { + "bbox": [ + 169, + 582, + 451, + 650 + ], + "lines": [ + { + "bbox": [ + 169, + 582, + 451, + 650 + ], + "spans": [ + { + "bbox": [ + 169, + 582, + 451, + 650 + ], + "type": "table", + "html": "
Pretraining Source#ImagesPose (↑)Seg(↑)Depth(↓)Normal(↓)
Random Initialization-30.240.30.72035.4
General-100M100M35.750.10.35127.5
General-300M300M37.352.80.34726.8
Humans-100M100M43.661.20.31624.0
Humans-300M (Full)300M47.066.50.28821.8
", + "image_path": "ddd9582152d5eca7d00db2c35b18ca23d810b5044a8668dd034015cc0cbd0a96.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 651, + 481, + 674 + ], + "lines": [ + { + "bbox": [ + 130, + 651, + 481, + 674 + ], + "spans": [ + { + "bbox": [ + 130, + 651, + 481, + 674 + ], + "type": "text", + "content": "Table 7: Comparison of Sapiens-0.3B pretrained on various data sources. A domain-specific pretraining yields superior results compared to general data sources." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 415, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 415, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 415, + 91, + 447, + 102 + ], + "type": "text", + "content": "Sapiens" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 246, + 186 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 186 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 186 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 186 + ], + "type": "image", + "image_path": "4d32dc066bdc97c77fd4eddfb2d6841af85623dff4abad49e54a0e317551bc00.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 149, + 186, + 231, + 194 + ], + "lines": [ + { + "bbox": [ + 149, + 186, + 231, + 194 + ], + "spans": [ + { + "bbox": [ + 149, + 186, + 231, + 194 + ], + "type": "text", + "content": "Multi-Human Generalization" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 137, + 201, + 474, + 213 + ], + "lines": [ + { + "bbox": [ + 137, + 201, + 474, + 213 + ], + "spans": [ + { + "bbox": [ + 137, + 201, + 474, + 213 + ], + "type": "text", + "content": "Fig. 11: Sapiens achieve broad generalization via large human-centric pretraining." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 246, + 114, + 353, + 186 + ], + "blocks": [ + { + "bbox": [ + 246, + 114, + 353, + 186 + ], + "lines": [ + { + "bbox": [ + 246, + 114, + 353, + 186 + ], + "spans": [ + { + "bbox": [ + 246, + 114, + 353, + 186 + ], + "type": "image", + "image_path": "f8b0e003be8dc2dcc8d3db3aeec1cc5f04a5fa88fccb1bd4bdc59fdf69468f49.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 272, + 186, + 328, + 195 + ], + "lines": [ + { + "bbox": [ + 272, + 186, + 328, + 195 + ], + "spans": [ + { + "bbox": [ + 272, + 186, + 328, + 195 + ], + "type": "text", + "content": "Age Generalization" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 354, + 114, + 480, + 186 + ], + "blocks": [ + { + "bbox": [ + 354, + 114, + 480, + 186 + ], + "lines": [ + { + "bbox": [ + 354, + 114, + 480, + 186 + ], + "spans": [ + { + "bbox": [ + 354, + 114, + 480, + 186 + ], + "type": "image", + "image_path": "1d80308ea50fc79ee73dafd05abf1e08af2f87c6e660abfac74723543ebf442c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 381, + 186, + 454, + 195 + ], + "lines": [ + { + "bbox": [ + 381, + 186, + 454, + 195 + ], + "spans": [ + { + "bbox": [ + 381, + 186, + 454, + 195 + ], + "type": "text", + "content": "Viewpoint Generalization" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 219, + 480, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 219, + 480, + 302 + ], + "spans": [ + { + "bbox": [ + 130, + 219, + 480, + 302 + ], + "type": "text", + "content": "Applications. Accurately predicting pose, body-part segmentation, depth and surface normals for in-the-wild human images enables a wide range of applications, such as 3D human scans and the generation of controllable images. We adapt ControlNet [118] to integrate keypoints for faces and hands predicted by Sapiens, and train this modified model on a subset of the Humans-300M dataset, as shown in Fig. 12. This grants more precise control over image generation, producing anatomically consistent human images." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 133, + 314, + 202, + 384 + ], + "blocks": [ + { + "bbox": [ + 133, + 314, + 202, + 384 + ], + "lines": [ + { + "bbox": [ + 133, + 314, + 202, + 384 + ], + "spans": [ + { + "bbox": [ + 133, + 314, + 202, + 384 + ], + "type": "image", + "image_path": "c2f1a7606945920b7e60e88c7c902901987a86846deaed5a998c4e02df204c1a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 154, + 384, + 176, + 393 + ], + "lines": [ + { + "bbox": [ + 154, + 384, + 176, + 393 + ], + "spans": [ + { + "bbox": [ + 154, + 384, + 176, + 393 + ], + "type": "text", + "content": "Image" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 133, + 399, + 478, + 411 + ], + "lines": [ + { + "bbox": [ + 133, + 399, + 478, + 411 + ], + "spans": [ + { + "bbox": [ + 133, + 399, + 478, + 411 + ], + "type": "text", + "content": "Fig. 12: Image synthesis using ControlNet [118] trained on Sapiens pose predictions." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 202, + 315, + 270, + 384 + ], + "blocks": [ + { + "bbox": [ + 202, + 315, + 270, + 384 + ], + "lines": [ + { + "bbox": [ + 202, + 315, + 270, + 384 + ], + "spans": [ + { + "bbox": [ + 202, + 315, + 270, + 384 + ], + "type": "image", + "image_path": "8c69392d70601b18d4164db9a9928890eb17689317cd4965cd91a2cdd318ceaa.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 211, + 384, + 260, + 393 + ], + "lines": [ + { + "bbox": [ + 211, + 384, + 260, + 393 + ], + "spans": [ + { + "bbox": [ + 211, + 384, + 260, + 393 + ], + "type": "text", + "content": "Predicted Pose" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 274, + 314, + 343, + 384 + ], + "blocks": [ + { + "bbox": [ + 274, + 314, + 343, + 384 + ], + "lines": [ + { + "bbox": [ + 274, + 314, + 343, + 384 + ], + "spans": [ + { + "bbox": [ + 274, + 314, + 343, + 384 + ], + "type": "image", + "image_path": "edf1a9a3e3b1767e500cbe94d93ddb0ffbb7d898a636e1dca733f4481eb12ff5.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 384, + 444, + 393 + ], + "lines": [ + { + "bbox": [ + 310, + 384, + 444, + 393 + ], + "spans": [ + { + "bbox": [ + 310, + 384, + 444, + 393 + ], + "type": "text", + "content": "Pose-conditioned Human Image Synthesis" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 343, + 315, + 412, + 384 + ], + "blocks": [ + { + "bbox": [ + 343, + 315, + 412, + 384 + ], + "lines": [ + { + "bbox": [ + 343, + 315, + 412, + 384 + ], + "spans": [ + { + "bbox": [ + 343, + 315, + 412, + 384 + ], + "type": "image", + "image_path": "b2845bd6118e2beec6fbbd1dc471d7fcbafc3b0bd8e12abaed6ce0f108bedf83.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 411, + 315, + 481, + 384 + ], + "blocks": [ + { + "bbox": [ + 411, + 315, + 481, + 384 + ], + "lines": [ + { + "bbox": [ + 411, + 315, + 481, + 384 + ], + "spans": [ + { + "bbox": [ + 411, + 315, + 481, + 384 + ], + "type": "image", + "image_path": "e0deb4fdb364e6ef015c964610534c3b14288f72b1f80b9ab58f475d273fd521.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 423, + 480, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 480, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 480, + 495 + ], + "type": "text", + "content": "Limitations. While our models generally perform well, they are not perfect. Human images with complex/rare poses, crowding, and severe occlusion are challenging (see supplemental for details). Although aggressive data augmentation and a detect-and-crop strategy could mitigate these issues, we envision our models as a tool for acquiring large-scale, real-world supervision with human-in-the-loop to develop the next generations of human vision models." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 511, + 220, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 511, + 220, + 523 + ], + "spans": [ + { + "bbox": [ + 132, + 511, + 220, + 523 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "type": "text", + "content": "Sapiens represents a significant step toward elevating human-centric vision models into the realm of foundation models. Our models demonstrate strong generalization capabilities on a variety of human-centric tasks. We attribute the state-of-the-art performance of our models to: (i) large-scale pretraining on a large curated dataset, which is specifically tailored to understanding humans, (ii) scaled high-resolution and high-capacity vision transformer backbones, and (iii) high-quality annotations on augmented studio and synthetic data. We believe that these models can become a key building block for a multitude of downstream tasks, and provide access to high-quality vision backbones to a significantly wider part of the community. A potential direction for future work would be extending Sapiens to 3D and multi-modal datasets." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 244, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 244, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 244, + 101 + ], + "type": "text", + "content": "R. Khirodkar et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 198, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 198, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 198, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 137, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 142, + 137, + 480, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 137, + 480, + 158 + ], + "spans": [ + { + "bbox": [ + 142, + 137, + 480, + 158 + ], + "type": "text", + "content": "1. Abnar, S., Dehghani, M., Neyshabur, B., Sedghi, H.: Exploring the limits of large scale pre-training. arXiv preprint arXiv:2110.02095 (2021) 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 142, + 159, + 480, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 159, + 480, + 190 + ], + "spans": [ + { + "bbox": [ + 142, + 159, + 480, + 190 + ], + "type": "text", + "content": "2. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023) 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 142, + 191, + 480, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 191, + 480, + 223 + ], + "spans": [ + { + "bbox": [ + 142, + 191, + 480, + 223 + ], + "type": "text", + "content": "3. Alldieck, T., Zanfir, M., Sminchisescu, C.: Photorealistic monocular 3d reconstruction of humans wearing clothing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1506-1515 (2022) 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 142, + 224, + 480, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 224, + 480, + 255 + ], + "spans": [ + { + "bbox": [ + 142, + 224, + 480, + 255 + ], + "type": "text", + "content": "4. Bai, Y., Geng, X., Mangalam, K., Bar, A., Yuille, A., Darrell, T., Malik, J., Efros, A.A.: Sequential modeling enables scalable learning for large vision models. arXiv preprint arXiv:2312.00785 (2023) 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 142, + 256, + 480, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 256, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 142, + 256, + 480, + 277 + ], + "type": "text", + "content": "5. Bao, H., Dong, L., Piao, S., Wei, F.: Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254 (2021) 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 142, + 277, + 480, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 277, + 480, + 309 + ], + "spans": [ + { + "bbox": [ + 142, + 277, + 480, + 309 + ], + "type": "text", + "content": "6. Barron, J.T., Malik, J.: Intrinsic scene properties from a single rgb-d image. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 17-24 (2013) 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 142, + 309, + 480, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 309, + 480, + 341 + ], + "spans": [ + { + "bbox": [ + 142, + 309, + 480, + 341 + ], + "type": "text", + "content": "7. Barron, J.T., Malik, J.: Shape, illumination, and reflectance from shading. IEEE transactions on pattern analysis and machine intelligence 37(8), 1670-1687 (2014) 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 142, + 342, + 480, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 342, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 142, + 342, + 480, + 363 + ], + "type": "text", + "content": "8. Bartol, K., Bojanic, D., Petkovic, T., Pribanic, T.: A review of body measurement using 3d scanning. IEEE Access 9, 67281-67301 (2021) 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 142, + 363, + 480, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 363, + 480, + 395 + ], + "spans": [ + { + "bbox": [ + 142, + 363, + 480, + 395 + ], + "type": "text", + "content": "9. Bhat, S.F., Alhashim, I., Wonka, P.: Adabins: Depth estimation using adaptive bins. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4009-4018 (2021) 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 396, + 480, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 396, + 480, + 427 + ], + "spans": [ + { + "bbox": [ + 138, + 396, + 480, + 427 + ], + "type": "text", + "content": "10. Bhat, S.F., Birkl, R., Wofk, D., Wonka, P., Müller, M.: Zoodepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288 (2023) 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 428, + 480, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 428, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 138, + 428, + 480, + 449 + ], + "type": "text", + "content": "11. Birkl, R., Wofk, D., Müller, M.: Midas v3. 1-a model zoo for robust monocular relative depth estimation. arXiv preprint arXiv:2307.14460 (2023) 11" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 449, + 480, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 449, + 480, + 481 + ], + "spans": [ + { + "bbox": [ + 138, + 449, + 480, + 481 + ], + "type": "text", + "content": "12. Bojic, L.: Metaverse through the prism of power and addiction: what will happen when the virtual world becomes more attractive than reality? European Journal of Futures Research 10(1), 1-24 (2022) 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 481, + 480, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 481, + 480, + 524 + ], + "spans": [ + { + "bbox": [ + 138, + 481, + 480, + 524 + ], + "type": "text", + "content": "13. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020) 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 525, + 480, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 525, + 480, + 556 + ], + "spans": [ + { + "bbox": [ + 138, + 525, + 480, + 556 + ], + "type": "text", + "content": "14. Cao, Z., Simon, T., Wei, S., Sheikh, Y.: Realtime multi-person 2d pose estimation using part affinity fields. corr abs/1611.08050. arXiv preprint arXiv:1611.08050 (2016) 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 557, + 480, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 557, + 480, + 589 + ], + "spans": [ + { + "bbox": [ + 138, + 557, + 480, + 589 + ], + "type": "text", + "content": "15. Cao, Z., Hidalgo, G., Simon, T., Wei, S.E., Sheikh, Y.: Openpose: real-time multi-person 2d pose estimation using part affinity fields. arXiv preprint arXiv:1812.08008 (2018) 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 589, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 589, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 589, + 480, + 632 + ], + "type": "text", + "content": "16. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021) 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "text", + "content": "17. Chan, C., Ginosar, S., Zhou, T., Efros, A.A.: Everybody dance now. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 5933-5942 (2019) 1" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 415, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 415, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 415, + 91, + 447, + 101 + ], + "type": "text", + "content": "Sapiens" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 138, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 481, + 149 + ], + "type": "text", + "content": "18. Chen, L.C., Zhu, Y., Papandreou, G., Schroff, F., Adam, H.: Encoder-decoder with atrous separable convolution for semantic image segmentation. In: Proceedings of the European conference on computer vision (ECCV). pp. 801-818 (2018) 10" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 151, + 481, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 151, + 481, + 184 + ], + "spans": [ + { + "bbox": [ + 138, + 151, + 481, + 184 + ], + "type": "text", + "content": "19. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International conference on machine learning. pp. 1597-1607. PMLR (2020) 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 185, + 481, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 185, + 481, + 217 + ], + "spans": [ + { + "bbox": [ + 138, + 185, + 481, + 217 + ], + "type": "text", + "content": "20. Chen, X., Wang, X., Changpinyo, S., Piergiovanni, A., Padlewski, P., Salz, D., Goodman, S., Grycner, A., Mustafa, B., Beyer, L., et al.: Pali: A jointly-scaled multilingual language-image model. arXiv preprint arXiv:2209.06794 (2022) 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 218, + 481, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 218, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 138, + 218, + 481, + 251 + ], + "type": "text", + "content": "21. Chen, Y., Wang, Z., Peng, Y., Zhang, Z., Yu, G., Sun, J.: Cascaded pyramid network for multi-person pose estimation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7103-7112 (2018) 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 251, + 481, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 251, + 481, + 284 + ], + "spans": [ + { + "bbox": [ + 138, + 251, + 481, + 284 + ], + "type": "text", + "content": "22. Cheng, B., Misra, I., Schwing, A., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. arxiv 2022. arXiv preprint arXiv:2112.01527 (2021) 10" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 285, + 481, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 285, + 481, + 308 + ], + "spans": [ + { + "bbox": [ + 138, + 285, + 481, + 308 + ], + "type": "text", + "content": "23. Couprie, C., Farabet, C., Najman, L., LeCun, Y.: Indoor semantic segmentation using depth information. arXiv preprint arXiv:1301.3572 (2013) 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 308, + 481, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 308, + 481, + 341 + ], + "spans": [ + { + "bbox": [ + 138, + 308, + 481, + 341 + ], + "type": "text", + "content": "24. Dai, Z., Liu, H., Le, Q.V., Tan, M.: Coatnet: Marrying convolution and attention for all data sizes. Advances in neural information processing systems 34, 3965-3977 (2021) 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 342, + 481, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 342, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 138, + 342, + 481, + 385 + ], + "type": "text", + "content": "25. Dehghani, M., Djolonga, J., Mustafa, B., Padlewski, P., Heek, J., Gilmer, J., Steiner, A.P., Caron, M., Geirhos, R., Alabdulmohsin, I., et al.: Scaling vision transformers to 22 billion parameters. In: International Conference on Machine Learning. pp. 7480-7512. PMLR (2023) 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 387, + 481, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 387, + 481, + 419 + ], + "spans": [ + { + "bbox": [ + 138, + 387, + 481, + 419 + ], + "type": "text", + "content": "26. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018) 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 420, + 481, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 420, + 481, + 463 + ], + "spans": [ + { + "bbox": [ + 138, + 420, + 481, + 463 + ], + "type": "text", + "content": "27. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) 2, 4, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 464, + 481, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 464, + 481, + 508 + ], + "spans": [ + { + "bbox": [ + 138, + 464, + 481, + 508 + ], + "type": "text", + "content": "28. Drobyshev, N., Chelishev, J., Khakhulin, T., Ivakhnenko, A., Lempitsky, V., Zakharov, E.: Megaportraits: One-shot megapixel neural head avatars. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 2663-2671 (2022) 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 509, + 481, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 509, + 481, + 553 + ], + "spans": [ + { + "bbox": [ + 138, + 509, + 481, + 553 + ], + "type": "text", + "content": "29. Du, Y., Kips, R., Pumarola, A., Starke, S., Thabet, A., Sanakoyeu, A.: Avatars grow legs: Generating smooth human motion from sparse tracking inputs with diffusion model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 481-490 (2023) 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 554, + 481, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 554, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 138, + 554, + 481, + 597 + ], + "type": "text", + "content": "30. Eftekhar, A., Sax, A., Malik, J., Zamir, A.: Omnidata: A scalable pipeline for making multi-task mid-level vision datasets from 3d scans. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10786-10796 (2021) 12" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 599, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 599, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 599, + 481, + 632 + ], + "type": "text", + "content": "31. Eigen, D., Fergus, R.: Predicting depth, surface normals and semantic labels with a common multi-scale convolutional architecture. In: Proceedings of the IEEE international conference on computer vision. pp. 2650-2658 (2015) 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "text", + "content": "32. Eigen, D., Puhrsch, C., Fergus, R.: Depth map prediction from a single image using a multi-scale deep network. Advances in neural information processing systems 27 (2014) 4, 7" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 244, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 244, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 244, + 100 + ], + "type": "text", + "content": "R. Khirodkar et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 480, + 149 + ], + "type": "text", + "content": "33. El-Nouby, A., Izacard, G., Touvron, H., Laptev, I., Jegou, H., Grave, E.: Are large-scale datasets necessary for self-supervised pre-training? arXiv preprint arXiv:2112.10740 (2021) 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 151, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 151, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 138, + 151, + 480, + 183 + ], + "type": "text", + "content": "34. El-Nouby, A., Klein, M., Zhai, S., Bautista, M.A., Toshev, A., Shankar, V., Susskind, J.M., Joulin, A.: Scalable pre-training of large autoregressive image models. arXiv preprint arXiv:2401.08541 (2024) 2, 3, 4, 9" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 184, + 480, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 184, + 480, + 217 + ], + "spans": [ + { + "bbox": [ + 138, + 184, + 480, + 217 + ], + "type": "text", + "content": "35. Fang, H.S., Lu, G., Fang, X., Xie, J., Tai, Y.W., Lu, C.: Weakly and semi supervised human body part parsing via pose-guided knowledge transfer. arXiv preprint arXiv:1805.04310 (2018) 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 218, + 480, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 218, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 138, + 218, + 480, + 251 + ], + "type": "text", + "content": "36. Fang, H.S., Xie, S., Tai, Y.W., Lu, C.: Rmpe: Regional multi-person pose estimation. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 2334-2343 (2017) 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 251, + 480, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 251, + 480, + 274 + ], + "spans": [ + { + "bbox": [ + 138, + 251, + 480, + 274 + ], + "type": "text", + "content": "37. Fang, Y., Sun, Q., Wang, X., Huang, T., Wang, X., Cao, Y.: Eva-02: A visual representation for neon genesis. arXiv preprint arXiv:2303.11331 (2023) 9" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 274, + 480, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 274, + 480, + 318 + ], + "spans": [ + { + "bbox": [ + 138, + 274, + 480, + 318 + ], + "type": "text", + "content": "38. Fang, Y., Wang, W., Xie, B., Sun, Q., Wu, L., Wang, X., Huang, T., Wang, X., Cao, Y.: Eva: Exploring the limits of masked visual representation learning at scale. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19358-19369 (2023) 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 319, + 480, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 319, + 480, + 351 + ], + "spans": [ + { + "bbox": [ + 138, + 319, + 480, + 351 + ], + "type": "text", + "content": "39. Fouhey, D.F., Gupta, A., Hebert, M.: Data-driven 3d primitives for single image understanding. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 3392-3399 (2013) 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 352, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 352, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 138, + 352, + 480, + 385 + ], + "type": "text", + "content": "40. Gong, K., Liang, X., Li, Y., Chen, Y., Yang, M., Lin, L.: Instance-level human parsing via part grouping network. In: Proceedings of the European conference on computer vision (ECCV). pp. 770-785 (2018) 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 386, + 480, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 386, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 138, + 386, + 480, + 430 + ], + "type": "text", + "content": "41. Gong, K., Liang, X., Zhang, D., Shen, X., Lin, L.: Look into person: Self-supervised structure-sensitive learning and a new benchmark for human parsing. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 932-940 (2017) 3, 4, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 430, + 480, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 430, + 480, + 463 + ], + "spans": [ + { + "bbox": [ + 138, + 430, + 480, + 463 + ], + "type": "text", + "content": "42. Goyal, P., Caron, M., Lefaudeauux, B., Xu, M., Wang, P., Pai, V., Singh, M., Liptchinsky, V., Misra, I., Joulin, A., et al.: Self-supervised pretraining of visual features in the wild. arXiv preprint arXiv:2103.01988 (2021) 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 464, + 480, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 464, + 480, + 497 + ], + "spans": [ + { + "bbox": [ + 138, + 464, + 480, + 497 + ], + "type": "text", + "content": "43. Guizilini, V., Vasiljevic, I., Chen, D., Ambrus, R., Gaidon, A.: Towards zero-shot scale-aware monocular depth estimation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9233-9243 (2023) 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 498, + 480, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 498, + 480, + 542 + ], + "spans": [ + { + "bbox": [ + 138, + 498, + 480, + 542 + ], + "type": "text", + "content": "44. Halstead, M.A., Barsky, B.A., Klein, S.A., Mandell, R.B.: Reconstructing curved surfaces from specular reflection patterns using spline surface fitting of normals. In: Proceedings of the 23rd annual conference on Computer graphics and interactive techniques. pp. 335-342 (1996) 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 543, + 480, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 543, + 480, + 575 + ], + "spans": [ + { + "bbox": [ + 138, + 543, + 480, + 575 + ], + "type": "text", + "content": "45. He, K., Chen, X., Xie, S., Li, Y., Dollár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022) 3, 4, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 576, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 576, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 138, + 576, + 480, + 609 + ], + "type": "text", + "content": "46. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729-9738 (2020) 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 609, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 609, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 609, + 480, + 632 + ], + "type": "text", + "content": "47. He, K., Gkioxari, G., Dollár, P., Girshick, R.B.: Mask r-cnn. corr abs/1703.06870 (2017). arXiv preprint arXiv:1703.06870 (2017) 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "text", + "content": "48. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) 5" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 415, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 415, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 415, + 91, + 447, + 101 + ], + "type": "text", + "content": "Sapiens" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 480, + 149 + ], + "type": "text", + "content": "49. Hoffmann, J., Borgeaud, S., Mensch, A., Buchatskaya, E., Cai, T., Rutherford, E., Casas, D.d.L., Hendricks, L.A., Welbl, J., Clark, A., et al.: Training compute-optimal large language models. arXiv preprint arXiv:2203.15556 (2022) 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 149, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 149, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 138, + 149, + 480, + 182 + ], + "type": "text", + "content": "50. Hu, L., Gao, X., Zhang, P., Sun, K., Zhang, B., Bo, L.: animate anyone: Consistent and controllable image-to-video synthesis for character animation. arXiv preprint arXiv:2311.17117 (2023) 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 182, + 480, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 182, + 480, + 214 + ], + "spans": [ + { + "bbox": [ + 138, + 182, + 480, + 214 + ], + "type": "text", + "content": "51. Huang, S., Gong, M., Tao, D.: A coarse-fine network for keypoint localization. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 3028-3037 (2017) 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 214, + 480, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 214, + 480, + 246 + ], + "spans": [ + { + "bbox": [ + 138, + 214, + 480, + 246 + ], + "type": "text", + "content": "52. Jafarian, Y., Park, H.S.: Learning high fidelity depths of dressed humans by watching social media dance videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12753-12762 (2021) 4, 12" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 246, + 480, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 246, + 480, + 278 + ], + "spans": [ + { + "bbox": [ + 138, + 246, + 480, + 278 + ], + "type": "text", + "content": "53. Jiang, A.Q., Sablayrolles, A., Mensch, A., Bamford, C., Chaplot, D.S., Casas, D.d.l., Bressand, F., Lengyel, G., Lample, G., Saulnier, L., et al.: Mistral 7b. arXiv preprint arXiv:2310.06825 (2023) 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 278, + 480, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 278, + 480, + 311 + ], + "spans": [ + { + "bbox": [ + 138, + 278, + 480, + 311 + ], + "type": "text", + "content": "54. Jiang, T., Lu, P., Zhang, L., Ma, N., Han, R., Lyu, C., Li, Y., Chen, K.: Rtmpose: Real-time multi-person pose estimation based on mmpose. arXiv preprint arXiv:2303.07399 (2023) 9" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 311, + 480, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 311, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 138, + 311, + 480, + 354 + ], + "type": "text", + "content": "55. Jin, S., Xu, L., Xu, J., Wang, C., Liu, W., Qian, C., Ouyang, W., Luo, P.: Whole-body human pose estimation in the wild. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IX 16. pp. 196-214. Springer (2020) 3, 6, 9" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 354, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 354, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 138, + 354, + 480, + 374 + ], + "type": "text", + "content": "56. Kato, N., Li, T., Nishino, K., Uchida, Y.: Improving multi-person pose estimation using label correction. arXiv preprint arXiv:1811.03331 (2018) 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 374, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 374, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 138, + 374, + 480, + 407 + ], + "type": "text", + "content": "57. Khirodkar, R., Chari, V., Agrawal, A., Tyagi, A.: Multi-instance pose networks: Rethinking top-down pose estimation. In: Proceedings of the IEEE/CVF International conference on computer vision. pp. 3122-3131 (2021) 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 407, + 480, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 407, + 480, + 439 + ], + "spans": [ + { + "bbox": [ + 138, + 407, + 480, + 439 + ], + "type": "text", + "content": "58. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023) 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 439, + 480, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 439, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 138, + 439, + 480, + 460 + ], + "type": "text", + "content": "59. Kocabas, M., Chang, J.H.R., Gabriel, J., Tuzel, O., Ranjan, A.: Hugs: Human gaussian splats. arXiv preprint arXiv:2311.17910 (2023) 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 460, + 480, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 460, + 480, + 503 + ], + "spans": [ + { + "bbox": [ + 138, + 460, + 480, + 503 + ], + "type": "text", + "content": "60. Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., et al.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision 123, 32-73 (2017) 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 503, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 503, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 138, + 503, + 480, + 525 + ], + "type": "text", + "content": "61. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images. arXiv preprint (2009) 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 525, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 525, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 525, + 480, + 567 + ], + "type": "text", + "content": "62. Ladicky, L., Zeisl, B., Pollefeys, M.: Discriminatively trained dense surface normal estimation. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 468-484. Springer (2014) 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 568, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 568, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 138, + 568, + 480, + 601 + ], + "type": "text", + "content": "63. Lawrence, J., Goldman, D.B., Achar, S., Blascovich, G.M., Desloge, J.G., Fortes, T., Gomez, E.M., Haberling, S., Hoppe, H., Huibers, A., et al.: Project starline: A high-fidelity telepresence system. arXiv preprint (2021) 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 601, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 480, + 643 + ], + "type": "text", + "content": "64. Levoy, M., Pulli, K., Curless, B., Rusinkiewicz, S., Koller, D., Pereira, L., Ginzton, M., Anderson, S., Davis, J., Ginsberg, J., et al.: The digital michelangelo project: 3d scanning of large statues. In: Proceedings of the 27th annual conference on Computer graphics and interactive techniques. pp. 131-144 (2000) 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "65. Lewkowycz, A.: How to decay your learning rate. arXiv preprint arXiv:2103.12682 (2021) 8" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 244, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 244, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 244, + 100 + ], + "type": "text", + "content": "R. Khirodkar et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 137, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 137, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 137, + 116, + 480, + 138 + ], + "type": "text", + "content": "66. Li, Z., Wang, X., Liu, X., Jiang, J.: Binsformer: Revisiting adaptive bins for monocular depth estimation. arXiv preprint arXiv:2204.00987 (2022) 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "type": "text", + "content": "67. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: European conference on computer vision. pp. 740-755. Springer (2014) 1, 4, 6, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 173, + 480, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 173, + 480, + 217 + ], + "spans": [ + { + "bbox": [ + 138, + 173, + 480, + 217 + ], + "type": "text", + "content": "68. Liu, Z., Hu, H., Lin, Y., Yao, Z., Xie, Z., Wei, Y., Ning, J., Cao, Y., Zhang, Z., Dong, L., et al.: Swin transformer v2: Scaling up capacity and resolution. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12009-12019 (2022) 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 218, + 480, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 218, + 480, + 250 + ], + "spans": [ + { + "bbox": [ + 138, + 218, + 480, + 250 + ], + "type": "text", + "content": "69. Lombardi, S., Simon, T., Saragih, J., Schwartz, G., Lehrmann, A., Sheikh, Y.: Neural volumes: Learning dynamic renderable volumes from images. arXiv preprint arXiv:1906.07751 (2019) 1, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 251, + 480, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 251, + 480, + 283 + ], + "spans": [ + { + "bbox": [ + 138, + 251, + 480, + 283 + ], + "type": "text", + "content": "70. Lombardi, S., Simon, T., Schwartz, G., Zollhoefer, M., Sheikh, Y., Saragih, J.: Mixture of volumetric primitives for efficient neural rendering. ACM Transactions on Graphics (ToG) 40(4), 1-13 (2021) 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 285, + 480, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 285, + 480, + 306 + ], + "spans": [ + { + "bbox": [ + 138, + 285, + 480, + 306 + ], + "type": "text", + "content": "71. Long, J., Shelhamer, E., Darrell, T., Berkeley, U.: Fully convolutional networks for semantic segmentation. arxiv 2015. arXiv preprint arXiv:1411.4038 (2014) 10" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 308, + 480, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 308, + 480, + 329 + ], + "spans": [ + { + "bbox": [ + 138, + 308, + 480, + 329 + ], + "type": "text", + "content": "72. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983 (2016) 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 331, + 480, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 331, + 480, + 352 + ], + "spans": [ + { + "bbox": [ + 138, + 331, + 480, + 352 + ], + "type": "text", + "content": "73. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017) 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 353, + 480, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 353, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 138, + 353, + 480, + 373 + ], + "type": "text", + "content": "74. Lowe, D.G.: Three-dimensional object recognition from single two-dimensional images. Artificial intelligence 31(3), 355-395 (1987) 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 376, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 376, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 138, + 376, + 480, + 407 + ], + "type": "text", + "content": "75. Luo, Y., Zheng, Z., Zheng, L., Guan, T., Yu, J., Yang, Y.: Macro-micro adversarial network for human parsing. In: Proceedings of the European conference on computer vision (ECCV). pp. 418-434 (2018) 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 409, + 480, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 409, + 480, + 441 + ], + "spans": [ + { + "bbox": [ + 138, + 409, + 480, + 441 + ], + "type": "text", + "content": "76. Ma, S., Simon, T., Saragih, J., Wang, D., Li, Y., De La Torre, F., Sheikh, Y.: Pixel codec avatars. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 64-73 (2021) 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 443, + 480, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 443, + 480, + 485 + ], + "spans": [ + { + "bbox": [ + 138, + 443, + 480, + 485 + ], + "type": "text", + "content": "77. Mahajan, D., Girshick, R., Ramanathan, V., He, K., Paluri, M., Li, Y., Bharambe, A., Van Der Maaten, L.: Exploring the limits of weakly supervised pretraining. In: Proceedings of the European conference on computer vision (ECCV). pp. 181-196 (2018) 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 487, + 480, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 487, + 480, + 517 + ], + "spans": [ + { + "bbox": [ + 138, + 487, + 480, + 517 + ], + "type": "text", + "content": "78. Newell, A., Yang, K., Deng, J.: Stacked hourglass networks for human pose estimation. In: European conference on computer vision. pp. 483-499. Springer (2016) 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 520, + 480, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 520, + 480, + 563 + ], + "spans": [ + { + "bbox": [ + 138, + 520, + 480, + 563 + ], + "type": "text", + "content": "79. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023) 2, 3, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 565, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 565, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 138, + 565, + 480, + 609 + ], + "type": "text", + "content": "80. Papandreou, G., Zhu, T., Kanazawa, N., Toshev, A., Tompson, J., Bregler, C., Murphy, K.: Towards accurate multi-person pose estimation in the wild. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4903-4911 (2017) 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 609, + 480, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 609, + 480, + 642 + ], + "spans": [ + { + "bbox": [ + 138, + 609, + 480, + 642 + ], + "type": "text", + "content": "81. Peebles, W., Xie, S.: Scalable diffusion models with transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4195-4205 (2023) 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "82. render people: 3d people for architectural visualization | renderpeople, https://renderpeople.com/3d-people, accessed: 2024-02-22 3" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 415, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 415, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 415, + 91, + 447, + 101 + ], + "type": "text", + "content": "Sapiens" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + }, + { + "para_blocks": [ + { + "bbox": [ + 137, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 137, + 116, + 481, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 116, + 481, + 161 + ], + "spans": [ + { + "bbox": [ + 137, + 116, + 481, + 161 + ], + "type": "text", + "content": "83. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 162, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 162, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 137, + 162, + 481, + 205 + ], + "type": "text", + "content": "84. Ranftl, R., Lasinger, K., Hafner, D., Schindler, K., Koltun, V.: Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence 44(3), 1623-1637 (2020) 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 137, + 206, + 481, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 206, + 481, + 249 + ], + "spans": [ + { + "bbox": [ + 137, + 206, + 481, + 249 + ], + "type": "text", + "content": "85. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022) 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 251, + 481, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 251, + 481, + 294 + ], + "spans": [ + { + "bbox": [ + 137, + 251, + 481, + 294 + ], + "type": "text", + "content": "86. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al.: Imagenet large scale visual recognition challenge. International journal of computer vision 115, 211-252 (2015) 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 296, + 481, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 296, + 481, + 340 + ], + "spans": [ + { + "bbox": [ + 137, + 296, + 481, + 340 + ], + "type": "text", + "content": "87. Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022) 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 341, + 481, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 341, + 481, + 384 + ], + "spans": [ + { + "bbox": [ + 137, + 341, + 481, + 384 + ], + "type": "text", + "content": "88. Saito, S., Huang, Z., Natsume, R., Morishima, S., Kanazawa, A., Li, H.: Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 2304-2314 (2019) 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 137, + 385, + 481, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 385, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 137, + 385, + 481, + 430 + ], + "type": "text", + "content": "89. Saito, S., Simon, T., Saragih, J., Joo, H.: Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 84-93 (2020) 1, 4, 12" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 137, + 430, + 481, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 430, + 481, + 474 + ], + "spans": [ + { + "bbox": [ + 137, + 430, + 481, + 474 + ], + "type": "text", + "content": "90. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022) 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 137, + 475, + 481, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 475, + 481, + 518 + ], + "spans": [ + { + "bbox": [ + 137, + 475, + 481, + 518 + ], + "type": "text", + "content": "91. Singh, M., Duval, Q., Alwala, K.V., Fan, H., Aggarwal, V., Adcock, A., Joulin, A., Dollar, P., Feichtenhofer, C., Girshick, R., et al.: The effectiveness of mae pretraining for billion-scale pretraining. arXiv preprint arXiv:2303.13496 (2023) 2, 3, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 137, + 519, + 481, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 519, + 481, + 553 + ], + "spans": [ + { + "bbox": [ + 137, + 519, + 481, + 553 + ], + "type": "text", + "content": "92. Sun, C., Shrivastava, A., Singh, S., Gupta, A.: Revisiting unreasonable effectiveness of data in deep learning era. In: Proceedings of the IEEE international conference on computer vision. pp. 843-852 (2017) 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 137, + 554, + 481, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 554, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 137, + 554, + 481, + 586 + ], + "type": "text", + "content": "93. Sun, K., Xiao, B., Liu, D., Wang, J.: Deep high-resolution representation learning for human pose estimation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5693-5703 (2019) 4, 9" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 137, + 587, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 587, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 137, + 587, + 481, + 632 + ], + "type": "text", + "content": "94. Taori, R., Gulrajani, I., Zhang, T., Dubois, Y., Li, X., Guestrin, C., Liang, P., Hashimoto, T.B.: Alpaca: A strong, replicable instruction-following model. Stanford Center for Research on Foundation Models. https://crfm.stanford.edu/2023/03/13/alpaca.html 3(6), 7 (2023) 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 137, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 137, + 632, + 481, + 666 + ], + "type": "text", + "content": "95. Tay, Y., Dehghani, M., Rao, J., Fedus, W., Abnar, S., Chung, H.W., Narang, S., Yogatama, D., Vaswani, A., Metzler, D.: Scale efficiently: Insights from pretraining and fine-tuning transformers. arXiv preprint arXiv:2109.10686 (2021) 4" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "20" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 244, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 244, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 244, + 100 + ], + "type": "text", + "content": "R. Khirodkar et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 19 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 136, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 136, + 116, + 480, + 149 + ], + "type": "text", + "content": "96. Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023) 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 151, + 481, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 151, + 481, + 184 + ], + "spans": [ + { + "bbox": [ + 138, + 151, + 481, + 184 + ], + "type": "text", + "content": "97. Thomee, B., Shamma, D.A., Friedland, G., Elizalde, B., Ni, K., Poland, D., Borth, D., Li, L.J.: Yfcc100m: The new data in multimedia research. Communications of the ACM 59(2), 64-73 (2016) 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 185, + 481, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 185, + 481, + 217 + ], + "spans": [ + { + "bbox": [ + 138, + 185, + 481, + 217 + ], + "type": "text", + "content": "98. Toshev, A., Szegedy, C.: Deepest: Human pose estimation via deep neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1653-1660 (2014) 9" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 218, + 481, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 218, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 138, + 218, + 481, + 251 + ], + "type": "text", + "content": "99. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023) 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 134, + 251, + 480, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 251, + 480, + 284 + ], + "spans": [ + { + "bbox": [ + 134, + 251, + 480, + 284 + ], + "type": "text", + "content": "100. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023) 4, 5, 9" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 134, + 285, + 480, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 285, + 480, + 318 + ], + "spans": [ + { + "bbox": [ + 134, + 285, + 480, + 318 + ], + "type": "text", + "content": "101. Wang, X., Fouhey, D., Gupta, A.: Designing deep networks for surface normal estimation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 539-547 (2015) 3, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 319, + 480, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 319, + 480, + 361 + ], + "spans": [ + { + "bbox": [ + 133, + 319, + 480, + 361 + ], + "type": "text", + "content": "102. Weng, C.Y., Curless, B., Srinivasan, P.P., Barron, J.T., Kemelmacher-Shlizerman, I.: Humannerf: Free-viewpoint rendering of moving people from monocular video. In: Proceedings of the IEEE/CVF conference on computer vision and pattern Recognition. pp. 16210-16220 (2022) 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 363, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 363, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 133, + 363, + 480, + 385 + ], + "type": "text", + "content": "103. Wu, Y., Kirillov, A., Massa, F., Lo, W.Y., Girshick, R.: Detector2. https://github.com/facebookresearch/detectron2 (2019) 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 386, + 480, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 386, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 133, + 386, + 480, + 430 + ], + "type": "text", + "content": "104. Xia, F., Wang, P., Chen, L.C., Yuille, A.L.: Zoom better to see clearer: Human and object parsing with hierarchical auto-zoom net. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14. pp. 648-663. Springer (2016) 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 430, + 480, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 430, + 480, + 463 + ], + "spans": [ + { + "bbox": [ + 133, + 430, + 480, + 463 + ], + "type": "text", + "content": "105. Xia, F., Wang, P., Chen, X., Yuille, A.L.: Joint multi-person pose estimation and semantic part segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6769-6778 (2017) 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 464, + 480, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 464, + 480, + 497 + ], + "spans": [ + { + "bbox": [ + 133, + 464, + 480, + 497 + ], + "type": "text", + "content": "106. Xiao, B., Wu, H., Wei, Y.: Simple baselines for human pose estimation and tracking. In: Proceedings of the European conference on computer vision (ECCV). pp. 466-481 (2018) 4, 9" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 498, + 480, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 498, + 480, + 530 + ], + "spans": [ + { + "bbox": [ + 133, + 498, + 480, + 530 + ], + "type": "text", + "content": "107. Xie, E., Wang, W., Yu, Z., Anandkumar, A., Alvarez, J.M., Luo, P.: Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in neural information processing systems 34, 12077-12090 (2021) 10" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 531, + 480, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 531, + 480, + 563 + ], + "spans": [ + { + "bbox": [ + 133, + 531, + 480, + 563 + ], + "type": "text", + "content": "108. Xiu, Y., Yang, J., Cao, X., Tzionas, D., Black, M.J.: Econ: Explicit clothed humans optimized via normal integration. arXiv preprint arXiv:2212.07422 (2022) 1, 4, 12" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 565, + 480, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 565, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 133, + 565, + 480, + 597 + ], + "type": "text", + "content": "109. Xiu, Y., Yang, J., Tzionas, D., Black, M.J.: Icon: Implicit clothed humans obtained from normals. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13286-13296. IEEE (2022) 1, 12" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 599, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 599, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 133, + 599, + 480, + 632 + ], + "type": "text", + "content": "110. Xu, L., Jin, S., Liu, W., Qian, C., Ouyang, W., Luo, P., Wang, X.: Zoomnas: searching for whole-body human pose estimation in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(4), 5296-5313 (2022) 9" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 632, + 480, + 665 + ], + "type": "text", + "content": "111. Xu, Y., Zhang, J., Zhang, Q., Tao, D.: Vitpose: Simple vision transformer baselines for human pose estimation. Advances in Neural Information Processing Systems 35, 38571-38584 (2022) 6, 9" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 415, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 415, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 415, + 91, + 447, + 101 + ], + "type": "text", + "content": "Sapiens" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "21" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 20 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 512 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 148 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 148 + ], + "type": "text", + "content": "112. Xu, Y., Zhang, J., Zhang, Q., Tao, D.: Vitpose+: Vision transformer foundation model for generic body pose estimation. arXiv preprint arXiv:2212.04246 (2022) 9, 13" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 149, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 480, + 182 + ], + "type": "text", + "content": "113. Yang, L., Kang, B., Huang, Z., Xu, X., Feng, J., Zhao, H.: Depth anything: Unleashing the power of large-scale unlabeled data. arXiv preprint arXiv:2401.10891 (2024) 1, 3, 4, 11" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 480, + 215 + ], + "type": "text", + "content": "114. Yang, Z., Zeng, A., Yuan, C., Li, Y.: Effective whole-body pose estimation with two-stages distillation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4210-4220 (2023) 9" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 215, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 215, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 215, + 480, + 248 + ], + "type": "text", + "content": "115. Yang, Z., Dai, Z., Yang, Y., Carbonell, J., Salakhutdinov, R.R., Le, Q.V.: Xlnet: Generalized autoregressive pretraining for language understanding. Advances in neural information processing systems 32 (2019) 9" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 248, + 480, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 248, + 480, + 291 + ], + "spans": [ + { + "bbox": [ + 132, + 248, + 480, + 291 + ], + "type": "text", + "content": "116. Yin, Y., Guo, C., Kaufmann, M., Zarate, J.J., Song, J., Hilliges, O.: Hi4d: 4d instance segmentation of close human interaction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17016-17027 (2023) 10, 12" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 335 + ], + "type": "text", + "content": "117. Yu, T., Zheng, Z., Guo, K., Liu, P., Dai, Q., Liu, Y.: Function4d: Real-time human volumetric capture from very sparse consumer rgbd sensors. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5746-5756 (2021) 10, 12" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 335, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 335, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 335, + 480, + 369 + ], + "type": "text", + "content": "118. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3836-3847 (2023) 1, 14" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 369, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 413 + ], + "type": "text", + "content": "119. Zhang, S.H., Li, R., Dong, X., Rosin, P., Cai, Z., Han, X., Yang, D., Huang, H., Hu, S.M.: Pose2seg: Detection free human instance segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 889-898 (2019) 1, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "type": "text", + "content": "120. Zhang, X., Lu, Y., Wang, W., Yan, A., Yan, J., Qin, L., Wang, H., Yan, X., Wang, W.Y., Petzold, L.R.: Gpt-4v (isdiction) as a generalist evaluator for vision-language tasks. arXiv preprint arXiv:2311.01361 (2023) 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 446, + 480, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 480, + 478 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 480, + 478 + ], + "type": "text", + "content": "121. Zhou, J., Wei, C., Wang, H., Shen, W., Xie, C., Yuille, A., Kong, T.: ibot: Image bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832 (2021) 3, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 479, + 480, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 512 + ], + "type": "text", + "content": "122. Zlateski, A., Jaroensri, R., Sharma, P., Durand, F.: On the importance of label quality for semantic segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1479-1487 (2018) 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "22" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 244, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 244, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 244, + 100 + ], + "type": "text", + "content": "R. Khirodkar et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 21 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Scalable Group Choreography via Variational Phase Manifold Learning/6dc9e2ab-125f-402e-97a2-dfb708ce273d_content_list.json b/2024/Scalable Group Choreography via Variational Phase Manifold Learning/6dc9e2ab-125f-402e-97a2-dfb708ce273d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7526b4899fb464aca362a48e5f5e5182c401f4cc --- /dev/null +++ b/2024/Scalable Group Choreography via Variational Phase Manifold Learning/6dc9e2ab-125f-402e-97a2-dfb708ce273d_content_list.json @@ -0,0 +1,1677 @@ +[ + { + "type": "text", + "text": "Scalable Group Choreography via Variational Phase Manifold Learning", + "text_level": 1, + "bbox": [ + 241, + 140, + 759, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Nhat $\\mathrm{Le}^{1}$ , Khoa $\\mathrm{Do}^{2,3}$ , Xuan Bui $^{2,3}$ , Tuong $\\mathrm{Do}^{1,4}$ , Erman Tjiputra $^{1}$ , Quang D. Tran $^{1}$ , and Anh Nguyen $^{4}$", + "bbox": [ + 225, + 210, + 774, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 AIOZ, Singapore", + "bbox": [ + 436, + 253, + 565, + 268 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "2 University of Science, Ho Chi Minh City, Vietnam", + "3 Vietnam National University, Ho Chi Minh City, Vietnam", + "4 University of Liverpool, Liverpool L69 3BX, United Kingdom" + ], + "bbox": [ + 290, + 268, + 710, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Generating group dance motion from the music is a challenging task with several industrial applications. Although several methods have been proposed to tackle this problem, most of them prioritize optimizing the fidelity in dancing movement, constrained by predetermined dancer counts in datasets. This limitation impedes adaptability to real-world applications. Our study addresses the scalability problem in group choreography while preserving naturalness and synchronization. In particular, we propose a phase-based variational generative model for group dance generation on learning a generative manifold. Our method achieves high-fidelity group dance motion and enables the generation with an unlimited number of dancers while consuming only a minimal and constant amount of memory. The intensive experiments on two public datasets show that our proposed method outperforms recent state-of-the-art approaches by a large margin and is scalable to a great number of dancers beyond the training data.", + "bbox": [ + 259, + 348, + 738, + 555 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 580, + 374, + 595 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The widespread availability of digital social media platforms has led to a significant increase in the popularity of creating dance videos. This heightened interest has resulted in the daily production and viewing of millions of dance videos on various online platforms [13, 32]. Recently, researchers from the computer vision and computer graphics community have focused on developing methods for generating authentic dance movements in response to music [6, 94]. These advancements have broad implications and find applications in diverse areas, including animation [49, 50], virtual idols [59], virtual metaverses [42], and dance education [2, 67, 72]. Such techniques empower artists, animators, and educators by providing powerful tools to enhance their creative pursuits and improve the overall dance experience for performers and audiences alike.", + "bbox": [ + 212, + 613, + 787, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, significant strides have been made in creating dance motions for solo dance performers [12,26,34,49,59,71,83]. However, generating synchronized group dance movements that are both lifelike and in harmony with music remains a complex challenge [9,27,87]. The introduction of AIOZ-GDance [40] stands as", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9a90bad30506cda01255b6a5143c3095f8e6726525ee39c3f98d8b43d43c3176.jpg", + "image_caption": [ + "Fig. 1: We present a new group dance generation method that can generate a large number of dancers within a fixed resource consumption. The illustration shows a generated group dance sample with 100 dancers." + ], + "image_footnote": [], + "bbox": [ + 230, + 143, + 408, + 224 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/61ccfb172c209e5731fa05a7051c23fa14c514d12c91f6ea07bc1bb086dc70a5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 411, + 143, + 589, + 224 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/cb1cbcb4bd334846ab2e10e79ee6ef277429dbb37d9d038b4efcefea77391613.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 591, + 143, + 771, + 224 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the extensive dataset to aid in generating group choreography. Different aspects of group dance motion such as the study of consistency and diversity among the movements of dancers are explored in [39]. However, despite recent advances in the field, existing methods can either only generate dances for a limited number of dancers given the input audio [40, 86, 90], or would consume an expensively large amount of memory due to their design nature [39, 40]. Typically, these methods are usually based on specialized collaborative mechanisms such as cross-entity attention [40, 51], group global attention [39], or residual communication block [66] achieve coherent multi-person synthesis. In other words, their innate architectural design requires the network to process and synthesize all motions simultaneously, leading to excessively growing computational/memory costs with the increasing number of inputs, and hence extremely difficult to scale up.", + "bbox": [ + 212, + 305, + 787, + 487 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite the comprehensive exploration of group dance motion generation, the predominant emphasis remains on optimizing the execution of the generated movements [14, 19, 31, 37, 52, 93]. These existing methodologies function within a confined scope, often tethered to a predetermined maximum number of dancers depicted in group choreography videos outlined by the predefined dataset [39]. This inherent limitation poses a potential constraint on the adaptability of trained models when transitioning to real-world applications [47]. In our pursuit, we center our efforts on addressing the critical aspect of scalability in group dance. Our primary objective revolves around not only expanding the number of dancers but also ensuring the preservation of the innate naturalness in the dance motions and the seamless synchronization between dancers throughout the entirety of their performance.", + "bbox": [ + 212, + 489, + 787, + 671 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the field of character motion control and synthesis, existing approaches can be broadly divided into two categories: deterministic [16, 26, 40, 49, 55, 86] and probabilistic [19, 39, 59, 61, 83]. The former tries to learn direct and deterministic mapping from the input conditioning signal (such as text, audio, or user controls) to the desired output motion. However, these methods usually regress towards the mean pose as they are typically trained by minimizing a regression objective under different outputs for similar inputs, resulting in freezing or drifting motion. On the other hand, probabilistic methods attempt to capture the distribution over possible motions for a given condition, allowing diverse and high-quality generation. In particular, Variational Autoencoder (VAE) [18, 47, 60, 61], or Diffusion Models [4, 39, 81, 83, 97] have been widely", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "N. Le et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "used. Although diffusion models have recently shown strong potential in generating high-quality motions [81], they generally operate on high-dimensional space (same as the original data) [39,81,83], hence rendering them extremely challenging to use for scalable group generation. Meanwhile, standard VAE models often rely on a Gaussian latent space of fixed dimension for sampling a single latent vector to generate motions, making it difficult to adapt to a wide range of architectures. Besides, the fixed-size single latent space might also not be sufficient to represent the temporal dynamics of the whole motion sequence [7,60].", + "bbox": [ + 212, + 146, + 782, + 267 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent studies have demonstrated the significant advantages of learning motion features in the frequency domain. Specifically, [22, 74-76] have found that different motion skills can be effectively represented by several phase variables that can faithfully capture the spatial-temporal alignment of a wide range of movements. This can be used to combine with common deep motion synthesis networks to enhance its capability, facilitating natural, and stable motion generations by enforcing a unidirectional motion transition (i.e., no backward direction in time) to prevent the motion from being stuck temporally [74]. Unlike those prior works, where the phases are utilized as auxiliary signals to assist in the motion synthesis process [68, 74, 88], we propose to learn to generate the phases conditioned on the input audio signal in an end-to-end manner, eliminating the need of training extra components. Our generative phase manifold can be used to generate new motions efficiently, from which the latent motion curves are sampled in frequency domain and decoded into the original motion space by a learned decoder.", + "bbox": [ + 212, + 272, + 784, + 498 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, our goal is to develop a scalable technique for group dance generation, a phase-based variational generative model for scalable group dance generation, namely Phase-conditioned Dance VAE (PDVAE). To our knowledge, PDVAE is the first method to represent the variational latent space using phase parameters in the frequency domain of the motion curves. Our method goes beyond the conventional VAE approach that typically relies on a single latent vector drawn from a Gaussian distribution, which is unable to adequately represent the temporal information of the motion sequence (e.g., the time dimension is squeezed out). Figure 1 illustrates an example of a dance motion sequence with 100 dancers generated by our model. In contrast to existing methods, our model can generate crowd-dance animations for an unlimited number of dancers without increasing computational burden (i.e., memory consumption remains constant), while still maintaining the performance and fidelity of the generated choreographies. To summarize, our key contributions are as follows:", + "bbox": [ + 212, + 503, + 785, + 715 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce PDVAE, a phase-based variational generative model for scalable group dance generation. The method focuses on generating large-scale group dance under limited resources.", + "- To effectively learn the manifold that is group-consistent (i.e., dancers within a group lie upon the same manifold), we propose a group consistency loss that enforces the networks to encode the latent phase manifold to be identical for the same group given the input music." + ], + "bbox": [ + 223, + 729, + 784, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Scalable Group Choreography via Variational Phase Manifold Learning", + "bbox": [ + 256, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- Extensive experiments along with thorough user study evaluations demonstrate the state-of-the-art performance of our model while achieving effective scalability.", + "bbox": [ + 223, + 146, + 785, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 210, + 387, + 226 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.1 Music-driven Choreography", + "text_level": 1, + "bbox": [ + 215, + 237, + 491, + 253 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Crafting natural human choreography derived from music presents a multifaceted challenge [30,62,89]. Certain methodologies integrate constraints based on music-motion similarity matching to ensure coherence between the generated motion and the music [10,33,35,43,65,69]. However, many of these approaches rely considerably on heuristic algorithms, stitching together pre-existing dance segments sourced from a limited music-dance database [10, 48]. While successful in generating extended and realistic dance sequences, these methods face limitations when endeavoring to create entirely novel dance fragments [57, 90].", + "bbox": [ + 212, + 257, + 785, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the recent period, advancements have been evident in the domain of converting music into dance movements through various techniques such as Convolutional Networks (CNN) [1,8,77,91,92,99], Recurrent Networks (RNN) [3,26,77,79,87], Graph Neural Networks (GNN) [5,12,64,95], Generative Adversarial Networks (GAN) [41,77], or Transformer models [34,44,45,49,59,71]. These methodologies typically depend on various inputs, including the present music and a concise history of previous dance movements, to forecast forthcoming sequences of human poses. This innovation introduces a music-text feature fusion module, engineered to amalgamate inputs into a motion decoder, thereby enabling the creation of dance sequences conditioned upon both musical and textual instructions [17].", + "bbox": [ + 212, + 378, + 785, + 545 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While these techniques show promise in generating authentic and lifelike dance motions, they often struggle to synchronize movements seamlessly across multiple dancers [40]. Specifically, achieving coordination and harmony among dancers necessitates consideration of spatial and temporal relationships, their interactions, and the overall choreographic structure [84]. Consequently, further advancements in this domain are being pursued to tackle these complexities [25, 47, 59, 98]. For instance, Perez et al. [59] integrate a multimodal transformer encoder with a normalizing-flow-based decoder to estimate a probability distribution encompassing potential subsequent poses. Additionally, Feng et al. [11] enable long-term generation by imposing a motion repeat constraint to forecast future frames while considering historical motions. A recent study by Le et al. [39] investigates consistency and diversity factors between the generated motions of two or more dancers within a timeframe. However, these approaches are limited to a predefined number of dancers, heavily constrained by the maximum number of dancers in the dataset-provided videos.", + "bbox": [ + 212, + 545, + 785, + 772 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Motion Manifold Learning", + "text_level": 1, + "bbox": [ + 215, + 787, + 480, + 805 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Motion manifold learning has attracted considerable attention in computer vision and artificial intelligence, with the primary goal of comprehending the funda-", + "bbox": [ + 215, + 809, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "N. Le et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "mental structures inherent in human movement and dynamics [16, 24, 56, 82]. Its distinctive ability to generate human movement patterns presents numerous opportunities to comprehend intrinsic motion dynamics, manage nonlinear relationships in motion data, and acquire contextual and hierarchical representations [29, 63, 88]. Consequently, various methodologies have emerged, each contributing distinct perspectives and techniques to advance the comprehension and synthesis of human motion [28, 38, 63, 68].", + "bbox": [ + 212, + 146, + 782, + 252 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For instance, Holden et al. [23] generate character movements by mapping high-level parameters to the human motion manifold, allowing diverse movements without manual preprocessing, and enabling post-generation editing for natural, smooth motion sequences. MotionCLIP [80] introduces a 3D human motion auto-encoder aligned with the Contrastive Language-Image Pre-training (CLIP) model's space, enabling semantic text-based motion generation, disentangled editing, and abstract language specification, leveraging CLIP's rich semantic knowledge within the motion manifold. Recently, Sun et al. [78] utilize a VQ-VAE to acquire a low-dimensional manifold, effectively cleansing the motion sequences. Regarding the scalability of group dance generation, the concept of the motion manifold emerges as a potential solution for addressing the restricted number of dancers in the dataset. This direction enables the learning of a distribution that facilitates the extraction of dance motions, allowing for the presentation of cohesive group dance sequences.", + "bbox": [ + 212, + 252, + 787, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Variational Phase Manifold Learning for Scalable Group Choreography", + "text_level": 1, + "bbox": [ + 214, + 487, + 733, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Task Definition", + "text_level": 1, + "bbox": [ + 214, + 537, + 388, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given an input music sequence $\\mathbf{a} = \\{a_1, a_t, \\dots, a_T\\}$ with $t = \\{1, \\dots, T\\}$ indicates the index of the music frames, our goal is to generate the group motion sequences of $N$ arbitrary dancers: $\\mathbf{x} = \\{x_1^1, \\dots, x_T^1; \\dots; x_1^N, \\dots, x_T^N\\}$ where $x_t^n$ is the pose of $n$ -th dancer at frame $t$ . We follow [19, 53] and represent dance as a sequence of poses in the 24-joint of the SMPL body model [54], using the 6D continuous rotation [96] for every joint, along with 3D joint positions and velocities. Additionally, the corresponding 3D root translation vectors are concatenated into the pose representations to involve the trajectory of motion. Previous group dance methods [40], which generate the whole group at once, cannot deal with the increasing number of dancers and can only create group sequences up to a pre-defined number of dancers, due to the vast complexity of the architecture. In contrast, we aim to generate group dance with an unlimited number of dancers.", + "bbox": [ + 212, + 564, + 787, + 746 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Phase-conditioned Dance VAE", + "text_level": 1, + "bbox": [ + 214, + 768, + 516, + 782 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our goal is to learn a continuous manifold such that the motion can be generated by sampling from this learned manifold. We assume that although different dancers within the same group may present visually distinctive movements, the", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Scalable Group Choreography via Variational Phase Manifold Learning", + "bbox": [ + 256, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a1e8235f870389f41b9939bcda93158b4a7097b0a339e531cc87b01f740444d5.jpg", + "image_caption": [ + "Fig. 2: Overview of our Phase-conditioned Dance VAE (PDVAE) for scalable group dance generation. It consists of an Encoder, a Prior, and a Decoder network. During training, we encode the corresponding motion and music inputs into a latent phase manifold, which is variational and parameterized by the frequency domain parameters of periodic functions. The latent phases can be sampled from the manifold and then decoded back to the original data space to obtain new motions. The consistency loss $\\mathcal{L}_{\\mathrm{csc}}$ is further imposed to constrain the manifold to be consistently unified for dancers that belong to the same group. At inference stage, only the Prior and the Decoder are used to synthesize group dances efficiently." + ], + "image_footnote": [], + "bbox": [ + 222, + 145, + 772, + 368 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "properties of their motions, such as timing, periodicity, or temporal alignment are intrinsically similar. Drawing inspiration from [74], we aim to learn a generative phase representation for each group of dancer in order to synthesize their motion indefinitely. Our generative model is built upon the conditional Variational Autoencoder architecture [73], thanks to its diverse generation capability and fast sampling speed. However, instead of directly encoding the data into a Gaussian latent distribution as in common VAE approaches [15, 41, 47, 60, 61], we model the latent variational distribution by the phase parameters extracted from the latent motion curve, which we call variational phase manifold. The latent phase manifold is well-structured and can well describe key characteristics of motion (such as its timing, local periodicity, and transition), which benefits learning motion features [74].", + "bbox": [ + 212, + 532, + 787, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The overview of our Phase-conditioned Dance VAE is illustrated in Figure 2. Specifically, the model contains three main networks: an encoder $\\mathcal{E}$ to capture the approximate posterior distribution conditioned on both motion and music $q_{\\phi}(\\mathbf{z}|\\mathbf{x},\\mathbf{a})$ , a prior network $\\mathcal{P}$ to learn the conditional prior given only the music $p_{\\theta}(\\mathbf{z}|\\mathbf{a})$ , and a decoder $\\mathcal{D}$ to learn to reconstruct the data from the latent distribution $p_{\\theta}(\\mathbf{x}|\\mathbf{z},\\mathbf{a})$ . The new motion is generated by sampling the frequency-domain parameters predicted by the prior network, which is then passed through the decoder network to reconstruct the motion in the original data space. Fur-", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "N. Le et al.", + "bbox": [ + 271, + 114, + 349, + 126 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "thermore, we adopt Transformer-based architecture [85] in each network to effectively capture long-range dependencies and holistic context of the dance sequence.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Encoder. The encoder $\\mathcal{E}$ is expected to take both the motion and music feature sequence as input, and produce a distribution over possible latent variables capturing the cross-modal relationship between them. To transform the joint input space into a learned phase manifold, we adopt the Transformer decoder architecture where the Cross-Attention mechanism [85] is utilized to learn the relationship between the motion and the music. Accordingly, the output of the encoder is a batch of latent curves (i.e., the activation sequences per channel) that can particularly capture different spatial and temporal aspects of the motion sequence. However, instead of training the model to directly reconstruct the input motion from the extracted latent curves, we further enforce each channel of the latent space to have a periodic functional form (i.e., sinusoidal). This enables us to effectively learn a compact parameterization for each latent channel from a small set of parameters in the frequency domain.", + "bbox": [ + 212, + 213, + 784, + 411 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Generative Variational Phase Manifold. Here we focus on learning the periodicity and non-linear temporal alignment of the motion in the latent space. In particular, given the output latent curves from the encoder $\\mathbf{L} = \\mathcal{E}(x,a)\\in \\mathbb{R}^{D\\times T}$ with $D$ is the number of desired phase channels to be extracted from the motion, we parameterize each latent curve in $\\mathbf{L}$ using a sinusoidal function with amplitude (A), frequency (F), offset (B) and phase shift (S) parameters [74]. To allow for variational phase manifold learning, we opt to predict two sets of parameters $\\mu_{\\varepsilon} = \\{\\mu^{A};\\mu^{F};\\mu^{B};\\mu^{S}\\}$ and $\\sigma_{\\varepsilon} = \\{\\sigma^{A};\\sigma^{F};\\sigma^{B};\\sigma^{S}\\}$ , which corresponds to the mean and variance of $\\mathbb{R}^{4D}$ dimensional Gaussian distribution:", + "bbox": [ + 212, + 431, + 782, + 566 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nq _ {\\phi} (\\mathbf {z} | \\mathbf {x}, \\mathbf {a}) = \\mathcal {N} (\\mathbf {z}; \\mu_ {\\varepsilon}, \\sigma_ {\\varepsilon}) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 411, + 580, + 784, + 595 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To do so, we first apply differentiable Fast Fourier Transform (FFT) to each channel of the latent curve $\\mathbf{L}$ and create the zero-indexed matrix of Fourier coefficients as $\\mathbf{c} = FFT(\\mathbf{L})$ with $\\mathbf{c} \\in \\mathbb{C}^{D \\times K + 1}$ , $K = \\left\\lfloor \\frac{T}{2} \\right\\rfloor$ . Correspondingly, we compute the per channel power spectrum $\\mathbf{p} \\in \\mathbb{R}^{D \\times K + 1}$ as $\\mathbf{p}_{i,j} = \\frac{2}{N} |\\mathbf{c}_{i,j}|^2$ , where $i$ is the channel index and $j$ is the index for the frequency bands. Correspondingly, the distributional mean parameters of the periodic sinusoidal function are then calculated as follows:", + "bbox": [ + 212, + 606, + 782, + 710 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mu_ {i} ^ {A} = \\sqrt {\\frac {2}{T} \\sum_ {j = 1} ^ {K} \\mathbf {p} _ {i , j}}, \\quad \\mu_ {i} ^ {F} = \\frac {\\sum_ {j = 1} ^ {K} \\mathbf {f} _ {j} \\cdot \\mathbf {p} _ {i , j}}{\\sum_ {j = 1} ^ {K} \\mathbf {p} _ {i , j}}, \\quad \\mu_ {i} ^ {B} = \\frac {\\mathbf {c} _ {i , 0}}{T}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 305, + 720, + 784, + 767 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathbf{f} = (0, \\frac{1}{T}, \\dots, \\frac{K}{T})$ is the frequencies vector. At the same time, the phase shift $\\mathbf{S}$ is predicted using a fully-connected (FC) layer with two arctan activation as:", + "bbox": [ + 212, + 777, + 782, + 821 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\left(s _ {y}, s _ {x}\\right) = \\operatorname {F C} \\left(\\mathbf {L} _ {i}\\right), \\quad \\mu_ {i} ^ {S} = \\arctan \\left(s _ {y}, s _ {x}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 356, + 824, + 784, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Scalable Group Choreography via Variational Phase Manifold Learning", + "bbox": [ + 256, + 114, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To predict the distributional variance of the phase amplitude and phase shift parameters $\\{\\sigma^A,\\sigma^S\\}$ , We additionally apply a separate two-layer MLP network over each channel of the latent curves, similar to Equation 3. The variational latent phase parameters are sampled by utilizing parameterization trick [36], i.e., $\\mathbf{A} \\sim \\mathcal{N}(\\mu^A,\\sigma^A)$ and $\\mathbf{S} \\sim \\mathcal{N}(\\mu^S,\\sigma^S)$ . In our experiments, we find that sampling the phase frequency $\\mathbf{F}$ and offset $\\mathbf{B}$ often produce unstable and noncoherent group movements. This might be because the frequency amplitudes of the dancers within the same group are likely to associate with the rhythmic pattern of the musical beats while the offsets capture their alignment, thereby should be consistent with each other. Therefore, we treat those parameters as deterministic by constraining their variance to zero.", + "bbox": [ + 212, + 146, + 787, + 313 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, the sampled set of phase parameters $\\mathbf{z} = \\{\\mathbf{A};\\mathbf{F};\\mathbf{B};\\mathbf{S}\\}$ are used to reconstruct a parametric latent space consisting of multiple periodic curves to represent each intrinsic property of the motion by:", + "bbox": [ + 212, + 316, + 787, + 364 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {L}} = \\mathbf {A} \\cdot \\sin (2 \\pi \\cdot (\\mathbf {F} \\cdot \\mathcal {T} - \\mathbf {S})) + \\mathbf {B} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 387, + 784, + 402 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathcal{T}$ is a known time window series obtained by evenly spacing the timesteps from 0 to $T$ . Intuitively this curve construction procedure can be viewed as a \"quantization\" layer to enforce the network to learn to represent the motion features in the frequency domain, which is useful in representing different aspects of human motion such as their timing and periodicity. In the last step, a decoder is utilized to reconstruct the original motion signals from the set of parametric latent curves.", + "bbox": [ + 212, + 426, + 787, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Decoder. To decode the latent space into the original motion space, previous works [60,61] have to use a sinusoidal positional encoding sequence with duration $T$ as the proxy input to the sequence decoder. This is because their latent space is only formed by single latent vectors following a Gaussian distribution, which cannot span the time dimension. However, we observe that it usually results in unstable and inconsistent movements, as the proxy sequence is generic and usually contains less meaningful information for the decoder. Meanwhile, our method does not suffer from this problem as our latent space is built on multiple curves that can represent the motion information through time, thanks to the phase parameters. Subsequently, our decoder $\\mathcal{D}$ is based on Transformer decoder architecture that takes the constructed parametric latent curve, as well as the music features as inputs, to reconstruct the corresponding dance motions. Here, we also utilize the cross-attention model [85] where we consider the sequence of and music features as key and value along with the sampled latent curves as the query. The output of the decoder is a sequence of $T$ vectors in $\\mathbb{R}^D$ , which is then projected back to the original motion dimensions through a linear layer, to obtain the reconstructed outputs $\\hat{\\mathbf{x}} = p_{\\theta}(\\mathbf{x}|\\mathbf{z},\\mathbf{a})$ . We additionally employ a global trajectory predictor to predict the global translation of the root joint based on the generated local motions [19, 46], in order to avoid intersection problems between dancers. We provide details of our global predictor in the supplemental.", + "bbox": [ + 212, + 537, + 787, + 843 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "N. Le et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Prior Network. Since the ground-truth motion is generally inaccessible at test time (i.e., we only have access to the music), we also need to learn a prior $\\mathcal{P}$ to match the posterior distribution of motion from which the latent phase can be sampled. Specifically, We follow the procedure similar to Section 3.2 (Equation 1-4) to predict the Gaussian distribution conditioned on the music sequence $\\mathbf{a}$ , which is then used for sampling the latent phases:", + "bbox": [ + 214, + 146, + 782, + 237 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} (\\mathbf {z} | \\mathbf {a}) = \\mathcal {N} (\\mathbf {z}; \\mu_ {\\mathcal {P}}, \\sigma_ {\\mathcal {P}}) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 416, + 246, + 785, + 263 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where a Transformer encoder is used to encode the input conditioning music sequence and predict the corresponding $\\mu_{\\mathcal{P}}$ and $\\sigma_{\\mathcal{P}}$ . We implement the prior network similarly to the encoder network, however, we use self-attention mechanism [85] to capture the global music context. Learning the conditional prior is crucial for the conditional VAE to generalize to diverse types of music and motion. Intuitively speaking, each latent variable $\\mathbf{z}$ is expected to represent possible dance motions $\\mathbf{x}$ conforming to the music context $\\mathbf{a}$ . Therefore, the prior should be able to encode different latent distributions given different musics.", + "bbox": [ + 214, + 272, + 784, + 393 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.3 Training", + "text_level": 1, + "bbox": [ + 215, + 414, + 333, + 429 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "During training, we consider the following variational lower bound [73] to mainly train our dance generation VAE model:", + "bbox": [ + 214, + 436, + 782, + 467 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\log p _ {\\theta} (\\mathbf {x} \\mid \\mathbf {a}) \\geq \\mathbb {E} _ {q _ {\\phi}} \\left[ \\log p _ {\\theta} (\\mathbf {x} \\mid \\mathbf {z}, \\mathbf {a}) \\right] - D _ {\\mathrm {K L}} \\left(q _ {\\phi} (\\mathbf {z} \\mid \\mathbf {x}, \\mathbf {a}) \\| p _ {\\theta} (\\mathbf {z} \\mid \\mathbf {a})\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 292, + 476, + 784, + 494 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In practice, we apply the conditional VAE loss as similar to [61], which is defined as the weighted sum $\\mathcal{L}_{\\mathrm{cvae}} = \\mathcal{L}_{\\mathrm{rec}} + \\lambda_{\\mathrm{KL}}\\mathcal{L}_{\\mathrm{KL}}$ . In particular, the reconstruction term $\\mathcal{L}_{\\mathrm{rec}}$ measures the motion reconstruction error given the decoder output (via a smooth-L1 loss). The KL divergence term $\\mathcal{L}_{\\mathrm{KL}}$ compares the divergence $D_{\\mathrm{KL}}$ between the posterior and the prior distribution to enforce them to be close to each other.", + "bbox": [ + 214, + 502, + 784, + 590 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The conditional VAE objective above is calculated for each dancer separately and cannot capture the correlation between dancers within a group. Therefore, it is essential to impose consistency among dancers and avoid strange effects such as unsynchronized dance. To this end, we propose a group consistency loss by enforcing the latent phase manifold to be similar for the same group, given the input music. Specifically, we first calculate the phase manifold features based on the frequency domain parameters as follows:", + "bbox": [ + 214, + 593, + 784, + 699 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} _ {2 i - 1} = \\mathbf {A} _ {i} \\sin (2 \\pi \\cdot \\mathbf {S} _ {i}), \\quad \\mathbf {P} _ {2 i} = \\mathbf {A} _ {i} \\cos (2 \\pi \\cdot \\mathbf {S} _ {i}) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 323, + 709, + 784, + 724 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\mathbf{P} \\in \\mathbb{R}^{2D}$ is the phase manifold vector that encodes the spatial-temporal information of the motion state. The phase feature may look similar to the positional encodings of transformers [85] in the sense that both use multi-resolution sinusoidal functions. However, the phase feature is much richer in terms of representation capacity since it learns to embed the spatial (body joints) and temporal (positions in time) information of the motion curves, whereas the positional encodings only encode the position of words. Finally, our consistency objective is", + "bbox": [ + 214, + 733, + 785, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Scalable Group Choreography via Variational Phase Manifold Learning", + "bbox": [ + 256, + 114, + 732, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "to constrain phase manifold between dancers within a group to be as close as possible to each other, which is formulated as:", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {c s c}} = D _ {\\mathrm {K L}} \\left(q _ {\\phi} (\\mathbf {z} | \\mathbf {x} ^ {m}, \\mathbf {a}) \\| \\left(q _ {\\phi} (\\mathbf {z} | \\mathbf {x} ^ {n}, \\mathbf {a})\\right) + \\| \\mathbf {P} ^ {m} - \\mathbf {P} ^ {n} \\| _ {2} ^ {2} \\right. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 308, + 189, + 785, + 208 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where the first term encourages the network to map different dancers belonging to the same group $(\\mathbf{x}^m$ and $\\mathbf{x}^n)$ into the same set of distributional phase parameters while the second term penalizes the discrepancy in their corresponding phase manifolds. In general, this loss is applied to ensure every dancer is embedded into a single unified manifold that can effectively represent their corresponding group. To summarize, our total training loss is defined as the combination of the VAE loss and the consistency loss $\\mathcal{L} = \\mathcal{L}_{\\mathrm{cvae}} + \\lambda_{\\mathrm{csc}}\\mathcal{L}_{\\mathrm{csc}}$ .", + "bbox": [ + 212, + 220, + 782, + 325 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "For testing, we can efficiently generate motions for an unlimited number of dancers by indefinitely drawing samples from the learned continuous group-consistent phase manifold. It is noteworthy that for inference, we only need to process the prior network once to obtain the latent distribution. To generate a new motion, we can sample from this latent (Gaussian) distribution and use the decoder to decode it back to the motion space. This approach is much more efficient and has significantly higher scalability than previous approaches [39, 40] that is limited by the number of dancers processed simultaneously by the entire network.", + "bbox": [ + 212, + 327, + 784, + 460 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 491, + 375, + 508 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Implementation Details", + "text_level": 1, + "bbox": [ + 214, + 527, + 455, + 541 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Our model was trained on 4 NVIDIA V100 GPUs using Adam optimizer with a fixed learning rate of $10^{-4}$ and a mini-batch size of 32 per GPU. For training losses, the weights are empirically set to $\\lambda_{\\mathrm{KL}} = 5\\times 10^{-4}$ and $\\lambda_{\\mathrm{csc}} = 10^{-4}$ , respectively. The Transformer encoders and decoders consist of 5 layers for both encoder, decoder, and prior Network with 512-dimensional hidden units. Meanwhile, the number of latent phase channels is set to 256. To further capture the periodic nature of the phase feature, we also use Siren activation following the initialization scheme as in [70]. This can effectively model the periodicity inherent in the motion data, and thus can benefit motion synthesis [68].", + "bbox": [ + 212, + 556, + 784, + 693 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Experimental Settings", + "text_level": 1, + "bbox": [ + 214, + 719, + 444, + 734 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Dataset In our experiments, we utilize the AIOZ-GDance [40] and AIST-M [90] datasets. AIOZ-GDance is the largest music-driven dataset focusing on group dance, encompassing paired music and 3D group motions extracted from in-the-wild videos through a semi-automatic process. This dataset spans 7 dance styles and 16 music genres. For consistency, we adhere to the training and testing split outlined in [40] during our experiments.", + "bbox": [ + 212, + 750, + 782, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "N. Le et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation Protocol We employ several metrics to assess the quality of individual dance motions, including Frechet Inception Distance (FID) [20, 49], Motion-Music Consistency (MMC) [49], and Generation Diversity (GenDiv) [26, 41, 49], along with the Physical Foot Contact score (PFC) [83]. Specifically, the FID score gauges the realism of individual dance movements concerning the ground-truth dance. MMC assesses the matching similarity between motion and music beats, reflecting how well-generated dances synchronize with the music's rhythm. GenDiv is computed as the average pairwise distance of kinetic features among motions [58]. PFC evaluates the physical plausibility of foot movements by determining the agreement between the acceleration of the character's center of mass and the foot's velocity.", + "bbox": [ + 212, + 146, + 787, + 311 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In assessing the group dance quality, we adopt three metrics outlined in [40]: Group Motion Realism (GMR), Group Motion Correlation (GMC), and Trajectory Intersection Frequency (TIF). Broadly, GMR gauges the realism of generated group motions in comparison to ground-truth data, employing Frechet Inception Distance on extracted group motion features. GMC evaluates the synchronization among dancers within the generated group by computing their cross-correlation. TIF quantifies the frequency of collisions among the generated dancers during their dance movements.", + "bbox": [ + 212, + 311, + 787, + 434 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Baselines Our method is subjected to comparison with various recent techniques in music-driven dance generation, namely FACT [49], Transflower [59], and EDGE [83]. These approaches are adapted for benchmarking within group dance generation context, as outlined in [40], considering that their original designs were tailored for single-dance scenarios. Additionally, our evaluation includes comparisons with GDanceR [40], GCD [39], and DanY [90]. All mentioned works are specifically designed for the group choreography generation.", + "bbox": [ + 212, + 453, + 787, + 559 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/ca468c86e3412e2a282545609367f4f0fb254fc2a1637d7d2c2f5bc1f7791017.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetMethodFID↓MMC↑GenDiv↑PFC↓GMR↓GMC↑TIF↓
AIOZ-GDANCE [40]FACT [49]56.200.2228.643.52101.5262.680.321
Transflower [59]37.730.2178.743.0781.1760.780.332
EDGE [83]31.400.2649.572.6363.3561.720.356
GDANCER [40]43.900.2509.233.0551.2779.010.217
GCD [39]31.160.26110.872.5331.4780.970.167
Ours31.010.27110.982.3330.0884.520.163
AIST-M [90]GDANCER [40]52.900.2226.521.9365.1360.560.121
GCD [39]35.360.24510.971.5242.5272.150.083
DanY [90]40.250.24011.401.6550.2963.530.137
Ours31.490.25711.811.4241.2478.640.076
", + "bbox": [ + 225, + 566, + 782, + 696 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 1: Performance comparison.", + "bbox": [ + 382, + 700, + 617, + 715 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Experimental Results", + "text_level": 1, + "bbox": [ + 215, + 741, + 441, + 757 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Quality Comparison Table 1 presents a comparison among the baselines FACT [49], Transflower [59], EDGE [83], GDanceR [40], GCD [39], and our proposed manifold-based method. The results clearly demonstrate that our model almost outperforms all baselines across all evaluations on two datasets AIOZ-GDANCE [40] and AIST-M [90]. We observe that recent diffusion-based dance", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Scalable Group Choreography via Variational Phase Manifold Learning", + "bbox": [ + 256, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "generation models such as EDGE or GCD yield competitive performance on both single-dance metrics (FID, MMC, GenDiv, and PFC) and group dance metrics (GMR, GMC, and TIF). However, due to limitations in their training procedures, they still struggle when dealing with generating multiple dancing motions in the context of a high quantity of dancers, as shown by their low performance compared to our proposal. This result implies that our method successfully preserves the performance of dancing motions when the number of dancers is increased. Besides, Figure 3 also shows that our proposed methods outperform other state-of-the-art methods such as GDanceR or GCD in dealing with monotonous, repetitive, sinking, and overlapping dance motions.", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/2d2603f27afe8456ee8f4803ace1a428fb7cb155c097f92c859d462ebae5b4fa.jpg", + "image_caption": [ + "Fig. 3: Visualization of a dancing sample between different methods. GDanceR displays monotonous, repetitive, or sinking dance motions. GCD exhibits more divergence in dance motions, yet dancers may intersect since their optimization does not address this issue explicitly. Blue boxes mark these issues. In contrast, our manifold-based solution ensures the divergence of dancing motions, while the phase motion path demonstrates its effectiveness in addressing floating and crossing issues in group dances." + ], + "image_footnote": [], + "bbox": [ + 225, + 311, + 781, + 494 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/00a18dd431b0bec6abb28d14e8a97bb6053554ec973e8fbaadef02f976c32430.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
#Generated DancersMethodFID↓MMC↑GenDiv↑GMR↓GMC↑TIF↓
5GDanceR [40]44.190.2498.9955.0578.720.218
GCD [39]35.080.2649.9238.4381.440.168
Ours31.350.26810.0532.5884.560.161
10GDanceR [40]57.930.2257.6373.3272.790.386
GCD [39]N/A (Memory)N/A (Memory)N/A (Memory)N/A (Memory)N/A (Memory)N/A (Memory)
Ours32.190.2698.9934.3286.960.193
100GDanceR [40]N/A (Memory)N/A (Memory)N/A (Memory)N/A (Memory)N/A (Memory)N/A (Memory)
GCD [39]N/A (Memory)N/A (Memory)N/A (Memory)N/A (Memory)N/A (Memory)N/A (Memory)
Ours30.970.2688.7638.1385.730.222
", + "bbox": [ + 222, + 609, + 782, + 718 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 2: Performance of group dance generation methods when we increase the number of generated dancers. The experiments are done with common consumer GPUs with 4GB memory. (N/A means models could not run due to inadequate memory footprint)", + "bbox": [ + 212, + 722, + 787, + 766 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Scalable Generation Analysis Table 2 illustrates the performance of different group dance generation methods (GCD, GDancer, and Ours) when generating dance movements with an increasing number of dancers. When the number of", + "bbox": [ + 212, + 794, + 787, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "N. Le et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "dancers is increased to 10, GCD appears to run out of memory, which is also observed in GDanceR when the number of dancers increases to 100. Regardless of the number of dancers, our method consistently achieves stable and competitive results. This implies that our proposed method successfully addresses the scalability issue in group dance generation without compromising the overall performance of each individual dance motion. Figure 4 illustrates the memory consumption to generate dance motions in groups for each method. Noticeably, our proposal still achieves the highest performance while consuming immensely fewer resources required for generating group dance motions (See Figure 5 for illustrations). This, again, indicates that our method successfully breaks the barrier of limited generated dancers by using the manifold.", + "bbox": [ + 212, + 146, + 787, + 313 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Ablation Analysis Table 3 presents the performance improvements achieved through the integration of consistency loss and phase manifold. Additionally, we showcase the effectiveness of our proposed approach across three different backbones: Transformer [49, 85], LSTM [21], and CNN [99]. Evaluation metrics including FID, GMR, and GMC are utilized. The results indicate that the absence of consistency loss leads to an increase in GMR and a decrease in GMC, suggesting a significant enhancement in the realism and correlation of group dance motions facilitated by the inclusion of the proposed objective. Meanwhile, with out the phase manifold, the model exhibits remarkably higher scores in both the FID and GMR metrics, suggesting that phase manifold can effectively improve the distinction in dance motions while maintaining the realism of group dances, even when the number of dancers in a group is large. In comparing three backbones—Transformer, LSTM, and CNN—we have observed that the chosen Transformer backbone achieved the best results compared to LSTM or CNN.", + "bbox": [ + 212, + 335, + 787, + 547 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/9c37744cff0978b08033b4584e15067eebd8ab549bd74b8c8a0336189fd0a50b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodFID↓GMR↓GMC↑
Ours31.0130.0884.52
Ours w/o Consistency Loss35.3557.6366.72
Ours w/o Phase Manifold41.7845.3277.93
Ours w LSTM41.2947.4771.82
Ours w CNN36.9944.9475.77
", + "bbox": [ + 223, + 585, + 488, + 670 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 3: Module contribution and loss analysis.", + "bbox": [ + 215, + 680, + 491, + 709 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/e83930fcbac387bcdfd345f08dec3d2db2014f9f814017447ff288ea36b23994.jpg", + "image_caption": [ + "Fig. 4: Memory usage vs. number of dancers in different dance generators." + ], + "image_footnote": [], + "bbox": [ + 509, + 560, + 787, + 681 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 User Study", + "text_level": 1, + "bbox": [ + 215, + 738, + 356, + 753 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "User studies are vital for evaluating generative models, as user perception is pivotal for downstream applications; thus, we conducted two studies with around 70 participants each, diverse in background, with experience in music and dance, aged between 20 to 40, consisting of approximately $47\\%$ females and $53\\%$ males, to assess the effectiveness of our approach in group choreography generation.", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Scalable Group Choreography via Variational Phase Manifold Learning", + "bbox": [ + 256, + 113, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In the user study, we aim to assess the realism of sample outputs with more and more dancers. Specifically, participants assign scores ranging from 0 to 10 to evaluate the realism of each dance clip with 2 to 10 dancers. Figure 6 shows that, across all methods, the more the number of dancers is increased, the lower the realism is found. However, the drop in realism of our proposed method is the least compared to GCD and GDanceR. The results indicate our method's good performance compared to other baselines when the number of dancers increases.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/7e085c9cc0289882b96eb06347ae0cab883ac551831d5e51d18109e8dc40c698.jpg", + "image_caption": [ + "Fig. 5: Visualization of scalable dancers." + ], + "image_footnote": [], + "bbox": [ + 225, + 267, + 496, + 478 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/b0e89c61c98ff2756242f7040614fc7693068996fb80a37866ced9d1e9e31c29.jpg", + "image_caption": [ + "Fig. 6: Realism between different methods when number of dancers is varied." + ], + "image_footnote": [], + "bbox": [ + 516, + 267, + 785, + 472 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Discussion", + "text_level": 1, + "bbox": [ + 215, + 531, + 354, + 546 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "While our approach leverages the VAE as a primary solution for generating a manifold, it is important to acknowledge certain inherent limitations associated with this choice. One notable challenge is the susceptibility to issues such as posterior collapse and unstable sampling within the VAE framework. These challenges can result in generated group dance motions that may not consistently meet performance expectations.", + "bbox": [ + 212, + 566, + 787, + 656 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "One specific manifestation of this limitation is the potential for false decoding when sampling points that lie too far from the learned distribution. This scenario can lead to unexpected rotations or disruptions in the physics of the generated content. The impact of this problem becomes evident in instances where the generated samples deviate significantly from the anticipated distribution, introducing inaccuracies and distortions.", + "bbox": [ + 212, + 657, + 787, + 747 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "To address these challenges, we recognize the need for ongoing efforts to mitigate the effects of posterior collapse and unstable sampling. While the problem is acknowledged, our approach incorporates measures to limit its impact. Future research directions could explore alternative generative models or additional techniques to enhance the robustness and reliability of the generated results in the face of these identified limitations.", + "bbox": [ + 212, + 750, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "N. Le et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Ahn, H., Kim, J., Kim, K., Oh, S.: Generative autoregressive networks for 3d dancing move synthesis from music. IEEE Robotics and Automation Letters (RA-L) (2020)", + "2. Alaoui, S.F., Henry, C., Jacquemin, C.: Physical modelling for interactive installations and the performing arts. International Journal of Performance Arts and Digital Media (2014)", + "3. Alemi, O., Françoise, J., Pasquier, P.: Groovenet: Real-time music-driven dance movement generation using artificial neural networks. Networks (2017)", + "4. Alexanderson, S., Nagy, R., Beskow, J., Henter, G.E.: Listen, denoise, action! audio-driven motion synthesis with diffusion models. ACM Transactions on Graphics (TOG) (2023)", + "5. Au, H.Y., Chen, J., Jiang, J., Guo, Y.: Choreograph: Music-conditioned automatic dance choreography over a style and tempo consistent dynamic graph. In: ACM International Conference on Multimedia (2022)", + "6. Bisig, D.: Generative dance-a taxonomy and survey. In: International Conference on Movement and Computing (2022)", + "7. Cervantes, P., Sekikawa, Y., Sato, I., Shinoda, K.: Implicit neural representations for variable length human motion generation. In: ECCV (2022)", + "8. Chan, C., Ginosar, S., Zhou, T., Efros, A.A.: Everybody dance now. In: ICCV (2019)", + "9. Chen, K., Tan, Z., Lei, J., Zhang, S.H., Guo, Y.C., Zhang, W., Hu, S.M.: Choreomaster: choreography-oriented music-driven dance synthesis. ACM Transactions on Graphics (TOG) (2021)", + "0. Fan, R., Xu, S., Geng, W.: Example-based automatic music-driven conventional dance motion synthesis. IEEE Transactions on Visualization and Computer Graphics (TVCG) (2011)", + "1. Feng, B., Ao, T., Liu, Z., Ju, W., Liu, L., Zhang, M.: Robust dancer: Long-term 3d dance synthesis using unpaired data. arXiv (2023)", + "2. Ferreira, J.P., Coutinho, T.M., Gomes, T.L., Neto, J.F., Azevedo, R., Martins, R., Nascimento, E.R.: Learning to dance: A graph convolutional adversarial network to generate realistic dance motions from audio. Computers & Graphics (2021)", + "3. Fink, B., Bläsing, B., Ravignani, A., Shackelford, T.K.: Evolution and functions of human dance. Evolution and Human Behavior (2021)", + "4. Gartner, E., Metz, L., Andriluka, M., Freeman, C.D., Sminchisescu, C.: Transformer-based learned optimization. In: CVPR (2023)", + "5. Ghorbani, S., Ferstl, Y., Holden, D., Troje, N.F., Carbonneau, M.A.: Zeroeggs: Zero-shot example-based gesture generation from speech. In: Computer Graphics Forum (2023)", + "6. Ghosh, A., Cheema, N., Oguz, C., Theobalt, C., Slusallek, P.: Synthesis of compositional animations from textual descriptions. In: ICCV (2021)", + "7. Gong, K., Lian, D., Chang, H., Guo, C., Zuo, X., Jiang, Z., Wang, X.: Tm2d: Bimodality driven 3d dance generation via music-text integration. In: ICCV (2023)", + "8. Guo, C., Zou, S., Zuo, X., Wang, S., Ji, W., Li, X., Cheng, L.: Generating diverse and natural 3d human motions from text. In: CVPR (2022)", + "9. He, C., Saito, J., Zachary, J., Rushmeier, H., Zhou, Y.: Nemf: Neural motion fields for kinematic animation. NeurIPS 35, 4244-4256 (2022)", + "20. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: NeurIPS (2017)" + ], + "bbox": [ + 225, + 172, + 785, + 840 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Scalable Group Choreography via Variational Phase Manifold Learning", + "bbox": [ + 256, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "21. Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural Computation (1997)", + "22. Holden, D., Komura, T., Saito, J.: Phase-functional neural networks for character control. ACM Transactions on Graphics (TOG) 36(4), 1-13 (2017)", + "23. Holden, D., Saito, J., Komura, T.: A deep learning framework for character motion synthesis and editing. ACM Transactions on Graphics (TOG) (2016)", + "24. Holden, D., Saito, J., Komura, T., Joyce, T.: Learning motion manifolds with convolutional autoencoders. In: SIGGRAPH Asia 2015 technical briefs (2015)", + "25. Hong, F., Zhang, M., Pan, L., Cai, Z., Yang, L., Liu, Z.: Avatarclip: zero-shot text-driven generation and animation of 3d avatars. ACM Transactions on Graphics (TOG) (2022)", + "26. Huang, R., Hu, H., Wu, W., Sawada, K., Zhang, M.: Dance revolution: Long sequence dance generation with music via curriculum learning. CoRR (2020)", + "27. Huang, Y., Zhang, J., Liu, S., Bao, Q., Zeng, D., Chen, Z., Liu, W.: Genre-conditioned long-term 3d dance generation driven by music. In: ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 4858-4862. IEEE (2022)", + "28. Jiang, C., Cornman, A., Park, C., Sapp, B., Zhou, Y., Anguelov, D., et al.: Motiondiffuser: Controllable multi-agent motion prediction using diffusion. In: CVPR (2023)", + "29. Jiang, Y., Won, J., Ye, Y., Liu, C.K.: Drop: Dynamics responses from human motion prior and projective dynamics. In: SIGGRAPH Asia 2023 Conference papers (2023)", + "30. Joshi, M., Chakrabarty, S.: An extensive review of computational dance automation techniques and applications. Proceedings of the Royal Society A (2021)", + "31. Karunratanakul, K., Preechakul, K., Suwajanakorn, S., Tang, S.: Guided motion diffusion for controllable human motion synthesis. In: ICCV (2023)", + "32. Kico, I., Grammalidis, N., Christidis, Y., Liarokapis, F.: Digitization and visualization of folk dances in cultural heritage: A review. Inventions (2018)", + "33. Kim, J.W., Fouad, H., Hahn, J.K.: Making them dance. In: AAAI Fall Symposium: Aurally Informed Performance (2006)", + "34. Kim, J., Oh, H., Kim, S., Tong, H., Lee, S.: A brand new dance partner: Music-conditioned pluralistic dancing controlled by multiple dance genres. In: CVPR (2022)", + "35. Kim, T.h., Park, S.I., Shin, S.Y.: Rhythmic-motion synthesis based on motion-beat analysis. ACM Transactions on Graphics (TOG) (2003)", + "36. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. In: ICLR (2014)", + "37. Kolotouros, N., Alldieck, T., Zanfir, A., Bazavan, E., Fieraru, M., Sminchisescu, C.: Dreamhuman: Animatable 3d avatars from text. In: NeurIPS (2024)", + "38. Lannan, N., Zhou, L., Fan, G.: Human motion enhancement via tobit kalman filter-assisted autoencoder. IEEE Access 10, 29233-29251 (2022)", + "39. Le, N., Do, T., Do, K., Nguyen, H., Tjiputra, E., Tran, Q.D., Nguyen, A.: Controllable group choreography using contrastive diffusion. ACM Transactions on Graphics (TOG) (2023)", + "40. Le, N., Pham, T., Do, T., Tjiputra, E., Tran, Q.D., Nguyen, A.: Music-driven group choreography. In: CVPR (2023)", + "41. Lee, H.Y., Yang, X., Liu, M.Y., Wang, T.C., Lu, Y.D., Yang, M.H., Kautz, J.: Dancing to music. In: NeurIPS (2019)", + "42. Lee, L., Lin, Z., Hu, R., Gong, Z., Kumar, A., Li, T., Li, S., Hui, P.: When creators meet the metaverse: A survey on computational arts. CoRR (2021)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "N. Le et al.", + "bbox": [ + 271, + 114, + 349, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "43. Lee, M., Lee, K., Park, J.: Music similarity-based approach to generating dance motion sequence. Multimedia tools and applications (2013)", + "44. Li, B., Zhao, Y., Sheng, L.: Dancenet3d: Music based dance generation with parametric motion transformer. In: AAAI (2022)", + "45. Li, B., Zhao, Y., Zhelun, S., Sheng, L.: Danceformer: Music conditioned 3d dance generation with parametric motion transformer. In: Proceedings of the AAAI Conference on Artificial Intelligence (2022)", + "46. Li, J., Villegas, R., Ceylan, D., Yang, J., Kuang, Z., Li, H., Zhao, Y.: Task-generic hierarchical human motion prior using vaes. In: International Conference on 3D Vision (3DV) (2021)", + "47. Li, J., Kang, D., Pei, W., Zhe, X., Zhang, Y., He, Z., Bao, L.: Audio2gestures: Generating diverse gestures from speech audio with conditional variational autoencoders. In: ICCV (2021)", + "48. Li, R., Zhao, J., Zhang, Y., Su, M., Ren, Z., Zhang, H., Tang, Y., Li, X.: Finedance: A fine-grained choreography dataset for 3d full body dance generation. In: ICCV (2023)", + "49. Li, R., Yang, S., Ross, D.A., Kanazawa, A.: Ai choreographer: Music conditioned 3d dance generation with aist++. In: ICCV (2021)", + "50. Li, W., Dai, B., Zhou, Z., Yao, Q., Wang, B.: Controlling character motions without observable driving source. In: WACV (2024)", + "51. Liang, H., Zhang, W., Li, W., Yu, J., Xu, L.: Intergen: Diffusion-based multi-human motion generation under complex interactions. arXiv preprint arXiv:2304.05684 (2023)", + "52. Lin, J., Zeng, A., Lu, S., Cai, Y., Zhang, R., Wang, H., Zhang, L.: Motion-x: A large-scale 3d expressive whole-body human motion dataset. NeurIPS (2024)", + "53. Ling, H.Y., Zinno, F., Cheng, G., Van De Panne, M.: Character controllers using motion vaes. ACM Transactions on Graphics (TOG) 39(4), 40-1 (2020)", + "54. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: SMPL: A skinned multi-person linear model. ACM Transactions on Graphics (TOG) (2015)", + "55. Martinez, J., Black, M.J., Romero, J.: On human motion prediction using recurrent neural networks. In: CVPR (2017)", + "56. Mo, C.A., Hu, K., Long, C., Wang, Z.: Continuous intermediate token learning with implicit motion manifold for keyframe based motion interpolation. In: CVPR (2023)", + "57. Ofli, F., Erzin, E., Yemez, Y., Tekalp, A.M.: Learn2dance: Learning statistical music-to-dance mappings for choreography synthesis. IEEE Transactions on Multimedia (TMM) (2011)", + "58. Onuma, K., Faloutsos, C., Hodgins, J.K.: Fmdistance: A fast and effective distance function for motion capture data. In: Eurographics (2008)", + "59. Perez, G.V., Beskow, J., Henter, G., Holzapfel, A., Oudeyer, P.Y., Alexanderson, S.: Transflower: probabilistic autoregressive dance generation with multimodal attention. ACM Transactions on Graphics (TOG) (2021)", + "60. Petrovich, M., Black, M.J., Varol, G.: Action-conditioned 3d human motion synthesis with transformer vae. In: ICCV (2021)", + "61. Petrovich, M., Black, M.J., Varol, G.: Temos: Generating diverse human motions from textual descriptions. In: ECCV (2022)", + "62. Qi, Q., Zhuo, L., Zhang, A., Liao, Y., Fang, F., Liu, S., Yan, S.: Diffdance: Cascaded human motion diffusion model for dance generation. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 1374-1382 (2023)", + "63. Raab, S., Leibovitch, I., Li, P., Aberman, K., Sorkine-Hornung, O., Cohen-Or, D.: Modi: Unconditional motion synthesis from diverse data. In: CVPR (2023)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Scalable Group Choreography via Variational Phase Manifold Learning", + "bbox": [ + 256, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "64. Ren, X., Li, H., Huang, Z., Chen, Q.: Self-supervised dance video synthesis conditioned on music. In: ACMMM (2020)", + "65. Safonova, A., Hodgins, J.K.: Construction and optimal search of interpolated motion graphs. In: SIGGRAPH (2007)", + "66. Shafir, Y., Tevet, G., Kapon, R., Bermano, A.H.: Human motion diffusion as a generative prior. arXiv (2023)", + "67. Shi, J.: Application of 3d computer aided system in dance creation and learning. In: International Conference on Machine Learning and Big Data Analytics for IoT Security and Privacy (2021)", + "68. Shi, M., Starke, S., Ye, Y., Komura, T., Won, J.: Phasemp: Robust 3d pose estimation via phase-conditioned human motion prior. In: ICCV (2023)", + "69. Shiratori, T., Nakazawa, A., Ikeuchi, K.: Dancing-to-music character animation. Computer Graphics Forum (2006)", + "70. Sitzmann, V., Martel, J., Bergman, A., Lindell, D., Wetzstein, G.: Implicit neural representations with periodic activation functions. In: NeurIPS (2020)", + "71. Siyao, L., Yu, W., Gu, T., Lin, C., Wang, Q., Qian, C., Loy, C.C., Liu, Z.: Bailando: 3d dance generation by actor-critic gpt with choreographic memory. In: CVPR (2022)", + "72. Soga, A., Umino, B., Longstaff, J.S.: Automatic composition of ballet sequences using a 3d motion archive. In: 1st South-Eastern European Digitization Initiative Conference (2005)", + "73. Sohn, K., Lee, H., Yan, X.: Learning structured output representation using deep conditional generative models. In: NeurIPS (2015)", + "74. Starke, S., Mason, I., Komura, T.: Deepphase: Periodic autoencoders for learning motion phase manifolds. ACM Transactions on Graphics (TOG) (2022)", + "75. Starke, S., Zhang, H., Komura, T., Saito, J.: Neural state machine for character-scene interactions. ACM Transactions on Graphics (TOG) 38(6), 209-1 (2019)", + "76. Starke, S., Zhao, Y., Komura, T., Zaman, K.: Local motion phases for learning multi-contact character movements. ACM Transactions on Graphics (TOG) 39(4), 54-1 (2020)", + "77. Sun, G., Wong, Y., Cheng, Z., Kankanhalli, M.S., Geng, W., Li, X.: Deepdance: Music-to-dance motion choreography with adversarial learning. IEEE Transactions on Multimedia (TMM) (2020)", + "78. Sun, J., Wang, C., Hu, H., Lai, H., Jin, Z., Hu, J.F.: You never stop dancing: Nonfreezing dance generation via bank-constrained manifold projection. In: NeurIPS (2022)", + "79. Tang, T., Jia, J., Mao, H.: Dance with melody: An LSTM-autoencoder approach to music-oriented dance synthesis. In: ACMMM (2018)", + "80. Tevet, G., Gordon, B., Hertz, A., Bermano, A.H., Cohen-Or, D.: Motionclip: Exposing human motion generation to clip space. In: European Conference on Computer Vision (2022)", + "81. Tevet, G., Raab, S., Gordon, B., Shafir, Y., Cohen-Or, D., Bermano, A.H.: Human motion diffusion model. In: ICLR (2023)", + "82. Tiwari, G., Antic, D., Lenssen, J.E., Sarafianos, N., Tung, T., Pons-Moll, G.: Posendf: Modeling human pose manifolds with neural distance fields. In: ECCV (2022)", + "83. Tseng, J., Castellon, R., Liu, C.K.: Edge: Editable dance generation from music. In: CVPR (2023)", + "84. Tsuchida, S., Fukayama, S., Hamasaki, M., Goto, M.: Aist dance video database: Multi-genre, multi-dancer, and multi-camera database for dance information processing. In: ISMIR (2019)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "N. Le et al.", + "bbox": [ + 271, + 114, + 349, + 126 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "85. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017)", + "86. Wang, Z., Jia, J., Wu, H., Xing, J., Cai, J., Meng, F., Chen, G., Wang, Y.: Group-dancer: Music to multi-people dance synthesis with style collaboration. In: ACM International Conference on Multimedia (2022)", + "87. Yalta, N., Watanabe, S., Nakadai, K., Ogata, T.: Weakly-supervised deep recurrent neural networks for basic dance step generation. In: IJCNN (2019)", + "88. Yang, S., Wu, Z., Li, M., Zhang, Z., Hao, L., Bao, W., Zhuang, H.: Qpgesture: Quantization-based and phase-guided motion matching for natural speech-driven gesture generation. In: CVPR (2023)", + "89. Yang, Z., Wen, Y.H., Chen, S.Y., Liu, X., Gao, Y., Liu, Y.J., Gao, L., Fu, H.: Keyframe control of music-driven 3d dance generation. IEEE Transactions on Visualization and Computer Graphics (2023)", + "90. Yao, S., Sun, M., Li, B., Yang, F., Wang, J., Zhang, R.: Dance with you: The diversity controllable dancer generation via diffusion models. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 8504-8514 (2023)", + "91. Ye, Z., Wu, H., Jia, J., Bu, Y., Chen, W., Meng, F., Wang, Y.: Choreonet: Towards music to dance synthesis with choreographic action unit. In: ACMMM (2020)", + "92. Yin, W., Yin, H., Baraka, K., Kragic, D., Björkman, M.: Dance style transfer with cross-modal transformer. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 5058-5067 (2023)", + "93. Yu, Z., Cheng, W., Liu, X., Wu, W., Lin, K.Y.: Monohuman: Animatable human neural field from monocular video. In: CVPR (2023)", + "94. Zhang, X., Yang, S., Xu, Y., Zhang, W., Gao, L.: Mining and applying composition knowledge of dance moves for style-concentrated dance generation. In: AAAI (2023)", + "95. Zhou, L., Luo, Y.: A spatio-temporal learning for music conditioned dance generation. In: International Conference on Multimodal Interaction (2022)", + "96. Zhou, Y., Barnes, C., Lu, J., Yang, J., Li, H.: On the continuity of rotation representations in neural networks. In: CVPR (2019)", + "97. Zhou, Z., Wang, B.: Ude: A unified driving engine for human motion generation. In: CVPR (2023)", + "98. Zhu, Y., Olszewski, K., Wu, Y., Achlioptas, P., Chai, M., Yan, Y., Tulyakov, S.: Quantized gan for complex music generation from dance videos. In: ECCV (2022)", + "99. Zhuang, W., Wang, C., Chai, J., Wang, Y., Shao, M., Xia, S.: Music2dance: Dancenet for music-driven dance generation. ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM) (2022)" + ], + "bbox": [ + 215, + 147, + 784, + 646 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "Scalable Group Choreography via Variational Phase Manifold Learning", + "bbox": [ + 256, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/Scalable Group Choreography via Variational Phase Manifold Learning/6dc9e2ab-125f-402e-97a2-dfb708ce273d_model.json b/2024/Scalable Group Choreography via Variational Phase Manifold Learning/6dc9e2ab-125f-402e-97a2-dfb708ce273d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2c0917623ce9f5fb0c184d2d7ffb3321890b2288 --- /dev/null +++ b/2024/Scalable Group Choreography via Variational Phase Manifold Learning/6dc9e2ab-125f-402e-97a2-dfb708ce273d_model.json @@ -0,0 +1,2746 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.242, + 0.141, + 0.761, + 0.187 + ], + "angle": 0, + "content": "Scalable Group Choreography via Variational Phase Manifold Learning" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.212, + 0.776, + 0.245 + ], + "angle": 0, + "content": "Nhat \\(\\mathrm{Le}^{1}\\), Khoa \\(\\mathrm{Do}^{2,3}\\), Xuan Bui\\(^{2,3}\\), Tuong \\(\\mathrm{Do}^{1,4}\\), Erman Tjiputra\\(^{1}\\), Quang D. Tran\\(^{1}\\), and Anh Nguyen\\(^{4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.437, + 0.255, + 0.566, + 0.269 + ], + "angle": 0, + "content": "1 AIOZ, Singapore" + }, + { + "type": "text", + "bbox": [ + 0.327, + 0.269, + 0.676, + 0.284 + ], + "angle": 0, + "content": "2 University of Science, Ho Chi Minh City, Vietnam" + }, + { + "type": "text", + "bbox": [ + 0.302, + 0.283, + 0.701, + 0.297 + ], + "angle": 0, + "content": "3 Vietnam National University, Ho Chi Minh City, Vietnam" + }, + { + "type": "text", + "bbox": [ + 0.291, + 0.297, + 0.712, + 0.311 + ], + "angle": 0, + "content": "4 University of Liverpool, Liverpool L69 3BX, United Kingdom" + }, + { + "type": "list", + "bbox": [ + 0.291, + 0.269, + 0.712, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.349, + 0.74, + 0.556 + ], + "angle": 0, + "content": "Abstract. Generating group dance motion from the music is a challenging task with several industrial applications. Although several methods have been proposed to tackle this problem, most of them prioritize optimizing the fidelity in dancing movement, constrained by predetermined dancer counts in datasets. This limitation impedes adaptability to real-world applications. Our study addresses the scalability problem in group choreography while preserving naturalness and synchronization. In particular, we propose a phase-based variational generative model for group dance generation on learning a generative manifold. Our method achieves high-fidelity group dance motion and enables the generation with an unlimited number of dancers while consuming only a minimal and constant amount of memory. The intensive experiments on two public datasets show that our proposed method outperforms recent state-of-the-art approaches by a large margin and is scalable to a great number of dancers beyond the training data." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.582, + 0.375, + 0.597 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.779 + ], + "angle": 0, + "content": "The widespread availability of digital social media platforms has led to a significant increase in the popularity of creating dance videos. This heightened interest has resulted in the daily production and viewing of millions of dance videos on various online platforms [13, 32]. Recently, researchers from the computer vision and computer graphics community have focused on developing methods for generating authentic dance movements in response to music [6, 94]. These advancements have broad implications and find applications in diverse areas, including animation [49, 50], virtual idols [59], virtual metaverses [42], and dance education [2, 67, 72]. Such techniques empower artists, animators, and educators by providing powerful tools to enhance their creative pursuits and improve the overall dance experience for performers and audiences alike." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Recently, significant strides have been made in creating dance motions for solo dance performers [12,26,34,49,59,71,83]. However, generating synchronized group dance movements that are both lifelike and in harmony with music remains a complex challenge [9,27,87]. The introduction of AIOZ-GDance [40] stands as" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.35, + 0.128 + ], + "angle": 0, + "content": "N. Le et al." + }, + { + "type": "image", + "bbox": [ + 0.232, + 0.145, + 0.41, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.413, + 0.145, + 0.59, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.145, + 0.772, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.239, + 0.788, + 0.281 + ], + "angle": 0, + "content": "Fig. 1: We present a new group dance generation method that can generate a large number of dancers within a fixed resource consumption. The illustration shows a generated group dance sample with 100 dancers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.306, + 0.789, + 0.488 + ], + "angle": 0, + "content": "the extensive dataset to aid in generating group choreography. Different aspects of group dance motion such as the study of consistency and diversity among the movements of dancers are explored in [39]. However, despite recent advances in the field, existing methods can either only generate dances for a limited number of dancers given the input audio [40, 86, 90], or would consume an expensively large amount of memory due to their design nature [39, 40]. Typically, these methods are usually based on specialized collaborative mechanisms such as cross-entity attention [40, 51], group global attention [39], or residual communication block [66] achieve coherent multi-person synthesis. In other words, their innate architectural design requires the network to process and synthesize all motions simultaneously, leading to excessively growing computational/memory costs with the increasing number of inputs, and hence extremely difficult to scale up." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.49, + 0.789, + 0.672 + ], + "angle": 0, + "content": "Despite the comprehensive exploration of group dance motion generation, the predominant emphasis remains on optimizing the execution of the generated movements [14, 19, 31, 37, 52, 93]. These existing methodologies function within a confined scope, often tethered to a predetermined maximum number of dancers depicted in group choreography videos outlined by the predefined dataset [39]. This inherent limitation poses a potential constraint on the adaptability of trained models when transitioning to real-world applications [47]. In our pursuit, we center our efforts on addressing the critical aspect of scalability in group dance. Our primary objective revolves around not only expanding the number of dancers but also ensuring the preservation of the innate naturalness in the dance motions and the seamless synchronization between dancers throughout the entirety of their performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.842 + ], + "angle": 0, + "content": "In the field of character motion control and synthesis, existing approaches can be broadly divided into two categories: deterministic [16, 26, 40, 49, 55, 86] and probabilistic [19, 39, 59, 61, 83]. The former tries to learn direct and deterministic mapping from the input conditioning signal (such as text, audio, or user controls) to the desired output motion. However, these methods usually regress towards the mean pose as they are typically trained by minimizing a regression objective under different outputs for similar inputs, resulting in freezing or drifting motion. On the other hand, probabilistic methods attempt to capture the distribution over possible motions for a given condition, allowing diverse and high-quality generation. In particular, Variational Autoencoder (VAE) [18, 47, 60, 61], or Diffusion Models [4, 39, 81, 83, 97] have been widely" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.257, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Scalable Group Choreography via Variational Phase Manifold Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.268 + ], + "angle": 0, + "content": "used. Although diffusion models have recently shown strong potential in generating high-quality motions [81], they generally operate on high-dimensional space (same as the original data) [39,81,83], hence rendering them extremely challenging to use for scalable group generation. Meanwhile, standard VAE models often rely on a Gaussian latent space of fixed dimension for sampling a single latent vector to generate motions, making it difficult to adapt to a wide range of architectures. Besides, the fixed-size single latent space might also not be sufficient to represent the temporal dynamics of the whole motion sequence [7,60]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.273, + 0.785, + 0.499 + ], + "angle": 0, + "content": "Recent studies have demonstrated the significant advantages of learning motion features in the frequency domain. Specifically, [22, 74-76] have found that different motion skills can be effectively represented by several phase variables that can faithfully capture the spatial-temporal alignment of a wide range of movements. This can be used to combine with common deep motion synthesis networks to enhance its capability, facilitating natural, and stable motion generations by enforcing a unidirectional motion transition (i.e., no backward direction in time) to prevent the motion from being stuck temporally [74]. Unlike those prior works, where the phases are utilized as auxiliary signals to assist in the motion synthesis process [68, 74, 88], we propose to learn to generate the phases conditioned on the input audio signal in an end-to-end manner, eliminating the need of training extra components. Our generative phase manifold can be used to generate new motions efficiently, from which the latent motion curves are sampled in frequency domain and decoded into the original motion space by a learned decoder." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.505, + 0.787, + 0.716 + ], + "angle": 0, + "content": "In this paper, our goal is to develop a scalable technique for group dance generation, a phase-based variational generative model for scalable group dance generation, namely Phase-conditioned Dance VAE (PDVAE). To our knowledge, PDVAE is the first method to represent the variational latent space using phase parameters in the frequency domain of the motion curves. Our method goes beyond the conventional VAE approach that typically relies on a single latent vector drawn from a Gaussian distribution, which is unable to adequately represent the temporal information of the motion sequence (e.g., the time dimension is squeezed out). Figure 1 illustrates an example of a dance motion sequence with 100 dancers generated by our model. In contrast to existing methods, our model can generate crowd-dance animations for an unlimited number of dancers without increasing computational burden (i.e., memory consumption remains constant), while still maintaining the performance and fidelity of the generated choreographies. To summarize, our key contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.73, + 0.784, + 0.774 + ], + "angle": 0, + "content": "- We introduce PDVAE, a phase-based variational generative model for scalable group dance generation. The method focuses on generating large-scale group dance under limited resources." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.78, + 0.785, + 0.84 + ], + "angle": 0, + "content": "- To effectively learn the manifold that is group-consistent (i.e., dancers within a group lie upon the same manifold), we propose a group consistency loss that enforces the networks to encode the latent phase manifold to be identical for the same group given the input music." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.73, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.35, + 0.128 + ], + "angle": 0, + "content": "N. Le et al." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.147, + 0.787, + 0.193 + ], + "angle": 0, + "content": "- Extensive experiments along with thorough user study evaluations demonstrate the state-of-the-art performance of our model while achieving effective scalability." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.211, + 0.388, + 0.227 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.238, + 0.492, + 0.254 + ], + "angle": 0, + "content": "2.1 Music-driven Choreography" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.258, + 0.787, + 0.379 + ], + "angle": 0, + "content": "Crafting natural human choreography derived from music presents a multifaceted challenge [30,62,89]. Certain methodologies integrate constraints based on music-motion similarity matching to ensure coherence between the generated motion and the music [10,33,35,43,65,69]. However, many of these approaches rely considerably on heuristic algorithms, stitching together pre-existing dance segments sourced from a limited music-dance database [10, 48]. While successful in generating extended and realistic dance sequences, these methods face limitations when endeavoring to create entirely novel dance fragments [57, 90]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.38, + 0.787, + 0.546 + ], + "angle": 0, + "content": "In the recent period, advancements have been evident in the domain of converting music into dance movements through various techniques such as Convolutional Networks (CNN) [1,8,77,91,92,99], Recurrent Networks (RNN) [3,26,77,79,87], Graph Neural Networks (GNN) [5,12,64,95], Generative Adversarial Networks (GAN) [41,77], or Transformer models [34,44,45,49,59,71]. These methodologies typically depend on various inputs, including the present music and a concise history of previous dance movements, to forecast forthcoming sequences of human poses. This innovation introduces a music-text feature fusion module, engineered to amalgamate inputs into a motion decoder, thereby enabling the creation of dance sequences conditioned upon both musical and textual instructions [17]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.546, + 0.787, + 0.773 + ], + "angle": 0, + "content": "While these techniques show promise in generating authentic and lifelike dance motions, they often struggle to synchronize movements seamlessly across multiple dancers [40]. Specifically, achieving coordination and harmony among dancers necessitates consideration of spatial and temporal relationships, their interactions, and the overall choreographic structure [84]. Consequently, further advancements in this domain are being pursued to tackle these complexities [25, 47, 59, 98]. For instance, Perez et al. [59] integrate a multimodal transformer encoder with a normalizing-flow-based decoder to estimate a probability distribution encompassing potential subsequent poses. Additionally, Feng et al. [11] enable long-term generation by imposing a motion repeat constraint to forecast future frames while considering historical motions. A recent study by Le et al. [39] investigates consistency and diversity factors between the generated motions of two or more dancers within a timeframe. However, these approaches are limited to a predefined number of dancers, heavily constrained by the maximum number of dancers in the dataset-provided videos." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.789, + 0.482, + 0.806 + ], + "angle": 0, + "content": "2.2 Motion Manifold Learning" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Motion manifold learning has attracted considerable attention in computer vision and artificial intelligence, with the primary goal of comprehending the funda-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.257, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Scalable Group Choreography via Variational Phase Manifold Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.253 + ], + "angle": 0, + "content": "mental structures inherent in human movement and dynamics [16, 24, 56, 82]. Its distinctive ability to generate human movement patterns presents numerous opportunities to comprehend intrinsic motion dynamics, manage nonlinear relationships in motion data, and acquire contextual and hierarchical representations [29, 63, 88]. Consequently, various methodologies have emerged, each contributing distinct perspectives and techniques to advance the comprehension and synthesis of human motion [28, 38, 63, 68]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.788, + 0.466 + ], + "angle": 0, + "content": "For instance, Holden et al. [23] generate character movements by mapping high-level parameters to the human motion manifold, allowing diverse movements without manual preprocessing, and enabling post-generation editing for natural, smooth motion sequences. MotionCLIP [80] introduces a 3D human motion auto-encoder aligned with the Contrastive Language-Image Pre-training (CLIP) model's space, enabling semantic text-based motion generation, disentangled editing, and abstract language specification, leveraging CLIP's rich semantic knowledge within the motion manifold. Recently, Sun et al. [78] utilize a VQ-VAE to acquire a low-dimensional manifold, effectively cleansing the motion sequences. Regarding the scalability of group dance generation, the concept of the motion manifold emerges as a potential solution for addressing the restricted number of dancers in the dataset. This direction enables the learning of a distribution that facilitates the extraction of dance motions, allowing for the presentation of cohesive group dance sequences." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.488, + 0.734, + 0.525 + ], + "angle": 0, + "content": "3 Variational Phase Manifold Learning for Scalable Group Choreography" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.538, + 0.39, + 0.552 + ], + "angle": 0, + "content": "3.1 Task Definition" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.565, + 0.788, + 0.747 + ], + "angle": 0, + "content": "Given an input music sequence \\(\\mathbf{a} = \\{a_1, a_t, \\dots, a_T\\}\\) with \\(t = \\{1, \\dots, T\\}\\) indicates the index of the music frames, our goal is to generate the group motion sequences of \\(N\\) arbitrary dancers: \\(\\mathbf{x} = \\{x_1^1, \\dots, x_T^1; \\dots; x_1^N, \\dots, x_T^N\\}\\) where \\(x_t^n\\) is the pose of \\(n\\)-th dancer at frame \\(t\\). We follow [19, 53] and represent dance as a sequence of poses in the 24-joint of the SMPL body model [54], using the 6D continuous rotation [96] for every joint, along with 3D joint positions and velocities. Additionally, the corresponding 3D root translation vectors are concatenated into the pose representations to involve the trajectory of motion. Previous group dance methods [40], which generate the whole group at once, cannot deal with the increasing number of dancers and can only create group sequences up to a pre-defined number of dancers, due to the vast complexity of the architecture. In contrast, we aim to generate group dance with an unlimited number of dancers." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.769, + 0.517, + 0.783 + ], + "angle": 0, + "content": "3.2 Phase-conditioned Dance VAE" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Our goal is to learn a continuous manifold such that the motion can be generated by sampling from this learned manifold. We assume that although different dancers within the same group may present visually distinctive movements, the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.35, + 0.127 + ], + "angle": 0, + "content": "N. Le et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.146, + 0.773, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.383, + 0.788, + 0.51 + ], + "angle": 0, + "content": "Fig. 2: Overview of our Phase-conditioned Dance VAE (PDVAE) for scalable group dance generation. It consists of an Encoder, a Prior, and a Decoder network. During training, we encode the corresponding motion and music inputs into a latent phase manifold, which is variational and parameterized by the frequency domain parameters of periodic functions. The latent phases can be sampled from the manifold and then decoded back to the original data space to obtain new motions. The consistency loss \\(\\mathcal{L}_{\\mathrm{csc}}\\) is further imposed to constrain the manifold to be consistently unified for dancers that belong to the same group. At inference stage, only the Prior and the Decoder are used to synthesize group dances efficiently." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.534, + 0.788, + 0.716 + ], + "angle": 0, + "content": "properties of their motions, such as timing, periodicity, or temporal alignment are intrinsically similar. Drawing inspiration from [74], we aim to learn a generative phase representation for each group of dancer in order to synthesize their motion indefinitely. Our generative model is built upon the conditional Variational Autoencoder architecture [73], thanks to its diverse generation capability and fast sampling speed. However, instead of directly encoding the data into a Gaussian latent distribution as in common VAE approaches [15, 41, 47, 60, 61], we model the latent variational distribution by the phase parameters extracted from the latent motion curve, which we call variational phase manifold. The latent phase manifold is well-structured and can well describe key characteristics of motion (such as its timing, local periodicity, and transition), which benefits learning motion features [74]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.842 + ], + "angle": 0, + "content": "The overview of our Phase-conditioned Dance VAE is illustrated in Figure 2. Specifically, the model contains three main networks: an encoder \\(\\mathcal{E}\\) to capture the approximate posterior distribution conditioned on both motion and music \\(q_{\\phi}(\\mathbf{z}|\\mathbf{x},\\mathbf{a})\\), a prior network \\(\\mathcal{P}\\) to learn the conditional prior given only the music \\(p_{\\theta}(\\mathbf{z}|\\mathbf{a})\\), and a decoder \\(\\mathcal{D}\\) to learn to reconstruct the data from the latent distribution \\(p_{\\theta}(\\mathbf{x}|\\mathbf{z},\\mathbf{a})\\). The new motion is generated by sampling the frequency-domain parameters predicted by the prior network, which is then passed through the decoder network to reconstruct the motion in the original data space. Fur-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.257, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Scalable Group Choreography via Variational Phase Manifold Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "thermore, we adopt Transformer-based architecture [85] in each network to effectively capture long-range dependencies and holistic context of the dance sequence." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.214, + 0.785, + 0.412 + ], + "angle": 0, + "content": "Encoder. The encoder \\(\\mathcal{E}\\) is expected to take both the motion and music feature sequence as input, and produce a distribution over possible latent variables capturing the cross-modal relationship between them. To transform the joint input space into a learned phase manifold, we adopt the Transformer decoder architecture where the Cross-Attention mechanism [85] is utilized to learn the relationship between the motion and the music. Accordingly, the output of the encoder is a batch of latent curves (i.e., the activation sequences per channel) that can particularly capture different spatial and temporal aspects of the motion sequence. However, instead of training the model to directly reconstruct the input motion from the extracted latent curves, we further enforce each channel of the latent space to have a periodic functional form (i.e., sinusoidal). This enables us to effectively learn a compact parameterization for each latent channel from a small set of parameters in the frequency domain." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.432, + 0.784, + 0.568 + ], + "angle": 0, + "content": "Generative Variational Phase Manifold. Here we focus on learning the periodicity and non-linear temporal alignment of the motion in the latent space. In particular, given the output latent curves from the encoder \\(\\mathbf{L} = \\mathcal{E}(x,a)\\in \\mathbb{R}^{D\\times T}\\) with \\(D\\) is the number of desired phase channels to be extracted from the motion, we parameterize each latent curve in \\(\\mathbf{L}\\) using a sinusoidal function with amplitude (A), frequency (F), offset (B) and phase shift (S) parameters [74]. To allow for variational phase manifold learning, we opt to predict two sets of parameters \\(\\mu_{\\varepsilon} = \\{\\mu^{A};\\mu^{F};\\mu^{B};\\mu^{S}\\}\\) and \\(\\sigma_{\\varepsilon} = \\{\\sigma^{A};\\sigma^{F};\\sigma^{B};\\sigma^{S}\\}\\), which corresponds to the mean and variance of \\(\\mathbb{R}^{4D}\\) dimensional Gaussian distribution:" + }, + { + "type": "equation", + "bbox": [ + 0.413, + 0.581, + 0.785, + 0.597 + ], + "angle": 0, + "content": "\\[\nq _ {\\phi} (\\mathbf {z} | \\mathbf {x}, \\mathbf {a}) = \\mathcal {N} (\\mathbf {z}; \\mu_ {\\varepsilon}, \\sigma_ {\\varepsilon}) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.607, + 0.784, + 0.711 + ], + "angle": 0, + "content": "To do so, we first apply differentiable Fast Fourier Transform (FFT) to each channel of the latent curve \\(\\mathbf{L}\\) and create the zero-indexed matrix of Fourier coefficients as \\(\\mathbf{c} = FFT(\\mathbf{L})\\) with \\(\\mathbf{c} \\in \\mathbb{C}^{D \\times K + 1}\\), \\(K = \\left\\lfloor \\frac{T}{2} \\right\\rfloor\\). Correspondingly, we compute the per channel power spectrum \\(\\mathbf{p} \\in \\mathbb{R}^{D \\times K + 1}\\) as \\(\\mathbf{p}_{i,j} = \\frac{2}{N} |\\mathbf{c}_{i,j}|^2\\), where \\(i\\) is the channel index and \\(j\\) is the index for the frequency bands. Correspondingly, the distributional mean parameters of the periodic sinusoidal function are then calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.306, + 0.722, + 0.785, + 0.768 + ], + "angle": 0, + "content": "\\[\n\\mu_ {i} ^ {A} = \\sqrt {\\frac {2}{T} \\sum_ {j = 1} ^ {K} \\mathbf {p} _ {i , j}}, \\quad \\mu_ {i} ^ {F} = \\frac {\\sum_ {j = 1} ^ {K} \\mathbf {f} _ {j} \\cdot \\mathbf {p} _ {i , j}}{\\sum_ {j = 1} ^ {K} \\mathbf {p} _ {i , j}}, \\quad \\mu_ {i} ^ {B} = \\frac {\\mathbf {c} _ {i , 0}}{T}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.779, + 0.784, + 0.822 + ], + "angle": 0, + "content": "where \\(\\mathbf{f} = (0, \\frac{1}{T}, \\dots, \\frac{K}{T})\\) is the frequencies vector. At the same time, the phase shift \\(\\mathbf{S}\\) is predicted using a fully-connected (FC) layer with two arctan activation as:" + }, + { + "type": "equation", + "bbox": [ + 0.357, + 0.825, + 0.785, + 0.842 + ], + "angle": 0, + "content": "\\[\n\\left(s _ {y}, s _ {x}\\right) = \\operatorname {F C} \\left(\\mathbf {L} _ {i}\\right), \\quad \\mu_ {i} ^ {S} = \\arctan \\left(s _ {y}, s _ {x}\\right), \\tag {3}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.35, + 0.128 + ], + "angle": 0, + "content": "N. Le et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.314 + ], + "angle": 0, + "content": "To predict the distributional variance of the phase amplitude and phase shift parameters \\(\\{\\sigma^A,\\sigma^S\\}\\), We additionally apply a separate two-layer MLP network over each channel of the latent curves, similar to Equation 3. The variational latent phase parameters are sampled by utilizing parameterization trick [36], i.e., \\(\\mathbf{A} \\sim \\mathcal{N}(\\mu^A,\\sigma^A)\\) and \\(\\mathbf{S} \\sim \\mathcal{N}(\\mu^S,\\sigma^S)\\). In our experiments, we find that sampling the phase frequency \\(\\mathbf{F}\\) and offset \\(\\mathbf{B}\\) often produce unstable and noncoherent group movements. This might be because the frequency amplitudes of the dancers within the same group are likely to associate with the rhythmic pattern of the musical beats while the offsets capture their alignment, thereby should be consistent with each other. Therefore, we treat those parameters as deterministic by constraining their variance to zero." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.318, + 0.788, + 0.365 + ], + "angle": 0, + "content": "Finally, the sampled set of phase parameters \\(\\mathbf{z} = \\{\\mathbf{A};\\mathbf{F};\\mathbf{B};\\mathbf{S}\\}\\) are used to reconstruct a parametric latent space consisting of multiple periodic curves to represent each intrinsic property of the motion by:" + }, + { + "type": "equation", + "bbox": [ + 0.385, + 0.388, + 0.785, + 0.404 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {L}} = \\mathbf {A} \\cdot \\sin (2 \\pi \\cdot (\\mathbf {F} \\cdot \\mathcal {T} - \\mathbf {S})) + \\mathbf {B} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.427, + 0.788, + 0.533 + ], + "angle": 0, + "content": "where \\(\\mathcal{T}\\) is a known time window series obtained by evenly spacing the timesteps from 0 to \\(T\\). Intuitively this curve construction procedure can be viewed as a \"quantization\" layer to enforce the network to learn to represent the motion features in the frequency domain, which is useful in representing different aspects of human motion such as their timing and periodicity. In the last step, a decoder is utilized to reconstruct the original motion signals from the set of parametric latent curves." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.789, + 0.844 + ], + "angle": 0, + "content": "Decoder. To decode the latent space into the original motion space, previous works [60,61] have to use a sinusoidal positional encoding sequence with duration \\( T \\) as the proxy input to the sequence decoder. This is because their latent space is only formed by single latent vectors following a Gaussian distribution, which cannot span the time dimension. However, we observe that it usually results in unstable and inconsistent movements, as the proxy sequence is generic and usually contains less meaningful information for the decoder. Meanwhile, our method does not suffer from this problem as our latent space is built on multiple curves that can represent the motion information through time, thanks to the phase parameters. Subsequently, our decoder \\( \\mathcal{D} \\) is based on Transformer decoder architecture that takes the constructed parametric latent curve, as well as the music features as inputs, to reconstruct the corresponding dance motions. Here, we also utilize the cross-attention model [85] where we consider the sequence of and music features as key and value along with the sampled latent curves as the query. The output of the decoder is a sequence of \\( T \\) vectors in \\( \\mathbb{R}^D \\), which is then projected back to the original motion dimensions through a linear layer, to obtain the reconstructed outputs \\( \\hat{\\mathbf{x}} = p_{\\theta}(\\mathbf{x}|\\mathbf{z},\\mathbf{a}) \\). We additionally employ a global trajectory predictor to predict the global translation of the root joint based on the generated local motions [19, 46], in order to avoid intersection problems between dancers. We provide details of our global predictor in the supplemental." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.257, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Scalable Group Choreography via Variational Phase Manifold Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.784, + 0.238 + ], + "angle": 0, + "content": "Prior Network. Since the ground-truth motion is generally inaccessible at test time (i.e., we only have access to the music), we also need to learn a prior \\(\\mathcal{P}\\) to match the posterior distribution of motion from which the latent phase can be sampled. Specifically, We follow the procedure similar to Section 3.2 (Equation 1-4) to predict the Gaussian distribution conditioned on the music sequence \\(\\mathbf{a}\\), which is then used for sampling the latent phases:" + }, + { + "type": "equation", + "bbox": [ + 0.417, + 0.247, + 0.786, + 0.265 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} (\\mathbf {z} | \\mathbf {a}) = \\mathcal {N} (\\mathbf {z}; \\mu_ {\\mathcal {P}}, \\sigma_ {\\mathcal {P}}) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.273, + 0.785, + 0.395 + ], + "angle": 0, + "content": "where a Transformer encoder is used to encode the input conditioning music sequence and predict the corresponding \\(\\mu_{\\mathcal{P}}\\) and \\(\\sigma_{\\mathcal{P}}\\). We implement the prior network similarly to the encoder network, however, we use self-attention mechanism [85] to capture the global music context. Learning the conditional prior is crucial for the conditional VAE to generalize to diverse types of music and motion. Intuitively speaking, each latent variable \\(\\mathbf{z}\\) is expected to represent possible dance motions \\(\\mathbf{x}\\) conforming to the music context \\(\\mathbf{a}\\). Therefore, the prior should be able to encode different latent distributions given different musics." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.415, + 0.334, + 0.43 + ], + "angle": 0, + "content": "3.3 Training" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.438, + 0.784, + 0.468 + ], + "angle": 0, + "content": "During training, we consider the following variational lower bound [73] to mainly train our dance generation VAE model:" + }, + { + "type": "equation", + "bbox": [ + 0.293, + 0.477, + 0.785, + 0.495 + ], + "angle": 0, + "content": "\\[\n\\log p _ {\\theta} (\\mathbf {x} \\mid \\mathbf {a}) \\geq \\mathbb {E} _ {q _ {\\phi}} \\left[ \\log p _ {\\theta} (\\mathbf {x} \\mid \\mathbf {z}, \\mathbf {a}) \\right] - D _ {\\mathrm {K L}} \\left(q _ {\\phi} (\\mathbf {z} \\mid \\mathbf {x}, \\mathbf {a}) \\| p _ {\\theta} (\\mathbf {z} \\mid \\mathbf {a})\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.503, + 0.785, + 0.592 + ], + "angle": 0, + "content": "In practice, we apply the conditional VAE loss as similar to [61], which is defined as the weighted sum \\(\\mathcal{L}_{\\mathrm{cvae}} = \\mathcal{L}_{\\mathrm{rec}} + \\lambda_{\\mathrm{KL}}\\mathcal{L}_{\\mathrm{KL}}\\). In particular, the reconstruction term \\(\\mathcal{L}_{\\mathrm{rec}}\\) measures the motion reconstruction error given the decoder output (via a smooth-L1 loss). The KL divergence term \\(\\mathcal{L}_{\\mathrm{KL}}\\) compares the divergence \\(D_{\\mathrm{KL}}\\) between the posterior and the prior distribution to enforce them to be close to each other." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.594, + 0.785, + 0.7 + ], + "angle": 0, + "content": "The conditional VAE objective above is calculated for each dancer separately and cannot capture the correlation between dancers within a group. Therefore, it is essential to impose consistency among dancers and avoid strange effects such as unsynchronized dance. To this end, we propose a group consistency loss by enforcing the latent phase manifold to be similar for the same group, given the input music. Specifically, we first calculate the phase manifold features based on the frequency domain parameters as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.325, + 0.71, + 0.785, + 0.726 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} _ {2 i - 1} = \\mathbf {A} _ {i} \\sin (2 \\pi \\cdot \\mathbf {S} _ {i}), \\quad \\mathbf {P} _ {2 i} = \\mathbf {A} _ {i} \\cos (2 \\pi \\cdot \\mathbf {S} _ {i}) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.734, + 0.786, + 0.841 + ], + "angle": 0, + "content": "where \\(\\mathbf{P} \\in \\mathbb{R}^{2D}\\) is the phase manifold vector that encodes the spatial-temporal information of the motion state. The phase feature may look similar to the positional encodings of transformers [85] in the sense that both use multi-resolution sinusoidal functions. However, the phase feature is much richer in terms of representation capacity since it learns to embed the spatial (body joints) and temporal (positions in time) information of the motion curves, whereas the positional encodings only encode the position of words. Finally, our consistency objective is" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.35, + 0.128 + ], + "angle": 0, + "content": "N. Le et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "to constrain phase manifold between dancers within a group to be as close as possible to each other, which is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.31, + 0.19, + 0.786, + 0.209 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {c s c}} = D _ {\\mathrm {K L}} \\left(q _ {\\phi} (\\mathbf {z} | \\mathbf {x} ^ {m}, \\mathbf {a}) \\| \\left(q _ {\\phi} (\\mathbf {z} | \\mathbf {x} ^ {n}, \\mathbf {a})\\right) + \\| \\mathbf {P} ^ {m} - \\mathbf {P} ^ {n} \\| _ {2} ^ {2} \\right. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.221, + 0.784, + 0.327 + ], + "angle": 0, + "content": "where the first term encourages the network to map different dancers belonging to the same group \\((\\mathbf{x}^m\\) and \\(\\mathbf{x}^n)\\) into the same set of distributional phase parameters while the second term penalizes the discrepancy in their corresponding phase manifolds. In general, this loss is applied to ensure every dancer is embedded into a single unified manifold that can effectively represent their corresponding group. To summarize, our total training loss is defined as the combination of the VAE loss and the consistency loss \\(\\mathcal{L} = \\mathcal{L}_{\\mathrm{cvae}} + \\lambda_{\\mathrm{csc}}\\mathcal{L}_{\\mathrm{csc}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.328, + 0.785, + 0.462 + ], + "angle": 0, + "content": "For testing, we can efficiently generate motions for an unlimited number of dancers by indefinitely drawing samples from the learned continuous group-consistent phase manifold. It is noteworthy that for inference, we only need to process the prior network once to obtain the latent distribution. To generate a new motion, we can sample from this latent (Gaussian) distribution and use the decoder to decode it back to the motion space. This approach is much more efficient and has significantly higher scalability than previous approaches [39, 40] that is limited by the number of dancers processed simultaneously by the entire network." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.492, + 0.376, + 0.509 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.528, + 0.457, + 0.542 + ], + "angle": 0, + "content": "4.1 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.557, + 0.785, + 0.694 + ], + "angle": 0, + "content": "Our model was trained on 4 NVIDIA V100 GPUs using Adam optimizer with a fixed learning rate of \\(10^{-4}\\) and a mini-batch size of 32 per GPU. For training losses, the weights are empirically set to \\(\\lambda_{\\mathrm{KL}} = 5\\times 10^{-4}\\) and \\(\\lambda_{\\mathrm{csc}} = 10^{-4}\\), respectively. The Transformer encoders and decoders consist of 5 layers for both encoder, decoder, and prior Network with 512-dimensional hidden units. Meanwhile, the number of latent phase channels is set to 256. To further capture the periodic nature of the phase feature, we also use Siren activation following the initialization scheme as in [70]. This can effectively model the periodicity inherent in the motion data, and thus can benefit motion synthesis [68]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.72, + 0.446, + 0.736 + ], + "angle": 0, + "content": "4.2 Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.784, + 0.841 + ], + "angle": 0, + "content": "Dataset In our experiments, we utilize the AIOZ-GDance [40] and AIST-M [90] datasets. AIOZ-GDance is the largest music-driven dataset focusing on group dance, encompassing paired music and 3D group motions extracted from in-the-wild videos through a semi-automatic process. This dataset spans 7 dance styles and 16 music genres. For consistency, we adhere to the training and testing split outlined in [40] during our experiments." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.257, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Scalable Group Choreography via Variational Phase Manifold Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.312 + ], + "angle": 0, + "content": "Evaluation Protocol We employ several metrics to assess the quality of individual dance motions, including Frechet Inception Distance (FID) [20, 49], Motion-Music Consistency (MMC) [49], and Generation Diversity (GenDiv) [26, 41, 49], along with the Physical Foot Contact score (PFC) [83]. Specifically, the FID score gauges the realism of individual dance movements concerning the ground-truth dance. MMC assesses the matching similarity between motion and music beats, reflecting how well-generated dances synchronize with the music's rhythm. GenDiv is computed as the average pairwise distance of kinetic features among motions [58]. PFC evaluates the physical plausibility of foot movements by determining the agreement between the acceleration of the character's center of mass and the foot's velocity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.313, + 0.788, + 0.435 + ], + "angle": 0, + "content": "In assessing the group dance quality, we adopt three metrics outlined in [40]: Group Motion Realism (GMR), Group Motion Correlation (GMC), and Trajectory Intersection Frequency (TIF). Broadly, GMR gauges the realism of generated group motions in comparison to ground-truth data, employing Frechet Inception Distance on extracted group motion features. GMC evaluates the synchronization among dancers within the generated group by computing their cross-correlation. TIF quantifies the frequency of collisions among the generated dancers during their dance movements." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.454, + 0.788, + 0.56 + ], + "angle": 0, + "content": "Baselines Our method is subjected to comparison with various recent techniques in music-driven dance generation, namely FACT [49], Transflower [59], and EDGE [83]. These approaches are adapted for benchmarking within group dance generation context, as outlined in [40], considering that their original designs were tailored for single-dance scenarios. Additionally, our evaluation includes comparisons with GDanceR [40], GCD [39], and DanY [90]. All mentioned works are specifically designed for the group choreography generation." + }, + { + "type": "table", + "bbox": [ + 0.226, + 0.568, + 0.784, + 0.698 + ], + "angle": 0, + "content": "
DatasetMethodFID↓MMC↑GenDiv↑PFC↓GMR↓GMC↑TIF↓
AIOZ-GDANCE [40]FACT [49]56.200.2228.643.52101.5262.680.321
Transflower [59]37.730.2178.743.0781.1760.780.332
EDGE [83]31.400.2649.572.6363.3561.720.356
GDANCER [40]43.900.2509.233.0551.2779.010.217
GCD [39]31.160.26110.872.5331.4780.970.167
Ours31.010.27110.982.3330.0884.520.163
AIST-M [90]GDANCER [40]52.900.2226.521.9365.1360.560.121
GCD [39]35.360.24510.971.5242.5272.150.083
DanY [90]40.250.24011.401.6550.2963.530.137
Ours31.490.25711.811.4241.2478.640.076
" + }, + { + "type": "table_caption", + "bbox": [ + 0.383, + 0.702, + 0.619, + 0.716 + ], + "angle": 0, + "content": "Table 1: Performance comparison." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.742, + 0.442, + 0.758 + ], + "angle": 0, + "content": "4.3 Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Quality Comparison Table 1 presents a comparison among the baselines FACT [49], Transflower [59], EDGE [83], GDanceR [40], GCD [39], and our proposed manifold-based method. The results clearly demonstrate that our model almost outperforms all baselines across all evaluations on two datasets AIOZ-GDANCE [40] and AIST-M [90]. We observe that recent diffusion-based dance" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.35, + 0.128 + ], + "angle": 0, + "content": "N. Le et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.299 + ], + "angle": 0, + "content": "generation models such as EDGE or GCD yield competitive performance on both single-dance metrics (FID, MMC, GenDiv, and PFC) and group dance metrics (GMR, GMC, and TIF). However, due to limitations in their training procedures, they still struggle when dealing with generating multiple dancing motions in the context of a high quantity of dancers, as shown by their low performance compared to our proposal. This result implies that our method successfully preserves the performance of dancing motions when the number of dancers is increased. Besides, Figure 3 also shows that our proposed methods outperform other state-of-the-art methods such as GDanceR or GCD in dealing with monotonous, repetitive, sinking, and overlapping dance motions." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.313, + 0.782, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.592 + ], + "angle": 0, + "content": "Fig. 3: Visualization of a dancing sample between different methods. GDanceR displays monotonous, repetitive, or sinking dance motions. GCD exhibits more divergence in dance motions, yet dancers may intersect since their optimization does not address this issue explicitly. Blue boxes mark these issues. In contrast, our manifold-based solution ensures the divergence of dancing motions, while the phase motion path demonstrates its effectiveness in addressing floating and crossing issues in group dances." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.611, + 0.784, + 0.719 + ], + "angle": 0, + "content": "
#Generated DancersMethodFID↓MMC↑GenDiv↑GMR↓GMC↑TIF↓
5GDanceR [40]44.190.2498.9955.0578.720.218
GCD [39]35.080.2649.9238.4381.440.168
Ours31.350.26810.0532.5884.560.161
10GDanceR [40]57.930.2257.6373.3272.790.386
GCD [39]N/A (Memory)N/A (Memory)N/A (Memory)N/A (Memory)N/A (Memory)N/A (Memory)
Ours32.190.2698.9934.3286.960.193
100GDanceR [40]N/A (Memory)N/A (Memory)N/A (Memory)N/A (Memory)N/A (Memory)N/A (Memory)
GCD [39]N/A (Memory)N/A (Memory)N/A (Memory)N/A (Memory)N/A (Memory)N/A (Memory)
Ours30.970.2688.7638.1385.730.222
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.723, + 0.788, + 0.767 + ], + "angle": 0, + "content": "Table 2: Performance of group dance generation methods when we increase the number of generated dancers. The experiments are done with common consumer GPUs with 4GB memory. (N/A means models could not run due to inadequate memory footprint)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Scalable Generation Analysis Table 2 illustrates the performance of different group dance generation methods (GCD, GDancer, and Ours) when generating dance movements with an increasing number of dancers. When the number of" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.257, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Scalable Group Choreography via Variational Phase Manifold Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.314 + ], + "angle": 0, + "content": "dancers is increased to 10, GCD appears to run out of memory, which is also observed in GDanceR when the number of dancers increases to 100. Regardless of the number of dancers, our method consistently achieves stable and competitive results. This implies that our proposed method successfully addresses the scalability issue in group dance generation without compromising the overall performance of each individual dance motion. Figure 4 illustrates the memory consumption to generate dance motions in groups for each method. Noticeably, our proposal still achieves the highest performance while consuming immensely fewer resources required for generating group dance motions (See Figure 5 for illustrations). This, again, indicates that our method successfully breaks the barrier of limited generated dancers by using the manifold." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.336, + 0.789, + 0.548 + ], + "angle": 0, + "content": "Ablation Analysis Table 3 presents the performance improvements achieved through the integration of consistency loss and phase manifold. Additionally, we showcase the effectiveness of our proposed approach across three different backbones: Transformer [49, 85], LSTM [21], and CNN [99]. Evaluation metrics including FID, GMR, and GMC are utilized. The results indicate that the absence of consistency loss leads to an increase in GMR and a decrease in GMC, suggesting a significant enhancement in the realism and correlation of group dance motions facilitated by the inclusion of the proposed objective. Meanwhile, with out the phase manifold, the model exhibits remarkably higher scores in both the FID and GMR metrics, suggesting that phase manifold can effectively improve the distinction in dance motions while maintaining the realism of group dances, even when the number of dancers in a group is large. In comparing three backbones—Transformer, LSTM, and CNN—we have observed that the chosen Transformer backbone achieved the best results compared to LSTM or CNN." + }, + { + "type": "table", + "bbox": [ + 0.224, + 0.587, + 0.49, + 0.671 + ], + "angle": 0, + "content": "
MethodFID↓GMR↓GMC↑
Ours31.0130.0884.52
Ours w/o Consistency Loss35.3557.6366.72
Ours w/o Phase Manifold41.7845.3277.93
Ours w LSTM41.2947.4771.82
Ours w CNN36.9944.9475.77
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.681, + 0.493, + 0.71 + ], + "angle": 0, + "content": "Table 3: Module contribution and loss analysis." + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.561, + 0.789, + 0.683 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.683, + 0.788, + 0.71 + ], + "angle": 0, + "content": "Fig. 4: Memory usage vs. number of dancers in different dance generators." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.739, + 0.357, + 0.755 + ], + "angle": 0, + "content": "4.4 User Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.842 + ], + "angle": 0, + "content": "User studies are vital for evaluating generative models, as user perception is pivotal for downstream applications; thus, we conducted two studies with around 70 participants each, diverse in background, with experience in music and dance, aged between 20 to 40, consisting of approximately \\(47\\%\\) females and \\(53\\%\\) males, to assess the effectiveness of our approach in group choreography generation." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.35, + 0.128 + ], + "angle": 0, + "content": "N. Le et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "In the user study, we aim to assess the realism of sample outputs with more and more dancers. Specifically, participants assign scores ranging from 0 to 10 to evaluate the realism of each dance clip with 2 to 10 dancers. Figure 6 shows that, across all methods, the more the number of dancers is increased, the lower the realism is found. However, the drop in realism of our proposed method is the least compared to GCD and GDanceR. The results indicate our method's good performance compared to other baselines when the number of dancers increases." + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.268, + 0.498, + 0.479 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.486, + 0.496, + 0.5 + ], + "angle": 0, + "content": "Fig. 5: Visualization of scalable dancers." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.268, + 0.787, + 0.473 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.516, + 0.474, + 0.785, + 0.499 + ], + "angle": 0, + "content": "Fig. 6: Realism between different methods when number of dancers is varied." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.532, + 0.355, + 0.547 + ], + "angle": 0, + "content": "5 Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.567, + 0.788, + 0.657 + ], + "angle": 0, + "content": "While our approach leverages the VAE as a primary solution for generating a manifold, it is important to acknowledge certain inherent limitations associated with this choice. One notable challenge is the susceptibility to issues such as posterior collapse and unstable sampling within the VAE framework. These challenges can result in generated group dance motions that may not consistently meet performance expectations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.748 + ], + "angle": 0, + "content": "One specific manifestation of this limitation is the potential for false decoding when sampling points that lie too far from the learned distribution. This scenario can lead to unexpected rotations or disruptions in the physics of the generated content. The impact of this problem becomes evident in instances where the generated samples deviate significantly from the anticipated distribution, introducing inaccuracies and distortions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.84 + ], + "angle": 0, + "content": "To address these challenges, we recognize the need for ongoing efforts to mitigate the effects of posterior collapse and unstable sampling. While the problem is acknowledged, our approach incorporates measures to limit its impact. Future research directions could explore alternative generative models or additional techniques to enhance the robustness and reliability of the generated results in the face of these identified limitations." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.257, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Scalable Group Choreography via Variational Phase Manifold Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.173, + 0.786, + 0.214 + ], + "angle": 0, + "content": "1. Ahn, H., Kim, J., Kim, K., Oh, S.: Generative autoregressive networks for 3d dancing move synthesis from music. IEEE Robotics and Automation Letters (RA-L) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.215, + 0.786, + 0.257 + ], + "angle": 0, + "content": "2. Alaoui, S.F., Henry, C., Jacquemin, C.: Physical modelling for interactive installations and the performing arts. International Journal of Performance Arts and Digital Media (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.257, + 0.786, + 0.283 + ], + "angle": 0, + "content": "3. Alemi, O., Françoise, J., Pasquier, P.: Groovenet: Real-time music-driven dance movement generation using artificial neural networks. Networks (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.283, + 0.786, + 0.324 + ], + "angle": 0, + "content": "4. Alexanderson, S., Nagy, R., Beskow, J., Henter, G.E.: Listen, denoise, action! audio-driven motion synthesis with diffusion models. ACM Transactions on Graphics (TOG) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.325, + 0.786, + 0.365 + ], + "angle": 0, + "content": "5. Au, H.Y., Chen, J., Jiang, J., Guo, Y.: Choreograph: Music-conditioned automatic dance choreography over a style and tempo consistent dynamic graph. In: ACM International Conference on Multimedia (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.365, + 0.786, + 0.392 + ], + "angle": 0, + "content": "6. Bisig, D.: Generative dance-a taxonomy and survey. In: International Conference on Movement and Computing (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.392, + 0.786, + 0.419 + ], + "angle": 0, + "content": "7. Cervantes, P., Sekikawa, Y., Sato, I., Shinoda, K.: Implicit neural representations for variable length human motion generation. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.419, + 0.786, + 0.446 + ], + "angle": 0, + "content": "8. Chan, C., Ginosar, S., Zhou, T., Efros, A.A.: Everybody dance now. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.446, + 0.786, + 0.487 + ], + "angle": 0, + "content": "9. Chen, K., Tan, Z., Lei, J., Zhang, S.H., Guo, Y.C., Zhang, W., Hu, S.M.: Choreomaster: choreography-oriented music-driven dance synthesis. ACM Transactions on Graphics (TOG) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.487, + 0.786, + 0.528 + ], + "angle": 0, + "content": "0. Fan, R., Xu, S., Geng, W.: Example-based automatic music-driven conventional dance motion synthesis. IEEE Transactions on Visualization and Computer Graphics (TVCG) (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.528, + 0.786, + 0.555 + ], + "angle": 0, + "content": "1. Feng, B., Ao, T., Liu, Z., Ju, W., Liu, L., Zhang, M.: Robust dancer: Long-term 3d dance synthesis using unpaired data. arXiv (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.555, + 0.786, + 0.596 + ], + "angle": 0, + "content": "2. Ferreira, J.P., Coutinho, T.M., Gomes, T.L., Neto, J.F., Azevedo, R., Martins, R., Nascimento, E.R.: Learning to dance: A graph convolutional adversarial network to generate realistic dance motions from audio. Computers & Graphics (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.596, + 0.786, + 0.623 + ], + "angle": 0, + "content": "3. Fink, B., Bläsing, B., Ravignani, A., Shackelford, T.K.: Evolution and functions of human dance. Evolution and Human Behavior (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.623, + 0.786, + 0.65 + ], + "angle": 0, + "content": "4. Gartner, E., Metz, L., Andriluka, M., Freeman, C.D., Sminchisescu, C.: Transformer-based learned optimization. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.65, + 0.786, + 0.69 + ], + "angle": 0, + "content": "5. Ghorbani, S., Ferstl, Y., Holden, D., Troje, N.F., Carbonneau, M.A.: Zeroeggs: Zero-shot example-based gesture generation from speech. In: Computer Graphics Forum (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.691, + 0.786, + 0.718 + ], + "angle": 0, + "content": "6. Ghosh, A., Cheema, N., Oguz, C., Theobalt, C., Slusallek, P.: Synthesis of compositional animations from textual descriptions. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.718, + 0.786, + 0.746 + ], + "angle": 0, + "content": "7. Gong, K., Lian, D., Chang, H., Guo, C., Zuo, X., Jiang, Z., Wang, X.: Tm2d: Bimodality driven 3d dance generation via music-text integration. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.746, + 0.786, + 0.772 + ], + "angle": 0, + "content": "8. Guo, C., Zou, S., Zuo, X., Wang, S., Ji, W., Li, X., Cheng, L.: Generating diverse and natural 3d human motions from text. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.786, + 0.799 + ], + "angle": 0, + "content": "9. He, C., Saito, J., Zachary, J., Rushmeier, H., Zhou, Y.: Nemf: Neural motion fields for kinematic animation. NeurIPS 35, 4244-4256 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.786, + 0.841 + ], + "angle": 0, + "content": "20. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: NeurIPS (2017)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.173, + 0.786, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.35, + 0.127 + ], + "angle": 0, + "content": "N. Le et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "21. Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural Computation (1997)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.204 + ], + "angle": 0, + "content": "22. Holden, D., Komura, T., Saito, J.: Phase-functional neural networks for character control. ACM Transactions on Graphics (TOG) 36(4), 1-13 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.205, + 0.785, + 0.231 + ], + "angle": 0, + "content": "23. Holden, D., Saito, J., Komura, T.: A deep learning framework for character motion synthesis and editing. ACM Transactions on Graphics (TOG) (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.232, + 0.785, + 0.259 + ], + "angle": 0, + "content": "24. Holden, D., Saito, J., Komura, T., Joyce, T.: Learning motion manifolds with convolutional autoencoders. In: SIGGRAPH Asia 2015 technical briefs (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.26, + 0.785, + 0.3 + ], + "angle": 0, + "content": "25. Hong, F., Zhang, M., Pan, L., Cai, Z., Yang, L., Liu, Z.: Avatarclip: zero-shot text-driven generation and animation of 3d avatars. ACM Transactions on Graphics (TOG) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.301, + 0.785, + 0.328 + ], + "angle": 0, + "content": "26. Huang, R., Hu, H., Wu, W., Sawada, K., Zhang, M.: Dance revolution: Long sequence dance generation with music via curriculum learning. CoRR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.329, + 0.785, + 0.384 + ], + "angle": 0, + "content": "27. Huang, Y., Zhang, J., Liu, S., Bao, Q., Zeng, D., Chen, Z., Liu, W.: Genre-conditioned long-term 3d dance generation driven by music. In: ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 4858-4862. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.385, + 0.785, + 0.424 + ], + "angle": 0, + "content": "28. Jiang, C., Cornman, A., Park, C., Sapp, B., Zhou, Y., Anguelov, D., et al.: Motiondiffuser: Controllable multi-agent motion prediction using diffusion. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.426, + 0.785, + 0.466 + ], + "angle": 0, + "content": "29. Jiang, Y., Won, J., Ye, Y., Liu, C.K.: Drop: Dynamics responses from human motion prior and projective dynamics. In: SIGGRAPH Asia 2023 Conference papers (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.468, + 0.785, + 0.494 + ], + "angle": 0, + "content": "30. Joshi, M., Chakrabarty, S.: An extensive review of computational dance automation techniques and applications. Proceedings of the Royal Society A (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.496, + 0.785, + 0.522 + ], + "angle": 0, + "content": "31. Karunratanakul, K., Preechakul, K., Suwajanakorn, S., Tang, S.: Guided motion diffusion for controllable human motion synthesis. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.523, + 0.785, + 0.55 + ], + "angle": 0, + "content": "32. Kico, I., Grammalidis, N., Christidis, Y., Liarokapis, F.: Digitization and visualization of folk dances in cultural heritage: A review. Inventions (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.551, + 0.785, + 0.577 + ], + "angle": 0, + "content": "33. Kim, J.W., Fouad, H., Hahn, J.K.: Making them dance. In: AAAI Fall Symposium: Aurally Informed Performance (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.578, + 0.785, + 0.618 + ], + "angle": 0, + "content": "34. Kim, J., Oh, H., Kim, S., Tong, H., Lee, S.: A brand new dance partner: Music-conditioned pluralistic dancing controlled by multiple dance genres. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.62, + 0.785, + 0.646 + ], + "angle": 0, + "content": "35. Kim, T.h., Park, S.I., Shin, S.Y.: Rhythmic-motion synthesis based on motion-beat analysis. ACM Transactions on Graphics (TOG) (2003)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.647, + 0.761, + 0.661 + ], + "angle": 0, + "content": "36. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. In: ICLR (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.662, + 0.785, + 0.688 + ], + "angle": 0, + "content": "37. Kolotouros, N., Alldieck, T., Zanfir, A., Bazavan, E., Fieraru, M., Sminchisescu, C.: Dreamhuman: Animatable 3d avatars from text. In: NeurIPS (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.689, + 0.785, + 0.715 + ], + "angle": 0, + "content": "38. Lannan, N., Zhou, L., Fan, G.: Human motion enhancement via tobit kalman filter-assisted autoencoder. IEEE Access 10, 29233-29251 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.717, + 0.785, + 0.757 + ], + "angle": 0, + "content": "39. Le, N., Do, T., Do, K., Nguyen, H., Tjiputra, E., Tran, Q.D., Nguyen, A.: Controllable group choreography using contrastive diffusion. ACM Transactions on Graphics (TOG) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.758, + 0.785, + 0.785 + ], + "angle": 0, + "content": "40. Le, N., Pham, T., Do, T., Tjiputra, E., Tran, Q.D., Nguyen, A.: Music-driven group choreography. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.786, + 0.785, + 0.812 + ], + "angle": 0, + "content": "41. Lee, H.Y., Yang, X., Liu, M.Y., Wang, T.C., Lu, Y.D., Yang, M.H., Kautz, J.: Dancing to music. In: NeurIPS (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "42. Lee, L., Lin, Z., Hu, R., Gong, Z., Kumar, A., Li, T., Li, S., Hui, P.: When creators meet the metaverse: A survey on computational arts. CoRR (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.257, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Scalable Group Choreography via Variational Phase Manifold Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "43. Lee, M., Lee, K., Park, J.: Music similarity-based approach to generating dance motion sequence. Multimedia tools and applications (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.785, + 0.202 + ], + "angle": 0, + "content": "44. Li, B., Zhao, Y., Sheng, L.: Dancenet3d: Music based dance generation with parametric motion transformer. In: AAAI (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.203, + 0.785, + 0.242 + ], + "angle": 0, + "content": "45. Li, B., Zhao, Y., Zhelun, S., Sheng, L.: Danceformer: Music conditioned 3d dance generation with parametric motion transformer. In: Proceedings of the AAAI Conference on Artificial Intelligence (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.243, + 0.785, + 0.284 + ], + "angle": 0, + "content": "46. Li, J., Villegas, R., Ceylan, D., Yang, J., Kuang, Z., Li, H., Zhao, Y.: Task-generic hierarchical human motion prior using vaes. In: International Conference on 3D Vision (3DV) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.285, + 0.785, + 0.324 + ], + "angle": 0, + "content": "47. Li, J., Kang, D., Pei, W., Zhe, X., Zhang, Y., He, Z., Bao, L.: Audio2gestures: Generating diverse gestures from speech audio with conditional variational autoencoders. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.326, + 0.785, + 0.366 + ], + "angle": 0, + "content": "48. Li, R., Zhao, J., Zhang, Y., Su, M., Ren, Z., Zhang, H., Tang, Y., Li, X.: Finedance: A fine-grained choreography dataset for 3d full body dance generation. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.367, + 0.785, + 0.393 + ], + "angle": 0, + "content": "49. Li, R., Yang, S., Ross, D.A., Kanazawa, A.: Ai choreographer: Music conditioned 3d dance generation with aist++. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.394, + 0.785, + 0.419 + ], + "angle": 0, + "content": "50. Li, W., Dai, B., Zhou, Z., Yao, Q., Wang, B.: Controlling character motions without observable driving source. In: WACV (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.42, + 0.785, + 0.46 + ], + "angle": 0, + "content": "51. Liang, H., Zhang, W., Li, W., Yu, J., Xu, L.: Intergen: Diffusion-based multi-human motion generation under complex interactions. arXiv preprint arXiv:2304.05684 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.461, + 0.785, + 0.487 + ], + "angle": 0, + "content": "52. Lin, J., Zeng, A., Lu, S., Cai, Y., Zhang, R., Wang, H., Zhang, L.: Motion-x: A large-scale 3d expressive whole-body human motion dataset. NeurIPS (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.488, + 0.785, + 0.514 + ], + "angle": 0, + "content": "53. Ling, H.Y., Zinno, F., Cheng, G., Van De Panne, M.: Character controllers using motion vaes. ACM Transactions on Graphics (TOG) 39(4), 40-1 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.515, + 0.785, + 0.542 + ], + "angle": 0, + "content": "54. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: SMPL: A skinned multi-person linear model. ACM Transactions on Graphics (TOG) (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.543, + 0.785, + 0.568 + ], + "angle": 0, + "content": "55. Martinez, J., Black, M.J., Romero, J.: On human motion prediction using recurrent neural networks. In: CVPR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.569, + 0.785, + 0.609 + ], + "angle": 0, + "content": "56. Mo, C.A., Hu, K., Long, C., Wang, Z.: Continuous intermediate token learning with implicit motion manifold for keyframe based motion interpolation. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.61, + 0.785, + 0.65 + ], + "angle": 0, + "content": "57. Ofli, F., Erzin, E., Yemez, Y., Tekalp, A.M.: Learn2dance: Learning statistical music-to-dance mappings for choreography synthesis. IEEE Transactions on Multimedia (TMM) (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.651, + 0.785, + 0.677 + ], + "angle": 0, + "content": "58. Onuma, K., Faloutsos, C., Hodgins, J.K.: Fmdistance: A fast and effective distance function for motion capture data. In: Eurographics (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.678, + 0.785, + 0.718 + ], + "angle": 0, + "content": "59. Perez, G.V., Beskow, J., Henter, G., Holzapfel, A., Oudeyer, P.Y., Alexanderson, S.: Transflower: probabilistic autoregressive dance generation with multimodal attention. ACM Transactions on Graphics (TOG) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.719, + 0.785, + 0.745 + ], + "angle": 0, + "content": "60. Petrovich, M., Black, M.J., Varol, G.: Action-conditioned 3d human motion synthesis with transformer vae. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.746, + 0.785, + 0.772 + ], + "angle": 0, + "content": "61. Petrovich, M., Black, M.J., Varol, G.: Temos: Generating diverse human motions from textual descriptions. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.773, + 0.785, + 0.813 + ], + "angle": 0, + "content": "62. Qi, Q., Zhuo, L., Zhang, A., Liao, Y., Fang, F., Liu, S., Yan, S.: Diffdance: Cascaded human motion diffusion model for dance generation. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 1374-1382 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "63. Raab, S., Leibovitch, I., Li, P., Aberman, K., Sorkine-Hornung, O., Cohen-Or, D.: Modi: Unconditional motion synthesis from diverse data. In: CVPR (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.35, + 0.127 + ], + "angle": 0, + "content": "N. Le et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "64. Ren, X., Li, H., Huang, Z., Chen, Q.: Self-supervised dance video synthesis conditioned on music. In: ACMMM (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.203 + ], + "angle": 0, + "content": "65. Safonova, A., Hodgins, J.K.: Construction and optimal search of interpolated motion graphs. In: SIGGRAPH (2007)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.785, + 0.231 + ], + "angle": 0, + "content": "66. Shafir, Y., Tevet, G., Kapon, R., Bermano, A.H.: Human motion diffusion as a generative prior. arXiv (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.232, + 0.785, + 0.272 + ], + "angle": 0, + "content": "67. Shi, J.: Application of 3d computer aided system in dance creation and learning. In: International Conference on Machine Learning and Big Data Analytics for IoT Security and Privacy (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.273, + 0.785, + 0.3 + ], + "angle": 0, + "content": "68. Shi, M., Starke, S., Ye, Y., Komura, T., Won, J.: Phasemp: Robust 3d pose estimation via phase-conditioned human motion prior. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.301, + 0.785, + 0.327 + ], + "angle": 0, + "content": "69. Shiratori, T., Nakazawa, A., Ikeuchi, K.: Dancing-to-music character animation. Computer Graphics Forum (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.328, + 0.785, + 0.355 + ], + "angle": 0, + "content": "70. Sitzmann, V., Martel, J., Bergman, A., Lindell, D., Wetzstein, G.: Implicit neural representations with periodic activation functions. In: NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.356, + 0.785, + 0.396 + ], + "angle": 0, + "content": "71. Siyao, L., Yu, W., Gu, T., Lin, C., Wang, Q., Qian, C., Loy, C.C., Liu, Z.: Bailando: 3d dance generation by actor-critic gpt with choreographic memory. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.398, + 0.785, + 0.437 + ], + "angle": 0, + "content": "72. Soga, A., Umino, B., Longstaff, J.S.: Automatic composition of ballet sequences using a 3d motion archive. In: 1st South-Eastern European Digitization Initiative Conference (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.439, + 0.785, + 0.466 + ], + "angle": 0, + "content": "73. Sohn, K., Lee, H., Yan, X.: Learning structured output representation using deep conditional generative models. In: NeurIPS (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.467, + 0.785, + 0.494 + ], + "angle": 0, + "content": "74. Starke, S., Mason, I., Komura, T.: Deepphase: Periodic autoencoders for learning motion phase manifolds. ACM Transactions on Graphics (TOG) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.495, + 0.785, + 0.522 + ], + "angle": 0, + "content": "75. Starke, S., Zhang, H., Komura, T., Saito, J.: Neural state machine for character-scene interactions. ACM Transactions on Graphics (TOG) 38(6), 209-1 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.523, + 0.785, + 0.562 + ], + "angle": 0, + "content": "76. Starke, S., Zhao, Y., Komura, T., Zaman, K.: Local motion phases for learning multi-contact character movements. ACM Transactions on Graphics (TOG) 39(4), 54-1 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.564, + 0.785, + 0.605 + ], + "angle": 0, + "content": "77. Sun, G., Wong, Y., Cheng, Z., Kankanhalli, M.S., Geng, W., Li, X.: Deepdance: Music-to-dance motion choreography with adversarial learning. IEEE Transactions on Multimedia (TMM) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.606, + 0.785, + 0.646 + ], + "angle": 0, + "content": "78. Sun, J., Wang, C., Hu, H., Lai, H., Jin, Z., Hu, J.F.: You never stop dancing: Nonfreezing dance generation via bank-constrained manifold projection. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.647, + 0.785, + 0.673 + ], + "angle": 0, + "content": "79. Tang, T., Jia, J., Mao, H.: Dance with melody: An LSTM-autoencoder approach to music-oriented dance synthesis. In: ACMMM (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.675, + 0.785, + 0.715 + ], + "angle": 0, + "content": "80. Tevet, G., Gordon, B., Hertz, A., Bermano, A.H., Cohen-Or, D.: Motionclip: Exposing human motion generation to clip space. In: European Conference on Computer Vision (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.716, + 0.785, + 0.743 + ], + "angle": 0, + "content": "81. Tevet, G., Raab, S., Gordon, B., Shafir, Y., Cohen-Or, D., Bermano, A.H.: Human motion diffusion model. In: ICLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.744, + 0.785, + 0.771 + ], + "angle": 0, + "content": "82. Tiwari, G., Antic, D., Lenssen, J.E., Sarafianos, N., Tung, T., Pons-Moll, G.: Posendf: Modeling human pose manifolds with neural distance fields. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.772, + 0.785, + 0.798 + ], + "angle": 0, + "content": "83. Tseng, J., Castellon, R., Liu, C.K.: Edge: Editable dance generation from music. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "84. Tsuchida, S., Fukayama, S., Hamasaki, M., Goto, M.: Aist dance video database: Multi-genre, multi-dancer, and multi-camera database for dance information processing. In: ISMIR (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.257, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Scalable Group Choreography via Variational Phase Manifold Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "85. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.217 + ], + "angle": 0, + "content": "86. Wang, Z., Jia, J., Wu, H., Xing, J., Cai, J., Meng, F., Chen, G., Wang, Y.: Group-dancer: Music to multi-people dance synthesis with style collaboration. In: ACM International Conference on Multimedia (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.785, + 0.244 + ], + "angle": 0, + "content": "87. Yalta, N., Watanabe, S., Nakadai, K., Ogata, T.: Weakly-supervised deep recurrent neural networks for basic dance step generation. In: IJCNN (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.785, + 0.286 + ], + "angle": 0, + "content": "88. Yang, S., Wu, Z., Li, M., Zhang, Z., Hao, L., Bao, W., Zhuang, H.: Qpgesture: Quantization-based and phase-guided motion matching for natural speech-driven gesture generation. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.327 + ], + "angle": 0, + "content": "89. Yang, Z., Wen, Y.H., Chen, S.Y., Liu, X., Gao, Y., Liu, Y.J., Gao, L., Fu, H.: Keyframe control of music-driven 3d dance generation. IEEE Transactions on Visualization and Computer Graphics (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.37 + ], + "angle": 0, + "content": "90. Yao, S., Sun, M., Li, B., Yang, F., Wang, J., Zhang, R.: Dance with you: The diversity controllable dancer generation via diffusion models. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 8504-8514 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.397 + ], + "angle": 0, + "content": "91. Ye, Z., Wu, H., Jia, J., Bu, Y., Chen, W., Meng, F., Wang, Y.: Choreonet: Towards music to dance synthesis with choreographic action unit. In: ACMMM (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.438 + ], + "angle": 0, + "content": "92. Yin, W., Yin, H., Baraka, K., Kragic, D., Björkman, M.: Dance style transfer with cross-modal transformer. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 5058-5067 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.466 + ], + "angle": 0, + "content": "93. Yu, Z., Cheng, W., Liu, X., Wu, W., Lin, K.Y.: Monohuman: Animatable human neural field from monocular video. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.494 + ], + "angle": 0, + "content": "94. Zhang, X., Yang, S., Xu, Y., Zhang, W., Gao, L.: Mining and applying composition knowledge of dance moves for style-concentrated dance generation. In: AAAI (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.785, + 0.522 + ], + "angle": 0, + "content": "95. Zhou, L., Luo, Y.: A spatio-temporal learning for music conditioned dance generation. In: International Conference on Multimodal Interaction (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.785, + 0.549 + ], + "angle": 0, + "content": "96. Zhou, Y., Barnes, C., Lu, J., Yang, J., Li, H.: On the continuity of rotation representations in neural networks. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.577 + ], + "angle": 0, + "content": "97. Zhou, Z., Wang, B.: Ude: A unified driving engine for human motion generation. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.605 + ], + "angle": 0, + "content": "98. Zhu, Y., Olszewski, K., Wu, Y., Achlioptas, P., Chai, M., Yan, Y., Tulyakov, S.: Quantized gan for complex music generation from dance videos. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.785, + 0.647 + ], + "angle": 0, + "content": "99. Zhuang, W., Wang, C., Chai, J., Wang, Y., Shao, M., Xia, S.: Music2dance: Dancenet for music-driven dance generation. ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM) (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.647 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Scalable Group Choreography via Variational Phase Manifold Learning/6dc9e2ab-125f-402e-97a2-dfb708ce273d_origin.pdf b/2024/Scalable Group Choreography via Variational Phase Manifold Learning/6dc9e2ab-125f-402e-97a2-dfb708ce273d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..38af23e422527b3a838c78bd5382a241ceb01b31 --- /dev/null +++ b/2024/Scalable Group Choreography via Variational Phase Manifold Learning/6dc9e2ab-125f-402e-97a2-dfb708ce273d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2b87248dd21853c9f9de433d4b07db1baabcd4e04666d595d247891db9951322 +size 6900640 diff --git a/2024/Scalable Group Choreography via Variational Phase Manifold Learning/full.md b/2024/Scalable Group Choreography via Variational Phase Manifold Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..69e346e7008cfd1a573fe971f8f12f35f05ef40a --- /dev/null +++ b/2024/Scalable Group Choreography via Variational Phase Manifold Learning/full.md @@ -0,0 +1,318 @@ +# Scalable Group Choreography via Variational Phase Manifold Learning + +Nhat $\mathrm{Le}^{1}$ , Khoa $\mathrm{Do}^{2,3}$ , Xuan Bui $^{2,3}$ , Tuong $\mathrm{Do}^{1,4}$ , Erman Tjiputra $^{1}$ , Quang D. Tran $^{1}$ , and Anh Nguyen $^{4}$ + +1 AIOZ, Singapore + +2 University of Science, Ho Chi Minh City, Vietnam +3 Vietnam National University, Ho Chi Minh City, Vietnam +4 University of Liverpool, Liverpool L69 3BX, United Kingdom + +Abstract. Generating group dance motion from the music is a challenging task with several industrial applications. Although several methods have been proposed to tackle this problem, most of them prioritize optimizing the fidelity in dancing movement, constrained by predetermined dancer counts in datasets. This limitation impedes adaptability to real-world applications. Our study addresses the scalability problem in group choreography while preserving naturalness and synchronization. In particular, we propose a phase-based variational generative model for group dance generation on learning a generative manifold. Our method achieves high-fidelity group dance motion and enables the generation with an unlimited number of dancers while consuming only a minimal and constant amount of memory. The intensive experiments on two public datasets show that our proposed method outperforms recent state-of-the-art approaches by a large margin and is scalable to a great number of dancers beyond the training data. + +# 1 Introduction + +The widespread availability of digital social media platforms has led to a significant increase in the popularity of creating dance videos. This heightened interest has resulted in the daily production and viewing of millions of dance videos on various online platforms [13, 32]. Recently, researchers from the computer vision and computer graphics community have focused on developing methods for generating authentic dance movements in response to music [6, 94]. These advancements have broad implications and find applications in diverse areas, including animation [49, 50], virtual idols [59], virtual metaverses [42], and dance education [2, 67, 72]. Such techniques empower artists, animators, and educators by providing powerful tools to enhance their creative pursuits and improve the overall dance experience for performers and audiences alike. + +Recently, significant strides have been made in creating dance motions for solo dance performers [12,26,34,49,59,71,83]. However, generating synchronized group dance movements that are both lifelike and in harmony with music remains a complex challenge [9,27,87]. The introduction of AIOZ-GDance [40] stands as + +![](images/9a90bad30506cda01255b6a5143c3095f8e6726525ee39c3f98d8b43d43c3176.jpg) +Fig. 1: We present a new group dance generation method that can generate a large number of dancers within a fixed resource consumption. The illustration shows a generated group dance sample with 100 dancers. + +![](images/61ccfb172c209e5731fa05a7051c23fa14c514d12c91f6ea07bc1bb086dc70a5.jpg) + +![](images/cb1cbcb4bd334846ab2e10e79ee6ef277429dbb37d9d038b4efcefea77391613.jpg) + +the extensive dataset to aid in generating group choreography. Different aspects of group dance motion such as the study of consistency and diversity among the movements of dancers are explored in [39]. However, despite recent advances in the field, existing methods can either only generate dances for a limited number of dancers given the input audio [40, 86, 90], or would consume an expensively large amount of memory due to their design nature [39, 40]. Typically, these methods are usually based on specialized collaborative mechanisms such as cross-entity attention [40, 51], group global attention [39], or residual communication block [66] achieve coherent multi-person synthesis. In other words, their innate architectural design requires the network to process and synthesize all motions simultaneously, leading to excessively growing computational/memory costs with the increasing number of inputs, and hence extremely difficult to scale up. + +Despite the comprehensive exploration of group dance motion generation, the predominant emphasis remains on optimizing the execution of the generated movements [14, 19, 31, 37, 52, 93]. These existing methodologies function within a confined scope, often tethered to a predetermined maximum number of dancers depicted in group choreography videos outlined by the predefined dataset [39]. This inherent limitation poses a potential constraint on the adaptability of trained models when transitioning to real-world applications [47]. In our pursuit, we center our efforts on addressing the critical aspect of scalability in group dance. Our primary objective revolves around not only expanding the number of dancers but also ensuring the preservation of the innate naturalness in the dance motions and the seamless synchronization between dancers throughout the entirety of their performance. + +In the field of character motion control and synthesis, existing approaches can be broadly divided into two categories: deterministic [16, 26, 40, 49, 55, 86] and probabilistic [19, 39, 59, 61, 83]. The former tries to learn direct and deterministic mapping from the input conditioning signal (such as text, audio, or user controls) to the desired output motion. However, these methods usually regress towards the mean pose as they are typically trained by minimizing a regression objective under different outputs for similar inputs, resulting in freezing or drifting motion. On the other hand, probabilistic methods attempt to capture the distribution over possible motions for a given condition, allowing diverse and high-quality generation. In particular, Variational Autoencoder (VAE) [18, 47, 60, 61], or Diffusion Models [4, 39, 81, 83, 97] have been widely + +used. Although diffusion models have recently shown strong potential in generating high-quality motions [81], they generally operate on high-dimensional space (same as the original data) [39,81,83], hence rendering them extremely challenging to use for scalable group generation. Meanwhile, standard VAE models often rely on a Gaussian latent space of fixed dimension for sampling a single latent vector to generate motions, making it difficult to adapt to a wide range of architectures. Besides, the fixed-size single latent space might also not be sufficient to represent the temporal dynamics of the whole motion sequence [7,60]. + +Recent studies have demonstrated the significant advantages of learning motion features in the frequency domain. Specifically, [22, 74-76] have found that different motion skills can be effectively represented by several phase variables that can faithfully capture the spatial-temporal alignment of a wide range of movements. This can be used to combine with common deep motion synthesis networks to enhance its capability, facilitating natural, and stable motion generations by enforcing a unidirectional motion transition (i.e., no backward direction in time) to prevent the motion from being stuck temporally [74]. Unlike those prior works, where the phases are utilized as auxiliary signals to assist in the motion synthesis process [68, 74, 88], we propose to learn to generate the phases conditioned on the input audio signal in an end-to-end manner, eliminating the need of training extra components. Our generative phase manifold can be used to generate new motions efficiently, from which the latent motion curves are sampled in frequency domain and decoded into the original motion space by a learned decoder. + +In this paper, our goal is to develop a scalable technique for group dance generation, a phase-based variational generative model for scalable group dance generation, namely Phase-conditioned Dance VAE (PDVAE). To our knowledge, PDVAE is the first method to represent the variational latent space using phase parameters in the frequency domain of the motion curves. Our method goes beyond the conventional VAE approach that typically relies on a single latent vector drawn from a Gaussian distribution, which is unable to adequately represent the temporal information of the motion sequence (e.g., the time dimension is squeezed out). Figure 1 illustrates an example of a dance motion sequence with 100 dancers generated by our model. In contrast to existing methods, our model can generate crowd-dance animations for an unlimited number of dancers without increasing computational burden (i.e., memory consumption remains constant), while still maintaining the performance and fidelity of the generated choreographies. To summarize, our key contributions are as follows: + +- We introduce PDVAE, a phase-based variational generative model for scalable group dance generation. The method focuses on generating large-scale group dance under limited resources. +- To effectively learn the manifold that is group-consistent (i.e., dancers within a group lie upon the same manifold), we propose a group consistency loss that enforces the networks to encode the latent phase manifold to be identical for the same group given the input music. + +- Extensive experiments along with thorough user study evaluations demonstrate the state-of-the-art performance of our model while achieving effective scalability. + +# 2 Related Work + +# 2.1 Music-driven Choreography + +Crafting natural human choreography derived from music presents a multifaceted challenge [30,62,89]. Certain methodologies integrate constraints based on music-motion similarity matching to ensure coherence between the generated motion and the music [10,33,35,43,65,69]. However, many of these approaches rely considerably on heuristic algorithms, stitching together pre-existing dance segments sourced from a limited music-dance database [10, 48]. While successful in generating extended and realistic dance sequences, these methods face limitations when endeavoring to create entirely novel dance fragments [57, 90]. + +In the recent period, advancements have been evident in the domain of converting music into dance movements through various techniques such as Convolutional Networks (CNN) [1,8,77,91,92,99], Recurrent Networks (RNN) [3,26,77,79,87], Graph Neural Networks (GNN) [5,12,64,95], Generative Adversarial Networks (GAN) [41,77], or Transformer models [34,44,45,49,59,71]. These methodologies typically depend on various inputs, including the present music and a concise history of previous dance movements, to forecast forthcoming sequences of human poses. This innovation introduces a music-text feature fusion module, engineered to amalgamate inputs into a motion decoder, thereby enabling the creation of dance sequences conditioned upon both musical and textual instructions [17]. + +While these techniques show promise in generating authentic and lifelike dance motions, they often struggle to synchronize movements seamlessly across multiple dancers [40]. Specifically, achieving coordination and harmony among dancers necessitates consideration of spatial and temporal relationships, their interactions, and the overall choreographic structure [84]. Consequently, further advancements in this domain are being pursued to tackle these complexities [25, 47, 59, 98]. For instance, Perez et al. [59] integrate a multimodal transformer encoder with a normalizing-flow-based decoder to estimate a probability distribution encompassing potential subsequent poses. Additionally, Feng et al. [11] enable long-term generation by imposing a motion repeat constraint to forecast future frames while considering historical motions. A recent study by Le et al. [39] investigates consistency and diversity factors between the generated motions of two or more dancers within a timeframe. However, these approaches are limited to a predefined number of dancers, heavily constrained by the maximum number of dancers in the dataset-provided videos. + +# 2.2 Motion Manifold Learning + +Motion manifold learning has attracted considerable attention in computer vision and artificial intelligence, with the primary goal of comprehending the funda- + +mental structures inherent in human movement and dynamics [16, 24, 56, 82]. Its distinctive ability to generate human movement patterns presents numerous opportunities to comprehend intrinsic motion dynamics, manage nonlinear relationships in motion data, and acquire contextual and hierarchical representations [29, 63, 88]. Consequently, various methodologies have emerged, each contributing distinct perspectives and techniques to advance the comprehension and synthesis of human motion [28, 38, 63, 68]. + +For instance, Holden et al. [23] generate character movements by mapping high-level parameters to the human motion manifold, allowing diverse movements without manual preprocessing, and enabling post-generation editing for natural, smooth motion sequences. MotionCLIP [80] introduces a 3D human motion auto-encoder aligned with the Contrastive Language-Image Pre-training (CLIP) model's space, enabling semantic text-based motion generation, disentangled editing, and abstract language specification, leveraging CLIP's rich semantic knowledge within the motion manifold. Recently, Sun et al. [78] utilize a VQ-VAE to acquire a low-dimensional manifold, effectively cleansing the motion sequences. Regarding the scalability of group dance generation, the concept of the motion manifold emerges as a potential solution for addressing the restricted number of dancers in the dataset. This direction enables the learning of a distribution that facilitates the extraction of dance motions, allowing for the presentation of cohesive group dance sequences. + +# 3 Variational Phase Manifold Learning for Scalable Group Choreography + +# 3.1 Task Definition + +Given an input music sequence $\mathbf{a} = \{a_1, a_t, \dots, a_T\}$ with $t = \{1, \dots, T\}$ indicates the index of the music frames, our goal is to generate the group motion sequences of $N$ arbitrary dancers: $\mathbf{x} = \{x_1^1, \dots, x_T^1; \dots; x_1^N, \dots, x_T^N\}$ where $x_t^n$ is the pose of $n$ -th dancer at frame $t$ . We follow [19, 53] and represent dance as a sequence of poses in the 24-joint of the SMPL body model [54], using the 6D continuous rotation [96] for every joint, along with 3D joint positions and velocities. Additionally, the corresponding 3D root translation vectors are concatenated into the pose representations to involve the trajectory of motion. Previous group dance methods [40], which generate the whole group at once, cannot deal with the increasing number of dancers and can only create group sequences up to a pre-defined number of dancers, due to the vast complexity of the architecture. In contrast, we aim to generate group dance with an unlimited number of dancers. + +# 3.2 Phase-conditioned Dance VAE + +Our goal is to learn a continuous manifold such that the motion can be generated by sampling from this learned manifold. We assume that although different dancers within the same group may present visually distinctive movements, the + +![](images/a1e8235f870389f41b9939bcda93158b4a7097b0a339e531cc87b01f740444d5.jpg) +Fig. 2: Overview of our Phase-conditioned Dance VAE (PDVAE) for scalable group dance generation. It consists of an Encoder, a Prior, and a Decoder network. During training, we encode the corresponding motion and music inputs into a latent phase manifold, which is variational and parameterized by the frequency domain parameters of periodic functions. The latent phases can be sampled from the manifold and then decoded back to the original data space to obtain new motions. The consistency loss $\mathcal{L}_{\mathrm{csc}}$ is further imposed to constrain the manifold to be consistently unified for dancers that belong to the same group. At inference stage, only the Prior and the Decoder are used to synthesize group dances efficiently. + +properties of their motions, such as timing, periodicity, or temporal alignment are intrinsically similar. Drawing inspiration from [74], we aim to learn a generative phase representation for each group of dancer in order to synthesize their motion indefinitely. Our generative model is built upon the conditional Variational Autoencoder architecture [73], thanks to its diverse generation capability and fast sampling speed. However, instead of directly encoding the data into a Gaussian latent distribution as in common VAE approaches [15, 41, 47, 60, 61], we model the latent variational distribution by the phase parameters extracted from the latent motion curve, which we call variational phase manifold. The latent phase manifold is well-structured and can well describe key characteristics of motion (such as its timing, local periodicity, and transition), which benefits learning motion features [74]. + +The overview of our Phase-conditioned Dance VAE is illustrated in Figure 2. Specifically, the model contains three main networks: an encoder $\mathcal{E}$ to capture the approximate posterior distribution conditioned on both motion and music $q_{\phi}(\mathbf{z}|\mathbf{x},\mathbf{a})$ , a prior network $\mathcal{P}$ to learn the conditional prior given only the music $p_{\theta}(\mathbf{z}|\mathbf{a})$ , and a decoder $\mathcal{D}$ to learn to reconstruct the data from the latent distribution $p_{\theta}(\mathbf{x}|\mathbf{z},\mathbf{a})$ . The new motion is generated by sampling the frequency-domain parameters predicted by the prior network, which is then passed through the decoder network to reconstruct the motion in the original data space. Fur- + +thermore, we adopt Transformer-based architecture [85] in each network to effectively capture long-range dependencies and holistic context of the dance sequence. + +Encoder. The encoder $\mathcal{E}$ is expected to take both the motion and music feature sequence as input, and produce a distribution over possible latent variables capturing the cross-modal relationship between them. To transform the joint input space into a learned phase manifold, we adopt the Transformer decoder architecture where the Cross-Attention mechanism [85] is utilized to learn the relationship between the motion and the music. Accordingly, the output of the encoder is a batch of latent curves (i.e., the activation sequences per channel) that can particularly capture different spatial and temporal aspects of the motion sequence. However, instead of training the model to directly reconstruct the input motion from the extracted latent curves, we further enforce each channel of the latent space to have a periodic functional form (i.e., sinusoidal). This enables us to effectively learn a compact parameterization for each latent channel from a small set of parameters in the frequency domain. + +Generative Variational Phase Manifold. Here we focus on learning the periodicity and non-linear temporal alignment of the motion in the latent space. In particular, given the output latent curves from the encoder $\mathbf{L} = \mathcal{E}(x,a)\in \mathbb{R}^{D\times T}$ with $D$ is the number of desired phase channels to be extracted from the motion, we parameterize each latent curve in $\mathbf{L}$ using a sinusoidal function with amplitude (A), frequency (F), offset (B) and phase shift (S) parameters [74]. To allow for variational phase manifold learning, we opt to predict two sets of parameters $\mu_{\varepsilon} = \{\mu^{A};\mu^{F};\mu^{B};\mu^{S}\}$ and $\sigma_{\varepsilon} = \{\sigma^{A};\sigma^{F};\sigma^{B};\sigma^{S}\}$ , which corresponds to the mean and variance of $\mathbb{R}^{4D}$ dimensional Gaussian distribution: + +$$ +q _ {\phi} (\mathbf {z} | \mathbf {x}, \mathbf {a}) = \mathcal {N} (\mathbf {z}; \mu_ {\varepsilon}, \sigma_ {\varepsilon}) \tag {1} +$$ + +To do so, we first apply differentiable Fast Fourier Transform (FFT) to each channel of the latent curve $\mathbf{L}$ and create the zero-indexed matrix of Fourier coefficients as $\mathbf{c} = FFT(\mathbf{L})$ with $\mathbf{c} \in \mathbb{C}^{D \times K + 1}$ , $K = \left\lfloor \frac{T}{2} \right\rfloor$ . Correspondingly, we compute the per channel power spectrum $\mathbf{p} \in \mathbb{R}^{D \times K + 1}$ as $\mathbf{p}_{i,j} = \frac{2}{N} |\mathbf{c}_{i,j}|^2$ , where $i$ is the channel index and $j$ is the index for the frequency bands. Correspondingly, the distributional mean parameters of the periodic sinusoidal function are then calculated as follows: + +$$ +\mu_ {i} ^ {A} = \sqrt {\frac {2}{T} \sum_ {j = 1} ^ {K} \mathbf {p} _ {i , j}}, \quad \mu_ {i} ^ {F} = \frac {\sum_ {j = 1} ^ {K} \mathbf {f} _ {j} \cdot \mathbf {p} _ {i , j}}{\sum_ {j = 1} ^ {K} \mathbf {p} _ {i , j}}, \quad \mu_ {i} ^ {B} = \frac {\mathbf {c} _ {i , 0}}{T}, \tag {2} +$$ + +where $\mathbf{f} = (0, \frac{1}{T}, \dots, \frac{K}{T})$ is the frequencies vector. At the same time, the phase shift $\mathbf{S}$ is predicted using a fully-connected (FC) layer with two arctan activation as: + +$$ +\left(s _ {y}, s _ {x}\right) = \operatorname {F C} \left(\mathbf {L} _ {i}\right), \quad \mu_ {i} ^ {S} = \arctan \left(s _ {y}, s _ {x}\right), \tag {3} +$$ + +To predict the distributional variance of the phase amplitude and phase shift parameters $\{\sigma^A,\sigma^S\}$ , We additionally apply a separate two-layer MLP network over each channel of the latent curves, similar to Equation 3. The variational latent phase parameters are sampled by utilizing parameterization trick [36], i.e., $\mathbf{A} \sim \mathcal{N}(\mu^A,\sigma^A)$ and $\mathbf{S} \sim \mathcal{N}(\mu^S,\sigma^S)$ . In our experiments, we find that sampling the phase frequency $\mathbf{F}$ and offset $\mathbf{B}$ often produce unstable and noncoherent group movements. This might be because the frequency amplitudes of the dancers within the same group are likely to associate with the rhythmic pattern of the musical beats while the offsets capture their alignment, thereby should be consistent with each other. Therefore, we treat those parameters as deterministic by constraining their variance to zero. + +Finally, the sampled set of phase parameters $\mathbf{z} = \{\mathbf{A};\mathbf{F};\mathbf{B};\mathbf{S}\}$ are used to reconstruct a parametric latent space consisting of multiple periodic curves to represent each intrinsic property of the motion by: + +$$ +\tilde {\mathbf {L}} = \mathbf {A} \cdot \sin (2 \pi \cdot (\mathbf {F} \cdot \mathcal {T} - \mathbf {S})) + \mathbf {B} \tag {4} +$$ + +where $\mathcal{T}$ is a known time window series obtained by evenly spacing the timesteps from 0 to $T$ . Intuitively this curve construction procedure can be viewed as a "quantization" layer to enforce the network to learn to represent the motion features in the frequency domain, which is useful in representing different aspects of human motion such as their timing and periodicity. In the last step, a decoder is utilized to reconstruct the original motion signals from the set of parametric latent curves. + +Decoder. To decode the latent space into the original motion space, previous works [60,61] have to use a sinusoidal positional encoding sequence with duration $T$ as the proxy input to the sequence decoder. This is because their latent space is only formed by single latent vectors following a Gaussian distribution, which cannot span the time dimension. However, we observe that it usually results in unstable and inconsistent movements, as the proxy sequence is generic and usually contains less meaningful information for the decoder. Meanwhile, our method does not suffer from this problem as our latent space is built on multiple curves that can represent the motion information through time, thanks to the phase parameters. Subsequently, our decoder $\mathcal{D}$ is based on Transformer decoder architecture that takes the constructed parametric latent curve, as well as the music features as inputs, to reconstruct the corresponding dance motions. Here, we also utilize the cross-attention model [85] where we consider the sequence of and music features as key and value along with the sampled latent curves as the query. The output of the decoder is a sequence of $T$ vectors in $\mathbb{R}^D$ , which is then projected back to the original motion dimensions through a linear layer, to obtain the reconstructed outputs $\hat{\mathbf{x}} = p_{\theta}(\mathbf{x}|\mathbf{z},\mathbf{a})$ . We additionally employ a global trajectory predictor to predict the global translation of the root joint based on the generated local motions [19, 46], in order to avoid intersection problems between dancers. We provide details of our global predictor in the supplemental. + +Prior Network. Since the ground-truth motion is generally inaccessible at test time (i.e., we only have access to the music), we also need to learn a prior $\mathcal{P}$ to match the posterior distribution of motion from which the latent phase can be sampled. Specifically, We follow the procedure similar to Section 3.2 (Equation 1-4) to predict the Gaussian distribution conditioned on the music sequence $\mathbf{a}$ , which is then used for sampling the latent phases: + +$$ +p _ {\theta} (\mathbf {z} | \mathbf {a}) = \mathcal {N} (\mathbf {z}; \mu_ {\mathcal {P}}, \sigma_ {\mathcal {P}}) \tag {5} +$$ + +where a Transformer encoder is used to encode the input conditioning music sequence and predict the corresponding $\mu_{\mathcal{P}}$ and $\sigma_{\mathcal{P}}$ . We implement the prior network similarly to the encoder network, however, we use self-attention mechanism [85] to capture the global music context. Learning the conditional prior is crucial for the conditional VAE to generalize to diverse types of music and motion. Intuitively speaking, each latent variable $\mathbf{z}$ is expected to represent possible dance motions $\mathbf{x}$ conforming to the music context $\mathbf{a}$ . Therefore, the prior should be able to encode different latent distributions given different musics. + +# 3.3 Training + +During training, we consider the following variational lower bound [73] to mainly train our dance generation VAE model: + +$$ +\log p _ {\theta} (\mathbf {x} \mid \mathbf {a}) \geq \mathbb {E} _ {q _ {\phi}} \left[ \log p _ {\theta} (\mathbf {x} \mid \mathbf {z}, \mathbf {a}) \right] - D _ {\mathrm {K L}} \left(q _ {\phi} (\mathbf {z} \mid \mathbf {x}, \mathbf {a}) \| p _ {\theta} (\mathbf {z} \mid \mathbf {a})\right) \tag {6} +$$ + +In practice, we apply the conditional VAE loss as similar to [61], which is defined as the weighted sum $\mathcal{L}_{\mathrm{cvae}} = \mathcal{L}_{\mathrm{rec}} + \lambda_{\mathrm{KL}}\mathcal{L}_{\mathrm{KL}}$ . In particular, the reconstruction term $\mathcal{L}_{\mathrm{rec}}$ measures the motion reconstruction error given the decoder output (via a smooth-L1 loss). The KL divergence term $\mathcal{L}_{\mathrm{KL}}$ compares the divergence $D_{\mathrm{KL}}$ between the posterior and the prior distribution to enforce them to be close to each other. + +The conditional VAE objective above is calculated for each dancer separately and cannot capture the correlation between dancers within a group. Therefore, it is essential to impose consistency among dancers and avoid strange effects such as unsynchronized dance. To this end, we propose a group consistency loss by enforcing the latent phase manifold to be similar for the same group, given the input music. Specifically, we first calculate the phase manifold features based on the frequency domain parameters as follows: + +$$ +\mathbf {P} _ {2 i - 1} = \mathbf {A} _ {i} \sin (2 \pi \cdot \mathbf {S} _ {i}), \quad \mathbf {P} _ {2 i} = \mathbf {A} _ {i} \cos (2 \pi \cdot \mathbf {S} _ {i}) \tag {7} +$$ + +where $\mathbf{P} \in \mathbb{R}^{2D}$ is the phase manifold vector that encodes the spatial-temporal information of the motion state. The phase feature may look similar to the positional encodings of transformers [85] in the sense that both use multi-resolution sinusoidal functions. However, the phase feature is much richer in terms of representation capacity since it learns to embed the spatial (body joints) and temporal (positions in time) information of the motion curves, whereas the positional encodings only encode the position of words. Finally, our consistency objective is + +to constrain phase manifold between dancers within a group to be as close as possible to each other, which is formulated as: + +$$ +\mathcal {L} _ {\mathrm {c s c}} = D _ {\mathrm {K L}} \left(q _ {\phi} (\mathbf {z} | \mathbf {x} ^ {m}, \mathbf {a}) \| \left(q _ {\phi} (\mathbf {z} | \mathbf {x} ^ {n}, \mathbf {a})\right) + \| \mathbf {P} ^ {m} - \mathbf {P} ^ {n} \| _ {2} ^ {2} \right. \tag {8} +$$ + +where the first term encourages the network to map different dancers belonging to the same group $(\mathbf{x}^m$ and $\mathbf{x}^n)$ into the same set of distributional phase parameters while the second term penalizes the discrepancy in their corresponding phase manifolds. In general, this loss is applied to ensure every dancer is embedded into a single unified manifold that can effectively represent their corresponding group. To summarize, our total training loss is defined as the combination of the VAE loss and the consistency loss $\mathcal{L} = \mathcal{L}_{\mathrm{cvae}} + \lambda_{\mathrm{csc}}\mathcal{L}_{\mathrm{csc}}$ . + +For testing, we can efficiently generate motions for an unlimited number of dancers by indefinitely drawing samples from the learned continuous group-consistent phase manifold. It is noteworthy that for inference, we only need to process the prior network once to obtain the latent distribution. To generate a new motion, we can sample from this latent (Gaussian) distribution and use the decoder to decode it back to the motion space. This approach is much more efficient and has significantly higher scalability than previous approaches [39, 40] that is limited by the number of dancers processed simultaneously by the entire network. + +# 4 Experiments + +# 4.1 Implementation Details + +Our model was trained on 4 NVIDIA V100 GPUs using Adam optimizer with a fixed learning rate of $10^{-4}$ and a mini-batch size of 32 per GPU. For training losses, the weights are empirically set to $\lambda_{\mathrm{KL}} = 5\times 10^{-4}$ and $\lambda_{\mathrm{csc}} = 10^{-4}$ , respectively. The Transformer encoders and decoders consist of 5 layers for both encoder, decoder, and prior Network with 512-dimensional hidden units. Meanwhile, the number of latent phase channels is set to 256. To further capture the periodic nature of the phase feature, we also use Siren activation following the initialization scheme as in [70]. This can effectively model the periodicity inherent in the motion data, and thus can benefit motion synthesis [68]. + +# 4.2 Experimental Settings + +Dataset In our experiments, we utilize the AIOZ-GDance [40] and AIST-M [90] datasets. AIOZ-GDance is the largest music-driven dataset focusing on group dance, encompassing paired music and 3D group motions extracted from in-the-wild videos through a semi-automatic process. This dataset spans 7 dance styles and 16 music genres. For consistency, we adhere to the training and testing split outlined in [40] during our experiments. + +Evaluation Protocol We employ several metrics to assess the quality of individual dance motions, including Frechet Inception Distance (FID) [20, 49], Motion-Music Consistency (MMC) [49], and Generation Diversity (GenDiv) [26, 41, 49], along with the Physical Foot Contact score (PFC) [83]. Specifically, the FID score gauges the realism of individual dance movements concerning the ground-truth dance. MMC assesses the matching similarity between motion and music beats, reflecting how well-generated dances synchronize with the music's rhythm. GenDiv is computed as the average pairwise distance of kinetic features among motions [58]. PFC evaluates the physical plausibility of foot movements by determining the agreement between the acceleration of the character's center of mass and the foot's velocity. + +In assessing the group dance quality, we adopt three metrics outlined in [40]: Group Motion Realism (GMR), Group Motion Correlation (GMC), and Trajectory Intersection Frequency (TIF). Broadly, GMR gauges the realism of generated group motions in comparison to ground-truth data, employing Frechet Inception Distance on extracted group motion features. GMC evaluates the synchronization among dancers within the generated group by computing their cross-correlation. TIF quantifies the frequency of collisions among the generated dancers during their dance movements. + +Baselines Our method is subjected to comparison with various recent techniques in music-driven dance generation, namely FACT [49], Transflower [59], and EDGE [83]. These approaches are adapted for benchmarking within group dance generation context, as outlined in [40], considering that their original designs were tailored for single-dance scenarios. Additionally, our evaluation includes comparisons with GDanceR [40], GCD [39], and DanY [90]. All mentioned works are specifically designed for the group choreography generation. + +
DatasetMethodFID↓MMC↑GenDiv↑PFC↓GMR↓GMC↑TIF↓
AIOZ-GDANCE [40]FACT [49]56.200.2228.643.52101.5262.680.321
Transflower [59]37.730.2178.743.0781.1760.780.332
EDGE [83]31.400.2649.572.6363.3561.720.356
GDANCER [40]43.900.2509.233.0551.2779.010.217
GCD [39]31.160.26110.872.5331.4780.970.167
Ours31.010.27110.982.3330.0884.520.163
AIST-M [90]GDANCER [40]52.900.2226.521.9365.1360.560.121
GCD [39]35.360.24510.971.5242.5272.150.083
DanY [90]40.250.24011.401.6550.2963.530.137
Ours31.490.25711.811.4241.2478.640.076
+ +Table 1: Performance comparison. + +# 4.3 Experimental Results + +Quality Comparison Table 1 presents a comparison among the baselines FACT [49], Transflower [59], EDGE [83], GDanceR [40], GCD [39], and our proposed manifold-based method. The results clearly demonstrate that our model almost outperforms all baselines across all evaluations on two datasets AIOZ-GDANCE [40] and AIST-M [90]. We observe that recent diffusion-based dance + +generation models such as EDGE or GCD yield competitive performance on both single-dance metrics (FID, MMC, GenDiv, and PFC) and group dance metrics (GMR, GMC, and TIF). However, due to limitations in their training procedures, they still struggle when dealing with generating multiple dancing motions in the context of a high quantity of dancers, as shown by their low performance compared to our proposal. This result implies that our method successfully preserves the performance of dancing motions when the number of dancers is increased. Besides, Figure 3 also shows that our proposed methods outperform other state-of-the-art methods such as GDanceR or GCD in dealing with monotonous, repetitive, sinking, and overlapping dance motions. + +![](images/2d2603f27afe8456ee8f4803ace1a428fb7cb155c097f92c859d462ebae5b4fa.jpg) +Fig. 3: Visualization of a dancing sample between different methods. GDanceR displays monotonous, repetitive, or sinking dance motions. GCD exhibits more divergence in dance motions, yet dancers may intersect since their optimization does not address this issue explicitly. Blue boxes mark these issues. In contrast, our manifold-based solution ensures the divergence of dancing motions, while the phase motion path demonstrates its effectiveness in addressing floating and crossing issues in group dances. + +
#Generated DancersMethodFID↓MMC↑GenDiv↑GMR↓GMC↑TIF↓
5GDanceR [40]44.190.2498.9955.0578.720.218
GCD [39]35.080.2649.9238.4381.440.168
Ours31.350.26810.0532.5884.560.161
10GDanceR [40]57.930.2257.6373.3272.790.386
GCD [39]N/A (Memory)N/A (Memory)N/A (Memory)N/A (Memory)N/A (Memory)N/A (Memory)
Ours32.190.2698.9934.3286.960.193
100GDanceR [40]N/A (Memory)N/A (Memory)N/A (Memory)N/A (Memory)N/A (Memory)N/A (Memory)
GCD [39]N/A (Memory)N/A (Memory)N/A (Memory)N/A (Memory)N/A (Memory)N/A (Memory)
Ours30.970.2688.7638.1385.730.222
+ +Table 2: Performance of group dance generation methods when we increase the number of generated dancers. The experiments are done with common consumer GPUs with 4GB memory. (N/A means models could not run due to inadequate memory footprint) + +Scalable Generation Analysis Table 2 illustrates the performance of different group dance generation methods (GCD, GDancer, and Ours) when generating dance movements with an increasing number of dancers. When the number of + +dancers is increased to 10, GCD appears to run out of memory, which is also observed in GDanceR when the number of dancers increases to 100. Regardless of the number of dancers, our method consistently achieves stable and competitive results. This implies that our proposed method successfully addresses the scalability issue in group dance generation without compromising the overall performance of each individual dance motion. Figure 4 illustrates the memory consumption to generate dance motions in groups for each method. Noticeably, our proposal still achieves the highest performance while consuming immensely fewer resources required for generating group dance motions (See Figure 5 for illustrations). This, again, indicates that our method successfully breaks the barrier of limited generated dancers by using the manifold. + +Ablation Analysis Table 3 presents the performance improvements achieved through the integration of consistency loss and phase manifold. Additionally, we showcase the effectiveness of our proposed approach across three different backbones: Transformer [49, 85], LSTM [21], and CNN [99]. Evaluation metrics including FID, GMR, and GMC are utilized. The results indicate that the absence of consistency loss leads to an increase in GMR and a decrease in GMC, suggesting a significant enhancement in the realism and correlation of group dance motions facilitated by the inclusion of the proposed objective. Meanwhile, with out the phase manifold, the model exhibits remarkably higher scores in both the FID and GMR metrics, suggesting that phase manifold can effectively improve the distinction in dance motions while maintaining the realism of group dances, even when the number of dancers in a group is large. In comparing three backbones—Transformer, LSTM, and CNN—we have observed that the chosen Transformer backbone achieved the best results compared to LSTM or CNN. + +
MethodFID↓GMR↓GMC↑
Ours31.0130.0884.52
Ours w/o Consistency Loss35.3557.6366.72
Ours w/o Phase Manifold41.7845.3277.93
Ours w LSTM41.2947.4771.82
Ours w CNN36.9944.9475.77
+ +Table 3: Module contribution and loss analysis. + +![](images/e83930fcbac387bcdfd345f08dec3d2db2014f9f814017447ff288ea36b23994.jpg) +Fig. 4: Memory usage vs. number of dancers in different dance generators. + +# 4.4 User Study + +User studies are vital for evaluating generative models, as user perception is pivotal for downstream applications; thus, we conducted two studies with around 70 participants each, diverse in background, with experience in music and dance, aged between 20 to 40, consisting of approximately $47\%$ females and $53\%$ males, to assess the effectiveness of our approach in group choreography generation. + +In the user study, we aim to assess the realism of sample outputs with more and more dancers. Specifically, participants assign scores ranging from 0 to 10 to evaluate the realism of each dance clip with 2 to 10 dancers. Figure 6 shows that, across all methods, the more the number of dancers is increased, the lower the realism is found. However, the drop in realism of our proposed method is the least compared to GCD and GDanceR. The results indicate our method's good performance compared to other baselines when the number of dancers increases. + +![](images/7e085c9cc0289882b96eb06347ae0cab883ac551831d5e51d18109e8dc40c698.jpg) +Fig. 5: Visualization of scalable dancers. + +![](images/b0e89c61c98ff2756242f7040614fc7693068996fb80a37866ced9d1e9e31c29.jpg) +Fig. 6: Realism between different methods when number of dancers is varied. + +# 5 Discussion + +While our approach leverages the VAE as a primary solution for generating a manifold, it is important to acknowledge certain inherent limitations associated with this choice. One notable challenge is the susceptibility to issues such as posterior collapse and unstable sampling within the VAE framework. These challenges can result in generated group dance motions that may not consistently meet performance expectations. + +One specific manifestation of this limitation is the potential for false decoding when sampling points that lie too far from the learned distribution. This scenario can lead to unexpected rotations or disruptions in the physics of the generated content. The impact of this problem becomes evident in instances where the generated samples deviate significantly from the anticipated distribution, introducing inaccuracies and distortions. + +To address these challenges, we recognize the need for ongoing efforts to mitigate the effects of posterior collapse and unstable sampling. While the problem is acknowledged, our approach incorporates measures to limit its impact. Future research directions could explore alternative generative models or additional techniques to enhance the robustness and reliability of the generated results in the face of these identified limitations. + +# References + +1. Ahn, H., Kim, J., Kim, K., Oh, S.: Generative autoregressive networks for 3d dancing move synthesis from music. IEEE Robotics and Automation Letters (RA-L) (2020) +2. Alaoui, S.F., Henry, C., Jacquemin, C.: Physical modelling for interactive installations and the performing arts. International Journal of Performance Arts and Digital Media (2014) +3. Alemi, O., Françoise, J., Pasquier, P.: Groovenet: Real-time music-driven dance movement generation using artificial neural networks. Networks (2017) +4. Alexanderson, S., Nagy, R., Beskow, J., Henter, G.E.: Listen, denoise, action! audio-driven motion synthesis with diffusion models. ACM Transactions on Graphics (TOG) (2023) +5. Au, H.Y., Chen, J., Jiang, J., Guo, Y.: Choreograph: Music-conditioned automatic dance choreography over a style and tempo consistent dynamic graph. In: ACM International Conference on Multimedia (2022) +6. Bisig, D.: Generative dance-a taxonomy and survey. In: International Conference on Movement and Computing (2022) +7. Cervantes, P., Sekikawa, Y., Sato, I., Shinoda, K.: Implicit neural representations for variable length human motion generation. In: ECCV (2022) +8. Chan, C., Ginosar, S., Zhou, T., Efros, A.A.: Everybody dance now. In: ICCV (2019) +9. Chen, K., Tan, Z., Lei, J., Zhang, S.H., Guo, Y.C., Zhang, W., Hu, S.M.: Choreomaster: choreography-oriented music-driven dance synthesis. ACM Transactions on Graphics (TOG) (2021) +0. Fan, R., Xu, S., Geng, W.: Example-based automatic music-driven conventional dance motion synthesis. IEEE Transactions on Visualization and Computer Graphics (TVCG) (2011) +1. Feng, B., Ao, T., Liu, Z., Ju, W., Liu, L., Zhang, M.: Robust dancer: Long-term 3d dance synthesis using unpaired data. arXiv (2023) +2. Ferreira, J.P., Coutinho, T.M., Gomes, T.L., Neto, J.F., Azevedo, R., Martins, R., Nascimento, E.R.: Learning to dance: A graph convolutional adversarial network to generate realistic dance motions from audio. Computers & Graphics (2021) +3. Fink, B., Bläsing, B., Ravignani, A., Shackelford, T.K.: Evolution and functions of human dance. Evolution and Human Behavior (2021) +4. Gartner, E., Metz, L., Andriluka, M., Freeman, C.D., Sminchisescu, C.: Transformer-based learned optimization. In: CVPR (2023) +5. Ghorbani, S., Ferstl, Y., Holden, D., Troje, N.F., Carbonneau, M.A.: Zeroeggs: Zero-shot example-based gesture generation from speech. In: Computer Graphics Forum (2023) +6. Ghosh, A., Cheema, N., Oguz, C., Theobalt, C., Slusallek, P.: Synthesis of compositional animations from textual descriptions. In: ICCV (2021) +7. Gong, K., Lian, D., Chang, H., Guo, C., Zuo, X., Jiang, Z., Wang, X.: Tm2d: Bimodality driven 3d dance generation via music-text integration. In: ICCV (2023) +8. Guo, C., Zou, S., Zuo, X., Wang, S., Ji, W., Li, X., Cheng, L.: Generating diverse and natural 3d human motions from text. In: CVPR (2022) +9. He, C., Saito, J., Zachary, J., Rushmeier, H., Zhou, Y.: Nemf: Neural motion fields for kinematic animation. NeurIPS 35, 4244-4256 (2022) +20. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: NeurIPS (2017) + +21. Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural Computation (1997) +22. Holden, D., Komura, T., Saito, J.: Phase-functional neural networks for character control. ACM Transactions on Graphics (TOG) 36(4), 1-13 (2017) +23. Holden, D., Saito, J., Komura, T.: A deep learning framework for character motion synthesis and editing. ACM Transactions on Graphics (TOG) (2016) +24. Holden, D., Saito, J., Komura, T., Joyce, T.: Learning motion manifolds with convolutional autoencoders. In: SIGGRAPH Asia 2015 technical briefs (2015) +25. Hong, F., Zhang, M., Pan, L., Cai, Z., Yang, L., Liu, Z.: Avatarclip: zero-shot text-driven generation and animation of 3d avatars. ACM Transactions on Graphics (TOG) (2022) +26. Huang, R., Hu, H., Wu, W., Sawada, K., Zhang, M.: Dance revolution: Long sequence dance generation with music via curriculum learning. CoRR (2020) +27. Huang, Y., Zhang, J., Liu, S., Bao, Q., Zeng, D., Chen, Z., Liu, W.: Genre-conditioned long-term 3d dance generation driven by music. In: ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 4858-4862. IEEE (2022) +28. Jiang, C., Cornman, A., Park, C., Sapp, B., Zhou, Y., Anguelov, D., et al.: Motiondiffuser: Controllable multi-agent motion prediction using diffusion. In: CVPR (2023) +29. Jiang, Y., Won, J., Ye, Y., Liu, C.K.: Drop: Dynamics responses from human motion prior and projective dynamics. In: SIGGRAPH Asia 2023 Conference papers (2023) +30. Joshi, M., Chakrabarty, S.: An extensive review of computational dance automation techniques and applications. Proceedings of the Royal Society A (2021) +31. Karunratanakul, K., Preechakul, K., Suwajanakorn, S., Tang, S.: Guided motion diffusion for controllable human motion synthesis. In: ICCV (2023) +32. Kico, I., Grammalidis, N., Christidis, Y., Liarokapis, F.: Digitization and visualization of folk dances in cultural heritage: A review. Inventions (2018) +33. Kim, J.W., Fouad, H., Hahn, J.K.: Making them dance. In: AAAI Fall Symposium: Aurally Informed Performance (2006) +34. Kim, J., Oh, H., Kim, S., Tong, H., Lee, S.: A brand new dance partner: Music-conditioned pluralistic dancing controlled by multiple dance genres. In: CVPR (2022) +35. Kim, T.h., Park, S.I., Shin, S.Y.: Rhythmic-motion synthesis based on motion-beat analysis. ACM Transactions on Graphics (TOG) (2003) +36. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. In: ICLR (2014) +37. Kolotouros, N., Alldieck, T., Zanfir, A., Bazavan, E., Fieraru, M., Sminchisescu, C.: Dreamhuman: Animatable 3d avatars from text. In: NeurIPS (2024) +38. Lannan, N., Zhou, L., Fan, G.: Human motion enhancement via tobit kalman filter-assisted autoencoder. IEEE Access 10, 29233-29251 (2022) +39. Le, N., Do, T., Do, K., Nguyen, H., Tjiputra, E., Tran, Q.D., Nguyen, A.: Controllable group choreography using contrastive diffusion. ACM Transactions on Graphics (TOG) (2023) +40. Le, N., Pham, T., Do, T., Tjiputra, E., Tran, Q.D., Nguyen, A.: Music-driven group choreography. In: CVPR (2023) +41. Lee, H.Y., Yang, X., Liu, M.Y., Wang, T.C., Lu, Y.D., Yang, M.H., Kautz, J.: Dancing to music. In: NeurIPS (2019) +42. Lee, L., Lin, Z., Hu, R., Gong, Z., Kumar, A., Li, T., Li, S., Hui, P.: When creators meet the metaverse: A survey on computational arts. CoRR (2021) + +43. Lee, M., Lee, K., Park, J.: Music similarity-based approach to generating dance motion sequence. Multimedia tools and applications (2013) +44. Li, B., Zhao, Y., Sheng, L.: Dancenet3d: Music based dance generation with parametric motion transformer. In: AAAI (2022) +45. Li, B., Zhao, Y., Zhelun, S., Sheng, L.: Danceformer: Music conditioned 3d dance generation with parametric motion transformer. In: Proceedings of the AAAI Conference on Artificial Intelligence (2022) +46. Li, J., Villegas, R., Ceylan, D., Yang, J., Kuang, Z., Li, H., Zhao, Y.: Task-generic hierarchical human motion prior using vaes. In: International Conference on 3D Vision (3DV) (2021) +47. Li, J., Kang, D., Pei, W., Zhe, X., Zhang, Y., He, Z., Bao, L.: Audio2gestures: Generating diverse gestures from speech audio with conditional variational autoencoders. In: ICCV (2021) +48. Li, R., Zhao, J., Zhang, Y., Su, M., Ren, Z., Zhang, H., Tang, Y., Li, X.: Finedance: A fine-grained choreography dataset for 3d full body dance generation. In: ICCV (2023) +49. Li, R., Yang, S., Ross, D.A., Kanazawa, A.: Ai choreographer: Music conditioned 3d dance generation with aist++. In: ICCV (2021) +50. Li, W., Dai, B., Zhou, Z., Yao, Q., Wang, B.: Controlling character motions without observable driving source. In: WACV (2024) +51. Liang, H., Zhang, W., Li, W., Yu, J., Xu, L.: Intergen: Diffusion-based multi-human motion generation under complex interactions. arXiv preprint arXiv:2304.05684 (2023) +52. Lin, J., Zeng, A., Lu, S., Cai, Y., Zhang, R., Wang, H., Zhang, L.: Motion-x: A large-scale 3d expressive whole-body human motion dataset. NeurIPS (2024) +53. Ling, H.Y., Zinno, F., Cheng, G., Van De Panne, M.: Character controllers using motion vaes. ACM Transactions on Graphics (TOG) 39(4), 40-1 (2020) +54. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: SMPL: A skinned multi-person linear model. ACM Transactions on Graphics (TOG) (2015) +55. Martinez, J., Black, M.J., Romero, J.: On human motion prediction using recurrent neural networks. In: CVPR (2017) +56. Mo, C.A., Hu, K., Long, C., Wang, Z.: Continuous intermediate token learning with implicit motion manifold for keyframe based motion interpolation. In: CVPR (2023) +57. Ofli, F., Erzin, E., Yemez, Y., Tekalp, A.M.: Learn2dance: Learning statistical music-to-dance mappings for choreography synthesis. IEEE Transactions on Multimedia (TMM) (2011) +58. Onuma, K., Faloutsos, C., Hodgins, J.K.: Fmdistance: A fast and effective distance function for motion capture data. In: Eurographics (2008) +59. Perez, G.V., Beskow, J., Henter, G., Holzapfel, A., Oudeyer, P.Y., Alexanderson, S.: Transflower: probabilistic autoregressive dance generation with multimodal attention. ACM Transactions on Graphics (TOG) (2021) +60. Petrovich, M., Black, M.J., Varol, G.: Action-conditioned 3d human motion synthesis with transformer vae. In: ICCV (2021) +61. Petrovich, M., Black, M.J., Varol, G.: Temos: Generating diverse human motions from textual descriptions. In: ECCV (2022) +62. Qi, Q., Zhuo, L., Zhang, A., Liao, Y., Fang, F., Liu, S., Yan, S.: Diffdance: Cascaded human motion diffusion model for dance generation. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 1374-1382 (2023) +63. Raab, S., Leibovitch, I., Li, P., Aberman, K., Sorkine-Hornung, O., Cohen-Or, D.: Modi: Unconditional motion synthesis from diverse data. In: CVPR (2023) + +64. Ren, X., Li, H., Huang, Z., Chen, Q.: Self-supervised dance video synthesis conditioned on music. In: ACMMM (2020) +65. Safonova, A., Hodgins, J.K.: Construction and optimal search of interpolated motion graphs. In: SIGGRAPH (2007) +66. Shafir, Y., Tevet, G., Kapon, R., Bermano, A.H.: Human motion diffusion as a generative prior. arXiv (2023) +67. Shi, J.: Application of 3d computer aided system in dance creation and learning. In: International Conference on Machine Learning and Big Data Analytics for IoT Security and Privacy (2021) +68. Shi, M., Starke, S., Ye, Y., Komura, T., Won, J.: Phasemp: Robust 3d pose estimation via phase-conditioned human motion prior. In: ICCV (2023) +69. Shiratori, T., Nakazawa, A., Ikeuchi, K.: Dancing-to-music character animation. Computer Graphics Forum (2006) +70. Sitzmann, V., Martel, J., Bergman, A., Lindell, D., Wetzstein, G.: Implicit neural representations with periodic activation functions. In: NeurIPS (2020) +71. Siyao, L., Yu, W., Gu, T., Lin, C., Wang, Q., Qian, C., Loy, C.C., Liu, Z.: Bailando: 3d dance generation by actor-critic gpt with choreographic memory. In: CVPR (2022) +72. Soga, A., Umino, B., Longstaff, J.S.: Automatic composition of ballet sequences using a 3d motion archive. In: 1st South-Eastern European Digitization Initiative Conference (2005) +73. Sohn, K., Lee, H., Yan, X.: Learning structured output representation using deep conditional generative models. In: NeurIPS (2015) +74. Starke, S., Mason, I., Komura, T.: Deepphase: Periodic autoencoders for learning motion phase manifolds. ACM Transactions on Graphics (TOG) (2022) +75. Starke, S., Zhang, H., Komura, T., Saito, J.: Neural state machine for character-scene interactions. ACM Transactions on Graphics (TOG) 38(6), 209-1 (2019) +76. Starke, S., Zhao, Y., Komura, T., Zaman, K.: Local motion phases for learning multi-contact character movements. ACM Transactions on Graphics (TOG) 39(4), 54-1 (2020) +77. Sun, G., Wong, Y., Cheng, Z., Kankanhalli, M.S., Geng, W., Li, X.: Deepdance: Music-to-dance motion choreography with adversarial learning. IEEE Transactions on Multimedia (TMM) (2020) +78. Sun, J., Wang, C., Hu, H., Lai, H., Jin, Z., Hu, J.F.: You never stop dancing: Nonfreezing dance generation via bank-constrained manifold projection. In: NeurIPS (2022) +79. Tang, T., Jia, J., Mao, H.: Dance with melody: An LSTM-autoencoder approach to music-oriented dance synthesis. In: ACMMM (2018) +80. Tevet, G., Gordon, B., Hertz, A., Bermano, A.H., Cohen-Or, D.: Motionclip: Exposing human motion generation to clip space. In: European Conference on Computer Vision (2022) +81. Tevet, G., Raab, S., Gordon, B., Shafir, Y., Cohen-Or, D., Bermano, A.H.: Human motion diffusion model. In: ICLR (2023) +82. Tiwari, G., Antic, D., Lenssen, J.E., Sarafianos, N., Tung, T., Pons-Moll, G.: Posendf: Modeling human pose manifolds with neural distance fields. In: ECCV (2022) +83. Tseng, J., Castellon, R., Liu, C.K.: Edge: Editable dance generation from music. In: CVPR (2023) +84. Tsuchida, S., Fukayama, S., Hamasaki, M., Goto, M.: Aist dance video database: Multi-genre, multi-dancer, and multi-camera database for dance information processing. In: ISMIR (2019) + +85. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017) +86. Wang, Z., Jia, J., Wu, H., Xing, J., Cai, J., Meng, F., Chen, G., Wang, Y.: Group-dancer: Music to multi-people dance synthesis with style collaboration. In: ACM International Conference on Multimedia (2022) +87. Yalta, N., Watanabe, S., Nakadai, K., Ogata, T.: Weakly-supervised deep recurrent neural networks for basic dance step generation. In: IJCNN (2019) +88. Yang, S., Wu, Z., Li, M., Zhang, Z., Hao, L., Bao, W., Zhuang, H.: Qpgesture: Quantization-based and phase-guided motion matching for natural speech-driven gesture generation. In: CVPR (2023) +89. Yang, Z., Wen, Y.H., Chen, S.Y., Liu, X., Gao, Y., Liu, Y.J., Gao, L., Fu, H.: Keyframe control of music-driven 3d dance generation. IEEE Transactions on Visualization and Computer Graphics (2023) +90. Yao, S., Sun, M., Li, B., Yang, F., Wang, J., Zhang, R.: Dance with you: The diversity controllable dancer generation via diffusion models. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 8504-8514 (2023) +91. Ye, Z., Wu, H., Jia, J., Bu, Y., Chen, W., Meng, F., Wang, Y.: Choreonet: Towards music to dance synthesis with choreographic action unit. In: ACMMM (2020) +92. Yin, W., Yin, H., Baraka, K., Kragic, D., Björkman, M.: Dance style transfer with cross-modal transformer. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 5058-5067 (2023) +93. Yu, Z., Cheng, W., Liu, X., Wu, W., Lin, K.Y.: Monohuman: Animatable human neural field from monocular video. In: CVPR (2023) +94. Zhang, X., Yang, S., Xu, Y., Zhang, W., Gao, L.: Mining and applying composition knowledge of dance moves for style-concentrated dance generation. In: AAAI (2023) +95. Zhou, L., Luo, Y.: A spatio-temporal learning for music conditioned dance generation. In: International Conference on Multimodal Interaction (2022) +96. Zhou, Y., Barnes, C., Lu, J., Yang, J., Li, H.: On the continuity of rotation representations in neural networks. In: CVPR (2019) +97. Zhou, Z., Wang, B.: Ude: A unified driving engine for human motion generation. In: CVPR (2023) +98. Zhu, Y., Olszewski, K., Wu, Y., Achlioptas, P., Chai, M., Yan, Y., Tulyakov, S.: Quantized gan for complex music generation from dance videos. In: ECCV (2022) +99. Zhuang, W., Wang, C., Chai, J., Wang, Y., Shao, M., Xia, S.: Music2dance: Dancenet for music-driven dance generation. ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM) (2022) \ No newline at end of file diff --git a/2024/Scalable Group Choreography via Variational Phase Manifold Learning/images.zip b/2024/Scalable Group Choreography via Variational Phase Manifold Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..45012f9360ffa2f80f638a2a158f6f446ddcaf6f --- /dev/null +++ b/2024/Scalable Group Choreography via Variational Phase Manifold Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6583279a30f3a627aba675396cb126befd10a893709acf6b4b68e34f2cb7c80c +size 371246 diff --git a/2024/Scalable Group Choreography via Variational Phase Manifold Learning/layout.json b/2024/Scalable Group Choreography via Variational Phase Manifold Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7cda031911614b3675a50f7c82614e87a3c265c4 --- /dev/null +++ b/2024/Scalable Group Choreography via Variational Phase Manifold Learning/layout.json @@ -0,0 +1,9752 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 148, + 111, + 465, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 111, + 465, + 148 + ], + "spans": [ + { + "bbox": [ + 148, + 111, + 465, + 148 + ], + "type": "text", + "content": "Scalable Group Choreography via Variational Phase Manifold Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 138, + 167, + 474, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 167, + 474, + 194 + ], + "spans": [ + { + "bbox": [ + 138, + 167, + 474, + 194 + ], + "type": "text", + "content": "Nhat " + }, + { + "bbox": [ + 138, + 167, + 474, + 194 + ], + "type": "inline_equation", + "content": "\\mathrm{Le}^{1}" + }, + { + "bbox": [ + 138, + 167, + 474, + 194 + ], + "type": "text", + "content": ", Khoa " + }, + { + "bbox": [ + 138, + 167, + 474, + 194 + ], + "type": "inline_equation", + "content": "\\mathrm{Do}^{2,3}" + }, + { + "bbox": [ + 138, + 167, + 474, + 194 + ], + "type": "text", + "content": ", Xuan Bui" + }, + { + "bbox": [ + 138, + 167, + 474, + 194 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 138, + 167, + 474, + 194 + ], + "type": "text", + "content": ", Tuong " + }, + { + "bbox": [ + 138, + 167, + 474, + 194 + ], + "type": "inline_equation", + "content": "\\mathrm{Do}^{1,4}" + }, + { + "bbox": [ + 138, + 167, + 474, + 194 + ], + "type": "text", + "content": ", Erman Tjiputra" + }, + { + "bbox": [ + 138, + 167, + 474, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 167, + 474, + 194 + ], + "type": "text", + "content": ", Quang D. Tran" + }, + { + "bbox": [ + 138, + 167, + 474, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 167, + 474, + 194 + ], + "type": "text", + "content": ", and Anh Nguyen" + }, + { + "bbox": [ + 138, + 167, + 474, + 194 + ], + "type": "inline_equation", + "content": "^{4}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 267, + 201, + 346, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 267, + 201, + 346, + 213 + ], + "spans": [ + { + "bbox": [ + 267, + 201, + 346, + 213 + ], + "type": "text", + "content": "1 AIOZ, Singapore" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 178, + 213, + 435, + 246 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 200, + 213, + 413, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 213, + 413, + 224 + ], + "spans": [ + { + "bbox": [ + 200, + 213, + 413, + 224 + ], + "type": "text", + "content": "2 University of Science, Ho Chi Minh City, Vietnam" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 184, + 224, + 429, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 224, + 429, + 235 + ], + "spans": [ + { + "bbox": [ + 184, + 224, + 429, + 235 + ], + "type": "text", + "content": "3 Vietnam National University, Ho Chi Minh City, Vietnam" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 178, + 235, + 435, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 235, + 435, + 246 + ], + "spans": [ + { + "bbox": [ + 178, + 235, + 435, + 246 + ], + "type": "text", + "content": "4 University of Liverpool, Liverpool L69 3BX, United Kingdom" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 159, + 276, + 452, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 276, + 452, + 440 + ], + "spans": [ + { + "bbox": [ + 159, + 276, + 452, + 440 + ], + "type": "text", + "content": "Abstract. Generating group dance motion from the music is a challenging task with several industrial applications. Although several methods have been proposed to tackle this problem, most of them prioritize optimizing the fidelity in dancing movement, constrained by predetermined dancer counts in datasets. This limitation impedes adaptability to real-world applications. Our study addresses the scalability problem in group choreography while preserving naturalness and synchronization. In particular, we propose a phase-based variational generative model for group dance generation on learning a generative manifold. Our method achieves high-fidelity group dance motion and enables the generation with an unlimited number of dancers while consuming only a minimal and constant amount of memory. The intensive experiments on two public datasets show that our proposed method outperforms recent state-of-the-art approaches by a large margin and is scalable to a great number of dancers beyond the training data." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 460, + 229, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 460, + 229, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 229, + 472 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 486, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 616 + ], + "type": "text", + "content": "The widespread availability of digital social media platforms has led to a significant increase in the popularity of creating dance videos. This heightened interest has resulted in the daily production and viewing of millions of dance videos on various online platforms [13, 32]. Recently, researchers from the computer vision and computer graphics community have focused on developing methods for generating authentic dance movements in response to music [6, 94]. These advancements have broad implications and find applications in diverse areas, including animation [49, 50], virtual idols [59], virtual metaverses [42], and dance education [2, 67, 72]. Such techniques empower artists, animators, and educators by providing powerful tools to enhance their creative pursuits and improve the overall dance experience for performers and audiences alike." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Recently, significant strides have been made in creating dance motions for solo dance performers [12,26,34,49,59,71,83]. However, generating synchronized group dance movements that are both lifelike and in harmony with music remains a complex challenge [9,27,87]. The introduction of AIOZ-GDance [40] stands as" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 141, + 114, + 250, + 178 + ], + "blocks": [ + { + "bbox": [ + 141, + 114, + 250, + 178 + ], + "lines": [ + { + "bbox": [ + 141, + 114, + 250, + 178 + ], + "spans": [ + { + "bbox": [ + 141, + 114, + 250, + 178 + ], + "type": "image", + "image_path": "9a90bad30506cda01255b6a5143c3095f8e6726525ee39c3f98d8b43d43c3176.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 189, + 482, + 222 + ], + "lines": [ + { + "bbox": [ + 130, + 189, + 482, + 222 + ], + "spans": [ + { + "bbox": [ + 130, + 189, + 482, + 222 + ], + "type": "text", + "content": "Fig. 1: We present a new group dance generation method that can generate a large number of dancers within a fixed resource consumption. The illustration shows a generated group dance sample with 100 dancers." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 252, + 114, + 361, + 178 + ], + "blocks": [ + { + "bbox": [ + 252, + 114, + 361, + 178 + ], + "lines": [ + { + "bbox": [ + 252, + 114, + 361, + 178 + ], + "spans": [ + { + "bbox": [ + 252, + 114, + 361, + 178 + ], + "type": "image", + "image_path": "61ccfb172c209e5731fa05a7051c23fa14c514d12c91f6ea07bc1bb086dc70a5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 362, + 114, + 472, + 178 + ], + "blocks": [ + { + "bbox": [ + 362, + 114, + 472, + 178 + ], + "lines": [ + { + "bbox": [ + 362, + 114, + 472, + 178 + ], + "spans": [ + { + "bbox": [ + 362, + 114, + 472, + 178 + ], + "type": "image", + "image_path": "cb1cbcb4bd334846ab2e10e79ee6ef277429dbb37d9d038b4efcefea77391613.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 242, + 482, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 242, + 482, + 386 + ], + "spans": [ + { + "bbox": [ + 130, + 242, + 482, + 386 + ], + "type": "text", + "content": "the extensive dataset to aid in generating group choreography. Different aspects of group dance motion such as the study of consistency and diversity among the movements of dancers are explored in [39]. However, despite recent advances in the field, existing methods can either only generate dances for a limited number of dancers given the input audio [40, 86, 90], or would consume an expensively large amount of memory due to their design nature [39, 40]. Typically, these methods are usually based on specialized collaborative mechanisms such as cross-entity attention [40, 51], group global attention [39], or residual communication block [66] achieve coherent multi-person synthesis. In other words, their innate architectural design requires the network to process and synthesize all motions simultaneously, leading to excessively growing computational/memory costs with the increasing number of inputs, and hence extremely difficult to scale up." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 388, + 482, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 482, + 532 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 482, + 532 + ], + "type": "text", + "content": "Despite the comprehensive exploration of group dance motion generation, the predominant emphasis remains on optimizing the execution of the generated movements [14, 19, 31, 37, 52, 93]. These existing methodologies function within a confined scope, often tethered to a predetermined maximum number of dancers depicted in group choreography videos outlined by the predefined dataset [39]. This inherent limitation poses a potential constraint on the adaptability of trained models when transitioning to real-world applications [47]. In our pursuit, we center our efforts on addressing the critical aspect of scalability in group dance. Our primary objective revolves around not only expanding the number of dancers but also ensuring the preservation of the innate naturalness in the dance motions and the seamless synchronization between dancers throughout the entirety of their performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "In the field of character motion control and synthesis, existing approaches can be broadly divided into two categories: deterministic [16, 26, 40, 49, 55, 86] and probabilistic [19, 39, 59, 61, 83]. The former tries to learn direct and deterministic mapping from the input conditioning signal (such as text, audio, or user controls) to the desired output motion. However, these methods usually regress towards the mean pose as they are typically trained by minimizing a regression objective under different outputs for similar inputs, resulting in freezing or drifting motion. On the other hand, probabilistic methods attempt to capture the distribution over possible motions for a given condition, allowing diverse and high-quality generation. In particular, Variational Autoencoder (VAE) [18, 47, 60, 61], or Diffusion Models [4, 39, 81, 83, 97] have been widely" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "N. Le et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 212 + ], + "type": "text", + "content": "used. Although diffusion models have recently shown strong potential in generating high-quality motions [81], they generally operate on high-dimensional space (same as the original data) [39,81,83], hence rendering them extremely challenging to use for scalable group generation. Meanwhile, standard VAE models often rely on a Gaussian latent space of fixed dimension for sampling a single latent vector to generate motions, making it difficult to adapt to a wide range of architectures. Besides, the fixed-size single latent space might also not be sufficient to represent the temporal dynamics of the whole motion sequence [7,60]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 216, + 480, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 216, + 480, + 395 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 480, + 395 + ], + "type": "text", + "content": "Recent studies have demonstrated the significant advantages of learning motion features in the frequency domain. Specifically, [22, 74-76] have found that different motion skills can be effectively represented by several phase variables that can faithfully capture the spatial-temporal alignment of a wide range of movements. This can be used to combine with common deep motion synthesis networks to enhance its capability, facilitating natural, and stable motion generations by enforcing a unidirectional motion transition (i.e., no backward direction in time) to prevent the motion from being stuck temporally [74]. Unlike those prior works, where the phases are utilized as auxiliary signals to assist in the motion synthesis process [68, 74, 88], we propose to learn to generate the phases conditioned on the input audio signal in an end-to-end manner, eliminating the need of training extra components. Our generative phase manifold can be used to generate new motions efficiently, from which the latent motion curves are sampled in frequency domain and decoded into the original motion space by a learned decoder." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 399, + 481, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 481, + 567 + ], + "type": "text", + "content": "In this paper, our goal is to develop a scalable technique for group dance generation, a phase-based variational generative model for scalable group dance generation, namely Phase-conditioned Dance VAE (PDVAE). To our knowledge, PDVAE is the first method to represent the variational latent space using phase parameters in the frequency domain of the motion curves. Our method goes beyond the conventional VAE approach that typically relies on a single latent vector drawn from a Gaussian distribution, which is unable to adequately represent the temporal information of the motion sequence (e.g., the time dimension is squeezed out). Figure 1 illustrates an example of a dance motion sequence with 100 dancers generated by our model. In contrast to existing methods, our model can generate crowd-dance animations for an unlimited number of dancers without increasing computational burden (i.e., memory consumption remains constant), while still maintaining the performance and fidelity of the generated choreographies. To summarize, our key contributions are as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 578, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 137, + 578, + 479, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 578, + 479, + 613 + ], + "spans": [ + { + "bbox": [ + 137, + 578, + 479, + 613 + ], + "type": "text", + "content": "- We introduce PDVAE, a phase-based variational generative model for scalable group dance generation. The method focuses on generating large-scale group dance under limited resources." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 617, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 617, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 137, + 617, + 480, + 665 + ], + "type": "text", + "content": "- To effectively learn the manifold that is group-consistent (i.e., dancers within a group lie upon the same manifold), we propose a group consistency loss that enforces the networks to encode the latent phase manifold to be identical for the same group given the input music." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "text", + "content": "Scalable Group Choreography via Variational Phase Manifold Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 137, + 116, + 481, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 116, + 481, + 152 + ], + "spans": [ + { + "bbox": [ + 137, + 116, + 481, + 152 + ], + "type": "text", + "content": "- Extensive experiments along with thorough user study evaluations demonstrate the state-of-the-art performance of our model while achieving effective scalability." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 167, + 237, + 179 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 167, + 237, + 179 + ], + "spans": [ + { + "bbox": [ + 132, + 167, + 237, + 179 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 188, + 301, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 188, + 301, + 201 + ], + "spans": [ + { + "bbox": [ + 132, + 188, + 301, + 201 + ], + "type": "text", + "content": "2.1 Music-driven Choreography" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 204, + 481, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 204, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 130, + 204, + 481, + 300 + ], + "type": "text", + "content": "Crafting natural human choreography derived from music presents a multifaceted challenge [30,62,89]. Certain methodologies integrate constraints based on music-motion similarity matching to ensure coherence between the generated motion and the music [10,33,35,43,65,69]. However, many of these approaches rely considerably on heuristic algorithms, stitching together pre-existing dance segments sourced from a limited music-dance database [10, 48]. While successful in generating extended and realistic dance sequences, these methods face limitations when endeavoring to create entirely novel dance fragments [57, 90]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 300, + 481, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 481, + 432 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 481, + 432 + ], + "type": "text", + "content": "In the recent period, advancements have been evident in the domain of converting music into dance movements through various techniques such as Convolutional Networks (CNN) [1,8,77,91,92,99], Recurrent Networks (RNN) [3,26,77,79,87], Graph Neural Networks (GNN) [5,12,64,95], Generative Adversarial Networks (GAN) [41,77], or Transformer models [34,44,45,49,59,71]. These methodologies typically depend on various inputs, including the present music and a concise history of previous dance movements, to forecast forthcoming sequences of human poses. This innovation introduces a music-text feature fusion module, engineered to amalgamate inputs into a motion decoder, thereby enabling the creation of dance sequences conditioned upon both musical and textual instructions [17]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 432, + 481, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 432, + 481, + 612 + ], + "spans": [ + { + "bbox": [ + 130, + 432, + 481, + 612 + ], + "type": "text", + "content": "While these techniques show promise in generating authentic and lifelike dance motions, they often struggle to synchronize movements seamlessly across multiple dancers [40]. Specifically, achieving coordination and harmony among dancers necessitates consideration of spatial and temporal relationships, their interactions, and the overall choreographic structure [84]. Consequently, further advancements in this domain are being pursued to tackle these complexities [25, 47, 59, 98]. For instance, Perez et al. [59] integrate a multimodal transformer encoder with a normalizing-flow-based decoder to estimate a probability distribution encompassing potential subsequent poses. Additionally, Feng et al. [11] enable long-term generation by imposing a motion repeat constraint to forecast future frames while considering historical motions. A recent study by Le et al. [39] investigates consistency and diversity factors between the generated motions of two or more dancers within a timeframe. However, these approaches are limited to a predefined number of dancers, heavily constrained by the maximum number of dancers in the dataset-provided videos." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 624, + 294, + 638 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 624, + 294, + 638 + ], + "spans": [ + { + "bbox": [ + 132, + 624, + 294, + 638 + ], + "type": "text", + "content": "2.2 Motion Manifold Learning" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 481, + 666 + ], + "type": "text", + "content": "Motion manifold learning has attracted considerable attention in computer vision and artificial intelligence, with the primary goal of comprehending the funda-" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "N. Le et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "content": "mental structures inherent in human movement and dynamics [16, 24, 56, 82]. Its distinctive ability to generate human movement patterns presents numerous opportunities to comprehend intrinsic motion dynamics, manage nonlinear relationships in motion data, and acquire contextual and hierarchical representations [29, 63, 88]. Consequently, various methodologies have emerged, each contributing distinct perspectives and techniques to advance the comprehension and synthesis of human motion [28, 38, 63, 68]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 482, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 482, + 369 + ], + "type": "text", + "content": "For instance, Holden et al. [23] generate character movements by mapping high-level parameters to the human motion manifold, allowing diverse movements without manual preprocessing, and enabling post-generation editing for natural, smooth motion sequences. MotionCLIP [80] introduces a 3D human motion auto-encoder aligned with the Contrastive Language-Image Pre-training (CLIP) model's space, enabling semantic text-based motion generation, disentangled editing, and abstract language specification, leveraging CLIP's rich semantic knowledge within the motion manifold. Recently, Sun et al. [78] utilize a VQ-VAE to acquire a low-dimensional manifold, effectively cleansing the motion sequences. Regarding the scalability of group dance generation, the concept of the motion manifold emerges as a potential solution for addressing the restricted number of dancers in the dataset. This direction enables the learning of a distribution that facilitates the extraction of dance motions, allowing for the presentation of cohesive group dance sequences." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 386, + 449, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 386, + 449, + 415 + ], + "spans": [ + { + "bbox": [ + 131, + 386, + 449, + 415 + ], + "type": "text", + "content": "3 Variational Phase Manifold Learning for Scalable Group Choreography" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 426, + 238, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 426, + 238, + 437 + ], + "spans": [ + { + "bbox": [ + 131, + 426, + 238, + 437 + ], + "type": "text", + "content": "3.1 Task Definition" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 447, + 482, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 447, + 482, + 591 + ], + "spans": [ + { + "bbox": [ + 130, + 447, + 482, + 591 + ], + "type": "text", + "content": "Given an input music sequence " + }, + { + "bbox": [ + 130, + 447, + 482, + 591 + ], + "type": "inline_equation", + "content": "\\mathbf{a} = \\{a_1, a_t, \\dots, a_T\\}" + }, + { + "bbox": [ + 130, + 447, + 482, + 591 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 447, + 482, + 591 + ], + "type": "inline_equation", + "content": "t = \\{1, \\dots, T\\}" + }, + { + "bbox": [ + 130, + 447, + 482, + 591 + ], + "type": "text", + "content": " indicates the index of the music frames, our goal is to generate the group motion sequences of " + }, + { + "bbox": [ + 130, + 447, + 482, + 591 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 447, + 482, + 591 + ], + "type": "text", + "content": " arbitrary dancers: " + }, + { + "bbox": [ + 130, + 447, + 482, + 591 + ], + "type": "inline_equation", + "content": "\\mathbf{x} = \\{x_1^1, \\dots, x_T^1; \\dots; x_1^N, \\dots, x_T^N\\}" + }, + { + "bbox": [ + 130, + 447, + 482, + 591 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 447, + 482, + 591 + ], + "type": "inline_equation", + "content": "x_t^n" + }, + { + "bbox": [ + 130, + 447, + 482, + 591 + ], + "type": "text", + "content": " is the pose of " + }, + { + "bbox": [ + 130, + 447, + 482, + 591 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 447, + 482, + 591 + ], + "type": "text", + "content": "-th dancer at frame " + }, + { + "bbox": [ + 130, + 447, + 482, + 591 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 447, + 482, + 591 + ], + "type": "text", + "content": ". We follow [19, 53] and represent dance as a sequence of poses in the 24-joint of the SMPL body model [54], using the 6D continuous rotation [96] for every joint, along with 3D joint positions and velocities. Additionally, the corresponding 3D root translation vectors are concatenated into the pose representations to involve the trajectory of motion. Previous group dance methods [40], which generate the whole group at once, cannot deal with the increasing number of dancers and can only create group sequences up to a pre-defined number of dancers, due to the vast complexity of the architecture. In contrast, we aim to generate group dance with an unlimited number of dancers." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 609, + 316, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 609, + 316, + 620 + ], + "spans": [ + { + "bbox": [ + 131, + 609, + 316, + 620 + ], + "type": "text", + "content": "3.2 Phase-conditioned Dance VAE" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "Our goal is to learn a continuous manifold such that the motion can be generated by sampling from this learned manifold. We assume that although different dancers within the same group may present visually distinctive movements, the" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "text", + "content": "Scalable Group Choreography via Variational Phase Manifold Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 115, + 473, + 292 + ], + "blocks": [ + { + "bbox": [ + 136, + 115, + 473, + 292 + ], + "lines": [ + { + "bbox": [ + 136, + 115, + 473, + 292 + ], + "spans": [ + { + "bbox": [ + 136, + 115, + 473, + 292 + ], + "type": "image", + "image_path": "a1e8235f870389f41b9939bcda93158b4a7097b0a339e531cc87b01f740444d5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 303, + 482, + 403 + ], + "lines": [ + { + "bbox": [ + 130, + 303, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 303, + 482, + 403 + ], + "type": "text", + "content": "Fig. 2: Overview of our Phase-conditioned Dance VAE (PDVAE) for scalable group dance generation. It consists of an Encoder, a Prior, and a Decoder network. During training, we encode the corresponding motion and music inputs into a latent phase manifold, which is variational and parameterized by the frequency domain parameters of periodic functions. The latent phases can be sampled from the manifold and then decoded back to the original data space to obtain new motions. The consistency loss " + }, + { + "bbox": [ + 130, + 303, + 482, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{csc}}" + }, + { + "bbox": [ + 130, + 303, + 482, + 403 + ], + "type": "text", + "content": " is further imposed to constrain the manifold to be consistently unified for dancers that belong to the same group. At inference stage, only the Prior and the Decoder are used to synthesize group dances efficiently." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 422, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 482, + 567 + ], + "type": "text", + "content": "properties of their motions, such as timing, periodicity, or temporal alignment are intrinsically similar. Drawing inspiration from [74], we aim to learn a generative phase representation for each group of dancer in order to synthesize their motion indefinitely. Our generative model is built upon the conditional Variational Autoencoder architecture [73], thanks to its diverse generation capability and fast sampling speed. However, instead of directly encoding the data into a Gaussian latent distribution as in common VAE approaches [15, 41, 47, 60, 61], we model the latent variational distribution by the phase parameters extracted from the latent motion curve, which we call variational phase manifold. The latent phase manifold is well-structured and can well describe key characteristics of motion (such as its timing, local periodicity, and transition), which benefits learning motion features [74]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "The overview of our Phase-conditioned Dance VAE is illustrated in Figure 2. Specifically, the model contains three main networks: an encoder " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " to capture the approximate posterior distribution conditioned on both motion and music " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "q_{\\phi}(\\mathbf{z}|\\mathbf{x},\\mathbf{a})" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": ", a prior network " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " to learn the conditional prior given only the music " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\mathbf{z}|\\mathbf{a})" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": ", and a decoder " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " to learn to reconstruct the data from the latent distribution " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\mathbf{x}|\\mathbf{z},\\mathbf{a})" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": ". The new motion is generated by sampling the frequency-domain parameters predicted by the prior network, which is then passed through the decoder network to reconstruct the motion in the original data space. Fur-" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "text", + "content": "N. Le et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "thermore, we adopt Transformer-based architecture [85] in each network to effectively capture long-range dependencies and holistic context of the dance sequence." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 169, + 480, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 169, + 480, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 169, + 480, + 326 + ], + "type": "text", + "content": "Encoder. The encoder " + }, + { + "bbox": [ + 130, + 169, + 480, + 326 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 130, + 169, + 480, + 326 + ], + "type": "text", + "content": " is expected to take both the motion and music feature sequence as input, and produce a distribution over possible latent variables capturing the cross-modal relationship between them. To transform the joint input space into a learned phase manifold, we adopt the Transformer decoder architecture where the Cross-Attention mechanism [85] is utilized to learn the relationship between the motion and the music. Accordingly, the output of the encoder is a batch of latent curves (i.e., the activation sequences per channel) that can particularly capture different spatial and temporal aspects of the motion sequence. However, instead of training the model to directly reconstruct the input motion from the extracted latent curves, we further enforce each channel of the latent space to have a periodic functional form (i.e., sinusoidal). This enables us to effectively learn a compact parameterization for each latent channel from a small set of parameters in the frequency domain." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 342, + 479, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 479, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 479, + 449 + ], + "type": "text", + "content": "Generative Variational Phase Manifold. Here we focus on learning the periodicity and non-linear temporal alignment of the motion in the latent space. In particular, given the output latent curves from the encoder " + }, + { + "bbox": [ + 130, + 342, + 479, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{L} = \\mathcal{E}(x,a)\\in \\mathbb{R}^{D\\times T}" + }, + { + "bbox": [ + 130, + 342, + 479, + 449 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 342, + 479, + 449 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 342, + 479, + 449 + ], + "type": "text", + "content": " is the number of desired phase channels to be extracted from the motion, we parameterize each latent curve in " + }, + { + "bbox": [ + 130, + 342, + 479, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{L}" + }, + { + "bbox": [ + 130, + 342, + 479, + 449 + ], + "type": "text", + "content": " using a sinusoidal function with amplitude (A), frequency (F), offset (B) and phase shift (S) parameters [74]. To allow for variational phase manifold learning, we opt to predict two sets of parameters " + }, + { + "bbox": [ + 130, + 342, + 479, + 449 + ], + "type": "inline_equation", + "content": "\\mu_{\\varepsilon} = \\{\\mu^{A};\\mu^{F};\\mu^{B};\\mu^{S}\\}" + }, + { + "bbox": [ + 130, + 342, + 479, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 342, + 479, + 449 + ], + "type": "inline_equation", + "content": "\\sigma_{\\varepsilon} = \\{\\sigma^{A};\\sigma^{F};\\sigma^{B};\\sigma^{S}\\}" + }, + { + "bbox": [ + 130, + 342, + 479, + 449 + ], + "type": "text", + "content": ", which corresponds to the mean and variance of " + }, + { + "bbox": [ + 130, + 342, + 479, + 449 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{4D}" + }, + { + "bbox": [ + 130, + 342, + 479, + 449 + ], + "type": "text", + "content": " dimensional Gaussian distribution:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 252, + 460, + 480, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 460, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 252, + 460, + 480, + 472 + ], + "type": "interline_equation", + "content": "q _ {\\phi} (\\mathbf {z} | \\mathbf {x}, \\mathbf {a}) = \\mathcal {N} (\\mathbf {z}; \\mu_ {\\varepsilon}, \\sigma_ {\\varepsilon}) \\tag {1}", + "image_path": "c1644fe9df0a9202d0e815e45e1b9b31dbfbd8a2432203f6e2d8754d41de5720.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 480, + 479, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 480, + 479, + 563 + ], + "spans": [ + { + "bbox": [ + 130, + 480, + 479, + 563 + ], + "type": "text", + "content": "To do so, we first apply differentiable Fast Fourier Transform (FFT) to each channel of the latent curve " + }, + { + "bbox": [ + 130, + 480, + 479, + 563 + ], + "type": "inline_equation", + "content": "\\mathbf{L}" + }, + { + "bbox": [ + 130, + 480, + 479, + 563 + ], + "type": "text", + "content": " and create the zero-indexed matrix of Fourier coefficients as " + }, + { + "bbox": [ + 130, + 480, + 479, + 563 + ], + "type": "inline_equation", + "content": "\\mathbf{c} = FFT(\\mathbf{L})" + }, + { + "bbox": [ + 130, + 480, + 479, + 563 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 480, + 479, + 563 + ], + "type": "inline_equation", + "content": "\\mathbf{c} \\in \\mathbb{C}^{D \\times K + 1}" + }, + { + "bbox": [ + 130, + 480, + 479, + 563 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 480, + 479, + 563 + ], + "type": "inline_equation", + "content": "K = \\left\\lfloor \\frac{T}{2} \\right\\rfloor" + }, + { + "bbox": [ + 130, + 480, + 479, + 563 + ], + "type": "text", + "content": ". Correspondingly, we compute the per channel power spectrum " + }, + { + "bbox": [ + 130, + 480, + 479, + 563 + ], + "type": "inline_equation", + "content": "\\mathbf{p} \\in \\mathbb{R}^{D \\times K + 1}" + }, + { + "bbox": [ + 130, + 480, + 479, + 563 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 480, + 479, + 563 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{i,j} = \\frac{2}{N} |\\mathbf{c}_{i,j}|^2" + }, + { + "bbox": [ + 130, + 480, + 479, + 563 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 480, + 479, + 563 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 480, + 479, + 563 + ], + "type": "text", + "content": " is the channel index and " + }, + { + "bbox": [ + 130, + 480, + 479, + 563 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 480, + 479, + 563 + ], + "type": "text", + "content": " is the index for the frequency bands. Correspondingly, the distributional mean parameters of the periodic sinusoidal function are then calculated as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 187, + 571, + 480, + 608 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 571, + 480, + 608 + ], + "spans": [ + { + "bbox": [ + 187, + 571, + 480, + 608 + ], + "type": "interline_equation", + "content": "\\mu_ {i} ^ {A} = \\sqrt {\\frac {2}{T} \\sum_ {j = 1} ^ {K} \\mathbf {p} _ {i , j}}, \\quad \\mu_ {i} ^ {F} = \\frac {\\sum_ {j = 1} ^ {K} \\mathbf {f} _ {j} \\cdot \\mathbf {p} _ {i , j}}{\\sum_ {j = 1} ^ {K} \\mathbf {p} _ {i , j}}, \\quad \\mu_ {i} ^ {B} = \\frac {\\mathbf {c} _ {i , 0}}{T}, \\tag {2}", + "image_path": "304db3da81144de38a60fdb029a98c25ced62d342f36dead3c0160ff70224440.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 616, + 479, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 616, + 479, + 651 + ], + "spans": [ + { + "bbox": [ + 130, + 616, + 479, + 651 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 616, + 479, + 651 + ], + "type": "inline_equation", + "content": "\\mathbf{f} = (0, \\frac{1}{T}, \\dots, \\frac{K}{T})" + }, + { + "bbox": [ + 130, + 616, + 479, + 651 + ], + "type": "text", + "content": " is the frequencies vector. At the same time, the phase shift " + }, + { + "bbox": [ + 130, + 616, + 479, + 651 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 130, + 616, + 479, + 651 + ], + "type": "text", + "content": " is predicted using a fully-connected (FC) layer with two arctan activation as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 218, + 653, + 480, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 218, + 653, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 218, + 653, + 480, + 666 + ], + "type": "interline_equation", + "content": "\\left(s _ {y}, s _ {x}\\right) = \\operatorname {F C} \\left(\\mathbf {L} _ {i}\\right), \\quad \\mu_ {i} ^ {S} = \\arctan \\left(s _ {y}, s _ {x}\\right), \\tag {3}", + "image_path": "bc5b6c5b97d5311eed4c5e5dadc6bcaf255c8be87b55e3a73507c390a60d3ad5.jpg" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 157, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 157, + 91, + 448, + 103 + ], + "type": "text", + "content": "Scalable Group Choreography via Variational Phase Manifold Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": "To predict the distributional variance of the phase amplitude and phase shift parameters " + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "inline_equation", + "content": "\\{\\sigma^A,\\sigma^S\\}" + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": ", We additionally apply a separate two-layer MLP network over each channel of the latent curves, similar to Equation 3. The variational latent phase parameters are sampled by utilizing parameterization trick [36], i.e., " + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "inline_equation", + "content": "\\mathbf{A} \\sim \\mathcal{N}(\\mu^A,\\sigma^A)" + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "inline_equation", + "content": "\\mathbf{S} \\sim \\mathcal{N}(\\mu^S,\\sigma^S)" + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": ". In our experiments, we find that sampling the phase frequency " + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": " and offset " + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": " often produce unstable and noncoherent group movements. This might be because the frequency amplitudes of the dancers within the same group are likely to associate with the rhythmic pattern of the musical beats while the offsets capture their alignment, thereby should be consistent with each other. Therefore, we treat those parameters as deterministic by constraining their variance to zero." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 251, + 482, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 251, + 482, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 482, + 289 + ], + "type": "text", + "content": "Finally, the sampled set of phase parameters " + }, + { + "bbox": [ + 130, + 251, + 482, + 289 + ], + "type": "inline_equation", + "content": "\\mathbf{z} = \\{\\mathbf{A};\\mathbf{F};\\mathbf{B};\\mathbf{S}\\}" + }, + { + "bbox": [ + 130, + 251, + 482, + 289 + ], + "type": "text", + "content": " are used to reconstruct a parametric latent space consisting of multiple periodic curves to represent each intrinsic property of the motion by:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 235, + 307, + 480, + 319 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 307, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 235, + 307, + 480, + 319 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {L}} = \\mathbf {A} \\cdot \\sin (2 \\pi \\cdot (\\mathbf {F} \\cdot \\mathcal {T} - \\mathbf {S})) + \\mathbf {B} \\tag {4}", + "image_path": "03327eec18f157641c869d7574f30eea43ccdc624324246f68b9cd44226d20c0.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 338, + 482, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 338, + 482, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 482, + 422 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 338, + 482, + 422 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 130, + 338, + 482, + 422 + ], + "type": "text", + "content": " is a known time window series obtained by evenly spacing the timesteps from 0 to " + }, + { + "bbox": [ + 130, + 338, + 482, + 422 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 338, + 482, + 422 + ], + "type": "text", + "content": ". Intuitively this curve construction procedure can be viewed as a \"quantization\" layer to enforce the network to learn to represent the motion features in the frequency domain, which is useful in representing different aspects of human motion such as their timing and periodicity. In the last step, a decoder is utilized to reconstruct the original motion signals from the set of parametric latent curves." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": "Decoder. To decode the latent space into the original motion space, previous works [60,61] have to use a sinusoidal positional encoding sequence with duration " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": " as the proxy input to the sequence decoder. This is because their latent space is only formed by single latent vectors following a Gaussian distribution, which cannot span the time dimension. However, we observe that it usually results in unstable and inconsistent movements, as the proxy sequence is generic and usually contains less meaningful information for the decoder. Meanwhile, our method does not suffer from this problem as our latent space is built on multiple curves that can represent the motion information through time, thanks to the phase parameters. Subsequently, our decoder " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": " is based on Transformer decoder architecture that takes the constructed parametric latent curve, as well as the music features as inputs, to reconstruct the corresponding dance motions. Here, we also utilize the cross-attention model [85] where we consider the sequence of and music features as key and value along with the sampled latent curves as the query. The output of the decoder is a sequence of " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": " vectors in " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^D" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": ", which is then projected back to the original motion dimensions through a linear layer, to obtain the reconstructed outputs " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}} = p_{\\theta}(\\mathbf{x}|\\mathbf{z},\\mathbf{a})" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": ". We additionally employ a global trajectory predictor to predict the global translation of the root joint based on the generated local motions [19, 46], in order to avoid intersection problems between dancers. We provide details of our global predictor in the supplemental." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "N. Le et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 479, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 479, + 188 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 479, + 188 + ], + "type": "text", + "content": "Prior Network. Since the ground-truth motion is generally inaccessible at test time (i.e., we only have access to the music), we also need to learn a prior " + }, + { + "bbox": [ + 131, + 116, + 479, + 188 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 131, + 116, + 479, + 188 + ], + "type": "text", + "content": " to match the posterior distribution of motion from which the latent phase can be sampled. Specifically, We follow the procedure similar to Section 3.2 (Equation 1-4) to predict the Gaussian distribution conditioned on the music sequence " + }, + { + "bbox": [ + 131, + 116, + 479, + 188 + ], + "type": "inline_equation", + "content": "\\mathbf{a}" + }, + { + "bbox": [ + 131, + 116, + 479, + 188 + ], + "type": "text", + "content": ", which is then used for sampling the latent phases:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 255, + 195, + 481, + 209 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 195, + 481, + 209 + ], + "spans": [ + { + "bbox": [ + 255, + 195, + 481, + 209 + ], + "type": "interline_equation", + "content": "p _ {\\theta} (\\mathbf {z} | \\mathbf {a}) = \\mathcal {N} (\\mathbf {z}; \\mu_ {\\mathcal {P}}, \\sigma_ {\\mathcal {P}}) \\tag {5}", + "image_path": "adc9aec89e2a02f2093e93e7e1cddbb9aed4fafc3948372cb7a6c0ab464671e2.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 216, + 480, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 216, + 480, + 312 + ], + "spans": [ + { + "bbox": [ + 131, + 216, + 480, + 312 + ], + "type": "text", + "content": "where a Transformer encoder is used to encode the input conditioning music sequence and predict the corresponding " + }, + { + "bbox": [ + 131, + 216, + 480, + 312 + ], + "type": "inline_equation", + "content": "\\mu_{\\mathcal{P}}" + }, + { + "bbox": [ + 131, + 216, + 480, + 312 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 216, + 480, + 312 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathcal{P}}" + }, + { + "bbox": [ + 131, + 216, + 480, + 312 + ], + "type": "text", + "content": ". We implement the prior network similarly to the encoder network, however, we use self-attention mechanism [85] to capture the global music context. Learning the conditional prior is crucial for the conditional VAE to generalize to diverse types of music and motion. Intuitively speaking, each latent variable " + }, + { + "bbox": [ + 131, + 216, + 480, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 131, + 216, + 480, + 312 + ], + "type": "text", + "content": " is expected to represent possible dance motions " + }, + { + "bbox": [ + 131, + 216, + 480, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 131, + 216, + 480, + 312 + ], + "type": "text", + "content": " conforming to the music context " + }, + { + "bbox": [ + 131, + 216, + 480, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{a}" + }, + { + "bbox": [ + 131, + 216, + 480, + 312 + ], + "type": "text", + "content": ". Therefore, the prior should be able to encode different latent distributions given different musics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 328, + 204, + 340 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 328, + 204, + 340 + ], + "spans": [ + { + "bbox": [ + 132, + 328, + 204, + 340 + ], + "type": "text", + "content": "3.3 Training" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 346, + 479, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 346, + 479, + 370 + ], + "spans": [ + { + "bbox": [ + 131, + 346, + 479, + 370 + ], + "type": "text", + "content": "During training, we consider the following variational lower bound [73] to mainly train our dance generation VAE model:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 179, + 377, + 480, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 377, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 179, + 377, + 480, + 392 + ], + "type": "interline_equation", + "content": "\\log p _ {\\theta} (\\mathbf {x} \\mid \\mathbf {a}) \\geq \\mathbb {E} _ {q _ {\\phi}} \\left[ \\log p _ {\\theta} (\\mathbf {x} \\mid \\mathbf {z}, \\mathbf {a}) \\right] - D _ {\\mathrm {K L}} \\left(q _ {\\phi} (\\mathbf {z} \\mid \\mathbf {x}, \\mathbf {a}) \\| p _ {\\theta} (\\mathbf {z} \\mid \\mathbf {a})\\right) \\tag {6}", + "image_path": "b99313157765505005e33fab6e965ff2d893b8e2e2dc83a5271b1f7915097e02.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 398, + 480, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 398, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 131, + 398, + 480, + 468 + ], + "type": "text", + "content": "In practice, we apply the conditional VAE loss as similar to [61], which is defined as the weighted sum " + }, + { + "bbox": [ + 131, + 398, + 480, + 468 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cvae}} = \\mathcal{L}_{\\mathrm{rec}} + \\lambda_{\\mathrm{KL}}\\mathcal{L}_{\\mathrm{KL}}" + }, + { + "bbox": [ + 131, + 398, + 480, + 468 + ], + "type": "text", + "content": ". In particular, the reconstruction term " + }, + { + "bbox": [ + 131, + 398, + 480, + 468 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{rec}}" + }, + { + "bbox": [ + 131, + 398, + 480, + 468 + ], + "type": "text", + "content": " measures the motion reconstruction error given the decoder output (via a smooth-L1 loss). The KL divergence term " + }, + { + "bbox": [ + 131, + 398, + 480, + 468 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{KL}}" + }, + { + "bbox": [ + 131, + 398, + 480, + 468 + ], + "type": "text", + "content": " compares the divergence " + }, + { + "bbox": [ + 131, + 398, + 480, + 468 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{KL}}" + }, + { + "bbox": [ + 131, + 398, + 480, + 468 + ], + "type": "text", + "content": " between the posterior and the prior distribution to enforce them to be close to each other." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 470, + 480, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 470, + 480, + 554 + ], + "spans": [ + { + "bbox": [ + 131, + 470, + 480, + 554 + ], + "type": "text", + "content": "The conditional VAE objective above is calculated for each dancer separately and cannot capture the correlation between dancers within a group. Therefore, it is essential to impose consistency among dancers and avoid strange effects such as unsynchronized dance. To this end, we propose a group consistency loss by enforcing the latent phase manifold to be similar for the same group, given the input music. Specifically, we first calculate the phase manifold features based on the frequency domain parameters as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 198, + 562, + 480, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 562, + 480, + 574 + ], + "spans": [ + { + "bbox": [ + 198, + 562, + 480, + 574 + ], + "type": "interline_equation", + "content": "\\mathbf {P} _ {2 i - 1} = \\mathbf {A} _ {i} \\sin (2 \\pi \\cdot \\mathbf {S} _ {i}), \\quad \\mathbf {P} _ {2 i} = \\mathbf {A} _ {i} \\cos (2 \\pi \\cdot \\mathbf {S} _ {i}) \\tag {7}", + "image_path": "77640c6425fdfd4d6791b78b29a0db5a5f840ea85cc108d064545872a5543e87.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 581, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 581, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 581, + 481, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 581, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{P} \\in \\mathbb{R}^{2D}" + }, + { + "bbox": [ + 131, + 581, + 481, + 666 + ], + "type": "text", + "content": " is the phase manifold vector that encodes the spatial-temporal information of the motion state. The phase feature may look similar to the positional encodings of transformers [85] in the sense that both use multi-resolution sinusoidal functions. However, the phase feature is much richer in terms of representation capacity since it learns to embed the spatial (body joints) and temporal (positions in time) information of the motion curves, whereas the positional encodings only encode the position of words. Finally, our consistency objective is" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 157, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 157, + 91, + 448, + 103 + ], + "type": "text", + "content": "Scalable Group Choreography via Variational Phase Manifold Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "to constrain phase manifold between dancers within a group to be as close as possible to each other, which is formulated as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 189, + 150, + 481, + 165 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 150, + 481, + 165 + ], + "spans": [ + { + "bbox": [ + 189, + 150, + 481, + 165 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {c s c}} = D _ {\\mathrm {K L}} \\left(q _ {\\phi} (\\mathbf {z} | \\mathbf {x} ^ {m}, \\mathbf {a}) \\| \\left(q _ {\\phi} (\\mathbf {z} | \\mathbf {x} ^ {n}, \\mathbf {a})\\right) + \\| \\mathbf {P} ^ {m} - \\mathbf {P} ^ {n} \\| _ {2} ^ {2} \\right. \\tag {8}", + "image_path": "ab1229e0c2b45bbe88916f8dab0da3fb049d0a0be38f541a7d23858170a29ecd.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 175, + 479, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 175, + 479, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 175, + 479, + 258 + ], + "type": "text", + "content": "where the first term encourages the network to map different dancers belonging to the same group " + }, + { + "bbox": [ + 130, + 175, + 479, + 258 + ], + "type": "inline_equation", + "content": "(\\mathbf{x}^m" + }, + { + "bbox": [ + 130, + 175, + 479, + 258 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 175, + 479, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^n)" + }, + { + "bbox": [ + 130, + 175, + 479, + 258 + ], + "type": "text", + "content": " into the same set of distributional phase parameters while the second term penalizes the discrepancy in their corresponding phase manifolds. In general, this loss is applied to ensure every dancer is embedded into a single unified manifold that can effectively represent their corresponding group. To summarize, our total training loss is defined as the combination of the VAE loss and the consistency loss " + }, + { + "bbox": [ + 130, + 175, + 479, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\mathcal{L}_{\\mathrm{cvae}} + \\lambda_{\\mathrm{csc}}\\mathcal{L}_{\\mathrm{csc}}" + }, + { + "bbox": [ + 130, + 175, + 479, + 258 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 259, + 480, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 480, + 365 + ], + "type": "text", + "content": "For testing, we can efficiently generate motions for an unlimited number of dancers by indefinitely drawing samples from the learned continuous group-consistent phase manifold. It is noteworthy that for inference, we only need to process the prior network once to obtain the latent distribution. To generate a new motion, we can sample from this latent (Gaussian) distribution and use the decoder to decode it back to the motion space. This approach is much more efficient and has significantly higher scalability than previous approaches [39, 40] that is limited by the number of dancers processed simultaneously by the entire network." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 389, + 230, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 389, + 230, + 403 + ], + "spans": [ + { + "bbox": [ + 131, + 389, + 230, + 403 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 418, + 279, + 429 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 418, + 279, + 429 + ], + "spans": [ + { + "bbox": [ + 131, + 418, + 279, + 429 + ], + "type": "text", + "content": "4.1 Implementation Details" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 441, + 480, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 441, + 480, + 549 + ], + "spans": [ + { + "bbox": [ + 130, + 441, + 480, + 549 + ], + "type": "text", + "content": "Our model was trained on 4 NVIDIA V100 GPUs using Adam optimizer with a fixed learning rate of " + }, + { + "bbox": [ + 130, + 441, + 480, + 549 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 130, + 441, + 480, + 549 + ], + "type": "text", + "content": " and a mini-batch size of 32 per GPU. For training losses, the weights are empirically set to " + }, + { + "bbox": [ + 130, + 441, + 480, + 549 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{KL}} = 5\\times 10^{-4}" + }, + { + "bbox": [ + 130, + 441, + 480, + 549 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 441, + 480, + 549 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{csc}} = 10^{-4}" + }, + { + "bbox": [ + 130, + 441, + 480, + 549 + ], + "type": "text", + "content": ", respectively. The Transformer encoders and decoders consist of 5 layers for both encoder, decoder, and prior Network with 512-dimensional hidden units. Meanwhile, the number of latent phase channels is set to 256. To further capture the periodic nature of the phase feature, we also use Siren activation following the initialization scheme as in [70]. This can effectively model the periodicity inherent in the motion data, and thus can benefit motion synthesis [68]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 570, + 272, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 570, + 272, + 582 + ], + "spans": [ + { + "bbox": [ + 131, + 570, + 272, + 582 + ], + "type": "text", + "content": "4.2 Experimental Settings" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 594, + 479, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 479, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 479, + 666 + ], + "type": "text", + "content": "Dataset In our experiments, we utilize the AIOZ-GDance [40] and AIST-M [90] datasets. AIOZ-GDance is the largest music-driven dataset focusing on group dance, encompassing paired music and 3D group motions extracted from in-the-wild videos through a semi-automatic process. This dataset spans 7 dance styles and 16 music genres. For consistency, we adhere to the training and testing split outlined in [40] during our experiments." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "N. Le et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": "Evaluation Protocol We employ several metrics to assess the quality of individual dance motions, including Frechet Inception Distance (FID) [20, 49], Motion-Music Consistency (MMC) [49], and Generation Diversity (GenDiv) [26, 41, 49], along with the Physical Foot Contact score (PFC) [83]. Specifically, the FID score gauges the realism of individual dance movements concerning the ground-truth dance. MMC assesses the matching similarity between motion and music beats, reflecting how well-generated dances synchronize with the music's rhythm. GenDiv is computed as the average pairwise distance of kinetic features among motions [58]. PFC evaluates the physical plausibility of foot movements by determining the agreement between the acceleration of the character's center of mass and the foot's velocity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 247, + 482, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 344 + ], + "type": "text", + "content": "In assessing the group dance quality, we adopt three metrics outlined in [40]: Group Motion Realism (GMR), Group Motion Correlation (GMC), and Trajectory Intersection Frequency (TIF). Broadly, GMR gauges the realism of generated group motions in comparison to ground-truth data, employing Frechet Inception Distance on extracted group motion features. GMC evaluates the synchronization among dancers within the generated group by computing their cross-correlation. TIF quantifies the frequency of collisions among the generated dancers during their dance movements." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 359, + 482, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 359, + 482, + 443 + ], + "spans": [ + { + "bbox": [ + 130, + 359, + 482, + 443 + ], + "type": "text", + "content": "Baselines Our method is subjected to comparison with various recent techniques in music-driven dance generation, namely FACT [49], Transflower [59], and EDGE [83]. These approaches are adapted for benchmarking within group dance generation context, as outlined in [40], considering that their original designs were tailored for single-dance scenarios. Additionally, our evaluation includes comparisons with GDanceR [40], GCD [39], and DanY [90]. All mentioned works are specifically designed for the group choreography generation." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 138, + 449, + 479, + 552 + ], + "blocks": [ + { + "bbox": [ + 138, + 449, + 479, + 552 + ], + "lines": [ + { + "bbox": [ + 138, + 449, + 479, + 552 + ], + "spans": [ + { + "bbox": [ + 138, + 449, + 479, + 552 + ], + "type": "table", + "html": "
DatasetMethodFID↓MMC↑GenDiv↑PFC↓GMR↓GMC↑TIF↓
AIOZ-GDANCE [40]FACT [49]56.200.2228.643.52101.5262.680.321
Transflower [59]37.730.2178.743.0781.1760.780.332
EDGE [83]31.400.2649.572.6363.3561.720.356
GDANCER [40]43.900.2509.233.0551.2779.010.217
GCD [39]31.160.26110.872.5331.4780.970.167
Ours31.010.27110.982.3330.0884.520.163
AIST-M [90]GDANCER [40]52.900.2226.521.9365.1360.560.121
GCD [39]35.360.24510.971.5242.5272.150.083
DanY [90]40.250.24011.401.6550.2963.530.137
Ours31.490.25711.811.4241.2478.640.076
", + "image_path": "ca468c86e3412e2a282545609367f4f0fb254fc2a1637d7d2c2f5bc1f7791017.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 234, + 555, + 378, + 567 + ], + "lines": [ + { + "bbox": [ + 234, + 555, + 378, + 567 + ], + "spans": [ + { + "bbox": [ + 234, + 555, + 378, + 567 + ], + "type": "text", + "content": "Table 1: Performance comparison." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 132, + 587, + 270, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 587, + 270, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 587, + 270, + 600 + ], + "type": "text", + "content": "4.3 Experimental Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Quality Comparison Table 1 presents a comparison among the baselines FACT [49], Transflower [59], EDGE [83], GDanceR [40], GCD [39], and our proposed manifold-based method. The results clearly demonstrate that our model almost outperforms all baselines across all evaluations on two datasets AIOZ-GDANCE [40] and AIST-M [90]. We observe that recent diffusion-based dance" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "text", + "content": "Scalable Group Choreography via Variational Phase Manifold Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "generation models such as EDGE or GCD yield competitive performance on both single-dance metrics (FID, MMC, GenDiv, and PFC) and group dance metrics (GMR, GMC, and TIF). However, due to limitations in their training procedures, they still struggle when dealing with generating multiple dancing motions in the context of a high quantity of dancers, as shown by their low performance compared to our proposal. This result implies that our method successfully preserves the performance of dancing motions when the number of dancers is increased. Besides, Figure 3 also shows that our proposed methods outperform other state-of-the-art methods such as GDanceR or GCD in dealing with monotonous, repetitive, sinking, and overlapping dance motions." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 138, + 247, + 478, + 392 + ], + "blocks": [ + { + "bbox": [ + 138, + 247, + 478, + 392 + ], + "lines": [ + { + "bbox": [ + 138, + 247, + 478, + 392 + ], + "spans": [ + { + "bbox": [ + 138, + 247, + 478, + 392 + ], + "type": "image", + "image_path": "2d2603f27afe8456ee8f4803ace1a428fb7cb155c097f92c859d462ebae5b4fa.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 402, + 482, + 468 + ], + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 468 + ], + "type": "text", + "content": "Fig. 3: Visualization of a dancing sample between different methods. GDanceR displays monotonous, repetitive, or sinking dance motions. GCD exhibits more divergence in dance motions, yet dancers may intersect since their optimization does not address this issue explicitly. Blue boxes mark these issues. In contrast, our manifold-based solution ensures the divergence of dancing motions, while the phase motion path demonstrates its effectiveness in addressing floating and crossing issues in group dances." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 136, + 483, + 479, + 569 + ], + "blocks": [ + { + "bbox": [ + 136, + 483, + 479, + 569 + ], + "lines": [ + { + "bbox": [ + 136, + 483, + 479, + 569 + ], + "spans": [ + { + "bbox": [ + 136, + 483, + 479, + 569 + ], + "type": "table", + "html": "
#Generated DancersMethodFID↓MMC↑GenDiv↑GMR↓GMC↑TIF↓
5GDanceR [40]44.190.2498.9955.0578.720.218
GCD [39]35.080.2649.9238.4381.440.168
Ours31.350.26810.0532.5884.560.161
10GDanceR [40]57.930.2257.6373.3272.790.386
GCD [39]N/A (Memory)N/A (Memory)N/A (Memory)N/A (Memory)N/A (Memory)N/A (Memory)
Ours32.190.2698.9934.3286.960.193
100GDanceR [40]N/A (Memory)N/A (Memory)N/A (Memory)N/A (Memory)N/A (Memory)N/A (Memory)
GCD [39]N/A (Memory)N/A (Memory)N/A (Memory)N/A (Memory)N/A (Memory)N/A (Memory)
Ours30.970.2688.7638.1385.730.222
", + "image_path": "00a18dd431b0bec6abb28d14e8a97bb6053554ec973e8fbaadef02f976c32430.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 572, + 482, + 607 + ], + "lines": [ + { + "bbox": [ + 130, + 572, + 482, + 607 + ], + "spans": [ + { + "bbox": [ + 130, + 572, + 482, + 607 + ], + "type": "text", + "content": "Table 2: Performance of group dance generation methods when we increase the number of generated dancers. The experiments are done with common consumer GPUs with 4GB memory. (N/A means models could not run due to inadequate memory footprint)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": "Scalable Generation Analysis Table 2 illustrates the performance of different group dance generation methods (GCD, GDancer, and Ours) when generating dance movements with an increasing number of dancers. When the number of" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "N. Le et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": "dancers is increased to 10, GCD appears to run out of memory, which is also observed in GDanceR when the number of dancers increases to 100. Regardless of the number of dancers, our method consistently achieves stable and competitive results. This implies that our proposed method successfully addresses the scalability issue in group dance generation without compromising the overall performance of each individual dance motion. Figure 4 illustrates the memory consumption to generate dance motions in groups for each method. Noticeably, our proposal still achieves the highest performance while consuming immensely fewer resources required for generating group dance motions (See Figure 5 for illustrations). This, again, indicates that our method successfully breaks the barrier of limited generated dancers by using the manifold." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 266, + 482, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 266, + 482, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 482, + 434 + ], + "type": "text", + "content": "Ablation Analysis Table 3 presents the performance improvements achieved through the integration of consistency loss and phase manifold. Additionally, we showcase the effectiveness of our proposed approach across three different backbones: Transformer [49, 85], LSTM [21], and CNN [99]. Evaluation metrics including FID, GMR, and GMC are utilized. The results indicate that the absence of consistency loss leads to an increase in GMR and a decrease in GMC, suggesting a significant enhancement in the realism and correlation of group dance motions facilitated by the inclusion of the proposed objective. Meanwhile, with out the phase manifold, the model exhibits remarkably higher scores in both the FID and GMR metrics, suggesting that phase manifold can effectively improve the distinction in dance motions while maintaining the realism of group dances, even when the number of dancers in a group is large. In comparing three backbones—Transformer, LSTM, and CNN—we have observed that the chosen Transformer backbone achieved the best results compared to LSTM or CNN." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 137, + 464, + 299, + 531 + ], + "blocks": [ + { + "bbox": [ + 137, + 464, + 299, + 531 + ], + "lines": [ + { + "bbox": [ + 137, + 464, + 299, + 531 + ], + "spans": [ + { + "bbox": [ + 137, + 464, + 299, + 531 + ], + "type": "table", + "html": "
MethodFID↓GMR↓GMC↑
Ours31.0130.0884.52
Ours w/o Consistency Loss35.3557.6366.72
Ours w/o Phase Manifold41.7845.3277.93
Ours w LSTM41.2947.4771.82
Ours w CNN36.9944.9475.77
", + "image_path": "9c37744cff0978b08033b4584e15067eebd8ab549bd74b8c8a0336189fd0a50b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 539, + 301, + 562 + ], + "lines": [ + { + "bbox": [ + 132, + 539, + 301, + 562 + ], + "spans": [ + { + "bbox": [ + 132, + 539, + 301, + 562 + ], + "type": "text", + "content": "Table 3: Module contribution and loss analysis." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 312, + 444, + 482, + 540 + ], + "blocks": [ + { + "bbox": [ + 312, + 444, + 482, + 540 + ], + "lines": [ + { + "bbox": [ + 312, + 444, + 482, + 540 + ], + "spans": [ + { + "bbox": [ + 312, + 444, + 482, + 540 + ], + "type": "image", + "image_path": "e83930fcbac387bcdfd345f08dec3d2db2014f9f814017447ff288ea36b23994.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 311, + 540, + 482, + 562 + ], + "lines": [ + { + "bbox": [ + 311, + 540, + 482, + 562 + ], + "spans": [ + { + "bbox": [ + 311, + 540, + 482, + 562 + ], + "type": "text", + "content": "Fig. 4: Memory usage vs. number of dancers in different dance generators." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 585, + 218, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 585, + 218, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 585, + 218, + 597 + ], + "type": "text", + "content": "4.4 User Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "User studies are vital for evaluating generative models, as user perception is pivotal for downstream applications; thus, we conducted two studies with around 70 participants each, diverse in background, with experience in music and dance, aged between 20 to 40, consisting of approximately " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "47\\%" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " females and " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "53\\%" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " males, to assess the effectiveness of our approach in group choreography generation." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 157, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 157, + 90, + 447, + 102 + ], + "type": "text", + "content": "Scalable Group Choreography via Variational Phase Manifold Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "In the user study, we aim to assess the realism of sample outputs with more and more dancers. Specifically, participants assign scores ranging from 0 to 10 to evaluate the realism of each dance clip with 2 to 10 dancers. Figure 6 shows that, across all methods, the more the number of dancers is increased, the lower the realism is found. However, the drop in realism of our proposed method is the least compared to GCD and GDanceR. The results indicate our method's good performance compared to other baselines when the number of dancers increases." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 138, + 212, + 304, + 379 + ], + "blocks": [ + { + "bbox": [ + 138, + 212, + 304, + 379 + ], + "lines": [ + { + "bbox": [ + 138, + 212, + 304, + 379 + ], + "spans": [ + { + "bbox": [ + 138, + 212, + 304, + 379 + ], + "type": "image", + "image_path": "7e085c9cc0289882b96eb06347ae0cab883ac551831d5e51d18109e8dc40c698.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 136, + 384, + 303, + 396 + ], + "lines": [ + { + "bbox": [ + 136, + 384, + 303, + 396 + ], + "spans": [ + { + "bbox": [ + 136, + 384, + 303, + 396 + ], + "type": "text", + "content": "Fig. 5: Visualization of scalable dancers." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 316, + 212, + 481, + 374 + ], + "blocks": [ + { + "bbox": [ + 316, + 212, + 481, + 374 + ], + "lines": [ + { + "bbox": [ + 316, + 212, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 212, + 481, + 374 + ], + "type": "image", + "image_path": "b0e89c61c98ff2756242f7040614fc7693068996fb80a37866ced9d1e9e31c29.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 375, + 480, + 395 + ], + "lines": [ + { + "bbox": [ + 315, + 375, + 480, + 395 + ], + "spans": [ + { + "bbox": [ + 315, + 375, + 480, + 395 + ], + "type": "text", + "content": "Fig. 6: Realism between different methods when number of dancers is varied." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 421, + 217, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 421, + 217, + 433 + ], + "spans": [ + { + "bbox": [ + 132, + 421, + 217, + 433 + ], + "type": "text", + "content": "5 Discussion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 449, + 482, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 520 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 520 + ], + "type": "text", + "content": "While our approach leverages the VAE as a primary solution for generating a manifold, it is important to acknowledge certain inherent limitations associated with this choice. One notable challenge is the susceptibility to issues such as posterior collapse and unstable sampling within the VAE framework. These challenges can result in generated group dance motions that may not consistently meet performance expectations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 521, + 482, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 592 + ], + "type": "text", + "content": "One specific manifestation of this limitation is the potential for false decoding when sampling points that lie too far from the learned distribution. This scenario can lead to unexpected rotations or disruptions in the physics of the generated content. The impact of this problem becomes evident in instances where the generated samples deviate significantly from the anticipated distribution, introducing inaccuracies and distortions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": "To address these challenges, we recognize the need for ongoing efforts to mitigate the effects of posterior collapse and unstable sampling. While the problem is acknowledged, our approach incorporates measures to limit its impact. Future research directions could explore alternative generative models or additional techniques to enhance the robustness and reliability of the generated results in the face of these identified limitations." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "N. Le et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 137, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 138, + 137, + 481, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 137, + 481, + 169 + ], + "spans": [ + { + "bbox": [ + 138, + 137, + 481, + 169 + ], + "type": "text", + "content": "1. Ahn, H., Kim, J., Kim, K., Oh, S.: Generative autoregressive networks for 3d dancing move synthesis from music. IEEE Robotics and Automation Letters (RA-L) (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 170, + 481, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 170, + 481, + 203 + ], + "spans": [ + { + "bbox": [ + 138, + 170, + 481, + 203 + ], + "type": "text", + "content": "2. Alaoui, S.F., Henry, C., Jacquemin, C.: Physical modelling for interactive installations and the performing arts. International Journal of Performance Arts and Digital Media (2014)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 203, + 481, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 203, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 138, + 203, + 481, + 224 + ], + "type": "text", + "content": "3. Alemi, O., Françoise, J., Pasquier, P.: Groovenet: Real-time music-driven dance movement generation using artificial neural networks. Networks (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 224, + 481, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 224, + 481, + 256 + ], + "spans": [ + { + "bbox": [ + 138, + 224, + 481, + 256 + ], + "type": "text", + "content": "4. Alexanderson, S., Nagy, R., Beskow, J., Henter, G.E.: Listen, denoise, action! audio-driven motion synthesis with diffusion models. ACM Transactions on Graphics (TOG) (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 257, + 481, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 257, + 481, + 289 + ], + "spans": [ + { + "bbox": [ + 138, + 257, + 481, + 289 + ], + "type": "text", + "content": "5. Au, H.Y., Chen, J., Jiang, J., Guo, Y.: Choreograph: Music-conditioned automatic dance choreography over a style and tempo consistent dynamic graph. In: ACM International Conference on Multimedia (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 289, + 481, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 289, + 481, + 310 + ], + "spans": [ + { + "bbox": [ + 138, + 289, + 481, + 310 + ], + "type": "text", + "content": "6. Bisig, D.: Generative dance-a taxonomy and survey. In: International Conference on Movement and Computing (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 310, + 481, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 310, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 138, + 310, + 481, + 331 + ], + "type": "text", + "content": "7. Cervantes, P., Sekikawa, Y., Sato, I., Shinoda, K.: Implicit neural representations for variable length human motion generation. In: ECCV (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 331, + 481, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 331, + 481, + 353 + ], + "spans": [ + { + "bbox": [ + 138, + 331, + 481, + 353 + ], + "type": "text", + "content": "8. Chan, C., Ginosar, S., Zhou, T., Efros, A.A.: Everybody dance now. In: ICCV (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 353, + 481, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 353, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 138, + 353, + 481, + 385 + ], + "type": "text", + "content": "9. Chen, K., Tan, Z., Lei, J., Zhang, S.H., Guo, Y.C., Zhang, W., Hu, S.M.: Choreomaster: choreography-oriented music-driven dance synthesis. ACM Transactions on Graphics (TOG) (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 385, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 385, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 138, + 385, + 481, + 418 + ], + "type": "text", + "content": "0. Fan, R., Xu, S., Geng, W.: Example-based automatic music-driven conventional dance motion synthesis. IEEE Transactions on Visualization and Computer Graphics (TVCG) (2011)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 418, + 481, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 418, + 481, + 439 + ], + "spans": [ + { + "bbox": [ + 138, + 418, + 481, + 439 + ], + "type": "text", + "content": "1. Feng, B., Ao, T., Liu, Z., Ju, W., Liu, L., Zhang, M.: Robust dancer: Long-term 3d dance synthesis using unpaired data. arXiv (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 439, + 481, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 439, + 481, + 472 + ], + "spans": [ + { + "bbox": [ + 138, + 439, + 481, + 472 + ], + "type": "text", + "content": "2. Ferreira, J.P., Coutinho, T.M., Gomes, T.L., Neto, J.F., Azevedo, R., Martins, R., Nascimento, E.R.: Learning to dance: A graph convolutional adversarial network to generate realistic dance motions from audio. Computers & Graphics (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 472, + 481, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 472, + 481, + 493 + ], + "spans": [ + { + "bbox": [ + 138, + 472, + 481, + 493 + ], + "type": "text", + "content": "3. Fink, B., Bläsing, B., Ravignani, A., Shackelford, T.K.: Evolution and functions of human dance. Evolution and Human Behavior (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 493, + 481, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 493, + 481, + 514 + ], + "spans": [ + { + "bbox": [ + 138, + 493, + 481, + 514 + ], + "type": "text", + "content": "4. Gartner, E., Metz, L., Andriluka, M., Freeman, C.D., Sminchisescu, C.: Transformer-based learned optimization. In: CVPR (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 514, + 481, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 514, + 481, + 546 + ], + "spans": [ + { + "bbox": [ + 138, + 514, + 481, + 546 + ], + "type": "text", + "content": "5. Ghorbani, S., Ferstl, Y., Holden, D., Troje, N.F., Carbonneau, M.A.: Zeroeggs: Zero-shot example-based gesture generation from speech. In: Computer Graphics Forum (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 547, + 481, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 547, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 138, + 547, + 481, + 568 + ], + "type": "text", + "content": "6. Ghosh, A., Cheema, N., Oguz, C., Theobalt, C., Slusallek, P.: Synthesis of compositional animations from textual descriptions. In: ICCV (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 568, + 481, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 568, + 481, + 590 + ], + "spans": [ + { + "bbox": [ + 138, + 568, + 481, + 590 + ], + "type": "text", + "content": "7. Gong, K., Lian, D., Chang, H., Guo, C., Zuo, X., Jiang, Z., Wang, X.: Tm2d: Bimodality driven 3d dance generation via music-text integration. In: ICCV (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 590, + 481, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 590, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 138, + 590, + 481, + 611 + ], + "type": "text", + "content": "8. Guo, C., Zou, S., Zuo, X., Wang, S., Ji, W., Li, X., Cheng, L.: Generating diverse and natural 3d human motions from text. In: CVPR (2022)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 611, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 481, + 632 + ], + "type": "text", + "content": "9. He, C., Saito, J., Zachary, J., Rushmeier, H., Zhou, Y.: Nemf: Neural motion fields for kinematic animation. NeurIPS 35, 4244-4256 (2022)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 138, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 481, + 666 + ], + "type": "text", + "content": "20. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: NeurIPS (2017)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "text", + "content": "Scalable Group Choreography via Variational Phase Manifold Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "text", + "content": "21. Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural Computation (1997)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 161 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 161 + ], + "type": "text", + "content": "22. Holden, D., Komura, T., Saito, J.: Phase-functional neural networks for character control. ACM Transactions on Graphics (TOG) 36(4), 1-13 (2017)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 162, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 162, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 162, + 480, + 182 + ], + "type": "text", + "content": "23. Holden, D., Saito, J., Komura, T.: A deep learning framework for character motion synthesis and editing. ACM Transactions on Graphics (TOG) (2016)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 183, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 183, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 183, + 480, + 205 + ], + "type": "text", + "content": "24. Holden, D., Saito, J., Komura, T., Joyce, T.: Learning motion manifolds with convolutional autoencoders. In: SIGGRAPH Asia 2015 technical briefs (2015)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 134, + 205, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 205, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 134, + 205, + 480, + 237 + ], + "type": "text", + "content": "25. Hong, F., Zhang, M., Pan, L., Cai, Z., Yang, L., Liu, Z.: Avatarclip: zero-shot text-driven generation and animation of 3d avatars. ACM Transactions on Graphics (TOG) (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 134, + 238, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 238, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 134, + 238, + 480, + 259 + ], + "type": "text", + "content": "26. Huang, R., Hu, H., Wu, W., Sawada, K., Zhang, M.: Dance revolution: Long sequence dance generation with music via curriculum learning. CoRR (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 134, + 260, + 480, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 260, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 134, + 260, + 480, + 304 + ], + "type": "text", + "content": "27. Huang, Y., Zhang, J., Liu, S., Bao, Q., Zeng, D., Chen, Z., Liu, W.: Genre-conditioned long-term 3d dance generation driven by music. In: ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 4858-4862. IEEE (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 134, + 304, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 304, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 134, + 304, + 480, + 335 + ], + "type": "text", + "content": "28. Jiang, C., Cornman, A., Park, C., Sapp, B., Zhou, Y., Anguelov, D., et al.: Motiondiffuser: Controllable multi-agent motion prediction using diffusion. In: CVPR (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 134, + 337, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 337, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 134, + 337, + 480, + 369 + ], + "type": "text", + "content": "29. Jiang, Y., Won, J., Ye, Y., Liu, C.K.: Drop: Dynamics responses from human motion prior and projective dynamics. In: SIGGRAPH Asia 2023 Conference papers (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 134, + 370, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 370, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 134, + 370, + 480, + 391 + ], + "type": "text", + "content": "30. Joshi, M., Chakrabarty, S.: An extensive review of computational dance automation techniques and applications. Proceedings of the Royal Society A (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 134, + 392, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 392, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 134, + 392, + 480, + 413 + ], + "type": "text", + "content": "31. Karunratanakul, K., Preechakul, K., Suwajanakorn, S., Tang, S.: Guided motion diffusion for controllable human motion synthesis. In: ICCV (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 414, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 414, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 134, + 414, + 480, + 435 + ], + "type": "text", + "content": "32. Kico, I., Grammalidis, N., Christidis, Y., Liarokapis, F.: Digitization and visualization of folk dances in cultural heritage: A review. Inventions (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 436, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 436, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 134, + 436, + 480, + 456 + ], + "type": "text", + "content": "33. Kim, J.W., Fouad, H., Hahn, J.K.: Making them dance. In: AAAI Fall Symposium: Aurally Informed Performance (2006)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 457, + 480, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 457, + 480, + 489 + ], + "spans": [ + { + "bbox": [ + 134, + 457, + 480, + 489 + ], + "type": "text", + "content": "34. Kim, J., Oh, H., Kim, S., Tong, H., Lee, S.: A brand new dance partner: Music-conditioned pluralistic dancing controlled by multiple dance genres. In: CVPR (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 491, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 491, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 134, + 491, + 480, + 511 + ], + "type": "text", + "content": "35. Kim, T.h., Park, S.I., Shin, S.Y.: Rhythmic-motion synthesis based on motion-beat analysis. ACM Transactions on Graphics (TOG) (2003)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 512, + 465, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 512, + 465, + 523 + ], + "spans": [ + { + "bbox": [ + 134, + 512, + 465, + 523 + ], + "type": "text", + "content": "36. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. In: ICLR (2014)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 524, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 524, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 134, + 524, + 480, + 544 + ], + "type": "text", + "content": "37. Kolotouros, N., Alldieck, T., Zanfir, A., Bazavan, E., Fieraru, M., Sminchisescu, C.: Dreamhuman: Animatable 3d avatars from text. In: NeurIPS (2024)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 134, + 545, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 545, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 134, + 545, + 480, + 566 + ], + "type": "text", + "content": "38. Lannan, N., Zhou, L., Fan, G.: Human motion enhancement via tobit kalman filter-assisted autoencoder. IEEE Access 10, 29233-29251 (2022)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 134, + 567, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 567, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 134, + 567, + 480, + 599 + ], + "type": "text", + "content": "39. Le, N., Do, T., Do, K., Nguyen, H., Tjiputra, E., Tran, Q.D., Nguyen, A.: Controllable group choreography using contrastive diffusion. ACM Transactions on Graphics (TOG) (2023)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 134, + 600, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 600, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 134, + 600, + 480, + 621 + ], + "type": "text", + "content": "40. Le, N., Pham, T., Do, T., Tjiputra, E., Tran, Q.D., Nguyen, A.: Music-driven group choreography. In: CVPR (2023)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 134, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 134, + 622, + 480, + 643 + ], + "type": "text", + "content": "41. Lee, H.Y., Yang, X., Liu, M.Y., Wang, T.C., Lu, Y.D., Yang, M.H., Kautz, J.: Dancing to music. In: NeurIPS (2019)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 134, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 643, + 480, + 665 + ], + "type": "text", + "content": "42. Lee, L., Lin, Z., Hu, R., Gong, Z., Kumar, A., Li, T., Li, S., Hui, P.: When creators meet the metaverse: A survey on computational arts. CoRR (2021)" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "text", + "content": "N. Le et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "43. Lee, M., Lee, K., Park, J.: Music similarity-based approach to generating dance motion sequence. Multimedia tools and applications (2013)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 480, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 480, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 480, + 159 + ], + "type": "text", + "content": "44. Li, B., Zhao, Y., Sheng, L.: Dancenet3d: Music based dance generation with parametric motion transformer. In: AAAI (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 160, + 480, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 160, + 480, + 191 + ], + "spans": [ + { + "bbox": [ + 133, + 160, + 480, + 191 + ], + "type": "text", + "content": "45. Li, B., Zhao, Y., Zhelun, S., Sheng, L.: Danceformer: Music conditioned 3d dance generation with parametric motion transformer. In: Proceedings of the AAAI Conference on Artificial Intelligence (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 192, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 192, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 133, + 192, + 480, + 224 + ], + "type": "text", + "content": "46. Li, J., Villegas, R., Ceylan, D., Yang, J., Kuang, Z., Li, H., Zhao, Y.: Task-generic hierarchical human motion prior using vaes. In: International Conference on 3D Vision (3DV) (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 225, + 480, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 225, + 480, + 256 + ], + "spans": [ + { + "bbox": [ + 133, + 225, + 480, + 256 + ], + "type": "text", + "content": "47. Li, J., Kang, D., Pei, W., Zhe, X., Zhang, Y., He, Z., Bao, L.: Audio2gestures: Generating diverse gestures from speech audio with conditional variational autoencoders. In: ICCV (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 258, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 258, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 133, + 258, + 480, + 289 + ], + "type": "text", + "content": "48. Li, R., Zhao, J., Zhang, Y., Su, M., Ren, Z., Zhang, H., Tang, Y., Li, X.: Finedance: A fine-grained choreography dataset for 3d full body dance generation. In: ICCV (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 290, + 480, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 290, + 480, + 311 + ], + "spans": [ + { + "bbox": [ + 133, + 290, + 480, + 311 + ], + "type": "text", + "content": "49. Li, R., Yang, S., Ross, D.A., Kanazawa, A.: Ai choreographer: Music conditioned 3d dance generation with aist++. In: ICCV (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 312, + 480, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 312, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 133, + 312, + 480, + 331 + ], + "type": "text", + "content": "50. Li, W., Dai, B., Zhou, Z., Yao, Q., Wang, B.: Controlling character motions without observable driving source. In: WACV (2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 332, + 480, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 332, + 480, + 364 + ], + "spans": [ + { + "bbox": [ + 133, + 332, + 480, + 364 + ], + "type": "text", + "content": "51. Liang, H., Zhang, W., Li, W., Yu, J., Xu, L.: Intergen: Diffusion-based multi-human motion generation under complex interactions. arXiv preprint arXiv:2304.05684 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 365, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 365, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 133, + 365, + 480, + 385 + ], + "type": "text", + "content": "52. Lin, J., Zeng, A., Lu, S., Cai, Y., Zhang, R., Wang, H., Zhang, L.: Motion-x: A large-scale 3d expressive whole-body human motion dataset. NeurIPS (2024)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 386, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 386, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 133, + 386, + 480, + 407 + ], + "type": "text", + "content": "53. Ling, H.Y., Zinno, F., Cheng, G., Van De Panne, M.: Character controllers using motion vaes. ACM Transactions on Graphics (TOG) 39(4), 40-1 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 407, + 480, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 407, + 480, + 429 + ], + "spans": [ + { + "bbox": [ + 133, + 407, + 480, + 429 + ], + "type": "text", + "content": "54. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: SMPL: A skinned multi-person linear model. ACM Transactions on Graphics (TOG) (2015)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 430, + 480, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 430, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 133, + 430, + 480, + 449 + ], + "type": "text", + "content": "55. Martinez, J., Black, M.J., Romero, J.: On human motion prediction using recurrent neural networks. In: CVPR (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 450, + 480, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 450, + 480, + 482 + ], + "spans": [ + { + "bbox": [ + 133, + 450, + 480, + 482 + ], + "type": "text", + "content": "56. Mo, C.A., Hu, K., Long, C., Wang, Z.: Continuous intermediate token learning with implicit motion manifold for keyframe based motion interpolation. In: CVPR (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 483, + 480, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 483, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 133, + 483, + 480, + 514 + ], + "type": "text", + "content": "57. Ofli, F., Erzin, E., Yemez, Y., Tekalp, A.M.: Learn2dance: Learning statistical music-to-dance mappings for choreography synthesis. IEEE Transactions on Multimedia (TMM) (2011)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 515, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 515, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 133, + 515, + 480, + 536 + ], + "type": "text", + "content": "58. Onuma, K., Faloutsos, C., Hodgins, J.K.: Fmdistance: A fast and effective distance function for motion capture data. In: Eurographics (2008)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 536, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 536, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 133, + 536, + 480, + 568 + ], + "type": "text", + "content": "59. Perez, G.V., Beskow, J., Henter, G., Holzapfel, A., Oudeyer, P.Y., Alexanderson, S.: Transflower: probabilistic autoregressive dance generation with multimodal attention. ACM Transactions on Graphics (TOG) (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 133, + 569, + 480, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 569, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 133, + 569, + 480, + 590 + ], + "type": "text", + "content": "60. Petrovich, M., Black, M.J., Varol, G.: Action-conditioned 3d human motion synthesis with transformer vae. In: ICCV (2021)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 133, + 590, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 590, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 133, + 590, + 480, + 611 + ], + "type": "text", + "content": "61. Petrovich, M., Black, M.J., Varol, G.: Temos: Generating diverse human motions from textual descriptions. In: ECCV (2022)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 133, + 612, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 612, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 133, + 612, + 480, + 643 + ], + "type": "text", + "content": "62. Qi, Q., Zhuo, L., Zhang, A., Liao, Y., Fang, F., Liu, S., Yan, S.: Diffdance: Cascaded human motion diffusion model for dance generation. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 1374-1382 (2023)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 133, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 644, + 480, + 665 + ], + "type": "text", + "content": "63. Raab, S., Leibovitch, I., Li, P., Aberman, K., Sorkine-Hornung, O., Cohen-Or, D.: Modi: Unconditional motion synthesis from diverse data. In: CVPR (2023)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "text", + "content": "Scalable Group Choreography via Variational Phase Manifold Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "text", + "content": "64. Ren, X., Li, H., Huang, Z., Chen, Q.: Self-supervised dance video synthesis conditioned on music. In: ACMMM (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "type": "text", + "content": "65. Safonova, A., Hodgins, J.K.: Construction and optimal search of interpolated motion graphs. In: SIGGRAPH (2007)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "type": "text", + "content": "66. Shafir, Y., Tevet, G., Kapon, R., Bermano, A.H.: Human motion diffusion as a generative prior. arXiv (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 183, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 183, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 133, + 183, + 480, + 215 + ], + "type": "text", + "content": "67. Shi, J.: Application of 3d computer aided system in dance creation and learning. In: International Conference on Machine Learning and Big Data Analytics for IoT Security and Privacy (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 216, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 216, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 133, + 216, + 480, + 237 + ], + "type": "text", + "content": "68. Shi, M., Starke, S., Ye, Y., Komura, T., Won, J.: Phasemp: Robust 3d pose estimation via phase-conditioned human motion prior. In: ICCV (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 238, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 238, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 133, + 238, + 480, + 258 + ], + "type": "text", + "content": "69. Shiratori, T., Nakazawa, A., Ikeuchi, K.: Dancing-to-music character animation. Computer Graphics Forum (2006)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 259, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 259, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 133, + 259, + 480, + 281 + ], + "type": "text", + "content": "70. Sitzmann, V., Martel, J., Bergman, A., Lindell, D., Wetzstein, G.: Implicit neural representations with periodic activation functions. In: NeurIPS (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 281, + 480, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 281, + 480, + 313 + ], + "spans": [ + { + "bbox": [ + 133, + 281, + 480, + 313 + ], + "type": "text", + "content": "71. Siyao, L., Yu, W., Gu, T., Lin, C., Wang, Q., Qian, C., Loy, C.C., Liu, Z.: Bailando: 3d dance generation by actor-critic gpt with choreographic memory. In: CVPR (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 315, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 315, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 133, + 315, + 480, + 346 + ], + "type": "text", + "content": "72. Soga, A., Umino, B., Longstaff, J.S.: Automatic composition of ballet sequences using a 3d motion archive. In: 1st South-Eastern European Digitization Initiative Conference (2005)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 347, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 347, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 133, + 347, + 480, + 369 + ], + "type": "text", + "content": "73. Sohn, K., Lee, H., Yan, X.: Learning structured output representation using deep conditional generative models. In: NeurIPS (2015)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 369, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 369, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 133, + 369, + 480, + 391 + ], + "type": "text", + "content": "74. Starke, S., Mason, I., Komura, T.: Deepphase: Periodic autoencoders for learning motion phase manifolds. ACM Transactions on Graphics (TOG) (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 392, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 392, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 133, + 392, + 480, + 413 + ], + "type": "text", + "content": "75. Starke, S., Zhang, H., Komura, T., Saito, J.: Neural state machine for character-scene interactions. ACM Transactions on Graphics (TOG) 38(6), 209-1 (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 414, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 414, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 133, + 414, + 480, + 445 + ], + "type": "text", + "content": "76. Starke, S., Zhao, Y., Komura, T., Zaman, K.: Local motion phases for learning multi-contact character movements. ACM Transactions on Graphics (TOG) 39(4), 54-1 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 446, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 446, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 133, + 446, + 480, + 479 + ], + "type": "text", + "content": "77. Sun, G., Wong, Y., Cheng, Z., Kankanhalli, M.S., Geng, W., Li, X.: Deepdance: Music-to-dance motion choreography with adversarial learning. IEEE Transactions on Multimedia (TMM) (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 479, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 479, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 133, + 479, + 480, + 511 + ], + "type": "text", + "content": "78. Sun, J., Wang, C., Hu, H., Lai, H., Jin, Z., Hu, J.F.: You never stop dancing: Nonfreezing dance generation via bank-constrained manifold projection. In: NeurIPS (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 512, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 512, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 133, + 512, + 480, + 533 + ], + "type": "text", + "content": "79. Tang, T., Jia, J., Mao, H.: Dance with melody: An LSTM-autoencoder approach to music-oriented dance synthesis. In: ACMMM (2018)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 534, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 534, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 133, + 534, + 480, + 566 + ], + "type": "text", + "content": "80. Tevet, G., Gordon, B., Hertz, A., Bermano, A.H., Cohen-Or, D.: Motionclip: Exposing human motion generation to clip space. In: European Conference on Computer Vision (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 133, + 567, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 567, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 133, + 567, + 480, + 588 + ], + "type": "text", + "content": "81. Tevet, G., Raab, S., Gordon, B., Shafir, Y., Cohen-Or, D., Bermano, A.H.: Human motion diffusion model. In: ICLR (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 133, + 589, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 589, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 133, + 589, + 480, + 610 + ], + "type": "text", + "content": "82. Tiwari, G., Antic, D., Lenssen, J.E., Sarafianos, N., Tung, T., Pons-Moll, G.: Posendf: Modeling human pose manifolds with neural distance fields. In: ECCV (2022)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 133, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 133, + 611, + 480, + 632 + ], + "type": "text", + "content": "83. Tseng, J., Castellon, R., Liu, C.K.: Edge: Editable dance generation from music. In: CVPR (2023)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 133, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 632, + 480, + 665 + ], + "type": "text", + "content": "84. Tsuchida, S., Fukayama, S., Hamasaki, M., Goto, M.: Aist dance video database: Multi-genre, multi-dancer, and multi-camera database for dance information processing. In: ISMIR (2019)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "text", + "content": "N. Le et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 512 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "text", + "content": "85. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 171 + ], + "type": "text", + "content": "86. Wang, Z., Jia, J., Wu, H., Xing, J., Cai, J., Meng, F., Chen, G., Wang, Y.: Group-dancer: Music to multi-people dance synthesis with style collaboration. In: ACM International Conference on Multimedia (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 480, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 480, + 193 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 480, + 193 + ], + "type": "text", + "content": "87. Yalta, N., Watanabe, S., Nakadai, K., Ogata, T.: Weakly-supervised deep recurrent neural networks for basic dance step generation. In: IJCNN (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "text", + "content": "88. Yang, S., Wu, Z., Li, M., Zhang, Z., Hao, L., Bao, W., Zhuang, H.: Qpgesture: Quantization-based and phase-guided motion matching for natural speech-driven gesture generation. In: CVPR (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "text", + "content": "89. Yang, Z., Wen, Y.H., Chen, S.Y., Liu, X., Gao, Y., Liu, Y.J., Gao, L., Fu, H.: Keyframe control of music-driven 3d dance generation. IEEE Transactions on Visualization and Computer Graphics (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "type": "text", + "content": "90. Yao, S., Sun, M., Li, B., Yang, F., Wang, J., Zhang, R.: Dance with you: The diversity controllable dancer generation via diffusion models. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 8504-8514 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "type": "text", + "content": "91. Ye, Z., Wu, H., Jia, J., Bu, Y., Chen, W., Meng, F., Wang, Y.: Choreonet: Towards music to dance synthesis with choreographic action unit. In: ACMMM (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "type": "text", + "content": "92. Yin, W., Yin, H., Baraka, K., Kragic, D., Björkman, M.: Dance style transfer with cross-modal transformer. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 5058-5067 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "type": "text", + "content": "93. Yu, Z., Cheng, W., Liu, X., Wu, W., Lin, K.Y.: Monohuman: Animatable human neural field from monocular video. In: CVPR (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "type": "text", + "content": "94. Zhang, X., Yang, S., Xu, Y., Zhang, W., Gao, L.: Mining and applying composition knowledge of dance moves for style-concentrated dance generation. In: AAAI (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "type": "text", + "content": "95. Zhou, L., Luo, Y.: A spatio-temporal learning for music conditioned dance generation. In: International Conference on Multimodal Interaction (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 413, + 480, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 480, + 434 + ], + "type": "text", + "content": "96. Zhou, Y., Barnes, C., Lu, J., Yang, J., Li, H.: On the continuity of rotation representations in neural networks. In: CVPR (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "type": "text", + "content": "97. Zhou, Z., Wang, B.: Ude: A unified driving engine for human motion generation. In: CVPR (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "type": "text", + "content": "98. Zhu, Y., Olszewski, K., Wu, Y., Achlioptas, P., Chai, M., Yan, Y., Tulyakov, S.: Quantized gan for complex music generation from dance videos. In: ECCV (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 479, + 480, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 512 + ], + "type": "text", + "content": "99. Zhuang, W., Wang, C., Chai, J., Wang, Y., Shao, M., Xia, S.: Music2dance: Dancenet for music-driven dance generation. ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM) (2022)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "text", + "content": "Scalable Group Choreography via Variational Phase Manifold Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Scalar Function Topology Divergence_ Comparing Topology of 3D Objects/228bb063-0fe3-438c-a000-0e252dd17cc3_content_list.json b/2024/Scalar Function Topology Divergence_ Comparing Topology of 3D Objects/228bb063-0fe3-438c-a000-0e252dd17cc3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..108e1151ca39af7daf428e1f26b3de325b32d63b --- /dev/null +++ b/2024/Scalar Function Topology Divergence_ Comparing Topology of 3D Objects/228bb063-0fe3-438c-a000-0e252dd17cc3_content_list.json @@ -0,0 +1,2399 @@ +[ + { + "type": "text", + "text": "Scalar Function Topology Divergence: Comparing Topology of 3D Objects", + "text_level": 1, + "bbox": [ + 218, + 140, + 782, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ilya Trofimov1, Daria Voronkova1,2, Eduard Tulchinskii1,4, Evgeny Burnaev1,2, and Serguei Barannikov1,3", + "bbox": [ + 254, + 210, + 746, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Skolkovo Institute of Science and Technology, Moscow, Russia {ilya.trofimov,darya.voronkova, eduard.tulchinskiy,e.burnaev, s.barannikov}@skoltech.ru", + "bbox": [ + 267, + 253, + 730, + 295 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 AIRI, Moscow, Russia", + "bbox": [ + 419, + 296, + 583, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ CNRS, IMJ, Paris Cité University, France", + "bbox": [ + 354, + 309, + 648, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 AI Foundation and Algorithm Lab, Moscow, Russia", + "bbox": [ + 321, + 324, + 679, + 337 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. We propose a new topological tool for computer vision - Scalar Function Topology Divergence (SFTD), which measures the dissimilarity of multi-scale topology between sublevel sets of two functions having a common domain. Functions can be defined on an undirected graph or Euclidean space of any dimensionality. Most of the existing methods for comparing topology are based on Wasserstein distance between persistence barcodes and they don't take into account the localization of topological features. The minimization of SFTD ensures that the corresponding topological features of scalar functions are located in the same places. The proposed tool provides useful visualizations depicting areas where functions have topological dissimilarities. We provide applications of the proposed method to 3D computer vision. In particular, experiments demonstrate that SFTD as an additional loss improves the reconstruction of cellular 3D shapes from 2D fluorescence microscopy images, and helps to identify topological errors in 3D segmentation. Additionally, we show that SFTD outperforms Betti matching loss in 2D segmentation problems. The code is publicly available: https://github.com/IlyaTrofimov/SFTD.", + "bbox": [ + 261, + 373, + 738, + 622 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Topological Data Analysis $\\cdot$ 3D Computer Vision $\\cdot$ Geometric Deep Learning", + "bbox": [ + 259, + 635, + 740, + 664 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 688, + 375, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A common situation in computer vision is a structural prediction defined over $\\mathbb{R}^n$ . Examples include image segmentation [24, 25, 31, 32] and 3D shape reconstruction [36, 37, 39] problems. However, in most of the cases the comparison between the prediction and a ground truth is done pixel(voxel)-wise by metrics like Dice score, cross-entropy, etc. These metrics are not capable of providing detailed topological correctness, i.e. preserving multi-scale shape patterns: cavities, holes, connectivity patterns, etc. Typically, predictions involve scalar functions $f:\\mathbb{R}^n\\to \\mathbb{R}$ and the final shape of an object is defined by thresholding $f(x)$ .", + "bbox": [ + 212, + 719, + 787, + 842 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "There is a need for topological-aware comparison metrics for scalar functions, which take into account the topology of sublevel sets $f^{-1}((-\\infty, \\varepsilon])$ (the analysis for superlevel sets is similar). Existing tools for this problem don't take into account the localization of topological features [24,36] or are limited to 2D space [32]. Our contributions are as follows:", + "bbox": [ + 212, + 146, + 787, + 223 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We introduce new topological tools: F-Cross-Barcodes and Scalar Function Topology Divergence (SFTD), for a multi-scale topology-aware comparison of scalar functions by tracking differences in their sublevel sets topology.", + "2. A distinctive feature of F-Cross-Barcodes and SFTD is that they take into account the localization of topological features;", + "3. We describe a visualization technique for F-Cross-Barcodes: a technique highlights with points areas where differences in topological features are localized:", + "4. We propose an algorithm for SFTD differentiation;", + "5. By doing computational experiments, we show that SFTD as an additional loss term improves 3D shape reconstruction, and helps to identify topological errors in 3D segmentation.", + "6. We also show that SFTD outperforms Betti matching loss in terms of topological accuracy in 2D segmentation while being $\\approx 35$ times faster, see Appendix G." + ], + "bbox": [ + 220, + 233, + 782, + 457 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Related work", + "text_level": 1, + "bbox": [ + 215, + 482, + 382, + 498 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Tools of topological data analysis find application in various machine learning problems: evaluation of deep generative models [8], comparison of neural representations [7], dimensionality reduction [29,34], graph neural networks [12,23,40], self-supervised learning [27], molecular data analysis [16], analysis of loss functions in deep learning [6,9], spatial-temporal graph neural network tasks [14,15], crowd localization [1], to name a few.", + "bbox": [ + 212, + 513, + 782, + 603 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Several works used topological tools to enhance image segmentation [20, 24, 25, 32]. However, proposed algorithms for comparison ground truth and predicted images either don't take into account localization of topological features [24], or are limited to 2D [32]. A method by [25] highlights in images topologically critical and error-prone structures using discrete Morse theory, and [20] proposes a method for inferring structure-aware uncertainty. The importance of topological interactions for multi-class medical image segmentation is shown in [19].", + "bbox": [ + 212, + 604, + 782, + 709 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 732, + 380, + 750 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We study scalar functions $f: \\mathbb{R}^n \\to \\mathbb{R}$ by assessing the topology of its sublevel sets. Consider a 1D function depicted in Figure 1. It has two local minima $y_1, y_2$ and a local maximum $y_3$ , such that $y_1 < y_2 < y_3$ . One can study sublevel sets", + "bbox": [ + 214, + 763, + 573, + 840 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3dd7950c826053b97be8ccddecdbe1c53dfc7631ee8e4200b6be564b361423e3.jpg", + "image_caption": [ + "Fig.1: 1D function and its sublevel sets." + ], + "image_footnote": [], + "bbox": [ + 586, + 733, + 782, + 829 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "I. Trofimov et al.", + "bbox": [ + 271, + 114, + 387, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "$X_{\\varepsilon} = f^{-1}((-\\infty ,\\varepsilon ])$ . For the scalar function $f$ , changes of $X_{\\varepsilon}$ for increasing $\\varepsilon$ can be intuitively understood as a \"filling the pool\", see Figure 1. There are four distinct situations:", + "bbox": [ + 212, + 145, + 782, + 191 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. For $\\varepsilon < y_{1}$ , $X_{\\varepsilon} = \\varnothing$", + "2. For $y_{1} \\leq \\varepsilon < y_{2}$ , $X_{\\varepsilon}$ is a single segment;", + "3. For $y_{2} \\leq \\varepsilon < y_{3}$ , $S_{\\varepsilon}$ consists of 2 segments;", + "4. Finally, for $\\varepsilon \\geq y_{3}$ , $X_{\\varepsilon}$ is a single segment." + ], + "bbox": [ + 220, + 204, + 550, + 266 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Note, that changes in topology of $X_{\\varepsilon}$ happen when $\\varepsilon$ is a critical point of $f$ . In the general case, smooth scalar functions defined on smooth manifolds are studied by Morse theory [6]. In practice, we assess functions through their values on a finite set of points, e.g. a lattice (square grid graph) in $\\mathbb{R}^n$ .", + "bbox": [ + 212, + 277, + 784, + 339 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1 Scalar functions defined on a graph", + "text_level": 1, + "bbox": [ + 214, + 364, + 552, + 381 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the most general form, we are interested in topology-aware comparison of scalar functions $f: \\mathcal{V} \\to \\mathbb{R}$ defined on vertices $\\mathcal{V}$ of an undirected graph $\\mathcal{G} = (\\mathcal{V}, \\mathcal{E})$ . Specifically, we study topological features of subgraphs $\\{v \\in \\mathcal{V} \\mid f(v) \\leq \\varepsilon\\}$ connected components, cycles, etc.", + "bbox": [ + 212, + 393, + 784, + 453 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For each clique $C$ in the graph $\\mathcal{G}$ , we define a filtration function $T(C) = \\max_{i\\in C}f(i)$ . By a clique complex $\\mathcal{C}$ , we mean a set of cliques where each sub-clique also belongs to a complex5. For some $\\varepsilon$ , $T(C)$ defines a clique complex $\\mathcal{C}_j = \\{C\\mid T(C)\\leq \\varepsilon \\}$ , where $f(j) = \\varepsilon$ . Topological features for a subgraph $\\{v\\in \\mathcal{V}\\mid f(v)\\leq \\varepsilon \\}$ are precisely expressed by homology groups. Overall, the filtration function $T(C)$ defines the lower-star filtration of the graph's clique complex:", + "bbox": [ + 212, + 454, + 784, + 559 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\varnothing \\subseteq \\mathcal {C} _ {0} \\subseteq \\mathcal {C} _ {1} \\subseteq \\dots \\subseteq \\mathcal {C} _ {n},\n$$\n", + "text_format": "latex", + "bbox": [ + 415, + 561, + 588, + 577 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{C}_n$ is a clique complex of $\\mathcal{G}$ . The filtration is said to be induced by the vertex function $f$ .", + "bbox": [ + 212, + 585, + 782, + 616 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A given topological feature \"appears\" at some $\\varepsilon_{1}$ and \"dies\" at some $\\varepsilon_{2}$ . These topological features, together with their birth and death times can be calculated via tools of topological data analysis. The multi-set of the corresponding intervals $(\\varepsilon_{1},\\varepsilon_{2})$ is called the persistence barcode, which depicts topological features of the graph for all scales. The whole theory is dubbed persistent homology [5, 13].", + "bbox": [ + 212, + 617, + 784, + 694 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4 Method", + "text_level": 1, + "bbox": [ + 214, + 719, + 330, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1 Scalar functions defined on a graph", + "text_level": 1, + "bbox": [ + 214, + 753, + 553, + 771 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To compare topological features of sublevel sets of functions $f,g:\\mathcal{G}\\to \\mathbb{R}$ we construct a graph $\\tilde{\\mathcal{G}} = (\\tilde{\\mathcal{V}},\\tilde{\\mathcal{E}})$ with doubled set of vertices from $\\mathcal{G}$ and one extra", + "bbox": [ + 212, + 782, + 784, + 813 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Scalar Function Topology Divergence", + "bbox": [ + 480, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "Every clique complex is an abstract simplicial complex, but the reserve is not true.", + "bbox": [ + 217, + 824, + 782, + 839 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ebe692964d3d7b3377c6e58b57b047b2eb36ef19282a72aac14b2a4969134bbc.jpg", + "image_caption": [ + "(a) The doubled graph $\\tilde{\\mathcal{G}}$" + ], + "image_footnote": [], + "bbox": [ + 294, + 146, + 421, + 276 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/747418fb38f3eac499b0fe55b32d5275d37f303be1480fb7a412b53b5d734265.jpg", + "image_caption": [ + "(b) The extended $n + 1$ dimensional lattice $\\tilde{\\mathcal{L}}$ .", + "Fig. 2: A visualization of $\\tilde{\\mathcal{G}}$ , $\\tilde{\\mathcal{L}}$ ." + ], + "image_footnote": [], + "bbox": [ + 566, + 178, + 714, + 275 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "vertex $O$ , see Figure 2a. Formally, $\\tilde{\\mathcal{V}} = \\{A_1,\\dots ,A_n,A_1',\\dots ,A_n',O\\}$ . Edges $\\tilde{\\mathcal{E}}$ are defined as follows:", + "bbox": [ + 212, + 349, + 784, + 380 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\left(A _ {i}, A _ {j}\\right) \\in \\mathcal {E} \\Rightarrow \\left(A _ {i}, A _ {j}\\right) \\in \\tilde {\\mathcal {E}}, \\left(A _ {i} ^ {\\prime}, A _ {j} ^ {\\prime}\\right) \\in \\tilde {\\mathcal {E}} \\\\ \\left(A _ {i}, A _ {j}\\right) \\in \\mathcal {E}, i < j \\Rightarrow \\left(A _ {i}, A _ {j} ^ {\\prime}\\right) \\in \\tilde {\\mathcal {E}} \\\\ \\left(O, A _ {i}\\right) \\in \\tilde {\\mathcal {E}}, \\left(A _ {i}, A _ {i} ^ {\\prime}\\right) \\in \\tilde {\\mathcal {E}} \\text {f o r a l l} i \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 349, + 388, + 650, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The vertex function $\\tilde{f}$ is defined as follows:", + "bbox": [ + 214, + 462, + 527, + 478 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\tilde {f} (A _ {i}) = f (i) \\\\ \\tilde {f} (A _ {i} ^ {\\prime}) = \\min (f (i), g (i)), \\tilde {f} (O) = \\min _ {f, g}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 359, + 487, + 643, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\min_{f,g}$ is the minimum value of both $f,g$ in the nodes of the graph $\\mathcal{G}$ .", + "bbox": [ + 212, + 537, + 763, + 553 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Definition. The $F$ -Cross-Barcode $_k(f, g)$ for two scalar functions $f, g$ defined on vertices of an undirected graph $\\mathcal{G}$ is the $k$ -th persistence barcode for the doubled graph $\\tilde{\\mathcal{G}}$ and the lower-star filtration induced by $\\tilde{f}$ .", + "bbox": [ + 212, + 553, + 784, + 597 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To put it simply, each interval in the $F$ -Cross-Barcode $_k(f, g)$ denotes a range of functions' sub-level sets where a given topological feature exists in the function $f$ but doesn't exist in the function $g$ , see further Appendix A. Technically, the desired persistence barcode is calculated via giotto-ph library [30]. For this software, values of a filtration function must be defined on vertices and edges. Algorithm 1 summarizes the aforementioned.", + "bbox": [ + 212, + 598, + 785, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2 Scalar functions defined on a lattice", + "text_level": 1, + "bbox": [ + 214, + 709, + 558, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A frequent situation is a function having a domain in $\\mathbb{R}^n$ . We assume that the function is approximated by its restriction on a sufficiently small grid. Algorithm 1 can also be applied in this case after a space triangulation. However, the number of simplices grows very quickly in high dimensions (see discussion in [35]), and Algorithm 1 has a high computational cost. That is why, we consider cubical complexes [26] instead of simplicial complexes. Specifically, we consider a regular square grid graph ( $n$ -dimensional lattice) and use values of functions", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "I. Trofimov et al.", + "bbox": [ + 269, + 114, + 387, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 F-Cross-Barcode $_k$ (scalar function $k$ -th cross-barcode)", + "text_level": 1, + "bbox": [ + 215, + 146, + 718, + 161 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Input: Scalar functions $f, g$ defined on vertices of a graph $\\mathcal{G} = (\\mathcal{V}, \\mathcal{E})$ . $|\\mathcal{V}| = n$ .", + "bbox": [ + 215, + 166, + 738, + 180 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Require: $\\mathsf{PB}_k(\\cdot)$ - an algorithm to calculate $k$ -th persistence barcode from a matrix containing values of a filtration function on graph: on vertices - on a diagonal, on edges - off a diagonal.", + "bbox": [ + 215, + 181, + 785, + 222 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Create a $n \\times n$ matrix $F$ for the function $f$ : $F_{ij} = f(i)$ for $i = j$ , $\\max(f(i), f(j))$ for $i \\neq j$ and $(i, j) \\in \\mathcal{E}$ , $+\\infty$ otherwise.", + "2. Create a $n \\times n$ matrix $G$ for the function $g$ analogously;", + "3. Initialize $c_{+\\infty}$ to a vector of $n$ elements, all equal to $+\\infty$ ;", + "4. Create $F_{+}$ , a matrix derived from $F$ with a lower-diagonal part replaced by $+\\infty$ ;", + "5. Create $F_{diag}$ , a vector containing the diagonal elements of $F$ ;", + "6. Compute $\\min (F,G)$ , the element-wise minimum of $F$ and $G$ ;", + "7. Compute $\\min_{F,G}$ , the minimum value of all elements in $F$ and $G$ ;", + "8. Compute the $(2n + 1)\\times (2n + 1)$ block matrix $m$ as follows:" + ], + "bbox": [ + 230, + 227, + 779, + 349 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nm \\leftarrow \\left( \\begin{array}{c c c} \\min (F, G) & F _ {+} ^ {\\intercal} & c _ {+ \\infty} ^ {\\intercal} \\\\ F _ {+} & F & F _ {d i a g} ^ {\\intercal} \\\\ c _ {+ \\infty} & F _ {d i a g} & \\min _ {F, G} \\end{array} \\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 359, + 616, + 405 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "9. Compute: F-Cross-Barcode $_k(f, g) \\gets \\mathsf{PB}_k(m)$", + "bbox": [ + 228, + 420, + 544, + 434 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Return: F-Cross-Barcode $_k(f, g)$ , a multi-set of (\"birth\", \"death\") intervals.", + "bbox": [ + 215, + 438, + 710, + 453 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "in nodes of this grid: $f(j_{1},\\ldots ,j_{n}),g(j_{1},\\ldots ,j_{n})$ . To compare two function $f,g$ we create an extended $n + 1$ dimensional lattice $\\tilde{\\mathcal{L}}$ with one extra dimension of size 3, see Figure 2b. The vertex function $\\tilde{f}:\\tilde{\\mathcal{L}}\\to \\mathbb{R}$ is defined as follows:", + "bbox": [ + 212, + 487, + 785, + 532 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {f} (i, j _ {1}, \\ldots , j _ {n}) = \\left\\{ \\begin{array}{l l} \\min _ {f, g} & \\text {f o r} i = 0 \\\\ f (j _ {1}, \\ldots , j _ {n}) & \\text {f o r} i = 1 \\\\ \\min (f (j _ {1}, \\ldots , j _ {n}), g (j _ {1}, \\ldots , j _ {n})) & \\text {f o r} i = 2 \\end{array} \\right.\n$$\n", + "text_format": "latex", + "bbox": [ + 279, + 544, + 718, + 601 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similarly to functions on graphs, $F$ -Cross-Barcode $_k(f,g)$ for functions on lattices is defined as the $k$ -th persistence barcode of the filtered cubical complex with the filtration induced by the vertex function $\\tilde{f}$ . We use the GUDHI $^6$ library which implements efficient algorithms for cubical complexes [35]. Note that $F$ -Cross-Barcode $_k(f,g)$ for functions restricted on a lattice and a triangulation of the same space are not equivalent, but would be similar.", + "bbox": [ + 212, + 613, + 787, + 704 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3 Scalar Function Topology Divergence", + "text_level": 1, + "bbox": [ + 215, + 728, + 568, + 743 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The $F$ -Cross-Barcode $_k(f, g)$ depicts topological differences of functions' sublevel sets at multiple scales. The $F$ -Cross-Barcode $_k(f, g)$ , as a regular barcode, is a multi-set of intervals. One requires some vectorization technique to inject it into a machine learning pipeline. If one needs to describe a discrepancy between $f, g$", + "bbox": [ + 212, + 753, + 787, + 816 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Scalar Function Topology Divergence", + "bbox": [ + 480, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "$^{6}$ https://gudhi.inria.fr/", + "bbox": [ + 217, + 824, + 413, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "by an integral numerical characteristic, we propose a scalar function topology divergence:", + "bbox": [ + 212, + 146, + 787, + 176 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {S F T D} _ {k} (f, g) = \\sum_ {(b, d) \\in \\mathrm {F} - \\text {C r o s s - B a r c o d e} _ {k} (f, g)} | | d - b | | ^ {p}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 325, + 186, + 785, + 220 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "that is the sum of lengths of intervals from the $F$ -Cross-Barcode $_k(f, g)$ to $p$ -th power, where typically $p = 1, 2$ . SFTD is not symmetric, and we often use an average of $\\mathrm{SFTD}_k(f, g)$ and $\\mathrm{SFTD}_k(g, f)$ in experiments below. For different dimensions $k$ , $\\mathrm{SFTD}_k(f, g)$ can be used either separately, or the sum can be taken to reflect discrepancies in all types of topological features (connected components, cycles, voids, etc.). Computational complexity of SFTD is discussed in Appendix C.", + "bbox": [ + 212, + 229, + 787, + 338 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4 SFTD basic properties", + "text_level": 1, + "bbox": [ + 215, + 358, + 450, + 375 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Proposition 1. If $SFTD_{k}(f,g) = SFTD_{k}(g,f) = 0$ for all $k\\geq 0$ , then the sublevel sets barcodes of functions $f$ and $g$ are the same in any degree. Moreover, in this case their topological features are located in the same places: the inclusions $f^{-1}((- \\infty ,\\varepsilon ])\\subseteq (\\min (f,g))^{-1}((- \\infty ,\\varepsilon ])$ , $g^{-1}((- \\infty ,\\varepsilon ])\\subseteq (\\min (f,g))^{-1}((- \\infty ,\\varepsilon ])$ , induce homology isomorphisms for any threshold $\\varepsilon$ .", + "bbox": [ + 214, + 382, + 789, + 458 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Proposition 2. (SFTD Stability)", + "text_level": 1, + "bbox": [ + 215, + 468, + 483, + 483 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(a) For any scalar functions $f, \\tilde{f}, g, \\tilde{g}$ on a grid, or on a graph vertex set:", + "bbox": [ + 240, + 484, + 777, + 500 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} d _ {B} (F - C r o s s - B a r c o d e _ {k} (f, g), F - C r o s s - B a r c o d e _ {k} \\left(f ^ {\\prime}, g ^ {\\prime}\\right)) \\leq \\\\ \\max (\\max _ {i} | f (i) - f ^ {\\prime} (i) |, \\max _ {i} | g (i) - g ^ {\\prime} (i) |). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 230, + 505, + 767, + 545 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(b) For any pair of scalar functions $f, \\hat{f}'$ :", + "bbox": [ + 240, + 542, + 544, + 558 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\| F \\text {- C r o s s - B a r c o d e} _ {k} (f, f ^ {\\prime}) \\| _ {B} \\leq \\max _ {i} | f (i) - f ^ {\\prime} (i) |,\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 563, + 674, + 585 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $d_B$ denotes the bottleneck distance between persistence barcodes [13]. We give the proofs in Appendix A. As a consequence, SFTD is robust to noise.", + "bbox": [ + 214, + 599, + 784, + 631 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.5 Visualization of F-Cross-Barcodes", + "text_level": 1, + "bbox": [ + 215, + 652, + 542, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$F$ -Cross-Barcodes can be used for localization and visualization of topological differences between scalar functions. Recall the filtered cubical7 complex:", + "bbox": [ + 214, + 676, + 785, + 708 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\varnothing \\subseteq \\mathcal {C} _ {0} \\subseteq \\mathcal {C} _ {1} \\subseteq \\dots \\subseteq \\mathcal {C} _ {n}.\n$$\n", + "text_format": "latex", + "bbox": [ + 411, + 718, + 588, + 734 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Since the filtration is induced by a vertex function $\\tilde{f}:\\tilde{\\mathcal{L}}\\to \\mathbb{R}$ , the filtered cubical complex corresponds to the sequence:", + "bbox": [ + 212, + 744, + 785, + 776 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {f} (i _ {0}) \\leq \\tilde {f} (i _ {1}) \\leq \\dots \\leq \\tilde {f} (i _ {n}).\n$$\n", + "text_format": "latex", + "bbox": [ + 398, + 784, + 601, + 804 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "I. Trofimov et al.", + "bbox": [ + 269, + 114, + 387, + 127 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "For ease of presentation, we discuss only the visualization for cubical complexes. The visualization for clique complexes is analogous.", + "bbox": [ + 217, + 810, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "If $\\tilde{f}(i_k) = \\tilde{f}(i_m)$ , the order is defined by indexes $i_k, i_m$ . Thus, each segment in the barcode - a birth/death pair - is associated with two vertices $v_b, v_d \\in \\tilde{\\mathcal{L}}$ . Recall that $\\tilde{\\mathcal{L}}$ is an extended $n + 1$ dimensional lattice (see Section 4.2), where the first dimension is of size 3, and the rest of dimensions are of the same size as the lattice inside the domain of $f, g$ . So, neglecting the first dimension, we can map the rest of indices $j_1, \\ldots, j_n$ of birth/death vertices $v_b, v_d$ right onto original $n$ -dimensional lattice. Later, in the experimental section, we shall provide such visualizations, where birth/death events are depicted by orange/red dots.", + "bbox": [ + 212, + 145, + 782, + 267 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.6 SFTD differentiation", + "text_level": 1, + "bbox": [ + 214, + 287, + 433, + 301 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Calculation of $\\mathrm{SFTD}_k(f,g)$ involves a filtered simplicial(cubical) complex $\\mathcal{C}$ with a vertex function $\\tilde{f} (i),i\\in \\tilde{\\mathcal{V}}$ . Also, since we apply the sub-level set filtration, for every simplex(cube) $\\sigma \\colon T(\\sigma) = \\tilde{f} (i_{\\sigma})$ , where $i_{\\sigma} = \\operatorname{argmax}_{i\\in \\sigma}\\tilde{f} (i)$ . Each interval $(b,d)$ in a $F$ -Cross-Barcode corresponds to some pair of simplices(cubes) $(\\sigma_1,\\sigma_2)$ which join the simplicial(cubical) complex $\\mathcal{C}$ at moments $b = T(\\sigma_1),d = T(\\sigma_2)$ Generally, we consider $f,g$ (and consequently $\\tilde{f}$ ) to be parametric functions of $w$ , for example, neural networks. By differentiating (1) we obtain:", + "bbox": [ + 212, + 309, + 784, + 415 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial \\operatorname {S F T D} _ {k} (f , g)}{\\partial w} = \\sum_ {\\left(\\sigma_ {1}, \\sigma_ {2}\\right) \\in \\Sigma_ {k}} p \\left(\\frac {\\partial \\tilde {f} \\left(i _ {\\sigma_ {2}}\\right)}{\\partial w} - \\frac {\\partial \\tilde {f} \\left(i _ {\\sigma_ {2}}\\right)}{\\partial w}\\right) ^ {p - 1}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 303, + 425, + 784, + 469 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\Sigma_{k}$ is a collection of \"birth\", \"death\" simplex pairs corresponding to intervals in the $F$ -Cross-Barcode $_k(f, g)$ .", + "bbox": [ + 212, + 477, + 782, + 507 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "By construction, $\\tilde{f}(i)$ can be either $f(i')$ or $g(i')$ for some mapping $i' = r(i)$ . Formally, introducing a binary function $c(\\cdot)$ for this choice:", + "bbox": [ + 212, + 507, + 782, + 537 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {f} (i) = c (i) f (r (i)) + (1 - c (i)) g (r (i)), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 370, + 546, + 784, + 564 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and", + "bbox": [ + 214, + 571, + 246, + 584 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial \\tilde {f} (i)}{\\partial w} = b (i) \\frac {\\partial f (r (i))}{\\partial w} + (1 - b (i)) \\frac {\\partial g (r (i))}{\\partial w}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 584, + 784, + 614 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Finally, the desirable gradient $\\partial \\mathrm{SFTD}_k(f,g) / \\partial w$ can be calculated by combining (2), (4) via a chain rule. With a common assumption that all the $f(i),g(i)$ are distinct, functions $c,r$ don't change in a vicinity of $w$ , and $\\mathrm{SFTD}_i(f,g)$ is differentiable w.r.t. $w$ [11, 22]. On a high level, the differentiation is done similarly to [11, 34].", + "bbox": [ + 212, + 618, + 782, + 694 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 214, + 714, + 375, + 732 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the experimental section, we are going to answer three research questions:", + "bbox": [ + 212, + 744, + 767, + 758 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "Q1: Does SFTD and F-Cross-Barcodes provide a better topologically-aware comparison of scalar functions than Wasserstein distance between regular persistence barcodes?", + "Q2: Does F-Cross-Barcodes help to visualize and localize topological differences?", + "Q3: Is SFTD, as a loss function, beneficial for computer vision problems?" + ], + "bbox": [ + 205, + 765, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Scalar Function Topology Divergence", + "bbox": [ + 480, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d2ab50479442e57f66504be1fe8b88937b063abfda85749b953715ee9645e657.jpg", + "image_caption": [ + "(a) Scalar func. $f_{1}$ ." + ], + "image_footnote": [], + "bbox": [ + 220, + 147, + 398, + 261 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b70f9280df963f3a824aa917709274b1ea09aeb08e0da3354e1cb892d49e6bbe.jpg", + "image_caption": [ + "(b) Scalar func. $f_{2}$ ." + ], + "image_footnote": [], + "bbox": [ + 421, + 148, + 599, + 261 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f3d2c1e646d9b98fe2032cb8621e3a02038d4db259739dae6c2b4765898ea058.jpg", + "image_caption": [ + "(c) Vis. diff." + ], + "image_footnote": [], + "bbox": [ + 630, + 146, + 782, + 261 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a754ee879426c914d789658c3eb148241a5dc345ad01d9c3cc564e6bda99edd1.jpg", + "image_caption": [ + "(d) Pers. barc. of $f_{1}$ ." + ], + "image_footnote": [], + "bbox": [ + 222, + 281, + 398, + 344 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b31e3284c1626ff870a0a5473ad4c7d58819b22497ea9e2786940ecaf07fbcd3.jpg", + "image_caption": [ + "(e) Pers. barc. of $f_{2}$ ." + ], + "image_footnote": [], + "bbox": [ + 413, + 281, + 588, + 344 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a6d5b7d1918b0c711b9b57a328283c4fd0a3d493a5ebbfb3fced4922e98715b9.jpg", + "image_caption": [ + "(f) F-Cross-Barcode", + "Fig. 3: Two scalar functions (a), (b) are indistinguishable by their persistence barcodes (d), (e). At the same time, $\\mathrm{SFTD}_0(f_1,f_2) = 4.6$ see (f)." + ], + "image_footnote": [], + "bbox": [ + 607, + 281, + 782, + 344 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.1 Synthetic data", + "text_level": 1, + "bbox": [ + 214, + 428, + 382, + 443 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To illustrate the performance of the proposed tools, we start with synthetic data. Functions under comparison have coinciding persistence barcodes of all degrees and they are indistinguishable by standard Wasserstein matching.", + "bbox": [ + 212, + 450, + 784, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Functions with local minima in 2D. First, we compare two scalar function on $64 \\times 64$ grid, each of them having 3 local minima but in different locations, see Figure 3, (a), (b). Persistence barcodes of both functions are equal, see Figure 3, (d), (e): they contain two segments from $H_0$ corresponding to local minima. Here and further, essential segments (which die at $+\\infty$ ) are omitted. The topology of the two functions is indistinguishable by standard persistence barcodes and their Wasserstein distance is zero. At the same time, F-Cross-Barcode between these scalar functions is non-empty for $H_0$ , see Figure 3, (f). The visualization technique correctly points at places with different topology - the centers and borders of the distinguishing local minima.", + "bbox": [ + 212, + 513, + 787, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Lattices in 2D. Next, we compare two lattice-style 2D scalar functions, see Figure 4, (a), (b). Lattices have defects in different places (top left and bottom right respectively). The persistence barcodes of both lattices are identical, see Figure 4, (d), (e): they contain 8 segments from $H_{1}$ corresponding to independent cycles. At the same time, $F$ -Cross-Barcode between two scalar functions is nonempty, see Figure 4, (f). The visualization technique correctly points at places with defects.", + "bbox": [ + 212, + 684, + 787, + 789 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Concentric spheres in 3D. Finally, we study the performance of the proposed method by comparing 3D scalar functions. We compare two scalar functions,", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "I. Trofimov et al.", + "bbox": [ + 269, + 114, + 387, + 126 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/69856d4fcb13b7dd2e2d930a3e5643239ccc2db98283f567f6541dd3a2580df2.jpg", + "image_caption": [ + "(a) Scalar func. $f_{1}$ ." + ], + "image_footnote": [], + "bbox": [ + 222, + 146, + 395, + 258 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/0d5fdc209618e10ffc645065dd5c4b756addc53b05a881ca4922b28ffd87d1bb.jpg", + "image_caption": [ + "(b) Scalar func. $f_{2}$ ." + ], + "image_footnote": [], + "bbox": [ + 424, + 146, + 602, + 258 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/f4ceb1ae7e305e8818135444cafec40f30f0f66a97538e2738035e2434c7fab6.jpg", + "image_caption": [ + "(c) Vis. diff." + ], + "image_footnote": [], + "bbox": [ + 633, + 146, + 782, + 258 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/3e51a20a812773a5e116b5370ae3c193c029babc5ccc44712bd4640b75b38b68.jpg", + "image_caption": [ + "(d) Pers. barc. of $f_1$ ." + ], + "image_footnote": [], + "bbox": [ + 222, + 277, + 395, + 339 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/eb57e2f6b538b268e340fd22af3b0a3bc60fd1edc058695c44775a6fd6c450de.jpg", + "image_caption": [ + "(e) Pers. barc. of $f_{2}$ ." + ], + "image_footnote": [], + "bbox": [ + 421, + 277, + 594, + 339 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/f465b508c3f4aa142f5e1a82f528282f6e0de4a9efb6e81032d4c98005f1a1a7.jpg", + "image_caption": [ + "(f) F-Cross-Barcode.", + "Fig. 4: Two scalar functions (a), (b) are indistinguishable by their persistence barcodes (c), (d). At the same time, $\\mathrm{SFTD}_1(f_1,f_2) = 1$ , see (e)." + ], + "image_footnote": [], + "bbox": [ + 607, + 277, + 779, + 339 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "which equal $-1$ in a vicinity of two concentric spheres and a bridge connecting them, which is depicted in green. The bridge is located in different places (above or below of the smallest sphere). In other areas of space, the scalar functions equal 0. We sample a $32^3$ grid from the unit cube, see Figure 5 (a), (b) and compare functions restricted on this grid via $F$ -Cross-Barcode. Again, regular persistence barcodes of two spheres with bridges are the same, see Figure 5, (d), (e): they contain 2 segments from $H_2$ corresponding to two cavities. But $F$ -Cross-Barcode is non-empty, see Figure 5, (f). The visualization technique correctly highlights areas with topological differences, see Figure 5 (c), where the central vertical slice is shown.", + "bbox": [ + 212, + 426, + 787, + 579 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.2 Comparing Eigenvectors of Graph Laplacian", + "text_level": 1, + "bbox": [ + 214, + 602, + 627, + 618 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In the most general form, the proposed tools can compare scalar functions $f: \\mathcal{V} \\to \\mathbb{R}$ defined on vertices $\\mathcal{V}$ of an undirected graph $\\mathcal{G} = (\\mathcal{V}, \\mathcal{E})$ . The normalized graph Laplacian is the matrix $L = D^{-1/2}(D - A)D^{-1/2}$ where $A$ is an adjacency matrix and $D$ is a diagonal matrix of node degrees. Eigenvectors $e_i$ and eigenvalues $\\lambda_i$ of $L$ describe intrinsic properties of the graph $\\mathcal{G}$ . Eigenvectors can be considered as scalar functions defined on graph's vertices $\\mathcal{V}$ . For eigenvectors: $(e_i, e_j) = 0$ and their norm is not defined; one can put $||e_i|| = 1$ for convenience. Thus, $e_i$ can't be distinguished by Euclidean distance or cosine distance. Alternatively, we can study $e_i$ as functions on vertices and compare them via SFTD.", + "bbox": [ + 212, + 628, + 787, + 777 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We consider a random graph with 30 nodes, see Figure 6. Figure 7 presents symmetrized $\\sum_{k=0,1} \\mathrm{SFTD}_k(e_i, e_j)$ via a heatmap, where $e_i, e_j$ are graph Laplacian eigenvectors ordered by eigenvalues. We observed the following interesting patterns: block structure, high topological similarity in the top left block, low", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Scalar Function Topology Divergence", + "bbox": [ + 480, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/e88065fab56b5ab72ee89ef7a28fdb8e425d191fe5a58b44b17bab39f47d5acf.jpg", + "image_caption": [ + "(a) Scalar func. $f_{1}$ ." + ], + "image_footnote": [], + "bbox": [ + 217, + 147, + 401, + 284 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/52af2e937923e1cb598b78cd11c5b3b55bafb10c4be3a871ab820e35f05ec85b.jpg", + "image_caption": [ + "(b) Scalar func. $f_{2}$ ." + ], + "image_footnote": [], + "bbox": [ + 408, + 147, + 591, + 284 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/288d13bb79d66d0d21002e1a4e481af7b027fbf7ca53b9cf52bd59526f78e706.jpg", + "image_caption": [ + "(c) Vis. diff." + ], + "image_footnote": [], + "bbox": [ + 604, + 146, + 782, + 281 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/00d52bcc64268c36432279fab4859f834b0afa42653cf9aa435796e4bb6fc218.jpg", + "image_caption": [ + "(d) Pers. barc. of $f_{1}$ ." + ], + "image_footnote": [], + "bbox": [ + 222, + 301, + 397, + 363 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/4c239fb411e8e02483fbb62d468775514cc6ab94496356f93451cf40ced6ace4.jpg", + "image_caption": [ + "(e) Pers. barc. of $f_{2}$ ." + ], + "image_footnote": [], + "bbox": [ + 421, + 301, + 596, + 363 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/02497798d1a0434f9433960f9c691ae3fb834b7714344ec13c6da181f562bac3.jpg", + "image_caption": [ + "(f) F-Cross-Barcode.", + "Fig. 5: Two scalar fields (a), (b) are indistinguishable by their persistence barcodes (d), (e). At the same time, $\\mathrm{SFTD}_1(f_1,f_2) = 1$ , see (f)." + ], + "image_footnote": [], + "bbox": [ + 607, + 301, + 781, + 363 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/9fa32e4376aa87517b402463bcbf351c9a17b4d5d28c960841d25bfd77aa8ea5.jpg", + "table_caption": [ + "Table 1: Errors of 3D shape reconstruction for \"Red blood cell\" dataset." + ], + "table_footnote": [], + "table_body": "
METHOD1-IoUVOLUMESURFACEROUGH.H0H1H2
SHAPR [37]0.4870.3530.2430.3610.0980.2410.065
SHAPR+W.D. [36]0.4670.2910.1840.2940.0980.2280.032
SHAPR+Σi=0,1,2SFTDi0.4490.2940.1760.2820.0790.2220.065
SHAPR+Σi=0,1,2,3SFTDi0.4530.2660.1890.2760.0790.2170.025
", + "bbox": [ + 215, + 486, + 787, + 575 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "topological similarity in the bottom right block, and very interesting diagonal lines parallel to the main diagonal. Figure 8 shows two the most similar eigenvectors, Figure 9 - two the most dissimilar ones, while their $F$ -Cross-Barcodes are shown on the right.", + "bbox": [ + 212, + 618, + 787, + 679 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.3 Enhancing 3D reconstruction", + "text_level": 1, + "bbox": [ + 214, + 705, + 503, + 720 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "SHAPR (SHApe PRediction) [37] is a deep learning model for reconstructing cellular and nuclear 3D shapes from 2D fluorescence microscopy images. Correct reconstruction of structural patterns in 3D shapes is critical for this problem. However, such information is not captured by established pixel/voxel-wise loss terms, e.g. Dice loss. Later, [36] improved SHAPR by adding contextual information about the overall shape of an object. The work [36] proposed to augment SHAPR with a topological loss term - a Wasserstein-2 distance (W.D.) between", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "I. Trofimov et al.", + "bbox": [ + 269, + 114, + 387, + 126 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/bd867d0aaeea8902623d0961ada0e57876244b0a3835a0dc86618c04ade28df3.jpg", + "image_caption": [ + "Fig. 6: A random Watts-Strogatz small-world graph [38] with 30 nodes." + ], + "image_footnote": [], + "bbox": [ + 238, + 165, + 470, + 289 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/01c184984f677ff2b714151d26999c779135968763d86a385dbb426dec059b9f.jpg", + "image_caption": [ + "Fig.7: Comparison of graph Laplacian eigenvectors via SFTD." + ], + "image_footnote": [], + "bbox": [ + 516, + 157, + 779, + 297 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/cefa2dcf58c0425a033f3f8a1e9a114b99a350cd8cf1619175ddc30d96e55cc2.jpg", + "image_caption": [ + "(a) Eigenvector 1." + ], + "image_footnote": [], + "bbox": [ + 228, + 376, + 344, + 446 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/37f366c109bdd92513fb6ab6aa7e6113f270569fe7733cd4e6f9a363c8e350dc.jpg", + "image_caption": [ + "(b) Eigenvector 4." + ], + "image_footnote": [], + "bbox": [ + 372, + 377, + 488, + 446 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/598c980bdb85d93577dfc354c882b5739a757599a06ed420f4bcb126566a9091.jpg", + "image_caption": [ + "(c) F-Cross-Barcode $(e_1,e_4)$", + "Fig.8: Comparison of $e_1$ and $e_4$ ." + ], + "image_footnote": [], + "bbox": [ + 504, + 380, + 638, + 441 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/5ca3151bd924b785f75f25a8749074107a2e1ee451d5e3201549dcb5206b6e5c.jpg", + "image_caption": [ + "(d) F-Cross-Barcode $(e_4,e_1)$" + ], + "image_footnote": [], + "bbox": [ + 648, + 380, + 782, + 441 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/9c86fd33eb424c0151530547f074d98b108dbf934c6682e6d701f2b3bf77648b.jpg", + "image_caption": [ + "(a) Eigenvector 0." + ], + "image_footnote": [], + "bbox": [ + 228, + 527, + 344, + 598 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/1f6069fd265f7d9040b62089adab0d6165009850ab21de3dd03577fab58ec3ef.jpg", + "image_caption": [ + "(b) Eigenvector 22.", + "Fig. 9: Comparison of $e_0$ and $e_{22}$ ." + ], + "image_footnote": [], + "bbox": [ + 372, + 529, + 488, + 598 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/d5a2ed65dae205e30160c04caf7d9bec0864528eea015ab1cef4b73b1f312670.jpg", + "image_caption": [ + "(c) F-Cross-Barcode $(e_0,e_{22})$" + ], + "image_footnote": [], + "bbox": [ + 504, + 534, + 638, + 594 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/08e9a06ef170114a2797ff618e0430ba32f54997672e04d202a631ec38dec65b.jpg", + "image_caption": [ + "(d) F-Cross-Barcode $(e_{22}, e_0)$ ." + ], + "image_footnote": [], + "bbox": [ + 648, + 537, + 782, + 593 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/b48c2302ce2a0dd8e88e8fea4e109602a0f69e3cde79064aded06ceda486aedb.jpg", + "table_caption": [ + "Table 2: Errors of 3D shape reconstruction for \"Cell nuclei\" dataset." + ], + "table_footnote": [], + "table_body": "
METHOD1-IoUVOLUMESURFACEROUGH.H0H1H2
SHAPR [37]0.6120.4820.2730.1910.0000.0070.038
SHAPR+W.D. [36]0.6120.4370.2510.2000.0000.0080.027
SHAPR+Σi=0,1,2SFTDi0.6180.4320.2440.1870.0000.0080.020
SHAPR+Σi=0,1,2,SFTDi0.5970.4370.2470.1910.0000.0110.011
", + "bbox": [ + 217, + 724, + 787, + 811 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Scalar Function Topology Divergence", + "bbox": [ + 480, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "persistence barcodes of ground truth and predicted 3D objects and a total persistence. However, as we discussed before, this loss term doesn't take into account the localization of topological features. We replaced the topological loss term in SHAPR with a symmetrized SFTD with $p = 2$ in two variants8: dimensions up to 2 and dimensions up to 3. Then, we carried out experiments with two datasets from [37].", + "bbox": [ + 212, + 146, + 787, + 236 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 1 shows experimental results for the \"Red blood cell\" dataset. We conclude that SHAPR augmented with the SFTD loss term is the best one in terms of 3D shape matching errors: IoU error, volume error, surface error, and roughness error (see details in [37]). Additionally, we calculated Wasserstein-2 distances between persistence barcodes $(H_0, H_1, H_2)$ of ground truth and generated 3D objects. Statistically significant improvements are underlined. The proposed SFTD topological loss term with 0-3 dimensions has the lowest Wasserstein-2 distances. This is the rare case of an application of higher homologies like $H_3$ for real-world applications.", + "bbox": [ + 212, + 237, + 787, + 372 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "For the \"Cell nuclei\" dataset results are presented in Table ??. Differences are quite small and are statistically significant only for two metrics: IoU error and W.D. $H_{2}$ , where SHAPR with SFTD loss is the best model.", + "bbox": [ + 212, + 373, + 787, + 417 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Figure 10 shows 3D visualizations of $2\\mathrm{D} \\rightarrow 3\\mathrm{D}$ reconstructions. We see that even with the topological regularization from [36], 3D reconstructions are sometimes erroneous and have serious topological defects: cavities, holes, and surface breaks. In experiments, the sublevel filtration was used and a density function equaled \"1\" inside the object and \"0\" in the ambient space. That is, a cavity inside a sphere corresponds to an extra bar in $H_0$ barcode, an extra connected component corresponds to an extra bar in $H_2$ barcode.", + "bbox": [ + 212, + 419, + 787, + 525 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.4 Identifying topological errors in 3D segmentation", + "text_level": 1, + "bbox": [ + 214, + 549, + 668, + 565 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Brain tumor 3D segmentation aims to reveal the sub-regions of tumor corresponding to the enhancing tumor, the edematous/invaded tissue, the necrosis. Availability of versatile similarity measures not only allows precise evaluation of deep learning 3D segmentation models, but also assists monitoring disease progression in clinical practice. As mentioned above, measuring the similarity of the two 3D segmentations by comparing their persistence barcodes does not account for localization of the topological features. In this section, we illustrate the ability of SFTD to complement the existing metrics with this additional information.", + "bbox": [ + 212, + 575, + 787, + 710 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In order to investigate the applicability of the proposed SFTD as a metric for 3D segmentation, we utilize the publicly available pretrained Swin UNETR model [21,33] and BraTS 21 dataset [2-4,28]. Figures 11a, 11b demonstrate the ground truth segmentation and model's prediction. While each of these two 3D segmentations has two voids inside, only two voids have the same localization", + "bbox": [ + 212, + 710, + 787, + 787 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "I. Trofimov et al.", + "bbox": [ + 271, + 114, + 387, + 127 + ], + "page_idx": 11 + }, + { + "type": "page_footnote", + "text": "8 A 3D cubical complex has empty $H_{3}$ homologies. But during the calculation of SFTD for 3D scalar functions, a 4D cubical complex is created, where $H_{3}$ homologies can exist.", + "bbox": [ + 217, + 796, + 787, + 839 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/41761ffd262887a54d63fd0727c7558abaddc1694f2bd5bbf16c3ad88a962044.jpg", + "image_caption": [ + "(a) Ground truth" + ], + "image_footnote": [], + "bbox": [ + 217, + 220, + 401, + 362 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/697ce54b929563dc668bd033838f372062628e902c0d47afc79e0203bfe85347.jpg", + "image_caption": [ + "(b) SHAPR+W.D." + ], + "image_footnote": [], + "bbox": [ + 405, + 222, + 591, + 362 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/cce19c6d5e2159634786337fd7d517e120134d377add3182811a9599fe002561.jpg", + "image_caption": [ + "(c) SHAPR+SFTD" + ], + "image_footnote": [], + "bbox": [ + 602, + 222, + 784, + 362 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/aff7a95aed9a7a7441a94a39017c440a152ebf12428b130b65c7778f4451b11f.jpg", + "image_caption": [ + "(d) Ground truth" + ], + "image_footnote": [], + "bbox": [ + 217, + 380, + 401, + 534 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/76fe38e46fe731ba4507148732ac5b1e86631b131e28d846827e201e569bd4db.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 416, + 378, + 598, + 532 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/1c487d91bb50c6554ad1c098fbfce44f5fdabd08b88dd89d19c00bc553a6c372.jpg", + "image_caption": [ + "(f) SHAPR+SFTD" + ], + "image_footnote": [], + "bbox": [ + 602, + 380, + 784, + 534 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/e4ce5524d6dc2fa9a370202961b2f3ff6e3d0a4b3e9c6d83bec94e26f508f045.jpg", + "image_caption": [ + "(g) Ground truth" + ], + "image_footnote": [], + "bbox": [ + 217, + 549, + 395, + 689 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/812345f3109fad08e623ae47e9aceeeec85d482fcd240a0b55a72cf854d63a71.jpg", + "image_caption": [ + "(e) SHAPR+W.D.", + "(h) SHAPR+W.D." + ], + "image_footnote": [], + "bbox": [ + 416, + 549, + 594, + 690 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/5f514d9dd00b360eb434c1558fff8e1a5945839b77374236590c63bfe6db2faa.jpg", + "image_caption": [ + "(i) $\\mathrm{SHAPR} + \\mathrm{SFTD}$", + "Fig. 10: Examples of topological errors. Top, $H_0$ : SHAPR+W.D. has a wrong cavity inside. Middle, $H_1$ : SHAPR+W.D. has a wrong hole. Bottom, $H_2$ : SHAPR+W.D. has a wrong surface break." + ], + "image_footnote": [], + "bbox": [ + 602, + 550, + 782, + 689 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Scalar Function Topology Divergence", + "bbox": [ + 480, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/9099adfac9472f400b60913b792f1470085a04b9e4399505e9c1b8820ffcf93b.jpg", + "image_caption": [ + "(a) Ground truth" + ], + "image_footnote": [], + "bbox": [ + 222, + 150, + 395, + 248 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/0810c284d85d4dfa652b7b3a30b41575108132af301e38f8ac400a741933f96a.jpg", + "image_caption": [ + "(b) Prediction" + ], + "image_footnote": [], + "bbox": [ + 418, + 148, + 596, + 250 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/8c54e9e47e0f1c8f3dcb4600ae08eb82ad5087dd9f04be4a28e6aa3db568c114.jpg", + "image_caption": [ + "(c) Pers. barc. of ground truth" + ], + "image_footnote": [], + "bbox": [ + 606, + 186, + 782, + 248 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/d5d25b88525fb4e0cd318060e0e8ff5a0d3e46c45d42f8fd6e1558b724704717.jpg", + "image_caption": [ + "(d) Pers. barc. of prediction" + ], + "image_footnote": [], + "bbox": [ + 222, + 268, + 398, + 332 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/f656680a8bc0f06d1eea9b3ff3bbf2eff033988e28b081d1d6f6000bbfbc0f51.jpg", + "image_caption": [ + "(e) F-Cross-Barcode", + "Fig. 11: Examples of topological errors. Two 3D segmentations - (a) - ground truth and (b) - prediction - are indistinguishable by their persistence barcodes (c), (d). At the same time, SFTD between ground truth and prediction equals to 1, see (e). A pair of points inside each object in (a), (b) corresponds to the dissimilar topological features, voids, where birth/death events are depicted by red/green dots." + ], + "image_footnote": [], + "bbox": [ + 606, + 268, + 782, + 332 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "which makes the two 3D segmentations different. As shown in Figures 11c, 11d, the two 3D segmentations are indistinguishable by their persistence barcodes. In contrary, SFTD is able to recognise the difference between these two cases as revealed by the F-Cross-Barcode in Figure 11e. In particular, SFTD indicates that the difference is precisely in locations of one pair of voids since the set of 1-dimensional topological features contains one segment.", + "bbox": [ + 212, + 458, + 787, + 549 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusions", + "text_level": 1, + "bbox": [ + 215, + 569, + 369, + 585 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we have proposed new topological tools: F-Cross-Barcode and Scalar Function Topology Divergence (SFTD), which measure the dissimilarity of multi-scale topology between sublevel sets of two functions having a common domain. Functions can be defined on an undirected graph or Euclidean space of any dimensionality. We have shown that our tools outperform the standard approach: evaluating Wasserstein distance between persistence barcodes, calculated independently for sublevel sets of scalar functions. F-Cross-Barcode and SFTD take into account the localization of topological features. SFTD, as a loss function, is beneficial in computer vision, since it helps to enforce topological similarity at multiple scales. Also, the proposed tools provide useful visualizations of areas of topological dissimilarities. In the experimental section, we demonstrated that SFTD improved the reconstruction of cellular 3D shapes from 2D fluorescence microscopy images and identified topological errors in 3D segmentation. Additionally, we showed that SFTD outperforms Betti matching loss in 2D segmentation problems. We believe that the proposed tools will have more applications in computer vision and learning from graph data.", + "bbox": [ + 212, + 598, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "I. Trofimov et al.", + "bbox": [ + 271, + 114, + 387, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "The work was supported by the Analytical center under the RF Government (subsidy agreement 000000D730321P5Q0002, Grant No. 70-2021-00145 02.11.2021)", + "bbox": [ + 215, + 172, + 799, + 204 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 224, + 323, + 239 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Abousamra, S., Hoai, M., Samaras, D., Chen, C.: Localization in the crowd with topological constraints. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 872-881 (2021)", + "2. Baid, Ujjwal, e.a.: The rsna-asnr-miccai brats 2021 benchmark on brain tumor segmentation and radiogenomic classification. arXiv preprint arXiv:2107.02314 (2021)", + "3. Bakas, S., Akbari, H., Sotiras, A., Bilello, M., Rozycki, M., Kirby, J.S., Freymann, J.B., Farahani, K., Davatzikos, C.: Advancing the cancer genome atlas glioma mri collections with expert segmentation labels and radiomic features. Nature Scientific Data (2017)", + "4. Bakas, S., Akbari, H., Sotiras, A., Bilello, M., Rozycki, M., Kirby, J.S., Freymann, J.B., Farahani, K., Davatzikos, C.: Segmentation labels and radiomic features for the pre-operative scans of the tcga-gbm collection. The Cancer Imaging Archive (2017)", + "5. Barannikov, S.: The framed Morse complex and its invariants. Advances in Soviet Mathematics 21, 93-116 (1994)", + "6. Barannikov, S., Korotin, A., Oganesyan, D., Emtsev, D., Burnaev, E.: Barcodes as summary of loss function topology. Doklady Rossijskoj akademii nauk. Matematika, informatika, processy upravlenia 514(2), 196-211 (2023)", + "7. Barannikov, S., Trofimov, I., Balabin, N., Burnaev, E.: Representation topology divergence: A method for comparing neural network representations. ICML 2022, arXiv:2201.00058 (2021)", + "8. Barannikov, S., Trofimov, I., Sotnikov, G., Trimbach, E., Korotin, A., Filippov, A., Burnaev, E.: Manifold topology divergence: a framework for comparing data manifolds. Advances in Neural Information Processing Systems 34, 7294-7305 (2021)", + "9. Barannikov, S., Voronkova, D., Trofimov, I., Korotin, A., Sotnikov, G., Burnaev, E.: Topological obstructions in neural networks learning. arXiv preprint arXiv:2012.15834 (2020)", + "10. Carpenter, A.E., Jones, T.R., Lamprecht, M.R., Clarke, C., Kang, I.H., Friman, O., Guertin, D.A., Chang, J.H., Lindquist, R.A., Moffat, J., et al.: Cellprofiler: image analysis software for identifying and quantifying cell phenotypes. Genome biology 7, 1-11 (2006)", + "11. Carrière, M., Chazal, F., Glisse, M., Ike, Y., Kannan, H., Umeda, Y.: Optimizing persistent homology based functions. In: International Conference on Machine Learning. pp. 1294-1303. PMLR (2021)", + "12. Carrière, M., Chazal, F., Ike, Y., Lacombe, T., Royer, M., Umeda, Y.: Perslay: A neural network layer for persistence diagrams and new graph topological signatures. In: International Conference on Artificial Intelligence and Statistics. pp. 2786-2796. PMLR (2020)", + "13. Chazal, F., Michel, B.: An introduction to topological data analysis: fundamental and practical aspects for data scientists. arXiv preprint arXiv:1710.04019 (2017)", + "14. Chen, Y., Segovia, I., Gel, Y.R.: Z-gcnets: Time zigzags at graph convolutional networks for time series forecasting. In: International Conference on Machine Learning. pp. 1684-1694. PMLR (2021)" + ], + "bbox": [ + 218, + 253, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Scalar Function Topology Divergence", + "bbox": [ + 480, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Chen, Y., Segovia Dominguez, I., Coskunuzer, B., Gel, Y.: Tamp-s2gcnets: When time-aware multipersistence meets spatio-supra graph convolutional nets while forecasting time series. In: The International Conference on Learning Representations (ICLR) (2022)", + "16. Demir, A., Coskunuzer, B., Gel, Y., Segovia-Dominguez, I., Chen, Y., Kiziltan, B.: Todd: Topological compound fingerprinting in computer-aided drug discovery. Advances in Neural Information Processing Systems 35, 27978-27993 (2022)", + "17. Funke, J., Tschopp, F., Grisaitis, W., Sheridan, A., Singh, C., Saalfeld, S., Turaga, S.C.: Large scale image segmentation with structured loss based deep learning for connectome reconstruction. IEEE transactions on pattern analysis and machine intelligence 41(7), 1669-1680 (2018)", + "18. Gelfand, S.I., Manin, Y.I.: Methods of homological algebra. Springer Science & Business Media (2002)", + "19. Gupta, S., Hu, X., Kaan, J., Jin, M., Mpoy, M., Chung, K., Singh, G., Saltz, M., Kurc, T., Saltz, J., et al.: Learning topological interactions for multi-class medical image segmentation. In: European Conference on Computer Vision. pp. 701-718. Springer (2022)", + "20. Gupta, S., Zhang, Y., Hu, X., Prasanna, P., Chen, C.: Topology-aware uncertainty for image segmentation. Advances in Neural Information Processing Systems 36 (2024)", + "21. Hatamizadeh, A., Nath, V., Tang, Y., Yang, D., Roth, H.R., Xu, D.: Swin unetr: Swin transformers for semantic segmentation of brain tumors in mri images. arXiv preprint arXiv:2201.01266v1 (2022)", + "22. Hofer, C., Graf, F., Rieck, B., Niethammer, M., Kwitt, R.: Graph filtration learning. In: International Conference on Machine Learning. pp. 4314-4323. PMLR (2020)", + "23. Horn, M., De Brouwer, E., Moor, M., Moreau, Y., Rieck, B., Borgwardt, K.: Topological graph neural networks. arXiv preprint arXiv:2102.07835 (2021)", + "24. Hu, X., Li, F., Samaras, D., Chen, C.: Topology-preserving deep image segmentation. Advances in neural information processing systems 32 (2019)", + "25. Hu, X., Wang, Y., Fuxin, L., Samaras, D., Chen, C.: Topology-aware segmentation using discrete Morse theory. arXiv preprint arXiv:2103.09992 (2021)", + "26. Kaczynski, T., Mischaikow, K.M., Mrozek, M.: Computational homology, vol. 3. Springer (2004)", + "27. Luo, Y., Shi, L., Thost, V.: Improving self-supervised molecular representation learning using persistent homology. arXiv preprint arXiv:2311.17327 (2023)", + "28. Menze, B.H., Jakab, A., Bauer, S., Kalpathy-Cramer, J., Farahani, K., Kirby, Justin, e.a.: The multimodal brain tumor image segmentation benchmark (brats). IEEE Transactions on Medical Imaging (2015)", + "29. Moor, M., Horn, M., Rieck, B., Borgwardt, K.: Topological autoencoders. In: International conference on machine learning. pp. 7045-7054. PMLR (2020)", + "30. Pérez, J.B., Hauke, S., Lupo, U., Caorsi, M., Dassatti, A.: giotto-ph: A python library for high-performance computation of persistent homology of vietoris-rips filtrations (2021)", + "31. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015)", + "32. Stucki, N., Paetzold, J.C., Shit, S., Menze, B., Bauer, U.: Topologically faithful image segmentation via induced matching of persistence barcodes. In: International Conference on Machine Learning. pp. 32698-32727. PMLR (2023)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "I. Trofimov et al.", + "bbox": [ + 271, + 114, + 385, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "33. Tang, Y., Yang, D., Li, W., Roth, H.R., Landman, B., Xu, D., Nath, V., Hatamizadeh, A.: Self-supervised pre-training of swim transformers for 3d medical image analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022)", + "34. Trofimov, I., Cherniavskii, D., Tulchinskii, E., Balabin, N., Burnaev, E., Barannikov, S.: Learning topology-preserving data representations. arXiv preprint arXiv:2302.00136 (2023)", + "35. Wagner, H., Chen, C., Vuçini, E.: Efficient computation of persistent homology for cubical data. In: Topological methods in data analysis and visualization II: theory, algorithms, and applications, pp. 91-106. Springer (2011)", + "36. Waibel, D.J., Atwell, S., Meier, M., Marr, C., Rieck, B.: Capturing shape information with multi-scale topological loss terms for 3d reconstruction. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 150-159. Springer (2022)", + "37. Waibel, D.J., Kiermeyer, N., Atwell, S., Sadafi, A., Meier, M., Marr, C.: Shapr predicts 3d cell shapes from 2d microscopic images. Iscience 25(11) (2022)", + "38. Watts, D.J., Strogatz, S.H.: Collective dynamics of 'small-world'networks. nature 393(6684), 440-442 (1998)", + "39. Wu, J., Wang, Y., Xue, T., Sun, X., Freeman, B., Tenenbaum, J.: Marrnet: 3d shape reconstruction via 2.5 d sketches. Advances in neural information processing systems 30 (2017)", + "40. Yan, Z., Ma, T., Gao, L., Tang, Z., Chen, C.: Link prediction with persistent homology: An interactive view. In: International conference on machine learning. pp. 11659-11669. PMLR (2021)" + ], + "bbox": [ + 212, + 146, + 787, + 481 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Scalar Function Topology Divergence", + "bbox": [ + 480, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Scalar Function Topology Divergence_ Comparing Topology of 3D Objects/228bb063-0fe3-438c-a000-0e252dd17cc3_model.json b/2024/Scalar Function Topology Divergence_ Comparing Topology of 3D Objects/228bb063-0fe3-438c-a000-0e252dd17cc3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..89798cbe5825410b6d558d5b906eb52a9e78a417 --- /dev/null +++ b/2024/Scalar Function Topology Divergence_ Comparing Topology of 3D Objects/228bb063-0fe3-438c-a000-0e252dd17cc3_model.json @@ -0,0 +1,3402 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.219, + 0.141, + 0.784, + 0.187 + ], + "angle": 0, + "content": "Scalar Function Topology Divergence: Comparing Topology of 3D Objects" + }, + { + "type": "text", + "bbox": [ + 0.256, + 0.212, + 0.747, + 0.244 + ], + "angle": 0, + "content": "Ilya Trofimov1, Daria Voronkova1,2, Eduard Tulchinskii1,4, Evgeny Burnaev1,2, and Serguei Barannikov1,3" + }, + { + "type": "text", + "bbox": [ + 0.268, + 0.255, + 0.732, + 0.296 + ], + "angle": 0, + "content": "\\(^{1}\\) Skolkovo Institute of Science and Technology, Moscow, Russia {ilya.trofimov,darya.voronkova, eduard.tulchinskiy,e.burnaev, s.barannikov}@skoltech.ru" + }, + { + "type": "text", + "bbox": [ + 0.421, + 0.297, + 0.584, + 0.31 + ], + "angle": 0, + "content": "2 AIRI, Moscow, Russia" + }, + { + "type": "text", + "bbox": [ + 0.355, + 0.31, + 0.649, + 0.325 + ], + "angle": 0, + "content": "\\(^{3}\\) CNRS, IMJ, Paris Cité University, France" + }, + { + "type": "text", + "bbox": [ + 0.323, + 0.325, + 0.681, + 0.338 + ], + "angle": 0, + "content": "4 AI Foundation and Algorithm Lab, Moscow, Russia" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.374, + 0.74, + 0.623 + ], + "angle": 0, + "content": "Abstract. We propose a new topological tool for computer vision - Scalar Function Topology Divergence (SFTD), which measures the dissimilarity of multi-scale topology between sublevel sets of two functions having a common domain. Functions can be defined on an undirected graph or Euclidean space of any dimensionality. Most of the existing methods for comparing topology are based on Wasserstein distance between persistence barcodes and they don't take into account the localization of topological features. The minimization of SFTD ensures that the corresponding topological features of scalar functions are located in the same places. The proposed tool provides useful visualizations depicting areas where functions have topological dissimilarities. We provide applications of the proposed method to 3D computer vision. In particular, experiments demonstrate that SFTD as an additional loss improves the reconstruction of cellular 3D shapes from 2D fluorescence microscopy images, and helps to identify topological errors in 3D segmentation. Additionally, we show that SFTD outperforms Betti matching loss in 2D segmentation problems. The code is publicly available: https://github.com/IlyaTrofimov/SFTD." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.636, + 0.741, + 0.665 + ], + "angle": 0, + "content": "Keywords: Topological Data Analysis \\(\\cdot\\) 3D Computer Vision \\(\\cdot\\) Geometric Deep Learning" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.689, + 0.377, + 0.705 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.843 + ], + "angle": 0, + "content": "A common situation in computer vision is a structural prediction defined over \\(\\mathbb{R}^n\\). Examples include image segmentation [24, 25, 31, 32] and 3D shape reconstruction [36, 37, 39] problems. However, in most of the cases the comparison between the prediction and a ground truth is done pixel(voxel)-wise by metrics like Dice score, cross-entropy, etc. These metrics are not capable of providing detailed topological correctness, i.e. preserving multi-scale shape patterns: cavities, holes, connectivity patterns, etc. Typically, predictions involve scalar functions \\(f:\\mathbb{R}^n\\to \\mathbb{R}\\) and the final shape of an object is defined by thresholding \\(f(x)\\)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.388, + 0.128 + ], + "angle": 0, + "content": "I. Trofimov et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "There is a need for topological-aware comparison metrics for scalar functions, which take into account the topology of sublevel sets \\( f^{-1}((-\\infty, \\varepsilon]) \\) (the analysis for superlevel sets is similar). Existing tools for this problem don't take into account the localization of topological features [24,36] or are limited to 2D space [32]. Our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.234, + 0.784, + 0.278 + ], + "angle": 0, + "content": "1. We introduce new topological tools: F-Cross-Barcodes and Scalar Function Topology Divergence (SFTD), for a multi-scale topology-aware comparison of scalar functions by tracking differences in their sublevel sets topology." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.28, + 0.784, + 0.308 + ], + "angle": 0, + "content": "2. A distinctive feature of F-Cross-Barcodes and SFTD is that they take into account the localization of topological features;" + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.31, + 0.784, + 0.351 + ], + "angle": 0, + "content": "3. We describe a visualization technique for F-Cross-Barcodes: a technique highlights with points areas where differences in topological features are localized:" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.354, + 0.612, + 0.367 + ], + "angle": 0, + "content": "4. We propose an algorithm for SFTD differentiation;" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.369, + 0.784, + 0.412 + ], + "angle": 0, + "content": "5. By doing computational experiments, we show that SFTD as an additional loss term improves 3D shape reconstruction, and helps to identify topological errors in 3D segmentation." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.415, + 0.784, + 0.458 + ], + "angle": 0, + "content": "6. We also show that SFTD outperforms Betti matching loss in terms of topological accuracy in 2D segmentation while being \\(\\approx 35\\) times faster, see Appendix G." + }, + { + "type": "list", + "bbox": [ + 0.222, + 0.234, + 0.784, + 0.458 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.483, + 0.383, + 0.499 + ], + "angle": 0, + "content": "2 Related work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.514, + 0.784, + 0.604 + ], + "angle": 0, + "content": "Tools of topological data analysis find application in various machine learning problems: evaluation of deep generative models [8], comparison of neural representations [7], dimensionality reduction [29,34], graph neural networks [12,23,40], self-supervised learning [27], molecular data analysis [16], analysis of loss functions in deep learning [6,9], spatial-temporal graph neural network tasks [14,15], crowd localization [1], to name a few." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.606, + 0.784, + 0.71 + ], + "angle": 0, + "content": "Several works used topological tools to enhance image segmentation [20, 24, 25, 32]. However, proposed algorithms for comparison ground truth and predicted images either don't take into account localization of topological features [24], or are limited to 2D [32]. A method by [25] highlights in images topologically critical and error-prone structures using discrete Morse theory, and [20] proposes a method for inferring structure-aware uncertainty. The importance of topological interactions for multi-class medical image segmentation is shown in [19]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.733, + 0.382, + 0.75 + ], + "angle": 0, + "content": "3 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.765, + 0.574, + 0.841 + ], + "angle": 0, + "content": "We study scalar functions \\( f: \\mathbb{R}^n \\to \\mathbb{R} \\) by assessing the topology of its sublevel sets. Consider a 1D function depicted in Figure 1. It has two local minima \\( y_1, y_2 \\) and a local maximum \\( y_3 \\), such that \\( y_1 < y_2 < y_3 \\). One can study sublevel sets" + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.734, + 0.784, + 0.83 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.583, + 0.845, + 0.788, + 0.872 + ], + "angle": 0, + "content": "Fig.1: 1D function and its sublevel sets." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.482, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Scalar Function Topology Divergence" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.193 + ], + "angle": 0, + "content": "\\(X_{\\varepsilon} = f^{-1}((-\\infty ,\\varepsilon ])\\). For the scalar function \\(f\\), changes of \\(X_{\\varepsilon}\\) for increasing \\(\\varepsilon\\) can be intuitively understood as a \"filling the pool\", see Figure 1. There are four distinct situations:" + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.205, + 0.393, + 0.22 + ], + "angle": 0, + "content": "1. For \\(\\varepsilon < y_{1}\\), \\(X_{\\varepsilon} = \\varnothing\\)" + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.221, + 0.533, + 0.236 + ], + "angle": 0, + "content": "2. For \\(y_{1} \\leq \\varepsilon < y_{2}\\), \\(X_{\\varepsilon}\\) is a single segment;" + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.237, + 0.551, + 0.251 + ], + "angle": 0, + "content": "3. For \\(y_{2} \\leq \\varepsilon < y_{3}\\), \\(S_{\\varepsilon}\\) consists of 2 segments;" + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.253, + 0.549, + 0.267 + ], + "angle": 0, + "content": "4. Finally, for \\(\\varepsilon \\geq y_{3}\\), \\(X_{\\varepsilon}\\) is a single segment." + }, + { + "type": "list", + "bbox": [ + 0.222, + 0.205, + 0.551, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.279, + 0.785, + 0.34 + ], + "angle": 0, + "content": "Note, that changes in topology of \\(X_{\\varepsilon}\\) happen when \\(\\varepsilon\\) is a critical point of \\(f\\). In the general case, smooth scalar functions defined on smooth manifolds are studied by Morse theory [6]. In practice, we assess functions through their values on a finite set of points, e.g. a lattice (square grid graph) in \\(\\mathbb{R}^n\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.366, + 0.553, + 0.382 + ], + "angle": 0, + "content": "3.1 Scalar functions defined on a graph" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.394, + 0.785, + 0.454 + ], + "angle": 0, + "content": "In the most general form, we are interested in topology-aware comparison of scalar functions \\( f: \\mathcal{V} \\to \\mathbb{R} \\) defined on vertices \\( \\mathcal{V} \\) of an undirected graph \\( \\mathcal{G} = (\\mathcal{V}, \\mathcal{E}) \\). Specifically, we study topological features of subgraphs \\( \\{v \\in \\mathcal{V} \\mid f(v) \\leq \\varepsilon\\} \\) connected components, cycles, etc." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.455, + 0.785, + 0.56 + ], + "angle": 0, + "content": "For each clique \\( C \\) in the graph \\( \\mathcal{G} \\), we define a filtration function \\( T(C) = \\max_{i\\in C}f(i) \\). By a clique complex \\( \\mathcal{C} \\), we mean a set of cliques where each sub-clique also belongs to a complex5. For some \\( \\varepsilon \\), \\( T(C) \\) defines a clique complex \\( \\mathcal{C}_j = \\{C\\mid T(C)\\leq \\varepsilon \\} \\), where \\( f(j) = \\varepsilon \\). Topological features for a subgraph \\( \\{v\\in \\mathcal{V}\\mid f(v)\\leq \\varepsilon \\} \\) are precisely expressed by homology groups. Overall, the filtration function \\( T(C) \\) defines the lower-star filtration of the graph's clique complex:" + }, + { + "type": "equation", + "bbox": [ + 0.416, + 0.563, + 0.589, + 0.578 + ], + "angle": 0, + "content": "\\[\n\\varnothing \\subseteq \\mathcal {C} _ {0} \\subseteq \\mathcal {C} _ {1} \\subseteq \\dots \\subseteq \\mathcal {C} _ {n},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.587, + 0.784, + 0.617 + ], + "angle": 0, + "content": "where \\(\\mathcal{C}_n\\) is a clique complex of \\(\\mathcal{G}\\). The filtration is said to be induced by the vertex function \\(f\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.618, + 0.785, + 0.695 + ], + "angle": 0, + "content": "A given topological feature \"appears\" at some \\(\\varepsilon_{1}\\) and \"dies\" at some \\(\\varepsilon_{2}\\). These topological features, together with their birth and death times can be calculated via tools of topological data analysis. The multi-set of the corresponding intervals \\((\\varepsilon_{1},\\varepsilon_{2})\\) is called the persistence barcode, which depicts topological features of the graph for all scales. The whole theory is dubbed persistent homology [5, 13]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.72, + 0.331, + 0.736 + ], + "angle": 0, + "content": "4 Method" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.755, + 0.554, + 0.772 + ], + "angle": 0, + "content": "4.1 Scalar functions defined on a graph" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.783, + 0.785, + 0.814 + ], + "angle": 0, + "content": "To compare topological features of sublevel sets of functions \\(f,g:\\mathcal{G}\\to \\mathbb{R}\\) we construct a graph \\(\\tilde{\\mathcal{G}} = (\\tilde{\\mathcal{V}},\\tilde{\\mathcal{E}})\\) with doubled set of vertices from \\(\\mathcal{G}\\) and one extra" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.783, + 0.84 + ], + "angle": 0, + "content": "Every clique complex is an abstract simplicial complex, but the reserve is not true." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.271, + 0.115, + 0.388, + 0.127 + ], + "angle": 0, + "content": "I. Trofimov et al." + }, + { + "type": "image", + "bbox": [ + 0.295, + 0.147, + 0.422, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.283, + 0.282, + 0.433, + 0.296 + ], + "angle": 0, + "content": "(a) The doubled graph \\(\\tilde{\\mathcal{G}}\\)" + }, + { + "type": "image", + "bbox": [ + 0.567, + 0.179, + 0.715, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.282, + 0.78, + 0.296 + ], + "angle": 0, + "content": "(b) The extended \\(n + 1\\) dimensional lattice \\(\\tilde{\\mathcal{L}}\\)." + }, + { + "type": "image_caption", + "bbox": [ + 0.395, + 0.306, + 0.607, + 0.322 + ], + "angle": 0, + "content": "Fig. 2: A visualization of \\(\\tilde{\\mathcal{G}}\\), \\(\\tilde{\\mathcal{L}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.35, + 0.785, + 0.381 + ], + "angle": 0, + "content": "vertex \\(O\\), see Figure 2a. Formally, \\(\\tilde{\\mathcal{V}} = \\{A_1,\\dots ,A_n,A_1',\\dots ,A_n',O\\}\\). Edges \\(\\tilde{\\mathcal{E}}\\) are defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.351, + 0.39, + 0.651, + 0.452 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\left(A _ {i}, A _ {j}\\right) \\in \\mathcal {E} \\Rightarrow \\left(A _ {i}, A _ {j}\\right) \\in \\tilde {\\mathcal {E}}, \\left(A _ {i} ^ {\\prime}, A _ {j} ^ {\\prime}\\right) \\in \\tilde {\\mathcal {E}} \\\\ \\left(A _ {i}, A _ {j}\\right) \\in \\mathcal {E}, i < j \\Rightarrow \\left(A _ {i}, A _ {j} ^ {\\prime}\\right) \\in \\tilde {\\mathcal {E}} \\\\ \\left(O, A _ {i}\\right) \\in \\tilde {\\mathcal {E}}, \\left(A _ {i}, A _ {i} ^ {\\prime}\\right) \\in \\tilde {\\mathcal {E}} \\text {f o r a l l} i \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.463, + 0.528, + 0.479 + ], + "angle": 0, + "content": "The vertex function \\(\\tilde{f}\\) is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.36, + 0.488, + 0.644, + 0.529 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\tilde {f} (A _ {i}) = f (i) \\\\ \\tilde {f} (A _ {i} ^ {\\prime}) = \\min (f (i), g (i)), \\tilde {f} (O) = \\min _ {f, g}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.764, + 0.554 + ], + "angle": 0, + "content": "where \\(\\min_{f,g}\\) is the minimum value of both \\(f,g\\) in the nodes of the graph \\(\\mathcal{G}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.554, + 0.785, + 0.598 + ], + "angle": 0, + "content": "Definition. The \\(F\\)-Cross-Barcode\\(_k(f, g)\\) for two scalar functions \\(f, g\\) defined on vertices of an undirected graph \\(\\mathcal{G}\\) is the \\(k\\)-th persistence barcode for the doubled graph \\(\\tilde{\\mathcal{G}}\\) and the lower-star filtration induced by \\(\\tilde{f}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.787, + 0.689 + ], + "angle": 0, + "content": "To put it simply, each interval in the \\( F \\)-Cross-Barcode\\(_k(f, g)\\) denotes a range of functions' sub-level sets where a given topological feature exists in the function \\( f \\) but doesn't exist in the function \\( g \\), see further Appendix A. Technically, the desired persistence barcode is calculated via giotto-ph library [30]. For this software, values of a filtration function must be defined on vertices and edges. Algorithm 1 summarizes the aforementioned." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.71, + 0.559, + 0.725 + ], + "angle": 0, + "content": "4.2 Scalar functions defined on a lattice" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "A frequent situation is a function having a domain in \\(\\mathbb{R}^n\\). We assume that the function is approximated by its restriction on a sufficiently small grid. Algorithm 1 can also be applied in this case after a space triangulation. However, the number of simplices grows very quickly in high dimensions (see discussion in [35]), and Algorithm 1 has a high computational cost. That is why, we consider cubical complexes [26] instead of simplicial complexes. Specifically, we consider a regular square grid graph (\\(n\\)-dimensional lattice) and use values of functions" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.482, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Scalar Function Topology Divergence" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.147, + 0.719, + 0.162 + ], + "angle": 0, + "content": "Algorithm 1 F-Cross-Barcode\\(_k\\) (scalar function \\( k \\)-th cross-barcode)" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.167, + 0.74, + 0.181 + ], + "angle": 0, + "content": "Input: Scalar functions \\( f, g \\) defined on vertices of a graph \\( \\mathcal{G} = (\\mathcal{V}, \\mathcal{E}) \\). \\( |\\mathcal{V}| = n \\)." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.182, + 0.786, + 0.223 + ], + "angle": 0, + "content": "Require: \\(\\mathsf{PB}_k(\\cdot)\\) - an algorithm to calculate \\(k\\)-th persistence barcode from a matrix containing values of a filtration function on graph: on vertices - on a diagonal, on edges - off a diagonal." + }, + { + "type": "text", + "bbox": [ + 0.231, + 0.228, + 0.745, + 0.255 + ], + "angle": 0, + "content": "1. Create a \\(n \\times n\\) matrix \\(F\\) for the function \\(f\\): \\(F_{ij} = f(i)\\) for \\(i = j\\), \\(\\max(f(i), f(j))\\) for \\(i \\neq j\\) and \\((i, j) \\in \\mathcal{E}\\), \\(+\\infty\\) otherwise." + }, + { + "type": "text", + "bbox": [ + 0.231, + 0.256, + 0.623, + 0.269 + ], + "angle": 0, + "content": "2. Create a \\(n \\times n\\) matrix \\(G\\) for the function \\(g\\) analogously;" + }, + { + "type": "text", + "bbox": [ + 0.231, + 0.27, + 0.631, + 0.282 + ], + "angle": 0, + "content": "3. Initialize \\( c_{+\\infty} \\) to a vector of \\( n \\) elements, all equal to \\( +\\infty \\);" + }, + { + "type": "text", + "bbox": [ + 0.231, + 0.284, + 0.78, + 0.296 + ], + "angle": 0, + "content": "4. Create \\( F_{+} \\), a matrix derived from \\( F \\) with a lower-diagonal part replaced by \\( +\\infty \\);" + }, + { + "type": "text", + "bbox": [ + 0.231, + 0.297, + 0.656, + 0.31 + ], + "angle": 0, + "content": "5. Create \\( F_{diag} \\), a vector containing the diagonal elements of \\( F \\);" + }, + { + "type": "text", + "bbox": [ + 0.231, + 0.311, + 0.654, + 0.324 + ], + "angle": 0, + "content": "6. Compute \\(\\min (F,G)\\), the element-wise minimum of \\(F\\) and \\(G\\);" + }, + { + "type": "text", + "bbox": [ + 0.231, + 0.325, + 0.685, + 0.338 + ], + "angle": 0, + "content": "7. Compute \\(\\min_{F,G}\\), the minimum value of all elements in \\(F\\) and \\(G\\);" + }, + { + "type": "text", + "bbox": [ + 0.231, + 0.339, + 0.649, + 0.351 + ], + "angle": 0, + "content": "8. Compute the \\((2n + 1)\\times (2n + 1)\\) block matrix \\(m\\) as follows:" + }, + { + "type": "list", + "bbox": [ + 0.231, + 0.228, + 0.78, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "equation", + "bbox": [ + 0.396, + 0.361, + 0.617, + 0.406 + ], + "angle": 0, + "content": "\\[\nm \\leftarrow \\left( \\begin{array}{c c c} \\min (F, G) & F _ {+} ^ {\\intercal} & c _ {+ \\infty} ^ {\\intercal} \\\\ F _ {+} & F & F _ {d i a g} ^ {\\intercal} \\\\ c _ {+ \\infty} & F _ {d i a g} & \\min _ {F, G} \\end{array} \\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.421, + 0.545, + 0.435 + ], + "angle": 0, + "content": "9. Compute: F-Cross-Barcode\\(_k(f, g) \\gets \\mathsf{PB}_k(m)\\)" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.439, + 0.712, + 0.454 + ], + "angle": 0, + "content": "Return: F-Cross-Barcode\\(_k(f, g)\\), a multi-set of (\"birth\", \"death\") intervals." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.488, + 0.787, + 0.534 + ], + "angle": 0, + "content": "in nodes of this grid: \\( f(j_{1},\\ldots ,j_{n}),g(j_{1},\\ldots ,j_{n}) \\). To compare two function \\( f,g \\) we create an extended \\( n + 1 \\) dimensional lattice \\( \\tilde{\\mathcal{L}} \\) with one extra dimension of size 3, see Figure 2b. The vertex function \\( \\tilde{f}:\\tilde{\\mathcal{L}}\\to \\mathbb{R} \\) is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.281, + 0.545, + 0.72, + 0.602 + ], + "angle": 0, + "content": "\\[\n\\tilde {f} (i, j _ {1}, \\ldots , j _ {n}) = \\left\\{ \\begin{array}{l l} \\min _ {f, g} & \\text {f o r} i = 0 \\\\ f (j _ {1}, \\ldots , j _ {n}) & \\text {f o r} i = 1 \\\\ \\min (f (j _ {1}, \\ldots , j _ {n}), g (j _ {1}, \\ldots , j _ {n})) & \\text {f o r} i = 2 \\end{array} \\right.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.705 + ], + "angle": 0, + "content": "Similarly to functions on graphs, \\( F \\)-Cross-Barcode\\(_k(f,g)\\) for functions on lattices is defined as the \\( k \\)-th persistence barcode of the filtered cubical complex with the filtration induced by the vertex function \\( \\tilde{f} \\). We use the GUDHI\\(^6\\) library which implements efficient algorithms for cubical complexes [35]. Note that \\( F \\)-Cross-Barcode\\(_k(f,g)\\) for functions restricted on a lattice and a triangulation of the same space are not equivalent, but would be similar." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.729, + 0.57, + 0.744 + ], + "angle": 0, + "content": "4.3 Scalar Function Topology Divergence" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.755, + 0.788, + 0.817 + ], + "angle": 0, + "content": "The \\( F \\)-Cross-Barcode\\(_k(f, g)\\) depicts topological differences of functions' sublevel sets at multiple scales. The \\( F \\)-Cross-Barcode\\(_k(f, g)\\), as a regular barcode, is a multi-set of intervals. One requires some vectorization technique to inject it into a machine learning pipeline. If one needs to describe a discrepancy between \\( f, g \\)" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.414, + 0.84 + ], + "angle": 0, + "content": "\\(^{6}\\) https://gudhi.inria.fr/" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.271, + 0.115, + 0.388, + 0.128 + ], + "angle": 0, + "content": "I. Trofimov et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.178 + ], + "angle": 0, + "content": "by an integral numerical characteristic, we propose a scalar function topology divergence:" + }, + { + "type": "equation", + "bbox": [ + 0.326, + 0.187, + 0.786, + 0.222 + ], + "angle": 0, + "content": "\\[\n\\mathrm {S F T D} _ {k} (f, g) = \\sum_ {(b, d) \\in \\mathrm {F} - \\text {C r o s s - B a r c o d e} _ {k} (f, g)} | | d - b | | ^ {p}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.231, + 0.789, + 0.339 + ], + "angle": 0, + "content": "that is the sum of lengths of intervals from the \\( F \\)-Cross-Barcode\\(_k(f, g)\\) to \\( p \\)-th power, where typically \\( p = 1, 2 \\). SFTD is not symmetric, and we often use an average of \\( \\mathrm{SFTD}_k(f, g) \\) and \\( \\mathrm{SFTD}_k(g, f) \\) in experiments below. For different dimensions \\( k \\), \\( \\mathrm{SFTD}_k(f, g) \\) can be used either separately, or the sum can be taken to reflect discrepancies in all types of topological features (connected components, cycles, voids, etc.). Computational complexity of SFTD is discussed in Appendix C." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.359, + 0.452, + 0.375 + ], + "angle": 0, + "content": "4.4 SFTD basic properties" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.383, + 0.79, + 0.459 + ], + "angle": 0, + "content": "Proposition 1. If \\( SFTD_{k}(f,g) = SFTD_{k}(g,f) = 0 \\) for all \\( k\\geq 0 \\), then the sublevel sets barcodes of functions \\( f \\) and \\( g \\) are the same in any degree. Moreover, in this case their topological features are located in the same places: the inclusions \\( f^{-1}((- \\infty ,\\varepsilon ])\\subseteq (\\min (f,g))^{-1}((- \\infty ,\\varepsilon ]) \\), \\( g^{-1}((- \\infty ,\\varepsilon ])\\subseteq (\\min (f,g))^{-1}((- \\infty ,\\varepsilon ]) \\), induce homology isomorphisms for any threshold \\( \\varepsilon \\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.469, + 0.484, + 0.484 + ], + "angle": 0, + "content": "Proposition 2. (SFTD Stability)" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.485, + 0.778, + 0.5 + ], + "angle": 0, + "content": "(a) For any scalar functions \\( f, \\tilde{f}, g, \\tilde{g} \\) on a grid, or on a graph vertex set:" + }, + { + "type": "equation", + "bbox": [ + 0.232, + 0.506, + 0.769, + 0.546 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} d _ {B} (F - C r o s s - B a r c o d e _ {k} (f, g), F - C r o s s - B a r c o d e _ {k} \\left(f ^ {\\prime}, g ^ {\\prime}\\right)) \\leq \\\\ \\max (\\max _ {i} | f (i) - f ^ {\\prime} (i) |, \\max _ {i} | g (i) - g ^ {\\prime} (i) |). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.544, + 0.545, + 0.559 + ], + "angle": 0, + "content": "(b) For any pair of scalar functions \\( f, \\hat{f}' \\):" + }, + { + "type": "equation", + "bbox": [ + 0.328, + 0.564, + 0.675, + 0.586 + ], + "angle": 0, + "content": "\\[\n\\| F \\text {- C r o s s - B a r c o d e} _ {k} (f, f ^ {\\prime}) \\| _ {B} \\leq \\max _ {i} | f (i) - f ^ {\\prime} (i) |,\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.601, + 0.785, + 0.632 + ], + "angle": 0, + "content": "where \\( d_B \\) denotes the bottleneck distance between persistence barcodes [13]. We give the proofs in Appendix A. As a consequence, SFTD is robust to noise." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.654, + 0.544, + 0.668 + ], + "angle": 0, + "content": "4.5 Visualization of F-Cross-Barcodes" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.678, + 0.787, + 0.709 + ], + "angle": 0, + "content": "\\(F\\)-Cross-Barcodes can be used for localization and visualization of topological differences between scalar functions. Recall the filtered cubical7 complex:" + }, + { + "type": "equation", + "bbox": [ + 0.412, + 0.719, + 0.589, + 0.736 + ], + "angle": 0, + "content": "\\[\n\\varnothing \\subseteq \\mathcal {C} _ {0} \\subseteq \\mathcal {C} _ {1} \\subseteq \\dots \\subseteq \\mathcal {C} _ {n}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.745, + 0.786, + 0.777 + ], + "angle": 0, + "content": "Since the filtration is induced by a vertex function \\(\\tilde{f}:\\tilde{\\mathcal{L}}\\to \\mathbb{R}\\), the filtered cubical complex corresponds to the sequence:" + }, + { + "type": "equation", + "bbox": [ + 0.4, + 0.785, + 0.602, + 0.805 + ], + "angle": 0, + "content": "\\[\n\\tilde {f} (i _ {0}) \\leq \\tilde {f} (i _ {1}) \\leq \\dots \\leq \\tilde {f} (i _ {n}).\n\\]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.787, + 0.841 + ], + "angle": 0, + "content": "For ease of presentation, we discuss only the visualization for cubical complexes. The visualization for clique complexes is analogous." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.482, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Scalar Function Topology Divergence" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.268 + ], + "angle": 0, + "content": "If \\(\\tilde{f}(i_k) = \\tilde{f}(i_m)\\), the order is defined by indexes \\(i_k, i_m\\). Thus, each segment in the barcode - a birth/death pair - is associated with two vertices \\(v_b, v_d \\in \\tilde{\\mathcal{L}}\\). Recall that \\(\\tilde{\\mathcal{L}}\\) is an extended \\(n + 1\\) dimensional lattice (see Section 4.2), where the first dimension is of size 3, and the rest of dimensions are of the same size as the lattice inside the domain of \\(f, g\\). So, neglecting the first dimension, we can map the rest of indices \\(j_1, \\ldots, j_n\\) of birth/death vertices \\(v_b, v_d\\) right onto original \\(n\\)-dimensional lattice. Later, in the experimental section, we shall provide such visualizations, where birth/death events are depicted by orange/red dots." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.288, + 0.434, + 0.302 + ], + "angle": 0, + "content": "4.6 SFTD differentiation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.31, + 0.785, + 0.416 + ], + "angle": 0, + "content": "Calculation of \\(\\mathrm{SFTD}_k(f,g)\\) involves a filtered simplicial(cubical) complex \\(\\mathcal{C}\\) with a vertex function \\(\\tilde{f} (i),i\\in \\tilde{\\mathcal{V}}\\) . Also, since we apply the sub-level set filtration, for every simplex(cube) \\(\\sigma \\colon T(\\sigma) = \\tilde{f} (i_{\\sigma})\\) , where \\(i_{\\sigma} = \\operatorname{argmax}_{i\\in \\sigma}\\tilde{f} (i)\\) . Each interval \\((b,d)\\) in a \\(F\\) -Cross-Barcode corresponds to some pair of simplices(cubes) \\((\\sigma_1,\\sigma_2)\\) which join the simplicial(cubical) complex \\(\\mathcal{C}\\) at moments \\(b = T(\\sigma_1),d = T(\\sigma_2)\\) Generally, we consider \\(f,g\\) (and consequently \\(\\tilde{f}\\)) to be parametric functions of \\(w\\) , for example, neural networks. By differentiating (1) we obtain:" + }, + { + "type": "equation", + "bbox": [ + 0.304, + 0.426, + 0.785, + 0.47 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial \\operatorname {S F T D} _ {k} (f , g)}{\\partial w} = \\sum_ {\\left(\\sigma_ {1}, \\sigma_ {2}\\right) \\in \\Sigma_ {k}} p \\left(\\frac {\\partial \\tilde {f} \\left(i _ {\\sigma_ {2}}\\right)}{\\partial w} - \\frac {\\partial \\tilde {f} \\left(i _ {\\sigma_ {2}}\\right)}{\\partial w}\\right) ^ {p - 1}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.478, + 0.784, + 0.508 + ], + "angle": 0, + "content": "where \\(\\Sigma_{k}\\) is a collection of \"birth\", \"death\" simplex pairs corresponding to intervals in the \\(F\\)-Cross-Barcode\\(_k(f, g)\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.784, + 0.539 + ], + "angle": 0, + "content": "By construction, \\(\\tilde{f}(i)\\) can be either \\(f(i')\\) or \\(g(i')\\) for some mapping \\(i' = r(i)\\). Formally, introducing a binary function \\(c(\\cdot)\\) for this choice:" + }, + { + "type": "equation", + "bbox": [ + 0.372, + 0.547, + 0.785, + 0.565 + ], + "angle": 0, + "content": "\\[\n\\tilde {f} (i) = c (i) f (r (i)) + (1 - c (i)) g (r (i)), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.572, + 0.248, + 0.585 + ], + "angle": 0, + "content": "and" + }, + { + "type": "equation", + "bbox": [ + 0.353, + 0.585, + 0.785, + 0.615 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial \\tilde {f} (i)}{\\partial w} = b (i) \\frac {\\partial f (r (i))}{\\partial w} + (1 - b (i)) \\frac {\\partial g (r (i))}{\\partial w}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.619, + 0.784, + 0.695 + ], + "angle": 0, + "content": "Finally, the desirable gradient \\(\\partial \\mathrm{SFTD}_k(f,g) / \\partial w\\) can be calculated by combining (2), (4) via a chain rule. With a common assumption that all the \\(f(i),g(i)\\) are distinct, functions \\(c,r\\) don't change in a vicinity of \\(w\\), and \\(\\mathrm{SFTD}_i(f,g)\\) is differentiable w.r.t. \\(w\\) [11, 22]. On a high level, the differentiation is done similarly to [11, 34]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.715, + 0.376, + 0.733 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.745, + 0.768, + 0.76 + ], + "angle": 0, + "content": "In the experimental section, we are going to answer three research questions:" + }, + { + "type": "text", + "bbox": [ + 0.207, + 0.766, + 0.784, + 0.809 + ], + "angle": 0, + "content": "Q1: Does SFTD and F-Cross-Barcodes provide a better topologically-aware comparison of scalar functions than Wasserstein distance between regular persistence barcodes?" + }, + { + "type": "text", + "bbox": [ + 0.207, + 0.811, + 0.784, + 0.825 + ], + "angle": 0, + "content": "Q2: Does F-Cross-Barcodes help to visualize and localize topological differences?" + }, + { + "type": "text", + "bbox": [ + 0.207, + 0.825, + 0.74, + 0.84 + ], + "angle": 0, + "content": "Q3: Is SFTD, as a loss function, beneficial for computer vision problems?" + }, + { + "type": "list", + "bbox": [ + 0.207, + 0.766, + 0.784, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.271, + 0.115, + 0.388, + 0.127 + ], + "angle": 0, + "content": "I. Trofimov et al." + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.148, + 0.399, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.254, + 0.267, + 0.366, + 0.279 + ], + "angle": 0, + "content": "(a) Scalar func. \\( f_{1} \\)." + }, + { + "type": "image", + "bbox": [ + 0.423, + 0.149, + 0.601, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.455, + 0.267, + 0.568, + 0.279 + ], + "angle": 0, + "content": "(b) Scalar func. \\( f_{2} \\)." + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.147, + 0.784, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.67, + 0.267, + 0.743, + 0.278 + ], + "angle": 0, + "content": "(c) Vis. diff." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.282, + 0.4, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.248, + 0.35, + 0.371, + 0.363 + ], + "angle": 0, + "content": "(d) Pers. barc. of \\( f_{1} \\)." + }, + { + "type": "image", + "bbox": [ + 0.414, + 0.282, + 0.589, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.439, + 0.351, + 0.56, + 0.363 + ], + "angle": 0, + "content": "(e) Pers. barc. of \\( f_{2} \\)." + }, + { + "type": "image", + "bbox": [ + 0.608, + 0.282, + 0.783, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.635, + 0.351, + 0.754, + 0.362 + ], + "angle": 0, + "content": "(f) F-Cross-Barcode" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.374, + 0.787, + 0.403 + ], + "angle": 0, + "content": "Fig. 3: Two scalar functions (a), (b) are indistinguishable by their persistence barcodes (d), (e). At the same time, \\(\\mathrm{SFTD}_0(f_1,f_2) = 4.6\\) see (f)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.429, + 0.383, + 0.444 + ], + "angle": 0, + "content": "5.1 Synthetic data" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.451, + 0.785, + 0.496 + ], + "angle": 0, + "content": "To illustrate the performance of the proposed tools, we start with synthetic data. Functions under comparison have coinciding persistence barcodes of all degrees and they are indistinguishable by standard Wasserstein matching." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.515, + 0.788, + 0.668 + ], + "angle": 0, + "content": "Functions with local minima in 2D. First, we compare two scalar function on \\(64 \\times 64\\) grid, each of them having 3 local minima but in different locations, see Figure 3, (a), (b). Persistence barcodes of both functions are equal, see Figure 3, (d), (e): they contain two segments from \\(H_0\\) corresponding to local minima. Here and further, essential segments (which die at \\(+\\infty\\)) are omitted. The topology of the two functions is indistinguishable by standard persistence barcodes and their Wasserstein distance is zero. At the same time, F-Cross-Barcode between these scalar functions is non-empty for \\(H_0\\), see Figure 3, (f). The visualization technique correctly points at places with different topology - the centers and borders of the distinguishing local minima." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.685, + 0.788, + 0.79 + ], + "angle": 0, + "content": "Lattices in 2D. Next, we compare two lattice-style 2D scalar functions, see Figure 4, (a), (b). Lattices have defects in different places (top left and bottom right respectively). The persistence barcodes of both lattices are identical, see Figure 4, (d), (e): they contain 8 segments from \\( H_{1} \\) corresponding to independent cycles. At the same time, \\( F \\)-Cross-Barcode between two scalar functions is nonempty, see Figure 4, (f). The visualization technique correctly points at places with defects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Concentric spheres in 3D. Finally, we study the performance of the proposed method by comparing 3D scalar functions. We compare two scalar functions," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.482, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Scalar Function Topology Divergence" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.147, + 0.396, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.254, + 0.264, + 0.365, + 0.276 + ], + "angle": 0, + "content": "(a) Scalar func. \\( f_{1} \\)." + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.147, + 0.603, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.457, + 0.264, + 0.571, + 0.276 + ], + "angle": 0, + "content": "(b) Scalar func. \\( f_{2} \\)." + }, + { + "type": "image", + "bbox": [ + 0.634, + 0.147, + 0.783, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.671, + 0.264, + 0.746, + 0.276 + ], + "angle": 0, + "content": "(c) Vis. diff." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.278, + 0.396, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.248, + 0.346, + 0.371, + 0.358 + ], + "angle": 0, + "content": "(d) Pers. barc. of \\( f_1 \\)." + }, + { + "type": "image", + "bbox": [ + 0.422, + 0.278, + 0.596, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.447, + 0.346, + 0.568, + 0.358 + ], + "angle": 0, + "content": "(e) Pers. barc. of \\( f_{2} \\)." + }, + { + "type": "image", + "bbox": [ + 0.608, + 0.278, + 0.781, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.633, + 0.346, + 0.755, + 0.358 + ], + "angle": 0, + "content": "(f) F-Cross-Barcode." + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.369, + 0.788, + 0.399 + ], + "angle": 0, + "content": "Fig. 4: Two scalar functions (a), (b) are indistinguishable by their persistence barcodes (c), (d). At the same time, \\(\\mathrm{SFTD}_1(f_1,f_2) = 1\\), see (e)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.428, + 0.789, + 0.58 + ], + "angle": 0, + "content": "which equal \\(-1\\) in a vicinity of two concentric spheres and a bridge connecting them, which is depicted in green. The bridge is located in different places (above or below of the smallest sphere). In other areas of space, the scalar functions equal 0. We sample a \\(32^3\\) grid from the unit cube, see Figure 5 (a), (b) and compare functions restricted on this grid via \\(F\\)-Cross-Barcode. Again, regular persistence barcodes of two spheres with bridges are the same, see Figure 5, (d), (e): they contain 2 segments from \\(H_2\\) corresponding to two cavities. But \\(F\\)-Cross-Barcode is non-empty, see Figure 5, (f). The visualization technique correctly highlights areas with topological differences, see Figure 5 (c), where the central vertical slice is shown." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.603, + 0.629, + 0.619 + ], + "angle": 0, + "content": "5.2 Comparing Eigenvectors of Graph Laplacian" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.779 + ], + "angle": 0, + "content": "In the most general form, the proposed tools can compare scalar functions \\( f: \\mathcal{V} \\to \\mathbb{R} \\) defined on vertices \\( \\mathcal{V} \\) of an undirected graph \\( \\mathcal{G} = (\\mathcal{V}, \\mathcal{E}) \\). The normalized graph Laplacian is the matrix \\( L = D^{-1/2}(D - A)D^{-1/2} \\) where \\( A \\) is an adjacency matrix and \\( D \\) is a diagonal matrix of node degrees. Eigenvectors \\( e_i \\) and eigenvalues \\( \\lambda_i \\) of \\( L \\) describe intrinsic properties of the graph \\( \\mathcal{G} \\). Eigenvectors can be considered as scalar functions defined on graph's vertices \\( \\mathcal{V} \\). For eigenvectors: \\( (e_i, e_j) = 0 \\) and their norm is not defined; one can put \\( ||e_i|| = 1 \\) for convenience. Thus, \\( e_i \\) can't be distinguished by Euclidean distance or cosine distance. Alternatively, we can study \\( e_i \\) as functions on vertices and compare them via SFTD." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We consider a random graph with 30 nodes, see Figure 6. Figure 7 presents symmetrized \\(\\sum_{k=0,1} \\mathrm{SFTD}_k(e_i, e_j)\\) via a heatmap, where \\(e_i, e_j\\) are graph Laplacian eigenvectors ordered by eigenvalues. We observed the following interesting patterns: block structure, high topological similarity in the top left block, low" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.271, + 0.115, + 0.388, + 0.127 + ], + "angle": 0, + "content": "I. Trofimov et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.148, + 0.402, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.254, + 0.287, + 0.366, + 0.299 + ], + "angle": 0, + "content": "(a) Scalar func. \\( f_{1} \\)." + }, + { + "type": "image", + "bbox": [ + 0.409, + 0.148, + 0.593, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.443, + 0.287, + 0.556, + 0.299 + ], + "angle": 0, + "content": "(b) Scalar func. \\( f_{2} \\)." + }, + { + "type": "image", + "bbox": [ + 0.606, + 0.147, + 0.783, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.658, + 0.287, + 0.73, + 0.298 + ], + "angle": 0, + "content": "(c) Vis. diff." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.302, + 0.398, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.249, + 0.37, + 0.371, + 0.383 + ], + "angle": 0, + "content": "(d) Pers. barc. of \\( f_{1} \\)." + }, + { + "type": "image", + "bbox": [ + 0.422, + 0.302, + 0.597, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.448, + 0.37, + 0.568, + 0.383 + ], + "angle": 0, + "content": "(e) Pers. barc. of \\( f_{2} \\)." + }, + { + "type": "image", + "bbox": [ + 0.608, + 0.302, + 0.782, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.634, + 0.37, + 0.755, + 0.382 + ], + "angle": 0, + "content": "(f) F-Cross-Barcode." + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.393, + 0.787, + 0.423 + ], + "angle": 0, + "content": "Fig. 5: Two scalar fields (a), (b) are indistinguishable by their persistence barcodes (d), (e). At the same time, \\(\\mathrm{SFTD}_1(f_1,f_2) = 1\\), see (f)." + }, + { + "type": "table_caption", + "bbox": [ + 0.258, + 0.438, + 0.744, + 0.452 + ], + "angle": 0, + "content": "Table 1: Errors of 3D shape reconstruction for \"Red blood cell\" dataset." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.487, + 0.789, + 0.577 + ], + "angle": 0, + "content": "
METHOD1-IoUVOLUMESURFACEROUGH.H0H1H2
SHAPR [37]0.4870.3530.2430.3610.0980.2410.065
SHAPR+W.D. [36]0.4670.2910.1840.2940.0980.2280.032
SHAPR+Σi=0,1,2SFTDi0.4490.2940.1760.2820.0790.2220.065
SHAPR+Σi=0,1,2,3SFTDi0.4530.2660.1890.2760.0790.2170.025
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.619, + 0.788, + 0.68 + ], + "angle": 0, + "content": "topological similarity in the bottom right block, and very interesting diagonal lines parallel to the main diagonal. Figure 8 shows two the most similar eigenvectors, Figure 9 - two the most dissimilar ones, while their \\( F \\)-Cross-Barcodes are shown on the right." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.706, + 0.504, + 0.721 + ], + "angle": 0, + "content": "5.3 Enhancing 3D reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "SHAPR (SHApe PRediction) [37] is a deep learning model for reconstructing cellular and nuclear 3D shapes from 2D fluorescence microscopy images. Correct reconstruction of structural patterns in 3D shapes is critical for this problem. However, such information is not captured by established pixel/voxel-wise loss terms, e.g. Dice loss. Later, [36] improved SHAPR by adding contextual information about the overall shape of an object. The work [36] proposed to augment SHAPR with a topological loss term - a Wasserstein-2 distance (W.D.) between" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.482, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Scalar Function Topology Divergence" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.24, + 0.166, + 0.471, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.312, + 0.493, + 0.34 + ], + "angle": 0, + "content": "Fig. 6: A random Watts-Strogatz small-world graph [38] with 30 nodes." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.158, + 0.78, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.312, + 0.785, + 0.34 + ], + "angle": 0, + "content": "Fig.7: Comparison of graph Laplacian eigenvectors via SFTD." + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.377, + 0.346, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.234, + 0.455, + 0.34, + 0.467 + ], + "angle": 0, + "content": "(a) Eigenvector 1." + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.378, + 0.49, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.378, + 0.455, + 0.486, + 0.467 + ], + "angle": 0, + "content": "(b) Eigenvector 4." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.381, + 0.64, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.505, + 0.445, + 0.64, + 0.468 + ], + "angle": 0, + "content": "(c) F-Cross-Barcode \\((e_1,e_4)\\)" + }, + { + "type": "image", + "bbox": [ + 0.649, + 0.381, + 0.784, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.649, + 0.445, + 0.785, + 0.467 + ], + "angle": 0, + "content": "(d) F-Cross-Barcode \\((e_4,e_1)\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.389, + 0.478, + 0.612, + 0.492 + ], + "angle": 0, + "content": "Fig.8: Comparison of \\(e_1\\) and \\(e_4\\)." + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.529, + 0.345, + 0.599 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.234, + 0.607, + 0.341, + 0.619 + ], + "angle": 0, + "content": "(a) Eigenvector 0." + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.53, + 0.49, + 0.599 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.375, + 0.607, + 0.488, + 0.619 + ], + "angle": 0, + "content": "(b) Eigenvector 22." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.535, + 0.64, + 0.595 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.505, + 0.598, + 0.64, + 0.62 + ], + "angle": 0, + "content": "(c) F-Cross-Barcode \\((e_0,e_{22})\\)" + }, + { + "type": "image", + "bbox": [ + 0.649, + 0.538, + 0.784, + 0.594 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.649, + 0.597, + 0.785, + 0.62 + ], + "angle": 0, + "content": "(d) F-Cross-Barcode \\((e_{22}, e_0)\\)." + }, + { + "type": "image_caption", + "bbox": [ + 0.386, + 0.63, + 0.615, + 0.644 + ], + "angle": 0, + "content": "Fig. 9: Comparison of \\( e_0 \\) and \\( e_{22} \\)." + }, + { + "type": "table_caption", + "bbox": [ + 0.27, + 0.676, + 0.731, + 0.69 + ], + "angle": 0, + "content": "Table 2: Errors of 3D shape reconstruction for \"Cell nuclei\" dataset." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.725, + 0.789, + 0.813 + ], + "angle": 0, + "content": "
METHOD1-IoUVOLUMESURFACEROUGH.H0H1H2
SHAPR [37]0.6120.4820.2730.1910.0000.0070.038
SHAPR+W.D. [36]0.6120.4370.2510.2000.0000.0080.027
SHAPR+Σi=0,1,2SFTDi0.6180.4320.2440.1870.0000.0080.020
SHAPR+Σi=0,1,2,SFTDi0.5970.4370.2470.1910.0000.0110.011
" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.388, + 0.128 + ], + "angle": 0, + "content": "I. Trofimov et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.237 + ], + "angle": 0, + "content": "persistence barcodes of ground truth and predicted 3D objects and a total persistence. However, as we discussed before, this loss term doesn't take into account the localization of topological features. We replaced the topological loss term in SHAPR with a symmetrized SFTD with \\( p = 2 \\) in two variants8: dimensions up to 2 and dimensions up to 3. Then, we carried out experiments with two datasets from [37]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.788, + 0.373 + ], + "angle": 0, + "content": "Table 1 shows experimental results for the \"Red blood cell\" dataset. We conclude that SHAPR augmented with the SFTD loss term is the best one in terms of 3D shape matching errors: IoU error, volume error, surface error, and roughness error (see details in [37]). Additionally, we calculated Wasserstein-2 distances between persistence barcodes \\((H_0, H_1, H_2)\\) of ground truth and generated 3D objects. Statistically significant improvements are underlined. The proposed SFTD topological loss term with 0-3 dimensions has the lowest Wasserstein-2 distances. This is the rare case of an application of higher homologies like \\(H_3\\) for real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.374, + 0.788, + 0.418 + ], + "angle": 0, + "content": "For the \"Cell nuclei\" dataset results are presented in Table ??. Differences are quite small and are statistically significant only for two metrics: IoU error and W.D. \\( H_{2} \\), where SHAPR with SFTD loss is the best model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.42, + 0.788, + 0.526 + ], + "angle": 0, + "content": "Figure 10 shows 3D visualizations of \\(2\\mathrm{D} \\rightarrow 3\\mathrm{D}\\) reconstructions. We see that even with the topological regularization from [36], 3D reconstructions are sometimes erroneous and have serious topological defects: cavities, holes, and surface breaks. In experiments, the sublevel filtration was used and a density function equaled \"1\" inside the object and \"0\" in the ambient space. That is, a cavity inside a sphere corresponds to an extra bar in \\(H_0\\) barcode, an extra connected component corresponds to an extra bar in \\(H_2\\) barcode." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.55, + 0.669, + 0.566 + ], + "angle": 0, + "content": "5.4 Identifying topological errors in 3D segmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.577, + 0.788, + 0.711 + ], + "angle": 0, + "content": "Brain tumor 3D segmentation aims to reveal the sub-regions of tumor corresponding to the enhancing tumor, the edematous/invaded tissue, the necrosis. Availability of versatile similarity measures not only allows precise evaluation of deep learning 3D segmentation models, but also assists monitoring disease progression in clinical practice. As mentioned above, measuring the similarity of the two 3D segmentations by comparing their persistence barcodes does not account for localization of the topological features. In this section, we illustrate the ability of SFTD to complement the existing metrics with this additional information." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.712, + 0.788, + 0.789 + ], + "angle": 0, + "content": "In order to investigate the applicability of the proposed SFTD as a metric for 3D segmentation, we utilize the publicly available pretrained Swin UNETR model [21,33] and BraTS 21 dataset [2-4,28]. Figures 11a, 11b demonstrate the ground truth segmentation and model's prediction. While each of these two 3D segmentations has two voids inside, only two voids have the same localization" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.797, + 0.788, + 0.84 + ], + "angle": 0, + "content": "8 A 3D cubical complex has empty \\( H_{3} \\) homologies. But during the calculation of SFTD for 3D scalar functions, a 4D cubical complex is created, where \\( H_{3} \\) homologies can exist." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.482, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Scalar Function Topology Divergence" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.222, + 0.402, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.259, + 0.365, + 0.362, + 0.377 + ], + "angle": 0, + "content": "(a) Ground truth" + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.223, + 0.593, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.444, + 0.365, + 0.555, + 0.377 + ], + "angle": 0, + "content": "(b) SHAPR+W.D." + }, + { + "type": "image", + "bbox": [ + 0.603, + 0.223, + 0.785, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.638, + 0.365, + 0.751, + 0.377 + ], + "angle": 0, + "content": "(c) SHAPR+SFTD" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.381, + 0.402, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.259, + 0.537, + 0.362, + 0.549 + ], + "angle": 0, + "content": "(d) Ground truth" + }, + { + "type": "image", + "bbox": [ + 0.417, + 0.379, + 0.599, + 0.534 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.453, + 0.537, + 0.564, + 0.548 + ], + "angle": 0, + "content": "(e) SHAPR+W.D." + }, + { + "type": "image", + "bbox": [ + 0.603, + 0.381, + 0.785, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.638, + 0.537, + 0.751, + 0.548 + ], + "angle": 0, + "content": "(f) SHAPR+SFTD" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.55, + 0.396, + 0.69 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.259, + 0.693, + 0.362, + 0.705 + ], + "angle": 0, + "content": "(g) Ground truth" + }, + { + "type": "image", + "bbox": [ + 0.417, + 0.55, + 0.596, + 0.691 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.453, + 0.693, + 0.564, + 0.705 + ], + "angle": 0, + "content": "(h) SHAPR+W.D." + }, + { + "type": "image", + "bbox": [ + 0.604, + 0.551, + 0.784, + 0.69 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.638, + 0.693, + 0.751, + 0.705 + ], + "angle": 0, + "content": "(i) \\(\\mathrm{SHAPR} + \\mathrm{SFTD}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.716, + 0.788, + 0.759 + ], + "angle": 0, + "content": "Fig. 10: Examples of topological errors. Top, \\( H_0 \\): SHAPR+W.D. has a wrong cavity inside. Middle, \\( H_1 \\): SHAPR+W.D. has a wrong hole. Bottom, \\( H_2 \\): SHAPR+W.D. has a wrong surface break." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.388, + 0.127 + ], + "angle": 0, + "content": "I. Trofimov et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.151, + 0.396, + 0.249 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.259, + 0.255, + 0.362, + 0.267 + ], + "angle": 0, + "content": "(a) Ground truth" + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.149, + 0.597, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.466, + 0.255, + 0.551, + 0.267 + ], + "angle": 0, + "content": "(b) Prediction" + }, + { + "type": "image", + "bbox": [ + 0.607, + 0.187, + 0.783, + 0.249 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.605, + 0.255, + 0.783, + 0.267 + ], + "angle": 0, + "content": "(c) Pers. barc. of ground truth" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.27, + 0.399, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.228, + 0.339, + 0.393, + 0.351 + ], + "angle": 0, + "content": "(d) Pers. barc. of prediction" + }, + { + "type": "image", + "bbox": [ + 0.607, + 0.27, + 0.783, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.635, + 0.339, + 0.754, + 0.351 + ], + "angle": 0, + "content": "(e) F-Cross-Barcode" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.362, + 0.788, + 0.432 + ], + "angle": 0, + "content": "Fig. 11: Examples of topological errors. Two 3D segmentations - (a) - ground truth and (b) - prediction - are indistinguishable by their persistence barcodes (c), (d). At the same time, SFTD between ground truth and prediction equals to 1, see (e). A pair of points inside each object in (a), (b) corresponds to the dissimilar topological features, voids, where birth/death events are depicted by red/green dots." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.459, + 0.788, + 0.55 + ], + "angle": 0, + "content": "which makes the two 3D segmentations different. As shown in Figures 11c, 11d, the two 3D segmentations are indistinguishable by their persistence barcodes. In contrary, SFTD is able to recognise the difference between these two cases as revealed by the F-Cross-Barcode in Figure 11e. In particular, SFTD indicates that the difference is precisely in locations of one pair of voids since the set of 1-dimensional topological features contains one segment." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.57, + 0.37, + 0.587 + ], + "angle": 0, + "content": "6 Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.789, + 0.842 + ], + "angle": 0, + "content": "In this paper, we have proposed new topological tools: F-Cross-Barcode and Scalar Function Topology Divergence (SFTD), which measure the dissimilarity of multi-scale topology between sublevel sets of two functions having a common domain. Functions can be defined on an undirected graph or Euclidean space of any dimensionality. We have shown that our tools outperform the standard approach: evaluating Wasserstein distance between persistence barcodes, calculated independently for sublevel sets of scalar functions. F-Cross-Barcode and SFTD take into account the localization of topological features. SFTD, as a loss function, is beneficial in computer vision, since it helps to enforce topological similarity at multiple scales. Also, the proposed tools provide useful visualizations of areas of topological dissimilarities. In the experimental section, we demonstrated that SFTD improved the reconstruction of cellular 3D shapes from 2D fluorescence microscopy images and identified topological errors in 3D segmentation. Additionally, we showed that SFTD outperforms Betti matching loss in 2D segmentation problems. We believe that the proposed tools will have more applications in computer vision and learning from graph data." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.482, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Scalar Function Topology Divergence" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.174, + 0.8, + 0.205 + ], + "angle": 0, + "content": "The work was supported by the Analytical center under the RF Government (subsidy agreement 000000D730321P5Q0002, Grant No. 70-2021-00145 02.11.2021)" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.226, + 0.324, + 0.241 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.254, + 0.785, + 0.295 + ], + "angle": 0, + "content": "1. Abousamra, S., Hoai, M., Samaras, D., Chen, C.: Localization in the crowd with topological constraints. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 872-881 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.296, + 0.785, + 0.322 + ], + "angle": 0, + "content": "2. Baid, Ujjwal, e.a.: The rsna-asnr-miccai brats 2021 benchmark on brain tumor segmentation and radiogenomic classification. arXiv preprint arXiv:2107.02314 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.323, + 0.785, + 0.377 + ], + "angle": 0, + "content": "3. Bakas, S., Akbari, H., Sotiras, A., Bilello, M., Rozycki, M., Kirby, J.S., Freymann, J.B., Farahani, K., Davatzikos, C.: Advancing the cancer genome atlas glioma mri collections with expert segmentation labels and radiomic features. Nature Scientific Data (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.378, + 0.785, + 0.431 + ], + "angle": 0, + "content": "4. Bakas, S., Akbari, H., Sotiras, A., Bilello, M., Rozycki, M., Kirby, J.S., Freymann, J.B., Farahani, K., Davatzikos, C.: Segmentation labels and radiomic features for the pre-operative scans of the tcga-gbm collection. The Cancer Imaging Archive (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.432, + 0.785, + 0.458 + ], + "angle": 0, + "content": "5. Barannikov, S.: The framed Morse complex and its invariants. Advances in Soviet Mathematics 21, 93-116 (1994)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.459, + 0.785, + 0.499 + ], + "angle": 0, + "content": "6. Barannikov, S., Korotin, A., Oganesyan, D., Emtsev, D., Burnaev, E.: Barcodes as summary of loss function topology. Doklady Rossijskoj akademii nauk. Matematika, informatika, processy upravlenia 514(2), 196-211 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.5, + 0.785, + 0.54 + ], + "angle": 0, + "content": "7. Barannikov, S., Trofimov, I., Balabin, N., Burnaev, E.: Representation topology divergence: A method for comparing neural network representations. ICML 2022, arXiv:2201.00058 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.541, + 0.785, + 0.581 + ], + "angle": 0, + "content": "8. Barannikov, S., Trofimov, I., Sotnikov, G., Trimbach, E., Korotin, A., Filippov, A., Burnaev, E.: Manifold topology divergence: a framework for comparing data manifolds. Advances in Neural Information Processing Systems 34, 7294-7305 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.582, + 0.785, + 0.622 + ], + "angle": 0, + "content": "9. Barannikov, S., Voronkova, D., Trofimov, I., Korotin, A., Sotnikov, G., Burnaev, E.: Topological obstructions in neural networks learning. arXiv preprint arXiv:2012.15834 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.623, + 0.785, + 0.676 + ], + "angle": 0, + "content": "10. Carpenter, A.E., Jones, T.R., Lamprecht, M.R., Clarke, C., Kang, I.H., Friman, O., Guertin, D.A., Chang, J.H., Lindquist, R.A., Moffat, J., et al.: Cellprofiler: image analysis software for identifying and quantifying cell phenotypes. Genome biology 7, 1-11 (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.677, + 0.785, + 0.717 + ], + "angle": 0, + "content": "11. Carrière, M., Chazal, F., Glisse, M., Ike, Y., Kannan, H., Umeda, Y.: Optimizing persistent homology based functions. In: International Conference on Machine Learning. pp. 1294-1303. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.718, + 0.785, + 0.771 + ], + "angle": 0, + "content": "12. Carrière, M., Chazal, F., Ike, Y., Lacombe, T., Royer, M., Umeda, Y.: Perslay: A neural network layer for persistence diagrams and new graph topological signatures. In: International Conference on Artificial Intelligence and Statistics. pp. 2786-2796. PMLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.772, + 0.785, + 0.8 + ], + "angle": 0, + "content": "13. Chazal, F., Michel, B.: An introduction to topological data analysis: fundamental and practical aspects for data scientists. arXiv preprint arXiv:1710.04019 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "14. Chen, Y., Segovia, I., Gel, Y.R.: Z-gcnets: Time zigzags at graph convolutional networks for time series forecasting. In: International Conference on Machine Learning. pp. 1684-1694. PMLR (2021)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.254, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.386, + 0.127 + ], + "angle": 0, + "content": "I. Trofimov et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.204 + ], + "angle": 0, + "content": "15. Chen, Y., Segovia Dominguez, I., Coskunuzer, B., Gel, Y.: Tamp-s2gcnets: When time-aware multipersistence meets spatio-supra graph convolutional nets while forecasting time series. In: The International Conference on Learning Representations (ICLR) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.205, + 0.787, + 0.245 + ], + "angle": 0, + "content": "16. Demir, A., Coskunuzer, B., Gel, Y., Segovia-Dominguez, I., Chen, Y., Kiziltan, B.: Todd: Topological compound fingerprinting in computer-aided drug discovery. Advances in Neural Information Processing Systems 35, 27978-27993 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.246, + 0.787, + 0.3 + ], + "angle": 0, + "content": "17. Funke, J., Tschopp, F., Grisaitis, W., Sheridan, A., Singh, C., Saalfeld, S., Turaga, S.C.: Large scale image segmentation with structured loss based deep learning for connectome reconstruction. IEEE transactions on pattern analysis and machine intelligence 41(7), 1669-1680 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.301, + 0.787, + 0.328 + ], + "angle": 0, + "content": "18. Gelfand, S.I., Manin, Y.I.: Methods of homological algebra. Springer Science & Business Media (2002)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.328, + 0.787, + 0.383 + ], + "angle": 0, + "content": "19. Gupta, S., Hu, X., Kaan, J., Jin, M., Mpoy, M., Chung, K., Singh, G., Saltz, M., Kurc, T., Saltz, J., et al.: Learning topological interactions for multi-class medical image segmentation. In: European Conference on Computer Vision. pp. 701-718. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.787, + 0.425 + ], + "angle": 0, + "content": "20. Gupta, S., Zhang, Y., Hu, X., Prasanna, P., Chen, C.: Topology-aware uncertainty for image segmentation. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.787, + 0.467 + ], + "angle": 0, + "content": "21. Hatamizadeh, A., Nath, V., Tang, Y., Yang, D., Roth, H.R., Xu, D.: Swin unetr: Swin transformers for semantic segmentation of brain tumors in mri images. arXiv preprint arXiv:2201.01266v1 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.787, + 0.495 + ], + "angle": 0, + "content": "22. Hofer, C., Graf, F., Rieck, B., Niethammer, M., Kwitt, R.: Graph filtration learning. In: International Conference on Machine Learning. pp. 4314-4323. PMLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.787, + 0.522 + ], + "angle": 0, + "content": "23. Horn, M., De Brouwer, E., Moor, M., Moreau, Y., Rieck, B., Borgwardt, K.: Topological graph neural networks. arXiv preprint arXiv:2102.07835 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.787, + 0.55 + ], + "angle": 0, + "content": "24. Hu, X., Li, F., Samaras, D., Chen, C.: Topology-preserving deep image segmentation. Advances in neural information processing systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.787, + 0.578 + ], + "angle": 0, + "content": "25. Hu, X., Wang, Y., Fuxin, L., Samaras, D., Chen, C.: Topology-aware segmentation using discrete Morse theory. arXiv preprint arXiv:2103.09992 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.787, + 0.605 + ], + "angle": 0, + "content": "26. Kaczynski, T., Mischaikow, K.M., Mrozek, M.: Computational homology, vol. 3. Springer (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.787, + 0.633 + ], + "angle": 0, + "content": "27. Luo, Y., Shi, L., Thost, V.: Improving self-supervised molecular representation learning using persistent homology. arXiv preprint arXiv:2311.17327 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.787, + 0.674 + ], + "angle": 0, + "content": "28. Menze, B.H., Jakab, A., Bauer, S., Kalpathy-Cramer, J., Farahani, K., Kirby, Justin, e.a.: The multimodal brain tumor image segmentation benchmark (brats). IEEE Transactions on Medical Imaging (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.674, + 0.787, + 0.702 + ], + "angle": 0, + "content": "29. Moor, M., Horn, M., Rieck, B., Borgwardt, K.: Topological autoencoders. In: International conference on machine learning. pp. 7045-7054. PMLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.787, + 0.743 + ], + "angle": 0, + "content": "30. Pérez, J.B., Hauke, S., Lupo, U., Caorsi, M., Dassatti, A.: giotto-ph: A python library for high-performance computation of persistent homology of vietoris-rips filtrations (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.787, + 0.799 + ], + "angle": 0, + "content": "31. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "32. Stucki, N., Paetzold, J.C., Shit, S., Menze, B., Bauer, U.: Topologically faithful image segmentation via induced matching of persistence barcodes. In: International Conference on Machine Learning. pp. 32698-32727. PMLR (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.482, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Scalar Function Topology Divergence" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.203 + ], + "angle": 0, + "content": "33. Tang, Y., Yang, D., Li, W., Roth, H.R., Landman, B., Xu, D., Nath, V., Hatamizadeh, A.: Self-supervised pre-training of swim transformers for 3d medical image analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.204, + 0.787, + 0.245 + ], + "angle": 0, + "content": "34. Trofimov, I., Cherniavskii, D., Tulchinskii, E., Balabin, N., Burnaev, E., Barannikov, S.: Learning topology-preserving data representations. arXiv preprint arXiv:2302.00136 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.246, + 0.786, + 0.286 + ], + "angle": 0, + "content": "35. Wagner, H., Chen, C., Vuçini, E.: Efficient computation of persistent homology for cubical data. In: Topological methods in data analysis and visualization II: theory, algorithms, and applications, pp. 91-106. Springer (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.787, + 0.342 + ], + "angle": 0, + "content": "36. Waibel, D.J., Atwell, S., Meier, M., Marr, C., Rieck, B.: Capturing shape information with multi-scale topological loss terms for 3d reconstruction. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 150-159. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.786, + 0.37 + ], + "angle": 0, + "content": "37. Waibel, D.J., Kiermeyer, N., Atwell, S., Sadafi, A., Meier, M., Marr, C.: Shapr predicts 3d cell shapes from 2d microscopic images. Iscience 25(11) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.786, + 0.397 + ], + "angle": 0, + "content": "38. Watts, D.J., Strogatz, S.H.: Collective dynamics of 'small-world'networks. nature 393(6684), 440-442 (1998)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.787, + 0.438 + ], + "angle": 0, + "content": "39. Wu, J., Wang, Y., Xue, T., Sun, X., Freeman, B., Tenenbaum, J.: Marrnet: 3d shape reconstruction via 2.5 d sketches. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.44, + 0.787, + 0.482 + ], + "angle": 0, + "content": "40. Yan, Z., Ma, T., Gao, L., Tang, Z., Chen, C.: Link prediction with persistent homology: An interactive view. In: International conference on machine learning. pp. 11659-11669. PMLR (2021)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.482 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Scalar Function Topology Divergence_ Comparing Topology of 3D Objects/228bb063-0fe3-438c-a000-0e252dd17cc3_origin.pdf b/2024/Scalar Function Topology Divergence_ Comparing Topology of 3D Objects/228bb063-0fe3-438c-a000-0e252dd17cc3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f2d534152f69f1a8dc1f2b1bd3bde7284a70cfae --- /dev/null +++ b/2024/Scalar Function Topology Divergence_ Comparing Topology of 3D Objects/228bb063-0fe3-438c-a000-0e252dd17cc3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:095256e03e4300c9925bc717b551bab4fbcc637e097edfe263838e3605b17abe +size 4807650 diff --git a/2024/Scalar Function Topology Divergence_ Comparing Topology of 3D Objects/full.md b/2024/Scalar Function Topology Divergence_ Comparing Topology of 3D Objects/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f7d78fba9fc05370ccd1edceb09a7df0d4950d70 --- /dev/null +++ b/2024/Scalar Function Topology Divergence_ Comparing Topology of 3D Objects/full.md @@ -0,0 +1,444 @@ +# Scalar Function Topology Divergence: Comparing Topology of 3D Objects + +Ilya Trofimov1, Daria Voronkova1,2, Eduard Tulchinskii1,4, Evgeny Burnaev1,2, and Serguei Barannikov1,3 + +$^{1}$ Skolkovo Institute of Science and Technology, Moscow, Russia {ilya.trofimov,darya.voronkova, eduard.tulchinskiy,e.burnaev, s.barannikov}@skoltech.ru + +2 AIRI, Moscow, Russia + +$^{3}$ CNRS, IMJ, Paris Cité University, France + +4 AI Foundation and Algorithm Lab, Moscow, Russia + +Abstract. We propose a new topological tool for computer vision - Scalar Function Topology Divergence (SFTD), which measures the dissimilarity of multi-scale topology between sublevel sets of two functions having a common domain. Functions can be defined on an undirected graph or Euclidean space of any dimensionality. Most of the existing methods for comparing topology are based on Wasserstein distance between persistence barcodes and they don't take into account the localization of topological features. The minimization of SFTD ensures that the corresponding topological features of scalar functions are located in the same places. The proposed tool provides useful visualizations depicting areas where functions have topological dissimilarities. We provide applications of the proposed method to 3D computer vision. In particular, experiments demonstrate that SFTD as an additional loss improves the reconstruction of cellular 3D shapes from 2D fluorescence microscopy images, and helps to identify topological errors in 3D segmentation. Additionally, we show that SFTD outperforms Betti matching loss in 2D segmentation problems. The code is publicly available: https://github.com/IlyaTrofimov/SFTD. + +Keywords: Topological Data Analysis $\cdot$ 3D Computer Vision $\cdot$ Geometric Deep Learning + +# 1 Introduction + +A common situation in computer vision is a structural prediction defined over $\mathbb{R}^n$ . Examples include image segmentation [24, 25, 31, 32] and 3D shape reconstruction [36, 37, 39] problems. However, in most of the cases the comparison between the prediction and a ground truth is done pixel(voxel)-wise by metrics like Dice score, cross-entropy, etc. These metrics are not capable of providing detailed topological correctness, i.e. preserving multi-scale shape patterns: cavities, holes, connectivity patterns, etc. Typically, predictions involve scalar functions $f:\mathbb{R}^n\to \mathbb{R}$ and the final shape of an object is defined by thresholding $f(x)$ . + +There is a need for topological-aware comparison metrics for scalar functions, which take into account the topology of sublevel sets $f^{-1}((-\infty, \varepsilon])$ (the analysis for superlevel sets is similar). Existing tools for this problem don't take into account the localization of topological features [24,36] or are limited to 2D space [32]. Our contributions are as follows: + +1. We introduce new topological tools: F-Cross-Barcodes and Scalar Function Topology Divergence (SFTD), for a multi-scale topology-aware comparison of scalar functions by tracking differences in their sublevel sets topology. +2. A distinctive feature of F-Cross-Barcodes and SFTD is that they take into account the localization of topological features; +3. We describe a visualization technique for F-Cross-Barcodes: a technique highlights with points areas where differences in topological features are localized: +4. We propose an algorithm for SFTD differentiation; +5. By doing computational experiments, we show that SFTD as an additional loss term improves 3D shape reconstruction, and helps to identify topological errors in 3D segmentation. +6. We also show that SFTD outperforms Betti matching loss in terms of topological accuracy in 2D segmentation while being $\approx 35$ times faster, see Appendix G. + +# 2 Related work + +Tools of topological data analysis find application in various machine learning problems: evaluation of deep generative models [8], comparison of neural representations [7], dimensionality reduction [29,34], graph neural networks [12,23,40], self-supervised learning [27], molecular data analysis [16], analysis of loss functions in deep learning [6,9], spatial-temporal graph neural network tasks [14,15], crowd localization [1], to name a few. + +Several works used topological tools to enhance image segmentation [20, 24, 25, 32]. However, proposed algorithms for comparison ground truth and predicted images either don't take into account localization of topological features [24], or are limited to 2D [32]. A method by [25] highlights in images topologically critical and error-prone structures using discrete Morse theory, and [20] proposes a method for inferring structure-aware uncertainty. The importance of topological interactions for multi-class medical image segmentation is shown in [19]. + +# 3 Preliminaries + +We study scalar functions $f: \mathbb{R}^n \to \mathbb{R}$ by assessing the topology of its sublevel sets. Consider a 1D function depicted in Figure 1. It has two local minima $y_1, y_2$ and a local maximum $y_3$ , such that $y_1 < y_2 < y_3$ . One can study sublevel sets + +![](images/3dd7950c826053b97be8ccddecdbe1c53dfc7631ee8e4200b6be564b361423e3.jpg) +Fig.1: 1D function and its sublevel sets. + +$X_{\varepsilon} = f^{-1}((-\infty ,\varepsilon ])$ . For the scalar function $f$ , changes of $X_{\varepsilon}$ for increasing $\varepsilon$ can be intuitively understood as a "filling the pool", see Figure 1. There are four distinct situations: + +1. For $\varepsilon < y_{1}$ , $X_{\varepsilon} = \varnothing$ +2. For $y_{1} \leq \varepsilon < y_{2}$ , $X_{\varepsilon}$ is a single segment; +3. For $y_{2} \leq \varepsilon < y_{3}$ , $S_{\varepsilon}$ consists of 2 segments; +4. Finally, for $\varepsilon \geq y_{3}$ , $X_{\varepsilon}$ is a single segment. + +Note, that changes in topology of $X_{\varepsilon}$ happen when $\varepsilon$ is a critical point of $f$ . In the general case, smooth scalar functions defined on smooth manifolds are studied by Morse theory [6]. In practice, we assess functions through their values on a finite set of points, e.g. a lattice (square grid graph) in $\mathbb{R}^n$ . + +# 3.1 Scalar functions defined on a graph + +In the most general form, we are interested in topology-aware comparison of scalar functions $f: \mathcal{V} \to \mathbb{R}$ defined on vertices $\mathcal{V}$ of an undirected graph $\mathcal{G} = (\mathcal{V}, \mathcal{E})$ . Specifically, we study topological features of subgraphs $\{v \in \mathcal{V} \mid f(v) \leq \varepsilon\}$ connected components, cycles, etc. + +For each clique $C$ in the graph $\mathcal{G}$ , we define a filtration function $T(C) = \max_{i\in C}f(i)$ . By a clique complex $\mathcal{C}$ , we mean a set of cliques where each sub-clique also belongs to a complex5. For some $\varepsilon$ , $T(C)$ defines a clique complex $\mathcal{C}_j = \{C\mid T(C)\leq \varepsilon \}$ , where $f(j) = \varepsilon$ . Topological features for a subgraph $\{v\in \mathcal{V}\mid f(v)\leq \varepsilon \}$ are precisely expressed by homology groups. Overall, the filtration function $T(C)$ defines the lower-star filtration of the graph's clique complex: + +$$ +\varnothing \subseteq \mathcal {C} _ {0} \subseteq \mathcal {C} _ {1} \subseteq \dots \subseteq \mathcal {C} _ {n}, +$$ + +where $\mathcal{C}_n$ is a clique complex of $\mathcal{G}$ . The filtration is said to be induced by the vertex function $f$ . + +A given topological feature "appears" at some $\varepsilon_{1}$ and "dies" at some $\varepsilon_{2}$ . These topological features, together with their birth and death times can be calculated via tools of topological data analysis. The multi-set of the corresponding intervals $(\varepsilon_{1},\varepsilon_{2})$ is called the persistence barcode, which depicts topological features of the graph for all scales. The whole theory is dubbed persistent homology [5, 13]. + +# 4 Method + +# 4.1 Scalar functions defined on a graph + +To compare topological features of sublevel sets of functions $f,g:\mathcal{G}\to \mathbb{R}$ we construct a graph $\tilde{\mathcal{G}} = (\tilde{\mathcal{V}},\tilde{\mathcal{E}})$ with doubled set of vertices from $\mathcal{G}$ and one extra + +![](images/ebe692964d3d7b3377c6e58b57b047b2eb36ef19282a72aac14b2a4969134bbc.jpg) +(a) The doubled graph $\tilde{\mathcal{G}}$ + +![](images/747418fb38f3eac499b0fe55b32d5275d37f303be1480fb7a412b53b5d734265.jpg) +(b) The extended $n + 1$ dimensional lattice $\tilde{\mathcal{L}}$ . +Fig. 2: A visualization of $\tilde{\mathcal{G}}$ , $\tilde{\mathcal{L}}$ . + +vertex $O$ , see Figure 2a. Formally, $\tilde{\mathcal{V}} = \{A_1,\dots ,A_n,A_1',\dots ,A_n',O\}$ . Edges $\tilde{\mathcal{E}}$ are defined as follows: + +$$ +\begin{array}{l} \left(A _ {i}, A _ {j}\right) \in \mathcal {E} \Rightarrow \left(A _ {i}, A _ {j}\right) \in \tilde {\mathcal {E}}, \left(A _ {i} ^ {\prime}, A _ {j} ^ {\prime}\right) \in \tilde {\mathcal {E}} \\ \left(A _ {i}, A _ {j}\right) \in \mathcal {E}, i < j \Rightarrow \left(A _ {i}, A _ {j} ^ {\prime}\right) \in \tilde {\mathcal {E}} \\ \left(O, A _ {i}\right) \in \tilde {\mathcal {E}}, \left(A _ {i}, A _ {i} ^ {\prime}\right) \in \tilde {\mathcal {E}} \text {f o r a l l} i \\ \end{array} +$$ + +The vertex function $\tilde{f}$ is defined as follows: + +$$ +\begin{array}{l} \tilde {f} (A _ {i}) = f (i) \\ \tilde {f} (A _ {i} ^ {\prime}) = \min (f (i), g (i)), \tilde {f} (O) = \min _ {f, g}, \\ \end{array} +$$ + +where $\min_{f,g}$ is the minimum value of both $f,g$ in the nodes of the graph $\mathcal{G}$ . + +Definition. The $F$ -Cross-Barcode $_k(f, g)$ for two scalar functions $f, g$ defined on vertices of an undirected graph $\mathcal{G}$ is the $k$ -th persistence barcode for the doubled graph $\tilde{\mathcal{G}}$ and the lower-star filtration induced by $\tilde{f}$ . + +To put it simply, each interval in the $F$ -Cross-Barcode $_k(f, g)$ denotes a range of functions' sub-level sets where a given topological feature exists in the function $f$ but doesn't exist in the function $g$ , see further Appendix A. Technically, the desired persistence barcode is calculated via giotto-ph library [30]. For this software, values of a filtration function must be defined on vertices and edges. Algorithm 1 summarizes the aforementioned. + +# 4.2 Scalar functions defined on a lattice + +A frequent situation is a function having a domain in $\mathbb{R}^n$ . We assume that the function is approximated by its restriction on a sufficiently small grid. Algorithm 1 can also be applied in this case after a space triangulation. However, the number of simplices grows very quickly in high dimensions (see discussion in [35]), and Algorithm 1 has a high computational cost. That is why, we consider cubical complexes [26] instead of simplicial complexes. Specifically, we consider a regular square grid graph ( $n$ -dimensional lattice) and use values of functions + +# Algorithm 1 F-Cross-Barcode $_k$ (scalar function $k$ -th cross-barcode) + +Input: Scalar functions $f, g$ defined on vertices of a graph $\mathcal{G} = (\mathcal{V}, \mathcal{E})$ . $|\mathcal{V}| = n$ . + +Require: $\mathsf{PB}_k(\cdot)$ - an algorithm to calculate $k$ -th persistence barcode from a matrix containing values of a filtration function on graph: on vertices - on a diagonal, on edges - off a diagonal. + +1. Create a $n \times n$ matrix $F$ for the function $f$ : $F_{ij} = f(i)$ for $i = j$ , $\max(f(i), f(j))$ for $i \neq j$ and $(i, j) \in \mathcal{E}$ , $+\infty$ otherwise. +2. Create a $n \times n$ matrix $G$ for the function $g$ analogously; +3. Initialize $c_{+\infty}$ to a vector of $n$ elements, all equal to $+\infty$ ; +4. Create $F_{+}$ , a matrix derived from $F$ with a lower-diagonal part replaced by $+\infty$ ; +5. Create $F_{diag}$ , a vector containing the diagonal elements of $F$ ; +6. Compute $\min (F,G)$ , the element-wise minimum of $F$ and $G$ ; +7. Compute $\min_{F,G}$ , the minimum value of all elements in $F$ and $G$ ; +8. Compute the $(2n + 1)\times (2n + 1)$ block matrix $m$ as follows: + +$$ +m \leftarrow \left( \begin{array}{c c c} \min (F, G) & F _ {+} ^ {\intercal} & c _ {+ \infty} ^ {\intercal} \\ F _ {+} & F & F _ {d i a g} ^ {\intercal} \\ c _ {+ \infty} & F _ {d i a g} & \min _ {F, G} \end{array} \right) +$$ + +9. Compute: F-Cross-Barcode $_k(f, g) \gets \mathsf{PB}_k(m)$ + +Return: F-Cross-Barcode $_k(f, g)$ , a multi-set of ("birth", "death") intervals. + +in nodes of this grid: $f(j_{1},\ldots ,j_{n}),g(j_{1},\ldots ,j_{n})$ . To compare two function $f,g$ we create an extended $n + 1$ dimensional lattice $\tilde{\mathcal{L}}$ with one extra dimension of size 3, see Figure 2b. The vertex function $\tilde{f}:\tilde{\mathcal{L}}\to \mathbb{R}$ is defined as follows: + +$$ +\tilde {f} (i, j _ {1}, \ldots , j _ {n}) = \left\{ \begin{array}{l l} \min _ {f, g} & \text {f o r} i = 0 \\ f (j _ {1}, \ldots , j _ {n}) & \text {f o r} i = 1 \\ \min (f (j _ {1}, \ldots , j _ {n}), g (j _ {1}, \ldots , j _ {n})) & \text {f o r} i = 2 \end{array} \right. +$$ + +Similarly to functions on graphs, $F$ -Cross-Barcode $_k(f,g)$ for functions on lattices is defined as the $k$ -th persistence barcode of the filtered cubical complex with the filtration induced by the vertex function $\tilde{f}$ . We use the GUDHI $^6$ library which implements efficient algorithms for cubical complexes [35]. Note that $F$ -Cross-Barcode $_k(f,g)$ for functions restricted on a lattice and a triangulation of the same space are not equivalent, but would be similar. + +# 4.3 Scalar Function Topology Divergence + +The $F$ -Cross-Barcode $_k(f, g)$ depicts topological differences of functions' sublevel sets at multiple scales. The $F$ -Cross-Barcode $_k(f, g)$ , as a regular barcode, is a multi-set of intervals. One requires some vectorization technique to inject it into a machine learning pipeline. If one needs to describe a discrepancy between $f, g$ + +by an integral numerical characteristic, we propose a scalar function topology divergence: + +$$ +\mathrm {S F T D} _ {k} (f, g) = \sum_ {(b, d) \in \mathrm {F} - \text {C r o s s - B a r c o d e} _ {k} (f, g)} | | d - b | | ^ {p}, \tag {1} +$$ + +that is the sum of lengths of intervals from the $F$ -Cross-Barcode $_k(f, g)$ to $p$ -th power, where typically $p = 1, 2$ . SFTD is not symmetric, and we often use an average of $\mathrm{SFTD}_k(f, g)$ and $\mathrm{SFTD}_k(g, f)$ in experiments below. For different dimensions $k$ , $\mathrm{SFTD}_k(f, g)$ can be used either separately, or the sum can be taken to reflect discrepancies in all types of topological features (connected components, cycles, voids, etc.). Computational complexity of SFTD is discussed in Appendix C. + +# 4.4 SFTD basic properties + +Proposition 1. If $SFTD_{k}(f,g) = SFTD_{k}(g,f) = 0$ for all $k\geq 0$ , then the sublevel sets barcodes of functions $f$ and $g$ are the same in any degree. Moreover, in this case their topological features are located in the same places: the inclusions $f^{-1}((- \infty ,\varepsilon ])\subseteq (\min (f,g))^{-1}((- \infty ,\varepsilon ])$ , $g^{-1}((- \infty ,\varepsilon ])\subseteq (\min (f,g))^{-1}((- \infty ,\varepsilon ])$ , induce homology isomorphisms for any threshold $\varepsilon$ . + +# Proposition 2. (SFTD Stability) + +(a) For any scalar functions $f, \tilde{f}, g, \tilde{g}$ on a grid, or on a graph vertex set: + +$$ +\begin{array}{l} d _ {B} (F - C r o s s - B a r c o d e _ {k} (f, g), F - C r o s s - B a r c o d e _ {k} \left(f ^ {\prime}, g ^ {\prime}\right)) \leq \\ \max (\max _ {i} | f (i) - f ^ {\prime} (i) |, \max _ {i} | g (i) - g ^ {\prime} (i) |). \\ \end{array} +$$ + +(b) For any pair of scalar functions $f, \hat{f}'$ : + +$$ +\| F \text {- C r o s s - B a r c o d e} _ {k} (f, f ^ {\prime}) \| _ {B} \leq \max _ {i} | f (i) - f ^ {\prime} (i) |, +$$ + +where $d_B$ denotes the bottleneck distance between persistence barcodes [13]. We give the proofs in Appendix A. As a consequence, SFTD is robust to noise. + +# 4.5 Visualization of F-Cross-Barcodes + +$F$ -Cross-Barcodes can be used for localization and visualization of topological differences between scalar functions. Recall the filtered cubical7 complex: + +$$ +\varnothing \subseteq \mathcal {C} _ {0} \subseteq \mathcal {C} _ {1} \subseteq \dots \subseteq \mathcal {C} _ {n}. +$$ + +Since the filtration is induced by a vertex function $\tilde{f}:\tilde{\mathcal{L}}\to \mathbb{R}$ , the filtered cubical complex corresponds to the sequence: + +$$ +\tilde {f} (i _ {0}) \leq \tilde {f} (i _ {1}) \leq \dots \leq \tilde {f} (i _ {n}). +$$ + +If $\tilde{f}(i_k) = \tilde{f}(i_m)$ , the order is defined by indexes $i_k, i_m$ . Thus, each segment in the barcode - a birth/death pair - is associated with two vertices $v_b, v_d \in \tilde{\mathcal{L}}$ . Recall that $\tilde{\mathcal{L}}$ is an extended $n + 1$ dimensional lattice (see Section 4.2), where the first dimension is of size 3, and the rest of dimensions are of the same size as the lattice inside the domain of $f, g$ . So, neglecting the first dimension, we can map the rest of indices $j_1, \ldots, j_n$ of birth/death vertices $v_b, v_d$ right onto original $n$ -dimensional lattice. Later, in the experimental section, we shall provide such visualizations, where birth/death events are depicted by orange/red dots. + +# 4.6 SFTD differentiation + +Calculation of $\mathrm{SFTD}_k(f,g)$ involves a filtered simplicial(cubical) complex $\mathcal{C}$ with a vertex function $\tilde{f} (i),i\in \tilde{\mathcal{V}}$ . Also, since we apply the sub-level set filtration, for every simplex(cube) $\sigma \colon T(\sigma) = \tilde{f} (i_{\sigma})$ , where $i_{\sigma} = \operatorname{argmax}_{i\in \sigma}\tilde{f} (i)$ . Each interval $(b,d)$ in a $F$ -Cross-Barcode corresponds to some pair of simplices(cubes) $(\sigma_1,\sigma_2)$ which join the simplicial(cubical) complex $\mathcal{C}$ at moments $b = T(\sigma_1),d = T(\sigma_2)$ Generally, we consider $f,g$ (and consequently $\tilde{f}$ ) to be parametric functions of $w$ , for example, neural networks. By differentiating (1) we obtain: + +$$ +\frac {\partial \operatorname {S F T D} _ {k} (f , g)}{\partial w} = \sum_ {\left(\sigma_ {1}, \sigma_ {2}\right) \in \Sigma_ {k}} p \left(\frac {\partial \tilde {f} \left(i _ {\sigma_ {2}}\right)}{\partial w} - \frac {\partial \tilde {f} \left(i _ {\sigma_ {2}}\right)}{\partial w}\right) ^ {p - 1}, \tag {2} +$$ + +where $\Sigma_{k}$ is a collection of "birth", "death" simplex pairs corresponding to intervals in the $F$ -Cross-Barcode $_k(f, g)$ . + +By construction, $\tilde{f}(i)$ can be either $f(i')$ or $g(i')$ for some mapping $i' = r(i)$ . Formally, introducing a binary function $c(\cdot)$ for this choice: + +$$ +\tilde {f} (i) = c (i) f (r (i)) + (1 - c (i)) g (r (i)), \tag {3} +$$ + +and + +$$ +\frac {\partial \tilde {f} (i)}{\partial w} = b (i) \frac {\partial f (r (i))}{\partial w} + (1 - b (i)) \frac {\partial g (r (i))}{\partial w}. \tag {4} +$$ + +Finally, the desirable gradient $\partial \mathrm{SFTD}_k(f,g) / \partial w$ can be calculated by combining (2), (4) via a chain rule. With a common assumption that all the $f(i),g(i)$ are distinct, functions $c,r$ don't change in a vicinity of $w$ , and $\mathrm{SFTD}_i(f,g)$ is differentiable w.r.t. $w$ [11, 22]. On a high level, the differentiation is done similarly to [11, 34]. + +# 5 Experiments + +In the experimental section, we are going to answer three research questions: + +Q1: Does SFTD and F-Cross-Barcodes provide a better topologically-aware comparison of scalar functions than Wasserstein distance between regular persistence barcodes? +Q2: Does F-Cross-Barcodes help to visualize and localize topological differences? +Q3: Is SFTD, as a loss function, beneficial for computer vision problems? + +![](images/d2ab50479442e57f66504be1fe8b88937b063abfda85749b953715ee9645e657.jpg) +(a) Scalar func. $f_{1}$ . + +![](images/b70f9280df963f3a824aa917709274b1ea09aeb08e0da3354e1cb892d49e6bbe.jpg) +(b) Scalar func. $f_{2}$ . + +![](images/f3d2c1e646d9b98fe2032cb8621e3a02038d4db259739dae6c2b4765898ea058.jpg) +(c) Vis. diff. + +![](images/a754ee879426c914d789658c3eb148241a5dc345ad01d9c3cc564e6bda99edd1.jpg) +(d) Pers. barc. of $f_{1}$ . + +![](images/b31e3284c1626ff870a0a5473ad4c7d58819b22497ea9e2786940ecaf07fbcd3.jpg) +(e) Pers. barc. of $f_{2}$ . + +![](images/a6d5b7d1918b0c711b9b57a328283c4fd0a3d493a5ebbfb3fced4922e98715b9.jpg) +(f) F-Cross-Barcode +Fig. 3: Two scalar functions (a), (b) are indistinguishable by their persistence barcodes (d), (e). At the same time, $\mathrm{SFTD}_0(f_1,f_2) = 4.6$ see (f). + +# 5.1 Synthetic data + +To illustrate the performance of the proposed tools, we start with synthetic data. Functions under comparison have coinciding persistence barcodes of all degrees and they are indistinguishable by standard Wasserstein matching. + +Functions with local minima in 2D. First, we compare two scalar function on $64 \times 64$ grid, each of them having 3 local minima but in different locations, see Figure 3, (a), (b). Persistence barcodes of both functions are equal, see Figure 3, (d), (e): they contain two segments from $H_0$ corresponding to local minima. Here and further, essential segments (which die at $+\infty$ ) are omitted. The topology of the two functions is indistinguishable by standard persistence barcodes and their Wasserstein distance is zero. At the same time, F-Cross-Barcode between these scalar functions is non-empty for $H_0$ , see Figure 3, (f). The visualization technique correctly points at places with different topology - the centers and borders of the distinguishing local minima. + +Lattices in 2D. Next, we compare two lattice-style 2D scalar functions, see Figure 4, (a), (b). Lattices have defects in different places (top left and bottom right respectively). The persistence barcodes of both lattices are identical, see Figure 4, (d), (e): they contain 8 segments from $H_{1}$ corresponding to independent cycles. At the same time, $F$ -Cross-Barcode between two scalar functions is nonempty, see Figure 4, (f). The visualization technique correctly points at places with defects. + +Concentric spheres in 3D. Finally, we study the performance of the proposed method by comparing 3D scalar functions. We compare two scalar functions, + +![](images/69856d4fcb13b7dd2e2d930a3e5643239ccc2db98283f567f6541dd3a2580df2.jpg) +(a) Scalar func. $f_{1}$ . + +![](images/0d5fdc209618e10ffc645065dd5c4b756addc53b05a881ca4922b28ffd87d1bb.jpg) +(b) Scalar func. $f_{2}$ . + +![](images/f4ceb1ae7e305e8818135444cafec40f30f0f66a97538e2738035e2434c7fab6.jpg) +(c) Vis. diff. + +![](images/3e51a20a812773a5e116b5370ae3c193c029babc5ccc44712bd4640b75b38b68.jpg) +(d) Pers. barc. of $f_1$ . + +![](images/eb57e2f6b538b268e340fd22af3b0a3bc60fd1edc058695c44775a6fd6c450de.jpg) +(e) Pers. barc. of $f_{2}$ . + +![](images/f465b508c3f4aa142f5e1a82f528282f6e0de4a9efb6e81032d4c98005f1a1a7.jpg) +(f) F-Cross-Barcode. +Fig. 4: Two scalar functions (a), (b) are indistinguishable by their persistence barcodes (c), (d). At the same time, $\mathrm{SFTD}_1(f_1,f_2) = 1$ , see (e). + +which equal $-1$ in a vicinity of two concentric spheres and a bridge connecting them, which is depicted in green. The bridge is located in different places (above or below of the smallest sphere). In other areas of space, the scalar functions equal 0. We sample a $32^3$ grid from the unit cube, see Figure 5 (a), (b) and compare functions restricted on this grid via $F$ -Cross-Barcode. Again, regular persistence barcodes of two spheres with bridges are the same, see Figure 5, (d), (e): they contain 2 segments from $H_2$ corresponding to two cavities. But $F$ -Cross-Barcode is non-empty, see Figure 5, (f). The visualization technique correctly highlights areas with topological differences, see Figure 5 (c), where the central vertical slice is shown. + +# 5.2 Comparing Eigenvectors of Graph Laplacian + +In the most general form, the proposed tools can compare scalar functions $f: \mathcal{V} \to \mathbb{R}$ defined on vertices $\mathcal{V}$ of an undirected graph $\mathcal{G} = (\mathcal{V}, \mathcal{E})$ . The normalized graph Laplacian is the matrix $L = D^{-1/2}(D - A)D^{-1/2}$ where $A$ is an adjacency matrix and $D$ is a diagonal matrix of node degrees. Eigenvectors $e_i$ and eigenvalues $\lambda_i$ of $L$ describe intrinsic properties of the graph $\mathcal{G}$ . Eigenvectors can be considered as scalar functions defined on graph's vertices $\mathcal{V}$ . For eigenvectors: $(e_i, e_j) = 0$ and their norm is not defined; one can put $||e_i|| = 1$ for convenience. Thus, $e_i$ can't be distinguished by Euclidean distance or cosine distance. Alternatively, we can study $e_i$ as functions on vertices and compare them via SFTD. + +We consider a random graph with 30 nodes, see Figure 6. Figure 7 presents symmetrized $\sum_{k=0,1} \mathrm{SFTD}_k(e_i, e_j)$ via a heatmap, where $e_i, e_j$ are graph Laplacian eigenvectors ordered by eigenvalues. We observed the following interesting patterns: block structure, high topological similarity in the top left block, low + +![](images/e88065fab56b5ab72ee89ef7a28fdb8e425d191fe5a58b44b17bab39f47d5acf.jpg) +(a) Scalar func. $f_{1}$ . + +![](images/52af2e937923e1cb598b78cd11c5b3b55bafb10c4be3a871ab820e35f05ec85b.jpg) +(b) Scalar func. $f_{2}$ . + +![](images/288d13bb79d66d0d21002e1a4e481af7b027fbf7ca53b9cf52bd59526f78e706.jpg) +(c) Vis. diff. + +![](images/00d52bcc64268c36432279fab4859f834b0afa42653cf9aa435796e4bb6fc218.jpg) +(d) Pers. barc. of $f_{1}$ . + +![](images/4c239fb411e8e02483fbb62d468775514cc6ab94496356f93451cf40ced6ace4.jpg) +(e) Pers. barc. of $f_{2}$ . + +![](images/02497798d1a0434f9433960f9c691ae3fb834b7714344ec13c6da181f562bac3.jpg) +(f) F-Cross-Barcode. +Fig. 5: Two scalar fields (a), (b) are indistinguishable by their persistence barcodes (d), (e). At the same time, $\mathrm{SFTD}_1(f_1,f_2) = 1$ , see (f). + +Table 1: Errors of 3D shape reconstruction for "Red blood cell" dataset. + +
METHOD1-IoUVOLUMESURFACEROUGH.H0H1H2
SHAPR [37]0.4870.3530.2430.3610.0980.2410.065
SHAPR+W.D. [36]0.4670.2910.1840.2940.0980.2280.032
SHAPR+Σi=0,1,2SFTDi0.4490.2940.1760.2820.0790.2220.065
SHAPR+Σi=0,1,2,3SFTDi0.4530.2660.1890.2760.0790.2170.025
+ +topological similarity in the bottom right block, and very interesting diagonal lines parallel to the main diagonal. Figure 8 shows two the most similar eigenvectors, Figure 9 - two the most dissimilar ones, while their $F$ -Cross-Barcodes are shown on the right. + +# 5.3 Enhancing 3D reconstruction + +SHAPR (SHApe PRediction) [37] is a deep learning model for reconstructing cellular and nuclear 3D shapes from 2D fluorescence microscopy images. Correct reconstruction of structural patterns in 3D shapes is critical for this problem. However, such information is not captured by established pixel/voxel-wise loss terms, e.g. Dice loss. Later, [36] improved SHAPR by adding contextual information about the overall shape of an object. The work [36] proposed to augment SHAPR with a topological loss term - a Wasserstein-2 distance (W.D.) between + +![](images/bd867d0aaeea8902623d0961ada0e57876244b0a3835a0dc86618c04ade28df3.jpg) +Fig. 6: A random Watts-Strogatz small-world graph [38] with 30 nodes. + +![](images/01c184984f677ff2b714151d26999c779135968763d86a385dbb426dec059b9f.jpg) +Fig.7: Comparison of graph Laplacian eigenvectors via SFTD. + +![](images/cefa2dcf58c0425a033f3f8a1e9a114b99a350cd8cf1619175ddc30d96e55cc2.jpg) +(a) Eigenvector 1. + +![](images/37f366c109bdd92513fb6ab6aa7e6113f270569fe7733cd4e6f9a363c8e350dc.jpg) +(b) Eigenvector 4. + +![](images/598c980bdb85d93577dfc354c882b5739a757599a06ed420f4bcb126566a9091.jpg) +(c) F-Cross-Barcode $(e_1,e_4)$ +Fig.8: Comparison of $e_1$ and $e_4$ . + +![](images/5ca3151bd924b785f75f25a8749074107a2e1ee451d5e3201549dcb5206b6e5c.jpg) +(d) F-Cross-Barcode $(e_4,e_1)$ + +![](images/9c86fd33eb424c0151530547f074d98b108dbf934c6682e6d701f2b3bf77648b.jpg) +(a) Eigenvector 0. + +![](images/1f6069fd265f7d9040b62089adab0d6165009850ab21de3dd03577fab58ec3ef.jpg) +(b) Eigenvector 22. +Fig. 9: Comparison of $e_0$ and $e_{22}$ . + +![](images/d5a2ed65dae205e30160c04caf7d9bec0864528eea015ab1cef4b73b1f312670.jpg) +(c) F-Cross-Barcode $(e_0,e_{22})$ + +![](images/08e9a06ef170114a2797ff618e0430ba32f54997672e04d202a631ec38dec65b.jpg) +(d) F-Cross-Barcode $(e_{22}, e_0)$ . + +Table 2: Errors of 3D shape reconstruction for "Cell nuclei" dataset. + +
METHOD1-IoUVOLUMESURFACEROUGH.H0H1H2
SHAPR [37]0.6120.4820.2730.1910.0000.0070.038
SHAPR+W.D. [36]0.6120.4370.2510.2000.0000.0080.027
SHAPR+Σi=0,1,2SFTDi0.6180.4320.2440.1870.0000.0080.020
SHAPR+Σi=0,1,2,SFTDi0.5970.4370.2470.1910.0000.0110.011
+ +persistence barcodes of ground truth and predicted 3D objects and a total persistence. However, as we discussed before, this loss term doesn't take into account the localization of topological features. We replaced the topological loss term in SHAPR with a symmetrized SFTD with $p = 2$ in two variants8: dimensions up to 2 and dimensions up to 3. Then, we carried out experiments with two datasets from [37]. + +Table 1 shows experimental results for the "Red blood cell" dataset. We conclude that SHAPR augmented with the SFTD loss term is the best one in terms of 3D shape matching errors: IoU error, volume error, surface error, and roughness error (see details in [37]). Additionally, we calculated Wasserstein-2 distances between persistence barcodes $(H_0, H_1, H_2)$ of ground truth and generated 3D objects. Statistically significant improvements are underlined. The proposed SFTD topological loss term with 0-3 dimensions has the lowest Wasserstein-2 distances. This is the rare case of an application of higher homologies like $H_3$ for real-world applications. + +For the "Cell nuclei" dataset results are presented in Table ??. Differences are quite small and are statistically significant only for two metrics: IoU error and W.D. $H_{2}$ , where SHAPR with SFTD loss is the best model. + +Figure 10 shows 3D visualizations of $2\mathrm{D} \rightarrow 3\mathrm{D}$ reconstructions. We see that even with the topological regularization from [36], 3D reconstructions are sometimes erroneous and have serious topological defects: cavities, holes, and surface breaks. In experiments, the sublevel filtration was used and a density function equaled "1" inside the object and "0" in the ambient space. That is, a cavity inside a sphere corresponds to an extra bar in $H_0$ barcode, an extra connected component corresponds to an extra bar in $H_2$ barcode. + +# 5.4 Identifying topological errors in 3D segmentation + +Brain tumor 3D segmentation aims to reveal the sub-regions of tumor corresponding to the enhancing tumor, the edematous/invaded tissue, the necrosis. Availability of versatile similarity measures not only allows precise evaluation of deep learning 3D segmentation models, but also assists monitoring disease progression in clinical practice. As mentioned above, measuring the similarity of the two 3D segmentations by comparing their persistence barcodes does not account for localization of the topological features. In this section, we illustrate the ability of SFTD to complement the existing metrics with this additional information. + +In order to investigate the applicability of the proposed SFTD as a metric for 3D segmentation, we utilize the publicly available pretrained Swin UNETR model [21,33] and BraTS 21 dataset [2-4,28]. Figures 11a, 11b demonstrate the ground truth segmentation and model's prediction. While each of these two 3D segmentations has two voids inside, only two voids have the same localization + +![](images/41761ffd262887a54d63fd0727c7558abaddc1694f2bd5bbf16c3ad88a962044.jpg) +(a) Ground truth + +![](images/697ce54b929563dc668bd033838f372062628e902c0d47afc79e0203bfe85347.jpg) +(b) SHAPR+W.D. + +![](images/cce19c6d5e2159634786337fd7d517e120134d377add3182811a9599fe002561.jpg) +(c) SHAPR+SFTD + +![](images/aff7a95aed9a7a7441a94a39017c440a152ebf12428b130b65c7778f4451b11f.jpg) +(d) Ground truth + +![](images/76fe38e46fe731ba4507148732ac5b1e86631b131e28d846827e201e569bd4db.jpg) + +![](images/1c487d91bb50c6554ad1c098fbfce44f5fdabd08b88dd89d19c00bc553a6c372.jpg) +(f) SHAPR+SFTD + +![](images/e4ce5524d6dc2fa9a370202961b2f3ff6e3d0a4b3e9c6d83bec94e26f508f045.jpg) +(g) Ground truth + +![](images/812345f3109fad08e623ae47e9aceeeec85d482fcd240a0b55a72cf854d63a71.jpg) +(e) SHAPR+W.D. +(h) SHAPR+W.D. + +![](images/5f514d9dd00b360eb434c1558fff8e1a5945839b77374236590c63bfe6db2faa.jpg) +(i) $\mathrm{SHAPR} + \mathrm{SFTD}$ +Fig. 10: Examples of topological errors. Top, $H_0$ : SHAPR+W.D. has a wrong cavity inside. Middle, $H_1$ : SHAPR+W.D. has a wrong hole. Bottom, $H_2$ : SHAPR+W.D. has a wrong surface break. + +![](images/9099adfac9472f400b60913b792f1470085a04b9e4399505e9c1b8820ffcf93b.jpg) +(a) Ground truth + +![](images/0810c284d85d4dfa652b7b3a30b41575108132af301e38f8ac400a741933f96a.jpg) +(b) Prediction + +![](images/8c54e9e47e0f1c8f3dcb4600ae08eb82ad5087dd9f04be4a28e6aa3db568c114.jpg) +(c) Pers. barc. of ground truth + +![](images/d5d25b88525fb4e0cd318060e0e8ff5a0d3e46c45d42f8fd6e1558b724704717.jpg) +(d) Pers. barc. of prediction + +![](images/f656680a8bc0f06d1eea9b3ff3bbf2eff033988e28b081d1d6f6000bbfbc0f51.jpg) +(e) F-Cross-Barcode +Fig. 11: Examples of topological errors. Two 3D segmentations - (a) - ground truth and (b) - prediction - are indistinguishable by their persistence barcodes (c), (d). At the same time, SFTD between ground truth and prediction equals to 1, see (e). A pair of points inside each object in (a), (b) corresponds to the dissimilar topological features, voids, where birth/death events are depicted by red/green dots. + +which makes the two 3D segmentations different. As shown in Figures 11c, 11d, the two 3D segmentations are indistinguishable by their persistence barcodes. In contrary, SFTD is able to recognise the difference between these two cases as revealed by the F-Cross-Barcode in Figure 11e. In particular, SFTD indicates that the difference is precisely in locations of one pair of voids since the set of 1-dimensional topological features contains one segment. + +# 6 Conclusions + +In this paper, we have proposed new topological tools: F-Cross-Barcode and Scalar Function Topology Divergence (SFTD), which measure the dissimilarity of multi-scale topology between sublevel sets of two functions having a common domain. Functions can be defined on an undirected graph or Euclidean space of any dimensionality. We have shown that our tools outperform the standard approach: evaluating Wasserstein distance between persistence barcodes, calculated independently for sublevel sets of scalar functions. F-Cross-Barcode and SFTD take into account the localization of topological features. SFTD, as a loss function, is beneficial in computer vision, since it helps to enforce topological similarity at multiple scales. Also, the proposed tools provide useful visualizations of areas of topological dissimilarities. In the experimental section, we demonstrated that SFTD improved the reconstruction of cellular 3D shapes from 2D fluorescence microscopy images and identified topological errors in 3D segmentation. Additionally, we showed that SFTD outperforms Betti matching loss in 2D segmentation problems. We believe that the proposed tools will have more applications in computer vision and learning from graph data. + +# Acknowledgements + +The work was supported by the Analytical center under the RF Government (subsidy agreement 000000D730321P5Q0002, Grant No. 70-2021-00145 02.11.2021) + +# References + +1. Abousamra, S., Hoai, M., Samaras, D., Chen, C.: Localization in the crowd with topological constraints. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 872-881 (2021) +2. Baid, Ujjwal, e.a.: The rsna-asnr-miccai brats 2021 benchmark on brain tumor segmentation and radiogenomic classification. arXiv preprint arXiv:2107.02314 (2021) +3. Bakas, S., Akbari, H., Sotiras, A., Bilello, M., Rozycki, M., Kirby, J.S., Freymann, J.B., Farahani, K., Davatzikos, C.: Advancing the cancer genome atlas glioma mri collections with expert segmentation labels and radiomic features. Nature Scientific Data (2017) +4. Bakas, S., Akbari, H., Sotiras, A., Bilello, M., Rozycki, M., Kirby, J.S., Freymann, J.B., Farahani, K., Davatzikos, C.: Segmentation labels and radiomic features for the pre-operative scans of the tcga-gbm collection. The Cancer Imaging Archive (2017) +5. Barannikov, S.: The framed Morse complex and its invariants. Advances in Soviet Mathematics 21, 93-116 (1994) +6. Barannikov, S., Korotin, A., Oganesyan, D., Emtsev, D., Burnaev, E.: Barcodes as summary of loss function topology. Doklady Rossijskoj akademii nauk. Matematika, informatika, processy upravlenia 514(2), 196-211 (2023) +7. Barannikov, S., Trofimov, I., Balabin, N., Burnaev, E.: Representation topology divergence: A method for comparing neural network representations. ICML 2022, arXiv:2201.00058 (2021) +8. Barannikov, S., Trofimov, I., Sotnikov, G., Trimbach, E., Korotin, A., Filippov, A., Burnaev, E.: Manifold topology divergence: a framework for comparing data manifolds. Advances in Neural Information Processing Systems 34, 7294-7305 (2021) +9. Barannikov, S., Voronkova, D., Trofimov, I., Korotin, A., Sotnikov, G., Burnaev, E.: Topological obstructions in neural networks learning. arXiv preprint arXiv:2012.15834 (2020) +10. Carpenter, A.E., Jones, T.R., Lamprecht, M.R., Clarke, C., Kang, I.H., Friman, O., Guertin, D.A., Chang, J.H., Lindquist, R.A., Moffat, J., et al.: Cellprofiler: image analysis software for identifying and quantifying cell phenotypes. Genome biology 7, 1-11 (2006) +11. Carrière, M., Chazal, F., Glisse, M., Ike, Y., Kannan, H., Umeda, Y.: Optimizing persistent homology based functions. In: International Conference on Machine Learning. pp. 1294-1303. PMLR (2021) +12. Carrière, M., Chazal, F., Ike, Y., Lacombe, T., Royer, M., Umeda, Y.: Perslay: A neural network layer for persistence diagrams and new graph topological signatures. In: International Conference on Artificial Intelligence and Statistics. pp. 2786-2796. PMLR (2020) +13. Chazal, F., Michel, B.: An introduction to topological data analysis: fundamental and practical aspects for data scientists. arXiv preprint arXiv:1710.04019 (2017) +14. Chen, Y., Segovia, I., Gel, Y.R.: Z-gcnets: Time zigzags at graph convolutional networks for time series forecasting. In: International Conference on Machine Learning. pp. 1684-1694. PMLR (2021) + +15. Chen, Y., Segovia Dominguez, I., Coskunuzer, B., Gel, Y.: Tamp-s2gcnets: When time-aware multipersistence meets spatio-supra graph convolutional nets while forecasting time series. In: The International Conference on Learning Representations (ICLR) (2022) +16. Demir, A., Coskunuzer, B., Gel, Y., Segovia-Dominguez, I., Chen, Y., Kiziltan, B.: Todd: Topological compound fingerprinting in computer-aided drug discovery. Advances in Neural Information Processing Systems 35, 27978-27993 (2022) +17. Funke, J., Tschopp, F., Grisaitis, W., Sheridan, A., Singh, C., Saalfeld, S., Turaga, S.C.: Large scale image segmentation with structured loss based deep learning for connectome reconstruction. IEEE transactions on pattern analysis and machine intelligence 41(7), 1669-1680 (2018) +18. Gelfand, S.I., Manin, Y.I.: Methods of homological algebra. Springer Science & Business Media (2002) +19. Gupta, S., Hu, X., Kaan, J., Jin, M., Mpoy, M., Chung, K., Singh, G., Saltz, M., Kurc, T., Saltz, J., et al.: Learning topological interactions for multi-class medical image segmentation. In: European Conference on Computer Vision. pp. 701-718. Springer (2022) +20. Gupta, S., Zhang, Y., Hu, X., Prasanna, P., Chen, C.: Topology-aware uncertainty for image segmentation. Advances in Neural Information Processing Systems 36 (2024) +21. Hatamizadeh, A., Nath, V., Tang, Y., Yang, D., Roth, H.R., Xu, D.: Swin unetr: Swin transformers for semantic segmentation of brain tumors in mri images. arXiv preprint arXiv:2201.01266v1 (2022) +22. Hofer, C., Graf, F., Rieck, B., Niethammer, M., Kwitt, R.: Graph filtration learning. In: International Conference on Machine Learning. pp. 4314-4323. PMLR (2020) +23. Horn, M., De Brouwer, E., Moor, M., Moreau, Y., Rieck, B., Borgwardt, K.: Topological graph neural networks. arXiv preprint arXiv:2102.07835 (2021) +24. Hu, X., Li, F., Samaras, D., Chen, C.: Topology-preserving deep image segmentation. Advances in neural information processing systems 32 (2019) +25. Hu, X., Wang, Y., Fuxin, L., Samaras, D., Chen, C.: Topology-aware segmentation using discrete Morse theory. arXiv preprint arXiv:2103.09992 (2021) +26. Kaczynski, T., Mischaikow, K.M., Mrozek, M.: Computational homology, vol. 3. Springer (2004) +27. Luo, Y., Shi, L., Thost, V.: Improving self-supervised molecular representation learning using persistent homology. arXiv preprint arXiv:2311.17327 (2023) +28. Menze, B.H., Jakab, A., Bauer, S., Kalpathy-Cramer, J., Farahani, K., Kirby, Justin, e.a.: The multimodal brain tumor image segmentation benchmark (brats). IEEE Transactions on Medical Imaging (2015) +29. Moor, M., Horn, M., Rieck, B., Borgwardt, K.: Topological autoencoders. In: International conference on machine learning. pp. 7045-7054. PMLR (2020) +30. Pérez, J.B., Hauke, S., Lupo, U., Caorsi, M., Dassatti, A.: giotto-ph: A python library for high-performance computation of persistent homology of vietoris-rips filtrations (2021) +31. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015) +32. Stucki, N., Paetzold, J.C., Shit, S., Menze, B., Bauer, U.: Topologically faithful image segmentation via induced matching of persistence barcodes. In: International Conference on Machine Learning. pp. 32698-32727. PMLR (2023) + +33. Tang, Y., Yang, D., Li, W., Roth, H.R., Landman, B., Xu, D., Nath, V., Hatamizadeh, A.: Self-supervised pre-training of swim transformers for 3d medical image analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022) +34. Trofimov, I., Cherniavskii, D., Tulchinskii, E., Balabin, N., Burnaev, E., Barannikov, S.: Learning topology-preserving data representations. arXiv preprint arXiv:2302.00136 (2023) +35. Wagner, H., Chen, C., Vuçini, E.: Efficient computation of persistent homology for cubical data. In: Topological methods in data analysis and visualization II: theory, algorithms, and applications, pp. 91-106. Springer (2011) +36. Waibel, D.J., Atwell, S., Meier, M., Marr, C., Rieck, B.: Capturing shape information with multi-scale topological loss terms for 3d reconstruction. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 150-159. Springer (2022) +37. Waibel, D.J., Kiermeyer, N., Atwell, S., Sadafi, A., Meier, M., Marr, C.: Shapr predicts 3d cell shapes from 2d microscopic images. Iscience 25(11) (2022) +38. Watts, D.J., Strogatz, S.H.: Collective dynamics of 'small-world'networks. nature 393(6684), 440-442 (1998) +39. Wu, J., Wang, Y., Xue, T., Sun, X., Freeman, B., Tenenbaum, J.: Marrnet: 3d shape reconstruction via 2.5 d sketches. Advances in neural information processing systems 30 (2017) +40. Yan, Z., Ma, T., Gao, L., Tang, Z., Chen, C.: Link prediction with persistent homology: An interactive view. In: International conference on machine learning. pp. 11659-11669. PMLR (2021) \ No newline at end of file diff --git a/2024/Scalar Function Topology Divergence_ Comparing Topology of 3D Objects/images.zip b/2024/Scalar Function Topology Divergence_ Comparing Topology of 3D Objects/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ae944152a62c52a12a0151998d82e04243c518e4 --- /dev/null +++ b/2024/Scalar Function Topology Divergence_ Comparing Topology of 3D Objects/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:05f44e0449b7414f0e873e7bb112dbafe1f6a36471331b5296a0febd4952cb01 +size 506841 diff --git a/2024/Scalar Function Topology Divergence_ Comparing Topology of 3D Objects/layout.json b/2024/Scalar Function Topology Divergence_ Comparing Topology of 3D Objects/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9daea8843d9905eae061edcc691c77d911024a1d --- /dev/null +++ b/2024/Scalar Function Topology Divergence_ Comparing Topology of 3D Objects/layout.json @@ -0,0 +1,15666 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 134, + 111, + 479, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 111, + 479, + 148 + ], + "spans": [ + { + "bbox": [ + 134, + 111, + 479, + 148 + ], + "type": "text", + "content": "Scalar Function Topology Divergence: Comparing Topology of 3D Objects" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 156, + 167, + 457, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 167, + 457, + 193 + ], + "spans": [ + { + "bbox": [ + 156, + 167, + 457, + 193 + ], + "type": "text", + "content": "Ilya Trofimov1, Daria Voronkova1,2, Eduard Tulchinskii1,4, Evgeny Burnaev1,2, and Serguei Barannikov1,3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 164, + 201, + 447, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 201, + 447, + 234 + ], + "spans": [ + { + "bbox": [ + 164, + 201, + 447, + 234 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 164, + 201, + 447, + 234 + ], + "type": "text", + "content": " Skolkovo Institute of Science and Technology, Moscow, Russia {ilya.trofimov,darya.voronkova, eduard.tulchinskiy,e.burnaev, s.barannikov}@skoltech.ru" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 257, + 235, + 357, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 235, + 357, + 245 + ], + "spans": [ + { + "bbox": [ + 257, + 235, + 357, + 245 + ], + "type": "text", + "content": "2 AIRI, Moscow, Russia" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 217, + 245, + 397, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 245, + 397, + 257 + ], + "spans": [ + { + "bbox": [ + 217, + 245, + 397, + 257 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 217, + 245, + 397, + 257 + ], + "type": "text", + "content": " CNRS, IMJ, Paris Cité University, France" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 197, + 257, + 416, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 257, + 416, + 267 + ], + "spans": [ + { + "bbox": [ + 197, + 257, + 416, + 267 + ], + "type": "text", + "content": "4 AI Foundation and Algorithm Lab, Moscow, Russia" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 160, + 296, + 452, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 296, + 452, + 493 + ], + "spans": [ + { + "bbox": [ + 160, + 296, + 452, + 493 + ], + "type": "text", + "content": "Abstract. We propose a new topological tool for computer vision - Scalar Function Topology Divergence (SFTD), which measures the dissimilarity of multi-scale topology between sublevel sets of two functions having a common domain. Functions can be defined on an undirected graph or Euclidean space of any dimensionality. Most of the existing methods for comparing topology are based on Wasserstein distance between persistence barcodes and they don't take into account the localization of topological features. The minimization of SFTD ensures that the corresponding topological features of scalar functions are located in the same places. The proposed tool provides useful visualizations depicting areas where functions have topological dissimilarities. We provide applications of the proposed method to 3D computer vision. In particular, experiments demonstrate that SFTD as an additional loss improves the reconstruction of cellular 3D shapes from 2D fluorescence microscopy images, and helps to identify topological errors in 3D segmentation. Additionally, we show that SFTD outperforms Betti matching loss in 2D segmentation problems. The code is publicly available: https://github.com/IlyaTrofimov/SFTD." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 159, + 503, + 453, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 503, + 453, + 526 + ], + "spans": [ + { + "bbox": [ + 159, + 503, + 453, + 526 + ], + "type": "text", + "content": "Keywords: Topological Data Analysis " + }, + { + "bbox": [ + 159, + 503, + 453, + 526 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 503, + 453, + 526 + ], + "type": "text", + "content": " 3D Computer Vision " + }, + { + "bbox": [ + 159, + 503, + 453, + 526 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 503, + 453, + 526 + ], + "type": "text", + "content": " Geometric Deep Learning" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 545, + 230, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 230, + 558 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 230, + 558 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": "A common situation in computer vision is a structural prediction defined over " + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^n" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": ". Examples include image segmentation [24, 25, 31, 32] and 3D shape reconstruction [36, 37, 39] problems. However, in most of the cases the comparison between the prediction and a ground truth is done pixel(voxel)-wise by metrics like Dice score, cross-entropy, etc. These metrics are not capable of providing detailed topological correctness, i.e. preserving multi-scale shape patterns: cavities, holes, connectivity patterns, etc. Typically, predictions involve scalar functions " + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "f:\\mathbb{R}^n\\to \\mathbb{R}" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": " and the final shape of an object is defined by thresholding " + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "f(x)" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": "There is a need for topological-aware comparison metrics for scalar functions, which take into account the topology of sublevel sets " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "f^{-1}((-\\infty, \\varepsilon])" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": " (the analysis for superlevel sets is similar). Existing tools for this problem don't take into account the localization of topological features [24,36] or are limited to 2D space [32]. Our contributions are as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 135, + 185, + 479, + 362 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 136, + 185, + 479, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 185, + 479, + 220 + ], + "spans": [ + { + "bbox": [ + 136, + 185, + 479, + 220 + ], + "type": "text", + "content": "1. We introduce new topological tools: F-Cross-Barcodes and Scalar Function Topology Divergence (SFTD), for a multi-scale topology-aware comparison of scalar functions by tracking differences in their sublevel sets topology." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 135, + 221, + 479, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 221, + 479, + 243 + ], + "spans": [ + { + "bbox": [ + 135, + 221, + 479, + 243 + ], + "type": "text", + "content": "2. A distinctive feature of F-Cross-Barcodes and SFTD is that they take into account the localization of topological features;" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 135, + 245, + 479, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 245, + 479, + 277 + ], + "spans": [ + { + "bbox": [ + 135, + 245, + 479, + 277 + ], + "type": "text", + "content": "3. We describe a visualization technique for F-Cross-Barcodes: a technique highlights with points areas where differences in topological features are localized:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 136, + 280, + 374, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 280, + 374, + 290 + ], + "spans": [ + { + "bbox": [ + 136, + 280, + 374, + 290 + ], + "type": "text", + "content": "4. We propose an algorithm for SFTD differentiation;" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 136, + 292, + 479, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 292, + 479, + 326 + ], + "spans": [ + { + "bbox": [ + 136, + 292, + 479, + 326 + ], + "type": "text", + "content": "5. By doing computational experiments, we show that SFTD as an additional loss term improves 3D shape reconstruction, and helps to identify topological errors in 3D segmentation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 135, + 328, + 479, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 328, + 479, + 362 + ], + "spans": [ + { + "bbox": [ + 135, + 328, + 479, + 362 + ], + "type": "text", + "content": "6. We also show that SFTD outperforms Betti matching loss in terms of topological accuracy in 2D segmentation while being " + }, + { + "bbox": [ + 135, + 328, + 479, + 362 + ], + "type": "inline_equation", + "content": "\\approx 35" + }, + { + "bbox": [ + 135, + 328, + 479, + 362 + ], + "type": "text", + "content": " times faster, see Appendix G." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 382, + 234, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 382, + 234, + 395 + ], + "spans": [ + { + "bbox": [ + 132, + 382, + 234, + 395 + ], + "type": "text", + "content": "2 Related work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 407, + 479, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 479, + 478 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 479, + 478 + ], + "type": "text", + "content": "Tools of topological data analysis find application in various machine learning problems: evaluation of deep generative models [8], comparison of neural representations [7], dimensionality reduction [29,34], graph neural networks [12,23,40], self-supervised learning [27], molecular data analysis [16], analysis of loss functions in deep learning [6,9], spatial-temporal graph neural network tasks [14,15], crowd localization [1], to name a few." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 479, + 479, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 479, + 562 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 479, + 562 + ], + "type": "text", + "content": "Several works used topological tools to enhance image segmentation [20, 24, 25, 32]. However, proposed algorithms for comparison ground truth and predicted images either don't take into account localization of topological features [24], or are limited to 2D [32]. A method by [25] highlights in images topologically critical and error-prone structures using discrete Morse theory, and [20] proposes a method for inferring structure-aware uncertainty. The importance of topological interactions for multi-class medical image segmentation is shown in [19]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 580, + 233, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 580, + 233, + 594 + ], + "spans": [ + { + "bbox": [ + 132, + 580, + 233, + 594 + ], + "type": "text", + "content": "3 Preliminaries" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 605, + 351, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 605, + 351, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 605, + 351, + 666 + ], + "type": "text", + "content": "We study scalar functions " + }, + { + "bbox": [ + 131, + 605, + 351, + 666 + ], + "type": "inline_equation", + "content": "f: \\mathbb{R}^n \\to \\mathbb{R}" + }, + { + "bbox": [ + 131, + 605, + 351, + 666 + ], + "type": "text", + "content": " by assessing the topology of its sublevel sets. Consider a 1D function depicted in Figure 1. It has two local minima " + }, + { + "bbox": [ + 131, + 605, + 351, + 666 + ], + "type": "inline_equation", + "content": "y_1, y_2" + }, + { + "bbox": [ + 131, + 605, + 351, + 666 + ], + "type": "text", + "content": " and a local maximum " + }, + { + "bbox": [ + 131, + 605, + 351, + 666 + ], + "type": "inline_equation", + "content": "y_3" + }, + { + "bbox": [ + 131, + 605, + 351, + 666 + ], + "type": "text", + "content": ", such that " + }, + { + "bbox": [ + 131, + 605, + 351, + 666 + ], + "type": "inline_equation", + "content": "y_1 < y_2 < y_3" + }, + { + "bbox": [ + 131, + 605, + 351, + 666 + ], + "type": "text", + "content": ". One can study sublevel sets" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 359, + 581, + 479, + 657 + ], + "blocks": [ + { + "bbox": [ + 359, + 581, + 479, + 657 + ], + "lines": [ + { + "bbox": [ + 359, + 581, + 479, + 657 + ], + "spans": [ + { + "bbox": [ + 359, + 581, + 479, + 657 + ], + "type": "image", + "image_path": "3dd7950c826053b97be8ccddecdbe1c53dfc7631ee8e4200b6be564b361423e3.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 356, + 669, + 482, + 690 + ], + "lines": [ + { + "bbox": [ + 356, + 669, + 482, + 690 + ], + "spans": [ + { + "bbox": [ + 356, + 669, + 482, + 690 + ], + "type": "text", + "content": "Fig.1: 1D function and its sublevel sets." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 237, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 237, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 237, + 101 + ], + "type": "text", + "content": "I. Trofimov et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 152 + ], + "type": "inline_equation", + "content": "X_{\\varepsilon} = f^{-1}((-\\infty ,\\varepsilon ])" + }, + { + "bbox": [ + 130, + 115, + 479, + 152 + ], + "type": "text", + "content": ". For the scalar function " + }, + { + "bbox": [ + 130, + 115, + 479, + 152 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 115, + 479, + 152 + ], + "type": "text", + "content": ", changes of " + }, + { + "bbox": [ + 130, + 115, + 479, + 152 + ], + "type": "inline_equation", + "content": "X_{\\varepsilon}" + }, + { + "bbox": [ + 130, + 115, + 479, + 152 + ], + "type": "text", + "content": " for increasing " + }, + { + "bbox": [ + 130, + 115, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 130, + 115, + 479, + 152 + ], + "type": "text", + "content": " can be intuitively understood as a \"filling the pool\", see Figure 1. There are four distinct situations:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 135, + 162, + 337, + 211 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 135, + 162, + 240, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 162, + 240, + 174 + ], + "spans": [ + { + "bbox": [ + 135, + 162, + 240, + 174 + ], + "type": "text", + "content": "1. For " + }, + { + "bbox": [ + 135, + 162, + 240, + 174 + ], + "type": "inline_equation", + "content": "\\varepsilon < y_{1}" + }, + { + "bbox": [ + 135, + 162, + 240, + 174 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 135, + 162, + 240, + 174 + ], + "type": "inline_equation", + "content": "X_{\\varepsilon} = \\varnothing" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 135, + 175, + 326, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 175, + 326, + 186 + ], + "spans": [ + { + "bbox": [ + 135, + 175, + 326, + 186 + ], + "type": "text", + "content": "2. For " + }, + { + "bbox": [ + 135, + 175, + 326, + 186 + ], + "type": "inline_equation", + "content": "y_{1} \\leq \\varepsilon < y_{2}" + }, + { + "bbox": [ + 135, + 175, + 326, + 186 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 135, + 175, + 326, + 186 + ], + "type": "inline_equation", + "content": "X_{\\varepsilon}" + }, + { + "bbox": [ + 135, + 175, + 326, + 186 + ], + "type": "text", + "content": " is a single segment;" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 135, + 187, + 337, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 187, + 337, + 198 + ], + "spans": [ + { + "bbox": [ + 135, + 187, + 337, + 198 + ], + "type": "text", + "content": "3. For " + }, + { + "bbox": [ + 135, + 187, + 337, + 198 + ], + "type": "inline_equation", + "content": "y_{2} \\leq \\varepsilon < y_{3}" + }, + { + "bbox": [ + 135, + 187, + 337, + 198 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 135, + 187, + 337, + 198 + ], + "type": "inline_equation", + "content": "S_{\\varepsilon}" + }, + { + "bbox": [ + 135, + 187, + 337, + 198 + ], + "type": "text", + "content": " consists of 2 segments;" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 135, + 200, + 335, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 200, + 335, + 211 + ], + "spans": [ + { + "bbox": [ + 135, + 200, + 335, + 211 + ], + "type": "text", + "content": "4. Finally, for " + }, + { + "bbox": [ + 135, + 200, + 335, + 211 + ], + "type": "inline_equation", + "content": "\\varepsilon \\geq y_{3}" + }, + { + "bbox": [ + 135, + 200, + 335, + 211 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 135, + 200, + 335, + 211 + ], + "type": "inline_equation", + "content": "X_{\\varepsilon}" + }, + { + "bbox": [ + 135, + 200, + 335, + 211 + ], + "type": "text", + "content": " is a single segment." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 220, + 480, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 220, + 480, + 269 + ], + "spans": [ + { + "bbox": [ + 130, + 220, + 480, + 269 + ], + "type": "text", + "content": "Note, that changes in topology of " + }, + { + "bbox": [ + 130, + 220, + 480, + 269 + ], + "type": "inline_equation", + "content": "X_{\\varepsilon}" + }, + { + "bbox": [ + 130, + 220, + 480, + 269 + ], + "type": "text", + "content": " happen when " + }, + { + "bbox": [ + 130, + 220, + 480, + 269 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 130, + 220, + 480, + 269 + ], + "type": "text", + "content": " is a critical point of " + }, + { + "bbox": [ + 130, + 220, + 480, + 269 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 220, + 480, + 269 + ], + "type": "text", + "content": ". In the general case, smooth scalar functions defined on smooth manifolds are studied by Morse theory [6]. In practice, we assess functions through their values on a finite set of points, e.g. a lattice (square grid graph) in " + }, + { + "bbox": [ + 130, + 220, + 480, + 269 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^n" + }, + { + "bbox": [ + 130, + 220, + 480, + 269 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 289, + 338, + 302 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 289, + 338, + 302 + ], + "spans": [ + { + "bbox": [ + 131, + 289, + 338, + 302 + ], + "type": "text", + "content": "3.1 Scalar functions defined on a graph" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 312, + 480, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 480, + 359 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 480, + 359 + ], + "type": "text", + "content": "In the most general form, we are interested in topology-aware comparison of scalar functions " + }, + { + "bbox": [ + 130, + 312, + 480, + 359 + ], + "type": "inline_equation", + "content": "f: \\mathcal{V} \\to \\mathbb{R}" + }, + { + "bbox": [ + 130, + 312, + 480, + 359 + ], + "type": "text", + "content": " defined on vertices " + }, + { + "bbox": [ + 130, + 312, + 480, + 359 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 130, + 312, + 480, + 359 + ], + "type": "text", + "content": " of an undirected graph " + }, + { + "bbox": [ + 130, + 312, + 480, + 359 + ], + "type": "inline_equation", + "content": "\\mathcal{G} = (\\mathcal{V}, \\mathcal{E})" + }, + { + "bbox": [ + 130, + 312, + 480, + 359 + ], + "type": "text", + "content": ". Specifically, we study topological features of subgraphs " + }, + { + "bbox": [ + 130, + 312, + 480, + 359 + ], + "type": "inline_equation", + "content": "\\{v \\in \\mathcal{V} \\mid f(v) \\leq \\varepsilon\\}" + }, + { + "bbox": [ + 130, + 312, + 480, + 359 + ], + "type": "text", + "content": " connected components, cycles, etc." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 360, + 480, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 360, + 480, + 443 + ], + "spans": [ + { + "bbox": [ + 130, + 360, + 480, + 443 + ], + "type": "text", + "content": "For each clique " + }, + { + "bbox": [ + 130, + 360, + 480, + 443 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 360, + 480, + 443 + ], + "type": "text", + "content": " in the graph " + }, + { + "bbox": [ + 130, + 360, + 480, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 130, + 360, + 480, + 443 + ], + "type": "text", + "content": ", we define a filtration function " + }, + { + "bbox": [ + 130, + 360, + 480, + 443 + ], + "type": "inline_equation", + "content": "T(C) = \\max_{i\\in C}f(i)" + }, + { + "bbox": [ + 130, + 360, + 480, + 443 + ], + "type": "text", + "content": ". By a clique complex " + }, + { + "bbox": [ + 130, + 360, + 480, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 360, + 480, + 443 + ], + "type": "text", + "content": ", we mean a set of cliques where each sub-clique also belongs to a complex5. For some " + }, + { + "bbox": [ + 130, + 360, + 480, + 443 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 130, + 360, + 480, + 443 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 360, + 480, + 443 + ], + "type": "inline_equation", + "content": "T(C)" + }, + { + "bbox": [ + 130, + 360, + 480, + 443 + ], + "type": "text", + "content": " defines a clique complex " + }, + { + "bbox": [ + 130, + 360, + 480, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_j = \\{C\\mid T(C)\\leq \\varepsilon \\}" + }, + { + "bbox": [ + 130, + 360, + 480, + 443 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 360, + 480, + 443 + ], + "type": "inline_equation", + "content": "f(j) = \\varepsilon" + }, + { + "bbox": [ + 130, + 360, + 480, + 443 + ], + "type": "text", + "content": ". Topological features for a subgraph " + }, + { + "bbox": [ + 130, + 360, + 480, + 443 + ], + "type": "inline_equation", + "content": "\\{v\\in \\mathcal{V}\\mid f(v)\\leq \\varepsilon \\}" + }, + { + "bbox": [ + 130, + 360, + 480, + 443 + ], + "type": "text", + "content": " are precisely expressed by homology groups. Overall, the filtration function " + }, + { + "bbox": [ + 130, + 360, + 480, + 443 + ], + "type": "inline_equation", + "content": "T(C)" + }, + { + "bbox": [ + 130, + 360, + 480, + 443 + ], + "type": "text", + "content": " defines the lower-star filtration of the graph's clique complex:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 254, + 445, + 360, + 457 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 254, + 445, + 360, + 457 + ], + "spans": [ + { + "bbox": [ + 254, + 445, + 360, + 457 + ], + "type": "interline_equation", + "content": "\\varnothing \\subseteq \\mathcal {C} _ {0} \\subseteq \\mathcal {C} _ {1} \\subseteq \\dots \\subseteq \\mathcal {C} _ {n},", + "image_path": "85fe8284fd748805a277fad9a8e5723e5802a7641d94a6266b838bbf7f86ff77.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 464, + 479, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 464, + 479, + 488 + ], + "spans": [ + { + "bbox": [ + 130, + 464, + 479, + 488 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 464, + 479, + 488 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_n" + }, + { + "bbox": [ + 130, + 464, + 479, + 488 + ], + "type": "text", + "content": " is a clique complex of " + }, + { + "bbox": [ + 130, + 464, + 479, + 488 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 130, + 464, + 479, + 488 + ], + "type": "text", + "content": ". The filtration is said to be induced by the vertex function " + }, + { + "bbox": [ + 130, + 464, + 479, + 488 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 464, + 479, + 488 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 489, + 480, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 489, + 480, + 550 + ], + "spans": [ + { + "bbox": [ + 130, + 489, + 480, + 550 + ], + "type": "text", + "content": "A given topological feature \"appears\" at some " + }, + { + "bbox": [ + 130, + 489, + 480, + 550 + ], + "type": "inline_equation", + "content": "\\varepsilon_{1}" + }, + { + "bbox": [ + 130, + 489, + 480, + 550 + ], + "type": "text", + "content": " and \"dies\" at some " + }, + { + "bbox": [ + 130, + 489, + 480, + 550 + ], + "type": "inline_equation", + "content": "\\varepsilon_{2}" + }, + { + "bbox": [ + 130, + 489, + 480, + 550 + ], + "type": "text", + "content": ". These topological features, together with their birth and death times can be calculated via tools of topological data analysis. The multi-set of the corresponding intervals " + }, + { + "bbox": [ + 130, + 489, + 480, + 550 + ], + "type": "inline_equation", + "content": "(\\varepsilon_{1},\\varepsilon_{2})" + }, + { + "bbox": [ + 130, + 489, + 480, + 550 + ], + "type": "text", + "content": " is called the persistence barcode, which depicts topological features of the graph for all scales. The whole theory is dubbed persistent homology [5, 13]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 570, + 202, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 570, + 202, + 582 + ], + "spans": [ + { + "bbox": [ + 131, + 570, + 202, + 582 + ], + "type": "text", + "content": "4 Method" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 131, + 597, + 339, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 597, + 339, + 611 + ], + "spans": [ + { + "bbox": [ + 131, + 597, + 339, + 611 + ], + "type": "text", + "content": "4.1 Scalar functions defined on a graph" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 620, + 480, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 620, + 480, + 644 + ], + "spans": [ + { + "bbox": [ + 130, + 620, + 480, + 644 + ], + "type": "text", + "content": "To compare topological features of sublevel sets of functions " + }, + { + "bbox": [ + 130, + 620, + 480, + 644 + ], + "type": "inline_equation", + "content": "f,g:\\mathcal{G}\\to \\mathbb{R}" + }, + { + "bbox": [ + 130, + 620, + 480, + 644 + ], + "type": "text", + "content": " we construct a graph " + }, + { + "bbox": [ + 130, + 620, + 480, + 644 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{G}} = (\\tilde{\\mathcal{V}},\\tilde{\\mathcal{E}})" + }, + { + "bbox": [ + 130, + 620, + 480, + 644 + ], + "type": "text", + "content": " with doubled set of vertices from " + }, + { + "bbox": [ + 130, + 620, + 480, + 644 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 130, + 620, + 480, + 644 + ], + "type": "text", + "content": " and one extra" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 294, + 91, + 447, + 102 + ], + "type": "text", + "content": "Scalar Function Topology Divergence" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 479, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 479, + 665 + ], + "type": "text", + "content": "Every clique complex is an abstract simplicial complex, but the reserve is not true." + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 180, + 116, + 258, + 219 + ], + "blocks": [ + { + "bbox": [ + 180, + 116, + 258, + 219 + ], + "lines": [ + { + "bbox": [ + 180, + 116, + 258, + 219 + ], + "spans": [ + { + "bbox": [ + 180, + 116, + 258, + 219 + ], + "type": "image", + "image_path": "ebe692964d3d7b3377c6e58b57b047b2eb36ef19282a72aac14b2a4969134bbc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 173, + 223, + 264, + 234 + ], + "lines": [ + { + "bbox": [ + 173, + 223, + 264, + 234 + ], + "spans": [ + { + "bbox": [ + 173, + 223, + 264, + 234 + ], + "type": "text", + "content": "(a) The doubled graph " + }, + { + "bbox": [ + 173, + 223, + 264, + 234 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{G}}" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 347, + 141, + 437, + 218 + ], + "blocks": [ + { + "bbox": [ + 347, + 141, + 437, + 218 + ], + "lines": [ + { + "bbox": [ + 347, + 141, + 437, + 218 + ], + "spans": [ + { + "bbox": [ + 347, + 141, + 437, + 218 + ], + "type": "image", + "image_path": "747418fb38f3eac499b0fe55b32d5275d37f303be1480fb7a412b53b5d734265.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 223, + 477, + 234 + ], + "lines": [ + { + "bbox": [ + 313, + 223, + 477, + 234 + ], + "spans": [ + { + "bbox": [ + 313, + 223, + 477, + 234 + ], + "type": "text", + "content": "(b) The extended " + }, + { + "bbox": [ + 313, + 223, + 477, + 234 + ], + "type": "inline_equation", + "content": "n + 1" + }, + { + "bbox": [ + 313, + 223, + 477, + 234 + ], + "type": "text", + "content": " dimensional lattice " + }, + { + "bbox": [ + 313, + 223, + 477, + 234 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{L}}" + }, + { + "bbox": [ + 313, + 223, + 477, + 234 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 241, + 242, + 371, + 255 + ], + "lines": [ + { + "bbox": [ + 241, + 242, + 371, + 255 + ], + "spans": [ + { + "bbox": [ + 241, + 242, + 371, + 255 + ], + "type": "text", + "content": "Fig. 2: A visualization of " + }, + { + "bbox": [ + 241, + 242, + 371, + 255 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{G}}" + }, + { + "bbox": [ + 241, + 242, + 371, + 255 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 241, + 242, + 371, + 255 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{L}}" + }, + { + "bbox": [ + 241, + 242, + 371, + 255 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 277, + 480, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 480, + 301 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 480, + 301 + ], + "type": "text", + "content": "vertex " + }, + { + "bbox": [ + 130, + 277, + 480, + 301 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 130, + 277, + 480, + 301 + ], + "type": "text", + "content": ", see Figure 2a. Formally, " + }, + { + "bbox": [ + 130, + 277, + 480, + 301 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{V}} = \\{A_1,\\dots ,A_n,A_1',\\dots ,A_n',O\\}" + }, + { + "bbox": [ + 130, + 277, + 480, + 301 + ], + "type": "text", + "content": ". Edges " + }, + { + "bbox": [ + 130, + 277, + 480, + 301 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{E}}" + }, + { + "bbox": [ + 130, + 277, + 480, + 301 + ], + "type": "text", + "content": " are defined as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 214, + 308, + 398, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 308, + 398, + 357 + ], + "spans": [ + { + "bbox": [ + 214, + 308, + 398, + 357 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\left(A _ {i}, A _ {j}\\right) \\in \\mathcal {E} \\Rightarrow \\left(A _ {i}, A _ {j}\\right) \\in \\tilde {\\mathcal {E}}, \\left(A _ {i} ^ {\\prime}, A _ {j} ^ {\\prime}\\right) \\in \\tilde {\\mathcal {E}} \\\\ \\left(A _ {i}, A _ {j}\\right) \\in \\mathcal {E}, i < j \\Rightarrow \\left(A _ {i}, A _ {j} ^ {\\prime}\\right) \\in \\tilde {\\mathcal {E}} \\\\ \\left(O, A _ {i}\\right) \\in \\tilde {\\mathcal {E}}, \\left(A _ {i}, A _ {i} ^ {\\prime}\\right) \\in \\tilde {\\mathcal {E}} \\text {f o r a l l} i \\\\ \\end{array}", + "image_path": "41f8a37b06df3b2967a283a35ea89b3230ca87aa32b5519eb8a4e299f8225e28.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 366, + 323, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 366, + 323, + 379 + ], + "spans": [ + { + "bbox": [ + 131, + 366, + 323, + 379 + ], + "type": "text", + "content": "The vertex function " + }, + { + "bbox": [ + 131, + 366, + 323, + 379 + ], + "type": "inline_equation", + "content": "\\tilde{f}" + }, + { + "bbox": [ + 131, + 366, + 323, + 379 + ], + "type": "text", + "content": " is defined as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 220, + 386, + 394, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 386, + 394, + 418 + ], + "spans": [ + { + "bbox": [ + 220, + 386, + 394, + 418 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\tilde {f} (A _ {i}) = f (i) \\\\ \\tilde {f} (A _ {i} ^ {\\prime}) = \\min (f (i), g (i)), \\tilde {f} (O) = \\min _ {f, g}, \\\\ \\end{array}", + "image_path": "43b8f37d0170f65e4f2a0f79043236c2c33720387c7e424eb8cf525e5b8c043c.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 426, + 467, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 467, + 438 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 467, + 438 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 426, + 467, + 438 + ], + "type": "inline_equation", + "content": "\\min_{f,g}" + }, + { + "bbox": [ + 130, + 426, + 467, + 438 + ], + "type": "text", + "content": " is the minimum value of both " + }, + { + "bbox": [ + 130, + 426, + 467, + 438 + ], + "type": "inline_equation", + "content": "f,g" + }, + { + "bbox": [ + 130, + 426, + 467, + 438 + ], + "type": "text", + "content": " in the nodes of the graph " + }, + { + "bbox": [ + 130, + 426, + 467, + 438 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 130, + 426, + 467, + 438 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 438, + 480, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 480, + 473 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 480, + 473 + ], + "type": "text", + "content": "Definition. The " + }, + { + "bbox": [ + 130, + 438, + 480, + 473 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 438, + 480, + 473 + ], + "type": "text", + "content": "-Cross-Barcode" + }, + { + "bbox": [ + 130, + 438, + 480, + 473 + ], + "type": "inline_equation", + "content": "_k(f, g)" + }, + { + "bbox": [ + 130, + 438, + 480, + 473 + ], + "type": "text", + "content": " for two scalar functions " + }, + { + "bbox": [ + 130, + 438, + 480, + 473 + ], + "type": "inline_equation", + "content": "f, g" + }, + { + "bbox": [ + 130, + 438, + 480, + 473 + ], + "type": "text", + "content": " defined on vertices of an undirected graph " + }, + { + "bbox": [ + 130, + 438, + 480, + 473 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 130, + 438, + 480, + 473 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 130, + 438, + 480, + 473 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 438, + 480, + 473 + ], + "type": "text", + "content": "-th persistence barcode for the doubled graph " + }, + { + "bbox": [ + 130, + 438, + 480, + 473 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{G}}" + }, + { + "bbox": [ + 130, + 438, + 480, + 473 + ], + "type": "text", + "content": " and the lower-star filtration induced by " + }, + { + "bbox": [ + 130, + 438, + 480, + 473 + ], + "type": "inline_equation", + "content": "\\tilde{f}" + }, + { + "bbox": [ + 130, + 438, + 480, + 473 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 474, + 481, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 481, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 481, + 545 + ], + "type": "text", + "content": "To put it simply, each interval in the " + }, + { + "bbox": [ + 130, + 474, + 481, + 545 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 474, + 481, + 545 + ], + "type": "text", + "content": "-Cross-Barcode" + }, + { + "bbox": [ + 130, + 474, + 481, + 545 + ], + "type": "inline_equation", + "content": "_k(f, g)" + }, + { + "bbox": [ + 130, + 474, + 481, + 545 + ], + "type": "text", + "content": " denotes a range of functions' sub-level sets where a given topological feature exists in the function " + }, + { + "bbox": [ + 130, + 474, + 481, + 545 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 474, + 481, + 545 + ], + "type": "text", + "content": " but doesn't exist in the function " + }, + { + "bbox": [ + 130, + 474, + 481, + 545 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 130, + 474, + 481, + 545 + ], + "type": "text", + "content": ", see further Appendix A. Technically, the desired persistence barcode is calculated via giotto-ph library [30]. For this software, values of a filtration function must be defined on vertices and edges. Algorithm 1 summarizes the aforementioned." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 562, + 342, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 562, + 342, + 574 + ], + "spans": [ + { + "bbox": [ + 131, + 562, + 342, + 574 + ], + "type": "text", + "content": "4.2 Scalar functions defined on a lattice" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "A frequent situation is a function having a domain in " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^n" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": ". We assume that the function is approximated by its restriction on a sufficiently small grid. Algorithm 1 can also be applied in this case after a space triangulation. However, the number of simplices grows very quickly in high dimensions (see discussion in [35]), and Algorithm 1 has a high computational cost. That is why, we consider cubical complexes [26] instead of simplicial complexes. Specifically, we consider a regular square grid graph (" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "-dimensional lattice) and use values of functions" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 91, + 237, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 237, + 100 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 237, + 100 + ], + "type": "text", + "content": "I. Trofimov et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 440, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 440, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 440, + 128 + ], + "type": "text", + "content": "Algorithm 1 F-Cross-Barcode" + }, + { + "bbox": [ + 132, + 116, + 440, + 128 + ], + "type": "inline_equation", + "content": "_k" + }, + { + "bbox": [ + 132, + 116, + 440, + 128 + ], + "type": "text", + "content": " (scalar function " + }, + { + "bbox": [ + 132, + 116, + 440, + 128 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 132, + 116, + 440, + 128 + ], + "type": "text", + "content": "-th cross-barcode)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 132, + 452, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 132, + 452, + 143 + ], + "spans": [ + { + "bbox": [ + 132, + 132, + 452, + 143 + ], + "type": "text", + "content": "Input: Scalar functions " + }, + { + "bbox": [ + 132, + 132, + 452, + 143 + ], + "type": "inline_equation", + "content": "f, g" + }, + { + "bbox": [ + 132, + 132, + 452, + 143 + ], + "type": "text", + "content": " defined on vertices of a graph " + }, + { + "bbox": [ + 132, + 132, + 452, + 143 + ], + "type": "inline_equation", + "content": "\\mathcal{G} = (\\mathcal{V}, \\mathcal{E})" + }, + { + "bbox": [ + 132, + 132, + 452, + 143 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 132, + 132, + 452, + 143 + ], + "type": "inline_equation", + "content": "|\\mathcal{V}| = n" + }, + { + "bbox": [ + 132, + 132, + 452, + 143 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 144, + 481, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 144, + 481, + 176 + ], + "spans": [ + { + "bbox": [ + 132, + 144, + 481, + 176 + ], + "type": "text", + "content": "Require: " + }, + { + "bbox": [ + 132, + 144, + 481, + 176 + ], + "type": "inline_equation", + "content": "\\mathsf{PB}_k(\\cdot)" + }, + { + "bbox": [ + 132, + 144, + 481, + 176 + ], + "type": "text", + "content": " - an algorithm to calculate " + }, + { + "bbox": [ + 132, + 144, + 481, + 176 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 132, + 144, + 481, + 176 + ], + "type": "text", + "content": "-th persistence barcode from a matrix containing values of a filtration function on graph: on vertices - on a diagonal, on edges - off a diagonal." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 141, + 180, + 477, + 277 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 141, + 180, + 455, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 180, + 455, + 201 + ], + "spans": [ + { + "bbox": [ + 141, + 180, + 455, + 201 + ], + "type": "text", + "content": "1. Create a " + }, + { + "bbox": [ + 141, + 180, + 455, + 201 + ], + "type": "inline_equation", + "content": "n \\times n" + }, + { + "bbox": [ + 141, + 180, + 455, + 201 + ], + "type": "text", + "content": " matrix " + }, + { + "bbox": [ + 141, + 180, + 455, + 201 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 141, + 180, + 455, + 201 + ], + "type": "text", + "content": " for the function " + }, + { + "bbox": [ + 141, + 180, + 455, + 201 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 141, + 180, + 455, + 201 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 141, + 180, + 455, + 201 + ], + "type": "inline_equation", + "content": "F_{ij} = f(i)" + }, + { + "bbox": [ + 141, + 180, + 455, + 201 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 141, + 180, + 455, + 201 + ], + "type": "inline_equation", + "content": "i = j" + }, + { + "bbox": [ + 141, + 180, + 455, + 201 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 141, + 180, + 455, + 201 + ], + "type": "inline_equation", + "content": "\\max(f(i), f(j))" + }, + { + "bbox": [ + 141, + 180, + 455, + 201 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 141, + 180, + 455, + 201 + ], + "type": "inline_equation", + "content": "i \\neq j" + }, + { + "bbox": [ + 141, + 180, + 455, + 201 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 141, + 180, + 455, + 201 + ], + "type": "inline_equation", + "content": "(i, j) \\in \\mathcal{E}" + }, + { + "bbox": [ + 141, + 180, + 455, + 201 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 141, + 180, + 455, + 201 + ], + "type": "inline_equation", + "content": "+\\infty" + }, + { + "bbox": [ + 141, + 180, + 455, + 201 + ], + "type": "text", + "content": " otherwise." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 202, + 381, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 202, + 381, + 213 + ], + "spans": [ + { + "bbox": [ + 141, + 202, + 381, + 213 + ], + "type": "text", + "content": "2. Create a " + }, + { + "bbox": [ + 141, + 202, + 381, + 213 + ], + "type": "inline_equation", + "content": "n \\times n" + }, + { + "bbox": [ + 141, + 202, + 381, + 213 + ], + "type": "text", + "content": " matrix " + }, + { + "bbox": [ + 141, + 202, + 381, + 213 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 141, + 202, + 381, + 213 + ], + "type": "text", + "content": " for the function " + }, + { + "bbox": [ + 141, + 202, + 381, + 213 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 141, + 202, + 381, + 213 + ], + "type": "text", + "content": " analogously;" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 213, + 386, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 213, + 386, + 223 + ], + "spans": [ + { + "bbox": [ + 141, + 213, + 386, + 223 + ], + "type": "text", + "content": "3. Initialize " + }, + { + "bbox": [ + 141, + 213, + 386, + 223 + ], + "type": "inline_equation", + "content": "c_{+\\infty}" + }, + { + "bbox": [ + 141, + 213, + 386, + 223 + ], + "type": "text", + "content": " to a vector of " + }, + { + "bbox": [ + 141, + 213, + 386, + 223 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 141, + 213, + 386, + 223 + ], + "type": "text", + "content": " elements, all equal to " + }, + { + "bbox": [ + 141, + 213, + 386, + 223 + ], + "type": "inline_equation", + "content": "+\\infty" + }, + { + "bbox": [ + 141, + 213, + 386, + 223 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 224, + 477, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 224, + 477, + 234 + ], + "spans": [ + { + "bbox": [ + 141, + 224, + 477, + 234 + ], + "type": "text", + "content": "4. Create " + }, + { + "bbox": [ + 141, + 224, + 477, + 234 + ], + "type": "inline_equation", + "content": "F_{+}" + }, + { + "bbox": [ + 141, + 224, + 477, + 234 + ], + "type": "text", + "content": ", a matrix derived from " + }, + { + "bbox": [ + 141, + 224, + 477, + 234 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 141, + 224, + 477, + 234 + ], + "type": "text", + "content": " with a lower-diagonal part replaced by " + }, + { + "bbox": [ + 141, + 224, + 477, + 234 + ], + "type": "inline_equation", + "content": "+\\infty" + }, + { + "bbox": [ + 141, + 224, + 477, + 234 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 235, + 401, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 235, + 401, + 245 + ], + "spans": [ + { + "bbox": [ + 141, + 235, + 401, + 245 + ], + "type": "text", + "content": "5. Create " + }, + { + "bbox": [ + 141, + 235, + 401, + 245 + ], + "type": "inline_equation", + "content": "F_{diag}" + }, + { + "bbox": [ + 141, + 235, + 401, + 245 + ], + "type": "text", + "content": ", a vector containing the diagonal elements of " + }, + { + "bbox": [ + 141, + 235, + 401, + 245 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 141, + 235, + 401, + 245 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 246, + 400, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 246, + 400, + 256 + ], + "spans": [ + { + "bbox": [ + 141, + 246, + 400, + 256 + ], + "type": "text", + "content": "6. Compute " + }, + { + "bbox": [ + 141, + 246, + 400, + 256 + ], + "type": "inline_equation", + "content": "\\min (F,G)" + }, + { + "bbox": [ + 141, + 246, + 400, + 256 + ], + "type": "text", + "content": ", the element-wise minimum of " + }, + { + "bbox": [ + 141, + 246, + 400, + 256 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 141, + 246, + 400, + 256 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 141, + 246, + 400, + 256 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 141, + 246, + 400, + 256 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 257, + 419, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 257, + 419, + 267 + ], + "spans": [ + { + "bbox": [ + 141, + 257, + 419, + 267 + ], + "type": "text", + "content": "7. Compute " + }, + { + "bbox": [ + 141, + 257, + 419, + 267 + ], + "type": "inline_equation", + "content": "\\min_{F,G}" + }, + { + "bbox": [ + 141, + 257, + 419, + 267 + ], + "type": "text", + "content": ", the minimum value of all elements in " + }, + { + "bbox": [ + 141, + 257, + 419, + 267 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 141, + 257, + 419, + 267 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 141, + 257, + 419, + 267 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 141, + 257, + 419, + 267 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 268, + 397, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 268, + 397, + 277 + ], + "spans": [ + { + "bbox": [ + 141, + 268, + 397, + 277 + ], + "type": "text", + "content": "8. Compute the " + }, + { + "bbox": [ + 141, + 268, + 397, + 277 + ], + "type": "inline_equation", + "content": "(2n + 1)\\times (2n + 1)" + }, + { + "bbox": [ + 141, + 268, + 397, + 277 + ], + "type": "text", + "content": " block matrix " + }, + { + "bbox": [ + 141, + 268, + 397, + 277 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 141, + 268, + 397, + 277 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 242, + 285, + 377, + 321 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 285, + 377, + 321 + ], + "spans": [ + { + "bbox": [ + 242, + 285, + 377, + 321 + ], + "type": "interline_equation", + "content": "m \\leftarrow \\left( \\begin{array}{c c c} \\min (F, G) & F _ {+} ^ {\\intercal} & c _ {+ \\infty} ^ {\\intercal} \\\\ F _ {+} & F & F _ {d i a g} ^ {\\intercal} \\\\ c _ {+ \\infty} & F _ {d i a g} & \\min _ {F, G} \\end{array} \\right)", + "image_path": "8023cde02843ebd7b50c10522b450e5f797bf7bbe25e1aeed2251935da06f82e.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 140, + 333, + 333, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 333, + 333, + 344 + ], + "spans": [ + { + "bbox": [ + 140, + 333, + 333, + 344 + ], + "type": "text", + "content": "9. Compute: F-Cross-Barcode" + }, + { + "bbox": [ + 140, + 333, + 333, + 344 + ], + "type": "inline_equation", + "content": "_k(f, g) \\gets \\mathsf{PB}_k(m)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 347, + 435, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 435, + 359 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 435, + 359 + ], + "type": "text", + "content": "Return: F-Cross-Barcode" + }, + { + "bbox": [ + 132, + 347, + 435, + 359 + ], + "type": "inline_equation", + "content": "_k(f, g)" + }, + { + "bbox": [ + 132, + 347, + 435, + 359 + ], + "type": "text", + "content": ", a multi-set of (\"birth\", \"death\") intervals." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 386, + 481, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 386, + 481, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 386, + 481, + 422 + ], + "type": "text", + "content": "in nodes of this grid: " + }, + { + "bbox": [ + 130, + 386, + 481, + 422 + ], + "type": "inline_equation", + "content": "f(j_{1},\\ldots ,j_{n}),g(j_{1},\\ldots ,j_{n})" + }, + { + "bbox": [ + 130, + 386, + 481, + 422 + ], + "type": "text", + "content": ". To compare two function " + }, + { + "bbox": [ + 130, + 386, + 481, + 422 + ], + "type": "inline_equation", + "content": "f,g" + }, + { + "bbox": [ + 130, + 386, + 481, + 422 + ], + "type": "text", + "content": " we create an extended " + }, + { + "bbox": [ + 130, + 386, + 481, + 422 + ], + "type": "inline_equation", + "content": "n + 1" + }, + { + "bbox": [ + 130, + 386, + 481, + 422 + ], + "type": "text", + "content": " dimensional lattice " + }, + { + "bbox": [ + 130, + 386, + 481, + 422 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{L}}" + }, + { + "bbox": [ + 130, + 386, + 481, + 422 + ], + "type": "text", + "content": " with one extra dimension of size 3, see Figure 2b. The vertex function " + }, + { + "bbox": [ + 130, + 386, + 481, + 422 + ], + "type": "inline_equation", + "content": "\\tilde{f}:\\tilde{\\mathcal{L}}\\to \\mathbb{R}" + }, + { + "bbox": [ + 130, + 386, + 481, + 422 + ], + "type": "text", + "content": " is defined as follows:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 171, + 431, + 440, + 476 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 431, + 440, + 476 + ], + "spans": [ + { + "bbox": [ + 171, + 431, + 440, + 476 + ], + "type": "interline_equation", + "content": "\\tilde {f} (i, j _ {1}, \\ldots , j _ {n}) = \\left\\{ \\begin{array}{l l} \\min _ {f, g} & \\text {f o r} i = 0 \\\\ f (j _ {1}, \\ldots , j _ {n}) & \\text {f o r} i = 1 \\\\ \\min (f (j _ {1}, \\ldots , j _ {n}), g (j _ {1}, \\ldots , j _ {n})) & \\text {f o r} i = 2 \\end{array} \\right.", + "image_path": "d3cccf565e07468c4e03eae3d230afa851714517fe45a6ca6137b221af62fad7.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 486, + 482, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 558 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 558 + ], + "type": "text", + "content": "Similarly to functions on graphs, " + }, + { + "bbox": [ + 130, + 486, + 482, + 558 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 486, + 482, + 558 + ], + "type": "text", + "content": "-Cross-Barcode" + }, + { + "bbox": [ + 130, + 486, + 482, + 558 + ], + "type": "inline_equation", + "content": "_k(f,g)" + }, + { + "bbox": [ + 130, + 486, + 482, + 558 + ], + "type": "text", + "content": " for functions on lattices is defined as the " + }, + { + "bbox": [ + 130, + 486, + 482, + 558 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 486, + 482, + 558 + ], + "type": "text", + "content": "-th persistence barcode of the filtered cubical complex with the filtration induced by the vertex function " + }, + { + "bbox": [ + 130, + 486, + 482, + 558 + ], + "type": "inline_equation", + "content": "\\tilde{f}" + }, + { + "bbox": [ + 130, + 486, + 482, + 558 + ], + "type": "text", + "content": ". We use the GUDHI" + }, + { + "bbox": [ + 130, + 486, + 482, + 558 + ], + "type": "inline_equation", + "content": "^6" + }, + { + "bbox": [ + 130, + 486, + 482, + 558 + ], + "type": "text", + "content": " library which implements efficient algorithms for cubical complexes [35]. Note that " + }, + { + "bbox": [ + 130, + 486, + 482, + 558 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 486, + 482, + 558 + ], + "type": "text", + "content": "-Cross-Barcode" + }, + { + "bbox": [ + 130, + 486, + 482, + 558 + ], + "type": "inline_equation", + "content": "_k(f,g)" + }, + { + "bbox": [ + 130, + 486, + 482, + 558 + ], + "type": "text", + "content": " for functions restricted on a lattice and a triangulation of the same space are not equivalent, but would be similar." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 577, + 348, + 589 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 577, + 348, + 589 + ], + "spans": [ + { + "bbox": [ + 132, + 577, + 348, + 589 + ], + "type": "text", + "content": "4.3 Scalar Function Topology Divergence" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 597, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 597, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 597, + 482, + 647 + ], + "type": "text", + "content": "The " + }, + { + "bbox": [ + 130, + 597, + 482, + 647 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 597, + 482, + 647 + ], + "type": "text", + "content": "-Cross-Barcode" + }, + { + "bbox": [ + 130, + 597, + 482, + 647 + ], + "type": "inline_equation", + "content": "_k(f, g)" + }, + { + "bbox": [ + 130, + 597, + 482, + 647 + ], + "type": "text", + "content": " depicts topological differences of functions' sublevel sets at multiple scales. The " + }, + { + "bbox": [ + 130, + 597, + 482, + 647 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 597, + 482, + 647 + ], + "type": "text", + "content": "-Cross-Barcode" + }, + { + "bbox": [ + 130, + 597, + 482, + 647 + ], + "type": "inline_equation", + "content": "_k(f, g)" + }, + { + "bbox": [ + 130, + 597, + 482, + 647 + ], + "type": "text", + "content": ", as a regular barcode, is a multi-set of intervals. One requires some vectorization technique to inject it into a machine learning pipeline. If one needs to describe a discrepancy between " + }, + { + "bbox": [ + 130, + 597, + 482, + 647 + ], + "type": "inline_equation", + "content": "f, g" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 294, + 91, + 447, + 102 + ], + "type": "text", + "content": "Scalar Function Topology Divergence" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 253, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 253, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 253, + 665 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 133, + 653, + 253, + 665 + ], + "type": "text", + "content": " https://gudhi.inria.fr/" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 140 + ], + "type": "text", + "content": "by an integral numerical characteristic, we propose a scalar function topology divergence:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 199, + 148, + 481, + 175 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 199, + 148, + 481, + 175 + ], + "spans": [ + { + "bbox": [ + 199, + 148, + 481, + 175 + ], + "type": "interline_equation", + "content": "\\mathrm {S F T D} _ {k} (f, g) = \\sum_ {(b, d) \\in \\mathrm {F} - \\text {C r o s s - B a r c o d e} _ {k} (f, g)} | | d - b | | ^ {p}, \\tag {1}", + "image_path": "17a93b8ae660a12e70756a1e1670c27bf316e0272564a3118d816bc9ad2c9583.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 182, + 482, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 482, + 268 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 482, + 268 + ], + "type": "text", + "content": "that is the sum of lengths of intervals from the " + }, + { + "bbox": [ + 130, + 182, + 482, + 268 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 182, + 482, + 268 + ], + "type": "text", + "content": "-Cross-Barcode" + }, + { + "bbox": [ + 130, + 182, + 482, + 268 + ], + "type": "inline_equation", + "content": "_k(f, g)" + }, + { + "bbox": [ + 130, + 182, + 482, + 268 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 182, + 482, + 268 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 182, + 482, + 268 + ], + "type": "text", + "content": "-th power, where typically " + }, + { + "bbox": [ + 130, + 182, + 482, + 268 + ], + "type": "inline_equation", + "content": "p = 1, 2" + }, + { + "bbox": [ + 130, + 182, + 482, + 268 + ], + "type": "text", + "content": ". SFTD is not symmetric, and we often use an average of " + }, + { + "bbox": [ + 130, + 182, + 482, + 268 + ], + "type": "inline_equation", + "content": "\\mathrm{SFTD}_k(f, g)" + }, + { + "bbox": [ + 130, + 182, + 482, + 268 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 182, + 482, + 268 + ], + "type": "inline_equation", + "content": "\\mathrm{SFTD}_k(g, f)" + }, + { + "bbox": [ + 130, + 182, + 482, + 268 + ], + "type": "text", + "content": " in experiments below. For different dimensions " + }, + { + "bbox": [ + 130, + 182, + 482, + 268 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 182, + 482, + 268 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 182, + 482, + 268 + ], + "type": "inline_equation", + "content": "\\mathrm{SFTD}_k(f, g)" + }, + { + "bbox": [ + 130, + 182, + 482, + 268 + ], + "type": "text", + "content": " can be used either separately, or the sum can be taken to reflect discrepancies in all types of topological features (connected components, cycles, voids, etc.). Computational complexity of SFTD is discussed in Appendix C." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 284, + 276, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 284, + 276, + 297 + ], + "spans": [ + { + "bbox": [ + 132, + 284, + 276, + 297 + ], + "type": "text", + "content": "4.4 SFTD basic properties" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 303, + 483, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 303, + 483, + 363 + ], + "spans": [ + { + "bbox": [ + 131, + 303, + 483, + 363 + ], + "type": "text", + "content": "Proposition 1. If " + }, + { + "bbox": [ + 131, + 303, + 483, + 363 + ], + "type": "inline_equation", + "content": "SFTD_{k}(f,g) = SFTD_{k}(g,f) = 0" + }, + { + "bbox": [ + 131, + 303, + 483, + 363 + ], + "type": "text", + "content": " for all " + }, + { + "bbox": [ + 131, + 303, + 483, + 363 + ], + "type": "inline_equation", + "content": "k\\geq 0" + }, + { + "bbox": [ + 131, + 303, + 483, + 363 + ], + "type": "text", + "content": ", then the sublevel sets barcodes of functions " + }, + { + "bbox": [ + 131, + 303, + 483, + 363 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 131, + 303, + 483, + 363 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 303, + 483, + 363 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 131, + 303, + 483, + 363 + ], + "type": "text", + "content": " are the same in any degree. Moreover, in this case their topological features are located in the same places: the inclusions " + }, + { + "bbox": [ + 131, + 303, + 483, + 363 + ], + "type": "inline_equation", + "content": "f^{-1}((- \\infty ,\\varepsilon ])\\subseteq (\\min (f,g))^{-1}((- \\infty ,\\varepsilon ])" + }, + { + "bbox": [ + 131, + 303, + 483, + 363 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 303, + 483, + 363 + ], + "type": "inline_equation", + "content": "g^{-1}((- \\infty ,\\varepsilon ])\\subseteq (\\min (f,g))^{-1}((- \\infty ,\\varepsilon ])" + }, + { + "bbox": [ + 131, + 303, + 483, + 363 + ], + "type": "text", + "content": ", induce homology isomorphisms for any threshold " + }, + { + "bbox": [ + 131, + 303, + 483, + 363 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 131, + 303, + 483, + 363 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 371, + 296, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 371, + 296, + 383 + ], + "spans": [ + { + "bbox": [ + 132, + 371, + 296, + 383 + ], + "type": "text", + "content": "Proposition 2. (SFTD Stability)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 147, + 384, + 476, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 384, + 476, + 396 + ], + "spans": [ + { + "bbox": [ + 147, + 384, + 476, + 396 + ], + "type": "text", + "content": "(a) For any scalar functions " + }, + { + "bbox": [ + 147, + 384, + 476, + 396 + ], + "type": "inline_equation", + "content": "f, \\tilde{f}, g, \\tilde{g}" + }, + { + "bbox": [ + 147, + 384, + 476, + 396 + ], + "type": "text", + "content": " on a grid, or on a graph vertex set:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 400, + 470, + 432 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 400, + 470, + 432 + ], + "spans": [ + { + "bbox": [ + 141, + 400, + 470, + 432 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} d _ {B} (F - C r o s s - B a r c o d e _ {k} (f, g), F - C r o s s - B a r c o d e _ {k} \\left(f ^ {\\prime}, g ^ {\\prime}\\right)) \\leq \\\\ \\max (\\max _ {i} | f (i) - f ^ {\\prime} (i) |, \\max _ {i} | g (i) - g ^ {\\prime} (i) |). \\\\ \\end{array}", + "image_path": "c145afc213cf4719bb2457b2a1348c6088d218d3ab53f6b2bc47d66c3d9c2048.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 147, + 430, + 333, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 430, + 333, + 442 + ], + "spans": [ + { + "bbox": [ + 147, + 430, + 333, + 442 + ], + "type": "text", + "content": "(b) For any pair of scalar functions " + }, + { + "bbox": [ + 147, + 430, + 333, + 442 + ], + "type": "inline_equation", + "content": "f, \\hat{f}'" + }, + { + "bbox": [ + 147, + 430, + 333, + 442 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 200, + 446, + 413, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 446, + 413, + 464 + ], + "spans": [ + { + "bbox": [ + 200, + 446, + 413, + 464 + ], + "type": "interline_equation", + "content": "\\| F \\text {- C r o s s - B a r c o d e} _ {k} (f, f ^ {\\prime}) \\| _ {B} \\leq \\max _ {i} | f (i) - f ^ {\\prime} (i) |,", + "image_path": "efbcefd694dbf24ac6893bc3146f387bacddcbf8d44f13af1ab41f515f3d6cf3.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 475, + 480, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 475, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 131, + 475, + 480, + 500 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 475, + 480, + 500 + ], + "type": "inline_equation", + "content": "d_B" + }, + { + "bbox": [ + 131, + 475, + 480, + 500 + ], + "type": "text", + "content": " denotes the bottleneck distance between persistence barcodes [13]. We give the proofs in Appendix A. As a consequence, SFTD is robust to noise." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 517, + 332, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 517, + 332, + 529 + ], + "spans": [ + { + "bbox": [ + 132, + 517, + 332, + 529 + ], + "type": "text", + "content": "4.5 Visualization of F-Cross-Barcodes" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 536, + 481, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 536, + 481, + 561 + ], + "spans": [ + { + "bbox": [ + 131, + 536, + 481, + 561 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 131, + 536, + 481, + 561 + ], + "type": "text", + "content": "-Cross-Barcodes can be used for localization and visualization of topological differences between scalar functions. Recall the filtered cubical7 complex:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 252, + 569, + 360, + 582 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 569, + 360, + 582 + ], + "spans": [ + { + "bbox": [ + 252, + 569, + 360, + 582 + ], + "type": "interline_equation", + "content": "\\varnothing \\subseteq \\mathcal {C} _ {0} \\subseteq \\mathcal {C} _ {1} \\subseteq \\dots \\subseteq \\mathcal {C} _ {n}.", + "image_path": "641826dba7d4f6c377974dd3d5a46b4e70ce5ce31a6be24e1db64547ddbe274e.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 590, + 481, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 590, + 481, + 615 + ], + "spans": [ + { + "bbox": [ + 130, + 590, + 481, + 615 + ], + "type": "text", + "content": "Since the filtration is induced by a vertex function " + }, + { + "bbox": [ + 130, + 590, + 481, + 615 + ], + "type": "inline_equation", + "content": "\\tilde{f}:\\tilde{\\mathcal{L}}\\to \\mathbb{R}" + }, + { + "bbox": [ + 130, + 590, + 481, + 615 + ], + "type": "text", + "content": ", the filtered cubical complex corresponds to the sequence:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 244, + 621, + 368, + 637 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 621, + 368, + 637 + ], + "spans": [ + { + "bbox": [ + 244, + 621, + 368, + 637 + ], + "type": "interline_equation", + "content": "\\tilde {f} (i _ {0}) \\leq \\tilde {f} (i _ {1}) \\leq \\dots \\leq \\tilde {f} (i _ {n}).", + "image_path": "ede55a10d629b2158c110cc416f9bc0a7b1e6949484cc1a240f73e18e1318698.jpg" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 91, + 237, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 237, + 101 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 237, + 101 + ], + "type": "text", + "content": "I. Trofimov et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 642, + 481, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 481, + 666 + ], + "type": "text", + "content": "For ease of presentation, we discuss only the visualization for cubical complexes. The visualization for clique complexes is analogous." + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 212 + ], + "type": "text", + "content": "If " + }, + { + "bbox": [ + 130, + 115, + 479, + 212 + ], + "type": "inline_equation", + "content": "\\tilde{f}(i_k) = \\tilde{f}(i_m)" + }, + { + "bbox": [ + 130, + 115, + 479, + 212 + ], + "type": "text", + "content": ", the order is defined by indexes " + }, + { + "bbox": [ + 130, + 115, + 479, + 212 + ], + "type": "inline_equation", + "content": "i_k, i_m" + }, + { + "bbox": [ + 130, + 115, + 479, + 212 + ], + "type": "text", + "content": ". Thus, each segment in the barcode - a birth/death pair - is associated with two vertices " + }, + { + "bbox": [ + 130, + 115, + 479, + 212 + ], + "type": "inline_equation", + "content": "v_b, v_d \\in \\tilde{\\mathcal{L}}" + }, + { + "bbox": [ + 130, + 115, + 479, + 212 + ], + "type": "text", + "content": ". Recall that " + }, + { + "bbox": [ + 130, + 115, + 479, + 212 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{L}}" + }, + { + "bbox": [ + 130, + 115, + 479, + 212 + ], + "type": "text", + "content": " is an extended " + }, + { + "bbox": [ + 130, + 115, + 479, + 212 + ], + "type": "inline_equation", + "content": "n + 1" + }, + { + "bbox": [ + 130, + 115, + 479, + 212 + ], + "type": "text", + "content": " dimensional lattice (see Section 4.2), where the first dimension is of size 3, and the rest of dimensions are of the same size as the lattice inside the domain of " + }, + { + "bbox": [ + 130, + 115, + 479, + 212 + ], + "type": "inline_equation", + "content": "f, g" + }, + { + "bbox": [ + 130, + 115, + 479, + 212 + ], + "type": "text", + "content": ". So, neglecting the first dimension, we can map the rest of indices " + }, + { + "bbox": [ + 130, + 115, + 479, + 212 + ], + "type": "inline_equation", + "content": "j_1, \\ldots, j_n" + }, + { + "bbox": [ + 130, + 115, + 479, + 212 + ], + "type": "text", + "content": " of birth/death vertices " + }, + { + "bbox": [ + 130, + 115, + 479, + 212 + ], + "type": "inline_equation", + "content": "v_b, v_d" + }, + { + "bbox": [ + 130, + 115, + 479, + 212 + ], + "type": "text", + "content": " right onto original " + }, + { + "bbox": [ + 130, + 115, + 479, + 212 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 115, + 479, + 212 + ], + "type": "text", + "content": "-dimensional lattice. Later, in the experimental section, we shall provide such visualizations, where birth/death events are depicted by orange/red dots." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 228, + 265, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 228, + 265, + 239 + ], + "spans": [ + { + "bbox": [ + 131, + 228, + 265, + 239 + ], + "type": "text", + "content": "4.6 SFTD differentiation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 245, + 480, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 245, + 480, + 329 + ], + "spans": [ + { + "bbox": [ + 130, + 245, + 480, + 329 + ], + "type": "text", + "content": "Calculation of " + }, + { + "bbox": [ + 130, + 245, + 480, + 329 + ], + "type": "inline_equation", + "content": "\\mathrm{SFTD}_k(f,g)" + }, + { + "bbox": [ + 130, + 245, + 480, + 329 + ], + "type": "text", + "content": " involves a filtered simplicial(cubical) complex " + }, + { + "bbox": [ + 130, + 245, + 480, + 329 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 245, + 480, + 329 + ], + "type": "text", + "content": " with a vertex function " + }, + { + "bbox": [ + 130, + 245, + 480, + 329 + ], + "type": "inline_equation", + "content": "\\tilde{f} (i),i\\in \\tilde{\\mathcal{V}}" + }, + { + "bbox": [ + 130, + 245, + 480, + 329 + ], + "type": "text", + "content": " . Also, since we apply the sub-level set filtration, for every simplex(cube) " + }, + { + "bbox": [ + 130, + 245, + 480, + 329 + ], + "type": "inline_equation", + "content": "\\sigma \\colon T(\\sigma) = \\tilde{f} (i_{\\sigma})" + }, + { + "bbox": [ + 130, + 245, + 480, + 329 + ], + "type": "text", + "content": " , where " + }, + { + "bbox": [ + 130, + 245, + 480, + 329 + ], + "type": "inline_equation", + "content": "i_{\\sigma} = \\operatorname{argmax}_{i\\in \\sigma}\\tilde{f} (i)" + }, + { + "bbox": [ + 130, + 245, + 480, + 329 + ], + "type": "text", + "content": " . Each interval " + }, + { + "bbox": [ + 130, + 245, + 480, + 329 + ], + "type": "inline_equation", + "content": "(b,d)" + }, + { + "bbox": [ + 130, + 245, + 480, + 329 + ], + "type": "text", + "content": " in a " + }, + { + "bbox": [ + 130, + 245, + 480, + 329 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 245, + 480, + 329 + ], + "type": "text", + "content": " -Cross-Barcode corresponds to some pair of simplices(cubes) " + }, + { + "bbox": [ + 130, + 245, + 480, + 329 + ], + "type": "inline_equation", + "content": "(\\sigma_1,\\sigma_2)" + }, + { + "bbox": [ + 130, + 245, + 480, + 329 + ], + "type": "text", + "content": " which join the simplicial(cubical) complex " + }, + { + "bbox": [ + 130, + 245, + 480, + 329 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 245, + 480, + 329 + ], + "type": "text", + "content": " at moments " + }, + { + "bbox": [ + 130, + 245, + 480, + 329 + ], + "type": "inline_equation", + "content": "b = T(\\sigma_1),d = T(\\sigma_2)" + }, + { + "bbox": [ + 130, + 245, + 480, + 329 + ], + "type": "text", + "content": " Generally, we consider " + }, + { + "bbox": [ + 130, + 245, + 480, + 329 + ], + "type": "inline_equation", + "content": "f,g" + }, + { + "bbox": [ + 130, + 245, + 480, + 329 + ], + "type": "text", + "content": " (and consequently " + }, + { + "bbox": [ + 130, + 245, + 480, + 329 + ], + "type": "inline_equation", + "content": "\\tilde{f}" + }, + { + "bbox": [ + 130, + 245, + 480, + 329 + ], + "type": "text", + "content": ") to be parametric functions of " + }, + { + "bbox": [ + 130, + 245, + 480, + 329 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 130, + 245, + 480, + 329 + ], + "type": "text", + "content": " , for example, neural networks. By differentiating (1) we obtain:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 186, + 337, + 480, + 372 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 337, + 480, + 372 + ], + "spans": [ + { + "bbox": [ + 186, + 337, + 480, + 372 + ], + "type": "interline_equation", + "content": "\\frac {\\partial \\operatorname {S F T D} _ {k} (f , g)}{\\partial w} = \\sum_ {\\left(\\sigma_ {1}, \\sigma_ {2}\\right) \\in \\Sigma_ {k}} p \\left(\\frac {\\partial \\tilde {f} \\left(i _ {\\sigma_ {2}}\\right)}{\\partial w} - \\frac {\\partial \\tilde {f} \\left(i _ {\\sigma_ {2}}\\right)}{\\partial w}\\right) ^ {p - 1}, \\tag {2}", + "image_path": "1dd481579fc15f50a609ac8d9d498e0b93abe067711f02b2078bc9c93373b800.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 378, + 479, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 378, + 479, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 378, + 479, + 402 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 378, + 479, + 402 + ], + "type": "inline_equation", + "content": "\\Sigma_{k}" + }, + { + "bbox": [ + 130, + 378, + 479, + 402 + ], + "type": "text", + "content": " is a collection of \"birth\", \"death\" simplex pairs corresponding to intervals in the " + }, + { + "bbox": [ + 130, + 378, + 479, + 402 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 378, + 479, + 402 + ], + "type": "text", + "content": "-Cross-Barcode" + }, + { + "bbox": [ + 130, + 378, + 479, + 402 + ], + "type": "inline_equation", + "content": "_k(f, g)" + }, + { + "bbox": [ + 130, + 378, + 479, + 402 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 402, + 479, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 479, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 479, + 426 + ], + "type": "text", + "content": "By construction, " + }, + { + "bbox": [ + 130, + 402, + 479, + 426 + ], + "type": "inline_equation", + "content": "\\tilde{f}(i)" + }, + { + "bbox": [ + 130, + 402, + 479, + 426 + ], + "type": "text", + "content": " can be either " + }, + { + "bbox": [ + 130, + 402, + 479, + 426 + ], + "type": "inline_equation", + "content": "f(i')" + }, + { + "bbox": [ + 130, + 402, + 479, + 426 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 130, + 402, + 479, + 426 + ], + "type": "inline_equation", + "content": "g(i')" + }, + { + "bbox": [ + 130, + 402, + 479, + 426 + ], + "type": "text", + "content": " for some mapping " + }, + { + "bbox": [ + 130, + 402, + 479, + 426 + ], + "type": "inline_equation", + "content": "i' = r(i)" + }, + { + "bbox": [ + 130, + 402, + 479, + 426 + ], + "type": "text", + "content": ". Formally, introducing a binary function " + }, + { + "bbox": [ + 130, + 402, + 479, + 426 + ], + "type": "inline_equation", + "content": "c(\\cdot)" + }, + { + "bbox": [ + 130, + 402, + 479, + 426 + ], + "type": "text", + "content": " for this choice:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 227, + 433, + 480, + 447 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 433, + 480, + 447 + ], + "spans": [ + { + "bbox": [ + 227, + 433, + 480, + 447 + ], + "type": "interline_equation", + "content": "\\tilde {f} (i) = c (i) f (r (i)) + (1 - c (i)) g (r (i)), \\tag {3}", + "image_path": "2d6eca865ec4e5054c6db7e454b8e464077709d5d902def6d5bfd8260d6f3b8e.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 453, + 151, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 453, + 151, + 463 + ], + "spans": [ + { + "bbox": [ + 131, + 453, + 151, + 463 + ], + "type": "text", + "content": "and" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 216, + 463, + 480, + 487 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 463, + 480, + 487 + ], + "spans": [ + { + "bbox": [ + 216, + 463, + 480, + 487 + ], + "type": "interline_equation", + "content": "\\frac {\\partial \\tilde {f} (i)}{\\partial w} = b (i) \\frac {\\partial f (r (i))}{\\partial w} + (1 - b (i)) \\frac {\\partial g (r (i))}{\\partial w}. \\tag {4}", + "image_path": "f51b63fdd1e2bca47bab0d223146ade69e7fa3881ad6693907dc9edb4b2ffe8e.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 490, + 479, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 490, + 479, + 550 + ], + "spans": [ + { + "bbox": [ + 130, + 490, + 479, + 550 + ], + "type": "text", + "content": "Finally, the desirable gradient " + }, + { + "bbox": [ + 130, + 490, + 479, + 550 + ], + "type": "inline_equation", + "content": "\\partial \\mathrm{SFTD}_k(f,g) / \\partial w" + }, + { + "bbox": [ + 130, + 490, + 479, + 550 + ], + "type": "text", + "content": " can be calculated by combining (2), (4) via a chain rule. With a common assumption that all the " + }, + { + "bbox": [ + 130, + 490, + 479, + 550 + ], + "type": "inline_equation", + "content": "f(i),g(i)" + }, + { + "bbox": [ + 130, + 490, + 479, + 550 + ], + "type": "text", + "content": " are distinct, functions " + }, + { + "bbox": [ + 130, + 490, + 479, + 550 + ], + "type": "inline_equation", + "content": "c,r" + }, + { + "bbox": [ + 130, + 490, + 479, + 550 + ], + "type": "text", + "content": " don't change in a vicinity of " + }, + { + "bbox": [ + 130, + 490, + 479, + 550 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 130, + 490, + 479, + 550 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 490, + 479, + 550 + ], + "type": "inline_equation", + "content": "\\mathrm{SFTD}_i(f,g)" + }, + { + "bbox": [ + 130, + 490, + 479, + 550 + ], + "type": "text", + "content": " is differentiable w.r.t. " + }, + { + "bbox": [ + 130, + 490, + 479, + 550 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 130, + 490, + 479, + 550 + ], + "type": "text", + "content": " [11, 22]. On a high level, the differentiation is done similarly to [11, 34]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 566, + 230, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 566, + 230, + 580 + ], + "spans": [ + { + "bbox": [ + 131, + 566, + 230, + 580 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 590, + 470, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 590, + 470, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 590, + 470, + 601 + ], + "type": "text", + "content": "In the experimental section, we are going to answer three research questions:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 126, + 606, + 479, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 126, + 606, + 479, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 606, + 479, + 640 + ], + "spans": [ + { + "bbox": [ + 126, + 606, + 479, + 640 + ], + "type": "text", + "content": "Q1: Does SFTD and F-Cross-Barcodes provide a better topologically-aware comparison of scalar functions than Wasserstein distance between regular persistence barcodes?" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 126, + 642, + 479, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 642, + 479, + 653 + ], + "spans": [ + { + "bbox": [ + 126, + 642, + 479, + 653 + ], + "type": "text", + "content": "Q2: Does F-Cross-Barcodes help to visualize and localize topological differences?" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 126, + 653, + 452, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 653, + 452, + 665 + ], + "spans": [ + { + "bbox": [ + 126, + 653, + 452, + 665 + ], + "type": "text", + "content": "Q3: Is SFTD, as a loss function, beneficial for computer vision problems?" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 294, + 91, + 447, + 102 + ], + "type": "text", + "content": "Scalar Function Topology Divergence" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 117, + 244, + 207 + ], + "blocks": [ + { + "bbox": [ + 135, + 117, + 244, + 207 + ], + "lines": [ + { + "bbox": [ + 135, + 117, + 244, + 207 + ], + "spans": [ + { + "bbox": [ + 135, + 117, + 244, + 207 + ], + "type": "image", + "image_path": "d2ab50479442e57f66504be1fe8b88937b063abfda85749b953715ee9645e657.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 155, + 211, + 223, + 220 + ], + "lines": [ + { + "bbox": [ + 155, + 211, + 223, + 220 + ], + "spans": [ + { + "bbox": [ + 155, + 211, + 223, + 220 + ], + "type": "text", + "content": "(a) Scalar func. " + }, + { + "bbox": [ + 155, + 211, + 223, + 220 + ], + "type": "inline_equation", + "content": "f_{1}" + }, + { + "bbox": [ + 155, + 211, + 223, + 220 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 258, + 118, + 367, + 207 + ], + "blocks": [ + { + "bbox": [ + 258, + 118, + 367, + 207 + ], + "lines": [ + { + "bbox": [ + 258, + 118, + 367, + 207 + ], + "spans": [ + { + "bbox": [ + 258, + 118, + 367, + 207 + ], + "type": "image", + "image_path": "b70f9280df963f3a824aa917709274b1ea09aeb08e0da3354e1cb892d49e6bbe.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 278, + 211, + 347, + 220 + ], + "lines": [ + { + "bbox": [ + 278, + 211, + 347, + 220 + ], + "spans": [ + { + "bbox": [ + 278, + 211, + 347, + 220 + ], + "type": "text", + "content": "(b) Scalar func. " + }, + { + "bbox": [ + 278, + 211, + 347, + 220 + ], + "type": "inline_equation", + "content": "f_{2}" + }, + { + "bbox": [ + 278, + 211, + 347, + 220 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 386, + 116, + 479, + 207 + ], + "blocks": [ + { + "bbox": [ + 386, + 116, + 479, + 207 + ], + "lines": [ + { + "bbox": [ + 386, + 116, + 479, + 207 + ], + "spans": [ + { + "bbox": [ + 386, + 116, + 479, + 207 + ], + "type": "image", + "image_path": "f3d2c1e646d9b98fe2032cb8621e3a02038d4db259739dae6c2b4765898ea058.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 410, + 211, + 454, + 220 + ], + "lines": [ + { + "bbox": [ + 410, + 211, + 454, + 220 + ], + "spans": [ + { + "bbox": [ + 410, + 211, + 454, + 220 + ], + "type": "text", + "content": "(c) Vis. diff." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 136, + 223, + 244, + 273 + ], + "blocks": [ + { + "bbox": [ + 136, + 223, + 244, + 273 + ], + "lines": [ + { + "bbox": [ + 136, + 223, + 244, + 273 + ], + "spans": [ + { + "bbox": [ + 136, + 223, + 244, + 273 + ], + "type": "image", + "image_path": "a754ee879426c914d789658c3eb148241a5dc345ad01d9c3cc564e6bda99edd1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 151, + 277, + 227, + 287 + ], + "lines": [ + { + "bbox": [ + 151, + 277, + 227, + 287 + ], + "spans": [ + { + "bbox": [ + 151, + 277, + 227, + 287 + ], + "type": "text", + "content": "(d) Pers. barc. of " + }, + { + "bbox": [ + 151, + 277, + 227, + 287 + ], + "type": "inline_equation", + "content": "f_{1}" + }, + { + "bbox": [ + 151, + 277, + 227, + 287 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 253, + 223, + 360, + 273 + ], + "blocks": [ + { + "bbox": [ + 253, + 223, + 360, + 273 + ], + "lines": [ + { + "bbox": [ + 253, + 223, + 360, + 273 + ], + "spans": [ + { + "bbox": [ + 253, + 223, + 360, + 273 + ], + "type": "image", + "image_path": "b31e3284c1626ff870a0a5473ad4c7d58819b22497ea9e2786940ecaf07fbcd3.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 268, + 277, + 342, + 287 + ], + "lines": [ + { + "bbox": [ + 268, + 277, + 342, + 287 + ], + "spans": [ + { + "bbox": [ + 268, + 277, + 342, + 287 + ], + "type": "text", + "content": "(e) Pers. barc. of " + }, + { + "bbox": [ + 268, + 277, + 342, + 287 + ], + "type": "inline_equation", + "content": "f_{2}" + }, + { + "bbox": [ + 268, + 277, + 342, + 287 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 372, + 223, + 479, + 273 + ], + "blocks": [ + { + "bbox": [ + 372, + 223, + 479, + 273 + ], + "lines": [ + { + "bbox": [ + 372, + 223, + 479, + 273 + ], + "spans": [ + { + "bbox": [ + 372, + 223, + 479, + 273 + ], + "type": "image", + "image_path": "a6d5b7d1918b0c711b9b57a328283c4fd0a3d493a5ebbfb3fced4922e98715b9.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 388, + 277, + 461, + 286 + ], + "lines": [ + { + "bbox": [ + 388, + 277, + 461, + 286 + ], + "spans": [ + { + "bbox": [ + 388, + 277, + 461, + 286 + ], + "type": "text", + "content": "(f) F-Cross-Barcode" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 296, + 481, + 319 + ], + "lines": [ + { + "bbox": [ + 131, + 296, + 481, + 319 + ], + "spans": [ + { + "bbox": [ + 131, + 296, + 481, + 319 + ], + "type": "text", + "content": "Fig. 3: Two scalar functions (a), (b) are indistinguishable by their persistence barcodes (d), (e). At the same time, " + }, + { + "bbox": [ + 131, + 296, + 481, + 319 + ], + "type": "inline_equation", + "content": "\\mathrm{SFTD}_0(f_1,f_2) = 4.6" + }, + { + "bbox": [ + 131, + 296, + 481, + 319 + ], + "type": "text", + "content": " see (f)." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 339, + 234, + 351 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 339, + 234, + 351 + ], + "spans": [ + { + "bbox": [ + 131, + 339, + 234, + 351 + ], + "type": "text", + "content": "5.1 Synthetic data" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 357, + 480, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 357, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 357, + 480, + 392 + ], + "type": "text", + "content": "To illustrate the performance of the proposed tools, we start with synthetic data. Functions under comparison have coinciding persistence barcodes of all degrees and they are indistinguishable by standard Wasserstein matching." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 407, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 482, + 529 + ], + "type": "text", + "content": "Functions with local minima in 2D. First, we compare two scalar function on " + }, + { + "bbox": [ + 130, + 407, + 482, + 529 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 130, + 407, + 482, + 529 + ], + "type": "text", + "content": " grid, each of them having 3 local minima but in different locations, see Figure 3, (a), (b). Persistence barcodes of both functions are equal, see Figure 3, (d), (e): they contain two segments from " + }, + { + "bbox": [ + 130, + 407, + 482, + 529 + ], + "type": "inline_equation", + "content": "H_0" + }, + { + "bbox": [ + 130, + 407, + 482, + 529 + ], + "type": "text", + "content": " corresponding to local minima. Here and further, essential segments (which die at " + }, + { + "bbox": [ + 130, + 407, + 482, + 529 + ], + "type": "inline_equation", + "content": "+\\infty" + }, + { + "bbox": [ + 130, + 407, + 482, + 529 + ], + "type": "text", + "content": ") are omitted. The topology of the two functions is indistinguishable by standard persistence barcodes and their Wasserstein distance is zero. At the same time, F-Cross-Barcode between these scalar functions is non-empty for " + }, + { + "bbox": [ + 130, + 407, + 482, + 529 + ], + "type": "inline_equation", + "content": "H_0" + }, + { + "bbox": [ + 130, + 407, + 482, + 529 + ], + "type": "text", + "content": ", see Figure 3, (f). The visualization technique correctly points at places with different topology - the centers and borders of the distinguishing local minima." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 542, + 482, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 542, + 482, + 625 + ], + "spans": [ + { + "bbox": [ + 130, + 542, + 482, + 625 + ], + "type": "text", + "content": "Lattices in 2D. Next, we compare two lattice-style 2D scalar functions, see Figure 4, (a), (b). Lattices have defects in different places (top left and bottom right respectively). The persistence barcodes of both lattices are identical, see Figure 4, (d), (e): they contain 8 segments from " + }, + { + "bbox": [ + 130, + 542, + 482, + 625 + ], + "type": "inline_equation", + "content": "H_{1}" + }, + { + "bbox": [ + 130, + 542, + 482, + 625 + ], + "type": "text", + "content": " corresponding to independent cycles. At the same time, " + }, + { + "bbox": [ + 130, + 542, + 482, + 625 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 542, + 482, + 625 + ], + "type": "text", + "content": "-Cross-Barcode between two scalar functions is nonempty, see Figure 4, (f). The visualization technique correctly points at places with defects." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Concentric spheres in 3D. Finally, we study the performance of the proposed method by comparing 3D scalar functions. We compare two scalar functions," + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 91, + 237, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 237, + 100 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 237, + 100 + ], + "type": "text", + "content": "I. Trofimov et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 116, + 242, + 205 + ], + "blocks": [ + { + "bbox": [ + 136, + 116, + 242, + 205 + ], + "lines": [ + { + "bbox": [ + 136, + 116, + 242, + 205 + ], + "spans": [ + { + "bbox": [ + 136, + 116, + 242, + 205 + ], + "type": "image", + "image_path": "69856d4fcb13b7dd2e2d930a3e5643239ccc2db98283f567f6541dd3a2580df2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 155, + 209, + 223, + 218 + ], + "lines": [ + { + "bbox": [ + 155, + 209, + 223, + 218 + ], + "spans": [ + { + "bbox": [ + 155, + 209, + 223, + 218 + ], + "type": "text", + "content": "(a) Scalar func. " + }, + { + "bbox": [ + 155, + 209, + 223, + 218 + ], + "type": "inline_equation", + "content": "f_{1}" + }, + { + "bbox": [ + 155, + 209, + 223, + 218 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 260, + 116, + 369, + 205 + ], + "blocks": [ + { + "bbox": [ + 260, + 116, + 369, + 205 + ], + "lines": [ + { + "bbox": [ + 260, + 116, + 369, + 205 + ], + "spans": [ + { + "bbox": [ + 260, + 116, + 369, + 205 + ], + "type": "image", + "image_path": "0d5fdc209618e10ffc645065dd5c4b756addc53b05a881ca4922b28ffd87d1bb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 279, + 209, + 349, + 218 + ], + "lines": [ + { + "bbox": [ + 279, + 209, + 349, + 218 + ], + "spans": [ + { + "bbox": [ + 279, + 209, + 349, + 218 + ], + "type": "text", + "content": "(b) Scalar func. " + }, + { + "bbox": [ + 279, + 209, + 349, + 218 + ], + "type": "inline_equation", + "content": "f_{2}" + }, + { + "bbox": [ + 279, + 209, + 349, + 218 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 388, + 116, + 479, + 205 + ], + "blocks": [ + { + "bbox": [ + 388, + 116, + 479, + 205 + ], + "lines": [ + { + "bbox": [ + 388, + 116, + 479, + 205 + ], + "spans": [ + { + "bbox": [ + 388, + 116, + 479, + 205 + ], + "type": "image", + "image_path": "f4ceb1ae7e305e8818135444cafec40f30f0f66a97538e2738035e2434c7fab6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 410, + 209, + 456, + 218 + ], + "lines": [ + { + "bbox": [ + 410, + 209, + 456, + 218 + ], + "spans": [ + { + "bbox": [ + 410, + 209, + 456, + 218 + ], + "type": "text", + "content": "(c) Vis. diff." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 136, + 220, + 242, + 269 + ], + "blocks": [ + { + "bbox": [ + 136, + 220, + 242, + 269 + ], + "lines": [ + { + "bbox": [ + 136, + 220, + 242, + 269 + ], + "spans": [ + { + "bbox": [ + 136, + 220, + 242, + 269 + ], + "type": "image", + "image_path": "3e51a20a812773a5e116b5370ae3c193c029babc5ccc44712bd4640b75b38b68.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 151, + 274, + 227, + 283 + ], + "lines": [ + { + "bbox": [ + 151, + 274, + 227, + 283 + ], + "spans": [ + { + "bbox": [ + 151, + 274, + 227, + 283 + ], + "type": "text", + "content": "(d) Pers. barc. of " + }, + { + "bbox": [ + 151, + 274, + 227, + 283 + ], + "type": "inline_equation", + "content": "f_1" + }, + { + "bbox": [ + 151, + 274, + 227, + 283 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 258, + 220, + 364, + 269 + ], + "blocks": [ + { + "bbox": [ + 258, + 220, + 364, + 269 + ], + "lines": [ + { + "bbox": [ + 258, + 220, + 364, + 269 + ], + "spans": [ + { + "bbox": [ + 258, + 220, + 364, + 269 + ], + "type": "image", + "image_path": "eb57e2f6b538b268e340fd22af3b0a3bc60fd1edc058695c44775a6fd6c450de.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 273, + 274, + 347, + 283 + ], + "lines": [ + { + "bbox": [ + 273, + 274, + 347, + 283 + ], + "spans": [ + { + "bbox": [ + 273, + 274, + 347, + 283 + ], + "type": "text", + "content": "(e) Pers. barc. of " + }, + { + "bbox": [ + 273, + 274, + 347, + 283 + ], + "type": "inline_equation", + "content": "f_{2}" + }, + { + "bbox": [ + 273, + 274, + 347, + 283 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 372, + 220, + 477, + 269 + ], + "blocks": [ + { + "bbox": [ + 372, + 220, + 477, + 269 + ], + "lines": [ + { + "bbox": [ + 372, + 220, + 477, + 269 + ], + "spans": [ + { + "bbox": [ + 372, + 220, + 477, + 269 + ], + "type": "image", + "image_path": "f465b508c3f4aa142f5e1a82f528282f6e0de4a9efb6e81032d4c98005f1a1a7.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 387, + 274, + 462, + 283 + ], + "lines": [ + { + "bbox": [ + 387, + 274, + 462, + 283 + ], + "spans": [ + { + "bbox": [ + 387, + 274, + 462, + 283 + ], + "type": "text", + "content": "(f) F-Cross-Barcode." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 292, + 482, + 316 + ], + "lines": [ + { + "bbox": [ + 131, + 292, + 482, + 316 + ], + "spans": [ + { + "bbox": [ + 131, + 292, + 482, + 316 + ], + "type": "text", + "content": "Fig. 4: Two scalar functions (a), (b) are indistinguishable by their persistence barcodes (c), (d). At the same time, " + }, + { + "bbox": [ + 131, + 292, + 482, + 316 + ], + "type": "inline_equation", + "content": "\\mathrm{SFTD}_1(f_1,f_2) = 1" + }, + { + "bbox": [ + 131, + 292, + 482, + 316 + ], + "type": "text", + "content": ", see (e)." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 338, + 482, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 338, + 482, + 459 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 482, + 459 + ], + "type": "text", + "content": "which equal " + }, + { + "bbox": [ + 130, + 338, + 482, + 459 + ], + "type": "inline_equation", + "content": "-1" + }, + { + "bbox": [ + 130, + 338, + 482, + 459 + ], + "type": "text", + "content": " in a vicinity of two concentric spheres and a bridge connecting them, which is depicted in green. The bridge is located in different places (above or below of the smallest sphere). In other areas of space, the scalar functions equal 0. We sample a " + }, + { + "bbox": [ + 130, + 338, + 482, + 459 + ], + "type": "inline_equation", + "content": "32^3" + }, + { + "bbox": [ + 130, + 338, + 482, + 459 + ], + "type": "text", + "content": " grid from the unit cube, see Figure 5 (a), (b) and compare functions restricted on this grid via " + }, + { + "bbox": [ + 130, + 338, + 482, + 459 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 338, + 482, + 459 + ], + "type": "text", + "content": "-Cross-Barcode. Again, regular persistence barcodes of two spheres with bridges are the same, see Figure 5, (d), (e): they contain 2 segments from " + }, + { + "bbox": [ + 130, + 338, + 482, + 459 + ], + "type": "inline_equation", + "content": "H_2" + }, + { + "bbox": [ + 130, + 338, + 482, + 459 + ], + "type": "text", + "content": " corresponding to two cavities. But " + }, + { + "bbox": [ + 130, + 338, + 482, + 459 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 338, + 482, + 459 + ], + "type": "text", + "content": "-Cross-Barcode is non-empty, see Figure 5, (f). The visualization technique correctly highlights areas with topological differences, see Figure 5 (c), where the central vertical slice is shown." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 131, + 477, + 384, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 477, + 384, + 490 + ], + "spans": [ + { + "bbox": [ + 131, + 477, + 384, + 490 + ], + "type": "text", + "content": "5.2 Comparing Eigenvectors of Graph Laplacian" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 498, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 616 + ], + "type": "text", + "content": "In the most general form, the proposed tools can compare scalar functions " + }, + { + "bbox": [ + 130, + 498, + 482, + 616 + ], + "type": "inline_equation", + "content": "f: \\mathcal{V} \\to \\mathbb{R}" + }, + { + "bbox": [ + 130, + 498, + 482, + 616 + ], + "type": "text", + "content": " defined on vertices " + }, + { + "bbox": [ + 130, + 498, + 482, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 130, + 498, + 482, + 616 + ], + "type": "text", + "content": " of an undirected graph " + }, + { + "bbox": [ + 130, + 498, + 482, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{G} = (\\mathcal{V}, \\mathcal{E})" + }, + { + "bbox": [ + 130, + 498, + 482, + 616 + ], + "type": "text", + "content": ". The normalized graph Laplacian is the matrix " + }, + { + "bbox": [ + 130, + 498, + 482, + 616 + ], + "type": "inline_equation", + "content": "L = D^{-1/2}(D - A)D^{-1/2}" + }, + { + "bbox": [ + 130, + 498, + 482, + 616 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 498, + 482, + 616 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 130, + 498, + 482, + 616 + ], + "type": "text", + "content": " is an adjacency matrix and " + }, + { + "bbox": [ + 130, + 498, + 482, + 616 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 498, + 482, + 616 + ], + "type": "text", + "content": " is a diagonal matrix of node degrees. Eigenvectors " + }, + { + "bbox": [ + 130, + 498, + 482, + 616 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 130, + 498, + 482, + 616 + ], + "type": "text", + "content": " and eigenvalues " + }, + { + "bbox": [ + 130, + 498, + 482, + 616 + ], + "type": "inline_equation", + "content": "\\lambda_i" + }, + { + "bbox": [ + 130, + 498, + 482, + 616 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 130, + 498, + 482, + 616 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 498, + 482, + 616 + ], + "type": "text", + "content": " describe intrinsic properties of the graph " + }, + { + "bbox": [ + 130, + 498, + 482, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 130, + 498, + 482, + 616 + ], + "type": "text", + "content": ". Eigenvectors can be considered as scalar functions defined on graph's vertices " + }, + { + "bbox": [ + 130, + 498, + 482, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 130, + 498, + 482, + 616 + ], + "type": "text", + "content": ". For eigenvectors: " + }, + { + "bbox": [ + 130, + 498, + 482, + 616 + ], + "type": "inline_equation", + "content": "(e_i, e_j) = 0" + }, + { + "bbox": [ + 130, + 498, + 482, + 616 + ], + "type": "text", + "content": " and their norm is not defined; one can put " + }, + { + "bbox": [ + 130, + 498, + 482, + 616 + ], + "type": "inline_equation", + "content": "||e_i|| = 1" + }, + { + "bbox": [ + 130, + 498, + 482, + 616 + ], + "type": "text", + "content": " for convenience. Thus, " + }, + { + "bbox": [ + 130, + 498, + 482, + 616 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 130, + 498, + 482, + 616 + ], + "type": "text", + "content": " can't be distinguished by Euclidean distance or cosine distance. Alternatively, we can study " + }, + { + "bbox": [ + 130, + 498, + 482, + 616 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 130, + 498, + 482, + 616 + ], + "type": "text", + "content": " as functions on vertices and compare them via SFTD." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "We consider a random graph with 30 nodes, see Figure 6. Figure 7 presents symmetrized " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\sum_{k=0,1} \\mathrm{SFTD}_k(e_i, e_j)" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " via a heatmap, where " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "e_i, e_j" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " are graph Laplacian eigenvectors ordered by eigenvalues. We observed the following interesting patterns: block structure, high topological similarity in the top left block, low" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 294, + 91, + 448, + 102 + ], + "type": "text", + "content": "Scalar Function Topology Divergence" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 117, + 246, + 225 + ], + "blocks": [ + { + "bbox": [ + 133, + 117, + 246, + 225 + ], + "lines": [ + { + "bbox": [ + 133, + 117, + 246, + 225 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 246, + 225 + ], + "type": "image", + "image_path": "e88065fab56b5ab72ee89ef7a28fdb8e425d191fe5a58b44b17bab39f47d5acf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 155, + 227, + 223, + 236 + ], + "lines": [ + { + "bbox": [ + 155, + 227, + 223, + 236 + ], + "spans": [ + { + "bbox": [ + 155, + 227, + 223, + 236 + ], + "type": "text", + "content": "(a) Scalar func. " + }, + { + "bbox": [ + 155, + 227, + 223, + 236 + ], + "type": "inline_equation", + "content": "f_{1}" + }, + { + "bbox": [ + 155, + 227, + 223, + 236 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 250, + 117, + 362, + 225 + ], + "blocks": [ + { + "bbox": [ + 250, + 117, + 362, + 225 + ], + "lines": [ + { + "bbox": [ + 250, + 117, + 362, + 225 + ], + "spans": [ + { + "bbox": [ + 250, + 117, + 362, + 225 + ], + "type": "image", + "image_path": "52af2e937923e1cb598b78cd11c5b3b55bafb10c4be3a871ab820e35f05ec85b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 271, + 227, + 340, + 236 + ], + "lines": [ + { + "bbox": [ + 271, + 227, + 340, + 236 + ], + "spans": [ + { + "bbox": [ + 271, + 227, + 340, + 236 + ], + "type": "text", + "content": "(b) Scalar func. " + }, + { + "bbox": [ + 271, + 227, + 340, + 236 + ], + "type": "inline_equation", + "content": "f_{2}" + }, + { + "bbox": [ + 271, + 227, + 340, + 236 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 370, + 116, + 479, + 223 + ], + "blocks": [ + { + "bbox": [ + 370, + 116, + 479, + 223 + ], + "lines": [ + { + "bbox": [ + 370, + 116, + 479, + 223 + ], + "spans": [ + { + "bbox": [ + 370, + 116, + 479, + 223 + ], + "type": "image", + "image_path": "288d13bb79d66d0d21002e1a4e481af7b027fbf7ca53b9cf52bd59526f78e706.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 402, + 227, + 446, + 236 + ], + "lines": [ + { + "bbox": [ + 402, + 227, + 446, + 236 + ], + "spans": [ + { + "bbox": [ + 402, + 227, + 446, + 236 + ], + "type": "text", + "content": "(c) Vis. diff." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 136, + 239, + 243, + 288 + ], + "blocks": [ + { + "bbox": [ + 136, + 239, + 243, + 288 + ], + "lines": [ + { + "bbox": [ + 136, + 239, + 243, + 288 + ], + "spans": [ + { + "bbox": [ + 136, + 239, + 243, + 288 + ], + "type": "image", + "image_path": "00d52bcc64268c36432279fab4859f834b0afa42653cf9aa435796e4bb6fc218.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 152, + 293, + 227, + 303 + ], + "lines": [ + { + "bbox": [ + 152, + 293, + 227, + 303 + ], + "spans": [ + { + "bbox": [ + 152, + 293, + 227, + 303 + ], + "type": "text", + "content": "(d) Pers. barc. of " + }, + { + "bbox": [ + 152, + 293, + 227, + 303 + ], + "type": "inline_equation", + "content": "f_{1}" + }, + { + "bbox": [ + 152, + 293, + 227, + 303 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 258, + 239, + 365, + 288 + ], + "blocks": [ + { + "bbox": [ + 258, + 239, + 365, + 288 + ], + "lines": [ + { + "bbox": [ + 258, + 239, + 365, + 288 + ], + "spans": [ + { + "bbox": [ + 258, + 239, + 365, + 288 + ], + "type": "image", + "image_path": "4c239fb411e8e02483fbb62d468775514cc6ab94496356f93451cf40ced6ace4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 274, + 293, + 347, + 303 + ], + "lines": [ + { + "bbox": [ + 274, + 293, + 347, + 303 + ], + "spans": [ + { + "bbox": [ + 274, + 293, + 347, + 303 + ], + "type": "text", + "content": "(e) Pers. barc. of " + }, + { + "bbox": [ + 274, + 293, + 347, + 303 + ], + "type": "inline_equation", + "content": "f_{2}" + }, + { + "bbox": [ + 274, + 293, + 347, + 303 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 372, + 239, + 478, + 288 + ], + "blocks": [ + { + "bbox": [ + 372, + 239, + 478, + 288 + ], + "lines": [ + { + "bbox": [ + 372, + 239, + 478, + 288 + ], + "spans": [ + { + "bbox": [ + 372, + 239, + 478, + 288 + ], + "type": "image", + "image_path": "02497798d1a0434f9433960f9c691ae3fb834b7714344ec13c6da181f562bac3.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 388, + 293, + 462, + 302 + ], + "lines": [ + { + "bbox": [ + 388, + 293, + 462, + 302 + ], + "spans": [ + { + "bbox": [ + 388, + 293, + 462, + 302 + ], + "type": "text", + "content": "(f) F-Cross-Barcode." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 311, + 481, + 335 + ], + "lines": [ + { + "bbox": [ + 131, + 311, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 131, + 311, + 481, + 335 + ], + "type": "text", + "content": "Fig. 5: Two scalar fields (a), (b) are indistinguishable by their persistence barcodes (d), (e). At the same time, " + }, + { + "bbox": [ + 131, + 311, + 481, + 335 + ], + "type": "inline_equation", + "content": "\\mathrm{SFTD}_1(f_1,f_2) = 1" + }, + { + "bbox": [ + 131, + 311, + 481, + 335 + ], + "type": "text", + "content": ", see (f)." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 132, + 385, + 482, + 456 + ], + "blocks": [ + { + "bbox": [ + 157, + 346, + 455, + 357 + ], + "lines": [ + { + "bbox": [ + 157, + 346, + 455, + 357 + ], + "spans": [ + { + "bbox": [ + 157, + 346, + 455, + 357 + ], + "type": "text", + "content": "Table 1: Errors of 3D shape reconstruction for \"Red blood cell\" dataset." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 385, + 482, + 456 + ], + "lines": [ + { + "bbox": [ + 132, + 385, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 385, + 482, + 456 + ], + "type": "table", + "html": "
METHOD1-IoUVOLUMESURFACEROUGH.H0H1H2
SHAPR [37]0.4870.3530.2430.3610.0980.2410.065
SHAPR+W.D. [36]0.4670.2910.1840.2940.0980.2280.032
SHAPR+Σi=0,1,2SFTDi0.4490.2940.1760.2820.0790.2220.065
SHAPR+Σi=0,1,2,3SFTDi0.4530.2660.1890.2760.0790.2170.025
", + "image_path": "9fa32e4376aa87517b402463bcbf351c9a17b4d5d28c960841d25bfd77aa8ea5.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 490, + 482, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 490, + 482, + 538 + ], + "spans": [ + { + "bbox": [ + 130, + 490, + 482, + 538 + ], + "type": "text", + "content": "topological similarity in the bottom right block, and very interesting diagonal lines parallel to the main diagonal. Figure 8 shows two the most similar eigenvectors, Figure 9 - two the most dissimilar ones, while their " + }, + { + "bbox": [ + 130, + 490, + 482, + 538 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 490, + 482, + 538 + ], + "type": "text", + "content": "-Cross-Barcodes are shown on the right." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 131, + 559, + 308, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 559, + 308, + 571 + ], + "spans": [ + { + "bbox": [ + 131, + 559, + 308, + 571 + ], + "type": "text", + "content": "5.3 Enhancing 3D reconstruction" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "SHAPR (SHApe PRediction) [37] is a deep learning model for reconstructing cellular and nuclear 3D shapes from 2D fluorescence microscopy images. Correct reconstruction of structural patterns in 3D shapes is critical for this problem. However, such information is not captured by established pixel/voxel-wise loss terms, e.g. Dice loss. Later, [36] improved SHAPR by adding contextual information about the overall shape of an object. The work [36] proposed to augment SHAPR with a topological loss term - a Wasserstein-2 distance (W.D.) between" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 91, + 237, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 237, + 100 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 237, + 100 + ], + "type": "text", + "content": "I. Trofimov et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 146, + 131, + 288, + 229 + ], + "blocks": [ + { + "bbox": [ + 146, + 131, + 288, + 229 + ], + "lines": [ + { + "bbox": [ + 146, + 131, + 288, + 229 + ], + "spans": [ + { + "bbox": [ + 146, + 131, + 288, + 229 + ], + "type": "image", + "image_path": "bd867d0aaeea8902623d0961ada0e57876244b0a3835a0dc86618c04ade28df3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 247, + 301, + 269 + ], + "lines": [ + { + "bbox": [ + 132, + 247, + 301, + 269 + ], + "spans": [ + { + "bbox": [ + 132, + 247, + 301, + 269 + ], + "type": "text", + "content": "Fig. 6: A random Watts-Strogatz small-world graph [38] with 30 nodes." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 316, + 125, + 477, + 236 + ], + "blocks": [ + { + "bbox": [ + 316, + 125, + 477, + 236 + ], + "lines": [ + { + "bbox": [ + 316, + 125, + 477, + 236 + ], + "spans": [ + { + "bbox": [ + 316, + 125, + 477, + 236 + ], + "type": "image", + "image_path": "01c184984f677ff2b714151d26999c779135968763d86a385dbb426dec059b9f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 311, + 247, + 480, + 269 + ], + "lines": [ + { + "bbox": [ + 311, + 247, + 480, + 269 + ], + "spans": [ + { + "bbox": [ + 311, + 247, + 480, + 269 + ], + "type": "text", + "content": "Fig.7: Comparison of graph Laplacian eigenvectors via SFTD." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 140, + 298, + 211, + 354 + ], + "blocks": [ + { + "bbox": [ + 140, + 298, + 211, + 354 + ], + "lines": [ + { + "bbox": [ + 140, + 298, + 211, + 354 + ], + "spans": [ + { + "bbox": [ + 140, + 298, + 211, + 354 + ], + "type": "image", + "image_path": "cefa2dcf58c0425a033f3f8a1e9a114b99a350cd8cf1619175ddc30d96e55cc2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 143, + 360, + 208, + 369 + ], + "lines": [ + { + "bbox": [ + 143, + 360, + 208, + 369 + ], + "spans": [ + { + "bbox": [ + 143, + 360, + 208, + 369 + ], + "type": "text", + "content": "(a) Eigenvector 1." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 228, + 299, + 299, + 354 + ], + "blocks": [ + { + "bbox": [ + 228, + 299, + 299, + 354 + ], + "lines": [ + { + "bbox": [ + 228, + 299, + 299, + 354 + ], + "spans": [ + { + "bbox": [ + 228, + 299, + 299, + 354 + ], + "type": "image", + "image_path": "37f366c109bdd92513fb6ab6aa7e6113f270569fe7733cd4e6f9a363c8e350dc.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 231, + 360, + 297, + 369 + ], + "lines": [ + { + "bbox": [ + 231, + 360, + 297, + 369 + ], + "spans": [ + { + "bbox": [ + 231, + 360, + 297, + 369 + ], + "type": "text", + "content": "(b) Eigenvector 4." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 309, + 301, + 391, + 350 + ], + "blocks": [ + { + "bbox": [ + 309, + 301, + 391, + 350 + ], + "lines": [ + { + "bbox": [ + 309, + 301, + 391, + 350 + ], + "spans": [ + { + "bbox": [ + 309, + 301, + 391, + 350 + ], + "type": "image", + "image_path": "598c980bdb85d93577dfc354c882b5739a757599a06ed420f4bcb126566a9091.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 309, + 352, + 391, + 370 + ], + "lines": [ + { + "bbox": [ + 309, + 352, + 391, + 370 + ], + "spans": [ + { + "bbox": [ + 309, + 352, + 391, + 370 + ], + "type": "text", + "content": "(c) F-Cross-Barcode " + }, + { + "bbox": [ + 309, + 352, + 391, + 370 + ], + "type": "inline_equation", + "content": "(e_1,e_4)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 238, + 378, + 374, + 389 + ], + "lines": [ + { + "bbox": [ + 238, + 378, + 374, + 389 + ], + "spans": [ + { + "bbox": [ + 238, + 378, + 374, + 389 + ], + "type": "text", + "content": "Fig.8: Comparison of " + }, + { + "bbox": [ + 238, + 378, + 374, + 389 + ], + "type": "inline_equation", + "content": "e_1" + }, + { + "bbox": [ + 238, + 378, + 374, + 389 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 238, + 378, + 374, + 389 + ], + "type": "inline_equation", + "content": "e_4" + }, + { + "bbox": [ + 238, + 378, + 374, + 389 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 397, + 301, + 479, + 350 + ], + "blocks": [ + { + "bbox": [ + 397, + 301, + 479, + 350 + ], + "lines": [ + { + "bbox": [ + 397, + 301, + 479, + 350 + ], + "spans": [ + { + "bbox": [ + 397, + 301, + 479, + 350 + ], + "type": "image", + "image_path": "5ca3151bd924b785f75f25a8749074107a2e1ee451d5e3201549dcb5206b6e5c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 397, + 352, + 480, + 369 + ], + "lines": [ + { + "bbox": [ + 397, + 352, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 397, + 352, + 480, + 369 + ], + "type": "text", + "content": "(d) F-Cross-Barcode " + }, + { + "bbox": [ + 397, + 352, + 480, + 369 + ], + "type": "inline_equation", + "content": "(e_4,e_1)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 140, + 418, + 211, + 474 + ], + "blocks": [ + { + "bbox": [ + 140, + 418, + 211, + 474 + ], + "lines": [ + { + "bbox": [ + 140, + 418, + 211, + 474 + ], + "spans": [ + { + "bbox": [ + 140, + 418, + 211, + 474 + ], + "type": "image", + "image_path": "9c86fd33eb424c0151530547f074d98b108dbf934c6682e6d701f2b3bf77648b.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 143, + 480, + 208, + 490 + ], + "lines": [ + { + "bbox": [ + 143, + 480, + 208, + 490 + ], + "spans": [ + { + "bbox": [ + 143, + 480, + 208, + 490 + ], + "type": "text", + "content": "(a) Eigenvector 0." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 228, + 419, + 299, + 474 + ], + "blocks": [ + { + "bbox": [ + 228, + 419, + 299, + 474 + ], + "lines": [ + { + "bbox": [ + 228, + 419, + 299, + 474 + ], + "spans": [ + { + "bbox": [ + 228, + 419, + 299, + 474 + ], + "type": "image", + "image_path": "1f6069fd265f7d9040b62089adab0d6165009850ab21de3dd03577fab58ec3ef.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 229, + 480, + 298, + 490 + ], + "lines": [ + { + "bbox": [ + 229, + 480, + 298, + 490 + ], + "spans": [ + { + "bbox": [ + 229, + 480, + 298, + 490 + ], + "type": "text", + "content": "(b) Eigenvector 22." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 236, + 498, + 376, + 510 + ], + "lines": [ + { + "bbox": [ + 236, + 498, + 376, + 510 + ], + "spans": [ + { + "bbox": [ + 236, + 498, + 376, + 510 + ], + "type": "text", + "content": "Fig. 9: Comparison of " + }, + { + "bbox": [ + 236, + 498, + 376, + 510 + ], + "type": "inline_equation", + "content": "e_0" + }, + { + "bbox": [ + 236, + 498, + 376, + 510 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 236, + 498, + 376, + 510 + ], + "type": "inline_equation", + "content": "e_{22}" + }, + { + "bbox": [ + 236, + 498, + 376, + 510 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 309, + 423, + 391, + 471 + ], + "blocks": [ + { + "bbox": [ + 309, + 423, + 391, + 471 + ], + "lines": [ + { + "bbox": [ + 309, + 423, + 391, + 471 + ], + "spans": [ + { + "bbox": [ + 309, + 423, + 391, + 471 + ], + "type": "image", + "image_path": "d5a2ed65dae205e30160c04caf7d9bec0864528eea015ab1cef4b73b1f312670.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 309, + 473, + 391, + 491 + ], + "lines": [ + { + "bbox": [ + 309, + 473, + 391, + 491 + ], + "spans": [ + { + "bbox": [ + 309, + 473, + 391, + 491 + ], + "type": "text", + "content": "(c) F-Cross-Barcode " + }, + { + "bbox": [ + 309, + 473, + 391, + 491 + ], + "type": "inline_equation", + "content": "(e_0,e_{22})" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 397, + 426, + 479, + 470 + ], + "blocks": [ + { + "bbox": [ + 397, + 426, + 479, + 470 + ], + "lines": [ + { + "bbox": [ + 397, + 426, + 479, + 470 + ], + "spans": [ + { + "bbox": [ + 397, + 426, + 479, + 470 + ], + "type": "image", + "image_path": "08e9a06ef170114a2797ff618e0430ba32f54997672e04d202a631ec38dec65b.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 397, + 472, + 480, + 491 + ], + "lines": [ + { + "bbox": [ + 397, + 472, + 480, + 491 + ], + "spans": [ + { + "bbox": [ + 397, + 472, + 480, + 491 + ], + "type": "text", + "content": "(d) F-Cross-Barcode " + }, + { + "bbox": [ + 397, + 472, + 480, + 491 + ], + "type": "inline_equation", + "content": "(e_{22}, e_0)" + }, + { + "bbox": [ + 397, + 472, + 480, + 491 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "table", + "bbox": [ + 133, + 574, + 482, + 643 + ], + "blocks": [ + { + "bbox": [ + 165, + 535, + 447, + 546 + ], + "lines": [ + { + "bbox": [ + 165, + 535, + 447, + 546 + ], + "spans": [ + { + "bbox": [ + 165, + 535, + 447, + 546 + ], + "type": "text", + "content": "Table 2: Errors of 3D shape reconstruction for \"Cell nuclei\" dataset." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 574, + 482, + 643 + ], + "lines": [ + { + "bbox": [ + 133, + 574, + 482, + 643 + ], + "spans": [ + { + "bbox": [ + 133, + 574, + 482, + 643 + ], + "type": "table", + "html": "
METHOD1-IoUVOLUMESURFACEROUGH.H0H1H2
SHAPR [37]0.6120.4820.2730.1910.0000.0070.038
SHAPR+W.D. [36]0.6120.4370.2510.2000.0000.0080.027
SHAPR+Σi=0,1,2SFTDi0.6180.4320.2440.1870.0000.0080.020
SHAPR+Σi=0,1,2,SFTDi0.5970.4370.2470.1910.0000.0110.011
", + "image_path": "b48c2302ce2a0dd8e88e8fea4e109602a0f69e3cde79064aded06ceda486aedb.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "table_body" + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 294, + 91, + 447, + 102 + ], + "type": "text", + "content": "Scalar Function Topology Divergence" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": "persistence barcodes of ground truth and predicted 3D objects and a total persistence. However, as we discussed before, this loss term doesn't take into account the localization of topological features. We replaced the topological loss term in SHAPR with a symmetrized SFTD with " + }, + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "inline_equation", + "content": "p = 2" + }, + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": " in two variants8: dimensions up to 2 and dimensions up to 3. Then, we carried out experiments with two datasets from [37]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 482, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 482, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 482, + 295 + ], + "type": "text", + "content": "Table 1 shows experimental results for the \"Red blood cell\" dataset. We conclude that SHAPR augmented with the SFTD loss term is the best one in terms of 3D shape matching errors: IoU error, volume error, surface error, and roughness error (see details in [37]). Additionally, we calculated Wasserstein-2 distances between persistence barcodes " + }, + { + "bbox": [ + 130, + 188, + 482, + 295 + ], + "type": "inline_equation", + "content": "(H_0, H_1, H_2)" + }, + { + "bbox": [ + 130, + 188, + 482, + 295 + ], + "type": "text", + "content": " of ground truth and generated 3D objects. Statistically significant improvements are underlined. The proposed SFTD topological loss term with 0-3 dimensions has the lowest Wasserstein-2 distances. This is the rare case of an application of higher homologies like " + }, + { + "bbox": [ + 130, + 188, + 482, + 295 + ], + "type": "inline_equation", + "content": "H_3" + }, + { + "bbox": [ + 130, + 188, + 482, + 295 + ], + "type": "text", + "content": " for real-world applications." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 296, + 482, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 296, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 482, + 331 + ], + "type": "text", + "content": "For the \"Cell nuclei\" dataset results are presented in Table ??. Differences are quite small and are statistically significant only for two metrics: IoU error and W.D. " + }, + { + "bbox": [ + 130, + 296, + 482, + 331 + ], + "type": "inline_equation", + "content": "H_{2}" + }, + { + "bbox": [ + 130, + 296, + 482, + 331 + ], + "type": "text", + "content": ", where SHAPR with SFTD loss is the best model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 332, + 482, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 332, + 482, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 332, + 482, + 416 + ], + "type": "text", + "content": "Figure 10 shows 3D visualizations of " + }, + { + "bbox": [ + 130, + 332, + 482, + 416 + ], + "type": "inline_equation", + "content": "2\\mathrm{D} \\rightarrow 3\\mathrm{D}" + }, + { + "bbox": [ + 130, + 332, + 482, + 416 + ], + "type": "text", + "content": " reconstructions. We see that even with the topological regularization from [36], 3D reconstructions are sometimes erroneous and have serious topological defects: cavities, holes, and surface breaks. In experiments, the sublevel filtration was used and a density function equaled \"1\" inside the object and \"0\" in the ambient space. That is, a cavity inside a sphere corresponds to an extra bar in " + }, + { + "bbox": [ + 130, + 332, + 482, + 416 + ], + "type": "inline_equation", + "content": "H_0" + }, + { + "bbox": [ + 130, + 332, + 482, + 416 + ], + "type": "text", + "content": " barcode, an extra connected component corresponds to an extra bar in " + }, + { + "bbox": [ + 130, + 332, + 482, + 416 + ], + "type": "inline_equation", + "content": "H_2" + }, + { + "bbox": [ + 130, + 332, + 482, + 416 + ], + "type": "text", + "content": " barcode." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 435, + 409, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 435, + 409, + 448 + ], + "spans": [ + { + "bbox": [ + 131, + 435, + 409, + 448 + ], + "type": "text", + "content": "5.4 Identifying topological errors in 3D segmentation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 456, + 482, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 456, + 482, + 563 + ], + "spans": [ + { + "bbox": [ + 130, + 456, + 482, + 563 + ], + "type": "text", + "content": "Brain tumor 3D segmentation aims to reveal the sub-regions of tumor corresponding to the enhancing tumor, the edematous/invaded tissue, the necrosis. Availability of versatile similarity measures not only allows precise evaluation of deep learning 3D segmentation models, but also assists monitoring disease progression in clinical practice. As mentioned above, measuring the similarity of the two 3D segmentations by comparing their persistence barcodes does not account for localization of the topological features. In this section, we illustrate the ability of SFTD to complement the existing metrics with this additional information." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 563, + 482, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 563, + 482, + 624 + ], + "spans": [ + { + "bbox": [ + 130, + 563, + 482, + 624 + ], + "type": "text", + "content": "In order to investigate the applicability of the proposed SFTD as a metric for 3D segmentation, we utilize the publicly available pretrained Swin UNETR model [21,33] and BraTS 21 dataset [2-4,28]. Figures 11a, 11b demonstrate the ground truth segmentation and model's prediction. While each of these two 3D segmentations has two voids inside, only two voids have the same localization" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 237, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 237, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 237, + 101 + ], + "type": "text", + "content": "I. Trofimov et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 631, + 482, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 631, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 631, + 482, + 665 + ], + "type": "text", + "content": "8 A 3D cubical complex has empty " + }, + { + "bbox": [ + 133, + 631, + 482, + 665 + ], + "type": "inline_equation", + "content": "H_{3}" + }, + { + "bbox": [ + 133, + 631, + 482, + 665 + ], + "type": "text", + "content": " homologies. But during the calculation of SFTD for 3D scalar functions, a 4D cubical complex is created, where " + }, + { + "bbox": [ + 133, + 631, + 482, + 665 + ], + "type": "inline_equation", + "content": "H_{3}" + }, + { + "bbox": [ + 133, + 631, + 482, + 665 + ], + "type": "text", + "content": " homologies can exist." + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 175, + 246, + 287 + ], + "blocks": [ + { + "bbox": [ + 133, + 175, + 246, + 287 + ], + "lines": [ + { + "bbox": [ + 133, + 175, + 246, + 287 + ], + "spans": [ + { + "bbox": [ + 133, + 175, + 246, + 287 + ], + "type": "image", + "image_path": "41761ffd262887a54d63fd0727c7558abaddc1694f2bd5bbf16c3ad88a962044.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 158, + 289, + 221, + 298 + ], + "lines": [ + { + "bbox": [ + 158, + 289, + 221, + 298 + ], + "spans": [ + { + "bbox": [ + 158, + 289, + 221, + 298 + ], + "type": "text", + "content": "(a) Ground truth" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 248, + 176, + 362, + 287 + ], + "blocks": [ + { + "bbox": [ + 248, + 176, + 362, + 287 + ], + "lines": [ + { + "bbox": [ + 248, + 176, + 362, + 287 + ], + "spans": [ + { + "bbox": [ + 248, + 176, + 362, + 287 + ], + "type": "image", + "image_path": "697ce54b929563dc668bd033838f372062628e902c0d47afc79e0203bfe85347.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 271, + 289, + 339, + 298 + ], + "lines": [ + { + "bbox": [ + 271, + 289, + 339, + 298 + ], + "spans": [ + { + "bbox": [ + 271, + 289, + 339, + 298 + ], + "type": "text", + "content": "(b) SHAPR+W.D." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 369, + 176, + 480, + 287 + ], + "blocks": [ + { + "bbox": [ + 369, + 176, + 480, + 287 + ], + "lines": [ + { + "bbox": [ + 369, + 176, + 480, + 287 + ], + "spans": [ + { + "bbox": [ + 369, + 176, + 480, + 287 + ], + "type": "image", + "image_path": "cce19c6d5e2159634786337fd7d517e120134d377add3182811a9599fe002561.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 390, + 289, + 459, + 298 + ], + "lines": [ + { + "bbox": [ + 390, + 289, + 459, + 298 + ], + "spans": [ + { + "bbox": [ + 390, + 289, + 459, + 298 + ], + "type": "text", + "content": "(c) SHAPR+SFTD" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 133, + 301, + 246, + 423 + ], + "blocks": [ + { + "bbox": [ + 133, + 301, + 246, + 423 + ], + "lines": [ + { + "bbox": [ + 133, + 301, + 246, + 423 + ], + "spans": [ + { + "bbox": [ + 133, + 301, + 246, + 423 + ], + "type": "image", + "image_path": "aff7a95aed9a7a7441a94a39017c440a152ebf12428b130b65c7778f4451b11f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 158, + 425, + 221, + 434 + ], + "lines": [ + { + "bbox": [ + 158, + 425, + 221, + 434 + ], + "spans": [ + { + "bbox": [ + 158, + 425, + 221, + 434 + ], + "type": "text", + "content": "(d) Ground truth" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 255, + 300, + 366, + 422 + ], + "blocks": [ + { + "bbox": [ + 255, + 300, + 366, + 422 + ], + "lines": [ + { + "bbox": [ + 255, + 300, + 366, + 422 + ], + "spans": [ + { + "bbox": [ + 255, + 300, + 366, + 422 + ], + "type": "image", + "image_path": "76fe38e46fe731ba4507148732ac5b1e86631b131e28d846827e201e569bd4db.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 369, + 301, + 480, + 423 + ], + "blocks": [ + { + "bbox": [ + 369, + 301, + 480, + 423 + ], + "lines": [ + { + "bbox": [ + 369, + 301, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 369, + 301, + 480, + 423 + ], + "type": "image", + "image_path": "1c487d91bb50c6554ad1c098fbfce44f5fdabd08b88dd89d19c00bc553a6c372.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 390, + 425, + 459, + 434 + ], + "lines": [ + { + "bbox": [ + 390, + 425, + 459, + 434 + ], + "spans": [ + { + "bbox": [ + 390, + 425, + 459, + 434 + ], + "type": "text", + "content": "(f) SHAPR+SFTD" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 133, + 435, + 242, + 546 + ], + "blocks": [ + { + "bbox": [ + 133, + 435, + 242, + 546 + ], + "lines": [ + { + "bbox": [ + 133, + 435, + 242, + 546 + ], + "spans": [ + { + "bbox": [ + 133, + 435, + 242, + 546 + ], + "type": "image", + "image_path": "e4ce5524d6dc2fa9a370202961b2f3ff6e3d0a4b3e9c6d83bec94e26f508f045.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 158, + 548, + 221, + 558 + ], + "lines": [ + { + "bbox": [ + 158, + 548, + 221, + 558 + ], + "spans": [ + { + "bbox": [ + 158, + 548, + 221, + 558 + ], + "type": "text", + "content": "(g) Ground truth" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 255, + 435, + 364, + 547 + ], + "blocks": [ + { + "bbox": [ + 277, + 425, + 345, + 434 + ], + "lines": [ + { + "bbox": [ + 277, + 425, + 345, + 434 + ], + "spans": [ + { + "bbox": [ + 277, + 425, + 345, + 434 + ], + "type": "text", + "content": "(e) SHAPR+W.D." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 255, + 435, + 364, + 547 + ], + "lines": [ + { + "bbox": [ + 255, + 435, + 364, + 547 + ], + "spans": [ + { + "bbox": [ + 255, + 435, + 364, + 547 + ], + "type": "image", + "image_path": "812345f3109fad08e623ae47e9aceeeec85d482fcd240a0b55a72cf854d63a71.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 277, + 548, + 345, + 558 + ], + "lines": [ + { + "bbox": [ + 277, + 548, + 345, + 558 + ], + "spans": [ + { + "bbox": [ + 277, + 548, + 345, + 558 + ], + "type": "text", + "content": "(h) SHAPR+W.D." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 369, + 436, + 479, + 546 + ], + "blocks": [ + { + "bbox": [ + 369, + 436, + 479, + 546 + ], + "lines": [ + { + "bbox": [ + 369, + 436, + 479, + 546 + ], + "spans": [ + { + "bbox": [ + 369, + 436, + 479, + 546 + ], + "type": "image", + "image_path": "5f514d9dd00b360eb434c1558fff8e1a5945839b77374236590c63bfe6db2faa.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 390, + 548, + 459, + 558 + ], + "lines": [ + { + "bbox": [ + 390, + 548, + 459, + 558 + ], + "spans": [ + { + "bbox": [ + 390, + 548, + 459, + 558 + ], + "type": "text", + "content": "(i) " + }, + { + "bbox": [ + 390, + 548, + 459, + 558 + ], + "type": "inline_equation", + "content": "\\mathrm{SHAPR} + \\mathrm{SFTD}" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 567, + 482, + 601 + ], + "lines": [ + { + "bbox": [ + 130, + 567, + 482, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 567, + 482, + 601 + ], + "type": "text", + "content": "Fig. 10: Examples of topological errors. Top, " + }, + { + "bbox": [ + 130, + 567, + 482, + 601 + ], + "type": "inline_equation", + "content": "H_0" + }, + { + "bbox": [ + 130, + 567, + 482, + 601 + ], + "type": "text", + "content": ": SHAPR+W.D. has a wrong cavity inside. Middle, " + }, + { + "bbox": [ + 130, + 567, + 482, + 601 + ], + "type": "inline_equation", + "content": "H_1" + }, + { + "bbox": [ + 130, + 567, + 482, + 601 + ], + "type": "text", + "content": ": SHAPR+W.D. has a wrong hole. Bottom, " + }, + { + "bbox": [ + 130, + 567, + 482, + 601 + ], + "type": "inline_equation", + "content": "H_2" + }, + { + "bbox": [ + 130, + 567, + 482, + 601 + ], + "type": "text", + "content": ": SHAPR+W.D. has a wrong surface break." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 294, + 91, + 447, + 102 + ], + "type": "text", + "content": "Scalar Function Topology Divergence" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 119, + 242, + 197 + ], + "blocks": [ + { + "bbox": [ + 136, + 119, + 242, + 197 + ], + "lines": [ + { + "bbox": [ + 136, + 119, + 242, + 197 + ], + "spans": [ + { + "bbox": [ + 136, + 119, + 242, + 197 + ], + "type": "image", + "image_path": "9099adfac9472f400b60913b792f1470085a04b9e4399505e9c1b8820ffcf93b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 158, + 201, + 221, + 211 + ], + "lines": [ + { + "bbox": [ + 158, + 201, + 221, + 211 + ], + "spans": [ + { + "bbox": [ + 158, + 201, + 221, + 211 + ], + "type": "text", + "content": "(a) Ground truth" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 256, + 118, + 365, + 198 + ], + "blocks": [ + { + "bbox": [ + 256, + 118, + 365, + 198 + ], + "lines": [ + { + "bbox": [ + 256, + 118, + 365, + 198 + ], + "spans": [ + { + "bbox": [ + 256, + 118, + 365, + 198 + ], + "type": "image", + "image_path": "0810c284d85d4dfa652b7b3a30b41575108132af301e38f8ac400a741933f96a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 285, + 201, + 337, + 211 + ], + "lines": [ + { + "bbox": [ + 285, + 201, + 337, + 211 + ], + "spans": [ + { + "bbox": [ + 285, + 201, + 337, + 211 + ], + "type": "text", + "content": "(b) Prediction" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 371, + 148, + 479, + 197 + ], + "blocks": [ + { + "bbox": [ + 371, + 148, + 479, + 197 + ], + "lines": [ + { + "bbox": [ + 371, + 148, + 479, + 197 + ], + "spans": [ + { + "bbox": [ + 371, + 148, + 479, + 197 + ], + "type": "image", + "image_path": "8c54e9e47e0f1c8f3dcb4600ae08eb82ad5087dd9f04be4a28e6aa3db568c114.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 370, + 201, + 479, + 211 + ], + "lines": [ + { + "bbox": [ + 370, + 201, + 479, + 211 + ], + "spans": [ + { + "bbox": [ + 370, + 201, + 479, + 211 + ], + "type": "text", + "content": "(c) Pers. barc. of ground truth" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 136, + 213, + 244, + 263 + ], + "blocks": [ + { + "bbox": [ + 136, + 213, + 244, + 263 + ], + "lines": [ + { + "bbox": [ + 136, + 213, + 244, + 263 + ], + "spans": [ + { + "bbox": [ + 136, + 213, + 244, + 263 + ], + "type": "image", + "image_path": "d5d25b88525fb4e0cd318060e0e8ff5a0d3e46c45d42f8fd6e1558b724704717.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 139, + 268, + 240, + 277 + ], + "lines": [ + { + "bbox": [ + 139, + 268, + 240, + 277 + ], + "spans": [ + { + "bbox": [ + 139, + 268, + 240, + 277 + ], + "type": "text", + "content": "(d) Pers. barc. of prediction" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 371, + 213, + 479, + 263 + ], + "blocks": [ + { + "bbox": [ + 371, + 213, + 479, + 263 + ], + "lines": [ + { + "bbox": [ + 371, + 213, + 479, + 263 + ], + "spans": [ + { + "bbox": [ + 371, + 213, + 479, + 263 + ], + "type": "image", + "image_path": "f656680a8bc0f06d1eea9b3ff3bbf2eff033988e28b081d1d6f6000bbfbc0f51.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 388, + 268, + 461, + 277 + ], + "lines": [ + { + "bbox": [ + 388, + 268, + 461, + 277 + ], + "spans": [ + { + "bbox": [ + 388, + 268, + 461, + 277 + ], + "type": "text", + "content": "(e) F-Cross-Barcode" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 286, + 482, + 342 + ], + "lines": [ + { + "bbox": [ + 130, + 286, + 482, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 286, + 482, + 342 + ], + "type": "text", + "content": "Fig. 11: Examples of topological errors. Two 3D segmentations - (a) - ground truth and (b) - prediction - are indistinguishable by their persistence barcodes (c), (d). At the same time, SFTD between ground truth and prediction equals to 1, see (e). A pair of points inside each object in (a), (b) corresponds to the dissimilar topological features, voids, where birth/death events are depicted by red/green dots." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 363, + 482, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 363, + 482, + 435 + ], + "spans": [ + { + "bbox": [ + 130, + 363, + 482, + 435 + ], + "type": "text", + "content": "which makes the two 3D segmentations different. As shown in Figures 11c, 11d, the two 3D segmentations are indistinguishable by their persistence barcodes. In contrary, SFTD is able to recognise the difference between these two cases as revealed by the F-Cross-Barcode in Figure 11e. In particular, SFTD indicates that the difference is precisely in locations of one pair of voids since the set of 1-dimensional topological features contains one segment." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 451, + 226, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 451, + 226, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 451, + 226, + 464 + ], + "type": "text", + "content": "6 Conclusions" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we have proposed new topological tools: F-Cross-Barcode and Scalar Function Topology Divergence (SFTD), which measure the dissimilarity of multi-scale topology between sublevel sets of two functions having a common domain. Functions can be defined on an undirected graph or Euclidean space of any dimensionality. We have shown that our tools outperform the standard approach: evaluating Wasserstein distance between persistence barcodes, calculated independently for sublevel sets of scalar functions. F-Cross-Barcode and SFTD take into account the localization of topological features. SFTD, as a loss function, is beneficial in computer vision, since it helps to enforce topological similarity at multiple scales. Also, the proposed tools provide useful visualizations of areas of topological dissimilarities. In the experimental section, we demonstrated that SFTD improved the reconstruction of cellular 3D shapes from 2D fluorescence microscopy images and identified topological errors in 3D segmentation. Additionally, we showed that SFTD outperforms Betti matching loss in 2D segmentation problems. We believe that the proposed tools will have more applications in computer vision and learning from graph data." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 237, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 237, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 237, + 100 + ], + "type": "text", + "content": "I. Trofimov et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 137, + 489, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 137, + 489, + 162 + ], + "spans": [ + { + "bbox": [ + 132, + 137, + 489, + 162 + ], + "type": "text", + "content": "The work was supported by the Analytical center under the RF Government (subsidy agreement 000000D730321P5Q0002, Grant No. 70-2021-00145 02.11.2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 178, + 198, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 178, + 198, + 190 + ], + "spans": [ + { + "bbox": [ + 133, + 178, + 198, + 190 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 201, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 201, + 480, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 201, + 480, + 233 + ], + "spans": [ + { + "bbox": [ + 138, + 201, + 480, + 233 + ], + "type": "text", + "content": "1. Abousamra, S., Hoai, M., Samaras, D., Chen, C.: Localization in the crowd with topological constraints. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 872-881 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 234, + 480, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 234, + 480, + 255 + ], + "spans": [ + { + "bbox": [ + 138, + 234, + 480, + 255 + ], + "type": "text", + "content": "2. Baid, Ujjwal, e.a.: The rsna-asnr-miccai brats 2021 benchmark on brain tumor segmentation and radiogenomic classification. arXiv preprint arXiv:2107.02314 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 255, + 480, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 255, + 480, + 298 + ], + "spans": [ + { + "bbox": [ + 138, + 255, + 480, + 298 + ], + "type": "text", + "content": "3. Bakas, S., Akbari, H., Sotiras, A., Bilello, M., Rozycki, M., Kirby, J.S., Freymann, J.B., Farahani, K., Davatzikos, C.: Advancing the cancer genome atlas glioma mri collections with expert segmentation labels and radiomic features. Nature Scientific Data (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 299, + 480, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 299, + 480, + 341 + ], + "spans": [ + { + "bbox": [ + 138, + 299, + 480, + 341 + ], + "type": "text", + "content": "4. Bakas, S., Akbari, H., Sotiras, A., Bilello, M., Rozycki, M., Kirby, J.S., Freymann, J.B., Farahani, K., Davatzikos, C.: Segmentation labels and radiomic features for the pre-operative scans of the tcga-gbm collection. The Cancer Imaging Archive (2017)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 342, + 480, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 342, + 480, + 362 + ], + "spans": [ + { + "bbox": [ + 138, + 342, + 480, + 362 + ], + "type": "text", + "content": "5. Barannikov, S.: The framed Morse complex and its invariants. Advances in Soviet Mathematics 21, 93-116 (1994)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 363, + 480, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 363, + 480, + 395 + ], + "spans": [ + { + "bbox": [ + 138, + 363, + 480, + 395 + ], + "type": "text", + "content": "6. Barannikov, S., Korotin, A., Oganesyan, D., Emtsev, D., Burnaev, E.: Barcodes as summary of loss function topology. Doklady Rossijskoj akademii nauk. Matematika, informatika, processy upravlenia 514(2), 196-211 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 396, + 480, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 396, + 480, + 427 + ], + "spans": [ + { + "bbox": [ + 138, + 396, + 480, + 427 + ], + "type": "text", + "content": "7. Barannikov, S., Trofimov, I., Balabin, N., Burnaev, E.: Representation topology divergence: A method for comparing neural network representations. ICML 2022, arXiv:2201.00058 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 428, + 480, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 428, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 138, + 428, + 480, + 460 + ], + "type": "text", + "content": "8. Barannikov, S., Trofimov, I., Sotnikov, G., Trimbach, E., Korotin, A., Filippov, A., Burnaev, E.: Manifold topology divergence: a framework for comparing data manifolds. Advances in Neural Information Processing Systems 34, 7294-7305 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 460, + 480, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 460, + 480, + 492 + ], + "spans": [ + { + "bbox": [ + 138, + 460, + 480, + 492 + ], + "type": "text", + "content": "9. Barannikov, S., Voronkova, D., Trofimov, I., Korotin, A., Sotnikov, G., Burnaev, E.: Topological obstructions in neural networks learning. arXiv preprint arXiv:2012.15834 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 493, + 480, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 493, + 480, + 535 + ], + "spans": [ + { + "bbox": [ + 134, + 493, + 480, + 535 + ], + "type": "text", + "content": "10. Carpenter, A.E., Jones, T.R., Lamprecht, M.R., Clarke, C., Kang, I.H., Friman, O., Guertin, D.A., Chang, J.H., Lindquist, R.A., Moffat, J., et al.: Cellprofiler: image analysis software for identifying and quantifying cell phenotypes. Genome biology 7, 1-11 (2006)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 536, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 536, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 134, + 536, + 480, + 567 + ], + "type": "text", + "content": "11. Carrière, M., Chazal, F., Glisse, M., Ike, Y., Kannan, H., Umeda, Y.: Optimizing persistent homology based functions. In: International Conference on Machine Learning. pp. 1294-1303. PMLR (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 568, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 568, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 134, + 568, + 480, + 610 + ], + "type": "text", + "content": "12. Carrière, M., Chazal, F., Ike, Y., Lacombe, T., Royer, M., Umeda, Y.: Perslay: A neural network layer for persistence diagrams and new graph topological signatures. In: International Conference on Artificial Intelligence and Statistics. pp. 2786-2796. PMLR (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 611, + 480, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 611, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 134, + 611, + 480, + 633 + ], + "type": "text", + "content": "13. Chazal, F., Michel, B.: An introduction to topological data analysis: fundamental and practical aspects for data scientists. arXiv preprint arXiv:1710.04019 (2017)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "text", + "content": "14. Chen, Y., Segovia, I., Gel, Y.R.: Z-gcnets: Time zigzags at graph convolutional networks for time series forecasting. In: International Conference on Machine Learning. pp. 1684-1694. PMLR (2021)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 294, + 91, + 447, + 102 + ], + "type": "text", + "content": "Scalar Function Topology Divergence" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "text", + "content": "15. Chen, Y., Segovia Dominguez, I., Coskunuzer, B., Gel, Y.: Tamp-s2gcnets: When time-aware multipersistence meets spatio-supra graph convolutional nets while forecasting time series. In: The International Conference on Learning Representations (ICLR) (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 162, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 162, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 162, + 481, + 194 + ], + "type": "text", + "content": "16. Demir, A., Coskunuzer, B., Gel, Y., Segovia-Dominguez, I., Chen, Y., Kiziltan, B.: Todd: Topological compound fingerprinting in computer-aided drug discovery. Advances in Neural Information Processing Systems 35, 27978-27993 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 194, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 194, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 133, + 194, + 481, + 237 + ], + "type": "text", + "content": "17. Funke, J., Tschopp, F., Grisaitis, W., Sheridan, A., Singh, C., Saalfeld, S., Turaga, S.C.: Large scale image segmentation with structured loss based deep learning for connectome reconstruction. IEEE transactions on pattern analysis and machine intelligence 41(7), 1669-1680 (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 238, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 238, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 133, + 238, + 481, + 259 + ], + "type": "text", + "content": "18. Gelfand, S.I., Manin, Y.I.: Methods of homological algebra. Springer Science & Business Media (2002)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 259, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 259, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 133, + 259, + 481, + 303 + ], + "type": "text", + "content": "19. Gupta, S., Hu, X., Kaan, J., Jin, M., Mpoy, M., Chung, K., Singh, G., Saltz, M., Kurc, T., Saltz, J., et al.: Learning topological interactions for multi-class medical image segmentation. In: European Conference on Computer Vision. pp. 701-718. Springer (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 304, + 481, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 336 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 336 + ], + "type": "text", + "content": "20. Gupta, S., Zhang, Y., Hu, X., Prasanna, P., Chen, C.: Topology-aware uncertainty for image segmentation. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "text", + "content": "21. Hatamizadeh, A., Nath, V., Tang, Y., Yang, D., Roth, H.R., Xu, D.: Swin unetr: Swin transformers for semantic segmentation of brain tumors in mri images. arXiv preprint arXiv:2201.01266v1 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 369, + 481, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 481, + 392 + ], + "type": "text", + "content": "22. Hofer, C., Graf, F., Rieck, B., Niethammer, M., Kwitt, R.: Graph filtration learning. In: International Conference on Machine Learning. pp. 4314-4323. PMLR (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 392, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 481, + 413 + ], + "type": "text", + "content": "23. Horn, M., De Brouwer, E., Moor, M., Moreau, Y., Rieck, B., Borgwardt, K.: Topological graph neural networks. arXiv preprint arXiv:2102.07835 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 413, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 481, + 435 + ], + "type": "text", + "content": "24. Hu, X., Li, F., Samaras, D., Chen, C.: Topology-preserving deep image segmentation. Advances in neural information processing systems 32 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 435, + 481, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 457 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 457 + ], + "type": "text", + "content": "25. Hu, X., Wang, Y., Fuxin, L., Samaras, D., Chen, C.: Topology-aware segmentation using discrete Morse theory. arXiv preprint arXiv:2103.09992 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 457, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 481, + 479 + ], + "type": "text", + "content": "26. Kaczynski, T., Mischaikow, K.M., Mrozek, M.: Computational homology, vol. 3. Springer (2004)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 479, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 501 + ], + "type": "text", + "content": "27. Luo, Y., Shi, L., Thost, V.: Improving self-supervised molecular representation learning using persistent homology. arXiv preprint arXiv:2311.17327 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 501, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 481, + 533 + ], + "type": "text", + "content": "28. Menze, B.H., Jakab, A., Bauer, S., Kalpathy-Cramer, J., Farahani, K., Kirby, Justin, e.a.: The multimodal brain tumor image segmentation benchmark (brats). IEEE Transactions on Medical Imaging (2015)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 533, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 481, + 555 + ], + "type": "text", + "content": "29. Moor, M., Horn, M., Rieck, B., Borgwardt, K.: Topological autoencoders. In: International conference on machine learning. pp. 7045-7054. PMLR (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 555, + 481, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 481, + 588 + ], + "type": "text", + "content": "30. Pérez, J.B., Hauke, S., Lupo, U., Caorsi, M., Dassatti, A.: giotto-ph: A python library for high-performance computation of persistent homology of vietoris-rips filtrations (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 589, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 481, + 632 + ], + "type": "text", + "content": "31. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "32. Stucki, N., Paetzold, J.C., Shit, S., Menze, B., Bauer, U.: Topologically faithful image segmentation via induced matching of persistence barcodes. In: International Conference on Machine Learning. pp. 32698-32727. PMLR (2023)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 236, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 236, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 236, + 100 + ], + "type": "text", + "content": "I. Trofimov et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 381 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "text", + "content": "33. Tang, Y., Yang, D., Li, W., Roth, H.R., Landman, B., Xu, D., Nath, V., Hatamizadeh, A.: Self-supervised pre-training of swim transformers for 3d medical image analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 161, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 161, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 161, + 481, + 194 + ], + "type": "text", + "content": "34. Trofimov, I., Cherniavskii, D., Tulchinskii, E., Balabin, N., Burnaev, E., Barannikov, S.: Learning topology-preserving data representations. arXiv preprint arXiv:2302.00136 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "text", + "content": "35. Wagner, H., Chen, C., Vuçini, E.: Efficient computation of persistent homology for cubical data. In: Topological methods in data analysis and visualization II: theory, algorithms, and applications, pp. 91-106. Springer (2011)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 270 + ], + "type": "text", + "content": "36. Waibel, D.J., Atwell, S., Meier, M., Marr, C., Rieck, B.: Capturing shape information with multi-scale topological loss terms for 3d reconstruction. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 150-159. Springer (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 271, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 481, + 293 + ], + "type": "text", + "content": "37. Waibel, D.J., Kiermeyer, N., Atwell, S., Sadafi, A., Meier, M., Marr, C.: Shapr predicts 3d cell shapes from 2d microscopic images. Iscience 25(11) (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 314 + ], + "type": "text", + "content": "38. Watts, D.J., Strogatz, S.H.: Collective dynamics of 'small-world'networks. nature 393(6684), 440-442 (1998)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "type": "text", + "content": "39. Wu, J., Wang, Y., Xue, T., Sun, X., Freeman, B., Tenenbaum, J.: Marrnet: 3d shape reconstruction via 2.5 d sketches. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 348, + 481, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 348, + 481, + 381 + ], + "spans": [ + { + "bbox": [ + 132, + 348, + 481, + 381 + ], + "type": "text", + "content": "40. Yan, Z., Ma, T., Gao, L., Tang, Z., Chen, C.: Link prediction with persistent homology: An interactive view. In: International conference on machine learning. pp. 11659-11669. PMLR (2021)" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 294, + 91, + 447, + 102 + ], + "type": "text", + "content": "Scalar Function Topology Divergence" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/ScaleDreamer_ Scalable Text-to-3D Synthesis with Asynchronous Score Distillation/0ae3774a-e55b-4154-9080-499a583a8778_content_list.json b/2024/ScaleDreamer_ Scalable Text-to-3D Synthesis with Asynchronous Score Distillation/0ae3774a-e55b-4154-9080-499a583a8778_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..585642bf7e1479baac3c13e1011ae3b5898ebac6 --- /dev/null +++ b/2024/ScaleDreamer_ Scalable Text-to-3D Synthesis with Asynchronous Score Distillation/0ae3774a-e55b-4154-9080-499a583a8778_content_list.json @@ -0,0 +1,1922 @@ +[ + { + "type": "text", + "text": "ScaleDreamer: Scalable Text-to-3D Synthesis with Asynchronous Score Distillation", + "text_level": 1, + "bbox": [ + 243, + 140, + 759, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhiyuan Ma $^{1,2}$ , Yuxiang Wei $^{1,5}$ , Yabin Zhang $^{1}$ ,", + "bbox": [ + 310, + 210, + 687, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiangyu Zhu $^{3,4}$ , Zhen Lei $^{1,2,3,4^{\\star}}$ , and Lei Zhang $^{1^{\\star}}$", + "bbox": [ + 303, + 231, + 697, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 The Hong Kong Polytechnic University, PolyU", + "bbox": [ + 336, + 261, + 663, + 275 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Center for Artificial Intelligence and Robotics, HKISI CAS", + "bbox": [ + 297, + 275, + 704, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 State Key Laboratory of Multimodal Artificial Intelligence Systems, CASIA", + "bbox": [ + 243, + 289, + 759, + 303 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ School of Artificial Intelligence, University of Chinese Academy of Sciences, UCAS", + "bbox": [ + 220, + 303, + 781, + 316 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "5 Harbin Institute of Technology, HIT", + "bbox": [ + 374, + 318, + 627, + 330 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. By leveraging the text-to-image diffusion prior, score distillation can synthesize 3D contents without paired text-3D training data. Instead of spending hours of online optimization per text prompt, recent studies have been focused on learning a text-to-3D generative network for amortizing multiple text-3D relations, which can synthesize 3D contents in seconds. However, existing score distillation methods are hard to scale up to a large amount of text prompts due to the difficulties in aligning pretrained diffusion prior with the distribution of rendered images from various text prompts. Current state-of-the-arts such as Variational Score Distillation finetune the pretrained diffusion model to minimize the noise prediction error so as to align the distributions, which are however unstable to train and will impair the model's comprehension capability to numerous text prompts. Based on the observation that the diffusion models tend to have lower noise prediction errors at earlier timesteps, we propose Asynchronous Score Distillation (ASD), which minimizes the noise prediction error by shifting the diffusion timestep to earlier ones. ASD is stable to train and can scale up to $100k$ prompts. It reduces the noise prediction error without changing the weights of pre-trained diffusion model, thus keeping its strong comprehension capability to prompts. We conduct extensive experiments using different text-to-3D architectures, including Hyper-iNGP and 3DConvNet. The results demonstrate ASD's effectiveness in stable 3D generator training, high-quality 3D content synthesis, and its superior prompt-consistency, especially under large prompt corpus. Code is available at https://github.com/theEricMa/ScaleDreamer.", + "bbox": [ + 259, + 361, + 738, + 708 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Text-to-3D $\\cdot$ Score Distillation $\\cdot$ Diffusion Model", + "bbox": [ + 261, + 720, + 673, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 757, + 374, + 772 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-to-3D aims to generate realistic 3D contents from the given textual descriptions [40], which is particularly useful in many applications such as virtual", + "bbox": [ + 212, + 786, + 785, + 818 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding authors.", + "bbox": [ + 217, + 824, + 392, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "reality [61] and game design [23]. The main challenge of this task, however, lies in how to generate high-quality 3D contents conditioned on the abstract and diverse textual descriptions. Many existing text-to-3D methods [11, 12, 26–29, 32, 35, 37, 40, 42, 58, 59, 75] are optimization-based ones, which distill the guidance from the powerful pretrained text-to-image diffusion models [5, 11, 26, 32, 42, 44, 73] via score distillation [40, 59, 62, 72]. In general, these methods employ the KL divergence to reduce the discrepancy between the distribution of rendered images and the desired image distribution embedded in the 2D diffusion prior, while they differ in how to use the pretrained diffusion prior to model the distribution of rendered images. Extensive efforts have been made to explore prompt-specific optimization of various 3D representations, including implicit radiance fields [40], explicit radiance fields [29, 37, 59], DmTets [56, 74] and 3D Gaussians [9]. Typically, tens of minutes to hours are needed to optimize a single 3D representation for one prompt to achieve the desired result.", + "bbox": [ + 212, + 146, + 787, + 357 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Compared to the aforementioned optimization-based text-to-3D methods, learning-based methods [6, 21, 31, 36, 43, 54, 65] can largely reduce the computational cost by training a text-conditioned 3D generative network. With the availability of 3D object collections [10, 63, 71], a deep network can be trained in a supervised manner so that 3D outputs can be generated in several seconds. Unfortunately, the size of existing text-3D datasets is far from sufficient compared to text-image datasets [45], limiting the text-to-3D generation performance of trained models. Inspired by the optimization-based text-to-3D methods that use pretrained 2D diffusion models, efforts have been made to train text-to-3D networks by using 2D diffusion models as supervisors [33, 41, 65] without using text-3D pairs. For example, a text-conditioned 3D hyper-network is trained in ATT3D [33] via Score Distillation Sampling (SDS) [40]. Nevertheless, this method suffers from numerical instability, which has been observed in subsequent studies [41, 65] that apply SDS to different 3D generator networks.", + "bbox": [ + 212, + 362, + 787, + 574 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite the success of score distillation in optimization-based text-to-3D generation [40, 59, 72], its application to learning-based text-to-3D frameworks is rather limited because of the unstable training or unsatisfactory results. We argue that the primary challenge lies in how to efficiently and effectively leverage the pretrained 2D diffusion prior to represent the distribution of images rendered by the 3D generator. For example, SDS [40] forces the rendered images to adhere to the Dirac distribution, which causes numerical instability in 3D generator training [33, 65]. Variational Score Distillation (VSD) [59] finetunes the 2D diffusion prior for distribution alignment via minimizing the noise prediction error. However, the finetuning changes the pretrained diffusion network and hurts its comprehension capability to numerous text prompts, leading to mode collapse when the size of text prompts is extended.", + "bbox": [ + 212, + 579, + 787, + 760 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the above mentioned issues, we propose Asynchronous Score Distillation (ASD). Like VSD, ASD aims to minimize the noise prediction error. Different from VSD, ASD does not finetune the pretrained 2D diffusion network; instead, it achieves the goal by shifting the diffusion timestep. This is based on the observation that diffusion networks will have smaller noise prediction errors", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Z.Ma et al.", + "bbox": [ + 271, + 114, + 346, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "in earlier timesteps [67]; therefore, we can shift the timestep to an earlier step to achieve a similar goal to VSD, i.e., reducing the noise prediction error. In this way, the diffusion network can be frozen in training and its strong text comprehension capability can be well-preserved. The shifted timesteps can be well sampled from a pre-defined range for most prompts. To evaluate the performance of ASD, we conduct extensive experiments by using two types of generator architectures, i.e. Hyper-iNGP [33] and 3DConv-Net [4], across various prompt corpus sizes. We conduct extensive experiments to evaluate the superiority of ASD to previous methods, including the stable training of 3D generators, the production of high-quality 3D outputs, the high content fidelity to input prompts, as well as its scalability to larger corpus sizes, e.g., 100k prompts. We also show that ASD can work with other 2D diffusion models such as MVDream [48], and can be used to train more 3D generators such as Triplane-Transformer [17].", + "bbox": [ + 212, + 146, + 787, + 344 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Literature Review", + "text_level": 1, + "bbox": [ + 215, + 368, + 429, + 383 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Text-to-3D with Score Distillation", + "text_level": 1, + "bbox": [ + 215, + 401, + 542, + 416 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text-to-3D takes text description, a.k.a. text prompt $y$ , as input, and outputs 3D representation $\\theta$ that renders high-fidelity images at any camera view $\\pi$ . Thanks to the powerful text-to-image diffusion models [32,42,44,48,73], we can optimize $\\theta$ to align with $y$ by computing the objective $\\mathcal{L}(\\boldsymbol{x},y)$ on the rendered image $\\boldsymbol{x} = g(\\theta ,\\pi)$ from camera view $\\pi$ . Through differential rendering, $\\theta$ can be updated with the gradient $\\nabla_{\\theta}\\mathcal{L}(\\theta ,y) = \\frac{\\partial\\mathcal{L}(\\boldsymbol{x},y)}{\\partial\\boldsymbol{x}}\\frac{\\partial\\boldsymbol{x}}{\\partial\\theta}$ . This technique is generally termed as score distillation. Unlike data-driven techniques [6,21,31,43,54], score distillation approaches [8, 19, 24, 29, 40, 53, 59, 72] can produce high-quality 3D content without the need for 3D training datasets.", + "bbox": [ + 212, + 429, + 787, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Prompt-Specific Text-to-3D. Existing score distillation methods [40, 59, 72] were originally developed to output a single 3D result $\\theta$ for a single text prompt $y$ via online optimization: $\\min_{\\theta} \\mathbb{E}_{\\pi, \\boldsymbol{x} = g(\\theta, \\pi)}[\\mathcal{L}(\\boldsymbol{x}, y)]$ . The utilized 3D representations, e.g., NeRF [39, 40], DmTet [47, 74], and 3D Gaussian [20, 30, 51, 53, 57, 70], are not designed to render scenes from varying text prompts. Therefore, the optimization has to be conducted again for newly provided text prompts. The optimization process typically costs tens of minutes to hours.", + "bbox": [ + 212, + 566, + 787, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Prompt-Amortized Text-to-3D. To mitigate the computational costs in prompt-specific methods, recent studies [25, 33, 41, 65] have attempted to use score distillation to train a text-to-3D generator $\\theta = \\mathcal{G}(y)$ , aiming to generate multiple 3D representations from a set of text prompts $S_{y} = \\{y\\}$ . These methods can generate 3D results from queried text prompt in seconds. As proposed by ATT3D [33], the 3D generator training is performed by minimizing $\\min_{\\mathcal{G}} \\mathbb{E}_{\\pi, y \\in S_{y}, x = g(\\mathcal{G}(y), \\pi)}[\\mathcal{L}(\\boldsymbol{x}, y)]$ over all text prompts. Unlike data-driven approaches [17, 52, 66], score distillation bypasses the scarcity of text-3D data pairs because the 2D diffusion prior can offer the guidance to align the 3D output with the input text prompt. However, its application is currently restricted to training the 3D generator within a limited range of text prompts.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "ScaleDreamer", + "bbox": [ + 637, + 114, + 730, + 127 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 Representative Score Distillation Methods", + "text_level": 1, + "bbox": [ + 215, + 146, + 609, + 161 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Denote by $\\phi$ the 2D diffusion prior [44,48] and by $p^{\\phi}(\\boldsymbol{x} \\mid y)$ the text-conditioned image distribution embedded within $\\phi$ , the objectives of most existing score distillation methods can be generally concluded as minimizing the objective $\\mathcal{L}(\\theta, y) = \\mathbb{E}_{\\pi, t, \\epsilon, \\boldsymbol{x} = g(\\theta, \\pi)}\\left[\\omega(t) D_{\\mathrm{KL}}\\left(q_t^\\theta(\\boldsymbol{x}_t \\mid \\pi) \\| p_t^\\phi(\\boldsymbol{x}_t \\mid y^\\pi)\\right)\\right]$ , where $D_{\\mathrm{KL}}$ denotes KL divergence, $q_t^\\theta(\\boldsymbol{x}_t \\mid \\pi)$ denotes the distribution of images $\\boldsymbol{x}$ rendered at camera view $\\pi$ at diffusion timestep $t$ [15], and the same for $p_t^\\phi(\\boldsymbol{x}_t \\mid y)$ . $\\omega(t)$ is a timestep-dependent weight [40]. $y^\\pi$ denotes the view-dependent strategy [44] or view-awareness [42, 48] to prompt the different camera views [40]. To minimize this objective, the gradient w.r.t. $\\theta$ can be calculated as per [59]:", + "bbox": [ + 212, + 172, + 787, + 319 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\theta} \\mathcal {L} (\\theta , y) = \\mathbb {E} _ {\\pi , t, \\epsilon} \\left[ \\omega (t) \\left(\\underbrace {- \\sigma_ {t} \\nabla_ {\\boldsymbol {x} _ {t}} \\log p _ {t} ^ {\\phi} (\\boldsymbol {x} _ {t} \\mid y ^ {\\pi})} _ {\\epsilon_ {\\phi} (\\boldsymbol {x} _ {t}; t, y ^ {\\pi})} - \\underbrace {(- \\sigma_ {t} \\nabla_ {\\boldsymbol {x} _ {t}} \\log q _ {t} ^ {\\theta} (\\boldsymbol {x} _ {t} \\mid \\pi))} _ {\\epsilon_ {\\theta} (\\boldsymbol {x} _ {t}; t, \\pi , y)}\\right) \\frac {\\partial \\boldsymbol {x}}{\\partial \\theta} \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 227, + 329, + 785, + 388 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the first term $-\\sigma_t\\nabla_{\\boldsymbol{x}_t}\\log p_t^\\phi (\\boldsymbol{x}_t\\mid y^\\pi)$ corresponds to the score function [50] of the desired image distribution, and it can be achieved by predicting the noise $\\epsilon \\sim \\mathcal{N}(0,I)$ in the noisy image $\\boldsymbol{x}_t = \\alpha_t\\boldsymbol {x} + \\sigma_t\\boldsymbol{\\epsilon}$ using the pretrained 2D diffusion model $\\epsilon_{\\phi}(\\pmb {x}_t;t,y^\\pi)$ [44,48]. Existing score distillation methods [40,59,72] mainly differ in how to model $-\\sigma_t\\nabla x_t\\log q_t^\\theta (x_t\\mid \\pi)$ , which corresponds to the score function of the distribution of rendered images $q^{\\theta}(\\pmb {x}\\mid \\pi)$ . We denote this term in Eq. 1 as $\\epsilon_{\\theta}(\\pmb {x}_t;t,\\pi ,y)$ in the following context, since it represents a diffusion model that corresponds to $\\theta$ . A summary of the objectives of major score distillation methods is shown in Tab. 1.", + "bbox": [ + 212, + 400, + 787, + 536 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The objective of Score Distillation Sampling (SDS) [40] is $\\nabla_{\\theta}\\mathcal{L}_{\\mathrm{SDS}}(\\theta ,y)\\triangleq$ $\\mathbb{E}_{\\pi ,t,\\epsilon}\\left[\\omega (t)(\\pmb{\\epsilon}_{\\phi}(\\pmb{x}_t;t,y^\\pi) - \\pmb {\\epsilon})\\frac{\\partial\\pmb{x}}{\\partial\\theta}\\right]$ , which approximates the term $\\pmb{\\epsilon}_{\\theta}(\\pmb {x}_t;t,\\pi ,y)$ in Eq. 1 as the ground-truth noise $\\pmb{\\epsilon}$ . That is, SDS assumes that $q^{\\theta}(\\pmb {x}\\mid \\pi)$ adheres to a Dirac distribution $\\delta (\\pmb {x} - g(\\theta ,\\pi))$ [59], which is characterized by a non-zero density at the singular point of $\\pmb {x} = g(\\theta ,\\pi)$ and zero density everywhere else. However, updating $\\theta$ under the Dirac distribution might be troublesome [59]. We may need to set the CFG (Classifier Free Guidance) [16] as high as 100 for model convergence, which will produce excessively large gradients and lead to unstable optimization. This problem is alleviated by Classifier Score Distillation (CSD) [72], which uses the classifier component [16] in SDS as the objective: $\\nabla_{\\theta}\\mathcal{L}_{\\mathrm{CSD}}(\\theta ,y)\\triangleq \\mathbb{E}_{\\pi ,t,\\epsilon}\\left[\\omega (t)(\\pmb{\\epsilon}_{\\phi}(\\pmb {x}_t;t,y^\\pi) - \\pmb{\\epsilon}_{\\phi}(\\pmb {x}_t;t))\\frac{\\partial\\pmb{x}}{\\partial\\theta}\\right]$ . CSD can be regraded as straightforwardly using the unconditional term of the diffusion prior $\\pmb{\\epsilon}_{\\phi}(\\pmb{x}_t;t)$ to represent $\\pmb{\\epsilon}_{\\theta}(\\pmb {x}_t;t,\\pi ,y)$ in Eq. 1. Unfortunately, in the case of prompt-amortized training, this term may not provide effective gradient, because $\\pmb{\\epsilon}_{\\phi}(\\pmb {x}_t;t)$ is unconditional to the provided text-prompts. In contrast, Variational Score Distillation (VSD) [59] models $\\pmb{\\epsilon}_{\\theta}(\\pmb {x}_t;t,\\pi ,y)$ with another text-aware diffusion model $\\pmb{\\epsilon}_{\\phi^{\\prime}}(\\mathbf{x}_t;t,\\pi ,y)$ , leading to $\\nabla_{\\theta}\\mathcal{L}_{\\mathrm{VSD}}(\\theta ,y)\\triangleq$ $\\mathbb{E}_{\\pi ,t,\\epsilon}\\left[\\omega (t)(\\pmb{\\epsilon}_{\\phi}(\\pmb {x}_t;t,y^\\pi) - \\pmb{\\epsilon}_{\\phi^{\\prime}}(\\pmb {x}_t;t,\\pi ,y))\\frac{\\partial\\pmb{x}}{\\partial\\theta}\\right]$ , where $\\pmb{\\epsilon}_{\\phi^{\\prime}}(\\mathbf{x}_t;t,\\pi ,y)$ is achieved by finetuning the pretrained 2D diffusion prior $\\pmb{\\epsilon}_{\\phi}(\\pmb {x}_t;t,y^\\pi)$ to align with the rendered image distribution $q^{\\theta}(\\pmb {x}\\mid \\pi)$ via parameter efficient adaptation [18]. In", + "bbox": [ + 212, + 537, + 787, + 843 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Z.Ma et al.", + "bbox": [ + 271, + 114, + 346, + 127 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/0ab3981ab01452839abaf8059133e5ad95175a690bbbd5686bc991283b5f04d0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodGradient of L(x,y) w.r.t. x = g(θ,π)
SDS [40]E_t,ε[ω(t) (εφ(xt; t, yπ) - ε)]
CSD [72]E_t,ε[ω(t) (εφ(xt; t, yπ) - εφ(xt, t))]
VSD [59]E_t,ε[ω(t) (εφ(xt; t, yπ) - εφ'(xt; t, π, y))]
ASD (Ours)E_t,ε[ω(t) (εφ(xt; t, yπ) - εφ(xt+Δt; t+Δt, yπ))]
", + "bbox": [ + 285, + 143, + 718, + 218 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1: Objectives of representative score distillation methods. ASD introduces $\\Delta t$ alongside $t$ to align with the rendered image distribution $q^{\\theta}(\\boldsymbol{x} \\mid \\pi)$ .", + "bbox": [ + 212, + 219, + 782, + 246 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "practice, this is conducted by alternatively optimizing $\\theta$ and finetuning $\\phi$ with the noise prediction objective $\\| \\pmb{\\epsilon}_{\\phi}(\\pmb{x}_t;t,y) - \\pmb{\\epsilon}\\| _2^2$ [15] such that:", + "bbox": [ + 212, + 253, + 782, + 286 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} _ {\\pi , t, \\epsilon} \\left[ \\| \\epsilon_ {\\phi J} \\left(\\boldsymbol {x} _ {t}; t, \\pi , y\\right) - \\epsilon \\| _ {2} ^ {2} \\right] \\leq \\mathbb {E} _ {\\pi , t, \\epsilon} \\left[ \\| \\epsilon_ {\\phi} \\left(\\boldsymbol {x} _ {t}; t, y ^ {\\pi}\\right) - \\epsilon \\| _ {2} ^ {2} \\right]. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 287, + 295, + 784, + 313 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The above equation reveals that a better alignment with the distribution of $q^{\\theta}(\\pmb{x} \\mid \\pi)$ can be achieved by a more accurate noise prediction.", + "bbox": [ + 212, + 323, + 782, + 353 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While VSD achieves state-of-the-art results in prompt-specific text-to-3D [14, 59], it changes the diffusion prior's parameters by alternately optimizing $\\theta$ and finetuning $\\phi$ . This forms a bi-level optimization, known to be problematic in generative adversarial training [55], and may be troublesome for training prompt-amortized text-to-3D models, because the change of pre-trained diffusion model might impairs its comprehension capability on a wide range of text-prompts. In specific, the pre-trained 2D diffusion model may have to sacrifice its generation capability in order to align with the distribution of rendered images, making it fail to produce good gradient for training the 3D generator.", + "bbox": [ + 212, + 354, + 784, + 489 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Asynchronous Score Distillation (ASD)", + "text_level": 1, + "bbox": [ + 215, + 512, + 630, + 530 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Objective of ASD", + "text_level": 1, + "bbox": [ + 215, + 542, + 408, + 558 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "From the above discussions in Sec. 2.2, it can be seen that one key issue in VSD is to minimize the noise prediction error so that the model output can be aligned with the desired distribution of rendered images. VSD achieves this goal via finetuning the pre-trained 2D diffusion model, which however sacrifices its comprehension capability on text prompts. One interesting question is: can we minimize the noise prediction error without changing the pre-trained diffusion network weights? Fortunately, we find that this is possible and in this section we present a new objective function to achieve this goal.", + "bbox": [ + 212, + 568, + 784, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Recall that diffusion models solve the stochastic differential equation [50] via reversing the noise added along different stages, a.k.a. diffusion timestep $t \\in \\{T_{\\max}, \\dots, T_{\\min}\\}$ via $\\boldsymbol{x}_t = \\alpha_t \\boldsymbol{x} + \\sigma_t \\boldsymbol{\\epsilon}$ [15]. The influence of the noise $\\epsilon \\sim \\mathcal{N}(0, I)$ on the image $\\boldsymbol{x}$ is incrementally reduced as the process progresses from the initial timestep $T_{\\max}$ to the final timestep $T_{\\min}$ , which is controlled by the scalars $\\alpha_t$ and $\\sigma_t$ . Consequently, the diffusion model's noise prediction accuracy will vary with the timestep $t$ , at which the identical noise $\\epsilon$ is added. To evaluate this, we consider a diffusion model with fixed image $\\boldsymbol{x}$ , noise $\\epsilon$ and condition $y$ , but varied timestep $t$ . We denote such a diffusion model as $\\epsilon(t)$ and explore how its prediction error, denoted by $e(t) = \\| \\epsilon(t) - \\epsilon \\|_2^2$ , changes with $t$ .", + "bbox": [ + 212, + 688, + 785, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "ScaleDreamer", + "bbox": [ + 637, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The model $\\epsilon(t)$ can be a pre-trained 2D diffusion model (such as Stable Diffusion [44]). We denote by $\\epsilon_{PT}(t)$ such a model, and investigate the behaviour of its noise prediction error, denoted by $e_{PT}(t)$ . In Fig. 1, we plot the curve (i.e., the blue colored curve) of $e_{PT}(t)$ versus $t$ . We use a corpus with 15 text prompts from Magic3D [40] to draw this curve. For each prompt $y$ , we generate 16 images with VSD [59]. Then for each image $\\pmb{x}$ , we apply one instance of Gaussian noise $\\pmb{\\epsilon}$ and conduct a single diffusion step with 100 distinct timesteps. The average noise reconstruction error is then calculated for these timesteps across all prompts and images. We can see from the curve of $e_{PT}(t)$ that earlier diffusion timesteps (e.g., timestep 600) will have lower noise prediction error than later timesteps (e.g., timestep 200). Such a trend holds for almost every image sample $\\pmb{x}$ and noise sample $\\pmb{\\epsilon}$ because the well-trained diffusion model is frozen in our case. Since the noise prediction error declines from $T_{\\mathrm{min}}$ (i.e., late diffusion timestep) to $T_{\\mathrm{max}}$ (i.e., early diffusion timestep), we can conclude that for a given timestep $t$ and a timestep shift $0 \\leq \\Delta t \\leq T_{\\mathrm{max}} - t$ , the following inequality holds:", + "bbox": [ + 212, + 146, + 787, + 375 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} _ {\\pi , t, \\epsilon} \\left[ \\| \\epsilon_ {\\phi} \\left(\\boldsymbol {x} _ {t + \\Delta t}; t + \\Delta t, y ^ {\\pi}\\right) - \\epsilon \\| _ {2} ^ {2} \\right] \\leq \\mathbb {E} _ {\\pi , t, \\epsilon} \\left[ \\| \\epsilon_ {\\phi} \\left(\\boldsymbol {x} _ {t}; t, y ^ {\\pi}\\right) - \\epsilon \\| _ {2} ^ {2} \\right], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 261, + 392, + 785, + 411 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "which implies that more accurate noise predictions can be achieved at earlier diffusion timesteps.", + "bbox": [ + 212, + 429, + 785, + 460 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The above property of diffusion models has also been observed by Yang et al. [68], who indicated that as the timestep shifts from $T_{\\mathrm{max}}$ towards $T_{\\mathrm{min}}$ , the variance in noise prediction increases, as evidenced by the rising Lipschitz constants, which suggests an increased instability in noise prediction and larger noise prediction errors. Such a behavior can be observed in both $\\epsilon$ -prediction and $\\pmb{v}$ -prediction models, as well as in 2D and 3D diffusion models (please refer to supplementary material for details). This can be intuitively explained as follows. When $t \\to T_{\\mathrm{max}}$ , $\\pmb{x}_t = \\alpha_t \\pmb{x} + \\sigma_t \\pmb{\\epsilon} \\to \\pmb{\\epsilon}$ , then it is easier to achieve $\\epsilon_{\\phi}(\\pmb{x}_t; t, y^{\\pi}) \\approx \\epsilon$ because the model can manage to copy the input as the output.", + "bbox": [ + 212, + 463, + 787, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The similarity between Eq. 3 and the fine-tuning objective of VSD in Eq. 2 inspires us to investigate whether simply shifting earlier the timestep could fulfill the fine-tuning requirements of VSD without modifying the pre-trained 2D diffusion network parameters. Specifically, we employ the pretrained 2D diffusion model with shifted timestep to approximate the diffusion model of rendered images in Eq. 1 as $\\pmb{\\epsilon}_{\\theta}(\\pmb{x}_t; t, \\pi, y) \\triangleq \\pmb{\\epsilon}_{\\phi}(\\pmb{x}_{t + \\Delta t}; t + \\Delta t, y^{\\pi})$ , resulting in the following Asynchronous Score Distillation (ASD) objective function:", + "bbox": [ + 212, + 603, + 787, + 709 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\theta} \\mathcal {L} _ {\\mathrm {A S D}} (\\theta , y) \\triangleq \\mathbb {E} _ {\\pi , t, \\epsilon} \\left[ \\omega (t) \\left(\\boldsymbol {\\epsilon} _ {\\phi} \\left(\\boldsymbol {x} _ {t}; t, y ^ {\\pi}\\right) - \\boldsymbol {\\epsilon} _ {\\phi} \\left(\\boldsymbol {x} _ {t + \\Delta t}; t + \\Delta t, y ^ {\\pi}\\right)\\right) \\frac {\\partial \\boldsymbol {x}}{\\partial \\theta} \\right]. \\quad (4)\n$$\n", + "text_format": "latex", + "bbox": [ + 233, + 728, + 785, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We can see that rather than iteratively fine-tuning the diffusion network as in VSD, ASD achieves similar goal by shifting the timestep $t$ with an interval $\\Delta t$ in each step, which is much more efficient. One key variable introduced in ASD is the timestep shift $\\Delta t$ , which will be discussed in the next subsection.", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Z.Ma et al.", + "bbox": [ + 271, + 114, + 346, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/42b5dd41704546434be15a59e10acbfba878ec1590bc5e05f91031f968d6966f.jpg", + "image_caption": [ + "Fig. 1: Illustration of the noise prediction error of the pre-trained 2D diffusion model $\\epsilon_{PT}(t)$ and that of the fine-tuned 2D diffusion model $\\epsilon_{FT}(t)$ . We can see that the curve of $e_{FT}(t)$ is positioned under that of $e_{PT}(t)$ , and we can shift the timestep of $\\epsilon_{PT}(t)$ to $\\epsilon_{PT}(t + \\Delta t)$ to approximate the noise prediction error of $\\epsilon_{FT}(t)$ ." + ], + "image_footnote": [], + "bbox": [ + 277, + 146, + 720, + 351 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 The Setting of Timestep Shift $\\Delta t$", + "text_level": 1, + "bbox": [ + 215, + 417, + 537, + 431 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Before discussing how to set the timestep shift $\\Delta t$ , let's plot another curve, i.e., the noise prediction error of $\\epsilon_{\\theta}(\\boldsymbol{x}_t; t, \\pi, y)$ w.r.t. timestep $t$ . Actually, in the process of generating $\\boldsymbol{x}$ with VSD, we will have the fine-tuned model $\\epsilon_{\\phi'}(\\boldsymbol{x}_t; t, \\pi, y)$ as the by-product, which is used to represent $\\epsilon_{\\theta}(\\boldsymbol{x}_t; t, \\pi, y)$ in Eq. 1. Therefore, with fixed $\\boldsymbol{x}$ , $\\epsilon$ and $y$ , the noise prediction error of the fine-tuned diffusion model, denoted by $\\epsilon_{FT}(t)$ , can be calculated as $e_{FT}(t) = \\| \\epsilon_{\\phi'}(t) - \\epsilon \\|_2^2$ .", + "bbox": [ + 212, + 446, + 787, + 537 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The curve of $e_{FT}(t)$ w.r.t. $t$ (i.e., the yellow curve) is plotted in Fig. 1 by using the same data as in plotting $e_{PT}(t)$ . We can see that the curve of $e_{FT}(t)$ is positioned under $e_{PT}(t)$ because $e_{FT}(t)$ is obtained by the fine-tuned diffusion model $\\epsilon_{FT}$ . However, as mentioned in Sec. 2.2, this fine-tuning changes the weights of pre-trained diffusion model and might damage its ability in comprehending text-image pairs. Therefore, we propose to fix the pre-trained model $\\epsilon_{PT}(t)$ but shift it to $\\epsilon_{PT}(t + \\Delta t)$ to approximate the desired $\\epsilon_{FT}(t)$ . Referring to Fig. 1, we could shift $\\epsilon_{PT}(t)$ to an earlier timestep to achieve this goal. For example, at timestep $t_0$ and with a time shift $\\Delta t_0 > 0$ , we can use $\\epsilon_{PT}(t_0 + \\Delta t_0)$ to approximate the noise prediction error of $\\epsilon_{FT}(t_0)$ .", + "bbox": [ + 212, + 537, + 787, + 688 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "On the other hand, the magnitude of $\\Delta t$ will vary with $t$ . Let's come to another timestep $t_1$ in Fig. 1, where $t_1$ is earlier than $t_0$ . Because the decreasing speeds of both $e_{PT}$ and $e_{FT}$ will be reduced with $t$ going to $T_{\\mathrm{max}}$ , the magnitude of $\\Delta t_1$ will be increased to approximate $e_{FT}(t_1)$ . In other words, the magnitude of $\\Delta t$ should grow when $t$ goes from $T_{\\mathrm{min}}$ to $T_{\\mathrm{max}}$ . We heuristically set this relationship as $\\Delta t = \\eta (t - T_{\\mathrm{min}})$ , where $\\eta \\in [0,1]$ is a hyper-parameter that controls the length of shift range. Finally, it should be pointed out that the curves in Fig. 1 will vary a little for different training iterations, rendered images $\\mathbf{x}$ and text prompts $y$ . Therefore, $\\Delta t$ should fall into some range $S(t)$ . In practice, we set $\\Delta t \\sim S(t) = \\mathcal{U}[0,\\eta (t - T_{\\mathrm{min}})]$ , which follows a uniform distribution within 0", + "bbox": [ + 212, + 689, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "ScaleDreamer", + "bbox": [ + 637, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a632f5b8e981ceef862e3fd7e172e560f6a486258493c5184b3f536ace0e9f65.jpg", + "image_caption": [ + "Fig. 2: Left and middle: 2D toy examples by SDS [40], CSD [72], VSD [59] and our proposed ASD. Right: Gradient norms generated by different methods." + ], + "image_footnote": [], + "bbox": [ + 250, + 147, + 753, + 284 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Algorithm 1 Asynchronous Score Distillation (ASD)", + "text_level": 1, + "bbox": [ + 215, + 320, + 604, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Input: 3D representation $\\theta$ ; Text prompt $y$ ; Hyperparameter $\\eta$ ; 2D diffusion prior $\\epsilon_{\\phi}$", + "bbox": [ + 215, + 338, + 776, + 351 + ], + "page_idx": 7 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [], + "code_body": "1 while not converged do \n2 Sample a camera pose $\\pi$ \n3 Render an image $\\pmb{x} = g(\\theta, \\pi)$ \n4 Sample a timestep $t \\sim \\mathcal{U}[T_{\\min}, T_{\\max}]$ , Gaussian noise $\\epsilon \\sim \\mathcal{N}(0, I)$ \n5 Sample a timestep shift $\\Delta t \\sim S(t) = \\mathcal{U}[0, \\eta(t - T_{\\min})]$ \n6 $\\pmb{x}_t \\gets \\alpha_t \\pmb{x} + \\sigma_t \\pmb{\\epsilon}$ , $\\pmb{x}_{t + \\Delta t} \\gets \\alpha_{t + \\Delta t} \\pmb{x} + \\sigma_{t + \\Delta t} \\pmb{\\epsilon}$ \n7 Update $\\theta$ with $\\Delta \\theta \\gets \\omega(t) (\\pmb{\\epsilon}_{\\phi}(\\pmb{x}_t; t, y^{\\pi}) - \\pmb{\\epsilon}_{\\phi}(\\pmb{x}_{t + \\Delta t}; t + \\Delta t, y^{\\pi})) \\frac{\\partial \\pmb{x}}{\\partial \\theta}$ \n8 end", + "guess_lang": "latex", + "bbox": [ + 202, + 353, + 689, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and $\\eta(t - T_{\\min})$ . The pseudo-code of ASD is summarized in Alg. 1, which can be applied to both prompt-specific and prompt-amortized text-to-3D tasks.", + "bbox": [ + 212, + 502, + 784, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "2D toy experiments. To verify the proposed timestep shift strategy, we follow the paradigm in [59] to test SDS, CSD, VSD and our ASD on 2D toy examples. The left column of Fig. 2 shows the results of SDS, CSD, VSD, and the middle column shows the results of ASD with different sampling strategies of $\\Delta t$ . One can see that the proposed sampling strategy $\\Delta t \\sim S(t) = \\mathcal{U}[0, \\eta(t - T_{\\min})]$ yields similar results to VSD [59]. Besides, we show the gradient norm produced by these score distillation methods in the right column of Fig. 2. One can see that the range of gradient norm produced by ASD is similar to that of VSD. However, the gradient norm of SDS is more than 10 times larger than ASD and VSD because it needs to set CFG=100 for convergence [40, 59, 72]. Such a large gradient may result in training instability. We append more 2D results in the supplementary material to further validate our proposed sampling strategy.", + "bbox": [ + 212, + 535, + 785, + 717 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Text-to-3D Synthesis with ASD. As a score distillation method, ASD is open to the selection of 3D generator architectures [4,17,22,33,39]. The general pipeline of ASD for text-to-3D synthesis is shown in Fig. 3. It takes a rendered image as input and diffuses it in two timesteps $t$ and $t + \\Delta t$ . The noise prediction difference is used as the gradient to optimize the 3D representation of generator. In this work, in addition to prompt-specific generation, as done in most existing score distillation works [16, 28, 40, 59, 64], we focus more on prompt-amortized text-to-3D and conduct thorough experiments to evaluate the effectiveness of", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Z.Ma et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/38adc03247ab655fdd85b769d22913e0486abe4f7fc76d04425fc119318d0119.jpg", + "image_caption": [ + "Fig. 3: Overview of Asynchronous Score Distillation (ASD). As illustrated in the left sub-figure, ASD can be employed for prompt-specific generation by optimizing 3D representations for each prompt, as well as for prompt-amortized generation by training a text-to-3D generator. The right sub-figure depicts how ASD uses the difference in noise predictions at asynchronous timesteps to update the 3D network parameters." + ], + "image_footnote": [], + "bbox": [ + 245, + 143, + 754, + 268 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "ASD with representative 3D generator architectures, including Hyper-iNGP and 3DConv-net, by using Stable Diffusion as the 2D diffusion model. HyperiNGP is adopted by ATT3D [33], which integrates a prompt-agnostic hash-grid spatial encoding [39] with prompt-conditioned decoding layers to output color and density. 3DConv-net [4] is a 3D generator that maps the provided condition to voxel using 3D convolution. They are chosen in our experiments because they represent two groups of 3D generators that inject the condition through implicit mapping function [3,21] or explicit space representation [7,17,46,49]. The results are shown in Sec. 4. We also conduct experiments to demonstrate that ASD can work with other 2D diffusion models such as MVDream [48], and can be used to train other 3D generator architectures such as Triplane-Transformer [17]. More details of these experiments can be found in the supplementary material.", + "bbox": [ + 212, + 345, + 787, + 527 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 547, + 375, + 565 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Experimental Settings", + "text_level": 1, + "bbox": [ + 215, + 575, + 444, + 590 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Comparison Methods. We compare ASD with state-of-the-art score distillation methods, including SDS [40], CSD [72] and VSD [59]. We adhere to their official codes for training prompt-amortized text-to-3D networks. For example, the CFG [16] values for SDS, CSD and VSD are configured to 100, 1, and 7.5, respectively. In addition, we compare with existing prompt-amortized method ATT3D [33] (whose code is not released yet) by replicating its reported results.", + "bbox": [ + 212, + 598, + 784, + 688 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Implementation Details. We employ VolSDF [69] to render images from the 3D generators. For Stable Diffusion, we employ SD-v2.1-base [1] for all score distillation methods for fair comparison. As configured in VSD [59], we set the CFG value as 7.5 for the pre-trained diffusion model in ASD, and 1 for the diffusion model of rendered images. The resolution of rendered images by HyperiNGP is set to $256 \\times 256$ , while that of 3DConv-net and Triplane-Transformer is set to $64 \\times 64$ for GPU memory considerations. Other details are in the supplementary material.", + "bbox": [ + 212, + 688, + 785, + 808 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Prompt Corpus. To thoroughly evaluate the capability of ASD in prompt-amortized text-to-3D synthesis, we employ multiple datasets encompassing a", + "bbox": [ + 214, + 809, + 785, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "ScaleDreamer", + "bbox": [ + 637, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "range of text prompt quantities. MG15 includes 15 prompts from Magic3D [29]; DF415 comprises 415 prompts from DreamFusion [40]; and AT2520 contains 2520 compositional prompts of animals from ATT3D [33]. DL17k contains 17k compositional prompts of human with daily activities, proposed by [25]. While AT2520 and DL17k provide a larger number of prompts than DF415, the prompt diversity of them is relatively low due to the predefined templates.", + "bbox": [ + 212, + 146, + 787, + 236 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To test ASD's performance with an even larger scale of prompts, we introduce a novel prompt corpus named CP100k. This corpus consists of 100,000 text prompts filtered from the image descriptions collected by Cap3D [34], which was developed to test text-to-image model performance. To the best of our knowledge, it is the first time to evaluate score distillation methods on such a scale of text prompts. Meanwhile, it should be clarified that this work is focused on examining the score distillation performance rather than prompt generalization, so the test prompts share the same distribution as training prompts. More details of the prompt corpus are in supplementary material.", + "bbox": [ + 212, + 237, + 787, + 372 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation Metrics. We render 120 surrounding view images as the 3D synthesis result from each prompt. Similar to previous text-to-3D works [25,33, 33,40], we compute the CLIP recall, i.e., the classification accuracy by applying CLIP model to the rendered images to predict the correct text prompt, as one performance metric, denoted by \"R@1\". Additionally, we calculate the CLIP text-image similarity between generated images and input prompts as another metric [54,60], denoted by \"Sim\".", + "bbox": [ + 212, + 373, + 787, + 479 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Evaluation Results", + "text_level": 1, + "bbox": [ + 214, + 500, + 418, + 513 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Results with iNGP/Hyper-iNGP as 3D Representation. The iNGP [39] architecture is designed for prompt-specific text-to-3D generation. Hyper-iNGP has the same spatial encoding as iNGP except that the weights of the decoding layer depend on the text prompt. To eliminate the effect caused by architecture difference as much as possible, we adopt iNGP for prompt-specific text-to-3D tasks, and Hyper-iNGP for prompt-amortized tasks. Our experiments are carried out on the MG15 dataset. For prompt-specific tasks, we optimize an individual iNGP [39] for each MG15 prompt; while for the prompt-amortized tasks, we train a single Hyper-iNGP [33] across all MG15 prompts. We also compare our results with ATT3D [33], which is among the first to apply Hyper-iNGP to prompt-amortized text-to-3D tasks. ATT3D employs SDS for training and uses soft-shading [40] (denoted as * in Tab. 2) for rendering.", + "bbox": [ + 212, + 521, + 787, + 703 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The qualitative and quantitative results are shown in Fig. 4 and Tab. 2, respectively. We can see that the existing methods suffer from performance decrease when transiting from prompt-specific to prompt-amortized tasks, as evidenced by the decreased CLIP similarity and recall in Tab. 2. It is worth mentioning that training Hyper-net with SDS requires turning on the spectral normalization [38] in the linear layers, otherwise the training will fail due to numerical instability. This observation is consistent with what reported in ATT3D [33]. This is because SDS suffers from large gradient norm (please also refer to Fig. 2 and the discussions therein), which makes Hyper-iNGP hard to", + "bbox": [ + 212, + 704, + 787, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Z.Ma et al.", + "bbox": [ + 271, + 114, + 346, + 127 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/408addbe1a5f9af089cabaa05c98be98d1cddd5e3eeb7aa992503005b8fc424a.jpg", + "image_caption": [ + "Fig. 4: Qualitative comparison on prompt-specific (with iNGP as the 3D representation) and prompt-amortized (with Hyper-iNGP as the 3D generator) text-to-3D results by SDS [40], CSD [72], VSD [59], ATT3D [33] and our ASD methods." + ], + "image_footnote": [], + "bbox": [ + 248, + 145, + 354, + 257 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/70f23e0a08667dcc4ee0b16f3e88208beb54d087ffb6d5993bfb832b8cd2dadb.jpg", + "image_caption": [ + "\"A DSLR photo of a peacok on a surfboard\" in MG15" + ], + "image_footnote": [], + "bbox": [ + 359, + 145, + 457, + 257 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/5ff548565ca51d53e4528d5ac75a260fc105de4301fd93dce9de5fdba658bf00.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 460, + 146, + 558, + 255 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/f1c78c07b9f778bee145766d543dc69bdf4e1170162c7308440824061e2e9a28.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 560, + 146, + 627, + 253 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/89da58504250c50a28819e51424c8b9344b20cf333d7a496a2765dcaa8610e40.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 629, + 146, + 758, + 253 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/e87999bd354efb422286b2753fe5b4c512770b1ed68a15e780a78512276a1f19.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ReferenceMethodSim ↑R@1 ↑MethodSim ↑R@1 ↑
ATT3D [33]---Hyper-iNGP* + SDS0.1950.468
DreamFusion [40]iNGP + SDS0.2881.000Hyper-iNGP + SDS0.2570.918
Classifier [16]iNGP + CSD0.2800.936Hyper-iNGP + CSD0.2640.972
ProlificDreamer [59]iNGP + VSD0.2760.932Hyper-iNGP + VSD0.2590.987
OursiNGP + ASD0.2891.000Hyper-iNGP + ASD0.2841.000
", + "bbox": [ + 215, + 319, + 787, + 425 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 2: Quantitative comparison on prompt-specific (with iNGP as the 3D representation) and prompt-amortized (with Hyper-iNGP as the 3D generator) text-to-3D results by SDS [40], CSD [72], VSD [59], ATT3D [33] and our ASD methods.", + "bbox": [ + 214, + 425, + 782, + 467 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "converge. As can be seen in Fig. 4, ATT3D results in wrong geometry by using soft shading and SDS for training. For CSD, we see that it fails to optimize the full geometry, as shown by the shrunk peacock in both prompt-amortized and prompt-amortized results. For VSD, it tends to generate content drifts [48], resulting in repetitive patterns and abnormal geometry. It may fail to generate reasonable contents in both prompt-specific and prompt-amortized tasks. In contrast, our proposed ASD works very stable across the two tasks, yielding not only outstanding quantitative scores but also high quality 3D contents.", + "bbox": [ + 212, + 487, + 784, + 607 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Results with 3DConv-net as 3D Generator. The issues of existing score distillation methods either persist or become more pronounced when replacing Hyper-iNGP to 3DConv-net as the 3D generator. We find that training SDS with 3DConv-net always fails within several thousand iterations, even using spectral or other normalization techniques. This issue stems from that deeper network is more sensitive to large gradients [13] caused by SDS. Therefore, we only compare the results of other methods in Fig. 5. We see that CSD outputs acceptable results on AT2520, but its results on DF415, which has more varied prompts, are consistently smaller than anticipated. Such a phenomenon has been observed when Hyper-iNGP is used as the generator, which underlines CSD's inability to reliably guide the 3D generator to produce geometries aligned with the text prompts. As for VSD, it leads to rather abnormal results, failing to match the text prompts. This can be attributed to its fine-tuning of the pre-trained 2D diffusion model, which severely compromises VSD's text-image comprehending ability. In comparison, our proposed ASD, with 3DConv-net as the generator,", + "bbox": [ + 212, + 613, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "ScaleDreamer", + "bbox": [ + 637, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/b3a37a77b484107a67ee819383b2523eaae0bb583d0f64c89985bbddfb578f32.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 248, + 145, + 388, + 209 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/3d868fa32efcdbdd740f76d8d754a5ac827cea9849b470efc4fe52b044cb5b4e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 395, + 145, + 542, + 209 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/24446fcc06bb1364faf18446ffdc62d4e0454581cde6a5660f0e42c79fdb78be.jpg", + "image_caption": [ + "Asynchronous Score Distillation (ASD, ours)", + "\"A squirrel holding a book wearing a sweater wearing a tophat\" in AT2520" + ], + "image_footnote": [], + "bbox": [ + 547, + 156, + 756, + 210 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/a5ca991def4b83545b4ae7a81043c95b557b01ff73bedf54aaef828fca7134b8.jpg", + "image_caption": [ + "Fig. 5: Qualitative comparison among CSD [72], VSD [59] and our ASD (with 3DConv-net as generator) on AT2520 and DF415 corpuses. SDS is not compared because it encounters numerical instability in this experiment." + ], + "image_footnote": [], + "bbox": [ + 248, + 219, + 316, + 271 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/44b3d42e96d1a3292a1a7e001c390c0e1b5848872b77cf845b9224fd7a1bbd04.jpg", + "image_caption": [ + "\"A DSLR photo of a cocker spaniel wearing a crown\" in DF415" + ], + "image_footnote": [], + "bbox": [ + 318, + 219, + 387, + 271 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/450fab972bbfc2c45476959b453c13a79c0fb0f539d28be8f77000d69e45cfde.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 388, + 219, + 468, + 271 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/e3394f68d5b3630cbecd4f07c52674033542965ca25c2e3262b6d00195c8aeea.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 470, + 219, + 540, + 271 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/164be7bcf11c40c71aacf25f2af2062f50ba2407999f6fea028fd83591c7d520.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 547, + 219, + 616, + 271 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/749cd727c6f06661375d504d3d893cba7c7d988f5d7e072f021e72f46949fc65.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 219, + 686, + 271 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/e662d35d7a692f9fb4d5eaa18a7187cf08b18d429301d6c631c0b6a5a94696be.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 219, + 756, + 271 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/beb5b096a6e65ba95f1aab24a7e7130555a6478251c19eec1de6b91d90c9eb70.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodDF415AT2520CP100k
Sim ↑R@1 ↑Sim ↑R@1 ↑Sim ↑R@1 ↑
SDS××××××
CSD0.1760.0620.2790.0370.1950.108
VSD0.1580.0020.1150.0010.1030.000
ASD (ours)0.2370.2760.2850.0580.1990.117
", + "bbox": [ + 276, + 325, + 728, + 429 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 3: Quantitative comparison on prompt-amortized text-to-3D with 3DConv-net as generator. Symbol $\\times$ denotes that the training fails due to numerical instability.", + "bbox": [ + 214, + 430, + 784, + 458 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "yields improved outcomes, as evidenced by the visual results in Fig. 5 and the enhanced metric scores in Tab. 3.", + "bbox": [ + 212, + 464, + 784, + 493 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Scalability. In this section, we evaluate the scalability of competing methods by using as many as 100k prompts in the CP100k dataset with 3DConv-net as the generator. The results are shown in Fig. 6 and Tab. 3. Due to the issue of numerical instability, SDS is not involved in this experiment. We can see that the outcomes of CSD are significantly diminished with uniformly small-sized shapes across all prompts. There is also a lack of variety since most outputs exhibit similar patterns. The results of VSD are also degenerated, displaying almost identical and anomalous outcomes for the text prompts. This resembles the phenomenon of mode collapse often encountered in bi-level optimization [55], which also highlights the importance of fixing the 2D diffusion model when training with such a large number of text prompts. In comparison, ASD is able to produce much higher quality outcomes across the text prompts, showcasing its capability in large-scale training with numerous text prompts as inputs.", + "bbox": [ + 212, + 494, + 785, + 691 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In addition to the above comparisons, we also evaluate the effectiveness of ASD when using other 2D diffusion models. We employ MVDream [48] to illustrate ASD's generality to diffusion prior models. Besides, we compare against data-driven methods to prove that ASD is advantageous in tackling diverse input of text prompts. Please see supplementary material for detail.", + "bbox": [ + 212, + 691, + 785, + 767 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 786, + 388, + 801 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In this section, we perform ablation studies to evaluate the settings of timestep shift $\\Delta t\\sim S(t) = \\mathcal{U}[0,\\eta (t - T_{\\mathrm{min}})]$ from several aspects. The qualitative and", + "bbox": [ + 212, + 809, + 784, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Z.Ma et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/fa6b09ba5048a47e5b6389bbe3d4e5b0b391bc2dfaffa711d11225f95d38282b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 243, + 143, + 406, + 262 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/d76ea56d08e5a8f55c72f62961598316a507b2285fadf685735f1b0893149965.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 418, + 145, + 581, + 261 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/c12bcf73513fd51a5dac999544ebd93ab04829e977d8f310604e724038b9dfc1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 586, + 145, + 751, + 262 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/0192162f68d2fc7bc887c2950ede62e400914c1b423010cba396e5a226cc99ba.jpg", + "image_caption": [ + "Fig. 6: The scalability comparison with CSD [72] and VSD [59] on CP100k corpus." + ], + "image_footnote": [], + "bbox": [ + 245, + 271, + 406, + 321 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/a78e82c0fd9d95b6c49ff83b72dc71e4cd27947f1461ab84884f1f753fa1b43e.jpg", + "image_caption": [ + "\"Small wooden brown bear figurine\"" + ], + "image_footnote": [], + "bbox": [ + 418, + 263, + 580, + 321 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/8da0bfae92e89486c3c7b9911273aa674f195a7ba6ad3986bfb4f3f8d63fe956.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 586, + 271, + 750, + 321 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "quantitative results are shown in Fig. 7 and Tab. 4, respectively. We provide extensive ablation studies in the supplementary material to show that the proposed strategy is also effective when using MVDream as the diffusion model.", + "bbox": [ + 212, + 356, + 784, + 400 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Importance of Timestep Shift. We use $\\eta = 0$ (i.e., no timestep shift) as a baseline to evaluate the necessity of introducing timestep shift $\\Delta t$ . From Fig. 7 and Tab. 4, we see that while it can generate plausible results, the model is prone to generating shapes that do not make sense, such as the so-called Janus problem [2]. Examples include a frog with an extra eye, robot face with block-like features, and a peacock with tails at both the front and back. This is because the non-shifted diffusion model will align more with the 2D image distribution, tending to generate redundant contents and unreasonable geometry along the training. By introducing a timestep shift, our proposed ASD demonstrates advantages in achieving more coherent and visually pleasing results.", + "bbox": [ + 212, + 402, + 784, + 551 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Range of Timestep Shift. By setting $\\eta = 0.2$ , we allow $\\Delta t$ to be sampled from a large range. However, this might not be a good choice. In the extreme case, for any timestep $t$ we can set a large interval $\\Delta t$ such that $t + \\Delta t = T_{\\max}$ , then the noise prediction becomes $\\epsilon_{\\phi}(\\boldsymbol{x}_t; t, y^\\pi) \\approx \\epsilon$ , so that ASD is degraded to SDS, which cannot perform well under CFG=7.5 [40]. In practice, we find a larger $\\eta$ tends to result 3D contents with larger size and rounded shapes, e.g., the peacock with closer views, or the frog with larger size, as shown in Fig. 7. Therefore, we set $\\eta = 0.1$ in all our experiments.", + "bbox": [ + 212, + 553, + 784, + 672 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Deterministic or Random Shift. If we set $\\Delta t = \\eta (t - T_{\\mathrm{min}})$ , it assumes that the diffusion model of rendered images can be approximated by the pretrained one with a fixed and deterministic timestep shift. As shown in Fig. 7 and Tab. 4, it reduces the chance to generate correct geometry and colors. Randomly sampling $\\Delta t$ in a range is more effective, which is adopted in our method.", + "bbox": [ + 212, + 675, + 784, + 750 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5 Conclusion and Limitations", + "text_level": 1, + "bbox": [ + 215, + 775, + 517, + 791 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In this paper, we presented Asynchronous Score Distillation (ASD), a novel score distillation method that can assist 2D diffusion prior in training 3D generators", + "bbox": [ + 212, + 809, + 784, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "ScaleDreamer", + "bbox": [ + 637, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/bab712512d378c9d64acee37b1fdfd89aa6daeef99c029a4aa80b8eebb1529c5.jpg", + "image_caption": [ + "Fig. 7: The qualitative results of the ablation study on the timestep interval $\\Delta t$ ." + ], + "image_footnote": [], + "bbox": [ + 233, + 143, + 774, + 383 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/df0fa4d5c6e3ffce1a87ada3d118b1dcfbc640b7c3685076b0ee621b16d0c98e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ParamSim ↑R@1 ↑
Δt = η(t - Tmin)η = 0.10.2140.178
η = 0.20.2140.180
η = 00.2350.267
Δt ~ U[0, η(t - Tmin)]η = 0.10.2370.276
η = 0.20.2290.237
", + "bbox": [ + 330, + 402, + 671, + 501 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Table 4: The quantitative results of the ablation study on the timestep interval $\\Delta t$ .", + "bbox": [ + 218, + 502, + 779, + 516 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "with a scalable size of text prompts. By shifting the diffusion timestep to earlier stages, our ASD can effectively predict the noise prediction error to align the diffusion model with the distribution of rendered images, while preserving the superior text comprehension capability of pre-trained models, thus facilitating stable training with high-fidelity generation results. Our extensive experiments revealed that ASD performed consistently well on datasets of various sizes, being able to manage as much as 100k prompts. Though ASD has shown improvements over earlier score distillation approaches, there remain some limitations.", + "bbox": [ + 212, + 530, + 787, + 651 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "For man-made objects that have very regular shapes, such as chairs or airplanes, the performance of our model will lag behind those data-driven methods, which benefit from an abundance of relevant data. We foresee opportunities to combine the advantages of data-driven and score distillation methodologies to improve text-to-3D capabilities in a more comprehensive manner in the future research.", + "bbox": [ + 212, + 652, + 787, + 742 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 215, + 772, + 393, + 789 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This work is supported in part by the Beijing Science and Technology Plan Project Z231100005923033, and the InnoHK program.", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Z.Ma et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Stable-diffusion-v2.1-base. https://huggingface.co/stabilityai/stable-diffusion-2-1-base", + "2. Armandpour, M., Zheng, H., Sadeghian, A., Sadeghian, A., Zhou, M.: Re-imagine the negative prompt algorithm: Transform 2d diffusion into 3d, alleviate janus problem and beyond. arXiv preprint arXiv:2304.04968 (2023)", + "3. Babu, S., Liu, R., Zhou, A., Maire, M., Shakhnarovich, G., Hanocka, R.: Hyperfields: Towards zero-shot generation of nerfs from text. arXiv preprint arXiv:2310.17075 (2023)", + "4. Bahmani, S., Park, J.J., Paschalidou, D., Yan, X., Wetzstein, G., Guibas, L., Tagliasacchi, A.: Cc3d: Layout-conditioned generation of compositional 3d scenes. arXiv preprint arXiv:2303.12074 (2023)", + "5. Balaji, Y., Nah, S., Huang, X., Vahdat, A., Song, J., Kreis, K., Aittala, M., Aila, T., Laine, S., Catanzaro, B., et al.: edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324 (2022)", + "6. Cao, Z., Hong, F., Wu, T., Pan, L., Liu, Z.: Large-vocabulary 3d diffusion model with transformer. arXiv preprint arXiv:2309.07920 (2023)", + "7. Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., De Mello, S., Gallo, O., Guibas, L.J., Tremblay, J., Khamis, S., et al.: Efficient geometry-aware 3d generative adversarial networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16123-16133 (2022)", + "8. Chen, R., Chen, Y., Jiao, N., Jia, K.: *Fantasia3d: Disentangling geometry* and appearance for high-quality text-to-3d content creation. arXiv preprint arXiv:2303.13873 (2023)", + "9. Chen, Z., Wang, F., Liu, H.: Text-to-3d using gaussian splatting. arXiv preprint arXiv:2309.16585 (2023)", + "10. Deitke, M., Liu, R., Wallingford, M., Ngo, H., Michel, O., Kusupati, A., Fan, A., Laforte, C., Voleti, V., Gadre, S.Y., et al.: Objaverse-xl: A universe of $10\\mathrm{m} + 3\\mathrm{d}$ objects. arXiv preprint arXiv:2307.05663 (2023)", + "11. Ding, L., Dong, S., Huang, Z., Wang, Z., Zhang, Y., Gong, K., Xu, D., Xue, T.: Text-to-3d generation with bidirectional diffusion using both 2d and 3d priors. arXiv preprint arXiv:2312.04963 (2023)", + "12. Guo, P., Hao, H., Caccavale, A., Ren, Z., Zhang, E., Shan, Q., Sankar, A., Schwing, A.G., Colburn, A., Ma, F.: Stabledreamer: Taming noisy score distillation sampling for text-to-3d. arXiv preprint arXiv:2312.02189 (2023)", + "13. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)", + "14. He, Y., Bai, Y., Lin, M., Zhao, W., Hu, Y., Sheng, J., Yi, R., Li, J., Liu, Y.J.: T $^3$ bench: Benchmarking current progress in text-to-3d generation. arXiv preprint arXiv:2310.02977 (2023)", + "15. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020)", + "16. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)", + "17. Hong, Y., Zhang, K., Gu, J., Bi, S., Zhou, Y., Liu, D., Liu, F., Sunkavalli, K., Bui, T., Tan, H.: Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400 (2023)" + ], + "bbox": [ + 218, + 179, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "ScaleDreamer", + "bbox": [ + 637, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021)", + "19. Huang, Y., Wang, J., Shi, Y., Qi, X., Zha, Z.J., Zhang, L.: Dreamtime: An improved optimization strategy for text-to-3d content creation. arXiv preprint arXiv:2306.12422 (2023)", + "20. Jiang, L., Wang, L.: Brightdreamer: Generic 3d gaussian generative framework for fast text-to-3d synthesis. arXiv preprint arXiv:2403.11273 (2024)", + "21. Jun, H., Nichol, A.: Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463 (2023)", + "22. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG) 42(4), 1-14 (2023)", + "23. Koster, R.: Theory of fun for game design. \"O'Reilly Media, Inc.\" (2013)", + "24. Lee, K., Sohn, K., Shin, J.: Dreamflow: High-quality text-to-3d generation by approximating probability flow. arXiv preprint arXiv:2403.14966 (2024)", + "25. Li, M., Zhou, P., Liu, J.W., Keppo, J., Lin, M., Yan, S., Xu, X.: Instant3d: Instant text-to-3d generation. arXiv preprint arXiv:2311.08403 (2023)", + "26. Li, W., Chen, R., Chen, X., Tan, P.: Sweetdreamer: Aligning geometric priors in 2d diffusion for consistent text-to-3d. arXiv preprint arXiv:2310.02596 (2023)", + "27. Li, Z., Chen, Y., Zhao, L., Liu, P.: Mvcontrol: Adding conditional control to multi-view diffusion for controllable text-to-3d generation. arXiv preprint arXiv:2311.14494 (2023)", + "28. Liang, Y., Yang, X., Lin, J., Li, H., Xu, X., Chen, Y.: Luciddreamer: Towards high-fidelity text-to-3d generation via interval score matching. arXiv preprint arXiv:2311.11284 (2023)", + "29. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3d: High-resolution text-to-3d content creation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 300–309 (2023)", + "30. Lin, Y., Clark, R., Torr, P.: Dreamampolisher: Towards high-quality text-to-3d generation via geometric diffusion. arXiv preprint arXiv:2403.17237 (2024)", + "31. Liu, Y.T., Luo, G., Sun, H., Yin, W., Guo, Y.C., Zhang, S.H.: Pi3d: Efficient text-to-3d generation with pseudo-image diffusion. arXiv preprint arXiv:2312.09069 (2023)", + "32. Liu, Z., Li, Y., Lin, Y., Yu, X., Peng, S., Cao, Y.P., Qi, X., Huang, X., Liang, D., Ouyang, W.: Unidream: Unifying diffusion priors for relightable text-to-3d generation. arXiv preprint arXiv:2312.08754 (2023)", + "33. Lorraine, J., Xie, K., Zeng, X., Lin, C.H., Takikawa, T., Sharp, N., Lin, T.Y., Liu, M.Y., Fidler, S., Lucas, J.: Att3d: Amortized text-to-3d object synthesis. arXiv preprint arXiv:2306.07349 (2023)", + "34. Luo, T., Rockwell, C., Lee, H., Johnson, J.: Scalable 3d captioning with pretrained models. Advances in Neural Information Processing Systems 36 (2024)", + "35. Ma, Y., Fan, Y., Ji, J., Wang, H., Sun, X., Jiang, G., Shu, A., Ji, R.: X-dreamer: Creating high-quality 3d content by bridging the domain gap between text-to-2d and text-to-3d generation. arXiv preprint arXiv:2312.00085 (2023)", + "36. Mercier, A., Nakhli, R., Reddy, M., Yasarla, R., Cai, H., Porikli, F., Berger, G.: Hexagen3d: Stablediffusion is just one step away from fast and diverse text-to-3d generation. arXiv preprint arXiv:2401.07727 (2024)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Z.Ma et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "37. Metzer, G., Richardson, E., Patashnik, O., Giryes, R., Cohen-Or, D.: Latent-nerf for shape-guided generation of 3d shapes and textures. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12663-12673 (2023)", + "38. Miyato, T., Kataoka, T., Koyama, M., Yoshida, Y.: Spectral normalization for generative adversarial networks. arXiv preprint arXiv:1802.05957 (2018)", + "39. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022)", + "40. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988 (2022)", + "41. Qian, G., Cao, J., Siarohin, A., Kant, Y., Wang, C., Vasilkovsky, M., Lee, H.Y., Fang, Y., Skorokhodov, I., Zhuang, P., et al.: Atom: Amortized text-to-mesh using 2d diffusion. arXiv preprint arXiv:2402.00867 (2024)", + "42. Qiu, L., Chen, G., Gu, X., Zuo, Q., Xu, M., Wu, Y., Yuan, W., Dong, Z., Bo, L., Han, X.: Richdreamer: A generalizable normal-depth diffusion model for detail richness in text-to-3d. arXiv preprint arXiv:2311.16918 (2023)", + "43. Ren, X., Huang, J., Zeng, X., Museth, K., Fidler, S., Williams, F.: Xcube: Large-scale 3d generative modeling using sparse voxel hierarchies. arXiv preprint (2023)", + "44. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)", + "45. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022)", + "46. Schwarz, K., Sauer, A., Niemeyer, M., Liao, Y., Geiger, A.: Voxgraf: Fast 3d-aware image synthesis with sparse voxel grids. Advances in Neural Information Processing Systems 35, 33999-34011 (2022)", + "47. Shen, T., Gao, J., Yin, K., Liu, M.Y., Fidler, S.: Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. Advances in Neural Information Processing Systems 34, 6087-6101 (2021)", + "48. Shi, Y., Wang, P., Ye, J., Long, M., Li, K., Yang, X.: Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512 (2023)", + "49. Sitzmann, V., Thies, J., Heide, F., Nießner, M., Wetzstein, G., Zollhofer, M.: Deepvoxels: Learning persistent 3d feature embeddings. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2437-2446 (2019)", + "50. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456 (2020)", + "51. Tang, B., Wang, J., Wu, Z., Zhang, L.: Stable score distillation for high-quality 3d generation. arXiv preprint arXiv:2312.09305 (2023)", + "52. Tang, J., Chen, Z., Chen, X., Wang, T., Zeng, G., Liu, Z.: Lgm: Large multi-view gaussian model for high-resolution 3d content creation. arXiv preprint arXiv:2402.05054 (2024)", + "53. Tang, J., Ren, J., Zhou, H., Liu, Z., Zeng, G.: Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653 (2023)", + "54. Tang, Z., Gu, S., Wang, C., Zhang, T., Bao, J., Chen, D., Guo, B.: Volumediffusion: Flexible text-to-3d generation with efficient volumetric encoder. arXiv preprint arXiv:2312.11459 (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "ScaleDreamer", + "bbox": [ + 637, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "55. Thanh-Tung, H., Tran, T.: Catastrophic forgetting and mode collapse in gans. In: 2020 international joint conference on neural networks (ijcnn). pp. 1-10. IEEE (2020)", + "56. Tsalicoglou, C., Manhardt, F., Tonioni, A., Niemeyer, M., Tombari, F.: Textmesh: Generation of realistic 3d meshes from text prompts. arXiv preprint arXiv:2304.12439 (2023)", + "57. Vilesov, A., Chari, P., Kadambi, A.: Cg3d: Compositional generation for text-to-3d via gaussian splatting. arXiv preprint arXiv:2311.17907 (2023)", + "58. Wang, H., Du, X., Li, J., Yeh, R.A., Shakhnarovich, G.: Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12619-12629 (2023)", + "59. Wang, Z., Lu, C., Wang, Y., Bao, F., Li, C., Su, H., Zhu, J.: Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213 (2023)", + "60. Wei, Y., Zhang, Y., Ji, Z., Bai, J., Zhang, L., Zuo, W.: Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. arXiv preprint arXiv:2302.13848 (2023)", + "61. Wohlgenannt, I., Simons, A., Stieglitz, S.: Virtual reality. Business & Information Systems Engineering 62, 455-461 (2020)", + "62. Wu, R., Sun, L., Ma, Z., Zhang, L.: One-step effective diffusion network for real-world image super-resolution. arXiv preprint arXiv:2406.08177 (2024)", + "63. Wu, T., Zhang, J., Fu, X., Wang, Y., Ren, J., Pan, L., Wu, W., Yang, L., Wang, J., Qian, C., et al.: Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 803-814 (2023)", + "64. Wu, Z., Zhou, P., Yi, X., Yuan, X., Zhang, H.: Consistent3d: Towards consistent high-fidelity text-to-3d generation with deterministic sampling prior. arXiv preprint arXiv:2401.09050 (2024)", + "65. Xie, K., Lorraine, J., Cao, T., Gao, J., Lucas, J., Torralba, A., Fidler, S., Zeng, X.: Latte3d: Large-scale amortized text-to-enhanced3d synthesis. arXiv preprint arXiv:2403.15385 (2024)", + "66. Xu, Y., Tan, H., Luan, F., Bi, S., Wang, P., Li, J., Shi, Z., Sunkavalli, K., Wetzstein, G., Xu, Z., et al.: Dmv3d: Denoising multi-view diffusion using 3d large reconstruction model. arXiv preprint arXiv:2311.09217 (2023)", + "67. Yang, Z., Feng, R., Zhang, H., Shen, Y., Zhu, K., Huang, L., Zhang, Y., Liu, Y., Zhao, D., Zhou, J., et al.: Eliminating lipschitz singularities in diffusion models. arXiv preprint arXiv:2306.11251 (2023)", + "68. Yang, Z., Feng, R., Zhang, H., Shen, Y., Zhu, K., Huang, L., Zhang, Y., Liu, Y., Zhao, D., Zhou, J., et al.: Lipschitz singularities in diffusion models. In: The Twelfth International Conference on Learning Representations (2023)", + "69. Yariv, L., Gu, J., Kasten, Y., Lipman, Y.: Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems 34, 4805-4815 (2021)", + "70. Yi, T., Fang, J., Wu, G., Xie, L., Zhang, X., Liu, W., Tian, Q., Wang, X.: Gaussian dreamer: Fast generation from text to 3d gaussian splatting with point cloud priors. arXiv preprint arXiv:2310.08529 (2023)", + "71. Yu, X., Xu, M., Zhang, Y., Liu, H., Ye, C., Wu, Y., Yan, Z., Zhu, C., Xiong, Z., Liang, T., et al.: Mvimgnet: A large-scale dataset of multi-view images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9150-9161 (2023)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Z.Ma et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "72. Yu, X., Guo, Y.C., Li, Y., Liang, D., Zhang, S.H., Qi, X.: Text-to-3d with classifier score distillation. arXiv preprint arXiv:2310.19415 (2023)", + "73. Zhao, M., Zhao, C., Liang, X., Li, L., Zhao, Z., Hu, Z., Fan, C., Yu, X.: Efficient-dreamer: High-fidelity and robust 3d creation via orthogonal-view diffusion prior. arXiv preprint arXiv:2308.13223 (2023)", + "74. Zhao, R., Wang, Z., Wang, Y., Zhou, Z., Zhu, J.: Flexidreamer: Single image-to-3d generation with flexicubes. arXiv preprint arXiv:2404.00987 (2024)", + "75. Zhou, L., Shih, A., Meng, C., Ermon, S.: Dreampropeller: Supercharge text-to-3d generation with parallel sampling. arXiv preprint arXiv:2311.17082 (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 273 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "ScaleDreamer", + "bbox": [ + 637, + 114, + 730, + 126 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/ScaleDreamer_ Scalable Text-to-3D Synthesis with Asynchronous Score Distillation/0ae3774a-e55b-4154-9080-499a583a8778_model.json b/2024/ScaleDreamer_ Scalable Text-to-3D Synthesis with Asynchronous Score Distillation/0ae3774a-e55b-4154-9080-499a583a8778_model.json new file mode 100644 index 0000000000000000000000000000000000000000..cbd273e04c8748379901d80f60905bf89a842f63 --- /dev/null +++ b/2024/ScaleDreamer_ Scalable Text-to-3D Synthesis with Asynchronous Score Distillation/0ae3774a-e55b-4154-9080-499a583a8778_model.json @@ -0,0 +1,2724 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.245, + 0.141, + 0.761, + 0.187 + ], + "angle": 0, + "content": "ScaleDreamer: Scalable Text-to-3D Synthesis with Asynchronous Score Distillation" + }, + { + "type": "text", + "bbox": [ + 0.312, + 0.212, + 0.688, + 0.228 + ], + "angle": 0, + "content": "Zhiyuan Ma\\(^{1,2}\\), Yuxiang Wei\\(^{1,5}\\), Yabin Zhang\\(^{1}\\)," + }, + { + "type": "text", + "bbox": [ + 0.304, + 0.232, + 0.699, + 0.251 + ], + "angle": 0, + "content": "Xiangyu Zhu\\(^{3,4}\\), Zhen Lei\\(^{1,2,3,4^{\\star}}\\), and Lei Zhang\\(^{1^{\\star}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.338, + 0.262, + 0.664, + 0.276 + ], + "angle": 0, + "content": "1 The Hong Kong Polytechnic University, PolyU" + }, + { + "type": "text", + "bbox": [ + 0.298, + 0.276, + 0.705, + 0.289 + ], + "angle": 0, + "content": "\\(^{2}\\) Center for Artificial Intelligence and Robotics, HKISI CAS" + }, + { + "type": "text", + "bbox": [ + 0.244, + 0.29, + 0.761, + 0.304 + ], + "angle": 0, + "content": "3 State Key Laboratory of Multimodal Artificial Intelligence Systems, CASIA" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.304, + 0.782, + 0.318 + ], + "angle": 0, + "content": "\\(^{4}\\) School of Artificial Intelligence, University of Chinese Academy of Sciences, UCAS" + }, + { + "type": "text", + "bbox": [ + 0.375, + 0.319, + 0.628, + 0.332 + ], + "angle": 0, + "content": "5 Harbin Institute of Technology, HIT" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.362, + 0.74, + 0.709 + ], + "angle": 0, + "content": "Abstract. By leveraging the text-to-image diffusion prior, score distillation can synthesize 3D contents without paired text-3D training data. Instead of spending hours of online optimization per text prompt, recent studies have been focused on learning a text-to-3D generative network for amortizing multiple text-3D relations, which can synthesize 3D contents in seconds. However, existing score distillation methods are hard to scale up to a large amount of text prompts due to the difficulties in aligning pretrained diffusion prior with the distribution of rendered images from various text prompts. Current state-of-the-arts such as Variational Score Distillation finetune the pretrained diffusion model to minimize the noise prediction error so as to align the distributions, which are however unstable to train and will impair the model's comprehension capability to numerous text prompts. Based on the observation that the diffusion models tend to have lower noise prediction errors at earlier timesteps, we propose Asynchronous Score Distillation (ASD), which minimizes the noise prediction error by shifting the diffusion timestep to earlier ones. ASD is stable to train and can scale up to \\(100k\\) prompts. It reduces the noise prediction error without changing the weights of pre-trained diffusion model, thus keeping its strong comprehension capability to prompts. We conduct extensive experiments using different text-to-3D architectures, including Hyper-iNGP and 3DConvNet. The results demonstrate ASD's effectiveness in stable 3D generator training, high-quality 3D content synthesis, and its superior prompt-consistency, especially under large prompt corpus. Code is available at https://github.com/theEricMa/ScaleDreamer." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.721, + 0.674, + 0.735 + ], + "angle": 0, + "content": "Keywords: Text-to-3D \\(\\cdot\\) Score Distillation \\(\\cdot\\) Diffusion Model" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.758, + 0.375, + 0.773 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.787, + 0.787, + 0.819 + ], + "angle": 0, + "content": "Text-to-3D aims to generate realistic 3D contents from the given textual descriptions [40], which is particularly useful in many applications such as virtual" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.393, + 0.84 + ], + "angle": 0, + "content": "* Corresponding authors." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.348, + 0.128 + ], + "angle": 0, + "content": "Z.Ma et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.358 + ], + "angle": 0, + "content": "reality [61] and game design [23]. The main challenge of this task, however, lies in how to generate high-quality 3D contents conditioned on the abstract and diverse textual descriptions. Many existing text-to-3D methods [11, 12, 26–29, 32, 35, 37, 40, 42, 58, 59, 75] are optimization-based ones, which distill the guidance from the powerful pretrained text-to-image diffusion models [5, 11, 26, 32, 42, 44, 73] via score distillation [40, 59, 62, 72]. In general, these methods employ the KL divergence to reduce the discrepancy between the distribution of rendered images and the desired image distribution embedded in the 2D diffusion prior, while they differ in how to use the pretrained diffusion prior to model the distribution of rendered images. Extensive efforts have been made to explore prompt-specific optimization of various 3D representations, including implicit radiance fields [40], explicit radiance fields [29, 37, 59], DmTets [56, 74] and 3D Gaussians [9]. Typically, tens of minutes to hours are needed to optimize a single 3D representation for one prompt to achieve the desired result." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.363, + 0.788, + 0.575 + ], + "angle": 0, + "content": "Compared to the aforementioned optimization-based text-to-3D methods, learning-based methods [6, 21, 31, 36, 43, 54, 65] can largely reduce the computational cost by training a text-conditioned 3D generative network. With the availability of 3D object collections [10, 63, 71], a deep network can be trained in a supervised manner so that 3D outputs can be generated in several seconds. Unfortunately, the size of existing text-3D datasets is far from sufficient compared to text-image datasets [45], limiting the text-to-3D generation performance of trained models. Inspired by the optimization-based text-to-3D methods that use pretrained 2D diffusion models, efforts have been made to train text-to-3D networks by using 2D diffusion models as supervisors [33, 41, 65] without using text-3D pairs. For example, a text-conditioned 3D hyper-network is trained in ATT3D [33] via Score Distillation Sampling (SDS) [40]. Nevertheless, this method suffers from numerical instability, which has been observed in subsequent studies [41, 65] that apply SDS to different 3D generator networks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.58, + 0.788, + 0.761 + ], + "angle": 0, + "content": "Despite the success of score distillation in optimization-based text-to-3D generation [40, 59, 72], its application to learning-based text-to-3D frameworks is rather limited because of the unstable training or unsatisfactory results. We argue that the primary challenge lies in how to efficiently and effectively leverage the pretrained 2D diffusion prior to represent the distribution of images rendered by the 3D generator. For example, SDS [40] forces the rendered images to adhere to the Dirac distribution, which causes numerical instability in 3D generator training [33, 65]. Variational Score Distillation (VSD) [59] finetunes the 2D diffusion prior for distribution alignment via minimizing the noise prediction error. However, the finetuning changes the pretrained diffusion network and hurts its comprehension capability to numerous text prompts, leading to mode collapse when the size of text prompts is extended." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "To address the above mentioned issues, we propose Asynchronous Score Distillation (ASD). Like VSD, ASD aims to minimize the noise prediction error. Different from VSD, ASD does not finetune the pretrained 2D diffusion network; instead, it achieves the goal by shifting the diffusion timestep. This is based on the observation that diffusion networks will have smaller noise prediction errors" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.638, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "ScaleDreamer" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.345 + ], + "angle": 0, + "content": "in earlier timesteps [67]; therefore, we can shift the timestep to an earlier step to achieve a similar goal to VSD, i.e., reducing the noise prediction error. In this way, the diffusion network can be frozen in training and its strong text comprehension capability can be well-preserved. The shifted timesteps can be well sampled from a pre-defined range for most prompts. To evaluate the performance of ASD, we conduct extensive experiments by using two types of generator architectures, i.e. Hyper-iNGP [33] and 3DConv-Net [4], across various prompt corpus sizes. We conduct extensive experiments to evaluate the superiority of ASD to previous methods, including the stable training of 3D generators, the production of high-quality 3D outputs, the high content fidelity to input prompts, as well as its scalability to larger corpus sizes, e.g., 100k prompts. We also show that ASD can work with other 2D diffusion models such as MVDream [48], and can be used to train more 3D generators such as Triplane-Transformer [17]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.369, + 0.43, + 0.385 + ], + "angle": 0, + "content": "2 Literature Review" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.402, + 0.544, + 0.417 + ], + "angle": 0, + "content": "2.1 Text-to-3D with Score Distillation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.788, + 0.568 + ], + "angle": 0, + "content": "Text-to-3D takes text description, a.k.a. text prompt \\( y \\), as input, and outputs 3D representation \\( \\theta \\) that renders high-fidelity images at any camera view \\( \\pi \\). Thanks to the powerful text-to-image diffusion models [32,42,44,48,73], we can optimize \\( \\theta \\) to align with \\( y \\) by computing the objective \\( \\mathcal{L}(\\boldsymbol{x},y) \\) on the rendered image \\( \\boldsymbol{x} = g(\\theta ,\\pi) \\) from camera view \\( \\pi \\). Through differential rendering, \\( \\theta \\) can be updated with the gradient \\( \\nabla_{\\theta}\\mathcal{L}(\\theta ,y) = \\frac{\\partial\\mathcal{L}(\\boldsymbol{x},y)}{\\partial\\boldsymbol{x}}\\frac{\\partial\\boldsymbol{x}}{\\partial\\theta} \\). This technique is generally termed as score distillation. Unlike data-driven techniques [6,21,31,43,54], score distillation approaches [8, 19, 24, 29, 40, 53, 59, 72] can produce high-quality 3D content without the need for 3D training datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.788, + 0.674 + ], + "angle": 0, + "content": "Prompt-Specific Text-to-3D. Existing score distillation methods [40, 59, 72] were originally developed to output a single 3D result \\(\\theta\\) for a single text prompt \\(y\\) via online optimization: \\(\\min_{\\theta} \\mathbb{E}_{\\pi, \\boldsymbol{x} = g(\\theta, \\pi)}[\\mathcal{L}(\\boldsymbol{x}, y)]\\). The utilized 3D representations, e.g., NeRF [39, 40], DmTet [47, 74], and 3D Gaussian [20, 30, 51, 53, 57, 70], are not designed to render scenes from varying text prompts. Therefore, the optimization has to be conducted again for newly provided text prompts. The optimization process typically costs tens of minutes to hours." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Prompt-Amortized Text-to-3D. To mitigate the computational costs in prompt-specific methods, recent studies [25, 33, 41, 65] have attempted to use score distillation to train a text-to-3D generator \\(\\theta = \\mathcal{G}(y)\\), aiming to generate multiple 3D representations from a set of text prompts \\(S_{y} = \\{y\\}\\). These methods can generate 3D results from queried text prompt in seconds. As proposed by ATT3D [33], the 3D generator training is performed by minimizing \\(\\min_{\\mathcal{G}} \\mathbb{E}_{\\pi, y \\in S_{y}, x = g(\\mathcal{G}(y), \\pi)}[\\mathcal{L}(\\boldsymbol{x}, y)]\\) over all text prompts. Unlike data-driven approaches [17, 52, 66], score distillation bypasses the scarcity of text-3D data pairs because the 2D diffusion prior can offer the guidance to align the 3D output with the input text prompt. However, its application is currently restricted to training the 3D generator within a limited range of text prompts." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.348, + 0.128 + ], + "angle": 0, + "content": "Z.Ma et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.611, + 0.162 + ], + "angle": 0, + "content": "2.2 Representative Score Distillation Methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.173, + 0.788, + 0.32 + ], + "angle": 0, + "content": "Denote by \\(\\phi\\) the 2D diffusion prior [44,48] and by \\(p^{\\phi}(\\boldsymbol{x} \\mid y)\\) the text-conditioned image distribution embedded within \\(\\phi\\), the objectives of most existing score distillation methods can be generally concluded as minimizing the objective \\(\\mathcal{L}(\\theta, y) = \\mathbb{E}_{\\pi, t, \\epsilon, \\boldsymbol{x} = g(\\theta, \\pi)}\\left[\\omega(t) D_{\\mathrm{KL}}\\left(q_t^\\theta(\\boldsymbol{x}_t \\mid \\pi) \\| p_t^\\phi(\\boldsymbol{x}_t \\mid y^\\pi)\\right)\\right]\\), where \\(D_{\\mathrm{KL}}\\) denotes KL divergence, \\(q_t^\\theta(\\boldsymbol{x}_t \\mid \\pi)\\) denotes the distribution of images \\(\\boldsymbol{x}\\) rendered at camera view \\(\\pi\\) at diffusion timestep \\(t\\) [15], and the same for \\(p_t^\\phi(\\boldsymbol{x}_t \\mid y)\\). \\(\\omega(t)\\) is a timestep-dependent weight [40]. \\(y^\\pi\\) denotes the view-dependent strategy [44] or view-awareness [42, 48] to prompt the different camera views [40]. To minimize this objective, the gradient w.r.t. \\(\\theta\\) can be calculated as per [59]:" + }, + { + "type": "equation", + "bbox": [ + 0.228, + 0.33, + 0.786, + 0.389 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\theta} \\mathcal {L} (\\theta , y) = \\mathbb {E} _ {\\pi , t, \\epsilon} \\left[ \\omega (t) \\left(\\underbrace {- \\sigma_ {t} \\nabla_ {\\boldsymbol {x} _ {t}} \\log p _ {t} ^ {\\phi} (\\boldsymbol {x} _ {t} \\mid y ^ {\\pi})} _ {\\epsilon_ {\\phi} (\\boldsymbol {x} _ {t}; t, y ^ {\\pi})} - \\underbrace {(- \\sigma_ {t} \\nabla_ {\\boldsymbol {x} _ {t}} \\log q _ {t} ^ {\\theta} (\\boldsymbol {x} _ {t} \\mid \\pi))} _ {\\epsilon_ {\\theta} (\\boldsymbol {x} _ {t}; t, \\pi , y)}\\right) \\frac {\\partial \\boldsymbol {x}}{\\partial \\theta} \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.401, + 0.788, + 0.537 + ], + "angle": 0, + "content": "where the first term \\(-\\sigma_t\\nabla_{\\boldsymbol{x}_t}\\log p_t^\\phi (\\boldsymbol{x}_t\\mid y^\\pi)\\) corresponds to the score function [50] of the desired image distribution, and it can be achieved by predicting the noise \\(\\epsilon \\sim \\mathcal{N}(0,I)\\) in the noisy image \\(\\boldsymbol{x}_t = \\alpha_t\\boldsymbol {x} + \\sigma_t\\boldsymbol{\\epsilon}\\) using the pretrained 2D diffusion model \\(\\epsilon_{\\phi}(\\pmb {x}_t;t,y^\\pi)\\) [44,48]. Existing score distillation methods [40,59,72] mainly differ in how to model \\(-\\sigma_t\\nabla x_t\\log q_t^\\theta (x_t\\mid \\pi)\\), which corresponds to the score function of the distribution of rendered images \\(q^{\\theta}(\\pmb {x}\\mid \\pi)\\). We denote this term in Eq. 1 as \\(\\epsilon_{\\theta}(\\pmb {x}_t;t,\\pi ,y)\\) in the following context, since it represents a diffusion model that corresponds to \\(\\theta\\). A summary of the objectives of major score distillation methods is shown in Tab. 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.789, + 0.844 + ], + "angle": 0, + "content": "The objective of Score Distillation Sampling (SDS) [40] is \\(\\nabla_{\\theta}\\mathcal{L}_{\\mathrm{SDS}}(\\theta ,y)\\triangleq\\) \\(\\mathbb{E}_{\\pi ,t,\\epsilon}\\left[\\omega (t)(\\pmb{\\epsilon}_{\\phi}(\\pmb{x}_t;t,y^\\pi) - \\pmb {\\epsilon})\\frac{\\partial\\pmb{x}}{\\partial\\theta}\\right]\\), which approximates the term \\(\\pmb{\\epsilon}_{\\theta}(\\pmb {x}_t;t,\\pi ,y)\\) in Eq. 1 as the ground-truth noise \\(\\pmb{\\epsilon}\\). That is, SDS assumes that \\(q^{\\theta}(\\pmb {x}\\mid \\pi)\\) adheres to a Dirac distribution \\(\\delta (\\pmb {x} - g(\\theta ,\\pi))\\) [59], which is characterized by a non-zero density at the singular point of \\(\\pmb {x} = g(\\theta ,\\pi)\\) and zero density everywhere else. However, updating \\(\\theta\\) under the Dirac distribution might be troublesome [59]. We may need to set the CFG (Classifier Free Guidance) [16] as high as 100 for model convergence, which will produce excessively large gradients and lead to unstable optimization. This problem is alleviated by Classifier Score Distillation (CSD) [72], which uses the classifier component [16] in SDS as the objective: \\(\\nabla_{\\theta}\\mathcal{L}_{\\mathrm{CSD}}(\\theta ,y)\\triangleq \\mathbb{E}_{\\pi ,t,\\epsilon}\\left[\\omega (t)(\\pmb{\\epsilon}_{\\phi}(\\pmb {x}_t;t,y^\\pi) - \\pmb{\\epsilon}_{\\phi}(\\pmb {x}_t;t))\\frac{\\partial\\pmb{x}}{\\partial\\theta}\\right]\\). CSD can be regraded as straightforwardly using the unconditional term of the diffusion prior \\(\\pmb{\\epsilon}_{\\phi}(\\pmb{x}_t;t)\\) to represent \\(\\pmb{\\epsilon}_{\\theta}(\\pmb {x}_t;t,\\pi ,y)\\) in Eq. 1. Unfortunately, in the case of prompt-amortized training, this term may not provide effective gradient, because \\(\\pmb{\\epsilon}_{\\phi}(\\pmb {x}_t;t)\\) is unconditional to the provided text-prompts. In contrast, Variational Score Distillation (VSD) [59] models \\(\\pmb{\\epsilon}_{\\theta}(\\pmb {x}_t;t,\\pi ,y)\\) with another text-aware diffusion model \\(\\pmb{\\epsilon}_{\\phi^{\\prime}}(\\mathbf{x}_t;t,\\pi ,y)\\), leading to \\(\\nabla_{\\theta}\\mathcal{L}_{\\mathrm{VSD}}(\\theta ,y)\\triangleq\\) \\(\\mathbb{E}_{\\pi ,t,\\epsilon}\\left[\\omega (t)(\\pmb{\\epsilon}_{\\phi}(\\pmb {x}_t;t,y^\\pi) - \\pmb{\\epsilon}_{\\phi^{\\prime}}(\\pmb {x}_t;t,\\pi ,y))\\frac{\\partial\\pmb{x}}{\\partial\\theta}\\right]\\), where \\(\\pmb{\\epsilon}_{\\phi^{\\prime}}(\\mathbf{x}_t;t,\\pi ,y)\\) is achieved by finetuning the pretrained 2D diffusion prior \\(\\pmb{\\epsilon}_{\\phi}(\\pmb {x}_t;t,y^\\pi)\\) to align with the rendered image distribution \\(q^{\\theta}(\\pmb {x}\\mid \\pi)\\) via parameter efficient adaptation [18]. In" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.638, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ScaleDreamer" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "table", + "bbox": [ + 0.286, + 0.144, + 0.719, + 0.219 + ], + "angle": 0, + "content": "
MethodGradient of L(x,y) w.r.t. x = g(θ,π)
SDS [40]E_t,ε[ω(t) (εφ(xt; t, yπ) - ε)]
CSD [72]E_t,ε[ω(t) (εφ(xt; t, yπ) - εφ(xt, t))]
VSD [59]E_t,ε[ω(t) (εφ(xt; t, yπ) - εφ'(xt; t, π, y))]
ASD (Ours)E_t,ε[ω(t) (εφ(xt; t, yπ) - εφ(xt+Δt; t+Δt, yπ))]
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.22, + 0.784, + 0.247 + ], + "angle": 0, + "content": "Table 1: Objectives of representative score distillation methods. ASD introduces \\(\\Delta t\\) alongside \\(t\\) to align with the rendered image distribution \\(q^{\\theta}(\\boldsymbol{x} \\mid \\pi)\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.255, + 0.784, + 0.287 + ], + "angle": 0, + "content": "practice, this is conducted by alternatively optimizing \\(\\theta\\) and finetuning \\(\\phi\\) with the noise prediction objective \\(\\| \\pmb{\\epsilon}_{\\phi}(\\pmb{x}_t;t,y) - \\pmb{\\epsilon}\\| _2^2\\) [15] such that:" + }, + { + "type": "equation", + "bbox": [ + 0.289, + 0.296, + 0.785, + 0.314 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} _ {\\pi , t, \\epsilon} \\left[ \\| \\epsilon_ {\\phi J} \\left(\\boldsymbol {x} _ {t}; t, \\pi , y\\right) - \\epsilon \\| _ {2} ^ {2} \\right] \\leq \\mathbb {E} _ {\\pi , t, \\epsilon} \\left[ \\| \\epsilon_ {\\phi} \\left(\\boldsymbol {x} _ {t}; t, y ^ {\\pi}\\right) - \\epsilon \\| _ {2} ^ {2} \\right]. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.324, + 0.784, + 0.354 + ], + "angle": 0, + "content": "The above equation reveals that a better alignment with the distribution of \\( q^{\\theta}(\\pmb{x} \\mid \\pi) \\) can be achieved by a more accurate noise prediction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.355, + 0.785, + 0.49 + ], + "angle": 0, + "content": "While VSD achieves state-of-the-art results in prompt-specific text-to-3D [14, 59], it changes the diffusion prior's parameters by alternately optimizing \\(\\theta\\) and finetuning \\(\\phi\\). This forms a bi-level optimization, known to be problematic in generative adversarial training [55], and may be troublesome for training prompt-amortized text-to-3D models, because the change of pre-trained diffusion model might impairs its comprehension capability on a wide range of text-prompts. In specific, the pre-trained 2D diffusion model may have to sacrifice its generation capability in order to align with the distribution of rendered images, making it fail to produce good gradient for training the 3D generator." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.513, + 0.631, + 0.531 + ], + "angle": 0, + "content": "3 Asynchronous Score Distillation (ASD)" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.544, + 0.409, + 0.559 + ], + "angle": 0, + "content": "3.1 Objective of ASD" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.785, + 0.689 + ], + "angle": 0, + "content": "From the above discussions in Sec. 2.2, it can be seen that one key issue in VSD is to minimize the noise prediction error so that the model output can be aligned with the desired distribution of rendered images. VSD achieves this goal via finetuning the pre-trained 2D diffusion model, which however sacrifices its comprehension capability on text prompts. One interesting question is: can we minimize the noise prediction error without changing the pre-trained diffusion network weights? Fortunately, we find that this is possible and in this section we present a new objective function to achieve this goal." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.786, + 0.841 + ], + "angle": 0, + "content": "Recall that diffusion models solve the stochastic differential equation [50] via reversing the noise added along different stages, a.k.a. diffusion timestep \\( t \\in \\{T_{\\max}, \\dots, T_{\\min}\\} \\) via \\( \\boldsymbol{x}_t = \\alpha_t \\boldsymbol{x} + \\sigma_t \\boldsymbol{\\epsilon} \\) [15]. The influence of the noise \\( \\epsilon \\sim \\mathcal{N}(0, I) \\) on the image \\( \\boldsymbol{x} \\) is incrementally reduced as the process progresses from the initial timestep \\( T_{\\max} \\) to the final timestep \\( T_{\\min} \\), which is controlled by the scalars \\( \\alpha_t \\) and \\( \\sigma_t \\). Consequently, the diffusion model's noise prediction accuracy will vary with the timestep \\( t \\), at which the identical noise \\( \\epsilon \\) is added. To evaluate this, we consider a diffusion model with fixed image \\( \\boldsymbol{x} \\), noise \\( \\epsilon \\) and condition \\( y \\), but varied timestep \\( t \\). We denote such a diffusion model as \\( \\epsilon(t) \\) and explore how its prediction error, denoted by \\( e(t) = \\| \\epsilon(t) - \\epsilon \\|_2^2 \\), changes with \\( t \\)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.348, + 0.128 + ], + "angle": 0, + "content": "Z.Ma et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.375 + ], + "angle": 0, + "content": "The model \\(\\epsilon(t)\\) can be a pre-trained 2D diffusion model (such as Stable Diffusion [44]). We denote by \\(\\epsilon_{PT}(t)\\) such a model, and investigate the behaviour of its noise prediction error, denoted by \\(e_{PT}(t)\\). In Fig. 1, we plot the curve (i.e., the blue colored curve) of \\(e_{PT}(t)\\) versus \\(t\\). We use a corpus with 15 text prompts from Magic3D [40] to draw this curve. For each prompt \\(y\\), we generate 16 images with VSD [59]. Then for each image \\(\\pmb{x}\\), we apply one instance of Gaussian noise \\(\\pmb{\\epsilon}\\) and conduct a single diffusion step with 100 distinct timesteps. The average noise reconstruction error is then calculated for these timesteps across all prompts and images. We can see from the curve of \\(e_{PT}(t)\\) that earlier diffusion timesteps (e.g., timestep 600) will have lower noise prediction error than later timesteps (e.g., timestep 200). Such a trend holds for almost every image sample \\(\\pmb{x}\\) and noise sample \\(\\pmb{\\epsilon}\\) because the well-trained diffusion model is frozen in our case. Since the noise prediction error declines from \\(T_{\\mathrm{min}}\\) (i.e., late diffusion timestep) to \\(T_{\\mathrm{max}}\\) (i.e., early diffusion timestep), we can conclude that for a given timestep \\(t\\) and a timestep shift \\(0 \\leq \\Delta t \\leq T_{\\mathrm{max}} - t\\), the following inequality holds:" + }, + { + "type": "equation", + "bbox": [ + 0.262, + 0.393, + 0.786, + 0.412 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} _ {\\pi , t, \\epsilon} \\left[ \\| \\epsilon_ {\\phi} \\left(\\boldsymbol {x} _ {t + \\Delta t}; t + \\Delta t, y ^ {\\pi}\\right) - \\epsilon \\| _ {2} ^ {2} \\right] \\leq \\mathbb {E} _ {\\pi , t, \\epsilon} \\left[ \\| \\epsilon_ {\\phi} \\left(\\boldsymbol {x} _ {t}; t, y ^ {\\pi}\\right) - \\epsilon \\| _ {2} ^ {2} \\right], \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.786, + 0.461 + ], + "angle": 0, + "content": "which implies that more accurate noise predictions can be achieved at earlier diffusion timesteps." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.464, + 0.788, + 0.601 + ], + "angle": 0, + "content": "The above property of diffusion models has also been observed by Yang et al. [68], who indicated that as the timestep shifts from \\( T_{\\mathrm{max}} \\) towards \\( T_{\\mathrm{min}} \\), the variance in noise prediction increases, as evidenced by the rising Lipschitz constants, which suggests an increased instability in noise prediction and larger noise prediction errors. Such a behavior can be observed in both \\( \\epsilon \\)-prediction and \\( \\pmb{v} \\)-prediction models, as well as in 2D and 3D diffusion models (please refer to supplementary material for details). This can be intuitively explained as follows. When \\( t \\to T_{\\mathrm{max}} \\), \\( \\pmb{x}_t = \\alpha_t \\pmb{x} + \\sigma_t \\pmb{\\epsilon} \\to \\pmb{\\epsilon} \\), then it is easier to achieve \\( \\epsilon_{\\phi}(\\pmb{x}_t; t, y^{\\pi}) \\approx \\epsilon \\) because the model can manage to copy the input as the output." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.604, + 0.788, + 0.71 + ], + "angle": 0, + "content": "The similarity between Eq. 3 and the fine-tuning objective of VSD in Eq. 2 inspires us to investigate whether simply shifting earlier the timestep could fulfill the fine-tuning requirements of VSD without modifying the pre-trained 2D diffusion network parameters. Specifically, we employ the pretrained 2D diffusion model with shifted timestep to approximate the diffusion model of rendered images in Eq. 1 as \\(\\pmb{\\epsilon}_{\\theta}(\\pmb{x}_t; t, \\pi, y) \\triangleq \\pmb{\\epsilon}_{\\phi}(\\pmb{x}_{t + \\Delta t}; t + \\Delta t, y^{\\pi})\\), resulting in the following Asynchronous Score Distillation (ASD) objective function:" + }, + { + "type": "equation", + "bbox": [ + 0.235, + 0.729, + 0.787, + 0.763 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\theta} \\mathcal {L} _ {\\mathrm {A S D}} (\\theta , y) \\triangleq \\mathbb {E} _ {\\pi , t, \\epsilon} \\left[ \\omega (t) \\left(\\boldsymbol {\\epsilon} _ {\\phi} \\left(\\boldsymbol {x} _ {t}; t, y ^ {\\pi}\\right) - \\boldsymbol {\\epsilon} _ {\\phi} \\left(\\boldsymbol {x} _ {t + \\Delta t}; t + \\Delta t, y ^ {\\pi}\\right)\\right) \\frac {\\partial \\boldsymbol {x}}{\\partial \\theta} \\right]. \\quad (4)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We can see that rather than iteratively fine-tuning the diffusion network as in VSD, ASD achieves similar goal by shifting the timestep \\( t \\) with an interval \\( \\Delta t \\) in each step, which is much more efficient. One key variable introduced in ASD is the timestep shift \\( \\Delta t \\), which will be discussed in the next subsection." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.638, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ScaleDreamer" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.279, + 0.147, + 0.722, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.353, + 0.788, + 0.409 + ], + "angle": 0, + "content": "Fig. 1: Illustration of the noise prediction error of the pre-trained 2D diffusion model \\(\\epsilon_{PT}(t)\\) and that of the fine-tuned 2D diffusion model \\(\\epsilon_{FT}(t)\\). We can see that the curve of \\(e_{FT}(t)\\) is positioned under that of \\(e_{PT}(t)\\), and we can shift the timestep of \\(\\epsilon_{PT}(t)\\) to \\(\\epsilon_{PT}(t + \\Delta t)\\) to approximate the noise prediction error of \\(\\epsilon_{FT}(t)\\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.419, + 0.538, + 0.433 + ], + "angle": 0, + "content": "3.2 The Setting of Timestep Shift \\(\\Delta t\\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.447, + 0.788, + 0.538 + ], + "angle": 0, + "content": "Before discussing how to set the timestep shift \\(\\Delta t\\), let's plot another curve, i.e., the noise prediction error of \\(\\epsilon_{\\theta}(\\boldsymbol{x}_t; t, \\pi, y)\\) w.r.t. timestep \\(t\\). Actually, in the process of generating \\(\\boldsymbol{x}\\) with VSD, we will have the fine-tuned model \\(\\epsilon_{\\phi'}(\\boldsymbol{x}_t; t, \\pi, y)\\) as the by-product, which is used to represent \\(\\epsilon_{\\theta}(\\boldsymbol{x}_t; t, \\pi, y)\\) in Eq. 1. Therefore, with fixed \\(\\boldsymbol{x}\\), \\(\\epsilon\\) and \\(y\\), the noise prediction error of the fine-tuned diffusion model, denoted by \\(\\epsilon_{FT}(t)\\), can be calculated as \\(e_{FT}(t) = \\| \\epsilon_{\\phi'}(t) - \\epsilon \\|_2^2\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.788, + 0.689 + ], + "angle": 0, + "content": "The curve of \\( e_{FT}(t) \\) w.r.t. \\( t \\) (i.e., the yellow curve) is plotted in Fig. 1 by using the same data as in plotting \\( e_{PT}(t) \\). We can see that the curve of \\( e_{FT}(t) \\) is positioned under \\( e_{PT}(t) \\) because \\( e_{FT}(t) \\) is obtained by the fine-tuned diffusion model \\( \\epsilon_{FT} \\). However, as mentioned in Sec. 2.2, this fine-tuning changes the weights of pre-trained diffusion model and might damage its ability in comprehending text-image pairs. Therefore, we propose to fix the pre-trained model \\( \\epsilon_{PT}(t) \\) but shift it to \\( \\epsilon_{PT}(t + \\Delta t) \\) to approximate the desired \\( \\epsilon_{FT}(t) \\). Referring to Fig. 1, we could shift \\( \\epsilon_{PT}(t) \\) to an earlier timestep to achieve this goal. For example, at timestep \\( t_0 \\) and with a time shift \\( \\Delta t_0 > 0 \\), we can use \\( \\epsilon_{PT}(t_0 + \\Delta t_0) \\) to approximate the noise prediction error of \\( \\epsilon_{FT}(t_0) \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.788, + 0.842 + ], + "angle": 0, + "content": "On the other hand, the magnitude of \\(\\Delta t\\) will vary with \\(t\\). Let's come to another timestep \\(t_1\\) in Fig. 1, where \\(t_1\\) is earlier than \\(t_0\\). Because the decreasing speeds of both \\(e_{PT}\\) and \\(e_{FT}\\) will be reduced with \\(t\\) going to \\(T_{\\mathrm{max}}\\), the magnitude of \\(\\Delta t_1\\) will be increased to approximate \\(e_{FT}(t_1)\\). In other words, the magnitude of \\(\\Delta t\\) should grow when \\(t\\) goes from \\(T_{\\mathrm{min}}\\) to \\(T_{\\mathrm{max}}\\). We heuristically set this relationship as \\(\\Delta t = \\eta (t - T_{\\mathrm{min}})\\), where \\(\\eta \\in [0,1]\\) is a hyper-parameter that controls the length of shift range. Finally, it should be pointed out that the curves in Fig. 1 will vary a little for different training iterations, rendered images \\(\\mathbf{x}\\) and text prompts \\(y\\). Therefore, \\(\\Delta t\\) should fall into some range \\(S(t)\\). In practice, we set \\(\\Delta t \\sim S(t) = \\mathcal{U}[0,\\eta (t - T_{\\mathrm{min}})]\\), which follows a uniform distribution within 0" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.348, + 0.127 + ], + "angle": 0, + "content": "Z.Ma et al." + }, + { + "type": "image", + "bbox": [ + 0.251, + 0.148, + 0.754, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.289, + 0.785, + 0.317 + ], + "angle": 0, + "content": "Fig. 2: Left and middle: 2D toy examples by SDS [40], CSD [72], VSD [59] and our proposed ASD. Right: Gradient norms generated by different methods." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.321, + 0.605, + 0.336 + ], + "angle": 0, + "content": "Algorithm 1 Asynchronous Score Distillation (ASD)" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.339, + 0.777, + 0.352 + ], + "angle": 0, + "content": "Input: 3D representation \\(\\theta\\); Text prompt \\(y\\); Hyperparameter \\(\\eta\\); 2D diffusion prior \\(\\epsilon_{\\phi}\\)" + }, + { + "type": "code", + "bbox": [ + 0.203, + 0.354, + 0.691, + 0.466 + ], + "angle": 0, + "content": "1 while not converged do \n2 Sample a camera pose \\(\\pi\\) \n3 Render an image \\(\\pmb{x} = g(\\theta, \\pi)\\) \n4 Sample a timestep \\(t \\sim \\mathcal{U}[T_{\\min}, T_{\\max}]\\), Gaussian noise \\(\\epsilon \\sim \\mathcal{N}(0, I)\\) \n5 Sample a timestep shift \\(\\Delta t \\sim S(t) = \\mathcal{U}[0, \\eta(t - T_{\\min})]\\) \n6 \\(\\pmb{x}_t \\gets \\alpha_t \\pmb{x} + \\sigma_t \\pmb{\\epsilon}\\), \\(\\pmb{x}_{t + \\Delta t} \\gets \\alpha_{t + \\Delta t} \\pmb{x} + \\sigma_{t + \\Delta t} \\pmb{\\epsilon}\\) \n7 Update \\(\\theta\\) with \\(\\Delta \\theta \\gets \\omega(t) (\\pmb{\\epsilon}_{\\phi}(\\pmb{x}_t; t, y^{\\pi}) - \\pmb{\\epsilon}_{\\phi}(\\pmb{x}_{t + \\Delta t}; t + \\Delta t, y^{\\pi})) \\frac{\\partial \\pmb{x}}{\\partial \\theta}\\) \n8 end" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.503, + 0.785, + 0.534 + ], + "angle": 0, + "content": "and \\(\\eta(t - T_{\\min})\\). The pseudo-code of ASD is summarized in Alg. 1, which can be applied to both prompt-specific and prompt-amortized text-to-3D tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.536, + 0.787, + 0.718 + ], + "angle": 0, + "content": "2D toy experiments. To verify the proposed timestep shift strategy, we follow the paradigm in [59] to test SDS, CSD, VSD and our ASD on 2D toy examples. The left column of Fig. 2 shows the results of SDS, CSD, VSD, and the middle column shows the results of ASD with different sampling strategies of \\(\\Delta t\\). One can see that the proposed sampling strategy \\(\\Delta t \\sim S(t) = \\mathcal{U}[0, \\eta(t - T_{\\min})]\\) yields similar results to VSD [59]. Besides, we show the gradient norm produced by these score distillation methods in the right column of Fig. 2. One can see that the range of gradient norm produced by ASD is similar to that of VSD. However, the gradient norm of SDS is more than 10 times larger than ASD and VSD because it needs to set CFG=100 for convergence [40, 59, 72]. Such a large gradient may result in training instability. We append more 2D results in the supplementary material to further validate our proposed sampling strategy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Text-to-3D Synthesis with ASD. As a score distillation method, ASD is open to the selection of 3D generator architectures [4,17,22,33,39]. The general pipeline of ASD for text-to-3D synthesis is shown in Fig. 3. It takes a rendered image as input and diffuses it in two timesteps \\(t\\) and \\(t + \\Delta t\\). The noise prediction difference is used as the gradient to optimize the 3D representation of generator. In this work, in addition to prompt-specific generation, as done in most existing score distillation works [16, 28, 40, 59, 64], we focus more on prompt-amortized text-to-3D and conduct thorough experiments to evaluate the effectiveness of" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.638, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ScaleDreamer" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.246, + 0.144, + 0.756, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.272, + 0.785, + 0.342 + ], + "angle": 0, + "content": "Fig. 3: Overview of Asynchronous Score Distillation (ASD). As illustrated in the left sub-figure, ASD can be employed for prompt-specific generation by optimizing 3D representations for each prompt, as well as for prompt-amortized generation by training a text-to-3D generator. The right sub-figure depicts how ASD uses the difference in noise predictions at asynchronous timesteps to update the 3D network parameters." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.346, + 0.788, + 0.528 + ], + "angle": 0, + "content": "ASD with representative 3D generator architectures, including Hyper-iNGP and 3DConv-net, by using Stable Diffusion as the 2D diffusion model. HyperiNGP is adopted by ATT3D [33], which integrates a prompt-agnostic hash-grid spatial encoding [39] with prompt-conditioned decoding layers to output color and density. 3DConv-net [4] is a 3D generator that maps the provided condition to voxel using 3D convolution. They are chosen in our experiments because they represent two groups of 3D generators that inject the condition through implicit mapping function [3,21] or explicit space representation [7,17,46,49]. The results are shown in Sec. 4. We also conduct experiments to demonstrate that ASD can work with other 2D diffusion models such as MVDream [48], and can be used to train other 3D generator architectures such as Triplane-Transformer [17]. More details of these experiments can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.548, + 0.376, + 0.566 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.577, + 0.446, + 0.592 + ], + "angle": 0, + "content": "4.1 Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.785, + 0.689 + ], + "angle": 0, + "content": "Comparison Methods. We compare ASD with state-of-the-art score distillation methods, including SDS [40], CSD [72] and VSD [59]. We adhere to their official codes for training prompt-amortized text-to-3D networks. For example, the CFG [16] values for SDS, CSD and VSD are configured to 100, 1, and 7.5, respectively. In addition, we compare with existing prompt-amortized method ATT3D [33] (whose code is not released yet) by replicating its reported results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.786, + 0.809 + ], + "angle": 0, + "content": "Implementation Details. We employ VolSDF [69] to render images from the 3D generators. For Stable Diffusion, we employ SD-v2.1-base [1] for all score distillation methods for fair comparison. As configured in VSD [59], we set the CFG value as 7.5 for the pre-trained diffusion model in ASD, and 1 for the diffusion model of rendered images. The resolution of rendered images by HyperiNGP is set to \\(256 \\times 256\\), while that of 3DConv-net and Triplane-Transformer is set to \\(64 \\times 64\\) for GPU memory considerations. Other details are in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.786, + 0.84 + ], + "angle": 0, + "content": "Prompt Corpus. To thoroughly evaluate the capability of ASD in prompt-amortized text-to-3D synthesis, we employ multiple datasets encompassing a" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.348, + 0.128 + ], + "angle": 0, + "content": "Z.Ma et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.237 + ], + "angle": 0, + "content": "range of text prompt quantities. MG15 includes 15 prompts from Magic3D [29]; DF415 comprises 415 prompts from DreamFusion [40]; and AT2520 contains 2520 compositional prompts of animals from ATT3D [33]. DL17k contains 17k compositional prompts of human with daily activities, proposed by [25]. While AT2520 and DL17k provide a larger number of prompts than DF415, the prompt diversity of them is relatively low due to the predefined templates." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.788, + 0.373 + ], + "angle": 0, + "content": "To test ASD's performance with an even larger scale of prompts, we introduce a novel prompt corpus named CP100k. This corpus consists of 100,000 text prompts filtered from the image descriptions collected by Cap3D [34], which was developed to test text-to-image model performance. To the best of our knowledge, it is the first time to evaluate score distillation methods on such a scale of text prompts. Meanwhile, it should be clarified that this work is focused on examining the score distillation performance rather than prompt generalization, so the test prompts share the same distribution as training prompts. More details of the prompt corpus are in supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.374, + 0.788, + 0.48 + ], + "angle": 0, + "content": "Evaluation Metrics. We render 120 surrounding view images as the 3D synthesis result from each prompt. Similar to previous text-to-3D works [25,33, 33,40], we compute the CLIP recall, i.e., the classification accuracy by applying CLIP model to the rendered images to predict the correct text prompt, as one performance metric, denoted by \"R@1\". Additionally, we calculate the CLIP text-image similarity between generated images and input prompts as another metric [54,60], denoted by \"Sim\"." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.5, + 0.419, + 0.515 + ], + "angle": 0, + "content": "4.2 Evaluation Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.522, + 0.788, + 0.704 + ], + "angle": 0, + "content": "Results with iNGP/Hyper-iNGP as 3D Representation. The iNGP [39] architecture is designed for prompt-specific text-to-3D generation. Hyper-iNGP has the same spatial encoding as iNGP except that the weights of the decoding layer depend on the text prompt. To eliminate the effect caused by architecture difference as much as possible, we adopt iNGP for prompt-specific text-to-3D tasks, and Hyper-iNGP for prompt-amortized tasks. Our experiments are carried out on the MG15 dataset. For prompt-specific tasks, we optimize an individual iNGP [39] for each MG15 prompt; while for the prompt-amortized tasks, we train a single Hyper-iNGP [33] across all MG15 prompts. We also compare our results with ATT3D [33], which is among the first to apply Hyper-iNGP to prompt-amortized text-to-3D tasks. ATT3D employs SDS for training and uses soft-shading [40] (denoted as * in Tab. 2) for rendering." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.84 + ], + "angle": 0, + "content": "The qualitative and quantitative results are shown in Fig. 4 and Tab. 2, respectively. We can see that the existing methods suffer from performance decrease when transiting from prompt-specific to prompt-amortized tasks, as evidenced by the decreased CLIP similarity and recall in Tab. 2. It is worth mentioning that training Hyper-net with SDS requires turning on the spectral normalization [38] in the linear layers, otherwise the training will fail due to numerical instability. This observation is consistent with what reported in ATT3D [33]. This is because SDS suffers from large gradient norm (please also refer to Fig. 2 and the discussions therein), which makes Hyper-iNGP hard to" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.638, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ScaleDreamer" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.249, + 0.146, + 0.356, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.146, + 0.458, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.462, + 0.147, + 0.559, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.561, + 0.147, + 0.628, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.63, + 0.147, + 0.759, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.419, + 0.26, + 0.605, + 0.268 + ], + "angle": 0, + "content": "\"A DSLR photo of a peacok on a surfboard\" in MG15" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.271, + 0.785, + 0.314 + ], + "angle": 0, + "content": "Fig. 4: Qualitative comparison on prompt-specific (with iNGP as the 3D representation) and prompt-amortized (with Hyper-iNGP as the 3D generator) text-to-3D results by SDS [40], CSD [72], VSD [59], ATT3D [33] and our ASD methods." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.32, + 0.789, + 0.426 + ], + "angle": 0, + "content": "
ReferenceMethodSim ↑R@1 ↑MethodSim ↑R@1 ↑
ATT3D [33]---Hyper-iNGP* + SDS0.1950.468
DreamFusion [40]iNGP + SDS0.2881.000Hyper-iNGP + SDS0.2570.918
Classifier [16]iNGP + CSD0.2800.936Hyper-iNGP + CSD0.2640.972
ProlificDreamer [59]iNGP + VSD0.2760.932Hyper-iNGP + VSD0.2590.987
OursiNGP + ASD0.2891.000Hyper-iNGP + ASD0.2841.000
" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.426, + 0.784, + 0.468 + ], + "angle": 0, + "content": "Table 2: Quantitative comparison on prompt-specific (with iNGP as the 3D representation) and prompt-amortized (with Hyper-iNGP as the 3D generator) text-to-3D results by SDS [40], CSD [72], VSD [59], ATT3D [33] and our ASD methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.488, + 0.785, + 0.608 + ], + "angle": 0, + "content": "converge. As can be seen in Fig. 4, ATT3D results in wrong geometry by using soft shading and SDS for training. For CSD, we see that it fails to optimize the full geometry, as shown by the shrunk peacock in both prompt-amortized and prompt-amortized results. For VSD, it tends to generate content drifts [48], resulting in repetitive patterns and abnormal geometry. It may fail to generate reasonable contents in both prompt-specific and prompt-amortized tasks. In contrast, our proposed ASD works very stable across the two tasks, yielding not only outstanding quantitative scores but also high quality 3D contents." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Results with 3DConv-net as 3D Generator. The issues of existing score distillation methods either persist or become more pronounced when replacing Hyper-iNGP to 3DConv-net as the 3D generator. We find that training SDS with 3DConv-net always fails within several thousand iterations, even using spectral or other normalization techniques. This issue stems from that deeper network is more sensitive to large gradients [13] caused by SDS. Therefore, we only compare the results of other methods in Fig. 5. We see that CSD outputs acceptable results on AT2520, but its results on DF415, which has more varied prompts, are consistently smaller than anticipated. Such a phenomenon has been observed when Hyper-iNGP is used as the generator, which underlines CSD's inability to reliably guide the 3D generator to produce geometries aligned with the text prompts. As for VSD, it leads to rather abnormal results, failing to match the text prompts. This can be attributed to its fine-tuning of the pre-trained 2D diffusion model, which severely compromises VSD's text-image comprehending ability. In comparison, our proposed ASD, with 3DConv-net as the generator," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.348, + 0.127 + ], + "angle": 0, + "content": "Z.Ma et al." + }, + { + "type": "image", + "bbox": [ + 0.249, + 0.146, + 0.389, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.396, + 0.146, + 0.544, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.562, + 0.147, + 0.744, + 0.156 + ], + "angle": 0, + "content": "Asynchronous Score Distillation (ASD, ours)" + }, + { + "type": "image", + "bbox": [ + 0.549, + 0.157, + 0.758, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.373, + 0.211, + 0.651, + 0.219 + ], + "angle": 0, + "content": "\"A squirrel holding a book wearing a sweater wearing a tophat\" in AT2520" + }, + { + "type": "image", + "bbox": [ + 0.249, + 0.22, + 0.318, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.319, + 0.22, + 0.388, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.39, + 0.22, + 0.47, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.472, + 0.22, + 0.541, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.549, + 0.22, + 0.617, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.618, + 0.22, + 0.687, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.69, + 0.22, + 0.758, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.386, + 0.273, + 0.622, + 0.283 + ], + "angle": 0, + "content": "\"A DSLR photo of a cocker spaniel wearing a crown\" in DF415" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.285, + 0.785, + 0.326 + ], + "angle": 0, + "content": "Fig. 5: Qualitative comparison among CSD [72], VSD [59] and our ASD (with 3DConv-net as generator) on AT2520 and DF415 corpuses. SDS is not compared because it encounters numerical instability in this experiment." + }, + { + "type": "table", + "bbox": [ + 0.277, + 0.327, + 0.729, + 0.43 + ], + "angle": 0, + "content": "
MethodDF415AT2520CP100k
Sim ↑R@1 ↑Sim ↑R@1 ↑Sim ↑R@1 ↑
SDS××××××
CSD0.1760.0620.2790.0370.1950.108
VSD0.1580.0020.1150.0010.1030.000
ASD (ours)0.2370.2760.2850.0580.1990.117
" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.431, + 0.785, + 0.459 + ], + "angle": 0, + "content": "Table 3: Quantitative comparison on prompt-amortized text-to-3D with 3DConv-net as generator. Symbol \\(\\times\\) denotes that the training fails due to numerical instability." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.465, + 0.785, + 0.494 + ], + "angle": 0, + "content": "yields improved outcomes, as evidenced by the visual results in Fig. 5 and the enhanced metric scores in Tab. 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.496, + 0.787, + 0.692 + ], + "angle": 0, + "content": "Scalability. In this section, we evaluate the scalability of competing methods by using as many as 100k prompts in the CP100k dataset with 3DConv-net as the generator. The results are shown in Fig. 6 and Tab. 3. Due to the issue of numerical instability, SDS is not involved in this experiment. We can see that the outcomes of CSD are significantly diminished with uniformly small-sized shapes across all prompts. There is also a lack of variety since most outputs exhibit similar patterns. The results of VSD are also degenerated, displaying almost identical and anomalous outcomes for the text prompts. This resembles the phenomenon of mode collapse often encountered in bi-level optimization [55], which also highlights the importance of fixing the 2D diffusion model when training with such a large number of text prompts. In comparison, ASD is able to produce much higher quality outcomes across the text prompts, showcasing its capability in large-scale training with numerous text prompts as inputs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.692, + 0.787, + 0.768 + ], + "angle": 0, + "content": "In addition to the above comparisons, we also evaluate the effectiveness of ASD when using other 2D diffusion models. We employ MVDream [48] to illustrate ASD's generality to diffusion prior models. Besides, we compare against data-driven methods to prove that ASD is advantageous in tackling diverse input of text prompts. Please see supplementary material for detail." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.787, + 0.389, + 0.803 + ], + "angle": 0, + "content": "4.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.785, + 0.841 + ], + "angle": 0, + "content": "In this section, we perform ablation studies to evaluate the settings of timestep shift \\(\\Delta t\\sim S(t) = \\mathcal{U}[0,\\eta (t - T_{\\mathrm{min}})]\\) from several aspects. The qualitative and" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.638, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ScaleDreamer" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.145, + 0.408, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.146, + 0.582, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.146, + 0.753, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.246, + 0.272, + 0.407, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.264, + 0.581, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.272, + 0.75, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.433, + 0.323, + 0.561, + 0.331 + ], + "angle": 0, + "content": "\"Small wooden brown bear figurine\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.333, + 0.778, + 0.348 + ], + "angle": 0, + "content": "Fig. 6: The scalability comparison with CSD [72] and VSD [59] on CP100k corpus." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.357, + 0.785, + 0.401 + ], + "angle": 0, + "content": "quantitative results are shown in Fig. 7 and Tab. 4, respectively. We provide extensive ablation studies in the supplementary material to show that the proposed strategy is also effective when using MVDream as the diffusion model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.403, + 0.785, + 0.553 + ], + "angle": 0, + "content": "Importance of Timestep Shift. We use \\(\\eta = 0\\) (i.e., no timestep shift) as a baseline to evaluate the necessity of introducing timestep shift \\(\\Delta t\\). From Fig. 7 and Tab. 4, we see that while it can generate plausible results, the model is prone to generating shapes that do not make sense, such as the so-called Janus problem [2]. Examples include a frog with an extra eye, robot face with block-like features, and a peacock with tails at both the front and back. This is because the non-shifted diffusion model will align more with the 2D image distribution, tending to generate redundant contents and unreasonable geometry along the training. By introducing a timestep shift, our proposed ASD demonstrates advantages in achieving more coherent and visually pleasing results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.554, + 0.785, + 0.674 + ], + "angle": 0, + "content": "Range of Timestep Shift. By setting \\(\\eta = 0.2\\), we allow \\(\\Delta t\\) to be sampled from a large range. However, this might not be a good choice. In the extreme case, for any timestep \\(t\\) we can set a large interval \\(\\Delta t\\) such that \\(t + \\Delta t = T_{\\max}\\), then the noise prediction becomes \\(\\epsilon_{\\phi}(\\boldsymbol{x}_t; t, y^\\pi) \\approx \\epsilon\\), so that ASD is degraded to SDS, which cannot perform well under CFG=7.5 [40]. In practice, we find a larger \\(\\eta\\) tends to result 3D contents with larger size and rounded shapes, e.g., the peacock with closer views, or the frog with larger size, as shown in Fig. 7. Therefore, we set \\(\\eta = 0.1\\) in all our experiments." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.676, + 0.785, + 0.751 + ], + "angle": 0, + "content": "Deterministic or Random Shift. If we set \\(\\Delta t = \\eta (t - T_{\\mathrm{min}})\\), it assumes that the diffusion model of rendered images can be approximated by the pretrained one with a fixed and deterministic timestep shift. As shown in Fig. 7 and Tab. 4, it reduces the chance to generate correct geometry and colors. Randomly sampling \\(\\Delta t\\) in a range is more effective, which is adopted in our method." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.776, + 0.519, + 0.792 + ], + "angle": 0, + "content": "5 Conclusion and Limitations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "In this paper, we presented Asynchronous Score Distillation (ASD), a novel score distillation method that can assist 2D diffusion prior in training 3D generators" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.348, + 0.127 + ], + "angle": 0, + "content": "Z.Ma et al." + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.145, + 0.775, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.232, + 0.387, + 0.77, + 0.402 + ], + "angle": 0, + "content": "Fig. 7: The qualitative results of the ablation study on the timestep interval \\(\\Delta t\\)." + }, + { + "type": "table", + "bbox": [ + 0.331, + 0.403, + 0.673, + 0.502 + ], + "angle": 0, + "content": "
ParamSim ↑R@1 ↑
Δt = η(t - Tmin)η = 0.10.2140.178
η = 0.20.2140.180
η = 00.2350.267
Δt ~ U[0, η(t - Tmin)]η = 0.10.2370.276
η = 0.20.2290.237
" + }, + { + "type": "table_caption", + "bbox": [ + 0.219, + 0.503, + 0.781, + 0.517 + ], + "angle": 0, + "content": "Table 4: The quantitative results of the ablation study on the timestep interval \\( \\Delta t \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.531, + 0.788, + 0.652 + ], + "angle": 0, + "content": "with a scalable size of text prompts. By shifting the diffusion timestep to earlier stages, our ASD can effectively predict the noise prediction error to align the diffusion model with the distribution of rendered images, while preserving the superior text comprehension capability of pre-trained models, thus facilitating stable training with high-fidelity generation results. Our extensive experiments revealed that ASD performed consistently well on datasets of various sizes, being able to manage as much as 100k prompts. Though ASD has shown improvements over earlier score distillation approaches, there remain some limitations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.653, + 0.788, + 0.743 + ], + "angle": 0, + "content": "For man-made objects that have very regular shapes, such as chairs or airplanes, the performance of our model will lag behind those data-driven methods, which benefit from an abundance of relevant data. We foresee opportunities to combine the advantages of data-driven and score distillation methodologies to improve text-to-3D capabilities in a more comprehensive manner in the future research." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.773, + 0.395, + 0.79 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "This work is supported in part by the Beijing Science and Technology Plan Project Z231100005923033, and the InnoHK program." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.638, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ScaleDreamer" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.18, + 0.784, + 0.206 + ], + "angle": 0, + "content": "1. Stable-diffusion-v2.1-base. https://huggingface.co/stabilityai/stable-diffusion-2-1-base" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.209, + 0.785, + 0.25 + ], + "angle": 0, + "content": "2. Armandpour, M., Zheng, H., Sadeghian, A., Sadeghian, A., Zhou, M.: Re-imagine the negative prompt algorithm: Transform 2d diffusion into 3d, alleviate janus problem and beyond. arXiv preprint arXiv:2304.04968 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.251, + 0.785, + 0.292 + ], + "angle": 0, + "content": "3. Babu, S., Liu, R., Zhou, A., Maire, M., Shakhnarovich, G., Hanocka, R.: Hyperfields: Towards zero-shot generation of nerfs from text. arXiv preprint arXiv:2310.17075 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.293, + 0.785, + 0.334 + ], + "angle": 0, + "content": "4. Bahmani, S., Park, J.J., Paschalidou, D., Yan, X., Wetzstein, G., Guibas, L., Tagliasacchi, A.: Cc3d: Layout-conditioned generation of compositional 3d scenes. arXiv preprint arXiv:2303.12074 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.335, + 0.785, + 0.376 + ], + "angle": 0, + "content": "5. Balaji, Y., Nah, S., Huang, X., Vahdat, A., Song, J., Kreis, K., Aittala, M., Aila, T., Laine, S., Catanzaro, B., et al.: edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.377, + 0.785, + 0.404 + ], + "angle": 0, + "content": "6. Cao, Z., Hong, F., Wu, T., Pan, L., Liu, Z.: Large-vocabulary 3d diffusion model with transformer. arXiv preprint arXiv:2309.07920 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.406, + 0.785, + 0.46 + ], + "angle": 0, + "content": "7. Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., De Mello, S., Gallo, O., Guibas, L.J., Tremblay, J., Khamis, S., et al.: Efficient geometry-aware 3d generative adversarial networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16123-16133 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.461, + 0.785, + 0.502 + ], + "angle": 0, + "content": "8. Chen, R., Chen, Y., Jiao, N., Jia, K.: *Fantasia3d: Disentangling geometry* and appearance for high-quality text-to-3d content creation. arXiv preprint arXiv:2303.13873 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.504, + 0.785, + 0.53 + ], + "angle": 0, + "content": "9. Chen, Z., Wang, F., Liu, H.: Text-to-3d using gaussian splatting. arXiv preprint arXiv:2309.16585 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.532, + 0.785, + 0.573 + ], + "angle": 0, + "content": "10. Deitke, M., Liu, R., Wallingford, M., Ngo, H., Michel, O., Kusupati, A., Fan, A., Laforte, C., Voleti, V., Gadre, S.Y., et al.: Objaverse-xl: A universe of \\(10\\mathrm{m} + 3\\mathrm{d}\\) objects. arXiv preprint arXiv:2307.05663 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.574, + 0.785, + 0.614 + ], + "angle": 0, + "content": "11. Ding, L., Dong, S., Huang, Z., Wang, Z., Zhang, Y., Gong, K., Xu, D., Xue, T.: Text-to-3d generation with bidirectional diffusion using both 2d and 3d priors. arXiv preprint arXiv:2312.04963 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.616, + 0.785, + 0.657 + ], + "angle": 0, + "content": "12. Guo, P., Hao, H., Caccavale, A., Ren, Z., Zhang, E., Shan, Q., Sankar, A., Schwing, A.G., Colburn, A., Ma, F.: Stabledreamer: Taming noisy score distillation sampling for text-to-3d. arXiv preprint arXiv:2312.02189 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.659, + 0.785, + 0.699 + ], + "angle": 0, + "content": "13. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.7, + 0.785, + 0.741 + ], + "angle": 0, + "content": "14. He, Y., Bai, Y., Lin, M., Zhao, W., Hu, Y., Sheng, J., Yi, R., Li, J., Liu, Y.J.: T\\(^3\\)bench: Benchmarking current progress in text-to-3d generation. arXiv preprint arXiv:2310.02977 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.742, + 0.785, + 0.77 + ], + "angle": 0, + "content": "15. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.771, + 0.785, + 0.798 + ], + "angle": 0, + "content": "16. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "17. Hong, Y., Zhang, K., Gu, J., Bi, S., Zhou, Y., Liu, D., Liu, F., Sunkavalli, K., Bui, T., Tan, H.: Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.18, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.127 + ], + "angle": 0, + "content": "Z.Ma et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "18. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.232 + ], + "angle": 0, + "content": "19. Huang, Y., Wang, J., Shi, Y., Qi, X., Zha, Z.J., Zhang, L.: Dreamtime: An improved optimization strategy for text-to-3d content creation. arXiv preprint arXiv:2306.12422 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.233, + 0.785, + 0.26 + ], + "angle": 0, + "content": "20. Jiang, L., Wang, L.: Brightdreamer: Generic 3d gaussian generative framework for fast text-to-3d synthesis. arXiv preprint arXiv:2403.11273 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.261, + 0.785, + 0.289 + ], + "angle": 0, + "content": "21. Jun, H., Nichol, A.: Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.29, + 0.785, + 0.331 + ], + "angle": 0, + "content": "22. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG) 42(4), 1-14 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.332, + 0.733, + 0.346 + ], + "angle": 0, + "content": "23. Koster, R.: Theory of fun for game design. \"O'Reilly Media, Inc.\" (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.347, + 0.785, + 0.374 + ], + "angle": 0, + "content": "24. Lee, K., Sohn, K., Shin, J.: Dreamflow: High-quality text-to-3d generation by approximating probability flow. arXiv preprint arXiv:2403.14966 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.375, + 0.785, + 0.403 + ], + "angle": 0, + "content": "25. Li, M., Zhou, P., Liu, J.W., Keppo, J., Lin, M., Yan, S., Xu, X.: Instant3d: Instant text-to-3d generation. arXiv preprint arXiv:2311.08403 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.404, + 0.785, + 0.432 + ], + "angle": 0, + "content": "26. Li, W., Chen, R., Chen, X., Tan, P.: Sweetdreamer: Aligning geometric priors in 2d diffusion for consistent text-to-3d. arXiv preprint arXiv:2310.02596 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.433, + 0.785, + 0.473 + ], + "angle": 0, + "content": "27. Li, Z., Chen, Y., Zhao, L., Liu, P.: Mvcontrol: Adding conditional control to multi-view diffusion for controllable text-to-3d generation. arXiv preprint arXiv:2311.14494 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.474, + 0.785, + 0.515 + ], + "angle": 0, + "content": "28. Liang, Y., Yang, X., Lin, J., Li, H., Xu, X., Chen, Y.: Luciddreamer: Towards high-fidelity text-to-3d generation via interval score matching. arXiv preprint arXiv:2311.11284 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.517, + 0.785, + 0.572 + ], + "angle": 0, + "content": "29. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3d: High-resolution text-to-3d content creation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 300–309 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.573, + 0.785, + 0.601 + ], + "angle": 0, + "content": "30. Lin, Y., Clark, R., Torr, P.: Dreamampolisher: Towards high-quality text-to-3d generation via geometric diffusion. arXiv preprint arXiv:2403.17237 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.602, + 0.785, + 0.642 + ], + "angle": 0, + "content": "31. Liu, Y.T., Luo, G., Sun, H., Yin, W., Guo, Y.C., Zhang, S.H.: Pi3d: Efficient text-to-3d generation with pseudo-image diffusion. arXiv preprint arXiv:2312.09069 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.643, + 0.785, + 0.685 + ], + "angle": 0, + "content": "32. Liu, Z., Li, Y., Lin, Y., Yu, X., Peng, S., Cao, Y.P., Qi, X., Huang, X., Liang, D., Ouyang, W.: Unidream: Unifying diffusion priors for relightable text-to-3d generation. arXiv preprint arXiv:2312.08754 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.686, + 0.785, + 0.728 + ], + "angle": 0, + "content": "33. Lorraine, J., Xie, K., Zeng, X., Lin, C.H., Takikawa, T., Sharp, N., Lin, T.Y., Liu, M.Y., Fidler, S., Lucas, J.: Att3d: Amortized text-to-3d object synthesis. arXiv preprint arXiv:2306.07349 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.729, + 0.785, + 0.756 + ], + "angle": 0, + "content": "34. Luo, T., Rockwell, C., Lee, H., Johnson, J.: Scalable 3d captioning with pretrained models. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.757, + 0.785, + 0.798 + ], + "angle": 0, + "content": "35. Ma, Y., Fan, Y., Ji, J., Wang, H., Sun, X., Jiang, G., Shu, A., Ji, R.: X-dreamer: Creating high-quality 3d content by bridging the domain gap between text-to-2d and text-to-3d generation. arXiv preprint arXiv:2312.00085 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "36. Mercier, A., Nakhli, R., Reddy, M., Yasarla, R., Cai, H., Porikli, F., Berger, G.: Hexagen3d: Stablediffusion is just one step away from fast and diverse text-to-3d generation. arXiv preprint arXiv:2401.07727 (2024)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.638, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ScaleDreamer" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.203 + ], + "angle": 0, + "content": "37. Metzer, G., Richardson, E., Patashnik, O., Giryes, R., Cohen-Or, D.: Latent-nerf for shape-guided generation of 3d shapes and textures. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12663-12673 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.203, + 0.787, + 0.23 + ], + "angle": 0, + "content": "38. Miyato, T., Kataoka, T., Koyama, M., Yoshida, Y.: Spectral normalization for generative adversarial networks. arXiv preprint arXiv:1802.05957 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.23, + 0.787, + 0.27 + ], + "angle": 0, + "content": "39. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.27, + 0.787, + 0.298 + ], + "angle": 0, + "content": "40. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.298, + 0.787, + 0.338 + ], + "angle": 0, + "content": "41. Qian, G., Cao, J., Siarohin, A., Kant, Y., Wang, C., Vasilkovsky, M., Lee, H.Y., Fang, Y., Skorokhodov, I., Zhuang, P., et al.: Atom: Amortized text-to-mesh using 2d diffusion. arXiv preprint arXiv:2402.00867 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.338, + 0.787, + 0.379 + ], + "angle": 0, + "content": "42. Qiu, L., Chen, G., Gu, X., Zuo, Q., Xu, M., Wu, Y., Yuan, W., Dong, Z., Bo, L., Han, X.: Richdreamer: A generalizable normal-depth diffusion model for detail richness in text-to-3d. arXiv preprint arXiv:2311.16918 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.379, + 0.787, + 0.406 + ], + "angle": 0, + "content": "43. Ren, X., Huang, J., Zeng, X., Museth, K., Fidler, S., Williams, F.: Xcube: Large-scale 3d generative modeling using sparse voxel hierarchies. arXiv preprint (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.406, + 0.787, + 0.446 + ], + "angle": 0, + "content": "44. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.446, + 0.787, + 0.5 + ], + "angle": 0, + "content": "45. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.501, + 0.787, + 0.542 + ], + "angle": 0, + "content": "46. Schwarz, K., Sauer, A., Niemeyer, M., Liao, Y., Geiger, A.: Voxgraf: Fast 3d-aware image synthesis with sparse voxel grids. Advances in Neural Information Processing Systems 35, 33999-34011 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.542, + 0.787, + 0.582 + ], + "angle": 0, + "content": "47. Shen, T., Gao, J., Yin, K., Liu, M.Y., Fidler, S.: Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. Advances in Neural Information Processing Systems 34, 6087-6101 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.582, + 0.787, + 0.61 + ], + "angle": 0, + "content": "48. Shi, Y., Wang, P., Ye, J., Long, M., Li, K., Yang, X.: Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.61, + 0.787, + 0.664 + ], + "angle": 0, + "content": "49. Sitzmann, V., Thies, J., Heide, F., Nießner, M., Wetzstein, G., Zollhofer, M.: Deepvoxels: Learning persistent 3d feature embeddings. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2437-2446 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.664, + 0.787, + 0.704 + ], + "angle": 0, + "content": "50. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.705, + 0.787, + 0.732 + ], + "angle": 0, + "content": "51. Tang, B., Wang, J., Wu, Z., Zhang, L.: Stable score distillation for high-quality 3d generation. arXiv preprint arXiv:2312.09305 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.732, + 0.787, + 0.772 + ], + "angle": 0, + "content": "52. Tang, J., Chen, Z., Chen, X., Wang, T., Zeng, G., Liu, Z.: Lgm: Large multi-view gaussian model for high-resolution 3d content creation. arXiv preprint arXiv:2402.05054 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.772, + 0.787, + 0.8 + ], + "angle": 0, + "content": "53. Tang, J., Ren, J., Zhou, H., Liu, Z., Zeng, G.: Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.8, + 0.787, + 0.84 + ], + "angle": 0, + "content": "54. Tang, Z., Gu, S., Wang, C., Zhang, T., Bao, J., Chen, D., Guo, B.: Volumediffusion: Flexible text-to-3d generation with efficient volumetric encoder. arXiv preprint arXiv:2312.11459 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.348, + 0.127 + ], + "angle": 0, + "content": "Z.Ma et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.188 + ], + "angle": 0, + "content": "55. Thanh-Tung, H., Tran, T.: Catastrophic forgetting and mode collapse in gans. In: 2020 international joint conference on neural networks (ijcnn). pp. 1-10. IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.231 + ], + "angle": 0, + "content": "56. Tsalicoglou, C., Manhardt, F., Tonioni, A., Niemeyer, M., Tombari, F.: Textmesh: Generation of realistic 3d meshes from text prompts. arXiv preprint arXiv:2304.12439 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.259 + ], + "angle": 0, + "content": "57. Vilesov, A., Chari, P., Kadambi, A.: Cg3d: Compositional generation for text-to-3d via gaussian splatting. arXiv preprint arXiv:2311.17907 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.785, + 0.313 + ], + "angle": 0, + "content": "58. Wang, H., Du, X., Li, J., Yeh, R.A., Shakhnarovich, G.: Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12619-12629 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.355 + ], + "angle": 0, + "content": "59. Wang, Z., Lu, C., Wang, Y., Bao, F., Li, C., Su, H., Zhu, J.: Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.785, + 0.397 + ], + "angle": 0, + "content": "60. Wei, Y., Zhang, Y., Ji, Z., Bai, J., Zhang, L., Zuo, W.: Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. arXiv preprint arXiv:2302.13848 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.424 + ], + "angle": 0, + "content": "61. Wohlgenannt, I., Simons, A., Stieglitz, S.: Virtual reality. Business & Information Systems Engineering 62, 455-461 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.426, + 0.785, + 0.452 + ], + "angle": 0, + "content": "62. Wu, R., Sun, L., Ma, Z., Zhang, L.: One-step effective diffusion network for real-world image super-resolution. arXiv preprint arXiv:2406.08177 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.508 + ], + "angle": 0, + "content": "63. Wu, T., Zhang, J., Fu, X., Wang, Y., Ren, J., Pan, L., Wu, W., Yang, L., Wang, J., Qian, C., et al.: Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 803-814 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.785, + 0.55 + ], + "angle": 0, + "content": "64. Wu, Z., Zhou, P., Yi, X., Yuan, X., Zhang, H.: Consistent3d: Towards consistent high-fidelity text-to-3d generation with deterministic sampling prior. arXiv preprint arXiv:2401.09050 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.551, + 0.785, + 0.59 + ], + "angle": 0, + "content": "65. Xie, K., Lorraine, J., Cao, T., Gao, J., Lucas, J., Torralba, A., Fidler, S., Zeng, X.: Latte3d: Large-scale amortized text-to-enhanced3d synthesis. arXiv preprint arXiv:2403.15385 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.785, + 0.632 + ], + "angle": 0, + "content": "66. Xu, Y., Tan, H., Luan, F., Bi, S., Wang, P., Li, J., Shi, Z., Sunkavalli, K., Wetzstein, G., Xu, Z., et al.: Dmv3d: Denoising multi-view diffusion using 3d large reconstruction model. arXiv preprint arXiv:2311.09217 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.674 + ], + "angle": 0, + "content": "67. Yang, Z., Feng, R., Zhang, H., Shen, Y., Zhu, K., Huang, L., Zhang, Y., Liu, Y., Zhao, D., Zhou, J., et al.: Eliminating lipschitz singularities in diffusion models. arXiv preprint arXiv:2306.11251 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.785, + 0.715 + ], + "angle": 0, + "content": "68. Yang, Z., Feng, R., Zhang, H., Shen, Y., Zhu, K., Huang, L., Zhang, Y., Liu, Y., Zhao, D., Zhou, J., et al.: Lipschitz singularities in diffusion models. In: The Twelfth International Conference on Learning Representations (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.785, + 0.743 + ], + "angle": 0, + "content": "69. Yariv, L., Gu, J., Kasten, Y., Lipman, Y.: Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems 34, 4805-4815 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.785, + 0.785 + ], + "angle": 0, + "content": "70. Yi, T., Fang, J., Wu, G., Xie, L., Zhang, X., Liu, W., Tian, Q., Wang, X.: Gaussian dreamer: Fast generation from text to 3d gaussian splatting with point cloud priors. arXiv preprint arXiv:2310.08529 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.785, + 0.84 + ], + "angle": 0, + "content": "71. Yu, X., Xu, M., Zhang, Y., Liu, H., Ye, C., Wu, Y., Yan, Z., Zhu, C., Xiong, Z., Liang, T., et al.: Mvimgnet: A large-scale dataset of multi-view images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9150-9161 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.638, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ScaleDreamer" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.175 + ], + "angle": 0, + "content": "72. Yu, X., Guo, Y.C., Li, Y., Liang, D., Zhang, S.H., Qi, X.: Text-to-3d with classifier score distillation. arXiv preprint arXiv:2310.19415 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.217 + ], + "angle": 0, + "content": "73. Zhao, M., Zhao, C., Liang, X., Li, L., Zhao, Z., Hu, Z., Fan, C., Yu, X.: Efficient-dreamer: High-fidelity and robust 3d creation via orthogonal-view diffusion prior. arXiv preprint arXiv:2308.13223 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.217, + 0.788, + 0.245 + ], + "angle": 0, + "content": "74. Zhao, R., Wang, Z., Wang, Y., Zhou, Z., Zhu, J.: Flexidreamer: Single image-to-3d generation with flexicubes. arXiv preprint arXiv:2404.00987 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.245, + 0.788, + 0.274 + ], + "angle": 0, + "content": "75. Zhou, L., Shih, A., Meng, C., Ermon, S.: Dreampropeller: Supercharge text-to-3d generation with parallel sampling. arXiv preprint arXiv:2311.17082 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.274 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/ScaleDreamer_ Scalable Text-to-3D Synthesis with Asynchronous Score Distillation/0ae3774a-e55b-4154-9080-499a583a8778_origin.pdf b/2024/ScaleDreamer_ Scalable Text-to-3D Synthesis with Asynchronous Score Distillation/0ae3774a-e55b-4154-9080-499a583a8778_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..993ccd498c6e2f20b19c30903d018b62d4e1dd98 --- /dev/null +++ b/2024/ScaleDreamer_ Scalable Text-to-3D Synthesis with Asynchronous Score Distillation/0ae3774a-e55b-4154-9080-499a583a8778_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:35077d595b88204525583373e931941c7a8497601a781402ced38647f429e116 +size 14577817 diff --git a/2024/ScaleDreamer_ Scalable Text-to-3D Synthesis with Asynchronous Score Distillation/full.md b/2024/ScaleDreamer_ Scalable Text-to-3D Synthesis with Asynchronous Score Distillation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3b10c6ba62394af46c37d01dfa85c7f600a12381 --- /dev/null +++ b/2024/ScaleDreamer_ Scalable Text-to-3D Synthesis with Asynchronous Score Distillation/full.md @@ -0,0 +1,338 @@ +# ScaleDreamer: Scalable Text-to-3D Synthesis with Asynchronous Score Distillation + +Zhiyuan Ma $^{1,2}$ , Yuxiang Wei $^{1,5}$ , Yabin Zhang $^{1}$ , + +Xiangyu Zhu $^{3,4}$ , Zhen Lei $^{1,2,3,4^{\star}}$ , and Lei Zhang $^{1^{\star}}$ + +1 The Hong Kong Polytechnic University, PolyU + +$^{2}$ Center for Artificial Intelligence and Robotics, HKISI CAS + +3 State Key Laboratory of Multimodal Artificial Intelligence Systems, CASIA + +$^{4}$ School of Artificial Intelligence, University of Chinese Academy of Sciences, UCAS + +5 Harbin Institute of Technology, HIT + +Abstract. By leveraging the text-to-image diffusion prior, score distillation can synthesize 3D contents without paired text-3D training data. Instead of spending hours of online optimization per text prompt, recent studies have been focused on learning a text-to-3D generative network for amortizing multiple text-3D relations, which can synthesize 3D contents in seconds. However, existing score distillation methods are hard to scale up to a large amount of text prompts due to the difficulties in aligning pretrained diffusion prior with the distribution of rendered images from various text prompts. Current state-of-the-arts such as Variational Score Distillation finetune the pretrained diffusion model to minimize the noise prediction error so as to align the distributions, which are however unstable to train and will impair the model's comprehension capability to numerous text prompts. Based on the observation that the diffusion models tend to have lower noise prediction errors at earlier timesteps, we propose Asynchronous Score Distillation (ASD), which minimizes the noise prediction error by shifting the diffusion timestep to earlier ones. ASD is stable to train and can scale up to $100k$ prompts. It reduces the noise prediction error without changing the weights of pre-trained diffusion model, thus keeping its strong comprehension capability to prompts. We conduct extensive experiments using different text-to-3D architectures, including Hyper-iNGP and 3DConvNet. The results demonstrate ASD's effectiveness in stable 3D generator training, high-quality 3D content synthesis, and its superior prompt-consistency, especially under large prompt corpus. Code is available at https://github.com/theEricMa/ScaleDreamer. + +Keywords: Text-to-3D $\cdot$ Score Distillation $\cdot$ Diffusion Model + +# 1 Introduction + +Text-to-3D aims to generate realistic 3D contents from the given textual descriptions [40], which is particularly useful in many applications such as virtual + +reality [61] and game design [23]. The main challenge of this task, however, lies in how to generate high-quality 3D contents conditioned on the abstract and diverse textual descriptions. Many existing text-to-3D methods [11, 12, 26–29, 32, 35, 37, 40, 42, 58, 59, 75] are optimization-based ones, which distill the guidance from the powerful pretrained text-to-image diffusion models [5, 11, 26, 32, 42, 44, 73] via score distillation [40, 59, 62, 72]. In general, these methods employ the KL divergence to reduce the discrepancy between the distribution of rendered images and the desired image distribution embedded in the 2D diffusion prior, while they differ in how to use the pretrained diffusion prior to model the distribution of rendered images. Extensive efforts have been made to explore prompt-specific optimization of various 3D representations, including implicit radiance fields [40], explicit radiance fields [29, 37, 59], DmTets [56, 74] and 3D Gaussians [9]. Typically, tens of minutes to hours are needed to optimize a single 3D representation for one prompt to achieve the desired result. + +Compared to the aforementioned optimization-based text-to-3D methods, learning-based methods [6, 21, 31, 36, 43, 54, 65] can largely reduce the computational cost by training a text-conditioned 3D generative network. With the availability of 3D object collections [10, 63, 71], a deep network can be trained in a supervised manner so that 3D outputs can be generated in several seconds. Unfortunately, the size of existing text-3D datasets is far from sufficient compared to text-image datasets [45], limiting the text-to-3D generation performance of trained models. Inspired by the optimization-based text-to-3D methods that use pretrained 2D diffusion models, efforts have been made to train text-to-3D networks by using 2D diffusion models as supervisors [33, 41, 65] without using text-3D pairs. For example, a text-conditioned 3D hyper-network is trained in ATT3D [33] via Score Distillation Sampling (SDS) [40]. Nevertheless, this method suffers from numerical instability, which has been observed in subsequent studies [41, 65] that apply SDS to different 3D generator networks. + +Despite the success of score distillation in optimization-based text-to-3D generation [40, 59, 72], its application to learning-based text-to-3D frameworks is rather limited because of the unstable training or unsatisfactory results. We argue that the primary challenge lies in how to efficiently and effectively leverage the pretrained 2D diffusion prior to represent the distribution of images rendered by the 3D generator. For example, SDS [40] forces the rendered images to adhere to the Dirac distribution, which causes numerical instability in 3D generator training [33, 65]. Variational Score Distillation (VSD) [59] finetunes the 2D diffusion prior for distribution alignment via minimizing the noise prediction error. However, the finetuning changes the pretrained diffusion network and hurts its comprehension capability to numerous text prompts, leading to mode collapse when the size of text prompts is extended. + +To address the above mentioned issues, we propose Asynchronous Score Distillation (ASD). Like VSD, ASD aims to minimize the noise prediction error. Different from VSD, ASD does not finetune the pretrained 2D diffusion network; instead, it achieves the goal by shifting the diffusion timestep. This is based on the observation that diffusion networks will have smaller noise prediction errors + +in earlier timesteps [67]; therefore, we can shift the timestep to an earlier step to achieve a similar goal to VSD, i.e., reducing the noise prediction error. In this way, the diffusion network can be frozen in training and its strong text comprehension capability can be well-preserved. The shifted timesteps can be well sampled from a pre-defined range for most prompts. To evaluate the performance of ASD, we conduct extensive experiments by using two types of generator architectures, i.e. Hyper-iNGP [33] and 3DConv-Net [4], across various prompt corpus sizes. We conduct extensive experiments to evaluate the superiority of ASD to previous methods, including the stable training of 3D generators, the production of high-quality 3D outputs, the high content fidelity to input prompts, as well as its scalability to larger corpus sizes, e.g., 100k prompts. We also show that ASD can work with other 2D diffusion models such as MVDream [48], and can be used to train more 3D generators such as Triplane-Transformer [17]. + +# 2 Literature Review + +# 2.1 Text-to-3D with Score Distillation + +Text-to-3D takes text description, a.k.a. text prompt $y$ , as input, and outputs 3D representation $\theta$ that renders high-fidelity images at any camera view $\pi$ . Thanks to the powerful text-to-image diffusion models [32,42,44,48,73], we can optimize $\theta$ to align with $y$ by computing the objective $\mathcal{L}(\boldsymbol{x},y)$ on the rendered image $\boldsymbol{x} = g(\theta ,\pi)$ from camera view $\pi$ . Through differential rendering, $\theta$ can be updated with the gradient $\nabla_{\theta}\mathcal{L}(\theta ,y) = \frac{\partial\mathcal{L}(\boldsymbol{x},y)}{\partial\boldsymbol{x}}\frac{\partial\boldsymbol{x}}{\partial\theta}$ . This technique is generally termed as score distillation. Unlike data-driven techniques [6,21,31,43,54], score distillation approaches [8, 19, 24, 29, 40, 53, 59, 72] can produce high-quality 3D content without the need for 3D training datasets. + +Prompt-Specific Text-to-3D. Existing score distillation methods [40, 59, 72] were originally developed to output a single 3D result $\theta$ for a single text prompt $y$ via online optimization: $\min_{\theta} \mathbb{E}_{\pi, \boldsymbol{x} = g(\theta, \pi)}[\mathcal{L}(\boldsymbol{x}, y)]$ . The utilized 3D representations, e.g., NeRF [39, 40], DmTet [47, 74], and 3D Gaussian [20, 30, 51, 53, 57, 70], are not designed to render scenes from varying text prompts. Therefore, the optimization has to be conducted again for newly provided text prompts. The optimization process typically costs tens of minutes to hours. + +Prompt-Amortized Text-to-3D. To mitigate the computational costs in prompt-specific methods, recent studies [25, 33, 41, 65] have attempted to use score distillation to train a text-to-3D generator $\theta = \mathcal{G}(y)$ , aiming to generate multiple 3D representations from a set of text prompts $S_{y} = \{y\}$ . These methods can generate 3D results from queried text prompt in seconds. As proposed by ATT3D [33], the 3D generator training is performed by minimizing $\min_{\mathcal{G}} \mathbb{E}_{\pi, y \in S_{y}, x = g(\mathcal{G}(y), \pi)}[\mathcal{L}(\boldsymbol{x}, y)]$ over all text prompts. Unlike data-driven approaches [17, 52, 66], score distillation bypasses the scarcity of text-3D data pairs because the 2D diffusion prior can offer the guidance to align the 3D output with the input text prompt. However, its application is currently restricted to training the 3D generator within a limited range of text prompts. + +# 2.2 Representative Score Distillation Methods + +Denote by $\phi$ the 2D diffusion prior [44,48] and by $p^{\phi}(\boldsymbol{x} \mid y)$ the text-conditioned image distribution embedded within $\phi$ , the objectives of most existing score distillation methods can be generally concluded as minimizing the objective $\mathcal{L}(\theta, y) = \mathbb{E}_{\pi, t, \epsilon, \boldsymbol{x} = g(\theta, \pi)}\left[\omega(t) D_{\mathrm{KL}}\left(q_t^\theta(\boldsymbol{x}_t \mid \pi) \| p_t^\phi(\boldsymbol{x}_t \mid y^\pi)\right)\right]$ , where $D_{\mathrm{KL}}$ denotes KL divergence, $q_t^\theta(\boldsymbol{x}_t \mid \pi)$ denotes the distribution of images $\boldsymbol{x}$ rendered at camera view $\pi$ at diffusion timestep $t$ [15], and the same for $p_t^\phi(\boldsymbol{x}_t \mid y)$ . $\omega(t)$ is a timestep-dependent weight [40]. $y^\pi$ denotes the view-dependent strategy [44] or view-awareness [42, 48] to prompt the different camera views [40]. To minimize this objective, the gradient w.r.t. $\theta$ can be calculated as per [59]: + +$$ +\nabla_ {\theta} \mathcal {L} (\theta , y) = \mathbb {E} _ {\pi , t, \epsilon} \left[ \omega (t) \left(\underbrace {- \sigma_ {t} \nabla_ {\boldsymbol {x} _ {t}} \log p _ {t} ^ {\phi} (\boldsymbol {x} _ {t} \mid y ^ {\pi})} _ {\epsilon_ {\phi} (\boldsymbol {x} _ {t}; t, y ^ {\pi})} - \underbrace {(- \sigma_ {t} \nabla_ {\boldsymbol {x} _ {t}} \log q _ {t} ^ {\theta} (\boldsymbol {x} _ {t} \mid \pi))} _ {\epsilon_ {\theta} (\boldsymbol {x} _ {t}; t, \pi , y)}\right) \frac {\partial \boldsymbol {x}}{\partial \theta} \right], \tag {1} +$$ + +where the first term $-\sigma_t\nabla_{\boldsymbol{x}_t}\log p_t^\phi (\boldsymbol{x}_t\mid y^\pi)$ corresponds to the score function [50] of the desired image distribution, and it can be achieved by predicting the noise $\epsilon \sim \mathcal{N}(0,I)$ in the noisy image $\boldsymbol{x}_t = \alpha_t\boldsymbol {x} + \sigma_t\boldsymbol{\epsilon}$ using the pretrained 2D diffusion model $\epsilon_{\phi}(\pmb {x}_t;t,y^\pi)$ [44,48]. Existing score distillation methods [40,59,72] mainly differ in how to model $-\sigma_t\nabla x_t\log q_t^\theta (x_t\mid \pi)$ , which corresponds to the score function of the distribution of rendered images $q^{\theta}(\pmb {x}\mid \pi)$ . We denote this term in Eq. 1 as $\epsilon_{\theta}(\pmb {x}_t;t,\pi ,y)$ in the following context, since it represents a diffusion model that corresponds to $\theta$ . A summary of the objectives of major score distillation methods is shown in Tab. 1. + +The objective of Score Distillation Sampling (SDS) [40] is $\nabla_{\theta}\mathcal{L}_{\mathrm{SDS}}(\theta ,y)\triangleq$ $\mathbb{E}_{\pi ,t,\epsilon}\left[\omega (t)(\pmb{\epsilon}_{\phi}(\pmb{x}_t;t,y^\pi) - \pmb {\epsilon})\frac{\partial\pmb{x}}{\partial\theta}\right]$ , which approximates the term $\pmb{\epsilon}_{\theta}(\pmb {x}_t;t,\pi ,y)$ in Eq. 1 as the ground-truth noise $\pmb{\epsilon}$ . That is, SDS assumes that $q^{\theta}(\pmb {x}\mid \pi)$ adheres to a Dirac distribution $\delta (\pmb {x} - g(\theta ,\pi))$ [59], which is characterized by a non-zero density at the singular point of $\pmb {x} = g(\theta ,\pi)$ and zero density everywhere else. However, updating $\theta$ under the Dirac distribution might be troublesome [59]. We may need to set the CFG (Classifier Free Guidance) [16] as high as 100 for model convergence, which will produce excessively large gradients and lead to unstable optimization. This problem is alleviated by Classifier Score Distillation (CSD) [72], which uses the classifier component [16] in SDS as the objective: $\nabla_{\theta}\mathcal{L}_{\mathrm{CSD}}(\theta ,y)\triangleq \mathbb{E}_{\pi ,t,\epsilon}\left[\omega (t)(\pmb{\epsilon}_{\phi}(\pmb {x}_t;t,y^\pi) - \pmb{\epsilon}_{\phi}(\pmb {x}_t;t))\frac{\partial\pmb{x}}{\partial\theta}\right]$ . CSD can be regraded as straightforwardly using the unconditional term of the diffusion prior $\pmb{\epsilon}_{\phi}(\pmb{x}_t;t)$ to represent $\pmb{\epsilon}_{\theta}(\pmb {x}_t;t,\pi ,y)$ in Eq. 1. Unfortunately, in the case of prompt-amortized training, this term may not provide effective gradient, because $\pmb{\epsilon}_{\phi}(\pmb {x}_t;t)$ is unconditional to the provided text-prompts. In contrast, Variational Score Distillation (VSD) [59] models $\pmb{\epsilon}_{\theta}(\pmb {x}_t;t,\pi ,y)$ with another text-aware diffusion model $\pmb{\epsilon}_{\phi^{\prime}}(\mathbf{x}_t;t,\pi ,y)$ , leading to $\nabla_{\theta}\mathcal{L}_{\mathrm{VSD}}(\theta ,y)\triangleq$ $\mathbb{E}_{\pi ,t,\epsilon}\left[\omega (t)(\pmb{\epsilon}_{\phi}(\pmb {x}_t;t,y^\pi) - \pmb{\epsilon}_{\phi^{\prime}}(\pmb {x}_t;t,\pi ,y))\frac{\partial\pmb{x}}{\partial\theta}\right]$ , where $\pmb{\epsilon}_{\phi^{\prime}}(\mathbf{x}_t;t,\pi ,y)$ is achieved by finetuning the pretrained 2D diffusion prior $\pmb{\epsilon}_{\phi}(\pmb {x}_t;t,y^\pi)$ to align with the rendered image distribution $q^{\theta}(\pmb {x}\mid \pi)$ via parameter efficient adaptation [18]. In + +
MethodGradient of L(x,y) w.r.t. x = g(θ,π)
SDS [40]E_t,ε[ω(t) (εφ(xt; t, yπ) - ε)]
CSD [72]E_t,ε[ω(t) (εφ(xt; t, yπ) - εφ(xt, t))]
VSD [59]E_t,ε[ω(t) (εφ(xt; t, yπ) - εφ'(xt; t, π, y))]
ASD (Ours)E_t,ε[ω(t) (εφ(xt; t, yπ) - εφ(xt+Δt; t+Δt, yπ))]
+ +Table 1: Objectives of representative score distillation methods. ASD introduces $\Delta t$ alongside $t$ to align with the rendered image distribution $q^{\theta}(\boldsymbol{x} \mid \pi)$ . + +practice, this is conducted by alternatively optimizing $\theta$ and finetuning $\phi$ with the noise prediction objective $\| \pmb{\epsilon}_{\phi}(\pmb{x}_t;t,y) - \pmb{\epsilon}\| _2^2$ [15] such that: + +$$ +\mathbb {E} _ {\pi , t, \epsilon} \left[ \| \epsilon_ {\phi J} \left(\boldsymbol {x} _ {t}; t, \pi , y\right) - \epsilon \| _ {2} ^ {2} \right] \leq \mathbb {E} _ {\pi , t, \epsilon} \left[ \| \epsilon_ {\phi} \left(\boldsymbol {x} _ {t}; t, y ^ {\pi}\right) - \epsilon \| _ {2} ^ {2} \right]. \tag {2} +$$ + +The above equation reveals that a better alignment with the distribution of $q^{\theta}(\pmb{x} \mid \pi)$ can be achieved by a more accurate noise prediction. + +While VSD achieves state-of-the-art results in prompt-specific text-to-3D [14, 59], it changes the diffusion prior's parameters by alternately optimizing $\theta$ and finetuning $\phi$ . This forms a bi-level optimization, known to be problematic in generative adversarial training [55], and may be troublesome for training prompt-amortized text-to-3D models, because the change of pre-trained diffusion model might impairs its comprehension capability on a wide range of text-prompts. In specific, the pre-trained 2D diffusion model may have to sacrifice its generation capability in order to align with the distribution of rendered images, making it fail to produce good gradient for training the 3D generator. + +# 3 Asynchronous Score Distillation (ASD) + +# 3.1 Objective of ASD + +From the above discussions in Sec. 2.2, it can be seen that one key issue in VSD is to minimize the noise prediction error so that the model output can be aligned with the desired distribution of rendered images. VSD achieves this goal via finetuning the pre-trained 2D diffusion model, which however sacrifices its comprehension capability on text prompts. One interesting question is: can we minimize the noise prediction error without changing the pre-trained diffusion network weights? Fortunately, we find that this is possible and in this section we present a new objective function to achieve this goal. + +Recall that diffusion models solve the stochastic differential equation [50] via reversing the noise added along different stages, a.k.a. diffusion timestep $t \in \{T_{\max}, \dots, T_{\min}\}$ via $\boldsymbol{x}_t = \alpha_t \boldsymbol{x} + \sigma_t \boldsymbol{\epsilon}$ [15]. The influence of the noise $\epsilon \sim \mathcal{N}(0, I)$ on the image $\boldsymbol{x}$ is incrementally reduced as the process progresses from the initial timestep $T_{\max}$ to the final timestep $T_{\min}$ , which is controlled by the scalars $\alpha_t$ and $\sigma_t$ . Consequently, the diffusion model's noise prediction accuracy will vary with the timestep $t$ , at which the identical noise $\epsilon$ is added. To evaluate this, we consider a diffusion model with fixed image $\boldsymbol{x}$ , noise $\epsilon$ and condition $y$ , but varied timestep $t$ . We denote such a diffusion model as $\epsilon(t)$ and explore how its prediction error, denoted by $e(t) = \| \epsilon(t) - \epsilon \|_2^2$ , changes with $t$ . + +The model $\epsilon(t)$ can be a pre-trained 2D diffusion model (such as Stable Diffusion [44]). We denote by $\epsilon_{PT}(t)$ such a model, and investigate the behaviour of its noise prediction error, denoted by $e_{PT}(t)$ . In Fig. 1, we plot the curve (i.e., the blue colored curve) of $e_{PT}(t)$ versus $t$ . We use a corpus with 15 text prompts from Magic3D [40] to draw this curve. For each prompt $y$ , we generate 16 images with VSD [59]. Then for each image $\pmb{x}$ , we apply one instance of Gaussian noise $\pmb{\epsilon}$ and conduct a single diffusion step with 100 distinct timesteps. The average noise reconstruction error is then calculated for these timesteps across all prompts and images. We can see from the curve of $e_{PT}(t)$ that earlier diffusion timesteps (e.g., timestep 600) will have lower noise prediction error than later timesteps (e.g., timestep 200). Such a trend holds for almost every image sample $\pmb{x}$ and noise sample $\pmb{\epsilon}$ because the well-trained diffusion model is frozen in our case. Since the noise prediction error declines from $T_{\mathrm{min}}$ (i.e., late diffusion timestep) to $T_{\mathrm{max}}$ (i.e., early diffusion timestep), we can conclude that for a given timestep $t$ and a timestep shift $0 \leq \Delta t \leq T_{\mathrm{max}} - t$ , the following inequality holds: + +$$ +\mathbb {E} _ {\pi , t, \epsilon} \left[ \| \epsilon_ {\phi} \left(\boldsymbol {x} _ {t + \Delta t}; t + \Delta t, y ^ {\pi}\right) - \epsilon \| _ {2} ^ {2} \right] \leq \mathbb {E} _ {\pi , t, \epsilon} \left[ \| \epsilon_ {\phi} \left(\boldsymbol {x} _ {t}; t, y ^ {\pi}\right) - \epsilon \| _ {2} ^ {2} \right], \tag {3} +$$ + +which implies that more accurate noise predictions can be achieved at earlier diffusion timesteps. + +The above property of diffusion models has also been observed by Yang et al. [68], who indicated that as the timestep shifts from $T_{\mathrm{max}}$ towards $T_{\mathrm{min}}$ , the variance in noise prediction increases, as evidenced by the rising Lipschitz constants, which suggests an increased instability in noise prediction and larger noise prediction errors. Such a behavior can be observed in both $\epsilon$ -prediction and $\pmb{v}$ -prediction models, as well as in 2D and 3D diffusion models (please refer to supplementary material for details). This can be intuitively explained as follows. When $t \to T_{\mathrm{max}}$ , $\pmb{x}_t = \alpha_t \pmb{x} + \sigma_t \pmb{\epsilon} \to \pmb{\epsilon}$ , then it is easier to achieve $\epsilon_{\phi}(\pmb{x}_t; t, y^{\pi}) \approx \epsilon$ because the model can manage to copy the input as the output. + +The similarity between Eq. 3 and the fine-tuning objective of VSD in Eq. 2 inspires us to investigate whether simply shifting earlier the timestep could fulfill the fine-tuning requirements of VSD without modifying the pre-trained 2D diffusion network parameters. Specifically, we employ the pretrained 2D diffusion model with shifted timestep to approximate the diffusion model of rendered images in Eq. 1 as $\pmb{\epsilon}_{\theta}(\pmb{x}_t; t, \pi, y) \triangleq \pmb{\epsilon}_{\phi}(\pmb{x}_{t + \Delta t}; t + \Delta t, y^{\pi})$ , resulting in the following Asynchronous Score Distillation (ASD) objective function: + +$$ +\nabla_ {\theta} \mathcal {L} _ {\mathrm {A S D}} (\theta , y) \triangleq \mathbb {E} _ {\pi , t, \epsilon} \left[ \omega (t) \left(\boldsymbol {\epsilon} _ {\phi} \left(\boldsymbol {x} _ {t}; t, y ^ {\pi}\right) - \boldsymbol {\epsilon} _ {\phi} \left(\boldsymbol {x} _ {t + \Delta t}; t + \Delta t, y ^ {\pi}\right)\right) \frac {\partial \boldsymbol {x}}{\partial \theta} \right]. \quad (4) +$$ + +We can see that rather than iteratively fine-tuning the diffusion network as in VSD, ASD achieves similar goal by shifting the timestep $t$ with an interval $\Delta t$ in each step, which is much more efficient. One key variable introduced in ASD is the timestep shift $\Delta t$ , which will be discussed in the next subsection. + +![](images/42b5dd41704546434be15a59e10acbfba878ec1590bc5e05f91031f968d6966f.jpg) +Fig. 1: Illustration of the noise prediction error of the pre-trained 2D diffusion model $\epsilon_{PT}(t)$ and that of the fine-tuned 2D diffusion model $\epsilon_{FT}(t)$ . We can see that the curve of $e_{FT}(t)$ is positioned under that of $e_{PT}(t)$ , and we can shift the timestep of $\epsilon_{PT}(t)$ to $\epsilon_{PT}(t + \Delta t)$ to approximate the noise prediction error of $\epsilon_{FT}(t)$ . + +# 3.2 The Setting of Timestep Shift $\Delta t$ + +Before discussing how to set the timestep shift $\Delta t$ , let's plot another curve, i.e., the noise prediction error of $\epsilon_{\theta}(\boldsymbol{x}_t; t, \pi, y)$ w.r.t. timestep $t$ . Actually, in the process of generating $\boldsymbol{x}$ with VSD, we will have the fine-tuned model $\epsilon_{\phi'}(\boldsymbol{x}_t; t, \pi, y)$ as the by-product, which is used to represent $\epsilon_{\theta}(\boldsymbol{x}_t; t, \pi, y)$ in Eq. 1. Therefore, with fixed $\boldsymbol{x}$ , $\epsilon$ and $y$ , the noise prediction error of the fine-tuned diffusion model, denoted by $\epsilon_{FT}(t)$ , can be calculated as $e_{FT}(t) = \| \epsilon_{\phi'}(t) - \epsilon \|_2^2$ . + +The curve of $e_{FT}(t)$ w.r.t. $t$ (i.e., the yellow curve) is plotted in Fig. 1 by using the same data as in plotting $e_{PT}(t)$ . We can see that the curve of $e_{FT}(t)$ is positioned under $e_{PT}(t)$ because $e_{FT}(t)$ is obtained by the fine-tuned diffusion model $\epsilon_{FT}$ . However, as mentioned in Sec. 2.2, this fine-tuning changes the weights of pre-trained diffusion model and might damage its ability in comprehending text-image pairs. Therefore, we propose to fix the pre-trained model $\epsilon_{PT}(t)$ but shift it to $\epsilon_{PT}(t + \Delta t)$ to approximate the desired $\epsilon_{FT}(t)$ . Referring to Fig. 1, we could shift $\epsilon_{PT}(t)$ to an earlier timestep to achieve this goal. For example, at timestep $t_0$ and with a time shift $\Delta t_0 > 0$ , we can use $\epsilon_{PT}(t_0 + \Delta t_0)$ to approximate the noise prediction error of $\epsilon_{FT}(t_0)$ . + +On the other hand, the magnitude of $\Delta t$ will vary with $t$ . Let's come to another timestep $t_1$ in Fig. 1, where $t_1$ is earlier than $t_0$ . Because the decreasing speeds of both $e_{PT}$ and $e_{FT}$ will be reduced with $t$ going to $T_{\mathrm{max}}$ , the magnitude of $\Delta t_1$ will be increased to approximate $e_{FT}(t_1)$ . In other words, the magnitude of $\Delta t$ should grow when $t$ goes from $T_{\mathrm{min}}$ to $T_{\mathrm{max}}$ . We heuristically set this relationship as $\Delta t = \eta (t - T_{\mathrm{min}})$ , where $\eta \in [0,1]$ is a hyper-parameter that controls the length of shift range. Finally, it should be pointed out that the curves in Fig. 1 will vary a little for different training iterations, rendered images $\mathbf{x}$ and text prompts $y$ . Therefore, $\Delta t$ should fall into some range $S(t)$ . In practice, we set $\Delta t \sim S(t) = \mathcal{U}[0,\eta (t - T_{\mathrm{min}})]$ , which follows a uniform distribution within 0 + +![](images/a632f5b8e981ceef862e3fd7e172e560f6a486258493c5184b3f536ace0e9f65.jpg) +Fig. 2: Left and middle: 2D toy examples by SDS [40], CSD [72], VSD [59] and our proposed ASD. Right: Gradient norms generated by different methods. + +# Algorithm 1 Asynchronous Score Distillation (ASD) + +Input: 3D representation $\theta$ ; Text prompt $y$ ; Hyperparameter $\eta$ ; 2D diffusion prior $\epsilon_{\phi}$ + +```latex +1 while not converged do +2 Sample a camera pose $\pi$ +3 Render an image $\pmb{x} = g(\theta, \pi)$ +4 Sample a timestep $t \sim \mathcal{U}[T_{\min}, T_{\max}]$ , Gaussian noise $\epsilon \sim \mathcal{N}(0, I)$ +5 Sample a timestep shift $\Delta t \sim S(t) = \mathcal{U}[0, \eta(t - T_{\min})]$ +6 $\pmb{x}_t \gets \alpha_t \pmb{x} + \sigma_t \pmb{\epsilon}$ , $\pmb{x}_{t + \Delta t} \gets \alpha_{t + \Delta t} \pmb{x} + \sigma_{t + \Delta t} \pmb{\epsilon}$ +7 Update $\theta$ with $\Delta \theta \gets \omega(t) (\pmb{\epsilon}_{\phi}(\pmb{x}_t; t, y^{\pi}) - \pmb{\epsilon}_{\phi}(\pmb{x}_{t + \Delta t}; t + \Delta t, y^{\pi})) \frac{\partial \pmb{x}}{\partial \theta}$ +8 end +``` + +and $\eta(t - T_{\min})$ . The pseudo-code of ASD is summarized in Alg. 1, which can be applied to both prompt-specific and prompt-amortized text-to-3D tasks. + +2D toy experiments. To verify the proposed timestep shift strategy, we follow the paradigm in [59] to test SDS, CSD, VSD and our ASD on 2D toy examples. The left column of Fig. 2 shows the results of SDS, CSD, VSD, and the middle column shows the results of ASD with different sampling strategies of $\Delta t$ . One can see that the proposed sampling strategy $\Delta t \sim S(t) = \mathcal{U}[0, \eta(t - T_{\min})]$ yields similar results to VSD [59]. Besides, we show the gradient norm produced by these score distillation methods in the right column of Fig. 2. One can see that the range of gradient norm produced by ASD is similar to that of VSD. However, the gradient norm of SDS is more than 10 times larger than ASD and VSD because it needs to set CFG=100 for convergence [40, 59, 72]. Such a large gradient may result in training instability. We append more 2D results in the supplementary material to further validate our proposed sampling strategy. + +Text-to-3D Synthesis with ASD. As a score distillation method, ASD is open to the selection of 3D generator architectures [4,17,22,33,39]. The general pipeline of ASD for text-to-3D synthesis is shown in Fig. 3. It takes a rendered image as input and diffuses it in two timesteps $t$ and $t + \Delta t$ . The noise prediction difference is used as the gradient to optimize the 3D representation of generator. In this work, in addition to prompt-specific generation, as done in most existing score distillation works [16, 28, 40, 59, 64], we focus more on prompt-amortized text-to-3D and conduct thorough experiments to evaluate the effectiveness of + +![](images/38adc03247ab655fdd85b769d22913e0486abe4f7fc76d04425fc119318d0119.jpg) +Fig. 3: Overview of Asynchronous Score Distillation (ASD). As illustrated in the left sub-figure, ASD can be employed for prompt-specific generation by optimizing 3D representations for each prompt, as well as for prompt-amortized generation by training a text-to-3D generator. The right sub-figure depicts how ASD uses the difference in noise predictions at asynchronous timesteps to update the 3D network parameters. + +ASD with representative 3D generator architectures, including Hyper-iNGP and 3DConv-net, by using Stable Diffusion as the 2D diffusion model. HyperiNGP is adopted by ATT3D [33], which integrates a prompt-agnostic hash-grid spatial encoding [39] with prompt-conditioned decoding layers to output color and density. 3DConv-net [4] is a 3D generator that maps the provided condition to voxel using 3D convolution. They are chosen in our experiments because they represent two groups of 3D generators that inject the condition through implicit mapping function [3,21] or explicit space representation [7,17,46,49]. The results are shown in Sec. 4. We also conduct experiments to demonstrate that ASD can work with other 2D diffusion models such as MVDream [48], and can be used to train other 3D generator architectures such as Triplane-Transformer [17]. More details of these experiments can be found in the supplementary material. + +# 4 Experiments + +# 4.1 Experimental Settings + +Comparison Methods. We compare ASD with state-of-the-art score distillation methods, including SDS [40], CSD [72] and VSD [59]. We adhere to their official codes for training prompt-amortized text-to-3D networks. For example, the CFG [16] values for SDS, CSD and VSD are configured to 100, 1, and 7.5, respectively. In addition, we compare with existing prompt-amortized method ATT3D [33] (whose code is not released yet) by replicating its reported results. + +Implementation Details. We employ VolSDF [69] to render images from the 3D generators. For Stable Diffusion, we employ SD-v2.1-base [1] for all score distillation methods for fair comparison. As configured in VSD [59], we set the CFG value as 7.5 for the pre-trained diffusion model in ASD, and 1 for the diffusion model of rendered images. The resolution of rendered images by HyperiNGP is set to $256 \times 256$ , while that of 3DConv-net and Triplane-Transformer is set to $64 \times 64$ for GPU memory considerations. Other details are in the supplementary material. + +Prompt Corpus. To thoroughly evaluate the capability of ASD in prompt-amortized text-to-3D synthesis, we employ multiple datasets encompassing a + +range of text prompt quantities. MG15 includes 15 prompts from Magic3D [29]; DF415 comprises 415 prompts from DreamFusion [40]; and AT2520 contains 2520 compositional prompts of animals from ATT3D [33]. DL17k contains 17k compositional prompts of human with daily activities, proposed by [25]. While AT2520 and DL17k provide a larger number of prompts than DF415, the prompt diversity of them is relatively low due to the predefined templates. + +To test ASD's performance with an even larger scale of prompts, we introduce a novel prompt corpus named CP100k. This corpus consists of 100,000 text prompts filtered from the image descriptions collected by Cap3D [34], which was developed to test text-to-image model performance. To the best of our knowledge, it is the first time to evaluate score distillation methods on such a scale of text prompts. Meanwhile, it should be clarified that this work is focused on examining the score distillation performance rather than prompt generalization, so the test prompts share the same distribution as training prompts. More details of the prompt corpus are in supplementary material. + +Evaluation Metrics. We render 120 surrounding view images as the 3D synthesis result from each prompt. Similar to previous text-to-3D works [25,33, 33,40], we compute the CLIP recall, i.e., the classification accuracy by applying CLIP model to the rendered images to predict the correct text prompt, as one performance metric, denoted by "R@1". Additionally, we calculate the CLIP text-image similarity between generated images and input prompts as another metric [54,60], denoted by "Sim". + +# 4.2 Evaluation Results + +Results with iNGP/Hyper-iNGP as 3D Representation. The iNGP [39] architecture is designed for prompt-specific text-to-3D generation. Hyper-iNGP has the same spatial encoding as iNGP except that the weights of the decoding layer depend on the text prompt. To eliminate the effect caused by architecture difference as much as possible, we adopt iNGP for prompt-specific text-to-3D tasks, and Hyper-iNGP for prompt-amortized tasks. Our experiments are carried out on the MG15 dataset. For prompt-specific tasks, we optimize an individual iNGP [39] for each MG15 prompt; while for the prompt-amortized tasks, we train a single Hyper-iNGP [33] across all MG15 prompts. We also compare our results with ATT3D [33], which is among the first to apply Hyper-iNGP to prompt-amortized text-to-3D tasks. ATT3D employs SDS for training and uses soft-shading [40] (denoted as * in Tab. 2) for rendering. + +The qualitative and quantitative results are shown in Fig. 4 and Tab. 2, respectively. We can see that the existing methods suffer from performance decrease when transiting from prompt-specific to prompt-amortized tasks, as evidenced by the decreased CLIP similarity and recall in Tab. 2. It is worth mentioning that training Hyper-net with SDS requires turning on the spectral normalization [38] in the linear layers, otherwise the training will fail due to numerical instability. This observation is consistent with what reported in ATT3D [33]. This is because SDS suffers from large gradient norm (please also refer to Fig. 2 and the discussions therein), which makes Hyper-iNGP hard to + +![](images/408addbe1a5f9af089cabaa05c98be98d1cddd5e3eeb7aa992503005b8fc424a.jpg) +Fig. 4: Qualitative comparison on prompt-specific (with iNGP as the 3D representation) and prompt-amortized (with Hyper-iNGP as the 3D generator) text-to-3D results by SDS [40], CSD [72], VSD [59], ATT3D [33] and our ASD methods. + +![](images/70f23e0a08667dcc4ee0b16f3e88208beb54d087ffb6d5993bfb832b8cd2dadb.jpg) +"A DSLR photo of a peacok on a surfboard" in MG15 + +![](images/5ff548565ca51d53e4528d5ac75a260fc105de4301fd93dce9de5fdba658bf00.jpg) + +![](images/f1c78c07b9f778bee145766d543dc69bdf4e1170162c7308440824061e2e9a28.jpg) + +![](images/89da58504250c50a28819e51424c8b9344b20cf333d7a496a2765dcaa8610e40.jpg) + +
ReferenceMethodSim ↑R@1 ↑MethodSim ↑R@1 ↑
ATT3D [33]---Hyper-iNGP* + SDS0.1950.468
DreamFusion [40]iNGP + SDS0.2881.000Hyper-iNGP + SDS0.2570.918
Classifier [16]iNGP + CSD0.2800.936Hyper-iNGP + CSD0.2640.972
ProlificDreamer [59]iNGP + VSD0.2760.932Hyper-iNGP + VSD0.2590.987
OursiNGP + ASD0.2891.000Hyper-iNGP + ASD0.2841.000
+ +Table 2: Quantitative comparison on prompt-specific (with iNGP as the 3D representation) and prompt-amortized (with Hyper-iNGP as the 3D generator) text-to-3D results by SDS [40], CSD [72], VSD [59], ATT3D [33] and our ASD methods. + +converge. As can be seen in Fig. 4, ATT3D results in wrong geometry by using soft shading and SDS for training. For CSD, we see that it fails to optimize the full geometry, as shown by the shrunk peacock in both prompt-amortized and prompt-amortized results. For VSD, it tends to generate content drifts [48], resulting in repetitive patterns and abnormal geometry. It may fail to generate reasonable contents in both prompt-specific and prompt-amortized tasks. In contrast, our proposed ASD works very stable across the two tasks, yielding not only outstanding quantitative scores but also high quality 3D contents. + +Results with 3DConv-net as 3D Generator. The issues of existing score distillation methods either persist or become more pronounced when replacing Hyper-iNGP to 3DConv-net as the 3D generator. We find that training SDS with 3DConv-net always fails within several thousand iterations, even using spectral or other normalization techniques. This issue stems from that deeper network is more sensitive to large gradients [13] caused by SDS. Therefore, we only compare the results of other methods in Fig. 5. We see that CSD outputs acceptable results on AT2520, but its results on DF415, which has more varied prompts, are consistently smaller than anticipated. Such a phenomenon has been observed when Hyper-iNGP is used as the generator, which underlines CSD's inability to reliably guide the 3D generator to produce geometries aligned with the text prompts. As for VSD, it leads to rather abnormal results, failing to match the text prompts. This can be attributed to its fine-tuning of the pre-trained 2D diffusion model, which severely compromises VSD's text-image comprehending ability. In comparison, our proposed ASD, with 3DConv-net as the generator, + +![](images/b3a37a77b484107a67ee819383b2523eaae0bb583d0f64c89985bbddfb578f32.jpg) + +![](images/3d868fa32efcdbdd740f76d8d754a5ac827cea9849b470efc4fe52b044cb5b4e.jpg) + +![](images/24446fcc06bb1364faf18446ffdc62d4e0454581cde6a5660f0e42c79fdb78be.jpg) +Asynchronous Score Distillation (ASD, ours) +"A squirrel holding a book wearing a sweater wearing a tophat" in AT2520 + +![](images/a5ca991def4b83545b4ae7a81043c95b557b01ff73bedf54aaef828fca7134b8.jpg) +Fig. 5: Qualitative comparison among CSD [72], VSD [59] and our ASD (with 3DConv-net as generator) on AT2520 and DF415 corpuses. SDS is not compared because it encounters numerical instability in this experiment. + +![](images/44b3d42e96d1a3292a1a7e001c390c0e1b5848872b77cf845b9224fd7a1bbd04.jpg) +"A DSLR photo of a cocker spaniel wearing a crown" in DF415 + +![](images/450fab972bbfc2c45476959b453c13a79c0fb0f539d28be8f77000d69e45cfde.jpg) + +![](images/e3394f68d5b3630cbecd4f07c52674033542965ca25c2e3262b6d00195c8aeea.jpg) + +![](images/164be7bcf11c40c71aacf25f2af2062f50ba2407999f6fea028fd83591c7d520.jpg) + +![](images/749cd727c6f06661375d504d3d893cba7c7d988f5d7e072f021e72f46949fc65.jpg) + +![](images/e662d35d7a692f9fb4d5eaa18a7187cf08b18d429301d6c631c0b6a5a94696be.jpg) + +
MethodDF415AT2520CP100k
Sim ↑R@1 ↑Sim ↑R@1 ↑Sim ↑R@1 ↑
SDS××××××
CSD0.1760.0620.2790.0370.1950.108
VSD0.1580.0020.1150.0010.1030.000
ASD (ours)0.2370.2760.2850.0580.1990.117
+ +Table 3: Quantitative comparison on prompt-amortized text-to-3D with 3DConv-net as generator. Symbol $\times$ denotes that the training fails due to numerical instability. + +yields improved outcomes, as evidenced by the visual results in Fig. 5 and the enhanced metric scores in Tab. 3. + +Scalability. In this section, we evaluate the scalability of competing methods by using as many as 100k prompts in the CP100k dataset with 3DConv-net as the generator. The results are shown in Fig. 6 and Tab. 3. Due to the issue of numerical instability, SDS is not involved in this experiment. We can see that the outcomes of CSD are significantly diminished with uniformly small-sized shapes across all prompts. There is also a lack of variety since most outputs exhibit similar patterns. The results of VSD are also degenerated, displaying almost identical and anomalous outcomes for the text prompts. This resembles the phenomenon of mode collapse often encountered in bi-level optimization [55], which also highlights the importance of fixing the 2D diffusion model when training with such a large number of text prompts. In comparison, ASD is able to produce much higher quality outcomes across the text prompts, showcasing its capability in large-scale training with numerous text prompts as inputs. + +In addition to the above comparisons, we also evaluate the effectiveness of ASD when using other 2D diffusion models. We employ MVDream [48] to illustrate ASD's generality to diffusion prior models. Besides, we compare against data-driven methods to prove that ASD is advantageous in tackling diverse input of text prompts. Please see supplementary material for detail. + +# 4.3 Ablation Study + +In this section, we perform ablation studies to evaluate the settings of timestep shift $\Delta t\sim S(t) = \mathcal{U}[0,\eta (t - T_{\mathrm{min}})]$ from several aspects. The qualitative and + +![](images/fa6b09ba5048a47e5b6389bbe3d4e5b0b391bc2dfaffa711d11225f95d38282b.jpg) + +![](images/d76ea56d08e5a8f55c72f62961598316a507b2285fadf685735f1b0893149965.jpg) + +![](images/c12bcf73513fd51a5dac999544ebd93ab04829e977d8f310604e724038b9dfc1.jpg) + +![](images/0192162f68d2fc7bc887c2950ede62e400914c1b423010cba396e5a226cc99ba.jpg) +Fig. 6: The scalability comparison with CSD [72] and VSD [59] on CP100k corpus. + +![](images/a78e82c0fd9d95b6c49ff83b72dc71e4cd27947f1461ab84884f1f753fa1b43e.jpg) +"Small wooden brown bear figurine" + +![](images/8da0bfae92e89486c3c7b9911273aa674f195a7ba6ad3986bfb4f3f8d63fe956.jpg) + +quantitative results are shown in Fig. 7 and Tab. 4, respectively. We provide extensive ablation studies in the supplementary material to show that the proposed strategy is also effective when using MVDream as the diffusion model. + +Importance of Timestep Shift. We use $\eta = 0$ (i.e., no timestep shift) as a baseline to evaluate the necessity of introducing timestep shift $\Delta t$ . From Fig. 7 and Tab. 4, we see that while it can generate plausible results, the model is prone to generating shapes that do not make sense, such as the so-called Janus problem [2]. Examples include a frog with an extra eye, robot face with block-like features, and a peacock with tails at both the front and back. This is because the non-shifted diffusion model will align more with the 2D image distribution, tending to generate redundant contents and unreasonable geometry along the training. By introducing a timestep shift, our proposed ASD demonstrates advantages in achieving more coherent and visually pleasing results. + +Range of Timestep Shift. By setting $\eta = 0.2$ , we allow $\Delta t$ to be sampled from a large range. However, this might not be a good choice. In the extreme case, for any timestep $t$ we can set a large interval $\Delta t$ such that $t + \Delta t = T_{\max}$ , then the noise prediction becomes $\epsilon_{\phi}(\boldsymbol{x}_t; t, y^\pi) \approx \epsilon$ , so that ASD is degraded to SDS, which cannot perform well under CFG=7.5 [40]. In practice, we find a larger $\eta$ tends to result 3D contents with larger size and rounded shapes, e.g., the peacock with closer views, or the frog with larger size, as shown in Fig. 7. Therefore, we set $\eta = 0.1$ in all our experiments. + +Deterministic or Random Shift. If we set $\Delta t = \eta (t - T_{\mathrm{min}})$ , it assumes that the diffusion model of rendered images can be approximated by the pretrained one with a fixed and deterministic timestep shift. As shown in Fig. 7 and Tab. 4, it reduces the chance to generate correct geometry and colors. Randomly sampling $\Delta t$ in a range is more effective, which is adopted in our method. + +# 5 Conclusion and Limitations + +In this paper, we presented Asynchronous Score Distillation (ASD), a novel score distillation method that can assist 2D diffusion prior in training 3D generators + +![](images/bab712512d378c9d64acee37b1fdfd89aa6daeef99c029a4aa80b8eebb1529c5.jpg) +Fig. 7: The qualitative results of the ablation study on the timestep interval $\Delta t$ . + +
ParamSim ↑R@1 ↑
Δt = η(t - Tmin)η = 0.10.2140.178
η = 0.20.2140.180
η = 00.2350.267
Δt ~ U[0, η(t - Tmin)]η = 0.10.2370.276
η = 0.20.2290.237
+ +Table 4: The quantitative results of the ablation study on the timestep interval $\Delta t$ . + +with a scalable size of text prompts. By shifting the diffusion timestep to earlier stages, our ASD can effectively predict the noise prediction error to align the diffusion model with the distribution of rendered images, while preserving the superior text comprehension capability of pre-trained models, thus facilitating stable training with high-fidelity generation results. Our extensive experiments revealed that ASD performed consistently well on datasets of various sizes, being able to manage as much as 100k prompts. Though ASD has shown improvements over earlier score distillation approaches, there remain some limitations. + +For man-made objects that have very regular shapes, such as chairs or airplanes, the performance of our model will lag behind those data-driven methods, which benefit from an abundance of relevant data. We foresee opportunities to combine the advantages of data-driven and score distillation methodologies to improve text-to-3D capabilities in a more comprehensive manner in the future research. + +# Acknowledgement + +This work is supported in part by the Beijing Science and Technology Plan Project Z231100005923033, and the InnoHK program. + +# References + +1. Stable-diffusion-v2.1-base. https://huggingface.co/stabilityai/stable-diffusion-2-1-base +2. Armandpour, M., Zheng, H., Sadeghian, A., Sadeghian, A., Zhou, M.: Re-imagine the negative prompt algorithm: Transform 2d diffusion into 3d, alleviate janus problem and beyond. arXiv preprint arXiv:2304.04968 (2023) +3. Babu, S., Liu, R., Zhou, A., Maire, M., Shakhnarovich, G., Hanocka, R.: Hyperfields: Towards zero-shot generation of nerfs from text. arXiv preprint arXiv:2310.17075 (2023) +4. Bahmani, S., Park, J.J., Paschalidou, D., Yan, X., Wetzstein, G., Guibas, L., Tagliasacchi, A.: Cc3d: Layout-conditioned generation of compositional 3d scenes. arXiv preprint arXiv:2303.12074 (2023) +5. Balaji, Y., Nah, S., Huang, X., Vahdat, A., Song, J., Kreis, K., Aittala, M., Aila, T., Laine, S., Catanzaro, B., et al.: edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324 (2022) +6. Cao, Z., Hong, F., Wu, T., Pan, L., Liu, Z.: Large-vocabulary 3d diffusion model with transformer. arXiv preprint arXiv:2309.07920 (2023) +7. Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., De Mello, S., Gallo, O., Guibas, L.J., Tremblay, J., Khamis, S., et al.: Efficient geometry-aware 3d generative adversarial networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16123-16133 (2022) +8. Chen, R., Chen, Y., Jiao, N., Jia, K.: *Fantasia3d: Disentangling geometry* and appearance for high-quality text-to-3d content creation. arXiv preprint arXiv:2303.13873 (2023) +9. Chen, Z., Wang, F., Liu, H.: Text-to-3d using gaussian splatting. arXiv preprint arXiv:2309.16585 (2023) +10. Deitke, M., Liu, R., Wallingford, M., Ngo, H., Michel, O., Kusupati, A., Fan, A., Laforte, C., Voleti, V., Gadre, S.Y., et al.: Objaverse-xl: A universe of $10\mathrm{m} + 3\mathrm{d}$ objects. arXiv preprint arXiv:2307.05663 (2023) +11. Ding, L., Dong, S., Huang, Z., Wang, Z., Zhang, Y., Gong, K., Xu, D., Xue, T.: Text-to-3d generation with bidirectional diffusion using both 2d and 3d priors. arXiv preprint arXiv:2312.04963 (2023) +12. Guo, P., Hao, H., Caccavale, A., Ren, Z., Zhang, E., Shan, Q., Sankar, A., Schwing, A.G., Colburn, A., Ma, F.: Stabledreamer: Taming noisy score distillation sampling for text-to-3d. arXiv preprint arXiv:2312.02189 (2023) +13. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) +14. He, Y., Bai, Y., Lin, M., Zhao, W., Hu, Y., Sheng, J., Yi, R., Li, J., Liu, Y.J.: T $^3$ bench: Benchmarking current progress in text-to-3d generation. arXiv preprint arXiv:2310.02977 (2023) +15. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020) +16. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022) +17. Hong, Y., Zhang, K., Gu, J., Bi, S., Zhou, Y., Liu, D., Liu, F., Sunkavalli, K., Bui, T., Tan, H.: Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400 (2023) + +18. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021) +19. Huang, Y., Wang, J., Shi, Y., Qi, X., Zha, Z.J., Zhang, L.: Dreamtime: An improved optimization strategy for text-to-3d content creation. arXiv preprint arXiv:2306.12422 (2023) +20. Jiang, L., Wang, L.: Brightdreamer: Generic 3d gaussian generative framework for fast text-to-3d synthesis. arXiv preprint arXiv:2403.11273 (2024) +21. Jun, H., Nichol, A.: Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463 (2023) +22. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG) 42(4), 1-14 (2023) +23. Koster, R.: Theory of fun for game design. "O'Reilly Media, Inc." (2013) +24. Lee, K., Sohn, K., Shin, J.: Dreamflow: High-quality text-to-3d generation by approximating probability flow. arXiv preprint arXiv:2403.14966 (2024) +25. Li, M., Zhou, P., Liu, J.W., Keppo, J., Lin, M., Yan, S., Xu, X.: Instant3d: Instant text-to-3d generation. arXiv preprint arXiv:2311.08403 (2023) +26. Li, W., Chen, R., Chen, X., Tan, P.: Sweetdreamer: Aligning geometric priors in 2d diffusion for consistent text-to-3d. arXiv preprint arXiv:2310.02596 (2023) +27. Li, Z., Chen, Y., Zhao, L., Liu, P.: Mvcontrol: Adding conditional control to multi-view diffusion for controllable text-to-3d generation. arXiv preprint arXiv:2311.14494 (2023) +28. Liang, Y., Yang, X., Lin, J., Li, H., Xu, X., Chen, Y.: Luciddreamer: Towards high-fidelity text-to-3d generation via interval score matching. arXiv preprint arXiv:2311.11284 (2023) +29. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3d: High-resolution text-to-3d content creation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 300–309 (2023) +30. Lin, Y., Clark, R., Torr, P.: Dreamampolisher: Towards high-quality text-to-3d generation via geometric diffusion. arXiv preprint arXiv:2403.17237 (2024) +31. Liu, Y.T., Luo, G., Sun, H., Yin, W., Guo, Y.C., Zhang, S.H.: Pi3d: Efficient text-to-3d generation with pseudo-image diffusion. arXiv preprint arXiv:2312.09069 (2023) +32. Liu, Z., Li, Y., Lin, Y., Yu, X., Peng, S., Cao, Y.P., Qi, X., Huang, X., Liang, D., Ouyang, W.: Unidream: Unifying diffusion priors for relightable text-to-3d generation. arXiv preprint arXiv:2312.08754 (2023) +33. Lorraine, J., Xie, K., Zeng, X., Lin, C.H., Takikawa, T., Sharp, N., Lin, T.Y., Liu, M.Y., Fidler, S., Lucas, J.: Att3d: Amortized text-to-3d object synthesis. arXiv preprint arXiv:2306.07349 (2023) +34. Luo, T., Rockwell, C., Lee, H., Johnson, J.: Scalable 3d captioning with pretrained models. Advances in Neural Information Processing Systems 36 (2024) +35. Ma, Y., Fan, Y., Ji, J., Wang, H., Sun, X., Jiang, G., Shu, A., Ji, R.: X-dreamer: Creating high-quality 3d content by bridging the domain gap between text-to-2d and text-to-3d generation. arXiv preprint arXiv:2312.00085 (2023) +36. Mercier, A., Nakhli, R., Reddy, M., Yasarla, R., Cai, H., Porikli, F., Berger, G.: Hexagen3d: Stablediffusion is just one step away from fast and diverse text-to-3d generation. arXiv preprint arXiv:2401.07727 (2024) + +37. Metzer, G., Richardson, E., Patashnik, O., Giryes, R., Cohen-Or, D.: Latent-nerf for shape-guided generation of 3d shapes and textures. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12663-12673 (2023) +38. Miyato, T., Kataoka, T., Koyama, M., Yoshida, Y.: Spectral normalization for generative adversarial networks. arXiv preprint arXiv:1802.05957 (2018) +39. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022) +40. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988 (2022) +41. Qian, G., Cao, J., Siarohin, A., Kant, Y., Wang, C., Vasilkovsky, M., Lee, H.Y., Fang, Y., Skorokhodov, I., Zhuang, P., et al.: Atom: Amortized text-to-mesh using 2d diffusion. arXiv preprint arXiv:2402.00867 (2024) +42. Qiu, L., Chen, G., Gu, X., Zuo, Q., Xu, M., Wu, Y., Yuan, W., Dong, Z., Bo, L., Han, X.: Richdreamer: A generalizable normal-depth diffusion model for detail richness in text-to-3d. arXiv preprint arXiv:2311.16918 (2023) +43. Ren, X., Huang, J., Zeng, X., Museth, K., Fidler, S., Williams, F.: Xcube: Large-scale 3d generative modeling using sparse voxel hierarchies. arXiv preprint (2023) +44. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022) +45. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022) +46. Schwarz, K., Sauer, A., Niemeyer, M., Liao, Y., Geiger, A.: Voxgraf: Fast 3d-aware image synthesis with sparse voxel grids. Advances in Neural Information Processing Systems 35, 33999-34011 (2022) +47. Shen, T., Gao, J., Yin, K., Liu, M.Y., Fidler, S.: Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. Advances in Neural Information Processing Systems 34, 6087-6101 (2021) +48. Shi, Y., Wang, P., Ye, J., Long, M., Li, K., Yang, X.: Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512 (2023) +49. Sitzmann, V., Thies, J., Heide, F., Nießner, M., Wetzstein, G., Zollhofer, M.: Deepvoxels: Learning persistent 3d feature embeddings. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2437-2446 (2019) +50. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456 (2020) +51. Tang, B., Wang, J., Wu, Z., Zhang, L.: Stable score distillation for high-quality 3d generation. arXiv preprint arXiv:2312.09305 (2023) +52. Tang, J., Chen, Z., Chen, X., Wang, T., Zeng, G., Liu, Z.: Lgm: Large multi-view gaussian model for high-resolution 3d content creation. arXiv preprint arXiv:2402.05054 (2024) +53. Tang, J., Ren, J., Zhou, H., Liu, Z., Zeng, G.: Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653 (2023) +54. Tang, Z., Gu, S., Wang, C., Zhang, T., Bao, J., Chen, D., Guo, B.: Volumediffusion: Flexible text-to-3d generation with efficient volumetric encoder. arXiv preprint arXiv:2312.11459 (2023) + +55. Thanh-Tung, H., Tran, T.: Catastrophic forgetting and mode collapse in gans. In: 2020 international joint conference on neural networks (ijcnn). pp. 1-10. IEEE (2020) +56. Tsalicoglou, C., Manhardt, F., Tonioni, A., Niemeyer, M., Tombari, F.: Textmesh: Generation of realistic 3d meshes from text prompts. arXiv preprint arXiv:2304.12439 (2023) +57. Vilesov, A., Chari, P., Kadambi, A.: Cg3d: Compositional generation for text-to-3d via gaussian splatting. arXiv preprint arXiv:2311.17907 (2023) +58. Wang, H., Du, X., Li, J., Yeh, R.A., Shakhnarovich, G.: Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12619-12629 (2023) +59. Wang, Z., Lu, C., Wang, Y., Bao, F., Li, C., Su, H., Zhu, J.: Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213 (2023) +60. Wei, Y., Zhang, Y., Ji, Z., Bai, J., Zhang, L., Zuo, W.: Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. arXiv preprint arXiv:2302.13848 (2023) +61. Wohlgenannt, I., Simons, A., Stieglitz, S.: Virtual reality. Business & Information Systems Engineering 62, 455-461 (2020) +62. Wu, R., Sun, L., Ma, Z., Zhang, L.: One-step effective diffusion network for real-world image super-resolution. arXiv preprint arXiv:2406.08177 (2024) +63. Wu, T., Zhang, J., Fu, X., Wang, Y., Ren, J., Pan, L., Wu, W., Yang, L., Wang, J., Qian, C., et al.: Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 803-814 (2023) +64. Wu, Z., Zhou, P., Yi, X., Yuan, X., Zhang, H.: Consistent3d: Towards consistent high-fidelity text-to-3d generation with deterministic sampling prior. arXiv preprint arXiv:2401.09050 (2024) +65. Xie, K., Lorraine, J., Cao, T., Gao, J., Lucas, J., Torralba, A., Fidler, S., Zeng, X.: Latte3d: Large-scale amortized text-to-enhanced3d synthesis. arXiv preprint arXiv:2403.15385 (2024) +66. Xu, Y., Tan, H., Luan, F., Bi, S., Wang, P., Li, J., Shi, Z., Sunkavalli, K., Wetzstein, G., Xu, Z., et al.: Dmv3d: Denoising multi-view diffusion using 3d large reconstruction model. arXiv preprint arXiv:2311.09217 (2023) +67. Yang, Z., Feng, R., Zhang, H., Shen, Y., Zhu, K., Huang, L., Zhang, Y., Liu, Y., Zhao, D., Zhou, J., et al.: Eliminating lipschitz singularities in diffusion models. arXiv preprint arXiv:2306.11251 (2023) +68. Yang, Z., Feng, R., Zhang, H., Shen, Y., Zhu, K., Huang, L., Zhang, Y., Liu, Y., Zhao, D., Zhou, J., et al.: Lipschitz singularities in diffusion models. In: The Twelfth International Conference on Learning Representations (2023) +69. Yariv, L., Gu, J., Kasten, Y., Lipman, Y.: Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems 34, 4805-4815 (2021) +70. Yi, T., Fang, J., Wu, G., Xie, L., Zhang, X., Liu, W., Tian, Q., Wang, X.: Gaussian dreamer: Fast generation from text to 3d gaussian splatting with point cloud priors. arXiv preprint arXiv:2310.08529 (2023) +71. Yu, X., Xu, M., Zhang, Y., Liu, H., Ye, C., Wu, Y., Yan, Z., Zhu, C., Xiong, Z., Liang, T., et al.: Mvimgnet: A large-scale dataset of multi-view images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9150-9161 (2023) + +72. Yu, X., Guo, Y.C., Li, Y., Liang, D., Zhang, S.H., Qi, X.: Text-to-3d with classifier score distillation. arXiv preprint arXiv:2310.19415 (2023) +73. Zhao, M., Zhao, C., Liang, X., Li, L., Zhao, Z., Hu, Z., Fan, C., Yu, X.: Efficient-dreamer: High-fidelity and robust 3d creation via orthogonal-view diffusion prior. arXiv preprint arXiv:2308.13223 (2023) +74. Zhao, R., Wang, Z., Wang, Y., Zhou, Z., Zhu, J.: Flexidreamer: Single image-to-3d generation with flexicubes. arXiv preprint arXiv:2404.00987 (2024) +75. Zhou, L., Shih, A., Meng, C., Ermon, S.: Dreampropeller: Supercharge text-to-3d generation with parallel sampling. arXiv preprint arXiv:2311.17082 (2023) \ No newline at end of file diff --git a/2024/ScaleDreamer_ Scalable Text-to-3D Synthesis with Asynchronous Score Distillation/images.zip b/2024/ScaleDreamer_ Scalable Text-to-3D Synthesis with Asynchronous Score Distillation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..97f13c46ad93263c9da96a1b3de21877152c66c4 --- /dev/null +++ b/2024/ScaleDreamer_ Scalable Text-to-3D Synthesis with Asynchronous Score Distillation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0204e2e601c3a327ca3226a0aaad111c0aaad280d4300a815d518c915073d683 +size 547126 diff --git a/2024/ScaleDreamer_ Scalable Text-to-3D Synthesis with Asynchronous Score Distillation/layout.json b/2024/ScaleDreamer_ Scalable Text-to-3D Synthesis with Asynchronous Score Distillation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b8a9f5d874f1a2b8f59175cd562caa981f6e5e55 --- /dev/null +++ b/2024/ScaleDreamer_ Scalable Text-to-3D Synthesis with Asynchronous Score Distillation/layout.json @@ -0,0 +1,12732 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 149, + 111, + 465, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 111, + 465, + 148 + ], + "spans": [ + { + "bbox": [ + 149, + 111, + 465, + 148 + ], + "type": "text", + "content": "ScaleDreamer: Scalable Text-to-3D Synthesis with Asynchronous Score Distillation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 190, + 167, + 421, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 167, + 421, + 180 + ], + "spans": [ + { + "bbox": [ + 190, + 167, + 421, + 180 + ], + "type": "text", + "content": "Zhiyuan Ma" + }, + { + "bbox": [ + 190, + 167, + 421, + 180 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 190, + 167, + 421, + 180 + ], + "type": "text", + "content": ", Yuxiang Wei" + }, + { + "bbox": [ + 190, + 167, + 421, + 180 + ], + "type": "inline_equation", + "content": "^{1,5}" + }, + { + "bbox": [ + 190, + 167, + 421, + 180 + ], + "type": "text", + "content": ", Yabin Zhang" + }, + { + "bbox": [ + 190, + 167, + 421, + 180 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 190, + 167, + 421, + 180 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 186, + 183, + 427, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 183, + 427, + 198 + ], + "spans": [ + { + "bbox": [ + 186, + 183, + 427, + 198 + ], + "type": "text", + "content": "Xiangyu Zhu" + }, + { + "bbox": [ + 186, + 183, + 427, + 198 + ], + "type": "inline_equation", + "content": "^{3,4}" + }, + { + "bbox": [ + 186, + 183, + 427, + 198 + ], + "type": "text", + "content": ", Zhen Lei" + }, + { + "bbox": [ + 186, + 183, + 427, + 198 + ], + "type": "inline_equation", + "content": "^{1,2,3,4^{\\star}}" + }, + { + "bbox": [ + 186, + 183, + 427, + 198 + ], + "type": "text", + "content": ", and Lei Zhang" + }, + { + "bbox": [ + 186, + 183, + 427, + 198 + ], + "type": "inline_equation", + "content": "^{1^{\\star}}" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 206, + 207, + 406, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 207, + 406, + 218 + ], + "spans": [ + { + "bbox": [ + 206, + 207, + 406, + 218 + ], + "type": "text", + "content": "1 The Hong Kong Polytechnic University, PolyU" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 182, + 218, + 431, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 218, + 431, + 228 + ], + "spans": [ + { + "bbox": [ + 182, + 218, + 431, + 228 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 182, + 218, + 431, + 228 + ], + "type": "text", + "content": " Center for Artificial Intelligence and Robotics, HKISI CAS" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 149, + 229, + 465, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 229, + 465, + 240 + ], + "spans": [ + { + "bbox": [ + 149, + 229, + 465, + 240 + ], + "type": "text", + "content": "3 State Key Laboratory of Multimodal Artificial Intelligence Systems, CASIA" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 135, + 240, + 478, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 240, + 478, + 251 + ], + "spans": [ + { + "bbox": [ + 135, + 240, + 478, + 251 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 135, + 240, + 478, + 251 + ], + "type": "text", + "content": " School of Artificial Intelligence, University of Chinese Academy of Sciences, UCAS" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 229, + 252, + 384, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 229, + 252, + 384, + 262 + ], + "spans": [ + { + "bbox": [ + 229, + 252, + 384, + 262 + ], + "type": "text", + "content": "5 Harbin Institute of Technology, HIT" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 159, + 286, + 452, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 286, + 452, + 561 + ], + "spans": [ + { + "bbox": [ + 159, + 286, + 452, + 561 + ], + "type": "text", + "content": "Abstract. By leveraging the text-to-image diffusion prior, score distillation can synthesize 3D contents without paired text-3D training data. Instead of spending hours of online optimization per text prompt, recent studies have been focused on learning a text-to-3D generative network for amortizing multiple text-3D relations, which can synthesize 3D contents in seconds. However, existing score distillation methods are hard to scale up to a large amount of text prompts due to the difficulties in aligning pretrained diffusion prior with the distribution of rendered images from various text prompts. Current state-of-the-arts such as Variational Score Distillation finetune the pretrained diffusion model to minimize the noise prediction error so as to align the distributions, which are however unstable to train and will impair the model's comprehension capability to numerous text prompts. Based on the observation that the diffusion models tend to have lower noise prediction errors at earlier timesteps, we propose Asynchronous Score Distillation (ASD), which minimizes the noise prediction error by shifting the diffusion timestep to earlier ones. ASD is stable to train and can scale up to " + }, + { + "bbox": [ + 159, + 286, + 452, + 561 + ], + "type": "inline_equation", + "content": "100k" + }, + { + "bbox": [ + 159, + 286, + 452, + 561 + ], + "type": "text", + "content": " prompts. It reduces the noise prediction error without changing the weights of pre-trained diffusion model, thus keeping its strong comprehension capability to prompts. We conduct extensive experiments using different text-to-3D architectures, including Hyper-iNGP and 3DConvNet. The results demonstrate ASD's effectiveness in stable 3D generator training, high-quality 3D content synthesis, and its superior prompt-consistency, especially under large prompt corpus. Code is available at https://github.com/theEricMa/ScaleDreamer." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 160, + 571, + 412, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 571, + 412, + 582 + ], + "spans": [ + { + "bbox": [ + 160, + 571, + 412, + 582 + ], + "type": "text", + "content": "Keywords: Text-to-3D " + }, + { + "bbox": [ + 160, + 571, + 412, + 582 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 571, + 412, + 582 + ], + "type": "text", + "content": " Score Distillation " + }, + { + "bbox": [ + 160, + 571, + 412, + 582 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 571, + 412, + 582 + ], + "type": "text", + "content": " Diffusion Model" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 600, + 229, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 229, + 612 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 229, + 612 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 623, + 481, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 623, + 481, + 648 + ], + "spans": [ + { + "bbox": [ + 130, + 623, + 481, + 648 + ], + "type": "text", + "content": "Text-to-3D aims to generate realistic 3D contents from the given textual descriptions [40], which is particularly useful in many applications such as virtual" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 240, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 240, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 240, + 665 + ], + "type": "text", + "content": "* Corresponding authors." + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 283 + ], + "type": "text", + "content": "reality [61] and game design [23]. The main challenge of this task, however, lies in how to generate high-quality 3D contents conditioned on the abstract and diverse textual descriptions. Many existing text-to-3D methods [11, 12, 26–29, 32, 35, 37, 40, 42, 58, 59, 75] are optimization-based ones, which distill the guidance from the powerful pretrained text-to-image diffusion models [5, 11, 26, 32, 42, 44, 73] via score distillation [40, 59, 62, 72]. In general, these methods employ the KL divergence to reduce the discrepancy between the distribution of rendered images and the desired image distribution embedded in the 2D diffusion prior, while they differ in how to use the pretrained diffusion prior to model the distribution of rendered images. Extensive efforts have been made to explore prompt-specific optimization of various 3D representations, including implicit radiance fields [40], explicit radiance fields [29, 37, 59], DmTets [56, 74] and 3D Gaussians [9]. Typically, tens of minutes to hours are needed to optimize a single 3D representation for one prompt to achieve the desired result." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 287, + 482, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 287, + 482, + 455 + ], + "spans": [ + { + "bbox": [ + 130, + 287, + 482, + 455 + ], + "type": "text", + "content": "Compared to the aforementioned optimization-based text-to-3D methods, learning-based methods [6, 21, 31, 36, 43, 54, 65] can largely reduce the computational cost by training a text-conditioned 3D generative network. With the availability of 3D object collections [10, 63, 71], a deep network can be trained in a supervised manner so that 3D outputs can be generated in several seconds. Unfortunately, the size of existing text-3D datasets is far from sufficient compared to text-image datasets [45], limiting the text-to-3D generation performance of trained models. Inspired by the optimization-based text-to-3D methods that use pretrained 2D diffusion models, efforts have been made to train text-to-3D networks by using 2D diffusion models as supervisors [33, 41, 65] without using text-3D pairs. For example, a text-conditioned 3D hyper-network is trained in ATT3D [33] via Score Distillation Sampling (SDS) [40]. Nevertheless, this method suffers from numerical instability, which has been observed in subsequent studies [41, 65] that apply SDS to different 3D generator networks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 459, + 482, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 459, + 482, + 602 + ], + "spans": [ + { + "bbox": [ + 130, + 459, + 482, + 602 + ], + "type": "text", + "content": "Despite the success of score distillation in optimization-based text-to-3D generation [40, 59, 72], its application to learning-based text-to-3D frameworks is rather limited because of the unstable training or unsatisfactory results. We argue that the primary challenge lies in how to efficiently and effectively leverage the pretrained 2D diffusion prior to represent the distribution of images rendered by the 3D generator. For example, SDS [40] forces the rendered images to adhere to the Dirac distribution, which causes numerical instability in 3D generator training [33, 65]. Variational Score Distillation (VSD) [59] finetunes the 2D diffusion prior for distribution alignment via minimizing the noise prediction error. However, the finetuning changes the pretrained diffusion network and hurts its comprehension capability to numerous text prompts, leading to mode collapse when the size of text prompts is extended." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "To address the above mentioned issues, we propose Asynchronous Score Distillation (ASD). Like VSD, ASD aims to minimize the noise prediction error. Different from VSD, ASD does not finetune the pretrained 2D diffusion network; instead, it achieves the goal by shifting the diffusion timestep. This is based on the observation that diffusion networks will have smaller noise prediction errors" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "text", + "content": "Z.Ma et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "text", + "content": "in earlier timesteps [67]; therefore, we can shift the timestep to an earlier step to achieve a similar goal to VSD, i.e., reducing the noise prediction error. In this way, the diffusion network can be frozen in training and its strong text comprehension capability can be well-preserved. The shifted timesteps can be well sampled from a pre-defined range for most prompts. To evaluate the performance of ASD, we conduct extensive experiments by using two types of generator architectures, i.e. Hyper-iNGP [33] and 3DConv-Net [4], across various prompt corpus sizes. We conduct extensive experiments to evaluate the superiority of ASD to previous methods, including the stable training of 3D generators, the production of high-quality 3D outputs, the high content fidelity to input prompts, as well as its scalability to larger corpus sizes, e.g., 100k prompts. We also show that ASD can work with other 2D diffusion models such as MVDream [48], and can be used to train more 3D generators such as Triplane-Transformer [17]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 292, + 263, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 292, + 263, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 292, + 263, + 304 + ], + "type": "text", + "content": "2 Literature Review" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 318, + 332, + 330 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 318, + 332, + 330 + ], + "spans": [ + { + "bbox": [ + 132, + 318, + 332, + 330 + ], + "type": "text", + "content": "2.1 Text-to-3D with Score Distillation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 340, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 482, + 449 + ], + "type": "text", + "content": "Text-to-3D takes text description, a.k.a. text prompt " + }, + { + "bbox": [ + 130, + 340, + 482, + 449 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 340, + 482, + 449 + ], + "type": "text", + "content": ", as input, and outputs 3D representation " + }, + { + "bbox": [ + 130, + 340, + 482, + 449 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 340, + 482, + 449 + ], + "type": "text", + "content": " that renders high-fidelity images at any camera view " + }, + { + "bbox": [ + 130, + 340, + 482, + 449 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 130, + 340, + 482, + 449 + ], + "type": "text", + "content": ". Thanks to the powerful text-to-image diffusion models [32,42,44,48,73], we can optimize " + }, + { + "bbox": [ + 130, + 340, + 482, + 449 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 340, + 482, + 449 + ], + "type": "text", + "content": " to align with " + }, + { + "bbox": [ + 130, + 340, + 482, + 449 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 340, + 482, + 449 + ], + "type": "text", + "content": " by computing the objective " + }, + { + "bbox": [ + 130, + 340, + 482, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\boldsymbol{x},y)" + }, + { + "bbox": [ + 130, + 340, + 482, + 449 + ], + "type": "text", + "content": " on the rendered image " + }, + { + "bbox": [ + 130, + 340, + 482, + 449 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x} = g(\\theta ,\\pi)" + }, + { + "bbox": [ + 130, + 340, + 482, + 449 + ], + "type": "text", + "content": " from camera view " + }, + { + "bbox": [ + 130, + 340, + 482, + 449 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 130, + 340, + 482, + 449 + ], + "type": "text", + "content": ". Through differential rendering, " + }, + { + "bbox": [ + 130, + 340, + 482, + 449 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 340, + 482, + 449 + ], + "type": "text", + "content": " can be updated with the gradient " + }, + { + "bbox": [ + 130, + 340, + 482, + 449 + ], + "type": "inline_equation", + "content": "\\nabla_{\\theta}\\mathcal{L}(\\theta ,y) = \\frac{\\partial\\mathcal{L}(\\boldsymbol{x},y)}{\\partial\\boldsymbol{x}}\\frac{\\partial\\boldsymbol{x}}{\\partial\\theta}" + }, + { + "bbox": [ + 130, + 340, + 482, + 449 + ], + "type": "text", + "content": ". This technique is generally termed as score distillation. Unlike data-driven techniques [6,21,31,43,54], score distillation approaches [8, 19, 24, 29, 40, 53, 59, 72] can produce high-quality 3D content without the need for 3D training datasets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 449, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 533 + ], + "type": "text", + "content": "Prompt-Specific Text-to-3D. Existing score distillation methods [40, 59, 72] were originally developed to output a single 3D result " + }, + { + "bbox": [ + 130, + 449, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 449, + 482, + 533 + ], + "type": "text", + "content": " for a single text prompt " + }, + { + "bbox": [ + 130, + 449, + 482, + 533 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 449, + 482, + 533 + ], + "type": "text", + "content": " via online optimization: " + }, + { + "bbox": [ + 130, + 449, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\min_{\\theta} \\mathbb{E}_{\\pi, \\boldsymbol{x} = g(\\theta, \\pi)}[\\mathcal{L}(\\boldsymbol{x}, y)]" + }, + { + "bbox": [ + 130, + 449, + 482, + 533 + ], + "type": "text", + "content": ". The utilized 3D representations, e.g., NeRF [39, 40], DmTet [47, 74], and 3D Gaussian [20, 30, 51, 53, 57, 70], are not designed to render scenes from varying text prompts. Therefore, the optimization has to be conducted again for newly provided text prompts. The optimization process typically costs tens of minutes to hours." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Prompt-Amortized Text-to-3D. To mitigate the computational costs in prompt-specific methods, recent studies [25, 33, 41, 65] have attempted to use score distillation to train a text-to-3D generator " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\theta = \\mathcal{G}(y)" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": ", aiming to generate multiple 3D representations from a set of text prompts " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "S_{y} = \\{y\\}" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": ". These methods can generate 3D results from queried text prompt in seconds. As proposed by ATT3D [33], the 3D generator training is performed by minimizing " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\min_{\\mathcal{G}} \\mathbb{E}_{\\pi, y \\in S_{y}, x = g(\\mathcal{G}(y), \\pi)}[\\mathcal{L}(\\boldsymbol{x}, y)]" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " over all text prompts. Unlike data-driven approaches [17, 52, 66], score distillation bypasses the scarcity of text-3D data pairs because the 2D diffusion prior can offer the guidance to align the 3D output with the input text prompt. However, its application is currently restricted to training the 3D generator within a limited range of text prompts." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 390, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 390, + 91, + 447, + 101 + ], + "type": "text", + "content": "ScaleDreamer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 373, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 373, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 373, + 128 + ], + "type": "text", + "content": "2.2 Representative Score Distillation Methods" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 482, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 482, + 253 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 482, + 253 + ], + "type": "text", + "content": "Denote by " + }, + { + "bbox": [ + 130, + 137, + 482, + 253 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 137, + 482, + 253 + ], + "type": "text", + "content": " the 2D diffusion prior [44,48] and by " + }, + { + "bbox": [ + 130, + 137, + 482, + 253 + ], + "type": "inline_equation", + "content": "p^{\\phi}(\\boldsymbol{x} \\mid y)" + }, + { + "bbox": [ + 130, + 137, + 482, + 253 + ], + "type": "text", + "content": " the text-conditioned image distribution embedded within " + }, + { + "bbox": [ + 130, + 137, + 482, + 253 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 137, + 482, + 253 + ], + "type": "text", + "content": ", the objectives of most existing score distillation methods can be generally concluded as minimizing the objective " + }, + { + "bbox": [ + 130, + 137, + 482, + 253 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\theta, y) = \\mathbb{E}_{\\pi, t, \\epsilon, \\boldsymbol{x} = g(\\theta, \\pi)}\\left[\\omega(t) D_{\\mathrm{KL}}\\left(q_t^\\theta(\\boldsymbol{x}_t \\mid \\pi) \\| p_t^\\phi(\\boldsymbol{x}_t \\mid y^\\pi)\\right)\\right]" + }, + { + "bbox": [ + 130, + 137, + 482, + 253 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 137, + 482, + 253 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{KL}}" + }, + { + "bbox": [ + 130, + 137, + 482, + 253 + ], + "type": "text", + "content": " denotes KL divergence, " + }, + { + "bbox": [ + 130, + 137, + 482, + 253 + ], + "type": "inline_equation", + "content": "q_t^\\theta(\\boldsymbol{x}_t \\mid \\pi)" + }, + { + "bbox": [ + 130, + 137, + 482, + 253 + ], + "type": "text", + "content": " denotes the distribution of images " + }, + { + "bbox": [ + 130, + 137, + 482, + 253 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}" + }, + { + "bbox": [ + 130, + 137, + 482, + 253 + ], + "type": "text", + "content": " rendered at camera view " + }, + { + "bbox": [ + 130, + 137, + 482, + 253 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 130, + 137, + 482, + 253 + ], + "type": "text", + "content": " at diffusion timestep " + }, + { + "bbox": [ + 130, + 137, + 482, + 253 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 137, + 482, + 253 + ], + "type": "text", + "content": " [15], and the same for " + }, + { + "bbox": [ + 130, + 137, + 482, + 253 + ], + "type": "inline_equation", + "content": "p_t^\\phi(\\boldsymbol{x}_t \\mid y)" + }, + { + "bbox": [ + 130, + 137, + 482, + 253 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 137, + 482, + 253 + ], + "type": "inline_equation", + "content": "\\omega(t)" + }, + { + "bbox": [ + 130, + 137, + 482, + 253 + ], + "type": "text", + "content": " is a timestep-dependent weight [40]. " + }, + { + "bbox": [ + 130, + 137, + 482, + 253 + ], + "type": "inline_equation", + "content": "y^\\pi" + }, + { + "bbox": [ + 130, + 137, + 482, + 253 + ], + "type": "text", + "content": " denotes the view-dependent strategy [44] or view-awareness [42, 48] to prompt the different camera views [40]. To minimize this objective, the gradient w.r.t. " + }, + { + "bbox": [ + 130, + 137, + 482, + 253 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 137, + 482, + 253 + ], + "type": "text", + "content": " can be calculated as per [59]:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 139, + 261, + 481, + 308 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 261, + 481, + 308 + ], + "spans": [ + { + "bbox": [ + 139, + 261, + 481, + 308 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\theta} \\mathcal {L} (\\theta , y) = \\mathbb {E} _ {\\pi , t, \\epsilon} \\left[ \\omega (t) \\left(\\underbrace {- \\sigma_ {t} \\nabla_ {\\boldsymbol {x} _ {t}} \\log p _ {t} ^ {\\phi} (\\boldsymbol {x} _ {t} \\mid y ^ {\\pi})} _ {\\epsilon_ {\\phi} (\\boldsymbol {x} _ {t}; t, y ^ {\\pi})} - \\underbrace {(- \\sigma_ {t} \\nabla_ {\\boldsymbol {x} _ {t}} \\log q _ {t} ^ {\\theta} (\\boldsymbol {x} _ {t} \\mid \\pi))} _ {\\epsilon_ {\\theta} (\\boldsymbol {x} _ {t}; t, \\pi , y)}\\right) \\frac {\\partial \\boldsymbol {x}}{\\partial \\theta} \\right], \\tag {1}", + "image_path": "d0c2ec8da4a9974cb580ff2cf2fac1f02b8268bfb94f6bb6584b9ed678ae95d9.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 317, + 482, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 317, + 482, + 425 + ], + "spans": [ + { + "bbox": [ + 130, + 317, + 482, + 425 + ], + "type": "text", + "content": "where the first term " + }, + { + "bbox": [ + 130, + 317, + 482, + 425 + ], + "type": "inline_equation", + "content": "-\\sigma_t\\nabla_{\\boldsymbol{x}_t}\\log p_t^\\phi (\\boldsymbol{x}_t\\mid y^\\pi)" + }, + { + "bbox": [ + 130, + 317, + 482, + 425 + ], + "type": "text", + "content": " corresponds to the score function [50] of the desired image distribution, and it can be achieved by predicting the noise " + }, + { + "bbox": [ + 130, + 317, + 482, + 425 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0,I)" + }, + { + "bbox": [ + 130, + 317, + 482, + 425 + ], + "type": "text", + "content": " in the noisy image " + }, + { + "bbox": [ + 130, + 317, + 482, + 425 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_t = \\alpha_t\\boldsymbol {x} + \\sigma_t\\boldsymbol{\\epsilon}" + }, + { + "bbox": [ + 130, + 317, + 482, + 425 + ], + "type": "text", + "content": " using the pretrained 2D diffusion model " + }, + { + "bbox": [ + 130, + 317, + 482, + 425 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}(\\pmb {x}_t;t,y^\\pi)" + }, + { + "bbox": [ + 130, + 317, + 482, + 425 + ], + "type": "text", + "content": " [44,48]. Existing score distillation methods [40,59,72] mainly differ in how to model " + }, + { + "bbox": [ + 130, + 317, + 482, + 425 + ], + "type": "inline_equation", + "content": "-\\sigma_t\\nabla x_t\\log q_t^\\theta (x_t\\mid \\pi)" + }, + { + "bbox": [ + 130, + 317, + 482, + 425 + ], + "type": "text", + "content": ", which corresponds to the score function of the distribution of rendered images " + }, + { + "bbox": [ + 130, + 317, + 482, + 425 + ], + "type": "inline_equation", + "content": "q^{\\theta}(\\pmb {x}\\mid \\pi)" + }, + { + "bbox": [ + 130, + 317, + 482, + 425 + ], + "type": "text", + "content": ". We denote this term in Eq. 1 as " + }, + { + "bbox": [ + 130, + 317, + 482, + 425 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\pmb {x}_t;t,\\pi ,y)" + }, + { + "bbox": [ + 130, + 317, + 482, + 425 + ], + "type": "text", + "content": " in the following context, since it represents a diffusion model that corresponds to " + }, + { + "bbox": [ + 130, + 317, + 482, + 425 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 317, + 482, + 425 + ], + "type": "text", + "content": ". A summary of the objectives of major score distillation methods is shown in Tab. 1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": "The objective of Score Distillation Sampling (SDS) [40] is " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\nabla_{\\theta}\\mathcal{L}_{\\mathrm{SDS}}(\\theta ,y)\\triangleq" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{\\pi ,t,\\epsilon}\\left[\\omega (t)(\\pmb{\\epsilon}_{\\phi}(\\pmb{x}_t;t,y^\\pi) - \\pmb {\\epsilon})\\frac{\\partial\\pmb{x}}{\\partial\\theta}\\right]" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": ", which approximates the term " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\pmb{\\epsilon}_{\\theta}(\\pmb {x}_t;t,\\pi ,y)" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": " in Eq. 1 as the ground-truth noise " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\pmb{\\epsilon}" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": ". That is, SDS assumes that " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "q^{\\theta}(\\pmb {x}\\mid \\pi)" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": " adheres to a Dirac distribution " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\delta (\\pmb {x} - g(\\theta ,\\pi))" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": " [59], which is characterized by a non-zero density at the singular point of " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\pmb {x} = g(\\theta ,\\pi)" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": " and zero density everywhere else. However, updating " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": " under the Dirac distribution might be troublesome [59]. We may need to set the CFG (Classifier Free Guidance) [16] as high as 100 for model convergence, which will produce excessively large gradients and lead to unstable optimization. This problem is alleviated by Classifier Score Distillation (CSD) [72], which uses the classifier component [16] in SDS as the objective: " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\nabla_{\\theta}\\mathcal{L}_{\\mathrm{CSD}}(\\theta ,y)\\triangleq \\mathbb{E}_{\\pi ,t,\\epsilon}\\left[\\omega (t)(\\pmb{\\epsilon}_{\\phi}(\\pmb {x}_t;t,y^\\pi) - \\pmb{\\epsilon}_{\\phi}(\\pmb {x}_t;t))\\frac{\\partial\\pmb{x}}{\\partial\\theta}\\right]" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": ". CSD can be regraded as straightforwardly using the unconditional term of the diffusion prior " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\pmb{\\epsilon}_{\\phi}(\\pmb{x}_t;t)" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": " to represent " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\pmb{\\epsilon}_{\\theta}(\\pmb {x}_t;t,\\pi ,y)" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": " in Eq. 1. Unfortunately, in the case of prompt-amortized training, this term may not provide effective gradient, because " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\pmb{\\epsilon}_{\\phi}(\\pmb {x}_t;t)" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": " is unconditional to the provided text-prompts. In contrast, Variational Score Distillation (VSD) [59] models " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\pmb{\\epsilon}_{\\theta}(\\pmb {x}_t;t,\\pi ,y)" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": " with another text-aware diffusion model " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\pmb{\\epsilon}_{\\phi^{\\prime}}(\\mathbf{x}_t;t,\\pi ,y)" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": ", leading to " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\nabla_{\\theta}\\mathcal{L}_{\\mathrm{VSD}}(\\theta ,y)\\triangleq" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{\\pi ,t,\\epsilon}\\left[\\omega (t)(\\pmb{\\epsilon}_{\\phi}(\\pmb {x}_t;t,y^\\pi) - \\pmb{\\epsilon}_{\\phi^{\\prime}}(\\pmb {x}_t;t,\\pi ,y))\\frac{\\partial\\pmb{x}}{\\partial\\theta}\\right]" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\pmb{\\epsilon}_{\\phi^{\\prime}}(\\mathbf{x}_t;t,\\pi ,y)" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": " is achieved by finetuning the pretrained 2D diffusion prior " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\pmb{\\epsilon}_{\\phi}(\\pmb {x}_t;t,y^\\pi)" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": " to align with the rendered image distribution " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "q^{\\theta}(\\pmb {x}\\mid \\pi)" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": " via parameter efficient adaptation [18]. In" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "text", + "content": "Z.Ma et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 175, + 114, + 440, + 173 + ], + "blocks": [ + { + "bbox": [ + 175, + 114, + 440, + 173 + ], + "lines": [ + { + "bbox": [ + 175, + 114, + 440, + 173 + ], + "spans": [ + { + "bbox": [ + 175, + 114, + 440, + 173 + ], + "type": "table", + "html": "
MethodGradient of L(x,y) w.r.t. x = g(θ,π)
SDS [40]E_t,ε[ω(t) (εφ(xt; t, yπ) - ε)]
CSD [72]E_t,ε[ω(t) (εφ(xt; t, yπ) - εφ(xt, t))]
VSD [59]E_t,ε[ω(t) (εφ(xt; t, yπ) - εφ'(xt; t, π, y))]
ASD (Ours)E_t,ε[ω(t) (εφ(xt; t, yπ) - εφ(xt+Δt; t+Δt, yπ))]
", + "image_path": "0ab3981ab01452839abaf8059133e5ad95175a690bbbd5686bc991283b5f04d0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 174, + 479, + 195 + ], + "lines": [ + { + "bbox": [ + 130, + 174, + 479, + 195 + ], + "spans": [ + { + "bbox": [ + 130, + 174, + 479, + 195 + ], + "type": "text", + "content": "Table 1: Objectives of representative score distillation methods. ASD introduces " + }, + { + "bbox": [ + 130, + 174, + 479, + 195 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 130, + 174, + 479, + 195 + ], + "type": "text", + "content": " alongside " + }, + { + "bbox": [ + 130, + 174, + 479, + 195 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 174, + 479, + 195 + ], + "type": "text", + "content": " to align with the rendered image distribution " + }, + { + "bbox": [ + 130, + 174, + 479, + 195 + ], + "type": "inline_equation", + "content": "q^{\\theta}(\\boldsymbol{x} \\mid \\pi)" + }, + { + "bbox": [ + 130, + 174, + 479, + 195 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 201, + 479, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 201, + 479, + 227 + ], + "spans": [ + { + "bbox": [ + 130, + 201, + 479, + 227 + ], + "type": "text", + "content": "practice, this is conducted by alternatively optimizing " + }, + { + "bbox": [ + 130, + 201, + 479, + 227 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 201, + 479, + 227 + ], + "type": "text", + "content": " and finetuning " + }, + { + "bbox": [ + 130, + 201, + 479, + 227 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 201, + 479, + 227 + ], + "type": "text", + "content": " with the noise prediction objective " + }, + { + "bbox": [ + 130, + 201, + 479, + 227 + ], + "type": "inline_equation", + "content": "\\| \\pmb{\\epsilon}_{\\phi}(\\pmb{x}_t;t,y) - \\pmb{\\epsilon}\\| _2^2" + }, + { + "bbox": [ + 130, + 201, + 479, + 227 + ], + "type": "text", + "content": " [15] such that:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 176, + 234, + 480, + 248 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 234, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 176, + 234, + 480, + 248 + ], + "type": "interline_equation", + "content": "\\mathbb {E} _ {\\pi , t, \\epsilon} \\left[ \\| \\epsilon_ {\\phi J} \\left(\\boldsymbol {x} _ {t}; t, \\pi , y\\right) - \\epsilon \\| _ {2} ^ {2} \\right] \\leq \\mathbb {E} _ {\\pi , t, \\epsilon} \\left[ \\| \\epsilon_ {\\phi} \\left(\\boldsymbol {x} _ {t}; t, y ^ {\\pi}\\right) - \\epsilon \\| _ {2} ^ {2} \\right]. \\tag {2}", + "image_path": "b39ba841915a1b338149930b1eb02174c71fa8d48a3093091d183d2a8ae63d36.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 256, + 479, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 256, + 479, + 280 + ], + "spans": [ + { + "bbox": [ + 130, + 256, + 479, + 280 + ], + "type": "text", + "content": "The above equation reveals that a better alignment with the distribution of " + }, + { + "bbox": [ + 130, + 256, + 479, + 280 + ], + "type": "inline_equation", + "content": "q^{\\theta}(\\pmb{x} \\mid \\pi)" + }, + { + "bbox": [ + 130, + 256, + 479, + 280 + ], + "type": "text", + "content": " can be achieved by a more accurate noise prediction." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 281, + 480, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 480, + 388 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 480, + 388 + ], + "type": "text", + "content": "While VSD achieves state-of-the-art results in prompt-specific text-to-3D [14, 59], it changes the diffusion prior's parameters by alternately optimizing " + }, + { + "bbox": [ + 130, + 281, + 480, + 388 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 281, + 480, + 388 + ], + "type": "text", + "content": " and finetuning " + }, + { + "bbox": [ + 130, + 281, + 480, + 388 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 281, + 480, + 388 + ], + "type": "text", + "content": ". This forms a bi-level optimization, known to be problematic in generative adversarial training [55], and may be troublesome for training prompt-amortized text-to-3D models, because the change of pre-trained diffusion model might impairs its comprehension capability on a wide range of text-prompts. In specific, the pre-trained 2D diffusion model may have to sacrifice its generation capability in order to align with the distribution of rendered images, making it fail to produce good gradient for training the 3D generator." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 406, + 386, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 406, + 386, + 420 + ], + "spans": [ + { + "bbox": [ + 132, + 406, + 386, + 420 + ], + "type": "text", + "content": "3 Asynchronous Score Distillation (ASD)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 430, + 250, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 250, + 442 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 250, + 442 + ], + "type": "text", + "content": "3.1 Objective of ASD" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 450, + 480, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 480, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 480, + 545 + ], + "type": "text", + "content": "From the above discussions in Sec. 2.2, it can be seen that one key issue in VSD is to minimize the noise prediction error so that the model output can be aligned with the desired distribution of rendered images. VSD achieves this goal via finetuning the pre-trained 2D diffusion model, which however sacrifices its comprehension capability on text prompts. One interesting question is: can we minimize the noise prediction error without changing the pre-trained diffusion network weights? Fortunately, we find that this is possible and in this section we present a new objective function to achieve this goal." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "content": "Recall that diffusion models solve the stochastic differential equation [50] via reversing the noise added along different stages, a.k.a. diffusion timestep " + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "inline_equation", + "content": "t \\in \\{T_{\\max}, \\dots, T_{\\min}\\}" + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "content": " via " + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_t = \\alpha_t \\boldsymbol{x} + \\sigma_t \\boldsymbol{\\epsilon}" + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "content": " [15]. The influence of the noise " + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0, I)" + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "content": " on the image " + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}" + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "content": " is incrementally reduced as the process progresses from the initial timestep " + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "inline_equation", + "content": "T_{\\max}" + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "content": " to the final timestep " + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "inline_equation", + "content": "T_{\\min}" + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "content": ", which is controlled by the scalars " + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\alpha_t" + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\sigma_t" + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "content": ". Consequently, the diffusion model's noise prediction accuracy will vary with the timestep " + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "content": ", at which the identical noise " + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "content": " is added. To evaluate this, we consider a diffusion model with fixed image " + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}" + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "content": ", noise " + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "content": " and condition " + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "content": ", but varied timestep " + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "content": ". We denote such a diffusion model as " + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\epsilon(t)" + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "content": " and explore how its prediction error, denoted by " + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "inline_equation", + "content": "e(t) = \\| \\epsilon(t) - \\epsilon \\|_2^2" + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "content": ", changes with " + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "type": "text", + "content": "ScaleDreamer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": "The model " + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "inline_equation", + "content": "\\epsilon(t)" + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": " can be a pre-trained 2D diffusion model (such as Stable Diffusion [44]). We denote by " + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "inline_equation", + "content": "\\epsilon_{PT}(t)" + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": " such a model, and investigate the behaviour of its noise prediction error, denoted by " + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "inline_equation", + "content": "e_{PT}(t)" + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": ". In Fig. 1, we plot the curve (i.e., the blue colored curve) of " + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "inline_equation", + "content": "e_{PT}(t)" + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": " versus " + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": ". We use a corpus with 15 text prompts from Magic3D [40] to draw this curve. For each prompt " + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": ", we generate 16 images with VSD [59]. Then for each image " + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": ", we apply one instance of Gaussian noise " + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "inline_equation", + "content": "\\pmb{\\epsilon}" + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": " and conduct a single diffusion step with 100 distinct timesteps. The average noise reconstruction error is then calculated for these timesteps across all prompts and images. We can see from the curve of " + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "inline_equation", + "content": "e_{PT}(t)" + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": " that earlier diffusion timesteps (e.g., timestep 600) will have lower noise prediction error than later timesteps (e.g., timestep 200). Such a trend holds for almost every image sample " + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": " and noise sample " + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "inline_equation", + "content": "\\pmb{\\epsilon}" + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": " because the well-trained diffusion model is frozen in our case. Since the noise prediction error declines from " + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{min}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": " (i.e., late diffusion timestep) to " + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{max}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": " (i.e., early diffusion timestep), we can conclude that for a given timestep " + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": " and a timestep shift " + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "inline_equation", + "content": "0 \\leq \\Delta t \\leq T_{\\mathrm{max}} - t" + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": ", the following inequality holds:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 311, + 481, + 326 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 311, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 160, + 311, + 481, + 326 + ], + "type": "interline_equation", + "content": "\\mathbb {E} _ {\\pi , t, \\epsilon} \\left[ \\| \\epsilon_ {\\phi} \\left(\\boldsymbol {x} _ {t + \\Delta t}; t + \\Delta t, y ^ {\\pi}\\right) - \\epsilon \\| _ {2} ^ {2} \\right] \\leq \\mathbb {E} _ {\\pi , t, \\epsilon} \\left[ \\| \\epsilon_ {\\phi} \\left(\\boldsymbol {x} _ {t}; t, y ^ {\\pi}\\right) - \\epsilon \\| _ {2} ^ {2} \\right], \\tag {3}", + "image_path": "cc1f5b9c355713367fd65df28caac2c8ddd279fae1ea86b12da5a2b35b0c1695.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 340, + 481, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 481, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 481, + 365 + ], + "type": "text", + "content": "which implies that more accurate noise predictions can be achieved at earlier diffusion timesteps." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 367, + 482, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 367, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 367, + 482, + 475 + ], + "type": "text", + "content": "The above property of diffusion models has also been observed by Yang et al. [68], who indicated that as the timestep shifts from " + }, + { + "bbox": [ + 130, + 367, + 482, + 475 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{max}}" + }, + { + "bbox": [ + 130, + 367, + 482, + 475 + ], + "type": "text", + "content": " towards " + }, + { + "bbox": [ + 130, + 367, + 482, + 475 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{min}}" + }, + { + "bbox": [ + 130, + 367, + 482, + 475 + ], + "type": "text", + "content": ", the variance in noise prediction increases, as evidenced by the rising Lipschitz constants, which suggests an increased instability in noise prediction and larger noise prediction errors. Such a behavior can be observed in both " + }, + { + "bbox": [ + 130, + 367, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 367, + 482, + 475 + ], + "type": "text", + "content": "-prediction and " + }, + { + "bbox": [ + 130, + 367, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\pmb{v}" + }, + { + "bbox": [ + 130, + 367, + 482, + 475 + ], + "type": "text", + "content": "-prediction models, as well as in 2D and 3D diffusion models (please refer to supplementary material for details). This can be intuitively explained as follows. When " + }, + { + "bbox": [ + 130, + 367, + 482, + 475 + ], + "type": "inline_equation", + "content": "t \\to T_{\\mathrm{max}}" + }, + { + "bbox": [ + 130, + 367, + 482, + 475 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 367, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t = \\alpha_t \\pmb{x} + \\sigma_t \\pmb{\\epsilon} \\to \\pmb{\\epsilon}" + }, + { + "bbox": [ + 130, + 367, + 482, + 475 + ], + "type": "text", + "content": ", then it is easier to achieve " + }, + { + "bbox": [ + 130, + 367, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}(\\pmb{x}_t; t, y^{\\pi}) \\approx \\epsilon" + }, + { + "bbox": [ + 130, + 367, + 482, + 475 + ], + "type": "text", + "content": " because the model can manage to copy the input as the output." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 478, + 482, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 478, + 482, + 562 + ], + "spans": [ + { + "bbox": [ + 130, + 478, + 482, + 562 + ], + "type": "text", + "content": "The similarity between Eq. 3 and the fine-tuning objective of VSD in Eq. 2 inspires us to investigate whether simply shifting earlier the timestep could fulfill the fine-tuning requirements of VSD without modifying the pre-trained 2D diffusion network parameters. Specifically, we employ the pretrained 2D diffusion model with shifted timestep to approximate the diffusion model of rendered images in Eq. 1 as " + }, + { + "bbox": [ + 130, + 478, + 482, + 562 + ], + "type": "inline_equation", + "content": "\\pmb{\\epsilon}_{\\theta}(\\pmb{x}_t; t, \\pi, y) \\triangleq \\pmb{\\epsilon}_{\\phi}(\\pmb{x}_{t + \\Delta t}; t + \\Delta t, y^{\\pi})" + }, + { + "bbox": [ + 130, + 478, + 482, + 562 + ], + "type": "text", + "content": ", resulting in the following Asynchronous Score Distillation (ASD) objective function:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 577, + 481, + 604 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 577, + 481, + 604 + ], + "spans": [ + { + "bbox": [ + 143, + 577, + 481, + 604 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\theta} \\mathcal {L} _ {\\mathrm {A S D}} (\\theta , y) \\triangleq \\mathbb {E} _ {\\pi , t, \\epsilon} \\left[ \\omega (t) \\left(\\boldsymbol {\\epsilon} _ {\\phi} \\left(\\boldsymbol {x} _ {t}; t, y ^ {\\pi}\\right) - \\boldsymbol {\\epsilon} _ {\\phi} \\left(\\boldsymbol {x} _ {t + \\Delta t}; t + \\Delta t, y ^ {\\pi}\\right)\\right) \\frac {\\partial \\boldsymbol {x}}{\\partial \\theta} \\right]. \\quad (4)", + "image_path": "83ffd7b834e41d61bcba98c13cd6ffc8985dfb909fd263a1d415f13358240f3e.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "We can see that rather than iteratively fine-tuning the diffusion network as in VSD, ASD achieves similar goal by shifting the timestep " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " with an interval " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " in each step, which is much more efficient. One key variable introduced in ASD is the timestep shift " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ", which will be discussed in the next subsection." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "text", + "content": "Z.Ma et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 170, + 116, + 441, + 278 + ], + "blocks": [ + { + "bbox": [ + 170, + 116, + 441, + 278 + ], + "lines": [ + { + "bbox": [ + 170, + 116, + 441, + 278 + ], + "spans": [ + { + "bbox": [ + 170, + 116, + 441, + 278 + ], + "type": "image", + "image_path": "42b5dd41704546434be15a59e10acbfba878ec1590bc5e05f91031f968d6966f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 279, + 482, + 323 + ], + "lines": [ + { + "bbox": [ + 130, + 279, + 482, + 323 + ], + "spans": [ + { + "bbox": [ + 130, + 279, + 482, + 323 + ], + "type": "text", + "content": "Fig. 1: Illustration of the noise prediction error of the pre-trained 2D diffusion model " + }, + { + "bbox": [ + 130, + 279, + 482, + 323 + ], + "type": "inline_equation", + "content": "\\epsilon_{PT}(t)" + }, + { + "bbox": [ + 130, + 279, + 482, + 323 + ], + "type": "text", + "content": " and that of the fine-tuned 2D diffusion model " + }, + { + "bbox": [ + 130, + 279, + 482, + 323 + ], + "type": "inline_equation", + "content": "\\epsilon_{FT}(t)" + }, + { + "bbox": [ + 130, + 279, + 482, + 323 + ], + "type": "text", + "content": ". We can see that the curve of " + }, + { + "bbox": [ + 130, + 279, + 482, + 323 + ], + "type": "inline_equation", + "content": "e_{FT}(t)" + }, + { + "bbox": [ + 130, + 279, + 482, + 323 + ], + "type": "text", + "content": " is positioned under that of " + }, + { + "bbox": [ + 130, + 279, + 482, + 323 + ], + "type": "inline_equation", + "content": "e_{PT}(t)" + }, + { + "bbox": [ + 130, + 279, + 482, + 323 + ], + "type": "text", + "content": ", and we can shift the timestep of " + }, + { + "bbox": [ + 130, + 279, + 482, + 323 + ], + "type": "inline_equation", + "content": "\\epsilon_{PT}(t)" + }, + { + "bbox": [ + 130, + 279, + 482, + 323 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 279, + 482, + 323 + ], + "type": "inline_equation", + "content": "\\epsilon_{PT}(t + \\Delta t)" + }, + { + "bbox": [ + 130, + 279, + 482, + 323 + ], + "type": "text", + "content": " to approximate the noise prediction error of " + }, + { + "bbox": [ + 130, + 279, + 482, + 323 + ], + "type": "inline_equation", + "content": "\\epsilon_{FT}(t)" + }, + { + "bbox": [ + 130, + 279, + 482, + 323 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 331, + 329, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 331, + 329, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 331, + 329, + 342 + ], + "type": "text", + "content": "3.2 The Setting of Timestep Shift " + }, + { + "bbox": [ + 132, + 331, + 329, + 342 + ], + "type": "inline_equation", + "content": "\\Delta t" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 354, + 482, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 426 + ], + "type": "text", + "content": "Before discussing how to set the timestep shift " + }, + { + "bbox": [ + 130, + 354, + 482, + 426 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 130, + 354, + 482, + 426 + ], + "type": "text", + "content": ", let's plot another curve, i.e., the noise prediction error of " + }, + { + "bbox": [ + 130, + 354, + 482, + 426 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\boldsymbol{x}_t; t, \\pi, y)" + }, + { + "bbox": [ + 130, + 354, + 482, + 426 + ], + "type": "text", + "content": " w.r.t. timestep " + }, + { + "bbox": [ + 130, + 354, + 482, + 426 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 354, + 482, + 426 + ], + "type": "text", + "content": ". Actually, in the process of generating " + }, + { + "bbox": [ + 130, + 354, + 482, + 426 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}" + }, + { + "bbox": [ + 130, + 354, + 482, + 426 + ], + "type": "text", + "content": " with VSD, we will have the fine-tuned model " + }, + { + "bbox": [ + 130, + 354, + 482, + 426 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi'}(\\boldsymbol{x}_t; t, \\pi, y)" + }, + { + "bbox": [ + 130, + 354, + 482, + 426 + ], + "type": "text", + "content": " as the by-product, which is used to represent " + }, + { + "bbox": [ + 130, + 354, + 482, + 426 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\boldsymbol{x}_t; t, \\pi, y)" + }, + { + "bbox": [ + 130, + 354, + 482, + 426 + ], + "type": "text", + "content": " in Eq. 1. Therefore, with fixed " + }, + { + "bbox": [ + 130, + 354, + 482, + 426 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}" + }, + { + "bbox": [ + 130, + 354, + 482, + 426 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 354, + 482, + 426 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 354, + 482, + 426 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 354, + 482, + 426 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 354, + 482, + 426 + ], + "type": "text", + "content": ", the noise prediction error of the fine-tuned diffusion model, denoted by " + }, + { + "bbox": [ + 130, + 354, + 482, + 426 + ], + "type": "inline_equation", + "content": "\\epsilon_{FT}(t)" + }, + { + "bbox": [ + 130, + 354, + 482, + 426 + ], + "type": "text", + "content": ", can be calculated as " + }, + { + "bbox": [ + 130, + 354, + 482, + 426 + ], + "type": "inline_equation", + "content": "e_{FT}(t) = \\| \\epsilon_{\\phi'}(t) - \\epsilon \\|_2^2" + }, + { + "bbox": [ + 130, + 354, + 482, + 426 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "content": "The curve of " + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "inline_equation", + "content": "e_{FT}(t)" + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "content": " w.r.t. " + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "content": " (i.e., the yellow curve) is plotted in Fig. 1 by using the same data as in plotting " + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "inline_equation", + "content": "e_{PT}(t)" + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "content": ". We can see that the curve of " + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "inline_equation", + "content": "e_{FT}(t)" + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "content": " is positioned under " + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "inline_equation", + "content": "e_{PT}(t)" + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "content": " because " + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "inline_equation", + "content": "e_{FT}(t)" + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "content": " is obtained by the fine-tuned diffusion model " + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "inline_equation", + "content": "\\epsilon_{FT}" + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "content": ". However, as mentioned in Sec. 2.2, this fine-tuning changes the weights of pre-trained diffusion model and might damage its ability in comprehending text-image pairs. Therefore, we propose to fix the pre-trained model " + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "inline_equation", + "content": "\\epsilon_{PT}(t)" + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "content": " but shift it to " + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "inline_equation", + "content": "\\epsilon_{PT}(t + \\Delta t)" + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "content": " to approximate the desired " + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "inline_equation", + "content": "\\epsilon_{FT}(t)" + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "content": ". Referring to Fig. 1, we could shift " + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "inline_equation", + "content": "\\epsilon_{PT}(t)" + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "content": " to an earlier timestep to achieve this goal. For example, at timestep " + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "inline_equation", + "content": "t_0" + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "content": " and with a time shift " + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "inline_equation", + "content": "\\Delta t_0 > 0" + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "content": ", we can use " + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "inline_equation", + "content": "\\epsilon_{PT}(t_0 + \\Delta t_0)" + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "content": " to approximate the noise prediction error of " + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "inline_equation", + "content": "\\epsilon_{FT}(t_0)" + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": "On the other hand, the magnitude of " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": " will vary with " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": ". Let's come to another timestep " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "t_1" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": " in Fig. 1, where " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "t_1" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": " is earlier than " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "t_0" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": ". Because the decreasing speeds of both " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "e_{PT}" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "e_{FT}" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": " will be reduced with " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": " going to " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{max}}" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": ", the magnitude of " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\Delta t_1" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": " will be increased to approximate " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "e_{FT}(t_1)" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": ". In other words, the magnitude of " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": " should grow when " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": " goes from " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{min}}" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{max}}" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": ". We heuristically set this relationship as " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\Delta t = \\eta (t - T_{\\mathrm{min}})" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\eta \\in [0,1]" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": " is a hyper-parameter that controls the length of shift range. Finally, it should be pointed out that the curves in Fig. 1 will vary a little for different training iterations, rendered images " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": " and text prompts " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": ". Therefore, " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": " should fall into some range " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "S(t)" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": ". In practice, we set " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\Delta t \\sim S(t) = \\mathcal{U}[0,\\eta (t - T_{\\mathrm{min}})]" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": ", which follows a uniform distribution within 0" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "type": "text", + "content": "ScaleDreamer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 153, + 117, + 461, + 225 + ], + "blocks": [ + { + "bbox": [ + 153, + 117, + 461, + 225 + ], + "lines": [ + { + "bbox": [ + 153, + 117, + 461, + 225 + ], + "spans": [ + { + "bbox": [ + 153, + 117, + 461, + 225 + ], + "type": "image", + "image_path": "a632f5b8e981ceef862e3fd7e172e560f6a486258493c5184b3f536ace0e9f65.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 228, + 480, + 251 + ], + "lines": [ + { + "bbox": [ + 131, + 228, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 131, + 228, + 480, + 251 + ], + "type": "text", + "content": "Fig. 2: Left and middle: 2D toy examples by SDS [40], CSD [72], VSD [59] and our proposed ASD. Right: Gradient norms generated by different methods." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 254, + 370, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 254, + 370, + 266 + ], + "spans": [ + { + "bbox": [ + 132, + 254, + 370, + 266 + ], + "type": "text", + "content": "Algorithm 1 Asynchronous Score Distillation (ASD)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 268, + 475, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 268, + 475, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 268, + 475, + 278 + ], + "type": "text", + "content": "Input: 3D representation " + }, + { + "bbox": [ + 132, + 268, + 475, + 278 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 132, + 268, + 475, + 278 + ], + "type": "text", + "content": "; Text prompt " + }, + { + "bbox": [ + 132, + 268, + 475, + 278 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 132, + 268, + 475, + 278 + ], + "type": "text", + "content": "; Hyperparameter " + }, + { + "bbox": [ + 132, + 268, + 475, + 278 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 132, + 268, + 475, + 278 + ], + "type": "text", + "content": "; 2D diffusion prior " + }, + { + "bbox": [ + 132, + 268, + 475, + 278 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}" + } + ] + } + ], + "index": 5 + }, + { + "type": "code", + "bbox": [ + 124, + 280, + 422, + 369 + ], + "blocks": [ + { + "bbox": [ + 124, + 280, + 422, + 369 + ], + "lines": [ + { + "bbox": [ + 124, + 280, + 422, + 369 + ], + "spans": [ + { + "bbox": [ + 124, + 280, + 422, + 369 + ], + "type": "text", + "content": "1 while not converged do \n2 Sample a camera pose " + }, + { + "bbox": [ + 124, + 280, + 422, + 369 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 124, + 280, + 422, + 369 + ], + "type": "text", + "content": " \n3 Render an image " + }, + { + "bbox": [ + 124, + 280, + 422, + 369 + ], + "type": "inline_equation", + "content": "\\pmb{x} = g(\\theta, \\pi)" + }, + { + "bbox": [ + 124, + 280, + 422, + 369 + ], + "type": "text", + "content": " \n4 Sample a timestep " + }, + { + "bbox": [ + 124, + 280, + 422, + 369 + ], + "type": "inline_equation", + "content": "t \\sim \\mathcal{U}[T_{\\min}, T_{\\max}]" + }, + { + "bbox": [ + 124, + 280, + 422, + 369 + ], + "type": "text", + "content": ", Gaussian noise " + }, + { + "bbox": [ + 124, + 280, + 422, + 369 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0, I)" + }, + { + "bbox": [ + 124, + 280, + 422, + 369 + ], + "type": "text", + "content": " \n5 Sample a timestep shift " + }, + { + "bbox": [ + 124, + 280, + 422, + 369 + ], + "type": "inline_equation", + "content": "\\Delta t \\sim S(t) = \\mathcal{U}[0, \\eta(t - T_{\\min})]" + }, + { + "bbox": [ + 124, + 280, + 422, + 369 + ], + "type": "text", + "content": " \n6 " + }, + { + "bbox": [ + 124, + 280, + 422, + 369 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t \\gets \\alpha_t \\pmb{x} + \\sigma_t \\pmb{\\epsilon}" + }, + { + "bbox": [ + 124, + 280, + 422, + 369 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 124, + 280, + 422, + 369 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{t + \\Delta t} \\gets \\alpha_{t + \\Delta t} \\pmb{x} + \\sigma_{t + \\Delta t} \\pmb{\\epsilon}" + }, + { + "bbox": [ + 124, + 280, + 422, + 369 + ], + "type": "text", + "content": " \n7 Update " + }, + { + "bbox": [ + 124, + 280, + 422, + 369 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 124, + 280, + 422, + 369 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 124, + 280, + 422, + 369 + ], + "type": "inline_equation", + "content": "\\Delta \\theta \\gets \\omega(t) (\\pmb{\\epsilon}_{\\phi}(\\pmb{x}_t; t, y^{\\pi}) - \\pmb{\\epsilon}_{\\phi}(\\pmb{x}_{t + \\Delta t}; t + \\Delta t, y^{\\pi})) \\frac{\\partial \\pmb{x}}{\\partial \\theta}" + }, + { + "bbox": [ + 124, + 280, + 422, + 369 + ], + "type": "text", + "content": " \n8 end" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "code_body" + } + ], + "index": 6, + "sub_type": "code", + "guess_lang": "latex" + }, + { + "bbox": [ + 130, + 398, + 480, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 398, + 480, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 398, + 480, + 422 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 130, + 398, + 480, + 422 + ], + "type": "inline_equation", + "content": "\\eta(t - T_{\\min})" + }, + { + "bbox": [ + 130, + 398, + 480, + 422 + ], + "type": "text", + "content": ". The pseudo-code of ASD is summarized in Alg. 1, which can be applied to both prompt-specific and prompt-amortized text-to-3D tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 424, + 481, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 424, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 424, + 481, + 568 + ], + "type": "text", + "content": "2D toy experiments. To verify the proposed timestep shift strategy, we follow the paradigm in [59] to test SDS, CSD, VSD and our ASD on 2D toy examples. The left column of Fig. 2 shows the results of SDS, CSD, VSD, and the middle column shows the results of ASD with different sampling strategies of " + }, + { + "bbox": [ + 130, + 424, + 481, + 568 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 130, + 424, + 481, + 568 + ], + "type": "text", + "content": ". One can see that the proposed sampling strategy " + }, + { + "bbox": [ + 130, + 424, + 481, + 568 + ], + "type": "inline_equation", + "content": "\\Delta t \\sim S(t) = \\mathcal{U}[0, \\eta(t - T_{\\min})]" + }, + { + "bbox": [ + 130, + 424, + 481, + 568 + ], + "type": "text", + "content": " yields similar results to VSD [59]. Besides, we show the gradient norm produced by these score distillation methods in the right column of Fig. 2. One can see that the range of gradient norm produced by ASD is similar to that of VSD. However, the gradient norm of SDS is more than 10 times larger than ASD and VSD because it needs to set CFG=100 for convergence [40, 59, 72]. Such a large gradient may result in training instability. We append more 2D results in the supplementary material to further validate our proposed sampling strategy." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "Text-to-3D Synthesis with ASD. As a score distillation method, ASD is open to the selection of 3D generator architectures [4,17,22,33,39]. The general pipeline of ASD for text-to-3D synthesis is shown in Fig. 3. It takes a rendered image as input and diffuses it in two timesteps " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "t + \\Delta t" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": ". The noise prediction difference is used as the gradient to optimize the 3D representation of generator. In this work, in addition to prompt-specific generation, as done in most existing score distillation works [16, 28, 40, 59, 64], we focus more on prompt-amortized text-to-3D and conduct thorough experiments to evaluate the effectiveness of" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "Z.Ma et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 150, + 114, + 462, + 213 + ], + "blocks": [ + { + "bbox": [ + 150, + 114, + 462, + 213 + ], + "lines": [ + { + "bbox": [ + 150, + 114, + 462, + 213 + ], + "spans": [ + { + "bbox": [ + 150, + 114, + 462, + 213 + ], + "type": "image", + "image_path": "38adc03247ab655fdd85b769d22913e0486abe4f7fc76d04425fc119318d0119.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 215, + 480, + 270 + ], + "lines": [ + { + "bbox": [ + 130, + 215, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 215, + 480, + 270 + ], + "type": "text", + "content": "Fig. 3: Overview of Asynchronous Score Distillation (ASD). As illustrated in the left sub-figure, ASD can be employed for prompt-specific generation by optimizing 3D representations for each prompt, as well as for prompt-amortized generation by training a text-to-3D generator. The right sub-figure depicts how ASD uses the difference in noise predictions at asynchronous timesteps to update the 3D network parameters." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 274, + 482, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 274, + 482, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 482, + 418 + ], + "type": "text", + "content": "ASD with representative 3D generator architectures, including Hyper-iNGP and 3DConv-net, by using Stable Diffusion as the 2D diffusion model. HyperiNGP is adopted by ATT3D [33], which integrates a prompt-agnostic hash-grid spatial encoding [39] with prompt-conditioned decoding layers to output color and density. 3DConv-net [4] is a 3D generator that maps the provided condition to voxel using 3D convolution. They are chosen in our experiments because they represent two groups of 3D generators that inject the condition through implicit mapping function [3,21] or explicit space representation [7,17,46,49]. The results are shown in Sec. 4. We also conduct experiments to demonstrate that ASD can work with other 2D diffusion models such as MVDream [48], and can be used to train other 3D generator architectures such as Triplane-Transformer [17]. More details of these experiments can be found in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 434, + 230, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 434, + 230, + 448 + ], + "spans": [ + { + "bbox": [ + 132, + 434, + 230, + 448 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 456, + 272, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 456, + 272, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 456, + 272, + 468 + ], + "type": "text", + "content": "4.1 Experimental Settings" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 474, + 480, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 480, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 480, + 545 + ], + "type": "text", + "content": "Comparison Methods. We compare ASD with state-of-the-art score distillation methods, including SDS [40], CSD [72] and VSD [59]. We adhere to their official codes for training prompt-amortized text-to-3D networks. For example, the CFG [16] values for SDS, CSD and VSD are configured to 100, 1, and 7.5, respectively. In addition, we compare with existing prompt-amortized method ATT3D [33] (whose code is not released yet) by replicating its reported results." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 481, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 481, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 481, + 640 + ], + "type": "text", + "content": "Implementation Details. We employ VolSDF [69] to render images from the 3D generators. For Stable Diffusion, we employ SD-v2.1-base [1] for all score distillation methods for fair comparison. As configured in VSD [59], we set the CFG value as 7.5 for the pre-trained diffusion model in ASD, and 1 for the diffusion model of rendered images. The resolution of rendered images by HyperiNGP is set to " + }, + { + "bbox": [ + 130, + 545, + 481, + 640 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 130, + 545, + 481, + 640 + ], + "type": "text", + "content": ", while that of 3DConv-net and Triplane-Transformer is set to " + }, + { + "bbox": [ + 130, + 545, + 481, + 640 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 130, + 545, + 481, + 640 + ], + "type": "text", + "content": " for GPU memory considerations. Other details are in the supplementary material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": "Prompt Corpus. To thoroughly evaluate the capability of ASD in prompt-amortized text-to-3D synthesis, we employ multiple datasets encompassing a" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "type": "text", + "content": "ScaleDreamer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": "range of text prompt quantities. MG15 includes 15 prompts from Magic3D [29]; DF415 comprises 415 prompts from DreamFusion [40]; and AT2520 contains 2520 compositional prompts of animals from ATT3D [33]. DL17k contains 17k compositional prompts of human with daily activities, proposed by [25]. While AT2520 and DL17k provide a larger number of prompts than DF415, the prompt diversity of them is relatively low due to the predefined templates." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 482, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 482, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 482, + 295 + ], + "type": "text", + "content": "To test ASD's performance with an even larger scale of prompts, we introduce a novel prompt corpus named CP100k. This corpus consists of 100,000 text prompts filtered from the image descriptions collected by Cap3D [34], which was developed to test text-to-image model performance. To the best of our knowledge, it is the first time to evaluate score distillation methods on such a scale of text prompts. Meanwhile, it should be clarified that this work is focused on examining the score distillation performance rather than prompt generalization, so the test prompts share the same distribution as training prompts. More details of the prompt corpus are in supplementary material." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 296, + 482, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 296, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 482, + 380 + ], + "type": "text", + "content": "Evaluation Metrics. We render 120 surrounding view images as the 3D synthesis result from each prompt. Similar to previous text-to-3D works [25,33, 33,40], we compute the CLIP recall, i.e., the classification accuracy by applying CLIP model to the rendered images to predict the correct text prompt, as one performance metric, denoted by \"R@1\". Additionally, we calculate the CLIP text-image similarity between generated images and input prompts as another metric [54,60], denoted by \"Sim\"." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 396, + 256, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 396, + 256, + 407 + ], + "spans": [ + { + "bbox": [ + 131, + 396, + 256, + 407 + ], + "type": "text", + "content": "4.2 Evaluation Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 413, + 482, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 482, + 557 + ], + "type": "text", + "content": "Results with iNGP/Hyper-iNGP as 3D Representation. The iNGP [39] architecture is designed for prompt-specific text-to-3D generation. Hyper-iNGP has the same spatial encoding as iNGP except that the weights of the decoding layer depend on the text prompt. To eliminate the effect caused by architecture difference as much as possible, we adopt iNGP for prompt-specific text-to-3D tasks, and Hyper-iNGP for prompt-amortized tasks. Our experiments are carried out on the MG15 dataset. For prompt-specific tasks, we optimize an individual iNGP [39] for each MG15 prompt; while for the prompt-amortized tasks, we train a single Hyper-iNGP [33] across all MG15 prompts. We also compare our results with ATT3D [33], which is among the first to apply Hyper-iNGP to prompt-amortized text-to-3D tasks. ATT3D employs SDS for training and uses soft-shading [40] (denoted as * in Tab. 2) for rendering." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": "The qualitative and quantitative results are shown in Fig. 4 and Tab. 2, respectively. We can see that the existing methods suffer from performance decrease when transiting from prompt-specific to prompt-amortized tasks, as evidenced by the decreased CLIP similarity and recall in Tab. 2. It is worth mentioning that training Hyper-net with SDS requires turning on the spectral normalization [38] in the linear layers, otherwise the training will fail due to numerical instability. This observation is consistent with what reported in ATT3D [33]. This is because SDS suffers from large gradient norm (please also refer to Fig. 2 and the discussions therein), which makes Hyper-iNGP hard to" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "text", + "content": "Z.Ma et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 152, + 115, + 217, + 204 + ], + "blocks": [ + { + "bbox": [ + 152, + 115, + 217, + 204 + ], + "lines": [ + { + "bbox": [ + 152, + 115, + 217, + 204 + ], + "spans": [ + { + "bbox": [ + 152, + 115, + 217, + 204 + ], + "type": "image", + "image_path": "408addbe1a5f9af089cabaa05c98be98d1cddd5e3eeb7aa992503005b8fc424a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 214, + 480, + 248 + ], + "lines": [ + { + "bbox": [ + 131, + 214, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 131, + 214, + 480, + 248 + ], + "type": "text", + "content": "Fig. 4: Qualitative comparison on prompt-specific (with iNGP as the 3D representation) and prompt-amortized (with Hyper-iNGP as the 3D generator) text-to-3D results by SDS [40], CSD [72], VSD [59], ATT3D [33] and our ASD methods." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 220, + 115, + 280, + 204 + ], + "blocks": [ + { + "bbox": [ + 220, + 115, + 280, + 204 + ], + "lines": [ + { + "bbox": [ + 220, + 115, + 280, + 204 + ], + "spans": [ + { + "bbox": [ + 220, + 115, + 280, + 204 + ], + "type": "image", + "image_path": "70f23e0a08667dcc4ee0b16f3e88208beb54d087ffb6d5993bfb832b8cd2dadb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 256, + 205, + 370, + 212 + ], + "lines": [ + { + "bbox": [ + 256, + 205, + 370, + 212 + ], + "spans": [ + { + "bbox": [ + 256, + 205, + 370, + 212 + ], + "type": "text", + "content": "\"A DSLR photo of a peacok on a surfboard\" in MG15" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 282, + 116, + 342, + 202 + ], + "blocks": [ + { + "bbox": [ + 282, + 116, + 342, + 202 + ], + "lines": [ + { + "bbox": [ + 282, + 116, + 342, + 202 + ], + "spans": [ + { + "bbox": [ + 282, + 116, + 342, + 202 + ], + "type": "image", + "image_path": "5ff548565ca51d53e4528d5ac75a260fc105de4301fd93dce9de5fdba658bf00.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 343, + 116, + 384, + 201 + ], + "blocks": [ + { + "bbox": [ + 343, + 116, + 384, + 201 + ], + "lines": [ + { + "bbox": [ + 343, + 116, + 384, + 201 + ], + "spans": [ + { + "bbox": [ + 343, + 116, + 384, + 201 + ], + "type": "image", + "image_path": "f1c78c07b9f778bee145766d543dc69bdf4e1170162c7308440824061e2e9a28.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 385, + 116, + 464, + 201 + ], + "blocks": [ + { + "bbox": [ + 385, + 116, + 464, + 201 + ], + "lines": [ + { + "bbox": [ + 385, + 116, + 464, + 201 + ], + "spans": [ + { + "bbox": [ + 385, + 116, + 464, + 201 + ], + "type": "image", + "image_path": "89da58504250c50a28819e51424c8b9344b20cf333d7a496a2765dcaa8610e40.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 132, + 253, + 482, + 337 + ], + "blocks": [ + { + "bbox": [ + 132, + 253, + 482, + 337 + ], + "lines": [ + { + "bbox": [ + 132, + 253, + 482, + 337 + ], + "spans": [ + { + "bbox": [ + 132, + 253, + 482, + 337 + ], + "type": "table", + "html": "
ReferenceMethodSim ↑R@1 ↑MethodSim ↑R@1 ↑
ATT3D [33]---Hyper-iNGP* + SDS0.1950.468
DreamFusion [40]iNGP + SDS0.2881.000Hyper-iNGP + SDS0.2570.918
Classifier [16]iNGP + CSD0.2800.936Hyper-iNGP + CSD0.2640.972
ProlificDreamer [59]iNGP + VSD0.2760.932Hyper-iNGP + VSD0.2590.987
OursiNGP + ASD0.2891.000Hyper-iNGP + ASD0.2841.000
", + "image_path": "e87999bd354efb422286b2753fe5b4c512770b1ed68a15e780a78512276a1f19.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 337, + 479, + 370 + ], + "lines": [ + { + "bbox": [ + 131, + 337, + 479, + 370 + ], + "spans": [ + { + "bbox": [ + 131, + 337, + 479, + 370 + ], + "type": "text", + "content": "Table 2: Quantitative comparison on prompt-specific (with iNGP as the 3D representation) and prompt-amortized (with Hyper-iNGP as the 3D generator) text-to-3D results by SDS [40], CSD [72], VSD [59], ATT3D [33] and our ASD methods." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 386, + 480, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 386, + 480, + 481 + ], + "spans": [ + { + "bbox": [ + 130, + 386, + 480, + 481 + ], + "type": "text", + "content": "converge. As can be seen in Fig. 4, ATT3D results in wrong geometry by using soft shading and SDS for training. For CSD, we see that it fails to optimize the full geometry, as shown by the shrunk peacock in both prompt-amortized and prompt-amortized results. For VSD, it tends to generate content drifts [48], resulting in repetitive patterns and abnormal geometry. It may fail to generate reasonable contents in both prompt-specific and prompt-amortized tasks. In contrast, our proposed ASD works very stable across the two tasks, yielding not only outstanding quantitative scores but also high quality 3D contents." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": "Results with 3DConv-net as 3D Generator. The issues of existing score distillation methods either persist or become more pronounced when replacing Hyper-iNGP to 3DConv-net as the 3D generator. We find that training SDS with 3DConv-net always fails within several thousand iterations, even using spectral or other normalization techniques. This issue stems from that deeper network is more sensitive to large gradients [13] caused by SDS. Therefore, we only compare the results of other methods in Fig. 5. We see that CSD outputs acceptable results on AT2520, but its results on DF415, which has more varied prompts, are consistently smaller than anticipated. Such a phenomenon has been observed when Hyper-iNGP is used as the generator, which underlines CSD's inability to reliably guide the 3D generator to produce geometries aligned with the text prompts. As for VSD, it leads to rather abnormal results, failing to match the text prompts. This can be attributed to its fine-tuning of the pre-trained 2D diffusion model, which severely compromises VSD's text-image comprehending ability. In comparison, our proposed ASD, with 3DConv-net as the generator," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "type": "text", + "content": "ScaleDreamer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 152, + 115, + 238, + 166 + ], + "blocks": [ + { + "bbox": [ + 152, + 115, + 238, + 166 + ], + "lines": [ + { + "bbox": [ + 152, + 115, + 238, + 166 + ], + "spans": [ + { + "bbox": [ + 152, + 115, + 238, + 166 + ], + "type": "image", + "image_path": "b3a37a77b484107a67ee819383b2523eaae0bb583d0f64c89985bbddfb578f32.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 242, + 115, + 332, + 166 + ], + "blocks": [ + { + "bbox": [ + 242, + 115, + 332, + 166 + ], + "lines": [ + { + "bbox": [ + 242, + 115, + 332, + 166 + ], + "spans": [ + { + "bbox": [ + 242, + 115, + 332, + 166 + ], + "type": "image", + "image_path": "3d868fa32efcdbdd740f76d8d754a5ac827cea9849b470efc4fe52b044cb5b4e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 335, + 124, + 463, + 167 + ], + "blocks": [ + { + "bbox": [ + 343, + 116, + 455, + 123 + ], + "lines": [ + { + "bbox": [ + 343, + 116, + 455, + 123 + ], + "spans": [ + { + "bbox": [ + 343, + 116, + 455, + 123 + ], + "type": "text", + "content": "Asynchronous Score Distillation (ASD, ours)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 335, + 124, + 463, + 167 + ], + "lines": [ + { + "bbox": [ + 335, + 124, + 463, + 167 + ], + "spans": [ + { + "bbox": [ + 335, + 124, + 463, + 167 + ], + "type": "image", + "image_path": "24446fcc06bb1364faf18446ffdc62d4e0454581cde6a5660f0e42c79fdb78be.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 228, + 167, + 398, + 173 + ], + "lines": [ + { + "bbox": [ + 228, + 167, + 398, + 173 + ], + "spans": [ + { + "bbox": [ + 228, + 167, + 398, + 173 + ], + "type": "text", + "content": "\"A squirrel holding a book wearing a sweater wearing a tophat\" in AT2520" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 152, + 174, + 194, + 215 + ], + "blocks": [ + { + "bbox": [ + 152, + 174, + 194, + 215 + ], + "lines": [ + { + "bbox": [ + 152, + 174, + 194, + 215 + ], + "spans": [ + { + "bbox": [ + 152, + 174, + 194, + 215 + ], + "type": "image", + "image_path": "a5ca991def4b83545b4ae7a81043c95b557b01ff73bedf54aaef828fca7134b8.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 225, + 480, + 258 + ], + "lines": [ + { + "bbox": [ + 131, + 225, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 131, + 225, + 480, + 258 + ], + "type": "text", + "content": "Fig. 5: Qualitative comparison among CSD [72], VSD [59] and our ASD (with 3DConv-net as generator) on AT2520 and DF415 corpuses. SDS is not compared because it encounters numerical instability in this experiment." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 195, + 174, + 237, + 215 + ], + "blocks": [ + { + "bbox": [ + 195, + 174, + 237, + 215 + ], + "lines": [ + { + "bbox": [ + 195, + 174, + 237, + 215 + ], + "spans": [ + { + "bbox": [ + 195, + 174, + 237, + 215 + ], + "type": "image", + "image_path": "44b3d42e96d1a3292a1a7e001c390c0e1b5848872b77cf845b9224fd7a1bbd04.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 236, + 216, + 380, + 224 + ], + "lines": [ + { + "bbox": [ + 236, + 216, + 380, + 224 + ], + "spans": [ + { + "bbox": [ + 236, + 216, + 380, + 224 + ], + "type": "text", + "content": "\"A DSLR photo of a cocker spaniel wearing a crown\" in DF415" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 238, + 174, + 287, + 215 + ], + "blocks": [ + { + "bbox": [ + 238, + 174, + 287, + 215 + ], + "lines": [ + { + "bbox": [ + 238, + 174, + 287, + 215 + ], + "spans": [ + { + "bbox": [ + 238, + 174, + 287, + 215 + ], + "type": "image", + "image_path": "450fab972bbfc2c45476959b453c13a79c0fb0f539d28be8f77000d69e45cfde.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 288, + 174, + 331, + 215 + ], + "blocks": [ + { + "bbox": [ + 288, + 174, + 331, + 215 + ], + "lines": [ + { + "bbox": [ + 288, + 174, + 331, + 215 + ], + "spans": [ + { + "bbox": [ + 288, + 174, + 331, + 215 + ], + "type": "image", + "image_path": "e3394f68d5b3630cbecd4f07c52674033542965ca25c2e3262b6d00195c8aeea.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 335, + 174, + 377, + 215 + ], + "blocks": [ + { + "bbox": [ + 335, + 174, + 377, + 215 + ], + "lines": [ + { + "bbox": [ + 335, + 174, + 377, + 215 + ], + "spans": [ + { + "bbox": [ + 335, + 174, + 377, + 215 + ], + "type": "image", + "image_path": "164be7bcf11c40c71aacf25f2af2062f50ba2407999f6fea028fd83591c7d520.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 378, + 174, + 420, + 215 + ], + "blocks": [ + { + "bbox": [ + 378, + 174, + 420, + 215 + ], + "lines": [ + { + "bbox": [ + 378, + 174, + 420, + 215 + ], + "spans": [ + { + "bbox": [ + 378, + 174, + 420, + 215 + ], + "type": "image", + "image_path": "749cd727c6f06661375d504d3d893cba7c7d988f5d7e072f021e72f46949fc65.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 422, + 174, + 463, + 215 + ], + "blocks": [ + { + "bbox": [ + 422, + 174, + 463, + 215 + ], + "lines": [ + { + "bbox": [ + 422, + 174, + 463, + 215 + ], + "spans": [ + { + "bbox": [ + 422, + 174, + 463, + 215 + ], + "type": "image", + "image_path": "e662d35d7a692f9fb4d5eaa18a7187cf08b18d429301d6c631c0b6a5a94696be.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 169, + 258, + 446, + 340 + ], + "blocks": [ + { + "bbox": [ + 169, + 258, + 446, + 340 + ], + "lines": [ + { + "bbox": [ + 169, + 258, + 446, + 340 + ], + "spans": [ + { + "bbox": [ + 169, + 258, + 446, + 340 + ], + "type": "table", + "html": "
MethodDF415AT2520CP100k
Sim ↑R@1 ↑Sim ↑R@1 ↑Sim ↑R@1 ↑
SDS××××××
CSD0.1760.0620.2790.0370.1950.108
VSD0.1580.0020.1150.0010.1030.000
ASD (ours)0.2370.2760.2850.0580.1990.117
", + "image_path": "beb5b096a6e65ba95f1aab24a7e7130555a6478251c19eec1de6b91d90c9eb70.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 131, + 341, + 480, + 363 + ], + "lines": [ + { + "bbox": [ + 131, + 341, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 131, + 341, + 480, + 363 + ], + "type": "text", + "content": "Table 3: Quantitative comparison on prompt-amortized text-to-3D with 3DConv-net as generator. Symbol " + }, + { + "bbox": [ + 131, + 341, + 480, + 363 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 131, + 341, + 480, + 363 + ], + "type": "text", + "content": " denotes that the training fails due to numerical instability." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 368, + 480, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 368, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 130, + 368, + 480, + 391 + ], + "type": "text", + "content": "yields improved outcomes, as evidenced by the visual results in Fig. 5 and the enhanced metric scores in Tab. 3." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 392, + 481, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 481, + 548 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 481, + 548 + ], + "type": "text", + "content": "Scalability. In this section, we evaluate the scalability of competing methods by using as many as 100k prompts in the CP100k dataset with 3DConv-net as the generator. The results are shown in Fig. 6 and Tab. 3. Due to the issue of numerical instability, SDS is not involved in this experiment. We can see that the outcomes of CSD are significantly diminished with uniformly small-sized shapes across all prompts. There is also a lack of variety since most outputs exhibit similar patterns. The results of VSD are also degenerated, displaying almost identical and anomalous outcomes for the text prompts. This resembles the phenomenon of mode collapse often encountered in bi-level optimization [55], which also highlights the importance of fixing the 2D diffusion model when training with such a large number of text prompts. In comparison, ASD is able to produce much higher quality outcomes across the text prompts, showcasing its capability in large-scale training with numerous text prompts as inputs." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 548, + 481, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 548, + 481, + 608 + ], + "spans": [ + { + "bbox": [ + 130, + 548, + 481, + 608 + ], + "type": "text", + "content": "In addition to the above comparisons, we also evaluate the effectiveness of ASD when using other 2D diffusion models. We employ MVDream [48] to illustrate ASD's generality to diffusion prior models. Besides, we compare against data-driven methods to prove that ASD is advantageous in tackling diverse input of text prompts. Please see supplementary material for detail." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 131, + 623, + 238, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 623, + 238, + 635 + ], + "spans": [ + { + "bbox": [ + 131, + 623, + 238, + 635 + ], + "type": "text", + "content": "4.3 Ablation Study" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 130, + 641, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 480, + 666 + ], + "type": "text", + "content": "In this section, we perform ablation studies to evaluate the settings of timestep shift " + }, + { + "bbox": [ + 130, + 641, + 480, + 666 + ], + "type": "inline_equation", + "content": "\\Delta t\\sim S(t) = \\mathcal{U}[0,\\eta (t - T_{\\mathrm{min}})]" + }, + { + "bbox": [ + 130, + 641, + 480, + 666 + ], + "type": "text", + "content": " from several aspects. The qualitative and" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "Z.Ma et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 114, + 249, + 208 + ], + "blocks": [ + { + "bbox": [ + 149, + 114, + 249, + 208 + ], + "lines": [ + { + "bbox": [ + 149, + 114, + 249, + 208 + ], + "spans": [ + { + "bbox": [ + 149, + 114, + 249, + 208 + ], + "type": "image", + "image_path": "fa6b09ba5048a47e5b6389bbe3d4e5b0b391bc2dfaffa711d11225f95d38282b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 256, + 115, + 356, + 207 + ], + "blocks": [ + { + "bbox": [ + 256, + 115, + 356, + 207 + ], + "lines": [ + { + "bbox": [ + 256, + 115, + 356, + 207 + ], + "spans": [ + { + "bbox": [ + 256, + 115, + 356, + 207 + ], + "type": "image", + "image_path": "d76ea56d08e5a8f55c72f62961598316a507b2285fadf685735f1b0893149965.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 359, + 115, + 460, + 208 + ], + "blocks": [ + { + "bbox": [ + 359, + 115, + 460, + 208 + ], + "lines": [ + { + "bbox": [ + 359, + 115, + 460, + 208 + ], + "spans": [ + { + "bbox": [ + 359, + 115, + 460, + 208 + ], + "type": "image", + "image_path": "c12bcf73513fd51a5dac999544ebd93ab04829e977d8f310604e724038b9dfc1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 150, + 215, + 249, + 255 + ], + "blocks": [ + { + "bbox": [ + 150, + 215, + 249, + 255 + ], + "lines": [ + { + "bbox": [ + 150, + 215, + 249, + 255 + ], + "spans": [ + { + "bbox": [ + 150, + 215, + 249, + 255 + ], + "type": "image", + "image_path": "0192162f68d2fc7bc887c2950ede62e400914c1b423010cba396e5a226cc99ba.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 136, + 263, + 476, + 275 + ], + "lines": [ + { + "bbox": [ + 136, + 263, + 476, + 275 + ], + "spans": [ + { + "bbox": [ + 136, + 263, + 476, + 275 + ], + "type": "text", + "content": "Fig. 6: The scalability comparison with CSD [72] and VSD [59] on CP100k corpus." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 256, + 209, + 355, + 255 + ], + "blocks": [ + { + "bbox": [ + 256, + 209, + 355, + 255 + ], + "lines": [ + { + "bbox": [ + 256, + 209, + 355, + 255 + ], + "spans": [ + { + "bbox": [ + 256, + 209, + 355, + 255 + ], + "type": "image", + "image_path": "a78e82c0fd9d95b6c49ff83b72dc71e4cd27947f1461ab84884f1f753fa1b43e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 264, + 255, + 343, + 262 + ], + "lines": [ + { + "bbox": [ + 264, + 255, + 343, + 262 + ], + "spans": [ + { + "bbox": [ + 264, + 255, + 343, + 262 + ], + "type": "text", + "content": "\"Small wooden brown bear figurine\"" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 359, + 215, + 459, + 255 + ], + "blocks": [ + { + "bbox": [ + 359, + 215, + 459, + 255 + ], + "lines": [ + { + "bbox": [ + 359, + 215, + 459, + 255 + ], + "spans": [ + { + "bbox": [ + 359, + 215, + 459, + 255 + ], + "type": "image", + "image_path": "8da0bfae92e89486c3c7b9911273aa674f195a7ba6ad3986bfb4f3f8d63fe956.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 282, + 480, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 282, + 480, + 317 + ], + "spans": [ + { + "bbox": [ + 130, + 282, + 480, + 317 + ], + "type": "text", + "content": "quantitative results are shown in Fig. 7 and Tab. 4, respectively. We provide extensive ablation studies in the supplementary material to show that the proposed strategy is also effective when using MVDream as the diffusion model." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 319, + 480, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 480, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 480, + 437 + ], + "type": "text", + "content": "Importance of Timestep Shift. We use " + }, + { + "bbox": [ + 130, + 319, + 480, + 437 + ], + "type": "inline_equation", + "content": "\\eta = 0" + }, + { + "bbox": [ + 130, + 319, + 480, + 437 + ], + "type": "text", + "content": " (i.e., no timestep shift) as a baseline to evaluate the necessity of introducing timestep shift " + }, + { + "bbox": [ + 130, + 319, + 480, + 437 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 130, + 319, + 480, + 437 + ], + "type": "text", + "content": ". From Fig. 7 and Tab. 4, we see that while it can generate plausible results, the model is prone to generating shapes that do not make sense, such as the so-called Janus problem [2]. Examples include a frog with an extra eye, robot face with block-like features, and a peacock with tails at both the front and back. This is because the non-shifted diffusion model will align more with the 2D image distribution, tending to generate redundant contents and unreasonable geometry along the training. By introducing a timestep shift, our proposed ASD demonstrates advantages in achieving more coherent and visually pleasing results." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 438, + 480, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 480, + 533 + ], + "type": "text", + "content": "Range of Timestep Shift. By setting " + }, + { + "bbox": [ + 130, + 438, + 480, + 533 + ], + "type": "inline_equation", + "content": "\\eta = 0.2" + }, + { + "bbox": [ + 130, + 438, + 480, + 533 + ], + "type": "text", + "content": ", we allow " + }, + { + "bbox": [ + 130, + 438, + 480, + 533 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 130, + 438, + 480, + 533 + ], + "type": "text", + "content": " to be sampled from a large range. However, this might not be a good choice. In the extreme case, for any timestep " + }, + { + "bbox": [ + 130, + 438, + 480, + 533 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 438, + 480, + 533 + ], + "type": "text", + "content": " we can set a large interval " + }, + { + "bbox": [ + 130, + 438, + 480, + 533 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 130, + 438, + 480, + 533 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 130, + 438, + 480, + 533 + ], + "type": "inline_equation", + "content": "t + \\Delta t = T_{\\max}" + }, + { + "bbox": [ + 130, + 438, + 480, + 533 + ], + "type": "text", + "content": ", then the noise prediction becomes " + }, + { + "bbox": [ + 130, + 438, + 480, + 533 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}(\\boldsymbol{x}_t; t, y^\\pi) \\approx \\epsilon" + }, + { + "bbox": [ + 130, + 438, + 480, + 533 + ], + "type": "text", + "content": ", so that ASD is degraded to SDS, which cannot perform well under CFG=7.5 [40]. In practice, we find a larger " + }, + { + "bbox": [ + 130, + 438, + 480, + 533 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 130, + 438, + 480, + 533 + ], + "type": "text", + "content": " tends to result 3D contents with larger size and rounded shapes, e.g., the peacock with closer views, or the frog with larger size, as shown in Fig. 7. Therefore, we set " + }, + { + "bbox": [ + 130, + 438, + 480, + 533 + ], + "type": "inline_equation", + "content": "\\eta = 0.1" + }, + { + "bbox": [ + 130, + 438, + 480, + 533 + ], + "type": "text", + "content": " in all our experiments." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 535, + 480, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 535, + 480, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 535, + 480, + 594 + ], + "type": "text", + "content": "Deterministic or Random Shift. If we set " + }, + { + "bbox": [ + 130, + 535, + 480, + 594 + ], + "type": "inline_equation", + "content": "\\Delta t = \\eta (t - T_{\\mathrm{min}})" + }, + { + "bbox": [ + 130, + 535, + 480, + 594 + ], + "type": "text", + "content": ", it assumes that the diffusion model of rendered images can be approximated by the pretrained one with a fixed and deterministic timestep shift. As shown in Fig. 7 and Tab. 4, it reduces the chance to generate correct geometry and colors. Randomly sampling " + }, + { + "bbox": [ + 130, + 535, + 480, + 594 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 130, + 535, + 480, + 594 + ], + "type": "text", + "content": " in a range is more effective, which is adopted in our method." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 614, + 317, + 627 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 614, + 317, + 627 + ], + "spans": [ + { + "bbox": [ + 132, + 614, + 317, + 627 + ], + "type": "text", + "content": "5 Conclusion and Limitations" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "content": "In this paper, we presented Asynchronous Score Distillation (ASD), a novel score distillation method that can assist 2D diffusion prior in training 3D generators" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "type": "text", + "content": "ScaleDreamer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 114, + 474, + 304 + ], + "blocks": [ + { + "bbox": [ + 143, + 114, + 474, + 304 + ], + "lines": [ + { + "bbox": [ + 143, + 114, + 474, + 304 + ], + "spans": [ + { + "bbox": [ + 143, + 114, + 474, + 304 + ], + "type": "image", + "image_path": "bab712512d378c9d64acee37b1fdfd89aa6daeef99c029a4aa80b8eebb1529c5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 141, + 306, + 471, + 318 + ], + "lines": [ + { + "bbox": [ + 141, + 306, + 471, + 318 + ], + "spans": [ + { + "bbox": [ + 141, + 306, + 471, + 318 + ], + "type": "text", + "content": "Fig. 7: The qualitative results of the ablation study on the timestep interval " + }, + { + "bbox": [ + 141, + 306, + 471, + 318 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 141, + 306, + 471, + 318 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 202, + 319, + 411, + 397 + ], + "blocks": [ + { + "bbox": [ + 202, + 319, + 411, + 397 + ], + "lines": [ + { + "bbox": [ + 202, + 319, + 411, + 397 + ], + "spans": [ + { + "bbox": [ + 202, + 319, + 411, + 397 + ], + "type": "table", + "html": "
ParamSim ↑R@1 ↑
Δt = η(t - Tmin)η = 0.10.2140.178
η = 0.20.2140.180
η = 00.2350.267
Δt ~ U[0, η(t - Tmin)]η = 0.10.2370.276
η = 0.20.2290.237
", + "image_path": "df0fa4d5c6e3ffce1a87ada3d118b1dcfbc640b7c3685076b0ee621b16d0c98e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 398, + 477, + 409 + ], + "lines": [ + { + "bbox": [ + 134, + 398, + 477, + 409 + ], + "spans": [ + { + "bbox": [ + 134, + 398, + 477, + 409 + ], + "type": "text", + "content": "Table 4: The quantitative results of the ablation study on the timestep interval " + }, + { + "bbox": [ + 134, + 398, + 477, + 409 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 134, + 398, + 477, + 409 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 420, + 482, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 420, + 482, + 516 + ], + "spans": [ + { + "bbox": [ + 130, + 420, + 482, + 516 + ], + "type": "text", + "content": "with a scalable size of text prompts. By shifting the diffusion timestep to earlier stages, our ASD can effectively predict the noise prediction error to align the diffusion model with the distribution of rendered images, while preserving the superior text comprehension capability of pre-trained models, thus facilitating stable training with high-fidelity generation results. Our extensive experiments revealed that ASD performed consistently well on datasets of various sizes, being able to manage as much as 100k prompts. Though ASD has shown improvements over earlier score distillation approaches, there remain some limitations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 517, + 482, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 517, + 482, + 588 + ], + "spans": [ + { + "bbox": [ + 130, + 517, + 482, + 588 + ], + "type": "text", + "content": "For man-made objects that have very regular shapes, such as chairs or airplanes, the performance of our model will lag behind those data-driven methods, which benefit from an abundance of relevant data. We foresee opportunities to combine the advantages of data-driven and score distillation methodologies to improve text-to-3D capabilities in a more comprehensive manner in the future research." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 612, + 241, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 612, + 241, + 625 + ], + "spans": [ + { + "bbox": [ + 132, + 612, + 241, + 625 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "This work is supported in part by the Beijing Science and Technology Plan Project Z231100005923033, and the InnoHK program." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "Z.Ma et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 142, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 142, + 479, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 142, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 138, + 142, + 479, + 163 + ], + "type": "text", + "content": "1. Stable-diffusion-v2.1-base. https://huggingface.co/stabilityai/stable-diffusion-2-1-base" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 165, + 480, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 165, + 480, + 198 + ], + "spans": [ + { + "bbox": [ + 138, + 165, + 480, + 198 + ], + "type": "text", + "content": "2. Armandpour, M., Zheng, H., Sadeghian, A., Sadeghian, A., Zhou, M.: Re-imagine the negative prompt algorithm: Transform 2d diffusion into 3d, alleviate janus problem and beyond. arXiv preprint arXiv:2304.04968 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 198, + 480, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 198, + 480, + 231 + ], + "spans": [ + { + "bbox": [ + 138, + 198, + 480, + 231 + ], + "type": "text", + "content": "3. Babu, S., Liu, R., Zhou, A., Maire, M., Shakhnarovich, G., Hanocka, R.: Hyperfields: Towards zero-shot generation of nerfs from text. arXiv preprint arXiv:2310.17075 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 232, + 480, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 232, + 480, + 264 + ], + "spans": [ + { + "bbox": [ + 138, + 232, + 480, + 264 + ], + "type": "text", + "content": "4. Bahmani, S., Park, J.J., Paschalidou, D., Yan, X., Wetzstein, G., Guibas, L., Tagliasacchi, A.: Cc3d: Layout-conditioned generation of compositional 3d scenes. arXiv preprint arXiv:2303.12074 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 265, + 480, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 265, + 480, + 297 + ], + "spans": [ + { + "bbox": [ + 138, + 265, + 480, + 297 + ], + "type": "text", + "content": "5. Balaji, Y., Nah, S., Huang, X., Vahdat, A., Song, J., Kreis, K., Aittala, M., Aila, T., Laine, S., Catanzaro, B., et al.: edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 298, + 480, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 298, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 138, + 298, + 480, + 319 + ], + "type": "text", + "content": "6. Cao, Z., Hong, F., Wu, T., Pan, L., Liu, Z.: Large-vocabulary 3d diffusion model with transformer. arXiv preprint arXiv:2309.07920 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 321, + 480, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 321, + 480, + 364 + ], + "spans": [ + { + "bbox": [ + 138, + 321, + 480, + 364 + ], + "type": "text", + "content": "7. Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., De Mello, S., Gallo, O., Guibas, L.J., Tremblay, J., Khamis, S., et al.: Efficient geometry-aware 3d generative adversarial networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16123-16133 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 365, + 480, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 365, + 480, + 397 + ], + "spans": [ + { + "bbox": [ + 138, + 365, + 480, + 397 + ], + "type": "text", + "content": "8. Chen, R., Chen, Y., Jiao, N., Jia, K.: *Fantasia3d: Disentangling geometry* and appearance for high-quality text-to-3d content creation. arXiv preprint arXiv:2303.13873 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 399, + 480, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 399, + 480, + 419 + ], + "spans": [ + { + "bbox": [ + 138, + 399, + 480, + 419 + ], + "type": "text", + "content": "9. Chen, Z., Wang, F., Liu, H.: Text-to-3d using gaussian splatting. arXiv preprint arXiv:2309.16585 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 134, + 421, + 480, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 421, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 134, + 421, + 480, + 453 + ], + "type": "text", + "content": "10. Deitke, M., Liu, R., Wallingford, M., Ngo, H., Michel, O., Kusupati, A., Fan, A., Laforte, C., Voleti, V., Gadre, S.Y., et al.: Objaverse-xl: A universe of " + }, + { + "bbox": [ + 134, + 421, + 480, + 453 + ], + "type": "inline_equation", + "content": "10\\mathrm{m} + 3\\mathrm{d}" + }, + { + "bbox": [ + 134, + 421, + 480, + 453 + ], + "type": "text", + "content": " objects. arXiv preprint arXiv:2307.05663 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 454, + 480, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 454, + 480, + 486 + ], + "spans": [ + { + "bbox": [ + 134, + 454, + 480, + 486 + ], + "type": "text", + "content": "11. Ding, L., Dong, S., Huang, Z., Wang, Z., Zhang, Y., Gong, K., Xu, D., Xue, T.: Text-to-3d generation with bidirectional diffusion using both 2d and 3d priors. arXiv preprint arXiv:2312.04963 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 487, + 480, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 487, + 480, + 520 + ], + "spans": [ + { + "bbox": [ + 134, + 487, + 480, + 520 + ], + "type": "text", + "content": "12. Guo, P., Hao, H., Caccavale, A., Ren, Z., Zhang, E., Shan, Q., Sankar, A., Schwing, A.G., Colburn, A., Ma, F.: Stabledreamer: Taming noisy score distillation sampling for text-to-3d. arXiv preprint arXiv:2312.02189 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 521, + 480, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 521, + 480, + 553 + ], + "spans": [ + { + "bbox": [ + 134, + 521, + 480, + 553 + ], + "type": "text", + "content": "13. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 554, + 480, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 554, + 480, + 586 + ], + "spans": [ + { + "bbox": [ + 134, + 554, + 480, + 586 + ], + "type": "text", + "content": "14. He, Y., Bai, Y., Lin, M., Zhao, W., Hu, Y., Sheng, J., Yi, R., Li, J., Liu, Y.J.: T" + }, + { + "bbox": [ + 134, + 554, + 480, + 586 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 134, + 554, + 480, + 586 + ], + "type": "text", + "content": "bench: Benchmarking current progress in text-to-3d generation. arXiv preprint arXiv:2310.02977 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 587, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 587, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 134, + 587, + 480, + 609 + ], + "type": "text", + "content": "15. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 610, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 610, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 610, + 480, + 632 + ], + "type": "text", + "content": "16. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 134, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 632, + 480, + 665 + ], + "type": "text", + "content": "17. Hong, Y., Zhang, K., Gu, J., Bi, S., Zhou, Y., Liu, D., Liu, F., Sunkavalli, K., Bui, T., Tan, H.: Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400 (2023)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "type": "text", + "content": "ScaleDreamer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "18. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "type": "text", + "content": "19. Huang, Y., Wang, J., Shi, Y., Qi, X., Zha, Z.J., Zhang, L.: Dreamtime: An improved optimization strategy for text-to-3d content creation. arXiv preprint arXiv:2306.12422 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 184, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 184, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 184, + 480, + 205 + ], + "type": "text", + "content": "20. Jiang, L., Wang, L.: Brightdreamer: Generic 3d gaussian generative framework for fast text-to-3d synthesis. arXiv preprint arXiv:2403.11273 (2024)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 206, + 480, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 206, + 480, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 206, + 480, + 228 + ], + "type": "text", + "content": "21. Jun, H., Nichol, A.: Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 229, + 480, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 229, + 480, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 229, + 480, + 262 + ], + "type": "text", + "content": "22. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG) 42(4), 1-14 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 262, + 448, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 262, + 448, + 274 + ], + "spans": [ + { + "bbox": [ + 132, + 262, + 448, + 274 + ], + "type": "text", + "content": "23. Koster, R.: Theory of fun for game design. \"O'Reilly Media, Inc.\" (2013)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 274, + 480, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 480, + 296 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 480, + 296 + ], + "type": "text", + "content": "24. Lee, K., Sohn, K., Shin, J.: Dreamflow: High-quality text-to-3d generation by approximating probability flow. arXiv preprint arXiv:2403.14966 (2024)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 297, + 480, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 297, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 480, + 319 + ], + "type": "text", + "content": "25. Li, M., Zhou, P., Liu, J.W., Keppo, J., Lin, M., Yan, S., Xu, X.: Instant3d: Instant text-to-3d generation. arXiv preprint arXiv:2311.08403 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 319, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 319, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 480, + 342 + ], + "type": "text", + "content": "26. Li, W., Chen, R., Chen, X., Tan, P.: Sweetdreamer: Aligning geometric priors in 2d diffusion for consistent text-to-3d. arXiv preprint arXiv:2310.02596 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 342, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 480, + 374 + ], + "type": "text", + "content": "27. Li, Z., Chen, Y., Zhao, L., Liu, P.: Mvcontrol: Adding conditional control to multi-view diffusion for controllable text-to-3d generation. arXiv preprint arXiv:2311.14494 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 375, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 375, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 375, + 480, + 407 + ], + "type": "text", + "content": "28. Liang, Y., Yang, X., Lin, J., Li, H., Xu, X., Chen, Y.: Luciddreamer: Towards high-fidelity text-to-3d generation via interval score matching. arXiv preprint arXiv:2311.11284 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 409, + 480, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 409, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 409, + 480, + 453 + ], + "type": "text", + "content": "29. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3d: High-resolution text-to-3d content creation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 300–309 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 453, + 480, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 480, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 480, + 475 + ], + "type": "text", + "content": "30. Lin, Y., Clark, R., Torr, P.: Dreamampolisher: Towards high-quality text-to-3d generation via geometric diffusion. arXiv preprint arXiv:2403.17237 (2024)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 476, + 480, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 476, + 480, + 508 + ], + "spans": [ + { + "bbox": [ + 132, + 476, + 480, + 508 + ], + "type": "text", + "content": "31. Liu, Y.T., Luo, G., Sun, H., Yin, W., Guo, Y.C., Zhang, S.H.: Pi3d: Efficient text-to-3d generation with pseudo-image diffusion. arXiv preprint arXiv:2312.09069 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 509, + 480, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 509, + 480, + 542 + ], + "spans": [ + { + "bbox": [ + 132, + 509, + 480, + 542 + ], + "type": "text", + "content": "32. Liu, Z., Li, Y., Lin, Y., Yu, X., Peng, S., Cao, Y.P., Qi, X., Huang, X., Liang, D., Ouyang, W.: Unidream: Unifying diffusion priors for relightable text-to-3d generation. arXiv preprint arXiv:2312.08754 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 543, + 480, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 543, + 480, + 576 + ], + "spans": [ + { + "bbox": [ + 132, + 543, + 480, + 576 + ], + "type": "text", + "content": "33. Lorraine, J., Xie, K., Zeng, X., Lin, C.H., Takikawa, T., Sharp, N., Lin, T.Y., Liu, M.Y., Fidler, S., Lucas, J.: Att3d: Amortized text-to-3d object synthesis. arXiv preprint arXiv:2306.07349 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 577, + 480, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 577, + 480, + 598 + ], + "spans": [ + { + "bbox": [ + 132, + 577, + 480, + 598 + ], + "type": "text", + "content": "34. Luo, T., Rockwell, C., Lee, H., Johnson, J.: Scalable 3d captioning with pretrained models. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 599, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 480, + 632 + ], + "type": "text", + "content": "35. Ma, Y., Fan, Y., Ji, J., Wang, H., Sun, X., Jiang, G., Shu, A., Ji, R.: X-dreamer: Creating high-quality 3d content by bridging the domain gap between text-to-2d and text-to-3d generation. arXiv preprint arXiv:2312.00085 (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "36. Mercier, A., Nakhli, R., Reddy, M., Yasarla, R., Cai, H., Porikli, F., Berger, G.: Hexagen3d: Stablediffusion is just one step away from fast and diverse text-to-3d generation. arXiv preprint arXiv:2401.07727 (2024)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "Z.Ma et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "text", + "content": "37. Metzer, G., Richardson, E., Patashnik, O., Giryes, R., Cohen-Or, D.: Latent-nerf for shape-guided generation of 3d shapes and textures. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12663-12673 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 160, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 160, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 160, + 481, + 182 + ], + "type": "text", + "content": "38. Miyato, T., Kataoka, T., Koyama, M., Yoshida, Y.: Spectral normalization for generative adversarial networks. arXiv preprint arXiv:1802.05957 (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 182, + 481, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 481, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 481, + 213 + ], + "type": "text", + "content": "39. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 213, + 481, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 481, + 236 + ], + "type": "text", + "content": "40. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 236, + 481, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 481, + 267 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 481, + 267 + ], + "type": "text", + "content": "41. Qian, G., Cao, J., Siarohin, A., Kant, Y., Wang, C., Vasilkovsky, M., Lee, H.Y., Fang, Y., Skorokhodov, I., Zhuang, P., et al.: Atom: Amortized text-to-mesh using 2d diffusion. arXiv preprint arXiv:2402.00867 (2024)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 267, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 267, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 130, + 267, + 481, + 300 + ], + "type": "text", + "content": "42. Qiu, L., Chen, G., Gu, X., Zuo, Q., Xu, M., Wu, Y., Yuan, W., Dong, Z., Bo, L., Han, X.: Richdreamer: A generalizable normal-depth diffusion model for detail richness in text-to-3d. arXiv preprint arXiv:2311.16918 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 300, + 481, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 481, + 321 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 481, + 321 + ], + "type": "text", + "content": "43. Ren, X., Huang, J., Zeng, X., Museth, K., Fidler, S., Williams, F.: Xcube: Large-scale 3d generative modeling using sparse voxel hierarchies. arXiv preprint (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 321, + 481, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 321, + 481, + 353 + ], + "spans": [ + { + "bbox": [ + 130, + 321, + 481, + 353 + ], + "type": "text", + "content": "44. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 353, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 353, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 353, + 481, + 396 + ], + "type": "text", + "content": "45. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 396, + 481, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 396, + 481, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 481, + 429 + ], + "type": "text", + "content": "46. Schwarz, K., Sauer, A., Niemeyer, M., Liao, Y., Geiger, A.: Voxgraf: Fast 3d-aware image synthesis with sparse voxel grids. Advances in Neural Information Processing Systems 35, 33999-34011 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 429, + 481, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 429, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 429, + 481, + 460 + ], + "type": "text", + "content": "47. Shen, T., Gao, J., Yin, K., Liu, M.Y., Fidler, S.: Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. Advances in Neural Information Processing Systems 34, 6087-6101 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 460, + 481, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 481, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 481, + 483 + ], + "type": "text", + "content": "48. Shi, Y., Wang, P., Ye, J., Long, M., Li, K., Yang, X.: Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 483, + 481, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 483, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 483, + 481, + 525 + ], + "type": "text", + "content": "49. Sitzmann, V., Thies, J., Heide, F., Nießner, M., Wetzstein, G., Zollhofer, M.: Deepvoxels: Learning persistent 3d feature embeddings. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2437-2446 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 525, + 481, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 525, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 525, + 481, + 557 + ], + "type": "text", + "content": "50. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 558, + 481, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 481, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 481, + 579 + ], + "type": "text", + "content": "51. Tang, B., Wang, J., Wu, Z., Zhang, L.: Stable score distillation for high-quality 3d generation. arXiv preprint arXiv:2312.09305 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 579, + 481, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 579, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 130, + 579, + 481, + 611 + ], + "type": "text", + "content": "52. Tang, J., Chen, Z., Chen, X., Wang, T., Zeng, G., Liu, Z.: Lgm: Large multi-view gaussian model for high-resolution 3d content creation. arXiv preprint arXiv:2402.05054 (2024)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 611, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 611, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 130, + 611, + 481, + 633 + ], + "type": "text", + "content": "53. Tang, J., Ren, J., Zhou, H., Liu, Z., Zeng, G.: Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 633, + 481, + 665 + ], + "type": "text", + "content": "54. Tang, Z., Gu, S., Wang, C., Zhang, T., Bao, J., Chen, D., Guo, B.: Volumediffusion: Flexible text-to-3d generation with efficient volumetric encoder. arXiv preprint arXiv:2312.11459 (2023)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "type": "text", + "content": "ScaleDreamer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 148 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 148 + ], + "type": "text", + "content": "55. Thanh-Tung, H., Tran, T.: Catastrophic forgetting and mode collapse in gans. In: 2020 international joint conference on neural networks (ijcnn). pp. 1-10. IEEE (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 182 + ], + "type": "text", + "content": "56. Tsalicoglou, C., Manhardt, F., Tonioni, A., Niemeyer, M., Tombari, F.: Textmesh: Generation of realistic 3d meshes from text prompts. arXiv preprint arXiv:2304.12439 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "type": "text", + "content": "57. Vilesov, A., Chari, P., Kadambi, A.: Cg3d: Compositional generation for text-to-3d via gaussian splatting. arXiv preprint arXiv:2311.17907 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 247 + ], + "type": "text", + "content": "58. Wang, H., Du, X., Li, J., Yeh, R.A., Shakhnarovich, G.: Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12619-12629 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "type": "text", + "content": "59. Wang, Z., Lu, C., Wang, Y., Bao, F., Li, C., Su, H., Zhu, J.: Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 282, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 480, + 314 + ], + "type": "text", + "content": "60. Wei, Y., Zhang, Y., Ji, Z., Bai, J., Zhang, L., Zuo, W.: Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. arXiv preprint arXiv:2302.13848 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "type": "text", + "content": "61. Wohlgenannt, I., Simons, A., Stieglitz, S.: Virtual reality. Business & Information Systems Engineering 62, 455-461 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 337, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 337, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 337, + 480, + 357 + ], + "type": "text", + "content": "62. Wu, R., Sun, L., Ma, Z., Zhang, L.: One-step effective diffusion network for real-world image super-resolution. arXiv preprint arXiv:2406.08177 (2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 358, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 402 + ], + "type": "text", + "content": "63. Wu, T., Zhang, J., Fu, X., Wang, Y., Ren, J., Pan, L., Wu, W., Yang, L., Wang, J., Qian, C., et al.: Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 803-814 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "type": "text", + "content": "64. Wu, Z., Zhou, P., Yi, X., Yuan, X., Zhang, H.: Consistent3d: Towards consistent high-fidelity text-to-3d generation with deterministic sampling prior. arXiv preprint arXiv:2401.09050 (2024)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 436, + 480, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 436, + 480, + 467 + ], + "spans": [ + { + "bbox": [ + 132, + 436, + 480, + 467 + ], + "type": "text", + "content": "65. Xie, K., Lorraine, J., Cao, T., Gao, J., Lucas, J., Torralba, A., Fidler, S., Zeng, X.: Latte3d: Large-scale amortized text-to-enhanced3d synthesis. arXiv preprint arXiv:2403.15385 (2024)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "type": "text", + "content": "66. Xu, Y., Tan, H., Luan, F., Bi, S., Wang, P., Li, J., Shi, Z., Sunkavalli, K., Wetzstein, G., Xu, Z., et al.: Dmv3d: Denoising multi-view diffusion using 3d large reconstruction model. arXiv preprint arXiv:2311.09217 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "type": "text", + "content": "67. Yang, Z., Feng, R., Zhang, H., Shen, Y., Zhu, K., Huang, L., Zhang, Y., Liu, Y., Zhao, D., Zhou, J., et al.: Eliminating lipschitz singularities in diffusion models. arXiv preprint arXiv:2306.11251 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "type": "text", + "content": "68. Yang, Z., Feng, R., Zhang, H., Shen, Y., Zhu, K., Huang, L., Zhang, Y., Liu, Y., Zhao, D., Zhou, J., et al.: Lipschitz singularities in diffusion models. In: The Twelfth International Conference on Learning Representations (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "type": "text", + "content": "69. Yariv, L., Gu, J., Kasten, Y., Lipman, Y.: Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems 34, 4805-4815 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 589, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 480, + 621 + ], + "type": "text", + "content": "70. Yi, T., Fang, J., Wu, G., Xie, L., Zhang, X., Liu, W., Tian, Q., Wang, X.: Gaussian dreamer: Fast generation from text to 3d gaussian splatting with point cloud priors. arXiv preprint arXiv:2310.08529 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 622, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 480, + 665 + ], + "type": "text", + "content": "71. Yu, X., Xu, M., Zhang, Y., Liu, H., Ye, C., Wu, Y., Yan, Z., Zhu, C., Xiong, Z., Liang, T., et al.: Mvimgnet: A large-scale dataset of multi-view images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9150-9161 (2023)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "Z.Ma et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 217 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "text", + "content": "72. Yu, X., Guo, Y.C., Li, Y., Liang, D., Zhang, S.H., Qi, X.: Text-to-3d with classifier score distillation. arXiv preprint arXiv:2310.19415 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 171 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 171 + ], + "type": "text", + "content": "73. Zhao, M., Zhao, C., Liang, X., Li, L., Zhao, Z., Hu, Z., Fan, C., Yu, X.: Efficient-dreamer: High-fidelity and robust 3d creation via orthogonal-view diffusion prior. arXiv preprint arXiv:2308.13223 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 171, + 482, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 171, + 482, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 171, + 482, + 194 + ], + "type": "text", + "content": "74. Zhao, R., Wang, Z., Wang, Y., Zhou, Z., Zhu, J.: Flexidreamer: Single image-to-3d generation with flexicubes. arXiv preprint arXiv:2404.00987 (2024)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 194, + 482, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 194, + 482, + 217 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 482, + 217 + ], + "type": "text", + "content": "75. Zhou, L., Shih, A., Meng, C., Ermon, S.: Dreampropeller: Supercharge text-to-3d generation with parallel sampling. arXiv preprint arXiv:2311.17082 (2023)" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "type": "text", + "content": "ScaleDreamer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Scaling Backwards_ Minimal Synthetic Pre-training_/82902d64-0219-4fba-9a2b-20c3a32a53cf_content_list.json b/2024/Scaling Backwards_ Minimal Synthetic Pre-training_/82902d64-0219-4fba-9a2b-20c3a32a53cf_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bc2a5c9879d819474761d52678382588aa58dc2b --- /dev/null +++ b/2024/Scaling Backwards_ Minimal Synthetic Pre-training_/82902d64-0219-4fba-9a2b-20c3a32a53cf_content_list.json @@ -0,0 +1,2031 @@ +[ + { + "type": "text", + "text": "Scaling Backwards: Minimal Synthetic Pre-training?", + "text_level": 1, + "bbox": [ + 313, + 141, + 686, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ryo Nakamura $^{1,*}$ , Ryu Tadokoro $^{2,*}$ , Ryosuke Yamada $^{1}$ , Yuki M. Asano $^{3}$ , Iro Laina $^{4}$ , Christian Rupprecht $^{4}$ , Nakamasa Inoue $^{5}$ , Rio Yokota $^{5}$ , and Hirokatsu Kataoka $^{1}$", + "bbox": [ + 225, + 212, + 774, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ National Institute of Advanced Industrial Science and Technology (AIST)", + "bbox": [ + 246, + 268, + 750, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Tohoku University", + "bbox": [ + 431, + 284, + 570, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 University of Amsterdam, now at University of Technology Nuremberg", + "bbox": [ + 256, + 297, + 743, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ University of Oxford", + "bbox": [ + 424, + 311, + 576, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "5 Tokyo Institute of Technology", + "bbox": [ + 393, + 325, + 607, + 339 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Pre-training and transfer learning are an important building block of current computer vision systems. While pre-training is usually performed on large real-world image datasets, in this paper we ask whether this is truly necessary. To this end, we search for a minimal, purely synthetic pre-training dataset that allows us to achieve performance similar to the 1 million images of ImageNet-1k. We construct such a dataset from a single fractal with perturbations. With this, we contribute three main findings. (i) We show that pre-training is effective even with minimal synthetic images, with performance on par with large-scale pre-training datasets like ImageNet-1k for full fine-tuning. (ii) We investigate the single parameter with which we construct artificial categories for our dataset. We find that while the shape differences can be indistinguishable to humans, they are crucial for obtaining strong performances. (iii) Finally, we investigate the minimal requirements for successful pre-training. Surprisingly, we find that a substantial reduction of synthetic images from 1k to 1 can even lead to an increase in pre-training performance, a motivation to further investigate \"scaling backwards\". Finally, we extend our method from synthetic images to real images to see if a single real image can show similar pre-training effect through shape augmentation. We find that the use of grayscale images and affine transformations allows even real images to \"scale backwards\". The code is available at https://github.com/SUPER-TADORY/1p-frac.", + "bbox": [ + 259, + 381, + 743, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Synthetic pre-training $\\cdot$ Limited data $\\cdot$ Vision transformers", + "bbox": [ + 261, + 700, + 740, + 714 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 743, + 375, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In image recognition, pre-training allows discovering fundamental visual representations for downstream task applications. Pre-training enhances the performance of visual tasks and enables the use of small-scale task-specific datasets.", + "bbox": [ + 212, + 768, + 785, + 814 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* These authors contributed equally", + "bbox": [ + 230, + 825, + 475, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b7a7bdbc988c958c7b3f3b6dca22bfcf87384e4cd9567b3d49c86ed52a1d8076.jpg", + "image_caption": [ + "Fig. 1: Comparison of ImageNet-1k, FractalDB and 1p-frac (ours). 1p-frac consists of only a single fractal for pre-training. With 1p-frac, neural networks learn to classify perturbations applied to the fractal. In our study \"single\" means a very narrow distribution over parameters that leads to images that are roughly equivalent from a human visual perspective. While the shape differences of perturbed images can be indistinguishable to humans, models pre-trained on 1p-frac achieve comparable performance with those pre-trained on ImageNet-1k or FractalDB." + ], + "image_footnote": [], + "bbox": [ + 223, + 146, + 782, + 277 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, pre-training has been used as a key technology to construct foundation models trained on massive datasets with over hundreds of millions of images. In some cases, a foundation model enables adaptation of zero-shot recognition without the need for additional data.", + "bbox": [ + 212, + 396, + 784, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Pre-training is often interpreted as discovering universal structures in large-scale datasets that later facilitate adaptation to down-stream tasks. In this paper, we challenge this interpretation by providing a minimal pre-training dataset, generated from a single fractal, that achieves similar downstream performance. At the heart of this investigation is the question of whether pre-training might simply be a better weight initialization rather than the discovery of useful visual concepts. If true, performing expensive pre-training with hundreds of millions of images might not be necessary. This additionally frees pre-training from licensing or ethical issues.", + "bbox": [ + 212, + 458, + 787, + 593 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Since the rise of deep neural networks, the ImageNet dataset [16] has been one of the most commonly used pre-training datasets. Originally, pre-training has been conducted through supervised learning (SL) with human-provided teacher labels. However, it has become clear that pre-training can also be achieved without human-provided labels, through self-supervised learning (SSL) [7,9-12,17, 21-23,37,38,56].", + "bbox": [ + 212, + 595, + 787, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this context, Asano et al. [3] successfully acquired visual representations while dramatically reducing the number of required images. They concluded that SSL can yield sufficient image representations even with a single training example, but only for earlier layers of a recognition model. However, it is unclear how these findings translate to modern architectures and representation learning methods. Based on this, vision transformers (ViT) [18] have been pre-trained with only 2,040 real images [52] through an instance discrimination learning signal.", + "bbox": [ + 212, + 686, + 787, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "More recently, it has been shown that basic visual representations can be acquired without using real images and human-provided labels. The trend of", + "bbox": [ + 214, + 809, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "R. Nakamura and R. Tadokoro et al.", + "bbox": [ + 271, + 114, + 517, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "artificially generated labeled images is on the rise for synthetic pre-training [5, 28, 30, 46]. Formula-driven supervised learning (FDSL) generates images from generative formulas, and labels from their parameters [30]. Under the FDSL framework, one can adapt the synthetic pre-training image dataset by altering formulas [24, 27, 29, 44]. While a million-order image dataset was constructed in FractalDB [30], our findings suggest that synthetic pre-training can be reduced to significantly fewer fractal images.", + "bbox": [ + 212, + 146, + 782, + 252 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Motivated by these findings, we believe it is possible to find the key essence of pre-training for image recognition. ViT training can be done with as few as 1,000 artificially generated images [35]. Here, we believe that equivalent performance can be achieved with even fewer images. This consideration is undoubtedly important as we approach a minimal synthetic pre-training dataset in image recognition, which goes against the trend of foundation models toward increasing the dataset scale.", + "bbox": [ + 212, + 253, + 784, + 358 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the present paper, we thus introduce a minimal synthetic dataset, namely 1-parameter Fractal as Data (1p-frac), which consists of a single fractal as shown in Figure 1, as well as a loss function for pre-training with it. Our contributions regarding minimal synthetic pre-training are as follows:", + "bbox": [ + 212, + 362, + 784, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Ordinal minimalism: We introduce the locally perturbed cross entropy (LPCE) loss for pre-training with a single fractal. It utilizes perturbed fractal images for training, where neural networks learn to classify small perturbations. In experiments, we demonstrate that pre-training can be performed even with a single fractal. The pre-training effect of 1p-frac is comparable to that of a million-scale labeled image dataset.", + "bbox": [ + 212, + 425, + 784, + 515 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Distributional minimalism: We introduce the locally integrated empirical (LIEP) distribution $p_{\\Delta}$ that has a controllable perturbation degree $\\Delta$ to investigate the minimal support of the probability density distribution of synthesized images. We observed positive pre-training effects even with a small $\\Delta$ producing shape differences that cannot be distinguished by humans. We also show that if $\\Delta$ is too small the visual pre-training collapses. From these observations, we establish the general bounds for generating good pre-training images from a mathematical formula.", + "bbox": [ + 212, + 518, + 782, + 637 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Instance minimalism: Based on the experimental results, the synthetic images should not simply contain complex shapes. The use of recursive image patterns, similar to objects in nature, should be applied in visual pretraining. Experiments with augmented categories from a real image have shown that good pre-training effects can be achieved by performing \"affine transformations on edge", + "bbox": [ + 212, + 641, + 537, + 775 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/4eb186fb94b5035971ec5682038b06c0d7eb5b07816e5e3608c0dcab4f097897.jpg", + "table_caption": [ + "Table 1: Scaling backwards in synthetic pre-training (Accuracies on CIFAR-100, Real: ImageNet, Synth: Fractal images)." + ], + "table_footnote": [], + "table_body": "
Type\\#Img11k1M
RealN/A76.985.5
Synth84.284.081.6
", + "bbox": [ + 540, + 705, + 772, + 767 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "emphasized objects in grayscale\". These operations are found to be almost synonymous with the configuration of the proposed 1p-frac.", + "bbox": [ + 212, + 777, + 782, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, we significantly reduce the size of the pre-training dataset from 1M images (fractal database (FractalDB) [30]) or 1k images (one-instance fractal", + "bbox": [ + 214, + 809, + 782, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Scaling Backwards", + "bbox": [ + 604, + 114, + 730, + 127 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "database (OFDB) [35]) down to 1 and show that this even improves the pretraining effect (Table 1), which motivates \"scaling backwards\".", + "bbox": [ + 215, + 146, + 784, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 196, + 388, + 214 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The present study focuses primarily on finding the minimal pre-training dataset with artificially generated images.", + "bbox": [ + 215, + 226, + 784, + 257 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pre-training in image recognition. Pre-training in image recognition has demonstrated significant success with vast amounts of labeled images. Pre-training helps neural networks to acquire fundamental visual representations which are essential for improving the performance of diverse downstream tasks in image recognition. In particular, researchers have used larger-scale image datasets for supervised learning (SL), starting from million-scale datasets (e.g., ImageNet [16], Places [59]). More recently, the largest datasets used are reaching billion-scale (e.g., IG-3.5B [34], JFT-300M/3B/4B [15, 42, 53]).", + "bbox": [ + 215, + 268, + 785, + 388 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pre-training can be also achieved without any human intervention. One of the most effective approaches is self-supervised learning (SSL [7,9-12,17,21-23,37, 38,56]), which has been developed in order to alleviate the manual annotations on large-amount of images. SSL has been studied as an expected alternative to supervised pre-training and has sometimes been shown to surpass supervised pre-training.", + "bbox": [ + 215, + 390, + 785, + 479 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Despite the development of supervised and self-supervised pre-training, the number of real images used in visual pre-training continues to increase. Exploratory research questions such as \"what is a minimal synthetic pre-training dataset?\" have received less attention. However, they could give us more insights on what is learned during pre-training and how it could be improved.", + "bbox": [ + 215, + 479, + 785, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pre-training with limited data. In order to address ethical issues such as privacy and fairness associated with real images, a pre-training with a very limited number of real images and synthetic pre-training without any real images have been on the rise. In this context, synthetic pre-training with copy-and-paste learning [19,39], domain randomization from 3D graphics [43,46], learning from primitive patterns [4,5,41], and dataset distillation [8,50,57] have been recognized as promising approaches to acquiring visual representations. More recently, a sophisticated approach, formula-driven supervised learning (FDSL) [28,30], enabled automatic construction of a labeled image dataset without any real images or human annotations. FDSL can generate images and their corresponding labels using a mathematical formula. Along these lines, One-instance FractalDB (OFDB) has successfully pre-trained a ViT using only 1,000 synthetic images generated from a mathematical formula, without using real images [35]. Moreover, several studies have been reported in the context of extremely limited data [3,49,52]. In particular, Asano et al. have shown that learning visual representations is possible using a single image for self-supervised learning. Venkataramanan et al. achieved a comparable method to ImageNet pre-training with only 10 walking videos and their cropped images.", + "bbox": [ + 215, + 566, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 230, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "R. Nakamura and R. Tadokoro et al.", + "bbox": [ + 271, + 114, + 517, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By analyzing the acquisition of visual representations through the use of minimal synthetic pre-training datasets, we believe that the FDSL framework will be key to discovering the essence of pre-training mechanisms in image recognition.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Scaling Backwards with a Single Fractal", + "text_level": 1, + "bbox": [ + 215, + 214, + 638, + 232 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This section presents the proposed methodology to explore the minimal requirements for successful pre-training by utilizing purely synthetic images of fractals. Specifically, we introduce the 1 parameter Fractal as Data (1p-frac), which consists of only a single fractal, with a method to pre-train neural networks on it. Our key idea is to introduce the locally integrated empirical (LIEP) distribution $p_{\\Delta}$ over perturbed fractal images, enabling pre-training even with a single fractal image. Since the LIEP distribution is designed so that it converges to the empirical distribution $p_{\\mathrm{data}}(x) = \\delta (x - I)$ of a single image $I$ when the perturbation degree $\\Delta \\in \\mathbb{R}_{\\geq 0}$ goes to zero, we can narrow down the support of the distribution by decreasing $\\Delta$ as shown in Figure 2a. Below, we begin with a preliminary for defining the empirical distributions of two fractal databases, namely FractalDB [30] and OFDB [35]. We then introduce 1p-frac with the LIEP distribution.", + "bbox": [ + 212, + 246, + 787, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Preliminary", + "text_level": 1, + "bbox": [ + 215, + 464, + 361, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "FractalDB [30]. Kataoka et al. have introduced a method for effectively pretraining neural networks with FractalDB, a set of fractal images generated by the iterated function systems (IFSs). Specifically, FractalDB $\\mathcal{F}$ consists of one million synthesized images: $\\mathcal{F} = \\{(\\varOmega_c,\\{I_i^c\\}_{i = 0}^{M - 1})\\}_{c = 0}^{C - 1}$ , where $\\varOmega_{c}$ is an IFS, $I_{i}^{c}$ is a fractal image generated by $\\varOmega_{c}$ , $C = 1,000$ is the number of fractal categories, and $M = 1,000$ is the number of images per category. Each IFS characterizes each fractal category $c$ and is defined as follows:", + "bbox": [ + 212, + 488, + 787, + 594 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Omega_ {c} = \\left\\{\\mathbb {R} ^ {2}; w _ {1}, w _ {2}, \\dots , w _ {N _ {c}}; p _ {1}, p _ {2}, \\dots , p _ {N _ {c}} \\right\\}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 348, + 606, + 784, + 622 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $w_{j}:\\mathbb{R}^{2}\\to \\mathbb{R}^{2}$ is a 2D affine transformation given by", + "bbox": [ + 214, + 633, + 640, + 648 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nw _ {j} (\\boldsymbol {v}) = \\left[ \\begin{array}{l l} a _ {j} & b _ {j} \\\\ c _ {j} & d _ {j} \\end{array} \\right] \\boldsymbol {v} + \\left[ \\begin{array}{l} e _ {j} \\\\ f _ {j} \\end{array} \\right] (\\boldsymbol {v} \\in \\mathbb {R} ^ {2}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 660, + 784, + 693 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and $p_j$ is a probability mass distribution. Each fractal image $I_i^c$ renders a fractal $F = \\{\\pmb{v}_t\\}_{t=0}^T \\subset \\mathbb{R}^2$ into a 2D image, where points $\\pmb{v}_t$ are determined by a recurrence relation that applies affine transformations as $\\pmb{v}_{t+1} = w_{\\sigma_t}(\\pmb{v}_t)$ for $t = 0, 1, 2, \\dots, T-1$ . Here, the initial point is set as $\\pmb{v}_0 = (0,0)^\\top$ , and the index $\\sigma_t$ is sampled at each $t$ independently from the probability mass distribution $p(\\sigma_t = j) = p_j$ . Pre-training with FractalDB utilizes the cross-entropy loss function:", + "bbox": [ + 214, + 702, + 787, + 806 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = - \\mathbb {E} _ {x, y \\sim p _ {\\mathrm {d a t a}}} [ \\log p _ {\\theta} (y | x) ], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 398, + 820, + 784, + 835 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Scaling Backwards", + "bbox": [ + 604, + 114, + 730, + 127 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d1a098d60ba4748ad3a145adc83d138d79f0d6283419d4c7ebafc33e335b120d.jpg", + "image_caption": [ + "Fig. 2: Scaling backwards from many images to a single synthetic image. (a) Empirical distribution $p_{\\mathrm{data}}$ . Colors indicate classes. With a single image, the distribution is given by a single Dirac's delta function. (b) LIEP distribution $p_{\\Delta}$ . The support of the distribution narrows as the degree of perturbation $\\Delta$ decreases. (c) $\\sigma$ -factor for investigating fractal shapes. A small $\\sigma$ produces complex fractals." + ], + "image_footnote": [], + "bbox": [ + 233, + 143, + 767, + 308 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $p_{\\theta}$ is the category distribution predicted by a neural network and $\\theta$ is a set of learnable parameters. The joint empirical distribution $p_{\\mathrm{data}}$ is defined over the dataset as follows:", + "bbox": [ + 212, + 398, + 787, + 443 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\mathrm {d a t a}} (x, y; \\mathcal {F}) = \\frac {1}{M C} \\sum_ {i = 0} ^ {M - 1} \\sum_ {c = 0} ^ {C - 1} \\delta \\left(x - I _ {i} ^ {c}\\right) \\delta (y - c) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 328, + 450, + 785, + 491 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\delta$ is the Dirac's delta function. Models pre-trained on this dataset perform comparable to those pre-trained on real-world image datasets, such as ImageNet-1k and Places365, in some downstream tasks.", + "bbox": [ + 212, + 500, + 787, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "OFDB [35]. This dataset consists of 1,000 fractal images. Specifically, OFDB $\\mathcal{F}_{\\mathrm{OF}}$ involves only one representative image per category, i.e., $\\mathcal{F}_{\\mathrm{OF}} = \\{\\varOmega_c,I_c\\}_{c = 0}^{C - 1}$ . Therefore, the joint empirical distribution reduces to", + "bbox": [ + 212, + 544, + 789, + 589 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\mathrm {d a t a}} (x, y; \\mathcal {F} _ {\\mathrm {O F}}) = \\frac {1}{C} \\sum_ {c = 0} ^ {C - 1} \\delta \\left(x - I _ {c}\\right) \\delta \\left(y - c\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 343, + 598, + 785, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Models pre-trained on this dataset perform comparable to or even better than those pre-trained on FractalDB. This work has shown that there exists a small but essential set of images for visual pre-training. However, reducing the number of fractals $C$ to less than 1,000 degrades the performance.", + "bbox": [ + 212, + 646, + 787, + 708 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Pre-training with a Single Fractal", + "text_level": 1, + "bbox": [ + 214, + 727, + 540, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Scaling backwards. To further facilitate the analysis of what images are minimally required for successful visual pre-training, we introduce 1p-frac, which ultimately reduces the number of IFSs and images to one each as $\\mathcal{F}_{\\mathrm{OP}} = (\\varOmega,I)$ . With this dataset, the empirical distribution is given by", + "bbox": [ + 212, + 750, + 787, + 811 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\mathrm {d a t a}} (x, y; \\mathcal {F} _ {\\mathrm {O P}}) = \\delta (x - I) \\delta (y). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 819, + 785, + 837 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "R. Nakamura and R. Tadokoro et al.", + "bbox": [ + 271, + 114, + 517, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "However, we notice that with this distribution, training neural networks using cross-entropy loss is not straightforward because $p_{\\theta}(y = 0|x) \\equiv 1 (\\forall x)$ gives a trivial solution to the loss minimization problem. To address this, we introduce locally perturbed cross entropy (LPCE) loss $\\mathcal{L}_{\\Delta}$ , a variant of the cross entropy loss defined with the LIEP distribution.", + "bbox": [ + 212, + 146, + 782, + 220 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Definition 1. Let $I_{\\epsilon} \\in \\mathcal{X}$ be a perturbed image, where $\\mathcal{X}$ is a set of images, $\\epsilon \\in \\mathbb{R}^d$ is a small perturbation with $d \\in \\mathbb{N}_{>0}$ and $I_0 = I$ is the original image. We define the LIEP distribution by", + "bbox": [ + 214, + 222, + 782, + 267 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\Delta} (x, y) = \\frac {1}{| \\mathcal {R} _ {\\Delta} |} \\int_ {\\mathcal {R} _ {\\Delta}} \\delta (x - I _ {\\epsilon}) \\delta (y - \\epsilon) d \\epsilon \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 354, + 277, + 782, + 309 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathcal{R}_{\\Delta} \\subset \\mathbb{R}^{d}$ is a compact set containing the origin and $|\\mathcal{R}_{\\Delta}|$ is its volume with order $O(|\\Delta|^{d})$ .", + "bbox": [ + 214, + 320, + 784, + 349 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Definition 2. We define the LPCE loss by", + "bbox": [ + 215, + 349, + 537, + 364 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\Delta} = - \\mathbb {E} _ {x, y \\sim p _ {\\Delta}} \\left[ \\log p _ {\\theta} (y | x) \\right], \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 398, + 378, + 782, + 393 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $p_{\\Delta}$ is the LIEP distribution.", + "bbox": [ + 214, + 402, + 465, + 417 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "If $\\mathcal{R}_{\\Delta}$ is a small hypercube or hypersphere, it is straightforward to see that $p_{\\Delta}$ approaches to the empirical distribution of Eq. (6) when $\\Delta$ goes to zero. Therefore, this loss allows us to analyze visual pre-training effects by narrowing the support of the distribution around a single image.", + "bbox": [ + 212, + 417, + 782, + 477 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "With 1p-frac, we apply perturbation to the affine transformations. As such, the perturbation $\\epsilon$ is in $\\mathbb{R}^{6*j}$ . A perturbed image $I_{\\epsilon}$ is obtained by the noisy affine transformations:", + "bbox": [ + 212, + 479, + 782, + 522 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nw _ {j} \\left(\\boldsymbol {v}; \\boldsymbol {\\epsilon} _ {j}\\right) = \\left(\\left[ \\begin{array}{l l l} a _ {j} & b _ {j} & e _ {j} \\\\ c _ {j} & d _ {j} & f _ {j} \\end{array} \\right] + \\boldsymbol {\\epsilon} _ {j}\\right) \\left[ \\begin{array}{l} \\boldsymbol {v} \\\\ 1 \\end{array} \\right] \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 532, + 782, + 564 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\epsilon_{j} \\in \\mathcal{R}_{\\Delta} = [-\\Delta / 2, \\Delta / 2]^{6 * j} \\subset \\mathbb{R}^{6 * j}$ is a hypercube with a side length $\\Delta$ , and $|\\mathcal{R}_{\\Delta}| = \\Delta^{6 * j}$ . Note that the numerical integration of Eq. (7) is used in practice, which is an approximation obtained by uniformly sampling $L$ points in $\\mathcal{R}_{\\Delta}$ , where we set $L = 1,000$ by default.", + "bbox": [ + 214, + 574, + 782, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visualization. Figure 2b shows examples of perturbed images to compute the LPCE loss. While most of shape differences are indistinguishable to humans, a neural network learns to distinguish the perturbations applied to the single image by minimizing the LPCE loss.", + "bbox": [ + 212, + 635, + 782, + 694 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Complexity of $\\Omega$ . We use the $\\sigma$ -factor proposed by Anderson et al. [1] to evaluate the complexity of IFSs. As shown in Figure 2c, small values of $\\sigma$ produce complex fractal shapes.", + "bbox": [ + 212, + 696, + 782, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 763, + 374, + 780 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Experimental setup. What is minimal visual pre-training, and what properties of images lend themselves to reducing their count? In order to clarify these questions, we conduct experiments following these steps:", + "bbox": [ + 212, + 794, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Scaling Backwards", + "bbox": [ + 604, + 114, + 730, + 127 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Exploration study (Tables 2-4): We verify whether a single fractal is sufficient for pre-training. Moreover, we compare the extent to which the shapes used in pre-training can be simplified, using object contours, Gaussian, and uniform distributions.", + "- Hyperparameter study (Tables 5-7): We investigate the effects of three hyperparameters $\\Delta$ , $\\sigma$ and $L$ .", + "- Scaling study (Table 8): We compare 1p-frac with other large-scale datasets in terms of fine-tuning accuracy on ImageNet-1k.", + "- Analysis and discussion (Tables 9-12): We discuss data augmentation, computational cost for synthesizing images, and pre-training using a single real-world image.", + "- Applications (Tables 13 and 14): We conduct comparisons in various fine-tuning datasets." + ], + "bbox": [ + 225, + 146, + 784, + 340 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Implementation details. In order to verify the effects of pre-training, we measure the accuracy of downstream tasks through fine-tuning. We use ViT [18] for all experiments. Specifically, we employ ViT-Tiny (ViT-T) in the exploration and hyperparameter studies, and employ ViT-Base (ViT-B) for the scaling study. The pre-training parameters and data augmentation methods for pre-training and fine-tuning follow conventional methods in OFDB. In OFDB, data augmentation is based on DeiT. For exploration studies, CIFAR-100 (C100) [32] and/or ImageNet-100 (IN100) [40] are used as fine-tuning datasets. In the comparisons, we assign seven representative datasets used in OFDB paper, namely C10/C100, Cars [31], Flowers [36], Pascal VOC 2012 (VOC12) [20], Places-30 (P30) [59], IN100. See supplementary material for more detailed experimental settings.", + "bbox": [ + 212, + 352, + 785, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Exploration Study", + "text_level": 1, + "bbox": [ + 215, + 542, + 415, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Pre-training with 1p-frac. Table 2 compares 1p-frac with FracdalDB (1M images, 1k categories) and OFDB (1k images, 1k categories). As can be seen, the pre-training effect of 1p-frac is comparable to that of FractalDB and OFDB. The number of fractal categories is significantly reduced from $C = 1,000$ to only 1. This shows the effectiveness and efficiency of 1p-frac and the LPCE loss.", + "bbox": [ + 212, + 566, + 782, + 640 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison with SOTA FDSL datasets. Table 3 compares 1p-frac with two SOTA FDSL datasets: (1) Radial Contour Database (RCDB) [28], which consists of 1 million images of synthetic polygons, and (2) Visual Atoms (VA) [45], which consists of 1 million images of parameterized wave functions. The \"1K\" counterparts in the table reduce the number of images per category to one. We also applied the LPCE loss to these two datasets by applying perturbations to the radios of a polygon or wave. As can be seen, 1p-frac performs the best when a single image is used for pre-training and outperforms the 1K FDSL datasets.", + "bbox": [ + 212, + 643, + 784, + 760 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Pre-training with a noisy image. To validate the necessity of a complex shape when pre-training with a single image, we applied the LPCE loss to two noisy images: an image of Gaussian noise and an image of uniform noise. Since these two noise images are generated by determining the parameters of the parametric distributions, Gaussian and uniform, the LPCE loss can be computed by", + "bbox": [ + 212, + 763, + 782, + 837 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "R. Nakamura and R. Tadokoro et al.", + "bbox": [ + 271, + 114, + 517, + 127 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2840e1eb4603019060f307471709a71fb8b07045f52991440f009c9c509e9c10.jpg", + "table_caption": [ + "Table 2: Comparison of 1p-frac with FractalDB [30] and OFDB [35]. $\\diamond$ indicates the use of LPCE loss. We report top-1 accuracies $(\\%)$ ." + ], + "table_footnote": [], + "table_body": "
Method#ImgC100IN100
Scratch-64.274.9
FractalDB1M81.688.5
OFDB1k84.088.6
1p-frac°184.289.0
", + "bbox": [ + 222, + 239, + 406, + 316 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/1d4a393406fe10db7573100ff361e25f7540c0bb4473864e3829d9db2fa007fa.jpg", + "table_caption": [ + "Table 3: Comparison with SOTA FDSL datasets. $\\diamond$ indicates the use of LPCE loss." + ], + "table_footnote": [], + "table_body": "
Method#ImgC100IN100
RCDB1M81.688.5
RCDB1K80.487.5
RCDB°182.587.9
VA1M84.990.3
VA1K82.188.5
VA°182.688.0
ip-frac°184.289.0
", + "bbox": [ + 416, + 202, + 594, + 327 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/ede00c18282aef3d6296d16bcc6fe6ea14332eb7b05a8f076778efa7dfec3b6f.jpg", + "table_caption": [ + "Table 4: Comparison with pre-training with a single noise image." + ], + "table_footnote": [], + "table_body": "
MethodC100IN100
Gaussian1.15.7
Uniform2.071.1
1p-frac84.289.0
", + "bbox": [ + 612, + 195, + 771, + 321 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/8467e1efaeaf1360a228aaf6f3487d9ee22af2da548f6616f43307edef5e49bf.jpg", + "table_caption": [ + "Table 5: Effects of pertur- Table 6: Effects of $\\sigma$ - bation degree $\\varDelta$ $(\\sigma=3.5)$ $(\\varDelta=0.1)$" + ], + "table_footnote": [], + "table_body": "
ΔC100IN100
0.0011.21.9
0.0119.961.8
0.0583.088.2
0.184.289.0
0.283.488.5
1.082.688.1
", + "bbox": [ + 220, + 366, + 401, + 474 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/c7295f7d4050335ab7d41ae39d91a08104122492bfc63457598fe6adb067b7a6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
σC100IN100
6.081.386.3
5.082.287.1
4.581.986.9
3.584.289.0
4.082.887.9
", + "bbox": [ + 415, + 373, + 571, + 474 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/b591faa3834f12288257d1dfee85354e25db11321e86e22abf7936938e12990a.jpg", + "table_caption": [ + "Table 7: Effects of number of sampling points $L$ for numerical integration." + ], + "table_footnote": [], + "table_body": "
LC100IN100
1678.785.3
6482.388.0
51282.488.0
100084.289.0
", + "bbox": [ + 598, + 388, + 766, + 473 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "giving small perturbations to the parameters. Table 4 shows these images and fine-tuning results. For example, here the $\\epsilon$ is added to the mean vector and covariance matrix of the Gaussian distribution. The results revealed that noise images fail in visual pre-training, performing worse than training from scratch. This suggests that acquiring fundamental visual representations through pretraining requires an image of a certain structured object, such as a fractal.", + "bbox": [ + 212, + 493, + 787, + 585 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Hyperparameter Study", + "text_level": 1, + "bbox": [ + 215, + 611, + 452, + 627 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Perturbation degree. Table 5 shows the results obtained by different perturbation degrees $\\Delta$ for the LIEP distribution. The results indicate that setting the value to 0.1 yields the highest performance. Interestingly, the performance significantly improved when $\\Delta$ was increased from 0.01 to 0.05. This suggests that the support of the empirical distribution must have a certain size to obtain positive pre-training effects.", + "bbox": [ + 212, + 641, + 787, + 732 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "$\\sigma$ -factor. Table 6 investigates the effects of the $\\sigma$ -factor, which measures the complexity of fractal shapes. As can be seen, the most complex shape yields the highest performance. It is worth noting that, even with a fractal of $\\sigma = 6.0$ , which looks like Gaussian noise, we observed positive pre-training effects. This suggests that the shapes obtained from IFS are crucial for pre-training.", + "bbox": [ + 212, + 733, + 787, + 808 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Numerical integration. Tables 7 reduces the number of sampling points $L$ of the numerical integration when approximately computing loss with Eq. (7).", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Scaling Backwards", + "bbox": [ + 604, + 114, + 730, + 127 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/7d47bd1f00a2430aeb3072df717c7900cb30dc88d21ae33966c860919401f33c.jpg", + "table_caption": [ + "Table 8: Comparison of pre-training datasets on ImageNet-1k fine-tuning. Fine-tuning accuracies calculated by using ViT-Base (B) are listed in the table. 21k indicates the number of categories in each pre-training dataset." + ], + "table_footnote": [], + "table_body": "
Pre-training#ImgTypeViT-B
Scratch--79.8
ImageNet-21k14MSL81.8
FractalDB-21k21MFDSL81.8
ExFractalDB-21k21MFDSL82.7
RCDB-21k21MFDSL82.4
VA-21k21MFDSL82.7
OFDB-21k21kFDSL82.2
3D-OFDB-21k21kFDSL82.7
1p-frac (ours)1FDSL82.1
", + "bbox": [ + 218, + 220, + 524, + 375 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/6456c106a50203e0a65bb61fb8bd2b00f4ffd4354646fc23fa82abf6e0c309c2.jpg", + "table_caption": [ + "Table 9: Ablation study on data augmentation methods. The DeiT augmentation [47] is used as a baseline." + ], + "table_footnote": [], + "table_body": "
MethodC100
Baseline84.2
w/o Random Aug. [14]83.4
w/o Random Crop80.1
w/o Rand Aspect83.6
w/o Rand Erasing [58]84.0
w/o Mixup [55]83.6
w/o Cutmix [51]83.3
w/o Mixup/Cutmix80.4
w/o Flipping84.3
w/o Color Jittering84.2
", + "bbox": [ + 545, + 200, + 777, + 375 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We see that larger numbers result in high performance. Interestingly, even with $L = 16$ , the pre-training effect was positive (better than training from scratch).", + "bbox": [ + 215, + 393, + 785, + 426 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Scaling Study", + "text_level": 1, + "bbox": [ + 215, + 446, + 375, + 464 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "ImageNet-1k fine-tuning (Table 8). While this work discusses minimal requirements for visual pre-training, increasing the approximation precision of the numerical integration could further improve the performance and benefit training of large models. In Table 8, we applied the LPCE loss with $L = 21,000$ to the ViT-B model and compare fine-tuning accuracy on ImageNet-1k. Surprisingly, and despite using a single fractal as data, our ViT-B pre-trained on 1p-frac outperforms pre-training on ImageNet-21k. This shows the strength of our approach in particular because we simply keep the same fine-tuning protocols as used for the original Imagenet-21k to ImageNet-1k transfer proposed in [18], putting us at a potential disadvantage.", + "bbox": [ + 212, + 473, + 787, + 625 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.4 Analysis and Discussion", + "text_level": 1, + "bbox": [ + 215, + 648, + 460, + 662 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Relationship with data augmentation (Table 9). We consider that the variation in image patterns also depends on data augmentation. While traditional methods primarily rely on the data augmentation techniques described in the DeiT paper [47], we investigated which data augmentation methods contribute significantly by excluding one or two techniques from the base data augmentation methods and observing the changes in fine-tuning accuracy on C100 dataset. The experimental results in Table 9 reveal that excluding Random Cropping (RandomCrop) significantly deteriorates accuracy, suggesting that omitting parts of the image region and causing partial loss facilitates the acquisition of the ability to focus on any part without relying on grasping the overall shape, greatly reducing the pre-training effect. Furthermore, when both Mixup and Cutmix are", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "R. Nakamura and R. Tadokoro et al.", + "bbox": [ + 271, + 114, + 517, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/e927b5c5bce39f41cc261afcd0d7e44c4cf4d3de29f6d647d4aee7fba64569e8.jpg", + "table_caption": [ + "Table 10: Dataset construction time (hours). FractalDB (1M images), OFDB (1k images) and 1p-frac (1k perturbations for one image) are compared. The processing time is shown separately for fractal category search (Search), image rendering (Render), and total time (Total)." + ], + "table_footnote": [], + "table_body": "
DatasetSearchRenderTotal
FractalDB2.3716.8619.23
OFDB2.370.412.78
1p-frac0.00220.0360.0382
", + "bbox": [ + 223, + 291, + 447, + 361 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/4c0c45ed6ad09ba8ef1a6b3811581f7e7e5eefd229e320905c14f902b1e7bd77.jpg", + "image_caption": [ + "Fig. 3: Linear probing following the experiments from [3]. We verified the performance with $\\{1,2,3,4,6\\}$ layers in the ViT-Tiny model on C100 dataset. Here, 2 and 4 show the use of up to 1-2 and 1-4 layers." + ], + "image_footnote": [], + "bbox": [ + 470, + 148, + 763, + 294 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "excluded, a degradation in accuracy is observed, indicating that mixing images and categories enables better feature recognition.", + "bbox": [ + 212, + 381, + 782, + 411 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Dataset construction times (Table 10). We calculated the processing times in fractal category search and image rendering. By comparing to the conventional approaches, our 1p-frac recorded much faster category search (0.0022 hours nearly equals to 8 seconds) and image rendering (0.036 hours nearly equals to 129 seconds). 1p-frac requires only one category with a parameter set. This very simple procedure leads to efficiently construct a minimum requirement for pre-training in image recognition.", + "bbox": [ + 212, + 412, + 784, + 518 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Linear probing (Figure 3). To compare supervised ImageNet-1k pre-training with our 1p-frac pre-training in the context of early layer visual representations, we executed a linear probing experiment on the C100 dataset. In the ViT-T model with $\\{1,2,3\\}$ layers, the 1p-frac pre-trained model outperformed the model pre-trained on ImageNet-1k. This demonstrates the superior quality of early, low-level representations learned from our dataset compared to ImageNet. In effect, layers up to layer 3 in ViT-T can be sufficiently pre-trained and frozen using our 1p-frac.", + "bbox": [ + 212, + 520, + 784, + 642 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Pre-training with a single real image and shape augmentation.", + "bbox": [ + 215, + 643, + 782, + 657 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "It was found that pretraining effects comparable to a dataset of one million real-world images can be achieved with 1p-frac, but is it possible to perform pre-training similarly with a real image? In this section, we verify whether", + "bbox": [ + 212, + 659, + 429, + 779 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/0bd2c3b5a3be8f5ed962b45e2d63c6123d606f9cc6805f4303e57c9cc3a5b56e.jpg", + "image_caption": [ + "Fig. 4: Shape augmentation with three geometry transformations from a single real image." + ], + "image_footnote": [], + "bbox": [ + 446, + 660, + 527, + 739 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/5ac48b223ad66e29540d7a9c91e7e54528fc10e51f617095edc6497d40f3eb62.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 531, + 661, + 612, + 737 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/d7505d95cded1adba0c183a93544a96bf82273c027b02e0831ebbce2b5d9d17c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 616, + 661, + 697, + 737 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/ecece888f4e4acfba674c84de78801f5c23d4a1bfb0f7bba46b793bd395af5b1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 700, + 661, + 784, + 737 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "visual features can be acquired through pre-training using a single real image (each image in Figure 5) with the LPCE loss using image representations similar to FDSL (Canny [6] images) and geometric transformations for perturbation (\\{affine, elastic, polynomial\\} transformations in Figure 4). The results of pre", + "bbox": [ + 212, + 779, + 784, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Scaling Backwards", + "bbox": [ + 604, + 114, + 730, + 127 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/858924f9b8dbfdcc015dda5ccb40b1d07b941fd790e07d2a5f52d7101df9d932.jpg", + "image_caption": [ + "Fig. 5: The five different images used in the pre-training with a single real image and shape augmentation." + ], + "image_footnote": [], + "bbox": [ + 217, + 146, + 785, + 203 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/32e09d909fb3581b18d764eeb64dd74642b9b765be1b0b09022fb658ed5c04b8.jpg", + "table_caption": [ + "Table 11: Comparison of pre-training effect between RGB and Canny images when shape augmentation is applied. The shape augmentation uses affine transformations." + ], + "table_footnote": [], + "table_body": "
ImageMean Accuracy
RGB image81.6
Canny image82.2
", + "bbox": [ + 228, + 325, + 475, + 385 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/ce742c89fc889e129780686fcfb55299fc5249999ad7fbe5adfb4562b29a82f7.jpg", + "table_caption": [ + "Table 12: Comparison of pre-training effects of affine, elastic, and polynomial image transformations. Real Img type with Canny applied to the images." + ], + "table_footnote": [], + "table_body": "
TransformationMean Accuracy
Affine trans.81.9
Elastic trans.37.0
Polynomial trans.81.3
", + "bbox": [ + 513, + 309, + 777, + 383 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "training effects with $L = 1,000$ are shown in Tables 11 and 12. Note that the results in these tables represent the average score on C100 over the five real images in Figure 5. See the supplementary material for results for each of image. According to the results, we find that contour-emphasized Canny images omitting color information and affine transformation perturbation yield better pre-training, as measured by the fine-tuning accuracies. The operations are similar to the classification of labeled fractal images via IFS.", + "bbox": [ + 212, + 386, + 787, + 492 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.5 Applications", + "text_level": 1, + "bbox": [ + 215, + 513, + 366, + 529 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Fine-tuning on commonly used datasets (Table 13). We compare fine-tuning accuracies in three types of dataset configurations. For supervised learning (SL) we compare pretraining on ImageNet-1k [16] and Places-365 [59], for self-supervised learning (SSL) we compare DINO [7] and masked auto-encoder (MAE) [22] pre-trained on ImageNet and PASS, and models trained on the fractal datasets FDSL on FractalDB-1k and RCDB-1k. Following the setting of one-instance per category [35], we also evaluate using random patch augmentation (\"w/ Aug.\") We compare our 1p-frac with 1-image SSL [3] with DINO and MAE (we use Real Img #1 of Figure 5). The results show that our 1p-frac pretraining with one single image demonstrated high fine-tuning accuracies (87.5% and 88.2% with and without random patch augmentation) which are close to the performance of self-supervised ImageNet-1k pre-training with MAE that utilizes 1.28M images and yields 88.5%. Surprisingly, despite using only one fractal image with perturbations, our 1p-frac without random patch augmentation has demonstrated fine-tuning accuracies that are equal to or even higher than that of conventional OFDB-1k/FractalDB-1k, which uses 1,000 parameter sets with fractal categories and 1,000/1,000,000 images: 1p-frac obtains mean performances of 87.5% vs OFDB-1k 87.2% vs FractalDB-1k 87.1%.", + "bbox": [ + 212, + 537, + 787, + 808 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Fine-tuning on specialized and structured datasets in VTAB [54] (Table 14). We also verify performance on a couple of specialized and structured", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "R. Nakamura and R. Tadokoro et al.", + "bbox": [ + 271, + 114, + 517, + 127 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/8626291342d5d868f4a016f8934a5bfe4de965208f2fd3fbbe3a0007da8588a1.jpg", + "table_caption": [ + "Table 13: Comparison among pre-training methods in fine-tuning accuracies. Best values at each dataset scale are in bold. ViT-T is used for all experiments. #OrgImg indicates number of collected images in real-image datasets or number of original images used in synthetic datasets. Type shows supervised types within supervised learning using cross-entropy loss (SL), self-supervised learning using DINO [7] or MAE [22] (SSL:D/SSL:M), and formula-supervised learning (FDSL)." + ], + "table_footnote": [], + "table_body": "
Pre-training#OrgImgTypeC10C100CarsFlowersVOC12P30IN100Mean
Scratch--80.362.113.771.456.276.274.862.1
Places-365 [59]1.80MSL97.683.989.299.384.6-89.4-
ImageNet-1k [16]1.28MSL98.085.589.999.488.780.0--
ImageNet-1k [16]1.28MSSL:D97.782.488.098.574.778.489.086.9
ImageNet-1k [16]1.28MSSL:M97.485.886.696.383.380.290.088.5
PASS [2]1.43MSSL:D97.584.086.498.682.979.082.987.8
FractalDB-1k [30]1.00MFDSL96.881.686.098.380.678.488.387.1
RCDB-1k [28]1.00MFDSL97.082.286.598.980.979.788.587.6
ImageNet-1k [35]1,000SL94.376.957.394.873.878.284.379.9
ImageNet-1k [35]1,000SSL:D94.978.071.294.675.578.684.982.5
OFDB-1k [35]1,000FDSL96.984.084.597.179.979.988.087.2
3D-OFDB-1k [35]1,000FDSL97.183.885.598.480.880.089.187.8
OFDB-1k w/ Aug. [35]1,000FDSL97.285.387.698.381.480.489.588.5
3D-OFDB-1k w/ Aug. [35]1,000FDSL97.084.785.698.381.279.888.987.9
1-image SSL [3]1SSL:D95.779.573.193.869.080.488.782.8
1-image SSL [3]1SSL:M97.284.782.194.678.580.389.086.6
1p-frac (ours)1FDSL96.984.284.597.480.580.689.087.5
1p-frac w/ Aug. (ours)1FDSL96.584.787.098.180.980.588.988.2
", + "bbox": [ + 217, + 233, + 789, + 500 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "datasets in VTAB (visual task adaptation benchmark; Retinopathy [26], Resisc45 [13], Camelyon [48], CLEVR-Count [25], and sNORB-Azim [33]). The experiment is more suitable to evaluate and compare the pre-training methods since the image domains and their labels are distinct from both real and fractal images on ImageNet and 1p-frac. According to Table 14, our 1p-frac pretrained model performed relatively similar fine-tuning accuracies on all listed datasets in the table. By comparing to self-supervised ImageNet-1k with MAE, our 1p-frac surpassed the fine-tuning accuracies on three (Retinopathy, Resisc45, and sNORB-Azim) out of the five datasets.", + "bbox": [ + 212, + 511, + 787, + 647 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5 Conclusion and Discussion", + "text_level": 1, + "bbox": [ + 215, + 671, + 508, + 686 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "This paper examined what a minimal dataset for synthetic pre-training might look like. We proposed the 1-parameter Fractal as Data (1p-frac) dataset that succeeds in pre-training even with a single fractal by utilizing the locally perturbed cross entropy (LPCE) loss. The following findings were presented:", + "bbox": [ + 212, + 704, + 785, + 763 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "What is a minimal synthetic pre-training dataset? In this paper, we applied IFS to generate labeled images. In this context, preparing a single set of parameters $a - f$ in IFS is the least informative. It recorded fine-tuning accuracy nearly equivalent to conventional methods such as FractalDB and OFDB (Tables 1 and 2), succeeding in pre-training with minimal image pattern utilization.", + "bbox": [ + 212, + 765, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Scaling Backwards", + "bbox": [ + 604, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/642c94a1c4e9918163fa05d2adb0e8d731fea4de5d596eee762c953eeb6c48e5.jpg", + "table_caption": [ + "Table 14: Performance comparisons on five fine-tuning datasets: Retinopathy (Retino) [26], Resisc45 [13], Camelyon [48], CLEVR-Count (CLEVR-C) [25], and sNORB-Azim (sNORB-A) [33]. Best and second-best values are in bold and underlined, respectively." + ], + "table_footnote": [], + "table_body": "
#OrgImgTypeRetinoResisc45CamelyonCLEVR-CsNORB-A
Scratch--66.989.274.446.411.5
ImageNet-1k [16]1.28MSL79.197.082.789.329.2
ImageNet-1k [16]1.28MSSL:M76.795.683.989.526.3
1p-frac (ours)1FDSL77.995.882.786.228.9
", + "bbox": [ + 228, + 200, + 774, + 282 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "tion. Moreover, it is not that any perturbation in image space is acceptable, as Gaussian/uniform distributions failed in pre-training (Table 4). It clarified the importance of geometric transformations with certain regularities like fractal geometry or contour shapes (Table 3).", + "bbox": [ + 212, + 295, + 782, + 354 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "What properties of images lead to image reduction? In 1p-frac, we investigated shape changes and pre-training effects by adopting the LPCE loss and the $\\sigma$ -factor that contribute to fractal shape variation. It was revealed that pre-training efficacy varies with shape variation (Tables 5, 6). For real images, it is not just about including more edge components, but about including affine transformations, which closely resembles the classification of labeled fractal images through IFS (Tables 11 and 12). By extensively abating data augmentation showed, we demonstrated the essential role of random cropping and MixUp.", + "bbox": [ + 212, + 354, + 784, + 474 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Pre-training effects with the proposed method. Supervised pre-training on ImageNet was more accurate for datasets where object labels are assigned to real images (Table 13), but in VTAB's Specialized/Structured datasets (Table 14), performances were relatively close. On the other hand, our 1p-frac pre-trained ViT performed a similar mean accuracy by comparing to MAE self-supervised ImageNet pre-training (MAE 88.5 vs. our 1p-frac w/ Aug. 88.2). Furthermore, increasing the number of sampling points $L$ showed performance improvement even with 21k categories and demonstrated good performance in ImageNet-1k fine-tuning (Table 8), indicating the potential for building large models with limited data resources.", + "bbox": [ + 212, + 477, + 785, + 626 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Broader impact. Our proposed dataset 1p-frac does not suffer from licensing or ethical issues such as biases and has a large potential to serve as a clean pretraining dataset. While not investigated in this paper, this approach also opens the possibility for significant gains in training speed by keeping the data on the GPU and applying transformations there, removing the GPU-CPU transfer bottleneck.", + "bbox": [ + 212, + 628, + 785, + 717 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitation. In this paper, we have explored achieving a minimal synthetic pre-training in image recognition based on fractal geometry with IFS (a single fractal, and their perturbations). Despite this, we believe it is also necessary to find a similarly minimal pre-training dataset containing real images - for some applications and increased interpretability. This might allow for quick and clear calibration of pre-training, similar to how calibration images are used in photography. We did not investigate this direction in this paper but leave this for future work.", + "bbox": [ + 212, + 719, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "R. Nakamura and R. Tadokoro et al.", + "bbox": [ + 271, + 114, + 517, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Computational resource of AI Bridging Cloud Infrastructure (ABCI) provided by National Institute of Advanced Industrial Science and Technology (AIST) was used.", + "bbox": [ + 212, + 179, + 782, + 223 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 242, + 321, + 258 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Anderson, C., Farrell, R.: Improving fractal pre-training. In: Proceedings of the 2022 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 1300-1309 (2022)", + "2. Asano, Y., Rupprecht, C., Zisserman, A., Vedaldi, A.: Pass: An imagenet replacement for self-Supervised pretraining without humans. In: Proceedings of the 2021 Neural Information Processing Systems Track on Datasets and Benchmarks (2021)", + "3. Asano, Y.M., Rupprecht, C., Vedaldi, A.: A critical analysis of self-supervision, or what we can learn from a single image. In: Proceedings of the 2020 International Conference on Learning Representations. (ICLR) (2020)", + "4. Baradad, M., Chen, C.F., Wulff, J., Wang, T., Feris, R., Torralba, A., Isola, P.: Procedural image programs for representation learning. In: Proceedings of the 2022 Advances in Neural Information Processing Systems (NeurIPS) (2022)", + "5. Baradad, M., Wulff, J., Wang, T., Isola, P., Torralba, A.: Learning to See by Looking at Noise. In: Proceedings of the 2021 Advances in Neural Information Processing Systems (NeurIPS). vol. 34, pp. 2556-2569 (2021)", + "6. Canny, J.: A computational approach to edge detection. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 8(6), 679-698 (1986)", + "7. Caron, M., Touvron, H., Misra, I., Jegou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the 2021 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 9650-9660 (2021)", + "8. Cazenavette, G., Wang, T., Torralba, A., Efros, A.A., Zhu, J.Y.: Dataset distillation by matching training trajectories. In: Proceedings of the 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10708-10717 (2022)", + "9. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A Simple Framework for Contrastive Learning of Visual Representations. In: Proceedings of the 2020 International Conference on Machine Learning (ICML). vol. 119, pp. 1597-1607 (2020)", + "0. Chen, X., Fan, H., Girshick, R., He, K.: Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297 (2020)", + "11. Chen, X., He, K.: Exploring simple siamese representation learning. In: Proceedings of the 2021 IEEE/CVF International Conference on Computer Vision and Pattern Recognition (CVPR). pp. 15750-15758 (2021)", + "12. Chen, X., Xie, S., He, K.: An empirical study of training self-supervised vision transformers. In: Proceedings of the 2021 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 9640-9649 (2021)", + "13. Cheng, G., Han, J., Lu, X.: Remote sensing image scene classification: Benchmark and state of the art. Proceedings of the IEEE 105(10), 1865-1883 (2017)", + "14. Cubuk, E.D., Zoph, B., Shlens, J., Le, Q.V.: Randaugment: Practical data augmentation with no separate search. arXiv preprint abs/2104.07918 (2019)" + ], + "bbox": [ + 218, + 277, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Scaling Backwards", + "bbox": [ + 604, + 114, + 730, + 127 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Dehghani, M., Djolonga, J., Mustafa, B., Padlewski, P., Heek, J., Gilmer, J., Steiner, A.P., Caron, M., Geirhos, R., Alabdulmohsin, I., Jenatton, R., Beyer, L., Tschannen, M., Arnab, A., Wang, X., Riquelme Ruiz, C., Minderer, M., Puigcerver, J., Evci, U., Kumar, M., Steenkiste, S.V., Elsayed, G.F., Mahendran, A., Yu, F., Oliver, A., Huot, F., Bastings, J., Collier, M., Gritsenko, A.A., Birodkar, V., Vasconcelos, C.N., Tay, Y., Mensink, T., Kolesnikov, A., Pavetic, F., Tran, D., Kipf, T., Lucic, M., Zhai, X., Keysers, D., Harmsen, J.J., Houlsby, N.: Scaling vision transformers to 22 billion parameters. In: Proceedings of the 2023 International Conference on Machine Learning. vol. 202, pp. 7480-7512 (2023)", + "16. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A large-scale hierarchical image database. In: Proceedings of the 2009 IEEE Computer Society Conference on Computer Vision and Pattern Recognition Workshops (CVPR Workshops). pp. 248-255 (2009)", + "17. Doersch, C., Gupta, A., Efros, A.A.: Unsupervised visual representation learning by context prediction. In: Proceedings of the 2015 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 1422-1430 (2015)", + "18. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: Proceedings of the 2021 International Conference on Learning Representation (ICLR) (2021)", + "19. Dwibedi, D., Misra, I., Hebert, M.: Cut, paste and learn: Surprisingly easy synthesis for instance detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 1310-1319 (2017)", + "20. Everingham, M., Eslami, S.M.A., Van Gool, L., Williams, C.K.I., Winn, J., Zisserman, A.: The Pascal Visual Object Classes Challenge: A Retrospective. International Journal of Computer Vision (IJCV) 111(1), 98-136 (2015)", + "21. Gidaris, S., Singh, P., Komodakis, N.: Unsupervised representation learning by predicting image rotations. In: International Conference on Learning Representation (ICLR) (2018)", + "22. He, K., Chen, X., Xie, S., Li, Y., Dólár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16000-16009 (June 2022)", + "23. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: IEEE/CVF International Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9729-9738 (2020)", + "24. Inoue, N., Yamagata, E., Kataoka, H.: Initialization using perlin noise for training networks with a limited amount of data. In: Proceedings of the 2020 International Conference on Pattern Recognition (ICPR). pp. 1023-1028 (2020)", + "25. Johnson, J., Hariharan, B., Maaten, L.v.d., Li, F.F., Zitnick, L., Girshick, R.: CLEVR: A diagnostic dataset for compositional language and elementary visual reasoning. In: Proceedings of the 2017 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1988-1997 (2017)", + "26. Kaggle, Eyepacs: Kaggle diabetic retinopathy detection (2015)", + "27. Kataoka, H., Hara, K., Hayashi, R., Yamagata, E., Inoue, N.: Spatiotemporal initialization for 3D CNNs with generated motion patterns. In: Proceedings of the 2022 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 737-746 (2022)", + "28. Kataoka, H., Hayamizu, R., Yamada, R., Nakashima, K., Takashima, S., Zhang, X., Martinez-Noriega, E.J., Inoue, N., Yokota, R.: Replacing labeled real-image" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "R. Nakamura and R. Tadokoro et al.", + "bbox": [ + 271, + 114, + 517, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "datasets with auto-generated contours. In: Proceedings of the 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 21232-21241 (2022)", + "29. Kataoka, H., Matsumoto, A., Yamada, R., Satoh, Y., Yamagata, E., Inoue, N.: Formula-driven supervised learning with recursive tiling patterns. In: Proceedings of the 2021 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 4098-4105 (2021)", + "30. Kataoka, H., Okayasu, K., Matsumoto, A., Yamagata, E., Yamada, R., Inoue, N., Nakamura, A., Satoh, Y.: Pre-training without natural images. International Journal of Computer Vision (IJCV) 130 (2022)", + "31. Krause, J., Stark, M., Deng, J., Fei-Fei, L.: 3d object representations for fine-grained categorization. In: Proceedings of the 2013 International IEEE Workshop on 3D Representation and Recognition (3DRR-13). pp. 554-561 (2013)", + "32. Krizhevsky, A., Hinton, G.: Learning Multiple Layers of Features from Tiny Images. Master's thesis, Department of Computer Science, University of Toronto (2009)", + "33. LeCun, Y., Huang, F.J., Bottou, L.: Learning methods for generic object recognition with invariance to pose and lighting. In: Proceedings of the 2004 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). vol. 2, pp. II-104 Vol.2 (2004)", + "34. Mahajan, D., Girshick, R., Ramanathan, V., He, K., Paluri, M., Li, Y., Bharambe, A., Maaten, L.v.d.: Exploring the Limits of Weakly Supervised Pretraining. In: Proceedings of the 2018 European Conference on Computer Vision (ECCV). pp. 181-196 (2018)", + "35. Nakamura, R., Kataoka, H., Takashima, S., Noriega, E.J.M., Yokota, R., Inoue, N.: Pre-training vision transformers with very limited synthesized images. In: Proceedings of the 2023 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 20303-20312. IEEE (2023)", + "36. Nilsback, M.E., Zisserman, A.: Automated flower classification over a large number of classes. In: Proceedings of the 2008 Indian Conference on Computer Vision, Graphics and Image Processing. pp. 722-729 (2008)", + "37. Noroozi, M., Favaro, P.: Unsupervised Learning of Visual Representations by Solving Jigsaw Puzzles. In: Proceedings of the 2016 European Conference on Computer Vision (ECCV). pp. 69-84 (2016)", + "38. Noroozi, M., Vinjimoor, A., Favaro, P., Pirsiavash, H.: Boosting Self-Supervised Learning via Knowledge Transfer. In: IEEE/CVF International Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9359-9367 (2018)", + "39. Remez, T., Huang, J., Brown, M.: Learning to Segment via Cut-and-Paste. In: Proceedings of the 2018 European Conference on Computer Vision (ECCV). pp. 39-54 (2018)", + "40. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., Berg, A.C., Fei-Fei, L.: ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV) 115(3) (2015)", + "41. Sharma, P., Rott Shaham, T., Baradad, M., Fu, S., Rodriguez-Munoz, A., Duggal, S., Isola, P., Torralba, A.: A vision check-up for language models. In: arXiv preprint arXiv:2401.01862 (2024)", + "42. Sun, C., Shrivastava, A., Singh, S., Gupta, A.: Revisiting unreasonable effectiveness of data in deep learning era. In: Proceedings of the 2017 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 843-852 (2017)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Scaling Backwards", + "bbox": [ + 604, + 114, + 730, + 127 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "43. Sundermeyer, M., Marton, Z.C., Durner, M., Brucker, M., Triebel, R.: Implicit 3D orientation learning for 6D object detection from RGB images. In: Proceedings of the 2018 European Conference on Computer Vision (ECCV). pp. 712-729 (2018)", + "44. Tadokoro, R., Yamada, R., Nakashima, K., Nakamura, R., Kataoka, H.: Primitive Geometry Segment Pre-training for 3D Medical Image Segmentation. In: Proceedings of the 2023 British Machine Vision Conference (BMVC) (2023)", + "45. Takashima, S., Hayamizu, R., Inoue, N., Kataoka, H., Yokota, R.: Visual atoms: Pre-training vision transformers with sinusoidal waves. In: Proceedings of the 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 18579-18588 (2023)", + "46. Tobin, J., Fong, R., Ray, A., Schneider, J., Zaremba, W., Abbeel, P.: Domain randomization for transferring deep neural networks from simulation to the real world. In: Proceedings of the 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 23-30 (2017)", + "47. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jegou, H.: Training data-efficient image transformers & distillation through attention. In: Proceedings of the 2021 International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 139, pp. 10347-10357 (2021)", + "48. Veeling, B.S., Linmans, J., Winkens, J., Cohen, T., Welling, M.: Rotation equivariant cnns for digital pathology. In: Proceedings of the 2018 International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI) (2018)", + "49. Venkataramanan, S., Rizve, M.N., Carreira, J., Asano, Y.M., Avrithis, Y.: Is image-net worth 1 video? learning strong image encoders from 1 long unlabelled video. arXiv preprint arXiv:2310.08584 (2023)", + "50. Wang, T., Zhu, J., Torralba, A., Efros, A.A.: Dataset Distillation (2018)", + "51. Yun, S., Han, D., Chun, S., Oh, S.J., Yoo, Y., Choe, J.: CutMix: Regularization strategy to train strong classifiers with localizable features. In: Proceedings of the 2019 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 6022-6031 (2019)", + "52. Yun-Hao Cao, H.Y., Wu, J.: Training vision transformers with only 2040 images. In: Proceedings of the 2022 European Conference on Computer Vision (ECCV). pp. 220-237 (2022)", + "53. Zhai, X., Kolesnikov, A., Houlsby, N., Beyer, L.: Scaling vision transformers. Proceedings of 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 1204-1213 (2022)", + "54. Zhai, X., Puigcerver, J., Kolesnikov, A., Ruyssen, P., Riquelme, C., Lucic, M., Djolonga, J., Pinto, A.S., Neumann, M., Dosovitskiy, A., Beyer, L., Bachem, O., Tschannen, M., Michalski, M., Bousquet, O., Gelly, S., Houlsby, N.: A Large-scale Study of Representation Learning with the Visual Task Adaptation Benchmark. arXiv preprint arXiv:1910:04867 (2019)", + "55. Zhang, H., Cisse, M., Dauphin, Y.N., Lopez-Paz, D.: Mixup: Beyond empirical risk minimization. arXiv preprint abs/1710.09412 (2017)", + "56. Zhang, R., Isola, P., Efros, A.A.: Colorful image colorization. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 649-666 (2016)", + "57. Zhao, B., Mopuri, K.R., Bilen, H.: Dataset condensation with gradient matching. In: Proceedings of the 2021 International Conference on Learning Representation (ICLR) (2021)", + "58. Zhong, Z., Zheng, L., Kang, G., Li, S., Yang, Y.: Random erasing data augmentation. arXiv preprint abs/1708.04896 (2017)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "R. Nakamura and R. Tadokoro et al.", + "bbox": [ + 271, + 114, + 517, + 127 + ], + "page_idx": 17 + }, + { + "type": "ref_text", + "text": "59. Zhou, B., Lapedriza, A., Khosla, A., Oliva, A., Torralba, A.: Places: A 10 million image database for scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 40(6), 1452-1464 (2017)", + "bbox": [ + 217, + 146, + 787, + 189 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "Scaling Backwards", + "bbox": [ + 604, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/Scaling Backwards_ Minimal Synthetic Pre-training_/82902d64-0219-4fba-9a2b-20c3a32a53cf_model.json b/2024/Scaling Backwards_ Minimal Synthetic Pre-training_/82902d64-0219-4fba-9a2b-20c3a32a53cf_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9ef2ca491ebbde6646736254a8d42e74a964ffa8 --- /dev/null +++ b/2024/Scaling Backwards_ Minimal Synthetic Pre-training_/82902d64-0219-4fba-9a2b-20c3a32a53cf_model.json @@ -0,0 +1,2779 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.315, + 0.142, + 0.687, + 0.187 + ], + "angle": 0, + "content": "Scaling Backwards: Minimal Synthetic Pre-training?" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.213, + 0.776, + 0.259 + ], + "angle": 0, + "content": "Ryo Nakamura\\(^{1,*}\\), Ryu Tadokoro\\(^{2,*}\\), Ryosuke Yamada\\(^{1}\\), Yuki M. Asano\\(^{3}\\), Iro Laina\\(^{4}\\), Christian Rupprecht\\(^{4}\\), Nakamasa Inoue\\(^{5}\\), Rio Yokota\\(^{5}\\), and Hirokatsu Kataoka\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.248, + 0.27, + 0.751, + 0.285 + ], + "angle": 0, + "content": "\\(^{1}\\) National Institute of Advanced Industrial Science and Technology (AIST)" + }, + { + "type": "text", + "bbox": [ + 0.432, + 0.285, + 0.571, + 0.299 + ], + "angle": 0, + "content": "2 Tohoku University" + }, + { + "type": "text", + "bbox": [ + 0.258, + 0.299, + 0.744, + 0.313 + ], + "angle": 0, + "content": "3 University of Amsterdam, now at University of Technology Nuremberg" + }, + { + "type": "text", + "bbox": [ + 0.425, + 0.313, + 0.578, + 0.326 + ], + "angle": 0, + "content": "\\(^{4}\\) University of Oxford" + }, + { + "type": "text", + "bbox": [ + 0.394, + 0.326, + 0.609, + 0.34 + ], + "angle": 0, + "content": "5 Tokyo Institute of Technology" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.382, + 0.744, + 0.688 + ], + "angle": 0, + "content": "Abstract. Pre-training and transfer learning are an important building block of current computer vision systems. While pre-training is usually performed on large real-world image datasets, in this paper we ask whether this is truly necessary. To this end, we search for a minimal, purely synthetic pre-training dataset that allows us to achieve performance similar to the 1 million images of ImageNet-1k. We construct such a dataset from a single fractal with perturbations. With this, we contribute three main findings. (i) We show that pre-training is effective even with minimal synthetic images, with performance on par with large-scale pre-training datasets like ImageNet-1k for full fine-tuning. (ii) We investigate the single parameter with which we construct artificial categories for our dataset. We find that while the shape differences can be indistinguishable to humans, they are crucial for obtaining strong performances. (iii) Finally, we investigate the minimal requirements for successful pre-training. Surprisingly, we find that a substantial reduction of synthetic images from 1k to 1 can even lead to an increase in pre-training performance, a motivation to further investigate \"scaling backwards\". Finally, we extend our method from synthetic images to real images to see if a single real image can show similar pre-training effect through shape augmentation. We find that the use of grayscale images and affine transformations allows even real images to \"scale backwards\". The code is available at https://github.com/SUPER-TADORY/1p-frac." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.701, + 0.741, + 0.715 + ], + "angle": 0, + "content": "Keywords: Synthetic pre-training \\(\\cdot\\) Limited data \\(\\cdot\\) Vision transformers" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.744, + 0.376, + 0.76 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.77, + 0.787, + 0.815 + ], + "angle": 0, + "content": "In image recognition, pre-training allows discovering fundamental visual representations for downstream task applications. Pre-training enhances the performance of visual tasks and enables the use of small-scale task-specific datasets." + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.826, + 0.477, + 0.84 + ], + "angle": 0, + "content": "* These authors contributed equally" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.518, + 0.128 + ], + "angle": 0, + "content": "R. Nakamura and R. Tadokoro et al." + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.147, + 0.784, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.285, + 0.788, + 0.382 + ], + "angle": 0, + "content": "Fig. 1: Comparison of ImageNet-1k, FractalDB and 1p-frac (ours). 1p-frac consists of only a single fractal for pre-training. With 1p-frac, neural networks learn to classify perturbations applied to the fractal. In our study \"single\" means a very narrow distribution over parameters that leads to images that are roughly equivalent from a human visual perspective. While the shape differences of perturbed images can be indistinguishable to humans, models pre-trained on 1p-frac achieve comparable performance with those pre-trained on ImageNet-1k or FractalDB." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.397, + 0.785, + 0.456 + ], + "angle": 0, + "content": "Recently, pre-training has been used as a key technology to construct foundation models trained on massive datasets with over hundreds of millions of images. In some cases, a foundation model enables adaptation of zero-shot recognition without the need for additional data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.459, + 0.788, + 0.594 + ], + "angle": 0, + "content": "Pre-training is often interpreted as discovering universal structures in large-scale datasets that later facilitate adaptation to down-stream tasks. In this paper, we challenge this interpretation by providing a minimal pre-training dataset, generated from a single fractal, that achieves similar downstream performance. At the heart of this investigation is the question of whether pre-training might simply be a better weight initialization rather than the discovery of useful visual concepts. If true, performing expensive pre-training with hundreds of millions of images might not be necessary. This additionally frees pre-training from licensing or ethical issues." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.596, + 0.788, + 0.687 + ], + "angle": 0, + "content": "Since the rise of deep neural networks, the ImageNet dataset [16] has been one of the most commonly used pre-training datasets. Originally, pre-training has been conducted through supervised learning (SL) with human-provided teacher labels. However, it has become clear that pre-training can also be achieved without human-provided labels, through self-supervised learning (SSL) [7,9-12,17, 21-23,37,38,56]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.688, + 0.788, + 0.809 + ], + "angle": 0, + "content": "In this context, Asano et al. [3] successfully acquired visual representations while dramatically reducing the number of required images. They concluded that SSL can yield sufficient image representations even with a single training example, but only for earlier layers of a recognition model. However, it is unclear how these findings translate to modern architectures and representation learning methods. Based on this, vision transformers (ViT) [18] have been pre-trained with only 2,040 real images [52] through an instance discrimination learning signal." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.788, + 0.841 + ], + "angle": 0, + "content": "More recently, it has been shown that basic visual representations can be acquired without using real images and human-provided labels. The trend of" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.606, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Scaling Backwards" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.253 + ], + "angle": 0, + "content": "artificially generated labeled images is on the rise for synthetic pre-training [5, 28, 30, 46]. Formula-driven supervised learning (FDSL) generates images from generative formulas, and labels from their parameters [30]. Under the FDSL framework, one can adapt the synthetic pre-training image dataset by altering formulas [24, 27, 29, 44]. While a million-order image dataset was constructed in FractalDB [30], our findings suggest that synthetic pre-training can be reduced to significantly fewer fractal images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.255, + 0.785, + 0.359 + ], + "angle": 0, + "content": "Motivated by these findings, we believe it is possible to find the key essence of pre-training for image recognition. ViT training can be done with as few as 1,000 artificially generated images [35]. Here, we believe that equivalent performance can be achieved with even fewer images. This consideration is undoubtedly important as we approach a minimal synthetic pre-training dataset in image recognition, which goes against the trend of foundation models toward increasing the dataset scale." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.363, + 0.785, + 0.424 + ], + "angle": 0, + "content": "In the present paper, we thus introduce a minimal synthetic dataset, namely 1-parameter Fractal as Data (1p-frac), which consists of a single fractal as shown in Figure 1, as well as a loss function for pre-training with it. Our contributions regarding minimal synthetic pre-training are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.426, + 0.785, + 0.516 + ], + "angle": 0, + "content": "Ordinal minimalism: We introduce the locally perturbed cross entropy (LPCE) loss for pre-training with a single fractal. It utilizes perturbed fractal images for training, where neural networks learn to classify small perturbations. In experiments, we demonstrate that pre-training can be performed even with a single fractal. The pre-training effect of 1p-frac is comparable to that of a million-scale labeled image dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.519, + 0.784, + 0.638 + ], + "angle": 0, + "content": "Distributional minimalism: We introduce the locally integrated empirical (LIEP) distribution \\( p_{\\Delta} \\) that has a controllable perturbation degree \\( \\Delta \\) to investigate the minimal support of the probability density distribution of synthesized images. We observed positive pre-training effects even with a small \\( \\Delta \\) producing shape differences that cannot be distinguished by humans. We also show that if \\( \\Delta \\) is too small the visual pre-training collapses. From these observations, we establish the general bounds for generating good pre-training images from a mathematical formula." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.642, + 0.538, + 0.776 + ], + "angle": 0, + "content": "Instance minimalism: Based on the experimental results, the synthetic images should not simply contain complex shapes. The use of recursive image patterns, similar to objects in nature, should be applied in visual pretraining. Experiments with augmented categories from a real image have shown that good pre-training effects can be achieved by performing \"affine transformations on edge" + }, + { + "type": "table_caption", + "bbox": [ + 0.547, + 0.649, + 0.784, + 0.706 + ], + "angle": 0, + "content": "Table 1: Scaling backwards in synthetic pre-training (Accuracies on CIFAR-100, Real: ImageNet, Synth: Fractal images)." + }, + { + "type": "table", + "bbox": [ + 0.541, + 0.707, + 0.774, + 0.768 + ], + "angle": 0, + "content": "
Type\\#Img11k1M
RealN/A76.985.5
Synth84.284.081.6
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.778, + 0.784, + 0.809 + ], + "angle": 0, + "content": "emphasized objects in grayscale\". These operations are found to be almost synonymous with the configuration of the proposed 1p-frac." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "In summary, we significantly reduce the size of the pre-training dataset from 1M images (fractal database (FractalDB) [30]) or 1k images (one-instance fractal" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.232, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.518, + 0.129 + ], + "angle": 0, + "content": "R. Nakamura and R. Tadokoro et al." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.178 + ], + "angle": 0, + "content": "database (OFDB) [35]) down to 1 and show that this even improves the pretraining effect (Table 1), which motivates \"scaling backwards\"." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.198, + 0.389, + 0.215 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.227, + 0.785, + 0.258 + ], + "angle": 0, + "content": "The present study focuses primarily on finding the minimal pre-training dataset with artificially generated images." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.269, + 0.787, + 0.39 + ], + "angle": 0, + "content": "Pre-training in image recognition. Pre-training in image recognition has demonstrated significant success with vast amounts of labeled images. Pre-training helps neural networks to acquire fundamental visual representations which are essential for improving the performance of diverse downstream tasks in image recognition. In particular, researchers have used larger-scale image datasets for supervised learning (SL), starting from million-scale datasets (e.g., ImageNet [16], Places [59]). More recently, the largest datasets used are reaching billion-scale (e.g., IG-3.5B [34], JFT-300M/3B/4B [15, 42, 53])." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.391, + 0.787, + 0.48 + ], + "angle": 0, + "content": "Pre-training can be also achieved without any human intervention. One of the most effective approaches is self-supervised learning (SSL [7,9-12,17,21-23,37, 38,56]), which has been developed in order to alleviate the manual annotations on large-amount of images. SSL has been studied as an expected alternative to supervised pre-training and has sometimes been shown to surpass supervised pre-training." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.481, + 0.787, + 0.557 + ], + "angle": 0, + "content": "Despite the development of supervised and self-supervised pre-training, the number of real images used in visual pre-training continues to increase. Exploratory research questions such as \"what is a minimal synthetic pre-training dataset?\" have received less attention. However, they could give us more insights on what is learned during pre-training and how it could be improved." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.568, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Pre-training with limited data. In order to address ethical issues such as privacy and fairness associated with real images, a pre-training with a very limited number of real images and synthetic pre-training without any real images have been on the rise. In this context, synthetic pre-training with copy-and-paste learning [19,39], domain randomization from 3D graphics [43,46], learning from primitive patterns [4,5,41], and dataset distillation [8,50,57] have been recognized as promising approaches to acquiring visual representations. More recently, a sophisticated approach, formula-driven supervised learning (FDSL) [28,30], enabled automatic construction of a labeled image dataset without any real images or human annotations. FDSL can generate images and their corresponding labels using a mathematical formula. Along these lines, One-instance FractalDB (OFDB) has successfully pre-trained a ViT using only 1,000 synthetic images generated from a mathematical formula, without using real images [35]. Moreover, several studies have been reported in the context of extremely limited data [3,49,52]. In particular, Asano et al. have shown that learning visual representations is possible using a single image for self-supervised learning. Venkataramanan et al. achieved a comparable method to ImageNet pre-training with only 10 walking videos and their cropped images." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.605, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Scaling Backwards" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "By analyzing the acquisition of visual representations through the use of minimal synthetic pre-training datasets, we believe that the FDSL framework will be key to discovering the essence of pre-training mechanisms in image recognition." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.215, + 0.64, + 0.233 + ], + "angle": 0, + "content": "3 Scaling Backwards with a Single Fractal" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.247, + 0.788, + 0.442 + ], + "angle": 0, + "content": "This section presents the proposed methodology to explore the minimal requirements for successful pre-training by utilizing purely synthetic images of fractals. Specifically, we introduce the 1 parameter Fractal as Data (1p-frac), which consists of only a single fractal, with a method to pre-train neural networks on it. Our key idea is to introduce the locally integrated empirical (LIEP) distribution \\( p_{\\Delta} \\) over perturbed fractal images, enabling pre-training even with a single fractal image. Since the LIEP distribution is designed so that it converges to the empirical distribution \\( p_{\\mathrm{data}}(x) = \\delta (x - I) \\) of a single image \\( I \\) when the perturbation degree \\( \\Delta \\in \\mathbb{R}_{\\geq 0} \\) goes to zero, we can narrow down the support of the distribution by decreasing \\( \\Delta \\) as shown in Figure 2a. Below, we begin with a preliminary for defining the empirical distributions of two fractal databases, namely FractalDB [30] and OFDB [35]. We then introduce 1p-frac with the LIEP distribution." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.465, + 0.362, + 0.48 + ], + "angle": 0, + "content": "3.1 Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.489, + 0.788, + 0.595 + ], + "angle": 0, + "content": "FractalDB [30]. Kataoka et al. have introduced a method for effectively pretraining neural networks with FractalDB, a set of fractal images generated by the iterated function systems (IFSs). Specifically, FractalDB \\(\\mathcal{F}\\) consists of one million synthesized images: \\(\\mathcal{F} = \\{(\\varOmega_c,\\{I_i^c\\}_{i = 0}^{M - 1})\\}_{c = 0}^{C - 1}\\), where \\(\\varOmega_{c}\\) is an IFS, \\(I_{i}^{c}\\) is a fractal image generated by \\(\\varOmega_{c}\\), \\(C = 1,000\\) is the number of fractal categories, and \\(M = 1,000\\) is the number of images per category. Each IFS characterizes each fractal category \\(c\\) and is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.349, + 0.607, + 0.785, + 0.623 + ], + "angle": 0, + "content": "\\[\n\\Omega_ {c} = \\left\\{\\mathbb {R} ^ {2}; w _ {1}, w _ {2}, \\dots , w _ {N _ {c}}; p _ {1}, p _ {2}, \\dots , p _ {N _ {c}} \\right\\}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.634, + 0.641, + 0.65 + ], + "angle": 0, + "content": "where \\(w_{j}:\\mathbb{R}^{2}\\to \\mathbb{R}^{2}\\) is a 2D affine transformation given by" + }, + { + "type": "equation", + "bbox": [ + 0.373, + 0.661, + 0.785, + 0.694 + ], + "angle": 0, + "content": "\\[\nw _ {j} (\\boldsymbol {v}) = \\left[ \\begin{array}{l l} a _ {j} & b _ {j} \\\\ c _ {j} & d _ {j} \\end{array} \\right] \\boldsymbol {v} + \\left[ \\begin{array}{l} e _ {j} \\\\ f _ {j} \\end{array} \\right] (\\boldsymbol {v} \\in \\mathbb {R} ^ {2}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.703, + 0.788, + 0.808 + ], + "angle": 0, + "content": "and \\( p_j \\) is a probability mass distribution. Each fractal image \\( I_i^c \\) renders a fractal \\( F = \\{\\pmb{v}_t\\}_{t=0}^T \\subset \\mathbb{R}^2 \\) into a 2D image, where points \\( \\pmb{v}_t \\) are determined by a recurrence relation that applies affine transformations as \\( \\pmb{v}_{t+1} = w_{\\sigma_t}(\\pmb{v}_t) \\) for \\( t = 0, 1, 2, \\dots, T-1 \\). Here, the initial point is set as \\( \\pmb{v}_0 = (0,0)^\\top \\), and the index \\( \\sigma_t \\) is sampled at each \\( t \\) independently from the probability mass distribution \\( p(\\sigma_t = j) = p_j \\). Pre-training with FractalDB utilizes the cross-entropy loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.4, + 0.821, + 0.785, + 0.837 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = - \\mathbb {E} _ {x, y \\sim p _ {\\mathrm {d a t a}}} [ \\log p _ {\\theta} (y | x) ], \\tag {3}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.518, + 0.128 + ], + "angle": 0, + "content": "R. Nakamura and R. Tadokoro et al." + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.145, + 0.768, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.321, + 0.789, + 0.391 + ], + "angle": 0, + "content": "Fig. 2: Scaling backwards from many images to a single synthetic image. (a) Empirical distribution \\( p_{\\mathrm{data}} \\). Colors indicate classes. With a single image, the distribution is given by a single Dirac's delta function. (b) LIEP distribution \\( p_{\\Delta} \\). The support of the distribution narrows as the degree of perturbation \\( \\Delta \\) decreases. (c) \\( \\sigma \\)-factor for investigating fractal shapes. A small \\( \\sigma \\) produces complex fractals." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.4, + 0.789, + 0.444 + ], + "angle": 0, + "content": "where \\( p_{\\theta} \\) is the category distribution predicted by a neural network and \\( \\theta \\) is a set of learnable parameters. The joint empirical distribution \\( p_{\\mathrm{data}} \\) is defined over the dataset as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.33, + 0.451, + 0.787, + 0.492 + ], + "angle": 0, + "content": "\\[\np _ {\\mathrm {d a t a}} (x, y; \\mathcal {F}) = \\frac {1}{M C} \\sum_ {i = 0} ^ {M - 1} \\sum_ {c = 0} ^ {C - 1} \\delta \\left(x - I _ {i} ^ {c}\\right) \\delta (y - c) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.5, + 0.789, + 0.544 + ], + "angle": 0, + "content": "where \\(\\delta\\) is the Dirac's delta function. Models pre-trained on this dataset perform comparable to those pre-trained on real-world image datasets, such as ImageNet-1k and Places365, in some downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.545, + 0.79, + 0.59 + ], + "angle": 0, + "content": "OFDB [35]. This dataset consists of 1,000 fractal images. Specifically, OFDB \\(\\mathcal{F}_{\\mathrm{OF}}\\) involves only one representative image per category, i.e., \\(\\mathcal{F}_{\\mathrm{OF}} = \\{\\varOmega_c,I_c\\}_{c = 0}^{C - 1}\\). Therefore, the joint empirical distribution reduces to" + }, + { + "type": "equation", + "bbox": [ + 0.344, + 0.599, + 0.787, + 0.64 + ], + "angle": 0, + "content": "\\[\np _ {\\mathrm {d a t a}} (x, y; \\mathcal {F} _ {\\mathrm {O F}}) = \\frac {1}{C} \\sum_ {c = 0} ^ {C - 1} \\delta \\left(x - I _ {c}\\right) \\delta \\left(y - c\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.647, + 0.789, + 0.709 + ], + "angle": 0, + "content": "Models pre-trained on this dataset perform comparable to or even better than those pre-trained on FractalDB. This work has shown that there exists a small but essential set of images for visual pre-training. However, reducing the number of fractals \\( C \\) to less than 1,000 degrades the performance." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.728, + 0.542, + 0.743 + ], + "angle": 0, + "content": "3.2 Pre-training with a Single Fractal" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.751, + 0.789, + 0.812 + ], + "angle": 0, + "content": "Scaling backwards. To further facilitate the analysis of what images are minimally required for successful visual pre-training, we introduce 1p-frac, which ultimately reduces the number of IFSs and images to one each as \\(\\mathcal{F}_{\\mathrm{OP}} = (\\varOmega,I)\\). With this dataset, the empirical distribution is given by" + }, + { + "type": "equation", + "bbox": [ + 0.384, + 0.82, + 0.787, + 0.838 + ], + "angle": 0, + "content": "\\[\np _ {\\mathrm {d a t a}} (x, y; \\mathcal {F} _ {\\mathrm {O P}}) = \\delta (x - I) \\delta (y). \\tag {6}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.605, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Scaling Backwards" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.221 + ], + "angle": 0, + "content": "However, we notice that with this distribution, training neural networks using cross-entropy loss is not straightforward because \\( p_{\\theta}(y = 0|x) \\equiv 1 (\\forall x) \\) gives a trivial solution to the loss minimization problem. To address this, we introduce locally perturbed cross entropy (LPCE) loss \\( \\mathcal{L}_{\\Delta} \\), a variant of the cross entropy loss defined with the LIEP distribution." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.223, + 0.784, + 0.268 + ], + "angle": 0, + "content": "Definition 1. Let \\( I_{\\epsilon} \\in \\mathcal{X} \\) be a perturbed image, where \\( \\mathcal{X} \\) is a set of images, \\( \\epsilon \\in \\mathbb{R}^d \\) is a small perturbation with \\( d \\in \\mathbb{N}_{>0} \\) and \\( I_0 = I \\) is the original image. We define the LIEP distribution by" + }, + { + "type": "equation", + "bbox": [ + 0.355, + 0.279, + 0.784, + 0.31 + ], + "angle": 0, + "content": "\\[\np _ {\\Delta} (x, y) = \\frac {1}{| \\mathcal {R} _ {\\Delta} |} \\int_ {\\mathcal {R} _ {\\Delta}} \\delta (x - I _ {\\epsilon}) \\delta (y - \\epsilon) d \\epsilon \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.321, + 0.785, + 0.351 + ], + "angle": 0, + "content": "where \\(\\mathcal{R}_{\\Delta} \\subset \\mathbb{R}^{d}\\) is a compact set containing the origin and \\(|\\mathcal{R}_{\\Delta}|\\) is its volume with order \\(O(|\\Delta|^{d})\\)." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.351, + 0.538, + 0.366 + ], + "angle": 0, + "content": "Definition 2. We define the LPCE loss by" + }, + { + "type": "equation", + "bbox": [ + 0.399, + 0.379, + 0.784, + 0.394 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\Delta} = - \\mathbb {E} _ {x, y \\sim p _ {\\Delta}} \\left[ \\log p _ {\\theta} (y | x) \\right], \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.404, + 0.467, + 0.418 + ], + "angle": 0, + "content": "where \\(p_{\\Delta}\\) is the LIEP distribution." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.419, + 0.784, + 0.478 + ], + "angle": 0, + "content": "If \\(\\mathcal{R}_{\\Delta}\\) is a small hypercube or hypersphere, it is straightforward to see that \\(p_{\\Delta}\\) approaches to the empirical distribution of Eq. (6) when \\(\\Delta\\) goes to zero. Therefore, this loss allows us to analyze visual pre-training effects by narrowing the support of the distribution around a single image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.48, + 0.784, + 0.523 + ], + "angle": 0, + "content": "With 1p-frac, we apply perturbation to the affine transformations. As such, the perturbation \\(\\epsilon\\) is in \\(\\mathbb{R}^{6*j}\\). A perturbed image \\(I_{\\epsilon}\\) is obtained by the noisy affine transformations:" + }, + { + "type": "equation", + "bbox": [ + 0.379, + 0.533, + 0.784, + 0.565 + ], + "angle": 0, + "content": "\\[\nw _ {j} \\left(\\boldsymbol {v}; \\boldsymbol {\\epsilon} _ {j}\\right) = \\left(\\left[ \\begin{array}{l l l} a _ {j} & b _ {j} & e _ {j} \\\\ c _ {j} & d _ {j} & f _ {j} \\end{array} \\right] + \\boldsymbol {\\epsilon} _ {j}\\right) \\left[ \\begin{array}{l} \\boldsymbol {v} \\\\ 1 \\end{array} \\right] \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.575, + 0.784, + 0.635 + ], + "angle": 0, + "content": "where \\(\\epsilon_{j} \\in \\mathcal{R}_{\\Delta} = [-\\Delta / 2, \\Delta / 2]^{6 * j} \\subset \\mathbb{R}^{6 * j}\\) is a hypercube with a side length \\(\\Delta\\), and \\(|\\mathcal{R}_{\\Delta}| = \\Delta^{6 * j}\\). Note that the numerical integration of Eq. (7) is used in practice, which is an approximation obtained by uniformly sampling \\(L\\) points in \\(\\mathcal{R}_{\\Delta}\\), where we set \\(L = 1,000\\) by default." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.636, + 0.784, + 0.695 + ], + "angle": 0, + "content": "Visualization. Figure 2b shows examples of perturbed images to compute the LPCE loss. While most of shape differences are indistinguishable to humans, a neural network learns to distinguish the perturbations applied to the single image by minimizing the LPCE loss." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.697, + 0.784, + 0.741 + ], + "angle": 0, + "content": "Complexity of \\(\\Omega\\). We use the \\(\\sigma\\)-factor proposed by Anderson et al. [1] to evaluate the complexity of IFSs. As shown in Figure 2c, small values of \\(\\sigma\\) produce complex fractal shapes." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.764, + 0.375, + 0.781 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Experimental setup. What is minimal visual pre-training, and what properties of images lend themselves to reducing their count? In order to clarify these questions, we conduct experiments following these steps:" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.518, + 0.128 + ], + "angle": 0, + "content": "R. Nakamura and R. Tadokoro et al." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.147, + 0.784, + 0.205 + ], + "angle": 0, + "content": "- Exploration study (Tables 2-4): We verify whether a single fractal is sufficient for pre-training. Moreover, we compare the extent to which the shapes used in pre-training can be simplified, using object contours, Gaussian, and uniform distributions." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.207, + 0.785, + 0.236 + ], + "angle": 0, + "content": "- Hyperparameter study (Tables 5-7): We investigate the effects of three hyperparameters \\(\\Delta\\), \\(\\sigma\\) and \\(L\\)." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.238, + 0.785, + 0.266 + ], + "angle": 0, + "content": "- Scaling study (Table 8): We compare 1p-frac with other large-scale datasets in terms of fine-tuning accuracy on ImageNet-1k." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.269, + 0.785, + 0.311 + ], + "angle": 0, + "content": "- Analysis and discussion (Tables 9-12): We discuss data augmentation, computational cost for synthesizing images, and pre-training using a single real-world image." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.313, + 0.784, + 0.342 + ], + "angle": 0, + "content": "- Applications (Tables 13 and 14): We conduct comparisons in various fine-tuning datasets." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.353, + 0.786, + 0.52 + ], + "angle": 0, + "content": "Implementation details. In order to verify the effects of pre-training, we measure the accuracy of downstream tasks through fine-tuning. We use ViT [18] for all experiments. Specifically, we employ ViT-Tiny (ViT-T) in the exploration and hyperparameter studies, and employ ViT-Base (ViT-B) for the scaling study. The pre-training parameters and data augmentation methods for pre-training and fine-tuning follow conventional methods in OFDB. In OFDB, data augmentation is based on DeiT. For exploration studies, CIFAR-100 (C100) [32] and/or ImageNet-100 (IN100) [40] are used as fine-tuning datasets. In the comparisons, we assign seven representative datasets used in OFDB paper, namely C10/C100, Cars [31], Flowers [36], Pascal VOC 2012 (VOC12) [20], Places-30 (P30) [59], IN100. See supplementary material for more detailed experimental settings." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.543, + 0.416, + 0.558 + ], + "angle": 0, + "content": "4.1 Exploration Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.784, + 0.641 + ], + "angle": 0, + "content": "Pre-training with 1p-frac. Table 2 compares 1p-frac with FracdalDB (1M images, 1k categories) and OFDB (1k images, 1k categories). As can be seen, the pre-training effect of 1p-frac is comparable to that of FractalDB and OFDB. The number of fractal categories is significantly reduced from \\( C = 1,000 \\) to only 1. This shows the effectiveness and efficiency of 1p-frac and the LPCE loss." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.785, + 0.761 + ], + "angle": 0, + "content": "Comparison with SOTA FDSL datasets. Table 3 compares 1p-frac with two SOTA FDSL datasets: (1) Radial Contour Database (RCDB) [28], which consists of 1 million images of synthetic polygons, and (2) Visual Atoms (VA) [45], which consists of 1 million images of parameterized wave functions. The \"1K\" counterparts in the table reduce the number of images per category to one. We also applied the LPCE loss to these two datasets by applying perturbations to the radios of a polygon or wave. As can be seen, 1p-frac performs the best when a single image is used for pre-training and outperforms the 1K FDSL datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.784, + 0.838 + ], + "angle": 0, + "content": "Pre-training with a noisy image. To validate the necessity of a complex shape when pre-training with a single image, we applied the LPCE loss to two noisy images: an image of Gaussian noise and an image of uniform noise. Since these two noise images are generated by determining the parameters of the parametric distributions, Gaussian and uniform, the LPCE loss can be computed by" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.606, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Scaling Backwards" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.218, + 0.156, + 0.406, + 0.239 + ], + "angle": 0, + "content": "Table 2: Comparison of 1p-frac with FractalDB [30] and OFDB [35]. \\(\\diamond\\) indicates the use of LPCE loss. We report top-1 accuracies \\((\\%)\\)." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.241, + 0.407, + 0.317 + ], + "angle": 0, + "content": "
Method#ImgC100IN100
Scratch-64.274.9
FractalDB1M81.688.5
OFDB1k84.088.6
1p-frac°184.289.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.413, + 0.145, + 0.599, + 0.199 + ], + "angle": 0, + "content": "Table 3: Comparison with SOTA FDSL datasets. \\(\\diamond\\) indicates the use of LPCE loss." + }, + { + "type": "table", + "bbox": [ + 0.418, + 0.203, + 0.596, + 0.328 + ], + "angle": 0, + "content": "
Method#ImgC100IN100
RCDB1M81.688.5
RCDB1K80.487.5
RCDB°182.587.9
VA1M84.990.3
VA1K82.188.5
VA°182.688.0
ip-frac°184.289.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.606, + 0.152, + 0.781, + 0.194 + ], + "angle": 0, + "content": "Table 4: Comparison with pre-training with a single noise image." + }, + { + "type": "table", + "bbox": [ + 0.613, + 0.196, + 0.772, + 0.323 + ], + "angle": 0, + "content": "
MethodC100IN100
Gaussian1.15.7
Uniform2.071.1
1p-frac84.289.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.219, + 0.333, + 0.588, + 0.362 + ], + "angle": 0, + "content": "Table 5: Effects of pertur- Table 6: Effects of \\(\\sigma\\) - bation degree \\(\\varDelta\\) \\((\\sigma=3.5)\\) \\((\\varDelta=0.1)\\)" + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.367, + 0.402, + 0.476 + ], + "angle": 0, + "content": "
ΔC100IN100
0.0011.21.9
0.0119.961.8
0.0583.088.2
0.184.289.0
0.283.488.5
1.082.688.1
" + }, + { + "type": "table", + "bbox": [ + 0.416, + 0.374, + 0.572, + 0.475 + ], + "angle": 0, + "content": "
σC100IN100
6.081.386.3
5.082.287.1
4.581.986.9
3.584.289.0
4.082.887.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.591, + 0.334, + 0.778, + 0.376 + ], + "angle": 0, + "content": "Table 7: Effects of number of sampling points \\( L \\) for numerical integration." + }, + { + "type": "table", + "bbox": [ + 0.599, + 0.389, + 0.767, + 0.474 + ], + "angle": 0, + "content": "
LC100IN100
1678.785.3
6482.388.0
51282.488.0
100084.289.0
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.494, + 0.788, + 0.586 + ], + "angle": 0, + "content": "giving small perturbations to the parameters. Table 4 shows these images and fine-tuning results. For example, here the \\(\\epsilon\\) is added to the mean vector and covariance matrix of the Gaussian distribution. The results revealed that noise images fail in visual pre-training, performing worse than training from scratch. This suggests that acquiring fundamental visual representations through pretraining requires an image of a certain structured object, such as a fractal." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.612, + 0.454, + 0.628 + ], + "angle": 0, + "content": "4.2 Hyperparameter Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.642, + 0.788, + 0.733 + ], + "angle": 0, + "content": "Perturbation degree. Table 5 shows the results obtained by different perturbation degrees \\(\\Delta\\) for the LIEP distribution. The results indicate that setting the value to 0.1 yields the highest performance. Interestingly, the performance significantly improved when \\(\\Delta\\) was increased from 0.01 to 0.05. This suggests that the support of the empirical distribution must have a certain size to obtain positive pre-training effects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.809 + ], + "angle": 0, + "content": "\\(\\sigma\\)-factor. Table 6 investigates the effects of the \\(\\sigma\\)-factor, which measures the complexity of fractal shapes. As can be seen, the most complex shape yields the highest performance. It is worth noting that, even with a fractal of \\(\\sigma = 6.0\\), which looks like Gaussian noise, we observed positive pre-training effects. This suggests that the shapes obtained from IFS are crucial for pre-training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Numerical integration. Tables 7 reduces the number of sampling points \\( L \\) of the numerical integration when approximately computing loss with Eq. (7)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.518, + 0.129 + ], + "angle": 0, + "content": "R. Nakamura and R. Tadokoro et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.218, + 0.148, + 0.53, + 0.219 + ], + "angle": 0, + "content": "Table 8: Comparison of pre-training datasets on ImageNet-1k fine-tuning. Fine-tuning accuracies calculated by using ViT-Base (B) are listed in the table. 21k indicates the number of categories in each pre-training dataset." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.221, + 0.526, + 0.376 + ], + "angle": 0, + "content": "
Pre-training#ImgTypeViT-B
Scratch--79.8
ImageNet-21k14MSL81.8
FractalDB-21k21MFDSL81.8
ExFractalDB-21k21MFDSL82.7
RCDB-21k21MFDSL82.4
VA-21k21MFDSL82.7
OFDB-21k21kFDSL82.2
3D-OFDB-21k21kFDSL82.7
1p-frac (ours)1FDSL82.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.544, + 0.145, + 0.778, + 0.199 + ], + "angle": 0, + "content": "Table 9: Ablation study on data augmentation methods. The DeiT augmentation [47] is used as a baseline." + }, + { + "type": "table", + "bbox": [ + 0.546, + 0.201, + 0.778, + 0.375 + ], + "angle": 0, + "content": "
MethodC100
Baseline84.2
w/o Random Aug. [14]83.4
w/o Random Crop80.1
w/o Rand Aspect83.6
w/o Rand Erasing [58]84.0
w/o Mixup [55]83.6
w/o Cutmix [51]83.3
w/o Mixup/Cutmix80.4
w/o Flipping84.3
w/o Color Jittering84.2
" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.395, + 0.787, + 0.427 + ], + "angle": 0, + "content": "We see that larger numbers result in high performance. Interestingly, even with \\( L = 16 \\), the pre-training effect was positive (better than training from scratch)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.448, + 0.377, + 0.465 + ], + "angle": 0, + "content": "4.3 Scaling Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.474, + 0.789, + 0.626 + ], + "angle": 0, + "content": "ImageNet-1k fine-tuning (Table 8). While this work discusses minimal requirements for visual pre-training, increasing the approximation precision of the numerical integration could further improve the performance and benefit training of large models. In Table 8, we applied the LPCE loss with \\( L = 21,000 \\) to the ViT-B model and compare fine-tuning accuracy on ImageNet-1k. Surprisingly, and despite using a single fractal as data, our ViT-B pre-trained on 1p-frac outperforms pre-training on ImageNet-21k. This shows the strength of our approach in particular because we simply keep the same fine-tuning protocols as used for the original Imagenet-21k to ImageNet-1k transfer proposed in [18], putting us at a potential disadvantage." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.649, + 0.461, + 0.664 + ], + "angle": 0, + "content": "4.4 Analysis and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Relationship with data augmentation (Table 9). We consider that the variation in image patterns also depends on data augmentation. While traditional methods primarily rely on the data augmentation techniques described in the DeiT paper [47], we investigated which data augmentation methods contribute significantly by excluding one or two techniques from the base data augmentation methods and observing the changes in fine-tuning accuracy on C100 dataset. The experimental results in Table 9 reveal that excluding Random Cropping (RandomCrop) significantly deteriorates accuracy, suggesting that omitting parts of the image region and causing partial loss facilitates the acquisition of the ability to focus on any part without relying on grasping the overall shape, greatly reducing the pre-training effect. Furthermore, when both Mixup and Cutmix are" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.606, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Scaling Backwards" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.218, + 0.16, + 0.45, + 0.284 + ], + "angle": 0, + "content": "Table 10: Dataset construction time (hours). FractalDB (1M images), OFDB (1k images) and 1p-frac (1k perturbations for one image) are compared. The processing time is shown separately for fractal category search (Search), image rendering (Render), and total time (Total)." + }, + { + "type": "table", + "bbox": [ + 0.224, + 0.292, + 0.449, + 0.362 + ], + "angle": 0, + "content": "
DatasetSearchRenderTotal
FractalDB2.3716.8619.23
OFDB2.370.412.78
1p-frac0.00220.0360.0382
" + }, + { + "type": "image", + "bbox": [ + 0.471, + 0.149, + 0.764, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.46, + 0.297, + 0.778, + 0.366 + ], + "angle": 0, + "content": "Fig. 3: Linear probing following the experiments from [3]. We verified the performance with \\(\\{1,2,3,4,6\\}\\) layers in the ViT-Tiny model on C100 dataset. Here, 2 and 4 show the use of up to 1-2 and 1-4 layers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.382, + 0.784, + 0.412 + ], + "angle": 0, + "content": "excluded, a degradation in accuracy is observed, indicating that mixing images and categories enables better feature recognition." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.413, + 0.785, + 0.52 + ], + "angle": 0, + "content": "Dataset construction times (Table 10). We calculated the processing times in fractal category search and image rendering. By comparing to the conventional approaches, our 1p-frac recorded much faster category search (0.0022 hours nearly equals to 8 seconds) and image rendering (0.036 hours nearly equals to 129 seconds). 1p-frac requires only one category with a parameter set. This very simple procedure leads to efficiently construct a minimum requirement for pre-training in image recognition." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.521, + 0.785, + 0.643 + ], + "angle": 0, + "content": "Linear probing (Figure 3). To compare supervised ImageNet-1k pre-training with our 1p-frac pre-training in the context of early layer visual representations, we executed a linear probing experiment on the C100 dataset. In the ViT-T model with \\(\\{1,2,3\\}\\) layers, the 1p-frac pre-trained model outperformed the model pre-trained on ImageNet-1k. This demonstrates the superior quality of early, low-level representations learned from our dataset compared to ImageNet. In effect, layers up to layer 3 in ViT-T can be sufficiently pre-trained and frozen using our 1p-frac." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.644, + 0.784, + 0.659 + ], + "angle": 0, + "content": "Pre-training with a single real image and shape augmentation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.66, + 0.431, + 0.78 + ], + "angle": 0, + "content": "It was found that pretraining effects comparable to a dataset of one million real-world images can be achieved with 1p-frac, but is it possible to perform pre-training similarly with a real image? In this section, we verify whether" + }, + { + "type": "image", + "bbox": [ + 0.447, + 0.661, + 0.528, + 0.74 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.662, + 0.614, + 0.738 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.662, + 0.699, + 0.738 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.662, + 0.785, + 0.738 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.442, + 0.75, + 0.785, + 0.778 + ], + "angle": 0, + "content": "Fig. 4: Shape augmentation with three geometry transformations from a single real image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.785, + 0.841 + ], + "angle": 0, + "content": "visual features can be acquired through pre-training using a single real image (each image in Figure 5) with the LPCE loss using image representations similar to FDSL (Canny [6] images) and geometric transformations for perturbation (\\{affine, elastic, polynomial\\} transformations in Figure 4). The results of pre" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.518, + 0.128 + ], + "angle": 0, + "content": "R. Nakamura and R. Tadokoro et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.208, + 0.785, + 0.237 + ], + "angle": 0, + "content": "Fig. 5: The five different images used in the pre-training with a single real image and shape augmentation." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.244, + 0.495, + 0.313 + ], + "angle": 0, + "content": "Table 11: Comparison of pre-training effect between RGB and Canny images when shape augmentation is applied. The shape augmentation uses affine transformations." + }, + { + "type": "table", + "bbox": [ + 0.229, + 0.326, + 0.476, + 0.386 + ], + "angle": 0, + "content": "
ImageMean Accuracy
RGB image81.6
Canny image82.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.509, + 0.246, + 0.788, + 0.304 + ], + "angle": 0, + "content": "Table 12: Comparison of pre-training effects of affine, elastic, and polynomial image transformations. Real Img type with Canny applied to the images." + }, + { + "type": "table", + "bbox": [ + 0.514, + 0.31, + 0.779, + 0.384 + ], + "angle": 0, + "content": "
TransformationMean Accuracy
Affine trans.81.9
Elastic trans.37.0
Polynomial trans.81.3
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.387, + 0.788, + 0.493 + ], + "angle": 0, + "content": "training effects with \\( L = 1,000 \\) are shown in Tables 11 and 12. Note that the results in these tables represent the average score on C100 over the five real images in Figure 5. See the supplementary material for results for each of image. According to the results, we find that contour-emphasized Canny images omitting color information and affine transformation perturbation yield better pre-training, as measured by the fine-tuning accuracies. The operations are similar to the classification of labeled fractal images via IFS." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.514, + 0.367, + 0.53 + ], + "angle": 0, + "content": "4.5 Applications" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.788, + 0.809 + ], + "angle": 0, + "content": "Fine-tuning on commonly used datasets (Table 13). We compare fine-tuning accuracies in three types of dataset configurations. For supervised learning (SL) we compare pretraining on ImageNet-1k [16] and Places-365 [59], for self-supervised learning (SSL) we compare DINO [7] and masked auto-encoder (MAE) [22] pre-trained on ImageNet and PASS, and models trained on the fractal datasets FDSL on FractalDB-1k and RCDB-1k. Following the setting of one-instance per category [35], we also evaluate using random patch augmentation (\"w/ Aug.\") We compare our 1p-frac with 1-image SSL [3] with DINO and MAE (we use Real Img #1 of Figure 5). The results show that our 1p-frac pretraining with one single image demonstrated high fine-tuning accuracies (87.5% and 88.2% with and without random patch augmentation) which are close to the performance of self-supervised ImageNet-1k pre-training with MAE that utilizes 1.28M images and yields 88.5%. Surprisingly, despite using only one fractal image with perturbations, our 1p-frac without random patch augmentation has demonstrated fine-tuning accuracies that are equal to or even higher than that of conventional OFDB-1k/FractalDB-1k, which uses 1,000 parameter sets with fractal categories and 1,000/1,000,000 images: 1p-frac obtains mean performances of 87.5% vs OFDB-1k 87.2% vs FractalDB-1k 87.1%." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Fine-tuning on specialized and structured datasets in VTAB [54] (Table 14). We also verify performance on a couple of specialized and structured" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.606, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Scaling Backwards" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.228 + ], + "angle": 0, + "content": "Table 13: Comparison among pre-training methods in fine-tuning accuracies. Best values at each dataset scale are in bold. ViT-T is used for all experiments. #OrgImg indicates number of collected images in real-image datasets or number of original images used in synthetic datasets. Type shows supervised types within supervised learning using cross-entropy loss (SL), self-supervised learning using DINO [7] or MAE [22] (SSL:D/SSL:M), and formula-supervised learning (FDSL)." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.234, + 0.79, + 0.501 + ], + "angle": 0, + "content": "
Pre-training#OrgImgTypeC10C100CarsFlowersVOC12P30IN100Mean
Scratch--80.362.113.771.456.276.274.862.1
Places-365 [59]1.80MSL97.683.989.299.384.6-89.4-
ImageNet-1k [16]1.28MSL98.085.589.999.488.780.0--
ImageNet-1k [16]1.28MSSL:D97.782.488.098.574.778.489.086.9
ImageNet-1k [16]1.28MSSL:M97.485.886.696.383.380.290.088.5
PASS [2]1.43MSSL:D97.584.086.498.682.979.082.987.8
FractalDB-1k [30]1.00MFDSL96.881.686.098.380.678.488.387.1
RCDB-1k [28]1.00MFDSL97.082.286.598.980.979.788.587.6
ImageNet-1k [35]1,000SL94.376.957.394.873.878.284.379.9
ImageNet-1k [35]1,000SSL:D94.978.071.294.675.578.684.982.5
OFDB-1k [35]1,000FDSL96.984.084.597.179.979.988.087.2
3D-OFDB-1k [35]1,000FDSL97.183.885.598.480.880.089.187.8
OFDB-1k w/ Aug. [35]1,000FDSL97.285.387.698.381.480.489.588.5
3D-OFDB-1k w/ Aug. [35]1,000FDSL97.084.785.698.381.279.888.987.9
1-image SSL [3]1SSL:D95.779.573.193.869.080.488.782.8
1-image SSL [3]1SSL:M97.284.782.194.678.580.389.086.6
1p-frac (ours)1FDSL96.984.284.597.480.580.689.087.5
1p-frac w/ Aug. (ours)1FDSL96.584.787.098.180.980.588.988.2
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.512, + 0.788, + 0.648 + ], + "angle": 0, + "content": "datasets in VTAB (visual task adaptation benchmark; Retinopathy [26], Resisc45 [13], Camelyon [48], CLEVR-Count [25], and sNORB-Azim [33]). The experiment is more suitable to evaluate and compare the pre-training methods since the image domains and their labels are distinct from both real and fractal images on ImageNet and 1p-frac. According to Table 14, our 1p-frac pretrained model performed relatively similar fine-tuning accuracies on all listed datasets in the table. By comparing to self-supervised ImageNet-1k with MAE, our 1p-frac surpassed the fine-tuning accuracies on three (Retinopathy, Resisc45, and sNORB-Azim) out of the five datasets." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.672, + 0.509, + 0.687 + ], + "angle": 0, + "content": "5 Conclusion and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.787, + 0.764 + ], + "angle": 0, + "content": "This paper examined what a minimal dataset for synthetic pre-training might look like. We proposed the 1-parameter Fractal as Data (1p-frac) dataset that succeeds in pre-training even with a single fractal by utilizing the locally perturbed cross entropy (LPCE) loss. The following findings were presented:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.788, + 0.841 + ], + "angle": 0, + "content": "What is a minimal synthetic pre-training dataset? In this paper, we applied IFS to generate labeled images. In this context, preparing a single set of parameters \\(a - f\\) in IFS is the least informative. It recorded fine-tuning accuracy nearly equivalent to conventional methods such as FractalDB and OFDB (Tables 1 and 2), succeeding in pre-training with minimal image pattern utilization." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.518, + 0.129 + ], + "angle": 0, + "content": "R. Nakamura and R. Tadokoro et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.2 + ], + "angle": 0, + "content": "Table 14: Performance comparisons on five fine-tuning datasets: Retinopathy (Retino) [26], Resisc45 [13], Camelyon [48], CLEVR-Count (CLEVR-C) [25], and sNORB-Azim (sNORB-A) [33]. Best and second-best values are in bold and underlined, respectively." + }, + { + "type": "table", + "bbox": [ + 0.23, + 0.201, + 0.776, + 0.283 + ], + "angle": 0, + "content": "
#OrgImgTypeRetinoResisc45CamelyonCLEVR-CsNORB-A
Scratch--66.989.274.446.411.5
ImageNet-1k [16]1.28MSL79.197.082.789.329.2
ImageNet-1k [16]1.28MSSL:M76.795.683.989.526.3
1p-frac (ours)1FDSL77.995.882.786.228.9
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.296, + 0.784, + 0.356 + ], + "angle": 0, + "content": "tion. Moreover, it is not that any perturbation in image space is acceptable, as Gaussian/uniform distributions failed in pre-training (Table 4). It clarified the importance of geometric transformations with certain regularities like fractal geometry or contour shapes (Table 3)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.356, + 0.785, + 0.476 + ], + "angle": 0, + "content": "What properties of images lead to image reduction? In 1p-frac, we investigated shape changes and pre-training effects by adopting the LPCE loss and the \\(\\sigma\\)-factor that contribute to fractal shape variation. It was revealed that pre-training efficacy varies with shape variation (Tables 5, 6). For real images, it is not just about including more edge components, but about including affine transformations, which closely resembles the classification of labeled fractal images through IFS (Tables 11 and 12). By extensively abating data augmentation showed, we demonstrated the essential role of random cropping and MixUp." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.478, + 0.787, + 0.627 + ], + "angle": 0, + "content": "Pre-training effects with the proposed method. Supervised pre-training on ImageNet was more accurate for datasets where object labels are assigned to real images (Table 13), but in VTAB's Specialized/Structured datasets (Table 14), performances were relatively close. On the other hand, our 1p-frac pre-trained ViT performed a similar mean accuracy by comparing to MAE self-supervised ImageNet pre-training (MAE 88.5 vs. our 1p-frac w/ Aug. 88.2). Furthermore, increasing the number of sampling points \\( L \\) showed performance improvement even with 21k categories and demonstrated good performance in ImageNet-1k fine-tuning (Table 8), indicating the potential for building large models with limited data resources." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.787, + 0.718 + ], + "angle": 0, + "content": "Broader impact. Our proposed dataset 1p-frac does not suffer from licensing or ethical issues such as biases and has a large potential to serve as a clean pretraining dataset. While not investigated in this paper, this approach also opens the possibility for significant gains in training speed by keeping the data on the GPU and applying transformations there, removing the GPU-CPU transfer bottleneck." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Limitation. In this paper, we have explored achieving a minimal synthetic pre-training in image recognition based on fractal geometry with IFS (a single fractal, and their perturbations). Despite this, we believe it is also necessary to find a similarly minimal pre-training dataset containing real images - for some applications and increased interpretability. This might allow for quick and clear calibration of pre-training, similar to how calibration images are used in photography. We did not investigate this direction in this paper but leave this for future work." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.606, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Scaling Backwards" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.18, + 0.784, + 0.224 + ], + "angle": 0, + "content": "Computational resource of AI Bridging Cloud Infrastructure (ABCI) provided by National Institute of Advanced Industrial Science and Technology (AIST) was used." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.243, + 0.323, + 0.259 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.278, + 0.785, + 0.319 + ], + "angle": 0, + "content": "1. Anderson, C., Farrell, R.: Improving fractal pre-training. In: Proceedings of the 2022 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 1300-1309 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.322, + 0.785, + 0.361 + ], + "angle": 0, + "content": "2. Asano, Y., Rupprecht, C., Zisserman, A., Vedaldi, A.: Pass: An imagenet replacement for self-Supervised pretraining without humans. In: Proceedings of the 2021 Neural Information Processing Systems Track on Datasets and Benchmarks (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.364, + 0.785, + 0.403 + ], + "angle": 0, + "content": "3. Asano, Y.M., Rupprecht, C., Vedaldi, A.: A critical analysis of self-supervision, or what we can learn from a single image. In: Proceedings of the 2020 International Conference on Learning Representations. (ICLR) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.406, + 0.785, + 0.445 + ], + "angle": 0, + "content": "4. Baradad, M., Chen, C.F., Wulff, J., Wang, T., Feris, R., Torralba, A., Isola, P.: Procedural image programs for representation learning. In: Proceedings of the 2022 Advances in Neural Information Processing Systems (NeurIPS) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.448, + 0.785, + 0.488 + ], + "angle": 0, + "content": "5. Baradad, M., Wulff, J., Wang, T., Isola, P., Torralba, A.: Learning to See by Looking at Noise. In: Proceedings of the 2021 Advances in Neural Information Processing Systems (NeurIPS). vol. 34, pp. 2556-2569 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.491, + 0.785, + 0.516 + ], + "angle": 0, + "content": "6. Canny, J.: A computational approach to edge detection. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 8(6), 679-698 (1986)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.519, + 0.785, + 0.571 + ], + "angle": 0, + "content": "7. Caron, M., Touvron, H., Misra, I., Jegou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the 2021 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 9650-9660 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.574, + 0.785, + 0.628 + ], + "angle": 0, + "content": "8. Cazenavette, G., Wang, T., Torralba, A., Efros, A.A., Zhu, J.Y.: Dataset distillation by matching training trajectories. In: Proceedings of the 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10708-10717 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.631, + 0.785, + 0.67 + ], + "angle": 0, + "content": "9. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A Simple Framework for Contrastive Learning of Visual Representations. In: Proceedings of the 2020 International Conference on Machine Learning (ICML). vol. 119, pp. 1597-1607 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.673, + 0.785, + 0.698 + ], + "angle": 0, + "content": "0. Chen, X., Fan, H., Girshick, R., He, K.: Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.701, + 0.785, + 0.74 + ], + "angle": 0, + "content": "11. Chen, X., He, K.: Exploring simple siamese representation learning. In: Proceedings of the 2021 IEEE/CVF International Conference on Computer Vision and Pattern Recognition (CVPR). pp. 15750-15758 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.743, + 0.785, + 0.783 + ], + "angle": 0, + "content": "12. Chen, X., Xie, S., He, K.: An empirical study of training self-supervised vision transformers. In: Proceedings of the 2021 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 9640-9649 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.786, + 0.785, + 0.811 + ], + "angle": 0, + "content": "13. Cheng, G., Han, J., Lu, X.: Remote sensing image scene classification: Benchmark and state of the art. Proceedings of the IEEE 105(10), 1865-1883 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "14. Cubuk, E.D., Zoph, B., Shlens, J., Le, Q.V.: Randaugment: Practical data augmentation with no separate search. arXiv preprint abs/2104.07918 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.278, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.518, + 0.128 + ], + "angle": 0, + "content": "R. Nakamura and R. Tadokoro et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.272 + ], + "angle": 0, + "content": "15. Dehghani, M., Djolonga, J., Mustafa, B., Padlewski, P., Heek, J., Gilmer, J., Steiner, A.P., Caron, M., Geirhos, R., Alabdulmohsin, I., Jenatton, R., Beyer, L., Tschannen, M., Arnab, A., Wang, X., Riquelme Ruiz, C., Minderer, M., Puigcerver, J., Evci, U., Kumar, M., Steenkiste, S.V., Elsayed, G.F., Mahendran, A., Yu, F., Oliver, A., Huot, F., Bastings, J., Collier, M., Gritsenko, A.A., Birodkar, V., Vasconcelos, C.N., Tay, Y., Mensink, T., Kolesnikov, A., Pavetic, F., Tran, D., Kipf, T., Lucic, M., Zhai, X., Keysers, D., Harmsen, J.J., Houlsby, N.: Scaling vision transformers to 22 billion parameters. In: Proceedings of the 2023 International Conference on Machine Learning. vol. 202, pp. 7480-7512 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.273, + 0.787, + 0.327 + ], + "angle": 0, + "content": "16. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A large-scale hierarchical image database. In: Proceedings of the 2009 IEEE Computer Society Conference on Computer Vision and Pattern Recognition Workshops (CVPR Workshops). pp. 248-255 (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.328, + 0.787, + 0.37 + ], + "angle": 0, + "content": "17. Doersch, C., Gupta, A., Efros, A.A.: Unsupervised visual representation learning by context prediction. In: Proceedings of the 2015 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 1422-1430 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.371, + 0.787, + 0.438 + ], + "angle": 0, + "content": "18. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: Proceedings of the 2021 International Conference on Learning Representation (ICLR) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.439, + 0.787, + 0.481 + ], + "angle": 0, + "content": "19. Dwibedi, D., Misra, I., Hebert, M.: Cut, paste and learn: Surprisingly easy synthesis for instance detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 1310-1319 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.787, + 0.522 + ], + "angle": 0, + "content": "20. Everingham, M., Eslami, S.M.A., Van Gool, L., Williams, C.K.I., Winn, J., Zisserman, A.: The Pascal Visual Object Classes Challenge: A Retrospective. International Journal of Computer Vision (IJCV) 111(1), 98-136 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.787, + 0.563 + ], + "angle": 0, + "content": "21. Gidaris, S., Singh, P., Komodakis, N.: Unsupervised representation learning by predicting image rotations. In: International Conference on Learning Representation (ICLR) (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.787, + 0.605 + ], + "angle": 0, + "content": "22. He, K., Chen, X., Xie, S., Li, Y., Dólár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16000-16009 (June 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.787, + 0.647 + ], + "angle": 0, + "content": "23. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: IEEE/CVF International Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9729-9738 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.787, + 0.688 + ], + "angle": 0, + "content": "24. Inoue, N., Yamagata, E., Kataoka, H.: Initialization using perlin noise for training networks with a limited amount of data. In: Proceedings of the 2020 International Conference on Pattern Recognition (ICPR). pp. 1023-1028 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.787, + 0.743 + ], + "angle": 0, + "content": "25. Johnson, J., Hariharan, B., Maaten, L.v.d., Li, F.F., Zitnick, L., Girshick, R.: CLEVR: A diagnostic dataset for compositional language and elementary visual reasoning. In: Proceedings of the 2017 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1988-1997 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.661, + 0.758 + ], + "angle": 0, + "content": "26. Kaggle, Eyepacs: Kaggle diabetic retinopathy detection (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.787, + 0.813 + ], + "angle": 0, + "content": "27. Kataoka, H., Hara, K., Hayashi, R., Yamagata, E., Inoue, N.: Spatiotemporal initialization for 3D CNNs with generated motion patterns. In: Proceedings of the 2022 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 737-746 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "28. Kataoka, H., Hayamizu, R., Yamada, R., Nakashima, K., Takashima, S., Zhang, X., Martinez-Noriega, E.J., Inoue, N., Yokota, R.: Replacing labeled real-image" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.606, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Scaling Backwards" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.241, + 0.147, + 0.786, + 0.189 + ], + "angle": 0, + "content": "datasets with auto-generated contours. In: Proceedings of the 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 21232-21241 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.191, + 0.787, + 0.245 + ], + "angle": 0, + "content": "29. Kataoka, H., Matsumoto, A., Yamada, R., Satoh, Y., Yamagata, E., Inoue, N.: Formula-driven supervised learning with recursive tiling patterns. In: Proceedings of the 2021 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 4098-4105 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.247, + 0.786, + 0.288 + ], + "angle": 0, + "content": "30. Kataoka, H., Okayasu, K., Matsumoto, A., Yamagata, E., Yamada, R., Inoue, N., Nakamura, A., Satoh, Y.: Pre-training without natural images. International Journal of Computer Vision (IJCV) 130 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.29, + 0.787, + 0.331 + ], + "angle": 0, + "content": "31. Krause, J., Stark, M., Deng, J., Fei-Fei, L.: 3d object representations for fine-grained categorization. In: Proceedings of the 2013 International IEEE Workshop on 3D Representation and Recognition (3DRR-13). pp. 554-561 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.333, + 0.786, + 0.36 + ], + "angle": 0, + "content": "32. Krizhevsky, A., Hinton, G.: Learning Multiple Layers of Features from Tiny Images. Master's thesis, Department of Computer Science, University of Toronto (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.361, + 0.787, + 0.415 + ], + "angle": 0, + "content": "33. LeCun, Y., Huang, F.J., Bottou, L.: Learning methods for generic object recognition with invariance to pose and lighting. In: Proceedings of the 2004 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). vol. 2, pp. II-104 Vol.2 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.417, + 0.787, + 0.471 + ], + "angle": 0, + "content": "34. Mahajan, D., Girshick, R., Ramanathan, V., He, K., Paluri, M., Li, Y., Bharambe, A., Maaten, L.v.d.: Exploring the Limits of Weakly Supervised Pretraining. In: Proceedings of the 2018 European Conference on Computer Vision (ECCV). pp. 181-196 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.787, + 0.529 + ], + "angle": 0, + "content": "35. Nakamura, R., Kataoka, H., Takashima, S., Noriega, E.J.M., Yokota, R., Inoue, N.: Pre-training vision transformers with very limited synthesized images. In: Proceedings of the 2023 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 20303-20312. IEEE (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.531, + 0.787, + 0.572 + ], + "angle": 0, + "content": "36. Nilsback, M.E., Zisserman, A.: Automated flower classification over a large number of classes. In: Proceedings of the 2008 Indian Conference on Computer Vision, Graphics and Image Processing. pp. 722-729 (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.573, + 0.787, + 0.613 + ], + "angle": 0, + "content": "37. Noroozi, M., Favaro, P.: Unsupervised Learning of Visual Representations by Solving Jigsaw Puzzles. In: Proceedings of the 2016 European Conference on Computer Vision (ECCV). pp. 69-84 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.615, + 0.787, + 0.657 + ], + "angle": 0, + "content": "38. Noroozi, M., Vinjimoor, A., Favaro, P., Pirsiavash, H.: Boosting Self-Supervised Learning via Knowledge Transfer. In: IEEE/CVF International Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9359-9367 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.658, + 0.787, + 0.698 + ], + "angle": 0, + "content": "39. Remez, T., Huang, J., Brown, M.: Learning to Segment via Cut-and-Paste. In: Proceedings of the 2018 European Conference on Computer Vision (ECCV). pp. 39-54 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.7, + 0.787, + 0.755 + ], + "angle": 0, + "content": "40. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., Berg, A.C., Fei-Fei, L.: ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV) 115(3) (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.757, + 0.787, + 0.797 + ], + "angle": 0, + "content": "41. Sharma, P., Rott Shaham, T., Baradad, M., Fu, S., Rodriguez-Munoz, A., Duggal, S., Isola, P., Torralba, A.: A vision check-up for language models. In: arXiv preprint arXiv:2401.01862 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "42. Sun, C., Shrivastava, A., Singh, S., Gupta, A.: Revisiting unreasonable effectiveness of data in deep learning era. In: Proceedings of the 2017 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 843-852 (2017)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.518, + 0.128 + ], + "angle": 0, + "content": "R. Nakamura and R. Tadokoro et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "43. Sundermeyer, M., Marton, Z.C., Durner, M., Brucker, M., Triebel, R.: Implicit 3D orientation learning for 6D object detection from RGB images. In: Proceedings of the 2018 European Conference on Computer Vision (ECCV). pp. 712-729 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.786, + 0.232 + ], + "angle": 0, + "content": "44. Tadokoro, R., Yamada, R., Nakashima, K., Nakamura, R., Kataoka, H.: Primitive Geometry Segment Pre-training for 3D Medical Image Segmentation. In: Proceedings of the 2023 British Machine Vision Conference (BMVC) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.787, + 0.286 + ], + "angle": 0, + "content": "45. Takashima, S., Hayamizu, R., Inoue, N., Kataoka, H., Yokota, R.: Visual atoms: Pre-training vision transformers with sinusoidal waves. In: Proceedings of the 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 18579-18588 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.787, + 0.342 + ], + "angle": 0, + "content": "46. Tobin, J., Fong, R., Ray, A., Schneider, J., Zaremba, W., Abbeel, P.: Domain randomization for transferring deep neural networks from simulation to the real world. In: Proceedings of the 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 23-30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.787, + 0.397 + ], + "angle": 0, + "content": "47. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jegou, H.: Training data-efficient image transformers & distillation through attention. In: Proceedings of the 2021 International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 139, pp. 10347-10357 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.787, + 0.452 + ], + "angle": 0, + "content": "48. Veeling, B.S., Linmans, J., Winkens, J., Cohen, T., Welling, M.: Rotation equivariant cnns for digital pathology. In: Proceedings of the 2018 International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI) (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.787, + 0.493 + ], + "angle": 0, + "content": "49. Venkataramanan, S., Rizve, M.N., Carreira, J., Asano, Y.M., Avrithis, Y.: Is image-net worth 1 video? learning strong image encoders from 1 long unlabelled video. arXiv preprint arXiv:2310.08584 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.494, + 0.725, + 0.508 + ], + "angle": 0, + "content": "50. Wang, T., Zhu, J., Torralba, A., Efros, A.A.: Dataset Distillation (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.787, + 0.563 + ], + "angle": 0, + "content": "51. Yun, S., Han, D., Chun, S., Oh, S.J., Yoo, Y., Choe, J.: CutMix: Regularization strategy to train strong classifiers with localizable features. In: Proceedings of the 2019 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 6022-6031 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.787, + 0.605 + ], + "angle": 0, + "content": "52. Yun-Hao Cao, H.Y., Wu, J.: Training vision transformers with only 2040 images. In: Proceedings of the 2022 European Conference on Computer Vision (ECCV). pp. 220-237 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.787, + 0.646 + ], + "angle": 0, + "content": "53. Zhai, X., Kolesnikov, A., Houlsby, N., Beyer, L.: Scaling vision transformers. Proceedings of 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 1204-1213 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.787, + 0.715 + ], + "angle": 0, + "content": "54. Zhai, X., Puigcerver, J., Kolesnikov, A., Ruyssen, P., Riquelme, C., Lucic, M., Djolonga, J., Pinto, A.S., Neumann, M., Dosovitskiy, A., Beyer, L., Bachem, O., Tschannen, M., Michalski, M., Bousquet, O., Gelly, S., Houlsby, N.: A Large-scale Study of Representation Learning with the Visual Task Adaptation Benchmark. arXiv preprint arXiv:1910:04867 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.787, + 0.743 + ], + "angle": 0, + "content": "55. Zhang, H., Cisse, M., Dauphin, Y.N., Lopez-Paz, D.: Mixup: Beyond empirical risk minimization. arXiv preprint abs/1710.09412 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.787, + 0.771 + ], + "angle": 0, + "content": "56. Zhang, R., Isola, P., Efros, A.A.: Colorful image colorization. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 649-666 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.787, + 0.812 + ], + "angle": 0, + "content": "57. Zhao, B., Mopuri, K.R., Bilen, H.: Dataset condensation with gradient matching. In: Proceedings of the 2021 International Conference on Learning Representation (ICLR) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "58. Zhong, Z., Zheng, L., Kang, G., Li, S., Yang, Y.: Random erasing data augmentation. arXiv preprint abs/1708.04896 (2017)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.606, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Scaling Backwards" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.788, + 0.19 + ], + "angle": 0, + "content": "59. Zhou, B., Lapedriza, A., Khosla, A., Oliva, A., Torralba, A.: Places: A 10 million image database for scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 40(6), 1452-1464 (2017)" + } + ] +] \ No newline at end of file diff --git a/2024/Scaling Backwards_ Minimal Synthetic Pre-training_/82902d64-0219-4fba-9a2b-20c3a32a53cf_origin.pdf b/2024/Scaling Backwards_ Minimal Synthetic Pre-training_/82902d64-0219-4fba-9a2b-20c3a32a53cf_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..bb78c38f805739da6107312f4e41d2912c599501 --- /dev/null +++ b/2024/Scaling Backwards_ Minimal Synthetic Pre-training_/82902d64-0219-4fba-9a2b-20c3a32a53cf_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:307ccac54202a616c3ca37c85f7279bd3172c8248e7d46456c24879aec46a21f +size 4675839 diff --git a/2024/Scaling Backwards_ Minimal Synthetic Pre-training_/full.md b/2024/Scaling Backwards_ Minimal Synthetic Pre-training_/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8072835a0aa691eb02e4dc27065907817b41abe8 --- /dev/null +++ b/2024/Scaling Backwards_ Minimal Synthetic Pre-training_/full.md @@ -0,0 +1,367 @@ +# Scaling Backwards: Minimal Synthetic Pre-training? + +Ryo Nakamura $^{1,*}$ , Ryu Tadokoro $^{2,*}$ , Ryosuke Yamada $^{1}$ , Yuki M. Asano $^{3}$ , Iro Laina $^{4}$ , Christian Rupprecht $^{4}$ , Nakamasa Inoue $^{5}$ , Rio Yokota $^{5}$ , and Hirokatsu Kataoka $^{1}$ + +$^{1}$ National Institute of Advanced Industrial Science and Technology (AIST) + +2 Tohoku University + +3 University of Amsterdam, now at University of Technology Nuremberg + +$^{4}$ University of Oxford + +5 Tokyo Institute of Technology + +Abstract. Pre-training and transfer learning are an important building block of current computer vision systems. While pre-training is usually performed on large real-world image datasets, in this paper we ask whether this is truly necessary. To this end, we search for a minimal, purely synthetic pre-training dataset that allows us to achieve performance similar to the 1 million images of ImageNet-1k. We construct such a dataset from a single fractal with perturbations. With this, we contribute three main findings. (i) We show that pre-training is effective even with minimal synthetic images, with performance on par with large-scale pre-training datasets like ImageNet-1k for full fine-tuning. (ii) We investigate the single parameter with which we construct artificial categories for our dataset. We find that while the shape differences can be indistinguishable to humans, they are crucial for obtaining strong performances. (iii) Finally, we investigate the minimal requirements for successful pre-training. Surprisingly, we find that a substantial reduction of synthetic images from 1k to 1 can even lead to an increase in pre-training performance, a motivation to further investigate "scaling backwards". Finally, we extend our method from synthetic images to real images to see if a single real image can show similar pre-training effect through shape augmentation. We find that the use of grayscale images and affine transformations allows even real images to "scale backwards". The code is available at https://github.com/SUPER-TADORY/1p-frac. + +Keywords: Synthetic pre-training $\cdot$ Limited data $\cdot$ Vision transformers + +# 1 Introduction + +In image recognition, pre-training allows discovering fundamental visual representations for downstream task applications. Pre-training enhances the performance of visual tasks and enables the use of small-scale task-specific datasets. + +![](images/b7a7bdbc988c958c7b3f3b6dca22bfcf87384e4cd9567b3d49c86ed52a1d8076.jpg) +Fig. 1: Comparison of ImageNet-1k, FractalDB and 1p-frac (ours). 1p-frac consists of only a single fractal for pre-training. With 1p-frac, neural networks learn to classify perturbations applied to the fractal. In our study "single" means a very narrow distribution over parameters that leads to images that are roughly equivalent from a human visual perspective. While the shape differences of perturbed images can be indistinguishable to humans, models pre-trained on 1p-frac achieve comparable performance with those pre-trained on ImageNet-1k or FractalDB. + +Recently, pre-training has been used as a key technology to construct foundation models trained on massive datasets with over hundreds of millions of images. In some cases, a foundation model enables adaptation of zero-shot recognition without the need for additional data. + +Pre-training is often interpreted as discovering universal structures in large-scale datasets that later facilitate adaptation to down-stream tasks. In this paper, we challenge this interpretation by providing a minimal pre-training dataset, generated from a single fractal, that achieves similar downstream performance. At the heart of this investigation is the question of whether pre-training might simply be a better weight initialization rather than the discovery of useful visual concepts. If true, performing expensive pre-training with hundreds of millions of images might not be necessary. This additionally frees pre-training from licensing or ethical issues. + +Since the rise of deep neural networks, the ImageNet dataset [16] has been one of the most commonly used pre-training datasets. Originally, pre-training has been conducted through supervised learning (SL) with human-provided teacher labels. However, it has become clear that pre-training can also be achieved without human-provided labels, through self-supervised learning (SSL) [7,9-12,17, 21-23,37,38,56]. + +In this context, Asano et al. [3] successfully acquired visual representations while dramatically reducing the number of required images. They concluded that SSL can yield sufficient image representations even with a single training example, but only for earlier layers of a recognition model. However, it is unclear how these findings translate to modern architectures and representation learning methods. Based on this, vision transformers (ViT) [18] have been pre-trained with only 2,040 real images [52] through an instance discrimination learning signal. + +More recently, it has been shown that basic visual representations can be acquired without using real images and human-provided labels. The trend of + +artificially generated labeled images is on the rise for synthetic pre-training [5, 28, 30, 46]. Formula-driven supervised learning (FDSL) generates images from generative formulas, and labels from their parameters [30]. Under the FDSL framework, one can adapt the synthetic pre-training image dataset by altering formulas [24, 27, 29, 44]. While a million-order image dataset was constructed in FractalDB [30], our findings suggest that synthetic pre-training can be reduced to significantly fewer fractal images. + +Motivated by these findings, we believe it is possible to find the key essence of pre-training for image recognition. ViT training can be done with as few as 1,000 artificially generated images [35]. Here, we believe that equivalent performance can be achieved with even fewer images. This consideration is undoubtedly important as we approach a minimal synthetic pre-training dataset in image recognition, which goes against the trend of foundation models toward increasing the dataset scale. + +In the present paper, we thus introduce a minimal synthetic dataset, namely 1-parameter Fractal as Data (1p-frac), which consists of a single fractal as shown in Figure 1, as well as a loss function for pre-training with it. Our contributions regarding minimal synthetic pre-training are as follows: + +Ordinal minimalism: We introduce the locally perturbed cross entropy (LPCE) loss for pre-training with a single fractal. It utilizes perturbed fractal images for training, where neural networks learn to classify small perturbations. In experiments, we demonstrate that pre-training can be performed even with a single fractal. The pre-training effect of 1p-frac is comparable to that of a million-scale labeled image dataset. + +Distributional minimalism: We introduce the locally integrated empirical (LIEP) distribution $p_{\Delta}$ that has a controllable perturbation degree $\Delta$ to investigate the minimal support of the probability density distribution of synthesized images. We observed positive pre-training effects even with a small $\Delta$ producing shape differences that cannot be distinguished by humans. We also show that if $\Delta$ is too small the visual pre-training collapses. From these observations, we establish the general bounds for generating good pre-training images from a mathematical formula. + +Instance minimalism: Based on the experimental results, the synthetic images should not simply contain complex shapes. The use of recursive image patterns, similar to objects in nature, should be applied in visual pretraining. Experiments with augmented categories from a real image have shown that good pre-training effects can be achieved by performing "affine transformations on edge + +Table 1: Scaling backwards in synthetic pre-training (Accuracies on CIFAR-100, Real: ImageNet, Synth: Fractal images). + +
Type\#Img11k1M
RealN/A76.985.5
Synth84.284.081.6
+ +emphasized objects in grayscale". These operations are found to be almost synonymous with the configuration of the proposed 1p-frac. + +In summary, we significantly reduce the size of the pre-training dataset from 1M images (fractal database (FractalDB) [30]) or 1k images (one-instance fractal + +database (OFDB) [35]) down to 1 and show that this even improves the pretraining effect (Table 1), which motivates "scaling backwards". + +# 2 Related Work + +The present study focuses primarily on finding the minimal pre-training dataset with artificially generated images. + +Pre-training in image recognition. Pre-training in image recognition has demonstrated significant success with vast amounts of labeled images. Pre-training helps neural networks to acquire fundamental visual representations which are essential for improving the performance of diverse downstream tasks in image recognition. In particular, researchers have used larger-scale image datasets for supervised learning (SL), starting from million-scale datasets (e.g., ImageNet [16], Places [59]). More recently, the largest datasets used are reaching billion-scale (e.g., IG-3.5B [34], JFT-300M/3B/4B [15, 42, 53]). + +Pre-training can be also achieved without any human intervention. One of the most effective approaches is self-supervised learning (SSL [7,9-12,17,21-23,37, 38,56]), which has been developed in order to alleviate the manual annotations on large-amount of images. SSL has been studied as an expected alternative to supervised pre-training and has sometimes been shown to surpass supervised pre-training. + +Despite the development of supervised and self-supervised pre-training, the number of real images used in visual pre-training continues to increase. Exploratory research questions such as "what is a minimal synthetic pre-training dataset?" have received less attention. However, they could give us more insights on what is learned during pre-training and how it could be improved. + +Pre-training with limited data. In order to address ethical issues such as privacy and fairness associated with real images, a pre-training with a very limited number of real images and synthetic pre-training without any real images have been on the rise. In this context, synthetic pre-training with copy-and-paste learning [19,39], domain randomization from 3D graphics [43,46], learning from primitive patterns [4,5,41], and dataset distillation [8,50,57] have been recognized as promising approaches to acquiring visual representations. More recently, a sophisticated approach, formula-driven supervised learning (FDSL) [28,30], enabled automatic construction of a labeled image dataset without any real images or human annotations. FDSL can generate images and their corresponding labels using a mathematical formula. Along these lines, One-instance FractalDB (OFDB) has successfully pre-trained a ViT using only 1,000 synthetic images generated from a mathematical formula, without using real images [35]. Moreover, several studies have been reported in the context of extremely limited data [3,49,52]. In particular, Asano et al. have shown that learning visual representations is possible using a single image for self-supervised learning. Venkataramanan et al. achieved a comparable method to ImageNet pre-training with only 10 walking videos and their cropped images. + +By analyzing the acquisition of visual representations through the use of minimal synthetic pre-training datasets, we believe that the FDSL framework will be key to discovering the essence of pre-training mechanisms in image recognition. + +# 3 Scaling Backwards with a Single Fractal + +This section presents the proposed methodology to explore the minimal requirements for successful pre-training by utilizing purely synthetic images of fractals. Specifically, we introduce the 1 parameter Fractal as Data (1p-frac), which consists of only a single fractal, with a method to pre-train neural networks on it. Our key idea is to introduce the locally integrated empirical (LIEP) distribution $p_{\Delta}$ over perturbed fractal images, enabling pre-training even with a single fractal image. Since the LIEP distribution is designed so that it converges to the empirical distribution $p_{\mathrm{data}}(x) = \delta (x - I)$ of a single image $I$ when the perturbation degree $\Delta \in \mathbb{R}_{\geq 0}$ goes to zero, we can narrow down the support of the distribution by decreasing $\Delta$ as shown in Figure 2a. Below, we begin with a preliminary for defining the empirical distributions of two fractal databases, namely FractalDB [30] and OFDB [35]. We then introduce 1p-frac with the LIEP distribution. + +# 3.1 Preliminary + +FractalDB [30]. Kataoka et al. have introduced a method for effectively pretraining neural networks with FractalDB, a set of fractal images generated by the iterated function systems (IFSs). Specifically, FractalDB $\mathcal{F}$ consists of one million synthesized images: $\mathcal{F} = \{(\varOmega_c,\{I_i^c\}_{i = 0}^{M - 1})\}_{c = 0}^{C - 1}$ , where $\varOmega_{c}$ is an IFS, $I_{i}^{c}$ is a fractal image generated by $\varOmega_{c}$ , $C = 1,000$ is the number of fractal categories, and $M = 1,000$ is the number of images per category. Each IFS characterizes each fractal category $c$ and is defined as follows: + +$$ +\Omega_ {c} = \left\{\mathbb {R} ^ {2}; w _ {1}, w _ {2}, \dots , w _ {N _ {c}}; p _ {1}, p _ {2}, \dots , p _ {N _ {c}} \right\}, \tag {1} +$$ + +where $w_{j}:\mathbb{R}^{2}\to \mathbb{R}^{2}$ is a 2D affine transformation given by + +$$ +w _ {j} (\boldsymbol {v}) = \left[ \begin{array}{l l} a _ {j} & b _ {j} \\ c _ {j} & d _ {j} \end{array} \right] \boldsymbol {v} + \left[ \begin{array}{l} e _ {j} \\ f _ {j} \end{array} \right] (\boldsymbol {v} \in \mathbb {R} ^ {2}), \tag {2} +$$ + +and $p_j$ is a probability mass distribution. Each fractal image $I_i^c$ renders a fractal $F = \{\pmb{v}_t\}_{t=0}^T \subset \mathbb{R}^2$ into a 2D image, where points $\pmb{v}_t$ are determined by a recurrence relation that applies affine transformations as $\pmb{v}_{t+1} = w_{\sigma_t}(\pmb{v}_t)$ for $t = 0, 1, 2, \dots, T-1$ . Here, the initial point is set as $\pmb{v}_0 = (0,0)^\top$ , and the index $\sigma_t$ is sampled at each $t$ independently from the probability mass distribution $p(\sigma_t = j) = p_j$ . Pre-training with FractalDB utilizes the cross-entropy loss function: + +$$ +\mathcal {L} = - \mathbb {E} _ {x, y \sim p _ {\mathrm {d a t a}}} [ \log p _ {\theta} (y | x) ], \tag {3} +$$ + +![](images/d1a098d60ba4748ad3a145adc83d138d79f0d6283419d4c7ebafc33e335b120d.jpg) +Fig. 2: Scaling backwards from many images to a single synthetic image. (a) Empirical distribution $p_{\mathrm{data}}$ . Colors indicate classes. With a single image, the distribution is given by a single Dirac's delta function. (b) LIEP distribution $p_{\Delta}$ . The support of the distribution narrows as the degree of perturbation $\Delta$ decreases. (c) $\sigma$ -factor for investigating fractal shapes. A small $\sigma$ produces complex fractals. + +where $p_{\theta}$ is the category distribution predicted by a neural network and $\theta$ is a set of learnable parameters. The joint empirical distribution $p_{\mathrm{data}}$ is defined over the dataset as follows: + +$$ +p _ {\mathrm {d a t a}} (x, y; \mathcal {F}) = \frac {1}{M C} \sum_ {i = 0} ^ {M - 1} \sum_ {c = 0} ^ {C - 1} \delta \left(x - I _ {i} ^ {c}\right) \delta (y - c) \tag {4} +$$ + +where $\delta$ is the Dirac's delta function. Models pre-trained on this dataset perform comparable to those pre-trained on real-world image datasets, such as ImageNet-1k and Places365, in some downstream tasks. + +OFDB [35]. This dataset consists of 1,000 fractal images. Specifically, OFDB $\mathcal{F}_{\mathrm{OF}}$ involves only one representative image per category, i.e., $\mathcal{F}_{\mathrm{OF}} = \{\varOmega_c,I_c\}_{c = 0}^{C - 1}$ . Therefore, the joint empirical distribution reduces to + +$$ +p _ {\mathrm {d a t a}} (x, y; \mathcal {F} _ {\mathrm {O F}}) = \frac {1}{C} \sum_ {c = 0} ^ {C - 1} \delta \left(x - I _ {c}\right) \delta \left(y - c\right). \tag {5} +$$ + +Models pre-trained on this dataset perform comparable to or even better than those pre-trained on FractalDB. This work has shown that there exists a small but essential set of images for visual pre-training. However, reducing the number of fractals $C$ to less than 1,000 degrades the performance. + +# 3.2 Pre-training with a Single Fractal + +Scaling backwards. To further facilitate the analysis of what images are minimally required for successful visual pre-training, we introduce 1p-frac, which ultimately reduces the number of IFSs and images to one each as $\mathcal{F}_{\mathrm{OP}} = (\varOmega,I)$ . With this dataset, the empirical distribution is given by + +$$ +p _ {\mathrm {d a t a}} (x, y; \mathcal {F} _ {\mathrm {O P}}) = \delta (x - I) \delta (y). \tag {6} +$$ + +However, we notice that with this distribution, training neural networks using cross-entropy loss is not straightforward because $p_{\theta}(y = 0|x) \equiv 1 (\forall x)$ gives a trivial solution to the loss minimization problem. To address this, we introduce locally perturbed cross entropy (LPCE) loss $\mathcal{L}_{\Delta}$ , a variant of the cross entropy loss defined with the LIEP distribution. + +Definition 1. Let $I_{\epsilon} \in \mathcal{X}$ be a perturbed image, where $\mathcal{X}$ is a set of images, $\epsilon \in \mathbb{R}^d$ is a small perturbation with $d \in \mathbb{N}_{>0}$ and $I_0 = I$ is the original image. We define the LIEP distribution by + +$$ +p _ {\Delta} (x, y) = \frac {1}{| \mathcal {R} _ {\Delta} |} \int_ {\mathcal {R} _ {\Delta}} \delta (x - I _ {\epsilon}) \delta (y - \epsilon) d \epsilon \tag {7} +$$ + +where $\mathcal{R}_{\Delta} \subset \mathbb{R}^{d}$ is a compact set containing the origin and $|\mathcal{R}_{\Delta}|$ is its volume with order $O(|\Delta|^{d})$ . + +Definition 2. We define the LPCE loss by + +$$ +\mathcal {L} _ {\Delta} = - \mathbb {E} _ {x, y \sim p _ {\Delta}} \left[ \log p _ {\theta} (y | x) \right], \tag {8} +$$ + +where $p_{\Delta}$ is the LIEP distribution. + +If $\mathcal{R}_{\Delta}$ is a small hypercube or hypersphere, it is straightforward to see that $p_{\Delta}$ approaches to the empirical distribution of Eq. (6) when $\Delta$ goes to zero. Therefore, this loss allows us to analyze visual pre-training effects by narrowing the support of the distribution around a single image. + +With 1p-frac, we apply perturbation to the affine transformations. As such, the perturbation $\epsilon$ is in $\mathbb{R}^{6*j}$ . A perturbed image $I_{\epsilon}$ is obtained by the noisy affine transformations: + +$$ +w _ {j} \left(\boldsymbol {v}; \boldsymbol {\epsilon} _ {j}\right) = \left(\left[ \begin{array}{l l l} a _ {j} & b _ {j} & e _ {j} \\ c _ {j} & d _ {j} & f _ {j} \end{array} \right] + \boldsymbol {\epsilon} _ {j}\right) \left[ \begin{array}{l} \boldsymbol {v} \\ 1 \end{array} \right] \tag {9} +$$ + +where $\epsilon_{j} \in \mathcal{R}_{\Delta} = [-\Delta / 2, \Delta / 2]^{6 * j} \subset \mathbb{R}^{6 * j}$ is a hypercube with a side length $\Delta$ , and $|\mathcal{R}_{\Delta}| = \Delta^{6 * j}$ . Note that the numerical integration of Eq. (7) is used in practice, which is an approximation obtained by uniformly sampling $L$ points in $\mathcal{R}_{\Delta}$ , where we set $L = 1,000$ by default. + +Visualization. Figure 2b shows examples of perturbed images to compute the LPCE loss. While most of shape differences are indistinguishable to humans, a neural network learns to distinguish the perturbations applied to the single image by minimizing the LPCE loss. + +Complexity of $\Omega$ . We use the $\sigma$ -factor proposed by Anderson et al. [1] to evaluate the complexity of IFSs. As shown in Figure 2c, small values of $\sigma$ produce complex fractal shapes. + +# 4 Experiments + +Experimental setup. What is minimal visual pre-training, and what properties of images lend themselves to reducing their count? In order to clarify these questions, we conduct experiments following these steps: + +- Exploration study (Tables 2-4): We verify whether a single fractal is sufficient for pre-training. Moreover, we compare the extent to which the shapes used in pre-training can be simplified, using object contours, Gaussian, and uniform distributions. +- Hyperparameter study (Tables 5-7): We investigate the effects of three hyperparameters $\Delta$ , $\sigma$ and $L$ . +- Scaling study (Table 8): We compare 1p-frac with other large-scale datasets in terms of fine-tuning accuracy on ImageNet-1k. +- Analysis and discussion (Tables 9-12): We discuss data augmentation, computational cost for synthesizing images, and pre-training using a single real-world image. +- Applications (Tables 13 and 14): We conduct comparisons in various fine-tuning datasets. + +Implementation details. In order to verify the effects of pre-training, we measure the accuracy of downstream tasks through fine-tuning. We use ViT [18] for all experiments. Specifically, we employ ViT-Tiny (ViT-T) in the exploration and hyperparameter studies, and employ ViT-Base (ViT-B) for the scaling study. The pre-training parameters and data augmentation methods for pre-training and fine-tuning follow conventional methods in OFDB. In OFDB, data augmentation is based on DeiT. For exploration studies, CIFAR-100 (C100) [32] and/or ImageNet-100 (IN100) [40] are used as fine-tuning datasets. In the comparisons, we assign seven representative datasets used in OFDB paper, namely C10/C100, Cars [31], Flowers [36], Pascal VOC 2012 (VOC12) [20], Places-30 (P30) [59], IN100. See supplementary material for more detailed experimental settings. + +# 4.1 Exploration Study + +Pre-training with 1p-frac. Table 2 compares 1p-frac with FracdalDB (1M images, 1k categories) and OFDB (1k images, 1k categories). As can be seen, the pre-training effect of 1p-frac is comparable to that of FractalDB and OFDB. The number of fractal categories is significantly reduced from $C = 1,000$ to only 1. This shows the effectiveness and efficiency of 1p-frac and the LPCE loss. + +Comparison with SOTA FDSL datasets. Table 3 compares 1p-frac with two SOTA FDSL datasets: (1) Radial Contour Database (RCDB) [28], which consists of 1 million images of synthetic polygons, and (2) Visual Atoms (VA) [45], which consists of 1 million images of parameterized wave functions. The "1K" counterparts in the table reduce the number of images per category to one. We also applied the LPCE loss to these two datasets by applying perturbations to the radios of a polygon or wave. As can be seen, 1p-frac performs the best when a single image is used for pre-training and outperforms the 1K FDSL datasets. + +Pre-training with a noisy image. To validate the necessity of a complex shape when pre-training with a single image, we applied the LPCE loss to two noisy images: an image of Gaussian noise and an image of uniform noise. Since these two noise images are generated by determining the parameters of the parametric distributions, Gaussian and uniform, the LPCE loss can be computed by + +Table 2: Comparison of 1p-frac with FractalDB [30] and OFDB [35]. $\diamond$ indicates the use of LPCE loss. We report top-1 accuracies $(\%)$ . + +
Method#ImgC100IN100
Scratch-64.274.9
FractalDB1M81.688.5
OFDB1k84.088.6
1p-frac°184.289.0
+ +Table 3: Comparison with SOTA FDSL datasets. $\diamond$ indicates the use of LPCE loss. + +
Method#ImgC100IN100
RCDB1M81.688.5
RCDB1K80.487.5
RCDB°182.587.9
VA1M84.990.3
VA1K82.188.5
VA°182.688.0
ip-frac°184.289.0
+ +Table 4: Comparison with pre-training with a single noise image. + +
MethodC100IN100
Gaussian1.15.7
Uniform2.071.1
1p-frac84.289.0
+ +Table 5: Effects of pertur- Table 6: Effects of $\sigma$ - bation degree $\varDelta$ $(\sigma=3.5)$ $(\varDelta=0.1)$ + +
ΔC100IN100
0.0011.21.9
0.0119.961.8
0.0583.088.2
0.184.289.0
0.283.488.5
1.082.688.1
+ +
σC100IN100
6.081.386.3
5.082.287.1
4.581.986.9
3.584.289.0
4.082.887.9
+ +Table 7: Effects of number of sampling points $L$ for numerical integration. + +
LC100IN100
1678.785.3
6482.388.0
51282.488.0
100084.289.0
+ +giving small perturbations to the parameters. Table 4 shows these images and fine-tuning results. For example, here the $\epsilon$ is added to the mean vector and covariance matrix of the Gaussian distribution. The results revealed that noise images fail in visual pre-training, performing worse than training from scratch. This suggests that acquiring fundamental visual representations through pretraining requires an image of a certain structured object, such as a fractal. + +# 4.2 Hyperparameter Study + +Perturbation degree. Table 5 shows the results obtained by different perturbation degrees $\Delta$ for the LIEP distribution. The results indicate that setting the value to 0.1 yields the highest performance. Interestingly, the performance significantly improved when $\Delta$ was increased from 0.01 to 0.05. This suggests that the support of the empirical distribution must have a certain size to obtain positive pre-training effects. + +$\sigma$ -factor. Table 6 investigates the effects of the $\sigma$ -factor, which measures the complexity of fractal shapes. As can be seen, the most complex shape yields the highest performance. It is worth noting that, even with a fractal of $\sigma = 6.0$ , which looks like Gaussian noise, we observed positive pre-training effects. This suggests that the shapes obtained from IFS are crucial for pre-training. + +Numerical integration. Tables 7 reduces the number of sampling points $L$ of the numerical integration when approximately computing loss with Eq. (7). + +Table 8: Comparison of pre-training datasets on ImageNet-1k fine-tuning. Fine-tuning accuracies calculated by using ViT-Base (B) are listed in the table. 21k indicates the number of categories in each pre-training dataset. + +
Pre-training#ImgTypeViT-B
Scratch--79.8
ImageNet-21k14MSL81.8
FractalDB-21k21MFDSL81.8
ExFractalDB-21k21MFDSL82.7
RCDB-21k21MFDSL82.4
VA-21k21MFDSL82.7
OFDB-21k21kFDSL82.2
3D-OFDB-21k21kFDSL82.7
1p-frac (ours)1FDSL82.1
+ +Table 9: Ablation study on data augmentation methods. The DeiT augmentation [47] is used as a baseline. + +
MethodC100
Baseline84.2
w/o Random Aug. [14]83.4
w/o Random Crop80.1
w/o Rand Aspect83.6
w/o Rand Erasing [58]84.0
w/o Mixup [55]83.6
w/o Cutmix [51]83.3
w/o Mixup/Cutmix80.4
w/o Flipping84.3
w/o Color Jittering84.2
+ +We see that larger numbers result in high performance. Interestingly, even with $L = 16$ , the pre-training effect was positive (better than training from scratch). + +# 4.3 Scaling Study + +ImageNet-1k fine-tuning (Table 8). While this work discusses minimal requirements for visual pre-training, increasing the approximation precision of the numerical integration could further improve the performance and benefit training of large models. In Table 8, we applied the LPCE loss with $L = 21,000$ to the ViT-B model and compare fine-tuning accuracy on ImageNet-1k. Surprisingly, and despite using a single fractal as data, our ViT-B pre-trained on 1p-frac outperforms pre-training on ImageNet-21k. This shows the strength of our approach in particular because we simply keep the same fine-tuning protocols as used for the original Imagenet-21k to ImageNet-1k transfer proposed in [18], putting us at a potential disadvantage. + +# 4.4 Analysis and Discussion + +Relationship with data augmentation (Table 9). We consider that the variation in image patterns also depends on data augmentation. While traditional methods primarily rely on the data augmentation techniques described in the DeiT paper [47], we investigated which data augmentation methods contribute significantly by excluding one or two techniques from the base data augmentation methods and observing the changes in fine-tuning accuracy on C100 dataset. The experimental results in Table 9 reveal that excluding Random Cropping (RandomCrop) significantly deteriorates accuracy, suggesting that omitting parts of the image region and causing partial loss facilitates the acquisition of the ability to focus on any part without relying on grasping the overall shape, greatly reducing the pre-training effect. Furthermore, when both Mixup and Cutmix are + +Table 10: Dataset construction time (hours). FractalDB (1M images), OFDB (1k images) and 1p-frac (1k perturbations for one image) are compared. The processing time is shown separately for fractal category search (Search), image rendering (Render), and total time (Total). + +
DatasetSearchRenderTotal
FractalDB2.3716.8619.23
OFDB2.370.412.78
1p-frac0.00220.0360.0382
+ +![](images/4c0c45ed6ad09ba8ef1a6b3811581f7e7e5eefd229e320905c14f902b1e7bd77.jpg) +Fig. 3: Linear probing following the experiments from [3]. We verified the performance with $\{1,2,3,4,6\}$ layers in the ViT-Tiny model on C100 dataset. Here, 2 and 4 show the use of up to 1-2 and 1-4 layers. + +excluded, a degradation in accuracy is observed, indicating that mixing images and categories enables better feature recognition. + +Dataset construction times (Table 10). We calculated the processing times in fractal category search and image rendering. By comparing to the conventional approaches, our 1p-frac recorded much faster category search (0.0022 hours nearly equals to 8 seconds) and image rendering (0.036 hours nearly equals to 129 seconds). 1p-frac requires only one category with a parameter set. This very simple procedure leads to efficiently construct a minimum requirement for pre-training in image recognition. + +Linear probing (Figure 3). To compare supervised ImageNet-1k pre-training with our 1p-frac pre-training in the context of early layer visual representations, we executed a linear probing experiment on the C100 dataset. In the ViT-T model with $\{1,2,3\}$ layers, the 1p-frac pre-trained model outperformed the model pre-trained on ImageNet-1k. This demonstrates the superior quality of early, low-level representations learned from our dataset compared to ImageNet. In effect, layers up to layer 3 in ViT-T can be sufficiently pre-trained and frozen using our 1p-frac. + +Pre-training with a single real image and shape augmentation. + +It was found that pretraining effects comparable to a dataset of one million real-world images can be achieved with 1p-frac, but is it possible to perform pre-training similarly with a real image? In this section, we verify whether + +![](images/0bd2c3b5a3be8f5ed962b45e2d63c6123d606f9cc6805f4303e57c9cc3a5b56e.jpg) +Fig. 4: Shape augmentation with three geometry transformations from a single real image. + +![](images/5ac48b223ad66e29540d7a9c91e7e54528fc10e51f617095edc6497d40f3eb62.jpg) + +![](images/d7505d95cded1adba0c183a93544a96bf82273c027b02e0831ebbce2b5d9d17c.jpg) + +![](images/ecece888f4e4acfba674c84de78801f5c23d4a1bfb0f7bba46b793bd395af5b1.jpg) + +visual features can be acquired through pre-training using a single real image (each image in Figure 5) with the LPCE loss using image representations similar to FDSL (Canny [6] images) and geometric transformations for perturbation (\{affine, elastic, polynomial\} transformations in Figure 4). The results of pre + +![](images/858924f9b8dbfdcc015dda5ccb40b1d07b941fd790e07d2a5f52d7101df9d932.jpg) +Fig. 5: The five different images used in the pre-training with a single real image and shape augmentation. + +Table 11: Comparison of pre-training effect between RGB and Canny images when shape augmentation is applied. The shape augmentation uses affine transformations. + +
ImageMean Accuracy
RGB image81.6
Canny image82.2
+ +Table 12: Comparison of pre-training effects of affine, elastic, and polynomial image transformations. Real Img type with Canny applied to the images. + +
TransformationMean Accuracy
Affine trans.81.9
Elastic trans.37.0
Polynomial trans.81.3
+ +training effects with $L = 1,000$ are shown in Tables 11 and 12. Note that the results in these tables represent the average score on C100 over the five real images in Figure 5. See the supplementary material for results for each of image. According to the results, we find that contour-emphasized Canny images omitting color information and affine transformation perturbation yield better pre-training, as measured by the fine-tuning accuracies. The operations are similar to the classification of labeled fractal images via IFS. + +# 4.5 Applications + +Fine-tuning on commonly used datasets (Table 13). We compare fine-tuning accuracies in three types of dataset configurations. For supervised learning (SL) we compare pretraining on ImageNet-1k [16] and Places-365 [59], for self-supervised learning (SSL) we compare DINO [7] and masked auto-encoder (MAE) [22] pre-trained on ImageNet and PASS, and models trained on the fractal datasets FDSL on FractalDB-1k and RCDB-1k. Following the setting of one-instance per category [35], we also evaluate using random patch augmentation ("w/ Aug.") We compare our 1p-frac with 1-image SSL [3] with DINO and MAE (we use Real Img #1 of Figure 5). The results show that our 1p-frac pretraining with one single image demonstrated high fine-tuning accuracies (87.5% and 88.2% with and without random patch augmentation) which are close to the performance of self-supervised ImageNet-1k pre-training with MAE that utilizes 1.28M images and yields 88.5%. Surprisingly, despite using only one fractal image with perturbations, our 1p-frac without random patch augmentation has demonstrated fine-tuning accuracies that are equal to or even higher than that of conventional OFDB-1k/FractalDB-1k, which uses 1,000 parameter sets with fractal categories and 1,000/1,000,000 images: 1p-frac obtains mean performances of 87.5% vs OFDB-1k 87.2% vs FractalDB-1k 87.1%. + +Fine-tuning on specialized and structured datasets in VTAB [54] (Table 14). We also verify performance on a couple of specialized and structured + +Table 13: Comparison among pre-training methods in fine-tuning accuracies. Best values at each dataset scale are in bold. ViT-T is used for all experiments. #OrgImg indicates number of collected images in real-image datasets or number of original images used in synthetic datasets. Type shows supervised types within supervised learning using cross-entropy loss (SL), self-supervised learning using DINO [7] or MAE [22] (SSL:D/SSL:M), and formula-supervised learning (FDSL). + +
Pre-training#OrgImgTypeC10C100CarsFlowersVOC12P30IN100Mean
Scratch--80.362.113.771.456.276.274.862.1
Places-365 [59]1.80MSL97.683.989.299.384.6-89.4-
ImageNet-1k [16]1.28MSL98.085.589.999.488.780.0--
ImageNet-1k [16]1.28MSSL:D97.782.488.098.574.778.489.086.9
ImageNet-1k [16]1.28MSSL:M97.485.886.696.383.380.290.088.5
PASS [2]1.43MSSL:D97.584.086.498.682.979.082.987.8
FractalDB-1k [30]1.00MFDSL96.881.686.098.380.678.488.387.1
RCDB-1k [28]1.00MFDSL97.082.286.598.980.979.788.587.6
ImageNet-1k [35]1,000SL94.376.957.394.873.878.284.379.9
ImageNet-1k [35]1,000SSL:D94.978.071.294.675.578.684.982.5
OFDB-1k [35]1,000FDSL96.984.084.597.179.979.988.087.2
3D-OFDB-1k [35]1,000FDSL97.183.885.598.480.880.089.187.8
OFDB-1k w/ Aug. [35]1,000FDSL97.285.387.698.381.480.489.588.5
3D-OFDB-1k w/ Aug. [35]1,000FDSL97.084.785.698.381.279.888.987.9
1-image SSL [3]1SSL:D95.779.573.193.869.080.488.782.8
1-image SSL [3]1SSL:M97.284.782.194.678.580.389.086.6
1p-frac (ours)1FDSL96.984.284.597.480.580.689.087.5
1p-frac w/ Aug. (ours)1FDSL96.584.787.098.180.980.588.988.2
+ +datasets in VTAB (visual task adaptation benchmark; Retinopathy [26], Resisc45 [13], Camelyon [48], CLEVR-Count [25], and sNORB-Azim [33]). The experiment is more suitable to evaluate and compare the pre-training methods since the image domains and their labels are distinct from both real and fractal images on ImageNet and 1p-frac. According to Table 14, our 1p-frac pretrained model performed relatively similar fine-tuning accuracies on all listed datasets in the table. By comparing to self-supervised ImageNet-1k with MAE, our 1p-frac surpassed the fine-tuning accuracies on three (Retinopathy, Resisc45, and sNORB-Azim) out of the five datasets. + +# 5 Conclusion and Discussion + +This paper examined what a minimal dataset for synthetic pre-training might look like. We proposed the 1-parameter Fractal as Data (1p-frac) dataset that succeeds in pre-training even with a single fractal by utilizing the locally perturbed cross entropy (LPCE) loss. The following findings were presented: + +What is a minimal synthetic pre-training dataset? In this paper, we applied IFS to generate labeled images. In this context, preparing a single set of parameters $a - f$ in IFS is the least informative. It recorded fine-tuning accuracy nearly equivalent to conventional methods such as FractalDB and OFDB (Tables 1 and 2), succeeding in pre-training with minimal image pattern utilization. + +Table 14: Performance comparisons on five fine-tuning datasets: Retinopathy (Retino) [26], Resisc45 [13], Camelyon [48], CLEVR-Count (CLEVR-C) [25], and sNORB-Azim (sNORB-A) [33]. Best and second-best values are in bold and underlined, respectively. + +
#OrgImgTypeRetinoResisc45CamelyonCLEVR-CsNORB-A
Scratch--66.989.274.446.411.5
ImageNet-1k [16]1.28MSL79.197.082.789.329.2
ImageNet-1k [16]1.28MSSL:M76.795.683.989.526.3
1p-frac (ours)1FDSL77.995.882.786.228.9
+ +tion. Moreover, it is not that any perturbation in image space is acceptable, as Gaussian/uniform distributions failed in pre-training (Table 4). It clarified the importance of geometric transformations with certain regularities like fractal geometry or contour shapes (Table 3). + +What properties of images lead to image reduction? In 1p-frac, we investigated shape changes and pre-training effects by adopting the LPCE loss and the $\sigma$ -factor that contribute to fractal shape variation. It was revealed that pre-training efficacy varies with shape variation (Tables 5, 6). For real images, it is not just about including more edge components, but about including affine transformations, which closely resembles the classification of labeled fractal images through IFS (Tables 11 and 12). By extensively abating data augmentation showed, we demonstrated the essential role of random cropping and MixUp. + +Pre-training effects with the proposed method. Supervised pre-training on ImageNet was more accurate for datasets where object labels are assigned to real images (Table 13), but in VTAB's Specialized/Structured datasets (Table 14), performances were relatively close. On the other hand, our 1p-frac pre-trained ViT performed a similar mean accuracy by comparing to MAE self-supervised ImageNet pre-training (MAE 88.5 vs. our 1p-frac w/ Aug. 88.2). Furthermore, increasing the number of sampling points $L$ showed performance improvement even with 21k categories and demonstrated good performance in ImageNet-1k fine-tuning (Table 8), indicating the potential for building large models with limited data resources. + +Broader impact. Our proposed dataset 1p-frac does not suffer from licensing or ethical issues such as biases and has a large potential to serve as a clean pretraining dataset. While not investigated in this paper, this approach also opens the possibility for significant gains in training speed by keeping the data on the GPU and applying transformations there, removing the GPU-CPU transfer bottleneck. + +Limitation. In this paper, we have explored achieving a minimal synthetic pre-training in image recognition based on fractal geometry with IFS (a single fractal, and their perturbations). Despite this, we believe it is also necessary to find a similarly minimal pre-training dataset containing real images - for some applications and increased interpretability. This might allow for quick and clear calibration of pre-training, similar to how calibration images are used in photography. We did not investigate this direction in this paper but leave this for future work. + +# Acknowledgements + +Computational resource of AI Bridging Cloud Infrastructure (ABCI) provided by National Institute of Advanced Industrial Science and Technology (AIST) was used. + +# References + +1. Anderson, C., Farrell, R.: Improving fractal pre-training. In: Proceedings of the 2022 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 1300-1309 (2022) +2. Asano, Y., Rupprecht, C., Zisserman, A., Vedaldi, A.: Pass: An imagenet replacement for self-Supervised pretraining without humans. In: Proceedings of the 2021 Neural Information Processing Systems Track on Datasets and Benchmarks (2021) +3. Asano, Y.M., Rupprecht, C., Vedaldi, A.: A critical analysis of self-supervision, or what we can learn from a single image. In: Proceedings of the 2020 International Conference on Learning Representations. (ICLR) (2020) +4. Baradad, M., Chen, C.F., Wulff, J., Wang, T., Feris, R., Torralba, A., Isola, P.: Procedural image programs for representation learning. In: Proceedings of the 2022 Advances in Neural Information Processing Systems (NeurIPS) (2022) +5. Baradad, M., Wulff, J., Wang, T., Isola, P., Torralba, A.: Learning to See by Looking at Noise. In: Proceedings of the 2021 Advances in Neural Information Processing Systems (NeurIPS). vol. 34, pp. 2556-2569 (2021) +6. Canny, J.: A computational approach to edge detection. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 8(6), 679-698 (1986) +7. Caron, M., Touvron, H., Misra, I., Jegou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the 2021 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 9650-9660 (2021) +8. Cazenavette, G., Wang, T., Torralba, A., Efros, A.A., Zhu, J.Y.: Dataset distillation by matching training trajectories. In: Proceedings of the 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10708-10717 (2022) +9. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A Simple Framework for Contrastive Learning of Visual Representations. In: Proceedings of the 2020 International Conference on Machine Learning (ICML). vol. 119, pp. 1597-1607 (2020) +0. Chen, X., Fan, H., Girshick, R., He, K.: Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297 (2020) +11. Chen, X., He, K.: Exploring simple siamese representation learning. In: Proceedings of the 2021 IEEE/CVF International Conference on Computer Vision and Pattern Recognition (CVPR). pp. 15750-15758 (2021) +12. Chen, X., Xie, S., He, K.: An empirical study of training self-supervised vision transformers. In: Proceedings of the 2021 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 9640-9649 (2021) +13. Cheng, G., Han, J., Lu, X.: Remote sensing image scene classification: Benchmark and state of the art. Proceedings of the IEEE 105(10), 1865-1883 (2017) +14. Cubuk, E.D., Zoph, B., Shlens, J., Le, Q.V.: Randaugment: Practical data augmentation with no separate search. arXiv preprint abs/2104.07918 (2019) + +15. Dehghani, M., Djolonga, J., Mustafa, B., Padlewski, P., Heek, J., Gilmer, J., Steiner, A.P., Caron, M., Geirhos, R., Alabdulmohsin, I., Jenatton, R., Beyer, L., Tschannen, M., Arnab, A., Wang, X., Riquelme Ruiz, C., Minderer, M., Puigcerver, J., Evci, U., Kumar, M., Steenkiste, S.V., Elsayed, G.F., Mahendran, A., Yu, F., Oliver, A., Huot, F., Bastings, J., Collier, M., Gritsenko, A.A., Birodkar, V., Vasconcelos, C.N., Tay, Y., Mensink, T., Kolesnikov, A., Pavetic, F., Tran, D., Kipf, T., Lucic, M., Zhai, X., Keysers, D., Harmsen, J.J., Houlsby, N.: Scaling vision transformers to 22 billion parameters. In: Proceedings of the 2023 International Conference on Machine Learning. vol. 202, pp. 7480-7512 (2023) +16. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A large-scale hierarchical image database. In: Proceedings of the 2009 IEEE Computer Society Conference on Computer Vision and Pattern Recognition Workshops (CVPR Workshops). pp. 248-255 (2009) +17. Doersch, C., Gupta, A., Efros, A.A.: Unsupervised visual representation learning by context prediction. In: Proceedings of the 2015 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 1422-1430 (2015) +18. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: Proceedings of the 2021 International Conference on Learning Representation (ICLR) (2021) +19. Dwibedi, D., Misra, I., Hebert, M.: Cut, paste and learn: Surprisingly easy synthesis for instance detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 1310-1319 (2017) +20. Everingham, M., Eslami, S.M.A., Van Gool, L., Williams, C.K.I., Winn, J., Zisserman, A.: The Pascal Visual Object Classes Challenge: A Retrospective. International Journal of Computer Vision (IJCV) 111(1), 98-136 (2015) +21. Gidaris, S., Singh, P., Komodakis, N.: Unsupervised representation learning by predicting image rotations. In: International Conference on Learning Representation (ICLR) (2018) +22. He, K., Chen, X., Xie, S., Li, Y., Dólár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16000-16009 (June 2022) +23. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: IEEE/CVF International Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9729-9738 (2020) +24. Inoue, N., Yamagata, E., Kataoka, H.: Initialization using perlin noise for training networks with a limited amount of data. In: Proceedings of the 2020 International Conference on Pattern Recognition (ICPR). pp. 1023-1028 (2020) +25. Johnson, J., Hariharan, B., Maaten, L.v.d., Li, F.F., Zitnick, L., Girshick, R.: CLEVR: A diagnostic dataset for compositional language and elementary visual reasoning. In: Proceedings of the 2017 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1988-1997 (2017) +26. Kaggle, Eyepacs: Kaggle diabetic retinopathy detection (2015) +27. Kataoka, H., Hara, K., Hayashi, R., Yamagata, E., Inoue, N.: Spatiotemporal initialization for 3D CNNs with generated motion patterns. In: Proceedings of the 2022 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 737-746 (2022) +28. Kataoka, H., Hayamizu, R., Yamada, R., Nakashima, K., Takashima, S., Zhang, X., Martinez-Noriega, E.J., Inoue, N., Yokota, R.: Replacing labeled real-image + +datasets with auto-generated contours. In: Proceedings of the 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 21232-21241 (2022) +29. Kataoka, H., Matsumoto, A., Yamada, R., Satoh, Y., Yamagata, E., Inoue, N.: Formula-driven supervised learning with recursive tiling patterns. In: Proceedings of the 2021 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 4098-4105 (2021) +30. Kataoka, H., Okayasu, K., Matsumoto, A., Yamagata, E., Yamada, R., Inoue, N., Nakamura, A., Satoh, Y.: Pre-training without natural images. International Journal of Computer Vision (IJCV) 130 (2022) +31. Krause, J., Stark, M., Deng, J., Fei-Fei, L.: 3d object representations for fine-grained categorization. In: Proceedings of the 2013 International IEEE Workshop on 3D Representation and Recognition (3DRR-13). pp. 554-561 (2013) +32. Krizhevsky, A., Hinton, G.: Learning Multiple Layers of Features from Tiny Images. Master's thesis, Department of Computer Science, University of Toronto (2009) +33. LeCun, Y., Huang, F.J., Bottou, L.: Learning methods for generic object recognition with invariance to pose and lighting. In: Proceedings of the 2004 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). vol. 2, pp. II-104 Vol.2 (2004) +34. Mahajan, D., Girshick, R., Ramanathan, V., He, K., Paluri, M., Li, Y., Bharambe, A., Maaten, L.v.d.: Exploring the Limits of Weakly Supervised Pretraining. In: Proceedings of the 2018 European Conference on Computer Vision (ECCV). pp. 181-196 (2018) +35. Nakamura, R., Kataoka, H., Takashima, S., Noriega, E.J.M., Yokota, R., Inoue, N.: Pre-training vision transformers with very limited synthesized images. In: Proceedings of the 2023 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 20303-20312. IEEE (2023) +36. Nilsback, M.E., Zisserman, A.: Automated flower classification over a large number of classes. In: Proceedings of the 2008 Indian Conference on Computer Vision, Graphics and Image Processing. pp. 722-729 (2008) +37. Noroozi, M., Favaro, P.: Unsupervised Learning of Visual Representations by Solving Jigsaw Puzzles. In: Proceedings of the 2016 European Conference on Computer Vision (ECCV). pp. 69-84 (2016) +38. Noroozi, M., Vinjimoor, A., Favaro, P., Pirsiavash, H.: Boosting Self-Supervised Learning via Knowledge Transfer. In: IEEE/CVF International Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9359-9367 (2018) +39. Remez, T., Huang, J., Brown, M.: Learning to Segment via Cut-and-Paste. In: Proceedings of the 2018 European Conference on Computer Vision (ECCV). pp. 39-54 (2018) +40. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., Berg, A.C., Fei-Fei, L.: ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV) 115(3) (2015) +41. Sharma, P., Rott Shaham, T., Baradad, M., Fu, S., Rodriguez-Munoz, A., Duggal, S., Isola, P., Torralba, A.: A vision check-up for language models. In: arXiv preprint arXiv:2401.01862 (2024) +42. Sun, C., Shrivastava, A., Singh, S., Gupta, A.: Revisiting unreasonable effectiveness of data in deep learning era. In: Proceedings of the 2017 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 843-852 (2017) + +43. Sundermeyer, M., Marton, Z.C., Durner, M., Brucker, M., Triebel, R.: Implicit 3D orientation learning for 6D object detection from RGB images. In: Proceedings of the 2018 European Conference on Computer Vision (ECCV). pp. 712-729 (2018) +44. Tadokoro, R., Yamada, R., Nakashima, K., Nakamura, R., Kataoka, H.: Primitive Geometry Segment Pre-training for 3D Medical Image Segmentation. In: Proceedings of the 2023 British Machine Vision Conference (BMVC) (2023) +45. Takashima, S., Hayamizu, R., Inoue, N., Kataoka, H., Yokota, R.: Visual atoms: Pre-training vision transformers with sinusoidal waves. In: Proceedings of the 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 18579-18588 (2023) +46. Tobin, J., Fong, R., Ray, A., Schneider, J., Zaremba, W., Abbeel, P.: Domain randomization for transferring deep neural networks from simulation to the real world. In: Proceedings of the 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 23-30 (2017) +47. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jegou, H.: Training data-efficient image transformers & distillation through attention. In: Proceedings of the 2021 International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 139, pp. 10347-10357 (2021) +48. Veeling, B.S., Linmans, J., Winkens, J., Cohen, T., Welling, M.: Rotation equivariant cnns for digital pathology. In: Proceedings of the 2018 International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI) (2018) +49. Venkataramanan, S., Rizve, M.N., Carreira, J., Asano, Y.M., Avrithis, Y.: Is image-net worth 1 video? learning strong image encoders from 1 long unlabelled video. arXiv preprint arXiv:2310.08584 (2023) +50. Wang, T., Zhu, J., Torralba, A., Efros, A.A.: Dataset Distillation (2018) +51. Yun, S., Han, D., Chun, S., Oh, S.J., Yoo, Y., Choe, J.: CutMix: Regularization strategy to train strong classifiers with localizable features. In: Proceedings of the 2019 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 6022-6031 (2019) +52. Yun-Hao Cao, H.Y., Wu, J.: Training vision transformers with only 2040 images. In: Proceedings of the 2022 European Conference on Computer Vision (ECCV). pp. 220-237 (2022) +53. Zhai, X., Kolesnikov, A., Houlsby, N., Beyer, L.: Scaling vision transformers. Proceedings of 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 1204-1213 (2022) +54. Zhai, X., Puigcerver, J., Kolesnikov, A., Ruyssen, P., Riquelme, C., Lucic, M., Djolonga, J., Pinto, A.S., Neumann, M., Dosovitskiy, A., Beyer, L., Bachem, O., Tschannen, M., Michalski, M., Bousquet, O., Gelly, S., Houlsby, N.: A Large-scale Study of Representation Learning with the Visual Task Adaptation Benchmark. arXiv preprint arXiv:1910:04867 (2019) +55. Zhang, H., Cisse, M., Dauphin, Y.N., Lopez-Paz, D.: Mixup: Beyond empirical risk minimization. arXiv preprint abs/1710.09412 (2017) +56. Zhang, R., Isola, P., Efros, A.A.: Colorful image colorization. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 649-666 (2016) +57. Zhao, B., Mopuri, K.R., Bilen, H.: Dataset condensation with gradient matching. In: Proceedings of the 2021 International Conference on Learning Representation (ICLR) (2021) +58. Zhong, Z., Zheng, L., Kang, G., Li, S., Yang, Y.: Random erasing data augmentation. arXiv preprint abs/1708.04896 (2017) + +59. Zhou, B., Lapedriza, A., Khosla, A., Oliva, A., Torralba, A.: Places: A 10 million image database for scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 40(6), 1452-1464 (2017) \ No newline at end of file diff --git a/2024/Scaling Backwards_ Minimal Synthetic Pre-training_/images.zip b/2024/Scaling Backwards_ Minimal Synthetic Pre-training_/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..82d176c2c2f682dcd7e6388c63469980226d7ace --- /dev/null +++ b/2024/Scaling Backwards_ Minimal Synthetic Pre-training_/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8a351c6799a0c685d75a315f59476ec773354f49600728b271a53c36bbc38221 +size 584372 diff --git a/2024/Scaling Backwards_ Minimal Synthetic Pre-training_/layout.json b/2024/Scaling Backwards_ Minimal Synthetic Pre-training_/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c4223513f8d09461e9777f95d684e4f6764b8ba8 --- /dev/null +++ b/2024/Scaling Backwards_ Minimal Synthetic Pre-training_/layout.json @@ -0,0 +1,10694 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 192, + 112, + 420, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 112, + 420, + 148 + ], + "spans": [ + { + "bbox": [ + 192, + 112, + 420, + 148 + ], + "type": "text", + "content": "Scaling Backwards: Minimal Synthetic Pre-training?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 138, + 168, + 474, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 168, + 474, + 205 + ], + "spans": [ + { + "bbox": [ + 138, + 168, + 474, + 205 + ], + "type": "text", + "content": "Ryo Nakamura" + }, + { + "bbox": [ + 138, + 168, + 474, + 205 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 138, + 168, + 474, + 205 + ], + "type": "text", + "content": ", Ryu Tadokoro" + }, + { + "bbox": [ + 138, + 168, + 474, + 205 + ], + "type": "inline_equation", + "content": "^{2,*}" + }, + { + "bbox": [ + 138, + 168, + 474, + 205 + ], + "type": "text", + "content": ", Ryosuke Yamada" + }, + { + "bbox": [ + 138, + 168, + 474, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 168, + 474, + 205 + ], + "type": "text", + "content": ", Yuki M. Asano" + }, + { + "bbox": [ + 138, + 168, + 474, + 205 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 138, + 168, + 474, + 205 + ], + "type": "text", + "content": ", Iro Laina" + }, + { + "bbox": [ + 138, + 168, + 474, + 205 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 138, + 168, + 474, + 205 + ], + "type": "text", + "content": ", Christian Rupprecht" + }, + { + "bbox": [ + 138, + 168, + 474, + 205 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 138, + 168, + 474, + 205 + ], + "type": "text", + "content": ", Nakamasa Inoue" + }, + { + "bbox": [ + 138, + 168, + 474, + 205 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 138, + 168, + 474, + 205 + ], + "type": "text", + "content": ", Rio Yokota" + }, + { + "bbox": [ + 138, + 168, + 474, + 205 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 138, + 168, + 474, + 205 + ], + "type": "text", + "content": ", and Hirokatsu Kataoka" + }, + { + "bbox": [ + 138, + 168, + 474, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 151, + 213, + 459, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 213, + 459, + 225 + ], + "spans": [ + { + "bbox": [ + 151, + 213, + 459, + 225 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 151, + 213, + 459, + 225 + ], + "type": "text", + "content": " National Institute of Advanced Industrial Science and Technology (AIST)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 264, + 225, + 349, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 225, + 349, + 236 + ], + "spans": [ + { + "bbox": [ + 264, + 225, + 349, + 236 + ], + "type": "text", + "content": "2 Tohoku University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 157, + 236, + 455, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 236, + 455, + 247 + ], + "spans": [ + { + "bbox": [ + 157, + 236, + 455, + 247 + ], + "type": "text", + "content": "3 University of Amsterdam, now at University of Technology Nuremberg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 260, + 247, + 353, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 247, + 353, + 258 + ], + "spans": [ + { + "bbox": [ + 260, + 247, + 353, + 258 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 260, + 247, + 353, + 258 + ], + "type": "text", + "content": " University of Oxford" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 241, + 258, + 372, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 258, + 372, + 269 + ], + "spans": [ + { + "bbox": [ + 241, + 258, + 372, + 269 + ], + "type": "text", + "content": "5 Tokyo Institute of Technology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 159, + 302, + 455, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 302, + 455, + 544 + ], + "spans": [ + { + "bbox": [ + 159, + 302, + 455, + 544 + ], + "type": "text", + "content": "Abstract. Pre-training and transfer learning are an important building block of current computer vision systems. While pre-training is usually performed on large real-world image datasets, in this paper we ask whether this is truly necessary. To this end, we search for a minimal, purely synthetic pre-training dataset that allows us to achieve performance similar to the 1 million images of ImageNet-1k. We construct such a dataset from a single fractal with perturbations. With this, we contribute three main findings. (i) We show that pre-training is effective even with minimal synthetic images, with performance on par with large-scale pre-training datasets like ImageNet-1k for full fine-tuning. (ii) We investigate the single parameter with which we construct artificial categories for our dataset. We find that while the shape differences can be indistinguishable to humans, they are crucial for obtaining strong performances. (iii) Finally, we investigate the minimal requirements for successful pre-training. Surprisingly, we find that a substantial reduction of synthetic images from 1k to 1 can even lead to an increase in pre-training performance, a motivation to further investigate \"scaling backwards\". Finally, we extend our method from synthetic images to real images to see if a single real image can show similar pre-training effect through shape augmentation. We find that the use of grayscale images and affine transformations allows even real images to \"scale backwards\". The code is available at https://github.com/SUPER-TADORY/1p-frac." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 160, + 555, + 453, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 555, + 453, + 566 + ], + "spans": [ + { + "bbox": [ + 160, + 555, + 453, + 566 + ], + "type": "text", + "content": "Keywords: Synthetic pre-training " + }, + { + "bbox": [ + 160, + 555, + 453, + 566 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 555, + 453, + 566 + ], + "type": "text", + "content": " Limited data " + }, + { + "bbox": [ + 160, + 555, + 453, + 566 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 555, + 453, + 566 + ], + "type": "text", + "content": " Vision transformers" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 589, + 230, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 230, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 230, + 601 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 609, + 481, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 609, + 481, + 645 + ], + "spans": [ + { + "bbox": [ + 130, + 609, + 481, + 645 + ], + "type": "text", + "content": "In image recognition, pre-training allows discovering fundamental visual representations for downstream task applications. Pre-training enhances the performance of visual tasks and enables the use of small-scale task-specific datasets." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 654, + 291, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 654, + 291, + 665 + ], + "spans": [ + { + "bbox": [ + 141, + 654, + 291, + 665 + ], + "type": "text", + "content": "* These authors contributed equally" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 116, + 479, + 220 + ], + "blocks": [ + { + "bbox": [ + 137, + 116, + 479, + 220 + ], + "lines": [ + { + "bbox": [ + 137, + 116, + 479, + 220 + ], + "spans": [ + { + "bbox": [ + 137, + 116, + 479, + 220 + ], + "type": "image", + "image_path": "b7a7bdbc988c958c7b3f3b6dca22bfcf87384e4cd9567b3d49c86ed52a1d8076.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 225, + 482, + 302 + ], + "lines": [ + { + "bbox": [ + 130, + 225, + 482, + 302 + ], + "spans": [ + { + "bbox": [ + 130, + 225, + 482, + 302 + ], + "type": "text", + "content": "Fig. 1: Comparison of ImageNet-1k, FractalDB and 1p-frac (ours). 1p-frac consists of only a single fractal for pre-training. With 1p-frac, neural networks learn to classify perturbations applied to the fractal. In our study \"single\" means a very narrow distribution over parameters that leads to images that are roughly equivalent from a human visual perspective. While the shape differences of perturbed images can be indistinguishable to humans, models pre-trained on 1p-frac achieve comparable performance with those pre-trained on ImageNet-1k or FractalDB." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 314, + 480, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 314, + 480, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 480, + 361 + ], + "type": "text", + "content": "Recently, pre-training has been used as a key technology to construct foundation models trained on massive datasets with over hundreds of millions of images. In some cases, a foundation model enables adaptation of zero-shot recognition without the need for additional data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 363, + 482, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 363, + 482, + 470 + ], + "spans": [ + { + "bbox": [ + 130, + 363, + 482, + 470 + ], + "type": "text", + "content": "Pre-training is often interpreted as discovering universal structures in large-scale datasets that later facilitate adaptation to down-stream tasks. In this paper, we challenge this interpretation by providing a minimal pre-training dataset, generated from a single fractal, that achieves similar downstream performance. At the heart of this investigation is the question of whether pre-training might simply be a better weight initialization rather than the discovery of useful visual concepts. If true, performing expensive pre-training with hundreds of millions of images might not be necessary. This additionally frees pre-training from licensing or ethical issues." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 472, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 544 + ], + "type": "text", + "content": "Since the rise of deep neural networks, the ImageNet dataset [16] has been one of the most commonly used pre-training datasets. Originally, pre-training has been conducted through supervised learning (SL) with human-provided teacher labels. However, it has become clear that pre-training can also be achieved without human-provided labels, through self-supervised learning (SSL) [7,9-12,17, 21-23,37,38,56]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 544, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 482, + 640 + ], + "type": "text", + "content": "In this context, Asano et al. [3] successfully acquired visual representations while dramatically reducing the number of required images. They concluded that SSL can yield sufficient image representations even with a single training example, but only for earlier layers of a recognition model. However, it is unclear how these findings translate to modern architectures and representation learning methods. Based on this, vision transformers (ViT) [18] have been pre-trained with only 2,040 real images [52] through an instance discrimination learning signal." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": "More recently, it has been shown that basic visual representations can be acquired without using real images and human-provided labels. The trend of" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 317, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 317, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 317, + 101 + ], + "type": "text", + "content": "R. Nakamura and R. Tadokoro et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "content": "artificially generated labeled images is on the rise for synthetic pre-training [5, 28, 30, 46]. Formula-driven supervised learning (FDSL) generates images from generative formulas, and labels from their parameters [30]. Under the FDSL framework, one can adapt the synthetic pre-training image dataset by altering formulas [24, 27, 29, 44]. While a million-order image dataset was constructed in FractalDB [30], our findings suggest that synthetic pre-training can be reduced to significantly fewer fractal images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 201, + 480, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 201, + 480, + 284 + ], + "spans": [ + { + "bbox": [ + 130, + 201, + 480, + 284 + ], + "type": "text", + "content": "Motivated by these findings, we believe it is possible to find the key essence of pre-training for image recognition. ViT training can be done with as few as 1,000 artificially generated images [35]. Here, we believe that equivalent performance can be achieved with even fewer images. This consideration is undoubtedly important as we approach a minimal synthetic pre-training dataset in image recognition, which goes against the trend of foundation models toward increasing the dataset scale." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 287, + 480, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 287, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 287, + 480, + 335 + ], + "type": "text", + "content": "In the present paper, we thus introduce a minimal synthetic dataset, namely 1-parameter Fractal as Data (1p-frac), which consists of a single fractal as shown in Figure 1, as well as a loss function for pre-training with it. Our contributions regarding minimal synthetic pre-training are as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 337, + 480, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 337, + 480, + 408 + ], + "spans": [ + { + "bbox": [ + 130, + 337, + 480, + 408 + ], + "type": "text", + "content": "Ordinal minimalism: We introduce the locally perturbed cross entropy (LPCE) loss for pre-training with a single fractal. It utilizes perturbed fractal images for training, where neural networks learn to classify small perturbations. In experiments, we demonstrate that pre-training can be performed even with a single fractal. The pre-training effect of 1p-frac is comparable to that of a million-scale labeled image dataset." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 411, + 479, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 479, + 505 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 479, + 505 + ], + "type": "text", + "content": "Distributional minimalism: We introduce the locally integrated empirical (LIEP) distribution " + }, + { + "bbox": [ + 130, + 411, + 479, + 505 + ], + "type": "inline_equation", + "content": "p_{\\Delta}" + }, + { + "bbox": [ + 130, + 411, + 479, + 505 + ], + "type": "text", + "content": " that has a controllable perturbation degree " + }, + { + "bbox": [ + 130, + 411, + 479, + 505 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 130, + 411, + 479, + 505 + ], + "type": "text", + "content": " to investigate the minimal support of the probability density distribution of synthesized images. We observed positive pre-training effects even with a small " + }, + { + "bbox": [ + 130, + 411, + 479, + 505 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 130, + 411, + 479, + 505 + ], + "type": "text", + "content": " producing shape differences that cannot be distinguished by humans. We also show that if " + }, + { + "bbox": [ + 130, + 411, + 479, + 505 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 130, + 411, + 479, + 505 + ], + "type": "text", + "content": " is too small the visual pre-training collapses. From these observations, we establish the general bounds for generating good pre-training images from a mathematical formula." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 508, + 329, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 508, + 329, + 614 + ], + "spans": [ + { + "bbox": [ + 130, + 508, + 329, + 614 + ], + "type": "text", + "content": "Instance minimalism: Based on the experimental results, the synthetic images should not simply contain complex shapes. The use of recursive image patterns, similar to objects in nature, should be applied in visual pretraining. Experiments with augmented categories from a real image have shown that good pre-training effects can be achieved by performing \"affine transformations on edge" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 331, + 559, + 473, + 608 + ], + "blocks": [ + { + "bbox": [ + 334, + 514, + 479, + 559 + ], + "lines": [ + { + "bbox": [ + 334, + 514, + 479, + 559 + ], + "spans": [ + { + "bbox": [ + 334, + 514, + 479, + 559 + ], + "type": "text", + "content": "Table 1: Scaling backwards in synthetic pre-training (Accuracies on CIFAR-100, Real: ImageNet, Synth: Fractal images)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 331, + 559, + 473, + 608 + ], + "lines": [ + { + "bbox": [ + 331, + 559, + 473, + 608 + ], + "spans": [ + { + "bbox": [ + 331, + 559, + 473, + 608 + ], + "type": "table", + "html": "
Type\\#Img11k1M
RealN/A76.985.5
Synth84.284.081.6
", + "image_path": "4eb186fb94b5035971ec5682038b06c0d7eb5b07816e5e3608c0dcab4f097897.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 616, + 479, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 616, + 479, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 616, + 479, + 640 + ], + "type": "text", + "content": "emphasized objects in grayscale\". These operations are found to be almost synonymous with the configuration of the proposed 1p-frac." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "content": "In summary, we significantly reduce the size of the pre-training dataset from 1M images (fractal database (FractalDB) [30]) or 1k images (one-instance fractal" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 370, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 370, + 91, + 447, + 101 + ], + "type": "text", + "content": "Scaling Backwards" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "content": "database (OFDB) [35]) down to 1 and show that this even improves the pretraining effect (Table 1), which motivates \"scaling backwards\"." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 156, + 238, + 170 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 156, + 238, + 170 + ], + "spans": [ + { + "bbox": [ + 132, + 156, + 238, + 170 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 179, + 480, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 179, + 480, + 204 + ], + "spans": [ + { + "bbox": [ + 132, + 179, + 480, + 204 + ], + "type": "text", + "content": "The present study focuses primarily on finding the minimal pre-training dataset with artificially generated images." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 213, + 481, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 213, + 481, + 308 + ], + "spans": [ + { + "bbox": [ + 132, + 213, + 481, + 308 + ], + "type": "text", + "content": "Pre-training in image recognition. Pre-training in image recognition has demonstrated significant success with vast amounts of labeled images. Pre-training helps neural networks to acquire fundamental visual representations which are essential for improving the performance of diverse downstream tasks in image recognition. In particular, researchers have used larger-scale image datasets for supervised learning (SL), starting from million-scale datasets (e.g., ImageNet [16], Places [59]). More recently, the largest datasets used are reaching billion-scale (e.g., IG-3.5B [34], JFT-300M/3B/4B [15, 42, 53])." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 309, + 481, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 309, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 309, + 481, + 380 + ], + "type": "text", + "content": "Pre-training can be also achieved without any human intervention. One of the most effective approaches is self-supervised learning (SSL [7,9-12,17,21-23,37, 38,56]), which has been developed in order to alleviate the manual annotations on large-amount of images. SSL has been studied as an expected alternative to supervised pre-training and has sometimes been shown to surpass supervised pre-training." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 380, + 481, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 441 + ], + "type": "text", + "content": "Despite the development of supervised and self-supervised pre-training, the number of real images used in visual pre-training continues to increase. Exploratory research questions such as \"what is a minimal synthetic pre-training dataset?\" have received less attention. However, they could give us more insights on what is learned during pre-training and how it could be improved." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 449, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 449, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 449, + 481, + 665 + ], + "type": "text", + "content": "Pre-training with limited data. In order to address ethical issues such as privacy and fairness associated with real images, a pre-training with a very limited number of real images and synthetic pre-training without any real images have been on the rise. In this context, synthetic pre-training with copy-and-paste learning [19,39], domain randomization from 3D graphics [43,46], learning from primitive patterns [4,5,41], and dataset distillation [8,50,57] have been recognized as promising approaches to acquiring visual representations. More recently, a sophisticated approach, formula-driven supervised learning (FDSL) [28,30], enabled automatic construction of a labeled image dataset without any real images or human annotations. FDSL can generate images and their corresponding labels using a mathematical formula. Along these lines, One-instance FractalDB (OFDB) has successfully pre-trained a ViT using only 1,000 synthetic images generated from a mathematical formula, without using real images [35]. Moreover, several studies have been reported in the context of extremely limited data [3,49,52]. In particular, Asano et al. have shown that learning visual representations is possible using a single image for self-supervised learning. Venkataramanan et al. achieved a comparable method to ImageNet pre-training with only 10 walking videos and their cropped images." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 317, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 317, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 317, + 102 + ], + "type": "text", + "content": "R. Nakamura and R. Tadokoro et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "By analyzing the acquisition of visual representations through the use of minimal synthetic pre-training datasets, we believe that the FDSL framework will be key to discovering the essence of pre-training mechanisms in image recognition." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 170, + 391, + 184 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 170, + 391, + 184 + ], + "spans": [ + { + "bbox": [ + 132, + 170, + 391, + 184 + ], + "type": "text", + "content": "3 Scaling Backwards with a Single Fractal" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 195, + 482, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 195, + 482, + 350 + ], + "spans": [ + { + "bbox": [ + 130, + 195, + 482, + 350 + ], + "type": "text", + "content": "This section presents the proposed methodology to explore the minimal requirements for successful pre-training by utilizing purely synthetic images of fractals. Specifically, we introduce the 1 parameter Fractal as Data (1p-frac), which consists of only a single fractal, with a method to pre-train neural networks on it. Our key idea is to introduce the locally integrated empirical (LIEP) distribution " + }, + { + "bbox": [ + 130, + 195, + 482, + 350 + ], + "type": "inline_equation", + "content": "p_{\\Delta}" + }, + { + "bbox": [ + 130, + 195, + 482, + 350 + ], + "type": "text", + "content": " over perturbed fractal images, enabling pre-training even with a single fractal image. Since the LIEP distribution is designed so that it converges to the empirical distribution " + }, + { + "bbox": [ + 130, + 195, + 482, + 350 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{data}}(x) = \\delta (x - I)" + }, + { + "bbox": [ + 130, + 195, + 482, + 350 + ], + "type": "text", + "content": " of a single image " + }, + { + "bbox": [ + 130, + 195, + 482, + 350 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 195, + 482, + 350 + ], + "type": "text", + "content": " when the perturbation degree " + }, + { + "bbox": [ + 130, + 195, + 482, + 350 + ], + "type": "inline_equation", + "content": "\\Delta \\in \\mathbb{R}_{\\geq 0}" + }, + { + "bbox": [ + 130, + 195, + 482, + 350 + ], + "type": "text", + "content": " goes to zero, we can narrow down the support of the distribution by decreasing " + }, + { + "bbox": [ + 130, + 195, + 482, + 350 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 130, + 195, + 482, + 350 + ], + "type": "text", + "content": " as shown in Figure 2a. Below, we begin with a preliminary for defining the empirical distributions of two fractal databases, namely FractalDB [30] and OFDB [35]. We then introduce 1p-frac with the LIEP distribution." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 368, + 221, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 368, + 221, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 368, + 221, + 380 + ], + "type": "text", + "content": "3.1 Preliminary" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 387, + 482, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 387, + 482, + 471 + ], + "spans": [ + { + "bbox": [ + 130, + 387, + 482, + 471 + ], + "type": "text", + "content": "FractalDB [30]. Kataoka et al. have introduced a method for effectively pretraining neural networks with FractalDB, a set of fractal images generated by the iterated function systems (IFSs). Specifically, FractalDB " + }, + { + "bbox": [ + 130, + 387, + 482, + 471 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 130, + 387, + 482, + 471 + ], + "type": "text", + "content": " consists of one million synthesized images: " + }, + { + "bbox": [ + 130, + 387, + 482, + 471 + ], + "type": "inline_equation", + "content": "\\mathcal{F} = \\{(\\varOmega_c,\\{I_i^c\\}_{i = 0}^{M - 1})\\}_{c = 0}^{C - 1}" + }, + { + "bbox": [ + 130, + 387, + 482, + 471 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 387, + 482, + 471 + ], + "type": "inline_equation", + "content": "\\varOmega_{c}" + }, + { + "bbox": [ + 130, + 387, + 482, + 471 + ], + "type": "text", + "content": " is an IFS, " + }, + { + "bbox": [ + 130, + 387, + 482, + 471 + ], + "type": "inline_equation", + "content": "I_{i}^{c}" + }, + { + "bbox": [ + 130, + 387, + 482, + 471 + ], + "type": "text", + "content": " is a fractal image generated by " + }, + { + "bbox": [ + 130, + 387, + 482, + 471 + ], + "type": "inline_equation", + "content": "\\varOmega_{c}" + }, + { + "bbox": [ + 130, + 387, + 482, + 471 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 387, + 482, + 471 + ], + "type": "inline_equation", + "content": "C = 1,000" + }, + { + "bbox": [ + 130, + 387, + 482, + 471 + ], + "type": "text", + "content": " is the number of fractal categories, and " + }, + { + "bbox": [ + 130, + 387, + 482, + 471 + ], + "type": "inline_equation", + "content": "M = 1,000" + }, + { + "bbox": [ + 130, + 387, + 482, + 471 + ], + "type": "text", + "content": " is the number of images per category. Each IFS characterizes each fractal category " + }, + { + "bbox": [ + 130, + 387, + 482, + 471 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 387, + 482, + 471 + ], + "type": "text", + "content": " and is defined as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 213, + 480, + 480, + 493 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 213, + 480, + 480, + 493 + ], + "spans": [ + { + "bbox": [ + 213, + 480, + 480, + 493 + ], + "type": "interline_equation", + "content": "\\Omega_ {c} = \\left\\{\\mathbb {R} ^ {2}; w _ {1}, w _ {2}, \\dots , w _ {N _ {c}}; p _ {1}, p _ {2}, \\dots , p _ {N _ {c}} \\right\\}, \\tag {1}", + "image_path": "79a3d63728247cad734bacd3399126e620133ae90cff8df5e27238e02ed47b54.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 502, + 392, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 502, + 392, + 514 + ], + "spans": [ + { + "bbox": [ + 131, + 502, + 392, + 514 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 502, + 392, + 514 + ], + "type": "inline_equation", + "content": "w_{j}:\\mathbb{R}^{2}\\to \\mathbb{R}^{2}" + }, + { + "bbox": [ + 131, + 502, + 392, + 514 + ], + "type": "text", + "content": " is a 2D affine transformation given by" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 228, + 523, + 480, + 549 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 523, + 480, + 549 + ], + "spans": [ + { + "bbox": [ + 228, + 523, + 480, + 549 + ], + "type": "interline_equation", + "content": "w _ {j} (\\boldsymbol {v}) = \\left[ \\begin{array}{l l} a _ {j} & b _ {j} \\\\ c _ {j} & d _ {j} \\end{array} \\right] \\boldsymbol {v} + \\left[ \\begin{array}{l} e _ {j} \\\\ f _ {j} \\end{array} \\right] (\\boldsymbol {v} \\in \\mathbb {R} ^ {2}), \\tag {2}", + "image_path": "870e57671fd0fab954f3536ad6ce14551bb61e666570b0273b6999b2837ef41e.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 556, + 482, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 556, + 482, + 639 + ], + "spans": [ + { + "bbox": [ + 131, + 556, + 482, + 639 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 131, + 556, + 482, + 639 + ], + "type": "inline_equation", + "content": "p_j" + }, + { + "bbox": [ + 131, + 556, + 482, + 639 + ], + "type": "text", + "content": " is a probability mass distribution. Each fractal image " + }, + { + "bbox": [ + 131, + 556, + 482, + 639 + ], + "type": "inline_equation", + "content": "I_i^c" + }, + { + "bbox": [ + 131, + 556, + 482, + 639 + ], + "type": "text", + "content": " renders a fractal " + }, + { + "bbox": [ + 131, + 556, + 482, + 639 + ], + "type": "inline_equation", + "content": "F = \\{\\pmb{v}_t\\}_{t=0}^T \\subset \\mathbb{R}^2" + }, + { + "bbox": [ + 131, + 556, + 482, + 639 + ], + "type": "text", + "content": " into a 2D image, where points " + }, + { + "bbox": [ + 131, + 556, + 482, + 639 + ], + "type": "inline_equation", + "content": "\\pmb{v}_t" + }, + { + "bbox": [ + 131, + 556, + 482, + 639 + ], + "type": "text", + "content": " are determined by a recurrence relation that applies affine transformations as " + }, + { + "bbox": [ + 131, + 556, + 482, + 639 + ], + "type": "inline_equation", + "content": "\\pmb{v}_{t+1} = w_{\\sigma_t}(\\pmb{v}_t)" + }, + { + "bbox": [ + 131, + 556, + 482, + 639 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 131, + 556, + 482, + 639 + ], + "type": "inline_equation", + "content": "t = 0, 1, 2, \\dots, T-1" + }, + { + "bbox": [ + 131, + 556, + 482, + 639 + ], + "type": "text", + "content": ". Here, the initial point is set as " + }, + { + "bbox": [ + 131, + 556, + 482, + 639 + ], + "type": "inline_equation", + "content": "\\pmb{v}_0 = (0,0)^\\top" + }, + { + "bbox": [ + 131, + 556, + 482, + 639 + ], + "type": "text", + "content": ", and the index " + }, + { + "bbox": [ + 131, + 556, + 482, + 639 + ], + "type": "inline_equation", + "content": "\\sigma_t" + }, + { + "bbox": [ + 131, + 556, + 482, + 639 + ], + "type": "text", + "content": " is sampled at each " + }, + { + "bbox": [ + 131, + 556, + 482, + 639 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 556, + 482, + 639 + ], + "type": "text", + "content": " independently from the probability mass distribution " + }, + { + "bbox": [ + 131, + 556, + 482, + 639 + ], + "type": "inline_equation", + "content": "p(\\sigma_t = j) = p_j" + }, + { + "bbox": [ + 131, + 556, + 482, + 639 + ], + "type": "text", + "content": ". Pre-training with FractalDB utilizes the cross-entropy loss function:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 244, + 650, + 480, + 662 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 650, + 480, + 662 + ], + "spans": [ + { + "bbox": [ + 244, + 650, + 480, + 662 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = - \\mathbb {E} _ {x, y \\sim p _ {\\mathrm {d a t a}}} [ \\log p _ {\\theta} (y | x) ], \\tag {3}", + "image_path": "22d51c313d4bee532da532264a4e22719c25a39745c0bbb567160002b08a9a64.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 370, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 370, + 91, + 447, + 101 + ], + "type": "text", + "content": "Scaling Backwards" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 114, + 470, + 244 + ], + "blocks": [ + { + "bbox": [ + 143, + 114, + 470, + 244 + ], + "lines": [ + { + "bbox": [ + 143, + 114, + 470, + 244 + ], + "spans": [ + { + "bbox": [ + 143, + 114, + 470, + 244 + ], + "type": "image", + "image_path": "d1a098d60ba4748ad3a145adc83d138d79f0d6283419d4c7ebafc33e335b120d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 254, + 482, + 309 + ], + "lines": [ + { + "bbox": [ + 130, + 254, + 482, + 309 + ], + "spans": [ + { + "bbox": [ + 130, + 254, + 482, + 309 + ], + "type": "text", + "content": "Fig. 2: Scaling backwards from many images to a single synthetic image. (a) Empirical distribution " + }, + { + "bbox": [ + 130, + 254, + 482, + 309 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{data}}" + }, + { + "bbox": [ + 130, + 254, + 482, + 309 + ], + "type": "text", + "content": ". Colors indicate classes. With a single image, the distribution is given by a single Dirac's delta function. (b) LIEP distribution " + }, + { + "bbox": [ + 130, + 254, + 482, + 309 + ], + "type": "inline_equation", + "content": "p_{\\Delta}" + }, + { + "bbox": [ + 130, + 254, + 482, + 309 + ], + "type": "text", + "content": ". The support of the distribution narrows as the degree of perturbation " + }, + { + "bbox": [ + 130, + 254, + 482, + 309 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 130, + 254, + 482, + 309 + ], + "type": "text", + "content": " decreases. (c) " + }, + { + "bbox": [ + 130, + 254, + 482, + 309 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 254, + 482, + 309 + ], + "type": "text", + "content": "-factor for investigating fractal shapes. A small " + }, + { + "bbox": [ + 130, + 254, + 482, + 309 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 254, + 482, + 309 + ], + "type": "text", + "content": " produces complex fractals." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 316, + 482, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 482, + 351 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 482, + 351 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 316, + 482, + 351 + ], + "type": "inline_equation", + "content": "p_{\\theta}" + }, + { + "bbox": [ + 130, + 316, + 482, + 351 + ], + "type": "text", + "content": " is the category distribution predicted by a neural network and " + }, + { + "bbox": [ + 130, + 316, + 482, + 351 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 316, + 482, + 351 + ], + "type": "text", + "content": " is a set of learnable parameters. The joint empirical distribution " + }, + { + "bbox": [ + 130, + 316, + 482, + 351 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{data}}" + }, + { + "bbox": [ + 130, + 316, + 482, + 351 + ], + "type": "text", + "content": " is defined over the dataset as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 201, + 357, + 481, + 389 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 357, + 481, + 389 + ], + "spans": [ + { + "bbox": [ + 201, + 357, + 481, + 389 + ], + "type": "interline_equation", + "content": "p _ {\\mathrm {d a t a}} (x, y; \\mathcal {F}) = \\frac {1}{M C} \\sum_ {i = 0} ^ {M - 1} \\sum_ {c = 0} ^ {C - 1} \\delta \\left(x - I _ {i} ^ {c}\\right) \\delta (y - c) \\tag {4}", + "image_path": "5dc7036a9f88babd1c6bfeaa9ed3d7f7e5c12d358ce96a45736b03313bfa72d8.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 396, + 482, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 396, + 482, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 482, + 430 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 396, + 482, + 430 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 130, + 396, + 482, + 430 + ], + "type": "text", + "content": " is the Dirac's delta function. Models pre-trained on this dataset perform comparable to those pre-trained on real-world image datasets, such as ImageNet-1k and Places365, in some downstream tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 431, + 483, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 431, + 483, + 467 + ], + "spans": [ + { + "bbox": [ + 130, + 431, + 483, + 467 + ], + "type": "text", + "content": "OFDB [35]. This dataset consists of 1,000 fractal images. Specifically, OFDB " + }, + { + "bbox": [ + 130, + 431, + 483, + 467 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{OF}}" + }, + { + "bbox": [ + 130, + 431, + 483, + 467 + ], + "type": "text", + "content": " involves only one representative image per category, i.e., " + }, + { + "bbox": [ + 130, + 431, + 483, + 467 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{OF}} = \\{\\varOmega_c,I_c\\}_{c = 0}^{C - 1}" + }, + { + "bbox": [ + 130, + 431, + 483, + 467 + ], + "type": "text", + "content": ". Therefore, the joint empirical distribution reduces to" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 210, + 474, + 481, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 210, + 474, + 481, + 506 + ], + "spans": [ + { + "bbox": [ + 210, + 474, + 481, + 506 + ], + "type": "interline_equation", + "content": "p _ {\\mathrm {d a t a}} (x, y; \\mathcal {F} _ {\\mathrm {O F}}) = \\frac {1}{C} \\sum_ {c = 0} ^ {C - 1} \\delta \\left(x - I _ {c}\\right) \\delta \\left(y - c\\right). \\tag {5}", + "image_path": "b56f2bba567693ca1b77a13cb2794f2718b4448152ca18a4dbd02cf1c237d0ee.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 512, + 482, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 512, + 482, + 561 + ], + "spans": [ + { + "bbox": [ + 130, + 512, + 482, + 561 + ], + "type": "text", + "content": "Models pre-trained on this dataset perform comparable to or even better than those pre-trained on FractalDB. This work has shown that there exists a small but essential set of images for visual pre-training. However, reducing the number of fractals " + }, + { + "bbox": [ + 130, + 512, + 482, + 561 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 512, + 482, + 561 + ], + "type": "text", + "content": " to less than 1,000 degrades the performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 576, + 331, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 576, + 331, + 588 + ], + "spans": [ + { + "bbox": [ + 131, + 576, + 331, + 588 + ], + "type": "text", + "content": "3.2 Pre-training with a Single Fractal" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 594, + 482, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 643 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 643 + ], + "type": "text", + "content": "Scaling backwards. To further facilitate the analysis of what images are minimally required for successful visual pre-training, we introduce 1p-frac, which ultimately reduces the number of IFSs and images to one each as " + }, + { + "bbox": [ + 130, + 594, + 482, + 643 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{OP}} = (\\varOmega,I)" + }, + { + "bbox": [ + 130, + 594, + 482, + 643 + ], + "type": "text", + "content": ". With this dataset, the empirical distribution is given by" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 235, + 649, + 481, + 663 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 649, + 481, + 663 + ], + "spans": [ + { + "bbox": [ + 235, + 649, + 481, + 663 + ], + "type": "interline_equation", + "content": "p _ {\\mathrm {d a t a}} (x, y; \\mathcal {F} _ {\\mathrm {O P}}) = \\delta (x - I) \\delta (y). \\tag {6}", + "image_path": "2865bdca4cf53098c84dc203021540cc6f1806b9a54cdae811be79b33d20a7dd.jpg" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 317, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 317, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 317, + 101 + ], + "type": "text", + "content": "R. Nakamura and R. Tadokoro et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": "However, we notice that with this distribution, training neural networks using cross-entropy loss is not straightforward because " + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "p_{\\theta}(y = 0|x) \\equiv 1 (\\forall x)" + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": " gives a trivial solution to the loss minimization problem. To address this, we introduce locally perturbed cross entropy (LPCE) loss " + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\Delta}" + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": ", a variant of the cross entropy loss defined with the LIEP distribution." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 176, + 479, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 176, + 479, + 212 + ], + "spans": [ + { + "bbox": [ + 131, + 176, + 479, + 212 + ], + "type": "text", + "content": "Definition 1. Let " + }, + { + "bbox": [ + 131, + 176, + 479, + 212 + ], + "type": "inline_equation", + "content": "I_{\\epsilon} \\in \\mathcal{X}" + }, + { + "bbox": [ + 131, + 176, + 479, + 212 + ], + "type": "text", + "content": " be a perturbed image, where " + }, + { + "bbox": [ + 131, + 176, + 479, + 212 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 131, + 176, + 479, + 212 + ], + "type": "text", + "content": " is a set of images, " + }, + { + "bbox": [ + 131, + 176, + 479, + 212 + ], + "type": "inline_equation", + "content": "\\epsilon \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 131, + 176, + 479, + 212 + ], + "type": "text", + "content": " is a small perturbation with " + }, + { + "bbox": [ + 131, + 176, + 479, + 212 + ], + "type": "inline_equation", + "content": "d \\in \\mathbb{N}_{>0}" + }, + { + "bbox": [ + 131, + 176, + 479, + 212 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 176, + 479, + 212 + ], + "type": "inline_equation", + "content": "I_0 = I" + }, + { + "bbox": [ + 131, + 176, + 479, + 212 + ], + "type": "text", + "content": " is the original image. We define the LIEP distribution by" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 217, + 220, + 479, + 245 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 220, + 479, + 245 + ], + "spans": [ + { + "bbox": [ + 217, + 220, + 479, + 245 + ], + "type": "interline_equation", + "content": "p _ {\\Delta} (x, y) = \\frac {1}{| \\mathcal {R} _ {\\Delta} |} \\int_ {\\mathcal {R} _ {\\Delta}} \\delta (x - I _ {\\epsilon}) \\delta (y - \\epsilon) d \\epsilon \\tag {7}", + "image_path": "fb7122ba276c2ca42a0ecb2fa39b7d590ee0f0ffe18bcd315496ddb428a44fef.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 254, + 480, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 254, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 131, + 254, + 480, + 277 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 254, + 480, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{\\Delta} \\subset \\mathbb{R}^{d}" + }, + { + "bbox": [ + 131, + 254, + 480, + 277 + ], + "type": "text", + "content": " is a compact set containing the origin and " + }, + { + "bbox": [ + 131, + 254, + 480, + 277 + ], + "type": "inline_equation", + "content": "|\\mathcal{R}_{\\Delta}|" + }, + { + "bbox": [ + 131, + 254, + 480, + 277 + ], + "type": "text", + "content": " is its volume with order " + }, + { + "bbox": [ + 131, + 254, + 480, + 277 + ], + "type": "inline_equation", + "content": "O(|\\Delta|^{d})" + }, + { + "bbox": [ + 131, + 254, + 480, + 277 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 277, + 329, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 277, + 329, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 277, + 329, + 289 + ], + "type": "text", + "content": "Definition 2. We define the LPCE loss by" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 244, + 300, + 479, + 312 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 300, + 479, + 312 + ], + "spans": [ + { + "bbox": [ + 244, + 300, + 479, + 312 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\Delta} = - \\mathbb {E} _ {x, y \\sim p _ {\\Delta}} \\left[ \\log p _ {\\theta} (y | x) \\right], \\tag {8}", + "image_path": "edbbe4627687c35414ff9853834eb8d68f6f498c4c58b8be03867ef7c16a6754.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 319, + 285, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 319, + 285, + 331 + ], + "spans": [ + { + "bbox": [ + 131, + 319, + 285, + 331 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 319, + 285, + 331 + ], + "type": "inline_equation", + "content": "p_{\\Delta}" + }, + { + "bbox": [ + 131, + 319, + 285, + 331 + ], + "type": "text", + "content": " is the LIEP distribution." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 331, + 479, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 479, + 378 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 479, + 378 + ], + "type": "text", + "content": "If " + }, + { + "bbox": [ + 130, + 331, + 479, + 378 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{\\Delta}" + }, + { + "bbox": [ + 130, + 331, + 479, + 378 + ], + "type": "text", + "content": " is a small hypercube or hypersphere, it is straightforward to see that " + }, + { + "bbox": [ + 130, + 331, + 479, + 378 + ], + "type": "inline_equation", + "content": "p_{\\Delta}" + }, + { + "bbox": [ + 130, + 331, + 479, + 378 + ], + "type": "text", + "content": " approaches to the empirical distribution of Eq. (6) when " + }, + { + "bbox": [ + 130, + 331, + 479, + 378 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 130, + 331, + 479, + 378 + ], + "type": "text", + "content": " goes to zero. Therefore, this loss allows us to analyze visual pre-training effects by narrowing the support of the distribution around a single image." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 380, + 479, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 479, + 414 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 479, + 414 + ], + "type": "text", + "content": "With 1p-frac, we apply perturbation to the affine transformations. As such, the perturbation " + }, + { + "bbox": [ + 130, + 380, + 479, + 414 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 380, + 479, + 414 + ], + "type": "text", + "content": " is in " + }, + { + "bbox": [ + 130, + 380, + 479, + 414 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{6*j}" + }, + { + "bbox": [ + 130, + 380, + 479, + 414 + ], + "type": "text", + "content": ". A perturbed image " + }, + { + "bbox": [ + 130, + 380, + 479, + 414 + ], + "type": "inline_equation", + "content": "I_{\\epsilon}" + }, + { + "bbox": [ + 130, + 380, + 479, + 414 + ], + "type": "text", + "content": " is obtained by the noisy affine transformations:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 231, + 422, + 479, + 447 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 422, + 479, + 447 + ], + "spans": [ + { + "bbox": [ + 231, + 422, + 479, + 447 + ], + "type": "interline_equation", + "content": "w _ {j} \\left(\\boldsymbol {v}; \\boldsymbol {\\epsilon} _ {j}\\right) = \\left(\\left[ \\begin{array}{l l l} a _ {j} & b _ {j} & e _ {j} \\\\ c _ {j} & d _ {j} & f _ {j} \\end{array} \\right] + \\boldsymbol {\\epsilon} _ {j}\\right) \\left[ \\begin{array}{l} \\boldsymbol {v} \\\\ 1 \\end{array} \\right] \\tag {9}", + "image_path": "48986cd75a1343170935b79660df7bc3d4866b02a74528194b715ca55f91ff02.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 455, + 479, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 455, + 479, + 502 + ], + "spans": [ + { + "bbox": [ + 131, + 455, + 479, + 502 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 455, + 479, + 502 + ], + "type": "inline_equation", + "content": "\\epsilon_{j} \\in \\mathcal{R}_{\\Delta} = [-\\Delta / 2, \\Delta / 2]^{6 * j} \\subset \\mathbb{R}^{6 * j}" + }, + { + "bbox": [ + 131, + 455, + 479, + 502 + ], + "type": "text", + "content": " is a hypercube with a side length " + }, + { + "bbox": [ + 131, + 455, + 479, + 502 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 131, + 455, + 479, + 502 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 455, + 479, + 502 + ], + "type": "inline_equation", + "content": "|\\mathcal{R}_{\\Delta}| = \\Delta^{6 * j}" + }, + { + "bbox": [ + 131, + 455, + 479, + 502 + ], + "type": "text", + "content": ". Note that the numerical integration of Eq. (7) is used in practice, which is an approximation obtained by uniformly sampling " + }, + { + "bbox": [ + 131, + 455, + 479, + 502 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 131, + 455, + 479, + 502 + ], + "type": "text", + "content": " points in " + }, + { + "bbox": [ + 131, + 455, + 479, + 502 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{\\Delta}" + }, + { + "bbox": [ + 131, + 455, + 479, + 502 + ], + "type": "text", + "content": ", where we set " + }, + { + "bbox": [ + 131, + 455, + 479, + 502 + ], + "type": "inline_equation", + "content": "L = 1,000" + }, + { + "bbox": [ + 131, + 455, + 479, + 502 + ], + "type": "text", + "content": " by default." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 503, + 479, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 503, + 479, + 550 + ], + "spans": [ + { + "bbox": [ + 130, + 503, + 479, + 550 + ], + "type": "text", + "content": "Visualization. Figure 2b shows examples of perturbed images to compute the LPCE loss. While most of shape differences are indistinguishable to humans, a neural network learns to distinguish the perturbations applied to the single image by minimizing the LPCE loss." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 552, + 479, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 552, + 479, + 586 + ], + "spans": [ + { + "bbox": [ + 130, + 552, + 479, + 586 + ], + "type": "text", + "content": "Complexity of " + }, + { + "bbox": [ + 130, + 552, + 479, + 586 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 130, + 552, + 479, + 586 + ], + "type": "text", + "content": ". We use the " + }, + { + "bbox": [ + 130, + 552, + 479, + 586 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 552, + 479, + 586 + ], + "type": "text", + "content": "-factor proposed by Anderson et al. [1] to evaluate the complexity of IFSs. As shown in Figure 2c, small values of " + }, + { + "bbox": [ + 130, + 552, + 479, + 586 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 552, + 479, + 586 + ], + "type": "text", + "content": " produce complex fractal shapes." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 605, + 229, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 605, + 229, + 618 + ], + "spans": [ + { + "bbox": [ + 131, + 605, + 229, + 618 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "content": "Experimental setup. What is minimal visual pre-training, and what properties of images lend themselves to reducing their count? In order to clarify these questions, we conduct experiments following these steps:" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 370, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 370, + 91, + 447, + 101 + ], + "type": "text", + "content": "Scaling Backwards" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 270 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 138, + 116, + 479, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 116, + 479, + 162 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 479, + 162 + ], + "type": "text", + "content": "- Exploration study (Tables 2-4): We verify whether a single fractal is sufficient for pre-training. Moreover, we compare the extent to which the shapes used in pre-training can be simplified, using object contours, Gaussian, and uniform distributions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 163, + 480, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 163, + 480, + 186 + ], + "spans": [ + { + "bbox": [ + 138, + 163, + 480, + 186 + ], + "type": "text", + "content": "- Hyperparameter study (Tables 5-7): We investigate the effects of three hyperparameters " + }, + { + "bbox": [ + 138, + 163, + 480, + 186 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 138, + 163, + 480, + 186 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 138, + 163, + 480, + 186 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 138, + 163, + 480, + 186 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 138, + 163, + 480, + 186 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 138, + 163, + 480, + 186 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 188, + 480, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 188, + 480, + 210 + ], + "spans": [ + { + "bbox": [ + 138, + 188, + 480, + 210 + ], + "type": "text", + "content": "- Scaling study (Table 8): We compare 1p-frac with other large-scale datasets in terms of fine-tuning accuracy on ImageNet-1k." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 213, + 480, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 213, + 480, + 246 + ], + "spans": [ + { + "bbox": [ + 138, + 213, + 480, + 246 + ], + "type": "text", + "content": "- Analysis and discussion (Tables 9-12): We discuss data augmentation, computational cost for synthesizing images, and pre-training using a single real-world image." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 247, + 479, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 247, + 479, + 270 + ], + "spans": [ + { + "bbox": [ + 138, + 247, + 479, + 270 + ], + "type": "text", + "content": "- Applications (Tables 13 and 14): We conduct comparisons in various fine-tuning datasets." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 279, + 481, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 279, + 481, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 279, + 481, + 411 + ], + "type": "text", + "content": "Implementation details. In order to verify the effects of pre-training, we measure the accuracy of downstream tasks through fine-tuning. We use ViT [18] for all experiments. Specifically, we employ ViT-Tiny (ViT-T) in the exploration and hyperparameter studies, and employ ViT-Base (ViT-B) for the scaling study. The pre-training parameters and data augmentation methods for pre-training and fine-tuning follow conventional methods in OFDB. In OFDB, data augmentation is based on DeiT. For exploration studies, CIFAR-100 (C100) [32] and/or ImageNet-100 (IN100) [40] are used as fine-tuning datasets. In the comparisons, we assign seven representative datasets used in OFDB paper, namely C10/C100, Cars [31], Flowers [36], Pascal VOC 2012 (VOC12) [20], Places-30 (P30) [59], IN100. See supplementary material for more detailed experimental settings." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 430, + 254, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 254, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 254, + 441 + ], + "type": "text", + "content": "4.1 Exploration Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 449, + 479, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 479, + 507 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 479, + 507 + ], + "type": "text", + "content": "Pre-training with 1p-frac. Table 2 compares 1p-frac with FracdalDB (1M images, 1k categories) and OFDB (1k images, 1k categories). As can be seen, the pre-training effect of 1p-frac is comparable to that of FractalDB and OFDB. The number of fractal categories is significantly reduced from " + }, + { + "bbox": [ + 130, + 449, + 479, + 507 + ], + "type": "inline_equation", + "content": "C = 1,000" + }, + { + "bbox": [ + 130, + 449, + 479, + 507 + ], + "type": "text", + "content": " to only 1. This shows the effectiveness and efficiency of 1p-frac and the LPCE loss." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 510, + 480, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 480, + 602 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 480, + 602 + ], + "type": "text", + "content": "Comparison with SOTA FDSL datasets. Table 3 compares 1p-frac with two SOTA FDSL datasets: (1) Radial Contour Database (RCDB) [28], which consists of 1 million images of synthetic polygons, and (2) Visual Atoms (VA) [45], which consists of 1 million images of parameterized wave functions. The \"1K\" counterparts in the table reduce the number of images per category to one. We also applied the LPCE loss to these two datasets by applying perturbations to the radios of a polygon or wave. As can be seen, 1p-frac performs the best when a single image is used for pre-training and outperforms the 1K FDSL datasets." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 605, + 479, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 479, + 663 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 479, + 663 + ], + "type": "text", + "content": "Pre-training with a noisy image. To validate the necessity of a complex shape when pre-training with a single image, we applied the LPCE loss to two noisy images: an image of Gaussian noise and an image of uniform noise. Since these two noise images are generated by determining the parameters of the parametric distributions, Gaussian and uniform, the LPCE loss can be computed by" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 317, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 317, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 317, + 101 + ], + "type": "text", + "content": "R. Nakamura and R. Tadokoro et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 190, + 249, + 251 + ], + "blocks": [ + { + "bbox": [ + 133, + 123, + 248, + 189 + ], + "lines": [ + { + "bbox": [ + 133, + 123, + 248, + 189 + ], + "spans": [ + { + "bbox": [ + 133, + 123, + 248, + 189 + ], + "type": "text", + "content": "Table 2: Comparison of 1p-frac with FractalDB [30] and OFDB [35]. " + }, + { + "bbox": [ + 133, + 123, + 248, + 189 + ], + "type": "inline_equation", + "content": "\\diamond" + }, + { + "bbox": [ + 133, + 123, + 248, + 189 + ], + "type": "text", + "content": " indicates the use of LPCE loss. We report top-1 accuracies " + }, + { + "bbox": [ + 133, + 123, + 248, + 189 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 133, + 123, + 248, + 189 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 190, + 249, + 251 + ], + "lines": [ + { + "bbox": [ + 136, + 190, + 249, + 251 + ], + "spans": [ + { + "bbox": [ + 136, + 190, + 249, + 251 + ], + "type": "table", + "html": "
Method#ImgC100IN100
Scratch-64.274.9
FractalDB1M81.688.5
OFDB1k84.088.6
1p-frac°184.289.0
", + "image_path": "2840e1eb4603019060f307471709a71fb8b07045f52991440f009c9c509e9c10.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 255, + 160, + 364, + 259 + ], + "blocks": [ + { + "bbox": [ + 252, + 114, + 366, + 157 + ], + "lines": [ + { + "bbox": [ + 252, + 114, + 366, + 157 + ], + "spans": [ + { + "bbox": [ + 252, + 114, + 366, + 157 + ], + "type": "text", + "content": "Table 3: Comparison with SOTA FDSL datasets. " + }, + { + "bbox": [ + 252, + 114, + 366, + 157 + ], + "type": "inline_equation", + "content": "\\diamond" + }, + { + "bbox": [ + 252, + 114, + 366, + 157 + ], + "type": "text", + "content": " indicates the use of LPCE loss." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 255, + 160, + 364, + 259 + ], + "lines": [ + { + "bbox": [ + 255, + 160, + 364, + 259 + ], + "spans": [ + { + "bbox": [ + 255, + 160, + 364, + 259 + ], + "type": "table", + "html": "
Method#ImgC100IN100
RCDB1M81.688.5
RCDB1K80.487.5
RCDB°182.587.9
VA1M84.990.3
VA1K82.188.5
VA°182.688.0
ip-frac°184.289.0
", + "image_path": "1d4a393406fe10db7573100ff361e25f7540c0bb4473864e3829d9db2fa007fa.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 375, + 155, + 472, + 255 + ], + "blocks": [ + { + "bbox": [ + 370, + 120, + 477, + 153 + ], + "lines": [ + { + "bbox": [ + 370, + 120, + 477, + 153 + ], + "spans": [ + { + "bbox": [ + 370, + 120, + 477, + 153 + ], + "type": "text", + "content": "Table 4: Comparison with pre-training with a single noise image." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 375, + 155, + 472, + 255 + ], + "lines": [ + { + "bbox": [ + 375, + 155, + 472, + 255 + ], + "spans": [ + { + "bbox": [ + 375, + 155, + 472, + 255 + ], + "type": "table", + "html": "
MethodC100IN100
Gaussian1.15.7
Uniform2.071.1
1p-frac84.289.0
", + "image_path": "ede00c18282aef3d6296d16bcc6fe6ea14332eb7b05a8f076778efa7dfec3b6f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 135, + 290, + 246, + 376 + ], + "blocks": [ + { + "bbox": [ + 134, + 263, + 359, + 286 + ], + "lines": [ + { + "bbox": [ + 134, + 263, + 359, + 286 + ], + "spans": [ + { + "bbox": [ + 134, + 263, + 359, + 286 + ], + "type": "text", + "content": "Table 5: Effects of pertur- Table 6: Effects of " + }, + { + "bbox": [ + 134, + 263, + 359, + 286 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 134, + 263, + 359, + 286 + ], + "type": "text", + "content": " - bation degree " + }, + { + "bbox": [ + 134, + 263, + 359, + 286 + ], + "type": "inline_equation", + "content": "\\varDelta" + }, + { + "bbox": [ + 134, + 263, + 359, + 286 + ], + "type": "inline_equation", + "content": "(\\sigma=3.5)" + }, + { + "bbox": [ + 134, + 263, + 359, + 286 + ], + "type": "inline_equation", + "content": "(\\varDelta=0.1)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 290, + 246, + 376 + ], + "lines": [ + { + "bbox": [ + 135, + 290, + 246, + 376 + ], + "spans": [ + { + "bbox": [ + 135, + 290, + 246, + 376 + ], + "type": "table", + "html": "
ΔC100IN100
0.0011.21.9
0.0119.961.8
0.0583.088.2
0.184.289.0
0.283.488.5
1.082.688.1
", + "image_path": "8467e1efaeaf1360a228aaf6f3487d9ee22af2da548f6616f43307edef5e49bf.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 254, + 296, + 350, + 376 + ], + "blocks": [ + { + "bbox": [ + 254, + 296, + 350, + 376 + ], + "lines": [ + { + "bbox": [ + 254, + 296, + 350, + 376 + ], + "spans": [ + { + "bbox": [ + 254, + 296, + 350, + 376 + ], + "type": "table", + "html": "
σC100IN100
6.081.386.3
5.082.287.1
4.581.986.9
3.584.289.0
4.082.887.9
", + "image_path": "c7295f7d4050335ab7d41ae39d91a08104122492bfc63457598fe6adb067b7a6.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 366, + 308, + 469, + 375 + ], + "blocks": [ + { + "bbox": [ + 361, + 264, + 476, + 297 + ], + "lines": [ + { + "bbox": [ + 361, + 264, + 476, + 297 + ], + "spans": [ + { + "bbox": [ + 361, + 264, + 476, + 297 + ], + "type": "text", + "content": "Table 7: Effects of number of sampling points " + }, + { + "bbox": [ + 361, + 264, + 476, + 297 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 361, + 264, + 476, + 297 + ], + "type": "text", + "content": " for numerical integration." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 366, + 308, + 469, + 375 + ], + "lines": [ + { + "bbox": [ + 366, + 308, + 469, + 375 + ], + "spans": [ + { + "bbox": [ + 366, + 308, + 469, + 375 + ], + "type": "table", + "html": "
LC100IN100
1678.785.3
6482.388.0
51282.488.0
100084.289.0
", + "image_path": "b591faa3834f12288257d1dfee85354e25db11321e86e22abf7936938e12990a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 391, + 482, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 391, + 482, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 482, + 464 + ], + "type": "text", + "content": "giving small perturbations to the parameters. Table 4 shows these images and fine-tuning results. For example, here the " + }, + { + "bbox": [ + 130, + 391, + 482, + 464 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 391, + 482, + 464 + ], + "type": "text", + "content": " is added to the mean vector and covariance matrix of the Gaussian distribution. The results revealed that noise images fail in visual pre-training, performing worse than training from scratch. This suggests that acquiring fundamental visual representations through pretraining requires an image of a certain structured object, such as a fractal." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 484, + 277, + 497 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 484, + 277, + 497 + ], + "spans": [ + { + "bbox": [ + 132, + 484, + 277, + 497 + ], + "type": "text", + "content": "4.2 Hyperparameter Study" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 508, + 482, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 508, + 482, + 580 + ], + "spans": [ + { + "bbox": [ + 130, + 508, + 482, + 580 + ], + "type": "text", + "content": "Perturbation degree. Table 5 shows the results obtained by different perturbation degrees " + }, + { + "bbox": [ + 130, + 508, + 482, + 580 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 130, + 508, + 482, + 580 + ], + "type": "text", + "content": " for the LIEP distribution. The results indicate that setting the value to 0.1 yields the highest performance. Interestingly, the performance significantly improved when " + }, + { + "bbox": [ + 130, + 508, + 482, + 580 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 130, + 508, + 482, + 580 + ], + "type": "text", + "content": " was increased from 0.01 to 0.05. This suggests that the support of the empirical distribution must have a certain size to obtain positive pre-training effects." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 581, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 581, + 482, + 640 + ], + "type": "text", + "content": "-factor. Table 6 investigates the effects of the " + }, + { + "bbox": [ + 130, + 581, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 581, + 482, + 640 + ], + "type": "text", + "content": "-factor, which measures the complexity of fractal shapes. As can be seen, the most complex shape yields the highest performance. It is worth noting that, even with a fractal of " + }, + { + "bbox": [ + 130, + 581, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\sigma = 6.0" + }, + { + "bbox": [ + 130, + 581, + 482, + 640 + ], + "type": "text", + "content": ", which looks like Gaussian noise, we observed positive pre-training effects. This suggests that the shapes obtained from IFS are crucial for pre-training." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Numerical integration. Tables 7 reduces the number of sampling points " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": " of the numerical integration when approximately computing loss with Eq. (7)." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 370, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 370, + 91, + 447, + 101 + ], + "type": "text", + "content": "Scaling Backwards" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 175, + 321, + 297 + ], + "blocks": [ + { + "bbox": [ + 133, + 117, + 324, + 173 + ], + "lines": [ + { + "bbox": [ + 133, + 117, + 324, + 173 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 324, + 173 + ], + "type": "text", + "content": "Table 8: Comparison of pre-training datasets on ImageNet-1k fine-tuning. Fine-tuning accuracies calculated by using ViT-Base (B) are listed in the table. 21k indicates the number of categories in each pre-training dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 175, + 321, + 297 + ], + "lines": [ + { + "bbox": [ + 134, + 175, + 321, + 297 + ], + "spans": [ + { + "bbox": [ + 134, + 175, + 321, + 297 + ], + "type": "table", + "html": "
Pre-training#ImgTypeViT-B
Scratch--79.8
ImageNet-21k14MSL81.8
FractalDB-21k21MFDSL81.8
ExFractalDB-21k21MFDSL82.7
RCDB-21k21MFDSL82.4
VA-21k21MFDSL82.7
OFDB-21k21kFDSL82.2
3D-OFDB-21k21kFDSL82.7
1p-frac (ours)1FDSL82.1
", + "image_path": "7d47bd1f00a2430aeb3072df717c7900cb30dc88d21ae33966c860919401f33c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 334, + 159, + 476, + 297 + ], + "blocks": [ + { + "bbox": [ + 332, + 114, + 476, + 157 + ], + "lines": [ + { + "bbox": [ + 332, + 114, + 476, + 157 + ], + "spans": [ + { + "bbox": [ + 332, + 114, + 476, + 157 + ], + "type": "text", + "content": "Table 9: Ablation study on data augmentation methods. The DeiT augmentation [47] is used as a baseline." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 334, + 159, + 476, + 297 + ], + "lines": [ + { + "bbox": [ + 334, + 159, + 476, + 297 + ], + "spans": [ + { + "bbox": [ + 334, + 159, + 476, + 297 + ], + "type": "table", + "html": "
MethodC100
Baseline84.2
w/o Random Aug. [14]83.4
w/o Random Crop80.1
w/o Rand Aspect83.6
w/o Rand Erasing [58]84.0
w/o Mixup [55]83.6
w/o Cutmix [51]83.3
w/o Mixup/Cutmix80.4
w/o Flipping84.3
w/o Color Jittering84.2
", + "image_path": "6456c106a50203e0a65bb61fb8bd2b00f4ffd4354646fc23fa82abf6e0c309c2.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 312, + 481, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 312, + 481, + 338 + ], + "spans": [ + { + "bbox": [ + 132, + 312, + 481, + 338 + ], + "type": "text", + "content": "We see that larger numbers result in high performance. Interestingly, even with " + }, + { + "bbox": [ + 132, + 312, + 481, + 338 + ], + "type": "inline_equation", + "content": "L = 16" + }, + { + "bbox": [ + 132, + 312, + 481, + 338 + ], + "type": "text", + "content": ", the pre-training effect was positive (better than training from scratch)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 354, + 230, + 368 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 230, + 368 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 230, + 368 + ], + "type": "text", + "content": "4.3 Scaling Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 375, + 482, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 375, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 375, + 482, + 495 + ], + "type": "text", + "content": "ImageNet-1k fine-tuning (Table 8). While this work discusses minimal requirements for visual pre-training, increasing the approximation precision of the numerical integration could further improve the performance and benefit training of large models. In Table 8, we applied the LPCE loss with " + }, + { + "bbox": [ + 130, + 375, + 482, + 495 + ], + "type": "inline_equation", + "content": "L = 21,000" + }, + { + "bbox": [ + 130, + 375, + 482, + 495 + ], + "type": "text", + "content": " to the ViT-B model and compare fine-tuning accuracy on ImageNet-1k. Surprisingly, and despite using a single fractal as data, our ViT-B pre-trained on 1p-frac outperforms pre-training on ImageNet-21k. This shows the strength of our approach in particular because we simply keep the same fine-tuning protocols as used for the original Imagenet-21k to ImageNet-1k transfer proposed in [18], putting us at a potential disadvantage." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 514, + 282, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 282, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 282, + 525 + ], + "type": "text", + "content": "4.4 Analysis and Discussion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Relationship with data augmentation (Table 9). We consider that the variation in image patterns also depends on data augmentation. While traditional methods primarily rely on the data augmentation techniques described in the DeiT paper [47], we investigated which data augmentation methods contribute significantly by excluding one or two techniques from the base data augmentation methods and observing the changes in fine-tuning accuracy on C100 dataset. The experimental results in Table 9 reveal that excluding Random Cropping (RandomCrop) significantly deteriorates accuracy, suggesting that omitting parts of the image region and causing partial loss facilitates the acquisition of the ability to focus on any part without relying on grasping the overall shape, greatly reducing the pre-training effect. Furthermore, when both Mixup and Cutmix are" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 317, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 317, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 317, + 102 + ], + "type": "text", + "content": "R. Nakamura and R. Tadokoro et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 137, + 231, + 274, + 286 + ], + "blocks": [ + { + "bbox": [ + 133, + 126, + 275, + 224 + ], + "lines": [ + { + "bbox": [ + 133, + 126, + 275, + 224 + ], + "spans": [ + { + "bbox": [ + 133, + 126, + 275, + 224 + ], + "type": "text", + "content": "Table 10: Dataset construction time (hours). FractalDB (1M images), OFDB (1k images) and 1p-frac (1k perturbations for one image) are compared. The processing time is shown separately for fractal category search (Search), image rendering (Render), and total time (Total)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 137, + 231, + 274, + 286 + ], + "lines": [ + { + "bbox": [ + 137, + 231, + 274, + 286 + ], + "spans": [ + { + "bbox": [ + 137, + 231, + 274, + 286 + ], + "type": "table", + "html": "
DatasetSearchRenderTotal
FractalDB2.3716.8619.23
OFDB2.370.412.78
1p-frac0.00220.0360.0382
", + "image_path": "e927b5c5bce39f41cc261afcd0d7e44c4cf4d3de29f6d647d4aee7fba64569e8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 288, + 118, + 467, + 233 + ], + "blocks": [ + { + "bbox": [ + 288, + 118, + 467, + 233 + ], + "lines": [ + { + "bbox": [ + 288, + 118, + 467, + 233 + ], + "spans": [ + { + "bbox": [ + 288, + 118, + 467, + 233 + ], + "type": "image", + "image_path": "4c0c45ed6ad09ba8ef1a6b3811581f7e7e5eefd229e320905c14f902b1e7bd77.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 281, + 235, + 476, + 289 + ], + "lines": [ + { + "bbox": [ + 281, + 235, + 476, + 289 + ], + "spans": [ + { + "bbox": [ + 281, + 235, + 476, + 289 + ], + "type": "text", + "content": "Fig. 3: Linear probing following the experiments from [3]. We verified the performance with " + }, + { + "bbox": [ + 281, + 235, + 476, + 289 + ], + "type": "inline_equation", + "content": "\\{1,2,3,4,6\\}" + }, + { + "bbox": [ + 281, + 235, + 476, + 289 + ], + "type": "text", + "content": " layers in the ViT-Tiny model on C100 dataset. Here, 2 and 4 show the use of up to 1-2 and 1-4 layers." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 302, + 479, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 302, + 479, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 302, + 479, + 326 + ], + "type": "text", + "content": "excluded, a degradation in accuracy is observed, indicating that mixing images and categories enables better feature recognition." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 327, + 480, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 480, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 480, + 411 + ], + "type": "text", + "content": "Dataset construction times (Table 10). We calculated the processing times in fractal category search and image rendering. By comparing to the conventional approaches, our 1p-frac recorded much faster category search (0.0022 hours nearly equals to 8 seconds) and image rendering (0.036 hours nearly equals to 129 seconds). 1p-frac requires only one category with a parameter set. This very simple procedure leads to efficiently construct a minimum requirement for pre-training in image recognition." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 412, + 480, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 412, + 480, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 412, + 480, + 509 + ], + "type": "text", + "content": "Linear probing (Figure 3). To compare supervised ImageNet-1k pre-training with our 1p-frac pre-training in the context of early layer visual representations, we executed a linear probing experiment on the C100 dataset. In the ViT-T model with " + }, + { + "bbox": [ + 130, + 412, + 480, + 509 + ], + "type": "inline_equation", + "content": "\\{1,2,3\\}" + }, + { + "bbox": [ + 130, + 412, + 480, + 509 + ], + "type": "text", + "content": " layers, the 1p-frac pre-trained model outperformed the model pre-trained on ImageNet-1k. This demonstrates the superior quality of early, low-level representations learned from our dataset compared to ImageNet. In effect, layers up to layer 3 in ViT-T can be sufficiently pre-trained and frozen using our 1p-frac." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 510, + 479, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 510, + 479, + 521 + ], + "spans": [ + { + "bbox": [ + 132, + 510, + 479, + 521 + ], + "type": "text", + "content": "Pre-training with a single real image and shape augmentation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 522, + 263, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 263, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 263, + 617 + ], + "type": "text", + "content": "It was found that pretraining effects comparable to a dataset of one million real-world images can be achieved with 1p-frac, but is it possible to perform pre-training similarly with a real image? In this section, we verify whether" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 273, + 523, + 323, + 586 + ], + "blocks": [ + { + "bbox": [ + 273, + 523, + 323, + 586 + ], + "lines": [ + { + "bbox": [ + 273, + 523, + 323, + 586 + ], + "spans": [ + { + "bbox": [ + 273, + 523, + 323, + 586 + ], + "type": "image", + "image_path": "0bd2c3b5a3be8f5ed962b45e2d63c6123d606f9cc6805f4303e57c9cc3a5b56e.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 270, + 594, + 480, + 616 + ], + "lines": [ + { + "bbox": [ + 270, + 594, + 480, + 616 + ], + "spans": [ + { + "bbox": [ + 270, + 594, + 480, + 616 + ], + "type": "text", + "content": "Fig. 4: Shape augmentation with three geometry transformations from a single real image." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 325, + 524, + 375, + 584 + ], + "blocks": [ + { + "bbox": [ + 325, + 524, + 375, + 584 + ], + "lines": [ + { + "bbox": [ + 325, + 524, + 375, + 584 + ], + "spans": [ + { + "bbox": [ + 325, + 524, + 375, + 584 + ], + "type": "image", + "image_path": "5ac48b223ad66e29540d7a9c91e7e54528fc10e51f617095edc6497d40f3eb62.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 377, + 524, + 427, + 584 + ], + "blocks": [ + { + "bbox": [ + 377, + 524, + 427, + 584 + ], + "lines": [ + { + "bbox": [ + 377, + 524, + 427, + 584 + ], + "spans": [ + { + "bbox": [ + 377, + 524, + 427, + 584 + ], + "type": "image", + "image_path": "d7505d95cded1adba0c183a93544a96bf82273c027b02e0831ebbce2b5d9d17c.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 429, + 524, + 480, + 584 + ], + "blocks": [ + { + "bbox": [ + 429, + 524, + 480, + 584 + ], + "lines": [ + { + "bbox": [ + 429, + 524, + 480, + 584 + ], + "spans": [ + { + "bbox": [ + 429, + 524, + 480, + 584 + ], + "type": "image", + "image_path": "ecece888f4e4acfba674c84de78801f5c23d4a1bfb0f7bba46b793bd395af5b1.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 617, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 480, + 666 + ], + "type": "text", + "content": "visual features can be acquired through pre-training using a single real image (each image in Figure 5) with the LPCE loss using image representations similar to FDSL (Canny [6] images) and geometric transformations for perturbation (\\{affine, elastic, polynomial\\} transformations in Figure 4). The results of pre" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 370, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 370, + 91, + 447, + 101 + ], + "type": "text", + "content": "Scaling Backwards" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 116, + 481, + 161 + ], + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "image", + "image_path": "858924f9b8dbfdcc015dda5ccb40b1d07b941fd790e07d2a5f52d7101df9d932.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 164, + 480, + 187 + ], + "lines": [ + { + "bbox": [ + 130, + 164, + 480, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 480, + 187 + ], + "type": "text", + "content": "Fig. 5: The five different images used in the pre-training with a single real image and shape augmentation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 140, + 258, + 291, + 305 + ], + "blocks": [ + { + "bbox": [ + 130, + 193, + 302, + 247 + ], + "lines": [ + { + "bbox": [ + 130, + 193, + 302, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 193, + 302, + 247 + ], + "type": "text", + "content": "Table 11: Comparison of pre-training effect between RGB and Canny images when shape augmentation is applied. The shape augmentation uses affine transformations." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 140, + 258, + 291, + 305 + ], + "lines": [ + { + "bbox": [ + 140, + 258, + 291, + 305 + ], + "spans": [ + { + "bbox": [ + 140, + 258, + 291, + 305 + ], + "type": "table", + "html": "
ImageMean Accuracy
RGB image81.6
Canny image82.2
", + "image_path": "32e09d909fb3581b18d764eeb64dd74642b9b765be1b0b09022fb658ed5c04b8.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 314, + 245, + 476, + 304 + ], + "blocks": [ + { + "bbox": [ + 311, + 194, + 482, + 240 + ], + "lines": [ + { + "bbox": [ + 311, + 194, + 482, + 240 + ], + "spans": [ + { + "bbox": [ + 311, + 194, + 482, + 240 + ], + "type": "text", + "content": "Table 12: Comparison of pre-training effects of affine, elastic, and polynomial image transformations. Real Img type with Canny applied to the images." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 314, + 245, + 476, + 304 + ], + "lines": [ + { + "bbox": [ + 314, + 245, + 476, + 304 + ], + "spans": [ + { + "bbox": [ + 314, + 245, + 476, + 304 + ], + "type": "table", + "html": "
TransformationMean Accuracy
Affine trans.81.9
Elastic trans.37.0
Polynomial trans.81.3
", + "image_path": "ce742c89fc889e129780686fcfb55299fc5249999ad7fbe5adfb4562b29a82f7.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 306, + 482, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 306, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 306, + 482, + 390 + ], + "type": "text", + "content": "training effects with " + }, + { + "bbox": [ + 130, + 306, + 482, + 390 + ], + "type": "inline_equation", + "content": "L = 1,000" + }, + { + "bbox": [ + 130, + 306, + 482, + 390 + ], + "type": "text", + "content": " are shown in Tables 11 and 12. Note that the results in these tables represent the average score on C100 over the five real images in Figure 5. See the supplementary material for results for each of image. According to the results, we find that contour-emphasized Canny images omitting color information and affine transformation perturbation yield better pre-training, as measured by the fine-tuning accuracies. The operations are similar to the classification of labeled fractal images via IFS." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 407, + 224, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 224, + 419 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 224, + 419 + ], + "type": "text", + "content": "4.5 Applications" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 426, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 640 + ], + "type": "text", + "content": "Fine-tuning on commonly used datasets (Table 13). We compare fine-tuning accuracies in three types of dataset configurations. For supervised learning (SL) we compare pretraining on ImageNet-1k [16] and Places-365 [59], for self-supervised learning (SSL) we compare DINO [7] and masked auto-encoder (MAE) [22] pre-trained on ImageNet and PASS, and models trained on the fractal datasets FDSL on FractalDB-1k and RCDB-1k. Following the setting of one-instance per category [35], we also evaluate using random patch augmentation (\"w/ Aug.\") We compare our 1p-frac with 1-image SSL [3] with DINO and MAE (we use Real Img #1 of Figure 5). The results show that our 1p-frac pretraining with one single image demonstrated high fine-tuning accuracies (87.5% and 88.2% with and without random patch augmentation) which are close to the performance of self-supervised ImageNet-1k pre-training with MAE that utilizes 1.28M images and yields 88.5%. Surprisingly, despite using only one fractal image with perturbations, our 1p-frac without random patch augmentation has demonstrated fine-tuning accuracies that are equal to or even higher than that of conventional OFDB-1k/FractalDB-1k, which uses 1,000 parameter sets with fractal categories and 1,000/1,000,000 images: 1p-frac obtains mean performances of 87.5% vs OFDB-1k 87.2% vs FractalDB-1k 87.1%." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Fine-tuning on specialized and structured datasets in VTAB [54] (Table 14). We also verify performance on a couple of specialized and structured" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 317, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 317, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 317, + 101 + ], + "type": "text", + "content": "R. Nakamura and R. Tadokoro et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 185, + 483, + 396 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "text", + "content": "Table 13: Comparison among pre-training methods in fine-tuning accuracies. Best values at each dataset scale are in bold. ViT-T is used for all experiments. #OrgImg indicates number of collected images in real-image datasets or number of original images used in synthetic datasets. Type shows supervised types within supervised learning using cross-entropy loss (SL), self-supervised learning using DINO [7] or MAE [22] (SSL:D/SSL:M), and formula-supervised learning (FDSL)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 185, + 483, + 396 + ], + "lines": [ + { + "bbox": [ + 133, + 185, + 483, + 396 + ], + "spans": [ + { + "bbox": [ + 133, + 185, + 483, + 396 + ], + "type": "table", + "html": "
Pre-training#OrgImgTypeC10C100CarsFlowersVOC12P30IN100Mean
Scratch--80.362.113.771.456.276.274.862.1
Places-365 [59]1.80MSL97.683.989.299.384.6-89.4-
ImageNet-1k [16]1.28MSL98.085.589.999.488.780.0--
ImageNet-1k [16]1.28MSSL:D97.782.488.098.574.778.489.086.9
ImageNet-1k [16]1.28MSSL:M97.485.886.696.383.380.290.088.5
PASS [2]1.43MSSL:D97.584.086.498.682.979.082.987.8
FractalDB-1k [30]1.00MFDSL96.881.686.098.380.678.488.387.1
RCDB-1k [28]1.00MFDSL97.082.286.598.980.979.788.587.6
ImageNet-1k [35]1,000SL94.376.957.394.873.878.284.379.9
ImageNet-1k [35]1,000SSL:D94.978.071.294.675.578.684.982.5
OFDB-1k [35]1,000FDSL96.984.084.597.179.979.988.087.2
3D-OFDB-1k [35]1,000FDSL97.183.885.598.480.880.089.187.8
OFDB-1k w/ Aug. [35]1,000FDSL97.285.387.698.381.480.489.588.5
3D-OFDB-1k w/ Aug. [35]1,000FDSL97.084.785.698.381.279.888.987.9
1-image SSL [3]1SSL:D95.779.573.193.869.080.488.782.8
1-image SSL [3]1SSL:M97.284.782.194.678.580.389.086.6
1p-frac (ours)1FDSL96.984.284.597.480.580.689.087.5
1p-frac w/ Aug. (ours)1FDSL96.584.787.098.180.980.588.988.2
", + "image_path": "8626291342d5d868f4a016f8934a5bfe4de965208f2fd3fbbe3a0007da8588a1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 405, + 482, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 405, + 482, + 513 + ], + "spans": [ + { + "bbox": [ + 130, + 405, + 482, + 513 + ], + "type": "text", + "content": "datasets in VTAB (visual task adaptation benchmark; Retinopathy [26], Resisc45 [13], Camelyon [48], CLEVR-Count [25], and sNORB-Azim [33]). The experiment is more suitable to evaluate and compare the pre-training methods since the image domains and their labels are distinct from both real and fractal images on ImageNet and 1p-frac. According to Table 14, our 1p-frac pretrained model performed relatively similar fine-tuning accuracies on all listed datasets in the table. By comparing to self-supervised ImageNet-1k with MAE, our 1p-frac surpassed the fine-tuning accuracies on three (Retinopathy, Resisc45, and sNORB-Azim) out of the five datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 532, + 311, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 532, + 311, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 532, + 311, + 544 + ], + "type": "text", + "content": "5 Conclusion and Discussion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 558, + 481, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 481, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 481, + 605 + ], + "type": "text", + "content": "This paper examined what a minimal dataset for synthetic pre-training might look like. We proposed the 1-parameter Fractal as Data (1p-frac) dataset that succeeds in pre-training even with a single fractal by utilizing the locally perturbed cross entropy (LPCE) loss. The following findings were presented:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "content": "What is a minimal synthetic pre-training dataset? In this paper, we applied IFS to generate labeled images. In this context, preparing a single set of parameters " + }, + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "inline_equation", + "content": "a - f" + }, + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "content": " in IFS is the least informative. It recorded fine-tuning accuracy nearly equivalent to conventional methods such as FractalDB and OFDB (Tables 1 and 2), succeeding in pre-training with minimal image pattern utilization." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 370, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 370, + 91, + 447, + 102 + ], + "type": "text", + "content": "Scaling Backwards" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 140, + 159, + 474, + 224 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 158 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 158 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 158 + ], + "type": "text", + "content": "Table 14: Performance comparisons on five fine-tuning datasets: Retinopathy (Retino) [26], Resisc45 [13], Camelyon [48], CLEVR-Count (CLEVR-C) [25], and sNORB-Azim (sNORB-A) [33]. Best and second-best values are in bold and underlined, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 140, + 159, + 474, + 224 + ], + "lines": [ + { + "bbox": [ + 140, + 159, + 474, + 224 + ], + "spans": [ + { + "bbox": [ + 140, + 159, + 474, + 224 + ], + "type": "table", + "html": "
#OrgImgTypeRetinoResisc45CamelyonCLEVR-CsNORB-A
Scratch--66.989.274.446.411.5
ImageNet-1k [16]1.28MSL79.197.082.789.329.2
ImageNet-1k [16]1.28MSSL:M76.795.683.989.526.3
1p-frac (ours)1FDSL77.995.882.786.228.9
", + "image_path": "642c94a1c4e9918163fa05d2adb0e8d731fea4de5d596eee762c953eeb6c48e5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 234, + 479, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 234, + 479, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 234, + 479, + 281 + ], + "type": "text", + "content": "tion. Moreover, it is not that any perturbation in image space is acceptable, as Gaussian/uniform distributions failed in pre-training (Table 4). It clarified the importance of geometric transformations with certain regularities like fractal geometry or contour shapes (Table 3)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 281, + 480, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 480, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 480, + 376 + ], + "type": "text", + "content": "What properties of images lead to image reduction? In 1p-frac, we investigated shape changes and pre-training effects by adopting the LPCE loss and the " + }, + { + "bbox": [ + 130, + 281, + 480, + 376 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 281, + 480, + 376 + ], + "type": "text", + "content": "-factor that contribute to fractal shape variation. It was revealed that pre-training efficacy varies with shape variation (Tables 5, 6). For real images, it is not just about including more edge components, but about including affine transformations, which closely resembles the classification of labeled fractal images through IFS (Tables 11 and 12). By extensively abating data augmentation showed, we demonstrated the essential role of random cropping and MixUp." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 378, + 481, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 378, + 481, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 378, + 481, + 496 + ], + "type": "text", + "content": "Pre-training effects with the proposed method. Supervised pre-training on ImageNet was more accurate for datasets where object labels are assigned to real images (Table 13), but in VTAB's Specialized/Structured datasets (Table 14), performances were relatively close. On the other hand, our 1p-frac pre-trained ViT performed a similar mean accuracy by comparing to MAE self-supervised ImageNet pre-training (MAE 88.5 vs. our 1p-frac w/ Aug. 88.2). Furthermore, increasing the number of sampling points " + }, + { + "bbox": [ + 130, + 378, + 481, + 496 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 378, + 481, + 496 + ], + "type": "text", + "content": " showed performance improvement even with 21k categories and demonstrated good performance in ImageNet-1k fine-tuning (Table 8), indicating the potential for building large models with limited data resources." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 498, + 481, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 481, + 568 + ], + "type": "text", + "content": "Broader impact. Our proposed dataset 1p-frac does not suffer from licensing or ethical issues such as biases and has a large potential to serve as a clean pretraining dataset. While not investigated in this paper, this approach also opens the possibility for significant gains in training speed by keeping the data on the GPU and applying transformations there, removing the GPU-CPU transfer bottleneck." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "content": "Limitation. In this paper, we have explored achieving a minimal synthetic pre-training in image recognition based on fractal geometry with IFS (a single fractal, and their perturbations). Despite this, we believe it is also necessary to find a similarly minimal pre-training dataset containing real images - for some applications and increased interpretability. This might allow for quick and clear calibration of pre-training, similar to how calibration images are used in photography. We did not investigate this direction in this paper but leave this for future work." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 317, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 317, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 317, + 102 + ], + "type": "text", + "content": "R. Nakamura and R. Tadokoro et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 142, + 479, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 142, + 479, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 142, + 479, + 177 + ], + "type": "text", + "content": "Computational resource of AI Bridging Cloud Infrastructure (ABCI) provided by National Institute of Advanced Industrial Science and Technology (AIST) was used." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 192, + 197, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 192, + 197, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 192, + 197, + 205 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 220, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 220, + 480, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 220, + 480, + 252 + ], + "spans": [ + { + "bbox": [ + 138, + 220, + 480, + 252 + ], + "type": "text", + "content": "1. Anderson, C., Farrell, R.: Improving fractal pre-training. In: Proceedings of the 2022 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 1300-1309 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 255, + 480, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 255, + 480, + 285 + ], + "spans": [ + { + "bbox": [ + 138, + 255, + 480, + 285 + ], + "type": "text", + "content": "2. Asano, Y., Rupprecht, C., Zisserman, A., Vedaldi, A.: Pass: An imagenet replacement for self-Supervised pretraining without humans. In: Proceedings of the 2021 Neural Information Processing Systems Track on Datasets and Benchmarks (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 288, + 480, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 288, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 138, + 288, + 480, + 319 + ], + "type": "text", + "content": "3. Asano, Y.M., Rupprecht, C., Vedaldi, A.: A critical analysis of self-supervision, or what we can learn from a single image. In: Proceedings of the 2020 International Conference on Learning Representations. (ICLR) (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 321, + 480, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 321, + 480, + 352 + ], + "spans": [ + { + "bbox": [ + 138, + 321, + 480, + 352 + ], + "type": "text", + "content": "4. Baradad, M., Chen, C.F., Wulff, J., Wang, T., Feris, R., Torralba, A., Isola, P.: Procedural image programs for representation learning. In: Proceedings of the 2022 Advances in Neural Information Processing Systems (NeurIPS) (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 354, + 480, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 354, + 480, + 386 + ], + "spans": [ + { + "bbox": [ + 138, + 354, + 480, + 386 + ], + "type": "text", + "content": "5. Baradad, M., Wulff, J., Wang, T., Isola, P., Torralba, A.: Learning to See by Looking at Noise. In: Proceedings of the 2021 Advances in Neural Information Processing Systems (NeurIPS). vol. 34, pp. 2556-2569 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 388, + 480, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 388, + 480, + 408 + ], + "spans": [ + { + "bbox": [ + 138, + 388, + 480, + 408 + ], + "type": "text", + "content": "6. Canny, J.: A computational approach to edge detection. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 8(6), 679-698 (1986)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 411, + 480, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 411, + 480, + 452 + ], + "spans": [ + { + "bbox": [ + 138, + 411, + 480, + 452 + ], + "type": "text", + "content": "7. Caron, M., Touvron, H., Misra, I., Jegou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the 2021 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 9650-9660 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 454, + 480, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 454, + 480, + 497 + ], + "spans": [ + { + "bbox": [ + 138, + 454, + 480, + 497 + ], + "type": "text", + "content": "8. Cazenavette, G., Wang, T., Torralba, A., Efros, A.A., Zhu, J.Y.: Dataset distillation by matching training trajectories. In: Proceedings of the 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10708-10717 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 499, + 480, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 499, + 480, + 530 + ], + "spans": [ + { + "bbox": [ + 138, + 499, + 480, + 530 + ], + "type": "text", + "content": "9. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A Simple Framework for Contrastive Learning of Visual Representations. In: Proceedings of the 2020 International Conference on Machine Learning (ICML). vol. 119, pp. 1597-1607 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 533, + 480, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 533, + 480, + 552 + ], + "spans": [ + { + "bbox": [ + 138, + 533, + 480, + 552 + ], + "type": "text", + "content": "0. Chen, X., Fan, H., Girshick, R., He, K.: Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 555, + 480, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 555, + 480, + 586 + ], + "spans": [ + { + "bbox": [ + 134, + 555, + 480, + 586 + ], + "type": "text", + "content": "11. Chen, X., He, K.: Exploring simple siamese representation learning. In: Proceedings of the 2021 IEEE/CVF International Conference on Computer Vision and Pattern Recognition (CVPR). pp. 15750-15758 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 588, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 588, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 134, + 588, + 480, + 620 + ], + "type": "text", + "content": "12. Chen, X., Xie, S., He, K.: An empirical study of training self-supervised vision transformers. In: Proceedings of the 2021 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 9640-9649 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 622, + 480, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 622, + 480, + 642 + ], + "spans": [ + { + "bbox": [ + 134, + 622, + 480, + 642 + ], + "type": "text", + "content": "13. Cheng, G., Han, J., Lu, X.: Remote sensing image scene classification: Benchmark and state of the art. Proceedings of the IEEE 105(10), 1865-1883 (2017)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 644, + 480, + 665 + ], + "type": "text", + "content": "14. Cubuk, E.D., Zoph, B., Shlens, J., Le, Q.V.: Randaugment: Practical data augmentation with no separate search. arXiv preprint abs/2104.07918 (2019)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 370, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 370, + 91, + 447, + 101 + ], + "type": "text", + "content": "Scaling Backwards" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 215 + ], + "type": "text", + "content": "15. Dehghani, M., Djolonga, J., Mustafa, B., Padlewski, P., Heek, J., Gilmer, J., Steiner, A.P., Caron, M., Geirhos, R., Alabdulmohsin, I., Jenatton, R., Beyer, L., Tschannen, M., Arnab, A., Wang, X., Riquelme Ruiz, C., Minderer, M., Puigcerver, J., Evci, U., Kumar, M., Steenkiste, S.V., Elsayed, G.F., Mahendran, A., Yu, F., Oliver, A., Huot, F., Bastings, J., Collier, M., Gritsenko, A.A., Birodkar, V., Vasconcelos, C.N., Tay, Y., Mensink, T., Kolesnikov, A., Pavetic, F., Tran, D., Kipf, T., Lucic, M., Zhai, X., Keysers, D., Harmsen, J.J., Houlsby, N.: Scaling vision transformers to 22 billion parameters. In: Proceedings of the 2023 International Conference on Machine Learning. vol. 202, pp. 7480-7512 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 216, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 216, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 133, + 216, + 481, + 258 + ], + "type": "text", + "content": "16. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A large-scale hierarchical image database. In: Proceedings of the 2009 IEEE Computer Society Conference on Computer Vision and Pattern Recognition Workshops (CVPR Workshops). pp. 248-255 (2009)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 259, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 259, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 133, + 259, + 481, + 293 + ], + "type": "text", + "content": "17. Doersch, C., Gupta, A., Efros, A.A.: Unsupervised visual representation learning by context prediction. In: Proceedings of the 2015 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 1422-1430 (2015)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 293, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 293, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 133, + 293, + 481, + 346 + ], + "type": "text", + "content": "18. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: Proceedings of the 2021 International Conference on Learning Representation (ICLR) (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 347, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 347, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 133, + 347, + 481, + 380 + ], + "type": "text", + "content": "19. Dwibedi, D., Misra, I., Hebert, M.: Cut, paste and learn: Surprisingly easy synthesis for instance detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 1310-1319 (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "text", + "content": "20. Everingham, M., Eslami, S.M.A., Van Gool, L., Williams, C.K.I., Winn, J., Zisserman, A.: The Pascal Visual Object Classes Challenge: A Retrospective. International Journal of Computer Vision (IJCV) 111(1), 98-136 (2015)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 413, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 481, + 445 + ], + "type": "text", + "content": "21. Gidaris, S., Singh, P., Komodakis, N.: Unsupervised representation learning by predicting image rotations. In: International Conference on Learning Representation (ICLR) (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "type": "text", + "content": "22. He, K., Chen, X., Xie, S., Li, Y., Dólár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16000-16009 (June 2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 479, + 481, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 512 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 512 + ], + "type": "text", + "content": "23. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: IEEE/CVF International Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9729-9738 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "text", + "content": "24. Inoue, N., Yamagata, E., Kataoka, H.: Initialization using perlin noise for training networks with a limited amount of data. In: Proceedings of the 2020 International Conference on Pattern Recognition (ICPR). pp. 1023-1028 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 545, + 481, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 481, + 588 + ], + "type": "text", + "content": "25. Johnson, J., Hariharan, B., Maaten, L.v.d., Li, F.F., Zitnick, L., Girshick, R.: CLEVR: A diagnostic dataset for compositional language and elementary visual reasoning. In: Proceedings of the 2017 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1988-1997 (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 589, + 404, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 404, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 404, + 600 + ], + "type": "text", + "content": "26. Kaggle, Eyepacs: Kaggle diabetic retinopathy detection (2015)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 600, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 481, + 643 + ], + "type": "text", + "content": "27. Kataoka, H., Hara, K., Hayashi, R., Yamagata, E., Inoue, N.: Spatiotemporal initialization for 3D CNNs with generated motion patterns. In: Proceedings of the 2022 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 737-746 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "text", + "content": "28. Kataoka, H., Hayamizu, R., Yamada, R., Nakashima, K., Takashima, S., Zhang, X., Martinez-Noriega, E.J., Inoue, N., Yokota, R.: Replacing labeled real-image" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 317, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 317, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 317, + 101 + ], + "type": "text", + "content": "R. Nakamura and R. Tadokoro et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 147, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 147, + 116, + 481, + 149 + ], + "type": "text", + "content": "datasets with auto-generated contours. In: Proceedings of the 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 21232-21241 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 151, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 151, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 151, + 481, + 194 + ], + "type": "text", + "content": "29. Kataoka, H., Matsumoto, A., Yamada, R., Satoh, Y., Yamagata, E., Inoue, N.: Formula-driven supervised learning with recursive tiling patterns. In: Proceedings of the 2021 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 4098-4105 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 195, + 481, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 195, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 195, + 481, + 228 + ], + "type": "text", + "content": "30. Kataoka, H., Okayasu, K., Matsumoto, A., Yamagata, E., Yamada, R., Inoue, N., Nakamura, A., Satoh, Y.: Pre-training without natural images. International Journal of Computer Vision (IJCV) 130 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 229, + 481, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 229, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 229, + 481, + 262 + ], + "type": "text", + "content": "31. Krause, J., Stark, M., Deng, J., Fei-Fei, L.: 3d object representations for fine-grained categorization. In: Proceedings of the 2013 International IEEE Workshop on 3D Representation and Recognition (3DRR-13). pp. 554-561 (2013)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 263, + 481, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 263, + 481, + 285 + ], + "spans": [ + { + "bbox": [ + 132, + 263, + 481, + 285 + ], + "type": "text", + "content": "32. Krizhevsky, A., Hinton, G.: Learning Multiple Layers of Features from Tiny Images. Master's thesis, Department of Computer Science, University of Toronto (2009)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 285, + 481, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 481, + 328 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 481, + 328 + ], + "type": "text", + "content": "33. LeCun, Y., Huang, F.J., Bottou, L.: Learning methods for generic object recognition with invariance to pose and lighting. In: Proceedings of the 2004 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). vol. 2, pp. II-104 Vol.2 (2004)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 330, + 481, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 330, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 330, + 481, + 373 + ], + "type": "text", + "content": "34. Mahajan, D., Girshick, R., Ramanathan, V., He, K., Paluri, M., Li, Y., Bharambe, A., Maaten, L.v.d.: Exploring the Limits of Weakly Supervised Pretraining. In: Proceedings of the 2018 European Conference on Computer Vision (ECCV). pp. 181-196 (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 374, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 481, + 418 + ], + "type": "text", + "content": "35. Nakamura, R., Kataoka, H., Takashima, S., Noriega, E.J.M., Yokota, R., Inoue, N.: Pre-training vision transformers with very limited synthesized images. In: Proceedings of the 2023 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 20303-20312. IEEE (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 420, + 481, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 420, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 420, + 481, + 453 + ], + "type": "text", + "content": "36. Nilsback, M.E., Zisserman, A.: Automated flower classification over a large number of classes. In: Proceedings of the 2008 Indian Conference on Computer Vision, Graphics and Image Processing. pp. 722-729 (2008)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 453, + 481, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 481, + 485 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 481, + 485 + ], + "type": "text", + "content": "37. Noroozi, M., Favaro, P.: Unsupervised Learning of Visual Representations by Solving Jigsaw Puzzles. In: Proceedings of the 2016 European Conference on Computer Vision (ECCV). pp. 69-84 (2016)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 487, + 481, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 481, + 520 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 481, + 520 + ], + "type": "text", + "content": "38. Noroozi, M., Vinjimoor, A., Favaro, P., Pirsiavash, H.: Boosting Self-Supervised Learning via Knowledge Transfer. In: IEEE/CVF International Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9359-9367 (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 521, + 481, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 521, + 481, + 552 + ], + "spans": [ + { + "bbox": [ + 132, + 521, + 481, + 552 + ], + "type": "text", + "content": "39. Remez, T., Huang, J., Brown, M.: Learning to Segment via Cut-and-Paste. In: Proceedings of the 2018 European Conference on Computer Vision (ECCV). pp. 39-54 (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 554, + 481, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 554, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 554, + 481, + 597 + ], + "type": "text", + "content": "40. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., Berg, A.C., Fei-Fei, L.: ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV) 115(3) (2015)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 599, + 481, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 481, + 631 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 481, + 631 + ], + "type": "text", + "content": "41. Sharma, P., Rott Shaham, T., Baradad, M., Fu, S., Rodriguez-Munoz, A., Duggal, S., Isola, P., Torralba, A.: A vision check-up for language models. In: arXiv preprint arXiv:2401.01862 (2024)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "text", + "content": "42. Sun, C., Shrivastava, A., Singh, S., Gupta, A.: Revisiting unreasonable effectiveness of data in deep learning era. In: Proceedings of the 2017 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 843-852 (2017)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 370, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 370, + 91, + 447, + 101 + ], + "type": "text", + "content": "Scaling Backwards" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "43. Sundermeyer, M., Marton, Z.C., Durner, M., Brucker, M., Triebel, R.: Implicit 3D orientation learning for 6D object detection from RGB images. In: Proceedings of the 2018 European Conference on Computer Vision (ECCV). pp. 712-729 (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "type": "text", + "content": "44. Tadokoro, R., Yamada, R., Nakashima, K., Nakamura, R., Kataoka, H.: Primitive Geometry Segment Pre-training for 3D Medical Image Segmentation. In: Proceedings of the 2023 British Machine Vision Conference (BMVC) (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 226 + ], + "type": "text", + "content": "45. Takashima, S., Hayamizu, R., Inoue, N., Kataoka, H., Yokota, R.: Visual atoms: Pre-training vision transformers with sinusoidal waves. In: Proceedings of the 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 18579-18588 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 270 + ], + "type": "text", + "content": "46. Tobin, J., Fong, R., Ray, A., Schneider, J., Zaremba, W., Abbeel, P.: Domain randomization for transferring deep neural networks from simulation to the real world. In: Proceedings of the 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 23-30 (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 270, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 481, + 314 + ], + "type": "text", + "content": "47. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jegou, H.: Training data-efficient image transformers & distillation through attention. In: Proceedings of the 2021 International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 139, pp. 10347-10357 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 315, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 357 + ], + "type": "text", + "content": "48. Veeling, B.S., Linmans, J., Winkens, J., Cohen, T., Welling, M.: Rotation equivariant cnns for digital pathology. In: Proceedings of the 2018 International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI) (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 358, + 481, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 390 + ], + "type": "text", + "content": "49. Venkataramanan, S., Rizve, M.N., Carreira, J., Asano, Y.M., Avrithis, Y.: Is image-net worth 1 video? learning strong image encoders from 1 long unlabelled video. arXiv preprint arXiv:2310.08584 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 391, + 443, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 391, + 443, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 391, + 443, + 402 + ], + "type": "text", + "content": "50. Wang, T., Zhu, J., Torralba, A., Efros, A.A.: Dataset Distillation (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 403, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 481, + 445 + ], + "type": "text", + "content": "51. Yun, S., Han, D., Chun, S., Oh, S.J., Yoo, Y., Choe, J.: CutMix: Regularization strategy to train strong classifiers with localizable features. In: Proceedings of the 2019 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 6022-6031 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "type": "text", + "content": "52. Yun-Hao Cao, H.Y., Wu, J.: Training vision transformers with only 2040 images. In: Proceedings of the 2022 European Conference on Computer Vision (ECCV). pp. 220-237 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "text", + "content": "53. Zhai, X., Kolesnikov, A., Houlsby, N., Beyer, L.: Scaling vision transformers. Proceedings of 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 1204-1213 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 512, + 481, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 481, + 566 + ], + "type": "text", + "content": "54. Zhai, X., Puigcerver, J., Kolesnikov, A., Ruyssen, P., Riquelme, C., Lucic, M., Djolonga, J., Pinto, A.S., Neumann, M., Dosovitskiy, A., Beyer, L., Bachem, O., Tschannen, M., Michalski, M., Bousquet, O., Gelly, S., Houlsby, N.: A Large-scale Study of Representation Learning with the Visual Task Adaptation Benchmark. arXiv preprint arXiv:1910:04867 (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 567, + 481, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 588 + ], + "type": "text", + "content": "55. Zhang, H., Cisse, M., Dauphin, Y.N., Lopez-Paz, D.: Mixup: Beyond empirical risk minimization. arXiv preprint abs/1710.09412 (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 589, + 481, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 481, + 610 + ], + "type": "text", + "content": "56. Zhang, R., Isola, P., Efros, A.A.: Colorful image colorization. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 649-666 (2016)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "text", + "content": "57. Zhao, B., Mopuri, K.R., Bilen, H.: Dataset condensation with gradient matching. In: Proceedings of the 2021 International Conference on Learning Representation (ICLR) (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "text", + "content": "58. Zhong, Z., Zheng, L., Kang, G., Li, S., Yang, Y.: Random erasing data augmentation. arXiv preprint abs/1708.04896 (2017)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 317, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 317, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 317, + 101 + ], + "type": "text", + "content": "R. Nakamura and R. Tadokoro et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 482, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 482, + 150 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 482, + 150 + ], + "type": "text", + "content": "59. Zhou, B., Lapedriza, A., Khosla, A., Oliva, A., Torralba, A.: Places: A 10 million image database for scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 40(6), 1452-1464 (2017)" + } + ] + } + ], + "index": 2 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 370, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 370, + 91, + 447, + 102 + ], + "type": "text", + "content": "Scaling Backwards" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Scaling Up Personalized Image Aesthetic Assessment via Task Vector Customization/77e3f8ef-24e2-4d25-ba6f-999ef6bc7191_content_list.json b/2024/Scaling Up Personalized Image Aesthetic Assessment via Task Vector Customization/77e3f8ef-24e2-4d25-ba6f-999ef6bc7191_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4ed61143f599073c9b4c65e262b9ad4a710d03a9 --- /dev/null +++ b/2024/Scaling Up Personalized Image Aesthetic Assessment via Task Vector Customization/77e3f8ef-24e2-4d25-ba6f-999ef6bc7191_content_list.json @@ -0,0 +1,1743 @@ +[ + { + "type": "text", + "text": "Scaling Up Personalized Image Aesthetic Assessment via Task Vector Customization", + "text_level": 1, + "bbox": [ + 256, + 140, + 746, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jooyeol Yun and Jaegul Choo", + "bbox": [ + 380, + 212, + 620, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Korea Advanced Institute of Science and Technology (KAIST) blizzard072@kaist.ac.kr jchoo@kaist.ac.kr", + "bbox": [ + 292, + 239, + 707, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. The task of personalized image aesthetic assessment seeks to tailor aesthetic score prediction models to match individual preferences with just a few user-provided inputs. However, the scalability and generalization capabilities of current approaches are considerably restricted by their reliance on an expensive curated database. To overcome this long-standing scalability challenge, we present a unique approach that leverages readily available databases for general image aesthetic assessment and image quality assessment. Specifically, we view each database as a distinct image score regression task that exhibits varying degrees of personalization potential. By determining optimal combinations of task vectors, known to represent specific traits of each database, we successfully create personalized models for individuals. This approach of integrating multiple models allows us to harness a substantial amount of data. Our extensive experiments demonstrate the effectiveness of our approach in generalizing to previously unseen domains—a challenge previous approaches have struggled to achieve—making it highly applicable to real-world scenarios. Our novel approach significantly advances the field by offering scalable solutions for personalized aesthetic assessment and establishing high standards for future research.", + "bbox": [ + 261, + 282, + 743, + 547 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 563, + 375, + 578 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Personalized image aesthetic assessment (PIAA) is an emerging field dedicated to developing aesthetic score prediction models that closely match an individual's aesthetic preferences based on a small set of user-provided samples. While they function as simple models that assign scores based on aesthetic quality, PIAA models enhance user experience in digital environments by personalizing tasks such as managing photo albums [17], curating web-scale databases like LAION-Aesthetics [33], and even guiding generative models such as Stable Diffusion [32] to create images that align with individual tastes [34].", + "bbox": [ + 212, + 593, + 787, + 714 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, there has been a notable trend in PIAA [21, 41, 44] that involves the integration of meta-learning techniques, as personalization is inherently related to few-shot learning. Specifically, these approaches entail pretraining a model on a database consisting of image collections from diverse individuals. However, these strategies encounter significant challenges in terms of scalability and generalization to unseen domains, primarily due to the prohibitive cost associated with data collection.", + "bbox": [ + 212, + 714, + 787, + 820 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1 https://yeolj00.github.io/personal-projects/personalized-aesthetics/", + "bbox": [ + 217, + 824, + 761, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2f8f78a387819845d903c928eb9b0ae98f9784f842142670e7d9fc764cbb51e7.jpg", + "image_caption": [ + "Fig. 1: Comparison between PIAA and other image regression databases. Our approach has the flexibility to leverage any image regression databases." + ], + "image_footnote": [], + "bbox": [ + 300, + 143, + 697, + 268 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While meta-learning typically demands an extensive number of meta-training tasks, often exceeding 10,000, to achieve high performance [1], currently available PIAA databases [19, 31] offer a limited number of tasks fewer than 200. Even with the recently collected PIAA database [40], which contains rich attributes from each annotator, the number of training tasks remains below 500. In the context of PIAA, a single meta-training task corresponds to learning the preferences of a single individual. Scaling such a database would require obtaining a vast number of images and corresponding personalized aesthetic scores from numerous individuals, which poses severe limitations in the scalability and generalization of existing approaches.", + "bbox": [ + 212, + 309, + 784, + 460 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we unveil a data-rich approach for personalizing aesthetic assessment models by leveraging existing general image aesthetic assessment (GIAA) and image quality assessment (IQA) databases. In contrast to prior work [20, 21, 41, 44, 45], our approach is not limited to training on databases that track individual annotators, as depicted in Figure 1. Instead, we have the flexibility to utilize multiple image score regression databases, which are readily available. By employing broader resources, we overcome the scalability and generalization challenges that have long been unaddressed in this field.", + "bbox": [ + 212, + 460, + 785, + 582 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/88277c622dc9f4f83ad4434af9fd83046166b75590aa18af1090d1ac57018053.jpg", + "image_caption": [ + "Fig.2: Zero-shot personalization performance for four users in the Flickr-AES database [31]. Each colored bar refers to the personalization performance of a model trained on a specific database." + ], + "image_footnote": [], + "bbox": [ + 287, + 589, + 709, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The key idea is to view each GIAA and IQA data as a unique image score regression task. This perspective arises from the distinct characteristics and tendencies of each database due to different data collection procedures and the population from which they originate. As visualized in Figure 2, image score regression models trained on four different GIAA [27,31] and IQA [6,12] databases", + "bbox": [ + 212, + 763, + 785, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "JY. Yun and J. Choo", + "bbox": [ + 271, + 114, + 415, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "exhibit varying levels of personalization for users within the PIAA database. For instance, User 3's aesthetic preferences closely align with the score regression model trained on the Flickr-AES [31] while diverging from the model trained on the SPAQ [6] database. In other words, even without any personalized training, we find that general image assessment models inherently possess different degrees of personalization capabilities.", + "bbox": [ + 212, + 146, + 782, + 236 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In light of these insights, we extend an emerging approach known as task arithmetic [14], which was originally devised for developing multi-task models. This method involves directly adding or subtracting the weights of models fine-tuned on specific tasks. In our context, task vectors from diverse image assessment models encapsulate distinct features and behaviors related to specific tasks, such as recognizing different aspects of image quality or aesthetics. By carefully combining task vectors, we explore the potential to selectively amplify or refine the capabilities of a pre-trained model.", + "bbox": [ + 212, + 239, + 784, + 359 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, there has been limited research on determining the optimal adjustments necessary to achieve a certain behavior. Thus, we present a personalization approach by introducing trainable coefficients for each task vector derived from models trained on distinct GIAA and IQA databases, thereby determining the weights for combining these vectors. Once the coefficients are trained with user-provided inputs, the weighted sum of task vectors creates a model tailored to the user's aesthetic preferences. Since task vectors obtained from large databases already capture preferences across various themes and contents, we find that training only the coefficients is sufficient for precise personalization, making our approach highly parameter-efficient. To the best of our knowledge, our approach marks the first to customize task vectors for specific model behaviors that were previously unattainable.", + "bbox": [ + 212, + 362, + 784, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "By carefully customizing task vectors, we incorporate large databases into the training process, granting our approach a strong generalization capability for unseen domains, a scenario often encountered in real-world applications. Furthermore, we find that our approach is also practical for learning a user's aesthetic preference with only a few samples, as our approach of merging task vectors guides the model to make well-informed updates and prevents overfitting. We provide extensive experiments demonstrating the efficacy of our approach for personalization in diverse scenarios.", + "bbox": [ + 212, + 544, + 784, + 665 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our contributions are threefold:", + "bbox": [ + 240, + 667, + 470, + 681 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a novel data-rich approach for PIAA that tackles the long-standing scalability challenge, eliminating the dependency on expensive, manually-curated databases that have hindered progress in this field.", + "- Demonstrating exceptional effectiveness in cross-database evaluations, our method showcases robust generalization capabilities that surpass those of existing approaches, setting high standards for future work.", + "- By learning optimal combinations of task vectors, we present a parameter-efficient approach for precise personalization, leveraging the comprehensive information embedded in these vectors." + ], + "bbox": [ + 227, + 693, + 784, + 838 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Scaling Up Personalized Image Aesthetic Assessment", + "bbox": [ + 377, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 143, + 387, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.1 Personalized Image Aesthetic Assessment", + "text_level": 1, + "bbox": [ + 215, + 172, + 604, + 188 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Personalized image aesthetic assessment (PIAA) [13,20,21,24,25,35,39-41,44,45] aims to learn the personal aesthetic preference of individual users based on the image-score pairs from one's image collection. Since collecting a large number of samples from a single user is impractical, PIAA is closely linked to few-shot learning methods. One line of work [20,40,45] incorporates additional attributes (e.g., personality traits, age, and gender) to facilitate user-specific aesthetic predictions. These approaches rely on the correlation between user attributes and aesthetic preferences, which may not always be clearly defined. Consequently, many high-performing approaches for PIAA [21, 41, 44] leverage meta-learning techniques to train models that can be easily fine-tuned to accommodate the aesthetic preferences of new users with only a limited number of samples.", + "bbox": [ + 212, + 196, + 785, + 364 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Nevertheless, applying meta-learning techniques to PIAA encounters scalability challenges owing to the limited number of tasks available during the meta-learning phase. In traditional meta-learning, achieving a high performance typically requires a substantial number of meta-training tasks [1], often exceeding 10,000. In PIAA databases [19,31], the number of training tasks ranges from 100 to fewer than 500, which can impose limitations on the performance. Expanding the task pool in PIAA is prohibitively expensive since each task corresponds to the image collection of a single user, posing significant scalability challenges that has long been unaddressed. We tackle this issue in previous PIAA approaches and present a scalable fine-tuning method that leverages existing general image aesthetic assessment (GIAA) and image quality assessment (IQA) databases.", + "bbox": [ + 212, + 364, + 785, + 531 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Task Vectors and Task Arithmetic", + "text_level": 1, + "bbox": [ + 215, + 547, + 545, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Ilharco et al. [14] presented a new paradigm for creating multi-task models, utilizing the parameters of multiple fine-tuned models. They define the task vector $\\tau$ as the difference between the fine-tuned and pre-trained weights,", + "bbox": [ + 215, + 573, + 785, + 619 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tau = \\theta_ {\\mathrm {f t}} - \\theta_ {\\mathrm {p r e}},\n$$\n", + "text_format": "latex", + "bbox": [ + 450, + 632, + 552, + 647 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\theta_{\\mathrm{ft}}$ and $\\theta_{\\mathrm{pre}}$ represent the weights of the fine-tuned and pre-trained model, respectively. They demonstrate that adding and subtracting task vectors to pretrained models can encourage or suppress the ability to perform specific tasks. For example, negating the task vector derived from a language model fine-tuned on toxic data reduces the frequency of generating toxic text. However, it remains unclear what the optimal adjustments are for achieving a specific behavior.", + "bbox": [ + 212, + 657, + 785, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our paper, we investigate methods to determine necessary adjustments to achieve a certain behavior defined by a few user-provided samples. By introducing learnable coefficients to task vectors, we effectively customize models to closely match individual preferences. We also demonstrate that finding the optimal coefficients is possible with only a few samples, as the task vectors encompass comprehensive knowledge derived from large databases.", + "bbox": [ + 212, + 750, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "JY. Yun and J. Choo", + "bbox": [ + 271, + 114, + 415, + 126 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0020b54c382bcb578b60cfb77bf859bbbc2620db6b71d62d99f591b3dcca21ba.jpg", + "image_caption": [ + "(a) Obtain task vectors from various databases" + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 779, + 210 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4f6382017d196c68d3c1153419e1c9131ca888072bd07f54c09a7dda609688b1.jpg", + "image_caption": [ + "(b) Learn optimal combinations for personalization", + "Fig. 3: (a) Multiple models are fine-tuned from a single pre-trained model on different tasks to obtain task vectors. For simplicity, we omit the discussion of layers. (b) Given a small number of user-provided samples, the coefficients corresponding to each task vector are optimized with the rank loss." + ], + "image_footnote": [], + "bbox": [ + 217, + 227, + 787, + 345 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 429, + 330, + 445 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Obtaining Layer-wise Task Vectors from Various Databases", + "text_level": 1, + "bbox": [ + 214, + 454, + 750, + 469 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To obtain task vectors, we train multiple models with the same architecture on various tasks (i.e., different databases), as illustrated in Figure 3 (a). Specifically, we fine-tune the weights from a pre-trained backbone, adjusting the parameters from $\\theta_{\\mathrm{pre}}$ to $\\theta_{\\mathrm{ft}_i}$ for each task $i$ . Moreover, while the specific roles of model layers in capturing unique features are inherently complex, their distinct contributions are undeniable [3,29]. Thus, we compute task vectors across all layers, enhancing the flexibility of our approach for personalization (See appendix Section A.2).", + "bbox": [ + 214, + 482, + 787, + 588 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Accordingly, we expand the concept of original task vectors to include layerwise differentiation by,", + "bbox": [ + 214, + 589, + 785, + 619 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tau_ {i} ^ {l} = \\theta_ {\\mathrm {f t} _ {i}} ^ {l} - \\theta_ {\\mathrm {p r e}} ^ {l}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 442, + 619, + 785, + 640 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tau_{i}^{l}$ is the task vector for the $l$ -th layer of the $i$ -th task. Each vector comprehensively integrates patterns learned from large databases, acting as basis vectors for robust and precise personalization.", + "bbox": [ + 214, + 647, + 785, + 694 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "It is important to note that this process is conducted only once and is not required again for personalizing to individual users, thereby not contributing to the count of learnable parameters for personalization.", + "bbox": [ + 214, + 695, + 785, + 742 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Learning Optimal Combinations for Personalization", + "text_level": 1, + "bbox": [ + 214, + 765, + 687, + 781 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the layer-wise task vectors, our goal is to find a combination of these vectors that accurately reflects a user's aesthetic preferences. To achieve this, we introduce learnable coefficients $\\alpha_{i}^{l}$ for each task vector $\\tau_{i}^{l}$ . Once the coefficients", + "bbox": [ + 214, + 794, + 785, + 842 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Scaling Up Personalized Image Aesthetic Assessment", + "bbox": [ + 377, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "are trained, the weights of the personalized model for each layer are derived as,", + "bbox": [ + 215, + 146, + 781, + 161 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {\\mathrm {p}} ^ {l} = \\theta_ {\\mathrm {p r e}} ^ {l} + \\sum_ {i = 1} ^ {n} \\alpha_ {i} ^ {l} \\tau_ {i} ^ {l} \\quad \\text {f o r} l = 1, 2, \\dots , L, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 174, + 785, + 212 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $n$ is the number of tasks, and $L$ is the number of layers in the model. We keep all parameters other than the $n \\times L$ coefficients frozen. This is essential since the task vectors contain comprehensive insights into aesthetic preferences, enabling the model to make well-informed updates. By harnessing the rich information embedded in these vectors, our method efficiently learns complex aesthetic preferences from merely a few user-provided samples.", + "bbox": [ + 212, + 222, + 782, + 313 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our objective function is designed to penalize the misordering of aesthetic scores. Specifically, we adopt the Bradley-Terry model [2], which estimates the probability of pairwise comparison between two samples. We frame the prediction of the two scores as a binary classification problem, where the objective is to predict which sample has the higher score. The coefficients of each task vector are trained to minimize the following loss function:", + "bbox": [ + 212, + 314, + 784, + 404 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {R}} = - \\log \\frac {\\exp \\hat {y} _ {1}}{\\exp \\hat {y} _ {1} + \\exp \\hat {y} _ {2}} (y _ {1} \\geq y _ {2}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 362, + 417, + 785, + 446 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\hat{y}_1$ and $\\hat{y}_2$ represent the aesthetic score predictions for images $x_1$ and $x_2$ , whose ground-truth scores are $y_1$ and $y_2$ . We prefer this loss function over the Earth Mover Distance (EMD) or the MSE loss, due to its sample efficiency and fast convergence (See appendix Section A.3).", + "bbox": [ + 214, + 455, + 787, + 518 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Adaptive Coefficient Initialization", + "text_level": 1, + "bbox": [ + 215, + 540, + 540, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We observe that appropriately initializing the coefficients plays a crucial role in determining the coefficients that best align with the user's aesthetic preferences. When a high coefficient is assigned to a specific task vector, it indicates a strong correlation between the user's aesthetic preference and that particular task. Thus, we adaptively initialize the coefficients based on the zero-shot personalization performances of each image score regression model associated with the task vector.", + "bbox": [ + 212, + 566, + 782, + 669 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To do this, we calculate the Spearman's rank-order correlation coefficient (SROCC) [28] between the user-provided samples and the scores predicted by each image score regression model. We then apply the softmax function to initialize the coefficients $\\alpha$ as,", + "bbox": [ + 214, + 672, + 784, + 734 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {i} ^ {l} = \\frac {e ^ {\\mathrm {S R O C C} (S , \\theta_ {\\mathrm {f t} _ {i}})}}{\\sum_ {j = 1} ^ {n} e ^ {\\mathrm {S R O C C} (S , \\theta_ {\\mathrm {f t} _ {j}})}} \\quad \\mathrm {f o r} l = 1, 2, \\ldots , L, \\qquad \\qquad (4)\n$$\n", + "text_format": "latex", + "bbox": [ + 338, + 744, + 785, + 790 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathrm{SROCC}(S,\\theta_{\\mathrm{ft}_i})$ indicates the SROCC calculated on the user-provided set $S$ using the model with parameters $\\theta_{\\mathrm{ft}_i}$ . Note that we initialize $\\alpha$ values for a single task $i$ with the same value for all layers.", + "bbox": [ + 214, + 797, + 787, + 845 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "JY. Yun and J. Choo", + "bbox": [ + 271, + 114, + 415, + 127 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8460e82d30767886a45d3bd844d834159c21402a86e5a5dfaa2b5fb8e0aae528.jpg", + "table_caption": [ + "Table 1: Description of the databases included in our experiments. The AADB and REAL-CUR databases are exclusively reserved as a test database." + ], + "table_footnote": [], + "table_body": "
DatabaseRegression taskAnnotator tagNumber of images
KonIQ-10K [12]IQA×10,073
SPAQ [6]IQA×11,125
AVA [27]GIAA×255,500
TAD66K [11]GIAA×67,125
Flickr-AES [31]GIAA/PIAA✓40,000
PARA [40]GIAA/PIAA✓31,220
AADB [19]GIAA/PIAA✓9,958
REAL-CUR [31]PIAA✓2,871
", + "bbox": [ + 222, + 174, + 776, + 329 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 342, + 374, + 358 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1 Database Overview", + "text_level": 1, + "bbox": [ + 215, + 371, + 419, + 385 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The key advantage of our approach is the flexibility to utilize multiple image score regression databases freely, even those without annotator tags. In our experiments, we fine-tune models using six different databases in Table 1, reserving the AADB [19] and REAL-CUR [31] database exclusively for testing. While the Flickr-AES [31] and PARA [40] databases do contain annotator tags, our approach does not require distinguishing between individuals. Therefore, we treat these two databases as GIAA databases. In total, we effectively leverage 415,043 samples from six databases, marking the most extensive utilization of samples in PIAA within our knowledge. Detailed descriptions and performance on these databases are provided in the appendix (See Section A.5). However, our approach is not limited to training solely on these six databases, having the scalability to include additional IQA and GIAA databases [4,8,10,16,42].", + "bbox": [ + 212, + 390, + 785, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For testing, we use the following three most widely-used PIAA test sets.", + "bbox": [ + 240, + 571, + 759, + 587 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Flickr-AES [31] database consists of 40,000 images collected from 210 users. For testing purposes, we utilized a test set containing 4,737 images rated by 37 users, with each user contributing image collections ranging from 105 to 171.", + "bbox": [ + 212, + 587, + 785, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "AADB [19] database consists of 9,958 images from 190 users. We selected a test set from 22 users, each of which contributed individual image collections containing 110 to 190 images.", + "bbox": [ + 212, + 633, + 785, + 678 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "REAL-CUR [31] database consists of personal image collections from 14 users, totaling 2,871 images. Each user's collection ranges from 197 to 222 images.", + "bbox": [ + 212, + 679, + 785, + 709 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 722, + 455, + 737 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To obtain task vectors from the models fine-tuned on each of the six databases, it is essential to start from an identical pre-trained model. Recent approaches in IQA [18], GIAA [38], and PIAA [13] emphasize the importance of pre-training on large databases for training powerful image score regression models. Thus, we choose the publicly available OpenCLIP models ViT-B/16 and ViT-L/14 [15] as the pre-trained model. Full details regarding the architectural design and", + "bbox": [ + 212, + 750, + 785, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Scaling Up Personalized Image Aesthetic Assessment", + "bbox": [ + 377, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1e4528043c1bb1590afd5b72a1635316b30e575890abbc618bf6d0424f322697.jpg", + "table_caption": [ + "Table 2: Cross-database evaluation on the REAL-CUR [31] database. For the baseline approaches, we report the best cross-domain performances." + ], + "table_footnote": [], + "table_body": "
MethodSROCC
10-shot100-shot
PA-IAA [20]0.443±0.0040.562±0.013
BLG-PIAA [44]0.448±0.0070.578±0.015
PIAA-SOA [45]0.487±0.0030.589±0.015
TAPP-PIAA [21]-0.580
MTCL [41]0.495±0.0070.599±0.012
Ours0.577±0.0050.621±0.007
", + "bbox": [ + 272, + 174, + 727, + 314 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "training hyperparameters are provided in the appendix (See Section A.5 and A.6). We will also release the codes upon acceptance.", + "bbox": [ + 212, + 329, + 782, + 359 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "If not stated otherwise, we use $n = 6$ fine-tuned models to derive the task vectors. Each model comprises $L = 296$ layers when using the ViT-L/14 model and $L = 152$ layers when using the ViT-B/16 model. While the number of 'layers' usually refers to the number of transformer blocks, we include all computational elements within these blocks, such as linear layers and normalization layers. The total number of learnable parameters, $n \\times L = 1776$ and $n \\times L = 912$ , is significantly lower compared to a full fine-tuning approach, which would require training over 23 million parameters even for small ResNet-50 models [9]. We do not consider the layer-wise task vector itself as a learnable parameter, as obtaining these vectors is not repeated for each individual.", + "bbox": [ + 212, + 359, + 784, + 510 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3 Evaluation Metric for PIAA", + "text_level": 1, + "bbox": [ + 214, + 518, + 495, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "PIAA approaches are often evaluated through the Spearman's rank-order correlation coefficient [28] (SROCC), which is the correlation coefficient between the predicted and ground-truth rank variables. Specifically, given the ground-truth and predicted scores $y$ and $\\hat{y}$ , the SROCC is calculated as,", + "bbox": [ + 212, + 537, + 782, + 598 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {S R O C C} = 1 - \\frac {6 \\sum_ {i = 1} ^ {N} \\left(r _ {i} - \\hat {r} _ {i}\\right) ^ {2}}{N \\left(N ^ {2} - 1\\right)}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 604, + 785, + 641 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $r_i$ and $\\hat{r}_i$ represent the rank of the $i$ -th sample within the ground-truth and predicted scores, respectively, and $N$ denotes the number of samples.", + "bbox": [ + 214, + 647, + 782, + 678 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4 Cross-database Evaluation", + "text_level": 1, + "bbox": [ + 214, + 696, + 478, + 710 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "PIAA evaluation can be categorized based on whether training and testing occur within the same database (i.e., intra-database evaluation) or involve different databases (i.e., cross-database evaluation). In most real-world scenarios involving personal image collections, user-provided samples often diverge from the statistics or tendencies of the training database. Therefore, cross-database evaluation becomes pivotal in assessing the generalization capabilities of personalization approaches, which is an aspect that remains relatively underdeveloped in comparison to intra-database evaluations.", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "JY. Yun and J. Choo", + "bbox": [ + 271, + 114, + 415, + 127 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b33b2d69c6d376f26b8be734fa991affb263ffd6b49917cbcd0afa98b4e51e06.jpg", + "table_caption": [ + "Table 3: Cross-database evaluation on the AADB [19] database. Results marked with * indicate intra-database performance (i.e., directly trained on the AADB database)." + ], + "table_footnote": [], + "table_body": "
MethodSROCC
10-shot100-shot
PA-IAA [20]0.469±0.0020.524±0.006
BLG-PIAA [44]0.486±0.0040.536±0.006
PIAA-SOA [45]0.509±0.0030.557±0.007
TAPP-PIAA [21]-0.540
MTCL [41]0.533±0.0040.572±0.007
*TAPP-PIAA [21]*0.534±0.004*0.612±0.007
*MTCL [41]*0.540±0.005*0.622±0.007
Ours0.556±0.0040.654±0.007
", + "bbox": [ + 272, + 174, + 727, + 348 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For cross-database evaluation, we employ the AADB [19] and REAL-CUR [31] databases as the test databases, neither of which were among the six databases from which we derived our task vectors. Following established protocols [31], we randomly select $K$ samples for each individual to form the training set, also referred to as the support set, while the remaining images form the test set. In our experiments, we set $K$ as either 10-shot or 100-shot. Given the potential variability inherent in few-shot learning depending on the chosen $K$ support set, we conduct 10 independent trials for each user and report the average SROCC with the standard deviation.", + "bbox": [ + 212, + 364, + 785, + 500 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "As illustrated in Table 2 and Table 3, our method significantly surpasses existing PIAA models in cross-database performance for both the 10-shot and 100-shot protocols. Remarkably, it even exceeds the performance of models trained explicitly on the AADB database, as compared in Table 3. Such results underscore the exceptional generalization capabilities of our approach, in contrast to previous methods that are constrained by their inability to leverage GIAA and IQA databases. This superior performance is primarily due to the scalability of our approach, which effectively harnesses an extensive amount of 415,043 samples across multiple databases, each contributing unique insights and tendencies.", + "bbox": [ + 212, + 500, + 787, + 636 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.5 Intra-database Evaluation", + "text_level": 1, + "bbox": [ + 215, + 650, + 475, + 664 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For intra-database evaluation, we test our approach for the 37 individuals in the Flickr-AES [31] database, which is the most widely used PIAA database. Since the Flickr-AES [31] database is one of the six databases used to derive the task vectors, this experiment falls under the category of intra-database evaluation. However, we do not use the annotator information in the Flickr-AES.", + "bbox": [ + 212, + 672, + 785, + 748 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In Table 4, we report the average SROCC for each approach along with the standard deviation across 10 independent trials. These experiments underscore the effectiveness of training personalized task vectors through the use of multiple fine-tuned models. Our approach consistently outperforms existing methods that rely on meta-learning techniques or incorporate additional user attributes (e.g., personality traits) to enhance personalized score predictions.", + "bbox": [ + 212, + 750, + 785, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Scaling Up Personalized Image Aesthetic Assessment", + "bbox": [ + 377, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/1cae3247ac5d6fae7c9bb31a0d6a30ff932a222f7a8621b3646a735427309c2f.jpg", + "table_caption": [ + "Table 4: 10-shot and 100-shot personalization results on the Flickr-AES database [31], with the average SROCC of 37 users and standard deviation across 10 trials." + ], + "table_footnote": [], + "table_body": "
MethodSROCC
10-shot100-shot
PAM (attribute only) [31]0.511±0.0040.516±0.003
PAM (content only) [31]0.512±0.0020.516±0.010
PAM [31]0.513±0.0030.524±0.007
PASS [35]0.516±0.0030.521±0.007
USAR-PPR [25]0.521±0.0020.544±0.007
USAR-PAD [25]0.520±0.0030.537±0.003
USAR-PPR&PAD [25]0.525±0.0040.552±0.015
PA-IAA [20]0.543±0.0030.639±0.011
BLG-PIAA [44]0.561±0.0050.669±0.013
UG-PIAA [24]0.559±0.0020.660±0.013
PIAA-SOA [45]0.618±0.0060.691±0.015
TAPP-PIAA [21]0.591±0.0070.685±0.012
IM-PIAA [13]0.6200.708
MTCL [21]0.667±0.0050.737±0.014
Ours0.668±0.0040.748±0.012
", + "bbox": [ + 259, + 174, + 738, + 439 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 Analysis and Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 455, + 542, + 472 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We perform an in-depth analysis and ablation studies on each component of our approach, including model size, scalability, and coefficient initialization.", + "bbox": [ + 215, + 483, + 782, + 513 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.1 Task Similarities Across Databases", + "text_level": 1, + "bbox": [ + 215, + 532, + 547, + 546 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To gain insights into the derived task vectors, we visualize the similarity matrix of the task vectors in Figure 4. For simplicity, we aggregate the $L$ task vectors derived from each layer of a single fine-tuned model as a single task vector by concatenating them into one vector. In line with the findings provided by Ilharco et al. [14], task vectors are typically orthogonal, even though the databases are collected with similar intentions (i.e., GIAA and IQA). This observation highlights that treating each database as a unique task remains a valid choice, owing to variations in their data collection methodologies and the diverse nature of the populations from which they are sourced.", + "bbox": [ + 215, + 560, + 785, + 696 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.2 Architecture Choices and Variations", + "text_level": 1, + "bbox": [ + 215, + 720, + 560, + 734 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Selecting the architecture. Weight mixing techniques, including task vectors [14, 36, 37], have been recognized for their effectiveness when fine-tuning models pre-trained on large databases, such as the LAION-5B [33] or DataComp [7]. Vision Transformers (ViT) [5] is a suitable architecture for accommodating web-scale databases, in contrast to ResNets [9], which often struggle to achieve high performance despite training on large databases. Therefore, given", + "bbox": [ + 215, + 750, + 785, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "JY. Yun and J. Choo", + "bbox": [ + 271, + 114, + 415, + 127 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/257fc4a21f7d034c3116344456b35f40df8843c93106eee83c6151a17a9c5f3b.jpg", + "image_caption": [ + "Fig. 4: Similarity matrix between task vectors derived from each database. The similarities are measured with the cosine similarity." + ], + "image_footnote": [], + "bbox": [ + 292, + 150, + 707, + 369 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/1c8f0d69225da917bc66f44ce89bd8c6ba521b6a9eccc733ed4cde65566f9e00.jpg", + "table_caption": [ + "Table 5: Performance and inference speed of the ViT-B/16 model and the Best-fit FT method (i.e., directly fine-tuning ViTs) on the REAL-CUR [31] database." + ], + "table_footnote": [], + "table_body": "
MethodSROCCFPS
10-shot100-shot
Best-fit FT0.303±0.0080.394±0.00894.41
PIAA-SOA [45]0.487±0.0030.589±0.015-
MTCL [41]0.495±0.0070.599±0.012359.05
Ours (ViT-B/16)0.562±0.0040.607±0.007417.95
Ours0.577±0.0050.621±0.00794.41
", + "bbox": [ + 250, + 436, + 750, + 563 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "its scalability and performance advantages, we choose the publicly available pretrained ViT as our architecture for our experiments. Regardless of the chosen architecture, our parameter-efficient approach only trains a minimal set of coefficients corresponding to each layer-wise task vector.", + "bbox": [ + 212, + 575, + 784, + 636 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Scaling down for efficient computation. While our primary experiments focus on combining task vectors derived from ViT-L/14 models, our approach is adaptable to smaller models as well. As shown in Table 5, our approach yields strong results even on smaller models like ViT-B/16, which are well-suited for deployment on commercial computers. Although fast inference was not our primary goal, we achieve real-time inference speeds ( $\\geq 30$ FPS) on the RTX 3090, as our weight mixing technique combines multiple models into a single ViT model, eliminating the need to pass through multiple models. Acceleration of ViTs [22, 23, 43] beyond our implementation is also actively supported in recent frameworks [26] and communities [30], which ensures the practicality of our approach across a wide range of computing platforms such as mobile devices.", + "bbox": [ + 212, + 640, + 787, + 805 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Vision Transformers without task vectors. Our experimental results in Table 5 demonstrate that the performance gains are not solely attributed to", + "bbox": [ + 214, + 809, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Scaling Up Personalized Image Aesthetic Assessment", + "bbox": [ + 377, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/fb42a99854c3c82dcd8107a702ae1cfd7a13d78c182c356f371608997fa0abd9.jpg", + "image_caption": [ + "Fig. 5: Performance gains with respect to the number of databases used. The dotted lines indicate the previous state-of-the-art performance." + ], + "image_footnote": [], + "bbox": [ + 320, + 148, + 678, + 337 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "the inherent strengths of the ViT architecture. When directly fine-tuning the aesthetic assessment model that best matches the user's preference (i.e., Best-fit FT), we observe distinctly low personalization performances despite using a large pre-trained model. This highlights the pivotal role of keeping the layer-wise task vectors frozen during personalization, as altering the vectors would indicate discarding the rich preferences learned from a large database.", + "bbox": [ + 214, + 380, + 787, + 470 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.3 Scalability of Our Approach", + "text_level": 1, + "bbox": [ + 215, + 487, + 493, + 503 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "How does each database contribute? We assess the scalability of our approach by varying the number of task vectors. Specifically, we evaluate our approach on the REAL-CUR database [31] by increasing the number of databases used. Starting from the GIAA model trained on the AVA [27] database, we accumulate the databases in the following order: Flicker-AES [31], TAD66K [11], PARA [40], KonIQ-10K [12], and SPAQ [6]. As demonstrated in Figure 5, both 10-shot and 100-shot performance significantly improves as the number of task vectors increases. This indicates that our scalable approach benefits from the expansion of task vectors, which allows us to capture a broader range of image characteristics and nuances.", + "bbox": [ + 214, + 517, + 787, + 667 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "How can we increase the number of task vectors? The advantage of our approach can be amplified with the availability of additional image score regression databases. For example, scaling up the number of task vectors can be further pursued by incorporating readily available IQA [4,8,42] and GIAA [10,16] databases not included in our current study. Additionally, partitioning large databases, such as AVA [27] and PaQ-2-PiQ [42], can be an alternative method for generating additional task vectors.", + "bbox": [ + 214, + 672, + 787, + 777 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "These strategies for obtaining a diverse set of task vectors may lead to performance gains in the 10-shot evaluation, which is considered highly challenging due to the limited number of user-provided samples. However, we leave this as future work, given that our current work already utilizes the largest number of", + "bbox": [ + 214, + 779, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "JY. Yun and J. Choo", + "bbox": [ + 271, + 114, + 415, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/2cccc0294fe52d3ac029fda7cdad6ed747a7a01e8bde055c3b02e8f530a84405.jpg", + "image_caption": [ + "Fig. 6: 2D-histogram of coefficients and zero-shot SROCC pairs. The size of the coefficients are correlated with the relevance of the task." + ], + "image_footnote": [], + "bbox": [ + 348, + 148, + 660, + 349 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "samples for PIAA, 10 times larger than the most widely used and highly curated Flickr-AES database [31].", + "bbox": [ + 214, + 397, + 782, + 428 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.4 Task Relevance and Coefficient Initialization", + "text_level": 1, + "bbox": [ + 214, + 444, + 627, + 458 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We conduct an analysis to better understand the distinguishing characteristics of coefficients, specifically, what sets apart those with high values from those with low values. For a straightforward examination, we calculate the average coefficient for each task (i.e., database), averaged across all the layers of the model, which results in six average coefficients for each user ( $n = 6$ ). Subsequently, we assess the six zero-shot SROCC scores associated with the average coefficients and plot the histogram of the coefficient-SROCC pairs from all test users in Figure 6. We observe that coefficients for tasks with a high zero-shot SROCC are trained to have high values, while those for less-relevant tasks (i.e., low zero-shot SROCC) exhibit low values. This analysis justifies our adaptive initialization of the coefficients, which assigns large values to coefficients associated with closely related tasks.", + "bbox": [ + 212, + 470, + 787, + 651 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We investigate two alternative initialization strategies: a uniform initialization and a best-fit initialization. Both cases can be viewed as an extreme case of the adaptive initialization, which can be written in a generalized form:", + "bbox": [ + 214, + 652, + 787, + 698 + ], + "page_idx": 12 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {i} ^ {l} = \\frac {e ^ {\\mathrm {S R O C C} (S , \\theta_ {\\mathrm {f t} _ {i}}) / T}}{\\sum_ {j = 1} ^ {n} e ^ {\\mathrm {S R O C C} (S , \\theta_ {\\mathrm {f t} _ {j}}) / T}} \\quad \\text {f o r} l = 1, 2, \\dots , L, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 325, + 708, + 785, + 755 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "where $T$ is the temperature for the softmax. The uniform initialization corresponds to the case where the $T \\to \\infty$ , while the best-fit initialization is when $T \\to 0$ . As demonstrated in Table 6, the adaptive initialization outperforms other initialization strategies. Nevertheless, we consider further exploration of the initialization strategies as a potential avenue for future research.", + "bbox": [ + 214, + 763, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Scaling Up Personalized Image Aesthetic Assessment", + "bbox": [ + 377, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/644c0b258854b00eaf86f8c79446b2f86c85a038fb064de52623e7b00b740d67.jpg", + "table_caption": [ + "Table 6: Comparison between three different coefficient initialization strategies tested on the REAL-CUR [31] database." + ], + "table_footnote": [], + "table_body": "
MethodSROCC
10-shot100-shot
Uniform init. (T→∞)0.520±0.0020.602±0.002
Best-fit init. (T→0)0.464±0.0120.596±0.002
Adaptive init. (T=1)0.577±0.0050.621±0.007
", + "bbox": [ + 271, + 174, + 725, + 267 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Limitation", + "text_level": 1, + "bbox": [ + 215, + 276, + 354, + 291 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We have observed that when the zero-shot SROCC for all six image regression models are low for an individual (i.e., $\\mathrm{SROCC} \\approx 0$ ), our approach may struggle to identify an optimal combination of task vectors for that user. This phenomenon suggests that when there is a limited correlation between the user's preferences and the predictions of these models, the task vectors derived from them may be insufficient for finding an appropriate personalized task vector.", + "bbox": [ + 212, + 297, + 782, + 387 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "However, we find this case to be extremely rare (See appendix Figure 2-7.), with only 1 among 73 cases having a low SROCC under 0.2, demonstrating the broad coverage of our approach. Also, increasing the number of task vectors will effectively mitigate this issue, leveraging the inherent scalability of our approach, which can accommodate a growing number of task vectors to improve personalization even in challenging scenarios.", + "bbox": [ + 212, + 388, + 784, + 479 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "7 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 492, + 359, + 508 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We address the scalability issue in PIAA, a critical yet previously unaddressed aspect necessary for enhancing the generalization performance. Unlike previous studies limited to training on PIAA databases curated by individual annotators, our approach enables the flexible use of multiple image score regression databases, thereby providing greater scalability and generalization capabilities.", + "bbox": [ + 212, + 513, + 782, + 588 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We demonstrate that GIAA and IQA databases exhibit distinct personalization potentials, which have motivated us to combine GIAA and IQA models to achieve desired behaviors. By introducing learnable parameters for optimal model combinations, we enable the precise personalization of models to individual users' aesthetic preferences. Our approach significantly outperforms existing approaches, demonstrating exceptional efficacy in generalizing to unseen domains which is a critical requirement for real-world applications. This work opens up new avenues for personalized image aesthetic assessment, offering valuable insights and practical solutions for this challenging task.", + "bbox": [ + 212, + 589, + 784, + 724 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 215, + 733, + 356, + 747 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This work was supported by the Institute for Information & communications Technology Promotion(IITP) grant funded by the Korea government(MSIT) (No.RS-2019-II190075 Artificial Intelligence Graduate School Program(KAIST)), the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. NRF-2022R1A2B5B02001913), and the Institute of Information & communications Technology Promotion (IITP) grant funded by the Korea government (MSIT) (RS-2021-II212068, Artificial Intelligence Innovation Hub).", + "bbox": [ + 212, + 751, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "JY. Yun and J. Choo", + "bbox": [ + 271, + 114, + 415, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 323, + 160 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Al-Shedivat, M., Li, L., Xing, E., Talwalkar, A.: On data efficiency of meta-learning. In: International Conference on Artificial Intelligence and Statistics. PMLR (2021)", + "2. Bradley, R.A., Terry, M.E.: Rank analysis of incomplete block designs: I. the method of paired comparisons. Biometrika (1952)", + "3. Chefer, H., Gur, S., Wolf, L.: Transformer interpretability beyond attention visualization. In: CVPR (2021)", + "4. Ciancio, A., da Silva, E.A., Said, A., Samadani, R., Obrador, P., et al.: No-reference blur assessment of digital pictures based on multifeature classifiers. IEEE Transactions on Image Processing (2010)", + "5. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2020)", + "6. Fang, Y., Zhu, H., Zeng, Y., Ma, K., Wang, Z.: Perceptual quality assessment of smartphone photography. In: CVPR (2020)", + "7. Gadre, S.Y., Ilharco, G., Fang, A., Hayase, J., Smyrnis, G., Nguyen, T., Marten, R., Wortsman, M., Ghosh, D., Zhang, J., et al.: Datacomp: In search of the next generation of multimodal datasets. arXiv preprint arXiv:2304.14108 (2023)", + "8. Ghadiyaram, D., Bovik, A.C.: Massive online crowdsourced study of subjective and objective picture quality. IEEE Transactions on Image Processing (2015)", + "9. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016)", + "10. He, S., Ming, A., Li, Y., Sun, J., Zheng, S., Ma, H.: Thinking image color aesthetics assessment: Models, datasets and benchmarks. In: ICCV (2023)", + "11. He, S., Zhang, Y., Xie, R., Jiang, D., Ming, A.: Rethinking image aesthetics assessment: Models, datasets and benchmarks. IJCAI (2022)", + "12. Hosu, V., Lin, H., Sziranyi, T., Saupe, D.: Koniq-10k: An ecologically valid database for deep learning of blind image quality assessment. IEEE Transactions on Image Processing (2020)", + "13. Hou, J., Lin, W., Yue, G., Liu, W., Zhao, B.: Interaction-matrix based personalized image aesthetics assessment. IEEE Transactions on Multimedia (2022)", + "14. Ilharco, G., Ribeiro, M.T., Wortsman, M., Schmidt, L., Hajishirzi, H., Farhadi, A.: Editing models with task arithmetic. In: ICLR (2023)", + "15. Ilharco, G., Wortsman, M., Wightman, R., Gordon, C., Carlini, N., Taori, R., Dave, A., Shankar, V., Namkoong, H., Miller, J., Hajishirzi, H., Farhadi, A., Schmidt, L.: Openclip (Jul 2021). https://doi.org/10.5281/zenodo.5143773, https://doi.org/10.5281/zenodo.5143773", + "16. Jin, X., Wu, L., Zhao, G., Li, X., Zhang, X., Ge, S., Zou, D., Zhou, B., Zhou, X.: Aesthetic attributes assessment of images. In: ACM MM (2019)", + "17. Karlsson, K., Jiang, W., Zhang, D.Q.: Mobile photo album management with multiscale timeline. In: ACM MM (2014)", + "18. Ke, J., Wang, Q., Wang, Y., Milanfar, P., Yang, F.: Musiq: Multi-scale image quality transformer. In: ICCV (2021)", + "19. Kong, S., Shen, X., Lin, Z., Mech, R., Fowlkes, C.: Photo aesthetics ranking network with attributes and content adaptation. In: ECCV. Springer (2016)", + "20. Li, L., Zhu, H., Zhao, S., Ding, G., Lin, W.: Personality-assisted multi-task learning for generic and personalized image aesthetics assessment. IEEE Transactions on Image Processing 29 (2020)" + ], + "bbox": [ + 217, + 178, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Scaling Up Personalized Image Aesthetic Assessment", + "bbox": [ + 377, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "21. Li, Y., Yang, Y., Li, H., Chen, H., Xu, L., Li, L., Li, Y., Guo, Y.: Transductive aesthetic preference propagation for personalized image aesthetics assessment. In: ACM MM. ACM (2022)", + "22. Lin, Y., Zhang, T., Sun, P., Li, Z., Zhou, S.: Fq-vit: Post-training quantization for fully quantized vision transformer. IJCAI (2022)", + "23. Liu, Z., Wang, Y., Han, K., Zhang, W., Ma, S., Gao, W.: Post-training quantization for vision transformer. NeurIPS (2021)", + "24. Lv, P., Fan, J., Nie, X., Dong, W., Jiang, X., Zhou, B., Xu, M., Xu, C.: User-guided personalized image aesthetic assessment based on deep reinforcement learning. IEEE Transactions on Multimedia (2021)", + "25. Lv, P., Wang, M., Xu, Y., Peng, Z., Sun, J., Su, S., Zhou, B., Xu, M.: Usar: An interactive user-specific aesthetic ranking framework for images. In: ACM MM. pp. 1328-1336 (2018)", + "26. Microsoft: Deepspeed (2023), https://www.deepspeed.ai/", + "27. Murray, N., Marchesotti, L., Perronnin, F.: Ava: A large-scale database for aesthetic visual analysis. In: CVPR. IEEE (2012)", + "28. Myers, J.L., Well, A.D., Lorch Jr, R.F.: Research design and statistical analysis. Routledge (2013)", + "29. Naseer, M.M., Ranasinghe, K., Khan, S.H., Hayat, M., Shahbaz Khan, F., Yang, M.H.: Intriguing properties of vision transformers. NeurIPS (2021)", + "30. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. NeurIPS (2019)", + "31. Ren, J., Shen, X., Lin, Z., Mech, R., Foran, D.J.: Personalized image aesthetics. In: ICCV (Oct 2017)", + "32. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. 2022 IEEE. In: CVPR (2021)", + "33. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. NeurIPS (2022)", + "34. Wallace, B., Gokul, A., Ermon, S., Naik, N.: End-to-end diffusion latent optimization improves classifier guidance. ICCV (2023)", + "35. Wang, G., Yan, J., Qin, Z.: Collaborative and attentive learning for personalized image aesthetic assessment. In: IJCAI. pp. 957-963 (2018)", + "36. Wortsman, M., Ilharco, G., Gadre, S.Y., Roelofs, R., Gontijo-Lopes, R., Morcos, A.S., Namkoong, H., Farhadi, A., Carmon, Y., Kornblith, S., et al.: Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In: ICML. PMLR (2022)", + "37. Wortsman, M., Ilharco, G., Li, M., Kim, J.W., Hajishirzi, H., Farhadi, A., Namkoong, H., Schmidt, L.: Robust fine-tuning of zero-shot models. 2022 IEEE. In: CVPR (2021)", + "38. Xu, L., Xu, J., Yang, Y., Huang, Y., Xie, Y., Li, Y.: Clip brings better features to visual aesthetics learners. arXiv preprint arXiv:2307.15640 (2023)", + "39. Yan, X., Shao, F., Chen, H., Jiang, Q.: Hybrid cnn-transformer based meta-learning approach for personalized image aesthetics assessment. Journal of Visual Communication and Image Representation 98 (2024)", + "40. Yang, Y., Xu, L., Li, L., Qie, N., Li, Y., Zhang, P., Guo, Y.: Personalized image aesthetics assessment with rich attributes. In: CVPR (2022)", + "41. Yang, Z., Li, L., Yang, Y., Li, Y., Lin, W.: Multi-level transitional contrast learning for personalized image aesthetics assessment. IEEE Transactions on Multimedia (2023)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "JY. Yun and J. Choo", + "bbox": [ + 271, + 114, + 415, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "42. Ying, Z., Niu, H., Gupta, P., Mahajan, D., Ghadiyaram, D., Bovik, A.: From patches to pictures (paq-2-piq): Mapping the perceptual space of picture quality. In: CVPR (2020)", + "43. Yuan, Z., Xue, C., Chen, Y., Wu, Q., Sun, G.: Ptq4vit: Post-training quantization for vision transformers with twin uniform quantization. In: ECCV. Springer (2022)", + "44. Zhu, H., Li, L., Wu, J., Zhao, S., Ding, G., Shi, G.: Personalized image aesthetics assessment via meta-learning with bilevel gradient optimization. IEEE Transactions on Cybernetics (2020)", + "45. Zhu, H., Zhou, Y., Li, L., Li, Y., Guo, Y.: Learning personalized image aesthetics from subjective and objective attributes. IEEE Transactions on Multimedia (2021)" + ], + "bbox": [ + 215, + 146, + 787, + 286 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Scaling Up Personalized Image Aesthetic Assessment", + "bbox": [ + 377, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Scaling Up Personalized Image Aesthetic Assessment via Task Vector Customization/77e3f8ef-24e2-4d25-ba6f-999ef6bc7191_model.json b/2024/Scaling Up Personalized Image Aesthetic Assessment via Task Vector Customization/77e3f8ef-24e2-4d25-ba6f-999ef6bc7191_model.json new file mode 100644 index 0000000000000000000000000000000000000000..355b9e2e19c25afede0cd49d356fb0b81e2f86e2 --- /dev/null +++ b/2024/Scaling Up Personalized Image Aesthetic Assessment via Task Vector Customization/77e3f8ef-24e2-4d25-ba6f-999ef6bc7191_model.json @@ -0,0 +1,2313 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.258, + 0.141, + 0.747, + 0.185 + ], + "angle": 0, + "content": "Scaling Up Personalized Image Aesthetic Assessment via Task Vector Customization" + }, + { + "type": "text", + "bbox": [ + 0.382, + 0.213, + 0.621, + 0.228 + ], + "angle": 0, + "content": "Jooyeol Yun and Jaegul Choo" + }, + { + "type": "text", + "bbox": [ + 0.293, + 0.24, + 0.709, + 0.269 + ], + "angle": 0, + "content": "Korea Advanced Institute of Science and Technology (KAIST) blizzard072@kaist.ac.kr jchoo@kaist.ac.kr" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.284, + 0.744, + 0.548 + ], + "angle": 0, + "content": "Abstract. The task of personalized image aesthetic assessment seeks to tailor aesthetic score prediction models to match individual preferences with just a few user-provided inputs. However, the scalability and generalization capabilities of current approaches are considerably restricted by their reliance on an expensive curated database. To overcome this long-standing scalability challenge, we present a unique approach that leverages readily available databases for general image aesthetic assessment and image quality assessment. Specifically, we view each database as a distinct image score regression task that exhibits varying degrees of personalization potential. By determining optimal combinations of task vectors, known to represent specific traits of each database, we successfully create personalized models for individuals. This approach of integrating multiple models allows us to harness a substantial amount of data. Our extensive experiments demonstrate the effectiveness of our approach in generalizing to previously unseen domains—a challenge previous approaches have struggled to achieve—making it highly applicable to real-world scenarios. Our novel approach significantly advances the field by offering scalable solutions for personalized aesthetic assessment and establishing high standards for future research." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.564, + 0.377, + 0.579 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.594, + 0.788, + 0.715 + ], + "angle": 0, + "content": "Personalized image aesthetic assessment (PIAA) is an emerging field dedicated to developing aesthetic score prediction models that closely match an individual's aesthetic preferences based on a small set of user-provided samples. While they function as simple models that assign scores based on aesthetic quality, PIAA models enhance user experience in digital environments by personalizing tasks such as managing photo albums [17], curating web-scale databases like LAION-Aesthetics [33], and even guiding generative models such as Stable Diffusion [32] to create images that align with individual tastes [34]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.715, + 0.788, + 0.821 + ], + "angle": 0, + "content": "In recent years, there has been a notable trend in PIAA [21, 41, 44] that involves the integration of meta-learning techniques, as personalization is inherently related to few-shot learning. Specifically, these approaches entail pretraining a model on a database consisting of image collections from diverse individuals. However, these strategies encounter significant challenges in terms of scalability and generalization to unseen domains, primarily due to the prohibitive cost associated with data collection." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.763, + 0.841 + ], + "angle": 0, + "content": "1 https://yeolj00.github.io/personal-projects/personalized-aesthetics/" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.416, + 0.127 + ], + "angle": 0, + "content": "JY. Yun and J. Choo" + }, + { + "type": "image", + "bbox": [ + 0.302, + 0.144, + 0.699, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.273, + 0.785, + 0.302 + ], + "angle": 0, + "content": "Fig. 1: Comparison between PIAA and other image regression databases. Our approach has the flexibility to leverage any image regression databases." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.31, + 0.785, + 0.461 + ], + "angle": 0, + "content": "While meta-learning typically demands an extensive number of meta-training tasks, often exceeding 10,000, to achieve high performance [1], currently available PIAA databases [19, 31] offer a limited number of tasks fewer than 200. Even with the recently collected PIAA database [40], which contains rich attributes from each annotator, the number of training tasks remains below 500. In the context of PIAA, a single meta-training task corresponds to learning the preferences of a single individual. Scaling such a database would require obtaining a vast number of images and corresponding personalized aesthetic scores from numerous individuals, which poses severe limitations in the scalability and generalization of existing approaches." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.787, + 0.583 + ], + "angle": 0, + "content": "In this paper, we unveil a data-rich approach for personalizing aesthetic assessment models by leveraging existing general image aesthetic assessment (GIAA) and image quality assessment (IQA) databases. In contrast to prior work [20, 21, 41, 44, 45], our approach is not limited to training on databases that track individual annotators, as depicted in Figure 1. Instead, we have the flexibility to utilize multiple image score regression databases, which are readily available. By employing broader resources, we overcome the scalability and generalization challenges that have long been unaddressed in this field." + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.59, + 0.71, + 0.716 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.718, + 0.785, + 0.759 + ], + "angle": 0, + "content": "Fig.2: Zero-shot personalization performance for four users in the Flickr-AES database [31]. Each colored bar refers to the personalization performance of a model trained on a specific database." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.786, + 0.841 + ], + "angle": 0, + "content": "The key idea is to view each GIAA and IQA data as a unique image score regression task. This perspective arises from the distinct characteristics and tendencies of each database due to different data collection procedures and the population from which they originate. As visualized in Figure 2, image score regression models trained on four different GIAA [27,31] and IQA [6,12] databases" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.379, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Scaling Up Personalized Image Aesthetic Assessment" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.237 + ], + "angle": 0, + "content": "exhibit varying levels of personalization for users within the PIAA database. For instance, User 3's aesthetic preferences closely align with the score regression model trained on the Flickr-AES [31] while diverging from the model trained on the SPAQ [6] database. In other words, even without any personalized training, we find that general image assessment models inherently possess different degrees of personalization capabilities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.24, + 0.785, + 0.36 + ], + "angle": 0, + "content": "In light of these insights, we extend an emerging approach known as task arithmetic [14], which was originally devised for developing multi-task models. This method involves directly adding or subtracting the weights of models fine-tuned on specific tasks. In our context, task vectors from diverse image assessment models encapsulate distinct features and behaviors related to specific tasks, such as recognizing different aspects of image quality or aesthetics. By carefully combining task vectors, we explore the potential to selectively amplify or refine the capabilities of a pre-trained model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.363, + 0.785, + 0.543 + ], + "angle": 0, + "content": "However, there has been limited research on determining the optimal adjustments necessary to achieve a certain behavior. Thus, we present a personalization approach by introducing trainable coefficients for each task vector derived from models trained on distinct GIAA and IQA databases, thereby determining the weights for combining these vectors. Once the coefficients are trained with user-provided inputs, the weighted sum of task vectors creates a model tailored to the user's aesthetic preferences. Since task vectors obtained from large databases already capture preferences across various themes and contents, we find that training only the coefficients is sufficient for precise personalization, making our approach highly parameter-efficient. To the best of our knowledge, our approach marks the first to customize task vectors for specific model behaviors that were previously unattainable." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.545, + 0.785, + 0.666 + ], + "angle": 0, + "content": "By carefully customizing task vectors, we incorporate large databases into the training process, granting our approach a strong generalization capability for unseen domains, a scenario often encountered in real-world applications. Furthermore, we find that our approach is also practical for learning a user's aesthetic preference with only a few samples, as our approach of merging task vectors guides the model to make well-informed updates and prevents overfitting. We provide extensive experiments demonstrating the efficacy of our approach for personalization in diverse scenarios." + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.668, + 0.472, + 0.682 + ], + "angle": 0, + "content": "Our contributions are threefold:" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.694, + 0.784, + 0.739 + ], + "angle": 0, + "content": "- We introduce a novel data-rich approach for PIAA that tackles the long-standing scalability challenge, eliminating the dependency on expensive, manually-curated databases that have hindered progress in this field." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.745, + 0.785, + 0.789 + ], + "angle": 0, + "content": "- Demonstrating exceptional effectiveness in cross-database evaluations, our method showcases robust generalization capabilities that surpass those of existing approaches, setting high standards for future work." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.795, + 0.784, + 0.839 + ], + "angle": 0, + "content": "- By learning optimal combinations of task vectors, we present a parameter-efficient approach for precise personalization, leveraging the comprehensive information embedded in these vectors." + }, + { + "type": "list", + "bbox": [ + 0.228, + 0.694, + 0.785, + 0.839 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.416, + 0.127 + ], + "angle": 0, + "content": "JY. Yun and J. Choo" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.388, + 0.161 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.173, + 0.605, + 0.189 + ], + "angle": 0, + "content": "2.1 Personalized Image Aesthetic Assessment" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.198, + 0.787, + 0.365 + ], + "angle": 0, + "content": "Personalized image aesthetic assessment (PIAA) [13,20,21,24,25,35,39-41,44,45] aims to learn the personal aesthetic preference of individual users based on the image-score pairs from one's image collection. Since collecting a large number of samples from a single user is impractical, PIAA is closely linked to few-shot learning methods. One line of work [20,40,45] incorporates additional attributes (e.g., personality traits, age, and gender) to facilitate user-specific aesthetic predictions. These approaches rely on the correlation between user attributes and aesthetic preferences, which may not always be clearly defined. Consequently, many high-performing approaches for PIAA [21, 41, 44] leverage meta-learning techniques to train models that can be easily fine-tuned to accommodate the aesthetic preferences of new users with only a limited number of samples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.365, + 0.787, + 0.532 + ], + "angle": 0, + "content": "Nevertheless, applying meta-learning techniques to PIAA encounters scalability challenges owing to the limited number of tasks available during the meta-learning phase. In traditional meta-learning, achieving a high performance typically requires a substantial number of meta-training tasks [1], often exceeding 10,000. In PIAA databases [19,31], the number of training tasks ranges from 100 to fewer than 500, which can impose limitations on the performance. Expanding the task pool in PIAA is prohibitively expensive since each task corresponds to the image collection of a single user, posing significant scalability challenges that has long been unaddressed. We tackle this issue in previous PIAA approaches and present a scalable fine-tuning method that leverages existing general image aesthetic assessment (GIAA) and image quality assessment (IQA) databases." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.549, + 0.546, + 0.563 + ], + "angle": 0, + "content": "2.2 Task Vectors and Task Arithmetic" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.574, + 0.787, + 0.62 + ], + "angle": 0, + "content": "Ilharco et al. [14] presented a new paradigm for creating multi-task models, utilizing the parameters of multiple fine-tuned models. They define the task vector \\(\\tau\\) as the difference between the fine-tuned and pre-trained weights," + }, + { + "type": "equation", + "bbox": [ + 0.451, + 0.633, + 0.553, + 0.648 + ], + "angle": 0, + "content": "\\[\n\\tau = \\theta_ {\\mathrm {f t}} - \\theta_ {\\mathrm {p r e}},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.787, + 0.749 + ], + "angle": 0, + "content": "where \\(\\theta_{\\mathrm{ft}}\\) and \\(\\theta_{\\mathrm{pre}}\\) represent the weights of the fine-tuned and pre-trained model, respectively. They demonstrate that adding and subtracting task vectors to pretrained models can encourage or suppress the ability to perform specific tasks. For example, negating the task vector derived from a language model fine-tuned on toxic data reduces the frequency of generating toxic text. However, it remains unclear what the optimal adjustments are for achieving a specific behavior." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In our paper, we investigate methods to determine necessary adjustments to achieve a certain behavior defined by a few user-provided samples. By introducing learnable coefficients to task vectors, we effectively customize models to closely match individual preferences. We also demonstrate that finding the optimal coefficients is possible with only a few samples, as the task vectors encompass comprehensive knowledge derived from large databases." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.378, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Scaling Up Personalized Image Aesthetic Assessment" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.78, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.373, + 0.214, + 0.638, + 0.226 + ], + "angle": 0, + "content": "(a) Obtain task vectors from various databases" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.228, + 0.789, + 0.347 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.36, + 0.348, + 0.651, + 0.362 + ], + "angle": 0, + "content": "(b) Learn optimal combinations for personalization" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.369, + 0.789, + 0.426 + ], + "angle": 0, + "content": "Fig. 3: (a) Multiple models are fine-tuned from a single pre-trained model on different tasks to obtain task vectors. For simplicity, we omit the discussion of layers. (b) Given a small number of user-provided samples, the coefficients corresponding to each task vector are optimized with the rank loss." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.43, + 0.331, + 0.446 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.455, + 0.75, + 0.47 + ], + "angle": 0, + "content": "3.1 Obtaining Layer-wise Task Vectors from Various Databases" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.483, + 0.788, + 0.589 + ], + "angle": 0, + "content": "To obtain task vectors, we train multiple models with the same architecture on various tasks (i.e., different databases), as illustrated in Figure 3 (a). Specifically, we fine-tune the weights from a pre-trained backbone, adjusting the parameters from \\(\\theta_{\\mathrm{pre}}\\) to \\(\\theta_{\\mathrm{ft}_i}\\) for each task \\(i\\). Moreover, while the specific roles of model layers in capturing unique features are inherently complex, their distinct contributions are undeniable [3,29]. Thus, we compute task vectors across all layers, enhancing the flexibility of our approach for personalization (See appendix Section A.2)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.59, + 0.787, + 0.62 + ], + "angle": 0, + "content": "Accordingly, we expand the concept of original task vectors to include layerwise differentiation by," + }, + { + "type": "equation", + "bbox": [ + 0.443, + 0.62, + 0.786, + 0.641 + ], + "angle": 0, + "content": "\\[\n\\tau_ {i} ^ {l} = \\theta_ {\\mathrm {f t} _ {i}} ^ {l} - \\theta_ {\\mathrm {p r e}} ^ {l}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.648, + 0.787, + 0.695 + ], + "angle": 0, + "content": "where \\(\\tau_{i}^{l}\\) is the task vector for the \\(l\\)-th layer of the \\(i\\)-th task. Each vector comprehensively integrates patterns learned from large databases, acting as basis vectors for robust and precise personalization." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.696, + 0.787, + 0.743 + ], + "angle": 0, + "content": "It is important to note that this process is conducted only once and is not required again for personalizing to individual users, thereby not contributing to the count of learnable parameters for personalization." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.766, + 0.688, + 0.782 + ], + "angle": 0, + "content": "3.2 Learning Optimal Combinations for Personalization" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.795, + 0.787, + 0.843 + ], + "angle": 0, + "content": "Given the layer-wise task vectors, our goal is to find a combination of these vectors that accurately reflects a user's aesthetic preferences. To achieve this, we introduce learnable coefficients \\(\\alpha_{i}^{l}\\) for each task vector \\(\\tau_{i}^{l}\\). Once the coefficients" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.416, + 0.128 + ], + "angle": 0, + "content": "JY. Yun and J. Choo" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.782, + 0.162 + ], + "angle": 0, + "content": "are trained, the weights of the personalized model for each layer are derived as," + }, + { + "type": "equation", + "bbox": [ + 0.359, + 0.175, + 0.786, + 0.213 + ], + "angle": 0, + "content": "\\[\n\\theta_ {\\mathrm {p}} ^ {l} = \\theta_ {\\mathrm {p r e}} ^ {l} + \\sum_ {i = 1} ^ {n} \\alpha_ {i} ^ {l} \\tau_ {i} ^ {l} \\quad \\text {f o r} l = 1, 2, \\dots , L, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.784, + 0.314 + ], + "angle": 0, + "content": "where \\( n \\) is the number of tasks, and \\( L \\) is the number of layers in the model. We keep all parameters other than the \\( n \\times L \\) coefficients frozen. This is essential since the task vectors contain comprehensive insights into aesthetic preferences, enabling the model to make well-informed updates. By harnessing the rich information embedded in these vectors, our method efficiently learns complex aesthetic preferences from merely a few user-provided samples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.315, + 0.785, + 0.405 + ], + "angle": 0, + "content": "Our objective function is designed to penalize the misordering of aesthetic scores. Specifically, we adopt the Bradley-Terry model [2], which estimates the probability of pairwise comparison between two samples. We frame the prediction of the two scores as a binary classification problem, where the objective is to predict which sample has the higher score. The coefficients of each task vector are trained to minimize the following loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.363, + 0.418, + 0.786, + 0.448 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {R}} = - \\log \\frac {\\exp \\hat {y} _ {1}}{\\exp \\hat {y} _ {1} + \\exp \\hat {y} _ {2}} (y _ {1} \\geq y _ {2}), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.457, + 0.788, + 0.519 + ], + "angle": 0, + "content": "where \\(\\hat{y}_1\\) and \\(\\hat{y}_2\\) represent the aesthetic score predictions for images \\(x_1\\) and \\(x_2\\), whose ground-truth scores are \\(y_1\\) and \\(y_2\\). We prefer this loss function over the Earth Mover Distance (EMD) or the MSE loss, due to its sample efficiency and fast convergence (See appendix Section A.3)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.541, + 0.542, + 0.557 + ], + "angle": 0, + "content": "3.3 Adaptive Coefficient Initialization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.567, + 0.784, + 0.67 + ], + "angle": 0, + "content": "We observe that appropriately initializing the coefficients plays a crucial role in determining the coefficients that best align with the user's aesthetic preferences. When a high coefficient is assigned to a specific task vector, it indicates a strong correlation between the user's aesthetic preference and that particular task. Thus, we adaptively initialize the coefficients based on the zero-shot personalization performances of each image score regression model associated with the task vector." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.673, + 0.785, + 0.735 + ], + "angle": 0, + "content": "To do this, we calculate the Spearman's rank-order correlation coefficient (SROCC) [28] between the user-provided samples and the scores predicted by each image score regression model. We then apply the softmax function to initialize the coefficients \\(\\alpha\\) as," + }, + { + "type": "equation", + "bbox": [ + 0.339, + 0.745, + 0.786, + 0.791 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {i} ^ {l} = \\frac {e ^ {\\mathrm {S R O C C} (S , \\theta_ {\\mathrm {f t} _ {i}})}}{\\sum_ {j = 1} ^ {n} e ^ {\\mathrm {S R O C C} (S , \\theta_ {\\mathrm {f t} _ {j}})}} \\quad \\mathrm {f o r} l = 1, 2, \\ldots , L, \\qquad \\qquad (4)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.799, + 0.788, + 0.846 + ], + "angle": 0, + "content": "where \\(\\mathrm{SROCC}(S,\\theta_{\\mathrm{ft}_i})\\) indicates the SROCC calculated on the user-provided set \\(S\\) using the model with parameters \\(\\theta_{\\mathrm{ft}_i}\\). Note that we initialize \\(\\alpha\\) values for a single task \\(i\\) with the same value for all layers." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.378, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Scaling Up Personalized Image Aesthetic Assessment" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.784, + 0.172 + ], + "angle": 0, + "content": "Table 1: Description of the databases included in our experiments. The AADB and REAL-CUR databases are exclusively reserved as a test database." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.175, + 0.777, + 0.33 + ], + "angle": 0, + "content": "
DatabaseRegression taskAnnotator tagNumber of images
KonIQ-10K [12]IQA×10,073
SPAQ [6]IQA×11,125
AVA [27]GIAA×255,500
TAD66K [11]GIAA×67,125
Flickr-AES [31]GIAA/PIAA✓40,000
PARA [40]GIAA/PIAA✓31,220
AADB [19]GIAA/PIAA✓9,958
REAL-CUR [31]PIAA✓2,871
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.343, + 0.375, + 0.359 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.372, + 0.421, + 0.386 + ], + "angle": 0, + "content": "4.1 Database Overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.391, + 0.787, + 0.572 + ], + "angle": 0, + "content": "The key advantage of our approach is the flexibility to utilize multiple image score regression databases freely, even those without annotator tags. In our experiments, we fine-tune models using six different databases in Table 1, reserving the AADB [19] and REAL-CUR [31] database exclusively for testing. While the Flickr-AES [31] and PARA [40] databases do contain annotator tags, our approach does not require distinguishing between individuals. Therefore, we treat these two databases as GIAA databases. In total, we effectively leverage 415,043 samples from six databases, marking the most extensive utilization of samples in PIAA within our knowledge. Detailed descriptions and performance on these databases are provided in the appendix (See Section A.5). However, our approach is not limited to training solely on these six databases, having the scalability to include additional IQA and GIAA databases [4,8,10,16,42]." + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.573, + 0.761, + 0.588 + ], + "angle": 0, + "content": "For testing, we use the following three most widely-used PIAA test sets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.588, + 0.786, + 0.634 + ], + "angle": 0, + "content": "Flickr-AES [31] database consists of 40,000 images collected from 210 users. For testing purposes, we utilized a test set containing 4,737 images rated by 37 users, with each user contributing image collections ranging from 105 to 171." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.634, + 0.786, + 0.679 + ], + "angle": 0, + "content": "AADB [19] database consists of 9,958 images from 190 users. We selected a test set from 22 users, each of which contributed individual image collections containing 110 to 190 images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.68, + 0.786, + 0.71 + ], + "angle": 0, + "content": "REAL-CUR [31] database consists of personal image collections from 14 users, totaling 2,871 images. Each user's collection ranges from 197 to 222 images." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.723, + 0.456, + 0.738 + ], + "angle": 0, + "content": "4.2 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.787, + 0.84 + ], + "angle": 0, + "content": "To obtain task vectors from the models fine-tuned on each of the six databases, it is essential to start from an identical pre-trained model. Recent approaches in IQA [18], GIAA [38], and PIAA [13] emphasize the importance of pre-training on large databases for training powerful image score regression models. Thus, we choose the publicly available OpenCLIP models ViT-B/16 and ViT-L/14 [15] as the pre-trained model. Full details regarding the architectural design and" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.416, + 0.128 + ], + "angle": 0, + "content": "JY. Yun and J. Choo" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.787, + 0.173 + ], + "angle": 0, + "content": "Table 2: Cross-database evaluation on the REAL-CUR [31] database. For the baseline approaches, we report the best cross-domain performances." + }, + { + "type": "table", + "bbox": [ + 0.273, + 0.175, + 0.728, + 0.315 + ], + "angle": 0, + "content": "
MethodSROCC
10-shot100-shot
PA-IAA [20]0.443±0.0040.562±0.013
BLG-PIAA [44]0.448±0.0070.578±0.015
PIAA-SOA [45]0.487±0.0030.589±0.015
TAPP-PIAA [21]-0.580
MTCL [41]0.495±0.0070.599±0.012
Ours0.577±0.0050.621±0.007
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.33, + 0.784, + 0.36 + ], + "angle": 0, + "content": "training hyperparameters are provided in the appendix (See Section A.5 and A.6). We will also release the codes upon acceptance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.36, + 0.785, + 0.511 + ], + "angle": 0, + "content": "If not stated otherwise, we use \\( n = 6 \\) fine-tuned models to derive the task vectors. Each model comprises \\( L = 296 \\) layers when using the ViT-L/14 model and \\( L = 152 \\) layers when using the ViT-B/16 model. While the number of 'layers' usually refers to the number of transformer blocks, we include all computational elements within these blocks, such as linear layers and normalization layers. The total number of learnable parameters, \\( n \\times L = 1776 \\) and \\( n \\times L = 912 \\), is significantly lower compared to a full fine-tuning approach, which would require training over 23 million parameters even for small ResNet-50 models [9]. We do not consider the layer-wise task vector itself as a learnable parameter, as obtaining these vectors is not repeated for each individual." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.52, + 0.496, + 0.533 + ], + "angle": 0, + "content": "4.3 Evaluation Metric for PIAA" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.784, + 0.599 + ], + "angle": 0, + "content": "PIAA approaches are often evaluated through the Spearman's rank-order correlation coefficient [28] (SROCC), which is the correlation coefficient between the predicted and ground-truth rank variables. Specifically, given the ground-truth and predicted scores \\( y \\) and \\( \\hat{y} \\), the SROCC is calculated as," + }, + { + "type": "equation", + "bbox": [ + 0.382, + 0.606, + 0.786, + 0.642 + ], + "angle": 0, + "content": "\\[\n\\mathrm {S R O C C} = 1 - \\frac {6 \\sum_ {i = 1} ^ {N} \\left(r _ {i} - \\hat {r} _ {i}\\right) ^ {2}}{N \\left(N ^ {2} - 1\\right)}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.648, + 0.784, + 0.679 + ], + "angle": 0, + "content": "where \\( r_i \\) and \\( \\hat{r}_i \\) represent the rank of the \\( i \\)-th sample within the ground-truth and predicted scores, respectively, and \\( N \\) denotes the number of samples." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.697, + 0.48, + 0.712 + ], + "angle": 0, + "content": "4.4 Cross-database Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "PIAA evaluation can be categorized based on whether training and testing occur within the same database (i.e., intra-database evaluation) or involve different databases (i.e., cross-database evaluation). In most real-world scenarios involving personal image collections, user-provided samples often diverge from the statistics or tendencies of the training database. Therefore, cross-database evaluation becomes pivotal in assessing the generalization capabilities of personalization approaches, which is an aspect that remains relatively underdeveloped in comparison to intra-database evaluations." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.378, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Scaling Up Personalized Image Aesthetic Assessment" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.784, + 0.173 + ], + "angle": 0, + "content": "Table 3: Cross-database evaluation on the AADB [19] database. Results marked with * indicate intra-database performance (i.e., directly trained on the AADB database)." + }, + { + "type": "table", + "bbox": [ + 0.273, + 0.175, + 0.728, + 0.349 + ], + "angle": 0, + "content": "
MethodSROCC
10-shot100-shot
PA-IAA [20]0.469±0.0020.524±0.006
BLG-PIAA [44]0.486±0.0040.536±0.006
PIAA-SOA [45]0.509±0.0030.557±0.007
TAPP-PIAA [21]-0.540
MTCL [41]0.533±0.0040.572±0.007
*TAPP-PIAA [21]*0.534±0.004*0.612±0.007
*MTCL [41]*0.540±0.005*0.622±0.007
Ours0.556±0.0040.654±0.007
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.365, + 0.787, + 0.501 + ], + "angle": 0, + "content": "For cross-database evaluation, we employ the AADB [19] and REAL-CUR [31] databases as the test databases, neither of which were among the six databases from which we derived our task vectors. Following established protocols [31], we randomly select \\( K \\) samples for each individual to form the training set, also referred to as the support set, while the remaining images form the test set. In our experiments, we set \\( K \\) as either 10-shot or 100-shot. Given the potential variability inherent in few-shot learning depending on the chosen \\( K \\) support set, we conduct 10 independent trials for each user and report the average SROCC with the standard deviation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.501, + 0.788, + 0.637 + ], + "angle": 0, + "content": "As illustrated in Table 2 and Table 3, our method significantly surpasses existing PIAA models in cross-database performance for both the 10-shot and 100-shot protocols. Remarkably, it even exceeds the performance of models trained explicitly on the AADB database, as compared in Table 3. Such results underscore the exceptional generalization capabilities of our approach, in contrast to previous methods that are constrained by their inability to leverage GIAA and IQA databases. This superior performance is primarily due to the scalability of our approach, which effectively harnesses an extensive amount of 415,043 samples across multiple databases, each contributing unique insights and tendencies." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.651, + 0.476, + 0.665 + ], + "angle": 0, + "content": "4.5 Intra-database Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.787, + 0.749 + ], + "angle": 0, + "content": "For intra-database evaluation, we test our approach for the 37 individuals in the Flickr-AES [31] database, which is the most widely used PIAA database. Since the Flickr-AES [31] database is one of the six databases used to derive the task vectors, this experiment falls under the category of intra-database evaluation. However, we do not use the annotator information in the Flickr-AES." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.787, + 0.841 + ], + "angle": 0, + "content": "In Table 4, we report the average SROCC for each approach along with the standard deviation across 10 independent trials. These experiments underscore the effectiveness of training personalized task vectors through the use of multiple fine-tuned models. Our approach consistently outperforms existing methods that rely on meta-learning techniques or incorporate additional user attributes (e.g., personality traits) to enhance personalized score predictions." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.416, + 0.128 + ], + "angle": 0, + "content": "JY. Yun and J. Choo" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.784, + 0.173 + ], + "angle": 0, + "content": "Table 4: 10-shot and 100-shot personalization results on the Flickr-AES database [31], with the average SROCC of 37 users and standard deviation across 10 trials." + }, + { + "type": "table", + "bbox": [ + 0.261, + 0.175, + 0.739, + 0.44 + ], + "angle": 0, + "content": "
MethodSROCC
10-shot100-shot
PAM (attribute only) [31]0.511±0.0040.516±0.003
PAM (content only) [31]0.512±0.0020.516±0.010
PAM [31]0.513±0.0030.524±0.007
PASS [35]0.516±0.0030.521±0.007
USAR-PPR [25]0.521±0.0020.544±0.007
USAR-PAD [25]0.520±0.0030.537±0.003
USAR-PPR&PAD [25]0.525±0.0040.552±0.015
PA-IAA [20]0.543±0.0030.639±0.011
BLG-PIAA [44]0.561±0.0050.669±0.013
UG-PIAA [24]0.559±0.0020.660±0.013
PIAA-SOA [45]0.618±0.0060.691±0.015
TAPP-PIAA [21]0.591±0.0070.685±0.012
IM-PIAA [13]0.6200.708
MTCL [21]0.667±0.0050.737±0.014
Ours0.668±0.0040.748±0.012
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.457, + 0.544, + 0.473 + ], + "angle": 0, + "content": "5 Analysis and Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.484, + 0.784, + 0.514 + ], + "angle": 0, + "content": "We perform an in-depth analysis and ablation studies on each component of our approach, including model size, scalability, and coefficient initialization." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.533, + 0.548, + 0.547 + ], + "angle": 0, + "content": "5.1 Task Similarities Across Databases" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.561, + 0.787, + 0.697 + ], + "angle": 0, + "content": "To gain insights into the derived task vectors, we visualize the similarity matrix of the task vectors in Figure 4. For simplicity, we aggregate the \\(L\\) task vectors derived from each layer of a single fine-tuned model as a single task vector by concatenating them into one vector. In line with the findings provided by Ilharco et al. [14], task vectors are typically orthogonal, even though the databases are collected with similar intentions (i.e., GIAA and IQA). This observation highlights that treating each database as a unique task remains a valid choice, owing to variations in their data collection methodologies and the diverse nature of the populations from which they are sourced." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.722, + 0.562, + 0.736 + ], + "angle": 0, + "content": "5.2 Architecture Choices and Variations" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.75, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Selecting the architecture. Weight mixing techniques, including task vectors [14, 36, 37], have been recognized for their effectiveness when fine-tuning models pre-trained on large databases, such as the LAION-5B [33] or DataComp [7]. Vision Transformers (ViT) [5] is a suitable architecture for accommodating web-scale databases, in contrast to ResNets [9], which often struggle to achieve high performance despite training on large databases. Therefore, given" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.379, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Scaling Up Personalized Image Aesthetic Assessment" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.294, + 0.151, + 0.709, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.374, + 0.785, + 0.401 + ], + "angle": 0, + "content": "Fig. 4: Similarity matrix between task vectors derived from each database. The similarities are measured with the cosine similarity." + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.406, + 0.785, + 0.434 + ], + "angle": 0, + "content": "Table 5: Performance and inference speed of the ViT-B/16 model and the Best-fit FT method (i.e., directly fine-tuning ViTs) on the REAL-CUR [31] database." + }, + { + "type": "table", + "bbox": [ + 0.25, + 0.437, + 0.75, + 0.564 + ], + "angle": 0, + "content": "
MethodSROCCFPS
10-shot100-shot
Best-fit FT0.303±0.0080.394±0.00894.41
PIAA-SOA [45]0.487±0.0030.589±0.015-
MTCL [41]0.495±0.0070.599±0.012359.05
Ours (ViT-B/16)0.562±0.0040.607±0.007417.95
Ours0.577±0.0050.621±0.00794.41
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.577, + 0.785, + 0.637 + ], + "angle": 0, + "content": "its scalability and performance advantages, we choose the publicly available pretrained ViT as our architecture for our experiments. Regardless of the chosen architecture, our parameter-efficient approach only trains a minimal set of coefficients corresponding to each layer-wise task vector." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.641, + 0.788, + 0.806 + ], + "angle": 0, + "content": "Scaling down for efficient computation. While our primary experiments focus on combining task vectors derived from ViT-L/14 models, our approach is adaptable to smaller models as well. As shown in Table 5, our approach yields strong results even on smaller models like ViT-B/16, which are well-suited for deployment on commercial computers. Although fast inference was not our primary goal, we achieve real-time inference speeds (\\(\\geq 30\\) FPS) on the RTX 3090, as our weight mixing technique combines multiple models into a single ViT model, eliminating the need to pass through multiple models. Acceleration of ViTs [22, 23, 43] beyond our implementation is also actively supported in recent frameworks [26] and communities [30], which ensures the practicality of our approach across a wide range of computing platforms such as mobile devices." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Vision Transformers without task vectors. Our experimental results in Table 5 demonstrate that the performance gains are not solely attributed to" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.416, + 0.127 + ], + "angle": 0, + "content": "JY. Yun and J. Choo" + }, + { + "type": "image", + "bbox": [ + 0.321, + 0.149, + 0.679, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.34, + 0.785, + 0.368 + ], + "angle": 0, + "content": "Fig. 5: Performance gains with respect to the number of databases used. The dotted lines indicate the previous state-of-the-art performance." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.381, + 0.788, + 0.472 + ], + "angle": 0, + "content": "the inherent strengths of the ViT architecture. When directly fine-tuning the aesthetic assessment model that best matches the user's preference (i.e., Best-fit FT), we observe distinctly low personalization performances despite using a large pre-trained model. This highlights the pivotal role of keeping the layer-wise task vectors frozen during personalization, as altering the vectors would indicate discarding the rich preferences learned from a large database." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.488, + 0.494, + 0.504 + ], + "angle": 0, + "content": "5.3 Scalability of Our Approach" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.518, + 0.788, + 0.669 + ], + "angle": 0, + "content": "How does each database contribute? We assess the scalability of our approach by varying the number of task vectors. Specifically, we evaluate our approach on the REAL-CUR database [31] by increasing the number of databases used. Starting from the GIAA model trained on the AVA [27] database, we accumulate the databases in the following order: Flicker-AES [31], TAD66K [11], PARA [40], KonIQ-10K [12], and SPAQ [6]. As demonstrated in Figure 5, both 10-shot and 100-shot performance significantly improves as the number of task vectors increases. This indicates that our scalable approach benefits from the expansion of task vectors, which allows us to capture a broader range of image characteristics and nuances." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.673, + 0.788, + 0.779 + ], + "angle": 0, + "content": "How can we increase the number of task vectors? The advantage of our approach can be amplified with the availability of additional image score regression databases. For example, scaling up the number of task vectors can be further pursued by incorporating readily available IQA [4,8,42] and GIAA [10,16] databases not included in our current study. Additionally, partitioning large databases, such as AVA [27] and PaQ-2-PiQ [42], can be an alternative method for generating additional task vectors." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "These strategies for obtaining a diverse set of task vectors may lead to performance gains in the 10-shot evaluation, which is considered highly challenging due to the limited number of user-provided samples. However, we leave this as future work, given that our current work already utilizes the largest number of" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.378, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Scaling Up Personalized Image Aesthetic Assessment" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.349, + 0.149, + 0.661, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.361, + 0.784, + 0.389 + ], + "angle": 0, + "content": "Fig. 6: 2D-histogram of coefficients and zero-shot SROCC pairs. The size of the coefficients are correlated with the relevance of the task." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.398, + 0.784, + 0.429 + ], + "angle": 0, + "content": "samples for PIAA, 10 times larger than the most widely used and highly curated Flickr-AES database [31]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.445, + 0.629, + 0.459 + ], + "angle": 0, + "content": "5.4 Task Relevance and Coefficient Initialization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.472, + 0.788, + 0.652 + ], + "angle": 0, + "content": "We conduct an analysis to better understand the distinguishing characteristics of coefficients, specifically, what sets apart those with high values from those with low values. For a straightforward examination, we calculate the average coefficient for each task (i.e., database), averaged across all the layers of the model, which results in six average coefficients for each user (\\(n = 6\\)). Subsequently, we assess the six zero-shot SROCC scores associated with the average coefficients and plot the histogram of the coefficient-SROCC pairs from all test users in Figure 6. We observe that coefficients for tasks with a high zero-shot SROCC are trained to have high values, while those for less-relevant tasks (i.e., low zero-shot SROCC) exhibit low values. This analysis justifies our adaptive initialization of the coefficients, which assigns large values to coefficients associated with closely related tasks." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.653, + 0.788, + 0.699 + ], + "angle": 0, + "content": "We investigate two alternative initialization strategies: a uniform initialization and a best-fit initialization. Both cases can be viewed as an extreme case of the adaptive initialization, which can be written in a generalized form:" + }, + { + "type": "equation", + "bbox": [ + 0.326, + 0.709, + 0.786, + 0.756 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {i} ^ {l} = \\frac {e ^ {\\mathrm {S R O C C} (S , \\theta_ {\\mathrm {f t} _ {i}}) / T}}{\\sum_ {j = 1} ^ {n} e ^ {\\mathrm {S R O C C} (S , \\theta_ {\\mathrm {f t} _ {j}}) / T}} \\quad \\text {f o r} l = 1, 2, \\dots , L, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "where \\( T \\) is the temperature for the softmax. The uniform initialization corresponds to the case where the \\( T \\to \\infty \\), while the best-fit initialization is when \\( T \\to 0 \\). As demonstrated in Table 6, the adaptive initialization outperforms other initialization strategies. Nevertheless, we consider further exploration of the initialization strategies as a potential avenue for future research." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.416, + 0.127 + ], + "angle": 0, + "content": "JY. Yun and J. Choo" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.784, + 0.173 + ], + "angle": 0, + "content": "Table 6: Comparison between three different coefficient initialization strategies tested on the REAL-CUR [31] database." + }, + { + "type": "table", + "bbox": [ + 0.272, + 0.175, + 0.727, + 0.268 + ], + "angle": 0, + "content": "
MethodSROCC
10-shot100-shot
Uniform init. (T→∞)0.520±0.0020.602±0.002
Best-fit init. (T→0)0.464±0.0120.596±0.002
Adaptive init. (T=1)0.577±0.0050.621±0.007
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.277, + 0.356, + 0.292 + ], + "angle": 0, + "content": "6 Limitation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.298, + 0.784, + 0.388 + ], + "angle": 0, + "content": "We have observed that when the zero-shot SROCC for all six image regression models are low for an individual (i.e., \\( \\mathrm{SROCC} \\approx 0 \\)), our approach may struggle to identify an optimal combination of task vectors for that user. This phenomenon suggests that when there is a limited correlation between the user's preferences and the predictions of these models, the task vectors derived from them may be insufficient for finding an appropriate personalized task vector." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.39, + 0.785, + 0.48 + ], + "angle": 0, + "content": "However, we find this case to be extremely rare (See appendix Figure 2-7.), with only 1 among 73 cases having a low SROCC under 0.2, demonstrating the broad coverage of our approach. Also, increasing the number of task vectors will effectively mitigate this issue, leveraging the inherent scalability of our approach, which can accommodate a growing number of task vectors to improve personalization even in challenging scenarios." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.493, + 0.36, + 0.509 + ], + "angle": 0, + "content": "7 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.514, + 0.784, + 0.589 + ], + "angle": 0, + "content": "We address the scalability issue in PIAA, a critical yet previously unaddressed aspect necessary for enhancing the generalization performance. Unlike previous studies limited to training on PIAA databases curated by individual annotators, our approach enables the flexible use of multiple image score regression databases, thereby providing greater scalability and generalization capabilities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.59, + 0.785, + 0.726 + ], + "angle": 0, + "content": "We demonstrate that GIAA and IQA databases exhibit distinct personalization potentials, which have motivated us to combine GIAA and IQA models to achieve desired behaviors. By introducing learnable parameters for optimal model combinations, we enable the precise personalization of models to individual users' aesthetic preferences. Our approach significantly outperforms existing approaches, demonstrating exceptional efficacy in generalizing to unseen domains which is a critical requirement for real-world applications. This work opens up new avenues for personalized image aesthetic assessment, offering valuable insights and practical solutions for this challenging task." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.734, + 0.357, + 0.748 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.752, + 0.787, + 0.84 + ], + "angle": 0, + "content": "This work was supported by the Institute for Information & communications Technology Promotion(IITP) grant funded by the Korea government(MSIT) (No.RS-2019-II190075 Artificial Intelligence Graduate School Program(KAIST)), the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. NRF-2022R1A2B5B02001913), and the Institute of Information & communications Technology Promotion (IITP) grant funded by the Korea government (MSIT) (RS-2021-II212068, Artificial Intelligence Innovation Hub)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.378, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Scaling Up Personalized Image Aesthetic Assessment" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.324, + 0.161 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.179, + 0.785, + 0.207 + ], + "angle": 0, + "content": "1. Al-Shedivat, M., Li, L., Xing, E., Talwalkar, A.: On data efficiency of meta-learning. In: International Conference on Artificial Intelligence and Statistics. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.209, + 0.785, + 0.235 + ], + "angle": 0, + "content": "2. Bradley, R.A., Terry, M.E.: Rank analysis of incomplete block designs: I. the method of paired comparisons. Biometrika (1952)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.236, + 0.785, + 0.263 + ], + "angle": 0, + "content": "3. Chefer, H., Gur, S., Wolf, L.: Transformer interpretability beyond attention visualization. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.264, + 0.785, + 0.305 + ], + "angle": 0, + "content": "4. Ciancio, A., da Silva, E.A., Said, A., Samadani, R., Obrador, P., et al.: No-reference blur assessment of digital pictures based on multifeature classifiers. IEEE Transactions on Image Processing (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.307, + 0.785, + 0.348 + ], + "angle": 0, + "content": "5. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.349, + 0.787, + 0.376 + ], + "angle": 0, + "content": "6. Fang, Y., Zhu, H., Zeng, Y., Ma, K., Wang, Z.: Perceptual quality assessment of smartphone photography. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.377, + 0.785, + 0.418 + ], + "angle": 0, + "content": "7. Gadre, S.Y., Ilharco, G., Fang, A., Hayase, J., Smyrnis, G., Nguyen, T., Marten, R., Wortsman, M., Ghosh, D., Zhang, J., et al.: Datacomp: In search of the next generation of multimodal datasets. arXiv preprint arXiv:2304.14108 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.419, + 0.785, + 0.447 + ], + "angle": 0, + "content": "8. Ghadiyaram, D., Bovik, A.C.: Massive online crowdsourced study of subjective and objective picture quality. IEEE Transactions on Image Processing (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.448, + 0.785, + 0.473 + ], + "angle": 0, + "content": "9. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.476, + 0.785, + 0.502 + ], + "angle": 0, + "content": "10. He, S., Ming, A., Li, Y., Sun, J., Zheng, S., Ma, H.: Thinking image color aesthetics assessment: Models, datasets and benchmarks. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.504, + 0.785, + 0.53 + ], + "angle": 0, + "content": "11. He, S., Zhang, Y., Xie, R., Jiang, D., Ming, A.: Rethinking image aesthetics assessment: Models, datasets and benchmarks. IJCAI (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.532, + 0.785, + 0.573 + ], + "angle": 0, + "content": "12. Hosu, V., Lin, H., Sziranyi, T., Saupe, D.: Koniq-10k: An ecologically valid database for deep learning of blind image quality assessment. IEEE Transactions on Image Processing (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.574, + 0.785, + 0.601 + ], + "angle": 0, + "content": "13. Hou, J., Lin, W., Yue, G., Liu, W., Zhao, B.: Interaction-matrix based personalized image aesthetics assessment. IEEE Transactions on Multimedia (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.602, + 0.785, + 0.629 + ], + "angle": 0, + "content": "14. Ilharco, G., Ribeiro, M.T., Wortsman, M., Schmidt, L., Hajishirzi, H., Farhadi, A.: Editing models with task arithmetic. In: ICLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.63, + 0.785, + 0.685 + ], + "angle": 0, + "content": "15. Ilharco, G., Wortsman, M., Wightman, R., Gordon, C., Carlini, N., Taori, R., Dave, A., Shankar, V., Namkoong, H., Miller, J., Hajishirzi, H., Farhadi, A., Schmidt, L.: Openclip (Jul 2021). https://doi.org/10.5281/zenodo.5143773, https://doi.org/10.5281/zenodo.5143773" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.686, + 0.785, + 0.713 + ], + "angle": 0, + "content": "16. Jin, X., Wu, L., Zhao, G., Li, X., Zhang, X., Ge, S., Zou, D., Zhou, B., Zhou, X.: Aesthetic attributes assessment of images. In: ACM MM (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.714, + 0.785, + 0.741 + ], + "angle": 0, + "content": "17. Karlsson, K., Jiang, W., Zhang, D.Q.: Mobile photo album management with multiscale timeline. In: ACM MM (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.742, + 0.785, + 0.77 + ], + "angle": 0, + "content": "18. Ke, J., Wang, Q., Wang, Y., Milanfar, P., Yang, F.: Musiq: Multi-scale image quality transformer. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.771, + 0.785, + 0.798 + ], + "angle": 0, + "content": "19. Kong, S., Shen, X., Lin, Z., Mech, R., Fowlkes, C.: Photo aesthetics ranking network with attributes and content adaptation. In: ECCV. Springer (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "20. Li, L., Zhu, H., Zhao, S., Ding, G., Lin, W.: Personality-assisted multi-task learning for generic and personalized image aesthetics assessment. IEEE Transactions on Image Processing 29 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.179, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.416, + 0.127 + ], + "angle": 0, + "content": "JY. Yun and J. Choo" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.786, + 0.189 + ], + "angle": 0, + "content": "21. Li, Y., Yang, Y., Li, H., Chen, H., Xu, L., Li, L., Li, Y., Guo, Y.: Transductive aesthetic preference propagation for personalized image aesthetics assessment. In: ACM MM. ACM (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.189, + 0.786, + 0.217 + ], + "angle": 0, + "content": "22. Lin, Y., Zhang, T., Sun, P., Li, Z., Zhou, S.: Fq-vit: Post-training quantization for fully quantized vision transformer. IJCAI (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.217, + 0.786, + 0.243 + ], + "angle": 0, + "content": "23. Liu, Z., Wang, Y., Han, K., Zhang, W., Ma, S., Gao, W.: Post-training quantization for vision transformer. NeurIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.243, + 0.786, + 0.284 + ], + "angle": 0, + "content": "24. Lv, P., Fan, J., Nie, X., Dong, W., Jiang, X., Zhou, B., Xu, M., Xu, C.: User-guided personalized image aesthetic assessment based on deep reinforcement learning. IEEE Transactions on Multimedia (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.284, + 0.786, + 0.325 + ], + "angle": 0, + "content": "25. Lv, P., Wang, M., Xu, Y., Peng, Z., Sun, J., Su, S., Zhou, B., Xu, M.: Usar: An interactive user-specific aesthetic ranking framework for images. In: ACM MM. pp. 1328-1336 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.325, + 0.638, + 0.338 + ], + "angle": 0, + "content": "26. Microsoft: Deepspeed (2023), https://www.deepspeed.ai/" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.338, + 0.786, + 0.365 + ], + "angle": 0, + "content": "27. Murray, N., Marchesotti, L., Perronnin, F.: Ava: A large-scale database for aesthetic visual analysis. In: CVPR. IEEE (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.365, + 0.786, + 0.392 + ], + "angle": 0, + "content": "28. Myers, J.L., Well, A.D., Lorch Jr, R.F.: Research design and statistical analysis. Routledge (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.392, + 0.786, + 0.419 + ], + "angle": 0, + "content": "29. Naseer, M.M., Ranasinghe, K., Khan, S.H., Hayat, M., Shahbaz Khan, F., Yang, M.H.: Intriguing properties of vision transformers. NeurIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.42, + 0.786, + 0.461 + ], + "angle": 0, + "content": "30. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. NeurIPS (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.461, + 0.786, + 0.487 + ], + "angle": 0, + "content": "31. Ren, J., Shen, X., Lin, Z., Mech, R., Foran, D.J.: Personalized image aesthetics. In: ICCV (Oct 2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.487, + 0.786, + 0.514 + ], + "angle": 0, + "content": "32. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. 2022 IEEE. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.514, + 0.786, + 0.555 + ], + "angle": 0, + "content": "33. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.555, + 0.786, + 0.582 + ], + "angle": 0, + "content": "34. Wallace, B., Gokul, A., Ermon, S., Naik, N.: End-to-end diffusion latent optimization improves classifier guidance. ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.582, + 0.786, + 0.609 + ], + "angle": 0, + "content": "35. Wang, G., Yan, J., Qin, Z.: Collaborative and attentive learning for personalized image aesthetic assessment. In: IJCAI. pp. 957-963 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.609, + 0.786, + 0.663 + ], + "angle": 0, + "content": "36. Wortsman, M., Ilharco, G., Gadre, S.Y., Roelofs, R., Gontijo-Lopes, R., Morcos, A.S., Namkoong, H., Farhadi, A., Carmon, Y., Kornblith, S., et al.: Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In: ICML. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.663, + 0.786, + 0.704 + ], + "angle": 0, + "content": "37. Wortsman, M., Ilharco, G., Li, M., Kim, J.W., Hajishirzi, H., Farhadi, A., Namkoong, H., Schmidt, L.: Robust fine-tuning of zero-shot models. 2022 IEEE. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.704, + 0.786, + 0.731 + ], + "angle": 0, + "content": "38. Xu, L., Xu, J., Yang, Y., Huang, Y., Xie, Y., Li, Y.: Clip brings better features to visual aesthetics learners. arXiv preprint arXiv:2307.15640 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.731, + 0.786, + 0.772 + ], + "angle": 0, + "content": "39. Yan, X., Shao, F., Chen, H., Jiang, Q.: Hybrid cnn-transformer based meta-learning approach for personalized image aesthetics assessment. Journal of Visual Communication and Image Representation 98 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.786, + 0.799 + ], + "angle": 0, + "content": "40. Yang, Y., Xu, L., Li, L., Qie, N., Li, Y., Zhang, P., Guo, Y.: Personalized image aesthetics assessment with rich attributes. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.786, + 0.841 + ], + "angle": 0, + "content": "41. Yang, Z., Li, L., Yang, Y., Li, Y., Lin, W.: Multi-level transitional contrast learning for personalized image aesthetics assessment. IEEE Transactions on Multimedia (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.786, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.378, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Scaling Up Personalized Image Aesthetic Assessment" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "42. Ying, Z., Niu, H., Gupta, P., Mahajan, D., Ghadiyaram, D., Bovik, A.: From patches to pictures (paq-2-piq): Mapping the perceptual space of picture quality. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.788, + 0.218 + ], + "angle": 0, + "content": "43. Yuan, Z., Xue, C., Chen, Y., Wu, Q., Sun, G.: Ptq4vit: Post-training quantization for vision transformers with twin uniform quantization. In: ECCV. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.788, + 0.259 + ], + "angle": 0, + "content": "44. Zhu, H., Li, L., Wu, J., Zhao, S., Ding, G., Shi, G.: Personalized image aesthetics assessment via meta-learning with bilevel gradient optimization. IEEE Transactions on Cybernetics (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.788, + 0.287 + ], + "angle": 0, + "content": "45. Zhu, H., Zhou, Y., Li, L., Li, Y., Guo, Y.: Learning personalized image aesthetics from subjective and objective attributes. IEEE Transactions on Multimedia (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.287 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Scaling Up Personalized Image Aesthetic Assessment via Task Vector Customization/77e3f8ef-24e2-4d25-ba6f-999ef6bc7191_origin.pdf b/2024/Scaling Up Personalized Image Aesthetic Assessment via Task Vector Customization/77e3f8ef-24e2-4d25-ba6f-999ef6bc7191_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9078b71faf71a9d14a642025ff2ab156776ee90a --- /dev/null +++ b/2024/Scaling Up Personalized Image Aesthetic Assessment via Task Vector Customization/77e3f8ef-24e2-4d25-ba6f-999ef6bc7191_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a2ca3a548221c5b85abc25c626557ead3a22aa80f916d43234a016d8f0a8d77d +size 671195 diff --git a/2024/Scaling Up Personalized Image Aesthetic Assessment via Task Vector Customization/full.md b/2024/Scaling Up Personalized Image Aesthetic Assessment via Task Vector Customization/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4dde94c354a358430172c7a251a098c69481338e --- /dev/null +++ b/2024/Scaling Up Personalized Image Aesthetic Assessment via Task Vector Customization/full.md @@ -0,0 +1,301 @@ +# Scaling Up Personalized Image Aesthetic Assessment via Task Vector Customization + +Jooyeol Yun and Jaegul Choo + +Korea Advanced Institute of Science and Technology (KAIST) blizzard072@kaist.ac.kr jchoo@kaist.ac.kr + +Abstract. The task of personalized image aesthetic assessment seeks to tailor aesthetic score prediction models to match individual preferences with just a few user-provided inputs. However, the scalability and generalization capabilities of current approaches are considerably restricted by their reliance on an expensive curated database. To overcome this long-standing scalability challenge, we present a unique approach that leverages readily available databases for general image aesthetic assessment and image quality assessment. Specifically, we view each database as a distinct image score regression task that exhibits varying degrees of personalization potential. By determining optimal combinations of task vectors, known to represent specific traits of each database, we successfully create personalized models for individuals. This approach of integrating multiple models allows us to harness a substantial amount of data. Our extensive experiments demonstrate the effectiveness of our approach in generalizing to previously unseen domains—a challenge previous approaches have struggled to achieve—making it highly applicable to real-world scenarios. Our novel approach significantly advances the field by offering scalable solutions for personalized aesthetic assessment and establishing high standards for future research. + +# 1 Introduction + +Personalized image aesthetic assessment (PIAA) is an emerging field dedicated to developing aesthetic score prediction models that closely match an individual's aesthetic preferences based on a small set of user-provided samples. While they function as simple models that assign scores based on aesthetic quality, PIAA models enhance user experience in digital environments by personalizing tasks such as managing photo albums [17], curating web-scale databases like LAION-Aesthetics [33], and even guiding generative models such as Stable Diffusion [32] to create images that align with individual tastes [34]. + +In recent years, there has been a notable trend in PIAA [21, 41, 44] that involves the integration of meta-learning techniques, as personalization is inherently related to few-shot learning. Specifically, these approaches entail pretraining a model on a database consisting of image collections from diverse individuals. However, these strategies encounter significant challenges in terms of scalability and generalization to unseen domains, primarily due to the prohibitive cost associated with data collection. + +![](images/2f8f78a387819845d903c928eb9b0ae98f9784f842142670e7d9fc764cbb51e7.jpg) +Fig. 1: Comparison between PIAA and other image regression databases. Our approach has the flexibility to leverage any image regression databases. + +While meta-learning typically demands an extensive number of meta-training tasks, often exceeding 10,000, to achieve high performance [1], currently available PIAA databases [19, 31] offer a limited number of tasks fewer than 200. Even with the recently collected PIAA database [40], which contains rich attributes from each annotator, the number of training tasks remains below 500. In the context of PIAA, a single meta-training task corresponds to learning the preferences of a single individual. Scaling such a database would require obtaining a vast number of images and corresponding personalized aesthetic scores from numerous individuals, which poses severe limitations in the scalability and generalization of existing approaches. + +In this paper, we unveil a data-rich approach for personalizing aesthetic assessment models by leveraging existing general image aesthetic assessment (GIAA) and image quality assessment (IQA) databases. In contrast to prior work [20, 21, 41, 44, 45], our approach is not limited to training on databases that track individual annotators, as depicted in Figure 1. Instead, we have the flexibility to utilize multiple image score regression databases, which are readily available. By employing broader resources, we overcome the scalability and generalization challenges that have long been unaddressed in this field. + +![](images/88277c622dc9f4f83ad4434af9fd83046166b75590aa18af1090d1ac57018053.jpg) +Fig.2: Zero-shot personalization performance for four users in the Flickr-AES database [31]. Each colored bar refers to the personalization performance of a model trained on a specific database. + +The key idea is to view each GIAA and IQA data as a unique image score regression task. This perspective arises from the distinct characteristics and tendencies of each database due to different data collection procedures and the population from which they originate. As visualized in Figure 2, image score regression models trained on four different GIAA [27,31] and IQA [6,12] databases + +exhibit varying levels of personalization for users within the PIAA database. For instance, User 3's aesthetic preferences closely align with the score regression model trained on the Flickr-AES [31] while diverging from the model trained on the SPAQ [6] database. In other words, even without any personalized training, we find that general image assessment models inherently possess different degrees of personalization capabilities. + +In light of these insights, we extend an emerging approach known as task arithmetic [14], which was originally devised for developing multi-task models. This method involves directly adding or subtracting the weights of models fine-tuned on specific tasks. In our context, task vectors from diverse image assessment models encapsulate distinct features and behaviors related to specific tasks, such as recognizing different aspects of image quality or aesthetics. By carefully combining task vectors, we explore the potential to selectively amplify or refine the capabilities of a pre-trained model. + +However, there has been limited research on determining the optimal adjustments necessary to achieve a certain behavior. Thus, we present a personalization approach by introducing trainable coefficients for each task vector derived from models trained on distinct GIAA and IQA databases, thereby determining the weights for combining these vectors. Once the coefficients are trained with user-provided inputs, the weighted sum of task vectors creates a model tailored to the user's aesthetic preferences. Since task vectors obtained from large databases already capture preferences across various themes and contents, we find that training only the coefficients is sufficient for precise personalization, making our approach highly parameter-efficient. To the best of our knowledge, our approach marks the first to customize task vectors for specific model behaviors that were previously unattainable. + +By carefully customizing task vectors, we incorporate large databases into the training process, granting our approach a strong generalization capability for unseen domains, a scenario often encountered in real-world applications. Furthermore, we find that our approach is also practical for learning a user's aesthetic preference with only a few samples, as our approach of merging task vectors guides the model to make well-informed updates and prevents overfitting. We provide extensive experiments demonstrating the efficacy of our approach for personalization in diverse scenarios. + +Our contributions are threefold: + +- We introduce a novel data-rich approach for PIAA that tackles the long-standing scalability challenge, eliminating the dependency on expensive, manually-curated databases that have hindered progress in this field. +- Demonstrating exceptional effectiveness in cross-database evaluations, our method showcases robust generalization capabilities that surpass those of existing approaches, setting high standards for future work. +- By learning optimal combinations of task vectors, we present a parameter-efficient approach for precise personalization, leveraging the comprehensive information embedded in these vectors. + +# 2 Related Work + +# 2.1 Personalized Image Aesthetic Assessment + +Personalized image aesthetic assessment (PIAA) [13,20,21,24,25,35,39-41,44,45] aims to learn the personal aesthetic preference of individual users based on the image-score pairs from one's image collection. Since collecting a large number of samples from a single user is impractical, PIAA is closely linked to few-shot learning methods. One line of work [20,40,45] incorporates additional attributes (e.g., personality traits, age, and gender) to facilitate user-specific aesthetic predictions. These approaches rely on the correlation between user attributes and aesthetic preferences, which may not always be clearly defined. Consequently, many high-performing approaches for PIAA [21, 41, 44] leverage meta-learning techniques to train models that can be easily fine-tuned to accommodate the aesthetic preferences of new users with only a limited number of samples. + +Nevertheless, applying meta-learning techniques to PIAA encounters scalability challenges owing to the limited number of tasks available during the meta-learning phase. In traditional meta-learning, achieving a high performance typically requires a substantial number of meta-training tasks [1], often exceeding 10,000. In PIAA databases [19,31], the number of training tasks ranges from 100 to fewer than 500, which can impose limitations on the performance. Expanding the task pool in PIAA is prohibitively expensive since each task corresponds to the image collection of a single user, posing significant scalability challenges that has long been unaddressed. We tackle this issue in previous PIAA approaches and present a scalable fine-tuning method that leverages existing general image aesthetic assessment (GIAA) and image quality assessment (IQA) databases. + +# 2.2 Task Vectors and Task Arithmetic + +Ilharco et al. [14] presented a new paradigm for creating multi-task models, utilizing the parameters of multiple fine-tuned models. They define the task vector $\tau$ as the difference between the fine-tuned and pre-trained weights, + +$$ +\tau = \theta_ {\mathrm {f t}} - \theta_ {\mathrm {p r e}}, +$$ + +where $\theta_{\mathrm{ft}}$ and $\theta_{\mathrm{pre}}$ represent the weights of the fine-tuned and pre-trained model, respectively. They demonstrate that adding and subtracting task vectors to pretrained models can encourage or suppress the ability to perform specific tasks. For example, negating the task vector derived from a language model fine-tuned on toxic data reduces the frequency of generating toxic text. However, it remains unclear what the optimal adjustments are for achieving a specific behavior. + +In our paper, we investigate methods to determine necessary adjustments to achieve a certain behavior defined by a few user-provided samples. By introducing learnable coefficients to task vectors, we effectively customize models to closely match individual preferences. We also demonstrate that finding the optimal coefficients is possible with only a few samples, as the task vectors encompass comprehensive knowledge derived from large databases. + +![](images/0020b54c382bcb578b60cfb77bf859bbbc2620db6b71d62d99f591b3dcca21ba.jpg) +(a) Obtain task vectors from various databases + +![](images/4f6382017d196c68d3c1153419e1c9131ca888072bd07f54c09a7dda609688b1.jpg) +(b) Learn optimal combinations for personalization +Fig. 3: (a) Multiple models are fine-tuned from a single pre-trained model on different tasks to obtain task vectors. For simplicity, we omit the discussion of layers. (b) Given a small number of user-provided samples, the coefficients corresponding to each task vector are optimized with the rank loss. + +# 3 Method + +# 3.1 Obtaining Layer-wise Task Vectors from Various Databases + +To obtain task vectors, we train multiple models with the same architecture on various tasks (i.e., different databases), as illustrated in Figure 3 (a). Specifically, we fine-tune the weights from a pre-trained backbone, adjusting the parameters from $\theta_{\mathrm{pre}}$ to $\theta_{\mathrm{ft}_i}$ for each task $i$ . Moreover, while the specific roles of model layers in capturing unique features are inherently complex, their distinct contributions are undeniable [3,29]. Thus, we compute task vectors across all layers, enhancing the flexibility of our approach for personalization (See appendix Section A.2). + +Accordingly, we expand the concept of original task vectors to include layerwise differentiation by, + +$$ +\tau_ {i} ^ {l} = \theta_ {\mathrm {f t} _ {i}} ^ {l} - \theta_ {\mathrm {p r e}} ^ {l}, \tag {1} +$$ + +where $\tau_{i}^{l}$ is the task vector for the $l$ -th layer of the $i$ -th task. Each vector comprehensively integrates patterns learned from large databases, acting as basis vectors for robust and precise personalization. + +It is important to note that this process is conducted only once and is not required again for personalizing to individual users, thereby not contributing to the count of learnable parameters for personalization. + +# 3.2 Learning Optimal Combinations for Personalization + +Given the layer-wise task vectors, our goal is to find a combination of these vectors that accurately reflects a user's aesthetic preferences. To achieve this, we introduce learnable coefficients $\alpha_{i}^{l}$ for each task vector $\tau_{i}^{l}$ . Once the coefficients + +are trained, the weights of the personalized model for each layer are derived as, + +$$ +\theta_ {\mathrm {p}} ^ {l} = \theta_ {\mathrm {p r e}} ^ {l} + \sum_ {i = 1} ^ {n} \alpha_ {i} ^ {l} \tau_ {i} ^ {l} \quad \text {f o r} l = 1, 2, \dots , L, \tag {2} +$$ + +where $n$ is the number of tasks, and $L$ is the number of layers in the model. We keep all parameters other than the $n \times L$ coefficients frozen. This is essential since the task vectors contain comprehensive insights into aesthetic preferences, enabling the model to make well-informed updates. By harnessing the rich information embedded in these vectors, our method efficiently learns complex aesthetic preferences from merely a few user-provided samples. + +Our objective function is designed to penalize the misordering of aesthetic scores. Specifically, we adopt the Bradley-Terry model [2], which estimates the probability of pairwise comparison between two samples. We frame the prediction of the two scores as a binary classification problem, where the objective is to predict which sample has the higher score. The coefficients of each task vector are trained to minimize the following loss function: + +$$ +L _ {\mathrm {R}} = - \log \frac {\exp \hat {y} _ {1}}{\exp \hat {y} _ {1} + \exp \hat {y} _ {2}} (y _ {1} \geq y _ {2}), \tag {3} +$$ + +where $\hat{y}_1$ and $\hat{y}_2$ represent the aesthetic score predictions for images $x_1$ and $x_2$ , whose ground-truth scores are $y_1$ and $y_2$ . We prefer this loss function over the Earth Mover Distance (EMD) or the MSE loss, due to its sample efficiency and fast convergence (See appendix Section A.3). + +# 3.3 Adaptive Coefficient Initialization + +We observe that appropriately initializing the coefficients plays a crucial role in determining the coefficients that best align with the user's aesthetic preferences. When a high coefficient is assigned to a specific task vector, it indicates a strong correlation between the user's aesthetic preference and that particular task. Thus, we adaptively initialize the coefficients based on the zero-shot personalization performances of each image score regression model associated with the task vector. + +To do this, we calculate the Spearman's rank-order correlation coefficient (SROCC) [28] between the user-provided samples and the scores predicted by each image score regression model. We then apply the softmax function to initialize the coefficients $\alpha$ as, + +$$ +\alpha_ {i} ^ {l} = \frac {e ^ {\mathrm {S R O C C} (S , \theta_ {\mathrm {f t} _ {i}})}}{\sum_ {j = 1} ^ {n} e ^ {\mathrm {S R O C C} (S , \theta_ {\mathrm {f t} _ {j}})}} \quad \mathrm {f o r} l = 1, 2, \ldots , L, \qquad \qquad (4) +$$ + +where $\mathrm{SROCC}(S,\theta_{\mathrm{ft}_i})$ indicates the SROCC calculated on the user-provided set $S$ using the model with parameters $\theta_{\mathrm{ft}_i}$ . Note that we initialize $\alpha$ values for a single task $i$ with the same value for all layers. + +Table 1: Description of the databases included in our experiments. The AADB and REAL-CUR databases are exclusively reserved as a test database. + +
DatabaseRegression taskAnnotator tagNumber of images
KonIQ-10K [12]IQA×10,073
SPAQ [6]IQA×11,125
AVA [27]GIAA×255,500
TAD66K [11]GIAA×67,125
Flickr-AES [31]GIAA/PIAA✓40,000
PARA [40]GIAA/PIAA✓31,220
AADB [19]GIAA/PIAA✓9,958
REAL-CUR [31]PIAA✓2,871
+ +# 4 Experiments + +# 4.1 Database Overview + +The key advantage of our approach is the flexibility to utilize multiple image score regression databases freely, even those without annotator tags. In our experiments, we fine-tune models using six different databases in Table 1, reserving the AADB [19] and REAL-CUR [31] database exclusively for testing. While the Flickr-AES [31] and PARA [40] databases do contain annotator tags, our approach does not require distinguishing between individuals. Therefore, we treat these two databases as GIAA databases. In total, we effectively leverage 415,043 samples from six databases, marking the most extensive utilization of samples in PIAA within our knowledge. Detailed descriptions and performance on these databases are provided in the appendix (See Section A.5). However, our approach is not limited to training solely on these six databases, having the scalability to include additional IQA and GIAA databases [4,8,10,16,42]. + +For testing, we use the following three most widely-used PIAA test sets. + +Flickr-AES [31] database consists of 40,000 images collected from 210 users. For testing purposes, we utilized a test set containing 4,737 images rated by 37 users, with each user contributing image collections ranging from 105 to 171. + +AADB [19] database consists of 9,958 images from 190 users. We selected a test set from 22 users, each of which contributed individual image collections containing 110 to 190 images. + +REAL-CUR [31] database consists of personal image collections from 14 users, totaling 2,871 images. Each user's collection ranges from 197 to 222 images. + +# 4.2 Implementation Details + +To obtain task vectors from the models fine-tuned on each of the six databases, it is essential to start from an identical pre-trained model. Recent approaches in IQA [18], GIAA [38], and PIAA [13] emphasize the importance of pre-training on large databases for training powerful image score regression models. Thus, we choose the publicly available OpenCLIP models ViT-B/16 and ViT-L/14 [15] as the pre-trained model. Full details regarding the architectural design and + +Table 2: Cross-database evaluation on the REAL-CUR [31] database. For the baseline approaches, we report the best cross-domain performances. + +
MethodSROCC
10-shot100-shot
PA-IAA [20]0.443±0.0040.562±0.013
BLG-PIAA [44]0.448±0.0070.578±0.015
PIAA-SOA [45]0.487±0.0030.589±0.015
TAPP-PIAA [21]-0.580
MTCL [41]0.495±0.0070.599±0.012
Ours0.577±0.0050.621±0.007
+ +training hyperparameters are provided in the appendix (See Section A.5 and A.6). We will also release the codes upon acceptance. + +If not stated otherwise, we use $n = 6$ fine-tuned models to derive the task vectors. Each model comprises $L = 296$ layers when using the ViT-L/14 model and $L = 152$ layers when using the ViT-B/16 model. While the number of 'layers' usually refers to the number of transformer blocks, we include all computational elements within these blocks, such as linear layers and normalization layers. The total number of learnable parameters, $n \times L = 1776$ and $n \times L = 912$ , is significantly lower compared to a full fine-tuning approach, which would require training over 23 million parameters even for small ResNet-50 models [9]. We do not consider the layer-wise task vector itself as a learnable parameter, as obtaining these vectors is not repeated for each individual. + +# 4.3 Evaluation Metric for PIAA + +PIAA approaches are often evaluated through the Spearman's rank-order correlation coefficient [28] (SROCC), which is the correlation coefficient between the predicted and ground-truth rank variables. Specifically, given the ground-truth and predicted scores $y$ and $\hat{y}$ , the SROCC is calculated as, + +$$ +\mathrm {S R O C C} = 1 - \frac {6 \sum_ {i = 1} ^ {N} \left(r _ {i} - \hat {r} _ {i}\right) ^ {2}}{N \left(N ^ {2} - 1\right)}, \tag {5} +$$ + +where $r_i$ and $\hat{r}_i$ represent the rank of the $i$ -th sample within the ground-truth and predicted scores, respectively, and $N$ denotes the number of samples. + +# 4.4 Cross-database Evaluation + +PIAA evaluation can be categorized based on whether training and testing occur within the same database (i.e., intra-database evaluation) or involve different databases (i.e., cross-database evaluation). In most real-world scenarios involving personal image collections, user-provided samples often diverge from the statistics or tendencies of the training database. Therefore, cross-database evaluation becomes pivotal in assessing the generalization capabilities of personalization approaches, which is an aspect that remains relatively underdeveloped in comparison to intra-database evaluations. + +Table 3: Cross-database evaluation on the AADB [19] database. Results marked with * indicate intra-database performance (i.e., directly trained on the AADB database). + +
MethodSROCC
10-shot100-shot
PA-IAA [20]0.469±0.0020.524±0.006
BLG-PIAA [44]0.486±0.0040.536±0.006
PIAA-SOA [45]0.509±0.0030.557±0.007
TAPP-PIAA [21]-0.540
MTCL [41]0.533±0.0040.572±0.007
*TAPP-PIAA [21]*0.534±0.004*0.612±0.007
*MTCL [41]*0.540±0.005*0.622±0.007
Ours0.556±0.0040.654±0.007
+ +For cross-database evaluation, we employ the AADB [19] and REAL-CUR [31] databases as the test databases, neither of which were among the six databases from which we derived our task vectors. Following established protocols [31], we randomly select $K$ samples for each individual to form the training set, also referred to as the support set, while the remaining images form the test set. In our experiments, we set $K$ as either 10-shot or 100-shot. Given the potential variability inherent in few-shot learning depending on the chosen $K$ support set, we conduct 10 independent trials for each user and report the average SROCC with the standard deviation. + +As illustrated in Table 2 and Table 3, our method significantly surpasses existing PIAA models in cross-database performance for both the 10-shot and 100-shot protocols. Remarkably, it even exceeds the performance of models trained explicitly on the AADB database, as compared in Table 3. Such results underscore the exceptional generalization capabilities of our approach, in contrast to previous methods that are constrained by their inability to leverage GIAA and IQA databases. This superior performance is primarily due to the scalability of our approach, which effectively harnesses an extensive amount of 415,043 samples across multiple databases, each contributing unique insights and tendencies. + +# 4.5 Intra-database Evaluation + +For intra-database evaluation, we test our approach for the 37 individuals in the Flickr-AES [31] database, which is the most widely used PIAA database. Since the Flickr-AES [31] database is one of the six databases used to derive the task vectors, this experiment falls under the category of intra-database evaluation. However, we do not use the annotator information in the Flickr-AES. + +In Table 4, we report the average SROCC for each approach along with the standard deviation across 10 independent trials. These experiments underscore the effectiveness of training personalized task vectors through the use of multiple fine-tuned models. Our approach consistently outperforms existing methods that rely on meta-learning techniques or incorporate additional user attributes (e.g., personality traits) to enhance personalized score predictions. + +Table 4: 10-shot and 100-shot personalization results on the Flickr-AES database [31], with the average SROCC of 37 users and standard deviation across 10 trials. + +
MethodSROCC
10-shot100-shot
PAM (attribute only) [31]0.511±0.0040.516±0.003
PAM (content only) [31]0.512±0.0020.516±0.010
PAM [31]0.513±0.0030.524±0.007
PASS [35]0.516±0.0030.521±0.007
USAR-PPR [25]0.521±0.0020.544±0.007
USAR-PAD [25]0.520±0.0030.537±0.003
USAR-PPR&PAD [25]0.525±0.0040.552±0.015
PA-IAA [20]0.543±0.0030.639±0.011
BLG-PIAA [44]0.561±0.0050.669±0.013
UG-PIAA [24]0.559±0.0020.660±0.013
PIAA-SOA [45]0.618±0.0060.691±0.015
TAPP-PIAA [21]0.591±0.0070.685±0.012
IM-PIAA [13]0.6200.708
MTCL [21]0.667±0.0050.737±0.014
Ours0.668±0.0040.748±0.012
+ +# 5 Analysis and Ablation Studies + +We perform an in-depth analysis and ablation studies on each component of our approach, including model size, scalability, and coefficient initialization. + +# 5.1 Task Similarities Across Databases + +To gain insights into the derived task vectors, we visualize the similarity matrix of the task vectors in Figure 4. For simplicity, we aggregate the $L$ task vectors derived from each layer of a single fine-tuned model as a single task vector by concatenating them into one vector. In line with the findings provided by Ilharco et al. [14], task vectors are typically orthogonal, even though the databases are collected with similar intentions (i.e., GIAA and IQA). This observation highlights that treating each database as a unique task remains a valid choice, owing to variations in their data collection methodologies and the diverse nature of the populations from which they are sourced. + +# 5.2 Architecture Choices and Variations + +Selecting the architecture. Weight mixing techniques, including task vectors [14, 36, 37], have been recognized for their effectiveness when fine-tuning models pre-trained on large databases, such as the LAION-5B [33] or DataComp [7]. Vision Transformers (ViT) [5] is a suitable architecture for accommodating web-scale databases, in contrast to ResNets [9], which often struggle to achieve high performance despite training on large databases. Therefore, given + +![](images/257fc4a21f7d034c3116344456b35f40df8843c93106eee83c6151a17a9c5f3b.jpg) +Fig. 4: Similarity matrix between task vectors derived from each database. The similarities are measured with the cosine similarity. + +Table 5: Performance and inference speed of the ViT-B/16 model and the Best-fit FT method (i.e., directly fine-tuning ViTs) on the REAL-CUR [31] database. + +
MethodSROCCFPS
10-shot100-shot
Best-fit FT0.303±0.0080.394±0.00894.41
PIAA-SOA [45]0.487±0.0030.589±0.015-
MTCL [41]0.495±0.0070.599±0.012359.05
Ours (ViT-B/16)0.562±0.0040.607±0.007417.95
Ours0.577±0.0050.621±0.00794.41
+ +its scalability and performance advantages, we choose the publicly available pretrained ViT as our architecture for our experiments. Regardless of the chosen architecture, our parameter-efficient approach only trains a minimal set of coefficients corresponding to each layer-wise task vector. + +Scaling down for efficient computation. While our primary experiments focus on combining task vectors derived from ViT-L/14 models, our approach is adaptable to smaller models as well. As shown in Table 5, our approach yields strong results even on smaller models like ViT-B/16, which are well-suited for deployment on commercial computers. Although fast inference was not our primary goal, we achieve real-time inference speeds ( $\geq 30$ FPS) on the RTX 3090, as our weight mixing technique combines multiple models into a single ViT model, eliminating the need to pass through multiple models. Acceleration of ViTs [22, 23, 43] beyond our implementation is also actively supported in recent frameworks [26] and communities [30], which ensures the practicality of our approach across a wide range of computing platforms such as mobile devices. + +Vision Transformers without task vectors. Our experimental results in Table 5 demonstrate that the performance gains are not solely attributed to + +![](images/fb42a99854c3c82dcd8107a702ae1cfd7a13d78c182c356f371608997fa0abd9.jpg) +Fig. 5: Performance gains with respect to the number of databases used. The dotted lines indicate the previous state-of-the-art performance. + +the inherent strengths of the ViT architecture. When directly fine-tuning the aesthetic assessment model that best matches the user's preference (i.e., Best-fit FT), we observe distinctly low personalization performances despite using a large pre-trained model. This highlights the pivotal role of keeping the layer-wise task vectors frozen during personalization, as altering the vectors would indicate discarding the rich preferences learned from a large database. + +# 5.3 Scalability of Our Approach + +How does each database contribute? We assess the scalability of our approach by varying the number of task vectors. Specifically, we evaluate our approach on the REAL-CUR database [31] by increasing the number of databases used. Starting from the GIAA model trained on the AVA [27] database, we accumulate the databases in the following order: Flicker-AES [31], TAD66K [11], PARA [40], KonIQ-10K [12], and SPAQ [6]. As demonstrated in Figure 5, both 10-shot and 100-shot performance significantly improves as the number of task vectors increases. This indicates that our scalable approach benefits from the expansion of task vectors, which allows us to capture a broader range of image characteristics and nuances. + +How can we increase the number of task vectors? The advantage of our approach can be amplified with the availability of additional image score regression databases. For example, scaling up the number of task vectors can be further pursued by incorporating readily available IQA [4,8,42] and GIAA [10,16] databases not included in our current study. Additionally, partitioning large databases, such as AVA [27] and PaQ-2-PiQ [42], can be an alternative method for generating additional task vectors. + +These strategies for obtaining a diverse set of task vectors may lead to performance gains in the 10-shot evaluation, which is considered highly challenging due to the limited number of user-provided samples. However, we leave this as future work, given that our current work already utilizes the largest number of + +![](images/2cccc0294fe52d3ac029fda7cdad6ed747a7a01e8bde055c3b02e8f530a84405.jpg) +Fig. 6: 2D-histogram of coefficients and zero-shot SROCC pairs. The size of the coefficients are correlated with the relevance of the task. + +samples for PIAA, 10 times larger than the most widely used and highly curated Flickr-AES database [31]. + +# 5.4 Task Relevance and Coefficient Initialization + +We conduct an analysis to better understand the distinguishing characteristics of coefficients, specifically, what sets apart those with high values from those with low values. For a straightforward examination, we calculate the average coefficient for each task (i.e., database), averaged across all the layers of the model, which results in six average coefficients for each user ( $n = 6$ ). Subsequently, we assess the six zero-shot SROCC scores associated with the average coefficients and plot the histogram of the coefficient-SROCC pairs from all test users in Figure 6. We observe that coefficients for tasks with a high zero-shot SROCC are trained to have high values, while those for less-relevant tasks (i.e., low zero-shot SROCC) exhibit low values. This analysis justifies our adaptive initialization of the coefficients, which assigns large values to coefficients associated with closely related tasks. + +We investigate two alternative initialization strategies: a uniform initialization and a best-fit initialization. Both cases can be viewed as an extreme case of the adaptive initialization, which can be written in a generalized form: + +$$ +\alpha_ {i} ^ {l} = \frac {e ^ {\mathrm {S R O C C} (S , \theta_ {\mathrm {f t} _ {i}}) / T}}{\sum_ {j = 1} ^ {n} e ^ {\mathrm {S R O C C} (S , \theta_ {\mathrm {f t} _ {j}}) / T}} \quad \text {f o r} l = 1, 2, \dots , L, \tag {6} +$$ + +where $T$ is the temperature for the softmax. The uniform initialization corresponds to the case where the $T \to \infty$ , while the best-fit initialization is when $T \to 0$ . As demonstrated in Table 6, the adaptive initialization outperforms other initialization strategies. Nevertheless, we consider further exploration of the initialization strategies as a potential avenue for future research. + +Table 6: Comparison between three different coefficient initialization strategies tested on the REAL-CUR [31] database. + +
MethodSROCC
10-shot100-shot
Uniform init. (T→∞)0.520±0.0020.602±0.002
Best-fit init. (T→0)0.464±0.0120.596±0.002
Adaptive init. (T=1)0.577±0.0050.621±0.007
+ +# 6 Limitation + +We have observed that when the zero-shot SROCC for all six image regression models are low for an individual (i.e., $\mathrm{SROCC} \approx 0$ ), our approach may struggle to identify an optimal combination of task vectors for that user. This phenomenon suggests that when there is a limited correlation between the user's preferences and the predictions of these models, the task vectors derived from them may be insufficient for finding an appropriate personalized task vector. + +However, we find this case to be extremely rare (See appendix Figure 2-7.), with only 1 among 73 cases having a low SROCC under 0.2, demonstrating the broad coverage of our approach. Also, increasing the number of task vectors will effectively mitigate this issue, leveraging the inherent scalability of our approach, which can accommodate a growing number of task vectors to improve personalization even in challenging scenarios. + +# 7 Conclusion + +We address the scalability issue in PIAA, a critical yet previously unaddressed aspect necessary for enhancing the generalization performance. Unlike previous studies limited to training on PIAA databases curated by individual annotators, our approach enables the flexible use of multiple image score regression databases, thereby providing greater scalability and generalization capabilities. + +We demonstrate that GIAA and IQA databases exhibit distinct personalization potentials, which have motivated us to combine GIAA and IQA models to achieve desired behaviors. By introducing learnable parameters for optimal model combinations, we enable the precise personalization of models to individual users' aesthetic preferences. Our approach significantly outperforms existing approaches, demonstrating exceptional efficacy in generalizing to unseen domains which is a critical requirement for real-world applications. This work opens up new avenues for personalized image aesthetic assessment, offering valuable insights and practical solutions for this challenging task. + +# Acknowledgement + +This work was supported by the Institute for Information & communications Technology Promotion(IITP) grant funded by the Korea government(MSIT) (No.RS-2019-II190075 Artificial Intelligence Graduate School Program(KAIST)), the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. NRF-2022R1A2B5B02001913), and the Institute of Information & communications Technology Promotion (IITP) grant funded by the Korea government (MSIT) (RS-2021-II212068, Artificial Intelligence Innovation Hub). + +# References + +1. Al-Shedivat, M., Li, L., Xing, E., Talwalkar, A.: On data efficiency of meta-learning. In: International Conference on Artificial Intelligence and Statistics. PMLR (2021) +2. Bradley, R.A., Terry, M.E.: Rank analysis of incomplete block designs: I. the method of paired comparisons. Biometrika (1952) +3. Chefer, H., Gur, S., Wolf, L.: Transformer interpretability beyond attention visualization. In: CVPR (2021) +4. Ciancio, A., da Silva, E.A., Said, A., Samadani, R., Obrador, P., et al.: No-reference blur assessment of digital pictures based on multifeature classifiers. IEEE Transactions on Image Processing (2010) +5. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2020) +6. Fang, Y., Zhu, H., Zeng, Y., Ma, K., Wang, Z.: Perceptual quality assessment of smartphone photography. In: CVPR (2020) +7. Gadre, S.Y., Ilharco, G., Fang, A., Hayase, J., Smyrnis, G., Nguyen, T., Marten, R., Wortsman, M., Ghosh, D., Zhang, J., et al.: Datacomp: In search of the next generation of multimodal datasets. arXiv preprint arXiv:2304.14108 (2023) +8. Ghadiyaram, D., Bovik, A.C.: Massive online crowdsourced study of subjective and objective picture quality. IEEE Transactions on Image Processing (2015) +9. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016) +10. He, S., Ming, A., Li, Y., Sun, J., Zheng, S., Ma, H.: Thinking image color aesthetics assessment: Models, datasets and benchmarks. In: ICCV (2023) +11. He, S., Zhang, Y., Xie, R., Jiang, D., Ming, A.: Rethinking image aesthetics assessment: Models, datasets and benchmarks. IJCAI (2022) +12. Hosu, V., Lin, H., Sziranyi, T., Saupe, D.: Koniq-10k: An ecologically valid database for deep learning of blind image quality assessment. IEEE Transactions on Image Processing (2020) +13. Hou, J., Lin, W., Yue, G., Liu, W., Zhao, B.: Interaction-matrix based personalized image aesthetics assessment. IEEE Transactions on Multimedia (2022) +14. Ilharco, G., Ribeiro, M.T., Wortsman, M., Schmidt, L., Hajishirzi, H., Farhadi, A.: Editing models with task arithmetic. In: ICLR (2023) +15. Ilharco, G., Wortsman, M., Wightman, R., Gordon, C., Carlini, N., Taori, R., Dave, A., Shankar, V., Namkoong, H., Miller, J., Hajishirzi, H., Farhadi, A., Schmidt, L.: Openclip (Jul 2021). https://doi.org/10.5281/zenodo.5143773, https://doi.org/10.5281/zenodo.5143773 +16. Jin, X., Wu, L., Zhao, G., Li, X., Zhang, X., Ge, S., Zou, D., Zhou, B., Zhou, X.: Aesthetic attributes assessment of images. In: ACM MM (2019) +17. Karlsson, K., Jiang, W., Zhang, D.Q.: Mobile photo album management with multiscale timeline. In: ACM MM (2014) +18. Ke, J., Wang, Q., Wang, Y., Milanfar, P., Yang, F.: Musiq: Multi-scale image quality transformer. In: ICCV (2021) +19. Kong, S., Shen, X., Lin, Z., Mech, R., Fowlkes, C.: Photo aesthetics ranking network with attributes and content adaptation. In: ECCV. Springer (2016) +20. Li, L., Zhu, H., Zhao, S., Ding, G., Lin, W.: Personality-assisted multi-task learning for generic and personalized image aesthetics assessment. IEEE Transactions on Image Processing 29 (2020) + +21. Li, Y., Yang, Y., Li, H., Chen, H., Xu, L., Li, L., Li, Y., Guo, Y.: Transductive aesthetic preference propagation for personalized image aesthetics assessment. In: ACM MM. ACM (2022) +22. Lin, Y., Zhang, T., Sun, P., Li, Z., Zhou, S.: Fq-vit: Post-training quantization for fully quantized vision transformer. IJCAI (2022) +23. Liu, Z., Wang, Y., Han, K., Zhang, W., Ma, S., Gao, W.: Post-training quantization for vision transformer. NeurIPS (2021) +24. Lv, P., Fan, J., Nie, X., Dong, W., Jiang, X., Zhou, B., Xu, M., Xu, C.: User-guided personalized image aesthetic assessment based on deep reinforcement learning. IEEE Transactions on Multimedia (2021) +25. Lv, P., Wang, M., Xu, Y., Peng, Z., Sun, J., Su, S., Zhou, B., Xu, M.: Usar: An interactive user-specific aesthetic ranking framework for images. In: ACM MM. pp. 1328-1336 (2018) +26. Microsoft: Deepspeed (2023), https://www.deepspeed.ai/ +27. Murray, N., Marchesotti, L., Perronnin, F.: Ava: A large-scale database for aesthetic visual analysis. In: CVPR. IEEE (2012) +28. Myers, J.L., Well, A.D., Lorch Jr, R.F.: Research design and statistical analysis. Routledge (2013) +29. Naseer, M.M., Ranasinghe, K., Khan, S.H., Hayat, M., Shahbaz Khan, F., Yang, M.H.: Intriguing properties of vision transformers. NeurIPS (2021) +30. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. NeurIPS (2019) +31. Ren, J., Shen, X., Lin, Z., Mech, R., Foran, D.J.: Personalized image aesthetics. In: ICCV (Oct 2017) +32. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. 2022 IEEE. In: CVPR (2021) +33. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. NeurIPS (2022) +34. Wallace, B., Gokul, A., Ermon, S., Naik, N.: End-to-end diffusion latent optimization improves classifier guidance. ICCV (2023) +35. Wang, G., Yan, J., Qin, Z.: Collaborative and attentive learning for personalized image aesthetic assessment. In: IJCAI. pp. 957-963 (2018) +36. Wortsman, M., Ilharco, G., Gadre, S.Y., Roelofs, R., Gontijo-Lopes, R., Morcos, A.S., Namkoong, H., Farhadi, A., Carmon, Y., Kornblith, S., et al.: Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In: ICML. PMLR (2022) +37. Wortsman, M., Ilharco, G., Li, M., Kim, J.W., Hajishirzi, H., Farhadi, A., Namkoong, H., Schmidt, L.: Robust fine-tuning of zero-shot models. 2022 IEEE. In: CVPR (2021) +38. Xu, L., Xu, J., Yang, Y., Huang, Y., Xie, Y., Li, Y.: Clip brings better features to visual aesthetics learners. arXiv preprint arXiv:2307.15640 (2023) +39. Yan, X., Shao, F., Chen, H., Jiang, Q.: Hybrid cnn-transformer based meta-learning approach for personalized image aesthetics assessment. Journal of Visual Communication and Image Representation 98 (2024) +40. Yang, Y., Xu, L., Li, L., Qie, N., Li, Y., Zhang, P., Guo, Y.: Personalized image aesthetics assessment with rich attributes. In: CVPR (2022) +41. Yang, Z., Li, L., Yang, Y., Li, Y., Lin, W.: Multi-level transitional contrast learning for personalized image aesthetics assessment. IEEE Transactions on Multimedia (2023) + +42. Ying, Z., Niu, H., Gupta, P., Mahajan, D., Ghadiyaram, D., Bovik, A.: From patches to pictures (paq-2-piq): Mapping the perceptual space of picture quality. In: CVPR (2020) +43. Yuan, Z., Xue, C., Chen, Y., Wu, Q., Sun, G.: Ptq4vit: Post-training quantization for vision transformers with twin uniform quantization. In: ECCV. Springer (2022) +44. Zhu, H., Li, L., Wu, J., Zhao, S., Ding, G., Shi, G.: Personalized image aesthetics assessment via meta-learning with bilevel gradient optimization. IEEE Transactions on Cybernetics (2020) +45. Zhu, H., Zhou, Y., Li, L., Li, Y., Guo, Y.: Learning personalized image aesthetics from subjective and objective attributes. IEEE Transactions on Multimedia (2021) \ No newline at end of file diff --git a/2024/Scaling Up Personalized Image Aesthetic Assessment via Task Vector Customization/images.zip b/2024/Scaling Up Personalized Image Aesthetic Assessment via Task Vector Customization/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..204c668fe84e68e81d7f7e1f27a899fae88a4e9b --- /dev/null +++ b/2024/Scaling Up Personalized Image Aesthetic Assessment via Task Vector Customization/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:291de62ee206a81a49e05ae2d72dc9623ad74892d1e4b980ea2386af1cdabd3e +size 503255 diff --git a/2024/Scaling Up Personalized Image Aesthetic Assessment via Task Vector Customization/layout.json b/2024/Scaling Up Personalized Image Aesthetic Assessment via Task Vector Customization/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5c3e3654261fd9050b244fcce70e45017a508139 --- /dev/null +++ b/2024/Scaling Up Personalized Image Aesthetic Assessment via Task Vector Customization/layout.json @@ -0,0 +1,8057 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 157, + 111, + 457, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 111, + 457, + 146 + ], + "spans": [ + { + "bbox": [ + 157, + 111, + 457, + 146 + ], + "type": "text", + "content": "Scaling Up Personalized Image Aesthetic Assessment via Task Vector Customization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 233, + 168, + 380, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 168, + 380, + 180 + ], + "spans": [ + { + "bbox": [ + 233, + 168, + 380, + 180 + ], + "type": "text", + "content": "Jooyeol Yun and Jaegul Choo" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 179, + 190, + 433, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 190, + 433, + 213 + ], + "spans": [ + { + "bbox": [ + 179, + 190, + 433, + 213 + ], + "type": "text", + "content": "Korea Advanced Institute of Science and Technology (KAIST) blizzard072@kaist.ac.kr jchoo@kaist.ac.kr" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 224, + 455, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 224, + 455, + 434 + ], + "spans": [ + { + "bbox": [ + 160, + 224, + 455, + 434 + ], + "type": "text", + "content": "Abstract. The task of personalized image aesthetic assessment seeks to tailor aesthetic score prediction models to match individual preferences with just a few user-provided inputs. However, the scalability and generalization capabilities of current approaches are considerably restricted by their reliance on an expensive curated database. To overcome this long-standing scalability challenge, we present a unique approach that leverages readily available databases for general image aesthetic assessment and image quality assessment. Specifically, we view each database as a distinct image score regression task that exhibits varying degrees of personalization potential. By determining optimal combinations of task vectors, known to represent specific traits of each database, we successfully create personalized models for individuals. This approach of integrating multiple models allows us to harness a substantial amount of data. Our extensive experiments demonstrate the effectiveness of our approach in generalizing to previously unseen domains—a challenge previous approaches have struggled to achieve—making it highly applicable to real-world scenarios. Our novel approach significantly advances the field by offering scalable solutions for personalized aesthetic assessment and establishing high standards for future research." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 446, + 230, + 458 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 230, + 458 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 230, + 458 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 470, + 482, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 470, + 482, + 566 + ], + "spans": [ + { + "bbox": [ + 130, + 470, + 482, + 566 + ], + "type": "text", + "content": "Personalized image aesthetic assessment (PIAA) is an emerging field dedicated to developing aesthetic score prediction models that closely match an individual's aesthetic preferences based on a small set of user-provided samples. While they function as simple models that assign scores based on aesthetic quality, PIAA models enhance user experience in digital environments by personalizing tasks such as managing photo albums [17], curating web-scale databases like LAION-Aesthetics [33], and even guiding generative models such as Stable Diffusion [32] to create images that align with individual tastes [34]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 566, + 482, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 566, + 482, + 650 + ], + "spans": [ + { + "bbox": [ + 130, + 566, + 482, + 650 + ], + "type": "text", + "content": "In recent years, there has been a notable trend in PIAA [21, 41, 44] that involves the integration of meta-learning techniques, as personalization is inherently related to few-shot learning. Specifically, these approaches entail pretraining a model on a database consisting of image collections from diverse individuals. However, these strategies encounter significant challenges in terms of scalability and generalization to unseen domains, primarily due to the prohibitive cost associated with data collection." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 466, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 466, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 466, + 666 + ], + "type": "text", + "content": "1 https://yeolj00.github.io/personal-projects/personalized-aesthetics/" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 184, + 114, + 427, + 213 + ], + "blocks": [ + { + "bbox": [ + 184, + 114, + 427, + 213 + ], + "lines": [ + { + "bbox": [ + 184, + 114, + 427, + 213 + ], + "spans": [ + { + "bbox": [ + 184, + 114, + 427, + 213 + ], + "type": "image", + "image_path": "2f8f78a387819845d903c928eb9b0ae98f9784f842142670e7d9fc764cbb51e7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 216, + 480, + 239 + ], + "lines": [ + { + "bbox": [ + 131, + 216, + 480, + 239 + ], + "spans": [ + { + "bbox": [ + 131, + 216, + 480, + 239 + ], + "type": "text", + "content": "Fig. 1: Comparison between PIAA and other image regression databases. Our approach has the flexibility to leverage any image regression databases." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 245, + 480, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 245, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 245, + 480, + 365 + ], + "type": "text", + "content": "While meta-learning typically demands an extensive number of meta-training tasks, often exceeding 10,000, to achieve high performance [1], currently available PIAA databases [19, 31] offer a limited number of tasks fewer than 200. Even with the recently collected PIAA database [40], which contains rich attributes from each annotator, the number of training tasks remains below 500. In the context of PIAA, a single meta-training task corresponds to learning the preferences of a single individual. Scaling such a database would require obtaining a vast number of images and corresponding personalized aesthetic scores from numerous individuals, which poses severe limitations in the scalability and generalization of existing approaches." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 365, + 481, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 481, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 481, + 461 + ], + "type": "text", + "content": "In this paper, we unveil a data-rich approach for personalizing aesthetic assessment models by leveraging existing general image aesthetic assessment (GIAA) and image quality assessment (IQA) databases. In contrast to prior work [20, 21, 41, 44, 45], our approach is not limited to training on databases that track individual annotators, as depicted in Figure 1. Instead, we have the flexibility to utilize multiple image score regression databases, which are readily available. By employing broader resources, we overcome the scalability and generalization challenges that have long been unaddressed in this field." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 176, + 467, + 434, + 567 + ], + "blocks": [ + { + "bbox": [ + 176, + 467, + 434, + 567 + ], + "lines": [ + { + "bbox": [ + 176, + 467, + 434, + 567 + ], + "spans": [ + { + "bbox": [ + 176, + 467, + 434, + 567 + ], + "type": "image", + "image_path": "88277c622dc9f4f83ad4434af9fd83046166b75590aa18af1090d1ac57018053.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 568, + 480, + 601 + ], + "lines": [ + { + "bbox": [ + 131, + 568, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 131, + 568, + 480, + 601 + ], + "type": "text", + "content": "Fig.2: Zero-shot personalization performance for four users in the Flickr-AES database [31]. Each colored bar refers to the personalization performance of a model trained on a specific database." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": "The key idea is to view each GIAA and IQA data as a unique image score regression task. This perspective arises from the distinct characteristics and tendencies of each database due to different data collection procedures and the population from which they originate. As visualized in Figure 2, image score regression models trained on four different GIAA [27,31] and IQA [6,12] databases" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 254, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 254, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 254, + 100 + ], + "type": "text", + "content": "JY. Yun and J. Choo" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "text", + "content": "exhibit varying levels of personalization for users within the PIAA database. For instance, User 3's aesthetic preferences closely align with the score regression model trained on the Flickr-AES [31] while diverging from the model trained on the SPAQ [6] database. In other words, even without any personalized training, we find that general image assessment models inherently possess different degrees of personalization capabilities." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 190, + 480, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 190, + 480, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 190, + 480, + 285 + ], + "type": "text", + "content": "In light of these insights, we extend an emerging approach known as task arithmetic [14], which was originally devised for developing multi-task models. This method involves directly adding or subtracting the weights of models fine-tuned on specific tasks. In our context, task vectors from diverse image assessment models encapsulate distinct features and behaviors related to specific tasks, such as recognizing different aspects of image quality or aesthetics. By carefully combining task vectors, we explore the potential to selectively amplify or refine the capabilities of a pre-trained model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 287, + 480, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 287, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 287, + 480, + 430 + ], + "type": "text", + "content": "However, there has been limited research on determining the optimal adjustments necessary to achieve a certain behavior. Thus, we present a personalization approach by introducing trainable coefficients for each task vector derived from models trained on distinct GIAA and IQA databases, thereby determining the weights for combining these vectors. Once the coefficients are trained with user-provided inputs, the weighted sum of task vectors creates a model tailored to the user's aesthetic preferences. Since task vectors obtained from large databases already capture preferences across various themes and contents, we find that training only the coefficients is sufficient for precise personalization, making our approach highly parameter-efficient. To the best of our knowledge, our approach marks the first to customize task vectors for specific model behaviors that were previously unattainable." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 431, + 480, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 431, + 480, + 527 + ], + "spans": [ + { + "bbox": [ + 130, + 431, + 480, + 527 + ], + "type": "text", + "content": "By carefully customizing task vectors, we incorporate large databases into the training process, granting our approach a strong generalization capability for unseen domains, a scenario often encountered in real-world applications. Furthermore, we find that our approach is also practical for learning a user's aesthetic preference with only a few samples, as our approach of merging task vectors guides the model to make well-informed updates and prevents overfitting. We provide extensive experiments demonstrating the efficacy of our approach for personalization in diverse scenarios." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 147, + 529, + 288, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 529, + 288, + 540 + ], + "spans": [ + { + "bbox": [ + 147, + 529, + 288, + 540 + ], + "type": "text", + "content": "Our contributions are threefold:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 139, + 549, + 480, + 664 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 139, + 549, + 479, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 549, + 479, + 585 + ], + "spans": [ + { + "bbox": [ + 139, + 549, + 479, + 585 + ], + "type": "text", + "content": "- We introduce a novel data-rich approach for PIAA that tackles the long-standing scalability challenge, eliminating the dependency on expensive, manually-curated databases that have hindered progress in this field." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 139, + 590, + 480, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 590, + 480, + 624 + ], + "spans": [ + { + "bbox": [ + 139, + 590, + 480, + 624 + ], + "type": "text", + "content": "- Demonstrating exceptional effectiveness in cross-database evaluations, our method showcases robust generalization capabilities that surpass those of existing approaches, setting high standards for future work." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 139, + 629, + 479, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 629, + 479, + 664 + ], + "spans": [ + { + "bbox": [ + 139, + 629, + 479, + 664 + ], + "type": "text", + "content": "- By learning optimal combinations of task vectors, we present a parameter-efficient approach for precise personalization, leveraging the comprehensive information embedded in these vectors." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 231, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 231, + 91, + 447, + 102 + ], + "type": "text", + "content": "Scaling Up Personalized Image Aesthetic Assessment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 137, + 370, + 149 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 137, + 370, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 137, + 370, + 149 + ], + "type": "text", + "content": "2.1 Personalized Image Aesthetic Assessment" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 156, + 481, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 156, + 481, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 156, + 481, + 289 + ], + "type": "text", + "content": "Personalized image aesthetic assessment (PIAA) [13,20,21,24,25,35,39-41,44,45] aims to learn the personal aesthetic preference of individual users based on the image-score pairs from one's image collection. Since collecting a large number of samples from a single user is impractical, PIAA is closely linked to few-shot learning methods. One line of work [20,40,45] incorporates additional attributes (e.g., personality traits, age, and gender) to facilitate user-specific aesthetic predictions. These approaches rely on the correlation between user attributes and aesthetic preferences, which may not always be clearly defined. Consequently, many high-performing approaches for PIAA [21, 41, 44] leverage meta-learning techniques to train models that can be easily fine-tuned to accommodate the aesthetic preferences of new users with only a limited number of samples." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 289, + 481, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 481, + 421 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 481, + 421 + ], + "type": "text", + "content": "Nevertheless, applying meta-learning techniques to PIAA encounters scalability challenges owing to the limited number of tasks available during the meta-learning phase. In traditional meta-learning, achieving a high performance typically requires a substantial number of meta-training tasks [1], often exceeding 10,000. In PIAA databases [19,31], the number of training tasks ranges from 100 to fewer than 500, which can impose limitations on the performance. Expanding the task pool in PIAA is prohibitively expensive since each task corresponds to the image collection of a single user, posing significant scalability challenges that has long been unaddressed. We tackle this issue in previous PIAA approaches and present a scalable fine-tuning method that leverages existing general image aesthetic assessment (GIAA) and image quality assessment (IQA) databases." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 434, + 334, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 434, + 334, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 434, + 334, + 445 + ], + "type": "text", + "content": "2.2 Task Vectors and Task Arithmetic" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 454, + 481, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 454, + 481, + 491 + ], + "spans": [ + { + "bbox": [ + 132, + 454, + 481, + 491 + ], + "type": "text", + "content": "Ilharco et al. [14] presented a new paradigm for creating multi-task models, utilizing the parameters of multiple fine-tuned models. They define the task vector " + }, + { + "bbox": [ + 132, + 454, + 481, + 491 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 132, + 454, + 481, + 491 + ], + "type": "text", + "content": " as the difference between the fine-tuned and pre-trained weights," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 276, + 501, + 338, + 513 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 501, + 338, + 513 + ], + "spans": [ + { + "bbox": [ + 276, + 501, + 338, + 513 + ], + "type": "interline_equation", + "content": "\\tau = \\theta_ {\\mathrm {f t}} - \\theta_ {\\mathrm {p r e}},", + "image_path": "fb466fd9c70d033a893830c54d738fbf26dcc1a9ca7faa304de142d235ab4913.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 521, + 481, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 481, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 481, + 593 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 521, + 481, + 593 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{ft}}" + }, + { + "bbox": [ + 130, + 521, + 481, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 521, + 481, + 593 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{pre}}" + }, + { + "bbox": [ + 130, + 521, + 481, + 593 + ], + "type": "text", + "content": " represent the weights of the fine-tuned and pre-trained model, respectively. They demonstrate that adding and subtracting task vectors to pretrained models can encourage or suppress the ability to perform specific tasks. For example, negating the task vector derived from a language model fine-tuned on toxic data reduces the frequency of generating toxic text. However, it remains unclear what the optimal adjustments are for achieving a specific behavior." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": "In our paper, we investigate methods to determine necessary adjustments to achieve a certain behavior defined by a few user-provided samples. By introducing learnable coefficients to task vectors, we effectively customize models to closely match individual preferences. We also demonstrate that finding the optimal coefficients is possible with only a few samples, as the task vectors encompass comprehensive knowledge derived from large databases." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 254, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 254, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 254, + 100 + ], + "type": "text", + "content": "JY. Yun and J. Choo" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 477, + 167 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 477, + 167 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 477, + 167 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 477, + 167 + ], + "type": "image", + "image_path": "0020b54c382bcb578b60cfb77bf859bbbc2620db6b71d62d99f591b3dcca21ba.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 228, + 169, + 390, + 178 + ], + "lines": [ + { + "bbox": [ + 228, + 169, + 390, + 178 + ], + "spans": [ + { + "bbox": [ + 228, + 169, + 390, + 178 + ], + "type": "text", + "content": "(a) Obtain task vectors from various databases" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 180, + 482, + 274 + ], + "blocks": [ + { + "bbox": [ + 133, + 180, + 482, + 274 + ], + "lines": [ + { + "bbox": [ + 133, + 180, + 482, + 274 + ], + "spans": [ + { + "bbox": [ + 133, + 180, + 482, + 274 + ], + "type": "image", + "image_path": "4f6382017d196c68d3c1153419e1c9131ca888072bd07f54c09a7dda609688b1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 275, + 398, + 286 + ], + "lines": [ + { + "bbox": [ + 220, + 275, + 398, + 286 + ], + "spans": [ + { + "bbox": [ + 220, + 275, + 398, + 286 + ], + "type": "text", + "content": "(b) Learn optimal combinations for personalization" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 292, + 482, + 337 + ], + "lines": [ + { + "bbox": [ + 131, + 292, + 482, + 337 + ], + "spans": [ + { + "bbox": [ + 131, + 292, + 482, + 337 + ], + "type": "text", + "content": "Fig. 3: (a) Multiple models are fine-tuned from a single pre-trained model on different tasks to obtain task vectors. For simplicity, we omit the discussion of layers. (b) Given a small number of user-provided samples, the coefficients corresponding to each task vector are optimized with the rank loss." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 340, + 202, + 353 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 340, + 202, + 353 + ], + "spans": [ + { + "bbox": [ + 132, + 340, + 202, + 353 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 360, + 459, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 360, + 459, + 372 + ], + "spans": [ + { + "bbox": [ + 131, + 360, + 459, + 372 + ], + "type": "text", + "content": "3.1 Obtaining Layer-wise Task Vectors from Various Databases" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 382, + 482, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 382, + 482, + 466 + ], + "spans": [ + { + "bbox": [ + 131, + 382, + 482, + 466 + ], + "type": "text", + "content": "To obtain task vectors, we train multiple models with the same architecture on various tasks (i.e., different databases), as illustrated in Figure 3 (a). Specifically, we fine-tune the weights from a pre-trained backbone, adjusting the parameters from " + }, + { + "bbox": [ + 131, + 382, + 482, + 466 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{pre}}" + }, + { + "bbox": [ + 131, + 382, + 482, + 466 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 131, + 382, + 482, + 466 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{ft}_i}" + }, + { + "bbox": [ + 131, + 382, + 482, + 466 + ], + "type": "text", + "content": " for each task " + }, + { + "bbox": [ + 131, + 382, + 482, + 466 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 382, + 482, + 466 + ], + "type": "text", + "content": ". Moreover, while the specific roles of model layers in capturing unique features are inherently complex, their distinct contributions are undeniable [3,29]. Thus, we compute task vectors across all layers, enhancing the flexibility of our approach for personalization (See appendix Section A.2)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 467, + 481, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 467, + 481, + 491 + ], + "spans": [ + { + "bbox": [ + 131, + 467, + 481, + 491 + ], + "type": "text", + "content": "Accordingly, we expand the concept of original task vectors to include layerwise differentiation by," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 271, + 491, + 481, + 507 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 271, + 491, + 481, + 507 + ], + "spans": [ + { + "bbox": [ + 271, + 491, + 481, + 507 + ], + "type": "interline_equation", + "content": "\\tau_ {i} ^ {l} = \\theta_ {\\mathrm {f t} _ {i}} ^ {l} - \\theta_ {\\mathrm {p r e}} ^ {l}, \\tag {1}", + "image_path": "13078ea7e931f9722ab4c6c7b140eeb01c331ed35b63d5ad1c4837205108680d.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 513, + 481, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 513, + 481, + 550 + ], + "spans": [ + { + "bbox": [ + 131, + 513, + 481, + 550 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 513, + 481, + 550 + ], + "type": "inline_equation", + "content": "\\tau_{i}^{l}" + }, + { + "bbox": [ + 131, + 513, + 481, + 550 + ], + "type": "text", + "content": " is the task vector for the " + }, + { + "bbox": [ + 131, + 513, + 481, + 550 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 131, + 513, + 481, + 550 + ], + "type": "text", + "content": "-th layer of the " + }, + { + "bbox": [ + 131, + 513, + 481, + 550 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 513, + 481, + 550 + ], + "type": "text", + "content": "-th task. Each vector comprehensively integrates patterns learned from large databases, acting as basis vectors for robust and precise personalization." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 551, + 481, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 551, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 131, + 551, + 481, + 588 + ], + "type": "text", + "content": "It is important to note that this process is conducted only once and is not required again for personalizing to individual users, thereby not contributing to the count of learnable parameters for personalization." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 606, + 421, + 619 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 606, + 421, + 619 + ], + "spans": [ + { + "bbox": [ + 131, + 606, + 421, + 619 + ], + "type": "text", + "content": "3.2 Learning Optimal Combinations for Personalization" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 629, + 481, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 629, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 131, + 629, + 481, + 667 + ], + "type": "text", + "content": "Given the layer-wise task vectors, our goal is to find a combination of these vectors that accurately reflects a user's aesthetic preferences. To achieve this, we introduce learnable coefficients " + }, + { + "bbox": [ + 131, + 629, + 481, + 667 + ], + "type": "inline_equation", + "content": "\\alpha_{i}^{l}" + }, + { + "bbox": [ + 131, + 629, + 481, + 667 + ], + "type": "text", + "content": " for each task vector " + }, + { + "bbox": [ + 131, + 629, + 481, + 667 + ], + "type": "inline_equation", + "content": "\\tau_{i}^{l}" + }, + { + "bbox": [ + 131, + 629, + 481, + 667 + ], + "type": "text", + "content": ". Once the coefficients" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 231, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 231, + 91, + 447, + 102 + ], + "type": "text", + "content": "Scaling Up Personalized Image Aesthetic Assessment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 478, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 478, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 478, + 128 + ], + "type": "text", + "content": "are trained, the weights of the personalized model for each layer are derived as," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 219, + 138, + 481, + 168 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 138, + 481, + 168 + ], + "spans": [ + { + "bbox": [ + 219, + 138, + 481, + 168 + ], + "type": "interline_equation", + "content": "\\theta_ {\\mathrm {p}} ^ {l} = \\theta_ {\\mathrm {p r e}} ^ {l} + \\sum_ {i = 1} ^ {n} \\alpha_ {i} ^ {l} \\tau_ {i} ^ {l} \\quad \\text {f o r} l = 1, 2, \\dots , L, \\tag {2}", + "image_path": "a71ef96de863cd4d46d8f34030e0592079bed14680e10131056e725da01dfaba.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 176, + 479, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 479, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 479, + 248 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 176, + 479, + 248 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 176, + 479, + 248 + ], + "type": "text", + "content": " is the number of tasks, and " + }, + { + "bbox": [ + 130, + 176, + 479, + 248 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 176, + 479, + 248 + ], + "type": "text", + "content": " is the number of layers in the model. We keep all parameters other than the " + }, + { + "bbox": [ + 130, + 176, + 479, + 248 + ], + "type": "inline_equation", + "content": "n \\times L" + }, + { + "bbox": [ + 130, + 176, + 479, + 248 + ], + "type": "text", + "content": " coefficients frozen. This is essential since the task vectors contain comprehensive insights into aesthetic preferences, enabling the model to make well-informed updates. By harnessing the rich information embedded in these vectors, our method efficiently learns complex aesthetic preferences from merely a few user-provided samples." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 249, + 480, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 249, + 480, + 320 + ], + "spans": [ + { + "bbox": [ + 130, + 249, + 480, + 320 + ], + "type": "text", + "content": "Our objective function is designed to penalize the misordering of aesthetic scores. Specifically, we adopt the Bradley-Terry model [2], which estimates the probability of pairwise comparison between two samples. We frame the prediction of the two scores as a binary classification problem, where the objective is to predict which sample has the higher score. The coefficients of each task vector are trained to minimize the following loss function:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 222, + 331, + 481, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 331, + 481, + 354 + ], + "spans": [ + { + "bbox": [ + 222, + 331, + 481, + 354 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {R}} = - \\log \\frac {\\exp \\hat {y} _ {1}}{\\exp \\hat {y} _ {1} + \\exp \\hat {y} _ {2}} (y _ {1} \\geq y _ {2}), \\tag {3}", + "image_path": "0cd0a9a805b00be94cf0cc78bc0614c470e7ad885e2c03a5763c8a583a12a57f.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 361, + 482, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 361, + 482, + 411 + ], + "spans": [ + { + "bbox": [ + 131, + 361, + 482, + 411 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 361, + 482, + 411 + ], + "type": "inline_equation", + "content": "\\hat{y}_1" + }, + { + "bbox": [ + 131, + 361, + 482, + 411 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 361, + 482, + 411 + ], + "type": "inline_equation", + "content": "\\hat{y}_2" + }, + { + "bbox": [ + 131, + 361, + 482, + 411 + ], + "type": "text", + "content": " represent the aesthetic score predictions for images " + }, + { + "bbox": [ + 131, + 361, + 482, + 411 + ], + "type": "inline_equation", + "content": "x_1" + }, + { + "bbox": [ + 131, + 361, + 482, + 411 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 361, + 482, + 411 + ], + "type": "inline_equation", + "content": "x_2" + }, + { + "bbox": [ + 131, + 361, + 482, + 411 + ], + "type": "text", + "content": ", whose ground-truth scores are " + }, + { + "bbox": [ + 131, + 361, + 482, + 411 + ], + "type": "inline_equation", + "content": "y_1" + }, + { + "bbox": [ + 131, + 361, + 482, + 411 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 361, + 482, + 411 + ], + "type": "inline_equation", + "content": "y_2" + }, + { + "bbox": [ + 131, + 361, + 482, + 411 + ], + "type": "text", + "content": ". We prefer this loss function over the Earth Mover Distance (EMD) or the MSE loss, due to its sample efficiency and fast convergence (See appendix Section A.3)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 428, + 331, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 428, + 331, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 428, + 331, + 441 + ], + "type": "text", + "content": "3.3 Adaptive Coefficient Initialization" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 449, + 479, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 479, + 530 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 479, + 530 + ], + "type": "text", + "content": "We observe that appropriately initializing the coefficients plays a crucial role in determining the coefficients that best align with the user's aesthetic preferences. When a high coefficient is assigned to a specific task vector, it indicates a strong correlation between the user's aesthetic preference and that particular task. Thus, we adaptively initialize the coefficients based on the zero-shot personalization performances of each image score regression model associated with the task vector." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 533, + 480, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 533, + 480, + 582 + ], + "spans": [ + { + "bbox": [ + 131, + 533, + 480, + 582 + ], + "type": "text", + "content": "To do this, we calculate the Spearman's rank-order correlation coefficient (SROCC) [28] between the user-provided samples and the scores predicted by each image score regression model. We then apply the softmax function to initialize the coefficients " + }, + { + "bbox": [ + 131, + 533, + 480, + 582 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 131, + 533, + 480, + 582 + ], + "type": "text", + "content": " as," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 207, + 590, + 481, + 626 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 590, + 481, + 626 + ], + "spans": [ + { + "bbox": [ + 207, + 590, + 481, + 626 + ], + "type": "interline_equation", + "content": "\\alpha_ {i} ^ {l} = \\frac {e ^ {\\mathrm {S R O C C} (S , \\theta_ {\\mathrm {f t} _ {i}})}}{\\sum_ {j = 1} ^ {n} e ^ {\\mathrm {S R O C C} (S , \\theta_ {\\mathrm {f t} _ {j}})}} \\quad \\mathrm {f o r} l = 1, 2, \\ldots , L, \\qquad \\qquad (4)", + "image_path": "0b3e28e79831b34b70ae405e0abfdfae2951fc9e26f7d429d3db797f5ae354d8.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 632, + 482, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 632, + 482, + 670 + ], + "spans": [ + { + "bbox": [ + 131, + 632, + 482, + 670 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 632, + 482, + 670 + ], + "type": "inline_equation", + "content": "\\mathrm{SROCC}(S,\\theta_{\\mathrm{ft}_i})" + }, + { + "bbox": [ + 131, + 632, + 482, + 670 + ], + "type": "text", + "content": " indicates the SROCC calculated on the user-provided set " + }, + { + "bbox": [ + 131, + 632, + 482, + 670 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 131, + 632, + 482, + 670 + ], + "type": "text", + "content": " using the model with parameters " + }, + { + "bbox": [ + 131, + 632, + 482, + 670 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{ft}_i}" + }, + { + "bbox": [ + 131, + 632, + 482, + 670 + ], + "type": "text", + "content": ". Note that we initialize " + }, + { + "bbox": [ + 131, + 632, + 482, + 670 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 131, + 632, + 482, + 670 + ], + "type": "text", + "content": " values for a single task " + }, + { + "bbox": [ + 131, + 632, + 482, + 670 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 632, + 482, + 670 + ], + "type": "text", + "content": " with the same value for all layers." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 254, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 254, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 254, + 101 + ], + "type": "text", + "content": "JY. Yun and J. Choo" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 138, + 475, + 261 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 136 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 136 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 136 + ], + "type": "text", + "content": "Table 1: Description of the databases included in our experiments. The AADB and REAL-CUR databases are exclusively reserved as a test database." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 138, + 475, + 261 + ], + "lines": [ + { + "bbox": [ + 136, + 138, + 475, + 261 + ], + "spans": [ + { + "bbox": [ + 136, + 138, + 475, + 261 + ], + "type": "table", + "html": "
DatabaseRegression taskAnnotator tagNumber of images
KonIQ-10K [12]IQA×10,073
SPAQ [6]IQA×11,125
AVA [27]GIAA×255,500
TAD66K [11]GIAA×67,125
Flickr-AES [31]GIAA/PIAA✓40,000
PARA [40]GIAA/PIAA✓31,220
AADB [19]GIAA/PIAA✓9,958
REAL-CUR [31]PIAA✓2,871
", + "image_path": "8460e82d30767886a45d3bd844d834159c21402a86e5a5dfaa2b5fb8e0aae528.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 271, + 229, + 284 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 229, + 284 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 229, + 284 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 294, + 257, + 305 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 294, + 257, + 305 + ], + "spans": [ + { + "bbox": [ + 132, + 294, + 257, + 305 + ], + "type": "text", + "content": "4.1 Database Overview" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 309, + 481, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 309, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 309, + 481, + 453 + ], + "type": "text", + "content": "The key advantage of our approach is the flexibility to utilize multiple image score regression databases freely, even those without annotator tags. In our experiments, we fine-tune models using six different databases in Table 1, reserving the AADB [19] and REAL-CUR [31] database exclusively for testing. While the Flickr-AES [31] and PARA [40] databases do contain annotator tags, our approach does not require distinguishing between individuals. Therefore, we treat these two databases as GIAA databases. In total, we effectively leverage 415,043 samples from six databases, marking the most extensive utilization of samples in PIAA within our knowledge. Detailed descriptions and performance on these databases are provided in the appendix (See Section A.5). However, our approach is not limited to training solely on these six databases, having the scalability to include additional IQA and GIAA databases [4,8,10,16,42]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 147, + 453, + 465, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 453, + 465, + 465 + ], + "spans": [ + { + "bbox": [ + 147, + 453, + 465, + 465 + ], + "type": "text", + "content": "For testing, we use the following three most widely-used PIAA test sets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 465, + 481, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 465, + 481, + 502 + ], + "spans": [ + { + "bbox": [ + 130, + 465, + 481, + 502 + ], + "type": "text", + "content": "Flickr-AES [31] database consists of 40,000 images collected from 210 users. For testing purposes, we utilized a test set containing 4,737 images rated by 37 users, with each user contributing image collections ranging from 105 to 171." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 502, + 481, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 502, + 481, + 537 + ], + "spans": [ + { + "bbox": [ + 130, + 502, + 481, + 537 + ], + "type": "text", + "content": "AADB [19] database consists of 9,958 images from 190 users. We selected a test set from 22 users, each of which contributed individual image collections containing 110 to 190 images." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 538, + 481, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 538, + 481, + 562 + ], + "spans": [ + { + "bbox": [ + 130, + 538, + 481, + 562 + ], + "type": "text", + "content": "REAL-CUR [31] database consists of personal image collections from 14 users, totaling 2,871 images. Each user's collection ranges from 197 to 222 images." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 572, + 279, + 584 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 572, + 279, + 584 + ], + "spans": [ + { + "bbox": [ + 132, + 572, + 279, + 584 + ], + "type": "text", + "content": "4.2 Implementation Details" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": "To obtain task vectors from the models fine-tuned on each of the six databases, it is essential to start from an identical pre-trained model. Recent approaches in IQA [18], GIAA [38], and PIAA [13] emphasize the importance of pre-training on large databases for training powerful image score regression models. Thus, we choose the publicly available OpenCLIP models ViT-B/16 and ViT-L/14 [15] as the pre-trained model. Full details regarding the architectural design and" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 231, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 231, + 91, + 447, + 102 + ], + "type": "text", + "content": "Scaling Up Personalized Image Aesthetic Assessment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 167, + 138, + 445, + 249 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "type": "text", + "content": "Table 2: Cross-database evaluation on the REAL-CUR [31] database. For the baseline approaches, we report the best cross-domain performances." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 167, + 138, + 445, + 249 + ], + "lines": [ + { + "bbox": [ + 167, + 138, + 445, + 249 + ], + "spans": [ + { + "bbox": [ + 167, + 138, + 445, + 249 + ], + "type": "table", + "html": "
MethodSROCC
10-shot100-shot
PA-IAA [20]0.443±0.0040.562±0.013
BLG-PIAA [44]0.448±0.0070.578±0.015
PIAA-SOA [45]0.487±0.0030.589±0.015
TAPP-PIAA [21]-0.580
MTCL [41]0.495±0.0070.599±0.012
Ours0.577±0.0050.621±0.007
", + "image_path": "1e4528043c1bb1590afd5b72a1635316b30e575890abbc618bf6d0424f322697.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 261, + 479, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 261, + 479, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 261, + 479, + 285 + ], + "type": "text", + "content": "training hyperparameters are provided in the appendix (See Section A.5 and A.6). We will also release the codes upon acceptance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 285, + 480, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 480, + 404 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 480, + 404 + ], + "type": "text", + "content": "If not stated otherwise, we use " + }, + { + "bbox": [ + 130, + 285, + 480, + 404 + ], + "type": "inline_equation", + "content": "n = 6" + }, + { + "bbox": [ + 130, + 285, + 480, + 404 + ], + "type": "text", + "content": " fine-tuned models to derive the task vectors. Each model comprises " + }, + { + "bbox": [ + 130, + 285, + 480, + 404 + ], + "type": "inline_equation", + "content": "L = 296" + }, + { + "bbox": [ + 130, + 285, + 480, + 404 + ], + "type": "text", + "content": " layers when using the ViT-L/14 model and " + }, + { + "bbox": [ + 130, + 285, + 480, + 404 + ], + "type": "inline_equation", + "content": "L = 152" + }, + { + "bbox": [ + 130, + 285, + 480, + 404 + ], + "type": "text", + "content": " layers when using the ViT-B/16 model. While the number of 'layers' usually refers to the number of transformer blocks, we include all computational elements within these blocks, such as linear layers and normalization layers. The total number of learnable parameters, " + }, + { + "bbox": [ + 130, + 285, + 480, + 404 + ], + "type": "inline_equation", + "content": "n \\times L = 1776" + }, + { + "bbox": [ + 130, + 285, + 480, + 404 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 285, + 480, + 404 + ], + "type": "inline_equation", + "content": "n \\times L = 912" + }, + { + "bbox": [ + 130, + 285, + 480, + 404 + ], + "type": "text", + "content": ", is significantly lower compared to a full fine-tuning approach, which would require training over 23 million parameters even for small ResNet-50 models [9]. We do not consider the layer-wise task vector itself as a learnable parameter, as obtaining these vectors is not repeated for each individual." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 411, + 303, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 411, + 303, + 422 + ], + "spans": [ + { + "bbox": [ + 131, + 411, + 303, + 422 + ], + "type": "text", + "content": "4.3 Evaluation Metric for PIAA" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 426, + 479, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 479, + 474 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 479, + 474 + ], + "type": "text", + "content": "PIAA approaches are often evaluated through the Spearman's rank-order correlation coefficient [28] (SROCC), which is the correlation coefficient between the predicted and ground-truth rank variables. Specifically, given the ground-truth and predicted scores " + }, + { + "bbox": [ + 130, + 426, + 479, + 474 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 426, + 479, + 474 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 426, + 479, + 474 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 130, + 426, + 479, + 474 + ], + "type": "text", + "content": ", the SROCC is calculated as," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 233, + 479, + 481, + 508 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 479, + 481, + 508 + ], + "spans": [ + { + "bbox": [ + 233, + 479, + 481, + 508 + ], + "type": "interline_equation", + "content": "\\mathrm {S R O C C} = 1 - \\frac {6 \\sum_ {i = 1} ^ {N} \\left(r _ {i} - \\hat {r} _ {i}\\right) ^ {2}}{N \\left(N ^ {2} - 1\\right)}, \\tag {5}", + "image_path": "348b62412bce0db6ff5c072b1402057c9675723c29930ddc61fd6953cb3a840a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 513, + 479, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 513, + 479, + 537 + ], + "spans": [ + { + "bbox": [ + 131, + 513, + 479, + 537 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 513, + 479, + 537 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 131, + 513, + 479, + 537 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 513, + 479, + 537 + ], + "type": "inline_equation", + "content": "\\hat{r}_i" + }, + { + "bbox": [ + 131, + 513, + 479, + 537 + ], + "type": "text", + "content": " represent the rank of the " + }, + { + "bbox": [ + 131, + 513, + 479, + 537 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 513, + 479, + 537 + ], + "type": "text", + "content": "-th sample within the ground-truth and predicted scores, respectively, and " + }, + { + "bbox": [ + 131, + 513, + 479, + 537 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 131, + 513, + 479, + 537 + ], + "type": "text", + "content": " denotes the number of samples." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 552, + 293, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 552, + 293, + 563 + ], + "spans": [ + { + "bbox": [ + 131, + 552, + 293, + 563 + ], + "type": "text", + "content": "4.4 Cross-database Evaluation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "PIAA evaluation can be categorized based on whether training and testing occur within the same database (i.e., intra-database evaluation) or involve different databases (i.e., cross-database evaluation). In most real-world scenarios involving personal image collections, user-provided samples often diverge from the statistics or tendencies of the training database. Therefore, cross-database evaluation becomes pivotal in assessing the generalization capabilities of personalization approaches, which is an aspect that remains relatively underdeveloped in comparison to intra-database evaluations." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 254, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 254, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 254, + 101 + ], + "type": "text", + "content": "JY. Yun and J. Choo" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 167, + 138, + 445, + 276 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "text", + "content": "Table 3: Cross-database evaluation on the AADB [19] database. Results marked with * indicate intra-database performance (i.e., directly trained on the AADB database)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 167, + 138, + 445, + 276 + ], + "lines": [ + { + "bbox": [ + 167, + 138, + 445, + 276 + ], + "spans": [ + { + "bbox": [ + 167, + 138, + 445, + 276 + ], + "type": "table", + "html": "
MethodSROCC
10-shot100-shot
PA-IAA [20]0.469±0.0020.524±0.006
BLG-PIAA [44]0.486±0.0040.536±0.006
PIAA-SOA [45]0.509±0.0030.557±0.007
TAPP-PIAA [21]-0.540
MTCL [41]0.533±0.0040.572±0.007
*TAPP-PIAA [21]*0.534±0.004*0.612±0.007
*MTCL [41]*0.540±0.005*0.622±0.007
Ours0.556±0.0040.654±0.007
", + "image_path": "b33b2d69c6d376f26b8be734fa991affb263ffd6b49917cbcd0afa98b4e51e06.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 289, + 481, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 481, + 396 + ], + "type": "text", + "content": "For cross-database evaluation, we employ the AADB [19] and REAL-CUR [31] databases as the test databases, neither of which were among the six databases from which we derived our task vectors. Following established protocols [31], we randomly select " + }, + { + "bbox": [ + 130, + 289, + 481, + 396 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 289, + 481, + 396 + ], + "type": "text", + "content": " samples for each individual to form the training set, also referred to as the support set, while the remaining images form the test set. In our experiments, we set " + }, + { + "bbox": [ + 130, + 289, + 481, + 396 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 289, + 481, + 396 + ], + "type": "text", + "content": " as either 10-shot or 100-shot. Given the potential variability inherent in few-shot learning depending on the chosen " + }, + { + "bbox": [ + 130, + 289, + 481, + 396 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 289, + 481, + 396 + ], + "type": "text", + "content": " support set, we conduct 10 independent trials for each user and report the average SROCC with the standard deviation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 396, + 482, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 396, + 482, + 504 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 482, + 504 + ], + "type": "text", + "content": "As illustrated in Table 2 and Table 3, our method significantly surpasses existing PIAA models in cross-database performance for both the 10-shot and 100-shot protocols. Remarkably, it even exceeds the performance of models trained explicitly on the AADB database, as compared in Table 3. Such results underscore the exceptional generalization capabilities of our approach, in contrast to previous methods that are constrained by their inability to leverage GIAA and IQA databases. This superior performance is primarily due to the scalability of our approach, which effectively harnesses an extensive amount of 415,043 samples across multiple databases, each contributing unique insights and tendencies." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 515, + 291, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 515, + 291, + 526 + ], + "spans": [ + { + "bbox": [ + 132, + 515, + 291, + 526 + ], + "type": "text", + "content": "4.5 Intra-database Evaluation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 481, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 481, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 481, + 593 + ], + "type": "text", + "content": "For intra-database evaluation, we test our approach for the 37 individuals in the Flickr-AES [31] database, which is the most widely used PIAA database. Since the Flickr-AES [31] database is one of the six databases used to derive the task vectors, this experiment falls under the category of intra-database evaluation. However, we do not use the annotator information in the Flickr-AES." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 481, + 666 + ], + "type": "text", + "content": "In Table 4, we report the average SROCC for each approach along with the standard deviation across 10 independent trials. These experiments underscore the effectiveness of training personalized task vectors through the use of multiple fine-tuned models. Our approach consistently outperforms existing methods that rely on meta-learning techniques or incorporate additional user attributes (e.g., personality traits) to enhance personalized score predictions." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 231, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 231, + 91, + 447, + 102 + ], + "type": "text", + "content": "Scaling Up Personalized Image Aesthetic Assessment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 159, + 138, + 452, + 348 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "text", + "content": "Table 4: 10-shot and 100-shot personalization results on the Flickr-AES database [31], with the average SROCC of 37 users and standard deviation across 10 trials." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 159, + 138, + 452, + 348 + ], + "lines": [ + { + "bbox": [ + 159, + 138, + 452, + 348 + ], + "spans": [ + { + "bbox": [ + 159, + 138, + 452, + 348 + ], + "type": "table", + "html": "
MethodSROCC
10-shot100-shot
PAM (attribute only) [31]0.511±0.0040.516±0.003
PAM (content only) [31]0.512±0.0020.516±0.010
PAM [31]0.513±0.0030.524±0.007
PASS [35]0.516±0.0030.521±0.007
USAR-PPR [25]0.521±0.0020.544±0.007
USAR-PAD [25]0.520±0.0030.537±0.003
USAR-PPR&PAD [25]0.525±0.0040.552±0.015
PA-IAA [20]0.543±0.0030.639±0.011
BLG-PIAA [44]0.561±0.0050.669±0.013
UG-PIAA [24]0.559±0.0020.660±0.013
PIAA-SOA [45]0.618±0.0060.691±0.015
TAPP-PIAA [21]0.591±0.0070.685±0.012
IM-PIAA [13]0.6200.708
MTCL [21]0.667±0.0050.737±0.014
Ours0.668±0.0040.748±0.012
", + "image_path": "1cae3247ac5d6fae7c9bb31a0d6a30ff932a222f7a8621b3646a735427309c2f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 361, + 332, + 374 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 361, + 332, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 361, + 332, + 374 + ], + "type": "text", + "content": "5 Analysis and Ablation Studies" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 383, + 479, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 383, + 479, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 383, + 479, + 407 + ], + "type": "text", + "content": "We perform an in-depth analysis and ablation studies on each component of our approach, including model size, scalability, and coefficient initialization." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 422, + 335, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 422, + 335, + 433 + ], + "spans": [ + { + "bbox": [ + 132, + 422, + 335, + 433 + ], + "type": "text", + "content": "5.1 Task Similarities Across Databases" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 444, + 481, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 444, + 481, + 552 + ], + "spans": [ + { + "bbox": [ + 132, + 444, + 481, + 552 + ], + "type": "text", + "content": "To gain insights into the derived task vectors, we visualize the similarity matrix of the task vectors in Figure 4. For simplicity, we aggregate the " + }, + { + "bbox": [ + 132, + 444, + 481, + 552 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 132, + 444, + 481, + 552 + ], + "type": "text", + "content": " task vectors derived from each layer of a single fine-tuned model as a single task vector by concatenating them into one vector. In line with the findings provided by Ilharco et al. [14], task vectors are typically orthogonal, even though the databases are collected with similar intentions (i.e., GIAA and IQA). This observation highlights that treating each database as a unique task remains a valid choice, owing to variations in their data collection methodologies and the diverse nature of the populations from which they are sourced." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 571, + 343, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 571, + 343, + 582 + ], + "spans": [ + { + "bbox": [ + 132, + 571, + 343, + 582 + ], + "type": "text", + "content": "5.2 Architecture Choices and Variations" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 594, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 594, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 594, + 481, + 666 + ], + "type": "text", + "content": "Selecting the architecture. Weight mixing techniques, including task vectors [14, 36, 37], have been recognized for their effectiveness when fine-tuning models pre-trained on large databases, such as the LAION-5B [33] or DataComp [7]. Vision Transformers (ViT) [5] is a suitable architecture for accommodating web-scale databases, in contrast to ResNets [9], which often struggle to achieve high performance despite training on large databases. Therefore, given" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 254, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 254, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 254, + 101 + ], + "type": "text", + "content": "JY. Yun and J. Choo" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 179, + 119, + 433, + 293 + ], + "blocks": [ + { + "bbox": [ + 179, + 119, + 433, + 293 + ], + "lines": [ + { + "bbox": [ + 179, + 119, + 433, + 293 + ], + "spans": [ + { + "bbox": [ + 179, + 119, + 433, + 293 + ], + "type": "image", + "image_path": "257fc4a21f7d034c3116344456b35f40df8843c93106eee83c6151a17a9c5f3b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 296, + 480, + 317 + ], + "lines": [ + { + "bbox": [ + 131, + 296, + 480, + 317 + ], + "spans": [ + { + "bbox": [ + 131, + 296, + 480, + 317 + ], + "type": "text", + "content": "Fig. 4: Similarity matrix between task vectors derived from each database. The similarities are measured with the cosine similarity." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 153, + 346, + 459, + 446 + ], + "blocks": [ + { + "bbox": [ + 131, + 321, + 480, + 343 + ], + "lines": [ + { + "bbox": [ + 131, + 321, + 480, + 343 + ], + "spans": [ + { + "bbox": [ + 131, + 321, + 480, + 343 + ], + "type": "text", + "content": "Table 5: Performance and inference speed of the ViT-B/16 model and the Best-fit FT method (i.e., directly fine-tuning ViTs) on the REAL-CUR [31] database." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 153, + 346, + 459, + 446 + ], + "lines": [ + { + "bbox": [ + 153, + 346, + 459, + 446 + ], + "spans": [ + { + "bbox": [ + 153, + 346, + 459, + 446 + ], + "type": "table", + "html": "
MethodSROCCFPS
10-shot100-shot
Best-fit FT0.303±0.0080.394±0.00894.41
PIAA-SOA [45]0.487±0.0030.589±0.015-
MTCL [41]0.495±0.0070.599±0.012359.05
Ours (ViT-B/16)0.562±0.0040.607±0.007417.95
Ours0.577±0.0050.621±0.00794.41
", + "image_path": "1c8f0d69225da917bc66f44ce89bd8c6ba521b6a9eccc733ed4cde65566f9e00.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 456, + 480, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 456, + 480, + 504 + ], + "spans": [ + { + "bbox": [ + 130, + 456, + 480, + 504 + ], + "type": "text", + "content": "its scalability and performance advantages, we choose the publicly available pretrained ViT as our architecture for our experiments. Regardless of the chosen architecture, our parameter-efficient approach only trains a minimal set of coefficients corresponding to each layer-wise task vector." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 507, + 482, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 507, + 482, + 638 + ], + "spans": [ + { + "bbox": [ + 130, + 507, + 482, + 638 + ], + "type": "text", + "content": "Scaling down for efficient computation. While our primary experiments focus on combining task vectors derived from ViT-L/14 models, our approach is adaptable to smaller models as well. As shown in Table 5, our approach yields strong results even on smaller models like ViT-B/16, which are well-suited for deployment on commercial computers. Although fast inference was not our primary goal, we achieve real-time inference speeds (" + }, + { + "bbox": [ + 130, + 507, + 482, + 638 + ], + "type": "inline_equation", + "content": "\\geq 30" + }, + { + "bbox": [ + 130, + 507, + 482, + 638 + ], + "type": "text", + "content": " FPS) on the RTX 3090, as our weight mixing technique combines multiple models into a single ViT model, eliminating the need to pass through multiple models. Acceleration of ViTs [22, 23, 43] beyond our implementation is also actively supported in recent frameworks [26] and communities [30], which ensures the practicality of our approach across a wide range of computing platforms such as mobile devices." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": "Vision Transformers without task vectors. Our experimental results in Table 5 demonstrate that the performance gains are not solely attributed to" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 231, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 231, + 91, + 447, + 102 + ], + "type": "text", + "content": "Scaling Up Personalized Image Aesthetic Assessment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 196, + 118, + 415, + 267 + ], + "blocks": [ + { + "bbox": [ + 196, + 118, + 415, + 267 + ], + "lines": [ + { + "bbox": [ + 196, + 118, + 415, + 267 + ], + "spans": [ + { + "bbox": [ + 196, + 118, + 415, + 267 + ], + "type": "image", + "image_path": "fb42a99854c3c82dcd8107a702ae1cfd7a13d78c182c356f371608997fa0abd9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 269, + 480, + 291 + ], + "lines": [ + { + "bbox": [ + 132, + 269, + 480, + 291 + ], + "spans": [ + { + "bbox": [ + 132, + 269, + 480, + 291 + ], + "type": "text", + "content": "Fig. 5: Performance gains with respect to the number of databases used. The dotted lines indicate the previous state-of-the-art performance." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 301, + 482, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 301, + 482, + 373 + ], + "spans": [ + { + "bbox": [ + 131, + 301, + 482, + 373 + ], + "type": "text", + "content": "the inherent strengths of the ViT architecture. When directly fine-tuning the aesthetic assessment model that best matches the user's preference (i.e., Best-fit FT), we observe distinctly low personalization performances despite using a large pre-trained model. This highlights the pivotal role of keeping the layer-wise task vectors frozen during personalization, as altering the vectors would indicate discarding the rich preferences learned from a large database." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 386, + 302, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 302, + 399 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 302, + 399 + ], + "type": "text", + "content": "5.3 Scalability of Our Approach" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 410, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 410, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 131, + 410, + 482, + 529 + ], + "type": "text", + "content": "How does each database contribute? We assess the scalability of our approach by varying the number of task vectors. Specifically, we evaluate our approach on the REAL-CUR database [31] by increasing the number of databases used. Starting from the GIAA model trained on the AVA [27] database, we accumulate the databases in the following order: Flicker-AES [31], TAD66K [11], PARA [40], KonIQ-10K [12], and SPAQ [6]. As demonstrated in Figure 5, both 10-shot and 100-shot performance significantly improves as the number of task vectors increases. This indicates that our scalable approach benefits from the expansion of task vectors, which allows us to capture a broader range of image characteristics and nuances." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 533, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 533, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 131, + 533, + 482, + 616 + ], + "type": "text", + "content": "How can we increase the number of task vectors? The advantage of our approach can be amplified with the availability of additional image score regression databases. For example, scaling up the number of task vectors can be further pursued by incorporating readily available IQA [4,8,42] and GIAA [10,16] databases not included in our current study. Additionally, partitioning large databases, such as AVA [27] and PaQ-2-PiQ [42], can be an alternative method for generating additional task vectors." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 617, + 482, + 666 + ], + "type": "text", + "content": "These strategies for obtaining a diverse set of task vectors may lead to performance gains in the 10-shot evaluation, which is considered highly challenging due to the limited number of user-provided samples. However, we leave this as future work, given that our current work already utilizes the largest number of" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 254, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 254, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 254, + 100 + ], + "type": "text", + "content": "JY. Yun and J. Choo" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 213, + 118, + 404, + 277 + ], + "blocks": [ + { + "bbox": [ + 213, + 118, + 404, + 277 + ], + "lines": [ + { + "bbox": [ + 213, + 118, + 404, + 277 + ], + "spans": [ + { + "bbox": [ + 213, + 118, + 404, + 277 + ], + "type": "image", + "image_path": "2cccc0294fe52d3ac029fda7cdad6ed747a7a01e8bde055c3b02e8f530a84405.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 285, + 479, + 308 + ], + "lines": [ + { + "bbox": [ + 131, + 285, + 479, + 308 + ], + "spans": [ + { + "bbox": [ + 131, + 285, + 479, + 308 + ], + "type": "text", + "content": "Fig. 6: 2D-histogram of coefficients and zero-shot SROCC pairs. The size of the coefficients are correlated with the relevance of the task." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 315, + 479, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 315, + 479, + 339 + ], + "spans": [ + { + "bbox": [ + 131, + 315, + 479, + 339 + ], + "type": "text", + "content": "samples for PIAA, 10 times larger than the most widely used and highly curated Flickr-AES database [31]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 352, + 384, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 352, + 384, + 363 + ], + "spans": [ + { + "bbox": [ + 131, + 352, + 384, + 363 + ], + "type": "text", + "content": "5.4 Task Relevance and Coefficient Initialization" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 373, + 482, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 373, + 482, + 516 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 482, + 516 + ], + "type": "text", + "content": "We conduct an analysis to better understand the distinguishing characteristics of coefficients, specifically, what sets apart those with high values from those with low values. For a straightforward examination, we calculate the average coefficient for each task (i.e., database), averaged across all the layers of the model, which results in six average coefficients for each user (" + }, + { + "bbox": [ + 130, + 373, + 482, + 516 + ], + "type": "inline_equation", + "content": "n = 6" + }, + { + "bbox": [ + 130, + 373, + 482, + 516 + ], + "type": "text", + "content": "). Subsequently, we assess the six zero-shot SROCC scores associated with the average coefficients and plot the histogram of the coefficient-SROCC pairs from all test users in Figure 6. We observe that coefficients for tasks with a high zero-shot SROCC are trained to have high values, while those for less-relevant tasks (i.e., low zero-shot SROCC) exhibit low values. This analysis justifies our adaptive initialization of the coefficients, which assigns large values to coefficients associated with closely related tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 517, + 482, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 517, + 482, + 553 + ], + "spans": [ + { + "bbox": [ + 131, + 517, + 482, + 553 + ], + "type": "text", + "content": "We investigate two alternative initialization strategies: a uniform initialization and a best-fit initialization. Both cases can be viewed as an extreme case of the adaptive initialization, which can be written in a generalized form:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 199, + 561, + 481, + 598 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 199, + 561, + 481, + 598 + ], + "spans": [ + { + "bbox": [ + 199, + 561, + 481, + 598 + ], + "type": "interline_equation", + "content": "\\alpha_ {i} ^ {l} = \\frac {e ^ {\\mathrm {S R O C C} (S , \\theta_ {\\mathrm {f t} _ {i}}) / T}}{\\sum_ {j = 1} ^ {n} e ^ {\\mathrm {S R O C C} (S , \\theta_ {\\mathrm {f t} _ {j}}) / T}} \\quad \\text {f o r} l = 1, 2, \\dots , L, \\tag {6}", + "image_path": "c8999080aeb7203b076406509a233c8cc97ca282fec4e165abee7c124bf86b56.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 605, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 131, + 605, + 482, + 666 + ], + "type": "text", + "content": " is the temperature for the softmax. The uniform initialization corresponds to the case where the " + }, + { + "bbox": [ + 131, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "T \\to \\infty" + }, + { + "bbox": [ + 131, + 605, + 482, + 666 + ], + "type": "text", + "content": ", while the best-fit initialization is when " + }, + { + "bbox": [ + 131, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "T \\to 0" + }, + { + "bbox": [ + 131, + 605, + 482, + 666 + ], + "type": "text", + "content": ". As demonstrated in Table 6, the adaptive initialization outperforms other initialization strategies. Nevertheless, we consider further exploration of the initialization strategies as a potential avenue for future research." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 231, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 231, + 91, + 447, + 102 + ], + "type": "text", + "content": "Scaling Up Personalized Image Aesthetic Assessment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 166, + 138, + 444, + 212 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "text", + "content": "Table 6: Comparison between three different coefficient initialization strategies tested on the REAL-CUR [31] database." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 166, + 138, + 444, + 212 + ], + "lines": [ + { + "bbox": [ + 166, + 138, + 444, + 212 + ], + "spans": [ + { + "bbox": [ + 166, + 138, + 444, + 212 + ], + "type": "table", + "html": "
MethodSROCC
10-shot100-shot
Uniform init. (T→∞)0.520±0.0020.602±0.002
Best-fit init. (T→0)0.464±0.0120.596±0.002
Adaptive init. (T=1)0.577±0.0050.621±0.007
", + "image_path": "644c0b258854b00eaf86f8c79446b2f86c85a038fb064de52623e7b00b740d67.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 219, + 217, + 231 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 219, + 217, + 231 + ], + "spans": [ + { + "bbox": [ + 132, + 219, + 217, + 231 + ], + "type": "text", + "content": "6 Limitation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 236, + 479, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 479, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 479, + 307 + ], + "type": "text", + "content": "We have observed that when the zero-shot SROCC for all six image regression models are low for an individual (i.e., " + }, + { + "bbox": [ + 130, + 236, + 479, + 307 + ], + "type": "inline_equation", + "content": "\\mathrm{SROCC} \\approx 0" + }, + { + "bbox": [ + 130, + 236, + 479, + 307 + ], + "type": "text", + "content": "), our approach may struggle to identify an optimal combination of task vectors for that user. This phenomenon suggests that when there is a limited correlation between the user's preferences and the predictions of these models, the task vectors derived from them may be insufficient for finding an appropriate personalized task vector." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 308, + 480, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 480, + 380 + ], + "type": "text", + "content": "However, we find this case to be extremely rare (See appendix Figure 2-7.), with only 1 among 73 cases having a low SROCC under 0.2, demonstrating the broad coverage of our approach. Also, increasing the number of task vectors will effectively mitigate this issue, leveraging the inherent scalability of our approach, which can accommodate a growing number of task vectors to improve personalization even in challenging scenarios." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 390, + 220, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 390, + 220, + 403 + ], + "spans": [ + { + "bbox": [ + 132, + 390, + 220, + 403 + ], + "type": "text", + "content": "7 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 407, + 479, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 479, + 466 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 479, + 466 + ], + "type": "text", + "content": "We address the scalability issue in PIAA, a critical yet previously unaddressed aspect necessary for enhancing the generalization performance. Unlike previous studies limited to training on PIAA databases curated by individual annotators, our approach enables the flexible use of multiple image score regression databases, thereby providing greater scalability and generalization capabilities." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 467, + 480, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 467, + 480, + 574 + ], + "spans": [ + { + "bbox": [ + 130, + 467, + 480, + 574 + ], + "type": "text", + "content": "We demonstrate that GIAA and IQA databases exhibit distinct personalization potentials, which have motivated us to combine GIAA and IQA models to achieve desired behaviors. By introducing learnable parameters for optimal model combinations, we enable the precise personalization of models to individual users' aesthetic preferences. Our approach significantly outperforms existing approaches, demonstrating exceptional efficacy in generalizing to unseen domains which is a critical requirement for real-world applications. This work opens up new avenues for personalized image aesthetic assessment, offering valuable insights and practical solutions for this challenging task." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 581, + 218, + 592 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 581, + 218, + 592 + ], + "spans": [ + { + "bbox": [ + 132, + 581, + 218, + 592 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 595, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 595, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 595, + 481, + 665 + ], + "type": "text", + "content": "This work was supported by the Institute for Information & communications Technology Promotion(IITP) grant funded by the Korea government(MSIT) (No.RS-2019-II190075 Artificial Intelligence Graduate School Program(KAIST)), the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. NRF-2022R1A2B5B02001913), and the Institute of Information & communications Technology Promotion (IITP) grant funded by the Korea government (MSIT) (RS-2021-II212068, Artificial Intelligence Innovation Hub)." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 254, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 254, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 254, + 100 + ], + "type": "text", + "content": "JY. Yun and J. Choo" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 141, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 138, + 141, + 480, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 141, + 480, + 163 + ], + "spans": [ + { + "bbox": [ + 138, + 141, + 480, + 163 + ], + "type": "text", + "content": "1. Al-Shedivat, M., Li, L., Xing, E., Talwalkar, A.: On data efficiency of meta-learning. In: International Conference on Artificial Intelligence and Statistics. PMLR (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 165, + 480, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 165, + 480, + 186 + ], + "spans": [ + { + "bbox": [ + 138, + 165, + 480, + 186 + ], + "type": "text", + "content": "2. Bradley, R.A., Terry, M.E.: Rank analysis of incomplete block designs: I. the method of paired comparisons. Biometrika (1952)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 186, + 480, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 186, + 480, + 208 + ], + "spans": [ + { + "bbox": [ + 138, + 186, + 480, + 208 + ], + "type": "text", + "content": "3. Chefer, H., Gur, S., Wolf, L.: Transformer interpretability beyond attention visualization. In: CVPR (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 209, + 480, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 209, + 480, + 241 + ], + "spans": [ + { + "bbox": [ + 138, + 209, + 480, + 241 + ], + "type": "text", + "content": "4. Ciancio, A., da Silva, E.A., Said, A., Samadani, R., Obrador, P., et al.: No-reference blur assessment of digital pictures based on multifeature classifiers. IEEE Transactions on Image Processing (2010)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 243, + 480, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 243, + 480, + 275 + ], + "spans": [ + { + "bbox": [ + 138, + 243, + 480, + 275 + ], + "type": "text", + "content": "5. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 276, + 481, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 276, + 481, + 297 + ], + "spans": [ + { + "bbox": [ + 138, + 276, + 481, + 297 + ], + "type": "text", + "content": "6. Fang, Y., Zhu, H., Zeng, Y., Ma, K., Wang, Z.: Perceptual quality assessment of smartphone photography. In: CVPR (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 298, + 480, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 298, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 138, + 298, + 480, + 331 + ], + "type": "text", + "content": "7. Gadre, S.Y., Ilharco, G., Fang, A., Hayase, J., Smyrnis, G., Nguyen, T., Marten, R., Wortsman, M., Ghosh, D., Zhang, J., et al.: Datacomp: In search of the next generation of multimodal datasets. arXiv preprint arXiv:2304.14108 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 331, + 480, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 331, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 138, + 331, + 480, + 354 + ], + "type": "text", + "content": "8. Ghadiyaram, D., Bovik, A.C.: Massive online crowdsourced study of subjective and objective picture quality. IEEE Transactions on Image Processing (2015)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 354, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 354, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 138, + 354, + 480, + 374 + ], + "type": "text", + "content": "9. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 134, + 376, + 480, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 376, + 480, + 397 + ], + "spans": [ + { + "bbox": [ + 134, + 376, + 480, + 397 + ], + "type": "text", + "content": "10. He, S., Ming, A., Li, Y., Sun, J., Zheng, S., Ma, H.: Thinking image color aesthetics assessment: Models, datasets and benchmarks. In: ICCV (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 399, + 480, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 399, + 480, + 419 + ], + "spans": [ + { + "bbox": [ + 133, + 399, + 480, + 419 + ], + "type": "text", + "content": "11. He, S., Zhang, Y., Xie, R., Jiang, D., Ming, A.: Rethinking image aesthetics assessment: Models, datasets and benchmarks. IJCAI (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 421, + 480, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 421, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 133, + 421, + 480, + 453 + ], + "type": "text", + "content": "12. Hosu, V., Lin, H., Sziranyi, T., Saupe, D.: Koniq-10k: An ecologically valid database for deep learning of blind image quality assessment. IEEE Transactions on Image Processing (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 454, + 480, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 454, + 480, + 475 + ], + "spans": [ + { + "bbox": [ + 133, + 454, + 480, + 475 + ], + "type": "text", + "content": "13. Hou, J., Lin, W., Yue, G., Liu, W., Zhao, B.: Interaction-matrix based personalized image aesthetics assessment. IEEE Transactions on Multimedia (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 476, + 480, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 476, + 480, + 498 + ], + "spans": [ + { + "bbox": [ + 133, + 476, + 480, + 498 + ], + "type": "text", + "content": "14. Ilharco, G., Ribeiro, M.T., Wortsman, M., Schmidt, L., Hajishirzi, H., Farhadi, A.: Editing models with task arithmetic. In: ICLR (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 498, + 480, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 498, + 480, + 542 + ], + "spans": [ + { + "bbox": [ + 133, + 498, + 480, + 542 + ], + "type": "text", + "content": "15. Ilharco, G., Wortsman, M., Wightman, R., Gordon, C., Carlini, N., Taori, R., Dave, A., Shankar, V., Namkoong, H., Miller, J., Hajishirzi, H., Farhadi, A., Schmidt, L.: Openclip (Jul 2021). https://doi.org/10.5281/zenodo.5143773, https://doi.org/10.5281/zenodo.5143773" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 543, + 480, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 543, + 480, + 564 + ], + "spans": [ + { + "bbox": [ + 133, + 543, + 480, + 564 + ], + "type": "text", + "content": "16. Jin, X., Wu, L., Zhao, G., Li, X., Zhang, X., Ge, S., Zou, D., Zhou, B., Zhou, X.: Aesthetic attributes assessment of images. In: ACM MM (2019)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 133, + 565, + 480, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 565, + 480, + 586 + ], + "spans": [ + { + "bbox": [ + 133, + 565, + 480, + 586 + ], + "type": "text", + "content": "17. Karlsson, K., Jiang, W., Zhang, D.Q.: Mobile photo album management with multiscale timeline. In: ACM MM (2014)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 133, + 587, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 587, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 133, + 587, + 480, + 609 + ], + "type": "text", + "content": "18. Ke, J., Wang, Q., Wang, Y., Milanfar, P., Yang, F.: Musiq: Multi-scale image quality transformer. In: ICCV (2021)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 133, + 610, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 610, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 133, + 610, + 480, + 632 + ], + "type": "text", + "content": "19. Kong, S., Shen, X., Lin, Z., Mech, R., Fowlkes, C.: Photo aesthetics ranking network with attributes and content adaptation. In: ECCV. Springer (2016)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 133, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 632, + 480, + 665 + ], + "type": "text", + "content": "20. Li, L., Zhu, H., Zhao, S., Ding, G., Lin, W.: Personality-assisted multi-task learning for generic and personalized image aesthetics assessment. IEEE Transactions on Image Processing 29 (2020)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 231, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 231, + 91, + 447, + 102 + ], + "type": "text", + "content": "Scaling Up Personalized Image Aesthetic Assessment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "type": "text", + "content": "21. Li, Y., Yang, Y., Li, H., Chen, H., Xu, L., Li, L., Li, Y., Guo, Y.: Transductive aesthetic preference propagation for personalized image aesthetics assessment. In: ACM MM. ACM (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 149, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 481, + 171 + ], + "type": "text", + "content": "22. Lin, Y., Zhang, T., Sun, P., Li, Z., Zhou, S.: Fq-vit: Post-training quantization for fully quantized vision transformer. IJCAI (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 171, + 481, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 171, + 481, + 192 + ], + "spans": [ + { + "bbox": [ + 132, + 171, + 481, + 192 + ], + "type": "text", + "content": "23. Liu, Z., Wang, Y., Han, K., Zhang, W., Ma, S., Gao, W.: Post-training quantization for vision transformer. NeurIPS (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 192, + 481, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 192, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 192, + 481, + 224 + ], + "type": "text", + "content": "24. Lv, P., Fan, J., Nie, X., Dong, W., Jiang, X., Zhou, B., Xu, M., Xu, C.: User-guided personalized image aesthetic assessment based on deep reinforcement learning. IEEE Transactions on Multimedia (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 224, + 481, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 481, + 257 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 481, + 257 + ], + "type": "text", + "content": "25. Lv, P., Wang, M., Xu, Y., Peng, Z., Sun, J., Su, S., Zhou, B., Xu, M.: Usar: An interactive user-specific aesthetic ranking framework for images. In: ACM MM. pp. 1328-1336 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 257, + 390, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 257, + 390, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 390, + 267 + ], + "type": "text", + "content": "26. Microsoft: Deepspeed (2023), https://www.deepspeed.ai/" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 267, + 481, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 267, + 481, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 267, + 481, + 289 + ], + "type": "text", + "content": "27. Murray, N., Marchesotti, L., Perronnin, F.: Ava: A large-scale database for aesthetic visual analysis. In: CVPR. IEEE (2012)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 289, + 481, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 481, + 310 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 481, + 310 + ], + "type": "text", + "content": "28. Myers, J.L., Well, A.D., Lorch Jr, R.F.: Research design and statistical analysis. Routledge (2013)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 310, + 481, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 310, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 132, + 310, + 481, + 331 + ], + "type": "text", + "content": "29. Naseer, M.M., Ranasinghe, K., Khan, S.H., Hayat, M., Shahbaz Khan, F., Yang, M.H.: Intriguing properties of vision transformers. NeurIPS (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 332, + 481, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 481, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 481, + 365 + ], + "type": "text", + "content": "30. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. NeurIPS (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 365, + 481, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 481, + 385 + ], + "type": "text", + "content": "31. Ren, J., Shen, X., Lin, Z., Mech, R., Foran, D.J.: Personalized image aesthetics. In: ICCV (Oct 2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 385, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 385, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 385, + 481, + 407 + ], + "type": "text", + "content": "32. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. 2022 IEEE. In: CVPR (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 407, + 481, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 481, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 481, + 439 + ], + "type": "text", + "content": "33. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. NeurIPS (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 439, + 481, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 439, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 439, + 481, + 460 + ], + "type": "text", + "content": "34. Wallace, B., Gokul, A., Ermon, S., Naik, N.: End-to-end diffusion latent optimization improves classifier guidance. ICCV (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 460, + 481, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 460, + 481, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 481, + 482 + ], + "type": "text", + "content": "35. Wang, G., Yan, J., Qin, Z.: Collaborative and attentive learning for personalized image aesthetic assessment. In: IJCAI. pp. 957-963 (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 482, + 481, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 482, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 482, + 481, + 525 + ], + "type": "text", + "content": "36. Wortsman, M., Ilharco, G., Gadre, S.Y., Roelofs, R., Gontijo-Lopes, R., Morcos, A.S., Namkoong, H., Farhadi, A., Carmon, Y., Kornblith, S., et al.: Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In: ICML. PMLR (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 525, + 481, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 481, + 557 + ], + "type": "text", + "content": "37. Wortsman, M., Ilharco, G., Li, M., Kim, J.W., Hajishirzi, H., Farhadi, A., Namkoong, H., Schmidt, L.: Robust fine-tuning of zero-shot models. 2022 IEEE. In: CVPR (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 557, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 481, + 578 + ], + "type": "text", + "content": "38. Xu, L., Xu, J., Yang, Y., Huang, Y., Xie, Y., Li, Y.: Clip brings better features to visual aesthetics learners. arXiv preprint arXiv:2307.15640 (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 578, + 481, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 481, + 611 + ], + "type": "text", + "content": "39. Yan, X., Shao, F., Chen, H., Jiang, Q.: Hybrid cnn-transformer based meta-learning approach for personalized image aesthetics assessment. Journal of Visual Communication and Image Representation 98 (2024)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 611, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 481, + 632 + ], + "type": "text", + "content": "40. Yang, Y., Xu, L., Li, L., Qie, N., Li, Y., Zhang, P., Guo, Y.: Personalized image aesthetics assessment with rich attributes. In: CVPR (2022)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "41. Yang, Z., Li, L., Yang, Y., Li, Y., Lin, W.: Multi-level transitional contrast learning for personalized image aesthetics assessment. IEEE Transactions on Multimedia (2023)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 254, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 254, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 254, + 100 + ], + "type": "text", + "content": "JY. Yun and J. Choo" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 227 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "text", + "content": "42. Ying, Z., Niu, H., Gupta, P., Mahajan, D., Ghadiyaram, D., Bovik, A.: From patches to pictures (paq-2-piq): Mapping the perceptual space of picture quality. In: CVPR (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 482, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 482, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 482, + 172 + ], + "type": "text", + "content": "43. Yuan, Z., Xue, C., Chen, Y., Wu, Q., Sun, G.: Ptq4vit: Post-training quantization for vision transformers with twin uniform quantization. In: ECCV. Springer (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 482, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 482, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 482, + 205 + ], + "type": "text", + "content": "44. Zhu, H., Li, L., Wu, J., Zhao, S., Ding, G., Shi, G.: Personalized image aesthetics assessment via meta-learning with bilevel gradient optimization. IEEE Transactions on Cybernetics (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 482, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 482, + 227 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 482, + 227 + ], + "type": "text", + "content": "45. Zhu, H., Zhou, Y., Li, L., Li, Y., Guo, Y.: Learning personalized image aesthetics from subjective and objective attributes. IEEE Transactions on Multimedia (2021)" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 231, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 231, + 91, + 447, + 102 + ], + "type": "text", + "content": "Scaling Up Personalized Image Aesthetic Assessment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/ScanReason_ Empowering 3D Visual Grounding with Reasoning Capabilities/c0158ad2-ca00-4982-8b2e-6cbbd910c1e7_content_list.json b/2024/ScanReason_ Empowering 3D Visual Grounding with Reasoning Capabilities/c0158ad2-ca00-4982-8b2e-6cbbd910c1e7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..459e85d0af5aa3ed86141c396995b5728d5b2876 --- /dev/null +++ b/2024/ScanReason_ Empowering 3D Visual Grounding with Reasoning Capabilities/c0158ad2-ca00-4982-8b2e-6cbbd910c1e7_content_list.json @@ -0,0 +1,1775 @@ +[ + { + "type": "text", + "text": "ScanReason: Empowering 3D Visual Grounding with Reasoning Capabilities", + "text_level": 1, + "bbox": [ + 228, + 140, + 774, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chenming Zhu $^{1,2}$ , Tai Wang $^{2}$ , Wenwei Zhang $^{2}$ , Kai Chen $^{2}$ , and Xihui Liu $^{1\\dagger}$", + "bbox": [ + 228, + 210, + 771, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 The University of Hong Kong \n2 Shanghai AI Laboratory \nchairzhu@connect.hku.hk \nhttps://zcmax.github.io/projects/ScanReason", + "bbox": [ + 331, + 238, + 669, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Although great progress has been made in 3D visual grounding, current models still rely on explicit textual descriptions for grounding and lack the ability to reason human intentions from implicit instructions. We propose a new task called 3D reasoning grounding and introduce a new benchmark ScanReason which provides over 10K question-answer-location pairs from five reasoning types that require the synchronization of reasoning and grounding. We further design our approach, ReGround3D, composed of the visual-centric reasoning module empowered by Multi-modal Large Language Model (MLLM) and the 3D grounding module to obtain accurate object locations by looking back to the enhanced geometry and fine-grained details from the 3D scenes. A chain-of-grounding mechanism is proposed to further boost the performance with interleaved reasoning and grounding steps during inference. Extensive experiments on the proposed benchmark validate the effectiveness of our proposed approach.", + "bbox": [ + 261, + 330, + 743, + 539 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: 3D reasoning grounding $\\cdot$ 3D visual grounding $\\cdot$ Multi-modal large language models", + "bbox": [ + 259, + 551, + 740, + 580 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 604, + 375, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Understanding and reasoning in the 3D visual world is critical for applications such as robotics and AR, where embodied agents are expected to understand the 3D layout and predict the 3D locations of objects based on human instructions. The example in Fig. 1 demonstrates a scenario where the question can only be solved with a comprehensive understanding of the 3D scene and joint reasoning of the question and the 3D environment. However, current 3D visual grounding models [24,39,43,50] trained on [1,7] localize objects based on explicit descriptions of the object category, attribute, and 3D spatial relationships, and lack the ability to reason the user intentions and predict object locations with implicit human instructions such as \"I am thirsty, can I have something to drink?\"", + "bbox": [ + 212, + 635, + 787, + 785 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To bridge the aforementioned gap and to push the boundaries of what embodied agents can understand and how they can interact with the 3D world, we", + "bbox": [ + 212, + 786, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author: xihuiliu@eee.hku.hk", + "bbox": [ + 217, + 824, + 522, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Question: I'm looking for a comfortable spot to watch TV immersively, where is the best place for me to lie down and watch TV?", + "bbox": [ + 259, + 150, + 735, + 172 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4c561b3f0105a6a0b0268ea384fa4a072cd3d46107e710ee8cfb9fad6d2e5abe.jpg", + "image_caption": [ + "GPT-4" + ], + "image_footnote": [], + "bbox": [ + 218, + 183, + 248, + 202 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The best spot for lying down to watch TV immersively is on a spacious, cushioned sectional sofa or an adjustable reclining bed in a room with a large screen TV mounted at an eye-level view from your lying position, ensuring comfort and optimal viewing angles. Ensure the room is dimly lit or has blackout curtains to enhance screen visibility and reduce glare.", + "bbox": [ + 261, + 186, + 464, + 275 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4a95c7dbd5de217e17e8a7e1b54d70cd020b4df923a278dce65612d360db09f6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 478, + 186, + 547, + 271 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ce635d6fa7f15884bccb2f950a6b0d5e412f2b3aa27caec0eedee6685f12390f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 550, + 186, + 620, + 270 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The best spot to lie down and watch TV immersively would be on the bed facing the TV in the cabinet; it provides a direct line of sight and the comfort of lying down.", + "bbox": [ + 624, + 186, + 733, + 265 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/aa38eb9871b82966b6329b67f248d43cfc47ccb26d6c4f553de5716685477204.jpg", + "image_caption": [ + "GPT-4V" + ], + "image_footnote": [], + "bbox": [ + 746, + 184, + 776, + 204 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0e30da9813c4b3ec62ff18a22440b8ac36f716ee37e9d7ca1bd6d6557adb0cdf.jpg", + "image_caption": [ + "ReGround3D" + ], + "image_footnote": [], + "bbox": [ + 220, + 277, + 248, + 300 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/97b7d2df696c624b87e1e831451244f95ed889ec992544662311448d0163f013.jpg", + "image_caption": [ + "Fig. 1: For an embodied agent, they not only need to be able to understand the 3D environment and complex human instructions but also localize the target objects for interaction and navigation. Although GPT-4 (GPT-4V) have strong text (multi-modal) reasoning abilities, they lack the ability to directly perceive the 3D scene, understand the 3D spatial relationships, and output corresponding target object locations. Instead, our proposed method ReGround3D has the 3D perception, reasoning, and grounding capabilities in the real-world 3D environment." + ], + "image_footnote": [], + "bbox": [ + 263, + 308, + 366, + 375 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For an optimal TV viewing experience, I recommend sitting on the couch. It is oriented towards the TV and closer to the TV compared with the bed, and it's large enough to lie down.", + "bbox": [ + 380, + 318, + 609, + 359 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/53f86f8df84306de5070778a32a6056b0d90a3b96eafe9bba48d951c14a45520.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 625, + 306, + 725, + 376 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "propose a new task of 3D reasoning grounding and introduce a new benchmark named ScanReason. The task requires the model to conduct joint reasoning on the question and the 3D environment before predicting the 3D locations of target objects. We define five categories of 3D reasoning: spatial reasoning, functional reasoning, logical reasoning, emotional reasoning, and safety reasoning. The first two categories focus on the fundamental understanding of the 3D physical world, while the last three categories are built upon fundamental abilities to address user-centric real-world challenges. The benchmark comprises more than 10K question-answer-3D bounding box pairs from 2K scenes belonging to the five reasoning types mentioned above. The GPT-4-assisted data annotation process largely increases the efficiency of curating such a dataset.", + "bbox": [ + 212, + 534, + 787, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose ReGround3D as an initial attempt to the new task of 3D Reasoning Grounding. Intuitively, for 3D grounding with implicit instructions, we need to first conduct reasoning on the language instructions and the coarse visual environment. Then with the idea of which object we want to find in mind, we look back to the 3D scene and ground the target object. For complex instructions, we may need to alternate the reasoning and look-back process for multiple iterations. Inspired by this intuition, our framework is composed of a visual-centric reasoning module and 3D grounding with geometry-enhanced look-back module, with a Chain-of-Grounding mechanism during inference to", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "C. Zhu et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "alternately conduct reasoning and grounding for multiple rounds. Specifically, the visual-centric reasoning module conducts joint reasoning of the 3D scene and instructions with an MLLM. This module predicts a special token representing the semantic and location information of the target object, which is used for the grounding module. The 3D grounding module uses the output token embedding from the previous reasoning module to locate the target object by looking back at the fine-grained 3D scene representation. Unlike previous MLLMs attempting to directly predict bounding box coordinates, our look-back mechanism enables the model to capture more comprehensive 3D geometry and fine-grained object details for accurate 3D grounding. The Chain-of-Grounding mechanism is proposed to synergize reasoning and grounding, which allows multiple rounds alternating between reasoning and grounding during inference.", + "bbox": [ + 212, + 145, + 787, + 327 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, our contributions are threefold: 1) We propose a new task of 3D reasoning grounding which requires the model to synergize reasoning and grounding. We further introduce a new benchmark ScanReason, which comprises five reasoning types (spatial reasoning, functional reasoning, logical reasoning, emotional reasoning, and safety reasoning) for the task of 3D reasoning grounding in 3D scenes. 2) We design a new framework ReGround3D with a visual-centric reasoning module and a 3D grounding module with geometry-enhanced look-back. We further introduce a Chain-of-Grounding mechanism to boost the 3D reasoning grounding ability with a chain of interleaved reasoning and grounding steps. 3) Extensive experiments demonstrate the effectiveness of the our ReGround3D on the ScanReason benchmark for 3D reasoning grounding.", + "bbox": [ + 212, + 329, + 787, + 494 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 515, + 387, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Vision and Language Learning 3D Vision-language learning (3D-VL) is garnering increasing attention, with many 3D-VL tasks focusing on how to connect the 3D world with natural language. Among them, 3D Question Answering (3D QA) [2,45] aims to enable models to provide text answers based on natural language questions. Situation Question Answering in 3D Scenes (SQA3D) [29] requires an agent to first understand its location based on a text description, then provide a reasonable answer based on the surrounding environment, which can be seen as an extension of 3D QA in the embodied AI area. 3D Visual Grounding [1,4,7,28,31,39] demands that models identify and locate target objects in a 3D scene based on given descriptions, outputting the objects' coordinates and 3D bounding boxes. These descriptions usually explicitly rely on the objects' attributes and their spatial relationships. 3D Dense Captioning [5,8,14,16,25,36,46,49] requires models to output a series of object coordinates and corresponding scene-based descriptions based on a given scene. Different from these 3D-VL tasks, the questions in 3D reasoning grounding could be more implicit and complex.", + "bbox": [ + 212, + 546, + 787, + 789 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Visual Grounding The task of 3D visual grounding is aimed at localizing the objects that are explicitly referred to by free-form guided language", + "bbox": [ + 212, + 808, + 785, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "ScanReason", + "bbox": [ + 648, + 114, + 730, + 127 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "expressions in the 3D scene. Inspired by the success of transformers in natural language processing, recent 3D visual grounding approaches [12,17,24,39,50,53] have started to adopt transformer [35] architectures for handling the complex relationships between language descriptions and 3D visual data. These methods leverage the self-attention mechanism of transformers to dynamically weigh the importance of different parts of the input data, facilitating a more effective grounding of textual descriptions in the 3D environment. Recent method [3] proposes a Chain-of-Thoughts module that predicts a chain of anchor objects that are subsequently utilized to localize the final target object. Compared with 3D visual grounding, our proposed 3D reasoning grounding requires the model to reason the complex question, ground target objects, and give the explanation at the same time.", + "bbox": [ + 212, + 146, + 787, + 327 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multi-modal Large Language Models Recently, there has been an increasing effort to extend the powerful complex reasoning and world knowledge capabilities of LLMs [18,34,47] to other modalities [6,9,10,19,26,44,51,52]. Among these works, some have aimed to enable LLMs to understand the 3D world. [20, 40] focus on delving into LLMs' ability to comprehend 3D objects, which can not be directly applied to 3D scenes. 3D-LLM [21] is the pioneering work that incorporates the 3D scene into LLM to carry out general 3D understanding tasks. However, by using 3D features constructed through projecting the 2D features of multi-view 2D images extracted by pre-trained 2D Vision-Language Models into 3D space, 3D-LLM struggles to directly capture the complex spatial relationships between objects and the structure of 3D scenes. [13,27] directly extract 3D features from the reconstructed 3D point cloud and support multi-modal visual prompts (text, images, 3D objects) in an MLLM. To alleviate the difficulty LLMs face in understanding complex 3D scenes, [22] choose to first explicitly segment the objects in the 3D scene and then perform multi-stage object-aware scenetest alignment to achieve 3D scene understanding. However, due to the lack of large-scale 3D-language alignment data and the intricate content of 3D scenes, although current MLLMs can achieve favorable performance in 3D scene understanding tasks, their localization performance is still significantly behind the 3D localization specialists. Our approach seeks to address this issue by introducing a 3D grounding model to enhance the localization capability of MLLMs.", + "bbox": [ + 212, + 347, + 787, + 664 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 ScanReason Benchmark", + "text_level": 1, + "bbox": [ + 214, + 683, + 485, + 699 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 3D Reasoning Grounding Task", + "text_level": 1, + "bbox": [ + 214, + 712, + 516, + 727 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a 3D scene and an implicit free-form query, 3D reasoning grounding requires the model to predict the 3D bounding boxes of the target objects as well as the textual answers and explanations. As shown in Fig. 2, different from traditional 3D visual grounding, the queries of 3D reasoning grounding are implicit and complex, requiring strong reasoning, commonsense, and world knowledge. The number of target objects in 3D reasoning grounding is flexible and any object satisfying the requirements should be considered as the target object.", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "C. Zhu et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8f87ed3d7733da89da0d9d3b3bf4ad2f2ebd2e4c9ff0528c031dc4fad4c9d162.jpg", + "image_caption": [ + "Fig. 2: The left side figure shows the overall of our ScanReason dataset. For each reasoning category, we designed different prompts to generate corresponding questions. And the right side figure shows the differences between the traditional 3D visual grounding task and our proposed 3D reasoning grounding task." + ], + "image_footnote": [], + "bbox": [ + 218, + 145, + 467, + 277 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3513ff74a02228af403122d8bdd0ee082eeb2998727007dc98136604b0aed7d1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 478, + 162, + 784, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Question Types", + "text_level": 1, + "bbox": [ + 215, + 378, + 390, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To comprehensively evaluate the 3D reasoning grounding abilities, we define 5 types of questions depending on which type of reasoning is required. Spatial reasoning and functional reasoning require a fundamental understanding of the 3D physical world, while logical reasoning, emotional reasoning, and safety reasoning are high-level reasoning skills built upon the two fundamental reasoning abilities to address user-centric real-world applications, as shown in Fig. 2.", + "bbox": [ + 212, + 411, + 784, + 501 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Spatial Reasoning measures models' understanding of 3D spatial relationships among objects in 3D scene. It encompasses the ability to comprehend the layout of the 3D scene and the 3D location of objects within it, which could serve as the foundation for navigating or planning movements in the 3D environment.", + "bbox": [ + 212, + 503, + 784, + 563 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Functional Reasoning involves understanding and inferring the purpose, function, or affordance of objects within the 3D scene. For example, functional reasoning allows an embodied agent to recognize that a chair is for sitting, a lamp is for lighting, and a refrigerator is for storing food at low temperatures. Such understanding enables the embodied agent to assist users and to perform complex tasks more effectively (e.g., turning on a lamp when the room gets dark, or navigating to a refrigerator to fetch a drink).", + "bbox": [ + 212, + 565, + 784, + 670 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Logical Reasoning allows an embodied agent to not only understand its environment but also to interact with it in a goal-directed manner. For example, given a question shown in Fig. 2: \"If I'm cooking dinner in the kitchen, where is the nearest place to throw the rubbish?\", an agent needs to use such reasoning ability to infer the location of objects (in this question, a rubbish bin) based on their function and spatial relationships under the specific setting (the kitchen).", + "bbox": [ + 212, + 672, + 784, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Emotional Reasoning plays a critical role in human-robot interaction, where the target objects are determined by understanding human emotions, preferences, and behavioral patterns. This ability makes the embodied agents more attuned to the emotional and psychological states of humans, allowing them to provide more personalized, empathetic, and contextually appropriate responses", + "bbox": [ + 212, + 763, + 784, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "ScanReason", + "bbox": [ + 648, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and solutions, such as: \"I'm very sad, is there something that can make me feel happy?\" shown in 2.", + "bbox": [ + 212, + 146, + 782, + 175 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Safety Reasoning focuses on preventing harm and ensuring the well-being of humans in the 3D environment. It requires the embodied agent to identify and assess the risk and make safety-aware decisions, such as: \"Where should first aid kits be placed for easy access but out of reach of children?\"", + "bbox": [ + 212, + 176, + 784, + 236 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Automatic Data Annotation with GPT-4", + "text_level": 1, + "bbox": [ + 214, + 258, + 599, + 273 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We leverage the 3D scenes and bounding box annotations from the EmbodiedScan dataset [37] and apply GPT-4 [30] to automatically generate question-answer-location pairs for the five question types respectively. Specifically, we provide GPT-4 with the categories and bounding box locations of all objects in the scene, and ask GPT-4 to generate questions and answers with the target object ids of the provided objects. The details can be found in the Appendix.", + "bbox": [ + 212, + 284, + 782, + 375 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In total, our ScanReason dataset consists of 12929 complex reasoning question-answer-location pairs from 1456 scenes, which are split into 11455 training pairs and 1474 validation pairs. All 1474 validation question-answer pairs have been manually verified, including 342 spatial reasoning questions, 287 functional reasoning questions, 581 logical reasoning questions, 211 emotional questions, and 53 safety reasoning questions. We provide detailed statistics and more examples of our dataset in the Appendix.", + "bbox": [ + 212, + 375, + 787, + 481 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4 Evaluation Metric", + "text_level": 1, + "bbox": [ + 214, + 503, + 411, + 517 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To evaluate the accuracy of predicted objects and their locations for a flexible number of ground-truth objects, we follow the evaluation metric of 3D object detection tasks. Specifically, we adopt Acc@kIoU as our metric, where $k$ is the threshold for the Intersection of Union (IoU) between positive predictions and ground truths. We evaluate the performance under $k = 0.25$ in our experiments.", + "bbox": [ + 212, + 527, + 782, + 603 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 Method", + "text_level": 1, + "bbox": [ + 214, + 627, + 330, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Solving the task of 3D reasoning grounding requires the synergization of the perception, reasoning, and grounding capability of the embodied agent. Intuitively, we can first conduct reasoning based on implicit instruction such as \"where should first aid kits be placed?\" and the visual environment. The reasoning process provides us with information about the rough location and semantics of the object we are looking for. Then keeping that information in mind, we look back to the 3D environment to precisely locate the object. For complex scenarios, alternate reasoning and looking back are required for multiple rounds until we obtain the final answer.", + "bbox": [ + 212, + 657, + 782, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Inspired by this intuition, we propose ReGround3D, consisting of a visual-centric reasoning module and a 3D grounding module with geometry-enhanced look-back, as illustrated in Fig. 3. The visual-centric reasoning module (Sec. 4.1)", + "bbox": [ + 212, + 795, + 782, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "C. Zhu et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "performs joint reasoning of language instruction and visual scene, and predicts a special $\\langle \\mathsf{LOC} \\rangle$ token representing the grounding information. The 3D grounding module (Sec. 4.2) looks back to the original 3D scene with comprehensive geometry information and fine-grained details. It takes the hidden embedding of the $\\langle \\mathsf{LOC} \\rangle$ token containing grounding-related information from the 3D features and eventually predicts the 3D locations of the target objects. Furthermore, we propose a Chain-of-Grounding mechanism (CoG) (Sec. 4.3), i.e., a chain of interleaved reasoning and grounding steps, to further synergize the grounding and reasoning capability for the 3D reasoning grounding task, as illustrated in Fig. 4.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1 Visual-Centric Reasoning", + "text_level": 1, + "bbox": [ + 215, + 306, + 472, + 321 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Due to the complexity of the 3D scene and user instructions, particularly given the implicit intention behind the human instruction, ReGround3D starts with a visual-centric reasoning module that can perceive the scene, comprehend the human instructions, and conduct joint reasoning of 3D scene and instructions. We believe the reasoning process eventually implies the grounding intention, i.e., implicitly encodes the information indicating the target object to solve the task. Thus, we design the visual-centric reasoning module to predict grounding queries for localizing the target objects in the following 3D grounding module.", + "bbox": [ + 212, + 334, + 785, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Specifically, for simplicity, we leverage 3D-LLM to serve as the visual-centric reasoning module because of its strong reasoning abilities inherited from the LLM. Based on the BLIP2 architecture [26], 3D-LLM uses pre-trained image encoders to extract multi-view 2D image features and back-project them into 3D spaces. The visual features are encoded by the Q-Former to produce 32 tokens as the visual input to the LLM. By leveraging the pre-trained image encoders and the Q-Former, the visual tokens encode rich semantics but lack the 3D structures, spatial interactions, and fine-grained details. Therefore, instead of directly predicting the object locations by the 3D-LLM, we ask the 3D-LLM to predict the feature representation as the output of the reasoning process, and the predicted feature is further used to ground the target object in the grounding module.", + "bbox": [ + 212, + 455, + 787, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To enable the prediction of the grounding feature, we expand the original vocabulary of the 3D-LLM with a special $\\langle \\mathsf{L0C} \\rangle$ token. The $\\langle \\mathsf{L0C} \\rangle$ token is laden with the contextual scene and the target object information which can guide the 3D grounding module to accurately localize target objects.", + "bbox": [ + 212, + 636, + 787, + 698 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2 3D Grounding with Geometry-Enhanced Look-Back", + "text_level": 1, + "bbox": [ + 215, + 722, + 692, + 738 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Once obtaining the token, ReGround3D extracts the last-layer embedding $h_{loc}$ of the token and sends it into the 3D grounding module to predict the 3D bounding boxes. The 3D grounding module is devised with a \"look-back\" mechanism which allows the model to access the 3D geometry and fine-grained details from a 3D point cloud encoder. The fine-grained geometry-enhanced 3D visual features and $h_{loc}$ are sent into a query selection module to retrieve the most", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "ScanReason", + "bbox": [ + 648, + 114, + 730, + 127 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a63d7526e0e73311efac638837eb1ad1a53e0edcda7dcf7e3c1305227a8d3141.jpg", + "image_caption": [ + "Fig. 3: The pipeline of ReGround3D. Given the 3D scene and human instruction, the visual reasoning module first performs joint 3D scene and instruction reasoning, and then guide the 3D grounding module to look-back the 3D scene and perform target object location." + ], + "image_footnote": [], + "bbox": [ + 225, + 146, + 781, + 299 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "relevant object features. Those features are further decoded into 3D bounding boxes with the 3D box decoder.", + "bbox": [ + 212, + 402, + 787, + 433 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3D Visual Encoder Unlike the 2D image encoder used for 3D-LLM, the 3D visual encoder directly extracts features from 3D point clouds to capture more geometric and spatial information about the 3D structures and fine-grained details. The powerful 3D visual encoder which captures comprehensive geometry, structure, layout, and fine-grained information is critical to accurate 3D grounding. Subsequently, the 3D features $f_{scene}$ produced by the 3D visual encoder and the grounding feature $h_{loc}$ from 3D-LLM are sent to the query selection module.", + "bbox": [ + 212, + 436, + 787, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Query Selection Module We adopt a cross-attention mechanism, where we treat $f_{\\text{scene}}$ as $Q$ (Query), and $h_{\\text{loc}}$ as both $K$ (Key) and $V$ (Value), to implicitly obtain a feature-level reasoning activation heatmap. During this scene look-back process, this module roughly locates scene features that have a high response to the token. We then select the $k$ most relevant features as the object query $f_{\\text{query}}$ .", + "bbox": [ + 212, + 546, + 789, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3D Box Decoder is a classical transformer decoder, which consists of M transformer decoder layers. In each decoder layer, the object queries $f_{query}$ go through the text feature $h_{loc}$ cross-attention layer and scene feature $f_{scene}$ cross-attention layer. Finally, the prediction head takes the updated object queries as input and predicts the final 3D locations and matching score.", + "bbox": [ + 212, + 640, + 787, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Discussion In comparison to previous works [11,13,15,21] that directly predicts the bounding boxes by the LLM, the extra grounding module has the following advantages: 1) Based on the MLLM reasoning results, it allows the grounding module to perceive the scene again and focus on the region under the implicit guidance of MLLM, adapted to the user queries and the reasoning results. 2) The scene representation perceived in the grounding module can be more precise and fine-grained, which is complementary to the visual-centric reasoning module. 3) The two-step reasoning-grounding pipeline is flexible and can generalized", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "C. Zhu et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7edbca3f9be2f8c93f5dd8fc66e31685c5b11300286c4d82ea9516bfe80a1150.jpg", + "image_caption": [ + "Fig. 4: Illustration of Chain-of-Grounding (CoG) Mechanism" + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 782, + 305 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "to other types of predicting formats such as segmentation masks (by simply replacing the 3D box decoder with a 3D mask decoder).", + "bbox": [ + 212, + 359, + 784, + 390 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3 Chain-of-Grounding Mechanism", + "text_level": 1, + "bbox": [ + 214, + 410, + 526, + 425 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The existing design conducts the reasoning and grounding process sequentially, i.e., the reasoning process is finished before grounding. We argue that the grounding results can also facilitate the reasoning process, especially for those requiring spatial information. Thus, to further synergize the reasoning and grounding process, inspired by chain-of-thought (CoT) [38], we propose Chain-of-Grounding (CoG), which introduces a chain of interleaved steps of reasoning and grounding to find the targeted objects during inference, as shown in Fig. 4. Such a process allows the model to actively find relevant objects that help solve the problem, and then conduct reasoning with the assistance of the additional information of these relevant objects so that the model can more precisely find the target objects.", + "bbox": [ + 212, + 433, + 787, + 598 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Specifically, given a question provided by users, CoG translates it into another question of finding the explicitly mentioned objects in the original question. The generated new question is sent into ReGround3D to ground the objects mentioned in the original question in the 3D scene with corresponding confidence scores. An object can be seen as a successfully located object when its confidence score is above the threshold, and the located object information could serve as explicit guidance for 3D-LLM in the next reasoning stage. As shown in Fig. 4, after obtaining the 3D information of objects explicitly mentioned in the original question, the located object information is inserted to update the question. The updated question is then sent to ReGround3D to perform reasoning and grounding to output the target object locations.", + "bbox": [ + 212, + 599, + 787, + 767 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.4 Instruction Tuning", + "text_level": 1, + "bbox": [ + 214, + 786, + 418, + 801 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Training Objective We use the pre-trained weights of 3D-LLM as the initialization for the visual-centric reasoning module. Except for freezing the 3D", + "bbox": [ + 214, + 809, + 785, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "ScanReason", + "bbox": [ + 648, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/73f7715e84ac56e717b9b4730eae7140c79d3bb296692c8441fa68f9dad71519.jpg", + "table_caption": [ + "Table 1: We list part of data templates used to train ReGround3D for each task." + ], + "table_footnote": [], + "table_body": "
Task NameText InstructionsOutput Type TemplatesExpected Output
3D Visual Grounding<scene> Here is a description Please answer the question only with outabout an object: "<expr>", whereis the object in the 3D scene?Please answer the question only with out-put 3D box prediction(s).It is <LOC>.
3D/Spatial Question Answering<scene> Answer the question Please answer the question only with text,do not output 3D box prediction(s).<question>Please answer the question only with text,do not output 3D box prediction(s).<answer>.
<scene><situation>, <question>Please answer the question only with text,do not output 3D box prediction(s).<answer>.
<scene><question>Please answer the question only with text,do not output 3D box prediction(s).<answer>.
3D Object Detection<scene> Where is the <category> Please answer the question only with out-in this 3D scene? Put 3D box prediction(s).Sure, <LOC>.
3D Reasoning Grounding<scene> Answer the question: Please answer the question with text andoutput 3D box prediction(s).Sure, <LOC>, <reason>
", + "bbox": [ + 218, + 171, + 784, + 299 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "visual encoder pre-trained on [37], the rest of the parameters in the visual-centric reasoning module and 3D grounding module in our framework are trained in an end-to-end manner. The training supervision is a weighted sum of the next token prediction loss from 3D-LLM and the 3D detection loss from the 3D grounding module.", + "bbox": [ + 212, + 323, + 784, + 398 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda_ {\\text {t e x t}} \\mathcal {L} _ {\\text {t e x t}} + \\lambda_ {\\text {d e t}} \\mathcal {L} _ {\\text {d e t}} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 412, + 785, + 429 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The 3D detection loss is defined following:", + "bbox": [ + 215, + 433, + 522, + 449 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {d e t}} = \\lambda_ {I O U} \\mathcal {L} _ {I O U} + \\lambda_ {\\text {c o n t r a s t}} \\mathcal {L} _ {\\text {c o n t r a s t}} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 362, + 463, + 785, + 479 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Instruction Tuning Dataset We load the pre-trained weights of 3D-LLM and the 3D visual encoder, and finetune the LoRA of 3D-LLM, the query selection module, and the 3D box decoder with an instruction tuning dataset. To construct the instruction tuning dataset, we reformulate the data annotations from existing 3D datasets into question-answer or question-answer-bounding-box pairs. Specifically, the 3D visual grounding data from ScanRefer [7], SR3D, NR3D [1] and the 3D object detection data from EmbodiedScan [37] are formulated into question-answer-bbox pairs, and the 3D/spatial question answering data from SR3D [1], CLEVR3D [41], SQA3D [29] are formulated into question-answer pairs without bounding boxes. The information shown in Tab. 1 illustrates how we unify the instruction and output with task-specific templates. More details can be found in the Appendix. The reformulated data combined with our proposed ScanReason dataset, serve as the instruction tuning dataset.", + "bbox": [ + 212, + 497, + 787, + 695 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 Experiment", + "text_level": 1, + "bbox": [ + 215, + 714, + 366, + 732 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.1 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 742, + 455, + 758 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Network Architecture For the 3D grounding module, we adopt the pretrained point cloud encoder as the 3D visual encoder. During the training stage, We use LoRA to efficiently finetune the 3D-LLM to preserve the original 3D scene understanding capability and reduce the computation costs. The number of object queries $k$ in the query selection module is set to 256.", + "bbox": [ + 212, + 763, + 784, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "C. Zhu et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Training Parameters The training is done on 8 NVIDIA A100 GPUs. We adopt the AdamW optimizer with a learning rate of 3e-4 and use a learning rate scheduler WarmupDecayLR with the warmup steps of 100. The total batch size is set to be 16. The loss weight parameters $\\lambda_{text}$ and $\\lambda_{det}$ in total loss $\\mathcal{L}$ are set to 1.0 and 1.0, respectively, and the weight $\\lambda_{IOU}$ and $\\lambda_{contrast}$ in $\\mathcal{L}_{\\mathrm{det}}$ are set to 1.0 and 1.0.", + "bbox": [ + 218, + 145, + 787, + 234 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.2 Results Comparison", + "text_level": 1, + "bbox": [ + 218, + 263, + 426, + 279 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Evaluation on 3D Visual Grounding In order to verify the superiority of our model in grounding ability and facilitate comparison between current models, we report the explicit grounding performance on the existing 3D visual grounding task. Since the evaluation settings of Nr3D and Sr3D [1] are based on ground-truth object proposals, while ScanRefer [7] requires models to output 3D bounding boxes, we choose ScanRefer as our benchmark for comparison. We divide the existing methods into two categories, one is the grounding model designed specifically for the 3D visual grounding task, and the other is generalist MLLMs which can understand a variety of 3D vision-language tasks. The original 3D-LLM embeds 3D locations in the vocabularies and represents the grounded 3D bounding boxes by a sequence of discrete location tokens. However, since the fine-tuned model of 3D-LLM on ScanRefer and related location tokens implementation are not accessible, we adapt 3D-LLM to directly output 3D numerical coordinates representing 3D bounding boxes by fine-tuning the pre-trained model on our reformulated 3D visual grounding data, denotes as 3D-LLM (vg). As shown in Tab. 2, current generalist MLLM models still lag behind the specialist models in terms of grounding ability. By incorporating the 3D grounding module into MLLM, ReGround3D shows the SOTA performance on the traditional 3D visual grounding task.", + "bbox": [ + 217, + 294, + 785, + 580 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/1b26609c0c3c352419e104b42bfe4469cc0a3cf471d6640ac86429ef00682a88.jpg", + "table_caption": [ + "Table 2: Results on 3D visual grounding task among ReGround3D (ours) and existing methods." + ], + "table_footnote": [], + "table_body": "
TypeMethodsAcc@0.25Acc@0.5
SpecialistsScanRefer [7]37.324.3
MVT [23]40.833.3
3DVG-Trans [48]45.934.5
ViL3DRel [12]47.937.7
BUTD-DETR [24]52.239.8
L3Det [50]52.840.2
Generalized MLLMsLLM-Grounder [42]17.15.3
3D-LLM [21]30.3-
3D-LLM(vg) [21]33.128.7
Chat3D-v2 [22]35.930.4
oursReGround3D53.141.1
", + "bbox": [ + 333, + 650, + 671, + 806 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "ScanReason", + "bbox": [ + 648, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/60b3b170fe061f406e68d6a7b3d27e431f9f77b9520caac20eb8c891668725f3.jpg", + "table_caption": [ + "Table 3: Results (Acc@0.25) on 3D reasoning grounding task among ReGround3D (ours) and existing methods." + ], + "table_footnote": [], + "table_body": "
MethodsLLMSpatial FunctionalLogicalEmotional SafetyOverall
Mask3D [32] + InternLM2-7B [33]InternLM2-7B10.3436.129.988.218.9914.86
3D-LLM(vg) [21]FlanT5XL-3B18.3117.4210.978.126.3313.29
Chat3D-v2 [22]Vicuna-7B20.2118.3911.327.989.8814.98
ReGround3D*FlanT5XL-3B30.7629.818.6719.2217.1223.27
ReGround3DFlanT5XL-3B32.9836.2326.9923.1222.9828.98
ReGround3D (CoG)FlanT5XL-3B34.7136.7929.1124.0323.2130.62
", + "bbox": [ + 217, + 184, + 785, + 282 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Evaluation on 3D Reasoning Grounding The various visual grounding models rely on explicit text-object alignment in the input object expression to achieve localization, which falls to be applied to 3D reasoning grounding task. We performed a comparison between our proposed method ReGround3D and existing MLLM methods, including 3D-LLM(vg) and Chat3D-v2 [22]. Besides, inspired by Chat-3D v2, which first segments objects, then equips them with unique object identifiers to conduct effective object grounding, we set up a LLM-based 3D reasoning grounding baseline: We first segment the objects from the scene using a 3D instance segmentor [32], then convert the segmented object information including their categories and 3D bounding boxes into text as input of LLM (InternLM2-7B [33]). Besides, to better validate the performance of our model and ensure a fair comparison, we remove the ScanReason dataset from the training data, denoted as ReGround3D*. As shown in Tab. 3, the LLM-based reasoning method (Mask3D [32] + InternLM2-7B [33]) possesses a very strong functional reasoning ability, but struggles to understand of 3D spatial relationship. Additionally, irrespective of whether ScanReason is used in training, our model significantly outperforms the existing MLLMs. By synergizing the reasoning and grounding process utilizing the Chain-of-Grounding (CoG) mechanism, the 3D reasoning grounding performance of ReGround3D can be further improved (from 28.98 to 30.62), especially on the spatial reasoning and logical reasoning questions. The qualitative results comparison shown in Fig. 5 demonstrates the superiority of our method in reasoning human complex instruction based on the 3D scene.", + "bbox": [ + 212, + 311, + 787, + 657 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.3 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 683, + 388, + 698 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In this section, we conduct an extensive ablation study to verify the effectiveness of each component.", + "bbox": [ + 212, + 709, + 785, + 739 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Effectiveness of 3D Grounding Module In order to more comprehensively verify the effectiveness of the 3D grounding module we proposed, we step by step verify the performance changes from 3D-LLM to ReGround3D. Apart from the 3D-LLM(vg), we fine-tuned the 3D-LLM model respectively on full reformulated existing data and all instruction tuning data, including ScanReason, denoted as", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "C. Zhu et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/784f1f2f5b6f232eba3420d872b915d4273734ee0cbb2c82f4a2d907c77908c0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 143, + 238, + 159 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "It's so cold outside, Where could I find my thick coat?", + "bbox": [ + 243, + 146, + 486, + 157 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/27dcfd88b158e59479e6576ac7425d522a093628a2d07ff552c4f872f11c92b2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 222, + 162, + 772, + 258 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/c9fa2146b542c300c78a46511bc63b7a9b3ae39054320d18b0a71467da5acf5e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 262, + 238, + 277 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Where can I find paper to wipe my hands?", + "bbox": [ + 241, + 265, + 429, + 277 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/4511c469625c1467fc4835c6fcd6ab695eebe2823a680f18daa75884368e1258.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 220, + 282, + 782, + 369 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/8d2c4f40c5021c3aa2fe6791c813035cbda4adb46952763b92b652089c94293b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 372, + 238, + 387 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "It's dark here, can you find something to light up the room?", + "bbox": [ + 243, + 375, + 509, + 387 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/631c09b5775357e613af7d1fbcb51c92ffc52729ec5abae40113360c2e24115e.jpg", + "image_caption": [ + "Fig. 5: Visualization comparison of 3D reasoning grounding capability between ReGround3D and 3D-LLM. Our method could achieve much more accurate grounding results which satisfy the implicit question intention, and give the corresponding explanation at the same time. More illustrations are given in the Appendix." + ], + "image_footnote": [], + "bbox": [ + 220, + 393, + 782, + 488 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "3D-LLM (full) and 3D-LLM (full+sr). All the tokens in answers of the data used for fine-tuning 3D-LLM have been converted to the numerical coordinates of the corresponding boxes. The results in Tab. 4a show that with the same tuning dataset, our ReGround3D achieves far superior performance to 3D-LLM (28.98 vs. 19.21) by introducing the 3D grounding module.", + "bbox": [ + 214, + 595, + 787, + 672 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effectiveness of Instruction Tuning Dataset Tab. 4b showcases the impact of different training data types on 3D visual grounding performance. 3D Object Detection (3D OD) provides the explicit 3D semantic category and visual alignment whereas 3D Question Answering (3D QA) data injects the basic 3D scene understanding ability into the model, which has certain benefits for the visual grounding ability of the model. In addition, we find that training with 3D reasoning grounding dataset can further improve the performance on 3D visual grounding.", + "bbox": [ + 212, + 718, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "ScanReason", + "bbox": [ + 648, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "(a) When using the same tuning dataset, ReGround3D achieves far better performance on ScanReason than 3D-LLM.", + "bbox": [ + 215, + 183, + 475, + 215 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "(b) Ablation study on training data. We evaluate through the metric of accuracy on the val set of the ScanRefer dataset.", + "bbox": [ + 482, + 183, + 781, + 215 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/66f53d6621b98ac4cd6d90f381b3e91320cc13af5a304578bb256809e79520fa.jpg", + "table_caption": [ + "Table 4: Ablation study on effectiveness of 3D grounding module and training data" + ], + "table_footnote": [], + "table_body": "
MethodsAcc@0.25
3D-LLM(vg) [21]13.29
3D-LLM(full) [21]15.31
3D-LLM(full+sr) [21]19.21
ReGround3D(full+sr)28.98
", + "bbox": [ + 232, + 227, + 455, + 311 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/efb50851875cbe0258e27622956b185a168af506f8a852d0eaf46a030f9db5a4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetAcc@0.25Acc@0.5
VGODVQARD
✓✓19.314.2
✓✓48.737.6
✓✓49.238.1
✓✓✓51.839.4
✓✓✓✓53.141.1
", + "bbox": [ + 522, + 227, + 736, + 319 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Discussion of CoG Mechanism While the CoG Mechanism boosts the performance with interleaved reasoning and grounding steps during inference, it uses the relevant object information explicitly presented in the question to help find the target objects. A natural question arises: if we input the information of all the objects instead of relevant objects into ReGround3D during CoG, will this make the model more accurately find the target objects? We first use the existing 3D Segmentor [32] to segment all objects in the scene, then update the original question using all the object 3D bounding boxes information according to the template in Sec. 4.3 and send into ReGround3D. However, experiments show that using all the object information will instead reduce the 3D reasoning grounding performance from 28.98 to 27.67. One possible reason could be the model's attention is dispersed by too many irrelevant objects.", + "bbox": [ + 212, + 347, + 787, + 529 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 551, + 359, + 568 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This paper introduces a new 3D vision language learning task: 3D reasoning grounding, which requires the model to perform active reasoning over complex and implicit human instruction, localize the target objects, and give corresponding explanations. Furthermore, we propose ScanReason, a new dataset and benchmark to further unlock and thoroughly evaluate the 3D reasoning grounding capability. Based on this dataset, we propose a novel approach: ReGround3D, which utilizes the strong reasoning capability of MLLM guiding the 3D grounding module to obtain accurate object locations, and a Chain of Grounding (CoG) mechanism is presented to further boost the performance with interleaved reasoning and grounding steps during inference. We believe that our work will further the natural interaction between embodied agents and humans in open 3D environments. For the current ScanReason benchmark, we find that the questions in three high-level 3D reasoning categories may have overlaps. For a certain reasoning question, similar questions may appear in one or two other categories. We leave the problem as a future challenge for better reasoning grounding ability evaluation.", + "bbox": [ + 212, + 584, + 787, + 825 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "C. Zhu et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work is supported in part by HKU Startup Fund, HKU Seed Fund for Basic Research, HKU Seed Fund for Translational and Applied Research, HKU IDS research Seed Fund, and HKU Fintech Academy R&D Funding.", + "bbox": [ + 212, + 178, + 787, + 224 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 248, + 323, + 263 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Achlioptas, P., Abdelreheem, A., Xia, F., Elhoseiny, M., Guibas, L.: Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16. pp. 422-440. Springer (2020)", + "2. Azuma, D., Miyanishi, T., Kurita, S., Kawanabe, M.: Scanqa: 3d question answering for spatial scene understanding. In: proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 19129-19139 (2022)", + "3. Bakr, E.M., Ayman, M., Ahmed, M., Slim, H., Elhoseiny, M.: Cot3dref: Chain-of-thoughts data-efficient 3d visual grounding. arXiv preprint arXiv:2310.06214 (2023)", + "4. Baroni, M.: Linguistic generalization and compositionality in modern artificial neural networks. Philosophical Transactions of the Royal Society B 375(1791), 20190307 (2020)", + "5. Cai, D., Zhao, L., Zhang, J., Sheng, L., Xu, D.: 3djcg: A unified framework for joint dense captioning and visual grounding on 3d point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16464-16473 (2022)", + "6. Chen, C., Qin, R., Luo, F., Mi, X., Li, P., Sun, M., Liu, Y.: Position-enhanced visual instruction tuning for multimodal large language models. arXiv preprint arXiv:2308.13437 (2023)", + "7. Chen, D.Z., Chang, A.X., Niefer, M.: Scanrefer: 3d object localization in rgb-d scans using natural language. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XX. pp. 202-221. Springer (2020)", + "8. Chen, D.Z., Wu, Q., Nießner, M., Chang, A.X.: D3net: A speaker-listener architecture for semi-supervised dense captioning and visual grounding in rgb-d scans. arXiv preprint arXiv:2112.01551 (2021)", + "9. Chen, F., Han, M., Zhao, H., Zhang, Q., Shi, J., Xu, S., Xu, B.: X-LLM: Bootstrapping advanced large language models by treating multi-modalities as foreign languages. arXiv preprint arXiv:2305.04160 (2023)", + "10. Chen, J., Zhu, D., Shen, X., Li, X., Liu, Z., Zhang, P., Krishnamoorthi, R., Chandra, V., Xiong, Y., Elhoseiny, M.: Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478 (2023)", + "11. Chen, K., Zhang, Z., Zeng, W., Zhang, R., Zhu, F., Zhao, R.: Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195 (2023)", + "12. Chen, S., Guhur, P.L., Tapaswi, M., Schmid, C., Laptev, I.: Language conditioned spatial relation reasoning for 3d object grounding. Advances in Neural Information Processing Systems 35, 20522-20535 (2022)" + ], + "bbox": [ + 217, + 281, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "ScanReason", + "bbox": [ + 648, + 114, + 730, + 127 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "13. Chen, S., Chen, X., Zhang, C., Li, M., Yu, G., Fei, H., Zhu, H., Fan, J., Chen, T.: Ll3da: Visual interactive instruction tuning for omni-3d understanding, reasoning, and planning. arXiv preprint arXiv:2311.18651 (2023)", + "14. Chen, S., Zhu, H., Chen, X., Lei, Y., Yu, G., Chen, T.: End-to-end 3d dense captioning with vote2cap-detr. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11124-11133 (2023)", + "15. Chen, T., Saxena, S., Li, L., Fleet, D.J., Hinton, G.: Pix2seq: A language modeling framework for object detection. arXiv preprint arXiv:2109.10852 (2021)", + "16. Chen, Z., Gholami, A., Nießner, M., Chang, A.X.: Scan2cap: Context-aware dense captioning in rgb-d scans. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3193-3203 (2021)", + "17. Chen, Z., Hu, R., Chen, X., Nießner, M., Chang, A.X.: Unit3d: A unified transformer for 3d dense captioning and visual grounding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 18109-18119 (2023)", + "18. Chung, H.W., Hou, L., Longpre, S., Zoph, B., Tay, Y., Fedus, W., Li, E., Wang, X., Dehghani, M., Brahma, S., et al.: Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416 (2022)", + "19. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning (2023)", + "20. Han, J., Zhang, R., Shao, W., Gao, P., Xu, P., Xiao, H., Zhang, K., Liu, C., Wen, S., Guo, Z., et al.: Imagebind-llm: Multi-modality instruction tuning. arXiv preprint arXiv:2309.03905 (2023)", + "21. Hong, Y., Zhen, H., Chen, P., Zheng, S., Du, Y., Chen, Z., Gan, C.: 3d-llm: Injecting the 3d world into large language models. Advances in Neural Information Processing Systems 36 (2024)", + "22. Huang, H., Wang, Z., Huang, R., Liu, L., Cheng, X., Zhao, Y., Jin, T., Zhao, Z.: Chat-3d v2: Bridging 3d scene and large language models with object identifiers. arXiv preprint arXiv:2312.08168 (2023)", + "23. Huang, S., Chen, Y., Jia, J., Wang, L.: Multi-view transformer for 3d visual grounding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15524-15533 (2022)", + "24. Jain, A., Gkanatsios, N., Mediratta, I., Fragkiadaki, K.: Bottom up top down detection transformers for language grounding in images and point clouds. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVI. pp. 417-433. Springer (2022)", + "25. Jiao, Y., Chen, S., Jie, Z., Chen, J., Ma, L., Jiang, Y.G.: More: Multi-order relation mining for dense captioning in 3d scenes. arXiv preprint arXiv:2203.05203 (2022)", + "26. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023)", + "27. Li, M., Chen, X., Zhang, C., Chen, S., Zhu, H., Yin, F., Yu, G., Chen, T.: M3dbench: Let's instruct large models with multi-modal 3d prompts. arXiv preprint arXiv:2312.10763 (2023)", + "28. Luo, J., Fu, J., Kong, X., Gao, C., Ren, H., Shen, H., Xia, H., Liu, S.: 3d-sps: Single-stage 3d visual grounding via referred point progressive selection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16454-16463 (2022)", + "29. Ma, X., Yong, S., Zheng, Z., Li, Q., Liang, Y., Zhu, S.C., Huang, S.: Sqa3d: Situated question answering in 3d scenes. arXiv preprint arXiv:2210.07474 (2022)" + ], + "bbox": [ + 217, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "C. Zhu et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "30. OpenAI: Gpt-4 technical report (2023)", + "31. Roh, J., Desingh, K., Farhadi, A., Fox, D.: Languagerefer: Spatial-language model for 3d visual grounding. In: Conference on Robot Learning. pp. 1046-1056. PMLR (2022)", + "32. Schult, J., Engelmann, F., Hermans, A., Litany, O., Tang, S., Leibe, B.: Mask3d: Mask transformer for 3d semantic instance segmentation. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 8216-8223. IEEE (2023)", + "33. Team, I.: Internl'm: A multilingual language model with progressively enhanced capabilities (2023)", + "34. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023)", + "35. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017)", + "36. Wang, H., Zhang, C., Yu, J., Cai, W.: Spatiality-guided transformer for 3d dense captioning on point clouds. arXiv preprint arXiv:2204.10688 (2022)", + "37. Wang, T., Mao, X., Zhu, C., Xu, R., Lyu, R., Li, P., Chen, X., Zhang, W., Chen, K., Xue, T., et al.: Embodiedscan: A holistic multi-modal 3d perception suite towards embodied ai. arXiv preprint arXiv:2312.16170 (2023)", + "38. Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q.V., Zhou, D., et al.: Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems 35, 24824-24837 (2022)", + "39. Wu, Y., Cheng, X., Zhang, R., Cheng, Z., Zhang, J.: Eda: Explicit textdecoupling and dense alignment for 3d visual and language learning. arXiv preprint arXiv:2209.14941 (2022)", + "40. Xu, R., Wang, X., Wang, T., Chen, Y., Pang, J., Lin, D.: Pointllm: Empowering large language models to understand point clouds. arXiv preprint arXiv:2308.16911 (2023)", + "41. Yan, X., Yuan, Z., Du, Y., Liao, Y., Guo, Y., Li, Z., Cui, S.: Clevr3d: Compositional language and elementary visual reasoning for question answering in 3d real-world scenes. arXiv preprint arXiv:2112.11691 (2021)", + "42. Yang, J., Chen, X., Qian, S., Madaan, N., Iyengar, M., Fouhey, D.F., Chai, J.: Llm-grounder: Open-vocabulary 3d visual grounding with large language model as an agent. arXiv preprint arXiv:2309.12311 (2023)", + "43. Yang, Z., Zhang, S., Wang, L., Luo, J.: Sat: 2d semantics assisted training for 3d visual grounding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1856-1866 (2021)", + "44. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178 (2023)", + "45. Ye, S., Chen, D., Han, S., Liao, J.: 3d question answering. IEEE Transactions on Visualization and Computer Graphics (2022)", + "46. Yuan, Z., Yan, X., Liao, Y., Guo, Y., Li, G., Cui, S., Li, Z.: X-trans2cap: Cross-modal knowledge transfer using transformer for 3d dense captioning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8563-8573 (2022)", + "47. Zhang, S., Roller, S., Goyal, N., Artetxe, M., Chen, M., Chen, S., Dewan, C., Diab, M., Li, X., Lin, X.V., et al.: Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068 (2022)" + ], + "bbox": [ + 215, + 147, + 785, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "ScanReason", + "bbox": [ + 648, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "48. Zhao, L., Cai, D., Sheng, L., Xu, D.: 3dvg-transformer: Relation modeling for visual grounding on point clouds. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2928-2937 (2021)", + "49. Zhong, Y., Xu, L., Luo, J., Ma, L.: Contextual modeling for 3d dense captioning on point clouds. arXiv preprint arXiv:2210.03925 (2022)", + "50. Zhu, C., Zhang, W., Wang, T., Liu, X., Chen, K.: Object2scene: Putting objects in context for open-vocabulary 3d detection. arXiv preprint arXiv:2309.09456 (2023)", + "51. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023)", + "52. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023)", + "53. Zhu, Z., Ma, X., Chen, Y., Deng, Z., Huang, S., Li, Q.: 3d-vista: Pre-trained transformer for 3d vision and text alignment. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2911-2921 (2023)" + ], + "bbox": [ + 215, + 146, + 785, + 369 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "C. Zhu et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/ScanReason_ Empowering 3D Visual Grounding with Reasoning Capabilities/c0158ad2-ca00-4982-8b2e-6cbbd910c1e7_model.json b/2024/ScanReason_ Empowering 3D Visual Grounding with Reasoning Capabilities/c0158ad2-ca00-4982-8b2e-6cbbd910c1e7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d93f7412ec977ee379a9ba1d8c06abe48afa92f3 --- /dev/null +++ b/2024/ScanReason_ Empowering 3D Visual Grounding with Reasoning Capabilities/c0158ad2-ca00-4982-8b2e-6cbbd910c1e7_model.json @@ -0,0 +1,2370 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.23, + 0.141, + 0.776, + 0.187 + ], + "angle": 0, + "content": "ScanReason: Empowering 3D Visual Grounding with Reasoning Capabilities" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.212, + 0.772, + 0.229 + ], + "angle": 0, + "content": "Chenming Zhu\\(^{1,2}\\), Tai Wang\\(^{2}\\), Wenwei Zhang\\(^{2}\\), Kai Chen\\(^{2}\\), and Xihui Liu\\(^{1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.333, + 0.239, + 0.671, + 0.298 + ], + "angle": 0, + "content": "1 The University of Hong Kong \n2 Shanghai AI Laboratory \nchairzhu@connect.hku.hk \nhttps://zcmax.github.io/projects/ScanReason" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.331, + 0.744, + 0.54 + ], + "angle": 0, + "content": "Abstract. Although great progress has been made in 3D visual grounding, current models still rely on explicit textual descriptions for grounding and lack the ability to reason human intentions from implicit instructions. We propose a new task called 3D reasoning grounding and introduce a new benchmark ScanReason which provides over 10K question-answer-location pairs from five reasoning types that require the synchronization of reasoning and grounding. We further design our approach, ReGround3D, composed of the visual-centric reasoning module empowered by Multi-modal Large Language Model (MLLM) and the 3D grounding module to obtain accurate object locations by looking back to the enhanced geometry and fine-grained details from the 3D scenes. A chain-of-grounding mechanism is proposed to further boost the performance with interleaved reasoning and grounding steps during inference. Extensive experiments on the proposed benchmark validate the effectiveness of our proposed approach." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.552, + 0.741, + 0.581 + ], + "angle": 0, + "content": "Keywords: 3D reasoning grounding \\(\\cdot\\) 3D visual grounding \\(\\cdot\\) Multi-modal large language models" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.605, + 0.377, + 0.621 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.636, + 0.788, + 0.786 + ], + "angle": 0, + "content": "Understanding and reasoning in the 3D visual world is critical for applications such as robotics and AR, where embodied agents are expected to understand the 3D layout and predict the 3D locations of objects based on human instructions. The example in Fig. 1 demonstrates a scenario where the question can only be solved with a comprehensive understanding of the 3D scene and joint reasoning of the question and the 3D environment. However, current 3D visual grounding models [24,39,43,50] trained on [1,7] localize objects based on explicit descriptions of the object category, attribute, and 3D spatial relationships, and lack the ability to reason the user intentions and predict object locations with implicit human instructions such as \"I am thirsty, can I have something to drink?\"" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.787, + 0.788, + 0.817 + ], + "angle": 0, + "content": "To bridge the aforementioned gap and to push the boundaries of what embodied agents can understand and how they can interact with the 3D world, we" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.524, + 0.84 + ], + "angle": 0, + "content": "† Corresponding author: xihuiliu@eee.hku.hk" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.359, + 0.127 + ], + "angle": 0, + "content": "C. Zhu et al." + }, + { + "type": "text", + "bbox": [ + 0.26, + 0.151, + 0.736, + 0.173 + ], + "angle": 0, + "content": "Question: I'm looking for a comfortable spot to watch TV immersively, where is the best place for me to lie down and watch TV?" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.184, + 0.249, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.209, + 0.253, + 0.219 + ], + "angle": 0, + "content": "GPT-4" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.188, + 0.465, + 0.276 + ], + "angle": 0, + "content": "The best spot for lying down to watch TV immersively is on a spacious, cushioned sectional sofa or an adjustable reclining bed in a room with a large screen TV mounted at an eye-level view from your lying position, ensuring comfort and optimal viewing angles. Ensure the room is dimly lit or has blackout curtains to enhance screen visibility and reduce glare." + }, + { + "type": "image", + "bbox": [ + 0.479, + 0.187, + 0.549, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.552, + 0.187, + 0.621, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.625, + 0.188, + 0.734, + 0.266 + ], + "angle": 0, + "content": "The best spot to lie down and watch TV immersively would be on the bed facing the TV in the cabinet; it provides a direct line of sight and the comfort of lying down." + }, + { + "type": "image", + "bbox": [ + 0.748, + 0.185, + 0.777, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.744, + 0.208, + 0.784, + 0.218 + ], + "angle": 0, + "content": "GPT-4V" + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.279, + 0.249, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.251, + 0.288, + 0.317, + 0.298 + ], + "angle": 0, + "content": "ReGround3D" + }, + { + "type": "image", + "bbox": [ + 0.264, + 0.309, + 0.367, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.382, + 0.319, + 0.611, + 0.36 + ], + "angle": 0, + "content": "For an optimal TV viewing experience, I recommend sitting on the couch. It is oriented towards the TV and closer to the TV compared with the bed, and it's large enough to lie down." + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.307, + 0.727, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.399, + 0.788, + 0.496 + ], + "angle": 0, + "content": "Fig. 1: For an embodied agent, they not only need to be able to understand the 3D environment and complex human instructions but also localize the target objects for interaction and navigation. Although GPT-4 (GPT-4V) have strong text (multi-modal) reasoning abilities, they lack the ability to directly perceive the 3D scene, understand the 3D spatial relationships, and output corresponding target object locations. Instead, our proposed method ReGround3D has the 3D perception, reasoning, and grounding capabilities in the real-world 3D environment." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.535, + 0.788, + 0.701 + ], + "angle": 0, + "content": "propose a new task of 3D reasoning grounding and introduce a new benchmark named ScanReason. The task requires the model to conduct joint reasoning on the question and the 3D environment before predicting the 3D locations of target objects. We define five categories of 3D reasoning: spatial reasoning, functional reasoning, logical reasoning, emotional reasoning, and safety reasoning. The first two categories focus on the fundamental understanding of the 3D physical world, while the last three categories are built upon fundamental abilities to address user-centric real-world challenges. The benchmark comprises more than 10K question-answer-3D bounding box pairs from 2K scenes belonging to the five reasoning types mentioned above. The GPT-4-assisted data annotation process largely increases the efficiency of curating such a dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We propose ReGround3D as an initial attempt to the new task of 3D Reasoning Grounding. Intuitively, for 3D grounding with implicit instructions, we need to first conduct reasoning on the language instructions and the coarse visual environment. Then with the idea of which object we want to find in mind, we look back to the 3D scene and ground the target object. For complex instructions, we may need to alternate the reasoning and look-back process for multiple iterations. Inspired by this intuition, our framework is composed of a visual-centric reasoning module and 3D grounding with geometry-enhanced look-back module, with a Chain-of-Grounding mechanism during inference to" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.65, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "ScanReason" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.328 + ], + "angle": 0, + "content": "alternately conduct reasoning and grounding for multiple rounds. Specifically, the visual-centric reasoning module conducts joint reasoning of the 3D scene and instructions with an MLLM. This module predicts a special token representing the semantic and location information of the target object, which is used for the grounding module. The 3D grounding module uses the output token embedding from the previous reasoning module to locate the target object by looking back at the fine-grained 3D scene representation. Unlike previous MLLMs attempting to directly predict bounding box coordinates, our look-back mechanism enables the model to capture more comprehensive 3D geometry and fine-grained object details for accurate 3D grounding. The Chain-of-Grounding mechanism is proposed to synergize reasoning and grounding, which allows multiple rounds alternating between reasoning and grounding during inference." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.33, + 0.788, + 0.495 + ], + "angle": 0, + "content": "In summary, our contributions are threefold: 1) We propose a new task of 3D reasoning grounding which requires the model to synergize reasoning and grounding. We further introduce a new benchmark ScanReason, which comprises five reasoning types (spatial reasoning, functional reasoning, logical reasoning, emotional reasoning, and safety reasoning) for the task of 3D reasoning grounding in 3D scenes. 2) We design a new framework ReGround3D with a visual-centric reasoning module and a 3D grounding module with geometry-enhanced look-back. We further introduce a Chain-of-Grounding mechanism to boost the 3D reasoning grounding ability with a chain of interleaved reasoning and grounding steps. 3) Extensive experiments demonstrate the effectiveness of the our ReGround3D on the ScanReason benchmark for 3D reasoning grounding." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.516, + 0.388, + 0.533 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.547, + 0.788, + 0.79 + ], + "angle": 0, + "content": "3D Vision and Language Learning 3D Vision-language learning (3D-VL) is garnering increasing attention, with many 3D-VL tasks focusing on how to connect the 3D world with natural language. Among them, 3D Question Answering (3D QA) [2,45] aims to enable models to provide text answers based on natural language questions. Situation Question Answering in 3D Scenes (SQA3D) [29] requires an agent to first understand its location based on a text description, then provide a reasonable answer based on the surrounding environment, which can be seen as an extension of 3D QA in the embodied AI area. 3D Visual Grounding [1,4,7,28,31,39] demands that models identify and locate target objects in a 3D scene based on given descriptions, outputting the objects' coordinates and 3D bounding boxes. These descriptions usually explicitly rely on the objects' attributes and their spatial relationships. 3D Dense Captioning [5,8,14,16,25,36,46,49] requires models to output a series of object coordinates and corresponding scene-based descriptions based on a given scene. Different from these 3D-VL tasks, the questions in 3D reasoning grounding could be more implicit and complex." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.809, + 0.787, + 0.842 + ], + "angle": 0, + "content": "3D Visual Grounding The task of 3D visual grounding is aimed at localizing the objects that are explicitly referred to by free-form guided language" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "C. Zhu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.328 + ], + "angle": 0, + "content": "expressions in the 3D scene. Inspired by the success of transformers in natural language processing, recent 3D visual grounding approaches [12,17,24,39,50,53] have started to adopt transformer [35] architectures for handling the complex relationships between language descriptions and 3D visual data. These methods leverage the self-attention mechanism of transformers to dynamically weigh the importance of different parts of the input data, facilitating a more effective grounding of textual descriptions in the 3D environment. Recent method [3] proposes a Chain-of-Thoughts module that predicts a chain of anchor objects that are subsequently utilized to localize the final target object. Compared with 3D visual grounding, our proposed 3D reasoning grounding requires the model to reason the complex question, ground target objects, and give the explanation at the same time." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.348, + 0.789, + 0.665 + ], + "angle": 0, + "content": "Multi-modal Large Language Models Recently, there has been an increasing effort to extend the powerful complex reasoning and world knowledge capabilities of LLMs [18,34,47] to other modalities [6,9,10,19,26,44,51,52]. Among these works, some have aimed to enable LLMs to understand the 3D world. [20, 40] focus on delving into LLMs' ability to comprehend 3D objects, which can not be directly applied to 3D scenes. 3D-LLM [21] is the pioneering work that incorporates the 3D scene into LLM to carry out general 3D understanding tasks. However, by using 3D features constructed through projecting the 2D features of multi-view 2D images extracted by pre-trained 2D Vision-Language Models into 3D space, 3D-LLM struggles to directly capture the complex spatial relationships between objects and the structure of 3D scenes. [13,27] directly extract 3D features from the reconstructed 3D point cloud and support multi-modal visual prompts (text, images, 3D objects) in an MLLM. To alleviate the difficulty LLMs face in understanding complex 3D scenes, [22] choose to first explicitly segment the objects in the 3D scene and then perform multi-stage object-aware scenetest alignment to achieve 3D scene understanding. However, due to the lack of large-scale 3D-language alignment data and the intricate content of 3D scenes, although current MLLMs can achieve favorable performance in 3D scene understanding tasks, their localization performance is still significantly behind the 3D localization specialists. Our approach seeks to address this issue by introducing a 3D grounding model to enhance the localization capability of MLLMs." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.684, + 0.486, + 0.7 + ], + "angle": 0, + "content": "3 ScanReason Benchmark" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.713, + 0.517, + 0.728 + ], + "angle": 0, + "content": "3.1 3D Reasoning Grounding Task" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Given a 3D scene and an implicit free-form query, 3D reasoning grounding requires the model to predict the 3D bounding boxes of the target objects as well as the textual answers and explanations. As shown in Fig. 2, different from traditional 3D visual grounding, the queries of 3D reasoning grounding are implicit and complex, requiring strong reasoning, commonsense, and world knowledge. The number of target objects in 3D reasoning grounding is flexible and any object satisfying the requirements should be considered as the target object." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.65, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ScanReason" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.146, + 0.468, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.479, + 0.163, + 0.785, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.29, + 0.785, + 0.346 + ], + "angle": 0, + "content": "Fig. 2: The left side figure shows the overall of our ScanReason dataset. For each reasoning category, we designed different prompts to generate corresponding questions. And the right side figure shows the differences between the traditional 3D visual grounding task and our proposed 3D reasoning grounding task." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.379, + 0.391, + 0.394 + ], + "angle": 0, + "content": "3.2 Question Types" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.412, + 0.785, + 0.502 + ], + "angle": 0, + "content": "To comprehensively evaluate the 3D reasoning grounding abilities, we define 5 types of questions depending on which type of reasoning is required. Spatial reasoning and functional reasoning require a fundamental understanding of the 3D physical world, while logical reasoning, emotional reasoning, and safety reasoning are high-level reasoning skills built upon the two fundamental reasoning abilities to address user-centric real-world applications, as shown in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.504, + 0.785, + 0.564 + ], + "angle": 0, + "content": "Spatial Reasoning measures models' understanding of 3D spatial relationships among objects in 3D scene. It encompasses the ability to comprehend the layout of the 3D scene and the 3D location of objects within it, which could serve as the foundation for navigating or planning movements in the 3D environment." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.566, + 0.785, + 0.671 + ], + "angle": 0, + "content": "Functional Reasoning involves understanding and inferring the purpose, function, or affordance of objects within the 3D scene. For example, functional reasoning allows an embodied agent to recognize that a chair is for sitting, a lamp is for lighting, and a refrigerator is for storing food at low temperatures. Such understanding enables the embodied agent to assist users and to perform complex tasks more effectively (e.g., turning on a lamp when the room gets dark, or navigating to a refrigerator to fetch a drink)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.673, + 0.785, + 0.763 + ], + "angle": 0, + "content": "Logical Reasoning allows an embodied agent to not only understand its environment but also to interact with it in a goal-directed manner. For example, given a question shown in Fig. 2: \"If I'm cooking dinner in the kitchen, where is the nearest place to throw the rubbish?\", an agent needs to use such reasoning ability to infer the location of objects (in this question, a rubbish bin) based on their function and spatial relationships under the specific setting (the kitchen)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Emotional Reasoning plays a critical role in human-robot interaction, where the target objects are determined by understanding human emotions, preferences, and behavioral patterns. This ability makes the embodied agents more attuned to the emotional and psychological states of humans, allowing them to provide more personalized, empathetic, and contextually appropriate responses" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "C. Zhu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.176 + ], + "angle": 0, + "content": "and solutions, such as: \"I'm very sad, is there something that can make me feel happy?\" shown in 2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.785, + 0.237 + ], + "angle": 0, + "content": "Safety Reasoning focuses on preventing harm and ensuring the well-being of humans in the 3D environment. It requires the embodied agent to identify and assess the risk and make safety-aware decisions, such as: \"Where should first aid kits be placed for easy access but out of reach of children?\"" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.26, + 0.601, + 0.274 + ], + "angle": 0, + "content": "3.3 Automatic Data Annotation with GPT-4" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.285, + 0.784, + 0.375 + ], + "angle": 0, + "content": "We leverage the 3D scenes and bounding box annotations from the EmbodiedScan dataset [37] and apply GPT-4 [30] to automatically generate question-answer-location pairs for the five question types respectively. Specifically, we provide GPT-4 with the categories and bounding box locations of all objects in the scene, and ask GPT-4 to generate questions and answers with the target object ids of the provided objects. The details can be found in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.376, + 0.788, + 0.482 + ], + "angle": 0, + "content": "In total, our ScanReason dataset consists of 12929 complex reasoning question-answer-location pairs from 1456 scenes, which are split into 11455 training pairs and 1474 validation pairs. All 1474 validation question-answer pairs have been manually verified, including 342 spatial reasoning questions, 287 functional reasoning questions, 581 logical reasoning questions, 211 emotional questions, and 53 safety reasoning questions. We provide detailed statistics and more examples of our dataset in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.504, + 0.412, + 0.518 + ], + "angle": 0, + "content": "3.4 Evaluation Metric" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.529, + 0.784, + 0.604 + ], + "angle": 0, + "content": "To evaluate the accuracy of predicted objects and their locations for a flexible number of ground-truth objects, we follow the evaluation metric of 3D object detection tasks. Specifically, we adopt Acc@kIoU as our metric, where \\( k \\) is the threshold for the Intersection of Union (IoU) between positive predictions and ground truths. We evaluate the performance under \\( k = 0.25 \\) in our experiments." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.628, + 0.331, + 0.643 + ], + "angle": 0, + "content": "4 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.784, + 0.794 + ], + "angle": 0, + "content": "Solving the task of 3D reasoning grounding requires the synergization of the perception, reasoning, and grounding capability of the embodied agent. Intuitively, we can first conduct reasoning based on implicit instruction such as \"where should first aid kits be placed?\" and the visual environment. The reasoning process provides us with information about the rough location and semantics of the object we are looking for. Then keeping that information in mind, we look back to the 3D environment to precisely locate the object. For complex scenarios, alternate reasoning and looking back are required for multiple rounds until we obtain the final answer." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Inspired by this intuition, we propose ReGround3D, consisting of a visual-centric reasoning module and a 3D grounding module with geometry-enhanced look-back, as illustrated in Fig. 3. The visual-centric reasoning module (Sec. 4.1)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.65, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "ScanReason" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.284 + ], + "angle": 0, + "content": "performs joint reasoning of language instruction and visual scene, and predicts a special \\( \\langle \\mathsf{LOC} \\rangle \\) token representing the grounding information. The 3D grounding module (Sec. 4.2) looks back to the original 3D scene with comprehensive geometry information and fine-grained details. It takes the hidden embedding of the \\( \\langle \\mathsf{LOC} \\rangle \\) token containing grounding-related information from the 3D features and eventually predicts the 3D locations of the target objects. Furthermore, we propose a Chain-of-Grounding mechanism (CoG) (Sec. 4.3), i.e., a chain of interleaved reasoning and grounding steps, to further synergize the grounding and reasoning capability for the 3D reasoning grounding task, as illustrated in Fig. 4." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.308, + 0.473, + 0.323 + ], + "angle": 0, + "content": "4.1 Visual-Centric Reasoning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.335, + 0.786, + 0.456 + ], + "angle": 0, + "content": "Due to the complexity of the 3D scene and user instructions, particularly given the implicit intention behind the human instruction, ReGround3D starts with a visual-centric reasoning module that can perceive the scene, comprehend the human instructions, and conduct joint reasoning of 3D scene and instructions. We believe the reasoning process eventually implies the grounding intention, i.e., implicitly encodes the information indicating the target object to solve the task. Thus, we design the visual-centric reasoning module to predict grounding queries for localizing the target objects in the following 3D grounding module." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.457, + 0.788, + 0.635 + ], + "angle": 0, + "content": "Specifically, for simplicity, we leverage 3D-LLM to serve as the visual-centric reasoning module because of its strong reasoning abilities inherited from the LLM. Based on the BLIP2 architecture [26], 3D-LLM uses pre-trained image encoders to extract multi-view 2D image features and back-project them into 3D spaces. The visual features are encoded by the Q-Former to produce 32 tokens as the visual input to the LLM. By leveraging the pre-trained image encoders and the Q-Former, the visual tokens encode rich semantics but lack the 3D structures, spatial interactions, and fine-grained details. Therefore, instead of directly predicting the object locations by the 3D-LLM, we ask the 3D-LLM to predict the feature representation as the output of the reasoning process, and the predicted feature is further used to ground the target object in the grounding module." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.637, + 0.788, + 0.699 + ], + "angle": 0, + "content": "To enable the prediction of the grounding feature, we expand the original vocabulary of the 3D-LLM with a special \\( \\langle \\mathsf{L0C} \\rangle \\) token. The \\( \\langle \\mathsf{L0C} \\rangle \\) token is laden with the contextual scene and the target object information which can guide the 3D grounding module to accurately localize target objects." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.723, + 0.694, + 0.739 + ], + "angle": 0, + "content": "4.2 3D Grounding with Geometry-Enhanced Look-Back" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Once obtaining the token, ReGround3D extracts the last-layer embedding \\( h_{loc} \\) of the token and sends it into the 3D grounding module to predict the 3D bounding boxes. The 3D grounding module is devised with a \"look-back\" mechanism which allows the model to access the 3D geometry and fine-grained details from a 3D point cloud encoder. The fine-grained geometry-enhanced 3D visual features and \\( h_{loc} \\) are sent into a query selection module to retrieve the most" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "C. Zhu et al." + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.147, + 0.782, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.312, + 0.789, + 0.368 + ], + "angle": 0, + "content": "Fig. 3: The pipeline of ReGround3D. Given the 3D scene and human instruction, the visual reasoning module first performs joint 3D scene and instruction reasoning, and then guide the 3D grounding module to look-back the 3D scene and perform target object location." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.404, + 0.789, + 0.434 + ], + "angle": 0, + "content": "relevant object features. Those features are further decoded into 3D bounding boxes with the 3D box decoder." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.438, + 0.789, + 0.544 + ], + "angle": 0, + "content": "3D Visual Encoder Unlike the 2D image encoder used for 3D-LLM, the 3D visual encoder directly extracts features from 3D point clouds to capture more geometric and spatial information about the 3D structures and fine-grained details. The powerful 3D visual encoder which captures comprehensive geometry, structure, layout, and fine-grained information is critical to accurate 3D grounding. Subsequently, the 3D features \\( f_{scene} \\) produced by the 3D visual encoder and the grounding feature \\( h_{loc} \\) from 3D-LLM are sent to the query selection module." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.547, + 0.79, + 0.639 + ], + "angle": 0, + "content": "Query Selection Module We adopt a cross-attention mechanism, where we treat \\( f_{\\text{scene}} \\) as \\( Q \\) (Query), and \\( h_{\\text{loc}} \\) as both \\( K \\) (Key) and \\( V \\) (Value), to implicitly obtain a feature-level reasoning activation heatmap. During this scene look-back process, this module roughly locates scene features that have a high response to the token. We then select the \\( k \\) most relevant features as the object query \\( f_{\\text{query}} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.641, + 0.789, + 0.717 + ], + "angle": 0, + "content": "3D Box Decoder is a classical transformer decoder, which consists of M transformer decoder layers. In each decoder layer, the object queries \\( f_{query} \\) go through the text feature \\( h_{loc} \\) cross-attention layer and scene feature \\( f_{scene} \\) cross-attention layer. Finally, the prediction head takes the updated object queries as input and predicts the final 3D locations and matching score." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Discussion In comparison to previous works [11,13,15,21] that directly predicts the bounding boxes by the LLM, the extra grounding module has the following advantages: 1) Based on the MLLM reasoning results, it allows the grounding module to perceive the scene again and focus on the region under the implicit guidance of MLLM, adapted to the user queries and the reasoning results. 2) The scene representation perceived in the grounding module can be more precise and fine-grained, which is complementary to the visual-centric reasoning module. 3) The two-step reasoning-grounding pipeline is flexible and can generalized" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.65, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ScanReason" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.145, + 0.784, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.296, + 0.318, + 0.706, + 0.333 + ], + "angle": 0, + "content": "Fig. 4: Illustration of Chain-of-Grounding (CoG) Mechanism" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.36, + 0.785, + 0.391 + ], + "angle": 0, + "content": "to other types of predicting formats such as segmentation masks (by simply replacing the 3D box decoder with a 3D mask decoder)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.411, + 0.527, + 0.426 + ], + "angle": 0, + "content": "4.3 Chain-of-Grounding Mechanism" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.434, + 0.788, + 0.599 + ], + "angle": 0, + "content": "The existing design conducts the reasoning and grounding process sequentially, i.e., the reasoning process is finished before grounding. We argue that the grounding results can also facilitate the reasoning process, especially for those requiring spatial information. Thus, to further synergize the reasoning and grounding process, inspired by chain-of-thought (CoT) [38], we propose Chain-of-Grounding (CoG), which introduces a chain of interleaved steps of reasoning and grounding to find the targeted objects during inference, as shown in Fig. 4. Such a process allows the model to actively find relevant objects that help solve the problem, and then conduct reasoning with the assistance of the additional information of these relevant objects so that the model can more precisely find the target objects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.6, + 0.788, + 0.768 + ], + "angle": 0, + "content": "Specifically, given a question provided by users, CoG translates it into another question of finding the explicitly mentioned objects in the original question. The generated new question is sent into ReGround3D to ground the objects mentioned in the original question in the 3D scene with corresponding confidence scores. An object can be seen as a successfully located object when its confidence score is above the threshold, and the located object information could serve as explicit guidance for 3D-LLM in the next reasoning stage. As shown in Fig. 4, after obtaining the 3D information of objects explicitly mentioned in the original question, the located object information is inserted to update the question. The updated question is then sent to ReGround3D to perform reasoning and grounding to output the target object locations." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.787, + 0.419, + 0.803 + ], + "angle": 0, + "content": "4.4 Instruction Tuning" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Training Objective We use the pre-trained weights of 3D-LLM as the initialization for the visual-centric reasoning module. Except for freezing the 3D" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.359, + 0.128 + ], + "angle": 0, + "content": "C. Zhu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.228, + 0.145, + 0.773, + 0.16 + ], + "angle": 0, + "content": "Table 1: We list part of data templates used to train ReGround3D for each task." + }, + { + "type": "table", + "bbox": [ + 0.219, + 0.172, + 0.785, + 0.3 + ], + "angle": 0, + "content": "
Task NameText InstructionsOutput Type TemplatesExpected Output
3D Visual Grounding<scene> Here is a description Please answer the question only with outabout an object: "<expr>", whereis the object in the 3D scene?Please answer the question only with out-put 3D box prediction(s).It is <LOC>.
3D/Spatial Question Answering<scene> Answer the question Please answer the question only with text,do not output 3D box prediction(s).<question>Please answer the question only with text,do not output 3D box prediction(s).<answer>.
<scene><situation>, <question>Please answer the question only with text,do not output 3D box prediction(s).<answer>.
<scene><question>Please answer the question only with text,do not output 3D box prediction(s).<answer>.
3D Object Detection<scene> Where is the <category> Please answer the question only with out-in this 3D scene? Put 3D box prediction(s).Sure, <LOC>.
3D Reasoning Grounding<scene> Answer the question: Please answer the question with text andoutput 3D box prediction(s).Sure, <LOC>, <reason>
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.324, + 0.785, + 0.399 + ], + "angle": 0, + "content": "visual encoder pre-trained on [37], the rest of the parameters in the visual-centric reasoning module and 3D grounding module in our framework are trained in an end-to-end manner. The training supervision is a weighted sum of the next token prediction loss from 3D-LLM and the 3D detection loss from the 3D grounding module." + }, + { + "type": "equation", + "bbox": [ + 0.408, + 0.414, + 0.786, + 0.43 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda_ {\\text {t e x t}} \\mathcal {L} _ {\\text {t e x t}} + \\lambda_ {\\text {d e t}} \\mathcal {L} _ {\\text {d e t}} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.434, + 0.523, + 0.45 + ], + "angle": 0, + "content": "The 3D detection loss is defined following:" + }, + { + "type": "equation", + "bbox": [ + 0.363, + 0.464, + 0.786, + 0.481 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {d e t}} = \\lambda_ {I O U} \\mathcal {L} _ {I O U} + \\lambda_ {\\text {c o n t r a s t}} \\mathcal {L} _ {\\text {c o n t r a s t}} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.498, + 0.788, + 0.696 + ], + "angle": 0, + "content": "Instruction Tuning Dataset We load the pre-trained weights of 3D-LLM and the 3D visual encoder, and finetune the LoRA of 3D-LLM, the query selection module, and the 3D box decoder with an instruction tuning dataset. To construct the instruction tuning dataset, we reformulate the data annotations from existing 3D datasets into question-answer or question-answer-bounding-box pairs. Specifically, the 3D visual grounding data from ScanRefer [7], SR3D, NR3D [1] and the 3D object detection data from EmbodiedScan [37] are formulated into question-answer-bbox pairs, and the 3D/spatial question answering data from SR3D [1], CLEVR3D [41], SQA3D [29] are formulated into question-answer pairs without bounding boxes. The information shown in Tab. 1 illustrates how we unify the instruction and output with task-specific templates. More details can be found in the Appendix. The reformulated data combined with our proposed ScanReason dataset, serve as the instruction tuning dataset." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.715, + 0.367, + 0.733 + ], + "angle": 0, + "content": "5 Experiment" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.743, + 0.457, + 0.759 + ], + "angle": 0, + "content": "5.1 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.785, + 0.841 + ], + "angle": 0, + "content": "Network Architecture For the 3D grounding module, we adopt the pretrained point cloud encoder as the 3D visual encoder. During the training stage, We use LoRA to efficiently finetune the 3D-LLM to preserve the original 3D scene understanding capability and reduce the computation costs. The number of object queries \\( k \\) in the query selection module is set to 256." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.65, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ScanReason" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.146, + 0.788, + 0.236 + ], + "angle": 0, + "content": "Training Parameters The training is done on 8 NVIDIA A100 GPUs. We adopt the AdamW optimizer with a learning rate of 3e-4 and use a learning rate scheduler WarmupDecayLR with the warmup steps of 100. The total batch size is set to be 16. The loss weight parameters \\(\\lambda_{text}\\) and \\(\\lambda_{det}\\) in total loss \\(\\mathcal{L}\\) are set to 1.0 and 1.0, respectively, and the weight \\(\\lambda_{IOU}\\) and \\(\\lambda_{contrast}\\) in \\(\\mathcal{L}_{\\mathrm{det}}\\) are set to 1.0 and 1.0." + }, + { + "type": "title", + "bbox": [ + 0.219, + 0.265, + 0.428, + 0.28 + ], + "angle": 0, + "content": "5.2 Results Comparison" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.295, + 0.786, + 0.581 + ], + "angle": 0, + "content": "Evaluation on 3D Visual Grounding In order to verify the superiority of our model in grounding ability and facilitate comparison between current models, we report the explicit grounding performance on the existing 3D visual grounding task. Since the evaluation settings of Nr3D and Sr3D [1] are based on ground-truth object proposals, while ScanRefer [7] requires models to output 3D bounding boxes, we choose ScanRefer as our benchmark for comparison. We divide the existing methods into two categories, one is the grounding model designed specifically for the 3D visual grounding task, and the other is generalist MLLMs which can understand a variety of 3D vision-language tasks. The original 3D-LLM embeds 3D locations in the vocabularies and represents the grounded 3D bounding boxes by a sequence of discrete location tokens. However, since the fine-tuned model of 3D-LLM on ScanRefer and related location tokens implementation are not accessible, we adapt 3D-LLM to directly output 3D numerical coordinates representing 3D bounding boxes by fine-tuning the pre-trained model on our reformulated 3D visual grounding data, denotes as 3D-LLM (vg). As shown in Tab. 2, current generalist MLLM models still lag behind the specialist models in terms of grounding ability. By incorporating the 3D grounding module into MLLM, ReGround3D shows the SOTA performance on the traditional 3D visual grounding task." + }, + { + "type": "table_caption", + "bbox": [ + 0.218, + 0.61, + 0.786, + 0.636 + ], + "angle": 0, + "content": "Table 2: Results on 3D visual grounding task among ReGround3D (ours) and existing methods." + }, + { + "type": "table", + "bbox": [ + 0.334, + 0.651, + 0.673, + 0.807 + ], + "angle": 0, + "content": "
TypeMethodsAcc@0.25Acc@0.5
SpecialistsScanRefer [7]37.324.3
MVT [23]40.833.3
3DVG-Trans [48]45.934.5
ViL3DRel [12]47.937.7
BUTD-DETR [24]52.239.8
L3Det [50]52.840.2
Generalized MLLMsLLM-Grounder [42]17.15.3
3D-LLM [21]30.3-
3D-LLM(vg) [21]33.128.7
Chat3D-v2 [22]35.930.4
oursReGround3D53.141.1
" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.359, + 0.127 + ], + "angle": 0, + "content": "C. Zhu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.785, + 0.174 + ], + "angle": 0, + "content": "Table 3: Results (Acc@0.25) on 3D reasoning grounding task among ReGround3D (ours) and existing methods." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.185, + 0.787, + 0.284 + ], + "angle": 0, + "content": "
MethodsLLMSpatial FunctionalLogicalEmotional SafetyOverall
Mask3D [32] + InternLM2-7B [33]InternLM2-7B10.3436.129.988.218.9914.86
3D-LLM(vg) [21]FlanT5XL-3B18.3117.4210.978.126.3313.29
Chat3D-v2 [22]Vicuna-7B20.2118.3911.327.989.8814.98
ReGround3D*FlanT5XL-3B30.7629.818.6719.2217.1223.27
ReGround3DFlanT5XL-3B32.9836.2326.9923.1222.9828.98
ReGround3D (CoG)FlanT5XL-3B34.7136.7929.1124.0323.2130.62
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.312, + 0.789, + 0.659 + ], + "angle": 0, + "content": "Evaluation on 3D Reasoning Grounding The various visual grounding models rely on explicit text-object alignment in the input object expression to achieve localization, which falls to be applied to 3D reasoning grounding task. We performed a comparison between our proposed method ReGround3D and existing MLLM methods, including 3D-LLM(vg) and Chat3D-v2 [22]. Besides, inspired by Chat-3D v2, which first segments objects, then equips them with unique object identifiers to conduct effective object grounding, we set up a LLM-based 3D reasoning grounding baseline: We first segment the objects from the scene using a 3D instance segmentor [32], then convert the segmented object information including their categories and 3D bounding boxes into text as input of LLM (InternLM2-7B [33]). Besides, to better validate the performance of our model and ensure a fair comparison, we remove the ScanReason dataset from the training data, denoted as ReGround3D*. As shown in Tab. 3, the LLM-based reasoning method (Mask3D [32] + InternLM2-7B [33]) possesses a very strong functional reasoning ability, but struggles to understand of 3D spatial relationship. Additionally, irrespective of whether ScanReason is used in training, our model significantly outperforms the existing MLLMs. By synergizing the reasoning and grounding process utilizing the Chain-of-Grounding (CoG) mechanism, the 3D reasoning grounding performance of ReGround3D can be further improved (from 28.98 to 30.62), especially on the spatial reasoning and logical reasoning questions. The qualitative results comparison shown in Fig. 5 demonstrates the superiority of our method in reasoning human complex instruction based on the 3D scene." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.684, + 0.389, + 0.699 + ], + "angle": 0, + "content": "5.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.71, + 0.786, + 0.741 + ], + "angle": 0, + "content": "In this section, we conduct an extensive ablation study to verify the effectiveness of each component." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Effectiveness of 3D Grounding Module In order to more comprehensively verify the effectiveness of the 3D grounding module we proposed, we step by step verify the performance changes from 3D-LLM to ReGround3D. Apart from the 3D-LLM(vg), we fine-tuned the 3D-LLM model respectively on full reformulated existing data and all instruction tuning data, including ScanReason, denoted as" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.65, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ScanReason" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.239, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.147, + 0.488, + 0.158 + ], + "angle": 0, + "content": "It's so cold outside, Where could I find my thick coat?" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.164, + 0.774, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.263, + 0.239, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.243, + 0.266, + 0.43, + 0.278 + ], + "angle": 0, + "content": "Where can I find paper to wipe my hands?" + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.283, + 0.784, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.373, + 0.239, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.376, + 0.51, + 0.388 + ], + "angle": 0, + "content": "It's dark here, can you find something to light up the room?" + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.394, + 0.784, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.5, + 0.789, + 0.557 + ], + "angle": 0, + "content": "Fig. 5: Visualization comparison of 3D reasoning grounding capability between ReGround3D and 3D-LLM. Our method could achieve much more accurate grounding results which satisfy the implicit question intention, and give the corresponding explanation at the same time. More illustrations are given in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.596, + 0.789, + 0.673 + ], + "angle": 0, + "content": "3D-LLM (full) and 3D-LLM (full+sr). All the tokens in answers of the data used for fine-tuning 3D-LLM have been converted to the numerical coordinates of the corresponding boxes. The results in Tab. 4a show that with the same tuning dataset, our ReGround3D achieves far superior performance to 3D-LLM (28.98 vs. 19.21) by introducing the 3D grounding module." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Effectiveness of Instruction Tuning Dataset Tab. 4b showcases the impact of different training data types on 3D visual grounding performance. 3D Object Detection (3D OD) provides the explicit 3D semantic category and visual alignment whereas 3D Question Answering (3D QA) data injects the basic 3D scene understanding ability into the model, which has certain benefits for the visual grounding ability of the model. In addition, we find that training with 3D reasoning grounding dataset can further improve the performance on 3D visual grounding." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "C. Zhu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.22, + 0.145, + 0.782, + 0.16 + ], + "angle": 0, + "content": "Table 4: Ablation study on effectiveness of 3D grounding module and training data" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.184, + 0.477, + 0.216 + ], + "angle": 0, + "content": "(a) When using the same tuning dataset, ReGround3D achieves far better performance on ScanReason than 3D-LLM." + }, + { + "type": "text", + "bbox": [ + 0.483, + 0.184, + 0.782, + 0.216 + ], + "angle": 0, + "content": "(b) Ablation study on training data. We evaluate through the metric of accuracy on the val set of the ScanRefer dataset." + }, + { + "type": "table", + "bbox": [ + 0.233, + 0.228, + 0.457, + 0.312 + ], + "angle": 0, + "content": "
MethodsAcc@0.25
3D-LLM(vg) [21]13.29
3D-LLM(full) [21]15.31
3D-LLM(full+sr) [21]19.21
ReGround3D(full+sr)28.98
" + }, + { + "type": "table", + "bbox": [ + 0.523, + 0.228, + 0.738, + 0.32 + ], + "angle": 0, + "content": "
DatasetAcc@0.25Acc@0.5
VGODVQARD
✓✓19.314.2
✓✓48.737.6
✓✓49.238.1
✓✓✓51.839.4
✓✓✓✓53.141.1
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.348, + 0.788, + 0.53 + ], + "angle": 0, + "content": "Discussion of CoG Mechanism While the CoG Mechanism boosts the performance with interleaved reasoning and grounding steps during inference, it uses the relevant object information explicitly presented in the question to help find the target objects. A natural question arises: if we input the information of all the objects instead of relevant objects into ReGround3D during CoG, will this make the model more accurately find the target objects? We first use the existing 3D Segmentor [32] to segment all objects in the scene, then update the original question using all the object 3D bounding boxes information according to the template in Sec. 4.3 and send into ReGround3D. However, experiments show that using all the object information will instead reduce the 3D reasoning grounding performance from 28.98 to 27.67. One possible reason could be the model's attention is dispersed by too many irrelevant objects." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.553, + 0.36, + 0.569 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.585, + 0.789, + 0.827 + ], + "angle": 0, + "content": "This paper introduces a new 3D vision language learning task: 3D reasoning grounding, which requires the model to perform active reasoning over complex and implicit human instruction, localize the target objects, and give corresponding explanations. Furthermore, we propose ScanReason, a new dataset and benchmark to further unlock and thoroughly evaluate the 3D reasoning grounding capability. Based on this dataset, we propose a novel approach: ReGround3D, which utilizes the strong reasoning capability of MLLM guiding the 3D grounding module to obtain accurate object locations, and a Chain of Grounding (CoG) mechanism is presented to further boost the performance with interleaved reasoning and grounding steps during inference. We believe that our work will further the natural interaction between embodied agents and humans in open 3D environments. For the current ScanReason benchmark, we find that the questions in three high-level 3D reasoning categories may have overlaps. For a certain reasoning question, similar questions may appear in one or two other categories. We leave the problem as a future challenge for better reasoning grounding ability evaluation." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.65, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "ScanReason" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.788, + 0.225 + ], + "angle": 0, + "content": "This work is supported in part by HKU Startup Fund, HKU Seed Fund for Basic Research, HKU Seed Fund for Translational and Applied Research, HKU IDS research Seed Fund, and HKU Fintech Academy R&D Funding." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.249, + 0.324, + 0.265 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.282, + 0.786, + 0.338 + ], + "angle": 0, + "content": "1. Achlioptas, P., Abdelreheem, A., Xia, F., Elhoseiny, M., Guibas, L.: Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16. pp. 422-440. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.34, + 0.787, + 0.38 + ], + "angle": 0, + "content": "2. Azuma, D., Miyanishi, T., Kurita, S., Kawanabe, M.: Scanqa: 3d question answering for spatial scene understanding. In: proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 19129-19139 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.381, + 0.787, + 0.422 + ], + "angle": 0, + "content": "3. Bakr, E.M., Ayman, M., Ahmed, M., Slim, H., Elhoseiny, M.: Cot3dref: Chain-of-thoughts data-efficient 3d visual grounding. arXiv preprint arXiv:2310.06214 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.423, + 0.787, + 0.463 + ], + "angle": 0, + "content": "4. Baroni, M.: Linguistic generalization and compositionality in modern artificial neural networks. Philosophical Transactions of the Royal Society B 375(1791), 20190307 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.465, + 0.787, + 0.519 + ], + "angle": 0, + "content": "5. Cai, D., Zhao, L., Zhang, J., Sheng, L., Xu, D.: 3djcg: A unified framework for joint dense captioning and visual grounding on 3d point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16464-16473 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.521, + 0.787, + 0.56 + ], + "angle": 0, + "content": "6. Chen, C., Qin, R., Luo, F., Mi, X., Li, P., Sun, M., Liu, Y.: Position-enhanced visual instruction tuning for multimodal large language models. arXiv preprint arXiv:2308.13437 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.562, + 0.787, + 0.617 + ], + "angle": 0, + "content": "7. Chen, D.Z., Chang, A.X., Niefer, M.: Scanrefer: 3d object localization in rgb-d scans using natural language. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XX. pp. 202-221. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.618, + 0.787, + 0.659 + ], + "angle": 0, + "content": "8. Chen, D.Z., Wu, Q., Nießner, M., Chang, A.X.: D3net: A speaker-listener architecture for semi-supervised dense captioning and visual grounding in rgb-d scans. arXiv preprint arXiv:2112.01551 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.66, + 0.787, + 0.701 + ], + "angle": 0, + "content": "9. Chen, F., Han, M., Zhao, H., Zhang, Q., Shi, J., Xu, S., Xu, B.: X-LLM: Bootstrapping advanced large language models by treating multi-modalities as foreign languages. arXiv preprint arXiv:2305.04160 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.702, + 0.787, + 0.756 + ], + "angle": 0, + "content": "10. Chen, J., Zhu, D., Shen, X., Li, X., Liu, Z., Zhang, P., Krishnamoorthi, R., Chandra, V., Xiong, Y., Elhoseiny, M.: Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.757, + 0.787, + 0.799 + ], + "angle": 0, + "content": "11. Chen, K., Zhang, Z., Zeng, W., Zhang, R., Zhu, F., Zhao, R.: Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "12. Chen, S., Guhur, P.L., Tapaswi, M., Schmid, C., Laptev, I.: Language conditioned spatial relation reasoning for 3d object grounding. Advances in Neural Information Processing Systems 35, 20522-20535 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.282, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "C. Zhu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "13. Chen, S., Chen, X., Zhang, C., Li, M., Yu, G., Fei, H., Zhu, H., Fan, J., Chen, T.: Ll3da: Visual interactive instruction tuning for omni-3d understanding, reasoning, and planning. arXiv preprint arXiv:2311.18651 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.231 + ], + "angle": 0, + "content": "14. Chen, S., Zhu, H., Chen, X., Lei, Y., Yu, G., Chen, T.: End-to-end 3d dense captioning with vote2cap-detr. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11124-11133 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.232, + 0.785, + 0.259 + ], + "angle": 0, + "content": "15. Chen, T., Saxena, S., Li, L., Fleet, D.J., Hinton, G.: Pix2seq: A language modeling framework for object detection. arXiv preprint arXiv:2109.10852 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.26, + 0.785, + 0.3 + ], + "angle": 0, + "content": "16. Chen, Z., Gholami, A., Nießner, M., Chang, A.X.: Scan2cap: Context-aware dense captioning in rgb-d scans. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3193-3203 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.301, + 0.785, + 0.342 + ], + "angle": 0, + "content": "17. Chen, Z., Hu, R., Chen, X., Nießner, M., Chang, A.X.: Unit3d: A unified transformer for 3d dense captioning and visual grounding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 18109-18119 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.343, + 0.785, + 0.383 + ], + "angle": 0, + "content": "18. Chung, H.W., Hou, L., Longpre, S., Zoph, B., Tay, Y., Fedus, W., Li, E., Wang, X., Dehghani, M., Brahma, S., et al.: Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.384, + 0.785, + 0.424 + ], + "angle": 0, + "content": "19. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.425, + 0.785, + 0.466 + ], + "angle": 0, + "content": "20. Han, J., Zhang, R., Shao, W., Gao, P., Xu, P., Xiao, H., Zhang, K., Liu, C., Wen, S., Guo, Z., et al.: Imagebind-llm: Multi-modality instruction tuning. arXiv preprint arXiv:2309.03905 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.467, + 0.785, + 0.508 + ], + "angle": 0, + "content": "21. Hong, Y., Zhen, H., Chen, P., Zheng, S., Du, Y., Chen, Z., Gan, C.: 3d-llm: Injecting the 3d world into large language models. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.509, + 0.785, + 0.55 + ], + "angle": 0, + "content": "22. Huang, H., Wang, Z., Huang, R., Liu, L., Cheng, X., Zhao, Y., Jin, T., Zhao, Z.: Chat-3d v2: Bridging 3d scene and large language models with object identifiers. arXiv preprint arXiv:2312.08168 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.55, + 0.785, + 0.591 + ], + "angle": 0, + "content": "23. Huang, S., Chen, Y., Jia, J., Wang, L.: Multi-view transformer for 3d visual grounding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15524-15533 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.592, + 0.785, + 0.647 + ], + "angle": 0, + "content": "24. Jain, A., Gkanatsios, N., Mediratta, I., Fragkiadaki, K.: Bottom up top down detection transformers for language grounding in images and point clouds. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVI. pp. 417-433. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.648, + 0.785, + 0.674 + ], + "angle": 0, + "content": "25. Jiao, Y., Chen, S., Jie, Z., Chen, J., Ma, L., Jiang, Y.G.: More: Multi-order relation mining for dense captioning in 3d scenes. arXiv preprint arXiv:2203.05203 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.675, + 0.785, + 0.715 + ], + "angle": 0, + "content": "26. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.716, + 0.785, + 0.758 + ], + "angle": 0, + "content": "27. Li, M., Chen, X., Zhang, C., Chen, S., Zhu, H., Yin, F., Yu, G., Chen, T.: M3dbench: Let's instruct large models with multi-modal 3d prompts. arXiv preprint arXiv:2312.10763 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.759, + 0.785, + 0.812 + ], + "angle": 0, + "content": "28. Luo, J., Fu, J., Kong, X., Gao, C., Ren, H., Shen, H., Xia, H., Liu, S.: 3d-sps: Single-stage 3d visual grounding via referred point progressive selection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16454-16463 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "29. Ma, X., Yong, S., Zheng, Z., Li, Q., Liang, Y., Zhu, S.C., Huang, S.: Sqa3d: Situated question answering in 3d scenes. arXiv preprint arXiv:2210.07474 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.65, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ScanReason" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.148, + 0.503, + 0.162 + ], + "angle": 0, + "content": "30. OpenAI: Gpt-4 technical report (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.162, + 0.787, + 0.203 + ], + "angle": 0, + "content": "31. Roh, J., Desingh, K., Farhadi, A., Fox, D.: Languagerefer: Spatial-language model for 3d visual grounding. In: Conference on Robot Learning. pp. 1046-1056. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.786, + 0.258 + ], + "angle": 0, + "content": "32. Schult, J., Engelmann, F., Hermans, A., Litany, O., Tang, S., Leibe, B.: Mask3d: Mask transformer for 3d semantic instance segmentation. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 8216-8223. IEEE (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.786, + 0.286 + ], + "angle": 0, + "content": "33. Team, I.: Internl'm: A multilingual language model with progressively enhanced capabilities (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.786, + 0.327 + ], + "angle": 0, + "content": "34. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.786, + 0.355 + ], + "angle": 0, + "content": "35. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.786, + 0.383 + ], + "angle": 0, + "content": "36. Wang, H., Zhang, C., Yu, J., Cai, W.: Spatiality-guided transformer for 3d dense captioning on point clouds. arXiv preprint arXiv:2204.10688 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.786, + 0.424 + ], + "angle": 0, + "content": "37. Wang, T., Mao, X., Zhu, C., Xu, R., Lyu, R., Li, P., Chen, X., Zhang, W., Chen, K., Xue, T., et al.: Embodiedscan: A holistic multi-modal 3d perception suite towards embodied ai. arXiv preprint arXiv:2312.16170 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.786, + 0.466 + ], + "angle": 0, + "content": "38. Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q.V., Zhou, D., et al.: Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems 35, 24824-24837 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.786, + 0.508 + ], + "angle": 0, + "content": "39. Wu, Y., Cheng, X., Zhang, R., Cheng, Z., Zhang, J.: Eda: Explicit textdecoupling and dense alignment for 3d visual and language learning. arXiv preprint arXiv:2209.14941 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.508, + 0.786, + 0.55 + ], + "angle": 0, + "content": "40. Xu, R., Wang, X., Wang, T., Chen, Y., Pang, J., Lin, D.: Pointllm: Empowering large language models to understand point clouds. arXiv preprint arXiv:2308.16911 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.786, + 0.591 + ], + "angle": 0, + "content": "41. Yan, X., Yuan, Z., Du, Y., Liao, Y., Guo, Y., Li, Z., Cui, S.: Clevr3d: Compositional language and elementary visual reasoning for question answering in 3d real-world scenes. arXiv preprint arXiv:2112.11691 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.591, + 0.786, + 0.632 + ], + "angle": 0, + "content": "42. Yang, J., Chen, X., Qian, S., Madaan, N., Iyengar, M., Fouhey, D.F., Chai, J.: Llm-grounder: Open-vocabulary 3d visual grounding with large language model as an agent. arXiv preprint arXiv:2309.12311 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.786, + 0.674 + ], + "angle": 0, + "content": "43. Yang, Z., Zhang, S., Wang, L., Luo, J.: Sat: 2d semantics assisted training for 3d visual grounding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1856-1866 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.674, + 0.786, + 0.715 + ], + "angle": 0, + "content": "44. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.786, + 0.743 + ], + "angle": 0, + "content": "45. Ye, S., Chen, D., Han, S., Liao, J.: 3d question answering. IEEE Transactions on Visualization and Computer Graphics (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.786, + 0.799 + ], + "angle": 0, + "content": "46. Yuan, Z., Yan, X., Liao, Y., Guo, Y., Li, G., Cui, S., Li, Z.: X-trans2cap: Cross-modal knowledge transfer using transformer for 3d dense captioning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8563-8573 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.786, + 0.841 + ], + "angle": 0, + "content": "47. Zhang, S., Roller, S., Goyal, N., Artetxe, M., Chen, M., Chen, S., Dewan, C., Diab, M., Li, X., Lin, X.V., et al.: Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.359, + 0.127 + ], + "angle": 0, + "content": "C. Zhu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "48. Zhao, L., Cai, D., Sheng, L., Xu, D.: 3dvg-transformer: Relation modeling for visual grounding on point clouds. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2928-2937 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.217 + ], + "angle": 0, + "content": "49. Zhong, Y., Xu, L., Luo, J., Ma, L.: Contextual modeling for 3d dense captioning on point clouds. arXiv preprint arXiv:2210.03925 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.787, + 0.244 + ], + "angle": 0, + "content": "50. Zhu, C., Zhang, W., Wang, T., Liu, X., Chen, K.: Object2scene: Putting objects in context for open-vocabulary 3d detection. arXiv preprint arXiv:2309.09456 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.787, + 0.285 + ], + "angle": 0, + "content": "51. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.787, + 0.327 + ], + "angle": 0, + "content": "52. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.787, + 0.37 + ], + "angle": 0, + "content": "53. Zhu, Z., Ma, X., Chen, Y., Deng, Z., Huang, S., Li, Q.: 3d-vista: Pre-trained transformer for 3d vision and text alignment. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2911-2921 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.37 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/ScanReason_ Empowering 3D Visual Grounding with Reasoning Capabilities/c0158ad2-ca00-4982-8b2e-6cbbd910c1e7_origin.pdf b/2024/ScanReason_ Empowering 3D Visual Grounding with Reasoning Capabilities/c0158ad2-ca00-4982-8b2e-6cbbd910c1e7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6d4a5b719926d0e952774d06ba508628f48c9548 --- /dev/null +++ b/2024/ScanReason_ Empowering 3D Visual Grounding with Reasoning Capabilities/c0158ad2-ca00-4982-8b2e-6cbbd910c1e7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9840483747c2994e8bc8dd64b76329b841621ab580fb5ce23c56b86ec44fd150 +size 4616010 diff --git a/2024/ScanReason_ Empowering 3D Visual Grounding with Reasoning Capabilities/full.md b/2024/ScanReason_ Empowering 3D Visual Grounding with Reasoning Capabilities/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6e1ff28bf4093ec5c1a0efaa0a069fbd9b6a0770 --- /dev/null +++ b/2024/ScanReason_ Empowering 3D Visual Grounding with Reasoning Capabilities/full.md @@ -0,0 +1,295 @@ +# ScanReason: Empowering 3D Visual Grounding with Reasoning Capabilities + +Chenming Zhu $^{1,2}$ , Tai Wang $^{2}$ , Wenwei Zhang $^{2}$ , Kai Chen $^{2}$ , and Xihui Liu $^{1\dagger}$ + +1 The University of Hong Kong +2 Shanghai AI Laboratory +chairzhu@connect.hku.hk +https://zcmax.github.io/projects/ScanReason + +Abstract. Although great progress has been made in 3D visual grounding, current models still rely on explicit textual descriptions for grounding and lack the ability to reason human intentions from implicit instructions. We propose a new task called 3D reasoning grounding and introduce a new benchmark ScanReason which provides over 10K question-answer-location pairs from five reasoning types that require the synchronization of reasoning and grounding. We further design our approach, ReGround3D, composed of the visual-centric reasoning module empowered by Multi-modal Large Language Model (MLLM) and the 3D grounding module to obtain accurate object locations by looking back to the enhanced geometry and fine-grained details from the 3D scenes. A chain-of-grounding mechanism is proposed to further boost the performance with interleaved reasoning and grounding steps during inference. Extensive experiments on the proposed benchmark validate the effectiveness of our proposed approach. + +Keywords: 3D reasoning grounding $\cdot$ 3D visual grounding $\cdot$ Multi-modal large language models + +# 1 Introduction + +Understanding and reasoning in the 3D visual world is critical for applications such as robotics and AR, where embodied agents are expected to understand the 3D layout and predict the 3D locations of objects based on human instructions. The example in Fig. 1 demonstrates a scenario where the question can only be solved with a comprehensive understanding of the 3D scene and joint reasoning of the question and the 3D environment. However, current 3D visual grounding models [24,39,43,50] trained on [1,7] localize objects based on explicit descriptions of the object category, attribute, and 3D spatial relationships, and lack the ability to reason the user intentions and predict object locations with implicit human instructions such as "I am thirsty, can I have something to drink?" + +To bridge the aforementioned gap and to push the boundaries of what embodied agents can understand and how they can interact with the 3D world, we + +Question: I'm looking for a comfortable spot to watch TV immersively, where is the best place for me to lie down and watch TV? + +![](images/4c561b3f0105a6a0b0268ea384fa4a072cd3d46107e710ee8cfb9fad6d2e5abe.jpg) +GPT-4 + +The best spot for lying down to watch TV immersively is on a spacious, cushioned sectional sofa or an adjustable reclining bed in a room with a large screen TV mounted at an eye-level view from your lying position, ensuring comfort and optimal viewing angles. Ensure the room is dimly lit or has blackout curtains to enhance screen visibility and reduce glare. + +![](images/4a95c7dbd5de217e17e8a7e1b54d70cd020b4df923a278dce65612d360db09f6.jpg) + +![](images/ce635d6fa7f15884bccb2f950a6b0d5e412f2b3aa27caec0eedee6685f12390f.jpg) + +The best spot to lie down and watch TV immersively would be on the bed facing the TV in the cabinet; it provides a direct line of sight and the comfort of lying down. + +![](images/aa38eb9871b82966b6329b67f248d43cfc47ccb26d6c4f553de5716685477204.jpg) +GPT-4V + +![](images/0e30da9813c4b3ec62ff18a22440b8ac36f716ee37e9d7ca1bd6d6557adb0cdf.jpg) +ReGround3D + +![](images/97b7d2df696c624b87e1e831451244f95ed889ec992544662311448d0163f013.jpg) +Fig. 1: For an embodied agent, they not only need to be able to understand the 3D environment and complex human instructions but also localize the target objects for interaction and navigation. Although GPT-4 (GPT-4V) have strong text (multi-modal) reasoning abilities, they lack the ability to directly perceive the 3D scene, understand the 3D spatial relationships, and output corresponding target object locations. Instead, our proposed method ReGround3D has the 3D perception, reasoning, and grounding capabilities in the real-world 3D environment. + +For an optimal TV viewing experience, I recommend sitting on the couch. It is oriented towards the TV and closer to the TV compared with the bed, and it's large enough to lie down. + +![](images/53f86f8df84306de5070778a32a6056b0d90a3b96eafe9bba48d951c14a45520.jpg) + +propose a new task of 3D reasoning grounding and introduce a new benchmark named ScanReason. The task requires the model to conduct joint reasoning on the question and the 3D environment before predicting the 3D locations of target objects. We define five categories of 3D reasoning: spatial reasoning, functional reasoning, logical reasoning, emotional reasoning, and safety reasoning. The first two categories focus on the fundamental understanding of the 3D physical world, while the last three categories are built upon fundamental abilities to address user-centric real-world challenges. The benchmark comprises more than 10K question-answer-3D bounding box pairs from 2K scenes belonging to the five reasoning types mentioned above. The GPT-4-assisted data annotation process largely increases the efficiency of curating such a dataset. + +We propose ReGround3D as an initial attempt to the new task of 3D Reasoning Grounding. Intuitively, for 3D grounding with implicit instructions, we need to first conduct reasoning on the language instructions and the coarse visual environment. Then with the idea of which object we want to find in mind, we look back to the 3D scene and ground the target object. For complex instructions, we may need to alternate the reasoning and look-back process for multiple iterations. Inspired by this intuition, our framework is composed of a visual-centric reasoning module and 3D grounding with geometry-enhanced look-back module, with a Chain-of-Grounding mechanism during inference to + +alternately conduct reasoning and grounding for multiple rounds. Specifically, the visual-centric reasoning module conducts joint reasoning of the 3D scene and instructions with an MLLM. This module predicts a special token representing the semantic and location information of the target object, which is used for the grounding module. The 3D grounding module uses the output token embedding from the previous reasoning module to locate the target object by looking back at the fine-grained 3D scene representation. Unlike previous MLLMs attempting to directly predict bounding box coordinates, our look-back mechanism enables the model to capture more comprehensive 3D geometry and fine-grained object details for accurate 3D grounding. The Chain-of-Grounding mechanism is proposed to synergize reasoning and grounding, which allows multiple rounds alternating between reasoning and grounding during inference. + +In summary, our contributions are threefold: 1) We propose a new task of 3D reasoning grounding which requires the model to synergize reasoning and grounding. We further introduce a new benchmark ScanReason, which comprises five reasoning types (spatial reasoning, functional reasoning, logical reasoning, emotional reasoning, and safety reasoning) for the task of 3D reasoning grounding in 3D scenes. 2) We design a new framework ReGround3D with a visual-centric reasoning module and a 3D grounding module with geometry-enhanced look-back. We further introduce a Chain-of-Grounding mechanism to boost the 3D reasoning grounding ability with a chain of interleaved reasoning and grounding steps. 3) Extensive experiments demonstrate the effectiveness of the our ReGround3D on the ScanReason benchmark for 3D reasoning grounding. + +# 2 Related Work + +3D Vision and Language Learning 3D Vision-language learning (3D-VL) is garnering increasing attention, with many 3D-VL tasks focusing on how to connect the 3D world with natural language. Among them, 3D Question Answering (3D QA) [2,45] aims to enable models to provide text answers based on natural language questions. Situation Question Answering in 3D Scenes (SQA3D) [29] requires an agent to first understand its location based on a text description, then provide a reasonable answer based on the surrounding environment, which can be seen as an extension of 3D QA in the embodied AI area. 3D Visual Grounding [1,4,7,28,31,39] demands that models identify and locate target objects in a 3D scene based on given descriptions, outputting the objects' coordinates and 3D bounding boxes. These descriptions usually explicitly rely on the objects' attributes and their spatial relationships. 3D Dense Captioning [5,8,14,16,25,36,46,49] requires models to output a series of object coordinates and corresponding scene-based descriptions based on a given scene. Different from these 3D-VL tasks, the questions in 3D reasoning grounding could be more implicit and complex. + +3D Visual Grounding The task of 3D visual grounding is aimed at localizing the objects that are explicitly referred to by free-form guided language + +expressions in the 3D scene. Inspired by the success of transformers in natural language processing, recent 3D visual grounding approaches [12,17,24,39,50,53] have started to adopt transformer [35] architectures for handling the complex relationships between language descriptions and 3D visual data. These methods leverage the self-attention mechanism of transformers to dynamically weigh the importance of different parts of the input data, facilitating a more effective grounding of textual descriptions in the 3D environment. Recent method [3] proposes a Chain-of-Thoughts module that predicts a chain of anchor objects that are subsequently utilized to localize the final target object. Compared with 3D visual grounding, our proposed 3D reasoning grounding requires the model to reason the complex question, ground target objects, and give the explanation at the same time. + +Multi-modal Large Language Models Recently, there has been an increasing effort to extend the powerful complex reasoning and world knowledge capabilities of LLMs [18,34,47] to other modalities [6,9,10,19,26,44,51,52]. Among these works, some have aimed to enable LLMs to understand the 3D world. [20, 40] focus on delving into LLMs' ability to comprehend 3D objects, which can not be directly applied to 3D scenes. 3D-LLM [21] is the pioneering work that incorporates the 3D scene into LLM to carry out general 3D understanding tasks. However, by using 3D features constructed through projecting the 2D features of multi-view 2D images extracted by pre-trained 2D Vision-Language Models into 3D space, 3D-LLM struggles to directly capture the complex spatial relationships between objects and the structure of 3D scenes. [13,27] directly extract 3D features from the reconstructed 3D point cloud and support multi-modal visual prompts (text, images, 3D objects) in an MLLM. To alleviate the difficulty LLMs face in understanding complex 3D scenes, [22] choose to first explicitly segment the objects in the 3D scene and then perform multi-stage object-aware scenetest alignment to achieve 3D scene understanding. However, due to the lack of large-scale 3D-language alignment data and the intricate content of 3D scenes, although current MLLMs can achieve favorable performance in 3D scene understanding tasks, their localization performance is still significantly behind the 3D localization specialists. Our approach seeks to address this issue by introducing a 3D grounding model to enhance the localization capability of MLLMs. + +# 3 ScanReason Benchmark + +# 3.1 3D Reasoning Grounding Task + +Given a 3D scene and an implicit free-form query, 3D reasoning grounding requires the model to predict the 3D bounding boxes of the target objects as well as the textual answers and explanations. As shown in Fig. 2, different from traditional 3D visual grounding, the queries of 3D reasoning grounding are implicit and complex, requiring strong reasoning, commonsense, and world knowledge. The number of target objects in 3D reasoning grounding is flexible and any object satisfying the requirements should be considered as the target object. + +![](images/8f87ed3d7733da89da0d9d3b3bf4ad2f2ebd2e4c9ff0528c031dc4fad4c9d162.jpg) +Fig. 2: The left side figure shows the overall of our ScanReason dataset. For each reasoning category, we designed different prompts to generate corresponding questions. And the right side figure shows the differences between the traditional 3D visual grounding task and our proposed 3D reasoning grounding task. + +![](images/3513ff74a02228af403122d8bdd0ee082eeb2998727007dc98136604b0aed7d1.jpg) + +# 3.2 Question Types + +To comprehensively evaluate the 3D reasoning grounding abilities, we define 5 types of questions depending on which type of reasoning is required. Spatial reasoning and functional reasoning require a fundamental understanding of the 3D physical world, while logical reasoning, emotional reasoning, and safety reasoning are high-level reasoning skills built upon the two fundamental reasoning abilities to address user-centric real-world applications, as shown in Fig. 2. + +Spatial Reasoning measures models' understanding of 3D spatial relationships among objects in 3D scene. It encompasses the ability to comprehend the layout of the 3D scene and the 3D location of objects within it, which could serve as the foundation for navigating or planning movements in the 3D environment. + +Functional Reasoning involves understanding and inferring the purpose, function, or affordance of objects within the 3D scene. For example, functional reasoning allows an embodied agent to recognize that a chair is for sitting, a lamp is for lighting, and a refrigerator is for storing food at low temperatures. Such understanding enables the embodied agent to assist users and to perform complex tasks more effectively (e.g., turning on a lamp when the room gets dark, or navigating to a refrigerator to fetch a drink). + +Logical Reasoning allows an embodied agent to not only understand its environment but also to interact with it in a goal-directed manner. For example, given a question shown in Fig. 2: "If I'm cooking dinner in the kitchen, where is the nearest place to throw the rubbish?", an agent needs to use such reasoning ability to infer the location of objects (in this question, a rubbish bin) based on their function and spatial relationships under the specific setting (the kitchen). + +Emotional Reasoning plays a critical role in human-robot interaction, where the target objects are determined by understanding human emotions, preferences, and behavioral patterns. This ability makes the embodied agents more attuned to the emotional and psychological states of humans, allowing them to provide more personalized, empathetic, and contextually appropriate responses + +and solutions, such as: "I'm very sad, is there something that can make me feel happy?" shown in 2. + +Safety Reasoning focuses on preventing harm and ensuring the well-being of humans in the 3D environment. It requires the embodied agent to identify and assess the risk and make safety-aware decisions, such as: "Where should first aid kits be placed for easy access but out of reach of children?" + +# 3.3 Automatic Data Annotation with GPT-4 + +We leverage the 3D scenes and bounding box annotations from the EmbodiedScan dataset [37] and apply GPT-4 [30] to automatically generate question-answer-location pairs for the five question types respectively. Specifically, we provide GPT-4 with the categories and bounding box locations of all objects in the scene, and ask GPT-4 to generate questions and answers with the target object ids of the provided objects. The details can be found in the Appendix. + +In total, our ScanReason dataset consists of 12929 complex reasoning question-answer-location pairs from 1456 scenes, which are split into 11455 training pairs and 1474 validation pairs. All 1474 validation question-answer pairs have been manually verified, including 342 spatial reasoning questions, 287 functional reasoning questions, 581 logical reasoning questions, 211 emotional questions, and 53 safety reasoning questions. We provide detailed statistics and more examples of our dataset in the Appendix. + +# 3.4 Evaluation Metric + +To evaluate the accuracy of predicted objects and their locations for a flexible number of ground-truth objects, we follow the evaluation metric of 3D object detection tasks. Specifically, we adopt Acc@kIoU as our metric, where $k$ is the threshold for the Intersection of Union (IoU) between positive predictions and ground truths. We evaluate the performance under $k = 0.25$ in our experiments. + +# 4 Method + +Solving the task of 3D reasoning grounding requires the synergization of the perception, reasoning, and grounding capability of the embodied agent. Intuitively, we can first conduct reasoning based on implicit instruction such as "where should first aid kits be placed?" and the visual environment. The reasoning process provides us with information about the rough location and semantics of the object we are looking for. Then keeping that information in mind, we look back to the 3D environment to precisely locate the object. For complex scenarios, alternate reasoning and looking back are required for multiple rounds until we obtain the final answer. + +Inspired by this intuition, we propose ReGround3D, consisting of a visual-centric reasoning module and a 3D grounding module with geometry-enhanced look-back, as illustrated in Fig. 3. The visual-centric reasoning module (Sec. 4.1) + +performs joint reasoning of language instruction and visual scene, and predicts a special $\langle \mathsf{LOC} \rangle$ token representing the grounding information. The 3D grounding module (Sec. 4.2) looks back to the original 3D scene with comprehensive geometry information and fine-grained details. It takes the hidden embedding of the $\langle \mathsf{LOC} \rangle$ token containing grounding-related information from the 3D features and eventually predicts the 3D locations of the target objects. Furthermore, we propose a Chain-of-Grounding mechanism (CoG) (Sec. 4.3), i.e., a chain of interleaved reasoning and grounding steps, to further synergize the grounding and reasoning capability for the 3D reasoning grounding task, as illustrated in Fig. 4. + +# 4.1 Visual-Centric Reasoning + +Due to the complexity of the 3D scene and user instructions, particularly given the implicit intention behind the human instruction, ReGround3D starts with a visual-centric reasoning module that can perceive the scene, comprehend the human instructions, and conduct joint reasoning of 3D scene and instructions. We believe the reasoning process eventually implies the grounding intention, i.e., implicitly encodes the information indicating the target object to solve the task. Thus, we design the visual-centric reasoning module to predict grounding queries for localizing the target objects in the following 3D grounding module. + +Specifically, for simplicity, we leverage 3D-LLM to serve as the visual-centric reasoning module because of its strong reasoning abilities inherited from the LLM. Based on the BLIP2 architecture [26], 3D-LLM uses pre-trained image encoders to extract multi-view 2D image features and back-project them into 3D spaces. The visual features are encoded by the Q-Former to produce 32 tokens as the visual input to the LLM. By leveraging the pre-trained image encoders and the Q-Former, the visual tokens encode rich semantics but lack the 3D structures, spatial interactions, and fine-grained details. Therefore, instead of directly predicting the object locations by the 3D-LLM, we ask the 3D-LLM to predict the feature representation as the output of the reasoning process, and the predicted feature is further used to ground the target object in the grounding module. + +To enable the prediction of the grounding feature, we expand the original vocabulary of the 3D-LLM with a special $\langle \mathsf{L0C} \rangle$ token. The $\langle \mathsf{L0C} \rangle$ token is laden with the contextual scene and the target object information which can guide the 3D grounding module to accurately localize target objects. + +# 4.2 3D Grounding with Geometry-Enhanced Look-Back + +Once obtaining the token, ReGround3D extracts the last-layer embedding $h_{loc}$ of the token and sends it into the 3D grounding module to predict the 3D bounding boxes. The 3D grounding module is devised with a "look-back" mechanism which allows the model to access the 3D geometry and fine-grained details from a 3D point cloud encoder. The fine-grained geometry-enhanced 3D visual features and $h_{loc}$ are sent into a query selection module to retrieve the most + +![](images/a63d7526e0e73311efac638837eb1ad1a53e0edcda7dcf7e3c1305227a8d3141.jpg) +Fig. 3: The pipeline of ReGround3D. Given the 3D scene and human instruction, the visual reasoning module first performs joint 3D scene and instruction reasoning, and then guide the 3D grounding module to look-back the 3D scene and perform target object location. + +relevant object features. Those features are further decoded into 3D bounding boxes with the 3D box decoder. + +3D Visual Encoder Unlike the 2D image encoder used for 3D-LLM, the 3D visual encoder directly extracts features from 3D point clouds to capture more geometric and spatial information about the 3D structures and fine-grained details. The powerful 3D visual encoder which captures comprehensive geometry, structure, layout, and fine-grained information is critical to accurate 3D grounding. Subsequently, the 3D features $f_{scene}$ produced by the 3D visual encoder and the grounding feature $h_{loc}$ from 3D-LLM are sent to the query selection module. + +Query Selection Module We adopt a cross-attention mechanism, where we treat $f_{\text{scene}}$ as $Q$ (Query), and $h_{\text{loc}}$ as both $K$ (Key) and $V$ (Value), to implicitly obtain a feature-level reasoning activation heatmap. During this scene look-back process, this module roughly locates scene features that have a high response to the token. We then select the $k$ most relevant features as the object query $f_{\text{query}}$ . + +3D Box Decoder is a classical transformer decoder, which consists of M transformer decoder layers. In each decoder layer, the object queries $f_{query}$ go through the text feature $h_{loc}$ cross-attention layer and scene feature $f_{scene}$ cross-attention layer. Finally, the prediction head takes the updated object queries as input and predicts the final 3D locations and matching score. + +Discussion In comparison to previous works [11,13,15,21] that directly predicts the bounding boxes by the LLM, the extra grounding module has the following advantages: 1) Based on the MLLM reasoning results, it allows the grounding module to perceive the scene again and focus on the region under the implicit guidance of MLLM, adapted to the user queries and the reasoning results. 2) The scene representation perceived in the grounding module can be more precise and fine-grained, which is complementary to the visual-centric reasoning module. 3) The two-step reasoning-grounding pipeline is flexible and can generalized + +![](images/7edbca3f9be2f8c93f5dd8fc66e31685c5b11300286c4d82ea9516bfe80a1150.jpg) +Fig. 4: Illustration of Chain-of-Grounding (CoG) Mechanism + +to other types of predicting formats such as segmentation masks (by simply replacing the 3D box decoder with a 3D mask decoder). + +# 4.3 Chain-of-Grounding Mechanism + +The existing design conducts the reasoning and grounding process sequentially, i.e., the reasoning process is finished before grounding. We argue that the grounding results can also facilitate the reasoning process, especially for those requiring spatial information. Thus, to further synergize the reasoning and grounding process, inspired by chain-of-thought (CoT) [38], we propose Chain-of-Grounding (CoG), which introduces a chain of interleaved steps of reasoning and grounding to find the targeted objects during inference, as shown in Fig. 4. Such a process allows the model to actively find relevant objects that help solve the problem, and then conduct reasoning with the assistance of the additional information of these relevant objects so that the model can more precisely find the target objects. + +Specifically, given a question provided by users, CoG translates it into another question of finding the explicitly mentioned objects in the original question. The generated new question is sent into ReGround3D to ground the objects mentioned in the original question in the 3D scene with corresponding confidence scores. An object can be seen as a successfully located object when its confidence score is above the threshold, and the located object information could serve as explicit guidance for 3D-LLM in the next reasoning stage. As shown in Fig. 4, after obtaining the 3D information of objects explicitly mentioned in the original question, the located object information is inserted to update the question. The updated question is then sent to ReGround3D to perform reasoning and grounding to output the target object locations. + +# 4.4 Instruction Tuning + +Training Objective We use the pre-trained weights of 3D-LLM as the initialization for the visual-centric reasoning module. Except for freezing the 3D + +Table 1: We list part of data templates used to train ReGround3D for each task. + +
Task NameText InstructionsOutput Type TemplatesExpected Output
3D Visual Grounding<scene> Here is a description Please answer the question only with outabout an object: "<expr>", whereis the object in the 3D scene?Please answer the question only with out-put 3D box prediction(s).It is <LOC>.
3D/Spatial Question Answering<scene> Answer the question Please answer the question only with text,do not output 3D box prediction(s).<question>Please answer the question only with text,do not output 3D box prediction(s).<answer>.
<scene><situation>, <question>Please answer the question only with text,do not output 3D box prediction(s).<answer>.
<scene><question>Please answer the question only with text,do not output 3D box prediction(s).<answer>.
3D Object Detection<scene> Where is the <category> Please answer the question only with out-in this 3D scene? Put 3D box prediction(s).Sure, <LOC>.
3D Reasoning Grounding<scene> Answer the question: Please answer the question with text andoutput 3D box prediction(s).Sure, <LOC>, <reason>
+ +visual encoder pre-trained on [37], the rest of the parameters in the visual-centric reasoning module and 3D grounding module in our framework are trained in an end-to-end manner. The training supervision is a weighted sum of the next token prediction loss from 3D-LLM and the 3D detection loss from the 3D grounding module. + +$$ +\mathcal {L} = \lambda_ {\text {t e x t}} \mathcal {L} _ {\text {t e x t}} + \lambda_ {\text {d e t}} \mathcal {L} _ {\text {d e t}} \tag {1} +$$ + +The 3D detection loss is defined following: + +$$ +\mathcal {L} _ {\text {d e t}} = \lambda_ {I O U} \mathcal {L} _ {I O U} + \lambda_ {\text {c o n t r a s t}} \mathcal {L} _ {\text {c o n t r a s t}} \tag {2} +$$ + +Instruction Tuning Dataset We load the pre-trained weights of 3D-LLM and the 3D visual encoder, and finetune the LoRA of 3D-LLM, the query selection module, and the 3D box decoder with an instruction tuning dataset. To construct the instruction tuning dataset, we reformulate the data annotations from existing 3D datasets into question-answer or question-answer-bounding-box pairs. Specifically, the 3D visual grounding data from ScanRefer [7], SR3D, NR3D [1] and the 3D object detection data from EmbodiedScan [37] are formulated into question-answer-bbox pairs, and the 3D/spatial question answering data from SR3D [1], CLEVR3D [41], SQA3D [29] are formulated into question-answer pairs without bounding boxes. The information shown in Tab. 1 illustrates how we unify the instruction and output with task-specific templates. More details can be found in the Appendix. The reformulated data combined with our proposed ScanReason dataset, serve as the instruction tuning dataset. + +# 5 Experiment + +# 5.1 Implementation Details + +Network Architecture For the 3D grounding module, we adopt the pretrained point cloud encoder as the 3D visual encoder. During the training stage, We use LoRA to efficiently finetune the 3D-LLM to preserve the original 3D scene understanding capability and reduce the computation costs. The number of object queries $k$ in the query selection module is set to 256. + +Training Parameters The training is done on 8 NVIDIA A100 GPUs. We adopt the AdamW optimizer with a learning rate of 3e-4 and use a learning rate scheduler WarmupDecayLR with the warmup steps of 100. The total batch size is set to be 16. The loss weight parameters $\lambda_{text}$ and $\lambda_{det}$ in total loss $\mathcal{L}$ are set to 1.0 and 1.0, respectively, and the weight $\lambda_{IOU}$ and $\lambda_{contrast}$ in $\mathcal{L}_{\mathrm{det}}$ are set to 1.0 and 1.0. + +# 5.2 Results Comparison + +Evaluation on 3D Visual Grounding In order to verify the superiority of our model in grounding ability and facilitate comparison between current models, we report the explicit grounding performance on the existing 3D visual grounding task. Since the evaluation settings of Nr3D and Sr3D [1] are based on ground-truth object proposals, while ScanRefer [7] requires models to output 3D bounding boxes, we choose ScanRefer as our benchmark for comparison. We divide the existing methods into two categories, one is the grounding model designed specifically for the 3D visual grounding task, and the other is generalist MLLMs which can understand a variety of 3D vision-language tasks. The original 3D-LLM embeds 3D locations in the vocabularies and represents the grounded 3D bounding boxes by a sequence of discrete location tokens. However, since the fine-tuned model of 3D-LLM on ScanRefer and related location tokens implementation are not accessible, we adapt 3D-LLM to directly output 3D numerical coordinates representing 3D bounding boxes by fine-tuning the pre-trained model on our reformulated 3D visual grounding data, denotes as 3D-LLM (vg). As shown in Tab. 2, current generalist MLLM models still lag behind the specialist models in terms of grounding ability. By incorporating the 3D grounding module into MLLM, ReGround3D shows the SOTA performance on the traditional 3D visual grounding task. + +Table 2: Results on 3D visual grounding task among ReGround3D (ours) and existing methods. + +
TypeMethodsAcc@0.25Acc@0.5
SpecialistsScanRefer [7]37.324.3
MVT [23]40.833.3
3DVG-Trans [48]45.934.5
ViL3DRel [12]47.937.7
BUTD-DETR [24]52.239.8
L3Det [50]52.840.2
Generalized MLLMsLLM-Grounder [42]17.15.3
3D-LLM [21]30.3-
3D-LLM(vg) [21]33.128.7
Chat3D-v2 [22]35.930.4
oursReGround3D53.141.1
+ +Table 3: Results (Acc@0.25) on 3D reasoning grounding task among ReGround3D (ours) and existing methods. + +
MethodsLLMSpatial FunctionalLogicalEmotional SafetyOverall
Mask3D [32] + InternLM2-7B [33]InternLM2-7B10.3436.129.988.218.9914.86
3D-LLM(vg) [21]FlanT5XL-3B18.3117.4210.978.126.3313.29
Chat3D-v2 [22]Vicuna-7B20.2118.3911.327.989.8814.98
ReGround3D*FlanT5XL-3B30.7629.818.6719.2217.1223.27
ReGround3DFlanT5XL-3B32.9836.2326.9923.1222.9828.98
ReGround3D (CoG)FlanT5XL-3B34.7136.7929.1124.0323.2130.62
+ +Evaluation on 3D Reasoning Grounding The various visual grounding models rely on explicit text-object alignment in the input object expression to achieve localization, which falls to be applied to 3D reasoning grounding task. We performed a comparison between our proposed method ReGround3D and existing MLLM methods, including 3D-LLM(vg) and Chat3D-v2 [22]. Besides, inspired by Chat-3D v2, which first segments objects, then equips them with unique object identifiers to conduct effective object grounding, we set up a LLM-based 3D reasoning grounding baseline: We first segment the objects from the scene using a 3D instance segmentor [32], then convert the segmented object information including their categories and 3D bounding boxes into text as input of LLM (InternLM2-7B [33]). Besides, to better validate the performance of our model and ensure a fair comparison, we remove the ScanReason dataset from the training data, denoted as ReGround3D*. As shown in Tab. 3, the LLM-based reasoning method (Mask3D [32] + InternLM2-7B [33]) possesses a very strong functional reasoning ability, but struggles to understand of 3D spatial relationship. Additionally, irrespective of whether ScanReason is used in training, our model significantly outperforms the existing MLLMs. By synergizing the reasoning and grounding process utilizing the Chain-of-Grounding (CoG) mechanism, the 3D reasoning grounding performance of ReGround3D can be further improved (from 28.98 to 30.62), especially on the spatial reasoning and logical reasoning questions. The qualitative results comparison shown in Fig. 5 demonstrates the superiority of our method in reasoning human complex instruction based on the 3D scene. + +# 5.3 Ablation Study + +In this section, we conduct an extensive ablation study to verify the effectiveness of each component. + +Effectiveness of 3D Grounding Module In order to more comprehensively verify the effectiveness of the 3D grounding module we proposed, we step by step verify the performance changes from 3D-LLM to ReGround3D. Apart from the 3D-LLM(vg), we fine-tuned the 3D-LLM model respectively on full reformulated existing data and all instruction tuning data, including ScanReason, denoted as + +![](images/784f1f2f5b6f232eba3420d872b915d4273734ee0cbb2c82f4a2d907c77908c0.jpg) + +It's so cold outside, Where could I find my thick coat? + +![](images/27dcfd88b158e59479e6576ac7425d522a093628a2d07ff552c4f872f11c92b2.jpg) + +![](images/c9fa2146b542c300c78a46511bc63b7a9b3ae39054320d18b0a71467da5acf5e.jpg) + +Where can I find paper to wipe my hands? + +![](images/4511c469625c1467fc4835c6fcd6ab695eebe2823a680f18daa75884368e1258.jpg) + +![](images/8d2c4f40c5021c3aa2fe6791c813035cbda4adb46952763b92b652089c94293b.jpg) + +It's dark here, can you find something to light up the room? + +![](images/631c09b5775357e613af7d1fbcb51c92ffc52729ec5abae40113360c2e24115e.jpg) +Fig. 5: Visualization comparison of 3D reasoning grounding capability between ReGround3D and 3D-LLM. Our method could achieve much more accurate grounding results which satisfy the implicit question intention, and give the corresponding explanation at the same time. More illustrations are given in the Appendix. + +3D-LLM (full) and 3D-LLM (full+sr). All the tokens in answers of the data used for fine-tuning 3D-LLM have been converted to the numerical coordinates of the corresponding boxes. The results in Tab. 4a show that with the same tuning dataset, our ReGround3D achieves far superior performance to 3D-LLM (28.98 vs. 19.21) by introducing the 3D grounding module. + +Effectiveness of Instruction Tuning Dataset Tab. 4b showcases the impact of different training data types on 3D visual grounding performance. 3D Object Detection (3D OD) provides the explicit 3D semantic category and visual alignment whereas 3D Question Answering (3D QA) data injects the basic 3D scene understanding ability into the model, which has certain benefits for the visual grounding ability of the model. In addition, we find that training with 3D reasoning grounding dataset can further improve the performance on 3D visual grounding. + +(a) When using the same tuning dataset, ReGround3D achieves far better performance on ScanReason than 3D-LLM. + +(b) Ablation study on training data. We evaluate through the metric of accuracy on the val set of the ScanRefer dataset. + +Table 4: Ablation study on effectiveness of 3D grounding module and training data + +
MethodsAcc@0.25
3D-LLM(vg) [21]13.29
3D-LLM(full) [21]15.31
3D-LLM(full+sr) [21]19.21
ReGround3D(full+sr)28.98
+ +
DatasetAcc@0.25Acc@0.5
VGODVQARD
✓✓19.314.2
✓✓48.737.6
✓✓49.238.1
✓✓✓51.839.4
✓✓✓✓53.141.1
+ +Discussion of CoG Mechanism While the CoG Mechanism boosts the performance with interleaved reasoning and grounding steps during inference, it uses the relevant object information explicitly presented in the question to help find the target objects. A natural question arises: if we input the information of all the objects instead of relevant objects into ReGround3D during CoG, will this make the model more accurately find the target objects? We first use the existing 3D Segmentor [32] to segment all objects in the scene, then update the original question using all the object 3D bounding boxes information according to the template in Sec. 4.3 and send into ReGround3D. However, experiments show that using all the object information will instead reduce the 3D reasoning grounding performance from 28.98 to 27.67. One possible reason could be the model's attention is dispersed by too many irrelevant objects. + +# 6 Conclusion + +This paper introduces a new 3D vision language learning task: 3D reasoning grounding, which requires the model to perform active reasoning over complex and implicit human instruction, localize the target objects, and give corresponding explanations. Furthermore, we propose ScanReason, a new dataset and benchmark to further unlock and thoroughly evaluate the 3D reasoning grounding capability. Based on this dataset, we propose a novel approach: ReGround3D, which utilizes the strong reasoning capability of MLLM guiding the 3D grounding module to obtain accurate object locations, and a Chain of Grounding (CoG) mechanism is presented to further boost the performance with interleaved reasoning and grounding steps during inference. We believe that our work will further the natural interaction between embodied agents and humans in open 3D environments. For the current ScanReason benchmark, we find that the questions in three high-level 3D reasoning categories may have overlaps. For a certain reasoning question, similar questions may appear in one or two other categories. We leave the problem as a future challenge for better reasoning grounding ability evaluation. + +# Acknowledgements + +This work is supported in part by HKU Startup Fund, HKU Seed Fund for Basic Research, HKU Seed Fund for Translational and Applied Research, HKU IDS research Seed Fund, and HKU Fintech Academy R&D Funding. + +# References + +1. Achlioptas, P., Abdelreheem, A., Xia, F., Elhoseiny, M., Guibas, L.: Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16. pp. 422-440. Springer (2020) +2. Azuma, D., Miyanishi, T., Kurita, S., Kawanabe, M.: Scanqa: 3d question answering for spatial scene understanding. In: proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 19129-19139 (2022) +3. Bakr, E.M., Ayman, M., Ahmed, M., Slim, H., Elhoseiny, M.: Cot3dref: Chain-of-thoughts data-efficient 3d visual grounding. arXiv preprint arXiv:2310.06214 (2023) +4. Baroni, M.: Linguistic generalization and compositionality in modern artificial neural networks. Philosophical Transactions of the Royal Society B 375(1791), 20190307 (2020) +5. Cai, D., Zhao, L., Zhang, J., Sheng, L., Xu, D.: 3djcg: A unified framework for joint dense captioning and visual grounding on 3d point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16464-16473 (2022) +6. Chen, C., Qin, R., Luo, F., Mi, X., Li, P., Sun, M., Liu, Y.: Position-enhanced visual instruction tuning for multimodal large language models. arXiv preprint arXiv:2308.13437 (2023) +7. Chen, D.Z., Chang, A.X., Niefer, M.: Scanrefer: 3d object localization in rgb-d scans using natural language. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XX. pp. 202-221. Springer (2020) +8. Chen, D.Z., Wu, Q., Nießner, M., Chang, A.X.: D3net: A speaker-listener architecture for semi-supervised dense captioning and visual grounding in rgb-d scans. arXiv preprint arXiv:2112.01551 (2021) +9. Chen, F., Han, M., Zhao, H., Zhang, Q., Shi, J., Xu, S., Xu, B.: X-LLM: Bootstrapping advanced large language models by treating multi-modalities as foreign languages. arXiv preprint arXiv:2305.04160 (2023) +10. Chen, J., Zhu, D., Shen, X., Li, X., Liu, Z., Zhang, P., Krishnamoorthi, R., Chandra, V., Xiong, Y., Elhoseiny, M.: Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478 (2023) +11. Chen, K., Zhang, Z., Zeng, W., Zhang, R., Zhu, F., Zhao, R.: Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195 (2023) +12. Chen, S., Guhur, P.L., Tapaswi, M., Schmid, C., Laptev, I.: Language conditioned spatial relation reasoning for 3d object grounding. Advances in Neural Information Processing Systems 35, 20522-20535 (2022) + +13. Chen, S., Chen, X., Zhang, C., Li, M., Yu, G., Fei, H., Zhu, H., Fan, J., Chen, T.: Ll3da: Visual interactive instruction tuning for omni-3d understanding, reasoning, and planning. arXiv preprint arXiv:2311.18651 (2023) +14. Chen, S., Zhu, H., Chen, X., Lei, Y., Yu, G., Chen, T.: End-to-end 3d dense captioning with vote2cap-detr. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11124-11133 (2023) +15. Chen, T., Saxena, S., Li, L., Fleet, D.J., Hinton, G.: Pix2seq: A language modeling framework for object detection. arXiv preprint arXiv:2109.10852 (2021) +16. Chen, Z., Gholami, A., Nießner, M., Chang, A.X.: Scan2cap: Context-aware dense captioning in rgb-d scans. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3193-3203 (2021) +17. Chen, Z., Hu, R., Chen, X., Nießner, M., Chang, A.X.: Unit3d: A unified transformer for 3d dense captioning and visual grounding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 18109-18119 (2023) +18. Chung, H.W., Hou, L., Longpre, S., Zoph, B., Tay, Y., Fedus, W., Li, E., Wang, X., Dehghani, M., Brahma, S., et al.: Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416 (2022) +19. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning (2023) +20. Han, J., Zhang, R., Shao, W., Gao, P., Xu, P., Xiao, H., Zhang, K., Liu, C., Wen, S., Guo, Z., et al.: Imagebind-llm: Multi-modality instruction tuning. arXiv preprint arXiv:2309.03905 (2023) +21. Hong, Y., Zhen, H., Chen, P., Zheng, S., Du, Y., Chen, Z., Gan, C.: 3d-llm: Injecting the 3d world into large language models. Advances in Neural Information Processing Systems 36 (2024) +22. Huang, H., Wang, Z., Huang, R., Liu, L., Cheng, X., Zhao, Y., Jin, T., Zhao, Z.: Chat-3d v2: Bridging 3d scene and large language models with object identifiers. arXiv preprint arXiv:2312.08168 (2023) +23. Huang, S., Chen, Y., Jia, J., Wang, L.: Multi-view transformer for 3d visual grounding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15524-15533 (2022) +24. Jain, A., Gkanatsios, N., Mediratta, I., Fragkiadaki, K.: Bottom up top down detection transformers for language grounding in images and point clouds. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVI. pp. 417-433. Springer (2022) +25. Jiao, Y., Chen, S., Jie, Z., Chen, J., Ma, L., Jiang, Y.G.: More: Multi-order relation mining for dense captioning in 3d scenes. arXiv preprint arXiv:2203.05203 (2022) +26. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023) +27. Li, M., Chen, X., Zhang, C., Chen, S., Zhu, H., Yin, F., Yu, G., Chen, T.: M3dbench: Let's instruct large models with multi-modal 3d prompts. arXiv preprint arXiv:2312.10763 (2023) +28. Luo, J., Fu, J., Kong, X., Gao, C., Ren, H., Shen, H., Xia, H., Liu, S.: 3d-sps: Single-stage 3d visual grounding via referred point progressive selection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16454-16463 (2022) +29. Ma, X., Yong, S., Zheng, Z., Li, Q., Liang, Y., Zhu, S.C., Huang, S.: Sqa3d: Situated question answering in 3d scenes. arXiv preprint arXiv:2210.07474 (2022) + +30. OpenAI: Gpt-4 technical report (2023) +31. Roh, J., Desingh, K., Farhadi, A., Fox, D.: Languagerefer: Spatial-language model for 3d visual grounding. In: Conference on Robot Learning. pp. 1046-1056. PMLR (2022) +32. Schult, J., Engelmann, F., Hermans, A., Litany, O., Tang, S., Leibe, B.: Mask3d: Mask transformer for 3d semantic instance segmentation. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 8216-8223. IEEE (2023) +33. Team, I.: Internl'm: A multilingual language model with progressively enhanced capabilities (2023) +34. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023) +35. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017) +36. Wang, H., Zhang, C., Yu, J., Cai, W.: Spatiality-guided transformer for 3d dense captioning on point clouds. arXiv preprint arXiv:2204.10688 (2022) +37. Wang, T., Mao, X., Zhu, C., Xu, R., Lyu, R., Li, P., Chen, X., Zhang, W., Chen, K., Xue, T., et al.: Embodiedscan: A holistic multi-modal 3d perception suite towards embodied ai. arXiv preprint arXiv:2312.16170 (2023) +38. Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q.V., Zhou, D., et al.: Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems 35, 24824-24837 (2022) +39. Wu, Y., Cheng, X., Zhang, R., Cheng, Z., Zhang, J.: Eda: Explicit textdecoupling and dense alignment for 3d visual and language learning. arXiv preprint arXiv:2209.14941 (2022) +40. Xu, R., Wang, X., Wang, T., Chen, Y., Pang, J., Lin, D.: Pointllm: Empowering large language models to understand point clouds. arXiv preprint arXiv:2308.16911 (2023) +41. Yan, X., Yuan, Z., Du, Y., Liao, Y., Guo, Y., Li, Z., Cui, S.: Clevr3d: Compositional language and elementary visual reasoning for question answering in 3d real-world scenes. arXiv preprint arXiv:2112.11691 (2021) +42. Yang, J., Chen, X., Qian, S., Madaan, N., Iyengar, M., Fouhey, D.F., Chai, J.: Llm-grounder: Open-vocabulary 3d visual grounding with large language model as an agent. arXiv preprint arXiv:2309.12311 (2023) +43. Yang, Z., Zhang, S., Wang, L., Luo, J.: Sat: 2d semantics assisted training for 3d visual grounding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1856-1866 (2021) +44. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178 (2023) +45. Ye, S., Chen, D., Han, S., Liao, J.: 3d question answering. IEEE Transactions on Visualization and Computer Graphics (2022) +46. Yuan, Z., Yan, X., Liao, Y., Guo, Y., Li, G., Cui, S., Li, Z.: X-trans2cap: Cross-modal knowledge transfer using transformer for 3d dense captioning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8563-8573 (2022) +47. Zhang, S., Roller, S., Goyal, N., Artetxe, M., Chen, M., Chen, S., Dewan, C., Diab, M., Li, X., Lin, X.V., et al.: Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068 (2022) + +48. Zhao, L., Cai, D., Sheng, L., Xu, D.: 3dvg-transformer: Relation modeling for visual grounding on point clouds. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2928-2937 (2021) +49. Zhong, Y., Xu, L., Luo, J., Ma, L.: Contextual modeling for 3d dense captioning on point clouds. arXiv preprint arXiv:2210.03925 (2022) +50. Zhu, C., Zhang, W., Wang, T., Liu, X., Chen, K.: Object2scene: Putting objects in context for open-vocabulary 3d detection. arXiv preprint arXiv:2309.09456 (2023) +51. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023) +52. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023) +53. Zhu, Z., Ma, X., Chen, Y., Deng, Z., Huang, S., Li, Q.: 3d-vista: Pre-trained transformer for 3d vision and text alignment. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2911-2921 (2023) \ No newline at end of file diff --git a/2024/ScanReason_ Empowering 3D Visual Grounding with Reasoning Capabilities/images.zip b/2024/ScanReason_ Empowering 3D Visual Grounding with Reasoning Capabilities/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..33866f9d41c9ada4e12ed7ee017633ad2a11aa8d --- /dev/null +++ b/2024/ScanReason_ Empowering 3D Visual Grounding with Reasoning Capabilities/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:36d225f6f7e1756e42dc5681280c6b0314c87a228ea30cea1e2cc10cd85decfe +size 407865 diff --git a/2024/ScanReason_ Empowering 3D Visual Grounding with Reasoning Capabilities/layout.json b/2024/ScanReason_ Empowering 3D Visual Grounding with Reasoning Capabilities/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9ac2e91a892f5c67d77d1c4ee8e73c7d12e4be9f --- /dev/null +++ b/2024/ScanReason_ Empowering 3D Visual Grounding with Reasoning Capabilities/layout.json @@ -0,0 +1,8065 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 140, + 111, + 474, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 111, + 474, + 148 + ], + "spans": [ + { + "bbox": [ + 140, + 111, + 474, + 148 + ], + "type": "text", + "content": "ScanReason: Empowering 3D Visual Grounding with Reasoning Capabilities" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 140, + 167, + 472, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 167, + 472, + 181 + ], + "spans": [ + { + "bbox": [ + 140, + 167, + 472, + 181 + ], + "type": "text", + "content": "Chenming Zhu" + }, + { + "bbox": [ + 140, + 167, + 472, + 181 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 140, + 167, + 472, + 181 + ], + "type": "text", + "content": ", Tai Wang" + }, + { + "bbox": [ + 140, + 167, + 472, + 181 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 140, + 167, + 472, + 181 + ], + "type": "text", + "content": ", Wenwei Zhang" + }, + { + "bbox": [ + 140, + 167, + 472, + 181 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 140, + 167, + 472, + 181 + ], + "type": "text", + "content": ", Kai Chen" + }, + { + "bbox": [ + 140, + 167, + 472, + 181 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 140, + 167, + 472, + 181 + ], + "type": "text", + "content": ", and Xihui Liu" + }, + { + "bbox": [ + 140, + 167, + 472, + 181 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 203, + 189, + 410, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 189, + 410, + 236 + ], + "spans": [ + { + "bbox": [ + 203, + 189, + 410, + 236 + ], + "type": "text", + "content": "1 The University of Hong Kong \n2 Shanghai AI Laboratory \nchairzhu@connect.hku.hk \nhttps://zcmax.github.io/projects/ScanReason" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 262, + 455, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 262, + 455, + 427 + ], + "spans": [ + { + "bbox": [ + 160, + 262, + 455, + 427 + ], + "type": "text", + "content": "Abstract. Although great progress has been made in 3D visual grounding, current models still rely on explicit textual descriptions for grounding and lack the ability to reason human intentions from implicit instructions. We propose a new task called 3D reasoning grounding and introduce a new benchmark ScanReason which provides over 10K question-answer-location pairs from five reasoning types that require the synchronization of reasoning and grounding. We further design our approach, ReGround3D, composed of the visual-centric reasoning module empowered by Multi-modal Large Language Model (MLLM) and the 3D grounding module to obtain accurate object locations by looking back to the enhanced geometry and fine-grained details from the 3D scenes. A chain-of-grounding mechanism is proposed to further boost the performance with interleaved reasoning and grounding steps during inference. Extensive experiments on the proposed benchmark validate the effectiveness of our proposed approach." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 437, + 453, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 437, + 453, + 460 + ], + "spans": [ + { + "bbox": [ + 159, + 437, + 453, + 460 + ], + "type": "text", + "content": "Keywords: 3D reasoning grounding " + }, + { + "bbox": [ + 159, + 437, + 453, + 460 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 437, + 453, + 460 + ], + "type": "text", + "content": " 3D visual grounding " + }, + { + "bbox": [ + 159, + 437, + 453, + 460 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 437, + 453, + 460 + ], + "type": "text", + "content": " Multi-modal large language models" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 479, + 230, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 230, + 491 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 230, + 491 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 503, + 482, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 503, + 482, + 622 + ], + "spans": [ + { + "bbox": [ + 130, + 503, + 482, + 622 + ], + "type": "text", + "content": "Understanding and reasoning in the 3D visual world is critical for applications such as robotics and AR, where embodied agents are expected to understand the 3D layout and predict the 3D locations of objects based on human instructions. The example in Fig. 1 demonstrates a scenario where the question can only be solved with a comprehensive understanding of the 3D scene and joint reasoning of the question and the 3D environment. However, current 3D visual grounding models [24,39,43,50] trained on [1,7] localize objects based on explicit descriptions of the object category, attribute, and 3D spatial relationships, and lack the ability to reason the user intentions and predict object locations with implicit human instructions such as \"I am thirsty, can I have something to drink?\"" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 623, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 623, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 623, + 482, + 647 + ], + "type": "text", + "content": "To bridge the aforementioned gap and to push the boundaries of what embodied agents can understand and how they can interact with the 3D world, we" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 320, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 320, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 320, + 665 + ], + "type": "text", + "content": "† Corresponding author: xihuiliu@eee.hku.hk" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 159, + 119, + 450, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 119, + 450, + 137 + ], + "spans": [ + { + "bbox": [ + 159, + 119, + 450, + 137 + ], + "type": "text", + "content": "Question: I'm looking for a comfortable spot to watch TV immersively, where is the best place for me to lie down and watch TV?" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 134, + 145, + 152, + 160 + ], + "blocks": [ + { + "bbox": [ + 134, + 145, + 152, + 160 + ], + "lines": [ + { + "bbox": [ + 134, + 145, + 152, + 160 + ], + "spans": [ + { + "bbox": [ + 134, + 145, + 152, + 160 + ], + "type": "image", + "image_path": "4c561b3f0105a6a0b0268ea384fa4a072cd3d46107e710ee8cfb9fad6d2e5abe.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 133, + 165, + 154, + 173 + ], + "lines": [ + { + "bbox": [ + 133, + 165, + 154, + 173 + ], + "spans": [ + { + "bbox": [ + 133, + 165, + 154, + 173 + ], + "type": "text", + "content": "GPT-4" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 148, + 284, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 148, + 284, + 218 + ], + "spans": [ + { + "bbox": [ + 160, + 148, + 284, + 218 + ], + "type": "text", + "content": "The best spot for lying down to watch TV immersively is on a spacious, cushioned sectional sofa or an adjustable reclining bed in a room with a large screen TV mounted at an eye-level view from your lying position, ensuring comfort and optimal viewing angles. Ensure the room is dimly lit or has blackout curtains to enhance screen visibility and reduce glare." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 293, + 148, + 335, + 215 + ], + "blocks": [ + { + "bbox": [ + 293, + 148, + 335, + 215 + ], + "lines": [ + { + "bbox": [ + 293, + 148, + 335, + 215 + ], + "spans": [ + { + "bbox": [ + 293, + 148, + 335, + 215 + ], + "type": "image", + "image_path": "4a95c7dbd5de217e17e8a7e1b54d70cd020b4df923a278dce65612d360db09f6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 337, + 148, + 380, + 214 + ], + "blocks": [ + { + "bbox": [ + 337, + 148, + 380, + 214 + ], + "lines": [ + { + "bbox": [ + 337, + 148, + 380, + 214 + ], + "spans": [ + { + "bbox": [ + 337, + 148, + 380, + 214 + ], + "type": "image", + "image_path": "ce635d6fa7f15884bccb2f950a6b0d5e412f2b3aa27caec0eedee6685f12390f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 382, + 148, + 449, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 148, + 449, + 210 + ], + "spans": [ + { + "bbox": [ + 382, + 148, + 449, + 210 + ], + "type": "text", + "content": "The best spot to lie down and watch TV immersively would be on the bed facing the TV in the cabinet; it provides a direct line of sight and the comfort of lying down." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 457, + 146, + 475, + 162 + ], + "blocks": [ + { + "bbox": [ + 457, + 146, + 475, + 162 + ], + "lines": [ + { + "bbox": [ + 457, + 146, + 475, + 162 + ], + "spans": [ + { + "bbox": [ + 457, + 146, + 475, + 162 + ], + "type": "image", + "image_path": "aa38eb9871b82966b6329b67f248d43cfc47ccb26d6c4f553de5716685477204.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 455, + 164, + 479, + 172 + ], + "lines": [ + { + "bbox": [ + 455, + 164, + 479, + 172 + ], + "spans": [ + { + "bbox": [ + 455, + 164, + 479, + 172 + ], + "type": "text", + "content": "GPT-4V" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 135, + 220, + 152, + 238 + ], + "blocks": [ + { + "bbox": [ + 135, + 220, + 152, + 238 + ], + "lines": [ + { + "bbox": [ + 135, + 220, + 152, + 238 + ], + "spans": [ + { + "bbox": [ + 135, + 220, + 152, + 238 + ], + "type": "image", + "image_path": "0e30da9813c4b3ec62ff18a22440b8ac36f716ee37e9d7ca1bd6d6557adb0cdf.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 228, + 194, + 236 + ], + "lines": [ + { + "bbox": [ + 153, + 228, + 194, + 236 + ], + "spans": [ + { + "bbox": [ + 153, + 228, + 194, + 236 + ], + "type": "text", + "content": "ReGround3D" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 161, + 244, + 224, + 297 + ], + "blocks": [ + { + "bbox": [ + 161, + 244, + 224, + 297 + ], + "lines": [ + { + "bbox": [ + 161, + 244, + 224, + 297 + ], + "spans": [ + { + "bbox": [ + 161, + 244, + 224, + 297 + ], + "type": "image", + "image_path": "97b7d2df696c624b87e1e831451244f95ed889ec992544662311448d0163f013.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 316, + 482, + 392 + ], + "lines": [ + { + "bbox": [ + 130, + 316, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 482, + 392 + ], + "type": "text", + "content": "Fig. 1: For an embodied agent, they not only need to be able to understand the 3D environment and complex human instructions but also localize the target objects for interaction and navigation. Although GPT-4 (GPT-4V) have strong text (multi-modal) reasoning abilities, they lack the ability to directly perceive the 3D scene, understand the 3D spatial relationships, and output corresponding target object locations. Instead, our proposed method ReGround3D has the 3D perception, reasoning, and grounding capabilities in the real-world 3D environment." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 233, + 252, + 373, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 252, + 373, + 285 + ], + "spans": [ + { + "bbox": [ + 233, + 252, + 373, + 285 + ], + "type": "text", + "content": "For an optimal TV viewing experience, I recommend sitting on the couch. It is oriented towards the TV and closer to the TV compared with the bed, and it's large enough to lie down." + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 383, + 243, + 444, + 298 + ], + "blocks": [ + { + "bbox": [ + 383, + 243, + 444, + 298 + ], + "lines": [ + { + "bbox": [ + 383, + 243, + 444, + 298 + ], + "spans": [ + { + "bbox": [ + 383, + 243, + 444, + 298 + ], + "type": "image", + "image_path": "53f86f8df84306de5070778a32a6056b0d90a3b96eafe9bba48d951c14a45520.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 423, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 482, + 555 + ], + "type": "text", + "content": "propose a new task of 3D reasoning grounding and introduce a new benchmark named ScanReason. The task requires the model to conduct joint reasoning on the question and the 3D environment before predicting the 3D locations of target objects. We define five categories of 3D reasoning: spatial reasoning, functional reasoning, logical reasoning, emotional reasoning, and safety reasoning. The first two categories focus on the fundamental understanding of the 3D physical world, while the last three categories are built upon fundamental abilities to address user-centric real-world challenges. The benchmark comprises more than 10K question-answer-3D bounding box pairs from 2K scenes belonging to the five reasoning types mentioned above. The GPT-4-assisted data annotation process largely increases the efficiency of curating such a dataset." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "We propose ReGround3D as an initial attempt to the new task of 3D Reasoning Grounding. Intuitively, for 3D grounding with implicit instructions, we need to first conduct reasoning on the language instructions and the coarse visual environment. Then with the idea of which object we want to find in mind, we look back to the 3D scene and ground the target object. For complex instructions, we may need to alternate the reasoning and look-back process for multiple iterations. Inspired by this intuition, our framework is composed of a visual-centric reasoning module and 3D grounding with geometry-enhanced look-back module, with a Chain-of-Grounding mechanism during inference to" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "C. Zhu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 259 + ], + "type": "text", + "content": "alternately conduct reasoning and grounding for multiple rounds. Specifically, the visual-centric reasoning module conducts joint reasoning of the 3D scene and instructions with an MLLM. This module predicts a special token representing the semantic and location information of the target object, which is used for the grounding module. The 3D grounding module uses the output token embedding from the previous reasoning module to locate the target object by looking back at the fine-grained 3D scene representation. Unlike previous MLLMs attempting to directly predict bounding box coordinates, our look-back mechanism enables the model to capture more comprehensive 3D geometry and fine-grained object details for accurate 3D grounding. The Chain-of-Grounding mechanism is proposed to synergize reasoning and grounding, which allows multiple rounds alternating between reasoning and grounding during inference." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 261, + 482, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 261, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 261, + 482, + 392 + ], + "type": "text", + "content": "In summary, our contributions are threefold: 1) We propose a new task of 3D reasoning grounding which requires the model to synergize reasoning and grounding. We further introduce a new benchmark ScanReason, which comprises five reasoning types (spatial reasoning, functional reasoning, logical reasoning, emotional reasoning, and safety reasoning) for the task of 3D reasoning grounding in 3D scenes. 2) We design a new framework ReGround3D with a visual-centric reasoning module and a 3D grounding module with geometry-enhanced look-back. We further introduce a Chain-of-Grounding mechanism to boost the 3D reasoning grounding ability with a chain of interleaved reasoning and grounding steps. 3) Extensive experiments demonstrate the effectiveness of the our ReGround3D on the ScanReason benchmark for 3D reasoning grounding." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 408, + 237, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 408, + 237, + 422 + ], + "spans": [ + { + "bbox": [ + 132, + 408, + 237, + 422 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 433, + 482, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 433, + 482, + 625 + ], + "spans": [ + { + "bbox": [ + 130, + 433, + 482, + 625 + ], + "type": "text", + "content": "3D Vision and Language Learning 3D Vision-language learning (3D-VL) is garnering increasing attention, with many 3D-VL tasks focusing on how to connect the 3D world with natural language. Among them, 3D Question Answering (3D QA) [2,45] aims to enable models to provide text answers based on natural language questions. Situation Question Answering in 3D Scenes (SQA3D) [29] requires an agent to first understand its location based on a text description, then provide a reasonable answer based on the surrounding environment, which can be seen as an extension of 3D QA in the embodied AI area. 3D Visual Grounding [1,4,7,28,31,39] demands that models identify and locate target objects in a 3D scene based on given descriptions, outputting the objects' coordinates and 3D bounding boxes. These descriptions usually explicitly rely on the objects' attributes and their spatial relationships. 3D Dense Captioning [5,8,14,16,25,36,46,49] requires models to output a series of object coordinates and corresponding scene-based descriptions based on a given scene. Different from these 3D-VL tasks, the questions in 3D reasoning grounding could be more implicit and complex." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 640, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 640, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 640, + 481, + 666 + ], + "type": "text", + "content": "3D Visual Grounding The task of 3D visual grounding is aimed at localizing the objects that are explicitly referred to by free-form guided language" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 397, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 397, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 397, + 91, + 447, + 101 + ], + "type": "text", + "content": "ScanReason" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "content": "expressions in the 3D scene. Inspired by the success of transformers in natural language processing, recent 3D visual grounding approaches [12,17,24,39,50,53] have started to adopt transformer [35] architectures for handling the complex relationships between language descriptions and 3D visual data. These methods leverage the self-attention mechanism of transformers to dynamically weigh the importance of different parts of the input data, facilitating a more effective grounding of textual descriptions in the 3D environment. Recent method [3] proposes a Chain-of-Thoughts module that predicts a chain of anchor objects that are subsequently utilized to localize the final target object. Compared with 3D visual grounding, our proposed 3D reasoning grounding requires the model to reason the complex question, ground target objects, and give the explanation at the same time." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 275, + 482, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 275, + 482, + 526 + ], + "spans": [ + { + "bbox": [ + 130, + 275, + 482, + 526 + ], + "type": "text", + "content": "Multi-modal Large Language Models Recently, there has been an increasing effort to extend the powerful complex reasoning and world knowledge capabilities of LLMs [18,34,47] to other modalities [6,9,10,19,26,44,51,52]. Among these works, some have aimed to enable LLMs to understand the 3D world. [20, 40] focus on delving into LLMs' ability to comprehend 3D objects, which can not be directly applied to 3D scenes. 3D-LLM [21] is the pioneering work that incorporates the 3D scene into LLM to carry out general 3D understanding tasks. However, by using 3D features constructed through projecting the 2D features of multi-view 2D images extracted by pre-trained 2D Vision-Language Models into 3D space, 3D-LLM struggles to directly capture the complex spatial relationships between objects and the structure of 3D scenes. [13,27] directly extract 3D features from the reconstructed 3D point cloud and support multi-modal visual prompts (text, images, 3D objects) in an MLLM. To alleviate the difficulty LLMs face in understanding complex 3D scenes, [22] choose to first explicitly segment the objects in the 3D scene and then perform multi-stage object-aware scenetest alignment to achieve 3D scene understanding. However, due to the lack of large-scale 3D-language alignment data and the intricate content of 3D scenes, although current MLLMs can achieve favorable performance in 3D scene understanding tasks, their localization performance is still significantly behind the 3D localization specialists. Our approach seeks to address this issue by introducing a 3D grounding model to enhance the localization capability of MLLMs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 541, + 297, + 554 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 541, + 297, + 554 + ], + "spans": [ + { + "bbox": [ + 131, + 541, + 297, + 554 + ], + "type": "text", + "content": "3 ScanReason Benchmark" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 564, + 316, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 564, + 316, + 576 + ], + "spans": [ + { + "bbox": [ + 131, + 564, + 316, + 576 + ], + "type": "text", + "content": "3.1 3D Reasoning Grounding Task" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Given a 3D scene and an implicit free-form query, 3D reasoning grounding requires the model to predict the 3D bounding boxes of the target objects as well as the textual answers and explanations. As shown in Fig. 2, different from traditional 3D visual grounding, the queries of 3D reasoning grounding are implicit and complex, requiring strong reasoning, commonsense, and world knowledge. The number of target objects in 3D reasoning grounding is flexible and any object satisfying the requirements should be considered as the target object." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "C. Zhu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 115, + 286, + 220 + ], + "blocks": [ + { + "bbox": [ + 134, + 115, + 286, + 220 + ], + "lines": [ + { + "bbox": [ + 134, + 115, + 286, + 220 + ], + "spans": [ + { + "bbox": [ + 134, + 115, + 286, + 220 + ], + "type": "image", + "image_path": "8f87ed3d7733da89da0d9d3b3bf4ad2f2ebd2e4c9ff0528c031dc4fad4c9d162.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 229, + 480, + 274 + ], + "lines": [ + { + "bbox": [ + 130, + 229, + 480, + 274 + ], + "spans": [ + { + "bbox": [ + 130, + 229, + 480, + 274 + ], + "type": "text", + "content": "Fig. 2: The left side figure shows the overall of our ScanReason dataset. For each reasoning category, we designed different prompts to generate corresponding questions. And the right side figure shows the differences between the traditional 3D visual grounding task and our proposed 3D reasoning grounding task." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 293, + 129, + 480, + 216 + ], + "blocks": [ + { + "bbox": [ + 293, + 129, + 480, + 216 + ], + "lines": [ + { + "bbox": [ + 293, + 129, + 480, + 216 + ], + "spans": [ + { + "bbox": [ + 293, + 129, + 480, + 216 + ], + "type": "image", + "image_path": "3513ff74a02228af403122d8bdd0ee082eeb2998727007dc98136604b0aed7d1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 300, + 239, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 239, + 312 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 239, + 312 + ], + "type": "text", + "content": "3.2 Question Types" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 326, + 480, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 326, + 480, + 397 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 480, + 397 + ], + "type": "text", + "content": "To comprehensively evaluate the 3D reasoning grounding abilities, we define 5 types of questions depending on which type of reasoning is required. Spatial reasoning and functional reasoning require a fundamental understanding of the 3D physical world, while logical reasoning, emotional reasoning, and safety reasoning are high-level reasoning skills built upon the two fundamental reasoning abilities to address user-centric real-world applications, as shown in Fig. 2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 399, + 480, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 480, + 446 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 480, + 446 + ], + "type": "text", + "content": "Spatial Reasoning measures models' understanding of 3D spatial relationships among objects in 3D scene. It encompasses the ability to comprehend the layout of the 3D scene and the 3D location of objects within it, which could serve as the foundation for navigating or planning movements in the 3D environment." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 448, + 480, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 448, + 480, + 531 + ], + "spans": [ + { + "bbox": [ + 130, + 448, + 480, + 531 + ], + "type": "text", + "content": "Functional Reasoning involves understanding and inferring the purpose, function, or affordance of objects within the 3D scene. For example, functional reasoning allows an embodied agent to recognize that a chair is for sitting, a lamp is for lighting, and a refrigerator is for storing food at low temperatures. Such understanding enables the embodied agent to assist users and to perform complex tasks more effectively (e.g., turning on a lamp when the room gets dark, or navigating to a refrigerator to fetch a drink)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 533, + 480, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 480, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 480, + 604 + ], + "type": "text", + "content": "Logical Reasoning allows an embodied agent to not only understand its environment but also to interact with it in a goal-directed manner. For example, given a question shown in Fig. 2: \"If I'm cooking dinner in the kitchen, where is the nearest place to throw the rubbish?\", an agent needs to use such reasoning ability to infer the location of objects (in this question, a rubbish bin) based on their function and spatial relationships under the specific setting (the kitchen)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "type": "text", + "content": "Emotional Reasoning plays a critical role in human-robot interaction, where the target objects are determined by understanding human emotions, preferences, and behavioral patterns. This ability makes the embodied agents more attuned to the emotional and psychological states of humans, allowing them to provide more personalized, empathetic, and contextually appropriate responses" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 397, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 397, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 397, + 91, + 447, + 100 + ], + "type": "text", + "content": "ScanReason" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "content": "and solutions, such as: \"I'm very sad, is there something that can make me feel happy?\" shown in 2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 480, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 480, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 480, + 187 + ], + "type": "text", + "content": "Safety Reasoning focuses on preventing harm and ensuring the well-being of humans in the 3D environment. It requires the embodied agent to identify and assess the risk and make safety-aware decisions, such as: \"Where should first aid kits be placed for easy access but out of reach of children?\"" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 205, + 367, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 205, + 367, + 217 + ], + "spans": [ + { + "bbox": [ + 131, + 205, + 367, + 217 + ], + "type": "text", + "content": "3.3 Automatic Data Annotation with GPT-4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 225, + 479, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 225, + 479, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 225, + 479, + 297 + ], + "type": "text", + "content": "We leverage the 3D scenes and bounding box annotations from the EmbodiedScan dataset [37] and apply GPT-4 [30] to automatically generate question-answer-location pairs for the five question types respectively. Specifically, we provide GPT-4 with the categories and bounding box locations of all objects in the scene, and ask GPT-4 to generate questions and answers with the target object ids of the provided objects. The details can be found in the Appendix." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 297, + 482, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 482, + 381 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 482, + 381 + ], + "type": "text", + "content": "In total, our ScanReason dataset consists of 12929 complex reasoning question-answer-location pairs from 1456 scenes, which are split into 11455 training pairs and 1474 validation pairs. All 1474 validation question-answer pairs have been manually verified, including 342 spatial reasoning questions, 287 functional reasoning questions, 581 logical reasoning questions, 211 emotional questions, and 53 safety reasoning questions. We provide detailed statistics and more examples of our dataset in the Appendix." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 399, + 252, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 399, + 252, + 410 + ], + "spans": [ + { + "bbox": [ + 131, + 399, + 252, + 410 + ], + "type": "text", + "content": "3.4 Evaluation Metric" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 418, + 479, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 479, + 478 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 479, + 478 + ], + "type": "text", + "content": "To evaluate the accuracy of predicted objects and their locations for a flexible number of ground-truth objects, we follow the evaluation metric of 3D object detection tasks. Specifically, we adopt Acc@kIoU as our metric, where " + }, + { + "bbox": [ + 130, + 418, + 479, + 478 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 418, + 479, + 478 + ], + "type": "text", + "content": " is the threshold for the Intersection of Union (IoU) between positive predictions and ground truths. We evaluate the performance under " + }, + { + "bbox": [ + 130, + 418, + 479, + 478 + ], + "type": "inline_equation", + "content": "k = 0.25" + }, + { + "bbox": [ + 130, + 418, + 479, + 478 + ], + "type": "text", + "content": " in our experiments." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 497, + 202, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 497, + 202, + 509 + ], + "spans": [ + { + "bbox": [ + 131, + 497, + 202, + 509 + ], + "type": "text", + "content": "4 Method" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 521, + 479, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 479, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 479, + 628 + ], + "type": "text", + "content": "Solving the task of 3D reasoning grounding requires the synergization of the perception, reasoning, and grounding capability of the embodied agent. Intuitively, we can first conduct reasoning based on implicit instruction such as \"where should first aid kits be placed?\" and the visual environment. The reasoning process provides us with information about the rough location and semantics of the object we are looking for. Then keeping that information in mind, we look back to the 3D environment to precisely locate the object. For complex scenarios, alternate reasoning and looking back are required for multiple rounds until we obtain the final answer." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 630, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 479, + 665 + ], + "type": "text", + "content": "Inspired by this intuition, we propose ReGround3D, consisting of a visual-centric reasoning module and a 3D grounding module with geometry-enhanced look-back, as illustrated in Fig. 3. The visual-centric reasoning module (Sec. 4.1)" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "C. Zhu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "performs joint reasoning of language instruction and visual scene, and predicts a special " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\langle \\mathsf{LOC} \\rangle" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " token representing the grounding information. The 3D grounding module (Sec. 4.2) looks back to the original 3D scene with comprehensive geometry information and fine-grained details. It takes the hidden embedding of the " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\langle \\mathsf{LOC} \\rangle" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " token containing grounding-related information from the 3D features and eventually predicts the 3D locations of the target objects. Furthermore, we propose a Chain-of-Grounding mechanism (CoG) (Sec. 4.3), i.e., a chain of interleaved reasoning and grounding steps, to further synergize the grounding and reasoning capability for the 3D reasoning grounding task, as illustrated in Fig. 4." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 243, + 289, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 243, + 289, + 255 + ], + "spans": [ + { + "bbox": [ + 132, + 243, + 289, + 255 + ], + "type": "text", + "content": "4.1 Visual-Centric Reasoning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 265, + 481, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 265, + 481, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 265, + 481, + 361 + ], + "type": "text", + "content": "Due to the complexity of the 3D scene and user instructions, particularly given the implicit intention behind the human instruction, ReGround3D starts with a visual-centric reasoning module that can perceive the scene, comprehend the human instructions, and conduct joint reasoning of 3D scene and instructions. We believe the reasoning process eventually implies the grounding intention, i.e., implicitly encodes the information indicating the target object to solve the task. Thus, we design the visual-centric reasoning module to predict grounding queries for localizing the target objects in the following 3D grounding module." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 361, + 482, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 482, + 502 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 482, + 502 + ], + "type": "text", + "content": "Specifically, for simplicity, we leverage 3D-LLM to serve as the visual-centric reasoning module because of its strong reasoning abilities inherited from the LLM. Based on the BLIP2 architecture [26], 3D-LLM uses pre-trained image encoders to extract multi-view 2D image features and back-project them into 3D spaces. The visual features are encoded by the Q-Former to produce 32 tokens as the visual input to the LLM. By leveraging the pre-trained image encoders and the Q-Former, the visual tokens encode rich semantics but lack the 3D structures, spatial interactions, and fine-grained details. Therefore, instead of directly predicting the object locations by the 3D-LLM, we ask the 3D-LLM to predict the feature representation as the output of the reasoning process, and the predicted feature is further used to ground the target object in the grounding module." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 504, + 482, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 504, + 482, + 553 + ], + "spans": [ + { + "bbox": [ + 130, + 504, + 482, + 553 + ], + "type": "text", + "content": "To enable the prediction of the grounding feature, we expand the original vocabulary of the 3D-LLM with a special " + }, + { + "bbox": [ + 130, + 504, + 482, + 553 + ], + "type": "inline_equation", + "content": "\\langle \\mathsf{L0C} \\rangle" + }, + { + "bbox": [ + 130, + 504, + 482, + 553 + ], + "type": "text", + "content": " token. The " + }, + { + "bbox": [ + 130, + 504, + 482, + 553 + ], + "type": "inline_equation", + "content": "\\langle \\mathsf{L0C} \\rangle" + }, + { + "bbox": [ + 130, + 504, + 482, + 553 + ], + "type": "text", + "content": " token is laden with the contextual scene and the target object information which can guide the 3D grounding module to accurately localize target objects." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 572, + 424, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 572, + 424, + 585 + ], + "spans": [ + { + "bbox": [ + 132, + 572, + 424, + 585 + ], + "type": "text", + "content": "4.2 3D Grounding with Geometry-Enhanced Look-Back" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Once obtaining the token, ReGround3D extracts the last-layer embedding " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "h_{loc}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " of the token and sends it into the 3D grounding module to predict the 3D bounding boxes. The 3D grounding module is devised with a \"look-back\" mechanism which allows the model to access the 3D geometry and fine-grained details from a 3D point cloud encoder. The fine-grained geometry-enhanced 3D visual features and " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "h_{loc}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " are sent into a query selection module to retrieve the most" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 397, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 397, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 397, + 91, + 447, + 101 + ], + "type": "text", + "content": "ScanReason" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 116, + 478, + 237 + ], + "blocks": [ + { + "bbox": [ + 138, + 116, + 478, + 237 + ], + "lines": [ + { + "bbox": [ + 138, + 116, + 478, + 237 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 478, + 237 + ], + "type": "image", + "image_path": "a63d7526e0e73311efac638837eb1ad1a53e0edcda7dcf7e3c1305227a8d3141.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 247, + 482, + 291 + ], + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 291 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 291 + ], + "type": "text", + "content": "Fig. 3: The pipeline of ReGround3D. Given the 3D scene and human instruction, the visual reasoning module first performs joint 3D scene and instruction reasoning, and then guide the 3D grounding module to look-back the 3D scene and perform target object location." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 319, + 482, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 482, + 343 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 482, + 343 + ], + "type": "text", + "content": "relevant object features. Those features are further decoded into 3D bounding boxes with the 3D box decoder." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 346, + 482, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 482, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 482, + 430 + ], + "type": "text", + "content": "3D Visual Encoder Unlike the 2D image encoder used for 3D-LLM, the 3D visual encoder directly extracts features from 3D point clouds to capture more geometric and spatial information about the 3D structures and fine-grained details. The powerful 3D visual encoder which captures comprehensive geometry, structure, layout, and fine-grained information is critical to accurate 3D grounding. Subsequently, the 3D features " + }, + { + "bbox": [ + 130, + 346, + 482, + 430 + ], + "type": "inline_equation", + "content": "f_{scene}" + }, + { + "bbox": [ + 130, + 346, + 482, + 430 + ], + "type": "text", + "content": " produced by the 3D visual encoder and the grounding feature " + }, + { + "bbox": [ + 130, + 346, + 482, + 430 + ], + "type": "inline_equation", + "content": "h_{loc}" + }, + { + "bbox": [ + 130, + 346, + 482, + 430 + ], + "type": "text", + "content": " from 3D-LLM are sent to the query selection module." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 433, + 483, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 433, + 483, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 433, + 483, + 506 + ], + "type": "text", + "content": "Query Selection Module We adopt a cross-attention mechanism, where we treat " + }, + { + "bbox": [ + 130, + 433, + 483, + 506 + ], + "type": "inline_equation", + "content": "f_{\\text{scene}}" + }, + { + "bbox": [ + 130, + 433, + 483, + 506 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 433, + 483, + 506 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 130, + 433, + 483, + 506 + ], + "type": "text", + "content": " (Query), and " + }, + { + "bbox": [ + 130, + 433, + 483, + 506 + ], + "type": "inline_equation", + "content": "h_{\\text{loc}}" + }, + { + "bbox": [ + 130, + 433, + 483, + 506 + ], + "type": "text", + "content": " as both " + }, + { + "bbox": [ + 130, + 433, + 483, + 506 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 433, + 483, + 506 + ], + "type": "text", + "content": " (Key) and " + }, + { + "bbox": [ + 130, + 433, + 483, + 506 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 433, + 483, + 506 + ], + "type": "text", + "content": " (Value), to implicitly obtain a feature-level reasoning activation heatmap. During this scene look-back process, this module roughly locates scene features that have a high response to the token. We then select the " + }, + { + "bbox": [ + 130, + 433, + 483, + 506 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 433, + 483, + 506 + ], + "type": "text", + "content": " most relevant features as the object query " + }, + { + "bbox": [ + 130, + 433, + 483, + 506 + ], + "type": "inline_equation", + "content": "f_{\\text{query}}" + }, + { + "bbox": [ + 130, + 433, + 483, + 506 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 507, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 507, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 507, + 482, + 567 + ], + "type": "text", + "content": "3D Box Decoder is a classical transformer decoder, which consists of M transformer decoder layers. In each decoder layer, the object queries " + }, + { + "bbox": [ + 130, + 507, + 482, + 567 + ], + "type": "inline_equation", + "content": "f_{query}" + }, + { + "bbox": [ + 130, + 507, + 482, + 567 + ], + "type": "text", + "content": " go through the text feature " + }, + { + "bbox": [ + 130, + 507, + 482, + 567 + ], + "type": "inline_equation", + "content": "h_{loc}" + }, + { + "bbox": [ + 130, + 507, + 482, + 567 + ], + "type": "text", + "content": " cross-attention layer and scene feature " + }, + { + "bbox": [ + 130, + 507, + 482, + 567 + ], + "type": "inline_equation", + "content": "f_{scene}" + }, + { + "bbox": [ + 130, + 507, + 482, + 567 + ], + "type": "text", + "content": " cross-attention layer. Finally, the prediction head takes the updated object queries as input and predicts the final 3D locations and matching score." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Discussion In comparison to previous works [11,13,15,21] that directly predicts the bounding boxes by the LLM, the extra grounding module has the following advantages: 1) Based on the MLLM reasoning results, it allows the grounding module to perceive the scene again and focus on the region under the implicit guidance of MLLM, adapted to the user queries and the reasoning results. 2) The scene representation perceived in the grounding module can be more precise and fine-grained, which is complementary to the visual-centric reasoning module. 3) The two-step reasoning-grounding pipeline is flexible and can generalized" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "C. Zhu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 479, + 242 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 479, + 242 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 479, + 242 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 479, + 242 + ], + "type": "image", + "image_path": "7edbca3f9be2f8c93f5dd8fc66e31685c5b11300286c4d82ea9516bfe80a1150.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 181, + 251, + 432, + 263 + ], + "lines": [ + { + "bbox": [ + 181, + 251, + 432, + 263 + ], + "spans": [ + { + "bbox": [ + 181, + 251, + 432, + 263 + ], + "type": "text", + "content": "Fig. 4: Illustration of Chain-of-Grounding (CoG) Mechanism" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 285, + 480, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 480, + 309 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 480, + 309 + ], + "type": "text", + "content": "to other types of predicting formats such as segmentation masks (by simply replacing the 3D box decoder with a 3D mask decoder)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 325, + 322, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 325, + 322, + 337 + ], + "spans": [ + { + "bbox": [ + 131, + 325, + 322, + 337 + ], + "type": "text", + "content": "4.3 Chain-of-Grounding Mechanism" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 343, + 482, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 343, + 482, + 474 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 482, + 474 + ], + "type": "text", + "content": "The existing design conducts the reasoning and grounding process sequentially, i.e., the reasoning process is finished before grounding. We argue that the grounding results can also facilitate the reasoning process, especially for those requiring spatial information. Thus, to further synergize the reasoning and grounding process, inspired by chain-of-thought (CoT) [38], we propose Chain-of-Grounding (CoG), which introduces a chain of interleaved steps of reasoning and grounding to find the targeted objects during inference, as shown in Fig. 4. Such a process allows the model to actively find relevant objects that help solve the problem, and then conduct reasoning with the assistance of the additional information of these relevant objects so that the model can more precisely find the target objects." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 475, + 482, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 475, + 482, + 608 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 482, + 608 + ], + "type": "text", + "content": "Specifically, given a question provided by users, CoG translates it into another question of finding the explicitly mentioned objects in the original question. The generated new question is sent into ReGround3D to ground the objects mentioned in the original question in the 3D scene with corresponding confidence scores. An object can be seen as a successfully located object when its confidence score is above the threshold, and the located object information could serve as explicit guidance for 3D-LLM in the next reasoning stage. As shown in Fig. 4, after obtaining the 3D information of objects explicitly mentioned in the original question, the located object information is inserted to update the question. The updated question is then sent to ReGround3D to perform reasoning and grounding to output the target object locations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 623, + 256, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 623, + 256, + 635 + ], + "spans": [ + { + "bbox": [ + 131, + 623, + 256, + 635 + ], + "type": "text", + "content": "4.4 Instruction Tuning" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": "Training Objective We use the pre-trained weights of 3D-LLM as the initialization for the visual-centric reasoning module. Except for freezing the 3D" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 397, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 397, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 397, + 91, + 447, + 100 + ], + "type": "text", + "content": "ScanReason" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 136, + 480, + 237 + ], + "blocks": [ + { + "bbox": [ + 139, + 114, + 473, + 126 + ], + "lines": [ + { + "bbox": [ + 139, + 114, + 473, + 126 + ], + "spans": [ + { + "bbox": [ + 139, + 114, + 473, + 126 + ], + "type": "text", + "content": "Table 1: We list part of data templates used to train ReGround3D for each task." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 136, + 480, + 237 + ], + "lines": [ + { + "bbox": [ + 134, + 136, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 134, + 136, + 480, + 237 + ], + "type": "table", + "html": "
Task NameText InstructionsOutput Type TemplatesExpected Output
3D Visual Grounding<scene> Here is a description Please answer the question only with outabout an object: "<expr>", whereis the object in the 3D scene?Please answer the question only with out-put 3D box prediction(s).It is <LOC>.
3D/Spatial Question Answering<scene> Answer the question Please answer the question only with text,do not output 3D box prediction(s).<question>Please answer the question only with text,do not output 3D box prediction(s).<answer>.
<scene><situation>, <question>Please answer the question only with text,do not output 3D box prediction(s).<answer>.
<scene><question>Please answer the question only with text,do not output 3D box prediction(s).<answer>.
3D Object Detection<scene> Where is the <category> Please answer the question only with out-in this 3D scene? Put 3D box prediction(s).Sure, <LOC>.
3D Reasoning Grounding<scene> Answer the question: Please answer the question with text andoutput 3D box prediction(s).Sure, <LOC>, <reason>
", + "image_path": "73f7715e84ac56e717b9b4730eae7140c79d3bb296692c8441fa68f9dad71519.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 256, + 480, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 256, + 480, + 316 + ], + "spans": [ + { + "bbox": [ + 130, + 256, + 480, + 316 + ], + "type": "text", + "content": "visual encoder pre-trained on [37], the rest of the parameters in the visual-centric reasoning module and 3D grounding module in our framework are trained in an end-to-end manner. The training supervision is a weighted sum of the next token prediction loss from 3D-LLM and the 3D detection loss from the 3D grounding module." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 249, + 327, + 481, + 340 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 327, + 481, + 340 + ], + "spans": [ + { + "bbox": [ + 249, + 327, + 481, + 340 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda_ {\\text {t e x t}} \\mathcal {L} _ {\\text {t e x t}} + \\lambda_ {\\text {d e t}} \\mathcal {L} _ {\\text {d e t}} \\tag {1}", + "image_path": "d461179be80f02d44d491633d05b8ab429e85937e01dfdedd7ba50f06520d0f5.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 343, + 320, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 320, + 356 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 320, + 356 + ], + "type": "text", + "content": "The 3D detection loss is defined following:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 222, + 367, + 481, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 367, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 222, + 367, + 481, + 380 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {d e t}} = \\lambda_ {I O U} \\mathcal {L} _ {I O U} + \\lambda_ {\\text {c o n t r a s t}} \\mathcal {L} _ {\\text {c o n t r a s t}} \\tag {2}", + "image_path": "710662b999f9fcd862b6b32daa8f95760ac1f61051d5dd1cfcc26ab3fe2a1a51.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 394, + 482, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 394, + 482, + 551 + ], + "spans": [ + { + "bbox": [ + 130, + 394, + 482, + 551 + ], + "type": "text", + "content": "Instruction Tuning Dataset We load the pre-trained weights of 3D-LLM and the 3D visual encoder, and finetune the LoRA of 3D-LLM, the query selection module, and the 3D box decoder with an instruction tuning dataset. To construct the instruction tuning dataset, we reformulate the data annotations from existing 3D datasets into question-answer or question-answer-bounding-box pairs. Specifically, the 3D visual grounding data from ScanRefer [7], SR3D, NR3D [1] and the 3D object detection data from EmbodiedScan [37] are formulated into question-answer-bbox pairs, and the 3D/spatial question answering data from SR3D [1], CLEVR3D [41], SQA3D [29] are formulated into question-answer pairs without bounding boxes. The information shown in Tab. 1 illustrates how we unify the instruction and output with task-specific templates. More details can be found in the Appendix. The reformulated data combined with our proposed ScanReason dataset, serve as the instruction tuning dataset." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 566, + 224, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 566, + 224, + 580 + ], + "spans": [ + { + "bbox": [ + 132, + 566, + 224, + 580 + ], + "type": "text", + "content": "5 Experiment" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 588, + 279, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 588, + 279, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 279, + 601 + ], + "type": "text", + "content": "5.1 Implementation Details" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 605, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 480, + 666 + ], + "type": "text", + "content": "Network Architecture For the 3D grounding module, we adopt the pretrained point cloud encoder as the 3D visual encoder. During the training stage, We use LoRA to efficiently finetune the 3D-LLM to preserve the original 3D scene understanding capability and reduce the computation costs. The number of object queries " + }, + { + "bbox": [ + 130, + 605, + 480, + 666 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 605, + 480, + 666 + ], + "type": "text", + "content": " in the query selection module is set to 256." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "C. Zhu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 134, + 115, + 482, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 115, + 482, + 186 + ], + "spans": [ + { + "bbox": [ + 134, + 115, + 482, + 186 + ], + "type": "text", + "content": "Training Parameters The training is done on 8 NVIDIA A100 GPUs. We adopt the AdamW optimizer with a learning rate of 3e-4 and use a learning rate scheduler WarmupDecayLR with the warmup steps of 100. The total batch size is set to be 16. The loss weight parameters " + }, + { + "bbox": [ + 134, + 115, + 482, + 186 + ], + "type": "inline_equation", + "content": "\\lambda_{text}" + }, + { + "bbox": [ + 134, + 115, + 482, + 186 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 134, + 115, + 482, + 186 + ], + "type": "inline_equation", + "content": "\\lambda_{det}" + }, + { + "bbox": [ + 134, + 115, + 482, + 186 + ], + "type": "text", + "content": " in total loss " + }, + { + "bbox": [ + 134, + 115, + 482, + 186 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 134, + 115, + 482, + 186 + ], + "type": "text", + "content": " are set to 1.0 and 1.0, respectively, and the weight " + }, + { + "bbox": [ + 134, + 115, + 482, + 186 + ], + "type": "inline_equation", + "content": "\\lambda_{IOU}" + }, + { + "bbox": [ + 134, + 115, + 482, + 186 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 134, + 115, + 482, + 186 + ], + "type": "inline_equation", + "content": "\\lambda_{contrast}" + }, + { + "bbox": [ + 134, + 115, + 482, + 186 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 134, + 115, + 482, + 186 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{det}}" + }, + { + "bbox": [ + 134, + 115, + 482, + 186 + ], + "type": "text", + "content": " are set to 1.0 and 1.0." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 209, + 261, + 221 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 209, + 261, + 221 + ], + "spans": [ + { + "bbox": [ + 134, + 209, + 261, + 221 + ], + "type": "text", + "content": "5.2 Results Comparison" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 233, + 481, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 233, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 133, + 233, + 481, + 460 + ], + "type": "text", + "content": "Evaluation on 3D Visual Grounding In order to verify the superiority of our model in grounding ability and facilitate comparison between current models, we report the explicit grounding performance on the existing 3D visual grounding task. Since the evaluation settings of Nr3D and Sr3D [1] are based on ground-truth object proposals, while ScanRefer [7] requires models to output 3D bounding boxes, we choose ScanRefer as our benchmark for comparison. We divide the existing methods into two categories, one is the grounding model designed specifically for the 3D visual grounding task, and the other is generalist MLLMs which can understand a variety of 3D vision-language tasks. The original 3D-LLM embeds 3D locations in the vocabularies and represents the grounded 3D bounding boxes by a sequence of discrete location tokens. However, since the fine-tuned model of 3D-LLM on ScanRefer and related location tokens implementation are not accessible, we adapt 3D-LLM to directly output 3D numerical coordinates representing 3D bounding boxes by fine-tuning the pre-trained model on our reformulated 3D visual grounding data, denotes as 3D-LLM (vg). As shown in Tab. 2, current generalist MLLM models still lag behind the specialist models in terms of grounding ability. By incorporating the 3D grounding module into MLLM, ReGround3D shows the SOTA performance on the traditional 3D visual grounding task." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 204, + 515, + 411, + 639 + ], + "blocks": [ + { + "bbox": [ + 133, + 483, + 481, + 503 + ], + "lines": [ + { + "bbox": [ + 133, + 483, + 481, + 503 + ], + "spans": [ + { + "bbox": [ + 133, + 483, + 481, + 503 + ], + "type": "text", + "content": "Table 2: Results on 3D visual grounding task among ReGround3D (ours) and existing methods." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 204, + 515, + 411, + 639 + ], + "lines": [ + { + "bbox": [ + 204, + 515, + 411, + 639 + ], + "spans": [ + { + "bbox": [ + 204, + 515, + 411, + 639 + ], + "type": "table", + "html": "
TypeMethodsAcc@0.25Acc@0.5
SpecialistsScanRefer [7]37.324.3
MVT [23]40.833.3
3DVG-Trans [48]45.934.5
ViL3DRel [12]47.937.7
BUTD-DETR [24]52.239.8
L3Det [50]52.840.2
Generalized MLLMsLLM-Grounder [42]17.15.3
3D-LLM [21]30.3-
3D-LLM(vg) [21]33.128.7
Chat3D-v2 [22]35.930.4
oursReGround3D53.141.1
", + "image_path": "1b26609c0c3c352419e104b42bfe4469cc0a3cf471d6640ac86429ef00682a88.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 397, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 397, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 397, + 91, + 447, + 100 + ], + "type": "text", + "content": "ScanReason" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 146, + 481, + 224 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 3: Results (Acc@0.25) on 3D reasoning grounding task among ReGround3D (ours) and existing methods." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 146, + 481, + 224 + ], + "lines": [ + { + "bbox": [ + 133, + 146, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 133, + 146, + 481, + 224 + ], + "type": "table", + "html": "
MethodsLLMSpatial FunctionalLogicalEmotional SafetyOverall
Mask3D [32] + InternLM2-7B [33]InternLM2-7B10.3436.129.988.218.9914.86
3D-LLM(vg) [21]FlanT5XL-3B18.3117.4210.978.126.3313.29
Chat3D-v2 [22]Vicuna-7B20.2118.3911.327.989.8814.98
ReGround3D*FlanT5XL-3B30.7629.818.6719.2217.1223.27
ReGround3DFlanT5XL-3B32.9836.2326.9923.1222.9828.98
ReGround3D (CoG)FlanT5XL-3B34.7136.7929.1124.0323.2130.62
", + "image_path": "60b3b170fe061f406e68d6a7b3d27e431f9f77b9520caac20eb8c891668725f3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 247, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 521 + ], + "type": "text", + "content": "Evaluation on 3D Reasoning Grounding The various visual grounding models rely on explicit text-object alignment in the input object expression to achieve localization, which falls to be applied to 3D reasoning grounding task. We performed a comparison between our proposed method ReGround3D and existing MLLM methods, including 3D-LLM(vg) and Chat3D-v2 [22]. Besides, inspired by Chat-3D v2, which first segments objects, then equips them with unique object identifiers to conduct effective object grounding, we set up a LLM-based 3D reasoning grounding baseline: We first segment the objects from the scene using a 3D instance segmentor [32], then convert the segmented object information including their categories and 3D bounding boxes into text as input of LLM (InternLM2-7B [33]). Besides, to better validate the performance of our model and ensure a fair comparison, we remove the ScanReason dataset from the training data, denoted as ReGround3D*. As shown in Tab. 3, the LLM-based reasoning method (Mask3D [32] + InternLM2-7B [33]) possesses a very strong functional reasoning ability, but struggles to understand of 3D spatial relationship. Additionally, irrespective of whether ScanReason is used in training, our model significantly outperforms the existing MLLMs. By synergizing the reasoning and grounding process utilizing the Chain-of-Grounding (CoG) mechanism, the 3D reasoning grounding performance of ReGround3D can be further improved (from 28.98 to 30.62), especially on the spatial reasoning and logical reasoning questions. The qualitative results comparison shown in Fig. 5 demonstrates the superiority of our method in reasoning human complex instruction based on the 3D scene." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 541, + 238, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 541, + 238, + 553 + ], + "spans": [ + { + "bbox": [ + 132, + 541, + 238, + 553 + ], + "type": "text", + "content": "5.3 Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 562, + 481, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 562, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 130, + 562, + 481, + 586 + ], + "type": "text", + "content": "In this section, we conduct an extensive ablation study to verify the effectiveness of each component." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Effectiveness of 3D Grounding Module In order to more comprehensively verify the effectiveness of the 3D grounding module we proposed, we step by step verify the performance changes from 3D-LLM to ReGround3D. Apart from the 3D-LLM(vg), we fine-tuned the 3D-LLM model respectively on full reformulated existing data and all instruction tuning data, including ScanReason, denoted as" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "C. Zhu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 146, + 126 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 146, + 126 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 146, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 146, + 126 + ], + "type": "image", + "image_path": "784f1f2f5b6f232eba3420d872b915d4273734ee0cbb2c82f4a2d907c77908c0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 149, + 116, + 298, + 125 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 116, + 298, + 125 + ], + "spans": [ + { + "bbox": [ + 149, + 116, + 298, + 125 + ], + "type": "text", + "content": "It's so cold outside, Where could I find my thick coat?" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 136, + 129, + 473, + 205 + ], + "blocks": [ + { + "bbox": [ + 136, + 129, + 473, + 205 + ], + "lines": [ + { + "bbox": [ + 136, + 129, + 473, + 205 + ], + "spans": [ + { + "bbox": [ + 136, + 129, + 473, + 205 + ], + "type": "image", + "image_path": "27dcfd88b158e59479e6576ac7425d522a093628a2d07ff552c4f872f11c92b2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 134, + 208, + 146, + 220 + ], + "blocks": [ + { + "bbox": [ + 134, + 208, + 146, + 220 + ], + "lines": [ + { + "bbox": [ + 134, + 208, + 146, + 220 + ], + "spans": [ + { + "bbox": [ + 134, + 208, + 146, + 220 + ], + "type": "image", + "image_path": "c9fa2146b542c300c78a46511bc63b7a9b3ae39054320d18b0a71467da5acf5e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 148, + 210, + 263, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 210, + 263, + 220 + ], + "spans": [ + { + "bbox": [ + 148, + 210, + 263, + 220 + ], + "type": "text", + "content": "Where can I find paper to wipe my hands?" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 135, + 224, + 479, + 293 + ], + "blocks": [ + { + "bbox": [ + 135, + 224, + 479, + 293 + ], + "lines": [ + { + "bbox": [ + 135, + 224, + 479, + 293 + ], + "spans": [ + { + "bbox": [ + 135, + 224, + 479, + 293 + ], + "type": "image", + "image_path": "4511c469625c1467fc4835c6fcd6ab695eebe2823a680f18daa75884368e1258.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 134, + 295, + 146, + 307 + ], + "blocks": [ + { + "bbox": [ + 134, + 295, + 146, + 307 + ], + "lines": [ + { + "bbox": [ + 134, + 295, + 146, + 307 + ], + "spans": [ + { + "bbox": [ + 134, + 295, + 146, + 307 + ], + "type": "image", + "image_path": "8d2c4f40c5021c3aa2fe6791c813035cbda4adb46952763b92b652089c94293b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 149, + 297, + 312, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 297, + 312, + 307 + ], + "spans": [ + { + "bbox": [ + 149, + 297, + 312, + 307 + ], + "type": "text", + "content": "It's dark here, can you find something to light up the room?" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 135, + 312, + 479, + 387 + ], + "blocks": [ + { + "bbox": [ + 135, + 312, + 479, + 387 + ], + "lines": [ + { + "bbox": [ + 135, + 312, + 479, + 387 + ], + "spans": [ + { + "bbox": [ + 135, + 312, + 479, + 387 + ], + "type": "image", + "image_path": "631c09b5775357e613af7d1fbcb51c92ffc52729ec5abae40113360c2e24115e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 396, + 482, + 441 + ], + "lines": [ + { + "bbox": [ + 131, + 396, + 482, + 441 + ], + "spans": [ + { + "bbox": [ + 131, + 396, + 482, + 441 + ], + "type": "text", + "content": "Fig. 5: Visualization comparison of 3D reasoning grounding capability between ReGround3D and 3D-LLM. Our method could achieve much more accurate grounding results which satisfy the implicit question intention, and give the corresponding explanation at the same time. More illustrations are given in the Appendix." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 472, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 472, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 131, + 472, + 482, + 533 + ], + "type": "text", + "content": "3D-LLM (full) and 3D-LLM (full+sr). All the tokens in answers of the data used for fine-tuning 3D-LLM have been converted to the numerical coordinates of the corresponding boxes. The results in Tab. 4a show that with the same tuning dataset, our ReGround3D achieves far superior performance to 3D-LLM (28.98 vs. 19.21) by introducing the 3D grounding module." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": "Effectiveness of Instruction Tuning Dataset Tab. 4b showcases the impact of different training data types on 3D visual grounding performance. 3D Object Detection (3D OD) provides the explicit 3D semantic category and visual alignment whereas 3D Question Answering (3D QA) data injects the basic 3D scene understanding ability into the model, which has certain benefits for the visual grounding ability of the model. In addition, we find that training with 3D reasoning grounding dataset can further improve the performance on 3D visual grounding." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 397, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 397, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 397, + 91, + 447, + 100 + ], + "type": "text", + "content": "ScanReason" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 145, + 291, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 145, + 291, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 145, + 291, + 171 + ], + "type": "text", + "content": "(a) When using the same tuning dataset, ReGround3D achieves far better performance on ScanReason than 3D-LLM." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 295, + 145, + 478, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 145, + 478, + 171 + ], + "spans": [ + { + "bbox": [ + 295, + 145, + 478, + 171 + ], + "type": "text", + "content": "(b) Ablation study on training data. We evaluate through the metric of accuracy on the val set of the ScanRefer dataset." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 142, + 180, + 279, + 247 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 478, + 126 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 478, + 126 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 478, + 126 + ], + "type": "text", + "content": "Table 4: Ablation study on effectiveness of 3D grounding module and training data" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 142, + 180, + 279, + 247 + ], + "lines": [ + { + "bbox": [ + 142, + 180, + 279, + 247 + ], + "spans": [ + { + "bbox": [ + 142, + 180, + 279, + 247 + ], + "type": "table", + "html": "
MethodsAcc@0.25
3D-LLM(vg) [21]13.29
3D-LLM(full) [21]15.31
3D-LLM(full+sr) [21]19.21
ReGround3D(full+sr)28.98
", + "image_path": "66f53d6621b98ac4cd6d90f381b3e91320cc13af5a304578bb256809e79520fa.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 320, + 180, + 451, + 253 + ], + "blocks": [ + { + "bbox": [ + 320, + 180, + 451, + 253 + ], + "lines": [ + { + "bbox": [ + 320, + 180, + 451, + 253 + ], + "spans": [ + { + "bbox": [ + 320, + 180, + 451, + 253 + ], + "type": "table", + "html": "
DatasetAcc@0.25Acc@0.5
VGODVQARD
✓✓19.314.2
✓✓48.737.6
✓✓49.238.1
✓✓✓51.839.4
✓✓✓✓53.141.1
", + "image_path": "efb50851875cbe0258e27622956b185a168af506f8a852d0eaf46a030f9db5a4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 275, + 482, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 275, + 482, + 419 + ], + "spans": [ + { + "bbox": [ + 130, + 275, + 482, + 419 + ], + "type": "text", + "content": "Discussion of CoG Mechanism While the CoG Mechanism boosts the performance with interleaved reasoning and grounding steps during inference, it uses the relevant object information explicitly presented in the question to help find the target objects. A natural question arises: if we input the information of all the objects instead of relevant objects into ReGround3D during CoG, will this make the model more accurately find the target objects? We first use the existing 3D Segmentor [32] to segment all objects in the scene, then update the original question using all the object 3D bounding boxes information according to the template in Sec. 4.3 and send into ReGround3D. However, experiments show that using all the object information will instead reduce the 3D reasoning grounding performance from 28.98 to 27.67. One possible reason could be the model's attention is dispersed by too many irrelevant objects." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 437, + 220, + 450 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 437, + 220, + 450 + ], + "spans": [ + { + "bbox": [ + 132, + 437, + 220, + 450 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 463, + 482, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 463, + 482, + 654 + ], + "spans": [ + { + "bbox": [ + 130, + 463, + 482, + 654 + ], + "type": "text", + "content": "This paper introduces a new 3D vision language learning task: 3D reasoning grounding, which requires the model to perform active reasoning over complex and implicit human instruction, localize the target objects, and give corresponding explanations. Furthermore, we propose ScanReason, a new dataset and benchmark to further unlock and thoroughly evaluate the 3D reasoning grounding capability. Based on this dataset, we propose a novel approach: ReGround3D, which utilizes the strong reasoning capability of MLLM guiding the 3D grounding module to obtain accurate object locations, and a Chain of Grounding (CoG) mechanism is presented to further boost the performance with interleaved reasoning and grounding steps during inference. We believe that our work will further the natural interaction between embodied agents and humans in open 3D environments. For the current ScanReason benchmark, we find that the questions in three high-level 3D reasoning categories may have overlaps. For a certain reasoning question, similar questions may appear in one or two other categories. We leave the problem as a future challenge for better reasoning grounding ability evaluation." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "C. Zhu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 482, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 482, + 178 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 482, + 178 + ], + "type": "text", + "content": "This work is supported in part by HKU Startup Fund, HKU Seed Fund for Basic Research, HKU Seed Fund for Translational and Applied Research, HKU IDS research Seed Fund, and HKU Fintech Academy R&D Funding." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 197, + 198, + 209 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 197, + 198, + 209 + ], + "spans": [ + { + "bbox": [ + 133, + 197, + 198, + 209 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 223, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 138, + 223, + 481, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 223, + 481, + 267 + ], + "spans": [ + { + "bbox": [ + 138, + 223, + 481, + 267 + ], + "type": "text", + "content": "1. Achlioptas, P., Abdelreheem, A., Xia, F., Elhoseiny, M., Guibas, L.: Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16. pp. 422-440. Springer (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 269, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 269, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 138, + 269, + 481, + 300 + ], + "type": "text", + "content": "2. Azuma, D., Miyanishi, T., Kurita, S., Kawanabe, M.: Scanqa: 3d question answering for spatial scene understanding. In: proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 19129-19139 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 301, + 481, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 301, + 481, + 334 + ], + "spans": [ + { + "bbox": [ + 138, + 301, + 481, + 334 + ], + "type": "text", + "content": "3. Bakr, E.M., Ayman, M., Ahmed, M., Slim, H., Elhoseiny, M.: Cot3dref: Chain-of-thoughts data-efficient 3d visual grounding. arXiv preprint arXiv:2310.06214 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 335, + 481, + 366 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 335, + 481, + 366 + ], + "spans": [ + { + "bbox": [ + 138, + 335, + 481, + 366 + ], + "type": "text", + "content": "4. Baroni, M.: Linguistic generalization and compositionality in modern artificial neural networks. Philosophical Transactions of the Royal Society B 375(1791), 20190307 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 368, + 481, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 368, + 481, + 411 + ], + "spans": [ + { + "bbox": [ + 138, + 368, + 481, + 411 + ], + "type": "text", + "content": "5. Cai, D., Zhao, L., Zhang, J., Sheng, L., Xu, D.: 3djcg: A unified framework for joint dense captioning and visual grounding on 3d point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16464-16473 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 412, + 481, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 412, + 481, + 443 + ], + "spans": [ + { + "bbox": [ + 138, + 412, + 481, + 443 + ], + "type": "text", + "content": "6. Chen, C., Qin, R., Luo, F., Mi, X., Li, P., Sun, M., Liu, Y.: Position-enhanced visual instruction tuning for multimodal large language models. arXiv preprint arXiv:2308.13437 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 445, + 481, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 445, + 481, + 488 + ], + "spans": [ + { + "bbox": [ + 138, + 445, + 481, + 488 + ], + "type": "text", + "content": "7. Chen, D.Z., Chang, A.X., Niefer, M.: Scanrefer: 3d object localization in rgb-d scans using natural language. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XX. pp. 202-221. Springer (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 489, + 481, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 489, + 481, + 521 + ], + "spans": [ + { + "bbox": [ + 138, + 489, + 481, + 521 + ], + "type": "text", + "content": "8. Chen, D.Z., Wu, Q., Nießner, M., Chang, A.X.: D3net: A speaker-listener architecture for semi-supervised dense captioning and visual grounding in rgb-d scans. arXiv preprint arXiv:2112.01551 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 522, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 522, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 522, + 481, + 555 + ], + "type": "text", + "content": "9. Chen, F., Han, M., Zhao, H., Zhang, Q., Shi, J., Xu, S., Xu, B.: X-LLM: Bootstrapping advanced large language models by treating multi-modalities as foreign languages. arXiv preprint arXiv:2305.04160 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 555, + 481, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 555, + 481, + 598 + ], + "spans": [ + { + "bbox": [ + 133, + 555, + 481, + 598 + ], + "type": "text", + "content": "10. Chen, J., Zhu, D., Shen, X., Li, X., Liu, Z., Zhang, P., Krishnamoorthi, R., Chandra, V., Xiong, Y., Elhoseiny, M.: Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 599, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 599, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 133, + 599, + 481, + 632 + ], + "type": "text", + "content": "11. Chen, K., Zhang, Z., Zeng, W., Zhang, R., Zhu, F., Zhao, R.: Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 632, + 481, + 665 + ], + "type": "text", + "content": "12. Chen, S., Guhur, P.L., Tapaswi, M., Schmid, C., Laptev, I.: Language conditioned spatial relation reasoning for 3d object grounding. Advances in Neural Information Processing Systems 35, 20522-20535 (2022)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 397, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 397, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 397, + 91, + 447, + 101 + ], + "type": "text", + "content": "ScanReason" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "13. Chen, S., Chen, X., Zhang, C., Li, M., Yu, G., Fei, H., Zhu, H., Fan, J., Chen, T.: Ll3da: Visual interactive instruction tuning for omni-3d understanding, reasoning, and planning. arXiv preprint arXiv:2311.18651 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 182 + ], + "type": "text", + "content": "14. Chen, S., Zhu, H., Chen, X., Lei, Y., Yu, G., Chen, T.: End-to-end 3d dense captioning with vote2cap-detr. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11124-11133 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 183, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 183, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 183, + 480, + 205 + ], + "type": "text", + "content": "15. Chen, T., Saxena, S., Li, L., Fleet, D.J., Hinton, G.: Pix2seq: A language modeling framework for object detection. arXiv preprint arXiv:2109.10852 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 205, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 205, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 134, + 205, + 480, + 237 + ], + "type": "text", + "content": "16. Chen, Z., Gholami, A., Nießner, M., Chang, A.X.: Scan2cap: Context-aware dense captioning in rgb-d scans. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3193-3203 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 238, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 238, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 133, + 238, + 480, + 270 + ], + "type": "text", + "content": "17. Chen, Z., Hu, R., Chen, X., Nießner, M., Chang, A.X.: Unit3d: A unified transformer for 3d dense captioning and visual grounding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 18109-18119 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 134, + 271, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 271, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 134, + 271, + 480, + 303 + ], + "type": "text", + "content": "18. Chung, H.W., Hou, L., Longpre, S., Zoph, B., Tay, Y., Fedus, W., Li, E., Wang, X., Dehghani, M., Brahma, S., et al.: Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 134, + 304, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 304, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 134, + 304, + 480, + 335 + ], + "type": "text", + "content": "19. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 336, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 336, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 133, + 336, + 480, + 369 + ], + "type": "text", + "content": "20. Han, J., Zhang, R., Shao, W., Gao, P., Xu, P., Xiao, H., Zhang, K., Liu, C., Wen, S., Guo, Z., et al.: Imagebind-llm: Multi-modality instruction tuning. arXiv preprint arXiv:2309.03905 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 369, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 369, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 133, + 369, + 480, + 402 + ], + "type": "text", + "content": "21. Hong, Y., Zhen, H., Chen, P., Zheng, S., Du, Y., Chen, Z., Gan, C.: 3d-llm: Injecting the 3d world into large language models. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 403, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 403, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 133, + 403, + 480, + 435 + ], + "type": "text", + "content": "22. Huang, H., Wang, Z., Huang, R., Liu, L., Cheng, X., Zhao, Y., Jin, T., Zhao, Z.: Chat-3d v2: Bridging 3d scene and large language models with object identifiers. arXiv preprint arXiv:2312.08168 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 435, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 435, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 133, + 435, + 480, + 468 + ], + "type": "text", + "content": "23. Huang, S., Chen, Y., Jia, J., Wang, L.: Multi-view transformer for 3d visual grounding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15524-15533 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 468, + 480, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 468, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 133, + 468, + 480, + 512 + ], + "type": "text", + "content": "24. Jain, A., Gkanatsios, N., Mediratta, I., Fragkiadaki, K.: Bottom up top down detection transformers for language grounding in images and point clouds. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVI. pp. 417-433. Springer (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 513, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 513, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 133, + 513, + 480, + 533 + ], + "type": "text", + "content": "25. Jiao, Y., Chen, S., Jie, Z., Chen, J., Ma, L., Jiang, Y.G.: More: Multi-order relation mining for dense captioning in 3d scenes. arXiv preprint arXiv:2203.05203 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 534, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 534, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 133, + 534, + 480, + 566 + ], + "type": "text", + "content": "26. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 567, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 567, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 133, + 567, + 480, + 600 + ], + "type": "text", + "content": "27. Li, M., Chen, X., Zhang, C., Chen, S., Zhu, H., Yin, F., Yu, G., Chen, T.: M3dbench: Let's instruct large models with multi-modal 3d prompts. arXiv preprint arXiv:2312.10763 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 601, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 601, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 133, + 601, + 480, + 643 + ], + "type": "text", + "content": "28. Luo, J., Fu, J., Kong, X., Gao, C., Ren, H., Shen, H., Xia, H., Liu, S.: 3d-sps: Single-stage 3d visual grounding via referred point progressive selection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16454-16463 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 643, + 480, + 665 + ], + "type": "text", + "content": "29. Ma, X., Yong, S., Zheng, Z., Li, Q., Liang, Y., Zhu, S.C., Huang, S.: Sqa3d: Situated question answering in 3d scenes. arXiv preprint arXiv:2210.07474 (2022)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "C. Zhu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 132, + 117, + 307, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 117, + 307, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 307, + 128 + ], + "type": "text", + "content": "30. OpenAI: Gpt-4 technical report (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 128, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 128, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 128, + 481, + 160 + ], + "type": "text", + "content": "31. Roh, J., Desingh, K., Farhadi, A., Fox, D.: Languagerefer: Spatial-language model for 3d visual grounding. In: Conference on Robot Learning. pp. 1046-1056. PMLR (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 161, + 481, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 481, + 204 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 481, + 204 + ], + "type": "text", + "content": "32. Schult, J., Engelmann, F., Hermans, A., Litany, O., Tang, S., Leibe, B.: Mask3d: Mask transformer for 3d semantic instance segmentation. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 8216-8223. IEEE (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 226 + ], + "type": "text", + "content": "33. Team, I.: Internl'm: A multilingual language model with progressively enhanced capabilities (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "type": "text", + "content": "34. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "type": "text", + "content": "35. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 281, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 481, + 303 + ], + "type": "text", + "content": "36. Wang, H., Zhang, C., Yu, J., Cai, W.: Spatiality-guided transformer for 3d dense captioning on point clouds. arXiv preprint arXiv:2204.10688 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "type": "text", + "content": "37. Wang, T., Mao, X., Zhu, C., Xu, R., Lyu, R., Li, P., Chen, X., Zhang, W., Chen, K., Xue, T., et al.: Embodiedscan: A holistic multi-modal 3d perception suite towards embodied ai. arXiv preprint arXiv:2312.16170 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "text", + "content": "38. Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q.V., Zhou, D., et al.: Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems 35, 24824-24837 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "type": "text", + "content": "39. Wu, Y., Cheng, X., Zhang, R., Cheng, Z., Zhang, J.: Eda: Explicit textdecoupling and dense alignment for 3d visual and language learning. arXiv preprint arXiv:2209.14941 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 402, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 481, + 435 + ], + "type": "text", + "content": "40. Xu, R., Wang, X., Wang, T., Chen, Y., Pang, J., Lin, D.: Pointllm: Empowering large language models to understand point clouds. arXiv preprint arXiv:2308.16911 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "type": "text", + "content": "41. Yan, X., Yuan, Z., Du, Y., Liao, Y., Guo, Y., Li, Z., Cui, S.: Clevr3d: Compositional language and elementary visual reasoning for question answering in 3d real-world scenes. arXiv preprint arXiv:2112.11691 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 468, + 481, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 500 + ], + "type": "text", + "content": "42. Yang, J., Chen, X., Qian, S., Madaan, N., Iyengar, M., Fouhey, D.F., Chai, J.: Llm-grounder: Open-vocabulary 3d visual grounding with large language model as an agent. arXiv preprint arXiv:2309.12311 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 501, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 481, + 533 + ], + "type": "text", + "content": "43. Yang, Z., Zhang, S., Wang, L., Luo, J.: Sat: 2d semantics assisted training for 3d visual grounding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1856-1866 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 533, + 481, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 481, + 566 + ], + "type": "text", + "content": "44. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 567, + 481, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 588 + ], + "type": "text", + "content": "45. Ye, S., Chen, D., Han, S., Liao, J.: 3d question answering. IEEE Transactions on Visualization and Computer Graphics (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 589, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 481, + 632 + ], + "type": "text", + "content": "46. Yuan, Z., Yan, X., Liao, Y., Guo, Y., Li, G., Cui, S., Li, Z.: X-trans2cap: Cross-modal knowledge transfer using transformer for 3d dense captioning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8563-8573 (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "47. Zhang, S., Roller, S., Goyal, N., Artetxe, M., Chen, M., Chen, S., Dewan, C., Diab, M., Li, X., Lin, X.V., et al.: Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068 (2022)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 397, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 397, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 397, + 91, + 447, + 100 + ], + "type": "text", + "content": "ScanReason" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 293 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "48. Zhao, L., Cai, D., Sheng, L., Xu, D.: 3dvg-transformer: Relation modeling for visual grounding on point clouds. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2928-2937 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 171 + ], + "type": "text", + "content": "49. Zhong, Y., Xu, L., Luo, J., Ma, L.: Contextual modeling for 3d dense captioning on point clouds. arXiv preprint arXiv:2210.03925 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 193 + ], + "type": "text", + "content": "50. Zhu, C., Zhang, W., Wang, T., Liu, X., Chen, K.: Object2scene: Putting objects in context for open-vocabulary 3d detection. arXiv preprint arXiv:2309.09456 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 481, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 225 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 225 + ], + "type": "text", + "content": "51. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "type": "text", + "content": "52. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "text", + "content": "53. Zhu, Z., Ma, X., Chen, Y., Deng, Z., Huang, S., Li, Q.: 3d-vista: Pre-trained transformer for 3d vision and text alignment. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2911-2921 (2023)" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "C. Zhu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/ScanTalk_ 3D Talking Heads from Unregistered Scans/34e5d7c9-005f-4b11-aa72-5d7f3ba885f0_content_list.json b/2024/ScanTalk_ 3D Talking Heads from Unregistered Scans/34e5d7c9-005f-4b11-aa72-5d7f3ba885f0_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..43ea74d3d2de41ee7572cfe2ecef37f9fd28cb25 --- /dev/null +++ b/2024/ScanTalk_ 3D Talking Heads from Unregistered Scans/34e5d7c9-005f-4b11-aa72-5d7f3ba885f0_content_list.json @@ -0,0 +1,1624 @@ +[ + { + "type": "text", + "text": "ScanTalk: 3D Talking Heads from Unregistered Scans", + "text_level": 1, + "bbox": [ + 338, + 140, + 666, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Federico Nocentini $^{1}$ , Thomas Besnier $^{2}$ , Claudio Ferrari $^{4}$ , Sylvain Arguillere $^{5}$ , Stefano Berretti $^{1}$ , and Mohamed Daoudi $^{2,3}$", + "bbox": [ + 251, + 214, + 750, + 246 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Media Integration and Communication Center (MICC), University of Florence, Italy federico.nocentini@unifi.it, stefano.berretti@unifi.it \n $^{2}$ Univ. Lille, CNRS, Centrale Lille, UMR 9189 CRISTAL, F-59000 Lille, France thomas.besnier@univ-lille.fr \n $^{3}$ IMT Nord Europe, Institut Mines-Télécom, Centre for Digital Systems mohamed.daoudi@imt-nord-europe.fr \n $^{4}$ Department of Architecture and Engineering University of Parma, Italy claudio.ferrari2@unipr.it \n $^{5}$ Univ. Lille, CNRS, UMR 8524 Laboratoire Paul Painlevé, Lille, F-59000, France sylvain,arguillere@univ-lille.fr", + "bbox": [ + 230, + 255, + 772, + 411 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Speech-driven 3D talking heads generation has emerged as a significant area of interest among researchers, presenting numerous challenges. Existing methods are constrained by animating faces with fixed topologies, wherein point-wise correspondence is established, and the number and order of points remains consistent across all identities the model can animate. In this work, we present ScanTalk, a novel framework capable of animating 3D faces in arbitrary topologies including scanned data. Our approach relies on the DiffusionNet architecture to overcome the fixed topology constraint, offering promising avenues for more flexible and realistic 3D animations. By leveraging the power of DiffusionNet, ScanTalk not only adapts to diverse facial structures but also maintains fidelity when dealing with scanned data, thereby enhancing the authenticity and versatility of generated 3D talking heads. Through comprehensive comparisons with state-of-the-art methods, we validate the efficacy of our approach, demonstrating its capacity to generate realistic talking heads comparable to existing techniques. While our primary objective is to develop a generic method free from topological constraints, all state-of-the-art methodologies are bound by such limitations. Code for reproducing our results, and the pre-trained model are available at https://github.com/miccunifi/ScanTalk.", + "bbox": [ + 261, + 465, + 743, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: 3D Talking Heads $\\cdot$ 3D Scans Animation $\\cdot$ DiffusionNet", + "bbox": [ + 261, + 758, + 720, + 773 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution", + "bbox": [ + 220, + 824, + 362, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/eb9a4099eaec2b0d67bb6cc8deadfa0b594b21bf6626422e4bb946d43f8fe39f.jpg", + "image_caption": [ + "Fig. 1: We present ScanTalk, a deep learning architecture to animate any 3D face mesh driven by a speech. ScanTalk is robust enough to learn on multiple unrelated datasets with a unique model, whilst allowing us to infer on unregistered face meshes." + ], + "image_footnote": [], + "bbox": [ + 217, + 147, + 782, + 286 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 368, + 375, + 383 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Human face animation is a complex task, widely explored in Computer Vision and Computer Graphics because of the broad range of applications drawn from it, spanning from virtual reality to video game graphics, and more. As 3D face models continue to improve, it becomes more and more relevant to incorporate multi-modal aspects such as speech with audio data. Speech-driven facial animation encounters challenges inherent to the fact that finding a cross-modality mapping from audio to geometric data is an ill-posed problem, dealing with the complex geometry of human faces and the limited availability of paired audio-visual 3D data. Another, less discussed aspect that touches a broader range of 3D graphics with meshes is the robustness to changes in the topology of the mesh, which refers to the arrangement of the vertices and how they are connected.", + "bbox": [ + 212, + 401, + 787, + 566 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Maintaining fidelity and coherence across different topologies is crucial for ensuring that facial animations remain realistic and expressive, regardless of variations in the underlying mesh structure. This challenge becomes particularly pronounced in speech-driven facial animation, where the dynamics of speech-related lip movements, and related face changes necessitate a high degree of adaptability within the mesh topology.", + "bbox": [ + 212, + 568, + 787, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Addressing these challenges requires innovative approaches able to navigate the complexities of facial geometry, while accommodating the nuances of speech-driven animation. In this regard, we present ScanTalk, a novel framework capable of animating faces in arbitrary topologies including scanned data. ScanTalk overcomes limitations associated with fixed topologies, offering promising avenues for more flexible and realistic 3D talking heads generation. Indeed, the aforementioned constraint limits the range of applications of current deep learning models for speech-driven motion synthesis. Because of this, deep models are required to train on large scale registered datasets such as FLAME [23] or Multiface [44] for human face data. Building these datasets is a costly procedure, and the trained model is then only usable within the same registered setting. This typically means that any newly acquired data needs to be fitted to the corre", + "bbox": [ + 212, + 659, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "F. Nocentini et al.", + "bbox": [ + 271, + 114, + 393, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "sponding topology before any animation can be predicted by a deep model. This extra preprocessing step usually prevents online applications, therefore canceling one of the key benefits of deep models, which is their inference speed.", + "bbox": [ + 212, + 146, + 784, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Aiming to address these limitations, in this paper we present a flexible deep learning framework built to generate speech-driven animations of 3D face meshes. In particular, it gathers several key elements:", + "bbox": [ + 212, + 191, + 784, + 237 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. A new robust approach to generate mesh deformation sequences based on DiffusionNet to compute intrinsic descriptors on 3D data;", + "2. A comprehensive architecture for learning speech-driven animations. Our approach works with meshes with different topologies, while showing competitive performance with respect to state-of-the-art models trained on an individual topology;", + "3. We show the generalizability of our model to unseen mesh topologies with qualitative examples." + ], + "bbox": [ + 220, + 250, + 784, + 371 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 396, + 387, + 412 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In literature, there are several methodologies to acquire 3D face meshes, either by extracting the geometry from images or videos [26,47] or by using a complex set of instruments in a controlled environment [16, 18, 33, 35, 44, 48]. The former offers advantages in terms of easier and more cost-effective data acquisition. However, these methods may sometimes fall short in capturing the complete 3D information from 2D data. With 3D scans, other challenges arise: they come unregistered and may present alterations such as holes and noise. Then, animating these objects directly becomes even more challenging. One can do it by registering the meshes onto a pre-defined topology [16, 22, 23, 28, 44] before animating the registrations with 3D morphable models [14] for example.", + "bbox": [ + 212, + 431, + 784, + 582 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Several deep learning models [3-5, 10, 25] proposed ways to learn a latent representation of face scans using robust encoders such as PointNet [6, 31] and Transformers but the resulting mesh is registered, which tends to smooth out some details from the scan geometry. However, this extra registration step may be handled efficiently with recent industrial applications such as MetaHuman from Epic Games. More recently, and closer to our goal, DiffusionNet [37] was combined with neural Jacobian fields [1] in Neural Face Rigging (NFR) [32] to learn a per-triangle deformation field on faces, allowing to transfer an animation from a sequence of unregistered face meshes to another. We stand out from this work by proposing a model that uses audio data to animate a given unregistered face mesh. While DiffusionNet has demonstrated its capability to generalize across varying triangulations in a static setting, our model is the first to exhibit similar properties in a multi-modal 4D setting.", + "bbox": [ + 212, + 583, + 785, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In recent years, numerous models and methodologies have emerged to address the challenge of synchronizing facial animations with speech audio. While significant progress has been made on 2D talking heads [2, 7, 12, 21, 42, 43, 49], only a handful of approaches can be seamlessly extended to 3D data. Procedural", + "bbox": [ + 212, + 780, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "ScanTalk: 3D Talking Heads from Unregistered Scans", + "bbox": [ + 372, + 114, + 732, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "techniques have been proposed to animate 3D faces [9, 13, 27, 41, 46], primarily relying on visemes (groups of phonemes) to drive the movements of facial muscles. However, many of these models necessitate re-targeting and struggle to generalize effectively without extensive processing steps. More recent works leverage the increasing availability of data to develop statistical methods, including deep learning strategies [11, 15, 19, 29, 34, 38-40, 45]. Currently, state-of-the-art deep learning approaches are limited to a fixed mesh topology, requiring it to be identical to the one observed during the training phase. Among these models, VOCA [11] pioneers the development of deep models trained on a large-scale dataset of registered meshes. It was outperformed when MeshTalk [34] introduced a larger registered dataset named Multiface [44] along with a more expressive model. Moving forward, FaceFormer [15] utilized the transformer architecture, optimized for temporal data, and leveraged the power of a large-scale pre-trained audio encoder called Wav2vec2 [36]. This strategy saw further development with CodeTalker [45] and SelfTalk [30]. More recently, FaceXHubert [19] and FaceDiffuser [38] utilized an improved audio encoder, Hubert [20], demonstrating superiority to predict cross-modality mappings from audio to face motions. While continuously enhancing performances, these models are bound to a fixed mesh topology, hindering real-world applications and limiting the quantity of usable data for a single model, both for training and inference.", + "bbox": [ + 212, + 146, + 787, + 446 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In response to these challenges, we present a novel framework in the landscape of 3D facial animation learning, avoiding the limitations posed by the specific topology of the face mesh to be animated. The proposed model stands out for its capability to animate any face mesh, including real-world 3D scans. This novel approach holds the potential to redefine the standards in the field of 3D facial animation, being applicable across diverse datasets and scenarios.", + "bbox": [ + 212, + 449, + 787, + 540 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Proposed approach: ScanTalk", + "text_level": 1, + "bbox": [ + 215, + 564, + 539, + 582 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we introduce Scantalk, a framework for animating 3D face meshes reproducing a spoken sentence contained in an audio file, which does not require the meshes to adhere to any specific topology. With ScanTalk, we push the state-of-the-art in 3D talking heads a step forward by allowing any 3D face, even raw scans, to be animated given a speech. To train ScanTalk, the only requirement is that the training meshes share a common topology within each sequence, but the topology may vary from one sequence to another. At inference time, any 3D face mesh in neutral state can be animated.", + "bbox": [ + 212, + 598, + 787, + 717 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ScanTalk is an Encoder-Decoder framework, which receives a neutral face mesh and an audio snippet as input, and outputs a sequence of per-vertex deformation fields, whose length depends on that of the audio. By summing each deformation field to the neutral input face, we ultimately obtain the animated sequence. The encoder is composed of two main modules: an audio encoder, which combines a pretrained encoder with a bi-directional LSTM that extracts audio features from the input speech, and a DiffusionNet encoder that computes surface descriptors from the neutral 3D face. These descriptors are then repli", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "F. Nocentini et al.", + "bbox": [ + 271, + 114, + 393, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ated and concatenated to the audio features and the resulting signal is fed to a DiffusionNet decoder that outputs the deformation to be applied to the neutral face. The framework is depicted in Fig. 2. Before providing the technical details in Sec. 3.1 and Sec. 3.2, below we introduce a few general notations.", + "bbox": [ + 212, + 146, + 787, + 207 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Let $L = \\{(M_i^{gt}, m_i^n, A_i)\\}_{i=0}^{N-1}$ denotes the training set comprising $N$ samples, where $A_i$ is an audio containing a spoken sentence, $M_i^{gt} = (m_i^0, \\ldots, m_i^{T_i-1}) \\in \\mathbb{R}^{T_i \\times V_i \\times 3}$ represents a sequence of 3D faces (same topology) of length $T_i$ synchronized with the spoken sentence in $A_i$ , and $m_i^n \\in \\mathbb{R}^{V_i \\times 3}$ is a 3D neutral face. $V_i = |m_i^n|$ is the number of vertices in the $i$ -th 3D face sequence that needs to be consistent across each mesh in the $i$ -th sequence together with the vertex connectivity, overall determining the topology of the surface. Our objective is to establish a mapping function that correlates an audio input $A_i$ , and a neutral 3D face $m_i^n$ , to the ground-truth sequence $M_i^{gt}$ , expressed as:", + "bbox": [ + 212, + 212, + 787, + 356 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS c a n t a l k \\left(A _ {i}, m _ {i} ^ {n}\\right) \\approx M _ {i} ^ {g t}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 378, + 785, + 397 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/505f7a57701bde4b66873761340bbc034e7dda107df59dd43688f187efcf7cbb.jpg", + "image_caption": [ + "Fig. 2: Architecture of ScanTalk. A novel Encoder-Decoder framework designed to dynamically animate any 3D face based on a spoken sentence from an audio file. The Encoder integrates the 3D neutral face $m_i^n$ , per-vertex surface features $P_i^n$ (crucial for DiffusionNet and precomputed by the operators $OP$ ), and the audio file $A_i$ , yielding a fusion of per-vertex and audio features. These combined descriptors, alongside $P_i^n$ , are then passed to the Decoder, which mirrors a reversed DiffusionNet encoder structure. The Decoder predicts the deformation of the 3D neutral face, which is then combined with the original 3D neutral face $m_i^n$ to generate the animated sequence." + ], + "image_footnote": [], + "bbox": [ + 215, + 470, + 787, + 674 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "ScanTalk: 3D Talking Heads from Unregistered Scans", + "bbox": [ + 372, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 ScanTalk Encoder", + "text_level": 1, + "bbox": [ + 215, + 146, + 410, + 159 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ScanTalk requires an audio snippet $A_{i}$ and a 3D face in neutral state $m_{i}^{n}$ as inputs. To process the two inputs, ScanTalk comprises two distinct encoders to process the mesh and the audio, respectively, as detailed below.", + "bbox": [ + 212, + 171, + 787, + 217 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Face mesh encoder. Several approaches are available for encoding face meshes or point clouds, yet traditional graph convolution-based models, like [17, 24], encounter limitations with varying graph structures, such as changes in mesh resolution. To address this, we employ DiffusionNet [37], a discretization-agnostic encoder that demonstrated to be effective for encoding face meshes as seen in [32]. DiffusionNet integrates multi-layer perceptrons (MLPs), learned diffusion, and spatial gradient features, offering a straightforward yet robust architecture for surface learning tasks. It bypasses the need for complex operations like explicit surface convolutions or pooling hierarchies.", + "bbox": [ + 212, + 239, + 787, + 375 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Critical to DiffusionNet's functionality are precomputed features of the face surface, namely the Cotangent Laplacian, Eigenbasis, Mass Matrix, and Spatial Gradient Matrix. Together, these operators enhance the architecture's robustness, flexibility, and expressiveness, making DiffusionNet suitable for diverse surface learning tasks. Notably, this architecture accommodates 3D faces of any topology, allowing for variations in the number and order of defining points.", + "bbox": [ + 212, + 376, + 787, + 465 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Let $P_{i}^{n} = OP(m_{i}^{n})$ represent the precomputed features obtained by applying the above surface closed-form operators $OP$ to the 3D neutral face $m_{i}^{n}$ . The DiffusionNet Encoder $DN_{e}$ , with latent size $h$ , is designed to process a neutral 3D face mesh $m_{i}^{n}$ , and requires the precomputed per-vertex features $P_{i}^{n}$ to extract a per-vertex descriptors $f_{i}^{n}$ , that is:", + "bbox": [ + 212, + 465, + 787, + 542 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nf _ {i} ^ {n} = D N _ {e} \\left(m _ {i} ^ {n}, P _ {i} ^ {n}\\right) \\in \\mathbb {R} ^ {V _ {i} \\times h}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 398, + 553, + 784, + 571 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The per-vertex descriptors $f_{i}^{n}$ , is capable of capturing intricate details of each vertex within the neutral 3D face.", + "bbox": [ + 212, + 582, + 785, + 612 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Audio encoder. Following [19, 38], the speech encoder adopts the architecture of the state-of-the-art pretrained speech model, HuBERT [20] that is a self-supervised speech representation learner utilizing an offline clustering step to provide aligned target labels for a BERT-like prediction loss. Using this module, followed by a Linear Layer, we obtain a per-frame audio representation:", + "bbox": [ + 212, + 635, + 787, + 710 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\na _ {i} = \\operatorname {S p e e c h E n c o d e r} \\left(A _ {i}\\right) \\in \\mathbb {R} ^ {T _ {i} \\times (h / 2)}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 366, + 722, + 784, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To ensure coherence between the speech representation, following the methodology outlined in [29], we concatenate the SpeechEncoder with a Multilayer Bidirectional-LSTM for temporal consistency, in a way such that the speech signal is projected into a temporal latent representation:", + "bbox": [ + 212, + 751, + 787, + 811 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nv _ {i} = B i L S T M \\left(a _ {i}\\right) \\in \\mathbb {R} ^ {T _ {i} \\times h}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 401, + 823, + 784, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "F. Nocentini et al.", + "bbox": [ + 271, + 114, + 393, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 ScanTalk Decoder", + "text_level": 1, + "bbox": [ + 215, + 146, + 410, + 159 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We combine per-vertex descriptors $f_{i}^{n}$ extracted from the neutral face, with the latent vector $v_{i}$ extracted from the Bidirectional-LSTM", + "bbox": [ + 212, + 172, + 782, + 202 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n(F _ {i} ^ {j}) _ {k} = \\left(f _ {i} ^ {n}\\right) _ {k} \\oplus v _ {i} ^ {j} \\in \\mathbb {R} ^ {h * 2}, \\quad \\forall k = 0, \\dots , V _ {i} - 1, \\quad \\forall j = 0, \\dots , T _ {i} - 1. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 227, + 214, + 785, + 232 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "With this concatenation, we obtain a combined latent $F_{i}^{j} \\in \\mathbb{R}^{V_{i} \\times h*2}$ that embeds both audio-related and geometry-related latents. To decode this sequence of combined latents for deforming the neutral face $m_{i}^{n}$ , we employ a DiffusionNet Decoder (which is essentially a reversed Encoder), denoted as $DN_{d}$ . The decoder module receives $F_{i}^{j}$ , and precomputed features $P_{i}^{n}$ derived from $m_{i}^{n}$ . It predicts the deformation of $m_{i}^{n}$ , denoted as $DN_{d}(F_{i}^{j}, P_{i}^{n})$ :", + "bbox": [ + 212, + 244, + 784, + 338 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {m _ {i} ^ {j}} = D N _ {d} \\left(F _ {i} ^ {j}, P _ {i} ^ {n}\\right) + m _ {i} ^ {n} \\in \\mathbb {R} ^ {V _ {i} \\times 3}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 351, + 784, + 368 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Here $\\widehat{m_i^j}$ represents the $j$ -th frame of the predicted sequence. The entire generated sequence is defined by $\\widehat{M_i} \\in \\mathbb{R}^{T_i \\times V_i \\times 3}$ . We opt to utilize ScanTalk for predicting the deformation of the neutral face $m_i^n$ rather than predicting the actual face. This decision aligns with previous works [15, 19, 30, 38, 40], and offers advantages in terms of training efficiency and resulting animation. Predicting face deformation makes the learning process easier, by focusing solely on speech-related motion, as opposed to incorporating the problem of predicting the entire face reconstruction. Essentially, the decoder learns to predict a per-vertex displacement field from a time-dependent per-vertex descriptors field.", + "bbox": [ + 212, + 381, + 782, + 518 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 ScanTalk Training", + "text_level": 1, + "bbox": [ + 215, + 541, + 413, + 556 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ScanTalk generates deformations of a neutral face; hence, a predicted sequence maintains a consistent topology across all frames. This property arises from the definition of the DiffusionNet Decoder, which necessitates knowledge of precomputed features $P_{i}^{n}$ , and the number of points in the 3D neutral face targeted for animation. For this reason, during a supervised training protocol, the ground-truth meshes within each sequence must adhere to a common topology. Despite the apparent specificity of this requirement, it proves to be non-prohibitive in practice. ScanTalk animates a neutral face in response to an audio input, producing a sequence of 3D faces sharing the same topology of the neutral face that is animated, which can be any topology (even different from those seen during training). This alignment between the training strategy and the desired inference outcome underscores the efficiency of ScanTalk, which is not restricted to the topology of the training faces.", + "bbox": [ + 212, + 566, + 784, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Notably, the model predicts per-vertex displacements of the neutral ground-truth mesh and the ground truth sequence is registered in a supervised setting. Consequently, the predicted sequence and the ground-truth sequence are aligned on the same topology but can vary from one sequence to another. Thus, during training, we minimize the average vertex-wise Mean Squared Error (MSE) over", + "bbox": [ + 212, + 765, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "ScanTalk: 3D Talking Heads from Unregistered Scans", + "bbox": [ + 372, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "a sequence of length $T_{i}$ between the ground truth $M_{i}^{gt}$ and the model prediction $\\widehat{M}_{i}$ . which is defined as:", + "bbox": [ + 215, + 145, + 784, + 178 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {M S E} = \\frac {1}{T _ {i} - 1} \\sum_ {j = 0} ^ {T _ {i} - 1} \\frac {1}{V _ {i} - 1} \\sum_ {k = 0} ^ {V _ {i} - 1} \\left\\| \\left(m _ {i} ^ {j}\\right) _ {k} - \\left(\\widehat {m _ {i}} ^ {j}\\right) _ {k} \\right\\| _ {2} ^ {2}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 316, + 191, + 784, + 232 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 253, + 375, + 272 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the following, we first introduce the datasets and the metrics used for evaluation, respectively in Sec. 4.1 and Sec. 4.2. Then, we report quantitative and qualitative results in comparison with state-of-the-art methods in Sec. 4.3 and Sec. 4.4. In Sec. 4.5 we present several ablation studies over the framework architecture. Finally, in Sec. 4.6, we report the results of an user study designed to compare our approach with state-of-the-art methods.", + "bbox": [ + 214, + 287, + 782, + 377 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Datasets", + "text_level": 1, + "bbox": [ + 215, + 401, + 333, + 414 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For training and quantitative evaluations, we rely on three classical datasets in 3D speech-driven motion synthesis: VOCAset [11], BIWI [16] and Multiface [44]. VOCAset gathers mesh sequences of 12 actors performing 40 speeches, captured at 60fps. Each sequence lasts for around 3 to 5 seconds, and each mesh is registered to the FLAME [23] topology with 5,023 vertices and 9,976 faces.", + "bbox": [ + 214, + 428, + 782, + 502 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "BIWI comprises 14 subjects articulating 40 sentences each, sampled at 25fps. Each sentence lasts around 5 seconds, and the meshes are registered to a fixed topology. Due to GPU limitations, we used a downsampled version of this dataset, called $\\mathrm{BIWI}_6$ , which has a fixed topology, with 3,895 vertices and 7,539 faces.", + "bbox": [ + 214, + 503, + 784, + 563 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Multiface includes 13 identities executing up to 50 speeches of around 4 seconds each, sampled at 30fps. The original multiface meshes have a fixed topology with 5,471 vertices and 10,837 faces.", + "bbox": [ + 214, + 563, + 782, + 607 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For training purposes, the meshes in both Multiface and BIWI $_6$ have been scaled and aligned with the meshes in the VOCAset dataset. The data splits can be found in the supplementary material. Since ScanTalk is the first topology-independent 3D talking head generator, for the sake of a comprehensive comparison with the state-of-the-art, we trained 4 different models: one is trained using all the three datasets together (multi-dataset), while the other three are trained on each dataset separately (single-dataset).", + "bbox": [ + 214, + 609, + 782, + 714 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 Evaluation Metrics", + "text_level": 1, + "bbox": [ + 215, + 738, + 419, + 752 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To evaluate the quality of the generated faces, we employ three standard metrics from previous works [30,34,38,45], namely: Lip vertex error (LVE) $\\times 10^{-5}(mm)$ Assesses the lip deviation in a sequence by comparing it to the ground truth. It is obtained by computing the maximum L2 error for all lip vertices in each frame, averaged across all frames.", + "bbox": [ + 214, + 763, + 787, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "F. Nocentini et al.", + "bbox": [ + 271, + 114, + 393, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Mean vertex error (MVE) $\\times 10^{-3}(mm)$ . Describes the quality of the overall face reconstruction by computing the maximal vertex L2 error for each frame and taking the mean across all frames.", + "bbox": [ + 212, + 146, + 782, + 190 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Upper-face dynamic deviation (FDD) $\\times 10^{-7}(mm)$ . Compares the standard deviation of upper vertices between the generated sequence and ground-truth.", + "bbox": [ + 212, + 191, + 782, + 220 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3 Quantitative Results", + "text_level": 1, + "bbox": [ + 215, + 273, + 431, + 287 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Although ScanTalk is the first deep learning model to animate unregistered 3D face meshes, comparisons with the state-of-the-art is possible when inferring our model on registered data. In Tab. 1, we present the performance evaluation of ScanTalk training on both a single-dataset and multi-datasets, juxtaposed against several state-of-the-art methods trained solely on a single dataset. While the augmentation of training data enhances the effectiveness of our model, addressing disparities in data composition becomes important for animating unseen faces. Particularly, the varied geometries and head dynamics inherent in $\\mathrm{BIWI}_6$ , VOCA, and Multiface meshes pose significant challenges, potentially limiting performance compared to a model trained exclusively on a single dataset.", + "bbox": [ + 212, + 311, + 784, + 462 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 1: ScanTalk performance, in both single-dataset (s-d) and multi-dataset (m-d) scenarios, in comparison with state-of-the-art methods. The heatmaps show the differences between the first frame and subsequent frames within sequences among the VOCAset, $\\mathrm{BIWI}_6$ , and Multiface datasets. Notably, in the VOCAset, primarily the lips display movement, whereas in $\\mathrm{BIWI}_6$ and Multiface datasets, substantial head and upper face movements are observed. The color gradient on the face meshes corresponds to the average per-vertex $L_2$ norm of the differences, where blue hues indicate lower values, and red hues indicate higher values.", + "bbox": [ + 212, + 506, + 784, + 617 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/1fe722d18ef68bdc0c0bc93ebd20e94ee33e2de08f9b34fd2762e984a10ffe48.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 369, + 632, + 419, + 694 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/8f329ff74c2ca8096ecc35d4960f69d09f901e004d9e634cc4b80402fccd7d35.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 526, + 637, + 568, + 681 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/571666e5c93356be6eb0d325bb1a06ac4fb02a042cdb3233d2c384868afa24f2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 674, + 628, + 730, + 700 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/94954a776536fe2d1b7cb37dfec6e8ee7bc6c9bec3ec2e7777f820870580a281.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
VOCAssetBIWI6Multiface
LVE ↓MVE ↓FDD ↓LVE ↓MVE ↓FDD ↓LVE ↓MVE ↓FDD ↓
VOCA6.9930.9832.6625.7432.59141.4824.9232.76155.781
FaceFormer6.1230.9352.1634.0852.16337.0912.4511.45320.239
SelfTalk5.6180.9182.3213.6282.06235.4702.2811.90137.434
CodeTalker3.5490.8882.2585.1902.64120.5994.0912.38247.905
FaceDiffuser4.3500.9012.4374.0222.12839.6043.5552.38829.157
ScanTalk m-d6.3750.9872.1014.0442.05740.0512.4351.67832.202
ScanTalk s-d3.0120.8612.4004.6512.14836.0342.6531.87164.451
", + "bbox": [ + 233, + 699, + 761, + 835 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "ScanTalk: 3D Talking Heads from Unregistered Scans", + "bbox": [ + 372, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This observation becomes evident when assessing Scantalk's performance on VOCAset, where a substantial improvement is observed when training on a single-dataset. Conversely, in the case of $\\mathrm{BIWI}_6$ and Multiface single-dataset training, a decline in results is noted compared to the multi-dataset version. This disparity is attributed to the presence of speech-related head and upper face movements within the ground truth talking head sequences of $\\mathrm{BIWI}_6$ and Multiface. Consequently, the multi-dataset training model demonstrates superior performance as it encounters a wider array of sequences featuring head and upper face movements. In contrast, VOCAset lacks such movements, with the 3D faces remaining static, while only the mouth moves, thereby enhancing performance for ScantTalk trained solely on VOCAset. This obviates the need to learn other movements of the head or upper face.", + "bbox": [ + 212, + 146, + 787, + 327 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The distinctions between VOCAset and the other two datasets are further underscored by the FDD metric, indicating that 3D faces in VOCAset remain stationary, with minimal activity in the upper facial area and head. Across all tested models, generating sequences with lower FDD is notably more achievable on VOCAset compared to the other datasets, thus emphasizing the disparities between VOCAset and the others.", + "bbox": [ + 212, + 329, + 787, + 417 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Nevertheless, the results presented in Tab. 1 demonstrate that ScanTalk produces outcomes comparable to state-of-the-art methods, whether undergoing single-dataset or multi-dataset training. Moreover, ScanTalk emerges as the first model capable of training in multi-dataset settings, demonstrating the ability to animate a 3D face regardless of its topology. An interesting characteristic of ScanTalk emerges when analyzing the GPU memory usage relative to the vertex count in the 3D facial model requiring animation. Fig. 3 illustrates a linear increase in GPU usage correlated with the number of vertices.", + "bbox": [ + 212, + 420, + 573, + 614 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/bd80a0f7c0b1ac3c572bd671c1a0bcf29a8349f9de5fdb3352912e1e4b6ea7f2.jpg", + "image_caption": [ + "Fig.3: ScanTalk GPU memory usage with respect to the mesh resolution." + ], + "image_footnote": [], + "bbox": [ + 591, + 449, + 777, + 549 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.4 Qualitative Results", + "text_level": 1, + "bbox": [ + 215, + 643, + 419, + 657 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In the domain of 3D speech-driven talking heads generation, qualitative evaluations hold as much significance as quantitative assessments. To evaluate the quantitative findings presented in Sec. 4.3, we introduce a straightforward test aimed at discerning disparities in head and upper face movements across the three datasets: for each dataset, we compute the average $L_{2}$ norm of the differences between the initial frame and subsequent frames within each sequence. This enables the quantification of both head dynamics and upper facial movements across mesh sequences. In the heatmaps of Tab. 1, we present the per-frame average difference for each dataset. Our analysis reveals that in the VOCAset, only the lower part of the face exhibits discernible movements, whereas in BIWI $_6$ and Multiface, the entire head or face moves.", + "bbox": [ + 212, + 672, + 787, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "F. Nocentini et al.", + "bbox": [ + 271, + 114, + 393, + 127 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/297b328e04cfc61facd6bd7b14328f3fcd4731285f89866ed76d9130a7879461.jpg", + "image_caption": [ + "Fig. 4: Relative norm of the per-vertex descriptors $f_{i}^{n}$ in Eq. (2) extracted by $DN_{e}$ displayed as a heatmap on a mesh from VOCAset (left), and a mesh from Multiface (right). For each mesh, we show the norm on the original topology, on a remeshed version, and on a further degraded mesh obtained by removing the back of the head and creating random holes. Here, pinker hues indicate lower values, and greener hues indicate higher values." + ], + "image_footnote": [], + "bbox": [ + 248, + 150, + 751, + 351 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/1342fd14a0bf9b6e8e99754d04b9a0a63353d62dd6c235986ad1646955114338.jpg", + "image_caption": [ + "Fig. 5: ScanTalk inference on 3D faces. The meshes have been rigidly aligned with the training data. The first row is an animation of a raw 3D scan (an hole for the mouth has been created), while the others two are animation of meshes in an arbitrary topology." + ], + "image_footnote": [], + "bbox": [ + 217, + 469, + 787, + 742 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "ScanTalk: 3D Talking Heads from Unregistered Scans", + "bbox": [ + 372, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In Fig. 4, we present a heatmap visualization representing the norm of geometric descriptors $f_{i}^{n}$ computed by the encoder $DN_{e}$ in Eq. (2) for various alterations of the same identity mesh. Our findings indicate that the ScanTalk encoder effectively extracts descriptors from input faces regardless of facial topology. Notably, despite differences in facial topology among the same identity, the extracted descriptors exhibit remarkable similarity.", + "bbox": [ + 212, + 146, + 782, + 234 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In Fig. 5, 3D faces in different topology, including a scan, animated using ScanTalk are presented. Our method demonstrates good generalization capabilities, successfully animating diverse 3D faces. Notably, enhanced animation quality is observed when a mouth aperture is present; however, our model performs well even in its absence. Nonetheless, without a mouth aperture, the model struggles to generate the corresponding mouth opening, despite accurate lip movement synchronization.", + "bbox": [ + 212, + 237, + 784, + 342 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.5 Ablations Studies", + "text_level": 1, + "bbox": [ + 215, + 366, + 408, + 380 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To investigate the impacts of various components within our architecture, we conducted extensive experiments across various configurations, encompassing both single-dataset and multi-dataset training paradigms. The results reported in Tab. 2 and Tab. 3 are computed by modifying ScanTalk's modules and training losses as described in Sec. 3.", + "bbox": [ + 212, + 392, + 782, + 467 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Audio encoder. In learning based speech-driven animation, some previous work used either Wav2vec2 [36] or Hubert [20], the latter being substantially better for this task according to [19,38]. For our purpose, we also tested WavLM [8] but obtained poorer efficiency. The results on both the single-dataset and multi-dataset training are displayed in Tab. 2a and Tab. 3 top. We can see that, in both cases, the usage of the Hubert Encoder [19] for audio features extraction leads to better results.", + "bbox": [ + 212, + 491, + 782, + 595 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Temporal consistency. In our investigation, as proven by [19,29,40], we found that supplementing pre-trained audio encoders with additional temporal consistency mechanisms such as Bidirectional-LSTM, Bidirectional-GRU, or an autoregressive Transformer Decoder (TD) significantly enhances model performance, as illustrated in Tab. 2b and in Tab. 3 middle. The BiLSTM model architecture is detailed in Sec. 3, while the BiGRU model mirrors the BiLSTM architecture but substitutes BiLSTM with BiGRU. Details about the Transformer Decoder can be found in the supplementary material. From the findings presented in Tab. 2b and in Tab. 3 middle, we observe that employing a multilayer Bidirectional-LSTM in the audio stream processing of ScanTalk yields the most favorable performance across both single-dataset training and multi-dataset training scenarios.", + "bbox": [ + 212, + 619, + 784, + 785 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Loss function. Previous research has extensively explored optimal objective functions for enhancing and refining the learning process. Inspired by [29,30,40],", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "F. Nocentini et al.", + "bbox": [ + 271, + 114, + 393, + 127 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "we experimented with Mean Square Error ( $L_{MSE}$ ), Masked ( $L_{mask}$ ), and Velocity ( $L_{vel}$ ) Loss. The former is a simple $L_2$ loss, the second employs Mean Square Error focused on lip vertices, while the latter aims to minimize differences between consecutive frames. While these loss functions have demonstrated efficacy in single-dataset training, as evidenced by [29,30,40], our findings, detailed in Tab. 3 bottom, indicate that employing a straightforward $L_2$ , on multi-dataset training, enhances the model's capacity to generate realistic talking heads. We attribute this to significant geometric variations observed across different datasets.", + "bbox": [ + 212, + 146, + 787, + 268 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/944ccbd6a36f4349563524f38b2408ca281973093ac4be9773a960887227200b.jpg", + "table_caption": [ + "Table 2: ScanTalk (ST) single-dataset ablation studies. Results obtained with a model trained just on VOCAsset.", + "(a) Audio Encoder Ablation." + ], + "table_footnote": [], + "table_body": "
LVE ↓MVE ↓FDD ↓
ST w WavLM3.6740.9372.413
ST w Wav2Vec23.3090.8602.244
ScanTalk3.0120.8612.400
", + "bbox": [ + 243, + 367, + 478, + 429 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/2d139170851efc67ce007d0a75659d1692a90cfea38e95adf4dab993ba835c44.jpg", + "table_caption": [ + "(b) Temporal consistency ablation." + ], + "table_footnote": [], + "table_body": "
LVE ↓MVE ↓FDD ↓
ST w/o3.3610.8702.365
ST w TD3.2910.8592.406
ST w BiGRU3.0360.8352.358
ScanTalk3.0120.8612.400
", + "bbox": [ + 532, + 367, + 751, + 439 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/4063bb6e5e28d1ddd2ae53723b40b65fc61013c337f27f0b0af3e83e1d4f6c86.jpg", + "table_caption": [ + "Table 3: ScanTalk (ST) multi-dataset ablation studies. Top: ST with different audio encoders. Middle: ST using different audio stream processing. Bottom: ST with different Loss function. The last row is the proposed ScanTalk" + ], + "table_footnote": [], + "table_body": "
VOCAsetBIWI6Multiface
LVE ↓MVE ↓FDD ↓LVE ↓MVE ↓FDD ↓LVE ↓MVE ↓FDD ↓
ST w Wav2Vec27.1271.2412.3494.4392.27641.3743.1982.32933.671
ST w WavLM6.7731.0042.2204.3332.16231.9222.9072.0048.500
ST w TD6.5120.9951.8584.8272.16836.7892.1361.72535.044
ST w BiGRU6.5840.9942.0544.4212.10341.4562.5341.98734.098
ST+Lmask+Lvel7.0311.0921.4734.7552.22739.4582.5471.89815.973
ST+Lmask7.4511.0840.8594.7482.19936.9392.8881.89132.501
ST+Lvel6.7400.9981.8994.5092.18033.6142.1031.77531.599
ScanTalk6.3750.9872.1014.0442.05740.0512.4351.67832.202
", + "bbox": [ + 235, + 549, + 767, + 704 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.6 User study", + "text_level": 1, + "bbox": [ + 215, + 753, + 354, + 768 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "To further evaluate our solution, we performed a study where human feedback is involved. This evaluation is conducted with two user-based studies involving 25 participants; $(i)$ For the first study, in alignment with prior research [15,19,30,34, 38,45], we designed an A/B test to compare ScanTalk with other state-of-the-art", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "ScanTalk: 3D Talking Heads from Unregistered Scans", + "bbox": [ + 372, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "models within a registered setting on both lip-syncing and naturalness criteria (Test 1); (ii) In the second test, we assessed the credibility of scan animations by asking participants to rate the animation quality of ten scans sourced from the COMA dataset [33], using a scale ranging from 1 to 10 (Test 2).", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The outcomes of Test 1 are depicted in the table on the left of Fig. 6. Notably, sequences produced using ScanTalk demonstrate levels of both naturalness and lip-syncing that are comparable to state-of-the-art methods. Specifically, our approach is preferred when compared with FaceFormer, CodeTalker, and FaceDiffuser on VOCAsset. However, SelfTalk and the ground truth are preferred over ScanTalk. Nevertheless, the percentage of users favoring ScanTalk remains nonnegligible, underscoring the efficacy of our approach in generating 3D talking heads with good realism and lip-syncing fidelity.", + "bbox": [ + 212, + 207, + 784, + 325 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Results of Test 2 are reported on the right of Fig. 6. These ratings are closely linked to both the scan quality and the fidelity of the mouth representation. Despite the scan quality not being optimal, as depicted in Fig. 6, our results indicate that we can achieve a good level of realism in the animated scans.", + "bbox": [ + 212, + 327, + 782, + 388 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/b5cad12d0dda874011b015b93bab35f4a60205631c74da298da269f31590d5e4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ComparisonNaturalness (%)Lip-sync (%)
ST vs. FaceFormer82.6780.67
ST vs. CodeTalker57.3356.00
ST vs. FaceDiffuser90.6794.00
ST vs. SelfTalk44.0043.33
ST vs. GT44.0042.00
", + "bbox": [ + 217, + 444, + 454, + 508 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/3d30e1e766fee7c38debca7d1ffd148c36d3676c9d9770c00fbb950612675710.jpg", + "image_caption": [ + "Fig. 6: The table on the left shows results of Test 1, ScanTalk (ST) vs. other methods and the ground truth (GT) on samples from the test set of VOCAset. The values denote the percentages of users who favored ScanTalk animations over others. The chart on the right shows results of Test 2 reported as a violin plot for the animation of scans from COMA [33]. The subject median rating is displayed as a black dot on each violin." + ], + "image_footnote": [], + "bbox": [ + 462, + 412, + 782, + 566 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusions", + "text_level": 1, + "bbox": [ + 215, + 699, + 366, + 714 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This paper introduces ScanTalk, a novel framework for speech-driven 3D facial animation. Unlike existing methods, ScanTalk possesses the unique capability to animate any 3D face, regardless of its topology, even if it differs from the ones on which it was trained. ScanTalk extends the applicability of deep speech-driven 3D facial animations by addressing the challenges of topology robustness. Additionally, our model demonstrates comparable quantitative and qualitative results with other state-of-the-art methods across three distinct datasets.", + "bbox": [ + 212, + 731, + 784, + 835 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "F. Nocentini et al.", + "bbox": [ + 271, + 114, + 393, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments This work is supported by the ANR project Human4D ANR-19-CE23-0020 and by the IRP CNRS project GeoGen3DHuman. This work was also partially supported by \"Partenariato FAIR (Future Artificial Intelligence Research) - PE00000013, CUP J33C22002830006\" funded by NextGenerationEU through the italian MUR within NRRP, project DL-MIG. This work was also partially funded by the ministerial decree n.352 of the 9th April 2022 NextGenerationEU through the italian MUR within NRRP. This work was also partially supported by Fédération de Recherche Mathématique des Hauts-de-France (FMHF, FR2037 du CNRS).", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 306, + 323, + 321 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Aigerman, N., Gupta, K., Kim, V.G., Chaudhuri, S., Saito, J., Groueix, T.: Neural jacobian fields: Learning intrinsic mappings of arbitrary meshes. ACM Trans. Graph. 41(4) (jul 2022). https://doi.org/10.1145/3528223.3530141, https://doi.org/10.1145/3528223.3530141", + "2. Alghamdi, M.M., Wang, H., Bulpitt, A.J., Hogg, D.C.: Talking head from speech audio using a pre-trained image generator. In: Proceedings of the 30th ACM International Conference on Multimedia (2022)", + "3. Bahri, M., O' Sullivan, E., Gong, S., Liu, F., Liu, X., Bronstein, M.M., Zafeiriou, S.: Shape my face: Registering 3d face scans by surface-to-surface translation. International Journal of Computer Vision (IJCV) (Sep 2021)", + "4. Besnier, T., Arguillère, S., Pierson, E., Daoudi, M.: Toward Mesh-Invariant 3D Generative Deep Learning with Geometric Measures. Computers and Graphics (2023). https://doi.org/10.1016/j.cag.2023.06.027, https://hal.science/hal-04143649", + "5. Chandran, P., Zoss, G., Gross, M., Gotardo, P., Bradley, D.: Shape transformers: Topology-independent 3d shape models using transformers. In: Computer Graphics Forum. vol. 41, pp. 195-207. Wiley Online Library (2022)", + "6. Charles, R.Q., Su, H., Kaichun, M., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). p. 77-85. IEEE (Jul 2017)", + "7. Chen, L., Cui, G., Liu, C., Li, Z., Kou, Z., Xu, Y., Xu, C.: Talking-head generation with rhythmic head motion. In: European Conference on Computer Vision (2020), https://api-semanticscholar.org/CorpusID:220633152", + "8. Chen, S., Wang, C., Chen, Z., Wu, Y., Liu, S., Chen, Z., Li, J., Kanda, N., Yoshioka, T., Xiao, X., Wu, J., Zhou, L., Ren, S., Qian, Y., Qian, Y., Zeng, M., Wei, F.: Wavlm: Large-scale self-supervised pre-training for full stack speech processing. IEEE Journal of Selected Topics in Signal Processing 16, 1505-1518 (2021), https://api-semanticscholar.org/CorpusID:239885872", + "9. Cosi, P., Caldognetto, E., Perin, G., Zmarich, C.: Labial coarticulation modeling for realistic facial animation. In: Proceedings. Fourth IEEE International Conference on Multimodal Interfaces. pp. 505-510 (2002). https://doi.org/10.1109/ICMI.2002.1167047", + "10. Croquet, B., Christiaens, D., Weinberg, S.M., Bronstein, M., Vandermeulen, D., Claes, P.: Unsupervised diffeomorphic surface registration and non-linear modelling. In: Medical Image Computing and Computer Assisted Intervention (MIC-CAI). p. 118–128. Springer (2021)" + ], + "bbox": [ + 217, + 339, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "ScanTalk: 3D Talking Heads from Unregistered Scans", + "bbox": [ + 372, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "11. Cudeiro, D., Bolkart, T., Laidlaw, C., Ranjan, A., Black, M.: Capture, learning, and synthesis of 3D speaking styles. In: Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR). pp. 10101-10111 (2019)", + "12. Das, D., Biswas, S., Sinha, S., Bhowmick, B.: Speech-driven facial animation using cascaded gans for learning of motion and texture. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J.M. (eds.) Computer Vision - ECCV 2020. pp. 408-424. Springer International Publishing, Cham (2020)", + "13. Edwards, P., Landreth, C., Fiume, E., Singh, K.: Jali: an animator-centric viseme model for expressive lip synchronization. ACM Trans. Graph. 35(4) (jul 2016). https://doi.org/10.1145/2897824.2925984, https://doi.org/10.1145/2897824.2925984", + "14. Egger, B., Smith, W.A.P., Tewari, A., Wuhrer, S., Zollhoefer, M., Beeler, T., Bernard, F., Bolkart, T., Kortylewski, A., Romdhani, S., Theobalt, C., Blanz, V., Vetter, T.: 3D Morphable Face Models - Past, Present and Future. ACM Transactions on Graphics 39(5), 157:1-38 (Aug 2020). https://doi.org/10.1145/3395208, https://inria.hal.science/hal-02280281", + "15. Fan, Y., Lin, Z., Saito, J., Wang, W., Komura, T.: Faceformer: Speech-driven 3d facial animation with transformers. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). p. 18749-18758. IEEE, New Orleans, LA, USA (Jun 2022). https://doi.org/10.1109/CVPR52688.2022.01821, https://ieeexplore.ieee.org/document/9878591/", + "16. Fanelli, G., Gall, J., Romsdorfer, H., Weise, T., Van Gool, L.: A 3-d audio-visual corpus of affective communication. IEEE Transactions on Multimedia 12(6), 591-598 (2010). https://doi.org/10.1109/TMM.2010.2052239", + "17. Gong, S., Chen, L., Bronstein, M., Zafeiriou, S.: Spiralnet++: A fast and highly efficient mesh convolution operator. In: Proceedings of the IEEE International Conference on Computer Vision Workshops. pp. 0-0 (2019)", + "18. Gupta, S., Castleman, K.R., Markey, M.K., Bovik, A.C.: Texas 3d face recognition database. In: 2010 IEEE Southwest Symposium on Image Analysis & Interpretation (SSIAI). pp. 97-100 (2010). https://doi.org/10.1109/SSIAI.2010.5483908", + "19. Haque, K.I., Yumak, Z.: Facexhubert: Text-less speech-driven e(x)pressive 3d facial animation synthesis using self-supervised speech representation learning. In: INTERNATIONAL CONFERENCE ON MULTIMODAL INTERACTION (ICMI '23). ACM, New York, NY, USA (2023). https://doi.org/10.1145/3577190.3614157, https://doi.org/10.1145/3577190.3614157", + "20. Hsu, W.N., Bolte, B., Tsai, Y.H.H., Lakhotia, K., Salakhutdinov, R., Mohamed, A.: Hubert: Self-supervised speech representation learning by masked prediction of hidden units. IEEE/ACM Trans. Audio, Speech and Lang. Proc. 29, 3451-3460 (oct 2021). https://doi.org/10.1109/TASLP.2021.3122291, https://doi.org/10.1109/TASLP.2021.3122291", + "21. Ji, X., Zhou, H., Wang, K., Wu, Q., Wu, W., Xu, F., Cao, X.: Eamm: One-shot emotional talking face via audio-based emotion-aware motion model. In: ACM SIGGRAPH 2022 Conference Proceedings. SIGGRAPH '22 (2022). https://doi.org/10.1145/3528233.3530745, https://doi.org/10.1145/3528233.3530745", + "22. Li, J., Kuang, Z., Zhao, Y., He, M., Bladin, K., Li, H.: Dynamic facial asset and rig generation from a single scan. ACM Trans. Graph. 39(6) (nov 2020). https://doi.org/10.1145/3414685.3417817, https://doi.org/10.1145/3414685.3417817", + "23. Li, T., Bolkart, T., Black, M.J., Li, H., Romero, J.: Learning a model of facial shape and expression from 4D scans. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia) 36(6), 194:1-194:17 (2017), https://doi.org/10.1145/3130800.3130813" + ], + "bbox": [ + 215, + 147, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "F. Nocentini et al.", + "bbox": [ + 271, + 114, + 393, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "24. Lim, I., Dielen, A., Campen, M., Kobbelt, L.: A simple approach to intrinsic correspondence learning on unstructured 3d meshes. p. 349-362. Springer-Verlag, Berlin, Heidelberg (2019). https://doi.org/10.1007/978-3-030-11015-4_26, https://doi.org/10.1007/978-3-030-11015-4_26", + "25. Liu, F., Tran, L., Liu, X.: 3d face modeling from diverse raw scan data. In: 2019 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 9407-9417. IEEE Computer Society, Los Alamitos, CA, USA (nov 2019). https://doi.org/10.1109/ICCV.2019.00950", + "26. Lugasri, C., Tang, J., Nash, H., McClanahan, C., Uboweja, E., Hays, M., Zhang, F., Chang, C.L., Yong, M., Lee, J., Chang, W.T., Hua, W., Georg, M., Grundmann, M.: Mediapipe: A framework for perceiving and processing reality. In: Third Workshop on Computer Vision for AR/VR at IEEE Computer Vision and Pattern Recognition (CVPR) 2019 (2019), https://mixedreality.cs.cornell.edu/s/NewTitle_May1_MediaPipe_CVPR_CV4ARVR_Workshop_2019.pdf", + "27. Massaro, D., Cohen, M., Tabain, M., Beskow, J., Clark, R.: Animated speech: Research progress and applications. Audiovisual Speech Processing (01 2001). https://doi.org/10.1017/cbo9780511843891.014", + "28. Muralikrishnan, S., Huang, C.H.P., Ceylan, D., Mitra, N.J.: Bliss: Bootstrapped linear shape space (2023)", + "29. Nocentini, F., Ferrari, C., Berretti, S.: Learning landmarks motion from speech for speaker-agnostic 3d talking heads generation. In: Foresti, G.L., Fusiello, A., Hancock, E. (eds.) Image Analysis and Processing - ICIAP 2023. pp. 340-351. Springer Nature Switzerland, Cham (2023)", + "30. Peng, Z., Luo, Y., Shi, Y., Xu, H., Zhu, X., Liu, H., He, J., Fan, Z.: Selftalk: A self-supervised commutative training diagram to comprehend 3d talking faces. In: Proceedings of the 31st ACM International Conference on Multimedia. p. 5292-5301 (2023). https://doi.org/10.1145/3581783.3611734", + "31. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems (NeurIPS) 30 (2017)", + "32. Qin, D., Saito, J., Aigerman, N., Thibault, G., Komura, T.: Neural face rigging for animating and retargeting facial meshes in the wild. In: SIGGRAPH 2023 Conference Papers (2023)", + "33. Ranjan, A., Bolkart, T., Sanyal, S., Black, M.J.: Generating 3D faces using convolutional mesh autoencoders. In: European Conference on Computer Vision (ECCV). pp. 725-741 (2018), http://coma.is.tue.mpg.de/", + "34. Richard, A., Zollhöfer, M., Wen, Y., de la Torre, F., Sheikh, Y.: Meshtalk: 3d face animation from speech using cross-modality disentanglement. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 1173-1182 (October 2021)", + "35. Savran, A., Alyüz, N., Dibeklioglu, H., Çeliktutan, O., Gökberk, B., Sankur, B., Akarun, L.: Bosphorus database for 3d face analysis. In: Schouten, B., Juul, N.C., Drygajlo, A., Tistarelli, M. (eds.) Biometrics and Identity Management. pp. 47-56. Springer Berlin Heidelberg, Berlin, Heidelberg (2008)", + "36. Schneider, S., Baevski, A., Collobert, R., Auli, M.: wav2vec: Unsupervised pretraining for speech recognition. In: Interspeech 2019. p. 3465-3469. ISCA (Sep 2019). https://doi.org/10.21437/Interspeech.2019-1873, https://www.iscaspeech.org/archive/interspeech_2019/schneider19_interspeech.html", + "37. Sharp, N., Attaiki, S., Crane, K., Ovsjanikov, M.: Diffusionnet: Discretization ag- nostic learning on surfaces. ACM Trans. Graph. 01(1) (2022)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "ScanTalk: 3D Talking Heads from Unregistered Scans", + "bbox": [ + 372, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "38. Stan, S., Haque, K.I., Yumak, Z.: Facediffuser: Speech-driven 3d facial animation synthesis using diffusion. In: ACM SIGGRAPH Conference on Motion, Interaction and Games (MIG '23), November 15-17, 2023, Rennes, France. ACM, New York, NY, USA (2023). https://doi.org/10.1145/3623264.3624447, https://doi.org/10.1145/3623264.3624447", + "39. Thambiraja, B., Aliakbarian, S., Cosker, D., Thies, J.: 3diface: Diffusion-based speech-driven 3d facial animation and editing (2023)", + "40. Thambiraja, B., Habibie, I., Aliakbarian, S., Cosker, D., Theobalt, C., Thies, J.: Imitator: Personalized speech-driven 3d facial animation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 20621-20631 (October 2023)", + "41. Wang, A., Emmi, M., Faloutsos, P.: Assembling an expressive facial animation system. In: Proceedings of the 2007 ACM SIGGRAPH Symposium on Video Games. p. 21-26. Sandbox '07, Association for Computing Machinery, New York, NY, USA (2007). https://doi.org/10.1145/1274940.1274947, https://doi.org/10.1145/1274940.1274947", + "42. Wang, J., Zhao, Y., Liu, L., Xu, T.S., Li, Q., Li, S.: Emotional talking head generation based on memory-sharing and attention-augmented networks. ArXiv abs/2306.03594 (2023), https://api_semanticscholar.org/CorpusID: 259089214", + "43. Wang, S., Li, L., Ding, Y., Yu, X.: One-shot talking face generation from single-speaker audio-visual correlation learning. In: AAAI 2022 (2022)", + "44. Wuu, C.h., Zheng, N., Ardisson, S., Bali, R., Belko, D., Brockmeyer, E., Evans, L., Godisart, T., Ha, H., Huang, X., Hypes, A., Koska, T., Krenn, S., Lombardi, S., Luo, X., McPhail, K., Millerschoen, L., Perdoch, M., Pitts, M., Richard, A., Saragih, J., Saragih, J., Shiratori, T., Simon, T., Stewart, M., Trimble, A., Weng, X., Whitewolf, D., Wu, C., Yu, S.I., Sheikh, Y.: Multiface: A dataset for neural face rendering. In: arXiv (2022). https://doi.org/10.48550/ARXIV.2207.11243, https://arxiv.org/abs/2207.11243", + "45. Xing, J., Xia, M., Zhang, Y., Cun, X., Wang, J., Wong, T.T.: Codetalker: Speechdriven 3d facial animation with discrete motion prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12780-12790 (2023)", + "46. Xu, Y., Feng, A.W., Marsella, S., Shapiro, A.: A practical and configurable lip sync method for games. In: Proceedings of Motion on Games. p. 131-140. MIG '13, Association for Computing Machinery, New York, NY, USA (2013). https://doi.org/10.1145/2522628.2522904, https://doi.org/10.1145/2522628.2522904", + "47. Yi, H., Liang, H., Liu, Y., Cao, Q., Wen, Y., Bolkart, T., Tao, D., Black, M.J.: Generating holistic 3d human motion from speech. In: CVPR (2023)", + "48. Yin, L., Wei, X., Sun, Y., Wang, J., Rosato, M.: A 3d facial expression database for facial behavior research. In: 7th International Conference on Automatic Face and Gesture Recognition (FGR06). pp. 211-216 (2006). https://doi.org/10.1109/FGR.2006.6", + "49. Zhong, W., Fang, C., Cai, Y., Wei, P., Zhao, G., Lin, L., Li, G.: Identity-preserving talking face generation with landmark and appearance priors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9729-9738 (June 2023)" + ], + "bbox": [ + 215, + 147, + 785, + 797 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "F. Nocentini et al.", + "bbox": [ + 271, + 114, + 393, + 127 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/ScanTalk_ 3D Talking Heads from Unregistered Scans/34e5d7c9-005f-4b11-aa72-5d7f3ba885f0_model.json b/2024/ScanTalk_ 3D Talking Heads from Unregistered Scans/34e5d7c9-005f-4b11-aa72-5d7f3ba885f0_model.json new file mode 100644 index 0000000000000000000000000000000000000000..cc9bc7cfe46a28e064576dec93295bcfc6eafa72 --- /dev/null +++ b/2024/ScanTalk_ 3D Talking Heads from Unregistered Scans/34e5d7c9-005f-4b11-aa72-5d7f3ba885f0_model.json @@ -0,0 +1,2205 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.339, + 0.141, + 0.668, + 0.188 + ], + "angle": 0, + "content": "ScanTalk: 3D Talking Heads from Unregistered Scans" + }, + { + "type": "text", + "bbox": [ + 0.253, + 0.215, + 0.751, + 0.247 + ], + "angle": 0, + "content": "Federico Nocentini\\(^{1}\\), Thomas Besnier\\(^{2}\\), Claudio Ferrari\\(^{4}\\), Sylvain Arguillere\\(^{5}\\), Stefano Berretti\\(^{1}\\), and Mohamed Daoudi\\(^{2,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.256, + 0.774, + 0.412 + ], + "angle": 0, + "content": "\\(^{1}\\) Media Integration and Communication Center (MICC), University of Florence, Italy federico.nocentini@unifi.it, stefano.berretti@unifi.it \n\\(^{2}\\) Univ. Lille, CNRS, Centrale Lille, UMR 9189 CRISTAL, F-59000 Lille, France thomas.besnier@univ-lille.fr \n\\(^{3}\\) IMT Nord Europe, Institut Mines-Télécom, Centre for Digital Systems mohamed.daoudi@imt-nord-europe.fr \n\\(^{4}\\) Department of Architecture and Engineering University of Parma, Italy claudio.ferrari2@unipr.it \n\\(^{5}\\) Univ. Lille, CNRS, UMR 8524 Laboratoire Paul Painlevé, Lille, F-59000, France sylvain,arguillere@univ-lille.fr" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.467, + 0.744, + 0.745 + ], + "angle": 0, + "content": "Abstract. Speech-driven 3D talking heads generation has emerged as a significant area of interest among researchers, presenting numerous challenges. Existing methods are constrained by animating faces with fixed topologies, wherein point-wise correspondence is established, and the number and order of points remains consistent across all identities the model can animate. In this work, we present ScanTalk, a novel framework capable of animating 3D faces in arbitrary topologies including scanned data. Our approach relies on the DiffusionNet architecture to overcome the fixed topology constraint, offering promising avenues for more flexible and realistic 3D animations. By leveraging the power of DiffusionNet, ScanTalk not only adapts to diverse facial structures but also maintains fidelity when dealing with scanned data, thereby enhancing the authenticity and versatility of generated 3D talking heads. Through comprehensive comparisons with state-of-the-art methods, we validate the efficacy of our approach, demonstrating its capacity to generate realistic talking heads comparable to existing techniques. While our primary objective is to develop a generic method free from topological constraints, all state-of-the-art methodologies are bound by such limitations. Code for reproducing our results, and the pre-trained model are available at https://github.com/miccunifi/ScanTalk." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.76, + 0.721, + 0.774 + ], + "angle": 0, + "content": "Keywords: 3D Talking Heads \\(\\cdot\\) 3D Scans Animation \\(\\cdot\\) DiffusionNet" + }, + { + "type": "page_footnote", + "bbox": [ + 0.221, + 0.825, + 0.363, + 0.84 + ], + "angle": 0, + "content": "* Equal contribution" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.395, + 0.127 + ], + "angle": 0, + "content": "F. Nocentini et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.148, + 0.784, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.297, + 0.788, + 0.34 + ], + "angle": 0, + "content": "Fig. 1: We present ScanTalk, a deep learning architecture to animate any 3D face mesh driven by a speech. ScanTalk is robust enough to learn on multiple unrelated datasets with a unique model, whilst allowing us to infer on unregistered face meshes." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.369, + 0.376, + 0.385 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.402, + 0.788, + 0.568 + ], + "angle": 0, + "content": "Human face animation is a complex task, widely explored in Computer Vision and Computer Graphics because of the broad range of applications drawn from it, spanning from virtual reality to video game graphics, and more. As 3D face models continue to improve, it becomes more and more relevant to incorporate multi-modal aspects such as speech with audio data. Speech-driven facial animation encounters challenges inherent to the fact that finding a cross-modality mapping from audio to geometric data is an ill-posed problem, dealing with the complex geometry of human faces and the limited availability of paired audio-visual 3D data. Another, less discussed aspect that touches a broader range of 3D graphics with meshes is the robustness to changes in the topology of the mesh, which refers to the arrangement of the vertices and how they are connected." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.788, + 0.659 + ], + "angle": 0, + "content": "Maintaining fidelity and coherence across different topologies is crucial for ensuring that facial animations remain realistic and expressive, regardless of variations in the underlying mesh structure. This challenge becomes particularly pronounced in speech-driven facial animation, where the dynamics of speech-related lip movements, and related face changes necessitate a high degree of adaptability within the mesh topology." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.66, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Addressing these challenges requires innovative approaches able to navigate the complexities of facial geometry, while accommodating the nuances of speech-driven animation. In this regard, we present ScanTalk, a novel framework capable of animating faces in arbitrary topologies including scanned data. ScanTalk overcomes limitations associated with fixed topologies, offering promising avenues for more flexible and realistic 3D talking heads generation. Indeed, the aforementioned constraint limits the range of applications of current deep learning models for speech-driven motion synthesis. Because of this, deep models are required to train on large scale registered datasets such as FLAME [23] or Multiface [44] for human face data. Building these datasets is a costly procedure, and the trained model is then only usable within the same registered setting. This typically means that any newly acquired data needs to be fitted to the corre" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.374, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "ScanTalk: 3D Talking Heads from Unregistered Scans" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.192 + ], + "angle": 0, + "content": "sponding topology before any animation can be predicted by a deep model. This extra preprocessing step usually prevents online applications, therefore canceling one of the key benefits of deep models, which is their inference speed." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.785, + 0.238 + ], + "angle": 0, + "content": "Aiming to address these limitations, in this paper we present a flexible deep learning framework built to generate speech-driven animations of 3D face meshes. In particular, it gathers several key elements:" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.25, + 0.784, + 0.279 + ], + "angle": 0, + "content": "1. A new robust approach to generate mesh deformation sequences based on DiffusionNet to compute intrinsic descriptors on 3D data;" + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.281, + 0.785, + 0.34 + ], + "angle": 0, + "content": "2. A comprehensive architecture for learning speech-driven animations. Our approach works with meshes with different topologies, while showing competitive performance with respect to state-of-the-art models trained on an individual topology;" + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.342, + 0.784, + 0.372 + ], + "angle": 0, + "content": "3. We show the generalizability of our model to unseen mesh topologies with qualitative examples." + }, + { + "type": "list", + "bbox": [ + 0.222, + 0.25, + 0.785, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.397, + 0.388, + 0.414 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.432, + 0.785, + 0.583 + ], + "angle": 0, + "content": "In literature, there are several methodologies to acquire 3D face meshes, either by extracting the geometry from images or videos [26,47] or by using a complex set of instruments in a controlled environment [16, 18, 33, 35, 44, 48]. The former offers advantages in terms of easier and more cost-effective data acquisition. However, these methods may sometimes fall short in capturing the complete 3D information from 2D data. With 3D scans, other challenges arise: they come unregistered and may present alterations such as holes and noise. Then, animating these objects directly becomes even more challenging. One can do it by registering the meshes onto a pre-defined topology [16, 22, 23, 28, 44] before animating the registrations with 3D morphable models [14] for example." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.786, + 0.78 + ], + "angle": 0, + "content": "Several deep learning models [3-5, 10, 25] proposed ways to learn a latent representation of face scans using robust encoders such as PointNet [6, 31] and Transformers but the resulting mesh is registered, which tends to smooth out some details from the scan geometry. However, this extra registration step may be handled efficiently with recent industrial applications such as MetaHuman from Epic Games. More recently, and closer to our goal, DiffusionNet [37] was combined with neural Jacobian fields [1] in Neural Face Rigging (NFR) [32] to learn a per-triangle deformation field on faces, allowing to transfer an animation from a sequence of unregistered face meshes to another. We stand out from this work by proposing a model that uses audio data to animate a given unregistered face mesh. While DiffusionNet has demonstrated its capability to generalize across varying triangulations in a static setting, our model is the first to exhibit similar properties in a multi-modal 4D setting." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.781, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In recent years, numerous models and methodologies have emerged to address the challenge of synchronizing facial animations with speech audio. While significant progress has been made on 2D talking heads [2, 7, 12, 21, 42, 43, 49], only a handful of approaches can be seamlessly extended to 3D data. Procedural" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.395, + 0.128 + ], + "angle": 0, + "content": "F. Nocentini et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.448 + ], + "angle": 0, + "content": "techniques have been proposed to animate 3D faces [9, 13, 27, 41, 46], primarily relying on visemes (groups of phonemes) to drive the movements of facial muscles. However, many of these models necessitate re-targeting and struggle to generalize effectively without extensive processing steps. More recent works leverage the increasing availability of data to develop statistical methods, including deep learning strategies [11, 15, 19, 29, 34, 38-40, 45]. Currently, state-of-the-art deep learning approaches are limited to a fixed mesh topology, requiring it to be identical to the one observed during the training phase. Among these models, VOCA [11] pioneers the development of deep models trained on a large-scale dataset of registered meshes. It was outperformed when MeshTalk [34] introduced a larger registered dataset named Multiface [44] along with a more expressive model. Moving forward, FaceFormer [15] utilized the transformer architecture, optimized for temporal data, and leveraged the power of a large-scale pre-trained audio encoder called Wav2vec2 [36]. This strategy saw further development with CodeTalker [45] and SelfTalk [30]. More recently, FaceXHubert [19] and FaceDiffuser [38] utilized an improved audio encoder, Hubert [20], demonstrating superiority to predict cross-modality mappings from audio to face motions. While continuously enhancing performances, these models are bound to a fixed mesh topology, hindering real-world applications and limiting the quantity of usable data for a single model, both for training and inference." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.45, + 0.788, + 0.541 + ], + "angle": 0, + "content": "In response to these challenges, we present a novel framework in the landscape of 3D facial animation learning, avoiding the limitations posed by the specific topology of the face mesh to be animated. The proposed model stands out for its capability to animate any face mesh, including real-world 3D scans. This novel approach holds the potential to redefine the standards in the field of 3D facial animation, being applicable across diverse datasets and scenarios." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.565, + 0.54, + 0.583 + ], + "angle": 0, + "content": "3 Proposed approach: ScanTalk" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.788, + 0.718 + ], + "angle": 0, + "content": "In this section, we introduce Scantalk, a framework for animating 3D face meshes reproducing a spoken sentence contained in an audio file, which does not require the meshes to adhere to any specific topology. With ScanTalk, we push the state-of-the-art in 3D talking heads a step forward by allowing any 3D face, even raw scans, to be animated given a speech. To train ScanTalk, the only requirement is that the training meshes share a common topology within each sequence, but the topology may vary from one sequence to another. At inference time, any 3D face mesh in neutral state can be animated." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "ScanTalk is an Encoder-Decoder framework, which receives a neutral face mesh and an audio snippet as input, and outputs a sequence of per-vertex deformation fields, whose length depends on that of the audio. By summing each deformation field to the neutral input face, we ultimately obtain the animated sequence. The encoder is composed of two main modules: an audio encoder, which combines a pretrained encoder with a bi-directional LSTM that extracts audio features from the input speech, and a DiffusionNet encoder that computes surface descriptors from the neutral 3D face. These descriptors are then repli" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.374, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "ScanTalk: 3D Talking Heads from Unregistered Scans" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.208 + ], + "angle": 0, + "content": "ated and concatenated to the audio features and the resulting signal is fed to a DiffusionNet decoder that outputs the deformation to be applied to the neutral face. The framework is depicted in Fig. 2. Before providing the technical details in Sec. 3.1 and Sec. 3.2, below we introduce a few general notations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.213, + 0.789, + 0.357 + ], + "angle": 0, + "content": "Let \\( L = \\{(M_i^{gt}, m_i^n, A_i)\\}_{i=0}^{N-1} \\) denotes the training set comprising \\( N \\) samples, where \\( A_i \\) is an audio containing a spoken sentence, \\( M_i^{gt} = (m_i^0, \\ldots, m_i^{T_i-1}) \\in \\mathbb{R}^{T_i \\times V_i \\times 3} \\) represents a sequence of 3D faces (same topology) of length \\( T_i \\) synchronized with the spoken sentence in \\( A_i \\), and \\( m_i^n \\in \\mathbb{R}^{V_i \\times 3} \\) is a 3D neutral face. \\( V_i = |m_i^n| \\) is the number of vertices in the \\( i \\)-th 3D face sequence that needs to be consistent across each mesh in the \\( i \\)-th sequence together with the vertex connectivity, overall determining the topology of the surface. Our objective is to establish a mapping function that correlates an audio input \\( A_i \\), and a neutral 3D face \\( m_i^n \\), to the ground-truth sequence \\( M_i^{gt} \\), expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.408, + 0.379, + 0.786, + 0.398 + ], + "angle": 0, + "content": "\\[\nS c a n t a l k \\left(A _ {i}, m _ {i} ^ {n}\\right) \\approx M _ {i} ^ {g t}. \\tag {1}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.472, + 0.789, + 0.675 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.684, + 0.789, + 0.797 + ], + "angle": 0, + "content": "Fig. 2: Architecture of ScanTalk. A novel Encoder-Decoder framework designed to dynamically animate any 3D face based on a spoken sentence from an audio file. The Encoder integrates the 3D neutral face \\( m_i^n \\), per-vertex surface features \\( P_i^n \\) (crucial for DiffusionNet and precomputed by the operators \\( OP \\)), and the audio file \\( A_i \\), yielding a fusion of per-vertex and audio features. These combined descriptors, alongside \\( P_i^n \\), are then passed to the Decoder, which mirrors a reversed DiffusionNet encoder structure. The Decoder predicts the deformation of the 3D neutral face, which is then combined with the original 3D neutral face \\( m_i^n \\) to generate the animated sequence." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.395, + 0.128 + ], + "angle": 0, + "content": "F. Nocentini et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.411, + 0.16 + ], + "angle": 0, + "content": "3.1 ScanTalk Encoder" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.172, + 0.788, + 0.218 + ], + "angle": 0, + "content": "ScanTalk requires an audio snippet \\( A_{i} \\) and a 3D face in neutral state \\( m_{i}^{n} \\) as inputs. To process the two inputs, ScanTalk comprises two distinct encoders to process the mesh and the audio, respectively, as detailed below." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.24, + 0.788, + 0.376 + ], + "angle": 0, + "content": "Face mesh encoder. Several approaches are available for encoding face meshes or point clouds, yet traditional graph convolution-based models, like [17, 24], encounter limitations with varying graph structures, such as changes in mesh resolution. To address this, we employ DiffusionNet [37], a discretization-agnostic encoder that demonstrated to be effective for encoding face meshes as seen in [32]. DiffusionNet integrates multi-layer perceptrons (MLPs), learned diffusion, and spatial gradient features, offering a straightforward yet robust architecture for surface learning tasks. It bypasses the need for complex operations like explicit surface convolutions or pooling hierarchies." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.377, + 0.788, + 0.466 + ], + "angle": 0, + "content": "Critical to DiffusionNet's functionality are precomputed features of the face surface, namely the Cotangent Laplacian, Eigenbasis, Mass Matrix, and Spatial Gradient Matrix. Together, these operators enhance the architecture's robustness, flexibility, and expressiveness, making DiffusionNet suitable for diverse surface learning tasks. Notably, this architecture accommodates 3D faces of any topology, allowing for variations in the number and order of defining points." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.467, + 0.788, + 0.543 + ], + "angle": 0, + "content": "Let \\( P_{i}^{n} = OP(m_{i}^{n}) \\) represent the precomputed features obtained by applying the above surface closed-form operators \\( OP \\) to the 3D neutral face \\( m_{i}^{n} \\). The DiffusionNet Encoder \\( DN_{e} \\), with latent size \\( h \\), is designed to process a neutral 3D face mesh \\( m_{i}^{n} \\), and requires the precomputed per-vertex features \\( P_{i}^{n} \\) to extract a per-vertex descriptors \\( f_{i}^{n} \\), that is:" + }, + { + "type": "equation", + "bbox": [ + 0.399, + 0.554, + 0.785, + 0.573 + ], + "angle": 0, + "content": "\\[\nf _ {i} ^ {n} = D N _ {e} \\left(m _ {i} ^ {n}, P _ {i} ^ {n}\\right) \\in \\mathbb {R} ^ {V _ {i} \\times h}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.787, + 0.613 + ], + "angle": 0, + "content": "The per-vertex descriptors \\( f_{i}^{n} \\), is capable of capturing intricate details of each vertex within the neutral 3D face." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.636, + 0.788, + 0.712 + ], + "angle": 0, + "content": "Audio encoder. Following [19, 38], the speech encoder adopts the architecture of the state-of-the-art pretrained speech model, HuBERT [20] that is a self-supervised speech representation learner utilizing an offline clustering step to provide aligned target labels for a BERT-like prediction loss. Using this module, followed by a Linear Layer, we obtain a per-frame audio representation:" + }, + { + "type": "equation", + "bbox": [ + 0.367, + 0.723, + 0.785, + 0.74 + ], + "angle": 0, + "content": "\\[\na _ {i} = \\operatorname {S p e e c h E n c o d e r} \\left(A _ {i}\\right) \\in \\mathbb {R} ^ {T _ {i} \\times (h / 2)}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.752, + 0.788, + 0.813 + ], + "angle": 0, + "content": "To ensure coherence between the speech representation, following the methodology outlined in [29], we concatenate the SpeechEncoder with a Multilayer Bidirectional-LSTM for temporal consistency, in a way such that the speech signal is projected into a temporal latent representation:" + }, + { + "type": "equation", + "bbox": [ + 0.402, + 0.824, + 0.785, + 0.842 + ], + "angle": 0, + "content": "\\[\nv _ {i} = B i L S T M \\left(a _ {i}\\right) \\in \\mathbb {R} ^ {T _ {i} \\times h}. \\tag {4}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.374, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "ScanTalk: 3D Talking Heads from Unregistered Scans" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.411, + 0.16 + ], + "angle": 0, + "content": "3.2 ScanTalk Decoder" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.173, + 0.784, + 0.203 + ], + "angle": 0, + "content": "We combine per-vertex descriptors \\( f_{i}^{n} \\) extracted from the neutral face, with the latent vector \\( v_{i} \\) extracted from the Bidirectional-LSTM" + }, + { + "type": "equation", + "bbox": [ + 0.228, + 0.215, + 0.786, + 0.233 + ], + "angle": 0, + "content": "\\[\n(F _ {i} ^ {j}) _ {k} = \\left(f _ {i} ^ {n}\\right) _ {k} \\oplus v _ {i} ^ {j} \\in \\mathbb {R} ^ {h * 2}, \\quad \\forall k = 0, \\dots , V _ {i} - 1, \\quad \\forall j = 0, \\dots , T _ {i} - 1. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.245, + 0.785, + 0.339 + ], + "angle": 0, + "content": "With this concatenation, we obtain a combined latent \\( F_{i}^{j} \\in \\mathbb{R}^{V_{i} \\times h*2} \\) that embeds both audio-related and geometry-related latents. To decode this sequence of combined latents for deforming the neutral face \\( m_{i}^{n} \\), we employ a DiffusionNet Decoder (which is essentially a reversed Encoder), denoted as \\( DN_{d} \\). The decoder module receives \\( F_{i}^{j} \\), and precomputed features \\( P_{i}^{n} \\) derived from \\( m_{i}^{n} \\). It predicts the deformation of \\( m_{i}^{n} \\), denoted as \\( DN_{d}(F_{i}^{j}, P_{i}^{n}) \\):" + }, + { + "type": "equation", + "bbox": [ + 0.374, + 0.352, + 0.785, + 0.369 + ], + "angle": 0, + "content": "\\[\n\\widehat {m _ {i} ^ {j}} = D N _ {d} \\left(F _ {i} ^ {j}, P _ {i} ^ {n}\\right) + m _ {i} ^ {n} \\in \\mathbb {R} ^ {V _ {i} \\times 3}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.382, + 0.784, + 0.519 + ], + "angle": 0, + "content": "Here \\(\\widehat{m_i^j}\\) represents the \\(j\\)-th frame of the predicted sequence. The entire generated sequence is defined by \\(\\widehat{M_i} \\in \\mathbb{R}^{T_i \\times V_i \\times 3}\\). We opt to utilize ScanTalk for predicting the deformation of the neutral face \\(m_i^n\\) rather than predicting the actual face. This decision aligns with previous works [15, 19, 30, 38, 40], and offers advantages in terms of training efficiency and resulting animation. Predicting face deformation makes the learning process easier, by focusing solely on speech-related motion, as opposed to incorporating the problem of predicting the entire face reconstruction. Essentially, the decoder learns to predict a per-vertex displacement field from a time-dependent per-vertex descriptors field." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.542, + 0.414, + 0.558 + ], + "angle": 0, + "content": "3.3 ScanTalk Training" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.785, + 0.764 + ], + "angle": 0, + "content": "ScanTalk generates deformations of a neutral face; hence, a predicted sequence maintains a consistent topology across all frames. This property arises from the definition of the DiffusionNet Decoder, which necessitates knowledge of precomputed features \\( P_{i}^{n} \\), and the number of points in the 3D neutral face targeted for animation. For this reason, during a supervised training protocol, the ground-truth meshes within each sequence must adhere to a common topology. Despite the apparent specificity of this requirement, it proves to be non-prohibitive in practice. ScanTalk animates a neutral face in response to an audio input, producing a sequence of 3D faces sharing the same topology of the neutral face that is animated, which can be any topology (even different from those seen during training). This alignment between the training strategy and the desired inference outcome underscores the efficiency of ScanTalk, which is not restricted to the topology of the training faces." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Notably, the model predicts per-vertex displacements of the neutral ground-truth mesh and the ground truth sequence is registered in a supervised setting. Consequently, the predicted sequence and the ground-truth sequence are aligned on the same topology but can vary from one sequence to another. Thus, during training, we minimize the average vertex-wise Mean Squared Error (MSE) over" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.395, + 0.128 + ], + "angle": 0, + "content": "F. Nocentini et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.146, + 0.785, + 0.179 + ], + "angle": 0, + "content": "a sequence of length \\( T_{i} \\) between the ground truth \\( M_{i}^{gt} \\) and the model prediction \\( \\widehat{M}_{i} \\). which is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.317, + 0.192, + 0.785, + 0.233 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {M S E} = \\frac {1}{T _ {i} - 1} \\sum_ {j = 0} ^ {T _ {i} - 1} \\frac {1}{V _ {i} - 1} \\sum_ {k = 0} ^ {V _ {i} - 1} \\left\\| \\left(m _ {i} ^ {j}\\right) _ {k} - \\left(\\widehat {m _ {i}} ^ {j}\\right) _ {k} \\right\\| _ {2} ^ {2}. \\tag {7}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.255, + 0.376, + 0.273 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.288, + 0.784, + 0.378 + ], + "angle": 0, + "content": "In the following, we first introduce the datasets and the metrics used for evaluation, respectively in Sec. 4.1 and Sec. 4.2. Then, we report quantitative and qualitative results in comparison with state-of-the-art methods in Sec. 4.3 and Sec. 4.4. In Sec. 4.5 we present several ablation studies over the framework architecture. Finally, in Sec. 4.6, we report the results of an user study designed to compare our approach with state-of-the-art methods." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.402, + 0.334, + 0.415 + ], + "angle": 0, + "content": "4.1 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.429, + 0.784, + 0.503 + ], + "angle": 0, + "content": "For training and quantitative evaluations, we rely on three classical datasets in 3D speech-driven motion synthesis: VOCAset [11], BIWI [16] and Multiface [44]. VOCAset gathers mesh sequences of 12 actors performing 40 speeches, captured at 60fps. Each sequence lasts for around 3 to 5 seconds, and each mesh is registered to the FLAME [23] topology with 5,023 vertices and 9,976 faces." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.504, + 0.785, + 0.564 + ], + "angle": 0, + "content": "BIWI comprises 14 subjects articulating 40 sentences each, sampled at 25fps. Each sentence lasts around 5 seconds, and the meshes are registered to a fixed topology. Due to GPU limitations, we used a downsampled version of this dataset, called \\(\\mathrm{BIWI}_6\\), which has a fixed topology, with 3,895 vertices and 7,539 faces." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.564, + 0.784, + 0.608 + ], + "angle": 0, + "content": "Multiface includes 13 identities executing up to 50 speeches of around 4 seconds each, sampled at 30fps. The original multiface meshes have a fixed topology with 5,471 vertices and 10,837 faces." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.61, + 0.784, + 0.715 + ], + "angle": 0, + "content": "For training purposes, the meshes in both Multiface and BIWI\\(_6\\) have been scaled and aligned with the meshes in the VOCAset dataset. The data splits can be found in the supplementary material. Since ScanTalk is the first topology-independent 3D talking head generator, for the sake of a comprehensive comparison with the state-of-the-art, we trained 4 different models: one is trained using all the three datasets together (multi-dataset), while the other three are trained on each dataset separately (single-dataset)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.739, + 0.42, + 0.753 + ], + "angle": 0, + "content": "4.2 Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "To evaluate the quality of the generated faces, we employ three standard metrics from previous works [30,34,38,45], namely: Lip vertex error (LVE) \\(\\times 10^{-5}(mm)\\) Assesses the lip deviation in a sequence by comparing it to the ground truth. It is obtained by computing the maximum L2 error for all lip vertices in each frame, averaged across all frames." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.374, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "ScanTalk: 3D Talking Heads from Unregistered Scans" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.191 + ], + "angle": 0, + "content": "Mean vertex error (MVE) \\(\\times 10^{-3}(mm)\\). Describes the quality of the overall face reconstruction by computing the maximal vertex L2 error for each frame and taking the mean across all frames." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.192, + 0.784, + 0.222 + ], + "angle": 0, + "content": "Upper-face dynamic deviation (FDD) \\(\\times 10^{-7}(mm)\\). Compares the standard deviation of upper vertices between the generated sequence and ground-truth." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.274, + 0.432, + 0.289 + ], + "angle": 0, + "content": "4.3 Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.312, + 0.785, + 0.463 + ], + "angle": 0, + "content": "Although ScanTalk is the first deep learning model to animate unregistered 3D face meshes, comparisons with the state-of-the-art is possible when inferring our model on registered data. In Tab. 1, we present the performance evaluation of ScanTalk training on both a single-dataset and multi-datasets, juxtaposed against several state-of-the-art methods trained solely on a single dataset. While the augmentation of training data enhances the effectiveness of our model, addressing disparities in data composition becomes important for animating unseen faces. Particularly, the varied geometries and head dynamics inherent in \\(\\mathrm{BIWI}_6\\), VOCA, and Multiface meshes pose significant challenges, potentially limiting performance compared to a model trained exclusively on a single dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.507, + 0.785, + 0.618 + ], + "angle": 0, + "content": "Table 1: ScanTalk performance, in both single-dataset (s-d) and multi-dataset (m-d) scenarios, in comparison with state-of-the-art methods. The heatmaps show the differences between the first frame and subsequent frames within sequences among the VOCAset, \\(\\mathrm{BIWI}_6\\), and Multiface datasets. Notably, in the VOCAset, primarily the lips display movement, whereas in \\(\\mathrm{BIWI}_6\\) and Multiface datasets, substantial head and upper face movements are observed. The color gradient on the face meshes corresponds to the average per-vertex \\(L_2\\) norm of the differences, where blue hues indicate lower values, and red hues indicate higher values." + }, + { + "type": "image", + "bbox": [ + 0.37, + 0.633, + 0.42, + 0.695 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.638, + 0.569, + 0.683 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.676, + 0.63, + 0.731, + 0.701 + ], + "angle": 0, + "content": null + }, + { + "type": "table", + "bbox": [ + 0.235, + 0.7, + 0.763, + 0.837 + ], + "angle": 0, + "content": "
VOCAssetBIWI6Multiface
LVE ↓MVE ↓FDD ↓LVE ↓MVE ↓FDD ↓LVE ↓MVE ↓FDD ↓
VOCA6.9930.9832.6625.7432.59141.4824.9232.76155.781
FaceFormer6.1230.9352.1634.0852.16337.0912.4511.45320.239
SelfTalk5.6180.9182.3213.6282.06235.4702.2811.90137.434
CodeTalker3.5490.8882.2585.1902.64120.5994.0912.38247.905
FaceDiffuser4.3500.9012.4374.0222.12839.6043.5552.38829.157
ScanTalk m-d6.3750.9872.1014.0442.05740.0512.4351.67832.202
ScanTalk s-d3.0120.8612.4004.6512.14836.0342.6531.87164.451
" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.395, + 0.128 + ], + "angle": 0, + "content": "F. Nocentini et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.328 + ], + "angle": 0, + "content": "This observation becomes evident when assessing Scantalk's performance on VOCAset, where a substantial improvement is observed when training on a single-dataset. Conversely, in the case of \\(\\mathrm{BIWI}_6\\) and Multiface single-dataset training, a decline in results is noted compared to the multi-dataset version. This disparity is attributed to the presence of speech-related head and upper face movements within the ground truth talking head sequences of \\(\\mathrm{BIWI}_6\\) and Multiface. Consequently, the multi-dataset training model demonstrates superior performance as it encounters a wider array of sequences featuring head and upper face movements. In contrast, VOCAset lacks such movements, with the 3D faces remaining static, while only the mouth moves, thereby enhancing performance for ScantTalk trained solely on VOCAset. This obviates the need to learn other movements of the head or upper face." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.33, + 0.788, + 0.419 + ], + "angle": 0, + "content": "The distinctions between VOCAset and the other two datasets are further underscored by the FDD metric, indicating that 3D faces in VOCAset remain stationary, with minimal activity in the upper facial area and head. Across all tested models, generating sequences with lower FDD is notably more achievable on VOCAset compared to the other datasets, thus emphasizing the disparities between VOCAset and the others." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.421, + 0.574, + 0.616 + ], + "angle": 0, + "content": "Nevertheless, the results presented in Tab. 1 demonstrate that ScanTalk produces outcomes comparable to state-of-the-art methods, whether undergoing single-dataset or multi-dataset training. Moreover, ScanTalk emerges as the first model capable of training in multi-dataset settings, demonstrating the ability to animate a 3D face regardless of its topology. An interesting characteristic of ScanTalk emerges when analyzing the GPU memory usage relative to the vertex count in the 3D facial model requiring animation. Fig. 3 illustrates a linear increase in GPU usage correlated with the number of vertices." + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.45, + 0.779, + 0.55 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.583, + 0.56, + 0.788, + 0.601 + ], + "angle": 0, + "content": "Fig.3: ScanTalk GPU memory usage with respect to the mesh resolution." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.644, + 0.421, + 0.659 + ], + "angle": 0, + "content": "4.4 Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.84 + ], + "angle": 0, + "content": "In the domain of 3D speech-driven talking heads generation, qualitative evaluations hold as much significance as quantitative assessments. To evaluate the quantitative findings presented in Sec. 4.3, we introduce a straightforward test aimed at discerning disparities in head and upper face movements across the three datasets: for each dataset, we compute the average \\( L_{2} \\) norm of the differences between the initial frame and subsequent frames within each sequence. This enables the quantification of both head dynamics and upper facial movements across mesh sequences. In the heatmaps of Tab. 1, we present the per-frame average difference for each dataset. Our analysis reveals that in the VOCAset, only the lower part of the face exhibits discernible movements, whereas in BIWI\\(_6\\) and Multiface, the entire head or face moves." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.374, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "ScanTalk: 3D Talking Heads from Unregistered Scans" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.249, + 0.151, + 0.752, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.364, + 0.788, + 0.448 + ], + "angle": 0, + "content": "Fig. 4: Relative norm of the per-vertex descriptors \\( f_{i}^{n} \\) in Eq. (2) extracted by \\( DN_{e} \\) displayed as a heatmap on a mesh from VOCAset (left), and a mesh from Multiface (right). For each mesh, we show the norm on the original topology, on a remeshed version, and on a further degraded mesh obtained by removing the back of the head and creating random holes. Here, pinker hues indicate lower values, and greener hues indicate higher values." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.47, + 0.789, + 0.743 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.755, + 0.788, + 0.798 + ], + "angle": 0, + "content": "Fig. 5: ScanTalk inference on 3D faces. The meshes have been rigidly aligned with the training data. The first row is an animation of a raw 3D scan (an hole for the mouth has been created), while the others two are animation of meshes in an arbitrary topology." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.395, + 0.128 + ], + "angle": 0, + "content": "F. Nocentini et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.236 + ], + "angle": 0, + "content": "In Fig. 4, we present a heatmap visualization representing the norm of geometric descriptors \\( f_{i}^{n} \\) computed by the encoder \\( DN_{e} \\) in Eq. (2) for various alterations of the same identity mesh. Our findings indicate that the ScanTalk encoder effectively extracts descriptors from input faces regardless of facial topology. Notably, despite differences in facial topology among the same identity, the extracted descriptors exhibit remarkable similarity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.785, + 0.343 + ], + "angle": 0, + "content": "In Fig. 5, 3D faces in different topology, including a scan, animated using ScanTalk are presented. Our method demonstrates good generalization capabilities, successfully animating diverse 3D faces. Notably, enhanced animation quality is observed when a mouth aperture is present; however, our model performs well even in its absence. Nonetheless, without a mouth aperture, the model struggles to generate the corresponding mouth opening, despite accurate lip movement synchronization." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.367, + 0.409, + 0.381 + ], + "angle": 0, + "content": "4.5 Ablations Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.393, + 0.784, + 0.468 + ], + "angle": 0, + "content": "To investigate the impacts of various components within our architecture, we conducted extensive experiments across various configurations, encompassing both single-dataset and multi-dataset training paradigms. The results reported in Tab. 2 and Tab. 3 are computed by modifying ScanTalk's modules and training losses as described in Sec. 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.492, + 0.784, + 0.596 + ], + "angle": 0, + "content": "Audio encoder. In learning based speech-driven animation, some previous work used either Wav2vec2 [36] or Hubert [20], the latter being substantially better for this task according to [19,38]. For our purpose, we also tested WavLM [8] but obtained poorer efficiency. The results on both the single-dataset and multi-dataset training are displayed in Tab. 2a and Tab. 3 top. We can see that, in both cases, the usage of the Hubert Encoder [19] for audio features extraction leads to better results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.62, + 0.785, + 0.786 + ], + "angle": 0, + "content": "Temporal consistency. In our investigation, as proven by [19,29,40], we found that supplementing pre-trained audio encoders with additional temporal consistency mechanisms such as Bidirectional-LSTM, Bidirectional-GRU, or an autoregressive Transformer Decoder (TD) significantly enhances model performance, as illustrated in Tab. 2b and in Tab. 3 middle. The BiLSTM model architecture is detailed in Sec. 3, while the BiGRU model mirrors the BiLSTM architecture but substitutes BiLSTM with BiGRU. Details about the Transformer Decoder can be found in the supplementary material. From the findings presented in Tab. 2b and in Tab. 3 middle, we observe that employing a multilayer Bidirectional-LSTM in the audio stream processing of ScanTalk yields the most favorable performance across both single-dataset training and multi-dataset training scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Loss function. Previous research has extensively explored optimal objective functions for enhancing and refining the learning process. Inspired by [29,30,40]," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.374, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "ScanTalk: 3D Talking Heads from Unregistered Scans" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.269 + ], + "angle": 0, + "content": "we experimented with Mean Square Error (\\( L_{MSE} \\)), Masked (\\( L_{mask} \\)), and Velocity (\\( L_{vel} \\)) Loss. The former is a simple \\( L_2 \\) loss, the second employs Mean Square Error focused on lip vertices, while the latter aims to minimize differences between consecutive frames. While these loss functions have demonstrated efficacy in single-dataset training, as evidenced by [29,30,40], our findings, detailed in Tab. 3 bottom, indicate that employing a straightforward \\( L_2 \\), on multi-dataset training, enhances the model's capacity to generate realistic talking heads. We attribute this to significant geometric variations observed across different datasets." + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.294, + 0.787, + 0.322 + ], + "angle": 0, + "content": "Table 2: ScanTalk (ST) single-dataset ablation studies. Results obtained with a model trained just on VOCAsset." + }, + { + "type": "table_caption", + "bbox": [ + 0.277, + 0.346, + 0.446, + 0.358 + ], + "angle": 0, + "content": "(a) Audio Encoder Ablation." + }, + { + "type": "table", + "bbox": [ + 0.245, + 0.368, + 0.479, + 0.43 + ], + "angle": 0, + "content": "
LVE ↓MVE ↓FDD ↓
ST w WavLM3.6740.9372.413
ST w Wav2Vec23.3090.8602.244
ScanTalk3.0120.8612.400
" + }, + { + "type": "table_caption", + "bbox": [ + 0.541, + 0.346, + 0.744, + 0.359 + ], + "angle": 0, + "content": "(b) Temporal consistency ablation." + }, + { + "type": "table", + "bbox": [ + 0.533, + 0.368, + 0.752, + 0.44 + ], + "angle": 0, + "content": "
LVE ↓MVE ↓FDD ↓
ST w/o3.3610.8702.365
ST w TD3.2910.8592.406
ST w BiGRU3.0360.8352.358
ScanTalk3.0120.8612.400
" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.495, + 0.788, + 0.539 + ], + "angle": 0, + "content": "Table 3: ScanTalk (ST) multi-dataset ablation studies. Top: ST with different audio encoders. Middle: ST using different audio stream processing. Bottom: ST with different Loss function. The last row is the proposed ScanTalk" + }, + { + "type": "table", + "bbox": [ + 0.236, + 0.55, + 0.769, + 0.705 + ], + "angle": 0, + "content": "
VOCAsetBIWI6Multiface
LVE ↓MVE ↓FDD ↓LVE ↓MVE ↓FDD ↓LVE ↓MVE ↓FDD ↓
ST w Wav2Vec27.1271.2412.3494.4392.27641.3743.1982.32933.671
ST w WavLM6.7731.0042.2204.3332.16231.9222.9072.0048.500
ST w TD6.5120.9951.8584.8272.16836.7892.1361.72535.044
ST w BiGRU6.5840.9942.0544.4212.10341.4562.5341.98734.098
ST+Lmask+Lvel7.0311.0921.4734.7552.22739.4582.5471.89815.973
ST+Lmask7.4511.0840.8594.7482.19936.9392.8881.89132.501
ST+Lvel6.7400.9981.8994.5092.18033.6142.1031.77531.599
ScanTalk6.3750.9872.1014.0442.05740.0512.4351.67832.202
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.755, + 0.355, + 0.77 + ], + "angle": 0, + "content": "4.6 User study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To further evaluate our solution, we performed a study where human feedback is involved. This evaluation is conducted with two user-based studies involving 25 participants; \\((i)\\) For the first study, in alignment with prior research [15,19,30,34, 38,45], we designed an A/B test to compare ScanTalk with other state-of-the-art" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.395, + 0.128 + ], + "angle": 0, + "content": "F. Nocentini et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "models within a registered setting on both lip-syncing and naturalness criteria (Test 1); (ii) In the second test, we assessed the credibility of scan animations by asking participants to rate the animation quality of ten scans sourced from the COMA dataset [33], using a scale ranging from 1 to 10 (Test 2)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.785, + 0.327 + ], + "angle": 0, + "content": "The outcomes of Test 1 are depicted in the table on the left of Fig. 6. Notably, sequences produced using ScanTalk demonstrate levels of both naturalness and lip-syncing that are comparable to state-of-the-art methods. Specifically, our approach is preferred when compared with FaceFormer, CodeTalker, and FaceDiffuser on VOCAsset. However, SelfTalk and the ground truth are preferred over ScanTalk. Nevertheless, the percentage of users favoring ScanTalk remains nonnegligible, underscoring the efficacy of our approach in generating 3D talking heads with good realism and lip-syncing fidelity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.328, + 0.784, + 0.389 + ], + "angle": 0, + "content": "Results of Test 2 are reported on the right of Fig. 6. These ratings are closely linked to both the scan quality and the fidelity of the mouth representation. Despite the scan quality not being optimal, as depicted in Fig. 6, our results indicate that we can achieve a good level of realism in the animated scans." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.445, + 0.455, + 0.51 + ], + "angle": 0, + "content": "
ComparisonNaturalness (%)Lip-sync (%)
ST vs. FaceFormer82.6780.67
ST vs. CodeTalker57.3356.00
ST vs. FaceDiffuser90.6794.00
ST vs. SelfTalk44.0043.33
ST vs. GT44.0042.00
" + }, + { + "type": "image", + "bbox": [ + 0.464, + 0.414, + 0.784, + 0.567 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.577, + 0.785, + 0.647 + ], + "angle": 0, + "content": "Fig. 6: The table on the left shows results of Test 1, ScanTalk (ST) vs. other methods and the ground truth (GT) on samples from the test set of VOCAset. The values denote the percentages of users who favored ScanTalk animations over others. The chart on the right shows results of Test 2 reported as a violin plot for the animation of scans from COMA [33]. The subject median rating is displayed as a black dot on each violin." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.7, + 0.367, + 0.715 + ], + "angle": 0, + "content": "5 Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.732, + 0.785, + 0.837 + ], + "angle": 0, + "content": "This paper introduces ScanTalk, a novel framework for speech-driven 3D facial animation. Unlike existing methods, ScanTalk possesses the unique capability to animate any 3D face, regardless of its topology, even if it differs from the ones on which it was trained. ScanTalk extends the applicability of deep speech-driven 3D facial animations by addressing the challenges of topology robustness. Additionally, our model demonstrates comparable quantitative and qualitative results with other state-of-the-art methods across three distinct datasets." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.374, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "ScanTalk: 3D Talking Heads from Unregistered Scans" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.284 + ], + "angle": 0, + "content": "Acknowledgments This work is supported by the ANR project Human4D ANR-19-CE23-0020 and by the IRP CNRS project GeoGen3DHuman. This work was also partially supported by \"Partenariato FAIR (Future Artificial Intelligence Research) - PE00000013, CUP J33C22002830006\" funded by NextGenerationEU through the italian MUR within NRRP, project DL-MIG. This work was also partially funded by the ministerial decree n.352 of the 9th April 2022 NextGenerationEU through the italian MUR within NRRP. This work was also partially supported by Fédération de Recherche Mathématique des Hauts-de-France (FMHF, FR2037 du CNRS)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.307, + 0.325, + 0.323 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.34, + 0.787, + 0.395 + ], + "angle": 0, + "content": "1. Aigerman, N., Gupta, K., Kim, V.G., Chaudhuri, S., Saito, J., Groueix, T.: Neural jacobian fields: Learning intrinsic mappings of arbitrary meshes. ACM Trans. Graph. 41(4) (jul 2022). https://doi.org/10.1145/3528223.3530141, https://doi.org/10.1145/3528223.3530141" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.396, + 0.787, + 0.437 + ], + "angle": 0, + "content": "2. Alghamdi, M.M., Wang, H., Bulpitt, A.J., Hogg, D.C.: Talking head from speech audio using a pre-trained image generator. In: Proceedings of the 30th ACM International Conference on Multimedia (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.438, + 0.787, + 0.479 + ], + "angle": 0, + "content": "3. Bahri, M., O' Sullivan, E., Gong, S., Liu, F., Liu, X., Bronstein, M.M., Zafeiriou, S.: Shape my face: Registering 3d face scans by surface-to-surface translation. International Journal of Computer Vision (IJCV) (Sep 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.48, + 0.787, + 0.533 + ], + "angle": 0, + "content": "4. Besnier, T., Arguillère, S., Pierson, E., Daoudi, M.: Toward Mesh-Invariant 3D Generative Deep Learning with Geometric Measures. Computers and Graphics (2023). https://doi.org/10.1016/j.cag.2023.06.027, https://hal.science/hal-04143649" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.535, + 0.787, + 0.577 + ], + "angle": 0, + "content": "5. Chandran, P., Zoss, G., Gross, M., Gotardo, P., Bradley, D.: Shape transformers: Topology-independent 3d shape models using transformers. In: Computer Graphics Forum. vol. 41, pp. 195-207. Wiley Online Library (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.578, + 0.787, + 0.619 + ], + "angle": 0, + "content": "6. Charles, R.Q., Su, H., Kaichun, M., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). p. 77-85. IEEE (Jul 2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.62, + 0.787, + 0.659 + ], + "angle": 0, + "content": "7. Chen, L., Cui, G., Liu, C., Li, Z., Kou, Z., Xu, Y., Xu, C.: Talking-head generation with rhythmic head motion. In: European Conference on Computer Vision (2020), https://api-semanticscholar.org/CorpusID:220633152" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.66, + 0.787, + 0.729 + ], + "angle": 0, + "content": "8. Chen, S., Wang, C., Chen, Z., Wu, Y., Liu, S., Chen, Z., Li, J., Kanda, N., Yoshioka, T., Xiao, X., Wu, J., Zhou, L., Ren, S., Qian, Y., Qian, Y., Zeng, M., Wei, F.: Wavlm: Large-scale self-supervised pre-training for full stack speech processing. IEEE Journal of Selected Topics in Signal Processing 16, 1505-1518 (2021), https://api-semanticscholar.org/CorpusID:239885872" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.73, + 0.787, + 0.783 + ], + "angle": 0, + "content": "9. Cosi, P., Caldognetto, E., Perin, G., Zmarich, C.: Labial coarticulation modeling for realistic facial animation. In: Proceedings. Fourth IEEE International Conference on Multimodal Interfaces. pp. 505-510 (2002). https://doi.org/10.1109/ICMI.2002.1167047" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.785, + 0.787, + 0.84 + ], + "angle": 0, + "content": "10. Croquet, B., Christiaens, D., Weinberg, S.M., Bronstein, M., Vandermeulen, D., Claes, P.: Unsupervised diffeomorphic surface registration and non-linear modelling. In: Medical Image Computing and Computer Assisted Intervention (MIC-CAI). p. 118–128. Springer (2021)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.34, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.395, + 0.127 + ], + "angle": 0, + "content": "F. Nocentini et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "11. Cudeiro, D., Bolkart, T., Laidlaw, C., Ranjan, A., Black, M.: Capture, learning, and synthesis of 3D speaking styles. In: Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR). pp. 10101-10111 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.245 + ], + "angle": 0, + "content": "12. Das, D., Biswas, S., Sinha, S., Bhowmick, B.: Speech-driven facial animation using cascaded gans for learning of motion and texture. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J.M. (eds.) Computer Vision - ECCV 2020. pp. 408-424. Springer International Publishing, Cham (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.246, + 0.785, + 0.299 + ], + "angle": 0, + "content": "13. Edwards, P., Landreth, C., Fiume, E., Singh, K.: Jali: an animator-centric viseme model for expressive lip synchronization. ACM Trans. Graph. 35(4) (jul 2016). https://doi.org/10.1145/2897824.2925984, https://doi.org/10.1145/2897824.2925984" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.785, + 0.369 + ], + "angle": 0, + "content": "14. Egger, B., Smith, W.A.P., Tewari, A., Wuhrer, S., Zollhoefer, M., Beeler, T., Bernard, F., Bolkart, T., Kortylewski, A., Romdhani, S., Theobalt, C., Blanz, V., Vetter, T.: 3D Morphable Face Models - Past, Present and Future. ACM Transactions on Graphics 39(5), 157:1-38 (Aug 2020). https://doi.org/10.1145/3395208, https://inria.hal.science/hal-02280281" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.37, + 0.785, + 0.438 + ], + "angle": 0, + "content": "15. Fan, Y., Lin, Z., Saito, J., Wang, W., Komura, T.: Faceformer: Speech-driven 3d facial animation with transformers. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). p. 18749-18758. IEEE, New Orleans, LA, USA (Jun 2022). https://doi.org/10.1109/CVPR52688.2022.01821, https://ieeexplore.ieee.org/document/9878591/" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.481 + ], + "angle": 0, + "content": "16. Fanelli, G., Gall, J., Romsdorfer, H., Weise, T., Van Gool, L.: A 3-d audio-visual corpus of affective communication. IEEE Transactions on Multimedia 12(6), 591-598 (2010). https://doi.org/10.1109/TMM.2010.2052239" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.522 + ], + "angle": 0, + "content": "17. Gong, S., Chen, L., Bronstein, M., Zafeiriou, S.: Spiralnet++: A fast and highly efficient mesh convolution operator. In: Proceedings of the IEEE International Conference on Computer Vision Workshops. pp. 0-0 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.785, + 0.564 + ], + "angle": 0, + "content": "18. Gupta, S., Castleman, K.R., Markey, M.K., Bovik, A.C.: Texas 3d face recognition database. In: 2010 IEEE Southwest Symposium on Image Analysis & Interpretation (SSIAI). pp. 97-100 (2010). https://doi.org/10.1109/SSIAI.2010.5483908" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.565, + 0.785, + 0.632 + ], + "angle": 0, + "content": "19. Haque, K.I., Yumak, Z.: Facexhubert: Text-less speech-driven e(x)pressive 3d facial animation synthesis using self-supervised speech representation learning. In: INTERNATIONAL CONFERENCE ON MULTIMODAL INTERACTION (ICMI '23). ACM, New York, NY, USA (2023). https://doi.org/10.1145/3577190.3614157, https://doi.org/10.1145/3577190.3614157" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.701 + ], + "angle": 0, + "content": "20. Hsu, W.N., Bolte, B., Tsai, Y.H.H., Lakhotia, K., Salakhutdinov, R., Mohamed, A.: Hubert: Self-supervised speech representation learning by masked prediction of hidden units. IEEE/ACM Trans. Audio, Speech and Lang. Proc. 29, 3451-3460 (oct 2021). https://doi.org/10.1109/TASLP.2021.3122291, https://doi.org/10.1109/TASLP.2021.3122291" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.785, + 0.758 + ], + "angle": 0, + "content": "21. Ji, X., Zhou, H., Wang, K., Wu, Q., Wu, W., Xu, F., Cao, X.: Eamm: One-shot emotional talking face via audio-based emotion-aware motion model. In: ACM SIGGRAPH 2022 Conference Proceedings. SIGGRAPH '22 (2022). https://doi.org/10.1145/3528233.3530745, https://doi.org/10.1145/3528233.3530745" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.785, + 0.799 + ], + "angle": 0, + "content": "22. Li, J., Kuang, Z., Zhao, Y., He, M., Bladin, K., Li, H.: Dynamic facial asset and rig generation from a single scan. ACM Trans. Graph. 39(6) (nov 2020). https://doi.org/10.1145/3414685.3417817, https://doi.org/10.1145/3414685.3417817" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.841 + ], + "angle": 0, + "content": "23. Li, T., Bolkart, T., Black, M.J., Li, H., Romero, J.: Learning a model of facial shape and expression from 4D scans. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia) 36(6), 194:1-194:17 (2017), https://doi.org/10.1145/3130800.3130813" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.374, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "ScanTalk: 3D Talking Heads from Unregistered Scans" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.203 + ], + "angle": 0, + "content": "24. Lim, I., Dielen, A., Campen, M., Kobbelt, L.: A simple approach to intrinsic correspondence learning on unstructured 3d meshes. p. 349-362. Springer-Verlag, Berlin, Heidelberg (2019). https://doi.org/10.1007/978-3-030-11015-4_26, https://doi.org/10.1007/978-3-030-11015-4_26" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.204, + 0.788, + 0.258 + ], + "angle": 0, + "content": "25. Liu, F., Tran, L., Liu, X.: 3d face modeling from diverse raw scan data. In: 2019 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 9407-9417. IEEE Computer Society, Los Alamitos, CA, USA (nov 2019). https://doi.org/10.1109/ICCV.2019.00950" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.259, + 0.788, + 0.342 + ], + "angle": 0, + "content": "26. Lugasri, C., Tang, J., Nash, H., McClanahan, C., Uboweja, E., Hays, M., Zhang, F., Chang, C.L., Yong, M., Lee, J., Chang, W.T., Hua, W., Georg, M., Grundmann, M.: Mediapipe: A framework for perceiving and processing reality. In: Third Workshop on Computer Vision for AR/VR at IEEE Computer Vision and Pattern Recognition (CVPR) 2019 (2019), https://mixedreality.cs.cornell.edu/s/NewTitle_May1_MediaPipe_CVPR_CV4ARVR_Workshop_2019.pdf" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.342, + 0.788, + 0.384 + ], + "angle": 0, + "content": "27. Massaro, D., Cohen, M., Tabain, M., Beskow, J., Clark, R.: Animated speech: Research progress and applications. Audiovisual Speech Processing (01 2001). https://doi.org/10.1017/cbo9780511843891.014" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.384, + 0.788, + 0.412 + ], + "angle": 0, + "content": "28. Muralikrishnan, S., Huang, C.H.P., Ceylan, D., Mitra, N.J.: Bliss: Bootstrapped linear shape space (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.412, + 0.788, + 0.467 + ], + "angle": 0, + "content": "29. Nocentini, F., Ferrari, C., Berretti, S.: Learning landmarks motion from speech for speaker-agnostic 3d talking heads generation. In: Foresti, G.L., Fusiello, A., Hancock, E. (eds.) Image Analysis and Processing - ICIAP 2023. pp. 340-351. Springer Nature Switzerland, Cham (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.467, + 0.788, + 0.522 + ], + "angle": 0, + "content": "30. Peng, Z., Luo, Y., Shi, Y., Xu, H., Zhu, X., Liu, H., He, J., Fan, Z.: Selftalk: A self-supervised commutative training diagram to comprehend 3d talking faces. In: Proceedings of the 31st ACM International Conference on Multimedia. p. 5292-5301 (2023). https://doi.org/10.1145/3581783.3611734" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.522, + 0.788, + 0.564 + ], + "angle": 0, + "content": "31. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems (NeurIPS) 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.564, + 0.788, + 0.605 + ], + "angle": 0, + "content": "32. Qin, D., Saito, J., Aigerman, N., Thibault, G., Komura, T.: Neural face rigging for animating and retargeting facial meshes in the wild. In: SIGGRAPH 2023 Conference Papers (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.605, + 0.788, + 0.647 + ], + "angle": 0, + "content": "33. Ranjan, A., Bolkart, T., Sanyal, S., Black, M.J.: Generating 3D faces using convolutional mesh autoencoders. In: European Conference on Computer Vision (ECCV). pp. 725-741 (2018), http://coma.is.tue.mpg.de/" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.647, + 0.788, + 0.702 + ], + "angle": 0, + "content": "34. Richard, A., Zollhöfer, M., Wen, Y., de la Torre, F., Sheikh, Y.: Meshtalk: 3d face animation from speech using cross-modality disentanglement. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 1173-1182 (October 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.702, + 0.788, + 0.758 + ], + "angle": 0, + "content": "35. Savran, A., Alyüz, N., Dibeklioglu, H., Çeliktutan, O., Gökberk, B., Sankur, B., Akarun, L.: Bosphorus database for 3d face analysis. In: Schouten, B., Juul, N.C., Drygajlo, A., Tistarelli, M. (eds.) Biometrics and Identity Management. pp. 47-56. Springer Berlin Heidelberg, Berlin, Heidelberg (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.758, + 0.788, + 0.813 + ], + "angle": 0, + "content": "36. Schneider, S., Baevski, A., Collobert, R., Auli, M.: wav2vec: Unsupervised pretraining for speech recognition. In: Interspeech 2019. p. 3465-3469. ISCA (Sep 2019). https://doi.org/10.21437/Interspeech.2019-1873, https://www.iscaspeech.org/archive/interspeech_2019/schneider19_interspeech.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.813, + 0.788, + 0.841 + ], + "angle": 0, + "content": "37. Sharp, N., Attaiki, S., Crane, K., Ovsjanikov, M.: Diffusionnet: Discretization ag- nostic learning on surfaces. ACM Trans. Graph. 01(1) (2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.395, + 0.128 + ], + "angle": 0, + "content": "F. Nocentini et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.787, + 0.217 + ], + "angle": 0, + "content": "38. Stan, S., Haque, K.I., Yumak, Z.: Facediffuser: Speech-driven 3d facial animation synthesis using diffusion. In: ACM SIGGRAPH Conference on Motion, Interaction and Games (MIG '23), November 15-17, 2023, Rennes, France. ACM, New York, NY, USA (2023). https://doi.org/10.1145/3623264.3624447, https://doi.org/10.1145/3623264.3624447" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.786, + 0.245 + ], + "angle": 0, + "content": "39. Thambiraja, B., Aliakbarian, S., Cosker, D., Thies, J.: 3diface: Diffusion-based speech-driven 3d facial animation and editing (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.246, + 0.787, + 0.3 + ], + "angle": 0, + "content": "40. Thambiraja, B., Habibie, I., Aliakbarian, S., Cosker, D., Theobalt, C., Thies, J.: Imitator: Personalized speech-driven 3d facial animation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 20621-20631 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.301, + 0.787, + 0.369 + ], + "angle": 0, + "content": "41. Wang, A., Emmi, M., Faloutsos, P.: Assembling an expressive facial animation system. In: Proceedings of the 2007 ACM SIGGRAPH Symposium on Video Games. p. 21-26. Sandbox '07, Association for Computing Machinery, New York, NY, USA (2007). https://doi.org/10.1145/1274940.1274947, https://doi.org/10.1145/1274940.1274947" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.37, + 0.787, + 0.424 + ], + "angle": 0, + "content": "42. Wang, J., Zhao, Y., Liu, L., Xu, T.S., Li, Q., Li, S.: Emotional talking head generation based on memory-sharing and attention-augmented networks. ArXiv abs/2306.03594 (2023), https://api_semanticscholar.org/CorpusID: 259089214" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.787, + 0.452 + ], + "angle": 0, + "content": "43. Wang, S., Li, L., Ding, Y., Yu, X.: One-shot talking face generation from single-speaker audio-visual correlation learning. In: AAAI 2022 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.453, + 0.787, + 0.549 + ], + "angle": 0, + "content": "44. Wuu, C.h., Zheng, N., Ardisson, S., Bali, R., Belko, D., Brockmeyer, E., Evans, L., Godisart, T., Ha, H., Huang, X., Hypes, A., Koska, T., Krenn, S., Lombardi, S., Luo, X., McPhail, K., Millerschoen, L., Perdoch, M., Pitts, M., Richard, A., Saragih, J., Saragih, J., Shiratori, T., Simon, T., Stewart, M., Trimble, A., Weng, X., Whitewolf, D., Wu, C., Yu, S.I., Sheikh, Y.: Multiface: A dataset for neural face rendering. In: arXiv (2022). https://doi.org/10.48550/ARXIV.2207.11243, https://arxiv.org/abs/2207.11243" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.55, + 0.787, + 0.605 + ], + "angle": 0, + "content": "45. Xing, J., Xia, M., Zhang, Y., Cun, X., Wang, J., Wong, T.T.: Codetalker: Speechdriven 3d facial animation with discrete motion prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12780-12790 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.606, + 0.787, + 0.66 + ], + "angle": 0, + "content": "46. Xu, Y., Feng, A.W., Marsella, S., Shapiro, A.: A practical and configurable lip sync method for games. In: Proceedings of Motion on Games. p. 131-140. MIG '13, Association for Computing Machinery, New York, NY, USA (2013). https://doi.org/10.1145/2522628.2522904, https://doi.org/10.1145/2522628.2522904" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.661, + 0.787, + 0.688 + ], + "angle": 0, + "content": "47. Yi, H., Liang, H., Liu, Y., Cao, Q., Wen, Y., Bolkart, T., Tao, D., Black, M.J.: Generating holistic 3d human motion from speech. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.689, + 0.787, + 0.742 + ], + "angle": 0, + "content": "48. Yin, L., Wei, X., Sun, Y., Wang, J., Rosato, M.: A 3d facial expression database for facial behavior research. In: 7th International Conference on Automatic Face and Gesture Recognition (FGR06). pp. 211-216 (2006). https://doi.org/10.1109/FGR.2006.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.743, + 0.787, + 0.799 + ], + "angle": 0, + "content": "49. Zhong, W., Fang, C., Cai, Y., Wei, P., Zhao, G., Lin, L., Li, G.: Identity-preserving talking face generation with landmark and appearance priors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9729-9738 (June 2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.799 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/ScanTalk_ 3D Talking Heads from Unregistered Scans/34e5d7c9-005f-4b11-aa72-5d7f3ba885f0_origin.pdf b/2024/ScanTalk_ 3D Talking Heads from Unregistered Scans/34e5d7c9-005f-4b11-aa72-5d7f3ba885f0_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..27673224e263ebc67227d13fa2fdef1722039822 --- /dev/null +++ b/2024/ScanTalk_ 3D Talking Heads from Unregistered Scans/34e5d7c9-005f-4b11-aa72-5d7f3ba885f0_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d97dbbca693b7d8738b8accc6f2a547ef94626cb5a9491341e1978a4e9b8fbdf +size 5724477 diff --git a/2024/ScanTalk_ 3D Talking Heads from Unregistered Scans/full.md b/2024/ScanTalk_ 3D Talking Heads from Unregistered Scans/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5a9993fcb91f6b1f8618692a379ad4c1ee180988 --- /dev/null +++ b/2024/ScanTalk_ 3D Talking Heads from Unregistered Scans/full.md @@ -0,0 +1,277 @@ +# ScanTalk: 3D Talking Heads from Unregistered Scans + +Federico Nocentini $^{1}$ , Thomas Besnier $^{2}$ , Claudio Ferrari $^{4}$ , Sylvain Arguillere $^{5}$ , Stefano Berretti $^{1}$ , and Mohamed Daoudi $^{2,3}$ + +$^{1}$ Media Integration and Communication Center (MICC), University of Florence, Italy federico.nocentini@unifi.it, stefano.berretti@unifi.it + $^{2}$ Univ. Lille, CNRS, Centrale Lille, UMR 9189 CRISTAL, F-59000 Lille, France thomas.besnier@univ-lille.fr + $^{3}$ IMT Nord Europe, Institut Mines-Télécom, Centre for Digital Systems mohamed.daoudi@imt-nord-europe.fr + $^{4}$ Department of Architecture and Engineering University of Parma, Italy claudio.ferrari2@unipr.it + $^{5}$ Univ. Lille, CNRS, UMR 8524 Laboratoire Paul Painlevé, Lille, F-59000, France sylvain,arguillere@univ-lille.fr + +Abstract. Speech-driven 3D talking heads generation has emerged as a significant area of interest among researchers, presenting numerous challenges. Existing methods are constrained by animating faces with fixed topologies, wherein point-wise correspondence is established, and the number and order of points remains consistent across all identities the model can animate. In this work, we present ScanTalk, a novel framework capable of animating 3D faces in arbitrary topologies including scanned data. Our approach relies on the DiffusionNet architecture to overcome the fixed topology constraint, offering promising avenues for more flexible and realistic 3D animations. By leveraging the power of DiffusionNet, ScanTalk not only adapts to diverse facial structures but also maintains fidelity when dealing with scanned data, thereby enhancing the authenticity and versatility of generated 3D talking heads. Through comprehensive comparisons with state-of-the-art methods, we validate the efficacy of our approach, demonstrating its capacity to generate realistic talking heads comparable to existing techniques. While our primary objective is to develop a generic method free from topological constraints, all state-of-the-art methodologies are bound by such limitations. Code for reproducing our results, and the pre-trained model are available at https://github.com/miccunifi/ScanTalk. + +Keywords: 3D Talking Heads $\cdot$ 3D Scans Animation $\cdot$ DiffusionNet + +![](images/eb9a4099eaec2b0d67bb6cc8deadfa0b594b21bf6626422e4bb946d43f8fe39f.jpg) +Fig. 1: We present ScanTalk, a deep learning architecture to animate any 3D face mesh driven by a speech. ScanTalk is robust enough to learn on multiple unrelated datasets with a unique model, whilst allowing us to infer on unregistered face meshes. + +# 1 Introduction + +Human face animation is a complex task, widely explored in Computer Vision and Computer Graphics because of the broad range of applications drawn from it, spanning from virtual reality to video game graphics, and more. As 3D face models continue to improve, it becomes more and more relevant to incorporate multi-modal aspects such as speech with audio data. Speech-driven facial animation encounters challenges inherent to the fact that finding a cross-modality mapping from audio to geometric data is an ill-posed problem, dealing with the complex geometry of human faces and the limited availability of paired audio-visual 3D data. Another, less discussed aspect that touches a broader range of 3D graphics with meshes is the robustness to changes in the topology of the mesh, which refers to the arrangement of the vertices and how they are connected. + +Maintaining fidelity and coherence across different topologies is crucial for ensuring that facial animations remain realistic and expressive, regardless of variations in the underlying mesh structure. This challenge becomes particularly pronounced in speech-driven facial animation, where the dynamics of speech-related lip movements, and related face changes necessitate a high degree of adaptability within the mesh topology. + +Addressing these challenges requires innovative approaches able to navigate the complexities of facial geometry, while accommodating the nuances of speech-driven animation. In this regard, we present ScanTalk, a novel framework capable of animating faces in arbitrary topologies including scanned data. ScanTalk overcomes limitations associated with fixed topologies, offering promising avenues for more flexible and realistic 3D talking heads generation. Indeed, the aforementioned constraint limits the range of applications of current deep learning models for speech-driven motion synthesis. Because of this, deep models are required to train on large scale registered datasets such as FLAME [23] or Multiface [44] for human face data. Building these datasets is a costly procedure, and the trained model is then only usable within the same registered setting. This typically means that any newly acquired data needs to be fitted to the corre + +sponding topology before any animation can be predicted by a deep model. This extra preprocessing step usually prevents online applications, therefore canceling one of the key benefits of deep models, which is their inference speed. + +Aiming to address these limitations, in this paper we present a flexible deep learning framework built to generate speech-driven animations of 3D face meshes. In particular, it gathers several key elements: + +1. A new robust approach to generate mesh deformation sequences based on DiffusionNet to compute intrinsic descriptors on 3D data; +2. A comprehensive architecture for learning speech-driven animations. Our approach works with meshes with different topologies, while showing competitive performance with respect to state-of-the-art models trained on an individual topology; +3. We show the generalizability of our model to unseen mesh topologies with qualitative examples. + +# 2 Related Work + +In literature, there are several methodologies to acquire 3D face meshes, either by extracting the geometry from images or videos [26,47] or by using a complex set of instruments in a controlled environment [16, 18, 33, 35, 44, 48]. The former offers advantages in terms of easier and more cost-effective data acquisition. However, these methods may sometimes fall short in capturing the complete 3D information from 2D data. With 3D scans, other challenges arise: they come unregistered and may present alterations such as holes and noise. Then, animating these objects directly becomes even more challenging. One can do it by registering the meshes onto a pre-defined topology [16, 22, 23, 28, 44] before animating the registrations with 3D morphable models [14] for example. + +Several deep learning models [3-5, 10, 25] proposed ways to learn a latent representation of face scans using robust encoders such as PointNet [6, 31] and Transformers but the resulting mesh is registered, which tends to smooth out some details from the scan geometry. However, this extra registration step may be handled efficiently with recent industrial applications such as MetaHuman from Epic Games. More recently, and closer to our goal, DiffusionNet [37] was combined with neural Jacobian fields [1] in Neural Face Rigging (NFR) [32] to learn a per-triangle deformation field on faces, allowing to transfer an animation from a sequence of unregistered face meshes to another. We stand out from this work by proposing a model that uses audio data to animate a given unregistered face mesh. While DiffusionNet has demonstrated its capability to generalize across varying triangulations in a static setting, our model is the first to exhibit similar properties in a multi-modal 4D setting. + +In recent years, numerous models and methodologies have emerged to address the challenge of synchronizing facial animations with speech audio. While significant progress has been made on 2D talking heads [2, 7, 12, 21, 42, 43, 49], only a handful of approaches can be seamlessly extended to 3D data. Procedural + +techniques have been proposed to animate 3D faces [9, 13, 27, 41, 46], primarily relying on visemes (groups of phonemes) to drive the movements of facial muscles. However, many of these models necessitate re-targeting and struggle to generalize effectively without extensive processing steps. More recent works leverage the increasing availability of data to develop statistical methods, including deep learning strategies [11, 15, 19, 29, 34, 38-40, 45]. Currently, state-of-the-art deep learning approaches are limited to a fixed mesh topology, requiring it to be identical to the one observed during the training phase. Among these models, VOCA [11] pioneers the development of deep models trained on a large-scale dataset of registered meshes. It was outperformed when MeshTalk [34] introduced a larger registered dataset named Multiface [44] along with a more expressive model. Moving forward, FaceFormer [15] utilized the transformer architecture, optimized for temporal data, and leveraged the power of a large-scale pre-trained audio encoder called Wav2vec2 [36]. This strategy saw further development with CodeTalker [45] and SelfTalk [30]. More recently, FaceXHubert [19] and FaceDiffuser [38] utilized an improved audio encoder, Hubert [20], demonstrating superiority to predict cross-modality mappings from audio to face motions. While continuously enhancing performances, these models are bound to a fixed mesh topology, hindering real-world applications and limiting the quantity of usable data for a single model, both for training and inference. + +In response to these challenges, we present a novel framework in the landscape of 3D facial animation learning, avoiding the limitations posed by the specific topology of the face mesh to be animated. The proposed model stands out for its capability to animate any face mesh, including real-world 3D scans. This novel approach holds the potential to redefine the standards in the field of 3D facial animation, being applicable across diverse datasets and scenarios. + +# 3 Proposed approach: ScanTalk + +In this section, we introduce Scantalk, a framework for animating 3D face meshes reproducing a spoken sentence contained in an audio file, which does not require the meshes to adhere to any specific topology. With ScanTalk, we push the state-of-the-art in 3D talking heads a step forward by allowing any 3D face, even raw scans, to be animated given a speech. To train ScanTalk, the only requirement is that the training meshes share a common topology within each sequence, but the topology may vary from one sequence to another. At inference time, any 3D face mesh in neutral state can be animated. + +ScanTalk is an Encoder-Decoder framework, which receives a neutral face mesh and an audio snippet as input, and outputs a sequence of per-vertex deformation fields, whose length depends on that of the audio. By summing each deformation field to the neutral input face, we ultimately obtain the animated sequence. The encoder is composed of two main modules: an audio encoder, which combines a pretrained encoder with a bi-directional LSTM that extracts audio features from the input speech, and a DiffusionNet encoder that computes surface descriptors from the neutral 3D face. These descriptors are then repli + +ated and concatenated to the audio features and the resulting signal is fed to a DiffusionNet decoder that outputs the deformation to be applied to the neutral face. The framework is depicted in Fig. 2. Before providing the technical details in Sec. 3.1 and Sec. 3.2, below we introduce a few general notations. + +Let $L = \{(M_i^{gt}, m_i^n, A_i)\}_{i=0}^{N-1}$ denotes the training set comprising $N$ samples, where $A_i$ is an audio containing a spoken sentence, $M_i^{gt} = (m_i^0, \ldots, m_i^{T_i-1}) \in \mathbb{R}^{T_i \times V_i \times 3}$ represents a sequence of 3D faces (same topology) of length $T_i$ synchronized with the spoken sentence in $A_i$ , and $m_i^n \in \mathbb{R}^{V_i \times 3}$ is a 3D neutral face. $V_i = |m_i^n|$ is the number of vertices in the $i$ -th 3D face sequence that needs to be consistent across each mesh in the $i$ -th sequence together with the vertex connectivity, overall determining the topology of the surface. Our objective is to establish a mapping function that correlates an audio input $A_i$ , and a neutral 3D face $m_i^n$ , to the ground-truth sequence $M_i^{gt}$ , expressed as: + +$$ +S c a n t a l k \left(A _ {i}, m _ {i} ^ {n}\right) \approx M _ {i} ^ {g t}. \tag {1} +$$ + +![](images/505f7a57701bde4b66873761340bbc034e7dda107df59dd43688f187efcf7cbb.jpg) +Fig. 2: Architecture of ScanTalk. A novel Encoder-Decoder framework designed to dynamically animate any 3D face based on a spoken sentence from an audio file. The Encoder integrates the 3D neutral face $m_i^n$ , per-vertex surface features $P_i^n$ (crucial for DiffusionNet and precomputed by the operators $OP$ ), and the audio file $A_i$ , yielding a fusion of per-vertex and audio features. These combined descriptors, alongside $P_i^n$ , are then passed to the Decoder, which mirrors a reversed DiffusionNet encoder structure. The Decoder predicts the deformation of the 3D neutral face, which is then combined with the original 3D neutral face $m_i^n$ to generate the animated sequence. + +# 3.1 ScanTalk Encoder + +ScanTalk requires an audio snippet $A_{i}$ and a 3D face in neutral state $m_{i}^{n}$ as inputs. To process the two inputs, ScanTalk comprises two distinct encoders to process the mesh and the audio, respectively, as detailed below. + +Face mesh encoder. Several approaches are available for encoding face meshes or point clouds, yet traditional graph convolution-based models, like [17, 24], encounter limitations with varying graph structures, such as changes in mesh resolution. To address this, we employ DiffusionNet [37], a discretization-agnostic encoder that demonstrated to be effective for encoding face meshes as seen in [32]. DiffusionNet integrates multi-layer perceptrons (MLPs), learned diffusion, and spatial gradient features, offering a straightforward yet robust architecture for surface learning tasks. It bypasses the need for complex operations like explicit surface convolutions or pooling hierarchies. + +Critical to DiffusionNet's functionality are precomputed features of the face surface, namely the Cotangent Laplacian, Eigenbasis, Mass Matrix, and Spatial Gradient Matrix. Together, these operators enhance the architecture's robustness, flexibility, and expressiveness, making DiffusionNet suitable for diverse surface learning tasks. Notably, this architecture accommodates 3D faces of any topology, allowing for variations in the number and order of defining points. + +Let $P_{i}^{n} = OP(m_{i}^{n})$ represent the precomputed features obtained by applying the above surface closed-form operators $OP$ to the 3D neutral face $m_{i}^{n}$ . The DiffusionNet Encoder $DN_{e}$ , with latent size $h$ , is designed to process a neutral 3D face mesh $m_{i}^{n}$ , and requires the precomputed per-vertex features $P_{i}^{n}$ to extract a per-vertex descriptors $f_{i}^{n}$ , that is: + +$$ +f _ {i} ^ {n} = D N _ {e} \left(m _ {i} ^ {n}, P _ {i} ^ {n}\right) \in \mathbb {R} ^ {V _ {i} \times h}. \tag {2} +$$ + +The per-vertex descriptors $f_{i}^{n}$ , is capable of capturing intricate details of each vertex within the neutral 3D face. + +Audio encoder. Following [19, 38], the speech encoder adopts the architecture of the state-of-the-art pretrained speech model, HuBERT [20] that is a self-supervised speech representation learner utilizing an offline clustering step to provide aligned target labels for a BERT-like prediction loss. Using this module, followed by a Linear Layer, we obtain a per-frame audio representation: + +$$ +a _ {i} = \operatorname {S p e e c h E n c o d e r} \left(A _ {i}\right) \in \mathbb {R} ^ {T _ {i} \times (h / 2)}. \tag {3} +$$ + +To ensure coherence between the speech representation, following the methodology outlined in [29], we concatenate the SpeechEncoder with a Multilayer Bidirectional-LSTM for temporal consistency, in a way such that the speech signal is projected into a temporal latent representation: + +$$ +v _ {i} = B i L S T M \left(a _ {i}\right) \in \mathbb {R} ^ {T _ {i} \times h}. \tag {4} +$$ + +# 3.2 ScanTalk Decoder + +We combine per-vertex descriptors $f_{i}^{n}$ extracted from the neutral face, with the latent vector $v_{i}$ extracted from the Bidirectional-LSTM + +$$ +(F _ {i} ^ {j}) _ {k} = \left(f _ {i} ^ {n}\right) _ {k} \oplus v _ {i} ^ {j} \in \mathbb {R} ^ {h * 2}, \quad \forall k = 0, \dots , V _ {i} - 1, \quad \forall j = 0, \dots , T _ {i} - 1. \tag {5} +$$ + +With this concatenation, we obtain a combined latent $F_{i}^{j} \in \mathbb{R}^{V_{i} \times h*2}$ that embeds both audio-related and geometry-related latents. To decode this sequence of combined latents for deforming the neutral face $m_{i}^{n}$ , we employ a DiffusionNet Decoder (which is essentially a reversed Encoder), denoted as $DN_{d}$ . The decoder module receives $F_{i}^{j}$ , and precomputed features $P_{i}^{n}$ derived from $m_{i}^{n}$ . It predicts the deformation of $m_{i}^{n}$ , denoted as $DN_{d}(F_{i}^{j}, P_{i}^{n})$ : + +$$ +\widehat {m _ {i} ^ {j}} = D N _ {d} \left(F _ {i} ^ {j}, P _ {i} ^ {n}\right) + m _ {i} ^ {n} \in \mathbb {R} ^ {V _ {i} \times 3}. \tag {6} +$$ + +Here $\widehat{m_i^j}$ represents the $j$ -th frame of the predicted sequence. The entire generated sequence is defined by $\widehat{M_i} \in \mathbb{R}^{T_i \times V_i \times 3}$ . We opt to utilize ScanTalk for predicting the deformation of the neutral face $m_i^n$ rather than predicting the actual face. This decision aligns with previous works [15, 19, 30, 38, 40], and offers advantages in terms of training efficiency and resulting animation. Predicting face deformation makes the learning process easier, by focusing solely on speech-related motion, as opposed to incorporating the problem of predicting the entire face reconstruction. Essentially, the decoder learns to predict a per-vertex displacement field from a time-dependent per-vertex descriptors field. + +# 3.3 ScanTalk Training + +ScanTalk generates deformations of a neutral face; hence, a predicted sequence maintains a consistent topology across all frames. This property arises from the definition of the DiffusionNet Decoder, which necessitates knowledge of precomputed features $P_{i}^{n}$ , and the number of points in the 3D neutral face targeted for animation. For this reason, during a supervised training protocol, the ground-truth meshes within each sequence must adhere to a common topology. Despite the apparent specificity of this requirement, it proves to be non-prohibitive in practice. ScanTalk animates a neutral face in response to an audio input, producing a sequence of 3D faces sharing the same topology of the neutral face that is animated, which can be any topology (even different from those seen during training). This alignment between the training strategy and the desired inference outcome underscores the efficiency of ScanTalk, which is not restricted to the topology of the training faces. + +Notably, the model predicts per-vertex displacements of the neutral ground-truth mesh and the ground truth sequence is registered in a supervised setting. Consequently, the predicted sequence and the ground-truth sequence are aligned on the same topology but can vary from one sequence to another. Thus, during training, we minimize the average vertex-wise Mean Squared Error (MSE) over + +a sequence of length $T_{i}$ between the ground truth $M_{i}^{gt}$ and the model prediction $\widehat{M}_{i}$ . which is defined as: + +$$ +\mathcal {L} _ {M S E} = \frac {1}{T _ {i} - 1} \sum_ {j = 0} ^ {T _ {i} - 1} \frac {1}{V _ {i} - 1} \sum_ {k = 0} ^ {V _ {i} - 1} \left\| \left(m _ {i} ^ {j}\right) _ {k} - \left(\widehat {m _ {i}} ^ {j}\right) _ {k} \right\| _ {2} ^ {2}. \tag {7} +$$ + +# 4 Experiments + +In the following, we first introduce the datasets and the metrics used for evaluation, respectively in Sec. 4.1 and Sec. 4.2. Then, we report quantitative and qualitative results in comparison with state-of-the-art methods in Sec. 4.3 and Sec. 4.4. In Sec. 4.5 we present several ablation studies over the framework architecture. Finally, in Sec. 4.6, we report the results of an user study designed to compare our approach with state-of-the-art methods. + +# 4.1 Datasets + +For training and quantitative evaluations, we rely on three classical datasets in 3D speech-driven motion synthesis: VOCAset [11], BIWI [16] and Multiface [44]. VOCAset gathers mesh sequences of 12 actors performing 40 speeches, captured at 60fps. Each sequence lasts for around 3 to 5 seconds, and each mesh is registered to the FLAME [23] topology with 5,023 vertices and 9,976 faces. + +BIWI comprises 14 subjects articulating 40 sentences each, sampled at 25fps. Each sentence lasts around 5 seconds, and the meshes are registered to a fixed topology. Due to GPU limitations, we used a downsampled version of this dataset, called $\mathrm{BIWI}_6$ , which has a fixed topology, with 3,895 vertices and 7,539 faces. + +Multiface includes 13 identities executing up to 50 speeches of around 4 seconds each, sampled at 30fps. The original multiface meshes have a fixed topology with 5,471 vertices and 10,837 faces. + +For training purposes, the meshes in both Multiface and BIWI $_6$ have been scaled and aligned with the meshes in the VOCAset dataset. The data splits can be found in the supplementary material. Since ScanTalk is the first topology-independent 3D talking head generator, for the sake of a comprehensive comparison with the state-of-the-art, we trained 4 different models: one is trained using all the three datasets together (multi-dataset), while the other three are trained on each dataset separately (single-dataset). + +# 4.2 Evaluation Metrics + +To evaluate the quality of the generated faces, we employ three standard metrics from previous works [30,34,38,45], namely: Lip vertex error (LVE) $\times 10^{-5}(mm)$ Assesses the lip deviation in a sequence by comparing it to the ground truth. It is obtained by computing the maximum L2 error for all lip vertices in each frame, averaged across all frames. + +Mean vertex error (MVE) $\times 10^{-3}(mm)$ . Describes the quality of the overall face reconstruction by computing the maximal vertex L2 error for each frame and taking the mean across all frames. + +Upper-face dynamic deviation (FDD) $\times 10^{-7}(mm)$ . Compares the standard deviation of upper vertices between the generated sequence and ground-truth. + +# 4.3 Quantitative Results + +Although ScanTalk is the first deep learning model to animate unregistered 3D face meshes, comparisons with the state-of-the-art is possible when inferring our model on registered data. In Tab. 1, we present the performance evaluation of ScanTalk training on both a single-dataset and multi-datasets, juxtaposed against several state-of-the-art methods trained solely on a single dataset. While the augmentation of training data enhances the effectiveness of our model, addressing disparities in data composition becomes important for animating unseen faces. Particularly, the varied geometries and head dynamics inherent in $\mathrm{BIWI}_6$ , VOCA, and Multiface meshes pose significant challenges, potentially limiting performance compared to a model trained exclusively on a single dataset. + +Table 1: ScanTalk performance, in both single-dataset (s-d) and multi-dataset (m-d) scenarios, in comparison with state-of-the-art methods. The heatmaps show the differences between the first frame and subsequent frames within sequences among the VOCAset, $\mathrm{BIWI}_6$ , and Multiface datasets. Notably, in the VOCAset, primarily the lips display movement, whereas in $\mathrm{BIWI}_6$ and Multiface datasets, substantial head and upper face movements are observed. The color gradient on the face meshes corresponds to the average per-vertex $L_2$ norm of the differences, where blue hues indicate lower values, and red hues indicate higher values. + +![](images/1fe722d18ef68bdc0c0bc93ebd20e94ee33e2de08f9b34fd2762e984a10ffe48.jpg) + +![](images/8f329ff74c2ca8096ecc35d4960f69d09f901e004d9e634cc4b80402fccd7d35.jpg) + +![](images/571666e5c93356be6eb0d325bb1a06ac4fb02a042cdb3233d2c384868afa24f2.jpg) + +
VOCAssetBIWI6Multiface
LVE ↓MVE ↓FDD ↓LVE ↓MVE ↓FDD ↓LVE ↓MVE ↓FDD ↓
VOCA6.9930.9832.6625.7432.59141.4824.9232.76155.781
FaceFormer6.1230.9352.1634.0852.16337.0912.4511.45320.239
SelfTalk5.6180.9182.3213.6282.06235.4702.2811.90137.434
CodeTalker3.5490.8882.2585.1902.64120.5994.0912.38247.905
FaceDiffuser4.3500.9012.4374.0222.12839.6043.5552.38829.157
ScanTalk m-d6.3750.9872.1014.0442.05740.0512.4351.67832.202
ScanTalk s-d3.0120.8612.4004.6512.14836.0342.6531.87164.451
+ +This observation becomes evident when assessing Scantalk's performance on VOCAset, where a substantial improvement is observed when training on a single-dataset. Conversely, in the case of $\mathrm{BIWI}_6$ and Multiface single-dataset training, a decline in results is noted compared to the multi-dataset version. This disparity is attributed to the presence of speech-related head and upper face movements within the ground truth talking head sequences of $\mathrm{BIWI}_6$ and Multiface. Consequently, the multi-dataset training model demonstrates superior performance as it encounters a wider array of sequences featuring head and upper face movements. In contrast, VOCAset lacks such movements, with the 3D faces remaining static, while only the mouth moves, thereby enhancing performance for ScantTalk trained solely on VOCAset. This obviates the need to learn other movements of the head or upper face. + +The distinctions between VOCAset and the other two datasets are further underscored by the FDD metric, indicating that 3D faces in VOCAset remain stationary, with minimal activity in the upper facial area and head. Across all tested models, generating sequences with lower FDD is notably more achievable on VOCAset compared to the other datasets, thus emphasizing the disparities between VOCAset and the others. + +Nevertheless, the results presented in Tab. 1 demonstrate that ScanTalk produces outcomes comparable to state-of-the-art methods, whether undergoing single-dataset or multi-dataset training. Moreover, ScanTalk emerges as the first model capable of training in multi-dataset settings, demonstrating the ability to animate a 3D face regardless of its topology. An interesting characteristic of ScanTalk emerges when analyzing the GPU memory usage relative to the vertex count in the 3D facial model requiring animation. Fig. 3 illustrates a linear increase in GPU usage correlated with the number of vertices. + +![](images/bd80a0f7c0b1ac3c572bd671c1a0bcf29a8349f9de5fdb3352912e1e4b6ea7f2.jpg) +Fig.3: ScanTalk GPU memory usage with respect to the mesh resolution. + +# 4.4 Qualitative Results + +In the domain of 3D speech-driven talking heads generation, qualitative evaluations hold as much significance as quantitative assessments. To evaluate the quantitative findings presented in Sec. 4.3, we introduce a straightforward test aimed at discerning disparities in head and upper face movements across the three datasets: for each dataset, we compute the average $L_{2}$ norm of the differences between the initial frame and subsequent frames within each sequence. This enables the quantification of both head dynamics and upper facial movements across mesh sequences. In the heatmaps of Tab. 1, we present the per-frame average difference for each dataset. Our analysis reveals that in the VOCAset, only the lower part of the face exhibits discernible movements, whereas in BIWI $_6$ and Multiface, the entire head or face moves. + +![](images/297b328e04cfc61facd6bd7b14328f3fcd4731285f89866ed76d9130a7879461.jpg) +Fig. 4: Relative norm of the per-vertex descriptors $f_{i}^{n}$ in Eq. (2) extracted by $DN_{e}$ displayed as a heatmap on a mesh from VOCAset (left), and a mesh from Multiface (right). For each mesh, we show the norm on the original topology, on a remeshed version, and on a further degraded mesh obtained by removing the back of the head and creating random holes. Here, pinker hues indicate lower values, and greener hues indicate higher values. + +![](images/1342fd14a0bf9b6e8e99754d04b9a0a63353d62dd6c235986ad1646955114338.jpg) +Fig. 5: ScanTalk inference on 3D faces. The meshes have been rigidly aligned with the training data. The first row is an animation of a raw 3D scan (an hole for the mouth has been created), while the others two are animation of meshes in an arbitrary topology. + +In Fig. 4, we present a heatmap visualization representing the norm of geometric descriptors $f_{i}^{n}$ computed by the encoder $DN_{e}$ in Eq. (2) for various alterations of the same identity mesh. Our findings indicate that the ScanTalk encoder effectively extracts descriptors from input faces regardless of facial topology. Notably, despite differences in facial topology among the same identity, the extracted descriptors exhibit remarkable similarity. + +In Fig. 5, 3D faces in different topology, including a scan, animated using ScanTalk are presented. Our method demonstrates good generalization capabilities, successfully animating diverse 3D faces. Notably, enhanced animation quality is observed when a mouth aperture is present; however, our model performs well even in its absence. Nonetheless, without a mouth aperture, the model struggles to generate the corresponding mouth opening, despite accurate lip movement synchronization. + +# 4.5 Ablations Studies + +To investigate the impacts of various components within our architecture, we conducted extensive experiments across various configurations, encompassing both single-dataset and multi-dataset training paradigms. The results reported in Tab. 2 and Tab. 3 are computed by modifying ScanTalk's modules and training losses as described in Sec. 3. + +Audio encoder. In learning based speech-driven animation, some previous work used either Wav2vec2 [36] or Hubert [20], the latter being substantially better for this task according to [19,38]. For our purpose, we also tested WavLM [8] but obtained poorer efficiency. The results on both the single-dataset and multi-dataset training are displayed in Tab. 2a and Tab. 3 top. We can see that, in both cases, the usage of the Hubert Encoder [19] for audio features extraction leads to better results. + +Temporal consistency. In our investigation, as proven by [19,29,40], we found that supplementing pre-trained audio encoders with additional temporal consistency mechanisms such as Bidirectional-LSTM, Bidirectional-GRU, or an autoregressive Transformer Decoder (TD) significantly enhances model performance, as illustrated in Tab. 2b and in Tab. 3 middle. The BiLSTM model architecture is detailed in Sec. 3, while the BiGRU model mirrors the BiLSTM architecture but substitutes BiLSTM with BiGRU. Details about the Transformer Decoder can be found in the supplementary material. From the findings presented in Tab. 2b and in Tab. 3 middle, we observe that employing a multilayer Bidirectional-LSTM in the audio stream processing of ScanTalk yields the most favorable performance across both single-dataset training and multi-dataset training scenarios. + +Loss function. Previous research has extensively explored optimal objective functions for enhancing and refining the learning process. Inspired by [29,30,40], + +we experimented with Mean Square Error ( $L_{MSE}$ ), Masked ( $L_{mask}$ ), and Velocity ( $L_{vel}$ ) Loss. The former is a simple $L_2$ loss, the second employs Mean Square Error focused on lip vertices, while the latter aims to minimize differences between consecutive frames. While these loss functions have demonstrated efficacy in single-dataset training, as evidenced by [29,30,40], our findings, detailed in Tab. 3 bottom, indicate that employing a straightforward $L_2$ , on multi-dataset training, enhances the model's capacity to generate realistic talking heads. We attribute this to significant geometric variations observed across different datasets. + +Table 2: ScanTalk (ST) single-dataset ablation studies. Results obtained with a model trained just on VOCAsset. +(a) Audio Encoder Ablation. + +
LVE ↓MVE ↓FDD ↓
ST w WavLM3.6740.9372.413
ST w Wav2Vec23.3090.8602.244
ScanTalk3.0120.8612.400
+ +(b) Temporal consistency ablation. + +
LVE ↓MVE ↓FDD ↓
ST w/o3.3610.8702.365
ST w TD3.2910.8592.406
ST w BiGRU3.0360.8352.358
ScanTalk3.0120.8612.400
+ +Table 3: ScanTalk (ST) multi-dataset ablation studies. Top: ST with different audio encoders. Middle: ST using different audio stream processing. Bottom: ST with different Loss function. The last row is the proposed ScanTalk + +
VOCAsetBIWI6Multiface
LVE ↓MVE ↓FDD ↓LVE ↓MVE ↓FDD ↓LVE ↓MVE ↓FDD ↓
ST w Wav2Vec27.1271.2412.3494.4392.27641.3743.1982.32933.671
ST w WavLM6.7731.0042.2204.3332.16231.9222.9072.0048.500
ST w TD6.5120.9951.8584.8272.16836.7892.1361.72535.044
ST w BiGRU6.5840.9942.0544.4212.10341.4562.5341.98734.098
ST+Lmask+Lvel7.0311.0921.4734.7552.22739.4582.5471.89815.973
ST+Lmask7.4511.0840.8594.7482.19936.9392.8881.89132.501
ST+Lvel6.7400.9981.8994.5092.18033.6142.1031.77531.599
ScanTalk6.3750.9872.1014.0442.05740.0512.4351.67832.202
+ +# 4.6 User study + +To further evaluate our solution, we performed a study where human feedback is involved. This evaluation is conducted with two user-based studies involving 25 participants; $(i)$ For the first study, in alignment with prior research [15,19,30,34, 38,45], we designed an A/B test to compare ScanTalk with other state-of-the-art + +models within a registered setting on both lip-syncing and naturalness criteria (Test 1); (ii) In the second test, we assessed the credibility of scan animations by asking participants to rate the animation quality of ten scans sourced from the COMA dataset [33], using a scale ranging from 1 to 10 (Test 2). + +The outcomes of Test 1 are depicted in the table on the left of Fig. 6. Notably, sequences produced using ScanTalk demonstrate levels of both naturalness and lip-syncing that are comparable to state-of-the-art methods. Specifically, our approach is preferred when compared with FaceFormer, CodeTalker, and FaceDiffuser on VOCAsset. However, SelfTalk and the ground truth are preferred over ScanTalk. Nevertheless, the percentage of users favoring ScanTalk remains nonnegligible, underscoring the efficacy of our approach in generating 3D talking heads with good realism and lip-syncing fidelity. + +Results of Test 2 are reported on the right of Fig. 6. These ratings are closely linked to both the scan quality and the fidelity of the mouth representation. Despite the scan quality not being optimal, as depicted in Fig. 6, our results indicate that we can achieve a good level of realism in the animated scans. + +
ComparisonNaturalness (%)Lip-sync (%)
ST vs. FaceFormer82.6780.67
ST vs. CodeTalker57.3356.00
ST vs. FaceDiffuser90.6794.00
ST vs. SelfTalk44.0043.33
ST vs. GT44.0042.00
+ +![](images/3d30e1e766fee7c38debca7d1ffd148c36d3676c9d9770c00fbb950612675710.jpg) +Fig. 6: The table on the left shows results of Test 1, ScanTalk (ST) vs. other methods and the ground truth (GT) on samples from the test set of VOCAset. The values denote the percentages of users who favored ScanTalk animations over others. The chart on the right shows results of Test 2 reported as a violin plot for the animation of scans from COMA [33]. The subject median rating is displayed as a black dot on each violin. + +# 5 Conclusions + +This paper introduces ScanTalk, a novel framework for speech-driven 3D facial animation. Unlike existing methods, ScanTalk possesses the unique capability to animate any 3D face, regardless of its topology, even if it differs from the ones on which it was trained. ScanTalk extends the applicability of deep speech-driven 3D facial animations by addressing the challenges of topology robustness. Additionally, our model demonstrates comparable quantitative and qualitative results with other state-of-the-art methods across three distinct datasets. + +Acknowledgments This work is supported by the ANR project Human4D ANR-19-CE23-0020 and by the IRP CNRS project GeoGen3DHuman. This work was also partially supported by "Partenariato FAIR (Future Artificial Intelligence Research) - PE00000013, CUP J33C22002830006" funded by NextGenerationEU through the italian MUR within NRRP, project DL-MIG. This work was also partially funded by the ministerial decree n.352 of the 9th April 2022 NextGenerationEU through the italian MUR within NRRP. This work was also partially supported by Fédération de Recherche Mathématique des Hauts-de-France (FMHF, FR2037 du CNRS). + +# References + +1. Aigerman, N., Gupta, K., Kim, V.G., Chaudhuri, S., Saito, J., Groueix, T.: Neural jacobian fields: Learning intrinsic mappings of arbitrary meshes. ACM Trans. Graph. 41(4) (jul 2022). https://doi.org/10.1145/3528223.3530141, https://doi.org/10.1145/3528223.3530141 +2. Alghamdi, M.M., Wang, H., Bulpitt, A.J., Hogg, D.C.: Talking head from speech audio using a pre-trained image generator. In: Proceedings of the 30th ACM International Conference on Multimedia (2022) +3. Bahri, M., O' Sullivan, E., Gong, S., Liu, F., Liu, X., Bronstein, M.M., Zafeiriou, S.: Shape my face: Registering 3d face scans by surface-to-surface translation. International Journal of Computer Vision (IJCV) (Sep 2021) +4. Besnier, T., Arguillère, S., Pierson, E., Daoudi, M.: Toward Mesh-Invariant 3D Generative Deep Learning with Geometric Measures. Computers and Graphics (2023). https://doi.org/10.1016/j.cag.2023.06.027, https://hal.science/hal-04143649 +5. Chandran, P., Zoss, G., Gross, M., Gotardo, P., Bradley, D.: Shape transformers: Topology-independent 3d shape models using transformers. In: Computer Graphics Forum. vol. 41, pp. 195-207. Wiley Online Library (2022) +6. Charles, R.Q., Su, H., Kaichun, M., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). p. 77-85. IEEE (Jul 2017) +7. Chen, L., Cui, G., Liu, C., Li, Z., Kou, Z., Xu, Y., Xu, C.: Talking-head generation with rhythmic head motion. In: European Conference on Computer Vision (2020), https://api-semanticscholar.org/CorpusID:220633152 +8. Chen, S., Wang, C., Chen, Z., Wu, Y., Liu, S., Chen, Z., Li, J., Kanda, N., Yoshioka, T., Xiao, X., Wu, J., Zhou, L., Ren, S., Qian, Y., Qian, Y., Zeng, M., Wei, F.: Wavlm: Large-scale self-supervised pre-training for full stack speech processing. IEEE Journal of Selected Topics in Signal Processing 16, 1505-1518 (2021), https://api-semanticscholar.org/CorpusID:239885872 +9. Cosi, P., Caldognetto, E., Perin, G., Zmarich, C.: Labial coarticulation modeling for realistic facial animation. In: Proceedings. Fourth IEEE International Conference on Multimodal Interfaces. pp. 505-510 (2002). https://doi.org/10.1109/ICMI.2002.1167047 +10. Croquet, B., Christiaens, D., Weinberg, S.M., Bronstein, M., Vandermeulen, D., Claes, P.: Unsupervised diffeomorphic surface registration and non-linear modelling. In: Medical Image Computing and Computer Assisted Intervention (MIC-CAI). p. 118–128. Springer (2021) + +11. Cudeiro, D., Bolkart, T., Laidlaw, C., Ranjan, A., Black, M.: Capture, learning, and synthesis of 3D speaking styles. In: Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR). pp. 10101-10111 (2019) +12. Das, D., Biswas, S., Sinha, S., Bhowmick, B.: Speech-driven facial animation using cascaded gans for learning of motion and texture. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J.M. (eds.) Computer Vision - ECCV 2020. pp. 408-424. Springer International Publishing, Cham (2020) +13. Edwards, P., Landreth, C., Fiume, E., Singh, K.: Jali: an animator-centric viseme model for expressive lip synchronization. ACM Trans. Graph. 35(4) (jul 2016). https://doi.org/10.1145/2897824.2925984, https://doi.org/10.1145/2897824.2925984 +14. Egger, B., Smith, W.A.P., Tewari, A., Wuhrer, S., Zollhoefer, M., Beeler, T., Bernard, F., Bolkart, T., Kortylewski, A., Romdhani, S., Theobalt, C., Blanz, V., Vetter, T.: 3D Morphable Face Models - Past, Present and Future. ACM Transactions on Graphics 39(5), 157:1-38 (Aug 2020). https://doi.org/10.1145/3395208, https://inria.hal.science/hal-02280281 +15. Fan, Y., Lin, Z., Saito, J., Wang, W., Komura, T.: Faceformer: Speech-driven 3d facial animation with transformers. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). p. 18749-18758. IEEE, New Orleans, LA, USA (Jun 2022). https://doi.org/10.1109/CVPR52688.2022.01821, https://ieeexplore.ieee.org/document/9878591/ +16. Fanelli, G., Gall, J., Romsdorfer, H., Weise, T., Van Gool, L.: A 3-d audio-visual corpus of affective communication. IEEE Transactions on Multimedia 12(6), 591-598 (2010). https://doi.org/10.1109/TMM.2010.2052239 +17. Gong, S., Chen, L., Bronstein, M., Zafeiriou, S.: Spiralnet++: A fast and highly efficient mesh convolution operator. In: Proceedings of the IEEE International Conference on Computer Vision Workshops. pp. 0-0 (2019) +18. Gupta, S., Castleman, K.R., Markey, M.K., Bovik, A.C.: Texas 3d face recognition database. In: 2010 IEEE Southwest Symposium on Image Analysis & Interpretation (SSIAI). pp. 97-100 (2010). https://doi.org/10.1109/SSIAI.2010.5483908 +19. Haque, K.I., Yumak, Z.: Facexhubert: Text-less speech-driven e(x)pressive 3d facial animation synthesis using self-supervised speech representation learning. In: INTERNATIONAL CONFERENCE ON MULTIMODAL INTERACTION (ICMI '23). ACM, New York, NY, USA (2023). https://doi.org/10.1145/3577190.3614157, https://doi.org/10.1145/3577190.3614157 +20. Hsu, W.N., Bolte, B., Tsai, Y.H.H., Lakhotia, K., Salakhutdinov, R., Mohamed, A.: Hubert: Self-supervised speech representation learning by masked prediction of hidden units. IEEE/ACM Trans. Audio, Speech and Lang. Proc. 29, 3451-3460 (oct 2021). https://doi.org/10.1109/TASLP.2021.3122291, https://doi.org/10.1109/TASLP.2021.3122291 +21. Ji, X., Zhou, H., Wang, K., Wu, Q., Wu, W., Xu, F., Cao, X.: Eamm: One-shot emotional talking face via audio-based emotion-aware motion model. In: ACM SIGGRAPH 2022 Conference Proceedings. SIGGRAPH '22 (2022). https://doi.org/10.1145/3528233.3530745, https://doi.org/10.1145/3528233.3530745 +22. Li, J., Kuang, Z., Zhao, Y., He, M., Bladin, K., Li, H.: Dynamic facial asset and rig generation from a single scan. ACM Trans. Graph. 39(6) (nov 2020). https://doi.org/10.1145/3414685.3417817, https://doi.org/10.1145/3414685.3417817 +23. Li, T., Bolkart, T., Black, M.J., Li, H., Romero, J.: Learning a model of facial shape and expression from 4D scans. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia) 36(6), 194:1-194:17 (2017), https://doi.org/10.1145/3130800.3130813 + +24. Lim, I., Dielen, A., Campen, M., Kobbelt, L.: A simple approach to intrinsic correspondence learning on unstructured 3d meshes. p. 349-362. Springer-Verlag, Berlin, Heidelberg (2019). https://doi.org/10.1007/978-3-030-11015-4_26, https://doi.org/10.1007/978-3-030-11015-4_26 +25. Liu, F., Tran, L., Liu, X.: 3d face modeling from diverse raw scan data. In: 2019 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 9407-9417. IEEE Computer Society, Los Alamitos, CA, USA (nov 2019). https://doi.org/10.1109/ICCV.2019.00950 +26. Lugasri, C., Tang, J., Nash, H., McClanahan, C., Uboweja, E., Hays, M., Zhang, F., Chang, C.L., Yong, M., Lee, J., Chang, W.T., Hua, W., Georg, M., Grundmann, M.: Mediapipe: A framework for perceiving and processing reality. In: Third Workshop on Computer Vision for AR/VR at IEEE Computer Vision and Pattern Recognition (CVPR) 2019 (2019), https://mixedreality.cs.cornell.edu/s/NewTitle_May1_MediaPipe_CVPR_CV4ARVR_Workshop_2019.pdf +27. Massaro, D., Cohen, M., Tabain, M., Beskow, J., Clark, R.: Animated speech: Research progress and applications. Audiovisual Speech Processing (01 2001). https://doi.org/10.1017/cbo9780511843891.014 +28. Muralikrishnan, S., Huang, C.H.P., Ceylan, D., Mitra, N.J.: Bliss: Bootstrapped linear shape space (2023) +29. Nocentini, F., Ferrari, C., Berretti, S.: Learning landmarks motion from speech for speaker-agnostic 3d talking heads generation. In: Foresti, G.L., Fusiello, A., Hancock, E. (eds.) Image Analysis and Processing - ICIAP 2023. pp. 340-351. Springer Nature Switzerland, Cham (2023) +30. Peng, Z., Luo, Y., Shi, Y., Xu, H., Zhu, X., Liu, H., He, J., Fan, Z.: Selftalk: A self-supervised commutative training diagram to comprehend 3d talking faces. In: Proceedings of the 31st ACM International Conference on Multimedia. p. 5292-5301 (2023). https://doi.org/10.1145/3581783.3611734 +31. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems (NeurIPS) 30 (2017) +32. Qin, D., Saito, J., Aigerman, N., Thibault, G., Komura, T.: Neural face rigging for animating and retargeting facial meshes in the wild. In: SIGGRAPH 2023 Conference Papers (2023) +33. Ranjan, A., Bolkart, T., Sanyal, S., Black, M.J.: Generating 3D faces using convolutional mesh autoencoders. In: European Conference on Computer Vision (ECCV). pp. 725-741 (2018), http://coma.is.tue.mpg.de/ +34. Richard, A., Zollhöfer, M., Wen, Y., de la Torre, F., Sheikh, Y.: Meshtalk: 3d face animation from speech using cross-modality disentanglement. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 1173-1182 (October 2021) +35. Savran, A., Alyüz, N., Dibeklioglu, H., Çeliktutan, O., Gökberk, B., Sankur, B., Akarun, L.: Bosphorus database for 3d face analysis. In: Schouten, B., Juul, N.C., Drygajlo, A., Tistarelli, M. (eds.) Biometrics and Identity Management. pp. 47-56. Springer Berlin Heidelberg, Berlin, Heidelberg (2008) +36. Schneider, S., Baevski, A., Collobert, R., Auli, M.: wav2vec: Unsupervised pretraining for speech recognition. In: Interspeech 2019. p. 3465-3469. ISCA (Sep 2019). https://doi.org/10.21437/Interspeech.2019-1873, https://www.iscaspeech.org/archive/interspeech_2019/schneider19_interspeech.html +37. Sharp, N., Attaiki, S., Crane, K., Ovsjanikov, M.: Diffusionnet: Discretization ag- nostic learning on surfaces. ACM Trans. Graph. 01(1) (2022) + +38. Stan, S., Haque, K.I., Yumak, Z.: Facediffuser: Speech-driven 3d facial animation synthesis using diffusion. In: ACM SIGGRAPH Conference on Motion, Interaction and Games (MIG '23), November 15-17, 2023, Rennes, France. ACM, New York, NY, USA (2023). https://doi.org/10.1145/3623264.3624447, https://doi.org/10.1145/3623264.3624447 +39. Thambiraja, B., Aliakbarian, S., Cosker, D., Thies, J.: 3diface: Diffusion-based speech-driven 3d facial animation and editing (2023) +40. Thambiraja, B., Habibie, I., Aliakbarian, S., Cosker, D., Theobalt, C., Thies, J.: Imitator: Personalized speech-driven 3d facial animation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 20621-20631 (October 2023) +41. Wang, A., Emmi, M., Faloutsos, P.: Assembling an expressive facial animation system. In: Proceedings of the 2007 ACM SIGGRAPH Symposium on Video Games. p. 21-26. Sandbox '07, Association for Computing Machinery, New York, NY, USA (2007). https://doi.org/10.1145/1274940.1274947, https://doi.org/10.1145/1274940.1274947 +42. Wang, J., Zhao, Y., Liu, L., Xu, T.S., Li, Q., Li, S.: Emotional talking head generation based on memory-sharing and attention-augmented networks. ArXiv abs/2306.03594 (2023), https://api_semanticscholar.org/CorpusID: 259089214 +43. Wang, S., Li, L., Ding, Y., Yu, X.: One-shot talking face generation from single-speaker audio-visual correlation learning. In: AAAI 2022 (2022) +44. Wuu, C.h., Zheng, N., Ardisson, S., Bali, R., Belko, D., Brockmeyer, E., Evans, L., Godisart, T., Ha, H., Huang, X., Hypes, A., Koska, T., Krenn, S., Lombardi, S., Luo, X., McPhail, K., Millerschoen, L., Perdoch, M., Pitts, M., Richard, A., Saragih, J., Saragih, J., Shiratori, T., Simon, T., Stewart, M., Trimble, A., Weng, X., Whitewolf, D., Wu, C., Yu, S.I., Sheikh, Y.: Multiface: A dataset for neural face rendering. In: arXiv (2022). https://doi.org/10.48550/ARXIV.2207.11243, https://arxiv.org/abs/2207.11243 +45. Xing, J., Xia, M., Zhang, Y., Cun, X., Wang, J., Wong, T.T.: Codetalker: Speechdriven 3d facial animation with discrete motion prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12780-12790 (2023) +46. Xu, Y., Feng, A.W., Marsella, S., Shapiro, A.: A practical and configurable lip sync method for games. In: Proceedings of Motion on Games. p. 131-140. MIG '13, Association for Computing Machinery, New York, NY, USA (2013). https://doi.org/10.1145/2522628.2522904, https://doi.org/10.1145/2522628.2522904 +47. Yi, H., Liang, H., Liu, Y., Cao, Q., Wen, Y., Bolkart, T., Tao, D., Black, M.J.: Generating holistic 3d human motion from speech. In: CVPR (2023) +48. Yin, L., Wei, X., Sun, Y., Wang, J., Rosato, M.: A 3d facial expression database for facial behavior research. In: 7th International Conference on Automatic Face and Gesture Recognition (FGR06). pp. 211-216 (2006). https://doi.org/10.1109/FGR.2006.6 +49. Zhong, W., Fang, C., Cai, Y., Wei, P., Zhao, G., Lin, L., Li, G.: Identity-preserving talking face generation with landmark and appearance priors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9729-9738 (June 2023) \ No newline at end of file diff --git a/2024/ScanTalk_ 3D Talking Heads from Unregistered Scans/images.zip b/2024/ScanTalk_ 3D Talking Heads from Unregistered Scans/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..05f06cbdd3194a4c40f427b828647fcd268d6618 --- /dev/null +++ b/2024/ScanTalk_ 3D Talking Heads from Unregistered Scans/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d03473b5c22fada11261cd70ccc43ccd47413839bdc332ae39ab32123c7fb6cb +size 460081 diff --git a/2024/ScanTalk_ 3D Talking Heads from Unregistered Scans/layout.json b/2024/ScanTalk_ 3D Talking Heads from Unregistered Scans/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7e6bf268d7ae981dfccfb8f15a3ea5b4c5629787 --- /dev/null +++ b/2024/ScanTalk_ 3D Talking Heads from Unregistered Scans/layout.json @@ -0,0 +1,8450 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 207, + 111, + 408, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 111, + 408, + 148 + ], + "spans": [ + { + "bbox": [ + 207, + 111, + 408, + 148 + ], + "type": "text", + "content": "ScanTalk: 3D Talking Heads from Unregistered Scans" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 154, + 170, + 459, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 170, + 459, + 195 + ], + "spans": [ + { + "bbox": [ + 154, + 170, + 459, + 195 + ], + "type": "text", + "content": "Federico Nocentini" + }, + { + "bbox": [ + 154, + 170, + 459, + 195 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 154, + 170, + 459, + 195 + ], + "type": "text", + "content": ", Thomas Besnier" + }, + { + "bbox": [ + 154, + 170, + 459, + 195 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 154, + 170, + 459, + 195 + ], + "type": "text", + "content": ", Claudio Ferrari" + }, + { + "bbox": [ + 154, + 170, + 459, + 195 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 154, + 170, + 459, + 195 + ], + "type": "text", + "content": ", Sylvain Arguillere" + }, + { + "bbox": [ + 154, + 170, + 459, + 195 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 154, + 170, + 459, + 195 + ], + "type": "text", + "content": ", Stefano Berretti" + }, + { + "bbox": [ + 154, + 170, + 459, + 195 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 154, + 170, + 459, + 195 + ], + "type": "text", + "content": ", and Mohamed Daoudi" + }, + { + "bbox": [ + 154, + 170, + 459, + 195 + ], + "type": "inline_equation", + "content": "^{2,3}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 141, + 202, + 473, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 202, + 473, + 326 + ], + "spans": [ + { + "bbox": [ + 141, + 202, + 473, + 326 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 141, + 202, + 473, + 326 + ], + "type": "text", + "content": " Media Integration and Communication Center (MICC), University of Florence, Italy federico.nocentini@unifi.it, stefano.berretti@unifi.it \n" + }, + { + "bbox": [ + 141, + 202, + 473, + 326 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 141, + 202, + 473, + 326 + ], + "type": "text", + "content": " Univ. Lille, CNRS, Centrale Lille, UMR 9189 CRISTAL, F-59000 Lille, France thomas.besnier@univ-lille.fr \n" + }, + { + "bbox": [ + 141, + 202, + 473, + 326 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 141, + 202, + 473, + 326 + ], + "type": "text", + "content": " IMT Nord Europe, Institut Mines-Télécom, Centre for Digital Systems mohamed.daoudi@imt-nord-europe.fr \n" + }, + { + "bbox": [ + 141, + 202, + 473, + 326 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 141, + 202, + 473, + 326 + ], + "type": "text", + "content": " Department of Architecture and Engineering University of Parma, Italy claudio.ferrari2@unipr.it \n" + }, + { + "bbox": [ + 141, + 202, + 473, + 326 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 141, + 202, + 473, + 326 + ], + "type": "text", + "content": " Univ. Lille, CNRS, UMR 8524 Laboratoire Paul Painlevé, Lille, F-59000, France sylvain,arguillere@univ-lille.fr" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 369, + 455, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 369, + 455, + 590 + ], + "spans": [ + { + "bbox": [ + 160, + 369, + 455, + 590 + ], + "type": "text", + "content": "Abstract. Speech-driven 3D talking heads generation has emerged as a significant area of interest among researchers, presenting numerous challenges. Existing methods are constrained by animating faces with fixed topologies, wherein point-wise correspondence is established, and the number and order of points remains consistent across all identities the model can animate. In this work, we present ScanTalk, a novel framework capable of animating 3D faces in arbitrary topologies including scanned data. Our approach relies on the DiffusionNet architecture to overcome the fixed topology constraint, offering promising avenues for more flexible and realistic 3D animations. By leveraging the power of DiffusionNet, ScanTalk not only adapts to diverse facial structures but also maintains fidelity when dealing with scanned data, thereby enhancing the authenticity and versatility of generated 3D talking heads. Through comprehensive comparisons with state-of-the-art methods, we validate the efficacy of our approach, demonstrating its capacity to generate realistic talking heads comparable to existing techniques. While our primary objective is to develop a generic method free from topological constraints, all state-of-the-art methodologies are bound by such limitations. Code for reproducing our results, and the pre-trained model are available at https://github.com/miccunifi/ScanTalk." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 601, + 441, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 601, + 441, + 613 + ], + "spans": [ + { + "bbox": [ + 160, + 601, + 441, + 613 + ], + "type": "text", + "content": "Keywords: 3D Talking Heads " + }, + { + "bbox": [ + 160, + 601, + 441, + 613 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 601, + 441, + 613 + ], + "type": "text", + "content": " 3D Scans Animation " + }, + { + "bbox": [ + 160, + 601, + 441, + 613 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 601, + 441, + 613 + ], + "type": "text", + "content": " DiffusionNet" + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 135, + 653, + 222, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 653, + 222, + 665 + ], + "spans": [ + { + "bbox": [ + 135, + 653, + 222, + 665 + ], + "type": "text", + "content": "* Equal contribution" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 117, + 479, + 227 + ], + "blocks": [ + { + "bbox": [ + 133, + 117, + 479, + 227 + ], + "lines": [ + { + "bbox": [ + 133, + 117, + 479, + 227 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 479, + 227 + ], + "type": "image", + "image_path": "eb9a4099eaec2b0d67bb6cc8deadfa0b594b21bf6626422e4bb946d43f8fe39f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 235, + 482, + 269 + ], + "lines": [ + { + "bbox": [ + 130, + 235, + 482, + 269 + ], + "spans": [ + { + "bbox": [ + 130, + 235, + 482, + 269 + ], + "type": "text", + "content": "Fig. 1: We present ScanTalk, a deep learning architecture to animate any 3D face mesh driven by a speech. ScanTalk is robust enough to learn on multiple unrelated datasets with a unique model, whilst allowing us to infer on unregistered face meshes." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 292, + 230, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 292, + 230, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 292, + 230, + 304 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 318, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 318, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 318, + 482, + 449 + ], + "type": "text", + "content": "Human face animation is a complex task, widely explored in Computer Vision and Computer Graphics because of the broad range of applications drawn from it, spanning from virtual reality to video game graphics, and more. As 3D face models continue to improve, it becomes more and more relevant to incorporate multi-modal aspects such as speech with audio data. Speech-driven facial animation encounters challenges inherent to the fact that finding a cross-modality mapping from audio to geometric data is an ill-posed problem, dealing with the complex geometry of human faces and the limited availability of paired audio-visual 3D data. Another, less discussed aspect that touches a broader range of 3D graphics with meshes is the robustness to changes in the topology of the mesh, which refers to the arrangement of the vertices and how they are connected." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 450, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 521 + ], + "type": "text", + "content": "Maintaining fidelity and coherence across different topologies is crucial for ensuring that facial animations remain realistic and expressive, regardless of variations in the underlying mesh structure. This challenge becomes particularly pronounced in speech-driven facial animation, where the dynamics of speech-related lip movements, and related face changes necessitate a high degree of adaptability within the mesh topology." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "type": "text", + "content": "Addressing these challenges requires innovative approaches able to navigate the complexities of facial geometry, while accommodating the nuances of speech-driven animation. In this regard, we present ScanTalk, a novel framework capable of animating faces in arbitrary topologies including scanned data. ScanTalk overcomes limitations associated with fixed topologies, offering promising avenues for more flexible and realistic 3D talking heads generation. Indeed, the aforementioned constraint limits the range of applications of current deep learning models for speech-driven motion synthesis. Because of this, deep models are required to train on large scale registered datasets such as FLAME [23] or Multiface [44] for human face data. Building these datasets is a costly procedure, and the trained model is then only usable within the same registered setting. This typically means that any newly acquired data needs to be fitted to the corre" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 241, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 241, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 241, + 100 + ], + "type": "text", + "content": "F. Nocentini et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "text", + "content": "sponding topology before any animation can be predicted by a deep model. This extra preprocessing step usually prevents online applications, therefore canceling one of the key benefits of deep models, which is their inference speed." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 480, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 480, + 188 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 480, + 188 + ], + "type": "text", + "content": "Aiming to address these limitations, in this paper we present a flexible deep learning framework built to generate speech-driven animations of 3D face meshes. In particular, it gathers several key elements:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 135, + 198, + 480, + 294 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 136, + 198, + 479, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 198, + 479, + 220 + ], + "spans": [ + { + "bbox": [ + 136, + 198, + 479, + 220 + ], + "type": "text", + "content": "1. A new robust approach to generate mesh deformation sequences based on DiffusionNet to compute intrinsic descriptors on 3D data;" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 135, + 222, + 480, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 222, + 480, + 269 + ], + "spans": [ + { + "bbox": [ + 135, + 222, + 480, + 269 + ], + "type": "text", + "content": "2. A comprehensive architecture for learning speech-driven animations. Our approach works with meshes with different topologies, while showing competitive performance with respect to state-of-the-art models trained on an individual topology;" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 135, + 270, + 479, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 270, + 479, + 294 + ], + "spans": [ + { + "bbox": [ + 135, + 270, + 479, + 294 + ], + "type": "text", + "content": "3. We show the generalizability of our model to unseen mesh topologies with qualitative examples." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 314, + 237, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 314, + 237, + 327 + ], + "spans": [ + { + "bbox": [ + 132, + 314, + 237, + 327 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 342, + 480, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 480, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 480, + 461 + ], + "type": "text", + "content": "In literature, there are several methodologies to acquire 3D face meshes, either by extracting the geometry from images or videos [26,47] or by using a complex set of instruments in a controlled environment [16, 18, 33, 35, 44, 48]. The former offers advantages in terms of easier and more cost-effective data acquisition. However, these methods may sometimes fall short in capturing the complete 3D information from 2D data. With 3D scans, other challenges arise: they come unregistered and may present alterations such as holes and noise. Then, animating these objects directly becomes even more challenging. One can do it by registering the meshes onto a pre-defined topology [16, 22, 23, 28, 44] before animating the registrations with 3D morphable models [14] for example." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 462, + 481, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 481, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 481, + 617 + ], + "type": "text", + "content": "Several deep learning models [3-5, 10, 25] proposed ways to learn a latent representation of face scans using robust encoders such as PointNet [6, 31] and Transformers but the resulting mesh is registered, which tends to smooth out some details from the scan geometry. However, this extra registration step may be handled efficiently with recent industrial applications such as MetaHuman from Epic Games. More recently, and closer to our goal, DiffusionNet [37] was combined with neural Jacobian fields [1] in Neural Face Rigging (NFR) [32] to learn a per-triangle deformation field on faces, allowing to transfer an animation from a sequence of unregistered face meshes to another. We stand out from this work by proposing a model that uses audio data to animate a given unregistered face mesh. While DiffusionNet has demonstrated its capability to generalize across varying triangulations in a static setting, our model is the first to exhibit similar properties in a multi-modal 4D setting." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 618, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 618, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 618, + 481, + 665 + ], + "type": "text", + "content": "In recent years, numerous models and methodologies have emerged to address the challenge of synchronizing facial animations with speech audio. While significant progress has been made on 2D talking heads [2, 7, 12, 21, 42, 43, 49], only a handful of approaches can be seamlessly extended to 3D data. Procedural" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 228, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 228, + 91, + 448, + 102 + ], + "type": "text", + "content": "ScanTalk: 3D Talking Heads from Unregistered Scans" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 354 + ], + "type": "text", + "content": "techniques have been proposed to animate 3D faces [9, 13, 27, 41, 46], primarily relying on visemes (groups of phonemes) to drive the movements of facial muscles. However, many of these models necessitate re-targeting and struggle to generalize effectively without extensive processing steps. More recent works leverage the increasing availability of data to develop statistical methods, including deep learning strategies [11, 15, 19, 29, 34, 38-40, 45]. Currently, state-of-the-art deep learning approaches are limited to a fixed mesh topology, requiring it to be identical to the one observed during the training phase. Among these models, VOCA [11] pioneers the development of deep models trained on a large-scale dataset of registered meshes. It was outperformed when MeshTalk [34] introduced a larger registered dataset named Multiface [44] along with a more expressive model. Moving forward, FaceFormer [15] utilized the transformer architecture, optimized for temporal data, and leveraged the power of a large-scale pre-trained audio encoder called Wav2vec2 [36]. This strategy saw further development with CodeTalker [45] and SelfTalk [30]. More recently, FaceXHubert [19] and FaceDiffuser [38] utilized an improved audio encoder, Hubert [20], demonstrating superiority to predict cross-modality mappings from audio to face motions. While continuously enhancing performances, these models are bound to a fixed mesh topology, hindering real-world applications and limiting the quantity of usable data for a single model, both for training and inference." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 356, + 482, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 356, + 482, + 428 + ], + "spans": [ + { + "bbox": [ + 130, + 356, + 482, + 428 + ], + "type": "text", + "content": "In response to these challenges, we present a novel framework in the landscape of 3D facial animation learning, avoiding the limitations posed by the specific topology of the face mesh to be animated. The proposed model stands out for its capability to animate any face mesh, including real-world 3D scans. This novel approach holds the potential to redefine the standards in the field of 3D facial animation, being applicable across diverse datasets and scenarios." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 447, + 330, + 461 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 330, + 461 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 330, + 461 + ], + "type": "text", + "content": "3 Proposed approach: ScanTalk" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 474, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 568 + ], + "type": "text", + "content": "In this section, we introduce Scantalk, a framework for animating 3D face meshes reproducing a spoken sentence contained in an audio file, which does not require the meshes to adhere to any specific topology. With ScanTalk, we push the state-of-the-art in 3D talking heads a step forward by allowing any 3D face, even raw scans, to be animated given a speech. To train ScanTalk, the only requirement is that the training meshes share a common topology within each sequence, but the topology may vary from one sequence to another. At inference time, any 3D face mesh in neutral state can be animated." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "ScanTalk is an Encoder-Decoder framework, which receives a neutral face mesh and an audio snippet as input, and outputs a sequence of per-vertex deformation fields, whose length depends on that of the audio. By summing each deformation field to the neutral input face, we ultimately obtain the animated sequence. The encoder is composed of two main modules: an audio encoder, which combines a pretrained encoder with a bi-directional LSTM that extracts audio features from the input speech, and a DiffusionNet encoder that computes surface descriptors from the neutral 3D face. These descriptors are then repli" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 241, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 241, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 241, + 101 + ], + "type": "text", + "content": "F. Nocentini et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": "ated and concatenated to the audio features and the resulting signal is fed to a DiffusionNet decoder that outputs the deformation to be applied to the neutral face. The framework is depicted in Fig. 2. Before providing the technical details in Sec. 3.1 and Sec. 3.2, below we introduce a few general notations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 168, + 482, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 168, + 482, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 168, + 482, + 282 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 130, + 168, + 482, + 282 + ], + "type": "inline_equation", + "content": "L = \\{(M_i^{gt}, m_i^n, A_i)\\}_{i=0}^{N-1}" + }, + { + "bbox": [ + 130, + 168, + 482, + 282 + ], + "type": "text", + "content": " denotes the training set comprising " + }, + { + "bbox": [ + 130, + 168, + 482, + 282 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 168, + 482, + 282 + ], + "type": "text", + "content": " samples, where " + }, + { + "bbox": [ + 130, + 168, + 482, + 282 + ], + "type": "inline_equation", + "content": "A_i" + }, + { + "bbox": [ + 130, + 168, + 482, + 282 + ], + "type": "text", + "content": " is an audio containing a spoken sentence, " + }, + { + "bbox": [ + 130, + 168, + 482, + 282 + ], + "type": "inline_equation", + "content": "M_i^{gt} = (m_i^0, \\ldots, m_i^{T_i-1}) \\in \\mathbb{R}^{T_i \\times V_i \\times 3}" + }, + { + "bbox": [ + 130, + 168, + 482, + 282 + ], + "type": "text", + "content": " represents a sequence of 3D faces (same topology) of length " + }, + { + "bbox": [ + 130, + 168, + 482, + 282 + ], + "type": "inline_equation", + "content": "T_i" + }, + { + "bbox": [ + 130, + 168, + 482, + 282 + ], + "type": "text", + "content": " synchronized with the spoken sentence in " + }, + { + "bbox": [ + 130, + 168, + 482, + 282 + ], + "type": "inline_equation", + "content": "A_i" + }, + { + "bbox": [ + 130, + 168, + 482, + 282 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 168, + 482, + 282 + ], + "type": "inline_equation", + "content": "m_i^n \\in \\mathbb{R}^{V_i \\times 3}" + }, + { + "bbox": [ + 130, + 168, + 482, + 282 + ], + "type": "text", + "content": " is a 3D neutral face. " + }, + { + "bbox": [ + 130, + 168, + 482, + 282 + ], + "type": "inline_equation", + "content": "V_i = |m_i^n|" + }, + { + "bbox": [ + 130, + 168, + 482, + 282 + ], + "type": "text", + "content": " is the number of vertices in the " + }, + { + "bbox": [ + 130, + 168, + 482, + 282 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 168, + 482, + 282 + ], + "type": "text", + "content": "-th 3D face sequence that needs to be consistent across each mesh in the " + }, + { + "bbox": [ + 130, + 168, + 482, + 282 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 168, + 482, + 282 + ], + "type": "text", + "content": "-th sequence together with the vertex connectivity, overall determining the topology of the surface. Our objective is to establish a mapping function that correlates an audio input " + }, + { + "bbox": [ + 130, + 168, + 482, + 282 + ], + "type": "inline_equation", + "content": "A_i" + }, + { + "bbox": [ + 130, + 168, + 482, + 282 + ], + "type": "text", + "content": ", and a neutral 3D face " + }, + { + "bbox": [ + 130, + 168, + 482, + 282 + ], + "type": "inline_equation", + "content": "m_i^n" + }, + { + "bbox": [ + 130, + 168, + 482, + 282 + ], + "type": "text", + "content": ", to the ground-truth sequence " + }, + { + "bbox": [ + 130, + 168, + 482, + 282 + ], + "type": "inline_equation", + "content": "M_i^{gt}" + }, + { + "bbox": [ + 130, + 168, + 482, + 282 + ], + "type": "text", + "content": ", expressed as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 249, + 300, + 481, + 315 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 300, + 481, + 315 + ], + "spans": [ + { + "bbox": [ + 249, + 300, + 481, + 315 + ], + "type": "interline_equation", + "content": "S c a n t a l k \\left(A _ {i}, m _ {i} ^ {n}\\right) \\approx M _ {i} ^ {g t}. \\tag {1}", + "image_path": "013d4692ec38bedf05af0a313a5cf6d75f31e5a79b1f2145b859ef687406c259.jpg" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 132, + 373, + 482, + 534 + ], + "blocks": [ + { + "bbox": [ + 132, + 373, + 482, + 534 + ], + "lines": [ + { + "bbox": [ + 132, + 373, + 482, + 534 + ], + "spans": [ + { + "bbox": [ + 132, + 373, + 482, + 534 + ], + "type": "image", + "image_path": "505f7a57701bde4b66873761340bbc034e7dda107df59dd43688f187efcf7cbb.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 541, + 482, + 631 + ], + "lines": [ + { + "bbox": [ + 130, + 541, + 482, + 631 + ], + "spans": [ + { + "bbox": [ + 130, + 541, + 482, + 631 + ], + "type": "text", + "content": "Fig. 2: Architecture of ScanTalk. A novel Encoder-Decoder framework designed to dynamically animate any 3D face based on a spoken sentence from an audio file. The Encoder integrates the 3D neutral face " + }, + { + "bbox": [ + 130, + 541, + 482, + 631 + ], + "type": "inline_equation", + "content": "m_i^n" + }, + { + "bbox": [ + 130, + 541, + 482, + 631 + ], + "type": "text", + "content": ", per-vertex surface features " + }, + { + "bbox": [ + 130, + 541, + 482, + 631 + ], + "type": "inline_equation", + "content": "P_i^n" + }, + { + "bbox": [ + 130, + 541, + 482, + 631 + ], + "type": "text", + "content": " (crucial for DiffusionNet and precomputed by the operators " + }, + { + "bbox": [ + 130, + 541, + 482, + 631 + ], + "type": "inline_equation", + "content": "OP" + }, + { + "bbox": [ + 130, + 541, + 482, + 631 + ], + "type": "text", + "content": "), and the audio file " + }, + { + "bbox": [ + 130, + 541, + 482, + 631 + ], + "type": "inline_equation", + "content": "A_i" + }, + { + "bbox": [ + 130, + 541, + 482, + 631 + ], + "type": "text", + "content": ", yielding a fusion of per-vertex and audio features. These combined descriptors, alongside " + }, + { + "bbox": [ + 130, + 541, + 482, + 631 + ], + "type": "inline_equation", + "content": "P_i^n" + }, + { + "bbox": [ + 130, + 541, + 482, + 631 + ], + "type": "text", + "content": ", are then passed to the Decoder, which mirrors a reversed DiffusionNet encoder structure. The Decoder predicts the deformation of the 3D neutral face, which is then combined with the original 3D neutral face " + }, + { + "bbox": [ + 130, + 541, + 482, + 631 + ], + "type": "inline_equation", + "content": "m_i^n" + }, + { + "bbox": [ + 130, + 541, + 482, + 631 + ], + "type": "text", + "content": " to generate the animated sequence." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 228, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 228, + 91, + 448, + 102 + ], + "type": "text", + "content": "ScanTalk: 3D Talking Heads from Unregistered Scans" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 251, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 251, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 251, + 126 + ], + "type": "text", + "content": "3.1 ScanTalk Encoder" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 136, + 482, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 136, + 482, + 172 + ], + "spans": [ + { + "bbox": [ + 130, + 136, + 482, + 172 + ], + "type": "text", + "content": "ScanTalk requires an audio snippet " + }, + { + "bbox": [ + 130, + 136, + 482, + 172 + ], + "type": "inline_equation", + "content": "A_{i}" + }, + { + "bbox": [ + 130, + 136, + 482, + 172 + ], + "type": "text", + "content": " and a 3D face in neutral state " + }, + { + "bbox": [ + 130, + 136, + 482, + 172 + ], + "type": "inline_equation", + "content": "m_{i}^{n}" + }, + { + "bbox": [ + 130, + 136, + 482, + 172 + ], + "type": "text", + "content": " as inputs. To process the two inputs, ScanTalk comprises two distinct encoders to process the mesh and the audio, respectively, as detailed below." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 190, + 482, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 190, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 190, + 482, + 297 + ], + "type": "text", + "content": "Face mesh encoder. Several approaches are available for encoding face meshes or point clouds, yet traditional graph convolution-based models, like [17, 24], encounter limitations with varying graph structures, such as changes in mesh resolution. To address this, we employ DiffusionNet [37], a discretization-agnostic encoder that demonstrated to be effective for encoding face meshes as seen in [32]. DiffusionNet integrates multi-layer perceptrons (MLPs), learned diffusion, and spatial gradient features, offering a straightforward yet robust architecture for surface learning tasks. It bypasses the need for complex operations like explicit surface convolutions or pooling hierarchies." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 298, + 482, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 298, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 298, + 482, + 369 + ], + "type": "text", + "content": "Critical to DiffusionNet's functionality are precomputed features of the face surface, namely the Cotangent Laplacian, Eigenbasis, Mass Matrix, and Spatial Gradient Matrix. Together, these operators enhance the architecture's robustness, flexibility, and expressiveness, making DiffusionNet suitable for diverse surface learning tasks. Notably, this architecture accommodates 3D faces of any topology, allowing for variations in the number and order of defining points." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 369, + 482, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 482, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 482, + 430 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 130, + 369, + 482, + 430 + ], + "type": "inline_equation", + "content": "P_{i}^{n} = OP(m_{i}^{n})" + }, + { + "bbox": [ + 130, + 369, + 482, + 430 + ], + "type": "text", + "content": " represent the precomputed features obtained by applying the above surface closed-form operators " + }, + { + "bbox": [ + 130, + 369, + 482, + 430 + ], + "type": "inline_equation", + "content": "OP" + }, + { + "bbox": [ + 130, + 369, + 482, + 430 + ], + "type": "text", + "content": " to the 3D neutral face " + }, + { + "bbox": [ + 130, + 369, + 482, + 430 + ], + "type": "inline_equation", + "content": "m_{i}^{n}" + }, + { + "bbox": [ + 130, + 369, + 482, + 430 + ], + "type": "text", + "content": ". The DiffusionNet Encoder " + }, + { + "bbox": [ + 130, + 369, + 482, + 430 + ], + "type": "inline_equation", + "content": "DN_{e}" + }, + { + "bbox": [ + 130, + 369, + 482, + 430 + ], + "type": "text", + "content": ", with latent size " + }, + { + "bbox": [ + 130, + 369, + 482, + 430 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 130, + 369, + 482, + 430 + ], + "type": "text", + "content": ", is designed to process a neutral 3D face mesh " + }, + { + "bbox": [ + 130, + 369, + 482, + 430 + ], + "type": "inline_equation", + "content": "m_{i}^{n}" + }, + { + "bbox": [ + 130, + 369, + 482, + 430 + ], + "type": "text", + "content": ", and requires the precomputed per-vertex features " + }, + { + "bbox": [ + 130, + 369, + 482, + 430 + ], + "type": "inline_equation", + "content": "P_{i}^{n}" + }, + { + "bbox": [ + 130, + 369, + 482, + 430 + ], + "type": "text", + "content": " to extract a per-vertex descriptors " + }, + { + "bbox": [ + 130, + 369, + 482, + 430 + ], + "type": "inline_equation", + "content": "f_{i}^{n}" + }, + { + "bbox": [ + 130, + 369, + 482, + 430 + ], + "type": "text", + "content": ", that is:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 244, + 438, + 480, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 438, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 244, + 438, + 480, + 453 + ], + "type": "interline_equation", + "content": "f _ {i} ^ {n} = D N _ {e} \\left(m _ {i} ^ {n}, P _ {i} ^ {n}\\right) \\in \\mathbb {R} ^ {V _ {i} \\times h}. \\tag {2}", + "image_path": "4219a4060a84a7755b7f4bbfd2e7410c5eb41f5d60010fe79c0f67d846fd4673.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 461, + 481, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 481, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 481, + 485 + ], + "type": "text", + "content": "The per-vertex descriptors " + }, + { + "bbox": [ + 130, + 461, + 481, + 485 + ], + "type": "inline_equation", + "content": "f_{i}^{n}" + }, + { + "bbox": [ + 130, + 461, + 481, + 485 + ], + "type": "text", + "content": ", is capable of capturing intricate details of each vertex within the neutral 3D face." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 503, + 482, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 503, + 482, + 563 + ], + "spans": [ + { + "bbox": [ + 130, + 503, + 482, + 563 + ], + "type": "text", + "content": "Audio encoder. Following [19, 38], the speech encoder adopts the architecture of the state-of-the-art pretrained speech model, HuBERT [20] that is a self-supervised speech representation learner utilizing an offline clustering step to provide aligned target labels for a BERT-like prediction loss. Using this module, followed by a Linear Layer, we obtain a per-frame audio representation:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 224, + 572, + 480, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 572, + 480, + 586 + ], + "spans": [ + { + "bbox": [ + 224, + 572, + 480, + 586 + ], + "type": "interline_equation", + "content": "a _ {i} = \\operatorname {S p e e c h E n c o d e r} \\left(A _ {i}\\right) \\in \\mathbb {R} ^ {T _ {i} \\times (h / 2)}. \\tag {3}", + "image_path": "64973c9c552184a7f301e93f9068005478176c27813fd5f83d87ccb3b208681d.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 595, + 482, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 595, + 482, + 643 + ], + "spans": [ + { + "bbox": [ + 130, + 595, + 482, + 643 + ], + "type": "text", + "content": "To ensure coherence between the speech representation, following the methodology outlined in [29], we concatenate the SpeechEncoder with a Multilayer Bidirectional-LSTM for temporal consistency, in a way such that the speech signal is projected into a temporal latent representation:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 246, + 652, + 480, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 652, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 246, + 652, + 480, + 666 + ], + "type": "interline_equation", + "content": "v _ {i} = B i L S T M \\left(a _ {i}\\right) \\in \\mathbb {R} ^ {T _ {i} \\times h}. \\tag {4}", + "image_path": "c2e492464095dd37ba3709dff6dc24e133c777a49134b7118b1e76fcb5ae2d04.jpg" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 241, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 241, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 241, + 101 + ], + "type": "text", + "content": "F. Nocentini et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 251, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 251, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 251, + 126 + ], + "type": "text", + "content": "3.2 ScanTalk Decoder" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 479, + 160 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 479, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 479, + 160 + ], + "type": "text", + "content": "We combine per-vertex descriptors " + }, + { + "bbox": [ + 130, + 137, + 479, + 160 + ], + "type": "inline_equation", + "content": "f_{i}^{n}" + }, + { + "bbox": [ + 130, + 137, + 479, + 160 + ], + "type": "text", + "content": " extracted from the neutral face, with the latent vector " + }, + { + "bbox": [ + 130, + 137, + 479, + 160 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 130, + 137, + 479, + 160 + ], + "type": "text", + "content": " extracted from the Bidirectional-LSTM" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 139, + 170, + 481, + 184 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 170, + 481, + 184 + ], + "spans": [ + { + "bbox": [ + 139, + 170, + 481, + 184 + ], + "type": "interline_equation", + "content": "(F _ {i} ^ {j}) _ {k} = \\left(f _ {i} ^ {n}\\right) _ {k} \\oplus v _ {i} ^ {j} \\in \\mathbb {R} ^ {h * 2}, \\quad \\forall k = 0, \\dots , V _ {i} - 1, \\quad \\forall j = 0, \\dots , T _ {i} - 1. \\tag {5}", + "image_path": "f8ec93fe9c02ad1d6b6d9bbef22008c462cde25ff2ef5f6f5d54c06dcd3bbdcb.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 194, + 480, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 194, + 480, + 268 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 480, + 268 + ], + "type": "text", + "content": "With this concatenation, we obtain a combined latent " + }, + { + "bbox": [ + 130, + 194, + 480, + 268 + ], + "type": "inline_equation", + "content": "F_{i}^{j} \\in \\mathbb{R}^{V_{i} \\times h*2}" + }, + { + "bbox": [ + 130, + 194, + 480, + 268 + ], + "type": "text", + "content": " that embeds both audio-related and geometry-related latents. To decode this sequence of combined latents for deforming the neutral face " + }, + { + "bbox": [ + 130, + 194, + 480, + 268 + ], + "type": "inline_equation", + "content": "m_{i}^{n}" + }, + { + "bbox": [ + 130, + 194, + 480, + 268 + ], + "type": "text", + "content": ", we employ a DiffusionNet Decoder (which is essentially a reversed Encoder), denoted as " + }, + { + "bbox": [ + 130, + 194, + 480, + 268 + ], + "type": "inline_equation", + "content": "DN_{d}" + }, + { + "bbox": [ + 130, + 194, + 480, + 268 + ], + "type": "text", + "content": ". The decoder module receives " + }, + { + "bbox": [ + 130, + 194, + 480, + 268 + ], + "type": "inline_equation", + "content": "F_{i}^{j}" + }, + { + "bbox": [ + 130, + 194, + 480, + 268 + ], + "type": "text", + "content": ", and precomputed features " + }, + { + "bbox": [ + 130, + 194, + 480, + 268 + ], + "type": "inline_equation", + "content": "P_{i}^{n}" + }, + { + "bbox": [ + 130, + 194, + 480, + 268 + ], + "type": "text", + "content": " derived from " + }, + { + "bbox": [ + 130, + 194, + 480, + 268 + ], + "type": "inline_equation", + "content": "m_{i}^{n}" + }, + { + "bbox": [ + 130, + 194, + 480, + 268 + ], + "type": "text", + "content": ". It predicts the deformation of " + }, + { + "bbox": [ + 130, + 194, + 480, + 268 + ], + "type": "inline_equation", + "content": "m_{i}^{n}" + }, + { + "bbox": [ + 130, + 194, + 480, + 268 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 130, + 194, + 480, + 268 + ], + "type": "inline_equation", + "content": "DN_{d}(F_{i}^{j}, P_{i}^{n})" + }, + { + "bbox": [ + 130, + 194, + 480, + 268 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 228, + 278, + 480, + 292 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 278, + 480, + 292 + ], + "spans": [ + { + "bbox": [ + 228, + 278, + 480, + 292 + ], + "type": "interline_equation", + "content": "\\widehat {m _ {i} ^ {j}} = D N _ {d} \\left(F _ {i} ^ {j}, P _ {i} ^ {n}\\right) + m _ {i} ^ {n} \\in \\mathbb {R} ^ {V _ {i} \\times 3}. \\tag {6}", + "image_path": "f5becc11060d682a9be3ff0a631da9b91b71ca5ba70ce71329f0d1aca63f93df.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 302, + 479, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 302, + 479, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 302, + 479, + 411 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 130, + 302, + 479, + 411 + ], + "type": "inline_equation", + "content": "\\widehat{m_i^j}" + }, + { + "bbox": [ + 130, + 302, + 479, + 411 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 130, + 302, + 479, + 411 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 302, + 479, + 411 + ], + "type": "text", + "content": "-th frame of the predicted sequence. The entire generated sequence is defined by " + }, + { + "bbox": [ + 130, + 302, + 479, + 411 + ], + "type": "inline_equation", + "content": "\\widehat{M_i} \\in \\mathbb{R}^{T_i \\times V_i \\times 3}" + }, + { + "bbox": [ + 130, + 302, + 479, + 411 + ], + "type": "text", + "content": ". We opt to utilize ScanTalk for predicting the deformation of the neutral face " + }, + { + "bbox": [ + 130, + 302, + 479, + 411 + ], + "type": "inline_equation", + "content": "m_i^n" + }, + { + "bbox": [ + 130, + 302, + 479, + 411 + ], + "type": "text", + "content": " rather than predicting the actual face. This decision aligns with previous works [15, 19, 30, 38, 40], and offers advantages in terms of training efficiency and resulting animation. Predicting face deformation makes the learning process easier, by focusing solely on speech-related motion, as opposed to incorporating the problem of predicting the entire face reconstruction. Essentially, the decoder learns to predict a per-vertex displacement field from a time-dependent per-vertex descriptors field." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 429, + 253, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 429, + 253, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 253, + 441 + ], + "type": "text", + "content": "3.3 ScanTalk Training" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 449, + 480, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 480, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 480, + 605 + ], + "type": "text", + "content": "ScanTalk generates deformations of a neutral face; hence, a predicted sequence maintains a consistent topology across all frames. This property arises from the definition of the DiffusionNet Decoder, which necessitates knowledge of precomputed features " + }, + { + "bbox": [ + 130, + 449, + 480, + 605 + ], + "type": "inline_equation", + "content": "P_{i}^{n}" + }, + { + "bbox": [ + 130, + 449, + 480, + 605 + ], + "type": "text", + "content": ", and the number of points in the 3D neutral face targeted for animation. For this reason, during a supervised training protocol, the ground-truth meshes within each sequence must adhere to a common topology. Despite the apparent specificity of this requirement, it proves to be non-prohibitive in practice. ScanTalk animates a neutral face in response to an audio input, producing a sequence of 3D faces sharing the same topology of the neutral face that is animated, which can be any topology (even different from those seen during training). This alignment between the training strategy and the desired inference outcome underscores the efficiency of ScanTalk, which is not restricted to the topology of the training faces." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 606, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 479, + 665 + ], + "type": "text", + "content": "Notably, the model predicts per-vertex displacements of the neutral ground-truth mesh and the ground truth sequence is registered in a supervised setting. Consequently, the predicted sequence and the ground-truth sequence are aligned on the same topology but can vary from one sequence to another. Thus, during training, we minimize the average vertex-wise Mean Squared Error (MSE) over" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 228, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 228, + 91, + 448, + 102 + ], + "type": "text", + "content": "ScanTalk: 3D Talking Heads from Unregistered Scans" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 115, + 480, + 141 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 115, + 480, + 141 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 480, + 141 + ], + "type": "text", + "content": "a sequence of length " + }, + { + "bbox": [ + 132, + 115, + 480, + 141 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 132, + 115, + 480, + 141 + ], + "type": "text", + "content": " between the ground truth " + }, + { + "bbox": [ + 132, + 115, + 480, + 141 + ], + "type": "inline_equation", + "content": "M_{i}^{gt}" + }, + { + "bbox": [ + 132, + 115, + 480, + 141 + ], + "type": "text", + "content": " and the model prediction " + }, + { + "bbox": [ + 132, + 115, + 480, + 141 + ], + "type": "inline_equation", + "content": "\\widehat{M}_{i}" + }, + { + "bbox": [ + 132, + 115, + 480, + 141 + ], + "type": "text", + "content": ". which is defined as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 194, + 152, + 480, + 184 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 152, + 480, + 184 + ], + "spans": [ + { + "bbox": [ + 194, + 152, + 480, + 184 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {M S E} = \\frac {1}{T _ {i} - 1} \\sum_ {j = 0} ^ {T _ {i} - 1} \\frac {1}{V _ {i} - 1} \\sum_ {k = 0} ^ {V _ {i} - 1} \\left\\| \\left(m _ {i} ^ {j}\\right) _ {k} - \\left(\\widehat {m _ {i}} ^ {j}\\right) _ {k} \\right\\| _ {2} ^ {2}. \\tag {7}", + "image_path": "9b65fdc72661b4d137ffa57e082123494cacb28aeff381f72ae4bef457ccc04e.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 201, + 230, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 201, + 230, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 201, + 230, + 216 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 228, + 479, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 228, + 479, + 299 + ], + "spans": [ + { + "bbox": [ + 131, + 228, + 479, + 299 + ], + "type": "text", + "content": "In the following, we first introduce the datasets and the metrics used for evaluation, respectively in Sec. 4.1 and Sec. 4.2. Then, we report quantitative and qualitative results in comparison with state-of-the-art methods in Sec. 4.3 and Sec. 4.4. In Sec. 4.5 we present several ablation studies over the framework architecture. Finally, in Sec. 4.6, we report the results of an user study designed to compare our approach with state-of-the-art methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 318, + 204, + 328 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 318, + 204, + 328 + ], + "spans": [ + { + "bbox": [ + 132, + 318, + 204, + 328 + ], + "type": "text", + "content": "4.1 Datasets" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 339, + 479, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 339, + 479, + 398 + ], + "spans": [ + { + "bbox": [ + 131, + 339, + 479, + 398 + ], + "type": "text", + "content": "For training and quantitative evaluations, we rely on three classical datasets in 3D speech-driven motion synthesis: VOCAset [11], BIWI [16] and Multiface [44]. VOCAset gathers mesh sequences of 12 actors performing 40 speeches, captured at 60fps. Each sequence lasts for around 3 to 5 seconds, and each mesh is registered to the FLAME [23] topology with 5,023 vertices and 9,976 faces." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 399, + 480, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 399, + 480, + 446 + ], + "spans": [ + { + "bbox": [ + 131, + 399, + 480, + 446 + ], + "type": "text", + "content": "BIWI comprises 14 subjects articulating 40 sentences each, sampled at 25fps. Each sentence lasts around 5 seconds, and the meshes are registered to a fixed topology. Due to GPU limitations, we used a downsampled version of this dataset, called " + }, + { + "bbox": [ + 131, + 399, + 480, + 446 + ], + "type": "inline_equation", + "content": "\\mathrm{BIWI}_6" + }, + { + "bbox": [ + 131, + 399, + 480, + 446 + ], + "type": "text", + "content": ", which has a fixed topology, with 3,895 vertices and 7,539 faces." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 446, + 479, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 446, + 479, + 481 + ], + "spans": [ + { + "bbox": [ + 131, + 446, + 479, + 481 + ], + "type": "text", + "content": "Multiface includes 13 identities executing up to 50 speeches of around 4 seconds each, sampled at 30fps. The original multiface meshes have a fixed topology with 5,471 vertices and 10,837 faces." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 483, + 479, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 483, + 479, + 566 + ], + "spans": [ + { + "bbox": [ + 131, + 483, + 479, + 566 + ], + "type": "text", + "content": "For training purposes, the meshes in both Multiface and BIWI" + }, + { + "bbox": [ + 131, + 483, + 479, + 566 + ], + "type": "inline_equation", + "content": "_6" + }, + { + "bbox": [ + 131, + 483, + 479, + 566 + ], + "type": "text", + "content": " have been scaled and aligned with the meshes in the VOCAset dataset. The data splits can be found in the supplementary material. Since ScanTalk is the first topology-independent 3D talking head generator, for the sake of a comprehensive comparison with the state-of-the-art, we trained 4 different models: one is trained using all the three datasets together (multi-dataset), while the other three are trained on each dataset separately (single-dataset)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 585, + 257, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 585, + 257, + 596 + ], + "spans": [ + { + "bbox": [ + 132, + 585, + 257, + 596 + ], + "type": "text", + "content": "4.2 Evaluation Metrics" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 605, + 482, + 665 + ], + "type": "text", + "content": "To evaluate the quality of the generated faces, we employ three standard metrics from previous works [30,34,38,45], namely: Lip vertex error (LVE) " + }, + { + "bbox": [ + 131, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\times 10^{-5}(mm)" + }, + { + "bbox": [ + 131, + 605, + 482, + 665 + ], + "type": "text", + "content": " Assesses the lip deviation in a sequence by comparing it to the ground truth. It is obtained by computing the maximum L2 error for all lip vertices in each frame, averaged across all frames." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 241, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 241, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 241, + 101 + ], + "type": "text", + "content": "F. Nocentini et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "content": "Mean vertex error (MVE) " + }, + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "inline_equation", + "content": "\\times 10^{-3}(mm)" + }, + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "content": ". Describes the quality of the overall face reconstruction by computing the maximal vertex L2 error for each frame and taking the mean across all frames." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 479, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 479, + 175 + ], + "type": "text", + "content": "Upper-face dynamic deviation (FDD) " + }, + { + "bbox": [ + 130, + 152, + 479, + 175 + ], + "type": "inline_equation", + "content": "\\times 10^{-7}(mm)" + }, + { + "bbox": [ + 130, + 152, + 479, + 175 + ], + "type": "text", + "content": ". Compares the standard deviation of upper vertices between the generated sequence and ground-truth." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 217, + 264, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 217, + 264, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 217, + 264, + 228 + ], + "type": "text", + "content": "4.3 Quantitative Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 247, + 480, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 480, + 366 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 480, + 366 + ], + "type": "text", + "content": "Although ScanTalk is the first deep learning model to animate unregistered 3D face meshes, comparisons with the state-of-the-art is possible when inferring our model on registered data. In Tab. 1, we present the performance evaluation of ScanTalk training on both a single-dataset and multi-datasets, juxtaposed against several state-of-the-art methods trained solely on a single dataset. While the augmentation of training data enhances the effectiveness of our model, addressing disparities in data composition becomes important for animating unseen faces. Particularly, the varied geometries and head dynamics inherent in " + }, + { + "bbox": [ + 130, + 247, + 480, + 366 + ], + "type": "inline_equation", + "content": "\\mathrm{BIWI}_6" + }, + { + "bbox": [ + 130, + 247, + 480, + 366 + ], + "type": "text", + "content": ", VOCA, and Multiface meshes pose significant challenges, potentially limiting performance compared to a model trained exclusively on a single dataset." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 401, + 480, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 401, + 480, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 480, + 489 + ], + "type": "text", + "content": "Table 1: ScanTalk performance, in both single-dataset (s-d) and multi-dataset (m-d) scenarios, in comparison with state-of-the-art methods. The heatmaps show the differences between the first frame and subsequent frames within sequences among the VOCAset, " + }, + { + "bbox": [ + 130, + 401, + 480, + 489 + ], + "type": "inline_equation", + "content": "\\mathrm{BIWI}_6" + }, + { + "bbox": [ + 130, + 401, + 480, + 489 + ], + "type": "text", + "content": ", and Multiface datasets. Notably, in the VOCAset, primarily the lips display movement, whereas in " + }, + { + "bbox": [ + 130, + 401, + 480, + 489 + ], + "type": "inline_equation", + "content": "\\mathrm{BIWI}_6" + }, + { + "bbox": [ + 130, + 401, + 480, + 489 + ], + "type": "text", + "content": " and Multiface datasets, substantial head and upper face movements are observed. The color gradient on the face meshes corresponds to the average per-vertex " + }, + { + "bbox": [ + 130, + 401, + 480, + 489 + ], + "type": "inline_equation", + "content": "L_2" + }, + { + "bbox": [ + 130, + 401, + 480, + 489 + ], + "type": "text", + "content": " norm of the differences, where blue hues indicate lower values, and red hues indicate higher values." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 226, + 501, + 257, + 550 + ], + "blocks": [ + { + "bbox": [ + 226, + 501, + 257, + 550 + ], + "lines": [ + { + "bbox": [ + 226, + 501, + 257, + 550 + ], + "spans": [ + { + "bbox": [ + 226, + 501, + 257, + 550 + ], + "type": "image", + "image_path": "1fe722d18ef68bdc0c0bc93ebd20e94ee33e2de08f9b34fd2762e984a10ffe48.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 322, + 505, + 348, + 540 + ], + "blocks": [ + { + "bbox": [ + 322, + 505, + 348, + 540 + ], + "lines": [ + { + "bbox": [ + 322, + 505, + 348, + 540 + ], + "spans": [ + { + "bbox": [ + 322, + 505, + 348, + 540 + ], + "type": "image", + "image_path": "8f329ff74c2ca8096ecc35d4960f69d09f901e004d9e634cc4b80402fccd7d35.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 413, + 498, + 447, + 555 + ], + "blocks": [ + { + "bbox": [ + 413, + 498, + 447, + 555 + ], + "lines": [ + { + "bbox": [ + 413, + 498, + 447, + 555 + ], + "spans": [ + { + "bbox": [ + 413, + 498, + 447, + 555 + ], + "type": "image", + "image_path": "571666e5c93356be6eb0d325bb1a06ac4fb02a042cdb3233d2c384868afa24f2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 143, + 554, + 466, + 662 + ], + "blocks": [ + { + "bbox": [ + 143, + 554, + 466, + 662 + ], + "lines": [ + { + "bbox": [ + 143, + 554, + 466, + 662 + ], + "spans": [ + { + "bbox": [ + 143, + 554, + 466, + 662 + ], + "type": "table", + "html": "
VOCAssetBIWI6Multiface
LVE ↓MVE ↓FDD ↓LVE ↓MVE ↓FDD ↓LVE ↓MVE ↓FDD ↓
VOCA6.9930.9832.6625.7432.59141.4824.9232.76155.781
FaceFormer6.1230.9352.1634.0852.16337.0912.4511.45320.239
SelfTalk5.6180.9182.3213.6282.06235.4702.2811.90137.434
CodeTalker3.5490.8882.2585.1902.64120.5994.0912.38247.905
FaceDiffuser4.3500.9012.4374.0222.12839.6043.5552.38829.157
ScanTalk m-d6.3750.9872.1014.0442.05740.0512.4351.67832.202
ScanTalk s-d3.0120.8612.4004.6512.14836.0342.6531.87164.451
", + "image_path": "94954a776536fe2d1b7cb37dfec6e8ee7bc6c9bec3ec2e7777f820870580a281.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 228, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 228, + 91, + 448, + 102 + ], + "type": "text", + "content": "ScanTalk: 3D Talking Heads from Unregistered Scans" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "content": "This observation becomes evident when assessing Scantalk's performance on VOCAset, where a substantial improvement is observed when training on a single-dataset. Conversely, in the case of " + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "inline_equation", + "content": "\\mathrm{BIWI}_6" + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "content": " and Multiface single-dataset training, a decline in results is noted compared to the multi-dataset version. This disparity is attributed to the presence of speech-related head and upper face movements within the ground truth talking head sequences of " + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "inline_equation", + "content": "\\mathrm{BIWI}_6" + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "content": " and Multiface. Consequently, the multi-dataset training model demonstrates superior performance as it encounters a wider array of sequences featuring head and upper face movements. In contrast, VOCAset lacks such movements, with the 3D faces remaining static, while only the mouth moves, thereby enhancing performance for ScantTalk trained solely on VOCAset. This obviates the need to learn other movements of the head or upper face." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 261, + 482, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 261, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 261, + 482, + 331 + ], + "type": "text", + "content": "The distinctions between VOCAset and the other two datasets are further underscored by the FDD metric, indicating that 3D faces in VOCAset remain stationary, with minimal activity in the upper facial area and head. Across all tested models, generating sequences with lower FDD is notably more achievable on VOCAset compared to the other datasets, thus emphasizing the disparities between VOCAset and the others." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 333, + 351, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 333, + 351, + 487 + ], + "spans": [ + { + "bbox": [ + 130, + 333, + 351, + 487 + ], + "type": "text", + "content": "Nevertheless, the results presented in Tab. 1 demonstrate that ScanTalk produces outcomes comparable to state-of-the-art methods, whether undergoing single-dataset or multi-dataset training. Moreover, ScanTalk emerges as the first model capable of training in multi-dataset settings, demonstrating the ability to animate a 3D face regardless of its topology. An interesting characteristic of ScanTalk emerges when analyzing the GPU memory usage relative to the vertex count in the 3D facial model requiring animation. Fig. 3 illustrates a linear increase in GPU usage correlated with the number of vertices." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 362, + 356, + 476, + 435 + ], + "blocks": [ + { + "bbox": [ + 362, + 356, + 476, + 435 + ], + "lines": [ + { + "bbox": [ + 362, + 356, + 476, + 435 + ], + "spans": [ + { + "bbox": [ + 362, + 356, + 476, + 435 + ], + "type": "image", + "image_path": "bd80a0f7c0b1ac3c572bd671c1a0bcf29a8349f9de5fdb3352912e1e4b6ea7f2.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 356, + 443, + 482, + 475 + ], + "lines": [ + { + "bbox": [ + 356, + 443, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 356, + 443, + 482, + 475 + ], + "type": "text", + "content": "Fig.3: ScanTalk GPU memory usage with respect to the mesh resolution." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 510, + 257, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 510, + 257, + 521 + ], + "spans": [ + { + "bbox": [ + 132, + 510, + 257, + 521 + ], + "type": "text", + "content": "4.4 Qualitative Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "type": "text", + "content": "In the domain of 3D speech-driven talking heads generation, qualitative evaluations hold as much significance as quantitative assessments. To evaluate the quantitative findings presented in Sec. 4.3, we introduce a straightforward test aimed at discerning disparities in head and upper face movements across the three datasets: for each dataset, we compute the average " + }, + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "type": "text", + "content": " norm of the differences between the initial frame and subsequent frames within each sequence. This enables the quantification of both head dynamics and upper facial movements across mesh sequences. In the heatmaps of Tab. 1, we present the per-frame average difference for each dataset. Our analysis reveals that in the VOCAset, only the lower part of the face exhibits discernible movements, whereas in BIWI" + }, + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "type": "inline_equation", + "content": "_6" + }, + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "type": "text", + "content": " and Multiface, the entire head or face moves." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 241, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 241, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 241, + 101 + ], + "type": "text", + "content": "F. Nocentini et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 152, + 119, + 460, + 278 + ], + "blocks": [ + { + "bbox": [ + 152, + 119, + 460, + 278 + ], + "lines": [ + { + "bbox": [ + 152, + 119, + 460, + 278 + ], + "spans": [ + { + "bbox": [ + 152, + 119, + 460, + 278 + ], + "type": "image", + "image_path": "297b328e04cfc61facd6bd7b14328f3fcd4731285f89866ed76d9130a7879461.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 288, + 482, + 354 + ], + "lines": [ + { + "bbox": [ + 130, + 288, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 288, + 482, + 354 + ], + "type": "text", + "content": "Fig. 4: Relative norm of the per-vertex descriptors " + }, + { + "bbox": [ + 130, + 288, + 482, + 354 + ], + "type": "inline_equation", + "content": "f_{i}^{n}" + }, + { + "bbox": [ + 130, + 288, + 482, + 354 + ], + "type": "text", + "content": " in Eq. (2) extracted by " + }, + { + "bbox": [ + 130, + 288, + 482, + 354 + ], + "type": "inline_equation", + "content": "DN_{e}" + }, + { + "bbox": [ + 130, + 288, + 482, + 354 + ], + "type": "text", + "content": " displayed as a heatmap on a mesh from VOCAset (left), and a mesh from Multiface (right). For each mesh, we show the norm on the original topology, on a remeshed version, and on a further degraded mesh obtained by removing the back of the head and creating random holes. Here, pinker hues indicate lower values, and greener hues indicate higher values." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 372, + 482, + 588 + ], + "blocks": [ + { + "bbox": [ + 133, + 372, + 482, + 588 + ], + "lines": [ + { + "bbox": [ + 133, + 372, + 482, + 588 + ], + "spans": [ + { + "bbox": [ + 133, + 372, + 482, + 588 + ], + "type": "image", + "image_path": "1342fd14a0bf9b6e8e99754d04b9a0a63353d62dd6c235986ad1646955114338.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 597, + 482, + 632 + ], + "lines": [ + { + "bbox": [ + 130, + 597, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 597, + 482, + 632 + ], + "type": "text", + "content": "Fig. 5: ScanTalk inference on 3D faces. The meshes have been rigidly aligned with the training data. The first row is an animation of a raw 3D scan (an hole for the mouth has been created), while the others two are animation of meshes in an arbitrary topology." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 228, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 228, + 91, + 448, + 102 + ], + "type": "text", + "content": "ScanTalk: 3D Talking Heads from Unregistered Scans" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "content": "In Fig. 4, we present a heatmap visualization representing the norm of geometric descriptors " + }, + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "inline_equation", + "content": "f_{i}^{n}" + }, + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "content": " computed by the encoder " + }, + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "inline_equation", + "content": "DN_{e}" + }, + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "content": " in Eq. (2) for various alterations of the same identity mesh. Our findings indicate that the ScanTalk encoder effectively extracts descriptors from input faces regardless of facial topology. Notably, despite differences in facial topology among the same identity, the extracted descriptors exhibit remarkable similarity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 480, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 480, + 271 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 480, + 271 + ], + "type": "text", + "content": "In Fig. 5, 3D faces in different topology, including a scan, animated using ScanTalk are presented. Our method demonstrates good generalization capabilities, successfully animating diverse 3D faces. Notably, enhanced animation quality is observed when a mouth aperture is present; however, our model performs well even in its absence. Nonetheless, without a mouth aperture, the model struggles to generate the corresponding mouth opening, despite accurate lip movement synchronization." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 290, + 250, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 290, + 250, + 301 + ], + "spans": [ + { + "bbox": [ + 132, + 290, + 250, + 301 + ], + "type": "text", + "content": "4.5 Ablations Studies" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 311, + 479, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 311, + 479, + 370 + ], + "spans": [ + { + "bbox": [ + 130, + 311, + 479, + 370 + ], + "type": "text", + "content": "To investigate the impacts of various components within our architecture, we conducted extensive experiments across various configurations, encompassing both single-dataset and multi-dataset training paradigms. The results reported in Tab. 2 and Tab. 3 are computed by modifying ScanTalk's modules and training losses as described in Sec. 3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 389, + 479, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 389, + 479, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 479, + 472 + ], + "type": "text", + "content": "Audio encoder. In learning based speech-driven animation, some previous work used either Wav2vec2 [36] or Hubert [20], the latter being substantially better for this task according to [19,38]. For our purpose, we also tested WavLM [8] but obtained poorer efficiency. The results on both the single-dataset and multi-dataset training are displayed in Tab. 2a and Tab. 3 top. We can see that, in both cases, the usage of the Hubert Encoder [19] for audio features extraction leads to better results." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 491, + 480, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 491, + 480, + 622 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 480, + 622 + ], + "type": "text", + "content": "Temporal consistency. In our investigation, as proven by [19,29,40], we found that supplementing pre-trained audio encoders with additional temporal consistency mechanisms such as Bidirectional-LSTM, Bidirectional-GRU, or an autoregressive Transformer Decoder (TD) significantly enhances model performance, as illustrated in Tab. 2b and in Tab. 3 middle. The BiLSTM model architecture is detailed in Sec. 3, while the BiGRU model mirrors the BiLSTM architecture but substitutes BiLSTM with BiGRU. Details about the Transformer Decoder can be found in the supplementary material. From the findings presented in Tab. 2b and in Tab. 3 middle, we observe that employing a multilayer Bidirectional-LSTM in the audio stream processing of ScanTalk yields the most favorable performance across both single-dataset training and multi-dataset training scenarios." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "Loss function. Previous research has extensively explored optimal objective functions for enhancing and refining the learning process. Inspired by [29,30,40]," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 241, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 241, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 241, + 101 + ], + "type": "text", + "content": "F. Nocentini et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "we experimented with Mean Square Error (" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "L_{MSE}" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "), Masked (" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "L_{mask}" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "), and Velocity (" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "L_{vel}" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": ") Loss. The former is a simple " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "L_2" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " loss, the second employs Mean Square Error focused on lip vertices, while the latter aims to minimize differences between consecutive frames. While these loss functions have demonstrated efficacy in single-dataset training, as evidenced by [29,30,40], our findings, detailed in Tab. 3 bottom, indicate that employing a straightforward " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "L_2" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": ", on multi-dataset training, enhances the model's capacity to generate realistic talking heads. We attribute this to significant geometric variations observed across different datasets." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 149, + 291, + 293, + 340 + ], + "blocks": [ + { + "bbox": [ + 131, + 232, + 481, + 255 + ], + "lines": [ + { + "bbox": [ + 131, + 232, + 481, + 255 + ], + "spans": [ + { + "bbox": [ + 131, + 232, + 481, + 255 + ], + "type": "text", + "content": "Table 2: ScanTalk (ST) single-dataset ablation studies. Results obtained with a model trained just on VOCAsset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 169, + 274, + 272, + 283 + ], + "lines": [ + { + "bbox": [ + 169, + 274, + 272, + 283 + ], + "spans": [ + { + "bbox": [ + 169, + 274, + 272, + 283 + ], + "type": "text", + "content": "(a) Audio Encoder Ablation." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 149, + 291, + 293, + 340 + ], + "lines": [ + { + "bbox": [ + 149, + 291, + 293, + 340 + ], + "spans": [ + { + "bbox": [ + 149, + 291, + 293, + 340 + ], + "type": "table", + "html": "
LVE ↓MVE ↓FDD ↓
ST w WavLM3.6740.9372.413
ST w Wav2Vec23.3090.8602.244
ScanTalk3.0120.8612.400
", + "image_path": "944ccbd6a36f4349563524f38b2408ca281973093ac4be9773a960887227200b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 326, + 291, + 460, + 348 + ], + "blocks": [ + { + "bbox": [ + 331, + 274, + 455, + 284 + ], + "lines": [ + { + "bbox": [ + 331, + 274, + 455, + 284 + ], + "spans": [ + { + "bbox": [ + 331, + 274, + 455, + 284 + ], + "type": "text", + "content": "(b) Temporal consistency ablation." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 326, + 291, + 460, + 348 + ], + "lines": [ + { + "bbox": [ + 326, + 291, + 460, + 348 + ], + "spans": [ + { + "bbox": [ + 326, + 291, + 460, + 348 + ], + "type": "table", + "html": "
LVE ↓MVE ↓FDD ↓
ST w/o3.3610.8702.365
ST w TD3.2910.8592.406
ST w BiGRU3.0360.8352.358
ScanTalk3.0120.8612.400
", + "image_path": "2d139170851efc67ce007d0a75659d1692a90cfea38e95adf4dab993ba835c44.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 144, + 435, + 470, + 558 + ], + "blocks": [ + { + "bbox": [ + 131, + 392, + 482, + 426 + ], + "lines": [ + { + "bbox": [ + 131, + 392, + 482, + 426 + ], + "spans": [ + { + "bbox": [ + 131, + 392, + 482, + 426 + ], + "type": "text", + "content": "Table 3: ScanTalk (ST) multi-dataset ablation studies. Top: ST with different audio encoders. Middle: ST using different audio stream processing. Bottom: ST with different Loss function. The last row is the proposed ScanTalk" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 144, + 435, + 470, + 558 + ], + "lines": [ + { + "bbox": [ + 144, + 435, + 470, + 558 + ], + "spans": [ + { + "bbox": [ + 144, + 435, + 470, + 558 + ], + "type": "table", + "html": "
VOCAsetBIWI6Multiface
LVE ↓MVE ↓FDD ↓LVE ↓MVE ↓FDD ↓LVE ↓MVE ↓FDD ↓
ST w Wav2Vec27.1271.2412.3494.4392.27641.3743.1982.32933.671
ST w WavLM6.7731.0042.2204.3332.16231.9222.9072.0048.500
ST w TD6.5120.9951.8584.8272.16836.7892.1361.72535.044
ST w BiGRU6.5840.9942.0544.4212.10341.4562.5341.98734.098
ST+Lmask+Lvel7.0311.0921.4734.7552.22739.4582.5471.89815.973
ST+Lmask7.4511.0840.8594.7482.19936.9392.8881.89132.501
ST+Lvel6.7400.9981.8994.5092.18033.6142.1031.77531.599
ScanTalk6.3750.9872.1014.0442.05740.0512.4351.67832.202
", + "image_path": "4063bb6e5e28d1ddd2ae53723b40b65fc61013c337f27f0b0af3e83e1d4f6c86.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 597, + 217, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 597, + 217, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 597, + 217, + 609 + ], + "type": "text", + "content": "4.6 User study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "To further evaluate our solution, we performed a study where human feedback is involved. This evaluation is conducted with two user-based studies involving 25 participants; " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "(i)" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " For the first study, in alignment with prior research [15,19,30,34, 38,45], we designed an A/B test to compare ScanTalk with other state-of-the-art" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 228, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 228, + 91, + 448, + 102 + ], + "type": "text", + "content": "ScanTalk: 3D Talking Heads from Unregistered Scans" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "models within a registered setting on both lip-syncing and naturalness criteria (Test 1); (ii) In the second test, we assessed the credibility of scan animations by asking participants to rate the animation quality of ten scans sourced from the COMA dataset [33], using a scale ranging from 1 to 10 (Test 2)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 480, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 480, + 258 + ], + "type": "text", + "content": "The outcomes of Test 1 are depicted in the table on the left of Fig. 6. Notably, sequences produced using ScanTalk demonstrate levels of both naturalness and lip-syncing that are comparable to state-of-the-art methods. Specifically, our approach is preferred when compared with FaceFormer, CodeTalker, and FaceDiffuser on VOCAsset. However, SelfTalk and the ground truth are preferred over ScanTalk. Nevertheless, the percentage of users favoring ScanTalk remains nonnegligible, underscoring the efficacy of our approach in generating 3D talking heads with good realism and lip-syncing fidelity." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 259, + 479, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 479, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 479, + 308 + ], + "type": "text", + "content": "Results of Test 2 are reported on the right of Fig. 6. These ratings are closely linked to both the scan quality and the fidelity of the mouth representation. Despite the scan quality not being optimal, as depicted in Fig. 6, our results indicate that we can achieve a good level of realism in the animated scans." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 133, + 352, + 278, + 403 + ], + "blocks": [ + { + "bbox": [ + 133, + 352, + 278, + 403 + ], + "lines": [ + { + "bbox": [ + 133, + 352, + 278, + 403 + ], + "spans": [ + { + "bbox": [ + 133, + 352, + 278, + 403 + ], + "type": "table", + "html": "
ComparisonNaturalness (%)Lip-sync (%)
ST vs. FaceFormer82.6780.67
ST vs. CodeTalker57.3356.00
ST vs. FaceDiffuser90.6794.00
ST vs. SelfTalk44.0043.33
ST vs. GT44.0042.00
", + "image_path": "b5cad12d0dda874011b015b93bab35f4a60205631c74da298da269f31590d5e4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 283, + 327, + 479, + 449 + ], + "blocks": [ + { + "bbox": [ + 283, + 327, + 479, + 449 + ], + "lines": [ + { + "bbox": [ + 283, + 327, + 479, + 449 + ], + "spans": [ + { + "bbox": [ + 283, + 327, + 479, + 449 + ], + "type": "image", + "image_path": "3d30e1e766fee7c38debca7d1ffd148c36d3676c9d9770c00fbb950612675710.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 456, + 480, + 512 + ], + "lines": [ + { + "bbox": [ + 130, + 456, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 130, + 456, + 480, + 512 + ], + "type": "text", + "content": "Fig. 6: The table on the left shows results of Test 1, ScanTalk (ST) vs. other methods and the ground truth (GT) on samples from the test set of VOCAset. The values denote the percentages of users who favored ScanTalk animations over others. The chart on the right shows results of Test 2 reported as a violin plot for the animation of scans from COMA [33]. The subject median rating is displayed as a black dot on each violin." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 554, + 224, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 554, + 224, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 554, + 224, + 566 + ], + "type": "text", + "content": "5 Conclusions" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 579, + 480, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 579, + 480, + 662 + ], + "spans": [ + { + "bbox": [ + 130, + 579, + 480, + 662 + ], + "type": "text", + "content": "This paper introduces ScanTalk, a novel framework for speech-driven 3D facial animation. Unlike existing methods, ScanTalk possesses the unique capability to animate any 3D face, regardless of its topology, even if it differs from the ones on which it was trained. ScanTalk extends the applicability of deep speech-driven 3D facial animations by addressing the challenges of topology robustness. Additionally, our model demonstrates comparable quantitative and qualitative results with other state-of-the-art methods across three distinct datasets." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 241, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 241, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 241, + 101 + ], + "type": "text", + "content": "F. Nocentini et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "Acknowledgments This work is supported by the ANR project Human4D ANR-19-CE23-0020 and by the IRP CNRS project GeoGen3DHuman. This work was also partially supported by \"Partenariato FAIR (Future Artificial Intelligence Research) - PE00000013, CUP J33C22002830006\" funded by NextGenerationEU through the italian MUR within NRRP, project DL-MIG. This work was also partially funded by the ministerial decree n.352 of the 9th April 2022 NextGenerationEU through the italian MUR within NRRP. This work was also partially supported by Fédération de Recherche Mathématique des Hauts-de-France (FMHF, FR2037 du CNRS)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 243, + 198, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 243, + 198, + 255 + ], + "spans": [ + { + "bbox": [ + 132, + 243, + 198, + 255 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 269, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 138, + 269, + 481, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 269, + 481, + 312 + ], + "spans": [ + { + "bbox": [ + 138, + 269, + 481, + 312 + ], + "type": "text", + "content": "1. Aigerman, N., Gupta, K., Kim, V.G., Chaudhuri, S., Saito, J., Groueix, T.: Neural jacobian fields: Learning intrinsic mappings of arbitrary meshes. ACM Trans. Graph. 41(4) (jul 2022). https://doi.org/10.1145/3528223.3530141, https://doi.org/10.1145/3528223.3530141" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 313, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 313, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 138, + 313, + 481, + 346 + ], + "type": "text", + "content": "2. Alghamdi, M.M., Wang, H., Bulpitt, A.J., Hogg, D.C.: Talking head from speech audio using a pre-trained image generator. In: Proceedings of the 30th ACM International Conference on Multimedia (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 346, + 481, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 346, + 481, + 379 + ], + "spans": [ + { + "bbox": [ + 138, + 346, + 481, + 379 + ], + "type": "text", + "content": "3. Bahri, M., O' Sullivan, E., Gong, S., Liu, F., Liu, X., Bronstein, M.M., Zafeiriou, S.: Shape my face: Registering 3d face scans by surface-to-surface translation. International Journal of Computer Vision (IJCV) (Sep 2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 380, + 481, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 380, + 481, + 422 + ], + "spans": [ + { + "bbox": [ + 138, + 380, + 481, + 422 + ], + "type": "text", + "content": "4. Besnier, T., Arguillère, S., Pierson, E., Daoudi, M.: Toward Mesh-Invariant 3D Generative Deep Learning with Geometric Measures. Computers and Graphics (2023). https://doi.org/10.1016/j.cag.2023.06.027, https://hal.science/hal-04143649" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 423, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 423, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 138, + 423, + 481, + 456 + ], + "type": "text", + "content": "5. Chandran, P., Zoss, G., Gross, M., Gotardo, P., Bradley, D.: Shape transformers: Topology-independent 3d shape models using transformers. In: Computer Graphics Forum. vol. 41, pp. 195-207. Wiley Online Library (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 457, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 457, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 138, + 457, + 481, + 490 + ], + "type": "text", + "content": "6. Charles, R.Q., Su, H., Kaichun, M., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). p. 77-85. IEEE (Jul 2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 491, + 481, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 491, + 481, + 521 + ], + "spans": [ + { + "bbox": [ + 138, + 491, + 481, + 521 + ], + "type": "text", + "content": "7. Chen, L., Cui, G., Liu, C., Li, Z., Kou, Z., Xu, Y., Xu, C.: Talking-head generation with rhythmic head motion. In: European Conference on Computer Vision (2020), https://api-semanticscholar.org/CorpusID:220633152" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 522, + 481, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 522, + 481, + 577 + ], + "spans": [ + { + "bbox": [ + 138, + 522, + 481, + 577 + ], + "type": "text", + "content": "8. Chen, S., Wang, C., Chen, Z., Wu, Y., Liu, S., Chen, Z., Li, J., Kanda, N., Yoshioka, T., Xiao, X., Wu, J., Zhou, L., Ren, S., Qian, Y., Qian, Y., Zeng, M., Wei, F.: Wavlm: Large-scale self-supervised pre-training for full stack speech processing. IEEE Journal of Selected Topics in Signal Processing 16, 1505-1518 (2021), https://api-semanticscholar.org/CorpusID:239885872" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 578, + 481, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 481, + 620 + ], + "type": "text", + "content": "9. Cosi, P., Caldognetto, E., Perin, G., Zmarich, C.: Labial coarticulation modeling for realistic facial animation. In: Proceedings. Fourth IEEE International Conference on Multimodal Interfaces. pp. 505-510 (2002). https://doi.org/10.1109/ICMI.2002.1167047" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 621, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 621, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 621, + 481, + 665 + ], + "type": "text", + "content": "10. Croquet, B., Christiaens, D., Weinberg, S.M., Bronstein, M., Vandermeulen, D., Claes, P.: Unsupervised diffeomorphic surface registration and non-linear modelling. In: Medical Image Computing and Computer Assisted Intervention (MIC-CAI). p. 118–128. Springer (2021)" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 228, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 228, + 91, + 447, + 102 + ], + "type": "text", + "content": "ScanTalk: 3D Talking Heads from Unregistered Scans" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "11. Cudeiro, D., Bolkart, T., Laidlaw, C., Ranjan, A., Black, M.: Capture, learning, and synthesis of 3D speaking styles. In: Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR). pp. 10101-10111 (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "type": "text", + "content": "12. Das, D., Biswas, S., Sinha, S., Bhowmick, B.: Speech-driven facial animation using cascaded gans for learning of motion and texture. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J.M. (eds.) Computer Vision - ECCV 2020. pp. 408-424. Springer International Publishing, Cham (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 194, + 480, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 194, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 133, + 194, + 480, + 236 + ], + "type": "text", + "content": "13. Edwards, P., Landreth, C., Fiume, E., Singh, K.: Jali: an animator-centric viseme model for expressive lip synchronization. ACM Trans. Graph. 35(4) (jul 2016). https://doi.org/10.1145/2897824.2925984, https://doi.org/10.1145/2897824.2925984" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 237, + 480, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 480, + 292 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 480, + 292 + ], + "type": "text", + "content": "14. Egger, B., Smith, W.A.P., Tewari, A., Wuhrer, S., Zollhoefer, M., Beeler, T., Bernard, F., Bolkart, T., Kortylewski, A., Romdhani, S., Theobalt, C., Blanz, V., Vetter, T.: 3D Morphable Face Models - Past, Present and Future. ACM Transactions on Graphics 39(5), 157:1-38 (Aug 2020). https://doi.org/10.1145/3395208, https://inria.hal.science/hal-02280281" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 293, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 346 + ], + "type": "text", + "content": "15. Fan, Y., Lin, Z., Saito, J., Wang, W., Komura, T.: Faceformer: Speech-driven 3d facial animation with transformers. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). p. 18749-18758. IEEE, New Orleans, LA, USA (Jun 2022). https://doi.org/10.1109/CVPR52688.2022.01821, https://ieeexplore.ieee.org/document/9878591/" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "text", + "content": "16. Fanelli, G., Gall, J., Romsdorfer, H., Weise, T., Van Gool, L.: A 3-d audio-visual corpus of affective communication. IEEE Transactions on Multimedia 12(6), 591-598 (2010). https://doi.org/10.1109/TMM.2010.2052239" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "text", + "content": "17. Gong, S., Chen, L., Bronstein, M., Zafeiriou, S.: Spiralnet++: A fast and highly efficient mesh convolution operator. In: Proceedings of the IEEE International Conference on Computer Vision Workshops. pp. 0-0 (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 413, + 480, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 480, + 446 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 480, + 446 + ], + "type": "text", + "content": "18. Gupta, S., Castleman, K.R., Markey, M.K., Bovik, A.C.: Texas 3d face recognition database. In: 2010 IEEE Southwest Symposium on Image Analysis & Interpretation (SSIAI). pp. 97-100 (2010). https://doi.org/10.1109/SSIAI.2010.5483908" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 447, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 447, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 133, + 447, + 480, + 500 + ], + "type": "text", + "content": "19. Haque, K.I., Yumak, Z.: Facexhubert: Text-less speech-driven e(x)pressive 3d facial animation synthesis using self-supervised speech representation learning. In: INTERNATIONAL CONFERENCE ON MULTIMODAL INTERACTION (ICMI '23). ACM, New York, NY, USA (2023). https://doi.org/10.1145/3577190.3614157, https://doi.org/10.1145/3577190.3614157" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 501, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 555 + ], + "type": "text", + "content": "20. Hsu, W.N., Bolte, B., Tsai, Y.H.H., Lakhotia, K., Salakhutdinov, R., Mohamed, A.: Hubert: Self-supervised speech representation learning by masked prediction of hidden units. IEEE/ACM Trans. Audio, Speech and Lang. Proc. 29, 3451-3460 (oct 2021). https://doi.org/10.1109/TASLP.2021.3122291, https://doi.org/10.1109/TASLP.2021.3122291" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 555, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 480, + 600 + ], + "type": "text", + "content": "21. Ji, X., Zhou, H., Wang, K., Wu, Q., Wu, W., Xu, F., Cao, X.: Eamm: One-shot emotional talking face via audio-based emotion-aware motion model. In: ACM SIGGRAPH 2022 Conference Proceedings. SIGGRAPH '22 (2022). https://doi.org/10.1145/3528233.3530745, https://doi.org/10.1145/3528233.3530745" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 601, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 480, + 632 + ], + "type": "text", + "content": "22. Li, J., Kuang, Z., Zhao, Y., He, M., Bladin, K., Li, H.: Dynamic facial asset and rig generation from a single scan. ACM Trans. Graph. 39(6) (nov 2020). https://doi.org/10.1145/3414685.3417817, https://doi.org/10.1145/3414685.3417817" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "text", + "content": "23. Li, T., Bolkart, T., Black, M.J., Li, H., Romero, J.: Learning a model of facial shape and expression from 4D scans. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia) 36(6), 194:1-194:17 (2017), https://doi.org/10.1145/3130800.3130813" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 241, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 241, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 241, + 100 + ], + "type": "text", + "content": "F. Nocentini et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "text", + "content": "24. Lim, I., Dielen, A., Campen, M., Kobbelt, L.: A simple approach to intrinsic correspondence learning on unstructured 3d meshes. p. 349-362. Springer-Verlag, Berlin, Heidelberg (2019). https://doi.org/10.1007/978-3-030-11015-4_26, https://doi.org/10.1007/978-3-030-11015-4_26" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 161, + 482, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 161, + 482, + 204 + ], + "spans": [ + { + "bbox": [ + 130, + 161, + 482, + 204 + ], + "type": "text", + "content": "25. Liu, F., Tran, L., Liu, X.: 3d face modeling from diverse raw scan data. In: 2019 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 9407-9417. IEEE Computer Society, Los Alamitos, CA, USA (nov 2019). https://doi.org/10.1109/ICCV.2019.00950" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 205, + 482, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 482, + 270 + ], + "type": "text", + "content": "26. Lugasri, C., Tang, J., Nash, H., McClanahan, C., Uboweja, E., Hays, M., Zhang, F., Chang, C.L., Yong, M., Lee, J., Chang, W.T., Hua, W., Georg, M., Grundmann, M.: Mediapipe: A framework for perceiving and processing reality. In: Third Workshop on Computer Vision for AR/VR at IEEE Computer Vision and Pattern Recognition (CVPR) 2019 (2019), https://mixedreality.cs.cornell.edu/s/NewTitle_May1_MediaPipe_CVPR_CV4ARVR_Workshop_2019.pdf" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 270, + 482, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 304 + ], + "type": "text", + "content": "27. Massaro, D., Cohen, M., Tabain, M., Beskow, J., Clark, R.: Animated speech: Research progress and applications. Audiovisual Speech Processing (01 2001). https://doi.org/10.1017/cbo9780511843891.014" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 304, + 482, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 482, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 482, + 326 + ], + "type": "text", + "content": "28. Muralikrishnan, S., Huang, C.H.P., Ceylan, D., Mitra, N.J.: Bliss: Bootstrapped linear shape space (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 326, + 482, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 326, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 482, + 369 + ], + "type": "text", + "content": "29. Nocentini, F., Ferrari, C., Berretti, S.: Learning landmarks motion from speech for speaker-agnostic 3d talking heads generation. In: Foresti, G.L., Fusiello, A., Hancock, E. (eds.) Image Analysis and Processing - ICIAP 2023. pp. 340-351. Springer Nature Switzerland, Cham (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 369, + 482, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 482, + 413 + ], + "type": "text", + "content": "30. Peng, Z., Luo, Y., Shi, Y., Xu, H., Zhu, X., Liu, H., He, J., Fan, Z.: Selftalk: A self-supervised commutative training diagram to comprehend 3d talking faces. In: Proceedings of the 31st ACM International Conference on Multimedia. p. 5292-5301 (2023). https://doi.org/10.1145/3581783.3611734" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 413, + 482, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 482, + 446 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 482, + 446 + ], + "type": "text", + "content": "31. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems (NeurIPS) 30 (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 446, + 482, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 446, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 446, + 482, + 479 + ], + "type": "text", + "content": "32. Qin, D., Saito, J., Aigerman, N., Thibault, G., Komura, T.: Neural face rigging for animating and retargeting facial meshes in the wild. In: SIGGRAPH 2023 Conference Papers (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 479, + 482, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 482, + 512 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 482, + 512 + ], + "type": "text", + "content": "33. Ranjan, A., Bolkart, T., Sanyal, S., Black, M.J.: Generating 3D faces using convolutional mesh autoencoders. In: European Conference on Computer Vision (ECCV). pp. 725-741 (2018), http://coma.is.tue.mpg.de/" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 512, + 482, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 512, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 512, + 482, + 555 + ], + "type": "text", + "content": "34. Richard, A., Zollhöfer, M., Wen, Y., de la Torre, F., Sheikh, Y.: Meshtalk: 3d face animation from speech using cross-modality disentanglement. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 1173-1182 (October 2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 555, + 482, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 555, + 482, + 600 + ], + "spans": [ + { + "bbox": [ + 130, + 555, + 482, + 600 + ], + "type": "text", + "content": "35. Savran, A., Alyüz, N., Dibeklioglu, H., Çeliktutan, O., Gökberk, B., Sankur, B., Akarun, L.: Bosphorus database for 3d face analysis. In: Schouten, B., Juul, N.C., Drygajlo, A., Tistarelli, M. (eds.) Biometrics and Identity Management. pp. 47-56. Springer Berlin Heidelberg, Berlin, Heidelberg (2008)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 600, + 482, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 600, + 482, + 643 + ], + "spans": [ + { + "bbox": [ + 130, + 600, + 482, + 643 + ], + "type": "text", + "content": "36. Schneider, S., Baevski, A., Collobert, R., Auli, M.: wav2vec: Unsupervised pretraining for speech recognition. In: Interspeech 2019. p. 3465-3469. ISCA (Sep 2019). https://doi.org/10.21437/Interspeech.2019-1873, https://www.iscaspeech.org/archive/interspeech_2019/schneider19_interspeech.html" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 643, + 482, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 643, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 643, + 482, + 666 + ], + "type": "text", + "content": "37. Sharp, N., Attaiki, S., Crane, K., Ovsjanikov, M.: Diffusionnet: Discretization ag- nostic learning on surfaces. ACM Trans. Graph. 01(1) (2022)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 228, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 228, + 91, + 447, + 102 + ], + "type": "text", + "content": "ScanTalk: 3D Talking Heads from Unregistered Scans" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 632 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 133, + 117, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 481, + 171 + ], + "type": "text", + "content": "38. Stan, S., Haque, K.I., Yumak, Z.: Facediffuser: Speech-driven 3d facial animation synthesis using diffusion. In: ACM SIGGRAPH Conference on Motion, Interaction and Games (MIG '23), November 15-17, 2023, Rennes, France. ACM, New York, NY, USA (2023). https://doi.org/10.1145/3623264.3624447, https://doi.org/10.1145/3623264.3624447" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 172, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 481, + 194 + ], + "type": "text", + "content": "39. Thambiraja, B., Aliakbarian, S., Cosker, D., Thies, J.: 3diface: Diffusion-based speech-driven 3d facial animation and editing (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 194, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 194, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 133, + 194, + 481, + 237 + ], + "type": "text", + "content": "40. Thambiraja, B., Habibie, I., Aliakbarian, S., Cosker, D., Theobalt, C., Thies, J.: Imitator: Personalized speech-driven 3d facial animation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 20621-20631 (October 2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 238, + 481, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 238, + 481, + 292 + ], + "spans": [ + { + "bbox": [ + 133, + 238, + 481, + 292 + ], + "type": "text", + "content": "41. Wang, A., Emmi, M., Faloutsos, P.: Assembling an expressive facial animation system. In: Proceedings of the 2007 ACM SIGGRAPH Symposium on Video Games. p. 21-26. Sandbox '07, Association for Computing Machinery, New York, NY, USA (2007). https://doi.org/10.1145/1274940.1274947, https://doi.org/10.1145/1274940.1274947" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 293, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 293, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 133, + 293, + 481, + 335 + ], + "type": "text", + "content": "42. Wang, J., Zhao, Y., Liu, L., Xu, T.S., Li, Q., Li, S.: Emotional talking head generation based on memory-sharing and attention-augmented networks. ArXiv abs/2306.03594 (2023), https://api_semanticscholar.org/CorpusID: 259089214" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 336, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 481, + 357 + ], + "type": "text", + "content": "43. Wang, S., Li, L., Ding, Y., Yu, X.: One-shot talking face generation from single-speaker audio-visual correlation learning. In: AAAI 2022 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 358, + 481, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 358, + 481, + 434 + ], + "spans": [ + { + "bbox": [ + 133, + 358, + 481, + 434 + ], + "type": "text", + "content": "44. Wuu, C.h., Zheng, N., Ardisson, S., Bali, R., Belko, D., Brockmeyer, E., Evans, L., Godisart, T., Ha, H., Huang, X., Hypes, A., Koska, T., Krenn, S., Lombardi, S., Luo, X., McPhail, K., Millerschoen, L., Perdoch, M., Pitts, M., Richard, A., Saragih, J., Saragih, J., Shiratori, T., Simon, T., Stewart, M., Trimble, A., Weng, X., Whitewolf, D., Wu, C., Yu, S.I., Sheikh, Y.: Multiface: A dataset for neural face rendering. In: arXiv (2022). https://doi.org/10.48550/ARXIV.2207.11243, https://arxiv.org/abs/2207.11243" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 435, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 435, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 133, + 435, + 481, + 479 + ], + "type": "text", + "content": "45. Xing, J., Xia, M., Zhang, Y., Cun, X., Wang, J., Wong, T.T.: Codetalker: Speechdriven 3d facial animation with discrete motion prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12780-12790 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 479, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 479, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 133, + 479, + 481, + 522 + ], + "type": "text", + "content": "46. Xu, Y., Feng, A.W., Marsella, S., Shapiro, A.: A practical and configurable lip sync method for games. In: Proceedings of Motion on Games. p. 131-140. MIG '13, Association for Computing Machinery, New York, NY, USA (2013). https://doi.org/10.1145/2522628.2522904, https://doi.org/10.1145/2522628.2522904" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 523, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 523, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 133, + 523, + 481, + 544 + ], + "type": "text", + "content": "47. Yi, H., Liang, H., Liu, Y., Cao, Q., Wen, Y., Bolkart, T., Tao, D., Black, M.J.: Generating holistic 3d human motion from speech. In: CVPR (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 545, + 481, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 545, + 481, + 587 + ], + "spans": [ + { + "bbox": [ + 133, + 545, + 481, + 587 + ], + "type": "text", + "content": "48. Yin, L., Wei, X., Sun, Y., Wang, J., Rosato, M.: A 3d facial expression database for facial behavior research. In: 7th International Conference on Automatic Face and Gesture Recognition (FGR06). pp. 211-216 (2006). https://doi.org/10.1109/FGR.2006.6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 588, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 588, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 133, + 588, + 481, + 632 + ], + "type": "text", + "content": "49. Zhong, W., Fang, C., Cai, Y., Wei, P., Zhao, G., Lin, L., Li, G.: Identity-preserving talking face generation with landmark and appearance priors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9729-9738 (June 2023)" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 241, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 241, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 241, + 101 + ], + "type": "text", + "content": "F. Nocentini et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/ScatterFormer_ Efficient Voxel Transformer with Scattered Linear Attention/1045fb37-bbfd-4adc-8de1-f127bb3078ca_content_list.json b/2024/ScatterFormer_ Efficient Voxel Transformer with Scattered Linear Attention/1045fb37-bbfd-4adc-8de1-f127bb3078ca_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6da2b720f112372ea0436d3e8c9598b0c50ce9b7 --- /dev/null +++ b/2024/ScatterFormer_ Efficient Voxel Transformer with Scattered Linear Attention/1045fb37-bbfd-4adc-8de1-f127bb3078ca_content_list.json @@ -0,0 +1,1631 @@ +[ + { + "type": "text", + "text": "ScatterFormer: Efficient Voxel Transformer with Scattered Linear Attention", + "text_level": 1, + "bbox": [ + 225, + 140, + 779, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chenhang He $^{1}$ , Ruihuang Li $^{1,2}$ , Guowen Zhang $^{1}$ , and Lei Zhang $^{1,2}$", + "bbox": [ + 261, + 212, + 743, + 229 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 The Hong Kong Polytechnic University, Hong Kong SAR, China", + "bbox": [ + 279, + 239, + 723, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ OPPO Research, Shenzhen, China chenhang.he@polyu.edu.hk csrhli@comp.polyu.edu.hk guowen.zhang@connect.polyu.edu.hk cslzhang@comp.polyu.edu.hk", + "bbox": [ + 374, + 253, + 630, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Window-based transformers excel in large-scale point cloud understanding by capturing context-aware representations with affordable attention computation in a more localized manner. However, the sparse nature of point clouds leads to a significant variance in the number of voxels per window. Existing methods group the voxels in each window into fixed-length sequences through extensive sorting and padding operations, resulting in a non-negligible computational and memory overhead. In this paper, we introduce ScatterFormer, which to the best of our knowledge, is the first to directly apply attention to voxels across different windows as a single sequence. The key of ScatterFormer is a Scattered Linear Attention (SLA) module, which leverages the pre-computation of key-value pairs in linear attention to enable parallel computation on the variable-length voxel sequences divided by windows. Leveraging the hierarchical structure of GPUs and shared memory, we propose a chunkwise algorithm that reduces the SLA module's latency to less than 1 millisecond on moderate GPUs. Furthermore, we develop a cross-window interaction module that improves the locality and connectivity of voxel features across different windows, eliminating the need for extensive window shifting. Our proposed ScatterFormer demonstrates 73.8 mAP (L2) on the Waymo Open Dataset and 72.4 NDS on the NuScenes dataset, running at an outstanding detection rate of 23 FPS. The code is available at https://github.com/skyhehe123/ScatterFormer.", + "bbox": [ + 261, + 367, + 743, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: 3D Object Detection $\\cdot$ Voxel Transformer", + "bbox": [ + 261, + 685, + 630, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 729, + 375, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the field of 3D object detection, the use of point clouds has become increasingly popular, especially for providing accurate and reliable perception results in autonomous systems. Unlike image data, point clouds obtained from LiDAR are often sparse and nonuniformly distributed, with varying density depending on their distances from the sensor. Earlier approaches utilize point-cloud operators", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/49d5b409f15b627611ec4e71b2a9f0cabab4fb09cb602404b2db8501d07c8322.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 315, + 143, + 687, + 257 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5adf862f1d88200a3e1cd0f498880ed5403aaf17d034e31c04402a7a5c650d03.jpg", + "image_caption": [ + "Fig. 1: Illustration of (a) group-based attention and (b) our scattered linear attention on variable-length sequences. The square in different colors represent voxels in different windows and the square with a cross represents the padding voxel." + ], + "image_footnote": [], + "bbox": [ + 295, + 268, + 702, + 328 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "like PointNet++ [31] for feature extraction in continuous space [37]. Some approaches [5,9,16,18,49,50,52,60] have transformed point clouds into voxel grids, which are then efficiently processed by a sparse convolutional neural network (SpCNN) [49].", + "bbox": [ + 212, + 383, + 784, + 445 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, the success of vision transformers [7, 21] has motivated a number of attention-based methods to process indoor point clouds [12, 22, 24, 25, 27, 57]. Inspired by SwinTransformer [21], various studies [8, 14, 23, 47] have advanced the application of window-based Transformers in large-scale 3D detection tasks within outdoor environments, achieving outstanding performance and highlighting their potential as alternatives to SpCNN. However, due to the inherent sparsity of point clouds, the number of features grouped by windows can vary significantly, hampering the parallelism in attention computation. To resolve this issue, SST [8] and SWFormer [40] group the voxels within a window into different batches and manages the attention computation in a hybrid serial-parallel manner. More effectively, DSVT [47] alternatively sort the voxels within a window from different axes and partition them into sequences of fixed length, which allow parallel computation on sparse voxels. Albeit effective, these group-based methods, as shown in Figure 1(a), require extensive sorting and padding operations, incurring substantial memory and computational overhead.", + "bbox": [ + 212, + 446, + 785, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we delve into the window-based voxel transformer where the voxels grouped by windows form variable-length sequences $\\{X_1\\in \\mathbb{R}^{n_1\\times d},X_2\\in$ $\\mathbb{R}^{n_2\\times d},\\dots,X_k\\in \\mathbb{R}^{n_k\\times d}\\}$ . Their corresponding self-attention matrices $\\{A_{1}\\in$ $\\mathbb{R}^{n_1\\times n_1}$ $A_{2}\\in \\mathbb{R}^{n_{2}\\times n_{2}}$ ..., $A_{k}\\in \\mathbb{R}^{n_{k}\\times n_{k}}\\}$ occupy irregular memory spaces, making it a challenge to perform attention in parallel for the voxels of the entire scene. Recent efforts on linear attention [1,13,17,48] have emerged a promising alternative to traditional attention. By approximating softmax operation with the kernel function, i.e., $\\phi (Q)\\cdot \\phi (K)^{\\top}\\approx \\mathrm{softmax}(QK^{\\top})$ , we can change the computation order from $(Q\\cdot K^{\\top})\\cdot V$ to $Q\\cdot (K^{\\top}\\cdot V)$ . This not only results in linear complexity but also yields a compressed hidden state $S\\in \\mathbb{R}^{d\\times d}$ that is invariant to the sequence length. Another attractive property of linear attention is that it", + "bbox": [ + 212, + 674, + 785, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "C. He et al.", + "bbox": [ + 271, + 114, + 351, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "can be converted into a \"recurrent\" form:", + "bbox": [ + 215, + 146, + 511, + 160 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\no _ {t} = q _ {t} S _ {t}; \\quad S _ {t} = S _ {t - 1} + k _ {t} ^ {\\mathsf {T}} v _ {t} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 397, + 176, + 785, + 195 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the sequences can be divided into non-overlapping chunks, $\\{q_j,k_j,v_j\\}_{j = 1:t}$ . The output can be calculated by scanning over the sequence, based on an updated hidden state $S_{t}$ . This enables the use of chunk-level computation to address the inability to parallelize at the sequence level due to varying lengths.", + "bbox": [ + 212, + 210, + 782, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inspired by this, we introduce the Scattered Linear Attention (SLA) module, which accommodates linear attention in a window-based voxel transformer. As shown in Figure 1(b), the SLA module treats the voxels of the entire scene into a single sequence and processes them directly without padding voxels. Using the recurrent form of linear attention, we develop an I/O-aware algorithm to further optimize matrix multiplication on voxel sequences. Specifically, we divide the voxel sequence into multiple chunks and loaded them into the shared memory (SRAM) of the GPU. The computation of the hidden state matrix in each window is then achieved through a series of chunk-wise matrix multiplications and cumulative sums. This optimized implementation significantly reduces I/O overhead and memory usage, resulting in extremely fast and memory-efficient attention computation.", + "bbox": [ + 212, + 273, + 784, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In addition, current window-based transformer models use window shifting to propagate the information across the windows. After performing an instruction-level analysis of existing implementations, we observe that the voxel permutation operations caused by window shifting consume significant computational overhead. To address this issue, we propose a cross-window interaction (CWI) module, which is composed of depth-wise convolutions with small 2D kernels and lengthy 1D kernels that allow the voxel features in each window to fully interact with the features in other windows. As a result, the proposed CWI module can improve both the locality and connectivity of the voxel features while requiring minimal computational effort.", + "bbox": [ + 212, + 457, + 784, + 608 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Building upon the SLA and CWI modules, we propose the ScatterFormer, an innovative voxel transformer for large-scale point cloud understanding. Our experiments show that ScatterFormer achieves linear complexity without compromising accuracy. It achieves superior results compared to the state-of-the-art model DSVT [47]. The latency of ScatterFormer is significantly lower than that of transformer-based detectors [8, 14, 47], which is comparable to that of sparse convolution-based detectors [52].", + "bbox": [ + 212, + 609, + 784, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In conclusion, we delve into attention on voxels grouped by windows, highlighting the challenges in memory allocation and computation for variable-length sequences. Then, we introduce an SLA module, which can directly process the voxels grouped by windows by facilitating the linear attention formula. A chunkwise matrix multiplication algorithm is proposed to further accelerate the attention computation in SLA. Finally, we present ScatterFormer, which has linear complexity and can efficiently process large-scale LiDAR scenes, achieving better accuracy with lower latency compared to existing transformer-based detectors.", + "bbox": [ + 212, + 719, + 784, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "ScatterFormer", + "bbox": [ + 633, + 114, + 730, + 127 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 127 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 143, + 388, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.1 Point Cloud-based 3D Object Detection", + "text_level": 1, + "bbox": [ + 215, + 180, + 591, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "There exist two primary point-cloud representations in 3D object detection, i.e., point-based and voxel-based ones. In point-based methods [28, 29, 37, 39, 55], point clouds are first passed through a point-based backbone network [30], in which the points are gradually sampled and features are learned by point cloud operators. F-Pointnet [29] first employs PoinNet to detect 3D objects based on the frustums lifted by 2D proposals. PointRCNN [37] direct generates 3D proposals from point-based features with foreground segmentation. VoteNet [28] clusters objects from the surface points using a deep Hough voting method. Other point operators based on the point graph [39, 55] and the range view [10] have also been developed for point cloud processing.", + "bbox": [ + 212, + 209, + 785, + 359 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Point-based methods are primarily limited by their inference efficiency and the absence of contextual features in a continuous space. On the other hand, voxel-based approaches [5, 16, 18, 49, 52, 58, 60] transform the entire point clouds into regular grids through voxelization, showcasing superior efficiency and context representation. VoxelNet [60] proposes a voxel feature encoding (VFE) module and combines it with 3D convolutions to extract voxel features in an end-to-end manner. SECOND [49] optimizes 3D convolution for sparse data, resulting in a significant reduction in both time and memory. PointPillars [18] demonstrates an efficient model by stacking voxels into vertical columns, subsequently processing them with naive 2D convolutions. There are also efforts [16, 35] that explore hybrid representations of point-based and voxel-based networks, demonstrating a better trade-off between speed and accuracy. However, a common limitation across these methods is their reliance on small convolution kernels with restricted receptive fields, making them less adept at capturing the global context for 3D object detection.", + "bbox": [ + 212, + 361, + 785, + 587 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this paper, we shed light on the voxel-based methods and introduce a Scattered Linear Attention module that enables dynamic modeling among voxels within each window with linear complexity, thereby efficiently extracting window-based contextual features.", + "bbox": [ + 212, + 588, + 785, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Transformer on Point Cloud", + "text_level": 1, + "bbox": [ + 215, + 675, + 496, + 689 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inspired by the significant success of self-attention in NLP [46] and CV [7, 45], Transformers have been adapted for 3D vision due to their ability to capture long-range dependencies. The Point Transformers [27, 57] employ attention to modulate point clouds for classification and segmentation tasks. PCT [12] presents an optimized offset attention module, which, when combined with the implicit Laplacian operator and normal estimation, becomes more adept at point-cloud processing. Some methods like [8, 22] opt for voxels or key points to optimize latency. 3DETR [25] proposes a query-based 3D object detection scheme. CT3D [33] enhances the region-based network using a channel-wise", + "bbox": [ + 212, + 703, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 230, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "C. He et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b774ed792faa3e7e125652a569fc710665377ac999a8cc3e7153910cf7ae9407.jpg", + "image_caption": [ + "Fig. 2: The macro design of ScatterFormer. The backbone comprises a Conditional Position Encoding (CPE) and six transformer blocks. Each block is composed by a Scattered Linear Attention (SLA) module, a Cross-Window Interaction (CWI) module, and a Feed-Forward Network (FFN)." + ], + "image_footnote": [], + "bbox": [ + 218, + 148, + 782, + 252 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Transformer framework. To achieve context-rich representations, several studies [8, 14, 23, 24, 47] have integrated the attention module into point- or voxel-based encoders. For instance, VoTr [24] utilizes dilated attention for expanded receptive fields; VoxSet [14] applies set attention for extracting point-based features in set-to-set translation; SST [8] employs local attention with shifted windows; and OcTr [59] adopts an Octree-based attention for efficient hierarchical context learning. Nevertheless, how to efficiently leverage global context from attention remains a challenge due to the intrinsic sparsity of point clouds. DSVT [47] and FlatFormer [23] group the voxels within each window into a series of fixed-length voxel sets, thus extracting the features in a fully parallel manner. Recently, a group-free state-space model [54] was proposed to directly process voxels as a sequence. However, these approaches more or less lose the spatial proximity and incur extensive computational overhead in grouping and sorting the voxels.", + "bbox": [ + 212, + 332, + 787, + 529 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 551, + 330, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overall architecture of our proposed ScatterFormer is depicted in Figure 2. It begins with the input point clouds, which are voxelized and transformed into high-dimensional embeddings using a VFE layer [60]. These embeddings are then processed through Conditional Positional Encoding (CPE) using a shallow convolutional network [6]. The encoded features enter the ScatterFormer backbone, consisting of six ScatterFormer blocks. Each block includes a Scattered Linear Attention (SLA) module, a Cross-Window Interaction (CWI) module, and a Feed-Forward Network (FFN), interspersed with Batch Normalization layers and skip connections. After three ScatterFormer blocks, the voxel features are downsampled via a sparse convolutional layer. The downsampled features are then converted into pillar features [47], generating compact BEV features for bounding-box prediction. ScatterFormer ScatterFormer stands out by not requiring voxel features to be organized into fixed-length sets [8, 23, 47], enabling flexible attention computation across windows. Additionally, the CWI module obviates the need for window shifting. These innovations allow ScatterFormer to avoid unnecessary memory allocation and permutation operations, achieving high efficiency comparable to CNN-based models.", + "bbox": [ + 212, + 582, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "ScatterFormer", + "bbox": [ + 633, + 114, + 730, + 127 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Linear Attention", + "text_level": 1, + "bbox": [ + 215, + 146, + 401, + 160 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We begin by revisiting the self-attention introduced by [46]. Given an input matrix $x \\in \\mathbb{R}^{N \\times d}$ , where $N$ denotes the number of tokens and $d$ denotes the dimensionality. Self-attention first linearly projects the input to queries, keys, and values using weight matrices $W_{Q}, W_{K}$ , and $W_{V}$ , such that $Q = xW_{Q}, K = xW_{K}, V = xW_{V}$ . Then it computes a Softmax-based attention map based on queries and keys i.e., $A(Q, K) = \\mathrm{Softmax}(QK^{\\top}) / \\sqrt{d}$ . The output of self-attention is defined as the weighted sum of $N$ values with the weights corresponding to the attention map, i.e., $O = A(Q, K)V$ . This approach involves calculating the similarity for all query-key pairs, which yields a computational complexity of $O(N^{2})$ .", + "bbox": [ + 212, + 172, + 789, + 323 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Linear attention [1, 17] offers a notable alternative to self-attention by significantly reducing its computational complexity from $O(N^2)$ to $O(N)$ . This efficiency is achieved through the application of kernel functions on the query $(Q)$ and key $(K)$ matrices, which effectively approximates the original attention map without relying on Softmax, i.e., $A(Q, K) = \\phi(Q)\\phi(K)^{\\mathsf{T}}$ . Taking advantage of the associative property of matrix multiplication, the computation shifts from $(\\phi(Q)\\phi(K)^{\\mathsf{T}})V$ to $\\phi(Q)(\\phi(K)^{\\mathsf{T}}V)$ . This modification leads to a computational complexity proportional to the number of tokens, maintaining an order of $O(N)$ .", + "bbox": [ + 212, + 324, + 787, + 446 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Scattered Linear Attention (SLA)", + "text_level": 1, + "bbox": [ + 215, + 468, + 540, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Instead of organizing the voxels within a window into fixed-length subsets, we arrange the voxels of the entire scene into a single flattened matrix, as illustrated in Figure 3(a). These voxels are ordered based on their window coordinates, ensuring that voxels from the same window form a sub-matrix in contiguous memory. Initially, we use a shared projection layer to map all the voxels in the scene to query, key, and value representations, denoted $Q$ , $K$ , $V$ . Subsequently, we perform attention computations on these submatrices separately, resulting in the scattered linear attention (SLA) formula as follows:", + "bbox": [ + 212, + 494, + 787, + 617 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S L A} (Q, K, V) = \\operatorname {C o n c a t} \\left[ \\operatorname {L A} \\left(Q ^ {j}, K ^ {j}, V ^ {j}\\right) \\right] _ {j = 1: M}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 331, + 628, + 784, + 646 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $M$ is the number of non-empty window in the current scene and LA is the linear attention formula used in [17]. Based on this formula, the output voxel features $O^j$ in the $j^{\\mathrm{th}}$ window can be defined as:", + "bbox": [ + 212, + 657, + 787, + 705 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {L A}: O ^ {j} = \\frac {\\phi (Q) \\sum_ {i = 1} ^ {m ^ {j}} \\phi \\left(K _ {i}\\right) ^ {\\top} V _ {i}}{\\phi (Q) \\sum_ {i = 1} ^ {m ^ {j}} \\phi \\left(K _ {i}\\right) ^ {\\top}} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 385, + 717, + 784, + 757 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $m^j$ is the number of voxel in the window and $\\phi (x)$ is the kernel function.", + "bbox": [ + 215, + 768, + 785, + 784 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Hardware Efficient Implementation. It should be noted that the implementation of Equation 2 is not straightforward, as the sub-matrices have different numbers of rows. One naive implementation is to cache the \"kv\" matrix of each", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "C. He et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5816bbdd784f56883d37f1190b2007a02448843582dd37f1427736e0a3847fbe.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 217, + 142, + 540, + 292 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/61611b9e5d4ca7f85444d115979bd35c71ff9bb30b2578de2216e84f4a722b90.jpg", + "image_caption": [ + "(b)", + "Fig. 3: (a) The voxel features are sorted by windows and flattened to a single matrix, which is further partitioned into multiple chunks of uniform length and sent to SLA kernel for attention computation. (b) In SLA, we first allocate individual threads to each window, each of them iterates over all corresponding key and value chunks, calculates, and accumulates their products to obtain the hidden state matrix of each window. Then, multiple threads are allocated to these query chunks $q_{i} \\in Q$ , calculating the chunk-wise output by multiplying $q_{i}$ with the corresponding hidden state matrix." + ], + "image_footnote": [], + "bbox": [ + 550, + 146, + 779, + 308 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "voxel and apply $scatter^3$ operator to accumulate them within each window. However, this would consume a high memory overhead and IO latency. Considering that GPUs feature a memory hierarchy that includes larger, slower global GPU memory (high-bandwidth memory; HBM) and smaller, faster shared memory (SRAM), optimal utilization of SRAM to minimize HBM I/O costs can lead to significant speed-ups. Based on this, we partition the flattened $Q, K, V$ into multiple chunks, load them from slow HBM to fast SRAM. Specifically, for each window, we allocate an single thread. Each thread iterates over all the key and value chunks, calculates, and accumulates their products to obtain the hidden state matrix of the current window. After calculating the hidden state matrix in window, we assign individual thread to each query chunk $q_i \\in Q^j$ and calculate the chunk-wise output can then be following Equation 3. Figure 3(b) illustrates this chunk-wise matrix multiplication approach. As can be seen, it not only avoids outputting large matrices but also allows for more flexible processing of variable-length sequences.", + "bbox": [ + 212, + 443, + 787, + 669 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In our implementation, we use Triton [44] to perform chunk-wise matrix multiplication. As illustrated in Figure 4, it outperforms the naive scattered-based operation, demonstrating improved speed and reduced memory usage when computing attention for extremely long sequences.", + "bbox": [ + 212, + 670, + 787, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Cross Window Interaction", + "text_level": 1, + "bbox": [ + 214, + 750, + 478, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Traditional window-based transformers [8,40,47] utilize window shifting for interwindow connection. For point cloud sequences, this requires recalculating window coordinates and rearranging the voxels. In DSVT, the computation of voxel", + "bbox": [ + 212, + 772, + 787, + 818 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "ScatterFormer", + "bbox": [ + 633, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "3 https://github.com/rustyls/pytorch_scatter", + "bbox": [ + 217, + 823, + 532, + 840 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e19d811d4b8025e814593bf484992f98c2b31d8126b02a975f810475fc70b9bb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 150, + 718, + 277 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/80784b1ee3048e8ecee1c136574d82adfefdd08e582f1395abe5b57500be5941.jpg", + "image_caption": [ + "Fig. 4: Comparison between between scatter-based implementation (yellow) and our chunk-wise implementation (blue)." + ], + "image_footnote": [], + "bbox": [ + 282, + 291, + 718, + 417 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "indices in subsets and the rearrangement of voxel matrices are particularly time consuming, accounting for $24\\%$ of the total backbone latency, as shown in Figure 5. To alleviate this unnecessary computational overhead from window shifting, we introduce the Cross-Window Interaction (CWI) module, which employs convolutions with large kernel to blend voxel features across windows.", + "bbox": [ + 212, + 484, + 782, + 559 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We adhere to the design principles of the Inception architecture [42], where the input features are divided along the channel dimension and processed through multiple branches using group-wise convolutions. Instead of employing a single large kernel for convolution, we adopt more efficient 1D kernels along different axes for feature aggregation. Specifically, we apply a $(S_h + 1) \\times 1 \\times 1$ convolution in one branch and a $1 \\times (S_w + 1) \\times 1$ convolution in another branch, where $(S_w, S_h)$ denotes the window size. Additionally, we incorporate a standard $3 \\times 3 \\times 3$ convolution in one branch to enhance locality, while another branch performs an identity mapping. Given input $X$ of channel $4c$ , the output $X'$ of the Cross Window Interaction module can be written as:", + "bbox": [ + 212, + 561, + 787, + 710 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nX ^ {\\prime} = \\operatorname {C o n c a t} \\left(X _ {k}, X _ {w}, X _ {h}, X ^ {3 c: 4 c}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 375, + 722, + 785, + 741 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 215, + 752, + 264, + 765 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nX _ {k} = \\operatorname {D W C o n v} _ {\\left(S _ {h} + 1\\right) \\times 1 \\times 1} \\left(X ^ {: c}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 779, + 784, + 796 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nX _ {w} = \\operatorname {D W C o n v} _ {1 \\times \\left(S _ {w} + 1\\right) \\times 1} \\left(X ^ {c: 2 c}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 797, + 784, + 816 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nX _ {h} = \\operatorname {D W C o n v} _ {3 \\times 3 \\times 3} \\left(X ^ {2 c: 3 c}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 819, + 784, + 837 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "C. He et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b80d62c5b3034d81653c564b43028cf806b236d6d2361938a5c9c4b42f7a6cee.jpg", + "image_caption": [ + "Fig. 5: Runtime decomposition in DSVT [47] and ScatterFormer backbones. The offsets generation means computing the starting address of each chunk in the flattened matrix." + ], + "image_footnote": [], + "bbox": [ + 313, + 143, + 498, + 297 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/43750b18e59eb1fa21fbec4d709e83e507896764ae1536406d913eae1fddbe25.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 511, + 143, + 694, + 297 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In our experiments, we found that this axis-decomposed large kernel design can achieve a better precision-latency trade-off. With these lengthy kernels, the voxel features can be efficiently blended among different windows.", + "bbox": [ + 212, + 352, + 784, + 398 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.4 Detection Head and Loss", + "text_level": 1, + "bbox": [ + 214, + 416, + 470, + 430 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To complement existing detection heads, ScatterFormer generates dense BEV feature maps from sparse voxel representations by placing them back to their spatial locations and filling the unoccupied positions with zeros. For the BEV network, we simply follow [53] by adopting a cascaded CNN with convolutional blocks at multiple levels of strides. In terms of detection head and loss function, we follow the design identical to that used in DSVT [47]. We train the detection model with heatmap estimation, bounding box regression, and incorporate an IoU loss for confidence calibration. On the nuScenes dataset, the detection head follows the architecture used in Transfusion [2].", + "bbox": [ + 212, + 438, + 787, + 575 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 594, + 375, + 611 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Datasets and Evaluation Metrics", + "text_level": 1, + "bbox": [ + 214, + 622, + 532, + 636 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Waymo Open Dataset (WOD). This dataset contains 230,000 annotated samples split into 160,000 for training, 40,000 for validation, and 30,000 for testing. It uses two metrics for 3D object detection: mean average precision (mAP) and mAP weighted by heading accuracy (mAPH), further categorized into Level 1 (L1) for objects detected by more than five LiDAR points and Level 2 (L2) for those detected with at least one point.", + "bbox": [ + 212, + 643, + 787, + 734 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "NuScenes. This dataset comprises 40,000 annotated samples, with 28,000 for training, 6,000 for validation, and 6,000 for testing. On this dataset, the model performance is measured by mean average precision (mAP) across multiple distance thresholds (0.5, 1, 2, and 4 meters) and the nuScenes detection score (NDS), which combines mAP with a weighted sum of five additional metrics assessing true positive predictions in translation, scale, orientation, velocity, and attribute accuracy.", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "ScatterFormer", + "bbox": [ + 633, + 114, + 730, + 127 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/1fd59f3f4ba6c1f1f382f901e3edfcb8ab732f8e1bb1ff93745ab0685f48f00b.jpg", + "image_caption": [ + "Fig. 6: The detection performance (mAPH/L2) vs. speed (FPS) of different methods on Waymo validation set. The speeds are measured on an NVIDIA A100 GPU." + ], + "image_footnote": [], + "bbox": [ + 289, + 148, + 707, + 388 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 438, + 457, + 454 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Our approach is implemented using the open-source framework OpenPCDet [43]. To construct ScatterFormer, we set the voxel size to $(0.32\\mathrm{m}, 0.32\\mathrm{m}, 0.1875\\mathrm{m})$ for the Waymo dataset and $(0.3\\mathrm{m}, 0.3\\mathrm{m}, 8\\mathrm{m})$ for the NuScenes dataset. The window sizes $(S_w, S_h)$ for the two datasets are set to $(12, 12)$ and $(20, 20)$ , respectively. We stack six building blocks for the backbone network. We configure our attention module to have 4 heads with a dimensionality of 128. ScatterFormer is trained for 24 epochs with a learning rate of 0.006 on Waymo Dataset and 20 epochs with a learning rate of 0.004 on NuScenes Dataset. In the last 4 epochs, we disabled the data augmentation strategy of ground-truth sampling. The model is trained on 8 RTX A6000 GPUs with a batch size of 32. Other settings for training and inference adhere strictly to DSVT [47].", + "bbox": [ + 212, + 462, + 787, + 628 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Comparison with State-of-the-Arts", + "text_level": 1, + "bbox": [ + 215, + 648, + 552, + 666 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Results on Waymo Open Dataset (WOD). We conduct a detailed comparison of ScatterFormer against the published results on the validation set of the Waymo Open Dataset. In line with standard practices, we distinctly categorize and list the methods utilizing single-frame and multi-frame approaches. To ensure comprehensive coverage, we also include comparisons with methods incorporating long-term temporal modeling, such as those described in [4,15,61]. As demonstrated in Table 1, our single-stage model outperforms most two-stage methods and achieves better performance than state-of-the-art methods such as DSVT [47] and HEDNet [53]. Moreover, our model achieves 76.0 and 76.7 level 2 mAPH on 3 and 4 frame settings, respectively, outperforming previous multiframe methods by a margin of 1.1. Furthermore, as shown in Figure 6,", + "bbox": [ + 212, + 672, + 787, + 843 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "C. He et al.", + "bbox": [ + 271, + 114, + 351, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/dbd41322d5fee2d2d964d0ffac534ba7a10f74057dfb00d9d16321d443125711.jpg", + "table_caption": [ + "Table 1: Performance comparison on the validation set of Waymo Open Dataset. Symbol \\* denotes the methods with temporal modeling, and - means that the result is not available." + ], + "table_footnote": [], + "table_body": "
MethodFramesALL (3D mAPH)Vehicle (AP/APH)Pedestrian (AP/APH)Cyclist (AP/APH)
L1L2L1L2L1L2L1L2
SECOND [49]163.0557.2372.27/71.6963.85/63.3368.70/58.1860.72/51.3160.62/59.2858.34/57.05
PointPillar [18]163.3357.5371.60/71.0063.10/62.5070.60/56.7062.90/50.2064.40/62.3061.90/59.90
IA-SSD [56]164.4858.0870.53/69.6761.55/60.8069.38/58.4760.30/50.7367.67/65.3064.98/62.71
LiDAR R-CNN [20]166.2060.1073.50/73.0064.70/64.2071.20/58.7063.10/51.7068.60/66.9066.10/64.40
RSN [41]1--75.10/74.6066.00/65.5077.80/72.7068.30/63.70--
Part-A2 [38]170.2563.8477.05/76.5168.47/67.9775.24/66.8766.18/58.6268.60/67.3666.13/64.93
Centerpoint [52]1-65.50-- /66.20--/62.60--/67.60
VoTR [24]1--74.95/74.2565.91/65.29----
VoxSeT [14]172.2466.2274.50/74.0365.99/65.5680.03/72.4272.45/65.3971.56/70.2968.95/67.73
SST-1f [8]1--76.22/75.7968.04/67.6481.39/74.0572.82/65.93--
SWFormer-1f [40]1--77.8/77.369.2/68.880.9/72.772.5/64.9--
PV-RCNN [35]169.6363.3377.51/76.8968.98/68.4175.01/65.6566.04/57.6167.81/66.3565.39/63.98
PillarNet [34]174.6068.4379.09/78.5970.92/70.4680.59/74.0172.28/66.1772.29/71.2169.72/68.67
PV-RCNN++ [35]175.2168.6179.10/78.6370.34/69.9180.62/74.6271.86/66.3073.49/72.3870.70/69.62
FlatFormer [23]1-67.2-69.0/68.6-71.5/65.3-68.6/67.5
PillarNext-1f [19]175.7469.7478.40/77.9070.27/69.8182.53/77.1474.90/69.8073.21/72.2070.58/69.62
VoxelNext [5]176.370.178.2/77.769.9/69.481.5/76.373.5/68.676.1/74.973.3/72.2
FSD [9]177.370.879.2/78.870.5/70.182.6/77.373.9/69.177.1/76.074.4/73.3
DSVT-1f [47]178.272.179.7/79.371.4/71.083.7/78.976.1/71.577.5/76.574.6/73.7
HEDNet-1f [53]179.573.481.1/80.673.2/72.784.4/80.076.8/72.678.7/77.775.8/74.9
ScatterFormer (ours)179.773.881.0/80.573.1/72.784.5/79.977.0/72.679.9/78.977.1/76.1
SST-3f [8]3--78.66/78.2169.98/69.5783.81/80.1475.94/72.37--
FlatFormer-3f [23]3-72.0-71.4/71.0-74.5/71.3-74.7/73.7
SWFormer-3f [40]3--79.4/78.971.1/70.682.9/79.074.8/71.1--
PillarNext-3f [19]380.074.580.6/80.172.9/72.485.0/82.178.0/75.278.9/77.976.7/75.7
DSVT-4f [47]481.375.681.8/81.474.1/73.685.6/82.878.6/75.980.4/79.678.1/77.3
ScatterFormer-3f (ours)381.776.082.0/81.475.0/74.185.7/83.279.0/76.180.6/79.578.1/77.7
ScatterFormer-4f (ours)481.976.782.4/81.975.2/74.786.1/83.479.3/76.681.2/80.478.9/78.1
*3D-MAN [51]16--74.53/74.0367.61/67.1471.7/67.762.6/59.0--
*CenterFormer [61]477.073.278.1/77.673.4/72.981.7/78.677.2/74.275.6/74.873.4/72.6
*CenterFormer [61]877.373.778.8/78.374.3/73.882.1/79.377.8/75.075.2/74.473.2/72.3
*MPPNet [4]479.8374.2281.54/81.0674.07/73.6184.56/81.9477.20/74.6777.15/76.5075.01/74.38
*MPPNet [4]1680.4074.8582.74/82.2875.41/74.9684.69/82.2577.43/75.0677.28/76.6675.13/74.52
*MSF [15]480.2074.6281.36/80.8773.81/73.3585.05/82.1077.92/75.1178.40/77.6176.17/75.40
*MSF [15]880.6575.4682.83/82.0175.76/75.3185.24/82.2178.32/75.6178.52/77.7476.32/75.47
", + "bbox": [ + 215, + 199, + 787, + 556 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "our method significantly enhances the detection runtime. This improvement is attributed to our use of Linear Attention, which eliminates the need for extensive voxel sorting and padding operations. Furthermore, as presented in Table 2, ScatterFormer achieves the highest Level 1 and Level 2 mAPH scores in all categories of the Waymo test set, highlighting its superior detection quality and precision. Remarkably, our approach, as a one-shot method, delivers performance that is on par with methods that require temporal modeling. This finding underscores the critical importance of network design and architecture in achieving high-performance architecture.", + "bbox": [ + 212, + 566, + 787, + 702 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Results on NuScenes. We compare ScatterFormer with the previous best performing methods on the nuScenes dataset. As shown in Table 3, ScatterFormer achieves the state-of-the-art performance in terms of val NDS (72.4) and mAP (68.3), surpassing HEDNet [53] by 1.0 and 1.6 respectively. Compared to TransFusion [2], ScatterFormer achieves a significant improvement of $2.8\\%$ in mAP, demonstrating that its backbone network is superior to the sparse convolutional network. ScatterFormer exhibits particularly strong performance on categories like Bus, Bike, Construction Vehicle and Trailer, suggesting its ability to effectively capture and encode detailed contextual features.", + "bbox": [ + 212, + 704, + 787, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "ScatterFormer", + "bbox": [ + 633, + 114, + 730, + 127 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/fcff4ef839ba2f630c72e3792b710fe627801b4b12e06777d3e0d4f02b0080a8.jpg", + "table_caption": [ + "Table 2: Performance comparison on the test set of Waymo Open Dataset. $\\cdot$ means that the result is not available." + ], + "table_footnote": [], + "table_body": "
MethodALL (3D mAPH)Vehicle (AP/APH)Pedestrian (AP/APH)Cyclist (AP/APH)
L1L2L1L2L1L2L1L2
PointPillar [18]--68.1060.1068.00/55.5061.40/50.10--
StarNet [26]--61.0054.5067.80/59.9061.10/54.00--
M3DETR [11]67.161.977.7/77.170.5/70.068.2/58.560.6/52.067.3/65.765.3/63.8
3D-MAN [51]--78.2869.9869.97/65.9863.98/60.26--
PV-RCNN++ [36]75.770.281.6/81.273.9/73.580.4/75.074.1/69.071.9/70.869.3/68.2
CenterPoint [52]77.271.981.1/80.673.4/73.080.5/77.374.6/71.574.6/73.772.2/71.3
RSN [41]--80.3071.6078.90/75.6070.70/67.80--
SST-3f [8]78.372.881.0/80.673.1/72.783.3/79.776.9/73.575.7/74.673.2/72.2
HEDNet [53]79.0573.7783.78/83.3976.33/75.9683.46/78.9877.53/73.2575.86/74.7873.13/72.09
PillarNext-3f [19]79.0074.0983.28/82.8376.18/75.7684.40/81.4478.84/75.9873.77/72.7371.56/70.55
ScatterFormer-4f (ours)80.7175.8485.60/85.1378.34/78.0784.58/81.6479.12/76.3276.20/75.3574.02/73.13
", + "bbox": [ + 217, + 178, + 785, + 308 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/13f4907d846e33a141b22a6f98790cce5416553f711092447b0cc9d85a7f45cc.jpg", + "table_caption": [ + "Table 3: The performance on the validation set of NuScenes." + ], + "table_footnote": [], + "table_body": "
MethodNDSmAPCarTruckBusT.L.C.V.Ped.M.T.BikeT.C.B.R.
CenterPoint [52]66.559.284.957.470.738.116.985.159.042.069.868.3
VoxelNeXt [5]66.760.583.955.570.538.121.184.662.850.069.469.4
TransFusion-L [2]70.165.586.960.873.143.425.287.572.957.377.270.3
PillarNext [19]68.462.285.057.467.635.620.686.868.653.177.369.7
DSVT [47]71.166.487.462.675.942.125.388.274.858.777.870.9
HEDNet [53]71.466.787.760.677.850.728.987.174.356.876.366.9
ScatterFormer (ours)72.468.388.665.479.345.429.188.774.761.578.272.3
", + "bbox": [ + 217, + 344, + 785, + 454 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 467, + 388, + 483 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In this section, we conduct ablation experiments in ScatterFormer using $20\\%$ of the Waymo training and validation data. Each model was trained for 24 epochs and run on default parameters. Table 4 shows four different configurations of ScatterFormer: (a) removing the Scattered Linear Attention module (SLA), (b) removing the Cross Window Interaction (CWI) module, (c) replacing the CWI with Shifted Window (SW) approach, and (d) removing Conditional Position Encoding (CPE) module.", + "bbox": [ + 212, + 492, + 787, + 598 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "As can be seen in Table 4, configuration (a) results in ScatterFormer experiencing a decrease of $(3.1\\%, 3.9\\%, \\text{and} 3.2\\%)$ APH on three categories. This highlights the effectiveness of the SLA, as it successfully captures long-range contexts and facilitates dynamic feature extraction for object detection. In configuration (b), ScatterFormer, which lacks cross-window interaction capability, exhibits a performance drop of $(1.9\\%, 1.0\\%, \\text{and} 2.0\\%)$ APH across the three categories. While the Shifted Window approach in configuration (c) enhances performance, the improvement is not as substantial as that achieved by the proposed CWI module. The results in configuration (d) indicate that CPE contributes to a certain degree of performance enhancement.", + "bbox": [ + 212, + 598, + 787, + 748 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Window Size. In Table 5, We tested the impact of different window sizes on the detection performance of various target types. It can be observed that ScatterFormer is not highly sensitive to window size. Slightly increasing the window size can enhance the detection performance for small targets, indicating that for objects like pedestrians and bicycles, where the point cloud is overly sparse, appropriate contextual information can enhance the recognition of these", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "C. He et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/fdd3990eb923f3fb19e0239113488a5c2ae51894b36edd983dab1d18c1c8685a.jpg", + "table_caption": [ + "Table 4: Ablation experiments on the validation set of Waymo Open Dataset. \"SLA\" and \"CWT\" refer to the proposed Scattered Linear Attention and Cross-Window Interaction modules, respectively. \"CPE\" refers to Conditional Position Embedding. AP and APH scores on LEVEL2 are reported." + ], + "table_footnote": [], + "table_body": "
AblationVeh.Ped.Cyc.
-baseline71.6/71.176.0/70.974.2/73.1
(a)w/o SLA module68.5/68.073.3/67.071.0/69.9
(b)w/o CWI module69.6/69.274.0/69.972.1/71.1
(c)CWI → SW69.9/69.575.0/70.473.8/72.4
(d)w/o CPE70.2/69.873.3/67.272.4/70.5
", + "bbox": [ + 313, + 205, + 686, + 303 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/f503a21c32aeb51147d230bffb9bdc7aa5b0e2ca5cf4f140f44059142530c64e.jpg", + "table_caption": [ + "Table 5: Comparison of model performance using different window sizes at various stages. The APH (L2) scores on different categories are reported." + ], + "table_footnote": [], + "table_body": "
Window SizeRegion SizeVeh. Ped. Cyc.
103.20 m70.2 69.2 71.0
123.84 m71.1 70.9 73.1
144.48 m71.3 69.9 72.5
165.12 m71.0 69.2 72.3
", + "bbox": [ + 348, + 347, + 647, + 430 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "objects. However, excessively large windows can degrade the performance. We speculate that this is because larger windows encompass more tokens, thereby mitigating the focusing power of linear attention.", + "bbox": [ + 212, + 449, + 787, + 494 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Kernel Design in CWI. Table 6 presents the performance under various kernel configurations in the Cross-Window Interaction (CWI) module. Increasing the number of parameters in the kernel significantly enhances performance. For example, using a large $7 \\times 7$ kernel brings notable performance improvements, but also introduces more latency. Additionally, larger kernel strides enhance the exchange rate of tokens across windows. For example, using dilated convolutions in CWI performs slightly better than using non-dilated convolutions. Instead of scaling up large, we decompose large kernel into more efficient 1D kernels, which achieves the optimal latency while keeping comparable performance. As the current Sconv library lacks support for depth-wise convolution, we have developed a modified version to address this limitation. $^4$", + "bbox": [ + 212, + 501, + 787, + 667 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.5 Comparison with Different Linear Attentions", + "text_level": 1, + "bbox": [ + 214, + 688, + 633, + 704 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We further explored the performance of ScatterFormer by conducting experiments with several other linear attention mechanisms, including Gated Linear Attention [32], Cross-covariance Attention (XCA) [1], and Focused Linear Attention [13]. The results are summarized in Table 7. Our findings suggest that the window-based attention framework employed by ScatterFormer is not highly sensitive to the specific form of linear attention used. Notably, Efficient Attention [3] performs slightly worse than XCA [1] in the Vehicle class but better in the", + "bbox": [ + 212, + 710, + 787, + 818 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "ScatterFormer", + "bbox": [ + 633, + 114, + 730, + 127 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_footnote", + "text": "4 https://github.com/skyhehe123/spconvv", + "bbox": [ + 217, + 824, + 514, + 840 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/1beea83c0bef77cd35b6c1d07c97779ddf13b5af87fb74019d59b9e331bb9bfc.jpg", + "table_caption": [ + "Table 6: Performance with various kernel configurations in the Cross-Window Integration (CWI) module. Each configuration is specified by the notation \"KxDy\", where \"Kx\" refers to the kernel size, and \"Dy\" denotes the dilation rate. The APH (L2) scores on different categories are reported." + ], + "table_footnote": [], + "table_body": "
KernelLatencyVeh.Ped.Cyc.
Conv1DK13 × 247ms71.170.973.1
Conv2DK5D149ms69.569.071.1
Conv2DK5D249ms70.770.472.8
Conv2DK7D165ms71.571.072.5
", + "bbox": [ + 352, + 205, + 640, + 287 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/1df3f6596a582e606f11d55510e8a1edd15190e76e9ba2934624709b097801e3.jpg", + "table_caption": [ + "Table 7: Performance with different linear attention designs. The APH (L2) scores on different categories are reported." + ], + "table_footnote": [], + "table_body": "
Linear AttentionVeh.Ped.Cyc.
Efficient Attn [3]71.170.973.1
Gated Linear Attn [32]69.969.072.5
Focused Linear Attn [13]70.469.572.3
XCA [1]71.570.472.8
", + "bbox": [ + 352, + 325, + 640, + 407 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Pedestrian and Cyclist classes. Focused Linear Attention introduces additional computational overhead, while Gated Linear Attention adds more learnable parameters to the model.", + "bbox": [ + 212, + 421, + 785, + 465 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "4.6 Limitations", + "text_level": 1, + "bbox": [ + 214, + 483, + 356, + 497 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "ScatterFormer relies on our customized operators, which have not yet been implemented as plugins in TensorRT. Therefore, deploying ScatterFormer on in-vehicle devices will require additional engineering efforts. Despite this, ScatterFormer is more efficient than current models based on sparse convolution and traditional attention when used on consumer-grade GPU cards. Furthermore, ScatterFormer can be optimized by dynamically partitioning matrices according to different GPU architectures to leverage TensorCore for hardware acceleration.", + "bbox": [ + 212, + 510, + 787, + 617 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 640, + 359, + 656 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We introduced ScatterFormer, an innovative architecture designed for 3D object detection using point clouds, specifically targeting the challenges associated with processing sparse and unevenly distributed data from LiDAR sensors. The cornerstone of our approach is the Scattered Linear Attention (SLA) module, which effectively addresses the limitations of conventional attention mechanisms in managing voxel features of varying lengths. SLA ingeniously combines linear attention with a chunk-wise matrix multiplication algorithm, tailored to meet the distinct requirements of processing voxels grouped by windows. By integrating SLA with a novel cross-window interaction (CWI) module, ScatterFormer achieves higher accuracy and lower latency, surpassing traditional transformer-based and sparse CNN-based detectors in extensive 3D detection tasks.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "C. He et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 323, + 160 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Ali, A., Touvron, H., Caron, M., Bojanowski, P., Douze, M., Joulin, A., Laptev, I., Neverova, N., Synnaeve, G., Verbeek, J., et al.: Xcit: Cross-covariance image transformers. Advances in neural information processing systems 34, 20014-20027 (2021)", + "2. Bai, X., Hu, Z., Zhu, X., Huang, Q., Chen, Y., Fu, H., Tai, C.L.: Transfusion: Robust lidar-camera fusion for 3d object detection with transformers. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1090-1099 (2022)", + "3. Cai, H., Gan, C., Han, S.: Efficientvit: Enhanced linear attention for high-resolution low-computation visual recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2023)", + "4. Chen, X., Shi, S., Zhu, B., Cheung, K.C., Xu, H., Li, H.: Mppnet: Multi-frame feature intertwining with proxy points for 3d temporal object detection. In: European Conference on Computer Vision. pp. 680-697. Springer (2022)", + "5. Chen, Y., Liu, J., Zhang, X., Qi, X., Jia, J.: Voxelnext: Fully sparse voxelnet for 3d object detection and tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21674-21683 (2023)", + "6. Chu, X., Tian, Z., Zhang, B., Wang, X., Shen, C.: Conditional positional encodings for vision transformers. ICLR (2023)", + "7. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. ICLR (2021)", + "8. Fan, L., Pang, Z., Zhang, T., Wang, Y.X., Zhao, H., Wang, F., Wang, N., Zhang, Z.: Embracing single stride 3d object detector with sparse transformer. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8458-8468 (2022)", + "9. Fan, L., Wang, F., Wang, N., Zhang, Z.: Fully sparse 3d object detection. Advances in Neural Information Processing Systems 35, 351-363 (2022)", + "0. Fan, L., Xiong, X., Wang, F., Wang, N., Zhang, Z.: Rangedet: In defense of range view for lidar-based 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2918-2927 (October 2021)", + "1. Guan, T., Wang, J., Lan, S., Chandra, R., Wu, Z., Davis, L., Manocha, D.: M3detr: Multi-representation, multi-scale, mutual-relation 3d object detection with transformers. In: Proceedings of the IEEE/CVF winter conference on applications of computer vision. pp. 772-782 (2022)", + "2. Guo, M.H., Cai, J.X., Liu, Z.N., Mu, T.J., Martin, R.R., Hu, S.M.: Pct: Point cloud transformer. Computational Visual Media 7(2), 187-199 (Apr 2021). https://doi.org/10.1007/s41095-021-0229-5, http://dx.doi.org/10.1007/s41095-021-0229-5", + "3. Han, D., Pan, X., Han, Y., Song, S., Huang, G.: Flatten transformer: Vision transformer using focused linear attention. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5961-5971 (2023)", + "4. He, C., Li, R., Li, S., Zhang, L.: Voxel set transformer: A set-to-set approach to 3d object detection from point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8417-8427 (2022)", + "5. He, C., Li, R., Zhang, Y., Li, S., Zhang, L.: Msf: Motion-guided sequential fusion for efficient 3d object detection from point cloud sequences. In: Proceedings of the" + ], + "bbox": [ + 225, + 175, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "ScatterFormer", + "bbox": [ + 633, + 114, + 730, + 127 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5196-5205 (2023)", + "16. He, C., Zeng, H., Huang, J., Hua, X.S., Zhang, L.: Structure aware single-stage 3d object detection from point cloud. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11873-11882 (2020)", + "17. Katharopoulos, A., Vyas, A., Pappas, N., Fleuret, F.: Transformers are rnns: Fast autoregressive transformers with linear attention. In: International conference on machine learning. pp. 5156-5165. PMLR (2020)", + "18. Lang, A.H., Vora, S., Caesar, H., Zhou, L., Yang, J., Beijbom, O.: Pointpillars: Fast encoders for object detection from point clouds. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 12697-12705 (2019)", + "19. Li, J., Luo, C., Yang, X.: Rethinking network designs for 3d object detection in lidar point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17567-17576 (2023)", + "20. Li, Z., Wang, F., Wang, N.: Lidar r-cnn: An efficient and universal 3d object detector. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021)", + "21. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 10012-10022 (October 2021)", + "22. Liu, Z., Zhang, Z., Cao, Y., Hu, H., Tong, X.: Group-free 3d object detection via transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2949-2958 (October 2021)", + "23. Liu, Z., Yang, X., Tang, H., Yang, S., Han, S.: Flatformer: Flattened window attention for efficient point cloud transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1200-1211 (2023)", + "24. Mao, J., Xue, Y., Niu, M., Bai, H., Feng, J., Liang, X., Xu, H., Xu, C.: Voxel transformer for 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 3164-3173 (October 2021)", + "25. Misra, I., Girdhar, R., Joulin, A.: An end-to-end transformer model for 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2906-2917 (October 2021)", + "26. Ngiam, J., Caine, B., Han, W., Yang, B., Chai, Y., Sun, P., Zhou, Y., Yi, X., Alsharif, O., Nguyen, P., et al.: Starnet: Targeted computation for object detection in point clouds. arXiv preprint arXiv:1908.11069 (2019)", + "27. Pan, X., Xia, Z., Song, S., Li, L.E., Huang, G.: 3d object detection with pointformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7463-7472 (June 2021)", + "28. Qi, C.R., Litany, O., He, K., Guibas, L.J.: Deep though voting for 3d object detection in point clouds. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (October 2019)", + "29. Qi, C.R., Liu, W., Wu, C., Su, H., Guibas, L.J.: Frustum pointnets for 3d object detection from rgb-d data. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 918-927 (2018)", + "30. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 652-660 (2017)", + "31. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In: Advances in neural information processing systems. pp. 5099-5108 (2017)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "C. He et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "32. Qin, Z., Li, D., Sun, W., Sun, W., Shen, X., Han, X., Wei, Y., Lv, B., Yuan, F., Luo, X., Qiao, Y., Zhong, Y.: Scaling transnormer to 175 billion parameters (2023)", + "33. Sheng, H., Cai, S., Liu, Y., Deng, B., Huang, J., Hua, X.S., Zhao, M.J.: Improving 3d object detection with channel-wise transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2743-2752 (October 2021)", + "34. Shi, G., Li, R., Ma, C.: Pillarnet: Real-time and high-performance pillar-based 3d object detection. In: European Conference on Computer Vision. pp. 35-52. Springer (2022)", + "35. Shi, S., Guo, C., Jiang, L., Wang, Z., Shi, J., Wang, X., Li, H.: Pv-rcnn: Pointvoxel feature set abstraction for 3d object detection. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)", + "36. Shi, S., Jiang, L., Deng, J., Wang, Z., Guo, C., Shi, J., Wang, X., Li, H.: Pvrnn++: Point-voxel feature set abstraction with local vector representation for 3d object detection. International Journal of Computer Vision 131(2), 531-551 (2023)", + "37. Shi, S., Wang, X., Li, H.: Pointrecnn: 3d object proposal generation and detection from point cloud. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 770-779 (2019)", + "38. Shi, S., Wang, Z., Wang, X., Li, H.: Part-a^2 net: 3d part-aware and aggregation neural network for object detection from point cloud. arXiv preprint arXiv:1907.03670 (2019)", + "39. Shi, W., Rajkumar, R.: Point-gnn: Graph neural network for 3d object detection in a point cloud. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1711-1719 (2020)", + "40. Sun, P., Tan, M., Wang, W., Liu, C., Xia, F., Leng, Z., Anguelov, D.: Swformer: Sparse window transformer for 3d object detection in point clouds. In: European Conference on Computer Vision. pp. 426-442. Springer (2022)", + "41. Sun, P., Wang, W., Chai, Y., Elsayed, G., Bewley, A., Zhang, X., Sminchisescu, C., Anguelov, D.: Rsn: Range sparse net for efficient, accurate lidar 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5725-5734 (2021)", + "42. Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., Wojna, Z.: Rethinking the inception architecture for computer vision. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2818-2826 (2016)", + "43. Team, O.D.: Openpcdet: An open-source toolbox for 3d object detection from point clouds. https://github.com/open-mmlab/OpenPCDet (2020)", + "44. Tilet, P., Kung, H.T., Cox, D.: Triton: An intermediate language and compiler for tiled neural network computations. In: Proceedings of the 3rd ACM SIGPLAN International Workshop on Machine Learning and Programming Languages. p. 10-19. MAPL 2019, Association for Computing Machinery, New York, NY, USA (2019). https://doi.org/10.1145/3315508.3329973, https://doi.org/10.1145/3315508.3329973", + "45. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jegou, H.: Training data-efficient image transformers & distillation through attention. In: International Conference on Machine Learning. vol. 139, pp. 10347-10357 (July 2021)", + "46. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: Advances in neural information processing systems. pp. 5998-6008 (2017)" + ], + "bbox": [ + 212, + 146, + 785, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "ScatterFormer", + "bbox": [ + 633, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "47. Wang, H., Shi, C., Shi, S., Lei, M., Wang, S., He, D., Schiele, B., Wang, L.: Dsvt: Dynamic sparse voxel transformer with rotated sets. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13520-13529 (2023)", + "48. Wang, S., Li, B.Z., Khabsa, M., Fang, H., Ma, H.: Linformer: Self-attention with linear complexity (2020)", + "49. Yan, Y., Mao, Y., Li, B.: Second: Sparsely embedded convolutional detection. Sensors 18(10), 3337 (2018)", + "50. Yang, B., Luo, W., Urtasun, R.: Pixor: Real-time 3d object detection from point clouds. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 7652-7660 (2018)", + "51. Yang, Z., Zhou, Y., Chen, Z., Ngiam, J.: 3d-man: 3d multi-frame attention network for object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1863-1872 (2021)", + "52. Yin, T., Zhou, X., Krahenbuhl, P.: Center-based 3d object detection and tracking. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11784-11793 (2021)", + "53. Zhang, G., Chen, J., Gao, G., Li, J., Hu, X.: Hednet: A hierarchical encoder-decoder network for 3d object detection in point clouds. arXiv preprint arXiv:2310.20234 (2023)", + "54. Zhang, G., Fan, L., He, C., Lei, Z., Zhang, Z., Zhang, L.: Voxel mamba: Group-free state space models for point cloud based 3d object detection. arXiv preprint arXiv:2406.10700 (2024)", + "55. Zhang, Y., Huang, D., Wang, Y.: Pc-rgnn: Point cloud completion and graph neural network for 3d object detection. In: Proceedings of the AAAI conference on artificial intelligence. vol. 35, pp. 3430-3437 (2021)", + "56. Zhang, Y., Hu, Q., Xu, G., Ma, Y., Wan, J., Guo, Y.: Not all points are equal: Learning highly efficient point-based detectors for 3d lidar point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18953-18962 (2022)", + "57. Zhao, H., Jiang, L., Jia, J., Torr, P.H., Koltun, V.: Point transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 16259-16268 (October 2021)", + "58. Zheng, W., Tang, W., Jiang, L., Fu, C.W.: Se-ssd: Self-ensembling single-stage object detector from point cloud. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14494-14503 (2021)", + "59. Zhou, C., Zhang, Y., Chen, J., Huang, D.: Octr: Octtree-based transformer for 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5166-5175 (2023)", + "60. Zhou, Y., Tuzel, O.: Voxelnet: End-to-end learning for point cloud based 3d object detection. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4490-4499 (2018)", + "61. Zhou, Z., Zhao, X., Wang, Y., Wang, P., Foroosh, H.: Centerformer: Center-based transformer for 3d object detection. In: European Conference on Computer Vision. pp. 496-513. Springer (2022)" + ], + "bbox": [ + 215, + 146, + 785, + 771 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "C. He et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/ScatterFormer_ Efficient Voxel Transformer with Scattered Linear Attention/1045fb37-bbfd-4adc-8de1-f127bb3078ca_model.json b/2024/ScatterFormer_ Efficient Voxel Transformer with Scattered Linear Attention/1045fb37-bbfd-4adc-8de1-f127bb3078ca_model.json new file mode 100644 index 0000000000000000000000000000000000000000..39ac055b6dd024bf691d744a8b38bd9ddbcd96d1 --- /dev/null +++ b/2024/ScatterFormer_ Efficient Voxel Transformer with Scattered Linear Attention/1045fb37-bbfd-4adc-8de1-f127bb3078ca_model.json @@ -0,0 +1,2348 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.227, + 0.141, + 0.78, + 0.187 + ], + "angle": 0, + "content": "ScatterFormer: Efficient Voxel Transformer with Scattered Linear Attention" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.213, + 0.744, + 0.23 + ], + "angle": 0, + "content": "Chenhang He\\(^{1}\\), Ruihuang Li\\(^{1,2}\\), Guowen Zhang\\(^{1}\\), and Lei Zhang\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.28, + 0.24, + 0.724, + 0.255 + ], + "angle": 0, + "content": "1 The Hong Kong Polytechnic University, Hong Kong SAR, China" + }, + { + "type": "text", + "bbox": [ + 0.375, + 0.255, + 0.631, + 0.325 + ], + "angle": 0, + "content": "\\(^{2}\\) OPPO Research, Shenzhen, China chenhang.he@polyu.edu.hk csrhli@comp.polyu.edu.hk guowen.zhang@connect.polyu.edu.hk cslzhang@comp.polyu.edu.hk" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.368, + 0.744, + 0.673 + ], + "angle": 0, + "content": "Abstract. Window-based transformers excel in large-scale point cloud understanding by capturing context-aware representations with affordable attention computation in a more localized manner. However, the sparse nature of point clouds leads to a significant variance in the number of voxels per window. Existing methods group the voxels in each window into fixed-length sequences through extensive sorting and padding operations, resulting in a non-negligible computational and memory overhead. In this paper, we introduce ScatterFormer, which to the best of our knowledge, is the first to directly apply attention to voxels across different windows as a single sequence. The key of ScatterFormer is a Scattered Linear Attention (SLA) module, which leverages the pre-computation of key-value pairs in linear attention to enable parallel computation on the variable-length voxel sequences divided by windows. Leveraging the hierarchical structure of GPUs and shared memory, we propose a chunkwise algorithm that reduces the SLA module's latency to less than 1 millisecond on moderate GPUs. Furthermore, we develop a cross-window interaction module that improves the locality and connectivity of voxel features across different windows, eliminating the need for extensive window shifting. Our proposed ScatterFormer demonstrates 73.8 mAP (L2) on the Waymo Open Dataset and 72.4 NDS on the NuScenes dataset, running at an outstanding detection rate of 23 FPS. The code is available at https://github.com/skyhehe123/ScatterFormer." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.686, + 0.631, + 0.701 + ], + "angle": 0, + "content": "Keywords: 3D Object Detection \\(\\cdot\\) Voxel Transformer" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.73, + 0.377, + 0.745 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.842 + ], + "angle": 0, + "content": "In the field of 3D object detection, the use of point clouds has become increasingly popular, especially for providing accurate and reliable perception results in autonomous systems. Unlike image data, point clouds obtained from LiDAR are often sparse and nonuniformly distributed, with varying density depending on their distances from the sensor. Earlier approaches utilize point-cloud operators" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.127 + ], + "angle": 0, + "content": "C. He et al." + }, + { + "type": "image", + "bbox": [ + 0.316, + 0.145, + 0.688, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.269, + 0.704, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.332, + 0.785, + 0.375 + ], + "angle": 0, + "content": "Fig. 1: Illustration of (a) group-based attention and (b) our scattered linear attention on variable-length sequences. The square in different colors represent voxels in different windows and the square with a cross represents the padding voxel." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.385, + 0.785, + 0.446 + ], + "angle": 0, + "content": "like PointNet++ [31] for feature extraction in continuous space [37]. Some approaches [5,9,16,18,49,50,52,60] have transformed point clouds into voxel grids, which are then efficiently processed by a sparse convolutional neural network (SpCNN) [49]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.447, + 0.786, + 0.674 + ], + "angle": 0, + "content": "Recently, the success of vision transformers [7, 21] has motivated a number of attention-based methods to process indoor point clouds [12, 22, 24, 25, 27, 57]. Inspired by SwinTransformer [21], various studies [8, 14, 23, 47] have advanced the application of window-based Transformers in large-scale 3D detection tasks within outdoor environments, achieving outstanding performance and highlighting their potential as alternatives to SpCNN. However, due to the inherent sparsity of point clouds, the number of features grouped by windows can vary significantly, hampering the parallelism in attention computation. To resolve this issue, SST [8] and SWFormer [40] group the voxels within a window into different batches and manages the attention computation in a hybrid serial-parallel manner. More effectively, DSVT [47] alternatively sort the voxels within a window from different axes and partition them into sequences of fixed length, which allow parallel computation on sparse voxels. Albeit effective, these group-based methods, as shown in Figure 1(a), require extensive sorting and padding operations, incurring substantial memory and computational overhead." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.787, + 0.841 + ], + "angle": 0, + "content": "In this paper, we delve into the window-based voxel transformer where the voxels grouped by windows form variable-length sequences \\(\\{X_1\\in \\mathbb{R}^{n_1\\times d},X_2\\in\\) \\(\\mathbb{R}^{n_2\\times d},\\dots,X_k\\in \\mathbb{R}^{n_k\\times d}\\}\\). Their corresponding self-attention matrices \\(\\{A_{1}\\in\\) \\(\\mathbb{R}^{n_1\\times n_1}\\) \\(A_{2}\\in \\mathbb{R}^{n_{2}\\times n_{2}}\\) ..., \\(A_{k}\\in \\mathbb{R}^{n_{k}\\times n_{k}}\\}\\) occupy irregular memory spaces, making it a challenge to perform attention in parallel for the voxels of the entire scene. Recent efforts on linear attention [1,13,17,48] have emerged a promising alternative to traditional attention. By approximating softmax operation with the kernel function, i.e., \\(\\phi (Q)\\cdot \\phi (K)^{\\top}\\approx \\mathrm{softmax}(QK^{\\top})\\), we can change the computation order from \\((Q\\cdot K^{\\top})\\cdot V\\) to \\(Q\\cdot (K^{\\top}\\cdot V)\\). This not only results in linear complexity but also yields a compressed hidden state \\(S\\in \\mathbb{R}^{d\\times d}\\) that is invariant to the sequence length. Another attractive property of linear attention is that it" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.634, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "ScatterFormer" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.128 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.513, + 0.161 + ], + "angle": 0, + "content": "can be converted into a \"recurrent\" form:" + }, + { + "type": "equation", + "bbox": [ + 0.398, + 0.178, + 0.786, + 0.196 + ], + "angle": 0, + "content": "\\[\no _ {t} = q _ {t} S _ {t}; \\quad S _ {t} = S _ {t - 1} + k _ {t} ^ {\\mathsf {T}} v _ {t} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.212, + 0.784, + 0.273 + ], + "angle": 0, + "content": "where the sequences can be divided into non-overlapping chunks, \\(\\{q_j,k_j,v_j\\}_{j = 1:t}\\). The output can be calculated by scanning over the sequence, based on an updated hidden state \\(S_{t}\\). This enables the use of chunk-level computation to address the inability to parallelize at the sequence level due to varying lengths." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.274, + 0.785, + 0.456 + ], + "angle": 0, + "content": "Inspired by this, we introduce the Scattered Linear Attention (SLA) module, which accommodates linear attention in a window-based voxel transformer. As shown in Figure 1(b), the SLA module treats the voxels of the entire scene into a single sequence and processes them directly without padding voxels. Using the recurrent form of linear attention, we develop an I/O-aware algorithm to further optimize matrix multiplication on voxel sequences. Specifically, we divide the voxel sequence into multiple chunks and loaded them into the shared memory (SRAM) of the GPU. The computation of the hidden state matrix in each window is then achieved through a series of chunk-wise matrix multiplications and cumulative sums. This optimized implementation significantly reduces I/O overhead and memory usage, resulting in extremely fast and memory-efficient attention computation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.458, + 0.785, + 0.609 + ], + "angle": 0, + "content": "In addition, current window-based transformer models use window shifting to propagate the information across the windows. After performing an instruction-level analysis of existing implementations, we observe that the voxel permutation operations caused by window shifting consume significant computational overhead. To address this issue, we propose a cross-window interaction (CWI) module, which is composed of depth-wise convolutions with small 2D kernels and lengthy 1D kernels that allow the voxel features in each window to fully interact with the features in other windows. As a result, the proposed CWI module can improve both the locality and connectivity of the voxel features while requiring minimal computational effort." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.611, + 0.785, + 0.718 + ], + "angle": 0, + "content": "Building upon the SLA and CWI modules, we propose the ScatterFormer, an innovative voxel transformer for large-scale point cloud understanding. Our experiments show that ScatterFormer achieves linear complexity without compromising accuracy. It achieves superior results compared to the state-of-the-art model DSVT [47]. The latency of ScatterFormer is significantly lower than that of transformer-based detectors [8, 14, 47], which is comparable to that of sparse convolution-based detectors [52]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.785, + 0.84 + ], + "angle": 0, + "content": "In conclusion, we delve into attention on voxels grouped by windows, highlighting the challenges in memory allocation and computation for variable-length sequences. Then, we introduce an SLA module, which can directly process the voxels grouped by windows by facilitating the linear attention formula. A chunkwise matrix multiplication algorithm is proposed to further accelerate the attention computation in SLA. Finally, we present ScatterFormer, which has linear complexity and can efficiently process large-scale LiDAR scenes, achieving better accuracy with lower latency compared to existing transformer-based detectors." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.232, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "C. He et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.389, + 0.161 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.181, + 0.592, + 0.197 + ], + "angle": 0, + "content": "2.1 Point Cloud-based 3D Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.21, + 0.787, + 0.361 + ], + "angle": 0, + "content": "There exist two primary point-cloud representations in 3D object detection, i.e., point-based and voxel-based ones. In point-based methods [28, 29, 37, 39, 55], point clouds are first passed through a point-based backbone network [30], in which the points are gradually sampled and features are learned by point cloud operators. F-Pointnet [29] first employs PoinNet to detect 3D objects based on the frustums lifted by 2D proposals. PointRCNN [37] direct generates 3D proposals from point-based features with foreground segmentation. VoteNet [28] clusters objects from the surface points using a deep Hough voting method. Other point operators based on the point graph [39, 55] and the range view [10] have also been developed for point cloud processing." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.362, + 0.787, + 0.588 + ], + "angle": 0, + "content": "Point-based methods are primarily limited by their inference efficiency and the absence of contextual features in a continuous space. On the other hand, voxel-based approaches [5, 16, 18, 49, 52, 58, 60] transform the entire point clouds into regular grids through voxelization, showcasing superior efficiency and context representation. VoxelNet [60] proposes a voxel feature encoding (VFE) module and combines it with 3D convolutions to extract voxel features in an end-to-end manner. SECOND [49] optimizes 3D convolution for sparse data, resulting in a significant reduction in both time and memory. PointPillars [18] demonstrates an efficient model by stacking voxels into vertical columns, subsequently processing them with naive 2D convolutions. There are also efforts [16, 35] that explore hybrid representations of point-based and voxel-based networks, demonstrating a better trade-off between speed and accuracy. However, a common limitation across these methods is their reliance on small convolution kernels with restricted receptive fields, making them less adept at capturing the global context for 3D object detection." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.589, + 0.787, + 0.649 + ], + "angle": 0, + "content": "In this paper, we shed light on the voxel-based methods and introduce a Scattered Linear Attention module that enables dynamic modeling among voxels within each window with linear complexity, thereby efficiently extracting window-based contextual features." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.676, + 0.497, + 0.69 + ], + "angle": 0, + "content": "2.2 Transformer on Point Cloud" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Inspired by the significant success of self-attention in NLP [46] and CV [7, 45], Transformers have been adapted for 3D vision due to their ability to capture long-range dependencies. The Point Transformers [27, 57] employ attention to modulate point clouds for classification and segmentation tasks. PCT [12] presents an optimized offset attention module, which, when combined with the implicit Laplacian operator and normal estimation, becomes more adept at point-cloud processing. Some methods like [8, 22] opt for voxels or key points to optimize latency. 3DETR [25] proposes a query-based 3D object detection scheme. CT3D [33] enhances the region-based network using a channel-wise" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.635, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "ScatterFormer" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.15, + 0.784, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.265, + 0.788, + 0.322 + ], + "angle": 0, + "content": "Fig. 2: The macro design of ScatterFormer. The backbone comprises a Conditional Position Encoding (CPE) and six transformer blocks. Each block is composed by a Scattered Linear Attention (SLA) module, a Cross-Window Interaction (CWI) module, and a Feed-Forward Network (FFN)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.333, + 0.788, + 0.53 + ], + "angle": 0, + "content": "Transformer framework. To achieve context-rich representations, several studies [8, 14, 23, 24, 47] have integrated the attention module into point- or voxel-based encoders. For instance, VoTr [24] utilizes dilated attention for expanded receptive fields; VoxSet [14] applies set attention for extracting point-based features in set-to-set translation; SST [8] employs local attention with shifted windows; and OcTr [59] adopts an Octree-based attention for efficient hierarchical context learning. Nevertheless, how to efficiently leverage global context from attention remains a challenge due to the intrinsic sparsity of point clouds. DSVT [47] and FlatFormer [23] group the voxels within each window into a series of fixed-length voxel sets, thus extracting the features in a fully parallel manner. Recently, a group-free state-space model [54] was proposed to directly process voxels as a sequence. However, these approaches more or less lose the spatial proximity and incur extensive computational overhead in grouping and sorting the voxels." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.552, + 0.331, + 0.568 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.789, + 0.841 + ], + "angle": 0, + "content": "The overall architecture of our proposed ScatterFormer is depicted in Figure 2. It begins with the input point clouds, which are voxelized and transformed into high-dimensional embeddings using a VFE layer [60]. These embeddings are then processed through Conditional Positional Encoding (CPE) using a shallow convolutional network [6]. The encoded features enter the ScatterFormer backbone, consisting of six ScatterFormer blocks. Each block includes a Scattered Linear Attention (SLA) module, a Cross-Window Interaction (CWI) module, and a Feed-Forward Network (FFN), interspersed with Batch Normalization layers and skip connections. After three ScatterFormer blocks, the voxel features are downsampled via a sparse convolutional layer. The downsampled features are then converted into pillar features [47], generating compact BEV features for bounding-box prediction. ScatterFormer ScatterFormer stands out by not requiring voxel features to be organized into fixed-length sets [8, 23, 47], enabling flexible attention computation across windows. Additionally, the CWI module obviates the need for window shifting. These innovations allow ScatterFormer to avoid unnecessary memory allocation and permutation operations, achieving high efficiency comparable to CNN-based models." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "C. He et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.403, + 0.161 + ], + "angle": 0, + "content": "3.1 Linear Attention" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.173, + 0.79, + 0.324 + ], + "angle": 0, + "content": "We begin by revisiting the self-attention introduced by [46]. Given an input matrix \\( x \\in \\mathbb{R}^{N \\times d} \\), where \\( N \\) denotes the number of tokens and \\( d \\) denotes the dimensionality. Self-attention first linearly projects the input to queries, keys, and values using weight matrices \\( W_{Q}, W_{K} \\), and \\( W_{V} \\), such that \\( Q = xW_{Q}, K = xW_{K}, V = xW_{V} \\). Then it computes a Softmax-based attention map based on queries and keys i.e., \\( A(Q, K) = \\mathrm{Softmax}(QK^{\\top}) / \\sqrt{d} \\). The output of self-attention is defined as the weighted sum of \\( N \\) values with the weights corresponding to the attention map, i.e., \\( O = A(Q, K)V \\). This approach involves calculating the similarity for all query-key pairs, which yields a computational complexity of \\( O(N^{2}) \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.325, + 0.789, + 0.448 + ], + "angle": 0, + "content": "Linear attention [1, 17] offers a notable alternative to self-attention by significantly reducing its computational complexity from \\( O(N^2) \\) to \\( O(N) \\). This efficiency is achieved through the application of kernel functions on the query \\( (Q) \\) and key \\( (K) \\) matrices, which effectively approximates the original attention map without relying on Softmax, i.e., \\( A(Q, K) = \\phi(Q)\\phi(K)^{\\mathsf{T}} \\). Taking advantage of the associative property of matrix multiplication, the computation shifts from \\( (\\phi(Q)\\phi(K)^{\\mathsf{T}})V \\) to \\( \\phi(Q)(\\phi(K)^{\\mathsf{T}}V) \\). This modification leads to a computational complexity proportional to the number of tokens, maintaining an order of \\( O(N) \\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.469, + 0.542, + 0.485 + ], + "angle": 0, + "content": "3.2 Scattered Linear Attention (SLA)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.496, + 0.789, + 0.618 + ], + "angle": 0, + "content": "Instead of organizing the voxels within a window into fixed-length subsets, we arrange the voxels of the entire scene into a single flattened matrix, as illustrated in Figure 3(a). These voxels are ordered based on their window coordinates, ensuring that voxels from the same window form a sub-matrix in contiguous memory. Initially, we use a shared projection layer to map all the voxels in the scene to query, key, and value representations, denoted \\( Q \\), \\( K \\), \\( V \\). Subsequently, we perform attention computations on these submatrices separately, resulting in the scattered linear attention (SLA) formula as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.333, + 0.63, + 0.785, + 0.647 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S L A} (Q, K, V) = \\operatorname {C o n c a t} \\left[ \\operatorname {L A} \\left(Q ^ {j}, K ^ {j}, V ^ {j}\\right) \\right] _ {j = 1: M}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.706 + ], + "angle": 0, + "content": "where \\( M \\) is the number of non-empty window in the current scene and LA is the linear attention formula used in [17]. Based on this formula, the output voxel features \\( O^j \\) in the \\( j^{\\mathrm{th}} \\) window can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.387, + 0.718, + 0.785, + 0.758 + ], + "angle": 0, + "content": "\\[\n\\mathrm {L A}: O ^ {j} = \\frac {\\phi (Q) \\sum_ {i = 1} ^ {m ^ {j}} \\phi \\left(K _ {i}\\right) ^ {\\top} V _ {i}}{\\phi (Q) \\sum_ {i = 1} ^ {m ^ {j}} \\phi \\left(K _ {i}\\right) ^ {\\top}} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.769, + 0.786, + 0.785 + ], + "angle": 0, + "content": "where \\(m^j\\) is the number of voxel in the window and \\(\\phi (x)\\) is the kernel function." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Hardware Efficient Implementation. It should be noted that the implementation of Equation 2 is not straightforward, as the sub-matrices have different numbers of rows. One naive implementation is to cache the \"kv\" matrix of each" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.635, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ScatterFormer" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.143, + 0.541, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.397, + 0.31, + 0.413, + 0.322 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.552, + 0.147, + 0.78, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.66, + 0.311, + 0.676, + 0.323 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.331, + 0.788, + 0.43 + ], + "angle": 0, + "content": "Fig. 3: (a) The voxel features are sorted by windows and flattened to a single matrix, which is further partitioned into multiple chunks of uniform length and sent to SLA kernel for attention computation. (b) In SLA, we first allocate individual threads to each window, each of them iterates over all corresponding key and value chunks, calculates, and accumulates their products to obtain the hidden state matrix of each window. Then, multiple threads are allocated to these query chunks \\( q_{i} \\in Q \\), calculating the chunk-wise output by multiplying \\( q_{i} \\) with the corresponding hidden state matrix." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.444, + 0.789, + 0.67 + ], + "angle": 0, + "content": "voxel and apply \\( scatter^3 \\) operator to accumulate them within each window. However, this would consume a high memory overhead and IO latency. Considering that GPUs feature a memory hierarchy that includes larger, slower global GPU memory (high-bandwidth memory; HBM) and smaller, faster shared memory (SRAM), optimal utilization of SRAM to minimize HBM I/O costs can lead to significant speed-ups. Based on this, we partition the flattened \\( Q, K, V \\) into multiple chunks, load them from slow HBM to fast SRAM. Specifically, for each window, we allocate an single thread. Each thread iterates over all the key and value chunks, calculates, and accumulates their products to obtain the hidden state matrix of the current window. After calculating the hidden state matrix in window, we assign individual thread to each query chunk \\( q_i \\in Q^j \\) and calculate the chunk-wise output can then be following Equation 3. Figure 3(b) illustrates this chunk-wise matrix multiplication approach. As can be seen, it not only avoids outputting large matrices but also allows for more flexible processing of variable-length sequences." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.671, + 0.788, + 0.734 + ], + "angle": 0, + "content": "In our implementation, we use Triton [44] to perform chunk-wise matrix multiplication. As illustrated in Figure 4, it outperforms the naive scattered-based operation, demonstrating improved speed and reduced memory usage when computing attention for extremely long sequences." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.751, + 0.48, + 0.765 + ], + "angle": 0, + "content": "3.3 Cross Window Interaction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.773, + 0.788, + 0.819 + ], + "angle": 0, + "content": "Traditional window-based transformers [8,40,47] utilize window shifting for interwindow connection. For point cloud sequences, this requires recalculating window coordinates and rearranging the voxels. In DSVT, the computation of voxel" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.824, + 0.533, + 0.841 + ], + "angle": 0, + "content": "3 https://github.com/rustyls/pytorch_scatter" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "C. He et al." + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.151, + 0.72, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.292, + 0.72, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.438, + 0.788, + 0.467 + ], + "angle": 0, + "content": "Fig. 4: Comparison between between scatter-based implementation (yellow) and our chunk-wise implementation (blue)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.486, + 0.784, + 0.56 + ], + "angle": 0, + "content": "indices in subsets and the rearrangement of voxel matrices are particularly time consuming, accounting for \\(24\\%\\) of the total backbone latency, as shown in Figure 5. To alleviate this unnecessary computational overhead from window shifting, we introduce the Cross-Window Interaction (CWI) module, which employs convolutions with large kernel to blend voxel features across windows." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.562, + 0.788, + 0.712 + ], + "angle": 0, + "content": "We adhere to the design principles of the Inception architecture [42], where the input features are divided along the channel dimension and processed through multiple branches using group-wise convolutions. Instead of employing a single large kernel for convolution, we adopt more efficient 1D kernels along different axes for feature aggregation. Specifically, we apply a \\((S_h + 1) \\times 1 \\times 1\\) convolution in one branch and a \\(1 \\times (S_w + 1) \\times 1\\) convolution in another branch, where \\((S_w, S_h)\\) denotes the window size. Additionally, we incorporate a standard \\(3 \\times 3 \\times 3\\) convolution in one branch to enhance locality, while another branch performs an identity mapping. Given input \\(X\\) of channel \\(4c\\), the output \\(X'\\) of the Cross Window Interaction module can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.377, + 0.723, + 0.786, + 0.742 + ], + "angle": 0, + "content": "\\[\nX ^ {\\prime} = \\operatorname {C o n c a t} \\left(X _ {k}, X _ {w}, X _ {h}, X ^ {3 c: 4 c}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.753, + 0.265, + 0.766 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.379, + 0.78, + 0.785, + 0.797 + ], + "angle": 0, + "content": "\\[\nX _ {k} = \\operatorname {D W C o n v} _ {\\left(S _ {h} + 1\\right) \\times 1 \\times 1} \\left(X ^ {: c}\\right), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.378, + 0.799, + 0.785, + 0.817 + ], + "angle": 0, + "content": "\\[\nX _ {w} = \\operatorname {D W C o n v} _ {1 \\times \\left(S _ {w} + 1\\right) \\times 1} \\left(X ^ {c: 2 c}\\right), \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.379, + 0.82, + 0.785, + 0.838 + ], + "angle": 0, + "content": "\\[\nX _ {h} = \\operatorname {D W C o n v} _ {3 \\times 3 \\times 3} \\left(X ^ {2 c: 3 c}\\right). \\tag {7}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.635, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "ScatterFormer" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.144, + 0.499, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.144, + 0.695, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.309, + 0.785, + 0.339 + ], + "angle": 0, + "content": "Fig. 5: Runtime decomposition in DSVT [47] and ScatterFormer backbones. The offsets generation means computing the starting address of each chunk in the flattened matrix." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.353, + 0.785, + 0.399 + ], + "angle": 0, + "content": "In our experiments, we found that this axis-decomposed large kernel design can achieve a better precision-latency trade-off. With these lengthy kernels, the voxel features can be efficiently blended among different windows." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.417, + 0.471, + 0.431 + ], + "angle": 0, + "content": "3.4 Detection Head and Loss" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.439, + 0.788, + 0.577 + ], + "angle": 0, + "content": "To complement existing detection heads, ScatterFormer generates dense BEV feature maps from sparse voxel representations by placing them back to their spatial locations and filling the unoccupied positions with zeros. For the BEV network, we simply follow [53] by adopting a cascaded CNN with convolutional blocks at multiple levels of strides. In terms of detection head and loss function, we follow the design identical to that used in DSVT [47]. We train the detection model with heatmap estimation, bounding box regression, and incorporate an IoU loss for confidence calibration. On the nuScenes dataset, the detection head follows the architecture used in Transfusion [2]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.595, + 0.376, + 0.612 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.623, + 0.533, + 0.637 + ], + "angle": 0, + "content": "4.1 Datasets and Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.735 + ], + "angle": 0, + "content": "Waymo Open Dataset (WOD). This dataset contains 230,000 annotated samples split into 160,000 for training, 40,000 for validation, and 30,000 for testing. It uses two metrics for 3D object detection: mean average precision (mAP) and mAP weighted by heading accuracy (mAPH), further categorized into Level 1 (L1) for objects detected by more than five LiDAR points and Level 2 (L2) for those detected with at least one point." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "NuScenes. This dataset comprises 40,000 annotated samples, with 28,000 for training, 6,000 for validation, and 6,000 for testing. On this dataset, the model performance is measured by mean average precision (mAP) across multiple distance thresholds (0.5, 1, 2, and 4 meters) and the nuScenes detection score (NDS), which combines mAP with a weighted sum of five additional metrics assessing true positive predictions in translation, scale, orientation, velocity, and attribute accuracy." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.127 + ], + "angle": 0, + "content": "C. He et al." + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.15, + 0.709, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.399, + 0.788, + 0.429 + ], + "angle": 0, + "content": "Fig. 6: The detection performance (mAPH/L2) vs. speed (FPS) of different methods on Waymo validation set. The speeds are measured on an NVIDIA A100 GPU." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.439, + 0.458, + 0.455 + ], + "angle": 0, + "content": "4.2 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.463, + 0.789, + 0.63 + ], + "angle": 0, + "content": "Our approach is implemented using the open-source framework OpenPCDet [43]. To construct ScatterFormer, we set the voxel size to \\((0.32\\mathrm{m}, 0.32\\mathrm{m}, 0.1875\\mathrm{m})\\) for the Waymo dataset and \\((0.3\\mathrm{m}, 0.3\\mathrm{m}, 8\\mathrm{m})\\) for the NuScenes dataset. The window sizes \\((S_w, S_h)\\) for the two datasets are set to \\((12, 12)\\) and \\((20, 20)\\), respectively. We stack six building blocks for the backbone network. We configure our attention module to have 4 heads with a dimensionality of 128. ScatterFormer is trained for 24 epochs with a learning rate of 0.006 on Waymo Dataset and 20 epochs with a learning rate of 0.004 on NuScenes Dataset. In the last 4 epochs, we disabled the data augmentation strategy of ground-truth sampling. The model is trained on 8 RTX A6000 GPUs with a batch size of 32. Other settings for training and inference adhere strictly to DSVT [47]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.65, + 0.553, + 0.667 + ], + "angle": 0, + "content": "4.3 Comparison with State-of-the-Arts" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.844 + ], + "angle": 0, + "content": "Results on Waymo Open Dataset (WOD). We conduct a detailed comparison of ScatterFormer against the published results on the validation set of the Waymo Open Dataset. In line with standard practices, we distinctly categorize and list the methods utilizing single-frame and multi-frame approaches. To ensure comprehensive coverage, we also include comparisons with methods incorporating long-term temporal modeling, such as those described in [4,15,61]. As demonstrated in Table 1, our single-stage model outperforms most two-stage methods and achieves better performance than state-of-the-art methods such as DSVT [47] and HEDNet [53]. Moreover, our model achieves 76.0 and 76.7 level 2 mAPH on 3 and 4 frame settings, respectively, outperforming previous multiframe methods by a margin of 1.1. Furthermore, as shown in Figure 6," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.635, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "ScatterFormer" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 1: Performance comparison on the validation set of Waymo Open Dataset. Symbol \\* denotes the methods with temporal modeling, and - means that the result is not available." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.2, + 0.789, + 0.558 + ], + "angle": 0, + "content": "
MethodFramesALL (3D mAPH)Vehicle (AP/APH)Pedestrian (AP/APH)Cyclist (AP/APH)
L1L2L1L2L1L2L1L2
SECOND [49]163.0557.2372.27/71.6963.85/63.3368.70/58.1860.72/51.3160.62/59.2858.34/57.05
PointPillar [18]163.3357.5371.60/71.0063.10/62.5070.60/56.7062.90/50.2064.40/62.3061.90/59.90
IA-SSD [56]164.4858.0870.53/69.6761.55/60.8069.38/58.4760.30/50.7367.67/65.3064.98/62.71
LiDAR R-CNN [20]166.2060.1073.50/73.0064.70/64.2071.20/58.7063.10/51.7068.60/66.9066.10/64.40
RSN [41]1--75.10/74.6066.00/65.5077.80/72.7068.30/63.70--
Part-A2 [38]170.2563.8477.05/76.5168.47/67.9775.24/66.8766.18/58.6268.60/67.3666.13/64.93
Centerpoint [52]1-65.50-- /66.20--/62.60--/67.60
VoTR [24]1--74.95/74.2565.91/65.29----
VoxSeT [14]172.2466.2274.50/74.0365.99/65.5680.03/72.4272.45/65.3971.56/70.2968.95/67.73
SST-1f [8]1--76.22/75.7968.04/67.6481.39/74.0572.82/65.93--
SWFormer-1f [40]1--77.8/77.369.2/68.880.9/72.772.5/64.9--
PV-RCNN [35]169.6363.3377.51/76.8968.98/68.4175.01/65.6566.04/57.6167.81/66.3565.39/63.98
PillarNet [34]174.6068.4379.09/78.5970.92/70.4680.59/74.0172.28/66.1772.29/71.2169.72/68.67
PV-RCNN++ [35]175.2168.6179.10/78.6370.34/69.9180.62/74.6271.86/66.3073.49/72.3870.70/69.62
FlatFormer [23]1-67.2-69.0/68.6-71.5/65.3-68.6/67.5
PillarNext-1f [19]175.7469.7478.40/77.9070.27/69.8182.53/77.1474.90/69.8073.21/72.2070.58/69.62
VoxelNext [5]176.370.178.2/77.769.9/69.481.5/76.373.5/68.676.1/74.973.3/72.2
FSD [9]177.370.879.2/78.870.5/70.182.6/77.373.9/69.177.1/76.074.4/73.3
DSVT-1f [47]178.272.179.7/79.371.4/71.083.7/78.976.1/71.577.5/76.574.6/73.7
HEDNet-1f [53]179.573.481.1/80.673.2/72.784.4/80.076.8/72.678.7/77.775.8/74.9
ScatterFormer (ours)179.773.881.0/80.573.1/72.784.5/79.977.0/72.679.9/78.977.1/76.1
SST-3f [8]3--78.66/78.2169.98/69.5783.81/80.1475.94/72.37--
FlatFormer-3f [23]3-72.0-71.4/71.0-74.5/71.3-74.7/73.7
SWFormer-3f [40]3--79.4/78.971.1/70.682.9/79.074.8/71.1--
PillarNext-3f [19]380.074.580.6/80.172.9/72.485.0/82.178.0/75.278.9/77.976.7/75.7
DSVT-4f [47]481.375.681.8/81.474.1/73.685.6/82.878.6/75.980.4/79.678.1/77.3
ScatterFormer-3f (ours)381.776.082.0/81.475.0/74.185.7/83.279.0/76.180.6/79.578.1/77.7
ScatterFormer-4f (ours)481.976.782.4/81.975.2/74.786.1/83.479.3/76.681.2/80.478.9/78.1
*3D-MAN [51]16--74.53/74.0367.61/67.1471.7/67.762.6/59.0--
*CenterFormer [61]477.073.278.1/77.673.4/72.981.7/78.677.2/74.275.6/74.873.4/72.6
*CenterFormer [61]877.373.778.8/78.374.3/73.882.1/79.377.8/75.075.2/74.473.2/72.3
*MPPNet [4]479.8374.2281.54/81.0674.07/73.6184.56/81.9477.20/74.6777.15/76.5075.01/74.38
*MPPNet [4]1680.4074.8582.74/82.2875.41/74.9684.69/82.2577.43/75.0677.28/76.6675.13/74.52
*MSF [15]480.2074.6281.36/80.8773.81/73.3585.05/82.1077.92/75.1178.40/77.6176.17/75.40
*MSF [15]880.6575.4682.83/82.0175.76/75.3185.24/82.2178.32/75.6178.52/77.7476.32/75.47
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.567, + 0.788, + 0.703 + ], + "angle": 0, + "content": "our method significantly enhances the detection runtime. This improvement is attributed to our use of Linear Attention, which eliminates the need for extensive voxel sorting and padding operations. Furthermore, as presented in Table 2, ScatterFormer achieves the highest Level 1 and Level 2 mAPH scores in all categories of the Waymo test set, highlighting its superior detection quality and precision. Remarkably, our approach, as a one-shot method, delivers performance that is on par with methods that require temporal modeling. This finding underscores the critical importance of network design and architecture in achieving high-performance architecture." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.789, + 0.84 + ], + "angle": 0, + "content": "Results on NuScenes. We compare ScatterFormer with the previous best performing methods on the nuScenes dataset. As shown in Table 3, ScatterFormer achieves the state-of-the-art performance in terms of val NDS (72.4) and mAP (68.3), surpassing HEDNet [53] by 1.0 and 1.6 respectively. Compared to TransFusion [2], ScatterFormer achieves a significant improvement of \\(2.8\\%\\) in mAP, demonstrating that its backbone network is superior to the sparse convolutional network. ScatterFormer exhibits particularly strong performance on categories like Bus, Bike, Construction Vehicle and Trailer, suggesting its ability to effectively capture and encode detailed contextual features." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "C. He et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.173 + ], + "angle": 0, + "content": "Table 2: Performance comparison on the test set of Waymo Open Dataset. \\(\\cdot\\) means that the result is not available." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.179, + 0.787, + 0.309 + ], + "angle": 0, + "content": "
MethodALL (3D mAPH)Vehicle (AP/APH)Pedestrian (AP/APH)Cyclist (AP/APH)
L1L2L1L2L1L2L1L2
PointPillar [18]--68.1060.1068.00/55.5061.40/50.10--
StarNet [26]--61.0054.5067.80/59.9061.10/54.00--
M3DETR [11]67.161.977.7/77.170.5/70.068.2/58.560.6/52.067.3/65.765.3/63.8
3D-MAN [51]--78.2869.9869.97/65.9863.98/60.26--
PV-RCNN++ [36]75.770.281.6/81.273.9/73.580.4/75.074.1/69.071.9/70.869.3/68.2
CenterPoint [52]77.271.981.1/80.673.4/73.080.5/77.374.6/71.574.6/73.772.2/71.3
RSN [41]--80.3071.6078.90/75.6070.70/67.80--
SST-3f [8]78.372.881.0/80.673.1/72.783.3/79.776.9/73.575.7/74.673.2/72.2
HEDNet [53]79.0573.7783.78/83.3976.33/75.9683.46/78.9877.53/73.2575.86/74.7873.13/72.09
PillarNext-3f [19]79.0074.0983.28/82.8376.18/75.7684.40/81.4478.84/75.9873.77/72.7371.56/70.55
ScatterFormer-4f (ours)80.7175.8485.60/85.1378.34/78.0784.58/81.6479.12/76.3276.20/75.3574.02/73.13
" + }, + { + "type": "table_caption", + "bbox": [ + 0.295, + 0.325, + 0.707, + 0.339 + ], + "angle": 0, + "content": "Table 3: The performance on the validation set of NuScenes." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.345, + 0.787, + 0.455 + ], + "angle": 0, + "content": "
MethodNDSmAPCarTruckBusT.L.C.V.Ped.M.T.BikeT.C.B.R.
CenterPoint [52]66.559.284.957.470.738.116.985.159.042.069.868.3
VoxelNeXt [5]66.760.583.955.570.538.121.184.662.850.069.469.4
TransFusion-L [2]70.165.586.960.873.143.425.287.572.957.377.270.3
PillarNext [19]68.462.285.057.467.635.620.686.868.653.177.369.7
DSVT [47]71.166.487.462.675.942.125.388.274.858.777.870.9
HEDNet [53]71.466.787.760.677.850.728.987.174.356.876.366.9
ScatterFormer (ours)72.468.388.665.479.345.429.188.774.761.578.272.3
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.468, + 0.39, + 0.484 + ], + "angle": 0, + "content": "4.4 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.788, + 0.599 + ], + "angle": 0, + "content": "In this section, we conduct ablation experiments in ScatterFormer using \\(20\\%\\) of the Waymo training and validation data. Each model was trained for 24 epochs and run on default parameters. Table 4 shows four different configurations of ScatterFormer: (a) removing the Scattered Linear Attention module (SLA), (b) removing the Cross Window Interaction (CWI) module, (c) replacing the CWI with Shifted Window (SW) approach, and (d) removing Conditional Position Encoding (CPE) module." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.788, + 0.749 + ], + "angle": 0, + "content": "As can be seen in Table 4, configuration (a) results in ScatterFormer experiencing a decrease of \\((3.1\\%, 3.9\\%, \\text{and} 3.2\\%)\\) APH on three categories. This highlights the effectiveness of the SLA, as it successfully captures long-range contexts and facilitates dynamic feature extraction for object detection. In configuration (b), ScatterFormer, which lacks cross-window interaction capability, exhibits a performance drop of \\((1.9\\%, 1.0\\%, \\text{and} 2.0\\%)\\) APH across the three categories. While the Shifted Window approach in configuration (c) enhances performance, the improvement is not as substantial as that achieved by the proposed CWI module. The results in configuration (d) indicate that CPE contributes to a certain degree of performance enhancement." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Window Size. In Table 5, We tested the impact of different window sizes on the detection performance of various target types. It can be observed that ScatterFormer is not highly sensitive to window size. Slightly increasing the window size can enhance the detection performance for small targets, indicating that for objects like pedestrians and bicycles, where the point cloud is overly sparse, appropriate contextual information can enhance the recognition of these" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.634, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "ScatterFormer" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 4: Ablation experiments on the validation set of Waymo Open Dataset. \"SLA\" and \"CWT\" refer to the proposed Scattered Linear Attention and Cross-Window Interaction modules, respectively. \"CPE\" refers to Conditional Position Embedding. AP and APH scores on LEVEL2 are reported." + }, + { + "type": "table", + "bbox": [ + 0.314, + 0.206, + 0.687, + 0.304 + ], + "angle": 0, + "content": "
AblationVeh.Ped.Cyc.
-baseline71.6/71.176.0/70.974.2/73.1
(a)w/o SLA module68.5/68.073.3/67.071.0/69.9
(b)w/o CWI module69.6/69.274.0/69.972.1/71.1
(c)CWI → SW69.9/69.575.0/70.473.8/72.4
(d)w/o CPE70.2/69.873.3/67.272.4/70.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.315, + 0.789, + 0.343 + ], + "angle": 0, + "content": "Table 5: Comparison of model performance using different window sizes at various stages. The APH (L2) scores on different categories are reported." + }, + { + "type": "table", + "bbox": [ + 0.349, + 0.348, + 0.648, + 0.431 + ], + "angle": 0, + "content": "
Window SizeRegion SizeVeh. Ped. Cyc.
103.20 m70.2 69.2 71.0
123.84 m71.1 70.9 73.1
144.48 m71.3 69.9 72.5
165.12 m71.0 69.2 72.3
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.45, + 0.788, + 0.495 + ], + "angle": 0, + "content": "objects. However, excessively large windows can degrade the performance. We speculate that this is because larger windows encompass more tokens, thereby mitigating the focusing power of linear attention." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.502, + 0.789, + 0.669 + ], + "angle": 0, + "content": "Kernel Design in CWI. Table 6 presents the performance under various kernel configurations in the Cross-Window Interaction (CWI) module. Increasing the number of parameters in the kernel significantly enhances performance. For example, using a large \\(7 \\times 7\\) kernel brings notable performance improvements, but also introduces more latency. Additionally, larger kernel strides enhance the exchange rate of tokens across windows. For example, using dilated convolutions in CWI performs slightly better than using non-dilated convolutions. Instead of scaling up large, we decompose large kernel into more efficient 1D kernels, which achieves the optimal latency while keeping comparable performance. As the current Sconv library lacks support for depth-wise convolution, we have developed a modified version to address this limitation.\\(^4\\)" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.689, + 0.634, + 0.705 + ], + "angle": 0, + "content": "4.5 Comparison with Different Linear Attentions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.711, + 0.788, + 0.819 + ], + "angle": 0, + "content": "We further explored the performance of ScatterFormer by conducting experiments with several other linear attention mechanisms, including Gated Linear Attention [32], Cross-covariance Attention (XCA) [1], and Focused Linear Attention [13]. The results are summarized in Table 7. Our findings suggest that the window-based attention framework employed by ScatterFormer is not highly sensitive to the specific form of linear attention used. Notably, Efficient Attention [3] performs slightly worse than XCA [1] in the Vehicle class but better in the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.516, + 0.841 + ], + "angle": 0, + "content": "4 https://github.com/skyhehe123/spconvv" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "C. He et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 6: Performance with various kernel configurations in the Cross-Window Integration (CWI) module. Each configuration is specified by the notation \"KxDy\", where \"Kx\" refers to the kernel size, and \"Dy\" denotes the dilation rate. The APH (L2) scores on different categories are reported." + }, + { + "type": "table", + "bbox": [ + 0.354, + 0.207, + 0.642, + 0.288 + ], + "angle": 0, + "content": "
KernelLatencyVeh.Ped.Cyc.
Conv1DK13 × 247ms71.170.973.1
Conv2DK5D149ms69.569.071.1
Conv2DK5D249ms70.770.472.8
Conv2DK7D165ms71.571.072.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.294, + 0.786, + 0.322 + ], + "angle": 0, + "content": "Table 7: Performance with different linear attention designs. The APH (L2) scores on different categories are reported." + }, + { + "type": "table", + "bbox": [ + 0.354, + 0.327, + 0.641, + 0.408 + ], + "angle": 0, + "content": "
Linear AttentionVeh.Ped.Cyc.
Efficient Attn [3]71.170.973.1
Gated Linear Attn [32]69.969.072.5
Focused Linear Attn [13]70.469.572.3
XCA [1]71.570.472.8
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.422, + 0.786, + 0.467 + ], + "angle": 0, + "content": "Pedestrian and Cyclist classes. Focused Linear Attention introduces additional computational overhead, while Gated Linear Attention adds more learnable parameters to the model." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.484, + 0.357, + 0.498 + ], + "angle": 0, + "content": "4.6 Limitations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.511, + 0.788, + 0.618 + ], + "angle": 0, + "content": "ScatterFormer relies on our customized operators, which have not yet been implemented as plugins in TensorRT. Therefore, deploying ScatterFormer on in-vehicle devices will require additional engineering efforts. Despite this, ScatterFormer is more efficient than current models based on sparse convolution and traditional attention when used on consumer-grade GPU cards. Furthermore, ScatterFormer can be optimized by dynamically partitioning matrices according to different GPU architectures to leverage TensorCore for hardware acceleration." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.641, + 0.36, + 0.657 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We introduced ScatterFormer, an innovative architecture designed for 3D object detection using point clouds, specifically targeting the challenges associated with processing sparse and unevenly distributed data from LiDAR sensors. The cornerstone of our approach is the Scattered Linear Attention (SLA) module, which effectively addresses the limitations of conventional attention mechanisms in managing voxel features of varying lengths. SLA ingeniously combines linear attention with a chunk-wise matrix multiplication algorithm, tailored to meet the distinct requirements of processing voxels grouped by windows. By integrating SLA with a novel cross-window interaction (CWI) module, ScatterFormer achieves higher accuracy and lower latency, surpassing traditional transformer-based and sparse CNN-based detectors in extensive 3D detection tasks." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.634, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "ScatterFormer" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.324, + 0.161 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.176, + 0.787, + 0.232 + ], + "angle": 0, + "content": "1. Ali, A., Touvron, H., Caron, M., Bojanowski, P., Douze, M., Joulin, A., Laptev, I., Neverova, N., Synnaeve, G., Verbeek, J., et al.: Xcit: Cross-covariance image transformers. Advances in neural information processing systems 34, 20014-20027 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.233, + 0.787, + 0.286 + ], + "angle": 0, + "content": "2. Bai, X., Hu, Z., Zhu, X., Huang, Q., Chen, Y., Fu, H., Tai, C.L.: Transfusion: Robust lidar-camera fusion for 3d object detection with transformers. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1090-1099 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.288, + 0.787, + 0.328 + ], + "angle": 0, + "content": "3. Cai, H., Gan, C., Han, S.: Efficientvit: Enhanced linear attention for high-resolution low-computation visual recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.329, + 0.787, + 0.37 + ], + "angle": 0, + "content": "4. Chen, X., Shi, S., Zhu, B., Cheung, K.C., Xu, H., Li, H.: Mppnet: Multi-frame feature intertwining with proxy points for 3d temporal object detection. In: European Conference on Computer Vision. pp. 680-697. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.371, + 0.787, + 0.412 + ], + "angle": 0, + "content": "5. Chen, Y., Liu, J., Zhang, X., Qi, X., Jia, J.: Voxelnext: Fully sparse voxelnet for 3d object detection and tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21674-21683 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.412, + 0.787, + 0.439 + ], + "angle": 0, + "content": "6. Chu, X., Tian, Z., Zhang, B., Wang, X., Shen, C.: Conditional positional encodings for vision transformers. ICLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.44, + 0.787, + 0.495 + ], + "angle": 0, + "content": "7. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. ICLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.496, + 0.787, + 0.55 + ], + "angle": 0, + "content": "8. Fan, L., Pang, Z., Zhang, T., Wang, Y.X., Zhao, H., Wang, F., Wang, N., Zhang, Z.: Embracing single stride 3d object detector with sparse transformer. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8458-8468 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.55, + 0.787, + 0.578 + ], + "angle": 0, + "content": "9. Fan, L., Wang, F., Wang, N., Zhang, Z.: Fully sparse 3d object detection. Advances in Neural Information Processing Systems 35, 351-363 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.579, + 0.787, + 0.619 + ], + "angle": 0, + "content": "0. Fan, L., Xiong, X., Wang, F., Wang, N., Zhang, Z.: Rangedet: In defense of range view for lidar-based 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2918-2927 (October 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.62, + 0.787, + 0.674 + ], + "angle": 0, + "content": "1. Guan, T., Wang, J., Lan, S., Chandra, R., Wu, Z., Davis, L., Manocha, D.: M3detr: Multi-representation, multi-scale, mutual-relation 3d object detection with transformers. In: Proceedings of the IEEE/CVF winter conference on applications of computer vision. pp. 772-782 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.675, + 0.787, + 0.728 + ], + "angle": 0, + "content": "2. Guo, M.H., Cai, J.X., Liu, Z.N., Mu, T.J., Martin, R.R., Hu, S.M.: Pct: Point cloud transformer. Computational Visual Media 7(2), 187-199 (Apr 2021). https://doi.org/10.1007/s41095-021-0229-5, http://dx.doi.org/10.1007/s41095-021-0229-5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.73, + 0.787, + 0.771 + ], + "angle": 0, + "content": "3. Han, D., Pan, X., Han, Y., Song, S., Huang, G.: Flatten transformer: Vision transformer using focused linear attention. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5961-5971 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.787, + 0.813 + ], + "angle": 0, + "content": "4. He, C., Li, R., Li, S., Zhang, L.: Voxel set transformer: A set-to-set approach to 3d object detection from point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8417-8427 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.814, + 0.787, + 0.84 + ], + "angle": 0, + "content": "5. He, C., Li, R., Zhang, Y., Li, S., Zhang, L.: Msf: Motion-guided sequential fusion for efficient 3d object detection from point cloud sequences. In: Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.176, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "C. He et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.241, + 0.147, + 0.786, + 0.175 + ], + "angle": 0, + "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5196-5205 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.175, + 0.787, + 0.216 + ], + "angle": 0, + "content": "16. He, C., Zeng, H., Huang, J., Hua, X.S., Zhang, L.: Structure aware single-stage 3d object detection from point cloud. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11873-11882 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.216, + 0.787, + 0.257 + ], + "angle": 0, + "content": "17. Katharopoulos, A., Vyas, A., Pappas, N., Fleuret, F.: Transformers are rnns: Fast autoregressive transformers with linear attention. In: International conference on machine learning. pp. 5156-5165. PMLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.257, + 0.787, + 0.298 + ], + "angle": 0, + "content": "18. Lang, A.H., Vora, S., Caesar, H., Zhou, L., Yang, J., Beijbom, O.: Pointpillars: Fast encoders for object detection from point clouds. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 12697-12705 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.298, + 0.787, + 0.338 + ], + "angle": 0, + "content": "19. Li, J., Luo, C., Yang, X.: Rethinking network designs for 3d object detection in lidar point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17567-17576 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.338, + 0.787, + 0.379 + ], + "angle": 0, + "content": "20. Li, Z., Wang, F., Wang, N.: Lidar r-cnn: An efficient and universal 3d object detector. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.787, + 0.433 + ], + "angle": 0, + "content": "21. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 10012-10022 (October 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.433, + 0.787, + 0.473 + ], + "angle": 0, + "content": "22. Liu, Z., Zhang, Z., Cao, Y., Hu, H., Tong, X.: Group-free 3d object detection via transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2949-2958 (October 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.787, + 0.514 + ], + "angle": 0, + "content": "23. Liu, Z., Yang, X., Tang, H., Yang, S., Han, S.: Flatformer: Flattened window attention for efficient point cloud transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1200-1211 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.514, + 0.787, + 0.555 + ], + "angle": 0, + "content": "24. Mao, J., Xue, Y., Niu, M., Bai, H., Feng, J., Liang, X., Xu, H., Xu, C.: Voxel transformer for 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 3164-3173 (October 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.555, + 0.787, + 0.596 + ], + "angle": 0, + "content": "25. Misra, I., Girdhar, R., Joulin, A.: An end-to-end transformer model for 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2906-2917 (October 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.596, + 0.787, + 0.637 + ], + "angle": 0, + "content": "26. Ngiam, J., Caine, B., Han, W., Yang, B., Chai, Y., Sun, P., Zhou, Y., Yi, X., Alsharif, O., Nguyen, P., et al.: Starnet: Targeted computation for object detection in point clouds. arXiv preprint arXiv:1908.11069 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.637, + 0.787, + 0.678 + ], + "angle": 0, + "content": "27. Pan, X., Xia, Z., Song, S., Li, L.E., Huang, G.: 3d object detection with pointformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7463-7472 (June 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.678, + 0.787, + 0.718 + ], + "angle": 0, + "content": "28. Qi, C.R., Litany, O., He, K., Guibas, L.J.: Deep though voting for 3d object detection in point clouds. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (October 2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.718, + 0.787, + 0.759 + ], + "angle": 0, + "content": "29. Qi, C.R., Liu, W., Wu, C., Su, H., Guibas, L.J.: Frustum pointnets for 3d object detection from rgb-d data. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 918-927 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.787, + 0.799 + ], + "angle": 0, + "content": "30. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 652-660 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "31. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In: Advances in neural information processing systems. pp. 5099-5108 (2017)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.635, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ScatterFormer" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "32. Qin, Z., Li, D., Sun, W., Sun, W., Shen, X., Han, X., Wei, Y., Lv, B., Yuan, F., Luo, X., Qiao, Y., Zhong, Y.: Scaling transnormer to 175 billion parameters (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.178, + 0.787, + 0.232 + ], + "angle": 0, + "content": "33. Sheng, H., Cai, S., Liu, Y., Deng, B., Huang, J., Hua, X.S., Zhao, M.J.: Improving 3d object detection with channel-wise transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2743-2752 (October 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.234, + 0.785, + 0.274 + ], + "angle": 0, + "content": "34. Shi, G., Li, R., Ma, C.: Pillarnet: Real-time and high-performance pillar-based 3d object detection. In: European Conference on Computer Vision. pp. 35-52. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.276, + 0.785, + 0.317 + ], + "angle": 0, + "content": "35. Shi, S., Guo, C., Jiang, L., Wang, Z., Shi, J., Wang, X., Li, H.: Pv-rcnn: Pointvoxel feature set abstraction for 3d object detection. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.319, + 0.785, + 0.374 + ], + "angle": 0, + "content": "36. Shi, S., Jiang, L., Deng, J., Wang, Z., Guo, C., Shi, J., Wang, X., Li, H.: Pvrnn++: Point-voxel feature set abstraction with local vector representation for 3d object detection. International Journal of Computer Vision 131(2), 531-551 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.376, + 0.785, + 0.416 + ], + "angle": 0, + "content": "37. Shi, S., Wang, X., Li, H.: Pointrecnn: 3d object proposal generation and detection from point cloud. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 770-779 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.418, + 0.785, + 0.459 + ], + "angle": 0, + "content": "38. Shi, S., Wang, Z., Wang, X., Li, H.: Part-a^2 net: 3d part-aware and aggregation neural network for object detection from point cloud. arXiv preprint arXiv:1907.03670 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.461, + 0.785, + 0.5 + ], + "angle": 0, + "content": "39. Shi, W., Rajkumar, R.: Point-gnn: Graph neural network for 3d object detection in a point cloud. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1711-1719 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.502, + 0.785, + 0.543 + ], + "angle": 0, + "content": "40. Sun, P., Tan, M., Wang, W., Liu, C., Xia, F., Leng, Z., Anguelov, D.: Swformer: Sparse window transformer for 3d object detection in point clouds. In: European Conference on Computer Vision. pp. 426-442. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.546, + 0.785, + 0.6 + ], + "angle": 0, + "content": "41. Sun, P., Wang, W., Chai, Y., Elsayed, G., Bewley, A., Zhang, X., Sminchisescu, C., Anguelov, D.: Rsn: Range sparse net for efficient, accurate lidar 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5725-5734 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.602, + 0.785, + 0.642 + ], + "angle": 0, + "content": "42. Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., Wojna, Z.: Rethinking the inception architecture for computer vision. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2818-2826 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.644, + 0.785, + 0.672 + ], + "angle": 0, + "content": "43. Team, O.D.: Openpcdet: An open-source toolbox for 3d object detection from point clouds. https://github.com/open-mmlab/OpenPCDet (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.673, + 0.785, + 0.754 + ], + "angle": 0, + "content": "44. Tilet, P., Kung, H.T., Cox, D.: Triton: An intermediate language and compiler for tiled neural network computations. In: Proceedings of the 3rd ACM SIGPLAN International Workshop on Machine Learning and Programming Languages. p. 10-19. MAPL 2019, Association for Computing Machinery, New York, NY, USA (2019). https://doi.org/10.1145/3315508.3329973, https://doi.org/10.1145/3315508.3329973" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.757, + 0.785, + 0.798 + ], + "angle": 0, + "content": "45. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jegou, H.: Training data-efficient image transformers & distillation through attention. In: International Conference on Machine Learning. vol. 139, pp. 10347-10357 (July 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "46. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: Advances in neural information processing systems. pp. 5998-6008 (2017)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "C. He et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.204 + ], + "angle": 0, + "content": "47. Wang, H., Shi, C., Shi, S., Lei, M., Wang, S., He, D., Schiele, B., Wang, L.: Dsvt: Dynamic sparse voxel transformer with rotated sets. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13520-13529 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.205, + 0.786, + 0.232 + ], + "angle": 0, + "content": "48. Wang, S., Li, B.Z., Khabsa, M., Fang, H., Ma, H.: Linformer: Self-attention with linear complexity (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.233, + 0.786, + 0.259 + ], + "angle": 0, + "content": "49. Yan, Y., Mao, Y., Li, B.: Second: Sparsely embedded convolutional detection. Sensors 18(10), 3337 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.26, + 0.787, + 0.3 + ], + "angle": 0, + "content": "50. Yang, B., Luo, W., Urtasun, R.: Pixor: Real-time 3d object detection from point clouds. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 7652-7660 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.301, + 0.786, + 0.342 + ], + "angle": 0, + "content": "51. Yang, Z., Zhou, Y., Chen, Z., Ngiam, J.: 3d-man: 3d multi-frame attention network for object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1863-1872 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.343, + 0.786, + 0.383 + ], + "angle": 0, + "content": "52. Yin, T., Zhou, X., Krahenbuhl, P.: Center-based 3d object detection and tracking. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11784-11793 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.384, + 0.786, + 0.424 + ], + "angle": 0, + "content": "53. Zhang, G., Chen, J., Gao, G., Li, J., Hu, X.: Hednet: A hierarchical encoder-decoder network for 3d object detection in point clouds. arXiv preprint arXiv:2310.20234 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.426, + 0.786, + 0.466 + ], + "angle": 0, + "content": "54. Zhang, G., Fan, L., He, C., Lei, Z., Zhang, Z., Zhang, L.: Voxel mamba: Group-free state space models for point cloud based 3d object detection. arXiv preprint arXiv:2406.10700 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.467, + 0.786, + 0.508 + ], + "angle": 0, + "content": "55. Zhang, Y., Huang, D., Wang, Y.: Pc-rgnn: Point cloud completion and graph neural network for 3d object detection. In: Proceedings of the AAAI conference on artificial intelligence. vol. 35, pp. 3430-3437 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.509, + 0.786, + 0.563 + ], + "angle": 0, + "content": "56. Zhang, Y., Hu, Q., Xu, G., Ma, Y., Wan, J., Guo, Y.: Not all points are equal: Learning highly efficient point-based detectors for 3d lidar point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18953-18962 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.565, + 0.786, + 0.605 + ], + "angle": 0, + "content": "57. Zhao, H., Jiang, L., Jia, J., Torr, P.H., Koltun, V.: Point transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 16259-16268 (October 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.606, + 0.786, + 0.646 + ], + "angle": 0, + "content": "58. Zheng, W., Tang, W., Jiang, L., Fu, C.W.: Se-ssd: Self-ensembling single-stage object detector from point cloud. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14494-14503 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.647, + 0.786, + 0.688 + ], + "angle": 0, + "content": "59. Zhou, C., Zhang, Y., Chen, J., Huang, D.: Octr: Octtree-based transformer for 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5166-5175 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.689, + 0.786, + 0.73 + ], + "angle": 0, + "content": "60. Zhou, Y., Tuzel, O.: Voxelnet: End-to-end learning for point cloud based 3d object detection. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4490-4499 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.731, + 0.786, + 0.772 + ], + "angle": 0, + "content": "61. Zhou, Z., Zhao, X., Wang, Y., Wang, P., Foroosh, H.: Centerformer: Center-based transformer for 3d object detection. In: European Conference on Computer Vision. pp. 496-513. Springer (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.772 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/ScatterFormer_ Efficient Voxel Transformer with Scattered Linear Attention/1045fb37-bbfd-4adc-8de1-f127bb3078ca_origin.pdf b/2024/ScatterFormer_ Efficient Voxel Transformer with Scattered Linear Attention/1045fb37-bbfd-4adc-8de1-f127bb3078ca_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f609f154b8837a385cfb526345979b84129ddefa --- /dev/null +++ b/2024/ScatterFormer_ Efficient Voxel Transformer with Scattered Linear Attention/1045fb37-bbfd-4adc-8de1-f127bb3078ca_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:db12b68cfda0954e73a87dc1254d1ab01cf85f8c27275bcbf6d378bcfb1e138c +size 1872657 diff --git a/2024/ScatterFormer_ Efficient Voxel Transformer with Scattered Linear Attention/full.md b/2024/ScatterFormer_ Efficient Voxel Transformer with Scattered Linear Attention/full.md new file mode 100644 index 0000000000000000000000000000000000000000..399953eff63bcba762b73255e6b6d34f8f602f14 --- /dev/null +++ b/2024/ScatterFormer_ Efficient Voxel Transformer with Scattered Linear Attention/full.md @@ -0,0 +1,287 @@ +# ScatterFormer: Efficient Voxel Transformer with Scattered Linear Attention + +Chenhang He $^{1}$ , Ruihuang Li $^{1,2}$ , Guowen Zhang $^{1}$ , and Lei Zhang $^{1,2}$ + +1 The Hong Kong Polytechnic University, Hong Kong SAR, China + +$^{2}$ OPPO Research, Shenzhen, China chenhang.he@polyu.edu.hk csrhli@comp.polyu.edu.hk guowen.zhang@connect.polyu.edu.hk cslzhang@comp.polyu.edu.hk + +Abstract. Window-based transformers excel in large-scale point cloud understanding by capturing context-aware representations with affordable attention computation in a more localized manner. However, the sparse nature of point clouds leads to a significant variance in the number of voxels per window. Existing methods group the voxels in each window into fixed-length sequences through extensive sorting and padding operations, resulting in a non-negligible computational and memory overhead. In this paper, we introduce ScatterFormer, which to the best of our knowledge, is the first to directly apply attention to voxels across different windows as a single sequence. The key of ScatterFormer is a Scattered Linear Attention (SLA) module, which leverages the pre-computation of key-value pairs in linear attention to enable parallel computation on the variable-length voxel sequences divided by windows. Leveraging the hierarchical structure of GPUs and shared memory, we propose a chunkwise algorithm that reduces the SLA module's latency to less than 1 millisecond on moderate GPUs. Furthermore, we develop a cross-window interaction module that improves the locality and connectivity of voxel features across different windows, eliminating the need for extensive window shifting. Our proposed ScatterFormer demonstrates 73.8 mAP (L2) on the Waymo Open Dataset and 72.4 NDS on the NuScenes dataset, running at an outstanding detection rate of 23 FPS. The code is available at https://github.com/skyhehe123/ScatterFormer. + +Keywords: 3D Object Detection $\cdot$ Voxel Transformer + +# 1 Introduction + +In the field of 3D object detection, the use of point clouds has become increasingly popular, especially for providing accurate and reliable perception results in autonomous systems. Unlike image data, point clouds obtained from LiDAR are often sparse and nonuniformly distributed, with varying density depending on their distances from the sensor. Earlier approaches utilize point-cloud operators + +![](images/49d5b409f15b627611ec4e71b2a9f0cabab4fb09cb602404b2db8501d07c8322.jpg) + +![](images/5adf862f1d88200a3e1cd0f498880ed5403aaf17d034e31c04402a7a5c650d03.jpg) +Fig. 1: Illustration of (a) group-based attention and (b) our scattered linear attention on variable-length sequences. The square in different colors represent voxels in different windows and the square with a cross represents the padding voxel. + +like PointNet++ [31] for feature extraction in continuous space [37]. Some approaches [5,9,16,18,49,50,52,60] have transformed point clouds into voxel grids, which are then efficiently processed by a sparse convolutional neural network (SpCNN) [49]. + +Recently, the success of vision transformers [7, 21] has motivated a number of attention-based methods to process indoor point clouds [12, 22, 24, 25, 27, 57]. Inspired by SwinTransformer [21], various studies [8, 14, 23, 47] have advanced the application of window-based Transformers in large-scale 3D detection tasks within outdoor environments, achieving outstanding performance and highlighting their potential as alternatives to SpCNN. However, due to the inherent sparsity of point clouds, the number of features grouped by windows can vary significantly, hampering the parallelism in attention computation. To resolve this issue, SST [8] and SWFormer [40] group the voxels within a window into different batches and manages the attention computation in a hybrid serial-parallel manner. More effectively, DSVT [47] alternatively sort the voxels within a window from different axes and partition them into sequences of fixed length, which allow parallel computation on sparse voxels. Albeit effective, these group-based methods, as shown in Figure 1(a), require extensive sorting and padding operations, incurring substantial memory and computational overhead. + +In this paper, we delve into the window-based voxel transformer where the voxels grouped by windows form variable-length sequences $\{X_1\in \mathbb{R}^{n_1\times d},X_2\in$ $\mathbb{R}^{n_2\times d},\dots,X_k\in \mathbb{R}^{n_k\times d}\}$ . Their corresponding self-attention matrices $\{A_{1}\in$ $\mathbb{R}^{n_1\times n_1}$ $A_{2}\in \mathbb{R}^{n_{2}\times n_{2}}$ ..., $A_{k}\in \mathbb{R}^{n_{k}\times n_{k}}\}$ occupy irregular memory spaces, making it a challenge to perform attention in parallel for the voxels of the entire scene. Recent efforts on linear attention [1,13,17,48] have emerged a promising alternative to traditional attention. By approximating softmax operation with the kernel function, i.e., $\phi (Q)\cdot \phi (K)^{\top}\approx \mathrm{softmax}(QK^{\top})$ , we can change the computation order from $(Q\cdot K^{\top})\cdot V$ to $Q\cdot (K^{\top}\cdot V)$ . This not only results in linear complexity but also yields a compressed hidden state $S\in \mathbb{R}^{d\times d}$ that is invariant to the sequence length. Another attractive property of linear attention is that it + +can be converted into a "recurrent" form: + +$$ +o _ {t} = q _ {t} S _ {t}; \quad S _ {t} = S _ {t - 1} + k _ {t} ^ {\mathsf {T}} v _ {t} \tag {1} +$$ + +where the sequences can be divided into non-overlapping chunks, $\{q_j,k_j,v_j\}_{j = 1:t}$ . The output can be calculated by scanning over the sequence, based on an updated hidden state $S_{t}$ . This enables the use of chunk-level computation to address the inability to parallelize at the sequence level due to varying lengths. + +Inspired by this, we introduce the Scattered Linear Attention (SLA) module, which accommodates linear attention in a window-based voxel transformer. As shown in Figure 1(b), the SLA module treats the voxels of the entire scene into a single sequence and processes them directly without padding voxels. Using the recurrent form of linear attention, we develop an I/O-aware algorithm to further optimize matrix multiplication on voxel sequences. Specifically, we divide the voxel sequence into multiple chunks and loaded them into the shared memory (SRAM) of the GPU. The computation of the hidden state matrix in each window is then achieved through a series of chunk-wise matrix multiplications and cumulative sums. This optimized implementation significantly reduces I/O overhead and memory usage, resulting in extremely fast and memory-efficient attention computation. + +In addition, current window-based transformer models use window shifting to propagate the information across the windows. After performing an instruction-level analysis of existing implementations, we observe that the voxel permutation operations caused by window shifting consume significant computational overhead. To address this issue, we propose a cross-window interaction (CWI) module, which is composed of depth-wise convolutions with small 2D kernels and lengthy 1D kernels that allow the voxel features in each window to fully interact with the features in other windows. As a result, the proposed CWI module can improve both the locality and connectivity of the voxel features while requiring minimal computational effort. + +Building upon the SLA and CWI modules, we propose the ScatterFormer, an innovative voxel transformer for large-scale point cloud understanding. Our experiments show that ScatterFormer achieves linear complexity without compromising accuracy. It achieves superior results compared to the state-of-the-art model DSVT [47]. The latency of ScatterFormer is significantly lower than that of transformer-based detectors [8, 14, 47], which is comparable to that of sparse convolution-based detectors [52]. + +In conclusion, we delve into attention on voxels grouped by windows, highlighting the challenges in memory allocation and computation for variable-length sequences. Then, we introduce an SLA module, which can directly process the voxels grouped by windows by facilitating the linear attention formula. A chunkwise matrix multiplication algorithm is proposed to further accelerate the attention computation in SLA. Finally, we present ScatterFormer, which has linear complexity and can efficiently process large-scale LiDAR scenes, achieving better accuracy with lower latency compared to existing transformer-based detectors. + +# 2 Related Work + +# 2.1 Point Cloud-based 3D Object Detection + +There exist two primary point-cloud representations in 3D object detection, i.e., point-based and voxel-based ones. In point-based methods [28, 29, 37, 39, 55], point clouds are first passed through a point-based backbone network [30], in which the points are gradually sampled and features are learned by point cloud operators. F-Pointnet [29] first employs PoinNet to detect 3D objects based on the frustums lifted by 2D proposals. PointRCNN [37] direct generates 3D proposals from point-based features with foreground segmentation. VoteNet [28] clusters objects from the surface points using a deep Hough voting method. Other point operators based on the point graph [39, 55] and the range view [10] have also been developed for point cloud processing. + +Point-based methods are primarily limited by their inference efficiency and the absence of contextual features in a continuous space. On the other hand, voxel-based approaches [5, 16, 18, 49, 52, 58, 60] transform the entire point clouds into regular grids through voxelization, showcasing superior efficiency and context representation. VoxelNet [60] proposes a voxel feature encoding (VFE) module and combines it with 3D convolutions to extract voxel features in an end-to-end manner. SECOND [49] optimizes 3D convolution for sparse data, resulting in a significant reduction in both time and memory. PointPillars [18] demonstrates an efficient model by stacking voxels into vertical columns, subsequently processing them with naive 2D convolutions. There are also efforts [16, 35] that explore hybrid representations of point-based and voxel-based networks, demonstrating a better trade-off between speed and accuracy. However, a common limitation across these methods is their reliance on small convolution kernels with restricted receptive fields, making them less adept at capturing the global context for 3D object detection. + +In this paper, we shed light on the voxel-based methods and introduce a Scattered Linear Attention module that enables dynamic modeling among voxels within each window with linear complexity, thereby efficiently extracting window-based contextual features. + +# 2.2 Transformer on Point Cloud + +Inspired by the significant success of self-attention in NLP [46] and CV [7, 45], Transformers have been adapted for 3D vision due to their ability to capture long-range dependencies. The Point Transformers [27, 57] employ attention to modulate point clouds for classification and segmentation tasks. PCT [12] presents an optimized offset attention module, which, when combined with the implicit Laplacian operator and normal estimation, becomes more adept at point-cloud processing. Some methods like [8, 22] opt for voxels or key points to optimize latency. 3DETR [25] proposes a query-based 3D object detection scheme. CT3D [33] enhances the region-based network using a channel-wise + +![](images/b774ed792faa3e7e125652a569fc710665377ac999a8cc3e7153910cf7ae9407.jpg) +Fig. 2: The macro design of ScatterFormer. The backbone comprises a Conditional Position Encoding (CPE) and six transformer blocks. Each block is composed by a Scattered Linear Attention (SLA) module, a Cross-Window Interaction (CWI) module, and a Feed-Forward Network (FFN). + +Transformer framework. To achieve context-rich representations, several studies [8, 14, 23, 24, 47] have integrated the attention module into point- or voxel-based encoders. For instance, VoTr [24] utilizes dilated attention for expanded receptive fields; VoxSet [14] applies set attention for extracting point-based features in set-to-set translation; SST [8] employs local attention with shifted windows; and OcTr [59] adopts an Octree-based attention for efficient hierarchical context learning. Nevertheless, how to efficiently leverage global context from attention remains a challenge due to the intrinsic sparsity of point clouds. DSVT [47] and FlatFormer [23] group the voxels within each window into a series of fixed-length voxel sets, thus extracting the features in a fully parallel manner. Recently, a group-free state-space model [54] was proposed to directly process voxels as a sequence. However, these approaches more or less lose the spatial proximity and incur extensive computational overhead in grouping and sorting the voxels. + +# 3 Method + +The overall architecture of our proposed ScatterFormer is depicted in Figure 2. It begins with the input point clouds, which are voxelized and transformed into high-dimensional embeddings using a VFE layer [60]. These embeddings are then processed through Conditional Positional Encoding (CPE) using a shallow convolutional network [6]. The encoded features enter the ScatterFormer backbone, consisting of six ScatterFormer blocks. Each block includes a Scattered Linear Attention (SLA) module, a Cross-Window Interaction (CWI) module, and a Feed-Forward Network (FFN), interspersed with Batch Normalization layers and skip connections. After three ScatterFormer blocks, the voxel features are downsampled via a sparse convolutional layer. The downsampled features are then converted into pillar features [47], generating compact BEV features for bounding-box prediction. ScatterFormer ScatterFormer stands out by not requiring voxel features to be organized into fixed-length sets [8, 23, 47], enabling flexible attention computation across windows. Additionally, the CWI module obviates the need for window shifting. These innovations allow ScatterFormer to avoid unnecessary memory allocation and permutation operations, achieving high efficiency comparable to CNN-based models. + +# 3.1 Linear Attention + +We begin by revisiting the self-attention introduced by [46]. Given an input matrix $x \in \mathbb{R}^{N \times d}$ , where $N$ denotes the number of tokens and $d$ denotes the dimensionality. Self-attention first linearly projects the input to queries, keys, and values using weight matrices $W_{Q}, W_{K}$ , and $W_{V}$ , such that $Q = xW_{Q}, K = xW_{K}, V = xW_{V}$ . Then it computes a Softmax-based attention map based on queries and keys i.e., $A(Q, K) = \mathrm{Softmax}(QK^{\top}) / \sqrt{d}$ . The output of self-attention is defined as the weighted sum of $N$ values with the weights corresponding to the attention map, i.e., $O = A(Q, K)V$ . This approach involves calculating the similarity for all query-key pairs, which yields a computational complexity of $O(N^{2})$ . + +Linear attention [1, 17] offers a notable alternative to self-attention by significantly reducing its computational complexity from $O(N^2)$ to $O(N)$ . This efficiency is achieved through the application of kernel functions on the query $(Q)$ and key $(K)$ matrices, which effectively approximates the original attention map without relying on Softmax, i.e., $A(Q, K) = \phi(Q)\phi(K)^{\mathsf{T}}$ . Taking advantage of the associative property of matrix multiplication, the computation shifts from $(\phi(Q)\phi(K)^{\mathsf{T}})V$ to $\phi(Q)(\phi(K)^{\mathsf{T}}V)$ . This modification leads to a computational complexity proportional to the number of tokens, maintaining an order of $O(N)$ . + +# 3.2 Scattered Linear Attention (SLA) + +Instead of organizing the voxels within a window into fixed-length subsets, we arrange the voxels of the entire scene into a single flattened matrix, as illustrated in Figure 3(a). These voxels are ordered based on their window coordinates, ensuring that voxels from the same window form a sub-matrix in contiguous memory. Initially, we use a shared projection layer to map all the voxels in the scene to query, key, and value representations, denoted $Q$ , $K$ , $V$ . Subsequently, we perform attention computations on these submatrices separately, resulting in the scattered linear attention (SLA) formula as follows: + +$$ +\operatorname {S L A} (Q, K, V) = \operatorname {C o n c a t} \left[ \operatorname {L A} \left(Q ^ {j}, K ^ {j}, V ^ {j}\right) \right] _ {j = 1: M}, \tag {2} +$$ + +where $M$ is the number of non-empty window in the current scene and LA is the linear attention formula used in [17]. Based on this formula, the output voxel features $O^j$ in the $j^{\mathrm{th}}$ window can be defined as: + +$$ +\mathrm {L A}: O ^ {j} = \frac {\phi (Q) \sum_ {i = 1} ^ {m ^ {j}} \phi \left(K _ {i}\right) ^ {\top} V _ {i}}{\phi (Q) \sum_ {i = 1} ^ {m ^ {j}} \phi \left(K _ {i}\right) ^ {\top}} \tag {3} +$$ + +where $m^j$ is the number of voxel in the window and $\phi (x)$ is the kernel function. + +Hardware Efficient Implementation. It should be noted that the implementation of Equation 2 is not straightforward, as the sub-matrices have different numbers of rows. One naive implementation is to cache the "kv" matrix of each + +![](images/5816bbdd784f56883d37f1190b2007a02448843582dd37f1427736e0a3847fbe.jpg) +(a) + +![](images/61611b9e5d4ca7f85444d115979bd35c71ff9bb30b2578de2216e84f4a722b90.jpg) +(b) +Fig. 3: (a) The voxel features are sorted by windows and flattened to a single matrix, which is further partitioned into multiple chunks of uniform length and sent to SLA kernel for attention computation. (b) In SLA, we first allocate individual threads to each window, each of them iterates over all corresponding key and value chunks, calculates, and accumulates their products to obtain the hidden state matrix of each window. Then, multiple threads are allocated to these query chunks $q_{i} \in Q$ , calculating the chunk-wise output by multiplying $q_{i}$ with the corresponding hidden state matrix. + +voxel and apply $scatter^3$ operator to accumulate them within each window. However, this would consume a high memory overhead and IO latency. Considering that GPUs feature a memory hierarchy that includes larger, slower global GPU memory (high-bandwidth memory; HBM) and smaller, faster shared memory (SRAM), optimal utilization of SRAM to minimize HBM I/O costs can lead to significant speed-ups. Based on this, we partition the flattened $Q, K, V$ into multiple chunks, load them from slow HBM to fast SRAM. Specifically, for each window, we allocate an single thread. Each thread iterates over all the key and value chunks, calculates, and accumulates their products to obtain the hidden state matrix of the current window. After calculating the hidden state matrix in window, we assign individual thread to each query chunk $q_i \in Q^j$ and calculate the chunk-wise output can then be following Equation 3. Figure 3(b) illustrates this chunk-wise matrix multiplication approach. As can be seen, it not only avoids outputting large matrices but also allows for more flexible processing of variable-length sequences. + +In our implementation, we use Triton [44] to perform chunk-wise matrix multiplication. As illustrated in Figure 4, it outperforms the naive scattered-based operation, demonstrating improved speed and reduced memory usage when computing attention for extremely long sequences. + +# 3.3 Cross Window Interaction + +Traditional window-based transformers [8,40,47] utilize window shifting for interwindow connection. For point cloud sequences, this requires recalculating window coordinates and rearranging the voxels. In DSVT, the computation of voxel + +![](images/e19d811d4b8025e814593bf484992f98c2b31d8126b02a975f810475fc70b9bb.jpg) + +![](images/80784b1ee3048e8ecee1c136574d82adfefdd08e582f1395abe5b57500be5941.jpg) +Fig. 4: Comparison between between scatter-based implementation (yellow) and our chunk-wise implementation (blue). + +indices in subsets and the rearrangement of voxel matrices are particularly time consuming, accounting for $24\%$ of the total backbone latency, as shown in Figure 5. To alleviate this unnecessary computational overhead from window shifting, we introduce the Cross-Window Interaction (CWI) module, which employs convolutions with large kernel to blend voxel features across windows. + +We adhere to the design principles of the Inception architecture [42], where the input features are divided along the channel dimension and processed through multiple branches using group-wise convolutions. Instead of employing a single large kernel for convolution, we adopt more efficient 1D kernels along different axes for feature aggregation. Specifically, we apply a $(S_h + 1) \times 1 \times 1$ convolution in one branch and a $1 \times (S_w + 1) \times 1$ convolution in another branch, where $(S_w, S_h)$ denotes the window size. Additionally, we incorporate a standard $3 \times 3 \times 3$ convolution in one branch to enhance locality, while another branch performs an identity mapping. Given input $X$ of channel $4c$ , the output $X'$ of the Cross Window Interaction module can be written as: + +$$ +X ^ {\prime} = \operatorname {C o n c a t} \left(X _ {k}, X _ {w}, X _ {h}, X ^ {3 c: 4 c}\right), \tag {4} +$$ + +where + +$$ +X _ {k} = \operatorname {D W C o n v} _ {\left(S _ {h} + 1\right) \times 1 \times 1} \left(X ^ {: c}\right), \tag {5} +$$ + +$$ +X _ {w} = \operatorname {D W C o n v} _ {1 \times \left(S _ {w} + 1\right) \times 1} \left(X ^ {c: 2 c}\right), \tag {6} +$$ + +$$ +X _ {h} = \operatorname {D W C o n v} _ {3 \times 3 \times 3} \left(X ^ {2 c: 3 c}\right). \tag {7} +$$ + +![](images/b80d62c5b3034d81653c564b43028cf806b236d6d2361938a5c9c4b42f7a6cee.jpg) +Fig. 5: Runtime decomposition in DSVT [47] and ScatterFormer backbones. The offsets generation means computing the starting address of each chunk in the flattened matrix. + +![](images/43750b18e59eb1fa21fbec4d709e83e507896764ae1536406d913eae1fddbe25.jpg) + +In our experiments, we found that this axis-decomposed large kernel design can achieve a better precision-latency trade-off. With these lengthy kernels, the voxel features can be efficiently blended among different windows. + +# 3.4 Detection Head and Loss + +To complement existing detection heads, ScatterFormer generates dense BEV feature maps from sparse voxel representations by placing them back to their spatial locations and filling the unoccupied positions with zeros. For the BEV network, we simply follow [53] by adopting a cascaded CNN with convolutional blocks at multiple levels of strides. In terms of detection head and loss function, we follow the design identical to that used in DSVT [47]. We train the detection model with heatmap estimation, bounding box regression, and incorporate an IoU loss for confidence calibration. On the nuScenes dataset, the detection head follows the architecture used in Transfusion [2]. + +# 4 Experiments + +# 4.1 Datasets and Evaluation Metrics + +Waymo Open Dataset (WOD). This dataset contains 230,000 annotated samples split into 160,000 for training, 40,000 for validation, and 30,000 for testing. It uses two metrics for 3D object detection: mean average precision (mAP) and mAP weighted by heading accuracy (mAPH), further categorized into Level 1 (L1) for objects detected by more than five LiDAR points and Level 2 (L2) for those detected with at least one point. + +NuScenes. This dataset comprises 40,000 annotated samples, with 28,000 for training, 6,000 for validation, and 6,000 for testing. On this dataset, the model performance is measured by mean average precision (mAP) across multiple distance thresholds (0.5, 1, 2, and 4 meters) and the nuScenes detection score (NDS), which combines mAP with a weighted sum of five additional metrics assessing true positive predictions in translation, scale, orientation, velocity, and attribute accuracy. + +![](images/1fd59f3f4ba6c1f1f382f901e3edfcb8ab732f8e1bb1ff93745ab0685f48f00b.jpg) +Fig. 6: The detection performance (mAPH/L2) vs. speed (FPS) of different methods on Waymo validation set. The speeds are measured on an NVIDIA A100 GPU. + +# 4.2 Implementation Details + +Our approach is implemented using the open-source framework OpenPCDet [43]. To construct ScatterFormer, we set the voxel size to $(0.32\mathrm{m}, 0.32\mathrm{m}, 0.1875\mathrm{m})$ for the Waymo dataset and $(0.3\mathrm{m}, 0.3\mathrm{m}, 8\mathrm{m})$ for the NuScenes dataset. The window sizes $(S_w, S_h)$ for the two datasets are set to $(12, 12)$ and $(20, 20)$ , respectively. We stack six building blocks for the backbone network. We configure our attention module to have 4 heads with a dimensionality of 128. ScatterFormer is trained for 24 epochs with a learning rate of 0.006 on Waymo Dataset and 20 epochs with a learning rate of 0.004 on NuScenes Dataset. In the last 4 epochs, we disabled the data augmentation strategy of ground-truth sampling. The model is trained on 8 RTX A6000 GPUs with a batch size of 32. Other settings for training and inference adhere strictly to DSVT [47]. + +# 4.3 Comparison with State-of-the-Arts + +Results on Waymo Open Dataset (WOD). We conduct a detailed comparison of ScatterFormer against the published results on the validation set of the Waymo Open Dataset. In line with standard practices, we distinctly categorize and list the methods utilizing single-frame and multi-frame approaches. To ensure comprehensive coverage, we also include comparisons with methods incorporating long-term temporal modeling, such as those described in [4,15,61]. As demonstrated in Table 1, our single-stage model outperforms most two-stage methods and achieves better performance than state-of-the-art methods such as DSVT [47] and HEDNet [53]. Moreover, our model achieves 76.0 and 76.7 level 2 mAPH on 3 and 4 frame settings, respectively, outperforming previous multiframe methods by a margin of 1.1. Furthermore, as shown in Figure 6, + +Table 1: Performance comparison on the validation set of Waymo Open Dataset. Symbol \* denotes the methods with temporal modeling, and - means that the result is not available. + +
MethodFramesALL (3D mAPH)Vehicle (AP/APH)Pedestrian (AP/APH)Cyclist (AP/APH)
L1L2L1L2L1L2L1L2
SECOND [49]163.0557.2372.27/71.6963.85/63.3368.70/58.1860.72/51.3160.62/59.2858.34/57.05
PointPillar [18]163.3357.5371.60/71.0063.10/62.5070.60/56.7062.90/50.2064.40/62.3061.90/59.90
IA-SSD [56]164.4858.0870.53/69.6761.55/60.8069.38/58.4760.30/50.7367.67/65.3064.98/62.71
LiDAR R-CNN [20]166.2060.1073.50/73.0064.70/64.2071.20/58.7063.10/51.7068.60/66.9066.10/64.40
RSN [41]1--75.10/74.6066.00/65.5077.80/72.7068.30/63.70--
Part-A2 [38]170.2563.8477.05/76.5168.47/67.9775.24/66.8766.18/58.6268.60/67.3666.13/64.93
Centerpoint [52]1-65.50-- /66.20--/62.60--/67.60
VoTR [24]1--74.95/74.2565.91/65.29----
VoxSeT [14]172.2466.2274.50/74.0365.99/65.5680.03/72.4272.45/65.3971.56/70.2968.95/67.73
SST-1f [8]1--76.22/75.7968.04/67.6481.39/74.0572.82/65.93--
SWFormer-1f [40]1--77.8/77.369.2/68.880.9/72.772.5/64.9--
PV-RCNN [35]169.6363.3377.51/76.8968.98/68.4175.01/65.6566.04/57.6167.81/66.3565.39/63.98
PillarNet [34]174.6068.4379.09/78.5970.92/70.4680.59/74.0172.28/66.1772.29/71.2169.72/68.67
PV-RCNN++ [35]175.2168.6179.10/78.6370.34/69.9180.62/74.6271.86/66.3073.49/72.3870.70/69.62
FlatFormer [23]1-67.2-69.0/68.6-71.5/65.3-68.6/67.5
PillarNext-1f [19]175.7469.7478.40/77.9070.27/69.8182.53/77.1474.90/69.8073.21/72.2070.58/69.62
VoxelNext [5]176.370.178.2/77.769.9/69.481.5/76.373.5/68.676.1/74.973.3/72.2
FSD [9]177.370.879.2/78.870.5/70.182.6/77.373.9/69.177.1/76.074.4/73.3
DSVT-1f [47]178.272.179.7/79.371.4/71.083.7/78.976.1/71.577.5/76.574.6/73.7
HEDNet-1f [53]179.573.481.1/80.673.2/72.784.4/80.076.8/72.678.7/77.775.8/74.9
ScatterFormer (ours)179.773.881.0/80.573.1/72.784.5/79.977.0/72.679.9/78.977.1/76.1
SST-3f [8]3--78.66/78.2169.98/69.5783.81/80.1475.94/72.37--
FlatFormer-3f [23]3-72.0-71.4/71.0-74.5/71.3-74.7/73.7
SWFormer-3f [40]3--79.4/78.971.1/70.682.9/79.074.8/71.1--
PillarNext-3f [19]380.074.580.6/80.172.9/72.485.0/82.178.0/75.278.9/77.976.7/75.7
DSVT-4f [47]481.375.681.8/81.474.1/73.685.6/82.878.6/75.980.4/79.678.1/77.3
ScatterFormer-3f (ours)381.776.082.0/81.475.0/74.185.7/83.279.0/76.180.6/79.578.1/77.7
ScatterFormer-4f (ours)481.976.782.4/81.975.2/74.786.1/83.479.3/76.681.2/80.478.9/78.1
*3D-MAN [51]16--74.53/74.0367.61/67.1471.7/67.762.6/59.0--
*CenterFormer [61]477.073.278.1/77.673.4/72.981.7/78.677.2/74.275.6/74.873.4/72.6
*CenterFormer [61]877.373.778.8/78.374.3/73.882.1/79.377.8/75.075.2/74.473.2/72.3
*MPPNet [4]479.8374.2281.54/81.0674.07/73.6184.56/81.9477.20/74.6777.15/76.5075.01/74.38
*MPPNet [4]1680.4074.8582.74/82.2875.41/74.9684.69/82.2577.43/75.0677.28/76.6675.13/74.52
*MSF [15]480.2074.6281.36/80.8773.81/73.3585.05/82.1077.92/75.1178.40/77.6176.17/75.40
*MSF [15]880.6575.4682.83/82.0175.76/75.3185.24/82.2178.32/75.6178.52/77.7476.32/75.47
+ +our method significantly enhances the detection runtime. This improvement is attributed to our use of Linear Attention, which eliminates the need for extensive voxel sorting and padding operations. Furthermore, as presented in Table 2, ScatterFormer achieves the highest Level 1 and Level 2 mAPH scores in all categories of the Waymo test set, highlighting its superior detection quality and precision. Remarkably, our approach, as a one-shot method, delivers performance that is on par with methods that require temporal modeling. This finding underscores the critical importance of network design and architecture in achieving high-performance architecture. + +Results on NuScenes. We compare ScatterFormer with the previous best performing methods on the nuScenes dataset. As shown in Table 3, ScatterFormer achieves the state-of-the-art performance in terms of val NDS (72.4) and mAP (68.3), surpassing HEDNet [53] by 1.0 and 1.6 respectively. Compared to TransFusion [2], ScatterFormer achieves a significant improvement of $2.8\%$ in mAP, demonstrating that its backbone network is superior to the sparse convolutional network. ScatterFormer exhibits particularly strong performance on categories like Bus, Bike, Construction Vehicle and Trailer, suggesting its ability to effectively capture and encode detailed contextual features. + +Table 2: Performance comparison on the test set of Waymo Open Dataset. $\cdot$ means that the result is not available. + +
MethodALL (3D mAPH)Vehicle (AP/APH)Pedestrian (AP/APH)Cyclist (AP/APH)
L1L2L1L2L1L2L1L2
PointPillar [18]--68.1060.1068.00/55.5061.40/50.10--
StarNet [26]--61.0054.5067.80/59.9061.10/54.00--
M3DETR [11]67.161.977.7/77.170.5/70.068.2/58.560.6/52.067.3/65.765.3/63.8
3D-MAN [51]--78.2869.9869.97/65.9863.98/60.26--
PV-RCNN++ [36]75.770.281.6/81.273.9/73.580.4/75.074.1/69.071.9/70.869.3/68.2
CenterPoint [52]77.271.981.1/80.673.4/73.080.5/77.374.6/71.574.6/73.772.2/71.3
RSN [41]--80.3071.6078.90/75.6070.70/67.80--
SST-3f [8]78.372.881.0/80.673.1/72.783.3/79.776.9/73.575.7/74.673.2/72.2
HEDNet [53]79.0573.7783.78/83.3976.33/75.9683.46/78.9877.53/73.2575.86/74.7873.13/72.09
PillarNext-3f [19]79.0074.0983.28/82.8376.18/75.7684.40/81.4478.84/75.9873.77/72.7371.56/70.55
ScatterFormer-4f (ours)80.7175.8485.60/85.1378.34/78.0784.58/81.6479.12/76.3276.20/75.3574.02/73.13
+ +Table 3: The performance on the validation set of NuScenes. + +
MethodNDSmAPCarTruckBusT.L.C.V.Ped.M.T.BikeT.C.B.R.
CenterPoint [52]66.559.284.957.470.738.116.985.159.042.069.868.3
VoxelNeXt [5]66.760.583.955.570.538.121.184.662.850.069.469.4
TransFusion-L [2]70.165.586.960.873.143.425.287.572.957.377.270.3
PillarNext [19]68.462.285.057.467.635.620.686.868.653.177.369.7
DSVT [47]71.166.487.462.675.942.125.388.274.858.777.870.9
HEDNet [53]71.466.787.760.677.850.728.987.174.356.876.366.9
ScatterFormer (ours)72.468.388.665.479.345.429.188.774.761.578.272.3
+ +# 4.4 Ablation Study + +In this section, we conduct ablation experiments in ScatterFormer using $20\%$ of the Waymo training and validation data. Each model was trained for 24 epochs and run on default parameters. Table 4 shows four different configurations of ScatterFormer: (a) removing the Scattered Linear Attention module (SLA), (b) removing the Cross Window Interaction (CWI) module, (c) replacing the CWI with Shifted Window (SW) approach, and (d) removing Conditional Position Encoding (CPE) module. + +As can be seen in Table 4, configuration (a) results in ScatterFormer experiencing a decrease of $(3.1\%, 3.9\%, \text{and} 3.2\%)$ APH on three categories. This highlights the effectiveness of the SLA, as it successfully captures long-range contexts and facilitates dynamic feature extraction for object detection. In configuration (b), ScatterFormer, which lacks cross-window interaction capability, exhibits a performance drop of $(1.9\%, 1.0\%, \text{and} 2.0\%)$ APH across the three categories. While the Shifted Window approach in configuration (c) enhances performance, the improvement is not as substantial as that achieved by the proposed CWI module. The results in configuration (d) indicate that CPE contributes to a certain degree of performance enhancement. + +Window Size. In Table 5, We tested the impact of different window sizes on the detection performance of various target types. It can be observed that ScatterFormer is not highly sensitive to window size. Slightly increasing the window size can enhance the detection performance for small targets, indicating that for objects like pedestrians and bicycles, where the point cloud is overly sparse, appropriate contextual information can enhance the recognition of these + +Table 4: Ablation experiments on the validation set of Waymo Open Dataset. "SLA" and "CWT" refer to the proposed Scattered Linear Attention and Cross-Window Interaction modules, respectively. "CPE" refers to Conditional Position Embedding. AP and APH scores on LEVEL2 are reported. + +
AblationVeh.Ped.Cyc.
-baseline71.6/71.176.0/70.974.2/73.1
(a)w/o SLA module68.5/68.073.3/67.071.0/69.9
(b)w/o CWI module69.6/69.274.0/69.972.1/71.1
(c)CWI → SW69.9/69.575.0/70.473.8/72.4
(d)w/o CPE70.2/69.873.3/67.272.4/70.5
+ +Table 5: Comparison of model performance using different window sizes at various stages. The APH (L2) scores on different categories are reported. + +
Window SizeRegion SizeVeh. Ped. Cyc.
103.20 m70.2 69.2 71.0
123.84 m71.1 70.9 73.1
144.48 m71.3 69.9 72.5
165.12 m71.0 69.2 72.3
+ +objects. However, excessively large windows can degrade the performance. We speculate that this is because larger windows encompass more tokens, thereby mitigating the focusing power of linear attention. + +Kernel Design in CWI. Table 6 presents the performance under various kernel configurations in the Cross-Window Interaction (CWI) module. Increasing the number of parameters in the kernel significantly enhances performance. For example, using a large $7 \times 7$ kernel brings notable performance improvements, but also introduces more latency. Additionally, larger kernel strides enhance the exchange rate of tokens across windows. For example, using dilated convolutions in CWI performs slightly better than using non-dilated convolutions. Instead of scaling up large, we decompose large kernel into more efficient 1D kernels, which achieves the optimal latency while keeping comparable performance. As the current Sconv library lacks support for depth-wise convolution, we have developed a modified version to address this limitation. $^4$ + +# 4.5 Comparison with Different Linear Attentions + +We further explored the performance of ScatterFormer by conducting experiments with several other linear attention mechanisms, including Gated Linear Attention [32], Cross-covariance Attention (XCA) [1], and Focused Linear Attention [13]. The results are summarized in Table 7. Our findings suggest that the window-based attention framework employed by ScatterFormer is not highly sensitive to the specific form of linear attention used. Notably, Efficient Attention [3] performs slightly worse than XCA [1] in the Vehicle class but better in the + +Table 6: Performance with various kernel configurations in the Cross-Window Integration (CWI) module. Each configuration is specified by the notation "KxDy", where "Kx" refers to the kernel size, and "Dy" denotes the dilation rate. The APH (L2) scores on different categories are reported. + +
KernelLatencyVeh.Ped.Cyc.
Conv1DK13 × 247ms71.170.973.1
Conv2DK5D149ms69.569.071.1
Conv2DK5D249ms70.770.472.8
Conv2DK7D165ms71.571.072.5
+ +Table 7: Performance with different linear attention designs. The APH (L2) scores on different categories are reported. + +
Linear AttentionVeh.Ped.Cyc.
Efficient Attn [3]71.170.973.1
Gated Linear Attn [32]69.969.072.5
Focused Linear Attn [13]70.469.572.3
XCA [1]71.570.472.8
+ +Pedestrian and Cyclist classes. Focused Linear Attention introduces additional computational overhead, while Gated Linear Attention adds more learnable parameters to the model. + +# 4.6 Limitations + +ScatterFormer relies on our customized operators, which have not yet been implemented as plugins in TensorRT. Therefore, deploying ScatterFormer on in-vehicle devices will require additional engineering efforts. Despite this, ScatterFormer is more efficient than current models based on sparse convolution and traditional attention when used on consumer-grade GPU cards. Furthermore, ScatterFormer can be optimized by dynamically partitioning matrices according to different GPU architectures to leverage TensorCore for hardware acceleration. + +# 5 Conclusion + +We introduced ScatterFormer, an innovative architecture designed for 3D object detection using point clouds, specifically targeting the challenges associated with processing sparse and unevenly distributed data from LiDAR sensors. The cornerstone of our approach is the Scattered Linear Attention (SLA) module, which effectively addresses the limitations of conventional attention mechanisms in managing voxel features of varying lengths. SLA ingeniously combines linear attention with a chunk-wise matrix multiplication algorithm, tailored to meet the distinct requirements of processing voxels grouped by windows. By integrating SLA with a novel cross-window interaction (CWI) module, ScatterFormer achieves higher accuracy and lower latency, surpassing traditional transformer-based and sparse CNN-based detectors in extensive 3D detection tasks. + +# References + +1. Ali, A., Touvron, H., Caron, M., Bojanowski, P., Douze, M., Joulin, A., Laptev, I., Neverova, N., Synnaeve, G., Verbeek, J., et al.: Xcit: Cross-covariance image transformers. Advances in neural information processing systems 34, 20014-20027 (2021) +2. Bai, X., Hu, Z., Zhu, X., Huang, Q., Chen, Y., Fu, H., Tai, C.L.: Transfusion: Robust lidar-camera fusion for 3d object detection with transformers. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1090-1099 (2022) +3. Cai, H., Gan, C., Han, S.: Efficientvit: Enhanced linear attention for high-resolution low-computation visual recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2023) +4. Chen, X., Shi, S., Zhu, B., Cheung, K.C., Xu, H., Li, H.: Mppnet: Multi-frame feature intertwining with proxy points for 3d temporal object detection. In: European Conference on Computer Vision. pp. 680-697. Springer (2022) +5. Chen, Y., Liu, J., Zhang, X., Qi, X., Jia, J.: Voxelnext: Fully sparse voxelnet for 3d object detection and tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21674-21683 (2023) +6. Chu, X., Tian, Z., Zhang, B., Wang, X., Shen, C.: Conditional positional encodings for vision transformers. ICLR (2023) +7. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. ICLR (2021) +8. Fan, L., Pang, Z., Zhang, T., Wang, Y.X., Zhao, H., Wang, F., Wang, N., Zhang, Z.: Embracing single stride 3d object detector with sparse transformer. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8458-8468 (2022) +9. Fan, L., Wang, F., Wang, N., Zhang, Z.: Fully sparse 3d object detection. Advances in Neural Information Processing Systems 35, 351-363 (2022) +0. Fan, L., Xiong, X., Wang, F., Wang, N., Zhang, Z.: Rangedet: In defense of range view for lidar-based 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2918-2927 (October 2021) +1. Guan, T., Wang, J., Lan, S., Chandra, R., Wu, Z., Davis, L., Manocha, D.: M3detr: Multi-representation, multi-scale, mutual-relation 3d object detection with transformers. In: Proceedings of the IEEE/CVF winter conference on applications of computer vision. pp. 772-782 (2022) +2. Guo, M.H., Cai, J.X., Liu, Z.N., Mu, T.J., Martin, R.R., Hu, S.M.: Pct: Point cloud transformer. Computational Visual Media 7(2), 187-199 (Apr 2021). https://doi.org/10.1007/s41095-021-0229-5, http://dx.doi.org/10.1007/s41095-021-0229-5 +3. Han, D., Pan, X., Han, Y., Song, S., Huang, G.: Flatten transformer: Vision transformer using focused linear attention. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5961-5971 (2023) +4. He, C., Li, R., Li, S., Zhang, L.: Voxel set transformer: A set-to-set approach to 3d object detection from point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8417-8427 (2022) +5. He, C., Li, R., Zhang, Y., Li, S., Zhang, L.: Msf: Motion-guided sequential fusion for efficient 3d object detection from point cloud sequences. In: Proceedings of the + +IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5196-5205 (2023) +16. He, C., Zeng, H., Huang, J., Hua, X.S., Zhang, L.: Structure aware single-stage 3d object detection from point cloud. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11873-11882 (2020) +17. Katharopoulos, A., Vyas, A., Pappas, N., Fleuret, F.: Transformers are rnns: Fast autoregressive transformers with linear attention. In: International conference on machine learning. pp. 5156-5165. PMLR (2020) +18. Lang, A.H., Vora, S., Caesar, H., Zhou, L., Yang, J., Beijbom, O.: Pointpillars: Fast encoders for object detection from point clouds. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 12697-12705 (2019) +19. Li, J., Luo, C., Yang, X.: Rethinking network designs for 3d object detection in lidar point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17567-17576 (2023) +20. Li, Z., Wang, F., Wang, N.: Lidar r-cnn: An efficient and universal 3d object detector. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021) +21. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 10012-10022 (October 2021) +22. Liu, Z., Zhang, Z., Cao, Y., Hu, H., Tong, X.: Group-free 3d object detection via transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2949-2958 (October 2021) +23. Liu, Z., Yang, X., Tang, H., Yang, S., Han, S.: Flatformer: Flattened window attention for efficient point cloud transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1200-1211 (2023) +24. Mao, J., Xue, Y., Niu, M., Bai, H., Feng, J., Liang, X., Xu, H., Xu, C.: Voxel transformer for 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 3164-3173 (October 2021) +25. Misra, I., Girdhar, R., Joulin, A.: An end-to-end transformer model for 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2906-2917 (October 2021) +26. Ngiam, J., Caine, B., Han, W., Yang, B., Chai, Y., Sun, P., Zhou, Y., Yi, X., Alsharif, O., Nguyen, P., et al.: Starnet: Targeted computation for object detection in point clouds. arXiv preprint arXiv:1908.11069 (2019) +27. Pan, X., Xia, Z., Song, S., Li, L.E., Huang, G.: 3d object detection with pointformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7463-7472 (June 2021) +28. Qi, C.R., Litany, O., He, K., Guibas, L.J.: Deep though voting for 3d object detection in point clouds. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (October 2019) +29. Qi, C.R., Liu, W., Wu, C., Su, H., Guibas, L.J.: Frustum pointnets for 3d object detection from rgb-d data. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 918-927 (2018) +30. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 652-660 (2017) +31. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In: Advances in neural information processing systems. pp. 5099-5108 (2017) + +32. Qin, Z., Li, D., Sun, W., Sun, W., Shen, X., Han, X., Wei, Y., Lv, B., Yuan, F., Luo, X., Qiao, Y., Zhong, Y.: Scaling transnormer to 175 billion parameters (2023) +33. Sheng, H., Cai, S., Liu, Y., Deng, B., Huang, J., Hua, X.S., Zhao, M.J.: Improving 3d object detection with channel-wise transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2743-2752 (October 2021) +34. Shi, G., Li, R., Ma, C.: Pillarnet: Real-time and high-performance pillar-based 3d object detection. In: European Conference on Computer Vision. pp. 35-52. Springer (2022) +35. Shi, S., Guo, C., Jiang, L., Wang, Z., Shi, J., Wang, X., Li, H.: Pv-rcnn: Pointvoxel feature set abstraction for 3d object detection. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020) +36. Shi, S., Jiang, L., Deng, J., Wang, Z., Guo, C., Shi, J., Wang, X., Li, H.: Pvrnn++: Point-voxel feature set abstraction with local vector representation for 3d object detection. International Journal of Computer Vision 131(2), 531-551 (2023) +37. Shi, S., Wang, X., Li, H.: Pointrecnn: 3d object proposal generation and detection from point cloud. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 770-779 (2019) +38. Shi, S., Wang, Z., Wang, X., Li, H.: Part-a^2 net: 3d part-aware and aggregation neural network for object detection from point cloud. arXiv preprint arXiv:1907.03670 (2019) +39. Shi, W., Rajkumar, R.: Point-gnn: Graph neural network for 3d object detection in a point cloud. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1711-1719 (2020) +40. Sun, P., Tan, M., Wang, W., Liu, C., Xia, F., Leng, Z., Anguelov, D.: Swformer: Sparse window transformer for 3d object detection in point clouds. In: European Conference on Computer Vision. pp. 426-442. Springer (2022) +41. Sun, P., Wang, W., Chai, Y., Elsayed, G., Bewley, A., Zhang, X., Sminchisescu, C., Anguelov, D.: Rsn: Range sparse net for efficient, accurate lidar 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5725-5734 (2021) +42. Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., Wojna, Z.: Rethinking the inception architecture for computer vision. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2818-2826 (2016) +43. Team, O.D.: Openpcdet: An open-source toolbox for 3d object detection from point clouds. https://github.com/open-mmlab/OpenPCDet (2020) +44. Tilet, P., Kung, H.T., Cox, D.: Triton: An intermediate language and compiler for tiled neural network computations. In: Proceedings of the 3rd ACM SIGPLAN International Workshop on Machine Learning and Programming Languages. p. 10-19. MAPL 2019, Association for Computing Machinery, New York, NY, USA (2019). https://doi.org/10.1145/3315508.3329973, https://doi.org/10.1145/3315508.3329973 +45. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jegou, H.: Training data-efficient image transformers & distillation through attention. In: International Conference on Machine Learning. vol. 139, pp. 10347-10357 (July 2021) +46. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: Advances in neural information processing systems. pp. 5998-6008 (2017) + +47. Wang, H., Shi, C., Shi, S., Lei, M., Wang, S., He, D., Schiele, B., Wang, L.: Dsvt: Dynamic sparse voxel transformer with rotated sets. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13520-13529 (2023) +48. Wang, S., Li, B.Z., Khabsa, M., Fang, H., Ma, H.: Linformer: Self-attention with linear complexity (2020) +49. Yan, Y., Mao, Y., Li, B.: Second: Sparsely embedded convolutional detection. Sensors 18(10), 3337 (2018) +50. Yang, B., Luo, W., Urtasun, R.: Pixor: Real-time 3d object detection from point clouds. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 7652-7660 (2018) +51. Yang, Z., Zhou, Y., Chen, Z., Ngiam, J.: 3d-man: 3d multi-frame attention network for object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1863-1872 (2021) +52. Yin, T., Zhou, X., Krahenbuhl, P.: Center-based 3d object detection and tracking. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11784-11793 (2021) +53. Zhang, G., Chen, J., Gao, G., Li, J., Hu, X.: Hednet: A hierarchical encoder-decoder network for 3d object detection in point clouds. arXiv preprint arXiv:2310.20234 (2023) +54. Zhang, G., Fan, L., He, C., Lei, Z., Zhang, Z., Zhang, L.: Voxel mamba: Group-free state space models for point cloud based 3d object detection. arXiv preprint arXiv:2406.10700 (2024) +55. Zhang, Y., Huang, D., Wang, Y.: Pc-rgnn: Point cloud completion and graph neural network for 3d object detection. In: Proceedings of the AAAI conference on artificial intelligence. vol. 35, pp. 3430-3437 (2021) +56. Zhang, Y., Hu, Q., Xu, G., Ma, Y., Wan, J., Guo, Y.: Not all points are equal: Learning highly efficient point-based detectors for 3d lidar point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18953-18962 (2022) +57. Zhao, H., Jiang, L., Jia, J., Torr, P.H., Koltun, V.: Point transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 16259-16268 (October 2021) +58. Zheng, W., Tang, W., Jiang, L., Fu, C.W.: Se-ssd: Self-ensembling single-stage object detector from point cloud. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14494-14503 (2021) +59. Zhou, C., Zhang, Y., Chen, J., Huang, D.: Octr: Octtree-based transformer for 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5166-5175 (2023) +60. Zhou, Y., Tuzel, O.: Voxelnet: End-to-end learning for point cloud based 3d object detection. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4490-4499 (2018) +61. Zhou, Z., Zhao, X., Wang, Y., Wang, P., Foroosh, H.: Centerformer: Center-based transformer for 3d object detection. In: European Conference on Computer Vision. pp. 496-513. Springer (2022) \ No newline at end of file diff --git a/2024/ScatterFormer_ Efficient Voxel Transformer with Scattered Linear Attention/images.zip b/2024/ScatterFormer_ Efficient Voxel Transformer with Scattered Linear Attention/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e82821581796bc164d5d06bddf33249c534daa47 --- /dev/null +++ b/2024/ScatterFormer_ Efficient Voxel Transformer with Scattered Linear Attention/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4a5e4e969f9c09b67ebdc6b6955951714f32a94b0407148d110dbd23d98ed962 +size 686309 diff --git a/2024/ScatterFormer_ Efficient Voxel Transformer with Scattered Linear Attention/layout.json b/2024/ScatterFormer_ Efficient Voxel Transformer with Scattered Linear Attention/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6a4a3dab7e1cedb65e2c19ac66997f6934bb133e --- /dev/null +++ b/2024/ScatterFormer_ Efficient Voxel Transformer with Scattered Linear Attention/layout.json @@ -0,0 +1,8595 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 138, + 111, + 477, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 111, + 477, + 148 + ], + "spans": [ + { + "bbox": [ + 138, + 111, + 477, + 148 + ], + "type": "text", + "content": "ScatterFormer: Efficient Voxel Transformer with Scattered Linear Attention" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 160, + 168, + 455, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 168, + 455, + 182 + ], + "spans": [ + { + "bbox": [ + 160, + 168, + 455, + 182 + ], + "type": "text", + "content": "Chenhang He" + }, + { + "bbox": [ + 160, + 168, + 455, + 182 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 160, + 168, + 455, + 182 + ], + "type": "text", + "content": ", Ruihuang Li" + }, + { + "bbox": [ + 160, + 168, + 455, + 182 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 160, + 168, + 455, + 182 + ], + "type": "text", + "content": ", Guowen Zhang" + }, + { + "bbox": [ + 160, + 168, + 455, + 182 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 160, + 168, + 455, + 182 + ], + "type": "text", + "content": ", and Lei Zhang" + }, + { + "bbox": [ + 160, + 168, + 455, + 182 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 171, + 190, + 443, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 190, + 443, + 201 + ], + "spans": [ + { + "bbox": [ + 171, + 190, + 443, + 201 + ], + "type": "text", + "content": "1 The Hong Kong Polytechnic University, Hong Kong SAR, China" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 229, + 201, + 386, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 229, + 201, + 386, + 257 + ], + "spans": [ + { + "bbox": [ + 229, + 201, + 386, + 257 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 229, + 201, + 386, + 257 + ], + "type": "text", + "content": " OPPO Research, Shenzhen, China chenhang.he@polyu.edu.hk csrhli@comp.polyu.edu.hk guowen.zhang@connect.polyu.edu.hk cslzhang@comp.polyu.edu.hk" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 291, + 455, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 291, + 455, + 533 + ], + "spans": [ + { + "bbox": [ + 160, + 291, + 455, + 533 + ], + "type": "text", + "content": "Abstract. Window-based transformers excel in large-scale point cloud understanding by capturing context-aware representations with affordable attention computation in a more localized manner. However, the sparse nature of point clouds leads to a significant variance in the number of voxels per window. Existing methods group the voxels in each window into fixed-length sequences through extensive sorting and padding operations, resulting in a non-negligible computational and memory overhead. In this paper, we introduce ScatterFormer, which to the best of our knowledge, is the first to directly apply attention to voxels across different windows as a single sequence. The key of ScatterFormer is a Scattered Linear Attention (SLA) module, which leverages the pre-computation of key-value pairs in linear attention to enable parallel computation on the variable-length voxel sequences divided by windows. Leveraging the hierarchical structure of GPUs and shared memory, we propose a chunkwise algorithm that reduces the SLA module's latency to less than 1 millisecond on moderate GPUs. Furthermore, we develop a cross-window interaction module that improves the locality and connectivity of voxel features across different windows, eliminating the need for extensive window shifting. Our proposed ScatterFormer demonstrates 73.8 mAP (L2) on the Waymo Open Dataset and 72.4 NDS on the NuScenes dataset, running at an outstanding detection rate of 23 FPS. The code is available at https://github.com/skyhehe123/ScatterFormer." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 543, + 386, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 543, + 386, + 555 + ], + "spans": [ + { + "bbox": [ + 160, + 543, + 386, + 555 + ], + "type": "text", + "content": "Keywords: 3D Object Detection " + }, + { + "bbox": [ + 160, + 543, + 386, + 555 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 543, + 386, + 555 + ], + "type": "text", + "content": " Voxel Transformer" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 578, + 230, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 230, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 230, + 590 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "In the field of 3D object detection, the use of point clouds has become increasingly popular, especially for providing accurate and reliable perception results in autonomous systems. Unlike image data, point clouds obtained from LiDAR are often sparse and nonuniformly distributed, with varying density depending on their distances from the sensor. Earlier approaches utilize point-cloud operators" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 193, + 114, + 421, + 204 + ], + "blocks": [ + { + "bbox": [ + 193, + 114, + 421, + 204 + ], + "lines": [ + { + "bbox": [ + 193, + 114, + 421, + 204 + ], + "spans": [ + { + "bbox": [ + 193, + 114, + 421, + 204 + ], + "type": "image", + "image_path": "49d5b409f15b627611ec4e71b2a9f0cabab4fb09cb602404b2db8501d07c8322.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 181, + 213, + 430, + 260 + ], + "blocks": [ + { + "bbox": [ + 181, + 213, + 430, + 260 + ], + "lines": [ + { + "bbox": [ + 181, + 213, + 430, + 260 + ], + "spans": [ + { + "bbox": [ + 181, + 213, + 430, + 260 + ], + "type": "image", + "image_path": "5adf862f1d88200a3e1cd0f498880ed5403aaf17d034e31c04402a7a5c650d03.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 262, + 480, + 297 + ], + "lines": [ + { + "bbox": [ + 130, + 262, + 480, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 480, + 297 + ], + "type": "text", + "content": "Fig. 1: Illustration of (a) group-based attention and (b) our scattered linear attention on variable-length sequences. The square in different colors represent voxels in different windows and the square with a cross represents the padding voxel." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 304, + 480, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 480, + 353 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 480, + 353 + ], + "type": "text", + "content": "like PointNet++ [31] for feature extraction in continuous space [37]. Some approaches [5,9,16,18,49,50,52,60] have transformed point clouds into voxel grids, which are then efficiently processed by a sparse convolutional neural network (SpCNN) [49]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 354, + 481, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 481, + 533 + ], + "type": "text", + "content": "Recently, the success of vision transformers [7, 21] has motivated a number of attention-based methods to process indoor point clouds [12, 22, 24, 25, 27, 57]. Inspired by SwinTransformer [21], various studies [8, 14, 23, 47] have advanced the application of window-based Transformers in large-scale 3D detection tasks within outdoor environments, achieving outstanding performance and highlighting their potential as alternatives to SpCNN. However, due to the inherent sparsity of point clouds, the number of features grouped by windows can vary significantly, hampering the parallelism in attention computation. To resolve this issue, SST [8] and SWFormer [40] group the voxels within a window into different batches and manages the attention computation in a hybrid serial-parallel manner. More effectively, DSVT [47] alternatively sort the voxels within a window from different axes and partition them into sequences of fixed length, which allow parallel computation on sparse voxels. Albeit effective, these group-based methods, as shown in Figure 1(a), require extensive sorting and padding operations, incurring substantial memory and computational overhead." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 534, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 481, + 666 + ], + "type": "text", + "content": "In this paper, we delve into the window-based voxel transformer where the voxels grouped by windows form variable-length sequences " + }, + { + "bbox": [ + 130, + 534, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\{X_1\\in \\mathbb{R}^{n_1\\times d},X_2\\in" + }, + { + "bbox": [ + 130, + 534, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{n_2\\times d},\\dots,X_k\\in \\mathbb{R}^{n_k\\times d}\\}" + }, + { + "bbox": [ + 130, + 534, + 481, + 666 + ], + "type": "text", + "content": ". Their corresponding self-attention matrices " + }, + { + "bbox": [ + 130, + 534, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\{A_{1}\\in" + }, + { + "bbox": [ + 130, + 534, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{n_1\\times n_1}" + }, + { + "bbox": [ + 130, + 534, + 481, + 666 + ], + "type": "inline_equation", + "content": "A_{2}\\in \\mathbb{R}^{n_{2}\\times n_{2}}" + }, + { + "bbox": [ + 130, + 534, + 481, + 666 + ], + "type": "text", + "content": " ..., " + }, + { + "bbox": [ + 130, + 534, + 481, + 666 + ], + "type": "inline_equation", + "content": "A_{k}\\in \\mathbb{R}^{n_{k}\\times n_{k}}\\}" + }, + { + "bbox": [ + 130, + 534, + 481, + 666 + ], + "type": "text", + "content": " occupy irregular memory spaces, making it a challenge to perform attention in parallel for the voxels of the entire scene. Recent efforts on linear attention [1,13,17,48] have emerged a promising alternative to traditional attention. By approximating softmax operation with the kernel function, i.e., " + }, + { + "bbox": [ + 130, + 534, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\phi (Q)\\cdot \\phi (K)^{\\top}\\approx \\mathrm{softmax}(QK^{\\top})" + }, + { + "bbox": [ + 130, + 534, + 481, + 666 + ], + "type": "text", + "content": ", we can change the computation order from " + }, + { + "bbox": [ + 130, + 534, + 481, + 666 + ], + "type": "inline_equation", + "content": "(Q\\cdot K^{\\top})\\cdot V" + }, + { + "bbox": [ + 130, + 534, + 481, + 666 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 534, + 481, + 666 + ], + "type": "inline_equation", + "content": "Q\\cdot (K^{\\top}\\cdot V)" + }, + { + "bbox": [ + 130, + 534, + 481, + 666 + ], + "type": "text", + "content": ". This not only results in linear complexity but also yields a compressed hidden state " + }, + { + "bbox": [ + 130, + 534, + 481, + 666 + ], + "type": "inline_equation", + "content": "S\\in \\mathbb{R}^{d\\times d}" + }, + { + "bbox": [ + 130, + 534, + 481, + 666 + ], + "type": "text", + "content": " that is invariant to the sequence length. Another attractive property of linear attention is that it" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "text", + "content": "C. He et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 313, + 127 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 313, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 313, + 127 + ], + "type": "text", + "content": "can be converted into a \"recurrent\" form:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 243, + 140, + 481, + 155 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 243, + 140, + 481, + 155 + ], + "spans": [ + { + "bbox": [ + 243, + 140, + 481, + 155 + ], + "type": "interline_equation", + "content": "o _ {t} = q _ {t} S _ {t}; \\quad S _ {t} = S _ {t - 1} + k _ {t} ^ {\\mathsf {T}} v _ {t} \\tag {1}", + "image_path": "a3bf48a21250d4a605663ac81fb822a76a84ae18a4820d76a0860f49b20f21d1.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 167, + 479, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 167, + 479, + 216 + ], + "spans": [ + { + "bbox": [ + 130, + 167, + 479, + 216 + ], + "type": "text", + "content": "where the sequences can be divided into non-overlapping chunks, " + }, + { + "bbox": [ + 130, + 167, + 479, + 216 + ], + "type": "inline_equation", + "content": "\\{q_j,k_j,v_j\\}_{j = 1:t}" + }, + { + "bbox": [ + 130, + 167, + 479, + 216 + ], + "type": "text", + "content": ". The output can be calculated by scanning over the sequence, based on an updated hidden state " + }, + { + "bbox": [ + 130, + 167, + 479, + 216 + ], + "type": "inline_equation", + "content": "S_{t}" + }, + { + "bbox": [ + 130, + 167, + 479, + 216 + ], + "type": "text", + "content": ". This enables the use of chunk-level computation to address the inability to parallelize at the sequence level due to varying lengths." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 217, + 480, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 217, + 480, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 217, + 480, + 361 + ], + "type": "text", + "content": "Inspired by this, we introduce the Scattered Linear Attention (SLA) module, which accommodates linear attention in a window-based voxel transformer. As shown in Figure 1(b), the SLA module treats the voxels of the entire scene into a single sequence and processes them directly without padding voxels. Using the recurrent form of linear attention, we develop an I/O-aware algorithm to further optimize matrix multiplication on voxel sequences. Specifically, we divide the voxel sequence into multiple chunks and loaded them into the shared memory (SRAM) of the GPU. The computation of the hidden state matrix in each window is then achieved through a series of chunk-wise matrix multiplications and cumulative sums. This optimized implementation significantly reduces I/O overhead and memory usage, resulting in extremely fast and memory-efficient attention computation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 362, + 480, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 362, + 480, + 482 + ], + "spans": [ + { + "bbox": [ + 130, + 362, + 480, + 482 + ], + "type": "text", + "content": "In addition, current window-based transformer models use window shifting to propagate the information across the windows. After performing an instruction-level analysis of existing implementations, we observe that the voxel permutation operations caused by window shifting consume significant computational overhead. To address this issue, we propose a cross-window interaction (CWI) module, which is composed of depth-wise convolutions with small 2D kernels and lengthy 1D kernels that allow the voxel features in each window to fully interact with the features in other windows. As a result, the proposed CWI module can improve both the locality and connectivity of the voxel features while requiring minimal computational effort." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 483, + 480, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 483, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 483, + 480, + 568 + ], + "type": "text", + "content": "Building upon the SLA and CWI modules, we propose the ScatterFormer, an innovative voxel transformer for large-scale point cloud understanding. Our experiments show that ScatterFormer achieves linear complexity without compromising accuracy. It achieves superior results compared to the state-of-the-art model DSVT [47]. The latency of ScatterFormer is significantly lower than that of transformer-based detectors [8, 14, 47], which is comparable to that of sparse convolution-based detectors [52]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 480, + 665 + ], + "type": "text", + "content": "In conclusion, we delve into attention on voxels grouped by windows, highlighting the challenges in memory allocation and computation for variable-length sequences. Then, we introduce an SLA module, which can directly process the voxels grouped by windows by facilitating the linear attention formula. A chunkwise matrix multiplication algorithm is proposed to further accelerate the attention computation in SLA. Finally, we present ScatterFormer, which has linear complexity and can efficiently process large-scale LiDAR scenes, achieving better accuracy with lower latency compared to existing transformer-based detectors." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 388, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 388, + 91, + 447, + 101 + ], + "type": "text", + "content": "ScatterFormer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 143, + 362, + 156 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 143, + 362, + 156 + ], + "spans": [ + { + "bbox": [ + 132, + 143, + 362, + 156 + ], + "type": "text", + "content": "2.1 Point Cloud-based 3D Object Detection" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 166, + 481, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 166, + 481, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 166, + 481, + 285 + ], + "type": "text", + "content": "There exist two primary point-cloud representations in 3D object detection, i.e., point-based and voxel-based ones. In point-based methods [28, 29, 37, 39, 55], point clouds are first passed through a point-based backbone network [30], in which the points are gradually sampled and features are learned by point cloud operators. F-Pointnet [29] first employs PoinNet to detect 3D objects based on the frustums lifted by 2D proposals. PointRCNN [37] direct generates 3D proposals from point-based features with foreground segmentation. VoteNet [28] clusters objects from the surface points using a deep Hough voting method. Other point operators based on the point graph [39, 55] and the range view [10] have also been developed for point cloud processing." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 286, + 481, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 286, + 481, + 465 + ], + "spans": [ + { + "bbox": [ + 130, + 286, + 481, + 465 + ], + "type": "text", + "content": "Point-based methods are primarily limited by their inference efficiency and the absence of contextual features in a continuous space. On the other hand, voxel-based approaches [5, 16, 18, 49, 52, 58, 60] transform the entire point clouds into regular grids through voxelization, showcasing superior efficiency and context representation. VoxelNet [60] proposes a voxel feature encoding (VFE) module and combines it with 3D convolutions to extract voxel features in an end-to-end manner. SECOND [49] optimizes 3D convolution for sparse data, resulting in a significant reduction in both time and memory. PointPillars [18] demonstrates an efficient model by stacking voxels into vertical columns, subsequently processing them with naive 2D convolutions. There are also efforts [16, 35] that explore hybrid representations of point-based and voxel-based networks, demonstrating a better trade-off between speed and accuracy. However, a common limitation across these methods is their reliance on small convolution kernels with restricted receptive fields, making them less adept at capturing the global context for 3D object detection." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 466, + 481, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 466, + 481, + 514 + ], + "spans": [ + { + "bbox": [ + 130, + 466, + 481, + 514 + ], + "type": "text", + "content": "In this paper, we shed light on the voxel-based methods and introduce a Scattered Linear Attention module that enables dynamic modeling among voxels within each window with linear complexity, thereby efficiently extracting window-based contextual features." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 535, + 304, + 546 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 535, + 304, + 546 + ], + "spans": [ + { + "bbox": [ + 132, + 535, + 304, + 546 + ], + "type": "text", + "content": "2.2 Transformer on Point Cloud" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 557, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 481, + 666 + ], + "type": "text", + "content": "Inspired by the significant success of self-attention in NLP [46] and CV [7, 45], Transformers have been adapted for 3D vision due to their ability to capture long-range dependencies. The Point Transformers [27, 57] employ attention to modulate point clouds for classification and segmentation tasks. PCT [12] presents an optimized offset attention module, which, when combined with the implicit Laplacian operator and normal estimation, becomes more adept at point-cloud processing. Some methods like [8, 22] opt for voxels or key points to optimize latency. 3DETR [25] proposes a query-based 3D object detection scheme. CT3D [33] enhances the region-based network using a channel-wise" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "C. He et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 118, + 479, + 200 + ], + "blocks": [ + { + "bbox": [ + 134, + 118, + 479, + 200 + ], + "lines": [ + { + "bbox": [ + 134, + 118, + 479, + 200 + ], + "spans": [ + { + "bbox": [ + 134, + 118, + 479, + 200 + ], + "type": "image", + "image_path": "b774ed792faa3e7e125652a569fc710665377ac999a8cc3e7153910cf7ae9407.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 209, + 482, + 255 + ], + "lines": [ + { + "bbox": [ + 130, + 209, + 482, + 255 + ], + "spans": [ + { + "bbox": [ + 130, + 209, + 482, + 255 + ], + "type": "text", + "content": "Fig. 2: The macro design of ScatterFormer. The backbone comprises a Conditional Position Encoding (CPE) and six transformer blocks. Each block is composed by a Scattered Linear Attention (SLA) module, a Cross-Window Interaction (CWI) module, and a Feed-Forward Network (FFN)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 263, + 482, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 263, + 482, + 419 + ], + "spans": [ + { + "bbox": [ + 130, + 263, + 482, + 419 + ], + "type": "text", + "content": "Transformer framework. To achieve context-rich representations, several studies [8, 14, 23, 24, 47] have integrated the attention module into point- or voxel-based encoders. For instance, VoTr [24] utilizes dilated attention for expanded receptive fields; VoxSet [14] applies set attention for extracting point-based features in set-to-set translation; SST [8] employs local attention with shifted windows; and OcTr [59] adopts an Octree-based attention for efficient hierarchical context learning. Nevertheless, how to efficiently leverage global context from attention remains a challenge due to the intrinsic sparsity of point clouds. DSVT [47] and FlatFormer [23] group the voxels within each window into a series of fixed-length voxel sets, thus extracting the features in a fully parallel manner. Recently, a group-free state-space model [54] was proposed to directly process voxels as a sequence. However, these approaches more or less lose the spatial proximity and incur extensive computational overhead in grouping and sorting the voxels." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 437, + 202, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 437, + 202, + 449 + ], + "spans": [ + { + "bbox": [ + 132, + 437, + 202, + 449 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 461, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 666 + ], + "type": "text", + "content": "The overall architecture of our proposed ScatterFormer is depicted in Figure 2. It begins with the input point clouds, which are voxelized and transformed into high-dimensional embeddings using a VFE layer [60]. These embeddings are then processed through Conditional Positional Encoding (CPE) using a shallow convolutional network [6]. The encoded features enter the ScatterFormer backbone, consisting of six ScatterFormer blocks. Each block includes a Scattered Linear Attention (SLA) module, a Cross-Window Interaction (CWI) module, and a Feed-Forward Network (FFN), interspersed with Batch Normalization layers and skip connections. After three ScatterFormer blocks, the voxel features are downsampled via a sparse convolutional layer. The downsampled features are then converted into pillar features [47], generating compact BEV features for bounding-box prediction. ScatterFormer ScatterFormer stands out by not requiring voxel features to be organized into fixed-length sets [8, 23, 47], enabling flexible attention computation across windows. Additionally, the CWI module obviates the need for window shifting. These innovations allow ScatterFormer to avoid unnecessary memory allocation and permutation operations, achieving high efficiency comparable to CNN-based models." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 388, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 388, + 91, + 447, + 101 + ], + "type": "text", + "content": "ScatterFormer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 246, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 246, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 246, + 127 + ], + "type": "text", + "content": "3.1 Linear Attention" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 483, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 483, + 256 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 483, + 256 + ], + "type": "text", + "content": "We begin by revisiting the self-attention introduced by [46]. Given an input matrix " + }, + { + "bbox": [ + 130, + 137, + 483, + 256 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^{N \\times d}" + }, + { + "bbox": [ + 130, + 137, + 483, + 256 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 137, + 483, + 256 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 137, + 483, + 256 + ], + "type": "text", + "content": " denotes the number of tokens and " + }, + { + "bbox": [ + 130, + 137, + 483, + 256 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 137, + 483, + 256 + ], + "type": "text", + "content": " denotes the dimensionality. Self-attention first linearly projects the input to queries, keys, and values using weight matrices " + }, + { + "bbox": [ + 130, + 137, + 483, + 256 + ], + "type": "inline_equation", + "content": "W_{Q}, W_{K}" + }, + { + "bbox": [ + 130, + 137, + 483, + 256 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 137, + 483, + 256 + ], + "type": "inline_equation", + "content": "W_{V}" + }, + { + "bbox": [ + 130, + 137, + 483, + 256 + ], + "type": "text", + "content": ", such that " + }, + { + "bbox": [ + 130, + 137, + 483, + 256 + ], + "type": "inline_equation", + "content": "Q = xW_{Q}, K = xW_{K}, V = xW_{V}" + }, + { + "bbox": [ + 130, + 137, + 483, + 256 + ], + "type": "text", + "content": ". Then it computes a Softmax-based attention map based on queries and keys i.e., " + }, + { + "bbox": [ + 130, + 137, + 483, + 256 + ], + "type": "inline_equation", + "content": "A(Q, K) = \\mathrm{Softmax}(QK^{\\top}) / \\sqrt{d}" + }, + { + "bbox": [ + 130, + 137, + 483, + 256 + ], + "type": "text", + "content": ". The output of self-attention is defined as the weighted sum of " + }, + { + "bbox": [ + 130, + 137, + 483, + 256 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 137, + 483, + 256 + ], + "type": "text", + "content": " values with the weights corresponding to the attention map, i.e., " + }, + { + "bbox": [ + 130, + 137, + 483, + 256 + ], + "type": "inline_equation", + "content": "O = A(Q, K)V" + }, + { + "bbox": [ + 130, + 137, + 483, + 256 + ], + "type": "text", + "content": ". This approach involves calculating the similarity for all query-key pairs, which yields a computational complexity of " + }, + { + "bbox": [ + 130, + 137, + 483, + 256 + ], + "type": "inline_equation", + "content": "O(N^{2})" + }, + { + "bbox": [ + 130, + 137, + 483, + 256 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 257, + 482, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 257, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 257, + 482, + 354 + ], + "type": "text", + "content": "Linear attention [1, 17] offers a notable alternative to self-attention by significantly reducing its computational complexity from " + }, + { + "bbox": [ + 130, + 257, + 482, + 354 + ], + "type": "inline_equation", + "content": "O(N^2)" + }, + { + "bbox": [ + 130, + 257, + 482, + 354 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 257, + 482, + 354 + ], + "type": "inline_equation", + "content": "O(N)" + }, + { + "bbox": [ + 130, + 257, + 482, + 354 + ], + "type": "text", + "content": ". This efficiency is achieved through the application of kernel functions on the query " + }, + { + "bbox": [ + 130, + 257, + 482, + 354 + ], + "type": "inline_equation", + "content": "(Q)" + }, + { + "bbox": [ + 130, + 257, + 482, + 354 + ], + "type": "text", + "content": " and key " + }, + { + "bbox": [ + 130, + 257, + 482, + 354 + ], + "type": "inline_equation", + "content": "(K)" + }, + { + "bbox": [ + 130, + 257, + 482, + 354 + ], + "type": "text", + "content": " matrices, which effectively approximates the original attention map without relying on Softmax, i.e., " + }, + { + "bbox": [ + 130, + 257, + 482, + 354 + ], + "type": "inline_equation", + "content": "A(Q, K) = \\phi(Q)\\phi(K)^{\\mathsf{T}}" + }, + { + "bbox": [ + 130, + 257, + 482, + 354 + ], + "type": "text", + "content": ". Taking advantage of the associative property of matrix multiplication, the computation shifts from " + }, + { + "bbox": [ + 130, + 257, + 482, + 354 + ], + "type": "inline_equation", + "content": "(\\phi(Q)\\phi(K)^{\\mathsf{T}})V" + }, + { + "bbox": [ + 130, + 257, + 482, + 354 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 257, + 482, + 354 + ], + "type": "inline_equation", + "content": "\\phi(Q)(\\phi(K)^{\\mathsf{T}}V)" + }, + { + "bbox": [ + 130, + 257, + 482, + 354 + ], + "type": "text", + "content": ". This modification leads to a computational complexity proportional to the number of tokens, maintaining an order of " + }, + { + "bbox": [ + 130, + 257, + 482, + 354 + ], + "type": "inline_equation", + "content": "O(N)" + }, + { + "bbox": [ + 130, + 257, + 482, + 354 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 371, + 331, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 371, + 331, + 384 + ], + "spans": [ + { + "bbox": [ + 132, + 371, + 331, + 384 + ], + "type": "text", + "content": "3.2 Scattered Linear Attention (SLA)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 392, + 482, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 482, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 482, + 489 + ], + "type": "text", + "content": "Instead of organizing the voxels within a window into fixed-length subsets, we arrange the voxels of the entire scene into a single flattened matrix, as illustrated in Figure 3(a). These voxels are ordered based on their window coordinates, ensuring that voxels from the same window form a sub-matrix in contiguous memory. Initially, we use a shared projection layer to map all the voxels in the scene to query, key, and value representations, denoted " + }, + { + "bbox": [ + 130, + 392, + 482, + 489 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 130, + 392, + 482, + 489 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 392, + 482, + 489 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 392, + 482, + 489 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 392, + 482, + 489 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 392, + 482, + 489 + ], + "type": "text", + "content": ". Subsequently, we perform attention computations on these submatrices separately, resulting in the scattered linear attention (SLA) formula as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 203, + 498, + 480, + 512 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 498, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 203, + 498, + 480, + 512 + ], + "type": "interline_equation", + "content": "\\operatorname {S L A} (Q, K, V) = \\operatorname {C o n c a t} \\left[ \\operatorname {L A} \\left(Q ^ {j}, K ^ {j}, V ^ {j}\\right) \\right] _ {j = 1: M}, \\tag {2}", + "image_path": "c87d2d840d19e902b6c61a56774e7c721efaa86976b11e3ee6a61aa01983fc3f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 521, + 482, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 559 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 521, + 482, + 559 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 521, + 482, + 559 + ], + "type": "text", + "content": " is the number of non-empty window in the current scene and LA is the linear attention formula used in [17]. Based on this formula, the output voxel features " + }, + { + "bbox": [ + 130, + 521, + 482, + 559 + ], + "type": "inline_equation", + "content": "O^j" + }, + { + "bbox": [ + 130, + 521, + 482, + 559 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 130, + 521, + 482, + 559 + ], + "type": "inline_equation", + "content": "j^{\\mathrm{th}}" + }, + { + "bbox": [ + 130, + 521, + 482, + 559 + ], + "type": "text", + "content": " window can be defined as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 236, + 568, + 480, + 600 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 568, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 236, + 568, + 480, + 600 + ], + "type": "interline_equation", + "content": "\\mathrm {L A}: O ^ {j} = \\frac {\\phi (Q) \\sum_ {i = 1} ^ {m ^ {j}} \\phi \\left(K _ {i}\\right) ^ {\\top} V _ {i}}{\\phi (Q) \\sum_ {i = 1} ^ {m ^ {j}} \\phi \\left(K _ {i}\\right) ^ {\\top}} \\tag {3}", + "image_path": "f7eb5887508a378431f55537f3f9d2abc68359b7744b49f49129ed16ef7f03bd.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 609, + 481, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 481, + 621 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 609, + 481, + 621 + ], + "type": "inline_equation", + "content": "m^j" + }, + { + "bbox": [ + 132, + 609, + 481, + 621 + ], + "type": "text", + "content": " is the number of voxel in the window and " + }, + { + "bbox": [ + 132, + 609, + 481, + 621 + ], + "type": "inline_equation", + "content": "\\phi (x)" + }, + { + "bbox": [ + 132, + 609, + 481, + 621 + ], + "type": "text", + "content": " is the kernel function." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Hardware Efficient Implementation. It should be noted that the implementation of Equation 2 is not straightforward, as the sub-matrices have different numbers of rows. One naive implementation is to cache the \"kv\" matrix of each" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "C. He et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 113, + 331, + 232 + ], + "blocks": [ + { + "bbox": [ + 133, + 113, + 331, + 232 + ], + "lines": [ + { + "bbox": [ + 133, + 113, + 331, + 232 + ], + "spans": [ + { + "bbox": [ + 133, + 113, + 331, + 232 + ], + "type": "image", + "image_path": "5816bbdd784f56883d37f1190b2007a02448843582dd37f1427736e0a3847fbe.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 242, + 245, + 252, + 255 + ], + "lines": [ + { + "bbox": [ + 242, + 245, + 252, + 255 + ], + "spans": [ + { + "bbox": [ + 242, + 245, + 252, + 255 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 337, + 116, + 477, + 244 + ], + "blocks": [ + { + "bbox": [ + 337, + 116, + 477, + 244 + ], + "lines": [ + { + "bbox": [ + 337, + 116, + 477, + 244 + ], + "spans": [ + { + "bbox": [ + 337, + 116, + 477, + 244 + ], + "type": "image", + "image_path": "61611b9e5d4ca7f85444d115979bd35c71ff9bb30b2578de2216e84f4a722b90.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 246, + 413, + 255 + ], + "lines": [ + { + "bbox": [ + 403, + 246, + 413, + 255 + ], + "spans": [ + { + "bbox": [ + 403, + 246, + 413, + 255 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 262, + 482, + 340 + ], + "lines": [ + { + "bbox": [ + 130, + 262, + 482, + 340 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 482, + 340 + ], + "type": "text", + "content": "Fig. 3: (a) The voxel features are sorted by windows and flattened to a single matrix, which is further partitioned into multiple chunks of uniform length and sent to SLA kernel for attention computation. (b) In SLA, we first allocate individual threads to each window, each of them iterates over all corresponding key and value chunks, calculates, and accumulates their products to obtain the hidden state matrix of each window. Then, multiple threads are allocated to these query chunks " + }, + { + "bbox": [ + 130, + 262, + 482, + 340 + ], + "type": "inline_equation", + "content": "q_{i} \\in Q" + }, + { + "bbox": [ + 130, + 262, + 482, + 340 + ], + "type": "text", + "content": ", calculating the chunk-wise output by multiplying " + }, + { + "bbox": [ + 130, + 262, + 482, + 340 + ], + "type": "inline_equation", + "content": "q_{i}" + }, + { + "bbox": [ + 130, + 262, + 482, + 340 + ], + "type": "text", + "content": " with the corresponding hidden state matrix." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 351, + 482, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 351, + 482, + 530 + ], + "spans": [ + { + "bbox": [ + 130, + 351, + 482, + 530 + ], + "type": "text", + "content": "voxel and apply " + }, + { + "bbox": [ + 130, + 351, + 482, + 530 + ], + "type": "inline_equation", + "content": "scatter^3" + }, + { + "bbox": [ + 130, + 351, + 482, + 530 + ], + "type": "text", + "content": " operator to accumulate them within each window. However, this would consume a high memory overhead and IO latency. Considering that GPUs feature a memory hierarchy that includes larger, slower global GPU memory (high-bandwidth memory; HBM) and smaller, faster shared memory (SRAM), optimal utilization of SRAM to minimize HBM I/O costs can lead to significant speed-ups. Based on this, we partition the flattened " + }, + { + "bbox": [ + 130, + 351, + 482, + 530 + ], + "type": "inline_equation", + "content": "Q, K, V" + }, + { + "bbox": [ + 130, + 351, + 482, + 530 + ], + "type": "text", + "content": " into multiple chunks, load them from slow HBM to fast SRAM. Specifically, for each window, we allocate an single thread. Each thread iterates over all the key and value chunks, calculates, and accumulates their products to obtain the hidden state matrix of the current window. After calculating the hidden state matrix in window, we assign individual thread to each query chunk " + }, + { + "bbox": [ + 130, + 351, + 482, + 530 + ], + "type": "inline_equation", + "content": "q_i \\in Q^j" + }, + { + "bbox": [ + 130, + 351, + 482, + 530 + ], + "type": "text", + "content": " and calculate the chunk-wise output can then be following Equation 3. Figure 3(b) illustrates this chunk-wise matrix multiplication approach. As can be seen, it not only avoids outputting large matrices but also allows for more flexible processing of variable-length sequences." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 531, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 531, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 531, + 482, + 581 + ], + "type": "text", + "content": "In our implementation, we use Triton [44] to perform chunk-wise matrix multiplication. As illustrated in Figure 4, it outperforms the naive scattered-based operation, demonstrating improved speed and reduced memory usage when computing attention for extremely long sequences." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 594, + 293, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 594, + 293, + 605 + ], + "spans": [ + { + "bbox": [ + 131, + 594, + 293, + 605 + ], + "type": "text", + "content": "3.3 Cross Window Interaction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 612, + 482, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 612, + 482, + 648 + ], + "spans": [ + { + "bbox": [ + 130, + 612, + 482, + 648 + ], + "type": "text", + "content": "Traditional window-based transformers [8,40,47] utilize window shifting for interwindow connection. For point cloud sequences, this requires recalculating window coordinates and rearranging the voxels. In DSVT, the computation of voxel" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 388, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 388, + 91, + 447, + 100 + ], + "type": "text", + "content": "ScatterFormer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 652, + 326, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 652, + 326, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 652, + 326, + 666 + ], + "type": "text", + "content": "3 https://github.com/rustyls/pytorch_scatter" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 173, + 119, + 440, + 220 + ], + "blocks": [ + { + "bbox": [ + 173, + 119, + 440, + 220 + ], + "lines": [ + { + "bbox": [ + 173, + 119, + 440, + 220 + ], + "spans": [ + { + "bbox": [ + 173, + 119, + 440, + 220 + ], + "type": "image", + "image_path": "e19d811d4b8025e814593bf484992f98c2b31d8126b02a975f810475fc70b9bb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 173, + 231, + 440, + 331 + ], + "blocks": [ + { + "bbox": [ + 173, + 231, + 440, + 331 + ], + "lines": [ + { + "bbox": [ + 173, + 231, + 440, + 331 + ], + "spans": [ + { + "bbox": [ + 173, + 231, + 440, + 331 + ], + "type": "image", + "image_path": "80784b1ee3048e8ecee1c136574d82adfefdd08e582f1395abe5b57500be5941.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 346, + 482, + 369 + ], + "lines": [ + { + "bbox": [ + 131, + 346, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 131, + 346, + 482, + 369 + ], + "type": "text", + "content": "Fig. 4: Comparison between between scatter-based implementation (yellow) and our chunk-wise implementation (blue)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 384, + 479, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 479, + 443 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 479, + 443 + ], + "type": "text", + "content": "indices in subsets and the rearrangement of voxel matrices are particularly time consuming, accounting for " + }, + { + "bbox": [ + 130, + 384, + 479, + 443 + ], + "type": "inline_equation", + "content": "24\\%" + }, + { + "bbox": [ + 130, + 384, + 479, + 443 + ], + "type": "text", + "content": " of the total backbone latency, as shown in Figure 5. To alleviate this unnecessary computational overhead from window shifting, we introduce the Cross-Window Interaction (CWI) module, which employs convolutions with large kernel to blend voxel features across windows." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 445, + 482, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 445, + 482, + 563 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 482, + 563 + ], + "type": "text", + "content": "We adhere to the design principles of the Inception architecture [42], where the input features are divided along the channel dimension and processed through multiple branches using group-wise convolutions. Instead of employing a single large kernel for convolution, we adopt more efficient 1D kernels along different axes for feature aggregation. Specifically, we apply a " + }, + { + "bbox": [ + 130, + 445, + 482, + 563 + ], + "type": "inline_equation", + "content": "(S_h + 1) \\times 1 \\times 1" + }, + { + "bbox": [ + 130, + 445, + 482, + 563 + ], + "type": "text", + "content": " convolution in one branch and a " + }, + { + "bbox": [ + 130, + 445, + 482, + 563 + ], + "type": "inline_equation", + "content": "1 \\times (S_w + 1) \\times 1" + }, + { + "bbox": [ + 130, + 445, + 482, + 563 + ], + "type": "text", + "content": " convolution in another branch, where " + }, + { + "bbox": [ + 130, + 445, + 482, + 563 + ], + "type": "inline_equation", + "content": "(S_w, S_h)" + }, + { + "bbox": [ + 130, + 445, + 482, + 563 + ], + "type": "text", + "content": " denotes the window size. Additionally, we incorporate a standard " + }, + { + "bbox": [ + 130, + 445, + 482, + 563 + ], + "type": "inline_equation", + "content": "3 \\times 3 \\times 3" + }, + { + "bbox": [ + 130, + 445, + 482, + 563 + ], + "type": "text", + "content": " convolution in one branch to enhance locality, while another branch performs an identity mapping. Given input " + }, + { + "bbox": [ + 130, + 445, + 482, + 563 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 445, + 482, + 563 + ], + "type": "text", + "content": " of channel " + }, + { + "bbox": [ + 130, + 445, + 482, + 563 + ], + "type": "inline_equation", + "content": "4c" + }, + { + "bbox": [ + 130, + 445, + 482, + 563 + ], + "type": "text", + "content": ", the output " + }, + { + "bbox": [ + 130, + 445, + 482, + 563 + ], + "type": "inline_equation", + "content": "X'" + }, + { + "bbox": [ + 130, + 445, + 482, + 563 + ], + "type": "text", + "content": " of the Cross Window Interaction module can be written as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 230, + 572, + 481, + 587 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 572, + 481, + 587 + ], + "spans": [ + { + "bbox": [ + 230, + 572, + 481, + 587 + ], + "type": "interline_equation", + "content": "X ^ {\\prime} = \\operatorname {C o n c a t} \\left(X _ {k}, X _ {w}, X _ {h}, X ^ {3 c: 4 c}\\right), \\tag {4}", + "image_path": "abde37f6a8aa73e06baf19d3a539442900234b1c3fcd6c2f1d68628572c10cb1.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 596, + 162, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 596, + 162, + 606 + ], + "spans": [ + { + "bbox": [ + 132, + 596, + 162, + 606 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 231, + 617, + 480, + 631 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 617, + 480, + 631 + ], + "spans": [ + { + "bbox": [ + 231, + 617, + 480, + 631 + ], + "type": "interline_equation", + "content": "X _ {k} = \\operatorname {D W C o n v} _ {\\left(S _ {h} + 1\\right) \\times 1 \\times 1} \\left(X ^ {: c}\\right), \\tag {5}", + "image_path": "126d7f2346fddc7e81509cc3d0f89058e03acf288598c9b161328f432df34b68.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 231, + 632, + 480, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 632, + 480, + 647 + ], + "spans": [ + { + "bbox": [ + 231, + 632, + 480, + 647 + ], + "type": "interline_equation", + "content": "X _ {w} = \\operatorname {D W C o n v} _ {1 \\times \\left(S _ {w} + 1\\right) \\times 1} \\left(X ^ {c: 2 c}\\right), \\tag {6}", + "image_path": "6aa7446fc4f21b8184a7524154014c1a84d9120a8555dabce7574a07930ac0b9.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 231, + 649, + 480, + 663 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 649, + 480, + 663 + ], + "spans": [ + { + "bbox": [ + 231, + 649, + 480, + 663 + ], + "type": "interline_equation", + "content": "X _ {h} = \\operatorname {D W C o n v} _ {3 \\times 3 \\times 3} \\left(X ^ {2 c: 3 c}\\right). \\tag {7}", + "image_path": "ace4d34e0cdd7263e3870c29e6d3eb304cfc157651d8e849a8b8961ddf8c8e7c.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "C. He et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 192, + 114, + 305, + 236 + ], + "blocks": [ + { + "bbox": [ + 192, + 114, + 305, + 236 + ], + "lines": [ + { + "bbox": [ + 192, + 114, + 305, + 236 + ], + "spans": [ + { + "bbox": [ + 192, + 114, + 305, + 236 + ], + "type": "image", + "image_path": "b80d62c5b3034d81653c564b43028cf806b236d6d2361938a5c9c4b42f7a6cee.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 244, + 480, + 268 + ], + "lines": [ + { + "bbox": [ + 130, + 244, + 480, + 268 + ], + "spans": [ + { + "bbox": [ + 130, + 244, + 480, + 268 + ], + "type": "text", + "content": "Fig. 5: Runtime decomposition in DSVT [47] and ScatterFormer backbones. The offsets generation means computing the starting address of each chunk in the flattened matrix." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 313, + 114, + 425, + 236 + ], + "blocks": [ + { + "bbox": [ + 313, + 114, + 425, + 236 + ], + "lines": [ + { + "bbox": [ + 313, + 114, + 425, + 236 + ], + "spans": [ + { + "bbox": [ + 313, + 114, + 425, + 236 + ], + "type": "image", + "image_path": "43750b18e59eb1fa21fbec4d709e83e507896764ae1536406d913eae1fddbe25.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 279, + 480, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 279, + 480, + 316 + ], + "spans": [ + { + "bbox": [ + 130, + 279, + 480, + 316 + ], + "type": "text", + "content": "In our experiments, we found that this axis-decomposed large kernel design can achieve a better precision-latency trade-off. With these lengthy kernels, the voxel features can be efficiently blended among different windows." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 330, + 288, + 341 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 330, + 288, + 341 + ], + "spans": [ + { + "bbox": [ + 131, + 330, + 288, + 341 + ], + "type": "text", + "content": "3.4 Detection Head and Loss" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 347, + 482, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 347, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 482, + 456 + ], + "type": "text", + "content": "To complement existing detection heads, ScatterFormer generates dense BEV feature maps from sparse voxel representations by placing them back to their spatial locations and filling the unoccupied positions with zeros. For the BEV network, we simply follow [53] by adopting a cascaded CNN with convolutional blocks at multiple levels of strides. In terms of detection head and loss function, we follow the design identical to that used in DSVT [47]. We train the detection model with heatmap estimation, bounding box regression, and incorporate an IoU loss for confidence calibration. On the nuScenes dataset, the detection head follows the architecture used in Transfusion [2]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 471, + 230, + 484 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 471, + 230, + 484 + ], + "spans": [ + { + "bbox": [ + 131, + 471, + 230, + 484 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 493, + 326, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 493, + 326, + 504 + ], + "spans": [ + { + "bbox": [ + 131, + 493, + 326, + 504 + ], + "type": "text", + "content": "4.1 Datasets and Evaluation Metrics" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 510, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 582 + ], + "type": "text", + "content": "Waymo Open Dataset (WOD). This dataset contains 230,000 annotated samples split into 160,000 for training, 40,000 for validation, and 30,000 for testing. It uses two metrics for 3D object detection: mean average precision (mAP) and mAP weighted by heading accuracy (mAPH), further categorized into Level 1 (L1) for objects detected by more than five LiDAR points and Level 2 (L2) for those detected with at least one point." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "NuScenes. This dataset comprises 40,000 annotated samples, with 28,000 for training, 6,000 for validation, and 6,000 for testing. On this dataset, the model performance is measured by mean average precision (mAP) across multiple distance thresholds (0.5, 1, 2, and 4 meters) and the nuScenes detection score (NDS), which combines mAP with a weighted sum of five additional metrics assessing true positive predictions in translation, scale, orientation, velocity, and attribute accuracy." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 388, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 388, + 91, + 447, + 101 + ], + "type": "text", + "content": "ScatterFormer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 177, + 118, + 433, + 308 + ], + "blocks": [ + { + "bbox": [ + 177, + 118, + 433, + 308 + ], + "lines": [ + { + "bbox": [ + 177, + 118, + 433, + 308 + ], + "spans": [ + { + "bbox": [ + 177, + 118, + 433, + 308 + ], + "type": "image", + "image_path": "1fd59f3f4ba6c1f1f382f901e3edfcb8ab732f8e1bb1ff93745ab0685f48f00b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 316, + 482, + 339 + ], + "lines": [ + { + "bbox": [ + 130, + 316, + 482, + 339 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 482, + 339 + ], + "type": "text", + "content": "Fig. 6: The detection performance (mAPH/L2) vs. speed (FPS) of different methods on Waymo validation set. The speeds are measured on an NVIDIA A100 GPU." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 347, + 280, + 360 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 280, + 360 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 280, + 360 + ], + "type": "text", + "content": "4.2 Implementation Details" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 366, + 482, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 366, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 366, + 482, + 498 + ], + "type": "text", + "content": "Our approach is implemented using the open-source framework OpenPCDet [43]. To construct ScatterFormer, we set the voxel size to " + }, + { + "bbox": [ + 130, + 366, + 482, + 498 + ], + "type": "inline_equation", + "content": "(0.32\\mathrm{m}, 0.32\\mathrm{m}, 0.1875\\mathrm{m})" + }, + { + "bbox": [ + 130, + 366, + 482, + 498 + ], + "type": "text", + "content": " for the Waymo dataset and " + }, + { + "bbox": [ + 130, + 366, + 482, + 498 + ], + "type": "inline_equation", + "content": "(0.3\\mathrm{m}, 0.3\\mathrm{m}, 8\\mathrm{m})" + }, + { + "bbox": [ + 130, + 366, + 482, + 498 + ], + "type": "text", + "content": " for the NuScenes dataset. The window sizes " + }, + { + "bbox": [ + 130, + 366, + 482, + 498 + ], + "type": "inline_equation", + "content": "(S_w, S_h)" + }, + { + "bbox": [ + 130, + 366, + 482, + 498 + ], + "type": "text", + "content": " for the two datasets are set to " + }, + { + "bbox": [ + 130, + 366, + 482, + 498 + ], + "type": "inline_equation", + "content": "(12, 12)" + }, + { + "bbox": [ + 130, + 366, + 482, + 498 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 366, + 482, + 498 + ], + "type": "inline_equation", + "content": "(20, 20)" + }, + { + "bbox": [ + 130, + 366, + 482, + 498 + ], + "type": "text", + "content": ", respectively. We stack six building blocks for the backbone network. We configure our attention module to have 4 heads with a dimensionality of 128. ScatterFormer is trained for 24 epochs with a learning rate of 0.006 on Waymo Dataset and 20 epochs with a learning rate of 0.004 on NuScenes Dataset. In the last 4 epochs, we disabled the data augmentation strategy of ground-truth sampling. The model is trained on 8 RTX A6000 GPUs with a batch size of 32. Other settings for training and inference adhere strictly to DSVT [47]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 514, + 338, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 338, + 528 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 338, + 528 + ], + "type": "text", + "content": "4.3 Comparison with State-of-the-Arts" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": "Results on Waymo Open Dataset (WOD). We conduct a detailed comparison of ScatterFormer against the published results on the validation set of the Waymo Open Dataset. In line with standard practices, we distinctly categorize and list the methods utilizing single-frame and multi-frame approaches. To ensure comprehensive coverage, we also include comparisons with methods incorporating long-term temporal modeling, such as those described in [4,15,61]. As demonstrated in Table 1, our single-stage model outperforms most two-stage methods and achieves better performance than state-of-the-art methods such as DSVT [47] and HEDNet [53]. Moreover, our model achieves 76.0 and 76.7 level 2 mAPH on 3 and 4 frame settings, respectively, outperforming previous multiframe methods by a margin of 1.1. Furthermore, as shown in Figure 6," + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "text", + "content": "C. He et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 158, + 482, + 441 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 1: Performance comparison on the validation set of Waymo Open Dataset. Symbol \\* denotes the methods with temporal modeling, and - means that the result is not available." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 158, + 482, + 441 + ], + "lines": [ + { + "bbox": [ + 132, + 158, + 482, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 158, + 482, + 441 + ], + "type": "table", + "html": "
MethodFramesALL (3D mAPH)Vehicle (AP/APH)Pedestrian (AP/APH)Cyclist (AP/APH)
L1L2L1L2L1L2L1L2
SECOND [49]163.0557.2372.27/71.6963.85/63.3368.70/58.1860.72/51.3160.62/59.2858.34/57.05
PointPillar [18]163.3357.5371.60/71.0063.10/62.5070.60/56.7062.90/50.2064.40/62.3061.90/59.90
IA-SSD [56]164.4858.0870.53/69.6761.55/60.8069.38/58.4760.30/50.7367.67/65.3064.98/62.71
LiDAR R-CNN [20]166.2060.1073.50/73.0064.70/64.2071.20/58.7063.10/51.7068.60/66.9066.10/64.40
RSN [41]1--75.10/74.6066.00/65.5077.80/72.7068.30/63.70--
Part-A2 [38]170.2563.8477.05/76.5168.47/67.9775.24/66.8766.18/58.6268.60/67.3666.13/64.93
Centerpoint [52]1-65.50-- /66.20--/62.60--/67.60
VoTR [24]1--74.95/74.2565.91/65.29----
VoxSeT [14]172.2466.2274.50/74.0365.99/65.5680.03/72.4272.45/65.3971.56/70.2968.95/67.73
SST-1f [8]1--76.22/75.7968.04/67.6481.39/74.0572.82/65.93--
SWFormer-1f [40]1--77.8/77.369.2/68.880.9/72.772.5/64.9--
PV-RCNN [35]169.6363.3377.51/76.8968.98/68.4175.01/65.6566.04/57.6167.81/66.3565.39/63.98
PillarNet [34]174.6068.4379.09/78.5970.92/70.4680.59/74.0172.28/66.1772.29/71.2169.72/68.67
PV-RCNN++ [35]175.2168.6179.10/78.6370.34/69.9180.62/74.6271.86/66.3073.49/72.3870.70/69.62
FlatFormer [23]1-67.2-69.0/68.6-71.5/65.3-68.6/67.5
PillarNext-1f [19]175.7469.7478.40/77.9070.27/69.8182.53/77.1474.90/69.8073.21/72.2070.58/69.62
VoxelNext [5]176.370.178.2/77.769.9/69.481.5/76.373.5/68.676.1/74.973.3/72.2
FSD [9]177.370.879.2/78.870.5/70.182.6/77.373.9/69.177.1/76.074.4/73.3
DSVT-1f [47]178.272.179.7/79.371.4/71.083.7/78.976.1/71.577.5/76.574.6/73.7
HEDNet-1f [53]179.573.481.1/80.673.2/72.784.4/80.076.8/72.678.7/77.775.8/74.9
ScatterFormer (ours)179.773.881.0/80.573.1/72.784.5/79.977.0/72.679.9/78.977.1/76.1
SST-3f [8]3--78.66/78.2169.98/69.5783.81/80.1475.94/72.37--
FlatFormer-3f [23]3-72.0-71.4/71.0-74.5/71.3-74.7/73.7
SWFormer-3f [40]3--79.4/78.971.1/70.682.9/79.074.8/71.1--
PillarNext-3f [19]380.074.580.6/80.172.9/72.485.0/82.178.0/75.278.9/77.976.7/75.7
DSVT-4f [47]481.375.681.8/81.474.1/73.685.6/82.878.6/75.980.4/79.678.1/77.3
ScatterFormer-3f (ours)381.776.082.0/81.475.0/74.185.7/83.279.0/76.180.6/79.578.1/77.7
ScatterFormer-4f (ours)481.976.782.4/81.975.2/74.786.1/83.479.3/76.681.2/80.478.9/78.1
*3D-MAN [51]16--74.53/74.0367.61/67.1471.7/67.762.6/59.0--
*CenterFormer [61]477.073.278.1/77.673.4/72.981.7/78.677.2/74.275.6/74.873.4/72.6
*CenterFormer [61]877.373.778.8/78.374.3/73.882.1/79.377.8/75.075.2/74.473.2/72.3
*MPPNet [4]479.8374.2281.54/81.0674.07/73.6184.56/81.9477.20/74.6777.15/76.5075.01/74.38
*MPPNet [4]1680.4074.8582.74/82.2875.41/74.9684.69/82.2577.43/75.0677.28/76.6675.13/74.52
*MSF [15]480.2074.6281.36/80.8773.81/73.3585.05/82.1077.92/75.1178.40/77.6176.17/75.40
*MSF [15]880.6575.4682.83/82.0175.76/75.3185.24/82.2178.32/75.6178.52/77.7476.32/75.47
", + "image_path": "dbd41322d5fee2d2d964d0ffac534ba7a10f74057dfb00d9d16321d443125711.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 449, + 482, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 556 + ], + "type": "text", + "content": "our method significantly enhances the detection runtime. This improvement is attributed to our use of Linear Attention, which eliminates the need for extensive voxel sorting and padding operations. Furthermore, as presented in Table 2, ScatterFormer achieves the highest Level 1 and Level 2 mAPH scores in all categories of the Waymo test set, highlighting its superior detection quality and precision. Remarkably, our approach, as a one-shot method, delivers performance that is on par with methods that require temporal modeling. This finding underscores the critical importance of network design and architecture in achieving high-performance architecture." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": "Results on NuScenes. We compare ScatterFormer with the previous best performing methods on the nuScenes dataset. As shown in Table 3, ScatterFormer achieves the state-of-the-art performance in terms of val NDS (72.4) and mAP (68.3), surpassing HEDNet [53] by 1.0 and 1.6 respectively. Compared to TransFusion [2], ScatterFormer achieves a significant improvement of " + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "inline_equation", + "content": "2.8\\%" + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": " in mAP, demonstrating that its backbone network is superior to the sparse convolutional network. ScatterFormer exhibits particularly strong performance on categories like Bus, Bike, Construction Vehicle and Trailer, suggesting its ability to effectively capture and encode detailed contextual features." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 388, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 388, + 91, + 447, + 101 + ], + "type": "text", + "content": "ScatterFormer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 141, + 481, + 244 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 2: Performance comparison on the test set of Waymo Open Dataset. " + }, + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": " means that the result is not available." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 141, + 481, + 244 + ], + "lines": [ + { + "bbox": [ + 133, + 141, + 481, + 244 + ], + "spans": [ + { + "bbox": [ + 133, + 141, + 481, + 244 + ], + "type": "table", + "html": "
MethodALL (3D mAPH)Vehicle (AP/APH)Pedestrian (AP/APH)Cyclist (AP/APH)
L1L2L1L2L1L2L1L2
PointPillar [18]--68.1060.1068.00/55.5061.40/50.10--
StarNet [26]--61.0054.5067.80/59.9061.10/54.00--
M3DETR [11]67.161.977.7/77.170.5/70.068.2/58.560.6/52.067.3/65.765.3/63.8
3D-MAN [51]--78.2869.9869.97/65.9863.98/60.26--
PV-RCNN++ [36]75.770.281.6/81.273.9/73.580.4/75.074.1/69.071.9/70.869.3/68.2
CenterPoint [52]77.271.981.1/80.673.4/73.080.5/77.374.6/71.574.6/73.772.2/71.3
RSN [41]--80.3071.6078.90/75.6070.70/67.80--
SST-3f [8]78.372.881.0/80.673.1/72.783.3/79.776.9/73.575.7/74.673.2/72.2
HEDNet [53]79.0573.7783.78/83.3976.33/75.9683.46/78.9877.53/73.2575.86/74.7873.13/72.09
PillarNext-3f [19]79.0074.0983.28/82.8376.18/75.7684.40/81.4478.84/75.9873.77/72.7371.56/70.55
ScatterFormer-4f (ours)80.7175.8485.60/85.1378.34/78.0784.58/81.6479.12/76.3276.20/75.3574.02/73.13
", + "image_path": "fcff4ef839ba2f630c72e3792b710fe627801b4b12e06777d3e0d4f02b0080a8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 133, + 273, + 481, + 360 + ], + "blocks": [ + { + "bbox": [ + 180, + 257, + 432, + 268 + ], + "lines": [ + { + "bbox": [ + 180, + 257, + 432, + 268 + ], + "spans": [ + { + "bbox": [ + 180, + 257, + 432, + 268 + ], + "type": "text", + "content": "Table 3: The performance on the validation set of NuScenes." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 273, + 481, + 360 + ], + "lines": [ + { + "bbox": [ + 133, + 273, + 481, + 360 + ], + "spans": [ + { + "bbox": [ + 133, + 273, + 481, + 360 + ], + "type": "table", + "html": "
MethodNDSmAPCarTruckBusT.L.C.V.Ped.M.T.BikeT.C.B.R.
CenterPoint [52]66.559.284.957.470.738.116.985.159.042.069.868.3
VoxelNeXt [5]66.760.583.955.570.538.121.184.662.850.069.469.4
TransFusion-L [2]70.165.586.960.873.143.425.287.572.957.377.270.3
PillarNext [19]68.462.285.057.467.635.620.686.868.653.177.369.7
DSVT [47]71.166.487.462.675.942.125.388.274.858.777.870.9
HEDNet [53]71.466.787.760.677.850.728.987.174.356.876.366.9
ScatterFormer (ours)72.468.388.665.479.345.429.188.774.761.578.272.3
", + "image_path": "13f4907d846e33a141b22a6f98790cce5416553f711092447b0cc9d85a7f45cc.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 370, + 238, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 370, + 238, + 383 + ], + "spans": [ + { + "bbox": [ + 132, + 370, + 238, + 383 + ], + "type": "text", + "content": "4.4 Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 390, + 482, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 482, + 474 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 482, + 474 + ], + "type": "text", + "content": "In this section, we conduct ablation experiments in ScatterFormer using " + }, + { + "bbox": [ + 130, + 390, + 482, + 474 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 390, + 482, + 474 + ], + "type": "text", + "content": " of the Waymo training and validation data. Each model was trained for 24 epochs and run on default parameters. Table 4 shows four different configurations of ScatterFormer: (a) removing the Scattered Linear Attention module (SLA), (b) removing the Cross Window Interaction (CWI) module, (c) replacing the CWI with Shifted Window (SW) approach, and (d) removing Conditional Position Encoding (CPE) module." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 474, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 593 + ], + "type": "text", + "content": "As can be seen in Table 4, configuration (a) results in ScatterFormer experiencing a decrease of " + }, + { + "bbox": [ + 130, + 474, + 482, + 593 + ], + "type": "inline_equation", + "content": "(3.1\\%, 3.9\\%, \\text{and} 3.2\\%)" + }, + { + "bbox": [ + 130, + 474, + 482, + 593 + ], + "type": "text", + "content": " APH on three categories. This highlights the effectiveness of the SLA, as it successfully captures long-range contexts and facilitates dynamic feature extraction for object detection. In configuration (b), ScatterFormer, which lacks cross-window interaction capability, exhibits a performance drop of " + }, + { + "bbox": [ + 130, + 474, + 482, + 593 + ], + "type": "inline_equation", + "content": "(1.9\\%, 1.0\\%, \\text{and} 2.0\\%)" + }, + { + "bbox": [ + 130, + 474, + 482, + 593 + ], + "type": "text", + "content": " APH across the three categories. While the Shifted Window approach in configuration (c) enhances performance, the improvement is not as substantial as that achieved by the proposed CWI module. The results in configuration (d) indicate that CPE contributes to a certain degree of performance enhancement." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Window Size. In Table 5, We tested the impact of different window sizes on the detection performance of various target types. It can be observed that ScatterFormer is not highly sensitive to window size. Slightly increasing the window size can enhance the detection performance for small targets, indicating that for objects like pedestrians and bicycles, where the point cloud is overly sparse, appropriate contextual information can enhance the recognition of these" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "C. He et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 192, + 163, + 420, + 240 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 4: Ablation experiments on the validation set of Waymo Open Dataset. \"SLA\" and \"CWT\" refer to the proposed Scattered Linear Attention and Cross-Window Interaction modules, respectively. \"CPE\" refers to Conditional Position Embedding. AP and APH scores on LEVEL2 are reported." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 192, + 163, + 420, + 240 + ], + "lines": [ + { + "bbox": [ + 192, + 163, + 420, + 240 + ], + "spans": [ + { + "bbox": [ + 192, + 163, + 420, + 240 + ], + "type": "table", + "html": "
AblationVeh.Ped.Cyc.
-baseline71.6/71.176.0/70.974.2/73.1
(a)w/o SLA module68.5/68.073.3/67.071.0/69.9
(b)w/o CWI module69.6/69.274.0/69.972.1/71.1
(c)CWI → SW69.9/69.575.0/70.473.8/72.4
(d)w/o CPE70.2/69.873.3/67.272.4/70.5
", + "image_path": "fdd3990eb923f3fb19e0239113488a5c2ae51894b36edd983dab1d18c1c8685a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 213, + 275, + 396, + 341 + ], + "blocks": [ + { + "bbox": [ + 130, + 249, + 482, + 271 + ], + "lines": [ + { + "bbox": [ + 130, + 249, + 482, + 271 + ], + "spans": [ + { + "bbox": [ + 130, + 249, + 482, + 271 + ], + "type": "text", + "content": "Table 5: Comparison of model performance using different window sizes at various stages. The APH (L2) scores on different categories are reported." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 213, + 275, + 396, + 341 + ], + "lines": [ + { + "bbox": [ + 213, + 275, + 396, + 341 + ], + "spans": [ + { + "bbox": [ + 213, + 275, + 396, + 341 + ], + "type": "table", + "html": "
Window SizeRegion SizeVeh. Ped. Cyc.
103.20 m70.2 69.2 71.0
123.84 m71.1 70.9 73.1
144.48 m71.3 69.9 72.5
165.12 m71.0 69.2 72.3
", + "image_path": "f503a21c32aeb51147d230bffb9bdc7aa5b0e2ca5cf4f140f44059142530c64e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 356, + 482, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 356, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 356, + 482, + 392 + ], + "type": "text", + "content": "objects. However, excessively large windows can degrade the performance. We speculate that this is because larger windows encompass more tokens, thereby mitigating the focusing power of linear attention." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 397, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 397, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 397, + 482, + 529 + ], + "type": "text", + "content": "Kernel Design in CWI. Table 6 presents the performance under various kernel configurations in the Cross-Window Interaction (CWI) module. Increasing the number of parameters in the kernel significantly enhances performance. For example, using a large " + }, + { + "bbox": [ + 130, + 397, + 482, + 529 + ], + "type": "inline_equation", + "content": "7 \\times 7" + }, + { + "bbox": [ + 130, + 397, + 482, + 529 + ], + "type": "text", + "content": " kernel brings notable performance improvements, but also introduces more latency. Additionally, larger kernel strides enhance the exchange rate of tokens across windows. For example, using dilated convolutions in CWI performs slightly better than using non-dilated convolutions. Instead of scaling up large, we decompose large kernel into more efficient 1D kernels, which achieves the optimal latency while keeping comparable performance. As the current Sconv library lacks support for depth-wise convolution, we have developed a modified version to address this limitation." + }, + { + "bbox": [ + 130, + 397, + 482, + 529 + ], + "type": "inline_equation", + "content": "^4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 545, + 388, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 545, + 388, + 558 + ], + "spans": [ + { + "bbox": [ + 131, + 545, + 388, + 558 + ], + "type": "text", + "content": "4.5 Comparison with Different Linear Attentions" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 563, + 482, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 563, + 482, + 648 + ], + "spans": [ + { + "bbox": [ + 130, + 563, + 482, + 648 + ], + "type": "text", + "content": "We further explored the performance of ScatterFormer by conducting experiments with several other linear attention mechanisms, including Gated Linear Attention [32], Cross-covariance Attention (XCA) [1], and Focused Linear Attention [13]. The results are summarized in Table 7. Our findings suggest that the window-based attention framework employed by ScatterFormer is not highly sensitive to the specific form of linear attention used. Notably, Efficient Attention [3] performs slightly worse than XCA [1] in the Vehicle class but better in the" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 388, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 388, + 91, + 447, + 101 + ], + "type": "text", + "content": "ScatterFormer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 315, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 315, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 315, + 666 + ], + "type": "text", + "content": "4 https://github.com/skyhehe123/spconvv" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 216, + 163, + 392, + 228 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 6: Performance with various kernel configurations in the Cross-Window Integration (CWI) module. Each configuration is specified by the notation \"KxDy\", where \"Kx\" refers to the kernel size, and \"Dy\" denotes the dilation rate. The APH (L2) scores on different categories are reported." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 216, + 163, + 392, + 228 + ], + "lines": [ + { + "bbox": [ + 216, + 163, + 392, + 228 + ], + "spans": [ + { + "bbox": [ + 216, + 163, + 392, + 228 + ], + "type": "table", + "html": "
KernelLatencyVeh.Ped.Cyc.
Conv1DK13 × 247ms71.170.973.1
Conv2DK5D149ms69.569.071.1
Conv2DK5D249ms70.770.472.8
Conv2DK7D165ms71.571.072.5
", + "image_path": "1beea83c0bef77cd35b6c1d07c97779ddf13b5af87fb74019d59b9e331bb9bfc.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 216, + 258, + 392, + 323 + ], + "blocks": [ + { + "bbox": [ + 130, + 232, + 481, + 255 + ], + "lines": [ + { + "bbox": [ + 130, + 232, + 481, + 255 + ], + "spans": [ + { + "bbox": [ + 130, + 232, + 481, + 255 + ], + "type": "text", + "content": "Table 7: Performance with different linear attention designs. The APH (L2) scores on different categories are reported." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 216, + 258, + 392, + 323 + ], + "lines": [ + { + "bbox": [ + 216, + 258, + 392, + 323 + ], + "spans": [ + { + "bbox": [ + 216, + 258, + 392, + 323 + ], + "type": "table", + "html": "
Linear AttentionVeh.Ped.Cyc.
Efficient Attn [3]71.170.973.1
Gated Linear Attn [32]69.969.072.5
Focused Linear Attn [13]70.469.572.3
XCA [1]71.570.472.8
", + "image_path": "1df3f6596a582e606f11d55510e8a1edd15190e76e9ba2934624709b097801e3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 334, + 481, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 334, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 334, + 481, + 369 + ], + "type": "text", + "content": "Pedestrian and Cyclist classes. Focused Linear Attention introduces additional computational overhead, while Gated Linear Attention adds more learnable parameters to the model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 383, + 218, + 394 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 383, + 218, + 394 + ], + "spans": [ + { + "bbox": [ + 131, + 383, + 218, + 394 + ], + "type": "text", + "content": "4.6 Limitations" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 404, + 482, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 404, + 482, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 404, + 482, + 489 + ], + "type": "text", + "content": "ScatterFormer relies on our customized operators, which have not yet been implemented as plugins in TensorRT. Therefore, deploying ScatterFormer on in-vehicle devices will require additional engineering efforts. Despite this, ScatterFormer is more efficient than current models based on sparse convolution and traditional attention when used on consumer-grade GPU cards. Furthermore, ScatterFormer can be optimized by dynamically partitioning matrices according to different GPU architectures to leverage TensorCore for hardware acceleration." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 507, + 220, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 507, + 220, + 520 + ], + "spans": [ + { + "bbox": [ + 131, + 507, + 220, + 520 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "We introduced ScatterFormer, an innovative architecture designed for 3D object detection using point clouds, specifically targeting the challenges associated with processing sparse and unevenly distributed data from LiDAR sensors. The cornerstone of our approach is the Scattered Linear Attention (SLA) module, which effectively addresses the limitations of conventional attention mechanisms in managing voxel features of varying lengths. SLA ingeniously combines linear attention with a chunk-wise matrix multiplication algorithm, tailored to meet the distinct requirements of processing voxels grouped by windows. By integrating SLA with a novel cross-window interaction (CWI) module, ScatterFormer achieves higher accuracy and lower latency, surpassing traditional transformer-based and sparse CNN-based detectors in extensive 3D detection tasks." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "C. He et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 139, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 139, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 139, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 138, + 139, + 481, + 183 + ], + "type": "text", + "content": "1. Ali, A., Touvron, H., Caron, M., Bojanowski, P., Douze, M., Joulin, A., Laptev, I., Neverova, N., Synnaeve, G., Verbeek, J., et al.: Xcit: Cross-covariance image transformers. Advances in neural information processing systems 34, 20014-20027 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 184, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 184, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 138, + 184, + 481, + 226 + ], + "type": "text", + "content": "2. Bai, X., Hu, Z., Zhu, X., Huang, Q., Chen, Y., Fu, H., Tai, C.L.: Transfusion: Robust lidar-camera fusion for 3d object detection with transformers. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1090-1099 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 228, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 481, + 259 + ], + "type": "text", + "content": "3. Cai, H., Gan, C., Han, S.: Efficientvit: Enhanced linear attention for high-resolution low-computation visual recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 260, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 260, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 138, + 260, + 481, + 293 + ], + "type": "text", + "content": "4. Chen, X., Shi, S., Zhu, B., Cheung, K.C., Xu, H., Li, H.: Mppnet: Multi-frame feature intertwining with proxy points for 3d temporal object detection. In: European Conference on Computer Vision. pp. 680-697. Springer (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 293, + 481, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 293, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 138, + 293, + 481, + 326 + ], + "type": "text", + "content": "5. Chen, Y., Liu, J., Zhang, X., Qi, X., Jia, J.: Voxelnext: Fully sparse voxelnet for 3d object detection and tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21674-21683 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 326, + 481, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 326, + 481, + 347 + ], + "spans": [ + { + "bbox": [ + 138, + 326, + 481, + 347 + ], + "type": "text", + "content": "6. Chu, X., Tian, Z., Zhang, B., Wang, X., Shen, C.: Conditional positional encodings for vision transformers. ICLR (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 348, + 481, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 348, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 138, + 348, + 481, + 392 + ], + "type": "text", + "content": "7. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. ICLR (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 392, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 392, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 138, + 392, + 481, + 435 + ], + "type": "text", + "content": "8. Fan, L., Pang, Z., Zhang, T., Wang, Y.X., Zhao, H., Wang, F., Wang, N., Zhang, Z.: Embracing single stride 3d object detector with sparse transformer. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8458-8468 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 435, + 481, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 435, + 481, + 457 + ], + "spans": [ + { + "bbox": [ + 138, + 435, + 481, + 457 + ], + "type": "text", + "content": "9. Fan, L., Wang, F., Wang, N., Zhang, Z.: Fully sparse 3d object detection. Advances in Neural Information Processing Systems 35, 351-363 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 458, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 458, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 138, + 458, + 481, + 490 + ], + "type": "text", + "content": "0. Fan, L., Xiong, X., Wang, F., Wang, N., Zhang, Z.: Rangedet: In defense of range view for lidar-based 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2918-2927 (October 2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 491, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 491, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 491, + 481, + 533 + ], + "type": "text", + "content": "1. Guan, T., Wang, J., Lan, S., Chandra, R., Wu, Z., Davis, L., Manocha, D.: M3detr: Multi-representation, multi-scale, mutual-relation 3d object detection with transformers. In: Proceedings of the IEEE/CVF winter conference on applications of computer vision. pp. 772-782 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 534, + 481, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 534, + 481, + 576 + ], + "spans": [ + { + "bbox": [ + 138, + 534, + 481, + 576 + ], + "type": "text", + "content": "2. Guo, M.H., Cai, J.X., Liu, Z.N., Mu, T.J., Martin, R.R., Hu, S.M.: Pct: Point cloud transformer. Computational Visual Media 7(2), 187-199 (Apr 2021). https://doi.org/10.1007/s41095-021-0229-5, http://dx.doi.org/10.1007/s41095-021-0229-5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 578, + 481, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 481, + 610 + ], + "type": "text", + "content": "3. Han, D., Pan, X., Han, Y., Song, S., Huang, G.: Flatten transformer: Vision transformer using focused linear attention. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5961-5971 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 611, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 481, + 643 + ], + "type": "text", + "content": "4. He, C., Li, R., Li, S., Zhang, L.: Voxel set transformer: A set-to-set approach to 3d object detection from point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8417-8427 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 644, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 644, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 644, + 481, + 665 + ], + "type": "text", + "content": "5. He, C., Li, R., Zhang, Y., Li, S., Zhang, L.: Msf: Motion-guided sequential fusion for efficient 3d object detection from point cloud sequences. In: Proceedings of the" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 388, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 388, + 91, + 447, + 101 + ], + "type": "text", + "content": "ScatterFormer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 147, + 116, + 481, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 116, + 481, + 138 + ], + "spans": [ + { + "bbox": [ + 147, + 116, + 481, + 138 + ], + "type": "text", + "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5196-5205 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 138, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 138, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 138, + 481, + 171 + ], + "type": "text", + "content": "16. He, C., Zeng, H., Huang, J., Hua, X.S., Zhang, L.: Structure aware single-stage 3d object detection from point cloud. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11873-11882 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 171, + 481, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 171, + 481, + 203 + ], + "spans": [ + { + "bbox": [ + 133, + 171, + 481, + 203 + ], + "type": "text", + "content": "17. Katharopoulos, A., Vyas, A., Pappas, N., Fleuret, F.: Transformers are rnns: Fast autoregressive transformers with linear attention. In: International conference on machine learning. pp. 5156-5165. PMLR (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 203, + 481, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 203, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 133, + 203, + 481, + 236 + ], + "type": "text", + "content": "18. Lang, A.H., Vora, S., Caesar, H., Zhou, L., Yang, J., Beijbom, O.: Pointpillars: Fast encoders for object detection from point clouds. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 12697-12705 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 236, + 481, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 236, + 481, + 267 + ], + "spans": [ + { + "bbox": [ + 133, + 236, + 481, + 267 + ], + "type": "text", + "content": "19. Li, J., Luo, C., Yang, X.: Rethinking network designs for 3d object detection in lidar point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17567-17576 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 267, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 267, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 133, + 267, + 481, + 300 + ], + "type": "text", + "content": "20. Li, Z., Wang, F., Wang, N.: Lidar r-cnn: An efficient and universal 3d object detector. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 300, + 481, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 481, + 342 + ], + "type": "text", + "content": "21. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 10012-10022 (October 2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 342, + 481, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 481, + 374 + ], + "type": "text", + "content": "22. Liu, Z., Zhang, Z., Cao, Y., Hu, H., Tong, X.: Group-free 3d object detection via transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2949-2958 (October 2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "type": "text", + "content": "23. Liu, Z., Yang, X., Tang, H., Yang, S., Han, S.: Flatformer: Flattened window attention for efficient point cloud transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1200-1211 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 407, + 481, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 481, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 481, + 439 + ], + "type": "text", + "content": "24. Mao, J., Xue, Y., Niu, M., Bai, H., Feng, J., Liang, X., Xu, H., Xu, C.: Voxel transformer for 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 3164-3173 (October 2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 439, + 481, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 439, + 481, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 439, + 481, + 472 + ], + "type": "text", + "content": "25. Misra, I., Girdhar, R., Joulin, A.: An end-to-end transformer model for 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2906-2917 (October 2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 472, + 481, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 481, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 481, + 504 + ], + "type": "text", + "content": "26. Ngiam, J., Caine, B., Han, W., Yang, B., Chai, Y., Sun, P., Zhou, Y., Yi, X., Alsharif, O., Nguyen, P., et al.: Starnet: Targeted computation for object detection in point clouds. arXiv preprint arXiv:1908.11069 (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 504, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 481, + 536 + ], + "type": "text", + "content": "27. Pan, X., Xia, Z., Song, S., Li, L.E., Huang, G.: 3d object detection with pointformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7463-7472 (June 2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 536, + 481, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 481, + 568 + ], + "type": "text", + "content": "28. Qi, C.R., Litany, O., He, K., Guibas, L.J.: Deep though voting for 3d object detection in point clouds. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (October 2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "type": "text", + "content": "29. Qi, C.R., Liu, W., Wu, C., Su, H., Guibas, L.J.: Frustum pointnets for 3d object detection from rgb-d data. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 918-927 (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 601, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 481, + 632 + ], + "type": "text", + "content": "30. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 652-660 (2017)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "text", + "content": "31. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In: Advances in neural information processing systems. pp. 5099-5108 (2017)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "C. He et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "type": "text", + "content": "32. Qin, Z., Li, D., Sun, W., Sun, W., Shen, X., Han, X., Wei, Y., Lv, B., Yuan, F., Luo, X., Qiao, Y., Zhong, Y.: Scaling transnormer to 175 billion parameters (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 481, + 183 + ], + "type": "text", + "content": "33. Sheng, H., Cai, S., Liu, Y., Deng, B., Huang, J., Hua, X.S., Zhao, M.J.: Improving 3d object detection with channel-wise transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2743-2752 (October 2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 185, + 480, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 185, + 480, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 185, + 480, + 217 + ], + "type": "text", + "content": "34. Shi, G., Li, R., Ma, C.: Pillarnet: Real-time and high-performance pillar-based 3d object detection. In: European Conference on Computer Vision. pp. 35-52. Springer (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 218, + 480, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 218, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 218, + 480, + 251 + ], + "type": "text", + "content": "35. Shi, S., Guo, C., Jiang, L., Wang, Z., Shi, J., Wang, X., Li, H.: Pv-rcnn: Pointvoxel feature set abstraction for 3d object detection. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 252, + 480, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 252, + 480, + 296 + ], + "spans": [ + { + "bbox": [ + 132, + 252, + 480, + 296 + ], + "type": "text", + "content": "36. Shi, S., Jiang, L., Deng, J., Wang, Z., Guo, C., Shi, J., Wang, X., Li, H.: Pvrnn++: Point-voxel feature set abstraction with local vector representation for 3d object detection. International Journal of Computer Vision 131(2), 531-551 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 297, + 480, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 297, + 480, + 329 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 480, + 329 + ], + "type": "text", + "content": "37. Shi, S., Wang, X., Li, H.: Pointrecnn: 3d object proposal generation and detection from point cloud. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 770-779 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 331, + 480, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 331, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 331, + 480, + 363 + ], + "type": "text", + "content": "38. Shi, S., Wang, Z., Wang, X., Li, H.: Part-a^2 net: 3d part-aware and aggregation neural network for object detection from point cloud. arXiv preprint arXiv:1907.03670 (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 365, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 480, + 396 + ], + "type": "text", + "content": "39. Shi, W., Rajkumar, R.: Point-gnn: Graph neural network for 3d object detection in a point cloud. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1711-1719 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 397, + 480, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 480, + 430 + ], + "type": "text", + "content": "40. Sun, P., Tan, M., Wang, W., Liu, C., Xia, F., Leng, Z., Anguelov, D.: Swformer: Sparse window transformer for 3d object detection in point clouds. In: European Conference on Computer Vision. pp. 426-442. Springer (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 432, + 480, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 432, + 480, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 432, + 480, + 475 + ], + "type": "text", + "content": "41. Sun, P., Wang, W., Chai, Y., Elsayed, G., Bewley, A., Zhang, X., Sminchisescu, C., Anguelov, D.: Rsn: Range sparse net for efficient, accurate lidar 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5725-5734 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 476, + 480, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 476, + 480, + 508 + ], + "spans": [ + { + "bbox": [ + 132, + 476, + 480, + 508 + ], + "type": "text", + "content": "42. Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., Wojna, Z.: Rethinking the inception architecture for computer vision. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2818-2826 (2016)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 510, + 480, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 510, + 480, + 532 + ], + "spans": [ + { + "bbox": [ + 132, + 510, + 480, + 532 + ], + "type": "text", + "content": "43. Team, O.D.: Openpcdet: An open-source toolbox for 3d object detection from point clouds. https://github.com/open-mmlab/OpenPCDet (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 533, + 480, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 480, + 597 + ], + "type": "text", + "content": "44. Tilet, P., Kung, H.T., Cox, D.: Triton: An intermediate language and compiler for tiled neural network computations. In: Proceedings of the 3rd ACM SIGPLAN International Workshop on Machine Learning and Programming Languages. p. 10-19. MAPL 2019, Association for Computing Machinery, New York, NY, USA (2019). https://doi.org/10.1145/3315508.3329973, https://doi.org/10.1145/3315508.3329973" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 599, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 480, + 632 + ], + "type": "text", + "content": "45. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jegou, H.: Training data-efficient image transformers & distillation through attention. In: International Conference on Machine Learning. vol. 139, pp. 10347-10357 (July 2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "text", + "content": "46. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: Advances in neural information processing systems. pp. 5998-6008 (2017)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 388, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 388, + 91, + 447, + 100 + ], + "type": "text", + "content": "ScatterFormer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 611 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "text", + "content": "47. Wang, H., Shi, C., Shi, S., Lei, M., Wang, S., He, D., Schiele, B., Wang, L.: Dsvt: Dynamic sparse voxel transformer with rotated sets. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13520-13529 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 162, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 162, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 162, + 481, + 183 + ], + "type": "text", + "content": "48. Wang, S., Li, B.Z., Khabsa, M., Fang, H., Ma, H.: Linformer: Self-attention with linear complexity (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 184, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 184, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 184, + 481, + 205 + ], + "type": "text", + "content": "49. Yan, Y., Mao, Y., Li, B.: Second: Sparsely embedded convolutional detection. Sensors 18(10), 3337 (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 205, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 205, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 134, + 205, + 481, + 237 + ], + "type": "text", + "content": "50. Yang, B., Luo, W., Urtasun, R.: Pixor: Real-time 3d object detection from point clouds. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 7652-7660 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 238, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 238, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 133, + 238, + 481, + 270 + ], + "type": "text", + "content": "51. Yang, Z., Zhou, Y., Chen, Z., Ngiam, J.: 3d-man: 3d multi-frame attention network for object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1863-1872 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 134, + 271, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 271, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 134, + 271, + 481, + 303 + ], + "type": "text", + "content": "52. Yin, T., Zhou, X., Krahenbuhl, P.: Center-based 3d object detection and tracking. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11784-11793 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 134, + 304, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 304, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 134, + 304, + 481, + 335 + ], + "type": "text", + "content": "53. Zhang, G., Chen, J., Gao, G., Li, J., Hu, X.: Hednet: A hierarchical encoder-decoder network for 3d object detection in point clouds. arXiv preprint arXiv:2310.20234 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 134, + 337, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 337, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 134, + 337, + 481, + 369 + ], + "type": "text", + "content": "54. Zhang, G., Fan, L., He, C., Lei, Z., Zhang, Z., Zhang, L.: Voxel mamba: Group-free state space models for point cloud based 3d object detection. arXiv preprint arXiv:2406.10700 (2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 134, + 369, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 369, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 134, + 369, + 481, + 402 + ], + "type": "text", + "content": "55. Zhang, Y., Huang, D., Wang, Y.: Pc-rgnn: Point cloud completion and graph neural network for 3d object detection. In: Proceedings of the AAAI conference on artificial intelligence. vol. 35, pp. 3430-3437 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 134, + 403, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 403, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 134, + 403, + 481, + 445 + ], + "type": "text", + "content": "56. Zhang, Y., Hu, Q., Xu, G., Ma, Y., Wan, J., Guo, Y.: Not all points are equal: Learning highly efficient point-based detectors for 3d lidar point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18953-18962 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 134, + 447, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 447, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 134, + 447, + 481, + 479 + ], + "type": "text", + "content": "57. Zhao, H., Jiang, L., Jia, J., Torr, P.H., Koltun, V.: Point transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 16259-16268 (October 2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 479, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 479, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 134, + 479, + 481, + 511 + ], + "type": "text", + "content": "58. Zheng, W., Tang, W., Jiang, L., Fu, C.W.: Se-ssd: Self-ensembling single-stage object detector from point cloud. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14494-14503 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 512, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 512, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 134, + 512, + 481, + 544 + ], + "type": "text", + "content": "59. Zhou, C., Zhang, Y., Chen, J., Huang, D.: Octr: Octtree-based transformer for 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5166-5175 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 545, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 545, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 134, + 545, + 481, + 578 + ], + "type": "text", + "content": "60. Zhou, Y., Tuzel, O.: Voxelnet: End-to-end learning for point cloud based 3d object detection. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4490-4499 (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 578, + 481, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 578, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 134, + 578, + 481, + 611 + ], + "type": "text", + "content": "61. Zhou, Z., Zhao, X., Wang, Y., Wang, P., Foroosh, H.: Centerformer: Center-based transformer for 3d object detection. In: European Conference on Computer Vision. pp. 496-513. Springer (2022)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "C. He et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Scene Coordinate Reconstruction_ Posing of Image Collections via Incremental Learning of a Relocalizer/c513fe61-35e0-435f-acdb-dbf8715686b2_content_list.json b/2024/Scene Coordinate Reconstruction_ Posing of Image Collections via Incremental Learning of a Relocalizer/c513fe61-35e0-435f-acdb-dbf8715686b2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a349e7fc6517432ab0860152f3cbc5ae3aca5395 --- /dev/null +++ b/2024/Scene Coordinate Reconstruction_ Posing of Image Collections via Incremental Learning of a Relocalizer/c513fe61-35e0-435f-acdb-dbf8715686b2_content_list.json @@ -0,0 +1,1716 @@ +[ + { + "type": "text", + "text": "Scene Coordinate Reconstruction: Posing of Image Collections via Incremental Learning of a Relocalizer", + "text_level": 1, + "bbox": [ + 250, + 141, + 754, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Eric Brachmann1, Jamie Wynn1, Shuai Chen2, Tommaso Cavallari1, Åron Monszpart1, Daniyar Turmukhambetov1, and Victor Adrian Prisacariu1,2", + "bbox": [ + 215, + 234, + 785, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Niantic $^{2}$ University of Oxford", + "bbox": [ + 423, + 276, + 580, + 305 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. We address the task of estimating camera parameters from a set of images depicting a scene. Popular feature-based structure-from-motion (SfM) tools solve this task by incremental reconstruction: they repeat triangulation of sparse 3D points and registration of more camera views to the sparse point cloud. We re-interpret incremental structure-from-motion as an iterated application and refinement of a visual relocalizer, that is, of a method that registers new views to the current state of the reconstruction. This perspective allows us to investigate alternative visual relocalizers that are not rooted in local feature matching. We show that scene coordinate regression, a learning-based relocalization approach, allows us to build implicit, neural scene representations from unposed images. Different from other learning-based reconstruction methods, we do not require pose priors nor sequential inputs, and we optimize efficiently over thousands of images. In many cases, our method, ACE0, estimates camera poses with an accuracy close to feature-based SfM, as demonstrated by novel view synthesis.", + "bbox": [ + 259, + 340, + 743, + 561 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Project page: https://nianticlabs.github.io/acezero/", + "bbox": [ + 261, + 563, + 651, + 575 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 601, + 375, + 616 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the beginning there was structure-from-motion.", + "bbox": [ + 321, + 632, + 679, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The genesis of numerous computer vision tasks lies in the estimation of camera poses and scene geometry from a set of images. It is the first fundamental step that lets us leave the image plane and venture into 3D. Since structure-from-motion (SfM) is such a central capability, we have researched it for decades. By now, refined open-source tools, such as COLMAP [80], and efficient commercial packages, such as RealityCapture [68], are available to us. Feature matching-based SfM is the gold standard for estimating poses from images, with a precision that makes its estimates occasionally considered \"ground truth\" [3,5,13,46,69].", + "bbox": [ + 212, + 657, + 787, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The success of Neural Radiance Fields (NeRF) [61] has renewed interest in the question of whether SfM can be solved differently, based on neural, implicit scene representations rather than 3D point clouds. There has been some progress in recent years but, thus far, learning-based approaches to camera pose recovery", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b4aa7e4e1c73fd6aa578d18ede7c9437c374cf0ce131026b0f4af869293101f9.jpg", + "image_caption": [ + "Fig. 1: Reconstructing 10,000 Images. Top: Starting from a single image and the identity pose, we train a learning-based visual relocalizer. The relocalizer allows us to estimate the poses of more views, and the additional views allow us to refine the relocalizer. We show three out of six iterations for this scene (7Scenes Office [81]). All 10k images have been posed in roughly 1 hour on a single GPU. In comparison NoPe-NeRF [10] needs two days to pose 200 images. The point cloud is a visualization of the implicit scene representation of the relocalizer. Camera positions are color coded by relocalization confidence from yellow (low) to green (high). Bottom: Point clouds from Nerfacto [89] trained on top of our poses for a few scenes from our experiments." + ], + "image_footnote": [], + "bbox": [ + 222, + 143, + 781, + 310 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "still have significant limitations. They either require coarse initial poses [45, 54, 98, 101], prior knowledge of the pose distribution [60] or sequential inputs [10, 11, 90]. In terms of the number of images that learning-based approaches can handle, they are either explicitly targeted at few-frame problems [53, 82, 95, 97, 99, 103] or they are computationally so demanding that they can realistically only be applied to a few hundred images at most [10, 54, 98]. We show that none of these limitations are an inherent consequence of using learning-based scene representations.", + "bbox": [ + 212, + 486, + 787, + 607 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our approach is inspired by incremental SfM and its relationship to another computer vision task: visual relocalization. Visual relocalization describes the problem of estimating the camera pose of a query image w.r.t. to an existing scene map. Incremental SfM can be re-interpreted as a loop of 1) do visual relocalization to register new views to the reconstruction, and 2) refine/extend the reconstruction using the newly registered views. Local feature matching is a traditional approach to visual relocalization [70,74-76]. In recent years, multiple learning-based relocalizers have been proposed that encode the scene implicitly in the weights of a neural network [3, 12, 14, 15, 17, 22, 47]. Not all of them are suitable for building a SfM pipeline. We need a relocalizer with high accuracy and good generalization. Training of the relocalizer has to be swift. We need to be able to bootstrap relocalization without ground truth poses. And we need to be able to tell whether registration of a new image was successful.", + "bbox": [ + 212, + 609, + 787, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We show that \"scene coordinate regression\", an approach to visual relocalization proposed a decade ago [81], has the desirable properties and can serve as", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "E. Brachmann et al.", + "bbox": [ + 271, + 114, + 408, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the core of a new approach to learning-based SfM: Scene coordinate reconstruction. Rather than optimizing over image-to-image matches, like feature-based SfM, scene coordinate reconstruction regresses image-to-scene correspondences directly. Rather than representing the scene as a 3D point cloud with high dimensional descriptors, we encode the scene into a lightweight neural network. Our approach works on unsorted images without pose priors and efficiently optimises over thousands of images, see Figure 1.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We summarize our contributions:", + "bbox": [ + 214, + 252, + 472, + 266 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Scene Coordinate Reconstruction, a new approach to SfM based on incremental learning of scene coordinate regression, a visual relocalization principle.", + "- We turn the fast-learning visual relocalizer ACE [12] into a SfM framework that is able to predict the camera poses of a set of unposed RGB images. We refer to this new SfM pipeline as $ACE0$ (ACE Zero).", + "- Compared to ACE [12], we add the capability to train in a self-supervised fashion. We start from a single image, and iterate between learning the map and registering new views. We expedite reconstruction times by early stopping, and increase reconstruction quality by pose refinement." + ], + "bbox": [ + 225, + 275, + 785, + 410 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 431, + 387, + 448 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Reconstruction. SfM pipelines either ingest a collection of unordered images [20,78,84] or an ordered image sequence [6,30,65,66,88] from a video to recover 3D structure of a scene and camera poses (\"motion\") of the images.", + "bbox": [ + 212, + 462, + 787, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SIFT [58] and other robust descriptors allow matching image features across wide baselines enabling systems to reconstruct large-scale scenes using Internet images [80,84,85,100]. Image-to-image matches can also be regressed directly in a detector-free setup [41,87]. Feature tracks across multiple images are built from image-to-image matches. Feature tracks and estimated relative poses are used to solve for the 3D feature coordinates, camera poses and calibrations (intrinsic matrices). This geometric optimization problem is mainly solved using bundle adjustment which was explored in photogrammetry and geodesy [19, 50] and became standard in the computer vision community [40,92]. Bundle adjustment relies on the initialization being close to the solution (i.e., camera poses and 3D points are already mostly accurate).", + "bbox": [ + 212, + 508, + 787, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "There are two main approaches to this problem. Incremental SfM [6, 66] starts the reconstruction from very few images to create a high-quality seed reconstruction that progressively grows by registering more images and refining the reconstruction until convergence. Global SfM approaches solve for \"global\" poses of all images using estimates of relative poses, i.e., motion averaging [37, 38], rotation averaging [39, 59] and pose-graph optimization [21]. Various techniques were proposed to improve SfM runtime for very large sets of images [1, 2, 8, 9, 28, 36, 42, 86, 91]. Our work is similar to Incremental SfM, as we also progressively register images to the reconstructed scene starting from a seed reconstruction. However, we do not explicitly compute image matches, nor feature tracks across images, which can be computationally expensive.", + "bbox": [ + 212, + 674, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Scene Coordinate Reconstruction", + "bbox": [ + 508, + 114, + 730, + 127 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Visual Relocalization. A reconstructed (or mapped) scene is a database of images with known camera poses. This database can be used by a visual relocalizer to estimate poses for new query images to \"relocalize\" a camera in the scene. Feature-based approaches extract 2D local features [31, 33, 56, 58, 71, 72, 87] from a query image and match them to 3D points to solve for the query pose using perspective-n-point (PnP) solvers [35], e.g., [73-75]; or match query features to 2D local features of mapped images to triangulate the query image, e.g., [104, 105]. For large scenes, matching features on a subset of database images relevant to the query can improve the speed and accuracy, e.g., [43, 67, 70, 76].", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Some learning-based approaches encode the map of a scene in the weights of a neural network. PoseNet [47] directly regresses the absolute camera pose given an input image using a CNN that was trained on image-posed pairs of the reconstructed scene. Sattler et al. show that extrapolating outside the mapped poses can be challenging for absolute pose regression approaches [77]. Relative pose regression networks [4, 32, 51, 93] estimate the relative pose for a pair of images, allowing triangulation of the query image from multiple map images, or estimate a scale-metric relative pose w.r.t. a single map image [3]. Scene coordinate regression [12, 14, 15, 17, 22, 23, 52, 81] directly predicts 3D coordinates of a point given a patch. This approach generalizes well as the camera pose is solved using PnP [35] within a RANSAC loop [34]. ACE [12] shows that training of scene coordinate regression can be greatly accelerated. In our work we train the ACE localizer, but we start from images without poses.", + "bbox": [ + 212, + 282, + 789, + 479 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Image-Based Rendering. In recent years, neural radiance fields (NeRFs) [61] have seen a lot of attention from the community. NeRFs allow photorealistic novel view synthesis when trained from image-pose pairs. Typically, estimation of the image poses is done in advance by using an SfM pipeline such as COLMAP, e.g., [5]. Nonetheless, research exists that estimates camera poses with NeRFs, facilitates camera localization for novel views after NeRF training [24-26, 62, 102], or simultaneously estimates camera poses during NeRF training from images alone [10, 27, 54, 82, 98]. However, these approaches either assume that the scene is captured from the front [45, 98, 101], that coarse poses are available for initialization [54], or that images are captured sequentially [10]. Techniques that rely on a multi-layer perceptron (MLP) representation of the scene, e.g. [10, 27], are slow to train, taking days to converge. While radiance fields can be trained faster using multi-layer hash grids [63] or Gaussian splats [48], their efficacy in pose estimation without approximate prior pose initialization [55, 62] or sequential image inputs [48] remains unproven. Concurrently, several learning-based camera pose estimation methods have been proposed [53, 97, 103]. Due to GPU memory constraints, these methods estimate poses for sparse sets of images.", + "bbox": [ + 212, + 494, + 789, + 752 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 776, + 331, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Preliminaries. The input to our system is a set of RGB images, denoted by $\\mathcal{I} = \\{I_i\\}$ , where $i$ refers to the image index. Our system estimates the corresponding", + "bbox": [ + 215, + 809, + 789, + 842 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "E. Brachmann et al.", + "bbox": [ + 271, + 114, + 408, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/326d8101c386fb5c2375761039db1a5410211ba682da210bdc544e843354ef62.jpg", + "image_caption": [ + "Fig. 2: ACE0 Framework. Top left: We loop between learning a reconstruction from the current set of images and poses (\"neural mapping\"), and estimating poses of more images (\"relocalization\"). Top right: During the mapping stage, we train a scene coordinate regression network as our scene representation. Camera poses of the last relocalization round and camera calibration parameters are refined during this process. We visualize scene coordinates by mapping XYZ to the RGB cube. Bottom: In the relocalization stage, we re-estimate poses of images using the scene coordinate regression network, including images that were previously not registered to the reconstruction. If the registration of an image succeeds, it will be used in the next iteration of the mapping stage; otherwise it will not." + ], + "image_footnote": [], + "bbox": [ + 220, + 143, + 777, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "set of camera parameters, both intrinsic and extrinsic: $\\mathcal{H} = \\{(K_i,T_i)\\}$ . Each $T_{i}$ refers to a $3\\times 4$ matrix containing a rotation and translation, while $K_{i}$ refers to a $3\\times 3$ matrix with the calibration parameters. We assume no particular image order or any prior knowledge about the pose distribution.", + "bbox": [ + 212, + 555, + 787, + 614 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We also want to recover the 3D structure of the scene: Each pixel $j$ in image $i$ with 2D pixel position $\\mathbf{p}_{ij}$ has a corresponding coordinate in 3D, denoted as $\\mathbf{y}_{ij}$ . The 2D pixel positions and 3D scene coordinates are related by the camera pose and the projection function $\\pi$ :", + "bbox": [ + 212, + 614, + 787, + 676 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {p} _ {i j} = \\boldsymbol {\\pi} \\left(K _ {i}, T _ {i}, \\mathbf {y} _ {i j}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 685, + 785, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $T_{i}$ maps camera coordinates to scene coordinates, and $K_{i}$ projects camera coordinates to the image plane.", + "bbox": [ + 212, + 709, + 787, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As our scene representation, we utilize a scene coordinate regression model [81], i.e., a learnable function $f_{\\mathrm{SCR}}$ that maps an image patch of image $I_{i}$ , centered around pixel position $\\mathbf{p}_{ij}$ to a scene coordinate: $\\mathbf{y}_{ij} = f_{\\mathrm{SCR}}(\\mathbf{p}_{ij},I_i)$ .", + "bbox": [ + 212, + 739, + 790, + 785 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given a set of 2D-3D correspondences predicted by $f_{\\mathrm{SCR}}$ for any image $I_i$ , we can recover this image's camera pose $T_i$ using a pose solver $g$ :", + "bbox": [ + 212, + 785, + 787, + 816 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nT _ {i} = g \\left(K _ {i}, \\left\\{\\left(\\mathbf {p} _ {i j}, \\mathbf {y} _ {i j}\\right) \\right\\}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 408, + 823, + 785, + 842 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Scene Coordinate Reconstruction", + "bbox": [ + 508, + 114, + 730, + 127 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since 2D-3D correspondences can be inaccurate, and contain incorrect predictions, $g$ combines a PnP solver [35] with a RANSAC loop [34].", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Normally, scene coordinate regression models are trained in a supervised fashion for the task of visual relocalization [12, 14, 15, 17, 22, 23, 52, 81]. That is, $f_{\\mathrm{SCR}}$ is trained using images with known ground truth camera parameters $\\{(I_i, T_i^{\\mathrm{GT}}, K_i^{\\mathrm{GT}})\\}$ , and used for estimating the poses of unseen query images. Instead, we show how these models can be trained self-supervised, without ground truth poses, to estimate the poses of the mapping images themselves. Thus, we turn scene coordinate regression into scene coordinate reconstruction, a learning-based SfM tool.", + "bbox": [ + 212, + 176, + 787, + 297 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.1 Neural Mapping", + "text_level": 1, + "bbox": [ + 215, + 320, + 398, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We train the scene coordinate regression model iteratively where we denote the current time step as $t$ and the corresponding scene model as $f_{\\mathrm{SCR}}^t$ . We iterate between training the scene model, and registering new views, see Figure 2.", + "bbox": [ + 212, + 347, + 782, + 391 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "At iteration $t$ we assume that a subset of images has already been registered to the scene, $\\mathcal{I}_{\\mathrm{Reg}}^t \\subset \\mathcal{I}$ , and where corresponding camera parameters, $T_{i}^{t}$ and $K_{i}^{t}$ , have already been estimated. Using these as pseudo ground truth, we train the scene model by minimizing the pixel-wise reprojection error:", + "bbox": [ + 214, + 392, + 785, + 455 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\sum_ {I _ {i} \\in \\mathcal {I} _ {\\text {R e g}} ^ {t}} \\sum_ {j \\in I _ {i}} \\| \\mathbf {p} _ {i j} - \\boldsymbol {\\pi} \\left(K _ {i} ^ {t}, T _ {i} ^ {t}, \\mathbf {y} _ {i j} ^ {t}\\right) \\|, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 375, + 465, + 785, + 501 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where the scene model $f_{\\mathrm{SCR}}^{t}$ predicts coordinates $\\mathbf{y}_{ij}^{t}$ .", + "bbox": [ + 212, + 513, + 599, + 530 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Mapping Framework. We optimize Eq. 3 using stochastic gradient descent, using the fast-learning scene coordinate regressor ACE [12] (Accelerated Coordinate Encoding). ACE is trained in minutes, even for thousands of views. The training speed is important since we have to train the model in multiple iterations. ACE approximates Eq. 3 by sampling random patches from the training set, $\\mathcal{I}_{\\mathrm{Reg}}^t$ . To do that efficiently, ACE employs a pre-trained encoder to precompute high dimensional features for a large number of training patches. The encoder stays frozen during mapping. The actual scene model, which is trained, is a multi-layer perceptron (MLP) that maps encoder features to scene coordinates; see Fig. 2 (top right) for a visual representation.", + "bbox": [ + 212, + 530, + 787, + 679 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pose Refinement. Differently from the ACE [12] protocol, the ground truth poses $T_{i}^{\\mathrm{GT}}$ are unknown during training. Instead, we have $T_{i}^{t}$ , estimates based on earlier iterations of the reconstruction. Since these estimates can be inaccurate, we add the ability to refine poses during mapping. We implement refinement using an MLP:", + "bbox": [ + 212, + 680, + 787, + 755 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nT _ {i} ^ {t} = f _ {\\text {P o s e}} ^ {t} \\left(\\hat {T} _ {i} ^ {t}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 444, + 755, + 785, + 772 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\tilde{T}_i^t$ denotes the initial pose estimate at the start of a mapping iteration. Inspired by [106], the refinement MLP ingests $\\tilde{T}_i^t$ as $3\\times 4 = 12$ values and predicts 12 additive offsets. We orthonormalize rotations using Gram-Schmidt [18]. We jointly optimize $f_{\\mathrm{Pose}}^t$ and $f_{\\mathrm{SCR}}^t$ to minimize the reprojection error of Eq. 3. We", + "bbox": [ + 212, + 777, + 787, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "E. Brachmann et al.", + "bbox": [ + 271, + 114, + 408, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e97ec08733e3676a4e5598bec4880fd665bb92cd0d8faad018b525ec29096720.jpg", + "image_caption": [ + "Fig. 3: Left: Pose Refinement. Since we register images based on a coarse and incomplete state of the reconstruction, we add the ability to refine poses during neural mapping. An MLP predicts pose updates relative to the initial poses, supervised by the reprojection error of scene coordinates. Right: Initialization. To start the reconstruction, we train the network using one image, the identity pose and a depth estimate, here ZoeDepth [7]. In this example, we register 33 views to the initial reconstruction. Depth estimates are only used for this step." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 500, + 292 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8c96ff0ceff1a5888f81096c6aa623935e3ebd2db8b011fa5891e3ca5993a194.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 143, + 784, + 292 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "show the impact of pose refinement for one iteration in Fig. 3 (left). We discard the refinement MLP after each mapping iteration. Its purpose is to enable the scene model, $f_{\\mathrm{SCR}}^{t}$ , to converge to a consistent scene representation.", + "bbox": [ + 212, + 439, + 784, + 484 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "With neither poses nor 2D-3D correspondences fixed in Eq. 3, the scene coordinate regressor could drift or degenerate. As regularization, we optimize the pose refiner $f_{\\mathrm{Pose}}^{t}$ using AdamW [57] with weight decay, biasing the MLP to predict small updates relative to the initial estimates $\\tilde{T}_i^t$ . This relies on the assumption that mapping images in $\\mathcal{I}_{\\mathrm{Reg}}^t$ have been registered close to their true position. If that assumption holds, the smoothness prior of the networks [94] encourages a multi-view consistent solution as shown in previous work [12,15,17].", + "bbox": [ + 212, + 487, + 787, + 593 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As an alternative to an MLP refiner, we could back-propagate directly to the input poses [98, 102]. However, when optimizing over thousands of views, the signal for each single pose becomes sparse. Cameras are correlated via the scene representation. If the optimization removes drift in the scene, multiple cameras need to move. The MLP refiner models the correlation of cameras.", + "bbox": [ + 212, + 595, + 787, + 671 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Calibration Refinement. We do not assume information about precise calibration parameters, although often reported by devices. We do assume that the principal point is in the center, that pixels are unskewed and square. We do not model image distortion. While these are reasonable assumptions for many data regimes, we cannot rely on the focal length to be given. Thus, we refine the focal length starting from a heuristic: $K_{i}^{t} = f_{\\mathrm{Calib}}^{t}(K_{i}^{\\mathrm{init}})$ . The refinement function $f_{\\mathrm{Calib}}^{t}$ entails a single learnable parameter $\\alpha^{t}$ such that the focal length $f_{i}^{t} = f^{\\mathrm{init}} \\cdot (1 + \\alpha^{t})$ . As before, superscript $t$ denotes the time step. We optimize $\\alpha^{t}$ using AdamW [57] with weight decay, biasing it towards a small relative scale factor. Estimates of $\\alpha$ are carried over across iterations. We set $f^{\\mathrm{init}}$ to 70% of the image diagonal and it is shared by all cameras in our experiments.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Scene Coordinate Reconstruction", + "bbox": [ + 508, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Relocalization", + "text_level": 1, + "bbox": [ + 215, + 146, + 380, + 160 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Given the scene model of iteration $t$ , we attempt to register more images to determine the training set for the next mapping iteration, $\\mathcal{I}_{\\mathrm{Reg}}^{t + 1}$ . We pass all images in $\\mathcal{I}$ to the scene coordinate regressor $f_{\\mathrm{SCR}}^t$ to gather 2D-3D correspondences, and solve for their poses using RANSAC and PnP:", + "bbox": [ + 212, + 169, + 787, + 231 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {T} _ {i} ^ {t + 1}, s _ {i} ^ {t + 1} = g \\left(K _ {i} ^ {t}, \\left\\{\\left(\\mathbf {p} _ {i j}, \\mathbf {y} _ {i j} ^ {t}\\right) \\right\\}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 239, + 785, + 258 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Here, we assume that the pose solver returns a confidence score $s_i^{t+1}$ alongside the pose itself that lets us decide whether the pose of the image has been estimated successfully. We simply utilize the inlier count as score $s_i^{t+1}$ and apply a threshold to form the training set of the next mapping iteration: $\\mathcal{I}_{\\mathrm{Reg}}^{t+1} = \\left\\{ I_i | s_i^{t+1} > \\tau_s \\right\\}$ . The relocalization process is depicted in Figure 2, bottom.", + "bbox": [ + 214, + 267, + 787, + 344 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 Initialization", + "text_level": 1, + "bbox": [ + 215, + 364, + 369, + 378 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We start the reconstruction with one image: $\\mathcal{I}_{\\mathrm{Reg}}^0 = \\{I_{\\mathrm{seed}}\\}$ . We set the seed pose $T_{\\mathrm{seed}}^0$ to identity, and we initialize the calibration $K_{\\mathrm{seed}}^{\\mathrm{init}}$ as explained above.", + "bbox": [ + 212, + 386, + 785, + 420 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We cannot train a scene coordinate regression network using Eq. 3 with a single image. The reprojection error is ill-conditioned without multiple views constraining the depth. Therefore, we optimize a different objective in the seed iteration, inspired by Map-free Relocalization [3]. Arnold et al. argue that a single image and a depth estimate allow to relocalize query images, albeit with limited accuracy. Our experiments show that such coarse relocalizations of a few images suffice as initialization for optimizing the reprojection error of Eq. 3.", + "bbox": [ + 212, + 420, + 785, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Let $d_{ij}$ be a depth value predicted for pixel $j$ of image $i$ . We derive a target scene coordinate by back-projection as $\\hat{\\mathbf{y}}_{ij} = d_{ij}(K_{\\mathrm{seed}}^{\\mathrm{init}})^{-1}\\mathbf{p}_{ij}$ . We train an initial scene coordinate regression network $f_{\\mathrm{SCR}}^{0}$ by optimizing $\\sum_{j\\in I^0}\\| \\hat{\\mathbf{y}}_{ij} - \\mathbf{y}_{ij}\\|$ . Fig. 3 (right) shows a scene coordinate point cloud learned from a depth estimate, and successful relocalization against it.", + "bbox": [ + 212, + 525, + 785, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We found our pipeline to be robust w.r.t. selecting the seed image. However, when a randomly selected image has little to no visual overlap with the remaining images, the whole reconstruction would fail. Selecting an unfortunate seed image, e.g., at the very end of a long camera trajectory, can increase reconstruction times. To decrease the probability of such incidents, we try 5 random seed images and choose the one with the highest relocalization rate across 1000 other mapping images. Since mapping seed images is fast, ca. 1 min on average, this poses no significant computational burden.", + "bbox": [ + 212, + 599, + 785, + 720 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Implementation", + "text_level": 1, + "bbox": [ + 215, + 741, + 393, + 756 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We base our pipeline on the public code of the ACE relocalizer [12]. ACE uses a convolutional feature backbone, pre-trained on ScanNet [29], that ingests images scaled to 480px height. On top of the backbone is the mapping network, a 9-layer MLP with 512 channels, consuming 4MB of weights in 16-bit floating point precision. This is our scene representation.", + "bbox": [ + 212, + 763, + 785, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "E. Brachmann et al.", + "bbox": [ + 271, + 114, + 408, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The ACE training process is reasonably fast, taking 5 minutes to train the scene network. Since we repeat training the scene network in multiple iterations, we expedite the process further to decrease our total reconstruction time.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Adaptive Sampling of Training Patches. ACE trains the scene representation based on 8M patches sampled from the mapping images, a process that takes 1 minute. This is excessive when having very few mapping images in the beginning of the reconstruction, thus we loop over the mapping images at most 10 times when sampling patches, or until 8M patches have been sampled. Using this adaptive strategy, sampling patches for the seed reconstruction where we have 1 image only takes 2 seconds instead of 1 minute.", + "bbox": [ + 212, + 191, + 785, + 297 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Adaptive Stopping. ACE trains the scene model using a fixed one-cycle learning rate schedule [83] to a total of 25k parameter updates. We make the training schedule adaptive since the network is likely to converge fast when trained on few images only. We monitor the reprojection error of scene coordinates within a mini-batch. If for 100 consecutive batches $70\\%$ of reprojection errors are below an inlier threshold of 10px, we stop training early. We approximate the one-cycle learning rate schedule in a linear fashion: we increase the learning rate in the first 1k iteration from $5 \\times 10^{-4}$ to $3 \\times 10^{-3}$ and when the early stopping criterion has been met, we decrease the learning rate to $5 \\times 10^{-4}$ within 5k iterations.", + "bbox": [ + 212, + 299, + 785, + 435 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We report more implementation details and hyper-parameters in the supplement. Our code is also publicly available to ensure reproducibility.", + "bbox": [ + 214, + 436, + 785, + 468 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 497, + 375, + 513 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We refer to our SfM pipeline as ACE0 (ACE Zero) since it builds on top of the ACE relocalizer [12] but adds the ability to train from scratch, without poses. We demonstrate the effectiveness of our approach on three datasets and 31 scenes in total. For all experiments, we rely on ZoeDepth [7] to initialize our reconstructions. All timings reported for ACE0 are based on a single V100 GPU.", + "bbox": [ + 212, + 534, + 787, + 611 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Baselines. We consider the pose estimates of COLMAP with default parameters as our pseudo ground truth, obtained by extensive processing. We also run COLMAP with parameter settings recommended for large image collections of 1k images and more [79]. This variation achieves much faster processing (denoted COLMAP fast). We use a V100 GPU for COLMAP feature extraction and matching, and we specify the exact parameters of COLMAP in the supplement. Furthermore, we show some results of RealityCapture [68], an efficient commercial feature-based SfM pipeline.", + "bbox": [ + 212, + 611, + 785, + 733 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We compare to learning-based SfM approaches that are not restricted to few-frame scenarios, namely to NeRF-based BARF [54] and NoPe-NeRF [10]. We also compare to DUSt3R [97], a non-NeRF learning-based SfM method. While we focus on reconstructing unsorted image collections, the datasets we consider allow for sequential processing. Hence, for context, we also show results of DROID-SLAM [90], a neural SLAM approach. Unless specified otherwise, we report timings based on a single V100 GPU.", + "bbox": [ + 212, + 734, + 785, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Scene Coordinate Reconstruction", + "bbox": [ + 508, + 114, + 730, + 127 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/7c1dc002114a01dbd79150f763da2146f7286fb718e64ef0d1745a6b71872aab.jpg", + "image_caption": [ + "Fig. 4: Reconstructed Poses. We show poses estimated by ACE0 for a selection of scenes. We color code the reconstruction iteration in which a particular view has been registered. We show the ACE0 point cloud as a representation of the scene. The seed image is shown as a gray frustum. We also compare our poses to poses estimated by COLMAP (Mip-NeRF 360, Tanks and Temples) and KinectFusion (7-Scenes)." + ], + "image_footnote": [], + "bbox": [ + 218, + 147, + 787, + 380 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Benchmark. We show results on 7-Scenes [81], a relocalization dataset, on Mip-NeRF 360 [5], a view synthesis dataset and on Tanks and Temples [49], a reconstruction dataset. Comparing poses on these datasets is problematic, as our pseudo ground truth is estimated rather than measured. For example, an approach might be more accurate than COLMAP on individual scenes. Computing pose errors w.r.t. COLMAP would result in incorrect conclusions. Therefore, we gauge the pose quality in a self-supervised way, using novel view synthesis [96].", + "bbox": [ + 212, + 492, + 787, + 598 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We let each method estimate the poses of all images of a scene. For evaluation, we split the images into training and test sets. We train a Nerfacto [89] model on the training set, and synthesize views for the test poses. We compare the synthesized images to the test images and report the difference as peak signal-to-noise ratio (PSNR). To ensure a fair comparison to NeRF-based competitors, we use these methods in the same way as the other SfM methods: we run them to pose all images, and train a Nerfacto model on top of their poses. This is to ensure that we solely compare the pose quality across methods, and not different capabilities in view synthesis. The quality of poses affects the PSNR numbers in two ways: Good training poses let the NeRF model fit a consistent scene representation. Good testing poses make sure that the synthesized images are aligned with the original image. We explain further details in the supplement which also includes additional, perceptual metrics. In spirit, our evaluation is similar to the Tanks and Temples benchmark which evaluates a derived scene mesh rather than camera poses. However, our evaluation can be applied to arbitrary datasets as it does not need ground truth geometry.", + "bbox": [ + 212, + 599, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "E. Brachmann et al.", + "bbox": [ + 271, + 114, + 408, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/bb9a0f72490a7c8e4033e223ec44867511e194457d6e99ee8fd67bb9146b8ed6.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1: 7-Scenes. We show the pose accuracy via view synthesis with Nerfacto [89] as PSNR in dB, and the reconstruction time. Results for All Frames are color coded w.r.t. similarity to the COLMAP pseudo ground truth: $>0.5$ dB better within $\\pm 0.5$ dB $>0.5$ dB worse $>1$ dB worse. For some competitors, we had to sub-sample the images due to their computational complexity (right side). Method needs sequential inputs. *Results on more powerful hardware." + ], + "table_body": "
FramesPseudo Ground TruthAll Frames200 Frames50 Frames
Kinect FusionCOLMAP (default)COLMAP (fast)DROID-SLAM† [90]ACE0 (ours)KF+ACE0 (ours)BARF [54]NoPE-NeRF† [10]ACE0 (ours)DUS3R [97]ACE0 (Ours)
Chess6k19.623.623.519.323.323.012.812.622.718.919.2
Fire4k19.222.622.613.022.322.312.711.822.118.819.5
Heads2k17.018.818.917.618.819.110.711.819.918.421.3
Office10k18.921.421.6failed21.121.511.910.919.812.513.7
Pumpkin6k19.924.123.818.324.123.819.614.224.721.722.3
RedKitchen12k17.621.421.410.920.820.911.611.218.913.813.7
Stairs3k19.016.721.013.017.719.915.815.918.815.315.4
Average18.721.221.8N/A21.221.513.612.621.017.117.9
Avg. Timerealtime38h13h18min1h7min8.5h47h27min4min*16min
", + "bbox": [ + 215, + 151, + 782, + 268 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.1 7-Scenes", + "text_level": 1, + "bbox": [ + 215, + 393, + 331, + 407 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The 7-Scenes dataset [81] consists of seven indoor scenes, scanned with a Kinect v1 camera. Multiple, disconnected scans are provided for each scene to a total of $2\\mathrm{k} - 12\\mathrm{k}$ images. For each method, we assume a shared focal length across scans and initialize with the default calibration of a Kinect v1. The dataset comes with pseudo ground truth camera poses estimated by KinectFusion [44, 64], a depth-based SLAM system. Individual scans were registered but not bundle-adjusted [13]. Inspired by [13], we recompute alternative, bundle-adjusted pseudo ground truth by running COLMAP with default parameters.", + "bbox": [ + 212, + 416, + 785, + 537 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Discussion. We show results in Table 1. Of both pseudo ground truth versions, KinectFusion achieves lower PSNR numbers than COLMAP, presumably due to the lack of global optimization. COLMAP with fast parameters shows PSNR numbers similar to COLMAP with default parameters, on average. Both versions of running COLMAP take considerable time to reconstruct each scene. We note that COLMAP has been optimised for quality, rather than speed. Not all acceleration strategies from the feature-based SfM literature have been implemented in COLMAP, so presumably comparable quality can be obtained faster. DROID-SLAM [90] does not perform well on 7-Scenes and partially fails altogether, presumably due to the jumps between individual scans of each scene.", + "bbox": [ + 212, + 537, + 785, + 688 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Our approach, ACE0, achieves a pose quality comparable to the COLMAP pseudo ground truth while reconstructing each scene in $\\sim 1$ hour despite the large number of images. We show qualitative examples in Figure 4 and in the supplement. We also demonstrate that ACE0 can swiftly optimize an initial set of approximate poses. When starting from KinectFusion poses, ACE0 increases PSNR significantly in less than 10 minutes per scene, see “ $\\mathrm{KF} + \\mathrm{ACE0}$ ” in Table 1. In the supplement, we include a parameter study on 7-Scenes to show that ACE0 is robust to the choice of depth estimator. We also show the positive impact of pose refinement on the reconstruction quality as well as the reconstruction speedup due to our early stopping schedule.", + "bbox": [ + 212, + 688, + 785, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Scene Coordinate Reconstruction", + "bbox": [ + 508, + 114, + 730, + 127 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/73c890157e77ec40554c46773547afaf2ca109859695b533d7aa2987b73a1723.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ACE [12]ACE [12]ACE0 (ours)
SupervisionKinectFusionCOLMAPself-supervised
Chess96.0%100.0%100.0%
Fire98.4%99.5%98.8%
Heads100.0%100.0%100.0%
Office36.9%100.0%99.1%
Pumpkin47.3%100.0%99.9%
Redlitchen47.8%98.9%98.1%
Stairs74.1%85.0%61.0%
Average71.5%97.6%93.8%
", + "bbox": [ + 217, + 143, + 454, + 241 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/283dcc312fb1a9015b1616daadce9fafc63f75e32aa9535cb43dcdc82d563f12.jpg", + "table_caption": [ + "Table 2 (a): Relocalization on 7-Scenes. % poses below $5\\mathrm{cm},5^{\\circ}$ error, computed w.r.t. COLMAP pseudo GT." + ], + "table_footnote": [], + "table_body": "
Pseudo GT (COLMAP)DROID-SLAM† [90]BARF [54]NoPe-NeRF† [10]ACEO (ours)
Bicycle21.510.911.912.218.7
Bonsai27.610.912.514.825.8
Counter25.512.911.911.624.5
Garden26.316.713.313.825.0
Kitchen27.413.913.314.426.1
Room28.011.311.914.319.8
Stump16.813.915.013.720.5
Average24.712.912.813.522.9
", + "bbox": [ + 459, + 143, + 787, + 241 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 2 (b): Mip-NeRF 360. Pose quality in PSNR, higher is better. Best in bold. Method needs sequential inputs.", + "bbox": [ + 464, + 242, + 781, + 275 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "For our learning-based competitors, we sub-sampled images due to their computational constraints, see right side of Table 1. Even using only 200 images, NoPe-NeRF [10] takes 2 days to fit a model and estimate poses. Despite these long processing times, we observe poor pose quality of BARF and NoPe-NeRF in terms of PSNR. BARF [54] requires pose initialisation. We provide identity poses since the scenes of 7Scenes are roughly forward-facing. Still, the camera motions are too complex for BARF to handle. NoPe-NeRF does not require pose initialisation but needs roughly sequential images, which we did provide. NoPe-NeRF relies upon successive images having similar poses to perform well, and struggles with the large jumps between the subsampled images.", + "bbox": [ + 212, + 314, + 784, + 465 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "For the comparison with DUSt3R [97], we had to subsample the sequences further, as we were only able to run it with 50 images at most, even when using an A100 GPU with 40GB memory. DUSt3R achieves reasonable PSNR numbers but consistently lower than ACE0.", + "bbox": [ + 212, + 465, + 785, + 525 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Relocalization. ACE0 is a learning-based SfM tool but it is also a self-supervised visual relocaliser. In Table 2 (a), we compare it to the supervised relocalizer ACE [12]. Using the scale-metric pseudo ground truth of [13], we train ACE with COLMAP mapping poses, and evaluate it against COLMAP query poses. Unsurprisingly, ACE achieves almost perfect relocalization under the usual $5\\mathrm{cm}$ , $5^{\\circ}$ error threshold. Interestingly, ACE0 achieves almost identical results when mapping the scene self-supervised, and evaluating the relocalized query poses against the COLMAP pseudo ground truth. For context, when training ACE with Kinect-Fusion mapping poses and evaluating against COLMAP pseudo ground truth, results are far worse. This signifies that ACE0 mapping poses are very similar to the COLMAP mapping poses, and less similar to KinectFusion mapping poses. We give more details about this experiment in the supplement.", + "bbox": [ + 212, + 526, + 787, + 708 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.2 Mip-NeRF 360", + "text_level": 1, + "bbox": [ + 215, + 726, + 387, + 742 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The Mip-NeRF 360 dataset [5] consists of seven small-scale scenes, both indoor and outdoor. The dataset was reconstructed with COLMAP and comes with intrinsics (which we ignore), camera poses and undistorted (pinhole camera) images. For each method, we assume a shared focal length per scene.", + "bbox": [ + 212, + 748, + 784, + 809 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Discussion. We present PSNR results in Table 2 (b). NoPe-NeRF does not perform well on this dataset, despite processing each scene for 2 days. The differences", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "E. Brachmann et al.", + "bbox": [ + 271, + 114, + 408, + 127 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/ef9ff5cc88193a46650a64c193de4569f68bdb0d49606daaaa34589846ed37a8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
FramesDROID-FramesDROID-SparseSparse
COLMAP (default)Reality CaptureSLAM† [90]ACE0 (ours)COLMAP (fast)Reality CaptureSLAM† [90]ACE0 (ours)COLMAP + Reloc + BACOLMAP + ACE0 (ours)
TrainingBarn41024.021.219.016.512.2k24.416.913.517.726.325.1
Caterpillar38317.115.916.616.911.4k18.617.918.918.618.718.8
Church50718.317.614.317.219.3k12.1-11.516.518.517.3
Ignatius26420.117.717.819.87.8k20.818.619.120.720.920.7
Meetingroom37118.618.115.618.011.1k19.418.217.116.620.820.3
Truck25121.119.018.320.17.5k23.619.120.623.023.423.1
Average36419.918.216.918.114.6k19.818.216.818.921.420.9
Avg Time1h3min5min1.1h74h14h18min2.2h8h1.8h
IntermediateFamily15219.518.817.619.04.4k21.219.819.818.021.321.3
Francis30221.620.720.720.17.8k19.920.421.821.722.522.7
Horse15119.219.016.319.56.0k21.620.719.221.722.622.3
Lighthouse30916.616.513.617.58.3k19.016.618.918.619.520.5
Playground30719.119.211.418.77.7k17.916.511.320.421.221.0
Train30116.815.413.816.212.6k19.614.415.618.519.818.5
Average25418.818.315.618.57.8k19.918.117.819.821.121.0
Avg Time32min2min3min1.3h48h11h14min2.2h5h1h
AdvancedAuditorium30219.612.216.718.713.6k13.7-16.620.021.419.8
Ballroom32416.318.313.117.910.8k17.2-10.418.918.015.6
Courtroom30118.217.212.317.112.6k14.6-10.216.318.717.8
Palace50914.211.710.810.721.9k13.8-8.611.015.312.3
Temple30218.115.711.89.717.5k13.3-11.914.819.616.1
Average34817.315.012.914.815.6k14.5-11.516.218.616.3
Avg Time1h2min4min1h71h27min2.8h10h2.1h
", + "bbox": [ + 215, + 152, + 782, + 438 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 3: Tanks and Temples. We show the pose accuracy via view synthesis with Nerfacto [89] as PSNR in dB, and the reconstruction time. We color code results compared to COLMAP, default and fast, respectively: $>0.5$ dB better within $\\pm 0.5$ dB $>0.5$ dB worse $>1$ dB worse. Method needs sequential inputs.", + "bbox": [ + 215, + 439, + 787, + 508 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "between sequential images are too large. DROID-SLAM fails for the same reason. BARF performs poorly because the identity pose is not a good initialisation for most scenes. In contrast, ACE0 reconstructs the dataset successfully. While it achieves slightly lower PSNR than COLMAP, its pose estimates are similar, $cf$ . Figure 4. The supp. shows that synthesized images based on ACE0 are visually close to those of COLMAP while our learning-based competitors are far off.", + "bbox": [ + 212, + 550, + 787, + 643 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.3 Tanks and Temples", + "text_level": 1, + "bbox": [ + 215, + 664, + 423, + 680 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The Tanks and Temples dataset [49] contains 21 diverse scenes, both indoors and outdoors, and with varying spatial extent. We remove two scenes (Panther, M60) with military associations. We also remove two scenes (Courthouse, Museum) where the COLMAP baseline did not finish the reconstruction after 5 days of processing or ran out of memory. For the latter two scenes, we provide ACE0 results in the supplement. The dataset provides 150-500 images per scene but also the original videos as a source for more frames. Thus, we consider each scene in two versions: Using 150-500 images and using $4\\mathrm{k} - 22\\mathrm{k}$ frames. For all methods, we assume a pinhole camera model with shared focal length across images, and images to be unordered. We found none of the learning-based SfM competitors", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Scene Coordinate Reconstruction", + "bbox": [ + 508, + 114, + 730, + 127 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "to be applicable to this dataset. NoPe-NeRF would run multiple days per scene, even when considering only a few hundred images. DUSt3R would run out of memory. BARF needs reasonable pose initializations which are not available.", + "bbox": [ + 212, + 146, + 785, + 191 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Discussion. We show PSNR numbers of RealityCapture, DROID-SLAM and ACE0 in Table 3, color-coded by similarity to the COLMAP pseudo GT. ACE0 achieves reasonable results when reconstructing scenes from a few hundred images (Table 3, left). ACE0 generally predicts plausible poses (cf., Figure 4), even if PSNR numbers are sometimes lower than those of COLMAP. RealityCapture performs similar to ACE0 while DROID-SLAM struggles on the sparse images.", + "bbox": [ + 212, + 191, + 785, + 282 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Next, we consider more than 1k images per scene (right side of Table 3). Here, we run COLMAP with parameters tuned for large images collections (fast) due to the extremely large image sets. ACE0 offers a reconstruction quality comparable to COLMAP on average while also being fast. We run RealityCapture on some of the scenes but it produces fractured reconstructions for these large images sets, leading to low PSNR numbers. DROID-SLAM still struggles on many scenes despite having access to sequential images that are temporally close.", + "bbox": [ + 212, + 284, + 785, + 388 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In the two rightmost columns, we initialise with a sparse COLMAP reconstruction from 150-500 images, and extend and refine it using all available frames. Firstly, using a feature-based baseline, we register the full set of frames using the relocalization mode of COLMAP, followed by a final round of bundle adjustment. Secondly, we run ACE0 initialised with the poses of the sparse COLMAP reconstruction. Both variants are considerably faster than running COLMAP from scratch on the full set of frames. ACE0 is able to register and refine all additional frames in 1-2 hours, on average. Again, we find the pose quality of ACE0 comparable to the feature-based alternative.", + "bbox": [ + 212, + 388, + 785, + 523 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion and Future Work", + "text_level": 1, + "bbox": [ + 214, + 549, + 532, + 566 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We have presented scene coordinate reconstruction, a new approach to learning-based SfM. We learn an implicit, neural scene representation from a set of unposed images. Our method, ACE0, is able to reconstruct a wide variety of scenes. In many cases, the accuracy of estimated poses is close to that of COLMAP and synthesized images visually similar. Unlike previous learning-based SfM methods, ACE0 can be applied to multiple thousand unsorted images, without pose priors, and reconstructs them within a few hours.", + "bbox": [ + 212, + 582, + 785, + 686 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations. We show some failure cases in the supplement. Scene coordinate regression struggles with repetitive structures since the network is not able to make multi-modal predictions for visually ambiguous inputs. Scene coordinate regression also struggles with representing large areas. The common solution is to use network ensembles based on pre-clustering of the scene [12, 16] which is difficult in a reconstruction setting. While scene coordinate regression generalises quite well, it has difficulties to bridge extreme view point or lighting changes, such as day versus night. In our experiments, we assumed a simple pinhole camera model with shared intrinsics across images. To the best of our knowledge, scene coordinate regression has not been coupled with image distortion, thus far.", + "bbox": [ + 212, + 688, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "E. Brachmann et al.", + "bbox": [ + 271, + 114, + 408, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Agarwal, S., Furukawa, Y., Snavely, N., Simon, I., Curless, B., Seitz, S.M., Szeliski, R.: Building Rome in a day. ACM TOG (2011)", + "2. Agarwal, S., Snavely, N., Seitz, S.M., Szeliski, R.: Bundle adjustment in the large. In: ECCV (2010)", + "3. Arnold, E., Wynn, J., Vicente, S., Garcia-Hernando, G., Monszpart, A., Prisacariu, V.A., Turmukhambetov, D., Brachmann, E.: Map-free visual relocalization: Metric pose relative to a single image. In: ECCV (2022)", + "4. Balntas, V., Li, S., Prisacariu, V.A.: RelocNet: Continuous metric learning relocalisation using neural nets. In: ECCV (2018)", + "5. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-NeRF 360: Unbounded anti-aliased neural radiance fields. In: CVPR (2022)", + "6. Beardsley, P.A., Zisserman, A., Murray, D.W.: Sequential updating of projective and affine structure from motion. IJCV (1997)", + "7. Bhat, S.F., Birkl, R., Wofk, D., Wonka, P., Müller, M.: ZoeDepth: Zero-shot transfer by combining relative and metric depth. arXiv (2023)", + "8. Bhowmick, B., Patra, S., Chatterjee, A., Govindu, V.M., Banerjee, S.: Divide and conquer: Efficient large-scale structure from motion using graph partitioning. In: ACCV (2015)", + "9. Bhowmick, B., Patra, S., Chatterjee, A., Govindu, V.M., Banerjee, S.: Divide and conquer: A hierarchical approach to large-scale structure-from-motion. CVIU (2017)", + "10. Bian, W., Wang, Z., Li, K., Bian, J.W., Prisacariu, V.A.: NoPe-NeRF: Optimising neural radiance field with no pose prior. In: CVPR (2023)", + "11. Bloesch, M., Czarnowski, J., Clark, R., Leutenegger, S., Davison, A.J.: CodeSLAM — Learning a compact, optimisable representation for dense visual SLAM. In: CVPR (2018)", + "12. Brachmann, E., Cavallari, T., Prisacariu, V.A.: Accelerated coordinate encoding: Learning to relocalize in minutes using RGB and poses. In: CVPR (2023)", + "13. Brachmann, E., Humenberger, M., Rother, C., Sattler, T.: On the limits of pseudo ground truth in visual camera re-localisation. In: ICCV (2021)", + "14. Brachmann, E., Krull, A., Nowozin, S., Shotton, J., Michel, F., Gumhold, S., Rother, C.: DSAC-differentiable RANSAC for camera localization. In: CVPR (2017)", + "15. Brachmann, E., Rother, C.: Learning less is more-6D camera localization via 3D surface regression. In: CVPR (2018)", + "16. Brachmann, E., Rother, C.: Expert sample consensus applied to camera relocalization. In: ICCV (2019)", + "17. Brachmann, E., Rother, C.: Visual camera re-localization from RGB and RGB-D images using DSAC. IEEE TPAMI (2021)", + "18. Brégier, R.: Deep regression on manifolds: a 3D rotation case study. In: 3DV (2021)", + "19. Brown, D.: The bundle adjustment-progress and prospect. In: Congr. of the Int. Soc. for Photogr. (1976)", + "20. Brown, M., Lowe, D.G.: Unsupervised 3D object recognition and reconstruction in unordered datasets. In: 3DIM (2005)", + "21. Carlone, L., Tron, R., Daniilidis, K., Dellaert, F.: Initialization techniques for 3D SLAM: A survey on rotation estimation and its use in pose graph optimization. In: ICRA (2015)" + ], + "bbox": [ + 225, + 176, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Scene Coordinate Reconstruction", + "bbox": [ + 508, + 114, + 730, + 127 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "22. Cavallari, T., Bertinetto, L., Mukhoti, J., Torr, P.H., Golodetz, S.: Let's take this online: Adapting scene coordinate regression network predictions for online RGB-D camera relocalisation. In: 3DV (2019)", + "23. Cavallari, T., Golodetz, S., Lord, N.A., Valentin, J., Di Stefano, L., Torr, P.H.: On-the-fly adaptation of regression forests for online camera relocalisation. In: CVPR (2017)", + "24. Chen, S., Bhalgat, Y., Li, X., Bian, J., Li, K., Wang, Z., Prisacariu, V.A.: Neural refinement for absolute pose regression with feature synthesis. In: CVPR (2024)", + "25. Chen, S., Li, X., Wang, Z., Prisacariu, V.: DFNet: Enhance absolute pose regression with direct feature matching. In: ECCV (2022)", + "26. Chen, S., Wang, Z., Prisacariu, V.: Direct-PoseNet: Absolute pose regression with photometric consistency. In: 3DV (2021)", + "27. Cheng, Z., Esteves, C., Jampani, V., Kar, A., Maji, S., Makadia, A.: LU-NeRF: Scene and pose estimation by synchronizing local unposed NeRFs. In: ICCV (2023)", + "28. Crandall, D., Owens, A., Snavely, N., Huttenlocher, D.: Discrete-continuous optimization for large-scale structure from motion. In: CVPR (2011)", + "29. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: ScanNet: Richly-annotated 3D reconstructions of indoor scenes. In: CVPR (2017)", + "30. Davison, A.J.: Real-time simultaneous localisation and mapping with a single camera. In: ICCV (2003)", + "31. DeTone, D., Malisiewicz, T., Rabinovich, A.: SuperPoint: Self-supervised interest point detection and description. In: CVPRW (2018)", + "32. Ding, M., Wang, Z., Sun, J., Shi, J., Luo, P.: CamNet: Coarse-to-fine retrieval for camera re-localization. In: ICCV (2019)", + "33. Dusmanu, M., Rocco, I., Pajdla, T., Pollefeys, M., Sivic, J., Torii, A., Sattler, T.: D2-net: A trainable CNN for joint description and detection of local features. In: CVPR (2019)", + "34. Fischler, M.A., Bolles, R.C.: Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Comm. of the ACM (1981)", + "35. Gao, X.S., Hou, X.R., Tang, J., Cheng, H.F.: Complete solution classification for the perspective-three-point problem. IEEE TPAMI (2003)", + "36. Gherardi, R., Farenzena, M., Fusiello, A.: Improving the efficiency of hierarchical structure-and-motion. In: CVPR (2010)", + "37. Govindu, V.M.: Combining two-view constraints for motion estimation. In: CVPR (2001)", + "38. Govindu, V.M.: Lie-algebraic averaging for globally consistent motion estimation. In: CVPR (2004)", + "39. Hartley, R., Trumpf, J., Dai, Y., Li, H.: Rotation averaging. IJCV (2013)", + "40. Hartley, R., Zisserman, A.: Multiple view geometry in computer vision. Cambridge University Press (2003)", + "41. He, X., Sun, J., Wang, Y., Peng, S., Huang, Q., Bao, H., Zhou, X.: Detector-free structure from motion. In: CVPR (2024)", + "42. Heinly, J., Schonberger, J.L., Dunn, E., Frahm, J.M.: Reconstructing the World in six days. In: CVPR (2015)", + "43. Humenberger, M., Cabon, Y., Pion, N., Weinzaepfel, P., Lee, D., Guérin, N., Sattler, T., Csurka, G.: Investigating the role of image retrieval for visual localization: An exhaustive benchmark. IJCV (2022)" + ], + "bbox": [ + 223, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "E. Brachmann et al.", + "bbox": [ + 271, + 114, + 408, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "44. Izadi, S., Kim, D., Hilliges, O., Molyneaux, D., Newcombe, R., Kohli, P., Shotton, J., Hodges, S., Freeman, D., Davison, A.J., Fitzgibbon, A.: KinectFusion: real-time 3D reconstruction and interaction using a moving depth camera. In: UIST (2011)", + "45. Jeong, Y., Ahn, S., Choy, C., Anandkumar, A., Cho, M., Park, J.: Self-calibrating neural radiance fields. In: ICCV (2021)", + "46. Jin, Y., Mishkin, D., Mishchuk, A., Matas, J., Fua, P., Yi, K.M., Trulls, E.: Image matching across wide baselines: From paper to practice. IJCV (2021)", + "47. Kendall, A., Grimes, M., Cipolla, R.: PoseNet: A convolutional network for real-time 6-DoF camera relocalization. In: ICCV (2015)", + "48. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3D gaussian splatting for real-time radiance field rendering. ACM TOG (2023)", + "49. Knapitsch, A., Park, J., Zhou, Q.Y., Koltun, V.: Tanks and Temples: Benchmarking large-scale scene reconstruction. ACM TOG (2017)", + "50. Kraus, K.: Photogrammetry. No. v. 1 in Photogrammetry, Ferdinand Dummlers Verlag (1993)", + "51. Laskar, Z., Melekhov, I., Kalia, S., Kannala, J.: Camera relocalization by computing pairwise relative poses using convolutional neural network. In: ICCV Workshops (2017)", + "52. Li, X., Wang, S., Zhao, Y., Verbeek, J., Kannala, J.: Hierarchical scene coordinate classification and regression for visual localization. In: CVPR (2020)", + "53. Lin, A., Zhang, J.Y., Ramanan, D., Tulsiani, S.: Relpose++: Recovering 6d poses from sparse-view observations. In: 3DV (2024)", + "54. Lin, C.H., Ma, W.C., Torralba, A., Lucey, S.: BARF: Bundle-adjusting neural radiance fields. In: ICCV (2021)", + "55. Lin, Y., Müller, T., Tremblay, J., Wen, B., Tyree, S., Evans, A., Vela, P.A., Birchfield, S.: Parallel inversion of neural radiance fields for robust pose estimation. In: ICRA (2023)", + "56. Lindenberger, P., Sarlin, P.E., Pollefeys, M.: LightGlue: Local feature matching at light speed. In: ICCV (2023)", + "57. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: ICLR (2019)", + "58. Lowe, D.G.: Distinctive image features from scale-invariant keypoints. IJCV (2004)", + "59. Martinec, D., Pajdla, T.: Robust rotation and translation estimation in multiview reconstruction. In: CVPR (2007)", + "60. Meng, Q., Chen, A., Luo, H., Wu, M., Su, H., Xu, L., He, X., Yu, J.: GNeRF: GAN-based Neural Radiance Field without Posed Camera. In: ICCV (2021)", + "61. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)", + "62. Moreau, A., Piasco, N., Bennehar, M., Tsishkou, D., Stanciulescu, B., de La Fortelle, A.: CROSSFIRE: Camera relocalization on self-supervised features from an implicit representation. ICCV (2023)", + "63. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM TOG (2022)", + "64. Newcombe, R., Izadi, S., Hilliges, O., Molyneaux, D., Kim, D., Davison, A.J., Kohi, P., Shotton, J., Hodges, S., Fitzgibbon, A.: KinectFusion: Real-time dense surface mapping and tracking. In: ISMAR (2011)", + "65. Nistér, D., Naroditsky, O., Bergen, J.: Visual odometry. In: CVPR (2004)", + "66. Pollefeys, M., Koch, R., Vergauwen, M., Van Gool, L.: Automated reconstruction of 3D scenes from sequences of images. J. of Photogr. and Rem. Sens. (2000)" + ], + "bbox": [ + 225, + 146, + 785, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Scene Coordinate Reconstruction", + "bbox": [ + 508, + 114, + 730, + 127 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "67. Rau, A., Garcia-Hernando, G., Stoyanov, D., Brostow, G.J., Turmukhambetov, D.: Predicting visual overlap of images through interpretable non-metric box embeddings. In: ECCV (2020)", + "68. Reality, C.: Reality Capture. https://www.capturingreality.com/ realitycapture (2016), [accessed 15-Nov-2023]", + "69. Reizenstein, J., Shapovalov, R., Henzler, P., Sbordone, L., Labatut, P., Novotny, D.: Common objects in 3D: Large-scale learning and evaluation of real-life 3D category reconstruction. In: ICCV (2021)", + "70. Sarlin, P.E., Cadena, C., Siegwart, R., Dymczyk, M.: From coarse to fine: Robust hierarchical localization at large scale. In: CVPR (2019)", + "71. Sarlin, P.E., DeTone, D., Malisiewicz, T., Rabinovich, A.: SuperGlue: Learning feature matching with graph neural networks. In: CVPR (2020)", + "72. Sarlin, P.E., Unagar, A., Larsson, M., Germain, H., Toft, C., Larsson, V., Pollefeys, M., Lepetit, V., Hammarstrand, L., Kahl, F., Sattler, T.: Back to the feature: Learning robust camera localization from pixels to pose. In: CVPR (2021)", + "73. Sattler, T., Leibe, B., Kobbelt, L.: Fast image-based localization using direct 2D-to-3D matching. In: ICCV (2011)", + "74. Sattler, T., Leibe, B., Kobbelt, L.: Improving image-based localization by active correspondence search. In: ECCV (2012)", + "75. Sattler, T., Leibe, B., Kobbelt, L.: Efficient & effective prioritized matching for large-scale image-based localization. IEEE TPAMI (2016)", + "76. Sattler, T., Torii, A., Sivic, J., Pollefeys, M., Taira, H., Okutomi, M., Pajdla, T.: Are large-scale 3D models really necessary for accurate visual localization? In: CVPR (2017)", + "77. Sattler, T., Zhou, Q., Pollefeys, M., Leal-Taixe, L.: Understanding the limitations of CNN-based absolute camera pose regression. In: CVPR (2019)", + "78. Schaffalitzky, F., Zisserman, A.: Multi-view matching for unordered image sets, or \"How do i organize my holiday snaps?\". In: ECCV (2002)", + "79. Schonberger, J.L.: Colmap Github Issues. https://github.com/colmap/colmap/issues/116#issuecomment-298926277 (2017), [accessed Nov/15/2023]", + "80. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: CVPR (2016)", + "81. Shotton, J., Glocker, B., Zach, C., Izadi, S., Criminisi, A., Fitzgibbon, A.: Scene coordinate regression forests for camera relocalization in RGB-D images. In: CVPR (2013)", + "82. Sinha, S., Zhang, J.Y., Tagliasacchi, A., Gilitschenski, I., Lindell, D.B.: Sparse-Pose: Sparse-view camera pose regression and refinement. In: CVPR (2023)", + "83. Smith, L.N., Topin, N.: Super-convergence: Very fast training of neural networks using large learning rates. In: Arti. Intell. and Mach. Learn. for Multi-Domain Operations Appl. (2019)", + "84. Snavely, N., Seitz, S.M., Szeliski, R.: Photo tourism: exploring photo collections in 3D. ACM TOG (2006)", + "85. Snavely, N., Seitz, S.M., Szeliski, R.: Modeling the world from internet photo collections. IJCV (2008)", + "86. Snavely, N., Seitz, S.M., Szeliski, R.: Skeletal graphs for efficient structure from motion. In: CVPR (2008)", + "87. Sun, J., Shen, Z., Wang, Y., Bao, H., Zhou, X.: LoFTR: Detector-free local feature matching with transformers. In: CVPR (2021)", + "88. Szeliski, R., Kang, S.B.: Recovering 3D shape and motion from image streams using nonlinear least squares. J. of Vis. Com. and Image Repr. (1994)" + ], + "bbox": [ + 225, + 146, + 784, + 840 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "E. Brachmann et al.", + "bbox": [ + 271, + 114, + 408, + 127 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "89. Tancik, M., Weber, E., Ng, E., Li, R., Yi, B., Kerr, J., Wang, T., Kristoffersen, A., Austin, J., Salahi, K., Ahuja, A., McAllister, D., Kanazawa, A.: Nerfstudio: A modular framework for neural radiance field development. In: ACM TOG (2023)", + "90. Teed, Z., Deng, J.: DROID-SLAM: Deep visual SLAM for monocular, stereo, and RGB-D cameras. In: NeurIPS (2021)", + "91. Toldo, R., Gherardi, R., Farenzena, M., Fusiello, A.: Hierarchical structure-and-motion recovery from uncalibrated images. CVIU (2015)", + "92. Triggs, B., McLauchlan, P.F., Hartley, R.I., Fitzgibbon, A.W.: Bundle adjustment—a modern synthesis. In: Int. Worksh. on Vis. Alg. (2000)", + "93. Türkoglu, M.Ö., Brachmann, E., Schindler, K., Brostow, G., Monszpart, A.: Visual camera re-localization using graph neural networks and relative pose supervision. In: 3DV (2021)", + "94. Ulyanov, D., Vedaldi, A., Lempitsky, V.: Deep image prior. In: CVPR (2018)", + "95. Ummenhofer, B., Zhou, H., Uhrig, J., Mayer, N., Ilg, E., Dosovitskiy, A., Brox, T.: DeMoN: Depth and motion network for learning monocular stereo. In: CVPR (2017)", + "96. Waechter, M., Beljan, M., Fuhrmann, S., Moehrle, N., Kopf, J., Goesele, M.: Virtual rephotography: Novel view prediction error for 3d reconstruction. ACM TOG (2017)", + "97. Wang, S., Leroy, V., Cabon, Y., Chidlovskii, B., Revaud, J.: DUSt3R: Geometric 3D vision made easy. In: CVPR (2024)", + "98. Wang, Z., Wu, S., Xie, W., Chen, M., Prisacariu, V.A.: NeRF--: Neural radiance fields without known camera parameters. arXiv (2021)", + "99. Wei, X., Zhang, Y., Li, Z., Fu, Y., Xue, X.: DeepSFM: Structure from motion via deep bundle adjustment. In: ECCV (2020)", + "00. Wu, C.: Towards linear-time incremental structure from motion. In: 3DV (2013)", + "01. Xia, Y., Tang, H., Timofte, R., Van Gool, L.: SiNeRF: Sinusoidal neural radiance fields for joint pose estimation and scene reconstruction. In: BMVC (2022)", + "02. Yen-Chen, L., Florence, P., Barron, J.T., Rodriguez, A., Isola, P., Lin, T.Y.: iNeRF: Inverting neural radiance fields for pose estimation. In: IROS (2021)", + "03. Zhang, J.Y., Lin, A., Kumar, M., Yang, T.H., Ramanan, D., Tulsiani, S.: Cameras as rays: Pose estimation via ray diffusion. In: ICLR (2024)", + "04. Zhang, W., Kosecka, J.: Image based localization in urban environments. In: 3DPVT (2006)", + "05. Zhou, Q., Sattler, T., Pollefeys, M., Leal-Taixe, L.: To learn or not to learn: Visual localization from essential matrices. In: ICRA (2020)", + "06. Zhou, Y., Barnes, C., Jingwan, L., Jimei, Y., Hao, L.: On the continuity of rotation representations in neural networks. In: CVPR (2019)" + ], + "bbox": [ + 225, + 146, + 784, + 672 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "Scene Coordinate Reconstruction", + "bbox": [ + 508, + 114, + 730, + 127 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/Scene Coordinate Reconstruction_ Posing of Image Collections via Incremental Learning of a Relocalizer/c513fe61-35e0-435f-acdb-dbf8715686b2_model.json b/2024/Scene Coordinate Reconstruction_ Posing of Image Collections via Incremental Learning of a Relocalizer/c513fe61-35e0-435f-acdb-dbf8715686b2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..62c45bc35a5a694da5cd4f66c36a175d86cdbd8c --- /dev/null +++ b/2024/Scene Coordinate Reconstruction_ Posing of Image Collections via Incremental Learning of a Relocalizer/c513fe61-35e0-435f-acdb-dbf8715686b2_model.json @@ -0,0 +1,2845 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.251, + 0.142, + 0.756, + 0.21 + ], + "angle": 0, + "content": "Scene Coordinate Reconstruction: Posing of Image Collections via Incremental Learning of a Relocalizer" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.235, + 0.787, + 0.267 + ], + "angle": 0, + "content": "Eric Brachmann1, Jamie Wynn1, Shuai Chen2, Tommaso Cavallari1, Åron Monszpart1, Daniyar Turmukhambetov1, and Victor Adrian Prisacariu1,2" + }, + { + "type": "text", + "bbox": [ + 0.424, + 0.277, + 0.581, + 0.306 + ], + "angle": 0, + "content": "\\(^{1}\\) Niantic \\(^{2}\\) University of Oxford" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.341, + 0.744, + 0.562 + ], + "angle": 0, + "content": "Abstract. We address the task of estimating camera parameters from a set of images depicting a scene. Popular feature-based structure-from-motion (SfM) tools solve this task by incremental reconstruction: they repeat triangulation of sparse 3D points and registration of more camera views to the sparse point cloud. We re-interpret incremental structure-from-motion as an iterated application and refinement of a visual relocalizer, that is, of a method that registers new views to the current state of the reconstruction. This perspective allows us to investigate alternative visual relocalizers that are not rooted in local feature matching. We show that scene coordinate regression, a learning-based relocalization approach, allows us to build implicit, neural scene representations from unposed images. Different from other learning-based reconstruction methods, we do not require pose priors nor sequential inputs, and we optimize efficiently over thousands of images. In many cases, our method, ACE0, estimates camera poses with an accuracy close to feature-based SfM, as demonstrated by novel view synthesis." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.564, + 0.653, + 0.577 + ], + "angle": 0, + "content": "Project page: https://nianticlabs.github.io/acezero/" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.602, + 0.377, + 0.617 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.322, + 0.633, + 0.681, + 0.649 + ], + "angle": 0, + "content": "In the beginning there was structure-from-motion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.78 + ], + "angle": 0, + "content": "The genesis of numerous computer vision tasks lies in the estimation of camera poses and scene geometry from a set of images. It is the first fundamental step that lets us leave the image plane and venture into 3D. Since structure-from-motion (SfM) is such a central capability, we have researched it for decades. By now, refined open-source tools, such as COLMAP [80], and efficient commercial packages, such as RealityCapture [68], are available to us. Feature matching-based SfM is the gold standard for estimating poses from images, with a precision that makes its estimates occasionally considered \"ground truth\" [3,5,13,46,69]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The success of Neural Radiance Fields (NeRF) [61] has renewed interest in the question of whether SfM can be solved differently, based on neural, implicit scene representations rather than 3D point clouds. There has been some progress in recent years but, thus far, learning-based approaches to camera pose recovery" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.409, + 0.128 + ], + "angle": 0, + "content": "E. Brachmann et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.145, + 0.782, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.325, + 0.788, + 0.451 + ], + "angle": 0, + "content": "Fig. 1: Reconstructing 10,000 Images. Top: Starting from a single image and the identity pose, we train a learning-based visual relocalizer. The relocalizer allows us to estimate the poses of more views, and the additional views allow us to refine the relocalizer. We show three out of six iterations for this scene (7Scenes Office [81]). All 10k images have been posed in roughly 1 hour on a single GPU. In comparison NoPe-NeRF [10] needs two days to pose 200 images. The point cloud is a visualization of the implicit scene representation of the relocalizer. Camera positions are color coded by relocalization confidence from yellow (low) to green (high). Bottom: Point clouds from Nerfacto [89] trained on top of our poses for a few scenes from our experiments." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.487, + 0.788, + 0.608 + ], + "angle": 0, + "content": "still have significant limitations. They either require coarse initial poses [45, 54, 98, 101], prior knowledge of the pose distribution [60] or sequential inputs [10, 11, 90]. In terms of the number of images that learning-based approaches can handle, they are either explicitly targeted at few-frame problems [53, 82, 95, 97, 99, 103] or they are computationally so demanding that they can realistically only be applied to a few hundred images at most [10, 54, 98]. We show that none of these limitations are an inherent consequence of using learning-based scene representations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.611, + 0.788, + 0.807 + ], + "angle": 0, + "content": "Our approach is inspired by incremental SfM and its relationship to another computer vision task: visual relocalization. Visual relocalization describes the problem of estimating the camera pose of a query image w.r.t. to an existing scene map. Incremental SfM can be re-interpreted as a loop of 1) do visual relocalization to register new views to the reconstruction, and 2) refine/extend the reconstruction using the newly registered views. Local feature matching is a traditional approach to visual relocalization [70,74-76]. In recent years, multiple learning-based relocalizers have been proposed that encode the scene implicitly in the weights of a neural network [3, 12, 14, 15, 17, 22, 47]. Not all of them are suitable for building a SfM pipeline. We need a relocalizer with high accuracy and good generalization. Training of the relocalizer has to be swift. We need to be able to bootstrap relocalization without ground truth poses. And we need to be able to tell whether registration of a new image was successful." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "We show that \"scene coordinate regression\", an approach to visual relocalization proposed a decade ago [81], has the desirable properties and can serve as" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.509, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Scene Coordinate Reconstruction" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "the core of a new approach to learning-based SfM: Scene coordinate reconstruction. Rather than optimizing over image-to-image matches, like feature-based SfM, scene coordinate reconstruction regresses image-to-scene correspondences directly. Rather than representing the scene as a 3D point cloud with high dimensional descriptors, we encode the scene into a lightweight neural network. Our approach works on unsorted images without pose priors and efficiently optimises over thousands of images, see Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.253, + 0.473, + 0.267 + ], + "angle": 0, + "content": "We summarize our contributions:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.276, + 0.786, + 0.305 + ], + "angle": 0, + "content": "- Scene Coordinate Reconstruction, a new approach to SfM based on incremental learning of scene coordinate regression, a visual relocalization principle." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.306, + 0.786, + 0.35 + ], + "angle": 0, + "content": "- We turn the fast-learning visual relocalizer ACE [12] into a SfM framework that is able to predict the camera poses of a set of unposed RGB images. We refer to this new SfM pipeline as \\(ACE0\\) (ACE Zero)." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.351, + 0.786, + 0.411 + ], + "angle": 0, + "content": "- Compared to ACE [12], we add the capability to train in a self-supervised fashion. We start from a single image, and iterate between learning the map and registering new views. We expedite reconstruction times by early stopping, and increase reconstruction quality by pose refinement." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.276, + 0.786, + 0.411 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.433, + 0.388, + 0.449 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.463, + 0.788, + 0.509 + ], + "angle": 0, + "content": "Reconstruction. SfM pipelines either ingest a collection of unordered images [20,78,84] or an ordered image sequence [6,30,65,66,88] from a video to recover 3D structure of a scene and camera poses (\"motion\") of the images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.509, + 0.788, + 0.674 + ], + "angle": 0, + "content": "SIFT [58] and other robust descriptors allow matching image features across wide baselines enabling systems to reconstruct large-scale scenes using Internet images [80,84,85,100]. Image-to-image matches can also be regressed directly in a detector-free setup [41,87]. Feature tracks across multiple images are built from image-to-image matches. Feature tracks and estimated relative poses are used to solve for the 3D feature coordinates, camera poses and calibrations (intrinsic matrices). This geometric optimization problem is mainly solved using bundle adjustment which was explored in photogrammetry and geodesy [19, 50] and became standard in the computer vision community [40,92]. Bundle adjustment relies on the initialization being close to the solution (i.e., camera poses and 3D points are already mostly accurate)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.788, + 0.841 + ], + "angle": 0, + "content": "There are two main approaches to this problem. Incremental SfM [6, 66] starts the reconstruction from very few images to create a high-quality seed reconstruction that progressively grows by registering more images and refining the reconstruction until convergence. Global SfM approaches solve for \"global\" poses of all images using estimates of relative poses, i.e., motion averaging [37, 38], rotation averaging [39, 59] and pose-graph optimization [21]. Various techniques were proposed to improve SfM runtime for very large sets of images [1, 2, 8, 9, 28, 36, 42, 86, 91]. Our work is similar to Incremental SfM, as we also progressively register images to the reconstructed scene starting from a seed reconstruction. However, we do not explicitly compute image matches, nor feature tracks across images, which can be computationally expensive." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.409, + 0.129 + ], + "angle": 0, + "content": "E. Brachmann et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.283 + ], + "angle": 0, + "content": "Visual Relocalization. A reconstructed (or mapped) scene is a database of images with known camera poses. This database can be used by a visual relocalizer to estimate poses for new query images to \"relocalize\" a camera in the scene. Feature-based approaches extract 2D local features [31, 33, 56, 58, 71, 72, 87] from a query image and match them to 3D points to solve for the query pose using perspective-n-point (PnP) solvers [35], e.g., [73-75]; or match query features to 2D local features of mapped images to triangulate the query image, e.g., [104, 105]. For large scenes, matching features on a subset of database images relevant to the query can improve the speed and accuracy, e.g., [43, 67, 70, 76]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.79, + 0.481 + ], + "angle": 0, + "content": "Some learning-based approaches encode the map of a scene in the weights of a neural network. PoseNet [47] directly regresses the absolute camera pose given an input image using a CNN that was trained on image-posed pairs of the reconstructed scene. Sattler et al. show that extrapolating outside the mapped poses can be challenging for absolute pose regression approaches [77]. Relative pose regression networks [4, 32, 51, 93] estimate the relative pose for a pair of images, allowing triangulation of the query image from multiple map images, or estimate a scale-metric relative pose w.r.t. a single map image [3]. Scene coordinate regression [12, 14, 15, 17, 22, 23, 52, 81] directly predicts 3D coordinates of a point given a patch. This approach generalizes well as the camera pose is solved using PnP [35] within a RANSAC loop [34]. ACE [12] shows that training of scene coordinate regression can be greatly accelerated. In our work we train the ACE localizer, but we start from images without poses." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.495, + 0.79, + 0.753 + ], + "angle": 0, + "content": "Image-Based Rendering. In recent years, neural radiance fields (NeRFs) [61] have seen a lot of attention from the community. NeRFs allow photorealistic novel view synthesis when trained from image-pose pairs. Typically, estimation of the image poses is done in advance by using an SfM pipeline such as COLMAP, e.g., [5]. Nonetheless, research exists that estimates camera poses with NeRFs, facilitates camera localization for novel views after NeRF training [24-26, 62, 102], or simultaneously estimates camera poses during NeRF training from images alone [10, 27, 54, 82, 98]. However, these approaches either assume that the scene is captured from the front [45, 98, 101], that coarse poses are available for initialization [54], or that images are captured sequentially [10]. Techniques that rely on a multi-layer perceptron (MLP) representation of the scene, e.g. [10, 27], are slow to train, taking days to converge. While radiance fields can be trained faster using multi-layer hash grids [63] or Gaussian splats [48], their efficacy in pose estimation without approximate prior pose initialization [55, 62] or sequential image inputs [48] remains unproven. Concurrently, several learning-based camera pose estimation methods have been proposed [53, 97, 103]. Due to GPU memory constraints, these methods estimate poses for sparse sets of images." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.777, + 0.333, + 0.794 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.81, + 0.79, + 0.843 + ], + "angle": 0, + "content": "Preliminaries. The input to our system is a set of RGB images, denoted by \\(\\mathcal{I} = \\{I_i\\}\\), where \\(i\\) refers to the image index. Our system estimates the corresponding" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.509, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Scene Coordinate Reconstruction" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.145, + 0.778, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.389, + 0.789, + 0.53 + ], + "angle": 0, + "content": "Fig. 2: ACE0 Framework. Top left: We loop between learning a reconstruction from the current set of images and poses (\"neural mapping\"), and estimating poses of more images (\"relocalization\"). Top right: During the mapping stage, we train a scene coordinate regression network as our scene representation. Camera poses of the last relocalization round and camera calibration parameters are refined during this process. We visualize scene coordinates by mapping XYZ to the RGB cube. Bottom: In the relocalization stage, we re-estimate poses of images using the scene coordinate regression network, including images that were previously not registered to the reconstruction. If the registration of an image succeeds, it will be used in the next iteration of the mapping stage; otherwise it will not." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.556, + 0.789, + 0.616 + ], + "angle": 0, + "content": "set of camera parameters, both intrinsic and extrinsic: \\(\\mathcal{H} = \\{(K_i,T_i)\\}\\). Each \\(T_{i}\\) refers to a \\(3\\times 4\\) matrix containing a rotation and translation, while \\(K_{i}\\) refers to a \\(3\\times 3\\) matrix with the calibration parameters. We assume no particular image order or any prior knowledge about the pose distribution." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.616, + 0.789, + 0.678 + ], + "angle": 0, + "content": "We also want to recover the 3D structure of the scene: Each pixel \\(j\\) in image \\(i\\) with 2D pixel position \\(\\mathbf{p}_{ij}\\) has a corresponding coordinate in 3D, denoted as \\(\\mathbf{y}_{ij}\\). The 2D pixel positions and 3D scene coordinates are related by the camera pose and the projection function \\(\\pi\\):" + }, + { + "type": "equation", + "bbox": [ + 0.427, + 0.686, + 0.787, + 0.704 + ], + "angle": 0, + "content": "\\[\n\\mathbf {p} _ {i j} = \\boldsymbol {\\pi} \\left(K _ {i}, T _ {i}, \\mathbf {y} _ {i j}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.71, + 0.789, + 0.741 + ], + "angle": 0, + "content": "where \\( T_{i} \\) maps camera coordinates to scene coordinates, and \\( K_{i} \\) projects camera coordinates to the image plane." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.741, + 0.792, + 0.786 + ], + "angle": 0, + "content": "As our scene representation, we utilize a scene coordinate regression model [81], i.e., a learnable function \\( f_{\\mathrm{SCR}} \\) that maps an image patch of image \\( I_{i} \\), centered around pixel position \\( \\mathbf{p}_{ij} \\) to a scene coordinate: \\( \\mathbf{y}_{ij} = f_{\\mathrm{SCR}}(\\mathbf{p}_{ij},I_i) \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.786, + 0.789, + 0.817 + ], + "angle": 0, + "content": "Given a set of 2D-3D correspondences predicted by \\( f_{\\mathrm{SCR}} \\) for any image \\( I_i \\), we can recover this image's camera pose \\( T_i \\) using a pose solver \\( g \\):" + }, + { + "type": "equation", + "bbox": [ + 0.41, + 0.824, + 0.787, + 0.843 + ], + "angle": 0, + "content": "\\[\nT _ {i} = g \\left(K _ {i}, \\left\\{\\left(\\mathbf {p} _ {i j}, \\mathbf {y} _ {i j}\\right) \\right\\}\\right). \\tag {2}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.409, + 0.128 + ], + "angle": 0, + "content": "E. Brachmann et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "Since 2D-3D correspondences can be inaccurate, and contain incorrect predictions, \\( g \\) combines a PnP solver [35] with a RANSAC loop [34]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.298 + ], + "angle": 0, + "content": "Normally, scene coordinate regression models are trained in a supervised fashion for the task of visual relocalization [12, 14, 15, 17, 22, 23, 52, 81]. That is, \\( f_{\\mathrm{SCR}} \\) is trained using images with known ground truth camera parameters \\( \\{(I_i, T_i^{\\mathrm{GT}}, K_i^{\\mathrm{GT}})\\} \\), and used for estimating the poses of unseen query images. Instead, we show how these models can be trained self-supervised, without ground truth poses, to estimate the poses of the mapping images themselves. Thus, we turn scene coordinate regression into scene coordinate reconstruction, a learning-based SfM tool." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.321, + 0.4, + 0.337 + ], + "angle": 0, + "content": "3.1 Neural Mapping" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.348, + 0.784, + 0.392 + ], + "angle": 0, + "content": "We train the scene coordinate regression model iteratively where we denote the current time step as \\( t \\) and the corresponding scene model as \\( f_{\\mathrm{SCR}}^t \\). We iterate between training the scene model, and registering new views, see Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.393, + 0.786, + 0.456 + ], + "angle": 0, + "content": "At iteration \\( t \\) we assume that a subset of images has already been registered to the scene, \\( \\mathcal{I}_{\\mathrm{Reg}}^t \\subset \\mathcal{I} \\), and where corresponding camera parameters, \\( T_{i}^{t} \\) and \\( K_{i}^{t} \\), have already been estimated. Using these as pseudo ground truth, we train the scene model by minimizing the pixel-wise reprojection error:" + }, + { + "type": "equation", + "bbox": [ + 0.377, + 0.466, + 0.786, + 0.502 + ], + "angle": 0, + "content": "\\[\n\\sum_ {I _ {i} \\in \\mathcal {I} _ {\\text {R e g}} ^ {t}} \\sum_ {j \\in I _ {i}} \\| \\mathbf {p} _ {i j} - \\boldsymbol {\\pi} \\left(K _ {i} ^ {t}, T _ {i} ^ {t}, \\mathbf {y} _ {i j} ^ {t}\\right) \\|, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.514, + 0.6, + 0.531 + ], + "angle": 0, + "content": "where the scene model \\( f_{\\mathrm{SCR}}^{t} \\) predicts coordinates \\( \\mathbf{y}_{ij}^{t} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.531, + 0.788, + 0.68 + ], + "angle": 0, + "content": "Mapping Framework. We optimize Eq. 3 using stochastic gradient descent, using the fast-learning scene coordinate regressor ACE [12] (Accelerated Coordinate Encoding). ACE is trained in minutes, even for thousands of views. The training speed is important since we have to train the model in multiple iterations. ACE approximates Eq. 3 by sampling random patches from the training set, \\(\\mathcal{I}_{\\mathrm{Reg}}^t\\). To do that efficiently, ACE employs a pre-trained encoder to precompute high dimensional features for a large number of training patches. The encoder stays frozen during mapping. The actual scene model, which is trained, is a multi-layer perceptron (MLP) that maps encoder features to scene coordinates; see Fig. 2 (top right) for a visual representation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.681, + 0.788, + 0.756 + ], + "angle": 0, + "content": "Pose Refinement. Differently from the ACE [12] protocol, the ground truth poses \\( T_{i}^{\\mathrm{GT}} \\) are unknown during training. Instead, we have \\( T_{i}^{t} \\), estimates based on earlier iterations of the reconstruction. Since these estimates can be inaccurate, we add the ability to refine poses during mapping. We implement refinement using an MLP:" + }, + { + "type": "equation", + "bbox": [ + 0.445, + 0.756, + 0.786, + 0.773 + ], + "angle": 0, + "content": "\\[\nT _ {i} ^ {t} = f _ {\\text {P o s e}} ^ {t} \\left(\\hat {T} _ {i} ^ {t}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.779, + 0.788, + 0.843 + ], + "angle": 0, + "content": "where \\(\\tilde{T}_i^t\\) denotes the initial pose estimate at the start of a mapping iteration. Inspired by [106], the refinement MLP ingests \\(\\tilde{T}_i^t\\) as \\(3\\times 4 = 12\\) values and predicts 12 additive offsets. We orthonormalize rotations using Gram-Schmidt [18]. We jointly optimize \\(f_{\\mathrm{Pose}}^t\\) and \\(f_{\\mathrm{SCR}}^t\\) to minimize the reprojection error of Eq. 3. We" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.509, + 0.115, + 0.731, + 0.127 + ], + "angle": 0, + "content": "Scene Coordinate Reconstruction" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.144, + 0.5, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.144, + 0.785, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.308, + 0.788, + 0.406 + ], + "angle": 0, + "content": "Fig. 3: Left: Pose Refinement. Since we register images based on a coarse and incomplete state of the reconstruction, we add the ability to refine poses during neural mapping. An MLP predicts pose updates relative to the initial poses, supervised by the reprojection error of scene coordinates. Right: Initialization. To start the reconstruction, we train the network using one image, the identity pose and a depth estimate, here ZoeDepth [7]. In this example, we register 33 views to the initial reconstruction. Depth estimates are only used for this step." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.44, + 0.785, + 0.486 + ], + "angle": 0, + "content": "show the impact of pose refinement for one iteration in Fig. 3 (left). We discard the refinement MLP after each mapping iteration. Its purpose is to enable the scene model, \\( f_{\\mathrm{SCR}}^{t} \\), to converge to a consistent scene representation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.488, + 0.788, + 0.594 + ], + "angle": 0, + "content": "With neither poses nor 2D-3D correspondences fixed in Eq. 3, the scene coordinate regressor could drift or degenerate. As regularization, we optimize the pose refiner \\( f_{\\mathrm{Pose}}^{t} \\) using AdamW [57] with weight decay, biasing the MLP to predict small updates relative to the initial estimates \\( \\tilde{T}_i^t \\). This relies on the assumption that mapping images in \\( \\mathcal{I}_{\\mathrm{Reg}}^t \\) have been registered close to their true position. If that assumption holds, the smoothness prior of the networks [94] encourages a multi-view consistent solution as shown in previous work [12,15,17]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.597, + 0.788, + 0.672 + ], + "angle": 0, + "content": "As an alternative to an MLP refiner, we could back-propagate directly to the input poses [98, 102]. However, when optimizing over thousands of views, the signal for each single pose becomes sparse. Cameras are correlated via the scene representation. If the optimization removes drift in the scene, multiple cameras need to move. The MLP refiner models the correlation of cameras." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Calibration Refinement. We do not assume information about precise calibration parameters, although often reported by devices. We do assume that the principal point is in the center, that pixels are unskewed and square. We do not model image distortion. While these are reasonable assumptions for many data regimes, we cannot rely on the focal length to be given. Thus, we refine the focal length starting from a heuristic: \\( K_{i}^{t} = f_{\\mathrm{Calib}}^{t}(K_{i}^{\\mathrm{init}}) \\). The refinement function \\( f_{\\mathrm{Calib}}^{t} \\) entails a single learnable parameter \\( \\alpha^{t} \\) such that the focal length \\( f_{i}^{t} = f^{\\mathrm{init}} \\cdot (1 + \\alpha^{t}) \\). As before, superscript \\( t \\) denotes the time step. We optimize \\( \\alpha^{t} \\) using AdamW [57] with weight decay, biasing it towards a small relative scale factor. Estimates of \\( \\alpha \\) are carried over across iterations. We set \\( f^{\\mathrm{init}} \\) to 70% of the image diagonal and it is shared by all cameras in our experiments." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.409, + 0.128 + ], + "angle": 0, + "content": "E. Brachmann et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.381, + 0.161 + ], + "angle": 0, + "content": "3.2 Relocalization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.17, + 0.788, + 0.232 + ], + "angle": 0, + "content": "Given the scene model of iteration \\(t\\), we attempt to register more images to determine the training set for the next mapping iteration, \\(\\mathcal{I}_{\\mathrm{Reg}}^{t + 1}\\). We pass all images in \\(\\mathcal{I}\\) to the scene coordinate regressor \\(f_{\\mathrm{SCR}}^t\\) to gather 2D-3D correspondences, and solve for their poses using RANSAC and PnP:" + }, + { + "type": "equation", + "bbox": [ + 0.382, + 0.24, + 0.786, + 0.26 + ], + "angle": 0, + "content": "\\[\n\\tilde {T} _ {i} ^ {t + 1}, s _ {i} ^ {t + 1} = g \\left(K _ {i} ^ {t}, \\left\\{\\left(\\mathbf {p} _ {i j}, \\mathbf {y} _ {i j} ^ {t}\\right) \\right\\}\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.268, + 0.789, + 0.345 + ], + "angle": 0, + "content": "Here, we assume that the pose solver returns a confidence score \\( s_i^{t+1} \\) alongside the pose itself that lets us decide whether the pose of the image has been estimated successfully. We simply utilize the inlier count as score \\( s_i^{t+1} \\) and apply a threshold to form the training set of the next mapping iteration: \\( \\mathcal{I}_{\\mathrm{Reg}}^{t+1} = \\left\\{ I_i | s_i^{t+1} > \\tau_s \\right\\} \\). The relocalization process is depicted in Figure 2, bottom." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.365, + 0.37, + 0.379 + ], + "angle": 0, + "content": "3.3 Initialization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.387, + 0.787, + 0.421 + ], + "angle": 0, + "content": "We start the reconstruction with one image: \\(\\mathcal{I}_{\\mathrm{Reg}}^0 = \\{I_{\\mathrm{seed}}\\}\\). We set the seed pose \\(T_{\\mathrm{seed}}^0\\) to identity, and we initialize the calibration \\(K_{\\mathrm{seed}}^{\\mathrm{init}}\\) as explained above." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.421, + 0.787, + 0.525 + ], + "angle": 0, + "content": "We cannot train a scene coordinate regression network using Eq. 3 with a single image. The reprojection error is ill-conditioned without multiple views constraining the depth. Therefore, we optimize a different objective in the seed iteration, inspired by Map-free Relocalization [3]. Arnold et al. argue that a single image and a depth estimate allow to relocalize query images, albeit with limited accuracy. Our experiments show that such coarse relocalizations of a few images suffice as initialization for optimizing the reprojection error of Eq. 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.526, + 0.787, + 0.601 + ], + "angle": 0, + "content": "Let \\(d_{ij}\\) be a depth value predicted for pixel \\(j\\) of image \\(i\\). We derive a target scene coordinate by back-projection as \\(\\hat{\\mathbf{y}}_{ij} = d_{ij}(K_{\\mathrm{seed}}^{\\mathrm{init}})^{-1}\\mathbf{p}_{ij}\\). We train an initial scene coordinate regression network \\(f_{\\mathrm{SCR}}^{0}\\) by optimizing \\(\\sum_{j\\in I^0}\\| \\hat{\\mathbf{y}}_{ij} - \\mathbf{y}_{ij}\\|\\). Fig. 3 (right) shows a scene coordinate point cloud learned from a depth estimate, and successful relocalization against it." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.601, + 0.787, + 0.722 + ], + "angle": 0, + "content": "We found our pipeline to be robust w.r.t. selecting the seed image. However, when a randomly selected image has little to no visual overlap with the remaining images, the whole reconstruction would fail. Selecting an unfortunate seed image, e.g., at the very end of a long camera trajectory, can increase reconstruction times. To decrease the probability of such incidents, we try 5 random seed images and choose the one with the highest relocalization rate across 1000 other mapping images. Since mapping seed images is fast, ca. 1 min on average, this poses no significant computational burden." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.742, + 0.395, + 0.757 + ], + "angle": 0, + "content": "3.4 Implementation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.842 + ], + "angle": 0, + "content": "We base our pipeline on the public code of the ACE relocalizer [12]. ACE uses a convolutional feature backbone, pre-trained on ScanNet [29], that ingests images scaled to 480px height. On top of the backbone is the mapping network, a 9-layer MLP with 512 channels, consuming 4MB of weights in 16-bit floating point precision. This is our scene representation." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.509, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Scene Coordinate Reconstruction" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "The ACE training process is reasonably fast, taking 5 minutes to train the scene network. Since we repeat training the scene network in multiple iterations, we expedite the process further to decrease our total reconstruction time." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.787, + 0.299 + ], + "angle": 0, + "content": "Adaptive Sampling of Training Patches. ACE trains the scene representation based on 8M patches sampled from the mapping images, a process that takes 1 minute. This is excessive when having very few mapping images in the beginning of the reconstruction, thus we loop over the mapping images at most 10 times when sampling patches, or until 8M patches have been sampled. Using this adaptive strategy, sampling patches for the seed reconstruction where we have 1 image only takes 2 seconds instead of 1 minute." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.3, + 0.787, + 0.436 + ], + "angle": 0, + "content": "Adaptive Stopping. ACE trains the scene model using a fixed one-cycle learning rate schedule [83] to a total of 25k parameter updates. We make the training schedule adaptive since the network is likely to converge fast when trained on few images only. We monitor the reprojection error of scene coordinates within a mini-batch. If for 100 consecutive batches \\(70\\%\\) of reprojection errors are below an inlier threshold of 10px, we stop training early. We approximate the one-cycle learning rate schedule in a linear fashion: we increase the learning rate in the first 1k iteration from \\(5 \\times 10^{-4}\\) to \\(3 \\times 10^{-3}\\) and when the early stopping criterion has been met, we decrease the learning rate to \\(5 \\times 10^{-4}\\) within 5k iterations." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.438, + 0.787, + 0.469 + ], + "angle": 0, + "content": "We report more implementation details and hyper-parameters in the supplement. Our code is also publicly available to ensure reproducibility." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.498, + 0.377, + 0.515 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.535, + 0.788, + 0.612 + ], + "angle": 0, + "content": "We refer to our SfM pipeline as ACE0 (ACE Zero) since it builds on top of the ACE relocalizer [12] but adds the ability to train from scratch, without poses. We demonstrate the effectiveness of our approach on three datasets and 31 scenes in total. For all experiments, we rely on ZoeDepth [7] to initialize our reconstructions. All timings reported for ACE0 are based on a single V100 GPU." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.612, + 0.787, + 0.734 + ], + "angle": 0, + "content": "Baselines. We consider the pose estimates of COLMAP with default parameters as our pseudo ground truth, obtained by extensive processing. We also run COLMAP with parameter settings recommended for large image collections of 1k images and more [79]. This variation achieves much faster processing (denoted COLMAP fast). We use a V100 GPU for COLMAP feature extraction and matching, and we specify the exact parameters of COLMAP in the supplement. Furthermore, we show some results of RealityCapture [68], an efficient commercial feature-based SfM pipeline." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.787, + 0.84 + ], + "angle": 0, + "content": "We compare to learning-based SfM approaches that are not restricted to few-frame scenarios, namely to NeRF-based BARF [54] and NoPe-NeRF [10]. We also compare to DUSt3R [97], a non-NeRF learning-based SfM method. While we focus on reconstructing unsorted image collections, the datasets we consider allow for sequential processing. Hence, for context, we also show results of DROID-SLAM [90], a neural SLAM approach. Unless specified otherwise, we report timings based on a single V100 GPU." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.409, + 0.128 + ], + "angle": 0, + "content": "E. Brachmann et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.148, + 0.788, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.393, + 0.789, + 0.464 + ], + "angle": 0, + "content": "Fig. 4: Reconstructed Poses. We show poses estimated by ACE0 for a selection of scenes. We color code the reconstruction iteration in which a particular view has been registered. We show the ACE0 point cloud as a representation of the scene. The seed image is shown as a gray frustum. We also compare our poses to poses estimated by COLMAP (Mip-NeRF 360, Tanks and Temples) and KinectFusion (7-Scenes)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.789, + 0.599 + ], + "angle": 0, + "content": "Benchmark. We show results on 7-Scenes [81], a relocalization dataset, on Mip-NeRF 360 [5], a view synthesis dataset and on Tanks and Temples [49], a reconstruction dataset. Comparing poses on these datasets is problematic, as our pseudo ground truth is estimated rather than measured. For example, an approach might be more accurate than COLMAP on individual scenes. Computing pose errors w.r.t. COLMAP would result in incorrect conclusions. Therefore, we gauge the pose quality in a self-supervised way, using novel view synthesis [96]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.6, + 0.789, + 0.841 + ], + "angle": 0, + "content": "We let each method estimate the poses of all images of a scene. For evaluation, we split the images into training and test sets. We train a Nerfacto [89] model on the training set, and synthesize views for the test poses. We compare the synthesized images to the test images and report the difference as peak signal-to-noise ratio (PSNR). To ensure a fair comparison to NeRF-based competitors, we use these methods in the same way as the other SfM methods: we run them to pose all images, and train a Nerfacto model on top of their poses. This is to ensure that we solely compare the pose quality across methods, and not different capabilities in view synthesis. The quality of poses affects the PSNR numbers in two ways: Good training poses let the NeRF model fit a consistent scene representation. Good testing poses make sure that the synthesized images are aligned with the original image. We explain further details in the supplement which also includes additional, perceptual metrics. In spirit, our evaluation is similar to the Tanks and Temples benchmark which evaluates a derived scene mesh rather than camera poses. However, our evaluation can be applied to arbitrary datasets as it does not need ground truth geometry." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.509, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Scene Coordinate Reconstruction" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.152, + 0.784, + 0.269 + ], + "angle": 0, + "content": "
FramesPseudo Ground TruthAll Frames200 Frames50 Frames
Kinect FusionCOLMAP (default)COLMAP (fast)DROID-SLAM† [90]ACE0 (ours)KF+ACE0 (ours)BARF [54]NoPE-NeRF† [10]ACE0 (ours)DUS3R [97]ACE0 (Ours)
Chess6k19.623.623.519.323.323.012.812.622.718.919.2
Fire4k19.222.622.613.022.322.312.711.822.118.819.5
Heads2k17.018.818.917.618.819.110.711.819.918.421.3
Office10k18.921.421.6failed21.121.511.910.919.812.513.7
Pumpkin6k19.924.123.818.324.123.819.614.224.721.722.3
RedKitchen12k17.621.421.410.920.820.911.611.218.913.813.7
Stairs3k19.016.721.013.017.719.915.815.918.815.315.4
Average18.721.221.8N/A21.221.513.612.621.017.117.9
Avg. Timerealtime38h13h18min1h7min8.5h47h27min4min*16min
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.216, + 0.27, + 0.787, + 0.352 + ], + "angle": 0, + "content": "Table 1: 7-Scenes. We show the pose accuracy via view synthesis with Nerfacto [89] as PSNR in dB, and the reconstruction time. Results for All Frames are color coded w.r.t. similarity to the COLMAP pseudo ground truth: \\(>0.5\\) dB better within \\(\\pm 0.5\\) dB \\(>0.5\\) dB worse \\(>1\\) dB worse. For some competitors, we had to sub-sample the images due to their computational complexity (right side). Method needs sequential inputs. *Results on more powerful hardware." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.394, + 0.333, + 0.408 + ], + "angle": 0, + "content": "4.1 7-Scenes" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.417, + 0.787, + 0.538 + ], + "angle": 0, + "content": "The 7-Scenes dataset [81] consists of seven indoor scenes, scanned with a Kinect v1 camera. Multiple, disconnected scans are provided for each scene to a total of \\(2\\mathrm{k} - 12\\mathrm{k}\\) images. For each method, we assume a shared focal length across scans and initialize with the default calibration of a Kinect v1. The dataset comes with pseudo ground truth camera poses estimated by KinectFusion [44, 64], a depth-based SLAM system. Individual scans were registered but not bundle-adjusted [13]. Inspired by [13], we recompute alternative, bundle-adjusted pseudo ground truth by running COLMAP with default parameters." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.539, + 0.787, + 0.689 + ], + "angle": 0, + "content": "Discussion. We show results in Table 1. Of both pseudo ground truth versions, KinectFusion achieves lower PSNR numbers than COLMAP, presumably due to the lack of global optimization. COLMAP with fast parameters shows PSNR numbers similar to COLMAP with default parameters, on average. Both versions of running COLMAP take considerable time to reconstruct each scene. We note that COLMAP has been optimised for quality, rather than speed. Not all acceleration strategies from the feature-based SfM literature have been implemented in COLMAP, so presumably comparable quality can be obtained faster. DROID-SLAM [90] does not perform well on 7-Scenes and partially fails altogether, presumably due to the jumps between individual scans of each scene." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Our approach, ACE0, achieves a pose quality comparable to the COLMAP pseudo ground truth while reconstructing each scene in \\(\\sim 1\\) hour despite the large number of images. We show qualitative examples in Figure 4 and in the supplement. We also demonstrate that ACE0 can swiftly optimize an initial set of approximate poses. When starting from KinectFusion poses, ACE0 increases PSNR significantly in less than 10 minutes per scene, see “\\(\\mathrm{KF} + \\mathrm{ACE0}\\)” in Table 1. In the supplement, we include a parameter study on 7-Scenes to show that ACE0 is robust to the choice of depth estimator. We also show the positive impact of pose refinement on the reconstruction quality as well as the reconstruction speedup due to our early stopping schedule." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.409, + 0.128 + ], + "angle": 0, + "content": "E. Brachmann et al." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.144, + 0.455, + 0.242 + ], + "angle": 0, + "content": "
ACE [12]ACE [12]ACE0 (ours)
SupervisionKinectFusionCOLMAPself-supervised
Chess96.0%100.0%100.0%
Fire98.4%99.5%98.8%
Heads100.0%100.0%100.0%
Office36.9%100.0%99.1%
Pumpkin47.3%100.0%99.9%
Redlitchen47.8%98.9%98.1%
Stairs74.1%85.0%61.0%
Average71.5%97.6%93.8%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.243, + 0.454, + 0.275 + ], + "angle": 0, + "content": "Table 2 (a): Relocalization on 7-Scenes. % poses below \\(5\\mathrm{cm},5^{\\circ}\\) error, computed w.r.t. COLMAP pseudo GT." + }, + { + "type": "table", + "bbox": [ + 0.46, + 0.144, + 0.788, + 0.242 + ], + "angle": 0, + "content": "
Pseudo GT (COLMAP)DROID-SLAM† [90]BARF [54]NoPe-NeRF† [10]ACEO (ours)
Bicycle21.510.911.912.218.7
Bonsai27.610.912.514.825.8
Counter25.512.911.911.624.5
Garden26.316.713.313.825.0
Kitchen27.413.913.314.426.1
Room28.011.311.914.319.8
Stump16.813.915.013.720.5
Average24.712.912.813.522.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.465, + 0.243, + 0.782, + 0.276 + ], + "angle": 0, + "content": "Table 2 (b): Mip-NeRF 360. Pose quality in PSNR, higher is better. Best in bold. Method needs sequential inputs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.315, + 0.785, + 0.466 + ], + "angle": 0, + "content": "For our learning-based competitors, we sub-sampled images due to their computational constraints, see right side of Table 1. Even using only 200 images, NoPe-NeRF [10] takes 2 days to fit a model and estimate poses. Despite these long processing times, we observe poor pose quality of BARF and NoPe-NeRF in terms of PSNR. BARF [54] requires pose initialisation. We provide identity poses since the scenes of 7Scenes are roughly forward-facing. Still, the camera motions are too complex for BARF to handle. NoPe-NeRF does not require pose initialisation but needs roughly sequential images, which we did provide. NoPe-NeRF relies upon successive images having similar poses to perform well, and struggles with the large jumps between the subsampled images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.467, + 0.787, + 0.526 + ], + "angle": 0, + "content": "For the comparison with DUSt3R [97], we had to subsample the sequences further, as we were only able to run it with 50 images at most, even when using an A100 GPU with 40GB memory. DUSt3R achieves reasonable PSNR numbers but consistently lower than ACE0." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.527, + 0.788, + 0.709 + ], + "angle": 0, + "content": "Relocalization. ACE0 is a learning-based SfM tool but it is also a self-supervised visual relocaliser. In Table 2 (a), we compare it to the supervised relocalizer ACE [12]. Using the scale-metric pseudo ground truth of [13], we train ACE with COLMAP mapping poses, and evaluate it against COLMAP query poses. Unsurprisingly, ACE achieves almost perfect relocalization under the usual \\(5\\mathrm{cm}\\), \\(5^{\\circ}\\) error threshold. Interestingly, ACE0 achieves almost identical results when mapping the scene self-supervised, and evaluating the relocalized query poses against the COLMAP pseudo ground truth. For context, when training ACE with Kinect-Fusion mapping poses and evaluating against COLMAP pseudo ground truth, results are far worse. This signifies that ACE0 mapping poses are very similar to the COLMAP mapping poses, and less similar to KinectFusion mapping poses. We give more details about this experiment in the supplement." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.727, + 0.388, + 0.743 + ], + "angle": 0, + "content": "4.2 Mip-NeRF 360" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.749, + 0.785, + 0.81 + ], + "angle": 0, + "content": "The Mip-NeRF 360 dataset [5] consists of seven small-scale scenes, both indoor and outdoor. The dataset was reconstructed with COLMAP and comes with intrinsics (which we ignore), camera poses and undistorted (pinhole camera) images. For each method, we assume a shared focal length per scene." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Discussion. We present PSNR results in Table 2 (b). NoPe-NeRF does not perform well on this dataset, despite processing each scene for 2 days. The differences" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.509, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Scene Coordinate Reconstruction" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.153, + 0.784, + 0.439 + ], + "angle": 0, + "content": "
FramesDROID-FramesDROID-SparseSparse
COLMAP (default)Reality CaptureSLAM† [90]ACE0 (ours)COLMAP (fast)Reality CaptureSLAM† [90]ACE0 (ours)COLMAP + Reloc + BACOLMAP + ACE0 (ours)
TrainingBarn41024.021.219.016.512.2k24.416.913.517.726.325.1
Caterpillar38317.115.916.616.911.4k18.617.918.918.618.718.8
Church50718.317.614.317.219.3k12.1-11.516.518.517.3
Ignatius26420.117.717.819.87.8k20.818.619.120.720.920.7
Meetingroom37118.618.115.618.011.1k19.418.217.116.620.820.3
Truck25121.119.018.320.17.5k23.619.120.623.023.423.1
Average36419.918.216.918.114.6k19.818.216.818.921.420.9
Avg Time1h3min5min1.1h74h14h18min2.2h8h1.8h
IntermediateFamily15219.518.817.619.04.4k21.219.819.818.021.321.3
Francis30221.620.720.720.17.8k19.920.421.821.722.522.7
Horse15119.219.016.319.56.0k21.620.719.221.722.622.3
Lighthouse30916.616.513.617.58.3k19.016.618.918.619.520.5
Playground30719.119.211.418.77.7k17.916.511.320.421.221.0
Train30116.815.413.816.212.6k19.614.415.618.519.818.5
Average25418.818.315.618.57.8k19.918.117.819.821.121.0
Avg Time32min2min3min1.3h48h11h14min2.2h5h1h
AdvancedAuditorium30219.612.216.718.713.6k13.7-16.620.021.419.8
Ballroom32416.318.313.117.910.8k17.2-10.418.918.015.6
Courtroom30118.217.212.317.112.6k14.6-10.216.318.717.8
Palace50914.211.710.810.721.9k13.8-8.611.015.312.3
Temple30218.115.711.89.717.5k13.3-11.914.819.616.1
Average34817.315.012.914.815.6k14.5-11.516.218.616.3
Avg Time1h2min4min1h71h27min2.8h10h2.1h
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.44, + 0.788, + 0.51 + ], + "angle": 0, + "content": "Table 3: Tanks and Temples. We show the pose accuracy via view synthesis with Nerfacto [89] as PSNR in dB, and the reconstruction time. We color code results compared to COLMAP, default and fast, respectively: \\(>0.5\\) dB better within \\(\\pm 0.5\\) dB \\(>0.5\\) dB worse \\(>1\\) dB worse. Method needs sequential inputs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.551, + 0.788, + 0.645 + ], + "angle": 0, + "content": "between sequential images are too large. DROID-SLAM fails for the same reason. BARF performs poorly because the identity pose is not a good initialisation for most scenes. In contrast, ACE0 reconstructs the dataset successfully. While it achieves slightly lower PSNR than COLMAP, its pose estimates are similar, \\( cf \\). Figure 4. The supp. shows that synthesized images based on ACE0 are visually close to those of COLMAP while our learning-based competitors are far off." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.665, + 0.424, + 0.681 + ], + "angle": 0, + "content": "4.3 Tanks and Temples" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.842 + ], + "angle": 0, + "content": "The Tanks and Temples dataset [49] contains 21 diverse scenes, both indoors and outdoors, and with varying spatial extent. We remove two scenes (Panther, M60) with military associations. We also remove two scenes (Courthouse, Museum) where the COLMAP baseline did not finish the reconstruction after 5 days of processing or ran out of memory. For the latter two scenes, we provide ACE0 results in the supplement. The dataset provides 150-500 images per scene but also the original videos as a source for more frames. Thus, we consider each scene in two versions: Using 150-500 images and using \\(4\\mathrm{k} - 22\\mathrm{k}\\) frames. For all methods, we assume a pinhole camera model with shared focal length across images, and images to be unordered. We found none of the learning-based SfM competitors" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.409, + 0.128 + ], + "angle": 0, + "content": "E. Brachmann et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.787, + 0.192 + ], + "angle": 0, + "content": "to be applicable to this dataset. NoPe-NeRF would run multiple days per scene, even when considering only a few hundred images. DUSt3R would run out of memory. BARF needs reasonable pose initializations which are not available." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.787, + 0.284 + ], + "angle": 0, + "content": "Discussion. We show PSNR numbers of RealityCapture, DROID-SLAM and ACE0 in Table 3, color-coded by similarity to the COLMAP pseudo GT. ACE0 achieves reasonable results when reconstructing scenes from a few hundred images (Table 3, left). ACE0 generally predicts plausible poses (cf., Figure 4), even if PSNR numbers are sometimes lower than those of COLMAP. RealityCapture performs similar to ACE0 while DROID-SLAM struggles on the sparse images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.285, + 0.787, + 0.389 + ], + "angle": 0, + "content": "Next, we consider more than 1k images per scene (right side of Table 3). Here, we run COLMAP with parameters tuned for large images collections (fast) due to the extremely large image sets. ACE0 offers a reconstruction quality comparable to COLMAP on average while also being fast. We run RealityCapture on some of the scenes but it produces fractured reconstructions for these large images sets, leading to low PSNR numbers. DROID-SLAM still struggles on many scenes despite having access to sequential images that are temporally close." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.39, + 0.787, + 0.525 + ], + "angle": 0, + "content": "In the two rightmost columns, we initialise with a sparse COLMAP reconstruction from 150-500 images, and extend and refine it using all available frames. Firstly, using a feature-based baseline, we register the full set of frames using the relocalization mode of COLMAP, followed by a final round of bundle adjustment. Secondly, we run ACE0 initialised with the poses of the sparse COLMAP reconstruction. Both variants are considerably faster than running COLMAP from scratch on the full set of frames. ACE0 is able to register and refine all additional frames in 1-2 hours, on average. Again, we find the pose quality of ACE0 comparable to the feature-based alternative." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.55, + 0.533, + 0.567 + ], + "angle": 0, + "content": "5 Conclusion and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.787, + 0.688 + ], + "angle": 0, + "content": "We have presented scene coordinate reconstruction, a new approach to learning-based SfM. We learn an implicit, neural scene representation from a set of unposed images. Our method, ACE0, is able to reconstruct a wide variety of scenes. In many cases, the accuracy of estimated poses is close to that of COLMAP and synthesized images visually similar. Unlike previous learning-based SfM methods, ACE0 can be applied to multiple thousand unsorted images, without pose priors, and reconstructs them within a few hours." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Limitations. We show some failure cases in the supplement. Scene coordinate regression struggles with repetitive structures since the network is not able to make multi-modal predictions for visually ambiguous inputs. Scene coordinate regression also struggles with representing large areas. The common solution is to use network ensembles based on pre-clustering of the scene [12, 16] which is difficult in a reconstruction setting. While scene coordinate regression generalises quite well, it has difficulties to bridge extreme view point or lighting changes, such as day versus night. In our experiments, we assumed a simple pinhole camera model with shared intrinsics across images. To the best of our knowledge, scene coordinate regression has not been coupled with image distortion, thus far." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.509, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Scene Coordinate Reconstruction" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.177, + 0.784, + 0.204 + ], + "angle": 0, + "content": "1. Agarwal, S., Furukawa, Y., Snavely, N., Simon, I., Curless, B., Seitz, S.M., Szeliski, R.: Building Rome in a day. ACM TOG (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.205, + 0.785, + 0.232 + ], + "angle": 0, + "content": "2. Agarwal, S., Snavely, N., Seitz, S.M., Szeliski, R.: Bundle adjustment in the large. In: ECCV (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.233, + 0.785, + 0.273 + ], + "angle": 0, + "content": "3. Arnold, E., Wynn, J., Vicente, S., Garcia-Hernando, G., Monszpart, A., Prisacariu, V.A., Turmukhambetov, D., Brachmann, E.: Map-free visual relocalization: Metric pose relative to a single image. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.274, + 0.785, + 0.3 + ], + "angle": 0, + "content": "4. Balntas, V., Li, S., Prisacariu, V.A.: RelocNet: Continuous metric learning relocalisation using neural nets. In: ECCV (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.301, + 0.785, + 0.328 + ], + "angle": 0, + "content": "5. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-NeRF 360: Unbounded anti-aliased neural radiance fields. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.33, + 0.785, + 0.355 + ], + "angle": 0, + "content": "6. Beardsley, P.A., Zisserman, A., Murray, D.W.: Sequential updating of projective and affine structure from motion. IJCV (1997)" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.357, + 0.785, + 0.384 + ], + "angle": 0, + "content": "7. Bhat, S.F., Birkl, R., Wofk, D., Wonka, P., Müller, M.: ZoeDepth: Zero-shot transfer by combining relative and metric depth. arXiv (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.385, + 0.785, + 0.424 + ], + "angle": 0, + "content": "8. Bhowmick, B., Patra, S., Chatterjee, A., Govindu, V.M., Banerjee, S.: Divide and conquer: Efficient large-scale structure from motion using graph partitioning. In: ACCV (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.426, + 0.785, + 0.467 + ], + "angle": 0, + "content": "9. Bhowmick, B., Patra, S., Chatterjee, A., Govindu, V.M., Banerjee, S.: Divide and conquer: A hierarchical approach to large-scale structure-from-motion. CVIU (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.468, + 0.785, + 0.494 + ], + "angle": 0, + "content": "10. Bian, W., Wang, Z., Li, K., Bian, J.W., Prisacariu, V.A.: NoPe-NeRF: Optimising neural radiance field with no pose prior. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.496, + 0.785, + 0.535 + ], + "angle": 0, + "content": "11. Bloesch, M., Czarnowski, J., Clark, R., Leutenegger, S., Davison, A.J.: CodeSLAM — Learning a compact, optimisable representation for dense visual SLAM. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.537, + 0.785, + 0.563 + ], + "angle": 0, + "content": "12. Brachmann, E., Cavallari, T., Prisacariu, V.A.: Accelerated coordinate encoding: Learning to relocalize in minutes using RGB and poses. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.565, + 0.785, + 0.591 + ], + "angle": 0, + "content": "13. Brachmann, E., Humenberger, M., Rother, C., Sattler, T.: On the limits of pseudo ground truth in visual camera re-localisation. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.593, + 0.785, + 0.632 + ], + "angle": 0, + "content": "14. Brachmann, E., Krull, A., Nowozin, S., Shotton, J., Michel, F., Gumhold, S., Rother, C.: DSAC-differentiable RANSAC for camera localization. In: CVPR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.634, + 0.785, + 0.66 + ], + "angle": 0, + "content": "15. Brachmann, E., Rother, C.: Learning less is more-6D camera localization via 3D surface regression. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.661, + 0.785, + 0.687 + ], + "angle": 0, + "content": "16. Brachmann, E., Rother, C.: Expert sample consensus applied to camera relocalization. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.689, + 0.785, + 0.715 + ], + "angle": 0, + "content": "17. Brachmann, E., Rother, C.: Visual camera re-localization from RGB and RGB-D images using DSAC. IEEE TPAMI (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.717, + 0.785, + 0.743 + ], + "angle": 0, + "content": "18. Brégier, R.: Deep regression on manifolds: a 3D rotation case study. In: 3DV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.745, + 0.785, + 0.77 + ], + "angle": 0, + "content": "19. Brown, D.: The bundle adjustment-progress and prospect. In: Congr. of the Int. Soc. for Photogr. (1976)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.785, + 0.798 + ], + "angle": 0, + "content": "20. Brown, M., Lowe, D.G.: Unsupervised 3D object recognition and reconstruction in unordered datasets. In: 3DIM (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "21. Carlone, L., Tron, R., Daniilidis, K., Dellaert, F.: Initialization techniques for 3D SLAM: A survey on rotation estimation and its use in pose graph optimization. In: ICRA (2015)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.409, + 0.128 + ], + "angle": 0, + "content": "E. Brachmann et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "22. Cavallari, T., Bertinetto, L., Mukhoti, J., Torr, P.H., Golodetz, S.: Let's take this online: Adapting scene coordinate regression network predictions for online RGB-D camera relocalisation. In: 3DV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "23. Cavallari, T., Golodetz, S., Lord, N.A., Valentin, J., Di Stefano, L., Torr, P.H.: On-the-fly adaptation of regression forests for online camera relocalisation. In: CVPR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.233, + 0.787, + 0.26 + ], + "angle": 0, + "content": "24. Chen, S., Bhalgat, Y., Li, X., Bian, J., Li, K., Wang, Z., Prisacariu, V.A.: Neural refinement for absolute pose regression with feature synthesis. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.261, + 0.787, + 0.288 + ], + "angle": 0, + "content": "25. Chen, S., Li, X., Wang, Z., Prisacariu, V.: DFNet: Enhance absolute pose regression with direct feature matching. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.289, + 0.787, + 0.317 + ], + "angle": 0, + "content": "26. Chen, S., Wang, Z., Prisacariu, V.: Direct-PoseNet: Absolute pose regression with photometric consistency. In: 3DV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.318, + 0.787, + 0.358 + ], + "angle": 0, + "content": "27. Cheng, Z., Esteves, C., Jampani, V., Kar, A., Maji, S., Makadia, A.: LU-NeRF: Scene and pose estimation by synchronizing local unposed NeRFs. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.36, + 0.787, + 0.387 + ], + "angle": 0, + "content": "28. Crandall, D., Owens, A., Snavely, N., Huttenlocher, D.: Discrete-continuous optimization for large-scale structure from motion. In: CVPR (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.388, + 0.787, + 0.416 + ], + "angle": 0, + "content": "29. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: ScanNet: Richly-annotated 3D reconstructions of indoor scenes. In: CVPR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.417, + 0.787, + 0.444 + ], + "angle": 0, + "content": "30. Davison, A.J.: Real-time simultaneous localisation and mapping with a single camera. In: ICCV (2003)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.445, + 0.787, + 0.472 + ], + "angle": 0, + "content": "31. DeTone, D., Malisiewicz, T., Rabinovich, A.: SuperPoint: Self-supervised interest point detection and description. In: CVPRW (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.473, + 0.787, + 0.501 + ], + "angle": 0, + "content": "32. Ding, M., Wang, Z., Sun, J., Shi, J., Luo, P.: CamNet: Coarse-to-fine retrieval for camera re-localization. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.502, + 0.787, + 0.543 + ], + "angle": 0, + "content": "33. Dusmanu, M., Rocco, I., Pajdla, T., Pollefeys, M., Sivic, J., Torii, A., Sattler, T.: D2-net: A trainable CNN for joint description and detection of local features. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.544, + 0.787, + 0.584 + ], + "angle": 0, + "content": "34. Fischler, M.A., Bolles, R.C.: Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Comm. of the ACM (1981)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.586, + 0.787, + 0.613 + ], + "angle": 0, + "content": "35. Gao, X.S., Hou, X.R., Tang, J., Cheng, H.F.: Complete solution classification for the perspective-three-point problem. IEEE TPAMI (2003)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.615, + 0.787, + 0.642 + ], + "angle": 0, + "content": "36. Gherardi, R., Farenzena, M., Fusiello, A.: Improving the efficiency of hierarchical structure-and-motion. In: CVPR (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.643, + 0.787, + 0.669 + ], + "angle": 0, + "content": "37. Govindu, V.M.: Combining two-view constraints for motion estimation. In: CVPR (2001)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.671, + 0.787, + 0.698 + ], + "angle": 0, + "content": "38. Govindu, V.M.: Lie-algebraic averaging for globally consistent motion estimation. In: CVPR (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.7, + 0.738, + 0.713 + ], + "angle": 0, + "content": "39. Hartley, R., Trumpf, J., Dai, Y., Li, H.: Rotation averaging. IJCV (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.715, + 0.787, + 0.741 + ], + "angle": 0, + "content": "40. Hartley, R., Zisserman, A.: Multiple view geometry in computer vision. Cambridge University Press (2003)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.742, + 0.787, + 0.769 + ], + "angle": 0, + "content": "41. He, X., Sun, J., Wang, Y., Peng, S., Huang, Q., Bao, H., Zhou, X.: Detector-free structure from motion. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.77, + 0.787, + 0.798 + ], + "angle": 0, + "content": "42. Heinly, J., Schonberger, J.L., Dunn, E., Frahm, J.M.: Reconstructing the World in six days. In: CVPR (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "43. Humenberger, M., Cabon, Y., Pion, N., Weinzaepfel, P., Lee, D., Guérin, N., Sattler, T., Csurka, G.: Investigating the role of image retrieval for visual localization: An exhaustive benchmark. IJCV (2022)" + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.509, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Scene Coordinate Reconstruction" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.147, + 0.786, + 0.203 + ], + "angle": 0, + "content": "44. Izadi, S., Kim, D., Hilliges, O., Molyneaux, D., Newcombe, R., Kohli, P., Shotton, J., Hodges, S., Freeman, D., Davison, A.J., Fitzgibbon, A.: KinectFusion: real-time 3D reconstruction and interaction using a moving depth camera. In: UIST (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.204, + 0.786, + 0.23 + ], + "angle": 0, + "content": "45. Jeong, Y., Ahn, S., Choy, C., Anandkumar, A., Cho, M., Park, J.: Self-calibrating neural radiance fields. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.23, + 0.786, + 0.258 + ], + "angle": 0, + "content": "46. Jin, Y., Mishkin, D., Mishchuk, A., Matas, J., Fua, P., Yi, K.M., Trulls, E.: Image matching across wide baselines: From paper to practice. IJCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.258, + 0.786, + 0.284 + ], + "angle": 0, + "content": "47. Kendall, A., Grimes, M., Cipolla, R.: PoseNet: A convolutional network for real-time 6-DoF camera relocalization. In: ICCV (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.284, + 0.786, + 0.311 + ], + "angle": 0, + "content": "48. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3D gaussian splatting for real-time radiance field rendering. ACM TOG (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.311, + 0.786, + 0.338 + ], + "angle": 0, + "content": "49. Knapitsch, A., Park, J., Zhou, Q.Y., Koltun, V.: Tanks and Temples: Benchmarking large-scale scene reconstruction. ACM TOG (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.339, + 0.786, + 0.366 + ], + "angle": 0, + "content": "50. Kraus, K.: Photogrammetry. No. v. 1 in Photogrammetry, Ferdinand Dummlers Verlag (1993)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.366, + 0.786, + 0.406 + ], + "angle": 0, + "content": "51. Laskar, Z., Melekhov, I., Kalia, S., Kannala, J.: Camera relocalization by computing pairwise relative poses using convolutional neural network. In: ICCV Workshops (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.406, + 0.786, + 0.433 + ], + "angle": 0, + "content": "52. Li, X., Wang, S., Zhao, Y., Verbeek, J., Kannala, J.: Hierarchical scene coordinate classification and regression for visual localization. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.433, + 0.786, + 0.461 + ], + "angle": 0, + "content": "53. Lin, A., Zhang, J.Y., Ramanan, D., Tulsiani, S.: Relpose++: Recovering 6d poses from sparse-view observations. In: 3DV (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.461, + 0.786, + 0.487 + ], + "angle": 0, + "content": "54. Lin, C.H., Ma, W.C., Torralba, A., Lucey, S.: BARF: Bundle-adjusting neural radiance fields. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.487, + 0.786, + 0.528 + ], + "angle": 0, + "content": "55. Lin, Y., Müller, T., Tremblay, J., Wen, B., Tyree, S., Evans, A., Vela, P.A., Birchfield, S.: Parallel inversion of neural radiance fields for robust pose estimation. In: ICRA (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.528, + 0.786, + 0.555 + ], + "angle": 0, + "content": "56. Lindenberger, P., Sarlin, P.E., Pollefeys, M.: LightGlue: Local feature matching at light speed. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.555, + 0.786, + 0.569 + ], + "angle": 0, + "content": "57. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: ICLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.569, + 0.786, + 0.596 + ], + "angle": 0, + "content": "58. Lowe, D.G.: Distinctive image features from scale-invariant keypoints. IJCV (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.596, + 0.786, + 0.623 + ], + "angle": 0, + "content": "59. Martinec, D., Pajdla, T.: Robust rotation and translation estimation in multiview reconstruction. In: CVPR (2007)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.623, + 0.786, + 0.65 + ], + "angle": 0, + "content": "60. Meng, Q., Chen, A., Luo, H., Wu, M., Su, H., Xu, L., He, X., Yu, J.: GNeRF: GAN-based Neural Radiance Field without Posed Camera. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.65, + 0.786, + 0.691 + ], + "angle": 0, + "content": "61. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.691, + 0.786, + 0.732 + ], + "angle": 0, + "content": "62. Moreau, A., Piasco, N., Bennehar, M., Tsishkou, D., Stanciulescu, B., de La Fortelle, A.: CROSSFIRE: Camera relocalization on self-supervised features from an implicit representation. ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.732, + 0.786, + 0.759 + ], + "angle": 0, + "content": "63. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM TOG (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.759, + 0.786, + 0.8 + ], + "angle": 0, + "content": "64. Newcombe, R., Izadi, S., Hilliges, O., Molyneaux, D., Kim, D., Davison, A.J., Kohi, P., Shotton, J., Hodges, S., Fitzgibbon, A.: KinectFusion: Real-time dense surface mapping and tracking. In: ISMAR (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.742, + 0.813 + ], + "angle": 0, + "content": "65. Nistér, D., Naroditsky, O., Bergen, J.: Visual odometry. In: CVPR (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.786, + 0.84 + ], + "angle": 0, + "content": "66. Pollefeys, M., Koch, R., Vergauwen, M., Van Gool, L.: Automated reconstruction of 3D scenes from sequences of images. J. of Photogr. and Rem. Sens. (2000)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.147, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.409, + 0.128 + ], + "angle": 0, + "content": "E. Brachmann et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "67. Rau, A., Garcia-Hernando, G., Stoyanov, D., Brostow, G.J., Turmukhambetov, D.: Predicting visual overlap of images through interpretable non-metric box embeddings. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.19, + 0.785, + 0.218 + ], + "angle": 0, + "content": "68. Reality, C.: Reality Capture. https://www.capturingreality.com/ realitycapture (2016), [accessed 15-Nov-2023]" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.218, + 0.785, + 0.259 + ], + "angle": 0, + "content": "69. Reizenstein, J., Shapovalov, R., Henzler, P., Sbordone, L., Labatut, P., Novotny, D.: Common objects in 3D: Large-scale learning and evaluation of real-life 3D category reconstruction. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.259, + 0.785, + 0.286 + ], + "angle": 0, + "content": "70. Sarlin, P.E., Cadena, C., Siegwart, R., Dymczyk, M.: From coarse to fine: Robust hierarchical localization at large scale. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.287, + 0.785, + 0.314 + ], + "angle": 0, + "content": "71. Sarlin, P.E., DeTone, D., Malisiewicz, T., Rabinovich, A.: SuperGlue: Learning feature matching with graph neural networks. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.315, + 0.785, + 0.357 + ], + "angle": 0, + "content": "72. Sarlin, P.E., Unagar, A., Larsson, M., Germain, H., Toft, C., Larsson, V., Pollefeys, M., Lepetit, V., Hammarstrand, L., Kahl, F., Sattler, T.: Back to the feature: Learning robust camera localization from pixels to pose. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.357, + 0.785, + 0.383 + ], + "angle": 0, + "content": "73. Sattler, T., Leibe, B., Kobbelt, L.: Fast image-based localization using direct 2D-to-3D matching. In: ICCV (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.383, + 0.785, + 0.411 + ], + "angle": 0, + "content": "74. Sattler, T., Leibe, B., Kobbelt, L.: Improving image-based localization by active correspondence search. In: ECCV (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.411, + 0.785, + 0.438 + ], + "angle": 0, + "content": "75. Sattler, T., Leibe, B., Kobbelt, L.: Efficient & effective prioritized matching for large-scale image-based localization. IEEE TPAMI (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.439, + 0.785, + 0.481 + ], + "angle": 0, + "content": "76. Sattler, T., Torii, A., Sivic, J., Pollefeys, M., Taira, H., Okutomi, M., Pajdla, T.: Are large-scale 3D models really necessary for accurate visual localization? In: CVPR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.481, + 0.785, + 0.509 + ], + "angle": 0, + "content": "77. Sattler, T., Zhou, Q., Pollefeys, M., Leal-Taixe, L.: Understanding the limitations of CNN-based absolute camera pose regression. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.509, + 0.785, + 0.536 + ], + "angle": 0, + "content": "78. Schaffalitzky, F., Zisserman, A.: Multi-view matching for unordered image sets, or \"How do i organize my holiday snaps?\". In: ECCV (2002)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.536, + 0.785, + 0.564 + ], + "angle": 0, + "content": "79. Schonberger, J.L.: Colmap Github Issues. https://github.com/colmap/colmap/issues/116#issuecomment-298926277 (2017), [accessed Nov/15/2023]" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.564, + 0.785, + 0.592 + ], + "angle": 0, + "content": "80. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: CVPR (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.592, + 0.785, + 0.632 + ], + "angle": 0, + "content": "81. Shotton, J., Glocker, B., Zach, C., Izadi, S., Criminisi, A., Fitzgibbon, A.: Scene coordinate regression forests for camera relocalization in RGB-D images. In: CVPR (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.632, + 0.785, + 0.661 + ], + "angle": 0, + "content": "82. Sinha, S., Zhang, J.Y., Tagliasacchi, A., Gilitschenski, I., Lindell, D.B.: Sparse-Pose: Sparse-view camera pose regression and refinement. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.661, + 0.785, + 0.702 + ], + "angle": 0, + "content": "83. Smith, L.N., Topin, N.: Super-convergence: Very fast training of neural networks using large learning rates. In: Arti. Intell. and Mach. Learn. for Multi-Domain Operations Appl. (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.702, + 0.785, + 0.73 + ], + "angle": 0, + "content": "84. Snavely, N., Seitz, S.M., Szeliski, R.: Photo tourism: exploring photo collections in 3D. ACM TOG (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.73, + 0.785, + 0.757 + ], + "angle": 0, + "content": "85. Snavely, N., Seitz, S.M., Szeliski, R.: Modeling the world from internet photo collections. IJCV (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.758, + 0.785, + 0.785 + ], + "angle": 0, + "content": "86. Snavely, N., Seitz, S.M., Szeliski, R.: Skeletal graphs for efficient structure from motion. In: CVPR (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.785, + 0.813 + ], + "angle": 0, + "content": "87. Sun, J., Shen, Z., Wang, Y., Bao, H., Zhou, X.: LoFTR: Detector-free local feature matching with transformers. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.785, + 0.841 + ], + "angle": 0, + "content": "88. Szeliski, R., Kang, S.B.: Recovering 3D shape and motion from image streams using nonlinear least squares. J. of Vis. Com. and Image Repr. (1994)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.509, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Scene Coordinate Reconstruction" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "89. Tancik, M., Weber, E., Ng, E., Li, R., Yi, B., Kerr, J., Wang, T., Kristoffersen, A., Austin, J., Salahi, K., Ahuja, A., McAllister, D., Kanazawa, A.: Nerfstudio: A modular framework for neural radiance field development. In: ACM TOG (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.19, + 0.785, + 0.217 + ], + "angle": 0, + "content": "90. Teed, Z., Deng, J.: DROID-SLAM: Deep visual SLAM for monocular, stereo, and RGB-D cameras. In: NeurIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.218, + 0.785, + 0.244 + ], + "angle": 0, + "content": "91. Toldo, R., Gherardi, R., Farenzena, M., Fusiello, A.: Hierarchical structure-and-motion recovery from uncalibrated images. CVIU (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.245, + 0.785, + 0.272 + ], + "angle": 0, + "content": "92. Triggs, B., McLauchlan, P.F., Hartley, R.I., Fitzgibbon, A.W.: Bundle adjustment—a modern synthesis. In: Int. Worksh. on Vis. Alg. (2000)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.273, + 0.785, + 0.313 + ], + "angle": 0, + "content": "93. Türkoglu, M.Ö., Brachmann, E., Schindler, K., Brostow, G., Monszpart, A.: Visual camera re-localization using graph neural networks and relative pose supervision. In: 3DV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.315, + 0.764, + 0.328 + ], + "angle": 0, + "content": "94. Ulyanov, D., Vedaldi, A., Lempitsky, V.: Deep image prior. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.329, + 0.785, + 0.369 + ], + "angle": 0, + "content": "95. Ummenhofer, B., Zhou, H., Uhrig, J., Mayer, N., Ilg, E., Dosovitskiy, A., Brox, T.: DeMoN: Depth and motion network for learning monocular stereo. In: CVPR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.37, + 0.785, + 0.411 + ], + "angle": 0, + "content": "96. Waechter, M., Beljan, M., Fuhrmann, S., Moehrle, N., Kopf, J., Goesele, M.: Virtual rephotography: Novel view prediction error for 3d reconstruction. ACM TOG (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.412, + 0.785, + 0.438 + ], + "angle": 0, + "content": "97. Wang, S., Leroy, V., Cabon, Y., Chidlovskii, B., Revaud, J.: DUSt3R: Geometric 3D vision made easy. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.439, + 0.785, + 0.466 + ], + "angle": 0, + "content": "98. Wang, Z., Wu, S., Xie, W., Chen, M., Prisacariu, V.A.: NeRF--: Neural radiance fields without known camera parameters. arXiv (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.467, + 0.785, + 0.494 + ], + "angle": 0, + "content": "99. Wei, X., Zhang, Y., Li, Z., Fu, Y., Xue, X.: DeepSFM: Structure from motion via deep bundle adjustment. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.495, + 0.785, + 0.508 + ], + "angle": 0, + "content": "00. Wu, C.: Towards linear-time incremental structure from motion. In: 3DV (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.509, + 0.785, + 0.535 + ], + "angle": 0, + "content": "01. Xia, Y., Tang, H., Timofte, R., Van Gool, L.: SiNeRF: Sinusoidal neural radiance fields for joint pose estimation and scene reconstruction. In: BMVC (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.536, + 0.785, + 0.563 + ], + "angle": 0, + "content": "02. Yen-Chen, L., Florence, P., Barron, J.T., Rodriguez, A., Isola, P., Lin, T.Y.: iNeRF: Inverting neural radiance fields for pose estimation. In: IROS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.565, + 0.785, + 0.591 + ], + "angle": 0, + "content": "03. Zhang, J.Y., Lin, A., Kumar, M., Yang, T.H., Ramanan, D., Tulsiani, S.: Cameras as rays: Pose estimation via ray diffusion. In: ICLR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.592, + 0.785, + 0.618 + ], + "angle": 0, + "content": "04. Zhang, W., Kosecka, J.: Image based localization in urban environments. In: 3DPVT (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.619, + 0.785, + 0.646 + ], + "angle": 0, + "content": "05. Zhou, Q., Sattler, T., Pollefeys, M., Leal-Taixe, L.: To learn or not to learn: Visual localization from essential matrices. In: ICRA (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.647, + 0.785, + 0.674 + ], + "angle": 0, + "content": "06. Zhou, Y., Barnes, C., Jingwan, L., Jimei, Y., Hao, L.: On the continuity of rotation representations in neural networks. In: CVPR (2019)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.674 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Scene Coordinate Reconstruction_ Posing of Image Collections via Incremental Learning of a Relocalizer/c513fe61-35e0-435f-acdb-dbf8715686b2_origin.pdf b/2024/Scene Coordinate Reconstruction_ Posing of Image Collections via Incremental Learning of a Relocalizer/c513fe61-35e0-435f-acdb-dbf8715686b2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..cdfcb8112a7168829522fcdd0877932489a21346 --- /dev/null +++ b/2024/Scene Coordinate Reconstruction_ Posing of Image Collections via Incremental Learning of a Relocalizer/c513fe61-35e0-435f-acdb-dbf8715686b2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0fd60d953638f93873b6f771934f6b2871f294f651ce950e7504b5f6dd66fcda +size 9741733 diff --git a/2024/Scene Coordinate Reconstruction_ Posing of Image Collections via Incremental Learning of a Relocalizer/full.md b/2024/Scene Coordinate Reconstruction_ Posing of Image Collections via Incremental Learning of a Relocalizer/full.md new file mode 100644 index 0000000000000000000000000000000000000000..cd0312ccbfdedf3229e1da54c61c5c573b753684 --- /dev/null +++ b/2024/Scene Coordinate Reconstruction_ Posing of Image Collections via Incremental Learning of a Relocalizer/full.md @@ -0,0 +1,328 @@ +# Scene Coordinate Reconstruction: Posing of Image Collections via Incremental Learning of a Relocalizer + +Eric Brachmann1, Jamie Wynn1, Shuai Chen2, Tommaso Cavallari1, Åron Monszpart1, Daniyar Turmukhambetov1, and Victor Adrian Prisacariu1,2 + +$^{1}$ Niantic $^{2}$ University of Oxford + +Abstract. We address the task of estimating camera parameters from a set of images depicting a scene. Popular feature-based structure-from-motion (SfM) tools solve this task by incremental reconstruction: they repeat triangulation of sparse 3D points and registration of more camera views to the sparse point cloud. We re-interpret incremental structure-from-motion as an iterated application and refinement of a visual relocalizer, that is, of a method that registers new views to the current state of the reconstruction. This perspective allows us to investigate alternative visual relocalizers that are not rooted in local feature matching. We show that scene coordinate regression, a learning-based relocalization approach, allows us to build implicit, neural scene representations from unposed images. Different from other learning-based reconstruction methods, we do not require pose priors nor sequential inputs, and we optimize efficiently over thousands of images. In many cases, our method, ACE0, estimates camera poses with an accuracy close to feature-based SfM, as demonstrated by novel view synthesis. + +Project page: https://nianticlabs.github.io/acezero/ + +# 1 Introduction + +In the beginning there was structure-from-motion. + +The genesis of numerous computer vision tasks lies in the estimation of camera poses and scene geometry from a set of images. It is the first fundamental step that lets us leave the image plane and venture into 3D. Since structure-from-motion (SfM) is such a central capability, we have researched it for decades. By now, refined open-source tools, such as COLMAP [80], and efficient commercial packages, such as RealityCapture [68], are available to us. Feature matching-based SfM is the gold standard for estimating poses from images, with a precision that makes its estimates occasionally considered "ground truth" [3,5,13,46,69]. + +The success of Neural Radiance Fields (NeRF) [61] has renewed interest in the question of whether SfM can be solved differently, based on neural, implicit scene representations rather than 3D point clouds. There has been some progress in recent years but, thus far, learning-based approaches to camera pose recovery + +![](images/b4aa7e4e1c73fd6aa578d18ede7c9437c374cf0ce131026b0f4af869293101f9.jpg) +Fig. 1: Reconstructing 10,000 Images. Top: Starting from a single image and the identity pose, we train a learning-based visual relocalizer. The relocalizer allows us to estimate the poses of more views, and the additional views allow us to refine the relocalizer. We show three out of six iterations for this scene (7Scenes Office [81]). All 10k images have been posed in roughly 1 hour on a single GPU. In comparison NoPe-NeRF [10] needs two days to pose 200 images. The point cloud is a visualization of the implicit scene representation of the relocalizer. Camera positions are color coded by relocalization confidence from yellow (low) to green (high). Bottom: Point clouds from Nerfacto [89] trained on top of our poses for a few scenes from our experiments. + +still have significant limitations. They either require coarse initial poses [45, 54, 98, 101], prior knowledge of the pose distribution [60] or sequential inputs [10, 11, 90]. In terms of the number of images that learning-based approaches can handle, they are either explicitly targeted at few-frame problems [53, 82, 95, 97, 99, 103] or they are computationally so demanding that they can realistically only be applied to a few hundred images at most [10, 54, 98]. We show that none of these limitations are an inherent consequence of using learning-based scene representations. + +Our approach is inspired by incremental SfM and its relationship to another computer vision task: visual relocalization. Visual relocalization describes the problem of estimating the camera pose of a query image w.r.t. to an existing scene map. Incremental SfM can be re-interpreted as a loop of 1) do visual relocalization to register new views to the reconstruction, and 2) refine/extend the reconstruction using the newly registered views. Local feature matching is a traditional approach to visual relocalization [70,74-76]. In recent years, multiple learning-based relocalizers have been proposed that encode the scene implicitly in the weights of a neural network [3, 12, 14, 15, 17, 22, 47]. Not all of them are suitable for building a SfM pipeline. We need a relocalizer with high accuracy and good generalization. Training of the relocalizer has to be swift. We need to be able to bootstrap relocalization without ground truth poses. And we need to be able to tell whether registration of a new image was successful. + +We show that "scene coordinate regression", an approach to visual relocalization proposed a decade ago [81], has the desirable properties and can serve as + +the core of a new approach to learning-based SfM: Scene coordinate reconstruction. Rather than optimizing over image-to-image matches, like feature-based SfM, scene coordinate reconstruction regresses image-to-scene correspondences directly. Rather than representing the scene as a 3D point cloud with high dimensional descriptors, we encode the scene into a lightweight neural network. Our approach works on unsorted images without pose priors and efficiently optimises over thousands of images, see Figure 1. + +We summarize our contributions: + +- Scene Coordinate Reconstruction, a new approach to SfM based on incremental learning of scene coordinate regression, a visual relocalization principle. +- We turn the fast-learning visual relocalizer ACE [12] into a SfM framework that is able to predict the camera poses of a set of unposed RGB images. We refer to this new SfM pipeline as $ACE0$ (ACE Zero). +- Compared to ACE [12], we add the capability to train in a self-supervised fashion. We start from a single image, and iterate between learning the map and registering new views. We expedite reconstruction times by early stopping, and increase reconstruction quality by pose refinement. + +# 2 Related Work + +Reconstruction. SfM pipelines either ingest a collection of unordered images [20,78,84] or an ordered image sequence [6,30,65,66,88] from a video to recover 3D structure of a scene and camera poses ("motion") of the images. + +SIFT [58] and other robust descriptors allow matching image features across wide baselines enabling systems to reconstruct large-scale scenes using Internet images [80,84,85,100]. Image-to-image matches can also be regressed directly in a detector-free setup [41,87]. Feature tracks across multiple images are built from image-to-image matches. Feature tracks and estimated relative poses are used to solve for the 3D feature coordinates, camera poses and calibrations (intrinsic matrices). This geometric optimization problem is mainly solved using bundle adjustment which was explored in photogrammetry and geodesy [19, 50] and became standard in the computer vision community [40,92]. Bundle adjustment relies on the initialization being close to the solution (i.e., camera poses and 3D points are already mostly accurate). + +There are two main approaches to this problem. Incremental SfM [6, 66] starts the reconstruction from very few images to create a high-quality seed reconstruction that progressively grows by registering more images and refining the reconstruction until convergence. Global SfM approaches solve for "global" poses of all images using estimates of relative poses, i.e., motion averaging [37, 38], rotation averaging [39, 59] and pose-graph optimization [21]. Various techniques were proposed to improve SfM runtime for very large sets of images [1, 2, 8, 9, 28, 36, 42, 86, 91]. Our work is similar to Incremental SfM, as we also progressively register images to the reconstructed scene starting from a seed reconstruction. However, we do not explicitly compute image matches, nor feature tracks across images, which can be computationally expensive. + +Visual Relocalization. A reconstructed (or mapped) scene is a database of images with known camera poses. This database can be used by a visual relocalizer to estimate poses for new query images to "relocalize" a camera in the scene. Feature-based approaches extract 2D local features [31, 33, 56, 58, 71, 72, 87] from a query image and match them to 3D points to solve for the query pose using perspective-n-point (PnP) solvers [35], e.g., [73-75]; or match query features to 2D local features of mapped images to triangulate the query image, e.g., [104, 105]. For large scenes, matching features on a subset of database images relevant to the query can improve the speed and accuracy, e.g., [43, 67, 70, 76]. + +Some learning-based approaches encode the map of a scene in the weights of a neural network. PoseNet [47] directly regresses the absolute camera pose given an input image using a CNN that was trained on image-posed pairs of the reconstructed scene. Sattler et al. show that extrapolating outside the mapped poses can be challenging for absolute pose regression approaches [77]. Relative pose regression networks [4, 32, 51, 93] estimate the relative pose for a pair of images, allowing triangulation of the query image from multiple map images, or estimate a scale-metric relative pose w.r.t. a single map image [3]. Scene coordinate regression [12, 14, 15, 17, 22, 23, 52, 81] directly predicts 3D coordinates of a point given a patch. This approach generalizes well as the camera pose is solved using PnP [35] within a RANSAC loop [34]. ACE [12] shows that training of scene coordinate regression can be greatly accelerated. In our work we train the ACE localizer, but we start from images without poses. + +Image-Based Rendering. In recent years, neural radiance fields (NeRFs) [61] have seen a lot of attention from the community. NeRFs allow photorealistic novel view synthesis when trained from image-pose pairs. Typically, estimation of the image poses is done in advance by using an SfM pipeline such as COLMAP, e.g., [5]. Nonetheless, research exists that estimates camera poses with NeRFs, facilitates camera localization for novel views after NeRF training [24-26, 62, 102], or simultaneously estimates camera poses during NeRF training from images alone [10, 27, 54, 82, 98]. However, these approaches either assume that the scene is captured from the front [45, 98, 101], that coarse poses are available for initialization [54], or that images are captured sequentially [10]. Techniques that rely on a multi-layer perceptron (MLP) representation of the scene, e.g. [10, 27], are slow to train, taking days to converge. While radiance fields can be trained faster using multi-layer hash grids [63] or Gaussian splats [48], their efficacy in pose estimation without approximate prior pose initialization [55, 62] or sequential image inputs [48] remains unproven. Concurrently, several learning-based camera pose estimation methods have been proposed [53, 97, 103]. Due to GPU memory constraints, these methods estimate poses for sparse sets of images. + +# 3 Method + +Preliminaries. The input to our system is a set of RGB images, denoted by $\mathcal{I} = \{I_i\}$ , where $i$ refers to the image index. Our system estimates the corresponding + +![](images/326d8101c386fb5c2375761039db1a5410211ba682da210bdc544e843354ef62.jpg) +Fig. 2: ACE0 Framework. Top left: We loop between learning a reconstruction from the current set of images and poses ("neural mapping"), and estimating poses of more images ("relocalization"). Top right: During the mapping stage, we train a scene coordinate regression network as our scene representation. Camera poses of the last relocalization round and camera calibration parameters are refined during this process. We visualize scene coordinates by mapping XYZ to the RGB cube. Bottom: In the relocalization stage, we re-estimate poses of images using the scene coordinate regression network, including images that were previously not registered to the reconstruction. If the registration of an image succeeds, it will be used in the next iteration of the mapping stage; otherwise it will not. + +set of camera parameters, both intrinsic and extrinsic: $\mathcal{H} = \{(K_i,T_i)\}$ . Each $T_{i}$ refers to a $3\times 4$ matrix containing a rotation and translation, while $K_{i}$ refers to a $3\times 3$ matrix with the calibration parameters. We assume no particular image order or any prior knowledge about the pose distribution. + +We also want to recover the 3D structure of the scene: Each pixel $j$ in image $i$ with 2D pixel position $\mathbf{p}_{ij}$ has a corresponding coordinate in 3D, denoted as $\mathbf{y}_{ij}$ . The 2D pixel positions and 3D scene coordinates are related by the camera pose and the projection function $\pi$ : + +$$ +\mathbf {p} _ {i j} = \boldsymbol {\pi} \left(K _ {i}, T _ {i}, \mathbf {y} _ {i j}\right), \tag {1} +$$ + +where $T_{i}$ maps camera coordinates to scene coordinates, and $K_{i}$ projects camera coordinates to the image plane. + +As our scene representation, we utilize a scene coordinate regression model [81], i.e., a learnable function $f_{\mathrm{SCR}}$ that maps an image patch of image $I_{i}$ , centered around pixel position $\mathbf{p}_{ij}$ to a scene coordinate: $\mathbf{y}_{ij} = f_{\mathrm{SCR}}(\mathbf{p}_{ij},I_i)$ . + +Given a set of 2D-3D correspondences predicted by $f_{\mathrm{SCR}}$ for any image $I_i$ , we can recover this image's camera pose $T_i$ using a pose solver $g$ : + +$$ +T _ {i} = g \left(K _ {i}, \left\{\left(\mathbf {p} _ {i j}, \mathbf {y} _ {i j}\right) \right\}\right). \tag {2} +$$ + +Since 2D-3D correspondences can be inaccurate, and contain incorrect predictions, $g$ combines a PnP solver [35] with a RANSAC loop [34]. + +Normally, scene coordinate regression models are trained in a supervised fashion for the task of visual relocalization [12, 14, 15, 17, 22, 23, 52, 81]. That is, $f_{\mathrm{SCR}}$ is trained using images with known ground truth camera parameters $\{(I_i, T_i^{\mathrm{GT}}, K_i^{\mathrm{GT}})\}$ , and used for estimating the poses of unseen query images. Instead, we show how these models can be trained self-supervised, without ground truth poses, to estimate the poses of the mapping images themselves. Thus, we turn scene coordinate regression into scene coordinate reconstruction, a learning-based SfM tool. + +# 3.1 Neural Mapping + +We train the scene coordinate regression model iteratively where we denote the current time step as $t$ and the corresponding scene model as $f_{\mathrm{SCR}}^t$ . We iterate between training the scene model, and registering new views, see Figure 2. + +At iteration $t$ we assume that a subset of images has already been registered to the scene, $\mathcal{I}_{\mathrm{Reg}}^t \subset \mathcal{I}$ , and where corresponding camera parameters, $T_{i}^{t}$ and $K_{i}^{t}$ , have already been estimated. Using these as pseudo ground truth, we train the scene model by minimizing the pixel-wise reprojection error: + +$$ +\sum_ {I _ {i} \in \mathcal {I} _ {\text {R e g}} ^ {t}} \sum_ {j \in I _ {i}} \| \mathbf {p} _ {i j} - \boldsymbol {\pi} \left(K _ {i} ^ {t}, T _ {i} ^ {t}, \mathbf {y} _ {i j} ^ {t}\right) \|, \tag {3} +$$ + +where the scene model $f_{\mathrm{SCR}}^{t}$ predicts coordinates $\mathbf{y}_{ij}^{t}$ . + +Mapping Framework. We optimize Eq. 3 using stochastic gradient descent, using the fast-learning scene coordinate regressor ACE [12] (Accelerated Coordinate Encoding). ACE is trained in minutes, even for thousands of views. The training speed is important since we have to train the model in multiple iterations. ACE approximates Eq. 3 by sampling random patches from the training set, $\mathcal{I}_{\mathrm{Reg}}^t$ . To do that efficiently, ACE employs a pre-trained encoder to precompute high dimensional features for a large number of training patches. The encoder stays frozen during mapping. The actual scene model, which is trained, is a multi-layer perceptron (MLP) that maps encoder features to scene coordinates; see Fig. 2 (top right) for a visual representation. + +Pose Refinement. Differently from the ACE [12] protocol, the ground truth poses $T_{i}^{\mathrm{GT}}$ are unknown during training. Instead, we have $T_{i}^{t}$ , estimates based on earlier iterations of the reconstruction. Since these estimates can be inaccurate, we add the ability to refine poses during mapping. We implement refinement using an MLP: + +$$ +T _ {i} ^ {t} = f _ {\text {P o s e}} ^ {t} \left(\hat {T} _ {i} ^ {t}\right), \tag {4} +$$ + +where $\tilde{T}_i^t$ denotes the initial pose estimate at the start of a mapping iteration. Inspired by [106], the refinement MLP ingests $\tilde{T}_i^t$ as $3\times 4 = 12$ values and predicts 12 additive offsets. We orthonormalize rotations using Gram-Schmidt [18]. We jointly optimize $f_{\mathrm{Pose}}^t$ and $f_{\mathrm{SCR}}^t$ to minimize the reprojection error of Eq. 3. We + +![](images/e97ec08733e3676a4e5598bec4880fd665bb92cd0d8faad018b525ec29096720.jpg) +Fig. 3: Left: Pose Refinement. Since we register images based on a coarse and incomplete state of the reconstruction, we add the ability to refine poses during neural mapping. An MLP predicts pose updates relative to the initial poses, supervised by the reprojection error of scene coordinates. Right: Initialization. To start the reconstruction, we train the network using one image, the identity pose and a depth estimate, here ZoeDepth [7]. In this example, we register 33 views to the initial reconstruction. Depth estimates are only used for this step. + +![](images/8c96ff0ceff1a5888f81096c6aa623935e3ebd2db8b011fa5891e3ca5993a194.jpg) + +show the impact of pose refinement for one iteration in Fig. 3 (left). We discard the refinement MLP after each mapping iteration. Its purpose is to enable the scene model, $f_{\mathrm{SCR}}^{t}$ , to converge to a consistent scene representation. + +With neither poses nor 2D-3D correspondences fixed in Eq. 3, the scene coordinate regressor could drift or degenerate. As regularization, we optimize the pose refiner $f_{\mathrm{Pose}}^{t}$ using AdamW [57] with weight decay, biasing the MLP to predict small updates relative to the initial estimates $\tilde{T}_i^t$ . This relies on the assumption that mapping images in $\mathcal{I}_{\mathrm{Reg}}^t$ have been registered close to their true position. If that assumption holds, the smoothness prior of the networks [94] encourages a multi-view consistent solution as shown in previous work [12,15,17]. + +As an alternative to an MLP refiner, we could back-propagate directly to the input poses [98, 102]. However, when optimizing over thousands of views, the signal for each single pose becomes sparse. Cameras are correlated via the scene representation. If the optimization removes drift in the scene, multiple cameras need to move. The MLP refiner models the correlation of cameras. + +Calibration Refinement. We do not assume information about precise calibration parameters, although often reported by devices. We do assume that the principal point is in the center, that pixels are unskewed and square. We do not model image distortion. While these are reasonable assumptions for many data regimes, we cannot rely on the focal length to be given. Thus, we refine the focal length starting from a heuristic: $K_{i}^{t} = f_{\mathrm{Calib}}^{t}(K_{i}^{\mathrm{init}})$ . The refinement function $f_{\mathrm{Calib}}^{t}$ entails a single learnable parameter $\alpha^{t}$ such that the focal length $f_{i}^{t} = f^{\mathrm{init}} \cdot (1 + \alpha^{t})$ . As before, superscript $t$ denotes the time step. We optimize $\alpha^{t}$ using AdamW [57] with weight decay, biasing it towards a small relative scale factor. Estimates of $\alpha$ are carried over across iterations. We set $f^{\mathrm{init}}$ to 70% of the image diagonal and it is shared by all cameras in our experiments. + +# 3.2 Relocalization + +Given the scene model of iteration $t$ , we attempt to register more images to determine the training set for the next mapping iteration, $\mathcal{I}_{\mathrm{Reg}}^{t + 1}$ . We pass all images in $\mathcal{I}$ to the scene coordinate regressor $f_{\mathrm{SCR}}^t$ to gather 2D-3D correspondences, and solve for their poses using RANSAC and PnP: + +$$ +\tilde {T} _ {i} ^ {t + 1}, s _ {i} ^ {t + 1} = g \left(K _ {i} ^ {t}, \left\{\left(\mathbf {p} _ {i j}, \mathbf {y} _ {i j} ^ {t}\right) \right\}\right). \tag {5} +$$ + +Here, we assume that the pose solver returns a confidence score $s_i^{t+1}$ alongside the pose itself that lets us decide whether the pose of the image has been estimated successfully. We simply utilize the inlier count as score $s_i^{t+1}$ and apply a threshold to form the training set of the next mapping iteration: $\mathcal{I}_{\mathrm{Reg}}^{t+1} = \left\{ I_i | s_i^{t+1} > \tau_s \right\}$ . The relocalization process is depicted in Figure 2, bottom. + +# 3.3 Initialization + +We start the reconstruction with one image: $\mathcal{I}_{\mathrm{Reg}}^0 = \{I_{\mathrm{seed}}\}$ . We set the seed pose $T_{\mathrm{seed}}^0$ to identity, and we initialize the calibration $K_{\mathrm{seed}}^{\mathrm{init}}$ as explained above. + +We cannot train a scene coordinate regression network using Eq. 3 with a single image. The reprojection error is ill-conditioned without multiple views constraining the depth. Therefore, we optimize a different objective in the seed iteration, inspired by Map-free Relocalization [3]. Arnold et al. argue that a single image and a depth estimate allow to relocalize query images, albeit with limited accuracy. Our experiments show that such coarse relocalizations of a few images suffice as initialization for optimizing the reprojection error of Eq. 3. + +Let $d_{ij}$ be a depth value predicted for pixel $j$ of image $i$ . We derive a target scene coordinate by back-projection as $\hat{\mathbf{y}}_{ij} = d_{ij}(K_{\mathrm{seed}}^{\mathrm{init}})^{-1}\mathbf{p}_{ij}$ . We train an initial scene coordinate regression network $f_{\mathrm{SCR}}^{0}$ by optimizing $\sum_{j\in I^0}\| \hat{\mathbf{y}}_{ij} - \mathbf{y}_{ij}\|$ . Fig. 3 (right) shows a scene coordinate point cloud learned from a depth estimate, and successful relocalization against it. + +We found our pipeline to be robust w.r.t. selecting the seed image. However, when a randomly selected image has little to no visual overlap with the remaining images, the whole reconstruction would fail. Selecting an unfortunate seed image, e.g., at the very end of a long camera trajectory, can increase reconstruction times. To decrease the probability of such incidents, we try 5 random seed images and choose the one with the highest relocalization rate across 1000 other mapping images. Since mapping seed images is fast, ca. 1 min on average, this poses no significant computational burden. + +# 3.4 Implementation + +We base our pipeline on the public code of the ACE relocalizer [12]. ACE uses a convolutional feature backbone, pre-trained on ScanNet [29], that ingests images scaled to 480px height. On top of the backbone is the mapping network, a 9-layer MLP with 512 channels, consuming 4MB of weights in 16-bit floating point precision. This is our scene representation. + +The ACE training process is reasonably fast, taking 5 minutes to train the scene network. Since we repeat training the scene network in multiple iterations, we expedite the process further to decrease our total reconstruction time. + +Adaptive Sampling of Training Patches. ACE trains the scene representation based on 8M patches sampled from the mapping images, a process that takes 1 minute. This is excessive when having very few mapping images in the beginning of the reconstruction, thus we loop over the mapping images at most 10 times when sampling patches, or until 8M patches have been sampled. Using this adaptive strategy, sampling patches for the seed reconstruction where we have 1 image only takes 2 seconds instead of 1 minute. + +Adaptive Stopping. ACE trains the scene model using a fixed one-cycle learning rate schedule [83] to a total of 25k parameter updates. We make the training schedule adaptive since the network is likely to converge fast when trained on few images only. We monitor the reprojection error of scene coordinates within a mini-batch. If for 100 consecutive batches $70\%$ of reprojection errors are below an inlier threshold of 10px, we stop training early. We approximate the one-cycle learning rate schedule in a linear fashion: we increase the learning rate in the first 1k iteration from $5 \times 10^{-4}$ to $3 \times 10^{-3}$ and when the early stopping criterion has been met, we decrease the learning rate to $5 \times 10^{-4}$ within 5k iterations. + +We report more implementation details and hyper-parameters in the supplement. Our code is also publicly available to ensure reproducibility. + +# 4 Experiments + +We refer to our SfM pipeline as ACE0 (ACE Zero) since it builds on top of the ACE relocalizer [12] but adds the ability to train from scratch, without poses. We demonstrate the effectiveness of our approach on three datasets and 31 scenes in total. For all experiments, we rely on ZoeDepth [7] to initialize our reconstructions. All timings reported for ACE0 are based on a single V100 GPU. + +Baselines. We consider the pose estimates of COLMAP with default parameters as our pseudo ground truth, obtained by extensive processing. We also run COLMAP with parameter settings recommended for large image collections of 1k images and more [79]. This variation achieves much faster processing (denoted COLMAP fast). We use a V100 GPU for COLMAP feature extraction and matching, and we specify the exact parameters of COLMAP in the supplement. Furthermore, we show some results of RealityCapture [68], an efficient commercial feature-based SfM pipeline. + +We compare to learning-based SfM approaches that are not restricted to few-frame scenarios, namely to NeRF-based BARF [54] and NoPe-NeRF [10]. We also compare to DUSt3R [97], a non-NeRF learning-based SfM method. While we focus on reconstructing unsorted image collections, the datasets we consider allow for sequential processing. Hence, for context, we also show results of DROID-SLAM [90], a neural SLAM approach. Unless specified otherwise, we report timings based on a single V100 GPU. + +![](images/7c1dc002114a01dbd79150f763da2146f7286fb718e64ef0d1745a6b71872aab.jpg) +Fig. 4: Reconstructed Poses. We show poses estimated by ACE0 for a selection of scenes. We color code the reconstruction iteration in which a particular view has been registered. We show the ACE0 point cloud as a representation of the scene. The seed image is shown as a gray frustum. We also compare our poses to poses estimated by COLMAP (Mip-NeRF 360, Tanks and Temples) and KinectFusion (7-Scenes). + +Benchmark. We show results on 7-Scenes [81], a relocalization dataset, on Mip-NeRF 360 [5], a view synthesis dataset and on Tanks and Temples [49], a reconstruction dataset. Comparing poses on these datasets is problematic, as our pseudo ground truth is estimated rather than measured. For example, an approach might be more accurate than COLMAP on individual scenes. Computing pose errors w.r.t. COLMAP would result in incorrect conclusions. Therefore, we gauge the pose quality in a self-supervised way, using novel view synthesis [96]. + +We let each method estimate the poses of all images of a scene. For evaluation, we split the images into training and test sets. We train a Nerfacto [89] model on the training set, and synthesize views for the test poses. We compare the synthesized images to the test images and report the difference as peak signal-to-noise ratio (PSNR). To ensure a fair comparison to NeRF-based competitors, we use these methods in the same way as the other SfM methods: we run them to pose all images, and train a Nerfacto model on top of their poses. This is to ensure that we solely compare the pose quality across methods, and not different capabilities in view synthesis. The quality of poses affects the PSNR numbers in two ways: Good training poses let the NeRF model fit a consistent scene representation. Good testing poses make sure that the synthesized images are aligned with the original image. We explain further details in the supplement which also includes additional, perceptual metrics. In spirit, our evaluation is similar to the Tanks and Temples benchmark which evaluates a derived scene mesh rather than camera poses. However, our evaluation can be applied to arbitrary datasets as it does not need ground truth geometry. + +
FramesPseudo Ground TruthAll Frames200 Frames50 Frames
Kinect FusionCOLMAP (default)COLMAP (fast)DROID-SLAM† [90]ACE0 (ours)KF+ACE0 (ours)BARF [54]NoPE-NeRF† [10]ACE0 (ours)DUS3R [97]ACE0 (Ours)
Chess6k19.623.623.519.323.323.012.812.622.718.919.2
Fire4k19.222.622.613.022.322.312.711.822.118.819.5
Heads2k17.018.818.917.618.819.110.711.819.918.421.3
Office10k18.921.421.6failed21.121.511.910.919.812.513.7
Pumpkin6k19.924.123.818.324.123.819.614.224.721.722.3
RedKitchen12k17.621.421.410.920.820.911.611.218.913.813.7
Stairs3k19.016.721.013.017.719.915.815.918.815.315.4
Average18.721.221.8N/A21.221.513.612.621.017.117.9
Avg. Timerealtime38h13h18min1h7min8.5h47h27min4min*16min
+ +Table 1: 7-Scenes. We show the pose accuracy via view synthesis with Nerfacto [89] as PSNR in dB, and the reconstruction time. Results for All Frames are color coded w.r.t. similarity to the COLMAP pseudo ground truth: $>0.5$ dB better within $\pm 0.5$ dB $>0.5$ dB worse $>1$ dB worse. For some competitors, we had to sub-sample the images due to their computational complexity (right side). Method needs sequential inputs. *Results on more powerful hardware. + +# 4.1 7-Scenes + +The 7-Scenes dataset [81] consists of seven indoor scenes, scanned with a Kinect v1 camera. Multiple, disconnected scans are provided for each scene to a total of $2\mathrm{k} - 12\mathrm{k}$ images. For each method, we assume a shared focal length across scans and initialize with the default calibration of a Kinect v1. The dataset comes with pseudo ground truth camera poses estimated by KinectFusion [44, 64], a depth-based SLAM system. Individual scans were registered but not bundle-adjusted [13]. Inspired by [13], we recompute alternative, bundle-adjusted pseudo ground truth by running COLMAP with default parameters. + +Discussion. We show results in Table 1. Of both pseudo ground truth versions, KinectFusion achieves lower PSNR numbers than COLMAP, presumably due to the lack of global optimization. COLMAP with fast parameters shows PSNR numbers similar to COLMAP with default parameters, on average. Both versions of running COLMAP take considerable time to reconstruct each scene. We note that COLMAP has been optimised for quality, rather than speed. Not all acceleration strategies from the feature-based SfM literature have been implemented in COLMAP, so presumably comparable quality can be obtained faster. DROID-SLAM [90] does not perform well on 7-Scenes and partially fails altogether, presumably due to the jumps between individual scans of each scene. + +Our approach, ACE0, achieves a pose quality comparable to the COLMAP pseudo ground truth while reconstructing each scene in $\sim 1$ hour despite the large number of images. We show qualitative examples in Figure 4 and in the supplement. We also demonstrate that ACE0 can swiftly optimize an initial set of approximate poses. When starting from KinectFusion poses, ACE0 increases PSNR significantly in less than 10 minutes per scene, see “ $\mathrm{KF} + \mathrm{ACE0}$ ” in Table 1. In the supplement, we include a parameter study on 7-Scenes to show that ACE0 is robust to the choice of depth estimator. We also show the positive impact of pose refinement on the reconstruction quality as well as the reconstruction speedup due to our early stopping schedule. + +
ACE [12]ACE [12]ACE0 (ours)
SupervisionKinectFusionCOLMAPself-supervised
Chess96.0%100.0%100.0%
Fire98.4%99.5%98.8%
Heads100.0%100.0%100.0%
Office36.9%100.0%99.1%
Pumpkin47.3%100.0%99.9%
Redlitchen47.8%98.9%98.1%
Stairs74.1%85.0%61.0%
Average71.5%97.6%93.8%
+ +Table 2 (a): Relocalization on 7-Scenes. % poses below $5\mathrm{cm},5^{\circ}$ error, computed w.r.t. COLMAP pseudo GT. + +
Pseudo GT (COLMAP)DROID-SLAM† [90]BARF [54]NoPe-NeRF† [10]ACEO (ours)
Bicycle21.510.911.912.218.7
Bonsai27.610.912.514.825.8
Counter25.512.911.911.624.5
Garden26.316.713.313.825.0
Kitchen27.413.913.314.426.1
Room28.011.311.914.319.8
Stump16.813.915.013.720.5
Average24.712.912.813.522.9
+ +Table 2 (b): Mip-NeRF 360. Pose quality in PSNR, higher is better. Best in bold. Method needs sequential inputs. + +For our learning-based competitors, we sub-sampled images due to their computational constraints, see right side of Table 1. Even using only 200 images, NoPe-NeRF [10] takes 2 days to fit a model and estimate poses. Despite these long processing times, we observe poor pose quality of BARF and NoPe-NeRF in terms of PSNR. BARF [54] requires pose initialisation. We provide identity poses since the scenes of 7Scenes are roughly forward-facing. Still, the camera motions are too complex for BARF to handle. NoPe-NeRF does not require pose initialisation but needs roughly sequential images, which we did provide. NoPe-NeRF relies upon successive images having similar poses to perform well, and struggles with the large jumps between the subsampled images. + +For the comparison with DUSt3R [97], we had to subsample the sequences further, as we were only able to run it with 50 images at most, even when using an A100 GPU with 40GB memory. DUSt3R achieves reasonable PSNR numbers but consistently lower than ACE0. + +Relocalization. ACE0 is a learning-based SfM tool but it is also a self-supervised visual relocaliser. In Table 2 (a), we compare it to the supervised relocalizer ACE [12]. Using the scale-metric pseudo ground truth of [13], we train ACE with COLMAP mapping poses, and evaluate it against COLMAP query poses. Unsurprisingly, ACE achieves almost perfect relocalization under the usual $5\mathrm{cm}$ , $5^{\circ}$ error threshold. Interestingly, ACE0 achieves almost identical results when mapping the scene self-supervised, and evaluating the relocalized query poses against the COLMAP pseudo ground truth. For context, when training ACE with Kinect-Fusion mapping poses and evaluating against COLMAP pseudo ground truth, results are far worse. This signifies that ACE0 mapping poses are very similar to the COLMAP mapping poses, and less similar to KinectFusion mapping poses. We give more details about this experiment in the supplement. + +# 4.2 Mip-NeRF 360 + +The Mip-NeRF 360 dataset [5] consists of seven small-scale scenes, both indoor and outdoor. The dataset was reconstructed with COLMAP and comes with intrinsics (which we ignore), camera poses and undistorted (pinhole camera) images. For each method, we assume a shared focal length per scene. + +Discussion. We present PSNR results in Table 2 (b). NoPe-NeRF does not perform well on this dataset, despite processing each scene for 2 days. The differences + +
FramesDROID-FramesDROID-SparseSparse
COLMAP (default)Reality CaptureSLAM† [90]ACE0 (ours)COLMAP (fast)Reality CaptureSLAM† [90]ACE0 (ours)COLMAP + Reloc + BACOLMAP + ACE0 (ours)
TrainingBarn41024.021.219.016.512.2k24.416.913.517.726.325.1
Caterpillar38317.115.916.616.911.4k18.617.918.918.618.718.8
Church50718.317.614.317.219.3k12.1-11.516.518.517.3
Ignatius26420.117.717.819.87.8k20.818.619.120.720.920.7
Meetingroom37118.618.115.618.011.1k19.418.217.116.620.820.3
Truck25121.119.018.320.17.5k23.619.120.623.023.423.1
Average36419.918.216.918.114.6k19.818.216.818.921.420.9
Avg Time1h3min5min1.1h74h14h18min2.2h8h1.8h
IntermediateFamily15219.518.817.619.04.4k21.219.819.818.021.321.3
Francis30221.620.720.720.17.8k19.920.421.821.722.522.7
Horse15119.219.016.319.56.0k21.620.719.221.722.622.3
Lighthouse30916.616.513.617.58.3k19.016.618.918.619.520.5
Playground30719.119.211.418.77.7k17.916.511.320.421.221.0
Train30116.815.413.816.212.6k19.614.415.618.519.818.5
Average25418.818.315.618.57.8k19.918.117.819.821.121.0
Avg Time32min2min3min1.3h48h11h14min2.2h5h1h
AdvancedAuditorium30219.612.216.718.713.6k13.7-16.620.021.419.8
Ballroom32416.318.313.117.910.8k17.2-10.418.918.015.6
Courtroom30118.217.212.317.112.6k14.6-10.216.318.717.8
Palace50914.211.710.810.721.9k13.8-8.611.015.312.3
Temple30218.115.711.89.717.5k13.3-11.914.819.616.1
Average34817.315.012.914.815.6k14.5-11.516.218.616.3
Avg Time1h2min4min1h71h27min2.8h10h2.1h
+ +Table 3: Tanks and Temples. We show the pose accuracy via view synthesis with Nerfacto [89] as PSNR in dB, and the reconstruction time. We color code results compared to COLMAP, default and fast, respectively: $>0.5$ dB better within $\pm 0.5$ dB $>0.5$ dB worse $>1$ dB worse. Method needs sequential inputs. + +between sequential images are too large. DROID-SLAM fails for the same reason. BARF performs poorly because the identity pose is not a good initialisation for most scenes. In contrast, ACE0 reconstructs the dataset successfully. While it achieves slightly lower PSNR than COLMAP, its pose estimates are similar, $cf$ . Figure 4. The supp. shows that synthesized images based on ACE0 are visually close to those of COLMAP while our learning-based competitors are far off. + +# 4.3 Tanks and Temples + +The Tanks and Temples dataset [49] contains 21 diverse scenes, both indoors and outdoors, and with varying spatial extent. We remove two scenes (Panther, M60) with military associations. We also remove two scenes (Courthouse, Museum) where the COLMAP baseline did not finish the reconstruction after 5 days of processing or ran out of memory. For the latter two scenes, we provide ACE0 results in the supplement. The dataset provides 150-500 images per scene but also the original videos as a source for more frames. Thus, we consider each scene in two versions: Using 150-500 images and using $4\mathrm{k} - 22\mathrm{k}$ frames. For all methods, we assume a pinhole camera model with shared focal length across images, and images to be unordered. We found none of the learning-based SfM competitors + +to be applicable to this dataset. NoPe-NeRF would run multiple days per scene, even when considering only a few hundred images. DUSt3R would run out of memory. BARF needs reasonable pose initializations which are not available. + +Discussion. We show PSNR numbers of RealityCapture, DROID-SLAM and ACE0 in Table 3, color-coded by similarity to the COLMAP pseudo GT. ACE0 achieves reasonable results when reconstructing scenes from a few hundred images (Table 3, left). ACE0 generally predicts plausible poses (cf., Figure 4), even if PSNR numbers are sometimes lower than those of COLMAP. RealityCapture performs similar to ACE0 while DROID-SLAM struggles on the sparse images. + +Next, we consider more than 1k images per scene (right side of Table 3). Here, we run COLMAP with parameters tuned for large images collections (fast) due to the extremely large image sets. ACE0 offers a reconstruction quality comparable to COLMAP on average while also being fast. We run RealityCapture on some of the scenes but it produces fractured reconstructions for these large images sets, leading to low PSNR numbers. DROID-SLAM still struggles on many scenes despite having access to sequential images that are temporally close. + +In the two rightmost columns, we initialise with a sparse COLMAP reconstruction from 150-500 images, and extend and refine it using all available frames. Firstly, using a feature-based baseline, we register the full set of frames using the relocalization mode of COLMAP, followed by a final round of bundle adjustment. Secondly, we run ACE0 initialised with the poses of the sparse COLMAP reconstruction. Both variants are considerably faster than running COLMAP from scratch on the full set of frames. ACE0 is able to register and refine all additional frames in 1-2 hours, on average. Again, we find the pose quality of ACE0 comparable to the feature-based alternative. + +# 5 Conclusion and Future Work + +We have presented scene coordinate reconstruction, a new approach to learning-based SfM. We learn an implicit, neural scene representation from a set of unposed images. Our method, ACE0, is able to reconstruct a wide variety of scenes. In many cases, the accuracy of estimated poses is close to that of COLMAP and synthesized images visually similar. Unlike previous learning-based SfM methods, ACE0 can be applied to multiple thousand unsorted images, without pose priors, and reconstructs them within a few hours. + +Limitations. We show some failure cases in the supplement. Scene coordinate regression struggles with repetitive structures since the network is not able to make multi-modal predictions for visually ambiguous inputs. Scene coordinate regression also struggles with representing large areas. The common solution is to use network ensembles based on pre-clustering of the scene [12, 16] which is difficult in a reconstruction setting. While scene coordinate regression generalises quite well, it has difficulties to bridge extreme view point or lighting changes, such as day versus night. In our experiments, we assumed a simple pinhole camera model with shared intrinsics across images. To the best of our knowledge, scene coordinate regression has not been coupled with image distortion, thus far. + +# References + +1. Agarwal, S., Furukawa, Y., Snavely, N., Simon, I., Curless, B., Seitz, S.M., Szeliski, R.: Building Rome in a day. ACM TOG (2011) +2. Agarwal, S., Snavely, N., Seitz, S.M., Szeliski, R.: Bundle adjustment in the large. In: ECCV (2010) +3. Arnold, E., Wynn, J., Vicente, S., Garcia-Hernando, G., Monszpart, A., Prisacariu, V.A., Turmukhambetov, D., Brachmann, E.: Map-free visual relocalization: Metric pose relative to a single image. In: ECCV (2022) +4. Balntas, V., Li, S., Prisacariu, V.A.: RelocNet: Continuous metric learning relocalisation using neural nets. In: ECCV (2018) +5. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-NeRF 360: Unbounded anti-aliased neural radiance fields. In: CVPR (2022) +6. Beardsley, P.A., Zisserman, A., Murray, D.W.: Sequential updating of projective and affine structure from motion. IJCV (1997) +7. Bhat, S.F., Birkl, R., Wofk, D., Wonka, P., Müller, M.: ZoeDepth: Zero-shot transfer by combining relative and metric depth. arXiv (2023) +8. Bhowmick, B., Patra, S., Chatterjee, A., Govindu, V.M., Banerjee, S.: Divide and conquer: Efficient large-scale structure from motion using graph partitioning. In: ACCV (2015) +9. Bhowmick, B., Patra, S., Chatterjee, A., Govindu, V.M., Banerjee, S.: Divide and conquer: A hierarchical approach to large-scale structure-from-motion. CVIU (2017) +10. Bian, W., Wang, Z., Li, K., Bian, J.W., Prisacariu, V.A.: NoPe-NeRF: Optimising neural radiance field with no pose prior. In: CVPR (2023) +11. Bloesch, M., Czarnowski, J., Clark, R., Leutenegger, S., Davison, A.J.: CodeSLAM — Learning a compact, optimisable representation for dense visual SLAM. In: CVPR (2018) +12. Brachmann, E., Cavallari, T., Prisacariu, V.A.: Accelerated coordinate encoding: Learning to relocalize in minutes using RGB and poses. In: CVPR (2023) +13. Brachmann, E., Humenberger, M., Rother, C., Sattler, T.: On the limits of pseudo ground truth in visual camera re-localisation. In: ICCV (2021) +14. Brachmann, E., Krull, A., Nowozin, S., Shotton, J., Michel, F., Gumhold, S., Rother, C.: DSAC-differentiable RANSAC for camera localization. In: CVPR (2017) +15. Brachmann, E., Rother, C.: Learning less is more-6D camera localization via 3D surface regression. In: CVPR (2018) +16. Brachmann, E., Rother, C.: Expert sample consensus applied to camera relocalization. In: ICCV (2019) +17. Brachmann, E., Rother, C.: Visual camera re-localization from RGB and RGB-D images using DSAC. IEEE TPAMI (2021) +18. Brégier, R.: Deep regression on manifolds: a 3D rotation case study. In: 3DV (2021) +19. Brown, D.: The bundle adjustment-progress and prospect. In: Congr. of the Int. Soc. for Photogr. (1976) +20. Brown, M., Lowe, D.G.: Unsupervised 3D object recognition and reconstruction in unordered datasets. In: 3DIM (2005) +21. Carlone, L., Tron, R., Daniilidis, K., Dellaert, F.: Initialization techniques for 3D SLAM: A survey on rotation estimation and its use in pose graph optimization. In: ICRA (2015) + +22. Cavallari, T., Bertinetto, L., Mukhoti, J., Torr, P.H., Golodetz, S.: Let's take this online: Adapting scene coordinate regression network predictions for online RGB-D camera relocalisation. In: 3DV (2019) +23. Cavallari, T., Golodetz, S., Lord, N.A., Valentin, J., Di Stefano, L., Torr, P.H.: On-the-fly adaptation of regression forests for online camera relocalisation. In: CVPR (2017) +24. Chen, S., Bhalgat, Y., Li, X., Bian, J., Li, K., Wang, Z., Prisacariu, V.A.: Neural refinement for absolute pose regression with feature synthesis. In: CVPR (2024) +25. Chen, S., Li, X., Wang, Z., Prisacariu, V.: DFNet: Enhance absolute pose regression with direct feature matching. In: ECCV (2022) +26. Chen, S., Wang, Z., Prisacariu, V.: Direct-PoseNet: Absolute pose regression with photometric consistency. In: 3DV (2021) +27. Cheng, Z., Esteves, C., Jampani, V., Kar, A., Maji, S., Makadia, A.: LU-NeRF: Scene and pose estimation by synchronizing local unposed NeRFs. In: ICCV (2023) +28. Crandall, D., Owens, A., Snavely, N., Huttenlocher, D.: Discrete-continuous optimization for large-scale structure from motion. In: CVPR (2011) +29. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: ScanNet: Richly-annotated 3D reconstructions of indoor scenes. In: CVPR (2017) +30. Davison, A.J.: Real-time simultaneous localisation and mapping with a single camera. In: ICCV (2003) +31. DeTone, D., Malisiewicz, T., Rabinovich, A.: SuperPoint: Self-supervised interest point detection and description. In: CVPRW (2018) +32. Ding, M., Wang, Z., Sun, J., Shi, J., Luo, P.: CamNet: Coarse-to-fine retrieval for camera re-localization. In: ICCV (2019) +33. Dusmanu, M., Rocco, I., Pajdla, T., Pollefeys, M., Sivic, J., Torii, A., Sattler, T.: D2-net: A trainable CNN for joint description and detection of local features. In: CVPR (2019) +34. Fischler, M.A., Bolles, R.C.: Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Comm. of the ACM (1981) +35. Gao, X.S., Hou, X.R., Tang, J., Cheng, H.F.: Complete solution classification for the perspective-three-point problem. IEEE TPAMI (2003) +36. Gherardi, R., Farenzena, M., Fusiello, A.: Improving the efficiency of hierarchical structure-and-motion. In: CVPR (2010) +37. Govindu, V.M.: Combining two-view constraints for motion estimation. In: CVPR (2001) +38. Govindu, V.M.: Lie-algebraic averaging for globally consistent motion estimation. In: CVPR (2004) +39. Hartley, R., Trumpf, J., Dai, Y., Li, H.: Rotation averaging. IJCV (2013) +40. Hartley, R., Zisserman, A.: Multiple view geometry in computer vision. Cambridge University Press (2003) +41. He, X., Sun, J., Wang, Y., Peng, S., Huang, Q., Bao, H., Zhou, X.: Detector-free structure from motion. In: CVPR (2024) +42. Heinly, J., Schonberger, J.L., Dunn, E., Frahm, J.M.: Reconstructing the World in six days. In: CVPR (2015) +43. Humenberger, M., Cabon, Y., Pion, N., Weinzaepfel, P., Lee, D., Guérin, N., Sattler, T., Csurka, G.: Investigating the role of image retrieval for visual localization: An exhaustive benchmark. IJCV (2022) + +44. Izadi, S., Kim, D., Hilliges, O., Molyneaux, D., Newcombe, R., Kohli, P., Shotton, J., Hodges, S., Freeman, D., Davison, A.J., Fitzgibbon, A.: KinectFusion: real-time 3D reconstruction and interaction using a moving depth camera. In: UIST (2011) +45. Jeong, Y., Ahn, S., Choy, C., Anandkumar, A., Cho, M., Park, J.: Self-calibrating neural radiance fields. In: ICCV (2021) +46. Jin, Y., Mishkin, D., Mishchuk, A., Matas, J., Fua, P., Yi, K.M., Trulls, E.: Image matching across wide baselines: From paper to practice. IJCV (2021) +47. Kendall, A., Grimes, M., Cipolla, R.: PoseNet: A convolutional network for real-time 6-DoF camera relocalization. In: ICCV (2015) +48. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3D gaussian splatting for real-time radiance field rendering. ACM TOG (2023) +49. Knapitsch, A., Park, J., Zhou, Q.Y., Koltun, V.: Tanks and Temples: Benchmarking large-scale scene reconstruction. ACM TOG (2017) +50. Kraus, K.: Photogrammetry. No. v. 1 in Photogrammetry, Ferdinand Dummlers Verlag (1993) +51. Laskar, Z., Melekhov, I., Kalia, S., Kannala, J.: Camera relocalization by computing pairwise relative poses using convolutional neural network. In: ICCV Workshops (2017) +52. Li, X., Wang, S., Zhao, Y., Verbeek, J., Kannala, J.: Hierarchical scene coordinate classification and regression for visual localization. In: CVPR (2020) +53. Lin, A., Zhang, J.Y., Ramanan, D., Tulsiani, S.: Relpose++: Recovering 6d poses from sparse-view observations. In: 3DV (2024) +54. Lin, C.H., Ma, W.C., Torralba, A., Lucey, S.: BARF: Bundle-adjusting neural radiance fields. In: ICCV (2021) +55. Lin, Y., Müller, T., Tremblay, J., Wen, B., Tyree, S., Evans, A., Vela, P.A., Birchfield, S.: Parallel inversion of neural radiance fields for robust pose estimation. In: ICRA (2023) +56. Lindenberger, P., Sarlin, P.E., Pollefeys, M.: LightGlue: Local feature matching at light speed. In: ICCV (2023) +57. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: ICLR (2019) +58. Lowe, D.G.: Distinctive image features from scale-invariant keypoints. IJCV (2004) +59. Martinec, D., Pajdla, T.: Robust rotation and translation estimation in multiview reconstruction. In: CVPR (2007) +60. Meng, Q., Chen, A., Luo, H., Wu, M., Su, H., Xu, L., He, X., Yu, J.: GNeRF: GAN-based Neural Radiance Field without Posed Camera. In: ICCV (2021) +61. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020) +62. Moreau, A., Piasco, N., Bennehar, M., Tsishkou, D., Stanciulescu, B., de La Fortelle, A.: CROSSFIRE: Camera relocalization on self-supervised features from an implicit representation. ICCV (2023) +63. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM TOG (2022) +64. Newcombe, R., Izadi, S., Hilliges, O., Molyneaux, D., Kim, D., Davison, A.J., Kohi, P., Shotton, J., Hodges, S., Fitzgibbon, A.: KinectFusion: Real-time dense surface mapping and tracking. In: ISMAR (2011) +65. Nistér, D., Naroditsky, O., Bergen, J.: Visual odometry. In: CVPR (2004) +66. Pollefeys, M., Koch, R., Vergauwen, M., Van Gool, L.: Automated reconstruction of 3D scenes from sequences of images. J. of Photogr. and Rem. Sens. (2000) + +67. Rau, A., Garcia-Hernando, G., Stoyanov, D., Brostow, G.J., Turmukhambetov, D.: Predicting visual overlap of images through interpretable non-metric box embeddings. In: ECCV (2020) +68. Reality, C.: Reality Capture. https://www.capturingreality.com/ realitycapture (2016), [accessed 15-Nov-2023] +69. Reizenstein, J., Shapovalov, R., Henzler, P., Sbordone, L., Labatut, P., Novotny, D.: Common objects in 3D: Large-scale learning and evaluation of real-life 3D category reconstruction. In: ICCV (2021) +70. Sarlin, P.E., Cadena, C., Siegwart, R., Dymczyk, M.: From coarse to fine: Robust hierarchical localization at large scale. In: CVPR (2019) +71. Sarlin, P.E., DeTone, D., Malisiewicz, T., Rabinovich, A.: SuperGlue: Learning feature matching with graph neural networks. In: CVPR (2020) +72. Sarlin, P.E., Unagar, A., Larsson, M., Germain, H., Toft, C., Larsson, V., Pollefeys, M., Lepetit, V., Hammarstrand, L., Kahl, F., Sattler, T.: Back to the feature: Learning robust camera localization from pixels to pose. In: CVPR (2021) +73. Sattler, T., Leibe, B., Kobbelt, L.: Fast image-based localization using direct 2D-to-3D matching. In: ICCV (2011) +74. Sattler, T., Leibe, B., Kobbelt, L.: Improving image-based localization by active correspondence search. In: ECCV (2012) +75. Sattler, T., Leibe, B., Kobbelt, L.: Efficient & effective prioritized matching for large-scale image-based localization. IEEE TPAMI (2016) +76. Sattler, T., Torii, A., Sivic, J., Pollefeys, M., Taira, H., Okutomi, M., Pajdla, T.: Are large-scale 3D models really necessary for accurate visual localization? In: CVPR (2017) +77. Sattler, T., Zhou, Q., Pollefeys, M., Leal-Taixe, L.: Understanding the limitations of CNN-based absolute camera pose regression. In: CVPR (2019) +78. Schaffalitzky, F., Zisserman, A.: Multi-view matching for unordered image sets, or "How do i organize my holiday snaps?". In: ECCV (2002) +79. Schonberger, J.L.: Colmap Github Issues. https://github.com/colmap/colmap/issues/116#issuecomment-298926277 (2017), [accessed Nov/15/2023] +80. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: CVPR (2016) +81. Shotton, J., Glocker, B., Zach, C., Izadi, S., Criminisi, A., Fitzgibbon, A.: Scene coordinate regression forests for camera relocalization in RGB-D images. In: CVPR (2013) +82. Sinha, S., Zhang, J.Y., Tagliasacchi, A., Gilitschenski, I., Lindell, D.B.: Sparse-Pose: Sparse-view camera pose regression and refinement. In: CVPR (2023) +83. Smith, L.N., Topin, N.: Super-convergence: Very fast training of neural networks using large learning rates. In: Arti. Intell. and Mach. Learn. for Multi-Domain Operations Appl. (2019) +84. Snavely, N., Seitz, S.M., Szeliski, R.: Photo tourism: exploring photo collections in 3D. ACM TOG (2006) +85. Snavely, N., Seitz, S.M., Szeliski, R.: Modeling the world from internet photo collections. IJCV (2008) +86. Snavely, N., Seitz, S.M., Szeliski, R.: Skeletal graphs for efficient structure from motion. In: CVPR (2008) +87. Sun, J., Shen, Z., Wang, Y., Bao, H., Zhou, X.: LoFTR: Detector-free local feature matching with transformers. In: CVPR (2021) +88. Szeliski, R., Kang, S.B.: Recovering 3D shape and motion from image streams using nonlinear least squares. J. of Vis. Com. and Image Repr. (1994) + +89. Tancik, M., Weber, E., Ng, E., Li, R., Yi, B., Kerr, J., Wang, T., Kristoffersen, A., Austin, J., Salahi, K., Ahuja, A., McAllister, D., Kanazawa, A.: Nerfstudio: A modular framework for neural radiance field development. In: ACM TOG (2023) +90. Teed, Z., Deng, J.: DROID-SLAM: Deep visual SLAM for monocular, stereo, and RGB-D cameras. In: NeurIPS (2021) +91. Toldo, R., Gherardi, R., Farenzena, M., Fusiello, A.: Hierarchical structure-and-motion recovery from uncalibrated images. CVIU (2015) +92. Triggs, B., McLauchlan, P.F., Hartley, R.I., Fitzgibbon, A.W.: Bundle adjustment—a modern synthesis. In: Int. Worksh. on Vis. Alg. (2000) +93. Türkoglu, M.Ö., Brachmann, E., Schindler, K., Brostow, G., Monszpart, A.: Visual camera re-localization using graph neural networks and relative pose supervision. In: 3DV (2021) +94. Ulyanov, D., Vedaldi, A., Lempitsky, V.: Deep image prior. In: CVPR (2018) +95. Ummenhofer, B., Zhou, H., Uhrig, J., Mayer, N., Ilg, E., Dosovitskiy, A., Brox, T.: DeMoN: Depth and motion network for learning monocular stereo. In: CVPR (2017) +96. Waechter, M., Beljan, M., Fuhrmann, S., Moehrle, N., Kopf, J., Goesele, M.: Virtual rephotography: Novel view prediction error for 3d reconstruction. ACM TOG (2017) +97. Wang, S., Leroy, V., Cabon, Y., Chidlovskii, B., Revaud, J.: DUSt3R: Geometric 3D vision made easy. In: CVPR (2024) +98. Wang, Z., Wu, S., Xie, W., Chen, M., Prisacariu, V.A.: NeRF--: Neural radiance fields without known camera parameters. arXiv (2021) +99. Wei, X., Zhang, Y., Li, Z., Fu, Y., Xue, X.: DeepSFM: Structure from motion via deep bundle adjustment. In: ECCV (2020) +00. Wu, C.: Towards linear-time incremental structure from motion. In: 3DV (2013) +01. Xia, Y., Tang, H., Timofte, R., Van Gool, L.: SiNeRF: Sinusoidal neural radiance fields for joint pose estimation and scene reconstruction. In: BMVC (2022) +02. Yen-Chen, L., Florence, P., Barron, J.T., Rodriguez, A., Isola, P., Lin, T.Y.: iNeRF: Inverting neural radiance fields for pose estimation. In: IROS (2021) +03. Zhang, J.Y., Lin, A., Kumar, M., Yang, T.H., Ramanan, D., Tulsiani, S.: Cameras as rays: Pose estimation via ray diffusion. In: ICLR (2024) +04. Zhang, W., Kosecka, J.: Image based localization in urban environments. In: 3DPVT (2006) +05. Zhou, Q., Sattler, T., Pollefeys, M., Leal-Taixe, L.: To learn or not to learn: Visual localization from essential matrices. In: ICRA (2020) +06. Zhou, Y., Barnes, C., Jingwan, L., Jimei, Y., Hao, L.: On the continuity of rotation representations in neural networks. In: CVPR (2019) \ No newline at end of file diff --git a/2024/Scene Coordinate Reconstruction_ Posing of Image Collections via Incremental Learning of a Relocalizer/images.zip b/2024/Scene Coordinate Reconstruction_ Posing of Image Collections via Incremental Learning of a Relocalizer/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6792cf680dad83d13ef3d54f04a6918e15cef9a0 --- /dev/null +++ b/2024/Scene Coordinate Reconstruction_ Posing of Image Collections via Incremental Learning of a Relocalizer/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:982b46b8eef96981ea805aa9d4bbab7128a5d4fd2533d78b02ac3b11a6db98c9 +size 617836 diff --git a/2024/Scene Coordinate Reconstruction_ Posing of Image Collections via Incremental Learning of a Relocalizer/layout.json b/2024/Scene Coordinate Reconstruction_ Posing of Image Collections via Incremental Learning of a Relocalizer/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..208bc4fd26de4c3150373716e12e8720701448bb --- /dev/null +++ b/2024/Scene Coordinate Reconstruction_ Posing of Image Collections via Incremental Learning of a Relocalizer/layout.json @@ -0,0 +1,10462 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 153, + 112, + 462, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 112, + 462, + 166 + ], + "spans": [ + { + "bbox": [ + 153, + 112, + 462, + 166 + ], + "type": "text", + "content": "Scene Coordinate Reconstruction: Posing of Image Collections via Incremental Learning of a Relocalizer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 132, + 186, + 481, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 186, + 481, + 211 + ], + "spans": [ + { + "bbox": [ + 132, + 186, + 481, + 211 + ], + "type": "text", + "content": "Eric Brachmann1, Jamie Wynn1, Shuai Chen2, Tommaso Cavallari1, Åron Monszpart1, Daniyar Turmukhambetov1, and Victor Adrian Prisacariu1,2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 259, + 219, + 355, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 259, + 219, + 355, + 242 + ], + "spans": [ + { + "bbox": [ + 259, + 219, + 355, + 242 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 259, + 219, + 355, + 242 + ], + "type": "text", + "content": " Niantic " + }, + { + "bbox": [ + 259, + 219, + 355, + 242 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 259, + 219, + 355, + 242 + ], + "type": "text", + "content": " University of Oxford" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 159, + 270, + 455, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 270, + 455, + 445 + ], + "spans": [ + { + "bbox": [ + 159, + 270, + 455, + 445 + ], + "type": "text", + "content": "Abstract. We address the task of estimating camera parameters from a set of images depicting a scene. Popular feature-based structure-from-motion (SfM) tools solve this task by incremental reconstruction: they repeat triangulation of sparse 3D points and registration of more camera views to the sparse point cloud. We re-interpret incremental structure-from-motion as an iterated application and refinement of a visual relocalizer, that is, of a method that registers new views to the current state of the reconstruction. This perspective allows us to investigate alternative visual relocalizers that are not rooted in local feature matching. We show that scene coordinate regression, a learning-based relocalization approach, allows us to build implicit, neural scene representations from unposed images. Different from other learning-based reconstruction methods, we do not require pose priors nor sequential inputs, and we optimize efficiently over thousands of images. In many cases, our method, ACE0, estimates camera poses with an accuracy close to feature-based SfM, as demonstrated by novel view synthesis." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 446, + 399, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 446, + 399, + 456 + ], + "spans": [ + { + "bbox": [ + 160, + 446, + 399, + 456 + ], + "type": "text", + "content": "Project page: https://nianticlabs.github.io/acezero/" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 476, + 230, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 476, + 230, + 488 + ], + "spans": [ + { + "bbox": [ + 132, + 476, + 230, + 488 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 197, + 501, + 416, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 501, + 416, + 514 + ], + "spans": [ + { + "bbox": [ + 197, + 501, + 416, + 514 + ], + "type": "text", + "content": "In the beginning there was structure-from-motion." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 521, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 617 + ], + "type": "text", + "content": "The genesis of numerous computer vision tasks lies in the estimation of camera poses and scene geometry from a set of images. It is the first fundamental step that lets us leave the image plane and venture into 3D. Since structure-from-motion (SfM) is such a central capability, we have researched it for decades. By now, refined open-source tools, such as COLMAP [80], and efficient commercial packages, such as RealityCapture [68], are available to us. Feature matching-based SfM is the gold standard for estimating poses from images, with a precision that makes its estimates occasionally considered \"ground truth\" [3,5,13,46,69]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "The success of Neural Radiance Fields (NeRF) [61] has renewed interest in the question of whether SfM can be solved differently, based on neural, implicit scene representations rather than 3D point clouds. There has been some progress in recent years but, thus far, learning-based approaches to camera pose recovery" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 478, + 246 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 478, + 246 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 478, + 246 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 478, + 246 + ], + "type": "image", + "image_path": "b4aa7e4e1c73fd6aa578d18ede7c9437c374cf0ce131026b0f4af869293101f9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 257, + 482, + 357 + ], + "lines": [ + { + "bbox": [ + 130, + 257, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 257, + 482, + 357 + ], + "type": "text", + "content": "Fig. 1: Reconstructing 10,000 Images. Top: Starting from a single image and the identity pose, we train a learning-based visual relocalizer. The relocalizer allows us to estimate the poses of more views, and the additional views allow us to refine the relocalizer. We show three out of six iterations for this scene (7Scenes Office [81]). All 10k images have been posed in roughly 1 hour on a single GPU. In comparison NoPe-NeRF [10] needs two days to pose 200 images. The point cloud is a visualization of the implicit scene representation of the relocalizer. Camera positions are color coded by relocalization confidence from yellow (low) to green (high). Bottom: Point clouds from Nerfacto [89] trained on top of our poses for a few scenes from our experiments." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 385, + 482, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 385, + 482, + 481 + ], + "spans": [ + { + "bbox": [ + 130, + 385, + 482, + 481 + ], + "type": "text", + "content": "still have significant limitations. They either require coarse initial poses [45, 54, 98, 101], prior knowledge of the pose distribution [60] or sequential inputs [10, 11, 90]. In terms of the number of images that learning-based approaches can handle, they are either explicitly targeted at few-frame problems [53, 82, 95, 97, 99, 103] or they are computationally so demanding that they can realistically only be applied to a few hundred images at most [10, 54, 98]. We show that none of these limitations are an inherent consequence of using learning-based scene representations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 483, + 482, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 483, + 482, + 639 + ], + "spans": [ + { + "bbox": [ + 130, + 483, + 482, + 639 + ], + "type": "text", + "content": "Our approach is inspired by incremental SfM and its relationship to another computer vision task: visual relocalization. Visual relocalization describes the problem of estimating the camera pose of a query image w.r.t. to an existing scene map. Incremental SfM can be re-interpreted as a loop of 1) do visual relocalization to register new views to the reconstruction, and 2) refine/extend the reconstruction using the newly registered views. Local feature matching is a traditional approach to visual relocalization [70,74-76]. In recent years, multiple learning-based relocalizers have been proposed that encode the scene implicitly in the weights of a neural network [3, 12, 14, 15, 17, 22, 47]. Not all of them are suitable for building a SfM pipeline. We need a relocalizer with high accuracy and good generalization. Training of the relocalizer has to be swift. We need to be able to bootstrap relocalization without ground truth poses. And we need to be able to tell whether registration of a new image was successful." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "We show that \"scene coordinate regression\", an approach to visual relocalization proposed a decade ago [81], has the desirable properties and can serve as" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 250, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 250, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 250, + 101 + ], + "type": "text", + "content": "E. Brachmann et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "the core of a new approach to learning-based SfM: Scene coordinate reconstruction. Rather than optimizing over image-to-image matches, like feature-based SfM, scene coordinate reconstruction regresses image-to-scene correspondences directly. Rather than representing the scene as a 3D point cloud with high dimensional descriptors, we encode the scene into a lightweight neural network. Our approach works on unsorted images without pose priors and efficiently optimises over thousands of images, see Figure 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 200, + 289, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 200, + 289, + 211 + ], + "spans": [ + { + "bbox": [ + 131, + 200, + 289, + 211 + ], + "type": "text", + "content": "We summarize our contributions:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 218, + 481, + 325 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 138, + 218, + 481, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 218, + 481, + 241 + ], + "spans": [ + { + "bbox": [ + 138, + 218, + 481, + 241 + ], + "type": "text", + "content": "- Scene Coordinate Reconstruction, a new approach to SfM based on incremental learning of scene coordinate regression, a visual relocalization principle." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 242, + 481, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 242, + 481, + 277 + ], + "spans": [ + { + "bbox": [ + 138, + 242, + 481, + 277 + ], + "type": "text", + "content": "- We turn the fast-learning visual relocalizer ACE [12] into a SfM framework that is able to predict the camera poses of a set of unposed RGB images. We refer to this new SfM pipeline as " + }, + { + "bbox": [ + 138, + 242, + 481, + 277 + ], + "type": "inline_equation", + "content": "ACE0" + }, + { + "bbox": [ + 138, + 242, + 481, + 277 + ], + "type": "text", + "content": " (ACE Zero)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 277, + 481, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 277, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 138, + 277, + 481, + 325 + ], + "type": "text", + "content": "- Compared to ACE [12], we add the capability to train in a self-supervised fashion. We start from a single image, and iterate between learning the map and registering new views. We expedite reconstruction times by early stopping, and increase reconstruction quality by pose refinement." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 342, + 237, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 237, + 355 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 237, + 355 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 366, + 482, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 366, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 366, + 482, + 403 + ], + "type": "text", + "content": "Reconstruction. SfM pipelines either ingest a collection of unordered images [20,78,84] or an ordered image sequence [6,30,65,66,88] from a video to recover 3D structure of a scene and camera poses (\"motion\") of the images." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 403, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 533 + ], + "type": "text", + "content": "SIFT [58] and other robust descriptors allow matching image features across wide baselines enabling systems to reconstruct large-scale scenes using Internet images [80,84,85,100]. Image-to-image matches can also be regressed directly in a detector-free setup [41,87]. Feature tracks across multiple images are built from image-to-image matches. Feature tracks and estimated relative poses are used to solve for the 3D feature coordinates, camera poses and calibrations (intrinsic matrices). This geometric optimization problem is mainly solved using bundle adjustment which was explored in photogrammetry and geodesy [19, 50] and became standard in the computer vision community [40,92]. Bundle adjustment relies on the initialization being close to the solution (i.e., camera poses and 3D points are already mostly accurate)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": "There are two main approaches to this problem. Incremental SfM [6, 66] starts the reconstruction from very few images to create a high-quality seed reconstruction that progressively grows by registering more images and refining the reconstruction until convergence. Global SfM approaches solve for \"global\" poses of all images using estimates of relative poses, i.e., motion averaging [37, 38], rotation averaging [39, 59] and pose-graph optimization [21]. Various techniques were proposed to improve SfM runtime for very large sets of images [1, 2, 8, 9, 28, 36, 42, 86, 91]. Our work is similar to Incremental SfM, as we also progressively register images to the reconstructed scene starting from a seed reconstruction. However, we do not explicitly compute image matches, nor feature tracks across images, which can be computationally expensive." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 311, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 311, + 91, + 447, + 101 + ], + "type": "text", + "content": "Scene Coordinate Reconstruction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "Visual Relocalization. A reconstructed (or mapped) scene is a database of images with known camera poses. This database can be used by a visual relocalizer to estimate poses for new query images to \"relocalize\" a camera in the scene. Feature-based approaches extract 2D local features [31, 33, 56, 58, 71, 72, 87] from a query image and match them to 3D points to solve for the query pose using perspective-n-point (PnP) solvers [35], e.g., [73-75]; or match query features to 2D local features of mapped images to triangulate the query image, e.g., [104, 105]. For large scenes, matching features on a subset of database images relevant to the query can improve the speed and accuracy, e.g., [43, 67, 70, 76]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 224, + 483, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 483, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 483, + 380 + ], + "type": "text", + "content": "Some learning-based approaches encode the map of a scene in the weights of a neural network. PoseNet [47] directly regresses the absolute camera pose given an input image using a CNN that was trained on image-posed pairs of the reconstructed scene. Sattler et al. show that extrapolating outside the mapped poses can be challenging for absolute pose regression approaches [77]. Relative pose regression networks [4, 32, 51, 93] estimate the relative pose for a pair of images, allowing triangulation of the query image from multiple map images, or estimate a scale-metric relative pose w.r.t. a single map image [3]. Scene coordinate regression [12, 14, 15, 17, 22, 23, 52, 81] directly predicts 3D coordinates of a point given a patch. This approach generalizes well as the camera pose is solved using PnP [35] within a RANSAC loop [34]. ACE [12] shows that training of scene coordinate regression can be greatly accelerated. In our work we train the ACE localizer, but we start from images without poses." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 392, + 483, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 483, + 596 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 483, + 596 + ], + "type": "text", + "content": "Image-Based Rendering. In recent years, neural radiance fields (NeRFs) [61] have seen a lot of attention from the community. NeRFs allow photorealistic novel view synthesis when trained from image-pose pairs. Typically, estimation of the image poses is done in advance by using an SfM pipeline such as COLMAP, e.g., [5]. Nonetheless, research exists that estimates camera poses with NeRFs, facilitates camera localization for novel views after NeRF training [24-26, 62, 102], or simultaneously estimates camera poses during NeRF training from images alone [10, 27, 54, 82, 98]. However, these approaches either assume that the scene is captured from the front [45, 98, 101], that coarse poses are available for initialization [54], or that images are captured sequentially [10]. Techniques that rely on a multi-layer perceptron (MLP) representation of the scene, e.g. [10, 27], are slow to train, taking days to converge. While radiance fields can be trained faster using multi-layer hash grids [63] or Gaussian splats [48], their efficacy in pose estimation without approximate prior pose initialization [55, 62] or sequential image inputs [48] remains unproven. Concurrently, several learning-based camera pose estimation methods have been proposed [53, 97, 103]. Due to GPU memory constraints, these methods estimate poses for sparse sets of images." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 615, + 203, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 615, + 203, + 628 + ], + "spans": [ + { + "bbox": [ + 132, + 615, + 203, + 628 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 641, + 483, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 483, + 667 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 483, + 667 + ], + "type": "text", + "content": "Preliminaries. The input to our system is a set of RGB images, denoted by " + }, + { + "bbox": [ + 132, + 641, + 483, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{I} = \\{I_i\\}" + }, + { + "bbox": [ + 132, + 641, + 483, + 667 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 132, + 641, + 483, + 667 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 132, + 641, + 483, + 667 + ], + "type": "text", + "content": " refers to the image index. Our system estimates the corresponding" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "text", + "content": "E. Brachmann et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 114, + 476, + 300 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 476, + 300 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 476, + 300 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 476, + 300 + ], + "type": "image", + "image_path": "326d8101c386fb5c2375761039db1a5410211ba682da210bdc544e843354ef62.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 308, + 482, + 419 + ], + "lines": [ + { + "bbox": [ + 130, + 308, + 482, + 419 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 482, + 419 + ], + "type": "text", + "content": "Fig. 2: ACE0 Framework. Top left: We loop between learning a reconstruction from the current set of images and poses (\"neural mapping\"), and estimating poses of more images (\"relocalization\"). Top right: During the mapping stage, we train a scene coordinate regression network as our scene representation. Camera poses of the last relocalization round and camera calibration parameters are refined during this process. We visualize scene coordinates by mapping XYZ to the RGB cube. Bottom: In the relocalization stage, we re-estimate poses of images using the scene coordinate regression network, including images that were previously not registered to the reconstruction. If the registration of an image succeeds, it will be used in the next iteration of the mapping stage; otherwise it will not." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 440, + 482, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 440, + 482, + 487 + ], + "spans": [ + { + "bbox": [ + 130, + 440, + 482, + 487 + ], + "type": "text", + "content": "set of camera parameters, both intrinsic and extrinsic: " + }, + { + "bbox": [ + 130, + 440, + 482, + 487 + ], + "type": "inline_equation", + "content": "\\mathcal{H} = \\{(K_i,T_i)\\}" + }, + { + "bbox": [ + 130, + 440, + 482, + 487 + ], + "type": "text", + "content": ". Each " + }, + { + "bbox": [ + 130, + 440, + 482, + 487 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 130, + 440, + 482, + 487 + ], + "type": "text", + "content": " refers to a " + }, + { + "bbox": [ + 130, + 440, + 482, + 487 + ], + "type": "inline_equation", + "content": "3\\times 4" + }, + { + "bbox": [ + 130, + 440, + 482, + 487 + ], + "type": "text", + "content": " matrix containing a rotation and translation, while " + }, + { + "bbox": [ + 130, + 440, + 482, + 487 + ], + "type": "inline_equation", + "content": "K_{i}" + }, + { + "bbox": [ + 130, + 440, + 482, + 487 + ], + "type": "text", + "content": " refers to a " + }, + { + "bbox": [ + 130, + 440, + 482, + 487 + ], + "type": "inline_equation", + "content": "3\\times 3" + }, + { + "bbox": [ + 130, + 440, + 482, + 487 + ], + "type": "text", + "content": " matrix with the calibration parameters. We assume no particular image order or any prior knowledge about the pose distribution." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 487, + 482, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 482, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 482, + 536 + ], + "type": "text", + "content": "We also want to recover the 3D structure of the scene: Each pixel " + }, + { + "bbox": [ + 130, + 487, + 482, + 536 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 487, + 482, + 536 + ], + "type": "text", + "content": " in image " + }, + { + "bbox": [ + 130, + 487, + 482, + 536 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 487, + 482, + 536 + ], + "type": "text", + "content": " with 2D pixel position " + }, + { + "bbox": [ + 130, + 487, + 482, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{ij}" + }, + { + "bbox": [ + 130, + 487, + 482, + 536 + ], + "type": "text", + "content": " has a corresponding coordinate in 3D, denoted as " + }, + { + "bbox": [ + 130, + 487, + 482, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_{ij}" + }, + { + "bbox": [ + 130, + 487, + 482, + 536 + ], + "type": "text", + "content": ". The 2D pixel positions and 3D scene coordinates are related by the camera pose and the projection function " + }, + { + "bbox": [ + 130, + 487, + 482, + 536 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 130, + 487, + 482, + 536 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 261, + 543, + 481, + 557 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 543, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 261, + 543, + 481, + 557 + ], + "type": "interline_equation", + "content": "\\mathbf {p} _ {i j} = \\boldsymbol {\\pi} \\left(K _ {i}, T _ {i}, \\mathbf {y} _ {i j}\\right), \\tag {1}", + "image_path": "d62378afd92194c6f4181c3c37ddef6973f8407512d0c6d35fc438525f32f3b8.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 562, + 482, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 562, + 482, + 586 + ], + "spans": [ + { + "bbox": [ + 130, + 562, + 482, + 586 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 562, + 482, + 586 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 130, + 562, + 482, + 586 + ], + "type": "text", + "content": " maps camera coordinates to scene coordinates, and " + }, + { + "bbox": [ + 130, + 562, + 482, + 586 + ], + "type": "inline_equation", + "content": "K_{i}" + }, + { + "bbox": [ + 130, + 562, + 482, + 586 + ], + "type": "text", + "content": " projects camera coordinates to the image plane." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 586, + 484, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 586, + 484, + 622 + ], + "spans": [ + { + "bbox": [ + 130, + 586, + 484, + 622 + ], + "type": "text", + "content": "As our scene representation, we utilize a scene coordinate regression model [81], i.e., a learnable function " + }, + { + "bbox": [ + 130, + 586, + 484, + 622 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{SCR}}" + }, + { + "bbox": [ + 130, + 586, + 484, + 622 + ], + "type": "text", + "content": " that maps an image patch of image " + }, + { + "bbox": [ + 130, + 586, + 484, + 622 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 130, + 586, + 484, + 622 + ], + "type": "text", + "content": ", centered around pixel position " + }, + { + "bbox": [ + 130, + 586, + 484, + 622 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{ij}" + }, + { + "bbox": [ + 130, + 586, + 484, + 622 + ], + "type": "text", + "content": " to a scene coordinate: " + }, + { + "bbox": [ + 130, + 586, + 484, + 622 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_{ij} = f_{\\mathrm{SCR}}(\\mathbf{p}_{ij},I_i)" + }, + { + "bbox": [ + 130, + 586, + 484, + 622 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 622, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 622, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 622, + 482, + 647 + ], + "type": "text", + "content": "Given a set of 2D-3D correspondences predicted by " + }, + { + "bbox": [ + 130, + 622, + 482, + 647 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{SCR}}" + }, + { + "bbox": [ + 130, + 622, + 482, + 647 + ], + "type": "text", + "content": " for any image " + }, + { + "bbox": [ + 130, + 622, + 482, + 647 + ], + "type": "inline_equation", + "content": "I_i" + }, + { + "bbox": [ + 130, + 622, + 482, + 647 + ], + "type": "text", + "content": ", we can recover this image's camera pose " + }, + { + "bbox": [ + 130, + 622, + 482, + 647 + ], + "type": "inline_equation", + "content": "T_i" + }, + { + "bbox": [ + 130, + 622, + 482, + 647 + ], + "type": "text", + "content": " using a pose solver " + }, + { + "bbox": [ + 130, + 622, + 482, + 647 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 130, + 622, + 482, + 647 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 250, + 652, + 481, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 652, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 250, + 652, + 481, + 667 + ], + "type": "interline_equation", + "content": "T _ {i} = g \\left(K _ {i}, \\left\\{\\left(\\mathbf {p} _ {i j}, \\mathbf {y} _ {i j}\\right) \\right\\}\\right). \\tag {2}", + "image_path": "c82df813ca31a9ae569d64fb93e3461b264392dff892b0a469a2254781fc509b.jpg" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 311, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 311, + 91, + 447, + 101 + ], + "type": "text", + "content": "Scene Coordinate Reconstruction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "Since 2D-3D correspondences can be inaccurate, and contain incorrect predictions, " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " combines a PnP solver [35] with a RANSAC loop [34]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 236 + ], + "type": "text", + "content": "Normally, scene coordinate regression models are trained in a supervised fashion for the task of visual relocalization [12, 14, 15, 17, 22, 23, 52, 81]. That is, " + }, + { + "bbox": [ + 130, + 140, + 482, + 236 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{SCR}}" + }, + { + "bbox": [ + 130, + 140, + 482, + 236 + ], + "type": "text", + "content": " is trained using images with known ground truth camera parameters " + }, + { + "bbox": [ + 130, + 140, + 482, + 236 + ], + "type": "inline_equation", + "content": "\\{(I_i, T_i^{\\mathrm{GT}}, K_i^{\\mathrm{GT}})\\}" + }, + { + "bbox": [ + 130, + 140, + 482, + 236 + ], + "type": "text", + "content": ", and used for estimating the poses of unseen query images. Instead, we show how these models can be trained self-supervised, without ground truth poses, to estimate the poses of the mapping images themselves. Thus, we turn scene coordinate regression into scene coordinate reconstruction, a learning-based SfM tool." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 254, + 244, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 254, + 244, + 266 + ], + "spans": [ + { + "bbox": [ + 132, + 254, + 244, + 266 + ], + "type": "text", + "content": "3.1 Neural Mapping" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 275, + 479, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 275, + 479, + 310 + ], + "spans": [ + { + "bbox": [ + 130, + 275, + 479, + 310 + ], + "type": "text", + "content": "We train the scene coordinate regression model iteratively where we denote the current time step as " + }, + { + "bbox": [ + 130, + 275, + 479, + 310 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 275, + 479, + 310 + ], + "type": "text", + "content": " and the corresponding scene model as " + }, + { + "bbox": [ + 130, + 275, + 479, + 310 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{SCR}}^t" + }, + { + "bbox": [ + 130, + 275, + 479, + 310 + ], + "type": "text", + "content": ". We iterate between training the scene model, and registering new views, see Figure 2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 311, + 481, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 311, + 481, + 361 + ], + "spans": [ + { + "bbox": [ + 131, + 311, + 481, + 361 + ], + "type": "text", + "content": "At iteration " + }, + { + "bbox": [ + 131, + 311, + 481, + 361 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 311, + 481, + 361 + ], + "type": "text", + "content": " we assume that a subset of images has already been registered to the scene, " + }, + { + "bbox": [ + 131, + 311, + 481, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\mathrm{Reg}}^t \\subset \\mathcal{I}" + }, + { + "bbox": [ + 131, + 311, + 481, + 361 + ], + "type": "text", + "content": ", and where corresponding camera parameters, " + }, + { + "bbox": [ + 131, + 311, + 481, + 361 + ], + "type": "inline_equation", + "content": "T_{i}^{t}" + }, + { + "bbox": [ + 131, + 311, + 481, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 311, + 481, + 361 + ], + "type": "inline_equation", + "content": "K_{i}^{t}" + }, + { + "bbox": [ + 131, + 311, + 481, + 361 + ], + "type": "text", + "content": ", have already been estimated. Using these as pseudo ground truth, we train the scene model by minimizing the pixel-wise reprojection error:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 230, + 369, + 481, + 397 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 369, + 481, + 397 + ], + "spans": [ + { + "bbox": [ + 230, + 369, + 481, + 397 + ], + "type": "interline_equation", + "content": "\\sum_ {I _ {i} \\in \\mathcal {I} _ {\\text {R e g}} ^ {t}} \\sum_ {j \\in I _ {i}} \\| \\mathbf {p} _ {i j} - \\boldsymbol {\\pi} \\left(K _ {i} ^ {t}, T _ {i} ^ {t}, \\mathbf {y} _ {i j} ^ {t}\\right) \\|, \\tag {3}", + "image_path": "350eb063004e2262286abcbfdf139773afb5ee5b3a2f1ac485155305305bb379.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 407, + 367, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 367, + 420 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 367, + 420 + ], + "type": "text", + "content": "where the scene model " + }, + { + "bbox": [ + 130, + 407, + 367, + 420 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{SCR}}^{t}" + }, + { + "bbox": [ + 130, + 407, + 367, + 420 + ], + "type": "text", + "content": " predicts coordinates " + }, + { + "bbox": [ + 130, + 407, + 367, + 420 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_{ij}^{t}" + }, + { + "bbox": [ + 130, + 407, + 367, + 420 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 420, + 482, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 420, + 482, + 538 + ], + "spans": [ + { + "bbox": [ + 130, + 420, + 482, + 538 + ], + "type": "text", + "content": "Mapping Framework. We optimize Eq. 3 using stochastic gradient descent, using the fast-learning scene coordinate regressor ACE [12] (Accelerated Coordinate Encoding). ACE is trained in minutes, even for thousands of views. The training speed is important since we have to train the model in multiple iterations. ACE approximates Eq. 3 by sampling random patches from the training set, " + }, + { + "bbox": [ + 130, + 420, + 482, + 538 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\mathrm{Reg}}^t" + }, + { + "bbox": [ + 130, + 420, + 482, + 538 + ], + "type": "text", + "content": ". To do that efficiently, ACE employs a pre-trained encoder to precompute high dimensional features for a large number of training patches. The encoder stays frozen during mapping. The actual scene model, which is trained, is a multi-layer perceptron (MLP) that maps encoder features to scene coordinates; see Fig. 2 (top right) for a visual representation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 539, + 482, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 539, + 482, + 598 + ], + "spans": [ + { + "bbox": [ + 130, + 539, + 482, + 598 + ], + "type": "text", + "content": "Pose Refinement. Differently from the ACE [12] protocol, the ground truth poses " + }, + { + "bbox": [ + 130, + 539, + 482, + 598 + ], + "type": "inline_equation", + "content": "T_{i}^{\\mathrm{GT}}" + }, + { + "bbox": [ + 130, + 539, + 482, + 598 + ], + "type": "text", + "content": " are unknown during training. Instead, we have " + }, + { + "bbox": [ + 130, + 539, + 482, + 598 + ], + "type": "inline_equation", + "content": "T_{i}^{t}" + }, + { + "bbox": [ + 130, + 539, + 482, + 598 + ], + "type": "text", + "content": ", estimates based on earlier iterations of the reconstruction. Since these estimates can be inaccurate, we add the ability to refine poses during mapping. We implement refinement using an MLP:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 272, + 598, + 481, + 612 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 272, + 598, + 481, + 612 + ], + "spans": [ + { + "bbox": [ + 272, + 598, + 481, + 612 + ], + "type": "interline_equation", + "content": "T _ {i} ^ {t} = f _ {\\text {P o s e}} ^ {t} \\left(\\hat {T} _ {i} ^ {t}\\right), \\tag {4}", + "image_path": "c40360b3e4b15fe6acba675b95f717c1f3bde9893b7852e150017972b89255bd.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 616, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 616, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 616, + 482, + 667 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 616, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\tilde{T}_i^t" + }, + { + "bbox": [ + 130, + 616, + 482, + 667 + ], + "type": "text", + "content": " denotes the initial pose estimate at the start of a mapping iteration. Inspired by [106], the refinement MLP ingests " + }, + { + "bbox": [ + 130, + 616, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\tilde{T}_i^t" + }, + { + "bbox": [ + 130, + 616, + 482, + 667 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 616, + 482, + 667 + ], + "type": "inline_equation", + "content": "3\\times 4 = 12" + }, + { + "bbox": [ + 130, + 616, + 482, + 667 + ], + "type": "text", + "content": " values and predicts 12 additive offsets. We orthonormalize rotations using Gram-Schmidt [18]. We jointly optimize " + }, + { + "bbox": [ + 130, + 616, + 482, + 667 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{Pose}}^t" + }, + { + "bbox": [ + 130, + 616, + 482, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 616, + 482, + 667 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{SCR}}^t" + }, + { + "bbox": [ + 130, + 616, + 482, + 667 + ], + "type": "text", + "content": " to minimize the reprojection error of Eq. 3. We" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 250, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 250, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 250, + 101 + ], + "type": "text", + "content": "E. Brachmann et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 306, + 232 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 306, + 232 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 306, + 232 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 306, + 232 + ], + "type": "image", + "image_path": "e97ec08733e3676a4e5598bec4880fd665bb92cd0d8faad018b525ec29096720.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 243, + 482, + 321 + ], + "lines": [ + { + "bbox": [ + 130, + 243, + 482, + 321 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 482, + 321 + ], + "type": "text", + "content": "Fig. 3: Left: Pose Refinement. Since we register images based on a coarse and incomplete state of the reconstruction, we add the ability to refine poses during neural mapping. An MLP predicts pose updates relative to the initial poses, supervised by the reprojection error of scene coordinates. Right: Initialization. To start the reconstruction, we train the network using one image, the identity pose and a depth estimate, here ZoeDepth [7]. In this example, we register 33 views to the initial reconstruction. Depth estimates are only used for this step." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 309, + 114, + 480, + 232 + ], + "blocks": [ + { + "bbox": [ + 309, + 114, + 480, + 232 + ], + "lines": [ + { + "bbox": [ + 309, + 114, + 480, + 232 + ], + "spans": [ + { + "bbox": [ + 309, + 114, + 480, + 232 + ], + "type": "image", + "image_path": "8c96ff0ceff1a5888f81096c6aa623935e3ebd2db8b011fa5891e3ca5993a194.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 348, + 480, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 348, + 480, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 348, + 480, + 384 + ], + "type": "text", + "content": "show the impact of pose refinement for one iteration in Fig. 3 (left). We discard the refinement MLP after each mapping iteration. Its purpose is to enable the scene model, " + }, + { + "bbox": [ + 130, + 348, + 480, + 384 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{SCR}}^{t}" + }, + { + "bbox": [ + 130, + 348, + 480, + 384 + ], + "type": "text", + "content": ", to converge to a consistent scene representation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 386, + 482, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 386, + 482, + 470 + ], + "spans": [ + { + "bbox": [ + 130, + 386, + 482, + 470 + ], + "type": "text", + "content": "With neither poses nor 2D-3D correspondences fixed in Eq. 3, the scene coordinate regressor could drift or degenerate. As regularization, we optimize the pose refiner " + }, + { + "bbox": [ + 130, + 386, + 482, + 470 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{Pose}}^{t}" + }, + { + "bbox": [ + 130, + 386, + 482, + 470 + ], + "type": "text", + "content": " using AdamW [57] with weight decay, biasing the MLP to predict small updates relative to the initial estimates " + }, + { + "bbox": [ + 130, + 386, + 482, + 470 + ], + "type": "inline_equation", + "content": "\\tilde{T}_i^t" + }, + { + "bbox": [ + 130, + 386, + 482, + 470 + ], + "type": "text", + "content": ". This relies on the assumption that mapping images in " + }, + { + "bbox": [ + 130, + 386, + 482, + 470 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\mathrm{Reg}}^t" + }, + { + "bbox": [ + 130, + 386, + 482, + 470 + ], + "type": "text", + "content": " have been registered close to their true position. If that assumption holds, the smoothness prior of the networks [94] encourages a multi-view consistent solution as shown in previous work [12,15,17]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 472, + 482, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 532 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 532 + ], + "type": "text", + "content": "As an alternative to an MLP refiner, we could back-propagate directly to the input poses [98, 102]. However, when optimizing over thousands of views, the signal for each single pose becomes sparse. Cameras are correlated via the scene representation. If the optimization removes drift in the scene, multiple cameras need to move. The MLP refiner models the correlation of cameras." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Calibration Refinement. We do not assume information about precise calibration parameters, although often reported by devices. We do assume that the principal point is in the center, that pixels are unskewed and square. We do not model image distortion. While these are reasonable assumptions for many data regimes, we cannot rely on the focal length to be given. Thus, we refine the focal length starting from a heuristic: " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "K_{i}^{t} = f_{\\mathrm{Calib}}^{t}(K_{i}^{\\mathrm{init}})" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": ". The refinement function " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{Calib}}^{t}" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " entails a single learnable parameter " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\alpha^{t}" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " such that the focal length " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "f_{i}^{t} = f^{\\mathrm{init}} \\cdot (1 + \\alpha^{t})" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": ". As before, superscript " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " denotes the time step. We optimize " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\alpha^{t}" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " using AdamW [57] with weight decay, biasing it towards a small relative scale factor. Estimates of " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " are carried over across iterations. We set " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "f^{\\mathrm{init}}" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " to 70% of the image diagonal and it is shared by all cameras in our experiments." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 311, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 311, + 91, + 447, + 100 + ], + "type": "text", + "content": "Scene Coordinate Reconstruction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 233, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 233, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 233, + 127 + ], + "type": "text", + "content": "3.2 Relocalization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 134, + 482, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 134, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 134, + 482, + 183 + ], + "type": "text", + "content": "Given the scene model of iteration " + }, + { + "bbox": [ + 130, + 134, + 482, + 183 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 134, + 482, + 183 + ], + "type": "text", + "content": ", we attempt to register more images to determine the training set for the next mapping iteration, " + }, + { + "bbox": [ + 130, + 134, + 482, + 183 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\mathrm{Reg}}^{t + 1}" + }, + { + "bbox": [ + 130, + 134, + 482, + 183 + ], + "type": "text", + "content": ". We pass all images in " + }, + { + "bbox": [ + 130, + 134, + 482, + 183 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 130, + 134, + 482, + 183 + ], + "type": "text", + "content": " to the scene coordinate regressor " + }, + { + "bbox": [ + 130, + 134, + 482, + 183 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{SCR}}^t" + }, + { + "bbox": [ + 130, + 134, + 482, + 183 + ], + "type": "text", + "content": " to gather 2D-3D correspondences, and solve for their poses using RANSAC and PnP:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 233, + 190, + 481, + 205 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 190, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 233, + 190, + 481, + 205 + ], + "type": "interline_equation", + "content": "\\tilde {T} _ {i} ^ {t + 1}, s _ {i} ^ {t + 1} = g \\left(K _ {i} ^ {t}, \\left\\{\\left(\\mathbf {p} _ {i j}, \\mathbf {y} _ {i j} ^ {t}\\right) \\right\\}\\right). \\tag {5}", + "image_path": "3db855609de40b76c09bf3a773bf54261f87453cb673297cb8337d6a6889fe82.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 212, + 482, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 212, + 482, + 273 + ], + "spans": [ + { + "bbox": [ + 131, + 212, + 482, + 273 + ], + "type": "text", + "content": "Here, we assume that the pose solver returns a confidence score " + }, + { + "bbox": [ + 131, + 212, + 482, + 273 + ], + "type": "inline_equation", + "content": "s_i^{t+1}" + }, + { + "bbox": [ + 131, + 212, + 482, + 273 + ], + "type": "text", + "content": " alongside the pose itself that lets us decide whether the pose of the image has been estimated successfully. We simply utilize the inlier count as score " + }, + { + "bbox": [ + 131, + 212, + 482, + 273 + ], + "type": "inline_equation", + "content": "s_i^{t+1}" + }, + { + "bbox": [ + 131, + 212, + 482, + 273 + ], + "type": "text", + "content": " and apply a threshold to form the training set of the next mapping iteration: " + }, + { + "bbox": [ + 131, + 212, + 482, + 273 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\mathrm{Reg}}^{t+1} = \\left\\{ I_i | s_i^{t+1} > \\tau_s \\right\\}" + }, + { + "bbox": [ + 131, + 212, + 482, + 273 + ], + "type": "text", + "content": ". The relocalization process is depicted in Figure 2, bottom." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 289, + 226, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 226, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 226, + 300 + ], + "type": "text", + "content": "3.3 Initialization" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 306, + 481, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 306, + 481, + 333 + ], + "spans": [ + { + "bbox": [ + 130, + 306, + 481, + 333 + ], + "type": "text", + "content": "We start the reconstruction with one image: " + }, + { + "bbox": [ + 130, + 306, + 481, + 333 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\mathrm{Reg}}^0 = \\{I_{\\mathrm{seed}}\\}" + }, + { + "bbox": [ + 130, + 306, + 481, + 333 + ], + "type": "text", + "content": ". We set the seed pose " + }, + { + "bbox": [ + 130, + 306, + 481, + 333 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{seed}}^0" + }, + { + "bbox": [ + 130, + 306, + 481, + 333 + ], + "type": "text", + "content": " to identity, and we initialize the calibration " + }, + { + "bbox": [ + 130, + 306, + 481, + 333 + ], + "type": "inline_equation", + "content": "K_{\\mathrm{seed}}^{\\mathrm{init}}" + }, + { + "bbox": [ + 130, + 306, + 481, + 333 + ], + "type": "text", + "content": " as explained above." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 333, + 481, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 333, + 481, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 333, + 481, + 415 + ], + "type": "text", + "content": "We cannot train a scene coordinate regression network using Eq. 3 with a single image. The reprojection error is ill-conditioned without multiple views constraining the depth. Therefore, we optimize a different objective in the seed iteration, inspired by Map-free Relocalization [3]. Arnold et al. argue that a single image and a depth estimate allow to relocalize query images, albeit with limited accuracy. Our experiments show that such coarse relocalizations of a few images suffice as initialization for optimizing the reprojection error of Eq. 3." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 416, + 481, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 416, + 481, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 416, + 481, + 475 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 130, + 416, + 481, + 475 + ], + "type": "inline_equation", + "content": "d_{ij}" + }, + { + "bbox": [ + 130, + 416, + 481, + 475 + ], + "type": "text", + "content": " be a depth value predicted for pixel " + }, + { + "bbox": [ + 130, + 416, + 481, + 475 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 416, + 481, + 475 + ], + "type": "text", + "content": " of image " + }, + { + "bbox": [ + 130, + 416, + 481, + 475 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 416, + 481, + 475 + ], + "type": "text", + "content": ". We derive a target scene coordinate by back-projection as " + }, + { + "bbox": [ + 130, + 416, + 481, + 475 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}}_{ij} = d_{ij}(K_{\\mathrm{seed}}^{\\mathrm{init}})^{-1}\\mathbf{p}_{ij}" + }, + { + "bbox": [ + 130, + 416, + 481, + 475 + ], + "type": "text", + "content": ". We train an initial scene coordinate regression network " + }, + { + "bbox": [ + 130, + 416, + 481, + 475 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{SCR}}^{0}" + }, + { + "bbox": [ + 130, + 416, + 481, + 475 + ], + "type": "text", + "content": " by optimizing " + }, + { + "bbox": [ + 130, + 416, + 481, + 475 + ], + "type": "inline_equation", + "content": "\\sum_{j\\in I^0}\\| \\hat{\\mathbf{y}}_{ij} - \\mathbf{y}_{ij}\\|" + }, + { + "bbox": [ + 130, + 416, + 481, + 475 + ], + "type": "text", + "content": ". Fig. 3 (right) shows a scene coordinate point cloud learned from a depth estimate, and successful relocalization against it." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 475, + 481, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 475, + 481, + 571 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 481, + 571 + ], + "type": "text", + "content": "We found our pipeline to be robust w.r.t. selecting the seed image. However, when a randomly selected image has little to no visual overlap with the remaining images, the whole reconstruction would fail. Selecting an unfortunate seed image, e.g., at the very end of a long camera trajectory, can increase reconstruction times. To decrease the probability of such incidents, we try 5 random seed images and choose the one with the highest relocalization rate across 1000 other mapping images. Since mapping seed images is fast, ca. 1 min on average, this poses no significant computational burden." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 587, + 241, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 587, + 241, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 587, + 241, + 599 + ], + "type": "text", + "content": "3.4 Implementation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": "We base our pipeline on the public code of the ACE relocalizer [12]. ACE uses a convolutional feature backbone, pre-trained on ScanNet [29], that ingests images scaled to 480px height. On top of the backbone is the mapping network, a 9-layer MLP with 512 channels, consuming 4MB of weights in 16-bit floating point precision. This is our scene representation." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 250, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 250, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 250, + 101 + ], + "type": "text", + "content": "E. Brachmann et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "The ACE training process is reasonably fast, taking 5 minutes to train the scene network. Since we repeat training the scene network in multiple iterations, we expedite the process further to decrease our total reconstruction time." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 481, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 481, + 236 + ], + "type": "text", + "content": "Adaptive Sampling of Training Patches. ACE trains the scene representation based on 8M patches sampled from the mapping images, a process that takes 1 minute. This is excessive when having very few mapping images in the beginning of the reconstruction, thus we loop over the mapping images at most 10 times when sampling patches, or until 8M patches have been sampled. Using this adaptive strategy, sampling patches for the seed reconstruction where we have 1 image only takes 2 seconds instead of 1 minute." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 237, + 481, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 481, + 345 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 481, + 345 + ], + "type": "text", + "content": "Adaptive Stopping. ACE trains the scene model using a fixed one-cycle learning rate schedule [83] to a total of 25k parameter updates. We make the training schedule adaptive since the network is likely to converge fast when trained on few images only. We monitor the reprojection error of scene coordinates within a mini-batch. If for 100 consecutive batches " + }, + { + "bbox": [ + 130, + 237, + 481, + 345 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 130, + 237, + 481, + 345 + ], + "type": "text", + "content": " of reprojection errors are below an inlier threshold of 10px, we stop training early. We approximate the one-cycle learning rate schedule in a linear fashion: we increase the learning rate in the first 1k iteration from " + }, + { + "bbox": [ + 130, + 237, + 481, + 345 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-4}" + }, + { + "bbox": [ + 130, + 237, + 481, + 345 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 237, + 481, + 345 + ], + "type": "inline_equation", + "content": "3 \\times 10^{-3}" + }, + { + "bbox": [ + 130, + 237, + 481, + 345 + ], + "type": "text", + "content": " and when the early stopping criterion has been met, we decrease the learning rate to " + }, + { + "bbox": [ + 130, + 237, + 481, + 345 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-4}" + }, + { + "bbox": [ + 130, + 237, + 481, + 345 + ], + "type": "text", + "content": " within 5k iterations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 346, + 481, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 346, + 481, + 371 + ], + "spans": [ + { + "bbox": [ + 131, + 346, + 481, + 371 + ], + "type": "text", + "content": "We report more implementation details and hyper-parameters in the supplement. Our code is also publicly available to ensure reproducibility." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 394, + 230, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 394, + 230, + 407 + ], + "spans": [ + { + "bbox": [ + 131, + 394, + 230, + 407 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 423, + 482, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 482, + 484 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 482, + 484 + ], + "type": "text", + "content": "We refer to our SfM pipeline as ACE0 (ACE Zero) since it builds on top of the ACE relocalizer [12] but adds the ability to train from scratch, without poses. We demonstrate the effectiveness of our approach on three datasets and 31 scenes in total. For all experiments, we rely on ZoeDepth [7] to initialize our reconstructions. All timings reported for ACE0 are based on a single V100 GPU." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 484, + 481, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 484, + 481, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 484, + 481, + 581 + ], + "type": "text", + "content": "Baselines. We consider the pose estimates of COLMAP with default parameters as our pseudo ground truth, obtained by extensive processing. We also run COLMAP with parameter settings recommended for large image collections of 1k images and more [79]. This variation achieves much faster processing (denoted COLMAP fast). We use a V100 GPU for COLMAP feature extraction and matching, and we specify the exact parameters of COLMAP in the supplement. Furthermore, we show some results of RealityCapture [68], an efficient commercial feature-based SfM pipeline." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": "We compare to learning-based SfM approaches that are not restricted to few-frame scenarios, namely to NeRF-based BARF [54] and NoPe-NeRF [10]. We also compare to DUSt3R [97], a non-NeRF learning-based SfM method. While we focus on reconstructing unsorted image collections, the datasets we consider allow for sequential processing. Hence, for context, we also show results of DROID-SLAM [90], a neural SLAM approach. Unless specified otherwise, we report timings based on a single V100 GPU." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 311, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 311, + 91, + 447, + 101 + ], + "type": "text", + "content": "Scene Coordinate Reconstruction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 117, + 482, + 301 + ], + "blocks": [ + { + "bbox": [ + 134, + 117, + 482, + 301 + ], + "lines": [ + { + "bbox": [ + 134, + 117, + 482, + 301 + ], + "spans": [ + { + "bbox": [ + 134, + 117, + 482, + 301 + ], + "type": "image", + "image_path": "7c1dc002114a01dbd79150f763da2146f7286fb718e64ef0d1745a6b71872aab.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 311, + 482, + 367 + ], + "lines": [ + { + "bbox": [ + 130, + 311, + 482, + 367 + ], + "spans": [ + { + "bbox": [ + 130, + 311, + 482, + 367 + ], + "type": "text", + "content": "Fig. 4: Reconstructed Poses. We show poses estimated by ACE0 for a selection of scenes. We color code the reconstruction iteration in which a particular view has been registered. We show the ACE0 point cloud as a representation of the scene. The seed image is shown as a gray frustum. We also compare our poses to poses estimated by COLMAP (Mip-NeRF 360, Tanks and Temples) and KinectFusion (7-Scenes)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 390, + 482, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 482, + 474 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 482, + 474 + ], + "type": "text", + "content": "Benchmark. We show results on 7-Scenes [81], a relocalization dataset, on Mip-NeRF 360 [5], a view synthesis dataset and on Tanks and Temples [49], a reconstruction dataset. Comparing poses on these datasets is problematic, as our pseudo ground truth is estimated rather than measured. For example, an approach might be more accurate than COLMAP on individual scenes. Computing pose errors w.r.t. COLMAP would result in incorrect conclusions. Therefore, we gauge the pose quality in a self-supervised way, using novel view synthesis [96]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 475, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 475, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 482, + 666 + ], + "type": "text", + "content": "We let each method estimate the poses of all images of a scene. For evaluation, we split the images into training and test sets. We train a Nerfacto [89] model on the training set, and synthesize views for the test poses. We compare the synthesized images to the test images and report the difference as peak signal-to-noise ratio (PSNR). To ensure a fair comparison to NeRF-based competitors, we use these methods in the same way as the other SfM methods: we run them to pose all images, and train a Nerfacto model on top of their poses. This is to ensure that we solely compare the pose quality across methods, and not different capabilities in view synthesis. The quality of poses affects the PSNR numbers in two ways: Good training poses let the NeRF model fit a consistent scene representation. Good testing poses make sure that the synthesized images are aligned with the original image. We explain further details in the supplement which also includes additional, perceptual metrics. In spirit, our evaluation is similar to the Tanks and Temples benchmark which evaluates a derived scene mesh rather than camera poses. However, our evaluation can be applied to arbitrary datasets as it does not need ground truth geometry." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 250, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 250, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 250, + 101 + ], + "type": "text", + "content": "E. Brachmann et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 120, + 479, + 213 + ], + "blocks": [ + { + "bbox": [ + 132, + 120, + 479, + 213 + ], + "lines": [ + { + "bbox": [ + 132, + 120, + 479, + 213 + ], + "spans": [ + { + "bbox": [ + 132, + 120, + 479, + 213 + ], + "type": "table", + "html": "
FramesPseudo Ground TruthAll Frames200 Frames50 Frames
Kinect FusionCOLMAP (default)COLMAP (fast)DROID-SLAM† [90]ACE0 (ours)KF+ACE0 (ours)BARF [54]NoPE-NeRF† [10]ACE0 (ours)DUS3R [97]ACE0 (Ours)
Chess6k19.623.623.519.323.323.012.812.622.718.919.2
Fire4k19.222.622.613.022.322.312.711.822.118.819.5
Heads2k17.018.818.917.618.819.110.711.819.918.421.3
Office10k18.921.421.6failed21.121.511.910.919.812.513.7
Pumpkin6k19.924.123.818.324.123.819.614.224.721.722.3
RedKitchen12k17.621.421.410.920.820.911.611.218.913.813.7
Stairs3k19.016.721.013.017.719.915.815.918.815.315.4
Average18.721.221.8N/A21.221.513.612.621.017.117.9
Avg. Timerealtime38h13h18min1h7min8.5h47h27min4min*16min
", + "image_path": "bb9a0f72490a7c8e4033e223ec44867511e194457d6e99ee8fd67bb9146b8ed6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 132, + 213, + 481, + 278 + ], + "lines": [ + { + "bbox": [ + 132, + 213, + 481, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 213, + 481, + 278 + ], + "type": "text", + "content": "Table 1: 7-Scenes. We show the pose accuracy via view synthesis with Nerfacto [89] as PSNR in dB, and the reconstruction time. Results for All Frames are color coded w.r.t. similarity to the COLMAP pseudo ground truth: " + }, + { + "bbox": [ + 132, + 213, + 481, + 278 + ], + "type": "inline_equation", + "content": ">0.5" + }, + { + "bbox": [ + 132, + 213, + 481, + 278 + ], + "type": "text", + "content": " dB better within " + }, + { + "bbox": [ + 132, + 213, + 481, + 278 + ], + "type": "inline_equation", + "content": "\\pm 0.5" + }, + { + "bbox": [ + 132, + 213, + 481, + 278 + ], + "type": "text", + "content": " dB " + }, + { + "bbox": [ + 132, + 213, + 481, + 278 + ], + "type": "inline_equation", + "content": ">0.5" + }, + { + "bbox": [ + 132, + 213, + 481, + 278 + ], + "type": "text", + "content": " dB worse " + }, + { + "bbox": [ + 132, + 213, + 481, + 278 + ], + "type": "inline_equation", + "content": ">1" + }, + { + "bbox": [ + 132, + 213, + 481, + 278 + ], + "type": "text", + "content": " dB worse. For some competitors, we had to sub-sample the images due to their computational complexity (right side). Method needs sequential inputs. *Results on more powerful hardware." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 312, + 203, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 312, + 203, + 323 + ], + "spans": [ + { + "bbox": [ + 132, + 312, + 203, + 323 + ], + "type": "text", + "content": "4.1 7-Scenes" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 330, + 481, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 330, + 481, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 330, + 481, + 426 + ], + "type": "text", + "content": "The 7-Scenes dataset [81] consists of seven indoor scenes, scanned with a Kinect v1 camera. Multiple, disconnected scans are provided for each scene to a total of " + }, + { + "bbox": [ + 130, + 330, + 481, + 426 + ], + "type": "inline_equation", + "content": "2\\mathrm{k} - 12\\mathrm{k}" + }, + { + "bbox": [ + 130, + 330, + 481, + 426 + ], + "type": "text", + "content": " images. For each method, we assume a shared focal length across scans and initialize with the default calibration of a Kinect v1. The dataset comes with pseudo ground truth camera poses estimated by KinectFusion [44, 64], a depth-based SLAM system. Individual scans were registered but not bundle-adjusted [13]. Inspired by [13], we recompute alternative, bundle-adjusted pseudo ground truth by running COLMAP with default parameters." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 426, + 481, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 481, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 481, + 545 + ], + "type": "text", + "content": "Discussion. We show results in Table 1. Of both pseudo ground truth versions, KinectFusion achieves lower PSNR numbers than COLMAP, presumably due to the lack of global optimization. COLMAP with fast parameters shows PSNR numbers similar to COLMAP with default parameters, on average. Both versions of running COLMAP take considerable time to reconstruct each scene. We note that COLMAP has been optimised for quality, rather than speed. Not all acceleration strategies from the feature-based SfM literature have been implemented in COLMAP, so presumably comparable quality can be obtained faster. DROID-SLAM [90] does not perform well on 7-Scenes and partially fails altogether, presumably due to the jumps between individual scans of each scene." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "content": "Our approach, ACE0, achieves a pose quality comparable to the COLMAP pseudo ground truth while reconstructing each scene in " + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\sim 1" + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "content": " hour despite the large number of images. We show qualitative examples in Figure 4 and in the supplement. We also demonstrate that ACE0 can swiftly optimize an initial set of approximate poses. When starting from KinectFusion poses, ACE0 increases PSNR significantly in less than 10 minutes per scene, see “" + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{KF} + \\mathrm{ACE0}" + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "content": "” in Table 1. In the supplement, we include a parameter study on 7-Scenes to show that ACE0 is robust to the choice of depth estimator. We also show the positive impact of pose refinement on the reconstruction quality as well as the reconstruction speedup due to our early stopping schedule." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 311, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 311, + 91, + 447, + 101 + ], + "type": "text", + "content": "Scene Coordinate Reconstruction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 114, + 278, + 191 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 278, + 191 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 278, + 191 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 278, + 191 + ], + "type": "table", + "html": "
ACE [12]ACE [12]ACE0 (ours)
SupervisionKinectFusionCOLMAPself-supervised
Chess96.0%100.0%100.0%
Fire98.4%99.5%98.8%
Heads100.0%100.0%100.0%
Office36.9%100.0%99.1%
Pumpkin47.3%100.0%99.9%
Redlitchen47.8%98.9%98.1%
Stairs74.1%85.0%61.0%
Average71.5%97.6%93.8%
", + "image_path": "73c890157e77ec40554c46773547afaf2ca109859695b533d7aa2987b73a1723.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 281, + 114, + 482, + 191 + ], + "blocks": [ + { + "bbox": [ + 132, + 192, + 277, + 217 + ], + "lines": [ + { + "bbox": [ + 132, + 192, + 277, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 192, + 277, + 217 + ], + "type": "text", + "content": "Table 2 (a): Relocalization on 7-Scenes. % poses below " + }, + { + "bbox": [ + 132, + 192, + 277, + 217 + ], + "type": "inline_equation", + "content": "5\\mathrm{cm},5^{\\circ}" + }, + { + "bbox": [ + 132, + 192, + 277, + 217 + ], + "type": "text", + "content": " error, computed w.r.t. COLMAP pseudo GT." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 281, + 114, + 482, + 191 + ], + "lines": [ + { + "bbox": [ + 281, + 114, + 482, + 191 + ], + "spans": [ + { + "bbox": [ + 281, + 114, + 482, + 191 + ], + "type": "table", + "html": "
Pseudo GT (COLMAP)DROID-SLAM† [90]BARF [54]NoPe-NeRF† [10]ACEO (ours)
Bicycle21.510.911.912.218.7
Bonsai27.610.912.514.825.8
Counter25.512.911.911.624.5
Garden26.316.713.313.825.0
Kitchen27.413.913.314.426.1
Room28.011.311.914.319.8
Stump16.813.915.013.720.5
Average24.712.912.813.522.9
", + "image_path": "283dcc312fb1a9015b1616daadce9fafc63f75e32aa9535cb43dcdc82d563f12.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 284, + 192, + 478, + 218 + ], + "lines": [ + { + "bbox": [ + 284, + 192, + 478, + 218 + ], + "spans": [ + { + "bbox": [ + 284, + 192, + 478, + 218 + ], + "type": "text", + "content": "Table 2 (b): Mip-NeRF 360. Pose quality in PSNR, higher is better. Best in bold. Method needs sequential inputs." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 249, + 480, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 249, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 249, + 480, + 369 + ], + "type": "text", + "content": "For our learning-based competitors, we sub-sampled images due to their computational constraints, see right side of Table 1. Even using only 200 images, NoPe-NeRF [10] takes 2 days to fit a model and estimate poses. Despite these long processing times, we observe poor pose quality of BARF and NoPe-NeRF in terms of PSNR. BARF [54] requires pose initialisation. We provide identity poses since the scenes of 7Scenes are roughly forward-facing. Still, the camera motions are too complex for BARF to handle. NoPe-NeRF does not require pose initialisation but needs roughly sequential images, which we did provide. NoPe-NeRF relies upon successive images having similar poses to perform well, and struggles with the large jumps between the subsampled images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 369, + 481, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 481, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 481, + 416 + ], + "type": "text", + "content": "For the comparison with DUSt3R [97], we had to subsample the sequences further, as we were only able to run it with 50 images at most, even when using an A100 GPU with 40GB memory. DUSt3R achieves reasonable PSNR numbers but consistently lower than ACE0." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 417, + 482, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 417, + 482, + 561 + ], + "spans": [ + { + "bbox": [ + 130, + 417, + 482, + 561 + ], + "type": "text", + "content": "Relocalization. ACE0 is a learning-based SfM tool but it is also a self-supervised visual relocaliser. In Table 2 (a), we compare it to the supervised relocalizer ACE [12]. Using the scale-metric pseudo ground truth of [13], we train ACE with COLMAP mapping poses, and evaluate it against COLMAP query poses. Unsurprisingly, ACE achieves almost perfect relocalization under the usual " + }, + { + "bbox": [ + 130, + 417, + 482, + 561 + ], + "type": "inline_equation", + "content": "5\\mathrm{cm}" + }, + { + "bbox": [ + 130, + 417, + 482, + 561 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 417, + 482, + 561 + ], + "type": "inline_equation", + "content": "5^{\\circ}" + }, + { + "bbox": [ + 130, + 417, + 482, + 561 + ], + "type": "text", + "content": " error threshold. Interestingly, ACE0 achieves almost identical results when mapping the scene self-supervised, and evaluating the relocalized query poses against the COLMAP pseudo ground truth. For context, when training ACE with Kinect-Fusion mapping poses and evaluating against COLMAP pseudo ground truth, results are far worse. This signifies that ACE0 mapping poses are very similar to the COLMAP mapping poses, and less similar to KinectFusion mapping poses. We give more details about this experiment in the supplement." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 575, + 237, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 575, + 237, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 575, + 237, + 588 + ], + "type": "text", + "content": "4.2 Mip-NeRF 360" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 593, + 480, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 593, + 480, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 593, + 480, + 641 + ], + "type": "text", + "content": "The Mip-NeRF 360 dataset [5] consists of seven small-scale scenes, both indoor and outdoor. The dataset was reconstructed with COLMAP and comes with intrinsics (which we ignore), camera poses and undistorted (pinhole camera) images. For each method, we assume a shared focal length per scene." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Discussion. We present PSNR results in Table 2 (b). NoPe-NeRF does not perform well on this dataset, despite processing each scene for 2 days. The differences" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 250, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 250, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 250, + 101 + ], + "type": "text", + "content": "E. Brachmann et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 121, + 479, + 347 + ], + "blocks": [ + { + "bbox": [ + 132, + 121, + 479, + 347 + ], + "lines": [ + { + "bbox": [ + 132, + 121, + 479, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 121, + 479, + 347 + ], + "type": "table", + "html": "
FramesDROID-FramesDROID-SparseSparse
COLMAP (default)Reality CaptureSLAM† [90]ACE0 (ours)COLMAP (fast)Reality CaptureSLAM† [90]ACE0 (ours)COLMAP + Reloc + BACOLMAP + ACE0 (ours)
TrainingBarn41024.021.219.016.512.2k24.416.913.517.726.325.1
Caterpillar38317.115.916.616.911.4k18.617.918.918.618.718.8
Church50718.317.614.317.219.3k12.1-11.516.518.517.3
Ignatius26420.117.717.819.87.8k20.818.619.120.720.920.7
Meetingroom37118.618.115.618.011.1k19.418.217.116.620.820.3
Truck25121.119.018.320.17.5k23.619.120.623.023.423.1
Average36419.918.216.918.114.6k19.818.216.818.921.420.9
Avg Time1h3min5min1.1h74h14h18min2.2h8h1.8h
IntermediateFamily15219.518.817.619.04.4k21.219.819.818.021.321.3
Francis30221.620.720.720.17.8k19.920.421.821.722.522.7
Horse15119.219.016.319.56.0k21.620.719.221.722.622.3
Lighthouse30916.616.513.617.58.3k19.016.618.918.619.520.5
Playground30719.119.211.418.77.7k17.916.511.320.421.221.0
Train30116.815.413.816.212.6k19.614.415.618.519.818.5
Average25418.818.315.618.57.8k19.918.117.819.821.121.0
Avg Time32min2min3min1.3h48h11h14min2.2h5h1h
AdvancedAuditorium30219.612.216.718.713.6k13.7-16.620.021.419.8
Ballroom32416.318.313.117.910.8k17.2-10.418.918.015.6
Courtroom30118.217.212.317.112.6k14.6-10.216.318.717.8
Palace50914.211.710.810.721.9k13.8-8.611.015.312.3
Temple30218.115.711.89.717.5k13.3-11.914.819.616.1
Average34817.315.012.914.815.6k14.5-11.516.218.616.3
Avg Time1h2min4min1h71h27min2.8h10h2.1h
", + "image_path": "ef9ff5cc88193a46650a64c193de4569f68bdb0d49606daaaa34589846ed37a8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 348, + 482, + 403 + ], + "lines": [ + { + "bbox": [ + 132, + 348, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 132, + 348, + 482, + 403 + ], + "type": "text", + "content": "Table 3: Tanks and Temples. We show the pose accuracy via view synthesis with Nerfacto [89] as PSNR in dB, and the reconstruction time. We color code results compared to COLMAP, default and fast, respectively: " + }, + { + "bbox": [ + 132, + 348, + 482, + 403 + ], + "type": "inline_equation", + "content": ">0.5" + }, + { + "bbox": [ + 132, + 348, + 482, + 403 + ], + "type": "text", + "content": " dB better within " + }, + { + "bbox": [ + 132, + 348, + 482, + 403 + ], + "type": "inline_equation", + "content": "\\pm 0.5" + }, + { + "bbox": [ + 132, + 348, + 482, + 403 + ], + "type": "text", + "content": " dB " + }, + { + "bbox": [ + 132, + 348, + 482, + 403 + ], + "type": "inline_equation", + "content": ">0.5" + }, + { + "bbox": [ + 132, + 348, + 482, + 403 + ], + "type": "text", + "content": " dB worse " + }, + { + "bbox": [ + 132, + 348, + 482, + 403 + ], + "type": "inline_equation", + "content": ">1" + }, + { + "bbox": [ + 132, + 348, + 482, + 403 + ], + "type": "text", + "content": " dB worse. Method needs sequential inputs." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 436, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 436, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 436, + 482, + 510 + ], + "type": "text", + "content": "between sequential images are too large. DROID-SLAM fails for the same reason. BARF performs poorly because the identity pose is not a good initialisation for most scenes. In contrast, ACE0 reconstructs the dataset successfully. While it achieves slightly lower PSNR than COLMAP, its pose estimates are similar, " + }, + { + "bbox": [ + 130, + 436, + 482, + 510 + ], + "type": "inline_equation", + "content": "cf" + }, + { + "bbox": [ + 130, + 436, + 482, + 510 + ], + "type": "text", + "content": ". Figure 4. The supp. shows that synthesized images based on ACE0 are visually close to those of COLMAP while our learning-based competitors are far off." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 526, + 259, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 526, + 259, + 539 + ], + "spans": [ + { + "bbox": [ + 132, + 526, + 259, + 539 + ], + "type": "text", + "content": "4.3 Tanks and Temples" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "The Tanks and Temples dataset [49] contains 21 diverse scenes, both indoors and outdoors, and with varying spatial extent. We remove two scenes (Panther, M60) with military associations. We also remove two scenes (Courthouse, Museum) where the COLMAP baseline did not finish the reconstruction after 5 days of processing or ran out of memory. For the latter two scenes, we provide ACE0 results in the supplement. The dataset provides 150-500 images per scene but also the original videos as a source for more frames. Thus, we consider each scene in two versions: Using 150-500 images and using " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "4\\mathrm{k} - 22\\mathrm{k}" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " frames. For all methods, we assume a pinhole camera model with shared focal length across images, and images to be unordered. We found none of the learning-based SfM competitors" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 311, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 311, + 91, + 447, + 101 + ], + "type": "text", + "content": "Scene Coordinate Reconstruction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 481, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 481, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 481, + 152 + ], + "type": "text", + "content": "to be applicable to this dataset. NoPe-NeRF would run multiple days per scene, even when considering only a few hundred images. DUSt3R would run out of memory. BARF needs reasonable pose initializations which are not available." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 481, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 481, + 224 + ], + "type": "text", + "content": "Discussion. We show PSNR numbers of RealityCapture, DROID-SLAM and ACE0 in Table 3, color-coded by similarity to the COLMAP pseudo GT. ACE0 achieves reasonable results when reconstructing scenes from a few hundred images (Table 3, left). ACE0 generally predicts plausible poses (cf., Figure 4), even if PSNR numbers are sometimes lower than those of COLMAP. RealityCapture performs similar to ACE0 while DROID-SLAM struggles on the sparse images." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 225, + 481, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 225, + 481, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 225, + 481, + 308 + ], + "type": "text", + "content": "Next, we consider more than 1k images per scene (right side of Table 3). Here, we run COLMAP with parameters tuned for large images collections (fast) due to the extremely large image sets. ACE0 offers a reconstruction quality comparable to COLMAP on average while also being fast. We run RealityCapture on some of the scenes but it produces fractured reconstructions for these large images sets, leading to low PSNR numbers. DROID-SLAM still struggles on many scenes despite having access to sequential images that are temporally close." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 308, + 481, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 481, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 481, + 415 + ], + "type": "text", + "content": "In the two rightmost columns, we initialise with a sparse COLMAP reconstruction from 150-500 images, and extend and refine it using all available frames. Firstly, using a feature-based baseline, we register the full set of frames using the relocalization mode of COLMAP, followed by a final round of bundle adjustment. Secondly, we run ACE0 initialised with the poses of the sparse COLMAP reconstruction. Both variants are considerably faster than running COLMAP from scratch on the full set of frames. ACE0 is able to register and refine all additional frames in 1-2 hours, on average. Again, we find the pose quality of ACE0 comparable to the feature-based alternative." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 435, + 326, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 435, + 326, + 449 + ], + "spans": [ + { + "bbox": [ + 131, + 435, + 326, + 449 + ], + "type": "text", + "content": "5 Conclusion and Future Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 461, + 481, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 481, + 544 + ], + "type": "text", + "content": "We have presented scene coordinate reconstruction, a new approach to learning-based SfM. We learn an implicit, neural scene representation from a set of unposed images. Our method, ACE0, is able to reconstruct a wide variety of scenes. In many cases, the accuracy of estimated poses is close to that of COLMAP and synthesized images visually similar. Unlike previous learning-based SfM methods, ACE0 can be applied to multiple thousand unsorted images, without pose priors, and reconstructs them within a few hours." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 481, + 665 + ], + "type": "text", + "content": "Limitations. We show some failure cases in the supplement. Scene coordinate regression struggles with repetitive structures since the network is not able to make multi-modal predictions for visually ambiguous inputs. Scene coordinate regression also struggles with representing large areas. The common solution is to use network ensembles based on pre-clustering of the scene [12, 16] which is difficult in a reconstruction setting. While scene coordinate regression generalises quite well, it has difficulties to bridge extreme view point or lighting changes, such as day versus night. In our experiments, we assumed a simple pinhole camera model with shared intrinsics across images. To the best of our knowledge, scene coordinate regression has not been coupled with image distortion, thus far." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 250, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 250, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 250, + 101 + ], + "type": "text", + "content": "E. Brachmann et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 142, + 140, + 479, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 140, + 479, + 161 + ], + "spans": [ + { + "bbox": [ + 142, + 140, + 479, + 161 + ], + "type": "text", + "content": "1. Agarwal, S., Furukawa, Y., Snavely, N., Simon, I., Curless, B., Seitz, S.M., Szeliski, R.: Building Rome in a day. ACM TOG (2011)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 142, + 162, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 162, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 142, + 162, + 480, + 183 + ], + "type": "text", + "content": "2. Agarwal, S., Snavely, N., Seitz, S.M., Szeliski, R.: Bundle adjustment in the large. In: ECCV (2010)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 142, + 184, + 480, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 184, + 480, + 216 + ], + "spans": [ + { + "bbox": [ + 142, + 184, + 480, + 216 + ], + "type": "text", + "content": "3. Arnold, E., Wynn, J., Vicente, S., Garcia-Hernando, G., Monszpart, A., Prisacariu, V.A., Turmukhambetov, D., Brachmann, E.: Map-free visual relocalization: Metric pose relative to a single image. In: ECCV (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 142, + 217, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 217, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 142, + 217, + 480, + 237 + ], + "type": "text", + "content": "4. Balntas, V., Li, S., Prisacariu, V.A.: RelocNet: Continuous metric learning relocalisation using neural nets. In: ECCV (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 142, + 238, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 238, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 142, + 238, + 480, + 259 + ], + "type": "text", + "content": "5. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-NeRF 360: Unbounded anti-aliased neural radiance fields. In: CVPR (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 142, + 261, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 261, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 142, + 261, + 480, + 281 + ], + "type": "text", + "content": "6. Beardsley, P.A., Zisserman, A., Murray, D.W.: Sequential updating of projective and affine structure from motion. IJCV (1997)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 142, + 282, + 480, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 282, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 142, + 282, + 480, + 304 + ], + "type": "text", + "content": "7. Bhat, S.F., Birkl, R., Wofk, D., Wonka, P., Müller, M.: ZoeDepth: Zero-shot transfer by combining relative and metric depth. arXiv (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 142, + 304, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 304, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 142, + 304, + 480, + 335 + ], + "type": "text", + "content": "8. Bhowmick, B., Patra, S., Chatterjee, A., Govindu, V.M., Banerjee, S.: Divide and conquer: Efficient large-scale structure from motion using graph partitioning. In: ACCV (2015)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 142, + 337, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 337, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 142, + 337, + 480, + 369 + ], + "type": "text", + "content": "9. Bhowmick, B., Patra, S., Chatterjee, A., Govindu, V.M., Banerjee, S.: Divide and conquer: A hierarchical approach to large-scale structure-from-motion. CVIU (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 370, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 370, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 138, + 370, + 480, + 391 + ], + "type": "text", + "content": "10. Bian, W., Wang, Z., Li, K., Bian, J.W., Prisacariu, V.A.: NoPe-NeRF: Optimising neural radiance field with no pose prior. In: CVPR (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 392, + 480, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 392, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 138, + 392, + 480, + 423 + ], + "type": "text", + "content": "11. Bloesch, M., Czarnowski, J., Clark, R., Leutenegger, S., Davison, A.J.: CodeSLAM — Learning a compact, optimisable representation for dense visual SLAM. In: CVPR (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 425, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 425, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 138, + 425, + 480, + 445 + ], + "type": "text", + "content": "12. Brachmann, E., Cavallari, T., Prisacariu, V.A.: Accelerated coordinate encoding: Learning to relocalize in minutes using RGB and poses. In: CVPR (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 447, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 447, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 447, + 480, + 468 + ], + "type": "text", + "content": "13. Brachmann, E., Humenberger, M., Rother, C., Sattler, T.: On the limits of pseudo ground truth in visual camera re-localisation. In: ICCV (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 469, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 469, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 138, + 469, + 480, + 500 + ], + "type": "text", + "content": "14. Brachmann, E., Krull, A., Nowozin, S., Shotton, J., Michel, F., Gumhold, S., Rother, C.: DSAC-differentiable RANSAC for camera localization. In: CVPR (2017)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 502, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 502, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 138, + 502, + 480, + 522 + ], + "type": "text", + "content": "15. Brachmann, E., Rother, C.: Learning less is more-6D camera localization via 3D surface regression. In: CVPR (2018)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 523, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 523, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 523, + 480, + 544 + ], + "type": "text", + "content": "16. Brachmann, E., Rother, C.: Expert sample consensus applied to camera relocalization. In: ICCV (2019)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 545, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 545, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 138, + 545, + 480, + 566 + ], + "type": "text", + "content": "17. Brachmann, E., Rother, C.: Visual camera re-localization from RGB and RGB-D images using DSAC. IEEE TPAMI (2021)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 567, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 480, + 588 + ], + "type": "text", + "content": "18. Brégier, R.: Deep regression on manifolds: a 3D rotation case study. In: 3DV (2021)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 590, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 590, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 138, + 590, + 480, + 609 + ], + "type": "text", + "content": "19. Brown, D.: The bundle adjustment-progress and prospect. In: Congr. of the Int. Soc. for Photogr. (1976)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "type": "text", + "content": "20. Brown, M., Lowe, D.G.: Unsupervised 3D object recognition and reconstruction in unordered datasets. In: 3DIM (2005)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "text", + "content": "21. Carlone, L., Tron, R., Daniilidis, K., Dellaert, F.: Initialization techniques for 3D SLAM: A survey on rotation estimation and its use in pose graph optimization. In: ICRA (2015)" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 311, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 311, + 91, + 447, + 101 + ], + "type": "text", + "content": "Scene Coordinate Reconstruction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 137, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 137, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 137, + 116, + 481, + 149 + ], + "type": "text", + "content": "22. Cavallari, T., Bertinetto, L., Mukhoti, J., Torr, P.H., Golodetz, S.: Let's take this online: Adapting scene coordinate regression network predictions for online RGB-D camera relocalisation. In: 3DV (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 137, + 150, + 481, + 183 + ], + "type": "text", + "content": "23. Cavallari, T., Golodetz, S., Lord, N.A., Valentin, J., Di Stefano, L., Torr, P.H.: On-the-fly adaptation of regression forests for online camera relocalisation. In: CVPR (2017)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 137, + 184, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 184, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 137, + 184, + 481, + 205 + ], + "type": "text", + "content": "24. Chen, S., Bhalgat, Y., Li, X., Bian, J., Li, K., Wang, Z., Prisacariu, V.A.: Neural refinement for absolute pose regression with feature synthesis. In: CVPR (2024)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 206, + 481, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 206, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 137, + 206, + 481, + 228 + ], + "type": "text", + "content": "25. Chen, S., Li, X., Wang, Z., Prisacariu, V.: DFNet: Enhance absolute pose regression with direct feature matching. In: ECCV (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 228, + 481, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 228, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 137, + 228, + 481, + 251 + ], + "type": "text", + "content": "26. Chen, S., Wang, Z., Prisacariu, V.: Direct-PoseNet: Absolute pose regression with photometric consistency. In: 3DV (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 251, + 481, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 251, + 481, + 283 + ], + "spans": [ + { + "bbox": [ + 137, + 251, + 481, + 283 + ], + "type": "text", + "content": "27. Cheng, Z., Esteves, C., Jampani, V., Kar, A., Maji, S., Makadia, A.: LU-NeRF: Scene and pose estimation by synchronizing local unposed NeRFs. In: ICCV (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 137, + 285, + 481, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 285, + 481, + 306 + ], + "spans": [ + { + "bbox": [ + 137, + 285, + 481, + 306 + ], + "type": "text", + "content": "28. Crandall, D., Owens, A., Snavely, N., Huttenlocher, D.: Discrete-continuous optimization for large-scale structure from motion. In: CVPR (2011)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 137, + 307, + 481, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 307, + 481, + 329 + ], + "spans": [ + { + "bbox": [ + 137, + 307, + 481, + 329 + ], + "type": "text", + "content": "29. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: ScanNet: Richly-annotated 3D reconstructions of indoor scenes. In: CVPR (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 137, + 330, + 481, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 330, + 481, + 351 + ], + "spans": [ + { + "bbox": [ + 137, + 330, + 481, + 351 + ], + "type": "text", + "content": "30. Davison, A.J.: Real-time simultaneous localisation and mapping with a single camera. In: ICCV (2003)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 137, + 352, + 481, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 352, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 137, + 352, + 481, + 373 + ], + "type": "text", + "content": "31. DeTone, D., Malisiewicz, T., Rabinovich, A.: SuperPoint: Self-supervised interest point detection and description. In: CVPRW (2018)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 137, + 374, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 374, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 137, + 374, + 481, + 396 + ], + "type": "text", + "content": "32. Ding, M., Wang, Z., Sun, J., Shi, J., Luo, P.: CamNet: Coarse-to-fine retrieval for camera re-localization. In: ICCV (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 137, + 397, + 481, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 397, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 137, + 397, + 481, + 430 + ], + "type": "text", + "content": "33. Dusmanu, M., Rocco, I., Pajdla, T., Pollefeys, M., Sivic, J., Torii, A., Sattler, T.: D2-net: A trainable CNN for joint description and detection of local features. In: CVPR (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 137, + 430, + 481, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 430, + 481, + 462 + ], + "spans": [ + { + "bbox": [ + 137, + 430, + 481, + 462 + ], + "type": "text", + "content": "34. Fischler, M.A., Bolles, R.C.: Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Comm. of the ACM (1981)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 137, + 464, + 481, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 464, + 481, + 485 + ], + "spans": [ + { + "bbox": [ + 137, + 464, + 481, + 485 + ], + "type": "text", + "content": "35. Gao, X.S., Hou, X.R., Tang, J., Cheng, H.F.: Complete solution classification for the perspective-three-point problem. IEEE TPAMI (2003)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 137, + 487, + 481, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 487, + 481, + 508 + ], + "spans": [ + { + "bbox": [ + 137, + 487, + 481, + 508 + ], + "type": "text", + "content": "36. Gherardi, R., Farenzena, M., Fusiello, A.: Improving the efficiency of hierarchical structure-and-motion. In: CVPR (2010)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 137, + 509, + 481, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 509, + 481, + 529 + ], + "spans": [ + { + "bbox": [ + 137, + 509, + 481, + 529 + ], + "type": "text", + "content": "37. Govindu, V.M.: Combining two-view constraints for motion estimation. In: CVPR (2001)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 137, + 531, + 481, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 531, + 481, + 552 + ], + "spans": [ + { + "bbox": [ + 137, + 531, + 481, + 552 + ], + "type": "text", + "content": "38. Govindu, V.M.: Lie-algebraic averaging for globally consistent motion estimation. In: CVPR (2004)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 137, + 554, + 451, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 554, + 451, + 564 + ], + "spans": [ + { + "bbox": [ + 137, + 554, + 451, + 564 + ], + "type": "text", + "content": "39. Hartley, R., Trumpf, J., Dai, Y., Li, H.: Rotation averaging. IJCV (2013)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 137, + 566, + 481, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 566, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 137, + 566, + 481, + 586 + ], + "type": "text", + "content": "40. Hartley, R., Zisserman, A.: Multiple view geometry in computer vision. Cambridge University Press (2003)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 137, + 587, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 587, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 137, + 587, + 481, + 609 + ], + "type": "text", + "content": "41. He, X., Sun, J., Wang, Y., Peng, S., Huang, Q., Bao, H., Zhou, X.: Detector-free structure from motion. In: CVPR (2024)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 137, + 609, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 609, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 137, + 609, + 481, + 632 + ], + "type": "text", + "content": "42. Heinly, J., Schonberger, J.L., Dunn, E., Frahm, J.M.: Reconstructing the World in six days. In: CVPR (2015)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 137, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 137, + 632, + 481, + 665 + ], + "type": "text", + "content": "43. Humenberger, M., Cabon, Y., Pion, N., Weinzaepfel, P., Lee, D., Guérin, N., Sattler, T., Csurka, G.: Investigating the role of image retrieval for visual localization: An exhaustive benchmark. IJCV (2022)" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 250, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 250, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 250, + 101 + ], + "type": "text", + "content": "E. Brachmann et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 138, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 481, + 160 + ], + "type": "text", + "content": "44. Izadi, S., Kim, D., Hilliges, O., Molyneaux, D., Newcombe, R., Kohli, P., Shotton, J., Hodges, S., Freeman, D., Davison, A.J., Fitzgibbon, A.: KinectFusion: real-time 3D reconstruction and interaction using a moving depth camera. In: UIST (2011)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 161, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 161, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 138, + 161, + 481, + 182 + ], + "type": "text", + "content": "45. Jeong, Y., Ahn, S., Choy, C., Anandkumar, A., Cho, M., Park, J.: Self-calibrating neural radiance fields. In: ICCV (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 182, + 481, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 182, + 481, + 204 + ], + "spans": [ + { + "bbox": [ + 138, + 182, + 481, + 204 + ], + "type": "text", + "content": "46. Jin, Y., Mishkin, D., Mishchuk, A., Matas, J., Fua, P., Yi, K.M., Trulls, E.: Image matching across wide baselines: From paper to practice. IJCV (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 204, + 481, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 204, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 138, + 204, + 481, + 224 + ], + "type": "text", + "content": "47. Kendall, A., Grimes, M., Cipolla, R.: PoseNet: A convolutional network for real-time 6-DoF camera relocalization. In: ICCV (2015)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 224, + 481, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 224, + 481, + 246 + ], + "spans": [ + { + "bbox": [ + 138, + 224, + 481, + 246 + ], + "type": "text", + "content": "48. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3D gaussian splatting for real-time radiance field rendering. ACM TOG (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 246, + 481, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 246, + 481, + 267 + ], + "spans": [ + { + "bbox": [ + 138, + 246, + 481, + 267 + ], + "type": "text", + "content": "49. Knapitsch, A., Park, J., Zhou, Q.Y., Koltun, V.: Tanks and Temples: Benchmarking large-scale scene reconstruction. ACM TOG (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 268, + 481, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 268, + 481, + 289 + ], + "spans": [ + { + "bbox": [ + 138, + 268, + 481, + 289 + ], + "type": "text", + "content": "50. Kraus, K.: Photogrammetry. No. v. 1 in Photogrammetry, Ferdinand Dummlers Verlag (1993)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 289, + 481, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 289, + 481, + 321 + ], + "spans": [ + { + "bbox": [ + 138, + 289, + 481, + 321 + ], + "type": "text", + "content": "51. Laskar, Z., Melekhov, I., Kalia, S., Kannala, J.: Camera relocalization by computing pairwise relative poses using convolutional neural network. In: ICCV Workshops (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 321, + 481, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 321, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 321, + 481, + 342 + ], + "type": "text", + "content": "52. Li, X., Wang, S., Zhao, Y., Verbeek, J., Kannala, J.: Hierarchical scene coordinate classification and regression for visual localization. In: CVPR (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 342, + 481, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 342, + 481, + 365 + ], + "spans": [ + { + "bbox": [ + 138, + 342, + 481, + 365 + ], + "type": "text", + "content": "53. Lin, A., Zhang, J.Y., Ramanan, D., Tulsiani, S.: Relpose++: Recovering 6d poses from sparse-view observations. In: 3DV (2024)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 365, + 481, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 365, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 138, + 365, + 481, + 385 + ], + "type": "text", + "content": "54. Lin, C.H., Ma, W.C., Torralba, A., Lucey, S.: BARF: Bundle-adjusting neural radiance fields. In: ICCV (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 385, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 385, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 138, + 385, + 481, + 418 + ], + "type": "text", + "content": "55. Lin, Y., Müller, T., Tremblay, J., Wen, B., Tyree, S., Evans, A., Vela, P.A., Birchfield, S.: Parallel inversion of neural radiance fields for robust pose estimation. In: ICRA (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 418, + 481, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 418, + 481, + 439 + ], + "spans": [ + { + "bbox": [ + 138, + 418, + 481, + 439 + ], + "type": "text", + "content": "56. Lindenberger, P., Sarlin, P.E., Pollefeys, M.: LightGlue: Local feature matching at light speed. In: ICCV (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 439, + 481, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 439, + 481, + 450 + ], + "spans": [ + { + "bbox": [ + 138, + 439, + 481, + 450 + ], + "type": "text", + "content": "57. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: ICLR (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 450, + 481, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 450, + 481, + 472 + ], + "spans": [ + { + "bbox": [ + 138, + 450, + 481, + 472 + ], + "type": "text", + "content": "58. Lowe, D.G.: Distinctive image features from scale-invariant keypoints. IJCV (2004)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 472, + 481, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 472, + 481, + 493 + ], + "spans": [ + { + "bbox": [ + 138, + 472, + 481, + 493 + ], + "type": "text", + "content": "59. Martinec, D., Pajdla, T.: Robust rotation and translation estimation in multiview reconstruction. In: CVPR (2007)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 493, + 481, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 493, + 481, + 514 + ], + "spans": [ + { + "bbox": [ + 138, + 493, + 481, + 514 + ], + "type": "text", + "content": "60. Meng, Q., Chen, A., Luo, H., Wu, M., Su, H., Xu, L., He, X., Yu, J.: GNeRF: GAN-based Neural Radiance Field without Posed Camera. In: ICCV (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 514, + 481, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 514, + 481, + 547 + ], + "spans": [ + { + "bbox": [ + 138, + 514, + 481, + 547 + ], + "type": "text", + "content": "61. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 547, + 481, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 547, + 481, + 579 + ], + "spans": [ + { + "bbox": [ + 138, + 547, + 481, + 579 + ], + "type": "text", + "content": "62. Moreau, A., Piasco, N., Bennehar, M., Tsishkou, D., Stanciulescu, B., de La Fortelle, A.: CROSSFIRE: Camera relocalization on self-supervised features from an implicit representation. ICCV (2023)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 579, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 579, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 138, + 579, + 481, + 601 + ], + "type": "text", + "content": "63. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM TOG (2022)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 138, + 601, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 481, + 633 + ], + "type": "text", + "content": "64. Newcombe, R., Izadi, S., Hilliges, O., Molyneaux, D., Kim, D., Davison, A.J., Kohi, P., Shotton, J., Hodges, S., Fitzgibbon, A.: KinectFusion: Real-time dense surface mapping and tracking. In: ISMAR (2011)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 138, + 633, + 454, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 454, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 454, + 643 + ], + "type": "text", + "content": "65. Nistér, D., Naroditsky, O., Bergen, J.: Visual odometry. In: CVPR (2004)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "type": "text", + "content": "66. Pollefeys, M., Koch, R., Vergauwen, M., Van Gool, L.: Automated reconstruction of 3D scenes from sequences of images. J. of Photogr. and Rem. Sens. (2000)" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 311, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 311, + 91, + 447, + 101 + ], + "type": "text", + "content": "Scene Coordinate Reconstruction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 480, + 149 + ], + "type": "text", + "content": "67. Rau, A., Garcia-Hernando, G., Stoyanov, D., Brostow, G.J., Turmukhambetov, D.: Predicting visual overlap of images through interpretable non-metric box embeddings. In: ECCV (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 150, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 150, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 138, + 150, + 480, + 172 + ], + "type": "text", + "content": "68. Reality, C.: Reality Capture. https://www.capturingreality.com/ realitycapture (2016), [accessed 15-Nov-2023]" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 172, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 172, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 138, + 172, + 480, + 205 + ], + "type": "text", + "content": "69. Reizenstein, J., Shapovalov, R., Henzler, P., Sbordone, L., Labatut, P., Novotny, D.: Common objects in 3D: Large-scale learning and evaluation of real-life 3D category reconstruction. In: ICCV (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 205, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 205, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 138, + 205, + 480, + 226 + ], + "type": "text", + "content": "70. Sarlin, P.E., Cadena, C., Siegwart, R., Dymczyk, M.: From coarse to fine: Robust hierarchical localization at large scale. In: CVPR (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 227, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 227, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 138, + 227, + 480, + 248 + ], + "type": "text", + "content": "71. Sarlin, P.E., DeTone, D., Malisiewicz, T., Rabinovich, A.: SuperGlue: Learning feature matching with graph neural networks. In: CVPR (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 249, + 480, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 249, + 480, + 282 + ], + "spans": [ + { + "bbox": [ + 138, + 249, + 480, + 282 + ], + "type": "text", + "content": "72. Sarlin, P.E., Unagar, A., Larsson, M., Germain, H., Toft, C., Larsson, V., Pollefeys, M., Lepetit, V., Hammarstrand, L., Kahl, F., Sattler, T.: Back to the feature: Learning robust camera localization from pixels to pose. In: CVPR (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 282, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 282, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 138, + 282, + 480, + 303 + ], + "type": "text", + "content": "73. Sattler, T., Leibe, B., Kobbelt, L.: Fast image-based localization using direct 2D-to-3D matching. In: ICCV (2011)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 303, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 303, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 138, + 303, + 480, + 325 + ], + "type": "text", + "content": "74. Sattler, T., Leibe, B., Kobbelt, L.: Improving image-based localization by active correspondence search. In: ECCV (2012)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 325, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 325, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 138, + 325, + 480, + 346 + ], + "type": "text", + "content": "75. Sattler, T., Leibe, B., Kobbelt, L.: Efficient & effective prioritized matching for large-scale image-based localization. IEEE TPAMI (2016)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 347, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 347, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 138, + 347, + 480, + 380 + ], + "type": "text", + "content": "76. Sattler, T., Torii, A., Sivic, J., Pollefeys, M., Taira, H., Okutomi, M., Pajdla, T.: Are large-scale 3D models really necessary for accurate visual localization? In: CVPR (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 380, + 480, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 380, + 480, + 403 + ], + "spans": [ + { + "bbox": [ + 138, + 380, + 480, + 403 + ], + "type": "text", + "content": "77. Sattler, T., Zhou, Q., Pollefeys, M., Leal-Taixe, L.: Understanding the limitations of CNN-based absolute camera pose regression. In: CVPR (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 403, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 403, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 138, + 403, + 480, + 424 + ], + "type": "text", + "content": "78. Schaffalitzky, F., Zisserman, A.: Multi-view matching for unordered image sets, or \"How do i organize my holiday snaps?\". In: ECCV (2002)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 424, + 480, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 424, + 480, + 446 + ], + "spans": [ + { + "bbox": [ + 138, + 424, + 480, + 446 + ], + "type": "text", + "content": "79. Schonberger, J.L.: Colmap Github Issues. https://github.com/colmap/colmap/issues/116#issuecomment-298926277 (2017), [accessed Nov/15/2023]" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 446, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 446, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 446, + 480, + 468 + ], + "type": "text", + "content": "80. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: CVPR (2016)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 468, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 468, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 138, + 468, + 480, + 500 + ], + "type": "text", + "content": "81. Shotton, J., Glocker, B., Zach, C., Izadi, S., Criminisi, A., Fitzgibbon, A.: Scene coordinate regression forests for camera relocalization in RGB-D images. In: CVPR (2013)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 500, + 480, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 500, + 480, + 523 + ], + "spans": [ + { + "bbox": [ + 138, + 500, + 480, + 523 + ], + "type": "text", + "content": "82. Sinha, S., Zhang, J.Y., Tagliasacchi, A., Gilitschenski, I., Lindell, D.B.: Sparse-Pose: Sparse-view camera pose regression and refinement. In: CVPR (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 523, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 523, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 523, + 480, + 555 + ], + "type": "text", + "content": "83. Smith, L.N., Topin, N.: Super-convergence: Very fast training of neural networks using large learning rates. In: Arti. Intell. and Mach. Learn. for Multi-Domain Operations Appl. (2019)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 555, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 480, + 578 + ], + "type": "text", + "content": "84. Snavely, N., Seitz, S.M., Szeliski, R.: Photo tourism: exploring photo collections in 3D. ACM TOG (2006)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 578, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 480, + 599 + ], + "type": "text", + "content": "85. Snavely, N., Seitz, S.M., Szeliski, R.: Modeling the world from internet photo collections. IJCV (2008)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 600, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 600, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 138, + 600, + 480, + 621 + ], + "type": "text", + "content": "86. Snavely, N., Seitz, S.M., Szeliski, R.: Skeletal graphs for efficient structure from motion. In: CVPR (2008)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 138, + 621, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 480, + 643 + ], + "type": "text", + "content": "87. Sun, J., Shen, Z., Wang, Y., Bao, H., Zhou, X.: LoFTR: Detector-free local feature matching with transformers. In: CVPR (2021)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 138, + 643, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 666 + ], + "type": "text", + "content": "88. Szeliski, R., Kang, S.B.: Recovering 3D shape and motion from image streams using nonlinear least squares. J. of Vis. Com. and Image Repr. (1994)" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 250, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 250, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 250, + 101 + ], + "type": "text", + "content": "E. Brachmann et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 533 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 480, + 149 + ], + "type": "text", + "content": "89. Tancik, M., Weber, E., Ng, E., Li, R., Yi, B., Kerr, J., Wang, T., Kristoffersen, A., Austin, J., Salahi, K., Ahuja, A., McAllister, D., Kanazawa, A.: Nerfstudio: A modular framework for neural radiance field development. In: ACM TOG (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 150, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 150, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 138, + 150, + 480, + 171 + ], + "type": "text", + "content": "90. Teed, Z., Deng, J.: DROID-SLAM: Deep visual SLAM for monocular, stereo, and RGB-D cameras. In: NeurIPS (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 172, + 480, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 172, + 480, + 193 + ], + "spans": [ + { + "bbox": [ + 138, + 172, + 480, + 193 + ], + "type": "text", + "content": "91. Toldo, R., Gherardi, R., Farenzena, M., Fusiello, A.: Hierarchical structure-and-motion recovery from uncalibrated images. CVIU (2015)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 194, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 194, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 138, + 194, + 480, + 215 + ], + "type": "text", + "content": "92. Triggs, B., McLauchlan, P.F., Hartley, R.I., Fitzgibbon, A.W.: Bundle adjustment—a modern synthesis. In: Int. Worksh. on Vis. Alg. (2000)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 216, + 480, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 216, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 138, + 216, + 480, + 247 + ], + "type": "text", + "content": "93. Türkoglu, M.Ö., Brachmann, E., Schindler, K., Brostow, G., Monszpart, A.: Visual camera re-localization using graph neural networks and relative pose supervision. In: 3DV (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 249, + 467, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 249, + 467, + 259 + ], + "spans": [ + { + "bbox": [ + 138, + 249, + 467, + 259 + ], + "type": "text", + "content": "94. Ulyanov, D., Vedaldi, A., Lempitsky, V.: Deep image prior. In: CVPR (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 260, + 480, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 260, + 480, + 292 + ], + "spans": [ + { + "bbox": [ + 138, + 260, + 480, + 292 + ], + "type": "text", + "content": "95. Ummenhofer, B., Zhou, H., Uhrig, J., Mayer, N., Ilg, E., Dosovitskiy, A., Brox, T.: DeMoN: Depth and motion network for learning monocular stereo. In: CVPR (2017)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 293, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 293, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 138, + 293, + 480, + 325 + ], + "type": "text", + "content": "96. Waechter, M., Beljan, M., Fuhrmann, S., Moehrle, N., Kopf, J., Goesele, M.: Virtual rephotography: Novel view prediction error for 3d reconstruction. ACM TOG (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 326, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 326, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 138, + 326, + 480, + 346 + ], + "type": "text", + "content": "97. Wang, S., Leroy, V., Cabon, Y., Chidlovskii, B., Revaud, J.: DUSt3R: Geometric 3D vision made easy. In: CVPR (2024)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 347, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 347, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 138, + 347, + 480, + 369 + ], + "type": "text", + "content": "98. Wang, Z., Wu, S., Xie, W., Chen, M., Prisacariu, V.A.: NeRF--: Neural radiance fields without known camera parameters. arXiv (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 369, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 369, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 138, + 369, + 480, + 391 + ], + "type": "text", + "content": "99. Wei, X., Zhang, Y., Li, Z., Fu, Y., Xue, X.: DeepSFM: Structure from motion via deep bundle adjustment. In: ECCV (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 392, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 392, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 138, + 392, + 480, + 402 + ], + "type": "text", + "content": "00. Wu, C.: Towards linear-time incremental structure from motion. In: 3DV (2013)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 403, + 480, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 403, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 138, + 403, + 480, + 423 + ], + "type": "text", + "content": "01. Xia, Y., Tang, H., Timofte, R., Van Gool, L.: SiNeRF: Sinusoidal neural radiance fields for joint pose estimation and scene reconstruction. In: BMVC (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 424, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 424, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 138, + 424, + 480, + 445 + ], + "type": "text", + "content": "02. Yen-Chen, L., Florence, P., Barron, J.T., Rodriguez, A., Isola, P., Lin, T.Y.: iNeRF: Inverting neural radiance fields for pose estimation. In: IROS (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 447, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 447, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 447, + 480, + 468 + ], + "type": "text", + "content": "03. Zhang, J.Y., Lin, A., Kumar, M., Yang, T.H., Ramanan, D., Tulsiani, S.: Cameras as rays: Pose estimation via ray diffusion. In: ICLR (2024)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 468, + 480, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 468, + 480, + 489 + ], + "spans": [ + { + "bbox": [ + 138, + 468, + 480, + 489 + ], + "type": "text", + "content": "04. Zhang, W., Kosecka, J.: Image based localization in urban environments. In: 3DPVT (2006)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 490, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 490, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 138, + 490, + 480, + 511 + ], + "type": "text", + "content": "05. Zhou, Q., Sattler, T., Pollefeys, M., Leal-Taixe, L.: To learn or not to learn: Visual localization from essential matrices. In: ICRA (2020)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 512, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 512, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 512, + 480, + 533 + ], + "type": "text", + "content": "06. Zhou, Y., Barnes, C., Jingwan, L., Jimei, Y., Hao, L.: On the continuity of rotation representations in neural networks. In: CVPR (2019)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 311, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 311, + 91, + 447, + 101 + ], + "type": "text", + "content": "Scene Coordinate Reconstruction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Scene-Conditional 3D Object Stylization and Composition/ed573969-4bd1-4a9b-b748-12ae2bb2b2d3_content_list.json b/2024/Scene-Conditional 3D Object Stylization and Composition/ed573969-4bd1-4a9b-b748-12ae2bb2b2d3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2d7d01478d7cb690b771f18a763213363a8125cf --- /dev/null +++ b/2024/Scene-Conditional 3D Object Stylization and Composition/ed573969-4bd1-4a9b-b748-12ae2bb2b2d3_content_list.json @@ -0,0 +1,1544 @@ +[ + { + "type": "text", + "text": "Scene-Conditional 3D Object Stylization and Composition", + "text_level": 1, + "bbox": [ + 274, + 141, + 727, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jinghao Zhou $^{1}$ , Tomas Jakab $^{1}$ , Philip Torr $^{1}$ , and Christian Rupprecht $^{2}$", + "bbox": [ + 238, + 213, + 764, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "University of Oxford {jinghao,tomj,chrisr}@robots.ox.ac.uk", + "bbox": [ + 357, + 239, + 643, + 268 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a322d6fb3ae17492b50af53c6a069cd9aa7f3ee71da505f94b738df7691db465.jpg", + "image_caption": [ + "Fig. 1: We present a framework that adapts a 3D object's appearance to a location in a 2D scene. It creates an image where the 3D object is seamlessly blended (left & bottom right), with its appearance influenced by the scene's environmental conditions and lighting effects. Moreover, the stylized object with adapted textures (top right), rendered here without the estimated lighting condition for illustrative purposes, can be further utilized as 3D assets for downstream tasks such as video games." + ], + "image_footnote": [], + "bbox": [ + 217, + 299, + 785, + 536 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Recently, 3D generative models have made impressive progress, enabling the generation of almost arbitrary 3D assets from text or image inputs. However, these approaches generate objects in isolation without any consideration for the scene where they will eventually be placed. In this paper, we propose a framework that allows for the stylization of an existing 3D asset to fit into a given 2D scene, and additionally produce a photorealistic composition as if the asset was placed within the environment. This not only opens up a new level of control for object stylization, for example, the same assets can be stylized to reflect changes in the environment, such as summer to winter or fantasy versus futuristic settings—but also makes the object-scene composition more controllable. We achieve this by combining modeling and optimizing the object's texture and environmental lighting through differentiable ray", + "bbox": [ + 259, + 659, + 750, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tracing with image priors from pre-trained text-to-image diffusion models. We demonstrate that our method applies to a wide variety of indoor and outdoor scenes and arbitrary objects. See also our project page.", + "bbox": [ + 259, + 147, + 743, + 189 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 209, + 375, + 224 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "After the success of generative models [20,43,46] for images in computer vision, the community is now interested in lifting these models to 3D [40,54] through advances in architectures, data, and training. These lifted 3D models demonstrate plausible 3D representations of objects conditioned on images or texts.", + "bbox": [ + 212, + 234, + 784, + 295 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While generating whole objects—e.g. an astronaut riding a horse—is a challenging problem that generative models are becoming able to solve, in practice, these assets are difficult to adapt for use in downstream applications or existing virtual environments such as video games as they are generated without context.", + "bbox": [ + 212, + 296, + 785, + 356 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This paper aims to provide a mechanism to close the appearance gap between 3D objects and environments. We formulate this problem through the lens of a creative tool. The intended goal of this paper is to align the appearance of a 3D object to a customized 2D scene as we place it into a specific location, a challenging task that requires fine-grained visual control for convincing results.", + "bbox": [ + 212, + 357, + 785, + 431 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We are interested in producing a photorealistic scene-object composition where the 3D object is seamlessly blended. For instance, an object placed in a rainy night scene should appear wet and dimly lit, and if the scene is muddy and sunny, the object should also look muddy and cast a strong shadow. To achieve a high level of realism on a fine-grained level, the mud on the object must match the scene's mud in appearance and the direction of the shadow must match where the sun is located. Instead of simply resorting to 2D blending operations [18, 50], we explicitly model an object's appearance and lighting in 3D space, which enables more accurate and regularized control of the object's appearance in the composed image.", + "bbox": [ + 212, + 431, + 785, + 583 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We also aim to preserve the underlying object geometry and appearance factors unrelated to the scene environment. This task is highly practical; for example, for media design, when presenting a 3D asset of a product within a realistic scene, we want the product to blend naturally while remaining recognizable. Manually achieving this would require a skilled artist to modify the object's texture to match the scene, a challenging and time-consuming process. Our paper introduces a method that automates this task. It accepts a 3D object with texture, a 2D scene, and the object's position within the scene and outputs an adapted texture consistent with the scene.", + "bbox": [ + 212, + 584, + 785, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To accomplish this, we leverage recent advances in generative diffusion models [46, 47] for images, which are trained on a large-scale dataset of internet images. We conduct an optimization procedure where we compose the 2D renderings of the 3D object with the 2D scene at a specified location and use a pre-trained diffusion model to \"critique\" the realism of the composed image, providing a gradient signal for optimization.", + "bbox": [ + 212, + 719, + 785, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The challenge is that diffusion models, such as Stable Diffusion [46], are trained on 2D images and lack any notion of 3D geometry. Naively optimizing", + "bbox": [ + 214, + 809, + 785, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "J. Zhou et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "or denoising in 2D space [18,50] would result in losing the object's 3D geometry. To prevent this, we optimize the object's content in 3D by randomly rotating the object during optimization and rendering it from various viewpoints, which helps to align the adapted texture with the object's geometry and avoid overfitting to a single viewpoint.", + "bbox": [ + 212, + 146, + 782, + 222 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Notably, randomly rotating the 3D object stops lighting effects from being baked into the texture, which is desirable. Yet, it also prevents the object from naturally blending with the scene in a single viewpoint. To counteract this, we predict the scene's lighting and the shadows cast by the object as separate components. For this purpose, we propose a novel technique inspired by a common physical approach from the computer graphics industry, where reflective and diffuse spheres with known material parameters are placed and photographed in a real-world scene to capture the environmental lighting conditions. We adopt this concept and place a virtual white diffuse sphere inside the scene during optimization while conditioning the generative model to expect a white sphere within the scene, thereby capturing the model's interpretation of the scene's lighting in an easily extracted form.", + "bbox": [ + 212, + 222, + 784, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Optimizing the object's appearance in 3D offers further significant advantages: The user can change the object's pose without re-running the optimization, as the texture is meaningful from all viewpoints, and view-dependent effects are modeled separately by the lighting component. Additionally, the 3D object with the adapted texture can be used as a standalone asset in other downstream tasks (e.g. in a video game).", + "bbox": [ + 212, + 402, + 784, + 494 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, we present a framework that (1) enables the stylization of a 3D object, which can also be used as a standalone asset, given a 2D scene and its location within it; (2) achieves photorealistic scene-object compositing with the help of a novel mechanism to estimate scene lighting from a single image.", + "bbox": [ + 212, + 494, + 784, + 555 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We further verify the effectiveness of our method in extensive experiments, demonstrating that our approach can realistically adapt objects to a diverse set of environments, laying the groundwork for practical applications.", + "bbox": [ + 212, + 555, + 784, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 214, + 619, + 387, + 635 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The purpose of this paper is to adapt the appearance of a 3D object to align with a customized 2D scene. This solves a more complicated problem than prior works, which often only tackle individual components, including but not limited to 2D scene-object compositing, light estimation, and texturing, but never as a whole. We compare with these methods next and detail their differences in Tab. 1.", + "bbox": [ + 212, + 647, + 782, + 723 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Scene-Object Compositing. Scene-object compositing is a challenging task that requires a delicate balance between visual transfer and control. It has been extensively researched [5,8,9,25]. Recent methods condition the denoising process of diffusion models [20,49] on object images. OST [50] aims to seamlessly integrate a 2D object image into a 2D scene by modifying the keys and values in the attention blocks. CDC [18] considers composing the 2D renderings of a 3D object with a 2D scene by directly modifying denoised latent.", + "bbox": [ + 212, + 734, + 784, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Scene-Conditional 3D Object Stylization and Composition", + "bbox": [ + 341, + 114, + 730, + 130 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/53d0cd996d163bd3a4717d79a5532e1a1763d7314b12d94e7e9cce75e212ebb3.jpg", + "table_caption": [ + "Table 1: Comparison with related work. Rep. denotes representation being 2D (□) or 3D (）。Env. Inf. denotes Environmental Influence and I.D. Pre. denotes Identity Preservation. Parameters for the lighting and the appearance model are either scene-agnostic constant (），scene-conditioned predictable (），or learnable (）。" + ], + "table_footnote": [], + "table_body": "
MethodSceneObjectCompose
Rep.LightingShadowRep.AppearanceEnv. Inf.I.D. Pre.
OST [50]□-×□-××Blend
CDC [18]□Ambient×Diffuse×××Blend
IGAN [28]□LDR✓Diffuse×××Copy & Paste
EVL [27]□HDR✓Diffuse×××Copy & Paste
PrDr [55]-Ambient×Diffuse✓×-
Fan3D [7]-HDR×PBR✓×-
Inst3D [26]-Point×Diffuse×✓-
Ours□HDR✓PBR✓✓Copy & Paste
", + "bbox": [ + 223, + 212, + 779, + 383 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, these methods rely entirely on the 2D denoising process and do not explicitly model the object's geometry, which often results in a significant loss of structural details and visual identity. This severely hampers these methods' ability to adapt to the environmental and lighting conditions of the scene.", + "bbox": [ + 212, + 392, + 784, + 454 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Light Estimation from Single-View Image. Estimating light conditions from a single-view 2D image is of practical use in the real world and has been extensively studied for both indoor [15,16] and outdoor scenes [22,23,58]. The distinction between these two settings has been significantly narrowed recently [27, 28], thanks to the outpainting capabilities of diffusion models, which have contributed to a unified framework. With additional training on large high dynamic range (HDR) datasets such as SUN360 [57], these methods can approximate the location and intensity of the light source, enabling realistic shadows.", + "bbox": [ + 212, + 464, + 784, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, these methods estimate only the lighting of the scene. They are typically indifferent to how objects should be stylized to match the scene, neglecting the potential environmental influences on the textures of objects.", + "bbox": [ + 212, + 585, + 785, + 631 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Mesh Texturing. Classic automated approaches [56] for mesh texturing are often limited and apply only simple texture patterns. In contrast, methods [14, 37, 48] based on Generative Adversarial Networks [17] (GANs) are more capable of synthesizing detailed textures. Recently, text-conditioned texture generation [34, 35] has been made possible through the optimization of CLIP-based objectives [42]. The latest works now employ powerful text-to-image diffusion models [20, 49], either through optimization [7, 29, 33, 51, 55] using Score Distillation Sampling [40] (SDS), or by wrapping the generated 2D textures onto the mesh [4, 6, 45]. Importantly, Variational Score Distillation [55] (VSD) as an improved optimization objective and the adoption of PBR model [7] have greatly boosted the quality of generated textures. Additionally, an increasing level of requirement for more visual control from users has been partially met [19, 26] through image-conditional diffusion models [3] to preserve objects' identities.", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "J. Zhou et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ae638e978c607d549deeff73f662f294c56565eeeaba25ed3e50646025032972.jpg", + "image_caption": [ + "Fig. 2: Framework. We learn an environment map and a texture map separately from the 2D supervision. We initialize (init.) the environment map with an LDR map estimated from the 2D scene and learn light multiplying scales bright areas, yielding an HDR map. We employ the PBR material model for texture maps, encoded via MLP with positional encoding. The object is rendered through a differentiable ray tracer and further composed with the scene background, receiving gradients from Stable Diffusion (SD) in the latent space." + ], + "image_footnote": [], + "bbox": [ + 243, + 143, + 753, + 280 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, these works focus on texturing the object in isolation from the scene and do not consider the homogeneity of the composition when these assets are placed in certain environments, which limits their practical usage.", + "bbox": [ + 212, + 397, + 787, + 444 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 467, + 330, + 483 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overall structure of our framework can be found in Fig. 2. Given a textured 3D object, a 2D image of a scene, and a 2D location where to place the object, we first adapt the texture of the 3D object (Secs. 3.2 and 3.3) to align with the scene. To enable a photorealistic composition of the object and the scene, we further estimate the lighting (Secs. 3.4 to 3.6) and apply it during the final rendering.", + "bbox": [ + 212, + 500, + 787, + 590 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Optimization", + "text_level": 1, + "bbox": [ + 214, + 614, + 372, + 630 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To enable end-to-end optimization, we adopt a differentiable renderer to render the object and score it using the priors from diffusion models.", + "bbox": [ + 212, + 641, + 785, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Differentiable Renderer. We employ the differential renderer Mitsuba3 [24] (denoted as $g$ ), due to its capabilities as a physically-based ray-tracing renderer equipped with versatile rendering options, such as diverse BSDFs, the path tracer integrator, and using an environment map as a light source. This enables us to produce photo-realistic renders complete with shadows and reflections, forming the foundation for robust texture adaptation and light estimation.", + "bbox": [ + 212, + 686, + 787, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Diffusion Model as Guidance. To score the rendered image, we make use of a generative diffusion model [46, 47] $\\phi$ . Diffusion models have learned powerful priors of the image formation process in the real world from large-scale training", + "bbox": [ + 212, + 794, + 787, + 842 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Scene-Conditional 3D Object Stylization and Composition", + "bbox": [ + 341, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "on billion-scale web data. Similar to prior work [40,54], we supervise the rendered image $\\pmb{x}_0 = g(\\pmb{\\theta})$ with a score-matching objective to obtain an image gradient", + "bbox": [ + 215, + 146, + 784, + 176 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\boldsymbol {\\theta}} \\mathcal {L} (\\boldsymbol {x} _ {0}) \\triangleq \\mathbb {E} _ {t, \\epsilon} \\left[ w _ {t} \\left(\\epsilon_ {\\phi} (\\boldsymbol {x} _ {t}) - \\hat {\\epsilon} (\\boldsymbol {x} _ {t})\\right) \\frac {\\partial \\boldsymbol {x} _ {0}}{\\partial \\boldsymbol {\\theta}} \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 351, + 189, + 785, + 222 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\pmb{x}_t = \\sqrt{\\alpha_t}\\pmb{x}_0 + \\sqrt{1 - \\alpha_t}\\epsilon$ is the noisy rendered image, $\\epsilon_{\\phi}(\\pmb{x}_t)$ is an approximated score for a noisy real image, $t\\in \\mathbb{N}$ indexes the discrete diffusion steps, and $\\hat{\\epsilon}$ is a scoring function for $\\pmb{x}_t$ . Specifically, $\\hat{\\epsilon}$ is set as constant $\\epsilon$ in Score Distillation Sampling [40] (SDS), and $\\epsilon_{\\psi}$ is an online LoRA-tuned network initialized from $\\epsilon_{\\phi}$ in Variational Score Sampling [55] (VSD). Notably, we employ their interpolated variant given by $\\hat{\\epsilon} = \\lambda \\epsilon_{\\psi} + (1 - \\lambda)\\epsilon$ with a $\\lambda$ annealed from 1 to $\\lambda_e$ , helping to counteract the over-texturization often observed in VSD, especially for objects with simple surfaces. With residual of noises $\\epsilon_{\\phi} - \\hat{\\epsilon} = (\\epsilon_{\\phi} - \\epsilon) - \\lambda (\\epsilon_{\\psi} - \\epsilon)$ , we highlight its relation with the image gradient obtained in traditional inverse rendering setup $\\nabla_{\\theta}\\mathcal{L}(\\pmb{x}_0) = \\pmb{x}_0 - \\pmb{x}_{\\mathrm{gt}}$ by reformulating it as", + "bbox": [ + 215, + 233, + 782, + 385 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon_ {\\phi} - \\hat {\\epsilon} = \\sqrt {\\frac {\\alpha_ {t}}{1 - \\alpha_ {t}}} [ (\\boldsymbol {x} _ {0} - \\hat {\\boldsymbol {x}} _ {\\phi}) - \\lambda (\\boldsymbol {x} _ {0} - \\hat {\\boldsymbol {x}} _ {\\psi}) ], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 398, + 784, + 430 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where the target used to supervise $\\pmb{x}_0$ is predicted clean image $\\hat{\\pmb{x}}_{\\phi}$ from $\\epsilon_{\\phi}$ as the positive and $\\hat{\\pmb{x}}_{\\psi}$ from $\\epsilon_{\\psi}$ as the negative. The technique enables nominal classifier-free guidance [21] and thus bypasses mode-seeking behaviors and saturated texture. Inspired by [45], we incorporate the inner product of the camera pose and object's surface normal $\\langle c, \\pmb{n} \\rangle$ into $w_t$ to penalize faces at small grazing angles. In practice, we use Stable Diffusion [46] (SD) and absorb the encoder into $g$ , optimizing in latent space.", + "bbox": [ + 215, + 441, + 784, + 547 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Appearance Model", + "text_level": 1, + "bbox": [ + 215, + 570, + 418, + 585 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Textured Mesh. We represent the 3D object as a textured mesh, which is the most common format for 3D assets in the industry and is compatible with many established workflows. Textured meshes are available in abundance in online datasets and marketplaces [1,10], and recently, one can also generate reasonable quality 3D meshes with text descriptions [7,29,55] or limited viewpoints [30,41].", + "bbox": [ + 215, + 597, + 782, + 674 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Neural Texture. We represent the appearance of objects using a texture map $\\mathcal{T} \\in \\mathbb{R}^{H_t \\times W_t \\times 5}$ , facilitating compatibility of textured results with widely-used rendering engines, such as Blender. We parameterize the texture using a coordinate neural network that maps 3D coordinates to the texture attributes. We first map the 3D coordinates of vertices using a rasterizer [36] based on their UV coordinates into the space of the texture map, resulting in $\\mathcal{C} \\in \\mathbb{R}^{H_t \\times W_t \\times 3}$ . We then use a 2-layer parameterized MLP without bias and Hash positional encoding $\\varPhi: \\mathcal{C} \\to \\mathcal{T}$ to map the coordinates to their texture attributes. For the final layer, we re-map the output activated after sigmoid according to pre-defined minimal and maximal values. The five channels in $\\mathcal{T}$ correspond to the Physically-Based", + "bbox": [ + 215, + 689, + 785, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "J. Zhou et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9d8e045379d92a98dbde3d91fb581c9b5641cbcd3210a98d1af91a471e966bf6.jpg", + "image_caption": [ + "Fig. 3: Pipeline for texture adaptation. We initialize the neural texture from the reference object and inject the feature of reference renderings to the U-Net of SD. We use both local-view and global-view guidance." + ], + "image_footnote": [], + "bbox": [ + 243, + 146, + 751, + 305 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Rendering (PBR) shading model [32], which includes a diffuse term $k_{\\mathrm{d}} \\in \\mathbb{R}^{3}$ and an isotropic specular GGX lobe [53] described by $k_{\\mathrm{rm}} = (k_r, k_m) \\in \\mathbb{R}^2$ , with $k_r$ representing roughness and $k_m$ indicating metalness. The PBR model, with $k_r$ affecting the scattering of light and $k_m$ influencing the reflective properties, creates realistic light interactions between surfaces.", + "bbox": [ + 214, + 371, + 784, + 446 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Scene-Conditional Guidance", + "text_level": 1, + "bbox": [ + 215, + 469, + 493, + 483 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Achieving a realistic texture adaptation for 3D objects placed into 2D scenes requires three components: Environmental Influence involves adjusting the texture to reflect the environmental impact realistically; Identity Preservation aims to preserve the objects' unique visual aspects, and Blending focuses on guiding the texture adaptation to match the visual characteristic of the scene, ensuring seamless integration with the surroundings and avoiding stark contrasts.", + "bbox": [ + 214, + 493, + 784, + 583 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Environmental Influence. To show the realistic effects of the scene environment on the object's appearance, we guide the diffusion model with text prompts. We find that a naive text prompt, combining a simple object description with a scene description, e.g., a leather sofa in a swamp, is insufficient to generate realistic textures that reflect the environmental impact on the object, as shown in the appendix. For instance, we would expect a sofa placed in a swamp to be dirty and possibly show signs of moss or other vegetation growth. Instead of manually crafting a detailed text prompt for each object and scene combination, we automatize the process by employing a large language model (LLM), specifically GPT-4 [38], to generate a text prompt for the image diffusion model. Given the simple description of the object (e.g., a leather sofa) and the scene (e.g., a swamp), we instruct the LLM first to analyze the potential impact of the scene's environment on the object's appearance and then to distil this information into a text prompt that describes the resulting appearance of the object (e.g., The leather sofa, partially submerged in the swamp, looks discolored and soggy, its once-polished surface marred by mud, moss, and the murky water of the swamp", + "bbox": [ + 214, + 598, + 785, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Scene-Conditional 3D Object Stylization and Composition", + "bbox": [ + 341, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/384078b29483dc05b10914722d91a27b165a49f4eebdafa8d70157f7f1e74900.jpg", + "image_caption": [ + "Fig. 4: Pipeline for estimating the LDR map. We utilize tailored pipelines for indoor and outdoor scenes. Areas masked red and blue correspond to the far light $\\mathcal{O}$ and near light region." + ], + "image_footnote": [], + "bbox": [ + 218, + 145, + 787, + 220 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "environment.). We then use this automatically generated text prompt to guide the diffusion model in adapting the texture. The specific prompt design for this automated process is detailed in the appendix.", + "bbox": [ + 212, + 285, + 787, + 330 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Identity Preservation. To prevent the transferred texture from drifting away from its original content, we regularize the denoising diffusion model with reference images, a technique that has been explored for controlled image generation [39, 52]. Suppose we have a rendered reference image $\\pmb{x}_0^r = g(\\pmb{\\theta}^r)$ and its noised version $\\pmb{x}_t^r$ as input forwarded to $\\epsilon_{\\phi}$ and $\\epsilon_{\\psi}$ , for the Multi-Head Self-Attention (MHSA) layer in each Transformer block of $\\epsilon_{\\phi}$ 's and $\\epsilon_{\\psi}$ 's U-Nets, we concatenate the key $K^r$ and value $V^r$ of $\\pmb{x}_t^r$ to the original $K$ and $V$ from $\\pmb{x}_t$ yielding an image-conditional attention output", + "bbox": [ + 212, + 356, + 787, + 478 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nA (\\boldsymbol {x} _ {t}, \\cdot , \\boldsymbol {x} _ {t} ^ {r}) = \\mathrm {M H S A} (Q, [ K; K ^ {r} ], [ V; V ^ {r} ]), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 349, + 493, + 785, + 512 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\cdot$ can either be text or $\\emptyset$ for text conditioning. The text-unconditional attention $A(\\pmb{x}_t, \\emptyset, \\pmb{x}_t^r)$ used for classifier-free guidance [21] is further modified as $s_c \\cdot A(\\pmb{x}_t, \\emptyset, \\emptyset) + (1 - s_c) \\cdot A(\\pmb{x}_t, \\emptyset, \\pmb{x}_t^r)$ weighted by a control guidance scale $s_c$ . With the percentage $p$ of blocks to be injected, the intensity of preservation can be controlled: $s_c = p = 1$ provide the most and $s_c = p = 0$ the least preservation.", + "bbox": [ + 212, + 526, + 787, + 603 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Blending. The two components previously mentioned directly utilize object renderings, referred to as the local view, as inputs for vanilla SD. To further enhance the adaptation of the object's appearance to its surroundings, we input the scene image, with the object rendered onto it, referred to as the global view, into an additional inpainting SD [46]. We have found that the inpainting model provides superior conditioning on the surrounding environment depicted in the image, resulting in improved blending. However, the inpainting model can sometimes attempt to undesirably modify the inserted object, removing parts of it by inpainting them with the background. To counter this, we compose the local view with randomly sample solid background colors, rather than the scene. This prevents the inpainting model from removing parts of the object as the object must look realistic when rendered in front of a random background. In practice, we utilize a set of global views created by cropping the original global view at various scales, ensuring that the crops encompass the entire local view.", + "bbox": [ + 212, + 628, + 789, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "J. Zhou et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/865fcc92eb9818599c57327de653788632139a206693e2bfba30d0f6a4851bce.jpg", + "image_caption": [ + "(a) Diverse Objects and Scenes" + ], + "image_footnote": [], + "bbox": [ + 223, + 155, + 777, + 589 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/f2a43d92e204323547645b405ac3fd1cc9f21e20c0e32dac4ecd4054bd961f8a.jpg", + "image_caption": [ + "(b) Small Objects", + "(c) Big Objects", + "(d) Different Locations", + "Fig. 5: Visual Results. (a) We showcase that our method applies to a diverse range of objects and scenes. The global view (top row) the overall composition quality and object-centric local view (bottom two rows) for the fidelity of stylized textures are demonstrated. For dim scenes, we additionally render objects without the estimated lighting condition $(w / o\\text{light})$ for illustrative purposes. Additionally, we showcase that our method applies to both (b) small and (c) big objects, as well as (d) different placing locations. The texture of the television, for example, adjusts the texture to match its surroundings." + ], + "image_footnote": [], + "bbox": [ + 223, + 604, + 777, + 679 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Scene-Conditional 3D Object Stylization and Composition", + "bbox": [ + 341, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.4 Lighting Model", + "text_level": 1, + "bbox": [ + 215, + 146, + 388, + 161 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To seamlessly integrate the object into the scene, we estimate the lighting conditions, as outlined in Fig. 4. To light the object, we utilize a high dynamic range (HDR) environment map $\\mathcal{E} \\in \\mathbb{R}^{H_e \\times W_e \\times 3}$ , which is well-suited for representing natural illumination. Since the given 2D scene image only captures a small angle of the full 360-degree environment map, we first use the image as input to estimate a low dynamic range (LDR) environment map. This process depends on the scene type (i.e., indoor vs. outdoor) and is explained in the following paragraphs. Given the LDR map, we then convert it to an HDR map by estimating a light scale (i.e., a scalar) for each bright region, which is thresholded from the LDR map directly (e.g., where intensity $\\mathcal{I}_{i,j} \\geq 0.8$ ). This scalar is then multiplied by the RGB values to approximate an HDR map. Note that we use the HDR map only to light the object. It thus does not need to be of high accuracy to make the object look plausible.", + "bbox": [ + 212, + 171, + 787, + 368 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Indoor Scenes. The well-bounded 2D visual geometry of indoor scenes is primarily characterized by enclosed spaces and interconnected planes, facilitating straightforward (rough) 3D reconstruction. To simulate the physical lighting of indoor scenes, we lift each pixel from the 2D scene to 3D space using its estimated depth $\\mathcal{D}$ [13, 44]. We perform world-to-camera and spherical-to-equirectangular coordinate transformations in succession, unwrapping 3D points into a latitude-longitude formatted LDR map. In post-processing, we automatically fill holes and remove small isolated regions. We fill the unseen region (i.e. areas behind the camera) with the average RGB value of the 2D scene. Two scalars are optimized: one for the far light region ( $\\triangleq \\{i,j \\mid \\mathcal{D}_{i,j} \\geq \\tau_d \\wedge \\mathcal{I}_{i,j} \\geq \\tau_f\\}$ ) and another for the near light region ( $\\triangleq \\{i,j \\mid \\mathcal{D}_{i,j} < \\tau_d \\wedge \\mathcal{I}_{i,j} \\geq \\tau_n\\}$ ).", + "bbox": [ + 212, + 383, + 787, + 551 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Outdoor Scenes. 3D reconstruction of outdoor scenes is challenging due to their inherently unbounded geometric structure. Furthermore, in most outdoor scenes, the sun is the dominant light source, and the position change of an object has a negligible effect on the angle shift of the light. Therefore, instead of painstakingly recovering location-specific viewpoints, we use a single-view to panorama outpainting model [2], taking its generated output as the initial LDR map. We optimize one scalar for the far light area with positive elevation $(\\bigtriangleup \\{i,j\\mid \\frac{i}{H_c} < 0.5\\vee \\mathcal{I}_{i,j}\\geq \\tau_o\\})$ , leaving the near light area $(\\bigtriangleup \\varnothing)$ unoptimized. While more sophisticated designs, such as a position estimator [27] or sky modeling [22], could be incorporated, we find our approach robust enough to provide a sufficient estimation.", + "bbox": [ + 212, + 565, + 787, + 731 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "3.5 Light-Capturing Apparatus", + "text_level": 1, + "bbox": [ + 215, + 753, + 488, + 768 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To estimate light scales, we place 3D objects with a fixed appearance model into the 2D scene and then optimize the lighting with VSD as described in Eq. (1) to anchor the lighting impact imposed by the scene. Inspired by the traditional inverse rendering setup [11] where the environment map can almost be perfectly", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "J. Zhou et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/17521f961a6554fcbf31cb7c62e6723b50a858eeee7ecbfb620b9a99635b6f06.jpg", + "image_caption": [ + "Fig. 6: Qualitative Comparison. We compare our the generated global view with the scene-object compositing method: CDC [18] and the local view with mesh texturing methods: Prolific Dreamer [55], Fantasia3D [7], and TEXTure [45]. s.gd. denotes similarity guidance. g.gd. and r.f.i denotes global-view guidance and reference feature injection, respectively. Our method achieves both seamless scene-object composition as well as texture adaptation of high fidelity. See sup. mat. for extensive comparison." + ], + "image_footnote": [], + "bbox": [ + 222, + 145, + 428, + 266 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/6b283a43f3b39b311ba05d1c1b0642bb08972f66ad04ebc31a69f7cb1890d016.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 429, + 151, + 500, + 261 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/6ef79b9c0fc85302e909a028449843bbc17886c21fcfc4dbe31a8fd51a9c9120.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 146, + 772, + 266 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "reconstructed from objects' reflections, we introduce a novel concept incorporating a virtual light-capturing apparatus alongside the object of interest during the optimization process. We insert a white sphere made of a smooth diffuse material into the scene and guide the lighting optimization with the diffusion model using the text prompt \"A gigantic diffuse white (spray-painted) sphere (ball)\". The white diffuse sphere proves advantageous in stabilizing lighting estimations for scenes with potentially strong light sources, as the reflected intensity from a white diffuse object closely approximates the intensity of the environmental lighting. In our initial experiments, we also tested spheres with mirrored and matte silver materials but found them less beneficial.", + "bbox": [ + 212, + 364, + 787, + 517 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "3.6 Light-Conditional Guidance", + "text_level": 1, + "bbox": [ + 215, + 537, + 491, + 553 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The absence of light conditions in text prompts makes the generated images only loosely constrained on lighting, which makes it challenging for diffusion models to score the renderings accurately. A white sphere looks gray in a darker environment, so light-agnostic prompts can engender an overestimation of the brightness for dim scenes. Therefore, we append \"in a dark environment\" if the average intensity of the background and light areas are below certain thresholds. We condition the LoRA-tuned model $\\epsilon_{\\psi}$ by concatenating the light scales with the camera extrinsic [55] as class embeddings for the U-Net. This improves light estimation, especially for outdoor scenes and dimly lit scenes. For scenes with atypical lighting that significantly alters the appearance of objects, we manually append color prompts such as \"in blue tint\" for the seabed bathed in blue light, or \"in red illumination\" for a nightclub enveloped in red atmospheric light.", + "bbox": [ + 212, + 561, + 787, + 743 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 763, + 375, + 781 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In this section, we show extensive qualitative and quantitative results of our method in various settings. Please see the appendix for many more examples, animations and discussion on limitations and negative impact.", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Scene-Conditional 3D Object Stylization and Composition", + "bbox": [ + 341, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/c0e7e34a89e1d7b3727f0ce0401f65198ca80b4fa9ea161895a8ea9d32147c2c.jpg", + "image_caption": [ + "Fig. 7: Ablation on Texture Adaptation. We ablate global-view guidance $(g.gd.)$ and reference feature injection $(r.f.i)$ . -ctrl. and $+\\mathrm{ctrl}$ denotes less or more control in feature injection (smaller or higher $s_{\\mathrm{f}}$ & $p$ ) for Identity Preservation requirement." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 501, + 226 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/622d9c7aeddf1d5a90381c5e85d951827bdf5b99bdd75c63928d963b80dbde80.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 157, + 784, + 228 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.1 Qualitative Evaluation", + "text_level": 1, + "bbox": [ + 215, + 291, + 447, + 306 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We first present visual results and a case study to demonstrate the generalizability of the proposed method. We then compare our method with others that specifically focus on either scene composition or object stylization.", + "bbox": [ + 212, + 325, + 782, + 372 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Main Results. To demonstrate the generalizability of the framework and the impact of different scenes on the inserted objects, we present visual results in Fig. 5. Our method successfully blends the objects into various environments, achieving photorealistic adaptation for both appearance and lighting. This includes scenarios drawn from both real-world and fantasy settings. For example, the dimly-lit umbrella in Fig. 5 casts a shadow that is perfectly aligned with its handle and the light direction of the scene, substantially increasing the photorealism of the composition; The Belweder TV set is wrapped with gray mud consistent with the scene while its original structural details get maintained.", + "bbox": [ + 212, + 395, + 787, + 532 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Comparison on Scene Composition. We showcase the result from CDC [18] using LLM-prompted text prompts for a fair comparison. As shown in Fig. 6, CDC loses the structural details of the object while providing limited adaptation for texture transfer. Moreover, the composed scene appears artificial and synthetic, giving the impression that the sofa is floating above the mud. Our experiments reveal that the CDC often leads to object removal, particularly when the re-painting [31] is employed. Therefore, we further incorporate similarity guidance similar to the classifier guidance [12] by adding $\\sqrt{1 - \\alpha_t} \\cdot \\frac{\\partial (||\\pmb{x}_0^r - \\hat{\\pmb{x}}_\\phi ||^2)}{\\partial \\pmb{x}_t}$ to the predicted gradient $\\epsilon_{\\phi}(\\pmb{x}_t)$ each denoising step referred as CDC w/s.gd., but find it provides limited improvement.", + "bbox": [ + 212, + 554, + 787, + 710 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Comparison on Object Stylization. We showcase results from mesh texturing method [7, 45, 55] using our LLM-prompted text prompts for a fair comparison. As shown in Fig. 6, although achieving realistic texture adaptation, these mesh texturing methods fall short in preserving the original structural details of the object and in utilizing scene-level information. Essentially, These methods only satisfy the Environmental Influence requirement via text prompts and do not meet all the criteria of our intended goal.", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "J. Zhou et al.", + "bbox": [ + 271, + 114, + 362, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/b74ac5b0c11eba0c7a91346562d1cf1238f02048e2fcea86ad685cd3894bc429.jpg", + "table_caption": [ + "Table 2: Quantitative Evaluation. (a) We measure LPIPS for controllability, $\\mathrm{CLIP}_{\\mathrm{dir}}$ and $\\mathrm{CLIP}_{\\mathrm{SI}}$ for composability. (b) We present the percentages of users who prefer our method over other methods: CDC [18] and InstructP2P [3]." + ], + "table_footnote": [ + "(a) Visual Metrics." + ], + "table_body": "
MethodLPIPS ↓CLIPdir ↑CLIPSI ↑
InstructP2P [3]0.00200.01750.1852
Ours (T.A.)0.00140.02010.2006
CDC [18]0.00530.00420.1996
Ours (T.A. + L.E.)0.00150.02060.2013
", + "bbox": [ + 217, + 198, + 514, + 266 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/a34a5791a67df410e9e061b5a04f11891674703dff7790493bd3e1d38a60354e.jpg", + "table_caption": [], + "table_footnote": [ + "(b) User study for human preference." + ], + "table_body": "
OursDrawOther
Ours vs. CDC [18]71%20%9%
Ours vs. InstructP2P [3]74%6%20%
", + "bbox": [ + 519, + 208, + 784, + 255 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.2 Quantitative Evaluation", + "text_level": 1, + "bbox": [ + 215, + 285, + 459, + 300 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We collected 34 composites between diverse objects and scenes, each with 3 random views, yielding a total of 102 images. Besides CDC [18], we compare with InstructPix2Pix [3] by directly copy-pasting the object with adapted texture onto the scene. We measure LPIPS for controllability, $\\mathrm{CLIP}_{\\mathrm{dir}}$ and $\\mathrm{CLIP}_{\\mathrm{SI}}$ for composability, similar to CDC, and report the results in Tab. 2. Compared with InstructPix2Pix our method for texture adaptation (T.A) alone excels in controllability and composability. Light estimation (L.E.) further improves results.", + "bbox": [ + 212, + 309, + 782, + 415 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We also conducted a user study in Tab. 2 with 42 participants to compare the realism of our compositing under the condition that the object identity must be preserved. Our method is preferred over prior work in over $70\\%$ of the cases.", + "bbox": [ + 212, + 416, + 784, + 460 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.3 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 483, + 388, + 500 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Identity Preservation. To illustrate how the Identity Preservation requirement is satisfied by reference feature injection (r.f.i.), we showcase the leather sofa in the swamp with less and more control (Fig. 7). The setup with less control drifts to a visually different sofa with too much mud wrapped around while more control generates a less muddy and adapted sofa compared to the reference image. Both adapted textures are less visually plausible. Additionally, when comparing Ours w/o g.gd. & r.f.i. to the reference image, it is evident that the identity of the sofa is lost.", + "bbox": [ + 212, + 508, + 787, + 628 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Global-View Guidance. To showcase the necessity of global-view guidance we conduct texture adaptation without it as Ours w/o g.gd. in Fig. 7. The adapted texture without seeing the full scene fails to adapt the color of the mud to that of the scene image, contrasting starkly with the scene. Experiments ablating the optimal global-view guidance set-up (e.g., solid-color background augmentation, usage of inpainting model, etc.) are detailed in the appendix.", + "bbox": [ + 214, + 643, + 787, + 737 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Light Estimation. In Fig. 1, the combination of accurately estimated lighting and shadows significantly enhances the photorealism of the composed image (compare the image on the left, which includes lighting, to those in the top right corner, which are without lighting). To further test the validity of the estimated light, we render images omitting the texture adaptation step. Fig. 8 demonstrates that the estimated light considers multiple factors, including its", + "bbox": [ + 214, + 750, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Scene-Conditional 3D Object Stylization and Composition", + "bbox": [ + 341, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/55d0f2c0196aa81bbf235a891b65a25da4524bad112f415ffa64c0df7d6f975b.jpg", + "image_caption": [ + "Fig. 8: Ablation on Light Estimation. We evaluate the effectiveness of light estimation by demonstrating the lighting effects on a sofa in various indoor and outdoor scenes. Overlaid in the left corners are objects rendered with the estimated light against a white background. In the right corners are objects before the light estimation." + ], + "image_footnote": [], + "bbox": [ + 272, + 143, + 725, + 494 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "direction relative to the specific locations of objects in indoor scenes and its impact on the appearance of objects under atmospheric colored lighting conditions. Importantly, the estimated HDR maps, paired with the proposed scene composition, facilitate the realistic integration of shadows.", + "bbox": [ + 212, + 578, + 787, + 640 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 675, + 359, + 690 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we are motivated by a practical goal—what if a 3D object is placed into a 2D scene?—and propose a novel framework that allows (1) stylizing the object with an adapted texture that aligns with the given scene; and (2) achieving a photorealistic scene composition with the aid of estimated light of the environment. This is enabled by physically-based differentiable ray tracing and diffusion models used as guidance. Our method allows for automatic adaptation and blending of existing objects into a variety of scenes, making it useful for both 2D and 3D downstream applications, such as visual media and video games.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "J. Zhou et al.", + "bbox": [ + 271, + 114, + 362, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement. This work is supported by the UKRI grant: Turing AI Fellowship EP/W002981/1 and EPSRC/MURI grant: EP/N019474/1. J. Zhou is also supported by the Horizon Robotics. We thank Luke Melas-Kyriazi and Fabio Pizzati for their helpful discussions.", + "bbox": [ + 215, + 146, + 787, + 205 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 229, + 321, + 244 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. https://polyhaven.com, https://polyhaven.com", + "2. Akimoto, N., Matsuo, Y., Aoki, Y.: Diverse plausible 360-degree image outpainting for efficient 3dgc background creation. In: CVPR (2022)", + "3. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: CVPR (2023)", + "4. Cao, T., Kreis, K., Fidler, S., Sharp, N., Yin, K.: Texfusion: Synthesizing 3d textures with text-guided image diffusion models. In: ICCV (2023)", + "5. Chen, B.C., Kae, A.: Toward realistic image compositing with adversarial learning. In: CVPR (2019)", + "6. Chen, D.Z., Siddiqui, Y., Lee, H.Y., Tulyakov, S., Nießner, M.: Text2tex: Text-driven texture synthesis via diffusion models. arXiv preprint arXiv:2303.11396 (2023)", + "7. Chen, R., Chen, Y., Jiao, N., Jia, K.: Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In: ICCV (2023)", + "8. Cong, W., Niu, L., Zhang, J., Liang, J., Zhang, L.: Bargainnet: Background-guided domain translation for image harmonization. In: ICME (2021)", + "9. Cong, W., Zhang, J., Niu, L., Liu, L., Ling, Z., Li, W., Zhang, L.: Dovenet: Deep image harmonization via domain verification. In: CVPR (2020)", + "0. Deitke, M., Schwenk, D., Salvador, J., Weihs, L., Michel, O., VanderBilt, E., Schmidt, L., Ehsani, K., Kembhavi, A., Farhadi, A.: Objaverse: A universe of annotated 3d objects. In: CVPR (2023)", + "1. Deng, Y., Li, X., Liu, S., Yang, M.H.: Dip: Differentiable interreflection-aware physics-based inverse rendering. arXiv preprint arXiv:2212.04705 (2022)", + "2. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. NeurIPS (2021)", + "3. Eftekhar, A., Sax, A., Malik, J., Zamir, A.: Omnidata: A scalable pipeline for making multi-task mid-level vision datasets from 3d scans. In: ICCV (2021)", + "4. Gao, J., Shen, T., Wang, Z., Chen, W., Yin, K., Li, D., Litany, O., Gojcic, Z., Fidler, S.: Get3d: A generative model of high quality 3d textured shapes learned from images. NeurIPS (2022)", + "5. Gardner, M.A., Hold-Geoffroy, Y., Sunkavalli, K., Gagné, C., Lalonde, J.F.: Deep parametric indoor lighting estimation. In: ICCV (2019)", + "6. Gardner, M.A., Sunkavalli, K., Yumer, E., Shen, X., Gambaretto, E., Gagné, C., Lalonde, J.F.: Learning to predict indoor illumination from a single image. arXiv preprint arXiv:1704.00090 (2017)", + "7. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. NeurIPS (2014)", + "8. Hachnochi, R., Zhao, M., Orzech, N., Gal, R., Mahdavi-Amiri, A., Cohen-Or, D., Bermano, A.H.: Cross-domain compositing with pretrained diffusion models. arXiv preprint arXiv:2302.10167 (2023)", + "9. Haque, A., Tancik, M., Efros, A., Holynski, A., Kanazawa, A.: Instruct-nerf2nerf: Editing 3d scenes with instructions. In: CVPR (2023)" + ], + "bbox": [ + 225, + 261, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Scene-Conditional 3D Object Stylization and Composition", + "bbox": [ + 341, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "20. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. NeurIPS (2020)", + "21. Ho, J., Salimans, T.: Classifier-free diffusion guidance. NeurIPSW (2021)", + "22. Hold-Geoffroy, Y., Athawale, A., Lalonde, J.F.: Deep sky modeling for single image outdoor lighting estimation. In: CVPR (2019)", + "23. Hold-Geoffroy, Y., Sunkavalli, K., Hadap, S., Gambaretto, E., Lalonde, J.F.: Deep outdoor illumination estimation. In: CVPR (2017)", + "24. Jakob, W., Speierer, S., Roussel, N., Vicini, D.: Dr.jit: A just-in-time compiler for differentiable rendering. TOG (2022)", + "25. Jiang, Y., Zhang, H., Zhang, J., Wang, Y., Lin, Z., Sunkavalli, K., Chen, S., Amirghodsi, S., Kong, S., Wang, Z.: Ssh: A self-supervised framework for image harmonization. In: ICCV (2021)", + "26. Kamata, H., Sakuma, Y., Hayakawa, A., Ishii, M., Narihira, T.: Instruct 3d-to-3d: Text instruction guided 3d-to-3d conversion. arXiv preprint arXiv:2303.15780 (2023)", + "27. Karimi Dastjerdi, M.R., Eisenmann, J., Hold-Geoffroy, Y., Lalonde, J.F.: Everlight: Indoor-outdoor editable hdr lighting estimation. In: ICCV (2023)", + "28. Karimi Dastjerdi, M.R., Hold-Geoffroy, Y., Eisenmann, J., Khodadadeh, S., Lalonde, J.F.: Guided co-modulated GAN for 360 field of view extrapolation. 3DV (2022)", + "29. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3d: High-resolution text-to-3d content creation. In: CVPR (2023)", + "30. Liu, M., Xu, C., Jin, H., Chen, L., Xu, Z., Su, H., et al.: One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. arXiv preprint arXiv:2306.16928 (2023)", + "31. Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Van Gool, L.: Repaint: Inpainting using denoising diffusion probabilistic models. In: CVPR (2022)", + "32. McAuley, S., Hill, S., Hoffman, N., Gotanda, Y., Smits, B., Burley, B., Martinez, A.: Practical physically-based shading in film and game production. In: SIGGRAPH (2012)", + "33. Metzer, G., Richardson, E., Patashnik, O., Giryes, R., Cohen-Or, D.: Latent-nerf for shape-guided generation of 3d shapes and textures. In: CVPR (2023)", + "34. Michel, O., Bar-On, R., Liu, R., Benaim, S., Hanocka, R.: Text2mesh: Text-driven neural stylization for meshes (2022)", + "35. Mohammad Khalid, N., Xie, T., Belilovsky, E., Popa, T.: Clip-mesh: Generating textured meshes from text using pretrained image-text models. In: SIGGRAPH Asia (2022)", + "36. Munkberg, J., Hasselgren, J., Shen, T., Gao, J., Chen, W., Evans, A., Müller, T., Fidler, S.: Extracting triangular 3d models, materials, and lighting from images. In: CVPR (2022)", + "37. Oechsle, M., Mescheder, L., Niemeyer, M., Strauss, T., Geiger, A.: Texture fields: Learning texture representations in function space. In: ICCV (2019)", + "38. OpenAI: Gpt-4 technical report (2023)", + "39. Parmar, G., Kumar Singh, K., Zhang, R., Li, Y., Lu, J., Zhu, J.Y.: Zero-shot image-to-image translation. In: SIGGRAPH (2023)", + "40. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. ICLR (2023)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "J. Zhou et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "41. Qian, G., Mai, J., Hamdi, A., Ren, J., Siarohin, A., Li, B., Lee, H.Y., Skorokhodov, I., Wonka, P., Tulyakov, S., Ghanem, B.: Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. arXiv preprint arXiv:2306.17843 (2023)", + "42. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021)", + "43. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: ICML (2021)", + "44. Ranftl, R., Lasinger, K., Hafner, D., Schindler, K., Koltun, V.: Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. PAMI (2022)", + "45. Richardson, E., Metzer, G., Alaluf, Y., Giryes, R., Cohen-Or, D.: Texture: Text-guided texturing of 3d shapes. SIGGRAPH (2023)", + "46. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models (2021)", + "47. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS (2022)", + "48. Siddiqui, Y., Thies, J., Ma, F., Shan, Q., Nießner, M., Dai, A.: Texturify: Generating textures on 3d shape surfaces. In: ECCV (2022)", + "49. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. In: ICLR (2021), https://openreview.net/forum?id=PxTIG12RRHS", + "50. Song, Y., Zhang, Z., Lin, Z., Cohen, S., Price, B., Zhang, J., Kim, S.Y., Aliaga, D.: Objectstitch: Generative object compositing. CVPR (2023)", + "51. Tsalicoglou, C., Manhardt, F., Tonioni, A., Niemeyer, M., Tombari, F.: Textmesh: Generation of realistic 3d meshes from text prompts (2024)", + "52. Tumanyan, N., Geyer, M., Bagon, S., Dekel, T.: Plug-and-play diffusion features for text-driven image-to-image translation. In: CVPR (2023)", + "53. Walter, B., Marschner, S.R., Li, H., Torrance, K.E.: Microfacet models for refraction through rough surfaces. In: EGSR (2007)", + "54. Wang, H., Du, X., Li, J., Yeh, R.A., Shakhnarovich, G.: Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In: CVPR (2023)", + "55. Wang, Z., Lu, C., Wang, Y., Bao, F., Li, C., Su, H., Zhu, J.: Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. NeurIPS (2023)", + "56. Wei, L.Y., Lefebvre, S., Kwatra, V., Turk, G.: State of the art in example-based texture synthesis. EG-STAR (2009)", + "57. Xiao, J., Ehinger, K.A., Oliva, A., Torralba, A.: Recognizing scene viewpoint using panoramic place representation. In: ICCV (2012)", + "58. Zhang, J., Sunkavalli, K., Hold-Geoffroy, Y., Hadap, S., Eisenman, J., Lalonde, J.F.: All-weather deep outdoor lighting estimation. In: CVPR (2019)" + ], + "bbox": [ + 212, + 146, + 787, + 742 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Scene-Conditional 3D Object Stylization and Composition", + "bbox": [ + 341, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Scene-Conditional 3D Object Stylization and Composition/ed573969-4bd1-4a9b-b748-12ae2bb2b2d3_model.json b/2024/Scene-Conditional 3D Object Stylization and Composition/ed573969-4bd1-4a9b-b748-12ae2bb2b2d3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6409f5a7f9b9426d16b412a67eb8d1323659bb41 --- /dev/null +++ b/2024/Scene-Conditional 3D Object Stylization and Composition/ed573969-4bd1-4a9b-b748-12ae2bb2b2d3_model.json @@ -0,0 +1,2247 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.276, + 0.142, + 0.728, + 0.187 + ], + "angle": 0, + "content": "Scene-Conditional 3D Object Stylization and Composition" + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.214, + 0.765, + 0.229 + ], + "angle": 0, + "content": "Jinghao Zhou\\(^{1}\\), Tomas Jakab\\(^{1}\\), Philip Torr\\(^{1}\\), and Christian Rupprecht\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.358, + 0.241, + 0.645, + 0.27 + ], + "angle": 0, + "content": "University of Oxford {jinghao,tomj,chrisr}@robots.ox.ac.uk" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.3, + 0.787, + 0.537 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.548, + 0.788, + 0.631 + ], + "angle": 0, + "content": "Fig. 1: We present a framework that adapts a 3D object's appearance to a location in a 2D scene. It creates an image where the 3D object is seamlessly blended (left & bottom right), with its appearance influenced by the scene's environmental conditions and lighting effects. Moreover, the stylized object with adapted textures (top right), rendered here without the estimated lighting condition for illustrative purposes, can be further utilized as 3D assets for downstream tasks such as video games." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.66, + 0.751, + 0.841 + ], + "angle": 0, + "content": "Abstract. Recently, 3D generative models have made impressive progress, enabling the generation of almost arbitrary 3D assets from text or image inputs. However, these approaches generate objects in isolation without any consideration for the scene where they will eventually be placed. In this paper, we propose a framework that allows for the stylization of an existing 3D asset to fit into a given 2D scene, and additionally produce a photorealistic composition as if the asset was placed within the environment. This not only opens up a new level of control for object stylization, for example, the same assets can be stylized to reflect changes in the environment, such as summer to winter or fantasy versus futuristic settings—but also makes the object-scene composition more controllable. We achieve this by combining modeling and optimizing the object's texture and environmental lighting through differentiable ray" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "J. Zhou et al." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.148, + 0.744, + 0.19 + ], + "angle": 0, + "content": "tracing with image priors from pre-trained text-to-image diffusion models. We demonstrate that our method applies to a wide variety of indoor and outdoor scenes and arbitrary objects. See also our project page." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.21, + 0.376, + 0.226 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.236, + 0.785, + 0.296 + ], + "angle": 0, + "content": "After the success of generative models [20,43,46] for images in computer vision, the community is now interested in lifting these models to 3D [40,54] through advances in architectures, data, and training. These lifted 3D models demonstrate plausible 3D representations of objects conditioned on images or texts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.297, + 0.787, + 0.357 + ], + "angle": 0, + "content": "While generating whole objects—e.g. an astronaut riding a horse—is a challenging problem that generative models are becoming able to solve, in practice, these assets are difficult to adapt for use in downstream applications or existing virtual environments such as video games as they are generated without context." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.358, + 0.787, + 0.433 + ], + "angle": 0, + "content": "This paper aims to provide a mechanism to close the appearance gap between 3D objects and environments. We formulate this problem through the lens of a creative tool. The intended goal of this paper is to align the appearance of a 3D object to a customized 2D scene as we place it into a specific location, a challenging task that requires fine-grained visual control for convincing results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.433, + 0.787, + 0.584 + ], + "angle": 0, + "content": "We are interested in producing a photorealistic scene-object composition where the 3D object is seamlessly blended. For instance, an object placed in a rainy night scene should appear wet and dimly lit, and if the scene is muddy and sunny, the object should also look muddy and cast a strong shadow. To achieve a high level of realism on a fine-grained level, the mud on the object must match the scene's mud in appearance and the direction of the shadow must match where the sun is located. Instead of simply resorting to 2D blending operations [18, 50], we explicitly model an object's appearance and lighting in 3D space, which enables more accurate and regularized control of the object's appearance in the composed image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.585, + 0.787, + 0.718 + ], + "angle": 0, + "content": "We also aim to preserve the underlying object geometry and appearance factors unrelated to the scene environment. This task is highly practical; for example, for media design, when presenting a 3D asset of a product within a realistic scene, we want the product to blend naturally while remaining recognizable. Manually achieving this would require a skilled artist to modify the object's texture to match the scene, a challenging and time-consuming process. Our paper introduces a method that automates this task. It accepts a 3D object with texture, a 2D scene, and the object's position within the scene and outputs an adapted texture consistent with the scene." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.787, + 0.809 + ], + "angle": 0, + "content": "To accomplish this, we leverage recent advances in generative diffusion models [46, 47] for images, which are trained on a large-scale dataset of internet images. We conduct an optimization procedure where we compose the 2D renderings of the 3D object with the 2D scene at a specified location and use a pre-trained diffusion model to \"critique\" the realism of the composed image, providing a gradient signal for optimization." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.841 + ], + "angle": 0, + "content": "The challenge is that diffusion models, such as Stable Diffusion [46], are trained on 2D images and lack any notion of 3D geometry. Naively optimizing" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.342, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "Scene-Conditional 3D Object Stylization and Composition" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.223 + ], + "angle": 0, + "content": "or denoising in 2D space [18,50] would result in losing the object's 3D geometry. To prevent this, we optimize the object's content in 3D by randomly rotating the object during optimization and rendering it from various viewpoints, which helps to align the adapted texture with the object's geometry and avoid overfitting to a single viewpoint." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.785, + 0.403 + ], + "angle": 0, + "content": "Notably, randomly rotating the 3D object stops lighting effects from being baked into the texture, which is desirable. Yet, it also prevents the object from naturally blending with the scene in a single viewpoint. To counteract this, we predict the scene's lighting and the shadows cast by the object as separate components. For this purpose, we propose a novel technique inspired by a common physical approach from the computer graphics industry, where reflective and diffuse spheres with known material parameters are placed and photographed in a real-world scene to capture the environmental lighting conditions. We adopt this concept and place a virtual white diffuse sphere inside the scene during optimization while conditioning the generative model to expect a white sphere within the scene, thereby capturing the model's interpretation of the scene's lighting in an easily extracted form." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.404, + 0.785, + 0.495 + ], + "angle": 0, + "content": "Optimizing the object's appearance in 3D offers further significant advantages: The user can change the object's pose without re-running the optimization, as the texture is meaningful from all viewpoints, and view-dependent effects are modeled separately by the lighting component. Additionally, the 3D object with the adapted texture can be used as a standalone asset in other downstream tasks (e.g. in a video game)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.495, + 0.785, + 0.556 + ], + "angle": 0, + "content": "In summary, we present a framework that (1) enables the stylization of a 3D object, which can also be used as a standalone asset, given a 2D scene and its location within it; (2) achieves photorealistic scene-object compositing with the help of a novel mechanism to estimate scene lighting from a single image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.556, + 0.785, + 0.601 + ], + "angle": 0, + "content": "We further verify the effectiveness of our method in extensive experiments, demonstrating that our approach can realistically adapt objects to a diverse set of environments, laying the groundwork for practical applications." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.62, + 0.388, + 0.636 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.648, + 0.784, + 0.724 + ], + "angle": 0, + "content": "The purpose of this paper is to adapt the appearance of a 3D object to align with a customized 2D scene. This solves a more complicated problem than prior works, which often only tackle individual components, including but not limited to 2D scene-object compositing, light estimation, and texturing, but never as a whole. We compare with these methods next and detail their differences in Tab. 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.785, + 0.841 + ], + "angle": 0, + "content": "Scene-Object Compositing. Scene-object compositing is a challenging task that requires a delicate balance between visual transfer and control. It has been extensively researched [5,8,9,25]. Recent methods condition the denoising process of diffusion models [20,49] on object images. OST [50] aims to seamlessly integrate a 2D object image into a 2D scene by modifying the keys and values in the attention blocks. CDC [18] considers composing the 2D renderings of a 3D object with a 2D scene by directly modifying denoised latent." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "J. Zhou et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.203 + ], + "angle": 0, + "content": "Table 1: Comparison with related work. Rep. denotes representation being 2D (□) or 3D (）。Env. Inf. denotes Environmental Influence and I.D. Pre. denotes Identity Preservation. Parameters for the lighting and the appearance model are either scene-agnostic constant (），scene-conditioned predictable (），or learnable (）。" + }, + { + "type": "table", + "bbox": [ + 0.224, + 0.213, + 0.781, + 0.384 + ], + "angle": 0, + "content": "
MethodSceneObjectCompose
Rep.LightingShadowRep.AppearanceEnv. Inf.I.D. Pre.
OST [50]□-×□-××Blend
CDC [18]□Ambient×Diffuse×××Blend
IGAN [28]□LDR✓Diffuse×××Copy & Paste
EVL [27]□HDR✓Diffuse×××Copy & Paste
PrDr [55]-Ambient×Diffuse✓×-
Fan3D [7]-HDR×PBR✓×-
Inst3D [26]-Point×Diffuse×✓-
Ours□HDR✓PBR✓✓Copy & Paste
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.393, + 0.785, + 0.455 + ], + "angle": 0, + "content": "However, these methods rely entirely on the 2D denoising process and do not explicitly model the object's geometry, which often results in a significant loss of structural details and visual identity. This severely hampers these methods' ability to adapt to the environmental and lighting conditions of the scene." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.465, + 0.785, + 0.587 + ], + "angle": 0, + "content": "Light Estimation from Single-View Image. Estimating light conditions from a single-view 2D image is of practical use in the real world and has been extensively studied for both indoor [15,16] and outdoor scenes [22,23,58]. The distinction between these two settings has been significantly narrowed recently [27, 28], thanks to the outpainting capabilities of diffusion models, which have contributed to a unified framework. With additional training on large high dynamic range (HDR) datasets such as SUN360 [57], these methods can approximate the location and intensity of the light source, enabling realistic shadows." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.587, + 0.787, + 0.632 + ], + "angle": 0, + "content": "However, these methods estimate only the lighting of the scene. They are typically indifferent to how objects should be stylized to match the scene, neglecting the potential environmental influences on the textures of objects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Mesh Texturing. Classic automated approaches [56] for mesh texturing are often limited and apply only simple texture patterns. In contrast, methods [14, 37, 48] based on Generative Adversarial Networks [17] (GANs) are more capable of synthesizing detailed textures. Recently, text-conditioned texture generation [34, 35] has been made possible through the optimization of CLIP-based objectives [42]. The latest works now employ powerful text-to-image diffusion models [20, 49], either through optimization [7, 29, 33, 51, 55] using Score Distillation Sampling [40] (SDS), or by wrapping the generated 2D textures onto the mesh [4, 6, 45]. Importantly, Variational Score Distillation [55] (VSD) as an improved optimization objective and the adoption of PBR model [7] have greatly boosted the quality of generated textures. Additionally, an increasing level of requirement for more visual control from users has been partially met [19, 26] through image-conditional diffusion models [3] to preserve objects' identities." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.342, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Scene-Conditional 3D Object Stylization and Composition" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.144, + 0.754, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.292, + 0.788, + 0.388 + ], + "angle": 0, + "content": "Fig. 2: Framework. We learn an environment map and a texture map separately from the 2D supervision. We initialize (init.) the environment map with an LDR map estimated from the 2D scene and learn light multiplying scales bright areas, yielding an HDR map. We employ the PBR material model for texture maps, encoded via MLP with positional encoding. The object is rendered through a differentiable ray tracer and further composed with the scene background, receiving gradients from Stable Diffusion (SD) in the latent space." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.398, + 0.788, + 0.445 + ], + "angle": 0, + "content": "However, these works focus on texturing the object in isolation from the scene and do not consider the homogeneity of the composition when these assets are placed in certain environments, which limits their practical usage." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.468, + 0.331, + 0.484 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.501, + 0.788, + 0.592 + ], + "angle": 0, + "content": "The overall structure of our framework can be found in Fig. 2. Given a textured 3D object, a 2D image of a scene, and a 2D location where to place the object, we first adapt the texture of the 3D object (Secs. 3.2 and 3.3) to align with the scene. To enable a photorealistic composition of the object and the scene, we further estimate the lighting (Secs. 3.4 to 3.6) and apply it during the final rendering." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.615, + 0.373, + 0.631 + ], + "angle": 0, + "content": "3.1 Optimization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.642, + 0.787, + 0.673 + ], + "angle": 0, + "content": "To enable end-to-end optimization, we adopt a differentiable renderer to render the object and score it using the priors from diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.688, + 0.788, + 0.78 + ], + "angle": 0, + "content": "Differentiable Renderer. We employ the differential renderer Mitsuba3 [24] (denoted as \\( g \\)), due to its capabilities as a physically-based ray-tracing renderer equipped with versatile rendering options, such as diverse BSDFs, the path tracer integrator, and using an environment map as a light source. This enables us to produce photo-realistic renders complete with shadows and reflections, forming the foundation for robust texture adaptation and light estimation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Diffusion Model as Guidance. To score the rendered image, we make use of a generative diffusion model [46, 47] \\(\\phi\\). Diffusion models have learned powerful priors of the image formation process in the real world from large-scale training" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "J. Zhou et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.785, + 0.178 + ], + "angle": 0, + "content": "on billion-scale web data. Similar to prior work [40,54], we supervise the rendered image \\(\\pmb{x}_0 = g(\\pmb{\\theta})\\) with a score-matching objective to obtain an image gradient" + }, + { + "type": "equation", + "bbox": [ + 0.352, + 0.19, + 0.786, + 0.223 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\boldsymbol {\\theta}} \\mathcal {L} (\\boldsymbol {x} _ {0}) \\triangleq \\mathbb {E} _ {t, \\epsilon} \\left[ w _ {t} \\left(\\epsilon_ {\\phi} (\\boldsymbol {x} _ {t}) - \\hat {\\epsilon} (\\boldsymbol {x} _ {t})\\right) \\frac {\\partial \\boldsymbol {x} _ {0}}{\\partial \\boldsymbol {\\theta}} \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.234, + 0.784, + 0.386 + ], + "angle": 0, + "content": "where \\(\\pmb{x}_t = \\sqrt{\\alpha_t}\\pmb{x}_0 + \\sqrt{1 - \\alpha_t}\\epsilon\\) is the noisy rendered image, \\(\\epsilon_{\\phi}(\\pmb{x}_t)\\) is an approximated score for a noisy real image, \\(t\\in \\mathbb{N}\\) indexes the discrete diffusion steps, and \\(\\hat{\\epsilon}\\) is a scoring function for \\(\\pmb{x}_t\\). Specifically, \\(\\hat{\\epsilon}\\) is set as constant \\(\\epsilon\\) in Score Distillation Sampling [40] (SDS), and \\(\\epsilon_{\\psi}\\) is an online LoRA-tuned network initialized from \\(\\epsilon_{\\phi}\\) in Variational Score Sampling [55] (VSD). Notably, we employ their interpolated variant given by \\(\\hat{\\epsilon} = \\lambda \\epsilon_{\\psi} + (1 - \\lambda)\\epsilon\\) with a \\(\\lambda\\) annealed from 1 to \\(\\lambda_e\\), helping to counteract the over-texturization often observed in VSD, especially for objects with simple surfaces. With residual of noises \\(\\epsilon_{\\phi} - \\hat{\\epsilon} = (\\epsilon_{\\phi} - \\epsilon) - \\lambda (\\epsilon_{\\psi} - \\epsilon)\\), we highlight its relation with the image gradient obtained in traditional inverse rendering setup \\(\\nabla_{\\theta}\\mathcal{L}(\\pmb{x}_0) = \\pmb{x}_0 - \\pmb{x}_{\\mathrm{gt}}\\) by reformulating it as" + }, + { + "type": "equation", + "bbox": [ + 0.346, + 0.399, + 0.785, + 0.431 + ], + "angle": 0, + "content": "\\[\n\\epsilon_ {\\phi} - \\hat {\\epsilon} = \\sqrt {\\frac {\\alpha_ {t}}{1 - \\alpha_ {t}}} [ (\\boldsymbol {x} _ {0} - \\hat {\\boldsymbol {x}} _ {\\phi}) - \\lambda (\\boldsymbol {x} _ {0} - \\hat {\\boldsymbol {x}} _ {\\psi}) ], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.443, + 0.785, + 0.548 + ], + "angle": 0, + "content": "where the target used to supervise \\(\\pmb{x}_0\\) is predicted clean image \\(\\hat{\\pmb{x}}_{\\phi}\\) from \\(\\epsilon_{\\phi}\\) as the positive and \\(\\hat{\\pmb{x}}_{\\psi}\\) from \\(\\epsilon_{\\psi}\\) as the negative. The technique enables nominal classifier-free guidance [21] and thus bypasses mode-seeking behaviors and saturated texture. Inspired by [45], we incorporate the inner product of the camera pose and object's surface normal \\(\\langle c, \\pmb{n} \\rangle\\) into \\(w_t\\) to penalize faces at small grazing angles. In practice, we use Stable Diffusion [46] (SD) and absorb the encoder into \\(g\\), optimizing in latent space." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.571, + 0.419, + 0.587 + ], + "angle": 0, + "content": "3.2 Appearance Model" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.598, + 0.784, + 0.675 + ], + "angle": 0, + "content": "Textured Mesh. We represent the 3D object as a textured mesh, which is the most common format for 3D assets in the industry and is compatible with many established workflows. Textured meshes are available in abundance in online datasets and marketplaces [1,10], and recently, one can also generate reasonable quality 3D meshes with text descriptions [7,29,55] or limited viewpoints [30,41]." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.69, + 0.786, + 0.84 + ], + "angle": 0, + "content": "Neural Texture. We represent the appearance of objects using a texture map \\(\\mathcal{T} \\in \\mathbb{R}^{H_t \\times W_t \\times 5}\\), facilitating compatibility of textured results with widely-used rendering engines, such as Blender. We parameterize the texture using a coordinate neural network that maps 3D coordinates to the texture attributes. We first map the 3D coordinates of vertices using a rasterizer [36] based on their UV coordinates into the space of the texture map, resulting in \\(\\mathcal{C} \\in \\mathbb{R}^{H_t \\times W_t \\times 3}\\). We then use a 2-layer parameterized MLP without bias and Hash positional encoding \\(\\varPhi: \\mathcal{C} \\to \\mathcal{T}\\) to map the coordinates to their texture attributes. For the final layer, we re-map the output activated after sigmoid according to pre-defined minimal and maximal values. The five channels in \\(\\mathcal{T}\\) correspond to the Physically-Based" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.342, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Scene-Conditional 3D Object Stylization and Composition" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.147, + 0.753, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.32, + 0.785, + 0.361 + ], + "angle": 0, + "content": "Fig. 3: Pipeline for texture adaptation. We initialize the neural texture from the reference object and inject the feature of reference renderings to the U-Net of SD. We use both local-view and global-view guidance." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.372, + 0.785, + 0.448 + ], + "angle": 0, + "content": "Rendering (PBR) shading model [32], which includes a diffuse term \\( k_{\\mathrm{d}} \\in \\mathbb{R}^{3} \\) and an isotropic specular GGX lobe [53] described by \\( k_{\\mathrm{rm}} = (k_r, k_m) \\in \\mathbb{R}^2 \\), with \\( k_r \\) representing roughness and \\( k_m \\) indicating metalness. The PBR model, with \\( k_r \\) affecting the scattering of light and \\( k_m \\) influencing the reflective properties, creates realistic light interactions between surfaces." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.47, + 0.494, + 0.484 + ], + "angle": 0, + "content": "3.3 Scene-Conditional Guidance" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.494, + 0.785, + 0.584 + ], + "angle": 0, + "content": "Achieving a realistic texture adaptation for 3D objects placed into 2D scenes requires three components: Environmental Influence involves adjusting the texture to reflect the environmental impact realistically; Identity Preservation aims to preserve the objects' unique visual aspects, and Blending focuses on guiding the texture adaptation to match the visual characteristic of the scene, ensuring seamless integration with the surroundings and avoiding stark contrasts." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.599, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Environmental Influence. To show the realistic effects of the scene environment on the object's appearance, we guide the diffusion model with text prompts. We find that a naive text prompt, combining a simple object description with a scene description, e.g., a leather sofa in a swamp, is insufficient to generate realistic textures that reflect the environmental impact on the object, as shown in the appendix. For instance, we would expect a sofa placed in a swamp to be dirty and possibly show signs of moss or other vegetation growth. Instead of manually crafting a detailed text prompt for each object and scene combination, we automatize the process by employing a large language model (LLM), specifically GPT-4 [38], to generate a text prompt for the image diffusion model. Given the simple description of the object (e.g., a leather sofa) and the scene (e.g., a swamp), we instruct the LLM first to analyze the potential impact of the scene's environment on the object's appearance and then to distil this information into a text prompt that describes the resulting appearance of the object (e.g., The leather sofa, partially submerged in the swamp, looks discolored and soggy, its once-polished surface marred by mud, moss, and the murky water of the swamp" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "J. Zhou et al." + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.146, + 0.788, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.23, + 0.788, + 0.272 + ], + "angle": 0, + "content": "Fig. 4: Pipeline for estimating the LDR map. We utilize tailored pipelines for indoor and outdoor scenes. Areas masked red and blue correspond to the far light \\(\\mathcal{O}\\) and near light region." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.286, + 0.789, + 0.332 + ], + "angle": 0, + "content": "environment.). We then use this automatically generated text prompt to guide the diffusion model in adapting the texture. The specific prompt design for this automated process is detailed in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.357, + 0.789, + 0.479 + ], + "angle": 0, + "content": "Identity Preservation. To prevent the transferred texture from drifting away from its original content, we regularize the denoising diffusion model with reference images, a technique that has been explored for controlled image generation [39, 52]. Suppose we have a rendered reference image \\(\\pmb{x}_0^r = g(\\pmb{\\theta}^r)\\) and its noised version \\(\\pmb{x}_t^r\\) as input forwarded to \\(\\epsilon_{\\phi}\\) and \\(\\epsilon_{\\psi}\\), for the Multi-Head Self-Attention (MHSA) layer in each Transformer block of \\(\\epsilon_{\\phi}\\)'s and \\(\\epsilon_{\\psi}\\)'s U-Nets, we concatenate the key \\(K^r\\) and value \\(V^r\\) of \\(\\pmb{x}_t^r\\) to the original \\(K\\) and \\(V\\) from \\(\\pmb{x}_t\\) yielding an image-conditional attention output" + }, + { + "type": "equation", + "bbox": [ + 0.35, + 0.494, + 0.787, + 0.513 + ], + "angle": 0, + "content": "\\[\nA (\\boldsymbol {x} _ {t}, \\cdot , \\boldsymbol {x} _ {t} ^ {r}) = \\mathrm {M H S A} (Q, [ K; K ^ {r} ], [ V; V ^ {r} ]), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.527, + 0.788, + 0.604 + ], + "angle": 0, + "content": "where \\(\\cdot\\) can either be text or \\(\\emptyset\\) for text conditioning. The text-unconditional attention \\(A(\\pmb{x}_t, \\emptyset, \\pmb{x}_t^r)\\) used for classifier-free guidance [21] is further modified as \\(s_c \\cdot A(\\pmb{x}_t, \\emptyset, \\emptyset) + (1 - s_c) \\cdot A(\\pmb{x}_t, \\emptyset, \\pmb{x}_t^r)\\) weighted by a control guidance scale \\(s_c\\). With the percentage \\(p\\) of blocks to be injected, the intensity of preservation can be controlled: \\(s_c = p = 1\\) provide the most and \\(s_c = p = 0\\) the least preservation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.79, + 0.842 + ], + "angle": 0, + "content": "Blending. The two components previously mentioned directly utilize object renderings, referred to as the local view, as inputs for vanilla SD. To further enhance the adaptation of the object's appearance to its surroundings, we input the scene image, with the object rendered onto it, referred to as the global view, into an additional inpainting SD [46]. We have found that the inpainting model provides superior conditioning on the surrounding environment depicted in the image, resulting in improved blending. However, the inpainting model can sometimes attempt to undesirably modify the inserted object, removing parts of it by inpainting them with the background. To counter this, we compose the local view with randomly sample solid background colors, rather than the scene. This prevents the inpainting model from removing parts of the object as the object must look realistic when rendered in front of a random background. In practice, we utilize a set of global views created by cropping the original global view at various scales, ensuring that the crops encompass the entire local view." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.342, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Scene-Conditional 3D Object Stylization and Composition" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.156, + 0.779, + 0.59 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.44, + 0.59, + 0.562, + 0.598 + ], + "angle": 0, + "content": "(a) Diverse Objects and Scenes" + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.605, + 0.779, + 0.68 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.283, + 0.681, + 0.35, + 0.689 + ], + "angle": 0, + "content": "(b) Small Objects" + }, + { + "type": "image_caption", + "bbox": [ + 0.472, + 0.681, + 0.53, + 0.689 + ], + "angle": 0, + "content": "(c) Big Objects" + }, + { + "type": "image_caption", + "bbox": [ + 0.637, + 0.681, + 0.726, + 0.689 + ], + "angle": 0, + "content": "(d) Different Locations" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.718, + 0.788, + 0.83 + ], + "angle": 0, + "content": "Fig. 5: Visual Results. (a) We showcase that our method applies to a diverse range of objects and scenes. The global view (top row) the overall composition quality and object-centric local view (bottom two rows) for the fidelity of stylized textures are demonstrated. For dim scenes, we additionally render objects without the estimated lighting condition \\((w / o\\text{light})\\) for illustrative purposes. Additionally, we showcase that our method applies to both (b) small and (c) big objects, as well as (d) different placing locations. The texture of the television, for example, adjusts the texture to match its surroundings." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "J. Zhou et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.39, + 0.162 + ], + "angle": 0, + "content": "3.4 Lighting Model" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.172, + 0.788, + 0.369 + ], + "angle": 0, + "content": "To seamlessly integrate the object into the scene, we estimate the lighting conditions, as outlined in Fig. 4. To light the object, we utilize a high dynamic range (HDR) environment map \\(\\mathcal{E} \\in \\mathbb{R}^{H_e \\times W_e \\times 3}\\), which is well-suited for representing natural illumination. Since the given 2D scene image only captures a small angle of the full 360-degree environment map, we first use the image as input to estimate a low dynamic range (LDR) environment map. This process depends on the scene type (i.e., indoor vs. outdoor) and is explained in the following paragraphs. Given the LDR map, we then convert it to an HDR map by estimating a light scale (i.e., a scalar) for each bright region, which is thresholded from the LDR map directly (e.g., where intensity \\(\\mathcal{I}_{i,j} \\geq 0.8\\)). This scalar is then multiplied by the RGB values to approximate an HDR map. Note that we use the HDR map only to light the object. It thus does not need to be of high accuracy to make the object look plausible." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.384, + 0.788, + 0.553 + ], + "angle": 0, + "content": "Indoor Scenes. The well-bounded 2D visual geometry of indoor scenes is primarily characterized by enclosed spaces and interconnected planes, facilitating straightforward (rough) 3D reconstruction. To simulate the physical lighting of indoor scenes, we lift each pixel from the 2D scene to 3D space using its estimated depth \\(\\mathcal{D}\\) [13, 44]. We perform world-to-camera and spherical-to-equirectangular coordinate transformations in succession, unwrapping 3D points into a latitude-longitude formatted LDR map. In post-processing, we automatically fill holes and remove small isolated regions. We fill the unseen region (i.e. areas behind the camera) with the average RGB value of the 2D scene. Two scalars are optimized: one for the far light region (\\(\\triangleq \\{i,j \\mid \\mathcal{D}_{i,j} \\geq \\tau_d \\wedge \\mathcal{I}_{i,j} \\geq \\tau_f\\}\\)) and another for the near light region (\\(\\triangleq \\{i,j \\mid \\mathcal{D}_{i,j} < \\tau_d \\wedge \\mathcal{I}_{i,j} \\geq \\tau_n\\}\\))." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.566, + 0.788, + 0.732 + ], + "angle": 0, + "content": "Outdoor Scenes. 3D reconstruction of outdoor scenes is challenging due to their inherently unbounded geometric structure. Furthermore, in most outdoor scenes, the sun is the dominant light source, and the position change of an object has a negligible effect on the angle shift of the light. Therefore, instead of painstakingly recovering location-specific viewpoints, we use a single-view to panorama outpainting model [2], taking its generated output as the initial LDR map. We optimize one scalar for the far light area with positive elevation \\((\\bigtriangleup \\{i,j\\mid \\frac{i}{H_c} < 0.5\\vee \\mathcal{I}_{i,j}\\geq \\tau_o\\})\\), leaving the near light area \\((\\bigtriangleup \\varnothing)\\) unoptimized. While more sophisticated designs, such as a position estimator [27] or sky modeling [22], could be incorporated, we find our approach robust enough to provide a sufficient estimation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.755, + 0.489, + 0.77 + ], + "angle": 0, + "content": "3.5 Light-Capturing Apparatus" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.842 + ], + "angle": 0, + "content": "To estimate light scales, we place 3D objects with a fixed appearance model into the 2D scene and then optimize the lighting with VSD as described in Eq. (1) to anchor the lighting impact imposed by the scene. Inspired by the traditional inverse rendering setup [11] where the environment map can almost be perfectly" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.342, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Scene-Conditional 3D Object Stylization and Composition" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.146, + 0.429, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.431, + 0.152, + 0.501, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.147, + 0.774, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.272, + 0.788, + 0.356 + ], + "angle": 0, + "content": "Fig. 6: Qualitative Comparison. We compare our the generated global view with the scene-object compositing method: CDC [18] and the local view with mesh texturing methods: Prolific Dreamer [55], Fantasia3D [7], and TEXTure [45]. s.gd. denotes similarity guidance. g.gd. and r.f.i denotes global-view guidance and reference feature injection, respectively. Our method achieves both seamless scene-object composition as well as texture adaptation of high fidelity. See sup. mat. for extensive comparison." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.366, + 0.789, + 0.518 + ], + "angle": 0, + "content": "reconstructed from objects' reflections, we introduce a novel concept incorporating a virtual light-capturing apparatus alongside the object of interest during the optimization process. We insert a white sphere made of a smooth diffuse material into the scene and guide the lighting optimization with the diffusion model using the text prompt \"A gigantic diffuse white (spray-painted) sphere (ball)\". The white diffuse sphere proves advantageous in stabilizing lighting estimations for scenes with potentially strong light sources, as the reflected intensity from a white diffuse object closely approximates the intensity of the environmental lighting. In our initial experiments, we also tested spheres with mirrored and matte silver materials but found them less beneficial." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.538, + 0.492, + 0.554 + ], + "angle": 0, + "content": "3.6 Light-Conditional Guidance" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.562, + 0.789, + 0.744 + ], + "angle": 0, + "content": "The absence of light conditions in text prompts makes the generated images only loosely constrained on lighting, which makes it challenging for diffusion models to score the renderings accurately. A white sphere looks gray in a darker environment, so light-agnostic prompts can engender an overestimation of the brightness for dim scenes. Therefore, we append \"in a dark environment\" if the average intensity of the background and light areas are below certain thresholds. We condition the LoRA-tuned model \\(\\epsilon_{\\psi}\\) by concatenating the light scales with the camera extrinsic [55] as class embeddings for the U-Net. This improves light estimation, especially for outdoor scenes and dimly lit scenes. For scenes with atypical lighting that significantly alters the appearance of objects, we manually append color prompts such as \"in blue tint\" for the seabed bathed in blue light, or \"in red illumination\" for a nightclub enveloped in red atmospheric light." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.765, + 0.377, + 0.782 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "In this section, we show extensive qualitative and quantitative results of our method in various settings. Please see the appendix for many more examples, animations and discussion on limitations and negative impact." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.127 + ], + "angle": 0, + "content": "J. Zhou et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.502, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.159, + 0.785, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.234, + 0.788, + 0.278 + ], + "angle": 0, + "content": "Fig. 7: Ablation on Texture Adaptation. We ablate global-view guidance \\((g.gd.)\\) and reference feature injection \\((r.f.i)\\). -ctrl. and \\(+\\mathrm{ctrl}\\) denotes less or more control in feature injection (smaller or higher \\(s_{\\mathrm{f}}\\) & \\(p\\)) for Identity Preservation requirement." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.292, + 0.449, + 0.308 + ], + "angle": 0, + "content": "4.1 Qualitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.326, + 0.784, + 0.373 + ], + "angle": 0, + "content": "We first present visual results and a case study to demonstrate the generalizability of the proposed method. We then compare our method with others that specifically focus on either scene composition or object stylization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.396, + 0.788, + 0.533 + ], + "angle": 0, + "content": "Main Results. To demonstrate the generalizability of the framework and the impact of different scenes on the inserted objects, we present visual results in Fig. 5. Our method successfully blends the objects into various environments, achieving photorealistic adaptation for both appearance and lighting. This includes scenarios drawn from both real-world and fantasy settings. For example, the dimly-lit umbrella in Fig. 5 casts a shadow that is perfectly aligned with its handle and the light direction of the scene, substantially increasing the photorealism of the composition; The Belweder TV set is wrapped with gray mud consistent with the scene while its original structural details get maintained." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.555, + 0.788, + 0.712 + ], + "angle": 0, + "content": "Comparison on Scene Composition. We showcase the result from CDC [18] using LLM-prompted text prompts for a fair comparison. As shown in Fig. 6, CDC loses the structural details of the object while providing limited adaptation for texture transfer. Moreover, the composed scene appears artificial and synthetic, giving the impression that the sofa is floating above the mud. Our experiments reveal that the CDC often leads to object removal, particularly when the re-painting [31] is employed. Therefore, we further incorporate similarity guidance similar to the classifier guidance [12] by adding \\(\\sqrt{1 - \\alpha_t} \\cdot \\frac{\\partial (||\\pmb{x}_0^r - \\hat{\\pmb{x}}_\\phi ||^2)}{\\partial \\pmb{x}_t}\\) to the predicted gradient \\(\\epsilon_{\\phi}(\\pmb{x}_t)\\) each denoising step referred as CDC w/s.gd., but find it provides limited improvement." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Comparison on Object Stylization. We showcase results from mesh texturing method [7, 45, 55] using our LLM-prompted text prompts for a fair comparison. As shown in Fig. 6, although achieving realistic texture adaptation, these mesh texturing methods fall short in preserving the original structural details of the object and in utilizing scene-level information. Essentially, These methods only satisfy the Environmental Influence requirement via text prompts and do not meet all the criteria of our intended goal." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.342, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Scene-Conditional 3D Object Stylization and Composition" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.188 + ], + "angle": 0, + "content": "Table 2: Quantitative Evaluation. (a) We measure LPIPS for controllability, \\(\\mathrm{CLIP}_{\\mathrm{dir}}\\) and \\(\\mathrm{CLIP}_{\\mathrm{SI}}\\) for composability. (b) We present the percentages of users who prefer our method over other methods: CDC [18] and InstructP2P [3]." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.199, + 0.516, + 0.267 + ], + "angle": 0, + "content": "
MethodLPIPS ↓CLIPdir ↑CLIPSI ↑
InstructP2P [3]0.00200.01750.1852
Ours (T.A.)0.00140.02010.2006
CDC [18]0.00530.00420.1996
Ours (T.A. + L.E.)0.00150.02060.2013
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.307, + 0.268, + 0.42, + 0.279 + ], + "angle": 0, + "content": "(a) Visual Metrics." + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.209, + 0.785, + 0.256 + ], + "angle": 0, + "content": "
OursDrawOther
Ours vs. CDC [18]71%20%9%
Ours vs. InstructP2P [3]74%6%20%
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.541, + 0.257, + 0.758, + 0.27 + ], + "angle": 0, + "content": "(b) User study for human preference." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.286, + 0.46, + 0.301 + ], + "angle": 0, + "content": "4.2 Quantitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.31, + 0.784, + 0.416 + ], + "angle": 0, + "content": "We collected 34 composites between diverse objects and scenes, each with 3 random views, yielding a total of 102 images. Besides CDC [18], we compare with InstructPix2Pix [3] by directly copy-pasting the object with adapted texture onto the scene. We measure LPIPS for controllability, \\(\\mathrm{CLIP}_{\\mathrm{dir}}\\) and \\(\\mathrm{CLIP}_{\\mathrm{SI}}\\) for composability, similar to CDC, and report the results in Tab. 2. Compared with InstructPix2Pix our method for texture adaptation (T.A) alone excels in controllability and composability. Light estimation (L.E.) further improves results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.417, + 0.785, + 0.462 + ], + "angle": 0, + "content": "We also conducted a user study in Tab. 2 with 42 participants to compare the realism of our compositing under the condition that the object identity must be preserved. Our method is preferred over prior work in over \\(70\\%\\) of the cases." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.484, + 0.39, + 0.5 + ], + "angle": 0, + "content": "4.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.509, + 0.788, + 0.63 + ], + "angle": 0, + "content": "Identity Preservation. To illustrate how the Identity Preservation requirement is satisfied by reference feature injection (r.f.i.), we showcase the leather sofa in the swamp with less and more control (Fig. 7). The setup with less control drifts to a visually different sofa with too much mud wrapped around while more control generates a less muddy and adapted sofa compared to the reference image. Both adapted textures are less visually plausible. Additionally, when comparing Ours w/o g.gd. & r.f.i. to the reference image, it is evident that the identity of the sofa is lost." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.645, + 0.788, + 0.738 + ], + "angle": 0, + "content": "Global-View Guidance. To showcase the necessity of global-view guidance we conduct texture adaptation without it as Ours w/o g.gd. in Fig. 7. The adapted texture without seeing the full scene fails to adapt the color of the mud to that of the scene image, contrasting starkly with the scene. Experiments ablating the optimal global-view guidance set-up (e.g., solid-color background augmentation, usage of inpainting model, etc.) are detailed in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Light Estimation. In Fig. 1, the combination of accurately estimated lighting and shadows significantly enhances the photorealism of the composed image (compare the image on the left, which includes lighting, to those in the top right corner, which are without lighting). To further test the validity of the estimated light, we render images omitting the texture adaptation step. Fig. 8 demonstrates that the estimated light considers multiple factors, including its" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.127 + ], + "angle": 0, + "content": "J. Zhou et al." + }, + { + "type": "image", + "bbox": [ + 0.273, + 0.145, + 0.726, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.506, + 0.788, + 0.562 + ], + "angle": 0, + "content": "Fig. 8: Ablation on Light Estimation. We evaluate the effectiveness of light estimation by demonstrating the lighting effects on a sofa in various indoor and outdoor scenes. Overlaid in the left corners are objects rendered with the estimated light against a white background. In the right corners are objects before the light estimation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.579, + 0.788, + 0.641 + ], + "angle": 0, + "content": "direction relative to the specific locations of objects in indoor scenes and its impact on the appearance of objects under atmospheric colored lighting conditions. Importantly, the estimated HDR maps, paired with the proposed scene composition, facilitate the realistic integration of shadows." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.676, + 0.36, + 0.691 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this paper, we are motivated by a practical goal—what if a 3D object is placed into a 2D scene?—and propose a novel framework that allows (1) stylizing the object with an adapted texture that aligns with the given scene; and (2) achieving a photorealistic scene composition with the aid of estimated light of the environment. This is enabled by physically-based differentiable ray tracing and diffusion models used as guidance. Our method allows for automatic adaptation and blending of existing objects into a variety of scenes, making it useful for both 2D and 3D downstream applications, such as visual media and video games." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.342, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Scene-Conditional 3D Object Stylization and Composition" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.788, + 0.207 + ], + "angle": 0, + "content": "Acknowledgement. This work is supported by the UKRI grant: Turing AI Fellowship EP/W002981/1 and EPSRC/MURI grant: EP/N019474/1. J. Zhou is also supported by the Horizon Robotics. We thank Luke Melas-Kyriazi and Fabio Pizzati for their helpful discussions." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.23, + 0.323, + 0.246 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.262, + 0.571, + 0.275 + ], + "angle": 0, + "content": "1. https://polyhaven.com, https://polyhaven.com" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.276, + 0.786, + 0.302 + ], + "angle": 0, + "content": "2. Akimoto, N., Matsuo, Y., Aoki, Y.: Diverse plausible 360-degree image outpainting for efficient 3dgc background creation. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.304, + 0.786, + 0.33 + ], + "angle": 0, + "content": "3. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.331, + 0.786, + 0.357 + ], + "angle": 0, + "content": "4. Cao, T., Kreis, K., Fidler, S., Sharp, N., Yin, K.: Texfusion: Synthesizing 3d textures with text-guided image diffusion models. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.359, + 0.786, + 0.385 + ], + "angle": 0, + "content": "5. Chen, B.C., Kae, A.: Toward realistic image compositing with adversarial learning. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.386, + 0.786, + 0.426 + ], + "angle": 0, + "content": "6. Chen, D.Z., Siddiqui, Y., Lee, H.Y., Tulyakov, S., Nießner, M.: Text2tex: Text-driven texture synthesis via diffusion models. arXiv preprint arXiv:2303.11396 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.428, + 0.786, + 0.454 + ], + "angle": 0, + "content": "7. Chen, R., Chen, Y., Jiao, N., Jia, K.: Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.456, + 0.786, + 0.481 + ], + "angle": 0, + "content": "8. Cong, W., Niu, L., Zhang, J., Liang, J., Zhang, L.: Bargainnet: Background-guided domain translation for image harmonization. In: ICME (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.483, + 0.786, + 0.509 + ], + "angle": 0, + "content": "9. Cong, W., Zhang, J., Niu, L., Liu, L., Ling, Z., Li, W., Zhang, L.: Dovenet: Deep image harmonization via domain verification. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.511, + 0.786, + 0.55 + ], + "angle": 0, + "content": "0. Deitke, M., Schwenk, D., Salvador, J., Weihs, L., Michel, O., VanderBilt, E., Schmidt, L., Ehsani, K., Kembhavi, A., Farhadi, A.: Objaverse: A universe of annotated 3d objects. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.552, + 0.786, + 0.578 + ], + "angle": 0, + "content": "1. Deng, Y., Li, X., Liu, S., Yang, M.H.: Dip: Differentiable interreflection-aware physics-based inverse rendering. arXiv preprint arXiv:2212.04705 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.58, + 0.786, + 0.605 + ], + "angle": 0, + "content": "2. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. NeurIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.607, + 0.786, + 0.634 + ], + "angle": 0, + "content": "3. Eftekhar, A., Sax, A., Malik, J., Zamir, A.: Omnidata: A scalable pipeline for making multi-task mid-level vision datasets from 3d scans. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.635, + 0.786, + 0.674 + ], + "angle": 0, + "content": "4. Gao, J., Shen, T., Wang, Z., Chen, W., Yin, K., Li, D., Litany, O., Gojcic, Z., Fidler, S.: Get3d: A generative model of high quality 3d textured shapes learned from images. NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.676, + 0.786, + 0.702 + ], + "angle": 0, + "content": "5. Gardner, M.A., Hold-Geoffroy, Y., Sunkavalli, K., Gagné, C., Lalonde, J.F.: Deep parametric indoor lighting estimation. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.704, + 0.786, + 0.743 + ], + "angle": 0, + "content": "6. Gardner, M.A., Sunkavalli, K., Yumer, E., Shen, X., Gambaretto, E., Gagné, C., Lalonde, J.F.: Learning to predict indoor illumination from a single image. arXiv preprint arXiv:1704.00090 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.745, + 0.786, + 0.771 + ], + "angle": 0, + "content": "7. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. NeurIPS (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.773, + 0.786, + 0.812 + ], + "angle": 0, + "content": "8. Hachnochi, R., Zhao, M., Orzech, N., Gal, R., Mahdavi-Amiri, A., Cohen-Or, D., Bermano, A.H.: Cross-domain compositing with pretrained diffusion models. arXiv preprint arXiv:2302.10167 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.814, + 0.786, + 0.84 + ], + "angle": 0, + "content": "9. Haque, A., Tancik, M., Efros, A., Holynski, A., Kanazawa, A.: Instruct-nerf2nerf: Editing 3d scenes with instructions. In: CVPR (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.262, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "J. Zhou et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.176 + ], + "angle": 0, + "content": "20. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.729, + 0.19 + ], + "angle": 0, + "content": "21. Ho, J., Salimans, T.: Classifier-free diffusion guidance. NeurIPSW (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.192, + 0.785, + 0.221 + ], + "angle": 0, + "content": "22. Hold-Geoffroy, Y., Athawale, A., Lalonde, J.F.: Deep sky modeling for single image outdoor lighting estimation. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.222, + 0.785, + 0.249 + ], + "angle": 0, + "content": "23. Hold-Geoffroy, Y., Sunkavalli, K., Hadap, S., Gambaretto, E., Lalonde, J.F.: Deep outdoor illumination estimation. In: CVPR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.25, + 0.785, + 0.278 + ], + "angle": 0, + "content": "24. Jakob, W., Speierer, S., Roussel, N., Vicini, D.: Dr.jit: A just-in-time compiler for differentiable rendering. TOG (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.28, + 0.785, + 0.321 + ], + "angle": 0, + "content": "25. Jiang, Y., Zhang, H., Zhang, J., Wang, Y., Lin, Z., Sunkavalli, K., Chen, S., Amirghodsi, S., Kong, S., Wang, Z.: Ssh: A self-supervised framework for image harmonization. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.322, + 0.785, + 0.364 + ], + "angle": 0, + "content": "26. Kamata, H., Sakuma, Y., Hayakawa, A., Ishii, M., Narihira, T.: Instruct 3d-to-3d: Text instruction guided 3d-to-3d conversion. arXiv preprint arXiv:2303.15780 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.365, + 0.785, + 0.393 + ], + "angle": 0, + "content": "27. Karimi Dastjerdi, M.R., Eisenmann, J., Hold-Geoffroy, Y., Lalonde, J.F.: Everlight: Indoor-outdoor editable hdr lighting estimation. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.395, + 0.785, + 0.436 + ], + "angle": 0, + "content": "28. Karimi Dastjerdi, M.R., Hold-Geoffroy, Y., Eisenmann, J., Khodadadeh, S., Lalonde, J.F.: Guided co-modulated GAN for 360 field of view extrapolation. 3DV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.437, + 0.785, + 0.479 + ], + "angle": 0, + "content": "29. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3d: High-resolution text-to-3d content creation. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.48, + 0.785, + 0.522 + ], + "angle": 0, + "content": "30. Liu, M., Xu, C., Jin, H., Chen, L., Xu, Z., Su, H., et al.: One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. arXiv preprint arXiv:2306.16928 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.785, + 0.551 + ], + "angle": 0, + "content": "31. Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Van Gool, L.: Repaint: Inpainting using denoising diffusion probabilistic models. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.552, + 0.785, + 0.593 + ], + "angle": 0, + "content": "32. McAuley, S., Hill, S., Hoffman, N., Gotanda, Y., Smits, B., Burley, B., Martinez, A.: Practical physically-based shading in film and game production. In: SIGGRAPH (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.595, + 0.785, + 0.623 + ], + "angle": 0, + "content": "33. Metzer, G., Richardson, E., Patashnik, O., Giryes, R., Cohen-Or, D.: Latent-nerf for shape-guided generation of 3d shapes and textures. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.624, + 0.785, + 0.652 + ], + "angle": 0, + "content": "34. Michel, O., Bar-On, R., Liu, R., Benaim, S., Hanocka, R.: Text2mesh: Text-driven neural stylization for meshes (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.653, + 0.785, + 0.695 + ], + "angle": 0, + "content": "35. Mohammad Khalid, N., Xie, T., Belilovsky, E., Popa, T.: Clip-mesh: Generating textured meshes from text using pretrained image-text models. In: SIGGRAPH Asia (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.696, + 0.785, + 0.738 + ], + "angle": 0, + "content": "36. Munkberg, J., Hasselgren, J., Shen, T., Gao, J., Chen, W., Evans, A., Müller, T., Fidler, S.: Extracting triangular 3d models, materials, and lighting from images. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.739, + 0.785, + 0.767 + ], + "angle": 0, + "content": "37. Oechsle, M., Mescheder, L., Niemeyer, M., Strauss, T., Geiger, A.: Texture fields: Learning texture representations in function space. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.768, + 0.503, + 0.782 + ], + "angle": 0, + "content": "38. OpenAI: Gpt-4 technical report (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.784, + 0.785, + 0.811 + ], + "angle": 0, + "content": "39. Parmar, G., Kumar Singh, K., Zhang, R., Li, Y., Lu, J., Zhu, J.Y.: Zero-shot image-to-image translation. In: SIGGRAPH (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.812, + 0.785, + 0.84 + ], + "angle": 0, + "content": "40. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. ICLR (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.342, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Scene-Conditional 3D Object Stylization and Composition" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.203 + ], + "angle": 0, + "content": "41. Qian, G., Mai, J., Hamdi, A., Ren, J., Siarohin, A., Li, B., Lee, H.Y., Skorokhodov, I., Wonka, P., Tulyakov, S., Ghanem, B.: Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. arXiv preprint arXiv:2306.17843 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.204, + 0.787, + 0.245 + ], + "angle": 0, + "content": "42. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.246, + 0.786, + 0.273 + ], + "angle": 0, + "content": "43. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: ICML (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.273, + 0.787, + 0.313 + ], + "angle": 0, + "content": "44. Ranftl, R., Lasinger, K., Hafner, D., Schindler, K., Koltun, V.: Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. PAMI (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.315, + 0.787, + 0.342 + ], + "angle": 0, + "content": "45. Richardson, E., Metzer, G., Alaluf, Y., Giryes, R., Cohen-Or, D.: Texture: Text-guided texturing of 3d shapes. SIGGRAPH (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.343, + 0.787, + 0.37 + ], + "angle": 0, + "content": "46. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.371, + 0.787, + 0.411 + ], + "angle": 0, + "content": "47. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.412, + 0.787, + 0.438 + ], + "angle": 0, + "content": "48. Siddiqui, Y., Thies, J., Ma, F., Shan, Q., Nießner, M., Dai, A.: Texturify: Generating textures on 3d shape surfaces. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.44, + 0.787, + 0.48 + ], + "angle": 0, + "content": "49. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. In: ICLR (2021), https://openreview.net/forum?id=PxTIG12RRHS" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.481, + 0.787, + 0.508 + ], + "angle": 0, + "content": "50. Song, Y., Zhang, Z., Lin, Z., Cohen, S., Price, B., Zhang, J., Kim, S.Y., Aliaga, D.: Objectstitch: Generative object compositing. CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.509, + 0.787, + 0.536 + ], + "angle": 0, + "content": "51. Tsalicoglou, C., Manhardt, F., Tonioni, A., Niemeyer, M., Tombari, F.: Textmesh: Generation of realistic 3d meshes from text prompts (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.537, + 0.787, + 0.563 + ], + "angle": 0, + "content": "52. Tumanyan, N., Geyer, M., Bagon, S., Dekel, T.: Plug-and-play diffusion features for text-driven image-to-image translation. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.565, + 0.787, + 0.591 + ], + "angle": 0, + "content": "53. Walter, B., Marschner, S.R., Li, H., Torrance, K.E.: Microfacet models for refraction through rough surfaces. In: EGSR (2007)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.592, + 0.787, + 0.619 + ], + "angle": 0, + "content": "54. Wang, H., Du, X., Li, J., Yeh, R.A., Shakhnarovich, G.: Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.62, + 0.787, + 0.66 + ], + "angle": 0, + "content": "55. Wang, Z., Lu, C., Wang, Y., Bao, F., Li, C., Su, H., Zhu, J.: Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.661, + 0.787, + 0.688 + ], + "angle": 0, + "content": "56. Wei, L.Y., Lefebvre, S., Kwatra, V., Turk, G.: State of the art in example-based texture synthesis. EG-STAR (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.689, + 0.787, + 0.716 + ], + "angle": 0, + "content": "57. Xiao, J., Ehinger, K.A., Oliva, A., Torralba, A.: Recognizing scene viewpoint using panoramic place representation. In: ICCV (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.717, + 0.787, + 0.743 + ], + "angle": 0, + "content": "58. Zhang, J., Sunkavalli, K., Hold-Geoffroy, Y., Hadap, S., Eisenman, J., Lalonde, J.F.: All-weather deep outdoor lighting estimation. In: CVPR (2019)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.743 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Scene-Conditional 3D Object Stylization and Composition/ed573969-4bd1-4a9b-b748-12ae2bb2b2d3_origin.pdf b/2024/Scene-Conditional 3D Object Stylization and Composition/ed573969-4bd1-4a9b-b748-12ae2bb2b2d3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ff8f30a7909ef9ae7eba5eb71875fe1b07ceb609 --- /dev/null +++ b/2024/Scene-Conditional 3D Object Stylization and Composition/ed573969-4bd1-4a9b-b748-12ae2bb2b2d3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7be7c4e92b61659d1114c98d86c5638cbd5335bff6aa14ec79cdc6f7c47b3c56 +size 15179149 diff --git a/2024/Scene-Conditional 3D Object Stylization and Composition/full.md b/2024/Scene-Conditional 3D Object Stylization and Composition/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f12cb652b46810688d1c19168f4d09bc38dd6653 --- /dev/null +++ b/2024/Scene-Conditional 3D Object Stylization and Composition/full.md @@ -0,0 +1,270 @@ +# Scene-Conditional 3D Object Stylization and Composition + +Jinghao Zhou $^{1}$ , Tomas Jakab $^{1}$ , Philip Torr $^{1}$ , and Christian Rupprecht $^{2}$ + +University of Oxford {jinghao,tomj,chrisr}@robots.ox.ac.uk + +![](images/a322d6fb3ae17492b50af53c6a069cd9aa7f3ee71da505f94b738df7691db465.jpg) +Fig. 1: We present a framework that adapts a 3D object's appearance to a location in a 2D scene. It creates an image where the 3D object is seamlessly blended (left & bottom right), with its appearance influenced by the scene's environmental conditions and lighting effects. Moreover, the stylized object with adapted textures (top right), rendered here without the estimated lighting condition for illustrative purposes, can be further utilized as 3D assets for downstream tasks such as video games. + +Abstract. Recently, 3D generative models have made impressive progress, enabling the generation of almost arbitrary 3D assets from text or image inputs. However, these approaches generate objects in isolation without any consideration for the scene where they will eventually be placed. In this paper, we propose a framework that allows for the stylization of an existing 3D asset to fit into a given 2D scene, and additionally produce a photorealistic composition as if the asset was placed within the environment. This not only opens up a new level of control for object stylization, for example, the same assets can be stylized to reflect changes in the environment, such as summer to winter or fantasy versus futuristic settings—but also makes the object-scene composition more controllable. We achieve this by combining modeling and optimizing the object's texture and environmental lighting through differentiable ray + +tracing with image priors from pre-trained text-to-image diffusion models. We demonstrate that our method applies to a wide variety of indoor and outdoor scenes and arbitrary objects. See also our project page. + +# 1 Introduction + +After the success of generative models [20,43,46] for images in computer vision, the community is now interested in lifting these models to 3D [40,54] through advances in architectures, data, and training. These lifted 3D models demonstrate plausible 3D representations of objects conditioned on images or texts. + +While generating whole objects—e.g. an astronaut riding a horse—is a challenging problem that generative models are becoming able to solve, in practice, these assets are difficult to adapt for use in downstream applications or existing virtual environments such as video games as they are generated without context. + +This paper aims to provide a mechanism to close the appearance gap between 3D objects and environments. We formulate this problem through the lens of a creative tool. The intended goal of this paper is to align the appearance of a 3D object to a customized 2D scene as we place it into a specific location, a challenging task that requires fine-grained visual control for convincing results. + +We are interested in producing a photorealistic scene-object composition where the 3D object is seamlessly blended. For instance, an object placed in a rainy night scene should appear wet and dimly lit, and if the scene is muddy and sunny, the object should also look muddy and cast a strong shadow. To achieve a high level of realism on a fine-grained level, the mud on the object must match the scene's mud in appearance and the direction of the shadow must match where the sun is located. Instead of simply resorting to 2D blending operations [18, 50], we explicitly model an object's appearance and lighting in 3D space, which enables more accurate and regularized control of the object's appearance in the composed image. + +We also aim to preserve the underlying object geometry and appearance factors unrelated to the scene environment. This task is highly practical; for example, for media design, when presenting a 3D asset of a product within a realistic scene, we want the product to blend naturally while remaining recognizable. Manually achieving this would require a skilled artist to modify the object's texture to match the scene, a challenging and time-consuming process. Our paper introduces a method that automates this task. It accepts a 3D object with texture, a 2D scene, and the object's position within the scene and outputs an adapted texture consistent with the scene. + +To accomplish this, we leverage recent advances in generative diffusion models [46, 47] for images, which are trained on a large-scale dataset of internet images. We conduct an optimization procedure where we compose the 2D renderings of the 3D object with the 2D scene at a specified location and use a pre-trained diffusion model to "critique" the realism of the composed image, providing a gradient signal for optimization. + +The challenge is that diffusion models, such as Stable Diffusion [46], are trained on 2D images and lack any notion of 3D geometry. Naively optimizing + +or denoising in 2D space [18,50] would result in losing the object's 3D geometry. To prevent this, we optimize the object's content in 3D by randomly rotating the object during optimization and rendering it from various viewpoints, which helps to align the adapted texture with the object's geometry and avoid overfitting to a single viewpoint. + +Notably, randomly rotating the 3D object stops lighting effects from being baked into the texture, which is desirable. Yet, it also prevents the object from naturally blending with the scene in a single viewpoint. To counteract this, we predict the scene's lighting and the shadows cast by the object as separate components. For this purpose, we propose a novel technique inspired by a common physical approach from the computer graphics industry, where reflective and diffuse spheres with known material parameters are placed and photographed in a real-world scene to capture the environmental lighting conditions. We adopt this concept and place a virtual white diffuse sphere inside the scene during optimization while conditioning the generative model to expect a white sphere within the scene, thereby capturing the model's interpretation of the scene's lighting in an easily extracted form. + +Optimizing the object's appearance in 3D offers further significant advantages: The user can change the object's pose without re-running the optimization, as the texture is meaningful from all viewpoints, and view-dependent effects are modeled separately by the lighting component. Additionally, the 3D object with the adapted texture can be used as a standalone asset in other downstream tasks (e.g. in a video game). + +In summary, we present a framework that (1) enables the stylization of a 3D object, which can also be used as a standalone asset, given a 2D scene and its location within it; (2) achieves photorealistic scene-object compositing with the help of a novel mechanism to estimate scene lighting from a single image. + +We further verify the effectiveness of our method in extensive experiments, demonstrating that our approach can realistically adapt objects to a diverse set of environments, laying the groundwork for practical applications. + +# 2 Related Work + +The purpose of this paper is to adapt the appearance of a 3D object to align with a customized 2D scene. This solves a more complicated problem than prior works, which often only tackle individual components, including but not limited to 2D scene-object compositing, light estimation, and texturing, but never as a whole. We compare with these methods next and detail their differences in Tab. 1. + +Scene-Object Compositing. Scene-object compositing is a challenging task that requires a delicate balance between visual transfer and control. It has been extensively researched [5,8,9,25]. Recent methods condition the denoising process of diffusion models [20,49] on object images. OST [50] aims to seamlessly integrate a 2D object image into a 2D scene by modifying the keys and values in the attention blocks. CDC [18] considers composing the 2D renderings of a 3D object with a 2D scene by directly modifying denoised latent. + +Table 1: Comparison with related work. Rep. denotes representation being 2D (□) or 3D (）。Env. Inf. denotes Environmental Influence and I.D. Pre. denotes Identity Preservation. Parameters for the lighting and the appearance model are either scene-agnostic constant (），scene-conditioned predictable (），or learnable (）。 + +
MethodSceneObjectCompose
Rep.LightingShadowRep.AppearanceEnv. Inf.I.D. Pre.
OST [50]□-×□-××Blend
CDC [18]□Ambient×Diffuse×××Blend
IGAN [28]□LDR✓Diffuse×××Copy & Paste
EVL [27]□HDR✓Diffuse×××Copy & Paste
PrDr [55]-Ambient×Diffuse✓×-
Fan3D [7]-HDR×PBR✓×-
Inst3D [26]-Point×Diffuse×✓-
Ours□HDR✓PBR✓✓Copy & Paste
+ +However, these methods rely entirely on the 2D denoising process and do not explicitly model the object's geometry, which often results in a significant loss of structural details and visual identity. This severely hampers these methods' ability to adapt to the environmental and lighting conditions of the scene. + +Light Estimation from Single-View Image. Estimating light conditions from a single-view 2D image is of practical use in the real world and has been extensively studied for both indoor [15,16] and outdoor scenes [22,23,58]. The distinction between these two settings has been significantly narrowed recently [27, 28], thanks to the outpainting capabilities of diffusion models, which have contributed to a unified framework. With additional training on large high dynamic range (HDR) datasets such as SUN360 [57], these methods can approximate the location and intensity of the light source, enabling realistic shadows. + +However, these methods estimate only the lighting of the scene. They are typically indifferent to how objects should be stylized to match the scene, neglecting the potential environmental influences on the textures of objects. + +Mesh Texturing. Classic automated approaches [56] for mesh texturing are often limited and apply only simple texture patterns. In contrast, methods [14, 37, 48] based on Generative Adversarial Networks [17] (GANs) are more capable of synthesizing detailed textures. Recently, text-conditioned texture generation [34, 35] has been made possible through the optimization of CLIP-based objectives [42]. The latest works now employ powerful text-to-image diffusion models [20, 49], either through optimization [7, 29, 33, 51, 55] using Score Distillation Sampling [40] (SDS), or by wrapping the generated 2D textures onto the mesh [4, 6, 45]. Importantly, Variational Score Distillation [55] (VSD) as an improved optimization objective and the adoption of PBR model [7] have greatly boosted the quality of generated textures. Additionally, an increasing level of requirement for more visual control from users has been partially met [19, 26] through image-conditional diffusion models [3] to preserve objects' identities. + +![](images/ae638e978c607d549deeff73f662f294c56565eeeaba25ed3e50646025032972.jpg) +Fig. 2: Framework. We learn an environment map and a texture map separately from the 2D supervision. We initialize (init.) the environment map with an LDR map estimated from the 2D scene and learn light multiplying scales bright areas, yielding an HDR map. We employ the PBR material model for texture maps, encoded via MLP with positional encoding. The object is rendered through a differentiable ray tracer and further composed with the scene background, receiving gradients from Stable Diffusion (SD) in the latent space. + +However, these works focus on texturing the object in isolation from the scene and do not consider the homogeneity of the composition when these assets are placed in certain environments, which limits their practical usage. + +# 3 Method + +The overall structure of our framework can be found in Fig. 2. Given a textured 3D object, a 2D image of a scene, and a 2D location where to place the object, we first adapt the texture of the 3D object (Secs. 3.2 and 3.3) to align with the scene. To enable a photorealistic composition of the object and the scene, we further estimate the lighting (Secs. 3.4 to 3.6) and apply it during the final rendering. + +# 3.1 Optimization + +To enable end-to-end optimization, we adopt a differentiable renderer to render the object and score it using the priors from diffusion models. + +Differentiable Renderer. We employ the differential renderer Mitsuba3 [24] (denoted as $g$ ), due to its capabilities as a physically-based ray-tracing renderer equipped with versatile rendering options, such as diverse BSDFs, the path tracer integrator, and using an environment map as a light source. This enables us to produce photo-realistic renders complete with shadows and reflections, forming the foundation for robust texture adaptation and light estimation. + +Diffusion Model as Guidance. To score the rendered image, we make use of a generative diffusion model [46, 47] $\phi$ . Diffusion models have learned powerful priors of the image formation process in the real world from large-scale training + +on billion-scale web data. Similar to prior work [40,54], we supervise the rendered image $\pmb{x}_0 = g(\pmb{\theta})$ with a score-matching objective to obtain an image gradient + +$$ +\nabla_ {\boldsymbol {\theta}} \mathcal {L} (\boldsymbol {x} _ {0}) \triangleq \mathbb {E} _ {t, \epsilon} \left[ w _ {t} \left(\epsilon_ {\phi} (\boldsymbol {x} _ {t}) - \hat {\epsilon} (\boldsymbol {x} _ {t})\right) \frac {\partial \boldsymbol {x} _ {0}}{\partial \boldsymbol {\theta}} \right], \tag {1} +$$ + +where $\pmb{x}_t = \sqrt{\alpha_t}\pmb{x}_0 + \sqrt{1 - \alpha_t}\epsilon$ is the noisy rendered image, $\epsilon_{\phi}(\pmb{x}_t)$ is an approximated score for a noisy real image, $t\in \mathbb{N}$ indexes the discrete diffusion steps, and $\hat{\epsilon}$ is a scoring function for $\pmb{x}_t$ . Specifically, $\hat{\epsilon}$ is set as constant $\epsilon$ in Score Distillation Sampling [40] (SDS), and $\epsilon_{\psi}$ is an online LoRA-tuned network initialized from $\epsilon_{\phi}$ in Variational Score Sampling [55] (VSD). Notably, we employ their interpolated variant given by $\hat{\epsilon} = \lambda \epsilon_{\psi} + (1 - \lambda)\epsilon$ with a $\lambda$ annealed from 1 to $\lambda_e$ , helping to counteract the over-texturization often observed in VSD, especially for objects with simple surfaces. With residual of noises $\epsilon_{\phi} - \hat{\epsilon} = (\epsilon_{\phi} - \epsilon) - \lambda (\epsilon_{\psi} - \epsilon)$ , we highlight its relation with the image gradient obtained in traditional inverse rendering setup $\nabla_{\theta}\mathcal{L}(\pmb{x}_0) = \pmb{x}_0 - \pmb{x}_{\mathrm{gt}}$ by reformulating it as + +$$ +\epsilon_ {\phi} - \hat {\epsilon} = \sqrt {\frac {\alpha_ {t}}{1 - \alpha_ {t}}} [ (\boldsymbol {x} _ {0} - \hat {\boldsymbol {x}} _ {\phi}) - \lambda (\boldsymbol {x} _ {0} - \hat {\boldsymbol {x}} _ {\psi}) ], \tag {2} +$$ + +where the target used to supervise $\pmb{x}_0$ is predicted clean image $\hat{\pmb{x}}_{\phi}$ from $\epsilon_{\phi}$ as the positive and $\hat{\pmb{x}}_{\psi}$ from $\epsilon_{\psi}$ as the negative. The technique enables nominal classifier-free guidance [21] and thus bypasses mode-seeking behaviors and saturated texture. Inspired by [45], we incorporate the inner product of the camera pose and object's surface normal $\langle c, \pmb{n} \rangle$ into $w_t$ to penalize faces at small grazing angles. In practice, we use Stable Diffusion [46] (SD) and absorb the encoder into $g$ , optimizing in latent space. + +# 3.2 Appearance Model + +Textured Mesh. We represent the 3D object as a textured mesh, which is the most common format for 3D assets in the industry and is compatible with many established workflows. Textured meshes are available in abundance in online datasets and marketplaces [1,10], and recently, one can also generate reasonable quality 3D meshes with text descriptions [7,29,55] or limited viewpoints [30,41]. + +Neural Texture. We represent the appearance of objects using a texture map $\mathcal{T} \in \mathbb{R}^{H_t \times W_t \times 5}$ , facilitating compatibility of textured results with widely-used rendering engines, such as Blender. We parameterize the texture using a coordinate neural network that maps 3D coordinates to the texture attributes. We first map the 3D coordinates of vertices using a rasterizer [36] based on their UV coordinates into the space of the texture map, resulting in $\mathcal{C} \in \mathbb{R}^{H_t \times W_t \times 3}$ . We then use a 2-layer parameterized MLP without bias and Hash positional encoding $\varPhi: \mathcal{C} \to \mathcal{T}$ to map the coordinates to their texture attributes. For the final layer, we re-map the output activated after sigmoid according to pre-defined minimal and maximal values. The five channels in $\mathcal{T}$ correspond to the Physically-Based + +![](images/9d8e045379d92a98dbde3d91fb581c9b5641cbcd3210a98d1af91a471e966bf6.jpg) +Fig. 3: Pipeline for texture adaptation. We initialize the neural texture from the reference object and inject the feature of reference renderings to the U-Net of SD. We use both local-view and global-view guidance. + +Rendering (PBR) shading model [32], which includes a diffuse term $k_{\mathrm{d}} \in \mathbb{R}^{3}$ and an isotropic specular GGX lobe [53] described by $k_{\mathrm{rm}} = (k_r, k_m) \in \mathbb{R}^2$ , with $k_r$ representing roughness and $k_m$ indicating metalness. The PBR model, with $k_r$ affecting the scattering of light and $k_m$ influencing the reflective properties, creates realistic light interactions between surfaces. + +# 3.3 Scene-Conditional Guidance + +Achieving a realistic texture adaptation for 3D objects placed into 2D scenes requires three components: Environmental Influence involves adjusting the texture to reflect the environmental impact realistically; Identity Preservation aims to preserve the objects' unique visual aspects, and Blending focuses on guiding the texture adaptation to match the visual characteristic of the scene, ensuring seamless integration with the surroundings and avoiding stark contrasts. + +Environmental Influence. To show the realistic effects of the scene environment on the object's appearance, we guide the diffusion model with text prompts. We find that a naive text prompt, combining a simple object description with a scene description, e.g., a leather sofa in a swamp, is insufficient to generate realistic textures that reflect the environmental impact on the object, as shown in the appendix. For instance, we would expect a sofa placed in a swamp to be dirty and possibly show signs of moss or other vegetation growth. Instead of manually crafting a detailed text prompt for each object and scene combination, we automatize the process by employing a large language model (LLM), specifically GPT-4 [38], to generate a text prompt for the image diffusion model. Given the simple description of the object (e.g., a leather sofa) and the scene (e.g., a swamp), we instruct the LLM first to analyze the potential impact of the scene's environment on the object's appearance and then to distil this information into a text prompt that describes the resulting appearance of the object (e.g., The leather sofa, partially submerged in the swamp, looks discolored and soggy, its once-polished surface marred by mud, moss, and the murky water of the swamp + +![](images/384078b29483dc05b10914722d91a27b165a49f4eebdafa8d70157f7f1e74900.jpg) +Fig. 4: Pipeline for estimating the LDR map. We utilize tailored pipelines for indoor and outdoor scenes. Areas masked red and blue correspond to the far light $\mathcal{O}$ and near light region. + +environment.). We then use this automatically generated text prompt to guide the diffusion model in adapting the texture. The specific prompt design for this automated process is detailed in the appendix. + +Identity Preservation. To prevent the transferred texture from drifting away from its original content, we regularize the denoising diffusion model with reference images, a technique that has been explored for controlled image generation [39, 52]. Suppose we have a rendered reference image $\pmb{x}_0^r = g(\pmb{\theta}^r)$ and its noised version $\pmb{x}_t^r$ as input forwarded to $\epsilon_{\phi}$ and $\epsilon_{\psi}$ , for the Multi-Head Self-Attention (MHSA) layer in each Transformer block of $\epsilon_{\phi}$ 's and $\epsilon_{\psi}$ 's U-Nets, we concatenate the key $K^r$ and value $V^r$ of $\pmb{x}_t^r$ to the original $K$ and $V$ from $\pmb{x}_t$ yielding an image-conditional attention output + +$$ +A (\boldsymbol {x} _ {t}, \cdot , \boldsymbol {x} _ {t} ^ {r}) = \mathrm {M H S A} (Q, [ K; K ^ {r} ], [ V; V ^ {r} ]), \tag {3} +$$ + +where $\cdot$ can either be text or $\emptyset$ for text conditioning. The text-unconditional attention $A(\pmb{x}_t, \emptyset, \pmb{x}_t^r)$ used for classifier-free guidance [21] is further modified as $s_c \cdot A(\pmb{x}_t, \emptyset, \emptyset) + (1 - s_c) \cdot A(\pmb{x}_t, \emptyset, \pmb{x}_t^r)$ weighted by a control guidance scale $s_c$ . With the percentage $p$ of blocks to be injected, the intensity of preservation can be controlled: $s_c = p = 1$ provide the most and $s_c = p = 0$ the least preservation. + +Blending. The two components previously mentioned directly utilize object renderings, referred to as the local view, as inputs for vanilla SD. To further enhance the adaptation of the object's appearance to its surroundings, we input the scene image, with the object rendered onto it, referred to as the global view, into an additional inpainting SD [46]. We have found that the inpainting model provides superior conditioning on the surrounding environment depicted in the image, resulting in improved blending. However, the inpainting model can sometimes attempt to undesirably modify the inserted object, removing parts of it by inpainting them with the background. To counter this, we compose the local view with randomly sample solid background colors, rather than the scene. This prevents the inpainting model from removing parts of the object as the object must look realistic when rendered in front of a random background. In practice, we utilize a set of global views created by cropping the original global view at various scales, ensuring that the crops encompass the entire local view. + +![](images/865fcc92eb9818599c57327de653788632139a206693e2bfba30d0f6a4851bce.jpg) +(a) Diverse Objects and Scenes + +![](images/f2a43d92e204323547645b405ac3fd1cc9f21e20c0e32dac4ecd4054bd961f8a.jpg) +(b) Small Objects +(c) Big Objects +(d) Different Locations +Fig. 5: Visual Results. (a) We showcase that our method applies to a diverse range of objects and scenes. The global view (top row) the overall composition quality and object-centric local view (bottom two rows) for the fidelity of stylized textures are demonstrated. For dim scenes, we additionally render objects without the estimated lighting condition $(w / o\text{light})$ for illustrative purposes. Additionally, we showcase that our method applies to both (b) small and (c) big objects, as well as (d) different placing locations. The texture of the television, for example, adjusts the texture to match its surroundings. + +# 3.4 Lighting Model + +To seamlessly integrate the object into the scene, we estimate the lighting conditions, as outlined in Fig. 4. To light the object, we utilize a high dynamic range (HDR) environment map $\mathcal{E} \in \mathbb{R}^{H_e \times W_e \times 3}$ , which is well-suited for representing natural illumination. Since the given 2D scene image only captures a small angle of the full 360-degree environment map, we first use the image as input to estimate a low dynamic range (LDR) environment map. This process depends on the scene type (i.e., indoor vs. outdoor) and is explained in the following paragraphs. Given the LDR map, we then convert it to an HDR map by estimating a light scale (i.e., a scalar) for each bright region, which is thresholded from the LDR map directly (e.g., where intensity $\mathcal{I}_{i,j} \geq 0.8$ ). This scalar is then multiplied by the RGB values to approximate an HDR map. Note that we use the HDR map only to light the object. It thus does not need to be of high accuracy to make the object look plausible. + +Indoor Scenes. The well-bounded 2D visual geometry of indoor scenes is primarily characterized by enclosed spaces and interconnected planes, facilitating straightforward (rough) 3D reconstruction. To simulate the physical lighting of indoor scenes, we lift each pixel from the 2D scene to 3D space using its estimated depth $\mathcal{D}$ [13, 44]. We perform world-to-camera and spherical-to-equirectangular coordinate transformations in succession, unwrapping 3D points into a latitude-longitude formatted LDR map. In post-processing, we automatically fill holes and remove small isolated regions. We fill the unseen region (i.e. areas behind the camera) with the average RGB value of the 2D scene. Two scalars are optimized: one for the far light region ( $\triangleq \{i,j \mid \mathcal{D}_{i,j} \geq \tau_d \wedge \mathcal{I}_{i,j} \geq \tau_f\}$ ) and another for the near light region ( $\triangleq \{i,j \mid \mathcal{D}_{i,j} < \tau_d \wedge \mathcal{I}_{i,j} \geq \tau_n\}$ ). + +Outdoor Scenes. 3D reconstruction of outdoor scenes is challenging due to their inherently unbounded geometric structure. Furthermore, in most outdoor scenes, the sun is the dominant light source, and the position change of an object has a negligible effect on the angle shift of the light. Therefore, instead of painstakingly recovering location-specific viewpoints, we use a single-view to panorama outpainting model [2], taking its generated output as the initial LDR map. We optimize one scalar for the far light area with positive elevation $(\bigtriangleup \{i,j\mid \frac{i}{H_c} < 0.5\vee \mathcal{I}_{i,j}\geq \tau_o\})$ , leaving the near light area $(\bigtriangleup \varnothing)$ unoptimized. While more sophisticated designs, such as a position estimator [27] or sky modeling [22], could be incorporated, we find our approach robust enough to provide a sufficient estimation. + +# 3.5 Light-Capturing Apparatus + +To estimate light scales, we place 3D objects with a fixed appearance model into the 2D scene and then optimize the lighting with VSD as described in Eq. (1) to anchor the lighting impact imposed by the scene. Inspired by the traditional inverse rendering setup [11] where the environment map can almost be perfectly + +![](images/17521f961a6554fcbf31cb7c62e6723b50a858eeee7ecbfb620b9a99635b6f06.jpg) +Fig. 6: Qualitative Comparison. We compare our the generated global view with the scene-object compositing method: CDC [18] and the local view with mesh texturing methods: Prolific Dreamer [55], Fantasia3D [7], and TEXTure [45]. s.gd. denotes similarity guidance. g.gd. and r.f.i denotes global-view guidance and reference feature injection, respectively. Our method achieves both seamless scene-object composition as well as texture adaptation of high fidelity. See sup. mat. for extensive comparison. + +![](images/6b283a43f3b39b311ba05d1c1b0642bb08972f66ad04ebc31a69f7cb1890d016.jpg) + +![](images/6ef79b9c0fc85302e909a028449843bbc17886c21fcfc4dbe31a8fd51a9c9120.jpg) + +reconstructed from objects' reflections, we introduce a novel concept incorporating a virtual light-capturing apparatus alongside the object of interest during the optimization process. We insert a white sphere made of a smooth diffuse material into the scene and guide the lighting optimization with the diffusion model using the text prompt "A gigantic diffuse white (spray-painted) sphere (ball)". The white diffuse sphere proves advantageous in stabilizing lighting estimations for scenes with potentially strong light sources, as the reflected intensity from a white diffuse object closely approximates the intensity of the environmental lighting. In our initial experiments, we also tested spheres with mirrored and matte silver materials but found them less beneficial. + +# 3.6 Light-Conditional Guidance + +The absence of light conditions in text prompts makes the generated images only loosely constrained on lighting, which makes it challenging for diffusion models to score the renderings accurately. A white sphere looks gray in a darker environment, so light-agnostic prompts can engender an overestimation of the brightness for dim scenes. Therefore, we append "in a dark environment" if the average intensity of the background and light areas are below certain thresholds. We condition the LoRA-tuned model $\epsilon_{\psi}$ by concatenating the light scales with the camera extrinsic [55] as class embeddings for the U-Net. This improves light estimation, especially for outdoor scenes and dimly lit scenes. For scenes with atypical lighting that significantly alters the appearance of objects, we manually append color prompts such as "in blue tint" for the seabed bathed in blue light, or "in red illumination" for a nightclub enveloped in red atmospheric light. + +# 4 Experiments + +In this section, we show extensive qualitative and quantitative results of our method in various settings. Please see the appendix for many more examples, animations and discussion on limitations and negative impact. + +![](images/c0e7e34a89e1d7b3727f0ce0401f65198ca80b4fa9ea161895a8ea9d32147c2c.jpg) +Fig. 7: Ablation on Texture Adaptation. We ablate global-view guidance $(g.gd.)$ and reference feature injection $(r.f.i)$ . -ctrl. and $+\mathrm{ctrl}$ denotes less or more control in feature injection (smaller or higher $s_{\mathrm{f}}$ & $p$ ) for Identity Preservation requirement. + +![](images/622d9c7aeddf1d5a90381c5e85d951827bdf5b99bdd75c63928d963b80dbde80.jpg) + +# 4.1 Qualitative Evaluation + +We first present visual results and a case study to demonstrate the generalizability of the proposed method. We then compare our method with others that specifically focus on either scene composition or object stylization. + +Main Results. To demonstrate the generalizability of the framework and the impact of different scenes on the inserted objects, we present visual results in Fig. 5. Our method successfully blends the objects into various environments, achieving photorealistic adaptation for both appearance and lighting. This includes scenarios drawn from both real-world and fantasy settings. For example, the dimly-lit umbrella in Fig. 5 casts a shadow that is perfectly aligned with its handle and the light direction of the scene, substantially increasing the photorealism of the composition; The Belweder TV set is wrapped with gray mud consistent with the scene while its original structural details get maintained. + +Comparison on Scene Composition. We showcase the result from CDC [18] using LLM-prompted text prompts for a fair comparison. As shown in Fig. 6, CDC loses the structural details of the object while providing limited adaptation for texture transfer. Moreover, the composed scene appears artificial and synthetic, giving the impression that the sofa is floating above the mud. Our experiments reveal that the CDC often leads to object removal, particularly when the re-painting [31] is employed. Therefore, we further incorporate similarity guidance similar to the classifier guidance [12] by adding $\sqrt{1 - \alpha_t} \cdot \frac{\partial (||\pmb{x}_0^r - \hat{\pmb{x}}_\phi ||^2)}{\partial \pmb{x}_t}$ to the predicted gradient $\epsilon_{\phi}(\pmb{x}_t)$ each denoising step referred as CDC w/s.gd., but find it provides limited improvement. + +Comparison on Object Stylization. We showcase results from mesh texturing method [7, 45, 55] using our LLM-prompted text prompts for a fair comparison. As shown in Fig. 6, although achieving realistic texture adaptation, these mesh texturing methods fall short in preserving the original structural details of the object and in utilizing scene-level information. Essentially, These methods only satisfy the Environmental Influence requirement via text prompts and do not meet all the criteria of our intended goal. + +Table 2: Quantitative Evaluation. (a) We measure LPIPS for controllability, $\mathrm{CLIP}_{\mathrm{dir}}$ and $\mathrm{CLIP}_{\mathrm{SI}}$ for composability. (b) We present the percentages of users who prefer our method over other methods: CDC [18] and InstructP2P [3]. + +
MethodLPIPS ↓CLIPdir ↑CLIPSI ↑
InstructP2P [3]0.00200.01750.1852
Ours (T.A.)0.00140.02010.2006
CDC [18]0.00530.00420.1996
Ours (T.A. + L.E.)0.00150.02060.2013
+ +(a) Visual Metrics. + +
OursDrawOther
Ours vs. CDC [18]71%20%9%
Ours vs. InstructP2P [3]74%6%20%
+ +(b) User study for human preference. + +# 4.2 Quantitative Evaluation + +We collected 34 composites between diverse objects and scenes, each with 3 random views, yielding a total of 102 images. Besides CDC [18], we compare with InstructPix2Pix [3] by directly copy-pasting the object with adapted texture onto the scene. We measure LPIPS for controllability, $\mathrm{CLIP}_{\mathrm{dir}}$ and $\mathrm{CLIP}_{\mathrm{SI}}$ for composability, similar to CDC, and report the results in Tab. 2. Compared with InstructPix2Pix our method for texture adaptation (T.A) alone excels in controllability and composability. Light estimation (L.E.) further improves results. + +We also conducted a user study in Tab. 2 with 42 participants to compare the realism of our compositing under the condition that the object identity must be preserved. Our method is preferred over prior work in over $70\%$ of the cases. + +# 4.3 Ablation Study + +Identity Preservation. To illustrate how the Identity Preservation requirement is satisfied by reference feature injection (r.f.i.), we showcase the leather sofa in the swamp with less and more control (Fig. 7). The setup with less control drifts to a visually different sofa with too much mud wrapped around while more control generates a less muddy and adapted sofa compared to the reference image. Both adapted textures are less visually plausible. Additionally, when comparing Ours w/o g.gd. & r.f.i. to the reference image, it is evident that the identity of the sofa is lost. + +Global-View Guidance. To showcase the necessity of global-view guidance we conduct texture adaptation without it as Ours w/o g.gd. in Fig. 7. The adapted texture without seeing the full scene fails to adapt the color of the mud to that of the scene image, contrasting starkly with the scene. Experiments ablating the optimal global-view guidance set-up (e.g., solid-color background augmentation, usage of inpainting model, etc.) are detailed in the appendix. + +Light Estimation. In Fig. 1, the combination of accurately estimated lighting and shadows significantly enhances the photorealism of the composed image (compare the image on the left, which includes lighting, to those in the top right corner, which are without lighting). To further test the validity of the estimated light, we render images omitting the texture adaptation step. Fig. 8 demonstrates that the estimated light considers multiple factors, including its + +![](images/55d0f2c0196aa81bbf235a891b65a25da4524bad112f415ffa64c0df7d6f975b.jpg) +Fig. 8: Ablation on Light Estimation. We evaluate the effectiveness of light estimation by demonstrating the lighting effects on a sofa in various indoor and outdoor scenes. Overlaid in the left corners are objects rendered with the estimated light against a white background. In the right corners are objects before the light estimation. + +direction relative to the specific locations of objects in indoor scenes and its impact on the appearance of objects under atmospheric colored lighting conditions. Importantly, the estimated HDR maps, paired with the proposed scene composition, facilitate the realistic integration of shadows. + +# 5 Conclusion + +In this paper, we are motivated by a practical goal—what if a 3D object is placed into a 2D scene?—and propose a novel framework that allows (1) stylizing the object with an adapted texture that aligns with the given scene; and (2) achieving a photorealistic scene composition with the aid of estimated light of the environment. This is enabled by physically-based differentiable ray tracing and diffusion models used as guidance. Our method allows for automatic adaptation and blending of existing objects into a variety of scenes, making it useful for both 2D and 3D downstream applications, such as visual media and video games. + +Acknowledgement. This work is supported by the UKRI grant: Turing AI Fellowship EP/W002981/1 and EPSRC/MURI grant: EP/N019474/1. J. Zhou is also supported by the Horizon Robotics. We thank Luke Melas-Kyriazi and Fabio Pizzati for their helpful discussions. + +# References + +1. https://polyhaven.com, https://polyhaven.com +2. Akimoto, N., Matsuo, Y., Aoki, Y.: Diverse plausible 360-degree image outpainting for efficient 3dgc background creation. In: CVPR (2022) +3. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: CVPR (2023) +4. Cao, T., Kreis, K., Fidler, S., Sharp, N., Yin, K.: Texfusion: Synthesizing 3d textures with text-guided image diffusion models. In: ICCV (2023) +5. Chen, B.C., Kae, A.: Toward realistic image compositing with adversarial learning. In: CVPR (2019) +6. Chen, D.Z., Siddiqui, Y., Lee, H.Y., Tulyakov, S., Nießner, M.: Text2tex: Text-driven texture synthesis via diffusion models. arXiv preprint arXiv:2303.11396 (2023) +7. Chen, R., Chen, Y., Jiao, N., Jia, K.: Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In: ICCV (2023) +8. Cong, W., Niu, L., Zhang, J., Liang, J., Zhang, L.: Bargainnet: Background-guided domain translation for image harmonization. In: ICME (2021) +9. Cong, W., Zhang, J., Niu, L., Liu, L., Ling, Z., Li, W., Zhang, L.: Dovenet: Deep image harmonization via domain verification. In: CVPR (2020) +0. Deitke, M., Schwenk, D., Salvador, J., Weihs, L., Michel, O., VanderBilt, E., Schmidt, L., Ehsani, K., Kembhavi, A., Farhadi, A.: Objaverse: A universe of annotated 3d objects. In: CVPR (2023) +1. Deng, Y., Li, X., Liu, S., Yang, M.H.: Dip: Differentiable interreflection-aware physics-based inverse rendering. arXiv preprint arXiv:2212.04705 (2022) +2. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. NeurIPS (2021) +3. Eftekhar, A., Sax, A., Malik, J., Zamir, A.: Omnidata: A scalable pipeline for making multi-task mid-level vision datasets from 3d scans. In: ICCV (2021) +4. Gao, J., Shen, T., Wang, Z., Chen, W., Yin, K., Li, D., Litany, O., Gojcic, Z., Fidler, S.: Get3d: A generative model of high quality 3d textured shapes learned from images. NeurIPS (2022) +5. Gardner, M.A., Hold-Geoffroy, Y., Sunkavalli, K., Gagné, C., Lalonde, J.F.: Deep parametric indoor lighting estimation. In: ICCV (2019) +6. Gardner, M.A., Sunkavalli, K., Yumer, E., Shen, X., Gambaretto, E., Gagné, C., Lalonde, J.F.: Learning to predict indoor illumination from a single image. arXiv preprint arXiv:1704.00090 (2017) +7. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. NeurIPS (2014) +8. Hachnochi, R., Zhao, M., Orzech, N., Gal, R., Mahdavi-Amiri, A., Cohen-Or, D., Bermano, A.H.: Cross-domain compositing with pretrained diffusion models. arXiv preprint arXiv:2302.10167 (2023) +9. Haque, A., Tancik, M., Efros, A., Holynski, A., Kanazawa, A.: Instruct-nerf2nerf: Editing 3d scenes with instructions. In: CVPR (2023) + +20. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. NeurIPS (2020) +21. Ho, J., Salimans, T.: Classifier-free diffusion guidance. NeurIPSW (2021) +22. Hold-Geoffroy, Y., Athawale, A., Lalonde, J.F.: Deep sky modeling for single image outdoor lighting estimation. In: CVPR (2019) +23. Hold-Geoffroy, Y., Sunkavalli, K., Hadap, S., Gambaretto, E., Lalonde, J.F.: Deep outdoor illumination estimation. In: CVPR (2017) +24. Jakob, W., Speierer, S., Roussel, N., Vicini, D.: Dr.jit: A just-in-time compiler for differentiable rendering. TOG (2022) +25. Jiang, Y., Zhang, H., Zhang, J., Wang, Y., Lin, Z., Sunkavalli, K., Chen, S., Amirghodsi, S., Kong, S., Wang, Z.: Ssh: A self-supervised framework for image harmonization. In: ICCV (2021) +26. Kamata, H., Sakuma, Y., Hayakawa, A., Ishii, M., Narihira, T.: Instruct 3d-to-3d: Text instruction guided 3d-to-3d conversion. arXiv preprint arXiv:2303.15780 (2023) +27. Karimi Dastjerdi, M.R., Eisenmann, J., Hold-Geoffroy, Y., Lalonde, J.F.: Everlight: Indoor-outdoor editable hdr lighting estimation. In: ICCV (2023) +28. Karimi Dastjerdi, M.R., Hold-Geoffroy, Y., Eisenmann, J., Khodadadeh, S., Lalonde, J.F.: Guided co-modulated GAN for 360 field of view extrapolation. 3DV (2022) +29. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3d: High-resolution text-to-3d content creation. In: CVPR (2023) +30. Liu, M., Xu, C., Jin, H., Chen, L., Xu, Z., Su, H., et al.: One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. arXiv preprint arXiv:2306.16928 (2023) +31. Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Van Gool, L.: Repaint: Inpainting using denoising diffusion probabilistic models. In: CVPR (2022) +32. McAuley, S., Hill, S., Hoffman, N., Gotanda, Y., Smits, B., Burley, B., Martinez, A.: Practical physically-based shading in film and game production. In: SIGGRAPH (2012) +33. Metzer, G., Richardson, E., Patashnik, O., Giryes, R., Cohen-Or, D.: Latent-nerf for shape-guided generation of 3d shapes and textures. In: CVPR (2023) +34. Michel, O., Bar-On, R., Liu, R., Benaim, S., Hanocka, R.: Text2mesh: Text-driven neural stylization for meshes (2022) +35. Mohammad Khalid, N., Xie, T., Belilovsky, E., Popa, T.: Clip-mesh: Generating textured meshes from text using pretrained image-text models. In: SIGGRAPH Asia (2022) +36. Munkberg, J., Hasselgren, J., Shen, T., Gao, J., Chen, W., Evans, A., Müller, T., Fidler, S.: Extracting triangular 3d models, materials, and lighting from images. In: CVPR (2022) +37. Oechsle, M., Mescheder, L., Niemeyer, M., Strauss, T., Geiger, A.: Texture fields: Learning texture representations in function space. In: ICCV (2019) +38. OpenAI: Gpt-4 technical report (2023) +39. Parmar, G., Kumar Singh, K., Zhang, R., Li, Y., Lu, J., Zhu, J.Y.: Zero-shot image-to-image translation. In: SIGGRAPH (2023) +40. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. ICLR (2023) + +41. Qian, G., Mai, J., Hamdi, A., Ren, J., Siarohin, A., Li, B., Lee, H.Y., Skorokhodov, I., Wonka, P., Tulyakov, S., Ghanem, B.: Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. arXiv preprint arXiv:2306.17843 (2023) +42. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021) +43. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: ICML (2021) +44. Ranftl, R., Lasinger, K., Hafner, D., Schindler, K., Koltun, V.: Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. PAMI (2022) +45. Richardson, E., Metzer, G., Alaluf, Y., Giryes, R., Cohen-Or, D.: Texture: Text-guided texturing of 3d shapes. SIGGRAPH (2023) +46. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models (2021) +47. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS (2022) +48. Siddiqui, Y., Thies, J., Ma, F., Shan, Q., Nießner, M., Dai, A.: Texturify: Generating textures on 3d shape surfaces. In: ECCV (2022) +49. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. In: ICLR (2021), https://openreview.net/forum?id=PxTIG12RRHS +50. Song, Y., Zhang, Z., Lin, Z., Cohen, S., Price, B., Zhang, J., Kim, S.Y., Aliaga, D.: Objectstitch: Generative object compositing. CVPR (2023) +51. Tsalicoglou, C., Manhardt, F., Tonioni, A., Niemeyer, M., Tombari, F.: Textmesh: Generation of realistic 3d meshes from text prompts (2024) +52. Tumanyan, N., Geyer, M., Bagon, S., Dekel, T.: Plug-and-play diffusion features for text-driven image-to-image translation. In: CVPR (2023) +53. Walter, B., Marschner, S.R., Li, H., Torrance, K.E.: Microfacet models for refraction through rough surfaces. In: EGSR (2007) +54. Wang, H., Du, X., Li, J., Yeh, R.A., Shakhnarovich, G.: Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In: CVPR (2023) +55. Wang, Z., Lu, C., Wang, Y., Bao, F., Li, C., Su, H., Zhu, J.: Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. NeurIPS (2023) +56. Wei, L.Y., Lefebvre, S., Kwatra, V., Turk, G.: State of the art in example-based texture synthesis. EG-STAR (2009) +57. Xiao, J., Ehinger, K.A., Oliva, A., Torralba, A.: Recognizing scene viewpoint using panoramic place representation. In: ICCV (2012) +58. Zhang, J., Sunkavalli, K., Hold-Geoffroy, Y., Hadap, S., Eisenman, J., Lalonde, J.F.: All-weather deep outdoor lighting estimation. In: CVPR (2019) \ No newline at end of file diff --git a/2024/Scene-Conditional 3D Object Stylization and Composition/images.zip b/2024/Scene-Conditional 3D Object Stylization and Composition/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9ae70d7751e0460730df58c6c4b0a6f2767337da --- /dev/null +++ b/2024/Scene-Conditional 3D Object Stylization and Composition/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f81918565ce15d5c5b1f47c3b0e542e428f5519ccaf9403111e77c895a1c355c +size 782709 diff --git a/2024/Scene-Conditional 3D Object Stylization and Composition/layout.json b/2024/Scene-Conditional 3D Object Stylization and Composition/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b6de9d74d43942f269f16c5040b73a226b7efb6c --- /dev/null +++ b/2024/Scene-Conditional 3D Object Stylization and Composition/layout.json @@ -0,0 +1,8564 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 168, + 112, + 445, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 112, + 445, + 148 + ], + "spans": [ + { + "bbox": [ + 168, + 112, + 445, + 148 + ], + "type": "text", + "content": "Scene-Conditional 3D Object Stylization and Composition" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 169, + 468, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 169, + 468, + 181 + ], + "spans": [ + { + "bbox": [ + 146, + 169, + 468, + 181 + ], + "type": "text", + "content": "Jinghao Zhou" + }, + { + "bbox": [ + 146, + 169, + 468, + 181 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 146, + 169, + 468, + 181 + ], + "type": "text", + "content": ", Tomas Jakab" + }, + { + "bbox": [ + 146, + 169, + 468, + 181 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 146, + 169, + 468, + 181 + ], + "type": "text", + "content": ", Philip Torr" + }, + { + "bbox": [ + 146, + 169, + 468, + 181 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 146, + 169, + 468, + 181 + ], + "type": "text", + "content": ", and Christian Rupprecht" + }, + { + "bbox": [ + 146, + 169, + 468, + 181 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 219, + 190, + 394, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 190, + 394, + 213 + ], + "spans": [ + { + "bbox": [ + 219, + 190, + 394, + 213 + ], + "type": "text", + "content": "University of Oxford {jinghao,tomj,chrisr}@robots.ox.ac.uk" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 237, + 481, + 425 + ], + "blocks": [ + { + "bbox": [ + 133, + 237, + 481, + 425 + ], + "lines": [ + { + "bbox": [ + 133, + 237, + 481, + 425 + ], + "spans": [ + { + "bbox": [ + 133, + 237, + 481, + 425 + ], + "type": "image", + "image_path": "a322d6fb3ae17492b50af53c6a069cd9aa7f3ee71da505f94b738df7691db465.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 434, + 482, + 499 + ], + "lines": [ + { + "bbox": [ + 130, + 434, + 482, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 482, + 499 + ], + "type": "text", + "content": "Fig. 1: We present a framework that adapts a 3D object's appearance to a location in a 2D scene. It creates an image where the 3D object is seamlessly blended (left & bottom right), with its appearance influenced by the scene's environmental conditions and lighting effects. Moreover, the stylized object with adapted textures (top right), rendered here without the estimated lighting condition for illustrative purposes, can be further utilized as 3D assets for downstream tasks such as video games." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 522, + 459, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 522, + 459, + 666 + ], + "spans": [ + { + "bbox": [ + 159, + 522, + 459, + 666 + ], + "type": "text", + "content": "Abstract. Recently, 3D generative models have made impressive progress, enabling the generation of almost arbitrary 3D assets from text or image inputs. However, these approaches generate objects in isolation without any consideration for the scene where they will eventually be placed. In this paper, we propose a framework that allows for the stylization of an existing 3D asset to fit into a given 2D scene, and additionally produce a photorealistic composition as if the asset was placed within the environment. This not only opens up a new level of control for object stylization, for example, the same assets can be stylized to reflect changes in the environment, such as summer to winter or fantasy versus futuristic settings—but also makes the object-scene composition more controllable. We achieve this by combining modeling and optimizing the object's texture and environmental lighting through differentiable ray" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 159, + 117, + 455, + 150 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 117, + 455, + 150 + ], + "spans": [ + { + "bbox": [ + 159, + 117, + 455, + 150 + ], + "type": "text", + "content": "tracing with image priors from pre-trained text-to-image diffusion models. We demonstrate that our method applies to a wide variety of indoor and outdoor scenes and arbitrary objects. See also our project page." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 166, + 230, + 178 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 166, + 230, + 178 + ], + "spans": [ + { + "bbox": [ + 132, + 166, + 230, + 178 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 186, + 480, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 186, + 480, + 234 + ], + "spans": [ + { + "bbox": [ + 130, + 186, + 480, + 234 + ], + "type": "text", + "content": "After the success of generative models [20,43,46] for images in computer vision, the community is now interested in lifting these models to 3D [40,54] through advances in architectures, data, and training. These lifted 3D models demonstrate plausible 3D representations of objects conditioned on images or texts." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 235, + 481, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 235, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 235, + 481, + 282 + ], + "type": "text", + "content": "While generating whole objects—e.g. an astronaut riding a horse—is a challenging problem that generative models are becoming able to solve, in practice, these assets are difficult to adapt for use in downstream applications or existing virtual environments such as video games as they are generated without context." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 283, + 481, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 283, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 481, + 342 + ], + "type": "text", + "content": "This paper aims to provide a mechanism to close the appearance gap between 3D objects and environments. We formulate this problem through the lens of a creative tool. The intended goal of this paper is to align the appearance of a 3D object to a customized 2D scene as we place it into a specific location, a challenging task that requires fine-grained visual control for convincing results." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 342, + 481, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 481, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 481, + 462 + ], + "type": "text", + "content": "We are interested in producing a photorealistic scene-object composition where the 3D object is seamlessly blended. For instance, an object placed in a rainy night scene should appear wet and dimly lit, and if the scene is muddy and sunny, the object should also look muddy and cast a strong shadow. To achieve a high level of realism on a fine-grained level, the mud on the object must match the scene's mud in appearance and the direction of the shadow must match where the sun is located. Instead of simply resorting to 2D blending operations [18, 50], we explicitly model an object's appearance and lighting in 3D space, which enables more accurate and regularized control of the object's appearance in the composed image." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 463, + 481, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 463, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 463, + 481, + 568 + ], + "type": "text", + "content": "We also aim to preserve the underlying object geometry and appearance factors unrelated to the scene environment. This task is highly practical; for example, for media design, when presenting a 3D asset of a product within a realistic scene, we want the product to blend naturally while remaining recognizable. Manually achieving this would require a skilled artist to modify the object's texture to match the scene, a challenging and time-consuming process. Our paper introduces a method that automates this task. It accepts a 3D object with texture, a 2D scene, and the object's position within the scene and outputs an adapted texture consistent with the scene." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 570, + 481, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 481, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 481, + 640 + ], + "type": "text", + "content": "To accomplish this, we leverage recent advances in generative diffusion models [46, 47] for images, which are trained on a large-scale dataset of internet images. We conduct an optimization procedure where we compose the 2D renderings of the 3D object with the 2D scene at a specified location and use a pre-trained diffusion model to \"critique\" the realism of the composed image, providing a gradient signal for optimization." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": "The challenge is that diffusion models, such as Stable Diffusion [46], are trained on 2D images and lack any notion of 3D geometry. Naively optimizing" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "J. Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "text", + "content": "or denoising in 2D space [18,50] would result in losing the object's 3D geometry. To prevent this, we optimize the object's content in 3D by randomly rotating the object during optimization and rendering it from various viewpoints, which helps to align the adapted texture with the object's geometry and avoid overfitting to a single viewpoint." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 480, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 480, + 319 + ], + "type": "text", + "content": "Notably, randomly rotating the 3D object stops lighting effects from being baked into the texture, which is desirable. Yet, it also prevents the object from naturally blending with the scene in a single viewpoint. To counteract this, we predict the scene's lighting and the shadows cast by the object as separate components. For this purpose, we propose a novel technique inspired by a common physical approach from the computer graphics industry, where reflective and diffuse spheres with known material parameters are placed and photographed in a real-world scene to capture the environmental lighting conditions. We adopt this concept and place a virtual white diffuse sphere inside the scene during optimization while conditioning the generative model to expect a white sphere within the scene, thereby capturing the model's interpretation of the scene's lighting in an easily extracted form." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 319, + 480, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 480, + 392 + ], + "type": "text", + "content": "Optimizing the object's appearance in 3D offers further significant advantages: The user can change the object's pose without re-running the optimization, as the texture is meaningful from all viewpoints, and view-dependent effects are modeled separately by the lighting component. Additionally, the 3D object with the adapted texture can be used as a standalone asset in other downstream tasks (e.g. in a video game)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 392, + 480, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 480, + 440 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 480, + 440 + ], + "type": "text", + "content": "In summary, we present a framework that (1) enables the stylization of a 3D object, which can also be used as a standalone asset, given a 2D scene and its location within it; (2) achieves photorealistic scene-object compositing with the help of a novel mechanism to estimate scene lighting from a single image." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 440, + 480, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 440, + 480, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 440, + 480, + 475 + ], + "type": "text", + "content": "We further verify the effectiveness of our method in extensive experiments, demonstrating that our approach can realistically adapt objects to a diverse set of environments, laying the groundwork for practical applications." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 491, + 237, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 491, + 237, + 503 + ], + "spans": [ + { + "bbox": [ + 131, + 491, + 237, + 503 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 513, + 479, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 513, + 479, + 573 + ], + "spans": [ + { + "bbox": [ + 130, + 513, + 479, + 573 + ], + "type": "text", + "content": "The purpose of this paper is to adapt the appearance of a 3D object to align with a customized 2D scene. This solves a more complicated problem than prior works, which often only tackle individual components, including but not limited to 2D scene-object compositing, light estimation, and texturing, but never as a whole. We compare with these methods next and detail their differences in Tab. 1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 582, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 480, + 666 + ], + "type": "text", + "content": "Scene-Object Compositing. Scene-object compositing is a challenging task that requires a delicate balance between visual transfer and control. It has been extensively researched [5,8,9,25]. Recent methods condition the denoising process of diffusion models [20,49] on object images. OST [50] aims to seamlessly integrate a 2D object image into a 2D scene by modifying the keys and values in the attention blocks. CDC [18] considers composing the 2D renderings of a 3D object with a 2D scene by directly modifying denoised latent." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 209, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 209, + 91, + 447, + 103 + ], + "type": "text", + "content": "Scene-Conditional 3D Object Stylization and Composition" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 137, + 168, + 477, + 304 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 160 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 160 + ], + "type": "text", + "content": "Table 1: Comparison with related work. Rep. denotes representation being 2D (□) or 3D (）。Env. Inf. denotes Environmental Influence and I.D. Pre. denotes Identity Preservation. Parameters for the lighting and the appearance model are either scene-agnostic constant (），scene-conditioned predictable (），or learnable (）。" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 137, + 168, + 477, + 304 + ], + "lines": [ + { + "bbox": [ + 137, + 168, + 477, + 304 + ], + "spans": [ + { + "bbox": [ + 137, + 168, + 477, + 304 + ], + "type": "table", + "html": "
MethodSceneObjectCompose
Rep.LightingShadowRep.AppearanceEnv. Inf.I.D. Pre.
OST [50]□-×□-××Blend
CDC [18]□Ambient×Diffuse×××Blend
IGAN [28]□LDR✓Diffuse×××Copy & Paste
EVL [27]□HDR✓Diffuse×××Copy & Paste
PrDr [55]-Ambient×Diffuse✓×-
Fan3D [7]-HDR×PBR✓×-
Inst3D [26]-Point×Diffuse×✓-
Ours□HDR✓PBR✓✓Copy & Paste
", + "image_path": "53d0cd996d163bd3a4717d79a5532e1a1763d7314b12d94e7e9cce75e212ebb3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 311, + 480, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 311, + 480, + 360 + ], + "spans": [ + { + "bbox": [ + 130, + 311, + 480, + 360 + ], + "type": "text", + "content": "However, these methods rely entirely on the 2D denoising process and do not explicitly model the object's geometry, which often results in a significant loss of structural details and visual identity. This severely hampers these methods' ability to adapt to the environmental and lighting conditions of the scene." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 368, + 480, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 368, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 368, + 480, + 464 + ], + "type": "text", + "content": "Light Estimation from Single-View Image. Estimating light conditions from a single-view 2D image is of practical use in the real world and has been extensively studied for both indoor [15,16] and outdoor scenes [22,23,58]. The distinction between these two settings has been significantly narrowed recently [27, 28], thanks to the outpainting capabilities of diffusion models, which have contributed to a unified framework. With additional training on large high dynamic range (HDR) datasets such as SUN360 [57], these methods can approximate the location and intensity of the light source, enabling realistic shadows." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 464, + 481, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 464, + 481, + 500 + ], + "spans": [ + { + "bbox": [ + 130, + 464, + 481, + 500 + ], + "type": "text", + "content": "However, these methods estimate only the lighting of the scene. They are typically indifferent to how objects should be stylized to match the scene, neglecting the potential environmental influences on the textures of objects." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "Mesh Texturing. Classic automated approaches [56] for mesh texturing are often limited and apply only simple texture patterns. In contrast, methods [14, 37, 48] based on Generative Adversarial Networks [17] (GANs) are more capable of synthesizing detailed textures. Recently, text-conditioned texture generation [34, 35] has been made possible through the optimization of CLIP-based objectives [42]. The latest works now employ powerful text-to-image diffusion models [20, 49], either through optimization [7, 29, 33, 51, 55] using Score Distillation Sampling [40] (SDS), or by wrapping the generated 2D textures onto the mesh [4, 6, 45]. Importantly, Variational Score Distillation [55] (VSD) as an improved optimization objective and the adoption of PBR model [7] have greatly boosted the quality of generated textures. Additionally, an increasing level of requirement for more visual control from users has been partially met [19, 26] through image-conditional diffusion models [3] to preserve objects' identities." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "J. Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 114, + 461, + 222 + ], + "blocks": [ + { + "bbox": [ + 149, + 114, + 461, + 222 + ], + "lines": [ + { + "bbox": [ + 149, + 114, + 461, + 222 + ], + "spans": [ + { + "bbox": [ + 149, + 114, + 461, + 222 + ], + "type": "image", + "image_path": "ae638e978c607d549deeff73f662f294c56565eeeaba25ed3e50646025032972.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 231, + 482, + 307 + ], + "lines": [ + { + "bbox": [ + 130, + 231, + 482, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 231, + 482, + 307 + ], + "type": "text", + "content": "Fig. 2: Framework. We learn an environment map and a texture map separately from the 2D supervision. We initialize (init.) the environment map with an LDR map estimated from the 2D scene and learn light multiplying scales bright areas, yielding an HDR map. We employ the PBR material model for texture maps, encoded via MLP with positional encoding. The object is rendered through a differentiable ray tracer and further composed with the scene background, receiving gradients from Stable Diffusion (SD) in the latent space." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 315, + 482, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 315, + 482, + 352 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 482, + 352 + ], + "type": "text", + "content": "However, these works focus on texturing the object in isolation from the scene and do not consider the homogeneity of the composition when these assets are placed in certain environments, which limits their practical usage." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 370, + 202, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 370, + 202, + 383 + ], + "spans": [ + { + "bbox": [ + 132, + 370, + 202, + 383 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 396, + 482, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 396, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 482, + 468 + ], + "type": "text", + "content": "The overall structure of our framework can be found in Fig. 2. Given a textured 3D object, a 2D image of a scene, and a 2D location where to place the object, we first adapt the texture of the 3D object (Secs. 3.2 and 3.3) to align with the scene. To enable a photorealistic composition of the object and the scene, we further estimate the lighting (Secs. 3.4 to 3.6) and apply it during the final rendering." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 487, + 228, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 487, + 228, + 499 + ], + "spans": [ + { + "bbox": [ + 131, + 487, + 228, + 499 + ], + "type": "text", + "content": "3.1 Optimization" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 508, + 481, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 508, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 508, + 481, + 533 + ], + "type": "text", + "content": "To enable end-to-end optimization, we adopt a differentiable renderer to render the object and score it using the priors from diffusion models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 544, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 482, + 617 + ], + "type": "text", + "content": "Differentiable Renderer. We employ the differential renderer Mitsuba3 [24] (denoted as " + }, + { + "bbox": [ + 130, + 544, + 482, + 617 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 130, + 544, + 482, + 617 + ], + "type": "text", + "content": "), due to its capabilities as a physically-based ray-tracing renderer equipped with versatile rendering options, such as diverse BSDFs, the path tracer integrator, and using an environment map as a light source. This enables us to produce photo-realistic renders complete with shadows and reflections, forming the foundation for robust texture adaptation and light estimation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "Diffusion Model as Guidance. To score the rendered image, we make use of a generative diffusion model [46, 47] " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": ". Diffusion models have learned powerful priors of the image formation process in the real world from large-scale training" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "type": "text", + "content": "Scene-Conditional 3D Object Stylization and Composition" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "content": "on billion-scale web data. Similar to prior work [40,54], we supervise the rendered image " + }, + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0 = g(\\pmb{\\theta})" + }, + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "content": " with a score-matching objective to obtain an image gradient" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 215, + 150, + 481, + 176 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 215, + 150, + 481, + 176 + ], + "spans": [ + { + "bbox": [ + 215, + 150, + 481, + 176 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\boldsymbol {\\theta}} \\mathcal {L} (\\boldsymbol {x} _ {0}) \\triangleq \\mathbb {E} _ {t, \\epsilon} \\left[ w _ {t} \\left(\\epsilon_ {\\phi} (\\boldsymbol {x} _ {t}) - \\hat {\\epsilon} (\\boldsymbol {x} _ {t})\\right) \\frac {\\partial \\boldsymbol {x} _ {0}}{\\partial \\boldsymbol {\\theta}} \\right], \\tag {1}", + "image_path": "b549514dd723113fa64b3e0d28544abab7069a1c37b066287840d4d9e860d54f.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 185, + 479, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 185, + 479, + 305 + ], + "spans": [ + { + "bbox": [ + 132, + 185, + 479, + 305 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 185, + 479, + 305 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t = \\sqrt{\\alpha_t}\\pmb{x}_0 + \\sqrt{1 - \\alpha_t}\\epsilon" + }, + { + "bbox": [ + 132, + 185, + 479, + 305 + ], + "type": "text", + "content": " is the noisy rendered image, " + }, + { + "bbox": [ + 132, + 185, + 479, + 305 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}(\\pmb{x}_t)" + }, + { + "bbox": [ + 132, + 185, + 479, + 305 + ], + "type": "text", + "content": " is an approximated score for a noisy real image, " + }, + { + "bbox": [ + 132, + 185, + 479, + 305 + ], + "type": "inline_equation", + "content": "t\\in \\mathbb{N}" + }, + { + "bbox": [ + 132, + 185, + 479, + 305 + ], + "type": "text", + "content": " indexes the discrete diffusion steps, and " + }, + { + "bbox": [ + 132, + 185, + 479, + 305 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}" + }, + { + "bbox": [ + 132, + 185, + 479, + 305 + ], + "type": "text", + "content": " is a scoring function for " + }, + { + "bbox": [ + 132, + 185, + 479, + 305 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t" + }, + { + "bbox": [ + 132, + 185, + 479, + 305 + ], + "type": "text", + "content": ". Specifically, " + }, + { + "bbox": [ + 132, + 185, + 479, + 305 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}" + }, + { + "bbox": [ + 132, + 185, + 479, + 305 + ], + "type": "text", + "content": " is set as constant " + }, + { + "bbox": [ + 132, + 185, + 479, + 305 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 132, + 185, + 479, + 305 + ], + "type": "text", + "content": " in Score Distillation Sampling [40] (SDS), and " + }, + { + "bbox": [ + 132, + 185, + 479, + 305 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\psi}" + }, + { + "bbox": [ + 132, + 185, + 479, + 305 + ], + "type": "text", + "content": " is an online LoRA-tuned network initialized from " + }, + { + "bbox": [ + 132, + 185, + 479, + 305 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}" + }, + { + "bbox": [ + 132, + 185, + 479, + 305 + ], + "type": "text", + "content": " in Variational Score Sampling [55] (VSD). Notably, we employ their interpolated variant given by " + }, + { + "bbox": [ + 132, + 185, + 479, + 305 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon} = \\lambda \\epsilon_{\\psi} + (1 - \\lambda)\\epsilon" + }, + { + "bbox": [ + 132, + 185, + 479, + 305 + ], + "type": "text", + "content": " with a " + }, + { + "bbox": [ + 132, + 185, + 479, + 305 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 132, + 185, + 479, + 305 + ], + "type": "text", + "content": " annealed from 1 to " + }, + { + "bbox": [ + 132, + 185, + 479, + 305 + ], + "type": "inline_equation", + "content": "\\lambda_e" + }, + { + "bbox": [ + 132, + 185, + 479, + 305 + ], + "type": "text", + "content": ", helping to counteract the over-texturization often observed in VSD, especially for objects with simple surfaces. With residual of noises " + }, + { + "bbox": [ + 132, + 185, + 479, + 305 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi} - \\hat{\\epsilon} = (\\epsilon_{\\phi} - \\epsilon) - \\lambda (\\epsilon_{\\psi} - \\epsilon)" + }, + { + "bbox": [ + 132, + 185, + 479, + 305 + ], + "type": "text", + "content": ", we highlight its relation with the image gradient obtained in traditional inverse rendering setup " + }, + { + "bbox": [ + 132, + 185, + 479, + 305 + ], + "type": "inline_equation", + "content": "\\nabla_{\\theta}\\mathcal{L}(\\pmb{x}_0) = \\pmb{x}_0 - \\pmb{x}_{\\mathrm{gt}}" + }, + { + "bbox": [ + 132, + 185, + 479, + 305 + ], + "type": "text", + "content": " by reformulating it as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 211, + 316, + 480, + 341 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 316, + 480, + 341 + ], + "spans": [ + { + "bbox": [ + 211, + 316, + 480, + 341 + ], + "type": "interline_equation", + "content": "\\epsilon_ {\\phi} - \\hat {\\epsilon} = \\sqrt {\\frac {\\alpha_ {t}}{1 - \\alpha_ {t}}} [ (\\boldsymbol {x} _ {0} - \\hat {\\boldsymbol {x}} _ {\\phi}) - \\lambda (\\boldsymbol {x} _ {0} - \\hat {\\boldsymbol {x}} _ {\\psi}) ], \\tag {2}", + "image_path": "169f0d79e6005607ea0407a9e88fed95090b13f162e3222417b3ae5a687880f6.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 350, + 480, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 350, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 350, + 480, + 434 + ], + "type": "text", + "content": "where the target used to supervise " + }, + { + "bbox": [ + 132, + 350, + 480, + 434 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0" + }, + { + "bbox": [ + 132, + 350, + 480, + 434 + ], + "type": "text", + "content": " is predicted clean image " + }, + { + "bbox": [ + 132, + 350, + 480, + 434 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}}_{\\phi}" + }, + { + "bbox": [ + 132, + 350, + 480, + 434 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 132, + 350, + 480, + 434 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}" + }, + { + "bbox": [ + 132, + 350, + 480, + 434 + ], + "type": "text", + "content": " as the positive and " + }, + { + "bbox": [ + 132, + 350, + 480, + 434 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}}_{\\psi}" + }, + { + "bbox": [ + 132, + 350, + 480, + 434 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 132, + 350, + 480, + 434 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\psi}" + }, + { + "bbox": [ + 132, + 350, + 480, + 434 + ], + "type": "text", + "content": " as the negative. The technique enables nominal classifier-free guidance [21] and thus bypasses mode-seeking behaviors and saturated texture. Inspired by [45], we incorporate the inner product of the camera pose and object's surface normal " + }, + { + "bbox": [ + 132, + 350, + 480, + 434 + ], + "type": "inline_equation", + "content": "\\langle c, \\pmb{n} \\rangle" + }, + { + "bbox": [ + 132, + 350, + 480, + 434 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 132, + 350, + 480, + 434 + ], + "type": "inline_equation", + "content": "w_t" + }, + { + "bbox": [ + 132, + 350, + 480, + 434 + ], + "type": "text", + "content": " to penalize faces at small grazing angles. In practice, we use Stable Diffusion [46] (SD) and absorb the encoder into " + }, + { + "bbox": [ + 132, + 350, + 480, + 434 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 132, + 350, + 480, + 434 + ], + "type": "text", + "content": ", optimizing in latent space." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 452, + 256, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 452, + 256, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 452, + 256, + 464 + ], + "type": "text", + "content": "3.2 Appearance Model" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 473, + 479, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 473, + 479, + 534 + ], + "spans": [ + { + "bbox": [ + 132, + 473, + 479, + 534 + ], + "type": "text", + "content": "Textured Mesh. We represent the 3D object as a textured mesh, which is the most common format for 3D assets in the industry and is compatible with many established workflows. Textured meshes are available in abundance in online datasets and marketplaces [1,10], and recently, one can also generate reasonable quality 3D meshes with text descriptions [7,29,55] or limited viewpoints [30,41]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 546, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 546, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 546, + 481, + 665 + ], + "type": "text", + "content": "Neural Texture. We represent the appearance of objects using a texture map " + }, + { + "bbox": [ + 132, + 546, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{T} \\in \\mathbb{R}^{H_t \\times W_t \\times 5}" + }, + { + "bbox": [ + 132, + 546, + 481, + 665 + ], + "type": "text", + "content": ", facilitating compatibility of textured results with widely-used rendering engines, such as Blender. We parameterize the texture using a coordinate neural network that maps 3D coordinates to the texture attributes. We first map the 3D coordinates of vertices using a rasterizer [36] based on their UV coordinates into the space of the texture map, resulting in " + }, + { + "bbox": [ + 132, + 546, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{C} \\in \\mathbb{R}^{H_t \\times W_t \\times 3}" + }, + { + "bbox": [ + 132, + 546, + 481, + 665 + ], + "type": "text", + "content": ". We then use a 2-layer parameterized MLP without bias and Hash positional encoding " + }, + { + "bbox": [ + 132, + 546, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\varPhi: \\mathcal{C} \\to \\mathcal{T}" + }, + { + "bbox": [ + 132, + 546, + 481, + 665 + ], + "type": "text", + "content": " to map the coordinates to their texture attributes. For the final layer, we re-map the output activated after sigmoid according to pre-defined minimal and maximal values. The five channels in " + }, + { + "bbox": [ + 132, + 546, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 132, + 546, + 481, + 665 + ], + "type": "text", + "content": " correspond to the Physically-Based" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "J. Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 116, + 460, + 242 + ], + "blocks": [ + { + "bbox": [ + 149, + 116, + 460, + 242 + ], + "lines": [ + { + "bbox": [ + 149, + 116, + 460, + 242 + ], + "spans": [ + { + "bbox": [ + 149, + 116, + 460, + 242 + ], + "type": "image", + "image_path": "9d8e045379d92a98dbde3d91fb581c9b5641cbcd3210a98d1af91a471e966bf6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 253, + 480, + 285 + ], + "lines": [ + { + "bbox": [ + 131, + 253, + 480, + 285 + ], + "spans": [ + { + "bbox": [ + 131, + 253, + 480, + 285 + ], + "type": "text", + "content": "Fig. 3: Pipeline for texture adaptation. We initialize the neural texture from the reference object and inject the feature of reference renderings to the U-Net of SD. We use both local-view and global-view guidance." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 294, + 480, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 294, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 131, + 294, + 480, + 354 + ], + "type": "text", + "content": "Rendering (PBR) shading model [32], which includes a diffuse term " + }, + { + "bbox": [ + 131, + 294, + 480, + 354 + ], + "type": "inline_equation", + "content": "k_{\\mathrm{d}} \\in \\mathbb{R}^{3}" + }, + { + "bbox": [ + 131, + 294, + 480, + 354 + ], + "type": "text", + "content": " and an isotropic specular GGX lobe [53] described by " + }, + { + "bbox": [ + 131, + 294, + 480, + 354 + ], + "type": "inline_equation", + "content": "k_{\\mathrm{rm}} = (k_r, k_m) \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 131, + 294, + 480, + 354 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 131, + 294, + 480, + 354 + ], + "type": "inline_equation", + "content": "k_r" + }, + { + "bbox": [ + 131, + 294, + 480, + 354 + ], + "type": "text", + "content": " representing roughness and " + }, + { + "bbox": [ + 131, + 294, + 480, + 354 + ], + "type": "inline_equation", + "content": "k_m" + }, + { + "bbox": [ + 131, + 294, + 480, + 354 + ], + "type": "text", + "content": " indicating metalness. The PBR model, with " + }, + { + "bbox": [ + 131, + 294, + 480, + 354 + ], + "type": "inline_equation", + "content": "k_r" + }, + { + "bbox": [ + 131, + 294, + 480, + 354 + ], + "type": "text", + "content": " affecting the scattering of light and " + }, + { + "bbox": [ + 131, + 294, + 480, + 354 + ], + "type": "inline_equation", + "content": "k_m" + }, + { + "bbox": [ + 131, + 294, + 480, + 354 + ], + "type": "text", + "content": " influencing the reflective properties, creates realistic light interactions between surfaces." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 372, + 302, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 372, + 302, + 383 + ], + "spans": [ + { + "bbox": [ + 132, + 372, + 302, + 383 + ], + "type": "text", + "content": "3.3 Scene-Conditional Guidance" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 391, + 480, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 391, + 480, + 462 + ], + "spans": [ + { + "bbox": [ + 131, + 391, + 480, + 462 + ], + "type": "text", + "content": "Achieving a realistic texture adaptation for 3D objects placed into 2D scenes requires three components: Environmental Influence involves adjusting the texture to reflect the environmental impact realistically; Identity Preservation aims to preserve the objects' unique visual aspects, and Blending focuses on guiding the texture adaptation to match the visual characteristic of the scene, ensuring seamless integration with the surroundings and avoiding stark contrasts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 474, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 474, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 474, + 481, + 666 + ], + "type": "text", + "content": "Environmental Influence. To show the realistic effects of the scene environment on the object's appearance, we guide the diffusion model with text prompts. We find that a naive text prompt, combining a simple object description with a scene description, e.g., a leather sofa in a swamp, is insufficient to generate realistic textures that reflect the environmental impact on the object, as shown in the appendix. For instance, we would expect a sofa placed in a swamp to be dirty and possibly show signs of moss or other vegetation growth. Instead of manually crafting a detailed text prompt for each object and scene combination, we automatize the process by employing a large language model (LLM), specifically GPT-4 [38], to generate a text prompt for the image diffusion model. Given the simple description of the object (e.g., a leather sofa) and the scene (e.g., a swamp), we instruct the LLM first to analyze the potential impact of the scene's environment on the object's appearance and then to distil this information into a text prompt that describes the resulting appearance of the object (e.g., The leather sofa, partially submerged in the swamp, looks discolored and soggy, its once-polished surface marred by mud, moss, and the murky water of the swamp" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "type": "text", + "content": "Scene-Conditional 3D Object Stylization and Composition" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 115, + 482, + 175 + ], + "blocks": [ + { + "bbox": [ + 134, + 115, + 482, + 175 + ], + "lines": [ + { + "bbox": [ + 134, + 115, + 482, + 175 + ], + "spans": [ + { + "bbox": [ + 134, + 115, + 482, + 175 + ], + "type": "image", + "image_path": "384078b29483dc05b10914722d91a27b165a49f4eebdafa8d70157f7f1e74900.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "lines": [ + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "type": "text", + "content": "Fig. 4: Pipeline for estimating the LDR map. We utilize tailored pipelines for indoor and outdoor scenes. Areas masked red and blue correspond to the far light " + }, + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "type": "text", + "content": " and near light region." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 226, + 482, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 226, + 482, + 262 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 482, + 262 + ], + "type": "text", + "content": "environment.). We then use this automatically generated text prompt to guide the diffusion model in adapting the texture. The specific prompt design for this automated process is detailed in the appendix." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 282, + 482, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 282, + 482, + 379 + ], + "spans": [ + { + "bbox": [ + 130, + 282, + 482, + 379 + ], + "type": "text", + "content": "Identity Preservation. To prevent the transferred texture from drifting away from its original content, we regularize the denoising diffusion model with reference images, a technique that has been explored for controlled image generation [39, 52]. Suppose we have a rendered reference image " + }, + { + "bbox": [ + 130, + 282, + 482, + 379 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0^r = g(\\pmb{\\theta}^r)" + }, + { + "bbox": [ + 130, + 282, + 482, + 379 + ], + "type": "text", + "content": " and its noised version " + }, + { + "bbox": [ + 130, + 282, + 482, + 379 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t^r" + }, + { + "bbox": [ + 130, + 282, + 482, + 379 + ], + "type": "text", + "content": " as input forwarded to " + }, + { + "bbox": [ + 130, + 282, + 482, + 379 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}" + }, + { + "bbox": [ + 130, + 282, + 482, + 379 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 282, + 482, + 379 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\psi}" + }, + { + "bbox": [ + 130, + 282, + 482, + 379 + ], + "type": "text", + "content": ", for the Multi-Head Self-Attention (MHSA) layer in each Transformer block of " + }, + { + "bbox": [ + 130, + 282, + 482, + 379 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}" + }, + { + "bbox": [ + 130, + 282, + 482, + 379 + ], + "type": "text", + "content": "'s and " + }, + { + "bbox": [ + 130, + 282, + 482, + 379 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\psi}" + }, + { + "bbox": [ + 130, + 282, + 482, + 379 + ], + "type": "text", + "content": "'s U-Nets, we concatenate the key " + }, + { + "bbox": [ + 130, + 282, + 482, + 379 + ], + "type": "inline_equation", + "content": "K^r" + }, + { + "bbox": [ + 130, + 282, + 482, + 379 + ], + "type": "text", + "content": " and value " + }, + { + "bbox": [ + 130, + 282, + 482, + 379 + ], + "type": "inline_equation", + "content": "V^r" + }, + { + "bbox": [ + 130, + 282, + 482, + 379 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 130, + 282, + 482, + 379 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t^r" + }, + { + "bbox": [ + 130, + 282, + 482, + 379 + ], + "type": "text", + "content": " to the original " + }, + { + "bbox": [ + 130, + 282, + 482, + 379 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 282, + 482, + 379 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 282, + 482, + 379 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 282, + 482, + 379 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 130, + 282, + 482, + 379 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t" + }, + { + "bbox": [ + 130, + 282, + 482, + 379 + ], + "type": "text", + "content": " yielding an image-conditional attention output" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 214, + 391, + 481, + 406 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 391, + 481, + 406 + ], + "spans": [ + { + "bbox": [ + 214, + 391, + 481, + 406 + ], + "type": "interline_equation", + "content": "A (\\boldsymbol {x} _ {t}, \\cdot , \\boldsymbol {x} _ {t} ^ {r}) = \\mathrm {M H S A} (Q, [ K; K ^ {r} ], [ V; V ^ {r} ]), \\tag {3}", + "image_path": "c68d9c603a58aa28dad8ae940180a6334ca149a91f71439047d5468499cef68a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 417, + 482, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 417, + 482, + 478 + ], + "spans": [ + { + "bbox": [ + 130, + 417, + 482, + 478 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 417, + 482, + 478 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 130, + 417, + 482, + 478 + ], + "type": "text", + "content": " can either be text or " + }, + { + "bbox": [ + 130, + 417, + 482, + 478 + ], + "type": "inline_equation", + "content": "\\emptyset" + }, + { + "bbox": [ + 130, + 417, + 482, + 478 + ], + "type": "text", + "content": " for text conditioning. The text-unconditional attention " + }, + { + "bbox": [ + 130, + 417, + 482, + 478 + ], + "type": "inline_equation", + "content": "A(\\pmb{x}_t, \\emptyset, \\pmb{x}_t^r)" + }, + { + "bbox": [ + 130, + 417, + 482, + 478 + ], + "type": "text", + "content": " used for classifier-free guidance [21] is further modified as " + }, + { + "bbox": [ + 130, + 417, + 482, + 478 + ], + "type": "inline_equation", + "content": "s_c \\cdot A(\\pmb{x}_t, \\emptyset, \\emptyset) + (1 - s_c) \\cdot A(\\pmb{x}_t, \\emptyset, \\pmb{x}_t^r)" + }, + { + "bbox": [ + 130, + 417, + 482, + 478 + ], + "type": "text", + "content": " weighted by a control guidance scale " + }, + { + "bbox": [ + 130, + 417, + 482, + 478 + ], + "type": "inline_equation", + "content": "s_c" + }, + { + "bbox": [ + 130, + 417, + 482, + 478 + ], + "type": "text", + "content": ". With the percentage " + }, + { + "bbox": [ + 130, + 417, + 482, + 478 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 417, + 482, + 478 + ], + "type": "text", + "content": " of blocks to be injected, the intensity of preservation can be controlled: " + }, + { + "bbox": [ + 130, + 417, + 482, + 478 + ], + "type": "inline_equation", + "content": "s_c = p = 1" + }, + { + "bbox": [ + 130, + 417, + 482, + 478 + ], + "type": "text", + "content": " provide the most and " + }, + { + "bbox": [ + 130, + 417, + 482, + 478 + ], + "type": "inline_equation", + "content": "s_c = p = 0" + }, + { + "bbox": [ + 130, + 417, + 482, + 478 + ], + "type": "text", + "content": " the least preservation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 498, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 483, + 666 + ], + "type": "text", + "content": "Blending. The two components previously mentioned directly utilize object renderings, referred to as the local view, as inputs for vanilla SD. To further enhance the adaptation of the object's appearance to its surroundings, we input the scene image, with the object rendered onto it, referred to as the global view, into an additional inpainting SD [46]. We have found that the inpainting model provides superior conditioning on the surrounding environment depicted in the image, resulting in improved blending. However, the inpainting model can sometimes attempt to undesirably modify the inserted object, removing parts of it by inpainting them with the background. To counter this, we compose the local view with randomly sample solid background colors, rather than the scene. This prevents the inpainting model from removing parts of the object as the object must look realistic when rendered in front of a random background. In practice, we utilize a set of global views created by cropping the original global view at various scales, ensuring that the crops encompass the entire local view." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "J. Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 123, + 476, + 467 + ], + "blocks": [ + { + "bbox": [ + 137, + 123, + 476, + 467 + ], + "lines": [ + { + "bbox": [ + 137, + 123, + 476, + 467 + ], + "spans": [ + { + "bbox": [ + 137, + 123, + 476, + 467 + ], + "type": "image", + "image_path": "865fcc92eb9818599c57327de653788632139a206693e2bfba30d0f6a4851bce.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 269, + 467, + 343, + 473 + ], + "lines": [ + { + "bbox": [ + 269, + 467, + 343, + 473 + ], + "spans": [ + { + "bbox": [ + 269, + 467, + 343, + 473 + ], + "type": "text", + "content": "(a) Diverse Objects and Scenes" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 137, + 479, + 476, + 538 + ], + "blocks": [ + { + "bbox": [ + 137, + 479, + 476, + 538 + ], + "lines": [ + { + "bbox": [ + 137, + 479, + 476, + 538 + ], + "spans": [ + { + "bbox": [ + 137, + 479, + 476, + 538 + ], + "type": "image", + "image_path": "f2a43d92e204323547645b405ac3fd1cc9f21e20c0e32dac4ecd4054bd961f8a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 173, + 539, + 214, + 545 + ], + "lines": [ + { + "bbox": [ + 173, + 539, + 214, + 545 + ], + "spans": [ + { + "bbox": [ + 173, + 539, + 214, + 545 + ], + "type": "text", + "content": "(b) Small Objects" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 288, + 539, + 324, + 545 + ], + "lines": [ + { + "bbox": [ + 288, + 539, + 324, + 545 + ], + "spans": [ + { + "bbox": [ + 288, + 539, + 324, + 545 + ], + "type": "text", + "content": "(c) Big Objects" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 389, + 539, + 444, + 545 + ], + "lines": [ + { + "bbox": [ + 389, + 539, + 444, + 545 + ], + "spans": [ + { + "bbox": [ + 389, + 539, + 444, + 545 + ], + "type": "text", + "content": "(d) Different Locations" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 568, + 482, + 657 + ], + "lines": [ + { + "bbox": [ + 130, + 568, + 482, + 657 + ], + "spans": [ + { + "bbox": [ + 130, + 568, + 482, + 657 + ], + "type": "text", + "content": "Fig. 5: Visual Results. (a) We showcase that our method applies to a diverse range of objects and scenes. The global view (top row) the overall composition quality and object-centric local view (bottom two rows) for the fidelity of stylized textures are demonstrated. For dim scenes, we additionally render objects without the estimated lighting condition " + }, + { + "bbox": [ + 130, + 568, + 482, + 657 + ], + "type": "inline_equation", + "content": "(w / o\\text{light})" + }, + { + "bbox": [ + 130, + 568, + 482, + 657 + ], + "type": "text", + "content": " for illustrative purposes. Additionally, we showcase that our method applies to both (b) small and (c) big objects, as well as (d) different placing locations. The texture of the television, for example, adjusts the texture to match its surroundings." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "type": "text", + "content": "Scene-Conditional 3D Object Stylization and Composition" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 238, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 238, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 238, + 128 + ], + "type": "text", + "content": "3.4 Lighting Model" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 136, + 482, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 136, + 482, + 292 + ], + "spans": [ + { + "bbox": [ + 130, + 136, + 482, + 292 + ], + "type": "text", + "content": "To seamlessly integrate the object into the scene, we estimate the lighting conditions, as outlined in Fig. 4. To light the object, we utilize a high dynamic range (HDR) environment map " + }, + { + "bbox": [ + 130, + 136, + 482, + 292 + ], + "type": "inline_equation", + "content": "\\mathcal{E} \\in \\mathbb{R}^{H_e \\times W_e \\times 3}" + }, + { + "bbox": [ + 130, + 136, + 482, + 292 + ], + "type": "text", + "content": ", which is well-suited for representing natural illumination. Since the given 2D scene image only captures a small angle of the full 360-degree environment map, we first use the image as input to estimate a low dynamic range (LDR) environment map. This process depends on the scene type (i.e., indoor vs. outdoor) and is explained in the following paragraphs. Given the LDR map, we then convert it to an HDR map by estimating a light scale (i.e., a scalar) for each bright region, which is thresholded from the LDR map directly (e.g., where intensity " + }, + { + "bbox": [ + 130, + 136, + 482, + 292 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{i,j} \\geq 0.8" + }, + { + "bbox": [ + 130, + 136, + 482, + 292 + ], + "type": "text", + "content": "). This scalar is then multiplied by the RGB values to approximate an HDR map. Note that we use the HDR map only to light the object. It thus does not need to be of high accuracy to make the object look plausible." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 304, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 482, + 437 + ], + "type": "text", + "content": "Indoor Scenes. The well-bounded 2D visual geometry of indoor scenes is primarily characterized by enclosed spaces and interconnected planes, facilitating straightforward (rough) 3D reconstruction. To simulate the physical lighting of indoor scenes, we lift each pixel from the 2D scene to 3D space using its estimated depth " + }, + { + "bbox": [ + 130, + 304, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 304, + 482, + 437 + ], + "type": "text", + "content": " [13, 44]. We perform world-to-camera and spherical-to-equirectangular coordinate transformations in succession, unwrapping 3D points into a latitude-longitude formatted LDR map. In post-processing, we automatically fill holes and remove small isolated regions. We fill the unseen region (i.e. areas behind the camera) with the average RGB value of the 2D scene. Two scalars are optimized: one for the far light region (" + }, + { + "bbox": [ + 130, + 304, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\triangleq \\{i,j \\mid \\mathcal{D}_{i,j} \\geq \\tau_d \\wedge \\mathcal{I}_{i,j} \\geq \\tau_f\\}" + }, + { + "bbox": [ + 130, + 304, + 482, + 437 + ], + "type": "text", + "content": ") and another for the near light region (" + }, + { + "bbox": [ + 130, + 304, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\triangleq \\{i,j \\mid \\mathcal{D}_{i,j} < \\tau_d \\wedge \\mathcal{I}_{i,j} \\geq \\tau_n\\}" + }, + { + "bbox": [ + 130, + 304, + 482, + 437 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 448, + 482, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 448, + 482, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 448, + 482, + 579 + ], + "type": "text", + "content": "Outdoor Scenes. 3D reconstruction of outdoor scenes is challenging due to their inherently unbounded geometric structure. Furthermore, in most outdoor scenes, the sun is the dominant light source, and the position change of an object has a negligible effect on the angle shift of the light. Therefore, instead of painstakingly recovering location-specific viewpoints, we use a single-view to panorama outpainting model [2], taking its generated output as the initial LDR map. We optimize one scalar for the far light area with positive elevation " + }, + { + "bbox": [ + 130, + 448, + 482, + 579 + ], + "type": "inline_equation", + "content": "(\\bigtriangleup \\{i,j\\mid \\frac{i}{H_c} < 0.5\\vee \\mathcal{I}_{i,j}\\geq \\tau_o\\})" + }, + { + "bbox": [ + 130, + 448, + 482, + 579 + ], + "type": "text", + "content": ", leaving the near light area " + }, + { + "bbox": [ + 130, + 448, + 482, + 579 + ], + "type": "inline_equation", + "content": "(\\bigtriangleup \\varnothing)" + }, + { + "bbox": [ + 130, + 448, + 482, + 579 + ], + "type": "text", + "content": " unoptimized. While more sophisticated designs, such as a position estimator [27] or sky modeling [22], could be incorporated, we find our approach robust enough to provide a sufficient estimation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 597, + 299, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 597, + 299, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 597, + 299, + 609 + ], + "type": "text", + "content": "3.5 Light-Capturing Apparatus" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "To estimate light scales, we place 3D objects with a fixed appearance model into the 2D scene and then optimize the lighting with VSD as described in Eq. (1) to anchor the lighting impact imposed by the scene. Inspired by the traditional inverse rendering setup [11] where the environment map can almost be perfectly" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "J. Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 115, + 262, + 211 + ], + "blocks": [ + { + "bbox": [ + 136, + 115, + 262, + 211 + ], + "lines": [ + { + "bbox": [ + 136, + 115, + 262, + 211 + ], + "spans": [ + { + "bbox": [ + 136, + 115, + 262, + 211 + ], + "type": "image", + "image_path": "17521f961a6554fcbf31cb7c62e6723b50a858eeee7ecbfb620b9a99635b6f06.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 215, + 482, + 281 + ], + "lines": [ + { + "bbox": [ + 130, + 215, + 482, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 215, + 482, + 281 + ], + "type": "text", + "content": "Fig. 6: Qualitative Comparison. We compare our the generated global view with the scene-object compositing method: CDC [18] and the local view with mesh texturing methods: Prolific Dreamer [55], Fantasia3D [7], and TEXTure [45]. s.gd. denotes similarity guidance. g.gd. and r.f.i denotes global-view guidance and reference feature injection, respectively. Our method achieves both seamless scene-object composition as well as texture adaptation of high fidelity. See sup. mat. for extensive comparison." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 263, + 120, + 306, + 207 + ], + "blocks": [ + { + "bbox": [ + 263, + 120, + 306, + 207 + ], + "lines": [ + { + "bbox": [ + 263, + 120, + 306, + 207 + ], + "spans": [ + { + "bbox": [ + 263, + 120, + 306, + 207 + ], + "type": "image", + "image_path": "6b283a43f3b39b311ba05d1c1b0642bb08972f66ad04ebc31a69f7cb1890d016.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 307, + 116, + 473, + 211 + ], + "blocks": [ + { + "bbox": [ + 307, + 116, + 473, + 211 + ], + "lines": [ + { + "bbox": [ + 307, + 116, + 473, + 211 + ], + "spans": [ + { + "bbox": [ + 307, + 116, + 473, + 211 + ], + "type": "image", + "image_path": "6ef79b9c0fc85302e909a028449843bbc17886c21fcfc4dbe31a8fd51a9c9120.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 289, + 482, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 410 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 410 + ], + "type": "text", + "content": "reconstructed from objects' reflections, we introduce a novel concept incorporating a virtual light-capturing apparatus alongside the object of interest during the optimization process. We insert a white sphere made of a smooth diffuse material into the scene and guide the lighting optimization with the diffusion model using the text prompt \"A gigantic diffuse white (spray-painted) sphere (ball)\". The white diffuse sphere proves advantageous in stabilizing lighting estimations for scenes with potentially strong light sources, as the reflected intensity from a white diffuse object closely approximates the intensity of the environmental lighting. In our initial experiments, we also tested spheres with mirrored and matte silver materials but found them less beneficial." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 426, + 301, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 426, + 301, + 438 + ], + "spans": [ + { + "bbox": [ + 132, + 426, + 301, + 438 + ], + "type": "text", + "content": "3.6 Light-Conditional Guidance" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 445, + 482, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 445, + 482, + 589 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 482, + 589 + ], + "type": "text", + "content": "The absence of light conditions in text prompts makes the generated images only loosely constrained on lighting, which makes it challenging for diffusion models to score the renderings accurately. A white sphere looks gray in a darker environment, so light-agnostic prompts can engender an overestimation of the brightness for dim scenes. Therefore, we append \"in a dark environment\" if the average intensity of the background and light areas are below certain thresholds. We condition the LoRA-tuned model " + }, + { + "bbox": [ + 130, + 445, + 482, + 589 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\psi}" + }, + { + "bbox": [ + 130, + 445, + 482, + 589 + ], + "type": "text", + "content": " by concatenating the light scales with the camera extrinsic [55] as class embeddings for the U-Net. This improves light estimation, especially for outdoor scenes and dimly lit scenes. For scenes with atypical lighting that significantly alters the appearance of objects, we manually append color prompts such as \"in blue tint\" for the seabed bathed in blue light, or \"in red illumination\" for a nightclub enveloped in red atmospheric light." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 605, + 230, + 619 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 605, + 230, + 619 + ], + "spans": [ + { + "bbox": [ + 132, + 605, + 230, + 619 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "In this section, we show extensive qualitative and quantitative results of our method in various settings. Please see the appendix for many more examples, animations and discussion on limitations and negative impact." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "type": "text", + "content": "Scene-Conditional 3D Object Stylization and Composition" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 307, + 179 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 307, + 179 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 307, + 179 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 307, + 179 + ], + "type": "image", + "image_path": "c0e7e34a89e1d7b3727f0ce0401f65198ca80b4fa9ea161895a8ea9d32147c2c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 185, + 482, + 220 + ], + "lines": [ + { + "bbox": [ + 130, + 185, + 482, + 220 + ], + "spans": [ + { + "bbox": [ + 130, + 185, + 482, + 220 + ], + "type": "text", + "content": "Fig. 7: Ablation on Texture Adaptation. We ablate global-view guidance " + }, + { + "bbox": [ + 130, + 185, + 482, + 220 + ], + "type": "inline_equation", + "content": "(g.gd.)" + }, + { + "bbox": [ + 130, + 185, + 482, + 220 + ], + "type": "text", + "content": " and reference feature injection " + }, + { + "bbox": [ + 130, + 185, + 482, + 220 + ], + "type": "inline_equation", + "content": "(r.f.i)" + }, + { + "bbox": [ + 130, + 185, + 482, + 220 + ], + "type": "text", + "content": ". -ctrl. and " + }, + { + "bbox": [ + 130, + 185, + 482, + 220 + ], + "type": "inline_equation", + "content": "+\\mathrm{ctrl}" + }, + { + "bbox": [ + 130, + 185, + 482, + 220 + ], + "type": "text", + "content": " denotes less or more control in feature injection (smaller or higher " + }, + { + "bbox": [ + 130, + 185, + 482, + 220 + ], + "type": "inline_equation", + "content": "s_{\\mathrm{f}}" + }, + { + "bbox": [ + 130, + 185, + 482, + 220 + ], + "type": "text", + "content": " & " + }, + { + "bbox": [ + 130, + 185, + 482, + 220 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 185, + 482, + 220 + ], + "type": "text", + "content": ") for Identity Preservation requirement." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 312, + 125, + 480, + 181 + ], + "blocks": [ + { + "bbox": [ + 312, + 125, + 480, + 181 + ], + "lines": [ + { + "bbox": [ + 312, + 125, + 480, + 181 + ], + "spans": [ + { + "bbox": [ + 312, + 125, + 480, + 181 + ], + "type": "image", + "image_path": "622d9c7aeddf1d5a90381c5e85d951827bdf5b99bdd75c63928d963b80dbde80.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 231, + 274, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 231, + 274, + 243 + ], + "spans": [ + { + "bbox": [ + 132, + 231, + 274, + 243 + ], + "type": "text", + "content": "4.1 Qualitative Evaluation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 258, + 479, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 479, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 479, + 295 + ], + "type": "text", + "content": "We first present visual results and a case study to demonstrate the generalizability of the proposed method. We then compare our method with others that specifically focus on either scene composition or object stylization." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 313, + 482, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 313, + 482, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 313, + 482, + 422 + ], + "type": "text", + "content": "Main Results. To demonstrate the generalizability of the framework and the impact of different scenes on the inserted objects, we present visual results in Fig. 5. Our method successfully blends the objects into various environments, achieving photorealistic adaptation for both appearance and lighting. This includes scenarios drawn from both real-world and fantasy settings. For example, the dimly-lit umbrella in Fig. 5 casts a shadow that is perfectly aligned with its handle and the light direction of the scene, substantially increasing the photorealism of the composition; The Belweder TV set is wrapped with gray mud consistent with the scene while its original structural details get maintained." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 439, + 482, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 439, + 482, + 563 + ], + "spans": [ + { + "bbox": [ + 130, + 439, + 482, + 563 + ], + "type": "text", + "content": "Comparison on Scene Composition. We showcase the result from CDC [18] using LLM-prompted text prompts for a fair comparison. As shown in Fig. 6, CDC loses the structural details of the object while providing limited adaptation for texture transfer. Moreover, the composed scene appears artificial and synthetic, giving the impression that the sofa is floating above the mud. Our experiments reveal that the CDC often leads to object removal, particularly when the re-painting [31] is employed. Therefore, we further incorporate similarity guidance similar to the classifier guidance [12] by adding " + }, + { + "bbox": [ + 130, + 439, + 482, + 563 + ], + "type": "inline_equation", + "content": "\\sqrt{1 - \\alpha_t} \\cdot \\frac{\\partial (||\\pmb{x}_0^r - \\hat{\\pmb{x}}_\\phi ||^2)}{\\partial \\pmb{x}_t}" + }, + { + "bbox": [ + 130, + 439, + 482, + 563 + ], + "type": "text", + "content": " to the predicted gradient " + }, + { + "bbox": [ + 130, + 439, + 482, + 563 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}(\\pmb{x}_t)" + }, + { + "bbox": [ + 130, + 439, + 482, + 563 + ], + "type": "text", + "content": " each denoising step referred as CDC w/s.gd., but find it provides limited improvement." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "Comparison on Object Stylization. We showcase results from mesh texturing method [7, 45, 55] using our LLM-prompted text prompts for a fair comparison. As shown in Fig. 6, although achieving realistic texture adaptation, these mesh texturing methods fall short in preserving the original structural details of the object and in utilizing scene-level information. Essentially, These methods only satisfy the Environmental Influence requirement via text prompts and do not meet all the criteria of our intended goal." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "text", + "content": "J. Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 157, + 315, + 211 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 2: Quantitative Evaluation. (a) We measure LPIPS for controllability, " + }, + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{dir}}" + }, + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{SI}}" + }, + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "type": "text", + "content": " for composability. (b) We present the percentages of users who prefer our method over other methods: CDC [18] and InstructP2P [3]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 157, + 315, + 211 + ], + "lines": [ + { + "bbox": [ + 133, + 157, + 315, + 211 + ], + "spans": [ + { + "bbox": [ + 133, + 157, + 315, + 211 + ], + "type": "table", + "html": "
MethodLPIPS ↓CLIPdir ↑CLIPSI ↑
InstructP2P [3]0.00200.01750.1852
Ours (T.A.)0.00140.02010.2006
CDC [18]0.00530.00420.1996
Ours (T.A. + L.E.)0.00150.02060.2013
", + "image_path": "b74ac5b0c11eba0c7a91346562d1cf1238f02048e2fcea86ad685cd3894bc429.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 187, + 212, + 257, + 220 + ], + "lines": [ + { + "bbox": [ + 187, + 212, + 257, + 220 + ], + "spans": [ + { + "bbox": [ + 187, + 212, + 257, + 220 + ], + "type": "text", + "content": "(a) Visual Metrics." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 318, + 165, + 480, + 202 + ], + "blocks": [ + { + "bbox": [ + 318, + 165, + 480, + 202 + ], + "lines": [ + { + "bbox": [ + 318, + 165, + 480, + 202 + ], + "spans": [ + { + "bbox": [ + 318, + 165, + 480, + 202 + ], + "type": "table", + "html": "
OursDrawOther
Ours vs. CDC [18]71%20%9%
Ours vs. InstructP2P [3]74%6%20%
", + "image_path": "a34a5791a67df410e9e061b5a04f11891674703dff7790493bd3e1d38a60354e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 331, + 203, + 463, + 213 + ], + "lines": [ + { + "bbox": [ + 331, + 203, + 463, + 213 + ], + "spans": [ + { + "bbox": [ + 331, + 203, + 463, + 213 + ], + "type": "text", + "content": "(b) User study for human preference." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 226, + 281, + 238 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 226, + 281, + 238 + ], + "spans": [ + { + "bbox": [ + 132, + 226, + 281, + 238 + ], + "type": "text", + "content": "4.2 Quantitative Evaluation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 245, + 479, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 245, + 479, + 329 + ], + "spans": [ + { + "bbox": [ + 130, + 245, + 479, + 329 + ], + "type": "text", + "content": "We collected 34 composites between diverse objects and scenes, each with 3 random views, yielding a total of 102 images. Besides CDC [18], we compare with InstructPix2Pix [3] by directly copy-pasting the object with adapted texture onto the scene. We measure LPIPS for controllability, " + }, + { + "bbox": [ + 130, + 245, + 479, + 329 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{dir}}" + }, + { + "bbox": [ + 130, + 245, + 479, + 329 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 245, + 479, + 329 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{SI}}" + }, + { + "bbox": [ + 130, + 245, + 479, + 329 + ], + "type": "text", + "content": " for composability, similar to CDC, and report the results in Tab. 2. Compared with InstructPix2Pix our method for texture adaptation (T.A) alone excels in controllability and composability. Light estimation (L.E.) further improves results." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 330, + 480, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 330, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 330, + 480, + 365 + ], + "type": "text", + "content": "We also conducted a user study in Tab. 2 with 42 participants to compare the realism of our compositing under the condition that the object identity must be preserved. Our method is preferred over prior work in over " + }, + { + "bbox": [ + 130, + 330, + 480, + 365 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 130, + 330, + 480, + 365 + ], + "type": "text", + "content": " of the cases." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 383, + 238, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 383, + 238, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 383, + 238, + 396 + ], + "type": "text", + "content": "4.3 Ablation Study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 403, + 482, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 498 + ], + "type": "text", + "content": "Identity Preservation. To illustrate how the Identity Preservation requirement is satisfied by reference feature injection (r.f.i.), we showcase the leather sofa in the swamp with less and more control (Fig. 7). The setup with less control drifts to a visually different sofa with too much mud wrapped around while more control generates a less muddy and adapted sofa compared to the reference image. Both adapted textures are less visually plausible. Additionally, when comparing Ours w/o g.gd. & r.f.i. to the reference image, it is evident that the identity of the sofa is lost." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 510, + 482, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 510, + 482, + 584 + ], + "spans": [ + { + "bbox": [ + 131, + 510, + 482, + 584 + ], + "type": "text", + "content": "Global-View Guidance. To showcase the necessity of global-view guidance we conduct texture adaptation without it as Ours w/o g.gd. in Fig. 7. The adapted texture without seeing the full scene fails to adapt the color of the mud to that of the scene image, contrasting starkly with the scene. Experiments ablating the optimal global-view guidance set-up (e.g., solid-color background augmentation, usage of inpainting model, etc.) are detailed in the appendix." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 594, + 482, + 666 + ], + "type": "text", + "content": "Light Estimation. In Fig. 1, the combination of accurately estimated lighting and shadows significantly enhances the photorealism of the composed image (compare the image on the left, which includes lighting, to those in the top right corner, which are without lighting). To further test the validity of the estimated light, we render images omitting the texture adaptation step. Fig. 8 demonstrates that the estimated light considers multiple factors, including its" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "type": "text", + "content": "Scene-Conditional 3D Object Stylization and Composition" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 167, + 114, + 444, + 392 + ], + "blocks": [ + { + "bbox": [ + 167, + 114, + 444, + 392 + ], + "lines": [ + { + "bbox": [ + 167, + 114, + 444, + 392 + ], + "spans": [ + { + "bbox": [ + 167, + 114, + 444, + 392 + ], + "type": "image", + "image_path": "55d0f2c0196aa81bbf235a891b65a25da4524bad112f415ffa64c0df7d6f975b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 400, + 482, + 445 + ], + "lines": [ + { + "bbox": [ + 130, + 400, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 400, + 482, + 445 + ], + "type": "text", + "content": "Fig. 8: Ablation on Light Estimation. We evaluate the effectiveness of light estimation by demonstrating the lighting effects on a sofa in various indoor and outdoor scenes. Overlaid in the left corners are objects rendered with the estimated light against a white background. In the right corners are objects before the light estimation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 458, + 482, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 458, + 482, + 507 + ], + "spans": [ + { + "bbox": [ + 130, + 458, + 482, + 507 + ], + "type": "text", + "content": "direction relative to the specific locations of objects in indoor scenes and its impact on the appearance of objects under atmospheric colored lighting conditions. Importantly, the estimated HDR maps, paired with the proposed scene composition, facilitate the realistic integration of shadows." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 535, + 220, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 535, + 220, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 535, + 220, + 547 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we are motivated by a practical goal—what if a 3D object is placed into a 2D scene?—and propose a novel framework that allows (1) stylizing the object with an adapted texture that aligns with the given scene; and (2) achieving a photorealistic scene composition with the aid of estimated light of the environment. This is enabled by physically-based differentiable ray tracing and diffusion models used as guidance. Our method allows for automatic adaptation and blending of existing objects into a variety of scenes, making it useful for both 2D and 3D downstream applications, such as visual media and video games." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "text", + "content": "J. Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 163 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 163 + ], + "type": "text", + "content": "Acknowledgement. This work is supported by the UKRI grant: Turing AI Fellowship EP/W002981/1 and EPSRC/MURI grant: EP/N019474/1. J. Zhou is also supported by the Horizon Robotics. We thank Luke Melas-Kyriazi and Fabio Pizzati for their helpful discussions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 182, + 197, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 197, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 197, + 194 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 207, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 138, + 207, + 349, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 207, + 349, + 217 + ], + "spans": [ + { + "bbox": [ + 138, + 207, + 349, + 217 + ], + "type": "text", + "content": "1. https://polyhaven.com, https://polyhaven.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 218, + 481, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 218, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 138, + 218, + 481, + 239 + ], + "type": "text", + "content": "2. Akimoto, N., Matsuo, Y., Aoki, Y.: Diverse plausible 360-degree image outpainting for efficient 3dgc background creation. In: CVPR (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 240, + 481, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 240, + 481, + 261 + ], + "spans": [ + { + "bbox": [ + 138, + 240, + 481, + 261 + ], + "type": "text", + "content": "3. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: CVPR (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 262, + 481, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 262, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 138, + 262, + 481, + 282 + ], + "type": "text", + "content": "4. Cao, T., Kreis, K., Fidler, S., Sharp, N., Yin, K.: Texfusion: Synthesizing 3d textures with text-guided image diffusion models. In: ICCV (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 284, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 284, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 138, + 284, + 481, + 304 + ], + "type": "text", + "content": "5. Chen, B.C., Kae, A.: Toward realistic image compositing with adversarial learning. In: CVPR (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 305, + 481, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 305, + 481, + 337 + ], + "spans": [ + { + "bbox": [ + 138, + 305, + 481, + 337 + ], + "type": "text", + "content": "6. Chen, D.Z., Siddiqui, Y., Lee, H.Y., Tulyakov, S., Nießner, M.: Text2tex: Text-driven texture synthesis via diffusion models. arXiv preprint arXiv:2303.11396 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 338, + 481, + 359 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 338, + 481, + 359 + ], + "spans": [ + { + "bbox": [ + 138, + 338, + 481, + 359 + ], + "type": "text", + "content": "7. Chen, R., Chen, Y., Jiao, N., Jia, K.: Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In: ICCV (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 361, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 361, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 138, + 361, + 481, + 380 + ], + "type": "text", + "content": "8. Cong, W., Niu, L., Zhang, J., Liang, J., Zhang, L.: Bargainnet: Background-guided domain translation for image harmonization. In: ICME (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 382, + 481, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 382, + 481, + 403 + ], + "spans": [ + { + "bbox": [ + 138, + 382, + 481, + 403 + ], + "type": "text", + "content": "9. Cong, W., Zhang, J., Niu, L., Liu, L., Ling, Z., Li, W., Zhang, L.: Dovenet: Deep image harmonization via domain verification. In: CVPR (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 404, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 404, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 138, + 404, + 481, + 435 + ], + "type": "text", + "content": "0. Deitke, M., Schwenk, D., Salvador, J., Weihs, L., Michel, O., VanderBilt, E., Schmidt, L., Ehsani, K., Kembhavi, A., Farhadi, A.: Objaverse: A universe of annotated 3d objects. In: CVPR (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 437, + 481, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 437, + 481, + 457 + ], + "spans": [ + { + "bbox": [ + 138, + 437, + 481, + 457 + ], + "type": "text", + "content": "1. Deng, Y., Li, X., Liu, S., Yang, M.H.: Dip: Differentiable interreflection-aware physics-based inverse rendering. arXiv preprint arXiv:2212.04705 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 459, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 459, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 138, + 459, + 481, + 479 + ], + "type": "text", + "content": "2. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. NeurIPS (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 480, + 481, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 480, + 481, + 502 + ], + "spans": [ + { + "bbox": [ + 138, + 480, + 481, + 502 + ], + "type": "text", + "content": "3. Eftekhar, A., Sax, A., Malik, J., Zamir, A.: Omnidata: A scalable pipeline for making multi-task mid-level vision datasets from 3d scans. In: ICCV (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 502, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 502, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 502, + 481, + 533 + ], + "type": "text", + "content": "4. Gao, J., Shen, T., Wang, Z., Chen, W., Yin, K., Li, D., Litany, O., Gojcic, Z., Fidler, S.: Get3d: A generative model of high quality 3d textured shapes learned from images. NeurIPS (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 535, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 535, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 535, + 481, + 555 + ], + "type": "text", + "content": "5. Gardner, M.A., Hold-Geoffroy, Y., Sunkavalli, K., Gagné, C., Lalonde, J.F.: Deep parametric indoor lighting estimation. In: ICCV (2019)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 557, + 481, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 557, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 138, + 557, + 481, + 588 + ], + "type": "text", + "content": "6. Gardner, M.A., Sunkavalli, K., Yumer, E., Shen, X., Gambaretto, E., Gagné, C., Lalonde, J.F.: Learning to predict indoor illumination from a single image. arXiv preprint arXiv:1704.00090 (2017)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 590, + 481, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 590, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 138, + 590, + 481, + 610 + ], + "type": "text", + "content": "7. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. NeurIPS (2014)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 612, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 612, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 612, + 481, + 643 + ], + "type": "text", + "content": "8. Hachnochi, R., Zhao, M., Orzech, N., Gal, R., Mahdavi-Amiri, A., Cohen-Or, D., Bermano, A.H.: Cross-domain compositing with pretrained diffusion models. arXiv preprint arXiv:2302.10167 (2023)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 138, + 644, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 644, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 644, + 481, + 665 + ], + "type": "text", + "content": "9. Haque, A., Tancik, M., Efros, A., Holynski, A., Kanazawa, A.: Instruct-nerf2nerf: Editing 3d scenes with instructions. In: CVPR (2023)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "type": "text", + "content": "Scene-Conditional 3D Object Stylization and Composition" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 139 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 139 + ], + "type": "text", + "content": "20. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. NeurIPS (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 446, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 446, + 150 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 446, + 150 + ], + "type": "text", + "content": "21. Ho, J., Salimans, T.: Classifier-free diffusion guidance. NeurIPSW (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 152, + 480, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 152, + 480, + 175 + ], + "spans": [ + { + "bbox": [ + 132, + 152, + 480, + 175 + ], + "type": "text", + "content": "22. Hold-Geoffroy, Y., Athawale, A., Lalonde, J.F.: Deep sky modeling for single image outdoor lighting estimation. In: CVPR (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 175, + 480, + 197 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 175, + 480, + 197 + ], + "spans": [ + { + "bbox": [ + 132, + 175, + 480, + 197 + ], + "type": "text", + "content": "23. Hold-Geoffroy, Y., Sunkavalli, K., Hadap, S., Gambaretto, E., Lalonde, J.F.: Deep outdoor illumination estimation. In: CVPR (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 198, + 480, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 198, + 480, + 220 + ], + "spans": [ + { + "bbox": [ + 132, + 198, + 480, + 220 + ], + "type": "text", + "content": "24. Jakob, W., Speierer, S., Roussel, N., Vicini, D.: Dr.jit: A just-in-time compiler for differentiable rendering. TOG (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 221, + 480, + 254 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 221, + 480, + 254 + ], + "spans": [ + { + "bbox": [ + 132, + 221, + 480, + 254 + ], + "type": "text", + "content": "25. Jiang, Y., Zhang, H., Zhang, J., Wang, Y., Lin, Z., Sunkavalli, K., Chen, S., Amirghodsi, S., Kong, S., Wang, Z.: Ssh: A self-supervised framework for image harmonization. In: ICCV (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 255, + 480, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 255, + 480, + 288 + ], + "spans": [ + { + "bbox": [ + 132, + 255, + 480, + 288 + ], + "type": "text", + "content": "26. Kamata, H., Sakuma, Y., Hayakawa, A., Ishii, M., Narihira, T.: Instruct 3d-to-3d: Text instruction guided 3d-to-3d conversion. arXiv preprint arXiv:2303.15780 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 289, + 480, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 480, + 311 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 480, + 311 + ], + "type": "text", + "content": "27. Karimi Dastjerdi, M.R., Eisenmann, J., Hold-Geoffroy, Y., Lalonde, J.F.: Everlight: Indoor-outdoor editable hdr lighting estimation. In: ICCV (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 312, + 480, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 312, + 480, + 345 + ], + "spans": [ + { + "bbox": [ + 132, + 312, + 480, + 345 + ], + "type": "text", + "content": "28. Karimi Dastjerdi, M.R., Hold-Geoffroy, Y., Eisenmann, J., Khodadadeh, S., Lalonde, J.F.: Guided co-modulated GAN for 360 field of view extrapolation. 3DV (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 346, + 480, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 346, + 480, + 379 + ], + "spans": [ + { + "bbox": [ + 132, + 346, + 480, + 379 + ], + "type": "text", + "content": "29. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3d: High-resolution text-to-3d content creation. In: CVPR (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "text", + "content": "30. Liu, M., Xu, C., Jin, H., Chen, L., Xu, Z., Su, H., et al.: One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. arXiv preprint arXiv:2306.16928 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 414, + 480, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 480, + 436 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 480, + 436 + ], + "type": "text", + "content": "31. Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Van Gool, L.: Repaint: Inpainting using denoising diffusion probabilistic models. In: CVPR (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 437, + 480, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 437, + 480, + 469 + ], + "spans": [ + { + "bbox": [ + 132, + 437, + 480, + 469 + ], + "type": "text", + "content": "32. McAuley, S., Hill, S., Hoffman, N., Gotanda, Y., Smits, B., Burley, B., Martinez, A.: Practical physically-based shading in film and game production. In: SIGGRAPH (2012)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 471, + 480, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 471, + 480, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 471, + 480, + 493 + ], + "type": "text", + "content": "33. Metzer, G., Richardson, E., Patashnik, O., Giryes, R., Cohen-Or, D.: Latent-nerf for shape-guided generation of 3d shapes and textures. In: CVPR (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 494, + 480, + 516 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 494, + 480, + 516 + ], + "spans": [ + { + "bbox": [ + 132, + 494, + 480, + 516 + ], + "type": "text", + "content": "34. Michel, O., Bar-On, R., Liu, R., Benaim, S., Hanocka, R.: Text2mesh: Text-driven neural stylization for meshes (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 517, + 480, + 550 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 517, + 480, + 550 + ], + "spans": [ + { + "bbox": [ + 132, + 517, + 480, + 550 + ], + "type": "text", + "content": "35. Mohammad Khalid, N., Xie, T., Belilovsky, E., Popa, T.: Clip-mesh: Generating textured meshes from text using pretrained image-text models. In: SIGGRAPH Asia (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 551, + 480, + 584 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 551, + 480, + 584 + ], + "spans": [ + { + "bbox": [ + 132, + 551, + 480, + 584 + ], + "type": "text", + "content": "36. Munkberg, J., Hasselgren, J., Shen, T., Gao, J., Chen, W., Evans, A., Müller, T., Fidler, S.: Extracting triangular 3d models, materials, and lighting from images. In: CVPR (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 585, + 480, + 607 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 585, + 480, + 607 + ], + "spans": [ + { + "bbox": [ + 132, + 585, + 480, + 607 + ], + "type": "text", + "content": "37. Oechsle, M., Mescheder, L., Niemeyer, M., Strauss, T., Geiger, A.: Texture fields: Learning texture representations in function space. In: ICCV (2019)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 608, + 307, + 619 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 608, + 307, + 619 + ], + "spans": [ + { + "bbox": [ + 132, + 608, + 307, + 619 + ], + "type": "text", + "content": "38. OpenAI: Gpt-4 technical report (2023)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 620, + 480, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 620, + 480, + 642 + ], + "spans": [ + { + "bbox": [ + 132, + 620, + 480, + 642 + ], + "type": "text", + "content": "39. Parmar, G., Kumar Singh, K., Zhang, R., Li, Y., Lu, J., Zhu, J.Y.: Zero-shot image-to-image translation. In: SIGGRAPH (2023)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "40. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. ICLR (2023)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "J. Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 588 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "text", + "content": "41. Qian, G., Mai, J., Hamdi, A., Ren, J., Siarohin, A., Li, B., Lee, H.Y., Skorokhodov, I., Wonka, P., Tulyakov, S., Ghanem, B.: Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. arXiv preprint arXiv:2306.17843 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 161, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 161, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 161, + 481, + 194 + ], + "type": "text", + "content": "42. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "text", + "content": "43. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: ICML (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 247 + ], + "type": "text", + "content": "44. Ranftl, R., Lasinger, K., Hafner, D., Schindler, K., Koltun, V.: Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. PAMI (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "type": "text", + "content": "45. Richardson, E., Metzer, G., Alaluf, Y., Giryes, R., Cohen-Or, D.: Texture: Text-guided texturing of 3d shapes. SIGGRAPH (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 271, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 481, + 293 + ], + "type": "text", + "content": "46. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 325 + ], + "type": "text", + "content": "47. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 326, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 346 + ], + "type": "text", + "content": "48. Siddiqui, Y., Thies, J., Ma, F., Shan, Q., Nießner, M., Dai, A.: Texturify: Generating textures on 3d shape surfaces. In: ECCV (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 348, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 348, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 348, + 481, + 380 + ], + "type": "text", + "content": "49. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. In: ICLR (2021), https://openreview.net/forum?id=PxTIG12RRHS" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 380, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 402 + ], + "type": "text", + "content": "50. Song, Y., Zhang, Z., Lin, Z., Cohen, S., Price, B., Zhang, J., Kim, S.Y., Aliaga, D.: Objectstitch: Generative object compositing. CVPR (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 403, + 481, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 481, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 481, + 424 + ], + "type": "text", + "content": "51. Tsalicoglou, C., Manhardt, F., Tonioni, A., Niemeyer, M., Tombari, F.: Textmesh: Generation of realistic 3d meshes from text prompts (2024)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 425, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 425, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 425, + 481, + 445 + ], + "type": "text", + "content": "52. Tumanyan, N., Geyer, M., Bagon, S., Dekel, T.: Plug-and-play diffusion features for text-driven image-to-image translation. In: CVPR (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 447, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 481, + 468 + ], + "type": "text", + "content": "53. Walter, B., Marschner, S.R., Li, H., Torrance, K.E.: Microfacet models for refraction through rough surfaces. In: EGSR (2007)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 468, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 490 + ], + "type": "text", + "content": "54. Wang, H., Du, X., Li, J., Yeh, R.A., Shakhnarovich, G.: Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In: CVPR (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 491, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 481, + 522 + ], + "type": "text", + "content": "55. Wang, Z., Lu, C., Wang, Y., Bao, F., Li, C., Su, H., Zhu, J.: Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. NeurIPS (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 523, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 481, + 544 + ], + "type": "text", + "content": "56. Wei, L.Y., Lefebvre, S., Kwatra, V., Turk, G.: State of the art in example-based texture synthesis. EG-STAR (2009)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 545, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 481, + 567 + ], + "type": "text", + "content": "57. Xiao, J., Ehinger, K.A., Oliva, A., Torralba, A.: Recognizing scene viewpoint using panoramic place representation. In: ICCV (2012)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 567, + 481, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 588 + ], + "type": "text", + "content": "58. Zhang, J., Sunkavalli, K., Hold-Geoffroy, Y., Hadap, S., Eisenman, J., Lalonde, J.F.: All-weather deep outdoor lighting estimation. In: CVPR (2019)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "type": "text", + "content": "Scene-Conditional 3D Object Stylization and Composition" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Scene-Graph ViT_ End-to-End Open-Vocabulary Visual Relationship Detection/07085046-a430-402b-be81-31056599f4f5_content_list.json b/2024/Scene-Graph ViT_ End-to-End Open-Vocabulary Visual Relationship Detection/07085046-a430-402b-be81-31056599f4f5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0ff6fefe28f35f49da70fd3e59f8c7103e2976e2 --- /dev/null +++ b/2024/Scene-Graph ViT_ End-to-End Open-Vocabulary Visual Relationship Detection/07085046-a430-402b-be81-31056599f4f5_content_list.json @@ -0,0 +1,1753 @@ +[ + { + "type": "text", + "text": "Scene-Graph ViT: End-to-End Open-Vocabulary Visual Relationship Detection", + "text_level": 1, + "bbox": [ + 223, + 140, + 779, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tim Salzmann $^{1,2}$ , Markus Ryll $^{2}$ , Alex Bewley $^{\\star 1}$ , and Matthias Minderer $^{\\star 1}$", + "bbox": [ + 339, + 210, + 663, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Google DeepMind {tsal, bewley, mjlm}@google.com $^{2}$ Technical University Munich {tim.salzmann, markus.ryll}@tum.d", + "bbox": [ + 369, + 253, + 625, + 310 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/76dc13707de447c668c91d07b4346bd00fc74461a2918d44f1240e6c8f250fb0.jpg", + "image_caption": [ + "Figure 1: Relationships detected by our method on an unseen image. The top relationships by confidence score are shown. Photo by Spacejoy on Unsplash." + ], + "image_footnote": [], + "bbox": [ + 217, + 332, + 558, + 507 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f8a13735bc1e2ecd2dd20a03f79190938d5c6a1b92da04791fe196e43dbb5792.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 560, + 330, + 787, + 507 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Visual relationship detection aims to identify objects and their relationships in images. Prior methods approach this task by adding separate relationship modules or decoders to existing object detection architectures. This separation increases complexity and hinders end-to-end training, which limits performance. We propose a simple and highly efficient decoder-free architecture for open-vocabulary visual relationship detection. Our model consists of a Transformer-based image encoder that represents objects as tokens and models their relationships implicitly. To extract relationship information, we introduce an attention mechanism that selects object pairs likely to form a relationship. We provide a single-stage recipe to train this model on a mixture of object and relationship detection data. Our approach achieves state-of-the-art relationship detection performance on Visual Genome and on the large-vocabulary GQA benchmark at real-time inference speeds. We provide ablations, real-world qualitative examples, and analyses of zero-shot performance.", + "bbox": [ + 259, + 565, + 743, + 773 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{\\dagger}$ Work done while at Google DeepMind. \n*Advising project leads in alphabetical order.", + "bbox": [ + 232, + 810, + 537, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 143, + 375, + 160 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A fundamental goal of computer vision is to decompose visual scenes into structured semantic representations. A commonly studied task towards this goal is object detection, in which objects in an image are localized by bounding boxes and classified into semantic categories. However, a scene description also includes the semantic relationships between objects. This gives rise to the task of visual relationship detection (VRD, [4,6,32,48,49,55]). In VRD, the model detects objects and infers pairwise relationships between them in the form of triplets3 [32].", + "bbox": [ + 212, + 179, + 787, + 300 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Detecting both objects and their relationships allows the construction of a scene graph [19,30,48] in which objects are represented as nodes and their relationships as edges. Scene graph generation (SGG) has wide-ranging applications in robotics [1,5,11,17,38,42] and image retrieval [19,27]. Increasingly, structured scene representations are also employed to provide grounding and explainability to multimodal large language models [15,36,59].", + "bbox": [ + 212, + 300, + 787, + 391 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Prior work typically draws a distinction between object detection and relationship prediction. Detection is performed either as a wholly separate step before relationship prediction [8, 32, 48, 52, 55, 57], or by separate model parts such as \"relationship decoders\" that are responsible for modeling the interactions between objects [7, 49, 61]. This distinction makes it hard to optimize such models end-to-end for VRD [61]. In contrast, we propose an encoder-only architecture that models objects and relationships jointly, directly in the image encoder. Our architecture performs open-vocabulary relationship detection and can be trained end-to-end on arbitrary mixtures of object detection and relationship annotations.", + "bbox": [ + 212, + 391, + 787, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our model builds on a Transformer-based encoder-only object detector [34] in which the output tokens of the image encoder directly represent object proposals. From these tokens, class embeddings and bounding boxes are decoded with lightweight heads. Our insight is that this architecture is perfectly set up to learn relationships between objects directly in the image encoder, without the need for additional relationship-specific stages. This is because the existing self-attention in the encoder already models all-to-all pairwise interactions between the object proposal tokens.", + "bbox": [ + 212, + 542, + 787, + 662 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To access information about the relationship between two of these tokens, we combine the embeddings corresponding to the and token using a new Relationship Attention layer. Obtaining relationship embeddings for all possible pairwise combinations of object proposal tokens would be computationally infeasible. To reduce the number of combinations, we introduce a self-supervised hard attention mechanism that selects the highest-confidence pairs at a computational cost comparable to a single self-attention layer. We show how to directly supervise the attention scores of this mechanism without the need to propagate gradients through the hard selection.", + "bbox": [ + 212, + 664, + 787, + 801 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "T. Salzmann et al.", + "bbox": [ + 271, + 114, + 395, + 127 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "3 We use to distinguish the grammatical terms and from generic use of the word object as in object detection.", + "bbox": [ + 217, + 810, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "An additional benefit of our design is that it disentangles object names from relationship predicates during inference. In contrast to prior open-vocabulary methods [53, 54, 61], our model can embed object and predicate texts separately and generate confidence scores efficiently for all possible combinations.", + "bbox": [ + 212, + 146, + 782, + 220 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, we make the following contributions:", + "bbox": [ + 238, + 222, + 599, + 234 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. An efficient architecture for open-vocabulary visual relationship detection.", + "2. A single-stage recipe for joint object and relationship detection training.", + "3. Efficient, disentangled object and relationship inference.", + "4. Analysis of inference speed, ablations, and qualitative examples." + ], + "bbox": [ + 220, + 243, + 772, + 301 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our method achieves state-of-the-art visual relationship detection performance on the Visual Genome dataset (26.1% mR@100 with graph constraint) and on the large-vocabulary GQA benchmark. It provides strong results while being significantly simpler than prior approaches to visual relationship detection.", + "bbox": [ + 212, + 308, + 782, + 382 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 404, + 385, + 420 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Below, we review the main approaches to VRD that are relevant to our work. For a comprehensive overview of the field, we refer to [27].", + "bbox": [ + 212, + 433, + 782, + 463 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Detector-Agnostic Relationship Detection. Historically, many works on VRD assume that object detections are given, such that the task reduces to inferring relationships between the objects. These \"detector-agnostic\" methods use off-the-shelf detectors such as Faster-RCNN [39] to obtain boxes and box embeddings and infer a scene graph from them. Early work on large-scale VRD employed word embeddings to improve relationship generalization [32] and graph inference methods such as message passing for SGG [48]. More recent methods leverage relationship co-occurrence statistics [55] and address the long-tailed nature of the training data [8, 52, 57].", + "bbox": [ + 212, + 468, + 784, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "End-to-End Relationship Detection. A fundamental limitation of the detector-agnostic approach is that object detection and relationship prediction are treated separately and cannot learn from each other. Overcoming this limitation has recently motivated the development of end-to-end architectures in which objects and relationships are predicted jointly by a single model [7, 20, 29, 61]. All of these models have in common that they use a Transformer [47] decoder to predict object and relationship embeddings. The decoder represents object proposals with query embeddings that can cross-attend into image embeddings. The literature on object detection suggests that this choice may be problematic, because the query embeddings can be difficult to initialize, which can lead to unstable and slow optimization [3, 34, 43, 50, 58, 63]. To avoid this issue, we design an encoder-only architecture in which both object and relationship embeddings are computed directly from the output tokens of a Vision Transformer [10] image encoder, building on an idea from object detection [34]. In contrast to prior models [61], this allows us to train the model end-to-end on object and relationship annotations simultaneously.", + "bbox": [ + 212, + 609, + 785, + 852 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Scene-Graph ViT", + "bbox": [ + 612, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Long-Tailed and Open-Vocabulary Relationship Detection. Object classes in the natural world follow a long-tailed distribution [12,62], and this is compounded in VRD due to the combination of subject, object, and predicate in a relationship [27,48,49]. A large number of VRD methods specifically aim to address this issue, for example with debiasing losses [52], data augmentation [57], or data resampling [8]. Additionally, while VRD models historically assumed a fixed set of object and predicate classes, recent works add open-vocabulary capabilities [13,41,54,61]. These models leverage pretrained vision-language models such as CLIP [18, 37, 51, 56] to inherit their natural-language understanding. Instead of learning a fixed set of classes, open-vocabulary models predict class embeddings that can be compared to the text embeddings of arbitrary object or predicate descriptions for classification. Here, we employ a pretrained vision-language backbone and combine it with a carefully designed, lightweight relationship detection head to preserve and transfer semantic knowledge from the backbone pretraining to VRD. This allows our model to achieve strong results on the large-vocabulary GQA benchmark [16] (Section 4.3).", + "bbox": [ + 212, + 145, + 787, + 388 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While we focus on architectural improvements, another line of work addresses the scarcity of training data for VRD. For example, RLIP [53] and RLIPv2 [54], which focus on human-object interaction [4], propose relationship-focused image-text pretraining and self-training, which yield large improvements in VRD performance. These works are orthogonal to our contributions and can be combined.", + "bbox": [ + 212, + 388, + 787, + 464 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Scene-Graph ViT", + "text_level": 1, + "bbox": [ + 215, + 486, + 423, + 505 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We propose an architecture for open-vocabulary relationship prediction in which both objects and the relationships between them are handled as \"first-class citizens\" directly in a single-stage process within the model backbone. We build on an encoder-only architecture for object detection [34] and adapt it to relationship prediction by adding a specialized attention layer. This layer exploits the pairwise structure between existing object embeddings to obtain relationship embeddings as schematized in Figure 2 and described below.", + "bbox": [ + 212, + 518, + 787, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Encoder-Only Open-Vocabulary Object Detection", + "text_level": 1, + "bbox": [ + 215, + 647, + 671, + 664 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We briefly review the detection architecture that forms the basis of our model [34]. The model consists of Transformer-based [10] image and text encoders that are contrastively pretrained on a large number of image and text pairs [37]. The image encoder is adapted to detection by removing the final pooling layer and adding heads that predict bounding boxes and class embeddings directly from the output tokens of the image encoder. For open-vocabulary object classification, the embeddings computed by the class prediction head are compared to text encoder embeddings of object descriptions. This architecture achieves strong open-vocabulary object detection performance [33, 34] and does not suffer from the training instabilities observed in some encoder-decoder detection models [3, 34, 43, 50, 58, 63].", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "T. Salzmann et al.", + "bbox": [ + 271, + 114, + 395, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/11851b4cb9f82a6929578514b366470e94ad3b87a5308ffb27f48cd4a4f5073a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 263, + 146, + 738, + 255 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f042bb5958900c574775ed1f1666401d8fbde29f3dc02abd99e8992bb68e8497.jpg", + "image_caption": [ + "Figure 2: For relationship selection, image tokens are first projected using two lightweight MLPs to produce and embeddings. A relationship score is then computed as the inner product between all and embeddings. Relationships are filtered by first selecting the top object instances, using the scores along the diagonal to represent instance likelihood. Among the remaining instances, the top pairs are selected using the off-diagonal scores. This yields a set of relationship triplets, each consisting of a index, an index, and a relationship embedding that is computed by summing the respective and embeddings. For classification, the relationship embeddings are compared against text embeddings of object class or predicate text descriptions." + ], + "image_footnote": [], + "bbox": [ + 279, + 256, + 704, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Extension to Relationship Prediction", + "text_level": 1, + "bbox": [ + 214, + 559, + 568, + 574 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the architecture described above, each image encoder token represents an object proposal that captures per-object information. Importantly, the encoder consists of self-attention layers which, by design, model all pairwise interactions between these tokens. We therefore hypothesize that information about object relationships can be learned directly in the image encoder. To extract this information in the form of triplets, we introduce two MLPs that transform vision encoder output tokens into embeddings $\\mathbf{s}_i$ and embeddings $\\mathbf{o}_j$ (Figure 2). Using different MLPs for and is crucial to break symmetry in subsequent processing (Appendix A.1).", + "bbox": [ + 212, + 583, + 787, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "An embedding representing the relationship between two object proposals (encoder tokens) can then be obtained simply by element-wise addition of their and embeddings. Obtaining relationship embeddings for all $N^2$ pairs, where $N$ is the number of object proposals, would be computationally infeasible. We therefore introduce a Relationship Attention layer that performs hard attention to select the pairs most likely to form a relationship. This layer computes an attention-like score $p_{ij} =$", + "bbox": [ + 212, + 734, + 787, + 842 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Scene-Graph ViT", + "bbox": [ + 612, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mathbf{s}_i \\mathbf{o}_j^T$ , where $\\mathbf{s}_i$ and $\\mathbf{o}_j$ are the embedding vectors of $<\\text { subject }>$ $i$ and $<\\text { object }>$ $j$ , and $p_{ij}$ represents the likelihood that a relationship between $<\\text { subject }>$ $i$ and $<\\text { object }>$ $j$ exists. By computing this score for all $<\\text { subject - object }>$ pairs, we obtain an $N \\times N$ matrix, from which we select the top $k$ pairs. For these $k$ pairs, we compute relationship embeddings $\\mathbf{r}_{ij} = \\text { LayerNorm }(\\mathbf{s}_i + \\mathbf{o}_j)$ . To classify the relationship of a pair, its relationship embedding is compared to text embeddings of relationship predicates. We use the embeddings where $<\\text { subject }>$ and $<\\text { object }>$ are identical ( $\\mathbf{r}_{i=j}$ ) to represent object instances, and predict object classes from them. Boxes are predicted from the corresponding image encoder output tokens (see Figure 2).", + "bbox": [ + 212, + 145, + 787, + 297 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The Relationship Attention layer therefore identifies object pairs to focus on for relationship classification, by computing a hard attention with embeddings as queries and embeddings as keys. Since the hard selection is not differentiable, gradients from the relationship prediction cannot be used directly to train this layer. Instead, the relationship score is self-supervised to predict the maximum class probability that will later be predicted for the relationship at the classification stage (Section 3.3).", + "bbox": [ + 212, + 297, + 787, + 404 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Training", + "text_level": 1, + "bbox": [ + 215, + 426, + 333, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The image and text encoders of our model are initialized from a vision-language model that was contrastively pretrained on a large number of image-text pairs [37]. After adding the Relationship Attention as well as the object bounding box and class prediction heads, the model is jointly trained on a mixture of object and relationship detection datasets in a single training stage (Section 4.1), using the losses described below.", + "bbox": [ + 212, + 452, + 789, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Bipartite Matching. Bipartite matching between object predictions $(\\mathbf{r}_{i = j})$ and ground-truth annotations is performed based on a cost consisting of the object classification loss and the box prediction losses as in [3]. This matching between objects also establishes a matching of predicted to ground-truth relationship predicates, since a relationship is uniquely identified by the and indices $i$ and $j$ . Unmatched predictions are trained to predict low scores for all classes and incur no box prediction loss.", + "bbox": [ + 212, + 542, + 787, + 648 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Box Prediction Losses. For bounding box regression, we use the L1 and generalized intersection-over-union (gIoU) losses described in [3].", + "bbox": [ + 212, + 648, + 787, + 680 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Object and Predicate Classification Loss. For both object category and predicate classification, we use a sigmoid cross-entropy loss as in [34]. This loss is computed between the ground truth classes and logits obtained by computing the inner product of the relationship embeddings $\\mathbf{r}_{ij}$ selected by the Relationship Attention layer with the text embeddings of the class names. For embeddings corresponding to individual objects $(\\mathbf{r}_{i = j})$ , the class name is the object category or description. For embeddings corresponding to relationships $(\\mathbf{r}_{i\\neq j})$ , we use the predicate text as class name. This differs from prior work [53, 54, 61], which", + "bbox": [ + 212, + 680, + 787, + 801 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "T. Salzmann et al.", + "bbox": [ + 271, + 114, + 395, + 127 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "This selection reduces the number of relationships that need to be processed, e.g., from $N^2 = 3600^2$ to $k = 16386$ (99.9% reduction) for our ViT-L/14 model.", + "bbox": [ + 217, + 810, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "uses the full triplet description. Our approach has the advantage of disentangling object category and predicate names, which allows for more efficient inference since object category and predicate texts are embedded separately.", + "bbox": [ + 212, + 146, + 787, + 205 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Relationship Score Loss. To select only the most promising embeddings for further processing, the Relationship Attention layer predicts a score $p_{ij}$ that represents the likelihood that the corresponding pair forms a relationship (if $i \\neq j$ ), or that the corresponding object exists in the image (if $i = j$ ). This score is trained with a sigmoid cross-entropy loss against targets provided by the model itself, namely the maximum probability predicted for any class for the corresponding $\\mathbf{r}_{ij}$ embedding [33]. In this way, the relationship score is trained to predict the class probability of a potential embedding $\\mathbf{r}_{ij}$ before actually computing and classifying the embedding. Note that this loss can only be computed for those objects and relationships that ultimately get selected for further processing. We found this to provide sufficient supervision.", + "bbox": [ + 212, + 207, + 787, + 372 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The final loss is an equally weighted sum of four components: (1) classification loss, (2) L1 box loss, (3) gIoU box loss, (4) relationship score loss.", + "bbox": [ + 212, + 373, + 787, + 405 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 428, + 375, + 446 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 214, + 460, + 428, + 478 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Datasets. We use the following datasets for training or evaluation:", + "bbox": [ + 214, + 488, + 705, + 503 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visual Genome [26] is the largest VRD dataset, labeled with 2.3M triplet relationships across 108K images. However, as Visual Genome includes noisy annotations, the community commonly evaluates VRD on a cleaned version of the dataset [55] where the label space is reduced to the 150 most frequent object classes and the 50 most frequent predicate classes. This dataset is commonly referred to as Visual Genome 150 (VG150).", + "bbox": [ + 212, + 503, + 787, + 593 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "GQA [16] uses the same image corpus as Visual Genome, but is more diversely labeled, with 1703 object classes (1704 including a background class) and 310 predicates. Spatial relationships like \"to the left of\" are automatically labeled based on the 2-D spatial arrangement of bounding boxes. Similar to VQ150, GQA200 is a reduced and cleaned version of GQA with 200 object classes and 100 predicate classes [9].", + "bbox": [ + 212, + 594, + 787, + 684 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "HICO-DET [4] is a dataset which focuses on a more specialized subset of visual relationships, specifically interactions between humans and objects. HICO-DET contains 50k images and is exhaustively annotated for 600 defined human-object interaction (HOIs).", + "bbox": [ + 212, + 685, + 787, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Objects365 [40] (O365) is a large scale object-detection dataset with 365 object categories across two million images.", + "bbox": [ + 212, + 744, + 787, + 776 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Open X-Embodiment [35] (OXE) dataset is targeted towards learning vision language action policies for robotics. OXE lacks bounding box annotations which precludes quantitative evaluation, but it represents both a visually different setting compared to the standard VRD benchmarks and a promising application area, so we use it for qualitative evaluation.", + "bbox": [ + 212, + 776, + 787, + 852 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Scene-Graph ViT", + "bbox": [ + 612, + 113, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/78b26de48b2c3d16bdca8e3dcec542e37b691cc232e06ba3a512f8866181f2ae.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
VG150VGHICOGQA200GQAObjects365
% in training mixture12.5%12.5%12.5%12.5%0%50%
# removed658717727729153611880115306
", + "bbox": [ + 230, + 143, + 772, + 189 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 1: Training data mixture. # removed indicates the number of images that were removed from the official training split due to overlap with data we evaluate on.", + "bbox": [ + 215, + 196, + 784, + 224 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Training. Unless otherwise noted, we train models on a mixture of VG, VG150, GQA200, HICO, and O365 in the proportions show in Table 1. The B/32 model is trained for 200'000 steps at batch size 256 (further details in Appendix A.1). Since the datasets we use share similar image sources, there is some overlap between the official training and evaluation splits. To ensure that no evaluation images are used for training, we rigorously filter all training datasets to remove images present in any of the test splits for all datasets we evaluate on, i.e. VG150, GQA, GQA200, and HICO. We use an image similarity filter that also identifies non-exact matches [24]. Table 1 shows the number of images removed from each training split. Note that not all prior work performs similar dedduplication.", + "bbox": [ + 212, + 242, + 787, + 395 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Evaluation Procedure and Metrics. Prior work presents a diverse range of evaluation metrics, often tailored to specific datasets. For non-exhaustively labelled datasets, like Visual Genome and GQA, using precision-based metrics is inherently inconclusive. Thus, the community has adapted a $Recall@K$ metric for these datasets, where $K$ denotes a fixed budget of triplets on which the recall is computed. However, this metric is biased towards the more frequent classes in the dataset [6,46]. Therefore, recent approaches use mean $Recall@K$ , which calculates $Recall@K$ separately for each class in the test data and then averages the results.", + "bbox": [ + 212, + 409, + 787, + 545 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For exhaustively labeled datasets like HICO, it is feasible to use precision-based metrics such as mean Average Precision (mAP), a well-established metric in object detection. For HICO, the mean is taken over the 600 possible HOI triplets. Besides the overall metric, we also report results separately for rare (< 10 occurrences) and non-rare ( $\\geq 10$ occurrences) HOIs as mAPr and mAPn.", + "bbox": [ + 212, + 546, + 787, + 622 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Further, evaluation can be \"graph-constrained\", where only a single prediction can be made per detected object pair, or \"graph-unconstrained\", where any number of predictions can be made per pair. Existing works evaluating on Visual Genome and GQA datasets partially lack a clear distinction between graph-constrained and unconstrained evaluation. For our models, we report graph-constrained results in the main paper, which reflects the dominant approach used in other studies. Results without graph-constraint are reported in Appendix A.2. For the HICO dataset, we follow the established evaluation procedure [4] that inherently represents a graph-unconstrained evaluation.", + "bbox": [ + 212, + 623, + 787, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Multiple implementations of the aforementioned metrics exist, often showing differences in results. To compare fairly and directly against a wide range of prior approaches we identified the PyTorch evaluation procedure of [45] as the most prevalent routine for recall-based metrics and replicated their procedure in JAX with numerical accuracy. Similarly, for HICO, we numerically reproduce the original Matlab evaluation procedure outlined in [4] in JAX.", + "bbox": [ + 212, + 760, + 787, + 851 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "T. Salzmann et al.", + "bbox": [ + 271, + 114, + 395, + 127 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9de4ecc27567bb3999b7891e1f07a0ce180fb938c80be318a3fdd4dd62831d7c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelBackbonemR@50mR@100
RelTR [7]ResNet-506.810.8
Transformer + CFA [28]ResNeXt-10112.314.6
VCTree [46] + IETrans + Rwt [57]ResNeXt-10112.014.9
Motif [55] + IETrans + Rwt [57]ResNeXt-10115.518.0
GPS-Net [31] + IETrans + Rwt [57]ResNeXt-10116.218.8
SG-Transformer [52] + IETrans + Rwt [57]ResNeXt-10115.018.1
Sgtr [29]ResNet-10115.820.1
DT2-ACBS [8]ResNet-10122.024.4
UniVRDCLIP: ViT-B/329.612.1
UniVRDCLIP: ViT-B/1610.913.2
UniVRDCLIP: ViT-L/1412.614.5
SG-ViTCLIP: ViT-B/3215.018.1
SG-ViTCLIP: ViT-B/1615.719.3
SG-ViTCLIP: ViT-L/1417.821.8
SG-ViT + simple predicate rebalancingCLIP: ViT-L/1422.326.1
", + "bbox": [ + 269, + 143, + 727, + 333 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 2: Performance on the Visual Genome 150 test set. We use graph-constrained evaluation (one predicate prediction per pair). For the last row (\"+ simple text rebalancing\"), the model was briefly fine-tuned on VG150 data with rebalanced predicate frequencies (see Section 4.2 and Appendix A.1).", + "bbox": [ + 212, + 339, + 787, + 395 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Exhaustive Relationship Evaluation. Previous methods [53,54,61] entangle objects and relationships in a single representation. Consequently, these methods score their embeddings against the full triplet. While this is feasible for datasets with a low number of possible triplets, like HICO with 600 defined relationships, it quickly becomes computationally intractable for datasets with a larger vocabulary of objects and predicates (e.g. VG, GQA). Prior works [61] therefore evaluate only on triplets which are known to be present in the test split, which may inflate metrics. In contrast, our method disentangles objects and predicates (Section 3). This allows for an exhaustive evaluation that considers all object and predicate combinations and enables applications where knowledge of possible relationships is unavailable.", + "bbox": [ + 212, + 410, + 787, + 578 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Relationship Detection Performance", + "text_level": 1, + "bbox": [ + 215, + 601, + 560, + 616 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We use Visual Genome as our main benchmark for relationship detection. Table 2 shows results for our models and prior work. Prior studies often achieve reported performance through the combination of a base model with a specialized approach to overcome the skewed data distribution (indicated as base-model + specialization in Table 2). Our models provide strong performance even without special treatment of the skewed predicate distribution. After brief fine-tuning on data in which the frequency of predicate annotations was rebalanced with simple rejection sampling (see Appendix A.1), our method improves on the prior best method DT2-ACBS [8] by 1.7 points mR@100 (26.1 vs. 24.4). This shows that our already-strong method may be further improved by combining it with more advanced rebalancing methods.", + "bbox": [ + 212, + 627, + 787, + 792 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We also note the large difference in performance to UniVRD [61], which builds on the same detection architecture as our method, but adds a Transformer decoder-based relationship module. To disentangle whether the improve", + "bbox": [ + 212, + 794, + 787, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Scene-Graph ViT", + "bbox": [ + 612, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/f390adbd569e1eb58b400e7d878db4fbdbc05aace5b222067c696053c7732b2f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelmR@50mR@100
LLM4SGG [21]5.36.5
Neural Motif w/ GCL [9]16.818.8
VCTree w/ GCL [9]15.617.8
SG-Transformer w/ CFA [28]13.415.3
SHA w/ GCL [9]17.820.1
SG-ViT (CLIP: ViT-B/32)15.718.6
SG-ViT (CLIP: ViT-B/16)17.720.5
SG-ViT (CLIP: ViT-L/14)19.322.9
", + "bbox": [ + 215, + 157, + 493, + 271 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/9d576e06afa31662f0e93e37f961b032f41f88d11d56a31d6cc6a8bad2bf12c2.jpg", + "table_caption": [ + "Table 3: Performance on the GQA200 test set. Graph-constrained evaluation." + ], + "table_footnote": [], + "table_body": "
ModelmR@50mR@100
Scene Graph Classification
IMP [25,44,48]0.50.7
Neural Motif [25,44,55]0.81.2
Unbiased TDE [25,45]-0.7
RTN [25]-1.4
MP [23]-2.8
Transformer w/ EBM [44]1.31.8
Scene Graph Generation
SG-ViT (CLIP: ViT-B/32)6.27.4
SG-ViT (CLIP: ViT-B/16)6.47.4
SG-ViT (CLIP: ViT-L/14)8.09.6
", + "bbox": [ + 511, + 143, + 777, + 287 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 4: Performance on the GQA test set. Graph-constrained evaluation.", + "bbox": [ + 511, + 289, + 781, + 316 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "ments over UniVRD are due to differences in architecture or training data, we also trained our B/32 model on the UniVRD data mixture, and still observe a large improvement (17.4 VG mR@100 for our method vs. 12.1 for UniVRD when trained on the same data; Table 6). This suggests that our encoder-only architecture is better suited for VRD than prior decoder-based architectures [7,20,29,61].", + "bbox": [ + 212, + 332, + 784, + 409 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Scaling to Large Vocabularies", + "text_level": 1, + "bbox": [ + 215, + 433, + 504, + 448 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Improving long-tail performance has been critical to the VRD field since the natural distribution of relationship triplets is highly skewed [27]. A substantial body of literature is devoted to this goal, often proposing complex loss debiasing or data augmentation approaches [8,27,52,57]. It is therefore of interest whether our method works well in this regime without special treatment of rare classes. To evaluate the performance of our method on large object and predicate vocabularies with a long tail of rare classes, we use the Graph Question Answering (GQA) dataset [16]. This dataset builds on VG and expands the number of classes and annotations. A simplified version of GQA with 200 object and 100 predicate classes, called GQA200 [9], is part of our training mixture. Our method surpasses prior results on the GQA200 test split (Table 3), despite lacking the specialized treatment of the data distribution that some prior methods use.", + "bbox": [ + 212, + 460, + 787, + 642 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The full GQA dataset has an even larger vocabulary with 1703 object and 311 predicate classes. To our knowledge, no prior methods evaluate scene graph generation on the full GQA dataset. Some prior methods report results on scene graph classification, where ground-truth boxes are provided. As shown in Table 4, our model achieves higher performance on the much harder task of scene graph generation than prior methods on classification. We further assess zero-shot generalization to unseen classes in Appendix A.2.", + "bbox": [ + 212, + 643, + 787, + 748 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We believe that two factors contribute to the strong performance in the open vocabulary regime: First, the simple design of the method does not impede transfer of semantic knowledge from the pre-trained VLM backbone because it adds only lightweight heads. In particular, we use no decoder, which could be difficult to train and could lead to \"forgetting\" of pretrained representations. Instead, most of the relationship modeling happens in the backbone, which has", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "T. Salzmann et al.", + "bbox": [ + 271, + 114, + 395, + 127 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "had the benefit of large-scale vision-language pretraining. Second, the open-vocabulary design allows end-to-end training on a mix of datasets, which was not possible for all prior methods.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.4 Human-Object Interaction", + "text_level": 1, + "bbox": [ + 214, + 215, + 482, + 231 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We also evaluate our model on the specialized task of human-object interaction, using the HICO benchmark [4]. The performance of our model is on par with the most comparable prior method [61], but does not show similarly large improvements as for VG or GQA200 (Table 5). HICO has a much narrower and more specialized vocabulary than VG or GQA (HICO evaluates on 600 prespecified triplets, whereas VG150 has $1.1 \\times 10^{6}$ possible triplet combinations). Performance on this specialized task may benefit less from transfer of pretrained representations and may instead be limited by the amount of task-specific training data. This is supported by the fact that the recent state-of-the-art method for HICO pretrains on large amounts of person-focused pseudo-labels [54].", + "bbox": [ + 212, + 241, + 787, + 393 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.5 Ablations", + "text_level": 1, + "bbox": [ + 214, + 415, + 344, + 429 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Inference Speed and Number of Predicted Relationships. In the following, we ablate training datasets and model components to study the interplay of data and architecture on the model performance.", + "bbox": [ + 212, + 438, + 602, + 497 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The primary hyperparameter of our method is the number of relationships $k$ that are selected by the Relationship Attention layer from all possible combinations for further processing. Figure 3 shows how VRD performance and inference speed depend on $k$ .", + "bbox": [ + 212, + 500, + 602, + 588 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "For training, we use $k = 2^{14}$ , which results in $71\\%$ (B/32) to $81\\%$ (L/14) of the speed of a pure object detection model [34] of the same size. Without top-k selection, the model would be several times slower (right end of top plot in Figure 3). Since the vast majority of pairs do not form relationships, VRD performance is essentially unaffected by top- $k$ selection at inference. For inference, we can therefore use $k = 2^{11}$ at no loss in accuracy to achieve $90\\%$ of the speed of the object detector (for B/32 and", + "bbox": [ + 212, + 589, + 602, + 739 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "L/14). At this $k$ , the B/32 model achieves 52.8 FPS at batch size 1 on an NVIDIA V100 GPU and is therefore suitable for real-time applications.", + "bbox": [ + 212, + 739, + 785, + 771 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/6ce65786911baddd0584536ae1881f94620020316915a4e5beca96624dd3b69e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 616, + 444, + 782, + 518 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/ee12aaefd35ea4036b65e7cfd654c798dde90954e058b06c94ce55efc763b626.jpg", + "image_caption": [ + "Figure 3: Model speed and VRD accuracy by number of selected relationships $k$ . Speed is relative to a non-VRD object detector [34]." + ], + "image_footnote": [], + "bbox": [ + 617, + 527, + 782, + 627 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Dataset Mixture. To assess how training data and architectural improvements contribute to the performance of our model compared to prior work, we train the models on a range of dataset mixtures (Table 6). The full mixture, which we use for all experiments unless otherwise noted, includes VG, VG150, GQA200,", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Scene-Graph ViT", + "bbox": [ + 612, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/7189384ec9b4a431c9263347626b2060294e3550b9ec7fc0665b71f9cab14225.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelBackbonemAPmAPrmAPn
HOTR [20]ResNet5025.1020.3325.86
RLIP [53]ResNet5032.8426.8534.63
RLIPv2 [54]ResNet5035.3829.6137.10
RLIPv2 [54]Swin-L45.0943.2345.64
UniVRD [60]CLIP: ViT-B/3229.9822.9432.02
UniVRD [60]CLIP: ViT-B/1629.9822.9432.02
UniVRD [60]CLIP: ViT-L/1437.4128.9039.95
SG-ViTCLIP: ViT-B/3228.8623.7230.39
SG-ViTCLIP: ViT-B/1631.9826.8333.52
SG-ViTCLIP: ViT-L/1438.1133.7139.42
", + "bbox": [ + 351, + 143, + 651, + 272 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/6336eef133c28a416384855a11e66d7c6a278eca3ef146ca2f8fe6350c40a875.jpg", + "table_caption": [ + "Table 5: Performance on the HICO test set. Our method is on par with the most comparable prior work, UniVRD, for overall mAP. However, our method performs better on \"rare\" classes (mAPr), indicating better generalization to rarely seen concepts. For reference, we also show RLIPv2, the state-of-the-art method on HICO. RLIPv2 proposes a self-training approach that is orthogonal and compatible to our method." + ], + "table_footnote": [], + "table_body": "
VGMixtureVG150 mR@100GQA200 mR@100HICO mAP
VG150GQA200HICOO365COCO
Our mixture12.5%12.5%12.5%12.5%50%-18.118.628.9
UniVRD [61] mix-10%-20%50%20%17.4 ↓ 4%5.1 ↓ 73%17.4 ↓ 8%
ours w/o O36525%25%25%25%--15.3 ↓ 15%15.6 ↓ 16%23.8 ↓ 18%
ours w/o VG-25%12.5%12.5%50%-14.2 ↓ 22%15.3 ↓ 18%26.8 ↓ 7%
", + "bbox": [ + 215, + 382, + 787, + 465 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 6: Dataset Ablation. Performance of SG-ViT (B/32) trained on different data mixtures. VG150 and GQA200 evaluation is graph-constrained, HICO is unconstrained.", + "bbox": [ + 212, + 474, + 787, + 503 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "HICO and O365. To allow direct comparison to UniVRD [61], we also train on their mixture (VG150, HICO, O365, COCO) and find that our model still performs better on VG150 by a large margin (18.1 mR@100 vs 12.1 for UniVRD for B/32 models). Since UniVRD uses the same image and text encoding backbones as our model, this result suggests that our encoder-only model with Relationship Attention is an advance over decoder-based relationship architectures. We further find that including both pure object detection (O365) and large-vocabulary VRD data (VG) benefit all metrics.", + "bbox": [ + 212, + 549, + 787, + 672 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Object Detection Performance. A necessary step towards good relationship detection performance is accurate object detection. In our architecture, object detection is treated as a special case of relationship detection in which the is identical to the . We validate this design by training the model only on detection datasets without relationship annotations, using the same dataset mixture as the OWL-ViT [34] model (Objects365 and Visual Genome, [34]). We find that our model achieves a similar object detection performance as OWL-ViT (21.9% vs. 22.1% mAP and 18.3% vs. 18.9% mAPr on LVIS [12]), indicating that the Relationship Attention design does not interfere with object detection.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "T. Salzmann et al.", + "bbox": [ + 271, + 114, + 393, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/f6264e2b1fc727434aab771caeec0d0fb4ceb3c09d21ec7a9cf8800abb6cbdaa.jpg", + "image_caption": [ + "Figure 4: Qualitative examples of difficult edge-cases from VG150 test split [26, 55]. From left to right: Ground Truth, SG-ViT (B/32), SG-ViT (L/14). In all cases the is lime and the is red." + ], + "image_footnote": [], + "bbox": [ + 243, + 142, + 759, + 459 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.6 Qualitative Results", + "text_level": 1, + "bbox": [ + 215, + 541, + 419, + 556 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In Figure 4, we visualize qualitative examples to highlight difficult edge cases. Even where the model output differs from the ground truth, it is often plausibly correct, e.g. where either the extent of an object differs with the ground truth or there are multiple appropriate, but unannotated, relations. For example, both models select the headlight on the bus in Figure 4b for \"light on bus\" as opposed to the sun reflected off the windshield. In Figure 4c both models focus more on the \"on\" in \"snow on hill\" than the distinction between a mountain and a hill. Figure 4d shows a challenging scene where the B/32 model selects the bottle inside the open fridge while the L/14 correctly selects a water bottle in the door for \"bottle in door\" while also capturing the extent of the door instance.", + "bbox": [ + 212, + 566, + 787, + 718 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Furthermore, to qualitatively assess the utility of this work beyond standard relation benchmark datasets, we also visualize some example relations applied to data from the robotics domain, namely the OXE dataset [35]. Figure 5 shows a real-life scene where multiple instances of the same semantic object categories are present within the scene. The model is able to match the to the instances in the desired configuration. One subjective failure-case is that the \"banana outside bowl\" text query had its highest scoring triplet involving a bowl on a shelf in the background rather than the adjacent bowl. Another charac", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Scene-Graph ViT", + "bbox": [ + 612, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/96211429d2725ae9e1b59b8a13180e29bd18aef541e76869374e0860caf512bb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 238, + 157, + 367, + 246 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/254994729b19ecd82ba91ce8146a55b6a254d15717861e67e505b8f10ec2d570.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 157, + 500, + 246 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/aeb341dbc46b61dda18fddfb4d8b9a969af189f75879cfa642faca97a3dd5b71.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 157, + 633, + 246 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/78b4213f16e188704cfe8e1fae668aed3ec2aee932542858578e5c5d5925f81b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 157, + 767, + 246 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/1e4342115292a3f92fbbfd5e311b2de5f1967605b4f32a18b54289b3c989ea20.jpg", + "image_caption": [ + "Figure 5: Qualitative examples showing SG-ViT (L/14) on out-of-distribution data from the OXE dataset [35]. In all cases the is lime and the is red. Note how the model correctly disambiguates several instances of the same class (e.g. \"bananas\" and \"bottle\") depending on their relationships." + ], + "image_footnote": [], + "bbox": [ + 238, + 248, + 367, + 335 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/a83cbc2f571ab5f2471b9a5a44ab9afb5748de82eb8c623823c3f7337ee5a21e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 248, + 500, + 335 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/906289885ef00cead1092aadcb6e8ec257528af9fdc0677df59a21814e53f5dc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 248, + 633, + 335 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/de418ba07b8a61d2a01bf4bd1a4ab2e4ad2ca3e10f1458c1001e6e19619aaf4d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 248, + 767, + 335 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "teristic is that the model occasionally places boxes over groups with multiple instances of the same semantic class when a singular word is used for either the or . This can be attributed to the similarity between word and their plural in the CLIP embeddings and that many human annotations from VG also confuse singular and plural (see Appendix A.3).", + "bbox": [ + 212, + 428, + 784, + 503 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Limitations", + "text_level": 1, + "bbox": [ + 215, + 522, + 362, + 537 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "While our model performs strongly on large-vocabulary VRD, its performance on specialized human-object interaction detection is only on par with prior models. We discuss this limitation in Section 4.4. Further error modes are explored qualitatively in Section 4.6 and Appendix A.3. A challenge for open-vocabulary relationship detection models as a whole is zero-shot generalization to unseen objects and predicates. While our model improves over prior approaches (Tables 3 and 4), there is still a large gap when it comes to entirely unseen classes (Appendix A.2). Future research should focus on closing this gap.", + "bbox": [ + 212, + 550, + 787, + 672 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 691, + 359, + 707 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We present an architecture for open-vocabulary visual relationship detection. By combining an encoder-only design with a novel Relationship Attention layer for efficient selection of high-confidence relationships, our architecture leverages VLM pretraining and multi-dataset VRD training. It achieves strong performance on standard and large-vocabulary VRD benchmarks while maintaining pure object detection performance and adding little extra computational cost. Due to its simplicity and strong performance, we believe that our method will be a useful basis for further research on visual relationship detection.", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "T. Salzmann et al.", + "bbox": [ + 271, + 114, + 395, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Amiri, S., Chandan, K., Zhang, S.: Reasoning with scene graphs for robot planning under partial observability. IEEE Robotics and Automation Letters 7(2), 5560-5567 (2022)", + "2. Ba, J.L., Kiros, J.R., Hinton, G.E.: Layer normalization. arXiv preprint arXiv:1607.06450 (2016)", + "3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: Conference on Computer Vision (ECCV). pp. 213-229. Springer (2020)", + "4. Chao, Y.W., Liu, Y., Liu, X., Zeng, H., Deng, J.: Learning to Detect Human-Object Interactions. In: IEEE Winter Conference on Applications of Computer Vision (2018)", + "5. Chen, B., Xia, F., Ichter, B., Rao, K., Gopalakrishnan, K., Ryoo, M.S., Stone, A., Kappler, D.: Open-vocabulary queryable scene representations for real world planning. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 11509-11522. IEEE (2023)", + "6. Chen, T., Yu, W., Chen, R., Lin, L.: Knowledge-embedded routing network for scene graph generation. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 6163-6171 (2019)", + "7. Cong, Y., Yang, M.Y., Rosenhahn, B.: Reltr: Relation transformer for scene graph generation. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)", + "8. Desai, A., Wu, T.Y., Tripathi, S., Vasconcelos, N.: Learning of visual relations: The devil is in the tails. In: IEEE/CVF International Conference on Computer Vision (ICCV). pp. 15404-15413 (2021)", + "9. Dong, X., Gan, T., Song, X., Wu, J., Cheng, Y., Nie, L.: Stacked hybrid-attention and group collaborative learning for unbiased scene graph generation. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 19427-19436 (2022)", + "10. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. ICLR (2021)", + "11. Gu, Q., Kuwajerwala, A., Morin, S., Jatavallabhula, K.M., Sen, B., Agarwal, A., Rivera, C., Paul, W., Ellis, K., Chellappa, R., et al.: Conceptgraphs: Open-vocabulary 3d scene graphs for perception and planning. arXiv preprint arXiv:2309.16650 (2023)", + "12. Gupta, A., Dollar, P., Girshick, R.: LVIS: A Dataset for Large Vocabulary Instance Segmentation. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2019)", + "13. He, T., Gao, L., Song, J., Li, Y.F.: Towards Open-Vocabulary Scene Graph Generation with Prompt-Based Finetuning. In: European Conference on Computer Vision (ECCV) (2022)", + "14. Hendrycks, D., Gimpel, K.: Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415 (2016)", + "15. Herzig, R., Mendelson, A., Karlinsky, L., Arbelle, A., Feris, R., Darrell, T., Globerson, A.: Incorporating Structured Representations into Pretrained Vision & Language Models Using Scene Graphs. In: The 2023 Conference on Empirical Methods in Natural Language Processing (2023)" + ], + "bbox": [ + 218, + 175, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Scene-Graph ViT", + "bbox": [ + 612, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Hudson, D.A., Manning, C.D.: Gqa: A new dataset for real-world visual reasoning and compositional question answering. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 6700-6709 (2019)", + "17. Hughes, N., Chang, Y., Carlone, L.: Hydra: A Real-time Spatial Perception System for 3D Scene Graph Construction and Optimization. Robotics: Science and Systems (RSS) (2022)", + "18. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International conference on machine learning. pp. 4904-4916. PMLR (2021)", + "19. Johnson, J., Krishna, R., Stark, M., Li, L.J., Shamma, D., Bernstein, M., Fei-Fei, L.: Image Retrieval Using Scene Graphs. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2015)", + "20. Kim, B., Lee, J., Kang, J., Kim, E.S., Kim, H.J.: Hotr: End-to-end human-object interaction detection with transformers. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 74-83 (2021)", + "21. Kim, K., Yoon, K., Jeon, J., In, Y., Moon, J., Kim, D., Park, C.: LLM4SGG: Large Language Model for Weakly Supervised Scene Graph Generation. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024)", + "22. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)", + "23. Knyazev, B., de Vries, H., Cangea, C., Taylor, G.W., Courville, A., Belilovsky, E.: Graph Density-Aware Losses for Novel Compositions in Scene Graph Generation. In: British Machine Vision Conference (BMVC) (2020)", + "24. Kolesnikov, A., Beyer, L., Zhai, X., Puigcerver, J., Yung, J., Gelly, S., Houlsby, N.: Big transfer (bit): General visual representation learning. In: European Conference on Computer Vision (ECCV). pp. 491-507 (2020)", + "25. Koner, R., Shit, S., Tresp, V.: Relation transformer network. arXiv preprint arXiv:2004.06193 (2020)", + "26. Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., et al.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. International Journal of Computer Vision 123, 32-73 (2017)", + "27. Li, H., Zhu, G., Zhang, L., Jiang, Y., Dang, Y., Hou, H., Shen, P., Zhao, X., Shah, S.A.A., Bennamoun, M.: Scene Graph Generation: A comprehensive survey. Neurocomputing (2024). https://doi.org/https://doi.org/10.1016/j.neucom.2023.127052", + "28. Li, L., Chen, G., Xiao, J., Yang, Y., Wang, C., Chen, L.: Compositional feature augmentation for unbiased scene graph generation. In: IEEE/CVF International Conference on Computer Vision (ICCV). pp. 21685-21695 (2023)", + "29. Li, R., Zhang, S., He, X.: Sgtr: End-to-end scene graph generation with transformer. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 19486-19496 (2022)", + "30. Li, Y., Ouyang, W., Zhou, B., Wang, K., Wang, X.: Scene graph generation from objects, phrases and region captions. In: IEEE/CVF International Conference on Computer Vision (ICCV). pp. 1261-1270 (2017)", + "31. Lin, X., Ding, C., Zeng, J., Tao, D.: Gps-net: Graph property sensing network for scene graph generation. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3746-3753 (2020)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "T. Salzmann et al.", + "bbox": [ + 271, + 114, + 395, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "32. Lu, C., Krishna, R., Bernstein, M., Fei-Fei, L.: Visual relationship detection with language priors. In: European Conference on Computer Vision (ECCV). pp. 852-869. Springer (2016)", + "33. Minderer, M., Gritsenko, A., Houlsby, N.: Scaling open-vocabulary object detection. Advances in Neural Information Processing Systems 36 (2024)", + "34. Minderer, M., Gritsenko, A., Stone, A., Neumann, M., Weissenborn, D., Dosovitskiy, A., Mahendran, A., Arnab, A., Dehghani, M., Shen, Z., et al.: Simple open-vocabulary object detection. In: European Conference on Computer Vision (ECCV). pp. 728-755. Springer (2022)", + "35. Open X-Embodiment Collaboration: Open X-Embodiment: Robotic Learning Datasets and RT-X Models (2023)", + "36. Peng, Z., Wang, W., Dong, L., Hao, Y., Huang, S., Ma, S., Ye, Q., Wei, F.: Grounding Multimodal Large Language Models to the World. In: International Conference on Learning Representations (2024)", + "37. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)", + "38. Rana, K., Haviland, J., Garg, S., Abou-Chakra, J., Reid, I., Suenderhauf, N.: Say-plan: Grounding large language models using 3d scene graphs for scalable robot task planning. In: 7th Annual Conference on Robot Learning (2023)", + "39. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems 28 (2015)", + "40. Shao, S., Li, Z., Zhang, T., Peng, C., Yu, G., Zhang, X., Li, J., Sun, J.: Objects365: A large-scale, high-quality dataset for object detection. In: IEEE/CVF International Conference on Computer Vision (ICCV). pp. 8430-8439 (2019)", + "41. Shi, H., Hayat, M., Cai, J.: Open-vocabulary object detection via scene graph discovery. In: 31st ACM International Conference on Multimedia. pp. 4012-4021 (2023)", + "42. Singh, K.P., Salvador, J., Weihs, L., Kembhavi, A.: Scene graph contrastive learning for embodied navigation. In: IEEE/CVF International Conference on Computer Vision (ICCV). pp. 10884-10894 (2023)", + "43. Song, H., Sun, D., Chun, S., Jampani, V., Han, D., Heo, B., Kim, W., Yang, M.H.: ViDT: An Efficient and Effective Fully Transformer-based Object Detector. In: International Conference on Learning Representations (2022)", + "44. Suhail, M., Mittal, A., Siddiquie, B., Broaddus, C., Eledath, J., Medioni, G., Sigal, L.: Energy-based learning for scene graph generation. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13936-13945 (2021)", + "45. Tang, K., Niu, Y., Huang, J., Shi, J., Zhang, H.: Unbiased scene graph generation from biased training. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3716-3725 (2020)", + "46. Tang, K., Zhang, H., Wu, B., Luo, W., Liu, W.: Learning to compose dynamic tree structures for visual contexts. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 6619-6628 (2019)", + "47. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)", + "48. Xu, D., Zhu, Y., Choy, C.B., Fei-Fei, L.: Scene graph generation by iterative message passing. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5410-5419 (2017)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Scene-Graph ViT", + "bbox": [ + 612, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "49. Yang, J., Lu, J., Lee, S., Batra, D., Parikh, D.: Graph r-cnn for scene graph generation. In: European Conference on Computer Vision (ECCV). pp. 670-685 (2018)", + "50. Yao, Z., Ai, J., Li, B., Zhang, C.: Efficient DETR: Improving End-to-End Object Detector With Dense Prior. arXiv preprint arXiv:2104.01318 (2021)", + "51. Yu, J., Wang, Z., Vasudevan, V., Yeung, L., Seyedhosseini, M., Wu, Y.: CoCa: Contrastive Captioners are Image-Text Foundation Models. Transactions on Machine Learning Research (2022)", + "52. Yu, J., Chai, Y., Wang, Y., Hu, Y., Wu, Q.: CogTree: Cognition Tree Loss for Unbiased Scene Graph Generation. In: International Joint Conference on Artificial Intelligence, IJCAI-21. International Joint Conferences on Artificial Intelligence Organization (8 2021)", + "53. Yuan, H., Jiang, J., Albaniae, S., Feng, T., Huang, Z., Ni, D., Tang, M.: Rlip: Relational language-image pre-training for human-object interaction detection. Advances in Neural Information Processing Systems 35, 37416-37431 (2022)", + "54. Yuan, H., Zhang, S., Wang, X., Albanie, S., Pan, Y., Feng, T., Jiang, J., Ni, D., Zhang, Y., Zhao, D.: RLIPv2: Fast Scaling of Relational Language-Image Pretraining. In: IEEE/CVF International Conference on Computer Vision (ICCV). pp. 21649-21661 (2023)", + "55. Zellers, R., Yatskar, M., Thomson, S., Choi, Y.: Neural motifs: Scene graph parsing with global context. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5831-5840 (2018)", + "56. Zhai, X., Wang, X., Mustafa, B., Steiner, A., Keysers, D., Kolesnikov, A., Beyer, L.: Lit: Zero-shot transfer with locked-image text tuning. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 18123-18133 (2022)", + "57. Zhang, A., Yao, Y., Chen, Q., Ji, W., Liu, Z., Sun, M., Chua, T.S.: Fine-grained scene graph generation with data transfer. In: European Conference on Computer Vision (ECCV). pp. 409-424. Springer (2022)", + "58. Zhang, H., Li, F., Liu, S., Zhang, L., Su, H., Zhu, J., Ni, L.M., Shum, H.Y.: Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605 (2022)", + "59. Zhang, Y., Ma, Z., Gao, X., Shakiah, S., Gao, Q., Chai, J.: GROUNDHOG: Grounding Large Language Models to Holistic Segmentation. In: Conference on Computer Vision and Pattern Recognition 2024 (2024), https://openreview.net/forum?id=PWW3ux1Ju", + "60. Zhao, L., Yuan, L., Gong, B., Cui, Y., Schroff, F., Yang, M.H., Adam, H., Liu, T.: Unified Visual Relationship Detection with Vision and Language Models. In: IEEE/CVF International Conference on Computer Vision (ICCV) (2023)", + "61. Zhao, L., Yuan, L., Gong, B., Cui, Y., Schroff, F., Yang, M.H., Adam, H., Liu, T.: Unified Visual Relationship Detection with Vision and Language Models. In: IEEE/CVF International Conference on Computer Vision (ICCV) (2023)", + "62. Zhou, X., Girdhar, R., Joulin, A., Krahenbuhl, P., Misra, I.: Detecting twenty-thousand classes using image-level supervision. In: European Conference on Computer Vision (ECCV). pp. 350-368 (2022)", + "63. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable DETR: Deformable Transformers for End-to-End Object Detection. In: International Conference on Learning Representations (2021)" + ], + "bbox": [ + 215, + 147, + 784, + 784 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "T. Salzmann et al.", + "bbox": [ + 271, + 114, + 395, + 127 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Scene-Graph ViT_ End-to-End Open-Vocabulary Visual Relationship Detection/07085046-a430-402b-be81-31056599f4f5_model.json b/2024/Scene-Graph ViT_ End-to-End Open-Vocabulary Visual Relationship Detection/07085046-a430-402b-be81-31056599f4f5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4f8729afc9c69132df54f80d90ba0bc2c12adf8e --- /dev/null +++ b/2024/Scene-Graph ViT_ End-to-End Open-Vocabulary Visual Relationship Detection/07085046-a430-402b-be81-31056599f4f5_model.json @@ -0,0 +1,2447 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.225, + 0.141, + 0.78, + 0.187 + ], + "angle": 0, + "content": "Scene-Graph ViT: End-to-End Open-Vocabulary Visual Relationship Detection" + }, + { + "type": "text", + "bbox": [ + 0.341, + 0.212, + 0.664, + 0.244 + ], + "angle": 0, + "content": "Tim Salzmann\\(^{1,2}\\), Markus Ryll\\(^{2}\\), Alex Bewley\\(^{\\star 1}\\), and Matthias Minderer\\(^{\\star 1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.37, + 0.254, + 0.626, + 0.311 + ], + "angle": 0, + "content": "\\(^{1}\\) Google DeepMind {tsal, bewley, mjlm}@google.com \\(^{2}\\) Technical University Munich {tim.salzmann, markus.ryll}@tum.d" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.333, + 0.559, + 0.508 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.561, + 0.332, + 0.789, + 0.508 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.511, + 0.788, + 0.54 + ], + "angle": 0, + "content": "Figure 1: Relationships detected by our method on an unseen image. The top relationships by confidence score are shown. Photo by Spacejoy on Unsplash." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.566, + 0.744, + 0.775 + ], + "angle": 0, + "content": "Abstract. Visual relationship detection aims to identify objects and their relationships in images. Prior methods approach this task by adding separate relationship modules or decoders to existing object detection architectures. This separation increases complexity and hinders end-to-end training, which limits performance. We propose a simple and highly efficient decoder-free architecture for open-vocabulary visual relationship detection. Our model consists of a Transformer-based image encoder that represents objects as tokens and models their relationships implicitly. To extract relationship information, we introduce an attention mechanism that selects object pairs likely to form a relationship. We provide a single-stage recipe to train this model on a mixture of object and relationship detection data. Our approach achieves state-of-the-art relationship detection performance on Visual Genome and on the large-vocabulary GQA benchmark at real-time inference speeds. We provide ablations, real-world qualitative examples, and analyses of zero-shot performance." + }, + { + "type": "page_footnote", + "bbox": [ + 0.233, + 0.811, + 0.538, + 0.841 + ], + "angle": 0, + "content": "\\(^{\\dagger}\\)Work done while at Google DeepMind. \n*Advising project leads in alphabetical order." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.397, + 0.128 + ], + "angle": 0, + "content": "T. Salzmann et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.376, + 0.161 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.18, + 0.788, + 0.301 + ], + "angle": 0, + "content": "A fundamental goal of computer vision is to decompose visual scenes into structured semantic representations. A commonly studied task towards this goal is object detection, in which objects in an image are localized by bounding boxes and classified into semantic categories. However, a scene description also includes the semantic relationships between objects. This gives rise to the task of visual relationship detection (VRD, [4,6,32,48,49,55]). In VRD, the model detects objects and infers pairwise relationships between them in the form of triplets3 [32]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.301, + 0.788, + 0.392 + ], + "angle": 0, + "content": "Detecting both objects and their relationships allows the construction of a scene graph [19,30,48] in which objects are represented as nodes and their relationships as edges. Scene graph generation (SGG) has wide-ranging applications in robotics [1,5,11,17,38,42] and image retrieval [19,27]. Increasingly, structured scene representations are also employed to provide grounding and explainability to multimodal large language models [15,36,59]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.392, + 0.788, + 0.543 + ], + "angle": 0, + "content": "Prior work typically draws a distinction between object detection and relationship prediction. Detection is performed either as a wholly separate step before relationship prediction [8, 32, 48, 52, 55, 57], or by separate model parts such as \"relationship decoders\" that are responsible for modeling the interactions between objects [7, 49, 61]. This distinction makes it hard to optimize such models end-to-end for VRD [61]. In contrast, we propose an encoder-only architecture that models objects and relationships jointly, directly in the image encoder. Our architecture performs open-vocabulary relationship detection and can be trained end-to-end on arbitrary mixtures of object detection and relationship annotations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.543, + 0.788, + 0.664 + ], + "angle": 0, + "content": "Our model builds on a Transformer-based encoder-only object detector [34] in which the output tokens of the image encoder directly represent object proposals. From these tokens, class embeddings and bounding boxes are decoded with lightweight heads. Our insight is that this architecture is perfectly set up to learn relationships between objects directly in the image encoder, without the need for additional relationship-specific stages. This is because the existing self-attention in the encoder already models all-to-all pairwise interactions between the object proposal tokens." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.665, + 0.788, + 0.803 + ], + "angle": 0, + "content": "To access information about the relationship between two of these tokens, we combine the embeddings corresponding to the and token using a new Relationship Attention layer. Obtaining relationship embeddings for all possible pairwise combinations of object proposal tokens would be computationally infeasible. To reduce the number of combinations, we introduce a self-supervised hard attention mechanism that selects the highest-confidence pairs at a computational cost comparable to a single self-attention layer. We show how to directly supervise the attention scores of this mechanism without the need to propagate gradients through the hard selection." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.788, + 0.84 + ], + "angle": 0, + "content": "3 We use to distinguish the grammatical terms and from generic use of the word object as in object detection." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.613, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Scene-Graph ViT" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.783, + 0.221 + ], + "angle": 0, + "content": "An additional benefit of our design is that it disentangles object names from relationship predicates during inference. In contrast to prior open-vocabulary methods [53, 54, 61], our model can embed object and predicate texts separately and generate confidence scores efficiently for all possible combinations." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.223, + 0.601, + 0.236 + ], + "angle": 0, + "content": "In summary, we make the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.244, + 0.774, + 0.258 + ], + "angle": 0, + "content": "1. An efficient architecture for open-vocabulary visual relationship detection." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.258, + 0.762, + 0.273 + ], + "angle": 0, + "content": "2. A single-stage recipe for joint object and relationship detection training." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.273, + 0.646, + 0.287 + ], + "angle": 0, + "content": "3. Efficient, disentangled object and relationship inference." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.287, + 0.704, + 0.302 + ], + "angle": 0, + "content": "4. Analysis of inference speed, ablations, and qualitative examples." + }, + { + "type": "list", + "bbox": [ + 0.222, + 0.244, + 0.774, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.309, + 0.783, + 0.383 + ], + "angle": 0, + "content": "Our method achieves state-of-the-art visual relationship detection performance on the Visual Genome dataset (26.1% mR@100 with graph constraint) and on the large-vocabulary GQA benchmark. It provides strong results while being significantly simpler than prior approaches to visual relationship detection." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.405, + 0.387, + 0.421 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.434, + 0.783, + 0.464 + ], + "angle": 0, + "content": "Below, we review the main approaches to VRD that are relevant to our work. For a comprehensive overview of the field, we refer to [27]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.469, + 0.785, + 0.606 + ], + "angle": 0, + "content": "Detector-Agnostic Relationship Detection. Historically, many works on VRD assume that object detections are given, such that the task reduces to inferring relationships between the objects. These \"detector-agnostic\" methods use off-the-shelf detectors such as Faster-RCNN [39] to obtain boxes and box embeddings and infer a scene graph from them. Early work on large-scale VRD employed word embeddings to improve relationship generalization [32] and graph inference methods such as message passing for SGG [48]. More recent methods leverage relationship co-occurrence statistics [55] and address the long-tailed nature of the training data [8, 52, 57]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.61, + 0.787, + 0.853 + ], + "angle": 0, + "content": "End-to-End Relationship Detection. A fundamental limitation of the detector-agnostic approach is that object detection and relationship prediction are treated separately and cannot learn from each other. Overcoming this limitation has recently motivated the development of end-to-end architectures in which objects and relationships are predicted jointly by a single model [7, 20, 29, 61]. All of these models have in common that they use a Transformer [47] decoder to predict object and relationship embeddings. The decoder represents object proposals with query embeddings that can cross-attend into image embeddings. The literature on object detection suggests that this choice may be problematic, because the query embeddings can be difficult to initialize, which can lead to unstable and slow optimization [3, 34, 43, 50, 58, 63]. To avoid this issue, we design an encoder-only architecture in which both object and relationship embeddings are computed directly from the output tokens of a Vision Transformer [10] image encoder, building on an idea from object detection [34]. In contrast to prior models [61], this allows us to train the model end-to-end on object and relationship annotations simultaneously." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.397, + 0.128 + ], + "angle": 0, + "content": "T. Salzmann et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.389 + ], + "angle": 0, + "content": "Long-Tailed and Open-Vocabulary Relationship Detection. Object classes in the natural world follow a long-tailed distribution [12,62], and this is compounded in VRD due to the combination of subject, object, and predicate in a relationship [27,48,49]. A large number of VRD methods specifically aim to address this issue, for example with debiasing losses [52], data augmentation [57], or data resampling [8]. Additionally, while VRD models historically assumed a fixed set of object and predicate classes, recent works add open-vocabulary capabilities [13,41,54,61]. These models leverage pretrained vision-language models such as CLIP [18, 37, 51, 56] to inherit their natural-language understanding. Instead of learning a fixed set of classes, open-vocabulary models predict class embeddings that can be compared to the text embeddings of arbitrary object or predicate descriptions for classification. Here, we employ a pretrained vision-language backbone and combine it with a carefully designed, lightweight relationship detection head to preserve and transfer semantic knowledge from the backbone pretraining to VRD. This allows our model to achieve strong results on the large-vocabulary GQA benchmark [16] (Section 4.3)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.389, + 0.788, + 0.465 + ], + "angle": 0, + "content": "While we focus on architectural improvements, another line of work addresses the scarcity of training data for VRD. For example, RLIP [53] and RLIPv2 [54], which focus on human-object interaction [4], propose relationship-focused image-text pretraining and self-training, which yield large improvements in VRD performance. These works are orthogonal to our contributions and can be combined." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.487, + 0.424, + 0.506 + ], + "angle": 0, + "content": "3 Scene-Graph ViT" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.52, + 0.788, + 0.626 + ], + "angle": 0, + "content": "We propose an architecture for open-vocabulary relationship prediction in which both objects and the relationships between them are handled as \"first-class citizens\" directly in a single-stage process within the model backbone. We build on an encoder-only architecture for object detection [34] and adapt it to relationship prediction by adding a specialized attention layer. This layer exploits the pairwise structure between existing object embeddings to obtain relationship embeddings as schematized in Figure 2 and described below." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.648, + 0.673, + 0.665 + ], + "angle": 0, + "content": "3.1 Encoder-Only Open-Vocabulary Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.841 + ], + "angle": 0, + "content": "We briefly review the detection architecture that forms the basis of our model [34]. The model consists of Transformer-based [10] image and text encoders that are contrastively pretrained on a large number of image and text pairs [37]. The image encoder is adapted to detection by removing the final pooling layer and adding heads that predict bounding boxes and class embeddings directly from the output tokens of the image encoder. For open-vocabulary object classification, the embeddings computed by the class prediction head are compared to text encoder embeddings of object descriptions. This architecture achieves strong open-vocabulary object detection performance [33, 34] and does not suffer from the training instabilities observed in some encoder-decoder detection models [3, 34, 43, 50, 58, 63]." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.613, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Scene-Graph ViT" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.264, + 0.147, + 0.74, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.257, + 0.705, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.393, + 0.788, + 0.532 + ], + "angle": 0, + "content": "Figure 2: For relationship selection, image tokens are first projected using two lightweight MLPs to produce and embeddings. A relationship score is then computed as the inner product between all and embeddings. Relationships are filtered by first selecting the top object instances, using the scores along the diagonal to represent instance likelihood. Among the remaining instances, the top pairs are selected using the off-diagonal scores. This yields a set of relationship triplets, each consisting of a index, an index, and a relationship embedding that is computed by summing the respective and embeddings. For classification, the relationship embeddings are compared against text embeddings of object class or predicate text descriptions." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.56, + 0.569, + 0.575 + ], + "angle": 0, + "content": "3.2 Extension to Relationship Prediction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.788, + 0.734 + ], + "angle": 0, + "content": "In the architecture described above, each image encoder token represents an object proposal that captures per-object information. Importantly, the encoder consists of self-attention layers which, by design, model all pairwise interactions between these tokens. We therefore hypothesize that information about object relationships can be learned directly in the image encoder. To extract this information in the form of triplets, we introduce two MLPs that transform vision encoder output tokens into embeddings \\(\\mathbf{s}_i\\) and embeddings \\(\\mathbf{o}_j\\) (Figure 2). Using different MLPs for and is crucial to break symmetry in subsequent processing (Appendix A.1)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.843 + ], + "angle": 0, + "content": "An embedding representing the relationship between two object proposals (encoder tokens) can then be obtained simply by element-wise addition of their and embeddings. Obtaining relationship embeddings for all \\( N^2 \\) pairs, where \\( N \\) is the number of object proposals, would be computationally infeasible. We therefore introduce a Relationship Attention layer that performs hard attention to select the pairs most likely to form a relationship. This layer computes an attention-like score \\( p_{ij} = \\)" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.397, + 0.128 + ], + "angle": 0, + "content": "T. Salzmann et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.298 + ], + "angle": 0, + "content": "\\(\\mathbf{s}_i \\mathbf{o}_j^T\\), where \\(\\mathbf{s}_i\\) and \\(\\mathbf{o}_j\\) are the embedding vectors of \\(<\\text { subject }>\\) \\(i\\) and \\(<\\text { object }>\\) \\(j\\), and \\(p_{ij}\\) represents the likelihood that a relationship between \\(<\\text { subject }>\\) \\(i\\) and \\(<\\text { object }>\\) \\(j\\) exists. By computing this score for all \\(<\\text { subject - object }>\\) pairs, we obtain an \\(N \\times N\\) matrix, from which we select the top \\(k\\) pairs. For these \\(k\\) pairs, we compute relationship embeddings \\(\\mathbf{r}_{ij} = \\text { LayerNorm }(\\mathbf{s}_i + \\mathbf{o}_j)\\). To classify the relationship of a pair, its relationship embedding is compared to text embeddings of relationship predicates. We use the embeddings where \\(<\\text { subject }>\\) and \\(<\\text { object }>\\) are identical (\\(\\mathbf{r}_{i=j}\\)) to represent object instances, and predict object classes from them. Boxes are predicted from the corresponding image encoder output tokens (see Figure 2)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.298, + 0.788, + 0.405 + ], + "angle": 0, + "content": "The Relationship Attention layer therefore identifies object pairs to focus on for relationship classification, by computing a hard attention with embeddings as queries and embeddings as keys. Since the hard selection is not differentiable, gradients from the relationship prediction cannot be used directly to train this layer. Instead, the relationship score is self-supervised to predict the maximum class probability that will later be predicted for the relationship at the classification stage (Section 3.3)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.428, + 0.334, + 0.443 + ], + "angle": 0, + "content": "3.3 Training" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.453, + 0.79, + 0.543 + ], + "angle": 0, + "content": "The image and text encoders of our model are initialized from a vision-language model that was contrastively pretrained on a large number of image-text pairs [37]. After adding the Relationship Attention as well as the object bounding box and class prediction heads, the model is jointly trained on a mixture of object and relationship detection datasets in a single training stage (Section 4.1), using the losses described below." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.544, + 0.788, + 0.65 + ], + "angle": 0, + "content": "Bipartite Matching. Bipartite matching between object predictions \\((\\mathbf{r}_{i = j})\\) and ground-truth annotations is performed based on a cost consisting of the object classification loss and the box prediction losses as in [3]. This matching between objects also establishes a matching of predicted to ground-truth relationship predicates, since a relationship is uniquely identified by the and indices \\(i\\) and \\(j\\). Unmatched predictions are trained to predict low scores for all classes and incur no box prediction loss." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.65, + 0.788, + 0.681 + ], + "angle": 0, + "content": "Box Prediction Losses. For bounding box regression, we use the L1 and generalized intersection-over-union (gIoU) losses described in [3]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.681, + 0.788, + 0.803 + ], + "angle": 0, + "content": "Object and Predicate Classification Loss. For both object category and predicate classification, we use a sigmoid cross-entropy loss as in [34]. This loss is computed between the ground truth classes and logits obtained by computing the inner product of the relationship embeddings \\(\\mathbf{r}_{ij}\\) selected by the Relationship Attention layer with the text embeddings of the class names. For embeddings corresponding to individual objects \\((\\mathbf{r}_{i = j})\\), the class name is the object category or description. For embeddings corresponding to relationships \\((\\mathbf{r}_{i\\neq j})\\), we use the predicate text as class name. This differs from prior work [53, 54, 61], which" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.787, + 0.842 + ], + "angle": 0, + "content": "This selection reduces the number of relationships that need to be processed, e.g., from \\( N^2 = 3600^2 \\) to \\( k = 16386 \\) (99.9% reduction) for our ViT-L/14 model." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.613, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Scene-Graph ViT" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.207 + ], + "angle": 0, + "content": "uses the full triplet description. Our approach has the advantage of disentangling object category and predicate names, which allows for more efficient inference since object category and predicate texts are embedded separately." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.789, + 0.373 + ], + "angle": 0, + "content": "Relationship Score Loss. To select only the most promising embeddings for further processing, the Relationship Attention layer predicts a score \\( p_{ij} \\) that represents the likelihood that the corresponding pair forms a relationship (if \\( i \\neq j \\)), or that the corresponding object exists in the image (if \\( i = j \\)). This score is trained with a sigmoid cross-entropy loss against targets provided by the model itself, namely the maximum probability predicted for any class for the corresponding \\( \\mathbf{r}_{ij} \\) embedding [33]. In this way, the relationship score is trained to predict the class probability of a potential embedding \\( \\mathbf{r}_{ij} \\) before actually computing and classifying the embedding. Note that this loss can only be computed for those objects and relationships that ultimately get selected for further processing. We found this to provide sufficient supervision." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.374, + 0.788, + 0.406 + ], + "angle": 0, + "content": "The final loss is an equally weighted sum of four components: (1) classification loss, (2) L1 box loss, (3) gIoU box loss, (4) relationship score loss." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.429, + 0.377, + 0.447 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.462, + 0.429, + 0.479 + ], + "angle": 0, + "content": "4.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.489, + 0.706, + 0.504 + ], + "angle": 0, + "content": "Datasets. We use the following datasets for training or evaluation:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.504, + 0.789, + 0.594 + ], + "angle": 0, + "content": "Visual Genome [26] is the largest VRD dataset, labeled with 2.3M triplet relationships across 108K images. However, as Visual Genome includes noisy annotations, the community commonly evaluates VRD on a cleaned version of the dataset [55] where the label space is reduced to the 150 most frequent object classes and the 50 most frequent predicate classes. This dataset is commonly referred to as Visual Genome 150 (VG150)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.595, + 0.789, + 0.685 + ], + "angle": 0, + "content": "GQA [16] uses the same image corpus as Visual Genome, but is more diversely labeled, with 1703 object classes (1704 including a background class) and 310 predicates. Spatial relationships like \"to the left of\" are automatically labeled based on the 2-D spatial arrangement of bounding boxes. Similar to VQ150, GQA200 is a reduced and cleaned version of GQA with 200 object classes and 100 predicate classes [9]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.686, + 0.789, + 0.746 + ], + "angle": 0, + "content": "HICO-DET [4] is a dataset which focuses on a more specialized subset of visual relationships, specifically interactions between humans and objects. HICO-DET contains 50k images and is exhaustively annotated for 600 defined human-object interaction (HOIs)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.746, + 0.789, + 0.777 + ], + "angle": 0, + "content": "Objects365 [40] (O365) is a large scale object-detection dataset with 365 object categories across two million images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.777, + 0.789, + 0.853 + ], + "angle": 0, + "content": "Open X-Embodiment [35] (OXE) dataset is targeted towards learning vision language action policies for robotics. OXE lacks bounding box annotations which precludes quantitative evaluation, but it represents both a visually different setting compared to the standard VRD benchmarks and a promising application area, so we use it for qualitative evaluation." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.397, + 0.128 + ], + "angle": 0, + "content": "T. Salzmann et al." + }, + { + "type": "table", + "bbox": [ + 0.232, + 0.144, + 0.773, + 0.19 + ], + "angle": 0, + "content": "
VG150VGHICOGQA200GQAObjects365
% in training mixture12.5%12.5%12.5%12.5%0%50%
# removed658717727729153611880115306
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.198, + 0.785, + 0.226 + ], + "angle": 0, + "content": "Table 1: Training data mixture. # removed indicates the number of images that were removed from the official training split due to overlap with data we evaluate on." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.243, + 0.788, + 0.396 + ], + "angle": 0, + "content": "Training. Unless otherwise noted, we train models on a mixture of VG, VG150, GQA200, HICO, and O365 in the proportions show in Table 1. The B/32 model is trained for 200'000 steps at batch size 256 (further details in Appendix A.1). Since the datasets we use share similar image sources, there is some overlap between the official training and evaluation splits. To ensure that no evaluation images are used for training, we rigorously filter all training datasets to remove images present in any of the test splits for all datasets we evaluate on, i.e. VG150, GQA, GQA200, and HICO. We use an image similarity filter that also identifies non-exact matches [24]. Table 1 shows the number of images removed from each training split. Note that not all prior work performs similar dedduplication." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.41, + 0.789, + 0.546 + ], + "angle": 0, + "content": "Evaluation Procedure and Metrics. Prior work presents a diverse range of evaluation metrics, often tailored to specific datasets. For non-exhaustively labelled datasets, like Visual Genome and GQA, using precision-based metrics is inherently inconclusive. Thus, the community has adapted a \\(Recall@K\\) metric for these datasets, where \\(K\\) denotes a fixed budget of triplets on which the recall is computed. However, this metric is biased towards the more frequent classes in the dataset [6,46]. Therefore, recent approaches use mean \\(Recall@K\\), which calculates \\(Recall@K\\) separately for each class in the test data and then averages the results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.547, + 0.789, + 0.623 + ], + "angle": 0, + "content": "For exhaustively labeled datasets like HICO, it is feasible to use precision-based metrics such as mean Average Precision (mAP), a well-established metric in object detection. For HICO, the mean is taken over the 600 possible HOI triplets. Besides the overall metric, we also report results separately for rare (< 10 occurrences) and non-rare (\\( \\geq 10 \\) occurrences) HOIs as mAPr and mAPn." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.624, + 0.789, + 0.76 + ], + "angle": 0, + "content": "Further, evaluation can be \"graph-constrained\", where only a single prediction can be made per detected object pair, or \"graph-unconstrained\", where any number of predictions can be made per pair. Existing works evaluating on Visual Genome and GQA datasets partially lack a clear distinction between graph-constrained and unconstrained evaluation. For our models, we report graph-constrained results in the main paper, which reflects the dominant approach used in other studies. Results without graph-constraint are reported in Appendix A.2. For the HICO dataset, we follow the established evaluation procedure [4] that inherently represents a graph-unconstrained evaluation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.761, + 0.789, + 0.852 + ], + "angle": 0, + "content": "Multiple implementations of the aforementioned metrics exist, often showing differences in results. To compare fairly and directly against a wide range of prior approaches we identified the PyTorch evaluation procedure of [45] as the most prevalent routine for recall-based metrics and replicated their procedure in JAX with numerical accuracy. Similarly, for HICO, we numerically reproduce the original Matlab evaluation procedure outlined in [4] in JAX." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.613, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Scene-Graph ViT" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table", + "bbox": [ + 0.271, + 0.144, + 0.728, + 0.334 + ], + "angle": 0, + "content": "
ModelBackbonemR@50mR@100
RelTR [7]ResNet-506.810.8
Transformer + CFA [28]ResNeXt-10112.314.6
VCTree [46] + IETrans + Rwt [57]ResNeXt-10112.014.9
Motif [55] + IETrans + Rwt [57]ResNeXt-10115.518.0
GPS-Net [31] + IETrans + Rwt [57]ResNeXt-10116.218.8
SG-Transformer [52] + IETrans + Rwt [57]ResNeXt-10115.018.1
Sgtr [29]ResNet-10115.820.1
DT2-ACBS [8]ResNet-10122.024.4
UniVRDCLIP: ViT-B/329.612.1
UniVRDCLIP: ViT-B/1610.913.2
UniVRDCLIP: ViT-L/1412.614.5
SG-ViTCLIP: ViT-B/3215.018.1
SG-ViTCLIP: ViT-B/1615.719.3
SG-ViTCLIP: ViT-L/1417.821.8
SG-ViT + simple predicate rebalancingCLIP: ViT-L/1422.326.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.34, + 0.788, + 0.396 + ], + "angle": 0, + "content": "Table 2: Performance on the Visual Genome 150 test set. We use graph-constrained evaluation (one predicate prediction per pair). For the last row (\"+ simple text rebalancing\"), the model was briefly fine-tuned on VG150 data with rebalanced predicate frequencies (see Section 4.2 and Appendix A.1)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.411, + 0.789, + 0.579 + ], + "angle": 0, + "content": "Exhaustive Relationship Evaluation. Previous methods [53,54,61] entangle objects and relationships in a single representation. Consequently, these methods score their embeddings against the full triplet. While this is feasible for datasets with a low number of possible triplets, like HICO with 600 defined relationships, it quickly becomes computationally intractable for datasets with a larger vocabulary of objects and predicates (e.g. VG, GQA). Prior works [61] therefore evaluate only on triplets which are known to be present in the test split, which may inflate metrics. In contrast, our method disentangles objects and predicates (Section 3). This allows for an exhaustive evaluation that considers all object and predicate combinations and enables applications where knowledge of possible relationships is unavailable." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.602, + 0.562, + 0.617 + ], + "angle": 0, + "content": "4.2 Relationship Detection Performance" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.788, + 0.794 + ], + "angle": 0, + "content": "We use Visual Genome as our main benchmark for relationship detection. Table 2 shows results for our models and prior work. Prior studies often achieve reported performance through the combination of a base model with a specialized approach to overcome the skewed data distribution (indicated as base-model + specialization in Table 2). Our models provide strong performance even without special treatment of the skewed predicate distribution. After brief fine-tuning on data in which the frequency of predicate annotations was rebalanced with simple rejection sampling (see Appendix A.1), our method improves on the prior best method DT2-ACBS [8] by 1.7 points mR@100 (26.1 vs. 24.4). This shows that our already-strong method may be further improved by combining it with more advanced rebalancing methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.84 + ], + "angle": 0, + "content": "We also note the large difference in performance to UniVRD [61], which builds on the same detection architecture as our method, but adds a Transformer decoder-based relationship module. To disentangle whether the improve" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.396, + 0.128 + ], + "angle": 0, + "content": "T. Salzmann et al." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.159, + 0.495, + 0.272 + ], + "angle": 0, + "content": "
ModelmR@50mR@100
LLM4SGG [21]5.36.5
Neural Motif w/ GCL [9]16.818.8
VCTree w/ GCL [9]15.617.8
SG-Transformer w/ CFA [28]13.415.3
SHA w/ GCL [9]17.820.1
SG-ViT (CLIP: ViT-B/32)15.718.6
SG-ViT (CLIP: ViT-B/16)17.720.5
SG-ViT (CLIP: ViT-L/14)19.322.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.276, + 0.487, + 0.317 + ], + "angle": 0, + "content": "Table 3: Performance on the GQA200 test set. Graph-constrained evaluation." + }, + { + "type": "table", + "bbox": [ + 0.513, + 0.144, + 0.778, + 0.288 + ], + "angle": 0, + "content": "
ModelmR@50mR@100
Scene Graph Classification
IMP [25,44,48]0.50.7
Neural Motif [25,44,55]0.81.2
Unbiased TDE [25,45]-0.7
RTN [25]-1.4
MP [23]-2.8
Transformer w/ EBM [44]1.31.8
Scene Graph Generation
SG-ViT (CLIP: ViT-B/32)6.27.4
SG-ViT (CLIP: ViT-B/16)6.47.4
SG-ViT (CLIP: ViT-L/14)8.09.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.29, + 0.782, + 0.318 + ], + "angle": 0, + "content": "Table 4: Performance on the GQA test set. Graph-constrained evaluation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.333, + 0.785, + 0.41 + ], + "angle": 0, + "content": "ments over UniVRD are due to differences in architecture or training data, we also trained our B/32 model on the UniVRD data mixture, and still observe a large improvement (17.4 VG mR@100 for our method vs. 12.1 for UniVRD when trained on the same data; Table 6). This suggests that our encoder-only architecture is better suited for VRD than prior decoder-based architectures [7,20,29,61]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.434, + 0.506, + 0.449 + ], + "angle": 0, + "content": "4.3 Scaling to Large Vocabularies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.788, + 0.643 + ], + "angle": 0, + "content": "Improving long-tail performance has been critical to the VRD field since the natural distribution of relationship triplets is highly skewed [27]. A substantial body of literature is devoted to this goal, often proposing complex loss debiasing or data augmentation approaches [8,27,52,57]. It is therefore of interest whether our method works well in this regime without special treatment of rare classes. To evaluate the performance of our method on large object and predicate vocabularies with a long tail of rare classes, we use the Graph Question Answering (GQA) dataset [16]. This dataset builds on VG and expands the number of classes and annotations. A simplified version of GQA with 200 object and 100 predicate classes, called GQA200 [9], is part of our training mixture. Our method surpasses prior results on the GQA200 test split (Table 3), despite lacking the specialized treatment of the data distribution that some prior methods use." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.749 + ], + "angle": 0, + "content": "The full GQA dataset has an even larger vocabulary with 1703 object and 311 predicate classes. To our knowledge, no prior methods evaluate scene graph generation on the full GQA dataset. Some prior methods report results on scene graph classification, where ground-truth boxes are provided. As shown in Table 4, our model achieves higher performance on the much harder task of scene graph generation than prior methods on classification. We further assess zero-shot generalization to unseen classes in Appendix A.2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We believe that two factors contribute to the strong performance in the open vocabulary regime: First, the simple design of the method does not impede transfer of semantic knowledge from the pre-trained VLM backbone because it adds only lightweight heads. In particular, we use no decoder, which could be difficult to train and could lead to \"forgetting\" of pretrained representations. Instead, most of the relationship modeling happens in the backbone, which has" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.613, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Scene-Graph ViT" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "had the benefit of large-scale vision-language pretraining. Second, the open-vocabulary design allows end-to-end training on a mix of datasets, which was not possible for all prior methods." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.216, + 0.483, + 0.232 + ], + "angle": 0, + "content": "4.4 Human-Object Interaction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.242, + 0.788, + 0.394 + ], + "angle": 0, + "content": "We also evaluate our model on the specialized task of human-object interaction, using the HICO benchmark [4]. The performance of our model is on par with the most comparable prior method [61], but does not show similarly large improvements as for VG or GQA200 (Table 5). HICO has a much narrower and more specialized vocabulary than VG or GQA (HICO evaluates on 600 prespecified triplets, whereas VG150 has \\(1.1 \\times 10^{6}\\) possible triplet combinations). Performance on this specialized task may benefit less from transfer of pretrained representations and may instead be limited by the amount of task-specific training data. This is supported by the fact that the recent state-of-the-art method for HICO pretrains on large amounts of person-focused pseudo-labels [54]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.416, + 0.345, + 0.43 + ], + "angle": 0, + "content": "4.5 Ablations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.439, + 0.603, + 0.498 + ], + "angle": 0, + "content": "Inference Speed and Number of Predicted Relationships. In the following, we ablate training datasets and model components to study the interplay of data and architecture on the model performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.5, + 0.603, + 0.589 + ], + "angle": 0, + "content": "The primary hyperparameter of our method is the number of relationships \\( k \\) that are selected by the Relationship Attention layer from all possible combinations for further processing. Figure 3 shows how VRD performance and inference speed depend on \\( k \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.59, + 0.603, + 0.741 + ], + "angle": 0, + "content": "For training, we use \\( k = 2^{14} \\), which results in \\( 71\\% \\) (B/32) to \\( 81\\% \\) (L/14) of the speed of a pure object detection model [34] of the same size. Without top-k selection, the model would be several times slower (right end of top plot in Figure 3). Since the vast majority of pairs do not form relationships, VRD performance is essentially unaffected by top- \\( k \\) selection at inference. For inference, we can therefore use \\( k = 2^{11} \\) at no loss in accuracy to achieve \\( 90\\% \\) of the speed of the object detector (for B/32 and" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.741, + 0.787, + 0.772 + ], + "angle": 0, + "content": "L/14). At this \\( k \\), the B/32 model achieves 52.8 FPS at batch size 1 on an NVIDIA V100 GPU and is therefore suitable for real-time applications." + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.445, + 0.784, + 0.52 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.528, + 0.784, + 0.628 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.611, + 0.643, + 0.788, + 0.728 + ], + "angle": 0, + "content": "Figure 3: Model speed and VRD accuracy by number of selected relationships \\( k \\). Speed is relative to a non-VRD object detector [34]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Dataset Mixture. To assess how training data and architectural improvements contribute to the performance of our model compared to prior work, we train the models on a range of dataset mixtures (Table 6). The full mixture, which we use for all experiments unless otherwise noted, includes VG, VG150, GQA200," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.395, + 0.127 + ], + "angle": 0, + "content": "T. Salzmann et al." + }, + { + "type": "table", + "bbox": [ + 0.352, + 0.144, + 0.653, + 0.273 + ], + "angle": 0, + "content": "
ModelBackbonemAPmAPrmAPn
HOTR [20]ResNet5025.1020.3325.86
RLIP [53]ResNet5032.8426.8534.63
RLIPv2 [54]ResNet5035.3829.6137.10
RLIPv2 [54]Swin-L45.0943.2345.64
UniVRD [60]CLIP: ViT-B/3229.9822.9432.02
UniVRD [60]CLIP: ViT-B/1629.9822.9432.02
UniVRD [60]CLIP: ViT-L/1437.4128.9039.95
SG-ViTCLIP: ViT-B/3228.8623.7230.39
SG-ViTCLIP: ViT-B/1631.9826.8333.52
SG-ViTCLIP: ViT-L/1438.1133.7139.42
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.282, + 0.788, + 0.354 + ], + "angle": 0, + "content": "Table 5: Performance on the HICO test set. Our method is on par with the most comparable prior work, UniVRD, for overall mAP. However, our method performs better on \"rare\" classes (mAPr), indicating better generalization to rarely seen concepts. For reference, we also show RLIPv2, the state-of-the-art method on HICO. RLIPv2 proposes a self-training approach that is orthogonal and compatible to our method." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.383, + 0.789, + 0.466 + ], + "angle": 0, + "content": "
VGMixtureVG150 mR@100GQA200 mR@100HICO mAP
VG150GQA200HICOO365COCO
Our mixture12.5%12.5%12.5%12.5%50%-18.118.628.9
UniVRD [61] mix-10%-20%50%20%17.4 ↓ 4%5.1 ↓ 73%17.4 ↓ 8%
ours w/o O36525%25%25%25%--15.3 ↓ 15%15.6 ↓ 16%23.8 ↓ 18%
ours w/o VG-25%12.5%12.5%50%-14.2 ↓ 22%15.3 ↓ 18%26.8 ↓ 7%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.476, + 0.788, + 0.505 + ], + "angle": 0, + "content": "Table 6: Dataset Ablation. Performance of SG-ViT (B/32) trained on different data mixtures. VG150 and GQA200 evaluation is graph-constrained, HICO is unconstrained." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.55, + 0.788, + 0.673 + ], + "angle": 0, + "content": "HICO and O365. To allow direct comparison to UniVRD [61], we also train on their mixture (VG150, HICO, O365, COCO) and find that our model still performs better on VG150 by a large margin (18.1 mR@100 vs 12.1 for UniVRD for B/32 models). Since UniVRD uses the same image and text encoding backbones as our model, this result suggests that our encoder-only model with Relationship Attention is an advance over decoder-based relationship architectures. We further find that including both pure object detection (O365) and large-vocabulary VRD data (VG) benefit all metrics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Object Detection Performance. A necessary step towards good relationship detection performance is accurate object detection. In our architecture, object detection is treated as a special case of relationship detection in which the is identical to the . We validate this design by training the model only on detection datasets without relationship annotations, using the same dataset mixture as the OWL-ViT [34] model (Objects365 and Visual Genome, [34]). We find that our model achieves a similar object detection performance as OWL-ViT (21.9% vs. 22.1% mAP and 18.3% vs. 18.9% mAPr on LVIS [12]), indicating that the Relationship Attention design does not interfere with object detection." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.613, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Scene-Graph ViT" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.143, + 0.76, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.469, + 0.788, + 0.514 + ], + "angle": 0, + "content": "Figure 4: Qualitative examples of difficult edge-cases from VG150 test split [26, 55]. From left to right: Ground Truth, SG-ViT (B/32), SG-ViT (L/14). In all cases the is lime and the is red." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.542, + 0.421, + 0.558 + ], + "angle": 0, + "content": "4.6 Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.788, + 0.719 + ], + "angle": 0, + "content": "In Figure 4, we visualize qualitative examples to highlight difficult edge cases. Even where the model output differs from the ground truth, it is often plausibly correct, e.g. where either the extent of an object differs with the ground truth or there are multiple appropriate, but unannotated, relations. For example, both models select the headlight on the bus in Figure 4b for \"light on bus\" as opposed to the sun reflected off the windshield. In Figure 4c both models focus more on the \"on\" in \"snow on hill\" than the distinction between a mountain and a hill. Figure 4d shows a challenging scene where the B/32 model selects the bottle inside the open fridge while the L/14 correctly selects a water bottle in the door for \"bottle in door\" while also capturing the extent of the door instance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Furthermore, to qualitatively assess the utility of this work beyond standard relation benchmark datasets, we also visualize some example relations applied to data from the robotics domain, namely the OXE dataset [35]. Figure 5 shows a real-life scene where multiple instances of the same semantic object categories are present within the scene. The model is able to match the to the instances in the desired configuration. One subjective failure-case is that the \"banana outside bowl\" text query had its highest scoring triplet involving a bowl on a shelf in the background rather than the adjacent bowl. Another charac" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.396, + 0.128 + ], + "angle": 0, + "content": "T. Salzmann et al." + }, + { + "type": "image", + "bbox": [ + 0.24, + 0.159, + 0.368, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.159, + 0.501, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.159, + 0.635, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.641, + 0.159, + 0.768, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.24, + 0.249, + 0.368, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.249, + 0.501, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.249, + 0.635, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.641, + 0.249, + 0.768, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.346, + 0.785, + 0.402 + ], + "angle": 0, + "content": "Figure 5: Qualitative examples showing SG-ViT (L/14) on out-of-distribution data from the OXE dataset [35]. In all cases the is lime and the is red. Note how the model correctly disambiguates several instances of the same class (e.g. \"bananas\" and \"bottle\") depending on their relationships." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.429, + 0.785, + 0.504 + ], + "angle": 0, + "content": "teristic is that the model occasionally places boxes over groups with multiple instances of the same semantic class when a singular word is used for either the or . This can be attributed to the similarity between word and their plural in the CLIP embeddings and that many human annotations from VG also confuse singular and plural (see Appendix A.3)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.523, + 0.364, + 0.538 + ], + "angle": 0, + "content": "5 Limitations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.551, + 0.788, + 0.673 + ], + "angle": 0, + "content": "While our model performs strongly on large-vocabulary VRD, its performance on specialized human-object interaction detection is only on par with prior models. We discuss this limitation in Section 4.4. Further error modes are explored qualitatively in Section 4.6 and Appendix A.3. A challenge for open-vocabulary relationship detection models as a whole is zero-shot generalization to unseen objects and predicates. While our model improves over prior approaches (Tables 3 and 4), there is still a large gap when it comes to entirely unseen classes (Appendix A.2). Future research should focus on closing this gap." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.692, + 0.36, + 0.708 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "We present an architecture for open-vocabulary visual relationship detection. By combining an encoder-only design with a novel Relationship Attention layer for efficient selection of high-confidence relationships, our architecture leverages VLM pretraining and multi-dataset VRD training. It achieves strong performance on standard and large-vocabulary VRD benchmarks while maintaining pure object detection performance and adding little extra computational cost. Due to its simplicity and strong performance, we believe that our method will be a useful basis for further research on visual relationship detection." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.613, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Scene-Graph ViT" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.176, + 0.787, + 0.218 + ], + "angle": 0, + "content": "1. Amiri, S., Chandan, K., Zhang, S.: Reasoning with scene graphs for robot planning under partial observability. IEEE Robotics and Automation Letters 7(2), 5560-5567 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.219, + 0.787, + 0.245 + ], + "angle": 0, + "content": "2. Ba, J.L., Kiros, J.R., Hinton, G.E.: Layer normalization. arXiv preprint arXiv:1607.06450 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.246, + 0.787, + 0.287 + ], + "angle": 0, + "content": "3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: Conference on Computer Vision (ECCV). pp. 213-229. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.288, + 0.787, + 0.328 + ], + "angle": 0, + "content": "4. Chao, Y.W., Liu, Y., Liu, X., Zeng, H., Deng, J.: Learning to Detect Human-Object Interactions. In: IEEE Winter Conference on Applications of Computer Vision (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.329, + 0.787, + 0.384 + ], + "angle": 0, + "content": "5. Chen, B., Xia, F., Ichter, B., Rao, K., Gopalakrishnan, K., Ryoo, M.S., Stone, A., Kappler, D.: Open-vocabulary queryable scene representations for real world planning. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 11509-11522. IEEE (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.384, + 0.787, + 0.425 + ], + "angle": 0, + "content": "6. Chen, T., Yu, W., Chen, R., Lin, L.: Knowledge-embedded routing network for scene graph generation. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 6163-6171 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.426, + 0.787, + 0.467 + ], + "angle": 0, + "content": "7. Cong, Y., Yang, M.Y., Rosenhahn, B.: Reltr: Relation transformer for scene graph generation. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.468, + 0.787, + 0.509 + ], + "angle": 0, + "content": "8. Desai, A., Wu, T.Y., Tripathi, S., Vasconcelos, N.: Learning of visual relations: The devil is in the tails. In: IEEE/CVF International Conference on Computer Vision (ICCV). pp. 15404-15413 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.509, + 0.787, + 0.564 + ], + "angle": 0, + "content": "9. Dong, X., Gan, T., Song, X., Wu, J., Cheng, Y., Nie, L.: Stacked hybrid-attention and group collaborative learning for unbiased scene graph generation. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 19427-19436 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.565, + 0.787, + 0.619 + ], + "angle": 0, + "content": "10. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. ICLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.62, + 0.787, + 0.674 + ], + "angle": 0, + "content": "11. Gu, Q., Kuwajerwala, A., Morin, S., Jatavallabhula, K.M., Sen, B., Agarwal, A., Rivera, C., Paul, W., Ellis, K., Chellappa, R., et al.: Conceptgraphs: Open-vocabulary 3d scene graphs for perception and planning. arXiv preprint arXiv:2309.16650 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.675, + 0.787, + 0.716 + ], + "angle": 0, + "content": "12. Gupta, A., Dollar, P., Girshick, R.: LVIS: A Dataset for Large Vocabulary Instance Segmentation. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.717, + 0.787, + 0.758 + ], + "angle": 0, + "content": "13. He, T., Gao, L., Song, J., Li, Y.F.: Towards Open-Vocabulary Scene Graph Generation with Prompt-Based Finetuning. In: European Conference on Computer Vision (ECCV) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.759, + 0.787, + 0.785 + ], + "angle": 0, + "content": "14. Hendrycks, D., Gimpel, K.: Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.786, + 0.787, + 0.84 + ], + "angle": 0, + "content": "15. Herzig, R., Mendelson, A., Karlinsky, L., Arbelle, A., Feris, R., Darrell, T., Globerson, A.: Incorporating Structured Representations into Pretrained Vision & Language Models Using Scene Graphs. In: The 2023 Conference on Empirical Methods in Natural Language Processing (2023)" + }, + { + "type": "list", + "bbox": [ + 0.219, + 0.176, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.396, + 0.127 + ], + "angle": 0, + "content": "T. Salzmann et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "16. Hudson, D.A., Manning, C.D.: Gqa: A new dataset for real-world visual reasoning and compositional question answering. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 6700-6709 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.233 + ], + "angle": 0, + "content": "17. Hughes, N., Chang, Y., Carlone, L.: Hydra: A Real-time Spatial Perception System for 3D Scene Graph Construction and Optimization. Robotics: Science and Systems (RSS) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.234, + 0.787, + 0.288 + ], + "angle": 0, + "content": "18. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International conference on machine learning. pp. 4904-4916. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.289, + 0.787, + 0.331 + ], + "angle": 0, + "content": "19. Johnson, J., Krishna, R., Stark, M., Li, L.J., Shamma, D., Bernstein, M., Fei-Fei, L.: Image Retrieval Using Scene Graphs. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.332, + 0.787, + 0.374 + ], + "angle": 0, + "content": "20. Kim, B., Lee, J., Kang, J., Kim, E.S., Kim, H.J.: Hotr: End-to-end human-object interaction detection with transformers. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 74-83 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.375, + 0.787, + 0.416 + ], + "angle": 0, + "content": "21. Kim, K., Yoon, K., Jeon, J., In, Y., Moon, J., Kim, D., Park, C.: LLM4SGG: Large Language Model for Weakly Supervised Scene Graph Generation. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.417, + 0.787, + 0.445 + ], + "angle": 0, + "content": "22. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.446, + 0.787, + 0.487 + ], + "angle": 0, + "content": "23. Knyazev, B., de Vries, H., Cangea, C., Taylor, G.W., Courville, A., Belilovsky, E.: Graph Density-Aware Losses for Novel Compositions in Scene Graph Generation. In: British Machine Vision Conference (BMVC) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.787, + 0.53 + ], + "angle": 0, + "content": "24. Kolesnikov, A., Beyer, L., Zhai, X., Puigcerver, J., Yung, J., Gelly, S., Houlsby, N.: Big transfer (bit): General visual representation learning. In: European Conference on Computer Vision (ECCV). pp. 491-507 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.531, + 0.787, + 0.558 + ], + "angle": 0, + "content": "25. Koner, R., Shit, S., Tresp, V.: Relation transformer network. arXiv preprint arXiv:2004.06193 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.559, + 0.787, + 0.614 + ], + "angle": 0, + "content": "26. Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., et al.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. International Journal of Computer Vision 123, 32-73 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.615, + 0.787, + 0.669 + ], + "angle": 0, + "content": "27. Li, H., Zhu, G., Zhang, L., Jiang, Y., Dang, Y., Hou, H., Shen, P., Zhao, X., Shah, S.A.A., Bennamoun, M.: Scene Graph Generation: A comprehensive survey. Neurocomputing (2024). https://doi.org/https://doi.org/10.1016/j.neucom.2023.127052" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.671, + 0.787, + 0.713 + ], + "angle": 0, + "content": "28. Li, L., Chen, G., Xiao, J., Yang, Y., Wang, C., Chen, L.: Compositional feature augmentation for unbiased scene graph generation. In: IEEE/CVF International Conference on Computer Vision (ICCV). pp. 21685-21695 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.714, + 0.787, + 0.756 + ], + "angle": 0, + "content": "29. Li, R., Zhang, S., He, X.: Sgtr: End-to-end scene graph generation with transformer. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 19486-19496 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.757, + 0.787, + 0.798 + ], + "angle": 0, + "content": "30. Li, Y., Ouyang, W., Zhou, B., Wang, K., Wang, X.: Scene graph generation from objects, phrases and region captions. In: IEEE/CVF International Conference on Computer Vision (ICCV). pp. 1261-1270 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "31. Lin, X., Ding, C., Zeng, J., Tao, D.: Gps-net: Graph property sensing network for scene graph generation. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3746-3753 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.613, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Scene-Graph ViT" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "32. Lu, C., Krishna, R., Bernstein, M., Fei-Fei, L.: Visual relationship detection with language priors. In: European Conference on Computer Vision (ECCV). pp. 852-869. Springer (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.787, + 0.217 + ], + "angle": 0, + "content": "33. Minderer, M., Gritsenko, A., Houlsby, N.: Scaling open-vocabulary object detection. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.217, + 0.787, + 0.271 + ], + "angle": 0, + "content": "34. Minderer, M., Gritsenko, A., Stone, A., Neumann, M., Weissenborn, D., Dosovitskiy, A., Mahendran, A., Arnab, A., Dehghani, M., Shen, Z., et al.: Simple open-vocabulary object detection. In: European Conference on Computer Vision (ECCV). pp. 728-755. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.271, + 0.787, + 0.298 + ], + "angle": 0, + "content": "35. Open X-Embodiment Collaboration: Open X-Embodiment: Robotic Learning Datasets and RT-X Models (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.298, + 0.787, + 0.338 + ], + "angle": 0, + "content": "36. Peng, Z., Wang, W., Dong, L., Hao, Y., Huang, S., Ma, S., Ye, Q., Wei, F.: Grounding Multimodal Large Language Models to the World. In: International Conference on Learning Representations (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.338, + 0.787, + 0.393 + ], + "angle": 0, + "content": "37. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.393, + 0.787, + 0.434 + ], + "angle": 0, + "content": "38. Rana, K., Haviland, J., Garg, S., Abou-Chakra, J., Reid, I., Suenderhauf, N.: Say-plan: Grounding large language models using 3d scene graphs for scalable robot task planning. In: 7th Annual Conference on Robot Learning (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.434, + 0.787, + 0.473 + ], + "angle": 0, + "content": "39. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems 28 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.473, + 0.787, + 0.515 + ], + "angle": 0, + "content": "40. Shao, S., Li, Z., Zhang, T., Peng, C., Yu, G., Zhang, X., Li, J., Sun, J.: Objects365: A large-scale, high-quality dataset for object detection. In: IEEE/CVF International Conference on Computer Vision (ICCV). pp. 8430-8439 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.515, + 0.787, + 0.555 + ], + "angle": 0, + "content": "41. Shi, H., Hayat, M., Cai, J.: Open-vocabulary object detection via scene graph discovery. In: 31st ACM International Conference on Multimedia. pp. 4012-4021 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.555, + 0.787, + 0.597 + ], + "angle": 0, + "content": "42. Singh, K.P., Salvador, J., Weihs, L., Kembhavi, A.: Scene graph contrastive learning for embodied navigation. In: IEEE/CVF International Conference on Computer Vision (ICCV). pp. 10884-10894 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.596, + 0.787, + 0.637 + ], + "angle": 0, + "content": "43. Song, H., Sun, D., Chun, S., Jampani, V., Han, D., Heo, B., Kim, W., Yang, M.H.: ViDT: An Efficient and Effective Fully Transformer-based Object Detector. In: International Conference on Learning Representations (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.637, + 0.787, + 0.678 + ], + "angle": 0, + "content": "44. Suhail, M., Mittal, A., Siddiquie, B., Broaddus, C., Eledath, J., Medioni, G., Sigal, L.: Energy-based learning for scene graph generation. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13936-13945 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.678, + 0.787, + 0.718 + ], + "angle": 0, + "content": "45. Tang, K., Niu, Y., Huang, J., Shi, J., Zhang, H.: Unbiased scene graph generation from biased training. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3716-3725 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.718, + 0.787, + 0.759 + ], + "angle": 0, + "content": "46. Tang, K., Zhang, H., Wu, B., Luo, W., Liu, W.: Learning to compose dynamic tree structures for visual contexts. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 6619-6628 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.759, + 0.787, + 0.8 + ], + "angle": 0, + "content": "47. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.8, + 0.787, + 0.841 + ], + "angle": 0, + "content": "48. Xu, D., Zhu, Y., Choy, C.B., Fei-Fei, L.: Scene graph generation by iterative message passing. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5410-5419 (2017)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.396, + 0.128 + ], + "angle": 0, + "content": "T. Salzmann et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "49. Yang, J., Lu, J., Lee, S., Batra, D., Parikh, D.: Graph r-cnn for scene graph generation. In: European Conference on Computer Vision (ECCV). pp. 670-685 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.785, + 0.203 + ], + "angle": 0, + "content": "50. Yao, Z., Ai, J., Li, B., Zhang, C.: Efficient DETR: Improving End-to-End Object Detector With Dense Prior. arXiv preprint arXiv:2104.01318 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.785, + 0.245 + ], + "angle": 0, + "content": "51. Yu, J., Wang, Z., Vasudevan, V., Yeung, L., Seyedhosseini, M., Wu, Y.: CoCa: Contrastive Captioners are Image-Text Foundation Models. Transactions on Machine Learning Research (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.785, + 0.299 + ], + "angle": 0, + "content": "52. Yu, J., Chai, Y., Wang, Y., Hu, Y., Wu, Q.: CogTree: Cognition Tree Loss for Unbiased Scene Graph Generation. In: International Joint Conference on Artificial Intelligence, IJCAI-21. International Joint Conferences on Artificial Intelligence Organization (8 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.785, + 0.342 + ], + "angle": 0, + "content": "53. Yuan, H., Jiang, J., Albaniae, S., Feng, T., Huang, Z., Ni, D., Tang, M.: Rlip: Relational language-image pre-training for human-object interaction detection. Advances in Neural Information Processing Systems 35, 37416-37431 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.785, + 0.397 + ], + "angle": 0, + "content": "54. Yuan, H., Zhang, S., Wang, X., Albanie, S., Pan, Y., Feng, T., Jiang, J., Ni, D., Zhang, Y., Zhao, D.: RLIPv2: Fast Scaling of Relational Language-Image Pretraining. In: IEEE/CVF International Conference on Computer Vision (ICCV). pp. 21649-21661 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.438 + ], + "angle": 0, + "content": "55. Zellers, R., Yatskar, M., Thomson, S., Choi, Y.: Neural motifs: Scene graph parsing with global context. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5831-5840 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.481 + ], + "angle": 0, + "content": "56. Zhai, X., Wang, X., Mustafa, B., Steiner, A., Keysers, D., Kolesnikov, A., Beyer, L.: Lit: Zero-shot transfer with locked-image text tuning. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 18123-18133 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.522 + ], + "angle": 0, + "content": "57. Zhang, A., Yao, Y., Chen, Q., Ji, W., Liu, Z., Sun, M., Chua, T.S.: Fine-grained scene graph generation with data transfer. In: European Conference on Computer Vision (ECCV). pp. 409-424. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.785, + 0.563 + ], + "angle": 0, + "content": "58. Zhang, H., Li, F., Liu, S., Zhang, L., Su, H., Zhu, J., Ni, L.M., Shum, H.Y.: Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.785, + 0.617 + ], + "angle": 0, + "content": "59. Zhang, Y., Ma, Z., Gao, X., Shakiah, S., Gao, Q., Chai, J.: GROUNDHOG: Grounding Large Language Models to Holistic Segmentation. In: Conference on Computer Vision and Pattern Recognition 2024 (2024), https://openreview.net/forum?id=PWW3ux1Ju" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.618, + 0.785, + 0.66 + ], + "angle": 0, + "content": "60. Zhao, L., Yuan, L., Gong, B., Cui, Y., Schroff, F., Yang, M.H., Adam, H., Liu, T.: Unified Visual Relationship Detection with Vision and Language Models. In: IEEE/CVF International Conference on Computer Vision (ICCV) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.66, + 0.785, + 0.702 + ], + "angle": 0, + "content": "61. Zhao, L., Yuan, L., Gong, B., Cui, Y., Schroff, F., Yang, M.H., Adam, H., Liu, T.: Unified Visual Relationship Detection with Vision and Language Models. In: IEEE/CVF International Conference on Computer Vision (ICCV) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.785, + 0.742 + ], + "angle": 0, + "content": "62. Zhou, X., Girdhar, R., Joulin, A., Krahenbuhl, P., Misra, I.: Detecting twenty-thousand classes using image-level supervision. In: European Conference on Computer Vision (ECCV). pp. 350-368 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.743, + 0.785, + 0.785 + ], + "angle": 0, + "content": "63. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable DETR: Deformable Transformers for End-to-End Object Detection. In: International Conference on Learning Representations (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.785 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Scene-Graph ViT_ End-to-End Open-Vocabulary Visual Relationship Detection/07085046-a430-402b-be81-31056599f4f5_origin.pdf b/2024/Scene-Graph ViT_ End-to-End Open-Vocabulary Visual Relationship Detection/07085046-a430-402b-be81-31056599f4f5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0962a76bdeeb57cca82d5dc48d906e53e0cd35e2 --- /dev/null +++ b/2024/Scene-Graph ViT_ End-to-End Open-Vocabulary Visual Relationship Detection/07085046-a430-402b-be81-31056599f4f5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:52bd20e4000d1c1c8a6a6298185fa33cf1bf3c507d4a358b82aee19d31b469d0 +size 11651071 diff --git a/2024/Scene-Graph ViT_ End-to-End Open-Vocabulary Visual Relationship Detection/full.md b/2024/Scene-Graph ViT_ End-to-End Open-Vocabulary Visual Relationship Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..deafb5cfb816a0fcaa5ec9ffebc432682b50ea3f --- /dev/null +++ b/2024/Scene-Graph ViT_ End-to-End Open-Vocabulary Visual Relationship Detection/full.md @@ -0,0 +1,286 @@ +# Scene-Graph ViT: End-to-End Open-Vocabulary Visual Relationship Detection + +Tim Salzmann $^{1,2}$ , Markus Ryll $^{2}$ , Alex Bewley $^{\star 1}$ , and Matthias Minderer $^{\star 1}$ + +$^{1}$ Google DeepMind {tsal, bewley, mjlm}@google.com $^{2}$ Technical University Munich {tim.salzmann, markus.ryll}@tum.d + +![](images/76dc13707de447c668c91d07b4346bd00fc74461a2918d44f1240e6c8f250fb0.jpg) +Figure 1: Relationships detected by our method on an unseen image. The top relationships by confidence score are shown. Photo by Spacejoy on Unsplash. + +![](images/f8a13735bc1e2ecd2dd20a03f79190938d5c6a1b92da04791fe196e43dbb5792.jpg) + +Abstract. Visual relationship detection aims to identify objects and their relationships in images. Prior methods approach this task by adding separate relationship modules or decoders to existing object detection architectures. This separation increases complexity and hinders end-to-end training, which limits performance. We propose a simple and highly efficient decoder-free architecture for open-vocabulary visual relationship detection. Our model consists of a Transformer-based image encoder that represents objects as tokens and models their relationships implicitly. To extract relationship information, we introduce an attention mechanism that selects object pairs likely to form a relationship. We provide a single-stage recipe to train this model on a mixture of object and relationship detection data. Our approach achieves state-of-the-art relationship detection performance on Visual Genome and on the large-vocabulary GQA benchmark at real-time inference speeds. We provide ablations, real-world qualitative examples, and analyses of zero-shot performance. + +# 1 Introduction + +A fundamental goal of computer vision is to decompose visual scenes into structured semantic representations. A commonly studied task towards this goal is object detection, in which objects in an image are localized by bounding boxes and classified into semantic categories. However, a scene description also includes the semantic relationships between objects. This gives rise to the task of visual relationship detection (VRD, [4,6,32,48,49,55]). In VRD, the model detects objects and infers pairwise relationships between them in the form of triplets3 [32]. + +Detecting both objects and their relationships allows the construction of a scene graph [19,30,48] in which objects are represented as nodes and their relationships as edges. Scene graph generation (SGG) has wide-ranging applications in robotics [1,5,11,17,38,42] and image retrieval [19,27]. Increasingly, structured scene representations are also employed to provide grounding and explainability to multimodal large language models [15,36,59]. + +Prior work typically draws a distinction between object detection and relationship prediction. Detection is performed either as a wholly separate step before relationship prediction [8, 32, 48, 52, 55, 57], or by separate model parts such as "relationship decoders" that are responsible for modeling the interactions between objects [7, 49, 61]. This distinction makes it hard to optimize such models end-to-end for VRD [61]. In contrast, we propose an encoder-only architecture that models objects and relationships jointly, directly in the image encoder. Our architecture performs open-vocabulary relationship detection and can be trained end-to-end on arbitrary mixtures of object detection and relationship annotations. + +Our model builds on a Transformer-based encoder-only object detector [34] in which the output tokens of the image encoder directly represent object proposals. From these tokens, class embeddings and bounding boxes are decoded with lightweight heads. Our insight is that this architecture is perfectly set up to learn relationships between objects directly in the image encoder, without the need for additional relationship-specific stages. This is because the existing self-attention in the encoder already models all-to-all pairwise interactions between the object proposal tokens. + +To access information about the relationship between two of these tokens, we combine the embeddings corresponding to the and token using a new Relationship Attention layer. Obtaining relationship embeddings for all possible pairwise combinations of object proposal tokens would be computationally infeasible. To reduce the number of combinations, we introduce a self-supervised hard attention mechanism that selects the highest-confidence pairs at a computational cost comparable to a single self-attention layer. We show how to directly supervise the attention scores of this mechanism without the need to propagate gradients through the hard selection. + +An additional benefit of our design is that it disentangles object names from relationship predicates during inference. In contrast to prior open-vocabulary methods [53, 54, 61], our model can embed object and predicate texts separately and generate confidence scores efficiently for all possible combinations. + +In summary, we make the following contributions: + +1. An efficient architecture for open-vocabulary visual relationship detection. +2. A single-stage recipe for joint object and relationship detection training. +3. Efficient, disentangled object and relationship inference. +4. Analysis of inference speed, ablations, and qualitative examples. + +Our method achieves state-of-the-art visual relationship detection performance on the Visual Genome dataset (26.1% mR@100 with graph constraint) and on the large-vocabulary GQA benchmark. It provides strong results while being significantly simpler than prior approaches to visual relationship detection. + +# 2 Related Work + +Below, we review the main approaches to VRD that are relevant to our work. For a comprehensive overview of the field, we refer to [27]. + +Detector-Agnostic Relationship Detection. Historically, many works on VRD assume that object detections are given, such that the task reduces to inferring relationships between the objects. These "detector-agnostic" methods use off-the-shelf detectors such as Faster-RCNN [39] to obtain boxes and box embeddings and infer a scene graph from them. Early work on large-scale VRD employed word embeddings to improve relationship generalization [32] and graph inference methods such as message passing for SGG [48]. More recent methods leverage relationship co-occurrence statistics [55] and address the long-tailed nature of the training data [8, 52, 57]. + +End-to-End Relationship Detection. A fundamental limitation of the detector-agnostic approach is that object detection and relationship prediction are treated separately and cannot learn from each other. Overcoming this limitation has recently motivated the development of end-to-end architectures in which objects and relationships are predicted jointly by a single model [7, 20, 29, 61]. All of these models have in common that they use a Transformer [47] decoder to predict object and relationship embeddings. The decoder represents object proposals with query embeddings that can cross-attend into image embeddings. The literature on object detection suggests that this choice may be problematic, because the query embeddings can be difficult to initialize, which can lead to unstable and slow optimization [3, 34, 43, 50, 58, 63]. To avoid this issue, we design an encoder-only architecture in which both object and relationship embeddings are computed directly from the output tokens of a Vision Transformer [10] image encoder, building on an idea from object detection [34]. In contrast to prior models [61], this allows us to train the model end-to-end on object and relationship annotations simultaneously. + +Long-Tailed and Open-Vocabulary Relationship Detection. Object classes in the natural world follow a long-tailed distribution [12,62], and this is compounded in VRD due to the combination of subject, object, and predicate in a relationship [27,48,49]. A large number of VRD methods specifically aim to address this issue, for example with debiasing losses [52], data augmentation [57], or data resampling [8]. Additionally, while VRD models historically assumed a fixed set of object and predicate classes, recent works add open-vocabulary capabilities [13,41,54,61]. These models leverage pretrained vision-language models such as CLIP [18, 37, 51, 56] to inherit their natural-language understanding. Instead of learning a fixed set of classes, open-vocabulary models predict class embeddings that can be compared to the text embeddings of arbitrary object or predicate descriptions for classification. Here, we employ a pretrained vision-language backbone and combine it with a carefully designed, lightweight relationship detection head to preserve and transfer semantic knowledge from the backbone pretraining to VRD. This allows our model to achieve strong results on the large-vocabulary GQA benchmark [16] (Section 4.3). + +While we focus on architectural improvements, another line of work addresses the scarcity of training data for VRD. For example, RLIP [53] and RLIPv2 [54], which focus on human-object interaction [4], propose relationship-focused image-text pretraining and self-training, which yield large improvements in VRD performance. These works are orthogonal to our contributions and can be combined. + +# 3 Scene-Graph ViT + +We propose an architecture for open-vocabulary relationship prediction in which both objects and the relationships between them are handled as "first-class citizens" directly in a single-stage process within the model backbone. We build on an encoder-only architecture for object detection [34] and adapt it to relationship prediction by adding a specialized attention layer. This layer exploits the pairwise structure between existing object embeddings to obtain relationship embeddings as schematized in Figure 2 and described below. + +# 3.1 Encoder-Only Open-Vocabulary Object Detection + +We briefly review the detection architecture that forms the basis of our model [34]. The model consists of Transformer-based [10] image and text encoders that are contrastively pretrained on a large number of image and text pairs [37]. The image encoder is adapted to detection by removing the final pooling layer and adding heads that predict bounding boxes and class embeddings directly from the output tokens of the image encoder. For open-vocabulary object classification, the embeddings computed by the class prediction head are compared to text encoder embeddings of object descriptions. This architecture achieves strong open-vocabulary object detection performance [33, 34] and does not suffer from the training instabilities observed in some encoder-decoder detection models [3, 34, 43, 50, 58, 63]. + +![](images/11851b4cb9f82a6929578514b366470e94ad3b87a5308ffb27f48cd4a4f5073a.jpg) + +![](images/f042bb5958900c574775ed1f1666401d8fbde29f3dc02abd99e8992bb68e8497.jpg) +Figure 2: For relationship selection, image tokens are first projected using two lightweight MLPs to produce and embeddings. A relationship score is then computed as the inner product between all and embeddings. Relationships are filtered by first selecting the top object instances, using the scores along the diagonal to represent instance likelihood. Among the remaining instances, the top pairs are selected using the off-diagonal scores. This yields a set of relationship triplets, each consisting of a index, an index, and a relationship embedding that is computed by summing the respective and embeddings. For classification, the relationship embeddings are compared against text embeddings of object class or predicate text descriptions. + +# 3.2 Extension to Relationship Prediction + +In the architecture described above, each image encoder token represents an object proposal that captures per-object information. Importantly, the encoder consists of self-attention layers which, by design, model all pairwise interactions between these tokens. We therefore hypothesize that information about object relationships can be learned directly in the image encoder. To extract this information in the form of triplets, we introduce two MLPs that transform vision encoder output tokens into embeddings $\mathbf{s}_i$ and embeddings $\mathbf{o}_j$ (Figure 2). Using different MLPs for and is crucial to break symmetry in subsequent processing (Appendix A.1). + +An embedding representing the relationship between two object proposals (encoder tokens) can then be obtained simply by element-wise addition of their and embeddings. Obtaining relationship embeddings for all $N^2$ pairs, where $N$ is the number of object proposals, would be computationally infeasible. We therefore introduce a Relationship Attention layer that performs hard attention to select the pairs most likely to form a relationship. This layer computes an attention-like score $p_{ij} =$ + +$\mathbf{s}_i \mathbf{o}_j^T$ , where $\mathbf{s}_i$ and $\mathbf{o}_j$ are the embedding vectors of $<\text { subject }>$ $i$ and $<\text { object }>$ $j$ , and $p_{ij}$ represents the likelihood that a relationship between $<\text { subject }>$ $i$ and $<\text { object }>$ $j$ exists. By computing this score for all $<\text { subject - object }>$ pairs, we obtain an $N \times N$ matrix, from which we select the top $k$ pairs. For these $k$ pairs, we compute relationship embeddings $\mathbf{r}_{ij} = \text { LayerNorm }(\mathbf{s}_i + \mathbf{o}_j)$ . To classify the relationship of a pair, its relationship embedding is compared to text embeddings of relationship predicates. We use the embeddings where $<\text { subject }>$ and $<\text { object }>$ are identical ( $\mathbf{r}_{i=j}$ ) to represent object instances, and predict object classes from them. Boxes are predicted from the corresponding image encoder output tokens (see Figure 2). + +The Relationship Attention layer therefore identifies object pairs to focus on for relationship classification, by computing a hard attention with embeddings as queries and embeddings as keys. Since the hard selection is not differentiable, gradients from the relationship prediction cannot be used directly to train this layer. Instead, the relationship score is self-supervised to predict the maximum class probability that will later be predicted for the relationship at the classification stage (Section 3.3). + +# 3.3 Training + +The image and text encoders of our model are initialized from a vision-language model that was contrastively pretrained on a large number of image-text pairs [37]. After adding the Relationship Attention as well as the object bounding box and class prediction heads, the model is jointly trained on a mixture of object and relationship detection datasets in a single training stage (Section 4.1), using the losses described below. + +Bipartite Matching. Bipartite matching between object predictions $(\mathbf{r}_{i = j})$ and ground-truth annotations is performed based on a cost consisting of the object classification loss and the box prediction losses as in [3]. This matching between objects also establishes a matching of predicted to ground-truth relationship predicates, since a relationship is uniquely identified by the and indices $i$ and $j$ . Unmatched predictions are trained to predict low scores for all classes and incur no box prediction loss. + +Box Prediction Losses. For bounding box regression, we use the L1 and generalized intersection-over-union (gIoU) losses described in [3]. + +Object and Predicate Classification Loss. For both object category and predicate classification, we use a sigmoid cross-entropy loss as in [34]. This loss is computed between the ground truth classes and logits obtained by computing the inner product of the relationship embeddings $\mathbf{r}_{ij}$ selected by the Relationship Attention layer with the text embeddings of the class names. For embeddings corresponding to individual objects $(\mathbf{r}_{i = j})$ , the class name is the object category or description. For embeddings corresponding to relationships $(\mathbf{r}_{i\neq j})$ , we use the predicate text as class name. This differs from prior work [53, 54, 61], which + +uses the full triplet description. Our approach has the advantage of disentangling object category and predicate names, which allows for more efficient inference since object category and predicate texts are embedded separately. + +Relationship Score Loss. To select only the most promising embeddings for further processing, the Relationship Attention layer predicts a score $p_{ij}$ that represents the likelihood that the corresponding pair forms a relationship (if $i \neq j$ ), or that the corresponding object exists in the image (if $i = j$ ). This score is trained with a sigmoid cross-entropy loss against targets provided by the model itself, namely the maximum probability predicted for any class for the corresponding $\mathbf{r}_{ij}$ embedding [33]. In this way, the relationship score is trained to predict the class probability of a potential embedding $\mathbf{r}_{ij}$ before actually computing and classifying the embedding. Note that this loss can only be computed for those objects and relationships that ultimately get selected for further processing. We found this to provide sufficient supervision. + +The final loss is an equally weighted sum of four components: (1) classification loss, (2) L1 box loss, (3) gIoU box loss, (4) relationship score loss. + +# 4 Experiments + +# 4.1 Experimental Setup + +Datasets. We use the following datasets for training or evaluation: + +Visual Genome [26] is the largest VRD dataset, labeled with 2.3M triplet relationships across 108K images. However, as Visual Genome includes noisy annotations, the community commonly evaluates VRD on a cleaned version of the dataset [55] where the label space is reduced to the 150 most frequent object classes and the 50 most frequent predicate classes. This dataset is commonly referred to as Visual Genome 150 (VG150). + +GQA [16] uses the same image corpus as Visual Genome, but is more diversely labeled, with 1703 object classes (1704 including a background class) and 310 predicates. Spatial relationships like "to the left of" are automatically labeled based on the 2-D spatial arrangement of bounding boxes. Similar to VQ150, GQA200 is a reduced and cleaned version of GQA with 200 object classes and 100 predicate classes [9]. + +HICO-DET [4] is a dataset which focuses on a more specialized subset of visual relationships, specifically interactions between humans and objects. HICO-DET contains 50k images and is exhaustively annotated for 600 defined human-object interaction (HOIs). + +Objects365 [40] (O365) is a large scale object-detection dataset with 365 object categories across two million images. + +Open X-Embodiment [35] (OXE) dataset is targeted towards learning vision language action policies for robotics. OXE lacks bounding box annotations which precludes quantitative evaluation, but it represents both a visually different setting compared to the standard VRD benchmarks and a promising application area, so we use it for qualitative evaluation. + +
VG150VGHICOGQA200GQAObjects365
% in training mixture12.5%12.5%12.5%12.5%0%50%
# removed658717727729153611880115306
+ +Table 1: Training data mixture. # removed indicates the number of images that were removed from the official training split due to overlap with data we evaluate on. + +Training. Unless otherwise noted, we train models on a mixture of VG, VG150, GQA200, HICO, and O365 in the proportions show in Table 1. The B/32 model is trained for 200'000 steps at batch size 256 (further details in Appendix A.1). Since the datasets we use share similar image sources, there is some overlap between the official training and evaluation splits. To ensure that no evaluation images are used for training, we rigorously filter all training datasets to remove images present in any of the test splits for all datasets we evaluate on, i.e. VG150, GQA, GQA200, and HICO. We use an image similarity filter that also identifies non-exact matches [24]. Table 1 shows the number of images removed from each training split. Note that not all prior work performs similar dedduplication. + +Evaluation Procedure and Metrics. Prior work presents a diverse range of evaluation metrics, often tailored to specific datasets. For non-exhaustively labelled datasets, like Visual Genome and GQA, using precision-based metrics is inherently inconclusive. Thus, the community has adapted a $Recall@K$ metric for these datasets, where $K$ denotes a fixed budget of triplets on which the recall is computed. However, this metric is biased towards the more frequent classes in the dataset [6,46]. Therefore, recent approaches use mean $Recall@K$ , which calculates $Recall@K$ separately for each class in the test data and then averages the results. + +For exhaustively labeled datasets like HICO, it is feasible to use precision-based metrics such as mean Average Precision (mAP), a well-established metric in object detection. For HICO, the mean is taken over the 600 possible HOI triplets. Besides the overall metric, we also report results separately for rare (< 10 occurrences) and non-rare ( $\geq 10$ occurrences) HOIs as mAPr and mAPn. + +Further, evaluation can be "graph-constrained", where only a single prediction can be made per detected object pair, or "graph-unconstrained", where any number of predictions can be made per pair. Existing works evaluating on Visual Genome and GQA datasets partially lack a clear distinction between graph-constrained and unconstrained evaluation. For our models, we report graph-constrained results in the main paper, which reflects the dominant approach used in other studies. Results without graph-constraint are reported in Appendix A.2. For the HICO dataset, we follow the established evaluation procedure [4] that inherently represents a graph-unconstrained evaluation. + +Multiple implementations of the aforementioned metrics exist, often showing differences in results. To compare fairly and directly against a wide range of prior approaches we identified the PyTorch evaluation procedure of [45] as the most prevalent routine for recall-based metrics and replicated their procedure in JAX with numerical accuracy. Similarly, for HICO, we numerically reproduce the original Matlab evaluation procedure outlined in [4] in JAX. + +
ModelBackbonemR@50mR@100
RelTR [7]ResNet-506.810.8
Transformer + CFA [28]ResNeXt-10112.314.6
VCTree [46] + IETrans + Rwt [57]ResNeXt-10112.014.9
Motif [55] + IETrans + Rwt [57]ResNeXt-10115.518.0
GPS-Net [31] + IETrans + Rwt [57]ResNeXt-10116.218.8
SG-Transformer [52] + IETrans + Rwt [57]ResNeXt-10115.018.1
Sgtr [29]ResNet-10115.820.1
DT2-ACBS [8]ResNet-10122.024.4
UniVRDCLIP: ViT-B/329.612.1
UniVRDCLIP: ViT-B/1610.913.2
UniVRDCLIP: ViT-L/1412.614.5
SG-ViTCLIP: ViT-B/3215.018.1
SG-ViTCLIP: ViT-B/1615.719.3
SG-ViTCLIP: ViT-L/1417.821.8
SG-ViT + simple predicate rebalancingCLIP: ViT-L/1422.326.1
+ +Table 2: Performance on the Visual Genome 150 test set. We use graph-constrained evaluation (one predicate prediction per pair). For the last row ("+ simple text rebalancing"), the model was briefly fine-tuned on VG150 data with rebalanced predicate frequencies (see Section 4.2 and Appendix A.1). + +Exhaustive Relationship Evaluation. Previous methods [53,54,61] entangle objects and relationships in a single representation. Consequently, these methods score their embeddings against the full triplet. While this is feasible for datasets with a low number of possible triplets, like HICO with 600 defined relationships, it quickly becomes computationally intractable for datasets with a larger vocabulary of objects and predicates (e.g. VG, GQA). Prior works [61] therefore evaluate only on triplets which are known to be present in the test split, which may inflate metrics. In contrast, our method disentangles objects and predicates (Section 3). This allows for an exhaustive evaluation that considers all object and predicate combinations and enables applications where knowledge of possible relationships is unavailable. + +# 4.2 Relationship Detection Performance + +We use Visual Genome as our main benchmark for relationship detection. Table 2 shows results for our models and prior work. Prior studies often achieve reported performance through the combination of a base model with a specialized approach to overcome the skewed data distribution (indicated as base-model + specialization in Table 2). Our models provide strong performance even without special treatment of the skewed predicate distribution. After brief fine-tuning on data in which the frequency of predicate annotations was rebalanced with simple rejection sampling (see Appendix A.1), our method improves on the prior best method DT2-ACBS [8] by 1.7 points mR@100 (26.1 vs. 24.4). This shows that our already-strong method may be further improved by combining it with more advanced rebalancing methods. + +We also note the large difference in performance to UniVRD [61], which builds on the same detection architecture as our method, but adds a Transformer decoder-based relationship module. To disentangle whether the improve + +
ModelmR@50mR@100
LLM4SGG [21]5.36.5
Neural Motif w/ GCL [9]16.818.8
VCTree w/ GCL [9]15.617.8
SG-Transformer w/ CFA [28]13.415.3
SHA w/ GCL [9]17.820.1
SG-ViT (CLIP: ViT-B/32)15.718.6
SG-ViT (CLIP: ViT-B/16)17.720.5
SG-ViT (CLIP: ViT-L/14)19.322.9
+ +Table 3: Performance on the GQA200 test set. Graph-constrained evaluation. + +
ModelmR@50mR@100
Scene Graph Classification
IMP [25,44,48]0.50.7
Neural Motif [25,44,55]0.81.2
Unbiased TDE [25,45]-0.7
RTN [25]-1.4
MP [23]-2.8
Transformer w/ EBM [44]1.31.8
Scene Graph Generation
SG-ViT (CLIP: ViT-B/32)6.27.4
SG-ViT (CLIP: ViT-B/16)6.47.4
SG-ViT (CLIP: ViT-L/14)8.09.6
+ +Table 4: Performance on the GQA test set. Graph-constrained evaluation. + +ments over UniVRD are due to differences in architecture or training data, we also trained our B/32 model on the UniVRD data mixture, and still observe a large improvement (17.4 VG mR@100 for our method vs. 12.1 for UniVRD when trained on the same data; Table 6). This suggests that our encoder-only architecture is better suited for VRD than prior decoder-based architectures [7,20,29,61]. + +# 4.3 Scaling to Large Vocabularies + +Improving long-tail performance has been critical to the VRD field since the natural distribution of relationship triplets is highly skewed [27]. A substantial body of literature is devoted to this goal, often proposing complex loss debiasing or data augmentation approaches [8,27,52,57]. It is therefore of interest whether our method works well in this regime without special treatment of rare classes. To evaluate the performance of our method on large object and predicate vocabularies with a long tail of rare classes, we use the Graph Question Answering (GQA) dataset [16]. This dataset builds on VG and expands the number of classes and annotations. A simplified version of GQA with 200 object and 100 predicate classes, called GQA200 [9], is part of our training mixture. Our method surpasses prior results on the GQA200 test split (Table 3), despite lacking the specialized treatment of the data distribution that some prior methods use. + +The full GQA dataset has an even larger vocabulary with 1703 object and 311 predicate classes. To our knowledge, no prior methods evaluate scene graph generation on the full GQA dataset. Some prior methods report results on scene graph classification, where ground-truth boxes are provided. As shown in Table 4, our model achieves higher performance on the much harder task of scene graph generation than prior methods on classification. We further assess zero-shot generalization to unseen classes in Appendix A.2. + +We believe that two factors contribute to the strong performance in the open vocabulary regime: First, the simple design of the method does not impede transfer of semantic knowledge from the pre-trained VLM backbone because it adds only lightweight heads. In particular, we use no decoder, which could be difficult to train and could lead to "forgetting" of pretrained representations. Instead, most of the relationship modeling happens in the backbone, which has + +had the benefit of large-scale vision-language pretraining. Second, the open-vocabulary design allows end-to-end training on a mix of datasets, which was not possible for all prior methods. + +# 4.4 Human-Object Interaction + +We also evaluate our model on the specialized task of human-object interaction, using the HICO benchmark [4]. The performance of our model is on par with the most comparable prior method [61], but does not show similarly large improvements as for VG or GQA200 (Table 5). HICO has a much narrower and more specialized vocabulary than VG or GQA (HICO evaluates on 600 prespecified triplets, whereas VG150 has $1.1 \times 10^{6}$ possible triplet combinations). Performance on this specialized task may benefit less from transfer of pretrained representations and may instead be limited by the amount of task-specific training data. This is supported by the fact that the recent state-of-the-art method for HICO pretrains on large amounts of person-focused pseudo-labels [54]. + +# 4.5 Ablations + +Inference Speed and Number of Predicted Relationships. In the following, we ablate training datasets and model components to study the interplay of data and architecture on the model performance. + +The primary hyperparameter of our method is the number of relationships $k$ that are selected by the Relationship Attention layer from all possible combinations for further processing. Figure 3 shows how VRD performance and inference speed depend on $k$ . + +For training, we use $k = 2^{14}$ , which results in $71\%$ (B/32) to $81\%$ (L/14) of the speed of a pure object detection model [34] of the same size. Without top-k selection, the model would be several times slower (right end of top plot in Figure 3). Since the vast majority of pairs do not form relationships, VRD performance is essentially unaffected by top- $k$ selection at inference. For inference, we can therefore use $k = 2^{11}$ at no loss in accuracy to achieve $90\%$ of the speed of the object detector (for B/32 and + +L/14). At this $k$ , the B/32 model achieves 52.8 FPS at batch size 1 on an NVIDIA V100 GPU and is therefore suitable for real-time applications. + +![](images/6ce65786911baddd0584536ae1881f94620020316915a4e5beca96624dd3b69e.jpg) + +![](images/ee12aaefd35ea4036b65e7cfd654c798dde90954e058b06c94ce55efc763b626.jpg) +Figure 3: Model speed and VRD accuracy by number of selected relationships $k$ . Speed is relative to a non-VRD object detector [34]. + +Dataset Mixture. To assess how training data and architectural improvements contribute to the performance of our model compared to prior work, we train the models on a range of dataset mixtures (Table 6). The full mixture, which we use for all experiments unless otherwise noted, includes VG, VG150, GQA200, + +
ModelBackbonemAPmAPrmAPn
HOTR [20]ResNet5025.1020.3325.86
RLIP [53]ResNet5032.8426.8534.63
RLIPv2 [54]ResNet5035.3829.6137.10
RLIPv2 [54]Swin-L45.0943.2345.64
UniVRD [60]CLIP: ViT-B/3229.9822.9432.02
UniVRD [60]CLIP: ViT-B/1629.9822.9432.02
UniVRD [60]CLIP: ViT-L/1437.4128.9039.95
SG-ViTCLIP: ViT-B/3228.8623.7230.39
SG-ViTCLIP: ViT-B/1631.9826.8333.52
SG-ViTCLIP: ViT-L/1438.1133.7139.42
+ +Table 5: Performance on the HICO test set. Our method is on par with the most comparable prior work, UniVRD, for overall mAP. However, our method performs better on "rare" classes (mAPr), indicating better generalization to rarely seen concepts. For reference, we also show RLIPv2, the state-of-the-art method on HICO. RLIPv2 proposes a self-training approach that is orthogonal and compatible to our method. + +
VGMixtureVG150 mR@100GQA200 mR@100HICO mAP
VG150GQA200HICOO365COCO
Our mixture12.5%12.5%12.5%12.5%50%-18.118.628.9
UniVRD [61] mix-10%-20%50%20%17.4 ↓ 4%5.1 ↓ 73%17.4 ↓ 8%
ours w/o O36525%25%25%25%--15.3 ↓ 15%15.6 ↓ 16%23.8 ↓ 18%
ours w/o VG-25%12.5%12.5%50%-14.2 ↓ 22%15.3 ↓ 18%26.8 ↓ 7%
+ +Table 6: Dataset Ablation. Performance of SG-ViT (B/32) trained on different data mixtures. VG150 and GQA200 evaluation is graph-constrained, HICO is unconstrained. + +HICO and O365. To allow direct comparison to UniVRD [61], we also train on their mixture (VG150, HICO, O365, COCO) and find that our model still performs better on VG150 by a large margin (18.1 mR@100 vs 12.1 for UniVRD for B/32 models). Since UniVRD uses the same image and text encoding backbones as our model, this result suggests that our encoder-only model with Relationship Attention is an advance over decoder-based relationship architectures. We further find that including both pure object detection (O365) and large-vocabulary VRD data (VG) benefit all metrics. + +Object Detection Performance. A necessary step towards good relationship detection performance is accurate object detection. In our architecture, object detection is treated as a special case of relationship detection in which the is identical to the . We validate this design by training the model only on detection datasets without relationship annotations, using the same dataset mixture as the OWL-ViT [34] model (Objects365 and Visual Genome, [34]). We find that our model achieves a similar object detection performance as OWL-ViT (21.9% vs. 22.1% mAP and 18.3% vs. 18.9% mAPr on LVIS [12]), indicating that the Relationship Attention design does not interfere with object detection. + +![](images/f6264e2b1fc727434aab771caeec0d0fb4ceb3c09d21ec7a9cf8800abb6cbdaa.jpg) +Figure 4: Qualitative examples of difficult edge-cases from VG150 test split [26, 55]. From left to right: Ground Truth, SG-ViT (B/32), SG-ViT (L/14). In all cases the is lime and the is red. + +# 4.6 Qualitative Results + +In Figure 4, we visualize qualitative examples to highlight difficult edge cases. Even where the model output differs from the ground truth, it is often plausibly correct, e.g. where either the extent of an object differs with the ground truth or there are multiple appropriate, but unannotated, relations. For example, both models select the headlight on the bus in Figure 4b for "light on bus" as opposed to the sun reflected off the windshield. In Figure 4c both models focus more on the "on" in "snow on hill" than the distinction between a mountain and a hill. Figure 4d shows a challenging scene where the B/32 model selects the bottle inside the open fridge while the L/14 correctly selects a water bottle in the door for "bottle in door" while also capturing the extent of the door instance. + +Furthermore, to qualitatively assess the utility of this work beyond standard relation benchmark datasets, we also visualize some example relations applied to data from the robotics domain, namely the OXE dataset [35]. Figure 5 shows a real-life scene where multiple instances of the same semantic object categories are present within the scene. The model is able to match the to the instances in the desired configuration. One subjective failure-case is that the "banana outside bowl" text query had its highest scoring triplet involving a bowl on a shelf in the background rather than the adjacent bowl. Another charac + +![](images/96211429d2725ae9e1b59b8a13180e29bd18aef541e76869374e0860caf512bb.jpg) + +![](images/254994729b19ecd82ba91ce8146a55b6a254d15717861e67e505b8f10ec2d570.jpg) + +![](images/aeb341dbc46b61dda18fddfb4d8b9a969af189f75879cfa642faca97a3dd5b71.jpg) + +![](images/78b4213f16e188704cfe8e1fae668aed3ec2aee932542858578e5c5d5925f81b.jpg) + +![](images/1e4342115292a3f92fbbfd5e311b2de5f1967605b4f32a18b54289b3c989ea20.jpg) +Figure 5: Qualitative examples showing SG-ViT (L/14) on out-of-distribution data from the OXE dataset [35]. In all cases the is lime and the is red. Note how the model correctly disambiguates several instances of the same class (e.g. "bananas" and "bottle") depending on their relationships. + +![](images/a83cbc2f571ab5f2471b9a5a44ab9afb5748de82eb8c623823c3f7337ee5a21e.jpg) + +![](images/906289885ef00cead1092aadcb6e8ec257528af9fdc0677df59a21814e53f5dc.jpg) + +![](images/de418ba07b8a61d2a01bf4bd1a4ab2e4ad2ca3e10f1458c1001e6e19619aaf4d.jpg) + +teristic is that the model occasionally places boxes over groups with multiple instances of the same semantic class when a singular word is used for either the or . This can be attributed to the similarity between word and their plural in the CLIP embeddings and that many human annotations from VG also confuse singular and plural (see Appendix A.3). + +# 5 Limitations + +While our model performs strongly on large-vocabulary VRD, its performance on specialized human-object interaction detection is only on par with prior models. We discuss this limitation in Section 4.4. Further error modes are explored qualitatively in Section 4.6 and Appendix A.3. A challenge for open-vocabulary relationship detection models as a whole is zero-shot generalization to unseen objects and predicates. While our model improves over prior approaches (Tables 3 and 4), there is still a large gap when it comes to entirely unseen classes (Appendix A.2). Future research should focus on closing this gap. + +# 6 Conclusion + +We present an architecture for open-vocabulary visual relationship detection. By combining an encoder-only design with a novel Relationship Attention layer for efficient selection of high-confidence relationships, our architecture leverages VLM pretraining and multi-dataset VRD training. It achieves strong performance on standard and large-vocabulary VRD benchmarks while maintaining pure object detection performance and adding little extra computational cost. Due to its simplicity and strong performance, we believe that our method will be a useful basis for further research on visual relationship detection. + +# References + +1. Amiri, S., Chandan, K., Zhang, S.: Reasoning with scene graphs for robot planning under partial observability. IEEE Robotics and Automation Letters 7(2), 5560-5567 (2022) +2. Ba, J.L., Kiros, J.R., Hinton, G.E.: Layer normalization. arXiv preprint arXiv:1607.06450 (2016) +3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: Conference on Computer Vision (ECCV). pp. 213-229. Springer (2020) +4. Chao, Y.W., Liu, Y., Liu, X., Zeng, H., Deng, J.: Learning to Detect Human-Object Interactions. In: IEEE Winter Conference on Applications of Computer Vision (2018) +5. Chen, B., Xia, F., Ichter, B., Rao, K., Gopalakrishnan, K., Ryoo, M.S., Stone, A., Kappler, D.: Open-vocabulary queryable scene representations for real world planning. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 11509-11522. IEEE (2023) +6. Chen, T., Yu, W., Chen, R., Lin, L.: Knowledge-embedded routing network for scene graph generation. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 6163-6171 (2019) +7. Cong, Y., Yang, M.Y., Rosenhahn, B.: Reltr: Relation transformer for scene graph generation. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023) +8. Desai, A., Wu, T.Y., Tripathi, S., Vasconcelos, N.: Learning of visual relations: The devil is in the tails. In: IEEE/CVF International Conference on Computer Vision (ICCV). pp. 15404-15413 (2021) +9. Dong, X., Gan, T., Song, X., Wu, J., Cheng, Y., Nie, L.: Stacked hybrid-attention and group collaborative learning for unbiased scene graph generation. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 19427-19436 (2022) +10. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. ICLR (2021) +11. Gu, Q., Kuwajerwala, A., Morin, S., Jatavallabhula, K.M., Sen, B., Agarwal, A., Rivera, C., Paul, W., Ellis, K., Chellappa, R., et al.: Conceptgraphs: Open-vocabulary 3d scene graphs for perception and planning. arXiv preprint arXiv:2309.16650 (2023) +12. Gupta, A., Dollar, P., Girshick, R.: LVIS: A Dataset for Large Vocabulary Instance Segmentation. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2019) +13. He, T., Gao, L., Song, J., Li, Y.F.: Towards Open-Vocabulary Scene Graph Generation with Prompt-Based Finetuning. In: European Conference on Computer Vision (ECCV) (2022) +14. Hendrycks, D., Gimpel, K.: Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415 (2016) +15. Herzig, R., Mendelson, A., Karlinsky, L., Arbelle, A., Feris, R., Darrell, T., Globerson, A.: Incorporating Structured Representations into Pretrained Vision & Language Models Using Scene Graphs. In: The 2023 Conference on Empirical Methods in Natural Language Processing (2023) + +16. Hudson, D.A., Manning, C.D.: Gqa: A new dataset for real-world visual reasoning and compositional question answering. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 6700-6709 (2019) +17. Hughes, N., Chang, Y., Carlone, L.: Hydra: A Real-time Spatial Perception System for 3D Scene Graph Construction and Optimization. Robotics: Science and Systems (RSS) (2022) +18. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International conference on machine learning. pp. 4904-4916. PMLR (2021) +19. Johnson, J., Krishna, R., Stark, M., Li, L.J., Shamma, D., Bernstein, M., Fei-Fei, L.: Image Retrieval Using Scene Graphs. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2015) +20. Kim, B., Lee, J., Kang, J., Kim, E.S., Kim, H.J.: Hotr: End-to-end human-object interaction detection with transformers. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 74-83 (2021) +21. Kim, K., Yoon, K., Jeon, J., In, Y., Moon, J., Kim, D., Park, C.: LLM4SGG: Large Language Model for Weakly Supervised Scene Graph Generation. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024) +22. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) +23. Knyazev, B., de Vries, H., Cangea, C., Taylor, G.W., Courville, A., Belilovsky, E.: Graph Density-Aware Losses for Novel Compositions in Scene Graph Generation. In: British Machine Vision Conference (BMVC) (2020) +24. Kolesnikov, A., Beyer, L., Zhai, X., Puigcerver, J., Yung, J., Gelly, S., Houlsby, N.: Big transfer (bit): General visual representation learning. In: European Conference on Computer Vision (ECCV). pp. 491-507 (2020) +25. Koner, R., Shit, S., Tresp, V.: Relation transformer network. arXiv preprint arXiv:2004.06193 (2020) +26. Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., et al.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. International Journal of Computer Vision 123, 32-73 (2017) +27. Li, H., Zhu, G., Zhang, L., Jiang, Y., Dang, Y., Hou, H., Shen, P., Zhao, X., Shah, S.A.A., Bennamoun, M.: Scene Graph Generation: A comprehensive survey. Neurocomputing (2024). https://doi.org/https://doi.org/10.1016/j.neucom.2023.127052 +28. Li, L., Chen, G., Xiao, J., Yang, Y., Wang, C., Chen, L.: Compositional feature augmentation for unbiased scene graph generation. In: IEEE/CVF International Conference on Computer Vision (ICCV). pp. 21685-21695 (2023) +29. Li, R., Zhang, S., He, X.: Sgtr: End-to-end scene graph generation with transformer. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 19486-19496 (2022) +30. Li, Y., Ouyang, W., Zhou, B., Wang, K., Wang, X.: Scene graph generation from objects, phrases and region captions. In: IEEE/CVF International Conference on Computer Vision (ICCV). pp. 1261-1270 (2017) +31. Lin, X., Ding, C., Zeng, J., Tao, D.: Gps-net: Graph property sensing network for scene graph generation. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3746-3753 (2020) + +32. Lu, C., Krishna, R., Bernstein, M., Fei-Fei, L.: Visual relationship detection with language priors. In: European Conference on Computer Vision (ECCV). pp. 852-869. Springer (2016) +33. Minderer, M., Gritsenko, A., Houlsby, N.: Scaling open-vocabulary object detection. Advances in Neural Information Processing Systems 36 (2024) +34. Minderer, M., Gritsenko, A., Stone, A., Neumann, M., Weissenborn, D., Dosovitskiy, A., Mahendran, A., Arnab, A., Dehghani, M., Shen, Z., et al.: Simple open-vocabulary object detection. In: European Conference on Computer Vision (ECCV). pp. 728-755. Springer (2022) +35. Open X-Embodiment Collaboration: Open X-Embodiment: Robotic Learning Datasets and RT-X Models (2023) +36. Peng, Z., Wang, W., Dong, L., Hao, Y., Huang, S., Ma, S., Ye, Q., Wei, F.: Grounding Multimodal Large Language Models to the World. In: International Conference on Learning Representations (2024) +37. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) +38. Rana, K., Haviland, J., Garg, S., Abou-Chakra, J., Reid, I., Suenderhauf, N.: Say-plan: Grounding large language models using 3d scene graphs for scalable robot task planning. In: 7th Annual Conference on Robot Learning (2023) +39. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems 28 (2015) +40. Shao, S., Li, Z., Zhang, T., Peng, C., Yu, G., Zhang, X., Li, J., Sun, J.: Objects365: A large-scale, high-quality dataset for object detection. In: IEEE/CVF International Conference on Computer Vision (ICCV). pp. 8430-8439 (2019) +41. Shi, H., Hayat, M., Cai, J.: Open-vocabulary object detection via scene graph discovery. In: 31st ACM International Conference on Multimedia. pp. 4012-4021 (2023) +42. Singh, K.P., Salvador, J., Weihs, L., Kembhavi, A.: Scene graph contrastive learning for embodied navigation. In: IEEE/CVF International Conference on Computer Vision (ICCV). pp. 10884-10894 (2023) +43. Song, H., Sun, D., Chun, S., Jampani, V., Han, D., Heo, B., Kim, W., Yang, M.H.: ViDT: An Efficient and Effective Fully Transformer-based Object Detector. In: International Conference on Learning Representations (2022) +44. Suhail, M., Mittal, A., Siddiquie, B., Broaddus, C., Eledath, J., Medioni, G., Sigal, L.: Energy-based learning for scene graph generation. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13936-13945 (2021) +45. Tang, K., Niu, Y., Huang, J., Shi, J., Zhang, H.: Unbiased scene graph generation from biased training. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3716-3725 (2020) +46. Tang, K., Zhang, H., Wu, B., Luo, W., Liu, W.: Learning to compose dynamic tree structures for visual contexts. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 6619-6628 (2019) +47. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017) +48. Xu, D., Zhu, Y., Choy, C.B., Fei-Fei, L.: Scene graph generation by iterative message passing. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5410-5419 (2017) + +49. Yang, J., Lu, J., Lee, S., Batra, D., Parikh, D.: Graph r-cnn for scene graph generation. In: European Conference on Computer Vision (ECCV). pp. 670-685 (2018) +50. Yao, Z., Ai, J., Li, B., Zhang, C.: Efficient DETR: Improving End-to-End Object Detector With Dense Prior. arXiv preprint arXiv:2104.01318 (2021) +51. Yu, J., Wang, Z., Vasudevan, V., Yeung, L., Seyedhosseini, M., Wu, Y.: CoCa: Contrastive Captioners are Image-Text Foundation Models. Transactions on Machine Learning Research (2022) +52. Yu, J., Chai, Y., Wang, Y., Hu, Y., Wu, Q.: CogTree: Cognition Tree Loss for Unbiased Scene Graph Generation. In: International Joint Conference on Artificial Intelligence, IJCAI-21. International Joint Conferences on Artificial Intelligence Organization (8 2021) +53. Yuan, H., Jiang, J., Albaniae, S., Feng, T., Huang, Z., Ni, D., Tang, M.: Rlip: Relational language-image pre-training for human-object interaction detection. Advances in Neural Information Processing Systems 35, 37416-37431 (2022) +54. Yuan, H., Zhang, S., Wang, X., Albanie, S., Pan, Y., Feng, T., Jiang, J., Ni, D., Zhang, Y., Zhao, D.: RLIPv2: Fast Scaling of Relational Language-Image Pretraining. In: IEEE/CVF International Conference on Computer Vision (ICCV). pp. 21649-21661 (2023) +55. Zellers, R., Yatskar, M., Thomson, S., Choi, Y.: Neural motifs: Scene graph parsing with global context. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5831-5840 (2018) +56. Zhai, X., Wang, X., Mustafa, B., Steiner, A., Keysers, D., Kolesnikov, A., Beyer, L.: Lit: Zero-shot transfer with locked-image text tuning. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 18123-18133 (2022) +57. Zhang, A., Yao, Y., Chen, Q., Ji, W., Liu, Z., Sun, M., Chua, T.S.: Fine-grained scene graph generation with data transfer. In: European Conference on Computer Vision (ECCV). pp. 409-424. Springer (2022) +58. Zhang, H., Li, F., Liu, S., Zhang, L., Su, H., Zhu, J., Ni, L.M., Shum, H.Y.: Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605 (2022) +59. Zhang, Y., Ma, Z., Gao, X., Shakiah, S., Gao, Q., Chai, J.: GROUNDHOG: Grounding Large Language Models to Holistic Segmentation. In: Conference on Computer Vision and Pattern Recognition 2024 (2024), https://openreview.net/forum?id=PWW3ux1Ju +60. Zhao, L., Yuan, L., Gong, B., Cui, Y., Schroff, F., Yang, M.H., Adam, H., Liu, T.: Unified Visual Relationship Detection with Vision and Language Models. In: IEEE/CVF International Conference on Computer Vision (ICCV) (2023) +61. Zhao, L., Yuan, L., Gong, B., Cui, Y., Schroff, F., Yang, M.H., Adam, H., Liu, T.: Unified Visual Relationship Detection with Vision and Language Models. In: IEEE/CVF International Conference on Computer Vision (ICCV) (2023) +62. Zhou, X., Girdhar, R., Joulin, A., Krahenbuhl, P., Misra, I.: Detecting twenty-thousand classes using image-level supervision. In: European Conference on Computer Vision (ECCV). pp. 350-368 (2022) +63. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable DETR: Deformable Transformers for End-to-End Object Detection. In: International Conference on Learning Representations (2021) \ No newline at end of file diff --git a/2024/Scene-Graph ViT_ End-to-End Open-Vocabulary Visual Relationship Detection/images.zip b/2024/Scene-Graph ViT_ End-to-End Open-Vocabulary Visual Relationship Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..cf49a41f3390f3b975664c75c3ad10b3c0219682 --- /dev/null +++ b/2024/Scene-Graph ViT_ End-to-End Open-Vocabulary Visual Relationship Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:66bb5c0e204e5698a7d963ebb11dc5cd3016976b7fe92ee6e31b013883caf154 +size 508950 diff --git a/2024/Scene-Graph ViT_ End-to-End Open-Vocabulary Visual Relationship Detection/layout.json b/2024/Scene-Graph ViT_ End-to-End Open-Vocabulary Visual Relationship Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..23fb8b95f125da79d26f1d4c9e058dd9e2f0783a --- /dev/null +++ b/2024/Scene-Graph ViT_ End-to-End Open-Vocabulary Visual Relationship Detection/layout.json @@ -0,0 +1,8724 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 137, + 111, + 477, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 111, + 477, + 148 + ], + "spans": [ + { + "bbox": [ + 137, + 111, + 477, + 148 + ], + "type": "text", + "content": "Scene-Graph ViT: End-to-End Open-Vocabulary Visual Relationship Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 208, + 167, + 406, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 167, + 406, + 193 + ], + "spans": [ + { + "bbox": [ + 208, + 167, + 406, + 193 + ], + "type": "text", + "content": "Tim Salzmann" + }, + { + "bbox": [ + 208, + 167, + 406, + 193 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 208, + 167, + 406, + 193 + ], + "type": "text", + "content": ", Markus Ryll" + }, + { + "bbox": [ + 208, + 167, + 406, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 208, + 167, + 406, + 193 + ], + "type": "text", + "content": ", Alex Bewley" + }, + { + "bbox": [ + 208, + 167, + 406, + 193 + ], + "type": "inline_equation", + "content": "^{\\star 1}" + }, + { + "bbox": [ + 208, + 167, + 406, + 193 + ], + "type": "text", + "content": ", and Matthias Minderer" + }, + { + "bbox": [ + 208, + 167, + 406, + 193 + ], + "type": "inline_equation", + "content": "^{\\star 1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 226, + 201, + 383, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 226, + 201, + 383, + 246 + ], + "spans": [ + { + "bbox": [ + 226, + 201, + 383, + 246 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 226, + 201, + 383, + 246 + ], + "type": "text", + "content": " Google DeepMind {tsal, bewley, mjlm}@google.com " + }, + { + "bbox": [ + 226, + 201, + 383, + 246 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 226, + 201, + 383, + 246 + ], + "type": "text", + "content": " Technical University Munich {tim.salzmann, markus.ryll}@tum.d" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 263, + 342, + 402 + ], + "blocks": [ + { + "bbox": [ + 133, + 263, + 342, + 402 + ], + "lines": [ + { + "bbox": [ + 133, + 263, + 342, + 402 + ], + "spans": [ + { + "bbox": [ + 133, + 263, + 342, + 402 + ], + "type": "image", + "image_path": "76dc13707de447c668c91d07b4346bd00fc74461a2918d44f1240e6c8f250fb0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 404, + 482, + 427 + ], + "lines": [ + { + "bbox": [ + 132, + 404, + 482, + 427 + ], + "spans": [ + { + "bbox": [ + 132, + 404, + 482, + 427 + ], + "type": "text", + "content": "Figure 1: Relationships detected by our method on an unseen image. The top relationships by confidence score are shown. Photo by Spacejoy on Unsplash." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 343, + 262, + 482, + 402 + ], + "blocks": [ + { + "bbox": [ + 343, + 262, + 482, + 402 + ], + "lines": [ + { + "bbox": [ + 343, + 262, + 482, + 402 + ], + "spans": [ + { + "bbox": [ + 343, + 262, + 482, + 402 + ], + "type": "image", + "image_path": "f8a13735bc1e2ecd2dd20a03f79190938d5c6a1b92da04791fe196e43dbb5792.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 448, + 455, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 448, + 455, + 613 + ], + "spans": [ + { + "bbox": [ + 159, + 448, + 455, + 613 + ], + "type": "text", + "content": "Abstract. Visual relationship detection aims to identify objects and their relationships in images. Prior methods approach this task by adding separate relationship modules or decoders to existing object detection architectures. This separation increases complexity and hinders end-to-end training, which limits performance. We propose a simple and highly efficient decoder-free architecture for open-vocabulary visual relationship detection. Our model consists of a Transformer-based image encoder that represents objects as tokens and models their relationships implicitly. To extract relationship information, we introduce an attention mechanism that selects object pairs likely to form a relationship. We provide a single-stage recipe to train this model on a mixture of object and relationship detection data. Our approach achieves state-of-the-art relationship detection performance on Visual Genome and on the large-vocabulary GQA benchmark at real-time inference speeds. We provide ablations, real-world qualitative examples, and analyses of zero-shot performance." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 142, + 642, + 329, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 642, + 329, + 666 + ], + "spans": [ + { + "bbox": [ + 142, + 642, + 329, + 666 + ], + "type": "inline_equation", + "content": "^{\\dagger}" + }, + { + "bbox": [ + 142, + 642, + 329, + 666 + ], + "type": "text", + "content": "Work done while at Google DeepMind. \n*Advising project leads in alphabetical order." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 230, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 230, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 230, + 127 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 142, + 482, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 142, + 482, + 238 + ], + "spans": [ + { + "bbox": [ + 130, + 142, + 482, + 238 + ], + "type": "text", + "content": "A fundamental goal of computer vision is to decompose visual scenes into structured semantic representations. A commonly studied task towards this goal is object detection, in which objects in an image are localized by bounding boxes and classified into semantic categories. However, a scene description also includes the semantic relationships between objects. This gives rise to the task of visual relationship detection (VRD, [4,6,32,48,49,55]). In VRD, the model detects objects and infers pairwise relationships between them in the form of triplets3 [32]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 238, + 482, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 238, + 482, + 310 + ], + "spans": [ + { + "bbox": [ + 130, + 238, + 482, + 310 + ], + "type": "text", + "content": "Detecting both objects and their relationships allows the construction of a scene graph [19,30,48] in which objects are represented as nodes and their relationships as edges. Scene graph generation (SGG) has wide-ranging applications in robotics [1,5,11,17,38,42] and image retrieval [19,27]. Increasingly, structured scene representations are also employed to provide grounding and explainability to multimodal large language models [15,36,59]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 310, + 482, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 310, + 482, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 310, + 482, + 430 + ], + "type": "text", + "content": "Prior work typically draws a distinction between object detection and relationship prediction. Detection is performed either as a wholly separate step before relationship prediction [8, 32, 48, 52, 55, 57], or by separate model parts such as \"relationship decoders\" that are responsible for modeling the interactions between objects [7, 49, 61]. This distinction makes it hard to optimize such models end-to-end for VRD [61]. In contrast, we propose an encoder-only architecture that models objects and relationships jointly, directly in the image encoder. Our architecture performs open-vocabulary relationship detection and can be trained end-to-end on arbitrary mixtures of object detection and relationship annotations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 430, + 482, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 482, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 482, + 525 + ], + "type": "text", + "content": "Our model builds on a Transformer-based encoder-only object detector [34] in which the output tokens of the image encoder directly represent object proposals. From these tokens, class embeddings and bounding boxes are decoded with lightweight heads. Our insight is that this architecture is perfectly set up to learn relationships between objects directly in the image encoder, without the need for additional relationship-specific stages. This is because the existing self-attention in the encoder already models all-to-all pairwise interactions between the object proposal tokens." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 526, + 482, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 526, + 482, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 526, + 482, + 635 + ], + "type": "text", + "content": "To access information about the relationship between two of these tokens, we combine the embeddings corresponding to the and token using a new Relationship Attention layer. Obtaining relationship embeddings for all possible pairwise combinations of object proposal tokens would be computationally infeasible. To reduce the number of combinations, we introduce a self-supervised hard attention mechanism that selects the highest-confidence pairs at a computational cost comparable to a single self-attention layer. We show how to directly supervise the attention scores of this mechanism without the need to propagate gradients through the hard selection." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "text", + "content": "T. Salzmann et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 642, + 482, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 482, + 665 + ], + "type": "text", + "content": "3 We use to distinguish the grammatical terms and from generic use of the word object as in object detection." + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": "An additional benefit of our design is that it disentangles object names from relationship predicates during inference. In contrast to prior open-vocabulary methods [53, 54, 61], our model can embed object and predicate texts separately and generate confidence scores efficiently for all possible combinations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 146, + 176, + 367, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 176, + 367, + 186 + ], + "spans": [ + { + "bbox": [ + 146, + 176, + 367, + 186 + ], + "type": "text", + "content": "In summary, we make the following contributions:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 135, + 193, + 473, + 239 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 135, + 193, + 473, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 193, + 473, + 204 + ], + "spans": [ + { + "bbox": [ + 135, + 193, + 473, + 204 + ], + "type": "text", + "content": "1. An efficient architecture for open-vocabulary visual relationship detection." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 135, + 204, + 466, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 204, + 466, + 216 + ], + "spans": [ + { + "bbox": [ + 135, + 204, + 466, + 216 + ], + "type": "text", + "content": "2. A single-stage recipe for joint object and relationship detection training." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 135, + 216, + 395, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 216, + 395, + 227 + ], + "spans": [ + { + "bbox": [ + 135, + 216, + 395, + 227 + ], + "type": "text", + "content": "3. Efficient, disentangled object and relationship inference." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 135, + 227, + 430, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 227, + 430, + 239 + ], + "spans": [ + { + "bbox": [ + 135, + 227, + 430, + 239 + ], + "type": "text", + "content": "4. Analysis of inference speed, ablations, and qualitative examples." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 244, + 479, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 244, + 479, + 303 + ], + "spans": [ + { + "bbox": [ + 130, + 244, + 479, + 303 + ], + "type": "text", + "content": "Our method achieves state-of-the-art visual relationship detection performance on the Visual Genome dataset (26.1% mR@100 with graph constraint) and on the large-vocabulary GQA benchmark. It provides strong results while being significantly simpler than prior approaches to visual relationship detection." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 320, + 236, + 333 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 320, + 236, + 333 + ], + "spans": [ + { + "bbox": [ + 132, + 320, + 236, + 333 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 343, + 479, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 343, + 479, + 367 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 479, + 367 + ], + "type": "text", + "content": "Below, we review the main approaches to VRD that are relevant to our work. For a comprehensive overview of the field, we refer to [27]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 371, + 480, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 371, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 371, + 480, + 479 + ], + "type": "text", + "content": "Detector-Agnostic Relationship Detection. Historically, many works on VRD assume that object detections are given, such that the task reduces to inferring relationships between the objects. These \"detector-agnostic\" methods use off-the-shelf detectors such as Faster-RCNN [39] to obtain boxes and box embeddings and infer a scene graph from them. Early work on large-scale VRD employed word embeddings to improve relationship generalization [32] and graph inference methods such as message passing for SGG [48]. More recent methods leverage relationship co-occurrence statistics [55] and address the long-tailed nature of the training data [8, 52, 57]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 483, + 481, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 483, + 481, + 675 + ], + "spans": [ + { + "bbox": [ + 130, + 483, + 481, + 675 + ], + "type": "text", + "content": "End-to-End Relationship Detection. A fundamental limitation of the detector-agnostic approach is that object detection and relationship prediction are treated separately and cannot learn from each other. Overcoming this limitation has recently motivated the development of end-to-end architectures in which objects and relationships are predicted jointly by a single model [7, 20, 29, 61]. All of these models have in common that they use a Transformer [47] decoder to predict object and relationship embeddings. The decoder represents object proposals with query embeddings that can cross-attend into image embeddings. The literature on object detection suggests that this choice may be problematic, because the query embeddings can be difficult to initialize, which can lead to unstable and slow optimization [3, 34, 43, 50, 58, 63]. To avoid this issue, we design an encoder-only architecture in which both object and relationship embeddings are computed directly from the output tokens of a Vision Transformer [10] image encoder, building on an idea from object detection [34]. In contrast to prior models [61], this allows us to train the model end-to-end on object and relationship annotations simultaneously." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 375, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 375, + 91, + 447, + 102 + ], + "type": "text", + "content": "Scene-Graph ViT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 308 + ], + "type": "text", + "content": "Long-Tailed and Open-Vocabulary Relationship Detection. Object classes in the natural world follow a long-tailed distribution [12,62], and this is compounded in VRD due to the combination of subject, object, and predicate in a relationship [27,48,49]. A large number of VRD methods specifically aim to address this issue, for example with debiasing losses [52], data augmentation [57], or data resampling [8]. Additionally, while VRD models historically assumed a fixed set of object and predicate classes, recent works add open-vocabulary capabilities [13,41,54,61]. These models leverage pretrained vision-language models such as CLIP [18, 37, 51, 56] to inherit their natural-language understanding. Instead of learning a fixed set of classes, open-vocabulary models predict class embeddings that can be compared to the text embeddings of arbitrary object or predicate descriptions for classification. Here, we employ a pretrained vision-language backbone and combine it with a carefully designed, lightweight relationship detection head to preserve and transfer semantic knowledge from the backbone pretraining to VRD. This allows our model to achieve strong results on the large-vocabulary GQA benchmark [16] (Section 4.3)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 308, + 482, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 482, + 368 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 482, + 368 + ], + "type": "text", + "content": "While we focus on architectural improvements, another line of work addresses the scarcity of training data for VRD. For example, RLIP [53] and RLIPv2 [54], which focus on human-object interaction [4], propose relationship-focused image-text pretraining and self-training, which yield large improvements in VRD performance. These works are orthogonal to our contributions and can be combined." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 385, + 259, + 400 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 385, + 259, + 400 + ], + "spans": [ + { + "bbox": [ + 132, + 385, + 259, + 400 + ], + "type": "text", + "content": "3 Scene-Graph ViT" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 411, + 482, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 482, + 495 + ], + "type": "text", + "content": "We propose an architecture for open-vocabulary relationship prediction in which both objects and the relationships between them are handled as \"first-class citizens\" directly in a single-stage process within the model backbone. We build on an encoder-only architecture for object detection [34] and adapt it to relationship prediction by adding a specialized attention layer. This layer exploits the pairwise structure between existing object embeddings to obtain relationship embeddings as schematized in Figure 2 and described below." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 513, + 411, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 513, + 411, + 526 + ], + "spans": [ + { + "bbox": [ + 132, + 513, + 411, + 526 + ], + "type": "text", + "content": "3.1 Encoder-Only Open-Vocabulary Object Detection" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "We briefly review the detection architecture that forms the basis of our model [34]. The model consists of Transformer-based [10] image and text encoders that are contrastively pretrained on a large number of image and text pairs [37]. The image encoder is adapted to detection by removing the final pooling layer and adding heads that predict bounding boxes and class embeddings directly from the output tokens of the image encoder. For open-vocabulary object classification, the embeddings computed by the class prediction head are compared to text encoder embeddings of object descriptions. This architecture achieves strong open-vocabulary object detection performance [33, 34] and does not suffer from the training instabilities observed in some encoder-decoder detection models [3, 34, 43, 50, 58, 63]." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "text", + "content": "T. Salzmann et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 161, + 116, + 452, + 202 + ], + "blocks": [ + { + "bbox": [ + 161, + 116, + 452, + 202 + ], + "lines": [ + { + "bbox": [ + 161, + 116, + 452, + 202 + ], + "spans": [ + { + "bbox": [ + 161, + 116, + 452, + 202 + ], + "type": "image", + "image_path": "11851b4cb9f82a6929578514b366470e94ad3b87a5308ffb27f48cd4a4f5073a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 171, + 203, + 431, + 300 + ], + "blocks": [ + { + "bbox": [ + 171, + 203, + 431, + 300 + ], + "lines": [ + { + "bbox": [ + 171, + 203, + 431, + 300 + ], + "spans": [ + { + "bbox": [ + 171, + 203, + 431, + 300 + ], + "type": "image", + "image_path": "f042bb5958900c574775ed1f1666401d8fbde29f3dc02abd99e8992bb68e8497.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 311, + 482, + 421 + ], + "lines": [ + { + "bbox": [ + 130, + 311, + 482, + 421 + ], + "spans": [ + { + "bbox": [ + 130, + 311, + 482, + 421 + ], + "type": "text", + "content": "Figure 2: For relationship selection, image tokens are first projected using two lightweight MLPs to produce and embeddings. A relationship score is then computed as the inner product between all and embeddings. Relationships are filtered by first selecting the top object instances, using the scores along the diagonal to represent instance likelihood. Among the remaining instances, the top pairs are selected using the off-diagonal scores. This yields a set of relationship triplets, each consisting of a index, an index, and a relationship embedding that is computed by summing the respective and embeddings. For classification, the relationship embeddings are compared against text embeddings of object class or predicate text descriptions." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 443, + 348, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 443, + 348, + 455 + ], + "spans": [ + { + "bbox": [ + 131, + 443, + 348, + 455 + ], + "type": "text", + "content": "3.2 Extension to Relationship Prediction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 462, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 581 + ], + "type": "text", + "content": "In the architecture described above, each image encoder token represents an object proposal that captures per-object information. Importantly, the encoder consists of self-attention layers which, by design, model all pairwise interactions between these tokens. We therefore hypothesize that information about object relationships can be learned directly in the image encoder. To extract this information in the form of triplets, we introduce two MLPs that transform vision encoder output tokens into embeddings " + }, + { + "bbox": [ + 130, + 462, + 482, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_i" + }, + { + "bbox": [ + 130, + 462, + 482, + 581 + ], + "type": "text", + "content": " and embeddings " + }, + { + "bbox": [ + 130, + 462, + 482, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{o}_j" + }, + { + "bbox": [ + 130, + 462, + 482, + 581 + ], + "type": "text", + "content": " (Figure 2). Using different MLPs for and is crucial to break symmetry in subsequent processing (Appendix A.1)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "text", + "content": "An embedding representing the relationship between two object proposals (encoder tokens) can then be obtained simply by element-wise addition of their and embeddings. Obtaining relationship embeddings for all " + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "inline_equation", + "content": "N^2" + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "text", + "content": " pairs, where " + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "text", + "content": " is the number of object proposals, would be computationally infeasible. We therefore introduce a Relationship Attention layer that performs hard attention to select the pairs most likely to form a relationship. This layer computes an attention-like score " + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "inline_equation", + "content": "p_{ij} =" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 375, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 375, + 91, + 447, + 102 + ], + "type": "text", + "content": "Scene-Graph ViT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 236 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_i \\mathbf{o}_j^T" + }, + { + "bbox": [ + 130, + 115, + 482, + 236 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 115, + 482, + 236 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_i" + }, + { + "bbox": [ + 130, + 115, + 482, + 236 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 115, + 482, + 236 + ], + "type": "inline_equation", + "content": "\\mathbf{o}_j" + }, + { + "bbox": [ + 130, + 115, + 482, + 236 + ], + "type": "text", + "content": " are the embedding vectors of " + }, + { + "bbox": [ + 130, + 115, + 482, + 236 + ], + "type": "inline_equation", + "content": "<\\text { subject }>" + }, + { + "bbox": [ + 130, + 115, + 482, + 236 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 115, + 482, + 236 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 115, + 482, + 236 + ], + "type": "inline_equation", + "content": "<\\text { object }>" + }, + { + "bbox": [ + 130, + 115, + 482, + 236 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 115, + 482, + 236 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 115, + 482, + 236 + ], + "type": "inline_equation", + "content": "p_{ij}" + }, + { + "bbox": [ + 130, + 115, + 482, + 236 + ], + "type": "text", + "content": " represents the likelihood that a relationship between " + }, + { + "bbox": [ + 130, + 115, + 482, + 236 + ], + "type": "inline_equation", + "content": "<\\text { subject }>" + }, + { + "bbox": [ + 130, + 115, + 482, + 236 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 115, + 482, + 236 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 115, + 482, + 236 + ], + "type": "inline_equation", + "content": "<\\text { object }>" + }, + { + "bbox": [ + 130, + 115, + 482, + 236 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 115, + 482, + 236 + ], + "type": "text", + "content": " exists. By computing this score for all " + }, + { + "bbox": [ + 130, + 115, + 482, + 236 + ], + "type": "inline_equation", + "content": "<\\text { subject - object }>" + }, + { + "bbox": [ + 130, + 115, + 482, + 236 + ], + "type": "text", + "content": " pairs, we obtain an " + }, + { + "bbox": [ + 130, + 115, + 482, + 236 + ], + "type": "inline_equation", + "content": "N \\times N" + }, + { + "bbox": [ + 130, + 115, + 482, + 236 + ], + "type": "text", + "content": " matrix, from which we select the top " + }, + { + "bbox": [ + 130, + 115, + 482, + 236 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 115, + 482, + 236 + ], + "type": "text", + "content": " pairs. For these " + }, + { + "bbox": [ + 130, + 115, + 482, + 236 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 115, + 482, + 236 + ], + "type": "text", + "content": " pairs, we compute relationship embeddings " + }, + { + "bbox": [ + 130, + 115, + 482, + 236 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{ij} = \\text { LayerNorm }(\\mathbf{s}_i + \\mathbf{o}_j)" + }, + { + "bbox": [ + 130, + 115, + 482, + 236 + ], + "type": "text", + "content": ". To classify the relationship of a pair, its relationship embedding is compared to text embeddings of relationship predicates. We use the embeddings where " + }, + { + "bbox": [ + 130, + 115, + 482, + 236 + ], + "type": "inline_equation", + "content": "<\\text { subject }>" + }, + { + "bbox": [ + 130, + 115, + 482, + 236 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 115, + 482, + 236 + ], + "type": "inline_equation", + "content": "<\\text { object }>" + }, + { + "bbox": [ + 130, + 115, + 482, + 236 + ], + "type": "text", + "content": " are identical (" + }, + { + "bbox": [ + 130, + 115, + 482, + 236 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{i=j}" + }, + { + "bbox": [ + 130, + 115, + 482, + 236 + ], + "type": "text", + "content": ") to represent object instances, and predict object classes from them. Boxes are predicted from the corresponding image encoder output tokens (see Figure 2)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 236, + 482, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 482, + 320 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 482, + 320 + ], + "type": "text", + "content": "The Relationship Attention layer therefore identifies object pairs to focus on for relationship classification, by computing a hard attention with embeddings as queries and embeddings as keys. Since the hard selection is not differentiable, gradients from the relationship prediction cannot be used directly to train this layer. Instead, the relationship score is self-supervised to predict the maximum class probability that will later be predicted for the relationship at the classification stage (Section 3.3)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 338, + 204, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 338, + 204, + 350 + ], + "spans": [ + { + "bbox": [ + 132, + 338, + 204, + 350 + ], + "type": "text", + "content": "3.3 Training" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 358, + 483, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 358, + 483, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 358, + 483, + 430 + ], + "type": "text", + "content": "The image and text encoders of our model are initialized from a vision-language model that was contrastively pretrained on a large number of image-text pairs [37]. After adding the Relationship Attention as well as the object bounding box and class prediction heads, the model is jointly trained on a mixture of object and relationship detection datasets in a single training stage (Section 4.1), using the losses described below." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 430, + 482, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 482, + 514 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 482, + 514 + ], + "type": "text", + "content": "Bipartite Matching. Bipartite matching between object predictions " + }, + { + "bbox": [ + 130, + 430, + 482, + 514 + ], + "type": "inline_equation", + "content": "(\\mathbf{r}_{i = j})" + }, + { + "bbox": [ + 130, + 430, + 482, + 514 + ], + "type": "text", + "content": " and ground-truth annotations is performed based on a cost consisting of the object classification loss and the box prediction losses as in [3]. This matching between objects also establishes a matching of predicted to ground-truth relationship predicates, since a relationship is uniquely identified by the and indices " + }, + { + "bbox": [ + 130, + 430, + 482, + 514 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 430, + 482, + 514 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 430, + 482, + 514 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 430, + 482, + 514 + ], + "type": "text", + "content": ". Unmatched predictions are trained to predict low scores for all classes and incur no box prediction loss." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 514, + 482, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 514, + 482, + 539 + ], + "spans": [ + { + "bbox": [ + 130, + 514, + 482, + 539 + ], + "type": "text", + "content": "Box Prediction Losses. For bounding box regression, we use the L1 and generalized intersection-over-union (gIoU) losses described in [3]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 539, + 482, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 539, + 482, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 539, + 482, + 635 + ], + "type": "text", + "content": "Object and Predicate Classification Loss. For both object category and predicate classification, we use a sigmoid cross-entropy loss as in [34]. This loss is computed between the ground truth classes and logits obtained by computing the inner product of the relationship embeddings " + }, + { + "bbox": [ + 130, + 539, + 482, + 635 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{ij}" + }, + { + "bbox": [ + 130, + 539, + 482, + 635 + ], + "type": "text", + "content": " selected by the Relationship Attention layer with the text embeddings of the class names. For embeddings corresponding to individual objects " + }, + { + "bbox": [ + 130, + 539, + 482, + 635 + ], + "type": "inline_equation", + "content": "(\\mathbf{r}_{i = j})" + }, + { + "bbox": [ + 130, + 539, + 482, + 635 + ], + "type": "text", + "content": ", the class name is the object category or description. For embeddings corresponding to relationships " + }, + { + "bbox": [ + 130, + 539, + 482, + 635 + ], + "type": "inline_equation", + "content": "(\\mathbf{r}_{i\\neq j})" + }, + { + "bbox": [ + 130, + 539, + 482, + 635 + ], + "type": "text", + "content": ", we use the predicate text as class name. This differs from prior work [53, 54, 61], which" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "text", + "content": "T. Salzmann et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 642, + 481, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 481, + 666 + ], + "type": "text", + "content": "This selection reduces the number of relationships that need to be processed, e.g., from " + }, + { + "bbox": [ + 133, + 642, + 481, + 666 + ], + "type": "inline_equation", + "content": "N^2 = 3600^2" + }, + { + "bbox": [ + 133, + 642, + 481, + 666 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 133, + 642, + 481, + 666 + ], + "type": "inline_equation", + "content": "k = 16386" + }, + { + "bbox": [ + 133, + 642, + 481, + 666 + ], + "type": "text", + "content": " (99.9% reduction) for our ViT-L/14 model." + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": "uses the full triplet description. Our approach has the advantage of disentangling object category and predicate names, which allows for more efficient inference since object category and predicate texts are embedded separately." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 482, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 482, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 482, + 295 + ], + "type": "text", + "content": "Relationship Score Loss. To select only the most promising embeddings for further processing, the Relationship Attention layer predicts a score " + }, + { + "bbox": [ + 130, + 164, + 482, + 295 + ], + "type": "inline_equation", + "content": "p_{ij}" + }, + { + "bbox": [ + 130, + 164, + 482, + 295 + ], + "type": "text", + "content": " that represents the likelihood that the corresponding pair forms a relationship (if " + }, + { + "bbox": [ + 130, + 164, + 482, + 295 + ], + "type": "inline_equation", + "content": "i \\neq j" + }, + { + "bbox": [ + 130, + 164, + 482, + 295 + ], + "type": "text", + "content": "), or that the corresponding object exists in the image (if " + }, + { + "bbox": [ + 130, + 164, + 482, + 295 + ], + "type": "inline_equation", + "content": "i = j" + }, + { + "bbox": [ + 130, + 164, + 482, + 295 + ], + "type": "text", + "content": "). This score is trained with a sigmoid cross-entropy loss against targets provided by the model itself, namely the maximum probability predicted for any class for the corresponding " + }, + { + "bbox": [ + 130, + 164, + 482, + 295 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{ij}" + }, + { + "bbox": [ + 130, + 164, + 482, + 295 + ], + "type": "text", + "content": " embedding [33]. In this way, the relationship score is trained to predict the class probability of a potential embedding " + }, + { + "bbox": [ + 130, + 164, + 482, + 295 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{ij}" + }, + { + "bbox": [ + 130, + 164, + 482, + 295 + ], + "type": "text", + "content": " before actually computing and classifying the embedding. Note that this loss can only be computed for those objects and relationships that ultimately get selected for further processing. We found this to provide sufficient supervision." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 296, + 482, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 296, + 482, + 321 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 482, + 321 + ], + "type": "text", + "content": "The final loss is an equally weighted sum of four components: (1) classification loss, (2) L1 box loss, (3) gIoU box loss, (4) relationship score loss." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 339, + 230, + 354 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 339, + 230, + 354 + ], + "spans": [ + { + "bbox": [ + 131, + 339, + 230, + 354 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 365, + 262, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 365, + 262, + 379 + ], + "spans": [ + { + "bbox": [ + 131, + 365, + 262, + 379 + ], + "type": "text", + "content": "4.1 Experimental Setup" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 387, + 432, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 387, + 432, + 399 + ], + "spans": [ + { + "bbox": [ + 131, + 387, + 432, + 399 + ], + "type": "text", + "content": "Datasets. We use the following datasets for training or evaluation:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 399, + 482, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 482, + 470 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 482, + 470 + ], + "type": "text", + "content": "Visual Genome [26] is the largest VRD dataset, labeled with 2.3M triplet relationships across 108K images. However, as Visual Genome includes noisy annotations, the community commonly evaluates VRD on a cleaned version of the dataset [55] where the label space is reduced to the 150 most frequent object classes and the 50 most frequent predicate classes. This dataset is commonly referred to as Visual Genome 150 (VG150)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 471, + 482, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 471, + 482, + 542 + ], + "spans": [ + { + "bbox": [ + 130, + 471, + 482, + 542 + ], + "type": "text", + "content": "GQA [16] uses the same image corpus as Visual Genome, but is more diversely labeled, with 1703 object classes (1704 including a background class) and 310 predicates. Spatial relationships like \"to the left of\" are automatically labeled based on the 2-D spatial arrangement of bounding boxes. Similar to VQ150, GQA200 is a reduced and cleaned version of GQA with 200 object classes and 100 predicate classes [9]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 543, + 482, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 543, + 482, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 543, + 482, + 590 + ], + "type": "text", + "content": "HICO-DET [4] is a dataset which focuses on a more specialized subset of visual relationships, specifically interactions between humans and objects. HICO-DET contains 50k images and is exhaustively annotated for 600 defined human-object interaction (HOIs)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 590, + 482, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 590, + 482, + 615 + ], + "spans": [ + { + "bbox": [ + 130, + 590, + 482, + 615 + ], + "type": "text", + "content": "Objects365 [40] (O365) is a large scale object-detection dataset with 365 object categories across two million images." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 615, + 482, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 615, + 482, + 675 + ], + "spans": [ + { + "bbox": [ + 130, + 615, + 482, + 675 + ], + "type": "text", + "content": "Open X-Embodiment [35] (OXE) dataset is targeted towards learning vision language action policies for robotics. OXE lacks bounding box annotations which precludes quantitative evaluation, but it represents both a visually different setting compared to the standard VRD benchmarks and a promising application area, so we use it for qualitative evaluation." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 375, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 375, + 90, + 447, + 102 + ], + "type": "text", + "content": "Scene-Graph ViT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 141, + 114, + 473, + 150 + ], + "blocks": [ + { + "bbox": [ + 141, + 114, + 473, + 150 + ], + "lines": [ + { + "bbox": [ + 141, + 114, + 473, + 150 + ], + "spans": [ + { + "bbox": [ + 141, + 114, + 473, + 150 + ], + "type": "table", + "html": "
VG150VGHICOGQA200GQAObjects365
% in training mixture12.5%12.5%12.5%12.5%0%50%
# removed658717727729153611880115306
", + "image_path": "78b26de48b2c3d16bdca8e3dcec542e37b691cc232e06ba3a512f8866181f2ae.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 156, + 480, + 178 + ], + "lines": [ + { + "bbox": [ + 132, + 156, + 480, + 178 + ], + "spans": [ + { + "bbox": [ + 132, + 156, + 480, + 178 + ], + "type": "text", + "content": "Table 1: Training data mixture. # removed indicates the number of images that were removed from the official training split due to overlap with data we evaluate on." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 192, + 482, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 192, + 482, + 313 + ], + "spans": [ + { + "bbox": [ + 130, + 192, + 482, + 313 + ], + "type": "text", + "content": "Training. Unless otherwise noted, we train models on a mixture of VG, VG150, GQA200, HICO, and O365 in the proportions show in Table 1. The B/32 model is trained for 200'000 steps at batch size 256 (further details in Appendix A.1). Since the datasets we use share similar image sources, there is some overlap between the official training and evaluation splits. To ensure that no evaluation images are used for training, we rigorously filter all training datasets to remove images present in any of the test splits for all datasets we evaluate on, i.e. VG150, GQA, GQA200, and HICO. We use an image similarity filter that also identifies non-exact matches [24]. Table 1 shows the number of images removed from each training split. Note that not all prior work performs similar dedduplication." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 324, + 482, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 324, + 482, + 432 + ], + "spans": [ + { + "bbox": [ + 130, + 324, + 482, + 432 + ], + "type": "text", + "content": "Evaluation Procedure and Metrics. Prior work presents a diverse range of evaluation metrics, often tailored to specific datasets. For non-exhaustively labelled datasets, like Visual Genome and GQA, using precision-based metrics is inherently inconclusive. Thus, the community has adapted a " + }, + { + "bbox": [ + 130, + 324, + 482, + 432 + ], + "type": "inline_equation", + "content": "Recall@K" + }, + { + "bbox": [ + 130, + 324, + 482, + 432 + ], + "type": "text", + "content": " metric for these datasets, where " + }, + { + "bbox": [ + 130, + 324, + 482, + 432 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 324, + 482, + 432 + ], + "type": "text", + "content": " denotes a fixed budget of triplets on which the recall is computed. However, this metric is biased towards the more frequent classes in the dataset [6,46]. Therefore, recent approaches use mean " + }, + { + "bbox": [ + 130, + 324, + 482, + 432 + ], + "type": "inline_equation", + "content": "Recall@K" + }, + { + "bbox": [ + 130, + 324, + 482, + 432 + ], + "type": "text", + "content": ", which calculates " + }, + { + "bbox": [ + 130, + 324, + 482, + 432 + ], + "type": "inline_equation", + "content": "Recall@K" + }, + { + "bbox": [ + 130, + 324, + 482, + 432 + ], + "type": "text", + "content": " separately for each class in the test data and then averages the results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 433, + 482, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 433, + 482, + 493 + ], + "spans": [ + { + "bbox": [ + 130, + 433, + 482, + 493 + ], + "type": "text", + "content": "For exhaustively labeled datasets like HICO, it is feasible to use precision-based metrics such as mean Average Precision (mAP), a well-established metric in object detection. For HICO, the mean is taken over the 600 possible HOI triplets. Besides the overall metric, we also report results separately for rare (< 10 occurrences) and non-rare (" + }, + { + "bbox": [ + 130, + 433, + 482, + 493 + ], + "type": "inline_equation", + "content": "\\geq 10" + }, + { + "bbox": [ + 130, + 433, + 482, + 493 + ], + "type": "text", + "content": " occurrences) HOIs as mAPr and mAPn." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 494, + 482, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 494, + 482, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 494, + 482, + 601 + ], + "type": "text", + "content": "Further, evaluation can be \"graph-constrained\", where only a single prediction can be made per detected object pair, or \"graph-unconstrained\", where any number of predictions can be made per pair. Existing works evaluating on Visual Genome and GQA datasets partially lack a clear distinction between graph-constrained and unconstrained evaluation. For our models, we report graph-constrained results in the main paper, which reflects the dominant approach used in other studies. Results without graph-constraint are reported in Appendix A.2. For the HICO dataset, we follow the established evaluation procedure [4] that inherently represents a graph-unconstrained evaluation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 602, + 482, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 602, + 482, + 674 + ], + "spans": [ + { + "bbox": [ + 130, + 602, + 482, + 674 + ], + "type": "text", + "content": "Multiple implementations of the aforementioned metrics exist, often showing differences in results. To compare fairly and directly against a wide range of prior approaches we identified the PyTorch evaluation procedure of [45] as the most prevalent routine for recall-based metrics and replicated their procedure in JAX with numerical accuracy. Similarly, for HICO, we numerically reproduce the original Matlab evaluation procedure outlined in [4] in JAX." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "text", + "content": "T. Salzmann et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 165, + 114, + 445, + 264 + ], + "blocks": [ + { + "bbox": [ + 165, + 114, + 445, + 264 + ], + "lines": [ + { + "bbox": [ + 165, + 114, + 445, + 264 + ], + "spans": [ + { + "bbox": [ + 165, + 114, + 445, + 264 + ], + "type": "table", + "html": "
ModelBackbonemR@50mR@100
RelTR [7]ResNet-506.810.8
Transformer + CFA [28]ResNeXt-10112.314.6
VCTree [46] + IETrans + Rwt [57]ResNeXt-10112.014.9
Motif [55] + IETrans + Rwt [57]ResNeXt-10115.518.0
GPS-Net [31] + IETrans + Rwt [57]ResNeXt-10116.218.8
SG-Transformer [52] + IETrans + Rwt [57]ResNeXt-10115.018.1
Sgtr [29]ResNet-10115.820.1
DT2-ACBS [8]ResNet-10122.024.4
UniVRDCLIP: ViT-B/329.612.1
UniVRDCLIP: ViT-B/1610.913.2
UniVRDCLIP: ViT-L/1412.614.5
SG-ViTCLIP: ViT-B/3215.018.1
SG-ViTCLIP: ViT-B/1615.719.3
SG-ViTCLIP: ViT-L/1417.821.8
SG-ViT + simple predicate rebalancingCLIP: ViT-L/1422.326.1
", + "image_path": "9de4ecc27567bb3999b7891e1f07a0ce180fb938c80be318a3fdd4dd62831d7c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 269, + 482, + 313 + ], + "lines": [ + { + "bbox": [ + 130, + 269, + 482, + 313 + ], + "spans": [ + { + "bbox": [ + 130, + 269, + 482, + 313 + ], + "type": "text", + "content": "Table 2: Performance on the Visual Genome 150 test set. We use graph-constrained evaluation (one predicate prediction per pair). For the last row (\"+ simple text rebalancing\"), the model was briefly fine-tuned on VG150 data with rebalanced predicate frequencies (see Section 4.2 and Appendix A.1)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 325, + 482, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 325, + 482, + 458 + ], + "spans": [ + { + "bbox": [ + 130, + 325, + 482, + 458 + ], + "type": "text", + "content": "Exhaustive Relationship Evaluation. Previous methods [53,54,61] entangle objects and relationships in a single representation. Consequently, these methods score their embeddings against the full triplet. While this is feasible for datasets with a low number of possible triplets, like HICO with 600 defined relationships, it quickly becomes computationally intractable for datasets with a larger vocabulary of objects and predicates (e.g. VG, GQA). Prior works [61] therefore evaluate only on triplets which are known to be present in the test split, which may inflate metrics. In contrast, our method disentangles objects and predicates (Section 3). This allows for an exhaustive evaluation that considers all object and predicate combinations and enables applications where knowledge of possible relationships is unavailable." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 476, + 343, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 476, + 343, + 488 + ], + "spans": [ + { + "bbox": [ + 132, + 476, + 343, + 488 + ], + "type": "text", + "content": "4.2 Relationship Detection Performance" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 497, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 482, + 628 + ], + "type": "text", + "content": "We use Visual Genome as our main benchmark for relationship detection. Table 2 shows results for our models and prior work. Prior studies often achieve reported performance through the combination of a base model with a specialized approach to overcome the skewed data distribution (indicated as base-model + specialization in Table 2). Our models provide strong performance even without special treatment of the skewed predicate distribution. After brief fine-tuning on data in which the frequency of predicate annotations was rebalanced with simple rejection sampling (see Appendix A.1), our method improves on the prior best method DT2-ACBS [8] by 1.7 points mR@100 (26.1 vs. 24.4). This shows that our already-strong method may be further improved by combining it with more advanced rebalancing methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": "We also note the large difference in performance to UniVRD [61], which builds on the same detection architecture as our method, but adds a Transformer decoder-based relationship module. To disentangle whether the improve" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 375, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 375, + 91, + 447, + 102 + ], + "type": "text", + "content": "Scene-Graph ViT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 125, + 302, + 215 + ], + "blocks": [ + { + "bbox": [ + 132, + 125, + 302, + 215 + ], + "lines": [ + { + "bbox": [ + 132, + 125, + 302, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 125, + 302, + 215 + ], + "type": "table", + "html": "
ModelmR@50mR@100
LLM4SGG [21]5.36.5
Neural Motif w/ GCL [9]16.818.8
VCTree w/ GCL [9]15.617.8
SG-Transformer w/ CFA [28]13.415.3
SHA w/ GCL [9]17.820.1
SG-ViT (CLIP: ViT-B/32)15.718.6
SG-ViT (CLIP: ViT-B/16)17.720.5
SG-ViT (CLIP: ViT-L/14)19.322.9
", + "image_path": "f390adbd569e1eb58b400e7d878db4fbdbc05aace5b222067c696053c7732b2f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 313, + 114, + 476, + 228 + ], + "blocks": [ + { + "bbox": [ + 132, + 218, + 298, + 251 + ], + "lines": [ + { + "bbox": [ + 132, + 218, + 298, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 218, + 298, + 251 + ], + "type": "text", + "content": "Table 3: Performance on the GQA200 test set. Graph-constrained evaluation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 114, + 476, + 228 + ], + "lines": [ + { + "bbox": [ + 313, + 114, + 476, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 114, + 476, + 228 + ], + "type": "table", + "html": "
ModelmR@50mR@100
Scene Graph Classification
IMP [25,44,48]0.50.7
Neural Motif [25,44,55]0.81.2
Unbiased TDE [25,45]-0.7
RTN [25]-1.4
MP [23]-2.8
Transformer w/ EBM [44]1.31.8
Scene Graph Generation
SG-ViT (CLIP: ViT-B/32)6.27.4
SG-ViT (CLIP: ViT-B/16)6.47.4
SG-ViT (CLIP: ViT-L/14)8.09.6
", + "image_path": "9d576e06afa31662f0e93e37f961b032f41f88d11d56a31d6cc6a8bad2bf12c2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 229, + 478, + 251 + ], + "lines": [ + { + "bbox": [ + 313, + 229, + 478, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 229, + 478, + 251 + ], + "type": "text", + "content": "Table 4: Performance on the GQA test set. Graph-constrained evaluation." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 263, + 480, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 263, + 480, + 324 + ], + "spans": [ + { + "bbox": [ + 130, + 263, + 480, + 324 + ], + "type": "text", + "content": "ments over UniVRD are due to differences in architecture or training data, we also trained our B/32 model on the UniVRD data mixture, and still observe a large improvement (17.4 VG mR@100 for our method vs. 12.1 for UniVRD when trained on the same data; Table 6). This suggests that our encoder-only architecture is better suited for VRD than prior decoder-based architectures [7,20,29,61]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 343, + 309, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 309, + 355 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 309, + 355 + ], + "type": "text", + "content": "4.3 Scaling to Large Vocabularies" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 365, + 482, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 509 + ], + "type": "text", + "content": "Improving long-tail performance has been critical to the VRD field since the natural distribution of relationship triplets is highly skewed [27]. A substantial body of literature is devoted to this goal, often proposing complex loss debiasing or data augmentation approaches [8,27,52,57]. It is therefore of interest whether our method works well in this regime without special treatment of rare classes. To evaluate the performance of our method on large object and predicate vocabularies with a long tail of rare classes, we use the Graph Question Answering (GQA) dataset [16]. This dataset builds on VG and expands the number of classes and annotations. A simplified version of GQA with 200 object and 100 predicate classes, called GQA200 [9], is part of our training mixture. Our method surpasses prior results on the GQA200 test split (Table 3), despite lacking the specialized treatment of the data distribution that some prior methods use." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 510, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 593 + ], + "type": "text", + "content": "The full GQA dataset has an even larger vocabulary with 1703 object and 311 predicate classes. To our knowledge, no prior methods evaluate scene graph generation on the full GQA dataset. Some prior methods report results on scene graph classification, where ground-truth boxes are provided. As shown in Table 4, our model achieves higher performance on the much harder task of scene graph generation than prior methods on classification. We further assess zero-shot generalization to unseen classes in Appendix A.2." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "We believe that two factors contribute to the strong performance in the open vocabulary regime: First, the simple design of the method does not impede transfer of semantic knowledge from the pre-trained VLM backbone because it adds only lightweight heads. In particular, we use no decoder, which could be difficult to train and could lead to \"forgetting\" of pretrained representations. Instead, most of the relationship modeling happens in the backbone, which has" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "text", + "content": "T. Salzmann et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "had the benefit of large-scale vision-language pretraining. Second, the open-vocabulary design allows end-to-end training on a mix of datasets, which was not possible for all prior methods." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 171, + 295, + 183 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 171, + 295, + 183 + ], + "spans": [ + { + "bbox": [ + 131, + 171, + 295, + 183 + ], + "type": "text", + "content": "4.4 Human-Object Interaction" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 191, + 482, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 191, + 482, + 312 + ], + "spans": [ + { + "bbox": [ + 130, + 191, + 482, + 312 + ], + "type": "text", + "content": "We also evaluate our model on the specialized task of human-object interaction, using the HICO benchmark [4]. The performance of our model is on par with the most comparable prior method [61], but does not show similarly large improvements as for VG or GQA200 (Table 5). HICO has a much narrower and more specialized vocabulary than VG or GQA (HICO evaluates on 600 prespecified triplets, whereas VG150 has " + }, + { + "bbox": [ + 130, + 191, + 482, + 312 + ], + "type": "inline_equation", + "content": "1.1 \\times 10^{6}" + }, + { + "bbox": [ + 130, + 191, + 482, + 312 + ], + "type": "text", + "content": " possible triplet combinations). Performance on this specialized task may benefit less from transfer of pretrained representations and may instead be limited by the amount of task-specific training data. This is supported by the fact that the recent state-of-the-art method for HICO pretrains on large amounts of person-focused pseudo-labels [54]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 329, + 211, + 340 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 329, + 211, + 340 + ], + "spans": [ + { + "bbox": [ + 131, + 329, + 211, + 340 + ], + "type": "text", + "content": "4.5 Ablations" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 347, + 369, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 347, + 369, + 394 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 369, + 394 + ], + "type": "text", + "content": "Inference Speed and Number of Predicted Relationships. In the following, we ablate training datasets and model components to study the interplay of data and architecture on the model performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 396, + 369, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 396, + 369, + 466 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 369, + 466 + ], + "type": "text", + "content": "The primary hyperparameter of our method is the number of relationships " + }, + { + "bbox": [ + 130, + 396, + 369, + 466 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 396, + 369, + 466 + ], + "type": "text", + "content": " that are selected by the Relationship Attention layer from all possible combinations for further processing. Figure 3 shows how VRD performance and inference speed depend on " + }, + { + "bbox": [ + 130, + 396, + 369, + 466 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 396, + 369, + 466 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 467, + 369, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 467, + 369, + 586 + ], + "spans": [ + { + "bbox": [ + 130, + 467, + 369, + 586 + ], + "type": "text", + "content": "For training, we use " + }, + { + "bbox": [ + 130, + 467, + 369, + 586 + ], + "type": "inline_equation", + "content": "k = 2^{14}" + }, + { + "bbox": [ + 130, + 467, + 369, + 586 + ], + "type": "text", + "content": ", which results in " + }, + { + "bbox": [ + 130, + 467, + 369, + 586 + ], + "type": "inline_equation", + "content": "71\\%" + }, + { + "bbox": [ + 130, + 467, + 369, + 586 + ], + "type": "text", + "content": " (B/32) to " + }, + { + "bbox": [ + 130, + 467, + 369, + 586 + ], + "type": "inline_equation", + "content": "81\\%" + }, + { + "bbox": [ + 130, + 467, + 369, + 586 + ], + "type": "text", + "content": " (L/14) of the speed of a pure object detection model [34] of the same size. Without top-k selection, the model would be several times slower (right end of top plot in Figure 3). Since the vast majority of pairs do not form relationships, VRD performance is essentially unaffected by top- " + }, + { + "bbox": [ + 130, + 467, + 369, + 586 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 467, + 369, + 586 + ], + "type": "text", + "content": " selection at inference. For inference, we can therefore use " + }, + { + "bbox": [ + 130, + 467, + 369, + 586 + ], + "type": "inline_equation", + "content": "k = 2^{11}" + }, + { + "bbox": [ + 130, + 467, + 369, + 586 + ], + "type": "text", + "content": " at no loss in accuracy to achieve " + }, + { + "bbox": [ + 130, + 467, + 369, + 586 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 130, + 467, + 369, + 586 + ], + "type": "text", + "content": " of the speed of the object detector (for B/32 and" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 586, + 481, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 586, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 130, + 586, + 481, + 611 + ], + "type": "text", + "content": "L/14). At this " + }, + { + "bbox": [ + 130, + 586, + 481, + 611 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 586, + 481, + 611 + ], + "type": "text", + "content": ", the B/32 model achieves 52.8 FPS at batch size 1 on an NVIDIA V100 GPU and is therefore suitable for real-time applications." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 377, + 352, + 479, + 411 + ], + "blocks": [ + { + "bbox": [ + 377, + 352, + 479, + 411 + ], + "lines": [ + { + "bbox": [ + 377, + 352, + 479, + 411 + ], + "spans": [ + { + "bbox": [ + 377, + 352, + 479, + 411 + ], + "type": "image", + "image_path": "6ce65786911baddd0584536ae1881f94620020316915a4e5beca96624dd3b69e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 378, + 418, + 479, + 497 + ], + "blocks": [ + { + "bbox": [ + 378, + 418, + 479, + 497 + ], + "lines": [ + { + "bbox": [ + 378, + 418, + 479, + 497 + ], + "spans": [ + { + "bbox": [ + 378, + 418, + 479, + 497 + ], + "type": "image", + "image_path": "ee12aaefd35ea4036b65e7cfd654c798dde90954e058b06c94ce55efc763b626.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 373, + 509, + 482, + 576 + ], + "lines": [ + { + "bbox": [ + 373, + 509, + 482, + 576 + ], + "spans": [ + { + "bbox": [ + 373, + 509, + 482, + 576 + ], + "type": "text", + "content": "Figure 3: Model speed and VRD accuracy by number of selected relationships " + }, + { + "bbox": [ + 373, + 509, + 482, + 576 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 373, + 509, + 482, + 576 + ], + "type": "text", + "content": ". Speed is relative to a non-VRD object detector [34]." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Dataset Mixture. To assess how training data and architectural improvements contribute to the performance of our model compared to prior work, we train the models on a range of dataset mixtures (Table 6). The full mixture, which we use for all experiments unless otherwise noted, includes VG, VG150, GQA200," + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 375, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 375, + 91, + 447, + 102 + ], + "type": "text", + "content": "Scene-Graph ViT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 215, + 114, + 399, + 216 + ], + "blocks": [ + { + "bbox": [ + 215, + 114, + 399, + 216 + ], + "lines": [ + { + "bbox": [ + 215, + 114, + 399, + 216 + ], + "spans": [ + { + "bbox": [ + 215, + 114, + 399, + 216 + ], + "type": "table", + "html": "
ModelBackbonemAPmAPrmAPn
HOTR [20]ResNet5025.1020.3325.86
RLIP [53]ResNet5032.8426.8534.63
RLIPv2 [54]ResNet5035.3829.6137.10
RLIPv2 [54]Swin-L45.0943.2345.64
UniVRD [60]CLIP: ViT-B/3229.9822.9432.02
UniVRD [60]CLIP: ViT-B/1629.9822.9432.02
UniVRD [60]CLIP: ViT-L/1437.4128.9039.95
SG-ViTCLIP: ViT-B/3228.8623.7230.39
SG-ViTCLIP: ViT-B/1631.9826.8333.52
SG-ViTCLIP: ViT-L/1438.1133.7139.42
", + "image_path": "7189384ec9b4a431c9263347626b2060294e3550b9ec7fc0665b71f9cab14225.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 132, + 303, + 482, + 369 + ], + "blocks": [ + { + "bbox": [ + 130, + 223, + 482, + 280 + ], + "lines": [ + { + "bbox": [ + 130, + 223, + 482, + 280 + ], + "spans": [ + { + "bbox": [ + 130, + 223, + 482, + 280 + ], + "type": "text", + "content": "Table 5: Performance on the HICO test set. Our method is on par with the most comparable prior work, UniVRD, for overall mAP. However, our method performs better on \"rare\" classes (mAPr), indicating better generalization to rarely seen concepts. For reference, we also show RLIPv2, the state-of-the-art method on HICO. RLIPv2 proposes a self-training approach that is orthogonal and compatible to our method." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 303, + 482, + 369 + ], + "lines": [ + { + "bbox": [ + 132, + 303, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 303, + 482, + 369 + ], + "type": "table", + "html": "
VGMixtureVG150 mR@100GQA200 mR@100HICO mAP
VG150GQA200HICOO365COCO
Our mixture12.5%12.5%12.5%12.5%50%-18.118.628.9
UniVRD [61] mix-10%-20%50%20%17.4 ↓ 4%5.1 ↓ 73%17.4 ↓ 8%
ours w/o O36525%25%25%25%--15.3 ↓ 15%15.6 ↓ 16%23.8 ↓ 18%
ours w/o VG-25%12.5%12.5%50%-14.2 ↓ 22%15.3 ↓ 18%26.8 ↓ 7%
", + "image_path": "6336eef133c28a416384855a11e66d7c6a278eca3ef146ca2f8fe6350c40a875.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 376, + 482, + 399 + ], + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 399 + ], + "type": "text", + "content": "Table 6: Dataset Ablation. Performance of SG-ViT (B/32) trained on different data mixtures. VG150 and GQA200 evaluation is graph-constrained, HICO is unconstrained." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 435, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 435, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 435, + 482, + 533 + ], + "type": "text", + "content": "HICO and O365. To allow direct comparison to UniVRD [61], we also train on their mixture (VG150, HICO, O365, COCO) and find that our model still performs better on VG150 by a large margin (18.1 mR@100 vs 12.1 for UniVRD for B/32 models). Since UniVRD uses the same image and text encoding backbones as our model, this result suggests that our encoder-only model with Relationship Attention is an advance over decoder-based relationship architectures. We further find that including both pure object detection (O365) and large-vocabulary VRD data (VG) benefit all metrics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Object Detection Performance. A necessary step towards good relationship detection performance is accurate object detection. In our architecture, object detection is treated as a special case of relationship detection in which the is identical to the . We validate this design by training the model only on detection datasets without relationship annotations, using the same dataset mixture as the OWL-ViT [34] model (Objects365 and Visual Genome, [34]). We find that our model achieves a similar object detection performance as OWL-ViT (21.9% vs. 22.1% mAP and 18.3% vs. 18.9% mAPr on LVIS [12]), indicating that the Relationship Attention design does not interfere with object detection." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 241, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 241, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 241, + 100 + ], + "type": "text", + "content": "T. Salzmann et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 113, + 465, + 364 + ], + "blocks": [ + { + "bbox": [ + 149, + 113, + 465, + 364 + ], + "lines": [ + { + "bbox": [ + 149, + 113, + 465, + 364 + ], + "spans": [ + { + "bbox": [ + 149, + 113, + 465, + 364 + ], + "type": "image", + "image_path": "f6264e2b1fc727434aab771caeec0d0fb4ceb3c09d21ec7a9cf8800abb6cbdaa.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 371, + 482, + 407 + ], + "lines": [ + { + "bbox": [ + 131, + 371, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 131, + 371, + 482, + 407 + ], + "type": "text", + "content": "Figure 4: Qualitative examples of difficult edge-cases from VG150 test split [26, 55]. From left to right: Ground Truth, SG-ViT (B/32), SG-ViT (L/14). In all cases the is lime and the is red." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 429, + 257, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 429, + 257, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 257, + 441 + ], + "type": "text", + "content": "4.6 Qualitative Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 449, + 482, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 569 + ], + "type": "text", + "content": "In Figure 4, we visualize qualitative examples to highlight difficult edge cases. Even where the model output differs from the ground truth, it is often plausibly correct, e.g. where either the extent of an object differs with the ground truth or there are multiple appropriate, but unannotated, relations. For example, both models select the headlight on the bus in Figure 4b for \"light on bus\" as opposed to the sun reflected off the windshield. In Figure 4c both models focus more on the \"on\" in \"snow on hill\" than the distinction between a mountain and a hill. Figure 4d shows a challenging scene where the B/32 model selects the bottle inside the open fridge while the L/14 correctly selects a water bottle in the door for \"bottle in door\" while also capturing the extent of the door instance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Furthermore, to qualitatively assess the utility of this work beyond standard relation benchmark datasets, we also visualize some example relations applied to data from the robotics domain, namely the OXE dataset [35]. Figure 5 shows a real-life scene where multiple instances of the same semantic object categories are present within the scene. The model is able to match the to the instances in the desired configuration. One subjective failure-case is that the \"banana outside bowl\" text query had its highest scoring triplet involving a bowl on a shelf in the background rather than the adjacent bowl. Another charac" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 375, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 375, + 91, + 447, + 102 + ], + "type": "text", + "content": "Scene-Graph ViT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 146, + 125, + 225, + 195 + ], + "blocks": [ + { + "bbox": [ + 146, + 125, + 225, + 195 + ], + "lines": [ + { + "bbox": [ + 146, + 125, + 225, + 195 + ], + "spans": [ + { + "bbox": [ + 146, + 125, + 225, + 195 + ], + "type": "image", + "image_path": "96211429d2725ae9e1b59b8a13180e29bd18aef541e76869374e0860caf512bb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 228, + 125, + 306, + 195 + ], + "blocks": [ + { + "bbox": [ + 228, + 125, + 306, + 195 + ], + "lines": [ + { + "bbox": [ + 228, + 125, + 306, + 195 + ], + "spans": [ + { + "bbox": [ + 228, + 125, + 306, + 195 + ], + "type": "image", + "image_path": "254994729b19ecd82ba91ce8146a55b6a254d15717861e67e505b8f10ec2d570.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 310, + 125, + 388, + 195 + ], + "blocks": [ + { + "bbox": [ + 310, + 125, + 388, + 195 + ], + "lines": [ + { + "bbox": [ + 310, + 125, + 388, + 195 + ], + "spans": [ + { + "bbox": [ + 310, + 125, + 388, + 195 + ], + "type": "image", + "image_path": "aeb341dbc46b61dda18fddfb4d8b9a969af189f75879cfa642faca97a3dd5b71.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 392, + 125, + 470, + 195 + ], + "blocks": [ + { + "bbox": [ + 392, + 125, + 470, + 195 + ], + "lines": [ + { + "bbox": [ + 392, + 125, + 470, + 195 + ], + "spans": [ + { + "bbox": [ + 392, + 125, + 470, + 195 + ], + "type": "image", + "image_path": "78b4213f16e188704cfe8e1fae668aed3ec2aee932542858578e5c5d5925f81b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 146, + 197, + 225, + 266 + ], + "blocks": [ + { + "bbox": [ + 146, + 197, + 225, + 266 + ], + "lines": [ + { + "bbox": [ + 146, + 197, + 225, + 266 + ], + "spans": [ + { + "bbox": [ + 146, + 197, + 225, + 266 + ], + "type": "image", + "image_path": "1e4342115292a3f92fbbfd5e311b2de5f1967605b4f32a18b54289b3c989ea20.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 274, + 480, + 318 + ], + "lines": [ + { + "bbox": [ + 131, + 274, + 480, + 318 + ], + "spans": [ + { + "bbox": [ + 131, + 274, + 480, + 318 + ], + "type": "text", + "content": "Figure 5: Qualitative examples showing SG-ViT (L/14) on out-of-distribution data from the OXE dataset [35]. In all cases the is lime and the is red. Note how the model correctly disambiguates several instances of the same class (e.g. \"bananas\" and \"bottle\") depending on their relationships." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 228, + 197, + 306, + 266 + ], + "blocks": [ + { + "bbox": [ + 228, + 197, + 306, + 266 + ], + "lines": [ + { + "bbox": [ + 228, + 197, + 306, + 266 + ], + "spans": [ + { + "bbox": [ + 228, + 197, + 306, + 266 + ], + "type": "image", + "image_path": "a83cbc2f571ab5f2471b9a5a44ab9afb5748de82eb8c623823c3f7337ee5a21e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 310, + 197, + 388, + 266 + ], + "blocks": [ + { + "bbox": [ + 310, + 197, + 388, + 266 + ], + "lines": [ + { + "bbox": [ + 310, + 197, + 388, + 266 + ], + "spans": [ + { + "bbox": [ + 310, + 197, + 388, + 266 + ], + "type": "image", + "image_path": "906289885ef00cead1092aadcb6e8ec257528af9fdc0677df59a21814e53f5dc.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 392, + 197, + 470, + 266 + ], + "blocks": [ + { + "bbox": [ + 392, + 197, + 470, + 266 + ], + "lines": [ + { + "bbox": [ + 392, + 197, + 470, + 266 + ], + "spans": [ + { + "bbox": [ + 392, + 197, + 470, + 266 + ], + "type": "image", + "image_path": "de418ba07b8a61d2a01bf4bd1a4ab2e4ad2ca3e10f1458c1001e6e19619aaf4d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 339, + 480, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 339, + 480, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 339, + 480, + 399 + ], + "type": "text", + "content": "teristic is that the model occasionally places boxes over groups with multiple instances of the same semantic class when a singular word is used for either the or . This can be attributed to the similarity between word and their plural in the CLIP embeddings and that many human annotations from VG also confuse singular and plural (see Appendix A.3)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 414, + 222, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 222, + 426 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 222, + 426 + ], + "type": "text", + "content": "5 Limitations" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 436, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 436, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 436, + 482, + 533 + ], + "type": "text", + "content": "While our model performs strongly on large-vocabulary VRD, its performance on specialized human-object interaction detection is only on par with prior models. We discuss this limitation in Section 4.4. Further error modes are explored qualitatively in Section 4.6 and Appendix A.3. A challenge for open-vocabulary relationship detection models as a whole is zero-shot generalization to unseen objects and predicates. While our model improves over prior approaches (Tables 3 and 4), there is still a large gap when it comes to entirely unseen classes (Appendix A.2). Future research should focus on closing this gap." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 548, + 220, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 548, + 220, + 560 + ], + "spans": [ + { + "bbox": [ + 132, + 548, + 220, + 560 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "We present an architecture for open-vocabulary visual relationship detection. By combining an encoder-only design with a novel Relationship Attention layer for efficient selection of high-confidence relationships, our architecture leverages VLM pretraining and multi-dataset VRD training. It achieves strong performance on standard and large-vocabulary VRD benchmarks while maintaining pure object detection performance and adding little extra computational cost. Due to its simplicity and strong performance, we believe that our method will be a useful basis for further research on visual relationship detection." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "text", + "content": "T. Salzmann et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 139, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 139, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 139, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 138, + 139, + 481, + 172 + ], + "type": "text", + "content": "1. Amiri, S., Chandan, K., Zhang, S.: Reasoning with scene graphs for robot planning under partial observability. IEEE Robotics and Automation Letters 7(2), 5560-5567 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 173, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 173, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 138, + 173, + 481, + 194 + ], + "type": "text", + "content": "2. Ba, J.L., Kiros, J.R., Hinton, G.E.: Layer normalization. arXiv preprint arXiv:1607.06450 (2016)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 194, + 481, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 194, + 481, + 227 + ], + "spans": [ + { + "bbox": [ + 138, + 194, + 481, + 227 + ], + "type": "text", + "content": "3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: Conference on Computer Vision (ECCV). pp. 213-229. Springer (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 228, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 481, + 259 + ], + "type": "text", + "content": "4. Chao, Y.W., Liu, Y., Liu, X., Zeng, H., Deng, J.: Learning to Detect Human-Object Interactions. In: IEEE Winter Conference on Applications of Computer Vision (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 260, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 260, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 138, + 260, + 481, + 304 + ], + "type": "text", + "content": "5. Chen, B., Xia, F., Ichter, B., Rao, K., Gopalakrishnan, K., Ryoo, M.S., Stone, A., Kappler, D.: Open-vocabulary queryable scene representations for real world planning. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 11509-11522. IEEE (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 304, + 481, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 304, + 481, + 336 + ], + "spans": [ + { + "bbox": [ + 138, + 304, + 481, + 336 + ], + "type": "text", + "content": "6. Chen, T., Yu, W., Chen, R., Lin, L.: Knowledge-embedded routing network for scene graph generation. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 6163-6171 (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 337, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 337, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 138, + 337, + 481, + 369 + ], + "type": "text", + "content": "7. Cong, Y., Yang, M.Y., Rosenhahn, B.: Reltr: Relation transformer for scene graph generation. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 370, + 481, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 370, + 481, + 403 + ], + "spans": [ + { + "bbox": [ + 138, + 370, + 481, + 403 + ], + "type": "text", + "content": "8. Desai, A., Wu, T.Y., Tripathi, S., Vasconcelos, N.: Learning of visual relations: The devil is in the tails. In: IEEE/CVF International Conference on Computer Vision (ICCV). pp. 15404-15413 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 403, + 481, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 403, + 481, + 446 + ], + "spans": [ + { + "bbox": [ + 138, + 403, + 481, + 446 + ], + "type": "text", + "content": "9. Dong, X., Gan, T., Song, X., Wu, J., Cheng, Y., Nie, L.: Stacked hybrid-attention and group collaborative learning for unbiased scene graph generation. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 19427-19436 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 134, + 447, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 447, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 134, + 447, + 481, + 490 + ], + "type": "text", + "content": "10. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. ICLR (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 491, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 491, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 134, + 491, + 481, + 533 + ], + "type": "text", + "content": "11. Gu, Q., Kuwajerwala, A., Morin, S., Jatavallabhula, K.M., Sen, B., Agarwal, A., Rivera, C., Paul, W., Ellis, K., Chellappa, R., et al.: Conceptgraphs: Open-vocabulary 3d scene graphs for perception and planning. arXiv preprint arXiv:2309.16650 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 534, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 534, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 134, + 534, + 481, + 567 + ], + "type": "text", + "content": "12. Gupta, A., Dollar, P., Girshick, R.: LVIS: A Dataset for Large Vocabulary Instance Segmentation. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 567, + 481, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 567, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 134, + 567, + 481, + 600 + ], + "type": "text", + "content": "13. He, T., Gao, L., Song, J., Li, Y.F.: Towards Open-Vocabulary Scene Graph Generation with Prompt-Based Finetuning. In: European Conference on Computer Vision (ECCV) (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 601, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 601, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 134, + 601, + 481, + 621 + ], + "type": "text", + "content": "14. Hendrycks, D., Gimpel, K.: Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415 (2016)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 622, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 622, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 622, + 481, + 665 + ], + "type": "text", + "content": "15. Herzig, R., Mendelson, A., Karlinsky, L., Arbelle, A., Feris, R., Darrell, T., Globerson, A.: Incorporating Structured Representations into Pretrained Vision & Language Models Using Scene Graphs. In: The 2023 Conference on Empirical Methods in Natural Language Processing (2023)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 375, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 375, + 91, + 447, + 102 + ], + "type": "text", + "content": "Scene-Graph ViT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "16. Hudson, D.A., Manning, C.D.: Gqa: A new dataset for real-world visual reasoning and compositional question answering. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 6700-6709 (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 184 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 184 + ], + "type": "text", + "content": "17. Hughes, N., Chang, Y., Carlone, L.: Hydra: A Real-time Spatial Perception System for 3D Scene Graph Construction and Optimization. Robotics: Science and Systems (RSS) (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 185, + 481, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 185, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 185, + 481, + 228 + ], + "type": "text", + "content": "18. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International conference on machine learning. pp. 4904-4916. PMLR (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 228, + 481, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 228, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 481, + 262 + ], + "type": "text", + "content": "19. Johnson, J., Krishna, R., Stark, M., Li, L.J., Shamma, D., Bernstein, M., Fei-Fei, L.: Image Retrieval Using Scene Graphs. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2015)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 262, + 481, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 262, + 481, + 296 + ], + "spans": [ + { + "bbox": [ + 132, + 262, + 481, + 296 + ], + "type": "text", + "content": "20. Kim, B., Lee, J., Kang, J., Kim, E.S., Kim, H.J.: Hotr: End-to-end human-object interaction detection with transformers. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 74-83 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 297, + 481, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 297, + 481, + 329 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 481, + 329 + ], + "type": "text", + "content": "21. Kim, K., Yoon, K., Jeon, J., In, Y., Moon, J., Kim, D., Park, C.: LLM4SGG: Large Language Model for Weakly Supervised Scene Graph Generation. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 330, + 481, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 330, + 481, + 352 + ], + "spans": [ + { + "bbox": [ + 132, + 330, + 481, + 352 + ], + "type": "text", + "content": "22. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 353, + 481, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 353, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 353, + 481, + 385 + ], + "type": "text", + "content": "23. Knyazev, B., de Vries, H., Cangea, C., Taylor, G.W., Courville, A., Belilovsky, E.: Graph Density-Aware Losses for Novel Compositions in Scene Graph Generation. In: British Machine Vision Conference (BMVC) (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 386, + 481, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 481, + 419 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 481, + 419 + ], + "type": "text", + "content": "24. Kolesnikov, A., Beyer, L., Zhai, X., Puigcerver, J., Yung, J., Gelly, S., Houlsby, N.: Big transfer (bit): General visual representation learning. In: European Conference on Computer Vision (ECCV). pp. 491-507 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 420, + 481, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 420, + 481, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 420, + 481, + 441 + ], + "type": "text", + "content": "25. Koner, R., Shit, S., Tresp, V.: Relation transformer network. arXiv preprint arXiv:2004.06193 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 442, + 481, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 442, + 481, + 486 + ], + "spans": [ + { + "bbox": [ + 132, + 442, + 481, + 486 + ], + "type": "text", + "content": "26. Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., et al.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. International Journal of Computer Vision 123, 32-73 (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 487, + 481, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 481, + 529 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 481, + 529 + ], + "type": "text", + "content": "27. Li, H., Zhu, G., Zhang, L., Jiang, Y., Dang, Y., Hou, H., Shen, P., Zhao, X., Shah, S.A.A., Bennamoun, M.: Scene Graph Generation: A comprehensive survey. Neurocomputing (2024). https://doi.org/https://doi.org/10.1016/j.neucom.2023.127052" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 531, + 481, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 531, + 481, + 564 + ], + "spans": [ + { + "bbox": [ + 132, + 531, + 481, + 564 + ], + "type": "text", + "content": "28. Li, L., Chen, G., Xiao, J., Yang, Y., Wang, C., Chen, L.: Compositional feature augmentation for unbiased scene graph generation. In: IEEE/CVF International Conference on Computer Vision (ICCV). pp. 21685-21695 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 565, + 481, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 565, + 481, + 598 + ], + "spans": [ + { + "bbox": [ + 132, + 565, + 481, + 598 + ], + "type": "text", + "content": "29. Li, R., Zhang, S., He, X.: Sgtr: End-to-end scene graph generation with transformer. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 19486-19496 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "type": "text", + "content": "30. Li, Y., Ouyang, W., Zhou, B., Wang, K., Wang, X.: Scene graph generation from objects, phrases and region captions. In: IEEE/CVF International Conference on Computer Vision (ICCV). pp. 1261-1270 (2017)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "31. Lin, X., Ding, C., Zeng, J., Tao, D.: Gps-net: Graph property sensing network for scene graph generation. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3746-3753 (2020)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 242, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 242, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 242, + 100 + ], + "type": "text", + "content": "T. Salzmann et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "32. Lu, C., Krishna, R., Bernstein, M., Fei-Fei, L.: Visual relationship detection with language priors. In: European Conference on Computer Vision (ECCV). pp. 852-869. Springer (2016)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 481, + 171 + ], + "type": "text", + "content": "33. Minderer, M., Gritsenko, A., Houlsby, N.: Scaling open-vocabulary object detection. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 171, + 481, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 171, + 481, + 214 + ], + "spans": [ + { + "bbox": [ + 130, + 171, + 481, + 214 + ], + "type": "text", + "content": "34. Minderer, M., Gritsenko, A., Stone, A., Neumann, M., Weissenborn, D., Dosovitskiy, A., Mahendran, A., Arnab, A., Dehghani, M., Shen, Z., et al.: Simple open-vocabulary object detection. In: European Conference on Computer Vision (ECCV). pp. 728-755. Springer (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 214, + 481, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 214, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 214, + 481, + 236 + ], + "type": "text", + "content": "35. Open X-Embodiment Collaboration: Open X-Embodiment: Robotic Learning Datasets and RT-X Models (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 236, + 481, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 481, + 267 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 481, + 267 + ], + "type": "text", + "content": "36. Peng, Z., Wang, W., Dong, L., Hao, Y., Huang, S., Ma, S., Ye, Q., Wei, F.: Grounding Multimodal Large Language Models to the World. In: International Conference on Learning Representations (2024)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 267, + 481, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 267, + 481, + 311 + ], + "spans": [ + { + "bbox": [ + 130, + 267, + 481, + 311 + ], + "type": "text", + "content": "37. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 311, + 481, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 311, + 481, + 343 + ], + "spans": [ + { + "bbox": [ + 130, + 311, + 481, + 343 + ], + "type": "text", + "content": "38. Rana, K., Haviland, J., Garg, S., Abou-Chakra, J., Reid, I., Suenderhauf, N.: Say-plan: Grounding large language models using 3d scene graphs for scalable robot task planning. In: 7th Annual Conference on Robot Learning (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 343, + 481, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 343, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 481, + 374 + ], + "type": "text", + "content": "39. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems 28 (2015)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 374, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 374, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 481, + 407 + ], + "type": "text", + "content": "40. Shao, S., Li, Z., Zhang, T., Peng, C., Yu, G., Zhang, X., Li, J., Sun, J.: Objects365: A large-scale, high-quality dataset for object detection. In: IEEE/CVF International Conference on Computer Vision (ICCV). pp. 8430-8439 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 407, + 481, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 481, + 439 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 481, + 439 + ], + "type": "text", + "content": "41. Shi, H., Hayat, M., Cai, J.: Open-vocabulary object detection via scene graph discovery. In: 31st ACM International Conference on Multimedia. pp. 4012-4021 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 439, + 481, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 439, + 481, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 439, + 481, + 472 + ], + "type": "text", + "content": "42. Singh, K.P., Salvador, J., Weihs, L., Kembhavi, A.: Scene graph contrastive learning for embodied navigation. In: IEEE/CVF International Conference on Computer Vision (ICCV). pp. 10884-10894 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 472, + 481, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 481, + 504 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 481, + 504 + ], + "type": "text", + "content": "43. Song, H., Sun, D., Chun, S., Jampani, V., Han, D., Heo, B., Kim, W., Yang, M.H.: ViDT: An Efficient and Effective Fully Transformer-based Object Detector. In: International Conference on Learning Representations (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 504, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 504, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 504, + 481, + 536 + ], + "type": "text", + "content": "44. Suhail, M., Mittal, A., Siddiquie, B., Broaddus, C., Eledath, J., Medioni, G., Sigal, L.: Energy-based learning for scene graph generation. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13936-13945 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 536, + 481, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 536, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 536, + 481, + 568 + ], + "type": "text", + "content": "45. Tang, K., Niu, Y., Huang, J., Shi, J., Zhang, H.: Unbiased scene graph generation from biased training. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3716-3725 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 568, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 568, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 568, + 481, + 601 + ], + "type": "text", + "content": "46. Tang, K., Zhang, H., Wu, B., Luo, W., Liu, W.: Learning to compose dynamic tree structures for visual contexts. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 6619-6628 (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 601, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 601, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 130, + 601, + 481, + 633 + ], + "type": "text", + "content": "47. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 633, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 633, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 633, + 481, + 666 + ], + "type": "text", + "content": "48. Xu, D., Zhu, Y., Choy, C.B., Fei-Fei, L.: Scene graph generation by iterative message passing. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5410-5419 (2017)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 375, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 375, + 91, + 447, + 102 + ], + "type": "text", + "content": "Scene-Graph ViT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 621 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "49. Yang, J., Lu, J., Lee, S., Batra, D., Parikh, D.: Graph r-cnn for scene graph generation. In: European Conference on Computer Vision (ECCV). pp. 670-685 (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "type": "text", + "content": "50. Yao, Z., Ai, J., Li, B., Zhang, C.: Efficient DETR: Improving End-to-End Object Detector With Dense Prior. arXiv preprint arXiv:2104.01318 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 161, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 480, + 194 + ], + "type": "text", + "content": "51. Yu, J., Wang, Z., Vasudevan, V., Yeung, L., Seyedhosseini, M., Wu, Y.: CoCa: Contrastive Captioners are Image-Text Foundation Models. Transactions on Machine Learning Research (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 236 + ], + "type": "text", + "content": "52. Yu, J., Chai, Y., Wang, Y., Hu, Y., Wu, Q.: CogTree: Cognition Tree Loss for Unbiased Scene Graph Generation. In: International Joint Conference on Artificial Intelligence, IJCAI-21. International Joint Conferences on Artificial Intelligence Organization (8 2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "text", + "content": "53. Yuan, H., Jiang, J., Albaniae, S., Feng, T., Huang, Z., Ni, D., Tang, M.: Rlip: Relational language-image pre-training for human-object interaction detection. Advances in Neural Information Processing Systems 35, 37416-37431 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 271, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 480, + 314 + ], + "type": "text", + "content": "54. Yuan, H., Zhang, S., Wang, X., Albanie, S., Pan, Y., Feng, T., Jiang, J., Ni, D., Zhang, Y., Zhao, D.: RLIPv2: Fast Scaling of Relational Language-Image Pretraining. In: IEEE/CVF International Conference on Computer Vision (ICCV). pp. 21649-21661 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "type": "text", + "content": "55. Zellers, R., Yatskar, M., Thomson, S., Choi, Y.: Neural motifs: Scene graph parsing with global context. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5831-5840 (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "text", + "content": "56. Zhai, X., Wang, X., Mustafa, B., Steiner, A., Keysers, D., Kolesnikov, A., Beyer, L.: Lit: Zero-shot transfer with locked-image text tuning. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 18123-18133 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "text", + "content": "57. Zhang, A., Yao, Y., Chen, Q., Ji, W., Liu, Z., Sun, M., Chua, T.S.: Fine-grained scene graph generation with data transfer. In: European Conference on Computer Vision (ECCV). pp. 409-424. Springer (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "type": "text", + "content": "58. Zhang, H., Li, F., Liu, S., Zhang, L., Su, H., Zhu, J., Ni, L.M., Shum, H.Y.: Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 446, + 480, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 480, + 488 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 480, + 488 + ], + "type": "text", + "content": "59. Zhang, Y., Ma, Z., Gao, X., Shakiah, S., Gao, Q., Chai, J.: GROUNDHOG: Grounding Large Language Models to Holistic Segmentation. In: Conference on Computer Vision and Pattern Recognition 2024 (2024), https://openreview.net/forum?id=PWW3ux1Ju" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 489, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 489, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 489, + 480, + 522 + ], + "type": "text", + "content": "60. Zhao, L., Yuan, L., Gong, B., Cui, Y., Schroff, F., Yang, M.H., Adam, H., Liu, T.: Unified Visual Relationship Detection with Vision and Language Models. In: IEEE/CVF International Conference on Computer Vision (ICCV) (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 522, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 480, + 555 + ], + "type": "text", + "content": "61. Zhao, L., Yuan, L., Gong, B., Cui, Y., Schroff, F., Yang, M.H., Adam, H., Liu, T.: Unified Visual Relationship Detection with Vision and Language Models. In: IEEE/CVF International Conference on Computer Vision (ICCV) (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 555, + 480, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 480, + 587 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 480, + 587 + ], + "type": "text", + "content": "62. Zhou, X., Girdhar, R., Joulin, A., Krahenbuhl, P., Misra, I.: Detecting twenty-thousand classes using image-level supervision. In: European Conference on Computer Vision (ECCV). pp. 350-368 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 588, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 588, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 480, + 621 + ], + "type": "text", + "content": "63. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable DETR: Deformable Transformers for End-to-End Object Detection. In: International Conference on Learning Representations (2021)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "text", + "content": "T. Salzmann et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Scene-aware Human Motion Forecasting via Mutual Distance Prediction/abecdb1c-2c3f-4eb7-820c-6ae2bc3ee016_content_list.json b/2024/Scene-aware Human Motion Forecasting via Mutual Distance Prediction/abecdb1c-2c3f-4eb7-820c-6ae2bc3ee016_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7b512adf12b368dd34e8c9a35c5cebc2f09bc875 --- /dev/null +++ b/2024/Scene-aware Human Motion Forecasting via Mutual Distance Prediction/abecdb1c-2c3f-4eb7-820c-6ae2bc3ee016_content_list.json @@ -0,0 +1,1636 @@ +[ + { + "type": "text", + "text": "Scene-aware Human Motion Forecasting via Mutual Distance Prediction", + "text_level": 1, + "bbox": [ + 250, + 141, + 754, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chaoyue Xing $^{1}$ , Wei Mao $^{1}$ , and Miaomiao Liu $^{1}$", + "bbox": [ + 328, + 210, + 671, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Australian National University, Canberra, Australia", + "bbox": [ + 323, + 239, + 676, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. In this paper, we address the issue of scene-aware 3D human motion forecasting. A key challenge in this task is to predict future human motions that are coherent with the scene by modeling human-scene interactions. While recent works have demonstrated that explicit constraints on human-scene interactions can prevent the occurrence of ghost motion, they only provide constraints on partial human motion e.g., the global motion of the human or a few joints contacting the scene, leaving the rest of unconstrained. To address this limitation, we propose to represent the human-scene interaction using the mutual distance between the human body and the scene. Such mutual distances constrain both the local and global human motion, resulting in a whole-body motion constrained prediction. In particular, mutual distance constraints consist of two components, the signed distance of each vertex on the human mesh to the scene surface and the distance of basis scene points to the human mesh. We further introduce a global scene representation learned from a signed distance function (SDF) volume to ensure coherence between the global scene representation and the explicit constraint from the mutual distance. We develop a pipeline with two sequential steps: predicting the future mutual distances first, followed by forecasting future human motion. We explicitly ensure consistency between predicted poses and mutual distances during training. Extensive testing on both synthetic and real datasets demonstrates that our method consistently surpasses the performance of current state-of-the-art techniques.", + "bbox": [ + 261, + 287, + 743, + 607 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Scene-aware $\\cdot$ Human Motion Forecasting $\\cdot$ Mutual Distance $\\cdot$ SDF Volume", + "bbox": [ + 259, + 619, + 740, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 672, + 375, + 690 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Human motion forecasting has numerous applications, including autonomous driving [23], human-robot interaction [15], animation [31] and virtual/augmented reality [29]. Despite being an active research field for decades [2,28,30,33], most previous methods [1, 3, 8, 16, 17, 21, 22, 32] focus solely on human motion and overlook the fact that humans interact with the 3D environment. In contrast, this paper addresses the problem of scene-aware human motion prediction.", + "bbox": [ + 212, + 704, + 787, + 794 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advancements in scene-aware human synthesis [11,34,37] and human motion prediction [4,5,20,40] have begun to model the human-scene interaction. Some approaches achieve this implicitly by incorporating scene encoding as an", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "additional input to their model e.g., 2D scene image [4], 3D scene point cloud [34] or a specific object [5]. However, without explicit constraints, those methods tend to predict motions with artifacts e.g., \"ghost motion\". Alternatively, other works [20, 27, 37, 40] start to explore the explicit representation of human-scene interaction. However, none of their representations are able to constrain the whole-body motion both locally and globally. For example, the proximity between human and scene used in [37] mainly constrains the global position of the human, while the human gaze employed in [40] only specifies the potential target location of the human. Although the contact map proposed in [20, 27] achieves detailed constraints on both the local pose and global position of certain human joints, their contact map can only constrain joints that contact the scene, neglecting the others, such as elbow joints, which often remain unsupported.", + "bbox": [ + 212, + 146, + 787, + 328 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/732e84fb3e1960611be0718822119a4755b9f3b1df65d655504dbbdd5309c682.jpg", + "image_caption": [ + "(a) The per-vertex signed distance (b) The per-basis point distance" + ], + "image_footnote": [], + "bbox": [ + 303, + 356, + 483, + 478 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2502a106d2db1d11117b5842c681cc1ca215eb8fbcc4c5f6f88b26536dd4a397.jpg", + "image_caption": [ + "Fig. 1: Our mutual distances. a) shows the per-vertex signed distance for sampled vertices on human mesh. Their color indicates the distance value. b) shows the per-basis point distance for the basis points which are not sampled on the scene surface. For both figures, the darker the color is, the smaller the distance is." + ], + "image_footnote": [], + "bbox": [ + 527, + 357, + 705, + 477 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To constrain the whole-body motion for better prediction, in this paper, we propose a mutual distance representation which captures the distances between the human body and the scene. We further introduce the global scene representation learned from SDF volume instead of point cloud to achieve consistent global and local scene representation learning. Specifically, our proposed mutual distance representation comprises two components: the per-vertex signed distance and the per-basis point distance. Given a 3D scene, the per-vertex distance is the minimum distance from each human body vertex to the scene surface, and its sign indicates whether such vertex penetrates the scene or not (see Fig. 1 (a)). Since the change in signed distance for each body vertex represents the motion of the corresponding body part, it provides a complete constraint on the entire body. However, as a signed distance corresponds to a level-set of the scene Signed Distance Field (SDF), a human mesh at multiple possible global locations can result in the same per-vertex signed distance.", + "bbox": [ + 212, + 595, + 787, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this, we leverage a per-basis point distance, which calculates the minimum distance between each point in a predefined set of basis points and the", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "C. Xing et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "human surface. Our per-basis point distance is similar to BPS body encoding in PLACE [37] with differences in choices of basis points. In particular, their basis points are partially scene-dependent making it not suitable for motion prediction, while our basis points are purely scene-independent and sampled with a fixed distance relative to origin, which can better capture the motion patterns.", + "bbox": [ + 212, + 146, + 782, + 220 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While our model is end-to-end trainable, we adopt a two-step training process followed by fine-tuning for better convergence. In the first step, we predict the mutual distance. Conditioning on the predicted distances, we then forecast the future human motion. Specifically, given a 3D scene, the past human poses, and past mutual distance, we use a strategy based on Graph Convolutional Network (GCN) [14], Discrete Cosine Transform (DCT), for the spatial and temporal encoding of the mutual distances and a simple 3D CNN for encoding of the SDF volume to predict the future ones. We subsequently take the predicted distances, the past human poses, and the scene as input to predict future human motion using an RNN-based model. A consistency prior is adopted to ensure that the predicted human motion is consistent with the input mutual distance.", + "bbox": [ + 212, + 222, + 784, + 387 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows. i) We propose an explicit representation of human-scene interaction using mutual distance to provide constraints for whole-body motion. ii) We explore different global scene representations and find the SDF volume to be more efficient than commonly used point cloud representation. Our experiments demonstrate that our approach outperforms the state-of-the-art scene-aware human motion prediction methods by a large margin on both existing real and synthesis benchmark datasets.", + "bbox": [ + 212, + 388, + 785, + 494 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 518, + 387, + 534 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Human motion prediction. Researchers have been studying 3D human motion modeling [2,28,30,33] for decades. While traditional methods, such as those employing Hidden Markov Models [2] or the Gaussian process latent variable model [33], are capable of handling periodic and simple non-periodic motions like walking and golf swings, more complex motions are typically modeled with deep neural networks [1,3,8,16,17,21,22,32]. However, these methods focus solely on the prediction of local poses, neglecting the global human motion and 3D scenes in which the motion takes place.", + "bbox": [ + 212, + 551, + 784, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "More recent works [4,5,20,27,40] have begun to incorporate scene context into the task of motion prediction. Specifically, Cao et al. [4] propose to represent the scene with a 2D image. The feature extracted from such an image is fed into the prediction model and lets the model learn the human-scene interaction implicitly. However, such 2D scene representation cannot handle occlusions and it does not ensure a valid human-scene interaction. In [5], a semantic graph is introduced to encode the interaction between the human and the object of interest. Although it is efficient to model a human interacting with one object, it is not suitable for human-scene interaction as the human may interact with multiple objects when he moves around. In [40], Zheng et al. employ eye gaze as additional information to guide the motion forecasting process. Although", + "bbox": [ + 212, + 674, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Scene-aware Human Motion Forecasting via Mutual Distance Prediction", + "bbox": [ + 250, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "eye gaze provides valuable hints about the goal location of the human motion, it constrains neither the local poses nor the human-scene interaction. Mao et al. [20] introduce a contact map that can explicitly encode the relation between scene surface points and human joints. Scofano et al. [27] propose a three-stage pipeline in a coarse-to-fine manner and utilizing the time information based on [20]. However, for forecasting, their contact map only constrains joints that contact the scene and disregard the rest. By contrast, our introduced mutual distance is able to constrain the whole-body motion.", + "bbox": [ + 212, + 146, + 787, + 267 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Scene-aware human synthesis. One of the main challenges in this task is to ensure a valid human-scene interaction. As such, existing works have explored various strategies to model such interaction [9, 11, 34, 35, 37, 39]. Similar to [4], the interaction is modeled implicitly in [34, 35, 39] via either Generative Adversarial Networks (GANs) [34], or Variational Autoencoders [34, 39]. Besides this, other works [9, 11, 37] adopt explicit representation of human-scene interaction. In particular, it was proposed in POSA [11] to develop a human-centric representation of the human-scene interaction which assigns scene semantic labels, such as floor, and chair, to each human surface vertex. Those labels indicate the probability of such vertex contacting what kind of object in the scene. Although it performs well in the task of placing a human in a scene, it lacks sufficient detail about the local pose of the human, rendering it unsuitable for human motion forecasting. In SAMP [9], Hassan et al. propose to only model the interaction with the object at the goal location. Given the final root translation and orientation, the task is then treated as a path planning problem. Similar to [5], such a strategy cannot constrain the interactions between the human and the scene when the human moves in the scene. The closest to our work is PLACE [37] where the human-scene interaction is represented by the minimum distance between a set of 3D scene points also known as Basis Point Sets (BSP) [26] to the human body surface. However, such representation alone only provides a prior on the global position of the human. For example, one can move its arm that is far from the scene surface while the scene to human proximity remains the same. By contrast, we combine the per-basis point distance with the per-vertex signed distance resulting in a detailed constraint on the whole-body motion both locally and globally.", + "bbox": [ + 212, + 267, + 787, + 645 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Approach", + "text_level": 1, + "bbox": [ + 215, + 666, + 346, + 683 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let us begin by formulating the task of scene-aware 3D human motion forecasting. Building on previous work [20], we represent a sequence of $T$ past human poses as $\\mathbf{X} = [\\mathbf{x}_1, \\mathbf{x}_2, \\dots, \\mathbf{x}_T] \\in \\mathbb{R}^{M \\times T^1}$ , where a human pose at time $t$ is represented as a vector $\\mathbf{x}_t \\in \\mathbb{R}^M$ of length $M$ . Such vector could be a parametric representation of human mesh such as SMPL-X [25] pose. To represent the 3D scene, we adopt a Signed Distance Volume $\\mathbf{S} \\in \\mathbb{R}^{N \\times N \\times N}$ , each entry of which stores the distance of the voxel to the nearest scene surface and its sign indicates whether it is inside or outside the scene. The representation learned from this", + "bbox": [ + 212, + 696, + 787, + 818 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "C. Xing et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1 Unless otherwise specified, in this paper, we use column vector convention.", + "bbox": [ + 217, + 824, + 730, + 839 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/29e607422490830b415c1545159ba801100ffb8ccf80748ee9e5074c897ab307.jpg", + "image_caption": [ + "Fig. 2: Network architecture. Given the 3D scene $\\mathbf{S}$ represented as a signed distance volume and the past motion $\\mathbf{X}$ shown in grey meshes, our approach first predicts the future per-vertex signed distance $\\hat{\\mathbf{D}}$ , and the future per-basis point distance $\\hat{\\mathbf{B}}$ from historical distance $\\mathbf{D}$ and $\\mathbf{B}$ , respectively. The two predicted future distances are then fed into a RNN-based network to predict the future motion $\\hat{\\mathbf{Y}}$ shown in orange meshes." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 787, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "volume is more consistent with our mutual distance representation than those learned from point cloud [20]. Moreover, such Signed Distance Volume allows us to query the signed distance of any given position easily via interpolation, which facilitates the fast calculation of the loss defined on the SDFs of the predicted pose sequence during training (see Eq. 14).", + "bbox": [ + 212, + 484, + 787, + 560 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our objective is predicting $U$ future human poses $\\mathbf{Y} = [\\mathbf{x}_{T + 1},\\mathbf{x}_{T + 2},\\dots ,\\mathbf{x}_{T + U}]\\in$ $\\mathbb{R}^{M\\times U}$ with history motion $\\mathbf{X}$ and 3D scene S. We introduce an innovative pipeline that predicts mutual distance representation first i.e., the signed minimum distance from each human body vertex to the scene surface (the per-vertex signed distance) and the minimum distance from each basis point to the human surface (the per-basis point distance) at each time frame. Conditioned on such mutual distance, our model then forecasts future poses. Our pipeline is illustrated in Fig. 2. We detail our mutual distance and the pipeline for predicting these distances and future motions below.", + "bbox": [ + 212, + 561, + 810, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Mutual Distance Representation", + "text_level": 1, + "bbox": [ + 214, + 720, + 532, + 736 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Let us first introduce our mutual distance representation, which consists of the per-vertex signed distance and the per-basis point distance.", + "bbox": [ + 212, + 750, + 785, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The per-vertex signed distance. Given the human pose $\\mathbf{x}_t$ at time $t$ , one can obtain the vertices of human mesh with a predefined mapping, such as SMPL-X [25]. However, constraining every single vertex may not be necessary due to their strong correlation. Hence, following [38], we select a subset of $K$", + "bbox": [ + 212, + 780, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Scene-aware Human Motion Forecasting via Mutual Distance Prediction", + "bbox": [ + 250, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "vertices denoted as $\\mathbf{v}_t \\in \\mathbb{R}^{K \\times 3}$ from the human mesh. Given a 3D scene and $\\mathbf{v}_t$ at time $t$ , the per-vertex signed distance can be computed as,", + "bbox": [ + 212, + 145, + 782, + 176 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nd _ {k t} = \\left\\{ \\begin{array}{l l} - \\min _ {\\mathbf {y} \\in \\partial S} \\| \\mathbf {v} _ {t k} - \\mathbf {y} \\| _ {2} & \\text {i f} \\mathbf {v} _ {t k} \\in \\mathcal {S} \\\\ \\min _ {\\mathbf {y} \\in \\partial S} \\| \\mathbf {v} _ {t k} - \\mathbf {y} \\| _ {2} & \\text {i f} \\mathbf {v} _ {t k} \\notin \\mathcal {S}, \\end{array} \\right. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 188, + 785, + 237 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{d}_t = [d_{1t}, d_{2t}, \\dots, d_{Kt}]^T \\in \\mathbb{R}^K$ is the per-vertex signed distance at time step $t$ , $\\mathcal{S}$ represents the volume inside the actual solid scene and $\\partial \\mathcal{S}$ is the surface points of such scene. The per-vertex signed distance provides information about the whole-body pose. However, the same human mesh at different global positions may result in the same per-vertex signed distance. To resolve such ambiguity, we further leverage the per-basis point distance which is described below.", + "bbox": [ + 212, + 247, + 782, + 352 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The per-basis point distance. Given a set of basis points from the scene denoted as $\\mathbf{p} \\in \\mathbb{R}^{P \\times 3}$ , the per-basis point distance is simply the minimum distance in each basis point to the human mesh. As the human moves within the scene, its motion is reflected in the continuous change of the per-basis point distance. Specifically, at time step $t$ , it can be computed as,", + "bbox": [ + 212, + 354, + 782, + 430 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nb _ {p t} = \\min _ {\\mathbf {y} \\in \\partial \\mathcal {H} _ {t}} \\| \\mathbf {p} _ {p} - \\mathbf {y} \\| _ {2}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 418, + 441, + 785, + 465 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\partial \\mathcal{H}_t$ is the human surface at time step $t$ and $\\mathbf{b}_t = [b_{1t}, b_{2t}, \\dots, b_{Pt}]^T \\in \\mathbb{R}^P$ . Note that our basis points are not necessary on the scene surface. In practice, for different motion sequences, our basis points can be fixed relatively to the root joint of the last observed pose. Such root joint is defined as the origin of that data sample. In particular, we employ the spherical Fibonacci lattice algorithm [7] to sample $P$ near-equidistant points from a sphere with radius $R$ . The per-basis point distance primarily captures the global location of the human, while the per-vertex signed distance constrains both the detailed whole-body pose and global location. By combining two distances, our mutual distance representation provides explicit constraints on both the whole-body local poses and the global human motion of the human.", + "bbox": [ + 212, + 474, + 782, + 641 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given the 3D scene $\\mathbf{S}$ , the past motion $\\mathbf{X}$ , we first pre-compute the past mutual distance denoted as $\\mathbf{D} = [\\mathbf{d}_1, \\mathbf{d}_2, \\dots, \\mathbf{d}_T] \\in \\mathbb{R}^{K \\times T}$ and $\\mathbf{B} = [\\mathbf{b}_1, \\mathbf{b}_2, \\dots, \\mathbf{b}_T] \\in \\mathbb{R}^{P \\times T}$ . Our goal now is to predict the future mutual distance given the above information. In the following section, we will introduce our mutual distance prediction module.", + "bbox": [ + 212, + 642, + 795, + 717 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Mutual Distance Prediction", + "text_level": 1, + "bbox": [ + 214, + 739, + 491, + 753 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We adopt a similar temporal and spatial encoding strategy proposed in [22], which has been proven to be effective in predicting smooth spatial-temporal data such as 3D human motion. Specifically, we use Discrete Cosine Transform (DCT) to represent a temporal sequence and Graph Convolutional Networks (GCN) to capture the spatial dependencies between different sequences.", + "bbox": [ + 212, + 763, + 782, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "C. Xing et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "With the DCT, we represent a sequence as a linear combination of pre-defined cosine bases. In particular, let us denote a $T$ -frame sequence of signed distance of the $k$ -th human vertex as $\\tilde{\\mathbf{d}}_k \\in [d_{1k}, d_{2k}, \\dots, d_{Tk}]^T \\in \\mathbb{R}^T$ , the $l$ -th DCT coefficient is then,", + "bbox": [ + 212, + 146, + 787, + 205 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nh _ {k l} = \\sqrt {\\frac {2}{T}} \\sum_ {t = 1} ^ {T} d _ {t k} \\frac {1}{\\sqrt {1 + \\delta_ {l 1}}} \\cos \\left(\\frac {\\pi}{2 T} (2 t - 1) (l - 1)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 308, + 217, + 785, + 258 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $l \\in \\{1, 2, \\dots, T\\}$ and $\\delta_{ij}$ denotes the Kronecker delta function, i.e.,", + "bbox": [ + 214, + 268, + 735, + 285 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\delta_ {i j} = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} i = j \\\\ 0 & \\text {i f} i \\neq j. \\end{array} \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 295, + 785, + 338 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Note that, with all cosine bases precomputed, the DCT can be simply represented as a linear operation,", + "bbox": [ + 214, + 347, + 782, + 375 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {h} _ {k} = \\mathbf {C} \\tilde {\\mathbf {d}} _ {k}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 455, + 376, + 785, + 393 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathbf{C} \\in \\mathbb{R}^{T \\times T}$ is an orthogonal matrix and each row of $\\mathbf{C}$ stores the precomputed the cosine basis. $\\mathbf{h}_k \\in \\mathbb{R}^T$ are the $T$ DCT coefficients. Given such coefficients $\\mathbf{h}_k$ , the original sequence can be losslessly recovered via Inverse DCT (IDCT) as,", + "bbox": [ + 212, + 398, + 782, + 458 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {d}} _ {k} = \\mathbf {C} ^ {T} \\mathbf {h} _ {k}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 450, + 458, + 785, + 476 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Similarly, we can obtain the DCT coefficients of all $K$ human vertices as $\\mathbf{H} = [\\mathbf{h}_1,\\mathbf{h}_2,\\dots ,\\mathbf{h}_K]^T\\in \\mathbb{R}^{K\\times T}$ . For the spatial encoding, we follow [22] to use GCN. Specifically, a GCN layer is defined as,", + "bbox": [ + 214, + 482, + 787, + 527 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} ^ {(n + 1)} = \\sigma \\left(\\mathbf {A} ^ {(n)} \\mathbf {F} ^ {(n)} \\mathbf {W} ^ {(n)}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 537, + 785, + 556 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\sigma$ is an activation function such as $\\tanh(\\cdot)$ , $\\mathbf{A}^{(n)} \\in \\mathbb{R}^{K \\times K}$ is the learnable adjacent matrix at the $n$ -th layer and $\\mathbf{W}^{(n)} \\in \\mathbb{R}^{T \\times F}$ is the feature extraction parameter. $\\mathbf{F}^{(n)} \\in \\mathbb{R}^{K \\times F'}$ and $\\mathbf{F}^{(n+1)} \\in \\mathbb{R}^{K \\times F}$ are the input and the output features of the $n$ -th GCN layer, respectively2. Same as [22], our prediction network consists of many such layers. Since our final goal is to predict the future per-vertex signed distance, we follow the strategy in [22] to first pad the sequence to make a sequence of length $T + U$ . Our prediction module then aims to predict a residual between the ground truth DCT coefficients of the whole sequence and those of the padded historical sequence given the scene $\\mathbf{S}$ and past human motion $\\mathbf{X}$ . More formally, the future per-vertex signed distance is predicted as follows,", + "bbox": [ + 212, + 565, + 787, + 731 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {D}} = \\left(\\mathcal {G} _ {d} (\\hat {\\mathbf {H}}, \\mathcal {G} _ {s} (\\mathbf {S}), \\mathcal {G} _ {x} (\\mathbf {X})) + \\hat {\\mathbf {H}}\\right) \\mathbf {C}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 732, + 785, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathcal{G}_d$ , $\\mathcal{G}_s$ and $\\mathcal{G}_x$ are the GCN predictor, the 3D scene encoding network, and the past motion encoding model, respectively. With a little abuse of notation, here, $\\mathbf{C} \\in \\mathbb{R}^{(T + U) \\times (T + U)}$ is the DCT bases for sequences with length $T + U$ . $\\hat{\\mathbf{H}} \\in \\mathbb{R}^{K \\times (T + U)}$ is the DCT coefficients of the padded historical sequence. $\\hat{\\mathbf{D}} \\in$", + "bbox": [ + 212, + 755, + 787, + 816 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Scene-aware Human Motion Forecasting via Mutual Distance Prediction", + "bbox": [ + 248, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "2 The first GCN layer will take the DCT coefficient matrix as input thus, $\\mathbf{F}^{(1)} = \\mathbf{H}$", + "bbox": [ + 217, + 824, + 781, + 840 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$\\mathbb{R}^{K \\times (T + U)}$ is the predicted per-vertex signed distance. Note that, here our model also tries to recover the past sequence.", + "bbox": [ + 215, + 145, + 782, + 176 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Similarly, we have another branch to predict the future per-basis point distance as,", + "bbox": [ + 215, + 176, + 782, + 205 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {B}} = \\left(\\mathcal {G} _ {b} (\\hat {\\mathbf {K}}, \\mathcal {G} _ {s} (\\mathbf {S}), \\mathcal {G} _ {x} (\\mathbf {X})) + \\hat {\\mathbf {K}}\\right) \\mathbf {C}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 205, + 784, + 222 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathcal{G}_b$ is also a GCN. $\\hat{\\mathbf{K}}$ is the DCT coefficients of padded past per-basis point distance sequence. $\\hat{\\mathbf{B}}\\in \\mathbb{R}^{P\\times (T + U)}$ is the predicted per-basis point distance.", + "bbox": [ + 215, + 229, + 782, + 260 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "During training, we use the average $\\ell_1$ loss between the ground-truth distances and the predicted ones. Formally, this loss is defined as,", + "bbox": [ + 215, + 261, + 782, + 290 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {\\text {d i s t}} = \\frac {1}{(T + U)} \\left(\\frac {1}{K} \\sum_ {t = 1} ^ {T + U} \\sum_ {k = 1} ^ {K} | \\hat {d} _ {k t} - d _ {k t} | + \\frac {1}{P} \\sum_ {t = 1} ^ {T + U} \\sum_ {p = 1} ^ {P} | \\hat {b} _ {p t} - b _ {p t} |\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 259, + 301, + 782, + 342 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\hat{d}_{kt}$ is predicted signed distance for $k$ -th vertex at time $t$ and $\\hat{b}_{pt}$ is predicted distance for $p$ -th basis point at time $t$ , $d_{kt}$ and $b_{pt}$ define the ground truth.", + "bbox": [ + 215, + 356, + 782, + 388 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 Motion Forecasting Network", + "text_level": 1, + "bbox": [ + 215, + 410, + 496, + 424 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Given the future mutual distance $\\hat{\\mathbf{D}}$ and $\\hat{\\mathbf{B}}$ , the past human motion $\\mathbf{X}$ , and the scene $\\mathbf{S}$ , our goal now is to predict the future human poses $\\hat{\\mathbf{Y}}$ . To this end, we use an RNN-based motion prediction model $(\\mathcal{G}_y)$ which predicts the future motion in an auto-regressive manner. More formally, it is defined as,", + "bbox": [ + 215, + 434, + 782, + 494 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {x}} _ {T + u} = \\mathcal {G} _ {y} \\left(\\hat {\\mathbf {x}} _ {T + u - 1}, \\hat {\\mathbf {d}} _ {T + u}, \\hat {\\mathbf {b}} _ {T + u}, \\mathcal {G} _ {x} (\\mathbf {X}), \\mathcal {G} _ {s} (\\mathbf {S})\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 330, + 506, + 782, + 522 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $u\\in \\{1,2,\\dots ,U\\}$ $\\hat{\\mathbf{x}}_{T + u}\\in \\mathbb{R}^{M}$ is the predicted $u$ -th future human pose.", + "bbox": [ + 215, + 535, + 774, + 550 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "During training, in addition to the human pose loss, we incorporate two extra losses to ensure the predicted future pose to be consistent with the ground truth mutual distance. For human pose loss, we employ global translation and orientation loss and the local pose loss. They are defined as follows,", + "bbox": [ + 215, + 551, + 782, + 609 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {\\text {g l o b a l}} = \\frac {1}{U} \\sum_ {u = 1} ^ {U} \\left| \\hat {\\mathbf {t}} _ {T + u} - \\mathbf {t} _ {T + u} \\right| + \\left| \\hat {\\mathbf {r}} _ {T + u} - \\mathbf {r} _ {T + u} \\right| \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 336, + 622, + 782, + 660 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {\\text {l o c a l}} = \\frac {1}{U} \\sum_ {u = 1} ^ {U} \\left| \\hat {\\mathbf {x}} _ {T + u} ^ {\\text {l o c a l}} - \\mathbf {x} _ {T + u} ^ {\\text {l o c a l}} \\right|, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 666, + 782, + 705 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{t}}_{T + u} \\in \\mathbb{R}^3$ and $\\mathbf{t}_{T + u} \\in \\mathbb{R}^3$ are the predicted and ground truth global translation, respectively. $\\hat{\\mathbf{r}}_{T + u} \\in \\mathbb{SO}(3)$ and $\\mathbf{r}_{T + u} \\in \\mathbb{SO}(3)$ are the predicted and ground truth global orientation, respectively. Here, we adopt the 6-D rotation representation [41]. $\\hat{\\mathbf{x}}_{T + u}^{\\mathrm{local}} \\in \\mathbb{R}^{M - 9}$ and $\\mathbf{x}_{T + u}^{\\mathrm{local}} \\in \\mathbb{R}^{M - 9}$ are the predicted and ground truth local human pose. To compute the two consistency losses, we first use the pre-defined mapping function such as SMPL-X [25] to obtain the sampled human surface vertices $\\hat{\\mathbf{v}}_{T + u} \\in \\mathbb{R}^{K \\times 3}$ and the set of all human surface points $\\partial \\hat{\\mathcal{H}}_{T + u}$ . Recall that we represent the 3D scene as a signed distance", + "bbox": [ + 215, + 718, + 784, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "C. Xing et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "volume. We can query the signed distance of the human surface vertices by interpolating within this volume. The per-vertex signed distance consistency loss is then defined as,", + "bbox": [ + 214, + 146, + 782, + 191 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {\\text {v e r t e x}} = \\frac {1}{U K} \\sum_ {u = 1} ^ {U} \\sum_ {k = 1} ^ {K} | \\tilde {d} _ {k (T + u)} - d _ {k (T + u)} |, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 204, + 785, + 244 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\tilde{d}_{k(T + u)}$ and $d_{k(T + u)}$ are the signed distance of the $k$ -th vertex from the $u$ -th predicted human mesh and ground truth human mesh, respectively.", + "bbox": [ + 214, + 257, + 782, + 287 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Similarly, given the set of human surface points $\\partial \\tilde{\\mathcal{H}}_{T + u}$ at time $T + u$ , we can compute the per-basis point distance $\\tilde{\\mathbf{b}}_{T + u} \\in \\mathbb{R}^P$ using Equation 2. The per-basis point distance consistency loss can then be computed as", + "bbox": [ + 214, + 289, + 784, + 335 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {\\text {b a s i s}} = \\frac {1}{U P} \\sum_ {u = 1} ^ {U} \\sum_ {p = 1} ^ {P} | \\tilde {b} _ {p (T + u)} - b _ {p (T + u)} |. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 345, + 785, + 388 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The final training loss for the motion forecasting network is the combination of all the above losses,", + "bbox": [ + 214, + 398, + 782, + 429 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {\\text {m o t i o n}} = \\lambda_ {1} \\ell_ {\\text {g l o b a l}} + \\lambda_ {2} \\ell_ {\\text {l o c a l}} + \\lambda_ {3} \\ell_ {\\text {v e r t e x}} + \\lambda_ {4} \\ell_ {\\text {b a s i s}}, \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 320, + 443, + 784, + 458 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\lambda_{i}$ with $i\\in \\{1,2,3,4\\}$ are the balancing weights.", + "bbox": [ + 214, + 469, + 609, + 484 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Here we adopt a stage-wise training strategy where we initially train the mutual distance prediction network and motion forecasting network separately. While our network is differentiable and can be trained from scratch in an end-to-end fashion, we observed that at early stage of the training process, the learning process can easily converge to a local minimum due to inaccurate prediction of mutual distance which serves as the input for the motion forecasting network. We thus propose to perform stage-wise training first. Note that, for stage-wise training process, the motion forecasting network is provided with the ground truth mutual distance as input. Subsequently, we combine both networks and fine-tune them in an end-to-end manner with both losses defined in Eq. 10 and 16.", + "bbox": [ + 214, + 486, + 785, + 636 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 660, + 375, + 678 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Datasets", + "text_level": 1, + "bbox": [ + 215, + 693, + 333, + 707 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We evaluate our method on four datasets, HUMANISE [36], GTA-IM [4], PROX [10] and GIMO [40].", + "bbox": [ + 214, + 718, + 803, + 748 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "GTA-IM. The GTA Indoor Motion dataset [4], building upon GTA game engine, is a synthetic human-scene interaction dataset. It comprises approximately 1 million RGB-D frames featuring 3D human poses for 50 distinct characters across 7 diverse scenes. To reconstruct the mesh, we fuse the depth maps of each scene and subsequently generate the SDF volume. We follow ContactAware [20] to use the same training-test split and 21 joints. The videos are 30 FPS. Both our", + "bbox": [ + 214, + 750, + 785, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Scene-aware Human Motion Forecasting via Mutual Distance Prediction", + "bbox": [ + 250, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "model and baseline models are trained to observe the past 30 frames to predict the future 60 frames following [20].", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "PROX. Proximal Relationships with Object eXclusion (PROX) [10] is a real-world dataset. This dataset contains human SMPL-X parameters [25] fitted from RGB-D frames and 20 distinct characters interacting with various scenes. Since the original motion sequences are jittery, we use the refined version provided by [20]. We adopt the same data split and the 22 joints as in [20]. The frame rate is $30\\mathrm{Hz}$ . Our model and baseline models input the past 30 frames to forecast the future 60 frames following [20].", + "bbox": [ + 212, + 176, + 784, + 281 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "HUMANISE. The HUMANISE dataset [36] is a synthetic human-scene interaction dataset. It is mainly built upon two primary datasets: the AMASS dataset [19] and the ScanNet dataset [6]. HUMANISE combines these datasets by incorporating different global motions to meet contact and collision constraints. Since the original split includes duplicated human motions in the training and test set which is not suitable for motion prediction, we create a new split to ensure that motions in the test set are never observed during training. Moreover, we ensure that the scenes paired with the motions in the test set are categorized into seen scenes, which include a set of scenes observed during the training, and unseen scenes consisting of scenes never observed during training. Our final split comprises approximately 20k sub-sequences for training and 6k for testing, which includes 5k sub-sequences from seen scenes and 1k from unseen scenes. The videos run at 30 frame-per-second (FPS). Both our model and baseline models are trained to observe past 15 frames (0.5 sec.) and predict future 30 frames (1 sec.) because the average length of motion sequences in the dataset is 2.2 sec. which is much shorter than 3 sec.", + "bbox": [ + 212, + 282, + 785, + 522 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "GIMO. The Gaze-Informed Human Motion dataset (GIMO) [40] is a real human-scene interaction dataset. The dataset captures human motion using an IMU-based motion capture system, and the 3D scene is scanned using a smartphone equipped with lidar sensors. Same as [36], GIMO also adopts SMPL-X [25] representation. It captures 11 subjects performing various actions e.g., open window, lie in bed, in 19 scenes, resulting in 127 long motion sequences with $192\\mathrm{k}$ frames. We follow the official training and test split, resulting in around 6k training sequences and 1k testing sequences. Same as HUMANISE [36], frame rate is $30\\mathrm{Hz}$ . Our model and baseline models predict future 30 frames with 15 past frames.", + "bbox": [ + 212, + 523, + 787, + 659 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Metrics,Baselines & Implementation", + "text_level": 1, + "bbox": [ + 214, + 680, + 568, + 695 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Metrics. We follow [20] to evaluate our method using path error and pose error. The path error is defined as the $l_{2}$ distance between the predicted and the ground truth translation; the pose error is the Mean Per Joint Position Error (MPJPE) [12] of local poses.", + "bbox": [ + 212, + 703, + 784, + 763 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Baselines. We compare our method with three existing scene-aware human motion prediction models: ContAware [20], GIMO [40] and STAG [27]. ContAware [20] and STAG [27] are state-of-the-art methods which first predict the contact map and then the future motion conditioning on the predicted contact map. We train their model using the official implementation on our datasets,", + "bbox": [ + 212, + 763, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "C. Xing et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "following their original settings to predict only human joints instead of SMPL-X poses. GIMO [40] is another state-of-the-art scene-aware human motion prediction method which mainly focuses on predicting long-term future poses given the past gaze information, and the frame rate of their motion is $2\\mathrm{Hz}$ . The original model of GIMO predicts future 10 frames (5 seconds) given the past 6 frames (3 seconds). We slightly adapt their official code to our settings for evaluation.", + "bbox": [ + 212, + 145, + 787, + 236 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Implementation details. Our models are developed using Pytorch [24] and optimized with the ADAM [13] optimizer. Initially, we train the mutual distance prediction model and the motion forecasting model separately for 40 epochs each. Subsequently, we fine-tune both models together for an additional epoch. The original learning rate is set to 0.0005 and decays by half at the 30-th epoch for both models. We train our model on a single RTX4090. For all datasets, the loss weights of the motion forecasting model $(\\lambda_1,\\lambda_2,\\lambda_3,\\lambda_4)$ are 1.0, 0.5, 1.0, 1.0. For each motion sequence, we crop the 3D scene to a region that is within 2.0 meters of the root joint of the last observed pose, and the root joint is used as the origin of the cropped scene. To compute the per-vertex signed distance, we follow [38] to use 67 vertices on the human body. For the basis point set, we select 150 points that are 2.0 meters away from the origin of the scene, and those points are fixed for different sequences.", + "bbox": [ + 212, + 237, + 787, + 434 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/4953412e93295536f0ff92cff47ecf1fa4fd69723fe0e185c556564ae94bc350.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
datasetmethodPath Error (mm)Pose Error (mm)
0.5s1.0s1.5s2.0smean0.5s1.0s1.5s2.0smean
GTA-IMContAware [20]†44.582.6125.6182.987.140.154.165.277.251.8
GIMO [40]*52.797.8160.6241.7110.347.960.771.182.759.9
STAG [27]†43.279.8119.9176.483.435.448.759.873.547.0
Ours34.465.9104.0155.672.031.046.858.970.744.6
ContAware w/GT34.546.049.767.641.736.344.348.753.941.3
STAG w/GT32.743.349.468.140.031.037.241.548.335.1
Ours w/GT15.818.920.523.022.420.423.825.828.217.6
PROXContAware [20]93.3187.2284.4381.2192.289.9127.5149.3167.5116.8
GIMO [40]*129.5233.2342.1455.0240.4112.6144.6168.0187.5140.7
STAG [27]†95.9186.4276.6369.8189.698.4137.8163.7185.7129.4
Ours80.9155.2238.6322.1162.683.9119.3142.0159.7110.7
ContAware w/GT73.9106.7104.6117.488.083.7112.9125.2132.9101.1
STAG w/GT64.081.693.4115.477.288.6106.0112.6125.498.3
Ours w/GT44.256.763.769.952.656.469.476.883.464.7
", + "bbox": [ + 218, + 457, + 784, + 696 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 1: Quantitative results on GTA-IM [4] and PROX [10]. Our model outperforms all baselines at all metrics, especially for the path error.", + "bbox": [ + 215, + 696, + 787, + 726 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Results", + "text_level": 1, + "bbox": [ + 215, + 786, + 323, + 800 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Quantitative results. We provide quantitative comparisons across three datasets: GTA-IM [4], PROX [10] and HUMANISE [36] presented in Tables 1 and 2. No", + "bbox": [ + 212, + 809, + 802, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Scene-aware Human Motion Forecasting via Mutual Distance Prediction", + "bbox": [ + 250, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/78d246db9c0b2665649f985cdc0da7504686b421cd38e126bcf932861a03a3b6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodSeen ScenesUnseen Scenes
PathError (mm)PoseError (mm)PathError (mm)PoseError (mm)
0.5s1.0smean0.5s1.0smean0.5s1.0smean0.5s1.0smean
ContAware [20]†52.8121.157.897.6141.492.953.1124.058.794.0139.190.3
GIMO [40]*70.1129.272.0141.4150.3140.277.7144.080.2146.3159.4146.5
STAG [27]†55.2124.660.788.3131.483.057.0131.563.289.9137.785.6
Ours41.593.545.683.7130.980.046.7100.250.184.3131.880.6
ContAware w/GT38.060.835.793.4119.186.438.261.836.490.3122.084.2
STAG w/GT37.563.236.483.5117.677.337.762.236.685.1122.779.4
Ours w/GT22.933.422.080.3122.076.324.835.523.680.2121.576.2
", + "bbox": [ + 220, + 143, + 787, + 297 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/3b0ee6bcc1627804c8560d40ae144e7590e798918c97d3ea05ab8a716ded0bc1.jpg", + "table_caption": [ + "Table 2: Quantitative results on HUMANISE [36]. Our model outperforms all baselines on both seen scenes and unseen scenes." + ], + "table_footnote": [], + "table_body": "
MethodSeen ScenesUnseen Scenes
PathError (mm)PoseError (mm)PathError (mm)PoseError (mm)
0.5s1.0smean0.5s1.0smean0.5s1.0smean0.5s1.0smean
Motion only53.4111.957.0102.0140.093.559.2119.062.0103.2145.395.5
Scene only48.4105.852.694.6132.987.654.3113.857.994.1136.988.3
Ours w/o D42.495.246.288.2134.683.447.9103.151.290.4140.186.0
Ours w/o B47.6101.350.685.7133.981.752.8109.155.886.4138.383.2
Ours w/ P45.297.249.390.8137.286.049.0101.752.789.8138.986.1
Ours w/ O50.1115.355.392.2141.286.755.0115.258.491.0141.386.2
Ours w/ BPS50.2108.953.9100.0153.694.552.0104.154.394.4148.290.3
Ours e2e47.6105.251.590.4141.086.252.1112.456.189.5142.085.7
Ours41.593.545.683.7130.980.046.7100.250.184.3131.880.6
", + "bbox": [ + 220, + 358, + 787, + 539 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 3: Ablation study on HUMANISE [36]. We show the ablation results of our model. D, B, P, O, BPS and $e2e$ indicate per-vertex signed distance, per-basis point distance, point cloud representation, occupancy representation, BPS encoding and end-to-end training respectively.", + "bbox": [ + 215, + 541, + 787, + 597 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "tably, these datasets lack gaze information, with the exception of GIMO [40], for which we report results without gaze, indicated by $*$ . Additionally, we have adapted our training strategy to baseline models for a fair comparison. We report the best performance for each model using our training strategy (denoted by $\\dagger$ ). Our method consistently outperforms the baseline approaches on all datasets and at all time steps by a substantial margin, both for path and pose predictions. Particularly noteworthy is our method's superiority in path prediction over baseline models on the real-world PROX [10] dataset. Our results demonstrate that our model with ground truth mutual distance significantly outperforms baseline models with ground truth contact (w/GT). This superior performance is evident in both global path and local pose metrics, indicating that our method provides a more effective constraint for human motion. More results on the GIMO [40] dataset can be found in the supplementary material.", + "bbox": [ + 212, + 642, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "C. Xing et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/57085e4aba0712c3ad0c59f9ac22f593ec4dd85ec138cea4c3ac27e9d48f4e31.jpg", + "image_caption": [ + "Fig. 3: This figure compares our method with baseline models on GTA-IM [4] (top row), PROX [10] (middle row), and HUMANISE [36] (bottom row). Our method predicts future motion closer to the ground truth." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 808, + 386 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/eaed32a321b1b18005b390bb959351aec0903028d49200fcc905b9dedc29fbcb.jpg", + "image_caption": [ + "Fig. 4: Ablation of the mutual distance. The three figures in the first three columns (from left to right) depict the ground truth pose for the last frame of future motion, results of our full model, and predictions of our model without mutual distance constraint. The sub-figure in gray is the last observed frame. Other sub-figures depict the predicted middle frame. The blue dot is the scene basis point, and the red dot is the sampled vertex on human mesh. The two graphs in the last two columns show the predicted per-vertex signed distance and the per-basis point distance for the red and blue point, respectively. As shown in the figures, with predicted mutual distance, we can forecast the 'stand-up' action more accurately." + ], + "image_footnote": [], + "bbox": [ + 215, + 468, + 787, + 554 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Qualitative results. We present qualitative comparisons with baseline methods on GTA-IM [4], HUMANISE [36], and PROX [10], in Fig. 3. The comparisons further demonstrate that our approach excels in predicting future human motions, aligning closely with ground truth data. In contrast, GIMO [40] exhibits issues such as motion freezing and scene penetration. Additionally, we show the predicted human skeletons generated by ContAware [20] and STAG [27], which struggle to handle complex motions.", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Scene-aware Human Motion Forecasting via Mutual Distance Prediction", + "bbox": [ + 250, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Ablation study. To gain a more comprehensive understanding, we present ablation results on HUMANISE [36] dataset in Table 3. We report results for our base models which are defined as our model without scene and mutual distance ('Motion only') and our model without mutual distance ('Scene only'). Notably, there are significant improvements in both path and pose prediction when we incorporate mutual distance into our model. Additionally, we assess the impact of two key components: the proposed per-vertex signed distance ('Ours w/o $\\mathbf{B}^{\\prime}$ ) and the per-basis point distance ('Ours w/o $\\mathbf{D}^{\\prime}$ ). Each distance individually contributes to the improvement for both path and pose prediction, with the most significant gains achieved when both distances are utilized. Furthermore, we examine the influence of the global scene representation: point cloud representation ('Ours w/ $P^{\\prime}$ ) and occupancy representation ('Ours w/ $O^{\\prime}$ ). For point cloud representation, similar to ContAware [20], we modified the official PVCNN [18] code to take point cloud as scene input instead of SDF volume. For occupancy representation, our method uses occupancy volume as scene input. Our model outperforms both the point cloud and occupancy representation across all metrics. We also compare the Basis Point Set (BPS) encoding in PLACE [37] with our per-basis point distance ('Ours w/ $BPS^{\\prime}$ ) by replacing the per-basis point distance with BPS encoding following the official implementation in [37]. Moreover, we show results from end-to-end training of our method ('Ours $e2e^{\\prime}$ ), which encounters a local minimum in the early epochs. Results shown in Fig. 4 further demonstrate the effectiveness of our method.", + "bbox": [ + 212, + 146, + 787, + 479 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 503, + 359, + 520 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we present an approach for scene-aware human motion forecasting that explicitly models the human-scene interaction. To constrain the whole-body motion, we have proposed a mutual distance representation: the per-vertex signed distance and the per-basis point distance. The mutual distance captures both the detailed local body movements and the global human motion. Thanks to the explicit mutual distance constraints, Our method generates future human motions that are more plausible and accurate than those produced by state-of-the-art scene-aware motion prediction techniques.", + "bbox": [ + 212, + 537, + 787, + 657 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Societal Impacts & Limitations. The motion prediction results depend on the prediction of the mutual distance is one limitation of our work. Our ablation study using the ground truth mutual distance indicates that enhancing the prediction of mutual distance results in improved motion prediction. This can therefore serve as a potential direction for future research. In real applications, a potential risk of our model is that without explicit real world physics modeling, it is possible for our model to predict future motions that have imbalanced forces. This may lead to unstable human motion when deploy to real life applications like human-robot interaction such as collision.", + "bbox": [ + 212, + 657, + 787, + 792 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements This research was supported in part by the Australia Research Council DECRA Fellowship (DE180100628) and ARC Discovery Grant (DP200102274).", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "C. Xing et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Aksan, E., Kaufmann, M., Hilliges, O.: Structured prediction helps 3d human motion modelling. In: ICCV. pp. 7144-7153 (2019) 1, 3", + "2. Brand, M., Hertzmann, A.: Style machines. In: Proceedings of the 27th annual conference on Computer graphics and interactive techniques. pp. 183-192. ACM Press/Addison-Wesley Publishing Co. (2000) 1, 3", + "3. Cai, Y., Huang, L., Wang, Y., Cham, T.J., Cai, J., Yuan, J., Liu, J., Yang, X., Zhu, Y., Shen, X., et al.: Learning progressive joint propagation for human motion prediction. In: ECCV. pp. 226-242. Springer (2020) 1, 3", + "4. Cao, Z., Gao, H., Mangalam, K., Cai, Q.Z., Vo, M., Malik, J.: Long-term human motion prediction with scene context. In: ECCV. pp. 387-404. Springer (2020) 1, 2, 3, 4, 9, 11, 13", + "5. Corona, E., Pumarola, A., Alenya, G., Moreno-Noguer, F.: Context-aware human motion prediction. In: CVPR. pp. 6992-7001 (2020) 1, 2, 3, 4", + "6. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: CVPR. pp. 5828-5839 (2017) 10", + "7. González, Á.: Measurement of areas on a sphere using fibonacci and latitude-longitude lattices. Mathematical Geosciences 42, 49-64 (2010) 6", + "8. Gopalakrishnan, A., Mali, A., Kifer, D., Giles, L., Ororbia, A.G.: A neural temporal model for human motion prediction. In: CVPR. pp. 12116-12125 (2019) 1, 3", + "9. Hassan, M., Ceylan, D., Villegas, R., Saito, J., Yang, J., Zhou, Y., Black, M.J.: Stochastic scene-aware motion prediction. In: ICCV. pp. 11374-11384 (2021) 4", + "0. Hassan, M., Choutas, V., Tzionas, D., Black, M.J.: Resolving 3d human pose ambiguities with 3d scene constraints. In: ICCV. pp. 2282-2292 (2019) 9, 10, 11, 12, 13", + "1. Hassan, M., Ghosh, P., Tesch, J., Tzionas, D., Black, M.J.: Populating 3d scenes by learning human-scene interaction. In: CVPR. pp. 14708-14718 (2021) 1, 4", + "2. Ionescu, C., Papava, D., Olaru, V., Sminchisescu, C.: Human3.6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. TPAMI 36(7), 1325-1339 (2014) 10", + "3. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. In: ICLR (2015) 11", + "4. Kipf, T.N., Welling, M.: Semi-supervised classification with graph convolutional networks. In: ICLR (2017) 3", + "5. Koppula, H.S., Saxena, A.: Anticipating human activities for reactive robotic response. In: IROS. p. 2071. Tokyo (2013) 1", + "6. Li, M., Chen, S., Zhao, Y., Zhang, Y., Wang, Y., Tian, Q.: Dynamic multiscale graph neural networks for 3d skeleton based human motion prediction. In: CVPR. pp. 214-223 (2020) 1, 3", + "7. Li, X., Li, H., Joo, H., Liu, Y., Sheikh, Y.: Structure from recurrent motion: From rigidity to recurrency. In: CVPR (2018) 1, 3", + "8. Liu, Z., Tang, H., Lin, Y., Han, S.: Point-voxel cnn for efficient 3d deep learning. Advances in Neural Information Processing Systems 32 (2019) 14", + "9. Mahmood, N., Ghorbani, N., Troje, N.F., Pons-Moll, G., Black, M.J.: Amass: Archive of motion capture as surface shapes. In: ICCV (2019) 10", + "20. Mao, W., Hartley, R.I., Salzmann, M., et al.: Contact-aware human motion forecasting. Advances in Neural Information Processing Systems 35, 7356-7367 (2022), 1, 2, 3, 4, 5, 9, 10, 11, 12, 13, 14" + ], + "bbox": [ + 225, + 176, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Scene-aware Human Motion Forecasting via Mutual Distance Prediction", + "bbox": [ + 250, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "21. Mao, W., Liu, M., Salzmann, M.: History repeats itself: Human motion prediction via motion attention. In: ECCV. pp. 474-489. Springer (2020) 1, 3", + "22. Mao, W., Liu, M., Salzmann, M., Li, H.: Learning trajectory dependencies for human motion prediction. In: ICCV. pp. 9489-9497 (2019) 1, 3, 6, 7", + "23. Paden, B., Cap, M., Yong, S.Z., Yershov, D., Frazzoli, E.: A survey of motion planning and control techniques for self-driving urban vehicles. IEEE Transactions on intelligent vehicles 1(1), 33-55 (2016) 1", + "24. Paszke, A., Gross, S., Chintala, S., Chanan, G., Yang, E., DeVito, Z., Lin, Z., Desmaison, A., Antiga, L., Lerer, A.: Automatic differentiation in pytorch. In: NeurIPS-W (2017) 11", + "25. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3D hands, face, and body from a single image. In: CVPR. pp. 10975-10985 (2019) 4, 5, 8, 10", + "26. Prokudin, S., Lassner, C., Romero, J.: Efficient learning on point clouds with basis point sets. In: ICCV. pp. 4332-4341 (2019) 4", + "27. Scofano, L., Sampieri, A., Schiele, E., De Matteis, E., Leal-Taixe, L., Galasso, F.: Staged contact-aware global human motion forecasting. In: BMVC (2023) 2, 3, 4, 10, 11, 12, 13", + "28. Sidenbladh, H., Black, M.J., Sigal, L.: Implicit probabilistic models of human motion for synthesis and tracking. In: ECCV. pp. 784-800. Springer (2002) 1, 3", + "29. Starke, S., Zhang, H., Komura, T., Saito, J.: Neural state machine for character-scene interactions. ACM Trans. Graph. 38(6), 209-1 (2019) 1", + "30. Taylor, G.W., Hinton, G.E., Roweis, S.T.: Modeling human motion using binary latent variables. In: NeurIPS. pp. 1345-1352 (2006) 1, 3", + "31. Van Welbergen, H., Van Basten, B.J., Egges, A., Ruttkay, Z.M., Overmars, M.H.: Real time animation of virtual humans: a trade-off between naturalness and control. In: Computer Graphics Forum. vol. 29, pp. 2530-2554. Wiley Online Library (2010) 1", + "32. Wang, B., Adeli, E., Chiu, H.k., Huang, D.A., Niebles, J.C.: Imitation learning for human pose prediction. In: ICCV. pp. 7124-7133 (2019) 1, 3", + "33. Wang, J.M., Fleet, D.J., Hertzmann, A.: Gaussian process dynamical models for human motion. TPAMI 30(2), 283-298 (2008) 1, 3", + "34. Wang, J., Xu, H., Xu, J., Liu, S., Wang, X.: Synthesizing long-term 3d human motion and interaction in 3d scenes. In: CVPR. pp. 9401-9411 (2021) 1, 2, 4", + "35. Wang, J., Yan, S., Dai, B., Lin, D.: Scene-aware generative network for human motion synthesis. In: CVPR. pp. 12206-12215 (2021) 4", + "36. Wang, Z., Chen, Y., Liu, T., Zhu, Y., Liang, W., Huang, S.: Humanise: Language-conditioned human motion generation in 3d scenes. Advances in Neural Information Processing Systems 35, 14959-14971 (2022) 9, 10, 11, 12, 13, 14", + "37. Zhang, S., Zhang, Y., Ma, Q., Black, M.J., Tang, S.: Place: Proximity learning of articulation and contact in 3d environments. In: 3DV. pp. 642-651. IEEE (2020) 1, 2, 3, 4, 14", + "38. Zhang, Y., Black, M.J., Tang, S.: We are more than our joints: Predicting how 3d bodies move. In: CVPR. pp. 3372-3382 (2021) 5, 11", + "39. Zhang, Y., Hassan, M., Neumann, H., Black, M.J., Tang, S.: Generating 3d people in scenes without people. In: CVPR. pp. 6194-6204 (2020) 4", + "40. Zheng, Y., Yang, Y., Mo, K., Li, J., Yu, T., Liu, Y., Liu, C.K., Guibas, L.J.: Gimo: Gaze-informed human motion prediction in context. In: ECCV. pp. 676-694. Springer (2022) 1, 2, 3, 9, 10, 11, 12, 13", + "41. Zhou, Y., Barnes, C., Lu, J., Yang, J., Li, H.: On the continuity of rotation representations in neural networks. In: CVPR. pp. 5745-5753 (2019) 8" + ], + "bbox": [ + 215, + 147, + 784, + 832 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "C. Xing et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 15 + } +] \ No newline at end of file diff --git a/2024/Scene-aware Human Motion Forecasting via Mutual Distance Prediction/abecdb1c-2c3f-4eb7-820c-6ae2bc3ee016_model.json b/2024/Scene-aware Human Motion Forecasting via Mutual Distance Prediction/abecdb1c-2c3f-4eb7-820c-6ae2bc3ee016_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2701bd78e673208e180afacfec0a699a3fa5037f --- /dev/null +++ b/2024/Scene-aware Human Motion Forecasting via Mutual Distance Prediction/abecdb1c-2c3f-4eb7-820c-6ae2bc3ee016_model.json @@ -0,0 +1,2080 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.251, + 0.142, + 0.756, + 0.187 + ], + "angle": 0, + "content": "Scene-aware Human Motion Forecasting via Mutual Distance Prediction" + }, + { + "type": "text", + "bbox": [ + 0.33, + 0.212, + 0.672, + 0.229 + ], + "angle": 0, + "content": "Chaoyue Xing\\(^{1}\\), Wei Mao\\(^{1}\\), and Miaomiao Liu\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.325, + 0.24, + 0.678, + 0.255 + ], + "angle": 0, + "content": "Australian National University, Canberra, Australia" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.289, + 0.744, + 0.608 + ], + "angle": 0, + "content": "Abstract. In this paper, we address the issue of scene-aware 3D human motion forecasting. A key challenge in this task is to predict future human motions that are coherent with the scene by modeling human-scene interactions. While recent works have demonstrated that explicit constraints on human-scene interactions can prevent the occurrence of ghost motion, they only provide constraints on partial human motion e.g., the global motion of the human or a few joints contacting the scene, leaving the rest of unconstrained. To address this limitation, we propose to represent the human-scene interaction using the mutual distance between the human body and the scene. Such mutual distances constrain both the local and global human motion, resulting in a whole-body motion constrained prediction. In particular, mutual distance constraints consist of two components, the signed distance of each vertex on the human mesh to the scene surface and the distance of basis scene points to the human mesh. We further introduce a global scene representation learned from a signed distance function (SDF) volume to ensure coherence between the global scene representation and the explicit constraint from the mutual distance. We develop a pipeline with two sequential steps: predicting the future mutual distances first, followed by forecasting future human motion. We explicitly ensure consistency between predicted poses and mutual distances during training. Extensive testing on both synthetic and real datasets demonstrates that our method consistently surpasses the performance of current state-of-the-art techniques." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.621, + 0.741, + 0.649 + ], + "angle": 0, + "content": "Keywords: Scene-aware \\(\\cdot\\) Human Motion Forecasting \\(\\cdot\\) Mutual Distance \\(\\cdot\\) SDF Volume" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.674, + 0.377, + 0.691 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.795 + ], + "angle": 0, + "content": "Human motion forecasting has numerous applications, including autonomous driving [23], human-robot interaction [15], animation [31] and virtual/augmented reality [29]. Despite being an active research field for decades [2,28,30,33], most previous methods [1, 3, 8, 16, 17, 21, 22, 32] focus solely on human motion and overlook the fact that humans interact with the 3D environment. In contrast, this paper addresses the problem of scene-aware human motion prediction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Recent advancements in scene-aware human synthesis [11,34,37] and human motion prediction [4,5,20,40] have begun to model the human-scene interaction. Some approaches achieve this implicitly by incorporating scene encoding as an" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "C. Xing et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.329 + ], + "angle": 0, + "content": "additional input to their model e.g., 2D scene image [4], 3D scene point cloud [34] or a specific object [5]. However, without explicit constraints, those methods tend to predict motions with artifacts e.g., \"ghost motion\". Alternatively, other works [20, 27, 37, 40] start to explore the explicit representation of human-scene interaction. However, none of their representations are able to constrain the whole-body motion both locally and globally. For example, the proximity between human and scene used in [37] mainly constrains the global position of the human, while the human gaze employed in [40] only specifies the potential target location of the human. Although the contact map proposed in [20, 27] achieves detailed constraints on both the local pose and global position of certain human joints, their contact map can only constrain joints that contact the scene, neglecting the others, such as elbow joints, which often remain unsupported." + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.357, + 0.484, + 0.479 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.278, + 0.481, + 0.726, + 0.496 + ], + "angle": 0, + "content": "(a) The per-vertex signed distance (b) The per-basis point distance" + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.358, + 0.706, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.507, + 0.788, + 0.564 + ], + "angle": 0, + "content": "Fig. 1: Our mutual distances. a) shows the per-vertex signed distance for sampled vertices on human mesh. Their color indicates the distance value. b) shows the per-basis point distance for the basis points which are not sampled on the scene surface. For both figures, the darker the color is, the smaller the distance is." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.597, + 0.788, + 0.809 + ], + "angle": 0, + "content": "To constrain the whole-body motion for better prediction, in this paper, we propose a mutual distance representation which captures the distances between the human body and the scene. We further introduce the global scene representation learned from SDF volume instead of point cloud to achieve consistent global and local scene representation learning. Specifically, our proposed mutual distance representation comprises two components: the per-vertex signed distance and the per-basis point distance. Given a 3D scene, the per-vertex distance is the minimum distance from each human body vertex to the scene surface, and its sign indicates whether such vertex penetrates the scene or not (see Fig. 1 (a)). Since the change in signed distance for each body vertex represents the motion of the corresponding body part, it provides a complete constraint on the entire body. However, as a signed distance corresponds to a level-set of the scene Signed Distance Field (SDF), a human mesh at multiple possible global locations can result in the same per-vertex signed distance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "To address this, we leverage a per-basis point distance, which calculates the minimum distance between each point in a predefined set of basis points and the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.25, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Scene-aware Human Motion Forecasting via Mutual Distance Prediction" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.221 + ], + "angle": 0, + "content": "human surface. Our per-basis point distance is similar to BPS body encoding in PLACE [37] with differences in choices of basis points. In particular, their basis points are partially scene-dependent making it not suitable for motion prediction, while our basis points are purely scene-independent and sampled with a fixed distance relative to origin, which can better capture the motion patterns." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.785, + 0.388 + ], + "angle": 0, + "content": "While our model is end-to-end trainable, we adopt a two-step training process followed by fine-tuning for better convergence. In the first step, we predict the mutual distance. Conditioning on the predicted distances, we then forecast the future human motion. Specifically, given a 3D scene, the past human poses, and past mutual distance, we use a strategy based on Graph Convolutional Network (GCN) [14], Discrete Cosine Transform (DCT), for the spatial and temporal encoding of the mutual distances and a simple 3D CNN for encoding of the SDF volume to predict the future ones. We subsequently take the predicted distances, the past human poses, and the scene as input to predict future human motion using an RNN-based model. A consistency prior is adopted to ensure that the predicted human motion is consistent with the input mutual distance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.39, + 0.787, + 0.495 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows. i) We propose an explicit representation of human-scene interaction using mutual distance to provide constraints for whole-body motion. ii) We explore different global scene representations and find the SDF volume to be more efficient than commonly used point cloud representation. Our experiments demonstrate that our approach outperforms the state-of-the-art scene-aware human motion prediction methods by a large margin on both existing real and synthesis benchmark datasets." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.519, + 0.388, + 0.535 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.785, + 0.673 + ], + "angle": 0, + "content": "Human motion prediction. Researchers have been studying 3D human motion modeling [2,28,30,33] for decades. While traditional methods, such as those employing Hidden Markov Models [2] or the Gaussian process latent variable model [33], are capable of handling periodic and simple non-periodic motions like walking and golf swings, more complex motions are typically modeled with deep neural networks [1,3,8,16,17,21,22,32]. However, these methods focus solely on the prediction of local poses, neglecting the global human motion and 3D scenes in which the motion takes place." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.787, + 0.84 + ], + "angle": 0, + "content": "More recent works [4,5,20,27,40] have begun to incorporate scene context into the task of motion prediction. Specifically, Cao et al. [4] propose to represent the scene with a 2D image. The feature extracted from such an image is fed into the prediction model and lets the model learn the human-scene interaction implicitly. However, such 2D scene representation cannot handle occlusions and it does not ensure a valid human-scene interaction. In [5], a semantic graph is introduced to encode the interaction between the human and the object of interest. Although it is efficient to model a human interacting with one object, it is not suitable for human-scene interaction as the human may interact with multiple objects when he moves around. In [40], Zheng et al. employ eye gaze as additional information to guide the motion forecasting process. Although" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "C. Xing et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.268 + ], + "angle": 0, + "content": "eye gaze provides valuable hints about the goal location of the human motion, it constrains neither the local poses nor the human-scene interaction. Mao et al. [20] introduce a contact map that can explicitly encode the relation between scene surface points and human joints. Scofano et al. [27] propose a three-stage pipeline in a coarse-to-fine manner and utilizing the time information based on [20]. However, for forecasting, their contact map only constrains joints that contact the scene and disregard the rest. By contrast, our introduced mutual distance is able to constrain the whole-body motion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.268, + 0.788, + 0.646 + ], + "angle": 0, + "content": "Scene-aware human synthesis. One of the main challenges in this task is to ensure a valid human-scene interaction. As such, existing works have explored various strategies to model such interaction [9, 11, 34, 35, 37, 39]. Similar to [4], the interaction is modeled implicitly in [34, 35, 39] via either Generative Adversarial Networks (GANs) [34], or Variational Autoencoders [34, 39]. Besides this, other works [9, 11, 37] adopt explicit representation of human-scene interaction. In particular, it was proposed in POSA [11] to develop a human-centric representation of the human-scene interaction which assigns scene semantic labels, such as floor, and chair, to each human surface vertex. Those labels indicate the probability of such vertex contacting what kind of object in the scene. Although it performs well in the task of placing a human in a scene, it lacks sufficient detail about the local pose of the human, rendering it unsuitable for human motion forecasting. In SAMP [9], Hassan et al. propose to only model the interaction with the object at the goal location. Given the final root translation and orientation, the task is then treated as a path planning problem. Similar to [5], such a strategy cannot constrain the interactions between the human and the scene when the human moves in the scene. The closest to our work is PLACE [37] where the human-scene interaction is represented by the minimum distance between a set of 3D scene points also known as Basis Point Sets (BSP) [26] to the human body surface. However, such representation alone only provides a prior on the global position of the human. For example, one can move its arm that is far from the scene surface while the scene to human proximity remains the same. By contrast, we combine the per-basis point distance with the per-vertex signed distance resulting in a detailed constraint on the whole-body motion both locally and globally." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.667, + 0.348, + 0.684 + ], + "angle": 0, + "content": "3 Approach" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.697, + 0.788, + 0.819 + ], + "angle": 0, + "content": "Let us begin by formulating the task of scene-aware 3D human motion forecasting. Building on previous work [20], we represent a sequence of \\(T\\) past human poses as \\(\\mathbf{X} = [\\mathbf{x}_1, \\mathbf{x}_2, \\dots, \\mathbf{x}_T] \\in \\mathbb{R}^{M \\times T^1}\\), where a human pose at time \\(t\\) is represented as a vector \\(\\mathbf{x}_t \\in \\mathbb{R}^M\\) of length \\(M\\). Such vector could be a parametric representation of human mesh such as SMPL-X [25] pose. To represent the 3D scene, we adopt a Signed Distance Volume \\(\\mathbf{S} \\in \\mathbb{R}^{N \\times N \\times N}\\), each entry of which stores the distance of the voxel to the nearest scene surface and its sign indicates whether it is inside or outside the scene. The representation learned from this" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.731, + 0.84 + ], + "angle": 0, + "content": "1 Unless otherwise specified, in this paper, we use column vector convention." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.25, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Scene-aware Human Motion Forecasting via Mutual Distance Prediction" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.144, + 0.789, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.385, + 0.788, + 0.456 + ], + "angle": 0, + "content": "Fig. 2: Network architecture. Given the 3D scene \\(\\mathbf{S}\\) represented as a signed distance volume and the past motion \\(\\mathbf{X}\\) shown in grey meshes, our approach first predicts the future per-vertex signed distance \\(\\hat{\\mathbf{D}}\\), and the future per-basis point distance \\(\\hat{\\mathbf{B}}\\) from historical distance \\(\\mathbf{D}\\) and \\(\\mathbf{B}\\), respectively. The two predicted future distances are then fed into a RNN-based network to predict the future motion \\(\\hat{\\mathbf{Y}}\\) shown in orange meshes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.486, + 0.788, + 0.561 + ], + "angle": 0, + "content": "volume is more consistent with our mutual distance representation than those learned from point cloud [20]. Moreover, such Signed Distance Volume allows us to query the signed distance of any given position easily via interpolation, which facilitates the fast calculation of the loss defined on the SDFs of the predicted pose sequence during training (see Eq. 14)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.562, + 0.812, + 0.697 + ], + "angle": 0, + "content": "Our objective is predicting \\(U\\) future human poses \\(\\mathbf{Y} = [\\mathbf{x}_{T + 1},\\mathbf{x}_{T + 2},\\dots ,\\mathbf{x}_{T + U}]\\in\\) \\(\\mathbb{R}^{M\\times U}\\) with history motion \\(\\mathbf{X}\\) and 3D scene S. We introduce an innovative pipeline that predicts mutual distance representation first i.e., the signed minimum distance from each human body vertex to the scene surface (the per-vertex signed distance) and the minimum distance from each basis point to the human surface (the per-basis point distance) at each time frame. Conditioned on such mutual distance, our model then forecasts future poses. Our pipeline is illustrated in Fig. 2. We detail our mutual distance and the pipeline for predicting these distances and future motions below." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.722, + 0.533, + 0.737 + ], + "angle": 0, + "content": "3.1 Mutual Distance Representation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.787, + 0.779 + ], + "angle": 0, + "content": "Let us first introduce our mutual distance representation, which consists of the per-vertex signed distance and the per-basis point distance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.781, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The per-vertex signed distance. Given the human pose \\(\\mathbf{x}_t\\) at time \\(t\\), one can obtain the vertices of human mesh with a predefined mapping, such as SMPL-X [25]. However, constraining every single vertex may not be necessary due to their strong correlation. Hence, following [38], we select a subset of \\(K\\)" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "C. Xing et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.178 + ], + "angle": 0, + "content": "vertices denoted as \\(\\mathbf{v}_t \\in \\mathbb{R}^{K \\times 3}\\) from the human mesh. Given a 3D scene and \\(\\mathbf{v}_t\\) at time \\(t\\), the per-vertex signed distance can be computed as," + }, + { + "type": "equation", + "bbox": [ + 0.358, + 0.189, + 0.786, + 0.238 + ], + "angle": 0, + "content": "\\[\nd _ {k t} = \\left\\{ \\begin{array}{l l} - \\min _ {\\mathbf {y} \\in \\partial S} \\| \\mathbf {v} _ {t k} - \\mathbf {y} \\| _ {2} & \\text {i f} \\mathbf {v} _ {t k} \\in \\mathcal {S} \\\\ \\min _ {\\mathbf {y} \\in \\partial S} \\| \\mathbf {v} _ {t k} - \\mathbf {y} \\| _ {2} & \\text {i f} \\mathbf {v} _ {t k} \\notin \\mathcal {S}, \\end{array} \\right. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.248, + 0.784, + 0.353 + ], + "angle": 0, + "content": "where \\(\\mathbf{d}_t = [d_{1t}, d_{2t}, \\dots, d_{Kt}]^T \\in \\mathbb{R}^K\\) is the per-vertex signed distance at time step \\(t\\), \\(\\mathcal{S}\\) represents the volume inside the actual solid scene and \\(\\partial \\mathcal{S}\\) is the surface points of such scene. The per-vertex signed distance provides information about the whole-body pose. However, the same human mesh at different global positions may result in the same per-vertex signed distance. To resolve such ambiguity, we further leverage the per-basis point distance which is described below." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.355, + 0.784, + 0.431 + ], + "angle": 0, + "content": "The per-basis point distance. Given a set of basis points from the scene denoted as \\(\\mathbf{p} \\in \\mathbb{R}^{P \\times 3}\\), the per-basis point distance is simply the minimum distance in each basis point to the human mesh. As the human moves within the scene, its motion is reflected in the continuous change of the per-basis point distance. Specifically, at time step \\(t\\), it can be computed as," + }, + { + "type": "equation", + "bbox": [ + 0.419, + 0.443, + 0.786, + 0.466 + ], + "angle": 0, + "content": "\\[\nb _ {p t} = \\min _ {\\mathbf {y} \\in \\partial \\mathcal {H} _ {t}} \\| \\mathbf {p} _ {p} - \\mathbf {y} \\| _ {2}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.476, + 0.784, + 0.642 + ], + "angle": 0, + "content": "where \\(\\partial \\mathcal{H}_t\\) is the human surface at time step \\(t\\) and \\(\\mathbf{b}_t = [b_{1t}, b_{2t}, \\dots, b_{Pt}]^T \\in \\mathbb{R}^P\\). Note that our basis points are not necessary on the scene surface. In practice, for different motion sequences, our basis points can be fixed relatively to the root joint of the last observed pose. Such root joint is defined as the origin of that data sample. In particular, we employ the spherical Fibonacci lattice algorithm [7] to sample \\(P\\) near-equidistant points from a sphere with radius \\(R\\). The per-basis point distance primarily captures the global location of the human, while the per-vertex signed distance constrains both the detailed whole-body pose and global location. By combining two distances, our mutual distance representation provides explicit constraints on both the whole-body local poses and the global human motion of the human." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.797, + 0.718 + ], + "angle": 0, + "content": "Given the 3D scene \\(\\mathbf{S}\\), the past motion \\(\\mathbf{X}\\), we first pre-compute the past mutual distance denoted as \\(\\mathbf{D} = [\\mathbf{d}_1, \\mathbf{d}_2, \\dots, \\mathbf{d}_T] \\in \\mathbb{R}^{K \\times T}\\) and \\(\\mathbf{B} = [\\mathbf{b}_1, \\mathbf{b}_2, \\dots, \\mathbf{b}_T] \\in \\mathbb{R}^{P \\times T}\\). Our goal now is to predict the future mutual distance given the above information. In the following section, we will introduce our mutual distance prediction module." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.74, + 0.492, + 0.755 + ], + "angle": 0, + "content": "3.2 Mutual Distance Prediction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.784, + 0.843 + ], + "angle": 0, + "content": "We adopt a similar temporal and spatial encoding strategy proposed in [22], which has been proven to be effective in predicting smooth spatial-temporal data such as 3D human motion. Specifically, we use Discrete Cosine Transform (DCT) to represent a temporal sequence and Graph Convolutional Networks (GCN) to capture the spatial dependencies between different sequences." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.249, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Scene-aware Human Motion Forecasting via Mutual Distance Prediction" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.207 + ], + "angle": 0, + "content": "With the DCT, we represent a sequence as a linear combination of pre-defined cosine bases. In particular, let us denote a \\(T\\)-frame sequence of signed distance of the \\(k\\)-th human vertex as \\(\\tilde{\\mathbf{d}}_k \\in [d_{1k}, d_{2k}, \\dots, d_{Tk}]^T \\in \\mathbb{R}^T\\), the \\(l\\)-th DCT coefficient is then," + }, + { + "type": "equation", + "bbox": [ + 0.31, + 0.218, + 0.786, + 0.259 + ], + "angle": 0, + "content": "\\[\nh _ {k l} = \\sqrt {\\frac {2}{T}} \\sum_ {t = 1} ^ {T} d _ {t k} \\frac {1}{\\sqrt {1 + \\delta_ {l 1}}} \\cos \\left(\\frac {\\pi}{2 T} (2 t - 1) (l - 1)\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.27, + 0.736, + 0.286 + ], + "angle": 0, + "content": "where \\(l \\in \\{1, 2, \\dots, T\\}\\) and \\(\\delta_{ij}\\) denotes the Kronecker delta function, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.428, + 0.296, + 0.786, + 0.339 + ], + "angle": 0, + "content": "\\[\n\\delta_ {i j} = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} i = j \\\\ 0 & \\text {i f} i \\neq j. \\end{array} \\right. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.348, + 0.784, + 0.376 + ], + "angle": 0, + "content": "Note that, with all cosine bases precomputed, the DCT can be simply represented as a linear operation," + }, + { + "type": "equation", + "bbox": [ + 0.457, + 0.377, + 0.786, + 0.394 + ], + "angle": 0, + "content": "\\[\n\\mathbf {h} _ {k} = \\mathbf {C} \\tilde {\\mathbf {d}} _ {k}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.399, + 0.784, + 0.459 + ], + "angle": 0, + "content": "where \\(\\mathbf{C} \\in \\mathbb{R}^{T \\times T}\\) is an orthogonal matrix and each row of \\(\\mathbf{C}\\) stores the precomputed the cosine basis. \\(\\mathbf{h}_k \\in \\mathbb{R}^T\\) are the \\(T\\) DCT coefficients. Given such coefficients \\(\\mathbf{h}_k\\), the original sequence can be losslessly recovered via Inverse DCT (IDCT) as," + }, + { + "type": "equation", + "bbox": [ + 0.452, + 0.459, + 0.786, + 0.477 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {d}} _ {k} = \\mathbf {C} ^ {T} \\mathbf {h} _ {k}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.483, + 0.788, + 0.528 + ], + "angle": 0, + "content": "Similarly, we can obtain the DCT coefficients of all \\(K\\) human vertices as \\(\\mathbf{H} = [\\mathbf{h}_1,\\mathbf{h}_2,\\dots ,\\mathbf{h}_K]^T\\in \\mathbb{R}^{K\\times T}\\). For the spatial encoding, we follow [22] to use GCN. Specifically, a GCN layer is defined as," + }, + { + "type": "equation", + "bbox": [ + 0.396, + 0.538, + 0.786, + 0.557 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} ^ {(n + 1)} = \\sigma \\left(\\mathbf {A} ^ {(n)} \\mathbf {F} ^ {(n)} \\mathbf {W} ^ {(n)}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.566, + 0.788, + 0.732 + ], + "angle": 0, + "content": "where \\(\\sigma\\) is an activation function such as \\(\\tanh(\\cdot)\\), \\(\\mathbf{A}^{(n)} \\in \\mathbb{R}^{K \\times K}\\) is the learnable adjacent matrix at the \\(n\\)-th layer and \\(\\mathbf{W}^{(n)} \\in \\mathbb{R}^{T \\times F}\\) is the feature extraction parameter. \\(\\mathbf{F}^{(n)} \\in \\mathbb{R}^{K \\times F'}\\) and \\(\\mathbf{F}^{(n+1)} \\in \\mathbb{R}^{K \\times F}\\) are the input and the output features of the \\(n\\)-th GCN layer, respectively2. Same as [22], our prediction network consists of many such layers. Since our final goal is to predict the future per-vertex signed distance, we follow the strategy in [22] to first pad the sequence to make a sequence of length \\(T + U\\). Our prediction module then aims to predict a residual between the ground truth DCT coefficients of the whole sequence and those of the padded historical sequence given the scene \\(\\mathbf{S}\\) and past human motion \\(\\mathbf{X}\\). More formally, the future per-vertex signed distance is predicted as follows," + }, + { + "type": "equation", + "bbox": [ + 0.373, + 0.733, + 0.786, + 0.75 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {D}} = \\left(\\mathcal {G} _ {d} (\\hat {\\mathbf {H}}, \\mathcal {G} _ {s} (\\mathbf {S}), \\mathcal {G} _ {x} (\\mathbf {X})) + \\hat {\\mathbf {H}}\\right) \\mathbf {C}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.756, + 0.788, + 0.817 + ], + "angle": 0, + "content": "where \\(\\mathcal{G}_d\\), \\(\\mathcal{G}_s\\) and \\(\\mathcal{G}_x\\) are the GCN predictor, the 3D scene encoding network, and the past motion encoding model, respectively. With a little abuse of notation, here, \\(\\mathbf{C} \\in \\mathbb{R}^{(T + U) \\times (T + U)}\\) is the DCT bases for sequences with length \\(T + U\\). \\(\\hat{\\mathbf{H}} \\in \\mathbb{R}^{K \\times (T + U)}\\) is the DCT coefficients of the padded historical sequence. \\(\\hat{\\mathbf{D}} \\in\\)" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.782, + 0.841 + ], + "angle": 0, + "content": "2 The first GCN layer will take the DCT coefficient matrix as input thus, \\(\\mathbf{F}^{(1)} = \\mathbf{H}\\)" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "C. Xing et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.146, + 0.784, + 0.177 + ], + "angle": 0, + "content": "\\(\\mathbb{R}^{K \\times (T + U)}\\) is the predicted per-vertex signed distance. Note that, here our model also tries to recover the past sequence." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.178, + 0.784, + 0.206 + ], + "angle": 0, + "content": "Similarly, we have another branch to predict the future per-basis point distance as," + }, + { + "type": "equation", + "bbox": [ + 0.379, + 0.207, + 0.785, + 0.223 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {B}} = \\left(\\mathcal {G} _ {b} (\\hat {\\mathbf {K}}, \\mathcal {G} _ {s} (\\mathbf {S}), \\mathcal {G} _ {x} (\\mathbf {X})) + \\hat {\\mathbf {K}}\\right) \\mathbf {C}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.231, + 0.784, + 0.261 + ], + "angle": 0, + "content": "where \\(\\mathcal{G}_b\\) is also a GCN. \\(\\hat{\\mathbf{K}}\\) is the DCT coefficients of padded past per-basis point distance sequence. \\(\\hat{\\mathbf{B}}\\in \\mathbb{R}^{P\\times (T + U)}\\) is the predicted per-basis point distance." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.262, + 0.784, + 0.291 + ], + "angle": 0, + "content": "During training, we use the average \\(\\ell_1\\) loss between the ground-truth distances and the predicted ones. Formally, this loss is defined as," + }, + { + "type": "equation", + "bbox": [ + 0.261, + 0.302, + 0.784, + 0.343 + ], + "angle": 0, + "content": "\\[\n\\ell_ {\\text {d i s t}} = \\frac {1}{(T + U)} \\left(\\frac {1}{K} \\sum_ {t = 1} ^ {T + U} \\sum_ {k = 1} ^ {K} | \\hat {d} _ {k t} - d _ {k t} | + \\frac {1}{P} \\sum_ {t = 1} ^ {T + U} \\sum_ {p = 1} ^ {P} | \\hat {b} _ {p t} - b _ {p t} |\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.357, + 0.784, + 0.389 + ], + "angle": 0, + "content": "where \\(\\hat{d}_{kt}\\) is predicted signed distance for \\(k\\)-th vertex at time \\(t\\) and \\(\\hat{b}_{pt}\\) is predicted distance for \\(p\\)-th basis point at time \\(t\\), \\(d_{kt}\\) and \\(b_{pt}\\) define the ground truth." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.411, + 0.498, + 0.425 + ], + "angle": 0, + "content": "3.3 Motion Forecasting Network" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.435, + 0.784, + 0.495 + ], + "angle": 0, + "content": "Given the future mutual distance \\(\\hat{\\mathbf{D}}\\) and \\(\\hat{\\mathbf{B}}\\), the past human motion \\(\\mathbf{X}\\), and the scene \\(\\mathbf{S}\\), our goal now is to predict the future human poses \\(\\hat{\\mathbf{Y}}\\). To this end, we use an RNN-based motion prediction model \\((\\mathcal{G}_y)\\) which predicts the future motion in an auto-regressive manner. More formally, it is defined as," + }, + { + "type": "equation", + "bbox": [ + 0.331, + 0.507, + 0.784, + 0.523 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {x}} _ {T + u} = \\mathcal {G} _ {y} \\left(\\hat {\\mathbf {x}} _ {T + u - 1}, \\hat {\\mathbf {d}} _ {T + u}, \\hat {\\mathbf {b}} _ {T + u}, \\mathcal {G} _ {x} (\\mathbf {X}), \\mathcal {G} _ {s} (\\mathbf {S})\\right), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.536, + 0.776, + 0.551 + ], + "angle": 0, + "content": "where \\(u\\in \\{1,2,\\dots ,U\\}\\) \\(\\hat{\\mathbf{x}}_{T + u}\\in \\mathbb{R}^{M}\\) is the predicted \\(u\\) -th future human pose." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.552, + 0.784, + 0.611 + ], + "angle": 0, + "content": "During training, in addition to the human pose loss, we incorporate two extra losses to ensure the predicted future pose to be consistent with the ground truth mutual distance. For human pose loss, we employ global translation and orientation loss and the local pose loss. They are defined as follows," + }, + { + "type": "equation", + "bbox": [ + 0.338, + 0.623, + 0.784, + 0.661 + ], + "angle": 0, + "content": "\\[\n\\ell_ {\\text {g l o b a l}} = \\frac {1}{U} \\sum_ {u = 1} ^ {U} \\left| \\hat {\\mathbf {t}} _ {T + u} - \\mathbf {t} _ {T + u} \\right| + \\left| \\hat {\\mathbf {r}} _ {T + u} - \\mathbf {r} _ {T + u} \\right| \\tag {12}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.345, + 0.667, + 0.784, + 0.706 + ], + "angle": 0, + "content": "\\[\n\\ell_ {\\text {l o c a l}} = \\frac {1}{U} \\sum_ {u = 1} ^ {U} \\left| \\hat {\\mathbf {x}} _ {T + u} ^ {\\text {l o c a l}} - \\mathbf {x} _ {T + u} ^ {\\text {l o c a l}} \\right|, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.719, + 0.785, + 0.84 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{t}}_{T + u} \\in \\mathbb{R}^3\\) and \\(\\mathbf{t}_{T + u} \\in \\mathbb{R}^3\\) are the predicted and ground truth global translation, respectively. \\(\\hat{\\mathbf{r}}_{T + u} \\in \\mathbb{SO}(3)\\) and \\(\\mathbf{r}_{T + u} \\in \\mathbb{SO}(3)\\) are the predicted and ground truth global orientation, respectively. Here, we adopt the 6-D rotation representation [41]. \\(\\hat{\\mathbf{x}}_{T + u}^{\\mathrm{local}} \\in \\mathbb{R}^{M - 9}\\) and \\(\\mathbf{x}_{T + u}^{\\mathrm{local}} \\in \\mathbb{R}^{M - 9}\\) are the predicted and ground truth local human pose. To compute the two consistency losses, we first use the pre-defined mapping function such as SMPL-X [25] to obtain the sampled human surface vertices \\(\\hat{\\mathbf{v}}_{T + u} \\in \\mathbb{R}^{K \\times 3}\\) and the set of all human surface points \\(\\partial \\hat{\\mathcal{H}}_{T + u}\\). Recall that we represent the 3D scene as a signed distance" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.25, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Scene-aware Human Motion Forecasting via Mutual Distance Prediction" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "volume. We can query the signed distance of the human surface vertices by interpolating within this volume. The per-vertex signed distance consistency loss is then defined as," + }, + { + "type": "equation", + "bbox": [ + 0.353, + 0.205, + 0.786, + 0.245 + ], + "angle": 0, + "content": "\\[\n\\ell_ {\\text {v e r t e x}} = \\frac {1}{U K} \\sum_ {u = 1} ^ {U} \\sum_ {k = 1} ^ {K} | \\tilde {d} _ {k (T + u)} - d _ {k (T + u)} |, \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.258, + 0.784, + 0.289 + ], + "angle": 0, + "content": "where \\(\\tilde{d}_{k(T + u)}\\) and \\(d_{k(T + u)}\\) are the signed distance of the \\(k\\)-th vertex from the \\(u\\)-th predicted human mesh and ground truth human mesh, respectively." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.29, + 0.785, + 0.336 + ], + "angle": 0, + "content": "Similarly, given the set of human surface points \\(\\partial \\tilde{\\mathcal{H}}_{T + u}\\) at time \\(T + u\\), we can compute the per-basis point distance \\(\\tilde{\\mathbf{b}}_{T + u} \\in \\mathbb{R}^P\\) using Equation 2. The per-basis point distance consistency loss can then be computed as" + }, + { + "type": "equation", + "bbox": [ + 0.359, + 0.347, + 0.786, + 0.389 + ], + "angle": 0, + "content": "\\[\n\\ell_ {\\text {b a s i s}} = \\frac {1}{U P} \\sum_ {u = 1} ^ {U} \\sum_ {p = 1} ^ {P} | \\tilde {b} _ {p (T + u)} - b _ {p (T + u)} |. \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.4, + 0.784, + 0.43 + ], + "angle": 0, + "content": "The final training loss for the motion forecasting network is the combination of all the above losses," + }, + { + "type": "equation", + "bbox": [ + 0.321, + 0.444, + 0.785, + 0.459 + ], + "angle": 0, + "content": "\\[\n\\ell_ {\\text {m o t i o n}} = \\lambda_ {1} \\ell_ {\\text {g l o b a l}} + \\lambda_ {2} \\ell_ {\\text {l o c a l}} + \\lambda_ {3} \\ell_ {\\text {v e r t e x}} + \\lambda_ {4} \\ell_ {\\text {b a s i s}}, \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.47, + 0.611, + 0.485 + ], + "angle": 0, + "content": "where \\(\\lambda_{i}\\) with \\(i\\in \\{1,2,3,4\\}\\) are the balancing weights." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.487, + 0.787, + 0.637 + ], + "angle": 0, + "content": "Here we adopt a stage-wise training strategy where we initially train the mutual distance prediction network and motion forecasting network separately. While our network is differentiable and can be trained from scratch in an end-to-end fashion, we observed that at early stage of the training process, the learning process can easily converge to a local minimum due to inaccurate prediction of mutual distance which serves as the input for the motion forecasting network. We thus propose to perform stage-wise training first. Note that, for stage-wise training process, the motion forecasting network is provided with the ground truth mutual distance as input. Subsequently, we combine both networks and fine-tune them in an end-to-end manner with both losses defined in Eq. 10 and 16." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.661, + 0.376, + 0.679 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.694, + 0.334, + 0.708 + ], + "angle": 0, + "content": "4.1 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.719, + 0.804, + 0.749 + ], + "angle": 0, + "content": "We evaluate our method on four datasets, HUMANISE [36], GTA-IM [4], PROX [10] and GIMO [40]." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.75, + 0.786, + 0.84 + ], + "angle": 0, + "content": "GTA-IM. The GTA Indoor Motion dataset [4], building upon GTA game engine, is a synthetic human-scene interaction dataset. It comprises approximately 1 million RGB-D frames featuring 3D human poses for 50 distinct characters across 7 diverse scenes. To reconstruct the mesh, we fuse the depth maps of each scene and subsequently generate the SDF volume. We follow ContactAware [20] to use the same training-test split and 21 joints. The videos are 30 FPS. Both our" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "C. Xing et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "model and baseline models are trained to observe the past 30 frames to predict the future 60 frames following [20]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.785, + 0.282 + ], + "angle": 0, + "content": "PROX. Proximal Relationships with Object eXclusion (PROX) [10] is a real-world dataset. This dataset contains human SMPL-X parameters [25] fitted from RGB-D frames and 20 distinct characters interacting with various scenes. Since the original motion sequences are jittery, we use the refined version provided by [20]. We adopt the same data split and the 22 joints as in [20]. The frame rate is \\(30\\mathrm{Hz}\\). Our model and baseline models input the past 30 frames to forecast the future 60 frames following [20]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.787, + 0.523 + ], + "angle": 0, + "content": "HUMANISE. The HUMANISE dataset [36] is a synthetic human-scene interaction dataset. It is mainly built upon two primary datasets: the AMASS dataset [19] and the ScanNet dataset [6]. HUMANISE combines these datasets by incorporating different global motions to meet contact and collision constraints. Since the original split includes duplicated human motions in the training and test set which is not suitable for motion prediction, we create a new split to ensure that motions in the test set are never observed during training. Moreover, we ensure that the scenes paired with the motions in the test set are categorized into seen scenes, which include a set of scenes observed during the training, and unseen scenes consisting of scenes never observed during training. Our final split comprises approximately 20k sub-sequences for training and 6k for testing, which includes 5k sub-sequences from seen scenes and 1k from unseen scenes. The videos run at 30 frame-per-second (FPS). Both our model and baseline models are trained to observe past 15 frames (0.5 sec.) and predict future 30 frames (1 sec.) because the average length of motion sequences in the dataset is 2.2 sec. which is much shorter than 3 sec." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.524, + 0.788, + 0.66 + ], + "angle": 0, + "content": "GIMO. The Gaze-Informed Human Motion dataset (GIMO) [40] is a real human-scene interaction dataset. The dataset captures human motion using an IMU-based motion capture system, and the 3D scene is scanned using a smartphone equipped with lidar sensors. Same as [36], GIMO also adopts SMPL-X [25] representation. It captures 11 subjects performing various actions e.g., open window, lie in bed, in 19 scenes, resulting in 127 long motion sequences with \\(192\\mathrm{k}\\) frames. We follow the official training and test split, resulting in around 6k training sequences and 1k testing sequences. Same as HUMANISE [36], frame rate is \\(30\\mathrm{Hz}\\). Our model and baseline models predict future 30 frames with 15 past frames." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.681, + 0.57, + 0.696 + ], + "angle": 0, + "content": "4.2 Metrics,Baselines & Implementation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.785, + 0.764 + ], + "angle": 0, + "content": "Metrics. We follow [20] to evaluate our method using path error and pose error. The path error is defined as the \\( l_{2} \\) distance between the predicted and the ground truth translation; the pose error is the Mean Per Joint Position Error (MPJPE) [12] of local poses." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Baselines. We compare our method with three existing scene-aware human motion prediction models: ContAware [20], GIMO [40] and STAG [27]. ContAware [20] and STAG [27] are state-of-the-art methods which first predict the contact map and then the future motion conditioning on the predicted contact map. We train their model using the official implementation on our datasets," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.25, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Scene-aware Human Motion Forecasting via Mutual Distance Prediction" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.237 + ], + "angle": 0, + "content": "following their original settings to predict only human joints instead of SMPL-X poses. GIMO [40] is another state-of-the-art scene-aware human motion prediction method which mainly focuses on predicting long-term future poses given the past gaze information, and the frame rate of their motion is \\(2\\mathrm{Hz}\\). The original model of GIMO predicts future 10 frames (5 seconds) given the past 6 frames (3 seconds). We slightly adapt their official code to our settings for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.789, + 0.435 + ], + "angle": 0, + "content": "Implementation details. Our models are developed using Pytorch [24] and optimized with the ADAM [13] optimizer. Initially, we train the mutual distance prediction model and the motion forecasting model separately for 40 epochs each. Subsequently, we fine-tune both models together for an additional epoch. The original learning rate is set to 0.0005 and decays by half at the 30-th epoch for both models. We train our model on a single RTX4090. For all datasets, the loss weights of the motion forecasting model \\((\\lambda_1,\\lambda_2,\\lambda_3,\\lambda_4)\\) are 1.0, 0.5, 1.0, 1.0. For each motion sequence, we crop the 3D scene to a region that is within 2.0 meters of the root joint of the last observed pose, and the root joint is used as the origin of the cropped scene. To compute the per-vertex signed distance, we follow [38] to use 67 vertices on the human body. For the basis point set, we select 150 points that are 2.0 meters away from the origin of the scene, and those points are fixed for different sequences." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.458, + 0.785, + 0.697 + ], + "angle": 0, + "content": "
datasetmethodPath Error (mm)Pose Error (mm)
0.5s1.0s1.5s2.0smean0.5s1.0s1.5s2.0smean
GTA-IMContAware [20]†44.582.6125.6182.987.140.154.165.277.251.8
GIMO [40]*52.797.8160.6241.7110.347.960.771.182.759.9
STAG [27]†43.279.8119.9176.483.435.448.759.873.547.0
Ours34.465.9104.0155.672.031.046.858.970.744.6
ContAware w/GT34.546.049.767.641.736.344.348.753.941.3
STAG w/GT32.743.349.468.140.031.037.241.548.335.1
Ours w/GT15.818.920.523.022.420.423.825.828.217.6
PROXContAware [20]93.3187.2284.4381.2192.289.9127.5149.3167.5116.8
GIMO [40]*129.5233.2342.1455.0240.4112.6144.6168.0187.5140.7
STAG [27]†95.9186.4276.6369.8189.698.4137.8163.7185.7129.4
Ours80.9155.2238.6322.1162.683.9119.3142.0159.7110.7
ContAware w/GT73.9106.7104.6117.488.083.7112.9125.2132.9101.1
STAG w/GT64.081.693.4115.477.288.6106.0112.6125.498.3
Ours w/GT44.256.763.769.952.656.469.476.883.464.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.697, + 0.788, + 0.727 + ], + "angle": 0, + "content": "Table 1: Quantitative results on GTA-IM [4] and PROX [10]. Our model outperforms all baselines at all metrics, especially for the path error." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.787, + 0.325, + 0.801 + ], + "angle": 0, + "content": "4.3 Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.803, + 0.842 + ], + "angle": 0, + "content": "Quantitative results. We provide quantitative comparisons across three datasets: GTA-IM [4], PROX [10] and HUMANISE [36] presented in Tables 1 and 2. No" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "C. Xing et al." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.144, + 0.788, + 0.299 + ], + "angle": 0, + "content": "
MethodSeen ScenesUnseen Scenes
PathError (mm)PoseError (mm)PathError (mm)PoseError (mm)
0.5s1.0smean0.5s1.0smean0.5s1.0smean0.5s1.0smean
ContAware [20]†52.8121.157.897.6141.492.953.1124.058.794.0139.190.3
GIMO [40]*70.1129.272.0141.4150.3140.277.7144.080.2146.3159.4146.5
STAG [27]†55.2124.660.788.3131.483.057.0131.563.289.9137.785.6
Ours41.593.545.683.7130.980.046.7100.250.184.3131.880.6
ContAware w/GT38.060.835.793.4119.186.438.261.836.490.3122.084.2
STAG w/GT37.563.236.483.5117.677.337.762.236.685.1122.779.4
Ours w/GT22.933.422.080.3122.076.324.835.523.680.2121.576.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.3, + 0.787, + 0.328 + ], + "angle": 0, + "content": "Table 2: Quantitative results on HUMANISE [36]. Our model outperforms all baselines on both seen scenes and unseen scenes." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.359, + 0.789, + 0.54 + ], + "angle": 0, + "content": "
MethodSeen ScenesUnseen Scenes
PathError (mm)PoseError (mm)PathError (mm)PoseError (mm)
0.5s1.0smean0.5s1.0smean0.5s1.0smean0.5s1.0smean
Motion only53.4111.957.0102.0140.093.559.2119.062.0103.2145.395.5
Scene only48.4105.852.694.6132.987.654.3113.857.994.1136.988.3
Ours w/o D42.495.246.288.2134.683.447.9103.151.290.4140.186.0
Ours w/o B47.6101.350.685.7133.981.752.8109.155.886.4138.383.2
Ours w/ P45.297.249.390.8137.286.049.0101.752.789.8138.986.1
Ours w/ O50.1115.355.392.2141.286.755.0115.258.491.0141.386.2
Ours w/ BPS50.2108.953.9100.0153.694.552.0104.154.394.4148.290.3
Ours e2e47.6105.251.590.4141.086.252.1112.456.189.5142.085.7
Ours41.593.545.683.7130.980.046.7100.250.184.3131.880.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.542, + 0.788, + 0.598 + ], + "angle": 0, + "content": "Table 3: Ablation study on HUMANISE [36]. We show the ablation results of our model. D, B, P, O, BPS and \\( e2e \\) indicate per-vertex signed distance, per-basis point distance, point cloud representation, occupancy representation, BPS encoding and end-to-end training respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.789, + 0.841 + ], + "angle": 0, + "content": "tably, these datasets lack gaze information, with the exception of GIMO [40], for which we report results without gaze, indicated by \\(*\\). Additionally, we have adapted our training strategy to baseline models for a fair comparison. We report the best performance for each model using our training strategy (denoted by \\(\\dagger\\)). Our method consistently outperforms the baseline approaches on all datasets and at all time steps by a substantial margin, both for path and pose predictions. Particularly noteworthy is our method's superiority in path prediction over baseline models on the real-world PROX [10] dataset. Our results demonstrate that our model with ground truth mutual distance significantly outperforms baseline models with ground truth contact (w/GT). This superior performance is evident in both global path and local pose metrics, indicating that our method provides a more effective constraint for human motion. More results on the GIMO [40] dataset can be found in the supplementary material." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.25, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Scene-aware Human Motion Forecasting via Mutual Distance Prediction" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.809, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.4, + 0.788, + 0.444 + ], + "angle": 0, + "content": "Fig. 3: This figure compares our method with baseline models on GTA-IM [4] (top row), PROX [10] (middle row), and HUMANISE [36] (bottom row). Our method predicts future motion closer to the ground truth." + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.469, + 0.788, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.567, + 0.789, + 0.693 + ], + "angle": 0, + "content": "Fig. 4: Ablation of the mutual distance. The three figures in the first three columns (from left to right) depict the ground truth pose for the last frame of future motion, results of our full model, and predictions of our model without mutual distance constraint. The sub-figure in gray is the last observed frame. Other sub-figures depict the predicted middle frame. The blue dot is the scene basis point, and the red dot is the sampled vertex on human mesh. The two graphs in the last two columns show the predicted per-vertex signed distance and the per-basis point distance for the red and blue point, respectively. As shown in the figures, with predicted mutual distance, we can forecast the 'stand-up' action more accurately." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Qualitative results. We present qualitative comparisons with baseline methods on GTA-IM [4], HUMANISE [36], and PROX [10], in Fig. 3. The comparisons further demonstrate that our approach excels in predicting future human motions, aligning closely with ground truth data. In contrast, GIMO [40] exhibits issues such as motion freezing and scene penetration. Additionally, we show the predicted human skeletons generated by ContAware [20] and STAG [27], which struggle to handle complex motions." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "C. Xing et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.481 + ], + "angle": 0, + "content": "Ablation study. To gain a more comprehensive understanding, we present ablation results on HUMANISE [36] dataset in Table 3. We report results for our base models which are defined as our model without scene and mutual distance ('Motion only') and our model without mutual distance ('Scene only'). Notably, there are significant improvements in both path and pose prediction when we incorporate mutual distance into our model. Additionally, we assess the impact of two key components: the proposed per-vertex signed distance ('Ours w/o \\(\\mathbf{B}^{\\prime}\\)) and the per-basis point distance ('Ours w/o \\(\\mathbf{D}^{\\prime}\\)). Each distance individually contributes to the improvement for both path and pose prediction, with the most significant gains achieved when both distances are utilized. Furthermore, we examine the influence of the global scene representation: point cloud representation ('Ours w/ \\(P^{\\prime}\\)) and occupancy representation ('Ours w/ \\(O^{\\prime}\\)). For point cloud representation, similar to ContAware [20], we modified the official PVCNN [18] code to take point cloud as scene input instead of SDF volume. For occupancy representation, our method uses occupancy volume as scene input. Our model outperforms both the point cloud and occupancy representation across all metrics. We also compare the Basis Point Set (BPS) encoding in PLACE [37] with our per-basis point distance ('Ours w/ \\(BPS^{\\prime}\\)) by replacing the per-basis point distance with BPS encoding following the official implementation in [37]. Moreover, we show results from end-to-end training of our method ('Ours \\(e2e^{\\prime}\\)), which encounters a local minimum in the early epochs. Results shown in Fig. 4 further demonstrate the effectiveness of our method." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.504, + 0.36, + 0.521 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.788, + 0.658 + ], + "angle": 0, + "content": "In this work, we present an approach for scene-aware human motion forecasting that explicitly models the human-scene interaction. To constrain the whole-body motion, we have proposed a mutual distance representation: the per-vertex signed distance and the per-basis point distance. The mutual distance captures both the detailed local body movements and the global human motion. Thanks to the explicit mutual distance constraints, Our method generates future human motions that are more plausible and accurate than those produced by state-of-the-art scene-aware motion prediction techniques." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.794 + ], + "angle": 0, + "content": "Societal Impacts & Limitations. The motion prediction results depend on the prediction of the mutual distance is one limitation of our work. Our ablation study using the ground truth mutual distance indicates that enhancing the prediction of mutual distance results in improved motion prediction. This can therefore serve as a potential direction for future research. In real applications, a potential risk of our model is that without explicit real world physics modeling, it is possible for our model to predict future motions that have imbalanced forces. This may lead to unstable human motion when deploy to real life applications like human-robot interaction such as collision." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Acknowledgements This research was supported in part by the Australia Research Council DECRA Fellowship (DE180100628) and ARC Discovery Grant (DP200102274)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.25, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Scene-aware Human Motion Forecasting via Mutual Distance Prediction" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.204 + ], + "angle": 0, + "content": "1. Aksan, E., Kaufmann, M., Hilliges, O.: Structured prediction helps 3d human motion modelling. In: ICCV. pp. 7144-7153 (2019) 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.205, + 0.785, + 0.245 + ], + "angle": 0, + "content": "2. Brand, M., Hertzmann, A.: Style machines. In: Proceedings of the 27th annual conference on Computer graphics and interactive techniques. pp. 183-192. ACM Press/Addison-Wesley Publishing Co. (2000) 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.247, + 0.785, + 0.287 + ], + "angle": 0, + "content": "3. Cai, Y., Huang, L., Wang, Y., Cham, T.J., Cai, J., Yuan, J., Liu, J., Yang, X., Zhu, Y., Shen, X., et al.: Learning progressive joint propagation for human motion prediction. In: ECCV. pp. 226-242. Springer (2020) 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.288, + 0.785, + 0.327 + ], + "angle": 0, + "content": "4. Cao, Z., Gao, H., Mangalam, K., Cai, Q.Z., Vo, M., Malik, J.: Long-term human motion prediction with scene context. In: ECCV. pp. 387-404. Springer (2020) 1, 2, 3, 4, 9, 11, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.33, + 0.785, + 0.355 + ], + "angle": 0, + "content": "5. Corona, E., Pumarola, A., Alenya, G., Moreno-Noguer, F.: Context-aware human motion prediction. In: CVPR. pp. 6992-7001 (2020) 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.357, + 0.785, + 0.396 + ], + "angle": 0, + "content": "6. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: CVPR. pp. 5828-5839 (2017) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.399, + 0.785, + 0.424 + ], + "angle": 0, + "content": "7. González, Á.: Measurement of areas on a sphere using fibonacci and latitude-longitude lattices. Mathematical Geosciences 42, 49-64 (2010) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.426, + 0.785, + 0.452 + ], + "angle": 0, + "content": "8. Gopalakrishnan, A., Mali, A., Kifer, D., Giles, L., Ororbia, A.G.: A neural temporal model for human motion prediction. In: CVPR. pp. 12116-12125 (2019) 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.454, + 0.785, + 0.48 + ], + "angle": 0, + "content": "9. Hassan, M., Ceylan, D., Villegas, R., Saito, J., Yang, J., Zhou, Y., Black, M.J.: Stochastic scene-aware motion prediction. In: ICCV. pp. 11374-11384 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.482, + 0.785, + 0.52 + ], + "angle": 0, + "content": "0. Hassan, M., Choutas, V., Tzionas, D., Black, M.J.: Resolving 3d human pose ambiguities with 3d scene constraints. In: ICCV. pp. 2282-2292 (2019) 9, 10, 11, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.523, + 0.785, + 0.55 + ], + "angle": 0, + "content": "1. Hassan, M., Ghosh, P., Tesch, J., Tzionas, D., Black, M.J.: Populating 3d scenes by learning human-scene interaction. In: CVPR. pp. 14708-14718 (2021) 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.551, + 0.785, + 0.591 + ], + "angle": 0, + "content": "2. Ionescu, C., Papava, D., Olaru, V., Sminchisescu, C.: Human3.6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. TPAMI 36(7), 1325-1339 (2014) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.593, + 0.785, + 0.619 + ], + "angle": 0, + "content": "3. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. In: ICLR (2015) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.62, + 0.785, + 0.646 + ], + "angle": 0, + "content": "4. Kipf, T.N., Welling, M.: Semi-supervised classification with graph convolutional networks. In: ICLR (2017) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.647, + 0.785, + 0.674 + ], + "angle": 0, + "content": "5. Koppula, H.S., Saxena, A.: Anticipating human activities for reactive robotic response. In: IROS. p. 2071. Tokyo (2013) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.675, + 0.785, + 0.716 + ], + "angle": 0, + "content": "6. Li, M., Chen, S., Zhao, Y., Zhang, Y., Wang, Y., Tian, Q.: Dynamic multiscale graph neural networks for 3d skeleton based human motion prediction. In: CVPR. pp. 214-223 (2020) 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.717, + 0.785, + 0.744 + ], + "angle": 0, + "content": "7. Li, X., Li, H., Joo, H., Liu, Y., Sheikh, Y.: Structure from recurrent motion: From rigidity to recurrency. In: CVPR (2018) 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.745, + 0.785, + 0.771 + ], + "angle": 0, + "content": "8. Liu, Z., Tang, H., Lin, Y., Han, S.: Point-voxel cnn for efficient 3d deep learning. Advances in Neural Information Processing Systems 32 (2019) 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.785, + 0.799 + ], + "angle": 0, + "content": "9. Mahmood, N., Ghorbani, N., Troje, N.F., Pons-Moll, G., Black, M.J.: Amass: Archive of motion capture as surface shapes. In: ICCV (2019) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "20. Mao, W., Hartley, R.I., Salzmann, M., et al.: Contact-aware human motion forecasting. Advances in Neural Information Processing Systems 35, 7356-7367 (2022), 1, 2, 3, 4, 5, 9, 10, 11, 12, 13, 14" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "C. Xing et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "21. Mao, W., Liu, M., Salzmann, M.: History repeats itself: Human motion prediction via motion attention. In: ECCV. pp. 474-489. Springer (2020) 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.785, + 0.202 + ], + "angle": 0, + "content": "22. Mao, W., Liu, M., Salzmann, M., Li, H.: Learning trajectory dependencies for human motion prediction. In: ICCV. pp. 9489-9497 (2019) 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.203, + 0.785, + 0.242 + ], + "angle": 0, + "content": "23. Paden, B., Cap, M., Yong, S.Z., Yershov, D., Frazzoli, E.: A survey of motion planning and control techniques for self-driving urban vehicles. IEEE Transactions on intelligent vehicles 1(1), 33-55 (2016) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.243, + 0.785, + 0.283 + ], + "angle": 0, + "content": "24. Paszke, A., Gross, S., Chintala, S., Chanan, G., Yang, E., DeVito, Z., Lin, Z., Desmaison, A., Antiga, L., Lerer, A.: Automatic differentiation in pytorch. In: NeurIPS-W (2017) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.284, + 0.785, + 0.323 + ], + "angle": 0, + "content": "25. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3D hands, face, and body from a single image. In: CVPR. pp. 10975-10985 (2019) 4, 5, 8, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.324, + 0.785, + 0.35 + ], + "angle": 0, + "content": "26. Prokudin, S., Lassner, C., Romero, J.: Efficient learning on point clouds with basis point sets. In: ICCV. pp. 4332-4341 (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.351, + 0.785, + 0.39 + ], + "angle": 0, + "content": "27. Scofano, L., Sampieri, A., Schiele, E., De Matteis, E., Leal-Taixe, L., Galasso, F.: Staged contact-aware global human motion forecasting. In: BMVC (2023) 2, 3, 4, 10, 11, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.391, + 0.785, + 0.418 + ], + "angle": 0, + "content": "28. Sidenbladh, H., Black, M.J., Sigal, L.: Implicit probabilistic models of human motion for synthesis and tracking. In: ECCV. pp. 784-800. Springer (2002) 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.418, + 0.785, + 0.444 + ], + "angle": 0, + "content": "29. Starke, S., Zhang, H., Komura, T., Saito, J.: Neural state machine for character-scene interactions. ACM Trans. Graph. 38(6), 209-1 (2019) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.445, + 0.785, + 0.47 + ], + "angle": 0, + "content": "30. Taylor, G.W., Hinton, G.E., Roweis, S.T.: Modeling human motion using binary latent variables. In: NeurIPS. pp. 1345-1352 (2006) 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.471, + 0.785, + 0.524 + ], + "angle": 0, + "content": "31. Van Welbergen, H., Van Basten, B.J., Egges, A., Ruttkay, Z.M., Overmars, M.H.: Real time animation of virtual humans: a trade-off between naturalness and control. In: Computer Graphics Forum. vol. 29, pp. 2530-2554. Wiley Online Library (2010) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.525, + 0.785, + 0.551 + ], + "angle": 0, + "content": "32. Wang, B., Adeli, E., Chiu, H.k., Huang, D.A., Niebles, J.C.: Imitation learning for human pose prediction. In: ICCV. pp. 7124-7133 (2019) 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.552, + 0.785, + 0.578 + ], + "angle": 0, + "content": "33. Wang, J.M., Fleet, D.J., Hertzmann, A.: Gaussian process dynamical models for human motion. TPAMI 30(2), 283-298 (2008) 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.579, + 0.785, + 0.605 + ], + "angle": 0, + "content": "34. Wang, J., Xu, H., Xu, J., Liu, S., Wang, X.: Synthesizing long-term 3d human motion and interaction in 3d scenes. In: CVPR. pp. 9401-9411 (2021) 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.785, + 0.632 + ], + "angle": 0, + "content": "35. Wang, J., Yan, S., Dai, B., Lin, D.: Scene-aware generative network for human motion synthesis. In: CVPR. pp. 12206-12215 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.672 + ], + "angle": 0, + "content": "36. Wang, Z., Chen, Y., Liu, T., Zhu, Y., Liang, W., Huang, S.: Humanise: Language-conditioned human motion generation in 3d scenes. Advances in Neural Information Processing Systems 35, 14959-14971 (2022) 9, 10, 11, 12, 13, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.673, + 0.785, + 0.711 + ], + "angle": 0, + "content": "37. Zhang, S., Zhang, Y., Ma, Q., Black, M.J., Tang, S.: Place: Proximity learning of articulation and contact in 3d environments. In: 3DV. pp. 642-651. IEEE (2020) 1, 2, 3, 4, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.712, + 0.785, + 0.739 + ], + "angle": 0, + "content": "38. Zhang, Y., Black, M.J., Tang, S.: We are more than our joints: Predicting how 3d bodies move. In: CVPR. pp. 3372-3382 (2021) 5, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.74, + 0.785, + 0.766 + ], + "angle": 0, + "content": "39. Zhang, Y., Hassan, M., Neumann, H., Black, M.J., Tang, S.: Generating 3d people in scenes without people. In: CVPR. pp. 6194-6204 (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.766, + 0.785, + 0.806 + ], + "angle": 0, + "content": "40. Zheng, Y., Yang, Y., Mo, K., Li, J., Yu, T., Liu, Y., Liu, C.K., Guibas, L.J.: Gimo: Gaze-informed human motion prediction in context. In: ECCV. pp. 676-694. Springer (2022) 1, 2, 3, 9, 10, 11, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.807, + 0.785, + 0.833 + ], + "angle": 0, + "content": "41. Zhou, Y., Barnes, C., Lu, J., Yang, J., Li, H.: On the continuity of rotation representations in neural networks. In: CVPR. pp. 5745-5753 (2019) 8" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.833 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Scene-aware Human Motion Forecasting via Mutual Distance Prediction/abecdb1c-2c3f-4eb7-820c-6ae2bc3ee016_origin.pdf b/2024/Scene-aware Human Motion Forecasting via Mutual Distance Prediction/abecdb1c-2c3f-4eb7-820c-6ae2bc3ee016_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9d56c49e15ecdc5b265a1b2ab64e056e51d76a0b --- /dev/null +++ b/2024/Scene-aware Human Motion Forecasting via Mutual Distance Prediction/abecdb1c-2c3f-4eb7-820c-6ae2bc3ee016_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1b61460d197466dee139d82e4c733c8232dd4c7fdf4869fb25d263038e6a4b4e +size 4913765 diff --git a/2024/Scene-aware Human Motion Forecasting via Mutual Distance Prediction/full.md b/2024/Scene-aware Human Motion Forecasting via Mutual Distance Prediction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..51046b59ab257b9c4540bce5f7d23804fc5e0fde --- /dev/null +++ b/2024/Scene-aware Human Motion Forecasting via Mutual Distance Prediction/full.md @@ -0,0 +1,291 @@ +# Scene-aware Human Motion Forecasting via Mutual Distance Prediction + +Chaoyue Xing $^{1}$ , Wei Mao $^{1}$ , and Miaomiao Liu $^{1}$ + +Australian National University, Canberra, Australia + +Abstract. In this paper, we address the issue of scene-aware 3D human motion forecasting. A key challenge in this task is to predict future human motions that are coherent with the scene by modeling human-scene interactions. While recent works have demonstrated that explicit constraints on human-scene interactions can prevent the occurrence of ghost motion, they only provide constraints on partial human motion e.g., the global motion of the human or a few joints contacting the scene, leaving the rest of unconstrained. To address this limitation, we propose to represent the human-scene interaction using the mutual distance between the human body and the scene. Such mutual distances constrain both the local and global human motion, resulting in a whole-body motion constrained prediction. In particular, mutual distance constraints consist of two components, the signed distance of each vertex on the human mesh to the scene surface and the distance of basis scene points to the human mesh. We further introduce a global scene representation learned from a signed distance function (SDF) volume to ensure coherence between the global scene representation and the explicit constraint from the mutual distance. We develop a pipeline with two sequential steps: predicting the future mutual distances first, followed by forecasting future human motion. We explicitly ensure consistency between predicted poses and mutual distances during training. Extensive testing on both synthetic and real datasets demonstrates that our method consistently surpasses the performance of current state-of-the-art techniques. + +Keywords: Scene-aware $\cdot$ Human Motion Forecasting $\cdot$ Mutual Distance $\cdot$ SDF Volume + +# 1 Introduction + +Human motion forecasting has numerous applications, including autonomous driving [23], human-robot interaction [15], animation [31] and virtual/augmented reality [29]. Despite being an active research field for decades [2,28,30,33], most previous methods [1, 3, 8, 16, 17, 21, 22, 32] focus solely on human motion and overlook the fact that humans interact with the 3D environment. In contrast, this paper addresses the problem of scene-aware human motion prediction. + +Recent advancements in scene-aware human synthesis [11,34,37] and human motion prediction [4,5,20,40] have begun to model the human-scene interaction. Some approaches achieve this implicitly by incorporating scene encoding as an + +additional input to their model e.g., 2D scene image [4], 3D scene point cloud [34] or a specific object [5]. However, without explicit constraints, those methods tend to predict motions with artifacts e.g., "ghost motion". Alternatively, other works [20, 27, 37, 40] start to explore the explicit representation of human-scene interaction. However, none of their representations are able to constrain the whole-body motion both locally and globally. For example, the proximity between human and scene used in [37] mainly constrains the global position of the human, while the human gaze employed in [40] only specifies the potential target location of the human. Although the contact map proposed in [20, 27] achieves detailed constraints on both the local pose and global position of certain human joints, their contact map can only constrain joints that contact the scene, neglecting the others, such as elbow joints, which often remain unsupported. + +![](images/732e84fb3e1960611be0718822119a4755b9f3b1df65d655504dbbdd5309c682.jpg) +(a) The per-vertex signed distance (b) The per-basis point distance + +![](images/2502a106d2db1d11117b5842c681cc1ca215eb8fbcc4c5f6f88b26536dd4a397.jpg) +Fig. 1: Our mutual distances. a) shows the per-vertex signed distance for sampled vertices on human mesh. Their color indicates the distance value. b) shows the per-basis point distance for the basis points which are not sampled on the scene surface. For both figures, the darker the color is, the smaller the distance is. + +To constrain the whole-body motion for better prediction, in this paper, we propose a mutual distance representation which captures the distances between the human body and the scene. We further introduce the global scene representation learned from SDF volume instead of point cloud to achieve consistent global and local scene representation learning. Specifically, our proposed mutual distance representation comprises two components: the per-vertex signed distance and the per-basis point distance. Given a 3D scene, the per-vertex distance is the minimum distance from each human body vertex to the scene surface, and its sign indicates whether such vertex penetrates the scene or not (see Fig. 1 (a)). Since the change in signed distance for each body vertex represents the motion of the corresponding body part, it provides a complete constraint on the entire body. However, as a signed distance corresponds to a level-set of the scene Signed Distance Field (SDF), a human mesh at multiple possible global locations can result in the same per-vertex signed distance. + +To address this, we leverage a per-basis point distance, which calculates the minimum distance between each point in a predefined set of basis points and the + +human surface. Our per-basis point distance is similar to BPS body encoding in PLACE [37] with differences in choices of basis points. In particular, their basis points are partially scene-dependent making it not suitable for motion prediction, while our basis points are purely scene-independent and sampled with a fixed distance relative to origin, which can better capture the motion patterns. + +While our model is end-to-end trainable, we adopt a two-step training process followed by fine-tuning for better convergence. In the first step, we predict the mutual distance. Conditioning on the predicted distances, we then forecast the future human motion. Specifically, given a 3D scene, the past human poses, and past mutual distance, we use a strategy based on Graph Convolutional Network (GCN) [14], Discrete Cosine Transform (DCT), for the spatial and temporal encoding of the mutual distances and a simple 3D CNN for encoding of the SDF volume to predict the future ones. We subsequently take the predicted distances, the past human poses, and the scene as input to predict future human motion using an RNN-based model. A consistency prior is adopted to ensure that the predicted human motion is consistent with the input mutual distance. + +Our contributions can be summarized as follows. i) We propose an explicit representation of human-scene interaction using mutual distance to provide constraints for whole-body motion. ii) We explore different global scene representations and find the SDF volume to be more efficient than commonly used point cloud representation. Our experiments demonstrate that our approach outperforms the state-of-the-art scene-aware human motion prediction methods by a large margin on both existing real and synthesis benchmark datasets. + +# 2 Related Work + +Human motion prediction. Researchers have been studying 3D human motion modeling [2,28,30,33] for decades. While traditional methods, such as those employing Hidden Markov Models [2] or the Gaussian process latent variable model [33], are capable of handling periodic and simple non-periodic motions like walking and golf swings, more complex motions are typically modeled with deep neural networks [1,3,8,16,17,21,22,32]. However, these methods focus solely on the prediction of local poses, neglecting the global human motion and 3D scenes in which the motion takes place. + +More recent works [4,5,20,27,40] have begun to incorporate scene context into the task of motion prediction. Specifically, Cao et al. [4] propose to represent the scene with a 2D image. The feature extracted from such an image is fed into the prediction model and lets the model learn the human-scene interaction implicitly. However, such 2D scene representation cannot handle occlusions and it does not ensure a valid human-scene interaction. In [5], a semantic graph is introduced to encode the interaction between the human and the object of interest. Although it is efficient to model a human interacting with one object, it is not suitable for human-scene interaction as the human may interact with multiple objects when he moves around. In [40], Zheng et al. employ eye gaze as additional information to guide the motion forecasting process. Although + +eye gaze provides valuable hints about the goal location of the human motion, it constrains neither the local poses nor the human-scene interaction. Mao et al. [20] introduce a contact map that can explicitly encode the relation between scene surface points and human joints. Scofano et al. [27] propose a three-stage pipeline in a coarse-to-fine manner and utilizing the time information based on [20]. However, for forecasting, their contact map only constrains joints that contact the scene and disregard the rest. By contrast, our introduced mutual distance is able to constrain the whole-body motion. + +Scene-aware human synthesis. One of the main challenges in this task is to ensure a valid human-scene interaction. As such, existing works have explored various strategies to model such interaction [9, 11, 34, 35, 37, 39]. Similar to [4], the interaction is modeled implicitly in [34, 35, 39] via either Generative Adversarial Networks (GANs) [34], or Variational Autoencoders [34, 39]. Besides this, other works [9, 11, 37] adopt explicit representation of human-scene interaction. In particular, it was proposed in POSA [11] to develop a human-centric representation of the human-scene interaction which assigns scene semantic labels, such as floor, and chair, to each human surface vertex. Those labels indicate the probability of such vertex contacting what kind of object in the scene. Although it performs well in the task of placing a human in a scene, it lacks sufficient detail about the local pose of the human, rendering it unsuitable for human motion forecasting. In SAMP [9], Hassan et al. propose to only model the interaction with the object at the goal location. Given the final root translation and orientation, the task is then treated as a path planning problem. Similar to [5], such a strategy cannot constrain the interactions between the human and the scene when the human moves in the scene. The closest to our work is PLACE [37] where the human-scene interaction is represented by the minimum distance between a set of 3D scene points also known as Basis Point Sets (BSP) [26] to the human body surface. However, such representation alone only provides a prior on the global position of the human. For example, one can move its arm that is far from the scene surface while the scene to human proximity remains the same. By contrast, we combine the per-basis point distance with the per-vertex signed distance resulting in a detailed constraint on the whole-body motion both locally and globally. + +# 3 Approach + +Let us begin by formulating the task of scene-aware 3D human motion forecasting. Building on previous work [20], we represent a sequence of $T$ past human poses as $\mathbf{X} = [\mathbf{x}_1, \mathbf{x}_2, \dots, \mathbf{x}_T] \in \mathbb{R}^{M \times T^1}$ , where a human pose at time $t$ is represented as a vector $\mathbf{x}_t \in \mathbb{R}^M$ of length $M$ . Such vector could be a parametric representation of human mesh such as SMPL-X [25] pose. To represent the 3D scene, we adopt a Signed Distance Volume $\mathbf{S} \in \mathbb{R}^{N \times N \times N}$ , each entry of which stores the distance of the voxel to the nearest scene surface and its sign indicates whether it is inside or outside the scene. The representation learned from this + +![](images/29e607422490830b415c1545159ba801100ffb8ccf80748ee9e5074c897ab307.jpg) +Fig. 2: Network architecture. Given the 3D scene $\mathbf{S}$ represented as a signed distance volume and the past motion $\mathbf{X}$ shown in grey meshes, our approach first predicts the future per-vertex signed distance $\hat{\mathbf{D}}$ , and the future per-basis point distance $\hat{\mathbf{B}}$ from historical distance $\mathbf{D}$ and $\mathbf{B}$ , respectively. The two predicted future distances are then fed into a RNN-based network to predict the future motion $\hat{\mathbf{Y}}$ shown in orange meshes. + +volume is more consistent with our mutual distance representation than those learned from point cloud [20]. Moreover, such Signed Distance Volume allows us to query the signed distance of any given position easily via interpolation, which facilitates the fast calculation of the loss defined on the SDFs of the predicted pose sequence during training (see Eq. 14). + +Our objective is predicting $U$ future human poses $\mathbf{Y} = [\mathbf{x}_{T + 1},\mathbf{x}_{T + 2},\dots ,\mathbf{x}_{T + U}]\in$ $\mathbb{R}^{M\times U}$ with history motion $\mathbf{X}$ and 3D scene S. We introduce an innovative pipeline that predicts mutual distance representation first i.e., the signed minimum distance from each human body vertex to the scene surface (the per-vertex signed distance) and the minimum distance from each basis point to the human surface (the per-basis point distance) at each time frame. Conditioned on such mutual distance, our model then forecasts future poses. Our pipeline is illustrated in Fig. 2. We detail our mutual distance and the pipeline for predicting these distances and future motions below. + +# 3.1 Mutual Distance Representation + +Let us first introduce our mutual distance representation, which consists of the per-vertex signed distance and the per-basis point distance. + +The per-vertex signed distance. Given the human pose $\mathbf{x}_t$ at time $t$ , one can obtain the vertices of human mesh with a predefined mapping, such as SMPL-X [25]. However, constraining every single vertex may not be necessary due to their strong correlation. Hence, following [38], we select a subset of $K$ + +vertices denoted as $\mathbf{v}_t \in \mathbb{R}^{K \times 3}$ from the human mesh. Given a 3D scene and $\mathbf{v}_t$ at time $t$ , the per-vertex signed distance can be computed as, + +$$ +d _ {k t} = \left\{ \begin{array}{l l} - \min _ {\mathbf {y} \in \partial S} \| \mathbf {v} _ {t k} - \mathbf {y} \| _ {2} & \text {i f} \mathbf {v} _ {t k} \in \mathcal {S} \\ \min _ {\mathbf {y} \in \partial S} \| \mathbf {v} _ {t k} - \mathbf {y} \| _ {2} & \text {i f} \mathbf {v} _ {t k} \notin \mathcal {S}, \end{array} \right. \tag {1} +$$ + +where $\mathbf{d}_t = [d_{1t}, d_{2t}, \dots, d_{Kt}]^T \in \mathbb{R}^K$ is the per-vertex signed distance at time step $t$ , $\mathcal{S}$ represents the volume inside the actual solid scene and $\partial \mathcal{S}$ is the surface points of such scene. The per-vertex signed distance provides information about the whole-body pose. However, the same human mesh at different global positions may result in the same per-vertex signed distance. To resolve such ambiguity, we further leverage the per-basis point distance which is described below. + +The per-basis point distance. Given a set of basis points from the scene denoted as $\mathbf{p} \in \mathbb{R}^{P \times 3}$ , the per-basis point distance is simply the minimum distance in each basis point to the human mesh. As the human moves within the scene, its motion is reflected in the continuous change of the per-basis point distance. Specifically, at time step $t$ , it can be computed as, + +$$ +b _ {p t} = \min _ {\mathbf {y} \in \partial \mathcal {H} _ {t}} \| \mathbf {p} _ {p} - \mathbf {y} \| _ {2}, \tag {2} +$$ + +where $\partial \mathcal{H}_t$ is the human surface at time step $t$ and $\mathbf{b}_t = [b_{1t}, b_{2t}, \dots, b_{Pt}]^T \in \mathbb{R}^P$ . Note that our basis points are not necessary on the scene surface. In practice, for different motion sequences, our basis points can be fixed relatively to the root joint of the last observed pose. Such root joint is defined as the origin of that data sample. In particular, we employ the spherical Fibonacci lattice algorithm [7] to sample $P$ near-equidistant points from a sphere with radius $R$ . The per-basis point distance primarily captures the global location of the human, while the per-vertex signed distance constrains both the detailed whole-body pose and global location. By combining two distances, our mutual distance representation provides explicit constraints on both the whole-body local poses and the global human motion of the human. + +Given the 3D scene $\mathbf{S}$ , the past motion $\mathbf{X}$ , we first pre-compute the past mutual distance denoted as $\mathbf{D} = [\mathbf{d}_1, \mathbf{d}_2, \dots, \mathbf{d}_T] \in \mathbb{R}^{K \times T}$ and $\mathbf{B} = [\mathbf{b}_1, \mathbf{b}_2, \dots, \mathbf{b}_T] \in \mathbb{R}^{P \times T}$ . Our goal now is to predict the future mutual distance given the above information. In the following section, we will introduce our mutual distance prediction module. + +# 3.2 Mutual Distance Prediction + +We adopt a similar temporal and spatial encoding strategy proposed in [22], which has been proven to be effective in predicting smooth spatial-temporal data such as 3D human motion. Specifically, we use Discrete Cosine Transform (DCT) to represent a temporal sequence and Graph Convolutional Networks (GCN) to capture the spatial dependencies between different sequences. + +With the DCT, we represent a sequence as a linear combination of pre-defined cosine bases. In particular, let us denote a $T$ -frame sequence of signed distance of the $k$ -th human vertex as $\tilde{\mathbf{d}}_k \in [d_{1k}, d_{2k}, \dots, d_{Tk}]^T \in \mathbb{R}^T$ , the $l$ -th DCT coefficient is then, + +$$ +h _ {k l} = \sqrt {\frac {2}{T}} \sum_ {t = 1} ^ {T} d _ {t k} \frac {1}{\sqrt {1 + \delta_ {l 1}}} \cos \left(\frac {\pi}{2 T} (2 t - 1) (l - 1)\right), \tag {3} +$$ + +where $l \in \{1, 2, \dots, T\}$ and $\delta_{ij}$ denotes the Kronecker delta function, i.e., + +$$ +\delta_ {i j} = \left\{ \begin{array}{l l} 1 & \text {i f} i = j \\ 0 & \text {i f} i \neq j. \end{array} \right. \tag {4} +$$ + +Note that, with all cosine bases precomputed, the DCT can be simply represented as a linear operation, + +$$ +\mathbf {h} _ {k} = \mathbf {C} \tilde {\mathbf {d}} _ {k}, \tag {5} +$$ + +where $\mathbf{C} \in \mathbb{R}^{T \times T}$ is an orthogonal matrix and each row of $\mathbf{C}$ stores the precomputed the cosine basis. $\mathbf{h}_k \in \mathbb{R}^T$ are the $T$ DCT coefficients. Given such coefficients $\mathbf{h}_k$ , the original sequence can be losslessly recovered via Inverse DCT (IDCT) as, + +$$ +\tilde {\mathbf {d}} _ {k} = \mathbf {C} ^ {T} \mathbf {h} _ {k}. \tag {6} +$$ + +Similarly, we can obtain the DCT coefficients of all $K$ human vertices as $\mathbf{H} = [\mathbf{h}_1,\mathbf{h}_2,\dots ,\mathbf{h}_K]^T\in \mathbb{R}^{K\times T}$ . For the spatial encoding, we follow [22] to use GCN. Specifically, a GCN layer is defined as, + +$$ +\mathbf {F} ^ {(n + 1)} = \sigma \left(\mathbf {A} ^ {(n)} \mathbf {F} ^ {(n)} \mathbf {W} ^ {(n)}\right), \tag {7} +$$ + +where $\sigma$ is an activation function such as $\tanh(\cdot)$ , $\mathbf{A}^{(n)} \in \mathbb{R}^{K \times K}$ is the learnable adjacent matrix at the $n$ -th layer and $\mathbf{W}^{(n)} \in \mathbb{R}^{T \times F}$ is the feature extraction parameter. $\mathbf{F}^{(n)} \in \mathbb{R}^{K \times F'}$ and $\mathbf{F}^{(n+1)} \in \mathbb{R}^{K \times F}$ are the input and the output features of the $n$ -th GCN layer, respectively2. Same as [22], our prediction network consists of many such layers. Since our final goal is to predict the future per-vertex signed distance, we follow the strategy in [22] to first pad the sequence to make a sequence of length $T + U$ . Our prediction module then aims to predict a residual between the ground truth DCT coefficients of the whole sequence and those of the padded historical sequence given the scene $\mathbf{S}$ and past human motion $\mathbf{X}$ . More formally, the future per-vertex signed distance is predicted as follows, + +$$ +\hat {\mathbf {D}} = \left(\mathcal {G} _ {d} (\hat {\mathbf {H}}, \mathcal {G} _ {s} (\mathbf {S}), \mathcal {G} _ {x} (\mathbf {X})) + \hat {\mathbf {H}}\right) \mathbf {C}, \tag {8} +$$ + +where $\mathcal{G}_d$ , $\mathcal{G}_s$ and $\mathcal{G}_x$ are the GCN predictor, the 3D scene encoding network, and the past motion encoding model, respectively. With a little abuse of notation, here, $\mathbf{C} \in \mathbb{R}^{(T + U) \times (T + U)}$ is the DCT bases for sequences with length $T + U$ . $\hat{\mathbf{H}} \in \mathbb{R}^{K \times (T + U)}$ is the DCT coefficients of the padded historical sequence. $\hat{\mathbf{D}} \in$ + +$\mathbb{R}^{K \times (T + U)}$ is the predicted per-vertex signed distance. Note that, here our model also tries to recover the past sequence. + +Similarly, we have another branch to predict the future per-basis point distance as, + +$$ +\tilde {\mathbf {B}} = \left(\mathcal {G} _ {b} (\hat {\mathbf {K}}, \mathcal {G} _ {s} (\mathbf {S}), \mathcal {G} _ {x} (\mathbf {X})) + \hat {\mathbf {K}}\right) \mathbf {C}, \tag {9} +$$ + +where $\mathcal{G}_b$ is also a GCN. $\hat{\mathbf{K}}$ is the DCT coefficients of padded past per-basis point distance sequence. $\hat{\mathbf{B}}\in \mathbb{R}^{P\times (T + U)}$ is the predicted per-basis point distance. + +During training, we use the average $\ell_1$ loss between the ground-truth distances and the predicted ones. Formally, this loss is defined as, + +$$ +\ell_ {\text {d i s t}} = \frac {1}{(T + U)} \left(\frac {1}{K} \sum_ {t = 1} ^ {T + U} \sum_ {k = 1} ^ {K} | \hat {d} _ {k t} - d _ {k t} | + \frac {1}{P} \sum_ {t = 1} ^ {T + U} \sum_ {p = 1} ^ {P} | \hat {b} _ {p t} - b _ {p t} |\right), \tag {10} +$$ + +where $\hat{d}_{kt}$ is predicted signed distance for $k$ -th vertex at time $t$ and $\hat{b}_{pt}$ is predicted distance for $p$ -th basis point at time $t$ , $d_{kt}$ and $b_{pt}$ define the ground truth. + +# 3.3 Motion Forecasting Network + +Given the future mutual distance $\hat{\mathbf{D}}$ and $\hat{\mathbf{B}}$ , the past human motion $\mathbf{X}$ , and the scene $\mathbf{S}$ , our goal now is to predict the future human poses $\hat{\mathbf{Y}}$ . To this end, we use an RNN-based motion prediction model $(\mathcal{G}_y)$ which predicts the future motion in an auto-regressive manner. More formally, it is defined as, + +$$ +\hat {\mathbf {x}} _ {T + u} = \mathcal {G} _ {y} \left(\hat {\mathbf {x}} _ {T + u - 1}, \hat {\mathbf {d}} _ {T + u}, \hat {\mathbf {b}} _ {T + u}, \mathcal {G} _ {x} (\mathbf {X}), \mathcal {G} _ {s} (\mathbf {S})\right), \tag {11} +$$ + +where $u\in \{1,2,\dots ,U\}$ $\hat{\mathbf{x}}_{T + u}\in \mathbb{R}^{M}$ is the predicted $u$ -th future human pose. + +During training, in addition to the human pose loss, we incorporate two extra losses to ensure the predicted future pose to be consistent with the ground truth mutual distance. For human pose loss, we employ global translation and orientation loss and the local pose loss. They are defined as follows, + +$$ +\ell_ {\text {g l o b a l}} = \frac {1}{U} \sum_ {u = 1} ^ {U} \left| \hat {\mathbf {t}} _ {T + u} - \mathbf {t} _ {T + u} \right| + \left| \hat {\mathbf {r}} _ {T + u} - \mathbf {r} _ {T + u} \right| \tag {12} +$$ + +$$ +\ell_ {\text {l o c a l}} = \frac {1}{U} \sum_ {u = 1} ^ {U} \left| \hat {\mathbf {x}} _ {T + u} ^ {\text {l o c a l}} - \mathbf {x} _ {T + u} ^ {\text {l o c a l}} \right|, \tag {13} +$$ + +where $\hat{\mathbf{t}}_{T + u} \in \mathbb{R}^3$ and $\mathbf{t}_{T + u} \in \mathbb{R}^3$ are the predicted and ground truth global translation, respectively. $\hat{\mathbf{r}}_{T + u} \in \mathbb{SO}(3)$ and $\mathbf{r}_{T + u} \in \mathbb{SO}(3)$ are the predicted and ground truth global orientation, respectively. Here, we adopt the 6-D rotation representation [41]. $\hat{\mathbf{x}}_{T + u}^{\mathrm{local}} \in \mathbb{R}^{M - 9}$ and $\mathbf{x}_{T + u}^{\mathrm{local}} \in \mathbb{R}^{M - 9}$ are the predicted and ground truth local human pose. To compute the two consistency losses, we first use the pre-defined mapping function such as SMPL-X [25] to obtain the sampled human surface vertices $\hat{\mathbf{v}}_{T + u} \in \mathbb{R}^{K \times 3}$ and the set of all human surface points $\partial \hat{\mathcal{H}}_{T + u}$ . Recall that we represent the 3D scene as a signed distance + +volume. We can query the signed distance of the human surface vertices by interpolating within this volume. The per-vertex signed distance consistency loss is then defined as, + +$$ +\ell_ {\text {v e r t e x}} = \frac {1}{U K} \sum_ {u = 1} ^ {U} \sum_ {k = 1} ^ {K} | \tilde {d} _ {k (T + u)} - d _ {k (T + u)} |, \tag {14} +$$ + +where $\tilde{d}_{k(T + u)}$ and $d_{k(T + u)}$ are the signed distance of the $k$ -th vertex from the $u$ -th predicted human mesh and ground truth human mesh, respectively. + +Similarly, given the set of human surface points $\partial \tilde{\mathcal{H}}_{T + u}$ at time $T + u$ , we can compute the per-basis point distance $\tilde{\mathbf{b}}_{T + u} \in \mathbb{R}^P$ using Equation 2. The per-basis point distance consistency loss can then be computed as + +$$ +\ell_ {\text {b a s i s}} = \frac {1}{U P} \sum_ {u = 1} ^ {U} \sum_ {p = 1} ^ {P} | \tilde {b} _ {p (T + u)} - b _ {p (T + u)} |. \tag {15} +$$ + +The final training loss for the motion forecasting network is the combination of all the above losses, + +$$ +\ell_ {\text {m o t i o n}} = \lambda_ {1} \ell_ {\text {g l o b a l}} + \lambda_ {2} \ell_ {\text {l o c a l}} + \lambda_ {3} \ell_ {\text {v e r t e x}} + \lambda_ {4} \ell_ {\text {b a s i s}}, \tag {16} +$$ + +where $\lambda_{i}$ with $i\in \{1,2,3,4\}$ are the balancing weights. + +Here we adopt a stage-wise training strategy where we initially train the mutual distance prediction network and motion forecasting network separately. While our network is differentiable and can be trained from scratch in an end-to-end fashion, we observed that at early stage of the training process, the learning process can easily converge to a local minimum due to inaccurate prediction of mutual distance which serves as the input for the motion forecasting network. We thus propose to perform stage-wise training first. Note that, for stage-wise training process, the motion forecasting network is provided with the ground truth mutual distance as input. Subsequently, we combine both networks and fine-tune them in an end-to-end manner with both losses defined in Eq. 10 and 16. + +# 4 Experiments + +# 4.1 Datasets + +We evaluate our method on four datasets, HUMANISE [36], GTA-IM [4], PROX [10] and GIMO [40]. + +GTA-IM. The GTA Indoor Motion dataset [4], building upon GTA game engine, is a synthetic human-scene interaction dataset. It comprises approximately 1 million RGB-D frames featuring 3D human poses for 50 distinct characters across 7 diverse scenes. To reconstruct the mesh, we fuse the depth maps of each scene and subsequently generate the SDF volume. We follow ContactAware [20] to use the same training-test split and 21 joints. The videos are 30 FPS. Both our + +model and baseline models are trained to observe the past 30 frames to predict the future 60 frames following [20]. + +PROX. Proximal Relationships with Object eXclusion (PROX) [10] is a real-world dataset. This dataset contains human SMPL-X parameters [25] fitted from RGB-D frames and 20 distinct characters interacting with various scenes. Since the original motion sequences are jittery, we use the refined version provided by [20]. We adopt the same data split and the 22 joints as in [20]. The frame rate is $30\mathrm{Hz}$ . Our model and baseline models input the past 30 frames to forecast the future 60 frames following [20]. + +HUMANISE. The HUMANISE dataset [36] is a synthetic human-scene interaction dataset. It is mainly built upon two primary datasets: the AMASS dataset [19] and the ScanNet dataset [6]. HUMANISE combines these datasets by incorporating different global motions to meet contact and collision constraints. Since the original split includes duplicated human motions in the training and test set which is not suitable for motion prediction, we create a new split to ensure that motions in the test set are never observed during training. Moreover, we ensure that the scenes paired with the motions in the test set are categorized into seen scenes, which include a set of scenes observed during the training, and unseen scenes consisting of scenes never observed during training. Our final split comprises approximately 20k sub-sequences for training and 6k for testing, which includes 5k sub-sequences from seen scenes and 1k from unseen scenes. The videos run at 30 frame-per-second (FPS). Both our model and baseline models are trained to observe past 15 frames (0.5 sec.) and predict future 30 frames (1 sec.) because the average length of motion sequences in the dataset is 2.2 sec. which is much shorter than 3 sec. + +GIMO. The Gaze-Informed Human Motion dataset (GIMO) [40] is a real human-scene interaction dataset. The dataset captures human motion using an IMU-based motion capture system, and the 3D scene is scanned using a smartphone equipped with lidar sensors. Same as [36], GIMO also adopts SMPL-X [25] representation. It captures 11 subjects performing various actions e.g., open window, lie in bed, in 19 scenes, resulting in 127 long motion sequences with $192\mathrm{k}$ frames. We follow the official training and test split, resulting in around 6k training sequences and 1k testing sequences. Same as HUMANISE [36], frame rate is $30\mathrm{Hz}$ . Our model and baseline models predict future 30 frames with 15 past frames. + +# 4.2 Metrics,Baselines & Implementation + +Metrics. We follow [20] to evaluate our method using path error and pose error. The path error is defined as the $l_{2}$ distance between the predicted and the ground truth translation; the pose error is the Mean Per Joint Position Error (MPJPE) [12] of local poses. + +Baselines. We compare our method with three existing scene-aware human motion prediction models: ContAware [20], GIMO [40] and STAG [27]. ContAware [20] and STAG [27] are state-of-the-art methods which first predict the contact map and then the future motion conditioning on the predicted contact map. We train their model using the official implementation on our datasets, + +following their original settings to predict only human joints instead of SMPL-X poses. GIMO [40] is another state-of-the-art scene-aware human motion prediction method which mainly focuses on predicting long-term future poses given the past gaze information, and the frame rate of their motion is $2\mathrm{Hz}$ . The original model of GIMO predicts future 10 frames (5 seconds) given the past 6 frames (3 seconds). We slightly adapt their official code to our settings for evaluation. + +Implementation details. Our models are developed using Pytorch [24] and optimized with the ADAM [13] optimizer. Initially, we train the mutual distance prediction model and the motion forecasting model separately for 40 epochs each. Subsequently, we fine-tune both models together for an additional epoch. The original learning rate is set to 0.0005 and decays by half at the 30-th epoch for both models. We train our model on a single RTX4090. For all datasets, the loss weights of the motion forecasting model $(\lambda_1,\lambda_2,\lambda_3,\lambda_4)$ are 1.0, 0.5, 1.0, 1.0. For each motion sequence, we crop the 3D scene to a region that is within 2.0 meters of the root joint of the last observed pose, and the root joint is used as the origin of the cropped scene. To compute the per-vertex signed distance, we follow [38] to use 67 vertices on the human body. For the basis point set, we select 150 points that are 2.0 meters away from the origin of the scene, and those points are fixed for different sequences. + +
datasetmethodPath Error (mm)Pose Error (mm)
0.5s1.0s1.5s2.0smean0.5s1.0s1.5s2.0smean
GTA-IMContAware [20]†44.582.6125.6182.987.140.154.165.277.251.8
GIMO [40]*52.797.8160.6241.7110.347.960.771.182.759.9
STAG [27]†43.279.8119.9176.483.435.448.759.873.547.0
Ours34.465.9104.0155.672.031.046.858.970.744.6
ContAware w/GT34.546.049.767.641.736.344.348.753.941.3
STAG w/GT32.743.349.468.140.031.037.241.548.335.1
Ours w/GT15.818.920.523.022.420.423.825.828.217.6
PROXContAware [20]93.3187.2284.4381.2192.289.9127.5149.3167.5116.8
GIMO [40]*129.5233.2342.1455.0240.4112.6144.6168.0187.5140.7
STAG [27]†95.9186.4276.6369.8189.698.4137.8163.7185.7129.4
Ours80.9155.2238.6322.1162.683.9119.3142.0159.7110.7
ContAware w/GT73.9106.7104.6117.488.083.7112.9125.2132.9101.1
STAG w/GT64.081.693.4115.477.288.6106.0112.6125.498.3
Ours w/GT44.256.763.769.952.656.469.476.883.464.7
+ +Table 1: Quantitative results on GTA-IM [4] and PROX [10]. Our model outperforms all baselines at all metrics, especially for the path error. + +# 4.3 Results + +Quantitative results. We provide quantitative comparisons across three datasets: GTA-IM [4], PROX [10] and HUMANISE [36] presented in Tables 1 and 2. No + +
MethodSeen ScenesUnseen Scenes
PathError (mm)PoseError (mm)PathError (mm)PoseError (mm)
0.5s1.0smean0.5s1.0smean0.5s1.0smean0.5s1.0smean
ContAware [20]†52.8121.157.897.6141.492.953.1124.058.794.0139.190.3
GIMO [40]*70.1129.272.0141.4150.3140.277.7144.080.2146.3159.4146.5
STAG [27]†55.2124.660.788.3131.483.057.0131.563.289.9137.785.6
Ours41.593.545.683.7130.980.046.7100.250.184.3131.880.6
ContAware w/GT38.060.835.793.4119.186.438.261.836.490.3122.084.2
STAG w/GT37.563.236.483.5117.677.337.762.236.685.1122.779.4
Ours w/GT22.933.422.080.3122.076.324.835.523.680.2121.576.2
+ +Table 2: Quantitative results on HUMANISE [36]. Our model outperforms all baselines on both seen scenes and unseen scenes. + +
MethodSeen ScenesUnseen Scenes
PathError (mm)PoseError (mm)PathError (mm)PoseError (mm)
0.5s1.0smean0.5s1.0smean0.5s1.0smean0.5s1.0smean
Motion only53.4111.957.0102.0140.093.559.2119.062.0103.2145.395.5
Scene only48.4105.852.694.6132.987.654.3113.857.994.1136.988.3
Ours w/o D42.495.246.288.2134.683.447.9103.151.290.4140.186.0
Ours w/o B47.6101.350.685.7133.981.752.8109.155.886.4138.383.2
Ours w/ P45.297.249.390.8137.286.049.0101.752.789.8138.986.1
Ours w/ O50.1115.355.392.2141.286.755.0115.258.491.0141.386.2
Ours w/ BPS50.2108.953.9100.0153.694.552.0104.154.394.4148.290.3
Ours e2e47.6105.251.590.4141.086.252.1112.456.189.5142.085.7
Ours41.593.545.683.7130.980.046.7100.250.184.3131.880.6
+ +Table 3: Ablation study on HUMANISE [36]. We show the ablation results of our model. D, B, P, O, BPS and $e2e$ indicate per-vertex signed distance, per-basis point distance, point cloud representation, occupancy representation, BPS encoding and end-to-end training respectively. + +tably, these datasets lack gaze information, with the exception of GIMO [40], for which we report results without gaze, indicated by $*$ . Additionally, we have adapted our training strategy to baseline models for a fair comparison. We report the best performance for each model using our training strategy (denoted by $\dagger$ ). Our method consistently outperforms the baseline approaches on all datasets and at all time steps by a substantial margin, both for path and pose predictions. Particularly noteworthy is our method's superiority in path prediction over baseline models on the real-world PROX [10] dataset. Our results demonstrate that our model with ground truth mutual distance significantly outperforms baseline models with ground truth contact (w/GT). This superior performance is evident in both global path and local pose metrics, indicating that our method provides a more effective constraint for human motion. More results on the GIMO [40] dataset can be found in the supplementary material. + +![](images/57085e4aba0712c3ad0c59f9ac22f593ec4dd85ec138cea4c3ac27e9d48f4e31.jpg) +Fig. 3: This figure compares our method with baseline models on GTA-IM [4] (top row), PROX [10] (middle row), and HUMANISE [36] (bottom row). Our method predicts future motion closer to the ground truth. + +![](images/eaed32a321b1b18005b390bb959351aec0903028d49200fcc905b9dedc29fbcb.jpg) +Fig. 4: Ablation of the mutual distance. The three figures in the first three columns (from left to right) depict the ground truth pose for the last frame of future motion, results of our full model, and predictions of our model without mutual distance constraint. The sub-figure in gray is the last observed frame. Other sub-figures depict the predicted middle frame. The blue dot is the scene basis point, and the red dot is the sampled vertex on human mesh. The two graphs in the last two columns show the predicted per-vertex signed distance and the per-basis point distance for the red and blue point, respectively. As shown in the figures, with predicted mutual distance, we can forecast the 'stand-up' action more accurately. + +Qualitative results. We present qualitative comparisons with baseline methods on GTA-IM [4], HUMANISE [36], and PROX [10], in Fig. 3. The comparisons further demonstrate that our approach excels in predicting future human motions, aligning closely with ground truth data. In contrast, GIMO [40] exhibits issues such as motion freezing and scene penetration. Additionally, we show the predicted human skeletons generated by ContAware [20] and STAG [27], which struggle to handle complex motions. + +Ablation study. To gain a more comprehensive understanding, we present ablation results on HUMANISE [36] dataset in Table 3. We report results for our base models which are defined as our model without scene and mutual distance ('Motion only') and our model without mutual distance ('Scene only'). Notably, there are significant improvements in both path and pose prediction when we incorporate mutual distance into our model. Additionally, we assess the impact of two key components: the proposed per-vertex signed distance ('Ours w/o $\mathbf{B}^{\prime}$ ) and the per-basis point distance ('Ours w/o $\mathbf{D}^{\prime}$ ). Each distance individually contributes to the improvement for both path and pose prediction, with the most significant gains achieved when both distances are utilized. Furthermore, we examine the influence of the global scene representation: point cloud representation ('Ours w/ $P^{\prime}$ ) and occupancy representation ('Ours w/ $O^{\prime}$ ). For point cloud representation, similar to ContAware [20], we modified the official PVCNN [18] code to take point cloud as scene input instead of SDF volume. For occupancy representation, our method uses occupancy volume as scene input. Our model outperforms both the point cloud and occupancy representation across all metrics. We also compare the Basis Point Set (BPS) encoding in PLACE [37] with our per-basis point distance ('Ours w/ $BPS^{\prime}$ ) by replacing the per-basis point distance with BPS encoding following the official implementation in [37]. Moreover, we show results from end-to-end training of our method ('Ours $e2e^{\prime}$ ), which encounters a local minimum in the early epochs. Results shown in Fig. 4 further demonstrate the effectiveness of our method. + +# 5 Conclusion + +In this work, we present an approach for scene-aware human motion forecasting that explicitly models the human-scene interaction. To constrain the whole-body motion, we have proposed a mutual distance representation: the per-vertex signed distance and the per-basis point distance. The mutual distance captures both the detailed local body movements and the global human motion. Thanks to the explicit mutual distance constraints, Our method generates future human motions that are more plausible and accurate than those produced by state-of-the-art scene-aware motion prediction techniques. + +Societal Impacts & Limitations. The motion prediction results depend on the prediction of the mutual distance is one limitation of our work. Our ablation study using the ground truth mutual distance indicates that enhancing the prediction of mutual distance results in improved motion prediction. This can therefore serve as a potential direction for future research. In real applications, a potential risk of our model is that without explicit real world physics modeling, it is possible for our model to predict future motions that have imbalanced forces. This may lead to unstable human motion when deploy to real life applications like human-robot interaction such as collision. + +Acknowledgements This research was supported in part by the Australia Research Council DECRA Fellowship (DE180100628) and ARC Discovery Grant (DP200102274). + +# References + +1. Aksan, E., Kaufmann, M., Hilliges, O.: Structured prediction helps 3d human motion modelling. In: ICCV. pp. 7144-7153 (2019) 1, 3 +2. Brand, M., Hertzmann, A.: Style machines. In: Proceedings of the 27th annual conference on Computer graphics and interactive techniques. pp. 183-192. ACM Press/Addison-Wesley Publishing Co. (2000) 1, 3 +3. Cai, Y., Huang, L., Wang, Y., Cham, T.J., Cai, J., Yuan, J., Liu, J., Yang, X., Zhu, Y., Shen, X., et al.: Learning progressive joint propagation for human motion prediction. In: ECCV. pp. 226-242. Springer (2020) 1, 3 +4. Cao, Z., Gao, H., Mangalam, K., Cai, Q.Z., Vo, M., Malik, J.: Long-term human motion prediction with scene context. In: ECCV. pp. 387-404. Springer (2020) 1, 2, 3, 4, 9, 11, 13 +5. Corona, E., Pumarola, A., Alenya, G., Moreno-Noguer, F.: Context-aware human motion prediction. In: CVPR. pp. 6992-7001 (2020) 1, 2, 3, 4 +6. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: CVPR. pp. 5828-5839 (2017) 10 +7. González, Á.: Measurement of areas on a sphere using fibonacci and latitude-longitude lattices. Mathematical Geosciences 42, 49-64 (2010) 6 +8. Gopalakrishnan, A., Mali, A., Kifer, D., Giles, L., Ororbia, A.G.: A neural temporal model for human motion prediction. In: CVPR. pp. 12116-12125 (2019) 1, 3 +9. Hassan, M., Ceylan, D., Villegas, R., Saito, J., Yang, J., Zhou, Y., Black, M.J.: Stochastic scene-aware motion prediction. In: ICCV. pp. 11374-11384 (2021) 4 +0. Hassan, M., Choutas, V., Tzionas, D., Black, M.J.: Resolving 3d human pose ambiguities with 3d scene constraints. In: ICCV. pp. 2282-2292 (2019) 9, 10, 11, 12, 13 +1. Hassan, M., Ghosh, P., Tesch, J., Tzionas, D., Black, M.J.: Populating 3d scenes by learning human-scene interaction. In: CVPR. pp. 14708-14718 (2021) 1, 4 +2. Ionescu, C., Papava, D., Olaru, V., Sminchisescu, C.: Human3.6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. TPAMI 36(7), 1325-1339 (2014) 10 +3. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. In: ICLR (2015) 11 +4. Kipf, T.N., Welling, M.: Semi-supervised classification with graph convolutional networks. In: ICLR (2017) 3 +5. Koppula, H.S., Saxena, A.: Anticipating human activities for reactive robotic response. In: IROS. p. 2071. Tokyo (2013) 1 +6. Li, M., Chen, S., Zhao, Y., Zhang, Y., Wang, Y., Tian, Q.: Dynamic multiscale graph neural networks for 3d skeleton based human motion prediction. In: CVPR. pp. 214-223 (2020) 1, 3 +7. Li, X., Li, H., Joo, H., Liu, Y., Sheikh, Y.: Structure from recurrent motion: From rigidity to recurrency. In: CVPR (2018) 1, 3 +8. Liu, Z., Tang, H., Lin, Y., Han, S.: Point-voxel cnn for efficient 3d deep learning. Advances in Neural Information Processing Systems 32 (2019) 14 +9. Mahmood, N., Ghorbani, N., Troje, N.F., Pons-Moll, G., Black, M.J.: Amass: Archive of motion capture as surface shapes. In: ICCV (2019) 10 +20. Mao, W., Hartley, R.I., Salzmann, M., et al.: Contact-aware human motion forecasting. Advances in Neural Information Processing Systems 35, 7356-7367 (2022), 1, 2, 3, 4, 5, 9, 10, 11, 12, 13, 14 + +21. Mao, W., Liu, M., Salzmann, M.: History repeats itself: Human motion prediction via motion attention. In: ECCV. pp. 474-489. Springer (2020) 1, 3 +22. Mao, W., Liu, M., Salzmann, M., Li, H.: Learning trajectory dependencies for human motion prediction. In: ICCV. pp. 9489-9497 (2019) 1, 3, 6, 7 +23. Paden, B., Cap, M., Yong, S.Z., Yershov, D., Frazzoli, E.: A survey of motion planning and control techniques for self-driving urban vehicles. IEEE Transactions on intelligent vehicles 1(1), 33-55 (2016) 1 +24. Paszke, A., Gross, S., Chintala, S., Chanan, G., Yang, E., DeVito, Z., Lin, Z., Desmaison, A., Antiga, L., Lerer, A.: Automatic differentiation in pytorch. In: NeurIPS-W (2017) 11 +25. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3D hands, face, and body from a single image. In: CVPR. pp. 10975-10985 (2019) 4, 5, 8, 10 +26. Prokudin, S., Lassner, C., Romero, J.: Efficient learning on point clouds with basis point sets. In: ICCV. pp. 4332-4341 (2019) 4 +27. Scofano, L., Sampieri, A., Schiele, E., De Matteis, E., Leal-Taixe, L., Galasso, F.: Staged contact-aware global human motion forecasting. In: BMVC (2023) 2, 3, 4, 10, 11, 12, 13 +28. Sidenbladh, H., Black, M.J., Sigal, L.: Implicit probabilistic models of human motion for synthesis and tracking. In: ECCV. pp. 784-800. Springer (2002) 1, 3 +29. Starke, S., Zhang, H., Komura, T., Saito, J.: Neural state machine for character-scene interactions. ACM Trans. Graph. 38(6), 209-1 (2019) 1 +30. Taylor, G.W., Hinton, G.E., Roweis, S.T.: Modeling human motion using binary latent variables. In: NeurIPS. pp. 1345-1352 (2006) 1, 3 +31. Van Welbergen, H., Van Basten, B.J., Egges, A., Ruttkay, Z.M., Overmars, M.H.: Real time animation of virtual humans: a trade-off between naturalness and control. In: Computer Graphics Forum. vol. 29, pp. 2530-2554. Wiley Online Library (2010) 1 +32. Wang, B., Adeli, E., Chiu, H.k., Huang, D.A., Niebles, J.C.: Imitation learning for human pose prediction. In: ICCV. pp. 7124-7133 (2019) 1, 3 +33. Wang, J.M., Fleet, D.J., Hertzmann, A.: Gaussian process dynamical models for human motion. TPAMI 30(2), 283-298 (2008) 1, 3 +34. Wang, J., Xu, H., Xu, J., Liu, S., Wang, X.: Synthesizing long-term 3d human motion and interaction in 3d scenes. In: CVPR. pp. 9401-9411 (2021) 1, 2, 4 +35. Wang, J., Yan, S., Dai, B., Lin, D.: Scene-aware generative network for human motion synthesis. In: CVPR. pp. 12206-12215 (2021) 4 +36. Wang, Z., Chen, Y., Liu, T., Zhu, Y., Liang, W., Huang, S.: Humanise: Language-conditioned human motion generation in 3d scenes. Advances in Neural Information Processing Systems 35, 14959-14971 (2022) 9, 10, 11, 12, 13, 14 +37. Zhang, S., Zhang, Y., Ma, Q., Black, M.J., Tang, S.: Place: Proximity learning of articulation and contact in 3d environments. In: 3DV. pp. 642-651. IEEE (2020) 1, 2, 3, 4, 14 +38. Zhang, Y., Black, M.J., Tang, S.: We are more than our joints: Predicting how 3d bodies move. In: CVPR. pp. 3372-3382 (2021) 5, 11 +39. Zhang, Y., Hassan, M., Neumann, H., Black, M.J., Tang, S.: Generating 3d people in scenes without people. In: CVPR. pp. 6194-6204 (2020) 4 +40. Zheng, Y., Yang, Y., Mo, K., Li, J., Yu, T., Liu, Y., Liu, C.K., Guibas, L.J.: Gimo: Gaze-informed human motion prediction in context. In: ECCV. pp. 676-694. Springer (2022) 1, 2, 3, 9, 10, 11, 12, 13 +41. Zhou, Y., Barnes, C., Lu, J., Yang, J., Li, H.: On the continuity of rotation representations in neural networks. In: CVPR. pp. 5745-5753 (2019) 8 \ No newline at end of file diff --git a/2024/Scene-aware Human Motion Forecasting via Mutual Distance Prediction/images.zip b/2024/Scene-aware Human Motion Forecasting via Mutual Distance Prediction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0d36cfa30197707bbdc48fa38aa7ff543bf6f0cb --- /dev/null +++ b/2024/Scene-aware Human Motion Forecasting via Mutual Distance Prediction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b4eacf56f9743dd095f8e74bdeee8bd17d454d57c946a2f78de4300ed95e4c9c +size 599749 diff --git a/2024/Scene-aware Human Motion Forecasting via Mutual Distance Prediction/layout.json b/2024/Scene-aware Human Motion Forecasting via Mutual Distance Prediction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8ed0041f1055c5f8820058da41189c4d4de05fa3 --- /dev/null +++ b/2024/Scene-aware Human Motion Forecasting via Mutual Distance Prediction/layout.json @@ -0,0 +1,8926 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 153, + 112, + 462, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 112, + 462, + 148 + ], + "spans": [ + { + "bbox": [ + 153, + 112, + 462, + 148 + ], + "type": "text", + "content": "Scene-aware Human Motion Forecasting via Mutual Distance Prediction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 201, + 167, + 411, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 167, + 411, + 181 + ], + "spans": [ + { + "bbox": [ + 201, + 167, + 411, + 181 + ], + "type": "text", + "content": "Chaoyue Xing" + }, + { + "bbox": [ + 201, + 167, + 411, + 181 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 201, + 167, + 411, + 181 + ], + "type": "text", + "content": ", Wei Mao" + }, + { + "bbox": [ + 201, + 167, + 411, + 181 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 201, + 167, + 411, + 181 + ], + "type": "text", + "content": ", and Miaomiao Liu" + }, + { + "bbox": [ + 201, + 167, + 411, + 181 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 198, + 190, + 414, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 190, + 414, + 201 + ], + "spans": [ + { + "bbox": [ + 198, + 190, + 414, + 201 + ], + "type": "text", + "content": "Australian National University, Canberra, Australia" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 228, + 455, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 228, + 455, + 481 + ], + "spans": [ + { + "bbox": [ + 160, + 228, + 455, + 481 + ], + "type": "text", + "content": "Abstract. In this paper, we address the issue of scene-aware 3D human motion forecasting. A key challenge in this task is to predict future human motions that are coherent with the scene by modeling human-scene interactions. While recent works have demonstrated that explicit constraints on human-scene interactions can prevent the occurrence of ghost motion, they only provide constraints on partial human motion e.g., the global motion of the human or a few joints contacting the scene, leaving the rest of unconstrained. To address this limitation, we propose to represent the human-scene interaction using the mutual distance between the human body and the scene. Such mutual distances constrain both the local and global human motion, resulting in a whole-body motion constrained prediction. In particular, mutual distance constraints consist of two components, the signed distance of each vertex on the human mesh to the scene surface and the distance of basis scene points to the human mesh. We further introduce a global scene representation learned from a signed distance function (SDF) volume to ensure coherence between the global scene representation and the explicit constraint from the mutual distance. We develop a pipeline with two sequential steps: predicting the future mutual distances first, followed by forecasting future human motion. We explicitly ensure consistency between predicted poses and mutual distances during training. Extensive testing on both synthetic and real datasets demonstrates that our method consistently surpasses the performance of current state-of-the-art techniques." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 491, + 453, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 491, + 453, + 514 + ], + "spans": [ + { + "bbox": [ + 159, + 491, + 453, + 514 + ], + "type": "text", + "content": "Keywords: Scene-aware " + }, + { + "bbox": [ + 159, + 491, + 453, + 514 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 491, + 453, + 514 + ], + "type": "text", + "content": " Human Motion Forecasting " + }, + { + "bbox": [ + 159, + 491, + 453, + 514 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 491, + 453, + 514 + ], + "type": "text", + "content": " Mutual Distance " + }, + { + "bbox": [ + 159, + 491, + 453, + 514 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 491, + 453, + 514 + ], + "type": "text", + "content": " SDF Volume" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 533, + 230, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 230, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 230, + 547 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 558, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 629 + ], + "type": "text", + "content": "Human motion forecasting has numerous applications, including autonomous driving [23], human-robot interaction [15], animation [31] and virtual/augmented reality [29]. Despite being an active research field for decades [2,28,30,33], most previous methods [1, 3, 8, 16, 17, 21, 22, 32] focus solely on human motion and overlook the fact that humans interact with the 3D environment. In contrast, this paper addresses the problem of scene-aware human motion prediction." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "Recent advancements in scene-aware human synthesis [11,34,37] and human motion prediction [4,5,20,40] have begun to model the human-scene interaction. Some approaches achieve this implicitly by incorporating scene encoding as an" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "content": "additional input to their model e.g., 2D scene image [4], 3D scene point cloud [34] or a specific object [5]. However, without explicit constraints, those methods tend to predict motions with artifacts e.g., \"ghost motion\". Alternatively, other works [20, 27, 37, 40] start to explore the explicit representation of human-scene interaction. However, none of their representations are able to constrain the whole-body motion both locally and globally. For example, the proximity between human and scene used in [37] mainly constrains the global position of the human, while the human gaze employed in [40] only specifies the potential target location of the human. Although the contact map proposed in [20, 27] achieves detailed constraints on both the local pose and global position of certain human joints, their contact map can only constrain joints that contact the scene, neglecting the others, such as elbow joints, which often remain unsupported." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 186, + 282, + 296, + 379 + ], + "blocks": [ + { + "bbox": [ + 186, + 282, + 296, + 379 + ], + "lines": [ + { + "bbox": [ + 186, + 282, + 296, + 379 + ], + "spans": [ + { + "bbox": [ + 186, + 282, + 296, + 379 + ], + "type": "image", + "image_path": "732e84fb3e1960611be0718822119a4755b9f3b1df65d655504dbbdd5309c682.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 170, + 380, + 444, + 392 + ], + "lines": [ + { + "bbox": [ + 170, + 380, + 444, + 392 + ], + "spans": [ + { + "bbox": [ + 170, + 380, + 444, + 392 + ], + "type": "text", + "content": "(a) The per-vertex signed distance (b) The per-basis point distance" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 323, + 283, + 432, + 378 + ], + "blocks": [ + { + "bbox": [ + 323, + 283, + 432, + 378 + ], + "lines": [ + { + "bbox": [ + 323, + 283, + 432, + 378 + ], + "spans": [ + { + "bbox": [ + 323, + 283, + 432, + 378 + ], + "type": "image", + "image_path": "2502a106d2db1d11117b5842c681cc1ca215eb8fbcc4c5f6f88b26536dd4a397.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 401, + 482, + 446 + ], + "lines": [ + { + "bbox": [ + 130, + 401, + 482, + 446 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 482, + 446 + ], + "type": "text", + "content": "Fig. 1: Our mutual distances. a) shows the per-vertex signed distance for sampled vertices on human mesh. Their color indicates the distance value. b) shows the per-basis point distance for the basis points which are not sampled on the scene surface. For both figures, the darker the color is, the smaller the distance is." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 472, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 640 + ], + "type": "text", + "content": "To constrain the whole-body motion for better prediction, in this paper, we propose a mutual distance representation which captures the distances between the human body and the scene. We further introduce the global scene representation learned from SDF volume instead of point cloud to achieve consistent global and local scene representation learning. Specifically, our proposed mutual distance representation comprises two components: the per-vertex signed distance and the per-basis point distance. Given a 3D scene, the per-vertex distance is the minimum distance from each human body vertex to the scene surface, and its sign indicates whether such vertex penetrates the scene or not (see Fig. 1 (a)). Since the change in signed distance for each body vertex represents the motion of the corresponding body part, it provides a complete constraint on the entire body. However, as a signed distance corresponds to a level-set of the scene Signed Distance Field (SDF), a human mesh at multiple possible global locations can result in the same per-vertex signed distance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "To address this, we leverage a per-basis point distance, which calculates the minimum distance between each point in a predefined set of basis points and the" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "C. Xing et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": "human surface. Our per-basis point distance is similar to BPS body encoding in PLACE [37] with differences in choices of basis points. In particular, their basis points are partially scene-dependent making it not suitable for motion prediction, while our basis points are purely scene-independent and sampled with a fixed distance relative to origin, which can better capture the motion patterns." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 480, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 480, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 480, + 307 + ], + "type": "text", + "content": "While our model is end-to-end trainable, we adopt a two-step training process followed by fine-tuning for better convergence. In the first step, we predict the mutual distance. Conditioning on the predicted distances, we then forecast the future human motion. Specifically, given a 3D scene, the past human poses, and past mutual distance, we use a strategy based on Graph Convolutional Network (GCN) [14], Discrete Cosine Transform (DCT), for the spatial and temporal encoding of the mutual distances and a simple 3D CNN for encoding of the SDF volume to predict the future ones. We subsequently take the predicted distances, the past human poses, and the scene as input to predict future human motion using an RNN-based model. A consistency prior is adopted to ensure that the predicted human motion is consistent with the input mutual distance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 308, + 481, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 481, + 392 + ], + "type": "text", + "content": "Our contributions can be summarized as follows. i) We propose an explicit representation of human-scene interaction using mutual distance to provide constraints for whole-body motion. ii) We explore different global scene representations and find the SDF volume to be more efficient than commonly used point cloud representation. Our experiments demonstrate that our approach outperforms the state-of-the-art scene-aware human motion prediction methods by a large margin on both existing real and synthesis benchmark datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 411, + 237, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 411, + 237, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 411, + 237, + 423 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 437, + 480, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 480, + 533 + ], + "type": "text", + "content": "Human motion prediction. Researchers have been studying 3D human motion modeling [2,28,30,33] for decades. While traditional methods, such as those employing Hidden Markov Models [2] or the Gaussian process latent variable model [33], are capable of handling periodic and simple non-periodic motions like walking and golf swings, more complex motions are typically modeled with deep neural networks [1,3,8,16,17,21,22,32]. However, these methods focus solely on the prediction of local poses, neglecting the global human motion and 3D scenes in which the motion takes place." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 534, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 481, + 665 + ], + "type": "text", + "content": "More recent works [4,5,20,27,40] have begun to incorporate scene context into the task of motion prediction. Specifically, Cao et al. [4] propose to represent the scene with a 2D image. The feature extracted from such an image is fed into the prediction model and lets the model learn the human-scene interaction implicitly. However, such 2D scene representation cannot handle occlusions and it does not ensure a valid human-scene interaction. In [5], a semantic graph is introduced to encode the interaction between the human and the object of interest. Although it is efficient to model a human interacting with one object, it is not suitable for human-scene interaction as the human may interact with multiple objects when he moves around. In [40], Zheng et al. employ eye gaze as additional information to guide the motion forecasting process. Although" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 153, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 153, + 91, + 447, + 102 + ], + "type": "text", + "content": "Scene-aware Human Motion Forecasting via Mutual Distance Prediction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": "eye gaze provides valuable hints about the goal location of the human motion, it constrains neither the local poses nor the human-scene interaction. Mao et al. [20] introduce a contact map that can explicitly encode the relation between scene surface points and human joints. Scofano et al. [27] propose a three-stage pipeline in a coarse-to-fine manner and utilizing the time information based on [20]. However, for forecasting, their contact map only constrains joints that contact the scene and disregard the rest. By contrast, our introduced mutual distance is able to constrain the whole-body motion." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 212, + 482, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 212, + 482, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 482, + 511 + ], + "type": "text", + "content": "Scene-aware human synthesis. One of the main challenges in this task is to ensure a valid human-scene interaction. As such, existing works have explored various strategies to model such interaction [9, 11, 34, 35, 37, 39]. Similar to [4], the interaction is modeled implicitly in [34, 35, 39] via either Generative Adversarial Networks (GANs) [34], or Variational Autoencoders [34, 39]. Besides this, other works [9, 11, 37] adopt explicit representation of human-scene interaction. In particular, it was proposed in POSA [11] to develop a human-centric representation of the human-scene interaction which assigns scene semantic labels, such as floor, and chair, to each human surface vertex. Those labels indicate the probability of such vertex contacting what kind of object in the scene. Although it performs well in the task of placing a human in a scene, it lacks sufficient detail about the local pose of the human, rendering it unsuitable for human motion forecasting. In SAMP [9], Hassan et al. propose to only model the interaction with the object at the goal location. Given the final root translation and orientation, the task is then treated as a path planning problem. Similar to [5], such a strategy cannot constrain the interactions between the human and the scene when the human moves in the scene. The closest to our work is PLACE [37] where the human-scene interaction is represented by the minimum distance between a set of 3D scene points also known as Basis Point Sets (BSP) [26] to the human body surface. However, such representation alone only provides a prior on the global position of the human. For example, one can move its arm that is far from the scene surface while the scene to human proximity remains the same. By contrast, we combine the per-basis point distance with the per-vertex signed distance resulting in a detailed constraint on the whole-body motion both locally and globally." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 528, + 212, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 528, + 212, + 541 + ], + "spans": [ + { + "bbox": [ + 132, + 528, + 212, + 541 + ], + "type": "text", + "content": "3 Approach" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 552, + 482, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 552, + 482, + 648 + ], + "spans": [ + { + "bbox": [ + 130, + 552, + 482, + 648 + ], + "type": "text", + "content": "Let us begin by formulating the task of scene-aware 3D human motion forecasting. Building on previous work [20], we represent a sequence of " + }, + { + "bbox": [ + 130, + 552, + 482, + 648 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 552, + 482, + 648 + ], + "type": "text", + "content": " past human poses as " + }, + { + "bbox": [ + 130, + 552, + 482, + 648 + ], + "type": "inline_equation", + "content": "\\mathbf{X} = [\\mathbf{x}_1, \\mathbf{x}_2, \\dots, \\mathbf{x}_T] \\in \\mathbb{R}^{M \\times T^1}" + }, + { + "bbox": [ + 130, + 552, + 482, + 648 + ], + "type": "text", + "content": ", where a human pose at time " + }, + { + "bbox": [ + 130, + 552, + 482, + 648 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 552, + 482, + 648 + ], + "type": "text", + "content": " is represented as a vector " + }, + { + "bbox": [ + 130, + 552, + 482, + 648 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t \\in \\mathbb{R}^M" + }, + { + "bbox": [ + 130, + 552, + 482, + 648 + ], + "type": "text", + "content": " of length " + }, + { + "bbox": [ + 130, + 552, + 482, + 648 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 552, + 482, + 648 + ], + "type": "text", + "content": ". Such vector could be a parametric representation of human mesh such as SMPL-X [25] pose. To represent the 3D scene, we adopt a Signed Distance Volume " + }, + { + "bbox": [ + 130, + 552, + 482, + 648 + ], + "type": "inline_equation", + "content": "\\mathbf{S} \\in \\mathbb{R}^{N \\times N \\times N}" + }, + { + "bbox": [ + 130, + 552, + 482, + 648 + ], + "type": "text", + "content": ", each entry of which stores the distance of the voxel to the nearest scene surface and its sign indicates whether it is inside or outside the scene. The representation learned from this" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "C. Xing et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 447, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 447, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 447, + 665 + ], + "type": "text", + "content": "1 Unless otherwise specified, in this paper, we use column vector convention." + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 482, + 297 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 297 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 297 + ], + "type": "image", + "image_path": "29e607422490830b415c1545159ba801100ffb8ccf80748ee9e5074c897ab307.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 304, + 482, + 361 + ], + "lines": [ + { + "bbox": [ + 130, + 304, + 482, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 482, + 361 + ], + "type": "text", + "content": "Fig. 2: Network architecture. Given the 3D scene " + }, + { + "bbox": [ + 130, + 304, + 482, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 130, + 304, + 482, + 361 + ], + "type": "text", + "content": " represented as a signed distance volume and the past motion " + }, + { + "bbox": [ + 130, + 304, + 482, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 130, + 304, + 482, + 361 + ], + "type": "text", + "content": " shown in grey meshes, our approach first predicts the future per-vertex signed distance " + }, + { + "bbox": [ + 130, + 304, + 482, + 361 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{D}}" + }, + { + "bbox": [ + 130, + 304, + 482, + 361 + ], + "type": "text", + "content": ", and the future per-basis point distance " + }, + { + "bbox": [ + 130, + 304, + 482, + 361 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{B}}" + }, + { + "bbox": [ + 130, + 304, + 482, + 361 + ], + "type": "text", + "content": " from historical distance " + }, + { + "bbox": [ + 130, + 304, + 482, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 130, + 304, + 482, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 304, + 482, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 130, + 304, + 482, + 361 + ], + "type": "text", + "content": ", respectively. The two predicted future distances are then fed into a RNN-based network to predict the future motion " + }, + { + "bbox": [ + 130, + 304, + 482, + 361 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Y}}" + }, + { + "bbox": [ + 130, + 304, + 482, + 361 + ], + "type": "text", + "content": " shown in orange meshes." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 384, + 482, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 482, + 444 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 482, + 444 + ], + "type": "text", + "content": "volume is more consistent with our mutual distance representation than those learned from point cloud [20]. Moreover, such Signed Distance Volume allows us to query the signed distance of any given position easily via interpolation, which facilitates the fast calculation of the loss defined on the SDFs of the predicted pose sequence during training (see Eq. 14)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 445, + 496, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 445, + 496, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 496, + 552 + ], + "type": "text", + "content": "Our objective is predicting " + }, + { + "bbox": [ + 130, + 445, + 496, + 552 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 130, + 445, + 496, + 552 + ], + "type": "text", + "content": " future human poses " + }, + { + "bbox": [ + 130, + 445, + 496, + 552 + ], + "type": "inline_equation", + "content": "\\mathbf{Y} = [\\mathbf{x}_{T + 1},\\mathbf{x}_{T + 2},\\dots ,\\mathbf{x}_{T + U}]\\in" + }, + { + "bbox": [ + 130, + 445, + 496, + 552 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{M\\times U}" + }, + { + "bbox": [ + 130, + 445, + 496, + 552 + ], + "type": "text", + "content": " with history motion " + }, + { + "bbox": [ + 130, + 445, + 496, + 552 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 130, + 445, + 496, + 552 + ], + "type": "text", + "content": " and 3D scene S. We introduce an innovative pipeline that predicts mutual distance representation first i.e., the signed minimum distance from each human body vertex to the scene surface (the per-vertex signed distance) and the minimum distance from each basis point to the human surface (the per-basis point distance) at each time frame. Conditioned on such mutual distance, our model then forecasts future poses. Our pipeline is illustrated in Fig. 2. We detail our mutual distance and the pipeline for predicting these distances and future motions below." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 571, + 326, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 571, + 326, + 583 + ], + "spans": [ + { + "bbox": [ + 131, + 571, + 326, + 583 + ], + "type": "text", + "content": "3.1 Mutual Distance Representation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 481, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 481, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 481, + 616 + ], + "type": "text", + "content": "Let us first introduce our mutual distance representation, which consists of the per-vertex signed distance and the per-basis point distance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 618, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 618, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 618, + 482, + 666 + ], + "type": "text", + "content": "The per-vertex signed distance. Given the human pose " + }, + { + "bbox": [ + 130, + 618, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 130, + 618, + 482, + 666 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 130, + 618, + 482, + 666 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 618, + 482, + 666 + ], + "type": "text", + "content": ", one can obtain the vertices of human mesh with a predefined mapping, such as SMPL-X [25]. However, constraining every single vertex may not be necessary due to their strong correlation. Hence, following [38], we select a subset of " + }, + { + "bbox": [ + 130, + 618, + 482, + 666 + ], + "type": "inline_equation", + "content": "K" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 153, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 153, + 91, + 447, + 102 + ], + "type": "text", + "content": "Scene-aware Human Motion Forecasting via Mutual Distance Prediction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "content": "vertices denoted as " + }, + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_t \\in \\mathbb{R}^{K \\times 3}" + }, + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "content": " from the human mesh. Given a 3D scene and " + }, + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_t" + }, + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "content": ", the per-vertex signed distance can be computed as," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 219, + 149, + 481, + 188 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 149, + 481, + 188 + ], + "spans": [ + { + "bbox": [ + 219, + 149, + 481, + 188 + ], + "type": "interline_equation", + "content": "d _ {k t} = \\left\\{ \\begin{array}{l l} - \\min _ {\\mathbf {y} \\in \\partial S} \\| \\mathbf {v} _ {t k} - \\mathbf {y} \\| _ {2} & \\text {i f} \\mathbf {v} _ {t k} \\in \\mathcal {S} \\\\ \\min _ {\\mathbf {y} \\in \\partial S} \\| \\mathbf {v} _ {t k} - \\mathbf {y} \\| _ {2} & \\text {i f} \\mathbf {v} _ {t k} \\notin \\mathcal {S}, \\end{array} \\right. \\tag {1}", + "image_path": "00b6ff8eb1911b65939ab84d75e08680ce868584c10b91957c5512f94ded3c18.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 196, + 479, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 196, + 479, + 279 + ], + "spans": [ + { + "bbox": [ + 130, + 196, + 479, + 279 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 196, + 479, + 279 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_t = [d_{1t}, d_{2t}, \\dots, d_{Kt}]^T \\in \\mathbb{R}^K" + }, + { + "bbox": [ + 130, + 196, + 479, + 279 + ], + "type": "text", + "content": " is the per-vertex signed distance at time step " + }, + { + "bbox": [ + 130, + 196, + 479, + 279 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 196, + 479, + 279 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 196, + 479, + 279 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 130, + 196, + 479, + 279 + ], + "type": "text", + "content": " represents the volume inside the actual solid scene and " + }, + { + "bbox": [ + 130, + 196, + 479, + 279 + ], + "type": "inline_equation", + "content": "\\partial \\mathcal{S}" + }, + { + "bbox": [ + 130, + 196, + 479, + 279 + ], + "type": "text", + "content": " is the surface points of such scene. The per-vertex signed distance provides information about the whole-body pose. However, the same human mesh at different global positions may result in the same per-vertex signed distance. To resolve such ambiguity, we further leverage the per-basis point distance which is described below." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 281, + 479, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 479, + 341 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 479, + 341 + ], + "type": "text", + "content": "The per-basis point distance. Given a set of basis points from the scene denoted as " + }, + { + "bbox": [ + 130, + 281, + 479, + 341 + ], + "type": "inline_equation", + "content": "\\mathbf{p} \\in \\mathbb{R}^{P \\times 3}" + }, + { + "bbox": [ + 130, + 281, + 479, + 341 + ], + "type": "text", + "content": ", the per-basis point distance is simply the minimum distance in each basis point to the human mesh. As the human moves within the scene, its motion is reflected in the continuous change of the per-basis point distance. Specifically, at time step " + }, + { + "bbox": [ + 130, + 281, + 479, + 341 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 281, + 479, + 341 + ], + "type": "text", + "content": ", it can be computed as," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 256, + 350, + 481, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 350, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 256, + 350, + 481, + 369 + ], + "type": "interline_equation", + "content": "b _ {p t} = \\min _ {\\mathbf {y} \\in \\partial \\mathcal {H} _ {t}} \\| \\mathbf {p} _ {p} - \\mathbf {y} \\| _ {2}, \\tag {2}", + "image_path": "0ac2e6ce13acac136e33aceb433366e0ae2862ca57f54e1420635136e7a5f76c.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 376, + 479, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 479, + 508 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 479, + 508 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 376, + 479, + 508 + ], + "type": "inline_equation", + "content": "\\partial \\mathcal{H}_t" + }, + { + "bbox": [ + 130, + 376, + 479, + 508 + ], + "type": "text", + "content": " is the human surface at time step " + }, + { + "bbox": [ + 130, + 376, + 479, + 508 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 376, + 479, + 508 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 376, + 479, + 508 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_t = [b_{1t}, b_{2t}, \\dots, b_{Pt}]^T \\in \\mathbb{R}^P" + }, + { + "bbox": [ + 130, + 376, + 479, + 508 + ], + "type": "text", + "content": ". Note that our basis points are not necessary on the scene surface. In practice, for different motion sequences, our basis points can be fixed relatively to the root joint of the last observed pose. Such root joint is defined as the origin of that data sample. In particular, we employ the spherical Fibonacci lattice algorithm [7] to sample " + }, + { + "bbox": [ + 130, + 376, + 479, + 508 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 130, + 376, + 479, + 508 + ], + "type": "text", + "content": " near-equidistant points from a sphere with radius " + }, + { + "bbox": [ + 130, + 376, + 479, + 508 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 130, + 376, + 479, + 508 + ], + "type": "text", + "content": ". The per-basis point distance primarily captures the global location of the human, while the per-vertex signed distance constrains both the detailed whole-body pose and global location. By combining two distances, our mutual distance representation provides explicit constraints on both the whole-body local poses and the global human motion of the human." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 509, + 487, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 487, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 487, + 568 + ], + "type": "text", + "content": "Given the 3D scene " + }, + { + "bbox": [ + 130, + 509, + 487, + 568 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 130, + 509, + 487, + 568 + ], + "type": "text", + "content": ", the past motion " + }, + { + "bbox": [ + 130, + 509, + 487, + 568 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 130, + 509, + 487, + 568 + ], + "type": "text", + "content": ", we first pre-compute the past mutual distance denoted as " + }, + { + "bbox": [ + 130, + 509, + 487, + 568 + ], + "type": "inline_equation", + "content": "\\mathbf{D} = [\\mathbf{d}_1, \\mathbf{d}_2, \\dots, \\mathbf{d}_T] \\in \\mathbb{R}^{K \\times T}" + }, + { + "bbox": [ + 130, + 509, + 487, + 568 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 509, + 487, + 568 + ], + "type": "inline_equation", + "content": "\\mathbf{B} = [\\mathbf{b}_1, \\mathbf{b}_2, \\dots, \\mathbf{b}_T] \\in \\mathbb{R}^{P \\times T}" + }, + { + "bbox": [ + 130, + 509, + 487, + 568 + ], + "type": "text", + "content": ". Our goal now is to predict the future mutual distance given the above information. In the following section, we will introduce our mutual distance prediction module." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 586, + 301, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 586, + 301, + 597 + ], + "spans": [ + { + "bbox": [ + 131, + 586, + 301, + 597 + ], + "type": "text", + "content": "3.2 Mutual Distance Prediction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 605, + 479, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 479, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 479, + 667 + ], + "type": "text", + "content": "We adopt a similar temporal and spatial encoding strategy proposed in [22], which has been proven to be effective in predicting smooth spatial-temporal data such as 3D human motion. Specifically, we use Discrete Cosine Transform (DCT) to represent a temporal sequence and Graph Convolutional Networks (GCN) to capture the spatial dependencies between different sequences." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "C. Xing et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": "With the DCT, we represent a sequence as a linear combination of pre-defined cosine bases. In particular, let us denote a " + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": "-frame sequence of signed distance of the " + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": "-th human vertex as " + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{d}}_k \\in [d_{1k}, d_{2k}, \\dots, d_{Tk}]^T \\in \\mathbb{R}^T" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": ", the " + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": "-th DCT coefficient is then," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 189, + 172, + 481, + 205 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 172, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 189, + 172, + 481, + 205 + ], + "type": "interline_equation", + "content": "h _ {k l} = \\sqrt {\\frac {2}{T}} \\sum_ {t = 1} ^ {T} d _ {t k} \\frac {1}{\\sqrt {1 + \\delta_ {l 1}}} \\cos \\left(\\frac {\\pi}{2 T} (2 t - 1) (l - 1)\\right), \\tag {3}", + "image_path": "24d928830adbb78cc1c5bfbdb3201c91580fb10ba35937ff2376c0c60b3949db.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 213, + 450, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 213, + 450, + 226 + ], + "spans": [ + { + "bbox": [ + 131, + 213, + 450, + 226 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 213, + 450, + 226 + ], + "type": "inline_equation", + "content": "l \\in \\{1, 2, \\dots, T\\}" + }, + { + "bbox": [ + 131, + 213, + 450, + 226 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 213, + 450, + 226 + ], + "type": "inline_equation", + "content": "\\delta_{ij}" + }, + { + "bbox": [ + 131, + 213, + 450, + 226 + ], + "type": "text", + "content": " denotes the Kronecker delta function, i.e.," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 261, + 234, + 481, + 268 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 234, + 481, + 268 + ], + "spans": [ + { + "bbox": [ + 261, + 234, + 481, + 268 + ], + "type": "interline_equation", + "content": "\\delta_ {i j} = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} i = j \\\\ 0 & \\text {i f} i \\neq j. \\end{array} \\right. \\tag {4}", + "image_path": "eba682eb4a820f6d5a6f8366dfa6a04e9b867f5c98755b272c861f6191ecdc52.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 275, + 479, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 275, + 479, + 297 + ], + "spans": [ + { + "bbox": [ + 131, + 275, + 479, + 297 + ], + "type": "text", + "content": "Note that, with all cosine bases precomputed, the DCT can be simply represented as a linear operation," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 279, + 298, + 481, + 312 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 298, + 481, + 312 + ], + "spans": [ + { + "bbox": [ + 279, + 298, + 481, + 312 + ], + "type": "interline_equation", + "content": "\\mathbf {h} _ {k} = \\mathbf {C} \\tilde {\\mathbf {d}} _ {k}, \\tag {5}", + "image_path": "feb3badad850c8a8501d9c4e266d96820b090ddbc0020230636f19ad1d533bd7.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 316, + 479, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 479, + 363 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 479, + 363 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 316, + 479, + 363 + ], + "type": "inline_equation", + "content": "\\mathbf{C} \\in \\mathbb{R}^{T \\times T}" + }, + { + "bbox": [ + 130, + 316, + 479, + 363 + ], + "type": "text", + "content": " is an orthogonal matrix and each row of " + }, + { + "bbox": [ + 130, + 316, + 479, + 363 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 130, + 316, + 479, + 363 + ], + "type": "text", + "content": " stores the precomputed the cosine basis. " + }, + { + "bbox": [ + 130, + 316, + 479, + 363 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_k \\in \\mathbb{R}^T" + }, + { + "bbox": [ + 130, + 316, + 479, + 363 + ], + "type": "text", + "content": " are the " + }, + { + "bbox": [ + 130, + 316, + 479, + 363 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 316, + 479, + 363 + ], + "type": "text", + "content": " DCT coefficients. Given such coefficients " + }, + { + "bbox": [ + 130, + 316, + 479, + 363 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_k" + }, + { + "bbox": [ + 130, + 316, + 479, + 363 + ], + "type": "text", + "content": ", the original sequence can be losslessly recovered via Inverse DCT (IDCT) as," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 276, + 363, + 481, + 377 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 363, + 481, + 377 + ], + "spans": [ + { + "bbox": [ + 276, + 363, + 481, + 377 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {d}} _ {k} = \\mathbf {C} ^ {T} \\mathbf {h} _ {k}. \\tag {6}", + "image_path": "9ea1f70e992a3295dcca555eabbfebc41684c79e116b64f71c77d9ed1839834a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 382, + 482, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 382, + 482, + 418 + ], + "spans": [ + { + "bbox": [ + 131, + 382, + 482, + 418 + ], + "type": "text", + "content": "Similarly, we can obtain the DCT coefficients of all " + }, + { + "bbox": [ + 131, + 382, + 482, + 418 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 131, + 382, + 482, + 418 + ], + "type": "text", + "content": " human vertices as " + }, + { + "bbox": [ + 131, + 382, + 482, + 418 + ], + "type": "inline_equation", + "content": "\\mathbf{H} = [\\mathbf{h}_1,\\mathbf{h}_2,\\dots ,\\mathbf{h}_K]^T\\in \\mathbb{R}^{K\\times T}" + }, + { + "bbox": [ + 131, + 382, + 482, + 418 + ], + "type": "text", + "content": ". For the spatial encoding, we follow [22] to use GCN. Specifically, a GCN layer is defined as," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 242, + 426, + 481, + 441 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 426, + 481, + 441 + ], + "spans": [ + { + "bbox": [ + 242, + 426, + 481, + 441 + ], + "type": "interline_equation", + "content": "\\mathbf {F} ^ {(n + 1)} = \\sigma \\left(\\mathbf {A} ^ {(n)} \\mathbf {F} ^ {(n)} \\mathbf {W} ^ {(n)}\\right), \\tag {7}", + "image_path": "6041ec3bf1ecc0c9b7137290af69a48b3e949689e6c0580c5956804bcad83e38.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 448, + 482, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 448, + 482, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 448, + 482, + 579 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 448, + 482, + 579 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 448, + 482, + 579 + ], + "type": "text", + "content": " is an activation function such as " + }, + { + "bbox": [ + 130, + 448, + 482, + 579 + ], + "type": "inline_equation", + "content": "\\tanh(\\cdot)" + }, + { + "bbox": [ + 130, + 448, + 482, + 579 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 448, + 482, + 579 + ], + "type": "inline_equation", + "content": "\\mathbf{A}^{(n)} \\in \\mathbb{R}^{K \\times K}" + }, + { + "bbox": [ + 130, + 448, + 482, + 579 + ], + "type": "text", + "content": " is the learnable adjacent matrix at the " + }, + { + "bbox": [ + 130, + 448, + 482, + 579 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 448, + 482, + 579 + ], + "type": "text", + "content": "-th layer and " + }, + { + "bbox": [ + 130, + 448, + 482, + 579 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^{(n)} \\in \\mathbb{R}^{T \\times F}" + }, + { + "bbox": [ + 130, + 448, + 482, + 579 + ], + "type": "text", + "content": " is the feature extraction parameter. " + }, + { + "bbox": [ + 130, + 448, + 482, + 579 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{(n)} \\in \\mathbb{R}^{K \\times F'}" + }, + { + "bbox": [ + 130, + 448, + 482, + 579 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 448, + 482, + 579 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{(n+1)} \\in \\mathbb{R}^{K \\times F}" + }, + { + "bbox": [ + 130, + 448, + 482, + 579 + ], + "type": "text", + "content": " are the input and the output features of the " + }, + { + "bbox": [ + 130, + 448, + 482, + 579 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 448, + 482, + 579 + ], + "type": "text", + "content": "-th GCN layer, respectively2. Same as [22], our prediction network consists of many such layers. Since our final goal is to predict the future per-vertex signed distance, we follow the strategy in [22] to first pad the sequence to make a sequence of length " + }, + { + "bbox": [ + 130, + 448, + 482, + 579 + ], + "type": "inline_equation", + "content": "T + U" + }, + { + "bbox": [ + 130, + 448, + 482, + 579 + ], + "type": "text", + "content": ". Our prediction module then aims to predict a residual between the ground truth DCT coefficients of the whole sequence and those of the padded historical sequence given the scene " + }, + { + "bbox": [ + 130, + 448, + 482, + 579 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 130, + 448, + 482, + 579 + ], + "type": "text", + "content": " and past human motion " + }, + { + "bbox": [ + 130, + 448, + 482, + 579 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 130, + 448, + 482, + 579 + ], + "type": "text", + "content": ". More formally, the future per-vertex signed distance is predicted as follows," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 228, + 580, + 481, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 580, + 481, + 594 + ], + "spans": [ + { + "bbox": [ + 228, + 580, + 481, + 594 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {D}} = \\left(\\mathcal {G} _ {d} (\\hat {\\mathbf {H}}, \\mathcal {G} _ {s} (\\mathbf {S}), \\mathcal {G} _ {x} (\\mathbf {X})) + \\hat {\\mathbf {H}}\\right) \\mathbf {C}, \\tag {8}", + "image_path": "41d84b50fe78d2cc6869d27bc943f2b0ca1c22863dcf2531b1da2c23745f9401.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_d" + }, + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_s" + }, + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_x" + }, + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "text", + "content": " are the GCN predictor, the 3D scene encoding network, and the past motion encoding model, respectively. With a little abuse of notation, here, " + }, + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\mathbf{C} \\in \\mathbb{R}^{(T + U) \\times (T + U)}" + }, + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "text", + "content": " is the DCT bases for sequences with length " + }, + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "inline_equation", + "content": "T + U" + }, + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{H}} \\in \\mathbb{R}^{K \\times (T + U)}" + }, + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "text", + "content": " is the DCT coefficients of the padded historical sequence. " + }, + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{D}} \\in" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 152, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 152, + 91, + 448, + 102 + ], + "type": "text", + "content": "Scene-aware Human Motion Forecasting via Mutual Distance Prediction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 478, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 478, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 478, + 666 + ], + "type": "text", + "content": "2 The first GCN layer will take the DCT coefficient matrix as input thus, " + }, + { + "bbox": [ + 133, + 653, + 478, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{(1)} = \\mathbf{H}" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 115, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 115, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{K \\times (T + U)}" + }, + { + "bbox": [ + 132, + 115, + 479, + 140 + ], + "type": "text", + "content": " is the predicted per-vertex signed distance. Note that, here our model also tries to recover the past sequence." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 479, + 163 + ], + "type": "text", + "content": "Similarly, we have another branch to predict the future per-basis point distance as," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 231, + 163, + 480, + 176 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 163, + 480, + 176 + ], + "spans": [ + { + "bbox": [ + 231, + 163, + 480, + 176 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {B}} = \\left(\\mathcal {G} _ {b} (\\hat {\\mathbf {K}}, \\mathcal {G} _ {s} (\\mathbf {S}), \\mathcal {G} _ {x} (\\mathbf {X})) + \\hat {\\mathbf {K}}\\right) \\mathbf {C}, \\tag {9}", + "image_path": "186fc1193c32f872867e05d6421c568b06737609b430069083146df51ff4964a.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 182, + 479, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 479, + 206 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 479, + 206 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 182, + 479, + 206 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_b" + }, + { + "bbox": [ + 132, + 182, + 479, + 206 + ], + "type": "text", + "content": " is also a GCN. " + }, + { + "bbox": [ + 132, + 182, + 479, + 206 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{K}}" + }, + { + "bbox": [ + 132, + 182, + 479, + 206 + ], + "type": "text", + "content": " is the DCT coefficients of padded past per-basis point distance sequence. " + }, + { + "bbox": [ + 132, + 182, + 479, + 206 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{B}}\\in \\mathbb{R}^{P\\times (T + U)}" + }, + { + "bbox": [ + 132, + 182, + 479, + 206 + ], + "type": "text", + "content": " is the predicted per-basis point distance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 207, + 479, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 207, + 479, + 230 + ], + "spans": [ + { + "bbox": [ + 132, + 207, + 479, + 230 + ], + "type": "text", + "content": "During training, we use the average " + }, + { + "bbox": [ + 132, + 207, + 479, + 230 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 132, + 207, + 479, + 230 + ], + "type": "text", + "content": " loss between the ground-truth distances and the predicted ones. Formally, this loss is defined as," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 159, + 239, + 479, + 271 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 239, + 479, + 271 + ], + "spans": [ + { + "bbox": [ + 159, + 239, + 479, + 271 + ], + "type": "interline_equation", + "content": "\\ell_ {\\text {d i s t}} = \\frac {1}{(T + U)} \\left(\\frac {1}{K} \\sum_ {t = 1} ^ {T + U} \\sum_ {k = 1} ^ {K} | \\hat {d} _ {k t} - d _ {k t} | + \\frac {1}{P} \\sum_ {t = 1} ^ {T + U} \\sum_ {p = 1} ^ {P} | \\hat {b} _ {p t} - b _ {p t} |\\right), \\tag {10}", + "image_path": "d13e8aa30cd5127de89a461edfae9be2a7c5c8b4d4333af8699d6f9dc67a2948.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 282, + 479, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 479, + 308 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 479, + 308 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 282, + 479, + 308 + ], + "type": "inline_equation", + "content": "\\hat{d}_{kt}" + }, + { + "bbox": [ + 132, + 282, + 479, + 308 + ], + "type": "text", + "content": " is predicted signed distance for " + }, + { + "bbox": [ + 132, + 282, + 479, + 308 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 132, + 282, + 479, + 308 + ], + "type": "text", + "content": "-th vertex at time " + }, + { + "bbox": [ + 132, + 282, + 479, + 308 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 132, + 282, + 479, + 308 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 282, + 479, + 308 + ], + "type": "inline_equation", + "content": "\\hat{b}_{pt}" + }, + { + "bbox": [ + 132, + 282, + 479, + 308 + ], + "type": "text", + "content": " is predicted distance for " + }, + { + "bbox": [ + 132, + 282, + 479, + 308 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 132, + 282, + 479, + 308 + ], + "type": "text", + "content": "-th basis point at time " + }, + { + "bbox": [ + 132, + 282, + 479, + 308 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 132, + 282, + 479, + 308 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 282, + 479, + 308 + ], + "type": "inline_equation", + "content": "d_{kt}" + }, + { + "bbox": [ + 132, + 282, + 479, + 308 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 282, + 479, + 308 + ], + "type": "inline_equation", + "content": "b_{pt}" + }, + { + "bbox": [ + 132, + 282, + 479, + 308 + ], + "type": "text", + "content": " define the ground truth." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 325, + 304, + 336 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 304, + 336 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 304, + 336 + ], + "type": "text", + "content": "3.3 Motion Forecasting Network" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 344, + 479, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 344, + 479, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 344, + 479, + 392 + ], + "type": "text", + "content": "Given the future mutual distance " + }, + { + "bbox": [ + 132, + 344, + 479, + 392 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{D}}" + }, + { + "bbox": [ + 132, + 344, + 479, + 392 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 344, + 479, + 392 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{B}}" + }, + { + "bbox": [ + 132, + 344, + 479, + 392 + ], + "type": "text", + "content": ", the past human motion " + }, + { + "bbox": [ + 132, + 344, + 479, + 392 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 132, + 344, + 479, + 392 + ], + "type": "text", + "content": ", and the scene " + }, + { + "bbox": [ + 132, + 344, + 479, + 392 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 132, + 344, + 479, + 392 + ], + "type": "text", + "content": ", our goal now is to predict the future human poses " + }, + { + "bbox": [ + 132, + 344, + 479, + 392 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Y}}" + }, + { + "bbox": [ + 132, + 344, + 479, + 392 + ], + "type": "text", + "content": ". To this end, we use an RNN-based motion prediction model " + }, + { + "bbox": [ + 132, + 344, + 479, + 392 + ], + "type": "inline_equation", + "content": "(\\mathcal{G}_y)" + }, + { + "bbox": [ + 132, + 344, + 479, + 392 + ], + "type": "text", + "content": " which predicts the future motion in an auto-regressive manner. More formally, it is defined as," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 202, + 401, + 479, + 414 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 401, + 479, + 414 + ], + "spans": [ + { + "bbox": [ + 202, + 401, + 479, + 414 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {x}} _ {T + u} = \\mathcal {G} _ {y} \\left(\\hat {\\mathbf {x}} _ {T + u - 1}, \\hat {\\mathbf {d}} _ {T + u}, \\hat {\\mathbf {b}} _ {T + u}, \\mathcal {G} _ {x} (\\mathbf {X}), \\mathcal {G} _ {s} (\\mathbf {S})\\right), \\tag {11}", + "image_path": "fb83793179129550526505ae3544e6a7e12a608399e19529849421fd97d70120.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 424, + 474, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 474, + 436 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 474, + 436 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 424, + 474, + 436 + ], + "type": "inline_equation", + "content": "u\\in \\{1,2,\\dots ,U\\}" + }, + { + "bbox": [ + 132, + 424, + 474, + 436 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_{T + u}\\in \\mathbb{R}^{M}" + }, + { + "bbox": [ + 132, + 424, + 474, + 436 + ], + "type": "text", + "content": " is the predicted " + }, + { + "bbox": [ + 132, + 424, + 474, + 436 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 132, + 424, + 474, + 436 + ], + "type": "text", + "content": " -th future human pose." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 437, + 479, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 437, + 479, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 437, + 479, + 483 + ], + "type": "text", + "content": "During training, in addition to the human pose loss, we incorporate two extra losses to ensure the predicted future pose to be consistent with the ground truth mutual distance. For human pose loss, we employ global translation and orientation loss and the local pose loss. They are defined as follows," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 206, + 493, + 479, + 523 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 493, + 479, + 523 + ], + "spans": [ + { + "bbox": [ + 206, + 493, + 479, + 523 + ], + "type": "interline_equation", + "content": "\\ell_ {\\text {g l o b a l}} = \\frac {1}{U} \\sum_ {u = 1} ^ {U} \\left| \\hat {\\mathbf {t}} _ {T + u} - \\mathbf {t} _ {T + u} \\right| + \\left| \\hat {\\mathbf {r}} _ {T + u} - \\mathbf {r} _ {T + u} \\right| \\tag {12}", + "image_path": "4581a3ec0e3097f3211ffaf42ef1bf8337d559cfbd138934277ffd6bfb668de0.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 211, + 528, + 479, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 528, + 479, + 559 + ], + "spans": [ + { + "bbox": [ + 211, + 528, + 479, + 559 + ], + "type": "interline_equation", + "content": "\\ell_ {\\text {l o c a l}} = \\frac {1}{U} \\sum_ {u = 1} ^ {U} \\left| \\hat {\\mathbf {x}} _ {T + u} ^ {\\text {l o c a l}} - \\mathbf {x} _ {T + u} ^ {\\text {l o c a l}} \\right|, \\tag {13}", + "image_path": "013ae0549b98bb81cbfcb8425b47b67132bb0a8175acbdcbd7066ede580ebe89.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 569, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 569, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 569, + 480, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 569, + 480, + 665 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{t}}_{T + u} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 132, + 569, + 480, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 569, + 480, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_{T + u} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 132, + 569, + 480, + 665 + ], + "type": "text", + "content": " are the predicted and ground truth global translation, respectively. " + }, + { + "bbox": [ + 132, + 569, + 480, + 665 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{r}}_{T + u} \\in \\mathbb{SO}(3)" + }, + { + "bbox": [ + 132, + 569, + 480, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 569, + 480, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{T + u} \\in \\mathbb{SO}(3)" + }, + { + "bbox": [ + 132, + 569, + 480, + 665 + ], + "type": "text", + "content": " are the predicted and ground truth global orientation, respectively. Here, we adopt the 6-D rotation representation [41]. " + }, + { + "bbox": [ + 132, + 569, + 480, + 665 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_{T + u}^{\\mathrm{local}} \\in \\mathbb{R}^{M - 9}" + }, + { + "bbox": [ + 132, + 569, + 480, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 569, + 480, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{T + u}^{\\mathrm{local}} \\in \\mathbb{R}^{M - 9}" + }, + { + "bbox": [ + 132, + 569, + 480, + 665 + ], + "type": "text", + "content": " are the predicted and ground truth local human pose. To compute the two consistency losses, we first use the pre-defined mapping function such as SMPL-X [25] to obtain the sampled human surface vertices " + }, + { + "bbox": [ + 132, + 569, + 480, + 665 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{v}}_{T + u} \\in \\mathbb{R}^{K \\times 3}" + }, + { + "bbox": [ + 132, + 569, + 480, + 665 + ], + "type": "text", + "content": " and the set of all human surface points " + }, + { + "bbox": [ + 132, + 569, + 480, + 665 + ], + "type": "inline_equation", + "content": "\\partial \\hat{\\mathcal{H}}_{T + u}" + }, + { + "bbox": [ + 132, + 569, + 480, + 665 + ], + "type": "text", + "content": ". Recall that we represent the 3D scene as a signed distance" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "C. Xing et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 479, + 152 + ], + "type": "text", + "content": "volume. We can query the signed distance of the human surface vertices by interpolating within this volume. The per-vertex signed distance consistency loss is then defined as," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 216, + 162, + 481, + 194 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 162, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 216, + 162, + 481, + 194 + ], + "type": "interline_equation", + "content": "\\ell_ {\\text {v e r t e x}} = \\frac {1}{U K} \\sum_ {u = 1} ^ {U} \\sum_ {k = 1} ^ {K} | \\tilde {d} _ {k (T + u)} - d _ {k (T + u)} |, \\tag {14}", + "image_path": "88ebba41d92a56243736eb936cc3d5c2dd3b1be3441bb9fe74ddf02e4df33cb1.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 204, + 479, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 204, + 479, + 228 + ], + "spans": [ + { + "bbox": [ + 131, + 204, + 479, + 228 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 204, + 479, + 228 + ], + "type": "inline_equation", + "content": "\\tilde{d}_{k(T + u)}" + }, + { + "bbox": [ + 131, + 204, + 479, + 228 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 204, + 479, + 228 + ], + "type": "inline_equation", + "content": "d_{k(T + u)}" + }, + { + "bbox": [ + 131, + 204, + 479, + 228 + ], + "type": "text", + "content": " are the signed distance of the " + }, + { + "bbox": [ + 131, + 204, + 479, + 228 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 131, + 204, + 479, + 228 + ], + "type": "text", + "content": "-th vertex from the " + }, + { + "bbox": [ + 131, + 204, + 479, + 228 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 131, + 204, + 479, + 228 + ], + "type": "text", + "content": "-th predicted human mesh and ground truth human mesh, respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 229, + 480, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 229, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 131, + 229, + 480, + 266 + ], + "type": "text", + "content": "Similarly, given the set of human surface points " + }, + { + "bbox": [ + 131, + 229, + 480, + 266 + ], + "type": "inline_equation", + "content": "\\partial \\tilde{\\mathcal{H}}_{T + u}" + }, + { + "bbox": [ + 131, + 229, + 480, + 266 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 131, + 229, + 480, + 266 + ], + "type": "inline_equation", + "content": "T + u" + }, + { + "bbox": [ + 131, + 229, + 480, + 266 + ], + "type": "text", + "content": ", we can compute the per-basis point distance " + }, + { + "bbox": [ + 131, + 229, + 480, + 266 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{b}}_{T + u} \\in \\mathbb{R}^P" + }, + { + "bbox": [ + 131, + 229, + 480, + 266 + ], + "type": "text", + "content": " using Equation 2. The per-basis point distance consistency loss can then be computed as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 219, + 274, + 481, + 308 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 274, + 481, + 308 + ], + "spans": [ + { + "bbox": [ + 219, + 274, + 481, + 308 + ], + "type": "interline_equation", + "content": "\\ell_ {\\text {b a s i s}} = \\frac {1}{U P} \\sum_ {u = 1} ^ {U} \\sum_ {p = 1} ^ {P} | \\tilde {b} _ {p (T + u)} - b _ {p (T + u)} |. \\tag {15}", + "image_path": "87dacd6944097153b2cc9e617176b9c553e342d58a3dd3d2f07910b9d384f0bf.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 316, + 479, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 316, + 479, + 340 + ], + "spans": [ + { + "bbox": [ + 131, + 316, + 479, + 340 + ], + "type": "text", + "content": "The final training loss for the motion forecasting network is the combination of all the above losses," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 196, + 351, + 480, + 363 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 351, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 196, + 351, + 480, + 363 + ], + "type": "interline_equation", + "content": "\\ell_ {\\text {m o t i o n}} = \\lambda_ {1} \\ell_ {\\text {g l o b a l}} + \\lambda_ {2} \\ell_ {\\text {l o c a l}} + \\lambda_ {3} \\ell_ {\\text {v e r t e x}} + \\lambda_ {4} \\ell_ {\\text {b a s i s}}, \\tag {16}", + "image_path": "f1e6a43358bf9f88435e5f41e58b56b3c20df08a8225aa017f4016cd981f169b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 372, + 373, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 372, + 373, + 384 + ], + "spans": [ + { + "bbox": [ + 131, + 372, + 373, + 384 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 372, + 373, + 384 + ], + "type": "inline_equation", + "content": "\\lambda_{i}" + }, + { + "bbox": [ + 131, + 372, + 373, + 384 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 131, + 372, + 373, + 384 + ], + "type": "inline_equation", + "content": "i\\in \\{1,2,3,4\\}" + }, + { + "bbox": [ + 131, + 372, + 373, + 384 + ], + "type": "text", + "content": " are the balancing weights." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 385, + 481, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 385, + 481, + 504 + ], + "spans": [ + { + "bbox": [ + 131, + 385, + 481, + 504 + ], + "type": "text", + "content": "Here we adopt a stage-wise training strategy where we initially train the mutual distance prediction network and motion forecasting network separately. While our network is differentiable and can be trained from scratch in an end-to-end fashion, we observed that at early stage of the training process, the learning process can easily converge to a local minimum due to inaccurate prediction of mutual distance which serves as the input for the motion forecasting network. We thus propose to perform stage-wise training first. Note that, for stage-wise training process, the motion forecasting network is provided with the ground truth mutual distance as input. Subsequently, we combine both networks and fine-tune them in an end-to-end manner with both losses defined in Eq. 10 and 16." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 523, + 230, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 230, + 537 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 230, + 537 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 549, + 204, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 549, + 204, + 560 + ], + "spans": [ + { + "bbox": [ + 132, + 549, + 204, + 560 + ], + "type": "text", + "content": "4.1 Datasets" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 569, + 492, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 569, + 492, + 593 + ], + "spans": [ + { + "bbox": [ + 131, + 569, + 492, + 593 + ], + "type": "text", + "content": "We evaluate our method on four datasets, HUMANISE [36], GTA-IM [4], PROX [10] and GIMO [40]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 594, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 594, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 594, + 481, + 665 + ], + "type": "text", + "content": "GTA-IM. The GTA Indoor Motion dataset [4], building upon GTA game engine, is a synthetic human-scene interaction dataset. It comprises approximately 1 million RGB-D frames featuring 3D human poses for 50 distinct characters across 7 diverse scenes. To reconstruct the mesh, we fuse the depth maps of each scene and subsequently generate the SDF volume. We follow ContactAware [20] to use the same training-test split and 21 joints. The videos are 30 FPS. Both our" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 153, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 153, + 91, + 448, + 102 + ], + "type": "text", + "content": "Scene-aware Human Motion Forecasting via Mutual Distance Prediction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "model and baseline models are trained to observe the past 30 frames to predict the future 60 frames following [20]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 480, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 480, + 223 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 480, + 223 + ], + "type": "text", + "content": "PROX. Proximal Relationships with Object eXclusion (PROX) [10] is a real-world dataset. This dataset contains human SMPL-X parameters [25] fitted from RGB-D frames and 20 distinct characters interacting with various scenes. Since the original motion sequences are jittery, we use the refined version provided by [20]. We adopt the same data split and the 22 joints as in [20]. The frame rate is " + }, + { + "bbox": [ + 130, + 140, + 480, + 223 + ], + "type": "inline_equation", + "content": "30\\mathrm{Hz}" + }, + { + "bbox": [ + 130, + 140, + 480, + 223 + ], + "type": "text", + "content": ". Our model and baseline models input the past 30 frames to forecast the future 60 frames following [20]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 224, + 481, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 481, + 414 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 481, + 414 + ], + "type": "text", + "content": "HUMANISE. The HUMANISE dataset [36] is a synthetic human-scene interaction dataset. It is mainly built upon two primary datasets: the AMASS dataset [19] and the ScanNet dataset [6]. HUMANISE combines these datasets by incorporating different global motions to meet contact and collision constraints. Since the original split includes duplicated human motions in the training and test set which is not suitable for motion prediction, we create a new split to ensure that motions in the test set are never observed during training. Moreover, we ensure that the scenes paired with the motions in the test set are categorized into seen scenes, which include a set of scenes observed during the training, and unseen scenes consisting of scenes never observed during training. Our final split comprises approximately 20k sub-sequences for training and 6k for testing, which includes 5k sub-sequences from seen scenes and 1k from unseen scenes. The videos run at 30 frame-per-second (FPS). Both our model and baseline models are trained to observe past 15 frames (0.5 sec.) and predict future 30 frames (1 sec.) because the average length of motion sequences in the dataset is 2.2 sec. which is much shorter than 3 sec." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 415, + 482, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 482, + 522 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 482, + 522 + ], + "type": "text", + "content": "GIMO. The Gaze-Informed Human Motion dataset (GIMO) [40] is a real human-scene interaction dataset. The dataset captures human motion using an IMU-based motion capture system, and the 3D scene is scanned using a smartphone equipped with lidar sensors. Same as [36], GIMO also adopts SMPL-X [25] representation. It captures 11 subjects performing various actions e.g., open window, lie in bed, in 19 scenes, resulting in 127 long motion sequences with " + }, + { + "bbox": [ + 130, + 415, + 482, + 522 + ], + "type": "inline_equation", + "content": "192\\mathrm{k}" + }, + { + "bbox": [ + 130, + 415, + 482, + 522 + ], + "type": "text", + "content": " frames. We follow the official training and test split, resulting in around 6k training sequences and 1k testing sequences. Same as HUMANISE [36], frame rate is " + }, + { + "bbox": [ + 130, + 415, + 482, + 522 + ], + "type": "inline_equation", + "content": "30\\mathrm{Hz}" + }, + { + "bbox": [ + 130, + 415, + 482, + 522 + ], + "type": "text", + "content": ". Our model and baseline models predict future 30 frames with 15 past frames." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 539, + 348, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 539, + 348, + 551 + ], + "spans": [ + { + "bbox": [ + 131, + 539, + 348, + 551 + ], + "type": "text", + "content": "4.2 Metrics,Baselines & Implementation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 557, + 480, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 480, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 480, + 605 + ], + "type": "text", + "content": "Metrics. We follow [20] to evaluate our method using path error and pose error. The path error is defined as the " + }, + { + "bbox": [ + 130, + 557, + 480, + 605 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 130, + 557, + 480, + 605 + ], + "type": "text", + "content": " distance between the predicted and the ground truth translation; the pose error is the Mean Per Joint Position Error (MPJPE) [12] of local poses." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": "Baselines. We compare our method with three existing scene-aware human motion prediction models: ContAware [20], GIMO [40] and STAG [27]. ContAware [20] and STAG [27] are state-of-the-art methods which first predict the contact map and then the future motion conditioning on the predicted contact map. We train their model using the official implementation on our datasets," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "C. Xing et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 187 + ], + "type": "text", + "content": "following their original settings to predict only human joints instead of SMPL-X poses. GIMO [40] is another state-of-the-art scene-aware human motion prediction method which mainly focuses on predicting long-term future poses given the past gaze information, and the frame rate of their motion is " + }, + { + "bbox": [ + 130, + 115, + 482, + 187 + ], + "type": "inline_equation", + "content": "2\\mathrm{Hz}" + }, + { + "bbox": [ + 130, + 115, + 482, + 187 + ], + "type": "text", + "content": ". The original model of GIMO predicts future 10 frames (5 seconds) given the past 6 frames (3 seconds). We slightly adapt their official code to our settings for evaluation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 482, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 482, + 344 + ], + "type": "text", + "content": "Implementation details. Our models are developed using Pytorch [24] and optimized with the ADAM [13] optimizer. Initially, we train the mutual distance prediction model and the motion forecasting model separately for 40 epochs each. Subsequently, we fine-tune both models together for an additional epoch. The original learning rate is set to 0.0005 and decays by half at the 30-th epoch for both models. We train our model on a single RTX4090. For all datasets, the loss weights of the motion forecasting model " + }, + { + "bbox": [ + 130, + 188, + 482, + 344 + ], + "type": "inline_equation", + "content": "(\\lambda_1,\\lambda_2,\\lambda_3,\\lambda_4)" + }, + { + "bbox": [ + 130, + 188, + 482, + 344 + ], + "type": "text", + "content": " are 1.0, 0.5, 1.0, 1.0. For each motion sequence, we crop the 3D scene to a region that is within 2.0 meters of the root joint of the last observed pose, and the root joint is used as the origin of the cropped scene. To compute the per-vertex signed distance, we follow [38] to use 67 vertices on the human body. For the basis point set, we select 150 points that are 2.0 meters away from the origin of the scene, and those points are fixed for different sequences." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 134, + 362, + 480, + 552 + ], + "blocks": [ + { + "bbox": [ + 134, + 362, + 480, + 552 + ], + "lines": [ + { + "bbox": [ + 134, + 362, + 480, + 552 + ], + "spans": [ + { + "bbox": [ + 134, + 362, + 480, + 552 + ], + "type": "table", + "html": "
datasetmethodPath Error (mm)Pose Error (mm)
0.5s1.0s1.5s2.0smean0.5s1.0s1.5s2.0smean
GTA-IMContAware [20]†44.582.6125.6182.987.140.154.165.277.251.8
GIMO [40]*52.797.8160.6241.7110.347.960.771.182.759.9
STAG [27]†43.279.8119.9176.483.435.448.759.873.547.0
Ours34.465.9104.0155.672.031.046.858.970.744.6
ContAware w/GT34.546.049.767.641.736.344.348.753.941.3
STAG w/GT32.743.349.468.140.031.037.241.548.335.1
Ours w/GT15.818.920.523.022.420.423.825.828.217.6
PROXContAware [20]93.3187.2284.4381.2192.289.9127.5149.3167.5116.8
GIMO [40]*129.5233.2342.1455.0240.4112.6144.6168.0187.5140.7
STAG [27]†95.9186.4276.6369.8189.698.4137.8163.7185.7129.4
Ours80.9155.2238.6322.1162.683.9119.3142.0159.7110.7
ContAware w/GT73.9106.7104.6117.488.083.7112.9125.2132.9101.1
STAG w/GT64.081.693.4115.477.288.6106.0112.6125.498.3
Ours w/GT44.256.763.769.952.656.469.476.883.464.7
", + "image_path": "4953412e93295536f0ff92cff47ecf1fa4fd69723fe0e185c556564ae94bc350.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 552, + 482, + 575 + ], + "lines": [ + { + "bbox": [ + 132, + 552, + 482, + 575 + ], + "spans": [ + { + "bbox": [ + 132, + 552, + 482, + 575 + ], + "type": "text", + "content": "Table 1: Quantitative results on GTA-IM [4] and PROX [10]. Our model outperforms all baselines at all metrics, especially for the path error." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 132, + 623, + 198, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 623, + 198, + 634 + ], + "spans": [ + { + "bbox": [ + 132, + 623, + 198, + 634 + ], + "type": "text", + "content": "4.3 Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 641, + 491, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 491, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 491, + 666 + ], + "type": "text", + "content": "Quantitative results. We provide quantitative comparisons across three datasets: GTA-IM [4], PROX [10] and HUMANISE [36] presented in Tables 1 and 2. No" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 153, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 153, + 91, + 447, + 102 + ], + "type": "text", + "content": "Scene-aware Human Motion Forecasting via Mutual Distance Prediction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 114, + 482, + 236 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 482, + 236 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 482, + 236 + ], + "type": "table", + "html": "
MethodSeen ScenesUnseen Scenes
PathError (mm)PoseError (mm)PathError (mm)PoseError (mm)
0.5s1.0smean0.5s1.0smean0.5s1.0smean0.5s1.0smean
ContAware [20]†52.8121.157.897.6141.492.953.1124.058.794.0139.190.3
GIMO [40]*70.1129.272.0141.4150.3140.277.7144.080.2146.3159.4146.5
STAG [27]†55.2124.660.788.3131.483.057.0131.563.289.9137.785.6
Ours41.593.545.683.7130.980.046.7100.250.184.3131.880.6
ContAware w/GT38.060.835.793.4119.186.438.261.836.490.3122.084.2
STAG w/GT37.563.236.483.5117.677.337.762.236.685.1122.779.4
Ours w/GT22.933.422.080.3122.076.324.835.523.680.2121.576.2
", + "image_path": "78d246db9c0b2665649f985cdc0da7504686b421cd38e126bcf932861a03a3b6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 135, + 284, + 482, + 427 + ], + "blocks": [ + { + "bbox": [ + 132, + 237, + 481, + 259 + ], + "lines": [ + { + "bbox": [ + 132, + 237, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 481, + 259 + ], + "type": "text", + "content": "Table 2: Quantitative results on HUMANISE [36]. Our model outperforms all baselines on both seen scenes and unseen scenes." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 284, + 482, + 427 + ], + "lines": [ + { + "bbox": [ + 135, + 284, + 482, + 427 + ], + "spans": [ + { + "bbox": [ + 135, + 284, + 482, + 427 + ], + "type": "table", + "html": "
MethodSeen ScenesUnseen Scenes
PathError (mm)PoseError (mm)PathError (mm)PoseError (mm)
0.5s1.0smean0.5s1.0smean0.5s1.0smean0.5s1.0smean
Motion only53.4111.957.0102.0140.093.559.2119.062.0103.2145.395.5
Scene only48.4105.852.694.6132.987.654.3113.857.994.1136.988.3
Ours w/o D42.495.246.288.2134.683.447.9103.151.290.4140.186.0
Ours w/o B47.6101.350.685.7133.981.752.8109.155.886.4138.383.2
Ours w/ P45.297.249.390.8137.286.049.0101.752.789.8138.986.1
Ours w/ O50.1115.355.392.2141.286.755.0115.258.491.0141.386.2
Ours w/ BPS50.2108.953.9100.0153.694.552.0104.154.394.4148.290.3
Ours e2e47.6105.251.590.4141.086.252.1112.456.189.5142.085.7
Ours41.593.545.683.7130.980.046.7100.250.184.3131.880.6
", + "image_path": "3b0ee6bcc1627804c8560d40ae144e7590e798918c97d3ea05ab8a716ded0bc1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 429, + 482, + 473 + ], + "lines": [ + { + "bbox": [ + 132, + 429, + 482, + 473 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 482, + 473 + ], + "type": "text", + "content": "Table 3: Ablation study on HUMANISE [36]. We show the ablation results of our model. D, B, P, O, BPS and " + }, + { + "bbox": [ + 132, + 429, + 482, + 473 + ], + "type": "inline_equation", + "content": "e2e" + }, + { + "bbox": [ + 132, + 429, + 482, + 473 + ], + "type": "text", + "content": " indicate per-vertex signed distance, per-basis point distance, point cloud representation, occupancy representation, BPS encoding and end-to-end training respectively." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "text", + "content": "tably, these datasets lack gaze information, with the exception of GIMO [40], for which we report results without gaze, indicated by " + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "inline_equation", + "content": "*" + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "text", + "content": ". Additionally, we have adapted our training strategy to baseline models for a fair comparison. We report the best performance for each model using our training strategy (denoted by " + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "text", + "content": "). Our method consistently outperforms the baseline approaches on all datasets and at all time steps by a substantial margin, both for path and pose predictions. Particularly noteworthy is our method's superiority in path prediction over baseline models on the real-world PROX [10] dataset. Our results demonstrate that our model with ground truth mutual distance significantly outperforms baseline models with ground truth contact (w/GT). This superior performance is evident in both global path and local pose metrics, indicating that our method provides a more effective constraint for human motion. More results on the GIMO [40] dataset can be found in the supplementary material." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "C. Xing et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 495, + 306 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 495, + 306 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 495, + 306 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 495, + 306 + ], + "type": "image", + "image_path": "57085e4aba0712c3ad0c59f9ac22f593ec4dd85ec138cea4c3ac27e9d48f4e31.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 316, + 482, + 351 + ], + "lines": [ + { + "bbox": [ + 130, + 316, + 482, + 351 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 482, + 351 + ], + "type": "text", + "content": "Fig. 3: This figure compares our method with baseline models on GTA-IM [4] (top row), PROX [10] (middle row), and HUMANISE [36] (bottom row). Our method predicts future motion closer to the ground truth." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 132, + 371, + 482, + 439 + ], + "blocks": [ + { + "bbox": [ + 132, + 371, + 482, + 439 + ], + "lines": [ + { + "bbox": [ + 132, + 371, + 482, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 371, + 482, + 439 + ], + "type": "image", + "image_path": "eaed32a321b1b18005b390bb959351aec0903028d49200fcc905b9dedc29fbcb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 449, + 482, + 548 + ], + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 548 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 548 + ], + "type": "text", + "content": "Fig. 4: Ablation of the mutual distance. The three figures in the first three columns (from left to right) depict the ground truth pose for the last frame of future motion, results of our full model, and predictions of our model without mutual distance constraint. The sub-figure in gray is the last observed frame. Other sub-figures depict the predicted middle frame. The blue dot is the scene basis point, and the red dot is the sampled vertex on human mesh. The two graphs in the last two columns show the predicted per-vertex signed distance and the per-basis point distance for the red and blue point, respectively. As shown in the figures, with predicted mutual distance, we can forecast the 'stand-up' action more accurately." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "Qualitative results. We present qualitative comparisons with baseline methods on GTA-IM [4], HUMANISE [36], and PROX [10], in Fig. 3. The comparisons further demonstrate that our approach excels in predicting future human motions, aligning closely with ground truth data. In contrast, GIMO [40] exhibits issues such as motion freezing and scene penetration. Additionally, we show the predicted human skeletons generated by ContAware [20] and STAG [27], which struggle to handle complex motions." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 153, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 153, + 91, + 447, + 102 + ], + "type": "text", + "content": "Scene-aware Human Motion Forecasting via Mutual Distance Prediction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 380 + ], + "type": "text", + "content": "Ablation study. To gain a more comprehensive understanding, we present ablation results on HUMANISE [36] dataset in Table 3. We report results for our base models which are defined as our model without scene and mutual distance ('Motion only') and our model without mutual distance ('Scene only'). Notably, there are significant improvements in both path and pose prediction when we incorporate mutual distance into our model. Additionally, we assess the impact of two key components: the proposed per-vertex signed distance ('Ours w/o " + }, + { + "bbox": [ + 130, + 116, + 482, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{B}^{\\prime}" + }, + { + "bbox": [ + 130, + 116, + 482, + 380 + ], + "type": "text", + "content": ") and the per-basis point distance ('Ours w/o " + }, + { + "bbox": [ + 130, + 116, + 482, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{D}^{\\prime}" + }, + { + "bbox": [ + 130, + 116, + 482, + 380 + ], + "type": "text", + "content": "). Each distance individually contributes to the improvement for both path and pose prediction, with the most significant gains achieved when both distances are utilized. Furthermore, we examine the influence of the global scene representation: point cloud representation ('Ours w/ " + }, + { + "bbox": [ + 130, + 116, + 482, + 380 + ], + "type": "inline_equation", + "content": "P^{\\prime}" + }, + { + "bbox": [ + 130, + 116, + 482, + 380 + ], + "type": "text", + "content": ") and occupancy representation ('Ours w/ " + }, + { + "bbox": [ + 130, + 116, + 482, + 380 + ], + "type": "inline_equation", + "content": "O^{\\prime}" + }, + { + "bbox": [ + 130, + 116, + 482, + 380 + ], + "type": "text", + "content": "). For point cloud representation, similar to ContAware [20], we modified the official PVCNN [18] code to take point cloud as scene input instead of SDF volume. For occupancy representation, our method uses occupancy volume as scene input. Our model outperforms both the point cloud and occupancy representation across all metrics. We also compare the Basis Point Set (BPS) encoding in PLACE [37] with our per-basis point distance ('Ours w/ " + }, + { + "bbox": [ + 130, + 116, + 482, + 380 + ], + "type": "inline_equation", + "content": "BPS^{\\prime}" + }, + { + "bbox": [ + 130, + 116, + 482, + 380 + ], + "type": "text", + "content": ") by replacing the per-basis point distance with BPS encoding following the official implementation in [37]. Moreover, we show results from end-to-end training of our method ('Ours " + }, + { + "bbox": [ + 130, + 116, + 482, + 380 + ], + "type": "inline_equation", + "content": "e2e^{\\prime}" + }, + { + "bbox": [ + 130, + 116, + 482, + 380 + ], + "type": "text", + "content": "), which encounters a local minimum in the early epochs. Results shown in Fig. 4 further demonstrate the effectiveness of our method." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 399, + 220, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 399, + 220, + 412 + ], + "spans": [ + { + "bbox": [ + 132, + 399, + 220, + 412 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 426, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 521 + ], + "type": "text", + "content": "In this work, we present an approach for scene-aware human motion forecasting that explicitly models the human-scene interaction. To constrain the whole-body motion, we have proposed a mutual distance representation: the per-vertex signed distance and the per-basis point distance. The mutual distance captures both the detailed local body movements and the global human motion. Thanks to the explicit mutual distance constraints, Our method generates future human motions that are more plausible and accurate than those produced by state-of-the-art scene-aware motion prediction techniques." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "type": "text", + "content": "Societal Impacts & Limitations. The motion prediction results depend on the prediction of the mutual distance is one limitation of our work. Our ablation study using the ground truth mutual distance indicates that enhancing the prediction of mutual distance results in improved motion prediction. This can therefore serve as a potential direction for future research. In real applications, a potential risk of our model is that without explicit real world physics modeling, it is possible for our model to predict future motions that have imbalanced forces. This may lead to unstable human motion when deploy to real life applications like human-robot interaction such as collision." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Acknowledgements This research was supported in part by the Australia Research Council DECRA Fellowship (DE180100628) and ARC Discovery Grant (DP200102274)." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "C. Xing et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "type": "text", + "content": "1. Aksan, E., Kaufmann, M., Hilliges, O.: Structured prediction helps 3d human motion modelling. In: ICCV. pp. 7144-7153 (2019) 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 162, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 162, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 138, + 162, + 480, + 194 + ], + "type": "text", + "content": "2. Brand, M., Hertzmann, A.: Style machines. In: Proceedings of the 27th annual conference on Computer graphics and interactive techniques. pp. 183-192. ACM Press/Addison-Wesley Publishing Co. (2000) 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 195, + 480, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 195, + 480, + 227 + ], + "spans": [ + { + "bbox": [ + 138, + 195, + 480, + 227 + ], + "type": "text", + "content": "3. Cai, Y., Huang, L., Wang, Y., Cham, T.J., Cai, J., Yuan, J., Liu, J., Yang, X., Zhu, Y., Shen, X., et al.: Learning progressive joint propagation for human motion prediction. In: ECCV. pp. 226-242. Springer (2020) 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 228, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 480, + 258 + ], + "type": "text", + "content": "4. Cao, Z., Gao, H., Mangalam, K., Cai, Q.Z., Vo, M., Malik, J.: Long-term human motion prediction with scene context. In: ECCV. pp. 387-404. Springer (2020) 1, 2, 3, 4, 9, 11, 13" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 261, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 261, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 138, + 261, + 480, + 281 + ], + "type": "text", + "content": "5. Corona, E., Pumarola, A., Alenya, G., Moreno-Noguer, F.: Context-aware human motion prediction. In: CVPR. pp. 6992-7001 (2020) 1, 2, 3, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 282, + 480, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 282, + 480, + 313 + ], + "spans": [ + { + "bbox": [ + 138, + 282, + 480, + 313 + ], + "type": "text", + "content": "6. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: CVPR. pp. 5828-5839 (2017) 10" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 316, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 316, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 138, + 316, + 480, + 335 + ], + "type": "text", + "content": "7. González, Á.: Measurement of areas on a sphere using fibonacci and latitude-longitude lattices. Mathematical Geosciences 42, 49-64 (2010) 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 337, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 337, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 138, + 337, + 480, + 357 + ], + "type": "text", + "content": "8. Gopalakrishnan, A., Mali, A., Kifer, D., Giles, L., Ororbia, A.G.: A neural temporal model for human motion prediction. In: CVPR. pp. 12116-12125 (2019) 1, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 359, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 359, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 138, + 359, + 480, + 380 + ], + "type": "text", + "content": "9. Hassan, M., Ceylan, D., Villegas, R., Saito, J., Yang, J., Zhou, Y., Black, M.J.: Stochastic scene-aware motion prediction. In: ICCV. pp. 11374-11384 (2021) 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 381, + 480, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 381, + 480, + 411 + ], + "spans": [ + { + "bbox": [ + 138, + 381, + 480, + 411 + ], + "type": "text", + "content": "0. Hassan, M., Choutas, V., Tzionas, D., Black, M.J.: Resolving 3d human pose ambiguities with 3d scene constraints. In: ICCV. pp. 2282-2292 (2019) 9, 10, 11, 12, 13" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 414, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 414, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 138, + 414, + 480, + 435 + ], + "type": "text", + "content": "1. Hassan, M., Ghosh, P., Tesch, J., Tzionas, D., Black, M.J.: Populating 3d scenes by learning human-scene interaction. In: CVPR. pp. 14708-14718 (2021) 1, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 436, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 436, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 436, + 480, + 468 + ], + "type": "text", + "content": "2. Ionescu, C., Papava, D., Olaru, V., Sminchisescu, C.: Human3.6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. TPAMI 36(7), 1325-1339 (2014) 10" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 469, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 469, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 138, + 469, + 480, + 490 + ], + "type": "text", + "content": "3. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. In: ICLR (2015) 11" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 491, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 491, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 138, + 491, + 480, + 511 + ], + "type": "text", + "content": "4. Kipf, T.N., Welling, M.: Semi-supervised classification with graph convolutional networks. In: ICLR (2017) 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 512, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 512, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 512, + 480, + 533 + ], + "type": "text", + "content": "5. Koppula, H.S., Saxena, A.: Anticipating human activities for reactive robotic response. In: IROS. p. 2071. Tokyo (2013) 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 534, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 534, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 534, + 480, + 567 + ], + "type": "text", + "content": "6. Li, M., Chen, S., Zhao, Y., Zhang, Y., Wang, Y., Tian, Q.: Dynamic multiscale graph neural networks for 3d skeleton based human motion prediction. In: CVPR. pp. 214-223 (2020) 1, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 567, + 480, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 480, + 589 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 480, + 589 + ], + "type": "text", + "content": "7. Li, X., Li, H., Joo, H., Liu, Y., Sheikh, Y.: Structure from recurrent motion: From rigidity to recurrency. In: CVPR (2018) 1, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 590, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 590, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 138, + 590, + 480, + 610 + ], + "type": "text", + "content": "8. Liu, Z., Tang, H., Lin, Y., Han, S.: Point-voxel cnn for efficient 3d deep learning. Advances in Neural Information Processing Systems 32 (2019) 14" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "type": "text", + "content": "9. Mahmood, N., Ghorbani, N., Troje, N.F., Pons-Moll, G., Black, M.J.: Amass: Archive of motion capture as surface shapes. In: ICCV (2019) 10" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "text", + "content": "20. Mao, W., Hartley, R.I., Salzmann, M., et al.: Contact-aware human motion forecasting. Advances in Neural Information Processing Systems 35, 7356-7367 (2022), 1, 2, 3, 4, 5, 9, 10, 11, 12, 13, 14" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 153, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 153, + 91, + 447, + 102 + ], + "type": "text", + "content": "Scene-aware Human Motion Forecasting via Mutual Distance Prediction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 659 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "text", + "content": "21. Mao, W., Liu, M., Salzmann, M.: History repeats itself: Human motion prediction via motion attention. In: ECCV. pp. 474-489. Springer (2020) 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 480, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 480, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 480, + 159 + ], + "type": "text", + "content": "22. Mao, W., Liu, M., Salzmann, M., Li, H.: Learning trajectory dependencies for human motion prediction. In: ICCV. pp. 9489-9497 (2019) 1, 3, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 160, + 480, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 480, + 191 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 480, + 191 + ], + "type": "text", + "content": "23. Paden, B., Cap, M., Yong, S.Z., Yershov, D., Frazzoli, E.: A survey of motion planning and control techniques for self-driving urban vehicles. IEEE Transactions on intelligent vehicles 1(1), 33-55 (2016) 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 192, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 192, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 192, + 480, + 224 + ], + "type": "text", + "content": "24. Paszke, A., Gross, S., Chintala, S., Chanan, G., Yang, E., DeVito, Z., Lin, Z., Desmaison, A., Antiga, L., Lerer, A.: Automatic differentiation in pytorch. In: NeurIPS-W (2017) 11" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 224, + 480, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 480, + 255 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 480, + 255 + ], + "type": "text", + "content": "25. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3D hands, face, and body from a single image. In: CVPR. pp. 10975-10985 (2019) 4, 5, 8, 10" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 256, + 480, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 256, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 132, + 256, + 480, + 277 + ], + "type": "text", + "content": "26. Prokudin, S., Lassner, C., Romero, J.: Efficient learning on point clouds with basis point sets. In: ICCV. pp. 4332-4341 (2019) 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 277, + 480, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 277, + 480, + 308 + ], + "spans": [ + { + "bbox": [ + 132, + 277, + 480, + 308 + ], + "type": "text", + "content": "27. Scofano, L., Sampieri, A., Schiele, E., De Matteis, E., Leal-Taixe, L., Galasso, F.: Staged contact-aware global human motion forecasting. In: BMVC (2023) 2, 3, 4, 10, 11, 12, 13" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 309, + 480, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 309, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 132, + 309, + 480, + 331 + ], + "type": "text", + "content": "28. Sidenbladh, H., Black, M.J., Sigal, L.: Implicit probabilistic models of human motion for synthesis and tracking. In: ECCV. pp. 784-800. Springer (2002) 1, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 331, + 480, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 331, + 480, + 351 + ], + "spans": [ + { + "bbox": [ + 132, + 331, + 480, + 351 + ], + "type": "text", + "content": "29. Starke, S., Zhang, H., Komura, T., Saito, J.: Neural state machine for character-scene interactions. ACM Trans. Graph. 38(6), 209-1 (2019) 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 352, + 480, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 352, + 480, + 372 + ], + "spans": [ + { + "bbox": [ + 132, + 352, + 480, + 372 + ], + "type": "text", + "content": "30. Taylor, G.W., Hinton, G.E., Roweis, S.T.: Modeling human motion using binary latent variables. In: NeurIPS. pp. 1345-1352 (2006) 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 373, + 480, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 373, + 480, + 415 + ], + "spans": [ + { + "bbox": [ + 132, + 373, + 480, + 415 + ], + "type": "text", + "content": "31. Van Welbergen, H., Van Basten, B.J., Egges, A., Ruttkay, Z.M., Overmars, M.H.: Real time animation of virtual humans: a trade-off between naturalness and control. In: Computer Graphics Forum. vol. 29, pp. 2530-2554. Wiley Online Library (2010) 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 415, + 480, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 415, + 480, + 436 + ], + "spans": [ + { + "bbox": [ + 132, + 415, + 480, + 436 + ], + "type": "text", + "content": "32. Wang, B., Adeli, E., Chiu, H.k., Huang, D.A., Niebles, J.C.: Imitation learning for human pose prediction. In: ICCV. pp. 7124-7133 (2019) 1, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 437, + 480, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 437, + 480, + 457 + ], + "spans": [ + { + "bbox": [ + 132, + 437, + 480, + 457 + ], + "type": "text", + "content": "33. Wang, J.M., Fleet, D.J., Hertzmann, A.: Gaussian process dynamical models for human motion. TPAMI 30(2), 283-298 (2008) 1, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 458, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 458, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 458, + 480, + 479 + ], + "type": "text", + "content": "34. Wang, J., Xu, H., Xu, J., Liu, S., Wang, X.: Synthesizing long-term 3d human motion and interaction in 3d scenes. In: CVPR. pp. 9401-9411 (2021) 1, 2, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 479, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 500 + ], + "type": "text", + "content": "35. Wang, J., Yan, S., Dai, B., Lin, D.: Scene-aware generative network for human motion synthesis. In: CVPR. pp. 12206-12215 (2021) 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 501, + 480, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 532 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 532 + ], + "type": "text", + "content": "36. Wang, Z., Chen, Y., Liu, T., Zhu, Y., Liang, W., Huang, S.: Humanise: Language-conditioned human motion generation in 3d scenes. Advances in Neural Information Processing Systems 35, 14959-14971 (2022) 9, 10, 11, 12, 13, 14" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 533, + 480, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 480, + 563 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 480, + 563 + ], + "type": "text", + "content": "37. Zhang, S., Zhang, Y., Ma, Q., Black, M.J., Tang, S.: Place: Proximity learning of articulation and contact in 3d environments. In: 3DV. pp. 642-651. IEEE (2020) 1, 2, 3, 4, 14" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 563, + 480, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 563, + 480, + 585 + ], + "spans": [ + { + "bbox": [ + 132, + 563, + 480, + 585 + ], + "type": "text", + "content": "38. Zhang, Y., Black, M.J., Tang, S.: We are more than our joints: Predicting how 3d bodies move. In: CVPR. pp. 3372-3382 (2021) 5, 11" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 586, + 480, + 606 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 586, + 480, + 606 + ], + "spans": [ + { + "bbox": [ + 132, + 586, + 480, + 606 + ], + "type": "text", + "content": "39. Zhang, Y., Hassan, M., Neumann, H., Black, M.J., Tang, S.: Generating 3d people in scenes without people. In: CVPR. pp. 6194-6204 (2020) 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 606, + 480, + 638 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 606, + 480, + 638 + ], + "spans": [ + { + "bbox": [ + 132, + 606, + 480, + 638 + ], + "type": "text", + "content": "40. Zheng, Y., Yang, Y., Mo, K., Li, J., Yu, T., Liu, Y., Liu, C.K., Guibas, L.J.: Gimo: Gaze-informed human motion prediction in context. In: ECCV. pp. 676-694. Springer (2022) 1, 2, 3, 9, 10, 11, 12, 13" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 639, + 480, + 659 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 639, + 480, + 659 + ], + "spans": [ + { + "bbox": [ + 132, + 639, + 480, + 659 + ], + "type": "text", + "content": "41. Zhou, Y., Barnes, C., Lu, J., Yang, J., Li, H.: On the continuity of rotation representations in neural networks. In: CVPR. pp. 5745-5753 (2019) 8" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "C. Xing et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/SceneGraphLoc_ Cross-Modal Coarse Visual Localization on 3D Scene Graphs/ca6ad507-bc20-40c4-b703-2093582a66a7_content_list.json b/2024/SceneGraphLoc_ Cross-Modal Coarse Visual Localization on 3D Scene Graphs/ca6ad507-bc20-40c4-b703-2093582a66a7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5157a0814138718e8d7ee5561f6e1f745388bd0d --- /dev/null +++ b/2024/SceneGraphLoc_ Cross-Modal Coarse Visual Localization on 3D Scene Graphs/ca6ad507-bc20-40c4-b703-2093582a66a7_content_list.json @@ -0,0 +1,1900 @@ +[ + { + "type": "text", + "text": "SceneGraphLoc: Cross-Modal Coarse Visual Localization on 3D Scene Graphs", + "text_level": 1, + "bbox": [ + 250, + 140, + 754, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yang Miao1, Francis Engelmann1,2, Olga Vysotska1, Federico Tombari2,3, Marc Pollefeys1,4, and Daniel Béla Baráth1", + "bbox": [ + 264, + 210, + 738, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1ETH Zurich 2Google 3TU Munich 4Microsoft", + "bbox": [ + 344, + 253, + 656, + 268 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b97064f8bd510427d17b3f0f97fd8f0aded53b6b81ff72b84f95c056655847ec.jpg", + "image_caption": [ + "Fig. 1: SceneGraphLoc addresses the novel problem of localizing a query image in a database of 3D scenes represented as compact multi-modal 3D scene graphs." + ], + "image_footnote": [], + "bbox": [ + 218, + 294, + 784, + 435 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. We introduce the task of localizing an input image within a multi-modal reference map represented by a collection of 3D scene graphs. These scene graphs comprise multiple modalities, including object-level point clouds, images, attributes, and relationships between objects, offering a lightweight and efficient alternative to conventional methods that rely on extensive image databases. Given these modalities, the proposed method SceneGraphLoc learns a fixed-sized embedding for each node (i.e., representing object instances) in the scene graph, enabling effective matching with the objects visible in the input query image. This strategy significantly outperforms other cross-modal methods, even without incorporating images into the map representation. With images, SceneGraphLoc achieves performance close to that of state-of-the-art techniques depending on large image databases, while requiring three orders-of-magnitude less storage and operating orders-of-magnitude faster. Code and models are available at https://scenegraphloc.github.io.", + "bbox": [ + 259, + 498, + 746, + 705 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Coarse Localization $\\cdot$ 3D Scene Graph $\\cdot$ Multi-modality", + "bbox": [ + 261, + 719, + 720, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 761, + 375, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Coarse visual localization, or place recognition, is a fundamental component in computer vision and robotics applications, defined as the task of identifying the approximate location where a query image was taken, given a certain tolerance", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6b731af108ffd1c2702a35f20734a8083aac8a368bca4f0292c01cee44ba433e.jpg", + "image_caption": [ + "Fig. 2: Overview. The training phase is represented by orange arrows, while blue arrows denote the inference phase. During training, a query image and its associated 3D scene graph form a positive sample within a contrastive learning framework, where negative samples are generated by associating scene graphs of different scenes with the same query image. The objective is to learn the embeddings of both the graph and the image so that embeddings of the positive pair are drawn closer, whereas those of the negative pair are pushed apart. In the inference phase, the task involves assigning the correct scene graph to a given query image from a selection of multiple graphs, achieved by optimizing the cosine similarity between their embeddings." + ], + "image_footnote": [], + "bbox": [ + 241, + 145, + 759, + 281 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "level [2,9,22,28,39,40,50,58,69,116,117,127,132,143]. This capability is crucial for estimating the state of robots and is widely utilized in autonomous, unmanned aerial, terrestrial, and underwater vehicles, as well as AR/VR devices. The task is typically approached as an image retrieval problem, where the image to be localized is compared against a large database of posed images and, optionally, a 3D reconstruction of the scene. The most similar images retrieved from the database are used to estimate the precise location of the query image.", + "bbox": [ + 212, + 468, + 787, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The challenge with current state-of-the-art image-based coarse localization methods, such as AnyLoc [55], is their dependency on extensive image databases, which are not only storage-heavy but also slow to query, despite optimizations through hashing and other tricks. Moreover, these methods typically necessitate that the query and database share the same modality, limiting the scope of their application. Cross-modal approaches, such as [100,142], which attempt to bridge different types of data, often restrict their scope to connecting two modalities at a time (e.g., image-to-point cloud or image-to-bird's eye view map), one for the query and one for the database, thus narrowing their potential applications.", + "bbox": [ + 212, + 579, + 787, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This paper addresses the novel challenge of localizing a query image within a database that is represented not by conventional images but by the 3D scene graph [3,124] that integrates a diverse set of modalities, including point clouds, images, semantics, object attributes, and relationships. We tackle this problem by learning to map these modalities into a unified embedding space, thus allowing us to represent indoor scenes compactly through their objects (e.g., table and wall). This method enables the creation of small, efficient databases and significantly accelerates the coarse localization process.", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Y. Miao et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contributions. The primary contributions of this paper are as follows:", + "bbox": [ + 215, + 146, + 733, + 161 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Introducing a novel problem: cross-modal localization of a query image within 3D scene graphs incorporating a mixture of modalities.", + "2. SceneGraphLoc, a new method for the coarse localization of an input image given a reference map represented by a database of 3D scene graphs." + ], + "bbox": [ + 220, + 171, + 785, + 231 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Even without incorporating images into the map, SceneGraphLoc largely outperforms other cross-modal methods on two large-scale, real-world indoor datasets. With images, SceneGraphLoc achieves performance close to that of state-of-the-art image-based methods while requiring three orders-of-magnitude less storage and operating orders-of-magnitude faster. The method is visualized in Fig. 1.", + "bbox": [ + 212, + 239, + 785, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 339, + 387, + 356 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Localization, the process of determining the position and orientation of an agent within a pre-built map, is pivotal across various domains such as mobile robots [42, 65, 139, 140], and augmented reality [16, 71]. The differentiation in localization techniques arises from their scene representation methods - be it through explicit 3D models [31, 32, 51, 64, 68, 72, 98, 99, 101, 102, 105, 112, 133, 135], sets of posed images [10, 88, 137, 141], or implicitly via neural network weights [6, 11-13, 17, 56, 57, 78, 118, 122] - and their approach to camera pose estimation, whether by 2D-3D [31, 32, 98, 99, 102, 112, 120, 126, 133] or 2D-2D [10, 141] matches, or through a composite of base poses [56, 57, 78, 88, 103, 122]. In practice, localization comprises two main steps: a coarse and precise stage. Here, we focus on the coarse step, finding potential locations of a query image.", + "bbox": [ + 212, + 371, + 785, + 537 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Coarse Localization (or place recognition) is often cast as an image retrieval problem [2,8,9,27,40,50,55,59,84,85,127] that consists of two phases. In the offline indexing phase, a reference map (image or point cloud database) is gathered. In the online retrieval phase, a query image - captured during a future traverse of the environment - is localized coarsely by retrieving the closest match to this image in the reference map. Recent methods perform the retrieval using learned embeddings that are produced by a feature extraction backbone equipped with a head that implements some form of aggregation or pooling, the most notable being NetVLAD [2]. While these methods achieve impressive results, they are limited to a single modality (e.g., images) and require large databases.", + "bbox": [ + 212, + 537, + 785, + 688 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Localization using multi-modal data. While dense mesh models are not as widely adopted as sparse Structure-from-Motion-based approaches, they have nonetheless been the focus of considerable research efforts [4,5,14,15,35,82,89,93, 107,109,110,115,138,142]. The body of prior work can be broadly segmented into two main strategies: The first entails the precise alignment of actual images with three-dimensional models (which may be coarse) through applying specialized techniques such as ones using contours [89] or skylines [93]. The second strategy emphasizes the identification and matching of local image features [4,5,14,35, 107,109,110,115,126,138,142], a method that has gained traction for its ability to match real-world images with non-photorealistic renderings of colored meshes, or", + "bbox": [ + 212, + 689, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "SceneGraphLoc", + "bbox": [ + 624, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "even meshes without color [14,82,115]. CAD and other models are also commonly used by object pose estimation [4, 30, 34, 36, 44-46, 62, 90]. Image to LiDAR localization [7,26,43,108] is also relevant, especially in robotics applications.", + "bbox": [ + 215, + 146, + 784, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Another variant of multi-modal data localization involves cross-view matching. This technique determines the camera position by finding correspondences between a ground-level query image and a two-dimensional bird's eye view map, such as a satellite image or a semantic landscape map [47,48,66,100,107,121,129]. Other cross-modal techniques were also proposed to involve semantics [28, 29] and event cameras [53] in image-based localization.", + "bbox": [ + 215, + 210, + 784, + 300 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "These approaches, while demonstrating promising localization results, are limited to interactions between two modalities - one for the query and one for the reference database. Our approach, in contrast, seeks the cross-modal coarse localization of a query image in a database composed of multiple modalities, e.g., 3D point clouds, images, semantics, object attributes, and relationships.", + "bbox": [ + 215, + 320, + 784, + 395 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Scene Representation, encapsulating various scene attributes has evolved significantly, yielding diverse surface representations from explicit forms (3D point clouds [75,91], meshes [37], surfels [111]) to implicit ones (occupancy [61], signed distance functions [20,52]). The advent of neural representations has introduced novel means of encoding geometry [73,83,87,128], appearance [76,80], and semantics [74,79]. A comprehensive review is provided by Tewari et al. [114]. The integration of directions/rays [77] and visibility encoding in surface reconstruction [104] has further enriched this domain. Armeni et al. [3] introduced the 3D scene graph structure as a multi-layer representation of a scene that captures geometry, semantics, objects, and camera poses in a unified manner. Subsequent efforts [96,124] have further advanced 3D scene graph learning and structure.", + "bbox": [ + 215, + 412, + 784, + 579 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The increasing interest in 3D scene graphs [3, 54, 60, 96, 124] underscores their potential as structured, rich descriptors for real-world scenes. Methods range from online incremental construction [49, 130] to offline generation from RGB-D imagery [3, 95, 124], and approaches for scene graph prediction [134, 136]. Their application spans embodied AI [94, 95, 106], task completion [1, 25], variability estimation [70], and SLAM [49, 95]. Recent studies like [131] introduce frameworks for localizing unseen objects by utilizing 3D scene graphs and graph neural networks for relation prediction, showcasing the utility of scene graphs in enhancing spatial understanding. Similarly, [97] offers new perspectives on 3D scene alignment, employing node matching within overlapping scene graphs to facilitate precise 3D map alignment. Kabalar et al. [54] assume that the query image has been coarsely localized and leverages a scene graph to identify dynamic objects and for precise localization with image features. Despite these significant advancements underscoring the value of scene graphs, their potential in multimodal localization remains largely untapped. In this paper, we use a scene graph representation of the map in which we aim to localize a query image.", + "bbox": [ + 215, + 598, + 784, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Y. Miao et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Visual Localization with 3D Scene Graph", + "text_level": 1, + "bbox": [ + 215, + 143, + 653, + 162 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Problem Statement. Let us assume that we are provided with a pre-constructed map of the environment, denoted as $\\mathcal{G}$ , which is represented as a set of $N \\in \\mathbb{N}^+$ 3D scene graphs $\\mathcal{G}_i = (\\mathcal{V}_i, \\mathcal{E}_i)$ , such that $\\mathcal{G} = \\{\\mathcal{G}_i\\}_{i \\in [0, N)}$ . Having separate graphs $\\mathcal{G}_i$ is analogous to the hierarchy levels presented in [49, 95], each representing a group of object instances constituting a place like a room or building. Moreover, this task can also be straightforwardly redefined as subgraph selection. Vertices $v \\in \\mathcal{V}_i$ symbolize instances of objects (e.g., chairs, tables) and large instances of semantic categories (e.g., walls, ground) within the scene, $i \\in [0, N)$ . Let $\\mathcal{V} = \\bigcup_{i \\in [0, N)} \\mathcal{V}_i$ aggregate all objects across the scenes. Edges $\\mathcal{E}_i = \\{(v_j^i, v_k^i) | v_j^i, v_k^i \\in \\mathcal{V}_i\\}$ delineate the relationships between objects, such as \"nearby\", \"standing on\", and \"attached to\".", + "bbox": [ + 212, + 176, + 787, + 343 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For each vertex $v$ , we introduce $M \\in \\mathbb{N}^{+}$ map modalities $f_{j}: \\mathcal{V} \\to \\mathcal{M}_{j}$ for $j \\in [0, M)$ , where $f_{j}$ maps vertex $v$ to the $j^{th}$ modality $\\mathcal{M}_{j}$ that may be $\\mathcal{M}_{j} \\in \\{\\text{position}, \\text{orientation}, \\text{pointcloud}, \\text{semantic category}, \\text{image}, \\text{attribute}\\}$ . While this paper focuses on these modalities, the set is easily extendable by incorporating additional ones, such as textual description or floor plan.", + "bbox": [ + 212, + 345, + 799, + 419 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given $\\mathcal{G}$ and an input query image $I$ , the objective is to identify the scene graph $\\mathcal{G}_i$ corresponding to the space depicted in image $I$ . Note that while we focus on having an input image in this paper, the method can be modified to other modalities as well, e.g., depth image. Formally, we aim to resolve problem:", + "bbox": [ + 212, + 420, + 787, + 481 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} _ {i ^ {*}} = \\underset {i \\in [ 0, N)} {\\arg \\max } \\llbracket \\text {c o n t a i n s} (\\mathcal {G} _ {i}, \\mathbf {p} _ {I}) \\rrbracket , \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 494, + 784, + 521 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{p}_I \\in \\mathbb{R}^3$ is the unknown 3D position of the image, and $[.]$ is the Iverson bracket which equals to 1 if the condition inside holds and 0 otherwise. It is important to note that our objective is coarse localization, opting for the selection of $\\mathcal{G}_i$ without needing precise estimation of $\\mathbf{p}_I$ .", + "bbox": [ + 212, + 527, + 787, + 588 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To this end, the optimization problem can be reformulated to incorporate the chirality constraint, asserting that if an object is visible in image $I$ , it must be positioned in front of the camera in 3D space, unobstructed by any entities (e.g., walls). Consequently, the problem becomes:", + "bbox": [ + 212, + 589, + 787, + 650 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} _ {i ^ {*}} = \\underset {i \\in [ 0, N)} {\\arg \\max } \\sum_ {o _ {I} \\in \\mathcal {O} _ {I}} \\llbracket \\text {v i s i b l e} (\\mathcal {G} _ {i}, o _ {I}) \\rrbracket \\approx \\underset {i \\in [ 0, N)} {\\arg \\max } \\sum_ {o _ {I} \\in \\mathcal {O} _ {I}} \\log P (o _ {I} \\mid \\mathcal {G} _ {i}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 246, + 661, + 784, + 694 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{O}_I$ is the object set visible in image $I$ , object $o_I \\subseteq \\{(x,y) \\in I\\}$ is a set of pixels in the image $(o_I \\in \\mathcal{O}_I)$ , visible: $\\mathcal{G}_i \\times \\mathcal{O}_I \\to \\{0,1\\}$ indicates if an object is visible in a scene graph, $P(o_I \\mid \\mathcal{G}_i)$ is the probability of $o_I$ stemming from $\\mathcal{G}_i$ .", + "bbox": [ + 212, + 703, + 787, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Function visible can be approached as an indicator of whether object $o_{I}$ appears in graph $G_{i}$ . It holds if and only if there exists a $v \\in \\mathcal{V}_{i}$ such that $v$ represents the same object as $o_{I}$ . Therefore, this problem can be approached as matching a set of image pixels $(o_{I})$ to a scene graph node $(v)$ that comprises a set of modalities. In the next sections, we will describe a way to learn a unified embedding space for both $o_{I}$ and $v$ such that they become matchable.", + "bbox": [ + 212, + 750, + 787, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "SceneGraphLoc", + "bbox": [ + 624, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proposed Pipeline is shown in Fig. 2. It consists of two concurrent stages: the first one generates object embeddings $e_{q} \\in \\mathbb{R}^{D}$ from patches $q \\in \\mathcal{Q}_I$ within the query image $I$ , and the second derives node embeddings $e_{v} \\in \\mathbb{R}^{D}$ for nodes $v \\in \\mathcal{V}_i$ in the scene graph $\\mathcal{G}_i$ . The training objective is to make $\\delta(e_{q}, e_{v}) = 0$ if and only if the object associated with node $v$ is directly visible (i.e., neither occluded nor outside the camera frustum) through the image patch $q$ . Here, $\\delta : \\mathbb{R}^D \\times \\mathbb{R}^D \\to \\mathbb{R}$ denotes inverse cosine similarity normalized to [0,1].", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After generating $e_{q}$ and $e_{v}$ , the model performs nearest neighbor matching (NN) for each patch in each scene graph $\\mathcal{G}_{i}$ , assigning node $v$ to $q$ such that", + "bbox": [ + 215, + 252, + 785, + 282 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {N N} (q, \\mathcal {V} _ {i}) = \\underset {v \\in \\mathcal {V} _ {i}} {\\arg \\min } \\delta \\left(e _ {q}, e _ {v}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 392, + 287, + 785, + 314 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Through matching, we establish patch-to-node correspondences $\\mathcal{C}_i = \\{(q,v) \\mid v = \\mathrm{NN}(q,\\mathcal{V}_i) \\in \\mathcal{V}_i, q \\in \\mathcal{Q}_I\\}$ . Based on $\\mathcal{C}_i$ , we devise an image-to-graph similarity score enabling us to deduce whether image $I$ corresponds to the space represented by $\\mathcal{G}_i$ . Finally, potential coarse locations of $I$ are selected by maximizing the similarity score across the stored scene graphs as depicted in Eq 2.", + "bbox": [ + 212, + 318, + 787, + 395 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.1 Object Embeddings in the Scene Graph", + "text_level": 1, + "bbox": [ + 214, + 412, + 589, + 428 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This section aims to obtain an embedding for each node within graph $\\mathcal{G}_i$ , encapsulating information from all available modalities. Our method builds upon the method of Sarkar et al. [97], with enhancements to include the image modality and to distill a unified embedding from all modalities rather than merely concatenating separate embeddings as in [97].", + "bbox": [ + 212, + 431, + 571, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Scene graphs are conceptualized as multimodal knowledge graphs, similar to those used in entity alignment, treating semantic and geometric", + "bbox": [ + 212, + 553, + 571, + 598 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "information as distinct modalities. The objective is to learn a joint multi-modal embedding from the individual modal encodings (uni-modal), ensuring nodes corresponding to the same object instance across different graphs are closely positioned. This involves the creation of uni-modal embeddings for the three primary types of 3D scene graph information: object embeddings encoding nodes in $\\mathcal{V}$ , structure embeddings $\\mathcal{S}$ representing edges in $\\mathcal{E}$ as a structured graph, and two meta modalities encoding attributes $(\\mathcal{A})$ and relationships $(\\mathcal{R})$ between objects as one-hot vectors. These uni-modal embeddings are then combined in a weighted manner and optimized jointly through knowledge distillation.", + "bbox": [ + 212, + 598, + 785, + 734 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Each of these modalities is processed separately to generate uni-modal embeddings, which are subsequently integrated to model complex inter-modal interactions within the joint embedding space.", + "bbox": [ + 212, + 734, + 785, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Object Embedding. Node $v \\in \\mathcal{V}$ may contain multiple modalities, such as point cloud $(\\mathcal{P})$ and image $(\\mathcal{I})$ . Point clouds contain rich geometric information about objects. The point cloud corresponding to each $v \\in \\mathcal{V}$ is inputted to the object encoder. We employ the PointNet architecture [91] as the", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7467a870839053004005c7f8889d65635d49e7aa88ab50b8c2311bb06ce46386.jpg", + "image_caption": [ + "Fig.3: The embedding of image modality $\\mathcal{I}$ for each object. The image crops of a pillow are shown as an example." + ], + "image_footnote": [], + "bbox": [ + 586, + 438, + 784, + 525 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Y. Miao et al.", + "bbox": [ + 271, + 114, + 364, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "object encoder to extract the geometric feature $e_v^{\\mathcal{P}}$ for every node. Furthermore, we integrate multi-level and multi-view visual embeddings to enrich the graph encoder with a more nuanced understanding of image information. The visual embedding pipeline is visualized in Fig. 3. For each node $v$ denoting a 3D object, the top $K_{view} \\in \\mathbb{N}^{+}$ images with largest visibility of $v$ is selected: $\\{I_{\\mathrm{db},k}^v \\mid k = [0,K)\\} \\subseteq \\mathcal{I}_{\\mathrm{db}}$ from the database $\\mathcal{I}_{\\mathrm{db}}$ . We can define an ordering over these images such that $\\phi(I_{\\mathrm{db},0}^v,v) \\geq \\phi(I_{\\mathrm{db},1}^v,v) \\geq \\ldots \\geq \\phi(I_{\\mathrm{db},K_{view}-1}^v,v)$ , with the visibility function $\\phi$ quantifying the extent of node $v$ observed in each image through pixel count. Visibility check is implemented by projecting the 3D model of $v$ to each posed image in $\\mathcal{I}_{\\mathrm{db}}$ , which are usually available when constructing the scene graph [49, 130]. Parameter $K_{view} = 10$ in all our experiments.", + "bbox": [ + 212, + 145, + 787, + 313 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Drawing inspiration from OpenMask3D [113], for any given view $I_{\\mathrm{db}}^v$ of object $v$ , initial steps include calculating the bounding box $b_{v,0}$ of $v$ within the image, followed by the generation of multi-level bounding boxes $\\{b_{v,l} \\mid l \\in [0,L)\\}$ through iterative enlargement of $b_{v,0}$ . This enlargement strategy aims to capture contextual information around object $v$ . Subsequently, Dino V2 [81] processes the image crops defined by $b_{v,l}$ , extracting multi-level features $\\{f_l \\mid l \\in [0,L)\\}$ . For each image $I_{\\mathrm{db},k}^v$ , an average pooling operation aggregates these multi-level crop features into a singular feature vector $f = \\operatorname{avg\\_pool}\\{f_l \\mid l \\in [0,L)\\}$ . The final step of this process involves the application of a Transformer encoder, which incorporates image poses as positional encodings. This step synthesizes multiview object tokens into a cohesive visual embedding $e_v^{\\mathcal{I}}$ , effectively integrating the diverse perspectives and levels of contextual information pertaining to each object within the database. Please note that the image database does not necessarily have to be stored after distilling the object embeddings.", + "bbox": [ + 212, + 319, + 787, + 531 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Structure Embedding. 3D Scene Graphs encapsulate the object relationships, which we exploit to encode their spatial configuration. This relational data is represented through a structure graph, where node features embody the relative translations between object instances, and edges denote these relationships. The relative translation is determined by calculating the distance from the object instance to any other object in the scene. To encapsulate this structural information within $\\mathcal{G}_i$ , a Graph Attention Network (GAT) [119] is utilized, with the weight matrix constrained to a diagonal form to reduce computational demands and enhance model scalability. Following the method in [97], the structural embedding $e_v^S$ is derived from the final layer of a two-layer GAT model.", + "bbox": [ + 212, + 537, + 787, + 690 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Meta Embeddings. In addition to geometry and structure, the object attributes and the inter-object relationships are captured in two distinct embeddings, $e_v^{\\mathcal{R}}$ and $e_v^{\\mathcal{A}}$ . The relationships an object maintains with others are conceptualized as a bag-of-words [19] feature vector, which is input to a feed-forward neural network, distilled in relational embedding $e_v^{\\mathcal{R}}$ . A similar approach is employed for the attributes associated with the objects, producing embedding $e_v^{\\mathcal{A}}$ .", + "bbox": [ + 212, + 696, + 787, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Joint Embedding. Similar to [97], we concatenate uni-modal features to a compact representation for each object $v$ . Contrary to [97], we encapsulate it in a multi-layer perceptron (MLP) to learn an embedding with size independent of", + "bbox": [ + 212, + 794, + 787, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "SceneGraphLoc", + "bbox": [ + 624, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the available modalities, fusing information from all. Embedding $e_v$ is as:", + "bbox": [ + 215, + 146, + 743, + 162 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\ne _ {v} = \\operatorname {M L P} \\left(\\oplus_ {k \\in \\mathcal {K}} \\left[ \\frac {\\exp (w _ {k})}{\\sum_ {j \\in \\mathcal {K}} \\exp (w _ {j})} e _ {v} ^ {k} \\right]\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 349, + 171, + 785, + 214 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\oplus$ is the concatenation operator, $\\mathcal{K} = \\{\\mathcal{P},\\mathcal{I},\\mathcal{S},\\mathcal{R},\\mathcal{A}\\}$ , and $w_{m}$ is a trainable attention weight for each modality $k\\in \\mathcal{K}$ . A two-layer MLP is applied to map the dimension of the concatenated multi-modal descriptors from $D^{\\mathcal{K}}$ to $D$ . We apply $L_{2}$ normalization to each uni-modal feature before concatenation.", + "bbox": [ + 212, + 223, + 782, + 282 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In practice, these independent modalities are only required and used in the mapping phase of the procedure. This stage involves the construction of an environmental map in the form of a 3D scene graph, during which each node $v$ is distilled into an embedding $e_v$ . During localization, we can ignore independent modalities and use only the fixed-sized embeddings $e_v$ .", + "bbox": [ + 212, + 284, + 784, + 359 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.2 Object Embeddings in the Query Image", + "text_level": 1, + "bbox": [ + 215, + 382, + 591, + 398 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To solve the optimization problem in Eq. 2, we need to find object instances $o_{I} \\in \\mathcal{O}_{I}$ in query image $I$ . A straightforward approach to do so would be to apply a 2D panoptic segmentation algorithm, e.g., Mask2Former [18]. However, we noticed in our experiments that such an approach is susceptible to inaccuracies and failures (over- and under-segmentation) in the segmentation, severely affecting the accuracy. Thus, we approach the problem by a visual Transformer (ViT), breaking up the image into rectangular patches $q \\in \\mathcal{Q}_I$ and distilling an independent embedding for each patch $q$ based on the object visible in $q$ . We use Dino V2 [81] as a backbone to obtain patch-level features. These features are passed through a patch encoder trained to create embeddings from the patch features considering the objects visible from each $q$ . For this encoder, we use a 4-layer convolution neural network (CNN) with residual blocks introduced in [41] on the Dino V2 features and a 3-layer MLP to further map the patch feature to dimension $D$ .", + "bbox": [ + 212, + 407, + 787, + 618 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 Contrastive Learning", + "text_level": 1, + "bbox": [ + 215, + 641, + 437, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We use contrastive learning to learn a joint embedding space for the scene graph nodes and image patches. To do so, we form query image and graph pairs $(I, \\mathcal{G}_I)$ . Real-world scenes are rarely static, e.g., objects move or undergo non-rigid deformations and illumination changes [123]. To ensure that the learned embedding is robust to such temporal changes, we use scene graph $\\mathcal{G}_I$ from the same temporal point when $I$ was captured, as positive samples, as well as a scene graph $\\mathcal{G}_I^t$ from another scan. Graph $\\mathcal{G}_I^t$ represents the same place as $\\mathcal{G}_I$ , but it undergoes temporal changes. An example is shown in Fig. 4.", + "bbox": [ + 212, + 666, + 658, + 818 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8d1a947c739d12d88cdf406f40339bac1ce79f7be3236c88c35980a1c617d118.jpg", + "image_caption": [ + "Fig.4: The same scene at different time steps $t$ ." + ], + "image_footnote": [], + "bbox": [ + 668, + 646, + 785, + 770 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Y. Miao et al.", + "bbox": [ + 271, + 114, + 364, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For a query image $I$ , a set of candidate graphs $\\{\\mathcal{G}_I, \\mathcal{G}_I^t, \\mathcal{G}_1, \\dots, \\mathcal{G}_N\\}$ is provided for training, where $\\{\\mathcal{G}_1, \\dots, \\mathcal{G}_N\\}$ act as $N$ negative samples, depicting different scenes than the target scene of the query image. We train our model by optimizing both a static loss and a temporal loss as:", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\alpha * \\mathcal {L} _ {\\text {s t a t i c}} + (1 - \\alpha) * \\mathcal {L} _ {\\text {t e m p}}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 215, + 784, + 232 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "During training, we assume that image patch to graph node pairs are available [123] such that $P_{I} = \\{(q,v) \\mid q \\in \\mathcal{Q}_{I}, v \\in \\mathcal{G}_{I}\\}$ and $P_{I}^{t} = \\{(q,v^{t}) \\mid q \\in \\mathcal{Q}_{I}, v \\in \\mathcal{G}_{I}^{t}\\}$ . For each pair $(q,v)$ from $P_{I}$ and each pair from $P_{I}^{t}$ , we use the following notation. Set $N_{q}^{\\mathcal{I}} = \\{q' \\mid q' \\in \\mathcal{Q}_{I}, v_{q'} \\neq v\\}$ contains patches seeing objects other than $v$ . $N_{v}^{\\mathcal{G}} = \\{v_{n} \\mid v_{n} \\in \\mathcal{V}_{I} \\cup \\mathcal{V}_{1} \\cup \\ldots \\cup \\mathcal{V}_{N}\\} \\setminus \\{v\\}$ contains the 3D objects of all candidate scene graphs other than $v$ , where $\\mathcal{V}_{I}$ represents the objects nodes of the target graph $\\mathcal{G}_{I}$ and $\\mathcal{V}_{i}$ is the nodes of other graphs $\\mathcal{G}_{i}$ . $N_{v}^{\\mathcal{G}_{I}^{t}} = \\{v_{n}|v_{n} \\in \\mathcal{V}_{I}^{t} \\cup \\mathcal{V}_{1} \\cup \\ldots \\cup \\mathcal{V}_{n}\\} \\setminus \\{v\\}$ , where $\\mathcal{V}_{I}^{t}$ represent the nodes of the target graph $\\mathcal{G}_{I}^{t}$ . The static loss is defined as bi-directional N-pair loss [67] as follows:", + "bbox": [ + 214, + 239, + 782, + 378 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\left. \\right. \\mathcal {L} _ {\\text {s t a t i c}} = E _ {P I \\in B} \\left[ E _ {(q, v) \\in P I} \\left[ - \\log \\left(\\frac {1}{2} p (q, v, N _ {q} ^ {\\mathcal {I}}, N _ {v} ^ {\\mathcal {G}}) + \\frac {1}{2} p (v, q, N _ {q} ^ {\\mathcal {I}}, N _ {v} ^ {\\mathcal {G}})\\right)\\right]\\right], \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 228, + 385, + 782, + 431 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $E_{P_I\\in B}$ represent loss averaging over a batch of query images and their corresponding candidate scene graphs; $p(q,v,N_q^\\mathcal{I},N_v^\\mathcal{G})$ and $p(v,q,N_q^\\mathcal{I},N_v^\\mathcal{G})$ represent the bi-directional probability distributions of the positive pair as:", + "bbox": [ + 212, + 431, + 782, + 477 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\np (q, v, N _ {q} ^ {\\mathcal {I}}, N _ {v} ^ {\\mathcal {G}}) = \\frac {f (e _ {q} , e _ {v})}{f (e _ {q} , e _ {v}) + \\sum_ {q _ {n} \\in N _ {q} ^ {\\mathcal {I}}} f (e _ {q} , e _ {q _ {n}}) + \\sum_ {v _ {n} \\in N _ {v} ^ {\\mathcal {G}}} f (e _ {q} , e _ {v _ {n}})},\n$$\n", + "text_format": "latex", + "bbox": [ + 250, + 484, + 750, + 520 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $f(e_q, e_v) = \\exp \\left( \\frac{-\\delta(e_q, e_v)}{\\tau} \\right)$ , $\\delta(e_q, e_v)$ represents the inverse cosine similarity between embeddings $e_q$ and $e_v$ , and $\\tau$ is a temperature parameter. Probability distribution $p(v, q, N_q^T, N_v^G)$ is written similarly. The temporal loss term is defined the same as Eq.6 but with $P_I^t$ and $N_v^{G^t}$ as follows:", + "bbox": [ + 212, + 529, + 782, + 595 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t e m p}} = E _ {P _ {I} ^ {t} \\in B} \\left[ E _ {(q, v) \\in P _ {I} ^ {t}} \\left[ - \\log \\left(\\frac {1}{2} p (q, v, N _ {q} ^ {\\mathcal {I}}, N _ {v} ^ {\\mathcal {G} ^ {t}}) + \\frac {1}{2} p (v, q, N _ {q} ^ {\\mathcal {I}}, N _ {v} ^ {\\mathcal {G} ^ {t}})\\right) \\right] \\right]. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 223, + 604, + 782, + 650 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "By minimizing Eq.5, the paired cross-modal embeddings $e_{q}$ and $e_{v}$ are pulled together while the embeddings from different objects are pushed apart.", + "bbox": [ + 212, + 651, + 782, + 681 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.4 Scene Graph Retrieval", + "text_level": 1, + "bbox": [ + 214, + 700, + 447, + 715 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Given a pre-established map of an environment represented through a collection of scene graphs $\\mathcal{G} = \\{\\mathcal{G}_i \\mid i \\in [0, N)\\}$ , where each node embedding has been precomputed, the goal during inference is to identify the top- $K$ scene graphs $\\mathcal{G}_i$ in which image $I$ was likely captured. The method to address this challenge involves calculating the similarity between a graph and the image as follows:", + "bbox": [ + 212, + 723, + 782, + 799 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\ns \\left(\\mathcal {G} _ {i}, I\\right) = \\frac {1}{\\left| \\mathcal {Q} _ {I} \\right|} \\sum_ {q \\in \\mathcal {Q} _ {I}} \\left[ 1 - \\delta \\left(e _ {q}, \\operatorname {N N} (q)\\right) \\right], \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 362, + 806, + 782, + 842 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "SceneGraphLoc", + "bbox": [ + 624, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/fdb92816cf9745f003cb305955fbc936562c082c727a27b71c1680fa28e2e8ce.jpg", + "image_caption": [ + "Fig.5: Qualitative Result of object-association-based scene retrieval from the 3RScan dataset [123]. The two left images show the ground truth (left) and predicted (right) patch-to-node associations of the query image. The right part of the figure illustrates the candidate scene graphs sorted by the image-graph similarity." + ], + "image_footnote": [], + "bbox": [ + 233, + 143, + 772, + 236 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\mathcal{Q}_I$ denotes the set of image patches in $I$ , and for each patch $q$ , $\\mathrm{NN}(q, \\mathcal{V}_i) \\in \\mathcal{V}_i$ represents the nearest node in terms of embedding similarity. The function $s$ , which is assumed to map values to interval [0, 1], facilitating the identification of the optimal scene graph $\\mathcal{G}_{i^*}$ that maximizes $s(\\mathcal{G}_i, I)$ . Graph $\\mathcal{G}_{i^*}$ is identified by iterating through all scene graphs and selecting the one with the highest similarity. This approach can be accelerated by using spatial partitioning techniques, such as kd-trees, to preprocess node embeddings within the map.", + "bbox": [ + 212, + 311, + 787, + 417 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 443, + 375, + 460 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Baselines. No existing methods directly tackle our task, but several recent advancements provide relevant baselines. LidarCLIP [43], designed for autonomous driving, transforms LiDAR clouds into global descriptors using the Single-stride Sparse Transformer [24] and matches them with CLIP image encoder embeddings [92]. Though not a perfect fit, it can be adapted for matching point clouds with CLIP embeddings of query images. LIP-Loc [108] employs a similar approach by converting LiDAR clouds into 2D range images for direct encoding and matching. Both methods were fine-tuned on our dataset for accurate comparison. We also explored open-vocabulary [23,33,86,113] object-retrieval-based baselines using OpenMask3D [113] and OpenSeg [33], with OpenMask3D assigning CLIP descriptors to 3D object instances from multiple observing images, and OpenSeg extracting pixel-level CLIP features from the query image. Despite their original purposes, both can be adapted for localization by matching query image descriptors with object instances. For comparison with state-of-the-art visual localization methods requiring large image databases, we included CVNet [63] and AnyLoc [55]. These methods offer advanced performance but demand significant storage for image descriptors and exhibit slower inference times.", + "bbox": [ + 212, + 477, + 787, + 733 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Map Generation. The mapping stage is executed offline as a preprocessing step for visual localization approaches, requiring specific mapping operations for each method before proceeding to localization. For our proposed method, this entails passing the point cloud, images, metadata, and relationships through the 3D scene graph encoder outlined in Section 3.1. For LIP-Loc, this step involves converting point clouds into range images and computing the embeddings of the range images and for LidarCLIP, point clouds are directly encoded into global", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Y. Miao et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/64e44c1d17447e40929ad03096ccdec33341a010fbcfd3143832698e7ab886c8.jpg", + "table_caption": [ + "Table 1: Retrieval recall on the test set of 3RScan dataset [123] (\\%; target scene ranked within the top 1, 3, and 5 of the retrieved list) and storage requirements (MB) for methods utilizing point clouds $(\\mathcal{P})$ , images $(\\mathcal{I})$ , and other modalities $(\\mathrm{O} = \\{\\mathcal{A}, \\mathcal{S}, \\mathcal{R}\\})$ for map representation. $R$ and $R^t$ represent the recall in the static and temporal scenarios respectively. Additionally, metrics for single-modal methods (CVNet and AnyLoc) reliant on extensive image datasets are presented. The results are reported for scenarios where the target room is chosen from a subset of 10 and 50 candidate scenes." + ], + "table_footnote": [], + "table_body": "
MethodMap modalities10 scenes50 scenesStorage (MB)
PIOR@1@3@5R@1@3@5R@1@3@5R@1@3@5
LidarCLIP [43]✓XX16.341.460.616.339.861.14.711.016.34.110.315.6
LIP-Loc [108]✓XX14.035.857.910.930.052.72.09.114.22.38.615.2
OpenMask3D [113]✓✓X---42.371.585.8---21.138.148.0
SceneGraphLoc (Ours)✓X✓53.681.992.850.576.888.430.250.261.228.246.256.4
SceneGraphLoc (Ours)✓✓✓---81.593.998.0---69.378.684.4
CVNet [63]X✓X---79.291.095.4---66.577.081.7
AnyLoc [55]X✓X---87.994.797.5---80.687.490.0
", + "bbox": [ + 215, + 253, + 785, + 351 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "descriptors. In OpenMask3D, the CLIP embeddings corresponding to each object are calculated and stored. For image-based methods like CVNet and AnyLoc, embeddings for all images in the database are precomputed.", + "bbox": [ + 212, + 383, + 784, + 428 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Metrics. To evaluate the accuracy of a method, we focus on the recall of scene selection. This entails analyzing the scenario where, given a query image and corresponding scene pair $(I, \\mathcal{G}_I)$ , alongside $N - 1$ alternative scenes from the database, an ordering is established for these scenes according to their computed similarity to $I$ as determined by the tested method. The metric Recall@ $K$ is employed to ascertain whether the target scene $\\mathcal{G}_i$ is ranked among the top $K$ scenes in terms of similarity as identified by the evaluated method. Additionally, we will report the inference time and storage requirements.", + "bbox": [ + 212, + 429, + 785, + 549 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Experiments on 3RScan. The 3RScan dataset [123] comprises 1335 annotated indoor scenes, representing 432 distinct rooms, with 1178 scenes (385 rooms) allocated for training and 157 (47 rooms) designated for validation. The training and validation sets include semantically annotated 3D point clouds for each scene, with some captured over extended periods (e.g., several months) showcasing environmental changes. Annotations for graphs within the 3RScan dataset are provided in [125]. Due to the absence of such annotations in the test set, it was excluded from our experiments. Thus, we reorganized the original validation set, allocating 34 scenes (17 rooms) for validation and 123 scenes (30 rooms) for testing. For full reproducibility, we will publish this split.", + "bbox": [ + 212, + 551, + 787, + 703 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "During testing, we examine all 123 scenes of 30 rooms within the test set, selecting query images from each scene. The next step involves matching this image against $N$ scenes (including the target) to ascertain whether the correct one could be identified by a method. This procedure is repeated for every image in each room. In total, all methods are tested on 30462 query images. Furthermore, we evaluate scene selection through two settings $N = 50$ and $N = 10$ . The latter setting emulates a scenario where a pre-selection strategy is employed, for example, utilizing a global scene descriptor. In image-based methods, we use all images from the database and determine the scene based on the retrieved image.", + "bbox": [ + 212, + 704, + 787, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "SceneGraphLoc", + "bbox": [ + 624, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/6f6815ca3de6c8122f9c57dcd4b9818ff5d60eddd9b97d4f4be5d6bee9aa9790.jpg", + "table_caption": [ + "Table 2: Retrieval recall in the temporal scenario on the test set of ScanNet dataset [21] (\\%; target scene ranked within the top 1, 3, and 5 of the retrieved list) and storage requirements (MB) for methods using point clouds $(\\mathcal{P})$ , images $(\\mathcal{I})$ , and other modalities $(S, \\mathcal{R})$ for map representation. The modality $\\mathcal{A}$ is not available in the scene graphs predicted from [130]. The results are reported for scenarios where the target room is chosen from a subset of 10, 50 and the complete set of all (210) scenes." + ], + "table_footnote": [], + "table_body": "
MethodMap modalities10 scenes50 scenesAll scenesStorage (MB)
PIORt@1at3at5Rt@1at3at5Rt@1at3at5
LidarCLIP [43]✓XX19.447.567.64.714.822.25.915.021.90.7
LIP-Loc [108]✓XX10.327.043.61.96.08.11.83.14.01.7
OpenMask3D [113]✓✓X54.984.894.031.351.363.216.527.234.517.8
SceneGraphLoc (Ours)✓X✓54.181.491.929.047.458.013.526.434.29.3
SceneGraphLoc (Ours)✓✓✓78.592.798.361.683.291.653.469.878.79.3
CVNet [63]X✓X96.598.999.692.696.097.089.993.494.6239.1
AnyLoc [55]X✓X98.499.499.896.598.198.695.196.997.45720.3
", + "bbox": [ + 217, + 239, + 785, + 353 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Additionally, the methods are evaluated under both static and temporal conditions. In the static scenario, the target scene graph for a given query image originates from the same scan, albeit from a different sequence, to ensure no image overlap. Conversely, in the temporal scenario, the scene graph is derived from a sequence captured at a different temporal stage than the query, introducing potential environmental changes. We do not show results for methods exploiting images in their maps in the static stage.", + "bbox": [ + 212, + 383, + 782, + 489 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The results are in Table 1. Despite being retrained, LidarCLIP and LIP-Loc are inaccurate, particularly in scenarios involving the selection of the target room from the entire scene set. LIP-Loc barely surpasses random selection. Although LidarCLIP exhibits marginally better accuracy, it remains inferior to alternative methods. The inferiority is because both methods learn shared embedding between entire scenes and query images of totally different objects due to viewpoint differences. OpenMask3D, while achieving better results than LidarCLIP and LIP-Loc, is less accurate than the proposed SceneGraphLoc. SceneGraphLoc, even when excluding the image modality $(\\mathcal{I})$ , outperforms other cross-modal strategies significantly. Incorporating images significantly enhances its performance, positioning it close to that of image-based approaches but with three orders of magnitude smaller storage requirements. Also, the storage of SceneGraphLoc with and without images is the same due to its design of distilling knowledge into fixed-sized embeddings. An example scene is shown in Fig. 5.", + "bbox": [ + 212, + 491, + 785, + 700 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Experiments on ScanNet. To evaluate the generalization ability of our methods in real-world applications when scene graph annotations are not available, we conduct further experiments in the ScanNet dataset [21]. ScanNet encompasses 1613 monocular sequences of room-scale 3D scenes, offering 3D mesh reconstructions alongside the RGBD frame sequences utilized for the reconstructions. Given the absence of scene graph annotations in ScanNet, we run the SceneGraphFusion [130] on the RGBD sequences of scans for 3D reconstruction and scene graph prediction with 3D instance segmentation and object relationships (i.e., graph edges) within these graphs. As the process of scene graph prediction", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Y. Miao et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/e32c99a69d262ee34890f79f40c0103b0c7871071b7d56a89dd56c8f0d592945.jpg", + "table_caption": [ + "Table 3: Average time (ms) of obtaining the query image embedding $(t_{e_q})$ and of the retrieval from 10, 50, and all scenes from the 3RScan [123] and ScanNet [21] datasets." + ], + "table_footnote": [], + "table_body": "
Method3DRScan [123]ScanNet [21]
teqt10retrt50retrteqt10retrt50retrtallretr
LidarCLIP [43]4.10.10.34.90.10.20.6
LIP-Loc [108]2.70.10.24.10.10.20.5
OpenMask3D [113]41.54.87.420.155.41.14.5
SceneGraphLoc (Ours)28.00.31.516.61.33.717.0
CVNet [63]14.39.060.054.010.674.1311.3
AnyLoc [55]658.4354.61826.4243.068.2329.01451.1
", + "bbox": [ + 277, + 184, + 725, + 301 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "uses the RGBD frames of each scan, we avoid using those RGB images to match to the scene graph predicted of the scan. Thus, we only measure recall in the temporal scenario. Additionally, unlike 3RScan, the frame rate of RGBD sequences in ScanNet is high, and motion between consecutive frames is small. Thus, each database image is selected from every 25 consecutive frames in the sequence, for image-based methods [55, 63]. For a fair comparison, all the methods only use the same selected images for training and evaluation.", + "bbox": [ + 212, + 339, + 787, + 446 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "For training, we use the official training set. We divide the official validation set, which includes 312 scenes, into two distinct subsets: the first 100 scenes form our validation set, while the remaining 212 are allocated for testing. To ensure full reproducibility, we will make this split publicly available.", + "bbox": [ + 212, + 450, + 787, + 513 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The results are in Table 2 for scenarios selecting the target room from subsets of 10, 50, and the entire set of 210 scenes. The performance of LIP-Loc shows a similar pattern to that observed on 3RScan, performing only slightly better than random selection. LidarCLIP shows a small improvement in accuracy. OpenMask3D attains an accuracy comparable to our proposed method without incorporating the image modality. Our proposed SceneGraphLoc, when including the image modality in its map, significantly outperforms all cross-modal approaches. Although there remains a gap in accuracy compared with methods that use extensive image collections as maps (such as CVNet and AnyLoc), SceneGraphLoc benefits from a database size three orders-of-magnitude smaller, highlighting its efficiency and effectiveness. We partly attribute this performance gap to the lack of object attributes in the dataset and the inaccurate instance segmentation predicted by [130]. More details can be found in the supp. mat.", + "bbox": [ + 212, + 517, + 787, + 715 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Processing time measured in milliseconds for various methods applied to the 3DRScan and ScanNet datasets, are detailed in Table 3. The computation time required to generate an embedding for the query image $(t_{e_q})$ is notably small across all methods, typically not exceeding a few tens of milliseconds, with the exception of AnyLoc, which runs for nearly a second. The retrieval phase for cross-modal approaches is generally limited to a few tens of milliseconds. However, methods such as CVNet and AnyLoc exhibit slower performance, due to searching through extensive image collections. When tasked with selecting from", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "SceneGraphLoc", + "bbox": [ + 624, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/2650bc5bdad8a365fc6486fb64e4edbfddf3d6a70f5c2ac969b80a0197ed4c79.jpg", + "table_caption": [ + "Table 4: Ablation study performed on the val. split of 3RScan [123], analysing map modalities ( $\\mathcal{P}$ - point cloud, $\\mathcal{I}$ - image, $\\mathcal{A}$ - attributes, $\\mathcal{S}$ - structure, $\\mathcal{R}$ - relationships) and the method (Dino v2 or GCVit) to obtain the image embeddings." + ], + "table_footnote": [], + "table_body": "
Map modalitiesDino v2 [81]GCVit [38]
PIASRR@1@3@5Rt@1@3@5R@1@3@5Rt@1@3@5
✓45.281.993.743.979.591.424.656.076.923.254.777.3
✓✓56.385.695.054.884.095.044.276.991.243.475.189.4
✓✓✓58.487.395.956.585.793.643.375.891.341.572.889.3
✓✓✓✓63.786.895.862.787.496.345.375.590.546.676.290.2
✓---80.296.099.0---69.487.493.7
✓✓---84.797.599.6---73.289.795.9
✓✓✓✓✓---88.597.799.6---72.188.895.7
", + "bbox": [ + 250, + 199, + 759, + 314 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "a large number of images, the processing times of these methods can extend into the range of several hundred milliseconds or even reach upwards of a second.", + "bbox": [ + 212, + 342, + 782, + 372 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Ablation Studies. To understand the impact of the integration of different modalities, we provide an ablation study on the localization performance with multiple combinations of modalities. We use the validation split of the 3DRScan dataset. Additional ablation studies can be found in supplementary material.", + "bbox": [ + 212, + 373, + 784, + 433 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Table 4 displays the coarse localization performance of our method under the incorporation of distinct modalities $(\\mathcal{P},\\mathcal{I},\\mathcal{A},\\mathcal{S},\\mathcal{R})$ within the pipeline. Additionally, the table reports results exploiting various backbones (Dino V2 [81] and GCVit [38]) for the extraction of image features. Similarly to the main experiments, we only show results on the temporal set when the map contains images. Employing Dino V2 for encoding both the query image and images within the map significantly enhances accuracy over using GCVit. The proposed method significantly outperforms both LidarCLIP and LIP-Loc (their results are in Table 1) even when using only point clouds in the map. Each additional modality contributes to the final accuracy, demonstrating that each plays a crucial role in enhancing the final performance when integrated into the pipeline.", + "bbox": [ + 212, + 434, + 787, + 601 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 623, + 359, + 638 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In conclusion, we introduce SceneGraphLoc, a novel method for solving the novel problem of localizing an input image within a 3D scene graph-based multi-modal reference map. This approach outperforms existing cross-modal methods by a large margin. It achieves comparable accuracy to state-of-the-art image-based techniques with significantly lower storage requirements and faster processing speeds. Our experiments across the 3RScan and ScanNet datasets demonstrate the effectiveness of SceneGraphLoc, with the best performance achieved when integrating all proposed modalities. We believe that SceneGraphLoc is a step towards lightweight and efficient localization.", + "bbox": [ + 212, + 654, + 787, + 791 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Y. Miao et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements.", + "text_level": 1, + "bbox": [ + 217, + 143, + 408, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "We would like to thank our colleagues Ganlin Zhang, Sayan Deb Sarkar, and Cathrin Elich for their valuable advice and insightful discussions throughout the course of this research. Their contributions and suggestions greatly enhanced the quality and depth of this work. This work was partially funded by Design++ initiative of ETH Zurich, by the ETH RobotX research grant, the Hasler Stiftung Research Grant via the ETH Zurich Foundation, an ETH AI Center postdoctoral research fellowship and an ETH Zurich Career Seed Award.", + "bbox": [ + 212, + 174, + 787, + 282 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 301, + 321, + 316 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Agia, C., Jatavallabhula, K.M., Khodeir, M., Miksik, O., Vineet, V., Mukadam, M., Paull, L., Shkurti, F.: Taskgraphy: Evaluating robot task planning over large 3D scene graphs. In: Conference on Robot Learning (CoRL) (2022) 4", + "2. Arandjelovic, R., Gronat, P., Torii, A., Pajdla, T., Sivic, J.: NetVLAD: CNN architecture for weakly supervised place recognition. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2016) 2, 3", + "3. Armeni, I., He, Z.Y., Gwak, J., Zamir, A.R., Fischer, M., Malik, J., Savarese, S.: 3D Scene Graph: A Structure for Unified semantics, 3d space, and camera. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2019) 2, 4", + "4. Aubry, M., Russell, B.C., Sivic, J.: Painting-to-3d model alignment via discriminative visual elements. ACM Transactions On Graphics (TOG) (2014) 3, 4", + "5. Aubry, M., Russell, B.C., Sivic, J.: Visual Geo-localization of Non-photographic Depictions via 2D-3D Alignment. In: Large-Scale Visual Geo-Localization (2016) 3", + "6. Balntas, V., Li, S., Prisacariu, V.: RelocNet: Continuous Metric Learning Relocalisation using Neural Nets. In: European Conference on Computer Vision (ECCV) (2018) 3", + "7. Bernreiter, L., Ott, L., Nieto, J., Siegwart, R., Cadena, C.: Spherical multi-modal place recognition for heterogeneous sensor systems. In: International Conference on Robotics and Automation (ICRA) (2021) 4", + "8. Berton, G., Masone, C., Caputo, B.: Rethinking visual geo-localization for large-scale applications. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2022) 3", + "9. Berton, G., Paolicelli, V., Masone, C., Caputo, B.: Adaptive-attentive geolocation from few queries: A hybrid approach. In: IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) (2021) 2, 3", + "10. Bhayani, S., Sattler, T., Barath, D., Beliansky, P., Heikkilä, J., Kukelova, Z.: Calibrated and partially calibrated semi-generalized homographies. In: International Conference on Computer Vision (ICCV) (2021) 3", + "11. Brachmann, E., Krull, A., Nowozin, S., Shotton, J., Michel, F., Gumhold, S., Rother, C.: DSAC - Differentiable RANSAC for Camera Localization. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2017) 3", + "12. Brachmann, E., Rother, C.: Learning Less is More - 6D Camera Localization via 3D Surface Regression. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2018) 3" + ], + "bbox": [ + 225, + 330, + 785, + 840 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "SceneGraphLoc", + "bbox": [ + 624, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "13. Brachmann, E., Rother, C.: Visual camera re-localization from RGB and RGB-D images using DSAC. Transactions on Pattern Analysis and Machine Intelligence (PAMI) (2021) 3", + "14. Brejcha, J., Lukac, M., Hold-Geoffroy, Y., Wang, O., Cadik, M.: LandscapeAR: Large Scale Outdoor Augmented Reality by Matching Photographs with Terrain Models Using Learned Descriptors. In: European Conference on Computer Vision (ECCV) (2020) 3, 4", + "15. Cadik, M., Sykora, D., Lee, S.: Automated outdoor depth-map generation and alignment. Computers & Graphics (2018) 3", + "16. Castle, R., Klein, G., Murray, D.W.: Video-rate localization in multiple maps for wearable augmented reality. In: IEEE International Symposium on Wearable Computers (2008) 3", + "17. Cavallari, T., Bertinetto, L., Mukhoti, J., Torr, P., Golodetz, S.: Let's take this online: Adapting scene coordinate regression network predictions for online RGB-D camera relocalisation. In: International Conference on 3D Vision (3DV) (2019) 3", + "18. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention Mask Transformer for Universal Image Segmentation. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2022) 8", + "19. Csurka, G., Dance, C., Fan, L., Willamowski, J., Bray, C.: Visual categorization with bags of keypoints. In: European Conference on Computer Vision (ECCV) Workshops (2004) 7", + "20. Curless, B., Levoy, M.: A volumetric method for building complex models from range images. In: Annual conference on Computer graphics and interactive techniques (1996) 4", + "21. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2017) 12, 13", + "22. Doan, A.D., Latif, Y., Chin, T.J., Liu, Y., Do, T.T., Reid, I.: Scalable place recognition under appearance change for autonomous driving. In: International Conference on Computer Vision (ICCV) (2019) 2", + "23. Engelmann, F., Manhardt, F., Niemeyer, M., Tateno, K., Tombari, F.: OpenNeRF: Open Set 3D Neural Scene Segmentation with Pixel-Wise Features and Rendered Novel Views. In: International Conference on Learning Representations (ICLR) (2024) 10", + "24. Fan, L., Pang, Z., Zhang, T., Wang, Y.X., Zhao, H., Wang, F., Wang, N., Zhang, Z.: Embracing single stride 3d object detector with sparse transformer. International Conference on Computer Vision and Pattern Recognition (CVPR) (2022) 10", + "25. Gadre, S.Y., Ehsani, K., Song, S., Mottaghi, R.: Continuous scene representations for embodied ai. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2022) 4", + "26. Gao, P., Liang, J., Shen, Y., Son, S., Lin, M.C.: Visual, Spatial, Geometric-Preserved Place Recognition for Cross-View and Cross-Modal Collaborative Perception. In: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) (2023) 4", + "27. Garg, S., Fischer, T., Milford, M.: Where is your place, visual place recognition? International Joint Conference on Artificial Intelligence (IJCAI) (2021) 3", + "28. Garg, S., Suenderhauf, N., Milford, M.: Semantic-geometric visual place recognition: a new perspective for reconciling opposing views. International Journal on Robotics Research (IJRR) (2019) 2, 4" + ], + "bbox": [ + 225, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Y. Miao et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "29. Garg, S., Sünderhauf, N., Dayoub, F., Morrison, D., Cosgun, A., Carneiro, G., Wu, Q., Chin, T.J., Reid, I., Gould, S., et al.: Semantics for robotic mapping, perception and interaction: A survey. Foundations and Trends in Robotics (2020) 4", + "30. Georgakis, G., Karanam, S., Wu, Z., Kosecka, J.: Learning local rgb-to-cad correspondences for object pose estimation. In: International Conference on Computer Vision (ICCV) (2019) 4", + "31. Germain, H., Bourmaud, G., Lepetit, V.: Sparse-to-Dense Hypercolumn Matching for Long-Term Visual Localization. In: International Conference on 3D Vision (3DV) (2019) 3", + "32. Germain, H., Bourmaud, G., Lepetit, V.: S2DNet: Learning Image Features for Accurate Sparse-to-Dense Matching. In: European Conference on Computer Vision (ECCV) (2020) 3", + "33. Ghiasi, G., Gu, X., Cui, Y., Lin, T.Y.: Scaling Open-Vocabulary Image Segmentation with Image-Level Labels. In: European Conference on Computer Vision (ECCV) (2022) 10", + "34. Grabner, A., Roth, P.M., Lepetit, V.: 3D Pose Estimation and 3D Model Retrieval for Objects in the Wild. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2018) 4", + "35. Grelsson, B., Robinson, A., Felsberg, M., Khan, F.S.: GPS-level accurate camera localization with HorizonNet. Journal of Field Robotics (2020) 3", + "36. Gumeli, C., Dai, A., Nießner, M.: ROCA: Robust CAD Model Retrieval and Alignment from a Single Image. arXiv preprint arXiv:2112.01988 (2021) 4", + "37. Hanocka, R., Metzer, G., Giryes, R., Cohen-Or, D.: Point2mesh: A self-prior for deformable meshes. arXiv preprint arXiv:2005.11084 (2020) 4", + "38. Hatamizadeh, A., Yin, H., Heinrich, G., Kautz, J., Molchanov, P.: Global context vision transformers. In: International Conference on Machine Learning (ICML) (2023) 14", + "39. Hausler, S., Jacobson, A., Milford, M.: Multi-process fusion: Visual place recognition using multiple image processing methods. IEEE Robotics and Automation Letters (RA-L) (2019) 2", + "40. Hausler, S., Garg, S., Xu, M., Milford, M., Fischer, T.: Patch-netvlad: Multiscale fusion of locally-global descriptors for place recognition. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2021) 2, 3", + "41. He, K., Zhang, X., Ren, S., Sun, J.: Deep Residual Learning for Image Recognition. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2016) 8", + "42. Heng, L., Choi, B., Cui, Z., Geppert, M., Hu, S., Kuan, B., Liu, P., Nguyen, R., Yeo, Y.C., Geiger, A., et al.: Project Autovision: Localization and 3D Scene Perception for an Autonomous Vehicle with a Multi-Camera System. In: International Conference on Robotics and Automation (ICRA) (2019) 3", + "43. Hess, G., Tondersski, A., Petersson, C., Åström, K., Svensson, L.: Lidarclip or: How i learned to talk to point clouds. In: IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) (2024) 4, 10, 11, 12, 13", + "44. Hodan, T.: Pose Estimation of Specific Rigid Objects. Ph.D. thesis (2021) 4", + "45. Hodan, T., Barath, D., Matas, J.: EPOS: Estimating 6D Pose of Objects With Symmetries. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2020) 4", + "46. Hodan, T., Zabulis, X., Lourakis, M.I.A., Obdrzalek, S., Matas, J.: Detection and fine 3D pose estimation of texture-less objects in RGB-D images. In: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) (2015) 4" + ], + "bbox": [ + 225, + 146, + 784, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "SceneGraphLoc", + "bbox": [ + 624, + 114, + 730, + 127 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "47. Hu, S., Feng, M., Nguyen, R.H.M., Lee, G.H.: CVM-Net: Cross-View Matching Network for Image-Based Ground-to-Aerial Geo-Localization. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2018) 4", + "48. Hu, S., Lee, G.H.: Image-Based GeoLocalization Using Satellite Imagery. International Journal on Computer Vision (IJCV) (2019) 4", + "49. Hughes, N., Chang, Y., Carlone, L.: Hydra: A real-time spatial perception system for 3D scene graph construction and optimization. arXiv preprint arXiv:2201.13360 (2022) 4, 5, 7", + "50. Ibrahimi, S., van Noord, N., Alpherts, T., Worring, M.: Inside out visual place recognition. In: British Machine Vision Conference (2021) 2, 3", + "51. Irschara, A., Zach, C., Frahm, J.M., Bischof, H.: From Structure-from-Motion Point Clouds to Fast Location Recognition. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2009) 3", + "52. Izadi, S., Kim, D., Hilliges, O., Molyneaux, D., Newcombe, R., Kohli, P., Shotton, J., Hodges, S., Freeman, D., Davison, A., et al.: Kinectfusion: real-time 3d reconstruction and interaction using a moving depth camera. In: ACM symposium on User interface software and technology (2011) 4", + "53. Ji, X., Wei, J., Wang, Y., Shang, H., Kneip, L.: Cross-modal Place Recognition in Image Databases using Event-based Sensors. arXiv preprint arXiv:2307.01047 (2023) 4", + "54. Kabalar, J., Wu, S.C., Wald, J., Tateno, K., Navab, N., Tombari, F.: Towards long-term retrieval-based visual localization in indoor environments with changes. IEEE Robotics and Automation Letters (2023) 4", + "55. Keetha, N., Mishra, A., Karhade, J., Jatavallabhula, K.M., Scherer, S., Krishna, M., Garg, S.: Anyloc: Towards universal visual place recognition. IEEE Robotics and Automation Letters (RA-L) (2023) 2, 3, 10, 11, 12, 13", + "56. Kendall, A., Cipolla, R.: Geometric loss functions for camera pose regression with deep learning. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2017) 3", + "57. Kendall, A., Grimes, M., Cipolla, R.: PoseNet: A Convolutional Network for Real-Time 6-DOF Camera Relocalization. In: International Conference on Computer Vision (ICCV) (2015) 3", + "58. Khaliq, A., Ehsan, S., Chen, Z., Milford, M., McDonald-Maier, K.: A holistic visual place recognition approach using lightweight CNNs for significant viewpoint and appearance changes. IEEE Transactions on Robotics (T-RO) (2020) 2", + "59. Kim, H.J., Dunn, E., Frahm, J.M.: Learned contextual feature reweighting for image geolocation. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2017) 3", + "60. Kim, U.H., Park, J.M., Song, T.J., Kim, J.H.: 3-D scene graph: A sparse and semantic representation of physical environments for intelligent agents. IEEE transactions on cybernetics (2019) 4", + "61. Kutulakos, K.N., Seitz, S.M.: A theory of shape by space carving. International Journal on Computer Vision (IJCV) (2000) 4", + "62. Labbe, Y., Carpentier, J., Aubry, M., Sivic, J.: CosyPose: Consistent multi-view multi-object 6D pose estimation. In: European Conference on Computer Vision (ECCV) (2020) 4", + "63. Lee, S., Seong, H., Lee, S., Kim, E.: Correlation verification for image retrieval. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2022) 10, 11, 12, 13" + ], + "bbox": [ + 223, + 146, + 785, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Y. Miao et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "64. Li, Y., Snavely, N., Huttenlocher, D.P., Fua, P.: Worldwide Pose Estimation Using 3D Point Clouds. In: European Conference on Computer Vision (ECCV) (2012) 3", + "65. Lim, H., Sinha, S.N., Cohen, M.F., Uysettendaele, M.: Real-time Image-based 6-DoF Localization in Large-scale Environments. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2012) 3", + "66. Lin, T.Y., Cui, Y., Belongie, S.J., Hays, J.: Learning deep representations for ground-to-aerial geolocation. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2015) 4", + "67. Lin, Z., Zhang, Z., Wang, M., Shi, Y., Wu, X., Zheng, Y.: Multi-modal contrastive representation learning for entity alignment. arXiv preprint arXiv:2209.00891 (2022) 9", + "68. Liu, L., Li, H., Dai, Y.: Efficient Global 2D-3D Matching for Camera Localization in a Large-Scale 3D Map. In: International Conference on Computer Vision (ICCV) (2017) 3", + "69. Liu, L., Li, H., Dai, Y.: Stochastic attraction-repulsion embedding for large scale image localization. In: International Conference on Computer Vision (ICCV) (2019) 2", + "70. Looper, S., Rodriguez-Puigvert, J., Siegwart, R., Cadena, C., Schmid, L.: 3D VSG: Long-term semantic scene change prediction through 3D variable scene graphs. In: International Conference on Robotics and Automation (ICRA) (2023) 4", + "71. Lynen, S., Zeisl, B., Aiger, D., Bosse, M., Hesch, J., Pollefeys, M., Siegwart, R., Sattler, T.: Large-scale, real-time visual-inertial localization revisited. International Journal on Robotics Research (IJRR) (2020) 3", + "72. Lynen, S., Zeisl, B., Aiger, D., Bosse, M., Hesch, J., Pollefeys, M., Siegwart, R., Sattler, T.: Large-scale, real-time visual-inertial localization revisited. International Journal on Robotics Research (IJRR) (2020) 3", + "73. Mescheder, L., Oechsle, M., Niemeyer, M., Nowozin, S., Geiger, A.: Occupancy networks: Learning 3d reconstruction in function space. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2019) 4", + "74. Miao, Y., Armeni, I., Pollefeys, M., Barath, D.: Volumetric Semantically Consistent 3D Panoptic Mapping. arXiv preprint arXiv:2309.14737 (2024) 4", + "75. Miao, Y., Li, C., Li, Z., Yang, Y., Yu, X.: A novel algorithm of ship structure modeling and target identification based on point cloud for automation in bulk cargo terminals. Measurement and Control (2021) 4", + "76. Mihajlovic, M., Weder, S., Pollefeys, M., Oswald, M.R.: DeepSurfels: Learning online appearance fusion. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2021) 4", + "77. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM (2021) 4", + "78. Moreau, A., Piasco, N., Tsishkou, D., Stanciulescu, B., de La Fortelle, A.: LENS: Localization enhanced by neRF synthesis. In: Conference on Robot Learning (CoRL) (2021) 3", + "79. Murez, Z., Van As, T., Bartolozzi, J., Sinha, A., Badrinarayanan, V., Rabinovich, A.: Atlas: End-to-end 3d scene reconstruction from posed images. In: European Conference on Computer Vision (ECCV) (2020) 4", + "80. Oechsle, M., Mescheder, L., Niemeyer, M., Strauss, T., Geiger, A.: Texture fields: Learning texture representations in function space. In: International Conference on Computer Vision (ICCV) (2019) 4" + ], + "bbox": [ + 225, + 146, + 784, + 839 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "SceneGraphLoc", + "bbox": [ + 624, + 114, + 730, + 127 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "81. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023) 7, 8, 14", + "82. Panek, V., Kukelova, Z., Sattler, T.: Meshloc: Mesh-based visual localization. In: European Conference on Computer Vision (ECCV) (2022) 3, 4", + "83. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2019) 4", + "84. Peng, G., Yue, Y., Zhang, J., Wu, Z., Tang, X., Wang, D.: Semantic reinforced attention learning for visual place recognition. In: International Conference on Robotics and Automation (ICRA) (2021) 3", + "85. Peng, G., Zhang, J., Li, H., Wang, D.: Attentional pyramid pooling of salient visual residuals for place recognition. In: International Conference on Computer Vision (ICCV) (2021) 3", + "86. Peng, S., Genova, K., Jiang, C.M., Tagliasacchi, A., Pollefeys, M., Funkhouser, T.: OpenScene: 3D Scene Understanding with Open Vocabularies. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2023) 10", + "87. Peng, S., Niemeyer, M., Mescheder, L., Pollefeys, M., Geiger, A.: Convolutional occupancy networks. In: European Conference on Computer Vision (ECCV) (2020) 4", + "88. Pion, N., Humenberger, M., Csurka, G., Cabon, Y., Sattler, T.: Benchmarking image retrieval for visual localization. In: International Conference on 3D Vision (3DV) (2020) 3", + "89. Plotz, T., Roth, S.: Automatic Registration of Images to Untextured Geometry Using Average Shading Gradients. International Journal on Computer Vision (IJCV) (2017) 3", + "90. Ponimatkin, G., Labbe, Y., Russell, B., Aubry, M., Sivic, J.: Focal length and object pose estimation via render and compare. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2022) 4", + "91. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2017) 4, 6", + "92. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. International Conference on Machine Learning (ICML) (2021) 10", + "93. Ramalingam, S., Bouaziz, S., Sturm, P.F., Brand, M.: SKYLINE2GPS: Localization in urban canyons using omni-skylines. In: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) (2010) 3", + "94. Ravichandran, Z., Peng, L., Hughes, N., Griffith, J., Carlone, L.: Hierarchical representations and explicit memory: Learning effective navigation policies on 3D scene graphs using graph neural networks. In: International Conference on Robotics and Automation (ICRA) (2022) 4", + "95. Rosinol, A., Violette, A., Abate, M., Hughes, N., Chang, Y., Shi, J., Gupta, A., Carlone, L.: Kimera: From SLAM to spatial perception with 3D dynamic scene graphs. International Journal on Robotics Research (IJRR) (2021) 4, 5", + "96. Rosinol, A., Gupta, A., Abate, M., Shi, J., Carlone, L.: 3D dynamic scene graphs: Actionable spatial perception with places, objects, and humans. arXiv preprint arXiv:2002.06289 (2020) 4" + ], + "bbox": [ + 223, + 146, + 785, + 840 + ], + "page_idx": 19 + }, + { + "type": "page_number", + "text": "20", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 19 + }, + { + "type": "header", + "text": "Y. Miao et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 19 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "97. Sarkar, S.D., Miksik, O., Pollefeys, M., Barath, D., Armeni, I.: SGAligner: 3D Scene Alignment with Scene Graphs. International Conference on Computer Vision (ICCV) (2023) 4, 6, 7", + "98. Sarlin, P.E., Cadena, C., Siegwart, R., Dymczyk, M.: From Coarse to Fine: Robust Hierarchical Localization at Large Scale. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2019) 3", + "99. Sarlin, P.E., DeTone, D., Malisiewicz, T., Rabinovich, A.: Superglue: Learning feature matching with graph neural networks. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2020) 3", + "100. Sarlin, P.E., DeTone, D., Yang, T.Y., Avetisyan, A., Straub, J., Malisiewicz, T., Bulò, S.R., Newcombe, R., Kontschieder, P., Balntas, V.: OrienterNet: Visual Localization in 2D Public Maps with Neural Matching. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2023) 2, 4", + "101. Sarlin, P.E., Unagar, A., Larsson, M., Germain, H., Toft, C., Larsson, V., Pollefeys, M., Lepetit, V., Hammarstrand, L., Kahl, F., et al.: Back to the Feature: Learning Robust Camera Localization from Pixels to Pose. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2021) 3", + "102. Sattler, T., Leibe, B., Kobbelt, L.: Efficient & Effective Prioritized Matching for Large-Scale Image-Based Localization. Transactions on Pattern Analysis and Machine Intelligence (PAMI) (2017) 3", + "103. Sattler, T., Zhou, Q., Pollefeys, M., Leal-Taixe, L.: Understanding the limitations of cnn-based absolute camera pose regression. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2019) 3", + "104. Savinov, N., Hane, C., Ladicky, L., Pollefeys, M.: Semantic 3d reconstruction with continuous regularization and ray potentials using a visibility consistency constraint. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2016) 4", + "105. Schonberger, J.L., Pollefeys, M., Geiger, A., Sattler, T.: Semantic visual localization. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2018) 3", + "106. Sepulveda, G., Niebles, J., Soto, A.: A deep learning based behavioral approach to indoor autonomous navigation. In: International Conference on Robotics and Automation (ICRA) (2018) 4", + "107. Shan, Q., Wu, C., Curless, B., Furukawa, Y., Hernandez, C., Seitz, S.M.: Accurate geo-registration by ground-to-aerial image matching. In: International Conference on 3D Vision (3DV) (2014) 3, 4", + "108. Shubodh, S., Omama, M., Zaidi, H., Parihar, U.S., Krishna, M.: Lip-loc: Lidar image pretraining for cross-modal localization. In: IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) (2024) 4, 10, 11, 12, 13", + "109. Sibling, D., Sattler, T., Leibe, B., Kobbelt, L.: SIFT-Realistic Rendering. In: International Conference on 3D Vision (3DV) (2013) 3", + "110. Steiger Mueller, M., Sattler, T., Pollefeys, M., Jutz, B.: Image-to-image translation for enhanced feature matching, image retrieval and visual localization. ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences (2019) 3", + "111. Stuckler, J., Behnke, S.: Multi-resolution surfel maps for efficient dense 3D modeling and tracking. Journal of Visual Communication and Image Representation (2014) 4", + "112. Svarm, L., Enqvist, O., Kahl, F., Oskarsson, M.: City-Scale Localization for Cameras with Known Vertical Direction. Transactions on Pattern Analysis and Machine Intelligence (PAMI) (2017) 3" + ], + "bbox": [ + 218, + 146, + 784, + 839 + ], + "page_idx": 20 + }, + { + "type": "header", + "text": "SceneGraphLoc", + "bbox": [ + 624, + 114, + 730, + 128 + ], + "page_idx": 20 + }, + { + "type": "page_number", + "text": "21", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 20 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "113. Takmaz, A., Fedele, E., Sumner, R.W., Pollefeys, M., Tombari, F., Engelmann, F.: OpenMask3D: Open-Vocabulary 3D Instance Segmentation. In: International Conference on Neural Information Processing Systems (NeurIPS) (2023) 7, 10, 11, 12, 13", + "114. Tewari, A., Thies, J., Mildenhall, B., Srinivasan, P., Tretschk, E., Yifan, W., Lassner, C., Sitzmann, V., Martin-Brualla, R., Lombardi, S., et al.: Advances in neural rendering. In: Computer Graphics Forum (2022) 4", + "115. Tomesek, J., Cadik, M., Brejcha, J.: CrossLocate: Cross-modal Large-scale Visual Geo-Localization in Natural Environments using Rendered Modalities. In: IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) (2022) 3, 4", + "116. Torii, A., Arandjelovic, R., Sivic, J., Okutomi, M., Pajdla, T.: 24/7 place recognition by view synthesis. Transactions on Pattern Analysis and Machine Intelligence (PAMI) (2018) 2", + "117. Torii, A., Taira, H., Sivic, J., Pollefeys, M., Okutomi, M., Pajdla, T., Sattler, T.: Are large-scale 3d models really necessary for accurate visual localization? Transactions on Pattern Analysis and Machine Intelligence (PAMI) (2021) 2", + "118. Valentin, J., Nießner, M., Shotton, J., Fitzgibbon, A., Izadi, S., Torr, P.: Exploiting Uncertainty in Regression Forests for Accurate Camera Relocalization. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2015) 3", + "119. Velicković, P., Cucurull, G., Casanova, A., Romero, A., Lio, P., Bengio, Y.: Graph Attention Networks. In: International Conference on Learning Representations (ICLR) (2018) 7", + "120. Ventura, J., Kukelova, Z., Sattler, T., Baráth, D.: Absolute Pose from One or Two Scaled and Oriented Features. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2024) 3", + "121. Viswanathan, A., Rodrigues Pires, B., Huber, D.F.: Vision based robot localization by ground to satellite matching in GPS-denied situations. In: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) (2014) 4", + "122. Walch, F., Hazirbas, C., Leal-Taixe, L., Sattler, T., Hilsenbeck, S., Cremers, D.: Image-Based Localization Using LSTMs for Structured Feature Correlation. In: International Conference on Computer Vision (ICCV) (2017) 3", + "123. Wald, J., Avetisyan, A., Navab, N., Tombari, F., Niefsner, M.: Rio: 3d object instance re-localization in changing indoor environments. In: International Conference on Computer Vision (ICCV) (2019) 8, 9, 10, 11, 13, 14", + "124. Wald, J., Dhamo, H., Navab, N., Tombari, F.: Learning 3d semantic scene graphs from 3d indoor reconstructions. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2020) 2, 4", + "125. Wald, J., Dhamo, H., Navab, N., Tombari, F.: Learning 3d semantic scene graphs from 3d indoor reconstructions. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2020) 11", + "126. Wang, S., Kannala, J., Barath, D.: DGC-GNN: Descriptor-free Geometric-Color Graph Neural Network for 2D-3D Matching. International Conference on Computer Vision and Pattern Recognition (CVPR) (2023) 3", + "127. Warburg, F., Hauberg, S., Lopez-Antequera, M., Gargallo, P., Kuang, Y., Civera, J.: Mapillary street-level sequences: A dataset for lifelong place recognition. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2020) 2, 3" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 21 + }, + { + "type": "page_number", + "text": "22", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 21 + }, + { + "type": "header", + "text": "Y. Miao et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 21 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "128. Weder, S., Schonberger, J.L., Pollefeys, M., Oswald, M.R.: Neuralfusion: Online depth fusion in latent space. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2021) 4", + "129. Workman, S., Souvenir, R., Jacobs, N.: Wide-Area Image Geolocation with Aerial Reference Imagery. In: International Conference on Computer Vision (ICCV) (2015) 4", + "130. Wu, S.C., Wald, J., Tateno, K., Navab, N., Tombari, F.: Scenegraphfusion: Incremental 3d scene graph prediction from rgb-d sequences. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2021) 4, 7, 12, 13", + "131. Ying, Z., Yuan, X., Yang, B., Song, Y., Xu, Q., Zhou, F., Sheng, W.: RP-SG: Relation Prediction in 3D Scene Graphs for Unobserved Objects Localization. IEEE Robotics and Automation Letters (RA-L) (2023) 4", + "132. Zaffar, M., Garg, S., Milford, M., et al.: Vpr-bench: An open-source visual place recognition evaluation framework with quantifiable viewpoint and appearance change. International Journal on Computer Vision (IJCV) (2021) 2", + "133. Zeisl, B., Sattler, T., Pollefeys, M.: Camera pose voting for large-scale image-based localization. In: International Conference on Computer Vision (ICCV) (2015) 3", + "134. Zhang, C., Yu, J., Song, Y., Cai, W.: Exploiting edge-oriented reasoning for 3D point-based scene graph analysis. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2021) 4", + "135. Zhang, G., Larsson, V., Barath, D.: Revisiting Rotation Averaging: Uncertainties and Robust Losses. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2023) 3", + "136. Zhang, S., Hao, A., Qin, H.: Knowledge-inspired 3D scene graph prediction in point cloud. International Conference on Neural Information Processing Systems (NeurIPS) (2021) 4", + "137. Zhang, W., Kosecka, J.: Image based localization in urban environments. In: International Symposium on 3D Data Processing, Visualization, and Transmission (2006) 3", + "138. Zhang, Z., Sattler, T., Scaramuzza, D.: Reference Pose Generation for Long-term Visual Localization via Learned Features and View Synthesis. International Journal on Computer Vision (IJCV) (2020) 3", + "139. Zhao, L., Gatsis, K., Papachristodoulou, A.: Stable and Safe Reinforcement Learning via a Barrier-Lyapunov Actor-Critic Approach. In: IEEE Conference on Decision and Control (CDC) (2023) 3", + "140. Zhao, L., Miao, K., Gatsis, K., Papachristodoulou, A.: Stable and safe human-aligned reinforcement learning through neural ordinary differential equations. arXiv preprint arXiv:2401.13148 (2024) 3", + "141. Zheng, E., Wu, C.: Structure From Motion Using Structure-Less Resection. In: International Conference on Computer Vision (ICCV) (2015) 3", + "142. Zhou, Q., Agostinho, S., Osep, A., Leal-Taixe, L.: Is Geometry Enough for Matching in Visual Localization? In: European Conference on Computer Vision (ECCV) (2022) 2, 3", + "143. Zurbrügg, R., Liu, Y., Engelmann, F., Kumar, S., Hutter, M., Patil, V., Yu, F.: ICGNet: A Unified Approach for Instance-Centric Grasping. In: International Conference on Robotics and Automation (ICRA) (2024) 2" + ], + "bbox": [ + 215, + 146, + 785, + 797 + ], + "page_idx": 22 + }, + { + "type": "header", + "text": "SceneGraphLoc", + "bbox": [ + 624, + 114, + 730, + 128 + ], + "page_idx": 22 + }, + { + "type": "page_number", + "text": "23", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 22 + } +] \ No newline at end of file diff --git a/2024/SceneGraphLoc_ Cross-Modal Coarse Visual Localization on 3D Scene Graphs/ca6ad507-bc20-40c4-b703-2093582a66a7_model.json b/2024/SceneGraphLoc_ Cross-Modal Coarse Visual Localization on 3D Scene Graphs/ca6ad507-bc20-40c4-b703-2093582a66a7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d31127d447eb9fc14c082b1ed329414c49c8b68a --- /dev/null +++ b/2024/SceneGraphLoc_ Cross-Modal Coarse Visual Localization on 3D Scene Graphs/ca6ad507-bc20-40c4-b703-2093582a66a7_model.json @@ -0,0 +1,3414 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.25, + 0.141, + 0.756, + 0.187 + ], + "angle": 0, + "content": "SceneGraphLoc: Cross-Modal Coarse Visual Localization on 3D Scene Graphs" + }, + { + "type": "text", + "bbox": [ + 0.265, + 0.212, + 0.74, + 0.244 + ], + "angle": 0, + "content": "Yang Miao1, Francis Engelmann1,2, Olga Vysotska1, Federico Tombari2,3, Marc Pollefeys1,4, and Daniel Béla Baráth1" + }, + { + "type": "text", + "bbox": [ + 0.346, + 0.255, + 0.657, + 0.27 + ], + "angle": 0, + "content": "1ETH Zurich 2Google 3TU Munich 4Microsoft" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.295, + 0.785, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.445, + 0.788, + 0.473 + ], + "angle": 0, + "content": "Fig. 1: SceneGraphLoc addresses the novel problem of localizing a query image in a database of 3D scenes represented as compact multi-modal 3D scene graphs." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.499, + 0.747, + 0.707 + ], + "angle": 0, + "content": "Abstract. We introduce the task of localizing an input image within a multi-modal reference map represented by a collection of 3D scene graphs. These scene graphs comprise multiple modalities, including object-level point clouds, images, attributes, and relationships between objects, offering a lightweight and efficient alternative to conventional methods that rely on extensive image databases. Given these modalities, the proposed method SceneGraphLoc learns a fixed-sized embedding for each node (i.e., representing object instances) in the scene graph, enabling effective matching with the objects visible in the input query image. This strategy significantly outperforms other cross-modal methods, even without incorporating images into the map representation. With images, SceneGraphLoc achieves performance close to that of state-of-the-art techniques depending on large image databases, while requiring three orders-of-magnitude less storage and operating orders-of-magnitude faster. Code and models are available at https://scenegraphloc.github.io." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.72, + 0.721, + 0.735 + ], + "angle": 0, + "content": "Keywords: Coarse Localization \\(\\cdot\\) 3D Scene Graph \\(\\cdot\\) Multi-modality" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.762, + 0.376, + 0.779 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Coarse visual localization, or place recognition, is a fundamental component in computer vision and robotics applications, defined as the task of identifying the approximate location where a query image was taken, given a certain tolerance" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.127 + ], + "angle": 0, + "content": "Y. Miao et al." + }, + { + "type": "image", + "bbox": [ + 0.242, + 0.146, + 0.761, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.294, + 0.788, + 0.419 + ], + "angle": 0, + "content": "Fig. 2: Overview. The training phase is represented by orange arrows, while blue arrows denote the inference phase. During training, a query image and its associated 3D scene graph form a positive sample within a contrastive learning framework, where negative samples are generated by associating scene graphs of different scenes with the same query image. The objective is to learn the embeddings of both the graph and the image so that embeddings of the positive pair are drawn closer, whereas those of the negative pair are pushed apart. In the inference phase, the task involves assigning the correct scene graph to a given query image from a selection of multiple graphs, achieved by optimizing the cosine similarity between their embeddings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.469, + 0.788, + 0.576 + ], + "angle": 0, + "content": "level [2,9,22,28,39,40,50,58,69,116,117,127,132,143]. This capability is crucial for estimating the state of robots and is widely utilized in autonomous, unmanned aerial, terrestrial, and underwater vehicles, as well as AR/VR devices. The task is typically approached as an image retrieval problem, where the image to be localized is compared against a large database of posed images and, optionally, a 3D reconstruction of the scene. The most similar images retrieved from the database are used to estimate the precise location of the query image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.58, + 0.788, + 0.716 + ], + "angle": 0, + "content": "The challenge with current state-of-the-art image-based coarse localization methods, such as AnyLoc [55], is their dependency on extensive image databases, which are not only storage-heavy but also slow to query, despite optimizations through hashing and other tricks. Moreover, these methods typically necessitate that the query and database share the same modality, limiting the scope of their application. Cross-modal approaches, such as [100,142], which attempt to bridge different types of data, often restrict their scope to connecting two modalities at a time (e.g., image-to-point cloud or image-to-bird's eye view map), one for the query and one for the database, thus narrowing their potential applications." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "This paper addresses the novel challenge of localizing a query image within a database that is represented not by conventional images but by the 3D scene graph [3,124] that integrates a diverse set of modalities, including point clouds, images, semantics, object attributes, and relationships. We tackle this problem by learning to map these modalities into a unified embedding space, thus allowing us to represent indoor scenes compactly through their objects (e.g., table and wall). This method enables the creation of small, efficient databases and significantly accelerates the coarse localization process." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.625, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SceneGraphLoc" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.147, + 0.734, + 0.162 + ], + "angle": 0, + "content": "Contributions. The primary contributions of this paper are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.172, + 0.786, + 0.201 + ], + "angle": 0, + "content": "1. Introducing a novel problem: cross-modal localization of a query image within 3D scene graphs incorporating a mixture of modalities." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.202, + 0.786, + 0.232 + ], + "angle": 0, + "content": "2. SceneGraphLoc, a new method for the coarse localization of an input image given a reference map represented by a database of 3D scene graphs." + }, + { + "type": "list", + "bbox": [ + 0.222, + 0.172, + 0.786, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.241, + 0.786, + 0.318 + ], + "angle": 0, + "content": "Even without incorporating images into the map, SceneGraphLoc largely outperforms other cross-modal methods on two large-scale, real-world indoor datasets. With images, SceneGraphLoc achieves performance close to that of state-of-the-art image-based methods while requiring three orders-of-magnitude less storage and operating orders-of-magnitude faster. The method is visualized in Fig. 1." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.34, + 0.388, + 0.357 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.372, + 0.786, + 0.538 + ], + "angle": 0, + "content": "Localization, the process of determining the position and orientation of an agent within a pre-built map, is pivotal across various domains such as mobile robots [42, 65, 139, 140], and augmented reality [16, 71]. The differentiation in localization techniques arises from their scene representation methods - be it through explicit 3D models [31, 32, 51, 64, 68, 72, 98, 99, 101, 102, 105, 112, 133, 135], sets of posed images [10, 88, 137, 141], or implicitly via neural network weights [6, 11-13, 17, 56, 57, 78, 118, 122] - and their approach to camera pose estimation, whether by 2D-3D [31, 32, 98, 99, 102, 112, 120, 126, 133] or 2D-2D [10, 141] matches, or through a composite of base poses [56, 57, 78, 88, 103, 122]. In practice, localization comprises two main steps: a coarse and precise stage. Here, we focus on the coarse step, finding potential locations of a query image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.539, + 0.786, + 0.689 + ], + "angle": 0, + "content": "Coarse Localization (or place recognition) is often cast as an image retrieval problem [2,8,9,27,40,50,55,59,84,85,127] that consists of two phases. In the offline indexing phase, a reference map (image or point cloud database) is gathered. In the online retrieval phase, a query image - captured during a future traverse of the environment - is localized coarsely by retrieving the closest match to this image in the reference map. Recent methods perform the retrieval using learned embeddings that are produced by a feature extraction backbone equipped with a head that implements some form of aggregation or pooling, the most notable being NetVLAD [2]. While these methods achieve impressive results, they are limited to a single modality (e.g., images) and require large databases." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.786, + 0.84 + ], + "angle": 0, + "content": "Localization using multi-modal data. While dense mesh models are not as widely adopted as sparse Structure-from-Motion-based approaches, they have nonetheless been the focus of considerable research efforts [4,5,14,15,35,82,89,93, 107,109,110,115,138,142]. The body of prior work can be broadly segmented into two main strategies: The first entails the precise alignment of actual images with three-dimensional models (which may be coarse) through applying specialized techniques such as ones using contours [89] or skylines [93]. The second strategy emphasizes the identification and matching of local image features [4,5,14,35, 107,109,110,115,126,138,142], a method that has gained traction for its ability to match real-world images with non-photorealistic renderings of colored meshes, or" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "Y. Miao et al." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.193 + ], + "angle": 0, + "content": "even meshes without color [14,82,115]. CAD and other models are also commonly used by object pose estimation [4, 30, 34, 36, 44-46, 62, 90]. Image to LiDAR localization [7,26,43,108] is also relevant, especially in robotics applications." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.211, + 0.785, + 0.301 + ], + "angle": 0, + "content": "Another variant of multi-modal data localization involves cross-view matching. This technique determines the camera position by finding correspondences between a ground-level query image and a two-dimensional bird's eye view map, such as a satellite image or a semantic landscape map [47,48,66,100,107,121,129]. Other cross-modal techniques were also proposed to involve semantics [28, 29] and event cameras [53] in image-based localization." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.321, + 0.785, + 0.396 + ], + "angle": 0, + "content": "These approaches, while demonstrating promising localization results, are limited to interactions between two modalities - one for the query and one for the reference database. Our approach, in contrast, seeks the cross-modal coarse localization of a query image in a database composed of multiple modalities, e.g., 3D point clouds, images, semantics, object attributes, and relationships." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.414, + 0.785, + 0.58 + ], + "angle": 0, + "content": "Scene Representation, encapsulating various scene attributes has evolved significantly, yielding diverse surface representations from explicit forms (3D point clouds [75,91], meshes [37], surfels [111]) to implicit ones (occupancy [61], signed distance functions [20,52]). The advent of neural representations has introduced novel means of encoding geometry [73,83,87,128], appearance [76,80], and semantics [74,79]. A comprehensive review is provided by Tewari et al. [114]. The integration of directions/rays [77] and visibility encoding in surface reconstruction [104] has further enriched this domain. Armeni et al. [3] introduced the 3D scene graph structure as a multi-layer representation of a scene that captures geometry, semantics, objects, and camera poses in a unified manner. Subsequent efforts [96,124] have further advanced 3D scene graph learning and structure." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.599, + 0.785, + 0.84 + ], + "angle": 0, + "content": "The increasing interest in 3D scene graphs [3, 54, 60, 96, 124] underscores their potential as structured, rich descriptors for real-world scenes. Methods range from online incremental construction [49, 130] to offline generation from RGB-D imagery [3, 95, 124], and approaches for scene graph prediction [134, 136]. Their application spans embodied AI [94, 95, 106], task completion [1, 25], variability estimation [70], and SLAM [49, 95]. Recent studies like [131] introduce frameworks for localizing unseen objects by utilizing 3D scene graphs and graph neural networks for relation prediction, showcasing the utility of scene graphs in enhancing spatial understanding. Similarly, [97] offers new perspectives on 3D scene alignment, employing node matching within overlapping scene graphs to facilitate precise 3D map alignment. Kabalar et al. [54] assume that the query image has been coarsely localized and leverages a scene graph to identify dynamic objects and for precise localization with image features. Despite these significant advancements underscoring the value of scene graphs, their potential in multimodal localization remains largely untapped. In this paper, we use a scene graph representation of the map in which we aim to localize a query image." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.625, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SceneGraphLoc" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.654, + 0.163 + ], + "angle": 0, + "content": "3 Visual Localization with 3D Scene Graph" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.177, + 0.788, + 0.344 + ], + "angle": 0, + "content": "Problem Statement. Let us assume that we are provided with a pre-constructed map of the environment, denoted as \\(\\mathcal{G}\\), which is represented as a set of \\(N \\in \\mathbb{N}^+\\) 3D scene graphs \\(\\mathcal{G}_i = (\\mathcal{V}_i, \\mathcal{E}_i)\\), such that \\(\\mathcal{G} = \\{\\mathcal{G}_i\\}_{i \\in [0, N)}\\). Having separate graphs \\(\\mathcal{G}_i\\) is analogous to the hierarchy levels presented in [49, 95], each representing a group of object instances constituting a place like a room or building. Moreover, this task can also be straightforwardly redefined as subgraph selection. Vertices \\(v \\in \\mathcal{V}_i\\) symbolize instances of objects (e.g., chairs, tables) and large instances of semantic categories (e.g., walls, ground) within the scene, \\(i \\in [0, N)\\). Let \\(\\mathcal{V} = \\bigcup_{i \\in [0, N)} \\mathcal{V}_i\\) aggregate all objects across the scenes. Edges \\(\\mathcal{E}_i = \\{(v_j^i, v_k^i) | v_j^i, v_k^i \\in \\mathcal{V}_i\\}\\) delineate the relationships between objects, such as \"nearby\", \"standing on\", and \"attached to\"." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.346, + 0.8, + 0.42 + ], + "angle": 0, + "content": "For each vertex \\( v \\), we introduce \\( M \\in \\mathbb{N}^{+} \\) map modalities \\( f_{j}: \\mathcal{V} \\to \\mathcal{M}_{j} \\) for \\( j \\in [0, M) \\), where \\( f_{j} \\) maps vertex \\( v \\) to the \\( j^{th} \\) modality \\( \\mathcal{M}_{j} \\) that may be \\( \\mathcal{M}_{j} \\in \\{\\text{position}, \\text{orientation}, \\text{pointcloud}, \\text{semantic category}, \\text{image}, \\text{attribute}\\} \\). While this paper focuses on these modalities, the set is easily extendable by incorporating additional ones, such as textual description or floor plan." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.421, + 0.788, + 0.482 + ], + "angle": 0, + "content": "Given \\(\\mathcal{G}\\) and an input query image \\(I\\), the objective is to identify the scene graph \\(\\mathcal{G}_i\\) corresponding to the space depicted in image \\(I\\). Note that while we focus on having an input image in this paper, the method can be modified to other modalities as well, e.g., depth image. Formally, we aim to resolve problem:" + }, + { + "type": "equation", + "bbox": [ + 0.381, + 0.496, + 0.785, + 0.522 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} _ {i ^ {*}} = \\underset {i \\in [ 0, N)} {\\arg \\max } \\llbracket \\text {c o n t a i n s} (\\mathcal {G} _ {i}, \\mathbf {p} _ {I}) \\rrbracket , \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.528, + 0.788, + 0.589 + ], + "angle": 0, + "content": "where \\(\\mathbf{p}_I \\in \\mathbb{R}^3\\) is the unknown 3D position of the image, and \\([.]\\) is the Iverson bracket which equals to 1 if the condition inside holds and 0 otherwise. It is important to note that our objective is coarse localization, opting for the selection of \\(\\mathcal{G}_i\\) without needing precise estimation of \\(\\mathbf{p}_I\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.59, + 0.788, + 0.651 + ], + "angle": 0, + "content": "To this end, the optimization problem can be reformulated to incorporate the chirality constraint, asserting that if an object is visible in image \\( I \\), it must be positioned in front of the camera in 3D space, unobstructed by any entities (e.g., walls). Consequently, the problem becomes:" + }, + { + "type": "equation", + "bbox": [ + 0.248, + 0.662, + 0.785, + 0.695 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} _ {i ^ {*}} = \\underset {i \\in [ 0, N)} {\\arg \\max } \\sum_ {o _ {I} \\in \\mathcal {O} _ {I}} \\llbracket \\text {v i s i b l e} (\\mathcal {G} _ {i}, o _ {I}) \\rrbracket \\approx \\underset {i \\in [ 0, N)} {\\arg \\max } \\sum_ {o _ {I} \\in \\mathcal {O} _ {I}} \\log P (o _ {I} \\mid \\mathcal {G} _ {i}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.75 + ], + "angle": 0, + "content": "where \\(\\mathcal{O}_I\\) is the object set visible in image \\(I\\), object \\(o_I \\subseteq \\{(x,y) \\in I\\}\\) is a set of pixels in the image \\((o_I \\in \\mathcal{O}_I)\\), visible: \\(\\mathcal{G}_i \\times \\mathcal{O}_I \\to \\{0,1\\}\\) indicates if an object is visible in a scene graph, \\(P(o_I \\mid \\mathcal{G}_i)\\) is the probability of \\(o_I\\) stemming from \\(\\mathcal{G}_i\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Function visible can be approached as an indicator of whether object \\( o_{I} \\) appears in graph \\( G_{i} \\). It holds if and only if there exists a \\( v \\in \\mathcal{V}_{i} \\) such that \\( v \\) represents the same object as \\( o_{I} \\). Therefore, this problem can be approached as matching a set of image pixels \\( (o_{I}) \\) to a scene graph node \\( (v) \\) that comprises a set of modalities. In the next sections, we will describe a way to learn a unified embedding space for both \\( o_{I} \\) and \\( v \\) such that they become matchable." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.128 + ], + "angle": 0, + "content": "Y. Miao et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "Proposed Pipeline is shown in Fig. 2. It consists of two concurrent stages: the first one generates object embeddings \\( e_{q} \\in \\mathbb{R}^{D} \\) from patches \\( q \\in \\mathcal{Q}_I \\) within the query image \\( I \\), and the second derives node embeddings \\( e_{v} \\in \\mathbb{R}^{D} \\) for nodes \\( v \\in \\mathcal{V}_i \\) in the scene graph \\( \\mathcal{G}_i \\). The training objective is to make \\( \\delta(e_{q}, e_{v}) = 0 \\) if and only if the object associated with node \\( v \\) is directly visible (i.e., neither occluded nor outside the camera frustum) through the image patch \\( q \\). Here, \\( \\delta : \\mathbb{R}^D \\times \\mathbb{R}^D \\to \\mathbb{R} \\) denotes inverse cosine similarity normalized to [0,1]." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.253, + 0.787, + 0.283 + ], + "angle": 0, + "content": "After generating \\( e_{q} \\) and \\( e_{v} \\), the model performs nearest neighbor matching (NN) for each patch in each scene graph \\( \\mathcal{G}_{i} \\), assigning node \\( v \\) to \\( q \\) such that" + }, + { + "type": "equation", + "bbox": [ + 0.393, + 0.289, + 0.786, + 0.315 + ], + "angle": 0, + "content": "\\[\n\\mathrm {N N} (q, \\mathcal {V} _ {i}) = \\underset {v \\in \\mathcal {V} _ {i}} {\\arg \\min } \\delta \\left(e _ {q}, e _ {v}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.319, + 0.788, + 0.396 + ], + "angle": 0, + "content": "Through matching, we establish patch-to-node correspondences \\(\\mathcal{C}_i = \\{(q,v) \\mid v = \\mathrm{NN}(q,\\mathcal{V}_i) \\in \\mathcal{V}_i, q \\in \\mathcal{Q}_I\\}\\). Based on \\(\\mathcal{C}_i\\), we devise an image-to-graph similarity score enabling us to deduce whether image \\(I\\) corresponds to the space represented by \\(\\mathcal{G}_i\\). Finally, potential coarse locations of \\(I\\) are selected by maximizing the similarity score across the stored scene graphs as depicted in Eq 2." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.413, + 0.591, + 0.429 + ], + "angle": 0, + "content": "3.1 Object Embeddings in the Scene Graph" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.433, + 0.573, + 0.553 + ], + "angle": 0, + "content": "This section aims to obtain an embedding for each node within graph \\( \\mathcal{G}_i \\), encapsulating information from all available modalities. Our method builds upon the method of Sarkar et al. [97], with enhancements to include the image modality and to distill a unified embedding from all modalities rather than merely concatenating separate embeddings as in [97]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.554, + 0.573, + 0.599 + ], + "angle": 0, + "content": "Scene graphs are conceptualized as multimodal knowledge graphs, similar to those used in entity alignment, treating semantic and geometric" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.787, + 0.735 + ], + "angle": 0, + "content": "information as distinct modalities. The objective is to learn a joint multi-modal embedding from the individual modal encodings (uni-modal), ensuring nodes corresponding to the same object instance across different graphs are closely positioned. This involves the creation of uni-modal embeddings for the three primary types of 3D scene graph information: object embeddings encoding nodes in \\(\\mathcal{V}\\), structure embeddings \\(\\mathcal{S}\\) representing edges in \\(\\mathcal{E}\\) as a structured graph, and two meta modalities encoding attributes \\((\\mathcal{A})\\) and relationships \\((\\mathcal{R})\\) between objects as one-hot vectors. These uni-modal embeddings are then combined in a weighted manner and optimized jointly through knowledge distillation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.787, + 0.78 + ], + "angle": 0, + "content": "Each of these modalities is processed separately to generate uni-modal embeddings, which are subsequently integrated to model complex inter-modal interactions within the joint embedding space." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Object Embedding. Node \\( v \\in \\mathcal{V} \\) may contain multiple modalities, such as point cloud \\( (\\mathcal{P}) \\) and image \\( (\\mathcal{I}) \\). Point clouds contain rich geometric information about objects. The point cloud corresponding to each \\( v \\in \\mathcal{V} \\) is inputted to the object encoder. We employ the PointNet architecture [91] as the" + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.439, + 0.785, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.583, + 0.531, + 0.788, + 0.587 + ], + "angle": 0, + "content": "Fig.3: The embedding of image modality \\(\\mathcal{I}\\) for each object. The image crops of a pillow are shown as an example." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.625, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SceneGraphLoc" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.314 + ], + "angle": 0, + "content": "object encoder to extract the geometric feature \\( e_v^{\\mathcal{P}} \\) for every node. Furthermore, we integrate multi-level and multi-view visual embeddings to enrich the graph encoder with a more nuanced understanding of image information. The visual embedding pipeline is visualized in Fig. 3. For each node \\( v \\) denoting a 3D object, the top \\( K_{view} \\in \\mathbb{N}^{+} \\)images with largest visibility of \\( v \\) is selected: \\( \\{I_{\\mathrm{db},k}^v \\mid k = [0,K)\\} \\subseteq \\mathcal{I}_{\\mathrm{db}} \\) from the database \\( \\mathcal{I}_{\\mathrm{db}} \\). We can define an ordering over these images such that \\( \\phi(I_{\\mathrm{db},0}^v,v) \\geq \\phi(I_{\\mathrm{db},1}^v,v) \\geq \\ldots \\geq \\phi(I_{\\mathrm{db},K_{view}-1}^v,v) \\), with the visibility function \\( \\phi \\) quantifying the extent of node \\( v \\) observed in each image through pixel count. Visibility check is implemented by projecting the 3D model of \\( v \\) to each posed image in \\( \\mathcal{I}_{\\mathrm{db}} \\), which are usually available when constructing the scene graph [49, 130]. Parameter \\( K_{view} = 10 \\) in all our experiments." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.32, + 0.788, + 0.532 + ], + "angle": 0, + "content": "Drawing inspiration from OpenMask3D [113], for any given view \\( I_{\\mathrm{db}}^v \\) of object \\( v \\), initial steps include calculating the bounding box \\( b_{v,0} \\) of \\( v \\) within the image, followed by the generation of multi-level bounding boxes \\( \\{b_{v,l} \\mid l \\in [0,L)\\} \\) through iterative enlargement of \\( b_{v,0} \\). This enlargement strategy aims to capture contextual information around object \\( v \\). Subsequently, Dino V2 [81] processes the image crops defined by \\( b_{v,l} \\), extracting multi-level features \\( \\{f_l \\mid l \\in [0,L)\\} \\). For each image \\( I_{\\mathrm{db},k}^v \\), an average pooling operation aggregates these multi-level crop features into a singular feature vector \\( f = \\operatorname{avg\\_pool}\\{f_l \\mid l \\in [0,L)\\} \\). The final step of this process involves the application of a Transformer encoder, which incorporates image poses as positional encodings. This step synthesizes multiview object tokens into a cohesive visual embedding \\( e_v^{\\mathcal{I}} \\), effectively integrating the diverse perspectives and levels of contextual information pertaining to each object within the database. Please note that the image database does not necessarily have to be stored after distilling the object embeddings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.539, + 0.788, + 0.691 + ], + "angle": 0, + "content": "Structure Embedding. 3D Scene Graphs encapsulate the object relationships, which we exploit to encode their spatial configuration. This relational data is represented through a structure graph, where node features embody the relative translations between object instances, and edges denote these relationships. The relative translation is determined by calculating the distance from the object instance to any other object in the scene. To encapsulate this structural information within \\(\\mathcal{G}_i\\), a Graph Attention Network (GAT) [119] is utilized, with the weight matrix constrained to a diagonal form to reduce computational demands and enhance model scalability. Following the method in [97], the structural embedding \\(e_v^S\\) is derived from the final layer of a two-layer GAT model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.697, + 0.788, + 0.789 + ], + "angle": 0, + "content": "Meta Embeddings. In addition to geometry and structure, the object attributes and the inter-object relationships are captured in two distinct embeddings, \\( e_v^{\\mathcal{R}} \\) and \\( e_v^{\\mathcal{A}} \\). The relationships an object maintains with others are conceptualized as a bag-of-words [19] feature vector, which is input to a feed-forward neural network, distilled in relational embedding \\( e_v^{\\mathcal{R}} \\). A similar approach is employed for the attributes associated with the objects, producing embedding \\( e_v^{\\mathcal{A}} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Joint Embedding. Similar to [97], we concatenate uni-modal features to a compact representation for each object \\( v \\). Contrary to [97], we encapsulate it in a multi-layer perceptron (MLP) to learn an embedding with size independent of" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.128 + ], + "angle": 0, + "content": "Y. Miao et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.744, + 0.163 + ], + "angle": 0, + "content": "the available modalities, fusing information from all. Embedding \\(e_v\\) is as:" + }, + { + "type": "equation", + "bbox": [ + 0.351, + 0.172, + 0.787, + 0.215 + ], + "angle": 0, + "content": "\\[\ne _ {v} = \\operatorname {M L P} \\left(\\oplus_ {k \\in \\mathcal {K}} \\left[ \\frac {\\exp (w _ {k})}{\\sum_ {j \\in \\mathcal {K}} \\exp (w _ {j})} e _ {v} ^ {k} \\right]\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.224, + 0.784, + 0.284 + ], + "angle": 0, + "content": "where \\(\\oplus\\) is the concatenation operator, \\(\\mathcal{K} = \\{\\mathcal{P},\\mathcal{I},\\mathcal{S},\\mathcal{R},\\mathcal{A}\\}\\), and \\(w_{m}\\) is a trainable attention weight for each modality \\(k\\in \\mathcal{K}\\). A two-layer MLP is applied to map the dimension of the concatenated multi-modal descriptors from \\(D^{\\mathcal{K}}\\) to \\(D\\). We apply \\(L_{2}\\) normalization to each uni-modal feature before concatenation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.285, + 0.785, + 0.361 + ], + "angle": 0, + "content": "In practice, these independent modalities are only required and used in the mapping phase of the procedure. This stage involves the construction of an environmental map in the form of a 3D scene graph, during which each node \\( v \\) is distilled into an embedding \\( e_v \\). During localization, we can ignore independent modalities and use only the fixed-sized embeddings \\( e_v \\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.383, + 0.593, + 0.399 + ], + "angle": 0, + "content": "3.2 Object Embeddings in the Query Image" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.408, + 0.789, + 0.619 + ], + "angle": 0, + "content": "To solve the optimization problem in Eq. 2, we need to find object instances \\( o_{I} \\in \\mathcal{O}_{I} \\) in query image \\( I \\). A straightforward approach to do so would be to apply a 2D panoptic segmentation algorithm, e.g., Mask2Former [18]. However, we noticed in our experiments that such an approach is susceptible to inaccuracies and failures (over- and under-segmentation) in the segmentation, severely affecting the accuracy. Thus, we approach the problem by a visual Transformer (ViT), breaking up the image into rectangular patches \\( q \\in \\mathcal{Q}_I \\) and distilling an independent embedding for each patch \\( q \\) based on the object visible in \\( q \\). We use Dino V2 [81] as a backbone to obtain patch-level features. These features are passed through a patch encoder trained to create embeddings from the patch features considering the objects visible from each \\( q \\). For this encoder, we use a 4-layer convolution neural network (CNN) with residual blocks introduced in [41] on the Dino V2 features and a 3-layer MLP to further map the patch feature to dimension \\( D \\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.642, + 0.439, + 0.658 + ], + "angle": 0, + "content": "3.3 Contrastive Learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.667, + 0.659, + 0.819 + ], + "angle": 0, + "content": "We use contrastive learning to learn a joint embedding space for the scene graph nodes and image patches. To do so, we form query image and graph pairs \\((I, \\mathcal{G}_I)\\). Real-world scenes are rarely static, e.g., objects move or undergo non-rigid deformations and illumination changes [123]. To ensure that the learned embedding is robust to such temporal changes, we use scene graph \\(\\mathcal{G}_I\\) from the same temporal point when \\(I\\) was captured, as positive samples, as well as a scene graph \\(\\mathcal{G}_I^t\\) from another scan. Graph \\(\\mathcal{G}_I^t\\) represents the same place as \\(\\mathcal{G}_I\\), but it undergoes temporal changes. An example is shown in Fig. 4." + }, + { + "type": "image", + "bbox": [ + 0.669, + 0.647, + 0.787, + 0.771 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.668, + 0.774, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Fig.4: The same scene at different time steps \\( t \\)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.625, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SceneGraphLoc" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "For a query image \\(I\\), a set of candidate graphs \\(\\{\\mathcal{G}_I, \\mathcal{G}_I^t, \\mathcal{G}_1, \\dots, \\mathcal{G}_N\\}\\) is provided for training, where \\(\\{\\mathcal{G}_1, \\dots, \\mathcal{G}_N\\}\\) act as \\(N\\) negative samples, depicting different scenes than the target scene of the query image. We train our model by optimizing both a static loss and a temporal loss as:" + }, + { + "type": "equation", + "bbox": [ + 0.385, + 0.217, + 0.785, + 0.233 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\alpha * \\mathcal {L} _ {\\text {s t a t i c}} + (1 - \\alpha) * \\mathcal {L} _ {\\text {t e m p}}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.241, + 0.784, + 0.379 + ], + "angle": 0, + "content": "During training, we assume that image patch to graph node pairs are available [123] such that \\( P_{I} = \\{(q,v) \\mid q \\in \\mathcal{Q}_{I}, v \\in \\mathcal{G}_{I}\\} \\) and \\( P_{I}^{t} = \\{(q,v^{t}) \\mid q \\in \\mathcal{Q}_{I}, v \\in \\mathcal{G}_{I}^{t}\\} \\). For each pair \\( (q,v) \\) from \\( P_{I} \\) and each pair from \\( P_{I}^{t} \\), we use the following notation. Set \\( N_{q}^{\\mathcal{I}} = \\{q' \\mid q' \\in \\mathcal{Q}_{I}, v_{q'} \\neq v\\} \\) contains patches seeing objects other than \\( v \\). \\( N_{v}^{\\mathcal{G}} = \\{v_{n} \\mid v_{n} \\in \\mathcal{V}_{I} \\cup \\mathcal{V}_{1} \\cup \\ldots \\cup \\mathcal{V}_{N}\\} \\setminus \\{v\\} \\) contains the 3D objects of all candidate scene graphs other than \\( v \\), where \\( \\mathcal{V}_{I} \\) represents the objects nodes of the target graph \\( \\mathcal{G}_{I} \\) and \\( \\mathcal{V}_{i} \\) is the nodes of other graphs \\( \\mathcal{G}_{i} \\). \\( N_{v}^{\\mathcal{G}_{I}^{t}} = \\{v_{n}|v_{n} \\in \\mathcal{V}_{I}^{t} \\cup \\mathcal{V}_{1} \\cup \\ldots \\cup \\mathcal{V}_{n}\\} \\setminus \\{v\\} \\), where \\( \\mathcal{V}_{I}^{t} \\) represent the nodes of the target graph \\( \\mathcal{G}_{I}^{t} \\). The static loss is defined as bi-directional N-pair loss [67] as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.23, + 0.386, + 0.784, + 0.432 + ], + "angle": 0, + "content": "\\[\n\\left. \\right. \\mathcal {L} _ {\\text {s t a t i c}} = E _ {P I \\in B} \\left[ E _ {(q, v) \\in P I} \\left[ - \\log \\left(\\frac {1}{2} p (q, v, N _ {q} ^ {\\mathcal {I}}, N _ {v} ^ {\\mathcal {G}}) + \\frac {1}{2} p (v, q, N _ {q} ^ {\\mathcal {I}}, N _ {v} ^ {\\mathcal {G}})\\right)\\right]\\right], \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.433, + 0.784, + 0.478 + ], + "angle": 0, + "content": "where \\(E_{P_I\\in B}\\) represent loss averaging over a batch of query images and their corresponding candidate scene graphs; \\(p(q,v,N_q^\\mathcal{I},N_v^\\mathcal{G})\\) and \\(p(v,q,N_q^\\mathcal{I},N_v^\\mathcal{G})\\) represent the bi-directional probability distributions of the positive pair as:" + }, + { + "type": "equation", + "bbox": [ + 0.251, + 0.485, + 0.75, + 0.521 + ], + "angle": 0, + "content": "\\[\np (q, v, N _ {q} ^ {\\mathcal {I}}, N _ {v} ^ {\\mathcal {G}}) = \\frac {f (e _ {q} , e _ {v})}{f (e _ {q} , e _ {v}) + \\sum_ {q _ {n} \\in N _ {q} ^ {\\mathcal {I}}} f (e _ {q} , e _ {q _ {n}}) + \\sum_ {v _ {n} \\in N _ {v} ^ {\\mathcal {G}}} f (e _ {q} , e _ {v _ {n}})},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.53, + 0.784, + 0.597 + ], + "angle": 0, + "content": "where \\( f(e_q, e_v) = \\exp \\left( \\frac{-\\delta(e_q, e_v)}{\\tau} \\right) \\), \\( \\delta(e_q, e_v) \\) represents the inverse cosine similarity between embeddings \\( e_q \\) and \\( e_v \\), and \\( \\tau \\) is a temperature parameter. Probability distribution \\( p(v, q, N_q^T, N_v^G) \\) is written similarly. The temporal loss term is defined the same as Eq.6 but with \\( P_I^t \\) and \\( N_v^{G^t} \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.225, + 0.605, + 0.784, + 0.651 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t e m p}} = E _ {P _ {I} ^ {t} \\in B} \\left[ E _ {(q, v) \\in P _ {I} ^ {t}} \\left[ - \\log \\left(\\frac {1}{2} p (q, v, N _ {q} ^ {\\mathcal {I}}, N _ {v} ^ {\\mathcal {G} ^ {t}}) + \\frac {1}{2} p (v, q, N _ {q} ^ {\\mathcal {I}}, N _ {v} ^ {\\mathcal {G} ^ {t}})\\right) \\right] \\right]. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.652, + 0.784, + 0.682 + ], + "angle": 0, + "content": "By minimizing Eq.5, the paired cross-modal embeddings \\( e_{q} \\) and \\( e_{v} \\) are pulled together while the embeddings from different objects are pushed apart." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.701, + 0.449, + 0.716 + ], + "angle": 0, + "content": "3.4 Scene Graph Retrieval" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.724, + 0.784, + 0.8 + ], + "angle": 0, + "content": "Given a pre-established map of an environment represented through a collection of scene graphs \\(\\mathcal{G} = \\{\\mathcal{G}_i \\mid i \\in [0, N)\\}\\), where each node embedding has been precomputed, the goal during inference is to identify the top-\\(K\\) scene graphs \\(\\mathcal{G}_i\\) in which image \\(I\\) was likely captured. The method to address this challenge involves calculating the similarity between a graph and the image as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.364, + 0.807, + 0.784, + 0.843 + ], + "angle": 0, + "content": "\\[\ns \\left(\\mathcal {G} _ {i}, I\\right) = \\frac {1}{\\left| \\mathcal {Q} _ {I} \\right|} \\sum_ {q \\in \\mathcal {Q} _ {I}} \\left[ 1 - \\delta \\left(e _ {q}, \\operatorname {N N} (q)\\right) \\right], \\tag {8}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.127 + ], + "angle": 0, + "content": "Y. Miao et al." + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.145, + 0.773, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.245, + 0.788, + 0.302 + ], + "angle": 0, + "content": "Fig.5: Qualitative Result of object-association-based scene retrieval from the 3RScan dataset [123]. The two left images show the ground truth (left) and predicted (right) patch-to-node associations of the query image. The right part of the figure illustrates the candidate scene graphs sorted by the image-graph similarity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.313, + 0.789, + 0.419 + ], + "angle": 0, + "content": "where \\(\\mathcal{Q}_I\\) denotes the set of image patches in \\(I\\), and for each patch \\(q\\), \\(\\mathrm{NN}(q, \\mathcal{V}_i) \\in \\mathcal{V}_i\\) represents the nearest node in terms of embedding similarity. The function \\(s\\), which is assumed to map values to interval [0, 1], facilitating the identification of the optimal scene graph \\(\\mathcal{G}_{i^*}\\) that maximizes \\(s(\\mathcal{G}_i, I)\\). Graph \\(\\mathcal{G}_{i^*}\\) is identified by iterating through all scene graphs and selecting the one with the highest similarity. This approach can be accelerated by using spatial partitioning techniques, such as kd-trees, to preprocess node embeddings within the map." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.444, + 0.376, + 0.461 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.478, + 0.789, + 0.734 + ], + "angle": 0, + "content": "Baselines. No existing methods directly tackle our task, but several recent advancements provide relevant baselines. LidarCLIP [43], designed for autonomous driving, transforms LiDAR clouds into global descriptors using the Single-stride Sparse Transformer [24] and matches them with CLIP image encoder embeddings [92]. Though not a perfect fit, it can be adapted for matching point clouds with CLIP embeddings of query images. LIP-Loc [108] employs a similar approach by converting LiDAR clouds into 2D range images for direct encoding and matching. Both methods were fine-tuned on our dataset for accurate comparison. We also explored open-vocabulary [23,33,86,113] object-retrieval-based baselines using OpenMask3D [113] and OpenSeg [33], with OpenMask3D assigning CLIP descriptors to 3D object instances from multiple observing images, and OpenSeg extracting pixel-level CLIP features from the query image. Despite their original purposes, both can be adapted for localization by matching query image descriptors with object instances. For comparison with state-of-the-art visual localization methods requiring large image databases, we included CVNet [63] and AnyLoc [55]. These methods offer advanced performance but demand significant storage for image descriptors and exhibit slower inference times." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Map Generation. The mapping stage is executed offline as a preprocessing step for visual localization approaches, requiring specific mapping operations for each method before proceeding to localization. For our proposed method, this entails passing the point cloud, images, metadata, and relationships through the 3D scene graph encoder outlined in Section 3.1. For LIP-Loc, this step involves converting point clouds into range images and computing the embeddings of the range images and for LidarCLIP, point clouds are directly encoded into global" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.625, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SceneGraphLoc" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.244 + ], + "angle": 0, + "content": "Table 1: Retrieval recall on the test set of 3RScan dataset [123] (\\%; target scene ranked within the top 1, 3, and 5 of the retrieved list) and storage requirements (MB) for methods utilizing point clouds \\((\\mathcal{P})\\), images \\((\\mathcal{I})\\), and other modalities \\((\\mathrm{O} = \\{\\mathcal{A}, \\mathcal{S}, \\mathcal{R}\\})\\) for map representation. \\(R\\) and \\(R^t\\) represent the recall in the static and temporal scenarios respectively. Additionally, metrics for single-modal methods (CVNet and AnyLoc) reliant on extensive image datasets are presented. The results are reported for scenarios where the target room is chosen from a subset of 10 and 50 candidate scenes." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.255, + 0.787, + 0.352 + ], + "angle": 0, + "content": "
MethodMap modalities10 scenes50 scenesStorage (MB)
PIOR@1@3@5R@1@3@5R@1@3@5R@1@3@5
LidarCLIP [43]✓XX16.341.460.616.339.861.14.711.016.34.110.315.6
LIP-Loc [108]✓XX14.035.857.910.930.052.72.09.114.22.38.615.2
OpenMask3D [113]✓✓X---42.371.585.8---21.138.148.0
SceneGraphLoc (Ours)✓X✓53.681.992.850.576.888.430.250.261.228.246.256.4
SceneGraphLoc (Ours)✓✓✓---81.593.998.0---69.378.684.4
CVNet [63]X✓X---79.291.095.4---66.577.081.7
AnyLoc [55]X✓X---87.994.797.5---80.687.490.0
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.384, + 0.785, + 0.429 + ], + "angle": 0, + "content": "descriptors. In OpenMask3D, the CLIP embeddings corresponding to each object are calculated and stored. For image-based methods like CVNet and AnyLoc, embeddings for all images in the database are precomputed." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.787, + 0.55 + ], + "angle": 0, + "content": "Metrics. To evaluate the accuracy of a method, we focus on the recall of scene selection. This entails analyzing the scenario where, given a query image and corresponding scene pair \\((I, \\mathcal{G}_I)\\), alongside \\(N - 1\\) alternative scenes from the database, an ordering is established for these scenes according to their computed similarity to \\(I\\) as determined by the tested method. The metric Recall@\\(K\\) is employed to ascertain whether the target scene \\(\\mathcal{G}_i\\) is ranked among the top \\(K\\) scenes in terms of similarity as identified by the evaluated method. Additionally, we will report the inference time and storage requirements." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.552, + 0.788, + 0.704 + ], + "angle": 0, + "content": "Experiments on 3RScan. The 3RScan dataset [123] comprises 1335 annotated indoor scenes, representing 432 distinct rooms, with 1178 scenes (385 rooms) allocated for training and 157 (47 rooms) designated for validation. The training and validation sets include semantically annotated 3D point clouds for each scene, with some captured over extended periods (e.g., several months) showcasing environmental changes. Annotations for graphs within the 3RScan dataset are provided in [125]. Due to the absence of such annotations in the test set, it was excluded from our experiments. Thus, we reorganized the original validation set, allocating 34 scenes (17 rooms) for validation and 123 scenes (30 rooms) for testing. For full reproducibility, we will publish this split." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.84 + ], + "angle": 0, + "content": "During testing, we examine all 123 scenes of 30 rooms within the test set, selecting query images from each scene. The next step involves matching this image against \\( N \\) scenes (including the target) to ascertain whether the correct one could be identified by a method. This procedure is repeated for every image in each room. In total, all methods are tested on 30462 query images. Furthermore, we evaluate scene selection through two settings \\( N = 50 \\) and \\( N = 10 \\). The latter setting emulates a scenario where a pre-selection strategy is employed, for example, utilizing a global scene descriptor. In image-based methods, we use all images from the database and determine the scene based on the retrieved image." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "Y. Miao et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.228 + ], + "angle": 0, + "content": "Table 2: Retrieval recall in the temporal scenario on the test set of ScanNet dataset [21] (\\%; target scene ranked within the top 1, 3, and 5 of the retrieved list) and storage requirements (MB) for methods using point clouds \\((\\mathcal{P})\\), images \\((\\mathcal{I})\\), and other modalities \\((S, \\mathcal{R})\\) for map representation. The modality \\(\\mathcal{A}\\) is not available in the scene graphs predicted from [130]. The results are reported for scenarios where the target room is chosen from a subset of 10, 50 and the complete set of all (210) scenes." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.241, + 0.787, + 0.354 + ], + "angle": 0, + "content": "
MethodMap modalities10 scenes50 scenesAll scenesStorage (MB)
PIORt@1at3at5Rt@1at3at5Rt@1at3at5
LidarCLIP [43]✓XX19.447.567.64.714.822.25.915.021.90.7
LIP-Loc [108]✓XX10.327.043.61.96.08.11.83.14.01.7
OpenMask3D [113]✓✓X54.984.894.031.351.363.216.527.234.517.8
SceneGraphLoc (Ours)✓X✓54.181.491.929.047.458.013.526.434.29.3
SceneGraphLoc (Ours)✓✓✓78.592.798.361.683.291.653.469.878.79.3
CVNet [63]X✓X96.598.999.692.696.097.089.993.494.6239.1
AnyLoc [55]X✓X98.499.499.896.598.198.695.196.997.45720.3
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.385, + 0.784, + 0.49 + ], + "angle": 0, + "content": "Additionally, the methods are evaluated under both static and temporal conditions. In the static scenario, the target scene graph for a given query image originates from the same scan, albeit from a different sequence, to ensure no image overlap. Conversely, in the temporal scenario, the scene graph is derived from a sequence captured at a different temporal stage than the query, introducing potential environmental changes. We do not show results for methods exploiting images in their maps in the static stage." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.492, + 0.787, + 0.702 + ], + "angle": 0, + "content": "The results are in Table 1. Despite being retrained, LidarCLIP and LIP-Loc are inaccurate, particularly in scenarios involving the selection of the target room from the entire scene set. LIP-Loc barely surpasses random selection. Although LidarCLIP exhibits marginally better accuracy, it remains inferior to alternative methods. The inferiority is because both methods learn shared embedding between entire scenes and query images of totally different objects due to viewpoint differences. OpenMask3D, while achieving better results than LidarCLIP and LIP-Loc, is less accurate than the proposed SceneGraphLoc. SceneGraphLoc, even when excluding the image modality \\((\\mathcal{I})\\), outperforms other cross-modal strategies significantly. Incorporating images significantly enhances its performance, positioning it close to that of image-based approaches but with three orders of magnitude smaller storage requirements. Also, the storage of SceneGraphLoc with and without images is the same due to its design of distilling knowledge into fixed-sized embeddings. An example scene is shown in Fig. 5." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Experiments on ScanNet. To evaluate the generalization ability of our methods in real-world applications when scene graph annotations are not available, we conduct further experiments in the ScanNet dataset [21]. ScanNet encompasses 1613 monocular sequences of room-scale 3D scenes, offering 3D mesh reconstructions alongside the RGBD frame sequences utilized for the reconstructions. Given the absence of scene graph annotations in ScanNet, we run the SceneGraphFusion [130] on the RGBD sequences of scans for 3D reconstruction and scene graph prediction with 3D instance segmentation and object relationships (i.e., graph edges) within these graphs. As the process of scene graph prediction" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.625, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SceneGraphLoc" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 3: Average time (ms) of obtaining the query image embedding \\((t_{e_q})\\) and of the retrieval from 10, 50, and all scenes from the 3RScan [123] and ScanNet [21] datasets." + }, + { + "type": "table", + "bbox": [ + 0.278, + 0.185, + 0.727, + 0.302 + ], + "angle": 0, + "content": "
Method3DRScan [123]ScanNet [21]
teqt10retrt50retrteqt10retrt50retrtallretr
LidarCLIP [43]4.10.10.34.90.10.20.6
LIP-Loc [108]2.70.10.24.10.10.20.5
OpenMask3D [113]41.54.87.420.155.41.14.5
SceneGraphLoc (Ours)28.00.31.516.61.33.717.0
CVNet [63]14.39.060.054.010.674.1311.3
AnyLoc [55]658.4354.61826.4243.068.2329.01451.1
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.34, + 0.788, + 0.448 + ], + "angle": 0, + "content": "uses the RGBD frames of each scan, we avoid using those RGB images to match to the scene graph predicted of the scan. Thus, we only measure recall in the temporal scenario. Additionally, unlike 3RScan, the frame rate of RGBD sequences in ScanNet is high, and motion between consecutive frames is small. Thus, each database image is selected from every 25 consecutive frames in the sequence, for image-based methods [55, 63]. For a fair comparison, all the methods only use the same selected images for training and evaluation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.452, + 0.788, + 0.514 + ], + "angle": 0, + "content": "For training, we use the official training set. We divide the official validation set, which includes 312 scenes, into two distinct subsets: the first 100 scenes form our validation set, while the remaining 212 are allocated for testing. To ensure full reproducibility, we will make this split publicly available." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.518, + 0.788, + 0.716 + ], + "angle": 0, + "content": "The results are in Table 2 for scenarios selecting the target room from subsets of 10, 50, and the entire set of 210 scenes. The performance of LIP-Loc shows a similar pattern to that observed on 3RScan, performing only slightly better than random selection. LidarCLIP shows a small improvement in accuracy. OpenMask3D attains an accuracy comparable to our proposed method without incorporating the image modality. Our proposed SceneGraphLoc, when including the image modality in its map, significantly outperforms all cross-modal approaches. Although there remains a gap in accuracy compared with methods that use extensive image collections as maps (such as CVNet and AnyLoc), SceneGraphLoc benefits from a database size three orders-of-magnitude smaller, highlighting its efficiency and effectiveness. We partly attribute this performance gap to the lack of object attributes in the dataset and the inaccurate instance segmentation predicted by [130]. More details can be found in the supp. mat." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Processing time measured in milliseconds for various methods applied to the 3DRScan and ScanNet datasets, are detailed in Table 3. The computation time required to generate an embedding for the query image \\((t_{e_q})\\) is notably small across all methods, typically not exceeding a few tens of milliseconds, with the exception of AnyLoc, which runs for nearly a second. The retrieval phase for cross-modal approaches is generally limited to a few tens of milliseconds. However, methods such as CVNet and AnyLoc exhibit slower performance, due to searching through extensive image collections. When tasked with selecting from" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.127 + ], + "angle": 0, + "content": "Y. Miao et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.189 + ], + "angle": 0, + "content": "Table 4: Ablation study performed on the val. split of 3RScan [123], analysing map modalities (\\(\\mathcal{P}\\) - point cloud, \\(\\mathcal{I}\\) - image, \\(\\mathcal{A}\\) - attributes, \\(\\mathcal{S}\\) - structure, \\(\\mathcal{R}\\) - relationships) and the method (Dino v2 or GCVit) to obtain the image embeddings." + }, + { + "type": "table", + "bbox": [ + 0.25, + 0.2, + 0.76, + 0.315 + ], + "angle": 0, + "content": "
Map modalitiesDino v2 [81]GCVit [38]
PIASRR@1@3@5Rt@1@3@5R@1@3@5Rt@1@3@5
✓45.281.993.743.979.591.424.656.076.923.254.777.3
✓✓56.385.695.054.884.095.044.276.991.243.475.189.4
✓✓✓58.487.395.956.585.793.643.375.891.341.572.889.3
✓✓✓✓63.786.895.862.787.496.345.375.590.546.676.290.2
✓---80.296.099.0---69.487.493.7
✓✓---84.797.599.6---73.289.795.9
✓✓✓✓✓---88.597.799.6---72.188.895.7
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.343, + 0.784, + 0.373 + ], + "angle": 0, + "content": "a large number of images, the processing times of these methods can extend into the range of several hundred milliseconds or even reach upwards of a second." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.374, + 0.785, + 0.434 + ], + "angle": 0, + "content": "Ablation Studies. To understand the impact of the integration of different modalities, we provide an ablation study on the localization performance with multiple combinations of modalities. We use the validation split of the 3DRScan dataset. Additional ablation studies can be found in supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.435, + 0.788, + 0.602 + ], + "angle": 0, + "content": "Table 4 displays the coarse localization performance of our method under the incorporation of distinct modalities \\((\\mathcal{P},\\mathcal{I},\\mathcal{A},\\mathcal{S},\\mathcal{R})\\) within the pipeline. Additionally, the table reports results exploiting various backbones (Dino V2 [81] and GCVit [38]) for the extraction of image features. Similarly to the main experiments, we only show results on the temporal set when the map contains images. Employing Dino V2 for encoding both the query image and images within the map significantly enhances accuracy over using GCVit. The proposed method significantly outperforms both LidarCLIP and LIP-Loc (their results are in Table 1) even when using only point clouds in the map. Each additional modality contributes to the final accuracy, demonstrating that each plays a crucial role in enhancing the final performance when integrated into the pipeline." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.624, + 0.36, + 0.64 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.655, + 0.788, + 0.792 + ], + "angle": 0, + "content": "In conclusion, we introduce SceneGraphLoc, a novel method for solving the novel problem of localizing an input image within a 3D scene graph-based multi-modal reference map. This approach outperforms existing cross-modal methods by a large margin. It achieves comparable accuracy to state-of-the-art image-based techniques with significantly lower storage requirements and faster processing speeds. Our experiments across the 3RScan and ScanNet datasets demonstrate the effectiveness of SceneGraphLoc, with the best performance achieved when integrating all proposed modalities. We believe that SceneGraphLoc is a step towards lightweight and efficient localization." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.625, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SceneGraphLoc" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.409, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.283 + ], + "angle": 0, + "content": "We would like to thank our colleagues Ganlin Zhang, Sayan Deb Sarkar, and Cathrin Elich for their valuable advice and insightful discussions throughout the course of this research. Their contributions and suggestions greatly enhanced the quality and depth of this work. This work was partially funded by Design++ initiative of ETH Zurich, by the ETH RobotX research grant, the Hasler Stiftung Research Grant via the ETH Zurich Foundation, an ETH AI Center postdoctoral research fellowship and an ETH Zurich Career Seed Award." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.303, + 0.323, + 0.318 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.231, + 0.332, + 0.787, + 0.375 + ], + "angle": 0, + "content": "1. Agia, C., Jatavallabhula, K.M., Khodeir, M., Miksik, O., Vineet, V., Mukadam, M., Paull, L., Shkurti, F.: Taskgraphy: Evaluating robot task planning over large 3D scene graphs. In: Conference on Robot Learning (CoRL) (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.231, + 0.374, + 0.787, + 0.416 + ], + "angle": 0, + "content": "2. Arandjelovic, R., Gronat, P., Torii, A., Pajdla, T., Sivic, J.: NetVLAD: CNN architecture for weakly supervised place recognition. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2016) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.231, + 0.415, + 0.787, + 0.47 + ], + "angle": 0, + "content": "3. Armeni, I., He, Z.Y., Gwak, J., Zamir, A.R., Fischer, M., Malik, J., Savarese, S.: 3D Scene Graph: A Structure for Unified semantics, 3d space, and camera. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2019) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.231, + 0.47, + 0.787, + 0.498 + ], + "angle": 0, + "content": "4. Aubry, M., Russell, B.C., Sivic, J.: Painting-to-3d model alignment via discriminative visual elements. ACM Transactions On Graphics (TOG) (2014) 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.231, + 0.498, + 0.787, + 0.538 + ], + "angle": 0, + "content": "5. Aubry, M., Russell, B.C., Sivic, J.: Visual Geo-localization of Non-photographic Depictions via 2D-3D Alignment. In: Large-Scale Visual Geo-Localization (2016) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.231, + 0.538, + 0.787, + 0.58 + ], + "angle": 0, + "content": "6. Balntas, V., Li, S., Prisacariu, V.: RelocNet: Continuous Metric Learning Relocalisation using Neural Nets. In: European Conference on Computer Vision (ECCV) (2018) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.231, + 0.58, + 0.787, + 0.621 + ], + "angle": 0, + "content": "7. Bernreiter, L., Ott, L., Nieto, J., Siegwart, R., Cadena, C.: Spherical multi-modal place recognition for heterogeneous sensor systems. In: International Conference on Robotics and Automation (ICRA) (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.231, + 0.621, + 0.787, + 0.662 + ], + "angle": 0, + "content": "8. Berton, G., Masone, C., Caputo, B.: Rethinking visual geo-localization for large-scale applications. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.231, + 0.662, + 0.787, + 0.703 + ], + "angle": 0, + "content": "9. Berton, G., Paolicelli, V., Masone, C., Caputo, B.: Adaptive-attentive geolocation from few queries: A hybrid approach. In: IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) (2021) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.703, + 0.787, + 0.744 + ], + "angle": 0, + "content": "10. Bhayani, S., Sattler, T., Barath, D., Beliansky, P., Heikkilä, J., Kukelova, Z.: Calibrated and partially calibrated semi-generalized homographies. In: International Conference on Computer Vision (ICCV) (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.744, + 0.787, + 0.799 + ], + "angle": 0, + "content": "11. Brachmann, E., Krull, A., Nowozin, S., Shotton, J., Michel, F., Gumhold, S., Rother, C.: DSAC - Differentiable RANSAC for Camera Localization. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2017) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.787, + 0.842 + ], + "angle": 0, + "content": "12. Brachmann, E., Rother, C.: Learning Less is More - 6D Camera Localization via 3D Surface Regression. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2018) 3" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.332, + 0.787, + 0.842 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.127 + ], + "angle": 0, + "content": "Y. Miao et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "13. Brachmann, E., Rother, C.: Visual camera re-localization from RGB and RGB-D images using DSAC. Transactions on Pattern Analysis and Machine Intelligence (PAMI) (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.189, + 0.787, + 0.244 + ], + "angle": 0, + "content": "14. Brejcha, J., Lukac, M., Hold-Geoffroy, Y., Wang, O., Cadik, M.: LandscapeAR: Large Scale Outdoor Augmented Reality by Matching Photographs with Terrain Models Using Learned Descriptors. In: European Conference on Computer Vision (ECCV) (2020) 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.244, + 0.787, + 0.271 + ], + "angle": 0, + "content": "15. Cadik, M., Sykora, D., Lee, S.: Automated outdoor depth-map generation and alignment. Computers & Graphics (2018) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.271, + 0.787, + 0.311 + ], + "angle": 0, + "content": "16. Castle, R., Klein, G., Murray, D.W.: Video-rate localization in multiple maps for wearable augmented reality. In: IEEE International Symposium on Wearable Computers (2008) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.311, + 0.787, + 0.364 + ], + "angle": 0, + "content": "17. Cavallari, T., Bertinetto, L., Mukhoti, J., Torr, P., Golodetz, S.: Let's take this online: Adapting scene coordinate regression network predictions for online RGB-D camera relocalisation. In: International Conference on 3D Vision (3DV) (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.365, + 0.787, + 0.407 + ], + "angle": 0, + "content": "18. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention Mask Transformer for Universal Image Segmentation. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2022) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.407, + 0.787, + 0.447 + ], + "angle": 0, + "content": "19. Csurka, G., Dance, C., Fan, L., Willamowski, J., Bray, C.: Visual categorization with bags of keypoints. In: European Conference on Computer Vision (ECCV) Workshops (2004) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.447, + 0.787, + 0.488 + ], + "angle": 0, + "content": "20. Curless, B., Levoy, M.: A volumetric method for building complex models from range images. In: Annual conference on Computer graphics and interactive techniques (1996) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.488, + 0.787, + 0.528 + ], + "angle": 0, + "content": "21. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2017) 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.528, + 0.787, + 0.569 + ], + "angle": 0, + "content": "22. Doan, A.D., Latif, Y., Chin, T.J., Liu, Y., Do, T.T., Reid, I.: Scalable place recognition under appearance change for autonomous driving. In: International Conference on Computer Vision (ICCV) (2019) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.569, + 0.787, + 0.623 + ], + "angle": 0, + "content": "23. Engelmann, F., Manhardt, F., Niemeyer, M., Tateno, K., Tombari, F.: OpenNeRF: Open Set 3D Neural Scene Segmentation with Pixel-Wise Features and Rendered Novel Views. In: International Conference on Learning Representations (ICLR) (2024) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.623, + 0.787, + 0.676 + ], + "angle": 0, + "content": "24. Fan, L., Pang, Z., Zhang, T., Wang, Y.X., Zhao, H., Wang, F., Wang, N., Zhang, Z.: Embracing single stride 3d object detector with sparse transformer. International Conference on Computer Vision and Pattern Recognition (CVPR) (2022) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.677, + 0.787, + 0.718 + ], + "angle": 0, + "content": "25. Gadre, S.Y., Ehsani, K., Song, S., Mottaghi, R.: Continuous scene representations for embodied ai. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.718, + 0.787, + 0.773 + ], + "angle": 0, + "content": "26. Gao, P., Liang, J., Shen, Y., Son, S., Lin, M.C.: Visual, Spatial, Geometric-Preserved Place Recognition for Cross-View and Cross-Modal Collaborative Perception. In: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.773, + 0.787, + 0.8 + ], + "angle": 0, + "content": "27. Garg, S., Fischer, T., Milford, M.: Where is your place, visual place recognition? International Joint Conference on Artificial Intelligence (IJCAI) (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.787, + 0.84 + ], + "angle": 0, + "content": "28. Garg, S., Suenderhauf, N., Milford, M.: Semantic-geometric visual place recognition: a new perspective for reconciling opposing views. International Journal on Robotics Research (IJRR) (2019) 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.625, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "SceneGraphLoc" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.202 + ], + "angle": 0, + "content": "29. Garg, S., Sünderhauf, N., Dayoub, F., Morrison, D., Cosgun, A., Carneiro, G., Wu, Q., Chin, T.J., Reid, I., Gould, S., et al.: Semantics for robotic mapping, perception and interaction: A survey. Foundations and Trends in Robotics (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.203, + 0.785, + 0.243 + ], + "angle": 0, + "content": "30. Georgakis, G., Karanam, S., Wu, Z., Kosecka, J.: Learning local rgb-to-cad correspondences for object pose estimation. In: International Conference on Computer Vision (ICCV) (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.243, + 0.785, + 0.285 + ], + "angle": 0, + "content": "31. Germain, H., Bourmaud, G., Lepetit, V.: Sparse-to-Dense Hypercolumn Matching for Long-Term Visual Localization. In: International Conference on 3D Vision (3DV) (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.285, + 0.785, + 0.325 + ], + "angle": 0, + "content": "32. Germain, H., Bourmaud, G., Lepetit, V.: S2DNet: Learning Image Features for Accurate Sparse-to-Dense Matching. In: European Conference on Computer Vision (ECCV) (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.325, + 0.785, + 0.367 + ], + "angle": 0, + "content": "33. Ghiasi, G., Gu, X., Cui, Y., Lin, T.Y.: Scaling Open-Vocabulary Image Segmentation with Image-Level Labels. In: European Conference on Computer Vision (ECCV) (2022) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.367, + 0.785, + 0.407 + ], + "angle": 0, + "content": "34. Grabner, A., Roth, P.M., Lepetit, V.: 3D Pose Estimation and 3D Model Retrieval for Objects in the Wild. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.407, + 0.785, + 0.434 + ], + "angle": 0, + "content": "35. Grelsson, B., Robinson, A., Felsberg, M., Khan, F.S.: GPS-level accurate camera localization with HorizonNet. Journal of Field Robotics (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.434, + 0.785, + 0.461 + ], + "angle": 0, + "content": "36. Gumeli, C., Dai, A., Nießner, M.: ROCA: Robust CAD Model Retrieval and Alignment from a Single Image. arXiv preprint arXiv:2112.01988 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.461, + 0.785, + 0.487 + ], + "angle": 0, + "content": "37. Hanocka, R., Metzer, G., Giryes, R., Cohen-Or, D.: Point2mesh: A self-prior for deformable meshes. arXiv preprint arXiv:2005.11084 (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.487, + 0.785, + 0.528 + ], + "angle": 0, + "content": "38. Hatamizadeh, A., Yin, H., Heinrich, G., Kautz, J., Molchanov, P.: Global context vision transformers. In: International Conference on Machine Learning (ICML) (2023) 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.528, + 0.785, + 0.569 + ], + "angle": 0, + "content": "39. Hausler, S., Jacobson, A., Milford, M.: Multi-process fusion: Visual place recognition using multiple image processing methods. IEEE Robotics and Automation Letters (RA-L) (2019) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.569, + 0.785, + 0.61 + ], + "angle": 0, + "content": "40. Hausler, S., Garg, S., Xu, M., Milford, M., Fischer, T.: Patch-netvlad: Multiscale fusion of locally-global descriptors for place recognition. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2021) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.61, + 0.785, + 0.651 + ], + "angle": 0, + "content": "41. He, K., Zhang, X., Ren, S., Sun, J.: Deep Residual Learning for Image Recognition. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2016) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.651, + 0.785, + 0.704 + ], + "angle": 0, + "content": "42. Heng, L., Choi, B., Cui, Z., Geppert, M., Hu, S., Kuan, B., Liu, P., Nguyen, R., Yeo, Y.C., Geiger, A., et al.: Project Autovision: Localization and 3D Scene Perception for an Autonomous Vehicle with a Multi-Camera System. In: International Conference on Robotics and Automation (ICRA) (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.705, + 0.785, + 0.746 + ], + "angle": 0, + "content": "43. Hess, G., Tondersski, A., Petersson, C., Åström, K., Svensson, L.: Lidarclip or: How i learned to talk to point clouds. In: IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) (2024) 4, 10, 11, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.746, + 0.76, + 0.759 + ], + "angle": 0, + "content": "44. Hodan, T.: Pose Estimation of Specific Rigid Objects. Ph.D. thesis (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.759, + 0.785, + 0.8 + ], + "angle": 0, + "content": "45. Hodan, T., Barath, D., Matas, J.: EPOS: Estimating 6D Pose of Objects With Symmetries. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.841 + ], + "angle": 0, + "content": "46. Hodan, T., Zabulis, X., Lourakis, M.I.A., Obdrzalek, S., Matas, J.: Detection and fine 3D pose estimation of texture-less objects in RGB-D images. In: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) (2015) 4" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.127 + ], + "angle": 0, + "content": "Y. Miao et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "47. Hu, S., Feng, M., Nguyen, R.H.M., Lee, G.H.: CVM-Net: Cross-View Matching Network for Image-Based Ground-to-Aerial Geo-Localization. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.191, + 0.786, + 0.219 + ], + "angle": 0, + "content": "48. Hu, S., Lee, G.H.: Image-Based GeoLocalization Using Satellite Imagery. International Journal on Computer Vision (IJCV) (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.22, + 0.787, + 0.26 + ], + "angle": 0, + "content": "49. Hughes, N., Chang, Y., Carlone, L.: Hydra: A real-time spatial perception system for 3D scene graph construction and optimization. arXiv preprint arXiv:2201.13360 (2022) 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.261, + 0.787, + 0.289 + ], + "angle": 0, + "content": "50. Ibrahimi, S., van Noord, N., Alpherts, T., Worring, M.: Inside out visual place recognition. In: British Machine Vision Conference (2021) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.29, + 0.787, + 0.331 + ], + "angle": 0, + "content": "51. Irschara, A., Zach, C., Frahm, J.M., Bischof, H.: From Structure-from-Motion Point Clouds to Fast Location Recognition. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2009) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.333, + 0.787, + 0.388 + ], + "angle": 0, + "content": "52. Izadi, S., Kim, D., Hilliges, O., Molyneaux, D., Newcombe, R., Kohli, P., Shotton, J., Hodges, S., Freeman, D., Davison, A., et al.: Kinectfusion: real-time 3d reconstruction and interaction using a moving depth camera. In: ACM symposium on User interface software and technology (2011) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.389, + 0.787, + 0.43 + ], + "angle": 0, + "content": "53. Ji, X., Wei, J., Wang, Y., Shang, H., Kneip, L.: Cross-modal Place Recognition in Image Databases using Event-based Sensors. arXiv preprint arXiv:2307.01047 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.431, + 0.787, + 0.472 + ], + "angle": 0, + "content": "54. Kabalar, J., Wu, S.C., Wald, J., Tateno, K., Navab, N., Tombari, F.: Towards long-term retrieval-based visual localization in indoor environments with changes. IEEE Robotics and Automation Letters (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.474, + 0.787, + 0.515 + ], + "angle": 0, + "content": "55. Keetha, N., Mishra, A., Karhade, J., Jatavallabhula, K.M., Scherer, S., Krishna, M., Garg, S.: Anyloc: Towards universal visual place recognition. IEEE Robotics and Automation Letters (RA-L) (2023) 2, 3, 10, 11, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.516, + 0.787, + 0.558 + ], + "angle": 0, + "content": "56. Kendall, A., Cipolla, R.: Geometric loss functions for camera pose regression with deep learning. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2017) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.559, + 0.787, + 0.599 + ], + "angle": 0, + "content": "57. Kendall, A., Grimes, M., Cipolla, R.: PoseNet: A Convolutional Network for Real-Time 6-DOF Camera Relocalization. In: International Conference on Computer Vision (ICCV) (2015) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.601, + 0.787, + 0.642 + ], + "angle": 0, + "content": "58. Khaliq, A., Ehsan, S., Chen, Z., Milford, M., McDonald-Maier, K.: A holistic visual place recognition approach using lightweight CNNs for significant viewpoint and appearance changes. IEEE Transactions on Robotics (T-RO) (2020) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.643, + 0.787, + 0.685 + ], + "angle": 0, + "content": "59. Kim, H.J., Dunn, E., Frahm, J.M.: Learned contextual feature reweighting for image geolocation. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2017) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.686, + 0.787, + 0.727 + ], + "angle": 0, + "content": "60. Kim, U.H., Park, J.M., Song, T.J., Kim, J.H.: 3-D scene graph: A sparse and semantic representation of physical environments for intelligent agents. IEEE transactions on cybernetics (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.728, + 0.787, + 0.755 + ], + "angle": 0, + "content": "61. Kutulakos, K.N., Seitz, S.M.: A theory of shape by space carving. International Journal on Computer Vision (IJCV) (2000) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.757, + 0.787, + 0.798 + ], + "angle": 0, + "content": "62. Labbe, Y., Carpentier, J., Aubry, M., Sivic, J.: CosyPose: Consistent multi-view multi-object 6D pose estimation. In: European Conference on Computer Vision (ECCV) (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "63. Lee, S., Seong, H., Lee, S., Kim, E.: Correlation verification for image retrieval. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2022) 10, 11, 12, 13" + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.625, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "SceneGraphLoc" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.188 + ], + "angle": 0, + "content": "64. Li, Y., Snavely, N., Huttenlocher, D.P., Fua, P.: Worldwide Pose Estimation Using 3D Point Clouds. In: European Conference on Computer Vision (ECCV) (2012) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.189, + 0.785, + 0.231 + ], + "angle": 0, + "content": "65. Lim, H., Sinha, S.N., Cohen, M.F., Uysettendaele, M.: Real-time Image-based 6-DoF Localization in Large-scale Environments. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2012) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.232, + 0.785, + 0.272 + ], + "angle": 0, + "content": "66. Lin, T.Y., Cui, Y., Belongie, S.J., Hays, J.: Learning deep representations for ground-to-aerial geolocation. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2015) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.273, + 0.785, + 0.313 + ], + "angle": 0, + "content": "67. Lin, Z., Zhang, Z., Wang, M., Shi, Y., Wu, X., Zheng, Y.: Multi-modal contrastive representation learning for entity alignment. arXiv preprint arXiv:2209.00891 (2022) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.315, + 0.785, + 0.355 + ], + "angle": 0, + "content": "68. Liu, L., Li, H., Dai, Y.: Efficient Global 2D-3D Matching for Camera Localization in a Large-Scale 3D Map. In: International Conference on Computer Vision (ICCV) (2017) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.357, + 0.785, + 0.396 + ], + "angle": 0, + "content": "69. Liu, L., Li, H., Dai, Y.: Stochastic attraction-repulsion embedding for large scale image localization. In: International Conference on Computer Vision (ICCV) (2019) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.398, + 0.785, + 0.438 + ], + "angle": 0, + "content": "70. Looper, S., Rodriguez-Puigvert, J., Siegwart, R., Cadena, C., Schmid, L.: 3D VSG: Long-term semantic scene change prediction through 3D variable scene graphs. In: International Conference on Robotics and Automation (ICRA) (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.439, + 0.785, + 0.481 + ], + "angle": 0, + "content": "71. Lynen, S., Zeisl, B., Aiger, D., Bosse, M., Hesch, J., Pollefeys, M., Siegwart, R., Sattler, T.: Large-scale, real-time visual-inertial localization revisited. International Journal on Robotics Research (IJRR) (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.481, + 0.785, + 0.522 + ], + "angle": 0, + "content": "72. Lynen, S., Zeisl, B., Aiger, D., Bosse, M., Hesch, J., Pollefeys, M., Siegwart, R., Sattler, T.: Large-scale, real-time visual-inertial localization revisited. International Journal on Robotics Research (IJRR) (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.523, + 0.785, + 0.563 + ], + "angle": 0, + "content": "73. Mescheder, L., Oechsle, M., Niemeyer, M., Nowozin, S., Geiger, A.: Occupancy networks: Learning 3d reconstruction in function space. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.565, + 0.785, + 0.591 + ], + "angle": 0, + "content": "74. Miao, Y., Armeni, I., Pollefeys, M., Barath, D.: Volumetric Semantically Consistent 3D Panoptic Mapping. arXiv preprint arXiv:2309.14737 (2024) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.592, + 0.785, + 0.633 + ], + "angle": 0, + "content": "75. Miao, Y., Li, C., Li, Z., Yang, Y., Yu, X.: A novel algorithm of ship structure modeling and target identification based on point cloud for automation in bulk cargo terminals. Measurement and Control (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.634, + 0.785, + 0.674 + ], + "angle": 0, + "content": "76. Mihajlovic, M., Weder, S., Pollefeys, M., Oswald, M.R.: DeepSurfels: Learning online appearance fusion. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.675, + 0.785, + 0.715 + ], + "angle": 0, + "content": "77. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.716, + 0.785, + 0.757 + ], + "angle": 0, + "content": "78. Moreau, A., Piasco, N., Tsishkou, D., Stanciulescu, B., de La Fortelle, A.: LENS: Localization enhanced by neRF synthesis. In: Conference on Robot Learning (CoRL) (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "79. Murez, Z., Van As, T., Bartolozzi, J., Sinha, A., Badrinarayanan, V., Rabinovich, A.: Atlas: End-to-end 3d scene reconstruction from posed images. In: European Conference on Computer Vision (ECCV) (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "80. Oechsle, M., Mescheder, L., Niemeyer, M., Strauss, T., Geiger, A.: Texture fields: Learning texture representations in function space. In: International Conference on Computer Vision (ICCV) (2019) 4" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "20" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.127 + ], + "angle": 0, + "content": "Y. Miao et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.147, + 0.787, + 0.202 + ], + "angle": 0, + "content": "81. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023) 7, 8, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.203, + 0.787, + 0.23 + ], + "angle": 0, + "content": "82. Panek, V., Kukelova, Z., Sattler, T.: Meshloc: Mesh-based visual localization. In: European Conference on Computer Vision (ECCV) (2022) 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.23, + 0.787, + 0.284 + ], + "angle": 0, + "content": "83. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.284, + 0.787, + 0.325 + ], + "angle": 0, + "content": "84. Peng, G., Yue, Y., Zhang, J., Wu, Z., Tang, X., Wang, D.: Semantic reinforced attention learning for visual place recognition. In: International Conference on Robotics and Automation (ICRA) (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.325, + 0.787, + 0.366 + ], + "angle": 0, + "content": "85. Peng, G., Zhang, J., Li, H., Wang, D.: Attentional pyramid pooling of salient visual residuals for place recognition. In: International Conference on Computer Vision (ICCV) (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.366, + 0.787, + 0.407 + ], + "angle": 0, + "content": "86. Peng, S., Genova, K., Jiang, C.M., Tagliasacchi, A., Pollefeys, M., Funkhouser, T.: OpenScene: 3D Scene Understanding with Open Vocabularies. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2023) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.407, + 0.787, + 0.447 + ], + "angle": 0, + "content": "87. Peng, S., Niemeyer, M., Mescheder, L., Pollefeys, M., Geiger, A.: Convolutional occupancy networks. In: European Conference on Computer Vision (ECCV) (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.447, + 0.787, + 0.488 + ], + "angle": 0, + "content": "88. Pion, N., Humenberger, M., Csurka, G., Cabon, Y., Sattler, T.: Benchmarking image retrieval for visual localization. In: International Conference on 3D Vision (3DV) (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.488, + 0.787, + 0.528 + ], + "angle": 0, + "content": "89. Plotz, T., Roth, S.: Automatic Registration of Images to Untextured Geometry Using Average Shading Gradients. International Journal on Computer Vision (IJCV) (2017) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.528, + 0.787, + 0.57 + ], + "angle": 0, + "content": "90. Ponimatkin, G., Labbe, Y., Russell, B., Aubry, M., Sivic, J.: Focal length and object pose estimation via render and compare. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.569, + 0.787, + 0.609 + ], + "angle": 0, + "content": "91. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2017) 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.609, + 0.787, + 0.664 + ], + "angle": 0, + "content": "92. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. International Conference on Machine Learning (ICML) (2021) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.664, + 0.787, + 0.704 + ], + "angle": 0, + "content": "93. Ramalingam, S., Bouaziz, S., Sturm, P.F., Brand, M.: SKYLINE2GPS: Localization in urban canyons using omni-skylines. In: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) (2010) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.704, + 0.787, + 0.759 + ], + "angle": 0, + "content": "94. Ravichandran, Z., Peng, L., Hughes, N., Griffith, J., Carlone, L.: Hierarchical representations and explicit memory: Learning effective navigation policies on 3D scene graphs using graph neural networks. In: International Conference on Robotics and Automation (ICRA) (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.759, + 0.787, + 0.8 + ], + "angle": 0, + "content": "95. Rosinol, A., Violette, A., Abate, M., Hughes, N., Chang, Y., Shi, J., Gupta, A., Carlone, L.: Kimera: From SLAM to spatial perception with 3D dynamic scene graphs. International Journal on Robotics Research (IJRR) (2021) 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.8, + 0.787, + 0.841 + ], + "angle": 0, + "content": "96. Rosinol, A., Gupta, A., Abate, M., Shi, J., Carlone, L.: 3D dynamic scene graphs: Actionable spatial perception with places, objects, and humans. arXiv preprint arXiv:2002.06289 (2020) 4" + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.625, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SceneGraphLoc" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "21" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "97. Sarkar, S.D., Miksik, O., Pollefeys, M., Barath, D., Armeni, I.: SGAligner: 3D Scene Alignment with Scene Graphs. International Conference on Computer Vision (ICCV) (2023) 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.189, + 0.785, + 0.23 + ], + "angle": 0, + "content": "98. Sarlin, P.E., Cadena, C., Siegwart, R., Dymczyk, M.: From Coarse to Fine: Robust Hierarchical Localization at Large Scale. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.23, + 0.785, + 0.271 + ], + "angle": 0, + "content": "99. Sarlin, P.E., DeTone, D., Malisiewicz, T., Rabinovich, A.: Superglue: Learning feature matching with graph neural networks. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.271, + 0.785, + 0.325 + ], + "angle": 0, + "content": "100. Sarlin, P.E., DeTone, D., Yang, T.Y., Avetisyan, A., Straub, J., Malisiewicz, T., Bulò, S.R., Newcombe, R., Kontschieder, P., Balntas, V.: OrienterNet: Visual Localization in 2D Public Maps with Neural Matching. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2023) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.325, + 0.785, + 0.379 + ], + "angle": 0, + "content": "101. Sarlin, P.E., Unagar, A., Larsson, M., Germain, H., Toft, C., Larsson, V., Pollefeys, M., Lepetit, V., Hammarstrand, L., Kahl, F., et al.: Back to the Feature: Learning Robust Camera Localization from Pixels to Pose. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.379, + 0.785, + 0.42 + ], + "angle": 0, + "content": "102. Sattler, T., Leibe, B., Kobbelt, L.: Efficient & Effective Prioritized Matching for Large-Scale Image-Based Localization. Transactions on Pattern Analysis and Machine Intelligence (PAMI) (2017) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.42, + 0.785, + 0.461 + ], + "angle": 0, + "content": "103. Sattler, T., Zhou, Q., Pollefeys, M., Leal-Taixe, L.: Understanding the limitations of cnn-based absolute camera pose regression. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.461, + 0.785, + 0.515 + ], + "angle": 0, + "content": "104. Savinov, N., Hane, C., Ladicky, L., Pollefeys, M.: Semantic 3d reconstruction with continuous regularization and ray potentials using a visibility consistency constraint. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2016) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.515, + 0.785, + 0.556 + ], + "angle": 0, + "content": "105. Schonberger, J.L., Pollefeys, M., Geiger, A., Sattler, T.: Semantic visual localization. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2018) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.556, + 0.785, + 0.597 + ], + "angle": 0, + "content": "106. Sepulveda, G., Niebles, J., Soto, A.: A deep learning based behavioral approach to indoor autonomous navigation. In: International Conference on Robotics and Automation (ICRA) (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.597, + 0.785, + 0.637 + ], + "angle": 0, + "content": "107. Shan, Q., Wu, C., Curless, B., Furukawa, Y., Hernandez, C., Seitz, S.M.: Accurate geo-registration by ground-to-aerial image matching. In: International Conference on 3D Vision (3DV) (2014) 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.637, + 0.785, + 0.678 + ], + "angle": 0, + "content": "108. Shubodh, S., Omama, M., Zaidi, H., Parihar, U.S., Krishna, M.: Lip-loc: Lidar image pretraining for cross-modal localization. In: IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) (2024) 4, 10, 11, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.678, + 0.785, + 0.704 + ], + "angle": 0, + "content": "109. Sibling, D., Sattler, T., Leibe, B., Kobbelt, L.: SIFT-Realistic Rendering. In: International Conference on 3D Vision (3DV) (2013) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.704, + 0.785, + 0.758 + ], + "angle": 0, + "content": "110. Steiger Mueller, M., Sattler, T., Pollefeys, M., Jutz, B.: Image-to-image translation for enhanced feature matching, image retrieval and visual localization. ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.759, + 0.785, + 0.799 + ], + "angle": 0, + "content": "111. Stuckler, J., Behnke, S.: Multi-resolution surfel maps for efficient dense 3D modeling and tracking. Journal of Visual Communication and Image Representation (2014) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "112. Svarm, L., Enqvist, O., Kahl, F., Oskarsson, M.: City-Scale Localization for Cameras with Known Vertical Direction. Transactions on Pattern Analysis and Machine Intelligence (PAMI) (2017) 3" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "22" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.127 + ], + "angle": 0, + "content": "Y. Miao et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.203 + ], + "angle": 0, + "content": "113. Takmaz, A., Fedele, E., Sumner, R.W., Pollefeys, M., Tombari, F., Engelmann, F.: OpenMask3D: Open-Vocabulary 3D Instance Segmentation. In: International Conference on Neural Information Processing Systems (NeurIPS) (2023) 7, 10, 11, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.787, + 0.247 + ], + "angle": 0, + "content": "114. Tewari, A., Thies, J., Mildenhall, B., Srinivasan, P., Tretschk, E., Yifan, W., Lassner, C., Sitzmann, V., Martin-Brualla, R., Lombardi, S., et al.: Advances in neural rendering. In: Computer Graphics Forum (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.247, + 0.787, + 0.302 + ], + "angle": 0, + "content": "115. Tomesek, J., Cadik, M., Brejcha, J.: CrossLocate: Cross-modal Large-scale Visual Geo-Localization in Natural Environments using Rendered Modalities. In: IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) (2022) 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.303, + 0.787, + 0.346 + ], + "angle": 0, + "content": "116. Torii, A., Arandjelovic, R., Sivic, J., Okutomi, M., Pajdla, T.: 24/7 place recognition by view synthesis. Transactions on Pattern Analysis and Machine Intelligence (PAMI) (2018) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.346, + 0.787, + 0.389 + ], + "angle": 0, + "content": "117. Torii, A., Taira, H., Sivic, J., Pollefeys, M., Okutomi, M., Pajdla, T., Sattler, T.: Are large-scale 3d models really necessary for accurate visual localization? Transactions on Pattern Analysis and Machine Intelligence (PAMI) (2021) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.389, + 0.787, + 0.444 + ], + "angle": 0, + "content": "118. Valentin, J., Nießner, M., Shotton, J., Fitzgibbon, A., Izadi, S., Torr, P.: Exploiting Uncertainty in Regression Forests for Accurate Camera Relocalization. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2015) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.445, + 0.787, + 0.487 + ], + "angle": 0, + "content": "119. Velicković, P., Cucurull, G., Casanova, A., Romero, A., Lio, P., Bengio, Y.: Graph Attention Networks. In: International Conference on Learning Representations (ICLR) (2018) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.487, + 0.787, + 0.53 + ], + "angle": 0, + "content": "120. Ventura, J., Kukelova, Z., Sattler, T., Baráth, D.: Absolute Pose from One or Two Scaled and Oriented Features. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2024) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.53, + 0.787, + 0.572 + ], + "angle": 0, + "content": "121. Viswanathan, A., Rodrigues Pires, B., Huber, D.F.: Vision based robot localization by ground to satellite matching in GPS-denied situations. In: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) (2014) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.572, + 0.787, + 0.614 + ], + "angle": 0, + "content": "122. Walch, F., Hazirbas, C., Leal-Taixe, L., Sattler, T., Hilsenbeck, S., Cremers, D.: Image-Based Localization Using LSTMs for Structured Feature Correlation. In: International Conference on Computer Vision (ICCV) (2017) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.614, + 0.787, + 0.657 + ], + "angle": 0, + "content": "123. Wald, J., Avetisyan, A., Navab, N., Tombari, F., Niefsner, M.: Rio: 3d object instance re-localization in changing indoor environments. In: International Conference on Computer Vision (ICCV) (2019) 8, 9, 10, 11, 13, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.657, + 0.787, + 0.699 + ], + "angle": 0, + "content": "124. Wald, J., Dhamo, H., Navab, N., Tombari, F.: Learning 3d semantic scene graphs from 3d indoor reconstructions. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2020) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.7, + 0.787, + 0.742 + ], + "angle": 0, + "content": "125. Wald, J., Dhamo, H., Navab, N., Tombari, F.: Learning 3d semantic scene graphs from 3d indoor reconstructions. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2020) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.742, + 0.787, + 0.785 + ], + "angle": 0, + "content": "126. Wang, S., Kannala, J., Barath, D.: DGC-GNN: Descriptor-free Geometric-Color Graph Neural Network for 2D-3D Matching. International Conference on Computer Vision and Pattern Recognition (CVPR) (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.787, + 0.84 + ], + "angle": 0, + "content": "127. Warburg, F., Hauberg, S., Lopez-Antequera, M., Gargallo, P., Kuang, Y., Civera, J.: Mapillary street-level sequences: A dataset for lifelong place recognition. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2020) 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.625, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SceneGraphLoc" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "23" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "128. Weder, S., Schonberger, J.L., Pollefeys, M., Oswald, M.R.: Neuralfusion: Online depth fusion in latent space. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.231 + ], + "angle": 0, + "content": "129. Workman, S., Souvenir, R., Jacobs, N.: Wide-Area Image Geolocation with Aerial Reference Imagery. In: International Conference on Computer Vision (ICCV) (2015) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.232, + 0.785, + 0.285 + ], + "angle": 0, + "content": "130. Wu, S.C., Wald, J., Tateno, K., Navab, N., Tombari, F.: Scenegraphfusion: Incremental 3d scene graph prediction from rgb-d sequences. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2021) 4, 7, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.327 + ], + "angle": 0, + "content": "131. Ying, Z., Yuan, X., Yang, B., Song, Y., Xu, Q., Zhou, F., Sheng, W.: RP-SG: Relation Prediction in 3D Scene Graphs for Unobserved Objects Localization. IEEE Robotics and Automation Letters (RA-L) (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.37 + ], + "angle": 0, + "content": "132. Zaffar, M., Garg, S., Milford, M., et al.: Vpr-bench: An open-source visual place recognition evaluation framework with quantifiable viewpoint and appearance change. International Journal on Computer Vision (IJCV) (2021) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.397 + ], + "angle": 0, + "content": "133. Zeisl, B., Sattler, T., Pollefeys, M.: Camera pose voting for large-scale image-based localization. In: International Conference on Computer Vision (ICCV) (2015) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.398, + 0.785, + 0.438 + ], + "angle": 0, + "content": "134. Zhang, C., Yu, J., Song, Y., Cai, W.: Exploiting edge-oriented reasoning for 3D point-based scene graph analysis. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.439, + 0.785, + 0.481 + ], + "angle": 0, + "content": "135. Zhang, G., Larsson, V., Barath, D.: Revisiting Rotation Averaging: Uncertainties and Robust Losses. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.481, + 0.785, + 0.522 + ], + "angle": 0, + "content": "136. Zhang, S., Hao, A., Qin, H.: Knowledge-inspired 3D scene graph prediction in point cloud. International Conference on Neural Information Processing Systems (NeurIPS) (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.522, + 0.785, + 0.563 + ], + "angle": 0, + "content": "137. Zhang, W., Kosecka, J.: Image based localization in urban environments. In: International Symposium on 3D Data Processing, Visualization, and Transmission (2006) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.565, + 0.785, + 0.605 + ], + "angle": 0, + "content": "138. Zhang, Z., Sattler, T., Scaramuzza, D.: Reference Pose Generation for Long-term Visual Localization via Learned Features and View Synthesis. International Journal on Computer Vision (IJCV) (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.606, + 0.785, + 0.646 + ], + "angle": 0, + "content": "139. Zhao, L., Gatsis, K., Papachristodoulou, A.: Stable and Safe Reinforcement Learning via a Barrier-Lyapunov Actor-Critic Approach. In: IEEE Conference on Decision and Control (CDC) (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.647, + 0.785, + 0.688 + ], + "angle": 0, + "content": "140. Zhao, L., Miao, K., Gatsis, K., Papachristodoulou, A.: Stable and safe human-aligned reinforcement learning through neural ordinary differential equations. arXiv preprint arXiv:2401.13148 (2024) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.689, + 0.785, + 0.716 + ], + "angle": 0, + "content": "141. Zheng, E., Wu, C.: Structure From Motion Using Structure-Less Resection. In: International Conference on Computer Vision (ICCV) (2015) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.717, + 0.785, + 0.757 + ], + "angle": 0, + "content": "142. Zhou, Q., Agostinho, S., Osep, A., Leal-Taixe, L.: Is Geometry Enough for Matching in Visual Localization? In: European Conference on Computer Vision (ECCV) (2022) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "143. Zurbrügg, R., Liu, Y., Engelmann, F., Kumar, S., Hutter, M., Patil, V., Yu, F.: ICGNet: A Unified Approach for Instance-Centric Grasping. In: International Conference on Robotics and Automation (ICRA) (2024) 2" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.799 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/SceneGraphLoc_ Cross-Modal Coarse Visual Localization on 3D Scene Graphs/ca6ad507-bc20-40c4-b703-2093582a66a7_origin.pdf b/2024/SceneGraphLoc_ Cross-Modal Coarse Visual Localization on 3D Scene Graphs/ca6ad507-bc20-40c4-b703-2093582a66a7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..efff463bcc607364d68609abf6efa31f2000b170 --- /dev/null +++ b/2024/SceneGraphLoc_ Cross-Modal Coarse Visual Localization on 3D Scene Graphs/ca6ad507-bc20-40c4-b703-2093582a66a7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d2f6934abccaa7e77272405cd80e6cc459aa9bf4a048507cb06467883be9219f +size 3009812 diff --git a/2024/SceneGraphLoc_ Cross-Modal Coarse Visual Localization on 3D Scene Graphs/full.md b/2024/SceneGraphLoc_ Cross-Modal Coarse Visual Localization on 3D Scene Graphs/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d54c85242a381c25718f0389361aeec379dc62ac --- /dev/null +++ b/2024/SceneGraphLoc_ Cross-Modal Coarse Visual Localization on 3D Scene Graphs/full.md @@ -0,0 +1,381 @@ +# SceneGraphLoc: Cross-Modal Coarse Visual Localization on 3D Scene Graphs + +Yang Miao1, Francis Engelmann1,2, Olga Vysotska1, Federico Tombari2,3, Marc Pollefeys1,4, and Daniel Béla Baráth1 + +1ETH Zurich 2Google 3TU Munich 4Microsoft + +![](images/b97064f8bd510427d17b3f0f97fd8f0aded53b6b81ff72b84f95c056655847ec.jpg) +Fig. 1: SceneGraphLoc addresses the novel problem of localizing a query image in a database of 3D scenes represented as compact multi-modal 3D scene graphs. + +Abstract. We introduce the task of localizing an input image within a multi-modal reference map represented by a collection of 3D scene graphs. These scene graphs comprise multiple modalities, including object-level point clouds, images, attributes, and relationships between objects, offering a lightweight and efficient alternative to conventional methods that rely on extensive image databases. Given these modalities, the proposed method SceneGraphLoc learns a fixed-sized embedding for each node (i.e., representing object instances) in the scene graph, enabling effective matching with the objects visible in the input query image. This strategy significantly outperforms other cross-modal methods, even without incorporating images into the map representation. With images, SceneGraphLoc achieves performance close to that of state-of-the-art techniques depending on large image databases, while requiring three orders-of-magnitude less storage and operating orders-of-magnitude faster. Code and models are available at https://scenegraphloc.github.io. + +Keywords: Coarse Localization $\cdot$ 3D Scene Graph $\cdot$ Multi-modality + +# 1 Introduction + +Coarse visual localization, or place recognition, is a fundamental component in computer vision and robotics applications, defined as the task of identifying the approximate location where a query image was taken, given a certain tolerance + +![](images/6b731af108ffd1c2702a35f20734a8083aac8a368bca4f0292c01cee44ba433e.jpg) +Fig. 2: Overview. The training phase is represented by orange arrows, while blue arrows denote the inference phase. During training, a query image and its associated 3D scene graph form a positive sample within a contrastive learning framework, where negative samples are generated by associating scene graphs of different scenes with the same query image. The objective is to learn the embeddings of both the graph and the image so that embeddings of the positive pair are drawn closer, whereas those of the negative pair are pushed apart. In the inference phase, the task involves assigning the correct scene graph to a given query image from a selection of multiple graphs, achieved by optimizing the cosine similarity between their embeddings. + +level [2,9,22,28,39,40,50,58,69,116,117,127,132,143]. This capability is crucial for estimating the state of robots and is widely utilized in autonomous, unmanned aerial, terrestrial, and underwater vehicles, as well as AR/VR devices. The task is typically approached as an image retrieval problem, where the image to be localized is compared against a large database of posed images and, optionally, a 3D reconstruction of the scene. The most similar images retrieved from the database are used to estimate the precise location of the query image. + +The challenge with current state-of-the-art image-based coarse localization methods, such as AnyLoc [55], is their dependency on extensive image databases, which are not only storage-heavy but also slow to query, despite optimizations through hashing and other tricks. Moreover, these methods typically necessitate that the query and database share the same modality, limiting the scope of their application. Cross-modal approaches, such as [100,142], which attempt to bridge different types of data, often restrict their scope to connecting two modalities at a time (e.g., image-to-point cloud or image-to-bird's eye view map), one for the query and one for the database, thus narrowing their potential applications. + +This paper addresses the novel challenge of localizing a query image within a database that is represented not by conventional images but by the 3D scene graph [3,124] that integrates a diverse set of modalities, including point clouds, images, semantics, object attributes, and relationships. We tackle this problem by learning to map these modalities into a unified embedding space, thus allowing us to represent indoor scenes compactly through their objects (e.g., table and wall). This method enables the creation of small, efficient databases and significantly accelerates the coarse localization process. + +Contributions. The primary contributions of this paper are as follows: + +1. Introducing a novel problem: cross-modal localization of a query image within 3D scene graphs incorporating a mixture of modalities. +2. SceneGraphLoc, a new method for the coarse localization of an input image given a reference map represented by a database of 3D scene graphs. + +Even without incorporating images into the map, SceneGraphLoc largely outperforms other cross-modal methods on two large-scale, real-world indoor datasets. With images, SceneGraphLoc achieves performance close to that of state-of-the-art image-based methods while requiring three orders-of-magnitude less storage and operating orders-of-magnitude faster. The method is visualized in Fig. 1. + +# 2 Related Work + +Localization, the process of determining the position and orientation of an agent within a pre-built map, is pivotal across various domains such as mobile robots [42, 65, 139, 140], and augmented reality [16, 71]. The differentiation in localization techniques arises from their scene representation methods - be it through explicit 3D models [31, 32, 51, 64, 68, 72, 98, 99, 101, 102, 105, 112, 133, 135], sets of posed images [10, 88, 137, 141], or implicitly via neural network weights [6, 11-13, 17, 56, 57, 78, 118, 122] - and their approach to camera pose estimation, whether by 2D-3D [31, 32, 98, 99, 102, 112, 120, 126, 133] or 2D-2D [10, 141] matches, or through a composite of base poses [56, 57, 78, 88, 103, 122]. In practice, localization comprises two main steps: a coarse and precise stage. Here, we focus on the coarse step, finding potential locations of a query image. + +Coarse Localization (or place recognition) is often cast as an image retrieval problem [2,8,9,27,40,50,55,59,84,85,127] that consists of two phases. In the offline indexing phase, a reference map (image or point cloud database) is gathered. In the online retrieval phase, a query image - captured during a future traverse of the environment - is localized coarsely by retrieving the closest match to this image in the reference map. Recent methods perform the retrieval using learned embeddings that are produced by a feature extraction backbone equipped with a head that implements some form of aggregation or pooling, the most notable being NetVLAD [2]. While these methods achieve impressive results, they are limited to a single modality (e.g., images) and require large databases. + +Localization using multi-modal data. While dense mesh models are not as widely adopted as sparse Structure-from-Motion-based approaches, they have nonetheless been the focus of considerable research efforts [4,5,14,15,35,82,89,93, 107,109,110,115,138,142]. The body of prior work can be broadly segmented into two main strategies: The first entails the precise alignment of actual images with three-dimensional models (which may be coarse) through applying specialized techniques such as ones using contours [89] or skylines [93]. The second strategy emphasizes the identification and matching of local image features [4,5,14,35, 107,109,110,115,126,138,142], a method that has gained traction for its ability to match real-world images with non-photorealistic renderings of colored meshes, or + +even meshes without color [14,82,115]. CAD and other models are also commonly used by object pose estimation [4, 30, 34, 36, 44-46, 62, 90]. Image to LiDAR localization [7,26,43,108] is also relevant, especially in robotics applications. + +Another variant of multi-modal data localization involves cross-view matching. This technique determines the camera position by finding correspondences between a ground-level query image and a two-dimensional bird's eye view map, such as a satellite image or a semantic landscape map [47,48,66,100,107,121,129]. Other cross-modal techniques were also proposed to involve semantics [28, 29] and event cameras [53] in image-based localization. + +These approaches, while demonstrating promising localization results, are limited to interactions between two modalities - one for the query and one for the reference database. Our approach, in contrast, seeks the cross-modal coarse localization of a query image in a database composed of multiple modalities, e.g., 3D point clouds, images, semantics, object attributes, and relationships. + +Scene Representation, encapsulating various scene attributes has evolved significantly, yielding diverse surface representations from explicit forms (3D point clouds [75,91], meshes [37], surfels [111]) to implicit ones (occupancy [61], signed distance functions [20,52]). The advent of neural representations has introduced novel means of encoding geometry [73,83,87,128], appearance [76,80], and semantics [74,79]. A comprehensive review is provided by Tewari et al. [114]. The integration of directions/rays [77] and visibility encoding in surface reconstruction [104] has further enriched this domain. Armeni et al. [3] introduced the 3D scene graph structure as a multi-layer representation of a scene that captures geometry, semantics, objects, and camera poses in a unified manner. Subsequent efforts [96,124] have further advanced 3D scene graph learning and structure. + +The increasing interest in 3D scene graphs [3, 54, 60, 96, 124] underscores their potential as structured, rich descriptors for real-world scenes. Methods range from online incremental construction [49, 130] to offline generation from RGB-D imagery [3, 95, 124], and approaches for scene graph prediction [134, 136]. Their application spans embodied AI [94, 95, 106], task completion [1, 25], variability estimation [70], and SLAM [49, 95]. Recent studies like [131] introduce frameworks for localizing unseen objects by utilizing 3D scene graphs and graph neural networks for relation prediction, showcasing the utility of scene graphs in enhancing spatial understanding. Similarly, [97] offers new perspectives on 3D scene alignment, employing node matching within overlapping scene graphs to facilitate precise 3D map alignment. Kabalar et al. [54] assume that the query image has been coarsely localized and leverages a scene graph to identify dynamic objects and for precise localization with image features. Despite these significant advancements underscoring the value of scene graphs, their potential in multimodal localization remains largely untapped. In this paper, we use a scene graph representation of the map in which we aim to localize a query image. + +# 3 Visual Localization with 3D Scene Graph + +Problem Statement. Let us assume that we are provided with a pre-constructed map of the environment, denoted as $\mathcal{G}$ , which is represented as a set of $N \in \mathbb{N}^+$ 3D scene graphs $\mathcal{G}_i = (\mathcal{V}_i, \mathcal{E}_i)$ , such that $\mathcal{G} = \{\mathcal{G}_i\}_{i \in [0, N)}$ . Having separate graphs $\mathcal{G}_i$ is analogous to the hierarchy levels presented in [49, 95], each representing a group of object instances constituting a place like a room or building. Moreover, this task can also be straightforwardly redefined as subgraph selection. Vertices $v \in \mathcal{V}_i$ symbolize instances of objects (e.g., chairs, tables) and large instances of semantic categories (e.g., walls, ground) within the scene, $i \in [0, N)$ . Let $\mathcal{V} = \bigcup_{i \in [0, N)} \mathcal{V}_i$ aggregate all objects across the scenes. Edges $\mathcal{E}_i = \{(v_j^i, v_k^i) | v_j^i, v_k^i \in \mathcal{V}_i\}$ delineate the relationships between objects, such as "nearby", "standing on", and "attached to". + +For each vertex $v$ , we introduce $M \in \mathbb{N}^{+}$ map modalities $f_{j}: \mathcal{V} \to \mathcal{M}_{j}$ for $j \in [0, M)$ , where $f_{j}$ maps vertex $v$ to the $j^{th}$ modality $\mathcal{M}_{j}$ that may be $\mathcal{M}_{j} \in \{\text{position}, \text{orientation}, \text{pointcloud}, \text{semantic category}, \text{image}, \text{attribute}\}$ . While this paper focuses on these modalities, the set is easily extendable by incorporating additional ones, such as textual description or floor plan. + +Given $\mathcal{G}$ and an input query image $I$ , the objective is to identify the scene graph $\mathcal{G}_i$ corresponding to the space depicted in image $I$ . Note that while we focus on having an input image in this paper, the method can be modified to other modalities as well, e.g., depth image. Formally, we aim to resolve problem: + +$$ +\mathcal {G} _ {i ^ {*}} = \underset {i \in [ 0, N)} {\arg \max } \llbracket \text {c o n t a i n s} (\mathcal {G} _ {i}, \mathbf {p} _ {I}) \rrbracket , \tag {1} +$$ + +where $\mathbf{p}_I \in \mathbb{R}^3$ is the unknown 3D position of the image, and $[.]$ is the Iverson bracket which equals to 1 if the condition inside holds and 0 otherwise. It is important to note that our objective is coarse localization, opting for the selection of $\mathcal{G}_i$ without needing precise estimation of $\mathbf{p}_I$ . + +To this end, the optimization problem can be reformulated to incorporate the chirality constraint, asserting that if an object is visible in image $I$ , it must be positioned in front of the camera in 3D space, unobstructed by any entities (e.g., walls). Consequently, the problem becomes: + +$$ +\mathcal {G} _ {i ^ {*}} = \underset {i \in [ 0, N)} {\arg \max } \sum_ {o _ {I} \in \mathcal {O} _ {I}} \llbracket \text {v i s i b l e} (\mathcal {G} _ {i}, o _ {I}) \rrbracket \approx \underset {i \in [ 0, N)} {\arg \max } \sum_ {o _ {I} \in \mathcal {O} _ {I}} \log P (o _ {I} \mid \mathcal {G} _ {i}), \tag {2} +$$ + +where $\mathcal{O}_I$ is the object set visible in image $I$ , object $o_I \subseteq \{(x,y) \in I\}$ is a set of pixels in the image $(o_I \in \mathcal{O}_I)$ , visible: $\mathcal{G}_i \times \mathcal{O}_I \to \{0,1\}$ indicates if an object is visible in a scene graph, $P(o_I \mid \mathcal{G}_i)$ is the probability of $o_I$ stemming from $\mathcal{G}_i$ . + +Function visible can be approached as an indicator of whether object $o_{I}$ appears in graph $G_{i}$ . It holds if and only if there exists a $v \in \mathcal{V}_{i}$ such that $v$ represents the same object as $o_{I}$ . Therefore, this problem can be approached as matching a set of image pixels $(o_{I})$ to a scene graph node $(v)$ that comprises a set of modalities. In the next sections, we will describe a way to learn a unified embedding space for both $o_{I}$ and $v$ such that they become matchable. + +Proposed Pipeline is shown in Fig. 2. It consists of two concurrent stages: the first one generates object embeddings $e_{q} \in \mathbb{R}^{D}$ from patches $q \in \mathcal{Q}_I$ within the query image $I$ , and the second derives node embeddings $e_{v} \in \mathbb{R}^{D}$ for nodes $v \in \mathcal{V}_i$ in the scene graph $\mathcal{G}_i$ . The training objective is to make $\delta(e_{q}, e_{v}) = 0$ if and only if the object associated with node $v$ is directly visible (i.e., neither occluded nor outside the camera frustum) through the image patch $q$ . Here, $\delta : \mathbb{R}^D \times \mathbb{R}^D \to \mathbb{R}$ denotes inverse cosine similarity normalized to [0,1]. + +After generating $e_{q}$ and $e_{v}$ , the model performs nearest neighbor matching (NN) for each patch in each scene graph $\mathcal{G}_{i}$ , assigning node $v$ to $q$ such that + +$$ +\mathrm {N N} (q, \mathcal {V} _ {i}) = \underset {v \in \mathcal {V} _ {i}} {\arg \min } \delta \left(e _ {q}, e _ {v}\right). \tag {3} +$$ + +Through matching, we establish patch-to-node correspondences $\mathcal{C}_i = \{(q,v) \mid v = \mathrm{NN}(q,\mathcal{V}_i) \in \mathcal{V}_i, q \in \mathcal{Q}_I\}$ . Based on $\mathcal{C}_i$ , we devise an image-to-graph similarity score enabling us to deduce whether image $I$ corresponds to the space represented by $\mathcal{G}_i$ . Finally, potential coarse locations of $I$ are selected by maximizing the similarity score across the stored scene graphs as depicted in Eq 2. + +# 3.1 Object Embeddings in the Scene Graph + +This section aims to obtain an embedding for each node within graph $\mathcal{G}_i$ , encapsulating information from all available modalities. Our method builds upon the method of Sarkar et al. [97], with enhancements to include the image modality and to distill a unified embedding from all modalities rather than merely concatenating separate embeddings as in [97]. + +Scene graphs are conceptualized as multimodal knowledge graphs, similar to those used in entity alignment, treating semantic and geometric + +information as distinct modalities. The objective is to learn a joint multi-modal embedding from the individual modal encodings (uni-modal), ensuring nodes corresponding to the same object instance across different graphs are closely positioned. This involves the creation of uni-modal embeddings for the three primary types of 3D scene graph information: object embeddings encoding nodes in $\mathcal{V}$ , structure embeddings $\mathcal{S}$ representing edges in $\mathcal{E}$ as a structured graph, and two meta modalities encoding attributes $(\mathcal{A})$ and relationships $(\mathcal{R})$ between objects as one-hot vectors. These uni-modal embeddings are then combined in a weighted manner and optimized jointly through knowledge distillation. + +Each of these modalities is processed separately to generate uni-modal embeddings, which are subsequently integrated to model complex inter-modal interactions within the joint embedding space. + +Object Embedding. Node $v \in \mathcal{V}$ may contain multiple modalities, such as point cloud $(\mathcal{P})$ and image $(\mathcal{I})$ . Point clouds contain rich geometric information about objects. The point cloud corresponding to each $v \in \mathcal{V}$ is inputted to the object encoder. We employ the PointNet architecture [91] as the + +![](images/7467a870839053004005c7f8889d65635d49e7aa88ab50b8c2311bb06ce46386.jpg) +Fig.3: The embedding of image modality $\mathcal{I}$ for each object. The image crops of a pillow are shown as an example. + +object encoder to extract the geometric feature $e_v^{\mathcal{P}}$ for every node. Furthermore, we integrate multi-level and multi-view visual embeddings to enrich the graph encoder with a more nuanced understanding of image information. The visual embedding pipeline is visualized in Fig. 3. For each node $v$ denoting a 3D object, the top $K_{view} \in \mathbb{N}^{+}$ images with largest visibility of $v$ is selected: $\{I_{\mathrm{db},k}^v \mid k = [0,K)\} \subseteq \mathcal{I}_{\mathrm{db}}$ from the database $\mathcal{I}_{\mathrm{db}}$ . We can define an ordering over these images such that $\phi(I_{\mathrm{db},0}^v,v) \geq \phi(I_{\mathrm{db},1}^v,v) \geq \ldots \geq \phi(I_{\mathrm{db},K_{view}-1}^v,v)$ , with the visibility function $\phi$ quantifying the extent of node $v$ observed in each image through pixel count. Visibility check is implemented by projecting the 3D model of $v$ to each posed image in $\mathcal{I}_{\mathrm{db}}$ , which are usually available when constructing the scene graph [49, 130]. Parameter $K_{view} = 10$ in all our experiments. + +Drawing inspiration from OpenMask3D [113], for any given view $I_{\mathrm{db}}^v$ of object $v$ , initial steps include calculating the bounding box $b_{v,0}$ of $v$ within the image, followed by the generation of multi-level bounding boxes $\{b_{v,l} \mid l \in [0,L)\}$ through iterative enlargement of $b_{v,0}$ . This enlargement strategy aims to capture contextual information around object $v$ . Subsequently, Dino V2 [81] processes the image crops defined by $b_{v,l}$ , extracting multi-level features $\{f_l \mid l \in [0,L)\}$ . For each image $I_{\mathrm{db},k}^v$ , an average pooling operation aggregates these multi-level crop features into a singular feature vector $f = \operatorname{avg\_pool}\{f_l \mid l \in [0,L)\}$ . The final step of this process involves the application of a Transformer encoder, which incorporates image poses as positional encodings. This step synthesizes multiview object tokens into a cohesive visual embedding $e_v^{\mathcal{I}}$ , effectively integrating the diverse perspectives and levels of contextual information pertaining to each object within the database. Please note that the image database does not necessarily have to be stored after distilling the object embeddings. + +Structure Embedding. 3D Scene Graphs encapsulate the object relationships, which we exploit to encode their spatial configuration. This relational data is represented through a structure graph, where node features embody the relative translations between object instances, and edges denote these relationships. The relative translation is determined by calculating the distance from the object instance to any other object in the scene. To encapsulate this structural information within $\mathcal{G}_i$ , a Graph Attention Network (GAT) [119] is utilized, with the weight matrix constrained to a diagonal form to reduce computational demands and enhance model scalability. Following the method in [97], the structural embedding $e_v^S$ is derived from the final layer of a two-layer GAT model. + +Meta Embeddings. In addition to geometry and structure, the object attributes and the inter-object relationships are captured in two distinct embeddings, $e_v^{\mathcal{R}}$ and $e_v^{\mathcal{A}}$ . The relationships an object maintains with others are conceptualized as a bag-of-words [19] feature vector, which is input to a feed-forward neural network, distilled in relational embedding $e_v^{\mathcal{R}}$ . A similar approach is employed for the attributes associated with the objects, producing embedding $e_v^{\mathcal{A}}$ . + +Joint Embedding. Similar to [97], we concatenate uni-modal features to a compact representation for each object $v$ . Contrary to [97], we encapsulate it in a multi-layer perceptron (MLP) to learn an embedding with size independent of + +the available modalities, fusing information from all. Embedding $e_v$ is as: + +$$ +e _ {v} = \operatorname {M L P} \left(\oplus_ {k \in \mathcal {K}} \left[ \frac {\exp (w _ {k})}{\sum_ {j \in \mathcal {K}} \exp (w _ {j})} e _ {v} ^ {k} \right]\right), \tag {4} +$$ + +where $\oplus$ is the concatenation operator, $\mathcal{K} = \{\mathcal{P},\mathcal{I},\mathcal{S},\mathcal{R},\mathcal{A}\}$ , and $w_{m}$ is a trainable attention weight for each modality $k\in \mathcal{K}$ . A two-layer MLP is applied to map the dimension of the concatenated multi-modal descriptors from $D^{\mathcal{K}}$ to $D$ . We apply $L_{2}$ normalization to each uni-modal feature before concatenation. + +In practice, these independent modalities are only required and used in the mapping phase of the procedure. This stage involves the construction of an environmental map in the form of a 3D scene graph, during which each node $v$ is distilled into an embedding $e_v$ . During localization, we can ignore independent modalities and use only the fixed-sized embeddings $e_v$ . + +# 3.2 Object Embeddings in the Query Image + +To solve the optimization problem in Eq. 2, we need to find object instances $o_{I} \in \mathcal{O}_{I}$ in query image $I$ . A straightforward approach to do so would be to apply a 2D panoptic segmentation algorithm, e.g., Mask2Former [18]. However, we noticed in our experiments that such an approach is susceptible to inaccuracies and failures (over- and under-segmentation) in the segmentation, severely affecting the accuracy. Thus, we approach the problem by a visual Transformer (ViT), breaking up the image into rectangular patches $q \in \mathcal{Q}_I$ and distilling an independent embedding for each patch $q$ based on the object visible in $q$ . We use Dino V2 [81] as a backbone to obtain patch-level features. These features are passed through a patch encoder trained to create embeddings from the patch features considering the objects visible from each $q$ . For this encoder, we use a 4-layer convolution neural network (CNN) with residual blocks introduced in [41] on the Dino V2 features and a 3-layer MLP to further map the patch feature to dimension $D$ . + +# 3.3 Contrastive Learning + +We use contrastive learning to learn a joint embedding space for the scene graph nodes and image patches. To do so, we form query image and graph pairs $(I, \mathcal{G}_I)$ . Real-world scenes are rarely static, e.g., objects move or undergo non-rigid deformations and illumination changes [123]. To ensure that the learned embedding is robust to such temporal changes, we use scene graph $\mathcal{G}_I$ from the same temporal point when $I$ was captured, as positive samples, as well as a scene graph $\mathcal{G}_I^t$ from another scan. Graph $\mathcal{G}_I^t$ represents the same place as $\mathcal{G}_I$ , but it undergoes temporal changes. An example is shown in Fig. 4. + +![](images/8d1a947c739d12d88cdf406f40339bac1ce79f7be3236c88c35980a1c617d118.jpg) +Fig.4: The same scene at different time steps $t$ . + +For a query image $I$ , a set of candidate graphs $\{\mathcal{G}_I, \mathcal{G}_I^t, \mathcal{G}_1, \dots, \mathcal{G}_N\}$ is provided for training, where $\{\mathcal{G}_1, \dots, \mathcal{G}_N\}$ act as $N$ negative samples, depicting different scenes than the target scene of the query image. We train our model by optimizing both a static loss and a temporal loss as: + +$$ +\mathcal {L} = \alpha * \mathcal {L} _ {\text {s t a t i c}} + (1 - \alpha) * \mathcal {L} _ {\text {t e m p}}. \tag {5} +$$ + +During training, we assume that image patch to graph node pairs are available [123] such that $P_{I} = \{(q,v) \mid q \in \mathcal{Q}_{I}, v \in \mathcal{G}_{I}\}$ and $P_{I}^{t} = \{(q,v^{t}) \mid q \in \mathcal{Q}_{I}, v \in \mathcal{G}_{I}^{t}\}$ . For each pair $(q,v)$ from $P_{I}$ and each pair from $P_{I}^{t}$ , we use the following notation. Set $N_{q}^{\mathcal{I}} = \{q' \mid q' \in \mathcal{Q}_{I}, v_{q'} \neq v\}$ contains patches seeing objects other than $v$ . $N_{v}^{\mathcal{G}} = \{v_{n} \mid v_{n} \in \mathcal{V}_{I} \cup \mathcal{V}_{1} \cup \ldots \cup \mathcal{V}_{N}\} \setminus \{v\}$ contains the 3D objects of all candidate scene graphs other than $v$ , where $\mathcal{V}_{I}$ represents the objects nodes of the target graph $\mathcal{G}_{I}$ and $\mathcal{V}_{i}$ is the nodes of other graphs $\mathcal{G}_{i}$ . $N_{v}^{\mathcal{G}_{I}^{t}} = \{v_{n}|v_{n} \in \mathcal{V}_{I}^{t} \cup \mathcal{V}_{1} \cup \ldots \cup \mathcal{V}_{n}\} \setminus \{v\}$ , where $\mathcal{V}_{I}^{t}$ represent the nodes of the target graph $\mathcal{G}_{I}^{t}$ . The static loss is defined as bi-directional N-pair loss [67] as follows: + +$$ +\left. \right. \mathcal {L} _ {\text {s t a t i c}} = E _ {P I \in B} \left[ E _ {(q, v) \in P I} \left[ - \log \left(\frac {1}{2} p (q, v, N _ {q} ^ {\mathcal {I}}, N _ {v} ^ {\mathcal {G}}) + \frac {1}{2} p (v, q, N _ {q} ^ {\mathcal {I}}, N _ {v} ^ {\mathcal {G}})\right)\right]\right], \tag {6} +$$ + +where $E_{P_I\in B}$ represent loss averaging over a batch of query images and their corresponding candidate scene graphs; $p(q,v,N_q^\mathcal{I},N_v^\mathcal{G})$ and $p(v,q,N_q^\mathcal{I},N_v^\mathcal{G})$ represent the bi-directional probability distributions of the positive pair as: + +$$ +p (q, v, N _ {q} ^ {\mathcal {I}}, N _ {v} ^ {\mathcal {G}}) = \frac {f (e _ {q} , e _ {v})}{f (e _ {q} , e _ {v}) + \sum_ {q _ {n} \in N _ {q} ^ {\mathcal {I}}} f (e _ {q} , e _ {q _ {n}}) + \sum_ {v _ {n} \in N _ {v} ^ {\mathcal {G}}} f (e _ {q} , e _ {v _ {n}})}, +$$ + +where $f(e_q, e_v) = \exp \left( \frac{-\delta(e_q, e_v)}{\tau} \right)$ , $\delta(e_q, e_v)$ represents the inverse cosine similarity between embeddings $e_q$ and $e_v$ , and $\tau$ is a temperature parameter. Probability distribution $p(v, q, N_q^T, N_v^G)$ is written similarly. The temporal loss term is defined the same as Eq.6 but with $P_I^t$ and $N_v^{G^t}$ as follows: + +$$ +\mathcal {L} _ {\text {t e m p}} = E _ {P _ {I} ^ {t} \in B} \left[ E _ {(q, v) \in P _ {I} ^ {t}} \left[ - \log \left(\frac {1}{2} p (q, v, N _ {q} ^ {\mathcal {I}}, N _ {v} ^ {\mathcal {G} ^ {t}}) + \frac {1}{2} p (v, q, N _ {q} ^ {\mathcal {I}}, N _ {v} ^ {\mathcal {G} ^ {t}})\right) \right] \right]. \tag {7} +$$ + +By minimizing Eq.5, the paired cross-modal embeddings $e_{q}$ and $e_{v}$ are pulled together while the embeddings from different objects are pushed apart. + +# 3.4 Scene Graph Retrieval + +Given a pre-established map of an environment represented through a collection of scene graphs $\mathcal{G} = \{\mathcal{G}_i \mid i \in [0, N)\}$ , where each node embedding has been precomputed, the goal during inference is to identify the top- $K$ scene graphs $\mathcal{G}_i$ in which image $I$ was likely captured. The method to address this challenge involves calculating the similarity between a graph and the image as follows: + +$$ +s \left(\mathcal {G} _ {i}, I\right) = \frac {1}{\left| \mathcal {Q} _ {I} \right|} \sum_ {q \in \mathcal {Q} _ {I}} \left[ 1 - \delta \left(e _ {q}, \operatorname {N N} (q)\right) \right], \tag {8} +$$ + +![](images/fdb92816cf9745f003cb305955fbc936562c082c727a27b71c1680fa28e2e8ce.jpg) +Fig.5: Qualitative Result of object-association-based scene retrieval from the 3RScan dataset [123]. The two left images show the ground truth (left) and predicted (right) patch-to-node associations of the query image. The right part of the figure illustrates the candidate scene graphs sorted by the image-graph similarity. + +where $\mathcal{Q}_I$ denotes the set of image patches in $I$ , and for each patch $q$ , $\mathrm{NN}(q, \mathcal{V}_i) \in \mathcal{V}_i$ represents the nearest node in terms of embedding similarity. The function $s$ , which is assumed to map values to interval [0, 1], facilitating the identification of the optimal scene graph $\mathcal{G}_{i^*}$ that maximizes $s(\mathcal{G}_i, I)$ . Graph $\mathcal{G}_{i^*}$ is identified by iterating through all scene graphs and selecting the one with the highest similarity. This approach can be accelerated by using spatial partitioning techniques, such as kd-trees, to preprocess node embeddings within the map. + +# 4 Experiments + +Baselines. No existing methods directly tackle our task, but several recent advancements provide relevant baselines. LidarCLIP [43], designed for autonomous driving, transforms LiDAR clouds into global descriptors using the Single-stride Sparse Transformer [24] and matches them with CLIP image encoder embeddings [92]. Though not a perfect fit, it can be adapted for matching point clouds with CLIP embeddings of query images. LIP-Loc [108] employs a similar approach by converting LiDAR clouds into 2D range images for direct encoding and matching. Both methods were fine-tuned on our dataset for accurate comparison. We also explored open-vocabulary [23,33,86,113] object-retrieval-based baselines using OpenMask3D [113] and OpenSeg [33], with OpenMask3D assigning CLIP descriptors to 3D object instances from multiple observing images, and OpenSeg extracting pixel-level CLIP features from the query image. Despite their original purposes, both can be adapted for localization by matching query image descriptors with object instances. For comparison with state-of-the-art visual localization methods requiring large image databases, we included CVNet [63] and AnyLoc [55]. These methods offer advanced performance but demand significant storage for image descriptors and exhibit slower inference times. + +Map Generation. The mapping stage is executed offline as a preprocessing step for visual localization approaches, requiring specific mapping operations for each method before proceeding to localization. For our proposed method, this entails passing the point cloud, images, metadata, and relationships through the 3D scene graph encoder outlined in Section 3.1. For LIP-Loc, this step involves converting point clouds into range images and computing the embeddings of the range images and for LidarCLIP, point clouds are directly encoded into global + +Table 1: Retrieval recall on the test set of 3RScan dataset [123] (\%; target scene ranked within the top 1, 3, and 5 of the retrieved list) and storage requirements (MB) for methods utilizing point clouds $(\mathcal{P})$ , images $(\mathcal{I})$ , and other modalities $(\mathrm{O} = \{\mathcal{A}, \mathcal{S}, \mathcal{R}\})$ for map representation. $R$ and $R^t$ represent the recall in the static and temporal scenarios respectively. Additionally, metrics for single-modal methods (CVNet and AnyLoc) reliant on extensive image datasets are presented. The results are reported for scenarios where the target room is chosen from a subset of 10 and 50 candidate scenes. + +
MethodMap modalities10 scenes50 scenesStorage (MB)
PIOR@1@3@5R@1@3@5R@1@3@5R@1@3@5
LidarCLIP [43]✓XX16.341.460.616.339.861.14.711.016.34.110.315.6
LIP-Loc [108]✓XX14.035.857.910.930.052.72.09.114.22.38.615.2
OpenMask3D [113]✓✓X---42.371.585.8---21.138.148.0
SceneGraphLoc (Ours)✓X✓53.681.992.850.576.888.430.250.261.228.246.256.4
SceneGraphLoc (Ours)✓✓✓---81.593.998.0---69.378.684.4
CVNet [63]X✓X---79.291.095.4---66.577.081.7
AnyLoc [55]X✓X---87.994.797.5---80.687.490.0
+ +descriptors. In OpenMask3D, the CLIP embeddings corresponding to each object are calculated and stored. For image-based methods like CVNet and AnyLoc, embeddings for all images in the database are precomputed. + +Metrics. To evaluate the accuracy of a method, we focus on the recall of scene selection. This entails analyzing the scenario where, given a query image and corresponding scene pair $(I, \mathcal{G}_I)$ , alongside $N - 1$ alternative scenes from the database, an ordering is established for these scenes according to their computed similarity to $I$ as determined by the tested method. The metric Recall@ $K$ is employed to ascertain whether the target scene $\mathcal{G}_i$ is ranked among the top $K$ scenes in terms of similarity as identified by the evaluated method. Additionally, we will report the inference time and storage requirements. + +Experiments on 3RScan. The 3RScan dataset [123] comprises 1335 annotated indoor scenes, representing 432 distinct rooms, with 1178 scenes (385 rooms) allocated for training and 157 (47 rooms) designated for validation. The training and validation sets include semantically annotated 3D point clouds for each scene, with some captured over extended periods (e.g., several months) showcasing environmental changes. Annotations for graphs within the 3RScan dataset are provided in [125]. Due to the absence of such annotations in the test set, it was excluded from our experiments. Thus, we reorganized the original validation set, allocating 34 scenes (17 rooms) for validation and 123 scenes (30 rooms) for testing. For full reproducibility, we will publish this split. + +During testing, we examine all 123 scenes of 30 rooms within the test set, selecting query images from each scene. The next step involves matching this image against $N$ scenes (including the target) to ascertain whether the correct one could be identified by a method. This procedure is repeated for every image in each room. In total, all methods are tested on 30462 query images. Furthermore, we evaluate scene selection through two settings $N = 50$ and $N = 10$ . The latter setting emulates a scenario where a pre-selection strategy is employed, for example, utilizing a global scene descriptor. In image-based methods, we use all images from the database and determine the scene based on the retrieved image. + +Table 2: Retrieval recall in the temporal scenario on the test set of ScanNet dataset [21] (\%; target scene ranked within the top 1, 3, and 5 of the retrieved list) and storage requirements (MB) for methods using point clouds $(\mathcal{P})$ , images $(\mathcal{I})$ , and other modalities $(S, \mathcal{R})$ for map representation. The modality $\mathcal{A}$ is not available in the scene graphs predicted from [130]. The results are reported for scenarios where the target room is chosen from a subset of 10, 50 and the complete set of all (210) scenes. + +
MethodMap modalities10 scenes50 scenesAll scenesStorage (MB)
PIORt@1at3at5Rt@1at3at5Rt@1at3at5
LidarCLIP [43]✓XX19.447.567.64.714.822.25.915.021.90.7
LIP-Loc [108]✓XX10.327.043.61.96.08.11.83.14.01.7
OpenMask3D [113]✓✓X54.984.894.031.351.363.216.527.234.517.8
SceneGraphLoc (Ours)✓X✓54.181.491.929.047.458.013.526.434.29.3
SceneGraphLoc (Ours)✓✓✓78.592.798.361.683.291.653.469.878.79.3
CVNet [63]X✓X96.598.999.692.696.097.089.993.494.6239.1
AnyLoc [55]X✓X98.499.499.896.598.198.695.196.997.45720.3
+ +Additionally, the methods are evaluated under both static and temporal conditions. In the static scenario, the target scene graph for a given query image originates from the same scan, albeit from a different sequence, to ensure no image overlap. Conversely, in the temporal scenario, the scene graph is derived from a sequence captured at a different temporal stage than the query, introducing potential environmental changes. We do not show results for methods exploiting images in their maps in the static stage. + +The results are in Table 1. Despite being retrained, LidarCLIP and LIP-Loc are inaccurate, particularly in scenarios involving the selection of the target room from the entire scene set. LIP-Loc barely surpasses random selection. Although LidarCLIP exhibits marginally better accuracy, it remains inferior to alternative methods. The inferiority is because both methods learn shared embedding between entire scenes and query images of totally different objects due to viewpoint differences. OpenMask3D, while achieving better results than LidarCLIP and LIP-Loc, is less accurate than the proposed SceneGraphLoc. SceneGraphLoc, even when excluding the image modality $(\mathcal{I})$ , outperforms other cross-modal strategies significantly. Incorporating images significantly enhances its performance, positioning it close to that of image-based approaches but with three orders of magnitude smaller storage requirements. Also, the storage of SceneGraphLoc with and without images is the same due to its design of distilling knowledge into fixed-sized embeddings. An example scene is shown in Fig. 5. + +Experiments on ScanNet. To evaluate the generalization ability of our methods in real-world applications when scene graph annotations are not available, we conduct further experiments in the ScanNet dataset [21]. ScanNet encompasses 1613 monocular sequences of room-scale 3D scenes, offering 3D mesh reconstructions alongside the RGBD frame sequences utilized for the reconstructions. Given the absence of scene graph annotations in ScanNet, we run the SceneGraphFusion [130] on the RGBD sequences of scans for 3D reconstruction and scene graph prediction with 3D instance segmentation and object relationships (i.e., graph edges) within these graphs. As the process of scene graph prediction + +Table 3: Average time (ms) of obtaining the query image embedding $(t_{e_q})$ and of the retrieval from 10, 50, and all scenes from the 3RScan [123] and ScanNet [21] datasets. + +
Method3DRScan [123]ScanNet [21]
teqt10retrt50retrteqt10retrt50retrtallretr
LidarCLIP [43]4.10.10.34.90.10.20.6
LIP-Loc [108]2.70.10.24.10.10.20.5
OpenMask3D [113]41.54.87.420.155.41.14.5
SceneGraphLoc (Ours)28.00.31.516.61.33.717.0
CVNet [63]14.39.060.054.010.674.1311.3
AnyLoc [55]658.4354.61826.4243.068.2329.01451.1
+ +uses the RGBD frames of each scan, we avoid using those RGB images to match to the scene graph predicted of the scan. Thus, we only measure recall in the temporal scenario. Additionally, unlike 3RScan, the frame rate of RGBD sequences in ScanNet is high, and motion between consecutive frames is small. Thus, each database image is selected from every 25 consecutive frames in the sequence, for image-based methods [55, 63]. For a fair comparison, all the methods only use the same selected images for training and evaluation. + +For training, we use the official training set. We divide the official validation set, which includes 312 scenes, into two distinct subsets: the first 100 scenes form our validation set, while the remaining 212 are allocated for testing. To ensure full reproducibility, we will make this split publicly available. + +The results are in Table 2 for scenarios selecting the target room from subsets of 10, 50, and the entire set of 210 scenes. The performance of LIP-Loc shows a similar pattern to that observed on 3RScan, performing only slightly better than random selection. LidarCLIP shows a small improvement in accuracy. OpenMask3D attains an accuracy comparable to our proposed method without incorporating the image modality. Our proposed SceneGraphLoc, when including the image modality in its map, significantly outperforms all cross-modal approaches. Although there remains a gap in accuracy compared with methods that use extensive image collections as maps (such as CVNet and AnyLoc), SceneGraphLoc benefits from a database size three orders-of-magnitude smaller, highlighting its efficiency and effectiveness. We partly attribute this performance gap to the lack of object attributes in the dataset and the inaccurate instance segmentation predicted by [130]. More details can be found in the supp. mat. + +Processing time measured in milliseconds for various methods applied to the 3DRScan and ScanNet datasets, are detailed in Table 3. The computation time required to generate an embedding for the query image $(t_{e_q})$ is notably small across all methods, typically not exceeding a few tens of milliseconds, with the exception of AnyLoc, which runs for nearly a second. The retrieval phase for cross-modal approaches is generally limited to a few tens of milliseconds. However, methods such as CVNet and AnyLoc exhibit slower performance, due to searching through extensive image collections. When tasked with selecting from + +Table 4: Ablation study performed on the val. split of 3RScan [123], analysing map modalities ( $\mathcal{P}$ - point cloud, $\mathcal{I}$ - image, $\mathcal{A}$ - attributes, $\mathcal{S}$ - structure, $\mathcal{R}$ - relationships) and the method (Dino v2 or GCVit) to obtain the image embeddings. + +
Map modalitiesDino v2 [81]GCVit [38]
PIASRR@1@3@5Rt@1@3@5R@1@3@5Rt@1@3@5
✓45.281.993.743.979.591.424.656.076.923.254.777.3
✓✓56.385.695.054.884.095.044.276.991.243.475.189.4
✓✓✓58.487.395.956.585.793.643.375.891.341.572.889.3
✓✓✓✓63.786.895.862.787.496.345.375.590.546.676.290.2
✓---80.296.099.0---69.487.493.7
✓✓---84.797.599.6---73.289.795.9
✓✓✓✓✓---88.597.799.6---72.188.895.7
+ +a large number of images, the processing times of these methods can extend into the range of several hundred milliseconds or even reach upwards of a second. + +Ablation Studies. To understand the impact of the integration of different modalities, we provide an ablation study on the localization performance with multiple combinations of modalities. We use the validation split of the 3DRScan dataset. Additional ablation studies can be found in supplementary material. + +Table 4 displays the coarse localization performance of our method under the incorporation of distinct modalities $(\mathcal{P},\mathcal{I},\mathcal{A},\mathcal{S},\mathcal{R})$ within the pipeline. Additionally, the table reports results exploiting various backbones (Dino V2 [81] and GCVit [38]) for the extraction of image features. Similarly to the main experiments, we only show results on the temporal set when the map contains images. Employing Dino V2 for encoding both the query image and images within the map significantly enhances accuracy over using GCVit. The proposed method significantly outperforms both LidarCLIP and LIP-Loc (their results are in Table 1) even when using only point clouds in the map. Each additional modality contributes to the final accuracy, demonstrating that each plays a crucial role in enhancing the final performance when integrated into the pipeline. + +# 5 Conclusion + +In conclusion, we introduce SceneGraphLoc, a novel method for solving the novel problem of localizing an input image within a 3D scene graph-based multi-modal reference map. This approach outperforms existing cross-modal methods by a large margin. It achieves comparable accuracy to state-of-the-art image-based techniques with significantly lower storage requirements and faster processing speeds. Our experiments across the 3RScan and ScanNet datasets demonstrate the effectiveness of SceneGraphLoc, with the best performance achieved when integrating all proposed modalities. We believe that SceneGraphLoc is a step towards lightweight and efficient localization. + +# Acknowledgements. + +We would like to thank our colleagues Ganlin Zhang, Sayan Deb Sarkar, and Cathrin Elich for their valuable advice and insightful discussions throughout the course of this research. Their contributions and suggestions greatly enhanced the quality and depth of this work. This work was partially funded by Design++ initiative of ETH Zurich, by the ETH RobotX research grant, the Hasler Stiftung Research Grant via the ETH Zurich Foundation, an ETH AI Center postdoctoral research fellowship and an ETH Zurich Career Seed Award. + +# References + +1. Agia, C., Jatavallabhula, K.M., Khodeir, M., Miksik, O., Vineet, V., Mukadam, M., Paull, L., Shkurti, F.: Taskgraphy: Evaluating robot task planning over large 3D scene graphs. In: Conference on Robot Learning (CoRL) (2022) 4 +2. Arandjelovic, R., Gronat, P., Torii, A., Pajdla, T., Sivic, J.: NetVLAD: CNN architecture for weakly supervised place recognition. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2016) 2, 3 +3. Armeni, I., He, Z.Y., Gwak, J., Zamir, A.R., Fischer, M., Malik, J., Savarese, S.: 3D Scene Graph: A Structure for Unified semantics, 3d space, and camera. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2019) 2, 4 +4. Aubry, M., Russell, B.C., Sivic, J.: Painting-to-3d model alignment via discriminative visual elements. ACM Transactions On Graphics (TOG) (2014) 3, 4 +5. Aubry, M., Russell, B.C., Sivic, J.: Visual Geo-localization of Non-photographic Depictions via 2D-3D Alignment. In: Large-Scale Visual Geo-Localization (2016) 3 +6. Balntas, V., Li, S., Prisacariu, V.: RelocNet: Continuous Metric Learning Relocalisation using Neural Nets. In: European Conference on Computer Vision (ECCV) (2018) 3 +7. Bernreiter, L., Ott, L., Nieto, J., Siegwart, R., Cadena, C.: Spherical multi-modal place recognition for heterogeneous sensor systems. In: International Conference on Robotics and Automation (ICRA) (2021) 4 +8. Berton, G., Masone, C., Caputo, B.: Rethinking visual geo-localization for large-scale applications. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2022) 3 +9. Berton, G., Paolicelli, V., Masone, C., Caputo, B.: Adaptive-attentive geolocation from few queries: A hybrid approach. In: IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) (2021) 2, 3 +10. Bhayani, S., Sattler, T., Barath, D., Beliansky, P., Heikkilä, J., Kukelova, Z.: Calibrated and partially calibrated semi-generalized homographies. In: International Conference on Computer Vision (ICCV) (2021) 3 +11. Brachmann, E., Krull, A., Nowozin, S., Shotton, J., Michel, F., Gumhold, S., Rother, C.: DSAC - Differentiable RANSAC for Camera Localization. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2017) 3 +12. Brachmann, E., Rother, C.: Learning Less is More - 6D Camera Localization via 3D Surface Regression. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2018) 3 + +13. Brachmann, E., Rother, C.: Visual camera re-localization from RGB and RGB-D images using DSAC. Transactions on Pattern Analysis and Machine Intelligence (PAMI) (2021) 3 +14. Brejcha, J., Lukac, M., Hold-Geoffroy, Y., Wang, O., Cadik, M.: LandscapeAR: Large Scale Outdoor Augmented Reality by Matching Photographs with Terrain Models Using Learned Descriptors. In: European Conference on Computer Vision (ECCV) (2020) 3, 4 +15. Cadik, M., Sykora, D., Lee, S.: Automated outdoor depth-map generation and alignment. Computers & Graphics (2018) 3 +16. Castle, R., Klein, G., Murray, D.W.: Video-rate localization in multiple maps for wearable augmented reality. In: IEEE International Symposium on Wearable Computers (2008) 3 +17. Cavallari, T., Bertinetto, L., Mukhoti, J., Torr, P., Golodetz, S.: Let's take this online: Adapting scene coordinate regression network predictions for online RGB-D camera relocalisation. In: International Conference on 3D Vision (3DV) (2019) 3 +18. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention Mask Transformer for Universal Image Segmentation. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2022) 8 +19. Csurka, G., Dance, C., Fan, L., Willamowski, J., Bray, C.: Visual categorization with bags of keypoints. In: European Conference on Computer Vision (ECCV) Workshops (2004) 7 +20. Curless, B., Levoy, M.: A volumetric method for building complex models from range images. In: Annual conference on Computer graphics and interactive techniques (1996) 4 +21. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2017) 12, 13 +22. Doan, A.D., Latif, Y., Chin, T.J., Liu, Y., Do, T.T., Reid, I.: Scalable place recognition under appearance change for autonomous driving. In: International Conference on Computer Vision (ICCV) (2019) 2 +23. Engelmann, F., Manhardt, F., Niemeyer, M., Tateno, K., Tombari, F.: OpenNeRF: Open Set 3D Neural Scene Segmentation with Pixel-Wise Features and Rendered Novel Views. In: International Conference on Learning Representations (ICLR) (2024) 10 +24. Fan, L., Pang, Z., Zhang, T., Wang, Y.X., Zhao, H., Wang, F., Wang, N., Zhang, Z.: Embracing single stride 3d object detector with sparse transformer. International Conference on Computer Vision and Pattern Recognition (CVPR) (2022) 10 +25. Gadre, S.Y., Ehsani, K., Song, S., Mottaghi, R.: Continuous scene representations for embodied ai. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2022) 4 +26. Gao, P., Liang, J., Shen, Y., Son, S., Lin, M.C.: Visual, Spatial, Geometric-Preserved Place Recognition for Cross-View and Cross-Modal Collaborative Perception. In: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) (2023) 4 +27. Garg, S., Fischer, T., Milford, M.: Where is your place, visual place recognition? International Joint Conference on Artificial Intelligence (IJCAI) (2021) 3 +28. Garg, S., Suenderhauf, N., Milford, M.: Semantic-geometric visual place recognition: a new perspective for reconciling opposing views. International Journal on Robotics Research (IJRR) (2019) 2, 4 + +29. Garg, S., Sünderhauf, N., Dayoub, F., Morrison, D., Cosgun, A., Carneiro, G., Wu, Q., Chin, T.J., Reid, I., Gould, S., et al.: Semantics for robotic mapping, perception and interaction: A survey. Foundations and Trends in Robotics (2020) 4 +30. Georgakis, G., Karanam, S., Wu, Z., Kosecka, J.: Learning local rgb-to-cad correspondences for object pose estimation. In: International Conference on Computer Vision (ICCV) (2019) 4 +31. Germain, H., Bourmaud, G., Lepetit, V.: Sparse-to-Dense Hypercolumn Matching for Long-Term Visual Localization. In: International Conference on 3D Vision (3DV) (2019) 3 +32. Germain, H., Bourmaud, G., Lepetit, V.: S2DNet: Learning Image Features for Accurate Sparse-to-Dense Matching. In: European Conference on Computer Vision (ECCV) (2020) 3 +33. Ghiasi, G., Gu, X., Cui, Y., Lin, T.Y.: Scaling Open-Vocabulary Image Segmentation with Image-Level Labels. In: European Conference on Computer Vision (ECCV) (2022) 10 +34. Grabner, A., Roth, P.M., Lepetit, V.: 3D Pose Estimation and 3D Model Retrieval for Objects in the Wild. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2018) 4 +35. Grelsson, B., Robinson, A., Felsberg, M., Khan, F.S.: GPS-level accurate camera localization with HorizonNet. Journal of Field Robotics (2020) 3 +36. Gumeli, C., Dai, A., Nießner, M.: ROCA: Robust CAD Model Retrieval and Alignment from a Single Image. arXiv preprint arXiv:2112.01988 (2021) 4 +37. Hanocka, R., Metzer, G., Giryes, R., Cohen-Or, D.: Point2mesh: A self-prior for deformable meshes. arXiv preprint arXiv:2005.11084 (2020) 4 +38. Hatamizadeh, A., Yin, H., Heinrich, G., Kautz, J., Molchanov, P.: Global context vision transformers. In: International Conference on Machine Learning (ICML) (2023) 14 +39. Hausler, S., Jacobson, A., Milford, M.: Multi-process fusion: Visual place recognition using multiple image processing methods. IEEE Robotics and Automation Letters (RA-L) (2019) 2 +40. Hausler, S., Garg, S., Xu, M., Milford, M., Fischer, T.: Patch-netvlad: Multiscale fusion of locally-global descriptors for place recognition. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2021) 2, 3 +41. He, K., Zhang, X., Ren, S., Sun, J.: Deep Residual Learning for Image Recognition. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2016) 8 +42. Heng, L., Choi, B., Cui, Z., Geppert, M., Hu, S., Kuan, B., Liu, P., Nguyen, R., Yeo, Y.C., Geiger, A., et al.: Project Autovision: Localization and 3D Scene Perception for an Autonomous Vehicle with a Multi-Camera System. In: International Conference on Robotics and Automation (ICRA) (2019) 3 +43. Hess, G., Tondersski, A., Petersson, C., Åström, K., Svensson, L.: Lidarclip or: How i learned to talk to point clouds. In: IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) (2024) 4, 10, 11, 12, 13 +44. Hodan, T.: Pose Estimation of Specific Rigid Objects. Ph.D. thesis (2021) 4 +45. Hodan, T., Barath, D., Matas, J.: EPOS: Estimating 6D Pose of Objects With Symmetries. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2020) 4 +46. Hodan, T., Zabulis, X., Lourakis, M.I.A., Obdrzalek, S., Matas, J.: Detection and fine 3D pose estimation of texture-less objects in RGB-D images. In: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) (2015) 4 + +47. Hu, S., Feng, M., Nguyen, R.H.M., Lee, G.H.: CVM-Net: Cross-View Matching Network for Image-Based Ground-to-Aerial Geo-Localization. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2018) 4 +48. Hu, S., Lee, G.H.: Image-Based GeoLocalization Using Satellite Imagery. International Journal on Computer Vision (IJCV) (2019) 4 +49. Hughes, N., Chang, Y., Carlone, L.: Hydra: A real-time spatial perception system for 3D scene graph construction and optimization. arXiv preprint arXiv:2201.13360 (2022) 4, 5, 7 +50. Ibrahimi, S., van Noord, N., Alpherts, T., Worring, M.: Inside out visual place recognition. In: British Machine Vision Conference (2021) 2, 3 +51. Irschara, A., Zach, C., Frahm, J.M., Bischof, H.: From Structure-from-Motion Point Clouds to Fast Location Recognition. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2009) 3 +52. Izadi, S., Kim, D., Hilliges, O., Molyneaux, D., Newcombe, R., Kohli, P., Shotton, J., Hodges, S., Freeman, D., Davison, A., et al.: Kinectfusion: real-time 3d reconstruction and interaction using a moving depth camera. In: ACM symposium on User interface software and technology (2011) 4 +53. Ji, X., Wei, J., Wang, Y., Shang, H., Kneip, L.: Cross-modal Place Recognition in Image Databases using Event-based Sensors. arXiv preprint arXiv:2307.01047 (2023) 4 +54. Kabalar, J., Wu, S.C., Wald, J., Tateno, K., Navab, N., Tombari, F.: Towards long-term retrieval-based visual localization in indoor environments with changes. IEEE Robotics and Automation Letters (2023) 4 +55. Keetha, N., Mishra, A., Karhade, J., Jatavallabhula, K.M., Scherer, S., Krishna, M., Garg, S.: Anyloc: Towards universal visual place recognition. IEEE Robotics and Automation Letters (RA-L) (2023) 2, 3, 10, 11, 12, 13 +56. Kendall, A., Cipolla, R.: Geometric loss functions for camera pose regression with deep learning. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2017) 3 +57. Kendall, A., Grimes, M., Cipolla, R.: PoseNet: A Convolutional Network for Real-Time 6-DOF Camera Relocalization. In: International Conference on Computer Vision (ICCV) (2015) 3 +58. Khaliq, A., Ehsan, S., Chen, Z., Milford, M., McDonald-Maier, K.: A holistic visual place recognition approach using lightweight CNNs for significant viewpoint and appearance changes. IEEE Transactions on Robotics (T-RO) (2020) 2 +59. Kim, H.J., Dunn, E., Frahm, J.M.: Learned contextual feature reweighting for image geolocation. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2017) 3 +60. Kim, U.H., Park, J.M., Song, T.J., Kim, J.H.: 3-D scene graph: A sparse and semantic representation of physical environments for intelligent agents. IEEE transactions on cybernetics (2019) 4 +61. Kutulakos, K.N., Seitz, S.M.: A theory of shape by space carving. International Journal on Computer Vision (IJCV) (2000) 4 +62. Labbe, Y., Carpentier, J., Aubry, M., Sivic, J.: CosyPose: Consistent multi-view multi-object 6D pose estimation. In: European Conference on Computer Vision (ECCV) (2020) 4 +63. Lee, S., Seong, H., Lee, S., Kim, E.: Correlation verification for image retrieval. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2022) 10, 11, 12, 13 + +64. Li, Y., Snavely, N., Huttenlocher, D.P., Fua, P.: Worldwide Pose Estimation Using 3D Point Clouds. In: European Conference on Computer Vision (ECCV) (2012) 3 +65. Lim, H., Sinha, S.N., Cohen, M.F., Uysettendaele, M.: Real-time Image-based 6-DoF Localization in Large-scale Environments. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2012) 3 +66. Lin, T.Y., Cui, Y., Belongie, S.J., Hays, J.: Learning deep representations for ground-to-aerial geolocation. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2015) 4 +67. Lin, Z., Zhang, Z., Wang, M., Shi, Y., Wu, X., Zheng, Y.: Multi-modal contrastive representation learning for entity alignment. arXiv preprint arXiv:2209.00891 (2022) 9 +68. Liu, L., Li, H., Dai, Y.: Efficient Global 2D-3D Matching for Camera Localization in a Large-Scale 3D Map. In: International Conference on Computer Vision (ICCV) (2017) 3 +69. Liu, L., Li, H., Dai, Y.: Stochastic attraction-repulsion embedding for large scale image localization. In: International Conference on Computer Vision (ICCV) (2019) 2 +70. Looper, S., Rodriguez-Puigvert, J., Siegwart, R., Cadena, C., Schmid, L.: 3D VSG: Long-term semantic scene change prediction through 3D variable scene graphs. In: International Conference on Robotics and Automation (ICRA) (2023) 4 +71. Lynen, S., Zeisl, B., Aiger, D., Bosse, M., Hesch, J., Pollefeys, M., Siegwart, R., Sattler, T.: Large-scale, real-time visual-inertial localization revisited. International Journal on Robotics Research (IJRR) (2020) 3 +72. Lynen, S., Zeisl, B., Aiger, D., Bosse, M., Hesch, J., Pollefeys, M., Siegwart, R., Sattler, T.: Large-scale, real-time visual-inertial localization revisited. International Journal on Robotics Research (IJRR) (2020) 3 +73. Mescheder, L., Oechsle, M., Niemeyer, M., Nowozin, S., Geiger, A.: Occupancy networks: Learning 3d reconstruction in function space. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2019) 4 +74. Miao, Y., Armeni, I., Pollefeys, M., Barath, D.: Volumetric Semantically Consistent 3D Panoptic Mapping. arXiv preprint arXiv:2309.14737 (2024) 4 +75. Miao, Y., Li, C., Li, Z., Yang, Y., Yu, X.: A novel algorithm of ship structure modeling and target identification based on point cloud for automation in bulk cargo terminals. Measurement and Control (2021) 4 +76. Mihajlovic, M., Weder, S., Pollefeys, M., Oswald, M.R.: DeepSurfels: Learning online appearance fusion. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2021) 4 +77. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM (2021) 4 +78. Moreau, A., Piasco, N., Tsishkou, D., Stanciulescu, B., de La Fortelle, A.: LENS: Localization enhanced by neRF synthesis. In: Conference on Robot Learning (CoRL) (2021) 3 +79. Murez, Z., Van As, T., Bartolozzi, J., Sinha, A., Badrinarayanan, V., Rabinovich, A.: Atlas: End-to-end 3d scene reconstruction from posed images. In: European Conference on Computer Vision (ECCV) (2020) 4 +80. Oechsle, M., Mescheder, L., Niemeyer, M., Strauss, T., Geiger, A.: Texture fields: Learning texture representations in function space. In: International Conference on Computer Vision (ICCV) (2019) 4 + +81. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023) 7, 8, 14 +82. Panek, V., Kukelova, Z., Sattler, T.: Meshloc: Mesh-based visual localization. In: European Conference on Computer Vision (ECCV) (2022) 3, 4 +83. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2019) 4 +84. Peng, G., Yue, Y., Zhang, J., Wu, Z., Tang, X., Wang, D.: Semantic reinforced attention learning for visual place recognition. In: International Conference on Robotics and Automation (ICRA) (2021) 3 +85. Peng, G., Zhang, J., Li, H., Wang, D.: Attentional pyramid pooling of salient visual residuals for place recognition. In: International Conference on Computer Vision (ICCV) (2021) 3 +86. Peng, S., Genova, K., Jiang, C.M., Tagliasacchi, A., Pollefeys, M., Funkhouser, T.: OpenScene: 3D Scene Understanding with Open Vocabularies. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2023) 10 +87. Peng, S., Niemeyer, M., Mescheder, L., Pollefeys, M., Geiger, A.: Convolutional occupancy networks. In: European Conference on Computer Vision (ECCV) (2020) 4 +88. Pion, N., Humenberger, M., Csurka, G., Cabon, Y., Sattler, T.: Benchmarking image retrieval for visual localization. In: International Conference on 3D Vision (3DV) (2020) 3 +89. Plotz, T., Roth, S.: Automatic Registration of Images to Untextured Geometry Using Average Shading Gradients. International Journal on Computer Vision (IJCV) (2017) 3 +90. Ponimatkin, G., Labbe, Y., Russell, B., Aubry, M., Sivic, J.: Focal length and object pose estimation via render and compare. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2022) 4 +91. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2017) 4, 6 +92. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. International Conference on Machine Learning (ICML) (2021) 10 +93. Ramalingam, S., Bouaziz, S., Sturm, P.F., Brand, M.: SKYLINE2GPS: Localization in urban canyons using omni-skylines. In: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) (2010) 3 +94. Ravichandran, Z., Peng, L., Hughes, N., Griffith, J., Carlone, L.: Hierarchical representations and explicit memory: Learning effective navigation policies on 3D scene graphs using graph neural networks. In: International Conference on Robotics and Automation (ICRA) (2022) 4 +95. Rosinol, A., Violette, A., Abate, M., Hughes, N., Chang, Y., Shi, J., Gupta, A., Carlone, L.: Kimera: From SLAM to spatial perception with 3D dynamic scene graphs. International Journal on Robotics Research (IJRR) (2021) 4, 5 +96. Rosinol, A., Gupta, A., Abate, M., Shi, J., Carlone, L.: 3D dynamic scene graphs: Actionable spatial perception with places, objects, and humans. arXiv preprint arXiv:2002.06289 (2020) 4 + +97. Sarkar, S.D., Miksik, O., Pollefeys, M., Barath, D., Armeni, I.: SGAligner: 3D Scene Alignment with Scene Graphs. International Conference on Computer Vision (ICCV) (2023) 4, 6, 7 +98. Sarlin, P.E., Cadena, C., Siegwart, R., Dymczyk, M.: From Coarse to Fine: Robust Hierarchical Localization at Large Scale. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2019) 3 +99. Sarlin, P.E., DeTone, D., Malisiewicz, T., Rabinovich, A.: Superglue: Learning feature matching with graph neural networks. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2020) 3 +100. Sarlin, P.E., DeTone, D., Yang, T.Y., Avetisyan, A., Straub, J., Malisiewicz, T., Bulò, S.R., Newcombe, R., Kontschieder, P., Balntas, V.: OrienterNet: Visual Localization in 2D Public Maps with Neural Matching. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2023) 2, 4 +101. Sarlin, P.E., Unagar, A., Larsson, M., Germain, H., Toft, C., Larsson, V., Pollefeys, M., Lepetit, V., Hammarstrand, L., Kahl, F., et al.: Back to the Feature: Learning Robust Camera Localization from Pixels to Pose. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2021) 3 +102. Sattler, T., Leibe, B., Kobbelt, L.: Efficient & Effective Prioritized Matching for Large-Scale Image-Based Localization. Transactions on Pattern Analysis and Machine Intelligence (PAMI) (2017) 3 +103. Sattler, T., Zhou, Q., Pollefeys, M., Leal-Taixe, L.: Understanding the limitations of cnn-based absolute camera pose regression. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2019) 3 +104. Savinov, N., Hane, C., Ladicky, L., Pollefeys, M.: Semantic 3d reconstruction with continuous regularization and ray potentials using a visibility consistency constraint. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2016) 4 +105. Schonberger, J.L., Pollefeys, M., Geiger, A., Sattler, T.: Semantic visual localization. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2018) 3 +106. Sepulveda, G., Niebles, J., Soto, A.: A deep learning based behavioral approach to indoor autonomous navigation. In: International Conference on Robotics and Automation (ICRA) (2018) 4 +107. Shan, Q., Wu, C., Curless, B., Furukawa, Y., Hernandez, C., Seitz, S.M.: Accurate geo-registration by ground-to-aerial image matching. In: International Conference on 3D Vision (3DV) (2014) 3, 4 +108. Shubodh, S., Omama, M., Zaidi, H., Parihar, U.S., Krishna, M.: Lip-loc: Lidar image pretraining for cross-modal localization. In: IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) (2024) 4, 10, 11, 12, 13 +109. Sibling, D., Sattler, T., Leibe, B., Kobbelt, L.: SIFT-Realistic Rendering. In: International Conference on 3D Vision (3DV) (2013) 3 +110. Steiger Mueller, M., Sattler, T., Pollefeys, M., Jutz, B.: Image-to-image translation for enhanced feature matching, image retrieval and visual localization. ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences (2019) 3 +111. Stuckler, J., Behnke, S.: Multi-resolution surfel maps for efficient dense 3D modeling and tracking. Journal of Visual Communication and Image Representation (2014) 4 +112. Svarm, L., Enqvist, O., Kahl, F., Oskarsson, M.: City-Scale Localization for Cameras with Known Vertical Direction. Transactions on Pattern Analysis and Machine Intelligence (PAMI) (2017) 3 + +113. Takmaz, A., Fedele, E., Sumner, R.W., Pollefeys, M., Tombari, F., Engelmann, F.: OpenMask3D: Open-Vocabulary 3D Instance Segmentation. In: International Conference on Neural Information Processing Systems (NeurIPS) (2023) 7, 10, 11, 12, 13 +114. Tewari, A., Thies, J., Mildenhall, B., Srinivasan, P., Tretschk, E., Yifan, W., Lassner, C., Sitzmann, V., Martin-Brualla, R., Lombardi, S., et al.: Advances in neural rendering. In: Computer Graphics Forum (2022) 4 +115. Tomesek, J., Cadik, M., Brejcha, J.: CrossLocate: Cross-modal Large-scale Visual Geo-Localization in Natural Environments using Rendered Modalities. In: IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) (2022) 3, 4 +116. Torii, A., Arandjelovic, R., Sivic, J., Okutomi, M., Pajdla, T.: 24/7 place recognition by view synthesis. Transactions on Pattern Analysis and Machine Intelligence (PAMI) (2018) 2 +117. Torii, A., Taira, H., Sivic, J., Pollefeys, M., Okutomi, M., Pajdla, T., Sattler, T.: Are large-scale 3d models really necessary for accurate visual localization? Transactions on Pattern Analysis and Machine Intelligence (PAMI) (2021) 2 +118. Valentin, J., Nießner, M., Shotton, J., Fitzgibbon, A., Izadi, S., Torr, P.: Exploiting Uncertainty in Regression Forests for Accurate Camera Relocalization. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2015) 3 +119. Velicković, P., Cucurull, G., Casanova, A., Romero, A., Lio, P., Bengio, Y.: Graph Attention Networks. In: International Conference on Learning Representations (ICLR) (2018) 7 +120. Ventura, J., Kukelova, Z., Sattler, T., Baráth, D.: Absolute Pose from One or Two Scaled and Oriented Features. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2024) 3 +121. Viswanathan, A., Rodrigues Pires, B., Huber, D.F.: Vision based robot localization by ground to satellite matching in GPS-denied situations. In: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) (2014) 4 +122. Walch, F., Hazirbas, C., Leal-Taixe, L., Sattler, T., Hilsenbeck, S., Cremers, D.: Image-Based Localization Using LSTMs for Structured Feature Correlation. In: International Conference on Computer Vision (ICCV) (2017) 3 +123. Wald, J., Avetisyan, A., Navab, N., Tombari, F., Niefsner, M.: Rio: 3d object instance re-localization in changing indoor environments. In: International Conference on Computer Vision (ICCV) (2019) 8, 9, 10, 11, 13, 14 +124. Wald, J., Dhamo, H., Navab, N., Tombari, F.: Learning 3d semantic scene graphs from 3d indoor reconstructions. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2020) 2, 4 +125. Wald, J., Dhamo, H., Navab, N., Tombari, F.: Learning 3d semantic scene graphs from 3d indoor reconstructions. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2020) 11 +126. Wang, S., Kannala, J., Barath, D.: DGC-GNN: Descriptor-free Geometric-Color Graph Neural Network for 2D-3D Matching. International Conference on Computer Vision and Pattern Recognition (CVPR) (2023) 3 +127. Warburg, F., Hauberg, S., Lopez-Antequera, M., Gargallo, P., Kuang, Y., Civera, J.: Mapillary street-level sequences: A dataset for lifelong place recognition. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2020) 2, 3 + +128. Weder, S., Schonberger, J.L., Pollefeys, M., Oswald, M.R.: Neuralfusion: Online depth fusion in latent space. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2021) 4 +129. Workman, S., Souvenir, R., Jacobs, N.: Wide-Area Image Geolocation with Aerial Reference Imagery. In: International Conference on Computer Vision (ICCV) (2015) 4 +130. Wu, S.C., Wald, J., Tateno, K., Navab, N., Tombari, F.: Scenegraphfusion: Incremental 3d scene graph prediction from rgb-d sequences. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2021) 4, 7, 12, 13 +131. Ying, Z., Yuan, X., Yang, B., Song, Y., Xu, Q., Zhou, F., Sheng, W.: RP-SG: Relation Prediction in 3D Scene Graphs for Unobserved Objects Localization. IEEE Robotics and Automation Letters (RA-L) (2023) 4 +132. Zaffar, M., Garg, S., Milford, M., et al.: Vpr-bench: An open-source visual place recognition evaluation framework with quantifiable viewpoint and appearance change. International Journal on Computer Vision (IJCV) (2021) 2 +133. Zeisl, B., Sattler, T., Pollefeys, M.: Camera pose voting for large-scale image-based localization. In: International Conference on Computer Vision (ICCV) (2015) 3 +134. Zhang, C., Yu, J., Song, Y., Cai, W.: Exploiting edge-oriented reasoning for 3D point-based scene graph analysis. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2021) 4 +135. Zhang, G., Larsson, V., Barath, D.: Revisiting Rotation Averaging: Uncertainties and Robust Losses. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2023) 3 +136. Zhang, S., Hao, A., Qin, H.: Knowledge-inspired 3D scene graph prediction in point cloud. International Conference on Neural Information Processing Systems (NeurIPS) (2021) 4 +137. Zhang, W., Kosecka, J.: Image based localization in urban environments. In: International Symposium on 3D Data Processing, Visualization, and Transmission (2006) 3 +138. Zhang, Z., Sattler, T., Scaramuzza, D.: Reference Pose Generation for Long-term Visual Localization via Learned Features and View Synthesis. International Journal on Computer Vision (IJCV) (2020) 3 +139. Zhao, L., Gatsis, K., Papachristodoulou, A.: Stable and Safe Reinforcement Learning via a Barrier-Lyapunov Actor-Critic Approach. In: IEEE Conference on Decision and Control (CDC) (2023) 3 +140. Zhao, L., Miao, K., Gatsis, K., Papachristodoulou, A.: Stable and safe human-aligned reinforcement learning through neural ordinary differential equations. arXiv preprint arXiv:2401.13148 (2024) 3 +141. Zheng, E., Wu, C.: Structure From Motion Using Structure-Less Resection. In: International Conference on Computer Vision (ICCV) (2015) 3 +142. Zhou, Q., Agostinho, S., Osep, A., Leal-Taixe, L.: Is Geometry Enough for Matching in Visual Localization? In: European Conference on Computer Vision (ECCV) (2022) 2, 3 +143. Zurbrügg, R., Liu, Y., Engelmann, F., Kumar, S., Hutter, M., Patil, V., Yu, F.: ICGNet: A Unified Approach for Instance-Centric Grasping. In: International Conference on Robotics and Automation (ICRA) (2024) 2 \ No newline at end of file diff --git a/2024/SceneGraphLoc_ Cross-Modal Coarse Visual Localization on 3D Scene Graphs/images.zip b/2024/SceneGraphLoc_ Cross-Modal Coarse Visual Localization on 3D Scene Graphs/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7ba6ab479e0fc67c364b289246e291344f8668c8 --- /dev/null +++ b/2024/SceneGraphLoc_ Cross-Modal Coarse Visual Localization on 3D Scene Graphs/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fd843c051f947491d5854e62f118fd0c95293eeb472d2e7820cbd7ea1f08565b +size 417580 diff --git a/2024/SceneGraphLoc_ Cross-Modal Coarse Visual Localization on 3D Scene Graphs/layout.json b/2024/SceneGraphLoc_ Cross-Modal Coarse Visual Localization on 3D Scene Graphs/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..fa9d9eaaf0d70b434fc7b7706f0b3983ffae4e93 --- /dev/null +++ b/2024/SceneGraphLoc_ Cross-Modal Coarse Visual Localization on 3D Scene Graphs/layout.json @@ -0,0 +1,14387 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 153, + 111, + 462, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 111, + 462, + 148 + ], + "spans": [ + { + "bbox": [ + 153, + 111, + 462, + 148 + ], + "type": "text", + "content": "SceneGraphLoc: Cross-Modal Coarse Visual Localization on 3D Scene Graphs" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 162, + 167, + 452, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 167, + 452, + 193 + ], + "spans": [ + { + "bbox": [ + 162, + 167, + 452, + 193 + ], + "type": "text", + "content": "Yang Miao1, Francis Engelmann1,2, Olga Vysotska1, Federico Tombari2,3, Marc Pollefeys1,4, and Daniel Béla Baráth1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 211, + 201, + 402, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 201, + 402, + 213 + ], + "spans": [ + { + "bbox": [ + 211, + 201, + 402, + 213 + ], + "type": "text", + "content": "1ETH Zurich 2Google 3TU Munich 4Microsoft" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 134, + 233, + 480, + 345 + ], + "blocks": [ + { + "bbox": [ + 134, + 233, + 480, + 345 + ], + "lines": [ + { + "bbox": [ + 134, + 233, + 480, + 345 + ], + "spans": [ + { + "bbox": [ + 134, + 233, + 480, + 345 + ], + "type": "image", + "image_path": "b97064f8bd510427d17b3f0f97fd8f0aded53b6b81ff72b84f95c056655847ec.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 352, + 482, + 374 + ], + "lines": [ + { + "bbox": [ + 132, + 352, + 482, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 352, + 482, + 374 + ], + "type": "text", + "content": "Fig. 1: SceneGraphLoc addresses the novel problem of localizing a query image in a database of 3D scenes represented as compact multi-modal 3D scene graphs." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 395, + 457, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 395, + 457, + 559 + ], + "spans": [ + { + "bbox": [ + 159, + 395, + 457, + 559 + ], + "type": "text", + "content": "Abstract. We introduce the task of localizing an input image within a multi-modal reference map represented by a collection of 3D scene graphs. These scene graphs comprise multiple modalities, including object-level point clouds, images, attributes, and relationships between objects, offering a lightweight and efficient alternative to conventional methods that rely on extensive image databases. Given these modalities, the proposed method SceneGraphLoc learns a fixed-sized embedding for each node (i.e., representing object instances) in the scene graph, enabling effective matching with the objects visible in the input query image. This strategy significantly outperforms other cross-modal methods, even without incorporating images into the map representation. With images, SceneGraphLoc achieves performance close to that of state-of-the-art techniques depending on large image databases, while requiring three orders-of-magnitude less storage and operating orders-of-magnitude faster. Code and models are available at https://scenegraphloc.github.io." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 160, + 570, + 441, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 570, + 441, + 582 + ], + "spans": [ + { + "bbox": [ + 160, + 570, + 441, + 582 + ], + "type": "text", + "content": "Keywords: Coarse Localization " + }, + { + "bbox": [ + 160, + 570, + 441, + 582 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 570, + 441, + 582 + ], + "type": "text", + "content": " 3D Scene Graph " + }, + { + "bbox": [ + 160, + 570, + 441, + 582 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 570, + 441, + 582 + ], + "type": "text", + "content": " Multi-modality" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 603, + 230, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 603, + 230, + 616 + ], + "spans": [ + { + "bbox": [ + 132, + 603, + 230, + 616 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Coarse visual localization, or place recognition, is a fundamental component in computer vision and robotics applications, defined as the task of identifying the approximate location where a query image was taken, given a certain tolerance" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 148, + 115, + 465, + 223 + ], + "blocks": [ + { + "bbox": [ + 148, + 115, + 465, + 223 + ], + "lines": [ + { + "bbox": [ + 148, + 115, + 465, + 223 + ], + "spans": [ + { + "bbox": [ + 148, + 115, + 465, + 223 + ], + "type": "image", + "image_path": "6b731af108ffd1c2702a35f20734a8083aac8a368bca4f0292c01cee44ba433e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 232, + 482, + 331 + ], + "lines": [ + { + "bbox": [ + 130, + 232, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 232, + 482, + 331 + ], + "type": "text", + "content": "Fig. 2: Overview. The training phase is represented by orange arrows, while blue arrows denote the inference phase. During training, a query image and its associated 3D scene graph form a positive sample within a contrastive learning framework, where negative samples are generated by associating scene graphs of different scenes with the same query image. The objective is to learn the embeddings of both the graph and the image so that embeddings of the positive pair are drawn closer, whereas those of the negative pair are pushed apart. In the inference phase, the task involves assigning the correct scene graph to a given query image from a selection of multiple graphs, achieved by optimizing the cosine similarity between their embeddings." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 371, + 482, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 371, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 371, + 482, + 456 + ], + "type": "text", + "content": "level [2,9,22,28,39,40,50,58,69,116,117,127,132,143]. This capability is crucial for estimating the state of robots and is widely utilized in autonomous, unmanned aerial, terrestrial, and underwater vehicles, as well as AR/VR devices. The task is typically approached as an image retrieval problem, where the image to be localized is compared against a large database of posed images and, optionally, a 3D reconstruction of the scene. The most similar images retrieved from the database are used to estimate the precise location of the query image." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 459, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 459, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 459, + 482, + 567 + ], + "type": "text", + "content": "The challenge with current state-of-the-art image-based coarse localization methods, such as AnyLoc [55], is their dependency on extensive image databases, which are not only storage-heavy but also slow to query, despite optimizations through hashing and other tricks. Moreover, these methods typically necessitate that the query and database share the same modality, limiting the scope of their application. Cross-modal approaches, such as [100,142], which attempt to bridge different types of data, often restrict their scope to connecting two modalities at a time (e.g., image-to-point cloud or image-to-bird's eye view map), one for the query and one for the database, thus narrowing their potential applications." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "This paper addresses the novel challenge of localizing a query image within a database that is represented not by conventional images but by the 3D scene graph [3,124] that integrates a diverse set of modalities, including point clouds, images, semantics, object attributes, and relationships. We tackle this problem by learning to map these modalities into a unified embedding space, thus allowing us to represent indoor scenes compactly through their objects (e.g., table and wall). This method enables the creation of small, efficient databases and significantly accelerates the coarse localization process." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "Y. Miao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 449, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 449, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 449, + 128 + ], + "type": "text", + "content": "Contributions. The primary contributions of this paper are as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 135, + 136, + 481, + 183 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 135, + 136, + 481, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 136, + 481, + 159 + ], + "spans": [ + { + "bbox": [ + 135, + 136, + 481, + 159 + ], + "type": "text", + "content": "1. Introducing a novel problem: cross-modal localization of a query image within 3D scene graphs incorporating a mixture of modalities." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 135, + 159, + 481, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 159, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 135, + 159, + 481, + 183 + ], + "type": "text", + "content": "2. SceneGraphLoc, a new method for the coarse localization of an input image given a reference map represented by a database of 3D scene graphs." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 190, + 481, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 190, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 130, + 190, + 481, + 251 + ], + "type": "text", + "content": "Even without incorporating images into the map, SceneGraphLoc largely outperforms other cross-modal methods on two large-scale, real-world indoor datasets. With images, SceneGraphLoc achieves performance close to that of state-of-the-art image-based methods while requiring three orders-of-magnitude less storage and operating orders-of-magnitude faster. The method is visualized in Fig. 1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 269, + 237, + 282 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 269, + 237, + 282 + ], + "spans": [ + { + "bbox": [ + 132, + 269, + 237, + 282 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 294, + 481, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 294, + 481, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 294, + 481, + 426 + ], + "type": "text", + "content": "Localization, the process of determining the position and orientation of an agent within a pre-built map, is pivotal across various domains such as mobile robots [42, 65, 139, 140], and augmented reality [16, 71]. The differentiation in localization techniques arises from their scene representation methods - be it through explicit 3D models [31, 32, 51, 64, 68, 72, 98, 99, 101, 102, 105, 112, 133, 135], sets of posed images [10, 88, 137, 141], or implicitly via neural network weights [6, 11-13, 17, 56, 57, 78, 118, 122] - and their approach to camera pose estimation, whether by 2D-3D [31, 32, 98, 99, 102, 112, 120, 126, 133] or 2D-2D [10, 141] matches, or through a composite of base poses [56, 57, 78, 88, 103, 122]. In practice, localization comprises two main steps: a coarse and precise stage. Here, we focus on the coarse step, finding potential locations of a query image." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 426, + 481, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 481, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 481, + 545 + ], + "type": "text", + "content": "Coarse Localization (or place recognition) is often cast as an image retrieval problem [2,8,9,27,40,50,55,59,84,85,127] that consists of two phases. In the offline indexing phase, a reference map (image or point cloud database) is gathered. In the online retrieval phase, a query image - captured during a future traverse of the environment - is localized coarsely by retrieving the closest match to this image in the reference map. Recent methods perform the retrieval using learned embeddings that are produced by a feature extraction backbone equipped with a head that implements some form of aggregation or pooling, the most notable being NetVLAD [2]. While these methods achieve impressive results, they are limited to a single modality (e.g., images) and require large databases." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 546, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 481, + 665 + ], + "type": "text", + "content": "Localization using multi-modal data. While dense mesh models are not as widely adopted as sparse Structure-from-Motion-based approaches, they have nonetheless been the focus of considerable research efforts [4,5,14,15,35,82,89,93, 107,109,110,115,138,142]. The body of prior work can be broadly segmented into two main strategies: The first entails the precise alignment of actual images with three-dimensional models (which may be coarse) through applying specialized techniques such as ones using contours [89] or skylines [93]. The second strategy emphasizes the identification and matching of local image features [4,5,14,35, 107,109,110,115,126,138,142], a method that has gained traction for its ability to match real-world images with non-photorealistic renderings of colored meshes, or" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "type": "text", + "content": "SceneGraphLoc" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 152 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 152 + ], + "type": "text", + "content": "even meshes without color [14,82,115]. CAD and other models are also commonly used by object pose estimation [4, 30, 34, 36, 44-46, 62, 90]. Image to LiDAR localization [7,26,43,108] is also relevant, especially in robotics applications." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 167, + 480, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 167, + 480, + 238 + ], + "spans": [ + { + "bbox": [ + 132, + 167, + 480, + 238 + ], + "type": "text", + "content": "Another variant of multi-modal data localization involves cross-view matching. This technique determines the camera position by finding correspondences between a ground-level query image and a two-dimensional bird's eye view map, such as a satellite image or a semantic landscape map [47,48,66,100,107,121,129]. Other cross-modal techniques were also proposed to involve semantics [28, 29] and event cameras [53] in image-based localization." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 254, + 480, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 254, + 480, + 313 + ], + "spans": [ + { + "bbox": [ + 132, + 254, + 480, + 313 + ], + "type": "text", + "content": "These approaches, while demonstrating promising localization results, are limited to interactions between two modalities - one for the query and one for the reference database. Our approach, in contrast, seeks the cross-modal coarse localization of a query image in a database composed of multiple modalities, e.g., 3D point clouds, images, semantics, object attributes, and relationships." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 327, + 480, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 327, + 480, + 459 + ], + "spans": [ + { + "bbox": [ + 132, + 327, + 480, + 459 + ], + "type": "text", + "content": "Scene Representation, encapsulating various scene attributes has evolved significantly, yielding diverse surface representations from explicit forms (3D point clouds [75,91], meshes [37], surfels [111]) to implicit ones (occupancy [61], signed distance functions [20,52]). The advent of neural representations has introduced novel means of encoding geometry [73,83,87,128], appearance [76,80], and semantics [74,79]. A comprehensive review is provided by Tewari et al. [114]. The integration of directions/rays [77] and visibility encoding in surface reconstruction [104] has further enriched this domain. Armeni et al. [3] introduced the 3D scene graph structure as a multi-layer representation of a scene that captures geometry, semantics, objects, and camera poses in a unified manner. Subsequent efforts [96,124] have further advanced 3D scene graph learning and structure." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 474, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 474, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 474, + 480, + 665 + ], + "type": "text", + "content": "The increasing interest in 3D scene graphs [3, 54, 60, 96, 124] underscores their potential as structured, rich descriptors for real-world scenes. Methods range from online incremental construction [49, 130] to offline generation from RGB-D imagery [3, 95, 124], and approaches for scene graph prediction [134, 136]. Their application spans embodied AI [94, 95, 106], task completion [1, 25], variability estimation [70], and SLAM [49, 95]. Recent studies like [131] introduce frameworks for localizing unseen objects by utilizing 3D scene graphs and graph neural networks for relation prediction, showcasing the utility of scene graphs in enhancing spatial understanding. Similarly, [97] offers new perspectives on 3D scene alignment, employing node matching within overlapping scene graphs to facilitate precise 3D map alignment. Kabalar et al. [54] assume that the query image has been coarsely localized and leverages a scene graph to identify dynamic objects and for precise localization with image features. Despite these significant advancements underscoring the value of scene graphs, their potential in multimodal localization remains largely untapped. In this paper, we use a scene graph representation of the map in which we aim to localize a query image." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "Y. Miao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 400, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 400, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 400, + 129 + ], + "type": "text", + "content": "3 Visual Localization with 3D Scene Graph" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 272 + ], + "type": "text", + "content": "Problem Statement. Let us assume that we are provided with a pre-constructed map of the environment, denoted as " + }, + { + "bbox": [ + 130, + 140, + 482, + 272 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 130, + 140, + 482, + 272 + ], + "type": "text", + "content": ", which is represented as a set of " + }, + { + "bbox": [ + 130, + 140, + 482, + 272 + ], + "type": "inline_equation", + "content": "N \\in \\mathbb{N}^+" + }, + { + "bbox": [ + 130, + 140, + 482, + 272 + ], + "type": "text", + "content": " 3D scene graphs " + }, + { + "bbox": [ + 130, + 140, + 482, + 272 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_i = (\\mathcal{V}_i, \\mathcal{E}_i)" + }, + { + "bbox": [ + 130, + 140, + 482, + 272 + ], + "type": "text", + "content": ", such that " + }, + { + "bbox": [ + 130, + 140, + 482, + 272 + ], + "type": "inline_equation", + "content": "\\mathcal{G} = \\{\\mathcal{G}_i\\}_{i \\in [0, N)}" + }, + { + "bbox": [ + 130, + 140, + 482, + 272 + ], + "type": "text", + "content": ". Having separate graphs " + }, + { + "bbox": [ + 130, + 140, + 482, + 272 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_i" + }, + { + "bbox": [ + 130, + 140, + 482, + 272 + ], + "type": "text", + "content": " is analogous to the hierarchy levels presented in [49, 95], each representing a group of object instances constituting a place like a room or building. Moreover, this task can also be straightforwardly redefined as subgraph selection. Vertices " + }, + { + "bbox": [ + 130, + 140, + 482, + 272 + ], + "type": "inline_equation", + "content": "v \\in \\mathcal{V}_i" + }, + { + "bbox": [ + 130, + 140, + 482, + 272 + ], + "type": "text", + "content": " symbolize instances of objects (e.g., chairs, tables) and large instances of semantic categories (e.g., walls, ground) within the scene, " + }, + { + "bbox": [ + 130, + 140, + 482, + 272 + ], + "type": "inline_equation", + "content": "i \\in [0, N)" + }, + { + "bbox": [ + 130, + 140, + 482, + 272 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 130, + 140, + 482, + 272 + ], + "type": "inline_equation", + "content": "\\mathcal{V} = \\bigcup_{i \\in [0, N)} \\mathcal{V}_i" + }, + { + "bbox": [ + 130, + 140, + 482, + 272 + ], + "type": "text", + "content": " aggregate all objects across the scenes. Edges " + }, + { + "bbox": [ + 130, + 140, + 482, + 272 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_i = \\{(v_j^i, v_k^i) | v_j^i, v_k^i \\in \\mathcal{V}_i\\}" + }, + { + "bbox": [ + 130, + 140, + 482, + 272 + ], + "type": "text", + "content": " delineate the relationships between objects, such as \"nearby\", \"standing on\", and \"attached to\"." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 274, + 489, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 274, + 489, + 332 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 489, + 332 + ], + "type": "text", + "content": "For each vertex " + }, + { + "bbox": [ + 130, + 274, + 489, + 332 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 130, + 274, + 489, + 332 + ], + "type": "text", + "content": ", we introduce " + }, + { + "bbox": [ + 130, + 274, + 489, + 332 + ], + "type": "inline_equation", + "content": "M \\in \\mathbb{N}^{+}" + }, + { + "bbox": [ + 130, + 274, + 489, + 332 + ], + "type": "text", + "content": " map modalities " + }, + { + "bbox": [ + 130, + 274, + 489, + 332 + ], + "type": "inline_equation", + "content": "f_{j}: \\mathcal{V} \\to \\mathcal{M}_{j}" + }, + { + "bbox": [ + 130, + 274, + 489, + 332 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 130, + 274, + 489, + 332 + ], + "type": "inline_equation", + "content": "j \\in [0, M)" + }, + { + "bbox": [ + 130, + 274, + 489, + 332 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 274, + 489, + 332 + ], + "type": "inline_equation", + "content": "f_{j}" + }, + { + "bbox": [ + 130, + 274, + 489, + 332 + ], + "type": "text", + "content": " maps vertex " + }, + { + "bbox": [ + 130, + 274, + 489, + 332 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 130, + 274, + 489, + 332 + ], + "type": "text", + "content": " to the " + }, + { + "bbox": [ + 130, + 274, + 489, + 332 + ], + "type": "inline_equation", + "content": "j^{th}" + }, + { + "bbox": [ + 130, + 274, + 489, + 332 + ], + "type": "text", + "content": " modality " + }, + { + "bbox": [ + 130, + 274, + 489, + 332 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{j}" + }, + { + "bbox": [ + 130, + 274, + 489, + 332 + ], + "type": "text", + "content": " that may be " + }, + { + "bbox": [ + 130, + 274, + 489, + 332 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{j} \\in \\{\\text{position}, \\text{orientation}, \\text{pointcloud}, \\text{semantic category}, \\text{image}, \\text{attribute}\\}" + }, + { + "bbox": [ + 130, + 274, + 489, + 332 + ], + "type": "text", + "content": ". While this paper focuses on these modalities, the set is easily extendable by incorporating additional ones, such as textual description or floor plan." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 333, + 482, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 333, + 482, + 381 + ], + "spans": [ + { + "bbox": [ + 130, + 333, + 482, + 381 + ], + "type": "text", + "content": "Given " + }, + { + "bbox": [ + 130, + 333, + 482, + 381 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 130, + 333, + 482, + 381 + ], + "type": "text", + "content": " and an input query image " + }, + { + "bbox": [ + 130, + 333, + 482, + 381 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 333, + 482, + 381 + ], + "type": "text", + "content": ", the objective is to identify the scene graph " + }, + { + "bbox": [ + 130, + 333, + 482, + 381 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_i" + }, + { + "bbox": [ + 130, + 333, + 482, + 381 + ], + "type": "text", + "content": " corresponding to the space depicted in image " + }, + { + "bbox": [ + 130, + 333, + 482, + 381 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 333, + 482, + 381 + ], + "type": "text", + "content": ". Note that while we focus on having an input image in this paper, the method can be modified to other modalities as well, e.g., depth image. Formally, we aim to resolve problem:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 233, + 392, + 480, + 413 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 392, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 233, + 392, + 480, + 413 + ], + "type": "interline_equation", + "content": "\\mathcal {G} _ {i ^ {*}} = \\underset {i \\in [ 0, N)} {\\arg \\max } \\llbracket \\text {c o n t a i n s} (\\mathcal {G} _ {i}, \\mathbf {p} _ {I}) \\rrbracket , \\tag {1}", + "image_path": "8ad9def15cf544a78e3ed9c9098a2eee65ea229b0d9ff6038b7f6d2f9fff2096.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 418, + 482, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 482, + 466 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 482, + 466 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 418, + 482, + 466 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_I \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 418, + 482, + 466 + ], + "type": "text", + "content": " is the unknown 3D position of the image, and " + }, + { + "bbox": [ + 130, + 418, + 482, + 466 + ], + "type": "inline_equation", + "content": "[.]" + }, + { + "bbox": [ + 130, + 418, + 482, + 466 + ], + "type": "text", + "content": " is the Iverson bracket which equals to 1 if the condition inside holds and 0 otherwise. It is important to note that our objective is coarse localization, opting for the selection of " + }, + { + "bbox": [ + 130, + 418, + 482, + 466 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_i" + }, + { + "bbox": [ + 130, + 418, + 482, + 466 + ], + "type": "text", + "content": " without needing precise estimation of " + }, + { + "bbox": [ + 130, + 418, + 482, + 466 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_I" + }, + { + "bbox": [ + 130, + 418, + 482, + 466 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 467, + 482, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 467, + 482, + 515 + ], + "spans": [ + { + "bbox": [ + 130, + 467, + 482, + 515 + ], + "type": "text", + "content": "To this end, the optimization problem can be reformulated to incorporate the chirality constraint, asserting that if an object is visible in image " + }, + { + "bbox": [ + 130, + 467, + 482, + 515 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 467, + 482, + 515 + ], + "type": "text", + "content": ", it must be positioned in front of the camera in 3D space, unobstructed by any entities (e.g., walls). Consequently, the problem becomes:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 151, + 524, + 480, + 550 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 524, + 480, + 550 + ], + "spans": [ + { + "bbox": [ + 151, + 524, + 480, + 550 + ], + "type": "interline_equation", + "content": "\\mathcal {G} _ {i ^ {*}} = \\underset {i \\in [ 0, N)} {\\arg \\max } \\sum_ {o _ {I} \\in \\mathcal {O} _ {I}} \\llbracket \\text {v i s i b l e} (\\mathcal {G} _ {i}, o _ {I}) \\rrbracket \\approx \\underset {i \\in [ 0, N)} {\\arg \\max } \\sum_ {o _ {I} \\in \\mathcal {O} _ {I}} \\log P (o _ {I} \\mid \\mathcal {G} _ {i}), \\tag {2}", + "image_path": "01e640a69fb1a71d92c914463f785df8ea299ce1916cd179c0812ed2204dc622.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 557, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 594 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 557, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_I" + }, + { + "bbox": [ + 130, + 557, + 482, + 594 + ], + "type": "text", + "content": " is the object set visible in image " + }, + { + "bbox": [ + 130, + 557, + 482, + 594 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 557, + 482, + 594 + ], + "type": "text", + "content": ", object " + }, + { + "bbox": [ + 130, + 557, + 482, + 594 + ], + "type": "inline_equation", + "content": "o_I \\subseteq \\{(x,y) \\in I\\}" + }, + { + "bbox": [ + 130, + 557, + 482, + 594 + ], + "type": "text", + "content": " is a set of pixels in the image " + }, + { + "bbox": [ + 130, + 557, + 482, + 594 + ], + "type": "inline_equation", + "content": "(o_I \\in \\mathcal{O}_I)" + }, + { + "bbox": [ + 130, + 557, + 482, + 594 + ], + "type": "text", + "content": ", visible: " + }, + { + "bbox": [ + 130, + 557, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_i \\times \\mathcal{O}_I \\to \\{0,1\\}" + }, + { + "bbox": [ + 130, + 557, + 482, + 594 + ], + "type": "text", + "content": " indicates if an object is visible in a scene graph, " + }, + { + "bbox": [ + 130, + 557, + 482, + 594 + ], + "type": "inline_equation", + "content": "P(o_I \\mid \\mathcal{G}_i)" + }, + { + "bbox": [ + 130, + 557, + 482, + 594 + ], + "type": "text", + "content": " is the probability of " + }, + { + "bbox": [ + 130, + 557, + 482, + 594 + ], + "type": "inline_equation", + "content": "o_I" + }, + { + "bbox": [ + 130, + 557, + 482, + 594 + ], + "type": "text", + "content": " stemming from " + }, + { + "bbox": [ + 130, + 557, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_i" + }, + { + "bbox": [ + 130, + 557, + 482, + 594 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": "Function visible can be approached as an indicator of whether object " + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "inline_equation", + "content": "o_{I}" + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": " appears in graph " + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "inline_equation", + "content": "G_{i}" + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": ". It holds if and only if there exists a " + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "inline_equation", + "content": "v \\in \\mathcal{V}_{i}" + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": " represents the same object as " + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "inline_equation", + "content": "o_{I}" + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": ". Therefore, this problem can be approached as matching a set of image pixels " + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "inline_equation", + "content": "(o_{I})" + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": " to a scene graph node " + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "inline_equation", + "content": "(v)" + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": " that comprises a set of modalities. In the next sections, we will describe a way to learn a unified embedding space for both " + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "inline_equation", + "content": "o_{I}" + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": " such that they become matchable." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "type": "text", + "content": "SceneGraphLoc" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "Proposed Pipeline is shown in Fig. 2. It consists of two concurrent stages: the first one generates object embeddings " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "e_{q} \\in \\mathbb{R}^{D}" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " from patches " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "q \\in \\mathcal{Q}_I" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " within the query image " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": ", and the second derives node embeddings " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "e_{v} \\in \\mathbb{R}^{D}" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " for nodes " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "v \\in \\mathcal{V}_i" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " in the scene graph " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_i" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": ". The training objective is to make " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "\\delta(e_{q}, e_{v}) = 0" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " if and only if the object associated with node " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " is directly visible (i.e., neither occluded nor outside the camera frustum) through the image patch " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "\\delta : \\mathbb{R}^D \\times \\mathbb{R}^D \\to \\mathbb{R}" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " denotes inverse cosine similarity normalized to [0,1]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 200, + 481, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 200, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 200, + 481, + 224 + ], + "type": "text", + "content": "After generating " + }, + { + "bbox": [ + 132, + 200, + 481, + 224 + ], + "type": "inline_equation", + "content": "e_{q}" + }, + { + "bbox": [ + 132, + 200, + 481, + 224 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 200, + 481, + 224 + ], + "type": "inline_equation", + "content": "e_{v}" + }, + { + "bbox": [ + 132, + 200, + 481, + 224 + ], + "type": "text", + "content": ", the model performs nearest neighbor matching (NN) for each patch in each scene graph " + }, + { + "bbox": [ + 132, + 200, + 481, + 224 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{i}" + }, + { + "bbox": [ + 132, + 200, + 481, + 224 + ], + "type": "text", + "content": ", assigning node " + }, + { + "bbox": [ + 132, + 200, + 481, + 224 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 132, + 200, + 481, + 224 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 132, + 200, + 481, + 224 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 132, + 200, + 481, + 224 + ], + "type": "text", + "content": " such that" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 240, + 228, + 481, + 249 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 240, + 228, + 481, + 249 + ], + "spans": [ + { + "bbox": [ + 240, + 228, + 481, + 249 + ], + "type": "interline_equation", + "content": "\\mathrm {N N} (q, \\mathcal {V} _ {i}) = \\underset {v \\in \\mathcal {V} _ {i}} {\\arg \\min } \\delta \\left(e _ {q}, e _ {v}\\right). \\tag {3}", + "image_path": "063682d495d9714c5354659ad68b0253e282acfc48e77b161fdaeeb188d4c73c.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 252, + 482, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 252, + 482, + 313 + ], + "spans": [ + { + "bbox": [ + 130, + 252, + 482, + 313 + ], + "type": "text", + "content": "Through matching, we establish patch-to-node correspondences " + }, + { + "bbox": [ + 130, + 252, + 482, + 313 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_i = \\{(q,v) \\mid v = \\mathrm{NN}(q,\\mathcal{V}_i) \\in \\mathcal{V}_i, q \\in \\mathcal{Q}_I\\}" + }, + { + "bbox": [ + 130, + 252, + 482, + 313 + ], + "type": "text", + "content": ". Based on " + }, + { + "bbox": [ + 130, + 252, + 482, + 313 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_i" + }, + { + "bbox": [ + 130, + 252, + 482, + 313 + ], + "type": "text", + "content": ", we devise an image-to-graph similarity score enabling us to deduce whether image " + }, + { + "bbox": [ + 130, + 252, + 482, + 313 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 252, + 482, + 313 + ], + "type": "text", + "content": " corresponds to the space represented by " + }, + { + "bbox": [ + 130, + 252, + 482, + 313 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_i" + }, + { + "bbox": [ + 130, + 252, + 482, + 313 + ], + "type": "text", + "content": ". Finally, potential coarse locations of " + }, + { + "bbox": [ + 130, + 252, + 482, + 313 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 252, + 482, + 313 + ], + "type": "text", + "content": " are selected by maximizing the similarity score across the stored scene graphs as depicted in Eq 2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 327, + 361, + 339 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 327, + 361, + 339 + ], + "spans": [ + { + "bbox": [ + 131, + 327, + 361, + 339 + ], + "type": "text", + "content": "3.1 Object Embeddings in the Scene Graph" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 342, + 350, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 350, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 350, + 437 + ], + "type": "text", + "content": "This section aims to obtain an embedding for each node within graph " + }, + { + "bbox": [ + 130, + 342, + 350, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_i" + }, + { + "bbox": [ + 130, + 342, + 350, + 437 + ], + "type": "text", + "content": ", encapsulating information from all available modalities. Our method builds upon the method of Sarkar et al. [97], with enhancements to include the image modality and to distill a unified embedding from all modalities rather than merely concatenating separate embeddings as in [97]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 438, + 350, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 350, + 474 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 350, + 474 + ], + "type": "text", + "content": "Scene graphs are conceptualized as multimodal knowledge graphs, similar to those used in entity alignment, treating semantic and geometric" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 474, + 481, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 481, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 481, + 582 + ], + "type": "text", + "content": "information as distinct modalities. The objective is to learn a joint multi-modal embedding from the individual modal encodings (uni-modal), ensuring nodes corresponding to the same object instance across different graphs are closely positioned. This involves the creation of uni-modal embeddings for the three primary types of 3D scene graph information: object embeddings encoding nodes in " + }, + { + "bbox": [ + 130, + 474, + 481, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 130, + 474, + 481, + 582 + ], + "type": "text", + "content": ", structure embeddings " + }, + { + "bbox": [ + 130, + 474, + 481, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 130, + 474, + 481, + 582 + ], + "type": "text", + "content": " representing edges in " + }, + { + "bbox": [ + 130, + 474, + 481, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 130, + 474, + 481, + 582 + ], + "type": "text", + "content": " as a structured graph, and two meta modalities encoding attributes " + }, + { + "bbox": [ + 130, + 474, + 481, + 582 + ], + "type": "inline_equation", + "content": "(\\mathcal{A})" + }, + { + "bbox": [ + 130, + 474, + 481, + 582 + ], + "type": "text", + "content": " and relationships " + }, + { + "bbox": [ + 130, + 474, + 481, + 582 + ], + "type": "inline_equation", + "content": "(\\mathcal{R})" + }, + { + "bbox": [ + 130, + 474, + 481, + 582 + ], + "type": "text", + "content": " between objects as one-hot vectors. These uni-modal embeddings are then combined in a weighted manner and optimized jointly through knowledge distillation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 582, + 481, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 481, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 481, + 617 + ], + "type": "text", + "content": "Each of these modalities is processed separately to generate uni-modal embeddings, which are subsequently integrated to model complex inter-modal interactions within the joint embedding space." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "Object Embedding. Node " + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "v \\in \\mathcal{V}" + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": " may contain multiple modalities, such as point cloud " + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "(\\mathcal{P})" + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": " and image " + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "(\\mathcal{I})" + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": ". Point clouds contain rich geometric information about objects. The point cloud corresponding to each " + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "v \\in \\mathcal{V}" + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": " is inputted to the object encoder. We employ the PointNet architecture [91] as the" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 359, + 347, + 480, + 416 + ], + "blocks": [ + { + "bbox": [ + 359, + 347, + 480, + 416 + ], + "lines": [ + { + "bbox": [ + 359, + 347, + 480, + 416 + ], + "spans": [ + { + "bbox": [ + 359, + 347, + 480, + 416 + ], + "type": "image", + "image_path": "7467a870839053004005c7f8889d65635d49e7aa88ab50b8c2311bb06ce46386.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 356, + 420, + 482, + 464 + ], + "lines": [ + { + "bbox": [ + 356, + 420, + 482, + 464 + ], + "spans": [ + { + "bbox": [ + 356, + 420, + 482, + 464 + ], + "type": "text", + "content": "Fig.3: The embedding of image modality " + }, + { + "bbox": [ + 356, + 420, + 482, + 464 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 356, + 420, + 482, + 464 + ], + "type": "text", + "content": " for each object. The image crops of a pillow are shown as an example." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "type": "text", + "content": "Y. Miao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 248 + ], + "type": "text", + "content": "object encoder to extract the geometric feature " + }, + { + "bbox": [ + 130, + 115, + 482, + 248 + ], + "type": "inline_equation", + "content": "e_v^{\\mathcal{P}}" + }, + { + "bbox": [ + 130, + 115, + 482, + 248 + ], + "type": "text", + "content": " for every node. Furthermore, we integrate multi-level and multi-view visual embeddings to enrich the graph encoder with a more nuanced understanding of image information. The visual embedding pipeline is visualized in Fig. 3. For each node " + }, + { + "bbox": [ + 130, + 115, + 482, + 248 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 130, + 115, + 482, + 248 + ], + "type": "text", + "content": " denoting a 3D object, the top " + }, + { + "bbox": [ + 130, + 115, + 482, + 248 + ], + "type": "inline_equation", + "content": "K_{view} \\in \\mathbb{N}^{+}" + }, + { + "bbox": [ + 130, + 115, + 482, + 248 + ], + "type": "text", + "content": "images with largest visibility of " + }, + { + "bbox": [ + 130, + 115, + 482, + 248 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 130, + 115, + 482, + 248 + ], + "type": "text", + "content": " is selected: " + }, + { + "bbox": [ + 130, + 115, + 482, + 248 + ], + "type": "inline_equation", + "content": "\\{I_{\\mathrm{db},k}^v \\mid k = [0,K)\\} \\subseteq \\mathcal{I}_{\\mathrm{db}}" + }, + { + "bbox": [ + 130, + 115, + 482, + 248 + ], + "type": "text", + "content": " from the database " + }, + { + "bbox": [ + 130, + 115, + 482, + 248 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\mathrm{db}}" + }, + { + "bbox": [ + 130, + 115, + 482, + 248 + ], + "type": "text", + "content": ". We can define an ordering over these images such that " + }, + { + "bbox": [ + 130, + 115, + 482, + 248 + ], + "type": "inline_equation", + "content": "\\phi(I_{\\mathrm{db},0}^v,v) \\geq \\phi(I_{\\mathrm{db},1}^v,v) \\geq \\ldots \\geq \\phi(I_{\\mathrm{db},K_{view}-1}^v,v)" + }, + { + "bbox": [ + 130, + 115, + 482, + 248 + ], + "type": "text", + "content": ", with the visibility function " + }, + { + "bbox": [ + 130, + 115, + 482, + 248 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 115, + 482, + 248 + ], + "type": "text", + "content": " quantifying the extent of node " + }, + { + "bbox": [ + 130, + 115, + 482, + 248 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 130, + 115, + 482, + 248 + ], + "type": "text", + "content": " observed in each image through pixel count. Visibility check is implemented by projecting the 3D model of " + }, + { + "bbox": [ + 130, + 115, + 482, + 248 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 130, + 115, + 482, + 248 + ], + "type": "text", + "content": " to each posed image in " + }, + { + "bbox": [ + 130, + 115, + 482, + 248 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\mathrm{db}}" + }, + { + "bbox": [ + 130, + 115, + 482, + 248 + ], + "type": "text", + "content": ", which are usually available when constructing the scene graph [49, 130]. Parameter " + }, + { + "bbox": [ + 130, + 115, + 482, + 248 + ], + "type": "inline_equation", + "content": "K_{view} = 10" + }, + { + "bbox": [ + 130, + 115, + 482, + 248 + ], + "type": "text", + "content": " in all our experiments." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 253, + 482, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 253, + 482, + 421 + ], + "spans": [ + { + "bbox": [ + 130, + 253, + 482, + 421 + ], + "type": "text", + "content": "Drawing inspiration from OpenMask3D [113], for any given view " + }, + { + "bbox": [ + 130, + 253, + 482, + 421 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{db}}^v" + }, + { + "bbox": [ + 130, + 253, + 482, + 421 + ], + "type": "text", + "content": " of object " + }, + { + "bbox": [ + 130, + 253, + 482, + 421 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 130, + 253, + 482, + 421 + ], + "type": "text", + "content": ", initial steps include calculating the bounding box " + }, + { + "bbox": [ + 130, + 253, + 482, + 421 + ], + "type": "inline_equation", + "content": "b_{v,0}" + }, + { + "bbox": [ + 130, + 253, + 482, + 421 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 130, + 253, + 482, + 421 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 130, + 253, + 482, + 421 + ], + "type": "text", + "content": " within the image, followed by the generation of multi-level bounding boxes " + }, + { + "bbox": [ + 130, + 253, + 482, + 421 + ], + "type": "inline_equation", + "content": "\\{b_{v,l} \\mid l \\in [0,L)\\}" + }, + { + "bbox": [ + 130, + 253, + 482, + 421 + ], + "type": "text", + "content": " through iterative enlargement of " + }, + { + "bbox": [ + 130, + 253, + 482, + 421 + ], + "type": "inline_equation", + "content": "b_{v,0}" + }, + { + "bbox": [ + 130, + 253, + 482, + 421 + ], + "type": "text", + "content": ". This enlargement strategy aims to capture contextual information around object " + }, + { + "bbox": [ + 130, + 253, + 482, + 421 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 130, + 253, + 482, + 421 + ], + "type": "text", + "content": ". Subsequently, Dino V2 [81] processes the image crops defined by " + }, + { + "bbox": [ + 130, + 253, + 482, + 421 + ], + "type": "inline_equation", + "content": "b_{v,l}" + }, + { + "bbox": [ + 130, + 253, + 482, + 421 + ], + "type": "text", + "content": ", extracting multi-level features " + }, + { + "bbox": [ + 130, + 253, + 482, + 421 + ], + "type": "inline_equation", + "content": "\\{f_l \\mid l \\in [0,L)\\}" + }, + { + "bbox": [ + 130, + 253, + 482, + 421 + ], + "type": "text", + "content": ". For each image " + }, + { + "bbox": [ + 130, + 253, + 482, + 421 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{db},k}^v" + }, + { + "bbox": [ + 130, + 253, + 482, + 421 + ], + "type": "text", + "content": ", an average pooling operation aggregates these multi-level crop features into a singular feature vector " + }, + { + "bbox": [ + 130, + 253, + 482, + 421 + ], + "type": "inline_equation", + "content": "f = \\operatorname{avg\\_pool}\\{f_l \\mid l \\in [0,L)\\}" + }, + { + "bbox": [ + 130, + 253, + 482, + 421 + ], + "type": "text", + "content": ". The final step of this process involves the application of a Transformer encoder, which incorporates image poses as positional encodings. This step synthesizes multiview object tokens into a cohesive visual embedding " + }, + { + "bbox": [ + 130, + 253, + 482, + 421 + ], + "type": "inline_equation", + "content": "e_v^{\\mathcal{I}}" + }, + { + "bbox": [ + 130, + 253, + 482, + 421 + ], + "type": "text", + "content": ", effectively integrating the diverse perspectives and levels of contextual information pertaining to each object within the database. Please note that the image database does not necessarily have to be stored after distilling the object embeddings." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 426, + 482, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 547 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 547 + ], + "type": "text", + "content": "Structure Embedding. 3D Scene Graphs encapsulate the object relationships, which we exploit to encode their spatial configuration. This relational data is represented through a structure graph, where node features embody the relative translations between object instances, and edges denote these relationships. The relative translation is determined by calculating the distance from the object instance to any other object in the scene. To encapsulate this structural information within " + }, + { + "bbox": [ + 130, + 426, + 482, + 547 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_i" + }, + { + "bbox": [ + 130, + 426, + 482, + 547 + ], + "type": "text", + "content": ", a Graph Attention Network (GAT) [119] is utilized, with the weight matrix constrained to a diagonal form to reduce computational demands and enhance model scalability. Following the method in [97], the structural embedding " + }, + { + "bbox": [ + 130, + 426, + 482, + 547 + ], + "type": "inline_equation", + "content": "e_v^S" + }, + { + "bbox": [ + 130, + 426, + 482, + 547 + ], + "type": "text", + "content": " is derived from the final layer of a two-layer GAT model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 552, + 482, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 552, + 482, + 624 + ], + "spans": [ + { + "bbox": [ + 130, + 552, + 482, + 624 + ], + "type": "text", + "content": "Meta Embeddings. In addition to geometry and structure, the object attributes and the inter-object relationships are captured in two distinct embeddings, " + }, + { + "bbox": [ + 130, + 552, + 482, + 624 + ], + "type": "inline_equation", + "content": "e_v^{\\mathcal{R}}" + }, + { + "bbox": [ + 130, + 552, + 482, + 624 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 552, + 482, + 624 + ], + "type": "inline_equation", + "content": "e_v^{\\mathcal{A}}" + }, + { + "bbox": [ + 130, + 552, + 482, + 624 + ], + "type": "text", + "content": ". The relationships an object maintains with others are conceptualized as a bag-of-words [19] feature vector, which is input to a feed-forward neural network, distilled in relational embedding " + }, + { + "bbox": [ + 130, + 552, + 482, + 624 + ], + "type": "inline_equation", + "content": "e_v^{\\mathcal{R}}" + }, + { + "bbox": [ + 130, + 552, + 482, + 624 + ], + "type": "text", + "content": ". A similar approach is employed for the attributes associated with the objects, producing embedding " + }, + { + "bbox": [ + 130, + 552, + 482, + 624 + ], + "type": "inline_equation", + "content": "e_v^{\\mathcal{A}}" + }, + { + "bbox": [ + 130, + 552, + 482, + 624 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": "Joint Embedding. Similar to [97], we concatenate uni-modal features to a compact representation for each object " + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": ". Contrary to [97], we encapsulate it in a multi-layer perceptron (MLP) to learn an embedding with size independent of" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "type": "text", + "content": "SceneGraphLoc" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 455, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 455, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 455, + 129 + ], + "type": "text", + "content": "the available modalities, fusing information from all. Embedding " + }, + { + "bbox": [ + 132, + 116, + 455, + 129 + ], + "type": "inline_equation", + "content": "e_v" + }, + { + "bbox": [ + 132, + 116, + 455, + 129 + ], + "type": "text", + "content": " is as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 214, + 136, + 481, + 170 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 136, + 481, + 170 + ], + "spans": [ + { + "bbox": [ + 214, + 136, + 481, + 170 + ], + "type": "interline_equation", + "content": "e _ {v} = \\operatorname {M L P} \\left(\\oplus_ {k \\in \\mathcal {K}} \\left[ \\frac {\\exp (w _ {k})}{\\sum_ {j \\in \\mathcal {K}} \\exp (w _ {j})} e _ {v} ^ {k} \\right]\\right), \\tag {4}", + "image_path": "f60084505300b226d710225798106306156ab72a334b2f6c43bdfb090c9f6019.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 177, + 479, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 177, + 479, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 177, + 479, + 224 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 177, + 479, + 224 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 130, + 177, + 479, + 224 + ], + "type": "text", + "content": " is the concatenation operator, " + }, + { + "bbox": [ + 130, + 177, + 479, + 224 + ], + "type": "inline_equation", + "content": "\\mathcal{K} = \\{\\mathcal{P},\\mathcal{I},\\mathcal{S},\\mathcal{R},\\mathcal{A}\\}" + }, + { + "bbox": [ + 130, + 177, + 479, + 224 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 177, + 479, + 224 + ], + "type": "inline_equation", + "content": "w_{m}" + }, + { + "bbox": [ + 130, + 177, + 479, + 224 + ], + "type": "text", + "content": " is a trainable attention weight for each modality " + }, + { + "bbox": [ + 130, + 177, + 479, + 224 + ], + "type": "inline_equation", + "content": "k\\in \\mathcal{K}" + }, + { + "bbox": [ + 130, + 177, + 479, + 224 + ], + "type": "text", + "content": ". A two-layer MLP is applied to map the dimension of the concatenated multi-modal descriptors from " + }, + { + "bbox": [ + 130, + 177, + 479, + 224 + ], + "type": "inline_equation", + "content": "D^{\\mathcal{K}}" + }, + { + "bbox": [ + 130, + 177, + 479, + 224 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 177, + 479, + 224 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 177, + 479, + 224 + ], + "type": "text", + "content": ". We apply " + }, + { + "bbox": [ + 130, + 177, + 479, + 224 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 130, + 177, + 479, + 224 + ], + "type": "text", + "content": " normalization to each uni-modal feature before concatenation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 225, + 480, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 225, + 480, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 225, + 480, + 285 + ], + "type": "text", + "content": "In practice, these independent modalities are only required and used in the mapping phase of the procedure. This stage involves the construction of an environmental map in the form of a 3D scene graph, during which each node " + }, + { + "bbox": [ + 130, + 225, + 480, + 285 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 130, + 225, + 480, + 285 + ], + "type": "text", + "content": " is distilled into an embedding " + }, + { + "bbox": [ + 130, + 225, + 480, + 285 + ], + "type": "inline_equation", + "content": "e_v" + }, + { + "bbox": [ + 130, + 225, + 480, + 285 + ], + "type": "text", + "content": ". During localization, we can ignore independent modalities and use only the fixed-sized embeddings " + }, + { + "bbox": [ + 130, + 225, + 480, + 285 + ], + "type": "inline_equation", + "content": "e_v" + }, + { + "bbox": [ + 130, + 225, + 480, + 285 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 303, + 362, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 303, + 362, + 316 + ], + "spans": [ + { + "bbox": [ + 132, + 303, + 362, + 316 + ], + "type": "text", + "content": "3.2 Object Embeddings in the Query Image" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 323, + 482, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 323, + 482, + 490 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 482, + 490 + ], + "type": "text", + "content": "To solve the optimization problem in Eq. 2, we need to find object instances " + }, + { + "bbox": [ + 130, + 323, + 482, + 490 + ], + "type": "inline_equation", + "content": "o_{I} \\in \\mathcal{O}_{I}" + }, + { + "bbox": [ + 130, + 323, + 482, + 490 + ], + "type": "text", + "content": " in query image " + }, + { + "bbox": [ + 130, + 323, + 482, + 490 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 323, + 482, + 490 + ], + "type": "text", + "content": ". A straightforward approach to do so would be to apply a 2D panoptic segmentation algorithm, e.g., Mask2Former [18]. However, we noticed in our experiments that such an approach is susceptible to inaccuracies and failures (over- and under-segmentation) in the segmentation, severely affecting the accuracy. Thus, we approach the problem by a visual Transformer (ViT), breaking up the image into rectangular patches " + }, + { + "bbox": [ + 130, + 323, + 482, + 490 + ], + "type": "inline_equation", + "content": "q \\in \\mathcal{Q}_I" + }, + { + "bbox": [ + 130, + 323, + 482, + 490 + ], + "type": "text", + "content": " and distilling an independent embedding for each patch " + }, + { + "bbox": [ + 130, + 323, + 482, + 490 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 130, + 323, + 482, + 490 + ], + "type": "text", + "content": " based on the object visible in " + }, + { + "bbox": [ + 130, + 323, + 482, + 490 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 130, + 323, + 482, + 490 + ], + "type": "text", + "content": ". We use Dino V2 [81] as a backbone to obtain patch-level features. These features are passed through a patch encoder trained to create embeddings from the patch features considering the objects visible from each " + }, + { + "bbox": [ + 130, + 323, + 482, + 490 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 130, + 323, + 482, + 490 + ], + "type": "text", + "content": ". For this encoder, we use a 4-layer convolution neural network (CNN) with residual blocks introduced in [41] on the Dino V2 features and a 3-layer MLP to further map the patch feature to dimension " + }, + { + "bbox": [ + 130, + 323, + 482, + 490 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 323, + 482, + 490 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 508, + 268, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 508, + 268, + 521 + ], + "spans": [ + { + "bbox": [ + 132, + 508, + 268, + 521 + ], + "type": "text", + "content": "3.3 Contrastive Learning" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 528, + 403, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 528, + 403, + 648 + ], + "spans": [ + { + "bbox": [ + 130, + 528, + 403, + 648 + ], + "type": "text", + "content": "We use contrastive learning to learn a joint embedding space for the scene graph nodes and image patches. To do so, we form query image and graph pairs " + }, + { + "bbox": [ + 130, + 528, + 403, + 648 + ], + "type": "inline_equation", + "content": "(I, \\mathcal{G}_I)" + }, + { + "bbox": [ + 130, + 528, + 403, + 648 + ], + "type": "text", + "content": ". Real-world scenes are rarely static, e.g., objects move or undergo non-rigid deformations and illumination changes [123]. To ensure that the learned embedding is robust to such temporal changes, we use scene graph " + }, + { + "bbox": [ + 130, + 528, + 403, + 648 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_I" + }, + { + "bbox": [ + 130, + 528, + 403, + 648 + ], + "type": "text", + "content": " from the same temporal point when " + }, + { + "bbox": [ + 130, + 528, + 403, + 648 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 528, + 403, + 648 + ], + "type": "text", + "content": " was captured, as positive samples, as well as a scene graph " + }, + { + "bbox": [ + 130, + 528, + 403, + 648 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_I^t" + }, + { + "bbox": [ + 130, + 528, + 403, + 648 + ], + "type": "text", + "content": " from another scan. Graph " + }, + { + "bbox": [ + 130, + 528, + 403, + 648 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_I^t" + }, + { + "bbox": [ + 130, + 528, + 403, + 648 + ], + "type": "text", + "content": " represents the same place as " + }, + { + "bbox": [ + 130, + 528, + 403, + 648 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_I" + }, + { + "bbox": [ + 130, + 528, + 403, + 648 + ], + "type": "text", + "content": ", but it undergoes temporal changes. An example is shown in Fig. 4." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 409, + 512, + 481, + 610 + ], + "blocks": [ + { + "bbox": [ + 409, + 512, + 481, + 610 + ], + "lines": [ + { + "bbox": [ + 409, + 512, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 409, + 512, + 481, + 610 + ], + "type": "image", + "image_path": "8d1a947c739d12d88cdf406f40339bac1ce79f7be3236c88c35980a1c617d118.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 408, + 613, + 482, + 647 + ], + "lines": [ + { + "bbox": [ + 408, + 613, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 408, + 613, + 482, + 647 + ], + "type": "text", + "content": "Fig.4: The same scene at different time steps " + }, + { + "bbox": [ + 408, + 613, + 482, + 647 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 408, + 613, + 482, + 647 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "type": "text", + "content": "Y. Miao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "For a query image " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": ", a set of candidate graphs " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{G}_I, \\mathcal{G}_I^t, \\mathcal{G}_1, \\dots, \\mathcal{G}_N\\}" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " is provided for training, where " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{G}_1, \\dots, \\mathcal{G}_N\\}" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " act as " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " negative samples, depicting different scenes than the target scene of the query image. We train our model by optimizing both a static loss and a temporal loss as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 235, + 171, + 480, + 184 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 171, + 480, + 184 + ], + "spans": [ + { + "bbox": [ + 235, + 171, + 480, + 184 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\alpha * \\mathcal {L} _ {\\text {s t a t i c}} + (1 - \\alpha) * \\mathcal {L} _ {\\text {t e m p}}. \\tag {5}", + "image_path": "913eb0326d1d28e5fd4d94a88277f339c893e569f9ac4b9d59b7a09704f01a66.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 190, + 479, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 190, + 479, + 300 + ], + "spans": [ + { + "bbox": [ + 131, + 190, + 479, + 300 + ], + "type": "text", + "content": "During training, we assume that image patch to graph node pairs are available [123] such that " + }, + { + "bbox": [ + 131, + 190, + 479, + 300 + ], + "type": "inline_equation", + "content": "P_{I} = \\{(q,v) \\mid q \\in \\mathcal{Q}_{I}, v \\in \\mathcal{G}_{I}\\}" + }, + { + "bbox": [ + 131, + 190, + 479, + 300 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 190, + 479, + 300 + ], + "type": "inline_equation", + "content": "P_{I}^{t} = \\{(q,v^{t}) \\mid q \\in \\mathcal{Q}_{I}, v \\in \\mathcal{G}_{I}^{t}\\}" + }, + { + "bbox": [ + 131, + 190, + 479, + 300 + ], + "type": "text", + "content": ". For each pair " + }, + { + "bbox": [ + 131, + 190, + 479, + 300 + ], + "type": "inline_equation", + "content": "(q,v)" + }, + { + "bbox": [ + 131, + 190, + 479, + 300 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 131, + 190, + 479, + 300 + ], + "type": "inline_equation", + "content": "P_{I}" + }, + { + "bbox": [ + 131, + 190, + 479, + 300 + ], + "type": "text", + "content": " and each pair from " + }, + { + "bbox": [ + 131, + 190, + 479, + 300 + ], + "type": "inline_equation", + "content": "P_{I}^{t}" + }, + { + "bbox": [ + 131, + 190, + 479, + 300 + ], + "type": "text", + "content": ", we use the following notation. Set " + }, + { + "bbox": [ + 131, + 190, + 479, + 300 + ], + "type": "inline_equation", + "content": "N_{q}^{\\mathcal{I}} = \\{q' \\mid q' \\in \\mathcal{Q}_{I}, v_{q'} \\neq v\\}" + }, + { + "bbox": [ + 131, + 190, + 479, + 300 + ], + "type": "text", + "content": " contains patches seeing objects other than " + }, + { + "bbox": [ + 131, + 190, + 479, + 300 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 131, + 190, + 479, + 300 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 131, + 190, + 479, + 300 + ], + "type": "inline_equation", + "content": "N_{v}^{\\mathcal{G}} = \\{v_{n} \\mid v_{n} \\in \\mathcal{V}_{I} \\cup \\mathcal{V}_{1} \\cup \\ldots \\cup \\mathcal{V}_{N}\\} \\setminus \\{v\\}" + }, + { + "bbox": [ + 131, + 190, + 479, + 300 + ], + "type": "text", + "content": " contains the 3D objects of all candidate scene graphs other than " + }, + { + "bbox": [ + 131, + 190, + 479, + 300 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 131, + 190, + 479, + 300 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 131, + 190, + 479, + 300 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{I}" + }, + { + "bbox": [ + 131, + 190, + 479, + 300 + ], + "type": "text", + "content": " represents the objects nodes of the target graph " + }, + { + "bbox": [ + 131, + 190, + 479, + 300 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{I}" + }, + { + "bbox": [ + 131, + 190, + 479, + 300 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 190, + 479, + 300 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{i}" + }, + { + "bbox": [ + 131, + 190, + 479, + 300 + ], + "type": "text", + "content": " is the nodes of other graphs " + }, + { + "bbox": [ + 131, + 190, + 479, + 300 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{i}" + }, + { + "bbox": [ + 131, + 190, + 479, + 300 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 131, + 190, + 479, + 300 + ], + "type": "inline_equation", + "content": "N_{v}^{\\mathcal{G}_{I}^{t}} = \\{v_{n}|v_{n} \\in \\mathcal{V}_{I}^{t} \\cup \\mathcal{V}_{1} \\cup \\ldots \\cup \\mathcal{V}_{n}\\} \\setminus \\{v\\}" + }, + { + "bbox": [ + 131, + 190, + 479, + 300 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 131, + 190, + 479, + 300 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{I}^{t}" + }, + { + "bbox": [ + 131, + 190, + 479, + 300 + ], + "type": "text", + "content": " represent the nodes of the target graph " + }, + { + "bbox": [ + 131, + 190, + 479, + 300 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{I}^{t}" + }, + { + "bbox": [ + 131, + 190, + 479, + 300 + ], + "type": "text", + "content": ". The static loss is defined as bi-directional N-pair loss [67] as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 140, + 305, + 479, + 342 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 305, + 479, + 342 + ], + "spans": [ + { + "bbox": [ + 140, + 305, + 479, + 342 + ], + "type": "interline_equation", + "content": "\\left. \\right. \\mathcal {L} _ {\\text {s t a t i c}} = E _ {P I \\in B} \\left[ E _ {(q, v) \\in P I} \\left[ - \\log \\left(\\frac {1}{2} p (q, v, N _ {q} ^ {\\mathcal {I}}, N _ {v} ^ {\\mathcal {G}}) + \\frac {1}{2} p (v, q, N _ {q} ^ {\\mathcal {I}}, N _ {v} ^ {\\mathcal {G}})\\right)\\right]\\right], \\tag {6}", + "image_path": "f7993bb3fd091bcba3228b73a04d00ab6de143791c54b76f99c585a4ab40d72c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 342, + 479, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 479, + 378 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 479, + 378 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 342, + 479, + 378 + ], + "type": "inline_equation", + "content": "E_{P_I\\in B}" + }, + { + "bbox": [ + 130, + 342, + 479, + 378 + ], + "type": "text", + "content": " represent loss averaging over a batch of query images and their corresponding candidate scene graphs; " + }, + { + "bbox": [ + 130, + 342, + 479, + 378 + ], + "type": "inline_equation", + "content": "p(q,v,N_q^\\mathcal{I},N_v^\\mathcal{G})" + }, + { + "bbox": [ + 130, + 342, + 479, + 378 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 342, + 479, + 378 + ], + "type": "inline_equation", + "content": "p(v,q,N_q^\\mathcal{I},N_v^\\mathcal{G})" + }, + { + "bbox": [ + 130, + 342, + 479, + 378 + ], + "type": "text", + "content": " represent the bi-directional probability distributions of the positive pair as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 153, + 384, + 459, + 412 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 384, + 459, + 412 + ], + "spans": [ + { + "bbox": [ + 153, + 384, + 459, + 412 + ], + "type": "interline_equation", + "content": "p (q, v, N _ {q} ^ {\\mathcal {I}}, N _ {v} ^ {\\mathcal {G}}) = \\frac {f (e _ {q} , e _ {v})}{f (e _ {q} , e _ {v}) + \\sum_ {q _ {n} \\in N _ {q} ^ {\\mathcal {I}}} f (e _ {q} , e _ {q _ {n}}) + \\sum_ {v _ {n} \\in N _ {v} ^ {\\mathcal {G}}} f (e _ {q} , e _ {v _ {n}})},", + "image_path": "a47de1dc5df66de841eb38a5351d8df3fe6eff9083e5192cd927512f13f60ebd.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 419, + 479, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 419, + 479, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 419, + 479, + 472 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 419, + 479, + 472 + ], + "type": "inline_equation", + "content": "f(e_q, e_v) = \\exp \\left( \\frac{-\\delta(e_q, e_v)}{\\tau} \\right)" + }, + { + "bbox": [ + 130, + 419, + 479, + 472 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 419, + 479, + 472 + ], + "type": "inline_equation", + "content": "\\delta(e_q, e_v)" + }, + { + "bbox": [ + 130, + 419, + 479, + 472 + ], + "type": "text", + "content": " represents the inverse cosine similarity between embeddings " + }, + { + "bbox": [ + 130, + 419, + 479, + 472 + ], + "type": "inline_equation", + "content": "e_q" + }, + { + "bbox": [ + 130, + 419, + 479, + 472 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 419, + 479, + 472 + ], + "type": "inline_equation", + "content": "e_v" + }, + { + "bbox": [ + 130, + 419, + 479, + 472 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 419, + 479, + 472 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 419, + 479, + 472 + ], + "type": "text", + "content": " is a temperature parameter. Probability distribution " + }, + { + "bbox": [ + 130, + 419, + 479, + 472 + ], + "type": "inline_equation", + "content": "p(v, q, N_q^T, N_v^G)" + }, + { + "bbox": [ + 130, + 419, + 479, + 472 + ], + "type": "text", + "content": " is written similarly. The temporal loss term is defined the same as Eq.6 but with " + }, + { + "bbox": [ + 130, + 419, + 479, + 472 + ], + "type": "inline_equation", + "content": "P_I^t" + }, + { + "bbox": [ + 130, + 419, + 479, + 472 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 419, + 479, + 472 + ], + "type": "inline_equation", + "content": "N_v^{G^t}" + }, + { + "bbox": [ + 130, + 419, + 479, + 472 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 137, + 479, + 479, + 515 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 479, + 479, + 515 + ], + "spans": [ + { + "bbox": [ + 137, + 479, + 479, + 515 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t e m p}} = E _ {P _ {I} ^ {t} \\in B} \\left[ E _ {(q, v) \\in P _ {I} ^ {t}} \\left[ - \\log \\left(\\frac {1}{2} p (q, v, N _ {q} ^ {\\mathcal {I}}, N _ {v} ^ {\\mathcal {G} ^ {t}}) + \\frac {1}{2} p (v, q, N _ {q} ^ {\\mathcal {I}}, N _ {v} ^ {\\mathcal {G} ^ {t}})\\right) \\right] \\right]. \\tag {7}", + "image_path": "da8e1bfb61eeb702103f181ecd8ea49b4be019f9338fb3bed0c09c34d7f264be.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 516, + 479, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 516, + 479, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 516, + 479, + 540 + ], + "type": "text", + "content": "By minimizing Eq.5, the paired cross-modal embeddings " + }, + { + "bbox": [ + 130, + 516, + 479, + 540 + ], + "type": "inline_equation", + "content": "e_{q}" + }, + { + "bbox": [ + 130, + 516, + 479, + 540 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 516, + 479, + 540 + ], + "type": "inline_equation", + "content": "e_{v}" + }, + { + "bbox": [ + 130, + 516, + 479, + 540 + ], + "type": "text", + "content": " are pulled together while the embeddings from different objects are pushed apart." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 555, + 274, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 555, + 274, + 567 + ], + "spans": [ + { + "bbox": [ + 131, + 555, + 274, + 567 + ], + "type": "text", + "content": "3.4 Scene Graph Retrieval" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 573, + 479, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 573, + 479, + 633 + ], + "spans": [ + { + "bbox": [ + 130, + 573, + 479, + 633 + ], + "type": "text", + "content": "Given a pre-established map of an environment represented through a collection of scene graphs " + }, + { + "bbox": [ + 130, + 573, + 479, + 633 + ], + "type": "inline_equation", + "content": "\\mathcal{G} = \\{\\mathcal{G}_i \\mid i \\in [0, N)\\}" + }, + { + "bbox": [ + 130, + 573, + 479, + 633 + ], + "type": "text", + "content": ", where each node embedding has been precomputed, the goal during inference is to identify the top-" + }, + { + "bbox": [ + 130, + 573, + 479, + 633 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 573, + 479, + 633 + ], + "type": "text", + "content": " scene graphs " + }, + { + "bbox": [ + 130, + 573, + 479, + 633 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_i" + }, + { + "bbox": [ + 130, + 573, + 479, + 633 + ], + "type": "text", + "content": " in which image " + }, + { + "bbox": [ + 130, + 573, + 479, + 633 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 573, + 479, + 633 + ], + "type": "text", + "content": " was likely captured. The method to address this challenge involves calculating the similarity between a graph and the image as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 222, + 639, + 479, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 639, + 479, + 667 + ], + "spans": [ + { + "bbox": [ + 222, + 639, + 479, + 667 + ], + "type": "interline_equation", + "content": "s \\left(\\mathcal {G} _ {i}, I\\right) = \\frac {1}{\\left| \\mathcal {Q} _ {I} \\right|} \\sum_ {q \\in \\mathcal {Q} _ {I}} \\left[ 1 - \\delta \\left(e _ {q}, \\operatorname {N N} (q)\\right) \\right], \\tag {8}", + "image_path": "e1ac4e8dfe527671d70576f1a8d21b813abedaf848ad1d5931dbb91dabe9f1ab.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "type": "text", + "content": "SceneGraphLoc" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 114, + 473, + 187 + ], + "blocks": [ + { + "bbox": [ + 143, + 114, + 473, + 187 + ], + "lines": [ + { + "bbox": [ + 143, + 114, + 473, + 187 + ], + "spans": [ + { + "bbox": [ + 143, + 114, + 473, + 187 + ], + "type": "image", + "image_path": "fdb92816cf9745f003cb305955fbc936562c082c727a27b71c1680fa28e2e8ce.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 194, + 482, + 239 + ], + "lines": [ + { + "bbox": [ + 130, + 194, + 482, + 239 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 482, + 239 + ], + "type": "text", + "content": "Fig.5: Qualitative Result of object-association-based scene retrieval from the 3RScan dataset [123]. The two left images show the ground truth (left) and predicted (right) patch-to-node associations of the query image. The right part of the figure illustrates the candidate scene graphs sorted by the image-graph similarity." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 247, + 482, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 331 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 247, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}_I" + }, + { + "bbox": [ + 130, + 247, + 482, + 331 + ], + "type": "text", + "content": " denotes the set of image patches in " + }, + { + "bbox": [ + 130, + 247, + 482, + 331 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 247, + 482, + 331 + ], + "type": "text", + "content": ", and for each patch " + }, + { + "bbox": [ + 130, + 247, + 482, + 331 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 130, + 247, + 482, + 331 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 247, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\mathrm{NN}(q, \\mathcal{V}_i) \\in \\mathcal{V}_i" + }, + { + "bbox": [ + 130, + 247, + 482, + 331 + ], + "type": "text", + "content": " represents the nearest node in terms of embedding similarity. The function " + }, + { + "bbox": [ + 130, + 247, + 482, + 331 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 247, + 482, + 331 + ], + "type": "text", + "content": ", which is assumed to map values to interval [0, 1], facilitating the identification of the optimal scene graph " + }, + { + "bbox": [ + 130, + 247, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{i^*}" + }, + { + "bbox": [ + 130, + 247, + 482, + 331 + ], + "type": "text", + "content": " that maximizes " + }, + { + "bbox": [ + 130, + 247, + 482, + 331 + ], + "type": "inline_equation", + "content": "s(\\mathcal{G}_i, I)" + }, + { + "bbox": [ + 130, + 247, + 482, + 331 + ], + "type": "text", + "content": ". Graph " + }, + { + "bbox": [ + 130, + 247, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{i^*}" + }, + { + "bbox": [ + 130, + 247, + 482, + 331 + ], + "type": "text", + "content": " is identified by iterating through all scene graphs and selecting the one with the highest similarity. This approach can be accelerated by using spatial partitioning techniques, such as kd-trees, to preprocess node embeddings within the map." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 351, + 230, + 365 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 351, + 230, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 351, + 230, + 365 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 378, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 378, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 378, + 482, + 581 + ], + "type": "text", + "content": "Baselines. No existing methods directly tackle our task, but several recent advancements provide relevant baselines. LidarCLIP [43], designed for autonomous driving, transforms LiDAR clouds into global descriptors using the Single-stride Sparse Transformer [24] and matches them with CLIP image encoder embeddings [92]. Though not a perfect fit, it can be adapted for matching point clouds with CLIP embeddings of query images. LIP-Loc [108] employs a similar approach by converting LiDAR clouds into 2D range images for direct encoding and matching. Both methods were fine-tuned on our dataset for accurate comparison. We also explored open-vocabulary [23,33,86,113] object-retrieval-based baselines using OpenMask3D [113] and OpenSeg [33], with OpenMask3D assigning CLIP descriptors to 3D object instances from multiple observing images, and OpenSeg extracting pixel-level CLIP features from the query image. Despite their original purposes, both can be adapted for localization by matching query image descriptors with object instances. For comparison with state-of-the-art visual localization methods requiring large image databases, we included CVNet [63] and AnyLoc [55]. These methods offer advanced performance but demand significant storage for image descriptors and exhibit slower inference times." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Map Generation. The mapping stage is executed offline as a preprocessing step for visual localization approaches, requiring specific mapping operations for each method before proceeding to localization. For our proposed method, this entails passing the point cloud, images, metadata, and relationships through the 3D scene graph encoder outlined in Section 3.1. For LIP-Loc, this step involves converting point clouds into range images and computing the embeddings of the range images and for LidarCLIP, point clouds are directly encoded into global" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "Y. Miao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 201, + 481, + 278 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 193 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 193 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 193 + ], + "type": "text", + "content": "Table 1: Retrieval recall on the test set of 3RScan dataset [123] (\\%; target scene ranked within the top 1, 3, and 5 of the retrieved list) and storage requirements (MB) for methods utilizing point clouds " + }, + { + "bbox": [ + 130, + 114, + 482, + 193 + ], + "type": "inline_equation", + "content": "(\\mathcal{P})" + }, + { + "bbox": [ + 130, + 114, + 482, + 193 + ], + "type": "text", + "content": ", images " + }, + { + "bbox": [ + 130, + 114, + 482, + 193 + ], + "type": "inline_equation", + "content": "(\\mathcal{I})" + }, + { + "bbox": [ + 130, + 114, + 482, + 193 + ], + "type": "text", + "content": ", and other modalities " + }, + { + "bbox": [ + 130, + 114, + 482, + 193 + ], + "type": "inline_equation", + "content": "(\\mathrm{O} = \\{\\mathcal{A}, \\mathcal{S}, \\mathcal{R}\\})" + }, + { + "bbox": [ + 130, + 114, + 482, + 193 + ], + "type": "text", + "content": " for map representation. " + }, + { + "bbox": [ + 130, + 114, + 482, + 193 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 130, + 114, + 482, + 193 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 114, + 482, + 193 + ], + "type": "inline_equation", + "content": "R^t" + }, + { + "bbox": [ + 130, + 114, + 482, + 193 + ], + "type": "text", + "content": " represent the recall in the static and temporal scenarios respectively. Additionally, metrics for single-modal methods (CVNet and AnyLoc) reliant on extensive image datasets are presented. The results are reported for scenarios where the target room is chosen from a subset of 10 and 50 candidate scenes." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 201, + 481, + 278 + ], + "lines": [ + { + "bbox": [ + 132, + 201, + 481, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 201, + 481, + 278 + ], + "type": "table", + "html": "
MethodMap modalities10 scenes50 scenesStorage (MB)
PIOR@1@3@5R@1@3@5R@1@3@5R@1@3@5
LidarCLIP [43]✓XX16.341.460.616.339.861.14.711.016.34.110.315.6
LIP-Loc [108]✓XX14.035.857.910.930.052.72.09.114.22.38.615.2
OpenMask3D [113]✓✓X---42.371.585.8---21.138.148.0
SceneGraphLoc (Ours)✓X✓53.681.992.850.576.888.430.250.261.228.246.256.4
SceneGraphLoc (Ours)✓✓✓---81.593.998.0---69.378.684.4
CVNet [63]X✓X---79.291.095.4---66.577.081.7
AnyLoc [55]X✓X---87.994.797.5---80.687.490.0
", + "image_path": "64e44c1d17447e40929ad03096ccdec33341a010fbcfd3143832698e7ab886c8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 304, + 480, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 480, + 339 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 480, + 339 + ], + "type": "text", + "content": "descriptors. In OpenMask3D, the CLIP embeddings corresponding to each object are calculated and stored. For image-based methods like CVNet and AnyLoc, embeddings for all images in the database are precomputed." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 340, + 481, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 481, + 435 + ], + "type": "text", + "content": "Metrics. To evaluate the accuracy of a method, we focus on the recall of scene selection. This entails analyzing the scenario where, given a query image and corresponding scene pair " + }, + { + "bbox": [ + 130, + 340, + 481, + 435 + ], + "type": "inline_equation", + "content": "(I, \\mathcal{G}_I)" + }, + { + "bbox": [ + 130, + 340, + 481, + 435 + ], + "type": "text", + "content": ", alongside " + }, + { + "bbox": [ + 130, + 340, + 481, + 435 + ], + "type": "inline_equation", + "content": "N - 1" + }, + { + "bbox": [ + 130, + 340, + 481, + 435 + ], + "type": "text", + "content": " alternative scenes from the database, an ordering is established for these scenes according to their computed similarity to " + }, + { + "bbox": [ + 130, + 340, + 481, + 435 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 340, + 481, + 435 + ], + "type": "text", + "content": " as determined by the tested method. The metric Recall@" + }, + { + "bbox": [ + 130, + 340, + 481, + 435 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 340, + 481, + 435 + ], + "type": "text", + "content": " is employed to ascertain whether the target scene " + }, + { + "bbox": [ + 130, + 340, + 481, + 435 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_i" + }, + { + "bbox": [ + 130, + 340, + 481, + 435 + ], + "type": "text", + "content": " is ranked among the top " + }, + { + "bbox": [ + 130, + 340, + 481, + 435 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 340, + 481, + 435 + ], + "type": "text", + "content": " scenes in terms of similarity as identified by the evaluated method. Additionally, we will report the inference time and storage requirements." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 437, + 482, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 557 + ], + "type": "text", + "content": "Experiments on 3RScan. The 3RScan dataset [123] comprises 1335 annotated indoor scenes, representing 432 distinct rooms, with 1178 scenes (385 rooms) allocated for training and 157 (47 rooms) designated for validation. The training and validation sets include semantically annotated 3D point clouds for each scene, with some captured over extended periods (e.g., several months) showcasing environmental changes. Annotations for graphs within the 3RScan dataset are provided in [125]. Due to the absence of such annotations in the test set, it was excluded from our experiments. Thus, we reorganized the original validation set, allocating 34 scenes (17 rooms) for validation and 123 scenes (30 rooms) for testing. For full reproducibility, we will publish this split." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": "During testing, we examine all 123 scenes of 30 rooms within the test set, selecting query images from each scene. The next step involves matching this image against " + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": " scenes (including the target) to ascertain whether the correct one could be identified by a method. This procedure is repeated for every image in each room. In total, all methods are tested on 30462 query images. Furthermore, we evaluate scene selection through two settings " + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "inline_equation", + "content": "N = 50" + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "inline_equation", + "content": "N = 10" + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": ". The latter setting emulates a scenario where a pre-selection strategy is employed, for example, utilizing a global scene descriptor. In image-based methods, we use all images from the database and determine the scene based on the retrieved image." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "type": "text", + "content": "SceneGraphLoc" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 190, + 481, + 280 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 180 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 180 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 180 + ], + "type": "text", + "content": "Table 2: Retrieval recall in the temporal scenario on the test set of ScanNet dataset [21] (\\%; target scene ranked within the top 1, 3, and 5 of the retrieved list) and storage requirements (MB) for methods using point clouds " + }, + { + "bbox": [ + 132, + 114, + 482, + 180 + ], + "type": "inline_equation", + "content": "(\\mathcal{P})" + }, + { + "bbox": [ + 132, + 114, + 482, + 180 + ], + "type": "text", + "content": ", images " + }, + { + "bbox": [ + 132, + 114, + 482, + 180 + ], + "type": "inline_equation", + "content": "(\\mathcal{I})" + }, + { + "bbox": [ + 132, + 114, + 482, + 180 + ], + "type": "text", + "content": ", and other modalities " + }, + { + "bbox": [ + 132, + 114, + 482, + 180 + ], + "type": "inline_equation", + "content": "(S, \\mathcal{R})" + }, + { + "bbox": [ + 132, + 114, + 482, + 180 + ], + "type": "text", + "content": " for map representation. The modality " + }, + { + "bbox": [ + 132, + 114, + 482, + 180 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 132, + 114, + 482, + 180 + ], + "type": "text", + "content": " is not available in the scene graphs predicted from [130]. The results are reported for scenarios where the target room is chosen from a subset of 10, 50 and the complete set of all (210) scenes." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 190, + 481, + 280 + ], + "lines": [ + { + "bbox": [ + 133, + 190, + 481, + 280 + ], + "spans": [ + { + "bbox": [ + 133, + 190, + 481, + 280 + ], + "type": "table", + "html": "
MethodMap modalities10 scenes50 scenesAll scenesStorage (MB)
PIORt@1at3at5Rt@1at3at5Rt@1at3at5
LidarCLIP [43]✓XX19.447.567.64.714.822.25.915.021.90.7
LIP-Loc [108]✓XX10.327.043.61.96.08.11.83.14.01.7
OpenMask3D [113]✓✓X54.984.894.031.351.363.216.527.234.517.8
SceneGraphLoc (Ours)✓X✓54.181.491.929.047.458.013.526.434.29.3
SceneGraphLoc (Ours)✓✓✓78.592.798.361.683.291.653.469.878.79.3
CVNet [63]X✓X96.598.999.692.696.097.089.993.494.6239.1
AnyLoc [55]X✓X98.499.499.896.598.198.695.196.997.45720.3
", + "image_path": "6f6815ca3de6c8122f9c57dcd4b9818ff5d60eddd9b97d4f4be5d6bee9aa9790.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 304, + 479, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 479, + 388 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 479, + 388 + ], + "type": "text", + "content": "Additionally, the methods are evaluated under both static and temporal conditions. In the static scenario, the target scene graph for a given query image originates from the same scan, albeit from a different sequence, to ensure no image overlap. Conversely, in the temporal scenario, the scene graph is derived from a sequence captured at a different temporal stage than the query, introducing potential environmental changes. We do not show results for methods exploiting images in their maps in the static stage." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 389, + 481, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 389, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 481, + 555 + ], + "type": "text", + "content": "The results are in Table 1. Despite being retrained, LidarCLIP and LIP-Loc are inaccurate, particularly in scenarios involving the selection of the target room from the entire scene set. LIP-Loc barely surpasses random selection. Although LidarCLIP exhibits marginally better accuracy, it remains inferior to alternative methods. The inferiority is because both methods learn shared embedding between entire scenes and query images of totally different objects due to viewpoint differences. OpenMask3D, while achieving better results than LidarCLIP and LIP-Loc, is less accurate than the proposed SceneGraphLoc. SceneGraphLoc, even when excluding the image modality " + }, + { + "bbox": [ + 130, + 389, + 481, + 555 + ], + "type": "inline_equation", + "content": "(\\mathcal{I})" + }, + { + "bbox": [ + 130, + 389, + 481, + 555 + ], + "type": "text", + "content": ", outperforms other cross-modal strategies significantly. Incorporating images significantly enhances its performance, positioning it close to that of image-based approaches but with three orders of magnitude smaller storage requirements. Also, the storage of SceneGraphLoc with and without images is the same due to its design of distilling knowledge into fixed-sized embeddings. An example scene is shown in Fig. 5." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "Experiments on ScanNet. To evaluate the generalization ability of our methods in real-world applications when scene graph annotations are not available, we conduct further experiments in the ScanNet dataset [21]. ScanNet encompasses 1613 monocular sequences of room-scale 3D scenes, offering 3D mesh reconstructions alongside the RGBD frame sequences utilized for the reconstructions. Given the absence of scene graph annotations in ScanNet, we run the SceneGraphFusion [130] on the RGBD sequences of scans for 3D reconstruction and scene graph prediction with 3D instance segmentation and object relationships (i.e., graph edges) within these graphs. As the process of scene graph prediction" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "Y. Miao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 170, + 146, + 444, + 239 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 3: Average time (ms) of obtaining the query image embedding " + }, + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "inline_equation", + "content": "(t_{e_q})" + }, + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": " and of the retrieval from 10, 50, and all scenes from the 3RScan [123] and ScanNet [21] datasets." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 170, + 146, + 444, + 239 + ], + "lines": [ + { + "bbox": [ + 170, + 146, + 444, + 239 + ], + "spans": [ + { + "bbox": [ + 170, + 146, + 444, + 239 + ], + "type": "table", + "html": "
Method3DRScan [123]ScanNet [21]
teqt10retrt50retrteqt10retrt50retrtallretr
LidarCLIP [43]4.10.10.34.90.10.20.6
LIP-Loc [108]2.70.10.24.10.10.20.5
OpenMask3D [113]41.54.87.420.155.41.14.5
SceneGraphLoc (Ours)28.00.31.516.61.33.717.0
CVNet [63]14.39.060.054.010.674.1311.3
AnyLoc [55]658.4354.61826.4243.068.2329.01451.1
", + "image_path": "e32c99a69d262ee34890f79f40c0103b0c7871071b7d56a89dd56c8f0d592945.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 269, + 482, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 269, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 269, + 482, + 354 + ], + "type": "text", + "content": "uses the RGBD frames of each scan, we avoid using those RGB images to match to the scene graph predicted of the scan. Thus, we only measure recall in the temporal scenario. Additionally, unlike 3RScan, the frame rate of RGBD sequences in ScanNet is high, and motion between consecutive frames is small. Thus, each database image is selected from every 25 consecutive frames in the sequence, for image-based methods [55, 63]. For a fair comparison, all the methods only use the same selected images for training and evaluation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 357, + 482, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 357, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 357, + 482, + 407 + ], + "type": "text", + "content": "For training, we use the official training set. We divide the official validation set, which includes 312 scenes, into two distinct subsets: the first 100 scenes form our validation set, while the remaining 212 are allocated for testing. To ensure full reproducibility, we will make this split publicly available." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 410, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 410, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 410, + 482, + 567 + ], + "type": "text", + "content": "The results are in Table 2 for scenarios selecting the target room from subsets of 10, 50, and the entire set of 210 scenes. The performance of LIP-Loc shows a similar pattern to that observed on 3RScan, performing only slightly better than random selection. LidarCLIP shows a small improvement in accuracy. OpenMask3D attains an accuracy comparable to our proposed method without incorporating the image modality. Our proposed SceneGraphLoc, when including the image modality in its map, significantly outperforms all cross-modal approaches. Although there remains a gap in accuracy compared with methods that use extensive image collections as maps (such as CVNet and AnyLoc), SceneGraphLoc benefits from a database size three orders-of-magnitude smaller, highlighting its efficiency and effectiveness. We partly attribute this performance gap to the lack of object attributes in the dataset and the inaccurate instance segmentation predicted by [130]. More details can be found in the supp. mat." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Processing time measured in milliseconds for various methods applied to the 3DRScan and ScanNet datasets, are detailed in Table 3. The computation time required to generate an embedding for the query image " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "(t_{e_q})" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " is notably small across all methods, typically not exceeding a few tens of milliseconds, with the exception of AnyLoc, which runs for nearly a second. The retrieval phase for cross-modal approaches is generally limited to a few tens of milliseconds. However, methods such as CVNet and AnyLoc exhibit slower performance, due to searching through extensive image collections. When tasked with selecting from" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "type": "text", + "content": "SceneGraphLoc" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 153, + 158, + 465, + 249 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 149 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 149 + ], + "type": "text", + "content": "Table 4: Ablation study performed on the val. split of 3RScan [123], analysing map modalities (" + }, + { + "bbox": [ + 132, + 114, + 482, + 149 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 132, + 114, + 482, + 149 + ], + "type": "text", + "content": " - point cloud, " + }, + { + "bbox": [ + 132, + 114, + 482, + 149 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 132, + 114, + 482, + 149 + ], + "type": "text", + "content": " - image, " + }, + { + "bbox": [ + 132, + 114, + 482, + 149 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 132, + 114, + 482, + 149 + ], + "type": "text", + "content": " - attributes, " + }, + { + "bbox": [ + 132, + 114, + 482, + 149 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 132, + 114, + 482, + 149 + ], + "type": "text", + "content": " - structure, " + }, + { + "bbox": [ + 132, + 114, + 482, + 149 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 132, + 114, + 482, + 149 + ], + "type": "text", + "content": " - relationships) and the method (Dino v2 or GCVit) to obtain the image embeddings." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 153, + 158, + 465, + 249 + ], + "lines": [ + { + "bbox": [ + 153, + 158, + 465, + 249 + ], + "spans": [ + { + "bbox": [ + 153, + 158, + 465, + 249 + ], + "type": "table", + "html": "
Map modalitiesDino v2 [81]GCVit [38]
PIASRR@1@3@5Rt@1@3@5R@1@3@5Rt@1@3@5
✓45.281.993.743.979.591.424.656.076.923.254.777.3
✓✓56.385.695.054.884.095.044.276.991.243.475.189.4
✓✓✓58.487.395.956.585.793.643.375.891.341.572.889.3
✓✓✓✓63.786.895.862.787.496.345.375.590.546.676.290.2
✓---80.296.099.0---69.487.493.7
✓✓---84.797.599.6---73.289.795.9
✓✓✓✓✓---88.597.799.6---72.188.895.7
", + "image_path": "2650bc5bdad8a365fc6486fb64e4edbfddf3d6a70f5c2ac969b80a0197ed4c79.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 271, + 479, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 271, + 479, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 271, + 479, + 295 + ], + "type": "text", + "content": "a large number of images, the processing times of these methods can extend into the range of several hundred milliseconds or even reach upwards of a second." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 296, + 480, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 296, + 480, + 343 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 480, + 343 + ], + "type": "text", + "content": "Ablation Studies. To understand the impact of the integration of different modalities, we provide an ablation study on the localization performance with multiple combinations of modalities. We use the validation split of the 3DRScan dataset. Additional ablation studies can be found in supplementary material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 344, + 482, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 344, + 482, + 476 + ], + "spans": [ + { + "bbox": [ + 130, + 344, + 482, + 476 + ], + "type": "text", + "content": "Table 4 displays the coarse localization performance of our method under the incorporation of distinct modalities " + }, + { + "bbox": [ + 130, + 344, + 482, + 476 + ], + "type": "inline_equation", + "content": "(\\mathcal{P},\\mathcal{I},\\mathcal{A},\\mathcal{S},\\mathcal{R})" + }, + { + "bbox": [ + 130, + 344, + 482, + 476 + ], + "type": "text", + "content": " within the pipeline. Additionally, the table reports results exploiting various backbones (Dino V2 [81] and GCVit [38]) for the extraction of image features. Similarly to the main experiments, we only show results on the temporal set when the map contains images. Employing Dino V2 for encoding both the query image and images within the map significantly enhances accuracy over using GCVit. The proposed method significantly outperforms both LidarCLIP and LIP-Loc (their results are in Table 1) even when using only point clouds in the map. Each additional modality contributes to the final accuracy, demonstrating that each plays a crucial role in enhancing the final performance when integrated into the pipeline." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 494, + 220, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 494, + 220, + 506 + ], + "spans": [ + { + "bbox": [ + 132, + 494, + 220, + 506 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 518, + 482, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 518, + 482, + 627 + ], + "spans": [ + { + "bbox": [ + 130, + 518, + 482, + 627 + ], + "type": "text", + "content": "In conclusion, we introduce SceneGraphLoc, a novel method for solving the novel problem of localizing an input image within a 3D scene graph-based multi-modal reference map. This approach outperforms existing cross-modal methods by a large margin. It achieves comparable accuracy to state-of-the-art image-based techniques with significantly lower storage requirements and faster processing speeds. Our experiments across the 3RScan and ScanNet datasets demonstrate the effectiveness of SceneGraphLoc, with the best performance achieved when integrating all proposed modalities. We believe that SceneGraphLoc is a step towards lightweight and efficient localization." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "Y. Miao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 250, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 250, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 250, + 129 + ], + "type": "text", + "content": "Acknowledgements." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 224 + ], + "type": "text", + "content": "We would like to thank our colleagues Ganlin Zhang, Sayan Deb Sarkar, and Cathrin Elich for their valuable advice and insightful discussions throughout the course of this research. Their contributions and suggestions greatly enhanced the quality and depth of this work. This work was partially funded by Design++ initiative of ETH Zurich, by the ETH RobotX research grant, the Hasler Stiftung Research Grant via the ETH Zurich Foundation, an ETH AI Center postdoctoral research fellowship and an ETH Zurich Career Seed Award." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 239, + 197, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 239, + 197, + 251 + ], + "spans": [ + { + "bbox": [ + 133, + 239, + 197, + 251 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 262, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 141, + 262, + 481, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 262, + 481, + 297 + ], + "spans": [ + { + "bbox": [ + 141, + 262, + 481, + 297 + ], + "type": "text", + "content": "1. Agia, C., Jatavallabhula, K.M., Khodeir, M., Miksik, O., Vineet, V., Mukadam, M., Paull, L., Shkurti, F.: Taskgraphy: Evaluating robot task planning over large 3D scene graphs. In: Conference on Robot Learning (CoRL) (2022) 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 296, + 481, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 296, + 481, + 329 + ], + "spans": [ + { + "bbox": [ + 141, + 296, + 481, + 329 + ], + "type": "text", + "content": "2. Arandjelovic, R., Gronat, P., Torii, A., Pajdla, T., Sivic, J.: NetVLAD: CNN architecture for weakly supervised place recognition. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2016) 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 328, + 481, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 328, + 481, + 372 + ], + "spans": [ + { + "bbox": [ + 141, + 328, + 481, + 372 + ], + "type": "text", + "content": "3. Armeni, I., He, Z.Y., Gwak, J., Zamir, A.R., Fischer, M., Malik, J., Savarese, S.: 3D Scene Graph: A Structure for Unified semantics, 3d space, and camera. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2019) 2, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 372, + 481, + 394 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 372, + 481, + 394 + ], + "spans": [ + { + "bbox": [ + 141, + 372, + 481, + 394 + ], + "type": "text", + "content": "4. Aubry, M., Russell, B.C., Sivic, J.: Painting-to-3d model alignment via discriminative visual elements. ACM Transactions On Graphics (TOG) (2014) 3, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 394, + 481, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 394, + 481, + 426 + ], + "spans": [ + { + "bbox": [ + 141, + 394, + 481, + 426 + ], + "type": "text", + "content": "5. Aubry, M., Russell, B.C., Sivic, J.: Visual Geo-localization of Non-photographic Depictions via 2D-3D Alignment. In: Large-Scale Visual Geo-Localization (2016) 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 426, + 481, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 426, + 481, + 459 + ], + "spans": [ + { + "bbox": [ + 141, + 426, + 481, + 459 + ], + "type": "text", + "content": "6. Balntas, V., Li, S., Prisacariu, V.: RelocNet: Continuous Metric Learning Relocalisation using Neural Nets. In: European Conference on Computer Vision (ECCV) (2018) 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 459, + 481, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 459, + 481, + 491 + ], + "spans": [ + { + "bbox": [ + 141, + 459, + 481, + 491 + ], + "type": "text", + "content": "7. Bernreiter, L., Ott, L., Nieto, J., Siegwart, R., Cadena, C.: Spherical multi-modal place recognition for heterogeneous sensor systems. In: International Conference on Robotics and Automation (ICRA) (2021) 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 491, + 481, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 491, + 481, + 524 + ], + "spans": [ + { + "bbox": [ + 141, + 491, + 481, + 524 + ], + "type": "text", + "content": "8. Berton, G., Masone, C., Caputo, B.: Rethinking visual geo-localization for large-scale applications. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2022) 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 524, + 481, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 524, + 481, + 556 + ], + "spans": [ + { + "bbox": [ + 141, + 524, + 481, + 556 + ], + "type": "text", + "content": "9. Berton, G., Paolicelli, V., Masone, C., Caputo, B.: Adaptive-attentive geolocation from few queries: A hybrid approach. In: IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) (2021) 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 556, + 481, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 556, + 481, + 589 + ], + "spans": [ + { + "bbox": [ + 138, + 556, + 481, + 589 + ], + "type": "text", + "content": "10. Bhayani, S., Sattler, T., Barath, D., Beliansky, P., Heikkilä, J., Kukelova, Z.: Calibrated and partially calibrated semi-generalized homographies. In: International Conference on Computer Vision (ICCV) (2021) 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 589, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 589, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 589, + 481, + 632 + ], + "type": "text", + "content": "11. Brachmann, E., Krull, A., Nowozin, S., Shotton, J., Michel, F., Gumhold, S., Rother, C.: DSAC - Differentiable RANSAC for Camera Localization. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2017) 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 481, + 666 + ], + "type": "text", + "content": "12. Brachmann, E., Rother, C.: Learning Less is More - 6D Camera Localization via 3D Surface Regression. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2018) 3" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "type": "text", + "content": "SceneGraphLoc" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 481, + 149 + ], + "type": "text", + "content": "13. Brachmann, E., Rother, C.: Visual camera re-localization from RGB and RGB-D images using DSAC. Transactions on Pattern Analysis and Machine Intelligence (PAMI) (2021) 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 149, + 481, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 149, + 481, + 193 + ], + "spans": [ + { + "bbox": [ + 138, + 149, + 481, + 193 + ], + "type": "text", + "content": "14. Brejcha, J., Lukac, M., Hold-Geoffroy, Y., Wang, O., Cadik, M.: LandscapeAR: Large Scale Outdoor Augmented Reality by Matching Photographs with Terrain Models Using Learned Descriptors. In: European Conference on Computer Vision (ECCV) (2020) 3, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 193, + 481, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 193, + 481, + 214 + ], + "spans": [ + { + "bbox": [ + 138, + 193, + 481, + 214 + ], + "type": "text", + "content": "15. Cadik, M., Sykora, D., Lee, S.: Automated outdoor depth-map generation and alignment. Computers & Graphics (2018) 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 214, + 481, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 214, + 481, + 246 + ], + "spans": [ + { + "bbox": [ + 138, + 214, + 481, + 246 + ], + "type": "text", + "content": "16. Castle, R., Klein, G., Murray, D.W.: Video-rate localization in multiple maps for wearable augmented reality. In: IEEE International Symposium on Wearable Computers (2008) 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 246, + 481, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 246, + 481, + 288 + ], + "spans": [ + { + "bbox": [ + 138, + 246, + 481, + 288 + ], + "type": "text", + "content": "17. Cavallari, T., Bertinetto, L., Mukhoti, J., Torr, P., Golodetz, S.: Let's take this online: Adapting scene coordinate regression network predictions for online RGB-D camera relocalisation. In: International Conference on 3D Vision (3DV) (2019) 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 289, + 481, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 289, + 481, + 322 + ], + "spans": [ + { + "bbox": [ + 138, + 289, + 481, + 322 + ], + "type": "text", + "content": "18. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention Mask Transformer for Universal Image Segmentation. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2022) 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 322, + 481, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 322, + 481, + 354 + ], + "spans": [ + { + "bbox": [ + 138, + 322, + 481, + 354 + ], + "type": "text", + "content": "19. Csurka, G., Dance, C., Fan, L., Willamowski, J., Bray, C.: Visual categorization with bags of keypoints. In: European Conference on Computer Vision (ECCV) Workshops (2004) 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 354, + 481, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 354, + 481, + 386 + ], + "spans": [ + { + "bbox": [ + 138, + 354, + 481, + 386 + ], + "type": "text", + "content": "20. Curless, B., Levoy, M.: A volumetric method for building complex models from range images. In: Annual conference on Computer graphics and interactive techniques (1996) 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 386, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 386, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 138, + 386, + 481, + 418 + ], + "type": "text", + "content": "21. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2017) 12, 13" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 418, + 481, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 418, + 481, + 450 + ], + "spans": [ + { + "bbox": [ + 138, + 418, + 481, + 450 + ], + "type": "text", + "content": "22. Doan, A.D., Latif, Y., Chin, T.J., Liu, Y., Do, T.T., Reid, I.: Scalable place recognition under appearance change for autonomous driving. In: International Conference on Computer Vision (ICCV) (2019) 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 450, + 481, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 450, + 481, + 493 + ], + "spans": [ + { + "bbox": [ + 138, + 450, + 481, + 493 + ], + "type": "text", + "content": "23. Engelmann, F., Manhardt, F., Niemeyer, M., Tateno, K., Tombari, F.: OpenNeRF: Open Set 3D Neural Scene Segmentation with Pixel-Wise Features and Rendered Novel Views. In: International Conference on Learning Representations (ICLR) (2024) 10" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 493, + 481, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 493, + 481, + 535 + ], + "spans": [ + { + "bbox": [ + 138, + 493, + 481, + 535 + ], + "type": "text", + "content": "24. Fan, L., Pang, Z., Zhang, T., Wang, Y.X., Zhao, H., Wang, F., Wang, N., Zhang, Z.: Embracing single stride 3d object detector with sparse transformer. International Conference on Computer Vision and Pattern Recognition (CVPR) (2022) 10" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 536, + 481, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 536, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 138, + 536, + 481, + 568 + ], + "type": "text", + "content": "25. Gadre, S.Y., Ehsani, K., Song, S., Mottaghi, R.: Continuous scene representations for embodied ai. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2022) 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 568, + 481, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 568, + 481, + 612 + ], + "spans": [ + { + "bbox": [ + 138, + 568, + 481, + 612 + ], + "type": "text", + "content": "26. Gao, P., Liang, J., Shen, Y., Son, S., Lin, M.C.: Visual, Spatial, Geometric-Preserved Place Recognition for Cross-View and Cross-Modal Collaborative Perception. In: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) (2023) 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 612, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 612, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 138, + 612, + 481, + 633 + ], + "type": "text", + "content": "27. Garg, S., Fischer, T., Milford, M.: Where is your place, visual place recognition? International Joint Conference on Artificial Intelligence (IJCAI) (2021) 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "type": "text", + "content": "28. Garg, S., Suenderhauf, N., Milford, M.: Semantic-geometric visual place recognition: a new perspective for reconciling opposing views. International Journal on Robotics Research (IJRR) (2019) 2, 4" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "Y. Miao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 116, + 480, + 159 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 480, + 159 + ], + "type": "text", + "content": "29. Garg, S., Sünderhauf, N., Dayoub, F., Morrison, D., Cosgun, A., Carneiro, G., Wu, Q., Chin, T.J., Reid, I., Gould, S., et al.: Semantics for robotic mapping, perception and interaction: A survey. Foundations and Trends in Robotics (2020) 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 160, + 480, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 160, + 480, + 192 + ], + "spans": [ + { + "bbox": [ + 138, + 160, + 480, + 192 + ], + "type": "text", + "content": "30. Georgakis, G., Karanam, S., Wu, Z., Kosecka, J.: Learning local rgb-to-cad correspondences for object pose estimation. In: International Conference on Computer Vision (ICCV) (2019) 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 192, + 480, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 192, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 138, + 192, + 480, + 225 + ], + "type": "text", + "content": "31. Germain, H., Bourmaud, G., Lepetit, V.: Sparse-to-Dense Hypercolumn Matching for Long-Term Visual Localization. In: International Conference on 3D Vision (3DV) (2019) 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 225, + 480, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 225, + 480, + 257 + ], + "spans": [ + { + "bbox": [ + 138, + 225, + 480, + 257 + ], + "type": "text", + "content": "32. Germain, H., Bourmaud, G., Lepetit, V.: S2DNet: Learning Image Features for Accurate Sparse-to-Dense Matching. In: European Conference on Computer Vision (ECCV) (2020) 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 257, + 480, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 257, + 480, + 290 + ], + "spans": [ + { + "bbox": [ + 138, + 257, + 480, + 290 + ], + "type": "text", + "content": "33. Ghiasi, G., Gu, X., Cui, Y., Lin, T.Y.: Scaling Open-Vocabulary Image Segmentation with Image-Level Labels. In: European Conference on Computer Vision (ECCV) (2022) 10" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 290, + 480, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 290, + 480, + 322 + ], + "spans": [ + { + "bbox": [ + 138, + 290, + 480, + 322 + ], + "type": "text", + "content": "34. Grabner, A., Roth, P.M., Lepetit, V.: 3D Pose Estimation and 3D Model Retrieval for Objects in the Wild. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2018) 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 322, + 480, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 322, + 480, + 343 + ], + "spans": [ + { + "bbox": [ + 138, + 322, + 480, + 343 + ], + "type": "text", + "content": "35. Grelsson, B., Robinson, A., Felsberg, M., Khan, F.S.: GPS-level accurate camera localization with HorizonNet. Journal of Field Robotics (2020) 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 343, + 480, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 343, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 138, + 343, + 480, + 365 + ], + "type": "text", + "content": "36. Gumeli, C., Dai, A., Nießner, M.: ROCA: Robust CAD Model Retrieval and Alignment from a Single Image. arXiv preprint arXiv:2112.01988 (2021) 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 365, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 365, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 138, + 365, + 480, + 385 + ], + "type": "text", + "content": "37. Hanocka, R., Metzer, G., Giryes, R., Cohen-Or, D.: Point2mesh: A self-prior for deformable meshes. arXiv preprint arXiv:2005.11084 (2020) 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 385, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 385, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 138, + 385, + 480, + 418 + ], + "type": "text", + "content": "38. Hatamizadeh, A., Yin, H., Heinrich, G., Kautz, J., Molchanov, P.: Global context vision transformers. In: International Conference on Machine Learning (ICML) (2023) 14" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 418, + 480, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 418, + 480, + 450 + ], + "spans": [ + { + "bbox": [ + 138, + 418, + 480, + 450 + ], + "type": "text", + "content": "39. Hausler, S., Jacobson, A., Milford, M.: Multi-process fusion: Visual place recognition using multiple image processing methods. IEEE Robotics and Automation Letters (RA-L) (2019) 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 450, + 480, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 450, + 480, + 483 + ], + "spans": [ + { + "bbox": [ + 138, + 450, + 480, + 483 + ], + "type": "text", + "content": "40. Hausler, S., Garg, S., Xu, M., Milford, M., Fischer, T.: Patch-netvlad: Multiscale fusion of locally-global descriptors for place recognition. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2021) 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 483, + 480, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 483, + 480, + 515 + ], + "spans": [ + { + "bbox": [ + 138, + 483, + 480, + 515 + ], + "type": "text", + "content": "41. He, K., Zhang, X., Ren, S., Sun, J.: Deep Residual Learning for Image Recognition. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2016) 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 515, + 480, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 515, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 138, + 515, + 480, + 557 + ], + "type": "text", + "content": "42. Heng, L., Choi, B., Cui, Z., Geppert, M., Hu, S., Kuan, B., Liu, P., Nguyen, R., Yeo, Y.C., Geiger, A., et al.: Project Autovision: Localization and 3D Scene Perception for an Autonomous Vehicle with a Multi-Camera System. In: International Conference on Robotics and Automation (ICRA) (2019) 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 558, + 480, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 558, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 138, + 558, + 480, + 590 + ], + "type": "text", + "content": "43. Hess, G., Tondersski, A., Petersson, C., Åström, K., Svensson, L.: Lidarclip or: How i learned to talk to point clouds. In: IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) (2024) 4, 10, 11, 12, 13" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 590, + 465, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 590, + 465, + 601 + ], + "spans": [ + { + "bbox": [ + 138, + 590, + 465, + 601 + ], + "type": "text", + "content": "44. Hodan, T.: Pose Estimation of Specific Rigid Objects. Ph.D. thesis (2021) 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 601, + 480, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 480, + 633 + ], + "type": "text", + "content": "45. Hodan, T., Barath, D., Matas, J.: EPOS: Estimating 6D Pose of Objects With Symmetries. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2020) 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 633, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 666 + ], + "type": "text", + "content": "46. Hodan, T., Zabulis, X., Lourakis, M.I.A., Obdrzalek, S., Matas, J.: Detection and fine 3D pose estimation of texture-less objects in RGB-D images. In: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) (2015) 4" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 382, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 382, + 91, + 447, + 101 + ], + "type": "text", + "content": "SceneGraphLoc" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 137, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 137, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 137, + 116, + 481, + 149 + ], + "type": "text", + "content": "47. Hu, S., Feng, M., Nguyen, R.H.M., Lee, G.H.: CVM-Net: Cross-View Matching Network for Image-Based Ground-to-Aerial Geo-Localization. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2018) 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 151, + 481, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 151, + 481, + 173 + ], + "spans": [ + { + "bbox": [ + 137, + 151, + 481, + 173 + ], + "type": "text", + "content": "48. Hu, S., Lee, G.H.: Image-Based GeoLocalization Using Satellite Imagery. International Journal on Computer Vision (IJCV) (2019) 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 174, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 174, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 138, + 174, + 481, + 205 + ], + "type": "text", + "content": "49. Hughes, N., Chang, Y., Carlone, L.: Hydra: A real-time spatial perception system for 3D scene graph construction and optimization. arXiv preprint arXiv:2201.13360 (2022) 4, 5, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 206, + 481, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 206, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 138, + 206, + 481, + 228 + ], + "type": "text", + "content": "50. Ibrahimi, S., van Noord, N., Alpherts, T., Worring, M.: Inside out visual place recognition. In: British Machine Vision Conference (2021) 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 229, + 481, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 229, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 138, + 229, + 481, + 262 + ], + "type": "text", + "content": "51. Irschara, A., Zach, C., Frahm, J.M., Bischof, H.: From Structure-from-Motion Point Clouds to Fast Location Recognition. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2009) 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 263, + 481, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 263, + 481, + 307 + ], + "spans": [ + { + "bbox": [ + 138, + 263, + 481, + 307 + ], + "type": "text", + "content": "52. Izadi, S., Kim, D., Hilliges, O., Molyneaux, D., Newcombe, R., Kohli, P., Shotton, J., Hodges, S., Freeman, D., Davison, A., et al.: Kinectfusion: real-time 3d reconstruction and interaction using a moving depth camera. In: ACM symposium on User interface software and technology (2011) 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 308, + 481, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 308, + 481, + 340 + ], + "spans": [ + { + "bbox": [ + 138, + 308, + 481, + 340 + ], + "type": "text", + "content": "53. Ji, X., Wei, J., Wang, Y., Shang, H., Kneip, L.: Cross-modal Place Recognition in Image Databases using Event-based Sensors. arXiv preprint arXiv:2307.01047 (2023) 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 341, + 481, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 341, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 138, + 341, + 481, + 373 + ], + "type": "text", + "content": "54. Kabalar, J., Wu, S.C., Wald, J., Tateno, K., Navab, N., Tombari, F.: Towards long-term retrieval-based visual localization in indoor environments with changes. IEEE Robotics and Automation Letters (2023) 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 375, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 375, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 138, + 375, + 481, + 407 + ], + "type": "text", + "content": "55. Keetha, N., Mishra, A., Karhade, J., Jatavallabhula, K.M., Scherer, S., Krishna, M., Garg, S.: Anyloc: Towards universal visual place recognition. IEEE Robotics and Automation Letters (RA-L) (2023) 2, 3, 10, 11, 12, 13" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 408, + 481, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 408, + 481, + 441 + ], + "spans": [ + { + "bbox": [ + 138, + 408, + 481, + 441 + ], + "type": "text", + "content": "56. Kendall, A., Cipolla, R.: Geometric loss functions for camera pose regression with deep learning. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2017) 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 442, + 481, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 442, + 481, + 474 + ], + "spans": [ + { + "bbox": [ + 138, + 442, + 481, + 474 + ], + "type": "text", + "content": "57. Kendall, A., Grimes, M., Cipolla, R.: PoseNet: A Convolutional Network for Real-Time 6-DOF Camera Relocalization. In: International Conference on Computer Vision (ICCV) (2015) 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 475, + 481, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 475, + 481, + 508 + ], + "spans": [ + { + "bbox": [ + 138, + 475, + 481, + 508 + ], + "type": "text", + "content": "58. Khaliq, A., Ehsan, S., Chen, Z., Milford, M., McDonald-Maier, K.: A holistic visual place recognition approach using lightweight CNNs for significant viewpoint and appearance changes. IEEE Transactions on Robotics (T-RO) (2020) 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 509, + 481, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 509, + 481, + 542 + ], + "spans": [ + { + "bbox": [ + 138, + 509, + 481, + 542 + ], + "type": "text", + "content": "59. Kim, H.J., Dunn, E., Frahm, J.M.: Learned contextual feature reweighting for image geolocation. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2017) 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 543, + 481, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 543, + 481, + 575 + ], + "spans": [ + { + "bbox": [ + 138, + 543, + 481, + 575 + ], + "type": "text", + "content": "60. Kim, U.H., Park, J.M., Song, T.J., Kim, J.H.: 3-D scene graph: A sparse and semantic representation of physical environments for intelligent agents. IEEE transactions on cybernetics (2019) 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 576, + 481, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 576, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 138, + 576, + 481, + 597 + ], + "type": "text", + "content": "61. Kutulakos, K.N., Seitz, S.M.: A theory of shape by space carving. International Journal on Computer Vision (IJCV) (2000) 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 599, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 599, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 599, + 481, + 632 + ], + "type": "text", + "content": "62. Labbe, Y., Carpentier, J., Aubry, M., Sivic, J.: CosyPose: Consistent multi-view multi-object 6D pose estimation. In: European Conference on Computer Vision (ECCV) (2020) 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "text", + "content": "63. Lee, S., Seong, H., Lee, S., Kim, E.: Correlation verification for image retrieval. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2022) 10, 11, 12, 13" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "Y. Miao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 116, + 480, + 148 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 480, + 148 + ], + "type": "text", + "content": "64. Li, Y., Snavely, N., Huttenlocher, D.P., Fua, P.: Worldwide Pose Estimation Using 3D Point Clouds. In: European Conference on Computer Vision (ECCV) (2012) 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 149, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 149, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 138, + 149, + 480, + 182 + ], + "type": "text", + "content": "65. Lim, H., Sinha, S.N., Cohen, M.F., Uysettendaele, M.: Real-time Image-based 6-DoF Localization in Large-scale Environments. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2012) 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 183, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 183, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 138, + 183, + 480, + 215 + ], + "type": "text", + "content": "66. Lin, T.Y., Cui, Y., Belongie, S.J., Hays, J.: Learning deep representations for ground-to-aerial geolocation. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2015) 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 216, + 480, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 216, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 138, + 216, + 480, + 247 + ], + "type": "text", + "content": "67. Lin, Z., Zhang, Z., Wang, M., Shi, Y., Wu, X., Zheng, Y.: Multi-modal contrastive representation learning for entity alignment. arXiv preprint arXiv:2209.00891 (2022) 9" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 249, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 249, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 138, + 249, + 480, + 281 + ], + "type": "text", + "content": "68. Liu, L., Li, H., Dai, Y.: Efficient Global 2D-3D Matching for Camera Localization in a Large-Scale 3D Map. In: International Conference on Computer Vision (ICCV) (2017) 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 282, + 480, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 282, + 480, + 313 + ], + "spans": [ + { + "bbox": [ + 138, + 282, + 480, + 313 + ], + "type": "text", + "content": "69. Liu, L., Li, H., Dai, Y.: Stochastic attraction-repulsion embedding for large scale image localization. In: International Conference on Computer Vision (ICCV) (2019) 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 315, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 315, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 138, + 315, + 480, + 346 + ], + "type": "text", + "content": "70. Looper, S., Rodriguez-Puigvert, J., Siegwart, R., Cadena, C., Schmid, L.: 3D VSG: Long-term semantic scene change prediction through 3D variable scene graphs. In: International Conference on Robotics and Automation (ICRA) (2023) 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 347, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 347, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 138, + 347, + 480, + 380 + ], + "type": "text", + "content": "71. Lynen, S., Zeisl, B., Aiger, D., Bosse, M., Hesch, J., Pollefeys, M., Siegwart, R., Sattler, T.: Large-scale, real-time visual-inertial localization revisited. International Journal on Robotics Research (IJRR) (2020) 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 380, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 380, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 138, + 380, + 480, + 413 + ], + "type": "text", + "content": "72. Lynen, S., Zeisl, B., Aiger, D., Bosse, M., Hesch, J., Pollefeys, M., Siegwart, R., Sattler, T.: Large-scale, real-time visual-inertial localization revisited. International Journal on Robotics Research (IJRR) (2020) 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 414, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 414, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 138, + 414, + 480, + 445 + ], + "type": "text", + "content": "73. Mescheder, L., Oechsle, M., Niemeyer, M., Nowozin, S., Geiger, A.: Occupancy networks: Learning 3d reconstruction in function space. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2019) 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 447, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 447, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 447, + 480, + 468 + ], + "type": "text", + "content": "74. Miao, Y., Armeni, I., Pollefeys, M., Barath, D.: Volumetric Semantically Consistent 3D Panoptic Mapping. arXiv preprint arXiv:2309.14737 (2024) 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 468, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 468, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 138, + 468, + 480, + 501 + ], + "type": "text", + "content": "75. Miao, Y., Li, C., Li, Z., Yang, Y., Yu, X.: A novel algorithm of ship structure modeling and target identification based on point cloud for automation in bulk cargo terminals. Measurement and Control (2021) 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 502, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 502, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 502, + 480, + 533 + ], + "type": "text", + "content": "76. Mihajlovic, M., Weder, S., Pollefeys, M., Oswald, M.R.: DeepSurfels: Learning online appearance fusion. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2021) 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 534, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 534, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 138, + 534, + 480, + 566 + ], + "type": "text", + "content": "77. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM (2021) 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 567, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 480, + 599 + ], + "type": "text", + "content": "78. Moreau, A., Piasco, N., Tsishkou, D., Stanciulescu, B., de La Fortelle, A.: LENS: Localization enhanced by neRF synthesis. In: Conference on Robot Learning (CoRL) (2021) 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "type": "text", + "content": "79. Murez, Z., Van As, T., Bartolozzi, J., Sinha, A., Badrinarayanan, V., Rabinovich, A.: Atlas: End-to-end 3d scene reconstruction from posed images. In: European Conference on Computer Vision (ECCV) (2020) 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "text", + "content": "80. Oechsle, M., Mescheder, L., Niemeyer, M., Strauss, T., Geiger, A.: Texture fields: Learning texture representations in function space. In: International Conference on Computer Vision (ICCV) (2019) 4" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 382, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 382, + 91, + 447, + 101 + ], + "type": "text", + "content": "SceneGraphLoc" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + }, + { + "para_blocks": [ + { + "bbox": [ + 137, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 137, + 116, + 481, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 116, + 481, + 159 + ], + "spans": [ + { + "bbox": [ + 137, + 116, + 481, + 159 + ], + "type": "text", + "content": "81. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023) 7, 8, 14" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 160, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 160, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 137, + 160, + 481, + 182 + ], + "type": "text", + "content": "82. Panek, V., Kukelova, Z., Sattler, T.: Meshloc: Mesh-based visual localization. In: European Conference on Computer Vision (ECCV) (2022) 3, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 137, + 182, + 481, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 182, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 137, + 182, + 481, + 224 + ], + "type": "text", + "content": "83. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2019) 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 224, + 481, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 224, + 481, + 257 + ], + "spans": [ + { + "bbox": [ + 137, + 224, + 481, + 257 + ], + "type": "text", + "content": "84. Peng, G., Yue, Y., Zhang, J., Wu, Z., Tang, X., Wang, D.: Semantic reinforced attention learning for visual place recognition. In: International Conference on Robotics and Automation (ICRA) (2021) 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 257, + 481, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 257, + 481, + 289 + ], + "spans": [ + { + "bbox": [ + 137, + 257, + 481, + 289 + ], + "type": "text", + "content": "85. Peng, G., Zhang, J., Li, H., Wang, D.: Attentional pyramid pooling of salient visual residuals for place recognition. In: International Conference on Computer Vision (ICCV) (2021) 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 289, + 481, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 289, + 481, + 322 + ], + "spans": [ + { + "bbox": [ + 137, + 289, + 481, + 322 + ], + "type": "text", + "content": "86. Peng, S., Genova, K., Jiang, C.M., Tagliasacchi, A., Pollefeys, M., Funkhouser, T.: OpenScene: 3D Scene Understanding with Open Vocabularies. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2023) 10" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 137, + 322, + 481, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 322, + 481, + 354 + ], + "spans": [ + { + "bbox": [ + 137, + 322, + 481, + 354 + ], + "type": "text", + "content": "87. Peng, S., Niemeyer, M., Mescheder, L., Pollefeys, M., Geiger, A.: Convolutional occupancy networks. In: European Conference on Computer Vision (ECCV) (2020) 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 137, + 354, + 481, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 354, + 481, + 386 + ], + "spans": [ + { + "bbox": [ + 137, + 354, + 481, + 386 + ], + "type": "text", + "content": "88. Pion, N., Humenberger, M., Csurka, G., Cabon, Y., Sattler, T.: Benchmarking image retrieval for visual localization. In: International Conference on 3D Vision (3DV) (2020) 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 137, + 386, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 386, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 137, + 386, + 481, + 418 + ], + "type": "text", + "content": "89. Plotz, T., Roth, S.: Automatic Registration of Images to Untextured Geometry Using Average Shading Gradients. International Journal on Computer Vision (IJCV) (2017) 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 137, + 418, + 481, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 418, + 481, + 451 + ], + "spans": [ + { + "bbox": [ + 137, + 418, + 481, + 451 + ], + "type": "text", + "content": "90. Ponimatkin, G., Labbe, Y., Russell, B., Aubry, M., Sivic, J.: Focal length and object pose estimation via render and compare. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2022) 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 137, + 450, + 481, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 450, + 481, + 482 + ], + "spans": [ + { + "bbox": [ + 137, + 450, + 481, + 482 + ], + "type": "text", + "content": "91. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2017) 4, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 137, + 482, + 481, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 482, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 137, + 482, + 481, + 525 + ], + "type": "text", + "content": "92. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. International Conference on Machine Learning (ICML) (2021) 10" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 137, + 525, + 481, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 525, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 137, + 525, + 481, + 557 + ], + "type": "text", + "content": "93. Ramalingam, S., Bouaziz, S., Sturm, P.F., Brand, M.: SKYLINE2GPS: Localization in urban canyons using omni-skylines. In: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) (2010) 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 137, + 557, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 557, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 137, + 557, + 481, + 601 + ], + "type": "text", + "content": "94. Ravichandran, Z., Peng, L., Hughes, N., Griffith, J., Carlone, L.: Hierarchical representations and explicit memory: Learning effective navigation policies on 3D scene graphs using graph neural networks. In: International Conference on Robotics and Automation (ICRA) (2022) 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 137, + 601, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 601, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 137, + 601, + 481, + 633 + ], + "type": "text", + "content": "95. Rosinol, A., Violette, A., Abate, M., Hughes, N., Chang, Y., Shi, J., Gupta, A., Carlone, L.: Kimera: From SLAM to spatial perception with 3D dynamic scene graphs. International Journal on Robotics Research (IJRR) (2021) 4, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 137, + 633, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 633, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 137, + 633, + 481, + 666 + ], + "type": "text", + "content": "96. Rosinol, A., Gupta, A., Abate, M., Shi, J., Carlone, L.: 3D dynamic scene graphs: Actionable spatial perception with places, objects, and humans. arXiv preprint arXiv:2002.06289 (2020) 4" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "20" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "Y. Miao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 19 + }, + { + "para_blocks": [ + { + "bbox": [ + 134, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 136, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 136, + 116, + 480, + 149 + ], + "type": "text", + "content": "97. Sarkar, S.D., Miksik, O., Pollefeys, M., Barath, D., Armeni, I.: SGAligner: 3D Scene Alignment with Scene Graphs. International Conference on Computer Vision (ICCV) (2023) 4, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 149, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 149, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 137, + 149, + 480, + 182 + ], + "type": "text", + "content": "98. Sarlin, P.E., Cadena, C., Siegwart, R., Dymczyk, M.: From Coarse to Fine: Robust Hierarchical Localization at Large Scale. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2019) 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 137, + 182, + 480, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 182, + 480, + 214 + ], + "spans": [ + { + "bbox": [ + 137, + 182, + 480, + 214 + ], + "type": "text", + "content": "99. Sarlin, P.E., DeTone, D., Malisiewicz, T., Rabinovich, A.: Superglue: Learning feature matching with graph neural networks. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2020) 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 214, + 480, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 214, + 480, + 257 + ], + "spans": [ + { + "bbox": [ + 134, + 214, + 480, + 257 + ], + "type": "text", + "content": "100. Sarlin, P.E., DeTone, D., Yang, T.Y., Avetisyan, A., Straub, J., Malisiewicz, T., Bulò, S.R., Newcombe, R., Kontschieder, P., Balntas, V.: OrienterNet: Visual Localization in 2D Public Maps with Neural Matching. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2023) 2, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 134, + 257, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 257, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 134, + 257, + 480, + 300 + ], + "type": "text", + "content": "101. Sarlin, P.E., Unagar, A., Larsson, M., Germain, H., Toft, C., Larsson, V., Pollefeys, M., Lepetit, V., Hammarstrand, L., Kahl, F., et al.: Back to the Feature: Learning Robust Camera Localization from Pixels to Pose. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2021) 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 134, + 300, + 480, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 300, + 480, + 332 + ], + "spans": [ + { + "bbox": [ + 134, + 300, + 480, + 332 + ], + "type": "text", + "content": "102. Sattler, T., Leibe, B., Kobbelt, L.: Efficient & Effective Prioritized Matching for Large-Scale Image-Based Localization. Transactions on Pattern Analysis and Machine Intelligence (PAMI) (2017) 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 134, + 332, + 480, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 332, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 134, + 332, + 480, + 365 + ], + "type": "text", + "content": "103. Sattler, T., Zhou, Q., Pollefeys, M., Leal-Taixe, L.: Understanding the limitations of cnn-based absolute camera pose regression. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2019) 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 134, + 365, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 365, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 134, + 365, + 480, + 407 + ], + "type": "text", + "content": "104. Savinov, N., Hane, C., Ladicky, L., Pollefeys, M.: Semantic 3d reconstruction with continuous regularization and ray potentials using a visibility consistency constraint. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2016) 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 134, + 407, + 480, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 407, + 480, + 440 + ], + "spans": [ + { + "bbox": [ + 134, + 407, + 480, + 440 + ], + "type": "text", + "content": "105. Schonberger, J.L., Pollefeys, M., Geiger, A., Sattler, T.: Semantic visual localization. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2018) 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 134, + 440, + 480, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 440, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 134, + 440, + 480, + 472 + ], + "type": "text", + "content": "106. Sepulveda, G., Niebles, J., Soto, A.: A deep learning based behavioral approach to indoor autonomous navigation. In: International Conference on Robotics and Automation (ICRA) (2018) 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 134, + 472, + 480, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 472, + 480, + 504 + ], + "spans": [ + { + "bbox": [ + 134, + 472, + 480, + 504 + ], + "type": "text", + "content": "107. Shan, Q., Wu, C., Curless, B., Furukawa, Y., Hernandez, C., Seitz, S.M.: Accurate geo-registration by ground-to-aerial image matching. In: International Conference on 3D Vision (3DV) (2014) 3, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 504, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 504, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 134, + 504, + 480, + 536 + ], + "type": "text", + "content": "108. Shubodh, S., Omama, M., Zaidi, H., Parihar, U.S., Krishna, M.: Lip-loc: Lidar image pretraining for cross-modal localization. In: IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) (2024) 4, 10, 11, 12, 13" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 536, + 480, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 536, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 134, + 536, + 480, + 557 + ], + "type": "text", + "content": "109. Sibling, D., Sattler, T., Leibe, B., Kobbelt, L.: SIFT-Realistic Rendering. In: International Conference on 3D Vision (3DV) (2013) 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 557, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 557, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 134, + 557, + 480, + 600 + ], + "type": "text", + "content": "110. Steiger Mueller, M., Sattler, T., Pollefeys, M., Jutz, B.: Image-to-image translation for enhanced feature matching, image retrieval and visual localization. ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences (2019) 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 601, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 601, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 601, + 480, + 632 + ], + "type": "text", + "content": "111. Stuckler, J., Behnke, S.: Multi-resolution surfel maps for efficient dense 3D modeling and tracking. Journal of Visual Communication and Image Representation (2014) 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 632, + 480, + 665 + ], + "type": "text", + "content": "112. Svarm, L., Enqvist, O., Kahl, F., Oskarsson, M.: City-Scale Localization for Cameras with Known Vertical Direction. Transactions on Pattern Analysis and Machine Intelligence (PAMI) (2017) 3" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "type": "text", + "content": "SceneGraphLoc" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "21" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 20 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "text", + "content": "113. Takmaz, A., Fedele, E., Sumner, R.W., Pollefeys, M., Tombari, F., Engelmann, F.: OpenMask3D: Open-Vocabulary 3D Instance Segmentation. In: International Conference on Neural Information Processing Systems (NeurIPS) (2023) 7, 10, 11, 12, 13" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 161, + 481, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 481, + 195 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 481, + 195 + ], + "type": "text", + "content": "114. Tewari, A., Thies, J., Mildenhall, B., Srinivasan, P., Tretschk, E., Yifan, W., Lassner, C., Sitzmann, V., Martin-Brualla, R., Lombardi, S., et al.: Advances in neural rendering. In: Computer Graphics Forum (2022) 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 195, + 481, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 195, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 195, + 481, + 239 + ], + "type": "text", + "content": "115. Tomesek, J., Cadik, M., Brejcha, J.: CrossLocate: Cross-modal Large-scale Visual Geo-Localization in Natural Environments using Rendered Modalities. In: IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) (2022) 3, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 239, + 481, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 239, + 481, + 274 + ], + "spans": [ + { + "bbox": [ + 132, + 239, + 481, + 274 + ], + "type": "text", + "content": "116. Torii, A., Arandjelovic, R., Sivic, J., Okutomi, M., Pajdla, T.: 24/7 place recognition by view synthesis. Transactions on Pattern Analysis and Machine Intelligence (PAMI) (2018) 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 274, + 481, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 481, + 308 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 481, + 308 + ], + "type": "text", + "content": "117. Torii, A., Taira, H., Sivic, J., Pollefeys, M., Okutomi, M., Pajdla, T., Sattler, T.: Are large-scale 3d models really necessary for accurate visual localization? Transactions on Pattern Analysis and Machine Intelligence (PAMI) (2021) 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 308, + 481, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 481, + 351 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 481, + 351 + ], + "type": "text", + "content": "118. Valentin, J., Nießner, M., Shotton, J., Fitzgibbon, A., Izadi, S., Torr, P.: Exploiting Uncertainty in Regression Forests for Accurate Camera Relocalization. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2015) 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 352, + 481, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 352, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 352, + 481, + 385 + ], + "type": "text", + "content": "119. Velicković, P., Cucurull, G., Casanova, A., Romero, A., Lio, P., Bengio, Y.: Graph Attention Networks. In: International Conference on Learning Representations (ICLR) (2018) 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 385, + 481, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 385, + 481, + 419 + ], + "spans": [ + { + "bbox": [ + 132, + 385, + 481, + 419 + ], + "type": "text", + "content": "120. Ventura, J., Kukelova, Z., Sattler, T., Baráth, D.: Absolute Pose from One or Two Scaled and Oriented Features. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2024) 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 419, + 481, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 419, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 419, + 481, + 453 + ], + "type": "text", + "content": "121. Viswanathan, A., Rodrigues Pires, B., Huber, D.F.: Vision based robot localization by ground to satellite matching in GPS-denied situations. In: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) (2014) 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 453, + 481, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 481, + 486 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 481, + 486 + ], + "type": "text", + "content": "122. Walch, F., Hazirbas, C., Leal-Taixe, L., Sattler, T., Hilsenbeck, S., Cremers, D.: Image-Based Localization Using LSTMs for Structured Feature Correlation. In: International Conference on Computer Vision (ICCV) (2017) 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 486, + 481, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 486, + 481, + 520 + ], + "spans": [ + { + "bbox": [ + 132, + 486, + 481, + 520 + ], + "type": "text", + "content": "123. Wald, J., Avetisyan, A., Navab, N., Tombari, F., Niefsner, M.: Rio: 3d object instance re-localization in changing indoor environments. In: International Conference on Computer Vision (ICCV) (2019) 8, 9, 10, 11, 13, 14" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 520, + 481, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 520, + 481, + 553 + ], + "spans": [ + { + "bbox": [ + 132, + 520, + 481, + 553 + ], + "type": "text", + "content": "124. Wald, J., Dhamo, H., Navab, N., Tombari, F.: Learning 3d semantic scene graphs from 3d indoor reconstructions. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2020) 2, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 554, + 481, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 554, + 481, + 587 + ], + "spans": [ + { + "bbox": [ + 132, + 554, + 481, + 587 + ], + "type": "text", + "content": "125. Wald, J., Dhamo, H., Navab, N., Tombari, F.: Learning 3d semantic scene graphs from 3d indoor reconstructions. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2020) 11" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 587, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 587, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 587, + 481, + 621 + ], + "type": "text", + "content": "126. Wang, S., Kannala, J., Barath, D.: DGC-GNN: Descriptor-free Geometric-Color Graph Neural Network for 2D-3D Matching. International Conference on Computer Vision and Pattern Recognition (CVPR) (2023) 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "text", + "content": "127. Warburg, F., Hauberg, S., Lopez-Antequera, M., Gargallo, P., Kuang, Y., Civera, J.: Mapillary street-level sequences: A dataset for lifelong place recognition. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2020) 2, 3" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "22" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "Y. Miao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 21 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 632 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "128. Weder, S., Schonberger, J.L., Pollefeys, M., Oswald, M.R.: Neuralfusion: Online depth fusion in latent space. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2021) 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 182 + ], + "type": "text", + "content": "129. Workman, S., Souvenir, R., Jacobs, N.: Wide-Area Image Geolocation with Aerial Reference Imagery. In: International Conference on Computer Vision (ICCV) (2015) 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 183, + 480, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 183, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 133, + 183, + 480, + 225 + ], + "type": "text", + "content": "130. Wu, S.C., Wald, J., Tateno, K., Navab, N., Tombari, F.: Scenegraphfusion: Incremental 3d scene graph prediction from rgb-d sequences. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2021) 4, 7, 12, 13" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "text", + "content": "131. Ying, Z., Yuan, X., Yang, B., Song, Y., Xu, Q., Zhou, F., Sheng, W.: RP-SG: Relation Prediction in 3D Scene Graphs for Unobserved Objects Localization. IEEE Robotics and Automation Letters (RA-L) (2023) 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "type": "text", + "content": "132. Zaffar, M., Garg, S., Milford, M., et al.: Vpr-bench: An open-source visual place recognition evaluation framework with quantifiable viewpoint and appearance change. International Journal on Computer Vision (IJCV) (2021) 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "type": "text", + "content": "133. Zeisl, B., Sattler, T., Pollefeys, M.: Camera pose voting for large-scale image-based localization. In: International Conference on Computer Vision (ICCV) (2015) 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 315, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 315, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 133, + 315, + 480, + 346 + ], + "type": "text", + "content": "134. Zhang, C., Yu, J., Song, Y., Cai, W.: Exploiting edge-oriented reasoning for 3D point-based scene graph analysis. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2021) 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 347, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 347, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 133, + 347, + 480, + 380 + ], + "type": "text", + "content": "135. Zhang, G., Larsson, V., Barath, D.: Revisiting Rotation Averaging: Uncertainties and Robust Losses. In: International Conference on Computer Vision and Pattern Recognition (CVPR) (2023) 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 380, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 380, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 133, + 380, + 480, + 413 + ], + "type": "text", + "content": "136. Zhang, S., Hao, A., Qin, H.: Knowledge-inspired 3D scene graph prediction in point cloud. International Conference on Neural Information Processing Systems (NeurIPS) (2021) 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 413, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 413, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 133, + 413, + 480, + 445 + ], + "type": "text", + "content": "137. Zhang, W., Kosecka, J.: Image based localization in urban environments. In: International Symposium on 3D Data Processing, Visualization, and Transmission (2006) 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 447, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 447, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 133, + 447, + 480, + 479 + ], + "type": "text", + "content": "138. Zhang, Z., Sattler, T., Scaramuzza, D.: Reference Pose Generation for Long-term Visual Localization via Learned Features and View Synthesis. International Journal on Computer Vision (IJCV) (2020) 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 479, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 479, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 133, + 479, + 480, + 511 + ], + "type": "text", + "content": "139. Zhao, L., Gatsis, K., Papachristodoulou, A.: Stable and Safe Reinforcement Learning via a Barrier-Lyapunov Actor-Critic Approach. In: IEEE Conference on Decision and Control (CDC) (2023) 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 512, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 512, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 133, + 512, + 480, + 544 + ], + "type": "text", + "content": "140. Zhao, L., Miao, K., Gatsis, K., Papachristodoulou, A.: Stable and safe human-aligned reinforcement learning through neural ordinary differential equations. arXiv preprint arXiv:2401.13148 (2024) 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 545, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 545, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 133, + 545, + 480, + 567 + ], + "type": "text", + "content": "141. Zheng, E., Wu, C.: Structure From Motion Using Structure-Less Resection. In: International Conference on Computer Vision (ICCV) (2015) 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 567, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 567, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 133, + 567, + 480, + 599 + ], + "type": "text", + "content": "142. Zhou, Q., Agostinho, S., Osep, A., Leal-Taixe, L.: Is Geometry Enough for Matching in Visual Localization? In: European Conference on Computer Vision (ECCV) (2022) 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 133, + 600, + 480, + 632 + ], + "type": "text", + "content": "143. Zurbrügg, R., Liu, Y., Engelmann, F., Kumar, S., Hutter, M., Patil, V., Yu, F.: ICGNet: A Unified Approach for Instance-Centric Grasping. In: International Conference on Robotics and Automation (ICRA) (2024) 2" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "type": "text", + "content": "SceneGraphLoc" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "23" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 22 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/SceneScript_ Reconstructing Scenes With An Autoregressive Structured Language Model/6d586386-d370-4aeb-a27d-9451b501b037_content_list.json b/2024/SceneScript_ Reconstructing Scenes With An Autoregressive Structured Language Model/6d586386-d370-4aeb-a27d-9451b501b037_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..092a86c76f8f8417ff704bb67cf36ce9ac6400bb --- /dev/null +++ b/2024/SceneScript_ Reconstructing Scenes With An Autoregressive Structured Language Model/6d586386-d370-4aeb-a27d-9451b501b037_content_list.json @@ -0,0 +1,1920 @@ +[ + { + "type": "text", + "text": "SceneScript: Reconstructing Scenes With An Autoregressive Structured Language Model", + "text_level": 1, + "bbox": [ + 243, + 77, + 758, + 121 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Armen Avetisyan1, Christopher Xie1, Henry Howard-Jenkins1, Tsun-Yi Yang1, Samir Aroudj1, Suvam Patra1, Fuyang Zhang2†, Duncan Frost1, Luke Holland1, Campbell Orme1, Jakob Engel1, Edward Miller1, Richard Newcombe1, and Vasileios Balntas1", + "bbox": [ + 218, + 133, + 782, + 193 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Meta Reality Labs Simon Fraser University https://projectaria.com/scenescript", + "bbox": [ + 328, + 205, + 671, + 234 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b5c2a70ef1294fc3992a7dc3fdcaccd62cf263fd13e44c1cc0c942d24d45769f.jpg", + "image_caption": [ + "Scene walkthrough" + ], + "image_footnote": [], + "bbox": [ + 217, + 250, + 367, + 316 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ef23cd8bcb4a5c2c1c1a1f60d8901c3cd6b9710c65b21064e5194dc91757e526.jpg", + "image_caption": [ + "#" + ], + "image_footnote": [], + "bbox": [ + 390, + 250, + 470, + 325 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0cb693aa5315d4bc1a3e0bf93e5a0e9a98bc6b360a4e1941ee206c96fa8328e4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 549, + 252, + 596, + 314 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b5e369cc3b6ed30fdeab00dfbfa8c06ad6193df51b8b6dcc2b5710698a0c791b.jpg", + "image_caption": [ + "Language to geometry interpreter" + ], + "image_footnote": [], + "bbox": [ + 635, + 251, + 784, + 315 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d53ad6538b00b5136035b7acfddfca2ec26dcb7d507edbd4acf542350e904719.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 215, + 334, + 328, + 388 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6cd6225ceb1d8f51b78851bed857cb49c69da71aa893bcc20e79194679c9c09e.jpg", + "image_caption": [ + "Fig. 1: (top) Given an egocentric video of an environment, SceneScript directly predicts a 3D scene representation consisting of structured scene language commands. (bottom) Our method generalizes on diverse real scenes while being solely trained on synthetic indoor environments. (last column, bottom) A notable advantage of our method is its capacity to easily adapt the structured language to represent novel scene entities. For example, by introducing a single new command, SceneScript can directly predict object parts jointly with the layout and bounding boxes." + ], + "image_footnote": [], + "bbox": [ + 217, + 391, + 328, + 446 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e6643aea7d8f149ea5b8d08e7e985345a028102d0b636820bb272dd4b1da4057.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 331, + 334, + 441, + 388 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/dd64e2665a6f9a6001da7faa4dcccd2412e132e6e9dc1d03409a84949dd48f23.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 331, + 391, + 442, + 446 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8c5baab5c2e2956dd9cfd5ea0a849af2e9c09c69c31da3c631169fae8ac4f47a.jpg", + "image_caption": [ + "Structured scene language" + ], + "image_footnote": [], + "bbox": [ + 446, + 334, + 555, + 388 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7d687e8dcf3da4e25121307cdc4481724f82426f8c04e4b66f0a19e81503f9f7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 444, + 391, + 555, + 446 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fc5afea0199592b657ee3cd9cddaf74bb01eb6309d2db95699ec84d92528365c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 558, + 334, + 669, + 388 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6f333f99c8811e8a9f9637f8934144af7678470a1981225f2cf9e8ff6d593fdc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 558, + 391, + 669, + 446 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4afa5e094cf97862c1ee0bcc43375ffece58e4f15bd994568b8095efb6bcbe1a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 673, + 334, + 784, + 388 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4fb9354133c22f600953d0c8f0537cf2a881486c6e2fb8b3a38bdb40cd53829e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 674, + 391, + 784, + 446 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. We introduce SceneScript, a method that directly produces full scene models as a sequence of structured language commands using an autoregressive, token-based approach. Our proposed scene representation is inspired by recent successes in transformers & LLMs, and departs from more traditional methods which commonly describe scenes as meshes, voxel grids, point clouds or radiance fields. Our method infers the set of structured language commands directly from encoded visual data using a scene language encoder-decoder architecture. To train SceneScript, we generate and release a large-scale synthetic dataset called Aria Synthetic Environments consisting of $100k$ high-quality indoor scenes, with photorealistic and ground-truth annotated renders of egocentric scene walkthroughs. Our method gives state-of-the-art results in architectural layout estimation, and competitive results in 3D object detection. Lastly, we explore an advantage for SceneScript, which is the ability to readily adapt to new commands via simple additions to the structured language, which we illustrate for tasks such as coarse 3D object part reconstruction.", + "bbox": [ + 259, + 582, + 740, + 818 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ Work done while the author was an intern at Meta.", + "bbox": [ + 230, + 824, + 583, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 143, + 375, + 161 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Scene representations play a crucial role in machine learning and computer vision applications, enabling accurate understanding of the environment. Over the years, researchers have explored various options such as meshes, voxel grids, point clouds, and implicit representations, aiming to represent complex real-world scenes with high-fidelity. Each of these exhibits distinct advantages and limitations that impact their suitability for different tasks. Meshes offer detailed geometric information but can be expensive in both computation and memory. Voxel grids provide a volumetric representation but suffer from a trade-off between resolution and memory requirements. Point clouds are efficient in representing sparse scenes but lack semantics and explicit connectivity information. Implicit representations, such as DeepSDF [22] and NeRF [15], can be infinitely precise but lack interpretability and editability. The selection of an appropriate representation directly impacts the performance and efficacy of various tasks including object recognition, scene understanding, and 3D reconstruction. In this paper, we propose a novel scene representation based on structured language commands as a more efficient and versatile solution.", + "bbox": [ + 212, + 175, + 787, + 416 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our motivation stems from the recent advancements in the field of Large Language Models (LLMs) and \"next token prediction\" autoregressive methods [20], coupled with recent works on exploring generation of sequences to represent geometric structures. For example, PolyGen [18] demonstrated the ability to describe 3D meshes as a sequence of vertices and faces generated using transformers [30]. Similarly, CAD-as-Language [11] showcased the effectiveness of generating Computer-Aided Design (CAD) primitives to represent 2D CAD sketches. Our main goal is to directly infer a metrically accurate representation of a full scene as a text-based sequence of specialized structured language commands.", + "bbox": [ + 212, + 417, + 787, + 566 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our method, denoted SceneScript, autoregressively predicts a language of hand-designed structured language commands in pure text form. This language offers several distinct advantages: 1) As pure text, it is compact and reduces memory requirements of a large scene to only a few bytes. 2) It is crisp and complete since the commands are designed to result in sharp and well-defined geometry (similar to scalable vector graphics). 3) It is interpretable, editable and semantically rich by design via the use of high-level parametric commands such as make_door(*door_parameters). 4) It can seamlessly integrate novel geometric entities by simply adding new structured commands to the language. 5) The fact that the scene representation is a series of language tokens similar to [20] opens up a plethora of potential new applications in the future such as ways to edit the scene, query it, or spin up chat interactions.", + "bbox": [ + 212, + 568, + 787, + 750 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We mainly focus on the problems of architectural layout estimation and object detection as proxy tasks for the efficacy of our SceneScript language as a scene representation. Architectural entities such as walls, doors, and windows are highly structured entities, making them an ideal test-bed. However, one notable drawback of language models is that they require vast amounts of data for training. Since there is no existing dataset of scene walkthroughs and their cor", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Avetisyan et al.", + "bbox": [ + 271, + 114, + 377, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "responding structured language commands, we publicly released Aria Synthetic Environments (ASE), a synthetically generated dataset of 100k unique interior scenes. For each scene, we simulate egocentric trajectories with an entire suite of sensor data from Project Aria [14]. We also release multiple sources of ground truth including depth and instance segmentations. Importantly, for each rendered egocentric sequence, the architectural layout ground truth is given in our proposed SceneScript language.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While architectural layout serves as a test-bed, we demonstrate that our method SceneScript can easily be extended to new tasks via simple extensions to SceneScript language while keeping both the visual input and network architecture fixed. We illustrate this on the problem of 3D object detection, which results in a method that jointly infers architectural layout and 3D oriented bounding boxes. Additionally, we demonstrate more proof-of-concept experiments that show that our method results in a significantly lower barrier to entry for new tasks including representing coarse 3D object reconstruction, curved entities, composition of entities, and entity states.", + "bbox": [ + 212, + 252, + 787, + 388 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our core contributions are:", + "bbox": [ + 240, + 388, + 437, + 402 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose SceneScript, a method that jointly predicts architectural layout and object bounding boxes of a scene in the form of structured language commands given a video stream.", + "- We demonstrate that SceneScript can easily be extended to completely new tasks with simple additions of commands to our structured language, significantly lowering the barrier to entry for new tasks.", + "- We release a large-scale synthetic dataset, named Aria Synthetic Environments, comprised of 100k unique high-quality 3D indoor scenes with GT, which will enable large scale ML training of scene understanding methods.", + "- We show that training SceneScript on Aria Synthetic Environments leads to generalization on real scenes (see videos/demos on the project page)." + ], + "bbox": [ + 225, + 412, + 784, + 579 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 601, + 395, + 617 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Layout Estimation", + "text_level": 1, + "bbox": [ + 215, + 633, + 415, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Layout estimation is an active research area, aiming to infer architectural elements. Scan2BIM [17] proposes heuristics for wall detection to produce 2D floorplans. Ochmann et al. [19] formulate layout inference as an integer linear program using constraints on detected walls. Shortest path algorithms around birds-eye view (BEV) free space [3] and wall plus room instance segmentation [5] have also been explored.", + "bbox": [ + 212, + 657, + 787, + 748 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Furukawa et al. [10] utilize a Manhattan world-based multi-view stereo algorithm [9] to merge axis-aligned depth maps into a full 3D mesh of building interiors. RoomNet [13] predicts layout keypoints while assuming that a fixed set of Manhattan layouts can occur in a single image of a room. LayoutNet [38] improves on this by predicting keypoints and optimising the Manhattan room layout inferred from them. Similarly, AtlantaNet [23] predicts a BEV floor or", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "SceneScript", + "bbox": [ + 651, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/d2a05526a400137c7748c0408ef09184df0a27647e68c51e1f51e918380c90f2.jpg", + "table_caption": [ + "Table 1: Complete set of structured language commands designed for detailing architectural layouts and object bounding boxes. Supported data types can include int, float, bool. It is important to note that the language's extensibility allows for easy augmentation by introducing new commands like make_prime, make_pillar, or enhancing existing commands, such as incorporating is-double_door (bool)." + ], + "table_footnote": [], + "table_body": "
make_WALL (int)make_door (int)make_window (int)make_bbox (int)
idintidintidintidint
a_xfloatwall0_idintwall0_idintclassint
a_yfloatwall1_idintwall1_idintposition_xfloat
a_zfloatposition_xfloatposition_xfloatposition_yfloat
b_xfloatposition_yfloatposition_yfloatposition_zfloat
b_yfloatposition_zfloatposition_zfloatangle_zfloat
b_zfloatwidthfloatwidthfloatscale_xfloat
heightfloatheightfloatheightfloatscale_yfloat
scale_zfloat
", + "bbox": [ + 287, + 226, + 705, + 333 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ceiling shape and approximates the shape contour with a polygon resulting in an Atlanta world prior. SceneCAD [1] uses a graph neural network to predict object-object and object-query relationships to refine its layout prediction.", + "bbox": [ + 212, + 364, + 784, + 411 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our approach stands out by requiring neither heuristics nor explicitly defined prior knowledge about architectural scene structure. In fact, our method demonstrates geometric understanding of the scene which emerges despite learning to predict the GT scene language as a sequence of text tokens.", + "bbox": [ + 212, + 411, + 785, + 472 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Geometric Sequence Modelling", + "text_level": 1, + "bbox": [ + 214, + 502, + 519, + 518 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recent works have explored transformers for generating objects as text-based sequences. PolyGen [18] models 3D meshes as a sequence of vertices and faces. CAD-as-Language [11] represents 2D CAD sketches as a sequence of triplets in protobuf representation, followed by a sequence of constraints. Both SketchGen [21] and SkexGen [33] use transformers to generate sketches. DeepSVG [4] learns a transformer-based variational autoencoder (VAE) that is capable of generating and interpolating through 2D vector graphics images. DeepCAD [32] proposes a low-level language and architecture similarly to DeepSVG, but applies it to 3D CAD models instead of 2D vector graphics. Our approach stands out by utilising high-level commands, offering interpretability and semantic richness. Additionally, while low-level commands can represent arbitrarily complex geometries, they lead to prohibitively longer sequences when representing a full scene.", + "bbox": [ + 212, + 536, + 787, + 731 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The closest work to ours is Pix2Seq [6]. Pix2Seq proposes a similar architecture to ours but experiments only with 2D object detection, thus requiring domain-specific augmentation strategies. Another closely related work is Point2Seq [34] that trains a recurrent network for autoregressively regressing continuous 3D bounding box parameters. Interestingly, they find the autoregressive ordering of parameters outperforms current standards for object detection architectures, including anchors [12] and centers [36].", + "bbox": [ + 212, + 734, + 787, + 842 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Avetisyan et al.", + "bbox": [ + 271, + 114, + 377, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1a1d1b552808797c3bc3fddf81646e7d68331207ca166790f8def6c7738630dc.jpg", + "image_caption": [ + "Fig. 2: Aria Synthetic Environments: (top) Random samples of generated scenes showing diversity of layouts, lights and object placements. (bottom - left to right) A top down view of a scene filled with objects, a simulated trajectory (blue path), renderings of depth, RGB, and object instances, and lastly a scene pointcloud." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 345 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 SceneScript Structured Language Commands", + "text_level": 1, + "bbox": [ + 215, + 439, + 687, + 458 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first describe our structured language commands that define a full scene representation including both layout and objects. After this, we introduce our corresponding large scale training dataset: Aria Synthetic Environments.", + "bbox": [ + 212, + 472, + 787, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Commands and Parameters", + "text_level": 1, + "bbox": [ + 215, + 541, + 491, + 555 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We begin with a parameterization that captures the most common layout elements. For this purpose we use three commands: make-wall, make_door, and make_window. Each command comes with a set of parameters that results in well-defined geometry. For example, the full set of parameters for make-wall specifies a gravity-aligned 2D plane, while make_door and make_window specify box-based cutouts from walls. It is worth noting that this parametrization is arbitrary, and is only made in the context of presenting a proof-of-concept SceneScript system. There are infinitely many parametrization schemes, in this work we opt for one that prioritizes ease of use and research iteration speed.", + "bbox": [ + 212, + 566, + 787, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In addition to representing these three major layout entities, we aim to jointly infer objects as oriented bounding boxes. Thus, we introduce a fourth command:", + "bbox": [ + 212, + 703, + 785, + 733 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [], + "code_body": "make_bbox: id, class, position_x, position_y, position_z, angle_z, scale_x, scale_y, scale_z", + "guess_lang": "python", + "bbox": [ + 215, + 742, + 723, + 772 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This simple parametrization represents an oriented 3D bounding box that is assumed to be aligned with gravity (assuming it points in the $-z$ direction). A summary of these commands and their respective parameters are shown in Table 1.", + "bbox": [ + 212, + 779, + 787, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "SceneScript", + "bbox": [ + 651, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While we have described just four commands to capture structure and objects in an indoor environment, importantly, this text-based parametrization can readily be extended geometrically and/or even semantically to include states or other functional aspects. For example, changing the mentioned makedoor command to include parameters such as open-degree allows the language to represent door states. In Section 6, we demonstrate how such extensions to the language allows for SceneScript to readily adapt to new tasks including coarse 3D object reconstruction.", + "bbox": [ + 212, + 146, + 787, + 268 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Scene Definition", + "text_level": 1, + "bbox": [ + 214, + 291, + 398, + 305 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A single scene can be described as a sequence of our proposed structured language commands. The sequence requires no specific ordering, and can be of arbitrary length. From this definition, the 3D scene can easily be obtained by parsing the commands through a simple custom interpreter.", + "bbox": [ + 212, + 316, + 785, + 380 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Training Dataset", + "text_level": 1, + "bbox": [ + 214, + 402, + 401, + 417 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To enable practical indoor scene reconstruction based on our structured language commands (Section 3.1), we publicly released a new dataset called Aria Synthetic Environments. It consists of a large number of training pairs of egocentric scene walkthroughs linked with corresponding ground truth command sequences.", + "bbox": [ + 212, + 428, + 785, + 489 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Since transformers require vast amounts of data, we generated $100k$ synthetic scenes, which is in comparison infeasible for real-world data. Each synthetic scene comes with a floor plan model, a corresponding complete 3D scene as well as a simulated agent trajectory and a photo-realistic rendering of this trajectory. Fig. 2 illustrates the basics of Aria Synthetic Environments. For brevity, we refer the reader to the supplemental material for further details.", + "bbox": [ + 212, + 489, + 787, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 SceneScript Network Architecture", + "text_level": 1, + "bbox": [ + 214, + 604, + 583, + 622 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our pipeline is a simple encoder-decoder architecture that consumes a video sequence and returns SceneScript language in a tokenized format. Figure 3 illustrates a high-level overview of our method.", + "bbox": [ + 212, + 638, + 785, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We examine three encoder variants: a pointcloud encoder, a posed image set encoder, and a combined encoder. The decoder remains the same in all cases.", + "bbox": [ + 212, + 684, + 785, + 714 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1 Input Modalities and Encoders", + "text_level": 1, + "bbox": [ + 214, + 737, + 519, + 753 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The encoder computes a latent scene code in the form of a 1D sequence from video walkthrough of the scene. The decoder is designed to consume these 1D sequences as input. This enables the integration of various input modalities within a unified framework. As a preliminary, for each scene we assume access to a set of $M$ posed camera images $\\{\\mathbf{I}_1,\\dots,\\mathbf{I}_{\\mathbf{M}}\\}$ , e.g., SLAM output.", + "bbox": [ + 212, + 763, + 785, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Avetisyan et al.", + "bbox": [ + 271, + 114, + 377, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c3b7e7ec3fe381af7c3e1f2cad895f41c4683d131632eaee1af83a60648d7867.jpg", + "image_caption": [ + "Fig. 3: SceneScript core pipeline overview. Raw images & pointcloud data are encoded into a latent code, which is then autoregressively decoded into a sequence of commands that describe the scene. Visualizations are shown using a customly built interpreter. Note that for the results in this paper, the the point clouds are computed from the images using Aria MPS [25] - i.e. are not using a dedicated RGB-D / Lidar sensor." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 529, + 289 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2dea9288ccf974cbfed2a66b1452436a7926daa4a216cbdccb1de8553d1dd53c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 532, + 142, + 785, + 290 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Point Clouds. A point cloud $\\mathbf{P} = \\{\\mathbf{p}_1, \\dots, \\mathbf{p_N}\\}$ consists of $N$ points, where $\\mathbf{p}_i$ is a 3D point coordinate. It can come from passive images using SLAM or SfM, or RGB-D / Lidar sensors.", + "bbox": [ + 212, + 392, + 782, + 436 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Specifically, we use the Semi-dense Pointclouds from Project Aria's Machine Perception Services [25], that are obtained from a visual-inertial SLAM system using Aria's monochrome cameras and IMUs. We discretize the point cloud to $5\\mathrm{cm}$ resolution, then employ a sparse 3D convolution library [26,27] to generate pooled features. The encoder $\\mathcal{E}_{geo}$ applies a series of down convolutions, resulting in a reduction of the number of points in the lowest level.", + "bbox": [ + 212, + 436, + 784, + 527 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {g e o} = \\mathcal {E} _ {g e o} (\\mathbf {P}), \\quad \\mathbf {P} \\in \\mathbb {R} ^ {N \\times 3}, \\mathbf {F} _ {g e o} \\in \\mathbb {R} ^ {K \\times 5 1 2} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 336, + 541, + 784, + 560 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $K \\ll N$ . $\\mathbf{F}_{geo}$ is a condensed latent representation of the point cloud that contains the necessary scene context. For later use in the transformer decoder, we treat $\\mathbf{F}_{geo}$ as a sequence of feature vectors where the entries $\\mathbf{f_i}$ , $i \\in 1\\dots K$ are sorted lexicographically according to the coordinate of the active site $\\mathbf{c_i}$ , $i \\in 1\\dots K$ . To incorporate positional encoding, we append the coordinates of the active sites to their respective feature vectors $\\mathbf{f}_i \\gets \\mathbf{cat}(\\mathbf{f_i},\\mathbf{c_i})$ .", + "bbox": [ + 212, + 565, + 787, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Point Clouds with Lifted Features. We additionally explore augmenting the point cloud with image features. From the original egocentric sequence and associated trajectory, we sample a set of $M$ keyframes, $\\mathbf{I}_i$ where $i\\in 1\\dots M$ , and compute a set of image features $\\mathbf{F}_i$ for each. We then project each point into the set of keyframe cameras and retrieve the feature vector (output by a CNN) at the pixel location:", + "bbox": [ + 212, + 678, + 787, + 770 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {i p} = F _ {i} (\\pi (\\mathbf {p})) \\quad \\mathbf {p} \\in \\mathbf {P}, i \\in 1 \\dots M, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 369, + 782, + 784, + 799 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\pi(\\cdot)$ represents the projection function of a 3D point into the camera. If $\\pi(\\mathbf{p})$ falls outside the image bounds, no feature is retrieved. We combine the set", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "SceneScript", + "bbox": [ + 651, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "of lifted features for a point through an average, resulting in a single feature vector for each point: $\\mathbf{f}_p = 1 / M\\sum_{i = 1}^{M}\\mathbf{f}_{ip}$", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We form our lifted-feature point cloud, $\\mathbf{P}' = \\{\\mathbf{p}_1',\\dots,\\mathbf{p}_{\\mathbf{M}}'\\}$ , by concatenating each point's lifted feature with the original XYZ location: $\\mathbf{p}' = \\mathrm{cat}(\\mathbf{f}_p,\\mathbf{p})$ . $\\mathbf{P}'$ is then encoded into a context sequence using sparse 3D convolutions, with only the input feature channels adjusted to match the new point feature dimension.", + "bbox": [ + 212, + 176, + 784, + 238 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "End-to-end Encoding of Posed Views. In order to encode the egocentric sequence more directly without a pre-computed point cloud, we adopt a $2\\mathrm{D} \\leftrightarrow 3\\mathrm{D}$ bidirectional transformer encoder following the form defined in RayTran [29].", + "bbox": [ + 212, + 262, + 782, + 306 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this formulation, we initialize a volume representation of the scene as a dense voxel grid of features, $\\mathbf{V}$ , that coincides with the scene geometry. In turn, we sample a subset of $M$ keyframes, $\\mathbf{I}_i$ where $i\\in 1\\dots M$ , from the full stream of posed images. And for each of these keyframes we compute image features from a CNN, $\\mathbf{F}_i$ . Repeated layers of bidirectional attention enable the image and voxel grid features to be refined iteratively in successive transformer blocks through the aggregation of view-point and global-scene information. As in RayTran [29], the interaction between the two representations is guided by the image-formation process by leveraging known camera parameters and poses. Attention in these transformer blocks is restricted by patch-voxel ray intersections, where each image patch attends to the voxels it observes and each voxel location attends to all the patches that observe it. The resulting voxel grid of features is flattened, concatenated with an encoded representation of its XYZ location, and passed to the decoder.", + "bbox": [ + 212, + 309, + 784, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 Language Decoder", + "text_level": 1, + "bbox": [ + 214, + 544, + 415, + 559 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We utilize a transformer decoder [30] to decode the scene latent code into a sequence of structured language commands. The sequence of tokens passes through an embedding layer, followed by a positional encoding layer. Together with the encoded scene code (Section 4.1), the embedded tokens are passed into the several transformer decoder layers where a causal attention mask is used to ensure autoregressive generation. More implementation details can be found in Appendix ??", + "bbox": [ + 212, + 571, + 782, + 676 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3 Language Tokenization", + "text_level": 1, + "bbox": [ + 214, + 703, + 450, + 718 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We refer to the serialization the structured language into a sequence of tokens as tokenization. The goal is to construct a bijective mapping between a sequence of structured language commands (Section 3) and a sequence of integer tokens that can be predicted by the transformer decoder architecture. We utilize the following schema:", + "bbox": [ + 212, + 731, + 782, + 806 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Avetisyan et al.", + "bbox": [ + 271, + 114, + 377, + 128 + ], + "page_idx": 7 + }, + { + "type": "footer", + "text": "[START, PART, CMD, PARAM_1, PARAM_2, ..., PARAM_N, PART, ..., STOP]", + "bbox": [ + 263, + 819, + 736, + 835 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For example, a sample sequence for a makedoor command may look like:", + "bbox": [ + 240, + 146, + 782, + 161 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "[START, PART, MAKE_DOOR, POSITION_X, POSITION_Y, POSITION_Z, WALLO_IDX, WALL1_IDX, WIDTH, HEIGHT, PART,..., STOP]", + "bbox": [ + 269, + 189, + 730, + 223 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This schema enables 1D packing of tokens without requiring fixed-size slots or padding like other sequence modelling methods such as [32]. Additionally, it does not impose any limitations on the number or the hierarchy of sub-sequences, as they are flexibly separated by a PART token. This allows for arbitrarily complex scene representations.", + "bbox": [ + 212, + 237, + 782, + 311 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The tokenized sequence is discretized into integers at a 5cm resolution, then translated into a sequence of embeddings via learnable lookup table. Note that by designing the SceneScript language, we also design the tokenization. This tokenization scheme is notably different from standard NLP tokenization, which involves Byte-Pair Encodings (BPE) [20].", + "bbox": [ + 212, + 313, + 782, + 388 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 Results", + "text_level": 1, + "bbox": [ + 215, + 412, + 323, + 428 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this section, we introduce the metrics we define to measure performance, and we discuss some qualitative and quantitative results that give insights to our proposed SceneScript method.", + "bbox": [ + 212, + 446, + 782, + 492 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.1 Metrics", + "text_level": 1, + "bbox": [ + 215, + 516, + 325, + 530 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To evaluate accuracy of the layout estimation, we make use of geometric metrics applied between the ground-truth room layout and our predicted SceneScript language. To do so, we define an entity distance, $d_{E}$ , between a pair of entities of the same class. Each entity, $E$ , is represented as a 3D plane segment comprising of 4 corners $\\{c_1, c_2, c_3, c_4\\}$ . The distance between two entities, $E$ and $E'$ is computed as the maximum Euclidean distance between each corner and its counterpart assigned via Hungarian mathcing, i.e.: $d_{E}(E, E') = \\max \\{||c_{i} - c_{\\pi(i)}'| : i = 1, \\dots, 4\\}$ , where $\\pi(i)$ is the permutation also found via Hungarian matching.", + "bbox": [ + 212, + 542, + 795, + 665 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We threshold $d_{E}$ to define the success criteria for the predicted entities. We compute the following metrics:", + "bbox": [ + 212, + 666, + 782, + 696 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- F1 Score @ threshold - the F1 score of the set of predictions is computed at a single $d_{E}$ threshold.", + "- Average F1 Score - the F1 score is computed across a range of entity distance thresholds and averaged." + ], + "bbox": [ + 223, + 707, + 782, + 767 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The scores are computed for each class independently and averaged to overall score. In addition, scores are computed for each scene and averaged across the dataset. We use the following range of thresholds (cm) for the average F1 scores: $T = \\{1, 2, \\dots, 9, 10, 15, 25, 30, 50, 75, 100\\}$ .", + "bbox": [ + 212, + 779, + 782, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "SceneScript", + "bbox": [ + 651, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/9c35d5d28bc1dbf8b979708ca1fbc4ae07d8432fc5235dbeec91b9e95e5ec278.jpg", + "image_caption": [ + "Fig. 4: Qualitative samples between our model and SOTA methods on Aria Synthetic Environments's test set. Hierarchical methods like SceneCAD suffer from error cascading which leads to missing elements in the edge prediction module. RoomFormer (a 2D method extruded to 3D) primarily suffers from lightly captured scene regions which leave a unnoticeable signal in the density map." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 359 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/ec7df5318e3f0c045a7474c26ea2953e38c0a2c51fea0eface63a8472b5447b9.jpg", + "table_caption": [ + "Table 2: Layout Estimation on Aria Synthetic Environments Quantitative comparison between our method and related recent work." + ], + "table_footnote": [], + "table_body": "
MethodF1 @5cmAvg F1
meanwalldoorwindowmeanwalldoorwindow
SceneCAD '20 [1]-0.048---0.275--
RoomFormer '23 [37]0.1390.1590.1480.1100.4640.5050.4810.407
Ours (Point cloud)0.8480.9300.9220.6920.7840.8160.8110.724
Ours (Lifted features)0.9030.9430.9590.8060.8010.8180.8220.764
Ours (Image-only)0.6610.6870.7980.4970.7190.7270.7720.658
", + "bbox": [ + 235, + 494, + 764, + 602 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.2 Layout Estimation", + "text_level": 1, + "bbox": [ + 215, + 628, + 415, + 643 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We perform scene layout estimation with the three encoder variants of SceneScript: a baseline model with sparse3d convolution pointcloud encoder, an RGB RayTran-based feature volume encoder [29], and our proposed lifted feature point encoder. The same transformer decoder is being used in all three scenarios.", + "bbox": [ + 212, + 657, + 787, + 732 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To provide comparison to existing works, we include results from two baseline methods, namely SceneCAD [1] and the recent RoomFormer [37]. For these experiments SceneCAD and RoomFormer were both trained on Aria Synthetic Environments. Note that SceneCAD only predicts walls.", + "bbox": [ + 212, + 733, + 787, + 794 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 2 shows the main results for our F1-based metrics on Aria Synthetic Environments. SceneScript exhibits a substantial performance advantage over SOTA layout estimation baselines across multiple metrics.", + "bbox": [ + 212, + 794, + 787, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Avetisyan et al.", + "bbox": [ + 271, + 114, + 377, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/4555838a8a76fdb99217d3189f976671d5f288e512ab2ba4226e8e7225e663a4.jpg", + "table_caption": [ + "Table 3: 3D Object Detection Performance comparison against state-of-the-art methods on an 3D object detection task trained and evaluated by F1-score at 0.25 and 0.5 IoU thresholds. By simply adding a make_bbox command SceneScript can achieve competitive object detection results.", + "(a) Aria Synthetic Environments" + ], + "table_footnote": [], + "table_body": "
MethodInputF1@.25 IoU @.50 IoU
3DETR '21 [16]Points0.2010.078
Cube R-CNN '23 [2]RGB0.3940.228
ImVoxelNet '22 [24]RGB0.5840.516
OursPoints0.6200.577
", + "bbox": [ + 225, + 247, + 496, + 325 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/7e5865182b493beef94be74a1d64d964e9a68fb0200d097b6dfed937c38ccedc.jpg", + "table_caption": [ + "(b) ScanNet [8]" + ], + "table_footnote": [], + "table_body": "
MethodInputF1
@.25 IoU@.50 IoU
3DETR '21 [16]Points0.4800.349
3DETR-m '21 [16]Points0.5360.407
SoftGroup '22 [31]RGB Points0.6220.573
OursRGB Points0.5060.406
", + "bbox": [ + 501, + 247, + 774, + 320 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Both baseline methods encounter a significant decline in accuracy when dealing with finer details. See Figure 4 for qualitative comparisons between our method and baseline methods.", + "bbox": [ + 212, + 354, + 782, + 400 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Encoder Ablation. The results demonstrate that SceneScript is robust to the encoding strategy chosen to encode the egocentric capture. It is able to infer highly accurate scene layouts in all configurations tested, and in each case SceneScript outperforms the included baselines by a significant margin.", + "bbox": [ + 212, + 425, + 784, + 484 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Relative comparison of the encoder strategies reveals that leveraging the point clouds from a highly specialized and well-tuned system is still able to offer advantages of an, almost, entirely learned approach such as RayTran [29]. A light extension in the form of lifted image features can widen this gap even further. In particular, we observe that the discrepancy between the encoders becomes particularly apparent as the complexity of the scene increases in the form of increased room count. In the Appendix, we also show a quantitative evaluation of per-entity error distances, which aids in further attribution of relative performance gains between the encoding methods.", + "bbox": [ + 212, + 486, + 787, + 622 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.3 Object Detection", + "text_level": 1, + "bbox": [ + 214, + 646, + 405, + 661 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In this section, we perform evaluation of SceneScript for object detection on both Aria Synthetic Environments and ScanNet [8]. For comparison, we include recent and state-of-the-art baseline methods, namely Cube-RCNN [2], ImVoxelNet [24], 3DETR [16], and SoftGroup [31].", + "bbox": [ + 212, + 672, + 785, + 733 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Worth noting is that SceneScript does not predict a confidence score for each make_bbox command. This means that fair computation of the conventional mAP metric for this task is not possible, as detections cannot be ranked across scenes. Among other issues, this results in a metric that varies with the order in which scenes are evaluated. We therefore report F1-score-based metrics, which do not exhibit this order variance. Further discussion of this, and mAP numbers for the baselines for reference, can be found in the Appendix.", + "bbox": [ + 212, + 734, + 787, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "SceneScript", + "bbox": [ + 651, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/814bbad6fa1cc611ca4aec168804d9a106aba6b1f4305b699bb40cb1ea365dc7.jpg", + "image_caption": [ + "Fig. 5: Example scene reconstructions on scenes from Aria Synthetic Environments. (left) Visualisation of the decomposed meshes used to create make_prim training pairs. (right) Views of full scene predictions, as well as close ups highlighting the fidelity of object reconstruction through the prediction volumetric primitives enabled by make_prim." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 330, + 266 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/3f9db45871fddf264da288563194d4b18fcf1c128a19683f51f9876000c4d842.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 331, + 143, + 787, + 267 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In Table 3a, all methods were trained on Aria Synthetic Environments. 3DETR performs poorly due to its encoder being designed for relatively clean datasets such as ScanNet [8], while semi-dense point clouds from Aria Synthetic Environments exhibit more noise and less uniformity (see Figure 3 for an example). Cube R-CNN and ImVoxelNet both operate on RGB images, and detections are tracked for the entire sequence via a tracker [2] to provide competitive performance. In Table 3b, our method provides similar performance to both 3DETR and SoftGroup.", + "bbox": [ + 212, + 364, + 787, + 484 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Through the addition of the make_bbox command, SceneScript demonstrates object detection performance on par with SOTA baselines on multiple object detection datasets. This result illustrates the extensibility of a token-based approach, and that our proposed SceneScript language representation does not suffer compared to specialised object detection network architectures.", + "bbox": [ + 212, + 486, + 787, + 561 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "6 Extending the SceneScript Structured Language", + "text_level": 1, + "bbox": [ + 214, + 587, + 718, + 606 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "A key advantage offered by SceneScript's structured language prediction paradigm is that the expressiveness of its reconstruction can be tailored without requiring a change to the method. Up to now, we have focussed on showcasing the efficacy of SceneScript for representing simple layout elements and objects as bounding boxes. In this section, we showcase this characteristic by increasing the fidelity of our scene representation by introducing coarse 3D object reconstruction.", + "bbox": [ + 212, + 621, + 803, + 712 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "6.1 Objects as Volumetric Primitives", + "text_level": 1, + "bbox": [ + 214, + 736, + 535, + 752 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We turn to a language based on volumetric primitives, motivated by works such as [28, 35]. Using simple primitives such as cuboids and extruded cylinders, enables us to coarsely represent arbitrary object categories while maintaining object semantics (e.g. tabletops can be represented by a single cuboid). Thus, this language can describe many object categories simultaneously.", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Avetisyan et al.", + "bbox": [ + 271, + 114, + 377, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/07e2ba398b0e28618caf4ace0b9efe68e8dc5331637f65fc02e672324a425a48.jpg", + "image_caption": [ + "Fig.6: Example scene reconstructions on real scenes with the addition of the make_prim command. Note that SceneScript was trained only on synthetic data." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 281 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "This representation requires only one additional command over the layout and box commands already discussed previously, namely:", + "bbox": [ + 212, + 349, + 782, + 380 + ], + "page_idx": 12 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [], + "code_body": "make_prime:_bbox_id, prim_num, class, center_x, center_y, center_z, angle_x, angle_y, angle_z, scale_x, scale_y, scale_z", + "guess_lang": "txt", + "bbox": [ + 215, + 388, + 781, + 434 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "This command and its parameters describe a volumetric primitive (cuboid or extruded cylinder) via its 3D center, 3D rotation, and 3D scale. The prim_num parameter can be associated with semantics, e.g. tabletops of different tables typically have the same prim_num.", + "bbox": [ + 212, + 441, + 787, + 502 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Dataset. To obtain ground truth make_prim commands that align with the objects in Aria Synthetic Environments, we first run an extension of Yang et al. [35] to obtain cuboid and extruded cylinder primitives of a database of 3D CAD models (ABO [7], which was used to populate Aria Synthetic Environments). See Figure 5 (left) for example decompositions. For this proof-of-concept experiment, we use three categories: table, chair, and sofa. We then convert these decomposed primitives into make_prim commands that are aligned with the objects in the dataset, which results in training pairs.", + "bbox": [ + 212, + 523, + 787, + 647 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Results. We show qualitative results on Aria Synthetic Environments in Figure 5. In Figure 6, we show inferences in a few real-world environments despite only having trained on our simulated dataset. These results demonstrate that SceneScript's general purpose architecture is capable of coarsely reconstructing objects of multiple categories through addition of a new command.", + "bbox": [ + 212, + 669, + 787, + 746 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "6.2 Further Extensions", + "text_level": 1, + "bbox": [ + 214, + 768, + 419, + 782 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In this section, we have explored just one extension of SceneScript's structured scene language in order to demonstrate its flexibility. The result was greatly increased expressiveness of the scene reconstruction. In the Appendix, we include", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "SceneScript", + "bbox": [ + 651, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "additional explorations that can further increase the fidelity and accuracy of the reconstructions. These explorations include reconstructing curved walls, inferring object states (e.g., door opening angles), as well as direct prediction of parametric models using object models deployed commonly by tech artists (e.g., blender geometry nodes) for object reconstruction.", + "bbox": [ + 212, + 146, + 787, + 223 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "7 Interactive Scene Reconstruction", + "text_level": 1, + "bbox": [ + 215, + 243, + 571, + 258 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Scene reconstruction often occurs offline, on pre-recorded walkthroughs of an environment, or derivatives of them, such as a point cloud. However, we take advantage of SceneScript's inference occurring at an interactive rate, which takes between 2-5s depending on the size of the environment, by implementing streaming of SceneScript's live reconstructions into a modern VR headset. This enables an interactive experience where the user can see the reconstruction overlaid onto the environment they are exploring in real-time. See the video recording on the website for examples. Visualisations of this interface are included in Figure 1.", + "bbox": [ + 212, + 271, + 787, + 409 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "8 Limitations and Future Work", + "text_level": 1, + "bbox": [ + 215, + 429, + 537, + 445 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "SceneScript exhibits certain limitations that should be acknowledged. First, the structured language commands are manually defined, which requires human intervention at this stage. Secondly, due to the higher-level nature of our commands, it can be challenging to capture fine-grained geometric details with extremely high precision (i.e. mm). As a consequence, the resulting reconstructions based on this representation tend to lead to simpler and coarser geometries, potentially missing intricate nuances at the very high detail level. These limitations potentially highlight areas for future research and optimization, aiming to automate the command definition process and explore techniques to enhance the representation's ability to capture intricate geometric details accurately. However, we believe that the ability to built scene representations that are based on structured language commands will be a key part in complex and efficient methods of scene reconstruction in the future.", + "bbox": [ + 212, + 457, + 787, + 654 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "9 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 674, + 359, + 690 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We introduced SceneScript, a novel reconstruction method that is based on a tokenized scene representation. SceneScript autoregressively predicts a sequence of structured scene language commands given a video stream of an indoor space. This tokenized scene representation is compact, editable and interpretable. In addition, we showed that a strength of SceneScript is the ability to extend to arbitrary elements (e.g. Bezier curves, object part decomposition) with minimal changes. This research opens up new directions in representing 3D scenes as language, bringing the 3D reconstruction community closer to the recent advances in large language models such as GPT-class models [20].", + "bbox": [ + 212, + 704, + 787, + 842 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Avetisyan et al.", + "bbox": [ + 271, + 114, + 377, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Avetisyan, A., Khanova, T., Choy, C., Dash, D., Dai, A., Nießner, M.: Scenecad: Predicting object alignments and layouts in rgb-d scans. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16. pp. 596-612. Springer (2020)", + "2. Brazil, G., Kumar, A., Straub, J., Ravi, N., Johnson, J., Gkioxari, G.: Omni3D: A large benchmark and model for 3D object detection in the wild. In: CVPR. IEEE, Vancouver, Canada (June 2023)", + "3. Cabral, R., Furukawa, Y.: Piecewise planar and compact floorplan reconstruction from images. In: 2014 IEEE Conference on Computer Vision and Pattern Recognition (2014)", + "4. Carlier, A., Danelljan, M., Alahi, A., Timofte, R.: Deepsvg: A hierarchical generative network for vector graphics animation. In: Advances in Neural Information Processing Systems (2020)", + "5. Chen, J., Liu, C., Wu, J., Furukawa, Y.: Floor-sp: Inverse cad for floorplans by sequential room-wise shortest path. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2019)", + "6. Chen, T., Saxena, S., Li, L., Fleet, D.J., Hinton, G.: Pix2seq: A language modeling framework for object detection. In: International Conference on Learning Representations (ICLR) (2022)", + "7. Collins, J., Goel, S., Deng, K., Luthra, A., Xu, L., Gundogdu, E., Zhang, X., Yago Vicente, T.F., Dideriksen, T., Arora, H., Guillaumin, M., Malik, J.: Abo: Dataset and benchmarks for real-world 3d object understanding. CVPR (2022)", + "8. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proc. Computer Vision and Pattern Recognition (CVPR), IEEE (2017)", + "9. Furukawa, Y., Curless, B., Seitz, S.M., Szeliski, R.: Manhattan-world stereo. In: 2009 IEEE Conference on Computer Vision and Pattern Recognition (2009)", + "0. Furukawa, Y., Curless, B., Seitz, S.M., Szeliski, R.: Reconstructing building interiors from images. In: 2009 IEEE 12th international conference on computer vision (2009)", + "1. Ganin, Y., Bartunov, S., Li, Y., Keller, E., Saliceti, S.: Computer-aided design as language. Advances in Neural Information Processing Systems 34, 5885-5897 (2021)", + "2. Girshick, R.: Fast r-cnn. In: Proceedings of the IEEE international conference on computer vision. pp. 1440-1448 (2015)", + "3. Lee, C.Y., Badrinarayanan, V., Malisiewicz, T., Rabinovich, A.: Roomnet: End-to-end room layout estimation. In: Proceedings of the IEEE international conference on computer vision (2017)", + "4. Meta: Project aria. https://projectaria.com/ (2022), accessed: 2023-08-30", + "5. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)", + "6. Misra, I., Girdhar, R., Joulin, A.: An end-to-end transformer model for 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2906-2917 (2021)", + "7. Murali, S., Speciale, P., Oswald, M.R., Pollefeys, M.: Indoor scan2bim: Building information models of house interiors. In: 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) (2017)" + ], + "bbox": [ + 225, + 176, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "SceneScript", + "bbox": [ + 651, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Nash, C., Ganin, Y., Eslami, S.A., Battaglia, P.: Polygen: An autoregressive generative model of 3d meshes. In: International conference on machine learning. pp. 7220-7229. PMLR (2020)", + "19. Ochmann, S., Vock, R., Klein, R.: Automatic reconstruction of fully volumetric 3d building models from oriented point clouds. ISPRS journal of photogrammetry and remote sensing 151, 251-262 (2019)", + "20. OpenAI: Gpt-4 technical report (2023)", + "21. Para, W., Bhat, S., Guerrero, P., Kelly, T., Mitra, N., Guibas, L.J., Wonka, P.: Sketchgen: Generating constrained cad sketches. In: Advances in Neural Information Processing Systems (2021)", + "22. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2019)", + "23. Pintore, G., Agus, M., Gobbetti, E.: Atlantanet: inferring the 3d indoor layout from a single 360 image beyond the Manhattan world assumption. In: Proceedings of the European conference on computer vision (ECCV) (2020)", + "24. Rukhovich, D., Vorontsova, A., Konushin, A.: Imvoxelnet: Image to voxels projection for monocular and multi-view general-purpose 3d object detection. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2397-2406 (2022)", + "25. Somasundaram, K., Dong, J., Tang, H., Straub, J., Yan, M., Goesele, M., Engel, J.J., De Nardi, R., Newcombe, R.: Project aria: A new tool for egocentric multimodal ai research. arXiv preprint arXiv:2308.13561 (2023)", + "26. Tang, H., Liu, Z., Li, X., Lin, Y., Han, S.: TorchSparse: Efficient Point Cloud Inference Engine. In: Conference on Machine Learning and Systems (MLSys) (2022)", + "27. Tang, H., Liu, Z., Zhao, S., Lin, Y., Lin, J., Wang, H., Han, S.: Searching Efficient 3D Architectures with Sparse Point-Voxel Convolution. In: European Conference on Computer Vision (ECCV) (2020)", + "28. Tulsiani, S., Su, H., Guibas, L.J., Efros, A.A., Malik, J.: Learning shape abstractions by assembling volumetric primitives. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2017)", + "29. Tyszkiiewicz, M.J., Maninis, K.K., Popov, S., Ferrari, V.: Raytran: 3d pose estimation and shape reconstruction of multiple objects from videos with ray-traced transformers. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part X. pp. 211-228. Springer (2022)", + "30. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. NeurIPS 30 (2017)", + "31. Vu, T., Kim, K., Luu, T.M., Nguyen, X.T., Yoo, C.D.: Softgroup for 3d instance segmentation on 3d point clouds. In: CVPR (2022)", + "32. Wu, R., Xiao, C., Zheng, C.: Deepcad: A deep generative network for computer-aided design models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2021)", + "33. Xu, X., Willis, K.D., Lambourne, J.G., Cheng, C.Y., Jayaraman, P.K., Furukawa, Y.: Skexgen: Autoregressive generation of cad construction sequences with disentangled codebooks. In: International Conference on Machine Learning (ICML) (2022)", + "34. Xue, Y., Mao, J., Niu, M., Xu, H., Mi, M.B., Zhang, W., Wang, X., Wang, X.: Point2seq: Detecting 3d objects as sequences. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8521-8530 (2022)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Avetisyan et al.", + "bbox": [ + 271, + 114, + 375, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "35. Yang, K., Chen, X.: Unsupervised learning for cuboid shape abstraction via joint segmentation from point clouds. ACM Transactions on Graphics (TOG) 40(4), 1-11 (2021)", + "36. Yin, T., Zhou, X., Krahenbuhl, P.: Center-based 3d object detection and tracking. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2021)", + "37. Yue, Y., Kontogianni, T., Schindler, K., Engelmann, F.: Connecting the dots: Floorplan reconstruction using two-level queries. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)", + "38. Zou, C., Colburn, A., Shan, Q., Hoiem, D.: Layoutnet: Reconstructing the 3d room layout from a single rgb image. In: Proceedings of the IEEE conference on computer vision and pattern recognition (2018)" + ], + "bbox": [ + 212, + 146, + 787, + 314 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "SceneScript", + "bbox": [ + 651, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/SceneScript_ Reconstructing Scenes With An Autoregressive Structured Language Model/6d586386-d370-4aeb-a27d-9451b501b037_model.json b/2024/SceneScript_ Reconstructing Scenes With An Autoregressive Structured Language Model/6d586386-d370-4aeb-a27d-9451b501b037_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bb4cc33a6bd804b0dfb7eb2a318ee519af680655 --- /dev/null +++ b/2024/SceneScript_ Reconstructing Scenes With An Autoregressive Structured Language Model/6d586386-d370-4aeb-a27d-9451b501b037_model.json @@ -0,0 +1,2434 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.245, + 0.078, + 0.759, + 0.122 + ], + "angle": 0, + "content": "SceneScript: Reconstructing Scenes With An Autoregressive Structured Language Model" + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.134, + 0.783, + 0.194 + ], + "angle": 0, + "content": "Armen Avetisyan1, Christopher Xie1, Henry Howard-Jenkins1, Tsun-Yi Yang1, Samir Aroudj1, Suvam Patra1, Fuyang Zhang2†, Duncan Frost1, Luke Holland1, Campbell Orme1, Jakob Engel1, Edward Miller1, Richard Newcombe1, and Vasileios Balntas1" + }, + { + "type": "text", + "bbox": [ + 0.33, + 0.206, + 0.673, + 0.236 + ], + "angle": 0, + "content": "Meta Reality Labs Simon Fraser University https://projectaria.com/scenescript" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.251, + 0.368, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.254, + 0.319, + 0.321, + 0.326 + ], + "angle": 0, + "content": "Scene walkthrough" + }, + { + "type": "image", + "bbox": [ + 0.391, + 0.251, + 0.471, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.493, + 0.278, + 0.511, + 0.292 + ], + "angle": 0, + "content": "#" + }, + { + "type": "image", + "bbox": [ + 0.55, + 0.253, + 0.598, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.319, + 0.62, + 0.327 + ], + "angle": 0, + "content": "Structured scene language" + }, + { + "type": "image", + "bbox": [ + 0.637, + 0.252, + 0.785, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.659, + 0.319, + 0.774, + 0.327 + ], + "angle": 0, + "content": "Language to geometry interpreter" + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.335, + 0.33, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.392, + 0.329, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.332, + 0.335, + 0.442, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.332, + 0.392, + 0.443, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.447, + 0.335, + 0.557, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.446, + 0.392, + 0.557, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.335, + 0.671, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.392, + 0.671, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.674, + 0.335, + 0.785, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.675, + 0.392, + 0.785, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.459, + 0.788, + 0.557 + ], + "angle": 0, + "content": "Fig. 1: (top) Given an egocentric video of an environment, SceneScript directly predicts a 3D scene representation consisting of structured scene language commands. (bottom) Our method generalizes on diverse real scenes while being solely trained on synthetic indoor environments. (last column, bottom) A notable advantage of our method is its capacity to easily adapt the structured language to represent novel scene entities. For example, by introducing a single new command, SceneScript can directly predict object parts jointly with the layout and bounding boxes." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.583, + 0.741, + 0.819 + ], + "angle": 0, + "content": "Abstract. We introduce SceneScript, a method that directly produces full scene models as a sequence of structured language commands using an autoregressive, token-based approach. Our proposed scene representation is inspired by recent successes in transformers & LLMs, and departs from more traditional methods which commonly describe scenes as meshes, voxel grids, point clouds or radiance fields. Our method infers the set of structured language commands directly from encoded visual data using a scene language encoder-decoder architecture. To train SceneScript, we generate and release a large-scale synthetic dataset called Aria Synthetic Environments consisting of \\(100k\\) high-quality indoor scenes, with photorealistic and ground-truth annotated renders of egocentric scene walkthroughs. Our method gives state-of-the-art results in architectural layout estimation, and competitive results in 3D object detection. Lastly, we explore an advantage for SceneScript, which is the ability to readily adapt to new commands via simple additions to the structured language, which we illustrate for tasks such as coarse 3D object part reconstruction." + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.825, + 0.584, + 0.839 + ], + "angle": 0, + "content": "\\(\\dagger\\)Work done while the author was an intern at Meta." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.129 + ], + "angle": 0, + "content": "Avetisyan et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.377, + 0.162 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.788, + 0.417 + ], + "angle": 0, + "content": "Scene representations play a crucial role in machine learning and computer vision applications, enabling accurate understanding of the environment. Over the years, researchers have explored various options such as meshes, voxel grids, point clouds, and implicit representations, aiming to represent complex real-world scenes with high-fidelity. Each of these exhibits distinct advantages and limitations that impact their suitability for different tasks. Meshes offer detailed geometric information but can be expensive in both computation and memory. Voxel grids provide a volumetric representation but suffer from a trade-off between resolution and memory requirements. Point clouds are efficient in representing sparse scenes but lack semantics and explicit connectivity information. Implicit representations, such as DeepSDF [22] and NeRF [15], can be infinitely precise but lack interpretability and editability. The selection of an appropriate representation directly impacts the performance and efficacy of various tasks including object recognition, scene understanding, and 3D reconstruction. In this paper, we propose a novel scene representation based on structured language commands as a more efficient and versatile solution." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.418, + 0.788, + 0.568 + ], + "angle": 0, + "content": "Our motivation stems from the recent advancements in the field of Large Language Models (LLMs) and \"next token prediction\" autoregressive methods [20], coupled with recent works on exploring generation of sequences to represent geometric structures. For example, PolyGen [18] demonstrated the ability to describe 3D meshes as a sequence of vertices and faces generated using transformers [30]. Similarly, CAD-as-Language [11] showcased the effectiveness of generating Computer-Aided Design (CAD) primitives to represent 2D CAD sketches. Our main goal is to directly infer a metrically accurate representation of a full scene as a text-based sequence of specialized structured language commands." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.788, + 0.75 + ], + "angle": 0, + "content": "Our method, denoted SceneScript, autoregressively predicts a language of hand-designed structured language commands in pure text form. This language offers several distinct advantages: 1) As pure text, it is compact and reduces memory requirements of a large scene to only a few bytes. 2) It is crisp and complete since the commands are designed to result in sharp and well-defined geometry (similar to scalable vector graphics). 3) It is interpretable, editable and semantically rich by design via the use of high-level parametric commands such as make_door(*door_parameters). 4) It can seamlessly integrate novel geometric entities by simply adding new structured commands to the language. 5) The fact that the scene representation is a series of language tokens similar to [20] opens up a plethora of potential new applications in the future such as ways to edit the scene, query it, or spin up chat interactions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We mainly focus on the problems of architectural layout estimation and object detection as proxy tasks for the efficacy of our SceneScript language as a scene representation. Architectural entities such as walls, doors, and windows are highly structured entities, making them an ideal test-bed. However, one notable drawback of language models is that they require vast amounts of data for training. Since there is no existing dataset of scene walkthroughs and their cor" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.652, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SceneScript" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "responding structured language commands, we publicly released Aria Synthetic Environments (ASE), a synthetically generated dataset of 100k unique interior scenes. For each scene, we simulate egocentric trajectories with an entire suite of sensor data from Project Aria [14]. We also release multiple sources of ground truth including depth and instance segmentations. Importantly, for each rendered egocentric sequence, the architectural layout ground truth is given in our proposed SceneScript language." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.788, + 0.389 + ], + "angle": 0, + "content": "While architectural layout serves as a test-bed, we demonstrate that our method SceneScript can easily be extended to new tasks via simple extensions to SceneScript language while keeping both the visual input and network architecture fixed. We illustrate this on the problem of 3D object detection, which results in a method that jointly infers architectural layout and 3D oriented bounding boxes. Additionally, we demonstrate more proof-of-concept experiments that show that our method results in a significantly lower barrier to entry for new tasks including representing coarse 3D object reconstruction, curved entities, composition of entities, and entity states." + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.39, + 0.438, + 0.403 + ], + "angle": 0, + "content": "Our core contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.414, + 0.784, + 0.458 + ], + "angle": 0, + "content": "- We propose SceneScript, a method that jointly predicts architectural layout and object bounding boxes of a scene in the form of structured language commands given a video stream." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.459, + 0.785, + 0.503 + ], + "angle": 0, + "content": "- We demonstrate that SceneScript can easily be extended to completely new tasks with simple additions of commands to our structured language, significantly lowering the barrier to entry for new tasks." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.504, + 0.785, + 0.549 + ], + "angle": 0, + "content": "- We release a large-scale synthetic dataset, named Aria Synthetic Environments, comprised of 100k unique high-quality 3D indoor scenes with GT, which will enable large scale ML training of scene understanding methods." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.549, + 0.785, + 0.58 + ], + "angle": 0, + "content": "- We show that training SceneScript on Aria Synthetic Environments leads to generalization on real scenes (see videos/demos on the project page)." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.414, + 0.785, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.602, + 0.396, + 0.618 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.634, + 0.416, + 0.649 + ], + "angle": 0, + "content": "2.1 Layout Estimation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.749 + ], + "angle": 0, + "content": "Layout estimation is an active research area, aiming to infer architectural elements. Scan2BIM [17] proposes heuristics for wall detection to produce 2D floorplans. Ochmann et al. [19] formulate layout inference as an integer linear program using constraints on detected walls. Shortest path algorithms around birds-eye view (BEV) free space [3] and wall plus room instance segmentation [5] have also been explored." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Furukawa et al. [10] utilize a Manhattan world-based multi-view stereo algorithm [9] to merge axis-aligned depth maps into a full 3D mesh of building interiors. RoomNet [13] predicts layout keypoints while assuming that a fixed set of Manhattan layouts can occur in a single image of a room. LayoutNet [38] improves on this by predicting keypoints and optimising the Manhattan room layout inferred from them. Similarly, AtlantaNet [23] predicts a BEV floor or" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.129 + ], + "angle": 0, + "content": "Avetisyan et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 1: Complete set of structured language commands designed for detailing architectural layouts and object bounding boxes. Supported data types can include int, float, bool. It is important to note that the language's extensibility allows for easy augmentation by introducing new commands like make_prime, make_pillar, or enhancing existing commands, such as incorporating is-double_door (bool)." + }, + { + "type": "table", + "bbox": [ + 0.289, + 0.227, + 0.706, + 0.334 + ], + "angle": 0, + "content": "
make_WALL (int)make_door (int)make_window (int)make_bbox (int)
idintidintidintidint
a_xfloatwall0_idintwall0_idintclassint
a_yfloatwall1_idintwall1_idintposition_xfloat
a_zfloatposition_xfloatposition_xfloatposition_yfloat
b_xfloatposition_yfloatposition_yfloatposition_zfloat
b_yfloatposition_zfloatposition_zfloatangle_zfloat
b_zfloatwidthfloatwidthfloatscale_xfloat
heightfloatheightfloatheightfloatscale_yfloat
scale_zfloat
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.366, + 0.785, + 0.412 + ], + "angle": 0, + "content": "ceiling shape and approximates the shape contour with a polygon resulting in an Atlanta world prior. SceneCAD [1] uses a graph neural network to predict object-object and object-query relationships to refine its layout prediction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.412, + 0.787, + 0.473 + ], + "angle": 0, + "content": "Our approach stands out by requiring neither heuristics nor explicitly defined prior knowledge about architectural scene structure. In fact, our method demonstrates geometric understanding of the scene which emerges despite learning to predict the GT scene language as a sequence of text tokens." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.503, + 0.52, + 0.52 + ], + "angle": 0, + "content": "2.2 Geometric Sequence Modelling" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.537, + 0.788, + 0.732 + ], + "angle": 0, + "content": "Recent works have explored transformers for generating objects as text-based sequences. PolyGen [18] models 3D meshes as a sequence of vertices and faces. CAD-as-Language [11] represents 2D CAD sketches as a sequence of triplets in protobuf representation, followed by a sequence of constraints. Both SketchGen [21] and SkexGen [33] use transformers to generate sketches. DeepSVG [4] learns a transformer-based variational autoencoder (VAE) that is capable of generating and interpolating through 2D vector graphics images. DeepCAD [32] proposes a low-level language and architecture similarly to DeepSVG, but applies it to 3D CAD models instead of 2D vector graphics. Our approach stands out by utilising high-level commands, offering interpretability and semantic richness. Additionally, while low-level commands can represent arbitrarily complex geometries, they lead to prohibitively longer sequences when representing a full scene." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.843 + ], + "angle": 0, + "content": "The closest work to ours is Pix2Seq [6]. Pix2Seq proposes a similar architecture to ours but experiments only with 2D object detection, thus requiring domain-specific augmentation strategies. Another closely related work is Point2Seq [34] that trains a recurrent network for autoregressively regressing continuous 3D bounding box parameters. Interestingly, they find the autoregressive ordering of parameters outperforms current standards for object detection architectures, including anchors [12] and centers [36]." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.652, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SceneScript" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.787, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.356, + 0.788, + 0.412 + ], + "angle": 0, + "content": "Fig. 2: Aria Synthetic Environments: (top) Random samples of generated scenes showing diversity of layouts, lights and object placements. (bottom - left to right) A top down view of a scene filled with objects, a simulated trajectory (blue path), renderings of depth, RGB, and object instances, and lastly a scene pointcloud." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.44, + 0.688, + 0.459 + ], + "angle": 0, + "content": "3 SceneScript Structured Language Commands" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.473, + 0.788, + 0.52 + ], + "angle": 0, + "content": "We first describe our structured language commands that define a full scene representation including both layout and objects. After this, we introduce our corresponding large scale training dataset: Aria Synthetic Environments." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.542, + 0.492, + 0.556 + ], + "angle": 0, + "content": "3.1 Commands and Parameters" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.788, + 0.703 + ], + "angle": 0, + "content": "We begin with a parameterization that captures the most common layout elements. For this purpose we use three commands: make-wall, make_door, and make_window. Each command comes with a set of parameters that results in well-defined geometry. For example, the full set of parameters for make-wall specifies a gravity-aligned 2D plane, while make_door and make_window specify box-based cutouts from walls. It is worth noting that this parametrization is arbitrary, and is only made in the context of presenting a proof-of-concept SceneScript system. There are infinitely many parametrization schemes, in this work we opt for one that prioritizes ease of use and research iteration speed." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.787, + 0.734 + ], + "angle": 0, + "content": "In addition to representing these three major layout entities, we aim to jointly infer objects as oriented bounding boxes. Thus, we introduce a fourth command:" + }, + { + "type": "code", + "bbox": [ + 0.216, + 0.743, + 0.725, + 0.773 + ], + "angle": 0, + "content": "make_bbox: id, class, position_x, position_y, position_z, angle_z, scale_x, scale_y, scale_z" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.84 + ], + "angle": 0, + "content": "This simple parametrization represents an oriented 3D bounding box that is assumed to be aligned with gravity (assuming it points in the \\(-z\\) direction). A summary of these commands and their respective parameters are shown in Table 1." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.129 + ], + "angle": 0, + "content": "Avetisyan et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.269 + ], + "angle": 0, + "content": "While we have described just four commands to capture structure and objects in an indoor environment, importantly, this text-based parametrization can readily be extended geometrically and/or even semantically to include states or other functional aspects. For example, changing the mentioned makedoor command to include parameters such as open-degree allows the language to represent door states. In Section 6, we demonstrate how such extensions to the language allows for SceneScript to readily adapt to new tasks including coarse 3D object reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.292, + 0.4, + 0.306 + ], + "angle": 0, + "content": "3.2 Scene Definition" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.318, + 0.787, + 0.381 + ], + "angle": 0, + "content": "A single scene can be described as a sequence of our proposed structured language commands. The sequence requires no specific ordering, and can be of arbitrary length. From this definition, the 3D scene can easily be obtained by parsing the commands through a simple custom interpreter." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.403, + 0.403, + 0.418 + ], + "angle": 0, + "content": "3.3 Training Dataset" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.429, + 0.787, + 0.49 + ], + "angle": 0, + "content": "To enable practical indoor scene reconstruction based on our structured language commands (Section 3.1), we publicly released a new dataset called Aria Synthetic Environments. It consists of a large number of training pairs of egocentric scene walkthroughs linked with corresponding ground truth command sequences." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.49, + 0.788, + 0.581 + ], + "angle": 0, + "content": "Since transformers require vast amounts of data, we generated \\(100k\\) synthetic scenes, which is in comparison infeasible for real-world data. Each synthetic scene comes with a floor plan model, a corresponding complete 3D scene as well as a simulated agent trajectory and a photo-realistic rendering of this trajectory. Fig. 2 illustrates the basics of Aria Synthetic Environments. For brevity, we refer the reader to the supplemental material for further details." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.606, + 0.584, + 0.623 + ], + "angle": 0, + "content": "4 SceneScript Network Architecture" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.639, + 0.787, + 0.683 + ], + "angle": 0, + "content": "Our pipeline is a simple encoder-decoder architecture that consumes a video sequence and returns SceneScript language in a tokenized format. Figure 3 illustrates a high-level overview of our method." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.685, + 0.787, + 0.715 + ], + "angle": 0, + "content": "We examine three encoder variants: a pointcloud encoder, a posed image set encoder, and a combined encoder. The decoder remains the same in all cases." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.738, + 0.52, + 0.754 + ], + "angle": 0, + "content": "4.1 Input Modalities and Encoders" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.843 + ], + "angle": 0, + "content": "The encoder computes a latent scene code in the form of a 1D sequence from video walkthrough of the scene. The decoder is designed to consume these 1D sequences as input. This enables the integration of various input modalities within a unified framework. As a preliminary, for each scene we assume access to a set of \\( M \\) posed camera images \\( \\{\\mathbf{I}_1,\\dots,\\mathbf{I}_{\\mathbf{M}}\\} \\), e.g., SLAM output." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.652, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SceneScript" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.144, + 0.531, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.143, + 0.787, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.293, + 0.788, + 0.364 + ], + "angle": 0, + "content": "Fig. 3: SceneScript core pipeline overview. Raw images & pointcloud data are encoded into a latent code, which is then autoregressively decoded into a sequence of commands that describe the scene. Visualizations are shown using a customly built interpreter. Note that for the results in this paper, the the point clouds are computed from the images using Aria MPS [25] - i.e. are not using a dedicated RGB-D / Lidar sensor." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.393, + 0.784, + 0.437 + ], + "angle": 0, + "content": "Point Clouds. A point cloud \\(\\mathbf{P} = \\{\\mathbf{p}_1, \\dots, \\mathbf{p_N}\\}\\) consists of \\(N\\) points, where \\(\\mathbf{p}_i\\) is a 3D point coordinate. It can come from passive images using SLAM or SfM, or RGB-D / Lidar sensors." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.438, + 0.785, + 0.528 + ], + "angle": 0, + "content": "Specifically, we use the Semi-dense Pointclouds from Project Aria's Machine Perception Services [25], that are obtained from a visual-inertial SLAM system using Aria's monochrome cameras and IMUs. We discretize the point cloud to \\(5\\mathrm{cm}\\) resolution, then employ a sparse 3D convolution library [26,27] to generate pooled features. The encoder \\(\\mathcal{E}_{geo}\\) applies a series of down convolutions, resulting in a reduction of the number of points in the lowest level." + }, + { + "type": "equation", + "bbox": [ + 0.338, + 0.542, + 0.785, + 0.561 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {g e o} = \\mathcal {E} _ {g e o} (\\mathbf {P}), \\quad \\mathbf {P} \\in \\mathbb {R} ^ {N \\times 3}, \\mathbf {F} _ {g e o} \\in \\mathbb {R} ^ {K \\times 5 1 2} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.566, + 0.788, + 0.658 + ], + "angle": 0, + "content": "where \\( K \\ll N \\). \\( \\mathbf{F}_{geo} \\) is a condensed latent representation of the point cloud that contains the necessary scene context. For later use in the transformer decoder, we treat \\( \\mathbf{F}_{geo} \\) as a sequence of feature vectors where the entries \\( \\mathbf{f_i} \\), \\( i \\in 1\\dots K \\) are sorted lexicographically according to the coordinate of the active site \\( \\mathbf{c_i} \\), \\( i \\in 1\\dots K \\). To incorporate positional encoding, we append the coordinates of the active sites to their respective feature vectors \\( \\mathbf{f}_i \\gets \\mathbf{cat}(\\mathbf{f_i},\\mathbf{c_i}) \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.679, + 0.788, + 0.771 + ], + "angle": 0, + "content": "Point Clouds with Lifted Features. We additionally explore augmenting the point cloud with image features. From the original egocentric sequence and associated trajectory, we sample a set of \\(M\\) keyframes, \\(\\mathbf{I}_i\\) where \\(i\\in 1\\dots M\\), and compute a set of image features \\(\\mathbf{F}_i\\) for each. We then project each point into the set of keyframe cameras and retrieve the feature vector (output by a CNN) at the pixel location:" + }, + { + "type": "equation", + "bbox": [ + 0.37, + 0.783, + 0.785, + 0.8 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {i p} = F _ {i} (\\pi (\\mathbf {p})) \\quad \\mathbf {p} \\in \\mathbf {P}, i \\in 1 \\dots M, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "where \\(\\pi(\\cdot)\\) represents the projection function of a 3D point into the camera. If \\(\\pi(\\mathbf{p})\\) falls outside the image bounds, no feature is retrieved. We combine the set" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.129 + ], + "angle": 0, + "content": "Avetisyan et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "of lifted features for a point through an average, resulting in a single feature vector for each point: \\(\\mathbf{f}_p = 1 / M\\sum_{i = 1}^{M}\\mathbf{f}_{ip}\\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.785, + 0.239 + ], + "angle": 0, + "content": "We form our lifted-feature point cloud, \\(\\mathbf{P}' = \\{\\mathbf{p}_1',\\dots,\\mathbf{p}_{\\mathbf{M}}'\\}\\), by concatenating each point's lifted feature with the original XYZ location: \\(\\mathbf{p}' = \\mathrm{cat}(\\mathbf{f}_p,\\mathbf{p})\\). \\(\\mathbf{P}'\\) is then encoded into a context sequence using sparse 3D convolutions, with only the input feature channels adjusted to match the new point feature dimension." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.263, + 0.784, + 0.308 + ], + "angle": 0, + "content": "End-to-end Encoding of Posed Views. In order to encode the egocentric sequence more directly without a pre-computed point cloud, we adopt a \\(2\\mathrm{D} \\leftrightarrow 3\\mathrm{D}\\) bidirectional transformer encoder following the form defined in RayTran [29]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.31, + 0.785, + 0.52 + ], + "angle": 0, + "content": "In this formulation, we initialize a volume representation of the scene as a dense voxel grid of features, \\(\\mathbf{V}\\), that coincides with the scene geometry. In turn, we sample a subset of \\(M\\) keyframes, \\(\\mathbf{I}_i\\) where \\(i\\in 1\\dots M\\), from the full stream of posed images. And for each of these keyframes we compute image features from a CNN, \\(\\mathbf{F}_i\\). Repeated layers of bidirectional attention enable the image and voxel grid features to be refined iteratively in successive transformer blocks through the aggregation of view-point and global-scene information. As in RayTran [29], the interaction between the two representations is guided by the image-formation process by leveraging known camera parameters and poses. Attention in these transformer blocks is restricted by patch-voxel ray intersections, where each image patch attends to the voxels it observes and each voxel location attends to all the patches that observe it. The resulting voxel grid of features is flattened, concatenated with an encoded representation of its XYZ location, and passed to the decoder." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.545, + 0.416, + 0.56 + ], + "angle": 0, + "content": "4.2 Language Decoder" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.573, + 0.784, + 0.678 + ], + "angle": 0, + "content": "We utilize a transformer decoder [30] to decode the scene latent code into a sequence of structured language commands. The sequence of tokens passes through an embedding layer, followed by a positional encoding layer. Together with the encoded scene code (Section 4.1), the embedded tokens are passed into the several transformer decoder layers where a causal attention mask is used to ensure autoregressive generation. More implementation details can be found in Appendix ??" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.704, + 0.452, + 0.719 + ], + "angle": 0, + "content": "4.3 Language Tokenization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.732, + 0.784, + 0.807 + ], + "angle": 0, + "content": "We refer to the serialization the structured language into a sequence of tokens as tokenization. The goal is to construct a bijective mapping between a sequence of structured language commands (Section 3) and a sequence of integer tokens that can be predicted by the transformer decoder architecture. We utilize the following schema:" + }, + { + "type": "footer", + "bbox": [ + 0.264, + 0.82, + 0.737, + 0.836 + ], + "angle": 0, + "content": "[START, PART, CMD, PARAM_1, PARAM_2, ..., PARAM_N, PART, ..., STOP]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.652, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SceneScript" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.147, + 0.784, + 0.162 + ], + "angle": 0, + "content": "For example, a sample sequence for a makedoor command may look like:" + }, + { + "type": "text", + "bbox": [ + 0.27, + 0.19, + 0.731, + 0.224 + ], + "angle": 0, + "content": "[START, PART, MAKE_DOOR, POSITION_X, POSITION_Y, POSITION_Z, WALLO_IDX, WALL1_IDX, WIDTH, HEIGHT, PART,..., STOP]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.784, + 0.312 + ], + "angle": 0, + "content": "This schema enables 1D packing of tokens without requiring fixed-size slots or padding like other sequence modelling methods such as [32]. Additionally, it does not impose any limitations on the number or the hierarchy of sub-sequences, as they are flexibly separated by a PART token. This allows for arbitrarily complex scene representations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.314, + 0.784, + 0.389 + ], + "angle": 0, + "content": "The tokenized sequence is discretized into integers at a 5cm resolution, then translated into a sequence of embeddings via learnable lookup table. Note that by designing the SceneScript language, we also design the tokenization. This tokenization scheme is notably different from standard NLP tokenization, which involves Byte-Pair Encodings (BPE) [20]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.414, + 0.325, + 0.429 + ], + "angle": 0, + "content": "5 Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.447, + 0.784, + 0.493 + ], + "angle": 0, + "content": "In this section, we introduce the metrics we define to measure performance, and we discuss some qualitative and quantitative results that give insights to our proposed SceneScript method." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.517, + 0.326, + 0.531 + ], + "angle": 0, + "content": "5.1 Metrics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.544, + 0.797, + 0.666 + ], + "angle": 0, + "content": "To evaluate accuracy of the layout estimation, we make use of geometric metrics applied between the ground-truth room layout and our predicted SceneScript language. To do so, we define an entity distance, \\( d_{E} \\), between a pair of entities of the same class. Each entity, \\( E \\), is represented as a 3D plane segment comprising of 4 corners \\( \\{c_1, c_2, c_3, c_4\\} \\). The distance between two entities, \\( E \\) and \\( E' \\) is computed as the maximum Euclidean distance between each corner and its counterpart assigned via Hungarian mathcing, i.e.: \\( d_{E}(E, E') = \\max \\{||c_{i} - c_{\\pi(i)}'| : i = 1, \\dots, 4\\} \\), where \\( \\pi(i) \\) is the permutation also found via Hungarian matching." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.667, + 0.784, + 0.697 + ], + "angle": 0, + "content": "We threshold \\( d_{E} \\) to define the success criteria for the predicted entities. We compute the following metrics:" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.708, + 0.784, + 0.737 + ], + "angle": 0, + "content": "- F1 Score @ threshold - the F1 score of the set of predictions is computed at a single \\( d_{E} \\) threshold." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.739, + 0.784, + 0.768 + ], + "angle": 0, + "content": "- Average F1 Score - the F1 score is computed across a range of entity distance thresholds and averaged." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.708, + 0.784, + 0.768 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.784, + 0.84 + ], + "angle": 0, + "content": "The scores are computed for each class independently and averaged to overall score. In addition, scores are computed for each scene and averaged across the dataset. We use the following range of thresholds (cm) for the average F1 scores: \\( T = \\{1, 2, \\dots, 9, 10, 15, 25, 30, 50, 75, 100\\} \\)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.129 + ], + "angle": 0, + "content": "Avetisyan et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.787, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.37, + 0.79, + 0.442 + ], + "angle": 0, + "content": "Fig. 4: Qualitative samples between our model and SOTA methods on Aria Synthetic Environments's test set. Hierarchical methods like SceneCAD suffer from error cascading which leads to missing elements in the edge prediction module. RoomFormer (a 2D method extruded to 3D) primarily suffers from lightly captured scene regions which leave a unnoticeable signal in the density map." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.456, + 0.79, + 0.484 + ], + "angle": 0, + "content": "Table 2: Layout Estimation on Aria Synthetic Environments Quantitative comparison between our method and related recent work." + }, + { + "type": "table", + "bbox": [ + 0.236, + 0.496, + 0.765, + 0.603 + ], + "angle": 0, + "content": "
MethodF1 @5cmAvg F1
meanwalldoorwindowmeanwalldoorwindow
SceneCAD '20 [1]-0.048---0.275--
RoomFormer '23 [37]0.1390.1590.1480.1100.4640.5050.4810.407
Ours (Point cloud)0.8480.9300.9220.6920.7840.8160.8110.724
Ours (Lifted features)0.9030.9430.9590.8060.8010.8180.8220.764
Ours (Image-only)0.6610.6870.7980.4970.7190.7270.7720.658
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.63, + 0.416, + 0.645 + ], + "angle": 0, + "content": "5.2 Layout Estimation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.658, + 0.788, + 0.733 + ], + "angle": 0, + "content": "We perform scene layout estimation with the three encoder variants of SceneScript: a baseline model with sparse3d convolution pointcloud encoder, an RGB RayTran-based feature volume encoder [29], and our proposed lifted feature point encoder. The same transformer decoder is being used in all three scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.795 + ], + "angle": 0, + "content": "To provide comparison to existing works, we include results from two baseline methods, namely SceneCAD [1] and the recent RoomFormer [37]. For these experiments SceneCAD and RoomFormer were both trained on Aria Synthetic Environments. Note that SceneCAD only predicts walls." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Table 2 shows the main results for our F1-based metrics on Aria Synthetic Environments. SceneScript exhibits a substantial performance advantage over SOTA layout estimation baselines across multiple metrics." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.652, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SceneScript" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 3: 3D Object Detection Performance comparison against state-of-the-art methods on an 3D object detection task trained and evaluated by F1-score at 0.25 and 0.5 IoU thresholds. By simply adding a make_bbox command SceneScript can achieve competitive object detection results." + }, + { + "type": "table_caption", + "bbox": [ + 0.264, + 0.225, + 0.464, + 0.239 + ], + "angle": 0, + "content": "(a) Aria Synthetic Environments" + }, + { + "type": "table", + "bbox": [ + 0.226, + 0.248, + 0.498, + 0.326 + ], + "angle": 0, + "content": "
MethodInputF1@.25 IoU @.50 IoU
3DETR '21 [16]Points0.2010.078
Cube R-CNN '23 [2]RGB0.3940.228
ImVoxelNet '22 [24]RGB0.5840.516
OursPoints0.6200.577
" + }, + { + "type": "table_caption", + "bbox": [ + 0.594, + 0.225, + 0.686, + 0.238 + ], + "angle": 0, + "content": "(b) ScanNet [8]" + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.248, + 0.776, + 0.321 + ], + "angle": 0, + "content": "
MethodInputF1
@.25 IoU@.50 IoU
3DETR '21 [16]Points0.4800.349
3DETR-m '21 [16]Points0.5360.407
SoftGroup '22 [31]RGB Points0.6220.573
OursRGB Points0.5060.406
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.356, + 0.784, + 0.401 + ], + "angle": 0, + "content": "Both baseline methods encounter a significant decline in accuracy when dealing with finer details. See Figure 4 for qualitative comparisons between our method and baseline methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.426, + 0.785, + 0.485 + ], + "angle": 0, + "content": "Encoder Ablation. The results demonstrate that SceneScript is robust to the encoding strategy chosen to encode the egocentric capture. It is able to infer highly accurate scene layouts in all configurations tested, and in each case SceneScript outperforms the included baselines by a significant margin." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.487, + 0.788, + 0.623 + ], + "angle": 0, + "content": "Relative comparison of the encoder strategies reveals that leveraging the point clouds from a highly specialized and well-tuned system is still able to offer advantages of an, almost, entirely learned approach such as RayTran [29]. A light extension in the form of lifted image features can widen this gap even further. In particular, we observe that the discrepancy between the encoders becomes particularly apparent as the complexity of the scene increases in the form of increased room count. In the Appendix, we also show a quantitative evaluation of per-entity error distances, which aids in further attribution of relative performance gains between the encoding methods." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.647, + 0.406, + 0.662 + ], + "angle": 0, + "content": "5.3 Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.787, + 0.734 + ], + "angle": 0, + "content": "In this section, we perform evaluation of SceneScript for object detection on both Aria Synthetic Environments and ScanNet [8]. For comparison, we include recent and state-of-the-art baseline methods, namely Cube-RCNN [2], ImVoxelNet [24], 3DETR [16], and SoftGroup [31]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Worth noting is that SceneScript does not predict a confidence score for each make_bbox command. This means that fair computation of the conventional mAP metric for this task is not possible, as detections cannot be ranked across scenes. Among other issues, this results in a metric that varies with the order in which scenes are evaluated. We therefore report F1-score-based metrics, which do not exhibit this order variance. Further discussion of this, and mAP numbers for the baselines for reference, can be found in the Appendix." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.129 + ], + "angle": 0, + "content": "Avetisyan et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.331, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.332, + 0.144, + 0.788, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.279, + 0.788, + 0.337 + ], + "angle": 0, + "content": "Fig. 5: Example scene reconstructions on scenes from Aria Synthetic Environments. (left) Visualisation of the decomposed meshes used to create make_prim training pairs. (right) Views of full scene predictions, as well as close ups highlighting the fidelity of object reconstruction through the prediction volumetric primitives enabled by make_prim." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.365, + 0.788, + 0.485 + ], + "angle": 0, + "content": "In Table 3a, all methods were trained on Aria Synthetic Environments. 3DETR performs poorly due to its encoder being designed for relatively clean datasets such as ScanNet [8], while semi-dense point clouds from Aria Synthetic Environments exhibit more noise and less uniformity (see Figure 3 for an example). Cube R-CNN and ImVoxelNet both operate on RGB images, and detections are tracked for the entire sequence via a tracker [2] to provide competitive performance. In Table 3b, our method provides similar performance to both 3DETR and SoftGroup." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.487, + 0.788, + 0.563 + ], + "angle": 0, + "content": "Through the addition of the make_bbox command, SceneScript demonstrates object detection performance on par with SOTA baselines on multiple object detection datasets. This result illustrates the extensibility of a token-based approach, and that our proposed SceneScript language representation does not suffer compared to specialised object detection network architectures." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.588, + 0.72, + 0.607 + ], + "angle": 0, + "content": "6 Extending the SceneScript Structured Language" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.622, + 0.804, + 0.713 + ], + "angle": 0, + "content": "A key advantage offered by SceneScript's structured language prediction paradigm is that the expressiveness of its reconstruction can be tailored without requiring a change to the method. Up to now, we have focussed on showcasing the efficacy of SceneScript for representing simple layout elements and objects as bounding boxes. In this section, we showcase this characteristic by increasing the fidelity of our scene representation by introducing coarse 3D object reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.737, + 0.536, + 0.753 + ], + "angle": 0, + "content": "6.1 Objects as Volumetric Primitives" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We turn to a language based on volumetric primitives, motivated by works such as [28, 35]. Using simple primitives such as cuboids and extruded cylinders, enables us to coarsely represent arbitrary object categories while maintaining object semantics (e.g. tabletops can be represented by a single cuboid). Thus, this language can describe many object categories simultaneously." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.652, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SceneScript" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.787, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.291, + 0.788, + 0.321 + ], + "angle": 0, + "content": "Fig.6: Example scene reconstructions on real scenes with the addition of the make_prim command. Note that SceneScript was trained only on synthetic data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.35, + 0.784, + 0.381 + ], + "angle": 0, + "content": "This representation requires only one additional command over the layout and box commands already discussed previously, namely:" + }, + { + "type": "code", + "bbox": [ + 0.216, + 0.389, + 0.782, + 0.435 + ], + "angle": 0, + "content": "make_prime:_bbox_id, prim_num, class, center_x, center_y, center_z, angle_x, angle_y, angle_z, scale_x, scale_y, scale_z" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.442, + 0.788, + 0.503 + ], + "angle": 0, + "content": "This command and its parameters describe a volumetric primitive (cuboid or extruded cylinder) via its 3D center, 3D rotation, and 3D scale. The prim_num parameter can be associated with semantics, e.g. tabletops of different tables typically have the same prim_num." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.525, + 0.788, + 0.648 + ], + "angle": 0, + "content": "Dataset. To obtain ground truth make_prim commands that align with the objects in Aria Synthetic Environments, we first run an extension of Yang et al. [35] to obtain cuboid and extruded cylinder primitives of a database of 3D CAD models (ABO [7], which was used to populate Aria Synthetic Environments). See Figure 5 (left) for example decompositions. For this proof-of-concept experiment, we use three categories: table, chair, and sofa. We then convert these decomposed primitives into make_prim commands that are aligned with the objects in the dataset, which results in training pairs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.67, + 0.788, + 0.747 + ], + "angle": 0, + "content": "Results. We show qualitative results on Aria Synthetic Environments in Figure 5. In Figure 6, we show inferences in a few real-world environments despite only having trained on our simulated dataset. These results demonstrate that SceneScript's general purpose architecture is capable of coarsely reconstructing objects of multiple categories through addition of a new command." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.769, + 0.421, + 0.784 + ], + "angle": 0, + "content": "6.2 Further Extensions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "In this section, we have explored just one extension of SceneScript's structured scene language in order to demonstrate its flexibility. The result was greatly increased expressiveness of the scene reconstruction. In the Appendix, we include" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.129 + ], + "angle": 0, + "content": "Avetisyan et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "additional explorations that can further increase the fidelity and accuracy of the reconstructions. These explorations include reconstructing curved walls, inferring object states (e.g., door opening angles), as well as direct prediction of parametric models using object models deployed commonly by tech artists (e.g., blender geometry nodes) for object reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.244, + 0.573, + 0.26 + ], + "angle": 0, + "content": "7 Interactive Scene Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.272, + 0.788, + 0.41 + ], + "angle": 0, + "content": "Scene reconstruction often occurs offline, on pre-recorded walkthroughs of an environment, or derivatives of them, such as a point cloud. However, we take advantage of SceneScript's inference occurring at an interactive rate, which takes between 2-5s depending on the size of the environment, by implementing streaming of SceneScript's live reconstructions into a modern VR headset. This enables an interactive experience where the user can see the reconstruction overlaid onto the environment they are exploring in real-time. See the video recording on the website for examples. Visualisations of this interface are included in Figure 1." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.43, + 0.538, + 0.446 + ], + "angle": 0, + "content": "8 Limitations and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.458, + 0.788, + 0.655 + ], + "angle": 0, + "content": "SceneScript exhibits certain limitations that should be acknowledged. First, the structured language commands are manually defined, which requires human intervention at this stage. Secondly, due to the higher-level nature of our commands, it can be challenging to capture fine-grained geometric details with extremely high precision (i.e. mm). As a consequence, the resulting reconstructions based on this representation tend to lead to simpler and coarser geometries, potentially missing intricate nuances at the very high detail level. These limitations potentially highlight areas for future research and optimization, aiming to automate the command definition process and explore techniques to enhance the representation's ability to capture intricate geometric details accurately. However, we believe that the ability to built scene representations that are based on structured language commands will be a key part in complex and efficient methods of scene reconstruction in the future." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.675, + 0.36, + 0.691 + ], + "angle": 0, + "content": "9 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.843 + ], + "angle": 0, + "content": "We introduced SceneScript, a novel reconstruction method that is based on a tokenized scene representation. SceneScript autoregressively predicts a sequence of structured scene language commands given a video stream of an indoor space. This tokenized scene representation is compact, editable and interpretable. In addition, we showed that a strength of SceneScript is the ability to extend to arbitrary elements (e.g. Bezier curves, object part decomposition) with minimal changes. This research opens up new directions in representing 3D scenes as language, bringing the 3D reconstruction community closer to the recent advances in large language models such as GPT-class models [20]." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.652, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SceneScript" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.786, + 0.232 + ], + "angle": 0, + "content": "1. Avetisyan, A., Khanova, T., Choy, C., Dash, D., Dai, A., Nießner, M.: Scenecad: Predicting object alignments and layouts in rgb-d scans. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16. pp. 596-612. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.233, + 0.786, + 0.273 + ], + "angle": 0, + "content": "2. Brazil, G., Kumar, A., Straub, J., Ravi, N., Johnson, J., Gkioxari, G.: Omni3D: A large benchmark and model for 3D object detection in the wild. In: CVPR. IEEE, Vancouver, Canada (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.274, + 0.786, + 0.314 + ], + "angle": 0, + "content": "3. Cabral, R., Furukawa, Y.: Piecewise planar and compact floorplan reconstruction from images. In: 2014 IEEE Conference on Computer Vision and Pattern Recognition (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.315, + 0.786, + 0.357 + ], + "angle": 0, + "content": "4. Carlier, A., Danelljan, M., Alahi, A., Timofte, R.: Deepsvg: A hierarchical generative network for vector graphics animation. In: Advances in Neural Information Processing Systems (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.357, + 0.786, + 0.398 + ], + "angle": 0, + "content": "5. Chen, J., Liu, C., Wu, J., Furukawa, Y.: Floor-sp: Inverse cad for floorplans by sequential room-wise shortest path. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.399, + 0.786, + 0.439 + ], + "angle": 0, + "content": "6. Chen, T., Saxena, S., Li, L., Fleet, D.J., Hinton, G.: Pix2seq: A language modeling framework for object detection. In: International Conference on Learning Representations (ICLR) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.44, + 0.786, + 0.481 + ], + "angle": 0, + "content": "7. Collins, J., Goel, S., Deng, K., Luthra, A., Xu, L., Gundogdu, E., Zhang, X., Yago Vicente, T.F., Dideriksen, T., Arora, H., Guillaumin, M., Malik, J.: Abo: Dataset and benchmarks for real-world 3d object understanding. CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.482, + 0.786, + 0.522 + ], + "angle": 0, + "content": "8. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proc. Computer Vision and Pattern Recognition (CVPR), IEEE (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.523, + 0.786, + 0.55 + ], + "angle": 0, + "content": "9. Furukawa, Y., Curless, B., Seitz, S.M., Szeliski, R.: Manhattan-world stereo. In: 2009 IEEE Conference on Computer Vision and Pattern Recognition (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.551, + 0.786, + 0.591 + ], + "angle": 0, + "content": "0. Furukawa, Y., Curless, B., Seitz, S.M., Szeliski, R.: Reconstructing building interiors from images. In: 2009 IEEE 12th international conference on computer vision (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.592, + 0.786, + 0.633 + ], + "angle": 0, + "content": "1. Ganin, Y., Bartunov, S., Li, Y., Keller, E., Saliceti, S.: Computer-aided design as language. Advances in Neural Information Processing Systems 34, 5885-5897 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.634, + 0.786, + 0.661 + ], + "angle": 0, + "content": "2. Girshick, R.: Fast r-cnn. In: Proceedings of the IEEE international conference on computer vision. pp. 1440-1448 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.661, + 0.786, + 0.702 + ], + "angle": 0, + "content": "3. Lee, C.Y., Badrinarayanan, V., Malisiewicz, T., Rabinovich, A.: Roomnet: End-to-end room layout estimation. In: Proceedings of the IEEE international conference on computer vision (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.703, + 0.753, + 0.716 + ], + "angle": 0, + "content": "4. Meta: Project aria. https://projectaria.com/ (2022), accessed: 2023-08-30" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.717, + 0.786, + 0.757 + ], + "angle": 0, + "content": "5. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.758, + 0.786, + 0.799 + ], + "angle": 0, + "content": "6. Misra, I., Girdhar, R., Joulin, A.: An end-to-end transformer model for 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2906-2917 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.786, + 0.84 + ], + "angle": 0, + "content": "7. Murali, S., Speciale, P., Oswald, M.R., Pollefeys, M.: Indoor scan2bim: Building information models of house interiors. In: 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) (2017)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.177, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.377, + 0.129 + ], + "angle": 0, + "content": "Avetisyan et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "18. Nash, C., Ganin, Y., Eslami, S.A., Battaglia, P.: Polygen: An autoregressive generative model of 3d meshes. In: International conference on machine learning. pp. 7220-7229. PMLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.231 + ], + "angle": 0, + "content": "19. Ochmann, S., Vock, R., Klein, R.: Automatic reconstruction of fully volumetric 3d building models from oriented point clouds. ISPRS journal of photogrammetry and remote sensing 151, 251-262 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.503, + 0.244 + ], + "angle": 0, + "content": "20. OpenAI: Gpt-4 technical report (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.785, + 0.286 + ], + "angle": 0, + "content": "21. Para, W., Bhat, S., Guerrero, P., Kelly, T., Mitra, N., Guibas, L.J., Wonka, P.: Sketchgen: Generating constrained cad sketches. In: Advances in Neural Information Processing Systems (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.288, + 0.785, + 0.327 + ], + "angle": 0, + "content": "22. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.329, + 0.785, + 0.37 + ], + "angle": 0, + "content": "23. Pintore, G., Agus, M., Gobbetti, E.: Atlantanet: inferring the 3d indoor layout from a single 360 image beyond the Manhattan world assumption. In: Proceedings of the European conference on computer vision (ECCV) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.424 + ], + "angle": 0, + "content": "24. Rukhovich, D., Vorontsova, A., Konushin, A.: Imvoxelnet: Image to voxels projection for monocular and multi-view general-purpose 3d object detection. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2397-2406 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.426, + 0.785, + 0.466 + ], + "angle": 0, + "content": "25. Somasundaram, K., Dong, J., Tang, H., Straub, J., Yan, M., Goesele, M., Engel, J.J., De Nardi, R., Newcombe, R.: Project aria: A new tool for egocentric multimodal ai research. arXiv preprint arXiv:2308.13561 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.494 + ], + "angle": 0, + "content": "26. Tang, H., Liu, Z., Li, X., Lin, Y., Han, S.: TorchSparse: Efficient Point Cloud Inference Engine. In: Conference on Machine Learning and Systems (MLSys) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.785, + 0.535 + ], + "angle": 0, + "content": "27. Tang, H., Liu, Z., Zhao, S., Lin, Y., Lin, J., Wang, H., Han, S.: Searching Efficient 3D Architectures with Sparse Point-Voxel Convolution. In: European Conference on Computer Vision (ECCV) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.537, + 0.785, + 0.577 + ], + "angle": 0, + "content": "28. Tulsiani, S., Su, H., Guibas, L.J., Efros, A.A., Malik, J.: Learning shape abstractions by assembling volumetric primitives. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.646 + ], + "angle": 0, + "content": "29. Tyszkiiewicz, M.J., Maninis, K.K., Popov, S., Ferrari, V.: Raytran: 3d pose estimation and shape reconstruction of multiple objects from videos with ray-traced transformers. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part X. pp. 211-228. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.673 + ], + "angle": 0, + "content": "30. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. NeurIPS 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.785, + 0.702 + ], + "angle": 0, + "content": "31. Vu, T., Kim, K., Luu, T.M., Nguyen, X.T., Yoo, C.D.: Softgroup for 3d instance segmentation on 3d point clouds. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.703, + 0.785, + 0.743 + ], + "angle": 0, + "content": "32. Wu, R., Xiao, C., Zheng, C.: Deepcad: A deep generative network for computer-aided design models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.785, + 0.798 + ], + "angle": 0, + "content": "33. Xu, X., Willis, K.D., Lambourne, J.G., Cheng, C.Y., Jayaraman, P.K., Furukawa, Y.: Skexgen: Autoregressive generation of cad construction sequences with disentangled codebooks. In: International Conference on Machine Learning (ICML) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "34. Xue, Y., Mao, J., Niu, M., Xu, H., Mi, M.B., Zhang, W., Wang, X., Wang, X.: Point2seq: Detecting 3d objects as sequences. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8521-8530 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.652, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SceneScript" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "35. Yang, K., Chen, X.: Unsupervised learning for cuboid shape abstraction via joint segmentation from point clouds. ACM Transactions on Graphics (TOG) 40(4), 1-11 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.231 + ], + "angle": 0, + "content": "36. Yin, T., Zhou, X., Krahenbuhl, P.: Center-based 3d object detection and tracking. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.232, + 0.788, + 0.273 + ], + "angle": 0, + "content": "37. Yue, Y., Kontogianni, T., Schindler, K., Engelmann, F.: Connecting the dots: Floorplan reconstruction using two-level queries. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.273, + 0.788, + 0.315 + ], + "angle": 0, + "content": "38. Zou, C., Colburn, A., Shan, Q., Hoiem, D.: Layoutnet: Reconstructing the 3d room layout from a single rgb image. In: Proceedings of the IEEE conference on computer vision and pattern recognition (2018)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.315 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/SceneScript_ Reconstructing Scenes With An Autoregressive Structured Language Model/6d586386-d370-4aeb-a27d-9451b501b037_origin.pdf b/2024/SceneScript_ Reconstructing Scenes With An Autoregressive Structured Language Model/6d586386-d370-4aeb-a27d-9451b501b037_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d8a31f7e07f817212938b072aa61bb655d3697dd --- /dev/null +++ b/2024/SceneScript_ Reconstructing Scenes With An Autoregressive Structured Language Model/6d586386-d370-4aeb-a27d-9451b501b037_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:76dbc651c5a54e48596f9f7af536a21d4cb374439bf4ec94e515fe619d1a1127 +size 7428007 diff --git a/2024/SceneScript_ Reconstructing Scenes With An Autoregressive Structured Language Model/full.md b/2024/SceneScript_ Reconstructing Scenes With An Autoregressive Structured Language Model/full.md new file mode 100644 index 0000000000000000000000000000000000000000..167a7bdf650f7e45238309842eb1826a0717e4bc --- /dev/null +++ b/2024/SceneScript_ Reconstructing Scenes With An Autoregressive Structured Language Model/full.md @@ -0,0 +1,315 @@ +# SceneScript: Reconstructing Scenes With An Autoregressive Structured Language Model + +Armen Avetisyan1, Christopher Xie1, Henry Howard-Jenkins1, Tsun-Yi Yang1, Samir Aroudj1, Suvam Patra1, Fuyang Zhang2†, Duncan Frost1, Luke Holland1, Campbell Orme1, Jakob Engel1, Edward Miller1, Richard Newcombe1, and Vasileios Balntas1 + +Meta Reality Labs Simon Fraser University https://projectaria.com/scenescript + +![](images/b5c2a70ef1294fc3992a7dc3fdcaccd62cf263fd13e44c1cc0c942d24d45769f.jpg) +Scene walkthrough + +![](images/ef23cd8bcb4a5c2c1c1a1f60d8901c3cd6b9710c65b21064e5194dc91757e526.jpg) +# + +![](images/0cb693aa5315d4bc1a3e0bf93e5a0e9a98bc6b360a4e1941ee206c96fa8328e4.jpg) + +![](images/b5e369cc3b6ed30fdeab00dfbfa8c06ad6193df51b8b6dcc2b5710698a0c791b.jpg) +Language to geometry interpreter + +![](images/d53ad6538b00b5136035b7acfddfca2ec26dcb7d507edbd4acf542350e904719.jpg) + +![](images/6cd6225ceb1d8f51b78851bed857cb49c69da71aa893bcc20e79194679c9c09e.jpg) +Fig. 1: (top) Given an egocentric video of an environment, SceneScript directly predicts a 3D scene representation consisting of structured scene language commands. (bottom) Our method generalizes on diverse real scenes while being solely trained on synthetic indoor environments. (last column, bottom) A notable advantage of our method is its capacity to easily adapt the structured language to represent novel scene entities. For example, by introducing a single new command, SceneScript can directly predict object parts jointly with the layout and bounding boxes. + +![](images/e6643aea7d8f149ea5b8d08e7e985345a028102d0b636820bb272dd4b1da4057.jpg) + +![](images/dd64e2665a6f9a6001da7faa4dcccd2412e132e6e9dc1d03409a84949dd48f23.jpg) + +![](images/8c5baab5c2e2956dd9cfd5ea0a849af2e9c09c69c31da3c631169fae8ac4f47a.jpg) +Structured scene language + +![](images/7d687e8dcf3da4e25121307cdc4481724f82426f8c04e4b66f0a19e81503f9f7.jpg) + +![](images/fc5afea0199592b657ee3cd9cddaf74bb01eb6309d2db95699ec84d92528365c.jpg) + +![](images/6f333f99c8811e8a9f9637f8934144af7678470a1981225f2cf9e8ff6d593fdc.jpg) + +![](images/4afa5e094cf97862c1ee0bcc43375ffece58e4f15bd994568b8095efb6bcbe1a.jpg) + +![](images/4fb9354133c22f600953d0c8f0537cf2a881486c6e2fb8b3a38bdb40cd53829e.jpg) + +Abstract. We introduce SceneScript, a method that directly produces full scene models as a sequence of structured language commands using an autoregressive, token-based approach. Our proposed scene representation is inspired by recent successes in transformers & LLMs, and departs from more traditional methods which commonly describe scenes as meshes, voxel grids, point clouds or radiance fields. Our method infers the set of structured language commands directly from encoded visual data using a scene language encoder-decoder architecture. To train SceneScript, we generate and release a large-scale synthetic dataset called Aria Synthetic Environments consisting of $100k$ high-quality indoor scenes, with photorealistic and ground-truth annotated renders of egocentric scene walkthroughs. Our method gives state-of-the-art results in architectural layout estimation, and competitive results in 3D object detection. Lastly, we explore an advantage for SceneScript, which is the ability to readily adapt to new commands via simple additions to the structured language, which we illustrate for tasks such as coarse 3D object part reconstruction. + +# 1 Introduction + +Scene representations play a crucial role in machine learning and computer vision applications, enabling accurate understanding of the environment. Over the years, researchers have explored various options such as meshes, voxel grids, point clouds, and implicit representations, aiming to represent complex real-world scenes with high-fidelity. Each of these exhibits distinct advantages and limitations that impact their suitability for different tasks. Meshes offer detailed geometric information but can be expensive in both computation and memory. Voxel grids provide a volumetric representation but suffer from a trade-off between resolution and memory requirements. Point clouds are efficient in representing sparse scenes but lack semantics and explicit connectivity information. Implicit representations, such as DeepSDF [22] and NeRF [15], can be infinitely precise but lack interpretability and editability. The selection of an appropriate representation directly impacts the performance and efficacy of various tasks including object recognition, scene understanding, and 3D reconstruction. In this paper, we propose a novel scene representation based on structured language commands as a more efficient and versatile solution. + +Our motivation stems from the recent advancements in the field of Large Language Models (LLMs) and "next token prediction" autoregressive methods [20], coupled with recent works on exploring generation of sequences to represent geometric structures. For example, PolyGen [18] demonstrated the ability to describe 3D meshes as a sequence of vertices and faces generated using transformers [30]. Similarly, CAD-as-Language [11] showcased the effectiveness of generating Computer-Aided Design (CAD) primitives to represent 2D CAD sketches. Our main goal is to directly infer a metrically accurate representation of a full scene as a text-based sequence of specialized structured language commands. + +Our method, denoted SceneScript, autoregressively predicts a language of hand-designed structured language commands in pure text form. This language offers several distinct advantages: 1) As pure text, it is compact and reduces memory requirements of a large scene to only a few bytes. 2) It is crisp and complete since the commands are designed to result in sharp and well-defined geometry (similar to scalable vector graphics). 3) It is interpretable, editable and semantically rich by design via the use of high-level parametric commands such as make_door(*door_parameters). 4) It can seamlessly integrate novel geometric entities by simply adding new structured commands to the language. 5) The fact that the scene representation is a series of language tokens similar to [20] opens up a plethora of potential new applications in the future such as ways to edit the scene, query it, or spin up chat interactions. + +We mainly focus on the problems of architectural layout estimation and object detection as proxy tasks for the efficacy of our SceneScript language as a scene representation. Architectural entities such as walls, doors, and windows are highly structured entities, making them an ideal test-bed. However, one notable drawback of language models is that they require vast amounts of data for training. Since there is no existing dataset of scene walkthroughs and their cor + +responding structured language commands, we publicly released Aria Synthetic Environments (ASE), a synthetically generated dataset of 100k unique interior scenes. For each scene, we simulate egocentric trajectories with an entire suite of sensor data from Project Aria [14]. We also release multiple sources of ground truth including depth and instance segmentations. Importantly, for each rendered egocentric sequence, the architectural layout ground truth is given in our proposed SceneScript language. + +While architectural layout serves as a test-bed, we demonstrate that our method SceneScript can easily be extended to new tasks via simple extensions to SceneScript language while keeping both the visual input and network architecture fixed. We illustrate this on the problem of 3D object detection, which results in a method that jointly infers architectural layout and 3D oriented bounding boxes. Additionally, we demonstrate more proof-of-concept experiments that show that our method results in a significantly lower barrier to entry for new tasks including representing coarse 3D object reconstruction, curved entities, composition of entities, and entity states. + +Our core contributions are: + +- We propose SceneScript, a method that jointly predicts architectural layout and object bounding boxes of a scene in the form of structured language commands given a video stream. +- We demonstrate that SceneScript can easily be extended to completely new tasks with simple additions of commands to our structured language, significantly lowering the barrier to entry for new tasks. +- We release a large-scale synthetic dataset, named Aria Synthetic Environments, comprised of 100k unique high-quality 3D indoor scenes with GT, which will enable large scale ML training of scene understanding methods. +- We show that training SceneScript on Aria Synthetic Environments leads to generalization on real scenes (see videos/demos on the project page). + +# 2 Related Works + +# 2.1 Layout Estimation + +Layout estimation is an active research area, aiming to infer architectural elements. Scan2BIM [17] proposes heuristics for wall detection to produce 2D floorplans. Ochmann et al. [19] formulate layout inference as an integer linear program using constraints on detected walls. Shortest path algorithms around birds-eye view (BEV) free space [3] and wall plus room instance segmentation [5] have also been explored. + +Furukawa et al. [10] utilize a Manhattan world-based multi-view stereo algorithm [9] to merge axis-aligned depth maps into a full 3D mesh of building interiors. RoomNet [13] predicts layout keypoints while assuming that a fixed set of Manhattan layouts can occur in a single image of a room. LayoutNet [38] improves on this by predicting keypoints and optimising the Manhattan room layout inferred from them. Similarly, AtlantaNet [23] predicts a BEV floor or + +Table 1: Complete set of structured language commands designed for detailing architectural layouts and object bounding boxes. Supported data types can include int, float, bool. It is important to note that the language's extensibility allows for easy augmentation by introducing new commands like make_prime, make_pillar, or enhancing existing commands, such as incorporating is-double_door (bool). + +
make_WALL (int)make_door (int)make_window (int)make_bbox (int)
idintidintidintidint
a_xfloatwall0_idintwall0_idintclassint
a_yfloatwall1_idintwall1_idintposition_xfloat
a_zfloatposition_xfloatposition_xfloatposition_yfloat
b_xfloatposition_yfloatposition_yfloatposition_zfloat
b_yfloatposition_zfloatposition_zfloatangle_zfloat
b_zfloatwidthfloatwidthfloatscale_xfloat
heightfloatheightfloatheightfloatscale_yfloat
scale_zfloat
+ +ceiling shape and approximates the shape contour with a polygon resulting in an Atlanta world prior. SceneCAD [1] uses a graph neural network to predict object-object and object-query relationships to refine its layout prediction. + +Our approach stands out by requiring neither heuristics nor explicitly defined prior knowledge about architectural scene structure. In fact, our method demonstrates geometric understanding of the scene which emerges despite learning to predict the GT scene language as a sequence of text tokens. + +# 2.2 Geometric Sequence Modelling + +Recent works have explored transformers for generating objects as text-based sequences. PolyGen [18] models 3D meshes as a sequence of vertices and faces. CAD-as-Language [11] represents 2D CAD sketches as a sequence of triplets in protobuf representation, followed by a sequence of constraints. Both SketchGen [21] and SkexGen [33] use transformers to generate sketches. DeepSVG [4] learns a transformer-based variational autoencoder (VAE) that is capable of generating and interpolating through 2D vector graphics images. DeepCAD [32] proposes a low-level language and architecture similarly to DeepSVG, but applies it to 3D CAD models instead of 2D vector graphics. Our approach stands out by utilising high-level commands, offering interpretability and semantic richness. Additionally, while low-level commands can represent arbitrarily complex geometries, they lead to prohibitively longer sequences when representing a full scene. + +The closest work to ours is Pix2Seq [6]. Pix2Seq proposes a similar architecture to ours but experiments only with 2D object detection, thus requiring domain-specific augmentation strategies. Another closely related work is Point2Seq [34] that trains a recurrent network for autoregressively regressing continuous 3D bounding box parameters. Interestingly, they find the autoregressive ordering of parameters outperforms current standards for object detection architectures, including anchors [12] and centers [36]. + +![](images/1a1d1b552808797c3bc3fddf81646e7d68331207ca166790f8def6c7738630dc.jpg) +Fig. 2: Aria Synthetic Environments: (top) Random samples of generated scenes showing diversity of layouts, lights and object placements. (bottom - left to right) A top down view of a scene filled with objects, a simulated trajectory (blue path), renderings of depth, RGB, and object instances, and lastly a scene pointcloud. + +# 3 SceneScript Structured Language Commands + +We first describe our structured language commands that define a full scene representation including both layout and objects. After this, we introduce our corresponding large scale training dataset: Aria Synthetic Environments. + +# 3.1 Commands and Parameters + +We begin with a parameterization that captures the most common layout elements. For this purpose we use three commands: make-wall, make_door, and make_window. Each command comes with a set of parameters that results in well-defined geometry. For example, the full set of parameters for make-wall specifies a gravity-aligned 2D plane, while make_door and make_window specify box-based cutouts from walls. It is worth noting that this parametrization is arbitrary, and is only made in the context of presenting a proof-of-concept SceneScript system. There are infinitely many parametrization schemes, in this work we opt for one that prioritizes ease of use and research iteration speed. + +In addition to representing these three major layout entities, we aim to jointly infer objects as oriented bounding boxes. Thus, we introduce a fourth command: + +```python +make_bbox: id, class, position_x, position_y, position_z, angle_z, scale_x, scale_y, scale_z +``` + +This simple parametrization represents an oriented 3D bounding box that is assumed to be aligned with gravity (assuming it points in the $-z$ direction). A summary of these commands and their respective parameters are shown in Table 1. + +While we have described just four commands to capture structure and objects in an indoor environment, importantly, this text-based parametrization can readily be extended geometrically and/or even semantically to include states or other functional aspects. For example, changing the mentioned makedoor command to include parameters such as open-degree allows the language to represent door states. In Section 6, we demonstrate how such extensions to the language allows for SceneScript to readily adapt to new tasks including coarse 3D object reconstruction. + +# 3.2 Scene Definition + +A single scene can be described as a sequence of our proposed structured language commands. The sequence requires no specific ordering, and can be of arbitrary length. From this definition, the 3D scene can easily be obtained by parsing the commands through a simple custom interpreter. + +# 3.3 Training Dataset + +To enable practical indoor scene reconstruction based on our structured language commands (Section 3.1), we publicly released a new dataset called Aria Synthetic Environments. It consists of a large number of training pairs of egocentric scene walkthroughs linked with corresponding ground truth command sequences. + +Since transformers require vast amounts of data, we generated $100k$ synthetic scenes, which is in comparison infeasible for real-world data. Each synthetic scene comes with a floor plan model, a corresponding complete 3D scene as well as a simulated agent trajectory and a photo-realistic rendering of this trajectory. Fig. 2 illustrates the basics of Aria Synthetic Environments. For brevity, we refer the reader to the supplemental material for further details. + +# 4 SceneScript Network Architecture + +Our pipeline is a simple encoder-decoder architecture that consumes a video sequence and returns SceneScript language in a tokenized format. Figure 3 illustrates a high-level overview of our method. + +We examine three encoder variants: a pointcloud encoder, a posed image set encoder, and a combined encoder. The decoder remains the same in all cases. + +# 4.1 Input Modalities and Encoders + +The encoder computes a latent scene code in the form of a 1D sequence from video walkthrough of the scene. The decoder is designed to consume these 1D sequences as input. This enables the integration of various input modalities within a unified framework. As a preliminary, for each scene we assume access to a set of $M$ posed camera images $\{\mathbf{I}_1,\dots,\mathbf{I}_{\mathbf{M}}\}$ , e.g., SLAM output. + +![](images/c3b7e7ec3fe381af7c3e1f2cad895f41c4683d131632eaee1af83a60648d7867.jpg) +Fig. 3: SceneScript core pipeline overview. Raw images & pointcloud data are encoded into a latent code, which is then autoregressively decoded into a sequence of commands that describe the scene. Visualizations are shown using a customly built interpreter. Note that for the results in this paper, the the point clouds are computed from the images using Aria MPS [25] - i.e. are not using a dedicated RGB-D / Lidar sensor. + +![](images/2dea9288ccf974cbfed2a66b1452436a7926daa4a216cbdccb1de8553d1dd53c.jpg) + +Point Clouds. A point cloud $\mathbf{P} = \{\mathbf{p}_1, \dots, \mathbf{p_N}\}$ consists of $N$ points, where $\mathbf{p}_i$ is a 3D point coordinate. It can come from passive images using SLAM or SfM, or RGB-D / Lidar sensors. + +Specifically, we use the Semi-dense Pointclouds from Project Aria's Machine Perception Services [25], that are obtained from a visual-inertial SLAM system using Aria's monochrome cameras and IMUs. We discretize the point cloud to $5\mathrm{cm}$ resolution, then employ a sparse 3D convolution library [26,27] to generate pooled features. The encoder $\mathcal{E}_{geo}$ applies a series of down convolutions, resulting in a reduction of the number of points in the lowest level. + +$$ +\mathbf {F} _ {g e o} = \mathcal {E} _ {g e o} (\mathbf {P}), \quad \mathbf {P} \in \mathbb {R} ^ {N \times 3}, \mathbf {F} _ {g e o} \in \mathbb {R} ^ {K \times 5 1 2} \tag {1} +$$ + +where $K \ll N$ . $\mathbf{F}_{geo}$ is a condensed latent representation of the point cloud that contains the necessary scene context. For later use in the transformer decoder, we treat $\mathbf{F}_{geo}$ as a sequence of feature vectors where the entries $\mathbf{f_i}$ , $i \in 1\dots K$ are sorted lexicographically according to the coordinate of the active site $\mathbf{c_i}$ , $i \in 1\dots K$ . To incorporate positional encoding, we append the coordinates of the active sites to their respective feature vectors $\mathbf{f}_i \gets \mathbf{cat}(\mathbf{f_i},\mathbf{c_i})$ . + +Point Clouds with Lifted Features. We additionally explore augmenting the point cloud with image features. From the original egocentric sequence and associated trajectory, we sample a set of $M$ keyframes, $\mathbf{I}_i$ where $i\in 1\dots M$ , and compute a set of image features $\mathbf{F}_i$ for each. We then project each point into the set of keyframe cameras and retrieve the feature vector (output by a CNN) at the pixel location: + +$$ +\mathbf {f} _ {i p} = F _ {i} (\pi (\mathbf {p})) \quad \mathbf {p} \in \mathbf {P}, i \in 1 \dots M, \tag {2} +$$ + +where $\pi(\cdot)$ represents the projection function of a 3D point into the camera. If $\pi(\mathbf{p})$ falls outside the image bounds, no feature is retrieved. We combine the set + +of lifted features for a point through an average, resulting in a single feature vector for each point: $\mathbf{f}_p = 1 / M\sum_{i = 1}^{M}\mathbf{f}_{ip}$ + +We form our lifted-feature point cloud, $\mathbf{P}' = \{\mathbf{p}_1',\dots,\mathbf{p}_{\mathbf{M}}'\}$ , by concatenating each point's lifted feature with the original XYZ location: $\mathbf{p}' = \mathrm{cat}(\mathbf{f}_p,\mathbf{p})$ . $\mathbf{P}'$ is then encoded into a context sequence using sparse 3D convolutions, with only the input feature channels adjusted to match the new point feature dimension. + +End-to-end Encoding of Posed Views. In order to encode the egocentric sequence more directly without a pre-computed point cloud, we adopt a $2\mathrm{D} \leftrightarrow 3\mathrm{D}$ bidirectional transformer encoder following the form defined in RayTran [29]. + +In this formulation, we initialize a volume representation of the scene as a dense voxel grid of features, $\mathbf{V}$ , that coincides with the scene geometry. In turn, we sample a subset of $M$ keyframes, $\mathbf{I}_i$ where $i\in 1\dots M$ , from the full stream of posed images. And for each of these keyframes we compute image features from a CNN, $\mathbf{F}_i$ . Repeated layers of bidirectional attention enable the image and voxel grid features to be refined iteratively in successive transformer blocks through the aggregation of view-point and global-scene information. As in RayTran [29], the interaction between the two representations is guided by the image-formation process by leveraging known camera parameters and poses. Attention in these transformer blocks is restricted by patch-voxel ray intersections, where each image patch attends to the voxels it observes and each voxel location attends to all the patches that observe it. The resulting voxel grid of features is flattened, concatenated with an encoded representation of its XYZ location, and passed to the decoder. + +# 4.2 Language Decoder + +We utilize a transformer decoder [30] to decode the scene latent code into a sequence of structured language commands. The sequence of tokens passes through an embedding layer, followed by a positional encoding layer. Together with the encoded scene code (Section 4.1), the embedded tokens are passed into the several transformer decoder layers where a causal attention mask is used to ensure autoregressive generation. More implementation details can be found in Appendix ?? + +# 4.3 Language Tokenization + +We refer to the serialization the structured language into a sequence of tokens as tokenization. The goal is to construct a bijective mapping between a sequence of structured language commands (Section 3) and a sequence of integer tokens that can be predicted by the transformer decoder architecture. We utilize the following schema: + +For example, a sample sequence for a makedoor command may look like: + +[START, PART, MAKE_DOOR, POSITION_X, POSITION_Y, POSITION_Z, WALLO_IDX, WALL1_IDX, WIDTH, HEIGHT, PART,..., STOP] + +This schema enables 1D packing of tokens without requiring fixed-size slots or padding like other sequence modelling methods such as [32]. Additionally, it does not impose any limitations on the number or the hierarchy of sub-sequences, as they are flexibly separated by a PART token. This allows for arbitrarily complex scene representations. + +The tokenized sequence is discretized into integers at a 5cm resolution, then translated into a sequence of embeddings via learnable lookup table. Note that by designing the SceneScript language, we also design the tokenization. This tokenization scheme is notably different from standard NLP tokenization, which involves Byte-Pair Encodings (BPE) [20]. + +# 5 Results + +In this section, we introduce the metrics we define to measure performance, and we discuss some qualitative and quantitative results that give insights to our proposed SceneScript method. + +# 5.1 Metrics + +To evaluate accuracy of the layout estimation, we make use of geometric metrics applied between the ground-truth room layout and our predicted SceneScript language. To do so, we define an entity distance, $d_{E}$ , between a pair of entities of the same class. Each entity, $E$ , is represented as a 3D plane segment comprising of 4 corners $\{c_1, c_2, c_3, c_4\}$ . The distance between two entities, $E$ and $E'$ is computed as the maximum Euclidean distance between each corner and its counterpart assigned via Hungarian mathcing, i.e.: $d_{E}(E, E') = \max \{||c_{i} - c_{\pi(i)}'| : i = 1, \dots, 4\}$ , where $\pi(i)$ is the permutation also found via Hungarian matching. + +We threshold $d_{E}$ to define the success criteria for the predicted entities. We compute the following metrics: + +- F1 Score @ threshold - the F1 score of the set of predictions is computed at a single $d_{E}$ threshold. +- Average F1 Score - the F1 score is computed across a range of entity distance thresholds and averaged. + +The scores are computed for each class independently and averaged to overall score. In addition, scores are computed for each scene and averaged across the dataset. We use the following range of thresholds (cm) for the average F1 scores: $T = \{1, 2, \dots, 9, 10, 15, 25, 30, 50, 75, 100\}$ . + +![](images/9c35d5d28bc1dbf8b979708ca1fbc4ae07d8432fc5235dbeec91b9e95e5ec278.jpg) +Fig. 4: Qualitative samples between our model and SOTA methods on Aria Synthetic Environments's test set. Hierarchical methods like SceneCAD suffer from error cascading which leads to missing elements in the edge prediction module. RoomFormer (a 2D method extruded to 3D) primarily suffers from lightly captured scene regions which leave a unnoticeable signal in the density map. + +Table 2: Layout Estimation on Aria Synthetic Environments Quantitative comparison between our method and related recent work. + +
MethodF1 @5cmAvg F1
meanwalldoorwindowmeanwalldoorwindow
SceneCAD '20 [1]-0.048---0.275--
RoomFormer '23 [37]0.1390.1590.1480.1100.4640.5050.4810.407
Ours (Point cloud)0.8480.9300.9220.6920.7840.8160.8110.724
Ours (Lifted features)0.9030.9430.9590.8060.8010.8180.8220.764
Ours (Image-only)0.6610.6870.7980.4970.7190.7270.7720.658
+ +# 5.2 Layout Estimation + +We perform scene layout estimation with the three encoder variants of SceneScript: a baseline model with sparse3d convolution pointcloud encoder, an RGB RayTran-based feature volume encoder [29], and our proposed lifted feature point encoder. The same transformer decoder is being used in all three scenarios. + +To provide comparison to existing works, we include results from two baseline methods, namely SceneCAD [1] and the recent RoomFormer [37]. For these experiments SceneCAD and RoomFormer were both trained on Aria Synthetic Environments. Note that SceneCAD only predicts walls. + +Table 2 shows the main results for our F1-based metrics on Aria Synthetic Environments. SceneScript exhibits a substantial performance advantage over SOTA layout estimation baselines across multiple metrics. + +Table 3: 3D Object Detection Performance comparison against state-of-the-art methods on an 3D object detection task trained and evaluated by F1-score at 0.25 and 0.5 IoU thresholds. By simply adding a make_bbox command SceneScript can achieve competitive object detection results. +(a) Aria Synthetic Environments + +
MethodInputF1@.25 IoU @.50 IoU
3DETR '21 [16]Points0.2010.078
Cube R-CNN '23 [2]RGB0.3940.228
ImVoxelNet '22 [24]RGB0.5840.516
OursPoints0.6200.577
+ +(b) ScanNet [8] + +
MethodInputF1
@.25 IoU@.50 IoU
3DETR '21 [16]Points0.4800.349
3DETR-m '21 [16]Points0.5360.407
SoftGroup '22 [31]RGB Points0.6220.573
OursRGB Points0.5060.406
+ +Both baseline methods encounter a significant decline in accuracy when dealing with finer details. See Figure 4 for qualitative comparisons between our method and baseline methods. + +Encoder Ablation. The results demonstrate that SceneScript is robust to the encoding strategy chosen to encode the egocentric capture. It is able to infer highly accurate scene layouts in all configurations tested, and in each case SceneScript outperforms the included baselines by a significant margin. + +Relative comparison of the encoder strategies reveals that leveraging the point clouds from a highly specialized and well-tuned system is still able to offer advantages of an, almost, entirely learned approach such as RayTran [29]. A light extension in the form of lifted image features can widen this gap even further. In particular, we observe that the discrepancy between the encoders becomes particularly apparent as the complexity of the scene increases in the form of increased room count. In the Appendix, we also show a quantitative evaluation of per-entity error distances, which aids in further attribution of relative performance gains between the encoding methods. + +# 5.3 Object Detection + +In this section, we perform evaluation of SceneScript for object detection on both Aria Synthetic Environments and ScanNet [8]. For comparison, we include recent and state-of-the-art baseline methods, namely Cube-RCNN [2], ImVoxelNet [24], 3DETR [16], and SoftGroup [31]. + +Worth noting is that SceneScript does not predict a confidence score for each make_bbox command. This means that fair computation of the conventional mAP metric for this task is not possible, as detections cannot be ranked across scenes. Among other issues, this results in a metric that varies with the order in which scenes are evaluated. We therefore report F1-score-based metrics, which do not exhibit this order variance. Further discussion of this, and mAP numbers for the baselines for reference, can be found in the Appendix. + +![](images/814bbad6fa1cc611ca4aec168804d9a106aba6b1f4305b699bb40cb1ea365dc7.jpg) +Fig. 5: Example scene reconstructions on scenes from Aria Synthetic Environments. (left) Visualisation of the decomposed meshes used to create make_prim training pairs. (right) Views of full scene predictions, as well as close ups highlighting the fidelity of object reconstruction through the prediction volumetric primitives enabled by make_prim. + +![](images/3f9db45871fddf264da288563194d4b18fcf1c128a19683f51f9876000c4d842.jpg) + +In Table 3a, all methods were trained on Aria Synthetic Environments. 3DETR performs poorly due to its encoder being designed for relatively clean datasets such as ScanNet [8], while semi-dense point clouds from Aria Synthetic Environments exhibit more noise and less uniformity (see Figure 3 for an example). Cube R-CNN and ImVoxelNet both operate on RGB images, and detections are tracked for the entire sequence via a tracker [2] to provide competitive performance. In Table 3b, our method provides similar performance to both 3DETR and SoftGroup. + +Through the addition of the make_bbox command, SceneScript demonstrates object detection performance on par with SOTA baselines on multiple object detection datasets. This result illustrates the extensibility of a token-based approach, and that our proposed SceneScript language representation does not suffer compared to specialised object detection network architectures. + +# 6 Extending the SceneScript Structured Language + +A key advantage offered by SceneScript's structured language prediction paradigm is that the expressiveness of its reconstruction can be tailored without requiring a change to the method. Up to now, we have focussed on showcasing the efficacy of SceneScript for representing simple layout elements and objects as bounding boxes. In this section, we showcase this characteristic by increasing the fidelity of our scene representation by introducing coarse 3D object reconstruction. + +# 6.1 Objects as Volumetric Primitives + +We turn to a language based on volumetric primitives, motivated by works such as [28, 35]. Using simple primitives such as cuboids and extruded cylinders, enables us to coarsely represent arbitrary object categories while maintaining object semantics (e.g. tabletops can be represented by a single cuboid). Thus, this language can describe many object categories simultaneously. + +![](images/07e2ba398b0e28618caf4ace0b9efe68e8dc5331637f65fc02e672324a425a48.jpg) +Fig.6: Example scene reconstructions on real scenes with the addition of the make_prim command. Note that SceneScript was trained only on synthetic data. + +This representation requires only one additional command over the layout and box commands already discussed previously, namely: + +```txt +make_prime:_bbox_id, prim_num, class, center_x, center_y, center_z, angle_x, angle_y, angle_z, scale_x, scale_y, scale_z +``` + +This command and its parameters describe a volumetric primitive (cuboid or extruded cylinder) via its 3D center, 3D rotation, and 3D scale. The prim_num parameter can be associated with semantics, e.g. tabletops of different tables typically have the same prim_num. + +Dataset. To obtain ground truth make_prim commands that align with the objects in Aria Synthetic Environments, we first run an extension of Yang et al. [35] to obtain cuboid and extruded cylinder primitives of a database of 3D CAD models (ABO [7], which was used to populate Aria Synthetic Environments). See Figure 5 (left) for example decompositions. For this proof-of-concept experiment, we use three categories: table, chair, and sofa. We then convert these decomposed primitives into make_prim commands that are aligned with the objects in the dataset, which results in training pairs. + +Results. We show qualitative results on Aria Synthetic Environments in Figure 5. In Figure 6, we show inferences in a few real-world environments despite only having trained on our simulated dataset. These results demonstrate that SceneScript's general purpose architecture is capable of coarsely reconstructing objects of multiple categories through addition of a new command. + +# 6.2 Further Extensions + +In this section, we have explored just one extension of SceneScript's structured scene language in order to demonstrate its flexibility. The result was greatly increased expressiveness of the scene reconstruction. In the Appendix, we include + +additional explorations that can further increase the fidelity and accuracy of the reconstructions. These explorations include reconstructing curved walls, inferring object states (e.g., door opening angles), as well as direct prediction of parametric models using object models deployed commonly by tech artists (e.g., blender geometry nodes) for object reconstruction. + +# 7 Interactive Scene Reconstruction + +Scene reconstruction often occurs offline, on pre-recorded walkthroughs of an environment, or derivatives of them, such as a point cloud. However, we take advantage of SceneScript's inference occurring at an interactive rate, which takes between 2-5s depending on the size of the environment, by implementing streaming of SceneScript's live reconstructions into a modern VR headset. This enables an interactive experience where the user can see the reconstruction overlaid onto the environment they are exploring in real-time. See the video recording on the website for examples. Visualisations of this interface are included in Figure 1. + +# 8 Limitations and Future Work + +SceneScript exhibits certain limitations that should be acknowledged. First, the structured language commands are manually defined, which requires human intervention at this stage. Secondly, due to the higher-level nature of our commands, it can be challenging to capture fine-grained geometric details with extremely high precision (i.e. mm). As a consequence, the resulting reconstructions based on this representation tend to lead to simpler and coarser geometries, potentially missing intricate nuances at the very high detail level. These limitations potentially highlight areas for future research and optimization, aiming to automate the command definition process and explore techniques to enhance the representation's ability to capture intricate geometric details accurately. However, we believe that the ability to built scene representations that are based on structured language commands will be a key part in complex and efficient methods of scene reconstruction in the future. + +# 9 Conclusion + +We introduced SceneScript, a novel reconstruction method that is based on a tokenized scene representation. SceneScript autoregressively predicts a sequence of structured scene language commands given a video stream of an indoor space. This tokenized scene representation is compact, editable and interpretable. In addition, we showed that a strength of SceneScript is the ability to extend to arbitrary elements (e.g. Bezier curves, object part decomposition) with minimal changes. This research opens up new directions in representing 3D scenes as language, bringing the 3D reconstruction community closer to the recent advances in large language models such as GPT-class models [20]. + +# References + +1. Avetisyan, A., Khanova, T., Choy, C., Dash, D., Dai, A., Nießner, M.: Scenecad: Predicting object alignments and layouts in rgb-d scans. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16. pp. 596-612. Springer (2020) +2. Brazil, G., Kumar, A., Straub, J., Ravi, N., Johnson, J., Gkioxari, G.: Omni3D: A large benchmark and model for 3D object detection in the wild. In: CVPR. IEEE, Vancouver, Canada (June 2023) +3. Cabral, R., Furukawa, Y.: Piecewise planar and compact floorplan reconstruction from images. In: 2014 IEEE Conference on Computer Vision and Pattern Recognition (2014) +4. Carlier, A., Danelljan, M., Alahi, A., Timofte, R.: Deepsvg: A hierarchical generative network for vector graphics animation. In: Advances in Neural Information Processing Systems (2020) +5. Chen, J., Liu, C., Wu, J., Furukawa, Y.: Floor-sp: Inverse cad for floorplans by sequential room-wise shortest path. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2019) +6. Chen, T., Saxena, S., Li, L., Fleet, D.J., Hinton, G.: Pix2seq: A language modeling framework for object detection. In: International Conference on Learning Representations (ICLR) (2022) +7. Collins, J., Goel, S., Deng, K., Luthra, A., Xu, L., Gundogdu, E., Zhang, X., Yago Vicente, T.F., Dideriksen, T., Arora, H., Guillaumin, M., Malik, J.: Abo: Dataset and benchmarks for real-world 3d object understanding. CVPR (2022) +8. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proc. Computer Vision and Pattern Recognition (CVPR), IEEE (2017) +9. Furukawa, Y., Curless, B., Seitz, S.M., Szeliski, R.: Manhattan-world stereo. In: 2009 IEEE Conference on Computer Vision and Pattern Recognition (2009) +0. Furukawa, Y., Curless, B., Seitz, S.M., Szeliski, R.: Reconstructing building interiors from images. In: 2009 IEEE 12th international conference on computer vision (2009) +1. Ganin, Y., Bartunov, S., Li, Y., Keller, E., Saliceti, S.: Computer-aided design as language. Advances in Neural Information Processing Systems 34, 5885-5897 (2021) +2. Girshick, R.: Fast r-cnn. In: Proceedings of the IEEE international conference on computer vision. pp. 1440-1448 (2015) +3. Lee, C.Y., Badrinarayanan, V., Malisiewicz, T., Rabinovich, A.: Roomnet: End-to-end room layout estimation. In: Proceedings of the IEEE international conference on computer vision (2017) +4. Meta: Project aria. https://projectaria.com/ (2022), accessed: 2023-08-30 +5. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021) +6. Misra, I., Girdhar, R., Joulin, A.: An end-to-end transformer model for 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2906-2917 (2021) +7. Murali, S., Speciale, P., Oswald, M.R., Pollefeys, M.: Indoor scan2bim: Building information models of house interiors. In: 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) (2017) + +18. Nash, C., Ganin, Y., Eslami, S.A., Battaglia, P.: Polygen: An autoregressive generative model of 3d meshes. In: International conference on machine learning. pp. 7220-7229. PMLR (2020) +19. Ochmann, S., Vock, R., Klein, R.: Automatic reconstruction of fully volumetric 3d building models from oriented point clouds. ISPRS journal of photogrammetry and remote sensing 151, 251-262 (2019) +20. OpenAI: Gpt-4 technical report (2023) +21. Para, W., Bhat, S., Guerrero, P., Kelly, T., Mitra, N., Guibas, L.J., Wonka, P.: Sketchgen: Generating constrained cad sketches. In: Advances in Neural Information Processing Systems (2021) +22. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2019) +23. Pintore, G., Agus, M., Gobbetti, E.: Atlantanet: inferring the 3d indoor layout from a single 360 image beyond the Manhattan world assumption. In: Proceedings of the European conference on computer vision (ECCV) (2020) +24. Rukhovich, D., Vorontsova, A., Konushin, A.: Imvoxelnet: Image to voxels projection for monocular and multi-view general-purpose 3d object detection. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2397-2406 (2022) +25. Somasundaram, K., Dong, J., Tang, H., Straub, J., Yan, M., Goesele, M., Engel, J.J., De Nardi, R., Newcombe, R.: Project aria: A new tool for egocentric multimodal ai research. arXiv preprint arXiv:2308.13561 (2023) +26. Tang, H., Liu, Z., Li, X., Lin, Y., Han, S.: TorchSparse: Efficient Point Cloud Inference Engine. In: Conference on Machine Learning and Systems (MLSys) (2022) +27. Tang, H., Liu, Z., Zhao, S., Lin, Y., Lin, J., Wang, H., Han, S.: Searching Efficient 3D Architectures with Sparse Point-Voxel Convolution. In: European Conference on Computer Vision (ECCV) (2020) +28. Tulsiani, S., Su, H., Guibas, L.J., Efros, A.A., Malik, J.: Learning shape abstractions by assembling volumetric primitives. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2017) +29. Tyszkiiewicz, M.J., Maninis, K.K., Popov, S., Ferrari, V.: Raytran: 3d pose estimation and shape reconstruction of multiple objects from videos with ray-traced transformers. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part X. pp. 211-228. Springer (2022) +30. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. NeurIPS 30 (2017) +31. Vu, T., Kim, K., Luu, T.M., Nguyen, X.T., Yoo, C.D.: Softgroup for 3d instance segmentation on 3d point clouds. In: CVPR (2022) +32. Wu, R., Xiao, C., Zheng, C.: Deepcad: A deep generative network for computer-aided design models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2021) +33. Xu, X., Willis, K.D., Lambourne, J.G., Cheng, C.Y., Jayaraman, P.K., Furukawa, Y.: Skexgen: Autoregressive generation of cad construction sequences with disentangled codebooks. In: International Conference on Machine Learning (ICML) (2022) +34. Xue, Y., Mao, J., Niu, M., Xu, H., Mi, M.B., Zhang, W., Wang, X., Wang, X.: Point2seq: Detecting 3d objects as sequences. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8521-8530 (2022) + +35. Yang, K., Chen, X.: Unsupervised learning for cuboid shape abstraction via joint segmentation from point clouds. ACM Transactions on Graphics (TOG) 40(4), 1-11 (2021) +36. Yin, T., Zhou, X., Krahenbuhl, P.: Center-based 3d object detection and tracking. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2021) +37. Yue, Y., Kontogianni, T., Schindler, K., Engelmann, F.: Connecting the dots: Floorplan reconstruction using two-level queries. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023) +38. Zou, C., Colburn, A., Shan, Q., Hoiem, D.: Layoutnet: Reconstructing the 3d room layout from a single rgb image. In: Proceedings of the IEEE conference on computer vision and pattern recognition (2018) \ No newline at end of file diff --git a/2024/SceneScript_ Reconstructing Scenes With An Autoregressive Structured Language Model/images.zip b/2024/SceneScript_ Reconstructing Scenes With An Autoregressive Structured Language Model/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a395d4b50b0ed44346ed114f8647d804b19be802 --- /dev/null +++ b/2024/SceneScript_ Reconstructing Scenes With An Autoregressive Structured Language Model/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ac4dac17ce65ee1019fb59169a7c83c7a133b9ec919dd4b0392cd479213136e1 +size 458967 diff --git a/2024/SceneScript_ Reconstructing Scenes With An Autoregressive Structured Language Model/layout.json b/2024/SceneScript_ Reconstructing Scenes With An Autoregressive Structured Language Model/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6bf3c40010c35c7391c66096c910629bd4284c76 --- /dev/null +++ b/2024/SceneScript_ Reconstructing Scenes With An Autoregressive Structured Language Model/layout.json @@ -0,0 +1,8506 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 149, + 61, + 464, + 96 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 61, + 464, + 96 + ], + "spans": [ + { + "bbox": [ + 149, + 61, + 464, + 96 + ], + "type": "text", + "content": "SceneScript: Reconstructing Scenes With An Autoregressive Structured Language Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 134, + 106, + 479, + 153 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 106, + 479, + 153 + ], + "spans": [ + { + "bbox": [ + 134, + 106, + 479, + 153 + ], + "type": "text", + "content": "Armen Avetisyan1, Christopher Xie1, Henry Howard-Jenkins1, Tsun-Yi Yang1, Samir Aroudj1, Suvam Patra1, Fuyang Zhang2†, Duncan Frost1, Luke Holland1, Campbell Orme1, Jakob Engel1, Edward Miller1, Richard Newcombe1, and Vasileios Balntas1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 201, + 163, + 411, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 163, + 411, + 186 + ], + "spans": [ + { + "bbox": [ + 201, + 163, + 411, + 186 + ], + "type": "text", + "content": "Meta Reality Labs Simon Fraser University https://projectaria.com/scenescript" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 198, + 225, + 251 + ], + "blocks": [ + { + "bbox": [ + 133, + 198, + 225, + 251 + ], + "lines": [ + { + "bbox": [ + 133, + 198, + 225, + 251 + ], + "spans": [ + { + "bbox": [ + 133, + 198, + 225, + 251 + ], + "type": "image", + "image_path": "b5c2a70ef1294fc3992a7dc3fdcaccd62cf263fd13e44c1cc0c942d24d45769f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 155, + 252, + 196, + 258 + ], + "lines": [ + { + "bbox": [ + 155, + 252, + 196, + 258 + ], + "spans": [ + { + "bbox": [ + 155, + 252, + 196, + 258 + ], + "type": "text", + "content": "Scene walkthrough" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 239, + 198, + 288, + 258 + ], + "blocks": [ + { + "bbox": [ + 239, + 198, + 288, + 258 + ], + "lines": [ + { + "bbox": [ + 239, + 198, + 288, + 258 + ], + "spans": [ + { + "bbox": [ + 239, + 198, + 288, + 258 + ], + "type": "image", + "image_path": "ef23cd8bcb4a5c2c1c1a1f60d8901c3cd6b9710c65b21064e5194dc91757e526.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 301, + 220, + 312, + 231 + ], + "lines": [ + { + "bbox": [ + 301, + 220, + 312, + 231 + ], + "spans": [ + { + "bbox": [ + 301, + 220, + 312, + 231 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 336, + 200, + 365, + 249 + ], + "blocks": [ + { + "bbox": [ + 336, + 200, + 365, + 249 + ], + "lines": [ + { + "bbox": [ + 336, + 200, + 365, + 249 + ], + "spans": [ + { + "bbox": [ + 336, + 200, + 365, + 249 + ], + "type": "image", + "image_path": "0cb693aa5315d4bc1a3e0bf93e5a0e9a98bc6b360a4e1941ee206c96fa8328e4.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 389, + 199, + 480, + 250 + ], + "blocks": [ + { + "bbox": [ + 389, + 199, + 480, + 250 + ], + "lines": [ + { + "bbox": [ + 389, + 199, + 480, + 250 + ], + "spans": [ + { + "bbox": [ + 389, + 199, + 480, + 250 + ], + "type": "image", + "image_path": "b5e369cc3b6ed30fdeab00dfbfa8c06ad6193df51b8b6dcc2b5710698a0c791b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 252, + 473, + 258 + ], + "lines": [ + { + "bbox": [ + 403, + 252, + 473, + 258 + ], + "spans": [ + { + "bbox": [ + 403, + 252, + 473, + 258 + ], + "type": "text", + "content": "Language to geometry interpreter" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 132, + 265, + 201, + 308 + ], + "blocks": [ + { + "bbox": [ + 132, + 265, + 201, + 308 + ], + "lines": [ + { + "bbox": [ + 132, + 265, + 201, + 308 + ], + "spans": [ + { + "bbox": [ + 132, + 265, + 201, + 308 + ], + "type": "image", + "image_path": "d53ad6538b00b5136035b7acfddfca2ec26dcb7d507edbd4acf542350e904719.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 133, + 310, + 201, + 354 + ], + "blocks": [ + { + "bbox": [ + 133, + 310, + 201, + 354 + ], + "lines": [ + { + "bbox": [ + 133, + 310, + 201, + 354 + ], + "spans": [ + { + "bbox": [ + 133, + 310, + 201, + 354 + ], + "type": "image", + "image_path": "6cd6225ceb1d8f51b78851bed857cb49c69da71aa893bcc20e79194679c9c09e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 363, + 482, + 441 + ], + "lines": [ + { + "bbox": [ + 130, + 363, + 482, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 363, + 482, + 441 + ], + "type": "text", + "content": "Fig. 1: (top) Given an egocentric video of an environment, SceneScript directly predicts a 3D scene representation consisting of structured scene language commands. (bottom) Our method generalizes on diverse real scenes while being solely trained on synthetic indoor environments. (last column, bottom) A notable advantage of our method is its capacity to easily adapt the structured language to represent novel scene entities. For example, by introducing a single new command, SceneScript can directly predict object parts jointly with the layout and bounding boxes." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 203, + 265, + 270, + 308 + ], + "blocks": [ + { + "bbox": [ + 203, + 265, + 270, + 308 + ], + "lines": [ + { + "bbox": [ + 203, + 265, + 270, + 308 + ], + "spans": [ + { + "bbox": [ + 203, + 265, + 270, + 308 + ], + "type": "image", + "image_path": "e6643aea7d8f149ea5b8d08e7e985345a028102d0b636820bb272dd4b1da4057.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 203, + 310, + 271, + 354 + ], + "blocks": [ + { + "bbox": [ + 203, + 310, + 271, + 354 + ], + "lines": [ + { + "bbox": [ + 203, + 310, + 271, + 354 + ], + "spans": [ + { + "bbox": [ + 203, + 310, + 271, + 354 + ], + "type": "image", + "image_path": "dd64e2665a6f9a6001da7faa4dcccd2412e132e6e9dc1d03409a84949dd48f23.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 273, + 265, + 340, + 308 + ], + "blocks": [ + { + "bbox": [ + 322, + 252, + 379, + 258 + ], + "lines": [ + { + "bbox": [ + 322, + 252, + 379, + 258 + ], + "spans": [ + { + "bbox": [ + 322, + 252, + 379, + 258 + ], + "type": "text", + "content": "Structured scene language" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 273, + 265, + 340, + 308 + ], + "lines": [ + { + "bbox": [ + 273, + 265, + 340, + 308 + ], + "spans": [ + { + "bbox": [ + 273, + 265, + 340, + 308 + ], + "type": "image", + "image_path": "8c5baab5c2e2956dd9cfd5ea0a849af2e9c09c69c31da3c631169fae8ac4f47a.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 272, + 310, + 340, + 354 + ], + "blocks": [ + { + "bbox": [ + 272, + 310, + 340, + 354 + ], + "lines": [ + { + "bbox": [ + 272, + 310, + 340, + 354 + ], + "spans": [ + { + "bbox": [ + 272, + 310, + 340, + 354 + ], + "type": "image", + "image_path": "7d687e8dcf3da4e25121307cdc4481724f82426f8c04e4b66f0a19e81503f9f7.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 342, + 265, + 410, + 308 + ], + "blocks": [ + { + "bbox": [ + 342, + 265, + 410, + 308 + ], + "lines": [ + { + "bbox": [ + 342, + 265, + 410, + 308 + ], + "spans": [ + { + "bbox": [ + 342, + 265, + 410, + 308 + ], + "type": "image", + "image_path": "fc5afea0199592b657ee3cd9cddaf74bb01eb6309d2db95699ec84d92528365c.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 342, + 310, + 410, + 354 + ], + "blocks": [ + { + "bbox": [ + 342, + 310, + 410, + 354 + ], + "lines": [ + { + "bbox": [ + 342, + 310, + 410, + 354 + ], + "spans": [ + { + "bbox": [ + 342, + 310, + 410, + 354 + ], + "type": "image", + "image_path": "6f333f99c8811e8a9f9637f8934144af7678470a1981225f2cf9e8ff6d593fdc.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 412, + 265, + 480, + 308 + ], + "blocks": [ + { + "bbox": [ + 412, + 265, + 480, + 308 + ], + "lines": [ + { + "bbox": [ + 412, + 265, + 480, + 308 + ], + "spans": [ + { + "bbox": [ + 412, + 265, + 480, + 308 + ], + "type": "image", + "image_path": "4afa5e094cf97862c1ee0bcc43375ffece58e4f15bd994568b8095efb6bcbe1a.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 413, + 310, + 480, + 354 + ], + "blocks": [ + { + "bbox": [ + 413, + 310, + 480, + 354 + ], + "lines": [ + { + "bbox": [ + 413, + 310, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 413, + 310, + 480, + 354 + ], + "type": "image", + "image_path": "4fb9354133c22f600953d0c8f0537cf2a881486c6e2fb8b3a38bdb40cd53829e.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 159, + 461, + 453, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 461, + 453, + 648 + ], + "spans": [ + { + "bbox": [ + 159, + 461, + 453, + 648 + ], + "type": "text", + "content": "Abstract. We introduce SceneScript, a method that directly produces full scene models as a sequence of structured language commands using an autoregressive, token-based approach. Our proposed scene representation is inspired by recent successes in transformers & LLMs, and departs from more traditional methods which commonly describe scenes as meshes, voxel grids, point clouds or radiance fields. Our method infers the set of structured language commands directly from encoded visual data using a scene language encoder-decoder architecture. To train SceneScript, we generate and release a large-scale synthetic dataset called Aria Synthetic Environments consisting of " + }, + { + "bbox": [ + 159, + 461, + 453, + 648 + ], + "type": "inline_equation", + "content": "100k" + }, + { + "bbox": [ + 159, + 461, + 453, + 648 + ], + "type": "text", + "content": " high-quality indoor scenes, with photorealistic and ground-truth annotated renders of egocentric scene walkthroughs. Our method gives state-of-the-art results in architectural layout estimation, and competitive results in 3D object detection. Lastly, we explore an advantage for SceneScript, which is the ability to readily adapt to new commands via simple additions to the structured language, which we illustrate for tasks such as coarse 3D object part reconstruction." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 653, + 357, + 664 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 653, + 357, + 664 + ], + "spans": [ + { + "bbox": [ + 141, + 653, + 357, + 664 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 141, + 653, + 357, + 664 + ], + "type": "text", + "content": "Work done while the author was an intern at Meta." + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 230, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 230, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 230, + 128 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 482, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 482, + 330 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 482, + 330 + ], + "type": "text", + "content": "Scene representations play a crucial role in machine learning and computer vision applications, enabling accurate understanding of the environment. Over the years, researchers have explored various options such as meshes, voxel grids, point clouds, and implicit representations, aiming to represent complex real-world scenes with high-fidelity. Each of these exhibits distinct advantages and limitations that impact their suitability for different tasks. Meshes offer detailed geometric information but can be expensive in both computation and memory. Voxel grids provide a volumetric representation but suffer from a trade-off between resolution and memory requirements. Point clouds are efficient in representing sparse scenes but lack semantics and explicit connectivity information. Implicit representations, such as DeepSDF [22] and NeRF [15], can be infinitely precise but lack interpretability and editability. The selection of an appropriate representation directly impacts the performance and efficacy of various tasks including object recognition, scene understanding, and 3D reconstruction. In this paper, we propose a novel scene representation based on structured language commands as a more efficient and versatile solution." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 331, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 482, + 449 + ], + "type": "text", + "content": "Our motivation stems from the recent advancements in the field of Large Language Models (LLMs) and \"next token prediction\" autoregressive methods [20], coupled with recent works on exploring generation of sequences to represent geometric structures. For example, PolyGen [18] demonstrated the ability to describe 3D meshes as a sequence of vertices and faces generated using transformers [30]. Similarly, CAD-as-Language [11] showcased the effectiveness of generating Computer-Aided Design (CAD) primitives to represent 2D CAD sketches. Our main goal is to directly infer a metrically accurate representation of a full scene as a text-based sequence of specialized structured language commands." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 450, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 594 + ], + "type": "text", + "content": "Our method, denoted SceneScript, autoregressively predicts a language of hand-designed structured language commands in pure text form. This language offers several distinct advantages: 1) As pure text, it is compact and reduces memory requirements of a large scene to only a few bytes. 2) It is crisp and complete since the commands are designed to result in sharp and well-defined geometry (similar to scalable vector graphics). 3) It is interpretable, editable and semantically rich by design via the use of high-level parametric commands such as make_door(*door_parameters). 4) It can seamlessly integrate novel geometric entities by simply adding new structured commands to the language. 5) The fact that the scene representation is a series of language tokens similar to [20] opens up a plethora of potential new applications in the future such as ways to edit the scene, query it, or spin up chat interactions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "We mainly focus on the problems of architectural layout estimation and object detection as proxy tasks for the efficacy of our SceneScript language as a scene representation. Architectural entities such as walls, doors, and windows are highly structured entities, making them an ideal test-bed. However, one notable drawback of language models is that they require vast amounts of data for training. Since there is no existing dataset of scene walkthroughs and their cor" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "text", + "content": "Avetisyan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "responding structured language commands, we publicly released Aria Synthetic Environments (ASE), a synthetically generated dataset of 100k unique interior scenes. For each scene, we simulate egocentric trajectories with an entire suite of sensor data from Project Aria [14]. We also release multiple sources of ground truth including depth and instance segmentations. Importantly, for each rendered egocentric sequence, the architectural layout ground truth is given in our proposed SceneScript language." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 482, + 308 + ], + "type": "text", + "content": "While architectural layout serves as a test-bed, we demonstrate that our method SceneScript can easily be extended to new tasks via simple extensions to SceneScript language while keeping both the visual input and network architecture fixed. We illustrate this on the problem of 3D object detection, which results in a method that jointly infers architectural layout and 3D oriented bounding boxes. Additionally, we demonstrate more proof-of-concept experiments that show that our method results in a significantly lower barrier to entry for new tasks including representing coarse 3D object reconstruction, curved entities, composition of entities, and entity states." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 147, + 308, + 268, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 308, + 268, + 319 + ], + "spans": [ + { + "bbox": [ + 147, + 308, + 268, + 319 + ], + "type": "text", + "content": "Our core contributions are:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 327, + 480, + 459 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 138, + 327, + 479, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 327, + 479, + 362 + ], + "spans": [ + { + "bbox": [ + 138, + 327, + 479, + 362 + ], + "type": "text", + "content": "- We propose SceneScript, a method that jointly predicts architectural layout and object bounding boxes of a scene in the form of structured language commands given a video stream." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 363, + 480, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 363, + 480, + 398 + ], + "spans": [ + { + "bbox": [ + 138, + 363, + 480, + 398 + ], + "type": "text", + "content": "- We demonstrate that SceneScript can easily be extended to completely new tasks with simple additions of commands to our structured language, significantly lowering the barrier to entry for new tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 399, + 480, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 399, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 138, + 399, + 480, + 434 + ], + "type": "text", + "content": "- We release a large-scale synthetic dataset, named Aria Synthetic Environments, comprised of 100k unique high-quality 3D indoor scenes with GT, which will enable large scale ML training of scene understanding methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 434, + 480, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 434, + 480, + 459 + ], + "spans": [ + { + "bbox": [ + 138, + 434, + 480, + 459 + ], + "type": "text", + "content": "- We show that training SceneScript on Aria Synthetic Environments leads to generalization on real scenes (see videos/demos on the project page)." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 476, + 242, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 476, + 242, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 476, + 242, + 489 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 502, + 254, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 254, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 254, + 514 + ], + "type": "text", + "content": "2.1 Layout Estimation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 521, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 593 + ], + "type": "text", + "content": "Layout estimation is an active research area, aiming to infer architectural elements. Scan2BIM [17] proposes heuristics for wall detection to produce 2D floorplans. Ochmann et al. [19] formulate layout inference as an integer linear program using constraints on detected walls. Shortest path algorithms around birds-eye view (BEV) free space [3] and wall plus room instance segmentation [5] have also been explored." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Furukawa et al. [10] utilize a Manhattan world-based multi-view stereo algorithm [9] to merge axis-aligned depth maps into a full 3D mesh of building interiors. RoomNet [13] predicts layout keypoints while assuming that a fixed set of Manhattan layouts can occur in a single image of a room. LayoutNet [38] improves on this by predicting keypoints and optimising the Manhattan room layout inferred from them. Similarly, AtlantaNet [23] predicts a BEV floor or" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 399, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 399, + 91, + 447, + 102 + ], + "type": "text", + "content": "SceneScript" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 176, + 179, + 432, + 264 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 1: Complete set of structured language commands designed for detailing architectural layouts and object bounding boxes. Supported data types can include int, float, bool. It is important to note that the language's extensibility allows for easy augmentation by introducing new commands like make_prime, make_pillar, or enhancing existing commands, such as incorporating is-double_door (bool)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 176, + 179, + 432, + 264 + ], + "lines": [ + { + "bbox": [ + 176, + 179, + 432, + 264 + ], + "spans": [ + { + "bbox": [ + 176, + 179, + 432, + 264 + ], + "type": "table", + "html": "
make_WALL (int)make_door (int)make_window (int)make_bbox (int)
idintidintidintidint
a_xfloatwall0_idintwall0_idintclassint
a_yfloatwall1_idintwall1_idintposition_xfloat
a_zfloatposition_xfloatposition_xfloatposition_yfloat
b_xfloatposition_yfloatposition_yfloatposition_zfloat
b_yfloatposition_zfloatposition_zfloatangle_zfloat
b_zfloatwidthfloatwidthfloatscale_xfloat
heightfloatheightfloatheightfloatscale_yfloat
scale_zfloat
", + "image_path": "d2a05526a400137c7748c0408ef09184df0a27647e68c51e1f51e918380c90f2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 289, + 480, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 480, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 480, + 326 + ], + "type": "text", + "content": "ceiling shape and approximates the shape contour with a polygon resulting in an Atlanta world prior. SceneCAD [1] uses a graph neural network to predict object-object and object-query relationships to refine its layout prediction." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 326, + 481, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 326, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 481, + 374 + ], + "type": "text", + "content": "Our approach stands out by requiring neither heuristics nor explicitly defined prior knowledge about architectural scene structure. In fact, our method demonstrates geometric understanding of the scene which emerges despite learning to predict the GT scene language as a sequence of text tokens." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 398, + 318, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 398, + 318, + 411 + ], + "spans": [ + { + "bbox": [ + 131, + 398, + 318, + 411 + ], + "type": "text", + "content": "2.2 Geometric Sequence Modelling" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 425, + 482, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 425, + 482, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 425, + 482, + 579 + ], + "type": "text", + "content": "Recent works have explored transformers for generating objects as text-based sequences. PolyGen [18] models 3D meshes as a sequence of vertices and faces. CAD-as-Language [11] represents 2D CAD sketches as a sequence of triplets in protobuf representation, followed by a sequence of constraints. Both SketchGen [21] and SkexGen [33] use transformers to generate sketches. DeepSVG [4] learns a transformer-based variational autoencoder (VAE) that is capable of generating and interpolating through 2D vector graphics images. DeepCAD [32] proposes a low-level language and architecture similarly to DeepSVG, but applies it to 3D CAD models instead of 2D vector graphics. Our approach stands out by utilising high-level commands, offering interpretability and semantic richness. Additionally, while low-level commands can represent arbitrarily complex geometries, they lead to prohibitively longer sequences when representing a full scene." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "text", + "content": "The closest work to ours is Pix2Seq [6]. Pix2Seq proposes a similar architecture to ours but experiments only with 2D object detection, thus requiring domain-specific augmentation strategies. Another closely related work is Point2Seq [34] that trains a recurrent network for autoregressively regressing continuous 3D bounding box parameters. Interestingly, they find the autoregressive ordering of parameters outperforms current standards for object detection architectures, including anchors [12] and centers [36]." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "text", + "content": "Avetisyan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 274 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 274 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 274 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 274 + ], + "type": "image", + "image_path": "1a1d1b552808797c3bc3fddf81646e7d68331207ca166790f8def6c7738630dc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 281, + 482, + 326 + ], + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 326 + ], + "type": "text", + "content": "Fig. 2: Aria Synthetic Environments: (top) Random samples of generated scenes showing diversity of layouts, lights and object placements. (bottom - left to right) A top down view of a scene filled with objects, a simulated trajectory (blue path), renderings of depth, RGB, and object instances, and lastly a scene pointcloud." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 348, + 421, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 348, + 421, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 348, + 421, + 363 + ], + "type": "text", + "content": "3 SceneScript Structured Language Commands" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 374, + 482, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 374, + 482, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 482, + 411 + ], + "type": "text", + "content": "We first describe our structured language commands that define a full scene representation including both layout and objects. After this, we introduce our corresponding large scale training dataset: Aria Synthetic Environments." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 429, + 301, + 440 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 429, + 301, + 440 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 301, + 440 + ], + "type": "text", + "content": "3.1 Commands and Parameters" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 449, + 482, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 556 + ], + "type": "text", + "content": "We begin with a parameterization that captures the most common layout elements. For this purpose we use three commands: make-wall, make_door, and make_window. Each command comes with a set of parameters that results in well-defined geometry. For example, the full set of parameters for make-wall specifies a gravity-aligned 2D plane, while make_door and make_window specify box-based cutouts from walls. It is worth noting that this parametrization is arbitrary, and is only made in the context of presenting a proof-of-concept SceneScript system. There are infinitely many parametrization schemes, in this work we opt for one that prioritizes ease of use and research iteration speed." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 557, + 481, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 481, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 481, + 581 + ], + "type": "text", + "content": "In addition to representing these three major layout entities, we aim to jointly infer objects as oriented bounding boxes. Thus, we introduce a fourth command:" + } + ] + } + ], + "index": 8 + }, + { + "type": "code", + "bbox": [ + 132, + 588, + 443, + 612 + ], + "blocks": [ + { + "bbox": [ + 132, + 588, + 443, + 612 + ], + "lines": [ + { + "bbox": [ + 132, + 588, + 443, + 612 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 443, + 612 + ], + "type": "text", + "content": "make_bbox: id, class, position_x, position_y, position_z, angle_z, scale_x, scale_y, scale_z" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "code_body" + } + ], + "index": 9, + "sub_type": "code", + "guess_lang": "python" + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "content": "This simple parametrization represents an oriented 3D bounding box that is assumed to be aligned with gravity (assuming it points in the " + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "inline_equation", + "content": "-z" + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "content": " direction). A summary of these commands and their respective parameters are shown in Table 1." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 399, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 399, + 91, + 447, + 102 + ], + "type": "text", + "content": "SceneScript" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "While we have described just four commands to capture structure and objects in an indoor environment, importantly, this text-based parametrization can readily be extended geometrically and/or even semantically to include states or other functional aspects. For example, changing the mentioned makedoor command to include parameters such as open-degree allows the language to represent door states. In Section 6, we demonstrate how such extensions to the language allows for SceneScript to readily adapt to new tasks including coarse 3D object reconstruction." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 231, + 244, + 242 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 231, + 244, + 242 + ], + "spans": [ + { + "bbox": [ + 131, + 231, + 244, + 242 + ], + "type": "text", + "content": "3.2 Scene Definition" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 251, + 481, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 251, + 481, + 301 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 481, + 301 + ], + "type": "text", + "content": "A single scene can be described as a sequence of our proposed structured language commands. The sequence requires no specific ordering, and can be of arbitrary length. From this definition, the 3D scene can easily be obtained by parsing the commands through a simple custom interpreter." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 319, + 246, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 319, + 246, + 331 + ], + "spans": [ + { + "bbox": [ + 131, + 319, + 246, + 331 + ], + "type": "text", + "content": "3.3 Training Dataset" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 339, + 481, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 339, + 481, + 388 + ], + "spans": [ + { + "bbox": [ + 130, + 339, + 481, + 388 + ], + "type": "text", + "content": "To enable practical indoor scene reconstruction based on our structured language commands (Section 3.1), we publicly released a new dataset called Aria Synthetic Environments. It consists of a large number of training pairs of egocentric scene walkthroughs linked with corresponding ground truth command sequences." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 388, + 482, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 482, + 460 + ], + "type": "text", + "content": "Since transformers require vast amounts of data, we generated " + }, + { + "bbox": [ + 130, + 388, + 482, + 460 + ], + "type": "inline_equation", + "content": "100k" + }, + { + "bbox": [ + 130, + 388, + 482, + 460 + ], + "type": "text", + "content": " synthetic scenes, which is in comparison infeasible for real-world data. Each synthetic scene comes with a floor plan model, a corresponding complete 3D scene as well as a simulated agent trajectory and a photo-realistic rendering of this trajectory. Fig. 2 illustrates the basics of Aria Synthetic Environments. For brevity, we refer the reader to the supplemental material for further details." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 479, + 357, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 479, + 357, + 493 + ], + "spans": [ + { + "bbox": [ + 131, + 479, + 357, + 493 + ], + "type": "text", + "content": "4 SceneScript Network Architecture" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 506, + 481, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 506, + 481, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 506, + 481, + 540 + ], + "type": "text", + "content": "Our pipeline is a simple encoder-decoder architecture that consumes a video sequence and returns SceneScript language in a tokenized format. Figure 3 illustrates a high-level overview of our method." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 542, + 481, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 542, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 130, + 542, + 481, + 566 + ], + "type": "text", + "content": "We examine three encoder variants: a pointcloud encoder, a posed image set encoder, and a combined encoder. The decoder remains the same in all cases." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 584, + 318, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 584, + 318, + 597 + ], + "spans": [ + { + "bbox": [ + 131, + 584, + 318, + 597 + ], + "type": "text", + "content": "4.1 Input Modalities and Encoders" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 605, + 481, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 667 + ], + "type": "text", + "content": "The encoder computes a latent scene code in the form of a 1D sequence from video walkthrough of the scene. The decoder is designed to consume these 1D sequences as input. This enables the integration of various input modalities within a unified framework. As a preliminary, for each scene we assume access to a set of " + }, + { + "bbox": [ + 130, + 605, + 481, + 667 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 605, + 481, + 667 + ], + "type": "text", + "content": " posed camera images " + }, + { + "bbox": [ + 130, + 605, + 481, + 667 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{I}_1,\\dots,\\mathbf{I}_{\\mathbf{M}}\\}" + }, + { + "bbox": [ + 130, + 605, + 481, + 667 + ], + "type": "text", + "content": ", e.g., SLAM output." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "text", + "content": "Avetisyan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 324, + 229 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 324, + 229 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 324, + 229 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 324, + 229 + ], + "type": "image", + "image_path": "c3b7e7ec3fe381af7c3e1f2cad895f41c4683d131632eaee1af83a60648d7867.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 232, + 482, + 288 + ], + "lines": [ + { + "bbox": [ + 130, + 232, + 482, + 288 + ], + "spans": [ + { + "bbox": [ + 130, + 232, + 482, + 288 + ], + "type": "text", + "content": "Fig. 3: SceneScript core pipeline overview. Raw images & pointcloud data are encoded into a latent code, which is then autoregressively decoded into a sequence of commands that describe the scene. Visualizations are shown using a customly built interpreter. Note that for the results in this paper, the the point clouds are computed from the images using Aria MPS [25] - i.e. are not using a dedicated RGB-D / Lidar sensor." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 326, + 113, + 481, + 230 + ], + "blocks": [ + { + "bbox": [ + 326, + 113, + 481, + 230 + ], + "lines": [ + { + "bbox": [ + 326, + 113, + 481, + 230 + ], + "spans": [ + { + "bbox": [ + 326, + 113, + 481, + 230 + ], + "type": "image", + "image_path": "2dea9288ccf974cbfed2a66b1452436a7926daa4a216cbdccb1de8553d1dd53c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 311, + 479, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 311, + 479, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 311, + 479, + 346 + ], + "type": "text", + "content": "Point Clouds. A point cloud " + }, + { + "bbox": [ + 130, + 311, + 479, + 346 + ], + "type": "inline_equation", + "content": "\\mathbf{P} = \\{\\mathbf{p}_1, \\dots, \\mathbf{p_N}\\}" + }, + { + "bbox": [ + 130, + 311, + 479, + 346 + ], + "type": "text", + "content": " consists of " + }, + { + "bbox": [ + 130, + 311, + 479, + 346 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 311, + 479, + 346 + ], + "type": "text", + "content": " points, where " + }, + { + "bbox": [ + 130, + 311, + 479, + 346 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 130, + 311, + 479, + 346 + ], + "type": "text", + "content": " is a 3D point coordinate. It can come from passive images using SLAM or SfM, or RGB-D / Lidar sensors." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 346, + 480, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 480, + 418 + ], + "type": "text", + "content": "Specifically, we use the Semi-dense Pointclouds from Project Aria's Machine Perception Services [25], that are obtained from a visual-inertial SLAM system using Aria's monochrome cameras and IMUs. We discretize the point cloud to " + }, + { + "bbox": [ + 130, + 346, + 480, + 418 + ], + "type": "inline_equation", + "content": "5\\mathrm{cm}" + }, + { + "bbox": [ + 130, + 346, + 480, + 418 + ], + "type": "text", + "content": " resolution, then employ a sparse 3D convolution library [26,27] to generate pooled features. The encoder " + }, + { + "bbox": [ + 130, + 346, + 480, + 418 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{geo}" + }, + { + "bbox": [ + 130, + 346, + 480, + 418 + ], + "type": "text", + "content": " applies a series of down convolutions, resulting in a reduction of the number of points in the lowest level." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 206, + 429, + 480, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 429, + 480, + 444 + ], + "spans": [ + { + "bbox": [ + 206, + 429, + 480, + 444 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {g e o} = \\mathcal {E} _ {g e o} (\\mathbf {P}), \\quad \\mathbf {P} \\in \\mathbb {R} ^ {N \\times 3}, \\mathbf {F} _ {g e o} \\in \\mathbb {R} ^ {K \\times 5 1 2} \\tag {1}", + "image_path": "ea3682509f9ec2ac918f8efe068edbd76ad90181fa80068e6dffa99960fac30f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 448, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 448, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 448, + 482, + 521 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 448, + 482, + 521 + ], + "type": "inline_equation", + "content": "K \\ll N" + }, + { + "bbox": [ + 130, + 448, + 482, + 521 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 448, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{geo}" + }, + { + "bbox": [ + 130, + 448, + 482, + 521 + ], + "type": "text", + "content": " is a condensed latent representation of the point cloud that contains the necessary scene context. For later use in the transformer decoder, we treat " + }, + { + "bbox": [ + 130, + 448, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{geo}" + }, + { + "bbox": [ + 130, + 448, + 482, + 521 + ], + "type": "text", + "content": " as a sequence of feature vectors where the entries " + }, + { + "bbox": [ + 130, + 448, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{f_i}" + }, + { + "bbox": [ + 130, + 448, + 482, + 521 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 448, + 482, + 521 + ], + "type": "inline_equation", + "content": "i \\in 1\\dots K" + }, + { + "bbox": [ + 130, + 448, + 482, + 521 + ], + "type": "text", + "content": " are sorted lexicographically according to the coordinate of the active site " + }, + { + "bbox": [ + 130, + 448, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{c_i}" + }, + { + "bbox": [ + 130, + 448, + 482, + 521 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 448, + 482, + 521 + ], + "type": "inline_equation", + "content": "i \\in 1\\dots K" + }, + { + "bbox": [ + 130, + 448, + 482, + 521 + ], + "type": "text", + "content": ". To incorporate positional encoding, we append the coordinates of the active sites to their respective feature vectors " + }, + { + "bbox": [ + 130, + 448, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i \\gets \\mathbf{cat}(\\mathbf{f_i},\\mathbf{c_i})" + }, + { + "bbox": [ + 130, + 448, + 482, + 521 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 537, + 482, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 537, + 482, + 610 + ], + "spans": [ + { + "bbox": [ + 130, + 537, + 482, + 610 + ], + "type": "text", + "content": "Point Clouds with Lifted Features. We additionally explore augmenting the point cloud with image features. From the original egocentric sequence and associated trajectory, we sample a set of " + }, + { + "bbox": [ + 130, + 537, + 482, + 610 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 537, + 482, + 610 + ], + "type": "text", + "content": " keyframes, " + }, + { + "bbox": [ + 130, + 537, + 482, + 610 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_i" + }, + { + "bbox": [ + 130, + 537, + 482, + 610 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 537, + 482, + 610 + ], + "type": "inline_equation", + "content": "i\\in 1\\dots M" + }, + { + "bbox": [ + 130, + 537, + 482, + 610 + ], + "type": "text", + "content": ", and compute a set of image features " + }, + { + "bbox": [ + 130, + 537, + 482, + 610 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i" + }, + { + "bbox": [ + 130, + 537, + 482, + 610 + ], + "type": "text", + "content": " for each. We then project each point into the set of keyframe cameras and retrieve the feature vector (output by a CNN) at the pixel location:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 226, + 620, + 480, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 226, + 620, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 226, + 620, + 480, + 633 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {i p} = F _ {i} (\\pi (\\mathbf {p})) \\quad \\mathbf {p} \\in \\mathbf {P}, i \\in 1 \\dots M, \\tag {2}", + "image_path": "a94bb2c8a60e9b403c4e8a0aeba61c16b1739e283b7915ae107dba76b58394c7.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\pi(\\cdot)" + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": " represents the projection function of a 3D point into the camera. If " + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\pi(\\mathbf{p})" + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": " falls outside the image bounds, no feature is retrieved. We combine the set" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 399, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 399, + 91, + 447, + 102 + ], + "type": "text", + "content": "SceneScript" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "of lifted features for a point through an average, resulting in a single feature vector for each point: " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_p = 1 / M\\sum_{i = 1}^{M}\\mathbf{f}_{ip}" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 480, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 480, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 480, + 189 + ], + "type": "text", + "content": "We form our lifted-feature point cloud, " + }, + { + "bbox": [ + 130, + 140, + 480, + 189 + ], + "type": "inline_equation", + "content": "\\mathbf{P}' = \\{\\mathbf{p}_1',\\dots,\\mathbf{p}_{\\mathbf{M}}'\\}" + }, + { + "bbox": [ + 130, + 140, + 480, + 189 + ], + "type": "text", + "content": ", by concatenating each point's lifted feature with the original XYZ location: " + }, + { + "bbox": [ + 130, + 140, + 480, + 189 + ], + "type": "inline_equation", + "content": "\\mathbf{p}' = \\mathrm{cat}(\\mathbf{f}_p,\\mathbf{p})" + }, + { + "bbox": [ + 130, + 140, + 480, + 189 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 140, + 480, + 189 + ], + "type": "inline_equation", + "content": "\\mathbf{P}'" + }, + { + "bbox": [ + 130, + 140, + 480, + 189 + ], + "type": "text", + "content": " is then encoded into a context sequence using sparse 3D convolutions, with only the input feature channels adjusted to match the new point feature dimension." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 208, + 479, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 208, + 479, + 243 + ], + "spans": [ + { + "bbox": [ + 130, + 208, + 479, + 243 + ], + "type": "text", + "content": "End-to-end Encoding of Posed Views. In order to encode the egocentric sequence more directly without a pre-computed point cloud, we adopt a " + }, + { + "bbox": [ + 130, + 208, + 479, + 243 + ], + "type": "inline_equation", + "content": "2\\mathrm{D} \\leftrightarrow 3\\mathrm{D}" + }, + { + "bbox": [ + 130, + 208, + 479, + 243 + ], + "type": "text", + "content": " bidirectional transformer encoder following the form defined in RayTran [29]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 245, + 480, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 245, + 480, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 245, + 480, + 411 + ], + "type": "text", + "content": "In this formulation, we initialize a volume representation of the scene as a dense voxel grid of features, " + }, + { + "bbox": [ + 130, + 245, + 480, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 130, + 245, + 480, + 411 + ], + "type": "text", + "content": ", that coincides with the scene geometry. In turn, we sample a subset of " + }, + { + "bbox": [ + 130, + 245, + 480, + 411 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 245, + 480, + 411 + ], + "type": "text", + "content": " keyframes, " + }, + { + "bbox": [ + 130, + 245, + 480, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_i" + }, + { + "bbox": [ + 130, + 245, + 480, + 411 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 245, + 480, + 411 + ], + "type": "inline_equation", + "content": "i\\in 1\\dots M" + }, + { + "bbox": [ + 130, + 245, + 480, + 411 + ], + "type": "text", + "content": ", from the full stream of posed images. And for each of these keyframes we compute image features from a CNN, " + }, + { + "bbox": [ + 130, + 245, + 480, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i" + }, + { + "bbox": [ + 130, + 245, + 480, + 411 + ], + "type": "text", + "content": ". Repeated layers of bidirectional attention enable the image and voxel grid features to be refined iteratively in successive transformer blocks through the aggregation of view-point and global-scene information. As in RayTran [29], the interaction between the two representations is guided by the image-formation process by leveraging known camera parameters and poses. Attention in these transformer blocks is restricted by patch-voxel ray intersections, where each image patch attends to the voxels it observes and each voxel location attends to all the patches that observe it. The resulting voxel grid of features is flattened, concatenated with an encoded representation of its XYZ location, and passed to the decoder." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 431, + 254, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 431, + 254, + 443 + ], + "spans": [ + { + "bbox": [ + 131, + 431, + 254, + 443 + ], + "type": "text", + "content": "4.2 Language Decoder" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 453, + 479, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 453, + 479, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 479, + 536 + ], + "type": "text", + "content": "We utilize a transformer decoder [30] to decode the scene latent code into a sequence of structured language commands. The sequence of tokens passes through an embedding layer, followed by a positional encoding layer. Together with the encoded scene code (Section 4.1), the embedded tokens are passed into the several transformer decoder layers where a causal attention mask is used to ensure autoregressive generation. More implementation details can be found in Appendix ??" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 557, + 276, + 569 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 557, + 276, + 569 + ], + "spans": [ + { + "bbox": [ + 131, + 557, + 276, + 569 + ], + "type": "text", + "content": "4.3 Language Tokenization" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 579, + 479, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 579, + 479, + 639 + ], + "spans": [ + { + "bbox": [ + 130, + 579, + 479, + 639 + ], + "type": "text", + "content": "We refer to the serialization the structured language into a sequence of tokens as tokenization. The goal is to construct a bijective mapping between a sequence of structured language commands (Section 3) and a sequence of integer tokens that can be predicted by the transformer decoder architecture. We utilize the following schema:" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "text", + "content": "Avetisyan et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 161, + 649, + 451, + 662 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 649, + 451, + 662 + ], + "spans": [ + { + "bbox": [ + 161, + 649, + 451, + 662 + ], + "type": "text", + "content": "[START, PART, CMD, PARAM_1, PARAM_2, ..., PARAM_N, PART, ..., STOP]" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 147, + 116, + 479, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 116, + 479, + 128 + ], + "spans": [ + { + "bbox": [ + 147, + 116, + 479, + 128 + ], + "type": "text", + "content": "For example, a sample sequence for a makedoor command may look like:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 165, + 150, + 447, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 150, + 447, + 177 + ], + "spans": [ + { + "bbox": [ + 165, + 150, + 447, + 177 + ], + "type": "text", + "content": "[START, PART, MAKE_DOOR, POSITION_X, POSITION_Y, POSITION_Z, WALLO_IDX, WALL1_IDX, WIDTH, HEIGHT, PART,..., STOP]" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 188, + 479, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 479, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 479, + 247 + ], + "type": "text", + "content": "This schema enables 1D packing of tokens without requiring fixed-size slots or padding like other sequence modelling methods such as [32]. Additionally, it does not impose any limitations on the number or the hierarchy of sub-sequences, as they are flexibly separated by a PART token. This allows for arbitrarily complex scene representations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 248, + 479, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 479, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 479, + 308 + ], + "type": "text", + "content": "The tokenized sequence is discretized into integers at a 5cm resolution, then translated into a sequence of embeddings via learnable lookup table. Note that by designing the SceneScript language, we also design the tokenization. This tokenization scheme is notably different from standard NLP tokenization, which involves Byte-Pair Encodings (BPE) [20]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 327, + 198, + 339 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 327, + 198, + 339 + ], + "spans": [ + { + "bbox": [ + 132, + 327, + 198, + 339 + ], + "type": "text", + "content": "5 Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 354, + 479, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 479, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 479, + 390 + ], + "type": "text", + "content": "In this section, we introduce the metrics we define to measure performance, and we discuss some qualitative and quantitative results that give insights to our proposed SceneScript method." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 409, + 199, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 409, + 199, + 420 + ], + "spans": [ + { + "bbox": [ + 132, + 409, + 199, + 420 + ], + "type": "text", + "content": "5.1 Metrics" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 430, + 487, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 487, + 527 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 487, + 527 + ], + "type": "text", + "content": "To evaluate accuracy of the layout estimation, we make use of geometric metrics applied between the ground-truth room layout and our predicted SceneScript language. To do so, we define an entity distance, " + }, + { + "bbox": [ + 130, + 430, + 487, + 527 + ], + "type": "inline_equation", + "content": "d_{E}" + }, + { + "bbox": [ + 130, + 430, + 487, + 527 + ], + "type": "text", + "content": ", between a pair of entities of the same class. Each entity, " + }, + { + "bbox": [ + 130, + 430, + 487, + 527 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 130, + 430, + 487, + 527 + ], + "type": "text", + "content": ", is represented as a 3D plane segment comprising of 4 corners " + }, + { + "bbox": [ + 130, + 430, + 487, + 527 + ], + "type": "inline_equation", + "content": "\\{c_1, c_2, c_3, c_4\\}" + }, + { + "bbox": [ + 130, + 430, + 487, + 527 + ], + "type": "text", + "content": ". The distance between two entities, " + }, + { + "bbox": [ + 130, + 430, + 487, + 527 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 130, + 430, + 487, + 527 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 430, + 487, + 527 + ], + "type": "inline_equation", + "content": "E'" + }, + { + "bbox": [ + 130, + 430, + 487, + 527 + ], + "type": "text", + "content": " is computed as the maximum Euclidean distance between each corner and its counterpart assigned via Hungarian mathcing, i.e.: " + }, + { + "bbox": [ + 130, + 430, + 487, + 527 + ], + "type": "inline_equation", + "content": "d_{E}(E, E') = \\max \\{||c_{i} - c_{\\pi(i)}'| : i = 1, \\dots, 4\\}" + }, + { + "bbox": [ + 130, + 430, + 487, + 527 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 430, + 487, + 527 + ], + "type": "inline_equation", + "content": "\\pi(i)" + }, + { + "bbox": [ + 130, + 430, + 487, + 527 + ], + "type": "text", + "content": " is the permutation also found via Hungarian matching." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 528, + 479, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 528, + 479, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 528, + 479, + 552 + ], + "type": "text", + "content": "We threshold " + }, + { + "bbox": [ + 130, + 528, + 479, + 552 + ], + "type": "inline_equation", + "content": "d_{E}" + }, + { + "bbox": [ + 130, + 528, + 479, + 552 + ], + "type": "text", + "content": " to define the success criteria for the predicted entities. We compute the following metrics:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 137, + 560, + 479, + 608 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 137, + 560, + 479, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 560, + 479, + 583 + ], + "spans": [ + { + "bbox": [ + 137, + 560, + 479, + 583 + ], + "type": "text", + "content": "- F1 Score @ threshold - the F1 score of the set of predictions is computed at a single " + }, + { + "bbox": [ + 137, + 560, + 479, + 583 + ], + "type": "inline_equation", + "content": "d_{E}" + }, + { + "bbox": [ + 137, + 560, + 479, + 583 + ], + "type": "text", + "content": " threshold." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 137, + 585, + 479, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 585, + 479, + 608 + ], + "spans": [ + { + "bbox": [ + 137, + 585, + 479, + 608 + ], + "type": "text", + "content": "- Average F1 Score - the F1 score is computed across a range of entity distance thresholds and averaged." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "type": "text", + "content": "The scores are computed for each class independently and averaged to overall score. In addition, scores are computed for each scene and averaged across the dataset. We use the following range of thresholds (cm) for the average F1 scores: " + }, + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "type": "inline_equation", + "content": "T = \\{1, 2, \\dots, 9, 10, 15, 25, 30, 50, 75, 100\\}" + }, + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 399, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 399, + 91, + 447, + 102 + ], + "type": "text", + "content": "SceneScript" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 285 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 285 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 285 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 285 + ], + "type": "image", + "image_path": "9c35d5d28bc1dbf8b979708ca1fbc4ae07d8432fc5235dbeec91b9e95e5ec278.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 293, + 483, + 350 + ], + "lines": [ + { + "bbox": [ + 130, + 293, + 483, + 350 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 483, + 350 + ], + "type": "text", + "content": "Fig. 4: Qualitative samples between our model and SOTA methods on Aria Synthetic Environments's test set. Hierarchical methods like SceneCAD suffer from error cascading which leads to missing elements in the edge prediction module. RoomFormer (a 2D method extruded to 3D) primarily suffers from lightly captured scene regions which leave a unnoticeable signal in the density map." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 144, + 392, + 468, + 477 + ], + "blocks": [ + { + "bbox": [ + 130, + 361, + 483, + 383 + ], + "lines": [ + { + "bbox": [ + 130, + 361, + 483, + 383 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 483, + 383 + ], + "type": "text", + "content": "Table 2: Layout Estimation on Aria Synthetic Environments Quantitative comparison between our method and related recent work." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 144, + 392, + 468, + 477 + ], + "lines": [ + { + "bbox": [ + 144, + 392, + 468, + 477 + ], + "spans": [ + { + "bbox": [ + 144, + 392, + 468, + 477 + ], + "type": "table", + "html": "
MethodF1 @5cmAvg F1
meanwalldoorwindowmeanwalldoorwindow
SceneCAD '20 [1]-0.048---0.275--
RoomFormer '23 [37]0.1390.1590.1480.1100.4640.5050.4810.407
Ours (Point cloud)0.8480.9300.9220.6920.7840.8160.8110.724
Ours (Lifted features)0.9030.9430.9590.8060.8010.8180.8220.764
Ours (Image-only)0.6610.6870.7980.4970.7190.7270.7720.658
", + "image_path": "ec7df5318e3f0c045a7474c26ea2953e38c0a2c51fea0eface63a8472b5447b9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 498, + 254, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 254, + 510 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 254, + 510 + ], + "type": "text", + "content": "5.2 Layout Estimation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 521, + 482, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 580 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 580 + ], + "type": "text", + "content": "We perform scene layout estimation with the three encoder variants of SceneScript: a baseline model with sparse3d convolution pointcloud encoder, an RGB RayTran-based feature volume encoder [29], and our proposed lifted feature point encoder. The same transformer decoder is being used in all three scenarios." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 581, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 629 + ], + "type": "text", + "content": "To provide comparison to existing works, we include results from two baseline methods, namely SceneCAD [1] and the recent RoomFormer [37]. For these experiments SceneCAD and RoomFormer were both trained on Aria Synthetic Environments. Note that SceneCAD only predicts walls." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": "Table 2 shows the main results for our F1-based metrics on Aria Synthetic Environments. SceneScript exhibits a substantial performance advantage over SOTA layout estimation baselines across multiple metrics." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "text", + "content": "Avetisyan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 138, + 196, + 304, + 258 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 3: 3D Object Detection Performance comparison against state-of-the-art methods on an 3D object detection task trained and evaluated by F1-score at 0.25 and 0.5 IoU thresholds. By simply adding a make_bbox command SceneScript can achieve competitive object detection results." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 161, + 178, + 283, + 189 + ], + "lines": [ + { + "bbox": [ + 161, + 178, + 283, + 189 + ], + "spans": [ + { + "bbox": [ + 161, + 178, + 283, + 189 + ], + "type": "text", + "content": "(a) Aria Synthetic Environments" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 196, + 304, + 258 + ], + "lines": [ + { + "bbox": [ + 138, + 196, + 304, + 258 + ], + "spans": [ + { + "bbox": [ + 138, + 196, + 304, + 258 + ], + "type": "table", + "html": "
MethodInputF1@.25 IoU @.50 IoU
3DETR '21 [16]Points0.2010.078
Cube R-CNN '23 [2]RGB0.3940.228
ImVoxelNet '22 [24]RGB0.5840.516
OursPoints0.6200.577
", + "image_path": "4555838a8a76fdb99217d3189f976671d5f288e512ab2ba4226e8e7225e663a4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 307, + 196, + 474, + 254 + ], + "blocks": [ + { + "bbox": [ + 363, + 178, + 419, + 188 + ], + "lines": [ + { + "bbox": [ + 363, + 178, + 419, + 188 + ], + "spans": [ + { + "bbox": [ + 363, + 178, + 419, + 188 + ], + "type": "text", + "content": "(b) ScanNet [8]" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 196, + 474, + 254 + ], + "lines": [ + { + "bbox": [ + 307, + 196, + 474, + 254 + ], + "spans": [ + { + "bbox": [ + 307, + 196, + 474, + 254 + ], + "type": "table", + "html": "
MethodInputF1
@.25 IoU@.50 IoU
3DETR '21 [16]Points0.4800.349
3DETR-m '21 [16]Points0.5360.407
SoftGroup '22 [31]RGB Points0.6220.573
OursRGB Points0.5060.406
", + "image_path": "7e5865182b493beef94be74a1d64d964e9a68fb0200d097b6dfed937c38ccedc.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 281, + 479, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 479, + 317 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 479, + 317 + ], + "type": "text", + "content": "Both baseline methods encounter a significant decline in accuracy when dealing with finer details. See Figure 4 for qualitative comparisons between our method and baseline methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 337, + 480, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 337, + 480, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 337, + 480, + 384 + ], + "type": "text", + "content": "Encoder Ablation. The results demonstrate that SceneScript is robust to the encoding strategy chosen to encode the egocentric capture. It is able to infer highly accurate scene layouts in all configurations tested, and in each case SceneScript outperforms the included baselines by a significant margin." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 385, + 482, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 385, + 482, + 493 + ], + "spans": [ + { + "bbox": [ + 130, + 385, + 482, + 493 + ], + "type": "text", + "content": "Relative comparison of the encoder strategies reveals that leveraging the point clouds from a highly specialized and well-tuned system is still able to offer advantages of an, almost, entirely learned approach such as RayTran [29]. A light extension in the form of lifted image features can widen this gap even further. In particular, we observe that the discrepancy between the encoders becomes particularly apparent as the complexity of the scene increases in the form of increased room count. In the Appendix, we also show a quantitative evaluation of per-entity error distances, which aids in further attribution of relative performance gains between the encoding methods." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 512, + 248, + 524 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 512, + 248, + 524 + ], + "spans": [ + { + "bbox": [ + 131, + 512, + 248, + 524 + ], + "type": "text", + "content": "5.3 Object Detection" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 533, + 481, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 481, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 481, + 581 + ], + "type": "text", + "content": "In this section, we perform evaluation of SceneScript for object detection on both Aria Synthetic Environments and ScanNet [8]. For comparison, we include recent and state-of-the-art baseline methods, namely Cube-RCNN [2], ImVoxelNet [24], 3DETR [16], and SoftGroup [31]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": "Worth noting is that SceneScript does not predict a confidence score for each make_bbox command. This means that fair computation of the conventional mAP metric for this task is not possible, as detections cannot be ranked across scenes. Among other issues, this results in a metric that varies with the order in which scenes are evaluated. We therefore report F1-score-based metrics, which do not exhibit this order variance. Further discussion of this, and mAP numbers for the baselines for reference, can be found in the Appendix." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 399, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 399, + 91, + 447, + 102 + ], + "type": "text", + "content": "SceneScript" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 202, + 211 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 202, + 211 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 202, + 211 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 202, + 211 + ], + "type": "image", + "image_path": "814bbad6fa1cc611ca4aec168804d9a106aba6b1f4305b699bb40cb1ea365dc7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 220, + 482, + 266 + ], + "lines": [ + { + "bbox": [ + 130, + 220, + 482, + 266 + ], + "spans": [ + { + "bbox": [ + 130, + 220, + 482, + 266 + ], + "type": "text", + "content": "Fig. 5: Example scene reconstructions on scenes from Aria Synthetic Environments. (left) Visualisation of the decomposed meshes used to create make_prim training pairs. (right) Views of full scene predictions, as well as close ups highlighting the fidelity of object reconstruction through the prediction volumetric primitives enabled by make_prim." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 203, + 114, + 482, + 212 + ], + "blocks": [ + { + "bbox": [ + 203, + 114, + 482, + 212 + ], + "lines": [ + { + "bbox": [ + 203, + 114, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 203, + 114, + 482, + 212 + ], + "type": "image", + "image_path": "3f9db45871fddf264da288563194d4b18fcf1c128a19683f51f9876000c4d842.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 289, + 482, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 384 + ], + "type": "text", + "content": "In Table 3a, all methods were trained on Aria Synthetic Environments. 3DETR performs poorly due to its encoder being designed for relatively clean datasets such as ScanNet [8], while semi-dense point clouds from Aria Synthetic Environments exhibit more noise and less uniformity (see Figure 3 for an example). Cube R-CNN and ImVoxelNet both operate on RGB images, and detections are tracked for the entire sequence via a tracker [2] to provide competitive performance. In Table 3b, our method provides similar performance to both 3DETR and SoftGroup." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 385, + 482, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 385, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 385, + 482, + 445 + ], + "type": "text", + "content": "Through the addition of the make_bbox command, SceneScript demonstrates object detection performance on par with SOTA baselines on multiple object detection datasets. This result illustrates the extensibility of a token-based approach, and that our proposed SceneScript language representation does not suffer compared to specialised object detection network architectures." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 465, + 440, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 465, + 440, + 480 + ], + "spans": [ + { + "bbox": [ + 131, + 465, + 440, + 480 + ], + "type": "text", + "content": "6 Extending the SceneScript Structured Language" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 492, + 492, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 492, + 492, + 564 + ], + "spans": [ + { + "bbox": [ + 130, + 492, + 492, + 564 + ], + "type": "text", + "content": "A key advantage offered by SceneScript's structured language prediction paradigm is that the expressiveness of its reconstruction can be tailored without requiring a change to the method. Up to now, we have focussed on showcasing the efficacy of SceneScript for representing simple layout elements and objects as bounding boxes. In this section, we showcase this characteristic by increasing the fidelity of our scene representation by introducing coarse 3D object reconstruction." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 583, + 328, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 583, + 328, + 596 + ], + "spans": [ + { + "bbox": [ + 131, + 583, + 328, + 596 + ], + "type": "text", + "content": "6.1 Objects as Volumetric Primitives" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "We turn to a language based on volumetric primitives, motivated by works such as [28, 35]. Using simple primitives such as cuboids and extruded cylinders, enables us to coarsely represent arbitrary object categories while maintaining object semantics (e.g. tabletops can be represented by a single cuboid). Thus, this language can describe many object categories simultaneously." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "text", + "content": "Avetisyan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 223 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 223 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 223 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 223 + ], + "type": "image", + "image_path": "07e2ba398b0e28618caf4ace0b9efe68e8dc5331637f65fc02e672324a425a48.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 230, + 482, + 254 + ], + "lines": [ + { + "bbox": [ + 130, + 230, + 482, + 254 + ], + "spans": [ + { + "bbox": [ + 130, + 230, + 482, + 254 + ], + "type": "text", + "content": "Fig.6: Example scene reconstructions on real scenes with the addition of the make_prim command. Note that SceneScript was trained only on synthetic data." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 277, + 479, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 479, + 301 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 479, + 301 + ], + "type": "text", + "content": "This representation requires only one additional command over the layout and box commands already discussed previously, namely:" + } + ] + } + ], + "index": 4 + }, + { + "type": "code", + "bbox": [ + 132, + 308, + 478, + 344 + ], + "blocks": [ + { + "bbox": [ + 132, + 308, + 478, + 344 + ], + "lines": [ + { + "bbox": [ + 132, + 308, + 478, + 344 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 478, + 344 + ], + "type": "text", + "content": "make_prime:_bbox_id, prim_num, class, center_x, center_y, center_z, angle_x, angle_y, angle_z, scale_x, scale_y, scale_z" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "code_body" + } + ], + "index": 5, + "sub_type": "code", + "guess_lang": "txt" + }, + { + "bbox": [ + 130, + 350, + 482, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 482, + 398 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 482, + 398 + ], + "type": "text", + "content": "This command and its parameters describe a volumetric primitive (cuboid or extruded cylinder) via its 3D center, 3D rotation, and 3D scale. The prim_num parameter can be associated with semantics, e.g. tabletops of different tables typically have the same prim_num." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 415, + 482, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 482, + 513 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 482, + 513 + ], + "type": "text", + "content": "Dataset. To obtain ground truth make_prim commands that align with the objects in Aria Synthetic Environments, we first run an extension of Yang et al. [35] to obtain cuboid and extruded cylinder primitives of a database of 3D CAD models (ABO [7], which was used to populate Aria Synthetic Environments). See Figure 5 (left) for example decompositions. For this proof-of-concept experiment, we use three categories: table, chair, and sofa. We then convert these decomposed primitives into make_prim commands that are aligned with the objects in the dataset, which results in training pairs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 530, + 482, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 530, + 482, + 591 + ], + "spans": [ + { + "bbox": [ + 130, + 530, + 482, + 591 + ], + "type": "text", + "content": "Results. We show qualitative results on Aria Synthetic Environments in Figure 5. In Figure 6, we show inferences in a few real-world environments despite only having trained on our simulated dataset. These results demonstrate that SceneScript's general purpose architecture is capable of coarsely reconstructing objects of multiple categories through addition of a new command." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 609, + 257, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 609, + 257, + 620 + ], + "spans": [ + { + "bbox": [ + 131, + 609, + 257, + 620 + ], + "type": "text", + "content": "6.2 Further Extensions" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "In this section, we have explored just one extension of SceneScript's structured scene language in order to demonstrate its flexibility. The result was greatly increased expressiveness of the scene reconstruction. In the Appendix, we include" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 399, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 399, + 91, + 447, + 102 + ], + "type": "text", + "content": "SceneScript" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": "additional explorations that can further increase the fidelity and accuracy of the reconstructions. These explorations include reconstructing curved walls, inferring object states (e.g., door opening angles), as well as direct prediction of parametric models using object models deployed commonly by tech artists (e.g., blender geometry nodes) for object reconstruction." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 193, + 350, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 193, + 350, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 193, + 350, + 205 + ], + "type": "text", + "content": "7 Interactive Scene Reconstruction" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 215, + 482, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 215, + 482, + 324 + ], + "spans": [ + { + "bbox": [ + 130, + 215, + 482, + 324 + ], + "type": "text", + "content": "Scene reconstruction often occurs offline, on pre-recorded walkthroughs of an environment, or derivatives of them, such as a point cloud. However, we take advantage of SceneScript's inference occurring at an interactive rate, which takes between 2-5s depending on the size of the environment, by implementing streaming of SceneScript's live reconstructions into a modern VR headset. This enables an interactive experience where the user can see the reconstruction overlaid onto the environment they are exploring in real-time. See the video recording on the website for examples. Visualisations of this interface are included in Figure 1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 340, + 329, + 353 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 340, + 329, + 353 + ], + "spans": [ + { + "bbox": [ + 132, + 340, + 329, + 353 + ], + "type": "text", + "content": "8 Limitations and Future Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 362, + 482, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 362, + 482, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 362, + 482, + 518 + ], + "type": "text", + "content": "SceneScript exhibits certain limitations that should be acknowledged. First, the structured language commands are manually defined, which requires human intervention at this stage. Secondly, due to the higher-level nature of our commands, it can be challenging to capture fine-grained geometric details with extremely high precision (i.e. mm). As a consequence, the resulting reconstructions based on this representation tend to lead to simpler and coarser geometries, potentially missing intricate nuances at the very high detail level. These limitations potentially highlight areas for future research and optimization, aiming to automate the command definition process and explore techniques to enhance the representation's ability to capture intricate geometric details accurately. However, we believe that the ability to built scene representations that are based on structured language commands will be a key part in complex and efficient methods of scene reconstruction in the future." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 534, + 220, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 220, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 220, + 547 + ], + "type": "text", + "content": "9 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 558, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 667 + ], + "type": "text", + "content": "We introduced SceneScript, a novel reconstruction method that is based on a tokenized scene representation. SceneScript autoregressively predicts a sequence of structured scene language commands given a video stream of an indoor space. This tokenized scene representation is compact, editable and interpretable. In addition, we showed that a strength of SceneScript is the ability to extend to arbitrary elements (e.g. Bezier curves, object part decomposition) with minimal changes. This research opens up new directions in representing 3D scenes as language, bringing the 3D reconstruction community closer to the recent advances in large language models such as GPT-class models [20]." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "text", + "content": "Avetisyan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 140, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 481, + 183 + ], + "type": "text", + "content": "1. Avetisyan, A., Khanova, T., Choy, C., Dash, D., Dai, A., Nießner, M.: Scenecad: Predicting object alignments and layouts in rgb-d scans. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16. pp. 596-612. Springer (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 184, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 184, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 138, + 184, + 481, + 216 + ], + "type": "text", + "content": "2. Brazil, G., Kumar, A., Straub, J., Ravi, N., Johnson, J., Gkioxari, G.: Omni3D: A large benchmark and model for 3D object detection in the wild. In: CVPR. IEEE, Vancouver, Canada (June 2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 217, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 217, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 138, + 217, + 481, + 248 + ], + "type": "text", + "content": "3. Cabral, R., Furukawa, Y.: Piecewise planar and compact floorplan reconstruction from images. In: 2014 IEEE Conference on Computer Vision and Pattern Recognition (2014)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 249, + 481, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 249, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 138, + 249, + 481, + 282 + ], + "type": "text", + "content": "4. Carlier, A., Danelljan, M., Alahi, A., Timofte, R.: Deepsvg: A hierarchical generative network for vector graphics animation. In: Advances in Neural Information Processing Systems (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 282, + 481, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 282, + 481, + 315 + ], + "spans": [ + { + "bbox": [ + 138, + 282, + 481, + 315 + ], + "type": "text", + "content": "5. Chen, J., Liu, C., Wu, J., Furukawa, Y.: Floor-sp: Inverse cad for floorplans by sequential room-wise shortest path. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 316, + 481, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 316, + 481, + 347 + ], + "spans": [ + { + "bbox": [ + 138, + 316, + 481, + 347 + ], + "type": "text", + "content": "6. Chen, T., Saxena, S., Li, L., Fleet, D.J., Hinton, G.: Pix2seq: A language modeling framework for object detection. In: International Conference on Learning Representations (ICLR) (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 348, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 348, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 138, + 348, + 481, + 380 + ], + "type": "text", + "content": "7. Collins, J., Goel, S., Deng, K., Luthra, A., Xu, L., Gundogdu, E., Zhang, X., Yago Vicente, T.F., Dideriksen, T., Arora, H., Guillaumin, M., Malik, J.: Abo: Dataset and benchmarks for real-world 3d object understanding. CVPR (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 381, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 381, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 138, + 381, + 481, + 413 + ], + "type": "text", + "content": "8. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proc. Computer Vision and Pattern Recognition (CVPR), IEEE (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 414, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 414, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 138, + 414, + 481, + 435 + ], + "type": "text", + "content": "9. Furukawa, Y., Curless, B., Seitz, S.M., Szeliski, R.: Manhattan-world stereo. In: 2009 IEEE Conference on Computer Vision and Pattern Recognition (2009)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 436, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 436, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 436, + 481, + 468 + ], + "type": "text", + "content": "0. Furukawa, Y., Curless, B., Seitz, S.M., Szeliski, R.: Reconstructing building interiors from images. In: 2009 IEEE 12th international conference on computer vision (2009)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 468, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 468, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 138, + 468, + 481, + 501 + ], + "type": "text", + "content": "1. Ganin, Y., Bartunov, S., Li, Y., Keller, E., Saliceti, S.: Computer-aided design as language. Advances in Neural Information Processing Systems 34, 5885-5897 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 502, + 481, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 502, + 481, + 523 + ], + "spans": [ + { + "bbox": [ + 138, + 502, + 481, + 523 + ], + "type": "text", + "content": "2. Girshick, R.: Fast r-cnn. In: Proceedings of the IEEE international conference on computer vision. pp. 1440-1448 (2015)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 523, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 523, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 523, + 481, + 555 + ], + "type": "text", + "content": "3. Lee, C.Y., Badrinarayanan, V., Malisiewicz, T., Rabinovich, A.: Roomnet: End-to-end room layout estimation. In: Proceedings of the IEEE international conference on computer vision (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 556, + 460, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 556, + 460, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 556, + 460, + 567 + ], + "type": "text", + "content": "4. Meta: Project aria. https://projectaria.com/ (2022), accessed: 2023-08-30" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 567, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 481, + 599 + ], + "type": "text", + "content": "5. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 600, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 600, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 600, + 481, + 632 + ], + "type": "text", + "content": "6. Misra, I., Girdhar, R., Joulin, A.: An end-to-end transformer model for 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2906-2917 (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "text", + "content": "7. Murali, S., Speciale, P., Oswald, M.R., Pollefeys, M.: Indoor scan2bim: Building information models of house interiors. In: 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) (2017)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 399, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 399, + 91, + 447, + 102 + ], + "type": "text", + "content": "SceneScript" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "18. Nash, C., Ganin, Y., Eslami, S.A., Battaglia, P.: Polygen: An autoregressive generative model of 3d meshes. In: International conference on machine learning. pp. 7220-7229. PMLR (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 182 + ], + "type": "text", + "content": "19. Ochmann, S., Vock, R., Klein, R.: Automatic reconstruction of fully volumetric 3d building models from oriented point clouds. ISPRS journal of photogrammetry and remote sensing 151, 251-262 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 307, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 307, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 307, + 193 + ], + "type": "text", + "content": "20. OpenAI: Gpt-4 technical report (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "text", + "content": "21. Para, W., Bhat, S., Guerrero, P., Kelly, T., Mitra, N., Guibas, L.J., Wonka, P.: Sketchgen: Generating constrained cad sketches. In: Advances in Neural Information Processing Systems (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 228, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 228, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 480, + 258 + ], + "type": "text", + "content": "22. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 260, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 260, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 260, + 480, + 293 + ], + "type": "text", + "content": "23. Pintore, G., Agus, M., Gobbetti, E.: Atlantanet: inferring the 3d indoor layout from a single 360 image beyond the Manhattan world assumption. In: Proceedings of the European conference on computer vision (ECCV) (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 335 + ], + "type": "text", + "content": "24. Rukhovich, D., Vorontsova, A., Konushin, A.: Imvoxelnet: Image to voxels projection for monocular and multi-view general-purpose 3d object detection. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2397-2406 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 337, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 337, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 337, + 480, + 369 + ], + "type": "text", + "content": "25. Somasundaram, K., Dong, J., Tang, H., Straub, J., Yan, M., Goesele, M., Engel, J.J., De Nardi, R., Newcombe, R.: Project aria: A new tool for egocentric multimodal ai research. arXiv preprint arXiv:2308.13561 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "type": "text", + "content": "26. Tang, H., Liu, Z., Li, X., Lin, Y., Han, S.: TorchSparse: Efficient Point Cloud Inference Engine. In: Conference on Machine Learning and Systems (MLSys) (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 392, + 480, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 423 + ], + "type": "text", + "content": "27. Tang, H., Liu, Z., Zhao, S., Lin, Y., Lin, J., Wang, H., Han, S.: Searching Efficient 3D Architectures with Sparse Point-Voxel Convolution. In: European Conference on Computer Vision (ECCV) (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 425, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 425, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 425, + 480, + 456 + ], + "type": "text", + "content": "28. Tulsiani, S., Su, H., Guibas, L.J., Efros, A.A., Malik, J.: Learning shape abstractions by assembling volumetric primitives. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 457, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 511 + ], + "type": "text", + "content": "29. Tyszkiiewicz, M.J., Maninis, K.K., Popov, S., Ferrari, V.: Raytran: 3d pose estimation and shape reconstruction of multiple objects from videos with ray-traced transformers. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part X. pp. 211-228. Springer (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "type": "text", + "content": "30. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. NeurIPS 30 (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "type": "text", + "content": "31. Vu, T., Kim, K., Luu, T.M., Nguyen, X.T., Yoo, C.D.: Softgroup for 3d instance segmentation on 3d point clouds. In: CVPR (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 556, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 556, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 556, + 480, + 588 + ], + "type": "text", + "content": "32. Wu, R., Xiao, C., Zheng, C.: Deepcad: A deep generative network for computer-aided design models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 589, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 480, + 632 + ], + "type": "text", + "content": "33. Xu, X., Willis, K.D., Lambourne, J.G., Cheng, C.Y., Jayaraman, P.K., Furukawa, Y.: Skexgen: Autoregressive generation of cad construction sequences with disentangled codebooks. In: International Conference on Machine Learning (ICML) (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "34. Xue, Y., Mao, J., Niu, M., Xu, H., Mi, M.B., Zhang, W., Wang, X., Wang, X.: Point2seq: Detecting 3d objects as sequences. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8521-8530 (2022)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "text", + "content": "Avetisyan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 249 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "35. Yang, K., Chen, X.: Unsupervised learning for cuboid shape abstraction via joint segmentation from point clouds. ACM Transactions on Graphics (TOG) 40(4), 1-11 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 182 + ], + "type": "text", + "content": "36. Yin, T., Zhou, X., Krahenbuhl, P.: Center-based 3d object detection and tracking. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 482, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 482, + 216 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 482, + 216 + ], + "type": "text", + "content": "37. Yue, Y., Kontogianni, T., Schindler, K., Engelmann, F.: Connecting the dots: Floorplan reconstruction using two-level queries. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 216, + 482, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 216, + 482, + 249 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 482, + 249 + ], + "type": "text", + "content": "38. Zou, C., Colburn, A., Shan, Q., Hoiem, D.: Layoutnet: Reconstructing the 3d room layout from a single rgb image. In: Proceedings of the IEEE conference on computer vision and pattern recognition (2018)" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 399, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 399, + 91, + 447, + 102 + ], + "type": "text", + "content": "SceneScript" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/SceneTeller_ Language-to-3D Scene Generation/dec651d5-2a06-4ad2-af7e-aeaf7a7dadd2_content_list.json b/2024/SceneTeller_ Language-to-3D Scene Generation/dec651d5-2a06-4ad2-af7e-aeaf7a7dadd2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..342f80fec387c4eb1c39cc104f76b0f9b5536320 --- /dev/null +++ b/2024/SceneTeller_ Language-to-3D Scene Generation/dec651d5-2a06-4ad2-af7e-aeaf7a7dadd2_content_list.json @@ -0,0 +1,1939 @@ +[ + { + "type": "text", + "text": "SceneTeller: Language-to-3D Scene Generation", + "text_level": 1, + "bbox": [ + 233, + 141, + 769, + 162 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bashak Melis Öcal1*, Maxim Tatarchenko2, Sezer Karaoglu1, and Theo Gevers1", + "bbox": [ + 217, + 191, + 782, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 UvA-Bosch Delta Lab, University of Amsterdam", + "bbox": [ + 330, + 218, + 671, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Bosch Center for AI, Robert Bosch GmbH", + "bbox": [ + 349, + 233, + 648, + 247 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/77fbd33969abdf6afa836f2b511b9d28687c1140c5661da868af46dd949cc097.jpg", + "image_caption": [ + "Fig. 1: With solely a textual prompt in natural language describing the desired spatial positions and orientations of objects within the scene, SceneTeller is able to generate realistic and high-quality 3D spaces. Additionally, by facilitating modifications to the style of the entire scene or individual objects within the scene through edit instructions, SceneTeller offers a practical and flexible framework for designing personalized rooms." + ], + "image_footnote": [], + "bbox": [ + 233, + 301, + 766, + 679 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "\\* Correspondence to $<$ b.m.ocal@uva.nl>", + "bbox": [ + 218, + 825, + 503, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Designing high-quality indoor 3D scenes is important in many practical applications, such as room planning or game development. Conventionally, this has been a time-consuming process which requires both artistic skill and familiarity with professional software, making it hardly accessible for layman users. However, recent advances in generative AI have established solid foundation for democratizing 3D design. In this paper, we propose a pioneering approach for text-based 3D room design. Given a prompt in natural language describing the object placement in the room, our method produces a high-quality 3D scene corresponding to it. With an additional text prompt the users can change the appearance of the entire scene or of individual objects in it. Built using in-context learning, CAD model retrieval and 3D-Gaussian-Splatting-based stylization, our turnkey pipeline produces state-of-the-art 3D scenes, while being easy to use even for novices. Our project page is available at https://sceneteller.github.io/.", + "bbox": [ + 261, + 147, + 743, + 356 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Keywords: Text-to-3D scene generation $\\cdot$ Language-guided generation $\\cdot$ 3D Gaussian splatting scene stylization", + "bbox": [ + 261, + 369, + 740, + 397 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 426, + 375, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D scene design has numerous applications in domains like architecture or game development. Designing high-quality 3D scenes, however, is a laborious and time-consuming process, challenging even for highly-skilled domain experts utilizing professional tools. This complexity often leaves novice users unable to customize their own spaces. Recent advancements in automatic text-to-image synthesis [22,41] have led to massive progress in text-to-3D creation [36-38], enabling the generation of 3D content only from text input. Such works create foundation for democratizing 3D design, supporting its accessibility for non-experts.", + "bbox": [ + 212, + 460, + 787, + 582 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A pioneering work by Poole et al. [36] introduced object-centric text-to-3D generation based on 2D diffusion models coupled with NeRFs [31]. Several subsequent works proposed ways to improve the geometric accuracy, multi-view consistency, and appearance quality [6,27,30,50]. Despite their success in object-level generation, hallucinating complex large-scale scenes with multiple objects is still remaining a challenging problem. Some recent 3D scene generation methods [11,35,52] require a 3D layout as input, limiting their usability for novices. Others [16,23] produce results that violate global object arrangements and are not multi-view consistent.", + "bbox": [ + 212, + 582, + 787, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Aiming to address both these shortcomings - insufficient user friendliness and unsatisfactory global consistency - we introduce SceneTeller, a framework for text-based 3D scene generation. 1) We start with a premise that a 3D scene generation framework should offer control over the placement and style of individual objects, allowing novice users to design their customized spaces through simple interactions with the system. Our approach achieves this by relying on text-based user interaction for scene and appearance specification. 2) In order to generate compositionally plausible 3D scenes, it is important to incorporate", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "B.M. Öcal et al.", + "bbox": [ + 271, + 113, + 380, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "some degree of global information guidance. We implement this by conditioning the method on global layout descriptions.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To this end, we split the task of generating high-quality 3D scenes into three steps: 3D layout generation, object population and style editing. 1) When presented with a prompt describing the desired spatial positions and orientations of objects, SceneTeller generates an initial 3D layout of the scene by exploiting the Large Language Models (LLMs) through in-context learning, to ensure global consistency. 2) The generated layout is then used to assemble a 3D scene and populate it with CAD models from a database. 3) To facilitate scene editing, the assembled scene is fitted with a 3D Gaussian Splitting [24] representation. This representation is then used to stylize the scene according to the user-provided edit prompt targeting the entire scene or individual objects within it.", + "bbox": [ + 212, + 176, + 785, + 327 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Extensive quantitative and qualitative evaluations show that SceneTeller performs significantly better than state-of-the-art methods, offering superior geometric fidelity and compositional plausibility. In brief, our contributions to the existing body of work can be summarized as follows:", + "bbox": [ + 212, + 328, + 785, + 388 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a turnkey pipeline for high-quality 3D scene generation yielding complex, realistic and globally consistent 3D scenes, given only scene descriptions in natural language. Our pipeline supports object- and scene-level appearance editing.", + "- We propose a novel LLM-based module for 3D layout generation trained through in-context learning, which provides control over individual object placement.", + "- We experimentally demonstrate that SceneTeller generates 3D scenes of higher quality and better consistency than state-of-the-art methods, while boasting unprecedented user friendliness and flexibility." + ], + "bbox": [ + 225, + 395, + 784, + 544 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 214, + 565, + 387, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 3D Scene Representations", + "text_level": 1, + "bbox": [ + 214, + 595, + 477, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Rendering 3D scenes using implicit neural networks has seen remarkable success, notably with NeRF's [31] photo-realistic image synthesis. Extensions of NeRF improve reconstruction quality [3, 25], training and rendering speed [5, 17, 32, 39, 43, 55], and handling large-scale scenes [7, 45]. Despite advancements, NeRF-based methods still struggle with slow rendering speeds. Recently, 3D Gaussian splatting (3DGS) was proposed for real-time, high-quality rendering using anisotropic 3D Gaussians and fast tile-based rasterization. Our framework employs 3DGS for scene representation and allows edits on these Gaussians, generating high-quality, 3D consistent scenes.", + "bbox": [ + 212, + 616, + 785, + 752 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 Text-to-3D Generation", + "text_level": 1, + "bbox": [ + 214, + 772, + 450, + 786 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Object-centric Generation. 3D generation from textual prompts has recently grown in popularity. DreamFusion [36] introduces Score Distillation Sampling (SDS) to optimize NeRF representations by leveraging an off-the-shelf", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "SceneTeller: Language-to-3D Scene Generation", + "bbox": [ + 418, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2D diffusion model, paving the way towards text-driven 3D content creation [29,37,38,57]. Following [36], Magic3D [27] employs a coarse-to-fine optimization scheme, exploiting low-resolution diffusion priors, and then enhancing the texture. To achieve high-fidelity mesh generation, Fantasia3D [6] disentangles the appearance from geometry. Latent-NeRF [30] optimizes a NeRF in the latent space of StableDiffusion [41], and additionally proposes incorporating 3D priors to generate a target geometry. To address the over-smoothing, over-saturation and lack of diversity problems within the SDS, ProfilicDreamer [50] proposes a principled particle-based variational framework, namely Variational Score Distillation (VSD). Following the same motivation, [49] presents Perturb-and-Average Scoring to estimate the score for non-noisy images.", + "bbox": [ + 212, + 146, + 787, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the recent introduction of 3D Gaussian Splatting [24], several works attempt to combine 3D Gs and diffusion models. GaussianDreamer [54] employs a 3D diffusion model to generate the initialized point clouds for 3D Gaussians, and then further optimize the Gaussians using SDS with a 2D diffusion model. [9, 47] follows a two-stage pipeline for geometry optimization and appearance refinement. GaussianDiffusion [26] focuses on the challenge of achieving multiview consistency in 3D generation and proposes combining structured noise with a variational 3DGS technique. Despite their success in generating object-level scenes, both NeRF-based and 3DGS-based methods fall short in hallucinating complex scenes with multiple objects.", + "bbox": [ + 212, + 313, + 789, + 464 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Scene Generation. To represent multiple objects within the scenes, existing works [11, 28, 35, 52] rely on compositional NeRFs that are initialized based on user-defined object proxies. However, they are usually restricted to produce only narrow inward-facing views of the rooms. Also, the requirement of having 3D layouts as input is restrictive for final users. Orthogonal to those, SceneScape [16] and Text2Room [23] generate scenes following the render-refine-repeat paradigm by iteratively warping and inpainting the previously generated image according to the specified camera motion. This strategy does not allow object-level control during scene generation, and the results tend to exhibit multi-view inconsistencies. LucidDreamer [10] employs the same paradigm for an initial point cloud reconstruction, and then further optimizes the scene using 3DGS. Recently, several works introduced techniques for 3DGS scene editing, by manipulating the training images using 2D diffusion models [8, 13, 48] and incorporating the edited images into 3DGS scenes for further optimization. Concurrent to our work, [56] initializes a 3DGS scene from the layouts created using LLMs and optimizes it using a diffusion model. However, [56] lacks control over the generated layout in terms of positions and orientations of objects.", + "bbox": [ + 212, + 482, + 789, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 LLMs for Vision", + "text_level": 1, + "bbox": [ + 214, + 757, + 398, + 772 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the inclusion of grounding information in LLMs, they demonstrate enhanced reasoning capabilities for visual concepts. Recently, the visual reasoning abilities of LLMs have been utilized for 3D tasks such as avatar generation [40], instruction-driven 3D modeling [44], 3D editing [14]. Exploiting the", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "B.M. Öcal et al.", + "bbox": [ + 271, + 113, + 380, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "spatial reasoning capabilities of LLMs, LayoutGPT [15] extends beyond its 2D counterparts [34] to generate random 3D layouts of specific room types through in-context learning. However, LayoutGPT lacks support for controlling the generated layout. The 3D layout generation module of SceneTeller is the first of its kind allowing control over individual object placement in accordance with the provided natural language prompts describing the desired positions of objects.", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 258, + 330, + 273 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7cea33ba983928d37ab833debb0d181fe011cb0a4db0f3442d8482c6bc6131a6.jpg", + "image_caption": [ + "Fig. 2: Overview of our method. Given a textual prompt in natural language delineating the desired spatial positions and orientations of objects within the scene, a 3D scene layout is generated using in-context learning. An initial 3D scene is assembled for the predicted layout, which is then fitted with a 3D Gaussian Splatting representation. This representation is then used to stylize the scene according to the user-provided text prompt, and subsequently render the final images of the scene." + ], + "image_footnote": [], + "bbox": [ + 225, + 313, + 777, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The pipeline of SceneTeller is divided into three main stages: language-driven 3D layout generation, 3D scene assembling from the layout and 3D scene stylization. In the first stage, a 3D scene layout is generated based on the given text prompt (Sec. 3.1). Then, an initial 3D scene is assembled for the predicted layout (Sec. 3.2). After that, a 3D Gaussian Splatting representation is fitted to the assembled scene. This representation is then used to stylize the scene according to the user-provided text prompt, and subsequently render the final images of the scene (Sec. 3.3. The scheme of the framework is illustrated in Fig. 2", + "bbox": [ + 212, + 643, + 787, + 765 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Language-driven 3D Layout Generation", + "text_level": 1, + "bbox": [ + 214, + 786, + 589, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The task of this module is to produce a collection of object bounding boxes conditioned on a brief textural description of their spatial positions within the", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "SceneTeller: Language-to-3D Scene Generation", + "bbox": [ + 418, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "scene in plain language. More formally, the input is a condition $\\mathcal{C}$ , consisting of a textual description $\\mathcal{V}$ , along with the scene type and scene dimensions information. The output is a 3D layout represented as a set of 3D bounding boxes $\\mathbf{b}_i \\in B = \\{\\mathbf{b}_i\\}_{i=1}^N$ for each of the $N$ referenced objects. Each 3D bounding box $\\mathbf{b}_i = (c_i, \\mathbf{t}_i, \\mathbf{s}_i, o_i)$ consists of the following components: category name $c_i$ , box center $\\mathbf{t}_i = (x_i, y_i, z_i) \\in \\mathbb{R}^3$ , box dimensions $\\mathbf{s}_i = (w_i, h_i, d_i) \\in \\mathbb{R}^3$ and orientation $o_i \\in \\mathbb{R}$ .", + "bbox": [ + 212, + 146, + 787, + 251 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The textual description $\\mathcal{V}$ is comprised of multiple sentences, each specifying the position and orientation of a single object within a canonical coordinate system. The canonical system is chosen to represent the scene in a 2D perspective, to ensure that the descriptions closely resemble those produced by a human observer. $\\mathcal{V}$ is initially generated using a rule-based approach formulated in three steps as follows: 1) The canonical coordinate system is partitioned into a $3\\times 3$ grid, creating rectangular patches. 2) Each of the predicted objects is assigned to a patch based on the center position $\\mathbf{t}_i$ and orientation $o_i$ of its bounding box within these patches. 3) Per-object sentences are then formulated according to which patch the object is assigned to and object's orientation, e.g. for the object assigned to the top-left patch, with category name $c_i$ and 90 degrees orientation, we generate: \"A $c_i$ is placed at the top-left corner of the room, with a perpendicular orientation.\" Given the relatively limited linguistic variation inherent in rule-based descriptions, initial textual templates are provided to an LLM to serve as a basis for paraphrasing.", + "bbox": [ + 212, + 252, + 787, + 478 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "LayoutGPT [15] focused on improving an LLM's interpretation of the spatial knowledge by using a structured format featuring attribute-value pairs. We follow a similar strategy and adopt the standard CSS (Cascading Style Sheets) format to describe $\\mathbf{b}_i$ . Bounding box values are mapped to standard CSS format attributes and category name is employed as the selector.", + "bbox": [ + 212, + 479, + 787, + 554 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Prompt Construction. Recent progress in the field has showcased the impressive in-context learning capabilities of LLMs, which entail conditioning these models with both natural language instructions and a select set of task demonstrations. Following this conditioning, the LLM is then tasked with completing further instances based on the provided demonstrations. In line with this, our prompts to LLM include three main parts: task specifications, in-context exemplars and the query condition:", + "bbox": [ + 212, + 575, + 787, + 681 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Task specifications: Following prior work [42, 51] on enhancing LLMs instruction following capabilities, a task description is incorporated at the beginning of each prompt, which explains the goal of the task, establishes a standard for the 3D layout format in CSS style and provides unit information for attributes. Additionally, constraints are integrated into the prompts to guide the LLM and minimize errors (e.g., predicting overlapping boxes or placing objects out of bounds), during task completion.", + "- In-context learning: Supporting exemplars for the in-context learning are selected by adopting the retrieval-based approach used in [2, 15, 53]. When provided with a set of supporting exemplars $S = \\{(\\mathcal{C}_m^s,\\mathbf{b}_m^s)\\}_{m = 1}^M$ and the" + ], + "bbox": [ + 223, + 688, + 787, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "B.M. Öcal et al.", + "bbox": [ + 271, + 113, + 380, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "queried condition $\\mathcal{C}_q$ , the function $f(\\mathcal{C}_k^s,\\mathcal{C}_q) = \\|rl_k - rl_q\\|^2 + \\|rw_k - rw_q\\|^2$ is computed between each element of the set and $\\mathcal{C}_q$ following [15], where $rl$ and $rw$ are the length and width of the scenes. Top- $k$ supporting exemplars with the shortest distance to $\\mathcal{C}_q$ are selected for in-context learning, provided to LLM with the same format with $\\mathcal{C}_q$ .", + "bbox": [ + 240, + 143, + 782, + 220 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "- Query condition: The inference condition $\\mathcal{C}_q$ , for which we want to predict the layout.", + "bbox": [ + 225, + 220, + 782, + 251 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 3D Scene Assembling from Layouts", + "text_level": 1, + "bbox": [ + 215, + 273, + 552, + 289 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To assemble a scene based on the generated layouts, we rely on object retrieval. More specifically, for each of the bounding boxes, the nearest 3D model of the corresponding predicted category is retrieved from the furniture dataset based on the Euclidean distance of the bounding box dimensions as follows:", + "bbox": [ + 212, + 299, + 782, + 359 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nd \\left(\\mathbf {s} _ {i}, \\mathbf {s} _ {\\text {c a n d}}\\right) = \\sqrt {\\left(w _ {i} - w _ {\\text {c a n d}}\\right) ^ {2} + \\left(h _ {i} - h _ {\\text {c a n d}}\\right) ^ {2} + \\left(d _ {i} - d _ {\\text {c a n d}}\\right) ^ {2}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 272, + 387, + 784, + 404 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathbf{s}_{\\text{cand}}$ denotes the bounding box dimensions for the candidate object from the dataset. The retrieved 3D model is placed within the scene according to the predicted box center $\\mathbf{t}_i$ and orientation angle $o_i$ .", + "bbox": [ + 212, + 415, + 782, + 460 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 3D Scene Stylization", + "text_level": 1, + "bbox": [ + 215, + 482, + 431, + 497 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Given a generated 3D scene, the users should be able to modify either the overall appearance of the scene or specific objects within it. Our approach facilitates scene editing by training a 3D Gaussian splatting model representing the generated scene and refining this representation according to the provided edit instruction, offering a flexible framework for 3D scene generation.", + "bbox": [ + 212, + 508, + 782, + 583 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3D Gaussian Splating Scene Representation. 3D Gaussian splatting (3DGS) [24] is a recent pioneering work for novel view generation, which represents the underlying scene as a collection of anisotropic 3D Gaussians defined by their center positions $\\mu \\in \\mathbb{R}^3$ and 3D covariance matrices $\\pmb{\\Sigma}$ parameterized as:", + "bbox": [ + 212, + 604, + 782, + 666 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\Sigma} = \\mathbf {R} \\mathbf {S} \\mathbf {S} ^ {\\mathrm {T}} \\mathbf {R} ^ {\\mathrm {T}}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 447, + 680, + 784, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$\\mathbf{R}$ denotes the rotation matrix and $\\mathbf{S}$ is the scale matrix. Each 3D Gaussian is assigned a color $c$ represented with spherical harmonics (SH) coefficients, to capture the view-dependent appearance. To allow $\\alpha$ -blending of splats, Gaussians are associated with an opacity value $\\alpha \\in \\mathbb{R}$ .", + "bbox": [ + 212, + 704, + 782, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Unlike NeRF [31] which relies on volume rendering, 3DGS enables faster training and rendering through differentiable rasterization. A set of 3D Gaussians is rendered by projecting into the camera's image plane as 2D Gaussians, and assigned to individual image tiles. The color of each pixel $\\mathbf{p}$ on the image plane is then determined as follows:", + "bbox": [ + 212, + 763, + 782, + 838 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "SceneTeller: Language-to-3D Scene Generation", + "bbox": [ + 418, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nC (\\mathbf {p}) = \\sum_ {i \\in \\mathcal {N}} c _ {i} \\sigma_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\sigma_ {j}), \\quad \\sigma_ {i} = \\alpha_ {i} e ^ {- \\frac {1}{2} \\left(\\mathbf {p} - \\mu_ {i}\\right) ^ {T} \\Sigma_ {i} ^ {- 1} \\left(\\mathbf {p} - \\mu_ {i}\\right)} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 289, + 159, + 784, + 200 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathcal{N}$ denotes the Gaussians in this tile, $\\sigma_{i}$ represents the influence of the Gaussian on the image pixel and $\\mu_{i},\\Sigma_{i},c_{i},\\alpha_{i}$ are the position, covariance, color and opacity of the $i$ -th Gaussian respectively. For optimization, a weighted combination of $\\mathcal{L}_1$ loss and SSIM loss is employed.", + "bbox": [ + 212, + 205, + 782, + 266 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To initialize the centers of the 3D Gaussians, we utilize the mesh vertices of the retrieved objects and create a scene point cloud. This generated scene point cloud replaces the Structure-from-Motion (SfM) points used in the original work. To obtain the required training images for 3DGS optimization, we render the generated scene into RGB images $I$ . We additionally render 2D segmentation masks $\\mathcal{M}$ to be used for 3D Gaussian scene editing.", + "bbox": [ + 212, + 267, + 784, + 357 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Scene Editing. For editing our 3DGS scenes based on the provided edit instruction, we build on top of Instruct-GS2GS [48] which in turn extends Instruct-NeRF2NeRF [20] for 3DGS scenes. Given the images used for training the 3DGS scene, Instruct-GS2GS updates the images individually using a diffusion model, namely Instruct-Pix2Pix [4]. The diffusion model takes three inputs: 1) an unedited conditioning image $I_0^v$ , a text instruction $c_{T}$ and the noisy version of the current render $I_i^v$ at optimization step $i$ , where $v$ denotes the viewpoint from which the images are captured. Formally, the process of updating a single image is defined as:", + "bbox": [ + 212, + 380, + 784, + 517 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nI _ {i + 1} ^ {v} \\leftarrow U _ {\\theta} \\left(I _ {i} ^ {v}, t; I _ {0} ^ {v}, c _ {T}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 418, + 532, + 784, + 549 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $t$ is the noise level within the constant range $[t_{\\mathrm{min}}, t_{\\mathrm{max}}]$ , $U_{\\theta}$ is the sampling process of DDIM [41] and $I_{i+1}^{v}$ is the edited image respecting the text instruction $c_{T}$ and the unedited conditioning image $I_{0}^{v}$ . The training is performed by editing all the training images during a dataset update at every 2.5k training iterations.", + "bbox": [ + 212, + 556, + 784, + 616 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Despite achieving general scene-level stylization, [48] lacks of support for more fine-grained object-level editing. To enable object-level edits, we use the binary masks $m_{\\{o\\}_{k=1}^{K}}$ . Each $m_{\\{o\\}_{k=1}^{K}}$ is obtained by binarizing 2D segmentation masks $\\mathcal{M}$ for the set of objects to edit $\\{o\\}_{k=1}^{K}$ . The set $\\{o\\}_{k=1}^{K}$ is defined by extracting the referenced category names from the text instruction $c_{T}$ . Having the unedited conditioning image $I_0^v$ , the edited image $I_{i+1}^v$ , and the binary mask $m_{\\{o\\}_{k=1}^{K}}$ , we keep only the edits at the pixels of the target object set:", + "bbox": [ + 212, + 617, + 784, + 728 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nI _ {i + 1} ^ {v} = m _ {\\{o \\} _ {k = 1} ^ {K}} \\odot I _ {i + 1} ^ {v} + \\left(1 - m _ {\\{o \\} _ {k = 1} ^ {K}}\\right) \\odot I _ {0} ^ {v}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 741, + 784, + 760 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\odot$ denotes element-wise multiplication of the image pixels. This way, the other pixels within $I_{i + 1}^{v}$ are set to their unedited versions, enabling an object-level editing of training images. While training the 3DGS scene with the edited images, the mask $m_{\\{o\\}_{k = 1}^K}$ is also applied for $\\mathcal{L}_1$ and SSIM losses, ensuring gradient propagation only to the target objects.", + "bbox": [ + 212, + 763, + 784, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "B.M. Öcal et al.", + "bbox": [ + 271, + 113, + 380, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 143, + 375, + 162 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Datasets. To create a set of supporting exemplars for the layout generation, we employ the 3D-FRONT dataset [18], a large-scale indoor scene dataset with 18,797 rooms furnished with 3D textured objects from 3D-FUTURE [19]. Following the same pre-processing steps as [33] and excluding non-rectangular floor plans from the test set similar to [15], we obtained 3397 train, 453 validation, 423 test bedroom and 690 train, 98 validation, 53 test living room scenes. During 3D scene assembling, 3D models are retrieved from the 3D-FUTURE dataset.", + "bbox": [ + 212, + 176, + 787, + 282 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Implementation Details. We use ChatCompletions API $^3$ , which enables a series of dialogue exchanges between the user and the LLM, to query GPT-4 [1]. BlenderProc $^4$ is employed to render 250 images per assembled scene with $512 \\times 512$ resolution. After relocating the center of each scene to the origin of the coordinate system, we sample camera positions on the upper hemisphere at a distance of $1.5r$ from the center where $r$ is the diagonal length of the room, with an elevation angle of 35 degrees. We use the Splatifto model from NeRFstudio [46] for 3DGS training, by optimizing each scene for a maximum of 20k iterations. For editing the 3DGS training images, we use InstructPix2Pix [4] as the diffusion model, with classifier guidance scales $s_T = [3.5, 12.5]$ for the text and $s_I = 1.5$ for the image. The dataset updates are performed every 2.5k iterations following Instruct-GS2GS [48]. The InstructGS2GS training takes approximately 15-20 minutes per scene on a single A6000, most of it spent on diffusion model sampling. The rendering speed is 101.21 fps.", + "bbox": [ + 212, + 303, + 789, + 517 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 3D Layout Generation", + "text_level": 1, + "bbox": [ + 215, + 537, + 447, + 553 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Evaluation Setup and Baselines. In this section we evaluate the first stage of our approach, therefore we only focus on assessing the generated layouts. To our knowledge, SceneTeller is the first approach that conducts conditional layout generation guided by text prompts to control the position and orientation of every object, targeting a user-friendly framework for novices. Hence, replicating an interaction mode that is natural for humans is pivotal for our framework. Based on this, our layout generation performance is compared with the performance of humans solving the same task. In particular, each participant of our user study is presented with a series of textual descriptions, each accompanied by an empty layout according to the room dimensions. For each text prompt, participants are requested to draw bounding boxes and label them with the object names. Each text prompt is given to at least 2, at most 3 participants. This study is conducted with 12 participants and 40 text prompts. To generate text prompts, 40 layouts are selected from the test set and their text prompts are generated by providing the initially rule-based formed sentences to GPT for paraphrasing, as described in Sec. 3.1.", + "bbox": [ + 212, + 563, + 787, + 805 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "SceneTeller: Language-to-3D Scene Generation", + "bbox": [ + 419, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_footnote", + "text": "3 See https://platform.openai.com/docs/api-reference/chat", + "bbox": [ + 217, + 809, + 614, + 825 + ], + "page_idx": 8 + }, + { + "type": "page_footnote", + "text": "See https://github.com/DLR-RM/BlenderProc", + "bbox": [ + 220, + 825, + 549, + 839 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Metrics. Based on the predicted number of bounding boxes in each category, we report recall, precision and accuracy [15]. An accuracy score of 100 indicates that the number of bounding boxes and their corresponding categories precisely match the ground-truth annotations. Otherwise, a score of 0 is assigned to the scene. All values are in percentage (\\%). We additionally report mean Intersection over Union (mIoU) between the predicted and ground-truth bounding boxes and out-of-bound rates, indicating the percentage of scenes where furniture extends beyond the floor plan boundary. We refer to the supplementary material for detailed metric descriptions.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Results. Tab. 1 presents the quantitative evaluation results of 3D layout generation. Our in-context-learning-based 3D layout generation module performs better than human participants in terms of mIoU, showcasing the effectiveness of LLMs' visual reasoning capabilities. In-context exemplars further enhance this performance by aiding in reasoning about the current data distribution. However, human participants possess their own understanding of the physical world, leading to variations in their notions of object sizes and prompt reasoning, resulting in slightly poorer performance in this assessment. Both our layout generation module and human participants performed excellently in terms of predicting the number of bounding boxes per category. Humans sometimes forget or miss objects mentioned in the prompt, which explains their slightly imperfect recall score. The ground-truth layouts from the dataset involve objects slightly exceeding the floor plan boundary, leading to $15.15\\%$ OOB rate. Human participants are able to place the objects within the layout more accurately.", + "bbox": [ + 212, + 306, + 787, + 518 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/ab0fd08d5f97041daedfd5d1daca3ffec9222702fe877f46fac5ea01322c021a.jpg", + "table_caption": [ + "Table 1: Quantitative evaluation of 3D layout generation. The proposed in-context learning based 3D layout generation module performs better than human participants in terms of predicting object sizes, enhanced by in-context exemplars aiding in reasoning about the current data distribution. All values are in percentage (\\%)." + ], + "table_footnote": [], + "table_body": "
MethodOOB ↓rec.↑prec.↑acc.↑mIoU ↑
Ours56.78100.00100.00100.0044.60
Human12.0998.79100.0095.0028.70
GT15.15100.00100.00100.00100.00
", + "bbox": [ + 344, + 612, + 661, + 686 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 3D Scene Generation", + "text_level": 1, + "bbox": [ + 215, + 738, + 437, + 752 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Baselines. Our approach is evaluated and compared with three state-of-the-art 3D scene generation methods. Set-the-scene [11] employs compositional NeRFs that are initialized based on user-defined object proxies to have more control over individual objects. To compare with [11], layouts generated by our method are provided as input to their pipeline. GSGEN [9] is a 3DGS-based approach,", + "bbox": [ + 212, + 763, + 787, + 842 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "B.M. Öcal et al.", + "bbox": [ + 271, + 113, + 380, + 127 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "following a two-stage pipeline for geometry optimization and appearance refinement. LucidDreamer [10] employs render-refine-repeat paradigm by iteratively warping and inpainting the previously generated image to create an initial point cloud, which is then used for 3DGS scene training.", + "bbox": [ + 214, + 146, + 787, + 205 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "a t a o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o .", + "bbox": [ + 222, + 234, + 779, + 265 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/b9cd48f6a5141197abf01aa23d3a76c3976289e99000f006b0869a0e967393be.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 223, + 267, + 303, + 328 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/9d0c2dd39d0b81c0c2b2daebe98265ee63c4fa095e9b32532f403be633440f50.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 320, + 267, + 401, + 328 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/e02baf7723bf937babf56b54897fdf4786d174d2548fa5781e7438366857e00d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 423, + 267, + 542, + 328 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/58e6f5f6c1e8785094486142bd3c64da1e2509215419e29e01e0a9e61db77e42.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 570, + 267, + 653, + 328 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/05f4520687f5ef3a8235ed2a3aaf0236035091f74f4f2615b4b2b171f282c844.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 681, + 267, + 764, + 328 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "\"The dining table is situated in the middle of the room, but slightly leaning towards the lower left wall, with its orientation reversed and perpendicular. One dining chair is positioned near the lower left corner, with its orientation being perpendicular. Another dining chair is placed in the middle of the left wall, a bit nearer to the lower wall, and is also perpendicular. Yet another dining chair is set up at the midpoint of the left wall, but it's slightly more towards the upper wall and is perpendicular. One dining chair is near the lower right corner, with its orientation reversed and perpendicular. Another dining chair can be found at the center of the right wall, a bit closer to the lower wall, and it is perpendicular but facing the opposite direction. Another dining chair is at the center of the right wall, but slightly more towards the upper wall and its orientation is perpendicular but reversed. One more dining chair is in the middle of the upper wall, closer to the left wall, and is oriented backwards. Finally, a dining chair is at the center of the lower wall, a bit closer to the left wall, and it is oriented without any rotation. Make it marble.\"", + "bbox": [ + 225, + 330, + 774, + 381 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/ef04450c06340b11e91bce98c9b579e3ab4f864fdf71d49f6290099d33b27bec.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 223, + 386, + 303, + 446 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/c3702bd7776a7f92ee6b846a80d566ccead06a59b44dee2cbf494d1e261bcc5a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 320, + 386, + 400, + 446 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/af387969c8cf0fadb7fe6d4d7e95acaee26d77881a2a7ee371bd4d0f770bd79e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 423, + 386, + 547, + 446 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/116e26e5cb13ab5e30d577d57d78bb755351d6d1394804aa0592c68e425942f2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 560, + 387, + 668, + 446 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/2739be99be80a1abbf9ed44d7bbc1f87d93a69f75b9cbeabd1632c379a21473c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 681, + 388, + 767, + 445 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The coffe table is sitd in the middle of the upper wall, albeit a tad nearer to the right wall, and is aligned perpendicularly facing the opposite direction. Another coffe table can be found in the middle of the right wall, a bit nearer to the lower wall, and is facing towards the rear. Nestled in the upper right corner is another coffe table, positioned without any rotation. In the upper left corner, a tv stand is positioned perpendicularly. A multi seat sofa is positioned near the upper right corner, facing perpendicularly but in the reverse direction. Close to the lower left corner, a dining table is situated without any rotation. A dinng chair can be found in the center of the lower wall, slightly nearer to the left wall, and is facing in its original direction. Another dinng chair is positioned near the lower left corner, also without any rotation. A third dinng chair is situated in the room's center, closer to the lower left wall, and oriented towards the rear. A fourth dinng chair is positioned in the middle of the left wall, slightly closer to the lower wall, positioned towards the rear. Lastly, a console table can be found in the lower left corner, placed without any rotation. Make it pink themed.\"", + "bbox": [ + 225, + 449, + 776, + 502 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/6e046d146f2c454a993350cdc2a17ceab9fd6d15af5809dfd2a5d7735c209aad.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 223, + 505, + 303, + 566 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/2e6a1815d743a8054c7934410e242013ea5396f686eaaff61a24b4c332d6fed9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 320, + 505, + 400, + 566 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/50c77c19e1fea3fdfe1bb02c96b14a416ec4d2e4ca561bb7cd745ecb96775de8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 423, + 505, + 547, + 566 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/6dbf722a40460f2451c1349b8d2eb6e4c4f1f5294aa73490309ba7674e860a72.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 555, + 508, + 669, + 569 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/004c5817ad26d584c0a3b933b66ce03264f5e7c00538d3ab224c763d53a589da.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 681, + 508, + 772, + 561 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "At the heart of the room, a double bed is positioned a tad nearer to the right wall, its orientation is perpendicular but in reverse. Adjacent to the top-right corner, a nightstand is placed, its orientation is also perpendicular but in reverse. A wardrobe can be found centrally aligned with the bottom wall, subtly closer to the right wall, with no rotation in its orientation. Make it Starry Night Van Gogh painting style.\"", + "bbox": [ + 225, + 570, + 776, + 592 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/5f26f22cc9e384574c1e4245a3895ea4bc62bf6972676f5f5ebf98209ec70fff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 223, + 594, + 303, + 656 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/43678837dad34e5db21bc61d0b262e05caac0452fa28cfc63903ce38b7f30282.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 320, + 594, + 401, + 657 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/d4a797edc559cd12f4e55bb9e1d797dee62a3ffff29dfa010341c075a90e14c0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 423, + 594, + 544, + 656 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/2ff0b3f0469ae4855ab16c84fba910428c6d002b72588f5934ee731a5d2555b7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 570, + 592, + 653, + 654 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/6938169294f0733a5742f89ded17979078fd393da12f0c8aa4e1e4e3a659a0c3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 681, + 592, + 763, + 654 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In the middle of the room, a double bed is placed, slightly more towards the bottom right wall, with its orientation being perpendicular and facing the reverse direction. A wardrobe is situated at the center of the top wall, a bit closer to the right wall, with its orientation facing backwards. Near the top right corner of the room, a table is situated, oriented perpendicularly but in the reverse direction. At the bottom left corner, a cabinet is positioned, perpendicular to the walls. Another table is situated near the bottom right corner, oriented perpendicularly but facing the opposite direction. Turn the double bed into golden.\"", + "bbox": [ + 220, + 657, + 777, + 688 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/812c1438c05f42d0ad6ef9153d9033e66cc7255a764ac85539cdda8666a19b1a.jpg", + "image_caption": [ + "GSGEN" + ], + "image_footnote": [], + "bbox": [ + 223, + 690, + 302, + 752 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/25193ae74de29f3d1da45e292352c7ef1d77361dfa4ae88caa2eba43d39bb9e9.jpg", + "image_caption": [ + "Set-the-scene" + ], + "image_footnote": [], + "bbox": [ + 323, + 691, + 403, + 753 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/576bb80a86cbf15a1999f877aa76aada3915f2cc2a0955e86cf85d51238bd8fe.jpg", + "image_caption": [ + "LucidDreamer" + ], + "image_footnote": [], + "bbox": [ + 424, + 691, + 542, + 752 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/9a97b20e054b0270228105f90ac4d371f5c734e2e58b09a39e22c1b6a998688b.jpg", + "image_caption": [ + "Fig. 3: Qualitative comparison with state-of-the-art text-to-3D scene generation methods. The top/upper walls are marked with arrows for your reference, if this information is available within the generations. SceneTeller is able to generate high-quality scenes, with superior geometric fidelity and 3D consistency." + ], + "image_footnote": [], + "bbox": [ + 571, + 691, + 647, + 752 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/e6e4b7a01a3ef283b5a2b88b38f5803ac6c27d6b38f1e7f0c2bbefc108ffb3c8.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 651, + 705, + 764, + 750 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "SceneTeller: Language-to-3D Scene Generation", + "bbox": [ + 418, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "User Study. We conduct a user study with 30 participants to assess the performance of our method in generating high-quality text-consistent realistic 3D scenes. Human evaluators are presented with 10 scenes, each accompanied by a textual prompt describing its layout and style. For each scene, they are requested to rate it on a scale from 1 to 5 (5 being the best) based on four criteria: (a) Realism, (b) Text Alignment, (c) Geometric Fidelity and (d) Compositional Plausibility.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We present the results of the study in Tab. 2. Our method consistently outperforms the existing state-of-the-art 3D scene generation methods across all the four criteria by a significant margin, showcasing the merits of SceneTeller in generating realistic high-quality 3D scenes solely from text prompts. For additional details of the study and extensive analysis of the results, we refer to the supplementary material.", + "bbox": [ + 212, + 252, + 787, + 343 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/f570d5131c9f6867bccca3a4ee05f1d30ac6fb499d749e5da7d06dd74aaaf27f.jpg", + "table_caption": [ + "Table 2: User study results. Study conducted with 30 participants to assess the performance of state-of-the-art methods in generating 3D scenes based on four criteria with a scale from 1 to 5 (5 being the best and vice versa). Our method consistently outperforms the existing state-of-the-art 3D scene generation methods across all the four criteria by a significant margin, showcasing the merits of SceneTeller in generating realistic high-quality 3D scenes solely from text prompts." + ], + "table_footnote": [], + "table_body": "
MethodInput TypeUser Study
RealismText AlignmentGeometric FidelityCompositional Plausibility
GSGEN [9]T1.3131.3751.4391.392
LucidDreamer [10]T3.5432.7513.0503.116
Set-the-scene [11]L + T1.9631.9002.1122.181
OursT4.1544.1894.4794.323
", + "bbox": [ + 238, + 465, + 766, + 584 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Quantitative Evaluation of Stylization. To evaluate the consistency of the generated scene style with textual descriptions, we utilize the CLIP Score [21] as the metric. Given that the CLIP score cannot reason about the parts of the prompts describing the layout, we only provide style information to CLIP in the format: 'A Barbie style bedroom' or 'A bedroom with Barbie style nightstands' for all methods. As shown in Tab. 3, our method consistently generates scenes adhering to the given style description.", + "bbox": [ + 212, + 635, + 787, + 742 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Qualitative Evaluation. Visual comparison of SceneTeller with state-of-the-art text-to-3D methods are provided in Fig. 3. SceneTeller yields high-quality 3D scene generations, surpassing the baselines in terms of text-alignment and 3D consistency by predicting an initial 3D layout conditioned on the textual descriptions. Unlike some of the state-of-the-art baselines producing unrealistic", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "B.M. Öcal et al.", + "bbox": [ + 271, + 113, + 380, + 127 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/c4bef7d7d9fa8c9c1a32cc3e178a76f0435561c11fcd865ab0d16fa9228b3b13.jpg", + "table_caption": [ + "Table 3: Quantitative comparison of stylization. We report the CLIP Score to evaluate the consistency of the generated scene style with textual descriptions. Our method consistently generates scenes adhering to the given style description." + ], + "table_footnote": [], + "table_body": "
GSGEN [9]LucidDreamer [10]Set-the-scene [11]Ours
CLIP Score ↑26.47830.71527.18830.923
", + "bbox": [ + 261, + 198, + 741, + 241 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "appearance (e.g. comic or cartoon-like characteristics), SceneTeller is able to produce more realistic object shapes and appearances.", + "bbox": [ + 212, + 268, + 785, + 301 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Realism of the Generated Scenes. We evaluate all methods using FID score with a real indoor dataset ScanNet [12] on 10 generated scenes by rendering 50 images from each of them. As shown in Tab. 4, our method consistently generates more realistic scenes.", + "bbox": [ + 212, + 323, + 785, + 383 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/bd60056536bd8594bada752833f7e479b4a66a764c6df398c7e2bb5891578896.jpg", + "table_caption": [ + "Table 4: FID score with the ScanNet dataset [12]." + ], + "table_footnote": [], + "table_body": "
GSGEN [9] LucidDreamer [10] Set-the-scene [11]Ours
FID ↓297.10238.75297.57167.27
", + "bbox": [ + 285, + 435, + 715, + 478 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Applicability to Real-world. While current design tools enable professionals to create floor plans, visualize 3D models, and render photo-realistic images, they are not suitable for novice users unfamiliar with such software tools. By enabling users to interact solely through plain natural language prompts, our approach also empowers novice users to flexibly design their own rooms.", + "bbox": [ + 212, + 532, + 787, + 609 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.3 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 632, + 390, + 648 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effect of design choices for 3D layout generation. To assess the effectiveness of individual components in our design, we conduct an ablation study. Specifically, we examine two aspects: 1) the prompt source for in-context exemplars, and 2) the technique for selecting in-context exemplars. Regarding the prompt source, we present exemplars where the textual description is generated either solely by the rule-based formulation (R) or by a combination of the rule-based formulation and GPT paraphrasing $(\\mathrm{R} + \\mathrm{GPT})$ to introduce linguistic variation as described in Sec. 3.1. For the exemplar selection technique, we experiment with a random and a retrieval-based exemplar selection following [15], $k = 8$ exemplars. Additionally, we experiment with a third set of exemplars, denoted as $pos + neg$ , where 4 exemplars are selected by retrieval, and the other 4 are chosen from negative exemplars involving out-of-bound (OOB)", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "SceneTeller: Language-to-3D Scene Generation", + "bbox": [ + 418, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "or overlapping objects. The negative ones are from the filtered samples during the pre-processing of the dataset, and the LLM is specifically instructed to avoid them. Tab. 5 verifies that providing in-context exemplars with linguistic variation $(\\mathrm{R} + \\mathrm{GPT})$ selected with a retrieval-based approach performs the best, especially in terms of OOB rate by constraining the in-context learning of LLM to similar layouts. Using the $pos + neg$ set also considerably improves the performance in terms of mIoU.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/af12552783cd9cbc1a4581f54f0123519530ca04504a58fa918724353d11212f.jpg", + "table_caption": [ + "Table 5: Effect of design choices for 3D layout generation. We report the 3D layout generation performance w.r.t different design choices such as the prompt source and the exemplar selection. Providing in-context exemplars to LLM with linguistic variation (R+GPT) selected with a retrieval-based approach, performs the best." + ], + "table_footnote": [], + "table_body": "
Prompt sourceExemplar SelectionOOB(%)Numerical ReasoningBbox
random retrieval pos+neg↓rec.↑prec.↑acc.↑
1R✓74.12100.00100.00100.0031.10
2R✓64.87100.00100.00100.0045.50
3R74.10100.00100.00100.0041.90
4R + GPT✓76.1199.1099.1099.1029.90
5R + GPT✓54.65100.00100.00100.0044.20
6R + GPT66.45100.00100.00100.0043.90
", + "bbox": [ + 218, + 342, + 784, + 488 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 537, + 359, + 553 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we presented a novel pipeline for 3D room generation from a textual description. Our approach supports style editing either on the individual object or on the entire scene level. Rooms generated by our method were preferred by human evaluators over the baseline results almost in all cases, confirming its efficacy on this challenging task. Importantly, our approach can be used by novices, thus making a step towards democratizing 3D scene design.", + "bbox": [ + 212, + 565, + 787, + 657 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 215, + 676, + 403, + 694 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We thank members of the Bosch-UvA Delta Lab and anonymous reviewers for helpful discussions and feedback. This project was generously supported by the Bosch Center for Artificial Intelligence.", + "bbox": [ + 212, + 704, + 785, + 751 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 770, + 323, + 787 + ], + "page_idx": 13 + }, + { + "type": "ref_text", + "text": "1. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)", + "bbox": [ + 222, + 797, + 785, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "B.M. Öcal et al.", + "bbox": [ + 271, + 113, + 380, + 127 + ], + "page_idx": 13 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "2. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems (NeurIPS) 35, 23716-23736 (2022)", + "3. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (CVPR). pp. 5855-5864 (2021)", + "4. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18392-18402 (2023)", + "5. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. In: European Conference on Computer Vision (ECCV) (2022)", + "6. Chen, R., Chen, Y., Jiao, N., Jia, K.: Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (October 2023)", + "7. Chen, X., Zhang, Q., Li, X., Chen, Y., Feng, Y., Wang, X., Wang, J.: Hallucinated neural radiance fields in the wild. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12943-12952 (2022)", + "8. Chen, Y., Chen, Z., Zhang, C., Wang, F., Yang, X., Wang, Y., Cai, Z., Yang, L., Liu, H., Lin, G.: Gaussianeditor: Swift and controllable 3d editing with gaussian splatting. arXiv preprint arXiv:2311.14521 (2023)", + "9. Chen, Z., Wang, F., Liu, H.: Text-to-3d using gaussian splatting. arXiv preprint arXiv:2309.16585 (2023)", + "10. Chung, J., Lee, S., Nam, H., Lee, J., Lee, K.M.: Luciddreamer: Domain-free generation of 3d gaussian splatting scenes. arXiv preprint arXiv:2311.13384 (2023)", + "11. Cohen-Bar, D., Richardson, E., Metzer, G., Giryes, R., Cohen-Or, D.: Set-the-scene: Global-local training for generating controllable nerf scenes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops. pp. 2920–2929 (October 2023)", + "12. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proc. Computer Vision and Pattern Recognition (CVPR), IEEE (2017)", + "13. Fang, J., Wang, J., Zhang, X., Xie, L., Tian, Q.: Gaussianeditor: Editing 3d gaussians delicately with text instructions. arXiv preprint arXiv:2311.16037 (2023)", + "14. Fang, J., Wang, J., Zhang, X., Xie, L., Tian, Q.: Gaussianeditor: Editing 3d gaussians delicately with text instructions. In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024)", + "15. Feng, W., Zhu, W., Fu, T.j., Jampani, V., Akula, A., He, X., Basu, S., Wang, X.E., Wang, W.Y.: Layoutgpt: Compositional visual planning and generation with large language models. Advances in Neural Information Processing Systems (NeurIPS) 36 (2024)", + "16. Fridman, R., Abecasis, A., Kasten, Y., Dekel, T.: Scenescape: Text-driven consistent scene generation. Advances in Neural Information Processing Systems (NeurIPS) 36 (2024)", + "17. Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5501-5510 (2022)" + ], + "bbox": [ + 218, + 146, + 787, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "SceneTeller: Language-to-3D Scene Generation", + "bbox": [ + 418, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Fu, H., Cai, B., Gao, L., Zhang, L.X., Wang, J., Li, C., Zeng, Q., Sun, C., Jia, R., Zhao, B., et al.: 3d-front: 3d furnished rooms with layouts and semantics. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10933-10942 (2021)", + "19. Fu, H., Jia, R., Gao, L., Gong, M., Zhao, B., Maybank, S., Tao, D.: 3d-future: 3d furniture shape with texture. International Journal of Computer Vision 129, 3313-3337 (2021)", + "20. Haque, A., Tancik, M., Efros, A., Holynski, A., Kanazawa, A.: Instruct-nerf2nerf: Editing 3d scenes with instructions. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (CVPR) (2023)", + "21. Hessel, J., Holtzman, A., Forbes, M., Le Bras, R., Choi, Y.: CLIPScore: A reference-free evaluation metric for image captioning. In: Moens, M.F., Huang, X., Specia, L., Yih, S.W.t. (eds.) Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. pp. 7514-7528. Association for Computational Linguistics, Online and Punta Cana, Dominican Republic (Nov 2021). https://doi.org/10.18653/v1/2021.emnlp-main.595, https://aclanthology.org/2021.emnlp-main.595", + "22. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020)", + "23. Hollein, L., Cao, A., Owens, A., Johnson, J., Nießner, M.: Text2room: Extracting textured 3d meshes from 2d text-to-image models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 7909-7920 (October 2023)", + "24. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (July 2023), https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/", + "25. Li, Q., Li, F., Guo, J., Guo, Y.: Uhdnerf: Ultra-high-definition neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (CVPR). pp. 23097-23108 (2023)", + "26. Li, X., Wang, H., Tseng, K.K.: Gaussiandiffusion: 3d gaussian splatting for denoising diffusion probabilistic models with structured noise. arXiv preprint arXiv:2311.11221 (2023)", + "27. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3d: High-resolution text-to-3d content creation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 300-309 (2023)", + "28. Lin, Y., Bai, H., Li, S., Lu, H., Lin, X., Xiong, H., Wang, L.: Componerf: Text-guided multi-object compositional nerf with editable 3d scene layout. arXiv preprint arXiv:2303.13843 (2023)", + "29. Lorraine, J., Xie, K., Zeng, X., Lin, C.H., Takikawa, T., Sharp, N., Lin, T.Y., Liu, M.Y., Fidler, S., Lucas, J.: Att3d: Amortized text-to-3d object synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 17946-17956 (October 2023)", + "30. Metzer, G., Richardson, E., Patashnik, O., Giryes, R., Cohen-Or, D.: Latent-nerf for shape-guided generation of 3d shapes and textures. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12663-12673 (2023)", + "31. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: Proceedings of the European Conference on Computer Vision (ECCV) (2020)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "B.M. Öcal et al.", + "bbox": [ + 271, + 113, + 380, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "32. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022)", + "33. Paschalidou, D., Kar, A., Shugrina, M., Kreis, K., Geiger, A., Fidler, S.: Atiss: Autoregressive transformers for indoor scene synthesis. Advances in Neural Information Processing Systems 34, 12013-12026 (2021)", + "34. Phung, Q., Ge, S., Huang, J.B.: Grounded text-to-image synthesis with attention refocusing. arXiv preprint arXiv:2306.05427 (2023)", + "35. Po, R., Wetzstein, G.: Compositional 3d scene generation using locally conditioned diffusion. ArXiv abs/2303.12218 (2023), https://api-semanticscholar.org/CorpusID:257663283", + "36. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. In: The Eleventh International Conference on Learning Representations (ICLR) (2023), https://openreview.net/forum?id=FjNys5c7VvY", + "37. Qian, G., Mai, J., Hamdi, A., Ren, J., Siarohin, A., Li, B., Lee, H.Y., Skorokhodov, I., Wonka, P., Tulyakov, S., Ghanem, B.: Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. In: The Twelfth International Conference on Learning Representations (ICLR) (2024), https://openreview.net/forum?id=0jHkUDyEO9", + "38. Raj, A., Kaza, S., Poole, B., Niemeyer, M., Mildenhall, B., Ruiz, N., Zada, S., Aberman, K., Rubenstein, M., Barron, J., Li, Y., Jampani, V.: Dreambooth3d: Subject-driven text-to-3d generation. The International Conference on Computer Vision (ICCV) (2023)", + "39. Reiser, C., Szeliski, R., Verbin, D., Srinivasan, P., Mildenhall, B., Geiger, A., Barron, J., Hedman, P.: Merf: Memory-efficient radiance fields for real-time view synthesis in unbounded scenes. ACM Transactions on Graphics (TOG) 42(4), 1-12 (2023)", + "40. Ren, J., He, C., Liu, L., Chen, J., Wang, Y., Song, Y., Li, J., Xue, T., Hu, S., Chen, T., Zheng, K., Xiang, J., Bo, L.: Make-a-character: High quality text-to-3d character generation within minutes. arXiv preprint arXiv:2312.15430 (2023)", + "41. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)", + "42. Sanh, V., Webson, A., Raffel, C., Bach, S., Sutawika, L., Ålyafeai, Z., Chaffin, Å., Stiegler, A., Raja, A., Dey, M., Bari, M.S., Xu, C., Thakker, U., Sharma, S.S., Szczechla, E., Kim, T., Chhablani, G., Nayak, N., Datta, D., Chang, J., Jiang, M.T.J., Wang, H., Manica, M., Shen, S., Yong, Z.X., Pandey, H., Bawden, R., Wang, T., Neeraj, T., Rozen, J., Sharma, A., Santilli, A., Fevry, T., Fries, J.A., Teehan, R., Scao, T.L., Biderman, S., Gao, L., Wolf, T., Rush, A.M.: Multitask prompted training enables zero-shot task generalization. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=9Vrb9DOWI4", + "43. Sun, C., Sun, M., Chen, H.T.: Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5459-5469 (2022)", + "44. Sun, C., Han, J., Deng, W., Wang, X., Qin, Z., Gould, S.: 3d-gpt: Procedural 3d modeling with large language models. arXiv preprint arXiv:2310.12945 (2023)", + "45. Tancik, M., Casser, V., Yan, X., Pradhan, S., Mildenhall, B., Srinivasan, P.P., Barron, J.T., Kretzschmar, H.: Block-nerf: Scalable large scene neural view synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8248-8258 (2022)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "SceneTeller: Language-to-3D Scene Generation", + "bbox": [ + 418, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "46. Tancik, M., Weber, E., Ng, E., Li, R., Yi, B., Kerr, J., Wang, T., Kristoffersen, A., Austin, J., Salahi, K., Ahuja, A., McAllister, D., Kanazawa, A.: Nerfstudio: A modular framework for neural radiance field development. In: ACM SIGGRAPH 2023 Conference Proceedings. SIGGRAPH '23 (2023)", + "47. Tang, J., Ren, J., Zhou, H., Liu, Z., Zeng, G.: Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653 (2023)", + "48. Vachha, C., Haque, A.: Instruct-gs2gs: Editing 3d gaussian splats with instructions (2024), https://instruct-gs2gs.github.io/", + "49. Wang, H., Du, X., Li, J., Yeh, R.A., Shakhnarovich, G.: Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 12619-12629 (2023)", + "50. Wang, Z., Lu, C., Wang, Y., Bao, F., Li, C., Su, H., Zhu, J.: Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. In: Advances in Neural Information Processing Systems (NeurIPS) (2023)", + "51. Wei, J., Bosma, M., Zhao, V., Guu, K., Yu, A.W., Lester, B., Du, N., Dai, A.M., Le, Q.V.: Finetuned language models are zero-shot learners. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=gEZrGcozdqR", + "52. Xu, Y., Chai, M., Shi, Z., Peng, S., Skorokhodov, I., Siarohin, A., Yang, C., Shen, Y., Lee, H.Y., Zhou, B., et al.: Discoscope: Spatially disentangled generative radiance fields for controllable 3d-aware scene synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4402-4412 (2023)", + "53. Yang, Z., Gan, Z., Wang, J., Hu, X., Lu, Y., Liu, Z., Wang, L.: An empirical study of gpt-3 for few-shot knowledge-based vqa. In: Proceedings of the AAAI Conference on Artificial Intelligence (AAAI). vol. 36, pp. 3081-3089 (2022)", + "54. Yi, T., Fang, J., Wu, G., Xie, L., Zhang, X., Liu, W., Tian, Q., Wang, X.: Gaussian dreamer: Fast generation from text to 3d gaussian splatting with point cloud priors. arXiv preprint arXiv:2310.08529 (2023)", + "55. Yu, A., Li, R., Tancik, M., Li, H., Ng, R., Kanazawa, A.: Plenoptrees for real-time rendering of neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (CVPR). pp. 5752-5761 (2021)", + "56. Zhou, X., Ran, X., Xiong, Y., He, J., Lin, Z., Wang, Y., Sun, D., Yang, M.H.: Gala3d: Towards text-to-3d complex scene generation via layout-guided generative gaussian splatting. arXiv preprint arXiv:2402.07207 (2024)", + "57. Zhu, J., Zhuang, P., Koyejo, S.: HIFA: High-fidelity text-to-3d generation with advanced diffusion guidance. In: The Twelfth International Conference on Learning Representations (ICLR) (2024), https://openreview.net/forum?id=IZMPWmcS3H" + ], + "bbox": [ + 215, + 146, + 785, + 686 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "B.M. Öcal et al.", + "bbox": [ + 271, + 113, + 380, + 127 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/SceneTeller_ Language-to-3D Scene Generation/dec651d5-2a06-4ad2-af7e-aeaf7a7dadd2_model.json b/2024/SceneTeller_ Language-to-3D Scene Generation/dec651d5-2a06-4ad2-af7e-aeaf7a7dadd2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f180a5c735b1c99f8ff14da26d3c9eb622de55df --- /dev/null +++ b/2024/SceneTeller_ Language-to-3D Scene Generation/dec651d5-2a06-4ad2-af7e-aeaf7a7dadd2_model.json @@ -0,0 +1,2601 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.234, + 0.142, + 0.77, + 0.163 + ], + "angle": 0, + "content": "SceneTeller: Language-to-3D Scene Generation" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.192, + 0.784, + 0.21 + ], + "angle": 0, + "content": "Bashak Melis Öcal1*, Maxim Tatarchenko2, Sezer Karaoglu1, and Theo Gevers1" + }, + { + "type": "text", + "bbox": [ + 0.331, + 0.219, + 0.672, + 0.234 + ], + "angle": 0, + "content": "1 UvA-Bosch Delta Lab, University of Amsterdam" + }, + { + "type": "text", + "bbox": [ + 0.351, + 0.234, + 0.65, + 0.248 + ], + "angle": 0, + "content": "\\(^{2}\\) Bosch Center for AI, Robert Bosch GmbH" + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.302, + 0.767, + 0.68 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.703, + 0.788, + 0.775 + ], + "angle": 0, + "content": "Fig. 1: With solely a textual prompt in natural language describing the desired spatial positions and orientations of objects within the scene, SceneTeller is able to generate realistic and high-quality 3D spaces. Additionally, by facilitating modifications to the style of the entire scene or individual objects within the scene through edit instructions, SceneTeller offers a practical and flexible framework for designing personalized rooms." + }, + { + "type": "page_footnote", + "bbox": [ + 0.22, + 0.826, + 0.504, + 0.84 + ], + "angle": 0, + "content": "\\* Correspondence to \\(< \\) b.m.ocal@uva.nl>" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.114, + 0.381, + 0.128 + ], + "angle": 0, + "content": "B.M. Öcal et al." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.148, + 0.744, + 0.357 + ], + "angle": 0, + "content": "Abstract. Designing high-quality indoor 3D scenes is important in many practical applications, such as room planning or game development. Conventionally, this has been a time-consuming process which requires both artistic skill and familiarity with professional software, making it hardly accessible for layman users. However, recent advances in generative AI have established solid foundation for democratizing 3D design. In this paper, we propose a pioneering approach for text-based 3D room design. Given a prompt in natural language describing the object placement in the room, our method produces a high-quality 3D scene corresponding to it. With an additional text prompt the users can change the appearance of the entire scene or of individual objects in it. Built using in-context learning, CAD model retrieval and 3D-Gaussian-Splatting-based stylization, our turnkey pipeline produces state-of-the-art 3D scenes, while being easy to use even for novices. Our project page is available at https://sceneteller.github.io/." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.37, + 0.741, + 0.398 + ], + "angle": 0, + "content": "Keywords: Text-to-3D scene generation \\(\\cdot\\) Language-guided generation \\(\\cdot\\) 3D Gaussian splatting scene stylization" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.427, + 0.377, + 0.442 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.788, + 0.583 + ], + "angle": 0, + "content": "3D scene design has numerous applications in domains like architecture or game development. Designing high-quality 3D scenes, however, is a laborious and time-consuming process, challenging even for highly-skilled domain experts utilizing professional tools. This complexity often leaves novice users unable to customize their own spaces. Recent advancements in automatic text-to-image synthesis [22,41] have led to massive progress in text-to-3D creation [36-38], enabling the generation of 3D content only from text input. Such works create foundation for democratizing 3D design, supporting its accessibility for non-experts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.788, + 0.718 + ], + "angle": 0, + "content": "A pioneering work by Poole et al. [36] introduced object-centric text-to-3D generation based on 2D diffusion models coupled with NeRFs [31]. Several subsequent works proposed ways to improve the geometric accuracy, multi-view consistency, and appearance quality [6,27,30,50]. Despite their success in object-level generation, hallucinating complex large-scale scenes with multiple objects is still remaining a challenging problem. Some recent 3D scene generation methods [11,35,52] require a 3D layout as input, limiting their usability for novices. Others [16,23] produce results that violate global object arrangements and are not multi-view consistent." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Aiming to address both these shortcomings - insufficient user friendliness and unsatisfactory global consistency - we introduce SceneTeller, a framework for text-based 3D scene generation. 1) We start with a premise that a 3D scene generation framework should offer control over the placement and style of individual objects, allowing novice users to design their customized spaces through simple interactions with the system. Our approach achieves this by relying on text-based user interaction for scene and appearance specification. 2) In order to generate compositionally plausible 3D scenes, it is important to incorporate" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.419, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SceneTeller: Language-to-3D Scene Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "some degree of global information guidance. We implement this by conditioning the method on global layout descriptions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.787, + 0.328 + ], + "angle": 0, + "content": "To this end, we split the task of generating high-quality 3D scenes into three steps: 3D layout generation, object population and style editing. 1) When presented with a prompt describing the desired spatial positions and orientations of objects, SceneTeller generates an initial 3D layout of the scene by exploiting the Large Language Models (LLMs) through in-context learning, to ensure global consistency. 2) The generated layout is then used to assemble a 3D scene and populate it with CAD models from a database. 3) To facilitate scene editing, the assembled scene is fitted with a 3D Gaussian Splitting [24] representation. This representation is then used to stylize the scene according to the user-provided edit prompt targeting the entire scene or individual objects within it." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.329, + 0.787, + 0.389 + ], + "angle": 0, + "content": "Extensive quantitative and qualitative evaluations show that SceneTeller performs significantly better than state-of-the-art methods, offering superior geometric fidelity and compositional plausibility. In brief, our contributions to the existing body of work can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.396, + 0.784, + 0.455 + ], + "angle": 0, + "content": "- We introduce a turnkey pipeline for high-quality 3D scene generation yielding complex, realistic and globally consistent 3D scenes, given only scene descriptions in natural language. Our pipeline supports object- and scene-level appearance editing." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.457, + 0.784, + 0.5 + ], + "angle": 0, + "content": "- We propose a novel LLM-based module for 3D layout generation trained through in-context learning, which provides control over individual object placement." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.501, + 0.785, + 0.545 + ], + "angle": 0, + "content": "- We experimentally demonstrate that SceneTeller generates 3D scenes of higher quality and better consistency than state-of-the-art methods, while boasting unprecedented user friendliness and flexibility." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.396, + 0.785, + 0.545 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.566, + 0.388, + 0.582 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.596, + 0.478, + 0.61 + ], + "angle": 0, + "content": "2.1 3D Scene Representations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.617, + 0.787, + 0.753 + ], + "angle": 0, + "content": "Rendering 3D scenes using implicit neural networks has seen remarkable success, notably with NeRF's [31] photo-realistic image synthesis. Extensions of NeRF improve reconstruction quality [3, 25], training and rendering speed [5, 17, 32, 39, 43, 55], and handling large-scale scenes [7, 45]. Despite advancements, NeRF-based methods still struggle with slow rendering speeds. Recently, 3D Gaussian splatting (3DGS) was proposed for real-time, high-quality rendering using anisotropic 3D Gaussians and fast tile-based rasterization. Our framework employs 3DGS for scene representation and allows edits on these Gaussians, generating high-quality, 3D consistent scenes." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.773, + 0.452, + 0.787 + ], + "angle": 0, + "content": "2.2 Text-to-3D Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Object-centric Generation. 3D generation from textual prompts has recently grown in popularity. DreamFusion [36] introduces Score Distillation Sampling (SDS) to optimize NeRF representations by leveraging an off-the-shelf" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.114, + 0.381, + 0.128 + ], + "angle": 0, + "content": "B.M. Öcal et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.313 + ], + "angle": 0, + "content": "2D diffusion model, paving the way towards text-driven 3D content creation [29,37,38,57]. Following [36], Magic3D [27] employs a coarse-to-fine optimization scheme, exploiting low-resolution diffusion priors, and then enhancing the texture. To achieve high-fidelity mesh generation, Fantasia3D [6] disentangles the appearance from geometry. Latent-NeRF [30] optimizes a NeRF in the latent space of StableDiffusion [41], and additionally proposes incorporating 3D priors to generate a target geometry. To address the over-smoothing, over-saturation and lack of diversity problems within the SDS, ProfilicDreamer [50] proposes a principled particle-based variational framework, namely Variational Score Distillation (VSD). Following the same motivation, [49] presents Perturb-and-Average Scoring to estimate the score for non-noisy images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.314, + 0.79, + 0.465 + ], + "angle": 0, + "content": "With the recent introduction of 3D Gaussian Splatting [24], several works attempt to combine 3D Gs and diffusion models. GaussianDreamer [54] employs a 3D diffusion model to generate the initialized point clouds for 3D Gaussians, and then further optimize the Gaussians using SDS with a 2D diffusion model. [9, 47] follows a two-stage pipeline for geometry optimization and appearance refinement. GaussianDiffusion [26] focuses on the challenge of achieving multiview consistency in 3D generation and proposes combining structured noise with a variational 3DGS technique. Despite their success in generating object-level scenes, both NeRF-based and 3DGS-based methods fall short in hallucinating complex scenes with multiple objects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.483, + 0.79, + 0.74 + ], + "angle": 0, + "content": "Scene Generation. To represent multiple objects within the scenes, existing works [11, 28, 35, 52] rely on compositional NeRFs that are initialized based on user-defined object proxies. However, they are usually restricted to produce only narrow inward-facing views of the rooms. Also, the requirement of having 3D layouts as input is restrictive for final users. Orthogonal to those, SceneScape [16] and Text2Room [23] generate scenes following the render-refine-repeat paradigm by iteratively warping and inpainting the previously generated image according to the specified camera motion. This strategy does not allow object-level control during scene generation, and the results tend to exhibit multi-view inconsistencies. LucidDreamer [10] employs the same paradigm for an initial point cloud reconstruction, and then further optimizes the scene using 3DGS. Recently, several works introduced techniques for 3DGS scene editing, by manipulating the training images using 2D diffusion models [8, 13, 48] and incorporating the edited images into 3DGS scenes for further optimization. Concurrent to our work, [56] initializes a 3DGS scene from the layouts created using LLMs and optimizes it using a diffusion model. However, [56] lacks control over the generated layout in terms of positions and orientations of objects." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.758, + 0.399, + 0.773 + ], + "angle": 0, + "content": "2.3 LLMs for Vision" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "With the inclusion of grounding information in LLMs, they demonstrate enhanced reasoning capabilities for visual concepts. Recently, the visual reasoning abilities of LLMs have been utilized for 3D tasks such as avatar generation [40], instruction-driven 3D modeling [44], 3D editing [14]. Exploiting the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.419, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SceneTeller: Language-to-3D Scene Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "spatial reasoning capabilities of LLMs, LayoutGPT [15] extends beyond its 2D counterparts [34] to generate random 3D layouts of specific room types through in-context learning. However, LayoutGPT lacks support for controlling the generated layout. The 3D layout generation module of SceneTeller is the first of its kind allowing control over individual object placement in accordance with the provided natural language prompts describing the desired positions of objects." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.259, + 0.331, + 0.275 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.314, + 0.778, + 0.52 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.529, + 0.788, + 0.614 + ], + "angle": 0, + "content": "Fig. 2: Overview of our method. Given a textual prompt in natural language delineating the desired spatial positions and orientations of objects within the scene, a 3D scene layout is generated using in-context learning. An initial 3D scene is assembled for the predicted layout, which is then fitted with a 3D Gaussian Splatting representation. This representation is then used to stylize the scene according to the user-provided text prompt, and subsequently render the final images of the scene." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.788, + 0.766 + ], + "angle": 0, + "content": "The pipeline of SceneTeller is divided into three main stages: language-driven 3D layout generation, 3D scene assembling from the layout and 3D scene stylization. In the first stage, a 3D scene layout is generated based on the given text prompt (Sec. 3.1). Then, an initial 3D scene is assembled for the predicted layout (Sec. 3.2). After that, a 3D Gaussian Splatting representation is fitted to the assembled scene. This representation is then used to stylize the scene according to the user-provided text prompt, and subsequently render the final images of the scene (Sec. 3.3. The scheme of the framework is illustrated in Fig. 2" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.787, + 0.591, + 0.803 + ], + "angle": 0, + "content": "3.1 Language-driven 3D Layout Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "The task of this module is to produce a collection of object bounding boxes conditioned on a brief textural description of their spatial positions within the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.114, + 0.381, + 0.128 + ], + "angle": 0, + "content": "B.M. Öcal et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.252 + ], + "angle": 0, + "content": "scene in plain language. More formally, the input is a condition \\(\\mathcal{C}\\), consisting of a textual description \\(\\mathcal{V}\\), along with the scene type and scene dimensions information. The output is a 3D layout represented as a set of 3D bounding boxes \\(\\mathbf{b}_i \\in B = \\{\\mathbf{b}_i\\}_{i=1}^N\\) for each of the \\(N\\) referenced objects. Each 3D bounding box \\(\\mathbf{b}_i = (c_i, \\mathbf{t}_i, \\mathbf{s}_i, o_i)\\) consists of the following components: category name \\(c_i\\), box center \\(\\mathbf{t}_i = (x_i, y_i, z_i) \\in \\mathbb{R}^3\\), box dimensions \\(\\mathbf{s}_i = (w_i, h_i, d_i) \\in \\mathbb{R}^3\\) and orientation \\(o_i \\in \\mathbb{R}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.788, + 0.479 + ], + "angle": 0, + "content": "The textual description \\(\\mathcal{V}\\) is comprised of multiple sentences, each specifying the position and orientation of a single object within a canonical coordinate system. The canonical system is chosen to represent the scene in a 2D perspective, to ensure that the descriptions closely resemble those produced by a human observer. \\(\\mathcal{V}\\) is initially generated using a rule-based approach formulated in three steps as follows: 1) The canonical coordinate system is partitioned into a \\(3\\times 3\\) grid, creating rectangular patches. 2) Each of the predicted objects is assigned to a patch based on the center position \\(\\mathbf{t}_i\\) and orientation \\(o_i\\) of its bounding box within these patches. 3) Per-object sentences are then formulated according to which patch the object is assigned to and object's orientation, e.g. for the object assigned to the top-left patch, with category name \\(c_i\\) and 90 degrees orientation, we generate: \"A \\(c_i\\) is placed at the top-left corner of the room, with a perpendicular orientation.\" Given the relatively limited linguistic variation inherent in rule-based descriptions, initial textual templates are provided to an LLM to serve as a basis for paraphrasing." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.48, + 0.788, + 0.555 + ], + "angle": 0, + "content": "LayoutGPT [15] focused on improving an LLM's interpretation of the spatial knowledge by using a structured format featuring attribute-value pairs. We follow a similar strategy and adopt the standard CSS (Cascading Style Sheets) format to describe \\(\\mathbf{b}_i\\). Bounding box values are mapped to standard CSS format attributes and category name is employed as the selector." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.576, + 0.788, + 0.682 + ], + "angle": 0, + "content": "Prompt Construction. Recent progress in the field has showcased the impressive in-context learning capabilities of LLMs, which entail conditioning these models with both natural language instructions and a select set of task demonstrations. Following this conditioning, the LLM is then tasked with completing further instances based on the provided demonstrations. In line with this, our prompts to LLM include three main parts: task specifications, in-context exemplars and the query condition:" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.689, + 0.788, + 0.796 + ], + "angle": 0, + "content": "- Task specifications: Following prior work [42, 51] on enhancing LLMs instruction following capabilities, a task description is incorporated at the beginning of each prompt, which explains the goal of the task, establishes a standard for the 3D layout format in CSS style and provides unit information for attributes. Additionally, constraints are integrated into the prompts to guide the LLM and minimize errors (e.g., predicting overlapping boxes or placing objects out of bounds), during task completion." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.796, + 0.788, + 0.843 + ], + "angle": 0, + "content": "- In-context learning: Supporting exemplars for the in-context learning are selected by adopting the retrieval-based approach used in [2, 15, 53]. When provided with a set of supporting exemplars \\( S = \\{(\\mathcal{C}_m^s,\\mathbf{b}_m^s)\\}_{m = 1}^M \\) and the" + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.689, + 0.788, + 0.843 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.419, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "SceneTeller: Language-to-3D Scene Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.145, + 0.784, + 0.221 + ], + "angle": 0, + "content": "queried condition \\(\\mathcal{C}_q\\), the function \\(f(\\mathcal{C}_k^s,\\mathcal{C}_q) = \\|rl_k - rl_q\\|^2 + \\|rw_k - rw_q\\|^2\\) is computed between each element of the set and \\(\\mathcal{C}_q\\) following [15], where \\(rl\\) and \\(rw\\) are the length and width of the scenes. Top-\\(k\\) supporting exemplars with the shortest distance to \\(\\mathcal{C}_q\\) are selected for in-context learning, provided to LLM with the same format with \\(\\mathcal{C}_q\\)." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.222, + 0.784, + 0.252 + ], + "angle": 0, + "content": "- Query condition: The inference condition \\(\\mathcal{C}_q\\), for which we want to predict the layout." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.275, + 0.553, + 0.29 + ], + "angle": 0, + "content": "3.2 3D Scene Assembling from Layouts" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.3, + 0.784, + 0.36 + ], + "angle": 0, + "content": "To assemble a scene based on the generated layouts, we rely on object retrieval. More specifically, for each of the bounding boxes, the nearest 3D model of the corresponding predicted category is retrieved from the furniture dataset based on the Euclidean distance of the bounding box dimensions as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.273, + 0.388, + 0.785, + 0.405 + ], + "angle": 0, + "content": "\\[\nd \\left(\\mathbf {s} _ {i}, \\mathbf {s} _ {\\text {c a n d}}\\right) = \\sqrt {\\left(w _ {i} - w _ {\\text {c a n d}}\\right) ^ {2} + \\left(h _ {i} - h _ {\\text {c a n d}}\\right) ^ {2} + \\left(d _ {i} - d _ {\\text {c a n d}}\\right) ^ {2}}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.416, + 0.784, + 0.461 + ], + "angle": 0, + "content": "where \\(\\mathbf{s}_{\\text{cand}}\\) denotes the bounding box dimensions for the candidate object from the dataset. The retrieved 3D model is placed within the scene according to the predicted box center \\(\\mathbf{t}_i\\) and orientation angle \\(o_i\\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.483, + 0.432, + 0.498 + ], + "angle": 0, + "content": "3.3 3D Scene Stylization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.509, + 0.784, + 0.584 + ], + "angle": 0, + "content": "Given a generated 3D scene, the users should be able to modify either the overall appearance of the scene or specific objects within it. Our approach facilitates scene editing by training a 3D Gaussian splatting model representing the generated scene and refining this representation according to the provided edit instruction, offering a flexible framework for 3D scene generation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.606, + 0.784, + 0.667 + ], + "angle": 0, + "content": "3D Gaussian Splating Scene Representation. 3D Gaussian splatting (3DGS) [24] is a recent pioneering work for novel view generation, which represents the underlying scene as a collection of anisotropic 3D Gaussians defined by their center positions \\(\\mu \\in \\mathbb{R}^3\\) and 3D covariance matrices \\(\\pmb{\\Sigma}\\) parameterized as:" + }, + { + "type": "equation", + "bbox": [ + 0.448, + 0.681, + 0.785, + 0.697 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\Sigma} = \\mathbf {R} \\mathbf {S} \\mathbf {S} ^ {\\mathrm {T}} \\mathbf {R} ^ {\\mathrm {T}}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.784, + 0.763 + ], + "angle": 0, + "content": "\\(\\mathbf{R}\\) denotes the rotation matrix and \\(\\mathbf{S}\\) is the scale matrix. Each 3D Gaussian is assigned a color \\(c\\) represented with spherical harmonics (SH) coefficients, to capture the view-dependent appearance. To allow \\(\\alpha\\)-blending of splats, Gaussians are associated with an opacity value \\(\\alpha \\in \\mathbb{R}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.784, + 0.839 + ], + "angle": 0, + "content": "Unlike NeRF [31] which relies on volume rendering, 3DGS enables faster training and rendering through differentiable rasterization. A set of 3D Gaussians is rendered by projecting into the camera's image plane as 2D Gaussians, and assigned to individual image tiles. The color of each pixel \\(\\mathbf{p}\\) on the image plane is then determined as follows:" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.114, + 0.381, + 0.128 + ], + "angle": 0, + "content": "B.M. Öcal et al." + }, + { + "type": "equation", + "bbox": [ + 0.29, + 0.16, + 0.785, + 0.201 + ], + "angle": 0, + "content": "\\[\nC (\\mathbf {p}) = \\sum_ {i \\in \\mathcal {N}} c _ {i} \\sigma_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\sigma_ {j}), \\quad \\sigma_ {i} = \\alpha_ {i} e ^ {- \\frac {1}{2} \\left(\\mathbf {p} - \\mu_ {i}\\right) ^ {T} \\Sigma_ {i} ^ {- 1} \\left(\\mathbf {p} - \\mu_ {i}\\right)} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.207, + 0.784, + 0.267 + ], + "angle": 0, + "content": "where \\(\\mathcal{N}\\) denotes the Gaussians in this tile, \\(\\sigma_{i}\\) represents the influence of the Gaussian on the image pixel and \\(\\mu_{i},\\Sigma_{i},c_{i},\\alpha_{i}\\) are the position, covariance, color and opacity of the \\(i\\)-th Gaussian respectively. For optimization, a weighted combination of \\(\\mathcal{L}_1\\) loss and SSIM loss is employed." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.268, + 0.785, + 0.358 + ], + "angle": 0, + "content": "To initialize the centers of the 3D Gaussians, we utilize the mesh vertices of the retrieved objects and create a scene point cloud. This generated scene point cloud replaces the Structure-from-Motion (SfM) points used in the original work. To obtain the required training images for 3DGS optimization, we render the generated scene into RGB images \\(I\\). We additionally render 2D segmentation masks \\(\\mathcal{M}\\) to be used for 3D Gaussian scene editing." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.381, + 0.785, + 0.518 + ], + "angle": 0, + "content": "Scene Editing. For editing our 3DGS scenes based on the provided edit instruction, we build on top of Instruct-GS2GS [48] which in turn extends Instruct-NeRF2NeRF [20] for 3DGS scenes. Given the images used for training the 3DGS scene, Instruct-GS2GS updates the images individually using a diffusion model, namely Instruct-Pix2Pix [4]. The diffusion model takes three inputs: 1) an unedited conditioning image \\( I_0^v \\), a text instruction \\( c_{T} \\) and the noisy version of the current render \\( I_i^v \\) at optimization step \\( i \\), where \\( v \\) denotes the viewpoint from which the images are captured. Formally, the process of updating a single image is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.419, + 0.534, + 0.785, + 0.55 + ], + "angle": 0, + "content": "\\[\nI _ {i + 1} ^ {v} \\leftarrow U _ {\\theta} \\left(I _ {i} ^ {v}, t; I _ {0} ^ {v}, c _ {T}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.557, + 0.785, + 0.617 + ], + "angle": 0, + "content": "where \\( t \\) is the noise level within the constant range \\( [t_{\\mathrm{min}}, t_{\\mathrm{max}}] \\), \\( U_{\\theta} \\) is the sampling process of DDIM [41] and \\( I_{i+1}^{v} \\) is the edited image respecting the text instruction \\( c_{T} \\) and the unedited conditioning image \\( I_{0}^{v} \\). The training is performed by editing all the training images during a dataset update at every 2.5k training iterations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.618, + 0.785, + 0.729 + ], + "angle": 0, + "content": "Despite achieving general scene-level stylization, [48] lacks of support for more fine-grained object-level editing. To enable object-level edits, we use the binary masks \\(m_{\\{o\\}_{k=1}^{K}}\\). Each \\(m_{\\{o\\}_{k=1}^{K}}\\) is obtained by binarizing 2D segmentation masks \\(\\mathcal{M}\\) for the set of objects to edit \\(\\{o\\}_{k=1}^{K}\\). The set \\(\\{o\\}_{k=1}^{K}\\) is defined by extracting the referenced category names from the text instruction \\(c_{T}\\). Having the unedited conditioning image \\(I_0^v\\), the edited image \\(I_{i+1}^v\\), and the binary mask \\(m_{\\{o\\}_{k=1}^{K}}\\), we keep only the edits at the pixels of the target object set:" + }, + { + "type": "equation", + "bbox": [ + 0.347, + 0.742, + 0.785, + 0.761 + ], + "angle": 0, + "content": "\\[\nI _ {i + 1} ^ {v} = m _ {\\{o \\} _ {k = 1} ^ {K}} \\odot I _ {i + 1} ^ {v} + \\left(1 - m _ {\\{o \\} _ {k = 1} ^ {K}}\\right) \\odot I _ {0} ^ {v}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.785, + 0.841 + ], + "angle": 0, + "content": "where \\(\\odot\\) denotes element-wise multiplication of the image pixels. This way, the other pixels within \\(I_{i + 1}^{v}\\) are set to their unedited versions, enabling an object-level editing of training images. While training the 3DGS scene with the edited images, the mask \\(m_{\\{o\\}_{k = 1}^K}\\) is also applied for \\(\\mathcal{L}_1\\) and SSIM losses, ensuring gradient propagation only to the target objects." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.42, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SceneTeller: Language-to-3D Scene Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.377, + 0.163 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.283 + ], + "angle": 0, + "content": "Datasets. To create a set of supporting exemplars for the layout generation, we employ the 3D-FRONT dataset [18], a large-scale indoor scene dataset with 18,797 rooms furnished with 3D textured objects from 3D-FUTURE [19]. Following the same pre-processing steps as [33] and excluding non-rectangular floor plans from the test set similar to [15], we obtained 3397 train, 453 validation, 423 test bedroom and 690 train, 98 validation, 53 test living room scenes. During 3D scene assembling, 3D models are retrieved from the 3D-FUTURE dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.304, + 0.79, + 0.518 + ], + "angle": 0, + "content": "Implementation Details. We use ChatCompletions API\\(^3\\), which enables a series of dialogue exchanges between the user and the LLM, to query GPT-4 [1]. BlenderProc\\(^4\\) is employed to render 250 images per assembled scene with \\(512 \\times 512\\) resolution. After relocating the center of each scene to the origin of the coordinate system, we sample camera positions on the upper hemisphere at a distance of \\(1.5r\\) from the center where \\(r\\) is the diagonal length of the room, with an elevation angle of 35 degrees. We use the Splatifto model from NeRFstudio [46] for 3DGS training, by optimizing each scene for a maximum of 20k iterations. For editing the 3DGS training images, we use InstructPix2Pix [4] as the diffusion model, with classifier guidance scales \\(s_T = [3.5, 12.5]\\) for the text and \\(s_I = 1.5\\) for the image. The dataset updates are performed every 2.5k iterations following Instruct-GS2GS [48]. The InstructGS2GS training takes approximately 15-20 minutes per scene on a single A6000, most of it spent on diffusion model sampling. The rendering speed is 101.21 fps." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.539, + 0.449, + 0.554 + ], + "angle": 0, + "content": "4.1 3D Layout Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.564, + 0.789, + 0.806 + ], + "angle": 0, + "content": "Evaluation Setup and Baselines. In this section we evaluate the first stage of our approach, therefore we only focus on assessing the generated layouts. To our knowledge, SceneTeller is the first approach that conducts conditional layout generation guided by text prompts to control the position and orientation of every object, targeting a user-friendly framework for novices. Hence, replicating an interaction mode that is natural for humans is pivotal for our framework. Based on this, our layout generation performance is compared with the performance of humans solving the same task. In particular, each participant of our user study is presented with a series of textual descriptions, each accompanied by an empty layout according to the room dimensions. For each text prompt, participants are requested to draw bounding boxes and label them with the object names. Each text prompt is given to at least 2, at most 3 participants. This study is conducted with 12 participants and 40 text prompts. To generate text prompts, 40 layouts are selected from the test set and their text prompts are generated by providing the initially rule-based formed sentences to GPT for paraphrasing, as described in Sec. 3.1." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.81, + 0.616, + 0.826 + ], + "angle": 0, + "content": "3 See https://platform.openai.com/docs/api-reference/chat" + }, + { + "type": "page_footnote", + "bbox": [ + 0.221, + 0.826, + 0.55, + 0.84 + ], + "angle": 0, + "content": "See https://github.com/DLR-RM/BlenderProc" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.81, + 0.616, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.114, + 0.381, + 0.128 + ], + "angle": 0, + "content": "B.M. Öcal et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.284 + ], + "angle": 0, + "content": "Metrics. Based on the predicted number of bounding boxes in each category, we report recall, precision and accuracy [15]. An accuracy score of 100 indicates that the number of bounding boxes and their corresponding categories precisely match the ground-truth annotations. Otherwise, a score of 0 is assigned to the scene. All values are in percentage (\\%). We additionally report mean Intersection over Union (mIoU) between the predicted and ground-truth bounding boxes and out-of-bound rates, indicating the percentage of scenes where furniture extends beyond the floor plan boundary. We refer to the supplementary material for detailed metric descriptions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.307, + 0.789, + 0.52 + ], + "angle": 0, + "content": "Results. Tab. 1 presents the quantitative evaluation results of 3D layout generation. Our in-context-learning-based 3D layout generation module performs better than human participants in terms of mIoU, showcasing the effectiveness of LLMs' visual reasoning capabilities. In-context exemplars further enhance this performance by aiding in reasoning about the current data distribution. However, human participants possess their own understanding of the physical world, leading to variations in their notions of object sizes and prompt reasoning, resulting in slightly poorer performance in this assessment. Both our layout generation module and human participants performed excellently in terms of predicting the number of bounding boxes per category. Humans sometimes forget or miss objects mentioned in the prompt, which explains their slightly imperfect recall score. The ground-truth layouts from the dataset involve objects slightly exceeding the floor plan boundary, leading to \\(15.15\\%\\) OOB rate. Human participants are able to place the objects within the layout more accurately." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.545, + 0.79, + 0.604 + ], + "angle": 0, + "content": "Table 1: Quantitative evaluation of 3D layout generation. The proposed in-context learning based 3D layout generation module performs better than human participants in terms of predicting object sizes, enhanced by in-context exemplars aiding in reasoning about the current data distribution. All values are in percentage (\\%)." + }, + { + "type": "table", + "bbox": [ + 0.345, + 0.613, + 0.663, + 0.687 + ], + "angle": 0, + "content": "
MethodOOB ↓rec.↑prec.↑acc.↑mIoU ↑
Ours56.78100.00100.00100.0044.60
Human12.0998.79100.0095.0028.70
GT15.15100.00100.00100.00100.00
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.739, + 0.439, + 0.753 + ], + "angle": 0, + "content": "4.2 3D Scene Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Baselines. Our approach is evaluated and compared with three state-of-the-art 3D scene generation methods. Set-the-scene [11] employs compositional NeRFs that are initialized based on user-defined object proxies to have more control over individual objects. To compare with [11], layouts generated by our method are provided as input to their pipeline. GSGEN [9] is a 3DGS-based approach," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.419, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SceneTeller: Language-to-3D Scene Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.788, + 0.207 + ], + "angle": 0, + "content": "following a two-stage pipeline for geometry optimization and appearance refinement. LucidDreamer [10] employs render-refine-repeat paradigm by iteratively warping and inpainting the previously generated image to create an initial point cloud, which is then used for 3DGS scene training." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.236, + 0.78, + 0.266 + ], + "angle": 0, + "content": "a t a o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o ." + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.268, + 0.304, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.321, + 0.268, + 0.402, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.424, + 0.268, + 0.544, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.268, + 0.655, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.683, + 0.268, + 0.765, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.331, + 0.776, + 0.382 + ], + "angle": 0, + "content": "\"The dining table is situated in the middle of the room, but slightly leaning towards the lower left wall, with its orientation reversed and perpendicular. One dining chair is positioned near the lower left corner, with its orientation being perpendicular. Another dining chair is placed in the middle of the left wall, a bit nearer to the lower wall, and is also perpendicular. Yet another dining chair is set up at the midpoint of the left wall, but it's slightly more towards the upper wall and is perpendicular. One dining chair is near the lower right corner, with its orientation reversed and perpendicular. Another dining chair can be found at the center of the right wall, a bit closer to the lower wall, and it is perpendicular but facing the opposite direction. Another dining chair is at the center of the right wall, but slightly more towards the upper wall and its orientation is perpendicular but reversed. One more dining chair is in the middle of the upper wall, closer to the left wall, and is oriented backwards. Finally, a dining chair is at the center of the lower wall, a bit closer to the left wall, and it is oriented without any rotation. Make it marble.\"" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.387, + 0.305, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.321, + 0.387, + 0.401, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.424, + 0.387, + 0.548, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.562, + 0.388, + 0.669, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.683, + 0.389, + 0.769, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.45, + 0.777, + 0.503 + ], + "angle": 0, + "content": "The coffe table is sitd in the middle of the upper wall, albeit a tad nearer to the right wall, and is aligned perpendicularly facing the opposite direction. Another coffe table can be found in the middle of the right wall, a bit nearer to the lower wall, and is facing towards the rear. Nestled in the upper right corner is another coffe table, positioned without any rotation. In the upper left corner, a tv stand is positioned perpendicularly. A multi seat sofa is positioned near the upper right corner, facing perpendicularly but in the reverse direction. Close to the lower left corner, a dining table is situated without any rotation. A dinng chair can be found in the center of the lower wall, slightly nearer to the left wall, and is facing in its original direction. Another dinng chair is positioned near the lower left corner, also without any rotation. A third dinng chair is situated in the room's center, closer to the lower left wall, and oriented towards the rear. A fourth dinng chair is positioned in the middle of the left wall, slightly closer to the lower wall, positioned towards the rear. Lastly, a console table can be found in the lower left corner, placed without any rotation. Make it pink themed.\"" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.506, + 0.305, + 0.568 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.321, + 0.506, + 0.401, + 0.568 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.424, + 0.506, + 0.548, + 0.568 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.509, + 0.671, + 0.57 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.683, + 0.51, + 0.774, + 0.562 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.571, + 0.777, + 0.593 + ], + "angle": 0, + "content": "At the heart of the room, a double bed is positioned a tad nearer to the right wall, its orientation is perpendicular but in reverse. Adjacent to the top-right corner, a nightstand is placed, its orientation is also perpendicular but in reverse. A wardrobe can be found centrally aligned with the bottom wall, subtly closer to the right wall, with no rotation in its orientation. Make it Starry Night Van Gogh painting style.\"" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.595, + 0.304, + 0.657 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.321, + 0.595, + 0.402, + 0.658 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.424, + 0.595, + 0.545, + 0.657 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.593, + 0.655, + 0.655 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.683, + 0.593, + 0.764, + 0.655 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.659, + 0.779, + 0.689 + ], + "angle": 0, + "content": "In the middle of the room, a double bed is placed, slightly more towards the bottom right wall, with its orientation being perpendicular and facing the reverse direction. A wardrobe is situated at the center of the top wall, a bit closer to the right wall, with its orientation facing backwards. Near the top right corner of the room, a table is situated, oriented perpendicularly but in the reverse direction. At the bottom left corner, a cabinet is positioned, perpendicular to the walls. Another table is situated near the bottom right corner, oriented perpendicularly but facing the opposite direction. Turn the double bed into golden.\"" + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.691, + 0.303, + 0.753 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.243, + 0.758, + 0.281, + 0.767 + ], + "angle": 0, + "content": "GSGEN" + }, + { + "type": "image", + "bbox": [ + 0.325, + 0.692, + 0.404, + 0.754 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.332, + 0.757, + 0.397, + 0.767 + ], + "angle": 0, + "content": "Set-the-scene" + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.692, + 0.544, + 0.753 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.45, + 0.757, + 0.518, + 0.767 + ], + "angle": 0, + "content": "LucidDreamer" + }, + { + "type": "image", + "bbox": [ + 0.572, + 0.692, + 0.648, + 0.753 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.706, + 0.765, + 0.75 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.649, + 0.758, + 0.675, + 0.767 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.781, + 0.788, + 0.837 + ], + "angle": 0, + "content": "Fig. 3: Qualitative comparison with state-of-the-art text-to-3D scene generation methods. The top/upper walls are marked with arrows for your reference, if this information is available within the generations. SceneTeller is able to generate high-quality scenes, with superior geometric fidelity and 3D consistency." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.114, + 0.381, + 0.128 + ], + "angle": 0, + "content": "B.M. Öcal et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "User Study. We conduct a user study with 30 participants to assess the performance of our method in generating high-quality text-consistent realistic 3D scenes. Human evaluators are presented with 10 scenes, each accompanied by a textual prompt describing its layout and style. For each scene, they are requested to rate it on a scale from 1 to 5 (5 being the best) based on four criteria: (a) Realism, (b) Text Alignment, (c) Geometric Fidelity and (d) Compositional Plausibility." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.788, + 0.344 + ], + "angle": 0, + "content": "We present the results of the study in Tab. 2. Our method consistently outperforms the existing state-of-the-art 3D scene generation methods across all the four criteria by a significant margin, showcasing the merits of SceneTeller in generating realistic high-quality 3D scenes solely from text prompts. For additional details of the study and extensive analysis of the results, we refer to the supplementary material." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.371, + 0.79, + 0.456 + ], + "angle": 0, + "content": "Table 2: User study results. Study conducted with 30 participants to assess the performance of state-of-the-art methods in generating 3D scenes based on four criteria with a scale from 1 to 5 (5 being the best and vice versa). Our method consistently outperforms the existing state-of-the-art 3D scene generation methods across all the four criteria by a significant margin, showcasing the merits of SceneTeller in generating realistic high-quality 3D scenes solely from text prompts." + }, + { + "type": "table", + "bbox": [ + 0.239, + 0.466, + 0.767, + 0.585 + ], + "angle": 0, + "content": "
MethodInput TypeUser Study
RealismText AlignmentGeometric FidelityCompositional Plausibility
GSGEN [9]T1.3131.3751.4391.392
LucidDreamer [10]T3.5432.7513.0503.116
Set-the-scene [11]L + T1.9631.9002.1122.181
OursT4.1544.1894.4794.323
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.636, + 0.788, + 0.743 + ], + "angle": 0, + "content": "Quantitative Evaluation of Stylization. To evaluate the consistency of the generated scene style with textual descriptions, we utilize the CLIP Score [21] as the metric. Given that the CLIP score cannot reason about the parts of the prompts describing the layout, we only provide style information to CLIP in the format: 'A Barbie style bedroom' or 'A bedroom with Barbie style nightstands' for all methods. As shown in Tab. 3, our method consistently generates scenes adhering to the given style description." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Qualitative Evaluation. Visual comparison of SceneTeller with state-of-the-art text-to-3D methods are provided in Fig. 3. SceneTeller yields high-quality 3D scene generations, surpassing the baselines in terms of text-alignment and 3D consistency by predicting an initial 3D layout conditioned on the textual descriptions. Unlike some of the state-of-the-art baselines producing unrealistic" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.419, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SceneTeller: Language-to-3D Scene Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 3: Quantitative comparison of stylization. We report the CLIP Score to evaluate the consistency of the generated scene style with textual descriptions. Our method consistently generates scenes adhering to the given style description." + }, + { + "type": "table", + "bbox": [ + 0.263, + 0.199, + 0.742, + 0.242 + ], + "angle": 0, + "content": "
GSGEN [9]LucidDreamer [10]Set-the-scene [11]Ours
CLIP Score ↑26.47830.71527.18830.923
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.27, + 0.786, + 0.302 + ], + "angle": 0, + "content": "appearance (e.g. comic or cartoon-like characteristics), SceneTeller is able to produce more realistic object shapes and appearances." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.324, + 0.787, + 0.385 + ], + "angle": 0, + "content": "Realism of the Generated Scenes. We evaluate all methods using FID score with a real indoor dataset ScanNet [12] on 10 generated scenes by rendering 50 images from each of them. As shown in Tab. 4, our method consistently generates more realistic scenes." + }, + { + "type": "table_caption", + "bbox": [ + 0.33, + 0.409, + 0.673, + 0.426 + ], + "angle": 0, + "content": "Table 4: FID score with the ScanNet dataset [12]." + }, + { + "type": "table", + "bbox": [ + 0.286, + 0.436, + 0.716, + 0.479 + ], + "angle": 0, + "content": "
GSGEN [9] LucidDreamer [10] Set-the-scene [11]Ours
FID ↓297.10238.75297.57167.27
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.533, + 0.788, + 0.61 + ], + "angle": 0, + "content": "Applicability to Real-world. While current design tools enable professionals to create floor plans, visualize 3D models, and render photo-realistic images, they are not suitable for novice users unfamiliar with such software tools. By enabling users to interact solely through plain natural language prompts, our approach also empowers novice users to flexibly design their own rooms." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.633, + 0.391, + 0.649 + ], + "angle": 0, + "content": "4.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Effect of design choices for 3D layout generation. To assess the effectiveness of individual components in our design, we conduct an ablation study. Specifically, we examine two aspects: 1) the prompt source for in-context exemplars, and 2) the technique for selecting in-context exemplars. Regarding the prompt source, we present exemplars where the textual description is generated either solely by the rule-based formulation (R) or by a combination of the rule-based formulation and GPT paraphrasing \\((\\mathrm{R} + \\mathrm{GPT})\\) to introduce linguistic variation as described in Sec. 3.1. For the exemplar selection technique, we experiment with a random and a retrieval-based exemplar selection following [15], \\(k = 8\\) exemplars. Additionally, we experiment with a third set of exemplars, denoted as \\(pos + neg\\), where 4 exemplars are selected by retrieval, and the other 4 are chosen from negative exemplars involving out-of-bound (OOB)" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.114, + 0.381, + 0.128 + ], + "angle": 0, + "content": "B.M. Öcal et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "or overlapping objects. The negative ones are from the filtered samples during the pre-processing of the dataset, and the LLM is specifically instructed to avoid them. Tab. 5 verifies that providing in-context exemplars with linguistic variation \\((\\mathrm{R} + \\mathrm{GPT})\\) selected with a retrieval-based approach performs the best, especially in terms of OOB rate by constraining the in-context learning of LLM to similar layouts. Using the \\(pos + neg\\) set also considerably improves the performance in terms of mIoU." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.274, + 0.788, + 0.334 + ], + "angle": 0, + "content": "Table 5: Effect of design choices for 3D layout generation. We report the 3D layout generation performance w.r.t different design choices such as the prompt source and the exemplar selection. Providing in-context exemplars to LLM with linguistic variation (R+GPT) selected with a retrieval-based approach, performs the best." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.343, + 0.785, + 0.489 + ], + "angle": 0, + "content": "
Prompt sourceExemplar SelectionOOB(%)Numerical ReasoningBbox
random retrieval pos+neg↓rec.↑prec.↑acc.↑
1R✓74.12100.00100.00100.0031.10
2R✓64.87100.00100.00100.0045.50
3R74.10100.00100.00100.0041.90
4R + GPT✓76.1199.1099.1099.1029.90
5R + GPT✓54.65100.00100.00100.0044.20
6R + GPT66.45100.00100.00100.0043.90
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.538, + 0.36, + 0.554 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.566, + 0.788, + 0.659 + ], + "angle": 0, + "content": "In this paper, we presented a novel pipeline for 3D room generation from a textual description. Our approach supports style editing either on the individual object or on the entire scene level. Rooms generated by our method were preferred by human evaluators over the baseline results almost in all cases, confirming its efficacy on this challenging task. Importantly, our approach can be used by novices, thus making a step towards democratizing 3D scene design." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.677, + 0.405, + 0.695 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.787, + 0.752 + ], + "angle": 0, + "content": "We thank members of the Bosch-UvA Delta Lab and anonymous reviewers for helpful discussions and feedback. This project was generously supported by the Bosch Center for Artificial Intelligence." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.771, + 0.325, + 0.788 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.798, + 0.787, + 0.842 + ], + "angle": 0, + "content": "1. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.419, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SceneTeller: Language-to-3D Scene Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.147, + 0.788, + 0.203 + ], + "angle": 0, + "content": "2. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems (NeurIPS) 35, 23716-23736 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.204, + 0.787, + 0.26 + ], + "angle": 0, + "content": "3. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (CVPR). pp. 5855-5864 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.261, + 0.787, + 0.302 + ], + "angle": 0, + "content": "4. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18392-18402 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.303, + 0.787, + 0.331 + ], + "angle": 0, + "content": "5. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. In: European Conference on Computer Vision (ECCV) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.332, + 0.787, + 0.374 + ], + "angle": 0, + "content": "6. Chen, R., Chen, Y., Jiao, N., Jia, K.: Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.375, + 0.787, + 0.416 + ], + "angle": 0, + "content": "7. Chen, X., Zhang, Q., Li, X., Chen, Y., Feng, Y., Wang, X., Wang, J.: Hallucinated neural radiance fields in the wild. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12943-12952 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.417, + 0.787, + 0.459 + ], + "angle": 0, + "content": "8. Chen, Y., Chen, Z., Zhang, C., Wang, F., Yang, X., Wang, Y., Cai, Z., Yang, L., Liu, H., Lin, G.: Gaussianeditor: Swift and controllable 3d editing with gaussian splatting. arXiv preprint arXiv:2311.14521 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.46, + 0.787, + 0.487 + ], + "angle": 0, + "content": "9. Chen, Z., Wang, F., Liu, H.: Text-to-3d using gaussian splatting. arXiv preprint arXiv:2309.16585 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.488, + 0.787, + 0.515 + ], + "angle": 0, + "content": "10. Chung, J., Lee, S., Nam, H., Lee, J., Lee, K.M.: Luciddreamer: Domain-free generation of 3d gaussian splatting scenes. arXiv preprint arXiv:2311.13384 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.516, + 0.787, + 0.572 + ], + "angle": 0, + "content": "11. Cohen-Bar, D., Richardson, E., Metzer, G., Giryes, R., Cohen-Or, D.: Set-the-scene: Global-local training for generating controllable nerf scenes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops. pp. 2920–2929 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.573, + 0.787, + 0.614 + ], + "angle": 0, + "content": "12. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proc. Computer Vision and Pattern Recognition (CVPR), IEEE (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.615, + 0.787, + 0.643 + ], + "angle": 0, + "content": "13. Fang, J., Wang, J., Zhang, X., Xie, L., Tian, Q.: Gaussianeditor: Editing 3d gaussians delicately with text instructions. arXiv preprint arXiv:2311.16037 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.644, + 0.787, + 0.685 + ], + "angle": 0, + "content": "14. Fang, J., Wang, J., Zhang, X., Xie, L., Tian, Q.: Gaussianeditor: Editing 3d gaussians delicately with text instructions. In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.686, + 0.787, + 0.741 + ], + "angle": 0, + "content": "15. Feng, W., Zhu, W., Fu, T.j., Jampani, V., Akula, A., He, X., Basu, S., Wang, X.E., Wang, W.Y.: Layoutgpt: Compositional visual planning and generation with large language models. Advances in Neural Information Processing Systems (NeurIPS) 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.742, + 0.787, + 0.784 + ], + "angle": 0, + "content": "16. Fridman, R., Abecasis, A., Kasten, Y., Dekel, T.: Scenescape: Text-driven consistent scene generation. Advances in Neural Information Processing Systems (NeurIPS) 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.785, + 0.787, + 0.84 + ], + "angle": 0, + "content": "17. Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5501-5510 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.219, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.114, + 0.381, + 0.128 + ], + "angle": 0, + "content": "B.M. Öcal et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.203 + ], + "angle": 0, + "content": "18. Fu, H., Cai, B., Gao, L., Zhang, L.X., Wang, J., Li, C., Zeng, Q., Sun, C., Jia, R., Zhao, B., et al.: 3d-front: 3d furnished rooms with layouts and semantics. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10933-10942 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.787, + 0.245 + ], + "angle": 0, + "content": "19. Fu, H., Jia, R., Gao, L., Gong, M., Zhao, B., Maybank, S., Tao, D.: 3d-future: 3d furniture shape with texture. International Journal of Computer Vision 129, 3313-3337 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.786, + 0.286 + ], + "angle": 0, + "content": "20. Haque, A., Tancik, M., Efros, A., Holynski, A., Kanazawa, A.: Instruct-nerf2nerf: Editing 3d scenes with instructions. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (CVPR) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.787, + 0.383 + ], + "angle": 0, + "content": "21. Hessel, J., Holtzman, A., Forbes, M., Le Bras, R., Choi, Y.: CLIPScore: A reference-free evaluation metric for image captioning. In: Moens, M.F., Huang, X., Specia, L., Yih, S.W.t. (eds.) Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. pp. 7514-7528. Association for Computational Linguistics, Online and Punta Cana, Dominican Republic (Nov 2021). https://doi.org/10.18653/v1/2021.emnlp-main.595, https://aclanthology.org/2021.emnlp-main.595" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.787, + 0.412 + ], + "angle": 0, + "content": "22. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.787, + 0.467 + ], + "angle": 0, + "content": "23. Hollein, L., Cao, A., Owens, A., Johnson, J., Nießner, M.: Text2room: Extracting textured 3d meshes from 2d text-to-image models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 7909-7920 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.468, + 0.787, + 0.509 + ], + "angle": 0, + "content": "24. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (July 2023), https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.787, + 0.55 + ], + "angle": 0, + "content": "25. Li, Q., Li, F., Guo, J., Guo, Y.: Uhdnerf: Ultra-high-definition neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (CVPR). pp. 23097-23108 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.787, + 0.591 + ], + "angle": 0, + "content": "26. Li, X., Wang, H., Tseng, K.K.: Gaussiandiffusion: 3d gaussian splatting for denoising diffusion probabilistic models with structured noise. arXiv preprint arXiv:2311.11221 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.591, + 0.787, + 0.647 + ], + "angle": 0, + "content": "27. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3d: High-resolution text-to-3d content creation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 300-309 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.787, + 0.688 + ], + "angle": 0, + "content": "28. Lin, Y., Bai, H., Li, S., Lu, H., Lin, X., Xiong, H., Wang, L.: Componerf: Text-guided multi-object compositional nerf with editable 3d scene layout. arXiv preprint arXiv:2303.13843 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.787, + 0.743 + ], + "angle": 0, + "content": "29. Lorraine, J., Xie, K., Zeng, X., Lin, C.H., Takikawa, T., Sharp, N., Lin, T.Y., Liu, M.Y., Fidler, S., Lucas, J.: Att3d: Amortized text-to-3d object synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 17946-17956 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.787, + 0.799 + ], + "angle": 0, + "content": "30. Metzer, G., Richardson, E., Patashnik, O., Giryes, R., Cohen-Or, D.: Latent-nerf for shape-guided generation of 3d shapes and textures. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12663-12673 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.842 + ], + "angle": 0, + "content": "31. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: Proceedings of the European Conference on Computer Vision (ECCV) (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.842 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.419, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SceneTeller: Language-to-3D Scene Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "32. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.788, + 0.229 + ], + "angle": 0, + "content": "33. Paschalidou, D., Kar, A., Shugrina, M., Kreis, K., Geiger, A., Fidler, S.: Atiss: Autoregressive transformers for indoor scene synthesis. Advances in Neural Information Processing Systems 34, 12013-12026 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.229, + 0.788, + 0.256 + ], + "angle": 0, + "content": "34. Phung, Q., Ge, S., Huang, J.B.: Grounded text-to-image synthesis with attention refocusing. arXiv preprint arXiv:2306.05427 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.256, + 0.788, + 0.297 + ], + "angle": 0, + "content": "35. Po, R., Wetzstein, G.: Compositional 3d scene generation using locally conditioned diffusion. ArXiv abs/2303.12218 (2023), https://api-semanticscholar.org/CorpusID:257663283" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.296, + 0.788, + 0.338 + ], + "angle": 0, + "content": "36. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. In: The Eleventh International Conference on Learning Representations (ICLR) (2023), https://openreview.net/forum?id=FjNys5c7VvY" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.337, + 0.788, + 0.405 + ], + "angle": 0, + "content": "37. Qian, G., Mai, J., Hamdi, A., Ren, J., Siarohin, A., Li, B., Lee, H.Y., Skorokhodov, I., Wonka, P., Tulyakov, S., Ghanem, B.: Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. In: The Twelfth International Conference on Learning Representations (ICLR) (2024), https://openreview.net/forum?id=0jHkUDyEO9" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.405, + 0.788, + 0.461 + ], + "angle": 0, + "content": "38. Raj, A., Kaza, S., Poole, B., Niemeyer, M., Mildenhall, B., Ruiz, N., Zada, S., Aberman, K., Rubenstein, M., Barron, J., Li, Y., Jampani, V.: Dreambooth3d: Subject-driven text-to-3d generation. The International Conference on Computer Vision (ICCV) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.459, + 0.788, + 0.514 + ], + "angle": 0, + "content": "39. Reiser, C., Szeliski, R., Verbin, D., Srinivasan, P., Mildenhall, B., Geiger, A., Barron, J., Hedman, P.: Merf: Memory-efficient radiance fields for real-time view synthesis in unbounded scenes. ACM Transactions on Graphics (TOG) 42(4), 1-12 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.514, + 0.788, + 0.556 + ], + "angle": 0, + "content": "40. Ren, J., He, C., Liu, L., Chen, J., Wang, Y., Song, Y., Li, J., Xue, T., Hu, S., Chen, T., Zheng, K., Xiang, J., Bo, L.: Make-a-character: High quality text-to-3d character generation within minutes. arXiv preprint arXiv:2312.15430 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.554, + 0.788, + 0.595 + ], + "angle": 0, + "content": "41. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.594, + 0.788, + 0.718 + ], + "angle": 0, + "content": "42. Sanh, V., Webson, A., Raffel, C., Bach, S., Sutawika, L., Ålyafeai, Z., Chaffin, Å., Stiegler, A., Raja, A., Dey, M., Bari, M.S., Xu, C., Thakker, U., Sharma, S.S., Szczechla, E., Kim, T., Chhablani, G., Nayak, N., Datta, D., Chang, J., Jiang, M.T.J., Wang, H., Manica, M., Shen, S., Yong, Z.X., Pandey, H., Bawden, R., Wang, T., Neeraj, T., Rozen, J., Sharma, A., Santilli, A., Fevry, T., Fries, J.A., Teehan, R., Scao, T.L., Biderman, S., Gao, L., Wolf, T., Rush, A.M.: Multitask prompted training enables zero-shot task generalization. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=9Vrb9DOWI4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.718, + 0.788, + 0.759 + ], + "angle": 0, + "content": "43. Sun, C., Sun, M., Chen, H.T.: Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5459-5469 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.759, + 0.788, + 0.786 + ], + "angle": 0, + "content": "44. Sun, C., Han, J., Deng, W., Wang, X., Qin, Z., Gould, S.: 3d-gpt: Procedural 3d modeling with large language models. arXiv preprint arXiv:2310.12945 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.785, + 0.788, + 0.841 + ], + "angle": 0, + "content": "45. Tancik, M., Casser, V., Yan, X., Pradhan, S., Mildenhall, B., Srinivasan, P.P., Barron, J.T., Kretzschmar, H.: Block-nerf: Scalable large scene neural view synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8248-8258 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.114, + 0.381, + 0.128 + ], + "angle": 0, + "content": "B.M. Öcal et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.204 + ], + "angle": 0, + "content": "46. Tancik, M., Weber, E., Ng, E., Li, R., Yi, B., Kerr, J., Wang, T., Kristoffersen, A., Austin, J., Salahi, K., Ahuja, A., McAllister, D., Kanazawa, A.: Nerfstudio: A modular framework for neural radiance field development. In: ACM SIGGRAPH 2023 Conference Proceedings. SIGGRAPH '23 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.205, + 0.786, + 0.232 + ], + "angle": 0, + "content": "47. Tang, J., Ren, J., Zhou, H., Liu, Z., Zeng, G.: Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.233, + 0.786, + 0.259 + ], + "angle": 0, + "content": "48. Vachha, C., Haque, A.: Instruct-gs2gs: Editing 3d gaussian splats with instructions (2024), https://instruct-gs2gs.github.io/" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.26, + 0.787, + 0.314 + ], + "angle": 0, + "content": "49. Wang, H., Du, X., Li, J., Yeh, R.A., Shakhnarovich, G.: Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 12619-12629 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.315, + 0.786, + 0.357 + ], + "angle": 0, + "content": "50. Wang, Z., Lu, C., Wang, Y., Bao, F., Li, C., Su, H., Zhu, J.: Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. In: Advances in Neural Information Processing Systems (NeurIPS) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.357, + 0.786, + 0.412 + ], + "angle": 0, + "content": "51. Wei, J., Bosma, M., Zhao, V., Guu, K., Yu, A.W., Lester, B., Du, N., Dai, A.M., Le, Q.V.: Finetuned language models are zero-shot learners. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=gEZrGcozdqR" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.412, + 0.786, + 0.48 + ], + "angle": 0, + "content": "52. Xu, Y., Chai, M., Shi, Z., Peng, S., Skorokhodov, I., Siarohin, A., Yang, C., Shen, Y., Lee, H.Y., Zhou, B., et al.: Discoscope: Spatially disentangled generative radiance fields for controllable 3d-aware scene synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4402-4412 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.481, + 0.786, + 0.522 + ], + "angle": 0, + "content": "53. Yang, Z., Gan, Z., Wang, J., Hu, X., Lu, Y., Liu, Z., Wang, L.: An empirical study of gpt-3 for few-shot knowledge-based vqa. In: Proceedings of the AAAI Conference on Artificial Intelligence (AAAI). vol. 36, pp. 3081-3089 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.522, + 0.786, + 0.564 + ], + "angle": 0, + "content": "54. Yi, T., Fang, J., Wu, G., Xie, L., Zhang, X., Liu, W., Tian, Q., Wang, X.: Gaussian dreamer: Fast generation from text to 3d gaussian splatting with point cloud priors. arXiv preprint arXiv:2310.08529 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.565, + 0.786, + 0.605 + ], + "angle": 0, + "content": "55. Yu, A., Li, R., Tancik, M., Li, H., Ng, R., Kanazawa, A.: Plenoptrees for real-time rendering of neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (CVPR). pp. 5752-5761 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.606, + 0.786, + 0.647 + ], + "angle": 0, + "content": "56. Zhou, X., Ran, X., Xiong, Y., He, J., Lin, Z., Wang, Y., Sun, D., Yang, M.H.: Gala3d: Towards text-to-3d complex scene generation via layout-guided generative gaussian splatting. arXiv preprint arXiv:2402.07207 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.647, + 0.786, + 0.688 + ], + "angle": 0, + "content": "57. Zhu, J., Zhuang, P., Koyejo, S.: HIFA: High-fidelity text-to-3d generation with advanced diffusion guidance. In: The Twelfth International Conference on Learning Representations (ICLR) (2024), https://openreview.net/forum?id=IZMPWmcS3H" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.688 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/SceneTeller_ Language-to-3D Scene Generation/dec651d5-2a06-4ad2-af7e-aeaf7a7dadd2_origin.pdf b/2024/SceneTeller_ Language-to-3D Scene Generation/dec651d5-2a06-4ad2-af7e-aeaf7a7dadd2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8e5a8daeb16089fc0878ac8c774d23cb8b88b61a --- /dev/null +++ b/2024/SceneTeller_ Language-to-3D Scene Generation/dec651d5-2a06-4ad2-af7e-aeaf7a7dadd2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0bfa4967326d673e6cf614561f619bc8964a08b8351c576cc8bd201998686250 +size 5483366 diff --git a/2024/SceneTeller_ Language-to-3D Scene Generation/full.md b/2024/SceneTeller_ Language-to-3D Scene Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..66ff593dfe904a7ad9eb13090c0d0de8a1541101 --- /dev/null +++ b/2024/SceneTeller_ Language-to-3D Scene Generation/full.md @@ -0,0 +1,325 @@ +# SceneTeller: Language-to-3D Scene Generation + +Bashak Melis Öcal1*, Maxim Tatarchenko2, Sezer Karaoglu1, and Theo Gevers1 + +1 UvA-Bosch Delta Lab, University of Amsterdam + +$^{2}$ Bosch Center for AI, Robert Bosch GmbH + +![](images/77fbd33969abdf6afa836f2b511b9d28687c1140c5661da868af46dd949cc097.jpg) +Fig. 1: With solely a textual prompt in natural language describing the desired spatial positions and orientations of objects within the scene, SceneTeller is able to generate realistic and high-quality 3D spaces. Additionally, by facilitating modifications to the style of the entire scene or individual objects within the scene through edit instructions, SceneTeller offers a practical and flexible framework for designing personalized rooms. + +Abstract. Designing high-quality indoor 3D scenes is important in many practical applications, such as room planning or game development. Conventionally, this has been a time-consuming process which requires both artistic skill and familiarity with professional software, making it hardly accessible for layman users. However, recent advances in generative AI have established solid foundation for democratizing 3D design. In this paper, we propose a pioneering approach for text-based 3D room design. Given a prompt in natural language describing the object placement in the room, our method produces a high-quality 3D scene corresponding to it. With an additional text prompt the users can change the appearance of the entire scene or of individual objects in it. Built using in-context learning, CAD model retrieval and 3D-Gaussian-Splatting-based stylization, our turnkey pipeline produces state-of-the-art 3D scenes, while being easy to use even for novices. Our project page is available at https://sceneteller.github.io/. + +Keywords: Text-to-3D scene generation $\cdot$ Language-guided generation $\cdot$ 3D Gaussian splatting scene stylization + +# 1 Introduction + +3D scene design has numerous applications in domains like architecture or game development. Designing high-quality 3D scenes, however, is a laborious and time-consuming process, challenging even for highly-skilled domain experts utilizing professional tools. This complexity often leaves novice users unable to customize their own spaces. Recent advancements in automatic text-to-image synthesis [22,41] have led to massive progress in text-to-3D creation [36-38], enabling the generation of 3D content only from text input. Such works create foundation for democratizing 3D design, supporting its accessibility for non-experts. + +A pioneering work by Poole et al. [36] introduced object-centric text-to-3D generation based on 2D diffusion models coupled with NeRFs [31]. Several subsequent works proposed ways to improve the geometric accuracy, multi-view consistency, and appearance quality [6,27,30,50]. Despite their success in object-level generation, hallucinating complex large-scale scenes with multiple objects is still remaining a challenging problem. Some recent 3D scene generation methods [11,35,52] require a 3D layout as input, limiting their usability for novices. Others [16,23] produce results that violate global object arrangements and are not multi-view consistent. + +Aiming to address both these shortcomings - insufficient user friendliness and unsatisfactory global consistency - we introduce SceneTeller, a framework for text-based 3D scene generation. 1) We start with a premise that a 3D scene generation framework should offer control over the placement and style of individual objects, allowing novice users to design their customized spaces through simple interactions with the system. Our approach achieves this by relying on text-based user interaction for scene and appearance specification. 2) In order to generate compositionally plausible 3D scenes, it is important to incorporate + +some degree of global information guidance. We implement this by conditioning the method on global layout descriptions. + +To this end, we split the task of generating high-quality 3D scenes into three steps: 3D layout generation, object population and style editing. 1) When presented with a prompt describing the desired spatial positions and orientations of objects, SceneTeller generates an initial 3D layout of the scene by exploiting the Large Language Models (LLMs) through in-context learning, to ensure global consistency. 2) The generated layout is then used to assemble a 3D scene and populate it with CAD models from a database. 3) To facilitate scene editing, the assembled scene is fitted with a 3D Gaussian Splitting [24] representation. This representation is then used to stylize the scene according to the user-provided edit prompt targeting the entire scene or individual objects within it. + +Extensive quantitative and qualitative evaluations show that SceneTeller performs significantly better than state-of-the-art methods, offering superior geometric fidelity and compositional plausibility. In brief, our contributions to the existing body of work can be summarized as follows: + +- We introduce a turnkey pipeline for high-quality 3D scene generation yielding complex, realistic and globally consistent 3D scenes, given only scene descriptions in natural language. Our pipeline supports object- and scene-level appearance editing. +- We propose a novel LLM-based module for 3D layout generation trained through in-context learning, which provides control over individual object placement. +- We experimentally demonstrate that SceneTeller generates 3D scenes of higher quality and better consistency than state-of-the-art methods, while boasting unprecedented user friendliness and flexibility. + +# 2 Related Work + +# 2.1 3D Scene Representations + +Rendering 3D scenes using implicit neural networks has seen remarkable success, notably with NeRF's [31] photo-realistic image synthesis. Extensions of NeRF improve reconstruction quality [3, 25], training and rendering speed [5, 17, 32, 39, 43, 55], and handling large-scale scenes [7, 45]. Despite advancements, NeRF-based methods still struggle with slow rendering speeds. Recently, 3D Gaussian splatting (3DGS) was proposed for real-time, high-quality rendering using anisotropic 3D Gaussians and fast tile-based rasterization. Our framework employs 3DGS for scene representation and allows edits on these Gaussians, generating high-quality, 3D consistent scenes. + +# 2.2 Text-to-3D Generation + +Object-centric Generation. 3D generation from textual prompts has recently grown in popularity. DreamFusion [36] introduces Score Distillation Sampling (SDS) to optimize NeRF representations by leveraging an off-the-shelf + +2D diffusion model, paving the way towards text-driven 3D content creation [29,37,38,57]. Following [36], Magic3D [27] employs a coarse-to-fine optimization scheme, exploiting low-resolution diffusion priors, and then enhancing the texture. To achieve high-fidelity mesh generation, Fantasia3D [6] disentangles the appearance from geometry. Latent-NeRF [30] optimizes a NeRF in the latent space of StableDiffusion [41], and additionally proposes incorporating 3D priors to generate a target geometry. To address the over-smoothing, over-saturation and lack of diversity problems within the SDS, ProfilicDreamer [50] proposes a principled particle-based variational framework, namely Variational Score Distillation (VSD). Following the same motivation, [49] presents Perturb-and-Average Scoring to estimate the score for non-noisy images. + +With the recent introduction of 3D Gaussian Splatting [24], several works attempt to combine 3D Gs and diffusion models. GaussianDreamer [54] employs a 3D diffusion model to generate the initialized point clouds for 3D Gaussians, and then further optimize the Gaussians using SDS with a 2D diffusion model. [9, 47] follows a two-stage pipeline for geometry optimization and appearance refinement. GaussianDiffusion [26] focuses on the challenge of achieving multiview consistency in 3D generation and proposes combining structured noise with a variational 3DGS technique. Despite their success in generating object-level scenes, both NeRF-based and 3DGS-based methods fall short in hallucinating complex scenes with multiple objects. + +Scene Generation. To represent multiple objects within the scenes, existing works [11, 28, 35, 52] rely on compositional NeRFs that are initialized based on user-defined object proxies. However, they are usually restricted to produce only narrow inward-facing views of the rooms. Also, the requirement of having 3D layouts as input is restrictive for final users. Orthogonal to those, SceneScape [16] and Text2Room [23] generate scenes following the render-refine-repeat paradigm by iteratively warping and inpainting the previously generated image according to the specified camera motion. This strategy does not allow object-level control during scene generation, and the results tend to exhibit multi-view inconsistencies. LucidDreamer [10] employs the same paradigm for an initial point cloud reconstruction, and then further optimizes the scene using 3DGS. Recently, several works introduced techniques for 3DGS scene editing, by manipulating the training images using 2D diffusion models [8, 13, 48] and incorporating the edited images into 3DGS scenes for further optimization. Concurrent to our work, [56] initializes a 3DGS scene from the layouts created using LLMs and optimizes it using a diffusion model. However, [56] lacks control over the generated layout in terms of positions and orientations of objects. + +# 2.3 LLMs for Vision + +With the inclusion of grounding information in LLMs, they demonstrate enhanced reasoning capabilities for visual concepts. Recently, the visual reasoning abilities of LLMs have been utilized for 3D tasks such as avatar generation [40], instruction-driven 3D modeling [44], 3D editing [14]. Exploiting the + +spatial reasoning capabilities of LLMs, LayoutGPT [15] extends beyond its 2D counterparts [34] to generate random 3D layouts of specific room types through in-context learning. However, LayoutGPT lacks support for controlling the generated layout. The 3D layout generation module of SceneTeller is the first of its kind allowing control over individual object placement in accordance with the provided natural language prompts describing the desired positions of objects. + +# 3 Method + +![](images/7cea33ba983928d37ab833debb0d181fe011cb0a4db0f3442d8482c6bc6131a6.jpg) +Fig. 2: Overview of our method. Given a textual prompt in natural language delineating the desired spatial positions and orientations of objects within the scene, a 3D scene layout is generated using in-context learning. An initial 3D scene is assembled for the predicted layout, which is then fitted with a 3D Gaussian Splatting representation. This representation is then used to stylize the scene according to the user-provided text prompt, and subsequently render the final images of the scene. + +The pipeline of SceneTeller is divided into three main stages: language-driven 3D layout generation, 3D scene assembling from the layout and 3D scene stylization. In the first stage, a 3D scene layout is generated based on the given text prompt (Sec. 3.1). Then, an initial 3D scene is assembled for the predicted layout (Sec. 3.2). After that, a 3D Gaussian Splatting representation is fitted to the assembled scene. This representation is then used to stylize the scene according to the user-provided text prompt, and subsequently render the final images of the scene (Sec. 3.3. The scheme of the framework is illustrated in Fig. 2 + +# 3.1 Language-driven 3D Layout Generation + +The task of this module is to produce a collection of object bounding boxes conditioned on a brief textural description of their spatial positions within the + +scene in plain language. More formally, the input is a condition $\mathcal{C}$ , consisting of a textual description $\mathcal{V}$ , along with the scene type and scene dimensions information. The output is a 3D layout represented as a set of 3D bounding boxes $\mathbf{b}_i \in B = \{\mathbf{b}_i\}_{i=1}^N$ for each of the $N$ referenced objects. Each 3D bounding box $\mathbf{b}_i = (c_i, \mathbf{t}_i, \mathbf{s}_i, o_i)$ consists of the following components: category name $c_i$ , box center $\mathbf{t}_i = (x_i, y_i, z_i) \in \mathbb{R}^3$ , box dimensions $\mathbf{s}_i = (w_i, h_i, d_i) \in \mathbb{R}^3$ and orientation $o_i \in \mathbb{R}$ . + +The textual description $\mathcal{V}$ is comprised of multiple sentences, each specifying the position and orientation of a single object within a canonical coordinate system. The canonical system is chosen to represent the scene in a 2D perspective, to ensure that the descriptions closely resemble those produced by a human observer. $\mathcal{V}$ is initially generated using a rule-based approach formulated in three steps as follows: 1) The canonical coordinate system is partitioned into a $3\times 3$ grid, creating rectangular patches. 2) Each of the predicted objects is assigned to a patch based on the center position $\mathbf{t}_i$ and orientation $o_i$ of its bounding box within these patches. 3) Per-object sentences are then formulated according to which patch the object is assigned to and object's orientation, e.g. for the object assigned to the top-left patch, with category name $c_i$ and 90 degrees orientation, we generate: "A $c_i$ is placed at the top-left corner of the room, with a perpendicular orientation." Given the relatively limited linguistic variation inherent in rule-based descriptions, initial textual templates are provided to an LLM to serve as a basis for paraphrasing. + +LayoutGPT [15] focused on improving an LLM's interpretation of the spatial knowledge by using a structured format featuring attribute-value pairs. We follow a similar strategy and adopt the standard CSS (Cascading Style Sheets) format to describe $\mathbf{b}_i$ . Bounding box values are mapped to standard CSS format attributes and category name is employed as the selector. + +Prompt Construction. Recent progress in the field has showcased the impressive in-context learning capabilities of LLMs, which entail conditioning these models with both natural language instructions and a select set of task demonstrations. Following this conditioning, the LLM is then tasked with completing further instances based on the provided demonstrations. In line with this, our prompts to LLM include three main parts: task specifications, in-context exemplars and the query condition: + +- Task specifications: Following prior work [42, 51] on enhancing LLMs instruction following capabilities, a task description is incorporated at the beginning of each prompt, which explains the goal of the task, establishes a standard for the 3D layout format in CSS style and provides unit information for attributes. Additionally, constraints are integrated into the prompts to guide the LLM and minimize errors (e.g., predicting overlapping boxes or placing objects out of bounds), during task completion. +- In-context learning: Supporting exemplars for the in-context learning are selected by adopting the retrieval-based approach used in [2, 15, 53]. When provided with a set of supporting exemplars $S = \{(\mathcal{C}_m^s,\mathbf{b}_m^s)\}_{m = 1}^M$ and the + +queried condition $\mathcal{C}_q$ , the function $f(\mathcal{C}_k^s,\mathcal{C}_q) = \|rl_k - rl_q\|^2 + \|rw_k - rw_q\|^2$ is computed between each element of the set and $\mathcal{C}_q$ following [15], where $rl$ and $rw$ are the length and width of the scenes. Top- $k$ supporting exemplars with the shortest distance to $\mathcal{C}_q$ are selected for in-context learning, provided to LLM with the same format with $\mathcal{C}_q$ . + +- Query condition: The inference condition $\mathcal{C}_q$ , for which we want to predict the layout. + +# 3.2 3D Scene Assembling from Layouts + +To assemble a scene based on the generated layouts, we rely on object retrieval. More specifically, for each of the bounding boxes, the nearest 3D model of the corresponding predicted category is retrieved from the furniture dataset based on the Euclidean distance of the bounding box dimensions as follows: + +$$ +d \left(\mathbf {s} _ {i}, \mathbf {s} _ {\text {c a n d}}\right) = \sqrt {\left(w _ {i} - w _ {\text {c a n d}}\right) ^ {2} + \left(h _ {i} - h _ {\text {c a n d}}\right) ^ {2} + \left(d _ {i} - d _ {\text {c a n d}}\right) ^ {2}}, \tag {1} +$$ + +where $\mathbf{s}_{\text{cand}}$ denotes the bounding box dimensions for the candidate object from the dataset. The retrieved 3D model is placed within the scene according to the predicted box center $\mathbf{t}_i$ and orientation angle $o_i$ . + +# 3.3 3D Scene Stylization + +Given a generated 3D scene, the users should be able to modify either the overall appearance of the scene or specific objects within it. Our approach facilitates scene editing by training a 3D Gaussian splatting model representing the generated scene and refining this representation according to the provided edit instruction, offering a flexible framework for 3D scene generation. + +3D Gaussian Splating Scene Representation. 3D Gaussian splatting (3DGS) [24] is a recent pioneering work for novel view generation, which represents the underlying scene as a collection of anisotropic 3D Gaussians defined by their center positions $\mu \in \mathbb{R}^3$ and 3D covariance matrices $\pmb{\Sigma}$ parameterized as: + +$$ +\boldsymbol {\Sigma} = \mathbf {R} \mathbf {S} \mathbf {S} ^ {\mathrm {T}} \mathbf {R} ^ {\mathrm {T}}. \tag {2} +$$ + +$\mathbf{R}$ denotes the rotation matrix and $\mathbf{S}$ is the scale matrix. Each 3D Gaussian is assigned a color $c$ represented with spherical harmonics (SH) coefficients, to capture the view-dependent appearance. To allow $\alpha$ -blending of splats, Gaussians are associated with an opacity value $\alpha \in \mathbb{R}$ . + +Unlike NeRF [31] which relies on volume rendering, 3DGS enables faster training and rendering through differentiable rasterization. A set of 3D Gaussians is rendered by projecting into the camera's image plane as 2D Gaussians, and assigned to individual image tiles. The color of each pixel $\mathbf{p}$ on the image plane is then determined as follows: + +$$ +C (\mathbf {p}) = \sum_ {i \in \mathcal {N}} c _ {i} \sigma_ {i} \prod_ {j = 1} ^ {i - 1} (1 - \sigma_ {j}), \quad \sigma_ {i} = \alpha_ {i} e ^ {- \frac {1}{2} \left(\mathbf {p} - \mu_ {i}\right) ^ {T} \Sigma_ {i} ^ {- 1} \left(\mathbf {p} - \mu_ {i}\right)} \tag {3} +$$ + +where $\mathcal{N}$ denotes the Gaussians in this tile, $\sigma_{i}$ represents the influence of the Gaussian on the image pixel and $\mu_{i},\Sigma_{i},c_{i},\alpha_{i}$ are the position, covariance, color and opacity of the $i$ -th Gaussian respectively. For optimization, a weighted combination of $\mathcal{L}_1$ loss and SSIM loss is employed. + +To initialize the centers of the 3D Gaussians, we utilize the mesh vertices of the retrieved objects and create a scene point cloud. This generated scene point cloud replaces the Structure-from-Motion (SfM) points used in the original work. To obtain the required training images for 3DGS optimization, we render the generated scene into RGB images $I$ . We additionally render 2D segmentation masks $\mathcal{M}$ to be used for 3D Gaussian scene editing. + +Scene Editing. For editing our 3DGS scenes based on the provided edit instruction, we build on top of Instruct-GS2GS [48] which in turn extends Instruct-NeRF2NeRF [20] for 3DGS scenes. Given the images used for training the 3DGS scene, Instruct-GS2GS updates the images individually using a diffusion model, namely Instruct-Pix2Pix [4]. The diffusion model takes three inputs: 1) an unedited conditioning image $I_0^v$ , a text instruction $c_{T}$ and the noisy version of the current render $I_i^v$ at optimization step $i$ , where $v$ denotes the viewpoint from which the images are captured. Formally, the process of updating a single image is defined as: + +$$ +I _ {i + 1} ^ {v} \leftarrow U _ {\theta} \left(I _ {i} ^ {v}, t; I _ {0} ^ {v}, c _ {T}\right), \tag {4} +$$ + +where $t$ is the noise level within the constant range $[t_{\mathrm{min}}, t_{\mathrm{max}}]$ , $U_{\theta}$ is the sampling process of DDIM [41] and $I_{i+1}^{v}$ is the edited image respecting the text instruction $c_{T}$ and the unedited conditioning image $I_{0}^{v}$ . The training is performed by editing all the training images during a dataset update at every 2.5k training iterations. + +Despite achieving general scene-level stylization, [48] lacks of support for more fine-grained object-level editing. To enable object-level edits, we use the binary masks $m_{\{o\}_{k=1}^{K}}$ . Each $m_{\{o\}_{k=1}^{K}}$ is obtained by binarizing 2D segmentation masks $\mathcal{M}$ for the set of objects to edit $\{o\}_{k=1}^{K}$ . The set $\{o\}_{k=1}^{K}$ is defined by extracting the referenced category names from the text instruction $c_{T}$ . Having the unedited conditioning image $I_0^v$ , the edited image $I_{i+1}^v$ , and the binary mask $m_{\{o\}_{k=1}^{K}}$ , we keep only the edits at the pixels of the target object set: + +$$ +I _ {i + 1} ^ {v} = m _ {\{o \} _ {k = 1} ^ {K}} \odot I _ {i + 1} ^ {v} + \left(1 - m _ {\{o \} _ {k = 1} ^ {K}}\right) \odot I _ {0} ^ {v}, \tag {5} +$$ + +where $\odot$ denotes element-wise multiplication of the image pixels. This way, the other pixels within $I_{i + 1}^{v}$ are set to their unedited versions, enabling an object-level editing of training images. While training the 3DGS scene with the edited images, the mask $m_{\{o\}_{k = 1}^K}$ is also applied for $\mathcal{L}_1$ and SSIM losses, ensuring gradient propagation only to the target objects. + +# 4 Experiments + +Datasets. To create a set of supporting exemplars for the layout generation, we employ the 3D-FRONT dataset [18], a large-scale indoor scene dataset with 18,797 rooms furnished with 3D textured objects from 3D-FUTURE [19]. Following the same pre-processing steps as [33] and excluding non-rectangular floor plans from the test set similar to [15], we obtained 3397 train, 453 validation, 423 test bedroom and 690 train, 98 validation, 53 test living room scenes. During 3D scene assembling, 3D models are retrieved from the 3D-FUTURE dataset. + +Implementation Details. We use ChatCompletions API $^3$ , which enables a series of dialogue exchanges between the user and the LLM, to query GPT-4 [1]. BlenderProc $^4$ is employed to render 250 images per assembled scene with $512 \times 512$ resolution. After relocating the center of each scene to the origin of the coordinate system, we sample camera positions on the upper hemisphere at a distance of $1.5r$ from the center where $r$ is the diagonal length of the room, with an elevation angle of 35 degrees. We use the Splatifto model from NeRFstudio [46] for 3DGS training, by optimizing each scene for a maximum of 20k iterations. For editing the 3DGS training images, we use InstructPix2Pix [4] as the diffusion model, with classifier guidance scales $s_T = [3.5, 12.5]$ for the text and $s_I = 1.5$ for the image. The dataset updates are performed every 2.5k iterations following Instruct-GS2GS [48]. The InstructGS2GS training takes approximately 15-20 minutes per scene on a single A6000, most of it spent on diffusion model sampling. The rendering speed is 101.21 fps. + +# 4.1 3D Layout Generation + +Evaluation Setup and Baselines. In this section we evaluate the first stage of our approach, therefore we only focus on assessing the generated layouts. To our knowledge, SceneTeller is the first approach that conducts conditional layout generation guided by text prompts to control the position and orientation of every object, targeting a user-friendly framework for novices. Hence, replicating an interaction mode that is natural for humans is pivotal for our framework. Based on this, our layout generation performance is compared with the performance of humans solving the same task. In particular, each participant of our user study is presented with a series of textual descriptions, each accompanied by an empty layout according to the room dimensions. For each text prompt, participants are requested to draw bounding boxes and label them with the object names. Each text prompt is given to at least 2, at most 3 participants. This study is conducted with 12 participants and 40 text prompts. To generate text prompts, 40 layouts are selected from the test set and their text prompts are generated by providing the initially rule-based formed sentences to GPT for paraphrasing, as described in Sec. 3.1. + +Metrics. Based on the predicted number of bounding boxes in each category, we report recall, precision and accuracy [15]. An accuracy score of 100 indicates that the number of bounding boxes and their corresponding categories precisely match the ground-truth annotations. Otherwise, a score of 0 is assigned to the scene. All values are in percentage (\%). We additionally report mean Intersection over Union (mIoU) between the predicted and ground-truth bounding boxes and out-of-bound rates, indicating the percentage of scenes where furniture extends beyond the floor plan boundary. We refer to the supplementary material for detailed metric descriptions. + +Results. Tab. 1 presents the quantitative evaluation results of 3D layout generation. Our in-context-learning-based 3D layout generation module performs better than human participants in terms of mIoU, showcasing the effectiveness of LLMs' visual reasoning capabilities. In-context exemplars further enhance this performance by aiding in reasoning about the current data distribution. However, human participants possess their own understanding of the physical world, leading to variations in their notions of object sizes and prompt reasoning, resulting in slightly poorer performance in this assessment. Both our layout generation module and human participants performed excellently in terms of predicting the number of bounding boxes per category. Humans sometimes forget or miss objects mentioned in the prompt, which explains their slightly imperfect recall score. The ground-truth layouts from the dataset involve objects slightly exceeding the floor plan boundary, leading to $15.15\%$ OOB rate. Human participants are able to place the objects within the layout more accurately. + +Table 1: Quantitative evaluation of 3D layout generation. The proposed in-context learning based 3D layout generation module performs better than human participants in terms of predicting object sizes, enhanced by in-context exemplars aiding in reasoning about the current data distribution. All values are in percentage (\%). + +
MethodOOB ↓rec.↑prec.↑acc.↑mIoU ↑
Ours56.78100.00100.00100.0044.60
Human12.0998.79100.0095.0028.70
GT15.15100.00100.00100.00100.00
+ +# 4.2 3D Scene Generation + +Baselines. Our approach is evaluated and compared with three state-of-the-art 3D scene generation methods. Set-the-scene [11] employs compositional NeRFs that are initialized based on user-defined object proxies to have more control over individual objects. To compare with [11], layouts generated by our method are provided as input to their pipeline. GSGEN [9] is a 3DGS-based approach, + +following a two-stage pipeline for geometry optimization and appearance refinement. LucidDreamer [10] employs render-refine-repeat paradigm by iteratively warping and inpainting the previously generated image to create an initial point cloud, which is then used for 3DGS scene training. + +a t a o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o . + +![](images/b9cd48f6a5141197abf01aa23d3a76c3976289e99000f006b0869a0e967393be.jpg) + +![](images/9d0c2dd39d0b81c0c2b2daebe98265ee63c4fa095e9b32532f403be633440f50.jpg) + +![](images/e02baf7723bf937babf56b54897fdf4786d174d2548fa5781e7438366857e00d.jpg) + +![](images/58e6f5f6c1e8785094486142bd3c64da1e2509215419e29e01e0a9e61db77e42.jpg) + +![](images/05f4520687f5ef3a8235ed2a3aaf0236035091f74f4f2615b4b2b171f282c844.jpg) + +"The dining table is situated in the middle of the room, but slightly leaning towards the lower left wall, with its orientation reversed and perpendicular. One dining chair is positioned near the lower left corner, with its orientation being perpendicular. Another dining chair is placed in the middle of the left wall, a bit nearer to the lower wall, and is also perpendicular. Yet another dining chair is set up at the midpoint of the left wall, but it's slightly more towards the upper wall and is perpendicular. One dining chair is near the lower right corner, with its orientation reversed and perpendicular. Another dining chair can be found at the center of the right wall, a bit closer to the lower wall, and it is perpendicular but facing the opposite direction. Another dining chair is at the center of the right wall, but slightly more towards the upper wall and its orientation is perpendicular but reversed. One more dining chair is in the middle of the upper wall, closer to the left wall, and is oriented backwards. Finally, a dining chair is at the center of the lower wall, a bit closer to the left wall, and it is oriented without any rotation. Make it marble." + +![](images/ef04450c06340b11e91bce98c9b579e3ab4f864fdf71d49f6290099d33b27bec.jpg) + +![](images/c3702bd7776a7f92ee6b846a80d566ccead06a59b44dee2cbf494d1e261bcc5a.jpg) + +![](images/af387969c8cf0fadb7fe6d4d7e95acaee26d77881a2a7ee371bd4d0f770bd79e.jpg) + +![](images/116e26e5cb13ab5e30d577d57d78bb755351d6d1394804aa0592c68e425942f2.jpg) + +![](images/2739be99be80a1abbf9ed44d7bbc1f87d93a69f75b9cbeabd1632c379a21473c.jpg) + +The coffe table is sitd in the middle of the upper wall, albeit a tad nearer to the right wall, and is aligned perpendicularly facing the opposite direction. Another coffe table can be found in the middle of the right wall, a bit nearer to the lower wall, and is facing towards the rear. Nestled in the upper right corner is another coffe table, positioned without any rotation. In the upper left corner, a tv stand is positioned perpendicularly. A multi seat sofa is positioned near the upper right corner, facing perpendicularly but in the reverse direction. Close to the lower left corner, a dining table is situated without any rotation. A dinng chair can be found in the center of the lower wall, slightly nearer to the left wall, and is facing in its original direction. Another dinng chair is positioned near the lower left corner, also without any rotation. A third dinng chair is situated in the room's center, closer to the lower left wall, and oriented towards the rear. A fourth dinng chair is positioned in the middle of the left wall, slightly closer to the lower wall, positioned towards the rear. Lastly, a console table can be found in the lower left corner, placed without any rotation. Make it pink themed." + +![](images/6e046d146f2c454a993350cdc2a17ceab9fd6d15af5809dfd2a5d7735c209aad.jpg) + +![](images/2e6a1815d743a8054c7934410e242013ea5396f686eaaff61a24b4c332d6fed9.jpg) + +![](images/50c77c19e1fea3fdfe1bb02c96b14a416ec4d2e4ca561bb7cd745ecb96775de8.jpg) + +![](images/6dbf722a40460f2451c1349b8d2eb6e4c4f1f5294aa73490309ba7674e860a72.jpg) + +![](images/004c5817ad26d584c0a3b933b66ce03264f5e7c00538d3ab224c763d53a589da.jpg) + +At the heart of the room, a double bed is positioned a tad nearer to the right wall, its orientation is perpendicular but in reverse. Adjacent to the top-right corner, a nightstand is placed, its orientation is also perpendicular but in reverse. A wardrobe can be found centrally aligned with the bottom wall, subtly closer to the right wall, with no rotation in its orientation. Make it Starry Night Van Gogh painting style." + +![](images/5f26f22cc9e384574c1e4245a3895ea4bc62bf6972676f5f5ebf98209ec70fff.jpg) + +![](images/43678837dad34e5db21bc61d0b262e05caac0452fa28cfc63903ce38b7f30282.jpg) + +![](images/d4a797edc559cd12f4e55bb9e1d797dee62a3ffff29dfa010341c075a90e14c0.jpg) + +![](images/2ff0b3f0469ae4855ab16c84fba910428c6d002b72588f5934ee731a5d2555b7.jpg) + +![](images/6938169294f0733a5742f89ded17979078fd393da12f0c8aa4e1e4e3a659a0c3.jpg) + +In the middle of the room, a double bed is placed, slightly more towards the bottom right wall, with its orientation being perpendicular and facing the reverse direction. A wardrobe is situated at the center of the top wall, a bit closer to the right wall, with its orientation facing backwards. Near the top right corner of the room, a table is situated, oriented perpendicularly but in the reverse direction. At the bottom left corner, a cabinet is positioned, perpendicular to the walls. Another table is situated near the bottom right corner, oriented perpendicularly but facing the opposite direction. Turn the double bed into golden." + +![](images/812c1438c05f42d0ad6ef9153d9033e66cc7255a764ac85539cdda8666a19b1a.jpg) +GSGEN + +![](images/25193ae74de29f3d1da45e292352c7ef1d77361dfa4ae88caa2eba43d39bb9e9.jpg) +Set-the-scene + +![](images/576bb80a86cbf15a1999f877aa76aada3915f2cc2a0955e86cf85d51238bd8fe.jpg) +LucidDreamer + +![](images/9a97b20e054b0270228105f90ac4d371f5c734e2e58b09a39e22c1b6a998688b.jpg) +Fig. 3: Qualitative comparison with state-of-the-art text-to-3D scene generation methods. The top/upper walls are marked with arrows for your reference, if this information is available within the generations. SceneTeller is able to generate high-quality scenes, with superior geometric fidelity and 3D consistency. + +![](images/e6e4b7a01a3ef283b5a2b88b38f5803ac6c27d6b38f1e7f0c2bbefc108ffb3c8.jpg) +Ours + +User Study. We conduct a user study with 30 participants to assess the performance of our method in generating high-quality text-consistent realistic 3D scenes. Human evaluators are presented with 10 scenes, each accompanied by a textual prompt describing its layout and style. For each scene, they are requested to rate it on a scale from 1 to 5 (5 being the best) based on four criteria: (a) Realism, (b) Text Alignment, (c) Geometric Fidelity and (d) Compositional Plausibility. + +We present the results of the study in Tab. 2. Our method consistently outperforms the existing state-of-the-art 3D scene generation methods across all the four criteria by a significant margin, showcasing the merits of SceneTeller in generating realistic high-quality 3D scenes solely from text prompts. For additional details of the study and extensive analysis of the results, we refer to the supplementary material. + +Table 2: User study results. Study conducted with 30 participants to assess the performance of state-of-the-art methods in generating 3D scenes based on four criteria with a scale from 1 to 5 (5 being the best and vice versa). Our method consistently outperforms the existing state-of-the-art 3D scene generation methods across all the four criteria by a significant margin, showcasing the merits of SceneTeller in generating realistic high-quality 3D scenes solely from text prompts. + +
MethodInput TypeUser Study
RealismText AlignmentGeometric FidelityCompositional Plausibility
GSGEN [9]T1.3131.3751.4391.392
LucidDreamer [10]T3.5432.7513.0503.116
Set-the-scene [11]L + T1.9631.9002.1122.181
OursT4.1544.1894.4794.323
+ +Quantitative Evaluation of Stylization. To evaluate the consistency of the generated scene style with textual descriptions, we utilize the CLIP Score [21] as the metric. Given that the CLIP score cannot reason about the parts of the prompts describing the layout, we only provide style information to CLIP in the format: 'A Barbie style bedroom' or 'A bedroom with Barbie style nightstands' for all methods. As shown in Tab. 3, our method consistently generates scenes adhering to the given style description. + +Qualitative Evaluation. Visual comparison of SceneTeller with state-of-the-art text-to-3D methods are provided in Fig. 3. SceneTeller yields high-quality 3D scene generations, surpassing the baselines in terms of text-alignment and 3D consistency by predicting an initial 3D layout conditioned on the textual descriptions. Unlike some of the state-of-the-art baselines producing unrealistic + +Table 3: Quantitative comparison of stylization. We report the CLIP Score to evaluate the consistency of the generated scene style with textual descriptions. Our method consistently generates scenes adhering to the given style description. + +
GSGEN [9]LucidDreamer [10]Set-the-scene [11]Ours
CLIP Score ↑26.47830.71527.18830.923
+ +appearance (e.g. comic or cartoon-like characteristics), SceneTeller is able to produce more realistic object shapes and appearances. + +Realism of the Generated Scenes. We evaluate all methods using FID score with a real indoor dataset ScanNet [12] on 10 generated scenes by rendering 50 images from each of them. As shown in Tab. 4, our method consistently generates more realistic scenes. + +Table 4: FID score with the ScanNet dataset [12]. + +
GSGEN [9] LucidDreamer [10] Set-the-scene [11]Ours
FID ↓297.10238.75297.57167.27
+ +Applicability to Real-world. While current design tools enable professionals to create floor plans, visualize 3D models, and render photo-realistic images, they are not suitable for novice users unfamiliar with such software tools. By enabling users to interact solely through plain natural language prompts, our approach also empowers novice users to flexibly design their own rooms. + +# 4.3 Ablation Study + +Effect of design choices for 3D layout generation. To assess the effectiveness of individual components in our design, we conduct an ablation study. Specifically, we examine two aspects: 1) the prompt source for in-context exemplars, and 2) the technique for selecting in-context exemplars. Regarding the prompt source, we present exemplars where the textual description is generated either solely by the rule-based formulation (R) or by a combination of the rule-based formulation and GPT paraphrasing $(\mathrm{R} + \mathrm{GPT})$ to introduce linguistic variation as described in Sec. 3.1. For the exemplar selection technique, we experiment with a random and a retrieval-based exemplar selection following [15], $k = 8$ exemplars. Additionally, we experiment with a third set of exemplars, denoted as $pos + neg$ , where 4 exemplars are selected by retrieval, and the other 4 are chosen from negative exemplars involving out-of-bound (OOB) + +or overlapping objects. The negative ones are from the filtered samples during the pre-processing of the dataset, and the LLM is specifically instructed to avoid them. Tab. 5 verifies that providing in-context exemplars with linguistic variation $(\mathrm{R} + \mathrm{GPT})$ selected with a retrieval-based approach performs the best, especially in terms of OOB rate by constraining the in-context learning of LLM to similar layouts. Using the $pos + neg$ set also considerably improves the performance in terms of mIoU. + +Table 5: Effect of design choices for 3D layout generation. We report the 3D layout generation performance w.r.t different design choices such as the prompt source and the exemplar selection. Providing in-context exemplars to LLM with linguistic variation (R+GPT) selected with a retrieval-based approach, performs the best. + +
Prompt sourceExemplar SelectionOOB(%)Numerical ReasoningBbox
random retrieval pos+neg↓rec.↑prec.↑acc.↑
1R✓74.12100.00100.00100.0031.10
2R✓64.87100.00100.00100.0045.50
3R74.10100.00100.00100.0041.90
4R + GPT✓76.1199.1099.1099.1029.90
5R + GPT✓54.65100.00100.00100.0044.20
6R + GPT66.45100.00100.00100.0043.90
+ +# 5 Conclusion + +In this paper, we presented a novel pipeline for 3D room generation from a textual description. Our approach supports style editing either on the individual object or on the entire scene level. Rooms generated by our method were preferred by human evaluators over the baseline results almost in all cases, confirming its efficacy on this challenging task. Importantly, our approach can be used by novices, thus making a step towards democratizing 3D scene design. + +# Acknowledgements + +We thank members of the Bosch-UvA Delta Lab and anonymous reviewers for helpful discussions and feedback. This project was generously supported by the Bosch Center for Artificial Intelligence. + +# References + +1. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023) + +2. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems (NeurIPS) 35, 23716-23736 (2022) +3. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (CVPR). pp. 5855-5864 (2021) +4. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18392-18402 (2023) +5. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. In: European Conference on Computer Vision (ECCV) (2022) +6. Chen, R., Chen, Y., Jiao, N., Jia, K.: Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (October 2023) +7. Chen, X., Zhang, Q., Li, X., Chen, Y., Feng, Y., Wang, X., Wang, J.: Hallucinated neural radiance fields in the wild. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12943-12952 (2022) +8. Chen, Y., Chen, Z., Zhang, C., Wang, F., Yang, X., Wang, Y., Cai, Z., Yang, L., Liu, H., Lin, G.: Gaussianeditor: Swift and controllable 3d editing with gaussian splatting. arXiv preprint arXiv:2311.14521 (2023) +9. Chen, Z., Wang, F., Liu, H.: Text-to-3d using gaussian splatting. arXiv preprint arXiv:2309.16585 (2023) +10. Chung, J., Lee, S., Nam, H., Lee, J., Lee, K.M.: Luciddreamer: Domain-free generation of 3d gaussian splatting scenes. arXiv preprint arXiv:2311.13384 (2023) +11. Cohen-Bar, D., Richardson, E., Metzer, G., Giryes, R., Cohen-Or, D.: Set-the-scene: Global-local training for generating controllable nerf scenes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops. pp. 2920–2929 (October 2023) +12. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proc. Computer Vision and Pattern Recognition (CVPR), IEEE (2017) +13. Fang, J., Wang, J., Zhang, X., Xie, L., Tian, Q.: Gaussianeditor: Editing 3d gaussians delicately with text instructions. arXiv preprint arXiv:2311.16037 (2023) +14. Fang, J., Wang, J., Zhang, X., Xie, L., Tian, Q.: Gaussianeditor: Editing 3d gaussians delicately with text instructions. In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024) +15. Feng, W., Zhu, W., Fu, T.j., Jampani, V., Akula, A., He, X., Basu, S., Wang, X.E., Wang, W.Y.: Layoutgpt: Compositional visual planning and generation with large language models. Advances in Neural Information Processing Systems (NeurIPS) 36 (2024) +16. Fridman, R., Abecasis, A., Kasten, Y., Dekel, T.: Scenescape: Text-driven consistent scene generation. Advances in Neural Information Processing Systems (NeurIPS) 36 (2024) +17. Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5501-5510 (2022) + +18. Fu, H., Cai, B., Gao, L., Zhang, L.X., Wang, J., Li, C., Zeng, Q., Sun, C., Jia, R., Zhao, B., et al.: 3d-front: 3d furnished rooms with layouts and semantics. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10933-10942 (2021) +19. Fu, H., Jia, R., Gao, L., Gong, M., Zhao, B., Maybank, S., Tao, D.: 3d-future: 3d furniture shape with texture. International Journal of Computer Vision 129, 3313-3337 (2021) +20. Haque, A., Tancik, M., Efros, A., Holynski, A., Kanazawa, A.: Instruct-nerf2nerf: Editing 3d scenes with instructions. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (CVPR) (2023) +21. Hessel, J., Holtzman, A., Forbes, M., Le Bras, R., Choi, Y.: CLIPScore: A reference-free evaluation metric for image captioning. In: Moens, M.F., Huang, X., Specia, L., Yih, S.W.t. (eds.) Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. pp. 7514-7528. Association for Computational Linguistics, Online and Punta Cana, Dominican Republic (Nov 2021). https://doi.org/10.18653/v1/2021.emnlp-main.595, https://aclanthology.org/2021.emnlp-main.595 +22. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020) +23. Hollein, L., Cao, A., Owens, A., Johnson, J., Nießner, M.: Text2room: Extracting textured 3d meshes from 2d text-to-image models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 7909-7920 (October 2023) +24. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (July 2023), https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/ +25. Li, Q., Li, F., Guo, J., Guo, Y.: Uhdnerf: Ultra-high-definition neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (CVPR). pp. 23097-23108 (2023) +26. Li, X., Wang, H., Tseng, K.K.: Gaussiandiffusion: 3d gaussian splatting for denoising diffusion probabilistic models with structured noise. arXiv preprint arXiv:2311.11221 (2023) +27. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3d: High-resolution text-to-3d content creation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 300-309 (2023) +28. Lin, Y., Bai, H., Li, S., Lu, H., Lin, X., Xiong, H., Wang, L.: Componerf: Text-guided multi-object compositional nerf with editable 3d scene layout. arXiv preprint arXiv:2303.13843 (2023) +29. Lorraine, J., Xie, K., Zeng, X., Lin, C.H., Takikawa, T., Sharp, N., Lin, T.Y., Liu, M.Y., Fidler, S., Lucas, J.: Att3d: Amortized text-to-3d object synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 17946-17956 (October 2023) +30. Metzer, G., Richardson, E., Patashnik, O., Giryes, R., Cohen-Or, D.: Latent-nerf for shape-guided generation of 3d shapes and textures. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12663-12673 (2023) +31. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: Proceedings of the European Conference on Computer Vision (ECCV) (2020) + +32. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022) +33. Paschalidou, D., Kar, A., Shugrina, M., Kreis, K., Geiger, A., Fidler, S.: Atiss: Autoregressive transformers for indoor scene synthesis. Advances in Neural Information Processing Systems 34, 12013-12026 (2021) +34. Phung, Q., Ge, S., Huang, J.B.: Grounded text-to-image synthesis with attention refocusing. arXiv preprint arXiv:2306.05427 (2023) +35. Po, R., Wetzstein, G.: Compositional 3d scene generation using locally conditioned diffusion. ArXiv abs/2303.12218 (2023), https://api-semanticscholar.org/CorpusID:257663283 +36. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. In: The Eleventh International Conference on Learning Representations (ICLR) (2023), https://openreview.net/forum?id=FjNys5c7VvY +37. Qian, G., Mai, J., Hamdi, A., Ren, J., Siarohin, A., Li, B., Lee, H.Y., Skorokhodov, I., Wonka, P., Tulyakov, S., Ghanem, B.: Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. In: The Twelfth International Conference on Learning Representations (ICLR) (2024), https://openreview.net/forum?id=0jHkUDyEO9 +38. Raj, A., Kaza, S., Poole, B., Niemeyer, M., Mildenhall, B., Ruiz, N., Zada, S., Aberman, K., Rubenstein, M., Barron, J., Li, Y., Jampani, V.: Dreambooth3d: Subject-driven text-to-3d generation. The International Conference on Computer Vision (ICCV) (2023) +39. Reiser, C., Szeliski, R., Verbin, D., Srinivasan, P., Mildenhall, B., Geiger, A., Barron, J., Hedman, P.: Merf: Memory-efficient radiance fields for real-time view synthesis in unbounded scenes. ACM Transactions on Graphics (TOG) 42(4), 1-12 (2023) +40. Ren, J., He, C., Liu, L., Chen, J., Wang, Y., Song, Y., Li, J., Xue, T., Hu, S., Chen, T., Zheng, K., Xiang, J., Bo, L.: Make-a-character: High quality text-to-3d character generation within minutes. arXiv preprint arXiv:2312.15430 (2023) +41. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022) +42. Sanh, V., Webson, A., Raffel, C., Bach, S., Sutawika, L., Ålyafeai, Z., Chaffin, Å., Stiegler, A., Raja, A., Dey, M., Bari, M.S., Xu, C., Thakker, U., Sharma, S.S., Szczechla, E., Kim, T., Chhablani, G., Nayak, N., Datta, D., Chang, J., Jiang, M.T.J., Wang, H., Manica, M., Shen, S., Yong, Z.X., Pandey, H., Bawden, R., Wang, T., Neeraj, T., Rozen, J., Sharma, A., Santilli, A., Fevry, T., Fries, J.A., Teehan, R., Scao, T.L., Biderman, S., Gao, L., Wolf, T., Rush, A.M.: Multitask prompted training enables zero-shot task generalization. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=9Vrb9DOWI4 +43. Sun, C., Sun, M., Chen, H.T.: Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5459-5469 (2022) +44. Sun, C., Han, J., Deng, W., Wang, X., Qin, Z., Gould, S.: 3d-gpt: Procedural 3d modeling with large language models. arXiv preprint arXiv:2310.12945 (2023) +45. Tancik, M., Casser, V., Yan, X., Pradhan, S., Mildenhall, B., Srinivasan, P.P., Barron, J.T., Kretzschmar, H.: Block-nerf: Scalable large scene neural view synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8248-8258 (2022) + +46. Tancik, M., Weber, E., Ng, E., Li, R., Yi, B., Kerr, J., Wang, T., Kristoffersen, A., Austin, J., Salahi, K., Ahuja, A., McAllister, D., Kanazawa, A.: Nerfstudio: A modular framework for neural radiance field development. In: ACM SIGGRAPH 2023 Conference Proceedings. SIGGRAPH '23 (2023) +47. Tang, J., Ren, J., Zhou, H., Liu, Z., Zeng, G.: Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653 (2023) +48. Vachha, C., Haque, A.: Instruct-gs2gs: Editing 3d gaussian splats with instructions (2024), https://instruct-gs2gs.github.io/ +49. Wang, H., Du, X., Li, J., Yeh, R.A., Shakhnarovich, G.: Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 12619-12629 (2023) +50. Wang, Z., Lu, C., Wang, Y., Bao, F., Li, C., Su, H., Zhu, J.: Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. In: Advances in Neural Information Processing Systems (NeurIPS) (2023) +51. Wei, J., Bosma, M., Zhao, V., Guu, K., Yu, A.W., Lester, B., Du, N., Dai, A.M., Le, Q.V.: Finetuned language models are zero-shot learners. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=gEZrGcozdqR +52. Xu, Y., Chai, M., Shi, Z., Peng, S., Skorokhodov, I., Siarohin, A., Yang, C., Shen, Y., Lee, H.Y., Zhou, B., et al.: Discoscope: Spatially disentangled generative radiance fields for controllable 3d-aware scene synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4402-4412 (2023) +53. Yang, Z., Gan, Z., Wang, J., Hu, X., Lu, Y., Liu, Z., Wang, L.: An empirical study of gpt-3 for few-shot knowledge-based vqa. In: Proceedings of the AAAI Conference on Artificial Intelligence (AAAI). vol. 36, pp. 3081-3089 (2022) +54. Yi, T., Fang, J., Wu, G., Xie, L., Zhang, X., Liu, W., Tian, Q., Wang, X.: Gaussian dreamer: Fast generation from text to 3d gaussian splatting with point cloud priors. arXiv preprint arXiv:2310.08529 (2023) +55. Yu, A., Li, R., Tancik, M., Li, H., Ng, R., Kanazawa, A.: Plenoptrees for real-time rendering of neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (CVPR). pp. 5752-5761 (2021) +56. Zhou, X., Ran, X., Xiong, Y., He, J., Lin, Z., Wang, Y., Sun, D., Yang, M.H.: Gala3d: Towards text-to-3d complex scene generation via layout-guided generative gaussian splatting. arXiv preprint arXiv:2402.07207 (2024) +57. Zhu, J., Zhuang, P., Koyejo, S.: HIFA: High-fidelity text-to-3d generation with advanced diffusion guidance. In: The Twelfth International Conference on Learning Representations (ICLR) (2024), https://openreview.net/forum?id=IZMPWmcS3H \ No newline at end of file diff --git a/2024/SceneTeller_ Language-to-3D Scene Generation/images.zip b/2024/SceneTeller_ Language-to-3D Scene Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..61a2dabf3ee0a9a0ec81ff8e4a1c505413c1dece --- /dev/null +++ b/2024/SceneTeller_ Language-to-3D Scene Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:66b286b6dff049752f055b68fa452ebef3995ce940947b02c17bf0d3b398866c +size 431526 diff --git a/2024/SceneTeller_ Language-to-3D Scene Generation/layout.json b/2024/SceneTeller_ Language-to-3D Scene Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9793509ae3cf30a9e8799822c04188e98be008a5 --- /dev/null +++ b/2024/SceneTeller_ Language-to-3D Scene Generation/layout.json @@ -0,0 +1,9840 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 143, + 112, + 471, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 112, + 471, + 129 + ], + "spans": [ + { + "bbox": [ + 143, + 112, + 471, + 129 + ], + "type": "text", + "content": "SceneTeller: Language-to-3D Scene Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 133, + 152, + 479, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 152, + 479, + 166 + ], + "spans": [ + { + "bbox": [ + 133, + 152, + 479, + 166 + ], + "type": "text", + "content": "Bashak Melis Öcal1*, Maxim Tatarchenko2, Sezer Karaoglu1, and Theo Gevers1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 202, + 173, + 411, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 173, + 411, + 185 + ], + "spans": [ + { + "bbox": [ + 202, + 173, + 411, + 185 + ], + "type": "text", + "content": "1 UvA-Bosch Delta Lab, University of Amsterdam" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 214, + 185, + 397, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 185, + 397, + 196 + ], + "spans": [ + { + "bbox": [ + 214, + 185, + 397, + 196 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 214, + 185, + 397, + 196 + ], + "type": "text", + "content": " Bosch Center for AI, Robert Bosch GmbH" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 143, + 239, + 469, + 538 + ], + "blocks": [ + { + "bbox": [ + 143, + 239, + 469, + 538 + ], + "lines": [ + { + "bbox": [ + 143, + 239, + 469, + 538 + ], + "spans": [ + { + "bbox": [ + 143, + 239, + 469, + 538 + ], + "type": "image", + "image_path": "77fbd33969abdf6afa836f2b511b9d28687c1140c5661da868af46dd949cc097.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 556, + 482, + 613 + ], + "lines": [ + { + "bbox": [ + 130, + 556, + 482, + 613 + ], + "spans": [ + { + "bbox": [ + 130, + 556, + 482, + 613 + ], + "type": "text", + "content": "Fig. 1: With solely a textual prompt in natural language describing the desired spatial positions and orientations of objects within the scene, SceneTeller is able to generate realistic and high-quality 3D spaces. Additionally, by facilitating modifications to the style of the entire scene or individual objects within the scene through edit instructions, SceneTeller offers a practical and flexible framework for designing personalized rooms." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 654, + 308, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 654, + 308, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 654, + 308, + 665 + ], + "type": "text", + "content": "\\* Correspondence to " + }, + { + "bbox": [ + 134, + 654, + 308, + 665 + ], + "type": "inline_equation", + "content": "<" + }, + { + "bbox": [ + 134, + 654, + 308, + 665 + ], + "type": "text", + "content": " b.m.ocal@uva.nl>" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 160, + 117, + 455, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 117, + 455, + 282 + ], + "spans": [ + { + "bbox": [ + 160, + 117, + 455, + 282 + ], + "type": "text", + "content": "Abstract. Designing high-quality indoor 3D scenes is important in many practical applications, such as room planning or game development. Conventionally, this has been a time-consuming process which requires both artistic skill and familiarity with professional software, making it hardly accessible for layman users. However, recent advances in generative AI have established solid foundation for democratizing 3D design. In this paper, we propose a pioneering approach for text-based 3D room design. Given a prompt in natural language describing the object placement in the room, our method produces a high-quality 3D scene corresponding to it. With an additional text prompt the users can change the appearance of the entire scene or of individual objects in it. Built using in-context learning, CAD model retrieval and 3D-Gaussian-Splatting-based stylization, our turnkey pipeline produces state-of-the-art 3D scenes, while being easy to use even for novices. Our project page is available at https://sceneteller.github.io/." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 293, + 453, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 293, + 453, + 315 + ], + "spans": [ + { + "bbox": [ + 160, + 293, + 453, + 315 + ], + "type": "text", + "content": "Keywords: Text-to-3D scene generation " + }, + { + "bbox": [ + 160, + 293, + 453, + 315 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 293, + 453, + 315 + ], + "type": "text", + "content": " Language-guided generation " + }, + { + "bbox": [ + 160, + 293, + 453, + 315 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 293, + 453, + 315 + ], + "type": "text", + "content": " 3D Gaussian splatting scene stylization" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 338, + 230, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 338, + 230, + 350 + ], + "spans": [ + { + "bbox": [ + 132, + 338, + 230, + 350 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 365, + 482, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 461 + ], + "type": "text", + "content": "3D scene design has numerous applications in domains like architecture or game development. Designing high-quality 3D scenes, however, is a laborious and time-consuming process, challenging even for highly-skilled domain experts utilizing professional tools. This complexity often leaves novice users unable to customize their own spaces. Recent advancements in automatic text-to-image synthesis [22,41] have led to massive progress in text-to-3D creation [36-38], enabling the generation of 3D content only from text input. Such works create foundation for democratizing 3D design, supporting its accessibility for non-experts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 461, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 568 + ], + "type": "text", + "content": "A pioneering work by Poole et al. [36] introduced object-centric text-to-3D generation based on 2D diffusion models coupled with NeRFs [31]. Several subsequent works proposed ways to improve the geometric accuracy, multi-view consistency, and appearance quality [6,27,30,50]. Despite their success in object-level generation, hallucinating complex large-scale scenes with multiple objects is still remaining a challenging problem. Some recent 3D scene generation methods [11,35,52] require a 3D layout as input, limiting their usability for novices. Others [16,23] produce results that violate global object arrangements and are not multi-view consistent." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Aiming to address both these shortcomings - insufficient user friendliness and unsatisfactory global consistency - we introduce SceneTeller, a framework for text-based 3D scene generation. 1) We start with a premise that a 3D scene generation framework should offer control over the placement and style of individual objects, allowing novice users to design their customized spaces through simple interactions with the system. Our approach achieves this by relying on text-based user interaction for scene and appearance specification. 2) In order to generate compositionally plausible 3D scenes, it is important to incorporate" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 90, + 233, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 90, + 233, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 90, + 233, + 101 + ], + "type": "text", + "content": "B.M. Öcal et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "some degree of global information guidance. We implement this by conditioning the method on global layout descriptions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 481, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 481, + 259 + ], + "type": "text", + "content": "To this end, we split the task of generating high-quality 3D scenes into three steps: 3D layout generation, object population and style editing. 1) When presented with a prompt describing the desired spatial positions and orientations of objects, SceneTeller generates an initial 3D layout of the scene by exploiting the Large Language Models (LLMs) through in-context learning, to ensure global consistency. 2) The generated layout is then used to assemble a 3D scene and populate it with CAD models from a database. 3) To facilitate scene editing, the assembled scene is fitted with a 3D Gaussian Splitting [24] representation. This representation is then used to stylize the scene according to the user-provided edit prompt targeting the entire scene or individual objects within it." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 260, + 481, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 260, + 481, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 260, + 481, + 308 + ], + "type": "text", + "content": "Extensive quantitative and qualitative evaluations show that SceneTeller performs significantly better than state-of-the-art methods, offering superior geometric fidelity and compositional plausibility. In brief, our contributions to the existing body of work can be summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 313, + 480, + 431 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 313, + 479, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 313, + 479, + 360 + ], + "spans": [ + { + "bbox": [ + 138, + 313, + 479, + 360 + ], + "type": "text", + "content": "- We introduce a turnkey pipeline for high-quality 3D scene generation yielding complex, realistic and globally consistent 3D scenes, given only scene descriptions in natural language. Our pipeline supports object- and scene-level appearance editing." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 361, + 479, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 361, + 479, + 396 + ], + "spans": [ + { + "bbox": [ + 138, + 361, + 479, + 396 + ], + "type": "text", + "content": "- We propose a novel LLM-based module for 3D layout generation trained through in-context learning, which provides control over individual object placement." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 396, + 480, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 396, + 480, + 431 + ], + "spans": [ + { + "bbox": [ + 138, + 396, + 480, + 431 + ], + "type": "text", + "content": "- We experimentally demonstrate that SceneTeller generates 3D scenes of higher quality and better consistency than state-of-the-art methods, while boasting unprecedented user friendliness and flexibility." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 131, + 448, + 237, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 448, + 237, + 460 + ], + "spans": [ + { + "bbox": [ + 131, + 448, + 237, + 460 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 472, + 292, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 472, + 292, + 483 + ], + "spans": [ + { + "bbox": [ + 131, + 472, + 292, + 483 + ], + "type": "text", + "content": "2.1 3D Scene Representations" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 488, + 481, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 488, + 481, + 596 + ], + "spans": [ + { + "bbox": [ + 130, + 488, + 481, + 596 + ], + "type": "text", + "content": "Rendering 3D scenes using implicit neural networks has seen remarkable success, notably with NeRF's [31] photo-realistic image synthesis. Extensions of NeRF improve reconstruction quality [3, 25], training and rendering speed [5, 17, 32, 39, 43, 55], and handling large-scale scenes [7, 45]. Despite advancements, NeRF-based methods still struggle with slow rendering speeds. Recently, 3D Gaussian splatting (3DGS) was proposed for real-time, high-quality rendering using anisotropic 3D Gaussians and fast tile-based rasterization. Our framework employs 3DGS for scene representation and allows edits on these Gaussians, generating high-quality, 3D consistent scenes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 612, + 276, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 612, + 276, + 623 + ], + "spans": [ + { + "bbox": [ + 131, + 612, + 276, + 623 + ], + "type": "text", + "content": "2.2 Text-to-3D Generation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "Object-centric Generation. 3D generation from textual prompts has recently grown in popularity. DreamFusion [36] introduces Score Distillation Sampling (SDS) to optimize NeRF representations by leveraging an off-the-shelf" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 256, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 256, + 91, + 447, + 102 + ], + "type": "text", + "content": "SceneTeller: Language-to-3D Scene Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": "2D diffusion model, paving the way towards text-driven 3D content creation [29,37,38,57]. Following [36], Magic3D [27] employs a coarse-to-fine optimization scheme, exploiting low-resolution diffusion priors, and then enhancing the texture. To achieve high-fidelity mesh generation, Fantasia3D [6] disentangles the appearance from geometry. Latent-NeRF [30] optimizes a NeRF in the latent space of StableDiffusion [41], and additionally proposes incorporating 3D priors to generate a target geometry. To address the over-smoothing, over-saturation and lack of diversity problems within the SDS, ProfilicDreamer [50] proposes a principled particle-based variational framework, namely Variational Score Distillation (VSD). Following the same motivation, [49] presents Perturb-and-Average Scoring to estimate the score for non-noisy images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 248, + 483, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 483, + 368 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 483, + 368 + ], + "type": "text", + "content": "With the recent introduction of 3D Gaussian Splatting [24], several works attempt to combine 3D Gs and diffusion models. GaussianDreamer [54] employs a 3D diffusion model to generate the initialized point clouds for 3D Gaussians, and then further optimize the Gaussians using SDS with a 2D diffusion model. [9, 47] follows a two-stage pipeline for geometry optimization and appearance refinement. GaussianDiffusion [26] focuses on the challenge of achieving multiview consistency in 3D generation and proposes combining structured noise with a variational 3DGS technique. Despite their success in generating object-level scenes, both NeRF-based and 3DGS-based methods fall short in hallucinating complex scenes with multiple objects." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 382, + 483, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 382, + 483, + 586 + ], + "spans": [ + { + "bbox": [ + 130, + 382, + 483, + 586 + ], + "type": "text", + "content": "Scene Generation. To represent multiple objects within the scenes, existing works [11, 28, 35, 52] rely on compositional NeRFs that are initialized based on user-defined object proxies. However, they are usually restricted to produce only narrow inward-facing views of the rooms. Also, the requirement of having 3D layouts as input is restrictive for final users. Orthogonal to those, SceneScape [16] and Text2Room [23] generate scenes following the render-refine-repeat paradigm by iteratively warping and inpainting the previously generated image according to the specified camera motion. This strategy does not allow object-level control during scene generation, and the results tend to exhibit multi-view inconsistencies. LucidDreamer [10] employs the same paradigm for an initial point cloud reconstruction, and then further optimizes the scene using 3DGS. Recently, several works introduced techniques for 3DGS scene editing, by manipulating the training images using 2D diffusion models [8, 13, 48] and incorporating the edited images into 3DGS scenes for further optimization. Concurrent to our work, [56] initializes a 3DGS scene from the layouts created using LLMs and optimizes it using a diffusion model. However, [56] lacks control over the generated layout in terms of positions and orientations of objects." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 600, + 244, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 600, + 244, + 612 + ], + "spans": [ + { + "bbox": [ + 131, + 600, + 244, + 612 + ], + "type": "text", + "content": "2.3 LLMs for Vision" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "With the inclusion of grounding information in LLMs, they demonstrate enhanced reasoning capabilities for visual concepts. Recently, the visual reasoning abilities of LLMs have been utilized for 3D tasks such as avatar generation [40], instruction-driven 3D modeling [44], 3D editing [14]. Exploiting the" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 90, + 233, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 90, + 233, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 90, + 233, + 101 + ], + "type": "text", + "content": "B.M. Öcal et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "spatial reasoning capabilities of LLMs, LayoutGPT [15] extends beyond its 2D counterparts [34] to generate random 3D layouts of specific room types through in-context learning. However, LayoutGPT lacks support for controlling the generated layout. The 3D layout generation module of SceneTeller is the first of its kind allowing control over individual object placement in accordance with the provided natural language prompts describing the desired positions of objects." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 205, + 202, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 202, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 202, + 217 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 138, + 248, + 476, + 411 + ], + "blocks": [ + { + "bbox": [ + 138, + 248, + 476, + 411 + ], + "lines": [ + { + "bbox": [ + 138, + 248, + 476, + 411 + ], + "spans": [ + { + "bbox": [ + 138, + 248, + 476, + 411 + ], + "type": "image", + "image_path": "7cea33ba983928d37ab833debb0d181fe011cb0a4db0f3442d8482c6bc6131a6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 418, + 482, + 486 + ], + "lines": [ + { + "bbox": [ + 130, + 418, + 482, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 482, + 486 + ], + "type": "text", + "content": "Fig. 2: Overview of our method. Given a textual prompt in natural language delineating the desired spatial positions and orientations of objects within the scene, a 3D scene layout is generated using in-context learning. An initial 3D scene is assembled for the predicted layout, which is then fitted with a 3D Gaussian Splatting representation. This representation is then used to stylize the scene according to the user-provided text prompt, and subsequently render the final images of the scene." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 510, + 482, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 606 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 606 + ], + "type": "text", + "content": "The pipeline of SceneTeller is divided into three main stages: language-driven 3D layout generation, 3D scene assembling from the layout and 3D scene stylization. In the first stage, a 3D scene layout is generated based on the given text prompt (Sec. 3.1). Then, an initial 3D scene is assembled for the predicted layout (Sec. 3.2). After that, a 3D Gaussian Splatting representation is fitted to the assembled scene. This representation is then used to stylize the scene according to the user-provided text prompt, and subsequently render the final images of the scene (Sec. 3.3. The scheme of the framework is illustrated in Fig. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 623, + 361, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 623, + 361, + 635 + ], + "spans": [ + { + "bbox": [ + 131, + 623, + 361, + 635 + ], + "type": "text", + "content": "3.1 Language-driven 3D Layout Generation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "The task of this module is to produce a collection of object bounding boxes conditioned on a brief textural description of their spatial positions within the" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 256, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 256, + 91, + 447, + 102 + ], + "type": "text", + "content": "SceneTeller: Language-to-3D Scene Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 199 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 199 + ], + "type": "text", + "content": "scene in plain language. More formally, the input is a condition " + }, + { + "bbox": [ + 130, + 116, + 482, + 199 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 116, + 482, + 199 + ], + "type": "text", + "content": ", consisting of a textual description " + }, + { + "bbox": [ + 130, + 116, + 482, + 199 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 130, + 116, + 482, + 199 + ], + "type": "text", + "content": ", along with the scene type and scene dimensions information. The output is a 3D layout represented as a set of 3D bounding boxes " + }, + { + "bbox": [ + 130, + 116, + 482, + 199 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_i \\in B = \\{\\mathbf{b}_i\\}_{i=1}^N" + }, + { + "bbox": [ + 130, + 116, + 482, + 199 + ], + "type": "text", + "content": " for each of the " + }, + { + "bbox": [ + 130, + 116, + 482, + 199 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 116, + 482, + 199 + ], + "type": "text", + "content": " referenced objects. Each 3D bounding box " + }, + { + "bbox": [ + 130, + 116, + 482, + 199 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_i = (c_i, \\mathbf{t}_i, \\mathbf{s}_i, o_i)" + }, + { + "bbox": [ + 130, + 116, + 482, + 199 + ], + "type": "text", + "content": " consists of the following components: category name " + }, + { + "bbox": [ + 130, + 116, + 482, + 199 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 130, + 116, + 482, + 199 + ], + "type": "text", + "content": ", box center " + }, + { + "bbox": [ + 130, + 116, + 482, + 199 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_i = (x_i, y_i, z_i) \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 116, + 482, + 199 + ], + "type": "text", + "content": ", box dimensions " + }, + { + "bbox": [ + 130, + 116, + 482, + 199 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_i = (w_i, h_i, d_i) \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 116, + 482, + 199 + ], + "type": "text", + "content": " and orientation " + }, + { + "bbox": [ + 130, + 116, + 482, + 199 + ], + "type": "inline_equation", + "content": "o_i \\in \\mathbb{R}" + }, + { + "bbox": [ + 130, + 116, + 482, + 199 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 482, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 482, + 379 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 482, + 379 + ], + "type": "text", + "content": "The textual description " + }, + { + "bbox": [ + 130, + 200, + 482, + 379 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 130, + 200, + 482, + 379 + ], + "type": "text", + "content": " is comprised of multiple sentences, each specifying the position and orientation of a single object within a canonical coordinate system. The canonical system is chosen to represent the scene in a 2D perspective, to ensure that the descriptions closely resemble those produced by a human observer. " + }, + { + "bbox": [ + 130, + 200, + 482, + 379 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 130, + 200, + 482, + 379 + ], + "type": "text", + "content": " is initially generated using a rule-based approach formulated in three steps as follows: 1) The canonical coordinate system is partitioned into a " + }, + { + "bbox": [ + 130, + 200, + 482, + 379 + ], + "type": "inline_equation", + "content": "3\\times 3" + }, + { + "bbox": [ + 130, + 200, + 482, + 379 + ], + "type": "text", + "content": " grid, creating rectangular patches. 2) Each of the predicted objects is assigned to a patch based on the center position " + }, + { + "bbox": [ + 130, + 200, + 482, + 379 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_i" + }, + { + "bbox": [ + 130, + 200, + 482, + 379 + ], + "type": "text", + "content": " and orientation " + }, + { + "bbox": [ + 130, + 200, + 482, + 379 + ], + "type": "inline_equation", + "content": "o_i" + }, + { + "bbox": [ + 130, + 200, + 482, + 379 + ], + "type": "text", + "content": " of its bounding box within these patches. 3) Per-object sentences are then formulated according to which patch the object is assigned to and object's orientation, e.g. for the object assigned to the top-left patch, with category name " + }, + { + "bbox": [ + 130, + 200, + 482, + 379 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 130, + 200, + 482, + 379 + ], + "type": "text", + "content": " and 90 degrees orientation, we generate: \"A " + }, + { + "bbox": [ + 130, + 200, + 482, + 379 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 130, + 200, + 482, + 379 + ], + "type": "text", + "content": " is placed at the top-left corner of the room, with a perpendicular orientation.\" Given the relatively limited linguistic variation inherent in rule-based descriptions, initial textual templates are provided to an LLM to serve as a basis for paraphrasing." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 380, + 482, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 439 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 439 + ], + "type": "text", + "content": "LayoutGPT [15] focused on improving an LLM's interpretation of the spatial knowledge by using a structured format featuring attribute-value pairs. We follow a similar strategy and adopt the standard CSS (Cascading Style Sheets) format to describe " + }, + { + "bbox": [ + 130, + 380, + 482, + 439 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_i" + }, + { + "bbox": [ + 130, + 380, + 482, + 439 + ], + "type": "text", + "content": ". Bounding box values are mapped to standard CSS format attributes and category name is employed as the selector." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 456, + 482, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 456, + 482, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 456, + 482, + 540 + ], + "type": "text", + "content": "Prompt Construction. Recent progress in the field has showcased the impressive in-context learning capabilities of LLMs, which entail conditioning these models with both natural language instructions and a select set of task demonstrations. Following this conditioning, the LLM is then tasked with completing further instances based on the provided demonstrations. In line with this, our prompts to LLM include three main parts: task specifications, in-context exemplars and the query condition:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 545, + 482, + 667 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 137, + 545, + 482, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 545, + 482, + 630 + ], + "spans": [ + { + "bbox": [ + 137, + 545, + 482, + 630 + ], + "type": "text", + "content": "- Task specifications: Following prior work [42, 51] on enhancing LLMs instruction following capabilities, a task description is incorporated at the beginning of each prompt, which explains the goal of the task, establishes a standard for the 3D layout format in CSS style and provides unit information for attributes. Additionally, constraints are integrated into the prompts to guide the LLM and minimize errors (e.g., predicting overlapping boxes or placing objects out of bounds), during task completion." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 630, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 630, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 137, + 630, + 482, + 667 + ], + "type": "text", + "content": "- In-context learning: Supporting exemplars for the in-context learning are selected by adopting the retrieval-based approach used in [2, 15, 53]. When provided with a set of supporting exemplars " + }, + { + "bbox": [ + 137, + 630, + 482, + 667 + ], + "type": "inline_equation", + "content": "S = \\{(\\mathcal{C}_m^s,\\mathbf{b}_m^s)\\}_{m = 1}^M" + }, + { + "bbox": [ + 137, + 630, + 482, + 667 + ], + "type": "text", + "content": " and the" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 90, + 233, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 90, + 233, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 90, + 233, + 101 + ], + "type": "text", + "content": "B.M. Öcal et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 147, + 114, + 479, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 114, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 147, + 114, + 479, + 175 + ], + "type": "text", + "content": "queried condition " + }, + { + "bbox": [ + 147, + 114, + 479, + 175 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_q" + }, + { + "bbox": [ + 147, + 114, + 479, + 175 + ], + "type": "text", + "content": ", the function " + }, + { + "bbox": [ + 147, + 114, + 479, + 175 + ], + "type": "inline_equation", + "content": "f(\\mathcal{C}_k^s,\\mathcal{C}_q) = \\|rl_k - rl_q\\|^2 + \\|rw_k - rw_q\\|^2" + }, + { + "bbox": [ + 147, + 114, + 479, + 175 + ], + "type": "text", + "content": " is computed between each element of the set and " + }, + { + "bbox": [ + 147, + 114, + 479, + 175 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_q" + }, + { + "bbox": [ + 147, + 114, + 479, + 175 + ], + "type": "text", + "content": " following [15], where " + }, + { + "bbox": [ + 147, + 114, + 479, + 175 + ], + "type": "inline_equation", + "content": "rl" + }, + { + "bbox": [ + 147, + 114, + 479, + 175 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 147, + 114, + 479, + 175 + ], + "type": "inline_equation", + "content": "rw" + }, + { + "bbox": [ + 147, + 114, + 479, + 175 + ], + "type": "text", + "content": " are the length and width of the scenes. Top-" + }, + { + "bbox": [ + 147, + 114, + 479, + 175 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 147, + 114, + 479, + 175 + ], + "type": "text", + "content": " supporting exemplars with the shortest distance to " + }, + { + "bbox": [ + 147, + 114, + 479, + 175 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_q" + }, + { + "bbox": [ + 147, + 114, + 479, + 175 + ], + "type": "text", + "content": " are selected for in-context learning, provided to LLM with the same format with " + }, + { + "bbox": [ + 147, + 114, + 479, + 175 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_q" + }, + { + "bbox": [ + 147, + 114, + 479, + 175 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 175, + 479, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 175, + 479, + 199 + ], + "spans": [ + { + "bbox": [ + 138, + 175, + 479, + 199 + ], + "type": "text", + "content": "- Query condition: The inference condition " + }, + { + "bbox": [ + 138, + 175, + 479, + 199 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_q" + }, + { + "bbox": [ + 138, + 175, + 479, + 199 + ], + "type": "text", + "content": ", for which we want to predict the layout." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 217, + 338, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 217, + 338, + 229 + ], + "spans": [ + { + "bbox": [ + 132, + 217, + 338, + 229 + ], + "type": "text", + "content": "3.2 3D Scene Assembling from Layouts" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 237, + 479, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 479, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 479, + 285 + ], + "type": "text", + "content": "To assemble a scene based on the generated layouts, we rely on object retrieval. More specifically, for each of the bounding boxes, the nearest 3D model of the corresponding predicted category is retrieved from the furniture dataset based on the Euclidean distance of the bounding box dimensions as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 167, + 307, + 480, + 320 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 307, + 480, + 320 + ], + "spans": [ + { + "bbox": [ + 167, + 307, + 480, + 320 + ], + "type": "interline_equation", + "content": "d \\left(\\mathbf {s} _ {i}, \\mathbf {s} _ {\\text {c a n d}}\\right) = \\sqrt {\\left(w _ {i} - w _ {\\text {c a n d}}\\right) ^ {2} + \\left(h _ {i} - h _ {\\text {c a n d}}\\right) ^ {2} + \\left(d _ {i} - d _ {\\text {c a n d}}\\right) ^ {2}}, \\tag {1}", + "image_path": "d96734d103a799b85eac9563a535fa06f77b1b39950d08c07affeea5ac63aede.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 329, + 479, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 329, + 479, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 329, + 479, + 365 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 329, + 479, + 365 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{\\text{cand}}" + }, + { + "bbox": [ + 130, + 329, + 479, + 365 + ], + "type": "text", + "content": " denotes the bounding box dimensions for the candidate object from the dataset. The retrieved 3D model is placed within the scene according to the predicted box center " + }, + { + "bbox": [ + 130, + 329, + 479, + 365 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_i" + }, + { + "bbox": [ + 130, + 329, + 479, + 365 + ], + "type": "text", + "content": " and orientation angle " + }, + { + "bbox": [ + 130, + 329, + 479, + 365 + ], + "type": "inline_equation", + "content": "o_i" + }, + { + "bbox": [ + 130, + 329, + 479, + 365 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 382, + 264, + 394 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 382, + 264, + 394 + ], + "spans": [ + { + "bbox": [ + 132, + 382, + 264, + 394 + ], + "type": "text", + "content": "3.3 3D Scene Stylization" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 403, + 479, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 479, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 479, + 462 + ], + "type": "text", + "content": "Given a generated 3D scene, the users should be able to modify either the overall appearance of the scene or specific objects within it. Our approach facilitates scene editing by training a 3D Gaussian splatting model representing the generated scene and refining this representation according to the provided edit instruction, offering a flexible framework for 3D scene generation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 479, + 479, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 479, + 528 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 479, + 528 + ], + "type": "text", + "content": "3D Gaussian Splating Scene Representation. 3D Gaussian splatting (3DGS) [24] is a recent pioneering work for novel view generation, which represents the underlying scene as a collection of anisotropic 3D Gaussians defined by their center positions " + }, + { + "bbox": [ + 130, + 479, + 479, + 528 + ], + "type": "inline_equation", + "content": "\\mu \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 479, + 479, + 528 + ], + "type": "text", + "content": " and 3D covariance matrices " + }, + { + "bbox": [ + 130, + 479, + 479, + 528 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}" + }, + { + "bbox": [ + 130, + 479, + 479, + 528 + ], + "type": "text", + "content": " parameterized as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 274, + 539, + 480, + 552 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 274, + 539, + 480, + 552 + ], + "spans": [ + { + "bbox": [ + 274, + 539, + 480, + 552 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\Sigma} = \\mathbf {R} \\mathbf {S} \\mathbf {S} ^ {\\mathrm {T}} \\mathbf {R} ^ {\\mathrm {T}}. \\tag {2}", + "image_path": "57227a6d630d85d92c9309c0c373deee7efcd34fc60e6d10fb488d6048f90008.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 558, + 479, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 479, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 479, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 130, + 558, + 479, + 604 + ], + "type": "text", + "content": " denotes the rotation matrix and " + }, + { + "bbox": [ + 130, + 558, + 479, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 130, + 558, + 479, + 604 + ], + "type": "text", + "content": " is the scale matrix. Each 3D Gaussian is assigned a color " + }, + { + "bbox": [ + 130, + 558, + 479, + 604 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 558, + 479, + 604 + ], + "type": "text", + "content": " represented with spherical harmonics (SH) coefficients, to capture the view-dependent appearance. To allow " + }, + { + "bbox": [ + 130, + 558, + 479, + 604 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 558, + 479, + 604 + ], + "type": "text", + "content": "-blending of splats, Gaussians are associated with an opacity value " + }, + { + "bbox": [ + 130, + 558, + 479, + 604 + ], + "type": "inline_equation", + "content": "\\alpha \\in \\mathbb{R}" + }, + { + "bbox": [ + 130, + 558, + 479, + 604 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 605, + 479, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 479, + 664 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 479, + 664 + ], + "type": "text", + "content": "Unlike NeRF [31] which relies on volume rendering, 3DGS enables faster training and rendering through differentiable rasterization. A set of 3D Gaussians is rendered by projecting into the camera's image plane as 2D Gaussians, and assigned to individual image tiles. The color of each pixel " + }, + { + "bbox": [ + 130, + 605, + 479, + 664 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 130, + 605, + 479, + 664 + ], + "type": "text", + "content": " on the image plane is then determined as follows:" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 256, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 256, + 91, + 447, + 102 + ], + "type": "text", + "content": "SceneTeller: Language-to-3D Scene Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 177, + 126, + 480, + 159 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 177, + 126, + 480, + 159 + ], + "spans": [ + { + "bbox": [ + 177, + 126, + 480, + 159 + ], + "type": "interline_equation", + "content": "C (\\mathbf {p}) = \\sum_ {i \\in \\mathcal {N}} c _ {i} \\sigma_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\sigma_ {j}), \\quad \\sigma_ {i} = \\alpha_ {i} e ^ {- \\frac {1}{2} \\left(\\mathbf {p} - \\mu_ {i}\\right) ^ {T} \\Sigma_ {i} ^ {- 1} \\left(\\mathbf {p} - \\mu_ {i}\\right)} \\tag {3}", + "image_path": "faf5ba53bd13a8db6d0f32c417fa27654c65c2d8f8e7fe3ec2948e76296bbb15.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 163, + 479, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 163, + 479, + 211 + ], + "spans": [ + { + "bbox": [ + 130, + 163, + 479, + 211 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 163, + 479, + 211 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 130, + 163, + 479, + 211 + ], + "type": "text", + "content": " denotes the Gaussians in this tile, " + }, + { + "bbox": [ + 130, + 163, + 479, + 211 + ], + "type": "inline_equation", + "content": "\\sigma_{i}" + }, + { + "bbox": [ + 130, + 163, + 479, + 211 + ], + "type": "text", + "content": " represents the influence of the Gaussian on the image pixel and " + }, + { + "bbox": [ + 130, + 163, + 479, + 211 + ], + "type": "inline_equation", + "content": "\\mu_{i},\\Sigma_{i},c_{i},\\alpha_{i}" + }, + { + "bbox": [ + 130, + 163, + 479, + 211 + ], + "type": "text", + "content": " are the position, covariance, color and opacity of the " + }, + { + "bbox": [ + 130, + 163, + 479, + 211 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 163, + 479, + 211 + ], + "type": "text", + "content": "-th Gaussian respectively. For optimization, a weighted combination of " + }, + { + "bbox": [ + 130, + 163, + 479, + 211 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 130, + 163, + 479, + 211 + ], + "type": "text", + "content": " loss and SSIM loss is employed." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 212, + 480, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 212, + 480, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 480, + 283 + ], + "type": "text", + "content": "To initialize the centers of the 3D Gaussians, we utilize the mesh vertices of the retrieved objects and create a scene point cloud. This generated scene point cloud replaces the Structure-from-Motion (SfM) points used in the original work. To obtain the required training images for 3DGS optimization, we render the generated scene into RGB images " + }, + { + "bbox": [ + 130, + 212, + 480, + 283 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 212, + 480, + 283 + ], + "type": "text", + "content": ". We additionally render 2D segmentation masks " + }, + { + "bbox": [ + 130, + 212, + 480, + 283 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 130, + 212, + 480, + 283 + ], + "type": "text", + "content": " to be used for 3D Gaussian scene editing." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 301, + 480, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 301, + 480, + 410 + ], + "spans": [ + { + "bbox": [ + 130, + 301, + 480, + 410 + ], + "type": "text", + "content": "Scene Editing. For editing our 3DGS scenes based on the provided edit instruction, we build on top of Instruct-GS2GS [48] which in turn extends Instruct-NeRF2NeRF [20] for 3DGS scenes. Given the images used for training the 3DGS scene, Instruct-GS2GS updates the images individually using a diffusion model, namely Instruct-Pix2Pix [4]. The diffusion model takes three inputs: 1) an unedited conditioning image " + }, + { + "bbox": [ + 130, + 301, + 480, + 410 + ], + "type": "inline_equation", + "content": "I_0^v" + }, + { + "bbox": [ + 130, + 301, + 480, + 410 + ], + "type": "text", + "content": ", a text instruction " + }, + { + "bbox": [ + 130, + 301, + 480, + 410 + ], + "type": "inline_equation", + "content": "c_{T}" + }, + { + "bbox": [ + 130, + 301, + 480, + 410 + ], + "type": "text", + "content": " and the noisy version of the current render " + }, + { + "bbox": [ + 130, + 301, + 480, + 410 + ], + "type": "inline_equation", + "content": "I_i^v" + }, + { + "bbox": [ + 130, + 301, + 480, + 410 + ], + "type": "text", + "content": " at optimization step " + }, + { + "bbox": [ + 130, + 301, + 480, + 410 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 301, + 480, + 410 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 301, + 480, + 410 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 130, + 301, + 480, + 410 + ], + "type": "text", + "content": " denotes the viewpoint from which the images are captured. Formally, the process of updating a single image is defined as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 256, + 422, + 480, + 435 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 422, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 256, + 422, + 480, + 435 + ], + "type": "interline_equation", + "content": "I _ {i + 1} ^ {v} \\leftarrow U _ {\\theta} \\left(I _ {i} ^ {v}, t; I _ {0} ^ {v}, c _ {T}\\right), \\tag {4}", + "image_path": "51f6aef3a241998e0e33ad51562a6f02451595da7dc32fbc14251ceabb2477ac.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 441, + 480, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 441, + 480, + 488 + ], + "spans": [ + { + "bbox": [ + 130, + 441, + 480, + 488 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 441, + 480, + 488 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 441, + 480, + 488 + ], + "type": "text", + "content": " is the noise level within the constant range " + }, + { + "bbox": [ + 130, + 441, + 480, + 488 + ], + "type": "inline_equation", + "content": "[t_{\\mathrm{min}}, t_{\\mathrm{max}}]" + }, + { + "bbox": [ + 130, + 441, + 480, + 488 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 441, + 480, + 488 + ], + "type": "inline_equation", + "content": "U_{\\theta}" + }, + { + "bbox": [ + 130, + 441, + 480, + 488 + ], + "type": "text", + "content": " is the sampling process of DDIM [41] and " + }, + { + "bbox": [ + 130, + 441, + 480, + 488 + ], + "type": "inline_equation", + "content": "I_{i+1}^{v}" + }, + { + "bbox": [ + 130, + 441, + 480, + 488 + ], + "type": "text", + "content": " is the edited image respecting the text instruction " + }, + { + "bbox": [ + 130, + 441, + 480, + 488 + ], + "type": "inline_equation", + "content": "c_{T}" + }, + { + "bbox": [ + 130, + 441, + 480, + 488 + ], + "type": "text", + "content": " and the unedited conditioning image " + }, + { + "bbox": [ + 130, + 441, + 480, + 488 + ], + "type": "inline_equation", + "content": "I_{0}^{v}" + }, + { + "bbox": [ + 130, + 441, + 480, + 488 + ], + "type": "text", + "content": ". The training is performed by editing all the training images during a dataset update at every 2.5k training iterations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 489, + 480, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 489, + 480, + 577 + ], + "spans": [ + { + "bbox": [ + 130, + 489, + 480, + 577 + ], + "type": "text", + "content": "Despite achieving general scene-level stylization, [48] lacks of support for more fine-grained object-level editing. To enable object-level edits, we use the binary masks " + }, + { + "bbox": [ + 130, + 489, + 480, + 577 + ], + "type": "inline_equation", + "content": "m_{\\{o\\}_{k=1}^{K}}" + }, + { + "bbox": [ + 130, + 489, + 480, + 577 + ], + "type": "text", + "content": ". Each " + }, + { + "bbox": [ + 130, + 489, + 480, + 577 + ], + "type": "inline_equation", + "content": "m_{\\{o\\}_{k=1}^{K}}" + }, + { + "bbox": [ + 130, + 489, + 480, + 577 + ], + "type": "text", + "content": " is obtained by binarizing 2D segmentation masks " + }, + { + "bbox": [ + 130, + 489, + 480, + 577 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 130, + 489, + 480, + 577 + ], + "type": "text", + "content": " for the set of objects to edit " + }, + { + "bbox": [ + 130, + 489, + 480, + 577 + ], + "type": "inline_equation", + "content": "\\{o\\}_{k=1}^{K}" + }, + { + "bbox": [ + 130, + 489, + 480, + 577 + ], + "type": "text", + "content": ". The set " + }, + { + "bbox": [ + 130, + 489, + 480, + 577 + ], + "type": "inline_equation", + "content": "\\{o\\}_{k=1}^{K}" + }, + { + "bbox": [ + 130, + 489, + 480, + 577 + ], + "type": "text", + "content": " is defined by extracting the referenced category names from the text instruction " + }, + { + "bbox": [ + 130, + 489, + 480, + 577 + ], + "type": "inline_equation", + "content": "c_{T}" + }, + { + "bbox": [ + 130, + 489, + 480, + 577 + ], + "type": "text", + "content": ". Having the unedited conditioning image " + }, + { + "bbox": [ + 130, + 489, + 480, + 577 + ], + "type": "inline_equation", + "content": "I_0^v" + }, + { + "bbox": [ + 130, + 489, + 480, + 577 + ], + "type": "text", + "content": ", the edited image " + }, + { + "bbox": [ + 130, + 489, + 480, + 577 + ], + "type": "inline_equation", + "content": "I_{i+1}^v" + }, + { + "bbox": [ + 130, + 489, + 480, + 577 + ], + "type": "text", + "content": ", and the binary mask " + }, + { + "bbox": [ + 130, + 489, + 480, + 577 + ], + "type": "inline_equation", + "content": "m_{\\{o\\}_{k=1}^{K}}" + }, + { + "bbox": [ + 130, + 489, + 480, + 577 + ], + "type": "text", + "content": ", we keep only the edits at the pixels of the target object set:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 212, + 587, + 480, + 602 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 587, + 480, + 602 + ], + "spans": [ + { + "bbox": [ + 212, + 587, + 480, + 602 + ], + "type": "interline_equation", + "content": "I _ {i + 1} ^ {v} = m _ {\\{o \\} _ {k = 1} ^ {K}} \\odot I _ {i + 1} ^ {v} + \\left(1 - m _ {\\{o \\} _ {k = 1} ^ {K}}\\right) \\odot I _ {0} ^ {v}, \\tag {5}", + "image_path": "5e0a75b98e6f0ae8feb43b8d59b2c6b5ed6a3aedbf4d07498340cb5042ce60ff.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 605, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 480, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 605, + 480, + 666 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 130, + 605, + 480, + 666 + ], + "type": "text", + "content": " denotes element-wise multiplication of the image pixels. This way, the other pixels within " + }, + { + "bbox": [ + 130, + 605, + 480, + 666 + ], + "type": "inline_equation", + "content": "I_{i + 1}^{v}" + }, + { + "bbox": [ + 130, + 605, + 480, + 666 + ], + "type": "text", + "content": " are set to their unedited versions, enabling an object-level editing of training images. While training the 3DGS scene with the edited images, the mask " + }, + { + "bbox": [ + 130, + 605, + 480, + 666 + ], + "type": "inline_equation", + "content": "m_{\\{o\\}_{k = 1}^K}" + }, + { + "bbox": [ + 130, + 605, + 480, + 666 + ], + "type": "text", + "content": " is also applied for " + }, + { + "bbox": [ + 130, + 605, + 480, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 130, + 605, + 480, + 666 + ], + "type": "text", + "content": " and SSIM losses, ensuring gradient propagation only to the target objects." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 90, + 233, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 90, + 233, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 90, + 233, + 101 + ], + "type": "text", + "content": "B.M. Öcal et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "content": "Datasets. To create a set of supporting exemplars for the layout generation, we employ the 3D-FRONT dataset [18], a large-scale indoor scene dataset with 18,797 rooms furnished with 3D textured objects from 3D-FUTURE [19]. Following the same pre-processing steps as [33] and excluding non-rectangular floor plans from the test set similar to [15], we obtained 3397 train, 453 validation, 423 test bedroom and 690 train, 98 validation, 53 test living room scenes. During 3D scene assembling, 3D models are retrieved from the 3D-FUTURE dataset." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 240, + 483, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 240, + 483, + 410 + ], + "spans": [ + { + "bbox": [ + 130, + 240, + 483, + 410 + ], + "type": "text", + "content": "Implementation Details. We use ChatCompletions API" + }, + { + "bbox": [ + 130, + 240, + 483, + 410 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 130, + 240, + 483, + 410 + ], + "type": "text", + "content": ", which enables a series of dialogue exchanges between the user and the LLM, to query GPT-4 [1]. BlenderProc" + }, + { + "bbox": [ + 130, + 240, + 483, + 410 + ], + "type": "inline_equation", + "content": "^4" + }, + { + "bbox": [ + 130, + 240, + 483, + 410 + ], + "type": "text", + "content": " is employed to render 250 images per assembled scene with " + }, + { + "bbox": [ + 130, + 240, + 483, + 410 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 130, + 240, + 483, + 410 + ], + "type": "text", + "content": " resolution. After relocating the center of each scene to the origin of the coordinate system, we sample camera positions on the upper hemisphere at a distance of " + }, + { + "bbox": [ + 130, + 240, + 483, + 410 + ], + "type": "inline_equation", + "content": "1.5r" + }, + { + "bbox": [ + 130, + 240, + 483, + 410 + ], + "type": "text", + "content": " from the center where " + }, + { + "bbox": [ + 130, + 240, + 483, + 410 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 240, + 483, + 410 + ], + "type": "text", + "content": " is the diagonal length of the room, with an elevation angle of 35 degrees. We use the Splatifto model from NeRFstudio [46] for 3DGS training, by optimizing each scene for a maximum of 20k iterations. For editing the 3DGS training images, we use InstructPix2Pix [4] as the diffusion model, with classifier guidance scales " + }, + { + "bbox": [ + 130, + 240, + 483, + 410 + ], + "type": "inline_equation", + "content": "s_T = [3.5, 12.5]" + }, + { + "bbox": [ + 130, + 240, + 483, + 410 + ], + "type": "text", + "content": " for the text and " + }, + { + "bbox": [ + 130, + 240, + 483, + 410 + ], + "type": "inline_equation", + "content": "s_I = 1.5" + }, + { + "bbox": [ + 130, + 240, + 483, + 410 + ], + "type": "text", + "content": " for the image. The dataset updates are performed every 2.5k iterations following Instruct-GS2GS [48]. The InstructGS2GS training takes approximately 15-20 minutes per scene on a single A6000, most of it spent on diffusion model sampling. The rendering speed is 101.21 fps." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 426, + 274, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 426, + 274, + 438 + ], + "spans": [ + { + "bbox": [ + 132, + 426, + 274, + 438 + ], + "type": "text", + "content": "4.1 3D Layout Generation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 446, + 482, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 446, + 482, + 638 + ], + "spans": [ + { + "bbox": [ + 130, + 446, + 482, + 638 + ], + "type": "text", + "content": "Evaluation Setup and Baselines. In this section we evaluate the first stage of our approach, therefore we only focus on assessing the generated layouts. To our knowledge, SceneTeller is the first approach that conducts conditional layout generation guided by text prompts to control the position and orientation of every object, targeting a user-friendly framework for novices. Hence, replicating an interaction mode that is natural for humans is pivotal for our framework. Based on this, our layout generation performance is compared with the performance of humans solving the same task. In particular, each participant of our user study is presented with a series of textual descriptions, each accompanied by an empty layout according to the room dimensions. For each text prompt, participants are requested to draw bounding boxes and label them with the object names. Each text prompt is given to at least 2, at most 3 participants. This study is conducted with 12 participants and 40 text prompts. To generate text prompts, 40 layouts are selected from the test set and their text prompts are generated by providing the initially rule-based formed sentences to GPT for paraphrasing, as described in Sec. 3.1." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 257, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 257, + 91, + 447, + 102 + ], + "type": "text", + "content": "SceneTeller: Language-to-3D Scene Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 641, + 376, + 654 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 641, + 376, + 654 + ], + "spans": [ + { + "bbox": [ + 133, + 641, + 376, + 654 + ], + "type": "text", + "content": "3 See https://platform.openai.com/docs/api-reference/chat" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 135, + 654, + 336, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 654, + 336, + 665 + ], + "spans": [ + { + "bbox": [ + 135, + 654, + 336, + 665 + ], + "type": "text", + "content": "See https://github.com/DLR-RM/BlenderProc" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "Metrics. Based on the predicted number of bounding boxes in each category, we report recall, precision and accuracy [15]. An accuracy score of 100 indicates that the number of bounding boxes and their corresponding categories precisely match the ground-truth annotations. Otherwise, a score of 0 is assigned to the scene. All values are in percentage (\\%). We additionally report mean Intersection over Union (mIoU) between the predicted and ground-truth bounding boxes and out-of-bound rates, indicating the percentage of scenes where furniture extends beyond the floor plan boundary. We refer to the supplementary material for detailed metric descriptions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 243, + 482, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 243, + 482, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 482, + 411 + ], + "type": "text", + "content": "Results. Tab. 1 presents the quantitative evaluation results of 3D layout generation. Our in-context-learning-based 3D layout generation module performs better than human participants in terms of mIoU, showcasing the effectiveness of LLMs' visual reasoning capabilities. In-context exemplars further enhance this performance by aiding in reasoning about the current data distribution. However, human participants possess their own understanding of the physical world, leading to variations in their notions of object sizes and prompt reasoning, resulting in slightly poorer performance in this assessment. Both our layout generation module and human participants performed excellently in terms of predicting the number of bounding boxes per category. Humans sometimes forget or miss objects mentioned in the prompt, which explains their slightly imperfect recall score. The ground-truth layouts from the dataset involve objects slightly exceeding the floor plan boundary, leading to " + }, + { + "bbox": [ + 130, + 243, + 482, + 411 + ], + "type": "inline_equation", + "content": "15.15\\%" + }, + { + "bbox": [ + 130, + 243, + 482, + 411 + ], + "type": "text", + "content": " OOB rate. Human participants are able to place the objects within the layout more accurately." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 211, + 485, + 405, + 544 + ], + "blocks": [ + { + "bbox": [ + 130, + 431, + 483, + 478 + ], + "lines": [ + { + "bbox": [ + 130, + 431, + 483, + 478 + ], + "spans": [ + { + "bbox": [ + 130, + 431, + 483, + 478 + ], + "type": "text", + "content": "Table 1: Quantitative evaluation of 3D layout generation. The proposed in-context learning based 3D layout generation module performs better than human participants in terms of predicting object sizes, enhanced by in-context exemplars aiding in reasoning about the current data distribution. All values are in percentage (\\%)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 211, + 485, + 405, + 544 + ], + "lines": [ + { + "bbox": [ + 211, + 485, + 405, + 544 + ], + "spans": [ + { + "bbox": [ + 211, + 485, + 405, + 544 + ], + "type": "table", + "html": "
MethodOOB ↓rec.↑prec.↑acc.↑mIoU ↑
Ours56.78100.00100.00100.0044.60
Human12.0998.79100.0095.0028.70
GT15.15100.00100.00100.00100.00
", + "image_path": "ab0fd08d5f97041daedfd5d1daca3ffec9222702fe877f46fac5ea01322c021a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 585, + 268, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 585, + 268, + 596 + ], + "spans": [ + { + "bbox": [ + 132, + 585, + 268, + 596 + ], + "type": "text", + "content": "4.2 3D Scene Generation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": "Baselines. Our approach is evaluated and compared with three state-of-the-art 3D scene generation methods. Set-the-scene [11] employs compositional NeRFs that are initialized based on user-defined object proxies to have more control over individual objects. To compare with [11], layouts generated by our method are provided as input to their pipeline. GSGEN [9] is a 3DGS-based approach," + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 90, + 233, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 90, + 233, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 90, + 233, + 101 + ], + "type": "text", + "content": "B.M. Öcal et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 482, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 482, + 163 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 482, + 163 + ], + "type": "text", + "content": "following a two-stage pipeline for geometry optimization and appearance refinement. LucidDreamer [10] employs render-refine-repeat paradigm by iteratively warping and inpainting the previously generated image to create an initial point cloud, which is then used for 3DGS scene training." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 136, + 186, + 477, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 186, + 477, + 210 + ], + "spans": [ + { + "bbox": [ + 136, + 186, + 477, + 210 + ], + "type": "text", + "content": "a t a o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o ." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 137, + 212, + 186, + 260 + ], + "blocks": [ + { + "bbox": [ + 137, + 212, + 186, + 260 + ], + "lines": [ + { + "bbox": [ + 137, + 212, + 186, + 260 + ], + "spans": [ + { + "bbox": [ + 137, + 212, + 186, + 260 + ], + "type": "image", + "image_path": "b9cd48f6a5141197abf01aa23d3a76c3976289e99000f006b0869a0e967393be.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 196, + 212, + 246, + 260 + ], + "blocks": [ + { + "bbox": [ + 196, + 212, + 246, + 260 + ], + "lines": [ + { + "bbox": [ + 196, + 212, + 246, + 260 + ], + "spans": [ + { + "bbox": [ + 196, + 212, + 246, + 260 + ], + "type": "image", + "image_path": "9d0c2dd39d0b81c0c2b2daebe98265ee63c4fa095e9b32532f403be633440f50.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 259, + 212, + 332, + 260 + ], + "blocks": [ + { + "bbox": [ + 259, + 212, + 332, + 260 + ], + "lines": [ + { + "bbox": [ + 259, + 212, + 332, + 260 + ], + "spans": [ + { + "bbox": [ + 259, + 212, + 332, + 260 + ], + "type": "image", + "image_path": "e02baf7723bf937babf56b54897fdf4786d174d2548fa5781e7438366857e00d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 349, + 212, + 400, + 260 + ], + "blocks": [ + { + "bbox": [ + 349, + 212, + 400, + 260 + ], + "lines": [ + { + "bbox": [ + 349, + 212, + 400, + 260 + ], + "spans": [ + { + "bbox": [ + 349, + 212, + 400, + 260 + ], + "type": "image", + "image_path": "58e6f5f6c1e8785094486142bd3c64da1e2509215419e29e01e0a9e61db77e42.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 417, + 212, + 468, + 260 + ], + "blocks": [ + { + "bbox": [ + 417, + 212, + 468, + 260 + ], + "lines": [ + { + "bbox": [ + 417, + 212, + 468, + 260 + ], + "spans": [ + { + "bbox": [ + 417, + 212, + 468, + 260 + ], + "type": "image", + "image_path": "05f4520687f5ef3a8235ed2a3aaf0236035091f74f4f2615b4b2b171f282c844.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 262, + 474, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 262, + 474, + 302 + ], + "spans": [ + { + "bbox": [ + 138, + 262, + 474, + 302 + ], + "type": "text", + "content": "\"The dining table is situated in the middle of the room, but slightly leaning towards the lower left wall, with its orientation reversed and perpendicular. One dining chair is positioned near the lower left corner, with its orientation being perpendicular. Another dining chair is placed in the middle of the left wall, a bit nearer to the lower wall, and is also perpendicular. Yet another dining chair is set up at the midpoint of the left wall, but it's slightly more towards the upper wall and is perpendicular. One dining chair is near the lower right corner, with its orientation reversed and perpendicular. Another dining chair can be found at the center of the right wall, a bit closer to the lower wall, and it is perpendicular but facing the opposite direction. Another dining chair is at the center of the right wall, but slightly more towards the upper wall and its orientation is perpendicular but reversed. One more dining chair is in the middle of the upper wall, closer to the left wall, and is oriented backwards. Finally, a dining chair is at the center of the lower wall, a bit closer to the left wall, and it is oriented without any rotation. Make it marble.\"" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 137, + 306, + 186, + 354 + ], + "blocks": [ + { + "bbox": [ + 137, + 306, + 186, + 354 + ], + "lines": [ + { + "bbox": [ + 137, + 306, + 186, + 354 + ], + "spans": [ + { + "bbox": [ + 137, + 306, + 186, + 354 + ], + "type": "image", + "image_path": "ef04450c06340b11e91bce98c9b579e3ab4f864fdf71d49f6290099d33b27bec.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 196, + 306, + 245, + 354 + ], + "blocks": [ + { + "bbox": [ + 196, + 306, + 245, + 354 + ], + "lines": [ + { + "bbox": [ + 196, + 306, + 245, + 354 + ], + "spans": [ + { + "bbox": [ + 196, + 306, + 245, + 354 + ], + "type": "image", + "image_path": "c3702bd7776a7f92ee6b846a80d566ccead06a59b44dee2cbf494d1e261bcc5a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 259, + 306, + 335, + 354 + ], + "blocks": [ + { + "bbox": [ + 259, + 306, + 335, + 354 + ], + "lines": [ + { + "bbox": [ + 259, + 306, + 335, + 354 + ], + "spans": [ + { + "bbox": [ + 259, + 306, + 335, + 354 + ], + "type": "image", + "image_path": "af387969c8cf0fadb7fe6d4d7e95acaee26d77881a2a7ee371bd4d0f770bd79e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 343, + 307, + 409, + 354 + ], + "blocks": [ + { + "bbox": [ + 343, + 307, + 409, + 354 + ], + "lines": [ + { + "bbox": [ + 343, + 307, + 409, + 354 + ], + "spans": [ + { + "bbox": [ + 343, + 307, + 409, + 354 + ], + "type": "image", + "image_path": "116e26e5cb13ab5e30d577d57d78bb755351d6d1394804aa0592c68e425942f2.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 417, + 308, + 470, + 353 + ], + "blocks": [ + { + "bbox": [ + 417, + 308, + 470, + 353 + ], + "lines": [ + { + "bbox": [ + 417, + 308, + 470, + 353 + ], + "spans": [ + { + "bbox": [ + 417, + 308, + 470, + 353 + ], + "type": "image", + "image_path": "2739be99be80a1abbf9ed44d7bbc1f87d93a69f75b9cbeabd1632c379a21473c.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 356, + 475, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 356, + 475, + 398 + ], + "spans": [ + { + "bbox": [ + 138, + 356, + 475, + 398 + ], + "type": "text", + "content": "The coffe table is sitd in the middle of the upper wall, albeit a tad nearer to the right wall, and is aligned perpendicularly facing the opposite direction. Another coffe table can be found in the middle of the right wall, a bit nearer to the lower wall, and is facing towards the rear. Nestled in the upper right corner is another coffe table, positioned without any rotation. In the upper left corner, a tv stand is positioned perpendicularly. A multi seat sofa is positioned near the upper right corner, facing perpendicularly but in the reverse direction. Close to the lower left corner, a dining table is situated without any rotation. A dinng chair can be found in the center of the lower wall, slightly nearer to the left wall, and is facing in its original direction. Another dinng chair is positioned near the lower left corner, also without any rotation. A third dinng chair is situated in the room's center, closer to the lower left wall, and oriented towards the rear. A fourth dinng chair is positioned in the middle of the left wall, slightly closer to the lower wall, positioned towards the rear. Lastly, a console table can be found in the lower left corner, placed without any rotation. Make it pink themed.\"" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 137, + 400, + 186, + 449 + ], + "blocks": [ + { + "bbox": [ + 137, + 400, + 186, + 449 + ], + "lines": [ + { + "bbox": [ + 137, + 400, + 186, + 449 + ], + "spans": [ + { + "bbox": [ + 137, + 400, + 186, + 449 + ], + "type": "image", + "image_path": "6e046d146f2c454a993350cdc2a17ceab9fd6d15af5809dfd2a5d7735c209aad.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 196, + 400, + 245, + 449 + ], + "blocks": [ + { + "bbox": [ + 196, + 400, + 245, + 449 + ], + "lines": [ + { + "bbox": [ + 196, + 400, + 245, + 449 + ], + "spans": [ + { + "bbox": [ + 196, + 400, + 245, + 449 + ], + "type": "image", + "image_path": "2e6a1815d743a8054c7934410e242013ea5396f686eaaff61a24b4c332d6fed9.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 259, + 400, + 335, + 449 + ], + "blocks": [ + { + "bbox": [ + 259, + 400, + 335, + 449 + ], + "lines": [ + { + "bbox": [ + 259, + 400, + 335, + 449 + ], + "spans": [ + { + "bbox": [ + 259, + 400, + 335, + 449 + ], + "type": "image", + "image_path": "50c77c19e1fea3fdfe1bb02c96b14a416ec4d2e4ca561bb7cd745ecb96775de8.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 340, + 403, + 410, + 451 + ], + "blocks": [ + { + "bbox": [ + 340, + 403, + 410, + 451 + ], + "lines": [ + { + "bbox": [ + 340, + 403, + 410, + 451 + ], + "spans": [ + { + "bbox": [ + 340, + 403, + 410, + 451 + ], + "type": "image", + "image_path": "6dbf722a40460f2451c1349b8d2eb6e4c4f1f5294aa73490309ba7674e860a72.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 417, + 403, + 473, + 445 + ], + "blocks": [ + { + "bbox": [ + 417, + 403, + 473, + 445 + ], + "lines": [ + { + "bbox": [ + 417, + 403, + 473, + 445 + ], + "spans": [ + { + "bbox": [ + 417, + 403, + 473, + 445 + ], + "type": "image", + "image_path": "004c5817ad26d584c0a3b933b66ce03264f5e7c00538d3ab224c763d53a589da.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 452, + 475, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 452, + 475, + 469 + ], + "spans": [ + { + "bbox": [ + 138, + 452, + 475, + 469 + ], + "type": "text", + "content": "At the heart of the room, a double bed is positioned a tad nearer to the right wall, its orientation is perpendicular but in reverse. Adjacent to the top-right corner, a nightstand is placed, its orientation is also perpendicular but in reverse. A wardrobe can be found centrally aligned with the bottom wall, subtly closer to the right wall, with no rotation in its orientation. Make it Starry Night Van Gogh painting style.\"" + } + ] + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 137, + 471, + 186, + 520 + ], + "blocks": [ + { + "bbox": [ + 137, + 471, + 186, + 520 + ], + "lines": [ + { + "bbox": [ + 137, + 471, + 186, + 520 + ], + "spans": [ + { + "bbox": [ + 137, + 471, + 186, + 520 + ], + "type": "image", + "image_path": "5f26f22cc9e384574c1e4245a3895ea4bc62bf6972676f5f5ebf98209ec70fff.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 196, + 471, + 246, + 521 + ], + "blocks": [ + { + "bbox": [ + 196, + 471, + 246, + 521 + ], + "lines": [ + { + "bbox": [ + 196, + 471, + 246, + 521 + ], + "spans": [ + { + "bbox": [ + 196, + 471, + 246, + 521 + ], + "type": "image", + "image_path": "43678837dad34e5db21bc61d0b262e05caac0452fa28cfc63903ce38b7f30282.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 259, + 471, + 333, + 520 + ], + "blocks": [ + { + "bbox": [ + 259, + 471, + 333, + 520 + ], + "lines": [ + { + "bbox": [ + 259, + 471, + 333, + 520 + ], + "spans": [ + { + "bbox": [ + 259, + 471, + 333, + 520 + ], + "type": "image", + "image_path": "d4a797edc559cd12f4e55bb9e1d797dee62a3ffff29dfa010341c075a90e14c0.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 349, + 469, + 400, + 518 + ], + "blocks": [ + { + "bbox": [ + 349, + 469, + 400, + 518 + ], + "lines": [ + { + "bbox": [ + 349, + 469, + 400, + 518 + ], + "spans": [ + { + "bbox": [ + 349, + 469, + 400, + 518 + ], + "type": "image", + "image_path": "2ff0b3f0469ae4855ab16c84fba910428c6d002b72588f5934ee731a5d2555b7.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 417, + 469, + 467, + 518 + ], + "blocks": [ + { + "bbox": [ + 417, + 469, + 467, + 518 + ], + "lines": [ + { + "bbox": [ + 417, + 469, + 467, + 518 + ], + "spans": [ + { + "bbox": [ + 417, + 469, + 467, + 518 + ], + "type": "image", + "image_path": "6938169294f0733a5742f89ded17979078fd393da12f0c8aa4e1e4e3a659a0c3.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "bbox": [ + 135, + 521, + 476, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 521, + 476, + 545 + ], + "spans": [ + { + "bbox": [ + 135, + 521, + 476, + 545 + ], + "type": "text", + "content": "In the middle of the room, a double bed is placed, slightly more towards the bottom right wall, with its orientation being perpendicular and facing the reverse direction. A wardrobe is situated at the center of the top wall, a bit closer to the right wall, with its orientation facing backwards. Near the top right corner of the room, a table is situated, oriented perpendicularly but in the reverse direction. At the bottom left corner, a cabinet is positioned, perpendicular to the walls. Another table is situated near the bottom right corner, oriented perpendicularly but facing the opposite direction. Turn the double bed into golden.\"" + } + ] + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 137, + 547, + 185, + 596 + ], + "blocks": [ + { + "bbox": [ + 137, + 547, + 185, + 596 + ], + "lines": [ + { + "bbox": [ + 137, + 547, + 185, + 596 + ], + "spans": [ + { + "bbox": [ + 137, + 547, + 185, + 596 + ], + "type": "image", + "image_path": "812c1438c05f42d0ad6ef9153d9033e66cc7255a764ac85539cdda8666a19b1a.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 148, + 600, + 171, + 607 + ], + "lines": [ + { + "bbox": [ + 148, + 600, + 171, + 607 + ], + "spans": [ + { + "bbox": [ + 148, + 600, + 171, + 607 + ], + "type": "text", + "content": "GSGEN" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 198, + 548, + 247, + 597 + ], + "blocks": [ + { + "bbox": [ + 198, + 548, + 247, + 597 + ], + "lines": [ + { + "bbox": [ + 198, + 548, + 247, + 597 + ], + "spans": [ + { + "bbox": [ + 198, + 548, + 247, + 597 + ], + "type": "image", + "image_path": "25193ae74de29f3d1da45e292352c7ef1d77361dfa4ae88caa2eba43d39bb9e9.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 203, + 599, + 242, + 607 + ], + "lines": [ + { + "bbox": [ + 203, + 599, + 242, + 607 + ], + "spans": [ + { + "bbox": [ + 203, + 599, + 242, + 607 + ], + "type": "text", + "content": "Set-the-scene" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 260, + 548, + 332, + 596 + ], + "blocks": [ + { + "bbox": [ + 260, + 548, + 332, + 596 + ], + "lines": [ + { + "bbox": [ + 260, + 548, + 332, + 596 + ], + "spans": [ + { + "bbox": [ + 260, + 548, + 332, + 596 + ], + "type": "image", + "image_path": "576bb80a86cbf15a1999f877aa76aada3915f2cc2a0955e86cf85d51238bd8fe.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 275, + 599, + 317, + 607 + ], + "lines": [ + { + "bbox": [ + 275, + 599, + 317, + 607 + ], + "spans": [ + { + "bbox": [ + 275, + 599, + 317, + 607 + ], + "type": "text", + "content": "LucidDreamer" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 350, + 548, + 396, + 596 + ], + "blocks": [ + { + "bbox": [ + 350, + 548, + 396, + 596 + ], + "lines": [ + { + "bbox": [ + 350, + 548, + 396, + 596 + ], + "spans": [ + { + "bbox": [ + 350, + 548, + 396, + 596 + ], + "type": "image", + "image_path": "9a97b20e054b0270228105f90ac4d371f5c734e2e58b09a39e22c1b6a998688b.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 618, + 482, + 662 + ], + "lines": [ + { + "bbox": [ + 131, + 618, + 482, + 662 + ], + "spans": [ + { + "bbox": [ + 131, + 618, + 482, + 662 + ], + "type": "text", + "content": "Fig. 3: Qualitative comparison with state-of-the-art text-to-3D scene generation methods. The top/upper walls are marked with arrows for your reference, if this information is available within the generations. SceneTeller is able to generate high-quality scenes, with superior geometric fidelity and 3D consistency." + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 399, + 559, + 468, + 594 + ], + "blocks": [ + { + "bbox": [ + 399, + 559, + 468, + 594 + ], + "lines": [ + { + "bbox": [ + 399, + 559, + 468, + 594 + ], + "spans": [ + { + "bbox": [ + 399, + 559, + 468, + 594 + ], + "type": "image", + "image_path": "e6e4b7a01a3ef283b5a2b88b38f5803ac6c27d6b38f1e7f0c2bbefc108ffb3c8.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 397, + 600, + 413, + 607 + ], + "lines": [ + { + "bbox": [ + 397, + 600, + 413, + 607 + ], + "spans": [ + { + "bbox": [ + 397, + 600, + 413, + 607 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 256, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 256, + 91, + 447, + 102 + ], + "type": "text", + "content": "SceneTeller: Language-to-3D Scene Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "User Study. We conduct a user study with 30 participants to assess the performance of our method in generating high-quality text-consistent realistic 3D scenes. Human evaluators are presented with 10 scenes, each accompanied by a textual prompt describing its layout and style. For each scene, they are requested to rate it on a scale from 1 to 5 (5 being the best) based on four criteria: (a) Realism, (b) Text Alignment, (c) Geometric Fidelity and (d) Compositional Plausibility." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 482, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 482, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 482, + 272 + ], + "type": "text", + "content": "We present the results of the study in Tab. 2. Our method consistently outperforms the existing state-of-the-art 3D scene generation methods across all the four criteria by a significant margin, showcasing the merits of SceneTeller in generating realistic high-quality 3D scenes solely from text prompts. For additional details of the study and extensive analysis of the results, we refer to the supplementary material." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 146, + 369, + 469, + 463 + ], + "blocks": [ + { + "bbox": [ + 130, + 293, + 483, + 361 + ], + "lines": [ + { + "bbox": [ + 130, + 293, + 483, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 483, + 361 + ], + "type": "text", + "content": "Table 2: User study results. Study conducted with 30 participants to assess the performance of state-of-the-art methods in generating 3D scenes based on four criteria with a scale from 1 to 5 (5 being the best and vice versa). Our method consistently outperforms the existing state-of-the-art 3D scene generation methods across all the four criteria by a significant margin, showcasing the merits of SceneTeller in generating realistic high-quality 3D scenes solely from text prompts." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 146, + 369, + 469, + 463 + ], + "lines": [ + { + "bbox": [ + 146, + 369, + 469, + 463 + ], + "spans": [ + { + "bbox": [ + 146, + 369, + 469, + 463 + ], + "type": "table", + "html": "
MethodInput TypeUser Study
RealismText AlignmentGeometric FidelityCompositional Plausibility
GSGEN [9]T1.3131.3751.4391.392
LucidDreamer [10]T3.5432.7513.0503.116
Set-the-scene [11]L + T1.9631.9002.1122.181
OursT4.1544.1894.4794.323
", + "image_path": "f570d5131c9f6867bccca3a4ee05f1d30ac6fb499d749e5da7d06dd74aaaf27f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 503, + 482, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 503, + 482, + 588 + ], + "spans": [ + { + "bbox": [ + 130, + 503, + 482, + 588 + ], + "type": "text", + "content": "Quantitative Evaluation of Stylization. To evaluate the consistency of the generated scene style with textual descriptions, we utilize the CLIP Score [21] as the metric. Given that the CLIP score cannot reason about the parts of the prompts describing the layout, we only provide style information to CLIP in the format: 'A Barbie style bedroom' or 'A bedroom with Barbie style nightstands' for all methods. As shown in Tab. 3, our method consistently generates scenes adhering to the given style description." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Qualitative Evaluation. Visual comparison of SceneTeller with state-of-the-art text-to-3D methods are provided in Fig. 3. SceneTeller yields high-quality 3D scene generations, surpassing the baselines in terms of text-alignment and 3D consistency by predicting an initial 3D layout conditioned on the textual descriptions. Unlike some of the state-of-the-art baselines producing unrealistic" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 90, + 233, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 90, + 233, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 90, + 233, + 101 + ], + "type": "text", + "content": "B.M. Öcal et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 160, + 157, + 454, + 191 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 3: Quantitative comparison of stylization. We report the CLIP Score to evaluate the consistency of the generated scene style with textual descriptions. Our method consistently generates scenes adhering to the given style description." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 160, + 157, + 454, + 191 + ], + "lines": [ + { + "bbox": [ + 160, + 157, + 454, + 191 + ], + "spans": [ + { + "bbox": [ + 160, + 157, + 454, + 191 + ], + "type": "table", + "html": "
GSGEN [9]LucidDreamer [10]Set-the-scene [11]Ours
CLIP Score ↑26.47830.71527.18830.923
", + "image_path": "c4bef7d7d9fa8c9c1a32cc3e178a76f0435561c11fcd865ab0d16fa9228b3b13.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 213, + 481, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 481, + 239 + ], + "type": "text", + "content": "appearance (e.g. comic or cartoon-like characteristics), SceneTeller is able to produce more realistic object shapes and appearances." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 256, + 481, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 256, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 256, + 481, + 304 + ], + "type": "text", + "content": "Realism of the Generated Scenes. We evaluate all methods using FID score with a real indoor dataset ScanNet [12] on 10 generated scenes by rendering 50 images from each of them. As shown in Tab. 4, our method consistently generates more realistic scenes." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 175, + 345, + 438, + 379 + ], + "blocks": [ + { + "bbox": [ + 201, + 323, + 411, + 337 + ], + "lines": [ + { + "bbox": [ + 201, + 323, + 411, + 337 + ], + "spans": [ + { + "bbox": [ + 201, + 323, + 411, + 337 + ], + "type": "text", + "content": "Table 4: FID score with the ScanNet dataset [12]." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 175, + 345, + 438, + 379 + ], + "lines": [ + { + "bbox": [ + 175, + 345, + 438, + 379 + ], + "spans": [ + { + "bbox": [ + 175, + 345, + 438, + 379 + ], + "type": "table", + "html": "
GSGEN [9] LucidDreamer [10] Set-the-scene [11]Ours
FID ↓297.10238.75297.57167.27
", + "image_path": "bd60056536bd8594bada752833f7e479b4a66a764c6df398c7e2bb5891578896.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 422, + 482, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 482, + 483 + ], + "type": "text", + "content": "Applicability to Real-world. While current design tools enable professionals to create floor plans, visualize 3D models, and render photo-realistic images, they are not suitable for novice users unfamiliar with such software tools. By enabling users to interact solely through plain natural language prompts, our approach also empowers novice users to flexibly design their own rooms." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 501, + 239, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 501, + 239, + 514 + ], + "spans": [ + { + "bbox": [ + 131, + 501, + 239, + 514 + ], + "type": "text", + "content": "4.3 Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "Effect of design choices for 3D layout generation. To assess the effectiveness of individual components in our design, we conduct an ablation study. Specifically, we examine two aspects: 1) the prompt source for in-context exemplars, and 2) the technique for selecting in-context exemplars. Regarding the prompt source, we present exemplars where the textual description is generated either solely by the rule-based formulation (R) or by a combination of the rule-based formulation and GPT paraphrasing " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "(\\mathrm{R} + \\mathrm{GPT})" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " to introduce linguistic variation as described in Sec. 3.1. For the exemplar selection technique, we experiment with a random and a retrieval-based exemplar selection following [15], " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "k = 8" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " exemplars. Additionally, we experiment with a third set of exemplars, denoted as " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "pos + neg" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": ", where 4 exemplars are selected by retrieval, and the other 4 are chosen from negative exemplars involving out-of-bound (OOB)" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 256, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 256, + 91, + 447, + 102 + ], + "type": "text", + "content": "SceneTeller: Language-to-3D Scene Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "or overlapping objects. The negative ones are from the filtered samples during the pre-processing of the dataset, and the LLM is specifically instructed to avoid them. Tab. 5 verifies that providing in-context exemplars with linguistic variation " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "(\\mathrm{R} + \\mathrm{GPT})" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " selected with a retrieval-based approach performs the best, especially in terms of OOB rate by constraining the in-context learning of LLM to similar layouts. Using the " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "pos + neg" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " set also considerably improves the performance in terms of mIoU." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 134, + 271, + 480, + 387 + ], + "blocks": [ + { + "bbox": [ + 130, + 217, + 482, + 264 + ], + "lines": [ + { + "bbox": [ + 130, + 217, + 482, + 264 + ], + "spans": [ + { + "bbox": [ + 130, + 217, + 482, + 264 + ], + "type": "text", + "content": "Table 5: Effect of design choices for 3D layout generation. We report the 3D layout generation performance w.r.t different design choices such as the prompt source and the exemplar selection. Providing in-context exemplars to LLM with linguistic variation (R+GPT) selected with a retrieval-based approach, performs the best." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 271, + 480, + 387 + ], + "lines": [ + { + "bbox": [ + 134, + 271, + 480, + 387 + ], + "spans": [ + { + "bbox": [ + 134, + 271, + 480, + 387 + ], + "type": "table", + "html": "
Prompt sourceExemplar SelectionOOB(%)Numerical ReasoningBbox
random retrieval pos+neg↓rec.↑prec.↑acc.↑
1R✓74.12100.00100.00100.0031.10
2R✓64.87100.00100.00100.0045.50
3R74.10100.00100.00100.0041.90
4R + GPT✓76.1199.1099.1099.1029.90
5R + GPT✓54.65100.00100.00100.0044.20
6R + GPT66.45100.00100.00100.0043.90
", + "image_path": "af12552783cd9cbc1a4581f54f0123519530ca04504a58fa918724353d11212f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 426, + 220, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 426, + 220, + 438 + ], + "spans": [ + { + "bbox": [ + 132, + 426, + 220, + 438 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 448, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 448, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 448, + 482, + 521 + ], + "type": "text", + "content": "In this paper, we presented a novel pipeline for 3D room generation from a textual description. Our approach supports style editing either on the individual object or on the entire scene level. Rooms generated by our method were preferred by human evaluators over the baseline results almost in all cases, confirming its efficacy on this challenging task. Importantly, our approach can be used by novices, thus making a step towards democratizing 3D scene design." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 536, + 247, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 247, + 550 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 247, + 550 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 558, + 481, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 481, + 595 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 481, + 595 + ], + "type": "text", + "content": "We thank members of the Bosch-UvA Delta Lab and anonymous reviewers for helpful discussions and feedback. This project was generously supported by the Bosch Center for Artificial Intelligence." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 610, + 198, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 198, + 624 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 198, + 624 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 136, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 136, + 632, + 481, + 666 + ], + "type": "text", + "content": "1. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 90, + 233, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 90, + 233, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 90, + 233, + 101 + ], + "type": "text", + "content": "B.M. Öcal et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 134, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 136, + 116, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 116, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 136, + 116, + 482, + 160 + ], + "type": "text", + "content": "2. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems (NeurIPS) 35, 23716-23736 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 161, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 161, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 138, + 161, + 481, + 205 + ], + "type": "text", + "content": "3. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (CVPR). pp. 5855-5864 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 206, + 481, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 206, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 138, + 206, + 481, + 239 + ], + "type": "text", + "content": "4. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18392-18402 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 239, + 481, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 239, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 138, + 239, + 481, + 262 + ], + "type": "text", + "content": "5. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. In: European Conference on Computer Vision (ECCV) (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 262, + 481, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 262, + 481, + 296 + ], + "spans": [ + { + "bbox": [ + 138, + 262, + 481, + 296 + ], + "type": "text", + "content": "6. Chen, R., Chen, Y., Jiao, N., Jia, K.: Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (October 2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 297, + 481, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 297, + 481, + 329 + ], + "spans": [ + { + "bbox": [ + 138, + 297, + 481, + 329 + ], + "type": "text", + "content": "7. Chen, X., Zhang, Q., Li, X., Chen, Y., Feng, Y., Wang, X., Wang, J.: Hallucinated neural radiance fields in the wild. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12943-12952 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 330, + 481, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 330, + 481, + 363 + ], + "spans": [ + { + "bbox": [ + 138, + 330, + 481, + 363 + ], + "type": "text", + "content": "8. Chen, Y., Chen, Z., Zhang, C., Wang, F., Yang, X., Wang, Y., Cai, Z., Yang, L., Liu, H., Lin, G.: Gaussianeditor: Swift and controllable 3d editing with gaussian splatting. arXiv preprint arXiv:2311.14521 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 364, + 481, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 364, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 138, + 364, + 481, + 385 + ], + "type": "text", + "content": "9. Chen, Z., Wang, F., Liu, H.: Text-to-3d using gaussian splatting. arXiv preprint arXiv:2309.16585 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 134, + 386, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 386, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 134, + 386, + 481, + 407 + ], + "type": "text", + "content": "10. Chung, J., Lee, S., Nam, H., Lee, J., Lee, K.M.: Luciddreamer: Domain-free generation of 3d gaussian splatting scenes. arXiv preprint arXiv:2311.13384 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 134, + 408, + 481, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 408, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 134, + 408, + 481, + 453 + ], + "type": "text", + "content": "11. Cohen-Bar, D., Richardson, E., Metzer, G., Giryes, R., Cohen-Or, D.: Set-the-scene: Global-local training for generating controllable nerf scenes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops. pp. 2920–2929 (October 2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 134, + 453, + 481, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 453, + 481, + 486 + ], + "spans": [ + { + "bbox": [ + 134, + 453, + 481, + 486 + ], + "type": "text", + "content": "12. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proc. Computer Vision and Pattern Recognition (CVPR), IEEE (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 487, + 481, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 487, + 481, + 509 + ], + "spans": [ + { + "bbox": [ + 134, + 487, + 481, + 509 + ], + "type": "text", + "content": "13. Fang, J., Wang, J., Zhang, X., Xie, L., Tian, Q.: Gaussianeditor: Editing 3d gaussians delicately with text instructions. arXiv preprint arXiv:2311.16037 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 510, + 481, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 510, + 481, + 542 + ], + "spans": [ + { + "bbox": [ + 134, + 510, + 481, + 542 + ], + "type": "text", + "content": "14. Fang, J., Wang, J., Zhang, X., Xie, L., Tian, Q.: Gaussianeditor: Editing 3d gaussians delicately with text instructions. In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 543, + 481, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 543, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 134, + 543, + 481, + 586 + ], + "type": "text", + "content": "15. Feng, W., Zhu, W., Fu, T.j., Jampani, V., Akula, A., He, X., Basu, S., Wang, X.E., Wang, W.Y.: Layoutgpt: Compositional visual planning and generation with large language models. Advances in Neural Information Processing Systems (NeurIPS) 36 (2024)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 587, + 481, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 587, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 134, + 587, + 481, + 620 + ], + "type": "text", + "content": "16. Fridman, R., Abecasis, A., Kasten, Y., Dekel, T.: Scenescape: Text-driven consistent scene generation. Advances in Neural Information Processing Systems (NeurIPS) 36 (2024)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 621, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 621, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 621, + 481, + 665 + ], + "type": "text", + "content": "17. Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5501-5510 (2022)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 256, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 256, + 91, + 447, + 102 + ], + "type": "text", + "content": "SceneTeller: Language-to-3D Scene Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "text", + "content": "18. Fu, H., Cai, B., Gao, L., Zhang, L.X., Wang, J., Li, C., Zeng, Q., Sun, C., Jia, R., Zhao, B., et al.: 3d-front: 3d furnished rooms with layouts and semantics. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10933-10942 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "type": "text", + "content": "19. Fu, H., Jia, R., Gao, L., Gong, M., Zhao, B., Maybank, S., Tao, D.: 3d-future: 3d furniture shape with texture. International Journal of Computer Vision 129, 3313-3337 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "text", + "content": "20. Haque, A., Tancik, M., Efros, A., Holynski, A., Kanazawa, A.: Instruct-nerf2nerf: Editing 3d scenes with instructions. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (CVPR) (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 303 + ], + "type": "text", + "content": "21. Hessel, J., Holtzman, A., Forbes, M., Le Bras, R., Choi, Y.: CLIPScore: A reference-free evaluation metric for image captioning. In: Moens, M.F., Huang, X., Specia, L., Yih, S.W.t. (eds.) Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. pp. 7514-7528. Association for Computational Linguistics, Online and Punta Cana, Dominican Republic (Nov 2021). https://doi.org/10.18653/v1/2021.emnlp-main.595, https://aclanthology.org/2021.emnlp-main.595" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 304, + 481, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 326 + ], + "type": "text", + "content": "22. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 326, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 369 + ], + "type": "text", + "content": "23. Hollein, L., Cao, A., Owens, A., Johnson, J., Nießner, M.: Text2room: Extracting textured 3d meshes from 2d text-to-image models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 7909-7920 (October 2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 370, + 481, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 370, + 481, + 403 + ], + "spans": [ + { + "bbox": [ + 132, + 370, + 481, + 403 + ], + "type": "text", + "content": "24. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (July 2023), https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "type": "text", + "content": "25. Li, Q., Li, F., Guo, J., Guo, Y.: Uhdnerf: Ultra-high-definition neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (CVPR). pp. 23097-23108 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "type": "text", + "content": "26. Li, X., Wang, H., Tseng, K.K.: Gaussiandiffusion: 3d gaussian splatting for denoising diffusion probabilistic models with structured noise. arXiv preprint arXiv:2311.11221 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 468, + 481, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 512 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 512 + ], + "type": "text", + "content": "27. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3d: High-resolution text-to-3d content creation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 300-309 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "text", + "content": "28. Lin, Y., Bai, H., Li, S., Lu, H., Lin, X., Xiong, H., Wang, L.: Componerf: Text-guided multi-object compositional nerf with editable 3d scene layout. arXiv preprint arXiv:2303.13843 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 545, + 481, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 481, + 588 + ], + "type": "text", + "content": "29. Lorraine, J., Xie, K., Zeng, X., Lin, C.H., Takikawa, T., Sharp, N., Lin, T.Y., Liu, M.Y., Fidler, S., Lucas, J.: Att3d: Amortized text-to-3d object synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 17946-17956 (October 2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 589, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 481, + 632 + ], + "type": "text", + "content": "30. Metzer, G., Richardson, E., Patashnik, O., Giryes, R., Cohen-Or, D.: Latent-nerf for shape-guided generation of 3d shapes and textures. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12663-12673 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "31. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: Proceedings of the European Conference on Computer Vision (ECCV) (2020)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 90, + 233, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 90, + 233, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 90, + 233, + 101 + ], + "type": "text", + "content": "B.M. Öcal et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "32. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 482, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 482, + 181 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 482, + 181 + ], + "type": "text", + "content": "33. Paschalidou, D., Kar, A., Shugrina, M., Kreis, K., Geiger, A., Fidler, S.: Atiss: Autoregressive transformers for indoor scene synthesis. Advances in Neural Information Processing Systems 34, 12013-12026 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 181, + 482, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 181, + 482, + 202 + ], + "spans": [ + { + "bbox": [ + 130, + 181, + 482, + 202 + ], + "type": "text", + "content": "34. Phung, Q., Ge, S., Huang, J.B.: Grounded text-to-image synthesis with attention refocusing. arXiv preprint arXiv:2306.05427 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 202, + 482, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 202, + 482, + 235 + ], + "spans": [ + { + "bbox": [ + 130, + 202, + 482, + 235 + ], + "type": "text", + "content": "35. Po, R., Wetzstein, G.: Compositional 3d scene generation using locally conditioned diffusion. ArXiv abs/2303.12218 (2023), https://api-semanticscholar.org/CorpusID:257663283" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 234, + 482, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 234, + 482, + 267 + ], + "spans": [ + { + "bbox": [ + 130, + 234, + 482, + 267 + ], + "type": "text", + "content": "36. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. In: The Eleventh International Conference on Learning Representations (ICLR) (2023), https://openreview.net/forum?id=FjNys5c7VvY" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 266, + 482, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 266, + 482, + 320 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 482, + 320 + ], + "type": "text", + "content": "37. Qian, G., Mai, J., Hamdi, A., Ren, J., Siarohin, A., Li, B., Lee, H.Y., Skorokhodov, I., Wonka, P., Tulyakov, S., Ghanem, B.: Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. In: The Twelfth International Conference on Learning Representations (ICLR) (2024), https://openreview.net/forum?id=0jHkUDyEO9" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 320, + 482, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 320, + 482, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 320, + 482, + 365 + ], + "type": "text", + "content": "38. Raj, A., Kaza, S., Poole, B., Niemeyer, M., Mildenhall, B., Ruiz, N., Zada, S., Aberman, K., Rubenstein, M., Barron, J., Li, Y., Jampani, V.: Dreambooth3d: Subject-driven text-to-3d generation. The International Conference on Computer Vision (ICCV) (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 363, + 482, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 363, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 363, + 482, + 407 + ], + "type": "text", + "content": "39. Reiser, C., Szeliski, R., Verbin, D., Srinivasan, P., Mildenhall, B., Geiger, A., Barron, J., Hedman, P.: Merf: Memory-efficient radiance fields for real-time view synthesis in unbounded scenes. ACM Transactions on Graphics (TOG) 42(4), 1-12 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 407, + 482, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 482, + 440 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 482, + 440 + ], + "type": "text", + "content": "40. Ren, J., He, C., Liu, L., Chen, J., Wang, Y., Song, Y., Li, J., Xue, T., Hu, S., Chen, T., Zheng, K., Xiang, J., Bo, L.: Make-a-character: High quality text-to-3d character generation within minutes. arXiv preprint arXiv:2312.15430 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 438, + 482, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 482, + 471 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 482, + 471 + ], + "type": "text", + "content": "41. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 470, + 482, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 470, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 470, + 482, + 568 + ], + "type": "text", + "content": "42. Sanh, V., Webson, A., Raffel, C., Bach, S., Sutawika, L., Ålyafeai, Z., Chaffin, Å., Stiegler, A., Raja, A., Dey, M., Bari, M.S., Xu, C., Thakker, U., Sharma, S.S., Szczechla, E., Kim, T., Chhablani, G., Nayak, N., Datta, D., Chang, J., Jiang, M.T.J., Wang, H., Manica, M., Shen, S., Yong, Z.X., Pandey, H., Bawden, R., Wang, T., Neeraj, T., Rozen, J., Sharma, A., Santilli, A., Fevry, T., Fries, J.A., Teehan, R., Scao, T.L., Biderman, S., Gao, L., Wolf, T., Rush, A.M.: Multitask prompted training enables zero-shot task generalization. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=9Vrb9DOWI4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 568, + 482, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 568, + 482, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 568, + 482, + 601 + ], + "type": "text", + "content": "43. Sun, C., Sun, M., Chen, H.T.: Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5459-5469 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 601, + 482, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 601, + 482, + 622 + ], + "spans": [ + { + "bbox": [ + 130, + 601, + 482, + 622 + ], + "type": "text", + "content": "44. Sun, C., Han, J., Deng, W., Wang, X., Qin, Z., Gould, S.: 3d-gpt: Procedural 3d modeling with large language models. arXiv preprint arXiv:2310.12945 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 621, + 482, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 621, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 621, + 482, + 666 + ], + "type": "text", + "content": "45. Tancik, M., Casser, V., Yan, X., Pradhan, S., Mildenhall, B., Srinivasan, P.P., Barron, J.T., Kretzschmar, H.: Block-nerf: Scalable large scene neural view synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8248-8258 (2022)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 256, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 256, + 91, + 447, + 102 + ], + "type": "text", + "content": "SceneTeller: Language-to-3D Scene Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 544 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "text", + "content": "46. Tancik, M., Weber, E., Ng, E., Li, R., Yi, B., Kerr, J., Wang, T., Kristoffersen, A., Austin, J., Salahi, K., Ahuja, A., McAllister, D., Kanazawa, A.: Nerfstudio: A modular framework for neural radiance field development. In: ACM SIGGRAPH 2023 Conference Proceedings. SIGGRAPH '23 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 162, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 162, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 162, + 481, + 183 + ], + "type": "text", + "content": "47. Tang, J., Ren, J., Zhou, H., Liu, Z., Zeng, G.: Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 184, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 184, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 184, + 481, + 205 + ], + "type": "text", + "content": "48. Vachha, C., Haque, A.: Instruct-gs2gs: Editing 3d gaussian splats with instructions (2024), https://instruct-gs2gs.github.io/" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 205, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 205, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 134, + 205, + 481, + 248 + ], + "type": "text", + "content": "49. Wang, H., Du, X., Li, J., Yeh, R.A., Shakhnarovich, G.: Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 12619-12629 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 249, + 481, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 249, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 133, + 249, + 481, + 282 + ], + "type": "text", + "content": "50. Wang, Z., Lu, C., Wang, Y., Bao, F., Li, C., Su, H., Zhu, J.: Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. In: Advances in Neural Information Processing Systems (NeurIPS) (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 282, + 481, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 282, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 133, + 282, + 481, + 326 + ], + "type": "text", + "content": "51. Wei, J., Bosma, M., Zhao, V., Guu, K., Yu, A.W., Lester, B., Du, N., Dai, A.M., Le, Q.V.: Finetuned language models are zero-shot learners. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=gEZrGcozdqR" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 326, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 326, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 133, + 326, + 481, + 380 + ], + "type": "text", + "content": "52. Xu, Y., Chai, M., Shi, Z., Peng, S., Skorokhodov, I., Siarohin, A., Yang, C., Shen, Y., Lee, H.Y., Zhou, B., et al.: Discoscope: Spatially disentangled generative radiance fields for controllable 3d-aware scene synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4402-4412 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 380, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 380, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 133, + 380, + 481, + 413 + ], + "type": "text", + "content": "53. Yang, Z., Gan, Z., Wang, J., Hu, X., Lu, Y., Liu, Z., Wang, L.: An empirical study of gpt-3 for few-shot knowledge-based vqa. In: Proceedings of the AAAI Conference on Artificial Intelligence (AAAI). vol. 36, pp. 3081-3089 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 413, + 481, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 413, + 481, + 446 + ], + "spans": [ + { + "bbox": [ + 133, + 413, + 481, + 446 + ], + "type": "text", + "content": "54. Yi, T., Fang, J., Wu, G., Xie, L., Zhang, X., Liu, W., Tian, Q., Wang, X.: Gaussian dreamer: Fast generation from text to 3d gaussian splatting with point cloud priors. arXiv preprint arXiv:2310.08529 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 447, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 447, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 133, + 447, + 481, + 479 + ], + "type": "text", + "content": "55. Yu, A., Li, R., Tancik, M., Li, H., Ng, R., Kanazawa, A.: Plenoptrees for real-time rendering of neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (CVPR). pp. 5752-5761 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 479, + 481, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 479, + 481, + 512 + ], + "spans": [ + { + "bbox": [ + 133, + 479, + 481, + 512 + ], + "type": "text", + "content": "56. Zhou, X., Ran, X., Xiong, Y., He, J., Lin, Z., Wang, Y., Sun, D., Yang, M.H.: Gala3d: Towards text-to-3d complex scene generation via layout-guided generative gaussian splatting. arXiv preprint arXiv:2402.07207 (2024)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 512, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 512, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 133, + 512, + 481, + 544 + ], + "type": "text", + "content": "57. Zhu, J., Zhuang, P., Koyejo, S.: HIFA: High-fidelity text-to-3d generation with advanced diffusion guidance. In: The Twelfth International Conference on Learning Representations (ICLR) (2024), https://openreview.net/forum?id=IZMPWmcS3H" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 90, + 233, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 90, + 233, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 90, + 233, + 101 + ], + "type": "text", + "content": "B.M. Öcal et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/SceneVerse_ Scaling 3D Vision-Language Learning for Grounded Scene Understanding/72c8327d-dc6a-436c-89b3-89891fd18764_content_list.json b/2024/SceneVerse_ Scaling 3D Vision-Language Learning for Grounded Scene Understanding/72c8327d-dc6a-436c-89b3-89891fd18764_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8f3758c1fe6e7d8c9cc1df70bd44712d254bfeab --- /dev/null +++ b/2024/SceneVerse_ Scaling 3D Vision-Language Learning for Grounded Scene Understanding/72c8327d-dc6a-436c-89b3-89891fd18764_content_list.json @@ -0,0 +1,2346 @@ +[ + { + "type": "text", + "text": "SCENEVERSE: Scaling 3D Vision-Language Learning for Grounded Scene Understanding", + "text_level": 1, + "bbox": [ + 243, + 140, + 756, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Baoxiong Jia\\*, Yixin Chen\\*, Huangyue Yu, Yan Wang, Xuesong Niu, Tengyu Liu, Qing Li, and Siyuan Huang", + "bbox": [ + 225, + 210, + 781, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "* indicates equal contribution\nState Key Laboratory of General Artificial Intelligence, BIGAI https://scene-verse.github.io", + "bbox": [ + 290, + 255, + 709, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. 3D vision-language (3D-VL) grounding, which aims to align language with 3D physical environments, stands as a cornerstone in developing embodied agents. In comparison to recent advancements in the 2D domain, grounding language in 3D scenes faces two significant challenges: (i) the scarcity of paired 3D-VL data to support grounded learning of 3D scenes, especially considering complexities within diverse object configurations, rich attributes, and intricate relationships; and (ii) the absence of a unified learning framework to distill knowledge from grounded 3D data. In this work, we aim to address these major challenges in 3D-VL by examining the potential of systematically upscaling 3D-VL learning in indoor scenes. We introduce the first million-scale 3D-VL dataset, ScENEVERSE, encompassing 68K indoor scenes and comprising 2.5M vision-language pairs collected from both human annotations and our scalable scene-graph-based generation approach. We demonstrate that this scaling allows for a unified pre-training framework, Grounded Pretraining for Scenes (GPS), for 3D-VL learning. Through extensive experiments, we showcase the effectiveness of GPS by achieving state-of-the-art performance on existing 3D visual grounding and question-answering benchmarks. We also show that the data scaling effect is not limited to GPS, but is generally beneficial for models on tasks like 3D semantic segmentation. The vast potential of ScENEVERSE and GPS is unveiled through zero-shot transfer experiments in challenging 3D-VL tasks.", + "bbox": [ + 261, + 332, + 743, + 638 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: 3D Vision-Language $\\cdot$ Data Scaling $\\cdot$ Grounded Scene Understanding", + "bbox": [ + 261, + 650, + 743, + 678 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 703, + 375, + 719 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The foundation of human cognitive development lies in the grounding of language within the physical world [53,81,108]. Recent progress in Large Language Models (LLMs) [10,11,83] has markedly promoted the alignment between vision and language [3,59,75] utilizing billion-scale vision-language datasets [79,107]. However, with these advancements predominantly focusing on the 2D domain, the grounded understanding of 3D physical environments remains in an incipient stage [1,5,16]. Recognizing the pivotal role of grounded 3D experiences in", + "bbox": [ + 212, + 733, + 789, + 843 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/17e5b5f02f00295aaa4e2093431eabf3f15e1646a4650c7a71a4a3b69e233eae.jpg", + "image_caption": [ + "Fig. 1: Overview of ScENEVERSE. A million-scale 3D vision-language dataset that comprises over 68K various 3D indoor scenes and 2.5M aligned scene-language pairs in the form of scene caption, object caption, and object referral." + ], + "image_footnote": [], + "bbox": [ + 225, + 143, + 781, + 406 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "shaping human cognition [7,8], there is a compelling need to focus on exploring vision-language learning in the context of 3D scenes.", + "bbox": [ + 212, + 460, + 785, + 491 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Seeking insights from success in 2D vision-language (2D-VL) learning, a major factor to the success was the notable scale-up of paired vision-language data [15, 52, 79]. However, applying this experience directly from 2D to 3D is fraught with challenges. Primarily, 3D data collection heavily relies on the scanning device, making it inherently much more complex and expensive than gathering 2D images. Despite steady efforts to increase the volume of 3D scene data [9, 24, 66, 98], most datasets remain limited to thousands of scenes, substantially lagging behind the scale of existing 2D datasets. This gap is further widened by the inherent complexities of 3D scenes, which feature a multitude of object instances with diverse attributes, varying arrangements, and intricate inter-object relationships. These unique aspects of 3D scenes not only make the accurate description of objects and their relations more challenging but also considerably increase the number of language descriptions required for thorough scene depiction. Consequently, this presents a significant challenge in gathering sufficient and high-quality paired scene-language data for grounded scene understanding.", + "bbox": [ + 212, + 491, + 787, + 718 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To confront these challenges, we propose SCENEVERSE, the first million-scale dataset aimed at advancing 3D vision-language (3D-VL) learning for grounded scene understanding. At the scene level, we unify 3D scene data from existing datasets [9,24,66,76,85], aligning scenes and annotations from various capturing sources, and supplement the collection with synthetic scenes [28,106]. This compilation represents the most extensive 3D scene data gathered to date, amounting to 68K scenes. For language, we first present 97K newly-annotated referring expressions, the most extensive thus far. We additionally propose an", + "bbox": [ + 212, + 719, + 789, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "B. Jia and Y. Chen et al.", + "bbox": [ + 271, + 114, + 441, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "automated generation pipeline utilizing 3D scene graphs [4,86] and LLMs to create comprehensive, high-quality scene-language pairs. This refined collection, totaling 2.5M scene-language pairs, provides detailed and comprehensive descriptions of both object-level and scene-level descriptions within the 3D scene.", + "bbox": [ + 212, + 146, + 787, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We thoroughly investigate the potential offered by SCENEVERSE with large-scale pre-training, introducing Grounded Pre-training for Scenes (GPS), a novel and unified pre-training framework designed for scene and object-level alignment without auxiliary losses. Through multi-level contrastive alignment, we achieve significant performance boosts on 3D-VL tasks, such as grounding and question answering, setting new state-of-the-art results via a simple and effective pretraining process. We unveil the vast possibilities offered by SCENEVERSE and GPS in 3D-VL tasks in a zero-shot transfer setting. Additionally, we show that the scaling effect in SCENEVERSE is not limited to GPS, but generally benefits models in tasks like 3D semantic segmentation. At last, we offer deeper insights into the data-scaling in SCENEVERSE through extensive ablative experiments, pointing out future directions. Our main contributions are as follows:", + "bbox": [ + 212, + 207, + 789, + 387 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We introduce SCENEVERSE, the first million-scale 3D-VL dataset for grounded scene understanding. SCENEVERSE encompasses 68K 3D scenes coupled with 2.5M scene-language pairs, sourced through a combination of human annotation and automated generation methods. This represents a significant improvement in terms of data diversity and scale compared to prior datasets.", + "2. We propose GPS, a transformer-based model trained with multi-level scenetest alignment that achieves state-of-the-art results on existing 3D-VL grounding and question-answering benchmarks by pre-training on ScENEVERSE.", + "3. We demonstrate that with the data scale-up and model design, our pre-trained GPS exhibit emerging zero-shot generalization capabilities in grounded scene understanding. We also show that this scaling effect is not limited to GPS, but is generally beneficial for models on tasks like semantic segmentation." + ], + "bbox": [ + 222, + 388, + 787, + 570 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 594, + 387, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Datasets for Grounded 3D Understanding Obtaining aligned 3D-language data is a inherently difficult task. In 3D object modeling, pioneering works like ShapeNet [14] sourced 3D assets from online repositories, leading to a proliferation of high-quality 3D object datasets [23, 69, 91]. Notably, recent developments include internet-scale data collection with Objaverse [26, 27], accompanied by the integration of object-level captions [93] for 3D-language alignment. Models trained on these datasets demonstrate an enhanced understanding of objects, evident in classification [60], generation [61], and captioning tasks [63].", + "bbox": [ + 212, + 627, + 787, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In contrast, developing datasets for grounded 3D scene understanding is even more challenging due to the extensive requirements for scene acquisition and annotation. Existing works curate RGB-D and scanned indoor scene datasets [9,13, 24,66,76,85] and synthetic scenes [28,50,95,106] used for benchmarking tasks like 3D object detection and segmentation [31,46,67,80,84]. These semantically labeled scenes are subsequently used in fine-grained scene grounding tasks like object", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Scaling 3D Vision-Language Learning for Grounded Scene Understanding", + "bbox": [ + 241, + 114, + 730, + 130 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/1ef0c5f8d02b680dafa1e8af6603cca974db298dbf4c96ede221e70067913dbe.jpg", + "table_caption": [ + "Table 1: Comparison of SCENEVERSE with existing 3DVL Datasets. SCENEVERSE expands the data scale of prior work by order of magnitude. \"VG\" stands for Visual Grounding, \"QA\" for Question Answering, \"PT\" for Pre-training and \"MT\" for Multi-tasking. \"Anno.\" denotes language from human annotations and \"Syn.\" for template-based or LLM generated descriptions." + ], + "table_footnote": [], + "table_body": "
Dataset3D Scene Obj.TaskObj. CaptionScene CaptionObj. ReferralQuality CheckNew Anno.Syn.Existing Anno.Syn.Total
ScanRefer [16]VGXX✓✓52K---52K
ReferIt3D [1]| |VGXX✓✓42K200K--242K
ScanQA [5]1.5K33KQA---✓27K---27K
SQA3D [65]QA---✓33K---33K
Multi3DRefer [104]| |VGXX✓✓-10K52K-62K
Cap3D [63]-666KVGX✓XX58K666K--724K
ScanScribe [109]3K56KPTXX✓X-90K94K94K278K
3D-LLM [40]1.5K186KMT✓✓✓X-659K--659K
EmbodiedScan [87]5K890KVGXX✓X-970K--970K
LEO [41]3K56KMT✓✓✓✓-188K235K90K513K
SCENEVERSE68K1.5MVG✓✓✓✓96K2.1M94K200K2.5M
", + "bbox": [ + 218, + 214, + 784, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "referral [1, 16, 104], captioning [17, 19, 22, 99], vision-language-navigation [39, 64, 73, 88] and reasoning [5, 40, 65]. Recent works exploit the representation of 3D scene graph (3DSG) [4, 20, 78, 86], which concisely describes scenes with hierarchical structures. This representation is notably advantageous for planning [2, 77] and captioning [35], owing to its compatibility with LLMs for flexible description generation [40, 41]. Nonetheless, as shown in Tab. 1, most datasets are constrained in both scene and language scales, underscoring the need for scaling up fine-grained and aligned scene-language data to enhance grounded scene understanding.", + "bbox": [ + 212, + 388, + 787, + 512 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Vision-Language Learning Recent years have witnessed tremendous progress in 2D-VL [3, 21, 25, 57, 59, 75], empowered by transformer-based pre-training models [11, 29, 72] and large-scale image-language datasets [15, 79]. A central theme across 2D-VL domains is the effectiveness of data scaling [49], as demonstrated by improved alignment and expanded capabilities in open-vocabulary understanding [33, 51, 54, 58] through a contrastive pre-training pipeline [75].", + "bbox": [ + 212, + 522, + 787, + 613 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, in grounded scene understanding, the primary challenge for models has been the limited availability of paired 3D scene-language data, which restricts the application of insights drawn from 2D-VL. Current models for 3D scene grounding [6,18,37,43,45,62,92,97,105] heavily rely on task-specific knowledge in both model and loss designs or advanced optimization strategies [109]. To bridge this gap, there has been a growing emphasis on employing pre-trained 2D-VL models for 3D-VL [36,38,74,82,93,102,103]. Yet, these models mostly draw information available from 2D-VL models (e.g., object attribute, affordance, etc.), falling short on capturing crucial 3D information like object spatial relationships which are necessary for more fine-grained tasks such as grounded human-scene [20, 44,47,48,89,90] and robot-scene interactions modeling [34,55,68,70]. This urges the need for a multi-level alignment between language and 3D scenes, particularly regarding 3D-specific information. Considering the nascent stage of existing 3D pre-training methods [30,94,109,110], we believe SCENEVERSE and GPS have the potential to spearhead new avenues in 3D-VL research.", + "bbox": [ + 212, + 613, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "B. Jia and Y. Chen et al.", + "bbox": [ + 271, + 114, + 441, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 SCENEVERSE", + "text_level": 1, + "bbox": [ + 215, + 143, + 380, + 160 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SCENEVERSE is designed for grounded scene understanding with 3D scenes curated from diverse existing datasets of both real and synthetic environments. Regarding language, we employ both human annotation and a novel automated generation pipeline to collect comprehensive and high-quality language for both object-level and scene-level descriptions. We provide details regarding data collection in the following sections.", + "bbox": [ + 212, + 172, + 787, + 265 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Scene Curation", + "text_level": 1, + "bbox": [ + 214, + 282, + 388, + 297 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address the scarcity of available 3D scene data, we construct SCENEVERSE by unifying 3D scene data from various existing datasets. We use real-world scene datasets, including ScanNet [24], ARKitScenes [9], HM3D [76], 3RScan [85] and MultiScan [66], alongside synthetic environments from Structured3D [106] and ProcTHOR [28]. The inclusion of these synthetic datasets is mainly motivated by their potential as scalable data sources for 3D-VL alignment. To ensure cohesion across various sources, we conduct preprocessing steps such as room segmentation, point subsampling, axis alignment, normalization, and semantic label alignment. Each scan is represented by a point cloud $\\mathbf{P} \\in \\mathbb{R}^{N \\times 8}$ , wherein each point is defined by its 3D coordinates, RGB color, instance id, and semantic label. In total, we curate 68,406 3D scenes in SCENEVERSE.", + "bbox": [ + 212, + 304, + 787, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Referral Annotation by Humans", + "text_level": 1, + "bbox": [ + 214, + 489, + 529, + 505 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the curated scenes of SCENEVERSE, we present the most comprehensive set of human-annotated, context-rich object referrals to date, serving as a valuable benchmark for assessing grounded scene understanding capabilities. The human annotations contain 96,863 descriptions in ARKitScenes [9], HM3D [76] and MultiScan [66]. During the annotation process, one human annotator was assigned to write at least 20 words to distinctly refer to a single 3D object within a 3D scene. Each referral text then undergoes independent verification by two additional reviewers, both mandated to accurately locate the referenced object based on the 3D scene and the annotated referral text. Any object referrals that do not pass the verification by either reviewer are flagged for re-annotation.", + "bbox": [ + 212, + 512, + 787, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3 3D Scene Graph Construction", + "text_level": 1, + "bbox": [ + 214, + 681, + 511, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our 3D scene graph is defined as a hierarchical graph $\\mathcal{G} = (\\mathcal{V},\\mathcal{E})$ . Each node $v\\in \\mathcal{V}$ represents one distinct 3D object instance, parameterized by its centroid $\\pmb {p}_i\\in \\mathbb{R}^3$ and bounding box size of $\\pmb {b}_i = (b_x,b_y,b_z)\\in \\mathbb{R}^3$ . The edges $\\mathcal{E}$ represent spatial relationships between nodes. To construct the scene graph $\\mathcal{G}$ , we instantiate the nodes with the instance annotation from the point clouds and assign object classes with their corresponding semantic labels. Following prior work [1,86], we consider the Vertical proximity, Horizontal proximity and Multi-object Relationships as spatial relations. For a more detailed description of the scene graph construction and relationship determination, please refer to supplementary.", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Scaling 3D Vision-Language Learning for Grounded Scene Understanding", + "bbox": [ + 241, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d6fe1d32398d30be9ead3924f24ac73f57a2feb2f73b4c29ecba65b7046d0bfd.jpg", + "image_caption": [ + "(a) 3D Scene", + "Fig. 2: SCENEVERSE collection and statistics. Given a 3D scene (a), our automated pipeline (c) generates three types of description including scene caption, object caption and object referral. (b) SCENEVERSE data comparison and composition." + ], + "image_footnote": [], + "bbox": [ + 222, + 145, + 439, + 248 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "一", + "bbox": [ + 253, + 263, + 271, + 270 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Scene Cap", + "bbox": [ + 276, + 263, + 326, + 270 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Sub-graph Context", + "bbox": [ + 225, + 273, + 303, + 282 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "{scene_type':Bedroom',", + "bbox": [ + 230, + 284, + 331, + 290 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "'object_count': {'nightstand':2,...},", + "bbox": [ + 233, + 291, + 331, + 297 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "'relation': {nightstand', 'on', floor},", + "bbox": [ + 233, + 299, + 333, + 305 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "['backback', in front of, b", + "bbox": [ + 233, + 306, + 333, + 311 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Summary", + "bbox": [ + 225, + 313, + 279, + 321 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Prompt: Provide a summary for a scene from a given", + "bbox": [ + 230, + 321, + 385, + 327 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "scene graph delimited by triple backticks,", + "bbox": [ + 230, + 328, + 352, + 334 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Response: In this bedroom, there are two nightstands, ...", + "bbox": [ + 230, + 335, + 403, + 340 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The backpack is in front of the nightstand as well . The roomapproves to be functional with the nightstands providing", + "bbox": [ + 230, + 342, + 403, + 348 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "storage space and the telephone for communication", + "bbox": [ + 230, + 349, + 380, + 356 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "", + "bbox": [ + 441, + 251, + 470, + 258 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(b) Language radar", + "bbox": [ + 473, + 251, + 537, + 258 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ar chart and", + "bbox": [ + 539, + 251, + 578, + 258 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Object Caption", + "bbox": [ + 470, + 263, + 537, + 270 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "BLIP2 Captions", + "bbox": [ + 416, + 273, + 478, + 282 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "1. A bed in a hotel room. (0.85)", + "bbox": [ + 419, + 282, + 513, + 289 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "2. A white comforter on a bed. (0.83)", + "bbox": [ + 419, + 289, + 532, + 296 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3. A bed with a striped comforter. (0.83)", + "bbox": [ + 419, + 296, + 532, + 301 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "N.A picture of cat. (0.63)", + "bbox": [ + 419, + 303, + 493, + 310 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Summary", + "bbox": [ + 416, + 311, + 478, + 321 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Prompt: Summarize the captions below. The summary", + "bbox": [ + 419, + 321, + 580, + 328 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "should be a description of the (object). Focus on the", + "bbox": [ + 419, + 328, + 571, + 334 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "{object's attributes, like color, shape, n", + "bbox": [ + 419, + 335, + 532, + 340 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Identity and correct the potential errors ...", + "bbox": [ + 419, + 342, + 575, + 348 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Response : The bed is in a motel room with a stripedcomforter . It has a white comforter and a blanket on it", + "bbox": [ + 419, + 349, + 573, + 356 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The bed is also in a room with a bedside table.", + "bbox": [ + 419, + 356, + 578, + 363 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(c) Automated language generation", + "bbox": [ + 444, + 367, + 573, + 376 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Object R", + "bbox": [ + 656, + 263, + 694, + 270 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Relationship Triplets", + "bbox": [ + 604, + 273, + 687, + 282 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "1. {table', 'chair', 'left'}", + "bbox": [ + 607, + 282, + 673, + 287 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "2.{'bed', ('lamp', 'mini fridge'), 'between'}", + "bbox": [ + 607, + 289, + 725, + 295 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Template-based Referral", + "bbox": [ + 607, + 296, + 702, + 301 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "1. The table is to the left of the chair. \n2. The chair is in the right of the chair.", + "bbox": [ + 604, + 303, + 740, + 310 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "2. It's a bed in the m", + "bbox": [ + 604, + 311, + 666, + 316 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Rephrasing", + "bbox": [ + 607, + 318, + 666, + 323 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Prompt: Rewrite the following sentence using one random word.", + "bbox": [ + 607, + 323, + 777, + 329 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "sentence structure. Focus on the", + "bbox": [ + 607, + 329, + 707, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "about the {target_object}", + "bbox": [ + 607, + 335, + 679, + 340 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Response:", + "bbox": [ + 607, + 342, + 642, + 345 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "1. The table is situated to the left of the armchair.", + "bbox": [ + 607, + 347, + 750, + 352 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "2. The bed occupies the space between the lamp and the wall.", + "bbox": [ + 607, + 352, + 774, + 359 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "mini fridge, creating a cozy atmosphere.", + "bbox": [ + 607, + 359, + 723, + 364 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4 Language Generation with LLMs", + "text_level": 1, + "bbox": [ + 215, + 431, + 534, + 446 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The scene-language pairs in ScENEVERSE aim to capture varying aspects of the 3D scene, including detailed object attributes in object captioning, spatial relationships between objects in object referral, and global scene descriptions in scene captioning. Based on 3D scene graphs, we utilize both templates and LLMs to automatically generate descriptions on these three granularities.", + "bbox": [ + 212, + 453, + 787, + 529 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Object Captioning Object captions aim to provide detailed descriptions of an object's visual and physical properties, facilitating object-level grounding with its distinctive features. Given the multi-view images, we utilize the point cloud of the object $v \\in V$ to identify its occurrence in the images through rendering. The images are then cropped with the rendered bounding boxes and processed through BLIP2 [56] to generate initial object captions. We select the top 10 sentences with the highest CLIP [75] similarity score and minimal occlusion and utilize an LLM to obtain a refined and coherent summary of the captions. The detailed object captioning pipeline is illustrated in supplementary.", + "bbox": [ + 212, + 532, + 787, + 669 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Object Referral Object relationship captions refer to objects by articulating their spatial relationships in the scene. Spatial relationship triplets $(v_{i}, v_{j}, e_{ij})$ are first extracted from the constructed 3D scene graph. We design various templates to generate descriptions for each relationship type, assigning the entities in the form of (target-object, spatial-relation, anchor-object(s)). This results in examples like \"the chair is next to the li2023behaviorarmchair\", \"facing the sofa, there is a suitcase far to the right of the shoes\", and \"the fridge is between cabinet and sofa\". Our designed templates span passive and active tenses, as well as inversion clauses, contributing to the richness of the generated text. To enhance the descriptions' naturalness, we employ LLM for sentence rephrasing. Statistics for the descriptions before and after rephrasing are presented in Fig. 2 (b).", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "B. Jia and Y. Chen et al.", + "bbox": [ + 271, + 114, + 441, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Scene Captioning The scene-level captions emphasize global information, portraying the key objects in the scene along with their attributes and functionalities. We use the constructed 3D scene graph and prompt LLMs to generate these captions. We random sample a subset of edges and nodes from the scene graph each time as the scene context to enhance the diversity of scene captions. The object counts are also provided as LLM prompts, together with the room type and object attributes if such annotations are available in the dataset.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.5 Data Quality and Statistics", + "text_level": 1, + "bbox": [ + 214, + 276, + 488, + 291 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Data Quality The 96K human-annotated set of ScENeVErSE is collected through AMT, where 82 humans are employed for annotation and 18 for verification. All final annotations passed the reference verification, with a re-annotation rate of $4.8\\%$ . For our automatic language generation pipeline, we conduct extensive prompt tuning and iterate with human feedback for LLMs on object captioning, summary, and rephrasing. To verify the efficacy of the pipeline, we conduct a quality check where 12K generated object-level descriptions are randomly selected for human verification. Results demonstrate a $96.93\\%$ pass rate, surpassing that in ReferIt3D [1] with $86.1\\%$ pass rate on 2K samples.", + "bbox": [ + 212, + 301, + 787, + 438 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Statistics In total, SCENEVERSE comprises a total of 68,406 room-level 3D scans, with the source composition shown in Fig. 2 (b). The dataset contains 1.5M object instances ranging in 2290 object categories. Our generated 3D scene graph comprises 21 types of relationships following prior work [1,86]. For the language descriptions, we generate 1M template-based texts and 1M sentences rephrased by Llama [83] and GPT-3.5 [71]. As can be seen from the radar chart and examples in Fig. 2, the diversity of the LLM-refined data, particularly in sentence length and variety, closely aligns with the characteristics of annotated descriptions, surpassing the template-based data. Together with existing sources (294K) and our newly annotated set (96K), SCENEVERSE contains 2.5M scene-language pairs in total. All the rephrasing and summary prompts, along with the complete set of relationships, are detailed in supplementary.", + "bbox": [ + 212, + 446, + 789, + 628 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Grounded Pre-training for Scenes", + "text_level": 1, + "bbox": [ + 214, + 652, + 578, + 669 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we introduce GPS, an efficient transformer-based model trained with multi-level contrastive losses for aligning 3D scenes and texts. As shown in Fig. 3, we echo the language descriptions collected at different granularities to form contrastive objectives at both object-level, referral-object-level, and scene-level in GPS. We describe the design of each level in the following sections.", + "bbox": [ + 212, + 684, + 787, + 760 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1 Object-level Grounding", + "text_level": 1, + "bbox": [ + 214, + 782, + 457, + 799 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Given a 3D scene point cloud $\\mathcal{S}$ , we use an off-the-shelf 3D object segmentation model to decompose it into a bag of $N$ objects $\\mathcal{S} = \\{\\pmb{o}_1, \\pmb{o}_2, \\dots, \\pmb{o}_n\\}_{i=1}^N$ . We extract", + "bbox": [ + 212, + 809, + 785, + 842 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Scaling 3D Vision-Language Learning for Grounded Scene Understanding", + "bbox": [ + 241, + 114, + 730, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/65c4acf2fbe695e949ed45de515b28bdd9cad216fdb934a32bfdd220c1fa25da.jpg", + "image_caption": [ + "Fig. 3: Overview of GPS model. We use contrastive alignment at three levels $\\mathcal{L}_{\\mathrm{obj}}$ , $\\mathcal{L}_{\\mathrm{scene}}$ , and $\\mathcal{L}_{\\mathrm{ref}}$ and a masked language modeling objective $\\mathcal{L}_{\\mathrm{MLM}}$ for model learning." + ], + "image_footnote": [], + "bbox": [ + 218, + 145, + 787, + 306 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "object features $\\{\\pmb{f}_i^O\\}$ with an object point cloud encoder and text features $\\{\\pmb{f}_i^T\\}$ by feeding object-captions $\\{\\pmb{T}_i^{\\mathrm{obj}}\\}$ into a frozen language model. Following [93], we perform cross-modal alignment on the object features and text features via:", + "bbox": [ + 212, + 349, + 787, + 398 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {o b j}} = - \\frac {1}{2} \\sum_ {(p, q)} \\left(\\log \\frac {\\exp \\left(D ^ {\\mathrm {o b j}} (p , q)\\right)}{\\sum_ {r} \\exp \\left(D ^ {\\mathrm {o b j}} (p , r)\\right)} + \\log \\frac {\\exp \\left(D ^ {\\mathrm {o b j}} (p , q)\\right)}{\\sum_ {r} \\exp \\left(D ^ {\\mathrm {o b j}} (r , q)\\right)}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 238, + 407, + 787, + 450 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $D^{\\mathrm{obj}}(p,q) = (\\pmb{f}_p^O\\pmb{f}_q^T /\\tau)$ denotes the dot product between object and text features and $(p,q)$ denotes a pair of aligned object-text pair in the training batch and $r$ iterates over all object-text pairs in the training batch. Similar to CLIP [75], we use a learnable temperature parameter $\\tau$ to facilitate model learning.", + "bbox": [ + 212, + 460, + 789, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 Scene-level Grounding", + "text_level": 1, + "bbox": [ + 215, + 545, + 447, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "With aligned object features, we encode the scene by incorporating object spatial locations into the extracted object features. Specifically, we use a spatial transformer model to encode extracted object features $\\{\\pmb{f}_i^O\\}$ with their spatial location features $\\{l_i\\}$ following [18, 109]:", + "bbox": [ + 212, + 570, + 787, + 632 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {f} ^ {S} = \\operatorname {S p a t i a l A t t n} (\\{\\boldsymbol {f} _ {i} ^ {O} \\}, \\{\\boldsymbol {l} _ {i} \\})\n$$\n", + "text_format": "latex", + "bbox": [ + 393, + 641, + 607, + 660 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\{\\pmb{f}_i^S\\}$ denotes the feature of object $\\pmb{o}_i$ after encoding with spatial location features. To perform scene-level alignment, we operate on these scene-level object features $\\{\\pmb{f}_i^S\\}$ and align it with the scene caption $\\pmb{T}^{\\mathrm{scene}}$ . Specifically, we feed the object features into a projection layer and use max-pooling over all object features to obtain the scene feature $\\pmb{g}^S$ . Similar to object-level grounding, we pass the scene caption through a tunable language model to obtain text feature $\\pmb{g}^T$ and perform scene-level contrastive alignment through:", + "bbox": [ + 212, + 670, + 787, + 777 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s c e n e}} = - \\frac {1}{2} \\sum_ {(p, q)} \\left(\\log \\frac {\\exp \\left(D ^ {\\text {s c e n e}} (p , q)\\right)}{\\sum_ {r} \\exp \\left(D ^ {\\text {s c e n e}} (p , r)\\right)} + \\log \\frac {\\exp \\left(D ^ {\\text {s c e n e}} (p , q)\\right)}{\\sum_ {r} \\exp \\left(D ^ {\\text {s c e n e}} (r , q)\\right)}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 230, + 785, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "B. Jia and Y. Chen et al.", + "bbox": [ + 271, + 114, + 441, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $D^{\\mathrm{scene}}(p,q) = (\\pmb{g}_p^S\\pmb{g}_q^T / \\tau)$ denotes the dot product between scene feature $\\pmb{g}_p^S$ and scene caption feature $\\pmb{g}_q^T$ for each pair of aligned scene-text pairs in the training batch and $r$ iterates over all scene-text pairs in the training batch.", + "bbox": [ + 212, + 145, + 787, + 194 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3 Referral-object-level Grounding", + "text_level": 1, + "bbox": [ + 215, + 210, + 526, + 227 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To model the relationships revealed in referring expressions, we employ a self-attention-based reasoning transformer for grounding object referrals in scenes. This transformer takes in scene-object features $\\{\\pmb{f}_i^S\\}$ and an object referral $T^{\\mathrm{ref}}$ and performs self-attention to learn relationships between text descriptions and object relationships. We use the same tunable language encoder as in scene-level grounding for extracting per-object referral features. We pass this text feature together with scene-object features into the self-attention transformer to obtain the aligned object features $\\pmb{h}_i^S$ and the sentence-level referral feature $\\pmb{h}^T$ . We then perform the referral-object-level contrastive alignment following:", + "bbox": [ + 212, + 231, + 787, + 368 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {r e f}} = - \\log \\frac {\\exp \\left(\\bar {\\boldsymbol {h}} ^ {S} \\boldsymbol {h} ^ {T} / \\tau\\right)}{\\sum_ {p} \\exp \\left(\\boldsymbol {h} _ {p} ^ {S} \\boldsymbol {h} ^ {T} / \\tau\\right)}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 390, + 373, + 787, + 411 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\bar{h}^S$ denotes the feature of the referred object, $p$ iterates over all objects within the same scene. Notably, in contrast to inter-scene contrast that was done in object- and scene-level alignment, we force the selection of positive pairs to be within the same scene to provide intra-scene contrast for fine-grained object grounding. This mimics the success of intra-image and inter-image contrasts commonly used for region-word alignment in 2D-VL models [101].", + "bbox": [ + 212, + 415, + 787, + 507 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To learn the multi-level alignment between 3D scenes and language, we first train the point cloud encoder with an object-level grounding objective to obtain a good feature initialization for grounding objects in scenes. During the scene grounding stage, we train our inter- and intra-scene objectives together with a masked language modeling loss $\\mathcal{L}_{\\mathrm{MLM}}$ over the inputted object-referral texts to tune the parameters within the language encoder and self-attention transformer. Above all, the learning of GPS could be summarized as optimizing:", + "bbox": [ + 212, + 507, + 789, + 613 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {o b j}} + \\mathcal {L} _ {\\mathrm {s c e n e}} + \\mathcal {L} _ {\\mathrm {r e f}} + \\mathcal {L} _ {\\mathrm {M L M}}.\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 619, + 624, + 636 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 654, + 375, + 671 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this section, we present experimental results addressing the following questions:", + "bbox": [ + 215, + 681, + 787, + 696 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. How effective is the data scaling in ScENEVERSE for 3D visual grounding? Does the scale-up benefit common 3D-VL tasks (e.g., 3D question answering, open-vocabulary 3D semantic segmentation) and pre-training-based models?", + "2. How well is the GPS pre-training pipeline for 3D-VL tasks? Does it exhibit similar properties of 2D-VL models in 3D-VL tasks?", + "3. What is offered by ScENEVERSE and GPS and what is missing?" + ], + "bbox": [ + 220, + 700, + 787, + 789 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In the following sections, we describe in detail the model performance regarding these key topics. Due to the page limit, we direct readers to the supplementary for implementation details, qualitative results, and more experimental analyses.", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Scaling 3D Vision-Language Learning for Grounded Scene Understanding", + "bbox": [ + 241, + 114, + 732, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/3eade42b36a095ddeb66dc8843922cdae95f3df7bc5b8a0601714842be8aa2b3.jpg", + "table_caption": [ + "Table 2: 3D visual grounding results on Nr3D, Sr3D, and ScanRefer. We use \"pre-train\" for our model trained on ScENeVERse w/o additional fine-tuning, and \"fine-tune\" for its data-specific fine-tuned version. Best results are highlighted in bold." + ], + "table_footnote": [], + "table_body": "
MethodNr3DSr3DScanRefer Acc@0.5
OverallEasyHardV-Dep.V-Indep.OverallEasyHardV-Dep.V-Indep.OverallUniqueMultiple
3DVG-Trans [105]40.848.534.834.843.751.454.244.944.651.734.760.628.4
TGNN [42]37.344.230.635.838.045.048.536.945.845.029.756.823.2
TransRefer3D [37]48.056.739.642.550.757.460.550.249.957.7---
InstanceRefer [100]38.846.031.834.541.948.051.140.545.848.132.966.824.7
FFL-3DOG [32]41.748.235.037.144.7-----34.067.925.7
LAR [6]48.958.442.347.452.159.463.051.250.059.1---
SAT [97]56.564.948.454.457.657.961.250.049.258.330.150.825.2
3D-SPS [62]51.558.145.148.053.262.656.265.449.263.237.066.729.8
3DJCG [12]----------37.364.330.8
BUTD-DETR [45]54.660.748.446.058.067.068.663.253.067.639.866.335.1
MVT [43]59.567.452.759.160.364.566.958.858.464.733.366.525.3
ViL3DRel [18]64.470.257.462.064.572.874.967.963.873.237.768.630.7
EDA [92]52.158.246.150.253.168.170.362.954.168.742.368.637.6
3D-VisTA (scratch) [109]57.565.949.453.759.469.672.163.657.970.141.570.934.8
3D-VisTA [109]64.272.156.761.565.176.478.871.358.977.345.875.139.1
Ours (scratch)58.767.050.955.859.868.470.563.453.169.040.471.334.7
Ours (pre-train)55.262.848.045.558.874.176.468.554.175.047.177.441.6
Ours (fine-tuned)64.972.557.856.967.977.580.171.662.878.248.177.942.7
", + "bbox": [ + 217, + 186, + 784, + 400 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.1 3D Visual Grounding", + "text_level": 1, + "bbox": [ + 215, + 422, + 439, + 436 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Settings We evaluate our model on three commonly-used datasets for 3D visual grounding: ScanRefer [16], Nr3D, and Sr3D [1]. For Nr3D and Sr3D, we follow Achlioptas et al. [1] and report the grounding accuracies of models using ground-truth object masks. For ScanRefer, we follow Zhu et al. [109] and use Mask3D [80] to generate object proposals. Results are reported as Acc@0.5 to evaluate the correctness of predictions whose object bounding boxes overlap the ground truth with IoU $>0.5$ . For comparisons, we compare with existing baselines by providing the results of pre-trained GPS and dataset-specific fine-tuned GPS. Please see more details in the supplementary.", + "bbox": [ + 212, + 465, + 787, + 603 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Results and Analyses As shown in Tab. 2, GPS trained on SCENEVERSE achieves state-of-the-art results on all existing 3D-VL grounding benchmarks. Initially, when GPS is trained directly on the training sets of benchmark datasets, labeled as Ours (scratch), it underperforms compared to existing models that employ more complex structures or loss designs. This result underscores the data-intensive nature of the contrastive alignment paradigm. However, when presented with extensive training data in SCENEVERSE, the results of our model without additional fine-tuning, i.e., Ours (pre-train), significantly improves and already achieves state-of-the-art results on benchmarks like ScanRefer. Moreover, the dataset-specific fine-tuned model, i.e., Ours (fine-tuned), consistently outperforms existing baselines with only a simple projection MLP added on top of the pretrained model, jointly optimized during fine-tuning without any other auxiliary architecture or loss objective. These results underscore the strong potential of both the SCENEVERSE and GPS for 3D-VL tasks.", + "bbox": [ + 212, + 628, + 787, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "B. Jia and Y. Chen et al.", + "bbox": [ + 271, + 114, + 439, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/02bd606ef257629e070b59138bca83c42b7b8d6f74874532d021ea6e0f12d5af.jpg", + "table_caption": [ + "Table 3: Zero-shot transfer on existing benchmarks. \"SR\" stands for ScanRefer." + ], + "table_footnote": [], + "table_body": "
MethodNr3DSr3DSR@0.25SR@0.5
3D-VisTA (scratch)57.569.645.941.5
3D-VisTA (zero-shot)35.231.233.229.6
3D-VisTA (zero-shot text)43.136.141.136.4
Ours (scratch)58.768.444.540.4
Ours (zero-shot)32.433.335.231.1
Ours (zero-shot text)41.938.140.735.8
", + "bbox": [ + 218, + 172, + 504, + 263 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/91857b136dd51597b4d4df6ef6d5e5f85a3a97076424c4002583d9fa5b69d03f.jpg", + "table_caption": [ + "Table 4: Zero-shot transfer on ScENeVERSE-val. Evaluation uses GT object proposals following Nr3D/Sr3D." + ], + "table_footnote": [], + "table_body": "
MethodOverallEasyHardV-Dep.V-Indep.
3D-VisTA (scratch)40.753.121.637.344.3
3D-VisTA (zero-shot)52.959.635.453.752.2
3D-VisTA (zero-shot text)58.170.039.652.564.1
Ours (scratch)38.550.220.833.743.9
Ours (zero-shot)59.269.444.053.166.3
Ours (zero-shot text)60.670.945.154.867.3
", + "bbox": [ + 521, + 186, + 784, + 262 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.2 Zero-Shot Transfer", + "text_level": 1, + "bbox": [ + 215, + 282, + 418, + 296 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Settings To better evaluate the effectiveness of both the SCENEVERSE data and the GPS model, we further perform zero-shot transfer experiments to test the models' capability in 4 benchmarks, ScanRefer, Sr3D, Nr3D, and SCENEVERSE- val. We create SCENEVERSE-val using $8.5K$ annotated object referrals of 271 scenes in MultiScan, and randomly split the scenes following a 4:1 train / test split for creating the held-out test set. We mainly consider 2 specific transfer settings in our experiments: (i) zero-shot: models trained by removing all the scenes from the target dataset, tested on held-out unseen scenes, and (ii) zero-shot text: Models trained on data that include the training set of scenes from the target dataset, yet tested exclusively with unseen scene-text distribution. Specifically, for the zero-shot text setting, we use the generated texts in SCENEVERSE as fine-tuning sources for the zero-shot model. We mainly compare our model against a recent pre-training-based model 3D-VisTA. See more details on experimental setting and implementation in the supplementary.", + "bbox": [ + 212, + 303, + 787, + 516 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Results and Analyses We present the results of zero-shot transfer experiments in Tab. 3 and Tab. 4 with the following key observations:", + "bbox": [ + 212, + 520, + 785, + 550 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Our GPS model demonstrates superior generalization to unseen scenes compared to the 3D-VisTA model. In zero-shot transfer scenarios, our model consistently outperforms 3D-VisTA across established benchmarks and ScEN-EVERSE-val. This indicates the effectiveness of contrastive alignment over traditional classification objectives, aligning with the advancements seen in 2D-VL models for open-vocabulary grounding and transfer capabilities", + "- SCENEVERSE dataset substantially enhances 3D-VL grounding capabilities through zero-shot transfer, especially when provided with relatively limited training data, i.e., SCENEVERSE-val. As demonstrated in Tab. 4, there is a significantly improved performance when comparing models trained on SCENEVERSE in a zero-shot manner to those trained from scratch. This indicates that SCENEVERSE can effectively capture knowledge for general 3D scene grounding. Consequently, this underscores its potential as a go-to pre-training dataset for 3D-VL tasks.", + "- The impact of our extensive collection and scalable generation of scene-text pairs is further evidenced by the results in the zero-shot text setting. Notably, as shown in Tab. 3, these automatically generated scene-text pairs supply ample knowledge for comprehending the scene distribution. This contributes significantly to the substantial improvement over the zero-shot performance." + ], + "bbox": [ + 225, + 551, + 787, + 838 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Scaling 3D Vision-Language Learning for Grounded Scene Understanding", + "bbox": [ + 241, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/c0887a789106b61f253d012ada1bbec8c39cd2a220e84ac7eda6443b80b81066.jpg", + "table_caption": [ + "Table 5: 3D question answering re- Table 6: Existing 3D backbones pre-sults on ScanQA and SQA3D. We re-trained on SCENEVERSE for openport EM@1 score on ScanQA and SQA3D vocabulary 3D semantic segmentation on evaluation sets. ScanNet. \"SPUNet\" denotes SparseUNet pro" + ], + "table_footnote": [], + "table_body": "
ModelScanQASQA3D
valw/objw/o obj
ScanRefer+MCAN [5]18.620.619.0-
ScanQA [5]20.323.520.946.6
SQA3D [65]---47.2
3D-VisTA [109]22.427.023.048.5
3D-LLM [40]20.519.1--
Ours22.725.023.549.9
", + "bbox": [ + 218, + 199, + 483, + 306 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/15fad004489a8bad2b2000fb494cc9b76d144cee90a07e36f80042b8b8505a7e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelNetworkmIoUΔmAccΔ
OpenScene [74]SPUNet1657.2-69.9-
PLA [30]SPUNet1617.7-33.5-
RegionPLC [94]SPUNet1656.9-75.6-
RegionPLC+SCENEVERSESPUNet1658.2+1.7%77.3+2.2%
OpenScene [74]SPUNet3257.8-70.3-
PLA [30]SPUNet3219.1-41.5-
RegionPLC [94]SPUNet3259.6-77.5-
RegionPLC+SCENEVERSESPUNet3261.0+2.3%79.7+2.8%
", + "bbox": [ + 488, + 214, + 774, + 310 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.3 Additional 3D-VL Tasks", + "text_level": 1, + "bbox": [ + 215, + 319, + 464, + 335 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Settings We evaluate the effectiveness of GPS and ScENEVERSE on additional 3D-VL tasks: (i) 3D question answering (3D-QA) on ScanQA [5] and SQA3D [65], and (ii) open-vocabulary 3D semantic segmentation (OV-Seg) on ScanNet.", + "bbox": [ + 212, + 340, + 785, + 386 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- In the 3D-QA task, we follow Zhu et al. [109] and evaluate models over the exact match metric (EM@1) on the validation and test sets of ScanQA, as well as the test set of SQA3D. We pre-train GPS on SCENEVERSE and fine-tune the model on the 3D-QA dataset to compare with state-of-the-art models.", + "- In the OV-Seg task, as GPS builds upon an object-centric design and thus is not directly applicable to semantic segmentation, we consider testing the effectiveness of ScENEVERSE on improving existing 3D models. Specifically, we follow the open-vocabulary semantic segmentation settings proposed by Yang et al. [94] and report the mIoU and mAcc score. We compare with existing works by pre-training the RegionPLC [94] model on ScENEVERSE." + ], + "bbox": [ + 225, + 387, + 787, + 537 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Results and Analyses We present the results of 3D-QA experiments in Tab. 5 and the results of OV-Seg experiments in Tab. 6. The analyses are as follows:", + "bbox": [ + 212, + 542, + 785, + 571 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "- As shown in Tab. 5, our model achieves state-of-the-art results on both benchmarks, outperforming recent strong pre-training-based baselines like 3D-VisTA and 3D-LLM. As SCENEVERSE currently contains only descriptions of objects and scenes, we believe involving more types of language descriptions (e.g., question-answer pairs, dialogues) is a promising direction for further improving model performance on these downstream tasks.", + "bbox": [ + 225, + 573, + 785, + 662 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "- As shown in Tab. 6, we observe consistent performance improvement of existing 3D backbone models on this task when pre-trained with SCENEVERSE data. This result validates that the collected data in SCENEVERSE can effectively boost the performance of existing models on scene understanding tasks. We further provide results of state-of-the-art 3D models that are pretrained on SCENEVERSE on the close-vocabulary 3D semantic segmentation task in the supplementary.", + "bbox": [ + 225, + 662, + 787, + 768 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.4 Ablative Studies and Discussion", + "text_level": 1, + "bbox": [ + 215, + 787, + 527, + 801 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In this section, we present further discussions on both the data collected in ScENEVERSE and the GPS model design. We aim to elucidate the effects of", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "B. Jia and Y. Chen et al.", + "bbox": [ + 271, + 114, + 441, + 127 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/5e2ab110ec62d74690b13d152711fe2ec4c7bf3cd4386d0ab8a3b7417fa5deef.jpg", + "image_caption": [ + "Fig.4: Model performance v.s. data scale. Plots show that models consistently improve in both the pre-train and zero-shot transfer settings on ScanRefer and ScENEV-ERSE-val with data scaling-up." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 362, + 257 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/1e80d2da8c1b18b851b85d99876ecdb44af4e1bc358562e8a8c10d2567e0a5bb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 143, + 509, + 257 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/461ec40e932f7f612a07ba18779ede6fc5e2dca0ed2b0623accc48279ddb41ba.jpg", + "table_caption": [ + "Table 8: Cross domain transfer results of models pre-trained on real and synthetic datasets. \"S3D\" stands for Structured3D." + ], + "table_footnote": [], + "table_body": "
RealSyntheticSCENEVERSE-valS3DProcTHOR
AllX64.837.143.4
XS3D7.085.116.1
XProcTHOR4.216.391.0
", + "bbox": [ + 222, + 375, + 509, + 434 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/1f68ae3786918d128019ced644d711e634b8ad8188c1caa40124b6aaab1d1518.jpg", + "table_caption": [ + "Table 7: Ablation on text data source used in model pre-training. All models are tested on ScanRefer with no additional finetuning." + ], + "table_footnote": [], + "table_body": "
TemplateLLMAnno.Acc@0.25Acc@0.5
XXX43.538.4
✓XX50.946.1
✓✓X51.146.3
✓✓✓52.047.1
", + "bbox": [ + 526, + 205, + 781, + 281 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/2f2286d2c16cc8ce226e10668ed7802c9f376b4e5e3e2577c47acce49768dec1.jpg", + "table_caption": [ + "Table 9: Ablation on model design on ScENeVERSE-val. We use \"Obj- lvl\", \"Scene-lvl\" to denote object and scene alignment loss, and \"MLM\" for the mask language modeling loss." + ], + "table_footnote": [], + "table_body": "
Obj-lvl.MLMScene-lvl.OverallEasyHard
XXX64.875.448.7
✓XX65.277.147.4
✓✓X62.473.445.8
✓✓✓66.977.850.3
", + "bbox": [ + 527, + 361, + 781, + 434 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "data scaling and show more clearly its effectiveness in 3D scene understanding. Regarding the experimental settings and more results discussion, refer to the supplementary. The following points are specifically discussed in this section:", + "bbox": [ + 212, + 455, + 787, + 502 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "How important is data-scaling? We conduct ablation studies over the amount of data used while pre-training GPS. We consider the model trained with $\\frac{1}{8}$ , $\\frac{1}{4}$ , $\\frac{1}{2}$ of ScENEVERSE to show the effectiveness of data-scaling on model performance in the pre-train and zero-shot transfer settings in ScanRefer and ScENEVERSE-val. As shown in Fig. 4, we observe consistent performance improvement over the increase of data scale for both settings. We provide additional experiments in the supplementary to show that such scaling effect is not only beneficial for 3D-VL grounding but also for other 3D tasks like semantic segmentation [80,96].", + "bbox": [ + 212, + 508, + 787, + 630 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "How is the generated data compared with human-annotated data? We assess the performance of models trained using various scene-text sources, specifically focusing on their performance in the ScanRefer dataset without additional fine-tuning. As shown in Tab. 7, models trained with our template-based generated texts and Large Language Model (LLM)-refined texts show significant improvements over models trained solely on ScanRefer. More importantly, these variants of our model already achieve state-of-the-art results compared with previous baselines. This indicates the effectiveness of our text-generation pipeline. Finally, we observe that adding human-annotated data is still beneficial for model performance. However, the improvement is relatively marginal over models trained on our generated data.", + "bbox": [ + 212, + 636, + 787, + 801 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "What is the role of the synthetic scenes in this scale-up process? With synthetic data providing large-scale and diverse scene data for 3D-VL tasks,", + "bbox": [ + 212, + 809, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Scaling 3D Vision-Language Learning for Grounded Scene Understanding", + "bbox": [ + 241, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "we evaluate the models' domain transfer (Sim2Real) capability. Specifically, we compare models trained on all real scenes in SCENEVERSE against models trained exclusively on two synthetic subsets of SCENEVERSE, i.e., Structured3D and ProcTHOR. As shown in Tab. 8, models trained on synthetic subsets demonstrate remarkable performance on their corresponding test sets while suffering when transferred to real or other synthetic scenes. In contrast, the model trained on real scene-text pairs exhibits less performance drop when generalizing to synthetic scenes. This result affirms the domain gap between real and synthetic scenes in 3D-VL grounding and shows that a simple scale-up in the number of scenes is insufficient when naturalness can not be guaranteed. Considering the scalability of our language generation pipeline and the scaling effect shown in our experiments, the rate-determining step for further scaling-up 3D-VL comes to the collection of diverse, high-quality, and realistic scenes that capture natural 3D scene distributions.", + "bbox": [ + 212, + 146, + 787, + 357 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "How important is the design of each module in GPS? We provide ablative analyses of our multi-level contrastive alignment design in Tab. 9. We mainly consider removing objectives in our model to reveal the effectiveness of each level of alignment. We choose the referral-object-level alignment objective as the default setting and consider removing: (i) object-level alignment objective, (ii) masked language modeling objective, and (iii) scene-level alignment objective. When removing the object-level alignment objective, we learn the object point cloud encoder with the referral-object-level alignment and without pre-training. As shown in Tab. 9, we test different models on the SCENEVERSE-val without additional fine-tuning. Results show that the scene-level alignment objective is crucial for referral object grounding in SCENEVERSE-val with the $\\sim 5\\%$ performance drop. Similar observations could be made for the model trained without object-level alignment ( $\\sim 2\\%$ drop) and masked language modeling objective ( $\\sim 1.5\\%$ drop). These results affirm the effectiveness of our model design.", + "bbox": [ + 212, + 369, + 789, + 582 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 608, + 359, + 625 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we scale up 3D-VL for grounded scene understanding. We present SCENEVERSE, a million-scale 3D-VL dataset covering various scenes and multi-level scene descriptions sourced from both human annotation and our proposed scene-text generation approach. Utilizing SCENEVERSE, we propose Grounded Pre-training for Scenes (GPS), a model trained with multi-level scene-language contrastive alignment. Through extensive experiments, we show that GPS achieves state-of-the-art results on common 3D-VL tasks including grounding and question answering. We further conduct zero-shot transfer experiments to show the improved generalization performances of GPS trained on SCENEVERSE compared with previous baselines. We also demonstrate that the scaling effect of SCENEVERSE is generally beneficial for existing 3D models on 3D-VL tasks like semantic segmentation. We hope our efforts and successful scale-up attempts in SCENEVERSE could pave the way for new research paradigms in 3D-VL.", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "B. Jia and Y. Chen et al.", + "bbox": [ + 271, + 114, + 439, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement The authors would like to thank Yaowei Zhang (BIGAI) for his help on online visualization and other colleagues from BIGAI General Vision Lab for fruitful discussions. The authors would also like to thank the anonymous reviewers for their constructive feedback.", + "bbox": [ + 212, + 146, + 787, + 205 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 227, + 321, + 242 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Achlioptas, P., Abdelreheem, A., Xia, F., Elhoseiny, M., Guibas, L.: Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In: Proceedings of European Conference on Computer Vision (ECCV) (2020)", + "2. Agia, C., Jatavallabhula, K.M., Khodeir, M., Miksik, O., Vineet, V., Mukadam, M., Paull, L., Shkurti, F.: Taskography: Evaluating robot task planning over large 3d scene graphs. In: Proceedings of Conference on Robot Learning (CoRL) (2022)", + "3. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2022)", + "4. Armeni, I., He, Z.Y., Gwak, J., Zamir, A.R., Fischer, M., Malik, J., Savarese, S.: 3d scene graph: A structure for unified semantics, 3d space, and camera. In: Proceedings of International Conference on Computer Vision (ICCV) (2019)", + "5. Azuma, D., Miyanishi, T., Kurita, S., Kawanabe, M.: Scanqa: 3d question answering for spatial scene understanding. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2022)", + "6. Bakr, E., Alsaedy, Y., Elhoseiny, M.: Look around and refer: 2d synthetic semantics knowledge distillation for 3d visual grounding. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2022)", + "7. Barsalou, L.W.: Perceptual symbol systems. Behavioral and brain sciences 22(4), 577-660 (1999)", + "8. Barsalou, L.W.: Grounded cognition. Annu. Rev. Psychol. 59, 617-645 (2008)", + "9. Baruch, G., Chen, Z., Dehghan, A., Dimry, T., Feigin, Y., Fu, P., Gebauer, T., Joffe, B., Kurz, D., Schwartz, A., et al.: Arkitscenes: A diverse real-world dataset for 3d indoor scene understanding using mobile rgb-d data. In: Proceedings of Advances in Neural Information Processing Systems Datasets and Benchmarks (NeurIPS Datasets and Benchmarks Track) (2021)", + "10. Bommasani, R., Hudson, D.A., Adeli, E., Altman, R., Arora, S., von Arx, S., Bernstein, M.S., Bohg, J., Bosselut, A., Brunskill, E., et al.: On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258 (2021)", + "11. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2020)", + "12. Cai, D., Zhao, L., Zhang, J., Sheng, L., Xu, D.: 3djcg: A unified framework for joint dense captioning and visual grounding on 3d point clouds. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2022)", + "13. Chang, A., Dai, A., Funkhouser, T., Halber, M., Niessner, M., Savva, M., Song, S., Zeng, A., Zhang, Y.: Matterport3d: Learning from rgb-d data in indoor environments. Proceedings of International Conference on 3D Vision (3DV) (2017)", + "14. Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., et al.: Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012 (2015)" + ], + "bbox": [ + 225, + 253, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Scaling 3D Vision-Language Learning for Grounded Scene Understanding", + "bbox": [ + 241, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Changpinyo, S., Sharma, P., Ding, N., Soricut, R.: Conceptual 12m: Pushing web-scale image-text pre-training to recognize long-tail visual concepts. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2021)", + "16. Chen, D.Z., Chang, A.X., Nießner, M.: Scanrefer: 3d object localization in rgb-d scans using natural language. In: Proceedings of European Conference on Computer Vision (ECCV) (2020)", + "17. Chen, D.Z., Wu, Q., Nießner, M., Chang, A.X.: D3net: a speaker-listener architecture for semi-supervised dense captioning and visual grounding in rgb-d scans. In: Proceedings of European Conference on Computer Vision (ECCV) (2022)", + "18. Chen, S., Guhur, P.L., Tapaswi, M., Schmid, C., Laptev, I.: Language conditioned spatial relation reasoning for 3d object grounding. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2022)", + "19. Chen, S., Zhu, H., Chen, X., Lei, Y., Yu, G., Chen, T.: End-to-end 3d dense captioning with vote2cap-detr. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2023)", + "20. Chen, Y., Huang, S., Yuan, T., Qi, S., Zhu, Y., Zhu, S.C.: Holistic++ scene understanding: Single-view 3d holistic scene parsing and human pose estimation with human-object interaction and physical commonsense. In: Proceedings of International Conference on Computer Vision (ICCV) (2019)", + "21. Chen, Y., Li, Q., Kong, D., Kei, Y.L., Zhu, S.C., Gao, T., Zhu, Y., Huang, S.: Yourfit: Embodied reference understanding with language and gesture. In: Proceedings of International Conference on Computer Vision (ICCV) (2021)", + "22. Chen, Z., Gholami, A., Nießner, M., Chang, A.X.: Scan2cap: Context-aware dense captioning in rgb-d scans. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2021)", + "23. Collins, J., Goel, S., Deng, K., Luthra, A., Xu, L., Gundogdu, E., Zhang, X., Vicente, T.F.Y., Dideriksen, T., Arora, H., et al.: Abo: Dataset and benchmarks for real-world 3d object understanding. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2022)", + "24. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2017)", + "25. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500 (2023)", + "26. Deitke, M., Liu, R., Wallingford, M., Ngo, H., Michel, O., Kusupati, A., Fan, A., Laforte, C., Voleti, V., Gadre, S.Y., et al.: Objverse-xl: A universe of $10\\mathrm{m} + 3\\mathrm{d}$ objects. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2023)", + "27. Deitke, M., Schwenk, D., Salvador, J., Weihs, L., Michel, O., VanderBilt, E., Schmidt, L., Ehsani, K., Kembhavi, A., Farhadi, A.: Objverse: A universe of annotated 3d objects. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2023)", + "28. Deitke, M., VanderBilt, E., Herrasti, A., Weihs, L., Ehsani, K., Salvador, J., Han, W., Kolve, E., Kembhavi, A., Mottaghi, R.: Procthor: Large-scale embodied ai using procedural generation. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2022)", + "29. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: Proceedings of Conference of the North American Chapter of the Association for Computational Linguistics (NAACL) (2018)" + ], + "bbox": [ + 225, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "B. Jia and Y. Chen et al.", + "bbox": [ + 271, + 114, + 441, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "30. Ding, R., Yang, J., Xue, C., Zhang, W., Bai, S., Qi, X.: Pla: Language-driven open-vocabulary 3d scene understanding. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2023)", + "31. Ding, Z., Han, X., Niethammer, M.: Volenet: A deep learning label fusion method for multi-atlas segmentation. In: Proceedings of International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI) (2019)", + "32. Feng, M., Li, Z., Li, Q., Zhang, L., Zhang, X., Zhu, G., Zhang, H., Wang, Y., Mian, A.: Free-form description guided 3d visual graph network for object grounding in point cloud. In: Proceedings of International Conference on Computer Vision (ICCV) (2021)", + "33. Ghiasi, G., Gu, X., Cui, Y., Lin, T.Y.: Scaling open-vocabulary image segmentation with image-level labels. In: Proceedings of European Conference on Computer Vision (ECCV) (2022)", + "34. Gong, R., Huang, J., Zhao, Y., Geng, H., Gao, X., Wu, Q., Ai, W., Zhou, Z., Terzopoulos, D., Zhu, S.C., Jia, B., Huang, S.: Arnold: A benchmark for language-grounded task learning with continuous states in realistic 3d scenes. In: Proceedings of International Conference on Computer Vision (ICCV) (2023)", + "35. Gu, Q., Kuwajerwala, A., Morin, S., Jatavallabhula, K.M., Sen, B., Agarwal, A., Rivera, C., Paul, W., Ellis, K., Chellappa, R., et al.: Conceptgraphs: Open-vocabulary 3d scene graphs for perception and planning. arXiv preprint arXiv:2309.16650 (2023)", + "36. Ha, H., Song, S.: Semantic abstraction: Open-world 3d scene understanding from 2d vision-language models. In: Proceedings of Conference on Robot Learning (CoRL) (2022)", + "37. He, D., Zhao, Y., Luo, J., Hui, T., Huang, S., Zhang, A., Liu, S.: Transrefer3d: Entity-and-relation aware transformer for fine-grained 3d visual grounding. In: Proceedings of ACM International Conference on Multimedia (MM) (2021)", + "38. Hegde, D., Valanarasu, J.M.J., Patel, V.: Clip goes 3d: Leveraging prompt tuning for language grounded 3d recognition. In: Proceedings of International Conference on Computer Vision (ICCV) (2023)", + "39. Hong, Y., Wu, Q., Qi, Y., Rodriguez-Opazo, C., Gould, S.: Vln bert: A recurrent vision-and-language bert for navigation. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2021)", + "40. Hong, Y., Lin, C., Du, Y., Chen, Z., Tenenbaum, J.B., Gan, C.: 3d concept learning and reasoning from multi-view images. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2023)", + "41. Huang, J., Yong, S., Ma, X., Linghu, X., Li, P., Wang, Y., Li, Q., Zhu, S.C., Jia, B., Huang, S.: An embodied generalist agent in 3d world. arXiv preprint arXiv:2311.12871 (2023)", + "42. Huang, P.H., Lee, H.H., Chen, H.T., Liu, T.L.: Text-guided graph neural networks for referring 3d instance segmentation. In: Proceedings of AAAI Conference on Artificial Intelligence (AAAI) (2021)", + "43. Huang, S., Chen, Y., Jia, J., Wang, L.: Multi-view transformer for 3d visual grounding. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2022)", + "44. Huang, S., Wang, Z., Li, P., Jia, B., Liu, T., Zhu, Y., Liang, W., Zhu, S.C.: Diffusion-based generation, optimization, and planning in 3d scenes. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2023)", + "45. Jain, A., Gkanatsios, N., Mediratta, I., Fragkiadaki, K.: Bottom up top down detection transformers for language grounding in images and point clouds. In: Proceedings of European Conference on Computer Vision (ECCV) (2022)" + ], + "bbox": [ + 223, + 146, + 785, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Scaling 3D Vision-Language Learning for Grounded Scene Understanding", + "bbox": [ + 241, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "46. Jiang, L., Zhao, H., Shi, S., Liu, S., Fu, C.W., Jia, J.: Pointgroup: Dual-set point grouping for 3d instance segmentation. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2020)", + "47. Jiang, N., Liu, T., Cao, Z., Cui, J., Zhang, Z., Chen, Y., Wang, H., Zhu, Y., Huang, S.: Full-body articulated human-object interaction. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2023)", + "48. Jiang, N., Zhang, Z., Li, H., Ma, X., Wang, Z., Chen, Y., Liu, T., Zhu, Y., Huang, S.: Scaling up dynamic human-scene interaction modeling. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2024)", + "49. Kaplan, J., McCandlish, S., Henighan, T., Brown, T.B., Chess, B., Child, R., Gray, S., Radford, A., Wu, J., Amodei, D.: Scaling laws for neural language models. arXiv preprint arXiv:2001.08361 (2020)", + "50. Khanna, M., Mao, Y., Jiang, H., Haresh, S., Shacklett, B., Batra, D., Clegg, A., Undersander, E., Chang, A.X., Savva, M.: Habitat synthetic scenes dataset (hssd-200): An analysis of 3d scene scale and realism tradeoffs for objectgoal navigation. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2024)", + "51. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: Proceedings of International Conference on Computer Vision (ICCV) (2023)", + "52. Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., et al.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. In: International Journal of Computer Vision (IJCV) (2017)", + "53. Lake, B.M., Ullman, T.D., Tenenbaum, J.B., Gershman, S.J.: Building machines that learn and think like people. Behavioral and brain sciences 40, e253 (2017)", + "54. Li, B., Weinberger, K.Q., Belongie, S., Koltun, V., Ranftl, R.: Language-driven semantic segmentation. In: Proceedings of International Conference on Learning Representations (ICLR) (2022)", + "55. Li, C., Zhang, R., Wong, J., Gokmen, C., Srivastava, S., Martin-Martín, R., Wang, C., Levine, G., Lingelbach, M., Sun, J., et al.: Behavior-1k: A benchmark for embodied ai with 1,000 everyday activities and realistic simulation. In: Proceedings of Conference on Robot Learning (CoRL) (2023)", + "56. Li, J., Li, D., Savarese, S., Hoi, S.: BLIP-2: bootstrapping language-image pretraining with frozen image encoders and large language models. In: Proceedings of International Conference on Machine Learning (ICML) (2023)", + "57. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: Proceedings of International Conference on Machine Learning (ICML) (2022)", + "58. Li, L.H., Zhang, P., Zhang, H., Yang, J., Li, C., Zhong, Y., Wang, L., Yuan, L., Zhang, L., Hwang, J.N., et al.: Grounded language-image pre-training. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2022)", + "59. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2023)", + "60. Liu, M., Shi, R., Kuang, K., Zhu, Y., Li, X., Han, S., Cai, H., Porikli, F., Su, H.: Openshape: Scaling up 3d shape representation towards open-world understanding. arXiv preprint arXiv:2305.10764 (2023)", + "61. Liu, R., Wu, R., Van Hoorick, B., Tokmakov, P., Zakharov, S., Vondrick, C.: Zero-1-to-3: Zero-shot one image to 3d object. In: Proceedings of International Conference on Computer Vision (ICCV) (2023)" + ], + "bbox": [ + 225, + 146, + 785, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "B. Jia and Y. Chen et al.", + "bbox": [ + 271, + 114, + 441, + 127 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "62. Luo, J., Fu, J., Kong, X., Gao, C., Ren, H., Shen, H., Xia, H., Liu, S.: 3d-sps: Single-stage 3d visual grounding via referred point progressive selection. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2022)", + "63. Luo, T., Rockwell, C., Lee, H., Johnson, J.: Scalable 3d captioning with pretrained models. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2023)", + "64. Ma, C.Y., Lu, J., Wu, Z., AlRegib, G., Kira, Z., Socher, R., Xiong, C.: Self-monitoring navigation agent via auxiliary progress estimation. In: Proceedings of International Conference on Learning Representations (ICLR) (2019)", + "65. Ma, X., Yong, S., Zheng, Z., Li, Q., Liang, Y., Zhu, S.C., Huang, S.: Sqa3d: Situated question answering in 3d scenes. In: Proceedings of International Conference on Learning Representations (ICLR) (2023)", + "66. Mao, Y., Zhang, Y., Jiang, H., Chang, A., Savva, M.: Multiscan: Scalable rgbd scanning for 3d environments with articulated objects. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2022)", + "67. Misra, I., Girdhar, R., Joulin, A.: An end-to-end transformer model for 3d object detection. In: Proceedings of International Conference on Computer Vision (ICCV) (2021)", + "68. Mittal, M., Yu, C., Yu, Q., Liu, J., Rudin, N., Hoeller, D., Yuan, J.L., Singh, R., Guo, Y., Mazhar, H., et al.: Orbit: A unified simulation framework for interactive robot learning environments. Robotics and Automation Letters (RA-L) (2023)", + "69. Mo, K., Zhu, S., Chang, A.X., Yi, L., Tripathi, S., Guibas, L.J., Su, H.: Partnet: A large-scale benchmark for fine-grained and hierarchical part-level 3d object understanding. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2019)", + "70. Mu, T., Ling, Z., Xiang, F., Yang, D., Li, X., Tao, S., Huang, Z., Jia, Z., Su, H.: Maniskill: Generalizable manipulation skill benchmark with large-scale demonstrations. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2021)", + "71. OpenAI: Introducing chatgpt. https://openai.com/blog/chatgpt (2022)", + "72. OpenAI: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)", + "73. Pashevich, A., Schmid, C., Sun, C.: Episodic transformer for vision-and-language navigation. In: Proceedings of International Conference on Computer Vision (ICCV) (2021)", + "74. Peng, S., Genova, K., Jiang, C., Tagliasacchi, A., Pollefeys, M., Funkhouser, T., et al.: Openscene: 3d scene understanding with open vocabularies. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2023)", + "75. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: Proceedings of International Conference on Machine Learning (ICML) (2021)", + "76. Ramakrishnan, S.K., Gokaslan, A., Wijmans, E., Maksymets, O., Clegg, A., Turner, J., Undersander, E., Galuba, W., Westbury, A., Chang, A.X., et al.: Habitatmatterport 3d dataset (hm3d): 1000 large-scale 3d environments for embodied ai. In: Proceedings of Advances in Neural Information Processing Systems Datasets and Benchmarks (NeurIPS Datasets and Benchmarks Track) (2021)", + "77. Rana, K., Haviland, J., Garg, S., Abou-Chakra, J., Reid, I., Suenderhauf, N.: Sayplan: Grounding large language models using 3d scene graphs for scalable robot task planning. In: Proceedings of Conference on Robot Learning (CoRL) (2023)" + ], + "bbox": [ + 223, + 146, + 787, + 839 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "Scaling 3D Vision-Language Learning for Grounded Scene Understanding", + "bbox": [ + 241, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 18 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "78. Rosinol, A., Violette, A., Abate, M., Hughes, N., Chang, Y., Shi, J., Gupta, A., Carlone, L.: Kimera: From slam to spatial perception with 3d dynamic scene graphs. International Journal of Robotics Research (IJRR) (2021)", + "79. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2022)", + "80. Schult, J., Engelmann, F., Hermans, A., Litany, O., Tang, S., Leibe, B.: Mask3d: Mask transformer for 3d semantic instance segmentation. In: Proceedings of International Conference on Robotics and Automation (ICRA) (2023)", + "81. Smith, L., Gasser, M.: The development of embodied cognition: Six lessons from babies. Artificial life 11(1-2), 13-29 (2005)", + "82. Takmaz, A., Fedele, E., Sumner, R.W., Pollefeys, M., Tombari, F., Engelmann, F.: Openmask3d: Open-vocabulary 3d instance segmentation. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2023)", + "83. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)", + "84. Vu, T., Kim, K., Luu, T.M., Nguyen, T., Yoo, C.D.: Softgroup for 3d instance segmentation on point clouds. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2022)", + "85. Wald, J., Avetisyan, A., Navab, N., Tombari, F., Nießner, M.: Rio: 3d object instance re-localization in changing indoor environments. In: Proceedings of International Conference on Computer Vision (ICCV) (2019)", + "86. Wald, J., Dhamo, H., Navab, N., Tombari, F.: Learning 3d semantic scene graphs from 3d indoor reconstructions. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2020)", + "87. Wang, T., Mao, X., Zhu, C., Xu, R., Lyu, R., Li, P., Chen, X., Zhang, W., Chen, K., Xue, T., et al.: Embodiedscan: A holistic multi-modal 3d perception suite towards embodied ai. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2024)", + "88. Wang, X., Huang, Q., Celikyilmaz, A., Gao, J., Shen, D., Wang, Y.F., Wang, W.Y., Zhang, L.: Reinforced cross-modal matching and self-supervised imitation learning for vision-language navigation. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2019)", + "89. Wang, Z., Chen, Y., Jia, B., Li, P., Zhang, J., Zhang, J., Liu, T., Zhu, Y., Liang, W., Huang, S.: Move as you say interact as you can: Language-guided human motion generation with scene affordance. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2024)", + "90. Wang, Z., Chen, Y., Liu, T., Zhu, Y., Liang, W., Huang, S.: Humanise: Language-conditioned human motion generation in 3d scenes. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2022)", + "91. Wu, T., Zhang, J., Fu, X., Wang, Y., Ren, J., Pan, L., Wu, W., Yang, L., Wang, J., Qian, C., et al.: Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2023)", + "92. Wu, Y., Cheng, X., Zhang, R., Cheng, Z., Zhang, J.: Eda: Explicit text-decoupling and dense alignment for 3d visual grounding. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + ], + "bbox": [ + 225, + 146, + 785, + 839 + ], + "page_idx": 19 + }, + { + "type": "page_number", + "text": "20", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 19 + }, + { + "type": "header", + "text": "B. Jia and Y. Chen et al.", + "bbox": [ + 271, + 114, + 441, + 127 + ], + "page_idx": 19 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "93. Xue, L., Gao, M., Xing, C., Martin-Martin, R., Wu, J., Xiong, C., Xu, R., Niebles, J.C., Savarese, S.: Ulip: Learning a unified representation of language, images, and point clouds for 3d understanding. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2023)", + "94. Yang, J., Ding, R., Wang, Z., Qi, X.: Regionplc: Regional point-language contrastive learning for open-world 3d scene understanding. arXiv preprint arXiv:2304.00962 (2023)", + "95. Yang, Y., Jia, B., Zhi, P., Huang, S.: Physcene: Physically interactable 3d scene synthesis for embodied ai. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2024)", + "96. Yang, Y.Q., Guo, Y.X., Xiong, J.Y., Liu, Y., Pan, H., Wang, P.S., Tong, X., Guo, B.: Swin3d: A pretrained transformer backbone for 3d indoor scene understanding. arXiv preprint arXiv:2304.06906 (2023)", + "97. Yang, Z., Zhang, S., Wang, L., Luo, J.: Sat: 2d semantics assisted training for 3d visual grounding. In: Proceedings of International Conference on Computer Vision (ICCV) (2021)", + "98. Yeshwanth, C., Liu, Y.C., Nießner, M., Dai, A.: Scannet++: A high-fidelity dataset of 3d indoor scenes. In: Proceedings of International Conference on Computer Vision (ICCV) (2023)", + "99. Yuan, Z., Yan, X., Liao, Y., Guo, Y., Li, G., Cui, S., Li, Z.: X-trans2cap: Cross-modal knowledge transfer using transformer for 3d dense captioning. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2022)", + "100. Yuan, Z., Yan, X., Liao, Y., Zhang, R., Wang, S., Li, Z., Cui, S.: Instancerefer: Cooperative holistic understanding for visual grounding on point clouds through instance multi-level contextual referring. In: Proceedings of International Conference on Computer Vision (ICCV) (2021)", + "101. Zhang, H., Zhang, P., Hu, X., Chen, Y.C., Li, L., Dai, X., Wang, L., Yuan, L., Hwang, J.N., Gao, J.: Glipv2: Unifying localization and vision-language understanding. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2022)", + "102. Zhang, R., Guo, Z., Zhang, W., Li, K., Miao, X., Cui, B., Qiao, Y., Gao, P., Li, H.: Pointclip: Point cloud understanding by clip. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2022)", + "103. Zhang, R., Wang, L., Qiao, Y., Gao, P., Li, H.: Learning 3d representations from 2d pre-trained models via image-to-point masked autoencoders. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2023)", + "104. Zhang, Y., Gong, Z., Chang, A.X.: Multi3drefer: Grounding text description to multiple 3d objects. In: Proceedings of International Conference on Computer Vision (ICCV) (2023)", + "105. Zhao, L., Cai, D., Sheng, L., Xu, D.: 3dvg-transformer: Relation modeling for visual grounding on point clouds. In: Proceedings of International Conference on Computer Vision (ICCV) (2021)", + "106. Zheng, J., Zhang, J., Li, J., Tang, R., Gao, S., Zhou, Z.: Structured3d: A large photo-realistic dataset for structured 3d modeling. In: Proceedings of European Conference on Computer Vision (ECCV) (2020)", + "107. Zhu, W., Hessel, J., Awadalla, A., Gadre, S.Y., Dodge, J., Fang, A., Yu, Y., Schmidt, L., Wang, W.Y., Choi, Y.: Multimodal c4: An open, billion-scale corpus of images interleaved with text. arXiv preprint arXiv:2304.06939 (2023)" + ], + "bbox": [ + 218, + 146, + 787, + 839 + ], + "page_idx": 20 + }, + { + "type": "header", + "text": "Scaling 3D Vision-Language Learning for Grounded Scene Understanding", + "bbox": [ + 241, + 114, + 730, + 128 + ], + "page_idx": 20 + }, + { + "type": "page_number", + "text": "21", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 20 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "108. Zhu, Y., Gao, T., Fan, L., Huang, S., Edmonds, M., Liu, H., Gao, F., Zhang, C., Qi, S., Wu, Y.N., et al.: Dark, beyond deep: A paradigm shift to cognitive ai with humanlike common sense. Engineering 6(3), 310-345 (2020)", + "109. Zhu, Z., Ma, X., Chen, Y., Deng, Z., Huang, S., Li, Q.: 3d-vista: Pre-trained transformer for 3d vision and text alignment. In: Proceedings of International Conference on Computer Vision (ICCV) (2023)", + "110. Zhu, Z., Zhang, Z., Ma, X., Niu, X., Chen, Y., Jia, B., Deng, Z., Huang, S., Li, Q.: Unifying 3d vision-language understanding via promptable queries. In: Proceedings of European Conference on Computer Vision (ECCV) (2024)" + ], + "bbox": [ + 217, + 146, + 785, + 272 + ], + "page_idx": 21 + }, + { + "type": "page_number", + "text": "22", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 21 + }, + { + "type": "header", + "text": "B. Jia and Y. Chen et al.", + "bbox": [ + 271, + 114, + 439, + 126 + ], + "page_idx": 21 + } +] \ No newline at end of file diff --git a/2024/SceneVerse_ Scaling 3D Vision-Language Learning for Grounded Scene Understanding/72c8327d-dc6a-436c-89b3-89891fd18764_model.json b/2024/SceneVerse_ Scaling 3D Vision-Language Learning for Grounded Scene Understanding/72c8327d-dc6a-436c-89b3-89891fd18764_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e1e13e9eb5d009864df1d7ef6d891b4ca1900324 --- /dev/null +++ b/2024/SceneVerse_ Scaling 3D Vision-Language Learning for Grounded Scene Understanding/72c8327d-dc6a-436c-89b3-89891fd18764_model.json @@ -0,0 +1,3632 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.245, + 0.141, + 0.758, + 0.187 + ], + "angle": 0, + "content": "SCENEVERSE: Scaling 3D Vision-Language Learning for Grounded Scene Understanding" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.212, + 0.782, + 0.245 + ], + "angle": 0, + "content": "Baoxiong Jia\\*, Yixin Chen\\*, Huangyue Yu, Yan Wang, Xuesong Niu, Tengyu Liu, Qing Li, and Siyuan Huang" + }, + { + "type": "text", + "bbox": [ + 0.292, + 0.256, + 0.71, + 0.298 + ], + "angle": 0, + "content": "* indicates equal contribution\nState Key Laboratory of General Artificial Intelligence, BIGAI https://scene-verse.github.io" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.333, + 0.744, + 0.639 + ], + "angle": 0, + "content": "Abstract. 3D vision-language (3D-VL) grounding, which aims to align language with 3D physical environments, stands as a cornerstone in developing embodied agents. In comparison to recent advancements in the 2D domain, grounding language in 3D scenes faces two significant challenges: (i) the scarcity of paired 3D-VL data to support grounded learning of 3D scenes, especially considering complexities within diverse object configurations, rich attributes, and intricate relationships; and (ii) the absence of a unified learning framework to distill knowledge from grounded 3D data. In this work, we aim to address these major challenges in 3D-VL by examining the potential of systematically upscaling 3D-VL learning in indoor scenes. We introduce the first million-scale 3D-VL dataset, ScENEVERSE, encompassing 68K indoor scenes and comprising 2.5M vision-language pairs collected from both human annotations and our scalable scene-graph-based generation approach. We demonstrate that this scaling allows for a unified pre-training framework, Grounded Pretraining for Scenes (GPS), for 3D-VL learning. Through extensive experiments, we showcase the effectiveness of GPS by achieving state-of-the-art performance on existing 3D visual grounding and question-answering benchmarks. We also show that the data scaling effect is not limited to GPS, but is generally beneficial for models on tasks like 3D semantic segmentation. The vast potential of ScENEVERSE and GPS is unveiled through zero-shot transfer experiments in challenging 3D-VL tasks." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.651, + 0.744, + 0.679 + ], + "angle": 0, + "content": "Keywords: 3D Vision-Language \\(\\cdot\\) Data Scaling \\(\\cdot\\) Grounded Scene Understanding" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.704, + 0.377, + 0.72 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.79, + 0.844 + ], + "angle": 0, + "content": "The foundation of human cognitive development lies in the grounding of language within the physical world [53,81,108]. Recent progress in Large Language Models (LLMs) [10,11,83] has markedly promoted the alignment between vision and language [3,59,75] utilizing billion-scale vision-language datasets [79,107]. However, with these advancements predominantly focusing on the 2D domain, the grounded understanding of 3D physical environments remains in an incipient stage [1,5,16]. Recognizing the pivotal role of grounded 3D experiences in" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.442, + 0.128 + ], + "angle": 0, + "content": "B. Jia and Y. Chen et al." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.144, + 0.782, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.41, + 0.788, + 0.453 + ], + "angle": 0, + "content": "Fig. 1: Overview of ScENEVERSE. A million-scale 3D vision-language dataset that comprises over 68K various 3D indoor scenes and 2.5M aligned scene-language pairs in the form of scene caption, object caption, and object referral." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.461, + 0.787, + 0.492 + ], + "angle": 0, + "content": "shaping human cognition [7,8], there is a compelling need to focus on exploring vision-language learning in the context of 3D scenes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.492, + 0.789, + 0.719 + ], + "angle": 0, + "content": "Seeking insights from success in 2D vision-language (2D-VL) learning, a major factor to the success was the notable scale-up of paired vision-language data [15, 52, 79]. However, applying this experience directly from 2D to 3D is fraught with challenges. Primarily, 3D data collection heavily relies on the scanning device, making it inherently much more complex and expensive than gathering 2D images. Despite steady efforts to increase the volume of 3D scene data [9, 24, 66, 98], most datasets remain limited to thousands of scenes, substantially lagging behind the scale of existing 2D datasets. This gap is further widened by the inherent complexities of 3D scenes, which feature a multitude of object instances with diverse attributes, varying arrangements, and intricate inter-object relationships. These unique aspects of 3D scenes not only make the accurate description of objects and their relations more challenging but also considerably increase the number of language descriptions required for thorough scene depiction. Consequently, this presents a significant challenge in gathering sufficient and high-quality paired scene-language data for grounded scene understanding." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.79, + 0.841 + ], + "angle": 0, + "content": "To confront these challenges, we propose SCENEVERSE, the first million-scale dataset aimed at advancing 3D vision-language (3D-VL) learning for grounded scene understanding. At the scene level, we unify 3D scene data from existing datasets [9,24,66,76,85], aligning scenes and annotations from various capturing sources, and supplement the collection with synthetic scenes [28,106]. This compilation represents the most extensive 3D scene data gathered to date, amounting to 68K scenes. For language, we first present 97K newly-annotated referring expressions, the most extensive thus far. We additionally propose an" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.242, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "Scaling 3D Vision-Language Learning for Grounded Scene Understanding" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.207 + ], + "angle": 0, + "content": "automated generation pipeline utilizing 3D scene graphs [4,86] and LLMs to create comprehensive, high-quality scene-language pairs. This refined collection, totaling 2.5M scene-language pairs, provides detailed and comprehensive descriptions of both object-level and scene-level descriptions within the 3D scene." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.79, + 0.388 + ], + "angle": 0, + "content": "We thoroughly investigate the potential offered by SCENEVERSE with large-scale pre-training, introducing Grounded Pre-training for Scenes (GPS), a novel and unified pre-training framework designed for scene and object-level alignment without auxiliary losses. Through multi-level contrastive alignment, we achieve significant performance boosts on 3D-VL tasks, such as grounding and question answering, setting new state-of-the-art results via a simple and effective pretraining process. We unveil the vast possibilities offered by SCENEVERSE and GPS in 3D-VL tasks in a zero-shot transfer setting. Additionally, we show that the scaling effect in SCENEVERSE is not limited to GPS, but generally benefits models in tasks like 3D semantic segmentation. At last, we offer deeper insights into the data-scaling in SCENEVERSE through extensive ablative experiments, pointing out future directions. Our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.39, + 0.788, + 0.465 + ], + "angle": 0, + "content": "1. We introduce SCENEVERSE, the first million-scale 3D-VL dataset for grounded scene understanding. SCENEVERSE encompasses 68K 3D scenes coupled with 2.5M scene-language pairs, sourced through a combination of human annotation and automated generation methods. This represents a significant improvement in terms of data diversity and scale compared to prior datasets." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.465, + 0.788, + 0.51 + ], + "angle": 0, + "content": "2. We propose GPS, a transformer-based model trained with multi-level scenetest alignment that achieves state-of-the-art results on existing 3D-VL grounding and question-answering benchmarks by pre-training on ScENEVERSE." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.51, + 0.788, + 0.571 + ], + "angle": 0, + "content": "3. We demonstrate that with the data scale-up and model design, our pre-trained GPS exhibit emerging zero-shot generalization capabilities in grounded scene understanding. We also show that this scaling effect is not limited to GPS, but is generally beneficial for models on tasks like semantic segmentation." + }, + { + "type": "list", + "bbox": [ + 0.223, + 0.39, + 0.788, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.595, + 0.388, + 0.611 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.788, + 0.75 + ], + "angle": 0, + "content": "Datasets for Grounded 3D Understanding Obtaining aligned 3D-language data is a inherently difficult task. In 3D object modeling, pioneering works like ShapeNet [14] sourced 3D assets from online repositories, leading to a proliferation of high-quality 3D object datasets [23, 69, 91]. Notably, recent developments include internet-scale data collection with Objaverse [26, 27], accompanied by the integration of object-level captions [93] for 3D-language alignment. Models trained on these datasets demonstrate an enhanced understanding of objects, evident in classification [60], generation [61], and captioning tasks [63]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.751, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In contrast, developing datasets for grounded 3D scene understanding is even more challenging due to the extensive requirements for scene acquisition and annotation. Existing works curate RGB-D and scanned indoor scene datasets [9,13, 24,66,76,85] and synthetic scenes [28,50,95,106] used for benchmarking tasks like 3D object detection and segmentation [31,46,67,80,84]. These semantically labeled scenes are subsequently used in fine-grained scene grounding tasks like object" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.442, + 0.128 + ], + "angle": 0, + "content": "B. Jia and Y. Chen et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.214 + ], + "angle": 0, + "content": "Table 1: Comparison of SCENEVERSE with existing 3DVL Datasets. SCENEVERSE expands the data scale of prior work by order of magnitude. \"VG\" stands for Visual Grounding, \"QA\" for Question Answering, \"PT\" for Pre-training and \"MT\" for Multi-tasking. \"Anno.\" denotes language from human annotations and \"Syn.\" for template-based or LLM generated descriptions." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.215, + 0.785, + 0.376 + ], + "angle": 0, + "content": "
Dataset3D Scene Obj.TaskObj. CaptionScene CaptionObj. ReferralQuality CheckNew Anno.Syn.Existing Anno.Syn.Total
ScanRefer [16]VGXX✓✓52K---52K
ReferIt3D [1]| |VGXX✓✓42K200K--242K
ScanQA [5]1.5K33KQA---✓27K---27K
SQA3D [65]QA---✓33K---33K
Multi3DRefer [104]| |VGXX✓✓-10K52K-62K
Cap3D [63]-666KVGX✓XX58K666K--724K
ScanScribe [109]3K56KPTXX✓X-90K94K94K278K
3D-LLM [40]1.5K186KMT✓✓✓X-659K--659K
EmbodiedScan [87]5K890KVGXX✓X-970K--970K
LEO [41]3K56KMT✓✓✓✓-188K235K90K513K
SCENEVERSE68K1.5MVG✓✓✓✓96K2.1M94K200K2.5M
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.389, + 0.788, + 0.513 + ], + "angle": 0, + "content": "referral [1, 16, 104], captioning [17, 19, 22, 99], vision-language-navigation [39, 64, 73, 88] and reasoning [5, 40, 65]. Recent works exploit the representation of 3D scene graph (3DSG) [4, 20, 78, 86], which concisely describes scenes with hierarchical structures. This representation is notably advantageous for planning [2, 77] and captioning [35], owing to its compatibility with LLMs for flexible description generation [40, 41]. Nonetheless, as shown in Tab. 1, most datasets are constrained in both scene and language scales, underscoring the need for scaling up fine-grained and aligned scene-language data to enhance grounded scene understanding." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.788, + 0.614 + ], + "angle": 0, + "content": "Vision-Language Learning Recent years have witnessed tremendous progress in 2D-VL [3, 21, 25, 57, 59, 75], empowered by transformer-based pre-training models [11, 29, 72] and large-scale image-language datasets [15, 79]. A central theme across 2D-VL domains is the effectiveness of data scaling [49], as demonstrated by improved alignment and expanded capabilities in open-vocabulary understanding [33, 51, 54, 58] through a contrastive pre-training pipeline [75]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.789, + 0.841 + ], + "angle": 0, + "content": "However, in grounded scene understanding, the primary challenge for models has been the limited availability of paired 3D scene-language data, which restricts the application of insights drawn from 2D-VL. Current models for 3D scene grounding [6,18,37,43,45,62,92,97,105] heavily rely on task-specific knowledge in both model and loss designs or advanced optimization strategies [109]. To bridge this gap, there has been a growing emphasis on employing pre-trained 2D-VL models for 3D-VL [36,38,74,82,93,102,103]. Yet, these models mostly draw information available from 2D-VL models (e.g., object attribute, affordance, etc.), falling short on capturing crucial 3D information like object spatial relationships which are necessary for more fine-grained tasks such as grounded human-scene [20, 44,47,48,89,90] and robot-scene interactions modeling [34,55,68,70]. This urges the need for a multi-level alignment between language and 3D scenes, particularly regarding 3D-specific information. Considering the nascent stage of existing 3D pre-training methods [30,94,109,110], we believe SCENEVERSE and GPS have the potential to spearhead new avenues in 3D-VL research." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.242, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Scaling 3D Vision-Language Learning for Grounded Scene Understanding" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.381, + 0.161 + ], + "angle": 0, + "content": "3 SCENEVERSE" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.174, + 0.788, + 0.266 + ], + "angle": 0, + "content": "SCENEVERSE is designed for grounded scene understanding with 3D scenes curated from diverse existing datasets of both real and synthetic environments. Regarding language, we employ both human annotation and a novel automated generation pipeline to collect comprehensive and high-quality language for both object-level and scene-level descriptions. We provide details regarding data collection in the following sections." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.284, + 0.39, + 0.298 + ], + "angle": 0, + "content": "3.1 Scene Curation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.305, + 0.789, + 0.472 + ], + "angle": 0, + "content": "To address the scarcity of available 3D scene data, we construct SCENEVERSE by unifying 3D scene data from various existing datasets. We use real-world scene datasets, including ScanNet [24], ARKitScenes [9], HM3D [76], 3RScan [85] and MultiScan [66], alongside synthetic environments from Structured3D [106] and ProcTHOR [28]. The inclusion of these synthetic datasets is mainly motivated by their potential as scalable data sources for 3D-VL alignment. To ensure cohesion across various sources, we conduct preprocessing steps such as room segmentation, point subsampling, axis alignment, normalization, and semantic label alignment. Each scan is represented by a point cloud \\(\\mathbf{P} \\in \\mathbb{R}^{N \\times 8}\\), wherein each point is defined by its 3D coordinates, RGB color, instance id, and semantic label. In total, we curate 68,406 3D scenes in SCENEVERSE." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.491, + 0.53, + 0.506 + ], + "angle": 0, + "content": "3.2 Referral Annotation by Humans" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.513, + 0.789, + 0.664 + ], + "angle": 0, + "content": "In the curated scenes of SCENEVERSE, we present the most comprehensive set of human-annotated, context-rich object referrals to date, serving as a valuable benchmark for assessing grounded scene understanding capabilities. The human annotations contain 96,863 descriptions in ARKitScenes [9], HM3D [76] and MultiScan [66]. During the annotation process, one human annotator was assigned to write at least 20 words to distinctly refer to a single 3D object within a 3D scene. Each referral text then undergoes independent verification by two additional reviewers, both mandated to accurately locate the referenced object based on the 3D scene and the annotated referral text. Any object referrals that do not pass the verification by either reviewer are flagged for re-annotation." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.683, + 0.512, + 0.698 + ], + "angle": 0, + "content": "3.3 3D Scene Graph Construction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Our 3D scene graph is defined as a hierarchical graph \\(\\mathcal{G} = (\\mathcal{V},\\mathcal{E})\\). Each node \\(v\\in \\mathcal{V}\\) represents one distinct 3D object instance, parameterized by its centroid \\(\\pmb {p}_i\\in \\mathbb{R}^3\\) and bounding box size of \\(\\pmb {b}_i = (b_x,b_y,b_z)\\in \\mathbb{R}^3\\). The edges \\(\\mathcal{E}\\) represent spatial relationships between nodes. To construct the scene graph \\(\\mathcal{G}\\), we instantiate the nodes with the instance annotation from the point clouds and assign object classes with their corresponding semantic labels. Following prior work [1,86], we consider the Vertical proximity, Horizontal proximity and Multi-object Relationships as spatial relations. For a more detailed description of the scene graph construction and relationship determination, please refer to supplementary." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.442, + 0.128 + ], + "angle": 0, + "content": "B. Jia and Y. Chen et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.146, + 0.44, + 0.249 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.306, + 0.252, + 0.353, + 0.26 + ], + "angle": 0, + "content": "(a) 3D Scene" + }, + { + "type": "text", + "bbox": [ + 0.254, + 0.264, + 0.272, + 0.271 + ], + "angle": 0, + "content": "一" + }, + { + "type": "text", + "bbox": [ + 0.277, + 0.264, + 0.327, + 0.271 + ], + "angle": 0, + "content": "Scene Cap" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.275, + 0.304, + 0.284 + ], + "angle": 0, + "content": "Sub-graph Context" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.285, + 0.333, + 0.291 + ], + "angle": 0, + "content": "{scene_type':Bedroom'," + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.292, + 0.333, + 0.299 + ], + "angle": 0, + "content": "'object_count': {'nightstand':2,...}," + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.3, + 0.334, + 0.306 + ], + "angle": 0, + "content": "'relation': {nightstand', 'on', floor}," + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.307, + 0.334, + 0.313 + ], + "angle": 0, + "content": "['backback', in front of, b" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.314, + 0.281, + 0.322 + ], + "angle": 0, + "content": "Summary" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.322, + 0.387, + 0.328 + ], + "angle": 0, + "content": "Prompt: Provide a summary for a scene from a given" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.329, + 0.353, + 0.335 + ], + "angle": 0, + "content": "scene graph delimited by triple backticks," + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.336, + 0.404, + 0.342 + ], + "angle": 0, + "content": "Response: In this bedroom, there are two nightstands, ..." + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.343, + 0.404, + 0.349 + ], + "angle": 0, + "content": "The backpack is in front of the nightstand as well . The roomapproves to be functional with the nightstands providing" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.35, + 0.382, + 0.357 + ], + "angle": 0, + "content": "storage space and the telephone for communication" + }, + { + "type": "text", + "bbox": [ + 0.442, + 0.252, + 0.471, + 0.26 + ], + "angle": 0, + "content": "" + }, + { + "type": "text", + "bbox": [ + 0.474, + 0.252, + 0.538, + 0.26 + ], + "angle": 0, + "content": "(b) Language radar" + }, + { + "type": "text", + "bbox": [ + 0.54, + 0.252, + 0.579, + 0.26 + ], + "angle": 0, + "content": "ar chart and" + }, + { + "type": "text", + "bbox": [ + 0.472, + 0.264, + 0.538, + 0.271 + ], + "angle": 0, + "content": "Object Caption" + }, + { + "type": "text", + "bbox": [ + 0.418, + 0.275, + 0.48, + 0.283 + ], + "angle": 0, + "content": "BLIP2 Captions" + }, + { + "type": "text", + "bbox": [ + 0.421, + 0.284, + 0.514, + 0.29 + ], + "angle": 0, + "content": "1. A bed in a hotel room. (0.85)" + }, + { + "type": "text", + "bbox": [ + 0.421, + 0.29, + 0.534, + 0.297 + ], + "angle": 0, + "content": "2. A white comforter on a bed. (0.83)" + }, + { + "type": "text", + "bbox": [ + 0.421, + 0.297, + 0.534, + 0.303 + ], + "angle": 0, + "content": "3. A bed with a striped comforter. (0.83)" + }, + { + "type": "text", + "bbox": [ + 0.421, + 0.304, + 0.495, + 0.311 + ], + "angle": 0, + "content": "N.A picture of cat. (0.63)" + }, + { + "type": "text", + "bbox": [ + 0.418, + 0.313, + 0.479, + 0.322 + ], + "angle": 0, + "content": "Summary" + }, + { + "type": "text", + "bbox": [ + 0.421, + 0.322, + 0.581, + 0.329 + ], + "angle": 0, + "content": "Prompt: Summarize the captions below. The summary" + }, + { + "type": "text", + "bbox": [ + 0.421, + 0.329, + 0.572, + 0.335 + ], + "angle": 0, + "content": "should be a description of the (object). Focus on the" + }, + { + "type": "text", + "bbox": [ + 0.421, + 0.336, + 0.534, + 0.342 + ], + "angle": 0, + "content": "{object's attributes, like color, shape, n" + }, + { + "type": "text", + "bbox": [ + 0.421, + 0.343, + 0.576, + 0.349 + ], + "angle": 0, + "content": "Identity and correct the potential errors ..." + }, + { + "type": "text", + "bbox": [ + 0.421, + 0.35, + 0.575, + 0.357 + ], + "angle": 0, + "content": "Response : The bed is in a motel room with a stripedcomforter . It has a white comforter and a blanket on it" + }, + { + "type": "text", + "bbox": [ + 0.421, + 0.357, + 0.58, + 0.364 + ], + "angle": 0, + "content": "The bed is also in a room with a bedside table." + }, + { + "type": "text", + "bbox": [ + 0.446, + 0.368, + 0.574, + 0.377 + ], + "angle": 0, + "content": "(c) Automated language generation" + }, + { + "type": "text", + "bbox": [ + 0.657, + 0.264, + 0.696, + 0.271 + ], + "angle": 0, + "content": "Object R" + }, + { + "type": "text", + "bbox": [ + 0.606, + 0.275, + 0.688, + 0.283 + ], + "angle": 0, + "content": "Relationship Triplets" + }, + { + "type": "text", + "bbox": [ + 0.609, + 0.284, + 0.674, + 0.289 + ], + "angle": 0, + "content": "1. {table', 'chair', 'left'}" + }, + { + "type": "text", + "bbox": [ + 0.608, + 0.29, + 0.726, + 0.296 + ], + "angle": 0, + "content": "2.{'bed', ('lamp', 'mini fridge'), 'between'}" + }, + { + "type": "text", + "bbox": [ + 0.608, + 0.297, + 0.703, + 0.303 + ], + "angle": 0, + "content": "Template-based Referral" + }, + { + "type": "text", + "bbox": [ + 0.606, + 0.304, + 0.741, + 0.311 + ], + "angle": 0, + "content": "1. The table is to the left of the chair. \n2. The chair is in the right of the chair." + }, + { + "type": "text", + "bbox": [ + 0.606, + 0.312, + 0.667, + 0.318 + ], + "angle": 0, + "content": "2. It's a bed in the m" + }, + { + "type": "text", + "bbox": [ + 0.609, + 0.319, + 0.667, + 0.324 + ], + "angle": 0, + "content": "Rephrasing" + }, + { + "type": "text", + "bbox": [ + 0.609, + 0.324, + 0.779, + 0.33 + ], + "angle": 0, + "content": "Prompt: Rewrite the following sentence using one random word." + }, + { + "type": "text", + "bbox": [ + 0.609, + 0.33, + 0.708, + 0.336 + ], + "angle": 0, + "content": "sentence structure. Focus on the" + }, + { + "type": "text", + "bbox": [ + 0.609, + 0.336, + 0.681, + 0.342 + ], + "angle": 0, + "content": "about the {target_object}" + }, + { + "type": "text", + "bbox": [ + 0.609, + 0.343, + 0.643, + 0.347 + ], + "angle": 0, + "content": "Response:" + }, + { + "type": "text", + "bbox": [ + 0.609, + 0.348, + 0.75, + 0.353 + ], + "angle": 0, + "content": "1. The table is situated to the left of the armchair." + }, + { + "type": "text", + "bbox": [ + 0.609, + 0.353, + 0.776, + 0.36 + ], + "angle": 0, + "content": "2. The bed occupies the space between the lamp and the wall." + }, + { + "type": "text", + "bbox": [ + 0.609, + 0.36, + 0.724, + 0.365 + ], + "angle": 0, + "content": "mini fridge, creating a cozy atmosphere." + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.381, + 0.787, + 0.424 + ], + "angle": 0, + "content": "Fig. 2: SCENEVERSE collection and statistics. Given a 3D scene (a), our automated pipeline (c) generates three types of description including scene caption, object caption and object referral. (b) SCENEVERSE data comparison and composition." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.432, + 0.535, + 0.447 + ], + "angle": 0, + "content": "3.4 Language Generation with LLMs" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.454, + 0.788, + 0.53 + ], + "angle": 0, + "content": "The scene-language pairs in ScENEVERSE aim to capture varying aspects of the 3D scene, including detailed object attributes in object captioning, spatial relationships between objects in object referral, and global scene descriptions in scene captioning. Based on 3D scene graphs, we utilize both templates and LLMs to automatically generate descriptions on these three granularities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.534, + 0.789, + 0.67 + ], + "angle": 0, + "content": "Object Captioning Object captions aim to provide detailed descriptions of an object's visual and physical properties, facilitating object-level grounding with its distinctive features. Given the multi-view images, we utilize the point cloud of the object \\( v \\in V \\) to identify its occurrence in the images through rendering. The images are then cropped with the rendered bounding boxes and processed through BLIP2 [56] to generate initial object captions. We select the top 10 sentences with the highest CLIP [75] similarity score and minimal occlusion and utilize an LLM to obtain a refined and coherent summary of the captions. The detailed object captioning pipeline is illustrated in supplementary." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Object Referral Object relationship captions refer to objects by articulating their spatial relationships in the scene. Spatial relationship triplets \\((v_{i}, v_{j}, e_{ij})\\) are first extracted from the constructed 3D scene graph. We design various templates to generate descriptions for each relationship type, assigning the entities in the form of (target-object, spatial-relation, anchor-object(s)). This results in examples like \"the chair is next to the li2023behaviorarmchair\", \"facing the sofa, there is a suitcase far to the right of the shoes\", and \"the fridge is between cabinet and sofa\". Our designed templates span passive and active tenses, as well as inversion clauses, contributing to the richness of the generated text. To enhance the descriptions' naturalness, we employ LLM for sentence rephrasing. Statistics for the descriptions before and after rephrasing are presented in Fig. 2 (b)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.242, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "Scaling 3D Vision-Language Learning for Grounded Scene Understanding" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "Scene Captioning The scene-level captions emphasize global information, portraying the key objects in the scene along with their attributes and functionalities. We use the constructed 3D scene graph and prompt LLMs to generate these captions. We random sample a subset of edges and nodes from the scene graph each time as the scene context to enhance the diversity of scene captions. The object counts are also provided as LLM prompts, together with the room type and object attributes if such annotations are available in the dataset." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.277, + 0.489, + 0.292 + ], + "angle": 0, + "content": "3.5 Data Quality and Statistics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.302, + 0.789, + 0.439 + ], + "angle": 0, + "content": "Data Quality The 96K human-annotated set of ScENeVErSE is collected through AMT, where 82 humans are employed for annotation and 18 for verification. All final annotations passed the reference verification, with a re-annotation rate of \\(4.8\\%\\). For our automatic language generation pipeline, we conduct extensive prompt tuning and iterate with human feedback for LLMs on object captioning, summary, and rephrasing. To verify the efficacy of the pipeline, we conduct a quality check where 12K generated object-level descriptions are randomly selected for human verification. Results demonstrate a \\(96.93\\%\\) pass rate, surpassing that in ReferIt3D [1] with \\(86.1\\%\\) pass rate on 2K samples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.447, + 0.79, + 0.629 + ], + "angle": 0, + "content": "Statistics In total, SCENEVERSE comprises a total of 68,406 room-level 3D scans, with the source composition shown in Fig. 2 (b). The dataset contains 1.5M object instances ranging in 2290 object categories. Our generated 3D scene graph comprises 21 types of relationships following prior work [1,86]. For the language descriptions, we generate 1M template-based texts and 1M sentences rephrased by Llama [83] and GPT-3.5 [71]. As can be seen from the radar chart and examples in Fig. 2, the diversity of the LLM-refined data, particularly in sentence length and variety, closely aligns with the characteristics of annotated descriptions, surpassing the template-based data. Together with existing sources (294K) and our newly annotated set (96K), SCENEVERSE contains 2.5M scene-language pairs in total. All the rephrasing and summary prompts, along with the complete set of relationships, are detailed in supplementary." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.653, + 0.58, + 0.67 + ], + "angle": 0, + "content": "4 Grounded Pre-training for Scenes" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.685, + 0.789, + 0.761 + ], + "angle": 0, + "content": "In this section, we introduce GPS, an efficient transformer-based model trained with multi-level contrastive losses for aligning 3D scenes and texts. As shown in Fig. 3, we echo the language descriptions collected at different granularities to form contrastive objectives at both object-level, referral-object-level, and scene-level in GPS. We describe the design of each level in the following sections." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.784, + 0.458, + 0.8 + ], + "angle": 0, + "content": "4.1 Object-level Grounding" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.843 + ], + "angle": 0, + "content": "Given a 3D scene point cloud \\(\\mathcal{S}\\), we use an off-the-shelf 3D object segmentation model to decompose it into a bag of \\(N\\) objects \\(\\mathcal{S} = \\{\\pmb{o}_1, \\pmb{o}_2, \\dots, \\pmb{o}_n\\}_{i=1}^N\\). We extract" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.442, + 0.128 + ], + "angle": 0, + "content": "B. Jia and Y. Chen et al." + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.146, + 0.789, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.311, + 0.789, + 0.34 + ], + "angle": 0, + "content": "Fig. 3: Overview of GPS model. We use contrastive alignment at three levels \\(\\mathcal{L}_{\\mathrm{obj}}\\), \\(\\mathcal{L}_{\\mathrm{scene}}\\), and \\(\\mathcal{L}_{\\mathrm{ref}}\\) and a masked language modeling objective \\(\\mathcal{L}_{\\mathrm{MLM}}\\) for model learning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.35, + 0.789, + 0.399 + ], + "angle": 0, + "content": "object features \\(\\{\\pmb{f}_i^O\\}\\) with an object point cloud encoder and text features \\(\\{\\pmb{f}_i^T\\}\\) by feeding object-captions \\(\\{\\pmb{T}_i^{\\mathrm{obj}}\\}\\) into a frozen language model. Following [93], we perform cross-modal alignment on the object features and text features via:" + }, + { + "type": "equation", + "bbox": [ + 0.24, + 0.409, + 0.789, + 0.451 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {o b j}} = - \\frac {1}{2} \\sum_ {(p, q)} \\left(\\log \\frac {\\exp \\left(D ^ {\\mathrm {o b j}} (p , q)\\right)}{\\sum_ {r} \\exp \\left(D ^ {\\mathrm {o b j}} (p , r)\\right)} + \\log \\frac {\\exp \\left(D ^ {\\mathrm {o b j}} (p , q)\\right)}{\\sum_ {r} \\exp \\left(D ^ {\\mathrm {o b j}} (r , q)\\right)}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.79, + 0.525 + ], + "angle": 0, + "content": "where \\( D^{\\mathrm{obj}}(p,q) = (\\pmb{f}_p^O\\pmb{f}_q^T /\\tau) \\) denotes the dot product between object and text features and \\((p,q)\\) denotes a pair of aligned object-text pair in the training batch and \\(r\\) iterates over all object-text pairs in the training batch. Similar to CLIP [75], we use a learnable temperature parameter \\(\\tau\\) to facilitate model learning." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.546, + 0.449, + 0.563 + ], + "angle": 0, + "content": "4.2 Scene-level Grounding" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.571, + 0.789, + 0.633 + ], + "angle": 0, + "content": "With aligned object features, we encode the scene by incorporating object spatial locations into the extracted object features. Specifically, we use a spatial transformer model to encode extracted object features \\(\\{\\pmb{f}_i^O\\}\\) with their spatial location features \\(\\{l_i\\}\\) following [18, 109]:" + }, + { + "type": "equation", + "bbox": [ + 0.394, + 0.642, + 0.609, + 0.661 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {f} ^ {S} = \\operatorname {S p a t i a l A t t n} (\\{\\boldsymbol {f} _ {i} ^ {O} \\}, \\{\\boldsymbol {l} _ {i} \\})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.671, + 0.789, + 0.779 + ], + "angle": 0, + "content": "where \\(\\{\\pmb{f}_i^S\\}\\) denotes the feature of object \\(\\pmb{o}_i\\) after encoding with spatial location features. To perform scene-level alignment, we operate on these scene-level object features \\(\\{\\pmb{f}_i^S\\}\\) and align it with the scene caption \\(\\pmb{T}^{\\mathrm{scene}}\\). Specifically, we feed the object features into a projection layer and use max-pooling over all object features to obtain the scene feature \\(\\pmb{g}^S\\). Similar to object-level grounding, we pass the scene caption through a tunable language model to obtain text feature \\(\\pmb{g}^T\\) and perform scene-level contrastive alignment through:" + }, + { + "type": "equation", + "bbox": [ + 0.232, + 0.786, + 0.788, + 0.842 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s c e n e}} = - \\frac {1}{2} \\sum_ {(p, q)} \\left(\\log \\frac {\\exp \\left(D ^ {\\text {s c e n e}} (p , q)\\right)}{\\sum_ {r} \\exp \\left(D ^ {\\text {s c e n e}} (p , r)\\right)} + \\log \\frac {\\exp \\left(D ^ {\\text {s c e n e}} (p , q)\\right)}{\\sum_ {r} \\exp \\left(D ^ {\\text {s c e n e}} (r , q)\\right)}\\right), \\tag {2}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.242, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Scaling 3D Vision-Language Learning for Grounded Scene Understanding" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.195 + ], + "angle": 0, + "content": "where \\( D^{\\mathrm{scene}}(p,q) = (\\pmb{g}_p^S\\pmb{g}_q^T / \\tau) \\) denotes the dot product between scene feature \\( \\pmb{g}_p^S \\) and scene caption feature \\( \\pmb{g}_q^T \\) for each pair of aligned scene-text pairs in the training batch and \\( r \\) iterates over all scene-text pairs in the training batch." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.212, + 0.527, + 0.228 + ], + "angle": 0, + "content": "4.3 Referral-object-level Grounding" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.232, + 0.789, + 0.369 + ], + "angle": 0, + "content": "To model the relationships revealed in referring expressions, we employ a self-attention-based reasoning transformer for grounding object referrals in scenes. This transformer takes in scene-object features \\(\\{\\pmb{f}_i^S\\}\\) and an object referral \\(T^{\\mathrm{ref}}\\) and performs self-attention to learn relationships between text descriptions and object relationships. We use the same tunable language encoder as in scene-level grounding for extracting per-object referral features. We pass this text feature together with scene-object features into the self-attention transformer to obtain the aligned object features \\(\\pmb{h}_i^S\\) and the sentence-level referral feature \\(\\pmb{h}^T\\). We then perform the referral-object-level contrastive alignment following:" + }, + { + "type": "equation", + "bbox": [ + 0.391, + 0.374, + 0.788, + 0.412 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {r e f}} = - \\log \\frac {\\exp \\left(\\bar {\\boldsymbol {h}} ^ {S} \\boldsymbol {h} ^ {T} / \\tau\\right)}{\\sum_ {p} \\exp \\left(\\boldsymbol {h} _ {p} ^ {S} \\boldsymbol {h} ^ {T} / \\tau\\right)}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.416, + 0.789, + 0.508 + ], + "angle": 0, + "content": "where \\(\\bar{h}^S\\) denotes the feature of the referred object, \\(p\\) iterates over all objects within the same scene. Notably, in contrast to inter-scene contrast that was done in object- and scene-level alignment, we force the selection of positive pairs to be within the same scene to provide intra-scene contrast for fine-grained object grounding. This mimics the success of intra-image and inter-image contrasts commonly used for region-word alignment in 2D-VL models [101]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.79, + 0.614 + ], + "angle": 0, + "content": "To learn the multi-level alignment between 3D scenes and language, we first train the point cloud encoder with an object-level grounding objective to obtain a good feature initialization for grounding objects in scenes. During the scene grounding stage, we train our inter- and intra-scene objectives together with a masked language modeling loss \\(\\mathcal{L}_{\\mathrm{MLM}}\\) over the inputted object-referral texts to tune the parameters within the language encoder and self-attention transformer. Above all, the learning of GPS could be summarized as optimizing:" + }, + { + "type": "equation", + "bbox": [ + 0.382, + 0.621, + 0.625, + 0.637 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {o b j}} + \\mathcal {L} _ {\\mathrm {s c e n e}} + \\mathcal {L} _ {\\mathrm {r e f}} + \\mathcal {L} _ {\\mathrm {M L M}}.\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.655, + 0.377, + 0.672 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.682, + 0.789, + 0.698 + ], + "angle": 0, + "content": "In this section, we present experimental results addressing the following questions:" + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.701, + 0.788, + 0.746 + ], + "angle": 0, + "content": "1. How effective is the data scaling in ScENEVERSE for 3D visual grounding? Does the scale-up benefit common 3D-VL tasks (e.g., 3D question answering, open-vocabulary 3D semantic segmentation) and pre-training-based models?" + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.746, + 0.788, + 0.774 + ], + "angle": 0, + "content": "2. How well is the GPS pre-training pipeline for 3D-VL tasks? Does it exhibit similar properties of 2D-VL models in 3D-VL tasks?" + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.775, + 0.71, + 0.79 + ], + "angle": 0, + "content": "3. What is offered by ScENEVERSE and GPS and what is missing?" + }, + { + "type": "list", + "bbox": [ + 0.222, + 0.701, + 0.788, + 0.79 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.789, + 0.841 + ], + "angle": 0, + "content": "In the following sections, we describe in detail the model performance regarding these key topics. Due to the page limit, we direct readers to the supplementary for implementation details, qualitative results, and more experimental analyses." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.441, + 0.128 + ], + "angle": 0, + "content": "B. Jia and Y. Chen et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 2: 3D visual grounding results on Nr3D, Sr3D, and ScanRefer. We use \"pre-train\" for our model trained on ScENeVERse w/o additional fine-tuning, and \"fine-tune\" for its data-specific fine-tuned version. Best results are highlighted in bold." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.188, + 0.785, + 0.401 + ], + "angle": 0, + "content": "
MethodNr3DSr3DScanRefer Acc@0.5
OverallEasyHardV-Dep.V-Indep.OverallEasyHardV-Dep.V-Indep.OverallUniqueMultiple
3DVG-Trans [105]40.848.534.834.843.751.454.244.944.651.734.760.628.4
TGNN [42]37.344.230.635.838.045.048.536.945.845.029.756.823.2
TransRefer3D [37]48.056.739.642.550.757.460.550.249.957.7---
InstanceRefer [100]38.846.031.834.541.948.051.140.545.848.132.966.824.7
FFL-3DOG [32]41.748.235.037.144.7-----34.067.925.7
LAR [6]48.958.442.347.452.159.463.051.250.059.1---
SAT [97]56.564.948.454.457.657.961.250.049.258.330.150.825.2
3D-SPS [62]51.558.145.148.053.262.656.265.449.263.237.066.729.8
3DJCG [12]----------37.364.330.8
BUTD-DETR [45]54.660.748.446.058.067.068.663.253.067.639.866.335.1
MVT [43]59.567.452.759.160.364.566.958.858.464.733.366.525.3
ViL3DRel [18]64.470.257.462.064.572.874.967.963.873.237.768.630.7
EDA [92]52.158.246.150.253.168.170.362.954.168.742.368.637.6
3D-VisTA (scratch) [109]57.565.949.453.759.469.672.163.657.970.141.570.934.8
3D-VisTA [109]64.272.156.761.565.176.478.871.358.977.345.875.139.1
Ours (scratch)58.767.050.955.859.868.470.563.453.169.040.471.334.7
Ours (pre-train)55.262.848.045.558.874.176.468.554.175.047.177.441.6
Ours (fine-tuned)64.972.557.856.967.977.580.171.662.878.248.177.942.7
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.423, + 0.44, + 0.438 + ], + "angle": 0, + "content": "5.1 3D Visual Grounding" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.466, + 0.788, + 0.604 + ], + "angle": 0, + "content": "Settings We evaluate our model on three commonly-used datasets for 3D visual grounding: ScanRefer [16], Nr3D, and Sr3D [1]. For Nr3D and Sr3D, we follow Achlioptas et al. [1] and report the grounding accuracies of models using ground-truth object masks. For ScanRefer, we follow Zhu et al. [109] and use Mask3D [80] to generate object proposals. Results are reported as Acc@0.5 to evaluate the correctness of predictions whose object bounding boxes overlap the ground truth with IoU \\(>0.5\\). For comparisons, we compare with existing baselines by providing the results of pre-trained GPS and dataset-specific fine-tuned GPS. Please see more details in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.789, + 0.84 + ], + "angle": 0, + "content": "Results and Analyses As shown in Tab. 2, GPS trained on SCENEVERSE achieves state-of-the-art results on all existing 3D-VL grounding benchmarks. Initially, when GPS is trained directly on the training sets of benchmark datasets, labeled as Ours (scratch), it underperforms compared to existing models that employ more complex structures or loss designs. This result underscores the data-intensive nature of the contrastive alignment paradigm. However, when presented with extensive training data in SCENEVERSE, the results of our model without additional fine-tuning, i.e., Ours (pre-train), significantly improves and already achieves state-of-the-art results on benchmarks like ScanRefer. Moreover, the dataset-specific fine-tuned model, i.e., Ours (fine-tuned), consistently outperforms existing baselines with only a simple projection MLP added on top of the pretrained model, jointly optimized during fine-tuning without any other auxiliary architecture or loss objective. These results underscore the strong potential of both the SCENEVERSE and GPS for 3D-VL tasks." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.242, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Scaling 3D Vision-Language Learning for Grounded Scene Understanding" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.508, + 0.171 + ], + "angle": 0, + "content": "Table 3: Zero-shot transfer on existing benchmarks. \"SR\" stands for ScanRefer." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.173, + 0.505, + 0.265 + ], + "angle": 0, + "content": "
MethodNr3DSr3DSR@0.25SR@0.5
3D-VisTA (scratch)57.569.645.941.5
3D-VisTA (zero-shot)35.231.233.229.6
3D-VisTA (zero-shot text)43.136.141.136.4
Ours (scratch)58.768.444.540.4
Ours (zero-shot)32.433.335.231.1
Ours (zero-shot text)41.938.140.735.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.518, + 0.145, + 0.788, + 0.186 + ], + "angle": 0, + "content": "Table 4: Zero-shot transfer on ScENeVERSE-val. Evaluation uses GT object proposals following Nr3D/Sr3D." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.187, + 0.785, + 0.263 + ], + "angle": 0, + "content": "
MethodOverallEasyHardV-Dep.V-Indep.
3D-VisTA (scratch)40.753.121.637.344.3
3D-VisTA (zero-shot)52.959.635.453.752.2
3D-VisTA (zero-shot text)58.170.039.652.564.1
Ours (scratch)38.550.220.833.743.9
Ours (zero-shot)59.269.444.053.166.3
Ours (zero-shot text)60.670.945.154.867.3
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.283, + 0.419, + 0.297 + ], + "angle": 0, + "content": "5.2 Zero-Shot Transfer" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.304, + 0.789, + 0.517 + ], + "angle": 0, + "content": "Settings To better evaluate the effectiveness of both the SCENEVERSE data and the GPS model, we further perform zero-shot transfer experiments to test the models' capability in 4 benchmarks, ScanRefer, Sr3D, Nr3D, and SCENEVERSE- val. We create SCENEVERSE-val using \\(8.5K\\) annotated object referrals of 271 scenes in MultiScan, and randomly split the scenes following a 4:1 train / test split for creating the held-out test set. We mainly consider 2 specific transfer settings in our experiments: (i) zero-shot: models trained by removing all the scenes from the target dataset, tested on held-out unseen scenes, and (ii) zero-shot text: Models trained on data that include the training set of scenes from the target dataset, yet tested exclusively with unseen scene-text distribution. Specifically, for the zero-shot text setting, we use the generated texts in SCENEVERSE as fine-tuning sources for the zero-shot model. We mainly compare our model against a recent pre-training-based model 3D-VisTA. See more details on experimental setting and implementation in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.521, + 0.786, + 0.551 + ], + "angle": 0, + "content": "Results and Analyses We present the results of zero-shot transfer experiments in Tab. 3 and Tab. 4 with the following key observations:" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.552, + 0.788, + 0.641 + ], + "angle": 0, + "content": "- Our GPS model demonstrates superior generalization to unseen scenes compared to the 3D-VisTA model. In zero-shot transfer scenarios, our model consistently outperforms 3D-VisTA across established benchmarks and ScEN-EVERSE-val. This indicates the effectiveness of contrastive alignment over traditional classification objectives, aligning with the advancements seen in 2D-VL models for open-vocabulary grounding and transfer capabilities" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.642, + 0.788, + 0.762 + ], + "angle": 0, + "content": "- SCENEVERSE dataset substantially enhances 3D-VL grounding capabilities through zero-shot transfer, especially when provided with relatively limited training data, i.e., SCENEVERSE-val. As demonstrated in Tab. 4, there is a significantly improved performance when comparing models trained on SCENEVERSE in a zero-shot manner to those trained from scratch. This indicates that SCENEVERSE can effectively capture knowledge for general 3D scene grounding. Consequently, this underscores its potential as a go-to pre-training dataset for 3D-VL tasks." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.763, + 0.788, + 0.839 + ], + "angle": 0, + "content": "- The impact of our extensive collection and scalable generation of scene-text pairs is further evidenced by the results in the zero-shot text setting. Notably, as shown in Tab. 3, these automatically generated scene-text pairs supply ample knowledge for comprehending the scene distribution. This contributes significantly to the substantial improvement over the zero-shot performance." + }, + { + "type": "list", + "bbox": [ + 0.227, + 0.552, + 0.788, + 0.839 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.442, + 0.128 + ], + "angle": 0, + "content": "B. Jia and Y. Chen et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.78, + 0.198 + ], + "angle": 0, + "content": "Table 5: 3D question answering re- Table 6: Existing 3D backbones pre-sults on ScanQA and SQA3D. We re-trained on SCENEVERSE for openport EM@1 score on ScanQA and SQA3D vocabulary 3D semantic segmentation on evaluation sets. ScanNet. \"SPUNet\" denotes SparseUNet pro" + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.2, + 0.484, + 0.308 + ], + "angle": 0, + "content": "
ModelScanQASQA3D
valw/objw/o obj
ScanRefer+MCAN [5]18.620.619.0-
ScanQA [5]20.323.520.946.6
SQA3D [65]---47.2
3D-VisTA [109]22.427.023.048.5
3D-LLM [40]20.519.1--
Ours22.725.023.549.9
" + }, + { + "type": "table", + "bbox": [ + 0.489, + 0.215, + 0.776, + 0.311 + ], + "angle": 0, + "content": "
ModelNetworkmIoUΔmAccΔ
OpenScene [74]SPUNet1657.2-69.9-
PLA [30]SPUNet1617.7-33.5-
RegionPLC [94]SPUNet1656.9-75.6-
RegionPLC+SCENEVERSESPUNet1658.2+1.7%77.3+2.2%
OpenScene [74]SPUNet3257.8-70.3-
PLA [30]SPUNet3219.1-41.5-
RegionPLC [94]SPUNet3259.6-77.5-
RegionPLC+SCENEVERSESPUNet3261.0+2.3%79.7+2.8%
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.32, + 0.465, + 0.336 + ], + "angle": 0, + "content": "5.3 Additional 3D-VL Tasks" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.341, + 0.787, + 0.387 + ], + "angle": 0, + "content": "Settings We evaluate the effectiveness of GPS and ScENEVERSE on additional 3D-VL tasks: (i) 3D question answering (3D-QA) on ScanQA [5] and SQA3D [65], and (ii) open-vocabulary 3D semantic segmentation (OV-Seg) on ScanNet." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.388, + 0.787, + 0.447 + ], + "angle": 0, + "content": "- In the 3D-QA task, we follow Zhu et al. [109] and evaluate models over the exact match metric (EM@1) on the validation and test sets of ScanQA, as well as the test set of SQA3D. We pre-train GPS on SCENEVERSE and fine-tune the model on the 3D-QA dataset to compare with state-of-the-art models." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.448, + 0.788, + 0.538 + ], + "angle": 0, + "content": "- In the OV-Seg task, as GPS builds upon an object-centric design and thus is not directly applicable to semantic segmentation, we consider testing the effectiveness of ScENEVERSE on improving existing 3D models. Specifically, we follow the open-vocabulary semantic segmentation settings proposed by Yang et al. [94] and report the mIoU and mAcc score. We compare with existing works by pre-training the RegionPLC [94] model on ScENEVERSE." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.388, + 0.788, + 0.538 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.543, + 0.786, + 0.573 + ], + "angle": 0, + "content": "Results and Analyses We present the results of 3D-QA experiments in Tab. 5 and the results of OV-Seg experiments in Tab. 6. The analyses are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.574, + 0.787, + 0.663 + ], + "angle": 0, + "content": "- As shown in Tab. 5, our model achieves state-of-the-art results on both benchmarks, outperforming recent strong pre-training-based baselines like 3D-VisTA and 3D-LLM. As SCENEVERSE currently contains only descriptions of objects and scenes, we believe involving more types of language descriptions (e.g., question-answer pairs, dialogues) is a promising direction for further improving model performance on these downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.664, + 0.788, + 0.77 + ], + "angle": 0, + "content": "- As shown in Tab. 6, we observe consistent performance improvement of existing 3D backbone models on this task when pre-trained with SCENEVERSE data. This result validates that the collected data in SCENEVERSE can effectively boost the performance of existing models on scene understanding tasks. We further provide results of state-of-the-art 3D models that are pretrained on SCENEVERSE on the close-vocabulary 3D semantic segmentation task in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.789, + 0.528, + 0.803 + ], + "angle": 0, + "content": "5.4 Ablative Studies and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In this section, we present further discussions on both the data collected in ScENEVERSE and the GPS model design. We aim to elucidate the effects of" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.242, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Scaling 3D Vision-Language Learning for Grounded Scene Understanding" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.363, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.369, + 0.145, + 0.511, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.259, + 0.516, + 0.328 + ], + "angle": 0, + "content": "Fig.4: Model performance v.s. data scale. Plots show that models consistently improve in both the pre-train and zero-shot transfer settings on ScanRefer and ScENEV-ERSE-val with data scaling-up." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.333, + 0.516, + 0.374 + ], + "angle": 0, + "content": "Table 8: Cross domain transfer results of models pre-trained on real and synthetic datasets. \"S3D\" stands for Structured3D." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.376, + 0.51, + 0.435 + ], + "angle": 0, + "content": "
RealSyntheticSCENEVERSE-valS3DProcTHOR
AllX64.837.143.4
XS3D7.085.116.1
XProcTHOR4.216.391.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.52, + 0.148, + 0.788, + 0.203 + ], + "angle": 0, + "content": "Table 7: Ablation on text data source used in model pre-training. All models are tested on ScanRefer with no additional finetuning." + }, + { + "type": "table", + "bbox": [ + 0.527, + 0.207, + 0.782, + 0.282 + ], + "angle": 0, + "content": "
TemplateLLMAnno.Acc@0.25Acc@0.5
XXX43.538.4
✓XX50.946.1
✓✓X51.146.3
✓✓✓52.047.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.52, + 0.291, + 0.788, + 0.36 + ], + "angle": 0, + "content": "Table 9: Ablation on model design on ScENeVERSE-val. We use \"Obj- lvl\", \"Scene-lvl\" to denote object and scene alignment loss, and \"MLM\" for the mask language modeling loss." + }, + { + "type": "table", + "bbox": [ + 0.528, + 0.362, + 0.782, + 0.435 + ], + "angle": 0, + "content": "
Obj-lvl.MLMScene-lvl.OverallEasyHard
XXX64.875.448.7
✓XX65.277.147.4
✓✓X62.473.445.8
✓✓✓66.977.850.3
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.457, + 0.788, + 0.503 + ], + "angle": 0, + "content": "data scaling and show more clearly its effectiveness in 3D scene understanding. Regarding the experimental settings and more results discussion, refer to the supplementary. The following points are specifically discussed in this section:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.509, + 0.788, + 0.631 + ], + "angle": 0, + "content": "How important is data-scaling? We conduct ablation studies over the amount of data used while pre-training GPS. We consider the model trained with \\(\\frac{1}{8}\\), \\(\\frac{1}{4}\\), \\(\\frac{1}{2}\\) of ScENEVERSE to show the effectiveness of data-scaling on model performance in the pre-train and zero-shot transfer settings in ScanRefer and ScENEVERSE-val. As shown in Fig. 4, we observe consistent performance improvement over the increase of data scale for both settings. We provide additional experiments in the supplementary to show that such scaling effect is not only beneficial for 3D-VL grounding but also for other 3D tasks like semantic segmentation [80,96]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.637, + 0.788, + 0.803 + ], + "angle": 0, + "content": "How is the generated data compared with human-annotated data? We assess the performance of models trained using various scene-text sources, specifically focusing on their performance in the ScanRefer dataset without additional fine-tuning. As shown in Tab. 7, models trained with our template-based generated texts and Large Language Model (LLM)-refined texts show significant improvements over models trained solely on ScanRefer. More importantly, these variants of our model already achieve state-of-the-art results compared with previous baselines. This indicates the effectiveness of our text-generation pipeline. Finally, we observe that adding human-annotated data is still beneficial for model performance. However, the improvement is relatively marginal over models trained on our generated data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.84 + ], + "angle": 0, + "content": "What is the role of the synthetic scenes in this scale-up process? With synthetic data providing large-scale and diverse scene data for 3D-VL tasks," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.441, + 0.128 + ], + "angle": 0, + "content": "B. Jia and Y. Chen et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.358 + ], + "angle": 0, + "content": "we evaluate the models' domain transfer (Sim2Real) capability. Specifically, we compare models trained on all real scenes in SCENEVERSE against models trained exclusively on two synthetic subsets of SCENEVERSE, i.e., Structured3D and ProcTHOR. As shown in Tab. 8, models trained on synthetic subsets demonstrate remarkable performance on their corresponding test sets while suffering when transferred to real or other synthetic scenes. In contrast, the model trained on real scene-text pairs exhibits less performance drop when generalizing to synthetic scenes. This result affirms the domain gap between real and synthetic scenes in 3D-VL grounding and shows that a simple scale-up in the number of scenes is insufficient when naturalness can not be guaranteed. Considering the scalability of our language generation pipeline and the scaling effect shown in our experiments, the rate-determining step for further scaling-up 3D-VL comes to the collection of diverse, high-quality, and realistic scenes that capture natural 3D scene distributions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.37, + 0.79, + 0.583 + ], + "angle": 0, + "content": "How important is the design of each module in GPS? We provide ablative analyses of our multi-level contrastive alignment design in Tab. 9. We mainly consider removing objectives in our model to reveal the effectiveness of each level of alignment. We choose the referral-object-level alignment objective as the default setting and consider removing: (i) object-level alignment objective, (ii) masked language modeling objective, and (iii) scene-level alignment objective. When removing the object-level alignment objective, we learn the object point cloud encoder with the referral-object-level alignment and without pre-training. As shown in Tab. 9, we test different models on the SCENEVERSE-val without additional fine-tuning. Results show that the scene-level alignment objective is crucial for referral object grounding in SCENEVERSE-val with the \\(\\sim 5\\%\\) performance drop. Similar observations could be made for the model trained without object-level alignment (\\(\\sim 2\\%\\) drop) and masked language modeling objective (\\(\\sim 1.5\\%\\) drop). These results affirm the effectiveness of our model design." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.609, + 0.36, + 0.625 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.789, + 0.841 + ], + "angle": 0, + "content": "In this work, we scale up 3D-VL for grounded scene understanding. We present SCENEVERSE, a million-scale 3D-VL dataset covering various scenes and multi-level scene descriptions sourced from both human annotation and our proposed scene-text generation approach. Utilizing SCENEVERSE, we propose Grounded Pre-training for Scenes (GPS), a model trained with multi-level scene-language contrastive alignment. Through extensive experiments, we show that GPS achieves state-of-the-art results on common 3D-VL tasks including grounding and question answering. We further conduct zero-shot transfer experiments to show the improved generalization performances of GPS trained on SCENEVERSE compared with previous baselines. We also demonstrate that the scaling effect of SCENEVERSE is generally beneficial for existing 3D models on 3D-VL tasks like semantic segmentation. We hope our efforts and successful scale-up attempts in SCENEVERSE could pave the way for new research paradigms in 3D-VL." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.242, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Scaling 3D Vision-Language Learning for Grounded Scene Understanding" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.207 + ], + "angle": 0, + "content": "Acknowledgement The authors would like to thank Yaowei Zhang (BIGAI) for his help on online visualization and other colleagues from BIGAI General Vision Lab for fruitful discussions. The authors would also like to thank the anonymous reviewers for their constructive feedback." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.228, + 0.323, + 0.243 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.231, + 0.255, + 0.786, + 0.296 + ], + "angle": 0, + "content": "1. Achlioptas, P., Abdelreheem, A., Xia, F., Elhoseiny, M., Guibas, L.: Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In: Proceedings of European Conference on Computer Vision (ECCV) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.231, + 0.297, + 0.787, + 0.338 + ], + "angle": 0, + "content": "2. Agia, C., Jatavallabhula, K.M., Khodeir, M., Miksik, O., Vineet, V., Mukadam, M., Paull, L., Shkurti, F.: Taskography: Evaluating robot task planning over large 3d scene graphs. In: Proceedings of Conference on Robot Learning (CoRL) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.231, + 0.338, + 0.786, + 0.392 + ], + "angle": 0, + "content": "3. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.231, + 0.392, + 0.786, + 0.433 + ], + "angle": 0, + "content": "4. Armeni, I., He, Z.Y., Gwak, J., Zamir, A.R., Fischer, M., Malik, J., Savarese, S.: 3d scene graph: A structure for unified semantics, 3d space, and camera. In: Proceedings of International Conference on Computer Vision (ICCV) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.231, + 0.433, + 0.786, + 0.473 + ], + "angle": 0, + "content": "5. Azuma, D., Miyanishi, T., Kurita, S., Kawanabe, M.: Scanqa: 3d question answering for spatial scene understanding. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.231, + 0.473, + 0.786, + 0.514 + ], + "angle": 0, + "content": "6. Bakr, E., Alsaedy, Y., Elhoseiny, M.: Look around and refer: 2d synthetic semantics knowledge distillation for 3d visual grounding. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.231, + 0.514, + 0.786, + 0.542 + ], + "angle": 0, + "content": "7. Barsalou, L.W.: Perceptual symbol systems. Behavioral and brain sciences 22(4), 577-660 (1999)" + }, + { + "type": "ref_text", + "bbox": [ + 0.231, + 0.542, + 0.786, + 0.554 + ], + "angle": 0, + "content": "8. Barsalou, L.W.: Grounded cognition. Annu. Rev. Psychol. 59, 617-645 (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.231, + 0.554, + 0.786, + 0.623 + ], + "angle": 0, + "content": "9. Baruch, G., Chen, Z., Dehghan, A., Dimry, T., Feigin, Y., Fu, P., Gebauer, T., Joffe, B., Kurz, D., Schwartz, A., et al.: Arkitscenes: A diverse real-world dataset for 3d indoor scene understanding using mobile rgb-d data. In: Proceedings of Advances in Neural Information Processing Systems Datasets and Benchmarks (NeurIPS Datasets and Benchmarks Track) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.623, + 0.786, + 0.663 + ], + "angle": 0, + "content": "10. Bommasani, R., Hudson, D.A., Adeli, E., Altman, R., Arora, S., von Arx, S., Bernstein, M.S., Bohg, J., Bosselut, A., Brunskill, E., et al.: On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.663, + 0.786, + 0.718 + ], + "angle": 0, + "content": "11. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.718, + 0.786, + 0.759 + ], + "angle": 0, + "content": "12. Cai, D., Zhao, L., Zhang, J., Sheng, L., Xu, D.: 3djcg: A unified framework for joint dense captioning and visual grounding on 3d point clouds. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.759, + 0.786, + 0.8 + ], + "angle": 0, + "content": "13. Chang, A., Dai, A., Funkhouser, T., Halber, M., Niessner, M., Savva, M., Song, S., Zeng, A., Zhang, Y.: Matterport3d: Learning from rgb-d data in indoor environments. Proceedings of International Conference on 3D Vision (3DV) (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.786, + 0.84 + ], + "angle": 0, + "content": "14. Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., et al.: Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012 (2015)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.255, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.442, + 0.128 + ], + "angle": 0, + "content": "B. Jia and Y. Chen et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "15. Changpinyo, S., Sharma, P., Ding, N., Soricut, R.: Conceptual 12m: Pushing web-scale image-text pre-training to recognize long-tail visual concepts. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.189, + 0.787, + 0.23 + ], + "angle": 0, + "content": "16. Chen, D.Z., Chang, A.X., Nießner, M.: Scanrefer: 3d object localization in rgb-d scans using natural language. In: Proceedings of European Conference on Computer Vision (ECCV) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.23, + 0.787, + 0.271 + ], + "angle": 0, + "content": "17. Chen, D.Z., Wu, Q., Nießner, M., Chang, A.X.: D3net: a speaker-listener architecture for semi-supervised dense captioning and visual grounding in rgb-d scans. In: Proceedings of European Conference on Computer Vision (ECCV) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.271, + 0.787, + 0.311 + ], + "angle": 0, + "content": "18. Chen, S., Guhur, P.L., Tapaswi, M., Schmid, C., Laptev, I.: Language conditioned spatial relation reasoning for 3d object grounding. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.311, + 0.787, + 0.352 + ], + "angle": 0, + "content": "19. Chen, S., Zhu, H., Chen, X., Lei, Y., Yu, G., Chen, T.: End-to-end 3d dense captioning with vote2cap-detr. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.352, + 0.787, + 0.406 + ], + "angle": 0, + "content": "20. Chen, Y., Huang, S., Yuan, T., Qi, S., Zhu, Y., Zhu, S.C.: Holistic++ scene understanding: Single-view 3d holistic scene parsing and human pose estimation with human-object interaction and physical commonsense. In: Proceedings of International Conference on Computer Vision (ICCV) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.406, + 0.787, + 0.447 + ], + "angle": 0, + "content": "21. Chen, Y., Li, Q., Kong, D., Kei, Y.L., Zhu, S.C., Gao, T., Zhu, Y., Huang, S.: Yourfit: Embodied reference understanding with language and gesture. In: Proceedings of International Conference on Computer Vision (ICCV) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.447, + 0.787, + 0.488 + ], + "angle": 0, + "content": "22. Chen, Z., Gholami, A., Nießner, M., Chang, A.X.: Scan2cap: Context-aware dense captioning in rgb-d scans. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.487, + 0.787, + 0.542 + ], + "angle": 0, + "content": "23. Collins, J., Goel, S., Deng, K., Luthra, A., Xu, L., Gundogdu, E., Zhang, X., Vicente, T.F.Y., Dideriksen, T., Arora, H., et al.: Abo: Dataset and benchmarks for real-world 3d object understanding. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.542, + 0.787, + 0.582 + ], + "angle": 0, + "content": "24. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.582, + 0.787, + 0.623 + ], + "angle": 0, + "content": "25. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.622, + 0.787, + 0.678 + ], + "angle": 0, + "content": "26. Deitke, M., Liu, R., Wallingford, M., Ngo, H., Michel, O., Kusupati, A., Fan, A., Laforte, C., Voleti, V., Gadre, S.Y., et al.: Objverse-xl: A universe of \\(10\\mathrm{m} + 3\\mathrm{d}\\) objects. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.677, + 0.787, + 0.731 + ], + "angle": 0, + "content": "27. Deitke, M., Schwenk, D., Salvador, J., Weihs, L., Michel, O., VanderBilt, E., Schmidt, L., Ehsani, K., Kembhavi, A., Farhadi, A.: Objverse: A universe of annotated 3d objects. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.731, + 0.787, + 0.786 + ], + "angle": 0, + "content": "28. Deitke, M., VanderBilt, E., Herrasti, A., Weihs, L., Ehsani, K., Salvador, J., Han, W., Kolve, E., Kembhavi, A., Mottaghi, R.: Procthor: Large-scale embodied ai using procedural generation. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.786, + 0.787, + 0.84 + ], + "angle": 0, + "content": "29. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: Proceedings of Conference of the North American Chapter of the Association for Computational Linguistics (NAACL) (2018)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.242, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Scaling 3D Vision-Language Learning for Grounded Scene Understanding" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.147, + 0.786, + 0.189 + ], + "angle": 0, + "content": "30. Ding, R., Yang, J., Xue, C., Zhang, W., Bai, S., Qi, X.: Pla: Language-driven open-vocabulary 3d scene understanding. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.189, + 0.787, + 0.23 + ], + "angle": 0, + "content": "31. Ding, Z., Han, X., Niethammer, M.: Volenet: A deep learning label fusion method for multi-atlas segmentation. In: Proceedings of International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.23, + 0.787, + 0.284 + ], + "angle": 0, + "content": "32. Feng, M., Li, Z., Li, Q., Zhang, L., Zhang, X., Zhu, G., Zhang, H., Wang, Y., Mian, A.: Free-form description guided 3d visual graph network for object grounding in point cloud. In: Proceedings of International Conference on Computer Vision (ICCV) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.284, + 0.787, + 0.325 + ], + "angle": 0, + "content": "33. Ghiasi, G., Gu, X., Cui, Y., Lin, T.Y.: Scaling open-vocabulary image segmentation with image-level labels. In: Proceedings of European Conference on Computer Vision (ECCV) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.325, + 0.787, + 0.379 + ], + "angle": 0, + "content": "34. Gong, R., Huang, J., Zhao, Y., Geng, H., Gao, X., Wu, Q., Ai, W., Zhou, Z., Terzopoulos, D., Zhu, S.C., Jia, B., Huang, S.: Arnold: A benchmark for language-grounded task learning with continuous states in realistic 3d scenes. In: Proceedings of International Conference on Computer Vision (ICCV) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.379, + 0.787, + 0.434 + ], + "angle": 0, + "content": "35. Gu, Q., Kuwajerwala, A., Morin, S., Jatavallabhula, K.M., Sen, B., Agarwal, A., Rivera, C., Paul, W., Ellis, K., Chellappa, R., et al.: Conceptgraphs: Open-vocabulary 3d scene graphs for perception and planning. arXiv preprint arXiv:2309.16650 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.434, + 0.787, + 0.474 + ], + "angle": 0, + "content": "36. Ha, H., Song, S.: Semantic abstraction: Open-world 3d scene understanding from 2d vision-language models. In: Proceedings of Conference on Robot Learning (CoRL) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.474, + 0.787, + 0.515 + ], + "angle": 0, + "content": "37. He, D., Zhao, Y., Luo, J., Hui, T., Huang, S., Zhang, A., Liu, S.: Transrefer3d: Entity-and-relation aware transformer for fine-grained 3d visual grounding. In: Proceedings of ACM International Conference on Multimedia (MM) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.515, + 0.787, + 0.556 + ], + "angle": 0, + "content": "38. Hegde, D., Valanarasu, J.M.J., Patel, V.: Clip goes 3d: Leveraging prompt tuning for language grounded 3d recognition. In: Proceedings of International Conference on Computer Vision (ICCV) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.556, + 0.787, + 0.597 + ], + "angle": 0, + "content": "39. Hong, Y., Wu, Q., Qi, Y., Rodriguez-Opazo, C., Gould, S.: Vln bert: A recurrent vision-and-language bert for navigation. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.597, + 0.787, + 0.637 + ], + "angle": 0, + "content": "40. Hong, Y., Lin, C., Du, Y., Chen, Z., Tenenbaum, J.B., Gan, C.: 3d concept learning and reasoning from multi-view images. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.637, + 0.787, + 0.678 + ], + "angle": 0, + "content": "41. Huang, J., Yong, S., Ma, X., Linghu, X., Li, P., Wang, Y., Li, Q., Zhu, S.C., Jia, B., Huang, S.: An embodied generalist agent in 3d world. arXiv preprint arXiv:2311.12871 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.678, + 0.787, + 0.718 + ], + "angle": 0, + "content": "42. Huang, P.H., Lee, H.H., Chen, H.T., Liu, T.L.: Text-guided graph neural networks for referring 3d instance segmentation. In: Proceedings of AAAI Conference on Artificial Intelligence (AAAI) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.718, + 0.787, + 0.759 + ], + "angle": 0, + "content": "43. Huang, S., Chen, Y., Jia, J., Wang, L.: Multi-view transformer for 3d visual grounding. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.759, + 0.787, + 0.8 + ], + "angle": 0, + "content": "44. Huang, S., Wang, Z., Li, P., Jia, B., Liu, T., Zhu, Y., Liang, W., Zhu, S.C.: Diffusion-based generation, optimization, and planning in 3d scenes. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.8, + 0.787, + 0.841 + ], + "angle": 0, + "content": "45. Jain, A., Gkanatsios, N., Mediratta, I., Fragkiadaki, K.: Bottom up top down detection transformers for language grounding in images and point clouds. In: Proceedings of European Conference on Computer Vision (ECCV) (2022)" + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.442, + 0.128 + ], + "angle": 0, + "content": "B. Jia and Y. Chen et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "46. Jiang, L., Zhao, H., Shi, S., Liu, S., Fu, C.W., Jia, J.: Pointgroup: Dual-set point grouping for 3d instance segmentation. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.189, + 0.787, + 0.23 + ], + "angle": 0, + "content": "47. Jiang, N., Liu, T., Cao, Z., Cui, J., Zhang, Z., Chen, Y., Wang, H., Zhu, Y., Huang, S.: Full-body articulated human-object interaction. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.23, + 0.787, + 0.271 + ], + "angle": 0, + "content": "48. Jiang, N., Zhang, Z., Li, H., Ma, X., Wang, Z., Chen, Y., Liu, T., Zhu, Y., Huang, S.: Scaling up dynamic human-scene interaction modeling. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.271, + 0.787, + 0.311 + ], + "angle": 0, + "content": "49. Kaplan, J., McCandlish, S., Henighan, T., Brown, T.B., Chess, B., Child, R., Gray, S., Radford, A., Wu, J., Amodei, D.: Scaling laws for neural language models. arXiv preprint arXiv:2001.08361 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.311, + 0.787, + 0.379 + ], + "angle": 0, + "content": "50. Khanna, M., Mao, Y., Jiang, H., Haresh, S., Shacklett, B., Batra, D., Clegg, A., Undersander, E., Chang, A.X., Savva, M.: Habitat synthetic scenes dataset (hssd-200): An analysis of 3d scene scale and realism tradeoffs for objectgoal navigation. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.379, + 0.787, + 0.42 + ], + "angle": 0, + "content": "51. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: Proceedings of International Conference on Computer Vision (ICCV) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.42, + 0.787, + 0.475 + ], + "angle": 0, + "content": "52. Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., et al.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. In: International Journal of Computer Vision (IJCV) (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.475, + 0.787, + 0.502 + ], + "angle": 0, + "content": "53. Lake, B.M., Ullman, T.D., Tenenbaum, J.B., Gershman, S.J.: Building machines that learn and think like people. Behavioral and brain sciences 40, e253 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.502, + 0.787, + 0.542 + ], + "angle": 0, + "content": "54. Li, B., Weinberger, K.Q., Belongie, S., Koltun, V., Ranftl, R.: Language-driven semantic segmentation. In: Proceedings of International Conference on Learning Representations (ICLR) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.542, + 0.787, + 0.597 + ], + "angle": 0, + "content": "55. Li, C., Zhang, R., Wong, J., Gokmen, C., Srivastava, S., Martin-Martín, R., Wang, C., Levine, G., Lingelbach, M., Sun, J., et al.: Behavior-1k: A benchmark for embodied ai with 1,000 everyday activities and realistic simulation. In: Proceedings of Conference on Robot Learning (CoRL) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.597, + 0.787, + 0.638 + ], + "angle": 0, + "content": "56. Li, J., Li, D., Savarese, S., Hoi, S.: BLIP-2: bootstrapping language-image pretraining with frozen image encoders and large language models. In: Proceedings of International Conference on Machine Learning (ICML) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.638, + 0.787, + 0.678 + ], + "angle": 0, + "content": "57. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: Proceedings of International Conference on Machine Learning (ICML) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.678, + 0.787, + 0.732 + ], + "angle": 0, + "content": "58. Li, L.H., Zhang, P., Zhang, H., Yang, J., Li, C., Zhong, Y., Wang, L., Yuan, L., Zhang, L., Hwang, J.N., et al.: Grounded language-image pre-training. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.732, + 0.787, + 0.759 + ], + "angle": 0, + "content": "59. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.759, + 0.787, + 0.8 + ], + "angle": 0, + "content": "60. Liu, M., Shi, R., Kuang, K., Zhu, Y., Li, X., Han, S., Cai, H., Porikli, F., Su, H.: Openshape: Scaling up 3d shape representation towards open-world understanding. arXiv preprint arXiv:2305.10764 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.787, + 0.84 + ], + "angle": 0, + "content": "61. Liu, R., Wu, R., Van Hoorick, B., Tokmakov, P., Zakharov, S., Vondrick, C.: Zero-1-to-3: Zero-shot one image to 3d object. In: Proceedings of International Conference on Computer Vision (ICCV) (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.242, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Scaling 3D Vision-Language Learning for Grounded Scene Understanding" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "62. Luo, J., Fu, J., Kong, X., Gao, C., Ren, H., Shen, H., Xia, H., Liu, S.: 3d-sps: Single-stage 3d visual grounding via referred point progressive selection. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.19, + 0.787, + 0.231 + ], + "angle": 0, + "content": "63. Luo, T., Rockwell, C., Lee, H., Johnson, J.: Scalable 3d captioning with pretrained models. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.231, + 0.787, + 0.272 + ], + "angle": 0, + "content": "64. Ma, C.Y., Lu, J., Wu, Z., AlRegib, G., Kira, Z., Socher, R., Xiong, C.: Self-monitoring navigation agent via auxiliary progress estimation. In: Proceedings of International Conference on Learning Representations (ICLR) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.273, + 0.787, + 0.314 + ], + "angle": 0, + "content": "65. Ma, X., Yong, S., Zheng, Z., Li, Q., Liang, Y., Zhu, S.C., Huang, S.: Sqa3d: Situated question answering in 3d scenes. In: Proceedings of International Conference on Learning Representations (ICLR) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.315, + 0.787, + 0.355 + ], + "angle": 0, + "content": "66. Mao, Y., Zhang, Y., Jiang, H., Chang, A., Savva, M.: Multiscan: Scalable rgbd scanning for 3d environments with articulated objects. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.356, + 0.787, + 0.397 + ], + "angle": 0, + "content": "67. Misra, I., Girdhar, R., Joulin, A.: An end-to-end transformer model for 3d object detection. In: Proceedings of International Conference on Computer Vision (ICCV) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.398, + 0.787, + 0.438 + ], + "angle": 0, + "content": "68. Mittal, M., Yu, C., Yu, Q., Liu, J., Rudin, N., Hoeller, D., Yuan, J.L., Singh, R., Guo, Y., Mazhar, H., et al.: Orbit: A unified simulation framework for interactive robot learning environments. Robotics and Automation Letters (RA-L) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.439, + 0.787, + 0.493 + ], + "angle": 0, + "content": "69. Mo, K., Zhu, S., Chang, A.X., Yi, L., Tripathi, S., Guibas, L.J., Su, H.: Partnet: A large-scale benchmark for fine-grained and hierarchical part-level 3d object understanding. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.494, + 0.787, + 0.549 + ], + "angle": 0, + "content": "70. Mu, T., Ling, Z., Xiang, F., Yang, D., Li, X., Tao, S., Huang, Z., Jia, Z., Su, H.: Maniskill: Generalizable manipulation skill benchmark with large-scale demonstrations. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.55, + 0.743, + 0.563 + ], + "angle": 0, + "content": "71. OpenAI: Introducing chatgpt. https://openai.com/blog/chatgpt (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.564, + 0.733, + 0.577 + ], + "angle": 0, + "content": "72. OpenAI: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.578, + 0.787, + 0.618 + ], + "angle": 0, + "content": "73. Pashevich, A., Schmid, C., Sun, C.: Episodic transformer for vision-and-language navigation. In: Proceedings of International Conference on Computer Vision (ICCV) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.619, + 0.787, + 0.66 + ], + "angle": 0, + "content": "74. Peng, S., Genova, K., Jiang, C., Tagliasacchi, A., Pollefeys, M., Funkhouser, T., et al.: Openscene: 3d scene understanding with open vocabularies. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.66, + 0.787, + 0.715 + ], + "angle": 0, + "content": "75. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: Proceedings of International Conference on Machine Learning (ICML) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.716, + 0.787, + 0.784 + ], + "angle": 0, + "content": "76. Ramakrishnan, S.K., Gokaslan, A., Wijmans, E., Maksymets, O., Clegg, A., Turner, J., Undersander, E., Galuba, W., Westbury, A., Chang, A.X., et al.: Habitatmatterport 3d dataset (hm3d): 1000 large-scale 3d environments for embodied ai. In: Proceedings of Advances in Neural Information Processing Systems Datasets and Benchmarks (NeurIPS Datasets and Benchmarks Track) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.787, + 0.84 + ], + "angle": 0, + "content": "77. Rana, K., Haviland, J., Garg, S., Abou-Chakra, J., Reid, I., Suenderhauf, N.: Sayplan: Grounding large language models using 3d scene graphs for scalable robot task planning. In: Proceedings of Conference on Robot Learning (CoRL) (2023)" + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "20" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.442, + 0.128 + ], + "angle": 0, + "content": "B. Jia and Y. Chen et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "78. Rosinol, A., Violette, A., Abate, M., Hughes, N., Chang, Y., Shi, J., Gupta, A., Carlone, L.: Kimera: From slam to spatial perception with 3d dynamic scene graphs. International Journal of Robotics Research (IJRR) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.191, + 0.787, + 0.245 + ], + "angle": 0, + "content": "79. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.247, + 0.787, + 0.288 + ], + "angle": 0, + "content": "80. Schult, J., Engelmann, F., Hermans, A., Litany, O., Tang, S., Leibe, B.: Mask3d: Mask transformer for 3d semantic instance segmentation. In: Proceedings of International Conference on Robotics and Automation (ICRA) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.29, + 0.787, + 0.317 + ], + "angle": 0, + "content": "81. Smith, L., Gasser, M.: The development of embodied cognition: Six lessons from babies. Artificial life 11(1-2), 13-29 (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.319, + 0.787, + 0.359 + ], + "angle": 0, + "content": "82. Takmaz, A., Fedele, E., Sumner, R.W., Pollefeys, M., Tombari, F., Engelmann, F.: Openmask3d: Open-vocabulary 3d instance segmentation. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.361, + 0.787, + 0.402 + ], + "angle": 0, + "content": "83. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.404, + 0.787, + 0.444 + ], + "angle": 0, + "content": "84. Vu, T., Kim, K., Luu, T.M., Nguyen, T., Yoo, C.D.: Softgroup for 3d instance segmentation on point clouds. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.446, + 0.787, + 0.487 + ], + "angle": 0, + "content": "85. Wald, J., Avetisyan, A., Navab, N., Tombari, F., Nießner, M.: Rio: 3d object instance re-localization in changing indoor environments. In: Proceedings of International Conference on Computer Vision (ICCV) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.489, + 0.787, + 0.53 + ], + "angle": 0, + "content": "86. Wald, J., Dhamo, H., Navab, N., Tombari, F.: Learning 3d semantic scene graphs from 3d indoor reconstructions. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.531, + 0.787, + 0.586 + ], + "angle": 0, + "content": "87. Wang, T., Mao, X., Zhu, C., Xu, R., Lyu, R., Li, P., Chen, X., Zhang, W., Chen, K., Xue, T., et al.: Embodiedscan: A holistic multi-modal 3d perception suite towards embodied ai. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.587, + 0.787, + 0.642 + ], + "angle": 0, + "content": "88. Wang, X., Huang, Q., Celikyilmaz, A., Gao, J., Shen, D., Wang, Y.F., Wang, W.Y., Zhang, L.: Reinforced cross-modal matching and self-supervised imitation learning for vision-language navigation. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.643, + 0.787, + 0.699 + ], + "angle": 0, + "content": "89. Wang, Z., Chen, Y., Jia, B., Li, P., Zhang, J., Zhang, J., Liu, T., Zhu, Y., Liang, W., Huang, S.: Move as you say interact as you can: Language-guided human motion generation with scene affordance. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.7, + 0.787, + 0.741 + ], + "angle": 0, + "content": "90. Wang, Z., Chen, Y., Liu, T., Zhu, Y., Liang, W., Huang, S.: Humanise: Language-conditioned human motion generation in 3d scenes. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.742, + 0.787, + 0.798 + ], + "angle": 0, + "content": "91. Wu, T., Zhang, J., Fu, X., Wang, Y., Ren, J., Pan, L., Wu, W., Yang, L., Wang, J., Qian, C., et al.: Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "92. Wu, Y., Cheng, X., Zhang, R., Cheng, Z., Zhang, J.: Eda: Explicit text-decoupling and dense alignment for 3d visual grounding. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.242, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Scaling 3D Vision-Language Learning for Grounded Scene Understanding" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "21" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.147, + 0.788, + 0.204 + ], + "angle": 0, + "content": "93. Xue, L., Gao, M., Xing, C., Martin-Martin, R., Wu, J., Xiong, C., Xu, R., Niebles, J.C., Savarese, S.: Ulip: Learning a unified representation of language, images, and point clouds for 3d understanding. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.205, + 0.787, + 0.247 + ], + "angle": 0, + "content": "94. Yang, J., Ding, R., Wang, Z., Qi, X.: Regionplc: Regional point-language contrastive learning for open-world 3d scene understanding. arXiv preprint arXiv:2304.00962 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.248, + 0.786, + 0.288 + ], + "angle": 0, + "content": "95. Yang, Y., Jia, B., Zhi, P., Huang, S.: Physcene: Physically interactable 3d scene synthesis for embodied ai. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.29, + 0.787, + 0.331 + ], + "angle": 0, + "content": "96. Yang, Y.Q., Guo, Y.X., Xiong, J.Y., Liu, Y., Pan, H., Wang, P.S., Tong, X., Guo, B.: Swin3d: A pretrained transformer backbone for 3d indoor scene understanding. arXiv preprint arXiv:2304.06906 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.333, + 0.786, + 0.374 + ], + "angle": 0, + "content": "97. Yang, Z., Zhang, S., Wang, L., Luo, J.: Sat: 2d semantics assisted training for 3d visual grounding. In: Proceedings of International Conference on Computer Vision (ICCV) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.376, + 0.786, + 0.416 + ], + "angle": 0, + "content": "98. Yeshwanth, C., Liu, Y.C., Nießner, M., Dai, A.: Scannet++: A high-fidelity dataset of 3d indoor scenes. In: Proceedings of International Conference on Computer Vision (ICCV) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.418, + 0.787, + 0.472 + ], + "angle": 0, + "content": "99. Yuan, Z., Yan, X., Liao, Y., Guo, Y., Li, G., Cui, S., Li, Z.: X-trans2cap: Cross-modal knowledge transfer using transformer for 3d dense captioning. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.474, + 0.787, + 0.529 + ], + "angle": 0, + "content": "100. Yuan, Z., Yan, X., Liao, Y., Zhang, R., Wang, S., Li, Z., Cui, S.: Instancerefer: Cooperative holistic understanding for visual grounding on point clouds through instance multi-level contextual referring. In: Proceedings of International Conference on Computer Vision (ICCV) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.531, + 0.787, + 0.585 + ], + "angle": 0, + "content": "101. Zhang, H., Zhang, P., Hu, X., Chen, Y.C., Li, L., Dai, X., Wang, L., Yuan, L., Hwang, J.N., Gao, J.: Glipv2: Unifying localization and vision-language understanding. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.221, + 0.587, + 0.787, + 0.628 + ], + "angle": 0, + "content": "102. Zhang, R., Guo, Z., Zhang, W., Li, K., Miao, X., Cui, B., Qiao, Y., Gao, P., Li, H.: Pointclip: Point cloud understanding by clip. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.63, + 0.787, + 0.671 + ], + "angle": 0, + "content": "103. Zhang, R., Wang, L., Qiao, Y., Gao, P., Li, H.: Learning 3d representations from 2d pre-trained models via image-to-point masked autoencoders. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.672, + 0.787, + 0.713 + ], + "angle": 0, + "content": "104. Zhang, Y., Gong, Z., Chang, A.X.: Multi3drefer: Grounding text description to multiple 3d objects. In: Proceedings of International Conference on Computer Vision (ICCV) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.715, + 0.787, + 0.755 + ], + "angle": 0, + "content": "105. Zhao, L., Cai, D., Sheng, L., Xu, D.: 3dvg-transformer: Relation modeling for visual grounding on point clouds. In: Proceedings of International Conference on Computer Vision (ICCV) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.757, + 0.787, + 0.798 + ], + "angle": 0, + "content": "106. Zheng, J., Zhang, J., Li, J., Tang, R., Gao, S., Zhou, Z.: Structured3d: A large photo-realistic dataset for structured 3d modeling. In: Proceedings of European Conference on Computer Vision (ECCV) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.8, + 0.787, + 0.84 + ], + "angle": 0, + "content": "107. Zhu, W., Hessel, J., Awadalla, A., Gadre, S.Y., Dodge, J., Fang, A., Yu, Y., Schmidt, L., Wang, W.Y., Choi, Y.: Multimodal c4: An open, billion-scale corpus of images interleaved with text. arXiv preprint arXiv:2304.06939 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "22" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.441, + 0.127 + ], + "angle": 0, + "content": "B. Jia and Y. Chen et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "108. Zhu, Y., Gao, T., Fan, L., Huang, S., Edmonds, M., Liu, H., Gao, F., Zhang, C., Qi, S., Wu, Y.N., et al.: Dark, beyond deep: A paradigm shift to cognitive ai with humanlike common sense. Engineering 6(3), 310-345 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.231 + ], + "angle": 0, + "content": "109. Zhu, Z., Ma, X., Chen, Y., Deng, Z., Huang, S., Li, Q.: 3d-vista: Pre-trained transformer for 3d vision and text alignment. In: Proceedings of International Conference on Computer Vision (ICCV) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.232, + 0.787, + 0.273 + ], + "angle": 0, + "content": "110. Zhu, Z., Zhang, Z., Ma, X., Niu, X., Chen, Y., Jia, B., Deng, Z., Huang, S., Li, Q.: Unifying 3d vision-language understanding via promptable queries. In: Proceedings of European Conference on Computer Vision (ECCV) (2024)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.273 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/SceneVerse_ Scaling 3D Vision-Language Learning for Grounded Scene Understanding/72c8327d-dc6a-436c-89b3-89891fd18764_origin.pdf b/2024/SceneVerse_ Scaling 3D Vision-Language Learning for Grounded Scene Understanding/72c8327d-dc6a-436c-89b3-89891fd18764_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f9df2a9a31e88fe78aac79c51b3f87d78b49a65f --- /dev/null +++ b/2024/SceneVerse_ Scaling 3D Vision-Language Learning for Grounded Scene Understanding/72c8327d-dc6a-436c-89b3-89891fd18764_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4747185c52fce827db2e1788a2e8e911d44f7a26163d4990bf47b26241b92a33 +size 12163065 diff --git a/2024/SceneVerse_ Scaling 3D Vision-Language Learning for Grounded Scene Understanding/full.md b/2024/SceneVerse_ Scaling 3D Vision-Language Learning for Grounded Scene Understanding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7acfe667b84f1ca5c7af8ebc4a8485bcb875411f --- /dev/null +++ b/2024/SceneVerse_ Scaling 3D Vision-Language Learning for Grounded Scene Understanding/full.md @@ -0,0 +1,439 @@ +# SCENEVERSE: Scaling 3D Vision-Language Learning for Grounded Scene Understanding + +Baoxiong Jia\*, Yixin Chen\*, Huangyue Yu, Yan Wang, Xuesong Niu, Tengyu Liu, Qing Li, and Siyuan Huang + +* indicates equal contribution +State Key Laboratory of General Artificial Intelligence, BIGAI https://scene-verse.github.io + +Abstract. 3D vision-language (3D-VL) grounding, which aims to align language with 3D physical environments, stands as a cornerstone in developing embodied agents. In comparison to recent advancements in the 2D domain, grounding language in 3D scenes faces two significant challenges: (i) the scarcity of paired 3D-VL data to support grounded learning of 3D scenes, especially considering complexities within diverse object configurations, rich attributes, and intricate relationships; and (ii) the absence of a unified learning framework to distill knowledge from grounded 3D data. In this work, we aim to address these major challenges in 3D-VL by examining the potential of systematically upscaling 3D-VL learning in indoor scenes. We introduce the first million-scale 3D-VL dataset, ScENEVERSE, encompassing 68K indoor scenes and comprising 2.5M vision-language pairs collected from both human annotations and our scalable scene-graph-based generation approach. We demonstrate that this scaling allows for a unified pre-training framework, Grounded Pretraining for Scenes (GPS), for 3D-VL learning. Through extensive experiments, we showcase the effectiveness of GPS by achieving state-of-the-art performance on existing 3D visual grounding and question-answering benchmarks. We also show that the data scaling effect is not limited to GPS, but is generally beneficial for models on tasks like 3D semantic segmentation. The vast potential of ScENEVERSE and GPS is unveiled through zero-shot transfer experiments in challenging 3D-VL tasks. + +Keywords: 3D Vision-Language $\cdot$ Data Scaling $\cdot$ Grounded Scene Understanding + +# 1 Introduction + +The foundation of human cognitive development lies in the grounding of language within the physical world [53,81,108]. Recent progress in Large Language Models (LLMs) [10,11,83] has markedly promoted the alignment between vision and language [3,59,75] utilizing billion-scale vision-language datasets [79,107]. However, with these advancements predominantly focusing on the 2D domain, the grounded understanding of 3D physical environments remains in an incipient stage [1,5,16]. Recognizing the pivotal role of grounded 3D experiences in + +![](images/17e5b5f02f00295aaa4e2093431eabf3f15e1646a4650c7a71a4a3b69e233eae.jpg) +Fig. 1: Overview of ScENEVERSE. A million-scale 3D vision-language dataset that comprises over 68K various 3D indoor scenes and 2.5M aligned scene-language pairs in the form of scene caption, object caption, and object referral. + +shaping human cognition [7,8], there is a compelling need to focus on exploring vision-language learning in the context of 3D scenes. + +Seeking insights from success in 2D vision-language (2D-VL) learning, a major factor to the success was the notable scale-up of paired vision-language data [15, 52, 79]. However, applying this experience directly from 2D to 3D is fraught with challenges. Primarily, 3D data collection heavily relies on the scanning device, making it inherently much more complex and expensive than gathering 2D images. Despite steady efforts to increase the volume of 3D scene data [9, 24, 66, 98], most datasets remain limited to thousands of scenes, substantially lagging behind the scale of existing 2D datasets. This gap is further widened by the inherent complexities of 3D scenes, which feature a multitude of object instances with diverse attributes, varying arrangements, and intricate inter-object relationships. These unique aspects of 3D scenes not only make the accurate description of objects and their relations more challenging but also considerably increase the number of language descriptions required for thorough scene depiction. Consequently, this presents a significant challenge in gathering sufficient and high-quality paired scene-language data for grounded scene understanding. + +To confront these challenges, we propose SCENEVERSE, the first million-scale dataset aimed at advancing 3D vision-language (3D-VL) learning for grounded scene understanding. At the scene level, we unify 3D scene data from existing datasets [9,24,66,76,85], aligning scenes and annotations from various capturing sources, and supplement the collection with synthetic scenes [28,106]. This compilation represents the most extensive 3D scene data gathered to date, amounting to 68K scenes. For language, we first present 97K newly-annotated referring expressions, the most extensive thus far. We additionally propose an + +automated generation pipeline utilizing 3D scene graphs [4,86] and LLMs to create comprehensive, high-quality scene-language pairs. This refined collection, totaling 2.5M scene-language pairs, provides detailed and comprehensive descriptions of both object-level and scene-level descriptions within the 3D scene. + +We thoroughly investigate the potential offered by SCENEVERSE with large-scale pre-training, introducing Grounded Pre-training for Scenes (GPS), a novel and unified pre-training framework designed for scene and object-level alignment without auxiliary losses. Through multi-level contrastive alignment, we achieve significant performance boosts on 3D-VL tasks, such as grounding and question answering, setting new state-of-the-art results via a simple and effective pretraining process. We unveil the vast possibilities offered by SCENEVERSE and GPS in 3D-VL tasks in a zero-shot transfer setting. Additionally, we show that the scaling effect in SCENEVERSE is not limited to GPS, but generally benefits models in tasks like 3D semantic segmentation. At last, we offer deeper insights into the data-scaling in SCENEVERSE through extensive ablative experiments, pointing out future directions. Our main contributions are as follows: + +1. We introduce SCENEVERSE, the first million-scale 3D-VL dataset for grounded scene understanding. SCENEVERSE encompasses 68K 3D scenes coupled with 2.5M scene-language pairs, sourced through a combination of human annotation and automated generation methods. This represents a significant improvement in terms of data diversity and scale compared to prior datasets. +2. We propose GPS, a transformer-based model trained with multi-level scenetest alignment that achieves state-of-the-art results on existing 3D-VL grounding and question-answering benchmarks by pre-training on ScENEVERSE. +3. We demonstrate that with the data scale-up and model design, our pre-trained GPS exhibit emerging zero-shot generalization capabilities in grounded scene understanding. We also show that this scaling effect is not limited to GPS, but is generally beneficial for models on tasks like semantic segmentation. + +# 2 Related Work + +Datasets for Grounded 3D Understanding Obtaining aligned 3D-language data is a inherently difficult task. In 3D object modeling, pioneering works like ShapeNet [14] sourced 3D assets from online repositories, leading to a proliferation of high-quality 3D object datasets [23, 69, 91]. Notably, recent developments include internet-scale data collection with Objaverse [26, 27], accompanied by the integration of object-level captions [93] for 3D-language alignment. Models trained on these datasets demonstrate an enhanced understanding of objects, evident in classification [60], generation [61], and captioning tasks [63]. + +In contrast, developing datasets for grounded 3D scene understanding is even more challenging due to the extensive requirements for scene acquisition and annotation. Existing works curate RGB-D and scanned indoor scene datasets [9,13, 24,66,76,85] and synthetic scenes [28,50,95,106] used for benchmarking tasks like 3D object detection and segmentation [31,46,67,80,84]. These semantically labeled scenes are subsequently used in fine-grained scene grounding tasks like object + +Table 1: Comparison of SCENEVERSE with existing 3DVL Datasets. SCENEVERSE expands the data scale of prior work by order of magnitude. "VG" stands for Visual Grounding, "QA" for Question Answering, "PT" for Pre-training and "MT" for Multi-tasking. "Anno." denotes language from human annotations and "Syn." for template-based or LLM generated descriptions. + +
Dataset3D Scene Obj.TaskObj. CaptionScene CaptionObj. ReferralQuality CheckNew Anno.Syn.Existing Anno.Syn.Total
ScanRefer [16]VGXX✓✓52K---52K
ReferIt3D [1]| |VGXX✓✓42K200K--242K
ScanQA [5]1.5K33KQA---✓27K---27K
SQA3D [65]QA---✓33K---33K
Multi3DRefer [104]| |VGXX✓✓-10K52K-62K
Cap3D [63]-666KVGX✓XX58K666K--724K
ScanScribe [109]3K56KPTXX✓X-90K94K94K278K
3D-LLM [40]1.5K186KMT✓✓✓X-659K--659K
EmbodiedScan [87]5K890KVGXX✓X-970K--970K
LEO [41]3K56KMT✓✓✓✓-188K235K90K513K
SCENEVERSE68K1.5MVG✓✓✓✓96K2.1M94K200K2.5M
+ +referral [1, 16, 104], captioning [17, 19, 22, 99], vision-language-navigation [39, 64, 73, 88] and reasoning [5, 40, 65]. Recent works exploit the representation of 3D scene graph (3DSG) [4, 20, 78, 86], which concisely describes scenes with hierarchical structures. This representation is notably advantageous for planning [2, 77] and captioning [35], owing to its compatibility with LLMs for flexible description generation [40, 41]. Nonetheless, as shown in Tab. 1, most datasets are constrained in both scene and language scales, underscoring the need for scaling up fine-grained and aligned scene-language data to enhance grounded scene understanding. + +Vision-Language Learning Recent years have witnessed tremendous progress in 2D-VL [3, 21, 25, 57, 59, 75], empowered by transformer-based pre-training models [11, 29, 72] and large-scale image-language datasets [15, 79]. A central theme across 2D-VL domains is the effectiveness of data scaling [49], as demonstrated by improved alignment and expanded capabilities in open-vocabulary understanding [33, 51, 54, 58] through a contrastive pre-training pipeline [75]. + +However, in grounded scene understanding, the primary challenge for models has been the limited availability of paired 3D scene-language data, which restricts the application of insights drawn from 2D-VL. Current models for 3D scene grounding [6,18,37,43,45,62,92,97,105] heavily rely on task-specific knowledge in both model and loss designs or advanced optimization strategies [109]. To bridge this gap, there has been a growing emphasis on employing pre-trained 2D-VL models for 3D-VL [36,38,74,82,93,102,103]. Yet, these models mostly draw information available from 2D-VL models (e.g., object attribute, affordance, etc.), falling short on capturing crucial 3D information like object spatial relationships which are necessary for more fine-grained tasks such as grounded human-scene [20, 44,47,48,89,90] and robot-scene interactions modeling [34,55,68,70]. This urges the need for a multi-level alignment between language and 3D scenes, particularly regarding 3D-specific information. Considering the nascent stage of existing 3D pre-training methods [30,94,109,110], we believe SCENEVERSE and GPS have the potential to spearhead new avenues in 3D-VL research. + +# 3 SCENEVERSE + +SCENEVERSE is designed for grounded scene understanding with 3D scenes curated from diverse existing datasets of both real and synthetic environments. Regarding language, we employ both human annotation and a novel automated generation pipeline to collect comprehensive and high-quality language for both object-level and scene-level descriptions. We provide details regarding data collection in the following sections. + +# 3.1 Scene Curation + +To address the scarcity of available 3D scene data, we construct SCENEVERSE by unifying 3D scene data from various existing datasets. We use real-world scene datasets, including ScanNet [24], ARKitScenes [9], HM3D [76], 3RScan [85] and MultiScan [66], alongside synthetic environments from Structured3D [106] and ProcTHOR [28]. The inclusion of these synthetic datasets is mainly motivated by their potential as scalable data sources for 3D-VL alignment. To ensure cohesion across various sources, we conduct preprocessing steps such as room segmentation, point subsampling, axis alignment, normalization, and semantic label alignment. Each scan is represented by a point cloud $\mathbf{P} \in \mathbb{R}^{N \times 8}$ , wherein each point is defined by its 3D coordinates, RGB color, instance id, and semantic label. In total, we curate 68,406 3D scenes in SCENEVERSE. + +# 3.2 Referral Annotation by Humans + +In the curated scenes of SCENEVERSE, we present the most comprehensive set of human-annotated, context-rich object referrals to date, serving as a valuable benchmark for assessing grounded scene understanding capabilities. The human annotations contain 96,863 descriptions in ARKitScenes [9], HM3D [76] and MultiScan [66]. During the annotation process, one human annotator was assigned to write at least 20 words to distinctly refer to a single 3D object within a 3D scene. Each referral text then undergoes independent verification by two additional reviewers, both mandated to accurately locate the referenced object based on the 3D scene and the annotated referral text. Any object referrals that do not pass the verification by either reviewer are flagged for re-annotation. + +# 3.3 3D Scene Graph Construction + +Our 3D scene graph is defined as a hierarchical graph $\mathcal{G} = (\mathcal{V},\mathcal{E})$ . Each node $v\in \mathcal{V}$ represents one distinct 3D object instance, parameterized by its centroid $\pmb {p}_i\in \mathbb{R}^3$ and bounding box size of $\pmb {b}_i = (b_x,b_y,b_z)\in \mathbb{R}^3$ . The edges $\mathcal{E}$ represent spatial relationships between nodes. To construct the scene graph $\mathcal{G}$ , we instantiate the nodes with the instance annotation from the point clouds and assign object classes with their corresponding semantic labels. Following prior work [1,86], we consider the Vertical proximity, Horizontal proximity and Multi-object Relationships as spatial relations. For a more detailed description of the scene graph construction and relationship determination, please refer to supplementary. + +![](images/d6fe1d32398d30be9ead3924f24ac73f57a2feb2f73b4c29ecba65b7046d0bfd.jpg) +(a) 3D Scene +Fig. 2: SCENEVERSE collection and statistics. Given a 3D scene (a), our automated pipeline (c) generates three types of description including scene caption, object caption and object referral. (b) SCENEVERSE data comparison and composition. + +一 + +Scene Cap + +Sub-graph Context + +{scene_type':Bedroom', + +'object_count': {'nightstand':2,...}, + +'relation': {nightstand', 'on', floor}, + +['backback', in front of, b + +Summary + +Prompt: Provide a summary for a scene from a given + +scene graph delimited by triple backticks, + +Response: In this bedroom, there are two nightstands, ... + +The backpack is in front of the nightstand as well . The roomapproves to be functional with the nightstands providing + +storage space and the telephone for communication + +(b) Language radar + +ar chart and + +Object Caption + +BLIP2 Captions + +1. A bed in a hotel room. (0.85) + +2. A white comforter on a bed. (0.83) + +3. A bed with a striped comforter. (0.83) + +N.A picture of cat. (0.63) + +Summary + +Prompt: Summarize the captions below. The summary + +should be a description of the (object). Focus on the + +{object's attributes, like color, shape, n + +Identity and correct the potential errors ... + +Response : The bed is in a motel room with a stripedcomforter . It has a white comforter and a blanket on it + +The bed is also in a room with a bedside table. + +(c) Automated language generation + +Object R + +Relationship Triplets + +1. {table', 'chair', 'left'} + +2.{'bed', ('lamp', 'mini fridge'), 'between'} + +Template-based Referral + +1. The table is to the left of the chair. +2. The chair is in the right of the chair. + +2. It's a bed in the m + +Rephrasing + +Prompt: Rewrite the following sentence using one random word. + +sentence structure. Focus on the + +about the {target_object} + +Response: + +1. The table is situated to the left of the armchair. + +2. The bed occupies the space between the lamp and the wall. + +mini fridge, creating a cozy atmosphere. + +# 3.4 Language Generation with LLMs + +The scene-language pairs in ScENEVERSE aim to capture varying aspects of the 3D scene, including detailed object attributes in object captioning, spatial relationships between objects in object referral, and global scene descriptions in scene captioning. Based on 3D scene graphs, we utilize both templates and LLMs to automatically generate descriptions on these three granularities. + +Object Captioning Object captions aim to provide detailed descriptions of an object's visual and physical properties, facilitating object-level grounding with its distinctive features. Given the multi-view images, we utilize the point cloud of the object $v \in V$ to identify its occurrence in the images through rendering. The images are then cropped with the rendered bounding boxes and processed through BLIP2 [56] to generate initial object captions. We select the top 10 sentences with the highest CLIP [75] similarity score and minimal occlusion and utilize an LLM to obtain a refined and coherent summary of the captions. The detailed object captioning pipeline is illustrated in supplementary. + +Object Referral Object relationship captions refer to objects by articulating their spatial relationships in the scene. Spatial relationship triplets $(v_{i}, v_{j}, e_{ij})$ are first extracted from the constructed 3D scene graph. We design various templates to generate descriptions for each relationship type, assigning the entities in the form of (target-object, spatial-relation, anchor-object(s)). This results in examples like "the chair is next to the li2023behaviorarmchair", "facing the sofa, there is a suitcase far to the right of the shoes", and "the fridge is between cabinet and sofa". Our designed templates span passive and active tenses, as well as inversion clauses, contributing to the richness of the generated text. To enhance the descriptions' naturalness, we employ LLM for sentence rephrasing. Statistics for the descriptions before and after rephrasing are presented in Fig. 2 (b). + +Scene Captioning The scene-level captions emphasize global information, portraying the key objects in the scene along with their attributes and functionalities. We use the constructed 3D scene graph and prompt LLMs to generate these captions. We random sample a subset of edges and nodes from the scene graph each time as the scene context to enhance the diversity of scene captions. The object counts are also provided as LLM prompts, together with the room type and object attributes if such annotations are available in the dataset. + +# 3.5 Data Quality and Statistics + +Data Quality The 96K human-annotated set of ScENeVErSE is collected through AMT, where 82 humans are employed for annotation and 18 for verification. All final annotations passed the reference verification, with a re-annotation rate of $4.8\%$ . For our automatic language generation pipeline, we conduct extensive prompt tuning and iterate with human feedback for LLMs on object captioning, summary, and rephrasing. To verify the efficacy of the pipeline, we conduct a quality check where 12K generated object-level descriptions are randomly selected for human verification. Results demonstrate a $96.93\%$ pass rate, surpassing that in ReferIt3D [1] with $86.1\%$ pass rate on 2K samples. + +Statistics In total, SCENEVERSE comprises a total of 68,406 room-level 3D scans, with the source composition shown in Fig. 2 (b). The dataset contains 1.5M object instances ranging in 2290 object categories. Our generated 3D scene graph comprises 21 types of relationships following prior work [1,86]. For the language descriptions, we generate 1M template-based texts and 1M sentences rephrased by Llama [83] and GPT-3.5 [71]. As can be seen from the radar chart and examples in Fig. 2, the diversity of the LLM-refined data, particularly in sentence length and variety, closely aligns with the characteristics of annotated descriptions, surpassing the template-based data. Together with existing sources (294K) and our newly annotated set (96K), SCENEVERSE contains 2.5M scene-language pairs in total. All the rephrasing and summary prompts, along with the complete set of relationships, are detailed in supplementary. + +# 4 Grounded Pre-training for Scenes + +In this section, we introduce GPS, an efficient transformer-based model trained with multi-level contrastive losses for aligning 3D scenes and texts. As shown in Fig. 3, we echo the language descriptions collected at different granularities to form contrastive objectives at both object-level, referral-object-level, and scene-level in GPS. We describe the design of each level in the following sections. + +# 4.1 Object-level Grounding + +Given a 3D scene point cloud $\mathcal{S}$ , we use an off-the-shelf 3D object segmentation model to decompose it into a bag of $N$ objects $\mathcal{S} = \{\pmb{o}_1, \pmb{o}_2, \dots, \pmb{o}_n\}_{i=1}^N$ . We extract + +![](images/65c4acf2fbe695e949ed45de515b28bdd9cad216fdb934a32bfdd220c1fa25da.jpg) +Fig. 3: Overview of GPS model. We use contrastive alignment at three levels $\mathcal{L}_{\mathrm{obj}}$ , $\mathcal{L}_{\mathrm{scene}}$ , and $\mathcal{L}_{\mathrm{ref}}$ and a masked language modeling objective $\mathcal{L}_{\mathrm{MLM}}$ for model learning. + +object features $\{\pmb{f}_i^O\}$ with an object point cloud encoder and text features $\{\pmb{f}_i^T\}$ by feeding object-captions $\{\pmb{T}_i^{\mathrm{obj}}\}$ into a frozen language model. Following [93], we perform cross-modal alignment on the object features and text features via: + +$$ +\mathcal {L} _ {\mathrm {o b j}} = - \frac {1}{2} \sum_ {(p, q)} \left(\log \frac {\exp \left(D ^ {\mathrm {o b j}} (p , q)\right)}{\sum_ {r} \exp \left(D ^ {\mathrm {o b j}} (p , r)\right)} + \log \frac {\exp \left(D ^ {\mathrm {o b j}} (p , q)\right)}{\sum_ {r} \exp \left(D ^ {\mathrm {o b j}} (r , q)\right)}\right), \tag {1} +$$ + +where $D^{\mathrm{obj}}(p,q) = (\pmb{f}_p^O\pmb{f}_q^T /\tau)$ denotes the dot product between object and text features and $(p,q)$ denotes a pair of aligned object-text pair in the training batch and $r$ iterates over all object-text pairs in the training batch. Similar to CLIP [75], we use a learnable temperature parameter $\tau$ to facilitate model learning. + +# 4.2 Scene-level Grounding + +With aligned object features, we encode the scene by incorporating object spatial locations into the extracted object features. Specifically, we use a spatial transformer model to encode extracted object features $\{\pmb{f}_i^O\}$ with their spatial location features $\{l_i\}$ following [18, 109]: + +$$ +\boldsymbol {f} ^ {S} = \operatorname {S p a t i a l A t t n} (\{\boldsymbol {f} _ {i} ^ {O} \}, \{\boldsymbol {l} _ {i} \}) +$$ + +where $\{\pmb{f}_i^S\}$ denotes the feature of object $\pmb{o}_i$ after encoding with spatial location features. To perform scene-level alignment, we operate on these scene-level object features $\{\pmb{f}_i^S\}$ and align it with the scene caption $\pmb{T}^{\mathrm{scene}}$ . Specifically, we feed the object features into a projection layer and use max-pooling over all object features to obtain the scene feature $\pmb{g}^S$ . Similar to object-level grounding, we pass the scene caption through a tunable language model to obtain text feature $\pmb{g}^T$ and perform scene-level contrastive alignment through: + +$$ +\mathcal {L} _ {\text {s c e n e}} = - \frac {1}{2} \sum_ {(p, q)} \left(\log \frac {\exp \left(D ^ {\text {s c e n e}} (p , q)\right)}{\sum_ {r} \exp \left(D ^ {\text {s c e n e}} (p , r)\right)} + \log \frac {\exp \left(D ^ {\text {s c e n e}} (p , q)\right)}{\sum_ {r} \exp \left(D ^ {\text {s c e n e}} (r , q)\right)}\right), \tag {2} +$$ + +where $D^{\mathrm{scene}}(p,q) = (\pmb{g}_p^S\pmb{g}_q^T / \tau)$ denotes the dot product between scene feature $\pmb{g}_p^S$ and scene caption feature $\pmb{g}_q^T$ for each pair of aligned scene-text pairs in the training batch and $r$ iterates over all scene-text pairs in the training batch. + +# 4.3 Referral-object-level Grounding + +To model the relationships revealed in referring expressions, we employ a self-attention-based reasoning transformer for grounding object referrals in scenes. This transformer takes in scene-object features $\{\pmb{f}_i^S\}$ and an object referral $T^{\mathrm{ref}}$ and performs self-attention to learn relationships between text descriptions and object relationships. We use the same tunable language encoder as in scene-level grounding for extracting per-object referral features. We pass this text feature together with scene-object features into the self-attention transformer to obtain the aligned object features $\pmb{h}_i^S$ and the sentence-level referral feature $\pmb{h}^T$ . We then perform the referral-object-level contrastive alignment following: + +$$ +\mathcal {L} _ {\mathrm {r e f}} = - \log \frac {\exp \left(\bar {\boldsymbol {h}} ^ {S} \boldsymbol {h} ^ {T} / \tau\right)}{\sum_ {p} \exp \left(\boldsymbol {h} _ {p} ^ {S} \boldsymbol {h} ^ {T} / \tau\right)}, \tag {3} +$$ + +where $\bar{h}^S$ denotes the feature of the referred object, $p$ iterates over all objects within the same scene. Notably, in contrast to inter-scene contrast that was done in object- and scene-level alignment, we force the selection of positive pairs to be within the same scene to provide intra-scene contrast for fine-grained object grounding. This mimics the success of intra-image and inter-image contrasts commonly used for region-word alignment in 2D-VL models [101]. + +To learn the multi-level alignment between 3D scenes and language, we first train the point cloud encoder with an object-level grounding objective to obtain a good feature initialization for grounding objects in scenes. During the scene grounding stage, we train our inter- and intra-scene objectives together with a masked language modeling loss $\mathcal{L}_{\mathrm{MLM}}$ over the inputted object-referral texts to tune the parameters within the language encoder and self-attention transformer. Above all, the learning of GPS could be summarized as optimizing: + +$$ +\mathcal {L} = \mathcal {L} _ {\mathrm {o b j}} + \mathcal {L} _ {\mathrm {s c e n e}} + \mathcal {L} _ {\mathrm {r e f}} + \mathcal {L} _ {\mathrm {M L M}}. +$$ + +# 5 Experiments + +In this section, we present experimental results addressing the following questions: + +1. How effective is the data scaling in ScENEVERSE for 3D visual grounding? Does the scale-up benefit common 3D-VL tasks (e.g., 3D question answering, open-vocabulary 3D semantic segmentation) and pre-training-based models? +2. How well is the GPS pre-training pipeline for 3D-VL tasks? Does it exhibit similar properties of 2D-VL models in 3D-VL tasks? +3. What is offered by ScENEVERSE and GPS and what is missing? + +In the following sections, we describe in detail the model performance regarding these key topics. Due to the page limit, we direct readers to the supplementary for implementation details, qualitative results, and more experimental analyses. + +Table 2: 3D visual grounding results on Nr3D, Sr3D, and ScanRefer. We use "pre-train" for our model trained on ScENeVERse w/o additional fine-tuning, and "fine-tune" for its data-specific fine-tuned version. Best results are highlighted in bold. + +
MethodNr3DSr3DScanRefer Acc@0.5
OverallEasyHardV-Dep.V-Indep.OverallEasyHardV-Dep.V-Indep.OverallUniqueMultiple
3DVG-Trans [105]40.848.534.834.843.751.454.244.944.651.734.760.628.4
TGNN [42]37.344.230.635.838.045.048.536.945.845.029.756.823.2
TransRefer3D [37]48.056.739.642.550.757.460.550.249.957.7---
InstanceRefer [100]38.846.031.834.541.948.051.140.545.848.132.966.824.7
FFL-3DOG [32]41.748.235.037.144.7-----34.067.925.7
LAR [6]48.958.442.347.452.159.463.051.250.059.1---
SAT [97]56.564.948.454.457.657.961.250.049.258.330.150.825.2
3D-SPS [62]51.558.145.148.053.262.656.265.449.263.237.066.729.8
3DJCG [12]----------37.364.330.8
BUTD-DETR [45]54.660.748.446.058.067.068.663.253.067.639.866.335.1
MVT [43]59.567.452.759.160.364.566.958.858.464.733.366.525.3
ViL3DRel [18]64.470.257.462.064.572.874.967.963.873.237.768.630.7
EDA [92]52.158.246.150.253.168.170.362.954.168.742.368.637.6
3D-VisTA (scratch) [109]57.565.949.453.759.469.672.163.657.970.141.570.934.8
3D-VisTA [109]64.272.156.761.565.176.478.871.358.977.345.875.139.1
Ours (scratch)58.767.050.955.859.868.470.563.453.169.040.471.334.7
Ours (pre-train)55.262.848.045.558.874.176.468.554.175.047.177.441.6
Ours (fine-tuned)64.972.557.856.967.977.580.171.662.878.248.177.942.7
+ +# 5.1 3D Visual Grounding + +Settings We evaluate our model on three commonly-used datasets for 3D visual grounding: ScanRefer [16], Nr3D, and Sr3D [1]. For Nr3D and Sr3D, we follow Achlioptas et al. [1] and report the grounding accuracies of models using ground-truth object masks. For ScanRefer, we follow Zhu et al. [109] and use Mask3D [80] to generate object proposals. Results are reported as Acc@0.5 to evaluate the correctness of predictions whose object bounding boxes overlap the ground truth with IoU $>0.5$ . For comparisons, we compare with existing baselines by providing the results of pre-trained GPS and dataset-specific fine-tuned GPS. Please see more details in the supplementary. + +Results and Analyses As shown in Tab. 2, GPS trained on SCENEVERSE achieves state-of-the-art results on all existing 3D-VL grounding benchmarks. Initially, when GPS is trained directly on the training sets of benchmark datasets, labeled as Ours (scratch), it underperforms compared to existing models that employ more complex structures or loss designs. This result underscores the data-intensive nature of the contrastive alignment paradigm. However, when presented with extensive training data in SCENEVERSE, the results of our model without additional fine-tuning, i.e., Ours (pre-train), significantly improves and already achieves state-of-the-art results on benchmarks like ScanRefer. Moreover, the dataset-specific fine-tuned model, i.e., Ours (fine-tuned), consistently outperforms existing baselines with only a simple projection MLP added on top of the pretrained model, jointly optimized during fine-tuning without any other auxiliary architecture or loss objective. These results underscore the strong potential of both the SCENEVERSE and GPS for 3D-VL tasks. + +Table 3: Zero-shot transfer on existing benchmarks. "SR" stands for ScanRefer. + +
MethodNr3DSr3DSR@0.25SR@0.5
3D-VisTA (scratch)57.569.645.941.5
3D-VisTA (zero-shot)35.231.233.229.6
3D-VisTA (zero-shot text)43.136.141.136.4
Ours (scratch)58.768.444.540.4
Ours (zero-shot)32.433.335.231.1
Ours (zero-shot text)41.938.140.735.8
+ +Table 4: Zero-shot transfer on ScENeVERSE-val. Evaluation uses GT object proposals following Nr3D/Sr3D. + +
MethodOverallEasyHardV-Dep.V-Indep.
3D-VisTA (scratch)40.753.121.637.344.3
3D-VisTA (zero-shot)52.959.635.453.752.2
3D-VisTA (zero-shot text)58.170.039.652.564.1
Ours (scratch)38.550.220.833.743.9
Ours (zero-shot)59.269.444.053.166.3
Ours (zero-shot text)60.670.945.154.867.3
+ +# 5.2 Zero-Shot Transfer + +Settings To better evaluate the effectiveness of both the SCENEVERSE data and the GPS model, we further perform zero-shot transfer experiments to test the models' capability in 4 benchmarks, ScanRefer, Sr3D, Nr3D, and SCENEVERSE- val. We create SCENEVERSE-val using $8.5K$ annotated object referrals of 271 scenes in MultiScan, and randomly split the scenes following a 4:1 train / test split for creating the held-out test set. We mainly consider 2 specific transfer settings in our experiments: (i) zero-shot: models trained by removing all the scenes from the target dataset, tested on held-out unseen scenes, and (ii) zero-shot text: Models trained on data that include the training set of scenes from the target dataset, yet tested exclusively with unseen scene-text distribution. Specifically, for the zero-shot text setting, we use the generated texts in SCENEVERSE as fine-tuning sources for the zero-shot model. We mainly compare our model against a recent pre-training-based model 3D-VisTA. See more details on experimental setting and implementation in the supplementary. + +Results and Analyses We present the results of zero-shot transfer experiments in Tab. 3 and Tab. 4 with the following key observations: + +- Our GPS model demonstrates superior generalization to unseen scenes compared to the 3D-VisTA model. In zero-shot transfer scenarios, our model consistently outperforms 3D-VisTA across established benchmarks and ScEN-EVERSE-val. This indicates the effectiveness of contrastive alignment over traditional classification objectives, aligning with the advancements seen in 2D-VL models for open-vocabulary grounding and transfer capabilities +- SCENEVERSE dataset substantially enhances 3D-VL grounding capabilities through zero-shot transfer, especially when provided with relatively limited training data, i.e., SCENEVERSE-val. As demonstrated in Tab. 4, there is a significantly improved performance when comparing models trained on SCENEVERSE in a zero-shot manner to those trained from scratch. This indicates that SCENEVERSE can effectively capture knowledge for general 3D scene grounding. Consequently, this underscores its potential as a go-to pre-training dataset for 3D-VL tasks. +- The impact of our extensive collection and scalable generation of scene-text pairs is further evidenced by the results in the zero-shot text setting. Notably, as shown in Tab. 3, these automatically generated scene-text pairs supply ample knowledge for comprehending the scene distribution. This contributes significantly to the substantial improvement over the zero-shot performance. + +Table 5: 3D question answering re- Table 6: Existing 3D backbones pre-sults on ScanQA and SQA3D. We re-trained on SCENEVERSE for openport EM@1 score on ScanQA and SQA3D vocabulary 3D semantic segmentation on evaluation sets. ScanNet. "SPUNet" denotes SparseUNet pro + +
ModelScanQASQA3D
valw/objw/o obj
ScanRefer+MCAN [5]18.620.619.0-
ScanQA [5]20.323.520.946.6
SQA3D [65]---47.2
3D-VisTA [109]22.427.023.048.5
3D-LLM [40]20.519.1--
Ours22.725.023.549.9
+ +
ModelNetworkmIoUΔmAccΔ
OpenScene [74]SPUNet1657.2-69.9-
PLA [30]SPUNet1617.7-33.5-
RegionPLC [94]SPUNet1656.9-75.6-
RegionPLC+SCENEVERSESPUNet1658.2+1.7%77.3+2.2%
OpenScene [74]SPUNet3257.8-70.3-
PLA [30]SPUNet3219.1-41.5-
RegionPLC [94]SPUNet3259.6-77.5-
RegionPLC+SCENEVERSESPUNet3261.0+2.3%79.7+2.8%
+ +# 5.3 Additional 3D-VL Tasks + +Settings We evaluate the effectiveness of GPS and ScENEVERSE on additional 3D-VL tasks: (i) 3D question answering (3D-QA) on ScanQA [5] and SQA3D [65], and (ii) open-vocabulary 3D semantic segmentation (OV-Seg) on ScanNet. + +- In the 3D-QA task, we follow Zhu et al. [109] and evaluate models over the exact match metric (EM@1) on the validation and test sets of ScanQA, as well as the test set of SQA3D. We pre-train GPS on SCENEVERSE and fine-tune the model on the 3D-QA dataset to compare with state-of-the-art models. +- In the OV-Seg task, as GPS builds upon an object-centric design and thus is not directly applicable to semantic segmentation, we consider testing the effectiveness of ScENEVERSE on improving existing 3D models. Specifically, we follow the open-vocabulary semantic segmentation settings proposed by Yang et al. [94] and report the mIoU and mAcc score. We compare with existing works by pre-training the RegionPLC [94] model on ScENEVERSE. + +Results and Analyses We present the results of 3D-QA experiments in Tab. 5 and the results of OV-Seg experiments in Tab. 6. The analyses are as follows: + +- As shown in Tab. 5, our model achieves state-of-the-art results on both benchmarks, outperforming recent strong pre-training-based baselines like 3D-VisTA and 3D-LLM. As SCENEVERSE currently contains only descriptions of objects and scenes, we believe involving more types of language descriptions (e.g., question-answer pairs, dialogues) is a promising direction for further improving model performance on these downstream tasks. + +- As shown in Tab. 6, we observe consistent performance improvement of existing 3D backbone models on this task when pre-trained with SCENEVERSE data. This result validates that the collected data in SCENEVERSE can effectively boost the performance of existing models on scene understanding tasks. We further provide results of state-of-the-art 3D models that are pretrained on SCENEVERSE on the close-vocabulary 3D semantic segmentation task in the supplementary. + +# 5.4 Ablative Studies and Discussion + +In this section, we present further discussions on both the data collected in ScENEVERSE and the GPS model design. We aim to elucidate the effects of + +![](images/5e2ab110ec62d74690b13d152711fe2ec4c7bf3cd4386d0ab8a3b7417fa5deef.jpg) +Fig.4: Model performance v.s. data scale. Plots show that models consistently improve in both the pre-train and zero-shot transfer settings on ScanRefer and ScENEV-ERSE-val with data scaling-up. + +![](images/1e80d2da8c1b18b851b85d99876ecdb44af4e1bc358562e8a8c10d2567e0a5bb.jpg) + +Table 8: Cross domain transfer results of models pre-trained on real and synthetic datasets. "S3D" stands for Structured3D. + +
RealSyntheticSCENEVERSE-valS3DProcTHOR
AllX64.837.143.4
XS3D7.085.116.1
XProcTHOR4.216.391.0
+ +Table 7: Ablation on text data source used in model pre-training. All models are tested on ScanRefer with no additional finetuning. + +
TemplateLLMAnno.Acc@0.25Acc@0.5
XXX43.538.4
✓XX50.946.1
✓✓X51.146.3
✓✓✓52.047.1
+ +Table 9: Ablation on model design on ScENeVERSE-val. We use "Obj- lvl", "Scene-lvl" to denote object and scene alignment loss, and "MLM" for the mask language modeling loss. + +
Obj-lvl.MLMScene-lvl.OverallEasyHard
XXX64.875.448.7
✓XX65.277.147.4
✓✓X62.473.445.8
✓✓✓66.977.850.3
+ +data scaling and show more clearly its effectiveness in 3D scene understanding. Regarding the experimental settings and more results discussion, refer to the supplementary. The following points are specifically discussed in this section: + +How important is data-scaling? We conduct ablation studies over the amount of data used while pre-training GPS. We consider the model trained with $\frac{1}{8}$ , $\frac{1}{4}$ , $\frac{1}{2}$ of ScENEVERSE to show the effectiveness of data-scaling on model performance in the pre-train and zero-shot transfer settings in ScanRefer and ScENEVERSE-val. As shown in Fig. 4, we observe consistent performance improvement over the increase of data scale for both settings. We provide additional experiments in the supplementary to show that such scaling effect is not only beneficial for 3D-VL grounding but also for other 3D tasks like semantic segmentation [80,96]. + +How is the generated data compared with human-annotated data? We assess the performance of models trained using various scene-text sources, specifically focusing on their performance in the ScanRefer dataset without additional fine-tuning. As shown in Tab. 7, models trained with our template-based generated texts and Large Language Model (LLM)-refined texts show significant improvements over models trained solely on ScanRefer. More importantly, these variants of our model already achieve state-of-the-art results compared with previous baselines. This indicates the effectiveness of our text-generation pipeline. Finally, we observe that adding human-annotated data is still beneficial for model performance. However, the improvement is relatively marginal over models trained on our generated data. + +What is the role of the synthetic scenes in this scale-up process? With synthetic data providing large-scale and diverse scene data for 3D-VL tasks, + +we evaluate the models' domain transfer (Sim2Real) capability. Specifically, we compare models trained on all real scenes in SCENEVERSE against models trained exclusively on two synthetic subsets of SCENEVERSE, i.e., Structured3D and ProcTHOR. As shown in Tab. 8, models trained on synthetic subsets demonstrate remarkable performance on their corresponding test sets while suffering when transferred to real or other synthetic scenes. In contrast, the model trained on real scene-text pairs exhibits less performance drop when generalizing to synthetic scenes. This result affirms the domain gap between real and synthetic scenes in 3D-VL grounding and shows that a simple scale-up in the number of scenes is insufficient when naturalness can not be guaranteed. Considering the scalability of our language generation pipeline and the scaling effect shown in our experiments, the rate-determining step for further scaling-up 3D-VL comes to the collection of diverse, high-quality, and realistic scenes that capture natural 3D scene distributions. + +How important is the design of each module in GPS? We provide ablative analyses of our multi-level contrastive alignment design in Tab. 9. We mainly consider removing objectives in our model to reveal the effectiveness of each level of alignment. We choose the referral-object-level alignment objective as the default setting and consider removing: (i) object-level alignment objective, (ii) masked language modeling objective, and (iii) scene-level alignment objective. When removing the object-level alignment objective, we learn the object point cloud encoder with the referral-object-level alignment and without pre-training. As shown in Tab. 9, we test different models on the SCENEVERSE-val without additional fine-tuning. Results show that the scene-level alignment objective is crucial for referral object grounding in SCENEVERSE-val with the $\sim 5\%$ performance drop. Similar observations could be made for the model trained without object-level alignment ( $\sim 2\%$ drop) and masked language modeling objective ( $\sim 1.5\%$ drop). These results affirm the effectiveness of our model design. + +# 6 Conclusion + +In this work, we scale up 3D-VL for grounded scene understanding. We present SCENEVERSE, a million-scale 3D-VL dataset covering various scenes and multi-level scene descriptions sourced from both human annotation and our proposed scene-text generation approach. Utilizing SCENEVERSE, we propose Grounded Pre-training for Scenes (GPS), a model trained with multi-level scene-language contrastive alignment. Through extensive experiments, we show that GPS achieves state-of-the-art results on common 3D-VL tasks including grounding and question answering. We further conduct zero-shot transfer experiments to show the improved generalization performances of GPS trained on SCENEVERSE compared with previous baselines. We also demonstrate that the scaling effect of SCENEVERSE is generally beneficial for existing 3D models on 3D-VL tasks like semantic segmentation. We hope our efforts and successful scale-up attempts in SCENEVERSE could pave the way for new research paradigms in 3D-VL. + +Acknowledgement The authors would like to thank Yaowei Zhang (BIGAI) for his help on online visualization and other colleagues from BIGAI General Vision Lab for fruitful discussions. The authors would also like to thank the anonymous reviewers for their constructive feedback. + +# References + +1. Achlioptas, P., Abdelreheem, A., Xia, F., Elhoseiny, M., Guibas, L.: Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In: Proceedings of European Conference on Computer Vision (ECCV) (2020) +2. Agia, C., Jatavallabhula, K.M., Khodeir, M., Miksik, O., Vineet, V., Mukadam, M., Paull, L., Shkurti, F.: Taskography: Evaluating robot task planning over large 3d scene graphs. In: Proceedings of Conference on Robot Learning (CoRL) (2022) +3. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2022) +4. Armeni, I., He, Z.Y., Gwak, J., Zamir, A.R., Fischer, M., Malik, J., Savarese, S.: 3d scene graph: A structure for unified semantics, 3d space, and camera. In: Proceedings of International Conference on Computer Vision (ICCV) (2019) +5. Azuma, D., Miyanishi, T., Kurita, S., Kawanabe, M.: Scanqa: 3d question answering for spatial scene understanding. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2022) +6. Bakr, E., Alsaedy, Y., Elhoseiny, M.: Look around and refer: 2d synthetic semantics knowledge distillation for 3d visual grounding. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2022) +7. Barsalou, L.W.: Perceptual symbol systems. Behavioral and brain sciences 22(4), 577-660 (1999) +8. Barsalou, L.W.: Grounded cognition. Annu. Rev. Psychol. 59, 617-645 (2008) +9. Baruch, G., Chen, Z., Dehghan, A., Dimry, T., Feigin, Y., Fu, P., Gebauer, T., Joffe, B., Kurz, D., Schwartz, A., et al.: Arkitscenes: A diverse real-world dataset for 3d indoor scene understanding using mobile rgb-d data. In: Proceedings of Advances in Neural Information Processing Systems Datasets and Benchmarks (NeurIPS Datasets and Benchmarks Track) (2021) +10. Bommasani, R., Hudson, D.A., Adeli, E., Altman, R., Arora, S., von Arx, S., Bernstein, M.S., Bohg, J., Bosselut, A., Brunskill, E., et al.: On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258 (2021) +11. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2020) +12. Cai, D., Zhao, L., Zhang, J., Sheng, L., Xu, D.: 3djcg: A unified framework for joint dense captioning and visual grounding on 3d point clouds. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2022) +13. Chang, A., Dai, A., Funkhouser, T., Halber, M., Niessner, M., Savva, M., Song, S., Zeng, A., Zhang, Y.: Matterport3d: Learning from rgb-d data in indoor environments. Proceedings of International Conference on 3D Vision (3DV) (2017) +14. Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., et al.: Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012 (2015) + +15. Changpinyo, S., Sharma, P., Ding, N., Soricut, R.: Conceptual 12m: Pushing web-scale image-text pre-training to recognize long-tail visual concepts. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2021) +16. Chen, D.Z., Chang, A.X., Nießner, M.: Scanrefer: 3d object localization in rgb-d scans using natural language. In: Proceedings of European Conference on Computer Vision (ECCV) (2020) +17. Chen, D.Z., Wu, Q., Nießner, M., Chang, A.X.: D3net: a speaker-listener architecture for semi-supervised dense captioning and visual grounding in rgb-d scans. In: Proceedings of European Conference on Computer Vision (ECCV) (2022) +18. Chen, S., Guhur, P.L., Tapaswi, M., Schmid, C., Laptev, I.: Language conditioned spatial relation reasoning for 3d object grounding. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2022) +19. Chen, S., Zhu, H., Chen, X., Lei, Y., Yu, G., Chen, T.: End-to-end 3d dense captioning with vote2cap-detr. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2023) +20. Chen, Y., Huang, S., Yuan, T., Qi, S., Zhu, Y., Zhu, S.C.: Holistic++ scene understanding: Single-view 3d holistic scene parsing and human pose estimation with human-object interaction and physical commonsense. In: Proceedings of International Conference on Computer Vision (ICCV) (2019) +21. Chen, Y., Li, Q., Kong, D., Kei, Y.L., Zhu, S.C., Gao, T., Zhu, Y., Huang, S.: Yourfit: Embodied reference understanding with language and gesture. In: Proceedings of International Conference on Computer Vision (ICCV) (2021) +22. Chen, Z., Gholami, A., Nießner, M., Chang, A.X.: Scan2cap: Context-aware dense captioning in rgb-d scans. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2021) +23. Collins, J., Goel, S., Deng, K., Luthra, A., Xu, L., Gundogdu, E., Zhang, X., Vicente, T.F.Y., Dideriksen, T., Arora, H., et al.: Abo: Dataset and benchmarks for real-world 3d object understanding. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2022) +24. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2017) +25. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500 (2023) +26. Deitke, M., Liu, R., Wallingford, M., Ngo, H., Michel, O., Kusupati, A., Fan, A., Laforte, C., Voleti, V., Gadre, S.Y., et al.: Objverse-xl: A universe of $10\mathrm{m} + 3\mathrm{d}$ objects. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2023) +27. Deitke, M., Schwenk, D., Salvador, J., Weihs, L., Michel, O., VanderBilt, E., Schmidt, L., Ehsani, K., Kembhavi, A., Farhadi, A.: Objverse: A universe of annotated 3d objects. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2023) +28. Deitke, M., VanderBilt, E., Herrasti, A., Weihs, L., Ehsani, K., Salvador, J., Han, W., Kolve, E., Kembhavi, A., Mottaghi, R.: Procthor: Large-scale embodied ai using procedural generation. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2022) +29. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: Proceedings of Conference of the North American Chapter of the Association for Computational Linguistics (NAACL) (2018) + +30. Ding, R., Yang, J., Xue, C., Zhang, W., Bai, S., Qi, X.: Pla: Language-driven open-vocabulary 3d scene understanding. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2023) +31. Ding, Z., Han, X., Niethammer, M.: Volenet: A deep learning label fusion method for multi-atlas segmentation. In: Proceedings of International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI) (2019) +32. Feng, M., Li, Z., Li, Q., Zhang, L., Zhang, X., Zhu, G., Zhang, H., Wang, Y., Mian, A.: Free-form description guided 3d visual graph network for object grounding in point cloud. In: Proceedings of International Conference on Computer Vision (ICCV) (2021) +33. Ghiasi, G., Gu, X., Cui, Y., Lin, T.Y.: Scaling open-vocabulary image segmentation with image-level labels. In: Proceedings of European Conference on Computer Vision (ECCV) (2022) +34. Gong, R., Huang, J., Zhao, Y., Geng, H., Gao, X., Wu, Q., Ai, W., Zhou, Z., Terzopoulos, D., Zhu, S.C., Jia, B., Huang, S.: Arnold: A benchmark for language-grounded task learning with continuous states in realistic 3d scenes. In: Proceedings of International Conference on Computer Vision (ICCV) (2023) +35. Gu, Q., Kuwajerwala, A., Morin, S., Jatavallabhula, K.M., Sen, B., Agarwal, A., Rivera, C., Paul, W., Ellis, K., Chellappa, R., et al.: Conceptgraphs: Open-vocabulary 3d scene graphs for perception and planning. arXiv preprint arXiv:2309.16650 (2023) +36. Ha, H., Song, S.: Semantic abstraction: Open-world 3d scene understanding from 2d vision-language models. In: Proceedings of Conference on Robot Learning (CoRL) (2022) +37. He, D., Zhao, Y., Luo, J., Hui, T., Huang, S., Zhang, A., Liu, S.: Transrefer3d: Entity-and-relation aware transformer for fine-grained 3d visual grounding. In: Proceedings of ACM International Conference on Multimedia (MM) (2021) +38. Hegde, D., Valanarasu, J.M.J., Patel, V.: Clip goes 3d: Leveraging prompt tuning for language grounded 3d recognition. In: Proceedings of International Conference on Computer Vision (ICCV) (2023) +39. Hong, Y., Wu, Q., Qi, Y., Rodriguez-Opazo, C., Gould, S.: Vln bert: A recurrent vision-and-language bert for navigation. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2021) +40. Hong, Y., Lin, C., Du, Y., Chen, Z., Tenenbaum, J.B., Gan, C.: 3d concept learning and reasoning from multi-view images. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2023) +41. Huang, J., Yong, S., Ma, X., Linghu, X., Li, P., Wang, Y., Li, Q., Zhu, S.C., Jia, B., Huang, S.: An embodied generalist agent in 3d world. arXiv preprint arXiv:2311.12871 (2023) +42. Huang, P.H., Lee, H.H., Chen, H.T., Liu, T.L.: Text-guided graph neural networks for referring 3d instance segmentation. In: Proceedings of AAAI Conference on Artificial Intelligence (AAAI) (2021) +43. Huang, S., Chen, Y., Jia, J., Wang, L.: Multi-view transformer for 3d visual grounding. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2022) +44. Huang, S., Wang, Z., Li, P., Jia, B., Liu, T., Zhu, Y., Liang, W., Zhu, S.C.: Diffusion-based generation, optimization, and planning in 3d scenes. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2023) +45. Jain, A., Gkanatsios, N., Mediratta, I., Fragkiadaki, K.: Bottom up top down detection transformers for language grounding in images and point clouds. In: Proceedings of European Conference on Computer Vision (ECCV) (2022) + +46. Jiang, L., Zhao, H., Shi, S., Liu, S., Fu, C.W., Jia, J.: Pointgroup: Dual-set point grouping for 3d instance segmentation. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2020) +47. Jiang, N., Liu, T., Cao, Z., Cui, J., Zhang, Z., Chen, Y., Wang, H., Zhu, Y., Huang, S.: Full-body articulated human-object interaction. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2023) +48. Jiang, N., Zhang, Z., Li, H., Ma, X., Wang, Z., Chen, Y., Liu, T., Zhu, Y., Huang, S.: Scaling up dynamic human-scene interaction modeling. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2024) +49. Kaplan, J., McCandlish, S., Henighan, T., Brown, T.B., Chess, B., Child, R., Gray, S., Radford, A., Wu, J., Amodei, D.: Scaling laws for neural language models. arXiv preprint arXiv:2001.08361 (2020) +50. Khanna, M., Mao, Y., Jiang, H., Haresh, S., Shacklett, B., Batra, D., Clegg, A., Undersander, E., Chang, A.X., Savva, M.: Habitat synthetic scenes dataset (hssd-200): An analysis of 3d scene scale and realism tradeoffs for objectgoal navigation. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2024) +51. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: Proceedings of International Conference on Computer Vision (ICCV) (2023) +52. Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., et al.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. In: International Journal of Computer Vision (IJCV) (2017) +53. Lake, B.M., Ullman, T.D., Tenenbaum, J.B., Gershman, S.J.: Building machines that learn and think like people. Behavioral and brain sciences 40, e253 (2017) +54. Li, B., Weinberger, K.Q., Belongie, S., Koltun, V., Ranftl, R.: Language-driven semantic segmentation. In: Proceedings of International Conference on Learning Representations (ICLR) (2022) +55. Li, C., Zhang, R., Wong, J., Gokmen, C., Srivastava, S., Martin-Martín, R., Wang, C., Levine, G., Lingelbach, M., Sun, J., et al.: Behavior-1k: A benchmark for embodied ai with 1,000 everyday activities and realistic simulation. In: Proceedings of Conference on Robot Learning (CoRL) (2023) +56. Li, J., Li, D., Savarese, S., Hoi, S.: BLIP-2: bootstrapping language-image pretraining with frozen image encoders and large language models. In: Proceedings of International Conference on Machine Learning (ICML) (2023) +57. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: Proceedings of International Conference on Machine Learning (ICML) (2022) +58. Li, L.H., Zhang, P., Zhang, H., Yang, J., Li, C., Zhong, Y., Wang, L., Yuan, L., Zhang, L., Hwang, J.N., et al.: Grounded language-image pre-training. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2022) +59. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2023) +60. Liu, M., Shi, R., Kuang, K., Zhu, Y., Li, X., Han, S., Cai, H., Porikli, F., Su, H.: Openshape: Scaling up 3d shape representation towards open-world understanding. arXiv preprint arXiv:2305.10764 (2023) +61. Liu, R., Wu, R., Van Hoorick, B., Tokmakov, P., Zakharov, S., Vondrick, C.: Zero-1-to-3: Zero-shot one image to 3d object. In: Proceedings of International Conference on Computer Vision (ICCV) (2023) + +62. Luo, J., Fu, J., Kong, X., Gao, C., Ren, H., Shen, H., Xia, H., Liu, S.: 3d-sps: Single-stage 3d visual grounding via referred point progressive selection. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2022) +63. Luo, T., Rockwell, C., Lee, H., Johnson, J.: Scalable 3d captioning with pretrained models. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2023) +64. Ma, C.Y., Lu, J., Wu, Z., AlRegib, G., Kira, Z., Socher, R., Xiong, C.: Self-monitoring navigation agent via auxiliary progress estimation. In: Proceedings of International Conference on Learning Representations (ICLR) (2019) +65. Ma, X., Yong, S., Zheng, Z., Li, Q., Liang, Y., Zhu, S.C., Huang, S.: Sqa3d: Situated question answering in 3d scenes. In: Proceedings of International Conference on Learning Representations (ICLR) (2023) +66. Mao, Y., Zhang, Y., Jiang, H., Chang, A., Savva, M.: Multiscan: Scalable rgbd scanning for 3d environments with articulated objects. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2022) +67. Misra, I., Girdhar, R., Joulin, A.: An end-to-end transformer model for 3d object detection. In: Proceedings of International Conference on Computer Vision (ICCV) (2021) +68. Mittal, M., Yu, C., Yu, Q., Liu, J., Rudin, N., Hoeller, D., Yuan, J.L., Singh, R., Guo, Y., Mazhar, H., et al.: Orbit: A unified simulation framework for interactive robot learning environments. Robotics and Automation Letters (RA-L) (2023) +69. Mo, K., Zhu, S., Chang, A.X., Yi, L., Tripathi, S., Guibas, L.J., Su, H.: Partnet: A large-scale benchmark for fine-grained and hierarchical part-level 3d object understanding. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2019) +70. Mu, T., Ling, Z., Xiang, F., Yang, D., Li, X., Tao, S., Huang, Z., Jia, Z., Su, H.: Maniskill: Generalizable manipulation skill benchmark with large-scale demonstrations. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2021) +71. OpenAI: Introducing chatgpt. https://openai.com/blog/chatgpt (2022) +72. OpenAI: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023) +73. Pashevich, A., Schmid, C., Sun, C.: Episodic transformer for vision-and-language navigation. In: Proceedings of International Conference on Computer Vision (ICCV) (2021) +74. Peng, S., Genova, K., Jiang, C., Tagliasacchi, A., Pollefeys, M., Funkhouser, T., et al.: Openscene: 3d scene understanding with open vocabularies. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2023) +75. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: Proceedings of International Conference on Machine Learning (ICML) (2021) +76. Ramakrishnan, S.K., Gokaslan, A., Wijmans, E., Maksymets, O., Clegg, A., Turner, J., Undersander, E., Galuba, W., Westbury, A., Chang, A.X., et al.: Habitatmatterport 3d dataset (hm3d): 1000 large-scale 3d environments for embodied ai. In: Proceedings of Advances in Neural Information Processing Systems Datasets and Benchmarks (NeurIPS Datasets and Benchmarks Track) (2021) +77. Rana, K., Haviland, J., Garg, S., Abou-Chakra, J., Reid, I., Suenderhauf, N.: Sayplan: Grounding large language models using 3d scene graphs for scalable robot task planning. In: Proceedings of Conference on Robot Learning (CoRL) (2023) + +78. Rosinol, A., Violette, A., Abate, M., Hughes, N., Chang, Y., Shi, J., Gupta, A., Carlone, L.: Kimera: From slam to spatial perception with 3d dynamic scene graphs. International Journal of Robotics Research (IJRR) (2021) +79. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2022) +80. Schult, J., Engelmann, F., Hermans, A., Litany, O., Tang, S., Leibe, B.: Mask3d: Mask transformer for 3d semantic instance segmentation. In: Proceedings of International Conference on Robotics and Automation (ICRA) (2023) +81. Smith, L., Gasser, M.: The development of embodied cognition: Six lessons from babies. Artificial life 11(1-2), 13-29 (2005) +82. Takmaz, A., Fedele, E., Sumner, R.W., Pollefeys, M., Tombari, F., Engelmann, F.: Openmask3d: Open-vocabulary 3d instance segmentation. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2023) +83. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023) +84. Vu, T., Kim, K., Luu, T.M., Nguyen, T., Yoo, C.D.: Softgroup for 3d instance segmentation on point clouds. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2022) +85. Wald, J., Avetisyan, A., Navab, N., Tombari, F., Nießner, M.: Rio: 3d object instance re-localization in changing indoor environments. In: Proceedings of International Conference on Computer Vision (ICCV) (2019) +86. Wald, J., Dhamo, H., Navab, N., Tombari, F.: Learning 3d semantic scene graphs from 3d indoor reconstructions. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2020) +87. Wang, T., Mao, X., Zhu, C., Xu, R., Lyu, R., Li, P., Chen, X., Zhang, W., Chen, K., Xue, T., et al.: Embodiedscan: A holistic multi-modal 3d perception suite towards embodied ai. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2024) +88. Wang, X., Huang, Q., Celikyilmaz, A., Gao, J., Shen, D., Wang, Y.F., Wang, W.Y., Zhang, L.: Reinforced cross-modal matching and self-supervised imitation learning for vision-language navigation. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2019) +89. Wang, Z., Chen, Y., Jia, B., Li, P., Zhang, J., Zhang, J., Liu, T., Zhu, Y., Liang, W., Huang, S.: Move as you say interact as you can: Language-guided human motion generation with scene affordance. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2024) +90. Wang, Z., Chen, Y., Liu, T., Zhu, Y., Liang, W., Huang, S.: Humanise: Language-conditioned human motion generation in 3d scenes. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2022) +91. Wu, T., Zhang, J., Fu, X., Wang, Y., Ren, J., Pan, L., Wu, W., Yang, L., Wang, J., Qian, C., et al.: Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2023) +92. Wu, Y., Cheng, X., Zhang, R., Cheng, Z., Zhang, J.: Eda: Explicit text-decoupling and dense alignment for 3d visual grounding. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2023) + +93. Xue, L., Gao, M., Xing, C., Martin-Martin, R., Wu, J., Xiong, C., Xu, R., Niebles, J.C., Savarese, S.: Ulip: Learning a unified representation of language, images, and point clouds for 3d understanding. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2023) +94. Yang, J., Ding, R., Wang, Z., Qi, X.: Regionplc: Regional point-language contrastive learning for open-world 3d scene understanding. arXiv preprint arXiv:2304.00962 (2023) +95. Yang, Y., Jia, B., Zhi, P., Huang, S.: Physcene: Physically interactable 3d scene synthesis for embodied ai. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2024) +96. Yang, Y.Q., Guo, Y.X., Xiong, J.Y., Liu, Y., Pan, H., Wang, P.S., Tong, X., Guo, B.: Swin3d: A pretrained transformer backbone for 3d indoor scene understanding. arXiv preprint arXiv:2304.06906 (2023) +97. Yang, Z., Zhang, S., Wang, L., Luo, J.: Sat: 2d semantics assisted training for 3d visual grounding. In: Proceedings of International Conference on Computer Vision (ICCV) (2021) +98. Yeshwanth, C., Liu, Y.C., Nießner, M., Dai, A.: Scannet++: A high-fidelity dataset of 3d indoor scenes. In: Proceedings of International Conference on Computer Vision (ICCV) (2023) +99. Yuan, Z., Yan, X., Liao, Y., Guo, Y., Li, G., Cui, S., Li, Z.: X-trans2cap: Cross-modal knowledge transfer using transformer for 3d dense captioning. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2022) +100. Yuan, Z., Yan, X., Liao, Y., Zhang, R., Wang, S., Li, Z., Cui, S.: Instancerefer: Cooperative holistic understanding for visual grounding on point clouds through instance multi-level contextual referring. In: Proceedings of International Conference on Computer Vision (ICCV) (2021) +101. Zhang, H., Zhang, P., Hu, X., Chen, Y.C., Li, L., Dai, X., Wang, L., Yuan, L., Hwang, J.N., Gao, J.: Glipv2: Unifying localization and vision-language understanding. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2022) +102. Zhang, R., Guo, Z., Zhang, W., Li, K., Miao, X., Cui, B., Qiao, Y., Gao, P., Li, H.: Pointclip: Point cloud understanding by clip. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2022) +103. Zhang, R., Wang, L., Qiao, Y., Gao, P., Li, H.: Learning 3d representations from 2d pre-trained models via image-to-point masked autoencoders. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2023) +104. Zhang, Y., Gong, Z., Chang, A.X.: Multi3drefer: Grounding text description to multiple 3d objects. In: Proceedings of International Conference on Computer Vision (ICCV) (2023) +105. Zhao, L., Cai, D., Sheng, L., Xu, D.: 3dvg-transformer: Relation modeling for visual grounding on point clouds. In: Proceedings of International Conference on Computer Vision (ICCV) (2021) +106. Zheng, J., Zhang, J., Li, J., Tang, R., Gao, S., Zhou, Z.: Structured3d: A large photo-realistic dataset for structured 3d modeling. In: Proceedings of European Conference on Computer Vision (ECCV) (2020) +107. Zhu, W., Hessel, J., Awadalla, A., Gadre, S.Y., Dodge, J., Fang, A., Yu, Y., Schmidt, L., Wang, W.Y., Choi, Y.: Multimodal c4: An open, billion-scale corpus of images interleaved with text. arXiv preprint arXiv:2304.06939 (2023) + +108. Zhu, Y., Gao, T., Fan, L., Huang, S., Edmonds, M., Liu, H., Gao, F., Zhang, C., Qi, S., Wu, Y.N., et al.: Dark, beyond deep: A paradigm shift to cognitive ai with humanlike common sense. Engineering 6(3), 310-345 (2020) +109. Zhu, Z., Ma, X., Chen, Y., Deng, Z., Huang, S., Li, Q.: 3d-vista: Pre-trained transformer for 3d vision and text alignment. In: Proceedings of International Conference on Computer Vision (ICCV) (2023) +110. Zhu, Z., Zhang, Z., Ma, X., Niu, X., Chen, Y., Jia, B., Deng, Z., Huang, S., Li, Q.: Unifying 3d vision-language understanding via promptable queries. In: Proceedings of European Conference on Computer Vision (ECCV) (2024) \ No newline at end of file diff --git a/2024/SceneVerse_ Scaling 3D Vision-Language Learning for Grounded Scene Understanding/images.zip b/2024/SceneVerse_ Scaling 3D Vision-Language Learning for Grounded Scene Understanding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..106af0c2bb1b91cc5a8cd14439cd96975a801cc4 --- /dev/null +++ b/2024/SceneVerse_ Scaling 3D Vision-Language Learning for Grounded Scene Understanding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:434f7f99a4cf26b91dbe36faf52aa16fb4911b5c131eb689b65833812e67de4e +size 530935 diff --git a/2024/SceneVerse_ Scaling 3D Vision-Language Learning for Grounded Scene Understanding/layout.json b/2024/SceneVerse_ Scaling 3D Vision-Language Learning for Grounded Scene Understanding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ebf49c20478ca03681330b12a34bdd7181cbc7c4 --- /dev/null +++ b/2024/SceneVerse_ Scaling 3D Vision-Language Learning for Grounded Scene Understanding/layout.json @@ -0,0 +1,12079 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 149, + 111, + 463, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 111, + 463, + 148 + ], + "spans": [ + { + "bbox": [ + 149, + 111, + 463, + 148 + ], + "type": "text", + "content": "SCENEVERSE: Scaling 3D Vision-Language Learning for Grounded Scene Understanding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 138, + 167, + 478, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 167, + 478, + 194 + ], + "spans": [ + { + "bbox": [ + 138, + 167, + 478, + 194 + ], + "type": "text", + "content": "Baoxiong Jia\\*, Yixin Chen\\*, Huangyue Yu, Yan Wang, Xuesong Niu, Tengyu Liu, Qing Li, and Siyuan Huang" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 178, + 202, + 434, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 202, + 434, + 236 + ], + "spans": [ + { + "bbox": [ + 178, + 202, + 434, + 236 + ], + "type": "text", + "content": "* indicates equal contribution\nState Key Laboratory of General Artificial Intelligence, BIGAI https://scene-verse.github.io" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 263, + 455, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 263, + 455, + 506 + ], + "spans": [ + { + "bbox": [ + 160, + 263, + 455, + 506 + ], + "type": "text", + "content": "Abstract. 3D vision-language (3D-VL) grounding, which aims to align language with 3D physical environments, stands as a cornerstone in developing embodied agents. In comparison to recent advancements in the 2D domain, grounding language in 3D scenes faces two significant challenges: (i) the scarcity of paired 3D-VL data to support grounded learning of 3D scenes, especially considering complexities within diverse object configurations, rich attributes, and intricate relationships; and (ii) the absence of a unified learning framework to distill knowledge from grounded 3D data. In this work, we aim to address these major challenges in 3D-VL by examining the potential of systematically upscaling 3D-VL learning in indoor scenes. We introduce the first million-scale 3D-VL dataset, ScENEVERSE, encompassing 68K indoor scenes and comprising 2.5M vision-language pairs collected from both human annotations and our scalable scene-graph-based generation approach. We demonstrate that this scaling allows for a unified pre-training framework, Grounded Pretraining for Scenes (GPS), for 3D-VL learning. Through extensive experiments, we showcase the effectiveness of GPS by achieving state-of-the-art performance on existing 3D visual grounding and question-answering benchmarks. We also show that the data scaling effect is not limited to GPS, but is generally beneficial for models on tasks like 3D semantic segmentation. The vast potential of ScENEVERSE and GPS is unveiled through zero-shot transfer experiments in challenging 3D-VL tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 515, + 455, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 515, + 455, + 537 + ], + "spans": [ + { + "bbox": [ + 160, + 515, + 455, + 537 + ], + "type": "text", + "content": "Keywords: 3D Vision-Language " + }, + { + "bbox": [ + 160, + 515, + 455, + 537 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 515, + 455, + 537 + ], + "type": "text", + "content": " Data Scaling " + }, + { + "bbox": [ + 160, + 515, + 455, + 537 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 515, + 455, + 537 + ], + "type": "text", + "content": " Grounded Scene Understanding" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 557, + 230, + 570 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 230, + 570 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 230, + 570 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 581, + 483, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 483, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 483, + 668 + ], + "type": "text", + "content": "The foundation of human cognitive development lies in the grounding of language within the physical world [53,81,108]. Recent progress in Large Language Models (LLMs) [10,11,83] has markedly promoted the alignment between vision and language [3,59,75] utilizing billion-scale vision-language datasets [79,107]. However, with these advancements predominantly focusing on the 2D domain, the grounded understanding of 3D physical environments remains in an incipient stage [1,5,16]. Recognizing the pivotal role of grounded 3D experiences in" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 114, + 478, + 322 + ], + "blocks": [ + { + "bbox": [ + 138, + 114, + 478, + 322 + ], + "lines": [ + { + "bbox": [ + 138, + 114, + 478, + 322 + ], + "spans": [ + { + "bbox": [ + 138, + 114, + 478, + 322 + ], + "type": "image", + "image_path": "17e5b5f02f00295aaa4e2093431eabf3f15e1646a4650c7a71a4a3b69e233eae.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 324, + 482, + 358 + ], + "lines": [ + { + "bbox": [ + 132, + 324, + 482, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 324, + 482, + 358 + ], + "type": "text", + "content": "Fig. 1: Overview of ScENEVERSE. A million-scale 3D vision-language dataset that comprises over 68K various 3D indoor scenes and 2.5M aligned scene-language pairs in the form of scene caption, object caption, and object referral." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 365, + 481, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 481, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 481, + 389 + ], + "type": "text", + "content": "shaping human cognition [7,8], there is a compelling need to focus on exploring vision-language learning in the context of 3D scenes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 389, + 482, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 389, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 482, + 569 + ], + "type": "text", + "content": "Seeking insights from success in 2D vision-language (2D-VL) learning, a major factor to the success was the notable scale-up of paired vision-language data [15, 52, 79]. However, applying this experience directly from 2D to 3D is fraught with challenges. Primarily, 3D data collection heavily relies on the scanning device, making it inherently much more complex and expensive than gathering 2D images. Despite steady efforts to increase the volume of 3D scene data [9, 24, 66, 98], most datasets remain limited to thousands of scenes, substantially lagging behind the scale of existing 2D datasets. This gap is further widened by the inherent complexities of 3D scenes, which feature a multitude of object instances with diverse attributes, varying arrangements, and intricate inter-object relationships. These unique aspects of 3D scenes not only make the accurate description of objects and their relations more challenging but also considerably increase the number of language descriptions required for thorough scene depiction. Consequently, this presents a significant challenge in gathering sufficient and high-quality paired scene-language data for grounded scene understanding." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 570, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 483, + 666 + ], + "type": "text", + "content": "To confront these challenges, we propose SCENEVERSE, the first million-scale dataset aimed at advancing 3D vision-language (3D-VL) learning for grounded scene understanding. At the scene level, we unify 3D scene data from existing datasets [9,24,66,76,85], aligning scenes and annotations from various capturing sources, and supplement the collection with synthetic scenes [28,106]. This compilation represents the most extensive 3D scene data gathered to date, amounting to 68K scenes. For language, we first present 97K newly-annotated referring expressions, the most extensive thus far. We additionally propose an" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 270, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 270, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 270, + 101 + ], + "type": "text", + "content": "B. Jia and Y. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": "automated generation pipeline utilizing 3D scene graphs [4,86] and LLMs to create comprehensive, high-quality scene-language pairs. This refined collection, totaling 2.5M scene-language pairs, provides detailed and comprehensive descriptions of both object-level and scene-level descriptions within the 3D scene." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 483, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 483, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 483, + 307 + ], + "type": "text", + "content": "We thoroughly investigate the potential offered by SCENEVERSE with large-scale pre-training, introducing Grounded Pre-training for Scenes (GPS), a novel and unified pre-training framework designed for scene and object-level alignment without auxiliary losses. Through multi-level contrastive alignment, we achieve significant performance boosts on 3D-VL tasks, such as grounding and question answering, setting new state-of-the-art results via a simple and effective pretraining process. We unveil the vast possibilities offered by SCENEVERSE and GPS in 3D-VL tasks in a zero-shot transfer setting. Additionally, we show that the scaling effect in SCENEVERSE is not limited to GPS, but generally benefits models in tasks like 3D semantic segmentation. At last, we offer deeper insights into the data-scaling in SCENEVERSE through extensive ablative experiments, pointing out future directions. Our main contributions are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 136, + 308, + 482, + 452 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 136, + 308, + 482, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 308, + 482, + 368 + ], + "spans": [ + { + "bbox": [ + 136, + 308, + 482, + 368 + ], + "type": "text", + "content": "1. We introduce SCENEVERSE, the first million-scale 3D-VL dataset for grounded scene understanding. SCENEVERSE encompasses 68K 3D scenes coupled with 2.5M scene-language pairs, sourced through a combination of human annotation and automated generation methods. This represents a significant improvement in terms of data diversity and scale compared to prior datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 136, + 368, + 482, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 368, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 136, + 368, + 482, + 403 + ], + "type": "text", + "content": "2. We propose GPS, a transformer-based model trained with multi-level scenetest alignment that achieves state-of-the-art results on existing 3D-VL grounding and question-answering benchmarks by pre-training on ScENEVERSE." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 136, + 403, + 482, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 403, + 482, + 452 + ], + "spans": [ + { + "bbox": [ + 136, + 403, + 482, + 452 + ], + "type": "text", + "content": "3. We demonstrate that with the data scale-up and model design, our pre-trained GPS exhibit emerging zero-shot generalization capabilities in grounded scene understanding. We also show that this scaling effect is not limited to GPS, but is generally beneficial for models on tasks like semantic segmentation." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 471, + 237, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 471, + 237, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 471, + 237, + 483 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 497, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 482, + 594 + ], + "type": "text", + "content": "Datasets for Grounded 3D Understanding Obtaining aligned 3D-language data is a inherently difficult task. In 3D object modeling, pioneering works like ShapeNet [14] sourced 3D assets from online repositories, leading to a proliferation of high-quality 3D object datasets [23, 69, 91]. Notably, recent developments include internet-scale data collection with Objaverse [26, 27], accompanied by the integration of object-level captions [93] for 3D-language alignment. Models trained on these datasets demonstrate an enhanced understanding of objects, evident in classification [60], generation [61], and captioning tasks [63]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "In contrast, developing datasets for grounded 3D scene understanding is even more challenging due to the extensive requirements for scene acquisition and annotation. Existing works curate RGB-D and scanned indoor scene datasets [9,13, 24,66,76,85] and synthetic scenes [28,50,95,106] used for benchmarking tasks like 3D object detection and segmentation [31,46,67,80,84]. These semantically labeled scenes are subsequently used in fine-grained scene grounding tasks like object" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 148, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 148, + 91, + 447, + 103 + ], + "type": "text", + "content": "Scaling 3D Vision-Language Learning for Grounded Scene Understanding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 170, + 480, + 297 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "type": "text", + "content": "Table 1: Comparison of SCENEVERSE with existing 3DVL Datasets. SCENEVERSE expands the data scale of prior work by order of magnitude. \"VG\" stands for Visual Grounding, \"QA\" for Question Answering, \"PT\" for Pre-training and \"MT\" for Multi-tasking. \"Anno.\" denotes language from human annotations and \"Syn.\" for template-based or LLM generated descriptions." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 170, + 480, + 297 + ], + "lines": [ + { + "bbox": [ + 134, + 170, + 480, + 297 + ], + "spans": [ + { + "bbox": [ + 134, + 170, + 480, + 297 + ], + "type": "table", + "html": "
Dataset3D Scene Obj.TaskObj. CaptionScene CaptionObj. ReferralQuality CheckNew Anno.Syn.Existing Anno.Syn.Total
ScanRefer [16]VGXX✓✓52K---52K
ReferIt3D [1]| |VGXX✓✓42K200K--242K
ScanQA [5]1.5K33KQA---✓27K---27K
SQA3D [65]QA---✓33K---33K
Multi3DRefer [104]| |VGXX✓✓-10K52K-62K
Cap3D [63]-666KVGX✓XX58K666K--724K
ScanScribe [109]3K56KPTXX✓X-90K94K94K278K
3D-LLM [40]1.5K186KMT✓✓✓X-659K--659K
EmbodiedScan [87]5K890KVGXX✓X-970K--970K
LEO [41]3K56KMT✓✓✓✓-188K235K90K513K
SCENEVERSE68K1.5MVG✓✓✓✓96K2.1M94K200K2.5M
", + "image_path": "1ef0c5f8d02b680dafa1e8af6603cca974db298dbf4c96ede221e70067913dbe.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 308, + 482, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 482, + 406 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 482, + 406 + ], + "type": "text", + "content": "referral [1, 16, 104], captioning [17, 19, 22, 99], vision-language-navigation [39, 64, 73, 88] and reasoning [5, 40, 65]. Recent works exploit the representation of 3D scene graph (3DSG) [4, 20, 78, 86], which concisely describes scenes with hierarchical structures. This representation is notably advantageous for planning [2, 77] and captioning [35], owing to its compatibility with LLMs for flexible description generation [40, 41]. Nonetheless, as shown in Tab. 1, most datasets are constrained in both scene and language scales, underscoring the need for scaling up fine-grained and aligned scene-language data to enhance grounded scene understanding." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 414, + 482, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 482, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 482, + 486 + ], + "type": "text", + "content": "Vision-Language Learning Recent years have witnessed tremendous progress in 2D-VL [3, 21, 25, 57, 59, 75], empowered by transformer-based pre-training models [11, 29, 72] and large-scale image-language datasets [15, 79]. A central theme across 2D-VL domains is the effectiveness of data scaling [49], as demonstrated by improved alignment and expanded capabilities in open-vocabulary understanding [33, 51, 54, 58] through a contrastive pre-training pipeline [75]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": "However, in grounded scene understanding, the primary challenge for models has been the limited availability of paired 3D scene-language data, which restricts the application of insights drawn from 2D-VL. Current models for 3D scene grounding [6,18,37,43,45,62,92,97,105] heavily rely on task-specific knowledge in both model and loss designs or advanced optimization strategies [109]. To bridge this gap, there has been a growing emphasis on employing pre-trained 2D-VL models for 3D-VL [36,38,74,82,93,102,103]. Yet, these models mostly draw information available from 2D-VL models (e.g., object attribute, affordance, etc.), falling short on capturing crucial 3D information like object spatial relationships which are necessary for more fine-grained tasks such as grounded human-scene [20, 44,47,48,89,90] and robot-scene interactions modeling [34,55,68,70]. This urges the need for a multi-level alignment between language and 3D scenes, particularly regarding 3D-specific information. Considering the nascent stage of existing 3D pre-training methods [30,94,109,110], we believe SCENEVERSE and GPS have the potential to spearhead new avenues in 3D-VL research." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 270, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 270, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 270, + 101 + ], + "type": "text", + "content": "B. Jia and Y. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 233, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 233, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 233, + 127 + ], + "type": "text", + "content": "3 SCENEVERSE" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 482, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 482, + 210 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 482, + 210 + ], + "type": "text", + "content": "SCENEVERSE is designed for grounded scene understanding with 3D scenes curated from diverse existing datasets of both real and synthetic environments. Regarding language, we employ both human annotation and a novel automated generation pipeline to collect comprehensive and high-quality language for both object-level and scene-level descriptions. We provide details regarding data collection in the following sections." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 224, + 238, + 236 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 224, + 238, + 236 + ], + "spans": [ + { + "bbox": [ + 131, + 224, + 238, + 236 + ], + "type": "text", + "content": "3.1 Scene Curation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 241, + 482, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 241, + 482, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 241, + 482, + 373 + ], + "type": "text", + "content": "To address the scarcity of available 3D scene data, we construct SCENEVERSE by unifying 3D scene data from various existing datasets. We use real-world scene datasets, including ScanNet [24], ARKitScenes [9], HM3D [76], 3RScan [85] and MultiScan [66], alongside synthetic environments from Structured3D [106] and ProcTHOR [28]. The inclusion of these synthetic datasets is mainly motivated by their potential as scalable data sources for 3D-VL alignment. To ensure cohesion across various sources, we conduct preprocessing steps such as room segmentation, point subsampling, axis alignment, normalization, and semantic label alignment. Each scan is represented by a point cloud " + }, + { + "bbox": [ + 130, + 241, + 482, + 373 + ], + "type": "inline_equation", + "content": "\\mathbf{P} \\in \\mathbb{R}^{N \\times 8}" + }, + { + "bbox": [ + 130, + 241, + 482, + 373 + ], + "type": "text", + "content": ", wherein each point is defined by its 3D coordinates, RGB color, instance id, and semantic label. In total, we curate 68,406 3D scenes in SCENEVERSE." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 388, + 324, + 400 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 388, + 324, + 400 + ], + "spans": [ + { + "bbox": [ + 131, + 388, + 324, + 400 + ], + "type": "text", + "content": "3.2 Referral Annotation by Humans" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 406, + 482, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 406, + 482, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 406, + 482, + 525 + ], + "type": "text", + "content": "In the curated scenes of SCENEVERSE, we present the most comprehensive set of human-annotated, context-rich object referrals to date, serving as a valuable benchmark for assessing grounded scene understanding capabilities. The human annotations contain 96,863 descriptions in ARKitScenes [9], HM3D [76] and MultiScan [66]. During the annotation process, one human annotator was assigned to write at least 20 words to distinctly refer to a single 3D object within a 3D scene. Each referral text then undergoes independent verification by two additional reviewers, both mandated to accurately locate the referenced object based on the 3D scene and the annotated referral text. Any object referrals that do not pass the verification by either reviewer are flagged for re-annotation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 540, + 313, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 540, + 313, + 552 + ], + "spans": [ + { + "bbox": [ + 131, + 540, + 313, + 552 + ], + "type": "text", + "content": "3.3 3D Scene Graph Construction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "Our 3D scene graph is defined as a hierarchical graph " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{G} = (\\mathcal{V},\\mathcal{E})" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": ". Each node " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "v\\in \\mathcal{V}" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " represents one distinct 3D object instance, parameterized by its centroid " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb {p}_i\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " and bounding box size of " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb {b}_i = (b_x,b_y,b_z)\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": ". The edges " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " represent spatial relationships between nodes. To construct the scene graph " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": ", we instantiate the nodes with the instance annotation from the point clouds and assign object classes with their corresponding semantic labels. Following prior work [1,86], we consider the Vertical proximity, Horizontal proximity and Multi-object Relationships as spatial relations. For a more detailed description of the scene graph construction and relationship determination, please refer to supplementary." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "type": "text", + "content": "Scaling 3D Vision-Language Learning for Grounded Scene Understanding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 115, + 269, + 197 + ], + "blocks": [ + { + "bbox": [ + 136, + 115, + 269, + 197 + ], + "lines": [ + { + "bbox": [ + 136, + 115, + 269, + 197 + ], + "spans": [ + { + "bbox": [ + 136, + 115, + 269, + 197 + ], + "type": "image", + "image_path": "d6fe1d32398d30be9ead3924f24ac73f57a2feb2f73b4c29ecba65b7046d0bfd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 187, + 199, + 216, + 205 + ], + "lines": [ + { + "bbox": [ + 187, + 199, + 216, + 205 + ], + "spans": [ + { + "bbox": [ + 187, + 199, + 216, + 205 + ], + "type": "text", + "content": "(a) 3D Scene" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 301, + 481, + 335 + ], + "lines": [ + { + "bbox": [ + 132, + 301, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 301, + 481, + 335 + ], + "type": "text", + "content": "Fig. 2: SCENEVERSE collection and statistics. Given a 3D scene (a), our automated pipeline (c) generates three types of description including scene caption, object caption and object referral. (b) SCENEVERSE data comparison and composition." + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 155, + 209, + 166, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 209, + 166, + 214 + ], + "spans": [ + { + "bbox": [ + 155, + 209, + 166, + 214 + ], + "type": "text", + "content": "一" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 169, + 209, + 200, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 209, + 200, + 214 + ], + "spans": [ + { + "bbox": [ + 169, + 209, + 200, + 214 + ], + "type": "text", + "content": "Scene Cap" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 217, + 186, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 217, + 186, + 224 + ], + "spans": [ + { + "bbox": [ + 138, + 217, + 186, + 224 + ], + "type": "text", + "content": "Sub-graph Context" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 225, + 203, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 225, + 203, + 230 + ], + "spans": [ + { + "bbox": [ + 141, + 225, + 203, + 230 + ], + "type": "text", + "content": "{scene_type':Bedroom'," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 231, + 203, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 203, + 236 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 203, + 236 + ], + "type": "text", + "content": "'object_count': {'nightstand':2,...}," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 237, + 204, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 237, + 204, + 242 + ], + "spans": [ + { + "bbox": [ + 143, + 237, + 204, + 242 + ], + "type": "text", + "content": "'relation': {nightstand', 'on', floor}," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 143, + 243, + 204, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 243, + 204, + 247 + ], + "spans": [ + { + "bbox": [ + 143, + 243, + 204, + 247 + ], + "type": "text", + "content": "['backback', in front of, b" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 248, + 171, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 248, + 171, + 255 + ], + "spans": [ + { + "bbox": [ + 138, + 248, + 171, + 255 + ], + "type": "text", + "content": "Summary" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 255, + 236, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 255, + 236, + 259 + ], + "spans": [ + { + "bbox": [ + 141, + 255, + 236, + 259 + ], + "type": "text", + "content": "Prompt: Provide a summary for a scene from a given" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 260, + 216, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 260, + 216, + 265 + ], + "spans": [ + { + "bbox": [ + 141, + 260, + 216, + 265 + ], + "type": "text", + "content": "scene graph delimited by triple backticks," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 266, + 247, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 266, + 247, + 270 + ], + "spans": [ + { + "bbox": [ + 141, + 266, + 247, + 270 + ], + "type": "text", + "content": "Response: In this bedroom, there are two nightstands, ..." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 271, + 247, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 271, + 247, + 276 + ], + "spans": [ + { + "bbox": [ + 141, + 271, + 247, + 276 + ], + "type": "text", + "content": "The backpack is in front of the nightstand as well . The roomapproves to be functional with the nightstands providing" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 277, + 233, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 277, + 233, + 282 + ], + "spans": [ + { + "bbox": [ + 141, + 277, + 233, + 282 + ], + "type": "text", + "content": "storage space and the telephone for communication" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 270, + 199, + 288, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 270, + 199, + 288, + 205 + ], + "spans": [ + { + "bbox": [ + 270, + 199, + 288, + 205 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 290, + 199, + 329, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 290, + 199, + 329, + 205 + ], + "spans": [ + { + "bbox": [ + 290, + 199, + 329, + 205 + ], + "type": "text", + "content": "(b) Language radar" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 330, + 199, + 354, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 199, + 354, + 205 + ], + "spans": [ + { + "bbox": [ + 330, + 199, + 354, + 205 + ], + "type": "text", + "content": "ar chart and" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 288, + 209, + 329, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 288, + 209, + 329, + 214 + ], + "spans": [ + { + "bbox": [ + 288, + 209, + 329, + 214 + ], + "type": "text", + "content": "Object Caption" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 255, + 217, + 293, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 217, + 293, + 224 + ], + "spans": [ + { + "bbox": [ + 255, + 217, + 293, + 224 + ], + "type": "text", + "content": "BLIP2 Captions" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 257, + 224, + 314, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 224, + 314, + 229 + ], + "spans": [ + { + "bbox": [ + 257, + 224, + 314, + 229 + ], + "type": "text", + "content": "1. A bed in a hotel room. (0.85)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 257, + 229, + 326, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 229, + 326, + 235 + ], + "spans": [ + { + "bbox": [ + 257, + 229, + 326, + 235 + ], + "type": "text", + "content": "2. A white comforter on a bed. (0.83)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 257, + 235, + 326, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 235, + 326, + 239 + ], + "spans": [ + { + "bbox": [ + 257, + 235, + 326, + 239 + ], + "type": "text", + "content": "3. A bed with a striped comforter. (0.83)" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 257, + 240, + 302, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 240, + 302, + 246 + ], + "spans": [ + { + "bbox": [ + 257, + 240, + 302, + 246 + ], + "type": "text", + "content": "N.A picture of cat. (0.63)" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 255, + 247, + 293, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 247, + 293, + 255 + ], + "spans": [ + { + "bbox": [ + 255, + 247, + 293, + 255 + ], + "type": "text", + "content": "Summary" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 257, + 255, + 355, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 255, + 355, + 260 + ], + "spans": [ + { + "bbox": [ + 257, + 255, + 355, + 260 + ], + "type": "text", + "content": "Prompt: Summarize the captions below. The summary" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 257, + 260, + 350, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 260, + 350, + 265 + ], + "spans": [ + { + "bbox": [ + 257, + 260, + 350, + 265 + ], + "type": "text", + "content": "should be a description of the (object). Focus on the" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 257, + 266, + 326, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 266, + 326, + 270 + ], + "spans": [ + { + "bbox": [ + 257, + 266, + 326, + 270 + ], + "type": "text", + "content": "{object's attributes, like color, shape, n" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 257, + 271, + 352, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 271, + 352, + 276 + ], + "spans": [ + { + "bbox": [ + 257, + 271, + 352, + 276 + ], + "type": "text", + "content": "Identity and correct the potential errors ..." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 257, + 277, + 351, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 277, + 351, + 282 + ], + "spans": [ + { + "bbox": [ + 257, + 277, + 351, + 282 + ], + "type": "text", + "content": "Response : The bed is in a motel room with a stripedcomforter . It has a white comforter and a blanket on it" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 257, + 282, + 354, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 282, + 354, + 288 + ], + "spans": [ + { + "bbox": [ + 257, + 282, + 354, + 288 + ], + "type": "text", + "content": "The bed is also in a room with a bedside table." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 272, + 291, + 351, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 272, + 291, + 351, + 298 + ], + "spans": [ + { + "bbox": [ + 272, + 291, + 351, + 298 + ], + "type": "text", + "content": "(c) Automated language generation" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 402, + 209, + 425, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 209, + 425, + 214 + ], + "spans": [ + { + "bbox": [ + 402, + 209, + 425, + 214 + ], + "type": "text", + "content": "Object R" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 370, + 217, + 421, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 217, + 421, + 224 + ], + "spans": [ + { + "bbox": [ + 370, + 217, + 421, + 224 + ], + "type": "text", + "content": "Relationship Triplets" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 372, + 224, + 412, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 224, + 412, + 228 + ], + "spans": [ + { + "bbox": [ + 372, + 224, + 412, + 228 + ], + "type": "text", + "content": "1. {table', 'chair', 'left'}" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 372, + 229, + 444, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 229, + 444, + 234 + ], + "spans": [ + { + "bbox": [ + 372, + 229, + 444, + 234 + ], + "type": "text", + "content": "2.{'bed', ('lamp', 'mini fridge'), 'between'}" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 372, + 235, + 430, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 235, + 430, + 239 + ], + "spans": [ + { + "bbox": [ + 372, + 235, + 430, + 239 + ], + "type": "text", + "content": "Template-based Referral" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 370, + 240, + 453, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 240, + 453, + 246 + ], + "spans": [ + { + "bbox": [ + 370, + 240, + 453, + 246 + ], + "type": "text", + "content": "1. The table is to the left of the chair. \n2. The chair is in the right of the chair." + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 370, + 247, + 408, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 247, + 408, + 251 + ], + "spans": [ + { + "bbox": [ + 370, + 247, + 408, + 251 + ], + "type": "text", + "content": "2. It's a bed in the m" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 372, + 252, + 408, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 252, + 408, + 256 + ], + "spans": [ + { + "bbox": [ + 372, + 252, + 408, + 256 + ], + "type": "text", + "content": "Rephrasing" + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 372, + 256, + 476, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 256, + 476, + 261 + ], + "spans": [ + { + "bbox": [ + 372, + 256, + 476, + 261 + ], + "type": "text", + "content": "Prompt: Rewrite the following sentence using one random word." + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 372, + 261, + 433, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 261, + 433, + 266 + ], + "spans": [ + { + "bbox": [ + 372, + 261, + 433, + 266 + ], + "type": "text", + "content": "sentence structure. Focus on the" + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 372, + 266, + 416, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 266, + 416, + 270 + ], + "spans": [ + { + "bbox": [ + 372, + 266, + 416, + 270 + ], + "type": "text", + "content": "about the {target_object}" + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 372, + 271, + 393, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 271, + 393, + 274 + ], + "spans": [ + { + "bbox": [ + 372, + 271, + 393, + 274 + ], + "type": "text", + "content": "Response:" + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 372, + 275, + 459, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 275, + 459, + 279 + ], + "spans": [ + { + "bbox": [ + 372, + 275, + 459, + 279 + ], + "type": "text", + "content": "1. The table is situated to the left of the armchair." + } + ] + } + ], + "index": 46 + }, + { + "bbox": [ + 372, + 279, + 474, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 279, + 474, + 285 + ], + "spans": [ + { + "bbox": [ + 372, + 279, + 474, + 285 + ], + "type": "text", + "content": "2. The bed occupies the space between the lamp and the wall." + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 372, + 285, + 443, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 285, + 443, + 289 + ], + "spans": [ + { + "bbox": [ + 372, + 285, + 443, + 289 + ], + "type": "text", + "content": "mini fridge, creating a cozy atmosphere." + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 132, + 342, + 327, + 354 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 327, + 354 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 327, + 354 + ], + "type": "text", + "content": "3.4 Language Generation with LLMs" + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 130, + 359, + 482, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 359, + 482, + 419 + ], + "spans": [ + { + "bbox": [ + 130, + 359, + 482, + 419 + ], + "type": "text", + "content": "The scene-language pairs in ScENEVERSE aim to capture varying aspects of the 3D scene, including detailed object attributes in object captioning, spatial relationships between objects in object referral, and global scene descriptions in scene captioning. Based on 3D scene graphs, we utilize both templates and LLMs to automatically generate descriptions on these three granularities." + } + ] + } + ], + "index": 51 + }, + { + "bbox": [ + 130, + 422, + 482, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 482, + 530 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 482, + 530 + ], + "type": "text", + "content": "Object Captioning Object captions aim to provide detailed descriptions of an object's visual and physical properties, facilitating object-level grounding with its distinctive features. Given the multi-view images, we utilize the point cloud of the object " + }, + { + "bbox": [ + 130, + 422, + 482, + 530 + ], + "type": "inline_equation", + "content": "v \\in V" + }, + { + "bbox": [ + 130, + 422, + 482, + 530 + ], + "type": "text", + "content": " to identify its occurrence in the images through rendering. The images are then cropped with the rendered bounding boxes and processed through BLIP2 [56] to generate initial object captions. We select the top 10 sentences with the highest CLIP [75] similarity score and minimal occlusion and utilize an LLM to obtain a refined and coherent summary of the captions. The detailed object captioning pipeline is illustrated in supplementary." + } + ] + } + ], + "index": 52 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Object Referral Object relationship captions refer to objects by articulating their spatial relationships in the scene. Spatial relationship triplets " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "(v_{i}, v_{j}, e_{ij})" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " are first extracted from the constructed 3D scene graph. We design various templates to generate descriptions for each relationship type, assigning the entities in the form of (target-object, spatial-relation, anchor-object(s)). This results in examples like \"the chair is next to the li2023behaviorarmchair\", \"facing the sofa, there is a suitcase far to the right of the shoes\", and \"the fridge is between cabinet and sofa\". Our designed templates span passive and active tenses, as well as inversion clauses, contributing to the richness of the generated text. To enhance the descriptions' naturalness, we employ LLM for sentence rephrasing. Statistics for the descriptions before and after rephrasing are presented in Fig. 2 (b)." + } + ] + } + ], + "index": 53 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 270, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 270, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 270, + 101 + ], + "type": "text", + "content": "B. Jia and Y. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "Scene Captioning The scene-level captions emphasize global information, portraying the key objects in the scene along with their attributes and functionalities. We use the constructed 3D scene graph and prompt LLMs to generate these captions. We random sample a subset of edges and nodes from the scene graph each time as the scene context to enhance the diversity of scene captions. The object counts are also provided as LLM prompts, together with the room type and object attributes if such annotations are available in the dataset." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 219, + 299, + 231 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 219, + 299, + 231 + ], + "spans": [ + { + "bbox": [ + 131, + 219, + 299, + 231 + ], + "type": "text", + "content": "3.5 Data Quality and Statistics" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 239, + 482, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 239, + 482, + 347 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 482, + 347 + ], + "type": "text", + "content": "Data Quality The 96K human-annotated set of ScENeVErSE is collected through AMT, where 82 humans are employed for annotation and 18 for verification. All final annotations passed the reference verification, with a re-annotation rate of " + }, + { + "bbox": [ + 130, + 239, + 482, + 347 + ], + "type": "inline_equation", + "content": "4.8\\%" + }, + { + "bbox": [ + 130, + 239, + 482, + 347 + ], + "type": "text", + "content": ". For our automatic language generation pipeline, we conduct extensive prompt tuning and iterate with human feedback for LLMs on object captioning, summary, and rephrasing. To verify the efficacy of the pipeline, we conduct a quality check where 12K generated object-level descriptions are randomly selected for human verification. Results demonstrate a " + }, + { + "bbox": [ + 130, + 239, + 482, + 347 + ], + "type": "inline_equation", + "content": "96.93\\%" + }, + { + "bbox": [ + 130, + 239, + 482, + 347 + ], + "type": "text", + "content": " pass rate, surpassing that in ReferIt3D [1] with " + }, + { + "bbox": [ + 130, + 239, + 482, + 347 + ], + "type": "inline_equation", + "content": "86.1\\%" + }, + { + "bbox": [ + 130, + 239, + 482, + 347 + ], + "type": "text", + "content": " pass rate on 2K samples." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 354, + 483, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 483, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 483, + 498 + ], + "type": "text", + "content": "Statistics In total, SCENEVERSE comprises a total of 68,406 room-level 3D scans, with the source composition shown in Fig. 2 (b). The dataset contains 1.5M object instances ranging in 2290 object categories. Our generated 3D scene graph comprises 21 types of relationships following prior work [1,86]. For the language descriptions, we generate 1M template-based texts and 1M sentences rephrased by Llama [83] and GPT-3.5 [71]. As can be seen from the radar chart and examples in Fig. 2, the diversity of the LLM-refined data, particularly in sentence length and variety, closely aligns with the characteristics of annotated descriptions, surpassing the template-based data. Together with existing sources (294K) and our newly annotated set (96K), SCENEVERSE contains 2.5M scene-language pairs in total. All the rephrasing and summary prompts, along with the complete set of relationships, are detailed in supplementary." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 517, + 354, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 517, + 354, + 530 + ], + "spans": [ + { + "bbox": [ + 131, + 517, + 354, + 530 + ], + "type": "text", + "content": "4 Grounded Pre-training for Scenes" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 542, + 482, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 542, + 482, + 602 + ], + "spans": [ + { + "bbox": [ + 130, + 542, + 482, + 602 + ], + "type": "text", + "content": "In this section, we introduce GPS, an efficient transformer-based model trained with multi-level contrastive losses for aligning 3D scenes and texts. As shown in Fig. 3, we echo the language descriptions collected at different granularities to form contrastive objectives at both object-level, referral-object-level, and scene-level in GPS. We describe the design of each level in the following sections." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 620, + 280, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 620, + 280, + 633 + ], + "spans": [ + { + "bbox": [ + 131, + 620, + 280, + 633 + ], + "type": "text", + "content": "4.1 Object-level Grounding" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 667 + ], + "type": "text", + "content": "Given a 3D scene point cloud " + }, + { + "bbox": [ + 130, + 641, + 481, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 130, + 641, + 481, + 667 + ], + "type": "text", + "content": ", we use an off-the-shelf 3D object segmentation model to decompose it into a bag of " + }, + { + "bbox": [ + 130, + 641, + 481, + 667 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 641, + 481, + 667 + ], + "type": "text", + "content": " objects " + }, + { + "bbox": [ + 130, + 641, + 481, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{S} = \\{\\pmb{o}_1, \\pmb{o}_2, \\dots, \\pmb{o}_n\\}_{i=1}^N" + }, + { + "bbox": [ + 130, + 641, + 481, + 667 + ], + "type": "text", + "content": ". We extract" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 148, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 148, + 91, + 447, + 103 + ], + "type": "text", + "content": "Scaling 3D Vision-Language Learning for Grounded Scene Understanding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 115, + 482, + 243 + ], + "blocks": [ + { + "bbox": [ + 134, + 115, + 482, + 243 + ], + "lines": [ + { + "bbox": [ + 134, + 115, + 482, + 243 + ], + "spans": [ + { + "bbox": [ + 134, + 115, + 482, + 243 + ], + "type": "image", + "image_path": "65c4acf2fbe695e949ed45de515b28bdd9cad216fdb934a32bfdd220c1fa25da.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 246, + 482, + 269 + ], + "lines": [ + { + "bbox": [ + 132, + 246, + 482, + 269 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 482, + 269 + ], + "type": "text", + "content": "Fig. 3: Overview of GPS model. We use contrastive alignment at three levels " + }, + { + "bbox": [ + 132, + 246, + 482, + 269 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{obj}}" + }, + { + "bbox": [ + 132, + 246, + 482, + 269 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 246, + 482, + 269 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{scene}}" + }, + { + "bbox": [ + 132, + 246, + 482, + 269 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 132, + 246, + 482, + 269 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ref}}" + }, + { + "bbox": [ + 132, + 246, + 482, + 269 + ], + "type": "text", + "content": " and a masked language modeling objective " + }, + { + "bbox": [ + 132, + 246, + 482, + 269 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{MLM}}" + }, + { + "bbox": [ + 132, + 246, + 482, + 269 + ], + "type": "text", + "content": " for model learning." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 277, + 482, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 482, + 316 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 482, + 316 + ], + "type": "text", + "content": "object features " + }, + { + "bbox": [ + 130, + 277, + 482, + 316 + ], + "type": "inline_equation", + "content": "\\{\\pmb{f}_i^O\\}" + }, + { + "bbox": [ + 130, + 277, + 482, + 316 + ], + "type": "text", + "content": " with an object point cloud encoder and text features " + }, + { + "bbox": [ + 130, + 277, + 482, + 316 + ], + "type": "inline_equation", + "content": "\\{\\pmb{f}_i^T\\}" + }, + { + "bbox": [ + 130, + 277, + 482, + 316 + ], + "type": "text", + "content": " by feeding object-captions " + }, + { + "bbox": [ + 130, + 277, + 482, + 316 + ], + "type": "inline_equation", + "content": "\\{\\pmb{T}_i^{\\mathrm{obj}}\\}" + }, + { + "bbox": [ + 130, + 277, + 482, + 316 + ], + "type": "text", + "content": " into a frozen language model. Following [93], we perform cross-modal alignment on the object features and text features via:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 146, + 323, + 482, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 323, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 146, + 323, + 482, + 357 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {o b j}} = - \\frac {1}{2} \\sum_ {(p, q)} \\left(\\log \\frac {\\exp \\left(D ^ {\\mathrm {o b j}} (p , q)\\right)}{\\sum_ {r} \\exp \\left(D ^ {\\mathrm {o b j}} (p , r)\\right)} + \\log \\frac {\\exp \\left(D ^ {\\mathrm {o b j}} (p , q)\\right)}{\\sum_ {r} \\exp \\left(D ^ {\\mathrm {o b j}} (r , q)\\right)}\\right), \\tag {1}", + "image_path": "5c5bfd4881a9d0686bffe746d52606c266daf44c79200e487947035d5e6a46da.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 365, + 483, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 483, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 483, + 415 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 365, + 483, + 415 + ], + "type": "inline_equation", + "content": "D^{\\mathrm{obj}}(p,q) = (\\pmb{f}_p^O\\pmb{f}_q^T /\\tau)" + }, + { + "bbox": [ + 130, + 365, + 483, + 415 + ], + "type": "text", + "content": " denotes the dot product between object and text features and " + }, + { + "bbox": [ + 130, + 365, + 483, + 415 + ], + "type": "inline_equation", + "content": "(p,q)" + }, + { + "bbox": [ + 130, + 365, + 483, + 415 + ], + "type": "text", + "content": " denotes a pair of aligned object-text pair in the training batch and " + }, + { + "bbox": [ + 130, + 365, + 483, + 415 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 365, + 483, + 415 + ], + "type": "text", + "content": " iterates over all object-text pairs in the training batch. Similar to CLIP [75], we use a learnable temperature parameter " + }, + { + "bbox": [ + 130, + 365, + 483, + 415 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 365, + 483, + 415 + ], + "type": "text", + "content": " to facilitate model learning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 432, + 274, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 432, + 274, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 432, + 274, + 445 + ], + "type": "text", + "content": "4.2 Scene-level Grounding" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 452, + 482, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 452, + 482, + 501 + ], + "spans": [ + { + "bbox": [ + 130, + 452, + 482, + 501 + ], + "type": "text", + "content": "With aligned object features, we encode the scene by incorporating object spatial locations into the extracted object features. Specifically, we use a spatial transformer model to encode extracted object features " + }, + { + "bbox": [ + 130, + 452, + 482, + 501 + ], + "type": "inline_equation", + "content": "\\{\\pmb{f}_i^O\\}" + }, + { + "bbox": [ + 130, + 452, + 482, + 501 + ], + "type": "text", + "content": " with their spatial location features " + }, + { + "bbox": [ + 130, + 452, + 482, + 501 + ], + "type": "inline_equation", + "content": "\\{l_i\\}" + }, + { + "bbox": [ + 130, + 452, + 482, + 501 + ], + "type": "text", + "content": " following [18, 109]:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 241, + 508, + 372, + 523 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 508, + 372, + 523 + ], + "spans": [ + { + "bbox": [ + 241, + 508, + 372, + 523 + ], + "type": "interline_equation", + "content": "\\boldsymbol {f} ^ {S} = \\operatorname {S p a t i a l A t t n} (\\{\\boldsymbol {f} _ {i} ^ {O} \\}, \\{\\boldsymbol {l} _ {i} \\})", + "image_path": "3ec6d80ebd5341708a30a9ca879da8009baa32d6d2314046e70dcfd31c96d592.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 531, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 531, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 531, + 482, + 616 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 531, + 482, + 616 + ], + "type": "inline_equation", + "content": "\\{\\pmb{f}_i^S\\}" + }, + { + "bbox": [ + 130, + 531, + 482, + 616 + ], + "type": "text", + "content": " denotes the feature of object " + }, + { + "bbox": [ + 130, + 531, + 482, + 616 + ], + "type": "inline_equation", + "content": "\\pmb{o}_i" + }, + { + "bbox": [ + 130, + 531, + 482, + 616 + ], + "type": "text", + "content": " after encoding with spatial location features. To perform scene-level alignment, we operate on these scene-level object features " + }, + { + "bbox": [ + 130, + 531, + 482, + 616 + ], + "type": "inline_equation", + "content": "\\{\\pmb{f}_i^S\\}" + }, + { + "bbox": [ + 130, + 531, + 482, + 616 + ], + "type": "text", + "content": " and align it with the scene caption " + }, + { + "bbox": [ + 130, + 531, + 482, + 616 + ], + "type": "inline_equation", + "content": "\\pmb{T}^{\\mathrm{scene}}" + }, + { + "bbox": [ + 130, + 531, + 482, + 616 + ], + "type": "text", + "content": ". Specifically, we feed the object features into a projection layer and use max-pooling over all object features to obtain the scene feature " + }, + { + "bbox": [ + 130, + 531, + 482, + 616 + ], + "type": "inline_equation", + "content": "\\pmb{g}^S" + }, + { + "bbox": [ + 130, + 531, + 482, + 616 + ], + "type": "text", + "content": ". Similar to object-level grounding, we pass the scene caption through a tunable language model to obtain text feature " + }, + { + "bbox": [ + 130, + 531, + 482, + 616 + ], + "type": "inline_equation", + "content": "\\pmb{g}^T" + }, + { + "bbox": [ + 130, + 531, + 482, + 616 + ], + "type": "text", + "content": " and perform scene-level contrastive alignment through:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 622, + 482, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 622, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 141, + 622, + 482, + 666 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s c e n e}} = - \\frac {1}{2} \\sum_ {(p, q)} \\left(\\log \\frac {\\exp \\left(D ^ {\\text {s c e n e}} (p , q)\\right)}{\\sum_ {r} \\exp \\left(D ^ {\\text {s c e n e}} (p , r)\\right)} + \\log \\frac {\\exp \\left(D ^ {\\text {s c e n e}} (p , q)\\right)}{\\sum_ {r} \\exp \\left(D ^ {\\text {s c e n e}} (r , q)\\right)}\\right), \\tag {2}", + "image_path": "1f0f65b4b98b4d661021e5dca13575129c7db1a2e940dc5b49e226aea4a38a7b.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 270, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 270, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 270, + 101 + ], + "type": "text", + "content": "B. Jia and Y. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 154 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 154 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 154 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 115, + 482, + 154 + ], + "type": "inline_equation", + "content": "D^{\\mathrm{scene}}(p,q) = (\\pmb{g}_p^S\\pmb{g}_q^T / \\tau)" + }, + { + "bbox": [ + 130, + 115, + 482, + 154 + ], + "type": "text", + "content": " denotes the dot product between scene feature " + }, + { + "bbox": [ + 130, + 115, + 482, + 154 + ], + "type": "inline_equation", + "content": "\\pmb{g}_p^S" + }, + { + "bbox": [ + 130, + 115, + 482, + 154 + ], + "type": "text", + "content": " and scene caption feature " + }, + { + "bbox": [ + 130, + 115, + 482, + 154 + ], + "type": "inline_equation", + "content": "\\pmb{g}_q^T" + }, + { + "bbox": [ + 130, + 115, + 482, + 154 + ], + "type": "text", + "content": " for each pair of aligned scene-text pairs in the training batch and " + }, + { + "bbox": [ + 130, + 115, + 482, + 154 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 115, + 482, + 154 + ], + "type": "text", + "content": " iterates over all scene-text pairs in the training batch." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 167, + 322, + 180 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 167, + 322, + 180 + ], + "spans": [ + { + "bbox": [ + 132, + 167, + 322, + 180 + ], + "type": "text", + "content": "4.3 Referral-object-level Grounding" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 482, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 482, + 292 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 482, + 292 + ], + "type": "text", + "content": "To model the relationships revealed in referring expressions, we employ a self-attention-based reasoning transformer for grounding object referrals in scenes. This transformer takes in scene-object features " + }, + { + "bbox": [ + 130, + 183, + 482, + 292 + ], + "type": "inline_equation", + "content": "\\{\\pmb{f}_i^S\\}" + }, + { + "bbox": [ + 130, + 183, + 482, + 292 + ], + "type": "text", + "content": " and an object referral " + }, + { + "bbox": [ + 130, + 183, + 482, + 292 + ], + "type": "inline_equation", + "content": "T^{\\mathrm{ref}}" + }, + { + "bbox": [ + 130, + 183, + 482, + 292 + ], + "type": "text", + "content": " and performs self-attention to learn relationships between text descriptions and object relationships. We use the same tunable language encoder as in scene-level grounding for extracting per-object referral features. We pass this text feature together with scene-object features into the self-attention transformer to obtain the aligned object features " + }, + { + "bbox": [ + 130, + 183, + 482, + 292 + ], + "type": "inline_equation", + "content": "\\pmb{h}_i^S" + }, + { + "bbox": [ + 130, + 183, + 482, + 292 + ], + "type": "text", + "content": " and the sentence-level referral feature " + }, + { + "bbox": [ + 130, + 183, + 482, + 292 + ], + "type": "inline_equation", + "content": "\\pmb{h}^T" + }, + { + "bbox": [ + 130, + 183, + 482, + 292 + ], + "type": "text", + "content": ". We then perform the referral-object-level contrastive alignment following:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 239, + 296, + 482, + 326 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 296, + 482, + 326 + ], + "spans": [ + { + "bbox": [ + 239, + 296, + 482, + 326 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {r e f}} = - \\log \\frac {\\exp \\left(\\bar {\\boldsymbol {h}} ^ {S} \\boldsymbol {h} ^ {T} / \\tau\\right)}{\\sum_ {p} \\exp \\left(\\boldsymbol {h} _ {p} ^ {S} \\boldsymbol {h} ^ {T} / \\tau\\right)}, \\tag {3}", + "image_path": "fdfdf5f053e4de58e5e8b15bc10aa2f56c42fd46a2e907d1f98488e9ef69ecd8.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 329, + 482, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 329, + 482, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 329, + 482, + 402 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 329, + 482, + 402 + ], + "type": "inline_equation", + "content": "\\bar{h}^S" + }, + { + "bbox": [ + 130, + 329, + 482, + 402 + ], + "type": "text", + "content": " denotes the feature of the referred object, " + }, + { + "bbox": [ + 130, + 329, + 482, + 402 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 329, + 482, + 402 + ], + "type": "text", + "content": " iterates over all objects within the same scene. Notably, in contrast to inter-scene contrast that was done in object- and scene-level alignment, we force the selection of positive pairs to be within the same scene to provide intra-scene contrast for fine-grained object grounding. This mimics the success of intra-image and inter-image contrasts commonly used for region-word alignment in 2D-VL models [101]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 402, + 483, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 483, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 483, + 486 + ], + "type": "text", + "content": "To learn the multi-level alignment between 3D scenes and language, we first train the point cloud encoder with an object-level grounding objective to obtain a good feature initialization for grounding objects in scenes. During the scene grounding stage, we train our inter- and intra-scene objectives together with a masked language modeling loss " + }, + { + "bbox": [ + 130, + 402, + 483, + 486 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{MLM}}" + }, + { + "bbox": [ + 130, + 402, + 483, + 486 + ], + "type": "text", + "content": " over the inputted object-referral texts to tune the parameters within the language encoder and self-attention transformer. Above all, the learning of GPS could be summarized as optimizing:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 233, + 491, + 382, + 504 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 491, + 382, + 504 + ], + "spans": [ + { + "bbox": [ + 233, + 491, + 382, + 504 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\mathrm {o b j}} + \\mathcal {L} _ {\\mathrm {s c e n e}} + \\mathcal {L} _ {\\mathrm {r e f}} + \\mathcal {L} _ {\\mathrm {M L M}}.", + "image_path": "efbd3c21248437044d1eb76bc5ebdb4ecc2ce100b9e1f65fab99176944c3cf45.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 518, + 230, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 518, + 230, + 532 + ], + "spans": [ + { + "bbox": [ + 132, + 518, + 230, + 532 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 540, + 482, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 540, + 482, + 552 + ], + "spans": [ + { + "bbox": [ + 132, + 540, + 482, + 552 + ], + "type": "text", + "content": "In this section, we present experimental results addressing the following questions:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 135, + 555, + 482, + 625 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 135, + 555, + 482, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 555, + 482, + 590 + ], + "spans": [ + { + "bbox": [ + 135, + 555, + 482, + 590 + ], + "type": "text", + "content": "1. How effective is the data scaling in ScENEVERSE for 3D visual grounding? Does the scale-up benefit common 3D-VL tasks (e.g., 3D question answering, open-vocabulary 3D semantic segmentation) and pre-training-based models?" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 135, + 590, + 482, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 590, + 482, + 613 + ], + "spans": [ + { + "bbox": [ + 135, + 590, + 482, + 613 + ], + "type": "text", + "content": "2. How well is the GPS pre-training pipeline for 3D-VL tasks? Does it exhibit similar properties of 2D-VL models in 3D-VL tasks?" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 135, + 613, + 434, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 613, + 434, + 625 + ], + "spans": [ + { + "bbox": [ + 135, + 613, + 434, + 625 + ], + "type": "text", + "content": "3. What is offered by ScENEVERSE and GPS and what is missing?" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "In the following sections, we describe in detail the model performance regarding these key topics. Due to the page limit, we direct readers to the supplementary for implementation details, qualitative results, and more experimental analyses." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 148, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 148, + 91, + 448, + 103 + ], + "type": "text", + "content": "Scaling 3D Vision-Language Learning for Grounded Scene Understanding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 148, + 480, + 317 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 2: 3D visual grounding results on Nr3D, Sr3D, and ScanRefer. We use \"pre-train\" for our model trained on ScENeVERse w/o additional fine-tuning, and \"fine-tune\" for its data-specific fine-tuned version. Best results are highlighted in bold." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 148, + 480, + 317 + ], + "lines": [ + { + "bbox": [ + 133, + 148, + 480, + 317 + ], + "spans": [ + { + "bbox": [ + 133, + 148, + 480, + 317 + ], + "type": "table", + "html": "
MethodNr3DSr3DScanRefer Acc@0.5
OverallEasyHardV-Dep.V-Indep.OverallEasyHardV-Dep.V-Indep.OverallUniqueMultiple
3DVG-Trans [105]40.848.534.834.843.751.454.244.944.651.734.760.628.4
TGNN [42]37.344.230.635.838.045.048.536.945.845.029.756.823.2
TransRefer3D [37]48.056.739.642.550.757.460.550.249.957.7---
InstanceRefer [100]38.846.031.834.541.948.051.140.545.848.132.966.824.7
FFL-3DOG [32]41.748.235.037.144.7-----34.067.925.7
LAR [6]48.958.442.347.452.159.463.051.250.059.1---
SAT [97]56.564.948.454.457.657.961.250.049.258.330.150.825.2
3D-SPS [62]51.558.145.148.053.262.656.265.449.263.237.066.729.8
3DJCG [12]----------37.364.330.8
BUTD-DETR [45]54.660.748.446.058.067.068.663.253.067.639.866.335.1
MVT [43]59.567.452.759.160.364.566.958.858.464.733.366.525.3
ViL3DRel [18]64.470.257.462.064.572.874.967.963.873.237.768.630.7
EDA [92]52.158.246.150.253.168.170.362.954.168.742.368.637.6
3D-VisTA (scratch) [109]57.565.949.453.759.469.672.163.657.970.141.570.934.8
3D-VisTA [109]64.272.156.761.565.176.478.871.358.977.345.875.139.1
Ours (scratch)58.767.050.955.859.868.470.563.453.169.040.471.334.7
Ours (pre-train)55.262.848.045.558.874.176.468.554.175.047.177.441.6
Ours (fine-tuned)64.972.557.856.967.977.580.171.662.878.248.177.942.7
", + "image_path": "3eade42b36a095ddeb66dc8843922cdae95f3df7bc5b8a0601714842be8aa2b3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 335, + 269, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 335, + 269, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 335, + 269, + 346 + ], + "type": "text", + "content": "5.1 3D Visual Grounding" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 369, + 482, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 482, + 478 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 482, + 478 + ], + "type": "text", + "content": "Settings We evaluate our model on three commonly-used datasets for 3D visual grounding: ScanRefer [16], Nr3D, and Sr3D [1]. For Nr3D and Sr3D, we follow Achlioptas et al. [1] and report the grounding accuracies of models using ground-truth object masks. For ScanRefer, we follow Zhu et al. [109] and use Mask3D [80] to generate object proposals. Results are reported as Acc@0.5 to evaluate the correctness of predictions whose object bounding boxes overlap the ground truth with IoU " + }, + { + "bbox": [ + 130, + 369, + 482, + 478 + ], + "type": "inline_equation", + "content": ">0.5" + }, + { + "bbox": [ + 130, + 369, + 482, + 478 + ], + "type": "text", + "content": ". For comparisons, we compare with existing baselines by providing the results of pre-trained GPS and dataset-specific fine-tuned GPS. Please see more details in the supplementary." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 498, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 665 + ], + "type": "text", + "content": "Results and Analyses As shown in Tab. 2, GPS trained on SCENEVERSE achieves state-of-the-art results on all existing 3D-VL grounding benchmarks. Initially, when GPS is trained directly on the training sets of benchmark datasets, labeled as Ours (scratch), it underperforms compared to existing models that employ more complex structures or loss designs. This result underscores the data-intensive nature of the contrastive alignment paradigm. However, when presented with extensive training data in SCENEVERSE, the results of our model without additional fine-tuning, i.e., Ours (pre-train), significantly improves and already achieves state-of-the-art results on benchmarks like ScanRefer. Moreover, the dataset-specific fine-tuned model, i.e., Ours (fine-tuned), consistently outperforms existing baselines with only a simple projection MLP added on top of the pretrained model, jointly optimized during fine-tuning without any other auxiliary architecture or loss objective. These results underscore the strong potential of both the SCENEVERSE and GPS for 3D-VL tasks." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 269, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 269, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 269, + 101 + ], + "type": "text", + "content": "B. Jia and Y. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 137, + 309, + 209 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 310, + 135 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 310, + 135 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 310, + 135 + ], + "type": "text", + "content": "Table 3: Zero-shot transfer on existing benchmarks. \"SR\" stands for ScanRefer." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 137, + 309, + 209 + ], + "lines": [ + { + "bbox": [ + 134, + 137, + 309, + 209 + ], + "spans": [ + { + "bbox": [ + 134, + 137, + 309, + 209 + ], + "type": "table", + "html": "
MethodNr3DSr3DSR@0.25SR@0.5
3D-VisTA (scratch)57.569.645.941.5
3D-VisTA (zero-shot)35.231.233.229.6
3D-VisTA (zero-shot text)43.136.141.136.4
Ours (scratch)58.768.444.540.4
Ours (zero-shot)32.433.335.231.1
Ours (zero-shot text)41.938.140.735.8
", + "image_path": "02bd606ef257629e070b59138bca83c42b7b8d6f74874532d021ea6e0f12d5af.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 319, + 148, + 480, + 208 + ], + "blocks": [ + { + "bbox": [ + 317, + 114, + 482, + 147 + ], + "lines": [ + { + "bbox": [ + 317, + 114, + 482, + 147 + ], + "spans": [ + { + "bbox": [ + 317, + 114, + 482, + 147 + ], + "type": "text", + "content": "Table 4: Zero-shot transfer on ScENeVERSE-val. Evaluation uses GT object proposals following Nr3D/Sr3D." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 148, + 480, + 208 + ], + "lines": [ + { + "bbox": [ + 319, + 148, + 480, + 208 + ], + "spans": [ + { + "bbox": [ + 319, + 148, + 480, + 208 + ], + "type": "table", + "html": "
MethodOverallEasyHardV-Dep.V-Indep.
3D-VisTA (scratch)40.753.121.637.344.3
3D-VisTA (zero-shot)52.959.635.453.752.2
3D-VisTA (zero-shot text)58.170.039.652.564.1
Ours (scratch)38.550.220.833.743.9
Ours (zero-shot)59.269.444.053.166.3
Ours (zero-shot text)60.670.945.154.867.3
", + "image_path": "91857b136dd51597b4d4df6ef6d5e5f85a3a97076424c4002583d9fa5b69d03f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 224, + 256, + 235 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 256, + 235 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 256, + 235 + ], + "type": "text", + "content": "5.2 Zero-Shot Transfer" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 240, + 482, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 240, + 482, + 409 + ], + "spans": [ + { + "bbox": [ + 130, + 240, + 482, + 409 + ], + "type": "text", + "content": "Settings To better evaluate the effectiveness of both the SCENEVERSE data and the GPS model, we further perform zero-shot transfer experiments to test the models' capability in 4 benchmarks, ScanRefer, Sr3D, Nr3D, and SCENEVERSE- val. We create SCENEVERSE-val using " + }, + { + "bbox": [ + 130, + 240, + 482, + 409 + ], + "type": "inline_equation", + "content": "8.5K" + }, + { + "bbox": [ + 130, + 240, + 482, + 409 + ], + "type": "text", + "content": " annotated object referrals of 271 scenes in MultiScan, and randomly split the scenes following a 4:1 train / test split for creating the held-out test set. We mainly consider 2 specific transfer settings in our experiments: (i) zero-shot: models trained by removing all the scenes from the target dataset, tested on held-out unseen scenes, and (ii) zero-shot text: Models trained on data that include the training set of scenes from the target dataset, yet tested exclusively with unseen scene-text distribution. Specifically, for the zero-shot text setting, we use the generated texts in SCENEVERSE as fine-tuning sources for the zero-shot model. We mainly compare our model against a recent pre-training-based model 3D-VisTA. See more details on experimental setting and implementation in the supplementary." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 412, + 481, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 412, + 481, + 436 + ], + "spans": [ + { + "bbox": [ + 130, + 412, + 481, + 436 + ], + "type": "text", + "content": "Results and Analyses We present the results of zero-shot transfer experiments in Tab. 3 and Tab. 4 with the following key observations:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 437, + 482, + 664 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 138, + 437, + 482, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 437, + 482, + 507 + ], + "spans": [ + { + "bbox": [ + 138, + 437, + 482, + 507 + ], + "type": "text", + "content": "- Our GPS model demonstrates superior generalization to unseen scenes compared to the 3D-VisTA model. In zero-shot transfer scenarios, our model consistently outperforms 3D-VisTA across established benchmarks and ScEN-EVERSE-val. This indicates the effectiveness of contrastive alignment over traditional classification objectives, aligning with the advancements seen in 2D-VL models for open-vocabulary grounding and transfer capabilities" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 139, + 508, + 482, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 508, + 482, + 603 + ], + "spans": [ + { + "bbox": [ + 139, + 508, + 482, + 603 + ], + "type": "text", + "content": "- SCENEVERSE dataset substantially enhances 3D-VL grounding capabilities through zero-shot transfer, especially when provided with relatively limited training data, i.e., SCENEVERSE-val. As demonstrated in Tab. 4, there is a significantly improved performance when comparing models trained on SCENEVERSE in a zero-shot manner to those trained from scratch. This indicates that SCENEVERSE can effectively capture knowledge for general 3D scene grounding. Consequently, this underscores its potential as a go-to pre-training dataset for 3D-VL tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 139, + 604, + 482, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 604, + 482, + 664 + ], + "spans": [ + { + "bbox": [ + 139, + 604, + 482, + 664 + ], + "type": "text", + "content": "- The impact of our extensive collection and scalable generation of scene-text pairs is further evidenced by the results in the zero-shot text setting. Notably, as shown in Tab. 3, these automatically generated scene-text pairs supply ample knowledge for comprehending the scene distribution. This contributes significantly to the substantial improvement over the zero-shot performance." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "type": "text", + "content": "Scaling 3D Vision-Language Learning for Grounded Scene Understanding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 158, + 296, + 243 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 477, + 156 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 477, + 156 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 477, + 156 + ], + "type": "text", + "content": "Table 5: 3D question answering re- Table 6: Existing 3D backbones pre-sults on ScanQA and SQA3D. We re-trained on SCENEVERSE for openport EM@1 score on ScanQA and SQA3D vocabulary 3D semantic segmentation on evaluation sets. ScanNet. \"SPUNet\" denotes SparseUNet pro" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 158, + 296, + 243 + ], + "lines": [ + { + "bbox": [ + 134, + 158, + 296, + 243 + ], + "spans": [ + { + "bbox": [ + 134, + 158, + 296, + 243 + ], + "type": "table", + "html": "
ModelScanQASQA3D
valw/objw/o obj
ScanRefer+MCAN [5]18.620.619.0-
ScanQA [5]20.323.520.946.6
SQA3D [65]---47.2
3D-VisTA [109]22.427.023.048.5
3D-LLM [40]20.519.1--
Ours22.725.023.549.9
", + "image_path": "c0887a789106b61f253d012ada1bbec8c39cd2a220e84ac7eda6443b80b81066.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 299, + 170, + 474, + 246 + ], + "blocks": [ + { + "bbox": [ + 299, + 170, + 474, + 246 + ], + "lines": [ + { + "bbox": [ + 299, + 170, + 474, + 246 + ], + "spans": [ + { + "bbox": [ + 299, + 170, + 474, + 246 + ], + "type": "table", + "html": "
ModelNetworkmIoUΔmAccΔ
OpenScene [74]SPUNet1657.2-69.9-
PLA [30]SPUNet1617.7-33.5-
RegionPLC [94]SPUNet1656.9-75.6-
RegionPLC+SCENEVERSESPUNet1658.2+1.7%77.3+2.2%
OpenScene [74]SPUNet3257.8-70.3-
PLA [30]SPUNet3219.1-41.5-
RegionPLC [94]SPUNet3259.6-77.5-
RegionPLC+SCENEVERSESPUNet3261.0+2.3%79.7+2.8%
", + "image_path": "15fad004489a8bad2b2000fb494cc9b76d144cee90a07e36f80042b8b8505a7e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 253, + 284, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 253, + 284, + 266 + ], + "spans": [ + { + "bbox": [ + 132, + 253, + 284, + 266 + ], + "type": "text", + "content": "5.3 Additional 3D-VL Tasks" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 270, + 481, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 481, + 306 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 481, + 306 + ], + "type": "text", + "content": "Settings We evaluate the effectiveness of GPS and ScENEVERSE on additional 3D-VL tasks: (i) 3D question answering (3D-QA) on ScanQA [5] and SQA3D [65], and (ii) open-vocabulary 3D semantic segmentation (OV-Seg) on ScanNet." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 307, + 482, + 426 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 138, + 307, + 481, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 307, + 481, + 354 + ], + "spans": [ + { + "bbox": [ + 138, + 307, + 481, + 354 + ], + "type": "text", + "content": "- In the 3D-QA task, we follow Zhu et al. [109] and evaluate models over the exact match metric (EM@1) on the validation and test sets of ScanQA, as well as the test set of SQA3D. We pre-train GPS on SCENEVERSE and fine-tune the model on the 3D-QA dataset to compare with state-of-the-art models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 139, + 354, + 482, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 354, + 482, + 426 + ], + "spans": [ + { + "bbox": [ + 139, + 354, + 482, + 426 + ], + "type": "text", + "content": "- In the OV-Seg task, as GPS builds upon an object-centric design and thus is not directly applicable to semantic segmentation, we consider testing the effectiveness of ScENEVERSE on improving existing 3D models. Specifically, we follow the open-vocabulary semantic segmentation settings proposed by Yang et al. [94] and report the mIoU and mAcc score. We compare with existing works by pre-training the RegionPLC [94] model on ScENEVERSE." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 430, + 481, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 481, + 453 + ], + "type": "text", + "content": "Results and Analyses We present the results of 3D-QA experiments in Tab. 5 and the results of OV-Seg experiments in Tab. 6. The analyses are as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 454, + 481, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 454, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 138, + 454, + 481, + 525 + ], + "type": "text", + "content": "- As shown in Tab. 5, our model achieves state-of-the-art results on both benchmarks, outperforming recent strong pre-training-based baselines like 3D-VisTA and 3D-LLM. As SCENEVERSE currently contains only descriptions of objects and scenes, we believe involving more types of language descriptions (e.g., question-answer pairs, dialogues) is a promising direction for further improving model performance on these downstream tasks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 525, + 482, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 525, + 482, + 609 + ], + "spans": [ + { + "bbox": [ + 138, + 525, + 482, + 609 + ], + "type": "text", + "content": "- As shown in Tab. 6, we observe consistent performance improvement of existing 3D backbone models on this task when pre-trained with SCENEVERSE data. This result validates that the collected data in SCENEVERSE can effectively boost the performance of existing models on scene understanding tasks. We further provide results of state-of-the-art 3D models that are pretrained on SCENEVERSE on the close-vocabulary 3D semantic segmentation task in the supplementary." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 624, + 323, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 624, + 323, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 624, + 323, + 635 + ], + "type": "text", + "content": "5.4 Ablative Studies and Discussion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "In this section, we present further discussions on both the data collected in ScENEVERSE and the GPS model design. We aim to elucidate the effects of" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 270, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 270, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 270, + 101 + ], + "type": "text", + "content": "B. Jia and Y. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 222, + 204 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 222, + 204 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 222, + 204 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 222, + 204 + ], + "type": "image", + "image_path": "5e2ab110ec62d74690b13d152711fe2ec4c7bf3cd4386d0ab8a3b7417fa5deef.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 205, + 315, + 259 + ], + "lines": [ + { + "bbox": [ + 132, + 205, + 315, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 315, + 259 + ], + "type": "text", + "content": "Fig.4: Model performance v.s. data scale. Plots show that models consistently improve in both the pre-train and zero-shot transfer settings on ScanRefer and ScENEV-ERSE-val with data scaling-up." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 225, + 114, + 312, + 204 + ], + "blocks": [ + { + "bbox": [ + 225, + 114, + 312, + 204 + ], + "lines": [ + { + "bbox": [ + 225, + 114, + 312, + 204 + ], + "spans": [ + { + "bbox": [ + 225, + 114, + 312, + 204 + ], + "type": "image", + "image_path": "1e80d2da8c1b18b851b85d99876ecdb44af4e1bc358562e8a8c10d2567e0a5bb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 136, + 297, + 312, + 344 + ], + "blocks": [ + { + "bbox": [ + 132, + 263, + 315, + 296 + ], + "lines": [ + { + "bbox": [ + 132, + 263, + 315, + 296 + ], + "spans": [ + { + "bbox": [ + 132, + 263, + 315, + 296 + ], + "type": "text", + "content": "Table 8: Cross domain transfer results of models pre-trained on real and synthetic datasets. \"S3D\" stands for Structured3D." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 297, + 312, + 344 + ], + "lines": [ + { + "bbox": [ + 136, + 297, + 312, + 344 + ], + "spans": [ + { + "bbox": [ + 136, + 297, + 312, + 344 + ], + "type": "table", + "html": "
RealSyntheticSCENEVERSE-valS3DProcTHOR
AllX64.837.143.4
XS3D7.085.116.1
XProcTHOR4.216.391.0
", + "image_path": "461ec40e932f7f612a07ba18779ede6fc5e2dca0ed2b0623accc48279ddb41ba.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 322, + 163, + 478, + 223 + ], + "blocks": [ + { + "bbox": [ + 318, + 117, + 482, + 160 + ], + "lines": [ + { + "bbox": [ + 318, + 117, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 318, + 117, + 482, + 160 + ], + "type": "text", + "content": "Table 7: Ablation on text data source used in model pre-training. All models are tested on ScanRefer with no additional finetuning." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 322, + 163, + 478, + 223 + ], + "lines": [ + { + "bbox": [ + 322, + 163, + 478, + 223 + ], + "spans": [ + { + "bbox": [ + 322, + 163, + 478, + 223 + ], + "type": "table", + "html": "
TemplateLLMAnno.Acc@0.25Acc@0.5
XXX43.538.4
✓XX50.946.1
✓✓X51.146.3
✓✓✓52.047.1
", + "image_path": "1f68ae3786918d128019ced644d711e634b8ad8188c1caa40124b6aaab1d1518.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 323, + 286, + 478, + 344 + ], + "blocks": [ + { + "bbox": [ + 318, + 230, + 482, + 285 + ], + "lines": [ + { + "bbox": [ + 318, + 230, + 482, + 285 + ], + "spans": [ + { + "bbox": [ + 318, + 230, + 482, + 285 + ], + "type": "text", + "content": "Table 9: Ablation on model design on ScENeVERSE-val. We use \"Obj- lvl\", \"Scene-lvl\" to denote object and scene alignment loss, and \"MLM\" for the mask language modeling loss." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 323, + 286, + 478, + 344 + ], + "lines": [ + { + "bbox": [ + 323, + 286, + 478, + 344 + ], + "spans": [ + { + "bbox": [ + 323, + 286, + 478, + 344 + ], + "type": "table", + "html": "
Obj-lvl.MLMScene-lvl.OverallEasyHard
XXX64.875.448.7
✓XX65.277.147.4
✓✓X62.473.445.8
✓✓✓66.977.850.3
", + "image_path": "2f2286d2c16cc8ce226e10668ed7802c9f376b4e5e3e2577c47acce49768dec1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 361, + 482, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 482, + 398 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 482, + 398 + ], + "type": "text", + "content": "data scaling and show more clearly its effectiveness in 3D scene understanding. Regarding the experimental settings and more results discussion, refer to the supplementary. The following points are specifically discussed in this section:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 403, + 482, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 499 + ], + "type": "text", + "content": "How important is data-scaling? We conduct ablation studies over the amount of data used while pre-training GPS. We consider the model trained with " + }, + { + "bbox": [ + 130, + 403, + 482, + 499 + ], + "type": "inline_equation", + "content": "\\frac{1}{8}" + }, + { + "bbox": [ + 130, + 403, + 482, + 499 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 403, + 482, + 499 + ], + "type": "inline_equation", + "content": "\\frac{1}{4}" + }, + { + "bbox": [ + 130, + 403, + 482, + 499 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 403, + 482, + 499 + ], + "type": "inline_equation", + "content": "\\frac{1}{2}" + }, + { + "bbox": [ + 130, + 403, + 482, + 499 + ], + "type": "text", + "content": " of ScENEVERSE to show the effectiveness of data-scaling on model performance in the pre-train and zero-shot transfer settings in ScanRefer and ScENEVERSE-val. As shown in Fig. 4, we observe consistent performance improvement over the increase of data scale for both settings. We provide additional experiments in the supplementary to show that such scaling effect is not only beneficial for 3D-VL grounding but also for other 3D tasks like semantic segmentation [80,96]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 504, + 482, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 504, + 482, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 504, + 482, + 635 + ], + "type": "text", + "content": "How is the generated data compared with human-annotated data? We assess the performance of models trained using various scene-text sources, specifically focusing on their performance in the ScanRefer dataset without additional fine-tuning. As shown in Tab. 7, models trained with our template-based generated texts and Large Language Model (LLM)-refined texts show significant improvements over models trained solely on ScanRefer. More importantly, these variants of our model already achieve state-of-the-art results compared with previous baselines. This indicates the effectiveness of our text-generation pipeline. Finally, we observe that adding human-annotated data is still beneficial for model performance. However, the improvement is relatively marginal over models trained on our generated data." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "content": "What is the role of the synthetic scenes in this scale-up process? With synthetic data providing large-scale and diverse scene data for 3D-VL tasks," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "type": "text", + "content": "Scaling 3D Vision-Language Learning for Grounded Scene Understanding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 283 + ], + "type": "text", + "content": "we evaluate the models' domain transfer (Sim2Real) capability. Specifically, we compare models trained on all real scenes in SCENEVERSE against models trained exclusively on two synthetic subsets of SCENEVERSE, i.e., Structured3D and ProcTHOR. As shown in Tab. 8, models trained on synthetic subsets demonstrate remarkable performance on their corresponding test sets while suffering when transferred to real or other synthetic scenes. In contrast, the model trained on real scene-text pairs exhibits less performance drop when generalizing to synthetic scenes. This result affirms the domain gap between real and synthetic scenes in 3D-VL grounding and shows that a simple scale-up in the number of scenes is insufficient when naturalness can not be guaranteed. Considering the scalability of our language generation pipeline and the scaling effect shown in our experiments, the rate-determining step for further scaling-up 3D-VL comes to the collection of diverse, high-quality, and realistic scenes that capture natural 3D scene distributions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 293, + 483, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 483, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 483, + 461 + ], + "type": "text", + "content": "How important is the design of each module in GPS? We provide ablative analyses of our multi-level contrastive alignment design in Tab. 9. We mainly consider removing objectives in our model to reveal the effectiveness of each level of alignment. We choose the referral-object-level alignment objective as the default setting and consider removing: (i) object-level alignment objective, (ii) masked language modeling objective, and (iii) scene-level alignment objective. When removing the object-level alignment objective, we learn the object point cloud encoder with the referral-object-level alignment and without pre-training. As shown in Tab. 9, we test different models on the SCENEVERSE-val without additional fine-tuning. Results show that the scene-level alignment objective is crucial for referral object grounding in SCENEVERSE-val with the " + }, + { + "bbox": [ + 130, + 293, + 483, + 461 + ], + "type": "inline_equation", + "content": "\\sim 5\\%" + }, + { + "bbox": [ + 130, + 293, + 483, + 461 + ], + "type": "text", + "content": " performance drop. Similar observations could be made for the model trained without object-level alignment (" + }, + { + "bbox": [ + 130, + 293, + 483, + 461 + ], + "type": "inline_equation", + "content": "\\sim 2\\%" + }, + { + "bbox": [ + 130, + 293, + 483, + 461 + ], + "type": "text", + "content": " drop) and masked language modeling objective (" + }, + { + "bbox": [ + 130, + 293, + 483, + 461 + ], + "type": "inline_equation", + "content": "\\sim 1.5\\%" + }, + { + "bbox": [ + 130, + 293, + 483, + 461 + ], + "type": "text", + "content": " drop). These results affirm the effectiveness of our model design." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 482, + 220, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 482, + 220, + 495 + ], + "spans": [ + { + "bbox": [ + 132, + 482, + 220, + 495 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "In this work, we scale up 3D-VL for grounded scene understanding. We present SCENEVERSE, a million-scale 3D-VL dataset covering various scenes and multi-level scene descriptions sourced from both human annotation and our proposed scene-text generation approach. Utilizing SCENEVERSE, we propose Grounded Pre-training for Scenes (GPS), a model trained with multi-level scene-language contrastive alignment. Through extensive experiments, we show that GPS achieves state-of-the-art results on common 3D-VL tasks including grounding and question answering. We further conduct zero-shot transfer experiments to show the improved generalization performances of GPS trained on SCENEVERSE compared with previous baselines. We also demonstrate that the scaling effect of SCENEVERSE is generally beneficial for existing 3D models on 3D-VL tasks like semantic segmentation. We hope our efforts and successful scale-up attempts in SCENEVERSE could pave the way for new research paradigms in 3D-VL." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 269, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 269, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 269, + 101 + ], + "type": "text", + "content": "B. Jia and Y. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": "Acknowledgement The authors would like to thank Yaowei Zhang (BIGAI) for his help on online visualization and other colleagues from BIGAI General Vision Lab for fruitful discussions. The authors would also like to thank the anonymous reviewers for their constructive feedback." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 180, + 197, + 192 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 180, + 197, + 192 + ], + "spans": [ + { + "bbox": [ + 132, + 180, + 197, + 192 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 201, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 141, + 201, + 481, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 201, + 481, + 234 + ], + "spans": [ + { + "bbox": [ + 141, + 201, + 481, + 234 + ], + "type": "text", + "content": "1. Achlioptas, P., Abdelreheem, A., Xia, F., Elhoseiny, M., Guibas, L.: Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In: Proceedings of European Conference on Computer Vision (ECCV) (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 141, + 235, + 481, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 235, + 481, + 267 + ], + "spans": [ + { + "bbox": [ + 141, + 235, + 481, + 267 + ], + "type": "text", + "content": "2. Agia, C., Jatavallabhula, K.M., Khodeir, M., Miksik, O., Vineet, V., Mukadam, M., Paull, L., Shkurti, F.: Taskography: Evaluating robot task planning over large 3d scene graphs. In: Proceedings of Conference on Robot Learning (CoRL) (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 267, + 481, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 267, + 481, + 310 + ], + "spans": [ + { + "bbox": [ + 141, + 267, + 481, + 310 + ], + "type": "text", + "content": "3. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 310, + 481, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 310, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 141, + 310, + 481, + 342 + ], + "type": "text", + "content": "4. Armeni, I., He, Z.Y., Gwak, J., Zamir, A.R., Fischer, M., Malik, J., Savarese, S.: 3d scene graph: A structure for unified semantics, 3d space, and camera. In: Proceedings of International Conference on Computer Vision (ICCV) (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 342, + 481, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 342, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 141, + 342, + 481, + 374 + ], + "type": "text", + "content": "5. Azuma, D., Miyanishi, T., Kurita, S., Kawanabe, M.: Scanqa: 3d question answering for spatial scene understanding. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 374, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 374, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 141, + 374, + 481, + 407 + ], + "type": "text", + "content": "6. Bakr, E., Alsaedy, Y., Elhoseiny, M.: Look around and refer: 2d synthetic semantics knowledge distillation for 3d visual grounding. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 407, + 481, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 407, + 481, + 429 + ], + "spans": [ + { + "bbox": [ + 141, + 407, + 481, + 429 + ], + "type": "text", + "content": "7. Barsalou, L.W.: Perceptual symbol systems. Behavioral and brain sciences 22(4), 577-660 (1999)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 429, + 481, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 429, + 481, + 438 + ], + "spans": [ + { + "bbox": [ + 141, + 429, + 481, + 438 + ], + "type": "text", + "content": "8. Barsalou, L.W.: Grounded cognition. Annu. Rev. Psychol. 59, 617-645 (2008)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 438, + 481, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 438, + 481, + 493 + ], + "spans": [ + { + "bbox": [ + 141, + 438, + 481, + 493 + ], + "type": "text", + "content": "9. Baruch, G., Chen, Z., Dehghan, A., Dimry, T., Feigin, Y., Fu, P., Gebauer, T., Joffe, B., Kurz, D., Schwartz, A., et al.: Arkitscenes: A diverse real-world dataset for 3d indoor scene understanding using mobile rgb-d data. In: Proceedings of Advances in Neural Information Processing Systems Datasets and Benchmarks (NeurIPS Datasets and Benchmarks Track) (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 493, + 481, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 493, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 138, + 493, + 481, + 525 + ], + "type": "text", + "content": "10. Bommasani, R., Hudson, D.A., Adeli, E., Altman, R., Arora, S., von Arx, S., Bernstein, M.S., Bohg, J., Bosselut, A., Brunskill, E., et al.: On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 525, + 481, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 525, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 138, + 525, + 481, + 568 + ], + "type": "text", + "content": "11. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 568, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 568, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 138, + 568, + 481, + 601 + ], + "type": "text", + "content": "12. Cai, D., Zhao, L., Zhang, J., Sheng, L., Xu, D.: 3djcg: A unified framework for joint dense captioning and visual grounding on 3d point clouds. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 601, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 481, + 633 + ], + "type": "text", + "content": "13. Chang, A., Dai, A., Funkhouser, T., Halber, M., Niessner, M., Savva, M., Song, S., Zeng, A., Zhang, Y.: Matterport3d: Learning from rgb-d data in indoor environments. Proceedings of International Conference on 3D Vision (3DV) (2017)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "type": "text", + "content": "14. Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., et al.: Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012 (2015)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "type": "text", + "content": "Scaling 3D Vision-Language Learning for Grounded Scene Understanding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 138, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 481, + 149 + ], + "type": "text", + "content": "15. Changpinyo, S., Sharma, P., Ding, N., Soricut, R.: Conceptual 12m: Pushing web-scale image-text pre-training to recognize long-tail visual concepts. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 149, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 149, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 138, + 149, + 481, + 182 + ], + "type": "text", + "content": "16. Chen, D.Z., Chang, A.X., Nießner, M.: Scanrefer: 3d object localization in rgb-d scans using natural language. In: Proceedings of European Conference on Computer Vision (ECCV) (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 182, + 481, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 182, + 481, + 214 + ], + "spans": [ + { + "bbox": [ + 138, + 182, + 481, + 214 + ], + "type": "text", + "content": "17. Chen, D.Z., Wu, Q., Nießner, M., Chang, A.X.: D3net: a speaker-listener architecture for semi-supervised dense captioning and visual grounding in rgb-d scans. In: Proceedings of European Conference on Computer Vision (ECCV) (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 214, + 481, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 214, + 481, + 246 + ], + "spans": [ + { + "bbox": [ + 138, + 214, + 481, + 246 + ], + "type": "text", + "content": "18. Chen, S., Guhur, P.L., Tapaswi, M., Schmid, C., Laptev, I.: Language conditioned spatial relation reasoning for 3d object grounding. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 246, + 481, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 246, + 481, + 278 + ], + "spans": [ + { + "bbox": [ + 138, + 246, + 481, + 278 + ], + "type": "text", + "content": "19. Chen, S., Zhu, H., Chen, X., Lei, Y., Yu, G., Chen, T.: End-to-end 3d dense captioning with vote2cap-detr. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 278, + 481, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 278, + 481, + 321 + ], + "spans": [ + { + "bbox": [ + 138, + 278, + 481, + 321 + ], + "type": "text", + "content": "20. Chen, Y., Huang, S., Yuan, T., Qi, S., Zhu, Y., Zhu, S.C.: Holistic++ scene understanding: Single-view 3d holistic scene parsing and human pose estimation with human-object interaction and physical commonsense. In: Proceedings of International Conference on Computer Vision (ICCV) (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 321, + 481, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 321, + 481, + 354 + ], + "spans": [ + { + "bbox": [ + 138, + 321, + 481, + 354 + ], + "type": "text", + "content": "21. Chen, Y., Li, Q., Kong, D., Kei, Y.L., Zhu, S.C., Gao, T., Zhu, Y., Huang, S.: Yourfit: Embodied reference understanding with language and gesture. In: Proceedings of International Conference on Computer Vision (ICCV) (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 354, + 481, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 354, + 481, + 386 + ], + "spans": [ + { + "bbox": [ + 138, + 354, + 481, + 386 + ], + "type": "text", + "content": "22. Chen, Z., Gholami, A., Nießner, M., Chang, A.X.: Scan2cap: Context-aware dense captioning in rgb-d scans. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 385, + 481, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 385, + 481, + 429 + ], + "spans": [ + { + "bbox": [ + 138, + 385, + 481, + 429 + ], + "type": "text", + "content": "23. Collins, J., Goel, S., Deng, K., Luthra, A., Xu, L., Gundogdu, E., Zhang, X., Vicente, T.F.Y., Dideriksen, T., Arora, H., et al.: Abo: Dataset and benchmarks for real-world 3d object understanding. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 429, + 481, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 429, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 138, + 429, + 481, + 460 + ], + "type": "text", + "content": "24. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 460, + 481, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 460, + 481, + 493 + ], + "spans": [ + { + "bbox": [ + 138, + 460, + 481, + 493 + ], + "type": "text", + "content": "25. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 492, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 492, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 138, + 492, + 481, + 536 + ], + "type": "text", + "content": "26. Deitke, M., Liu, R., Wallingford, M., Ngo, H., Michel, O., Kusupati, A., Fan, A., Laforte, C., Voleti, V., Gadre, S.Y., et al.: Objverse-xl: A universe of " + }, + { + "bbox": [ + 138, + 492, + 481, + 536 + ], + "type": "inline_equation", + "content": "10\\mathrm{m} + 3\\mathrm{d}" + }, + { + "bbox": [ + 138, + 492, + 481, + 536 + ], + "type": "text", + "content": " objects. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 536, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 536, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 536, + 481, + 578 + ], + "type": "text", + "content": "27. Deitke, M., Schwenk, D., Salvador, J., Weihs, L., Michel, O., VanderBilt, E., Schmidt, L., Ehsani, K., Kembhavi, A., Farhadi, A.: Objverse: A universe of annotated 3d objects. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 578, + 481, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 481, + 622 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 481, + 622 + ], + "type": "text", + "content": "28. Deitke, M., VanderBilt, E., Herrasti, A., Weihs, L., Ehsani, K., Salvador, J., Han, W., Kolve, E., Kembhavi, A., Mottaghi, R.: Procthor: Large-scale embodied ai using procedural generation. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 622, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 622, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 622, + 481, + 665 + ], + "type": "text", + "content": "29. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: Proceedings of Conference of the North American Chapter of the Association for Computational Linguistics (NAACL) (2018)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 270, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 270, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 270, + 101 + ], + "type": "text", + "content": "B. Jia and Y. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 137, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 137, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 137, + 116, + 481, + 149 + ], + "type": "text", + "content": "30. Ding, R., Yang, J., Xue, C., Zhang, W., Bai, S., Qi, X.: Pla: Language-driven open-vocabulary 3d scene understanding. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 149, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 149, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 137, + 149, + 481, + 182 + ], + "type": "text", + "content": "31. Ding, Z., Han, X., Niethammer, M.: Volenet: A deep learning label fusion method for multi-atlas segmentation. In: Proceedings of International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI) (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 137, + 182, + 481, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 182, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 137, + 182, + 481, + 224 + ], + "type": "text", + "content": "32. Feng, M., Li, Z., Li, Q., Zhang, L., Zhang, X., Zhu, G., Zhang, H., Wang, Y., Mian, A.: Free-form description guided 3d visual graph network for object grounding in point cloud. In: Proceedings of International Conference on Computer Vision (ICCV) (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 224, + 481, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 224, + 481, + 257 + ], + "spans": [ + { + "bbox": [ + 137, + 224, + 481, + 257 + ], + "type": "text", + "content": "33. Ghiasi, G., Gu, X., Cui, Y., Lin, T.Y.: Scaling open-vocabulary image segmentation with image-level labels. In: Proceedings of European Conference on Computer Vision (ECCV) (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 257, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 257, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 137, + 257, + 481, + 300 + ], + "type": "text", + "content": "34. Gong, R., Huang, J., Zhao, Y., Geng, H., Gao, X., Wu, Q., Ai, W., Zhou, Z., Terzopoulos, D., Zhu, S.C., Jia, B., Huang, S.: Arnold: A benchmark for language-grounded task learning with continuous states in realistic 3d scenes. In: Proceedings of International Conference on Computer Vision (ICCV) (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 300, + 481, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 300, + 481, + 343 + ], + "spans": [ + { + "bbox": [ + 137, + 300, + 481, + 343 + ], + "type": "text", + "content": "35. Gu, Q., Kuwajerwala, A., Morin, S., Jatavallabhula, K.M., Sen, B., Agarwal, A., Rivera, C., Paul, W., Ellis, K., Chellappa, R., et al.: Conceptgraphs: Open-vocabulary 3d scene graphs for perception and planning. arXiv preprint arXiv:2309.16650 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 137, + 343, + 481, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 343, + 481, + 375 + ], + "spans": [ + { + "bbox": [ + 137, + 343, + 481, + 375 + ], + "type": "text", + "content": "36. Ha, H., Song, S.: Semantic abstraction: Open-world 3d scene understanding from 2d vision-language models. In: Proceedings of Conference on Robot Learning (CoRL) (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 137, + 375, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 375, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 137, + 375, + 481, + 407 + ], + "type": "text", + "content": "37. He, D., Zhao, Y., Luo, J., Hui, T., Huang, S., Zhang, A., Liu, S.: Transrefer3d: Entity-and-relation aware transformer for fine-grained 3d visual grounding. In: Proceedings of ACM International Conference on Multimedia (MM) (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 137, + 407, + 481, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 407, + 481, + 440 + ], + "spans": [ + { + "bbox": [ + 137, + 407, + 481, + 440 + ], + "type": "text", + "content": "38. Hegde, D., Valanarasu, J.M.J., Patel, V.: Clip goes 3d: Leveraging prompt tuning for language grounded 3d recognition. In: Proceedings of International Conference on Computer Vision (ICCV) (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 137, + 440, + 481, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 440, + 481, + 472 + ], + "spans": [ + { + "bbox": [ + 137, + 440, + 481, + 472 + ], + "type": "text", + "content": "39. Hong, Y., Wu, Q., Qi, Y., Rodriguez-Opazo, C., Gould, S.: Vln bert: A recurrent vision-and-language bert for navigation. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 137, + 472, + 481, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 472, + 481, + 504 + ], + "spans": [ + { + "bbox": [ + 137, + 472, + 481, + 504 + ], + "type": "text", + "content": "40. Hong, Y., Lin, C., Du, Y., Chen, Z., Tenenbaum, J.B., Gan, C.: 3d concept learning and reasoning from multi-view images. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 137, + 504, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 504, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 137, + 504, + 481, + 536 + ], + "type": "text", + "content": "41. Huang, J., Yong, S., Ma, X., Linghu, X., Li, P., Wang, Y., Li, Q., Zhu, S.C., Jia, B., Huang, S.: An embodied generalist agent in 3d world. arXiv preprint arXiv:2311.12871 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 137, + 536, + 481, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 536, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 137, + 536, + 481, + 568 + ], + "type": "text", + "content": "42. Huang, P.H., Lee, H.H., Chen, H.T., Liu, T.L.: Text-guided graph neural networks for referring 3d instance segmentation. In: Proceedings of AAAI Conference on Artificial Intelligence (AAAI) (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 137, + 568, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 568, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 137, + 568, + 481, + 601 + ], + "type": "text", + "content": "43. Huang, S., Chen, Y., Jia, J., Wang, L.: Multi-view transformer for 3d visual grounding. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 137, + 601, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 601, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 137, + 601, + 481, + 633 + ], + "type": "text", + "content": "44. Huang, S., Wang, Z., Li, P., Jia, B., Liu, T., Zhu, Y., Liang, W., Zhu, S.C.: Diffusion-based generation, optimization, and planning in 3d scenes. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 137, + 633, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 633, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 137, + 633, + 481, + 666 + ], + "type": "text", + "content": "45. Jain, A., Gkanatsios, N., Mediratta, I., Fragkiadaki, K.: Bottom up top down detection transformers for language grounding in images and point clouds. In: Proceedings of European Conference on Computer Vision (ECCV) (2022)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "type": "text", + "content": "Scaling 3D Vision-Language Learning for Grounded Scene Understanding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 480, + 149 + ], + "type": "text", + "content": "46. Jiang, L., Zhao, H., Shi, S., Liu, S., Fu, C.W., Jia, J.: Pointgroup: Dual-set point grouping for 3d instance segmentation. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 149, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 149, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 138, + 149, + 481, + 182 + ], + "type": "text", + "content": "47. Jiang, N., Liu, T., Cao, Z., Cui, J., Zhang, Z., Chen, Y., Wang, H., Zhu, Y., Huang, S.: Full-body articulated human-object interaction. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 182, + 481, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 182, + 481, + 214 + ], + "spans": [ + { + "bbox": [ + 138, + 182, + 481, + 214 + ], + "type": "text", + "content": "48. Jiang, N., Zhang, Z., Li, H., Ma, X., Wang, Z., Chen, Y., Liu, T., Zhu, Y., Huang, S.: Scaling up dynamic human-scene interaction modeling. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2024)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 214, + 481, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 214, + 481, + 246 + ], + "spans": [ + { + "bbox": [ + 138, + 214, + 481, + 246 + ], + "type": "text", + "content": "49. Kaplan, J., McCandlish, S., Henighan, T., Brown, T.B., Chess, B., Child, R., Gray, S., Radford, A., Wu, J., Amodei, D.: Scaling laws for neural language models. arXiv preprint arXiv:2001.08361 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 246, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 246, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 138, + 246, + 481, + 300 + ], + "type": "text", + "content": "50. Khanna, M., Mao, Y., Jiang, H., Haresh, S., Shacklett, B., Batra, D., Clegg, A., Undersander, E., Chang, A.X., Savva, M.: Habitat synthetic scenes dataset (hssd-200): An analysis of 3d scene scale and realism tradeoffs for objectgoal navigation. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2024)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 300, + 481, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 300, + 481, + 332 + ], + "spans": [ + { + "bbox": [ + 138, + 300, + 481, + 332 + ], + "type": "text", + "content": "51. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: Proceedings of International Conference on Computer Vision (ICCV) (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 332, + 481, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 332, + 481, + 376 + ], + "spans": [ + { + "bbox": [ + 138, + 332, + 481, + 376 + ], + "type": "text", + "content": "52. Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., et al.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. In: International Journal of Computer Vision (IJCV) (2017)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 376, + 481, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 376, + 481, + 397 + ], + "spans": [ + { + "bbox": [ + 138, + 376, + 481, + 397 + ], + "type": "text", + "content": "53. Lake, B.M., Ullman, T.D., Tenenbaum, J.B., Gershman, S.J.: Building machines that learn and think like people. Behavioral and brain sciences 40, e253 (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 397, + 481, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 397, + 481, + 429 + ], + "spans": [ + { + "bbox": [ + 138, + 397, + 481, + 429 + ], + "type": "text", + "content": "54. Li, B., Weinberger, K.Q., Belongie, S., Koltun, V., Ranftl, R.: Language-driven semantic segmentation. In: Proceedings of International Conference on Learning Representations (ICLR) (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 429, + 481, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 429, + 481, + 472 + ], + "spans": [ + { + "bbox": [ + 138, + 429, + 481, + 472 + ], + "type": "text", + "content": "55. Li, C., Zhang, R., Wong, J., Gokmen, C., Srivastava, S., Martin-Martín, R., Wang, C., Levine, G., Lingelbach, M., Sun, J., et al.: Behavior-1k: A benchmark for embodied ai with 1,000 everyday activities and realistic simulation. In: Proceedings of Conference on Robot Learning (CoRL) (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 472, + 481, + 505 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 472, + 481, + 505 + ], + "spans": [ + { + "bbox": [ + 138, + 472, + 481, + 505 + ], + "type": "text", + "content": "56. Li, J., Li, D., Savarese, S., Hoi, S.: BLIP-2: bootstrapping language-image pretraining with frozen image encoders and large language models. In: Proceedings of International Conference on Machine Learning (ICML) (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 505, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 505, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 138, + 505, + 481, + 536 + ], + "type": "text", + "content": "57. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: Proceedings of International Conference on Machine Learning (ICML) (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 536, + 481, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 536, + 481, + 579 + ], + "spans": [ + { + "bbox": [ + 138, + 536, + 481, + 579 + ], + "type": "text", + "content": "58. Li, L.H., Zhang, P., Zhang, H., Yang, J., Li, C., Zhong, Y., Wang, L., Yuan, L., Zhang, L., Hwang, J.N., et al.: Grounded language-image pre-training. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 579, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 579, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 138, + 579, + 481, + 601 + ], + "type": "text", + "content": "59. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 601, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 481, + 633 + ], + "type": "text", + "content": "60. Liu, M., Shi, R., Kuang, K., Zhu, Y., Li, X., Han, S., Cai, H., Porikli, F., Su, H.: Openshape: Scaling up 3d shape representation towards open-world understanding. arXiv preprint arXiv:2305.10764 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "type": "text", + "content": "61. Liu, R., Wu, R., Van Hoorick, B., Tokmakov, P., Zakharov, S., Vondrick, C.: Zero-1-to-3: Zero-shot one image to 3d object. In: Proceedings of International Conference on Computer Vision (ICCV) (2023)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 270, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 270, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 270, + 101 + ], + "type": "text", + "content": "B. Jia and Y. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 137, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 137, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 137, + 116, + 482, + 149 + ], + "type": "text", + "content": "62. Luo, J., Fu, J., Kong, X., Gao, C., Ren, H., Shen, H., Xia, H., Liu, S.: 3d-sps: Single-stage 3d visual grounding via referred point progressive selection. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 150, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 150, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 137, + 150, + 481, + 182 + ], + "type": "text", + "content": "63. Luo, T., Rockwell, C., Lee, H., Johnson, J.: Scalable 3d captioning with pretrained models. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 137, + 182, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 182, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 137, + 182, + 481, + 215 + ], + "type": "text", + "content": "64. Ma, C.Y., Lu, J., Wu, Z., AlRegib, G., Kira, Z., Socher, R., Xiong, C.: Self-monitoring navigation agent via auxiliary progress estimation. In: Proceedings of International Conference on Learning Representations (ICLR) (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 216, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 216, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 137, + 216, + 481, + 248 + ], + "type": "text", + "content": "65. Ma, X., Yong, S., Zheng, Z., Li, Q., Liang, Y., Zhu, S.C., Huang, S.: Sqa3d: Situated question answering in 3d scenes. In: Proceedings of International Conference on Learning Representations (ICLR) (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 249, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 249, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 137, + 249, + 481, + 281 + ], + "type": "text", + "content": "66. Mao, Y., Zhang, Y., Jiang, H., Chang, A., Savva, M.: Multiscan: Scalable rgbd scanning for 3d environments with articulated objects. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 281, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 281, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 137, + 281, + 481, + 314 + ], + "type": "text", + "content": "67. Misra, I., Girdhar, R., Joulin, A.: An end-to-end transformer model for 3d object detection. In: Proceedings of International Conference on Computer Vision (ICCV) (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 137, + 315, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 315, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 137, + 315, + 481, + 346 + ], + "type": "text", + "content": "68. Mittal, M., Yu, C., Yu, Q., Liu, J., Rudin, N., Hoeller, D., Yuan, J.L., Singh, R., Guo, Y., Mazhar, H., et al.: Orbit: A unified simulation framework for interactive robot learning environments. Robotics and Automation Letters (RA-L) (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 137, + 347, + 481, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 347, + 481, + 390 + ], + "spans": [ + { + "bbox": [ + 137, + 347, + 481, + 390 + ], + "type": "text", + "content": "69. Mo, K., Zhu, S., Chang, A.X., Yi, L., Tripathi, S., Guibas, L.J., Su, H.: Partnet: A large-scale benchmark for fine-grained and hierarchical part-level 3d object understanding. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 137, + 391, + 481, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 391, + 481, + 434 + ], + "spans": [ + { + "bbox": [ + 137, + 391, + 481, + 434 + ], + "type": "text", + "content": "70. Mu, T., Ling, Z., Xiang, F., Yang, D., Li, X., Tao, S., Huang, Z., Jia, Z., Su, H.: Maniskill: Generalizable manipulation skill benchmark with large-scale demonstrations. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 435, + 454, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 435, + 454, + 445 + ], + "spans": [ + { + "bbox": [ + 138, + 435, + 454, + 445 + ], + "type": "text", + "content": "71. OpenAI: Introducing chatgpt. https://openai.com/blog/chatgpt (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 446, + 448, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 446, + 448, + 456 + ], + "spans": [ + { + "bbox": [ + 138, + 446, + 448, + 456 + ], + "type": "text", + "content": "72. OpenAI: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 457, + 481, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 457, + 481, + 489 + ], + "spans": [ + { + "bbox": [ + 138, + 457, + 481, + 489 + ], + "type": "text", + "content": "73. Pashevich, A., Schmid, C., Sun, C.: Episodic transformer for vision-and-language navigation. In: Proceedings of International Conference on Computer Vision (ICCV) (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 490, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 490, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 138, + 490, + 481, + 522 + ], + "type": "text", + "content": "74. Peng, S., Genova, K., Jiang, C., Tagliasacchi, A., Pollefeys, M., Funkhouser, T., et al.: Openscene: 3d scene understanding with open vocabularies. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 522, + 481, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 522, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 138, + 522, + 481, + 566 + ], + "type": "text", + "content": "75. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: Proceedings of International Conference on Machine Learning (ICML) (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 567, + 481, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 481, + 620 + ], + "type": "text", + "content": "76. Ramakrishnan, S.K., Gokaslan, A., Wijmans, E., Maksymets, O., Clegg, A., Turner, J., Undersander, E., Galuba, W., Westbury, A., Chang, A.X., et al.: Habitatmatterport 3d dataset (hm3d): 1000 large-scale 3d environments for embodied ai. In: Proceedings of Advances in Neural Information Processing Systems Datasets and Benchmarks (NeurIPS Datasets and Benchmarks Track) (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 621, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 481, + 665 + ], + "type": "text", + "content": "77. Rana, K., Haviland, J., Garg, S., Abou-Chakra, J., Reid, I., Suenderhauf, N.: Sayplan: Grounding large language models using 3d scene graphs for scalable robot task planning. In: Proceedings of Conference on Robot Learning (CoRL) (2023)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "type": "text", + "content": "Scaling 3D Vision-Language Learning for Grounded Scene Understanding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 138, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 481, + 149 + ], + "type": "text", + "content": "78. Rosinol, A., Violette, A., Abate, M., Hughes, N., Chang, Y., Shi, J., Gupta, A., Carlone, L.: Kimera: From slam to spatial perception with 3d dynamic scene graphs. International Journal of Robotics Research (IJRR) (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 151, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 151, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 138, + 151, + 481, + 194 + ], + "type": "text", + "content": "79. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 195, + 481, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 195, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 138, + 195, + 481, + 228 + ], + "type": "text", + "content": "80. Schult, J., Engelmann, F., Hermans, A., Litany, O., Tang, S., Leibe, B.: Mask3d: Mask transformer for 3d semantic instance segmentation. In: Proceedings of International Conference on Robotics and Automation (ICRA) (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 229, + 481, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 229, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 138, + 229, + 481, + 251 + ], + "type": "text", + "content": "81. Smith, L., Gasser, M.: The development of embodied cognition: Six lessons from babies. Artificial life 11(1-2), 13-29 (2005)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 252, + 481, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 252, + 481, + 284 + ], + "spans": [ + { + "bbox": [ + 138, + 252, + 481, + 284 + ], + "type": "text", + "content": "82. Takmaz, A., Fedele, E., Sumner, R.W., Pollefeys, M., Tombari, F., Engelmann, F.: Openmask3d: Open-vocabulary 3d instance segmentation. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 285, + 481, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 285, + 481, + 318 + ], + "spans": [ + { + "bbox": [ + 138, + 285, + 481, + 318 + ], + "type": "text", + "content": "83. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 319, + 481, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 319, + 481, + 351 + ], + "spans": [ + { + "bbox": [ + 138, + 319, + 481, + 351 + ], + "type": "text", + "content": "84. Vu, T., Kim, K., Luu, T.M., Nguyen, T., Yoo, C.D.: Softgroup for 3d instance segmentation on point clouds. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 353, + 481, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 353, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 138, + 353, + 481, + 385 + ], + "type": "text", + "content": "85. Wald, J., Avetisyan, A., Navab, N., Tombari, F., Nießner, M.: Rio: 3d object instance re-localization in changing indoor environments. In: Proceedings of International Conference on Computer Vision (ICCV) (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 387, + 481, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 387, + 481, + 419 + ], + "spans": [ + { + "bbox": [ + 138, + 387, + 481, + 419 + ], + "type": "text", + "content": "86. Wald, J., Dhamo, H., Navab, N., Tombari, F.: Learning 3d semantic scene graphs from 3d indoor reconstructions. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 420, + 481, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 420, + 481, + 464 + ], + "spans": [ + { + "bbox": [ + 138, + 420, + 481, + 464 + ], + "type": "text", + "content": "87. Wang, T., Mao, X., Zhu, C., Xu, R., Lyu, R., Li, P., Chen, X., Zhang, W., Chen, K., Xue, T., et al.: Embodiedscan: A holistic multi-modal 3d perception suite towards embodied ai. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2024)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 464, + 481, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 464, + 481, + 508 + ], + "spans": [ + { + "bbox": [ + 138, + 464, + 481, + 508 + ], + "type": "text", + "content": "88. Wang, X., Huang, Q., Celikyilmaz, A., Gao, J., Shen, D., Wang, Y.F., Wang, W.Y., Zhang, L.: Reinforced cross-modal matching and self-supervised imitation learning for vision-language navigation. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 509, + 481, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 509, + 481, + 553 + ], + "spans": [ + { + "bbox": [ + 138, + 509, + 481, + 553 + ], + "type": "text", + "content": "89. Wang, Z., Chen, Y., Jia, B., Li, P., Zhang, J., Zhang, J., Liu, T., Zhu, Y., Liang, W., Huang, S.: Move as you say interact as you can: Language-guided human motion generation with scene affordance. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2024)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 554, + 481, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 554, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 138, + 554, + 481, + 586 + ], + "type": "text", + "content": "90. Wang, Z., Chen, Y., Liu, T., Zhu, Y., Liang, W., Huang, S.: Humanise: Language-conditioned human motion generation in 3d scenes. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 587, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 587, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 587, + 481, + 632 + ], + "type": "text", + "content": "91. Wu, T., Zhang, J., Fu, X., Wang, Y., Ren, J., Pan, L., Wu, W., Yang, L., Wang, J., Qian, C., et al.: Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "text", + "content": "92. Wu, Y., Cheng, X., Zhang, R., Cheng, Z., Zhang, J.: Eda: Explicit text-decoupling and dense alignment for 3d visual grounding. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "20" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 270, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 270, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 270, + 101 + ], + "type": "text", + "content": "B. Jia and Y. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 19 + }, + { + "para_blocks": [ + { + "bbox": [ + 134, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 136, + 116, + 482, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 116, + 482, + 161 + ], + "spans": [ + { + "bbox": [ + 136, + 116, + 482, + 161 + ], + "type": "text", + "content": "93. Xue, L., Gao, M., Xing, C., Martin-Martin, R., Wu, J., Xiong, C., Xu, R., Niebles, J.C., Savarese, S.: Ulip: Learning a unified representation of language, images, and point clouds for 3d understanding. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 162, + 481, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 162, + 481, + 195 + ], + "spans": [ + { + "bbox": [ + 138, + 162, + 481, + 195 + ], + "type": "text", + "content": "94. Yang, J., Ding, R., Wang, Z., Qi, X.: Regionplc: Regional point-language contrastive learning for open-world 3d scene understanding. arXiv preprint arXiv:2304.00962 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 196, + 481, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 196, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 138, + 196, + 481, + 228 + ], + "type": "text", + "content": "95. Yang, Y., Jia, B., Zhi, P., Huang, S.: Physcene: Physically interactable 3d scene synthesis for embodied ai. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2024)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 229, + 481, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 229, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 138, + 229, + 481, + 262 + ], + "type": "text", + "content": "96. Yang, Y.Q., Guo, Y.X., Xiong, J.Y., Liu, Y., Pan, H., Wang, P.S., Tong, X., Guo, B.: Swin3d: A pretrained transformer backbone for 3d indoor scene understanding. arXiv preprint arXiv:2304.06906 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 263, + 481, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 263, + 481, + 296 + ], + "spans": [ + { + "bbox": [ + 138, + 263, + 481, + 296 + ], + "type": "text", + "content": "97. Yang, Z., Zhang, S., Wang, L., Luo, J.: Sat: 2d semantics assisted training for 3d visual grounding. In: Proceedings of International Conference on Computer Vision (ICCV) (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 297, + 481, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 297, + 481, + 329 + ], + "spans": [ + { + "bbox": [ + 138, + 297, + 481, + 329 + ], + "type": "text", + "content": "98. Yeshwanth, C., Liu, Y.C., Nießner, M., Dai, A.: Scannet++: A high-fidelity dataset of 3d indoor scenes. In: Proceedings of International Conference on Computer Vision (ICCV) (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 331, + 481, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 331, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 138, + 331, + 481, + 373 + ], + "type": "text", + "content": "99. Yuan, Z., Yan, X., Liao, Y., Guo, Y., Li, G., Cui, S., Li, Z.: X-trans2cap: Cross-modal knowledge transfer using transformer for 3d dense captioning. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 134, + 375, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 375, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 134, + 375, + 481, + 418 + ], + "type": "text", + "content": "100. Yuan, Z., Yan, X., Liao, Y., Zhang, R., Wang, S., Li, Z., Cui, S.: Instancerefer: Cooperative holistic understanding for visual grounding on point clouds through instance multi-level contextual referring. In: Proceedings of International Conference on Computer Vision (ICCV) (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 134, + 420, + 481, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 420, + 481, + 463 + ], + "spans": [ + { + "bbox": [ + 134, + 420, + 481, + 463 + ], + "type": "text", + "content": "101. Zhang, H., Zhang, P., Hu, X., Chen, Y.C., Li, L., Dai, X., Wang, L., Yuan, L., Hwang, J.N., Gao, J.: Glipv2: Unifying localization and vision-language understanding. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 135, + 464, + 481, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 464, + 481, + 497 + ], + "spans": [ + { + "bbox": [ + 135, + 464, + 481, + 497 + ], + "type": "text", + "content": "102. Zhang, R., Guo, Z., Zhang, W., Li, K., Miao, X., Cui, B., Qiao, Y., Gao, P., Li, H.: Pointclip: Point cloud understanding by clip. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 134, + 498, + 481, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 498, + 481, + 531 + ], + "spans": [ + { + "bbox": [ + 134, + 498, + 481, + 531 + ], + "type": "text", + "content": "103. Zhang, R., Wang, L., Qiao, Y., Gao, P., Li, H.: Learning 3d representations from 2d pre-trained models via image-to-point masked autoencoders. In: Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 532, + 481, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 532, + 481, + 564 + ], + "spans": [ + { + "bbox": [ + 134, + 532, + 481, + 564 + ], + "type": "text", + "content": "104. Zhang, Y., Gong, Z., Chang, A.X.: Multi3drefer: Grounding text description to multiple 3d objects. In: Proceedings of International Conference on Computer Vision (ICCV) (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 566, + 481, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 566, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 134, + 566, + 481, + 597 + ], + "type": "text", + "content": "105. Zhao, L., Cai, D., Sheng, L., Xu, D.: 3dvg-transformer: Relation modeling for visual grounding on point clouds. In: Proceedings of International Conference on Computer Vision (ICCV) (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 599, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 599, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 599, + 481, + 632 + ], + "type": "text", + "content": "106. Zheng, J., Zhang, J., Li, J., Tang, R., Gao, S., Zhou, Z.: Structured3d: A large photo-realistic dataset for structured 3d modeling. In: Proceedings of European Conference on Computer Vision (ECCV) (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 633, + 481, + 665 + ], + "type": "text", + "content": "107. Zhu, W., Hessel, J., Awadalla, A., Gadre, S.Y., Dodge, J., Fang, A., Yu, Y., Schmidt, L., Wang, W.Y., Choi, Y.: Multimodal c4: An open, billion-scale corpus of images interleaved with text. arXiv preprint arXiv:2304.06939 (2023)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "type": "text", + "content": "Scaling 3D Vision-Language Learning for Grounded Scene Understanding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "21" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 20 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 216 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "108. Zhu, Y., Gao, T., Fan, L., Huang, S., Edmonds, M., Liu, H., Gao, F., Zhang, C., Qi, S., Wu, Y.N., et al.: Dark, beyond deep: A paradigm shift to cognitive ai with humanlike common sense. Engineering 6(3), 310-345 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 182 + ], + "type": "text", + "content": "109. Zhu, Z., Ma, X., Chen, Y., Deng, Z., Huang, S., Li, Q.: 3d-vista: Pre-trained transformer for 3d vision and text alignment. In: Proceedings of International Conference on Computer Vision (ICCV) (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 183, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 183, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 133, + 183, + 481, + 216 + ], + "type": "text", + "content": "110. Zhu, Z., Zhang, Z., Ma, X., Niu, X., Chen, Y., Jia, B., Deng, Z., Huang, S., Li, Q.: Unifying 3d vision-language understanding via promptable queries. In: Proceedings of European Conference on Computer Vision (ECCV) (2024)" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "22" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 269, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 269, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 269, + 100 + ], + "type": "text", + "content": "B. Jia and Y. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 21 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Scissorhands_ Scrub Data Influence via Connection Sensitivity in Networks/43a51c48-5759-4f46-b600-93f450455ce7_content_list.json b/2024/Scissorhands_ Scrub Data Influence via Connection Sensitivity in Networks/43a51c48-5759-4f46-b600-93f450455ce7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e4d11444ed0011a8e72bd374f036703492a8a07b --- /dev/null +++ b/2024/Scissorhands_ Scrub Data Influence via Connection Sensitivity in Networks/43a51c48-5759-4f46-b600-93f450455ce7_content_list.json @@ -0,0 +1,1659 @@ +[ + { + "type": "text", + "text": "Scissorhands: Scrub Data Influence via Connection Sensitivity in Networks", + "text_level": 1, + "bbox": [ + 279, + 140, + 725, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jing Wu1, Mehrtash Harandi1", + "bbox": [ + 379, + 210, + 622, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Department of Electrical and Computer Systems Engineering Monash University, Australia {jing.wu1, mehrtash.harandi}@monash.edu", + "bbox": [ + 292, + 239, + 710, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Machine unlearning has become a pivotal task to erase the influence of data from a trained model. It adheres to recent data regulation standards and enhances the privacy and security of machine learning applications. In this work, we present a new machine unlearning approach Scissorhands. Initially, Scissorhands identifies the most pertinent parameters in the given model relative to the forgetting data via connection sensitivity. By reinitializing the most influential top- $k$ percent of these parameters, a trimmed model for erasing the influence of the forgetting data is obtained. Subsequently, Scissorhands fine-tunes the trimmed model with a gradient projection-based approach, seeking parameters that preserve information on the remaining data while discarding information related to the forgetting data. Our experimental results, conducted across image classification and image generation tasks, demonstrate that Scissorhands, showcases competitive performance when compared to existing methods. Source code is available at https://github.com/JingWu321/Scissorhands.", + "bbox": [ + 261, + 311, + 743, + 534 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Machine unlearning $\\cdot$ Connection sensitivity $\\cdot$ Diffusion model", + "bbox": [ + 261, + 546, + 751, + 561 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Warning: This paper contains explicit sexual imagery that may be offensive.", + "bbox": [ + 259, + 579, + 740, + 606 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 630, + 375, + 645 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we aim to propose an effective machine unlearning method. Under data regulations like the European Union General Data Protection Regulation (GDPR) [53] and the California Consumer Privacy Act (CCPA) [22], all users are granted the right to be forgotten. In machine learning, these legal provisions empower data owners with the right not only to withdraw their data from trained models but also to ensure that their data's influence on these models is erased.", + "bbox": [ + 212, + 657, + 787, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The most direct approach to accomplish this objective is to retrain the model from scratch, excluding the data requested for deletion from the training process. Retraining from scratch is typically considered the gold standard in the field of machine unlearning [12, 52]. Yet, this poses a challenge as numerous in-production models require prolonged training periods and substantial computing resources. While retraining is feasible, it is often impractical.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Consequently, the proposal and development of efficient approximate unlearning methods [13, 23, 26, 50, 51] have become essential. Currently, most approximate unlearning techniques achieve forgetting by adding normally distributed noise to the parameters [19-21] or by estimating the influence of a particular data point on the model's predictions [23, 37, 38, 48] based on the influence function method [10]. Jia et al. [27] demonstrate that model sparsity can help to unlearn and fuse model sparsity into the unlearning process.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Fan et al. [12] recently raised concerns regarding the instability in approximate unlearning methods, and highlighted that the current machine unlearning approaches, initially crafted for image classification tasks, fall short in effectively tackling the challenges of machine unlearning within the realm of image generation. The authors introduce a novel unlearning method SalUn that can effectively perform forgetting in both image classification and generation tasks. Their unlearning mechanism includes finding salient weights and then fine-tuning these weights using forgetting data assigned with random labels. However, studies like [59] show that the model can still learn the true data distribution even using data with random labels. While SalUn has shown state-of-the-art performance, it could potentially memorize information about forgetting data points, albeit associating them with random labels but still can be undesired in sensitive applications. In § 4, results also show that SalUn tends to memorize knowledge about the forgetting data.", + "bbox": [ + 212, + 252, + 787, + 464 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This work. We present an unlearning approach, Scissorhands, designed to erase data influence in the classification models and eliminate a particular concept from a text-to-image model. To achieve this, inspired by [27], our key insight is to first erase the critical influence of forgetting data in models and then relearn. Unlike previous methods that employ the forgetting data with noisy labels as a part of relearning, we propose to use an efficient gradient projection method to relearn the critical features and patterns while ensuring the exclusion of influences associated with the forgetting data. Through a series of experiments and evaluations, including the classification on SVHN [39], CIFAR-10 and CIFAR-100 [30], CelebAMASK-HQ [31] datasets, as well as the open-source stable diffusion [44] text-to-image model, results demonstrate the viability and effectiveness of our technique in forgetting the influence of random samples, discrete classes, and sensitive content such as nudity.", + "bbox": [ + 212, + 464, + 787, + 660 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Methodology", + "text_level": 1, + "bbox": [ + 215, + 684, + 380, + 702 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we propose Scissorhands, our unlearning framework that scrubs data from a model by eliminating their influence on model parameters. Throughout the paper, we denote scalars and vectors/matrices by lowercase and bold symbols, respectively (e.g., $a$ , $a$ , and $A$ ). We show a $d$ -dimensional vector of ones by $\\mathbf{1}_d$ , and use $\\odot$ to denote the Hadamard product.", + "bbox": [ + 212, + 717, + 787, + 794 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Overview. Consider a model $f$ with parameters $\\pmb{\\theta} \\in \\mathbb{R}^d$ trained on a dataset $\\mathcal{D} = \\{\\mathbf{x}_i, \\mathbf{y}_i\\}_{i=1}^N$ . Suppose a subset of $\\mathcal{D}$ , denoted as $\\mathcal{D}_f$ , is requested for deletion.", + "bbox": [ + 215, + 809, + 785, + 842 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Jing Wu, Mehrtash Harandi", + "bbox": [ + 271, + 114, + 460, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The remaining data is defined as $\\mathcal{D}_r\\coloneqq \\mathcal{D}\\setminus \\mathcal{D}_f$ . Our objective is to develop the unlearning algorithm to generate a scrubbed model $f_{u}$ , effectively removing the influence of $\\mathcal{D}_f$ from model $f$ , while maintaining its utility over $\\mathcal{D}_r$ .", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In Scissorhands, we initially identify critical parameters w.r.t. the forgetting data via connection sensitivity analysis [32, 33]. We then reinitialize the top- $k$ percent of these key parameters, resulting in a trimmed model with parameters $\\pmb{\\theta}_t$ . This step aims to diminish the model's memory of the forgetting data $\\mathcal{D}_f$ . Nevertheless, this process runs the risk of erasing valuable information from the data set $\\mathcal{D}_r$ that we aim to retain.", + "bbox": [ + 212, + 191, + 784, + 280 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address this, the next phase of Scissorhands concentrates on restoring the performance of the trimmed model on remaining data $\\mathcal{D}_r$ , while ensuring that the influence of the forgetting data $\\mathcal{D}_f$ remains excluded. We accomplish this balance and obtain the final scrubbed model with parameters $\\theta^{\\star}$ through a gradient projection-based approach.", + "bbox": [ + 212, + 282, + 784, + 357 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Trimming", + "text_level": 1, + "bbox": [ + 215, + 378, + 344, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To effectively identify salient parameters of a network with respect to $\\mathcal{D}_f$ , one has to establish a good criterion to determine such salient connections. A commonly used criterion is the magnitude of the weights, with those exceeding a certain threshold deemed salient as suggested in [25]. Our requirement is more nuanced, as we need a measure that specifically discerns saliency based on $\\mathcal{D}_f$ . Therefore, we adopt a single-shot approach as proposed in [33], which we outline briefly to ensure clarity and completeness in our methodology. Consider", + "bbox": [ + 212, + 401, + 782, + 506 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {s} _ {j} (\\mathcal {D}) := \\mathbb {E} _ {\\boldsymbol {x}, \\boldsymbol {y} \\sim \\mathcal {D}} \\left[ \\ell (\\boldsymbol {\\theta}; \\boldsymbol {x}, \\boldsymbol {y}) - \\ell \\left(\\left(\\mathbf {1} _ {d} - \\boldsymbol {e} _ {j}\\right) \\odot \\boldsymbol {\\theta}; \\boldsymbol {x}, \\boldsymbol {y}\\right) \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 316, + 515, + 784, + 540 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "which measures the influence of parameter $j \\in \\{1, \\dots, d\\}$ on a model in terms of the empirical risk for a given dataset $\\mathcal{D}$ . Here, $\\ell : \\mathbb{R}^d \\times \\mathcal{X} \\times \\mathcal{Y} \\to \\mathbb{R}_+$ is the loss of the model and $e_j$ is the indicator vector of element $j$ (i.e., a binary vector with its $j$ -th component set to one). Note that computing $\\mathbf{s}_j$ for each $j \\in \\{1, \\dots, d\\}$ is prohibitively expensive, as it demands $d + 1$ forward passes through the dataset. An approximation to Equation (1) [25] is in the form of", + "bbox": [ + 214, + 547, + 787, + 638 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {s} _ {j} (\\mathcal {D}) \\approx \\mathbb {E} _ {\\boldsymbol {x}, \\boldsymbol {y} \\sim \\mathcal {D}} \\left[ \\frac {\\partial \\ell (\\boldsymbol {\\theta} ; \\boldsymbol {x} , \\boldsymbol {y})}{\\partial \\boldsymbol {\\theta} _ {j}} \\boldsymbol {\\theta} _ {j} \\right]. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 648, + 784, + 680 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Detailed proofs can be found in Appendix A. Equation (2) defines the sensitivity of parameter $j$ as the average (i.e., expectation) of a product, the gradient of the loss and the current value of the parameter. As such, $\\mathbf{s}_j$ encodes the loss sensitivity according to the parameter's magnitude. Intuitively, a parameter with a large value that significantly affects the loss, as captured by the gradient, is considered more influential or salient. The form in Equation (2) offers several benefits, most importantly, the ability to calculate the sensitivity for all parameters with just a single sweep through the dataset.", + "bbox": [ + 212, + 688, + 782, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To scrub the influence of the forgetting data $\\mathcal{D}_f$ in the model, we first obtain $\\mathfrak{s}_j(\\mathcal{D}_f)$ , saliency of parameters w.r.t. the forgetting data. We then re-initialize", + "bbox": [ + 214, + 809, + 782, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Scissorhands: Scrub Data Influence via Connection Sensitivity in Networks", + "bbox": [ + 233, + 114, + 732, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the top- $k\\%$ of the parameters based on their saliency rankings. This process is akin to performing a targeted 'lobotomy' on the model specifically concerning the data $\\mathcal{D}_f$ , thereby selectively erasing their influence. Unfortunately, this aggressive approach of reinitializing parameters can detrimentally impact the model's performance on the remaining dataset $\\mathcal{D}_r$ , which we will correct in the next phase of Scissorhands.", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Remark 1. In principle, Scissorhands can benefit from any algorithms that can identify important connections w.r.t. $\\mathcal{D}_f$ . The main reason behind choosing SNIP is its single-shot nature. As presented in the study [12], the gradient of loss w.r.t. the model parameters over the forgetting data can help identify important parameters w.r.t. $\\mathcal{D}_f$ and hence help with erasing.", + "bbox": [ + 212, + 248, + 787, + 325 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Remark 2. In practice, we have observed that even a small subset of $\\mathcal{D}_f$ can be sufficient for the trimming. For example, to unlearn a large set of data on CIFAR-100, our algorithm outperforms baselines by utilizing approximately $3\\%$ of the forgetting data for trimming.", + "bbox": [ + 212, + 334, + 787, + 397 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Remark 3. Initializing weights of the trimmed neurons can take various forms. Empirically, we observed initializing with a uniform distribution is particularly effective. Details on the influence of different initialization strategies and the choice of $\\mathrm{k}\\%$ trimming on the model's performance will be discussed in $\\S 4.2$ .", + "bbox": [ + 212, + 406, + 787, + 468 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Repairing", + "text_level": 1, + "bbox": [ + 214, + 488, + 344, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following the parameter-trimming process, aimed at mitigating the influence of the forgetting data $\\mathcal{D}_f$ , a challenge presents itself: the potential erasure of crucial information associated with the remaining data $\\mathcal{D}_r$ . A straightforward solution to recover the model utility is to relearn the model using the remaining data. However, this approach risks biasing the model towards features of the remaining data and inadvertently reintegrating information about the forgetting samples. In the quest to ensure that models retain essential information from the remaining data $\\mathcal{D}_r$ while effectively forgetting $\\mathcal{D}_f$ , our strategy involves maximizing the loss over the forgetting data $\\mathcal{D}_f$ while concurrently minimizing the loss over the remaining data $\\mathcal{D}_r$ . Therefore, to achieve the balance between these goals, we propose the objective of an efficient practice for unlearning:", + "bbox": [ + 212, + 513, + 787, + 681 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL (\\boldsymbol {\\theta}, \\mathcal {D} _ {r}, \\mathcal {D} _ {f}) := \\mathcal {L} (\\boldsymbol {\\theta}; \\mathcal {D} _ {r}) - \\lambda \\mathcal {L} (\\boldsymbol {\\theta}; \\mathcal {D} _ {f}). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 364, + 691, + 784, + 709 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We aim to optimize Equation (3) w.r.t. $\\pmb{\\theta}$ through gradient descent, ensuring minimal reintroduction of information about $\\mathcal{D}_f$ . This is particularly challenging when there are similarities between samples in $\\mathcal{D}_f$ and $\\mathcal{D}_r$ . To achieve this desideratum, it is vital that updates to the model do not improve (i.e., reduce) $\\mathcal{L}(\\pmb{\\theta};\\mathcal{D}_f)$ . To this end, we define $\\pmb{g}_f = \\nabla_{\\pmb{\\theta}}\\mathcal{L}(\\pmb{\\theta};\\mathcal{D}_f)$ as the gradient direction for $\\mathcal{D}_f$ and $\\pmb{g}_o = \\nabla_{\\pmb{\\theta}}L(\\pmb{\\theta},\\mathcal{D}_r,\\mathcal{D}_f)$ as the gradient direction for optimizing Equation (3). The optimal descent direction $\\pmb{g} \\in \\mathbb{R}^d$ should thus satisfy two criteria: 1. It should exhibit maximum similarity to $\\pmb{g}_o$ to ensure swift convergence. The", + "bbox": [ + 212, + 719, + 787, + 842 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Jing Wu, Mehrtash Harandi", + "bbox": [ + 271, + 114, + 460, + 128 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 The procedure of Scissorhands." + ], + "code_body": "1: procedure TRIMMING $(\\pmb {\\theta},\\mathcal{D}_f)$ 11: for $e = 0$ to $E - 1$ do \n2: $\\pmb {X},\\pmb {y}\\gets \\{\\pmb {x}_i,\\pmb {y}_i\\sim \\mathcal{D}_f|i\\in [B]\\} .$ 12: Get $L(\\pmb {\\theta},\\mathcal{D}_r,\\mathcal{D}_f)$ (cf. Eq. (3)). \n3: // Compute connection sensitivity 13: Get $\\pmb{g_o}$ and $\\pmb{g_f}$ \n4: Get $\\mathbf{s}_j(\\pmb {\\theta};\\pmb {X},\\pmb {y}),\\forall j\\in [1,d]$ 14: //Get the optimal direction using Eq.(2). 15: if $\\langle g_o,g_f\\rangle >0$ then \n5: // Re-initialization 16: Compute $v^{*}$ (cf. Eq. (5)). \n6: Re-initialize parameters according 17: $\\pmb {g} = \\pmb {g}_o - v^*\\pmb {g}_f$ to the top- $k \\%$ value of $\\mathfrak{s}_j$ 18: $\\pmb {g}_o\\gets \\pmb {g}$ \n7: return $\\pmb{\\theta}_{t}$ 19: end if \n8: end procedure 20: $\\pmb{\\theta}^{e + 1}\\leftarrow \\pmb{\\theta}^{e} - \\eta \\pmb{g}_{o}$ \n21: end for \n9: procedure REPAIRING $(\\pmb {\\theta}_t,\\mathcal{D}_f,\\mathcal{D}_r)$ 22: return $\\pmb{\\theta}^{*} = \\pmb{\\theta}^{E}$ \n10: $\\pmb{\\theta}^{0} = \\pmb{\\theta}_{t}$ 23: end procedure", + "bbox": [ + 215, + 162, + 787, + 377 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "notion of similarity can be captured by $\\pmb{g}_o$ using $\\| \\pmb{g} - \\pmb{g}_o \\|_2^2$ . 2. It should not align with $\\pmb{g}_f$ to prevent improving $\\mathcal{L}(\\pmb{\\theta}; \\mathcal{D}_f)$ . This can be mathematically formalized as $\\langle \\pmb{g}_f, \\pmb{g} \\rangle \\leq 0$ . Consequently, after computing $\\pmb{g}_o = \\nabla_{\\pmb{\\theta}} L(\\pmb{\\theta}, \\mathcal{D}_r, \\mathcal{D}_f)$ and $\\pmb{g}_f = \\nabla_{\\pmb{\\theta}} \\mathcal{L}(\\pmb{\\theta}; \\mathcal{D}_f)$ in each iteration, we formulate the following optimization challenge to identify the model update direction:", + "bbox": [ + 212, + 407, + 787, + 484 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\underset {\\boldsymbol {g}} {\\arg \\min} \\frac {1}{2} \\| \\boldsymbol {g} - \\boldsymbol {g} _ {o} \\| _ {2} ^ {2}, \\\\ \\text {s . t .} \\quad \\langle \\boldsymbol {g} _ {f}, \\boldsymbol {g} \\rangle \\leq 0. \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 416, + 494, + 785, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This problem may be addressed using the Frank-Wolfe algorithm. Given the high dimensionality of gradients in neural networks, directly addressing the constraint optimization in Equation (4) could become overwhelming. Inspired by [36], we propose to make use of the dual formulation of the problem Equation (4) as:", + "bbox": [ + 212, + 558, + 787, + 621 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\sup _ {v} \\frac {1}{2} \\boldsymbol {g} _ {f} ^ {\\top} \\boldsymbol {g} _ {f} v ^ {2} - \\boldsymbol {g} _ {o} ^ {\\top} \\boldsymbol {g} _ {f} v, \\\\ \\begin{array}{l l} \\text {s . t .} & v \\geq 0. \\end{array} \\tag {5} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 408, + 643, + 785, + 694 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, the optimal descent direction $\\pmb{g}$ is given from the solution $v^{*}$ in Equation (5) as $\\pmb {g} = \\pmb{g}_{o} - v^{*}\\pmb{g}_{f}$ . Detailed proofs can be found in Appendix A. By doing so, we ensure a balance between unlearning and retaining utility.", + "bbox": [ + 212, + 705, + 787, + 752 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Remark 4. The primal problem (cf. Equation (4)) involves optimizing over the high-dimensional space of gradient vectors, while the dual problem (cf. Equation (5)) simplifies this problem to optimize over a single scalar variable $v \\in \\mathbb{R}$ . This dimensionality reduction can significantly decrease the computational complexity, especially for large-scale neural networks.", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Scissorhands: Scrub Data Influence via Connection Sensitivity in Networks", + "bbox": [ + 233, + 113, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2.3 Algorithm Description", + "text_level": 1, + "bbox": [ + 215, + 146, + 447, + 162 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Algorithm 1 describes the procedure of Scissorhands in detail. We first identify the important parameters w.r.t. the forgetting data in single-shot via connection sensitivity in networks, then re-initialize these parameters for erasing the influence of the forgetting data $\\mathcal{D}_f$ in the given model. To repair the performance of the trimmed model on the remaining data $\\mathcal{D}_r$ , while ensuring the exclusion of information associated with the forgetting data $\\mathcal{D}_f$ , we employ the gradient projection-based optimization algorithm. By doing so, we can scrub the influence of the forgetting data while retaining model utility on the remaining data. The loss depicted in Equation (3) can be viewed as a scalarization of a Multi-Objective Optimization (MOO) problem, namely minimizing $\\left(\\mathcal{L}(\\theta; \\mathcal{D}_r), -\\mathcal{L}(\\theta; \\mathcal{D}_f)\\right)^{\\top}$ . A common issue in MOO is the gradient conflict, as optimizing one objective could hinder another one. Our proposal on gradient projection is designed to address and mitigate the gradient conflict. Empirically (see Table 3), we studied the effect of removing gradient projection.", + "bbox": [ + 212, + 167, + 787, + 383 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3 Related Work", + "text_level": 1, + "bbox": [ + 215, + 402, + 388, + 419 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The development of efficient machine unlearning methods [2, 3, 11, 17, 18, 20, 23, 27, 28, 37, 40, 42, 45, 48, 51, 58, 61] has gained prominence. Applications of machine unlearning span various domains, including regression tasks [51], federated learning [24, 34, 35, 54, 56], graph neural networks [5, 7], and diffusion models [12, 15, 16, 26, 57, 60, 62], as well as scenarios where training data are not available [8, 50].", + "bbox": [ + 212, + 431, + 785, + 521 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Retraining the model from scratch without forgetting data is typically considered the standard gold unlearning algorithm [12, 52]. However, this is often deemed impractical as most in-production models require extensive training periods and considerable computing resources. While fine-tuning the model for a new task might lead to forgetting previous knowledge (i.e., catastrophic forgetting) [36], it fails to adequately erase the data influence in the models. Approximate unlearning methods, as such, become attractive alternatives. We will briefly discuss these methods in image classification and image generation.", + "bbox": [ + 212, + 523, + 787, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Unlearning in Classification Models. Most unlearning algorithms are based on the influence function [23,38,48] and the Fisher Information Matrix (FIM) [19-21]. Influence function-based methods estimate the influence of the particular training data points on the model's predictions, and Fisher unlearning methods assume that the unlearned model and the retrained model are close to each other, Golatkar et al. [20] leverage FIM and hide the difference between the unlearned models and retrained models via adding noise. Golatkar et al. [20] also present an upper bound for the amount of retained information, offering a quantifiable measure of the effectiveness of the unlearning process. Such approximate unlearning methods need to compute the Hessian matrix, or FIM w.r.t. the data. This may render them impractical for certain scenarios, such as federated learning, where computational resources are distributed and limited. To mitigate these computational demands, Mehta et al. [37] introduce L-CODEC, a strategy aimed at", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Jing Wu, Mehrtash Harandi", + "bbox": [ + 271, + 114, + 459, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "pre-selecting a subset of parameters, allowing for more efficient computation. Jia et al. [27] recently propose to fuse the model sparsity into the unlearning algorithm, helping improve approximate unlearning methods, and achieving effective and efficient unlearning effects.", + "bbox": [ + 212, + 146, + 784, + 205 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Concept Erasure in Generative Models. The advent of generative models, particularly those converting text to images, has been a significant milestone in the field of artificial intelligence. A notable concern is the risk of it being tainted or manipulated [6,43], resulting in not-safe-for-work (NSFW) generations [46], as these models leverage training data from a wide array of open sources. To address these challenges, data censoring [1, 14, 41, 47] to exclude black-listed images is employed. Studies [15, 46] introduce methods to update the models away from the inappropriate concepts. Heng and Soh [26] recently propose to adopt Elastic Weight Consolidation (EWC) and Generative Replay (GR) to effectively unlearn without access to the training data for a wide range of generative models.", + "bbox": [ + 212, + 207, + 785, + 357 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Unlearning across Domains. Recently, Fan et al. [12] highlight the difficulty in cross-domain applicability of the machine unlearning algorithms, i.e., existing machine unlearning methods designed for classification tasks cannot be effective when applied to image generation. To address this gap, they introduce a novel method, Saliency unlearning (SalUn), which shifts attention from the entire model to target parameters. SalUn achieves state-of-the-art performance and is effective in both image classification and image generation tasks.", + "bbox": [ + 212, + 359, + 785, + 463 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this work, we propose Scissorhands which could achieve improved performance compared to existing methods across various scenarios. Inspired by [27] that apply $\\ell_1$ regularization to simply fuse sparsity into the unlearning methods, we identify and re-initialize the key model parameters w.r.t. the forgetting data via connection sensitivity. Concurrently, Fan et al. [12] also explore unlearning via salience scores, determining weight salience via the gradient of the forgetting loss w.r.t. model parameters, and employing random labeling for model finetuning. While there are similarities in our overarching goals, the methodologies present unique perspectives and solutions to the problem of unlearning. Fan et al. [12] only focus on the salient weights, while our mechanism aims to first 'destroy' the key information and then relearn for the whole model. Additionally, using the forgetting data with random labels could potentially lead to memorizing information about the forgetting data points [59]. In contrast, Scissorhands excludes knowledge about the forgetting data via a gradient projection-based approach, resulting in the effective erasure of information about the forgetting data while preserving the knowledge about the retained data.", + "bbox": [ + 212, + 465, + 787, + 705 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Experimental Evaluation", + "text_level": 1, + "bbox": [ + 215, + 731, + 493, + 748 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we illustrate how Scissorhands effectively eliminates the data influence in the models. For sample-wise forgetting, where the forgetting data shares the same distribution as the training data, we evaluate Scissorhands on SVHN [39], CIFAR-10 and CIFAR-100 [30]. We further extend our evaluation on CelebAMask-HQ [31] where targeting to remove the entirety of specific identities.", + "bbox": [ + 212, + 763, + 785, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Scissorhands: Scrub Data Influence via Connection Sensitivity in Networks", + "bbox": [ + 233, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Additionally, our experimentation encompasses the open-source text-to-image model, Stable Diffusion v1.4 [44], which is conditioned on CLIP text embeddings through the cross-attention mechanism. Further detailed experimental setups and additional results can be found in Appendix B.", + "bbox": [ + 212, + 146, + 787, + 207 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Baselines. We primarily compare against the following standard baselines that are frequently employed in machine unlearning, alongside recently proposed SOTA methods: (i) Retrain: models obtained by only retraining from scratch on $\\mathcal{D}_r$ . This will provide the performance of the oracle for us. (ii) Fine-tuning (FT) [55]: models that are fine-tuned on $\\mathcal{D}_r$ , i.e., taking advantage from catastrophic forgetting in neural networks to unlearn. (iii) Gradient ascent (GA) [52]: gradient ascent on $\\mathcal{D}_f$ . This will provide a simple way of unlearning by making the performance of the model worse on $\\mathcal{D}_f$ . (iv) Influence unlearning (IU) [29]: utilizes the influence function method to estimate the change in model parameters when transitioning from the unscrubbed model to the retrained model. (v) Boundary shrink (BS) [4] and (vi) Boundary expanding (BE) [4]: shift the decision boundary of the original model to imitate that of the retrained model. (vii) $\\ell_1$ -sparse [27]: fine-tuning models on $\\mathcal{D}_r$ with $\\ell_1$ -norm sparse regularization. (viii) Saliency unlearning (SalUn) [12]: adopt weight saliency and random labeling for unlearning.", + "bbox": [ + 212, + 212, + 795, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Metrics. To assess the unlearning algorithms, we employ several metrics: (i) Accuracy: the accuracy of the model on $\\mathcal{D}_f$ (denoted as $\\mathbf{Acc}_{\\mathcal{D}_f}$ ), $\\mathcal{D}_r$ (denoted as $\\mathbf{Acc}_{\\mathcal{D}_r}$ ) and the test set (denoted as $\\mathbf{Acc}_{\\mathcal{D}_t}$ ). It provides insight into how well the model performs after undergoing the unlearning process. (ii) Membership inference attack (MIA): a standard metric for verifying the unlearning effect. A classifier $\\phi$ is trained using both training data (marked with a label of 1, indicating data points included in training) and test data (marked with a label of 0, indicating data points not seen during training), and subsequently assessed on $\\mathcal{D}_f$ . An effective unlearning method should make it challenging to infer whether a particular sample was part of the training set, mitigating potential privacy breaches. In this context, we define MIA as $P_{\\phi}(\\pmb{y} = 0|\\pmb{x}_f)$ , representing the probability assigned by the classifier $\\phi$ that a given sample $\\pmb{x}_f$ was not included in the training data. (iii) Avg. Gap [12]: the average of the performance gaps measured in accuracy-related metrics: Avg. Gap $= (|\\mathrm{Acc}_{\\mathcal{D}_t} - \\mathrm{Acc}_{\\mathcal{D}_t}^*| + |\\mathrm{Acc}_{\\mathcal{D}_f} - \\mathrm{Acc}_{\\mathcal{D}_f}^*| + |\\mathrm{Acc}_{\\mathcal{D}_r} - \\mathrm{Acc}_{\\mathcal{D}_r}^*| + |\\mathrm{MIA} - \\mathrm{MIA}^*|)/4$ , where $\\mathrm{Acc}_{\\mathcal{D}_t}^*$ , $\\mathrm{Acc}_{\\mathcal{D}_f}^*$ , $\\mathrm{Acc}_{\\mathcal{D}_r}^*$ and $\\mathrm{MIA}^*$ are metric values of the retrained model. A better performance of an unlearning method corresponds to a lower performance gap with retraining, measuring how close the unlearned model is to the retrained model. (iv) Run-time efficiency (RTE) [27]: evaluate the computational efficiency of the unlearning method. Specifically, we have $\\mathrm{RTE} = \\mathrm{T} / \\mathrm{T}_r$ , where $\\mathrm{T}$ and $\\mathrm{T}_r$ present the time for the unlearning methods to get the scrubbed models and the time for retraining from scratch with $\\mathcal{D}_r$ , respectively. An efficient unlearning method should aim for minimal computational overhead compared to the baseline. (v) Relearn time [8,20,51]: the number of epochs to fine-tune the scrubbed models for regaining performance on $\\mathcal{D}_f$ (i.e., reach the original models' accuracy on $\\mathcal{D}_f$ ).", + "bbox": [ + 212, + 446, + 789, + 843 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Jing Wu, Mehrtash Harandi", + "bbox": [ + 271, + 114, + 460, + 128 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/632b4b1f7d52b5d5b2e1911a3b33d77c170d99be741ff15af519f19cd7224e30.jpg", + "table_caption": [ + "Table 1: Quantitative results for forgetting ${10}\\%$ data." + ], + "table_footnote": [], + "table_body": "
MethodAccDf(↓)AccDt(↑)AccDr(↑)MIA(↑)Avg. Gap
CIFAR-100Retrain75.13±0.8574.69±0.0899.98±0.0150.22±0.62-
FT [55]97.98±1.3675.28±0.1299.95±0.029.64±3.616.01
GA [52]98.00±1.3475.59±0.1198.24±1.165.00±2.2517.68
IU [29]95.67±4.8272.13±4.5896.14±4.519.43±5.9816.93
BE [4]97.94±1.3874.16±0.0998.12±1.247.60±3.0516.96
BS [4]97.65±1.4873.20±0.1897.93±1.308.24±3.2317.01
l1-sparse [27]96.35±0.6770.06±0.4696.35±0.6721.33±1.9514.59
SalUn [12]88.56±1.1871.34±0.4899.40±0.3574.66±2.4810.45
Ours68.76±1.8173.17±0.2499.24±0.3042.42±2.064.11
CIFAR-10Retrain94.81±0.5394.26±0.14100.0±0.0013.05±0.64-
FT [55]99.15±0.4693.83±0.4599.84±0.113.01±0.933.74
GA [52]99.66±0.2394.57±0.0199.62±0.250.91±0.294.42
IU [29]98.08±2.1091.91±2.7398.01±2.264.01±3.444.16
BE [4]99.41±0.3893.79±0.1599.41±0.3816.16±0.782.19
BS [4]99.60±0.2594.24±0.0799.56±0.544.46±0.333.46
l1-sparse [27]94.17±0.4990.64±0.5296.64±0.5411.87±0.612.20
SalUn [12]98.07±0.4293.92±0.2599.89±0.0717.93±0.372.15
Ours95.40±1.4892.92±0.4898.93±0.579.56±2.131.62
SVHNRetrain91.81±1.1191.17±1.7797.73±1.1215.74±1.28-
FT [55]99.60±0.2495.28±0.0499.99±0.003.85±0.46.51
GA [52]98.41±0.2392.87±0.0698.52±0.295.96±0.384.72
IU [29]92.47±1.6286.92±2.0693.36±1.8217.00±2.672.63
BE [4]98.48±0.3392.62±0.0598.44±0.296.77±0.364.45
BS [4]98.29±0.3292.48±0.0298.36±0.296.74±0.394.36
l1-sparse [27]97.79±0.1093.59±3.1899.49±0.237.57±0.294.65
SalUn [12]96.29±4.8093.59±3.1897.41±4.9624.47±4.083.99
Ours91.62±1.1893.05±0.5699.51±0.1417.62±2.541.43
", + "bbox": [ + 235, + 170, + 767, + 604 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Results on Classification Task", + "text_level": 1, + "bbox": [ + 215, + 633, + 508, + 647 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this experiment on SVHN, CIFAR-10, and CIFAR-100, we try to forget randomly selected $10\\%$ of the training data; on CelebAMask-HQ, we attempt to forget randomly selected $10\\%$ identities among 307 identities. We further apply GradCAM [49] to visualize regions where models focus on $\\mathbf{w}/$ and $\\mathbf{w} / \\mathbf{o}$ machine unlearning algorithms. In brief, the results suggest that Scissorhands has successfully induced forgetting for the relevant samples and classes, with minor degradation in model performance over the remaining data and classes. We will discuss the results in depth below.", + "bbox": [ + 212, + 657, + 787, + 779 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Sample-wise unlearning. Table 1 presents the results when forgetting randomly selected samples. Scissorhands achieves the lowest average performance gap on all the presented datasets and shows good generalization ability compared to", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Scissorhands: Scrub Data Influence via Connection Sensitivity in Networks", + "bbox": [ + 233, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/1020f8174793d2d1f244284ff718413d2df0ef903a2abd6edb5301a59b5e3a0a.jpg", + "table_caption": [ + "Table 2: Quantitative results for forgetting ${10}\\%$ identities on the CelebAMask-HQ." + ], + "table_footnote": [], + "table_body": "
MethodAccDf(↓)AccDt(↑)AccDr(↑)MIA(↑)Avg. Gap
Retrain0.00±0.0087.02±0.8099.96±0.01100.0±0.00-
FT [55]99.94±0.1288.59±0.5999.97±7.025.28±2.0349.06
GA [52]87.60±8.7181.22±2.1199.74±0.2651.37±5.9635.56
IU [29]88.92±10.2570.24±11.7795.27±5.0729.59±18.5945.20
BE [4]69.07±2.7344.11±2.0895.58±1.2346.24±5.9042.53
BS [4]98.18±1.9281.92±0.2799.86±0.0345.93±5.1139.36
l1-sparse [27]98.81±0.7289.37±0.7099.97±0.0076.78±5.6631.10
SalUn [12]0.00±0.0078.36±1.3496.90±1.11100.0±0.002.93
Ours1.52±2.7380.18±6.6097.20±3.8199.83±0.352.82
", + "bbox": [ + 254, + 171, + 748, + 333 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/a785f5a123cef880a8214a26247f8d55b6be1f71058b3f264d6dd46c0648a217.jpg", + "image_caption": [ + "Fig. 1: Visualizations of regions where models focus on generated by GradCAM [49]." + ], + "image_footnote": [], + "bbox": [ + 259, + 349, + 488, + 445 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/44b0c00ac887072f2a3e599bc8d2653d331a7a141fb1fbc3f2eb13b8495116f3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 514, + 349, + 743, + 445 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Retrain. For example, on SVHN, Scissorhands achieves $\\sim 91.62\\%$ accuracy on the forgetting data $\\mathcal{D}_f$ , closely matching the $91.81\\%$ accuracy of a fully retrained model. Furthermore, the retrained model reaches around $91\\%$ accuracy on the test dataset, Scissorhands exhibits an accuracy of roughly $93\\%$ , showcasing the enhanced generalization of our scrubbed model.", + "bbox": [ + 212, + 503, + 785, + 580 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "When evaluating solely based on forgetting accuracy on CIFAR-10, the $\\ell_1$ -sparse method might seem to be the most effective baseline, achieving approximately $94.17\\%$ . Nonetheless, this perceived advantage is offset by a decline in both test accuracy and the accuracy of retained data. In contrast, Scissorhands shows a good trade-off between preserving model utility on the remaining data, generalization on unseen sets, and erasing the influence of the forgetting data.", + "bbox": [ + 212, + 580, + 787, + 672 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Similarly, on CIFAR-100, SalUn may initially stand out with its impressive MIA accuracy of approximately $74.66\\%$ , relying solely on a single metric to gauge the performance of machine unlearning can be misleading, as it might not fully capture the method's effectiveness [12]. While SalUn achieves a forgetting accuracy of about $88.56\\%$ and a test accuracy of $71.34\\%$ , Scissorhands demonstrates a slightly lower forgetting accuracy of around $68.76\\%$ but achieves a higher test accuracy of $73.17\\%$ . This indicates that Scissorhands achieves the best trade-off between eradicating data traces and generalizing to unseen data. Soon, we will show that on this dataset, the relearning time of Scissorhands is greater than 200 to regain the performance on $\\mathcal{D}_f$ , which indicates that Scissorhands has efficiently scrubbed information about $\\mathcal{D}_f$ .", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Jing Wu, Mehrtash Harandi", + "bbox": [ + 271, + 114, + 460, + 128 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Overall, Scissorhands illustrates its capability to adeptly balance the objectives of data removal, model utility preservation, and generalization to unseen data, ensuring that the influence of forgotten data is minimized without significantly compromising the overall performance and utility of the model.", + "bbox": [ + 212, + 146, + 787, + 207 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Class-wise unlearning. Furthermore, Table 2 presents the outcomes of our analysis for the scenario where forgetting $10\\%$ of identities on the CelebAMask-HQ. Images are rescaled to $224 \\times 224$ , and a model pre-trained with ImageNet1K [9] is employed. In this context, Scissorhands manifests the smallest average performance gap with retrained models. Notably, the Retrain method achieves perfect metrics in terms of MIA accuracy, which is $100\\%$ , and completely erases the identity information from the dataset, as indicated by a $0.00\\%$ accuracy on the forgotten data $\\mathcal{D}_f$ . However, retraining proves impractical and limited for unlearning purposes in general, functioning merely as an oracle in our experiments.", + "bbox": [ + 212, + 229, + 787, + 366 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Among the baselines, fine-tuning (FT) and $\\ell_1$ -sparse exhibit higher accuracies on the remaining data $\\mathcal{D}_f$ and unseen data $\\mathcal{D}_t$ ; yet, their accuracies on forgetting remain high as well, highlighting a less effective balance between eradicating data traces and generalizing to unseen data. In contrast, Scissorhands and SalUn demonstrate superior capabilities in effectively forgetting identities with minimal impact on the models' overall performance. SalUn achieves an MIA accuracy of $100\\%$ and a forgetting accuracy of $0.00\\%$ , effectively nullifying the identity information similar to the retrained models. Scissorhands slightly surpasses SalUn in terms of the average performance gap with retrained models, maintaining a high MIA accuracy of $99.83\\%$ and a significantly low accuracy on the forgotten data $\\mathcal{D}_f$ (i.e., $1.52\\%$ ), while achieving a test accuracy of $80.18\\%$ and an accuracy of $97.20\\%$ on the retained data.", + "bbox": [ + 212, + 368, + 787, + 547 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We further employ the GradCAM [49] to illustrate the focus area of models $\\mathbf{w}/$ and $\\mathbf{w} / \\mathbf{o}$ machine unlearning algorithms. As shown in Figure 1, when evaluated on the remaining data $\\mathcal{D}_r$ , both SalUn and Scissorhands even pay more attention to the facial features for identity classification than the retrained model. Conversely, when evaluated on the forgetting data $\\mathcal{D}_f$ , the scrubbed model by Scissorhands shifts its attention towards regions that are least associated with the facial features. We hypothesize that in the scenario where the task involves forgetting identities, both the remaining data and the forgetting data are face images, the scrubbed models would need a more nuanced approach to distinguish between individuals. Specifically, because the task at hand requires the discernment of subtle differences across facial features to effectively differentiate and forget specific identities while retaining others, the models adapt by developing a refined sensitivity towards those facial characteristics that are most indicative of individual identity. This adaptation enables the models to maintain high accuracy on the remaining data by focusing more intently on the facial features that are crucial for identity recognition, thereby enhancing their ability to generalize and accurately classify identities.", + "bbox": [ + 212, + 551, + 787, + 808 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "These results underscore our proposed machine unlearning algorithm Scisserhands's superior capabilities in balancing identity forgetting and model utility", + "bbox": [ + 215, + 809, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Scissorhands: Scrub Data Influence via Connection Sensitivity in Networks", + "bbox": [ + 233, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 127 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/41241239017a32499826cd1338fe3b60b7ed4972549993a96bc8e6d4b1036d12.jpg", + "table_caption": [ + "Table 3: Influence of projection and initialization strategies (re-initializing parameters with uniform distribution $\\mathcal{U}$ , Gaussian distribution $\\mathcal{N}$ , and constant value of $\\mathbf{0} / \\mathbf{1}$ )." + ], + "table_footnote": [], + "table_body": "
ProjectionInitializationAccDf(↓)AccDt(↑)AccDr(↑)MIA(↑)Avg. Gap
✓U95.40±1.4892.92±0.4898.93±0.579.56±2.131.62
XU98.14±0.6093.14±0.1699.73±0.166.08±0.782.92
✓N94.68±2.5092.13±2.2097.94±2.7110.18±3.811.80
✓096.10±2.0592.92±0.9999.09±1.198.76±3.301.96
X184.28±0.4983.38±0.5887.18±0.9020.67±0.4110.46
", + "bbox": [ + 241, + 184, + 763, + 272 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/49ce2eb998962f055f1c1505c0eab18abf65a9e97e0ffcfdbbd67a2a5ba70f36.jpg", + "image_caption": [ + "Fig. 2: Influence of the percent value $k$ in the trimming process, the balance term $\\lambda$ of Equation (3), and the ratio of forgetting data used in the trimming process." + ], + "image_footnote": [], + "bbox": [ + 223, + 287, + 408, + 388 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/42debbda1a0e6c1c2adbc107457577a88ce1b522f4d8e5506df22c64479722eb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 413, + 287, + 594, + 387 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/3b54eb9b20dc279b9ad11a69362034f579e60d0e6a084e51e7c9813020958b2b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 598, + 287, + 779, + 388 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "performance, Scissorhands not only minimizes privacy risks but also maintains the integrity and applicability of the model to unseen data.", + "bbox": [ + 212, + 460, + 787, + 492 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.2 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 512, + 388, + 527 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Component analysis. As detailed in Table 3, the inclusion of different procedures in Scissorhands has a noticeable impact. Notably, the employment of the projection mechanism plays a crucial role in eliminating the influence of forgetting data $\\mathcal{D}_f$ , affecting the overall utility of the model. The rationale behind the projection process is to preclude the reacquisition of critical information about the forgetting data $\\mathcal{D}_f$ during the repair phase. Moreover, Table 3 reveals that initializing parameters with a uniform distribution $\\mathcal{U}$ or opting for zero initialization facilitates effective unlearning, in contrast to re-initializing parameters with a constant value of one, which impedes finding a viable solution to Equation (5).", + "bbox": [ + 212, + 537, + 787, + 676 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Hyper-parameter impact. Figure 2 presents the effects of varying hyperparameters, such as the percentage value $k\\%$ for the trimming process, the balance term $\\lambda$ of Equation (3) in the repairing phase, and the ratio of forgetting data used in the trimming process. An increase $k$ results in higher accuracies on data, as fewer parameters associated with the forgetting data $\\mathcal{D}_f$ are reinitialized. Empirically, we observe choosing a value of $k \\geq 0.9$ will result in stable performances. That said, this is a hyperparameter of the algorithm and can be chosen by measuring the performances on $D_{f}$ and $D_{r}$ during training. The hyper-parameter $\\lambda$ controls the balance between retaining and forgetting data, the accuracies on test data and remaining data show a peak at $\\lambda = 0.05$", + "bbox": [ + 212, + 688, + 787, + 843 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Jing Wu, Mehrtash Harandi", + "bbox": [ + 271, + 114, + 460, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/3fcc8d75600a3a75b9f3649de707efdfc9679a9c06b8e6795ed4f33fbc87b2b3.jpg", + "image_caption": [ + "Fig.3: Quantity of nudity content detected using the NudeNet classifier from 1K sampled images and I2P data. We observed a high false positive rate for exposed female genitalia/breast using the NudeNet classifier on generated I2P images. The flagged images can be found in Appendix B." + ], + "image_footnote": [], + "bbox": [ + 225, + 145, + 504, + 273 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/f1cae026a9d6fe32206f7d1e78adee6957744618fdf13d1bdca542d8c70f0ed5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 145, + 776, + 272 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "SD v1.4", + "bbox": [ + 222, + 354, + 263, + 364 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/713a1ee117321f659b6b8469b69a2fca897a11456e862636577b74899a7e1ef3.jpg", + "image_caption": [ + "Fig. 4: Sample images with the prompt from $c_f = \\{ \\text{nudity}, \\text{naked}, \\text'erotic', \\text'sexual' \\}$ generated by SDs w/ and w/o machine unlearning algorithms. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 223, + 364, + 406, + 412 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "SalUn", + "bbox": [ + 410, + 356, + 439, + 364 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/2689f618d6fa3b839545dc332ea13e39ec40792a50a53b9884239f241373cb04.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 410, + 364, + 591, + 412 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Ours", + "bbox": [ + 594, + 357, + 619, + 364 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/50e51cbd01bfd27c35570d50cbcd24fc19546610e8d2930a676a0ce37f2124bc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 594, + 364, + 647, + 412 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "$\\star$ Added by authors for publication", + "bbox": [ + 656, + 356, + 779, + 364 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/f0dde109d4f8e9617c60fb4672f44a226e66ce417cac5949c9bb12b0d8e60f08.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 656, + 364, + 777, + 412 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "when forgetting $10\\%$ data on CIFAR-10. Generally, as $\\lambda$ escalates, it places greater emphasis on the process of data omission, leading to a decrease in forgetting accuracy. As the ratio $p$ increases, all accuracy metrics exhibit a decline, suggesting that employing more forgetting data for the purpose of forgetting in order to isolate critical knowledge about the forgetting data also inadvertently results in the loss of pertinent information regarding the remaining data. This phenomenon occurs particularly when the data designated for forgetting shares the same distribution as the data that is retained. In essence, while aiming to enhance the specificity of the forgetting process by pinpointing essential details about the data to be forgotten, there's a risk of simultaneously diminishing knowledge pertinent to the retained data.", + "bbox": [ + 212, + 484, + 787, + 650 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.3 Case Study: Stable Diffusion", + "text_level": 1, + "bbox": [ + 215, + 675, + 498, + 690 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We also employed our algorithm to mitigate the generation of inappropriate content in Stable Diffusion (SD) models [44]. Specifically, our primary objective is to effectively eliminate the concept of nudity from the model's generative capabilities. To this end, followed [26] and employed SD v1.4 for sampling with 50 time steps. We evaluate on 1K generated images with prompts $c_f = \\{ \\text{nudity}, \\text{naked}, \\text'erotic', \\text'sexual' \\}$ and 4703 generated images with I2P [46] using the open-source NudeNet classifier.", + "bbox": [ + 212, + 703, + 787, + 808 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Figures 3 and 4 present the results for SD v2.1 (trained on a dataset filtered for nudity), SalUn, and Scissorhands. In Figure 3, y-axis denotes the number of", + "bbox": [ + 212, + 809, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Scissorhands: Scrub Data Influence via Connection Sensitivity in Networks", + "bbox": [ + 233, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "exposed body parts generated by the SD v1.4 model, it presents the percentage change in exposed body parts w.r.t. SD v1.4. When comparing on I2P data, our algorithms outperform SalUn on 'Feet', while coming short on 'Female Breast and Genitalia'. While this could suggest the superiority of SalUn on this data, the gap is not that big in reality, as out of the 9 images flagged for containing exposed female breasts, 7 were inaccurately identified, and none of the flagged images depicted exposed female genitalia. The use of the NudeNet classifier on generated I2P images exhibits a significant rate of false positives, as highlighted in [26] as well. When we use prompts $c_f$ explicitly associated with nudity, no exposed sensitive content is detected for both Scissorhands and SalUn methods. We can conclude that, Scissorhands significantly reduces the amount of nudity content compared to SD v1.4, SD v2.1, showing comparable results to SalUn.", + "bbox": [ + 212, + 146, + 787, + 329 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Limitations and Broader Impact", + "text_level": 1, + "bbox": [ + 215, + 349, + 565, + 367 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Our proposed unlearning algorithm, Scissorhands, can achieve comparable performance to state-of-the-art unlearning methods. We addressed potential issues of conflicting gradients arising from direct conflicts between features or patterns in the forgetting and remaining data, ensuring training stability. However, the limitation is the incomplete concept erasure shown in Figure 8 in Appendix B, and the potential for biased models due to the selective deletion of information. This is a crucial consideration, especially when unlearning algorithms are applied with malicious intent. Furthermore, in experiments on stable diffusion, while our method can successfully erase specific concepts, finding an optimal balance between forgetting and maintaining high image quality remains challenging. There is also a risk that unlearning algorithms might be used to alter or remove concepts in unethical or malicious ways. It is our hope that Scissorhands offers a fresh perspective on practical and effective machine unlearning, emphasizing the need for ethical guidelines to prevent misuse in its application.", + "bbox": [ + 212, + 380, + 787, + 592 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 613, + 359, + 628 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we proposed Scissorhands, an effective and practical machine unlearning algorithm. Our algorithm identified the important parameters w.r.t. the forgetting data in single-shot via connection sensitivity, then re-initialized these parameters for scrubbing the influence of the forgetting data. An effective gradient projection-based technique is applied to enhance the model utility while excluding the information w.r.t. the forgetting data. Our evaluations on various datasets showed that, compared with other methods, our approach offers superior performance across image classification and image generation tasks. Future works could explore how Scissorhands performs on regression and NLP tasks, and further investigation into the scenario where data are not available, as well as unlearning with sequential/time-series data where could introduce unique challenges, e.g., the need to consider temporal dependencies, which may require new frameworks specifically tailored for sequential data.", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Jing Wu, Mehrtash Harandi", + "bbox": [ + 271, + 114, + 460, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Mehrtash Harandi is supported by funding from the Australian Research Council (ARC) Discovery Program DP230101176.", + "bbox": [ + 215, + 176, + 784, + 205 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 229, + 321, + 244 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Birhane, A., Prabhu, V.U.: Large image datasets: A pyrrhic win for computer vision? In: 2021 IEEE Winter Conference on Applications of Computer Vision (WACV). pp. 1536-1546. IEEE (2021)", + "2. Bourtoule, L., Chandrasekaran, V., Choquette-Choo, C.A., Jia, H., Travers, A., Zhang, B., Lie, D., Papernot, N.: Machine unlearning. In: 2021 IEEE Symposium on Security and Privacy (SP). pp. 141-159 (2021)", + "3. Cao, Y., Yang, J.: Towards making systems forget with machine unlearning. In: 2015 IEEE Symposium on Security and Privacy (SP). pp. 463-480 (2015)", + "4. Chen, M., Gao, W., Liu, G., Peng, K., Wang, C.: Boundary unlearning: Rapid forgetting of deep networks via shifting the decision boundary. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7766-7775 (2023)", + "5. Chen, M., Zhang, Z., Wang, T., Backes, M., Humbert, M., Zhang, Y.: Graph unlearning. In: Proceedings of the 2022 ACM SIGSAC Conference on Computer and Communications Security. pp. 499-513 (2022)", + "6. Chen, W., Song, D., Li, B.: Trojdiff: Trojan attacks on diffusion models with diverse targets. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4035-4044 (2023)", + "7. Cheng, J., Dasoulas, G., He, H., Agarwal, C., Zitnik, M.: Gnndelete: A general strategy for unlearning in graph neural networks. In: International Conference on Learning Representations (ICLR) (2023)", + "8. Chandawat, V.S., Tarun, A.K., Mandal, M., Kankanhalli, M.: Zero-shot machine unlearning. IEEE Transactions on Information Forensics and Security (2023)", + "9. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)", + "0. Dennis, C.R., Weisberg, S.: Residuals and influence in regression (1982)", + "1. Fan, C., Liu, J., Hero, A., Liu, S.: Challenging forgets: Unveiling the worst-case forget sets in machine unlearning. arXiv preprint arXiv:2403.07362 (2024)", + "2. Fan, C., Liu, J., Zhang, Y., Wei, D., Wong, E., Liu, S.: Salun: Empowering machine unlearning via gradient-based weight saliency in both image classification and generation. In: International Conference on Learning Representations (ICLR) (2024)", + "3. Foster, J., Schoepf, S., Brintrup, A.: Fast machine unlearning without retraining through selective synaptic dampening. arXiv preprint arXiv:2308.07707 (2023)", + "4. Gandhi, S., Kokkula, S., Chaudhuri, A., Magnani, A., Stanley, T., Ahmadi, B., Kandaswamy, V., Ovenc, O., Mannor, S.: Scalable detection of offensive and noncompliant content/logo in product images. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2247-2256 (2020)", + "5. Gandikota, R., Materzynska, J., Fiotto-Kaufman, J., Bau, D.: Erasing concepts from diffusion models. In: 2023 IEEE International Conference on Computer Vision (ICCV) (2023)" + ], + "bbox": [ + 225, + 258, + 785, + 840 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Scissorhands: Scrub Data Influence via Connection Sensitivity in Networks", + "bbox": [ + 233, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Gandikota, R., Orgad, H., Belinkov, Y., Materzyńska, J., Bau, D.: Unified concept editing in diffusion models. arXiv preprint arXiv:2308.14761 (2023)", + "17. Ginart, A., Guan, M., Valiant, G., Zou, J.Y.: Making ai forget you: Data deletion in machine learning. In: Advances in Neural Information Processing Systems (NeurIPS). vol. 32 (2019)", + "18. Goel, S., Prabhu, A., Sanyal, A., Lim, S.N., Torr, P., Kumaraguru, P.: Towards adversarial evaluations for inexact machine unlearning. arXiv preprint arXiv:2201.06640 (2022)", + "19. Golatkar, A., Achille, A., Ravichandran, A., Polito, M., Soatto, S.: Mixed-privacy forgetting in deep networks. In: 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 792-801 (2021)", + "20. Golatkar, A., Achille, A., Soatto, S.: Eternal sunshine of the spotless net: Selective forgetting in deep networks. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9301-9309 (2020)", + "21. Golatkar, A., Achille, A., Soatto, S.: Forgetting outside the box: Scrubbing deep networks of information accessible from input-output observations. In: European Conference on Computer Vision (ECCV). pp. 383-398. Springer (2020)", + "22. Goldman, E.: An introduction to the california consumer privacy act (ccpa). Santa Clara Univ. Legal Studies Research Paper (2020)", + "23. Guo, C., Goldstein, T., Hannun, A., Van Der Maaten, L.: Certified data removal from machine learning models. In: Proceedings of the 37th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 119, pp. 3832-3842. PMLR (2020)", + "24. Halimi, A., Kadhe, S., Rawat, A., Baracaldo, N.: Federated unlearning: How to efficiently erase a client in fl? In: International conference on machine learning. PMLR (2022)", + "25. Han, S., Pool, J., Tran, J., Dally, W.: Learning both weights and connections for efficient neural network. vol. 28 (2015)", + "26. Heng, A., Soh, H.: Selective amnesia: A continual learning approach to forgetting in deep generative models. In: Advances in Neural Information Processing Systems (NeurIPS) (2023)", + "27. Jia, J., Liu, J., Ram, P., Yao, Y., Liu, G., Liu, Y., Sharma, P., Liu, S.: Model sparsification can simplify machine unlearning. In: Advances in Neural Information Processing Systems (NeurIPS) (2023)", + "28. Karasuyama, M., Takeuchi, I.: Multiple incremental decremental learning of support vector machines. IEEE Transactions on Neural Networks 21(7), 1048-1059 (2010)", + "29. Koh, P.W., Liang, P.: Understanding black-box predictions via influence functions. In: International conference on machine learning. pp. 1885-1894. PMLR (2017)", + "30. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)", + "31. Lee, C.H., Liu, Z., Wu, L., Luo, P.: Maskgan: Towards diverse and interactive facial image manipulation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5549-5558 (2020)", + "32. Lee, N., Ajanthan, T., Gould, S., Torr, P.H.: A signal propagation perspective for pruning neural networks at initialization. In: International Conference on Learning Representations (ICLR) (2020)", + "33. Lee, N., Ajanthan, T., Torr, P.H.: Snip: Single-shot network pruning based on connection sensitivity. In: International Conference on Learning Representations (ICLR) (2019)" + ], + "bbox": [ + 215, + 146, + 784, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Jing Wu, Mehrtash Harandi", + "bbox": [ + 271, + 114, + 459, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "34. Liu, G., Ma, X., Yang, Y., Wang, C., Liu, J.: Federaser: Enabling efficient client-level data removal from federated learning models. In: 2021 IEEE/ACM 29th International Symposium on Quality of Service (IWQOS). pp. 1-10 (2021)", + "35. Liu, Y., Xu, L., Yuan, X., Wang, C., Li, B.: The right to be forgotten in federated learning: An efficient realization with rapid retraining. In: IEEE INFOCOM 2022 - IEEE Conference on Computer Communications. pp. 1749-1758 (2022)", + "36. Lopez-Paz, D., Ranzato, M.: Gradient episodic memory for continual learning. Advances in neural information processing systems 30 (2017)", + "37. Mehta, R., Pal, S., Singh, V., Ravi, S.N.: Deep unlearning via randomized conditionally independent hessians. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10412-10421 (2022)", + "38. Neel, S., Roth, A., Sharifi-Malvajerdi, S.: Descent-to-delete: Gradient-based methods for machine unlearning. In: Proceedings of the 32nd International Conference on Algorithmic Learning Theory. pp. 931-962. Proceedings of Machine Learning Research, PMLR (2021)", + "39. Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B., Ng, A.Y.: Reading digits in natural images with unsupervised feature learning (2011)", + "40. Nguyen, T.T., Huynh, T.T., Nguyen, P.L., Liew, A.W.C., Yin, H., Nguyen, Q.V.H.: A survey of machine unlearning. arXiv preprint arXiv:2209.02299 (2022)", + "41. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741 (2021)", + "42. Peste, A., Alistarh, D., Lampert, C.H.: Ssse: Efficiently erasing samples from trained machine learning models. arXiv preprint arXiv:2107.03860 (2021)", + "43. Rando, J., Paleka, D., Lindner, D., Heim, L., Tramér, F.: Red-teaming the stable diffusion safety filter. arXiv preprint arXiv:2210.04610 (2022)", + "44. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)", + "45. Romero, E., Barrio, I., Belanche, L.: Incremental and decremental learning for linear support vector machines. In: International Conference on Artificial Neural Networks. pp. 209-218. Springer (2007)", + "46. Schramowski, P., Brack, M., Deiseroth, B., Kersting, K.: Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22522-22531 (2023)", + "47. Schramowski, P., Tauchmann, C., Kersting, K.: Can machines help us answering question 16 in datasheets, and in turn reflecting on inappropriate content? In: Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency. pp. 1350-1361 (2022)", + "48. Sekhari, A., Acharya, J., Kamath, G., Suresh, A.T.: Remember what you want to forget: Algorithms for machine unlearning. Advances in Neural Information Processing Systems (NeurIPS) 34, 18075-18086 (2021)", + "49. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Gradcam: Visual explanations from deep networks via gradient-based localization. In: 2017 IEEE International Conference on Computer Vision (ICCV). pp. 618-626 (2017)", + "50. Tarun, A.K., Chundawat, V.S., Mandal, M., Kankanhalli, M.: Fast yet effective machine unlearning. IEEE Transactions on Neural Networks and Learning Systems (2023)" + ], + "bbox": [ + 215, + 146, + 784, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Scissorhands: Scrub Data Influence via Connection Sensitivity in Networks", + "bbox": [ + 233, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "51. Tarun, A.K., Chundawat, V.S., Mandal, M., Kankanhalli, M.: Deep regression unlearning. In: International Conference on Machine Learning. pp. 33921-33939. PMLR (2023)", + "52. Thudi, A., Deza, G., Chandrasekaran, V., Papernot, N.: Unrolling sgd: Understanding factors influencing machine unlearning. In: 2022 IEEE 7th European Symposium on Security and Privacy (EuroS&P). pp. 303-319. IEEE (2022)", + "53. Voigt, P., Von dem Bussche, A.: The eu general data protection regulation (gdpr). A Practical Guide, 1st Ed., Cham: Springer International Publishing 10(3152676), 10-5555 (2017)", + "54. Wang, J., Guo, S., Xie, X., Qi, H.: Federated unlearning via class-discriminative pruning. In: Proceedings of the ACM Web Conference 2022. pp. 622-632 (2022)", + "55. Warnecke, A., Pirch, L., Wressnegger, C., Rieck, K.: Machine unlearning of features and labels. arXiv preprint arXiv:2108.11577 (2021)", + "56. Wu, C., Zhu, S., Mitra, P.: Federated unlearning with knowledge distillation. arXiv preprint arXiv:2201.09441 (2022)", + "57. Wu, J., Le, T., Hayat, M., Harandi, M.: Erasediff: Erasing data influence in diffusion models. arXiv preprint arXiv:2401.05779 (2024)", + "58. Wu, Y., Dobriban, E., Davidson, S.: DeltaGrad: Rapid retraining of machine learning models. In: Proceedings of the 37th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 119, pp. 10355-10366. PMLR (13-18 Jul 2020)", + "59. Zhang, C., Bengio, S., Hardt, M., Recht, B., Vinyals, O.: Understanding deep learning (still) requires rethinking generalization. In: International Conference on Learning Representations (ICLR) (2017)", + "60. Zhang, E., Wang, K., Xu, X., Wang, Z., Shi, H.: Forget-me-not: Learning to forget in text-to-image diffusion models. arXiv preprint arXiv:2303.17591 (2023)", + "61. Zhang, Y., Zhang, Y., Yao, Y., Jia, J., Liu, J., Liu, X., Liu, S.: Unlearncanvas: A stylized image dataset to benchmark machine unlearning for diffusion models. arXiv preprint arXiv:2402.11846 (2024)", + "62. Zhang, Y., Chen, X., Jia, J., Zhang, Y., Fan, C., Liu, J., Hong, M., Ding, K., Liu, S.: Defensive unlearning with adversarial training for robust concept erasure in diffusion models. arXiv preprint arXiv:2405.15234 (2024)" + ], + "bbox": [ + 215, + 146, + 784, + 590 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Jing Wu, Mehrtash Harandi", + "bbox": [ + 271, + 114, + 459, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Scissorhands_ Scrub Data Influence via Connection Sensitivity in Networks/43a51c48-5759-4f46-b600-93f450455ce7_model.json b/2024/Scissorhands_ Scrub Data Influence via Connection Sensitivity in Networks/43a51c48-5759-4f46-b600-93f450455ce7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..16a7166ae2d32e3335655df2226c46ebc284a2ea --- /dev/null +++ b/2024/Scissorhands_ Scrub Data Influence via Connection Sensitivity in Networks/43a51c48-5759-4f46-b600-93f450455ce7_model.json @@ -0,0 +1,2326 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.28, + 0.141, + 0.726, + 0.187 + ], + "angle": 0, + "content": "Scissorhands: Scrub Data Influence via Connection Sensitivity in Networks" + }, + { + "type": "text", + "bbox": [ + 0.38, + 0.212, + 0.624, + 0.228 + ], + "angle": 0, + "content": "Jing Wu1, Mehrtash Harandi1" + }, + { + "type": "text", + "bbox": [ + 0.294, + 0.24, + 0.711, + 0.282 + ], + "angle": 0, + "content": "\\(^{1}\\)Department of Electrical and Computer Systems Engineering Monash University, Australia {jing.wu1, mehrtash.harandi}@monash.edu" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.313, + 0.744, + 0.535 + ], + "angle": 0, + "content": "Abstract. Machine unlearning has become a pivotal task to erase the influence of data from a trained model. It adheres to recent data regulation standards and enhances the privacy and security of machine learning applications. In this work, we present a new machine unlearning approach Scissorhands. Initially, Scissorhands identifies the most pertinent parameters in the given model relative to the forgetting data via connection sensitivity. By reinitializing the most influential top-\\(k\\) percent of these parameters, a trimmed model for erasing the influence of the forgetting data is obtained. Subsequently, Scissorhands fine-tunes the trimmed model with a gradient projection-based approach, seeking parameters that preserve information on the remaining data while discarding information related to the forgetting data. Our experimental results, conducted across image classification and image generation tasks, demonstrate that Scissorhands, showcases competitive performance when compared to existing methods. Source code is available at https://github.com/JingWu321/Scissorhands." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.547, + 0.752, + 0.562 + ], + "angle": 0, + "content": "Keywords: Machine unlearning \\(\\cdot\\) Connection sensitivity \\(\\cdot\\) Diffusion model" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.58, + 0.741, + 0.607 + ], + "angle": 0, + "content": "Warning: This paper contains explicit sexual imagery that may be offensive." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.631, + 0.377, + 0.646 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.749 + ], + "angle": 0, + "content": "In this work, we aim to propose an effective machine unlearning method. Under data regulations like the European Union General Data Protection Regulation (GDPR) [53] and the California Consumer Privacy Act (CCPA) [22], all users are granted the right to be forgotten. In machine learning, these legal provisions empower data owners with the right not only to withdraw their data from trained models but also to ensure that their data's influence on these models is erased." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The most direct approach to accomplish this objective is to retrain the model from scratch, excluding the data requested for deletion from the training process. Retraining from scratch is typically considered the gold standard in the field of machine unlearning [12, 52]. Yet, this poses a challenge as numerous in-production models require prolonged training periods and substantial computing resources. While retraining is feasible, it is often impractical." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.461, + 0.129 + ], + "angle": 0, + "content": "Jing Wu, Mehrtash Harandi" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "Consequently, the proposal and development of efficient approximate unlearning methods [13, 23, 26, 50, 51] have become essential. Currently, most approximate unlearning techniques achieve forgetting by adding normally distributed noise to the parameters [19-21] or by estimating the influence of a particular data point on the model's predictions [23, 37, 38, 48] based on the influence function method [10]. Jia et al. [27] demonstrate that model sparsity can help to unlearn and fuse model sparsity into the unlearning process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.788, + 0.465 + ], + "angle": 0, + "content": "Fan et al. [12] recently raised concerns regarding the instability in approximate unlearning methods, and highlighted that the current machine unlearning approaches, initially crafted for image classification tasks, fall short in effectively tackling the challenges of machine unlearning within the realm of image generation. The authors introduce a novel unlearning method SalUn that can effectively perform forgetting in both image classification and generation tasks. Their unlearning mechanism includes finding salient weights and then fine-tuning these weights using forgetting data assigned with random labels. However, studies like [59] show that the model can still learn the true data distribution even using data with random labels. While SalUn has shown state-of-the-art performance, it could potentially memorize information about forgetting data points, albeit associating them with random labels but still can be undesired in sensitive applications. In § 4, results also show that SalUn tends to memorize knowledge about the forgetting data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.465, + 0.788, + 0.661 + ], + "angle": 0, + "content": "This work. We present an unlearning approach, Scissorhands, designed to erase data influence in the classification models and eliminate a particular concept from a text-to-image model. To achieve this, inspired by [27], our key insight is to first erase the critical influence of forgetting data in models and then relearn. Unlike previous methods that employ the forgetting data with noisy labels as a part of relearning, we propose to use an efficient gradient projection method to relearn the critical features and patterns while ensuring the exclusion of influences associated with the forgetting data. Through a series of experiments and evaluations, including the classification on SVHN [39], CIFAR-10 and CIFAR-100 [30], CelebAMASK-HQ [31] datasets, as well as the open-source stable diffusion [44] text-to-image model, results demonstrate the viability and effectiveness of our technique in forgetting the influence of random samples, discrete classes, and sensitive content such as nudity." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.685, + 0.381, + 0.703 + ], + "angle": 0, + "content": "2 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.718, + 0.788, + 0.795 + ], + "angle": 0, + "content": "In this section, we propose Scissorhands, our unlearning framework that scrubs data from a model by eliminating their influence on model parameters. Throughout the paper, we denote scalars and vectors/matrices by lowercase and bold symbols, respectively (e.g., \\(a\\), \\(a\\), and \\(A\\)). We show a \\(d\\)-dimensional vector of ones by \\(\\mathbf{1}_d\\), and use \\(\\odot\\) to denote the Hadamard product." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.81, + 0.787, + 0.843 + ], + "angle": 0, + "content": "Overview. Consider a model \\(f\\) with parameters \\(\\pmb{\\theta} \\in \\mathbb{R}^d\\) trained on a dataset \\(\\mathcal{D} = \\{\\mathbf{x}_i, \\mathbf{y}_i\\}_{i=1}^N\\). Suppose a subset of \\(\\mathcal{D}\\), denoted as \\(\\mathcal{D}_f\\), is requested for deletion." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.234, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Scissorhands: Scrub Data Influence via Connection Sensitivity in Networks" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "The remaining data is defined as \\(\\mathcal{D}_r\\coloneqq \\mathcal{D}\\setminus \\mathcal{D}_f\\). Our objective is to develop the unlearning algorithm to generate a scrubbed model \\(f_{u}\\), effectively removing the influence of \\(\\mathcal{D}_f\\) from model \\(f\\), while maintaining its utility over \\(\\mathcal{D}_r\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.785, + 0.281 + ], + "angle": 0, + "content": "In Scissorhands, we initially identify critical parameters w.r.t. the forgetting data via connection sensitivity analysis [32, 33]. We then reinitialize the top-\\(k\\) percent of these key parameters, resulting in a trimmed model with parameters \\(\\pmb{\\theta}_t\\). This step aims to diminish the model's memory of the forgetting data \\(\\mathcal{D}_f\\). Nevertheless, this process runs the risk of erasing valuable information from the data set \\(\\mathcal{D}_r\\) that we aim to retain." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.785, + 0.358 + ], + "angle": 0, + "content": "To address this, the next phase of Scissorhands concentrates on restoring the performance of the trimmed model on remaining data \\(\\mathcal{D}_r\\), while ensuring that the influence of the forgetting data \\(\\mathcal{D}_f\\) remains excluded. We accomplish this balance and obtain the final scrubbed model with parameters \\(\\theta^{\\star}\\) through a gradient projection-based approach." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.379, + 0.345, + 0.394 + ], + "angle": 0, + "content": "2.1 Trimming" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.402, + 0.784, + 0.507 + ], + "angle": 0, + "content": "To effectively identify salient parameters of a network with respect to \\(\\mathcal{D}_f\\), one has to establish a good criterion to determine such salient connections. A commonly used criterion is the magnitude of the weights, with those exceeding a certain threshold deemed salient as suggested in [25]. Our requirement is more nuanced, as we need a measure that specifically discerns saliency based on \\(\\mathcal{D}_f\\). Therefore, we adopt a single-shot approach as proposed in [33], which we outline briefly to ensure clarity and completeness in our methodology. Consider" + }, + { + "type": "equation", + "bbox": [ + 0.317, + 0.516, + 0.785, + 0.541 + ], + "angle": 0, + "content": "\\[\n\\mathbf {s} _ {j} (\\mathcal {D}) := \\mathbb {E} _ {\\boldsymbol {x}, \\boldsymbol {y} \\sim \\mathcal {D}} \\left[ \\ell (\\boldsymbol {\\theta}; \\boldsymbol {x}, \\boldsymbol {y}) - \\ell \\left(\\left(\\mathbf {1} _ {d} - \\boldsymbol {e} _ {j}\\right) \\odot \\boldsymbol {\\theta}; \\boldsymbol {x}, \\boldsymbol {y}\\right) \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.549, + 0.788, + 0.64 + ], + "angle": 0, + "content": "which measures the influence of parameter \\( j \\in \\{1, \\dots, d\\} \\) on a model in terms of the empirical risk for a given dataset \\( \\mathcal{D} \\). Here, \\( \\ell : \\mathbb{R}^d \\times \\mathcal{X} \\times \\mathcal{Y} \\to \\mathbb{R}_+ \\) is the loss of the model and \\( e_j \\) is the indicator vector of element \\( j \\) (i.e., a binary vector with its \\( j \\)-th component set to one). Note that computing \\( \\mathbf{s}_j \\) for each \\( j \\in \\{1, \\dots, d\\} \\) is prohibitively expensive, as it demands \\( d + 1 \\) forward passes through the dataset. An approximation to Equation (1) [25] is in the form of" + }, + { + "type": "equation", + "bbox": [ + 0.385, + 0.649, + 0.785, + 0.681 + ], + "angle": 0, + "content": "\\[\n\\mathbf {s} _ {j} (\\mathcal {D}) \\approx \\mathbb {E} _ {\\boldsymbol {x}, \\boldsymbol {y} \\sim \\mathcal {D}} \\left[ \\frac {\\partial \\ell (\\boldsymbol {\\theta} ; \\boldsymbol {x} , \\boldsymbol {y})}{\\partial \\boldsymbol {\\theta} _ {j}} \\boldsymbol {\\theta} _ {j} \\right]. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.784, + 0.809 + ], + "angle": 0, + "content": "Detailed proofs can be found in Appendix A. Equation (2) defines the sensitivity of parameter \\( j \\) as the average (i.e., expectation) of a product, the gradient of the loss and the current value of the parameter. As such, \\( \\mathbf{s}_j \\) encodes the loss sensitivity according to the parameter's magnitude. Intuitively, a parameter with a large value that significantly affects the loss, as captured by the gradient, is considered more influential or salient. The form in Equation (2) offers several benefits, most importantly, the ability to calculate the sensitivity for all parameters with just a single sweep through the dataset." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.784, + 0.841 + ], + "angle": 0, + "content": "To scrub the influence of the forgetting data \\(\\mathcal{D}_f\\) in the model, we first obtain \\(\\mathfrak{s}_j(\\mathcal{D}_f)\\), saliency of parameters w.r.t. the forgetting data. We then re-initialize" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.461, + 0.129 + ], + "angle": 0, + "content": "Jing Wu, Mehrtash Harandi" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "the top-\\(k\\%\\) of the parameters based on their saliency rankings. This process is akin to performing a targeted 'lobotomy' on the model specifically concerning the data \\(\\mathcal{D}_f\\), thereby selectively erasing their influence. Unfortunately, this aggressive approach of reinitializing parameters can detrimentally impact the model's performance on the remaining dataset \\(\\mathcal{D}_r\\), which we will correct in the next phase of Scissorhands." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.249, + 0.788, + 0.327 + ], + "angle": 0, + "content": "Remark 1. In principle, Scissorhands can benefit from any algorithms that can identify important connections w.r.t. \\(\\mathcal{D}_f\\). The main reason behind choosing SNIP is its single-shot nature. As presented in the study [12], the gradient of loss w.r.t. the model parameters over the forgetting data can help identify important parameters w.r.t. \\(\\mathcal{D}_f\\) and hence help with erasing." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.335, + 0.788, + 0.398 + ], + "angle": 0, + "content": "Remark 2. In practice, we have observed that even a small subset of \\(\\mathcal{D}_f\\) can be sufficient for the trimming. For example, to unlearn a large set of data on CIFAR-100, our algorithm outperforms baselines by utilizing approximately \\(3\\%\\) of the forgetting data for trimming." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.407, + 0.788, + 0.469 + ], + "angle": 0, + "content": "Remark 3. Initializing weights of the trimmed neurons can take various forms. Empirically, we observed initializing with a uniform distribution is particularly effective. Details on the influence of different initialization strategies and the choice of \\(\\mathrm{k}\\%\\) trimming on the model's performance will be discussed in \\(\\S 4.2\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.489, + 0.345, + 0.505 + ], + "angle": 0, + "content": "2.2 Repairing" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.514, + 0.789, + 0.683 + ], + "angle": 0, + "content": "Following the parameter-trimming process, aimed at mitigating the influence of the forgetting data \\(\\mathcal{D}_f\\), a challenge presents itself: the potential erasure of crucial information associated with the remaining data \\(\\mathcal{D}_r\\). A straightforward solution to recover the model utility is to relearn the model using the remaining data. However, this approach risks biasing the model towards features of the remaining data and inadvertently reintegrating information about the forgetting samples. In the quest to ensure that models retain essential information from the remaining data \\(\\mathcal{D}_r\\) while effectively forgetting \\(\\mathcal{D}_f\\), our strategy involves maximizing the loss over the forgetting data \\(\\mathcal{D}_f\\) while concurrently minimizing the loss over the remaining data \\(\\mathcal{D}_r\\). Therefore, to achieve the balance between these goals, we propose the objective of an efficient practice for unlearning:" + }, + { + "type": "equation", + "bbox": [ + 0.365, + 0.693, + 0.785, + 0.71 + ], + "angle": 0, + "content": "\\[\nL (\\boldsymbol {\\theta}, \\mathcal {D} _ {r}, \\mathcal {D} _ {f}) := \\mathcal {L} (\\boldsymbol {\\theta}; \\mathcal {D} _ {r}) - \\lambda \\mathcal {L} (\\boldsymbol {\\theta}; \\mathcal {D} _ {f}). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.843 + ], + "angle": 0, + "content": "We aim to optimize Equation (3) w.r.t. \\(\\pmb{\\theta}\\) through gradient descent, ensuring minimal reintroduction of information about \\(\\mathcal{D}_f\\). This is particularly challenging when there are similarities between samples in \\(\\mathcal{D}_f\\) and \\(\\mathcal{D}_r\\). To achieve this desideratum, it is vital that updates to the model do not improve (i.e., reduce) \\(\\mathcal{L}(\\pmb{\\theta};\\mathcal{D}_f)\\). To this end, we define \\(\\pmb{g}_f = \\nabla_{\\pmb{\\theta}}\\mathcal{L}(\\pmb{\\theta};\\mathcal{D}_f)\\) as the gradient direction for \\(\\mathcal{D}_f\\) and \\(\\pmb{g}_o = \\nabla_{\\pmb{\\theta}}L(\\pmb{\\theta},\\mathcal{D}_r,\\mathcal{D}_f)\\) as the gradient direction for optimizing Equation (3). The optimal descent direction \\(\\pmb{g} \\in \\mathbb{R}^d\\) should thus satisfy two criteria: 1. It should exhibit maximum similarity to \\(\\pmb{g}_o\\) to ensure swift convergence. The" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.234, + 0.114, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Scissorhands: Scrub Data Influence via Connection Sensitivity in Networks" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "code_caption", + "bbox": [ + 0.218, + 0.147, + 0.551, + 0.162 + ], + "angle": 0, + "content": "Algorithm 1 The procedure of Scissorhands." + }, + { + "type": "algorithm", + "bbox": [ + 0.217, + 0.163, + 0.788, + 0.378 + ], + "angle": 0, + "content": "1: procedure TRIMMING \\((\\pmb {\\theta},\\mathcal{D}_f)\\) 11: for \\(e = 0\\) to \\(E - 1\\) do \n2: \\(\\pmb {X},\\pmb {y}\\gets \\{\\pmb {x}_i,\\pmb {y}_i\\sim \\mathcal{D}_f|i\\in [B]\\} .\\) 12: Get \\(L(\\pmb {\\theta},\\mathcal{D}_r,\\mathcal{D}_f)\\) (cf. Eq. (3)). \n3: // Compute connection sensitivity 13: Get \\(\\pmb{g_o}\\) and \\(\\pmb{g_f}\\) \n4: Get \\(\\mathbf{s}_j(\\pmb {\\theta};\\pmb {X},\\pmb {y}),\\forall j\\in [1,d]\\) 14: //Get the optimal direction using Eq.(2). 15: if \\(\\langle g_o,g_f\\rangle >0\\) then \n5: // Re-initialization 16: Compute \\(v^{*}\\) (cf. Eq. (5)). \n6: Re-initialize parameters according 17: \\(\\pmb {g} = \\pmb {g}_o - v^*\\pmb {g}_f\\) to the top- \\(k \\%\\) value of \\(\\mathfrak{s}_j\\) 18: \\(\\pmb {g}_o\\gets \\pmb {g}\\) \n7: return \\(\\pmb{\\theta}_{t}\\) 19: end if \n8: end procedure 20: \\(\\pmb{\\theta}^{e + 1}\\leftarrow \\pmb{\\theta}^{e} - \\eta \\pmb{g}_{o}\\) \n21: end for \n9: procedure REPAIRING \\((\\pmb {\\theta}_t,\\mathcal{D}_f,\\mathcal{D}_r)\\) 22: return \\(\\pmb{\\theta}^{*} = \\pmb{\\theta}^{E}\\) \n10: \\(\\pmb{\\theta}^{0} = \\pmb{\\theta}_{t}\\) 23: end procedure" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.408, + 0.789, + 0.485 + ], + "angle": 0, + "content": "notion of similarity can be captured by \\( \\pmb{g}_o \\) using \\( \\| \\pmb{g} - \\pmb{g}_o \\|_2^2 \\). 2. It should not align with \\( \\pmb{g}_f \\) to prevent improving \\( \\mathcal{L}(\\pmb{\\theta}; \\mathcal{D}_f) \\). This can be mathematically formalized as \\( \\langle \\pmb{g}_f, \\pmb{g} \\rangle \\leq 0 \\). Consequently, after computing \\( \\pmb{g}_o = \\nabla_{\\pmb{\\theta}} L(\\pmb{\\theta}, \\mathcal{D}_r, \\mathcal{D}_f) \\) and \\( \\pmb{g}_f = \\nabla_{\\pmb{\\theta}} \\mathcal{L}(\\pmb{\\theta}; \\mathcal{D}_f) \\) in each iteration, we formulate the following optimization challenge to identify the model update direction:" + }, + { + "type": "equation", + "bbox": [ + 0.418, + 0.495, + 0.786, + 0.549 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\underset {\\boldsymbol {g}} {\\arg \\min} \\frac {1}{2} \\| \\boldsymbol {g} - \\boldsymbol {g} _ {o} \\| _ {2} ^ {2}, \\\\ \\text {s . t .} \\quad \\langle \\boldsymbol {g} _ {f}, \\boldsymbol {g} \\rangle \\leq 0. \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.559, + 0.788, + 0.622 + ], + "angle": 0, + "content": "This problem may be addressed using the Frank-Wolfe algorithm. Given the high dimensionality of gradients in neural networks, directly addressing the constraint optimization in Equation (4) could become overwhelming. Inspired by [36], we propose to make use of the dual formulation of the problem Equation (4) as:" + }, + { + "type": "equation", + "bbox": [ + 0.409, + 0.644, + 0.786, + 0.695 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\sup _ {v} \\frac {1}{2} \\boldsymbol {g} _ {f} ^ {\\top} \\boldsymbol {g} _ {f} v ^ {2} - \\boldsymbol {g} _ {o} ^ {\\top} \\boldsymbol {g} _ {f} v, \\\\ \\begin{array}{l l} \\text {s . t .} & v \\geq 0. \\end{array} \\tag {5} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.706, + 0.788, + 0.753 + ], + "angle": 0, + "content": "Then, the optimal descent direction \\(\\pmb{g}\\) is given from the solution \\(v^{*}\\) in Equation (5) as \\(\\pmb {g} = \\pmb{g}_{o} - v^{*}\\pmb{g}_{f}\\). Detailed proofs can be found in Appendix A. By doing so, we ensure a balance between unlearning and retaining utility." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Remark 4. The primal problem (cf. Equation (4)) involves optimizing over the high-dimensional space of gradient vectors, while the dual problem (cf. Equation (5)) simplifies this problem to optimize over a single scalar variable \\( v \\in \\mathbb{R} \\). This dimensionality reduction can significantly decrease the computational complexity, especially for large-scale neural networks." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.46, + 0.129 + ], + "angle": 0, + "content": "Jing Wu, Mehrtash Harandi" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.449, + 0.163 + ], + "angle": 0, + "content": "2.3 Algorithm Description" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.169, + 0.788, + 0.385 + ], + "angle": 0, + "content": "Algorithm 1 describes the procedure of Scissorhands in detail. We first identify the important parameters w.r.t. the forgetting data in single-shot via connection sensitivity in networks, then re-initialize these parameters for erasing the influence of the forgetting data \\(\\mathcal{D}_f\\) in the given model. To repair the performance of the trimmed model on the remaining data \\(\\mathcal{D}_r\\), while ensuring the exclusion of information associated with the forgetting data \\(\\mathcal{D}_f\\), we employ the gradient projection-based optimization algorithm. By doing so, we can scrub the influence of the forgetting data while retaining model utility on the remaining data. The loss depicted in Equation (3) can be viewed as a scalarization of a Multi-Objective Optimization (MOO) problem, namely minimizing \\(\\left(\\mathcal{L}(\\theta; \\mathcal{D}_r), -\\mathcal{L}(\\theta; \\mathcal{D}_f)\\right)^{\\top}\\). A common issue in MOO is the gradient conflict, as optimizing one objective could hinder another one. Our proposal on gradient projection is designed to address and mitigate the gradient conflict. Empirically (see Table 3), we studied the effect of removing gradient projection." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.404, + 0.389, + 0.42 + ], + "angle": 0, + "content": "3 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.432, + 0.787, + 0.522 + ], + "angle": 0, + "content": "The development of efficient machine unlearning methods [2, 3, 11, 17, 18, 20, 23, 27, 28, 37, 40, 42, 45, 48, 51, 58, 61] has gained prominence. Applications of machine unlearning span various domains, including regression tasks [51], federated learning [24, 34, 35, 54, 56], graph neural networks [5, 7], and diffusion models [12, 15, 16, 26, 57, 60, 62], as well as scenarios where training data are not available [8, 50]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.524, + 0.788, + 0.644 + ], + "angle": 0, + "content": "Retraining the model from scratch without forgetting data is typically considered the standard gold unlearning algorithm [12, 52]. However, this is often deemed impractical as most in-production models require extensive training periods and considerable computing resources. While fine-tuning the model for a new task might lead to forgetting previous knowledge (i.e., catastrophic forgetting) [36], it fails to adequately erase the data influence in the models. Approximate unlearning methods, as such, become attractive alternatives. We will briefly discuss these methods in image classification and image generation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Unlearning in Classification Models. Most unlearning algorithms are based on the influence function [23,38,48] and the Fisher Information Matrix (FIM) [19-21]. Influence function-based methods estimate the influence of the particular training data points on the model's predictions, and Fisher unlearning methods assume that the unlearned model and the retrained model are close to each other, Golatkar et al. [20] leverage FIM and hide the difference between the unlearned models and retrained models via adding noise. Golatkar et al. [20] also present an upper bound for the amount of retained information, offering a quantifiable measure of the effectiveness of the unlearning process. Such approximate unlearning methods need to compute the Hessian matrix, or FIM w.r.t. the data. This may render them impractical for certain scenarios, such as federated learning, where computational resources are distributed and limited. To mitigate these computational demands, Mehta et al. [37] introduce L-CODEC, a strategy aimed at" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.234, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Scissorhands: Scrub Data Influence via Connection Sensitivity in Networks" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.207 + ], + "angle": 0, + "content": "pre-selecting a subset of parameters, allowing for more efficient computation. Jia et al. [27] recently propose to fuse the model sparsity into the unlearning algorithm, helping improve approximate unlearning methods, and achieving effective and efficient unlearning effects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.787, + 0.358 + ], + "angle": 0, + "content": "Concept Erasure in Generative Models. The advent of generative models, particularly those converting text to images, has been a significant milestone in the field of artificial intelligence. A notable concern is the risk of it being tainted or manipulated [6,43], resulting in not-safe-for-work (NSFW) generations [46], as these models leverage training data from a wide array of open sources. To address these challenges, data censoring [1, 14, 41, 47] to exclude black-listed images is employed. Studies [15, 46] introduce methods to update the models away from the inappropriate concepts. Heng and Soh [26] recently propose to adopt Elastic Weight Consolidation (EWC) and Generative Replay (GR) to effectively unlearn without access to the training data for a wide range of generative models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.36, + 0.787, + 0.464 + ], + "angle": 0, + "content": "Unlearning across Domains. Recently, Fan et al. [12] highlight the difficulty in cross-domain applicability of the machine unlearning algorithms, i.e., existing machine unlearning methods designed for classification tasks cannot be effective when applied to image generation. To address this gap, they introduce a novel method, Saliency unlearning (SalUn), which shifts attention from the entire model to target parameters. SalUn achieves state-of-the-art performance and is effective in both image classification and image generation tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.466, + 0.788, + 0.707 + ], + "angle": 0, + "content": "In this work, we propose Scissorhands which could achieve improved performance compared to existing methods across various scenarios. Inspired by [27] that apply \\(\\ell_1\\) regularization to simply fuse sparsity into the unlearning methods, we identify and re-initialize the key model parameters w.r.t. the forgetting data via connection sensitivity. Concurrently, Fan et al. [12] also explore unlearning via salience scores, determining weight salience via the gradient of the forgetting loss w.r.t. model parameters, and employing random labeling for model finetuning. While there are similarities in our overarching goals, the methodologies present unique perspectives and solutions to the problem of unlearning. Fan et al. [12] only focus on the salient weights, while our mechanism aims to first 'destroy' the key information and then relearn for the whole model. Additionally, using the forgetting data with random labels could potentially lead to memorizing information about the forgetting data points [59]. In contrast, Scissorhands excludes knowledge about the forgetting data via a gradient projection-based approach, resulting in the effective erasure of information about the forgetting data while preserving the knowledge about the retained data." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.732, + 0.494, + 0.749 + ], + "angle": 0, + "content": "4 Experimental Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.841 + ], + "angle": 0, + "content": "In this section, we illustrate how Scissorhands effectively eliminates the data influence in the models. For sample-wise forgetting, where the forgetting data shares the same distribution as the training data, we evaluate Scissorhands on SVHN [39], CIFAR-10 and CIFAR-100 [30]. We further extend our evaluation on CelebAMask-HQ [31] where targeting to remove the entirety of specific identities." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.461, + 0.129 + ], + "angle": 0, + "content": "Jing Wu, Mehrtash Harandi" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.208 + ], + "angle": 0, + "content": "Additionally, our experimentation encompasses the open-source text-to-image model, Stable Diffusion v1.4 [44], which is conditioned on CLIP text embeddings through the cross-attention mechanism. Further detailed experimental setups and additional results can be found in Appendix B." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.213, + 0.796, + 0.442 + ], + "angle": 0, + "content": "Baselines. We primarily compare against the following standard baselines that are frequently employed in machine unlearning, alongside recently proposed SOTA methods: (i) Retrain: models obtained by only retraining from scratch on \\(\\mathcal{D}_r\\). This will provide the performance of the oracle for us. (ii) Fine-tuning (FT) [55]: models that are fine-tuned on \\(\\mathcal{D}_r\\), i.e., taking advantage from catastrophic forgetting in neural networks to unlearn. (iii) Gradient ascent (GA) [52]: gradient ascent on \\(\\mathcal{D}_f\\). This will provide a simple way of unlearning by making the performance of the model worse on \\(\\mathcal{D}_f\\). (iv) Influence unlearning (IU) [29]: utilizes the influence function method to estimate the change in model parameters when transitioning from the unscrubbed model to the retrained model. (v) Boundary shrink (BS) [4] and (vi) Boundary expanding (BE) [4]: shift the decision boundary of the original model to imitate that of the retrained model. (vii) \\(\\ell_1\\)-sparse [27]: fine-tuning models on \\(\\mathcal{D}_r\\) with \\(\\ell_1\\)-norm sparse regularization. (viii) Saliency unlearning (SalUn) [12]: adopt weight saliency and random labeling for unlearning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.447, + 0.79, + 0.844 + ], + "angle": 0, + "content": "Metrics. To assess the unlearning algorithms, we employ several metrics: (i) Accuracy: the accuracy of the model on \\(\\mathcal{D}_f\\) (denoted as \\(\\mathbf{Acc}_{\\mathcal{D}_f}\\)), \\(\\mathcal{D}_r\\) (denoted as \\(\\mathbf{Acc}_{\\mathcal{D}_r}\\)) and the test set (denoted as \\(\\mathbf{Acc}_{\\mathcal{D}_t}\\)). It provides insight into how well the model performs after undergoing the unlearning process. (ii) Membership inference attack (MIA): a standard metric for verifying the unlearning effect. A classifier \\(\\phi\\) is trained using both training data (marked with a label of 1, indicating data points included in training) and test data (marked with a label of 0, indicating data points not seen during training), and subsequently assessed on \\(\\mathcal{D}_f\\). An effective unlearning method should make it challenging to infer whether a particular sample was part of the training set, mitigating potential privacy breaches. In this context, we define MIA as \\(P_{\\phi}(\\pmb{y} = 0|\\pmb{x}_f)\\), representing the probability assigned by the classifier \\(\\phi\\) that a given sample \\(\\pmb{x}_f\\) was not included in the training data. (iii) Avg. Gap [12]: the average of the performance gaps measured in accuracy-related metrics: Avg. Gap \\(= (|\\mathrm{Acc}_{\\mathcal{D}_t} - \\mathrm{Acc}_{\\mathcal{D}_t}^*| + |\\mathrm{Acc}_{\\mathcal{D}_f} - \\mathrm{Acc}_{\\mathcal{D}_f}^*| + |\\mathrm{Acc}_{\\mathcal{D}_r} - \\mathrm{Acc}_{\\mathcal{D}_r}^*| + |\\mathrm{MIA} - \\mathrm{MIA}^*|)/4\\), where \\(\\mathrm{Acc}_{\\mathcal{D}_t}^*\\), \\(\\mathrm{Acc}_{\\mathcal{D}_f}^*\\), \\(\\mathrm{Acc}_{\\mathcal{D}_r}^*\\) and \\(\\mathrm{MIA}^*\\) are metric values of the retrained model. A better performance of an unlearning method corresponds to a lower performance gap with retraining, measuring how close the unlearned model is to the retrained model. (iv) Run-time efficiency (RTE) [27]: evaluate the computational efficiency of the unlearning method. Specifically, we have \\(\\mathrm{RTE} = \\mathrm{T} / \\mathrm{T}_r\\), where \\(\\mathrm{T}\\) and \\(\\mathrm{T}_r\\) present the time for the unlearning methods to get the scrubbed models and the time for retraining from scratch with \\(\\mathcal{D}_r\\), respectively. An efficient unlearning method should aim for minimal computational overhead compared to the baseline. (v) Relearn time [8,20,51]: the number of epochs to fine-tune the scrubbed models for regaining performance on \\(\\mathcal{D}_f\\) (i.e., reach the original models' accuracy on \\(\\mathcal{D}_f\\))." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.235, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Scissorhands: Scrub Data Influence via Connection Sensitivity in Networks" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.319, + 0.145, + 0.684, + 0.16 + ], + "angle": 0, + "content": "Table 1: Quantitative results for forgetting \\( {10}\\% \\) data." + }, + { + "type": "table", + "bbox": [ + 0.236, + 0.171, + 0.769, + 0.606 + ], + "angle": 0, + "content": "
MethodAccDf(↓)AccDt(↑)AccDr(↑)MIA(↑)Avg. Gap
CIFAR-100Retrain75.13±0.8574.69±0.0899.98±0.0150.22±0.62-
FT [55]97.98±1.3675.28±0.1299.95±0.029.64±3.616.01
GA [52]98.00±1.3475.59±0.1198.24±1.165.00±2.2517.68
IU [29]95.67±4.8272.13±4.5896.14±4.519.43±5.9816.93
BE [4]97.94±1.3874.16±0.0998.12±1.247.60±3.0516.96
BS [4]97.65±1.4873.20±0.1897.93±1.308.24±3.2317.01
l1-sparse [27]96.35±0.6770.06±0.4696.35±0.6721.33±1.9514.59
SalUn [12]88.56±1.1871.34±0.4899.40±0.3574.66±2.4810.45
Ours68.76±1.8173.17±0.2499.24±0.3042.42±2.064.11
CIFAR-10Retrain94.81±0.5394.26±0.14100.0±0.0013.05±0.64-
FT [55]99.15±0.4693.83±0.4599.84±0.113.01±0.933.74
GA [52]99.66±0.2394.57±0.0199.62±0.250.91±0.294.42
IU [29]98.08±2.1091.91±2.7398.01±2.264.01±3.444.16
BE [4]99.41±0.3893.79±0.1599.41±0.3816.16±0.782.19
BS [4]99.60±0.2594.24±0.0799.56±0.544.46±0.333.46
l1-sparse [27]94.17±0.4990.64±0.5296.64±0.5411.87±0.612.20
SalUn [12]98.07±0.4293.92±0.2599.89±0.0717.93±0.372.15
Ours95.40±1.4892.92±0.4898.93±0.579.56±2.131.62
SVHNRetrain91.81±1.1191.17±1.7797.73±1.1215.74±1.28-
FT [55]99.60±0.2495.28±0.0499.99±0.003.85±0.46.51
GA [52]98.41±0.2392.87±0.0698.52±0.295.96±0.384.72
IU [29]92.47±1.6286.92±2.0693.36±1.8217.00±2.672.63
BE [4]98.48±0.3392.62±0.0598.44±0.296.77±0.364.45
BS [4]98.29±0.3292.48±0.0298.36±0.296.74±0.394.36
l1-sparse [27]97.79±0.1093.59±3.1899.49±0.237.57±0.294.65
SalUn [12]96.29±4.8093.59±3.1897.41±4.9624.47±4.083.99
Ours91.62±1.1893.05±0.5699.51±0.1417.62±2.541.43
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.635, + 0.509, + 0.648 + ], + "angle": 0, + "content": "4.1 Results on Classification Task" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.78 + ], + "angle": 0, + "content": "In this experiment on SVHN, CIFAR-10, and CIFAR-100, we try to forget randomly selected \\(10\\%\\) of the training data; on CelebAMask-HQ, we attempt to forget randomly selected \\(10\\%\\) identities among 307 identities. We further apply GradCAM [49] to visualize regions where models focus on \\(\\mathbf{w}/\\) and \\(\\mathbf{w} / \\mathbf{o}\\) machine unlearning algorithms. In brief, the results suggest that Scissorhands has successfully induced forgetting for the relevant samples and classes, with minor degradation in model performance over the remaining data and classes. We will discuss the results in depth below." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Sample-wise unlearning. Table 1 presents the results when forgetting randomly selected samples. Scissorhands achieves the lowest average performance gap on all the presented datasets and shows good generalization ability compared to" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.461, + 0.129 + ], + "angle": 0, + "content": "Jing Wu, Mehrtash Harandi" + }, + { + "type": "table_caption", + "bbox": [ + 0.221, + 0.145, + 0.78, + 0.16 + ], + "angle": 0, + "content": "Table 2: Quantitative results for forgetting \\( {10}\\% \\) identities on the CelebAMask-HQ." + }, + { + "type": "table", + "bbox": [ + 0.256, + 0.172, + 0.749, + 0.334 + ], + "angle": 0, + "content": "
MethodAccDf(↓)AccDt(↑)AccDr(↑)MIA(↑)Avg. Gap
Retrain0.00±0.0087.02±0.8099.96±0.01100.0±0.00-
FT [55]99.94±0.1288.59±0.5999.97±7.025.28±2.0349.06
GA [52]87.60±8.7181.22±2.1199.74±0.2651.37±5.9635.56
IU [29]88.92±10.2570.24±11.7795.27±5.0729.59±18.5945.20
BE [4]69.07±2.7344.11±2.0895.58±1.2346.24±5.9042.53
BS [4]98.18±1.9281.92±0.2799.86±0.0345.93±5.1139.36
l1-sparse [27]98.81±0.7289.37±0.7099.97±0.0076.78±5.6631.10
SalUn [12]0.00±0.0078.36±1.3496.90±1.11100.0±0.002.93
Ours1.52±2.7380.18±6.6097.20±3.8199.83±0.352.82
" + }, + { + "type": "image", + "bbox": [ + 0.26, + 0.35, + 0.49, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.35, + 0.744, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.217, + 0.458, + 0.784, + 0.473 + ], + "angle": 0, + "content": "Fig. 1: Visualizations of regions where models focus on generated by GradCAM [49]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.505, + 0.787, + 0.581 + ], + "angle": 0, + "content": "Retrain. For example, on SVHN, Scissorhands achieves \\(\\sim 91.62\\%\\) accuracy on the forgetting data \\(\\mathcal{D}_f\\), closely matching the \\(91.81\\%\\) accuracy of a fully retrained model. Furthermore, the retrained model reaches around \\(91\\%\\) accuracy on the test dataset, Scissorhands exhibits an accuracy of roughly \\(93\\%\\), showcasing the enhanced generalization of our scrubbed model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.582, + 0.788, + 0.673 + ], + "angle": 0, + "content": "When evaluating solely based on forgetting accuracy on CIFAR-10, the \\(\\ell_1\\)-sparse method might seem to be the most effective baseline, achieving approximately \\(94.17\\%\\). Nonetheless, this perceived advantage is offset by a decline in both test accuracy and the accuracy of retained data. In contrast, Scissorhands shows a good trade-off between preserving model utility on the remaining data, generalization on unseen sets, and erasing the influence of the forgetting data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Similarly, on CIFAR-100, SalUn may initially stand out with its impressive MIA accuracy of approximately \\(74.66\\%\\), relying solely on a single metric to gauge the performance of machine unlearning can be misleading, as it might not fully capture the method's effectiveness [12]. While SalUn achieves a forgetting accuracy of about \\(88.56\\%\\) and a test accuracy of \\(71.34\\%\\), Scissorhands demonstrates a slightly lower forgetting accuracy of around \\(68.76\\%\\) but achieves a higher test accuracy of \\(73.17\\%\\). This indicates that Scissorhands achieves the best trade-off between eradicating data traces and generalizing to unseen data. Soon, we will show that on this dataset, the relearning time of Scissorhands is greater than 200 to regain the performance on \\(\\mathcal{D}_f\\), which indicates that Scissorhands has efficiently scrubbed information about \\(\\mathcal{D}_f\\)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.234, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Scissorhands: Scrub Data Influence via Connection Sensitivity in Networks" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.128 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.208 + ], + "angle": 0, + "content": "Overall, Scissorhands illustrates its capability to adeptly balance the objectives of data removal, model utility preservation, and generalization to unseen data, ensuring that the influence of forgotten data is minimized without significantly compromising the overall performance and utility of the model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.231, + 0.788, + 0.367 + ], + "angle": 0, + "content": "Class-wise unlearning. Furthermore, Table 2 presents the outcomes of our analysis for the scenario where forgetting \\(10\\%\\) of identities on the CelebAMask-HQ. Images are rescaled to \\(224 \\times 224\\), and a model pre-trained with ImageNet1K [9] is employed. In this context, Scissorhands manifests the smallest average performance gap with retrained models. Notably, the Retrain method achieves perfect metrics in terms of MIA accuracy, which is \\(100\\%\\), and completely erases the identity information from the dataset, as indicated by a \\(0.00\\%\\) accuracy on the forgotten data \\(\\mathcal{D}_f\\). However, retraining proves impractical and limited for unlearning purposes in general, functioning merely as an oracle in our experiments." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.369, + 0.789, + 0.549 + ], + "angle": 0, + "content": "Among the baselines, fine-tuning (FT) and \\(\\ell_1\\)-sparse exhibit higher accuracies on the remaining data \\(\\mathcal{D}_f\\) and unseen data \\(\\mathcal{D}_t\\); yet, their accuracies on forgetting remain high as well, highlighting a less effective balance between eradicating data traces and generalizing to unseen data. In contrast, Scissorhands and SalUn demonstrate superior capabilities in effectively forgetting identities with minimal impact on the models' overall performance. SalUn achieves an MIA accuracy of \\(100\\%\\) and a forgetting accuracy of \\(0.00\\%\\), effectively nullifying the identity information similar to the retrained models. Scissorhands slightly surpasses SalUn in terms of the average performance gap with retrained models, maintaining a high MIA accuracy of \\(99.83\\%\\) and a significantly low accuracy on the forgotten data \\(\\mathcal{D}_f\\) (i.e., \\(1.52\\%\\)), while achieving a test accuracy of \\(80.18\\%\\) and an accuracy of \\(97.20\\%\\) on the retained data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.552, + 0.789, + 0.809 + ], + "angle": 0, + "content": "We further employ the GradCAM [49] to illustrate the focus area of models \\(\\mathbf{w}/\\) and \\(\\mathbf{w} / \\mathbf{o}\\) machine unlearning algorithms. As shown in Figure 1, when evaluated on the remaining data \\(\\mathcal{D}_r\\), both SalUn and Scissorhands even pay more attention to the facial features for identity classification than the retrained model. Conversely, when evaluated on the forgetting data \\(\\mathcal{D}_f\\), the scrubbed model by Scissorhands shifts its attention towards regions that are least associated with the facial features. We hypothesize that in the scenario where the task involves forgetting identities, both the remaining data and the forgetting data are face images, the scrubbed models would need a more nuanced approach to distinguish between individuals. Specifically, because the task at hand requires the discernment of subtle differences across facial features to effectively differentiate and forget specific identities while retaining others, the models adapt by developing a refined sensitivity towards those facial characteristics that are most indicative of individual identity. This adaptation enables the models to maintain high accuracy on the remaining data by focusing more intently on the facial features that are crucial for identity recognition, thereby enhancing their ability to generalize and accurately classify identities." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.81, + 0.789, + 0.841 + ], + "angle": 0, + "content": "These results underscore our proposed machine unlearning algorithm Scisserhands's superior capabilities in balancing identity forgetting and model utility" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.461, + 0.129 + ], + "angle": 0, + "content": "Jing Wu, Mehrtash Harandi" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 3: Influence of projection and initialization strategies (re-initializing parameters with uniform distribution \\(\\mathcal{U}\\), Gaussian distribution \\(\\mathcal{N}\\), and constant value of \\(\\mathbf{0} / \\mathbf{1}\\))." + }, + { + "type": "table", + "bbox": [ + 0.242, + 0.185, + 0.764, + 0.273 + ], + "angle": 0, + "content": "
ProjectionInitializationAccDf(↓)AccDt(↑)AccDr(↑)MIA(↑)Avg. Gap
✓U95.40±1.4892.92±0.4898.93±0.579.56±2.131.62
XU98.14±0.6093.14±0.1699.73±0.166.08±0.782.92
✓N94.68±2.5092.13±2.2097.94±2.7110.18±3.811.80
✓096.10±2.0592.92±0.9999.09±1.198.76±3.301.96
X184.28±0.4983.38±0.5887.18±0.9020.67±0.4110.46
" + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.288, + 0.41, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.414, + 0.288, + 0.596, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.288, + 0.781, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.403, + 0.789, + 0.434 + ], + "angle": 0, + "content": "Fig. 2: Influence of the percent value \\( k \\) in the trimming process, the balance term \\( \\lambda \\) of Equation (3), and the ratio of forgetting data used in the trimming process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.461, + 0.788, + 0.493 + ], + "angle": 0, + "content": "performance, Scissorhands not only minimizes privacy risks but also maintains the integrity and applicability of the model to unseen data." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.513, + 0.39, + 0.529 + ], + "angle": 0, + "content": "4.2 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.539, + 0.788, + 0.678 + ], + "angle": 0, + "content": "Component analysis. As detailed in Table 3, the inclusion of different procedures in Scissorhands has a noticeable impact. Notably, the employment of the projection mechanism plays a crucial role in eliminating the influence of forgetting data \\(\\mathcal{D}_f\\), affecting the overall utility of the model. The rationale behind the projection process is to preclude the reacquisition of critical information about the forgetting data \\(\\mathcal{D}_f\\) during the repair phase. Moreover, Table 3 reveals that initializing parameters with a uniform distribution \\(\\mathcal{U}\\) or opting for zero initialization facilitates effective unlearning, in contrast to re-initializing parameters with a constant value of one, which impedes finding a viable solution to Equation (5)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.844 + ], + "angle": 0, + "content": "Hyper-parameter impact. Figure 2 presents the effects of varying hyperparameters, such as the percentage value \\( k\\% \\) for the trimming process, the balance term \\( \\lambda \\) of Equation (3) in the repairing phase, and the ratio of forgetting data used in the trimming process. An increase \\( k \\) results in higher accuracies on data, as fewer parameters associated with the forgetting data \\( \\mathcal{D}_f \\) are reinitialized. Empirically, we observe choosing a value of \\( k \\geq 0.9 \\) will result in stable performances. That said, this is a hyperparameter of the algorithm and can be chosen by measuring the performances on \\( D_{f} \\) and \\( D_{r} \\) during training. The hyper-parameter \\( \\lambda \\) controls the balance between retaining and forgetting data, the accuracies on test data and remaining data show a peak at \\( \\lambda = 0.05 \\)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.235, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Scissorhands: Scrub Data Influence via Connection Sensitivity in Networks" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.146, + 0.505, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.146, + 0.777, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.286, + 0.788, + 0.342 + ], + "angle": 0, + "content": "Fig.3: Quantity of nudity content detected using the NudeNet classifier from 1K sampled images and I2P data. We observed a high false positive rate for exposed female genitalia/breast using the NudeNet classifier on generated I2P images. The flagged images can be found in Appendix B." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.356, + 0.264, + 0.366 + ], + "angle": 0, + "content": "SD v1.4" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.366, + 0.408, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.411, + 0.357, + 0.44, + 0.366 + ], + "angle": 0, + "content": "SalUn" + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.366, + 0.593, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.595, + 0.358, + 0.62, + 0.366 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.595, + 0.366, + 0.648, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.657, + 0.357, + 0.78, + 0.366 + ], + "angle": 0, + "content": "\\(\\star\\) Added by authors for publication" + }, + { + "type": "image", + "bbox": [ + 0.657, + 0.366, + 0.779, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.426, + 0.785, + 0.455 + ], + "angle": 0, + "content": "Fig. 4: Sample images with the prompt from \\( c_f = \\{ \\text{nudity}, \\text{naked}, \\text'erotic', \\text'sexual' \\} \\) generated by SDs w/ and w/o machine unlearning algorithms. Best viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.485, + 0.789, + 0.651 + ], + "angle": 0, + "content": "when forgetting \\(10\\%\\) data on CIFAR-10. Generally, as \\(\\lambda\\) escalates, it places greater emphasis on the process of data omission, leading to a decrease in forgetting accuracy. As the ratio \\(p\\) increases, all accuracy metrics exhibit a decline, suggesting that employing more forgetting data for the purpose of forgetting in order to isolate critical knowledge about the forgetting data also inadvertently results in the loss of pertinent information regarding the remaining data. This phenomenon occurs particularly when the data designated for forgetting shares the same distribution as the data that is retained. In essence, while aiming to enhance the specificity of the forgetting process by pinpointing essential details about the data to be forgotten, there's a risk of simultaneously diminishing knowledge pertinent to the retained data." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.676, + 0.499, + 0.691 + ], + "angle": 0, + "content": "4.3 Case Study: Stable Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.809 + ], + "angle": 0, + "content": "We also employed our algorithm to mitigate the generation of inappropriate content in Stable Diffusion (SD) models [44]. Specifically, our primary objective is to effectively eliminate the concept of nudity from the model's generative capabilities. To this end, followed [26] and employed SD v1.4 for sampling with 50 time steps. We evaluate on 1K generated images with prompts \\( c_f = \\{ \\text{nudity}, \\text{naked}, \\text'erotic', \\text'sexual' \\} \\) and 4703 generated images with I2P [46] using the open-source NudeNet classifier." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.789, + 0.84 + ], + "angle": 0, + "content": "Figures 3 and 4 present the results for SD v2.1 (trained on a dataset filtered for nudity), SalUn, and Scissorhands. In Figure 3, y-axis denotes the number of" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.461, + 0.129 + ], + "angle": 0, + "content": "Jing Wu, Mehrtash Harandi" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.33 + ], + "angle": 0, + "content": "exposed body parts generated by the SD v1.4 model, it presents the percentage change in exposed body parts w.r.t. SD v1.4. When comparing on I2P data, our algorithms outperform SalUn on 'Feet', while coming short on 'Female Breast and Genitalia'. While this could suggest the superiority of SalUn on this data, the gap is not that big in reality, as out of the 9 images flagged for containing exposed female breasts, 7 were inaccurately identified, and none of the flagged images depicted exposed female genitalia. The use of the NudeNet classifier on generated I2P images exhibits a significant rate of false positives, as highlighted in [26] as well. When we use prompts \\( c_f \\) explicitly associated with nudity, no exposed sensitive content is detected for both Scissorhands and SalUn methods. We can conclude that, Scissorhands significantly reduces the amount of nudity content compared to SD v1.4, SD v2.1, showing comparable results to SalUn." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.35, + 0.566, + 0.368 + ], + "angle": 0, + "content": "5 Limitations and Broader Impact" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.381, + 0.789, + 0.593 + ], + "angle": 0, + "content": "Our proposed unlearning algorithm, Scissorhands, can achieve comparable performance to state-of-the-art unlearning methods. We addressed potential issues of conflicting gradients arising from direct conflicts between features or patterns in the forgetting and remaining data, ensuring training stability. However, the limitation is the incomplete concept erasure shown in Figure 8 in Appendix B, and the potential for biased models due to the selective deletion of information. This is a crucial consideration, especially when unlearning algorithms are applied with malicious intent. Furthermore, in experiments on stable diffusion, while our method can successfully erase specific concepts, finding an optimal balance between forgetting and maintaining high image quality remains challenging. There is also a risk that unlearning algorithms might be used to alter or remove concepts in unethical or malicious ways. It is our hope that Scissorhands offers a fresh perspective on practical and effective machine unlearning, emphasizing the need for ethical guidelines to prevent misuse in its application." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.614, + 0.36, + 0.63 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this work, we proposed Scissorhands, an effective and practical machine unlearning algorithm. Our algorithm identified the important parameters w.r.t. the forgetting data in single-shot via connection sensitivity, then re-initialized these parameters for scrubbing the influence of the forgetting data. An effective gradient projection-based technique is applied to enhance the model utility while excluding the information w.r.t. the forgetting data. Our evaluations on various datasets showed that, compared with other methods, our approach offers superior performance across image classification and image generation tasks. Future works could explore how Scissorhands performs on regression and NLP tasks, and further investigation into the scenario where data are not available, as well as unlearning with sequential/time-series data where could introduce unique challenges, e.g., the need to consider temporal dependencies, which may require new frameworks specifically tailored for sequential data." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.235, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Scissorhands: Scrub Data Influence via Connection Sensitivity in Networks" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.207 + ], + "angle": 0, + "content": "Mehrtash Harandi is supported by funding from the Australian Research Council (ARC) Discovery Program DP230101176." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.23, + 0.323, + 0.245 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.26, + 0.786, + 0.302 + ], + "angle": 0, + "content": "1. Birhane, A., Prabhu, V.U.: Large image datasets: A pyrrhic win for computer vision? In: 2021 IEEE Winter Conference on Applications of Computer Vision (WACV). pp. 1536-1546. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.302, + 0.786, + 0.344 + ], + "angle": 0, + "content": "2. Bourtoule, L., Chandrasekaran, V., Choquette-Choo, C.A., Jia, H., Travers, A., Zhang, B., Lie, D., Papernot, N.: Machine unlearning. In: 2021 IEEE Symposium on Security and Privacy (SP). pp. 141-159 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.344, + 0.786, + 0.372 + ], + "angle": 0, + "content": "3. Cao, Y., Yang, J.: Towards making systems forget with machine unlearning. In: 2015 IEEE Symposium on Security and Privacy (SP). pp. 463-480 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.372, + 0.786, + 0.426 + ], + "angle": 0, + "content": "4. Chen, M., Gao, W., Liu, G., Peng, K., Wang, C.: Boundary unlearning: Rapid forgetting of deep networks via shifting the decision boundary. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7766-7775 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.426, + 0.786, + 0.468 + ], + "angle": 0, + "content": "5. Chen, M., Zhang, Z., Wang, T., Backes, M., Humbert, M., Zhang, Y.: Graph unlearning. In: Proceedings of the 2022 ACM SIGSAC Conference on Computer and Communications Security. pp. 499-513 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.468, + 0.786, + 0.51 + ], + "angle": 0, + "content": "6. Chen, W., Song, D., Li, B.: Trojdiff: Trojan attacks on diffusion models with diverse targets. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4035-4044 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.51, + 0.786, + 0.551 + ], + "angle": 0, + "content": "7. Cheng, J., Dasoulas, G., He, H., Agarwal, C., Zitnik, M.: Gnndelete: A general strategy for unlearning in graph neural networks. In: International Conference on Learning Representations (ICLR) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.551, + 0.786, + 0.579 + ], + "angle": 0, + "content": "8. Chandawat, V.S., Tarun, A.K., Mandal, M., Kankanhalli, M.: Zero-shot machine unlearning. IEEE Transactions on Information Forensics and Security (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.579, + 0.786, + 0.62 + ], + "angle": 0, + "content": "9. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.62, + 0.721, + 0.634 + ], + "angle": 0, + "content": "0. Dennis, C.R., Weisberg, S.: Residuals and influence in regression (1982)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.634, + 0.786, + 0.661 + ], + "angle": 0, + "content": "1. Fan, C., Liu, J., Hero, A., Liu, S.: Challenging forgets: Unveiling the worst-case forget sets in machine unlearning. arXiv preprint arXiv:2403.07362 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.661, + 0.786, + 0.716 + ], + "angle": 0, + "content": "2. Fan, C., Liu, J., Zhang, Y., Wei, D., Wong, E., Liu, S.: Salun: Empowering machine unlearning via gradient-based weight saliency in both image classification and generation. In: International Conference on Learning Representations (ICLR) (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.716, + 0.786, + 0.744 + ], + "angle": 0, + "content": "3. Foster, J., Schoepf, S., Brintrup, A.: Fast machine unlearning without retraining through selective synaptic dampening. arXiv preprint arXiv:2308.07707 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.744, + 0.786, + 0.799 + ], + "angle": 0, + "content": "4. Gandhi, S., Kokkula, S., Chaudhuri, A., Magnani, A., Stanley, T., Ahmadi, B., Kandaswamy, V., Ovenc, O., Mannor, S.: Scalable detection of offensive and noncompliant content/logo in product images. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2247-2256 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.786, + 0.841 + ], + "angle": 0, + "content": "5. Gandikota, R., Materzynska, J., Fiotto-Kaufman, J., Bau, D.: Erasing concepts from diffusion models. In: 2023 IEEE International Conference on Computer Vision (ICCV) (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.26, + 0.786, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.46, + 0.129 + ], + "angle": 0, + "content": "Jing Wu, Mehrtash Harandi" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "16. Gandikota, R., Orgad, H., Belinkov, Y., Materzyńska, J., Bau, D.: Unified concept editing in diffusion models. arXiv preprint arXiv:2308.14761 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.785, + 0.217 + ], + "angle": 0, + "content": "17. Ginart, A., Guan, M., Valiant, G., Zou, J.Y.: Making ai forget you: Data deletion in machine learning. In: Advances in Neural Information Processing Systems (NeurIPS). vol. 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.785, + 0.258 + ], + "angle": 0, + "content": "18. Goel, S., Prabhu, A., Sanyal, A., Lim, S.N., Torr, P., Kumaraguru, P.: Towards adversarial evaluations for inexact machine unlearning. arXiv preprint arXiv:2201.06640 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.259, + 0.785, + 0.3 + ], + "angle": 0, + "content": "19. Golatkar, A., Achille, A., Ravichandran, A., Polito, M., Soatto, S.: Mixed-privacy forgetting in deep networks. In: 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 792-801 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.3, + 0.785, + 0.342 + ], + "angle": 0, + "content": "20. Golatkar, A., Achille, A., Soatto, S.: Eternal sunshine of the spotless net: Selective forgetting in deep networks. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9301-9309 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.785, + 0.384 + ], + "angle": 0, + "content": "21. Golatkar, A., Achille, A., Soatto, S.: Forgetting outside the box: Scrubbing deep networks of information accessible from input-output observations. In: European Conference on Computer Vision (ECCV). pp. 383-398. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.412 + ], + "angle": 0, + "content": "22. Goldman, E.: An introduction to the california consumer privacy act (ccpa). Santa Clara Univ. Legal Studies Research Paper (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.466 + ], + "angle": 0, + "content": "23. Guo, C., Goldstein, T., Hannun, A., Van Der Maaten, L.: Certified data removal from machine learning models. In: Proceedings of the 37th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 119, pp. 3832-3842. PMLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.466, + 0.785, + 0.508 + ], + "angle": 0, + "content": "24. Halimi, A., Kadhe, S., Rawat, A., Baracaldo, N.: Federated unlearning: How to efficiently erase a client in fl? In: International conference on machine learning. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.508, + 0.785, + 0.536 + ], + "angle": 0, + "content": "25. Han, S., Pool, J., Tran, J., Dally, W.: Learning both weights and connections for efficient neural network. vol. 28 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.785, + 0.577 + ], + "angle": 0, + "content": "26. Heng, A., Soh, H.: Selective amnesia: A continual learning approach to forgetting in deep generative models. In: Advances in Neural Information Processing Systems (NeurIPS) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.577, + 0.785, + 0.619 + ], + "angle": 0, + "content": "27. Jia, J., Liu, J., Ram, P., Yao, Y., Liu, G., Liu, Y., Sharma, P., Liu, S.: Model sparsification can simplify machine unlearning. In: Advances in Neural Information Processing Systems (NeurIPS) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.785, + 0.66 + ], + "angle": 0, + "content": "28. Karasuyama, M., Takeuchi, I.: Multiple incremental decremental learning of support vector machines. IEEE Transactions on Neural Networks 21(7), 1048-1059 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.66, + 0.785, + 0.689 + ], + "angle": 0, + "content": "29. Koh, P.W., Liang, P.: Understanding black-box predictions via influence functions. In: International conference on machine learning. pp. 1885-1894. PMLR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.785, + 0.716 + ], + "angle": 0, + "content": "30. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.785, + 0.758 + ], + "angle": 0, + "content": "31. Lee, C.H., Liu, Z., Wu, L., Luo, P.: Maskgan: Towards diverse and interactive facial image manipulation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5549-5558 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "32. Lee, N., Ajanthan, T., Gould, S., Torr, P.H.: A signal propagation perspective for pruning neural networks at initialization. In: International Conference on Learning Representations (ICLR) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.841 + ], + "angle": 0, + "content": "33. Lee, N., Ajanthan, T., Torr, P.H.: Snip: Single-shot network pruning based on connection sensitivity. In: International Conference on Learning Representations (ICLR) (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.235, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Scissorhands: Scrub Data Influence via Connection Sensitivity in Networks" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "34. Liu, G., Ma, X., Yang, Y., Wang, C., Liu, J.: Federaser: Enabling efficient client-level data removal from federated learning models. In: 2021 IEEE/ACM 29th International Symposium on Quality of Service (IWQOS). pp. 1-10 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.231 + ], + "angle": 0, + "content": "35. Liu, Y., Xu, L., Yuan, X., Wang, C., Li, B.: The right to be forgotten in federated learning: An efficient realization with rapid retraining. In: IEEE INFOCOM 2022 - IEEE Conference on Computer Communications. pp. 1749-1758 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.259 + ], + "angle": 0, + "content": "36. Lopez-Paz, D., Ranzato, M.: Gradient episodic memory for continual learning. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.785, + 0.3 + ], + "angle": 0, + "content": "37. Mehta, R., Pal, S., Singh, V., Ravi, S.N.: Deep unlearning via randomized conditionally independent hessians. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10412-10421 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.301, + 0.785, + 0.355 + ], + "angle": 0, + "content": "38. Neel, S., Roth, A., Sharifi-Malvajerdi, S.: Descent-to-delete: Gradient-based methods for machine unlearning. In: Proceedings of the 32nd International Conference on Algorithmic Learning Theory. pp. 931-962. Proceedings of Machine Learning Research, PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.785, + 0.383 + ], + "angle": 0, + "content": "39. Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B., Ng, A.Y.: Reading digits in natural images with unsupervised feature learning (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.411 + ], + "angle": 0, + "content": "40. Nguyen, T.T., Huynh, T.T., Nguyen, P.L., Liew, A.W.C., Yin, H., Nguyen, Q.V.H.: A survey of machine unlearning. arXiv preprint arXiv:2209.02299 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.452 + ], + "angle": 0, + "content": "41. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.479 + ], + "angle": 0, + "content": "42. Peste, A., Alistarh, D., Lampert, C.H.: Ssse: Efficiently erasing samples from trained machine learning models. arXiv preprint arXiv:2107.03860 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.508 + ], + "angle": 0, + "content": "43. Rando, J., Paleka, D., Lindner, D., Heim, L., Tramér, F.: Red-teaming the stable diffusion safety filter. arXiv preprint arXiv:2210.04610 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.785, + 0.55 + ], + "angle": 0, + "content": "44. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.551, + 0.785, + 0.591 + ], + "angle": 0, + "content": "45. Romero, E., Barrio, I., Belanche, L.: Incremental and decremental learning for linear support vector machines. In: International Conference on Artificial Neural Networks. pp. 209-218. Springer (2007)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.785, + 0.646 + ], + "angle": 0, + "content": "46. Schramowski, P., Brack, M., Deiseroth, B., Kersting, K.: Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22522-22531 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.702 + ], + "angle": 0, + "content": "47. Schramowski, P., Tauchmann, C., Kersting, K.: Can machines help us answering question 16 in datasheets, and in turn reflecting on inappropriate content? In: Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency. pp. 1350-1361 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.785, + 0.742 + ], + "angle": 0, + "content": "48. Sekhari, A., Acharya, J., Kamath, G., Suresh, A.T.: Remember what you want to forget: Algorithms for machine unlearning. Advances in Neural Information Processing Systems (NeurIPS) 34, 18075-18086 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.743, + 0.785, + 0.799 + ], + "angle": 0, + "content": "49. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Gradcam: Visual explanations from deep networks via gradient-based localization. In: 2017 IEEE International Conference on Computer Vision (ICCV). pp. 618-626 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.841 + ], + "angle": 0, + "content": "50. Tarun, A.K., Chundawat, V.S., Mandal, M., Kankanhalli, M.: Fast yet effective machine unlearning. IEEE Transactions on Neural Networks and Learning Systems (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.46, + 0.129 + ], + "angle": 0, + "content": "Jing Wu, Mehrtash Harandi" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.188 + ], + "angle": 0, + "content": "51. Tarun, A.K., Chundawat, V.S., Mandal, M., Kankanhalli, M.: Deep regression unlearning. In: International Conference on Machine Learning. pp. 33921-33939. PMLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.189, + 0.785, + 0.231 + ], + "angle": 0, + "content": "52. Thudi, A., Deza, G., Chandrasekaran, V., Papernot, N.: Unrolling sgd: Understanding factors influencing machine unlearning. In: 2022 IEEE 7th European Symposium on Security and Privacy (EuroS&P). pp. 303-319. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.231, + 0.785, + 0.272 + ], + "angle": 0, + "content": "53. Voigt, P., Von dem Bussche, A.: The eu general data protection regulation (gdpr). A Practical Guide, 1st Ed., Cham: Springer International Publishing 10(3152676), 10-5555 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.272, + 0.785, + 0.3 + ], + "angle": 0, + "content": "54. Wang, J., Guo, S., Xie, X., Qi, H.: Federated unlearning via class-discriminative pruning. In: Proceedings of the ACM Web Conference 2022. pp. 622-632 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.785, + 0.328 + ], + "angle": 0, + "content": "55. Warnecke, A., Pirch, L., Wressnegger, C., Rieck, K.: Machine unlearning of features and labels. arXiv preprint arXiv:2108.11577 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.356 + ], + "angle": 0, + "content": "56. Wu, C., Zhu, S., Mitra, P.: Federated unlearning with knowledge distillation. arXiv preprint arXiv:2201.09441 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.785, + 0.383 + ], + "angle": 0, + "content": "57. Wu, J., Le, T., Hayat, M., Harandi, M.: Erasediff: Erasing data influence in diffusion models. arXiv preprint arXiv:2401.05779 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.383, + 0.785, + 0.438 + ], + "angle": 0, + "content": "58. Wu, Y., Dobriban, E., Davidson, S.: DeltaGrad: Rapid retraining of machine learning models. In: Proceedings of the 37th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 119, pp. 10355-10366. PMLR (13-18 Jul 2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.481 + ], + "angle": 0, + "content": "59. Zhang, C., Bengio, S., Hardt, M., Recht, B., Vinyals, O.: Understanding deep learning (still) requires rethinking generalization. In: International Conference on Learning Representations (ICLR) (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.508 + ], + "angle": 0, + "content": "60. Zhang, E., Wang, K., Xu, X., Wang, Z., Shi, H.: Forget-me-not: Learning to forget in text-to-image diffusion models. arXiv preprint arXiv:2303.17591 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.508, + 0.785, + 0.55 + ], + "angle": 0, + "content": "61. Zhang, Y., Zhang, Y., Yao, Y., Jia, J., Liu, J., Liu, X., Liu, S.: Unlearncanvas: A stylized image dataset to benchmark machine unlearning for diffusion models. arXiv preprint arXiv:2402.11846 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.591 + ], + "angle": 0, + "content": "62. Zhang, Y., Chen, X., Jia, J., Zhang, Y., Fan, C., Liu, J., Hong, M., Ding, K., Liu, S.: Defensive unlearning with adversarial training for robust concept erasure in diffusion models. arXiv preprint arXiv:2405.15234 (2024)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.591 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Scissorhands_ Scrub Data Influence via Connection Sensitivity in Networks/43a51c48-5759-4f46-b600-93f450455ce7_origin.pdf b/2024/Scissorhands_ Scrub Data Influence via Connection Sensitivity in Networks/43a51c48-5759-4f46-b600-93f450455ce7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f87f4242500a352646f592ad5c73160849af3327 --- /dev/null +++ b/2024/Scissorhands_ Scrub Data Influence via Connection Sensitivity in Networks/43a51c48-5759-4f46-b600-93f450455ce7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b0960dc423a0ba56ac2188b5d4f05851ad92727fab4a96ac30d9e8b1fdb65008 +size 1501477 diff --git a/2024/Scissorhands_ Scrub Data Influence via Connection Sensitivity in Networks/full.md b/2024/Scissorhands_ Scrub Data Influence via Connection Sensitivity in Networks/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b9046ca1041a5e9fc6b68b81e66d4d62bd196ee0 --- /dev/null +++ b/2024/Scissorhands_ Scrub Data Influence via Connection Sensitivity in Networks/full.md @@ -0,0 +1,296 @@ +# Scissorhands: Scrub Data Influence via Connection Sensitivity in Networks + +Jing Wu1, Mehrtash Harandi1 + +$^{1}$ Department of Electrical and Computer Systems Engineering Monash University, Australia {jing.wu1, mehrtash.harandi}@monash.edu + +Abstract. Machine unlearning has become a pivotal task to erase the influence of data from a trained model. It adheres to recent data regulation standards and enhances the privacy and security of machine learning applications. In this work, we present a new machine unlearning approach Scissorhands. Initially, Scissorhands identifies the most pertinent parameters in the given model relative to the forgetting data via connection sensitivity. By reinitializing the most influential top- $k$ percent of these parameters, a trimmed model for erasing the influence of the forgetting data is obtained. Subsequently, Scissorhands fine-tunes the trimmed model with a gradient projection-based approach, seeking parameters that preserve information on the remaining data while discarding information related to the forgetting data. Our experimental results, conducted across image classification and image generation tasks, demonstrate that Scissorhands, showcases competitive performance when compared to existing methods. Source code is available at https://github.com/JingWu321/Scissorhands. + +Keywords: Machine unlearning $\cdot$ Connection sensitivity $\cdot$ Diffusion model + +Warning: This paper contains explicit sexual imagery that may be offensive. + +# 1 Introduction + +In this work, we aim to propose an effective machine unlearning method. Under data regulations like the European Union General Data Protection Regulation (GDPR) [53] and the California Consumer Privacy Act (CCPA) [22], all users are granted the right to be forgotten. In machine learning, these legal provisions empower data owners with the right not only to withdraw their data from trained models but also to ensure that their data's influence on these models is erased. + +The most direct approach to accomplish this objective is to retrain the model from scratch, excluding the data requested for deletion from the training process. Retraining from scratch is typically considered the gold standard in the field of machine unlearning [12, 52]. Yet, this poses a challenge as numerous in-production models require prolonged training periods and substantial computing resources. While retraining is feasible, it is often impractical. + +Consequently, the proposal and development of efficient approximate unlearning methods [13, 23, 26, 50, 51] have become essential. Currently, most approximate unlearning techniques achieve forgetting by adding normally distributed noise to the parameters [19-21] or by estimating the influence of a particular data point on the model's predictions [23, 37, 38, 48] based on the influence function method [10]. Jia et al. [27] demonstrate that model sparsity can help to unlearn and fuse model sparsity into the unlearning process. + +Fan et al. [12] recently raised concerns regarding the instability in approximate unlearning methods, and highlighted that the current machine unlearning approaches, initially crafted for image classification tasks, fall short in effectively tackling the challenges of machine unlearning within the realm of image generation. The authors introduce a novel unlearning method SalUn that can effectively perform forgetting in both image classification and generation tasks. Their unlearning mechanism includes finding salient weights and then fine-tuning these weights using forgetting data assigned with random labels. However, studies like [59] show that the model can still learn the true data distribution even using data with random labels. While SalUn has shown state-of-the-art performance, it could potentially memorize information about forgetting data points, albeit associating them with random labels but still can be undesired in sensitive applications. In § 4, results also show that SalUn tends to memorize knowledge about the forgetting data. + +This work. We present an unlearning approach, Scissorhands, designed to erase data influence in the classification models and eliminate a particular concept from a text-to-image model. To achieve this, inspired by [27], our key insight is to first erase the critical influence of forgetting data in models and then relearn. Unlike previous methods that employ the forgetting data with noisy labels as a part of relearning, we propose to use an efficient gradient projection method to relearn the critical features and patterns while ensuring the exclusion of influences associated with the forgetting data. Through a series of experiments and evaluations, including the classification on SVHN [39], CIFAR-10 and CIFAR-100 [30], CelebAMASK-HQ [31] datasets, as well as the open-source stable diffusion [44] text-to-image model, results demonstrate the viability and effectiveness of our technique in forgetting the influence of random samples, discrete classes, and sensitive content such as nudity. + +# 2 Methodology + +In this section, we propose Scissorhands, our unlearning framework that scrubs data from a model by eliminating their influence on model parameters. Throughout the paper, we denote scalars and vectors/matrices by lowercase and bold symbols, respectively (e.g., $a$ , $a$ , and $A$ ). We show a $d$ -dimensional vector of ones by $\mathbf{1}_d$ , and use $\odot$ to denote the Hadamard product. + +Overview. Consider a model $f$ with parameters $\pmb{\theta} \in \mathbb{R}^d$ trained on a dataset $\mathcal{D} = \{\mathbf{x}_i, \mathbf{y}_i\}_{i=1}^N$ . Suppose a subset of $\mathcal{D}$ , denoted as $\mathcal{D}_f$ , is requested for deletion. + +The remaining data is defined as $\mathcal{D}_r\coloneqq \mathcal{D}\setminus \mathcal{D}_f$ . Our objective is to develop the unlearning algorithm to generate a scrubbed model $f_{u}$ , effectively removing the influence of $\mathcal{D}_f$ from model $f$ , while maintaining its utility over $\mathcal{D}_r$ . + +In Scissorhands, we initially identify critical parameters w.r.t. the forgetting data via connection sensitivity analysis [32, 33]. We then reinitialize the top- $k$ percent of these key parameters, resulting in a trimmed model with parameters $\pmb{\theta}_t$ . This step aims to diminish the model's memory of the forgetting data $\mathcal{D}_f$ . Nevertheless, this process runs the risk of erasing valuable information from the data set $\mathcal{D}_r$ that we aim to retain. + +To address this, the next phase of Scissorhands concentrates on restoring the performance of the trimmed model on remaining data $\mathcal{D}_r$ , while ensuring that the influence of the forgetting data $\mathcal{D}_f$ remains excluded. We accomplish this balance and obtain the final scrubbed model with parameters $\theta^{\star}$ through a gradient projection-based approach. + +# 2.1 Trimming + +To effectively identify salient parameters of a network with respect to $\mathcal{D}_f$ , one has to establish a good criterion to determine such salient connections. A commonly used criterion is the magnitude of the weights, with those exceeding a certain threshold deemed salient as suggested in [25]. Our requirement is more nuanced, as we need a measure that specifically discerns saliency based on $\mathcal{D}_f$ . Therefore, we adopt a single-shot approach as proposed in [33], which we outline briefly to ensure clarity and completeness in our methodology. Consider + +$$ +\mathbf {s} _ {j} (\mathcal {D}) := \mathbb {E} _ {\boldsymbol {x}, \boldsymbol {y} \sim \mathcal {D}} \left[ \ell (\boldsymbol {\theta}; \boldsymbol {x}, \boldsymbol {y}) - \ell \left(\left(\mathbf {1} _ {d} - \boldsymbol {e} _ {j}\right) \odot \boldsymbol {\theta}; \boldsymbol {x}, \boldsymbol {y}\right) \right], \tag {1} +$$ + +which measures the influence of parameter $j \in \{1, \dots, d\}$ on a model in terms of the empirical risk for a given dataset $\mathcal{D}$ . Here, $\ell : \mathbb{R}^d \times \mathcal{X} \times \mathcal{Y} \to \mathbb{R}_+$ is the loss of the model and $e_j$ is the indicator vector of element $j$ (i.e., a binary vector with its $j$ -th component set to one). Note that computing $\mathbf{s}_j$ for each $j \in \{1, \dots, d\}$ is prohibitively expensive, as it demands $d + 1$ forward passes through the dataset. An approximation to Equation (1) [25] is in the form of + +$$ +\mathbf {s} _ {j} (\mathcal {D}) \approx \mathbb {E} _ {\boldsymbol {x}, \boldsymbol {y} \sim \mathcal {D}} \left[ \frac {\partial \ell (\boldsymbol {\theta} ; \boldsymbol {x} , \boldsymbol {y})}{\partial \boldsymbol {\theta} _ {j}} \boldsymbol {\theta} _ {j} \right]. \tag {2} +$$ + +Detailed proofs can be found in Appendix A. Equation (2) defines the sensitivity of parameter $j$ as the average (i.e., expectation) of a product, the gradient of the loss and the current value of the parameter. As such, $\mathbf{s}_j$ encodes the loss sensitivity according to the parameter's magnitude. Intuitively, a parameter with a large value that significantly affects the loss, as captured by the gradient, is considered more influential or salient. The form in Equation (2) offers several benefits, most importantly, the ability to calculate the sensitivity for all parameters with just a single sweep through the dataset. + +To scrub the influence of the forgetting data $\mathcal{D}_f$ in the model, we first obtain $\mathfrak{s}_j(\mathcal{D}_f)$ , saliency of parameters w.r.t. the forgetting data. We then re-initialize + +the top- $k\%$ of the parameters based on their saliency rankings. This process is akin to performing a targeted 'lobotomy' on the model specifically concerning the data $\mathcal{D}_f$ , thereby selectively erasing their influence. Unfortunately, this aggressive approach of reinitializing parameters can detrimentally impact the model's performance on the remaining dataset $\mathcal{D}_r$ , which we will correct in the next phase of Scissorhands. + +Remark 1. In principle, Scissorhands can benefit from any algorithms that can identify important connections w.r.t. $\mathcal{D}_f$ . The main reason behind choosing SNIP is its single-shot nature. As presented in the study [12], the gradient of loss w.r.t. the model parameters over the forgetting data can help identify important parameters w.r.t. $\mathcal{D}_f$ and hence help with erasing. + +Remark 2. In practice, we have observed that even a small subset of $\mathcal{D}_f$ can be sufficient for the trimming. For example, to unlearn a large set of data on CIFAR-100, our algorithm outperforms baselines by utilizing approximately $3\%$ of the forgetting data for trimming. + +Remark 3. Initializing weights of the trimmed neurons can take various forms. Empirically, we observed initializing with a uniform distribution is particularly effective. Details on the influence of different initialization strategies and the choice of $\mathrm{k}\%$ trimming on the model's performance will be discussed in $\S 4.2$ . + +# 2.2 Repairing + +Following the parameter-trimming process, aimed at mitigating the influence of the forgetting data $\mathcal{D}_f$ , a challenge presents itself: the potential erasure of crucial information associated with the remaining data $\mathcal{D}_r$ . A straightforward solution to recover the model utility is to relearn the model using the remaining data. However, this approach risks biasing the model towards features of the remaining data and inadvertently reintegrating information about the forgetting samples. In the quest to ensure that models retain essential information from the remaining data $\mathcal{D}_r$ while effectively forgetting $\mathcal{D}_f$ , our strategy involves maximizing the loss over the forgetting data $\mathcal{D}_f$ while concurrently minimizing the loss over the remaining data $\mathcal{D}_r$ . Therefore, to achieve the balance between these goals, we propose the objective of an efficient practice for unlearning: + +$$ +L (\boldsymbol {\theta}, \mathcal {D} _ {r}, \mathcal {D} _ {f}) := \mathcal {L} (\boldsymbol {\theta}; \mathcal {D} _ {r}) - \lambda \mathcal {L} (\boldsymbol {\theta}; \mathcal {D} _ {f}). \tag {3} +$$ + +We aim to optimize Equation (3) w.r.t. $\pmb{\theta}$ through gradient descent, ensuring minimal reintroduction of information about $\mathcal{D}_f$ . This is particularly challenging when there are similarities between samples in $\mathcal{D}_f$ and $\mathcal{D}_r$ . To achieve this desideratum, it is vital that updates to the model do not improve (i.e., reduce) $\mathcal{L}(\pmb{\theta};\mathcal{D}_f)$ . To this end, we define $\pmb{g}_f = \nabla_{\pmb{\theta}}\mathcal{L}(\pmb{\theta};\mathcal{D}_f)$ as the gradient direction for $\mathcal{D}_f$ and $\pmb{g}_o = \nabla_{\pmb{\theta}}L(\pmb{\theta},\mathcal{D}_r,\mathcal{D}_f)$ as the gradient direction for optimizing Equation (3). The optimal descent direction $\pmb{g} \in \mathbb{R}^d$ should thus satisfy two criteria: 1. It should exhibit maximum similarity to $\pmb{g}_o$ to ensure swift convergence. The + +Algorithm 1 The procedure of Scissorhands. +1: procedure TRIMMING $(\pmb {\theta},\mathcal{D}_f)$ 11: for $e = 0$ to $E - 1$ do +2: $\pmb {X},\pmb {y}\gets \{\pmb {x}_i,\pmb {y}_i\sim \mathcal{D}_f|i\in [B]\} .$ 12: Get $L(\pmb {\theta},\mathcal{D}_r,\mathcal{D}_f)$ (cf. Eq. (3)). +3: // Compute connection sensitivity 13: Get $\pmb{g_o}$ and $\pmb{g_f}$ +4: Get $\mathbf{s}_j(\pmb {\theta};\pmb {X},\pmb {y}),\forall j\in [1,d]$ 14: //Get the optimal direction using Eq.(2). 15: if $\langle g_o,g_f\rangle >0$ then +5: // Re-initialization 16: Compute $v^{*}$ (cf. Eq. (5)). +6: Re-initialize parameters according 17: $\pmb {g} = \pmb {g}_o - v^*\pmb {g}_f$ to the top- $k \%$ value of $\mathfrak{s}_j$ 18: $\pmb {g}_o\gets \pmb {g}$ +7: return $\pmb{\theta}_{t}$ 19: end if +8: end procedure 20: $\pmb{\theta}^{e + 1}\leftarrow \pmb{\theta}^{e} - \eta \pmb{g}_{o}$ +21: end for +9: procedure REPAIRING $(\pmb {\theta}_t,\mathcal{D}_f,\mathcal{D}_r)$ 22: return $\pmb{\theta}^{*} = \pmb{\theta}^{E}$ +10: $\pmb{\theta}^{0} = \pmb{\theta}_{t}$ 23: end procedure + +notion of similarity can be captured by $\pmb{g}_o$ using $\| \pmb{g} - \pmb{g}_o \|_2^2$ . 2. It should not align with $\pmb{g}_f$ to prevent improving $\mathcal{L}(\pmb{\theta}; \mathcal{D}_f)$ . This can be mathematically formalized as $\langle \pmb{g}_f, \pmb{g} \rangle \leq 0$ . Consequently, after computing $\pmb{g}_o = \nabla_{\pmb{\theta}} L(\pmb{\theta}, \mathcal{D}_r, \mathcal{D}_f)$ and $\pmb{g}_f = \nabla_{\pmb{\theta}} \mathcal{L}(\pmb{\theta}; \mathcal{D}_f)$ in each iteration, we formulate the following optimization challenge to identify the model update direction: + +$$ +\begin{array}{l} \underset {\boldsymbol {g}} {\arg \min} \frac {1}{2} \| \boldsymbol {g} - \boldsymbol {g} _ {o} \| _ {2} ^ {2}, \\ \text {s . t .} \quad \langle \boldsymbol {g} _ {f}, \boldsymbol {g} \rangle \leq 0. \tag {4} \\ \end{array} +$$ + +This problem may be addressed using the Frank-Wolfe algorithm. Given the high dimensionality of gradients in neural networks, directly addressing the constraint optimization in Equation (4) could become overwhelming. Inspired by [36], we propose to make use of the dual formulation of the problem Equation (4) as: + +$$ +\begin{array}{l} \sup _ {v} \frac {1}{2} \boldsymbol {g} _ {f} ^ {\top} \boldsymbol {g} _ {f} v ^ {2} - \boldsymbol {g} _ {o} ^ {\top} \boldsymbol {g} _ {f} v, \\ \begin{array}{l l} \text {s . t .} & v \geq 0. \end{array} \tag {5} \\ \end{array} +$$ + +Then, the optimal descent direction $\pmb{g}$ is given from the solution $v^{*}$ in Equation (5) as $\pmb {g} = \pmb{g}_{o} - v^{*}\pmb{g}_{f}$ . Detailed proofs can be found in Appendix A. By doing so, we ensure a balance between unlearning and retaining utility. + +Remark 4. The primal problem (cf. Equation (4)) involves optimizing over the high-dimensional space of gradient vectors, while the dual problem (cf. Equation (5)) simplifies this problem to optimize over a single scalar variable $v \in \mathbb{R}$ . This dimensionality reduction can significantly decrease the computational complexity, especially for large-scale neural networks. + +# 2.3 Algorithm Description + +Algorithm 1 describes the procedure of Scissorhands in detail. We first identify the important parameters w.r.t. the forgetting data in single-shot via connection sensitivity in networks, then re-initialize these parameters for erasing the influence of the forgetting data $\mathcal{D}_f$ in the given model. To repair the performance of the trimmed model on the remaining data $\mathcal{D}_r$ , while ensuring the exclusion of information associated with the forgetting data $\mathcal{D}_f$ , we employ the gradient projection-based optimization algorithm. By doing so, we can scrub the influence of the forgetting data while retaining model utility on the remaining data. The loss depicted in Equation (3) can be viewed as a scalarization of a Multi-Objective Optimization (MOO) problem, namely minimizing $\left(\mathcal{L}(\theta; \mathcal{D}_r), -\mathcal{L}(\theta; \mathcal{D}_f)\right)^{\top}$ . A common issue in MOO is the gradient conflict, as optimizing one objective could hinder another one. Our proposal on gradient projection is designed to address and mitigate the gradient conflict. Empirically (see Table 3), we studied the effect of removing gradient projection. + +# 3 Related Work + +The development of efficient machine unlearning methods [2, 3, 11, 17, 18, 20, 23, 27, 28, 37, 40, 42, 45, 48, 51, 58, 61] has gained prominence. Applications of machine unlearning span various domains, including regression tasks [51], federated learning [24, 34, 35, 54, 56], graph neural networks [5, 7], and diffusion models [12, 15, 16, 26, 57, 60, 62], as well as scenarios where training data are not available [8, 50]. + +Retraining the model from scratch without forgetting data is typically considered the standard gold unlearning algorithm [12, 52]. However, this is often deemed impractical as most in-production models require extensive training periods and considerable computing resources. While fine-tuning the model for a new task might lead to forgetting previous knowledge (i.e., catastrophic forgetting) [36], it fails to adequately erase the data influence in the models. Approximate unlearning methods, as such, become attractive alternatives. We will briefly discuss these methods in image classification and image generation. + +Unlearning in Classification Models. Most unlearning algorithms are based on the influence function [23,38,48] and the Fisher Information Matrix (FIM) [19-21]. Influence function-based methods estimate the influence of the particular training data points on the model's predictions, and Fisher unlearning methods assume that the unlearned model and the retrained model are close to each other, Golatkar et al. [20] leverage FIM and hide the difference between the unlearned models and retrained models via adding noise. Golatkar et al. [20] also present an upper bound for the amount of retained information, offering a quantifiable measure of the effectiveness of the unlearning process. Such approximate unlearning methods need to compute the Hessian matrix, or FIM w.r.t. the data. This may render them impractical for certain scenarios, such as federated learning, where computational resources are distributed and limited. To mitigate these computational demands, Mehta et al. [37] introduce L-CODEC, a strategy aimed at + +pre-selecting a subset of parameters, allowing for more efficient computation. Jia et al. [27] recently propose to fuse the model sparsity into the unlearning algorithm, helping improve approximate unlearning methods, and achieving effective and efficient unlearning effects. + +Concept Erasure in Generative Models. The advent of generative models, particularly those converting text to images, has been a significant milestone in the field of artificial intelligence. A notable concern is the risk of it being tainted or manipulated [6,43], resulting in not-safe-for-work (NSFW) generations [46], as these models leverage training data from a wide array of open sources. To address these challenges, data censoring [1, 14, 41, 47] to exclude black-listed images is employed. Studies [15, 46] introduce methods to update the models away from the inappropriate concepts. Heng and Soh [26] recently propose to adopt Elastic Weight Consolidation (EWC) and Generative Replay (GR) to effectively unlearn without access to the training data for a wide range of generative models. + +Unlearning across Domains. Recently, Fan et al. [12] highlight the difficulty in cross-domain applicability of the machine unlearning algorithms, i.e., existing machine unlearning methods designed for classification tasks cannot be effective when applied to image generation. To address this gap, they introduce a novel method, Saliency unlearning (SalUn), which shifts attention from the entire model to target parameters. SalUn achieves state-of-the-art performance and is effective in both image classification and image generation tasks. + +In this work, we propose Scissorhands which could achieve improved performance compared to existing methods across various scenarios. Inspired by [27] that apply $\ell_1$ regularization to simply fuse sparsity into the unlearning methods, we identify and re-initialize the key model parameters w.r.t. the forgetting data via connection sensitivity. Concurrently, Fan et al. [12] also explore unlearning via salience scores, determining weight salience via the gradient of the forgetting loss w.r.t. model parameters, and employing random labeling for model finetuning. While there are similarities in our overarching goals, the methodologies present unique perspectives and solutions to the problem of unlearning. Fan et al. [12] only focus on the salient weights, while our mechanism aims to first 'destroy' the key information and then relearn for the whole model. Additionally, using the forgetting data with random labels could potentially lead to memorizing information about the forgetting data points [59]. In contrast, Scissorhands excludes knowledge about the forgetting data via a gradient projection-based approach, resulting in the effective erasure of information about the forgetting data while preserving the knowledge about the retained data. + +# 4 Experimental Evaluation + +In this section, we illustrate how Scissorhands effectively eliminates the data influence in the models. For sample-wise forgetting, where the forgetting data shares the same distribution as the training data, we evaluate Scissorhands on SVHN [39], CIFAR-10 and CIFAR-100 [30]. We further extend our evaluation on CelebAMask-HQ [31] where targeting to remove the entirety of specific identities. + +Additionally, our experimentation encompasses the open-source text-to-image model, Stable Diffusion v1.4 [44], which is conditioned on CLIP text embeddings through the cross-attention mechanism. Further detailed experimental setups and additional results can be found in Appendix B. + +Baselines. We primarily compare against the following standard baselines that are frequently employed in machine unlearning, alongside recently proposed SOTA methods: (i) Retrain: models obtained by only retraining from scratch on $\mathcal{D}_r$ . This will provide the performance of the oracle for us. (ii) Fine-tuning (FT) [55]: models that are fine-tuned on $\mathcal{D}_r$ , i.e., taking advantage from catastrophic forgetting in neural networks to unlearn. (iii) Gradient ascent (GA) [52]: gradient ascent on $\mathcal{D}_f$ . This will provide a simple way of unlearning by making the performance of the model worse on $\mathcal{D}_f$ . (iv) Influence unlearning (IU) [29]: utilizes the influence function method to estimate the change in model parameters when transitioning from the unscrubbed model to the retrained model. (v) Boundary shrink (BS) [4] and (vi) Boundary expanding (BE) [4]: shift the decision boundary of the original model to imitate that of the retrained model. (vii) $\ell_1$ -sparse [27]: fine-tuning models on $\mathcal{D}_r$ with $\ell_1$ -norm sparse regularization. (viii) Saliency unlearning (SalUn) [12]: adopt weight saliency and random labeling for unlearning. + +Metrics. To assess the unlearning algorithms, we employ several metrics: (i) Accuracy: the accuracy of the model on $\mathcal{D}_f$ (denoted as $\mathbf{Acc}_{\mathcal{D}_f}$ ), $\mathcal{D}_r$ (denoted as $\mathbf{Acc}_{\mathcal{D}_r}$ ) and the test set (denoted as $\mathbf{Acc}_{\mathcal{D}_t}$ ). It provides insight into how well the model performs after undergoing the unlearning process. (ii) Membership inference attack (MIA): a standard metric for verifying the unlearning effect. A classifier $\phi$ is trained using both training data (marked with a label of 1, indicating data points included in training) and test data (marked with a label of 0, indicating data points not seen during training), and subsequently assessed on $\mathcal{D}_f$ . An effective unlearning method should make it challenging to infer whether a particular sample was part of the training set, mitigating potential privacy breaches. In this context, we define MIA as $P_{\phi}(\pmb{y} = 0|\pmb{x}_f)$ , representing the probability assigned by the classifier $\phi$ that a given sample $\pmb{x}_f$ was not included in the training data. (iii) Avg. Gap [12]: the average of the performance gaps measured in accuracy-related metrics: Avg. Gap $= (|\mathrm{Acc}_{\mathcal{D}_t} - \mathrm{Acc}_{\mathcal{D}_t}^*| + |\mathrm{Acc}_{\mathcal{D}_f} - \mathrm{Acc}_{\mathcal{D}_f}^*| + |\mathrm{Acc}_{\mathcal{D}_r} - \mathrm{Acc}_{\mathcal{D}_r}^*| + |\mathrm{MIA} - \mathrm{MIA}^*|)/4$ , where $\mathrm{Acc}_{\mathcal{D}_t}^*$ , $\mathrm{Acc}_{\mathcal{D}_f}^*$ , $\mathrm{Acc}_{\mathcal{D}_r}^*$ and $\mathrm{MIA}^*$ are metric values of the retrained model. A better performance of an unlearning method corresponds to a lower performance gap with retraining, measuring how close the unlearned model is to the retrained model. (iv) Run-time efficiency (RTE) [27]: evaluate the computational efficiency of the unlearning method. Specifically, we have $\mathrm{RTE} = \mathrm{T} / \mathrm{T}_r$ , where $\mathrm{T}$ and $\mathrm{T}_r$ present the time for the unlearning methods to get the scrubbed models and the time for retraining from scratch with $\mathcal{D}_r$ , respectively. An efficient unlearning method should aim for minimal computational overhead compared to the baseline. (v) Relearn time [8,20,51]: the number of epochs to fine-tune the scrubbed models for regaining performance on $\mathcal{D}_f$ (i.e., reach the original models' accuracy on $\mathcal{D}_f$ ). + +Table 1: Quantitative results for forgetting ${10}\%$ data. + +
MethodAccDf(↓)AccDt(↑)AccDr(↑)MIA(↑)Avg. Gap
CIFAR-100Retrain75.13±0.8574.69±0.0899.98±0.0150.22±0.62-
FT [55]97.98±1.3675.28±0.1299.95±0.029.64±3.616.01
GA [52]98.00±1.3475.59±0.1198.24±1.165.00±2.2517.68
IU [29]95.67±4.8272.13±4.5896.14±4.519.43±5.9816.93
BE [4]97.94±1.3874.16±0.0998.12±1.247.60±3.0516.96
BS [4]97.65±1.4873.20±0.1897.93±1.308.24±3.2317.01
l1-sparse [27]96.35±0.6770.06±0.4696.35±0.6721.33±1.9514.59
SalUn [12]88.56±1.1871.34±0.4899.40±0.3574.66±2.4810.45
Ours68.76±1.8173.17±0.2499.24±0.3042.42±2.064.11
CIFAR-10Retrain94.81±0.5394.26±0.14100.0±0.0013.05±0.64-
FT [55]99.15±0.4693.83±0.4599.84±0.113.01±0.933.74
GA [52]99.66±0.2394.57±0.0199.62±0.250.91±0.294.42
IU [29]98.08±2.1091.91±2.7398.01±2.264.01±3.444.16
BE [4]99.41±0.3893.79±0.1599.41±0.3816.16±0.782.19
BS [4]99.60±0.2594.24±0.0799.56±0.544.46±0.333.46
l1-sparse [27]94.17±0.4990.64±0.5296.64±0.5411.87±0.612.20
SalUn [12]98.07±0.4293.92±0.2599.89±0.0717.93±0.372.15
Ours95.40±1.4892.92±0.4898.93±0.579.56±2.131.62
SVHNRetrain91.81±1.1191.17±1.7797.73±1.1215.74±1.28-
FT [55]99.60±0.2495.28±0.0499.99±0.003.85±0.46.51
GA [52]98.41±0.2392.87±0.0698.52±0.295.96±0.384.72
IU [29]92.47±1.6286.92±2.0693.36±1.8217.00±2.672.63
BE [4]98.48±0.3392.62±0.0598.44±0.296.77±0.364.45
BS [4]98.29±0.3292.48±0.0298.36±0.296.74±0.394.36
l1-sparse [27]97.79±0.1093.59±3.1899.49±0.237.57±0.294.65
SalUn [12]96.29±4.8093.59±3.1897.41±4.9624.47±4.083.99
Ours91.62±1.1893.05±0.5699.51±0.1417.62±2.541.43
+ +# 4.1 Results on Classification Task + +In this experiment on SVHN, CIFAR-10, and CIFAR-100, we try to forget randomly selected $10\%$ of the training data; on CelebAMask-HQ, we attempt to forget randomly selected $10\%$ identities among 307 identities. We further apply GradCAM [49] to visualize regions where models focus on $\mathbf{w}/$ and $\mathbf{w} / \mathbf{o}$ machine unlearning algorithms. In brief, the results suggest that Scissorhands has successfully induced forgetting for the relevant samples and classes, with minor degradation in model performance over the remaining data and classes. We will discuss the results in depth below. + +Sample-wise unlearning. Table 1 presents the results when forgetting randomly selected samples. Scissorhands achieves the lowest average performance gap on all the presented datasets and shows good generalization ability compared to + +Table 2: Quantitative results for forgetting ${10}\%$ identities on the CelebAMask-HQ. + +
MethodAccDf(↓)AccDt(↑)AccDr(↑)MIA(↑)Avg. Gap
Retrain0.00±0.0087.02±0.8099.96±0.01100.0±0.00-
FT [55]99.94±0.1288.59±0.5999.97±7.025.28±2.0349.06
GA [52]87.60±8.7181.22±2.1199.74±0.2651.37±5.9635.56
IU [29]88.92±10.2570.24±11.7795.27±5.0729.59±18.5945.20
BE [4]69.07±2.7344.11±2.0895.58±1.2346.24±5.9042.53
BS [4]98.18±1.9281.92±0.2799.86±0.0345.93±5.1139.36
l1-sparse [27]98.81±0.7289.37±0.7099.97±0.0076.78±5.6631.10
SalUn [12]0.00±0.0078.36±1.3496.90±1.11100.0±0.002.93
Ours1.52±2.7380.18±6.6097.20±3.8199.83±0.352.82
+ +![](images/a785f5a123cef880a8214a26247f8d55b6be1f71058b3f264d6dd46c0648a217.jpg) +Fig. 1: Visualizations of regions where models focus on generated by GradCAM [49]. + +![](images/44b0c00ac887072f2a3e599bc8d2653d331a7a141fb1fbc3f2eb13b8495116f3.jpg) + +Retrain. For example, on SVHN, Scissorhands achieves $\sim 91.62\%$ accuracy on the forgetting data $\mathcal{D}_f$ , closely matching the $91.81\%$ accuracy of a fully retrained model. Furthermore, the retrained model reaches around $91\%$ accuracy on the test dataset, Scissorhands exhibits an accuracy of roughly $93\%$ , showcasing the enhanced generalization of our scrubbed model. + +When evaluating solely based on forgetting accuracy on CIFAR-10, the $\ell_1$ -sparse method might seem to be the most effective baseline, achieving approximately $94.17\%$ . Nonetheless, this perceived advantage is offset by a decline in both test accuracy and the accuracy of retained data. In contrast, Scissorhands shows a good trade-off between preserving model utility on the remaining data, generalization on unseen sets, and erasing the influence of the forgetting data. + +Similarly, on CIFAR-100, SalUn may initially stand out with its impressive MIA accuracy of approximately $74.66\%$ , relying solely on a single metric to gauge the performance of machine unlearning can be misleading, as it might not fully capture the method's effectiveness [12]. While SalUn achieves a forgetting accuracy of about $88.56\%$ and a test accuracy of $71.34\%$ , Scissorhands demonstrates a slightly lower forgetting accuracy of around $68.76\%$ but achieves a higher test accuracy of $73.17\%$ . This indicates that Scissorhands achieves the best trade-off between eradicating data traces and generalizing to unseen data. Soon, we will show that on this dataset, the relearning time of Scissorhands is greater than 200 to regain the performance on $\mathcal{D}_f$ , which indicates that Scissorhands has efficiently scrubbed information about $\mathcal{D}_f$ . + +Overall, Scissorhands illustrates its capability to adeptly balance the objectives of data removal, model utility preservation, and generalization to unseen data, ensuring that the influence of forgotten data is minimized without significantly compromising the overall performance and utility of the model. + +Class-wise unlearning. Furthermore, Table 2 presents the outcomes of our analysis for the scenario where forgetting $10\%$ of identities on the CelebAMask-HQ. Images are rescaled to $224 \times 224$ , and a model pre-trained with ImageNet1K [9] is employed. In this context, Scissorhands manifests the smallest average performance gap with retrained models. Notably, the Retrain method achieves perfect metrics in terms of MIA accuracy, which is $100\%$ , and completely erases the identity information from the dataset, as indicated by a $0.00\%$ accuracy on the forgotten data $\mathcal{D}_f$ . However, retraining proves impractical and limited for unlearning purposes in general, functioning merely as an oracle in our experiments. + +Among the baselines, fine-tuning (FT) and $\ell_1$ -sparse exhibit higher accuracies on the remaining data $\mathcal{D}_f$ and unseen data $\mathcal{D}_t$ ; yet, their accuracies on forgetting remain high as well, highlighting a less effective balance between eradicating data traces and generalizing to unseen data. In contrast, Scissorhands and SalUn demonstrate superior capabilities in effectively forgetting identities with minimal impact on the models' overall performance. SalUn achieves an MIA accuracy of $100\%$ and a forgetting accuracy of $0.00\%$ , effectively nullifying the identity information similar to the retrained models. Scissorhands slightly surpasses SalUn in terms of the average performance gap with retrained models, maintaining a high MIA accuracy of $99.83\%$ and a significantly low accuracy on the forgotten data $\mathcal{D}_f$ (i.e., $1.52\%$ ), while achieving a test accuracy of $80.18\%$ and an accuracy of $97.20\%$ on the retained data. + +We further employ the GradCAM [49] to illustrate the focus area of models $\mathbf{w}/$ and $\mathbf{w} / \mathbf{o}$ machine unlearning algorithms. As shown in Figure 1, when evaluated on the remaining data $\mathcal{D}_r$ , both SalUn and Scissorhands even pay more attention to the facial features for identity classification than the retrained model. Conversely, when evaluated on the forgetting data $\mathcal{D}_f$ , the scrubbed model by Scissorhands shifts its attention towards regions that are least associated with the facial features. We hypothesize that in the scenario where the task involves forgetting identities, both the remaining data and the forgetting data are face images, the scrubbed models would need a more nuanced approach to distinguish between individuals. Specifically, because the task at hand requires the discernment of subtle differences across facial features to effectively differentiate and forget specific identities while retaining others, the models adapt by developing a refined sensitivity towards those facial characteristics that are most indicative of individual identity. This adaptation enables the models to maintain high accuracy on the remaining data by focusing more intently on the facial features that are crucial for identity recognition, thereby enhancing their ability to generalize and accurately classify identities. + +These results underscore our proposed machine unlearning algorithm Scisserhands's superior capabilities in balancing identity forgetting and model utility + +Table 3: Influence of projection and initialization strategies (re-initializing parameters with uniform distribution $\mathcal{U}$ , Gaussian distribution $\mathcal{N}$ , and constant value of $\mathbf{0} / \mathbf{1}$ ). + +
ProjectionInitializationAccDf(↓)AccDt(↑)AccDr(↑)MIA(↑)Avg. Gap
✓U95.40±1.4892.92±0.4898.93±0.579.56±2.131.62
XU98.14±0.6093.14±0.1699.73±0.166.08±0.782.92
✓N94.68±2.5092.13±2.2097.94±2.7110.18±3.811.80
✓096.10±2.0592.92±0.9999.09±1.198.76±3.301.96
X184.28±0.4983.38±0.5887.18±0.9020.67±0.4110.46
+ +![](images/49ce2eb998962f055f1c1505c0eab18abf65a9e97e0ffcfdbbd67a2a5ba70f36.jpg) +Fig. 2: Influence of the percent value $k$ in the trimming process, the balance term $\lambda$ of Equation (3), and the ratio of forgetting data used in the trimming process. + +![](images/42debbda1a0e6c1c2adbc107457577a88ce1b522f4d8e5506df22c64479722eb.jpg) + +![](images/3b54eb9b20dc279b9ad11a69362034f579e60d0e6a084e51e7c9813020958b2b.jpg) + +performance, Scissorhands not only minimizes privacy risks but also maintains the integrity and applicability of the model to unseen data. + +# 4.2 Ablation Study + +Component analysis. As detailed in Table 3, the inclusion of different procedures in Scissorhands has a noticeable impact. Notably, the employment of the projection mechanism plays a crucial role in eliminating the influence of forgetting data $\mathcal{D}_f$ , affecting the overall utility of the model. The rationale behind the projection process is to preclude the reacquisition of critical information about the forgetting data $\mathcal{D}_f$ during the repair phase. Moreover, Table 3 reveals that initializing parameters with a uniform distribution $\mathcal{U}$ or opting for zero initialization facilitates effective unlearning, in contrast to re-initializing parameters with a constant value of one, which impedes finding a viable solution to Equation (5). + +Hyper-parameter impact. Figure 2 presents the effects of varying hyperparameters, such as the percentage value $k\%$ for the trimming process, the balance term $\lambda$ of Equation (3) in the repairing phase, and the ratio of forgetting data used in the trimming process. An increase $k$ results in higher accuracies on data, as fewer parameters associated with the forgetting data $\mathcal{D}_f$ are reinitialized. Empirically, we observe choosing a value of $k \geq 0.9$ will result in stable performances. That said, this is a hyperparameter of the algorithm and can be chosen by measuring the performances on $D_{f}$ and $D_{r}$ during training. The hyper-parameter $\lambda$ controls the balance between retaining and forgetting data, the accuracies on test data and remaining data show a peak at $\lambda = 0.05$ + +![](images/3fcc8d75600a3a75b9f3649de707efdfc9679a9c06b8e6795ed4f33fbc87b2b3.jpg) +Fig.3: Quantity of nudity content detected using the NudeNet classifier from 1K sampled images and I2P data. We observed a high false positive rate for exposed female genitalia/breast using the NudeNet classifier on generated I2P images. The flagged images can be found in Appendix B. + +![](images/f1cae026a9d6fe32206f7d1e78adee6957744618fdf13d1bdca542d8c70f0ed5.jpg) + +SD v1.4 + +![](images/713a1ee117321f659b6b8469b69a2fca897a11456e862636577b74899a7e1ef3.jpg) +Fig. 4: Sample images with the prompt from $c_f = \{ \text{nudity}, \text{naked}, \text'erotic', \text'sexual' \}$ generated by SDs w/ and w/o machine unlearning algorithms. Best viewed in color. + +SalUn + +![](images/2689f618d6fa3b839545dc332ea13e39ec40792a50a53b9884239f241373cb04.jpg) + +Ours + +![](images/50e51cbd01bfd27c35570d50cbcd24fc19546610e8d2930a676a0ce37f2124bc.jpg) + +$\star$ Added by authors for publication + +![](images/f0dde109d4f8e9617c60fb4672f44a226e66ce417cac5949c9bb12b0d8e60f08.jpg) + +when forgetting $10\%$ data on CIFAR-10. Generally, as $\lambda$ escalates, it places greater emphasis on the process of data omission, leading to a decrease in forgetting accuracy. As the ratio $p$ increases, all accuracy metrics exhibit a decline, suggesting that employing more forgetting data for the purpose of forgetting in order to isolate critical knowledge about the forgetting data also inadvertently results in the loss of pertinent information regarding the remaining data. This phenomenon occurs particularly when the data designated for forgetting shares the same distribution as the data that is retained. In essence, while aiming to enhance the specificity of the forgetting process by pinpointing essential details about the data to be forgotten, there's a risk of simultaneously diminishing knowledge pertinent to the retained data. + +# 4.3 Case Study: Stable Diffusion + +We also employed our algorithm to mitigate the generation of inappropriate content in Stable Diffusion (SD) models [44]. Specifically, our primary objective is to effectively eliminate the concept of nudity from the model's generative capabilities. To this end, followed [26] and employed SD v1.4 for sampling with 50 time steps. We evaluate on 1K generated images with prompts $c_f = \{ \text{nudity}, \text{naked}, \text'erotic', \text'sexual' \}$ and 4703 generated images with I2P [46] using the open-source NudeNet classifier. + +Figures 3 and 4 present the results for SD v2.1 (trained on a dataset filtered for nudity), SalUn, and Scissorhands. In Figure 3, y-axis denotes the number of + +exposed body parts generated by the SD v1.4 model, it presents the percentage change in exposed body parts w.r.t. SD v1.4. When comparing on I2P data, our algorithms outperform SalUn on 'Feet', while coming short on 'Female Breast and Genitalia'. While this could suggest the superiority of SalUn on this data, the gap is not that big in reality, as out of the 9 images flagged for containing exposed female breasts, 7 were inaccurately identified, and none of the flagged images depicted exposed female genitalia. The use of the NudeNet classifier on generated I2P images exhibits a significant rate of false positives, as highlighted in [26] as well. When we use prompts $c_f$ explicitly associated with nudity, no exposed sensitive content is detected for both Scissorhands and SalUn methods. We can conclude that, Scissorhands significantly reduces the amount of nudity content compared to SD v1.4, SD v2.1, showing comparable results to SalUn. + +# 5 Limitations and Broader Impact + +Our proposed unlearning algorithm, Scissorhands, can achieve comparable performance to state-of-the-art unlearning methods. We addressed potential issues of conflicting gradients arising from direct conflicts between features or patterns in the forgetting and remaining data, ensuring training stability. However, the limitation is the incomplete concept erasure shown in Figure 8 in Appendix B, and the potential for biased models due to the selective deletion of information. This is a crucial consideration, especially when unlearning algorithms are applied with malicious intent. Furthermore, in experiments on stable diffusion, while our method can successfully erase specific concepts, finding an optimal balance between forgetting and maintaining high image quality remains challenging. There is also a risk that unlearning algorithms might be used to alter or remove concepts in unethical or malicious ways. It is our hope that Scissorhands offers a fresh perspective on practical and effective machine unlearning, emphasizing the need for ethical guidelines to prevent misuse in its application. + +# 6 Conclusion + +In this work, we proposed Scissorhands, an effective and practical machine unlearning algorithm. Our algorithm identified the important parameters w.r.t. the forgetting data in single-shot via connection sensitivity, then re-initialized these parameters for scrubbing the influence of the forgetting data. An effective gradient projection-based technique is applied to enhance the model utility while excluding the information w.r.t. the forgetting data. Our evaluations on various datasets showed that, compared with other methods, our approach offers superior performance across image classification and image generation tasks. Future works could explore how Scissorhands performs on regression and NLP tasks, and further investigation into the scenario where data are not available, as well as unlearning with sequential/time-series data where could introduce unique challenges, e.g., the need to consider temporal dependencies, which may require new frameworks specifically tailored for sequential data. + +# Acknowledgements + +Mehrtash Harandi is supported by funding from the Australian Research Council (ARC) Discovery Program DP230101176. + +# References + +1. Birhane, A., Prabhu, V.U.: Large image datasets: A pyrrhic win for computer vision? In: 2021 IEEE Winter Conference on Applications of Computer Vision (WACV). pp. 1536-1546. IEEE (2021) +2. Bourtoule, L., Chandrasekaran, V., Choquette-Choo, C.A., Jia, H., Travers, A., Zhang, B., Lie, D., Papernot, N.: Machine unlearning. In: 2021 IEEE Symposium on Security and Privacy (SP). pp. 141-159 (2021) +3. Cao, Y., Yang, J.: Towards making systems forget with machine unlearning. In: 2015 IEEE Symposium on Security and Privacy (SP). pp. 463-480 (2015) +4. Chen, M., Gao, W., Liu, G., Peng, K., Wang, C.: Boundary unlearning: Rapid forgetting of deep networks via shifting the decision boundary. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7766-7775 (2023) +5. Chen, M., Zhang, Z., Wang, T., Backes, M., Humbert, M., Zhang, Y.: Graph unlearning. In: Proceedings of the 2022 ACM SIGSAC Conference on Computer and Communications Security. pp. 499-513 (2022) +6. Chen, W., Song, D., Li, B.: Trojdiff: Trojan attacks on diffusion models with diverse targets. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4035-4044 (2023) +7. Cheng, J., Dasoulas, G., He, H., Agarwal, C., Zitnik, M.: Gnndelete: A general strategy for unlearning in graph neural networks. In: International Conference on Learning Representations (ICLR) (2023) +8. Chandawat, V.S., Tarun, A.K., Mandal, M., Kankanhalli, M.: Zero-shot machine unlearning. IEEE Transactions on Information Forensics and Security (2023) +9. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009) +0. Dennis, C.R., Weisberg, S.: Residuals and influence in regression (1982) +1. Fan, C., Liu, J., Hero, A., Liu, S.: Challenging forgets: Unveiling the worst-case forget sets in machine unlearning. arXiv preprint arXiv:2403.07362 (2024) +2. Fan, C., Liu, J., Zhang, Y., Wei, D., Wong, E., Liu, S.: Salun: Empowering machine unlearning via gradient-based weight saliency in both image classification and generation. In: International Conference on Learning Representations (ICLR) (2024) +3. Foster, J., Schoepf, S., Brintrup, A.: Fast machine unlearning without retraining through selective synaptic dampening. arXiv preprint arXiv:2308.07707 (2023) +4. Gandhi, S., Kokkula, S., Chaudhuri, A., Magnani, A., Stanley, T., Ahmadi, B., Kandaswamy, V., Ovenc, O., Mannor, S.: Scalable detection of offensive and noncompliant content/logo in product images. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2247-2256 (2020) +5. Gandikota, R., Materzynska, J., Fiotto-Kaufman, J., Bau, D.: Erasing concepts from diffusion models. In: 2023 IEEE International Conference on Computer Vision (ICCV) (2023) + +16. Gandikota, R., Orgad, H., Belinkov, Y., Materzyńska, J., Bau, D.: Unified concept editing in diffusion models. arXiv preprint arXiv:2308.14761 (2023) +17. Ginart, A., Guan, M., Valiant, G., Zou, J.Y.: Making ai forget you: Data deletion in machine learning. In: Advances in Neural Information Processing Systems (NeurIPS). vol. 32 (2019) +18. Goel, S., Prabhu, A., Sanyal, A., Lim, S.N., Torr, P., Kumaraguru, P.: Towards adversarial evaluations for inexact machine unlearning. arXiv preprint arXiv:2201.06640 (2022) +19. Golatkar, A., Achille, A., Ravichandran, A., Polito, M., Soatto, S.: Mixed-privacy forgetting in deep networks. In: 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 792-801 (2021) +20. Golatkar, A., Achille, A., Soatto, S.: Eternal sunshine of the spotless net: Selective forgetting in deep networks. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9301-9309 (2020) +21. Golatkar, A., Achille, A., Soatto, S.: Forgetting outside the box: Scrubbing deep networks of information accessible from input-output observations. In: European Conference on Computer Vision (ECCV). pp. 383-398. Springer (2020) +22. Goldman, E.: An introduction to the california consumer privacy act (ccpa). Santa Clara Univ. Legal Studies Research Paper (2020) +23. Guo, C., Goldstein, T., Hannun, A., Van Der Maaten, L.: Certified data removal from machine learning models. In: Proceedings of the 37th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 119, pp. 3832-3842. PMLR (2020) +24. Halimi, A., Kadhe, S., Rawat, A., Baracaldo, N.: Federated unlearning: How to efficiently erase a client in fl? In: International conference on machine learning. PMLR (2022) +25. Han, S., Pool, J., Tran, J., Dally, W.: Learning both weights and connections for efficient neural network. vol. 28 (2015) +26. Heng, A., Soh, H.: Selective amnesia: A continual learning approach to forgetting in deep generative models. In: Advances in Neural Information Processing Systems (NeurIPS) (2023) +27. Jia, J., Liu, J., Ram, P., Yao, Y., Liu, G., Liu, Y., Sharma, P., Liu, S.: Model sparsification can simplify machine unlearning. In: Advances in Neural Information Processing Systems (NeurIPS) (2023) +28. Karasuyama, M., Takeuchi, I.: Multiple incremental decremental learning of support vector machines. IEEE Transactions on Neural Networks 21(7), 1048-1059 (2010) +29. Koh, P.W., Liang, P.: Understanding black-box predictions via influence functions. In: International conference on machine learning. pp. 1885-1894. PMLR (2017) +30. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009) +31. Lee, C.H., Liu, Z., Wu, L., Luo, P.: Maskgan: Towards diverse and interactive facial image manipulation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5549-5558 (2020) +32. Lee, N., Ajanthan, T., Gould, S., Torr, P.H.: A signal propagation perspective for pruning neural networks at initialization. In: International Conference on Learning Representations (ICLR) (2020) +33. Lee, N., Ajanthan, T., Torr, P.H.: Snip: Single-shot network pruning based on connection sensitivity. In: International Conference on Learning Representations (ICLR) (2019) + +34. Liu, G., Ma, X., Yang, Y., Wang, C., Liu, J.: Federaser: Enabling efficient client-level data removal from federated learning models. In: 2021 IEEE/ACM 29th International Symposium on Quality of Service (IWQOS). pp. 1-10 (2021) +35. Liu, Y., Xu, L., Yuan, X., Wang, C., Li, B.: The right to be forgotten in federated learning: An efficient realization with rapid retraining. In: IEEE INFOCOM 2022 - IEEE Conference on Computer Communications. pp. 1749-1758 (2022) +36. Lopez-Paz, D., Ranzato, M.: Gradient episodic memory for continual learning. Advances in neural information processing systems 30 (2017) +37. Mehta, R., Pal, S., Singh, V., Ravi, S.N.: Deep unlearning via randomized conditionally independent hessians. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10412-10421 (2022) +38. Neel, S., Roth, A., Sharifi-Malvajerdi, S.: Descent-to-delete: Gradient-based methods for machine unlearning. In: Proceedings of the 32nd International Conference on Algorithmic Learning Theory. pp. 931-962. Proceedings of Machine Learning Research, PMLR (2021) +39. Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B., Ng, A.Y.: Reading digits in natural images with unsupervised feature learning (2011) +40. Nguyen, T.T., Huynh, T.T., Nguyen, P.L., Liew, A.W.C., Yin, H., Nguyen, Q.V.H.: A survey of machine unlearning. arXiv preprint arXiv:2209.02299 (2022) +41. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741 (2021) +42. Peste, A., Alistarh, D., Lampert, C.H.: Ssse: Efficiently erasing samples from trained machine learning models. arXiv preprint arXiv:2107.03860 (2021) +43. Rando, J., Paleka, D., Lindner, D., Heim, L., Tramér, F.: Red-teaming the stable diffusion safety filter. arXiv preprint arXiv:2210.04610 (2022) +44. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022) +45. Romero, E., Barrio, I., Belanche, L.: Incremental and decremental learning for linear support vector machines. In: International Conference on Artificial Neural Networks. pp. 209-218. Springer (2007) +46. Schramowski, P., Brack, M., Deiseroth, B., Kersting, K.: Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22522-22531 (2023) +47. Schramowski, P., Tauchmann, C., Kersting, K.: Can machines help us answering question 16 in datasheets, and in turn reflecting on inappropriate content? In: Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency. pp. 1350-1361 (2022) +48. Sekhari, A., Acharya, J., Kamath, G., Suresh, A.T.: Remember what you want to forget: Algorithms for machine unlearning. Advances in Neural Information Processing Systems (NeurIPS) 34, 18075-18086 (2021) +49. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Gradcam: Visual explanations from deep networks via gradient-based localization. In: 2017 IEEE International Conference on Computer Vision (ICCV). pp. 618-626 (2017) +50. Tarun, A.K., Chundawat, V.S., Mandal, M., Kankanhalli, M.: Fast yet effective machine unlearning. IEEE Transactions on Neural Networks and Learning Systems (2023) + +51. Tarun, A.K., Chundawat, V.S., Mandal, M., Kankanhalli, M.: Deep regression unlearning. In: International Conference on Machine Learning. pp. 33921-33939. PMLR (2023) +52. Thudi, A., Deza, G., Chandrasekaran, V., Papernot, N.: Unrolling sgd: Understanding factors influencing machine unlearning. In: 2022 IEEE 7th European Symposium on Security and Privacy (EuroS&P). pp. 303-319. IEEE (2022) +53. Voigt, P., Von dem Bussche, A.: The eu general data protection regulation (gdpr). A Practical Guide, 1st Ed., Cham: Springer International Publishing 10(3152676), 10-5555 (2017) +54. Wang, J., Guo, S., Xie, X., Qi, H.: Federated unlearning via class-discriminative pruning. In: Proceedings of the ACM Web Conference 2022. pp. 622-632 (2022) +55. Warnecke, A., Pirch, L., Wressnegger, C., Rieck, K.: Machine unlearning of features and labels. arXiv preprint arXiv:2108.11577 (2021) +56. Wu, C., Zhu, S., Mitra, P.: Federated unlearning with knowledge distillation. arXiv preprint arXiv:2201.09441 (2022) +57. Wu, J., Le, T., Hayat, M., Harandi, M.: Erasediff: Erasing data influence in diffusion models. arXiv preprint arXiv:2401.05779 (2024) +58. Wu, Y., Dobriban, E., Davidson, S.: DeltaGrad: Rapid retraining of machine learning models. In: Proceedings of the 37th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 119, pp. 10355-10366. PMLR (13-18 Jul 2020) +59. Zhang, C., Bengio, S., Hardt, M., Recht, B., Vinyals, O.: Understanding deep learning (still) requires rethinking generalization. In: International Conference on Learning Representations (ICLR) (2017) +60. Zhang, E., Wang, K., Xu, X., Wang, Z., Shi, H.: Forget-me-not: Learning to forget in text-to-image diffusion models. arXiv preprint arXiv:2303.17591 (2023) +61. Zhang, Y., Zhang, Y., Yao, Y., Jia, J., Liu, J., Liu, X., Liu, S.: Unlearncanvas: A stylized image dataset to benchmark machine unlearning for diffusion models. arXiv preprint arXiv:2402.11846 (2024) +62. Zhang, Y., Chen, X., Jia, J., Zhang, Y., Fan, C., Liu, J., Hong, M., Ding, K., Liu, S.: Defensive unlearning with adversarial training for robust concept erasure in diffusion models. arXiv preprint arXiv:2405.15234 (2024) \ No newline at end of file diff --git a/2024/Scissorhands_ Scrub Data Influence via Connection Sensitivity in Networks/images.zip b/2024/Scissorhands_ Scrub Data Influence via Connection Sensitivity in Networks/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..fb87bc866643b8a6a42562229f73d7a7be909dcb --- /dev/null +++ b/2024/Scissorhands_ Scrub Data Influence via Connection Sensitivity in Networks/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c5c6689478271052c9025ce90eb40cfbcf9580ec1d3bb8014ab567f322b1d6c5 +size 442510 diff --git a/2024/Scissorhands_ Scrub Data Influence via Connection Sensitivity in Networks/layout.json b/2024/Scissorhands_ Scrub Data Influence via Connection Sensitivity in Networks/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d9272f083f7e0485ba82a2626c0db42449dcebdc --- /dev/null +++ b/2024/Scissorhands_ Scrub Data Influence via Connection Sensitivity in Networks/layout.json @@ -0,0 +1,11010 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 171, + 111, + 444, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 111, + 444, + 148 + ], + "spans": [ + { + "bbox": [ + 171, + 111, + 444, + 148 + ], + "type": "text", + "content": "Scissorhands: Scrub Data Influence via Connection Sensitivity in Networks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 232, + 167, + 381, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 232, + 167, + 381, + 180 + ], + "spans": [ + { + "bbox": [ + 232, + 167, + 381, + 180 + ], + "type": "text", + "content": "Jing Wu1, Mehrtash Harandi1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 179, + 190, + 435, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 190, + 435, + 223 + ], + "spans": [ + { + "bbox": [ + 179, + 190, + 435, + 223 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 179, + 190, + 435, + 223 + ], + "type": "text", + "content": "Department of Electrical and Computer Systems Engineering Monash University, Australia {jing.wu1, mehrtash.harandi}@monash.edu" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 247, + 455, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 247, + 455, + 423 + ], + "spans": [ + { + "bbox": [ + 160, + 247, + 455, + 423 + ], + "type": "text", + "content": "Abstract. Machine unlearning has become a pivotal task to erase the influence of data from a trained model. It adheres to recent data regulation standards and enhances the privacy and security of machine learning applications. In this work, we present a new machine unlearning approach Scissorhands. Initially, Scissorhands identifies the most pertinent parameters in the given model relative to the forgetting data via connection sensitivity. By reinitializing the most influential top-" + }, + { + "bbox": [ + 160, + 247, + 455, + 423 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 160, + 247, + 455, + 423 + ], + "type": "text", + "content": " percent of these parameters, a trimmed model for erasing the influence of the forgetting data is obtained. Subsequently, Scissorhands fine-tunes the trimmed model with a gradient projection-based approach, seeking parameters that preserve information on the remaining data while discarding information related to the forgetting data. Our experimental results, conducted across image classification and image generation tasks, demonstrate that Scissorhands, showcases competitive performance when compared to existing methods. Source code is available at https://github.com/JingWu321/Scissorhands." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 433, + 460, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 433, + 460, + 445 + ], + "spans": [ + { + "bbox": [ + 160, + 433, + 460, + 445 + ], + "type": "text", + "content": "Keywords: Machine unlearning " + }, + { + "bbox": [ + 160, + 433, + 460, + 445 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 433, + 460, + 445 + ], + "type": "text", + "content": " Connection sensitivity " + }, + { + "bbox": [ + 160, + 433, + 460, + 445 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 433, + 460, + 445 + ], + "type": "text", + "content": " Diffusion model" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 459, + 453, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 459, + 453, + 480 + ], + "spans": [ + { + "bbox": [ + 159, + 459, + 453, + 480 + ], + "type": "text", + "content": "Warning: This paper contains explicit sexual imagery that may be offensive." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 499, + 230, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 499, + 230, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 499, + 230, + 511 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 521, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 593 + ], + "type": "text", + "content": "In this work, we aim to propose an effective machine unlearning method. Under data regulations like the European Union General Data Protection Regulation (GDPR) [53] and the California Consumer Privacy Act (CCPA) [22], all users are granted the right to be forgotten. In machine learning, these legal provisions empower data owners with the right not only to withdraw their data from trained models but also to ensure that their data's influence on these models is erased." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "The most direct approach to accomplish this objective is to retrain the model from scratch, excluding the data requested for deletion from the training process. Retraining from scratch is typically considered the gold standard in the field of machine unlearning [12, 52]. Yet, this poses a challenge as numerous in-production models require prolonged training periods and substantial computing resources. While retraining is feasible, it is often impractical." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "Consequently, the proposal and development of efficient approximate unlearning methods [13, 23, 26, 50, 51] have become essential. Currently, most approximate unlearning techniques achieve forgetting by adding normally distributed noise to the parameters [19-21] or by estimating the influence of a particular data point on the model's predictions [23, 37, 38, 48] based on the influence function method [10]. Jia et al. [27] demonstrate that model sparsity can help to unlearn and fuse model sparsity into the unlearning process." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 482, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 482, + 368 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 482, + 368 + ], + "type": "text", + "content": "Fan et al. [12] recently raised concerns regarding the instability in approximate unlearning methods, and highlighted that the current machine unlearning approaches, initially crafted for image classification tasks, fall short in effectively tackling the challenges of machine unlearning within the realm of image generation. The authors introduce a novel unlearning method SalUn that can effectively perform forgetting in both image classification and generation tasks. Their unlearning mechanism includes finding salient weights and then fine-tuning these weights using forgetting data assigned with random labels. However, studies like [59] show that the model can still learn the true data distribution even using data with random labels. While SalUn has shown state-of-the-art performance, it could potentially memorize information about forgetting data points, albeit associating them with random labels but still can be undesired in sensitive applications. In § 4, results also show that SalUn tends to memorize knowledge about the forgetting data." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 368, + 482, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 368, + 482, + 523 + ], + "spans": [ + { + "bbox": [ + 130, + 368, + 482, + 523 + ], + "type": "text", + "content": "This work. We present an unlearning approach, Scissorhands, designed to erase data influence in the classification models and eliminate a particular concept from a text-to-image model. To achieve this, inspired by [27], our key insight is to first erase the critical influence of forgetting data in models and then relearn. Unlike previous methods that employ the forgetting data with noisy labels as a part of relearning, we propose to use an efficient gradient projection method to relearn the critical features and patterns while ensuring the exclusion of influences associated with the forgetting data. Through a series of experiments and evaluations, including the classification on SVHN [39], CIFAR-10 and CIFAR-100 [30], CelebAMASK-HQ [31] datasets, as well as the open-source stable diffusion [44] text-to-image model, results demonstrate the viability and effectiveness of our technique in forgetting the influence of random samples, discrete classes, and sensitive content such as nudity." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 542, + 233, + 556 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 542, + 233, + 556 + ], + "spans": [ + { + "bbox": [ + 132, + 542, + 233, + 556 + ], + "type": "text", + "content": "2 Methodology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 568, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 568, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 568, + 482, + 629 + ], + "type": "text", + "content": "In this section, we propose Scissorhands, our unlearning framework that scrubs data from a model by eliminating their influence on model parameters. Throughout the paper, we denote scalars and vectors/matrices by lowercase and bold symbols, respectively (e.g., " + }, + { + "bbox": [ + 130, + 568, + 482, + 629 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 130, + 568, + 482, + 629 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 568, + 482, + 629 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 130, + 568, + 482, + 629 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 568, + 482, + 629 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 130, + 568, + 482, + 629 + ], + "type": "text", + "content": "). We show a " + }, + { + "bbox": [ + 130, + 568, + 482, + 629 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 568, + 482, + 629 + ], + "type": "text", + "content": "-dimensional vector of ones by " + }, + { + "bbox": [ + 130, + 568, + 482, + 629 + ], + "type": "inline_equation", + "content": "\\mathbf{1}_d" + }, + { + "bbox": [ + 130, + 568, + 482, + 629 + ], + "type": "text", + "content": ", and use " + }, + { + "bbox": [ + 130, + 568, + 482, + 629 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 130, + 568, + 482, + 629 + ], + "type": "text", + "content": " to denote the Hadamard product." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 641, + 481, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 481, + 667 + ], + "type": "text", + "content": "Overview. Consider a model " + }, + { + "bbox": [ + 132, + 641, + 481, + 667 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 132, + 641, + 481, + 667 + ], + "type": "text", + "content": " with parameters " + }, + { + "bbox": [ + 132, + 641, + 481, + 667 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta} \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 132, + 641, + 481, + 667 + ], + "type": "text", + "content": " trained on a dataset " + }, + { + "bbox": [ + 132, + 641, + 481, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{\\mathbf{x}_i, \\mathbf{y}_i\\}_{i=1}^N" + }, + { + "bbox": [ + 132, + 641, + 481, + 667 + ], + "type": "text", + "content": ". Suppose a subset of " + }, + { + "bbox": [ + 132, + 641, + 481, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 132, + 641, + 481, + 667 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 132, + 641, + 481, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 132, + 641, + 481, + 667 + ], + "type": "text", + "content": ", is requested for deletion." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 282, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 282, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 282, + 102 + ], + "type": "text", + "content": "Jing Wu, Mehrtash Harandi" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "The remaining data is defined as " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r\\coloneqq \\mathcal{D}\\setminus \\mathcal{D}_f" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": ". Our objective is to develop the unlearning algorithm to generate a scrubbed model " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "f_{u}" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": ", effectively removing the influence of " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " from model " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": ", while maintaining its utility over " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 480, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 480, + 222 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 480, + 222 + ], + "type": "text", + "content": "In Scissorhands, we initially identify critical parameters w.r.t. the forgetting data via connection sensitivity analysis [32, 33]. We then reinitialize the top-" + }, + { + "bbox": [ + 130, + 152, + 480, + 222 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 152, + 480, + 222 + ], + "type": "text", + "content": " percent of these key parameters, resulting in a trimmed model with parameters " + }, + { + "bbox": [ + 130, + 152, + 480, + 222 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}_t" + }, + { + "bbox": [ + 130, + 152, + 480, + 222 + ], + "type": "text", + "content": ". This step aims to diminish the model's memory of the forgetting data " + }, + { + "bbox": [ + 130, + 152, + 480, + 222 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 130, + 152, + 480, + 222 + ], + "type": "text", + "content": ". Nevertheless, this process runs the risk of erasing valuable information from the data set " + }, + { + "bbox": [ + 130, + 152, + 480, + 222 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r" + }, + { + "bbox": [ + 130, + 152, + 480, + 222 + ], + "type": "text", + "content": " that we aim to retain." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 224, + 480, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 480, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 480, + 283 + ], + "type": "text", + "content": "To address this, the next phase of Scissorhands concentrates on restoring the performance of the trimmed model on remaining data " + }, + { + "bbox": [ + 130, + 224, + 480, + 283 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r" + }, + { + "bbox": [ + 130, + 224, + 480, + 283 + ], + "type": "text", + "content": ", while ensuring that the influence of the forgetting data " + }, + { + "bbox": [ + 130, + 224, + 480, + 283 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 130, + 224, + 480, + 283 + ], + "type": "text", + "content": " remains excluded. We accomplish this balance and obtain the final scrubbed model with parameters " + }, + { + "bbox": [ + 130, + 224, + 480, + 283 + ], + "type": "inline_equation", + "content": "\\theta^{\\star}" + }, + { + "bbox": [ + 130, + 224, + 480, + 283 + ], + "type": "text", + "content": " through a gradient projection-based approach." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 300, + 211, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 211, + 312 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 211, + 312 + ], + "type": "text", + "content": "2.1 Trimming" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 318, + 479, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 318, + 479, + 401 + ], + "spans": [ + { + "bbox": [ + 130, + 318, + 479, + 401 + ], + "type": "text", + "content": "To effectively identify salient parameters of a network with respect to " + }, + { + "bbox": [ + 130, + 318, + 479, + 401 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 130, + 318, + 479, + 401 + ], + "type": "text", + "content": ", one has to establish a good criterion to determine such salient connections. A commonly used criterion is the magnitude of the weights, with those exceeding a certain threshold deemed salient as suggested in [25]. Our requirement is more nuanced, as we need a measure that specifically discerns saliency based on " + }, + { + "bbox": [ + 130, + 318, + 479, + 401 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 130, + 318, + 479, + 401 + ], + "type": "text", + "content": ". Therefore, we adopt a single-shot approach as proposed in [33], which we outline briefly to ensure clarity and completeness in our methodology. Consider" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 194, + 408, + 480, + 428 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 408, + 480, + 428 + ], + "spans": [ + { + "bbox": [ + 194, + 408, + 480, + 428 + ], + "type": "interline_equation", + "content": "\\mathbf {s} _ {j} (\\mathcal {D}) := \\mathbb {E} _ {\\boldsymbol {x}, \\boldsymbol {y} \\sim \\mathcal {D}} \\left[ \\ell (\\boldsymbol {\\theta}; \\boldsymbol {x}, \\boldsymbol {y}) - \\ell \\left(\\left(\\mathbf {1} _ {d} - \\boldsymbol {e} _ {j}\\right) \\odot \\boldsymbol {\\theta}; \\boldsymbol {x}, \\boldsymbol {y}\\right) \\right], \\tag {1}", + "image_path": "aa876a5e221fa01a36ecc1622880f735c77a356a1495d8ba0643486460a79fbf.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 434, + 482, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 434, + 482, + 506 + ], + "spans": [ + { + "bbox": [ + 131, + 434, + 482, + 506 + ], + "type": "text", + "content": "which measures the influence of parameter " + }, + { + "bbox": [ + 131, + 434, + 482, + 506 + ], + "type": "inline_equation", + "content": "j \\in \\{1, \\dots, d\\}" + }, + { + "bbox": [ + 131, + 434, + 482, + 506 + ], + "type": "text", + "content": " on a model in terms of the empirical risk for a given dataset " + }, + { + "bbox": [ + 131, + 434, + 482, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 131, + 434, + 482, + 506 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 131, + 434, + 482, + 506 + ], + "type": "inline_equation", + "content": "\\ell : \\mathbb{R}^d \\times \\mathcal{X} \\times \\mathcal{Y} \\to \\mathbb{R}_+" + }, + { + "bbox": [ + 131, + 434, + 482, + 506 + ], + "type": "text", + "content": " is the loss of the model and " + }, + { + "bbox": [ + 131, + 434, + 482, + 506 + ], + "type": "inline_equation", + "content": "e_j" + }, + { + "bbox": [ + 131, + 434, + 482, + 506 + ], + "type": "text", + "content": " is the indicator vector of element " + }, + { + "bbox": [ + 131, + 434, + 482, + 506 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 131, + 434, + 482, + 506 + ], + "type": "text", + "content": " (i.e., a binary vector with its " + }, + { + "bbox": [ + 131, + 434, + 482, + 506 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 131, + 434, + 482, + 506 + ], + "type": "text", + "content": "-th component set to one). Note that computing " + }, + { + "bbox": [ + 131, + 434, + 482, + 506 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_j" + }, + { + "bbox": [ + 131, + 434, + 482, + 506 + ], + "type": "text", + "content": " for each " + }, + { + "bbox": [ + 131, + 434, + 482, + 506 + ], + "type": "inline_equation", + "content": "j \\in \\{1, \\dots, d\\}" + }, + { + "bbox": [ + 131, + 434, + 482, + 506 + ], + "type": "text", + "content": " is prohibitively expensive, as it demands " + }, + { + "bbox": [ + 131, + 434, + 482, + 506 + ], + "type": "inline_equation", + "content": "d + 1" + }, + { + "bbox": [ + 131, + 434, + 482, + 506 + ], + "type": "text", + "content": " forward passes through the dataset. An approximation to Equation (1) [25] is in the form of" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 235, + 514, + 480, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 514, + 480, + 539 + ], + "spans": [ + { + "bbox": [ + 235, + 514, + 480, + 539 + ], + "type": "interline_equation", + "content": "\\mathbf {s} _ {j} (\\mathcal {D}) \\approx \\mathbb {E} _ {\\boldsymbol {x}, \\boldsymbol {y} \\sim \\mathcal {D}} \\left[ \\frac {\\partial \\ell (\\boldsymbol {\\theta} ; \\boldsymbol {x} , \\boldsymbol {y})}{\\partial \\boldsymbol {\\theta} _ {j}} \\boldsymbol {\\theta} _ {j} \\right]. \\tag {2}", + "image_path": "de83ba3ccf902411ad186c330ac6977fc688e69e646a61c76312fb0d1dd82902.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 545, + 479, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 479, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 479, + 640 + ], + "type": "text", + "content": "Detailed proofs can be found in Appendix A. Equation (2) defines the sensitivity of parameter " + }, + { + "bbox": [ + 130, + 545, + 479, + 640 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 545, + 479, + 640 + ], + "type": "text", + "content": " as the average (i.e., expectation) of a product, the gradient of the loss and the current value of the parameter. As such, " + }, + { + "bbox": [ + 130, + 545, + 479, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_j" + }, + { + "bbox": [ + 130, + 545, + 479, + 640 + ], + "type": "text", + "content": " encodes the loss sensitivity according to the parameter's magnitude. Intuitively, a parameter with a large value that significantly affects the loss, as captured by the gradient, is considered more influential or salient. The form in Equation (2) offers several benefits, most importantly, the ability to calculate the sensitivity for all parameters with just a single sweep through the dataset." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 641, + 479, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 479, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 479, + 666 + ], + "type": "text", + "content": "To scrub the influence of the forgetting data " + }, + { + "bbox": [ + 131, + 641, + 479, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 131, + 641, + 479, + 666 + ], + "type": "text", + "content": " in the model, we first obtain " + }, + { + "bbox": [ + 131, + 641, + 479, + 666 + ], + "type": "inline_equation", + "content": "\\mathfrak{s}_j(\\mathcal{D}_f)" + }, + { + "bbox": [ + 131, + 641, + 479, + 666 + ], + "type": "text", + "content": ", saliency of parameters w.r.t. the forgetting data. We then re-initialize" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 143, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 143, + 91, + 448, + 102 + ], + "type": "text", + "content": "Scissorhands: Scrub Data Influence via Connection Sensitivity in Networks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "the top-" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "k\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": " of the parameters based on their saliency rankings. This process is akin to performing a targeted 'lobotomy' on the model specifically concerning the data " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": ", thereby selectively erasing their influence. Unfortunately, this aggressive approach of reinitializing parameters can detrimentally impact the model's performance on the remaining dataset " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": ", which we will correct in the next phase of Scissorhands." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 197, + 482, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 197, + 482, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 197, + 482, + 258 + ], + "type": "text", + "content": "Remark 1. In principle, Scissorhands can benefit from any algorithms that can identify important connections w.r.t. " + }, + { + "bbox": [ + 130, + 197, + 482, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 130, + 197, + 482, + 258 + ], + "type": "text", + "content": ". The main reason behind choosing SNIP is its single-shot nature. As presented in the study [12], the gradient of loss w.r.t. the model parameters over the forgetting data can help identify important parameters w.r.t. " + }, + { + "bbox": [ + 130, + 197, + 482, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 130, + 197, + 482, + 258 + ], + "type": "text", + "content": " and hence help with erasing." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 265, + 482, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 265, + 482, + 315 + ], + "spans": [ + { + "bbox": [ + 130, + 265, + 482, + 315 + ], + "type": "text", + "content": "Remark 2. In practice, we have observed that even a small subset of " + }, + { + "bbox": [ + 130, + 265, + 482, + 315 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 130, + 265, + 482, + 315 + ], + "type": "text", + "content": " can be sufficient for the trimming. For example, to unlearn a large set of data on CIFAR-100, our algorithm outperforms baselines by utilizing approximately " + }, + { + "bbox": [ + 130, + 265, + 482, + 315 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 130, + 265, + 482, + 315 + ], + "type": "text", + "content": " of the forgetting data for trimming." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 322, + 482, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 322, + 482, + 371 + ], + "spans": [ + { + "bbox": [ + 130, + 322, + 482, + 371 + ], + "type": "text", + "content": "Remark 3. Initializing weights of the trimmed neurons can take various forms. Empirically, we observed initializing with a uniform distribution is particularly effective. Details on the influence of different initialization strategies and the choice of " + }, + { + "bbox": [ + 130, + 322, + 482, + 371 + ], + "type": "inline_equation", + "content": "\\mathrm{k}\\%" + }, + { + "bbox": [ + 130, + 322, + 482, + 371 + ], + "type": "text", + "content": " trimming on the model's performance will be discussed in " + }, + { + "bbox": [ + 130, + 322, + 482, + 371 + ], + "type": "inline_equation", + "content": "\\S 4.2" + }, + { + "bbox": [ + 130, + 322, + 482, + 371 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 387, + 211, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 387, + 211, + 399 + ], + "spans": [ + { + "bbox": [ + 131, + 387, + 211, + 399 + ], + "type": "text", + "content": "2.2 Repairing" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 407, + 482, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 482, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 482, + 540 + ], + "type": "text", + "content": "Following the parameter-trimming process, aimed at mitigating the influence of the forgetting data " + }, + { + "bbox": [ + 130, + 407, + 482, + 540 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 130, + 407, + 482, + 540 + ], + "type": "text", + "content": ", a challenge presents itself: the potential erasure of crucial information associated with the remaining data " + }, + { + "bbox": [ + 130, + 407, + 482, + 540 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r" + }, + { + "bbox": [ + 130, + 407, + 482, + 540 + ], + "type": "text", + "content": ". A straightforward solution to recover the model utility is to relearn the model using the remaining data. However, this approach risks biasing the model towards features of the remaining data and inadvertently reintegrating information about the forgetting samples. In the quest to ensure that models retain essential information from the remaining data " + }, + { + "bbox": [ + 130, + 407, + 482, + 540 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r" + }, + { + "bbox": [ + 130, + 407, + 482, + 540 + ], + "type": "text", + "content": " while effectively forgetting " + }, + { + "bbox": [ + 130, + 407, + 482, + 540 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 130, + 407, + 482, + 540 + ], + "type": "text", + "content": ", our strategy involves maximizing the loss over the forgetting data " + }, + { + "bbox": [ + 130, + 407, + 482, + 540 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 130, + 407, + 482, + 540 + ], + "type": "text", + "content": " while concurrently minimizing the loss over the remaining data " + }, + { + "bbox": [ + 130, + 407, + 482, + 540 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r" + }, + { + "bbox": [ + 130, + 407, + 482, + 540 + ], + "type": "text", + "content": ". Therefore, to achieve the balance between these goals, we propose the objective of an efficient practice for unlearning:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 223, + 548, + 480, + 562 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 548, + 480, + 562 + ], + "spans": [ + { + "bbox": [ + 223, + 548, + 480, + 562 + ], + "type": "interline_equation", + "content": "L (\\boldsymbol {\\theta}, \\mathcal {D} _ {r}, \\mathcal {D} _ {f}) := \\mathcal {L} (\\boldsymbol {\\theta}; \\mathcal {D} _ {r}) - \\lambda \\mathcal {L} (\\boldsymbol {\\theta}; \\mathcal {D} _ {f}). \\tag {3}", + "image_path": "74edd5bca4c396b55a02ca152ed713f24691ff578def30bcde1a373bdc90830c.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": "We aim to optimize Equation (3) w.r.t. " + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": " through gradient descent, ensuring minimal reintroduction of information about " + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": ". This is particularly challenging when there are similarities between samples in " + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": ". To achieve this desideratum, it is vital that updates to the model do not improve (i.e., reduce) " + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\pmb{\\theta};\\mathcal{D}_f)" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": ". To this end, we define " + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\pmb{g}_f = \\nabla_{\\pmb{\\theta}}\\mathcal{L}(\\pmb{\\theta};\\mathcal{D}_f)" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": " as the gradient direction for " + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\pmb{g}_o = \\nabla_{\\pmb{\\theta}}L(\\pmb{\\theta},\\mathcal{D}_r,\\mathcal{D}_f)" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": " as the gradient direction for optimizing Equation (3). The optimal descent direction " + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\pmb{g} \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": " should thus satisfy two criteria: 1. It should exhibit maximum similarity to " + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\pmb{g}_o" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": " to ensure swift convergence. The" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 282, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 282, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 282, + 102 + ], + "type": "text", + "content": "Jing Wu, Mehrtash Harandi" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 132, + 129, + 482, + 299 + ], + "blocks": [ + { + "bbox": [ + 133, + 116, + 337, + 128 + ], + "lines": [ + { + "bbox": [ + 133, + 116, + 337, + 128 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 337, + 128 + ], + "type": "text", + "content": "Algorithm 1 The procedure of Scissorhands." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 132, + 129, + 482, + 299 + ], + "lines": [ + { + "bbox": [ + 132, + 129, + 482, + 299 + ], + "spans": [ + { + "bbox": [ + 132, + 129, + 482, + 299 + ], + "type": "text", + "content": "1: procedure TRIMMING " + }, + { + "bbox": [ + 132, + 129, + 482, + 299 + ], + "type": "inline_equation", + "content": "(\\pmb {\\theta},\\mathcal{D}_f)" + }, + { + "bbox": [ + 132, + 129, + 482, + 299 + ], + "type": "text", + "content": " 11: for " + }, + { + "bbox": [ + 132, + 129, + 482, + 299 + ], + "type": "inline_equation", + "content": "e = 0" + }, + { + "bbox": [ + 132, + 129, + 482, + 299 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 132, + 129, + 482, + 299 + ], + "type": "inline_equation", + "content": "E - 1" + }, + { + "bbox": [ + 132, + 129, + 482, + 299 + ], + "type": "text", + "content": " do \n2: " + }, + { + "bbox": [ + 132, + 129, + 482, + 299 + ], + "type": "inline_equation", + "content": "\\pmb {X},\\pmb {y}\\gets \\{\\pmb {x}_i,\\pmb {y}_i\\sim \\mathcal{D}_f|i\\in [B]\\} ." + }, + { + "bbox": [ + 132, + 129, + 482, + 299 + ], + "type": "text", + "content": " 12: Get " + }, + { + "bbox": [ + 132, + 129, + 482, + 299 + ], + "type": "inline_equation", + "content": "L(\\pmb {\\theta},\\mathcal{D}_r,\\mathcal{D}_f)" + }, + { + "bbox": [ + 132, + 129, + 482, + 299 + ], + "type": "text", + "content": " (cf. Eq. (3)). \n3: // Compute connection sensitivity 13: Get " + }, + { + "bbox": [ + 132, + 129, + 482, + 299 + ], + "type": "inline_equation", + "content": "\\pmb{g_o}" + }, + { + "bbox": [ + 132, + 129, + 482, + 299 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 129, + 482, + 299 + ], + "type": "inline_equation", + "content": "\\pmb{g_f}" + }, + { + "bbox": [ + 132, + 129, + 482, + 299 + ], + "type": "text", + "content": " \n4: Get " + }, + { + "bbox": [ + 132, + 129, + 482, + 299 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_j(\\pmb {\\theta};\\pmb {X},\\pmb {y}),\\forall j\\in [1,d]" + }, + { + "bbox": [ + 132, + 129, + 482, + 299 + ], + "type": "text", + "content": " 14: //Get the optimal direction using Eq.(2). 15: if " + }, + { + "bbox": [ + 132, + 129, + 482, + 299 + ], + "type": "inline_equation", + "content": "\\langle g_o,g_f\\rangle >0" + }, + { + "bbox": [ + 132, + 129, + 482, + 299 + ], + "type": "text", + "content": " then \n5: // Re-initialization 16: Compute " + }, + { + "bbox": [ + 132, + 129, + 482, + 299 + ], + "type": "inline_equation", + "content": "v^{*}" + }, + { + "bbox": [ + 132, + 129, + 482, + 299 + ], + "type": "text", + "content": " (cf. Eq. (5)). \n6: Re-initialize parameters according 17: " + }, + { + "bbox": [ + 132, + 129, + 482, + 299 + ], + "type": "inline_equation", + "content": "\\pmb {g} = \\pmb {g}_o - v^*\\pmb {g}_f" + }, + { + "bbox": [ + 132, + 129, + 482, + 299 + ], + "type": "text", + "content": " to the top- " + }, + { + "bbox": [ + 132, + 129, + 482, + 299 + ], + "type": "inline_equation", + "content": "k \\%" + }, + { + "bbox": [ + 132, + 129, + 482, + 299 + ], + "type": "text", + "content": " value of " + }, + { + "bbox": [ + 132, + 129, + 482, + 299 + ], + "type": "inline_equation", + "content": "\\mathfrak{s}_j" + }, + { + "bbox": [ + 132, + 129, + 482, + 299 + ], + "type": "text", + "content": " 18: " + }, + { + "bbox": [ + 132, + 129, + 482, + 299 + ], + "type": "inline_equation", + "content": "\\pmb {g}_o\\gets \\pmb {g}" + }, + { + "bbox": [ + 132, + 129, + 482, + 299 + ], + "type": "text", + "content": " \n7: return " + }, + { + "bbox": [ + 132, + 129, + 482, + 299 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}_{t}" + }, + { + "bbox": [ + 132, + 129, + 482, + 299 + ], + "type": "text", + "content": " 19: end if \n8: end procedure 20: " + }, + { + "bbox": [ + 132, + 129, + 482, + 299 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}^{e + 1}\\leftarrow \\pmb{\\theta}^{e} - \\eta \\pmb{g}_{o}" + }, + { + "bbox": [ + 132, + 129, + 482, + 299 + ], + "type": "text", + "content": " \n21: end for \n9: procedure REPAIRING " + }, + { + "bbox": [ + 132, + 129, + 482, + 299 + ], + "type": "inline_equation", + "content": "(\\pmb {\\theta}_t,\\mathcal{D}_f,\\mathcal{D}_r)" + }, + { + "bbox": [ + 132, + 129, + 482, + 299 + ], + "type": "text", + "content": " 22: return " + }, + { + "bbox": [ + 132, + 129, + 482, + 299 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}^{*} = \\pmb{\\theta}^{E}" + }, + { + "bbox": [ + 132, + 129, + 482, + 299 + ], + "type": "text", + "content": " \n10: " + }, + { + "bbox": [ + 132, + 129, + 482, + 299 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}^{0} = \\pmb{\\theta}_{t}" + }, + { + "bbox": [ + 132, + 129, + 482, + 299 + ], + "type": "text", + "content": " 23: end procedure" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "algorithm" + }, + { + "bbox": [ + 130, + 323, + 482, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 323, + 482, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 482, + 384 + ], + "type": "text", + "content": "notion of similarity can be captured by " + }, + { + "bbox": [ + 130, + 323, + 482, + 384 + ], + "type": "inline_equation", + "content": "\\pmb{g}_o" + }, + { + "bbox": [ + 130, + 323, + 482, + 384 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 130, + 323, + 482, + 384 + ], + "type": "inline_equation", + "content": "\\| \\pmb{g} - \\pmb{g}_o \\|_2^2" + }, + { + "bbox": [ + 130, + 323, + 482, + 384 + ], + "type": "text", + "content": ". 2. It should not align with " + }, + { + "bbox": [ + 130, + 323, + 482, + 384 + ], + "type": "inline_equation", + "content": "\\pmb{g}_f" + }, + { + "bbox": [ + 130, + 323, + 482, + 384 + ], + "type": "text", + "content": " to prevent improving " + }, + { + "bbox": [ + 130, + 323, + 482, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\pmb{\\theta}; \\mathcal{D}_f)" + }, + { + "bbox": [ + 130, + 323, + 482, + 384 + ], + "type": "text", + "content": ". This can be mathematically formalized as " + }, + { + "bbox": [ + 130, + 323, + 482, + 384 + ], + "type": "inline_equation", + "content": "\\langle \\pmb{g}_f, \\pmb{g} \\rangle \\leq 0" + }, + { + "bbox": [ + 130, + 323, + 482, + 384 + ], + "type": "text", + "content": ". Consequently, after computing " + }, + { + "bbox": [ + 130, + 323, + 482, + 384 + ], + "type": "inline_equation", + "content": "\\pmb{g}_o = \\nabla_{\\pmb{\\theta}} L(\\pmb{\\theta}, \\mathcal{D}_r, \\mathcal{D}_f)" + }, + { + "bbox": [ + 130, + 323, + 482, + 384 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 323, + 482, + 384 + ], + "type": "inline_equation", + "content": "\\pmb{g}_f = \\nabla_{\\pmb{\\theta}} \\mathcal{L}(\\pmb{\\theta}; \\mathcal{D}_f)" + }, + { + "bbox": [ + 130, + 323, + 482, + 384 + ], + "type": "text", + "content": " in each iteration, we formulate the following optimization challenge to identify the model update direction:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 255, + 392, + 481, + 434 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 392, + 481, + 434 + ], + "spans": [ + { + "bbox": [ + 255, + 392, + 481, + 434 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\underset {\\boldsymbol {g}} {\\arg \\min} \\frac {1}{2} \\| \\boldsymbol {g} - \\boldsymbol {g} _ {o} \\| _ {2} ^ {2}, \\\\ \\text {s . t .} \\quad \\langle \\boldsymbol {g} _ {f}, \\boldsymbol {g} \\rangle \\leq 0. \\tag {4} \\\\ \\end{array}", + "image_path": "8756ef2c56c4b569fccb5919c68b4584219701653954aa3df5da6ddeff81e09d.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 442, + 482, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 442, + 482, + 492 + ], + "spans": [ + { + "bbox": [ + 130, + 442, + 482, + 492 + ], + "type": "text", + "content": "This problem may be addressed using the Frank-Wolfe algorithm. Given the high dimensionality of gradients in neural networks, directly addressing the constraint optimization in Equation (4) could become overwhelming. Inspired by [36], we propose to make use of the dual formulation of the problem Equation (4) as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 250, + 510, + 481, + 550 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 510, + 481, + 550 + ], + "spans": [ + { + "bbox": [ + 250, + 510, + 481, + 550 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\sup _ {v} \\frac {1}{2} \\boldsymbol {g} _ {f} ^ {\\top} \\boldsymbol {g} _ {f} v ^ {2} - \\boldsymbol {g} _ {o} ^ {\\top} \\boldsymbol {g} _ {f} v, \\\\ \\begin{array}{l l} \\text {s . t .} & v \\geq 0. \\end{array} \\tag {5} \\\\ \\end{array}", + "image_path": "e7c1de67dd3928794a866962deb5274a88e1b1cf7945e190d322e2498c606615.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 559, + 482, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 559, + 482, + 596 + ], + "spans": [ + { + "bbox": [ + 130, + 559, + 482, + 596 + ], + "type": "text", + "content": "Then, the optimal descent direction " + }, + { + "bbox": [ + 130, + 559, + 482, + 596 + ], + "type": "inline_equation", + "content": "\\pmb{g}" + }, + { + "bbox": [ + 130, + 559, + 482, + 596 + ], + "type": "text", + "content": " is given from the solution " + }, + { + "bbox": [ + 130, + 559, + 482, + 596 + ], + "type": "inline_equation", + "content": "v^{*}" + }, + { + "bbox": [ + 130, + 559, + 482, + 596 + ], + "type": "text", + "content": " in Equation (5) as " + }, + { + "bbox": [ + 130, + 559, + 482, + 596 + ], + "type": "inline_equation", + "content": "\\pmb {g} = \\pmb{g}_{o} - v^{*}\\pmb{g}_{f}" + }, + { + "bbox": [ + 130, + 559, + 482, + 596 + ], + "type": "text", + "content": ". Detailed proofs can be found in Appendix A. By doing so, we ensure a balance between unlearning and retaining utility." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Remark 4. The primal problem (cf. Equation (4)) involves optimizing over the high-dimensional space of gradient vectors, while the dual problem (cf. Equation (5)) simplifies this problem to optimize over a single scalar variable " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "v \\in \\mathbb{R}" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": ". This dimensionality reduction can significantly decrease the computational complexity, especially for large-scale neural networks." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 143, + 90, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 90, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 143, + 90, + 448, + 102 + ], + "type": "text", + "content": "Scissorhands: Scrub Data Influence via Connection Sensitivity in Networks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 274, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 274, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 274, + 129 + ], + "type": "text", + "content": "2.3 Algorithm Description" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 133, + 482, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 133, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 133, + 482, + 304 + ], + "type": "text", + "content": "Algorithm 1 describes the procedure of Scissorhands in detail. We first identify the important parameters w.r.t. the forgetting data in single-shot via connection sensitivity in networks, then re-initialize these parameters for erasing the influence of the forgetting data " + }, + { + "bbox": [ + 130, + 133, + 482, + 304 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 130, + 133, + 482, + 304 + ], + "type": "text", + "content": " in the given model. To repair the performance of the trimmed model on the remaining data " + }, + { + "bbox": [ + 130, + 133, + 482, + 304 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r" + }, + { + "bbox": [ + 130, + 133, + 482, + 304 + ], + "type": "text", + "content": ", while ensuring the exclusion of information associated with the forgetting data " + }, + { + "bbox": [ + 130, + 133, + 482, + 304 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 130, + 133, + 482, + 304 + ], + "type": "text", + "content": ", we employ the gradient projection-based optimization algorithm. By doing so, we can scrub the influence of the forgetting data while retaining model utility on the remaining data. The loss depicted in Equation (3) can be viewed as a scalarization of a Multi-Objective Optimization (MOO) problem, namely minimizing " + }, + { + "bbox": [ + 130, + 133, + 482, + 304 + ], + "type": "inline_equation", + "content": "\\left(\\mathcal{L}(\\theta; \\mathcal{D}_r), -\\mathcal{L}(\\theta; \\mathcal{D}_f)\\right)^{\\top}" + }, + { + "bbox": [ + 130, + 133, + 482, + 304 + ], + "type": "text", + "content": ". A common issue in MOO is the gradient conflict, as optimizing one objective could hinder another one. Our proposal on gradient projection is designed to address and mitigate the gradient conflict. Empirically (see Table 3), we studied the effect of removing gradient projection." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 319, + 238, + 332 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 319, + 238, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 238, + 332 + ], + "type": "text", + "content": "3 Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 342, + 481, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 481, + 413 + ], + "type": "text", + "content": "The development of efficient machine unlearning methods [2, 3, 11, 17, 18, 20, 23, 27, 28, 37, 40, 42, 45, 48, 51, 58, 61] has gained prominence. Applications of machine unlearning span various domains, including regression tasks [51], federated learning [24, 34, 35, 54, 56], graph neural networks [5, 7], and diffusion models [12, 15, 16, 26, 57, 60, 62], as well as scenarios where training data are not available [8, 50]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 415, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 482, + 510 + ], + "type": "text", + "content": "Retraining the model from scratch without forgetting data is typically considered the standard gold unlearning algorithm [12, 52]. However, this is often deemed impractical as most in-production models require extensive training periods and considerable computing resources. While fine-tuning the model for a new task might lead to forgetting previous knowledge (i.e., catastrophic forgetting) [36], it fails to adequately erase the data influence in the models. Approximate unlearning methods, as such, become attractive alternatives. We will briefly discuss these methods in image classification and image generation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "Unlearning in Classification Models. Most unlearning algorithms are based on the influence function [23,38,48] and the Fisher Information Matrix (FIM) [19-21]. Influence function-based methods estimate the influence of the particular training data points on the model's predictions, and Fisher unlearning methods assume that the unlearned model and the retrained model are close to each other, Golatkar et al. [20] leverage FIM and hide the difference between the unlearned models and retrained models via adding noise. Golatkar et al. [20] also present an upper bound for the amount of retained information, offering a quantifiable measure of the effectiveness of the unlearning process. Such approximate unlearning methods need to compute the Hessian matrix, or FIM w.r.t. the data. This may render them impractical for certain scenarios, such as federated learning, where computational resources are distributed and limited. To mitigate these computational demands, Mehta et al. [37] introduce L-CODEC, a strategy aimed at" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 281, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 281, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 281, + 102 + ], + "type": "text", + "content": "Jing Wu, Mehrtash Harandi" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 163 + ], + "type": "text", + "content": "pre-selecting a subset of parameters, allowing for more efficient computation. Jia et al. [27] recently propose to fuse the model sparsity into the unlearning algorithm, helping improve approximate unlearning methods, and achieving effective and efficient unlearning effects." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 481, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 481, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 481, + 283 + ], + "type": "text", + "content": "Concept Erasure in Generative Models. The advent of generative models, particularly those converting text to images, has been a significant milestone in the field of artificial intelligence. A notable concern is the risk of it being tainted or manipulated [6,43], resulting in not-safe-for-work (NSFW) generations [46], as these models leverage training data from a wide array of open sources. To address these challenges, data censoring [1, 14, 41, 47] to exclude black-listed images is employed. Studies [15, 46] introduce methods to update the models away from the inappropriate concepts. Heng and Soh [26] recently propose to adopt Elastic Weight Consolidation (EWC) and Generative Replay (GR) to effectively unlearn without access to the training data for a wide range of generative models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 285, + 481, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 481, + 367 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 481, + 367 + ], + "type": "text", + "content": "Unlearning across Domains. Recently, Fan et al. [12] highlight the difficulty in cross-domain applicability of the machine unlearning algorithms, i.e., existing machine unlearning methods designed for classification tasks cannot be effective when applied to image generation. To address this gap, they introduce a novel method, Saliency unlearning (SalUn), which shifts attention from the entire model to target parameters. SalUn achieves state-of-the-art performance and is effective in both image classification and image generation tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 369, + 482, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 482, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 482, + 559 + ], + "type": "text", + "content": "In this work, we propose Scissorhands which could achieve improved performance compared to existing methods across various scenarios. Inspired by [27] that apply " + }, + { + "bbox": [ + 130, + 369, + 482, + 559 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 130, + 369, + 482, + 559 + ], + "type": "text", + "content": " regularization to simply fuse sparsity into the unlearning methods, we identify and re-initialize the key model parameters w.r.t. the forgetting data via connection sensitivity. Concurrently, Fan et al. [12] also explore unlearning via salience scores, determining weight salience via the gradient of the forgetting loss w.r.t. model parameters, and employing random labeling for model finetuning. While there are similarities in our overarching goals, the methodologies present unique perspectives and solutions to the problem of unlearning. Fan et al. [12] only focus on the salient weights, while our mechanism aims to first 'destroy' the key information and then relearn for the whole model. Additionally, using the forgetting data with random labels could potentially lead to memorizing information about the forgetting data points [59]. In contrast, Scissorhands excludes knowledge about the forgetting data via a gradient projection-based approach, resulting in the effective erasure of information about the forgetting data while preserving the knowledge about the retained data." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 579, + 302, + 593 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 579, + 302, + 593 + ], + "spans": [ + { + "bbox": [ + 132, + 579, + 302, + 593 + ], + "type": "text", + "content": "4 Experimental Evaluation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": "In this section, we illustrate how Scissorhands effectively eliminates the data influence in the models. For sample-wise forgetting, where the forgetting data shares the same distribution as the training data, we evaluate Scissorhands on SVHN [39], CIFAR-10 and CIFAR-100 [30]. We further extend our evaluation on CelebAMask-HQ [31] where targeting to remove the entirety of specific identities." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 143, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 143, + 91, + 448, + 102 + ], + "type": "text", + "content": "Scissorhands: Scrub Data Influence via Connection Sensitivity in Networks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": "Additionally, our experimentation encompasses the open-source text-to-image model, Stable Diffusion v1.4 [44], which is conditioned on CLIP text embeddings through the cross-attention mechanism. Further detailed experimental setups and additional results can be found in Appendix B." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 168, + 487, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 168, + 487, + 350 + ], + "spans": [ + { + "bbox": [ + 130, + 168, + 487, + 350 + ], + "type": "text", + "content": "Baselines. We primarily compare against the following standard baselines that are frequently employed in machine unlearning, alongside recently proposed SOTA methods: (i) Retrain: models obtained by only retraining from scratch on " + }, + { + "bbox": [ + 130, + 168, + 487, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r" + }, + { + "bbox": [ + 130, + 168, + 487, + 350 + ], + "type": "text", + "content": ". This will provide the performance of the oracle for us. (ii) Fine-tuning (FT) [55]: models that are fine-tuned on " + }, + { + "bbox": [ + 130, + 168, + 487, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r" + }, + { + "bbox": [ + 130, + 168, + 487, + 350 + ], + "type": "text", + "content": ", i.e., taking advantage from catastrophic forgetting in neural networks to unlearn. (iii) Gradient ascent (GA) [52]: gradient ascent on " + }, + { + "bbox": [ + 130, + 168, + 487, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 130, + 168, + 487, + 350 + ], + "type": "text", + "content": ". This will provide a simple way of unlearning by making the performance of the model worse on " + }, + { + "bbox": [ + 130, + 168, + 487, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 130, + 168, + 487, + 350 + ], + "type": "text", + "content": ". (iv) Influence unlearning (IU) [29]: utilizes the influence function method to estimate the change in model parameters when transitioning from the unscrubbed model to the retrained model. (v) Boundary shrink (BS) [4] and (vi) Boundary expanding (BE) [4]: shift the decision boundary of the original model to imitate that of the retrained model. (vii) " + }, + { + "bbox": [ + 130, + 168, + 487, + 350 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 130, + 168, + 487, + 350 + ], + "type": "text", + "content": "-sparse [27]: fine-tuning models on " + }, + { + "bbox": [ + 130, + 168, + 487, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r" + }, + { + "bbox": [ + 130, + 168, + 487, + 350 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 168, + 487, + 350 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 130, + 168, + 487, + 350 + ], + "type": "text", + "content": "-norm sparse regularization. (viii) Saliency unlearning (SalUn) [12]: adopt weight saliency and random labeling for unlearning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 354, + 483, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 483, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 483, + 668 + ], + "type": "text", + "content": "Metrics. To assess the unlearning algorithms, we employ several metrics: (i) Accuracy: the accuracy of the model on " + }, + { + "bbox": [ + 130, + 354, + 483, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 130, + 354, + 483, + 668 + ], + "type": "text", + "content": " (denoted as " + }, + { + "bbox": [ + 130, + 354, + 483, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{Acc}_{\\mathcal{D}_f}" + }, + { + "bbox": [ + 130, + 354, + 483, + 668 + ], + "type": "text", + "content": "), " + }, + { + "bbox": [ + 130, + 354, + 483, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r" + }, + { + "bbox": [ + 130, + 354, + 483, + 668 + ], + "type": "text", + "content": " (denoted as " + }, + { + "bbox": [ + 130, + 354, + 483, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{Acc}_{\\mathcal{D}_r}" + }, + { + "bbox": [ + 130, + 354, + 483, + 668 + ], + "type": "text", + "content": ") and the test set (denoted as " + }, + { + "bbox": [ + 130, + 354, + 483, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{Acc}_{\\mathcal{D}_t}" + }, + { + "bbox": [ + 130, + 354, + 483, + 668 + ], + "type": "text", + "content": "). It provides insight into how well the model performs after undergoing the unlearning process. (ii) Membership inference attack (MIA): a standard metric for verifying the unlearning effect. A classifier " + }, + { + "bbox": [ + 130, + 354, + 483, + 668 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 354, + 483, + 668 + ], + "type": "text", + "content": " is trained using both training data (marked with a label of 1, indicating data points included in training) and test data (marked with a label of 0, indicating data points not seen during training), and subsequently assessed on " + }, + { + "bbox": [ + 130, + 354, + 483, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 130, + 354, + 483, + 668 + ], + "type": "text", + "content": ". An effective unlearning method should make it challenging to infer whether a particular sample was part of the training set, mitigating potential privacy breaches. In this context, we define MIA as " + }, + { + "bbox": [ + 130, + 354, + 483, + 668 + ], + "type": "inline_equation", + "content": "P_{\\phi}(\\pmb{y} = 0|\\pmb{x}_f)" + }, + { + "bbox": [ + 130, + 354, + 483, + 668 + ], + "type": "text", + "content": ", representing the probability assigned by the classifier " + }, + { + "bbox": [ + 130, + 354, + 483, + 668 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 354, + 483, + 668 + ], + "type": "text", + "content": " that a given sample " + }, + { + "bbox": [ + 130, + 354, + 483, + 668 + ], + "type": "inline_equation", + "content": "\\pmb{x}_f" + }, + { + "bbox": [ + 130, + 354, + 483, + 668 + ], + "type": "text", + "content": " was not included in the training data. (iii) Avg. Gap [12]: the average of the performance gaps measured in accuracy-related metrics: Avg. Gap " + }, + { + "bbox": [ + 130, + 354, + 483, + 668 + ], + "type": "inline_equation", + "content": "= (|\\mathrm{Acc}_{\\mathcal{D}_t} - \\mathrm{Acc}_{\\mathcal{D}_t}^*| + |\\mathrm{Acc}_{\\mathcal{D}_f} - \\mathrm{Acc}_{\\mathcal{D}_f}^*| + |\\mathrm{Acc}_{\\mathcal{D}_r} - \\mathrm{Acc}_{\\mathcal{D}_r}^*| + |\\mathrm{MIA} - \\mathrm{MIA}^*|)/4" + }, + { + "bbox": [ + 130, + 354, + 483, + 668 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 354, + 483, + 668 + ], + "type": "inline_equation", + "content": "\\mathrm{Acc}_{\\mathcal{D}_t}^*" + }, + { + "bbox": [ + 130, + 354, + 483, + 668 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 354, + 483, + 668 + ], + "type": "inline_equation", + "content": "\\mathrm{Acc}_{\\mathcal{D}_f}^*" + }, + { + "bbox": [ + 130, + 354, + 483, + 668 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 354, + 483, + 668 + ], + "type": "inline_equation", + "content": "\\mathrm{Acc}_{\\mathcal{D}_r}^*" + }, + { + "bbox": [ + 130, + 354, + 483, + 668 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 354, + 483, + 668 + ], + "type": "inline_equation", + "content": "\\mathrm{MIA}^*" + }, + { + "bbox": [ + 130, + 354, + 483, + 668 + ], + "type": "text", + "content": " are metric values of the retrained model. A better performance of an unlearning method corresponds to a lower performance gap with retraining, measuring how close the unlearned model is to the retrained model. (iv) Run-time efficiency (RTE) [27]: evaluate the computational efficiency of the unlearning method. Specifically, we have " + }, + { + "bbox": [ + 130, + 354, + 483, + 668 + ], + "type": "inline_equation", + "content": "\\mathrm{RTE} = \\mathrm{T} / \\mathrm{T}_r" + }, + { + "bbox": [ + 130, + 354, + 483, + 668 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 354, + 483, + 668 + ], + "type": "inline_equation", + "content": "\\mathrm{T}" + }, + { + "bbox": [ + 130, + 354, + 483, + 668 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 354, + 483, + 668 + ], + "type": "inline_equation", + "content": "\\mathrm{T}_r" + }, + { + "bbox": [ + 130, + 354, + 483, + 668 + ], + "type": "text", + "content": " present the time for the unlearning methods to get the scrubbed models and the time for retraining from scratch with " + }, + { + "bbox": [ + 130, + 354, + 483, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r" + }, + { + "bbox": [ + 130, + 354, + 483, + 668 + ], + "type": "text", + "content": ", respectively. An efficient unlearning method should aim for minimal computational overhead compared to the baseline. (v) Relearn time [8,20,51]: the number of epochs to fine-tune the scrubbed models for regaining performance on " + }, + { + "bbox": [ + 130, + 354, + 483, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 130, + 354, + 483, + 668 + ], + "type": "text", + "content": " (i.e., reach the original models' accuracy on " + }, + { + "bbox": [ + 130, + 354, + 483, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 130, + 354, + 483, + 668 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 282, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 282, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 282, + 102 + ], + "type": "text", + "content": "Jing Wu, Mehrtash Harandi" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 144, + 135, + 470, + 479 + ], + "blocks": [ + { + "bbox": [ + 195, + 114, + 418, + 126 + ], + "lines": [ + { + "bbox": [ + 195, + 114, + 418, + 126 + ], + "spans": [ + { + "bbox": [ + 195, + 114, + 418, + 126 + ], + "type": "text", + "content": "Table 1: Quantitative results for forgetting " + }, + { + "bbox": [ + 195, + 114, + 418, + 126 + ], + "type": "inline_equation", + "content": "{10}\\%" + }, + { + "bbox": [ + 195, + 114, + 418, + 126 + ], + "type": "text", + "content": " data." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 144, + 135, + 470, + 479 + ], + "lines": [ + { + "bbox": [ + 144, + 135, + 470, + 479 + ], + "spans": [ + { + "bbox": [ + 144, + 135, + 470, + 479 + ], + "type": "table", + "html": "
MethodAccDf(↓)AccDt(↑)AccDr(↑)MIA(↑)Avg. Gap
CIFAR-100Retrain75.13±0.8574.69±0.0899.98±0.0150.22±0.62-
FT [55]97.98±1.3675.28±0.1299.95±0.029.64±3.616.01
GA [52]98.00±1.3475.59±0.1198.24±1.165.00±2.2517.68
IU [29]95.67±4.8272.13±4.5896.14±4.519.43±5.9816.93
BE [4]97.94±1.3874.16±0.0998.12±1.247.60±3.0516.96
BS [4]97.65±1.4873.20±0.1897.93±1.308.24±3.2317.01
l1-sparse [27]96.35±0.6770.06±0.4696.35±0.6721.33±1.9514.59
SalUn [12]88.56±1.1871.34±0.4899.40±0.3574.66±2.4810.45
Ours68.76±1.8173.17±0.2499.24±0.3042.42±2.064.11
CIFAR-10Retrain94.81±0.5394.26±0.14100.0±0.0013.05±0.64-
FT [55]99.15±0.4693.83±0.4599.84±0.113.01±0.933.74
GA [52]99.66±0.2394.57±0.0199.62±0.250.91±0.294.42
IU [29]98.08±2.1091.91±2.7398.01±2.264.01±3.444.16
BE [4]99.41±0.3893.79±0.1599.41±0.3816.16±0.782.19
BS [4]99.60±0.2594.24±0.0799.56±0.544.46±0.333.46
l1-sparse [27]94.17±0.4990.64±0.5296.64±0.5411.87±0.612.20
SalUn [12]98.07±0.4293.92±0.2599.89±0.0717.93±0.372.15
Ours95.40±1.4892.92±0.4898.93±0.579.56±2.131.62
SVHNRetrain91.81±1.1191.17±1.7797.73±1.1215.74±1.28-
FT [55]99.60±0.2495.28±0.0499.99±0.003.85±0.46.51
GA [52]98.41±0.2392.87±0.0698.52±0.295.96±0.384.72
IU [29]92.47±1.6286.92±2.0693.36±1.8217.00±2.672.63
BE [4]98.48±0.3392.62±0.0598.44±0.296.77±0.364.45
BS [4]98.29±0.3292.48±0.0298.36±0.296.74±0.394.36
l1-sparse [27]97.79±0.1093.59±3.1899.49±0.237.57±0.294.65
SalUn [12]96.29±4.8093.59±3.1897.41±4.9624.47±4.083.99
Ours91.62±1.1893.05±0.5699.51±0.1417.62±2.541.43
", + "image_path": "632b4b1f7d52b5d5b2e1911a3b33d77c170d99be741ff15af519f19cd7224e30.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 502, + 311, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 311, + 513 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 311, + 513 + ], + "type": "text", + "content": "4.1 Results on Classification Task" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 521, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 617 + ], + "type": "text", + "content": "In this experiment on SVHN, CIFAR-10, and CIFAR-100, we try to forget randomly selected " + }, + { + "bbox": [ + 130, + 521, + 482, + 617 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 521, + 482, + 617 + ], + "type": "text", + "content": " of the training data; on CelebAMask-HQ, we attempt to forget randomly selected " + }, + { + "bbox": [ + 130, + 521, + 482, + 617 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 521, + 482, + 617 + ], + "type": "text", + "content": " identities among 307 identities. We further apply GradCAM [49] to visualize regions where models focus on " + }, + { + "bbox": [ + 130, + 521, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{w}/" + }, + { + "bbox": [ + 130, + 521, + 482, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 521, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{w} / \\mathbf{o}" + }, + { + "bbox": [ + 130, + 521, + 482, + 617 + ], + "type": "text", + "content": " machine unlearning algorithms. In brief, the results suggest that Scissorhands has successfully induced forgetting for the relevant samples and classes, with minor degradation in model performance over the remaining data and classes. We will discuss the results in depth below." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Sample-wise unlearning. Table 1 presents the results when forgetting randomly selected samples. Scissorhands achieves the lowest average performance gap on all the presented datasets and shows good generalization ability compared to" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 143, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 143, + 91, + 447, + 102 + ], + "type": "text", + "content": "Scissorhands: Scrub Data Influence via Connection Sensitivity in Networks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 156, + 136, + 458, + 264 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 477, + 126 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 477, + 126 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 477, + 126 + ], + "type": "text", + "content": "Table 2: Quantitative results for forgetting " + }, + { + "bbox": [ + 135, + 114, + 477, + 126 + ], + "type": "inline_equation", + "content": "{10}\\%" + }, + { + "bbox": [ + 135, + 114, + 477, + 126 + ], + "type": "text", + "content": " identities on the CelebAMask-HQ." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 156, + 136, + 458, + 264 + ], + "lines": [ + { + "bbox": [ + 156, + 136, + 458, + 264 + ], + "spans": [ + { + "bbox": [ + 156, + 136, + 458, + 264 + ], + "type": "table", + "html": "
MethodAccDf(↓)AccDt(↑)AccDr(↑)MIA(↑)Avg. Gap
Retrain0.00±0.0087.02±0.8099.96±0.01100.0±0.00-
FT [55]99.94±0.1288.59±0.5999.97±7.025.28±2.0349.06
GA [52]87.60±8.7181.22±2.1199.74±0.2651.37±5.9635.56
IU [29]88.92±10.2570.24±11.7795.27±5.0729.59±18.5945.20
BE [4]69.07±2.7344.11±2.0895.58±1.2346.24±5.9042.53
BS [4]98.18±1.9281.92±0.2799.86±0.0345.93±5.1139.36
l1-sparse [27]98.81±0.7289.37±0.7099.97±0.0076.78±5.6631.10
SalUn [12]0.00±0.0078.36±1.3496.90±1.11100.0±0.002.93
Ours1.52±2.7380.18±6.6097.20±3.8199.83±0.352.82
", + "image_path": "1020f8174793d2d1f244284ff718413d2df0ef903a2abd6edb5301a59b5e3a0a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 159, + 277, + 299, + 353 + ], + "blocks": [ + { + "bbox": [ + 159, + 277, + 299, + 353 + ], + "lines": [ + { + "bbox": [ + 159, + 277, + 299, + 353 + ], + "spans": [ + { + "bbox": [ + 159, + 277, + 299, + 353 + ], + "type": "image", + "image_path": "a785f5a123cef880a8214a26247f8d55b6be1f71058b3f264d6dd46c0648a217.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 362, + 479, + 374 + ], + "lines": [ + { + "bbox": [ + 132, + 362, + 479, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 362, + 479, + 374 + ], + "type": "text", + "content": "Fig. 1: Visualizations of regions where models focus on generated by GradCAM [49]." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 315, + 277, + 455, + 353 + ], + "blocks": [ + { + "bbox": [ + 315, + 277, + 455, + 353 + ], + "lines": [ + { + "bbox": [ + 315, + 277, + 455, + 353 + ], + "spans": [ + { + "bbox": [ + 315, + 277, + 455, + 353 + ], + "type": "image", + "image_path": "44b0c00ac887072f2a3e599bc8d2653d331a7a141fb1fbc3f2eb13b8495116f3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 399, + 481, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 481, + 460 + ], + "type": "text", + "content": "Retrain. For example, on SVHN, Scissorhands achieves " + }, + { + "bbox": [ + 130, + 399, + 481, + 460 + ], + "type": "inline_equation", + "content": "\\sim 91.62\\%" + }, + { + "bbox": [ + 130, + 399, + 481, + 460 + ], + "type": "text", + "content": " accuracy on the forgetting data " + }, + { + "bbox": [ + 130, + 399, + 481, + 460 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 130, + 399, + 481, + 460 + ], + "type": "text", + "content": ", closely matching the " + }, + { + "bbox": [ + 130, + 399, + 481, + 460 + ], + "type": "inline_equation", + "content": "91.81\\%" + }, + { + "bbox": [ + 130, + 399, + 481, + 460 + ], + "type": "text", + "content": " accuracy of a fully retrained model. Furthermore, the retrained model reaches around " + }, + { + "bbox": [ + 130, + 399, + 481, + 460 + ], + "type": "inline_equation", + "content": "91\\%" + }, + { + "bbox": [ + 130, + 399, + 481, + 460 + ], + "type": "text", + "content": " accuracy on the test dataset, Scissorhands exhibits an accuracy of roughly " + }, + { + "bbox": [ + 130, + 399, + 481, + 460 + ], + "type": "inline_equation", + "content": "93\\%" + }, + { + "bbox": [ + 130, + 399, + 481, + 460 + ], + "type": "text", + "content": ", showcasing the enhanced generalization of our scrubbed model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 460, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 482, + 533 + ], + "type": "text", + "content": "When evaluating solely based on forgetting accuracy on CIFAR-10, the " + }, + { + "bbox": [ + 130, + 460, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 130, + 460, + 482, + 533 + ], + "type": "text", + "content": "-sparse method might seem to be the most effective baseline, achieving approximately " + }, + { + "bbox": [ + 130, + 460, + 482, + 533 + ], + "type": "inline_equation", + "content": "94.17\\%" + }, + { + "bbox": [ + 130, + 460, + 482, + 533 + ], + "type": "text", + "content": ". Nonetheless, this perceived advantage is offset by a decline in both test accuracy and the accuracy of retained data. In contrast, Scissorhands shows a good trade-off between preserving model utility on the remaining data, generalization on unseen sets, and erasing the influence of the forgetting data." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Similarly, on CIFAR-100, SalUn may initially stand out with its impressive MIA accuracy of approximately " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "74.66\\%" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": ", relying solely on a single metric to gauge the performance of machine unlearning can be misleading, as it might not fully capture the method's effectiveness [12]. While SalUn achieves a forgetting accuracy of about " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "88.56\\%" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " and a test accuracy of " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "71.34\\%" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": ", Scissorhands demonstrates a slightly lower forgetting accuracy of around " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "68.76\\%" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " but achieves a higher test accuracy of " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "73.17\\%" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": ". This indicates that Scissorhands achieves the best trade-off between eradicating data traces and generalizing to unseen data. Soon, we will show that on this dataset, the relearning time of Scissorhands is greater than 200 to regain the performance on " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": ", which indicates that Scissorhands has efficiently scrubbed information about " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 282, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 282, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 282, + 102 + ], + "type": "text", + "content": "Jing Wu, Mehrtash Harandi" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": "Overall, Scissorhands illustrates its capability to adeptly balance the objectives of data removal, model utility preservation, and generalization to unseen data, ensuring that the influence of forgotten data is minimized without significantly compromising the overall performance and utility of the model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 182, + 482, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 482, + 290 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 482, + 290 + ], + "type": "text", + "content": "Class-wise unlearning. Furthermore, Table 2 presents the outcomes of our analysis for the scenario where forgetting " + }, + { + "bbox": [ + 130, + 182, + 482, + 290 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 182, + 482, + 290 + ], + "type": "text", + "content": " of identities on the CelebAMask-HQ. Images are rescaled to " + }, + { + "bbox": [ + 130, + 182, + 482, + 290 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 130, + 182, + 482, + 290 + ], + "type": "text", + "content": ", and a model pre-trained with ImageNet1K [9] is employed. In this context, Scissorhands manifests the smallest average performance gap with retrained models. Notably, the Retrain method achieves perfect metrics in terms of MIA accuracy, which is " + }, + { + "bbox": [ + 130, + 182, + 482, + 290 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 130, + 182, + 482, + 290 + ], + "type": "text", + "content": ", and completely erases the identity information from the dataset, as indicated by a " + }, + { + "bbox": [ + 130, + 182, + 482, + 290 + ], + "type": "inline_equation", + "content": "0.00\\%" + }, + { + "bbox": [ + 130, + 182, + 482, + 290 + ], + "type": "text", + "content": " accuracy on the forgotten data " + }, + { + "bbox": [ + 130, + 182, + 482, + 290 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 130, + 182, + 482, + 290 + ], + "type": "text", + "content": ". However, retraining proves impractical and limited for unlearning purposes in general, functioning merely as an oracle in our experiments." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 292, + 482, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 292, + 482, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 292, + 482, + 434 + ], + "type": "text", + "content": "Among the baselines, fine-tuning (FT) and " + }, + { + "bbox": [ + 130, + 292, + 482, + 434 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 130, + 292, + 482, + 434 + ], + "type": "text", + "content": "-sparse exhibit higher accuracies on the remaining data " + }, + { + "bbox": [ + 130, + 292, + 482, + 434 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 130, + 292, + 482, + 434 + ], + "type": "text", + "content": " and unseen data " + }, + { + "bbox": [ + 130, + 292, + 482, + 434 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_t" + }, + { + "bbox": [ + 130, + 292, + 482, + 434 + ], + "type": "text", + "content": "; yet, their accuracies on forgetting remain high as well, highlighting a less effective balance between eradicating data traces and generalizing to unseen data. In contrast, Scissorhands and SalUn demonstrate superior capabilities in effectively forgetting identities with minimal impact on the models' overall performance. SalUn achieves an MIA accuracy of " + }, + { + "bbox": [ + 130, + 292, + 482, + 434 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 130, + 292, + 482, + 434 + ], + "type": "text", + "content": " and a forgetting accuracy of " + }, + { + "bbox": [ + 130, + 292, + 482, + 434 + ], + "type": "inline_equation", + "content": "0.00\\%" + }, + { + "bbox": [ + 130, + 292, + 482, + 434 + ], + "type": "text", + "content": ", effectively nullifying the identity information similar to the retrained models. Scissorhands slightly surpasses SalUn in terms of the average performance gap with retrained models, maintaining a high MIA accuracy of " + }, + { + "bbox": [ + 130, + 292, + 482, + 434 + ], + "type": "inline_equation", + "content": "99.83\\%" + }, + { + "bbox": [ + 130, + 292, + 482, + 434 + ], + "type": "text", + "content": " and a significantly low accuracy on the forgotten data " + }, + { + "bbox": [ + 130, + 292, + 482, + 434 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 130, + 292, + 482, + 434 + ], + "type": "text", + "content": " (i.e., " + }, + { + "bbox": [ + 130, + 292, + 482, + 434 + ], + "type": "inline_equation", + "content": "1.52\\%" + }, + { + "bbox": [ + 130, + 292, + 482, + 434 + ], + "type": "text", + "content": "), while achieving a test accuracy of " + }, + { + "bbox": [ + 130, + 292, + 482, + 434 + ], + "type": "inline_equation", + "content": "80.18\\%" + }, + { + "bbox": [ + 130, + 292, + 482, + 434 + ], + "type": "text", + "content": " and an accuracy of " + }, + { + "bbox": [ + 130, + 292, + 482, + 434 + ], + "type": "inline_equation", + "content": "97.20\\%" + }, + { + "bbox": [ + 130, + 292, + 482, + 434 + ], + "type": "text", + "content": " on the retained data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 437, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 640 + ], + "type": "text", + "content": "We further employ the GradCAM [49] to illustrate the focus area of models " + }, + { + "bbox": [ + 130, + 437, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{w}/" + }, + { + "bbox": [ + 130, + 437, + 482, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 437, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{w} / \\mathbf{o}" + }, + { + "bbox": [ + 130, + 437, + 482, + 640 + ], + "type": "text", + "content": " machine unlearning algorithms. As shown in Figure 1, when evaluated on the remaining data " + }, + { + "bbox": [ + 130, + 437, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r" + }, + { + "bbox": [ + 130, + 437, + 482, + 640 + ], + "type": "text", + "content": ", both SalUn and Scissorhands even pay more attention to the facial features for identity classification than the retrained model. Conversely, when evaluated on the forgetting data " + }, + { + "bbox": [ + 130, + 437, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 130, + 437, + 482, + 640 + ], + "type": "text", + "content": ", the scrubbed model by Scissorhands shifts its attention towards regions that are least associated with the facial features. We hypothesize that in the scenario where the task involves forgetting identities, both the remaining data and the forgetting data are face images, the scrubbed models would need a more nuanced approach to distinguish between individuals. Specifically, because the task at hand requires the discernment of subtle differences across facial features to effectively differentiate and forget specific identities while retaining others, the models adapt by developing a refined sensitivity towards those facial characteristics that are most indicative of individual identity. This adaptation enables the models to maintain high accuracy on the remaining data by focusing more intently on the facial features that are crucial for identity recognition, thereby enhancing their ability to generalize and accurately classify identities." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 482, + 666 + ], + "type": "text", + "content": "These results underscore our proposed machine unlearning algorithm Scisserhands's superior capabilities in balancing identity forgetting and model utility" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 143, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 143, + 91, + 448, + 102 + ], + "type": "text", + "content": "Scissorhands: Scrub Data Influence via Connection Sensitivity in Networks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 101 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 101 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 148, + 146, + 467, + 216 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 3: Influence of projection and initialization strategies (re-initializing parameters with uniform distribution " + }, + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": ", Gaussian distribution " + }, + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": ", and constant value of " + }, + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "inline_equation", + "content": "\\mathbf{0} / \\mathbf{1}" + }, + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 148, + 146, + 467, + 216 + ], + "lines": [ + { + "bbox": [ + 148, + 146, + 467, + 216 + ], + "spans": [ + { + "bbox": [ + 148, + 146, + 467, + 216 + ], + "type": "table", + "html": "
ProjectionInitializationAccDf(↓)AccDt(↑)AccDr(↑)MIA(↑)Avg. Gap
✓U95.40±1.4892.92±0.4898.93±0.579.56±2.131.62
XU98.14±0.6093.14±0.1699.73±0.166.08±0.782.92
✓N94.68±2.5092.13±2.2097.94±2.7110.18±3.811.80
✓096.10±2.0592.92±0.9999.09±1.198.76±3.301.96
X184.28±0.4983.38±0.5887.18±0.9020.67±0.4110.46
", + "image_path": "41241239017a32499826cd1338fe3b60b7ed4972549993a96bc8e6d4b1036d12.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 137, + 228, + 250, + 308 + ], + "blocks": [ + { + "bbox": [ + 137, + 228, + 250, + 308 + ], + "lines": [ + { + "bbox": [ + 137, + 228, + 250, + 308 + ], + "spans": [ + { + "bbox": [ + 137, + 228, + 250, + 308 + ], + "type": "image", + "image_path": "49ce2eb998962f055f1c1505c0eab18abf65a9e97e0ffcfdbbd67a2a5ba70f36.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 319, + 482, + 343 + ], + "lines": [ + { + "bbox": [ + 130, + 319, + 482, + 343 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 482, + 343 + ], + "type": "text", + "content": "Fig. 2: Influence of the percent value " + }, + { + "bbox": [ + 130, + 319, + 482, + 343 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 319, + 482, + 343 + ], + "type": "text", + "content": " in the trimming process, the balance term " + }, + { + "bbox": [ + 130, + 319, + 482, + 343 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 319, + 482, + 343 + ], + "type": "text", + "content": " of Equation (3), and the ratio of forgetting data used in the trimming process." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 253, + 228, + 364, + 307 + ], + "blocks": [ + { + "bbox": [ + 253, + 228, + 364, + 307 + ], + "lines": [ + { + "bbox": [ + 253, + 228, + 364, + 307 + ], + "spans": [ + { + "bbox": [ + 253, + 228, + 364, + 307 + ], + "type": "image", + "image_path": "42debbda1a0e6c1c2adbc107457577a88ce1b522f4d8e5506df22c64479722eb.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 366, + 228, + 477, + 308 + ], + "blocks": [ + { + "bbox": [ + 366, + 228, + 477, + 308 + ], + "lines": [ + { + "bbox": [ + 366, + 228, + 477, + 308 + ], + "spans": [ + { + "bbox": [ + 366, + 228, + 477, + 308 + ], + "type": "image", + "image_path": "3b54eb9b20dc279b9ad11a69362034f579e60d0e6a084e51e7c9813020958b2b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 365, + 482, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 390 + ], + "type": "text", + "content": "performance, Scissorhands not only minimizes privacy risks but also maintains the integrity and applicability of the model to unseen data." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 406, + 238, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 406, + 238, + 418 + ], + "spans": [ + { + "bbox": [ + 131, + 406, + 238, + 418 + ], + "type": "text", + "content": "4.2 Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 426, + 482, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 536 + ], + "type": "text", + "content": "Component analysis. As detailed in Table 3, the inclusion of different procedures in Scissorhands has a noticeable impact. Notably, the employment of the projection mechanism plays a crucial role in eliminating the influence of forgetting data " + }, + { + "bbox": [ + 130, + 426, + 482, + 536 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 130, + 426, + 482, + 536 + ], + "type": "text", + "content": ", affecting the overall utility of the model. The rationale behind the projection process is to preclude the reacquisition of critical information about the forgetting data " + }, + { + "bbox": [ + 130, + 426, + 482, + 536 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 130, + 426, + 482, + 536 + ], + "type": "text", + "content": " during the repair phase. Moreover, Table 3 reveals that initializing parameters with a uniform distribution " + }, + { + "bbox": [ + 130, + 426, + 482, + 536 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 130, + 426, + 482, + 536 + ], + "type": "text", + "content": " or opting for zero initialization facilitates effective unlearning, in contrast to re-initializing parameters with a constant value of one, which impedes finding a viable solution to Equation (5)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "text", + "content": "Hyper-parameter impact. Figure 2 presents the effects of varying hyperparameters, such as the percentage value " + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "inline_equation", + "content": "k\\%" + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "text", + "content": " for the trimming process, the balance term " + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "text", + "content": " of Equation (3) in the repairing phase, and the ratio of forgetting data used in the trimming process. An increase " + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "text", + "content": " results in higher accuracies on data, as fewer parameters associated with the forgetting data " + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "text", + "content": " are reinitialized. Empirically, we observe choosing a value of " + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "inline_equation", + "content": "k \\geq 0.9" + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "text", + "content": " will result in stable performances. That said, this is a hyperparameter of the algorithm and can be chosen by measuring the performances on " + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "inline_equation", + "content": "D_{f}" + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "inline_equation", + "content": "D_{r}" + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "text", + "content": " during training. The hyper-parameter " + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "text", + "content": " controls the balance between retaining and forgetting data, the accuracies on test data and remaining data show a peak at " + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\lambda = 0.05" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 282, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 282, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 282, + 102 + ], + "type": "text", + "content": "Jing Wu, Mehrtash Harandi" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 115, + 309, + 217 + ], + "blocks": [ + { + "bbox": [ + 138, + 115, + 309, + 217 + ], + "lines": [ + { + "bbox": [ + 138, + 115, + 309, + 217 + ], + "spans": [ + { + "bbox": [ + 138, + 115, + 309, + 217 + ], + "type": "image", + "image_path": "3fcc8d75600a3a75b9f3649de707efdfc9679a9c06b8e6795ed4f33fbc87b2b3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 226, + 482, + 270 + ], + "lines": [ + { + "bbox": [ + 130, + 226, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 482, + 270 + ], + "type": "text", + "content": "Fig.3: Quantity of nudity content detected using the NudeNet classifier from 1K sampled images and I2P data. We observed a high false positive rate for exposed female genitalia/breast using the NudeNet classifier on generated I2P images. The flagged images can be found in Appendix B." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 309, + 115, + 475, + 216 + ], + "blocks": [ + { + "bbox": [ + 309, + 115, + 475, + 216 + ], + "lines": [ + { + "bbox": [ + 309, + 115, + 475, + 216 + ], + "spans": [ + { + "bbox": [ + 309, + 115, + 475, + 216 + ], + "type": "image", + "image_path": "f1cae026a9d6fe32206f7d1e78adee6957744618fdf13d1bdca542d8c70f0ed5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 136, + 281, + 161, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 281, + 161, + 289 + ], + "spans": [ + { + "bbox": [ + 136, + 281, + 161, + 289 + ], + "type": "text", + "content": "SD v1.4" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 137, + 289, + 249, + 327 + ], + "blocks": [ + { + "bbox": [ + 137, + 289, + 249, + 327 + ], + "lines": [ + { + "bbox": [ + 137, + 289, + 249, + 327 + ], + "spans": [ + { + "bbox": [ + 137, + 289, + 249, + 327 + ], + "type": "image", + "image_path": "713a1ee117321f659b6b8469b69a2fca897a11456e862636577b74899a7e1ef3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 337, + 480, + 360 + ], + "lines": [ + { + "bbox": [ + 131, + 337, + 480, + 360 + ], + "spans": [ + { + "bbox": [ + 131, + 337, + 480, + 360 + ], + "type": "text", + "content": "Fig. 4: Sample images with the prompt from " + }, + { + "bbox": [ + 131, + 337, + 480, + 360 + ], + "type": "inline_equation", + "content": "c_f = \\{ \\text{nudity}, \\text{naked}, \\text'erotic', \\text'sexual' \\}" + }, + { + "bbox": [ + 131, + 337, + 480, + 360 + ], + "type": "text", + "content": " generated by SDs w/ and w/o machine unlearning algorithms. Best viewed in color." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 251, + 282, + 269, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 282, + 269, + 289 + ], + "spans": [ + { + "bbox": [ + 251, + 282, + 269, + 289 + ], + "type": "text", + "content": "SalUn" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 251, + 289, + 362, + 327 + ], + "blocks": [ + { + "bbox": [ + 251, + 289, + 362, + 327 + ], + "lines": [ + { + "bbox": [ + 251, + 289, + 362, + 327 + ], + "spans": [ + { + "bbox": [ + 251, + 289, + 362, + 327 + ], + "type": "image", + "image_path": "2689f618d6fa3b839545dc332ea13e39ec40792a50a53b9884239f241373cb04.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 364, + 283, + 379, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 283, + 379, + 289 + ], + "spans": [ + { + "bbox": [ + 364, + 283, + 379, + 289 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 364, + 289, + 396, + 327 + ], + "blocks": [ + { + "bbox": [ + 364, + 289, + 396, + 327 + ], + "lines": [ + { + "bbox": [ + 364, + 289, + 396, + 327 + ], + "spans": [ + { + "bbox": [ + 364, + 289, + 396, + 327 + ], + "type": "image", + "image_path": "50e51cbd01bfd27c35570d50cbcd24fc19546610e8d2930a676a0ce37f2124bc.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 402, + 282, + 477, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 282, + 477, + 289 + ], + "spans": [ + { + "bbox": [ + 402, + 282, + 477, + 289 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 402, + 282, + 477, + 289 + ], + "type": "text", + "content": " Added by authors for publication" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 402, + 289, + 476, + 327 + ], + "blocks": [ + { + "bbox": [ + 402, + 289, + 476, + 327 + ], + "lines": [ + { + "bbox": [ + 402, + 289, + 476, + 327 + ], + "spans": [ + { + "bbox": [ + 402, + 289, + 476, + 327 + ], + "type": "image", + "image_path": "f0dde109d4f8e9617c60fb4672f44a226e66ce417cac5949c9bb12b0d8e60f08.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 384, + 482, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 482, + 515 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 482, + 515 + ], + "type": "text", + "content": "when forgetting " + }, + { + "bbox": [ + 130, + 384, + 482, + 515 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 384, + 482, + 515 + ], + "type": "text", + "content": " data on CIFAR-10. Generally, as " + }, + { + "bbox": [ + 130, + 384, + 482, + 515 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 384, + 482, + 515 + ], + "type": "text", + "content": " escalates, it places greater emphasis on the process of data omission, leading to a decrease in forgetting accuracy. As the ratio " + }, + { + "bbox": [ + 130, + 384, + 482, + 515 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 384, + 482, + 515 + ], + "type": "text", + "content": " increases, all accuracy metrics exhibit a decline, suggesting that employing more forgetting data for the purpose of forgetting in order to isolate critical knowledge about the forgetting data also inadvertently results in the loss of pertinent information regarding the remaining data. This phenomenon occurs particularly when the data designated for forgetting shares the same distribution as the data that is retained. In essence, while aiming to enhance the specificity of the forgetting process by pinpointing essential details about the data to be forgotten, there's a risk of simultaneously diminishing knowledge pertinent to the retained data." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 535, + 305, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 535, + 305, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 535, + 305, + 547 + ], + "type": "text", + "content": "4.3 Case Study: Stable Diffusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 557, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 640 + ], + "type": "text", + "content": "We also employed our algorithm to mitigate the generation of inappropriate content in Stable Diffusion (SD) models [44]. Specifically, our primary objective is to effectively eliminate the concept of nudity from the model's generative capabilities. To this end, followed [26] and employed SD v1.4 for sampling with 50 time steps. We evaluate on 1K generated images with prompts " + }, + { + "bbox": [ + 130, + 557, + 482, + 640 + ], + "type": "inline_equation", + "content": "c_f = \\{ \\text{nudity}, \\text{naked}, \\text'erotic', \\text'sexual' \\}" + }, + { + "bbox": [ + 130, + 557, + 482, + 640 + ], + "type": "text", + "content": " and 4703 generated images with I2P [46] using the open-source NudeNet classifier." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "content": "Figures 3 and 4 present the results for SD v2.1 (trained on a dataset filtered for nudity), SalUn, and Scissorhands. In Figure 3, y-axis denotes the number of" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 143, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 143, + 91, + 447, + 102 + ], + "type": "text", + "content": "Scissorhands: Scrub Data Influence via Connection Sensitivity in Networks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 261 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 261 + ], + "type": "text", + "content": "exposed body parts generated by the SD v1.4 model, it presents the percentage change in exposed body parts w.r.t. SD v1.4. When comparing on I2P data, our algorithms outperform SalUn on 'Feet', while coming short on 'Female Breast and Genitalia'. While this could suggest the superiority of SalUn on this data, the gap is not that big in reality, as out of the 9 images flagged for containing exposed female breasts, 7 were inaccurately identified, and none of the flagged images depicted exposed female genitalia. The use of the NudeNet classifier on generated I2P images exhibits a significant rate of false positives, as highlighted in [26] as well. When we use prompts " + }, + { + "bbox": [ + 130, + 116, + 482, + 261 + ], + "type": "inline_equation", + "content": "c_f" + }, + { + "bbox": [ + 130, + 116, + 482, + 261 + ], + "type": "text", + "content": " explicitly associated with nudity, no exposed sensitive content is detected for both Scissorhands and SalUn methods. We can conclude that, Scissorhands significantly reduces the amount of nudity content compared to SD v1.4, SD v2.1, showing comparable results to SalUn." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 277, + 346, + 291 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 277, + 346, + 291 + ], + "spans": [ + { + "bbox": [ + 132, + 277, + 346, + 291 + ], + "type": "text", + "content": "5 Limitations and Broader Impact" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 301, + 482, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 301, + 482, + 469 + ], + "spans": [ + { + "bbox": [ + 130, + 301, + 482, + 469 + ], + "type": "text", + "content": "Our proposed unlearning algorithm, Scissorhands, can achieve comparable performance to state-of-the-art unlearning methods. We addressed potential issues of conflicting gradients arising from direct conflicts between features or patterns in the forgetting and remaining data, ensuring training stability. However, the limitation is the incomplete concept erasure shown in Figure 8 in Appendix B, and the potential for biased models due to the selective deletion of information. This is a crucial consideration, especially when unlearning algorithms are applied with malicious intent. Furthermore, in experiments on stable diffusion, while our method can successfully erase specific concepts, finding an optimal balance between forgetting and maintaining high image quality remains challenging. There is also a risk that unlearning algorithms might be used to alter or remove concepts in unethical or malicious ways. It is our hope that Scissorhands offers a fresh perspective on practical and effective machine unlearning, emphasizing the need for ethical guidelines to prevent misuse in its application." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 486, + 220, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 486, + 220, + 498 + ], + "spans": [ + { + "bbox": [ + 132, + 486, + 220, + 498 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "In this work, we proposed Scissorhands, an effective and practical machine unlearning algorithm. Our algorithm identified the important parameters w.r.t. the forgetting data in single-shot via connection sensitivity, then re-initialized these parameters for scrubbing the influence of the forgetting data. An effective gradient projection-based technique is applied to enhance the model utility while excluding the information w.r.t. the forgetting data. Our evaluations on various datasets showed that, compared with other methods, our approach offers superior performance across image classification and image generation tasks. Future works could explore how Scissorhands performs on regression and NLP tasks, and further investigation into the scenario where data are not available, as well as unlearning with sequential/time-series data where could introduce unique challenges, e.g., the need to consider temporal dependencies, which may require new frameworks specifically tailored for sequential data." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 282, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 282, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 282, + 102 + ], + "type": "text", + "content": "Jing Wu, Mehrtash Harandi" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 163 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 163 + ], + "type": "text", + "content": "Mehrtash Harandi is supported by funding from the Australian Research Council (ARC) Discovery Program DP230101176." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 182, + 197, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 182, + 197, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 182, + 197, + 194 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 205, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 205, + 481, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 205, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 138, + 205, + 481, + 239 + ], + "type": "text", + "content": "1. Birhane, A., Prabhu, V.U.: Large image datasets: A pyrrhic win for computer vision? In: 2021 IEEE Winter Conference on Applications of Computer Vision (WACV). pp. 1536-1546. IEEE (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 239, + 481, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 239, + 481, + 272 + ], + "spans": [ + { + "bbox": [ + 138, + 239, + 481, + 272 + ], + "type": "text", + "content": "2. Bourtoule, L., Chandrasekaran, V., Choquette-Choo, C.A., Jia, H., Travers, A., Zhang, B., Lie, D., Papernot, N.: Machine unlearning. In: 2021 IEEE Symposium on Security and Privacy (SP). pp. 141-159 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 272, + 481, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 272, + 481, + 294 + ], + "spans": [ + { + "bbox": [ + 138, + 272, + 481, + 294 + ], + "type": "text", + "content": "3. Cao, Y., Yang, J.: Towards making systems forget with machine unlearning. In: 2015 IEEE Symposium on Security and Privacy (SP). pp. 463-480 (2015)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 294, + 481, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 294, + 481, + 337 + ], + "spans": [ + { + "bbox": [ + 138, + 294, + 481, + 337 + ], + "type": "text", + "content": "4. Chen, M., Gao, W., Liu, G., Peng, K., Wang, C.: Boundary unlearning: Rapid forgetting of deep networks via shifting the decision boundary. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7766-7775 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 337, + 481, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 337, + 481, + 370 + ], + "spans": [ + { + "bbox": [ + 138, + 337, + 481, + 370 + ], + "type": "text", + "content": "5. Chen, M., Zhang, Z., Wang, T., Backes, M., Humbert, M., Zhang, Y.: Graph unlearning. In: Proceedings of the 2022 ACM SIGSAC Conference on Computer and Communications Security. pp. 499-513 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 370, + 481, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 370, + 481, + 403 + ], + "spans": [ + { + "bbox": [ + 138, + 370, + 481, + 403 + ], + "type": "text", + "content": "6. Chen, W., Song, D., Li, B.: Trojdiff: Trojan attacks on diffusion models with diverse targets. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4035-4044 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 403, + 481, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 403, + 481, + 436 + ], + "spans": [ + { + "bbox": [ + 138, + 403, + 481, + 436 + ], + "type": "text", + "content": "7. Cheng, J., Dasoulas, G., He, H., Agarwal, C., Zitnik, M.: Gnndelete: A general strategy for unlearning in graph neural networks. In: International Conference on Learning Representations (ICLR) (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 436, + 481, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 436, + 481, + 458 + ], + "spans": [ + { + "bbox": [ + 138, + 436, + 481, + 458 + ], + "type": "text", + "content": "8. Chandawat, V.S., Tarun, A.K., Mandal, M., Kankanhalli, M.: Zero-shot machine unlearning. IEEE Transactions on Information Forensics and Security (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 458, + 481, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 458, + 481, + 491 + ], + "spans": [ + { + "bbox": [ + 138, + 458, + 481, + 491 + ], + "type": "text", + "content": "9. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 491, + 441, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 491, + 441, + 502 + ], + "spans": [ + { + "bbox": [ + 138, + 491, + 441, + 502 + ], + "type": "text", + "content": "0. Dennis, C.R., Weisberg, S.: Residuals and influence in regression (1982)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 502, + 481, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 502, + 481, + 523 + ], + "spans": [ + { + "bbox": [ + 138, + 502, + 481, + 523 + ], + "type": "text", + "content": "1. Fan, C., Liu, J., Hero, A., Liu, S.: Challenging forgets: Unveiling the worst-case forget sets in machine unlearning. arXiv preprint arXiv:2403.07362 (2024)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 523, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 523, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 523, + 481, + 567 + ], + "type": "text", + "content": "2. Fan, C., Liu, J., Zhang, Y., Wei, D., Wong, E., Liu, S.: Salun: Empowering machine unlearning via gradient-based weight saliency in both image classification and generation. In: International Conference on Learning Representations (ICLR) (2024)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 567, + 481, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 481, + 589 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 481, + 589 + ], + "type": "text", + "content": "3. Foster, J., Schoepf, S., Brintrup, A.: Fast machine unlearning without retraining through selective synaptic dampening. arXiv preprint arXiv:2308.07707 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 589, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 589, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 589, + 481, + 632 + ], + "type": "text", + "content": "4. Gandhi, S., Kokkula, S., Chaudhuri, A., Magnani, A., Stanley, T., Ahmadi, B., Kandaswamy, V., Ovenc, O., Mannor, S.: Scalable detection of offensive and noncompliant content/logo in product images. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2247-2256 (2020)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 481, + 666 + ], + "type": "text", + "content": "5. Gandikota, R., Materzynska, J., Fiotto-Kaufman, J., Bau, D.: Erasing concepts from diffusion models. In: 2023 IEEE International Conference on Computer Vision (ICCV) (2023)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 143, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 143, + 91, + 447, + 102 + ], + "type": "text", + "content": "Scissorhands: Scrub Data Influence via Connection Sensitivity in Networks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "16. Gandikota, R., Orgad, H., Belinkov, Y., Materzyńska, J., Bau, D.: Unified concept editing in diffusion models. arXiv preprint arXiv:2308.14761 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 480, + 171 + ], + "type": "text", + "content": "17. Ginart, A., Guan, M., Valiant, G., Zou, J.Y.: Making ai forget you: Data deletion in machine learning. In: Advances in Neural Information Processing Systems (NeurIPS). vol. 32 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 480, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 480, + 204 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 480, + 204 + ], + "type": "text", + "content": "18. Goel, S., Prabhu, A., Sanyal, A., Lim, S.N., Torr, P., Kumaraguru, P.: Towards adversarial evaluations for inexact machine unlearning. arXiv preprint arXiv:2201.06640 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 205, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 205, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 134, + 205, + 480, + 237 + ], + "type": "text", + "content": "19. Golatkar, A., Achille, A., Ravichandran, A., Polito, M., Soatto, S.: Mixed-privacy forgetting in deep networks. In: 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 792-801 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 237, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 237, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 133, + 237, + 480, + 270 + ], + "type": "text", + "content": "20. Golatkar, A., Achille, A., Soatto, S.: Eternal sunshine of the spotless net: Selective forgetting in deep networks. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9301-9309 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 270, + 480, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 480, + 304 + ], + "type": "text", + "content": "21. Golatkar, A., Achille, A., Soatto, S.: Forgetting outside the box: Scrubbing deep networks of information accessible from input-output observations. In: European Conference on Computer Vision (ECCV). pp. 383-398. Springer (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 480, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 326 + ], + "type": "text", + "content": "22. Goldman, E.: An introduction to the california consumer privacy act (ccpa). Santa Clara Univ. Legal Studies Research Paper (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 326, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 369 + ], + "type": "text", + "content": "23. Guo, C., Goldstein, T., Hannun, A., Van Der Maaten, L.: Certified data removal from machine learning models. In: Proceedings of the 37th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 119, pp. 3832-3842. PMLR (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "text", + "content": "24. Halimi, A., Kadhe, S., Rawat, A., Baracaldo, N.: Federated unlearning: How to efficiently erase a client in fl? In: International conference on machine learning. PMLR (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 402, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 480, + 424 + ], + "type": "text", + "content": "25. Han, S., Pool, J., Tran, J., Dally, W.: Learning both weights and connections for efficient neural network. vol. 28 (2015)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "type": "text", + "content": "26. Heng, A., Soh, H.: Selective amnesia: A continual learning approach to forgetting in deep generative models. In: Advances in Neural Information Processing Systems (NeurIPS) (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 456, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 456, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 456, + 480, + 490 + ], + "type": "text", + "content": "27. Jia, J., Liu, J., Ram, P., Yao, Y., Liu, G., Liu, Y., Sharma, P., Liu, S.: Model sparsification can simplify machine unlearning. In: Advances in Neural Information Processing Systems (NeurIPS) (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 490, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 480, + 522 + ], + "type": "text", + "content": "28. Karasuyama, M., Takeuchi, I.: Multiple incremental decremental learning of support vector machines. IEEE Transactions on Neural Networks 21(7), 1048-1059 (2010)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 522, + 480, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 480, + 545 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 480, + 545 + ], + "type": "text", + "content": "29. Koh, P.W., Liang, P.: Understanding black-box predictions via influence functions. In: International conference on machine learning. pp. 1885-1894. PMLR (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 545, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 480, + 567 + ], + "type": "text", + "content": "30. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 567, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 600 + ], + "type": "text", + "content": "31. Lee, C.H., Liu, Z., Wu, L., Luo, P.: Maskgan: Towards diverse and interactive facial image manipulation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5549-5558 (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "text", + "content": "32. Lee, N., Ajanthan, T., Gould, S., Torr, P.H.: A signal propagation perspective for pruning neural networks at initialization. In: International Conference on Learning Representations (ICLR) (2020)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "text", + "content": "33. Lee, N., Ajanthan, T., Torr, P.H.: Snip: Single-shot network pruning based on connection sensitivity. In: International Conference on Learning Representations (ICLR) (2019)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 281, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 281, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 281, + 102 + ], + "type": "text", + "content": "Jing Wu, Mehrtash Harandi" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "34. Liu, G., Ma, X., Yang, Y., Wang, C., Liu, J.: Federaser: Enabling efficient client-level data removal from federated learning models. In: 2021 IEEE/ACM 29th International Symposium on Quality of Service (IWQOS). pp. 1-10 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 182 + ], + "type": "text", + "content": "35. Liu, Y., Xu, L., Yuan, X., Wang, C., Li, B.: The right to be forgotten in federated learning: An efficient realization with rapid retraining. In: IEEE INFOCOM 2022 - IEEE Conference on Computer Communications. pp. 1749-1758 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "type": "text", + "content": "36. Lopez-Paz, D., Ranzato, M.: Gradient episodic memory for continual learning. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "text", + "content": "37. Mehta, R., Pal, S., Singh, V., Ravi, S.N.: Deep unlearning via randomized conditionally independent hessians. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10412-10421 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 238, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 480, + 281 + ], + "type": "text", + "content": "38. Neel, S., Roth, A., Sharifi-Malvajerdi, S.: Descent-to-delete: Gradient-based methods for machine unlearning. In: Proceedings of the 32nd International Conference on Algorithmic Learning Theory. pp. 931-962. Proceedings of Machine Learning Research, PMLR (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 282, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 480, + 303 + ], + "type": "text", + "content": "39. Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B., Ng, A.Y.: Reading digits in natural images with unsupervised feature learning (2011)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "text", + "content": "40. Nguyen, T.T., Huynh, T.T., Nguyen, P.L., Liew, A.W.C., Yin, H., Nguyen, Q.V.H.: A survey of machine unlearning. arXiv preprint arXiv:2209.02299 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "type": "text", + "content": "41. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "type": "text", + "content": "42. Peste, A., Alistarh, D., Lampert, C.H.: Ssse: Efficiently erasing samples from trained machine learning models. arXiv preprint arXiv:2107.03860 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "text", + "content": "43. Rando, J., Paleka, D., Lindner, D., Heim, L., Tramér, F.: Red-teaming the stable diffusion safety filter. arXiv preprint arXiv:2210.04610 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "type": "text", + "content": "44. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 436, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 436, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 436, + 480, + 468 + ], + "type": "text", + "content": "45. Romero, E., Barrio, I., Belanche, L.: Incremental and decremental learning for linear support vector machines. In: International Conference on Artificial Neural Networks. pp. 209-218. Springer (2007)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 468, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 511 + ], + "type": "text", + "content": "46. Schramowski, P., Brack, M., Deiseroth, B., Kersting, K.: Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22522-22531 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 512, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 555 + ], + "type": "text", + "content": "47. Schramowski, P., Tauchmann, C., Kersting, K.: Can machines help us answering question 16 in datasheets, and in turn reflecting on inappropriate content? In: Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency. pp. 1350-1361 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 555, + 480, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 480, + 587 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 480, + 587 + ], + "type": "text", + "content": "48. Sekhari, A., Acharya, J., Kamath, G., Suresh, A.T.: Remember what you want to forget: Algorithms for machine unlearning. Advances in Neural Information Processing Systems (NeurIPS) 34, 18075-18086 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 588, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 588, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 480, + 632 + ], + "type": "text", + "content": "49. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Gradcam: Visual explanations from deep networks via gradient-based localization. In: 2017 IEEE International Conference on Computer Vision (ICCV). pp. 618-626 (2017)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "text", + "content": "50. Tarun, A.K., Chundawat, V.S., Mandal, M., Kankanhalli, M.: Fast yet effective machine unlearning. IEEE Transactions on Neural Networks and Learning Systems (2023)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 143, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 143, + 91, + 447, + 102 + ], + "type": "text", + "content": "Scissorhands: Scrub Data Influence via Connection Sensitivity in Networks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 468 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 148 + ], + "type": "text", + "content": "51. Tarun, A.K., Chundawat, V.S., Mandal, M., Kankanhalli, M.: Deep regression unlearning. In: International Conference on Machine Learning. pp. 33921-33939. PMLR (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 149, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 480, + 182 + ], + "type": "text", + "content": "52. Thudi, A., Deza, G., Chandrasekaran, V., Papernot, N.: Unrolling sgd: Understanding factors influencing machine unlearning. In: 2022 IEEE 7th European Symposium on Security and Privacy (EuroS&P). pp. 303-319. IEEE (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 480, + 215 + ], + "type": "text", + "content": "53. Voigt, P., Von dem Bussche, A.: The eu general data protection regulation (gdpr). A Practical Guide, 1st Ed., Cham: Springer International Publishing 10(3152676), 10-5555 (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 215, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 215, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 215, + 480, + 237 + ], + "type": "text", + "content": "54. Wang, J., Guo, S., Xie, X., Qi, H.: Federated unlearning via class-discriminative pruning. In: Proceedings of the ACM Web Conference 2022. pp. 622-632 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 480, + 259 + ], + "type": "text", + "content": "55. Warnecke, A., Pirch, L., Wressnegger, C., Rieck, K.: Machine unlearning of features and labels. arXiv preprint arXiv:2108.11577 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "text", + "content": "56. Wu, C., Zhu, S., Mitra, P.: Federated unlearning with knowledge distillation. arXiv preprint arXiv:2201.09441 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 281, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 480, + 303 + ], + "type": "text", + "content": "57. Wu, J., Le, T., Hayat, M., Harandi, M.: Erasediff: Erasing data influence in diffusion models. arXiv preprint arXiv:2401.05779 (2024)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 303, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 303, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 303, + 480, + 346 + ], + "type": "text", + "content": "58. Wu, Y., Dobriban, E., Davidson, S.: DeltaGrad: Rapid retraining of machine learning models. In: Proceedings of the 37th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 119, pp. 10355-10366. PMLR (13-18 Jul 2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "text", + "content": "59. Zhang, C., Bengio, S., Hardt, M., Recht, B., Vinyals, O.: Understanding deep learning (still) requires rethinking generalization. In: International Conference on Learning Representations (ICLR) (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "text", + "content": "60. Zhang, E., Wang, K., Xu, X., Wang, Z., Shi, H.: Forget-me-not: Learning to forget in text-to-image diffusion models. arXiv preprint arXiv:2303.17591 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 402, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 480, + 435 + ], + "type": "text", + "content": "61. Zhang, Y., Zhang, Y., Yao, Y., Jia, J., Liu, J., Liu, X., Liu, S.: Unlearncanvas: A stylized image dataset to benchmark machine unlearning for diffusion models. arXiv preprint arXiv:2402.11846 (2024)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "type": "text", + "content": "62. Zhang, Y., Chen, X., Jia, J., Zhang, Y., Fan, C., Liu, J., Hong, M., Ding, K., Liu, S.: Defensive unlearning with adversarial training for robust concept erasure in diffusion models. arXiv preprint arXiv:2405.15234 (2024)" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 281, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 281, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 281, + 102 + ], + "type": "text", + "content": "Jing Wu, Mehrtash Harandi" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Score Distillation Sampling with Learned Manifold Corrective/71b1df1d-3510-4a6f-9a35-a2f80ebbbb53_content_list.json b/2024/Score Distillation Sampling with Learned Manifold Corrective/71b1df1d-3510-4a6f-9a35-a2f80ebbbb53_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4dc4d7232fa4beb4eda7ea027bffea2e5e83efb6 --- /dev/null +++ b/2024/Score Distillation Sampling with Learned Manifold Corrective/71b1df1d-3510-4a6f-9a35-a2f80ebbbb53_content_list.json @@ -0,0 +1,1656 @@ +[ + { + "type": "text", + "text": "Score Distillation Sampling with Learned Manifold Corrective", + "text_level": 1, + "bbox": [ + 315, + 140, + 686, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Thiemo Alldieck, Nikos Kolotouros, and Cristian Sminchisescu", + "bbox": [ + 259, + 212, + 743, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Google Research\\*", + "bbox": [ + 441, + 239, + 560, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Score Distillation Sampling (SDS) is a recent but already widely popular method that relies on an image diffusion model to control optimization problems using text prompts. aIn this paper, we conduct an in-depth analysis of the SDS loss function, identify an inherent problem with its formulation, and propose a surprisingly easy but effective fix. Specifically, we decompose the loss into different factors and isolate the component responsible for noisy gradients. In the original formulation, high text guidance is used to account for the noise, leading to unwanted side effects such as oversaturation or repeated detail. Instead, we train a shallow network mimicking the timestep-dependent frequency bias of the image diffusion model in order to effectively factor it out. We demonstrate the versatility and the effectiveness of our novel loss formulation through qualitative and quantitative experiments, including optimization-based image synthesis and editing, zero-shot image translation network training, and text-to-3D synthesis.", + "bbox": [ + 261, + 277, + 743, + 484 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 505, + 375, + 520 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image diffusion models [10] have recently become the de facto standard for image generation. Especially text-to-image models [27,30,31,34] have emerged as powerful tools for high fidelity, diverse, image synthesis. By being controlled through natural language, these models are extremely easy to use and thus open up a wide range of creative applications, without the need for special training. Beyond image synthesis, image diffusion models have been successfully deployed in applications like image restoration, inpating, editing, super-resolution, or colorization [4,7,14,25], among others. Image diffusion models are typically obtained using very large training sets and thus inherently represent the distribution of natural images. While applications typically make use of the generation process of diffusion models, e.g. by inserting images into the process [24,25] or by altering the denoising function [4,5,14], relatively little research has been conducted on how diffusion models can be used as rich, general purpose image priors. Score Distillation Sampling (SDS) proposed in DreamFusion [29] is one of the few exceptions. The SDS loss formulation uses a pretrained text-to-image model [34] to measure how well a given text matches an observation. In DreamFusion this loss is being used to generate 3D assets from textual descriptions, an idea that was quickly adopted [17,20,23,26,32,37] and essentially established a new research direction: text-to-3D synthesis. While being mostly used in this context, the design", + "bbox": [ + 212, + 527, + 787, + 818 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* now at Google DeepMind", + "bbox": [ + 218, + 825, + 405, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/92fd2c0d7eb99c8b112827956c346c8f9d50ae16387f90e9f92d2ffd741bb748.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 143, + 480, + 212 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/89828352ee95179d799d5a86f7af7c80ca6479f57dce5d81a46b1c5d5d1eb690.jpg", + "image_caption": [ + "Fig. 1: Left: Visualization of SDS and LMC-SDS gradients w.r.t. pixel values and estimated denoised images for the given image $\\mathbf{z}$ , the prompt $\\mathbf{y} = \\text{\"autumn\"}$ , and $t = 0.5$ . We visualize the negative gradient, i.e. the direction of change. Right: Power spectra of denoised images $\\hat{\\mathbf{z}}_t$ for varying time-step $t$ compared to the power spectrum of natural images $\\mathbf{z}$ . See Sec. 3 for details." + ], + "image_footnote": [], + "bbox": [ + 217, + 212, + 480, + 297 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/318684887935c2babf608634a96a64b2ebee988815453627f685e267600e3a20.jpg", + "image_caption": [ + "Frequency" + ], + "image_footnote": [], + "bbox": [ + 493, + 148, + 787, + 282 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "of the SDS loss is by no means only tailored to text-to-3D applications. In fact, SDS is an image loss and can be used in much wider contexts [8,12]. However, in its original formulation the SDS loss may degrade the observation, be too eager to match the text prompt, or provide meaningless gradients which inject noise into the objective. In this paper, we conduct an extensive analysis of the SDS loss, identify an inherent problem with its formulation, propose a surprisingly easy but effective fix, and demonstrate its effectiveness in several applications, including optimization-based image synthesis and editing, zero-shot training of image translation networks, and text-to-3D synthesis. Concretely, our new loss formulation - Score Distillation Sampling with Learned Manifold Corrective, or LMC-SDS for short - aims to provide better gradients along the direction of the learned manifold of real images. We present evidence that gradients towards the learned manifold are extremely noisy in SDS and that high text guidance is needed to compensate for the noisy signal. Further, we show that high guidance - or lack thereof - is a possible explanation for the artifacts observed with SDS, cf. Sec. 3. In contrast, applications relying on our novel formulation benefit from meaningful manifold gradients, may use lower text guidance, and produce results of overall higher visual fidelity.", + "bbox": [ + 212, + 383, + 787, + 656 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are: (1) conducting an in-depth analysis of the SDS loss, (2) identifying an inherent problem with the formulation, and (3) proposing a novel LMC-SDS loss, providing cleaner gradients and eliminating the need for high guidance. (4) In extensive experiments, we demonstrate the effectiveness and versatility of the new LMC-SDS loss.", + "bbox": [ + 212, + 657, + 787, + 733 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 214, + 744, + 387, + 758 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Methods that use image diffusion models as priors for tasks other than image synthesis can be broadly categorized into two lines of work.", + "bbox": [ + 212, + 763, + 785, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The first class of methods leverage image diffusion networks to solve specific tasks by relying on the diffusion sampling process. We discuss here a number of representative methods and refer the reader to [6, 19] for detailed discussions.", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Alldieck et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "DDRM [14] tackles linear inverse problems such as deblurring, super-resolution, colorization, or inpainting by defining a modified variational posterior distribution conditioned on the observation signal, which can then be used for sampling. DPS [4] focuses on general inverse problems and proposes to augment the score prediction at each denoising step with an additional term that models the likelihood of the observations. MCG [5] builds on top and improves the fidelity of generation by introducing additional manifold constraints. SDEdit [25] uses diffusion models for image editing by first adding noise to the input image and then denoising it using the text-to-image diffusion prior. RePaint [24] uses unconditional image diffusion models as a prior for solving image inpainting. All these methods use pre-trained diffusion models. In contrast, methods like Palette [33] or Imagic [15] require training a diffusion model from scratch or fine-tuning a pretrained one, respectively. This line of work exclusively deals with image editing problems and is typically restricted to the resolution of the diffusion models.", + "bbox": [ + 212, + 145, + 787, + 357 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "More related to our method is the second category: This body of work uses pre-trained diffusion models within general purpose loss functions in iterative optimization settings. Score Jacobian Chaining [37] propagates the score of a text-to-image diffusion model through a differentiable renderer to supervise the generation of 3D models. DreamFusion [29] takes an orthogonal approach and proposes the Score Distillation Sampling loss. SDS is a modification of the diffusion training objective that does not require expensive backpropagation through the diffusion model and encourages the alignment of the generated image with a conditioning signal. DDS [8] proposes an improved version of SDS, specifically designed for image editing and aims to reduce the artifacts introduced by noisy SDS gradients. In similar spirit and in concurrent work, NFSD [13] aims to reduce noise in the gradients by adding negative conditioning. In contrast, our method neither requires a source prompt nor a negative prompt. SparseFusion [40] introduces a multi-step SDS loss that performs multiple denoising steps instead of one, which adds significant computational overhead. ProlificDreamer [38] proposes a generalization of SDS from point estimates to distributions, but unlike our approach requires computationally expensive fine-tuning of a diffusion model during optimization. Collaborative Score Distillation [16] generalizes SDS to multiple samples and is inspired by Stein variational gradient descent [22]. Category Score Distillation Sampling (C-SDS) focuses on 3D generation and replaces the noise estimation error in SDS with a difference between the noise predictions of a standard and a multi-view consistent diffusion model [41]. Unlike our method, C-SDS is only applicable to 3D generation tasks. In contrast, we propose a general purpose loss function with very little computational overhead.", + "bbox": [ + 212, + 358, + 787, + 720 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 Analysis", + "text_level": 1, + "bbox": [ + 215, + 738, + 334, + 756 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "An image diffusion model is a generative model that is trained to reverse the diffusion process that transforms a natural image into pure noise [10]. An image is formed by iteratively removing small amounts of Gaussian noise over a fixed variance schedule $\\bar{\\alpha}_t$ until an image is created. Text-to-image models are additionally conditioned on text to steer the denoising process in a direction matching", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "LMC-SDS", + "bbox": [ + 658, + 113, + 730, + 127 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the textural description. For a given image $\\mathbf{z}$ , its textual description $\\pmb{y}$ , a randomly sampled timestep $t\\sim \\mathcal{U}(0,1)$ , and random Gaussian noise $\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ , a denoising model $\\epsilon_{\\phi}$ parameterized by $\\phi$ can be trained by minimizing the diffusion loss", + "bbox": [ + 212, + 146, + 782, + 204 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {d i f f}} = w (t) \\left\\| \\epsilon_ {\\phi} ^ {\\omega} (\\mathbf {z} _ {t}, \\mathbf {y}, t) - \\epsilon \\right\\| _ {2} ^ {2}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 390, + 205, + 785, + 224 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{z}_t = \\sqrt{\\bar{\\alpha}_t}\\mathbf{z} + \\sqrt{1 - \\bar{\\alpha}_t}\\epsilon$ refers to the noisy version of $\\mathbf{z}$ and $w(t)$ is a weighting function, omitted in the sequel. DreamFusion [29] showed that given a pre-trained model $\\epsilon_{\\phi}$ the diffusion loss can be utilized for optimization problems. For an arbitrary differentiable rendering function returning $\\mathbf{z}$ and parameterized by $\\pmb{\\theta}$ , the gradient of the denoising function w.r.t. $\\pmb{\\theta}$ is given by", + "bbox": [ + 214, + 231, + 782, + 306 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\boldsymbol {\\theta}} \\mathcal {L} _ {\\text {d i f f}} = \\left(\\boldsymbol {\\epsilon} _ {\\phi} ^ {\\omega} (\\mathbf {z} _ {t}, \\boldsymbol {y}, t) - \\boldsymbol {\\epsilon}\\right) \\frac {\\partial \\boldsymbol {\\epsilon} _ {\\phi} ^ {\\omega} (\\mathbf {z} , \\boldsymbol {y} , t)}{\\partial \\mathbf {z} _ {t}} \\frac {\\partial \\mathbf {z} _ {t}}{\\partial \\boldsymbol {\\theta}}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 319, + 785, + 351 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In practice, the Jacobian term is omitted to avoid backpropagating through the denoising model and the gradient is approximated as", + "bbox": [ + 214, + 361, + 782, + 391 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\boldsymbol {\\theta}} \\mathcal {L} _ {\\mathrm {S D S}} = \\left(\\boldsymbol {\\epsilon} _ {\\phi} ^ {\\omega} (\\mathbf {z} _ {t}, \\boldsymbol {y}, t) - \\boldsymbol {\\epsilon}\\right) \\frac {\\partial \\mathbf {z} _ {t}}{\\partial \\boldsymbol {\\theta}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 385, + 402, + 785, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "resulting in the gradient of original SDS loss. We will now rewrite Eq. (3) and conduct an analysis of the derived components.", + "bbox": [ + 212, + 441, + 782, + 470 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Using classifier-free guidance [11], the predicted noise is the sum of the $\\pmb{y}$ -conditioned and the unconditioned noise prediction weighted by the guidance weight $\\omega$", + "bbox": [ + 212, + 472, + 782, + 516 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\epsilon} _ {\\phi} ^ {\\omega} \\left(\\mathbf {z} _ {t}, \\boldsymbol {y}, t\\right) = \\omega \\boldsymbol {\\epsilon} _ {\\phi} \\left(\\mathbf {z} _ {t}, \\boldsymbol {y}, t\\right) + (1 - \\omega) \\boldsymbol {\\epsilon} _ {\\phi} \\left(\\mathbf {z} _ {t}, t\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 343, + 518, + 784, + 534 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which we can rewrite as", + "bbox": [ + 214, + 540, + 390, + 553 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\epsilon} _ {\\phi} ^ {\\omega} \\left(\\mathbf {z} _ {t}, \\boldsymbol {y}, t\\right) = \\omega \\left(\\boldsymbol {\\epsilon} _ {\\phi} \\left(\\mathbf {z} _ {t}, \\boldsymbol {y}, t\\right) - \\boldsymbol {\\epsilon} _ {\\phi} \\left(\\mathbf {z} _ {t}, t\\right)\\right) + \\boldsymbol {\\epsilon} _ {\\phi} \\left(\\mathbf {z} _ {t}, t\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 321, + 568, + 784, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By inserting Eq. (5) into Eq. (3) we can express $\\nabla_{\\pmb{\\theta}}\\mathcal{L}_{\\mathrm{SDS}}$ as", + "bbox": [ + 214, + 595, + 647, + 611 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\boldsymbol {\\theta}} \\mathcal {L} _ {\\mathrm {S D S}} = \\nabla_ {\\boldsymbol {\\theta}} \\mathcal {L} _ {\\text {c o n d}} + \\nabla_ {\\boldsymbol {\\theta}} \\mathcal {L} _ {\\text {p r o j}} \\tag {6a}\n$$\n", + "text_format": "latex", + "bbox": [ + 392, + 625, + 784, + 638 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\boldsymbol {\\theta}} \\mathcal {L} _ {\\text {c o n d}} = \\omega \\left(\\epsilon_ {\\phi} \\left(\\mathbf {z} _ {t}, \\boldsymbol {y}, t\\right) - \\epsilon_ {\\phi} \\left(\\mathbf {z} _ {t}, t\\right)\\right) \\frac {\\partial \\mathbf {z} _ {t}}{\\partial \\boldsymbol {\\theta}} \\tag {6b}\n$$\n", + "text_format": "latex", + "bbox": [ + 354, + 651, + 784, + 680 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\boldsymbol {\\theta}} \\mathcal {L} _ {\\text {p r o j}} = \\left(\\epsilon_ {\\phi} \\left(\\mathbf {z} _ {t}, t\\right) - \\boldsymbol {\\epsilon}\\right) \\frac {\\partial \\mathbf {z} _ {t}}{\\partial \\boldsymbol {\\theta}}. \\tag {6c}\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 686, + 784, + 714 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The two loss components can be interpreted as follows: $\\mathcal{L}_{\\mathrm{cond}}$ maximizes the agreement of the image with the text prompt by providing gradients towards images formed through conditioning on $\\pmb{y}$ . $\\mathcal{L}_{\\mathrm{proj}}$ performs a single denoising step and provides gradients informing how the image $\\mathbf{z}$ (and thereby parameters $\\pmb{\\theta}$ ) should change such that the image can be better denoised. As the denoising function was trained to minimize the expected squared error between the denoised and the original data, $\\nabla_{\\pmb{\\theta}}\\mathcal{L}_{\\mathrm{proj}}$ can be understood as the gradient direction towards the manifold of natural images.", + "bbox": [ + 212, + 719, + 782, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Aldieck et al.", + "bbox": [ + 271, + 114, + 364, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "DreamFusion [29] originally proposed to set $\\omega = 100$ . This means that, following our derivation, $\\mathcal{L}_{\\mathrm{cond}}$ is weighted proportionally much higher than $\\mathcal{L}_{\\mathrm{proj}}$ . This can be problematic as large guidance weights have been identified to be an important cause of over-exposure in diffusion models [21]. In fact, DreamFusion produces detailed results but also has a tendency towards unrealistically saturated colors, see also Fig. 8. On the other hand, setting $\\omega$ to lower values has been reported to result in very blurry images [8], a behavior that we also demonstrate in Sec. 5. Both behaviors, (1) the tendency to produce overly saturated results when using high guidance as well as (2) blurry results for low guidance, can be explained by looking at the components of $\\mathcal{L}_{\\mathrm{SDS}}$ in isolation. To this end, we consider the case where the rendering function is a no-op and simply returns pixel values, i.e. $\\mathbf{z} = \\boldsymbol{\\theta}$ . In Fig. 1 (left), we visualize $\\nabla_{\\boldsymbol{\\theta}}\\mathcal{L}_{\\mathrm{cond}}$ and $\\nabla_{\\boldsymbol{\\theta}}\\mathcal{L}_{\\mathrm{proj}}$ for this case. $\\nabla_{\\boldsymbol{\\theta}}\\mathcal{L}_{\\mathrm{cond}}$ returns the direction towards an image with altered colors. However, as $\\nabla_{\\boldsymbol{\\theta}}\\mathcal{L}_{\\mathrm{cond}}$ is not anchored in the manifold of natural images, the gradient may eventually point away from that manifold. $\\nabla_{\\boldsymbol{\\theta}}\\mathcal{L}_{\\mathrm{cond}}$ will always point towards changes that \"better\" respect the prompt and eventually produce saturated colors or other artifacts (cf. Sec. 5). $\\nabla_{\\boldsymbol{\\theta}}\\mathcal{L}_{\\mathrm{proj}}$ , on the other hand, incorrectly marks high-frequency detail originating from $\\mathbf{z}$ for removal - the reason for blurriness. This behavior can be understood when rewriting $\\nabla_{\\boldsymbol{\\theta}}\\mathcal{L}_{\\mathrm{proj}}$ as", + "bbox": [ + 212, + 145, + 787, + 435 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\pmb {\\theta}} \\mathcal {L} _ {\\mathrm {p r o j}} = \\left(\\frac {\\sqrt {\\alpha_ {t}}}{\\sqrt {1 - \\alpha_ {t}}} \\left(\\mathbf {z} - \\mathbf {z} _ {\\phi} (\\mathbf {z} _ {t}, t)\\right)\\right) \\frac {\\partial \\mathbf {z} _ {t}}{\\partial \\pmb {\\theta}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 349, + 446, + 785, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which is equivalent to comparing the predicted and true noise. By looking at Eq. (7), we can now see a possible explanation why $\\mathcal{L}_{\\mathrm{proj}}$ might not be optimal: $\\hat{\\mathbf{z}}_t = \\mathbf{z}_\\phi (\\mathbf{z}_t,t)$ is an approximation of $\\mathbf{z}$ and most probably not a perfect replica of $\\mathbf{z}$ , even when $\\mathbf{z}$ already lies on the natural image manifold, see Fig. 1 (left). This is especially the case for large values of $t$ , where the denoising model has to reconstruct the true image from almost pure noise. More formally, we can observe a frequency bias in $\\mathbf{z}_\\phi$ dependent on $t$ when comparing images $\\mathbf{z}$ with their denoised counterparts $\\hat{\\mathbf{z}}_t$ , see Fig. 1 (right).", + "bbox": [ + 212, + 491, + 787, + 611 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We conclude the following: (1) The reason for oversaturated colors and other artifacts is due to $\\mathcal{L}_{\\mathrm{cond}}$ being \"over-eager\" to maximize the agreement of the image with the conditioning variable. (2) $\\mathcal{L}_{\\mathrm{proj}}$ provides deficient gradients ultimately removing high-frequency detail instead of \"correcting\" the gradient towards on-manifold solutions due to a frequency bias in $\\mathbf{z}_{\\phi}$ . Thus, $\\mathcal{L}_{\\mathrm{proj}}$ appears to be a non-optimal choice and we provide an alternative in the sequel. But first we use our decomposition to discuss other approaches to \"fix\" the SDS loss.", + "bbox": [ + 212, + 612, + 787, + 717 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "DDS [8] proposes to use the difference between two SDS losses for image editing, one computed on the current state w.r.t. the target prompt and one computed on the initial image w.r.t. a textual description of that image. By computing an SDS difference, DDS entirely removes $\\mathcal{L}_{\\mathrm{proj}}$ and introduces a negative gradient pointing away from the initial state. However, the need for an initial image is limiting possible use-cases. In concurrent work, NFSD [13] aims to obtain better results by replacing $\\mathcal{L}_{\\mathrm{proj}}$ with a $\\mathcal{L}_{\\mathrm{cond}}$ variant using negative prompts for most timesteps, and a loss based on $\\epsilon$ for small $t$ . While the root", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "LMC-SDS", + "bbox": [ + 658, + 113, + 730, + 127 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "for blurry results is removed by eliminating $\\mathcal{L}_{\\mathrm{proj}}$ , both DDS and NFSD do not anchor the optimization along the learned manifold. Other methods aim for a better noise prediction, where by improving $\\epsilon_{\\phi}(\\mathbf{z},t)$ , one may also improve $\\mathcal{L}_{\\mathrm{proj}}$ ( $\\mathcal{L}_{\\mathrm{proj}}$ is the difference between the predicted and true noise). E.g. SparseFusion [40] introduces multi-step SDS. Hereby, instead of denoising in one step, up to 50 steps are used. While this may produce more informative gradients in $\\mathcal{L}_{\\mathrm{proj}}$ , this is also significantly more expensive. Through different means but towards the same goal, the concurrent work ProlificDreamer [38] learns a low-rank adaptation of the text-to-image model. While this approach is effective, it also adds the overhead of fine-tuning an additional diffusion model. Finally, some recent methods aim to alleviate artifacts by augmenting the original SDS formulation with advanced schedules for sampling $t$ or with negative prompts [1,35]. In contrast, we aim for a generally applicable solution.", + "bbox": [ + 212, + 146, + 787, + 344 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 Method", + "text_level": 1, + "bbox": [ + 215, + 354, + 330, + 369 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We now present our solution to eliminating the frequency bias in $\\mathcal{L}_{\\mathrm{proj}}$ that we identified earlier. In an attempt to better understand the behavior of the denoising model given $\\mathbf{z}$ and $t$ , we model it as a two step process, as follows", + "bbox": [ + 212, + 375, + 785, + 421 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} _ {\\phi} \\left(\\mathbf {z} _ {t}, t\\right) = b \\circ p \\left(\\mathbf {z} _ {t}, t\\right) = b \\left(\\tilde {\\mathbf {z}}, t\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 392, + 431, + 784, + 446 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $p$ projects $\\mathbf{z}$ onto the learned natural image manifold, $\\tilde{\\mathbf{z}}$ is the projected image, and $\\tilde{\\mathbf{z}} = \\mathbf{z}$ for images on the manifold. $b$ is the error or frequency bias of the denoising model introduced by the current time step $t$ . In this two-step model we are only interested in $p$ and would like to neglect any gradients coming from $b$ . As it is unfeasible to correct for $b$ , we propose to cancel out effects of $b$ by approximating $b$ with $\\hat{b}$ and applying $\\hat{b}$ on $\\mathbf{z}$ before comparing $\\mathbf{z}$ with $\\hat{\\mathbf{z}}_t$ :", + "bbox": [ + 212, + 457, + 787, + 549 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\boldsymbol {\\theta}} \\mathcal {L} _ {\\mathrm {l m c}} = \\left(\\hat {b} (\\mathbf {z}, t) - \\hat {\\mathbf {z}} _ {t}\\right) \\frac {\\partial \\mathbf {z} _ {t}}{\\partial \\boldsymbol {\\theta}} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 556, + 785, + 583 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this formulation we omit both the Jacobian term from the denoising model (like in the original formulation, cf. Eq. (3)) as well as the Jacobian from $\\hat{b}$ . Also, we empirically found that by dropping the weighting term $\\frac{\\sqrt{\\alpha_t}}{\\sqrt{1 - \\alpha_t}}$ we can further improve visual fidelity – which is similar to the common practice of predicting the original image instead of the noise when learning the denoising model. As demonstrated in Fig. 1 (left), given an adequate model $\\hat{b}$ , $\\nabla_{\\theta} \\mathcal{L}_{\\mathrm{Imc}}$ focuses on global changes, whereas $\\nabla_{\\theta} \\mathcal{L}_{\\mathrm{proj}}$ is dominated by high-frequent detail. We now explain how to obtain $\\hat{b}$ .", + "bbox": [ + 212, + 588, + 787, + 714 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We propose to model $\\hat{b}$ as a learnable function $\\hat{b}_{\\psi}(\\mathbf{z},t)$ parameterized by $\\psi$ . We can observe $b$ as a function of $t$ by sampling triplets $(\\mathbf{z},\\hat{\\mathbf{z}}_t,t)$ , where $\\mathbf{z}$ are random natural images assumed to be approximately on manifold images, i.e. $\\mathbf{z} \\approx \\tilde{\\mathbf{z}}$ . Using these triplets we can learn $\\hat{b}_{\\psi}$ by minimizing $\\| \\hat{\\mathbf{z}}_t - \\hat{b}_{\\mathbf{z},t}\\| _2^2$ , with $\\hat{b}_{\\mathbf{z},t} = \\hat{b}_{\\psi}(\\mathbf{z},t)$ . Modeled with a deterministic model, $\\hat{b}_{\\psi}$ is expected to underfit the highly complex and probabilistic nature of $b$ . In fact, for a given pair $(\\mathbf{z},t)$ at best it can learn the mean over all possible $\\hat{\\mathbf{z}}_t$ . This is desired, because even after approximating $b(\\mathbf{z},t)$ through $\\hat{b}_{\\psi}(\\mathbf{z},t)$ , we will be able to observe a", + "bbox": [ + 212, + 715, + 787, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Aldieck et al.", + "bbox": [ + 271, + 114, + 364, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "gradient in $\\mathcal{L}_{\\mathrm{lmc}}$ : a gradient towards a specific instance of $\\hat{\\mathbf{z}}_t$ . Further, $\\hat{b}_{\\psi}$ only learns to \"blur\" images and neither generates new content nor is informed by the diffusion models' image manifold. Thus, it will perform similar operations for on-manifold and off-manifold images and $\\mathcal{L}_{\\mathrm{lmc}}$ may correct non-manifold images. However, naively applying a $\\hat{b}_{\\psi}$ in Eq. (9), learned that way, is causing an undesired effect, namely a drift in image statistics. Concretely we observed that images produced by $\\hat{b}_{\\psi}$ tend to underestimate the true image dynamics, especially for large $t$ where $\\hat{\\mathbf{z}}_t$ may be vastly different to $\\mathbf{z}$ . By underestimating image dynamics, however, $\\nabla_{\\theta} \\mathcal{L}_{\\mathrm{lmc}}$ contains a gradient to \"correct\" for that and we have found images obtained via optimization using this gradient to be overly saturated and contrasting. To this end, we make $\\hat{b}_{\\psi}$ agnostic to the global image statistics by computing the training loss in a normalized space", + "bbox": [ + 212, + 145, + 787, + 329 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {k}} = \\left\\| \\hat {\\mathbf {z}} _ {t} - \\left(\\frac {\\sigma \\left(\\hat {\\mathbf {z}} _ {t}\\right)}{\\sigma \\left(\\hat {b} _ {\\mathbf {z} , t}\\right)} \\left(\\hat {b} _ {\\mathbf {z}, t} - \\mu \\left(\\hat {b} _ {\\mathbf {z}, t}\\right)\\right) + \\mu \\left(\\hat {b} _ {\\mathbf {z}, t}\\right)\\right) \\right\\| _ {2} ^ {2}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 342, + 785, + 380 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "with standard deviation $\\sigma$ and mean $\\mu$ . Predicting in a normalized space naturally requires to perform the same (normalisation) operation in Eq. (9), which we assume to be absorbed into $\\hat{b}$ as follows.", + "bbox": [ + 212, + 393, + 785, + 439 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Increasing Sample Diversity. The original SDS formulation has been reported to have mode-seeking behaviour. Consequently, optimization results obtained via $\\mathcal{L}_{\\mathrm{SDS}}$ have the tendency to look very similar, even when based on different random seeds. We hypothesize this is due to optimizing over many steps and hence averaging the effects of varying noise $\\epsilon$ . To prevent such averaging behaviour, we propose to fix $\\epsilon$ over the course of optimization - which is in spirit similar to DDIM sampling [36]. Empirically, we found that optimizing with fixed $\\epsilon$ is often unstable in the original SDS formulation. We hypothesize this is rooted in $\\mathcal{L}_{\\mathrm{proj}}$ not providing proper gradients towards the image manifold and at the same time $\\mathcal{L}_{\\mathrm{cond}}$ providing too strong gradients towards prompt agreement. The optimization state may leave the manifold of natural images and drift towards a state associated more to an \"adversarial example\" for the given prompt. While our loss formulation allows the use of a fixed $\\epsilon$ , probably due to both the usage of smaller $\\omega$ and better manifold correctives, we observe that we can further robustify the formulation by only fixing $\\epsilon$ in $\\mathcal{L}_{\\mathrm{cond}}$ , while continue to sample $\\epsilon$ in $\\mathcal{L}_{\\mathrm{lmc}}$ . This results in high quality, yet diverse, results, cf. Sec. 5.", + "bbox": [ + 212, + 441, + 787, + 684 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation Details. We use an image diffusion model producing images of $128 \\times 128$ pixel resolution and trained on internal data sources [34] as the pre-trained text-to-image diffusion model $\\epsilon_{\\phi}^{\\omega}$ or $\\mathbf{z}_{\\phi}^{\\omega}$ , respectively. We model $\\hat{b}_{\\psi}$ as a standard U-Net with four encoder layers with two Conv/ReLU/MaxPool blocks per layer, and skip connections to the decoder. The first layer contains 32 filters and the number of filters is doubled in each layer in the encoder and halved in the decoder. $\\hat{b}_{\\psi}$ has been trained until convergence, using $\\mathcal{L}_{\\mathrm{k}}$ as the only loss. For creating training triplets $(\\mathbf{z},\\hat{\\mathbf{z}}_t,t)$ , we sampled random images $\\mathbf{z}$ from OpenImages [18], random time-steps $t \\sim \\mathcal{U}(0.02,0.98)$ , and passed both into $\\mathbf{z}_{\\phi}$ obtaining $\\hat{\\mathbf{z}}_t$ . We train $\\hat{b}_{\\psi}$ once and keep it fixed for all experiments.", + "bbox": [ + 212, + 686, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "LMC-SDS", + "bbox": [ + 658, + 113, + 730, + 127 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f7dbffefb133f9df63588e53364cf03df4fc4cb598085ab761b4f1e89e6c5dc5.jpg", + "image_caption": [ + "Fig. 2: Optimization-based image synthesis results. We optimize an empty image to match a given prompt using our LMC-SDS, the original SDS loss, and $\\mathcal{L}_{\\mathrm{cond}}$ . SDS struggles to create detailed content when using low guidance $\\omega$ . High $\\omega$ produces detailed results but colors may be oversaturated (chimpanzee face), fake detail may appear (2nd mouse tail), or artifacts emerge. $\\mathcal{L}_{\\mathrm{cond}}$ is unstable to optimize and produces unrealistic colors. In contrast, our method produces detailed results with balanced colors." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 255 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 359, + 375, + 378 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We evaluate our proposed LMC-SDS loss in a number of qualitative and quantitative experiments, including image synthesis, image editing, image-to-image translation network training, and 3D asset generation.", + "bbox": [ + 212, + 383, + 785, + 430 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Set-up. If not noted otherwise, we scale the total loss by $\\frac{1}{\\omega}$ to be able to ablate $\\omega$ without changing the gradient magnitude. Fixed $\\epsilon$ sampling is only used when explicitly stated. The pre-trained diffusion model we use in our experiments produces images of $128\\times 128\\mathrm{px}$ resolution. We enable arbitrary higher resolutions in our image experiments through the following strategy: We downsample the optimization state to $128\\times 128\\mathrm{px}$ for the first iterations. This ensures that the whole image is affected by the loss, but the signal may be blurry. Accounting for this, we sample random patches of arbitrary resolutions, from the image, for the remaining optimization steps. To prevent the loss from producing fake detail in sampled image patches, we lower the learning rate during this phase through Cosine decay.", + "bbox": [ + 212, + 430, + 787, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Baselines. This paper presents a generic novel loss function which uses an image diffusion model. To this end, we compare with the original SDS formulation and with related work focused on improving SDS. Specifically, we compare against DDS [8], NFSD [13], and multi-step SDS (MS-SDS) [40]. DDS, NFSD, and MS-SDS originally were proposed using a latent diffusion model [31] and operate in latent space. For fair comparisons, we adapted all methods to operate in pixel space and to use the same image diffusion model. Note that optimizing in the (ambient) pixel space is generally considered harder due to higher-dimensionality. Finally, we evaluate the effectiveness of our LMC-SDS loss for 3D asset generation using DreamFusion. Hereby we focus on the improvements of our loss and deliberately exclude recent improvements over DreamFusion rooted in other factors, e.g. novel-view synthesis networks or additional perceptual losses. While being important findings, those factors are orthogonal and complementary to our approach.", + "bbox": [ + 212, + 598, + 787, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Metrics. We use the following metrics for evaluating our methods and baselines: We report Learned Perceptual Image Patch Similarity (LPIPS $\\downarrow$ ) [39] w.r.t.", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 116, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Alldieck et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1d6092aeb8609f06ab3caf1cd0e7f721c37d0181de7d816454a7d2f4df65167a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 143, + 408, + 218 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/d73dc3ff9c116bb001e1265c3faf4f3d5b9589d64196a0af7530469bc9af4424.jpg", + "image_caption": [ + "\"a tiny dragon\"" + ], + "image_footnote": [], + "bbox": [ + 411, + 143, + 596, + 218 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/2ded896864c0a31568640bd99067ac9f219526c503718d207b762403cac11a8f.jpg", + "image_caption": [ + "\"ahoneybee\"" + ], + "image_footnote": [], + "bbox": [ + 596, + 143, + 784, + 218 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/e6dc849adc76137711c827c0dca50a77f6e0c79ec8e2df46c4ef22fdcb51deaf.jpg", + "image_caption": [ + "\"oil paiting of a landscape\"", + "\"bananas in a basket\"" + ], + "image_footnote": [], + "bbox": [ + 217, + 228, + 408, + 301 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/27aaf75c6f56b59494840075b1a977a2614ccb209aba8ff75cd9e858bbd9215f.jpg", + "image_caption": [ + "\"akangaroo\"" + ], + "image_footnote": [], + "bbox": [ + 408, + 228, + 594, + 301 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/565bae8b01313f8243b2efe6281e6cf9bee7ee460f0b2e5ab8bf3bb04d3716f8.jpg", + "image_caption": [ + "\"amarble bust of Caesar\"" + ], + "image_footnote": [], + "bbox": [ + 594, + 229, + 784, + 301 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/c16e2a6890e18ac399300e3cd87ff18736a0b2490a9a871df13b132974434c06.jpg", + "image_caption": [ + "Fig. 3: Examples of optimization-based image editing results. We show pairs of input images (left) and editing result (right).", + "Input", + "Fig. 4: By fixing $\\epsilon$ in $\\mathcal{L}_{\\mathrm{cond}}$ we can obtain diverse editing results. We show four variants of optimization-based editing results of the input image under the given prompt \"a mountain during sunset\"." + ], + "image_footnote": [], + "bbox": [ + 217, + 348, + 330, + 436 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/51ca10230060baa3e20691f7bddca0c52a8ad2028cb0be05056157e887e9459a.jpg", + "image_caption": [ + "seed = 1" + ], + "image_footnote": [], + "bbox": [ + 331, + 348, + 555, + 436 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/f33e7d7377930a3685268d68b695aeb2c26820d595ed9b3eef2d161e21122c84.jpg", + "image_caption": [ + "seed $= 3$" + ], + "image_footnote": [], + "bbox": [ + 555, + 348, + 671, + 436 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/383c261f0463ed4501e3ce7e2c83562b12642e98ced2904a7d81e92403a67d0b.jpg", + "image_caption": [ + "seed $= 4$" + ], + "image_footnote": [], + "bbox": [ + 671, + 348, + 784, + 436 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "the source image, thus capturing how much an image was altered. Note that high LPIPS can be a result of either successful editing or image corruption. To disambiguate these, we additionally report the CLIP score $(\\uparrow)$ [9] w.r.t. the target prompt. As CLIP can be insensitive, we additionally report CLIP-R-Precision $(\\uparrow)$ [28], the retrieval accuracy using the CLIP model, where feasible.", + "bbox": [ + 212, + 508, + 784, + 585 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.1 Image Synthesis & Editing", + "text_level": 1, + "bbox": [ + 215, + 608, + 483, + 625 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We first explore our LMC-SDS loss for optimization-based image synthesis and image editing.", + "bbox": [ + 212, + 627, + 782, + 657 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Synthesis. To be able to evaluate and ablate the original SDS loss and our LMC-SDS in isolation (without additional interfering losses), we propose the task of optimization-based image synthesis. Instead of using the denoising schedule of a diffusion model, we initialize an image with $50\\%$ grey and optimize its pixel values to match a given prompt, following the patching strategy described earlier. While it is not expected that images obtained using this strategy are on par with those obtained via the generation process, the results allow to draw meaningful conclusions about the properties of the loss functions. We show results of our loss and SDS in Fig. 2. We additionally include results obtained by only using $\\mathcal{L}_{\\mathrm{cond}}$ , ablating the influence of our proposed $\\mathcal{L}_{\\mathrm{lmc}}$ . Optimizing solely with $\\mathcal{L}_{\\mathrm{cond}}$ is unstable and the final images feature clearly visible defects, such as oversaturated colors or noisy patterns. The prompt is also not always well re", + "bbox": [ + 212, + 657, + 785, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "LMC-SDS", + "bbox": [ + 661, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/bc0d8b288b67ae2f6b435606a31c9244b89f05562a29df3b2848945e444c1bb0.jpg", + "image_caption": [ + "Fig. 5: Quantitative results for optimization-based editing under varying $\\omega$ . Left: Our method results in the highest CLIP score over all baselines for all $\\omega$ . Right: We plot LPIPS over CLIP for further performance insights: DDS stays close to the original image (lowest LPIPS) by performing only small edits (low CLIP). SDS and MS-SDS respect the prompt better (higher CLIP), but corrupt the image (high LPIPS). NFSD corrupts the image less (lower LPIPS), but exhibits weak editing capabilities (low CLIP). Our method shows the strongest editing capabilities (highest CLIP), while staying close to the original structure." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 454, + 236 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/305999ff26a8456ced7cbf81e8ee6e5f8b7cf3e2b347c5e7bbdc631b0629c942.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 465, + 142, + 781, + 237 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/00911d2ae640ca2bf9d3a78c40ee89b5b2f08ac2d884acda9be80dc1c898df60.jpg", + "image_caption": [ + "Fig. 6: Comparison of image statistics before (Original) and after editing with $\\omega = 5$ . We plot the average power spectra (left) and vertical derivative histograms (right). Ours, DDS, and NFSD preserve the image statistics well, while SDS-based methods introduce significant blur." + ], + "image_footnote": [], + "bbox": [ + 285, + 356, + 496, + 459 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/af68a175d2317401efacf7ec6fce64dcaf4699a0c98d9443d0b39e0282850923.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 354, + 715, + 460 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "spected as e.g. in the case of the gibbon. The original SDS formulation produces blurry and desaturated images when using low guidance $\\omega$ . High $\\omega$ leads to better performance but artifacts are present, e.g. fake detail like the second mouse tail, oversaturated colors like in the chimpanzee's face, or blocky artifacts. In contrast, our method produces images of higher visual fidelity which respect the prompt well. Additionally, we empirically found that problems of the individual losses become more pronounced when optimizing longer (here we only optimize for 500 steps, cf. Sec. 5.2 for effects after 50K steps).", + "bbox": [ + 212, + 534, + 787, + 654 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Editing. Our loss can also be used for optimization-based image editing. We load existing natural images and optimize their pixel values directly. Optionally, we may constrain the optimization with an L2 loss w.r.t. the original pixel values. In Fig. 3 (more in Supp. Mat.) we show qualitative results of so edited images along with the corresponding target prompt. All results faithfully respect the target prompt while keeping the image structure intact. We can also vary the result by fixing $\\epsilon$ over the course of the optimization, cf. Sec. 4. In Fig. 4, we show an input image along with four result variants. All results respect the target prompt well, while at the same time being significantly different from each other.", + "bbox": [ + 212, + 657, + 787, + 792 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "For a numerical evaluation, we use an internal dataset of 1,000 images with text editing instructions in form of source and target prompts, similar to the dataset used in InstructPix2Pix [2]. We optimize images to match the target", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Aldieck et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/cc7936a0c758293b65df5dadd3d2a8ad515440d03153fac15ced5467624406d6.jpg", + "image_caption": [ + "Fig. 7: Results obtained with cats-to-others image translation networks trained with different losses. The network trained with SDS and NFSD fail to create detailed results. The DDS network produces rather uniform results, whereas the network trained with our LMC-SDS creates realistic and detailed images." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 787, + 325 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/74dabe669053e10d97d71989757fa2bf781a92ddafc4ae2a0991c419a994ad11.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
LPIPS ↓CLIP Score ↑CLIP-R-Precision ↑
DDS0.1430.2530.333
NFSD0.3320.2550.434
SDS0.4010.2560.355
SDS, ω = 1000.2010.2590.482
Ours0.1780.2610.592
", + "bbox": [ + 305, + 395, + 694, + 481 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 1: Numerical results of the cats-to-others network training experiment. Note that LPIPS is computed w.r.t. the source image, meaning a low score may mean the image has not been edited enough. This is the case for DDS, as evident from low R-Precision.", + "bbox": [ + 212, + 484, + 787, + 539 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "prompt solely using our LMC-SDS and baselines DDS, NFSD, SDS, and MS-SDS. For MS-SDS we chose to use five step denoising. We also experimented with more steps, but didn't observe significantly different performance. We report numerical results with varying $\\omega$ (without loss rescaling) in Fig. 5. We found DDS to be least effective and producing results closest to the original image. We hypothesize that the negative gradient w.r.t. to the original image may not always be helpful. This could also explain a behavior reported by original authors, where editing results on real images are \"inferior\" over those on synthetic images. A possible reason being, that since for synthetic images the original prompts are describing the image well, the negative gradients are more helpful in that case. SDS can effectively edit the image, but also corrupts the image. This is also evident from the altered image statistics, see Fig. 6. NFSD corrupts the image less, but exhibits weak editing capabilities. In our experiments, MS-SDS performs very similar to SDS, corrupting the image in similar manner, but also being slightly less effective for editing. MS-SDS produces clear and detailed denoised images as part of the loss computation. We hypothesize that since denoised images and the current optimization state may differ drastically, and as denoised images vary considerably at each step, the computed gradients are not always very meaningful. The original paper proposes MS-SDS", + "bbox": [ + 212, + 551, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "LMC-SDS", + "bbox": [ + 661, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "to be used with a view-conditioned diffusion model, where the denoised images probably vary much less and computed gradients may be more directed. Our loss respects the prompt best (highest CLIP), and finds a good balance between preserving the original structure and performing significant edits.", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To quantify the how much the result diversity can be increased through fixing $\\epsilon$ , we re-ran the experiment using LMC-SDS ( $\\omega = 30$ ) ten times with different seeds, both using the standard strategy and with fixing $\\epsilon$ in $\\mathcal{L}_{\\mathrm{cond}}$ . We then computed the LPIPS (lower is more similar) of nine runs w.r.t. to the 10th (reference) run. The results using the standard strategy are all very similar with a per-run average of $0.041 \\pm 0.007$ w.r.t. the reference run. In contrast, our new strategy produces much more diverse results: $0.158 \\pm 0.018$ .", + "bbox": [ + 212, + 207, + 787, + 311 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.2 Image-to-image Translation Network Training", + "text_level": 1, + "bbox": [ + 214, + 330, + 640, + 345 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Optimization-based editing is not always an optimal choice, e.g. when building interactive editing tools where execution speed matters. Further, optimization can be unstable or perform better or worse, depending on the input images or target prompts. To this end, we explore to what extend LMC-SDS can be used to train a reliable and fast image-to-image translation network. Given a known distribution of images, we train a network to translate images from that distribution to an unseen target distribution, only described by a target caption. Inspired by the authors of DDS [8], we train a cats-to-others network, with \"others\" being bears, dogs, lions, and squirrels. The network transforms an image of a cat into an image of one of the other species. In DDS, the authors fine-tune a pre-trained diffusion model into a single-step feedforward model using images created by that same diffusion model. In contrast, we train a standard U-Net from scratch and use a dataset of real images, a much more practical and versatile approach. Concretely, we train using cat images from AFHQ [3] and evaluate on $5\\%$ held out test images. We use the target prompt \"a photo of a [class]\" with $\\omega = 15$ for all methods and the source prompt \"a photo of a cat\" for DDS. Moreover, we stabilize the training using an L2 loss on the original image and use Cosine scheduling to let the network first focus on reconstructing the original image and then on edits, similarly to the experiment in DDS. The class is passed via a frequency encoded class id concatenated to each pixel. We report numerical results in Tab. 1 and show qualitative examples in Fig. 7. Additional analysis is provided in the Suppl. Mat. SDS performs similarly to previous experiments for this task. While the target class is evident in the resulting images, the results feature strong deficiencies. Again DDS struggles with significant edits and stays close to the source. In contrast, the network trained with our LMC-SDS loss produces realistic and convincing edits. This is also evident from the significantly higher CLIP-R-Precision among all network variants.", + "bbox": [ + 212, + 349, + 787, + 758 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.3 Text-to-3D Asset Generation", + "text_level": 1, + "bbox": [ + 214, + 775, + 503, + 789 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In our final experiment, we demonstrate how LMC-SDS can improve text-to-3D models in the style of DreamFusion. In Sec. 3, we discussed how large guidance weights $\\omega$ may affect the quality of results. To provide further evidence for this", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Alldieck et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/519d9e2b6fbed0e4548a1383d1ae7c738c897efadf3dd49418c64d405d424df8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 212, + 143, + 782, + 226 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/489a39cdbd337c00259befbf8d224feb8653f9132242bd102f2cfada4ab6b047.jpg", + "image_caption": [ + "Fig. 8: The effect of increasing guidance weight $\\omega$ in DreamFusion using the original SDS loss (even rows) and our proposed LMC-SDS (odd rows). Using larger $\\omega$ saturates colors (chair, motorcycle, ship) and worsens the Janus problem (third wing and claw for the eagle). For low $\\omega$ , vanilla DreamFusion fails to produce detailed geometry (ship, eagle) or geometry at all (motorcycle), while our approach performs well for $\\omega = 20$ , often even for $\\omega = 10$ . Notice how our approach constantly produces more detailed results for all $\\omega$ ." + ], + "image_footnote": [], + "bbox": [ + 212, + 229, + 784, + 323 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "observation, we produce 3D assets with varying $\\omega$ using the original DreamFusion formulation and a variant using our LMC-SDS. The results in Fig. 8 also show the trend of high $\\omega$ having the tendency to produce oversaturated colors, albeit not as strongly as in the original paper, possibly due the use of a different image diffusion model. But we also observed an additional issue: the infamous Janus problem, the tendency to repeat detail in each view, becomes more severe with bigger $\\omega$ , as evident from the eagle example. On the other hand, when using low $\\omega$ in the original SDS loss, we observe that DreamFusion struggles to create detailed structure or even structure at all. In contrast, when using LMC-SDS, DreamFusion successfully produces detailed results, even when using up to only one tenth of the originally proposed $\\omega$ . In Fig. 9 (more in Suppl. Mat.), we show additional side-by-side comparisons of the original SDS-based DreamFusion and our LMC-SDS version. Using our loss, the results are consistently sharper and feature far more detail. Finally in Fig. 10, we show that fixed $\\epsilon$ sampling may also be used in 3D asset generation, to obtain diverse results.", + "bbox": [ + 212, + 429, + 787, + 656 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "6 Discussion & Conclusion", + "text_level": 1, + "bbox": [ + 215, + 676, + 490, + 691 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We proposed a novel loss formulation called LMC-SDS. LMC-SDS allows to effectively utilize text-to-image diffusion models as rich image priors in several optimization and network training tasks. We derived LMC-SDS by analyzing DreamFusion's SDS loss and by identifying an inherent problem in its formulation. Concretely, we expressed the SDS gradient as the sum of two terms. We identified that the second term $\\mathcal{L}_{\\mathrm{proj}}$ provides deficient gradients and consequently was weighted down in DreamFusion. We argued that this term, however, should provide gradients towards the learned manifold of natural images associated to the diffusion model. We also provided evidence that the lack thereof", + "bbox": [ + 212, + 703, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "LMC-SDS", + "bbox": [ + 661, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/c1d0b6406303682d3d553725f0151343131c5d6569271a7909e8064cee2e251c.jpg", + "image_caption": [ + "Fig. 9: Results of DreamFusion using the original SDS loss (right) and our proposed LMC-SDS (left) for $\\omega = 20$ . Our results are much sharper, contain more detail, and feature more realistic colors in all examples." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 321 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/da2d69e7c59d505d9c97f752e6ff124b46b4a62ff470de2e627258c1af0cd2eb.jpg", + "image_caption": [ + "Fig. 10: DreamFusion results using our LMC-SDS loss and fixed $\\epsilon$ sampling for the prompts \"a mannequin wearing a [hoodie, hat]. Note the diversity in colors, poses, and mannequin styles." + ], + "image_footnote": [], + "bbox": [ + 217, + 369, + 779, + 421 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "makes the optimization unstable, thus negatively impacting the fidelity of results. To this end, we proposed to model the denoising models' time step dependent frequency bias. This allowed us to factor it out and obtain the cleaner gradient $\\nabla_{\\theta}\\mathcal{L}_{\\mathrm{lmc}}$ . In extensive experiments, we have demonstrated the effectiveness and versatility of our novel LMC-SDS loss. Furthermore, we discussed how LMC-SDS differs from other SDS-style losses and demonstrated how its performance is superior both qualitatively and quantitatively.", + "bbox": [ + 212, + 484, + 787, + 590 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations. LMC-SDS finds its limits where either the signal from the diffusion model is too weak and ambiguous or when the optimization goes off the natural image manifold by too much. Concretely, LMC-SDS can only utilize whatever signal is coming from the diffusion model. This means, that when the diffusion model does not \"understand\" the prompt, LMC-SDS (and all other SDS-based approaches) will not be able to provide meaningful gradients. Further, whenever the current optimization state is too corrupted, LMC-SDS will not be able to correct for that and guide the optimization back to the manifold. Also, LMC-SDS sometimes overcompensates the frequency bias. Nevertheless, LMC-SDS is an important first step towards more stable and meaningful usage of the inherent image prior associated to image diffusion models.", + "bbox": [ + 212, + 592, + 787, + 758 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We proposed LMC-SDS, a novel SDS-based loss formulation with cleaner gradients towards the image manifold. In the future, we would like to continue to investigate how the manifold corrective can be further improved. We also want to leverage our findings in concrete applications like text-to-3D or in image editing and stylization tasks.", + "bbox": [ + 212, + 758, + 787, + 834 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Aldieck et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 323, + 160 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Armandpour, M., Zheng, H., Sadeghian, A., Sadeghian, A., Zhou, M.: Re-imagine the negative prompt algorithm: Transform 2d diffusion into 3d, alleviate janus problem and beyond. arXiv preprint arXiv:2304.04968 (2023)", + "2. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)", + "3. Choi, Y., Uh, Y., Yoo, J., Ha, J.W.: Stargan v2: Diverse image synthesis for multiple domains. In: IEEE Conf. Comput. Vis. Pattern Recog. (2020)", + "4. Chung, H., Kim, J., Mccann, M.T., Klasky, M.L., Ye, J.C.: Diffusion posterior sampling for general noisy inverse problems. In: Int. Conf. Learn. Represent. (2022)", + "5. Chung, H., Sim, B., Ryu, D., Ye, J.C.: Improving diffusion models for inverse problems using manifold constraints. Adv. Neural Inform. Process. Syst. 35, 25683-25696 (2022)", + "6. Croitoru, F.A., Hondru, V., Ionescu, R.T., Shah, M.: Diffusion models in vision: A survey. IEEE Trans. Pattern Anal. Mach. Intell. (2023)", + "7. Gao, S., Liu, X., Zeng, B., Xu, S., Li, Y., Luo, X., Liu, J., Zhen, X., Zhang, B.: Implicit diffusion models for continuous super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10021-10030 (2023)", + "8. Hertz, A., Aberman, K., Cohen-Or, D.: Delta denoising score. In: Int. Conf. Comput. Vis. (2023)", + "9. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718 (2021)", + "0. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Adv. Neural Inform. Process. Syst. 33, 6840-6851 (2020)", + "1. Ho, J., Salimans, T.: Classifier-free diffusion guidance. In: NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications (2021)", + "2. Jain, A., Xie, A., Abbeel, P.: Vectorfusion: Text-to-svq by abstracting pixel-based diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1911-1920 (2023)", + "3. Katzir, O., Patashnik, O., Cohen-Or, D., Lischinski, D.: Noise-free score distillation. arXiv preprint arXiv:2310.17590 (2023)", + "4. Kawar, B., Elad, M., Ermon, S., Song, J.: Denoising diffusion restoration models. In: Adv. Neural Inform. Process. Syst. (2022)", + "5. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 6007-6017 (2023)", + "6. Kim, S., Lee, K., Choi, J.S., Jeong, J., Sohn, K., Shin, J.: Collaborative score distillation for consistent visual synthesis. arXiv preprint arXiv:2307.04787 (2023)", + "7. Kolotouros, N., Alldieck, T., Zanfir, A., Bazavan, E.G., Fieraru, M., Sminchisescu, C.: Dreamhuman: Animatable 3d avatars from text. In: Adv. Neural Inform. Process. Syst. (2023)", + "8. Kuznetsova, A., Rom, H., Alldrin, N., Uijlings, J., Krasin, I., Pont-Tuset, J., Kamali, S., Popov, S., Malloci, M., Kolesnikov, A., Duerig, T., Ferrari, V.: The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. Int. J. Comput. Vis. (2020)", + "9. Li, X., Ren, Y., Jin, X., Lan, C., Wang, X., Zeng, W., Wang, X., Chen, Z.: Diffusion models for image restoration and enhancement-a comprehensive survey. arXiv preprint arXiv:2308.09388 (2023)" + ], + "bbox": [ + 225, + 176, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "LMC-SDS", + "bbox": [ + 658, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "20. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3d: High-resolution text-to-3d content creation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 300–309 (2023)", + "21. Lin, S., Liu, B., Li, J., Yang, X.: Common diffusion noise schedules and sample steps are flawed. arXiv preprint arXiv:2305.08891 (2023)", + "22. Liu, Q., Wang, D.: Stein variational gradient descent: A general purpose bayesian inference algorithm. Advances in neural information processing systems 29 (2016)", + "23. Liu, R., Wu, R., Van Hoorick, B., Tokmakov, P., Zakharov, S., Vondrick, C.: Zero-1-to-3: Zero-shot one image to 3d object. In: Int. Conf. Comput. Vis. pp. 9298-9309 (2023)", + "24. Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Van Gool, L.: Repaint: Inpainting using denoising diffusion probabilistic models. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 11461-11471 (2022)", + "25. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: SDEdit: Guided image synthesis and editing with stochastic differential equations. In: Int. Conf. Learn. Represent. (2022)", + "26. Metzer, G., Richardson, E., Patashnik, O., Giryes, R., Cohen-Or, D.: Latent-nerf for shape-guided generation of 3d shapes and textures. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 12663-12673 (2023)", + "27. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In: Int. Conf. on Mach. Learn. (2022)", + "28. Park, D.H., Azadi, S., Liu, X., Darrell, T., Rohrbach, A.: Benchmark for compositional text-to-image synthesis. In: Conf. on Neural Inf. Proc. Systems Datasets and Benchmarks Track (Round 1) (2021)", + "29. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. In: Int. Conf. Learn. Represent. (2022)", + "30. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022)", + "31. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 10684-10695 (2022)", + "32. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 22500-22510 (2023)", + "33. Sahara, C., Chan, W., Chang, H., Lee, C., Ho, J., Salimans, T., Fleet, D., Norouzi, M.: Palette: Image-to-image diffusion models. In: ACM SIGGRAPH 2022 Conference Proceedings. pp. 1-10 (2022)", + "34. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Adv. Neural Inform. Process. Syst. 35, 36479-36494 (2022)", + "35. Shi, Y., Wang, P., Ye, J., Mai, L., Li, K., Yang, X.: Mvdream: Multi-view diffusion for 3d generation. arXiv:2308.16512 (2023)", + "36. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. In: Int. Conf. Learn. Represent. (2020)", + "37. Wang, H., Du, X., Li, J., Yeh, R.A., Shakhnarovich, G.: Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 12619-12629 (2023)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Aldieck et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "38. Wang, Z., Lu, C., Wang, Y., Bao, F., Li, C., Su, H., Zhu, J.: Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213 (2023)", + "39. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: IEEE Conf. Comput. Vis. Pattern Recog. (2018)", + "40. Zhou, Z., Tulsiani, S.: Sparsefusion: Distilling view-conditioned diffusion for 3d reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12588-12597 (2023)", + "41. Zou, Z.X., Cheng, W., Cao, Y.P., Huang, S.S., Shan, Y., Zhang, S.H.: Sparse3d: Distilling multiview-consistent diffusion for object reconstruction from sparse views. arXiv preprint arXiv:2308.14078 (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 314 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "LMC-SDS", + "bbox": [ + 660, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Score Distillation Sampling with Learned Manifold Corrective/71b1df1d-3510-4a6f-9a35-a2f80ebbbb53_model.json b/2024/Score Distillation Sampling with Learned Manifold Corrective/71b1df1d-3510-4a6f-9a35-a2f80ebbbb53_model.json new file mode 100644 index 0000000000000000000000000000000000000000..fd509a7ddf140a04da2520a3fd340cf14d65b2e8 --- /dev/null +++ b/2024/Score Distillation Sampling with Learned Manifold Corrective/71b1df1d-3510-4a6f-9a35-a2f80ebbbb53_model.json @@ -0,0 +1,2214 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.316, + 0.141, + 0.687, + 0.185 + ], + "angle": 0, + "content": "Score Distillation Sampling with Learned Manifold Corrective" + }, + { + "type": "text", + "bbox": [ + 0.26, + 0.213, + 0.744, + 0.228 + ], + "angle": 0, + "content": "Thiemo Alldieck, Nikos Kolotouros, and Cristian Sminchisescu" + }, + { + "type": "text", + "bbox": [ + 0.442, + 0.24, + 0.561, + 0.255 + ], + "angle": 0, + "content": "Google Research\\*" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.278, + 0.744, + 0.486 + ], + "angle": 0, + "content": "Abstract. Score Distillation Sampling (SDS) is a recent but already widely popular method that relies on an image diffusion model to control optimization problems using text prompts. aIn this paper, we conduct an in-depth analysis of the SDS loss function, identify an inherent problem with its formulation, and propose a surprisingly easy but effective fix. Specifically, we decompose the loss into different factors and isolate the component responsible for noisy gradients. In the original formulation, high text guidance is used to account for the noise, leading to unwanted side effects such as oversaturation or repeated detail. Instead, we train a shallow network mimicking the timestep-dependent frequency bias of the image diffusion model in order to effectively factor it out. We demonstrate the versatility and the effectiveness of our novel loss formulation through qualitative and quantitative experiments, including optimization-based image synthesis and editing, zero-shot image translation network training, and text-to-3D synthesis." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.506, + 0.377, + 0.521 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.529, + 0.789, + 0.819 + ], + "angle": 0, + "content": "Image diffusion models [10] have recently become the de facto standard for image generation. Especially text-to-image models [27,30,31,34] have emerged as powerful tools for high fidelity, diverse, image synthesis. By being controlled through natural language, these models are extremely easy to use and thus open up a wide range of creative applications, without the need for special training. Beyond image synthesis, image diffusion models have been successfully deployed in applications like image restoration, inpating, editing, super-resolution, or colorization [4,7,14,25], among others. Image diffusion models are typically obtained using very large training sets and thus inherently represent the distribution of natural images. While applications typically make use of the generation process of diffusion models, e.g. by inserting images into the process [24,25] or by altering the denoising function [4,5,14], relatively little research has been conducted on how diffusion models can be used as rich, general purpose image priors. Score Distillation Sampling (SDS) proposed in DreamFusion [29] is one of the few exceptions. The SDS loss formulation uses a pretrained text-to-image model [34] to measure how well a given text matches an observation. In DreamFusion this loss is being used to generate 3D assets from textual descriptions, an idea that was quickly adopted [17,20,23,26,32,37] and essentially established a new research direction: text-to-3D synthesis. While being mostly used in this context, the design" + }, + { + "type": "page_footnote", + "bbox": [ + 0.22, + 0.826, + 0.406, + 0.84 + ], + "angle": 0, + "content": "* now at Google DeepMind" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.127 + ], + "angle": 0, + "content": "Alldieck et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.481, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.213, + 0.481, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.495, + 0.149, + 0.788, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.645, + 0.287, + 0.701, + 0.298 + ], + "angle": 0, + "content": "Frequency" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.302, + 0.788, + 0.372 + ], + "angle": 0, + "content": "Fig. 1: Left: Visualization of SDS and LMC-SDS gradients w.r.t. pixel values and estimated denoised images for the given image \\(\\mathbf{z}\\), the prompt \\(\\mathbf{y} = \\text{\"autumn\"}\\), and \\(t = 0.5\\). We visualize the negative gradient, i.e. the direction of change. Right: Power spectra of denoised images \\(\\hat{\\mathbf{z}}_t\\) for varying time-step \\(t\\) compared to the power spectrum of natural images \\(\\mathbf{z}\\). See Sec. 3 for details." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.385, + 0.789, + 0.657 + ], + "angle": 0, + "content": "of the SDS loss is by no means only tailored to text-to-3D applications. In fact, SDS is an image loss and can be used in much wider contexts [8,12]. However, in its original formulation the SDS loss may degrade the observation, be too eager to match the text prompt, or provide meaningless gradients which inject noise into the objective. In this paper, we conduct an extensive analysis of the SDS loss, identify an inherent problem with its formulation, propose a surprisingly easy but effective fix, and demonstrate its effectiveness in several applications, including optimization-based image synthesis and editing, zero-shot training of image translation networks, and text-to-3D synthesis. Concretely, our new loss formulation - Score Distillation Sampling with Learned Manifold Corrective, or LMC-SDS for short - aims to provide better gradients along the direction of the learned manifold of real images. We present evidence that gradients towards the learned manifold are extremely noisy in SDS and that high text guidance is needed to compensate for the noisy signal. Further, we show that high guidance - or lack thereof - is a possible explanation for the artifacts observed with SDS, cf. Sec. 3. In contrast, applications relying on our novel formulation benefit from meaningful manifold gradients, may use lower text guidance, and produce results of overall higher visual fidelity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.658, + 0.788, + 0.734 + ], + "angle": 0, + "content": "In summary, our contributions are: (1) conducting an in-depth analysis of the SDS loss, (2) identifying an inherent problem with the formulation, and (3) proposing a novel LMC-SDS loss, providing cleaner gradients and eliminating the need for high guidance. (4) In extensive experiments, we demonstrate the effectiveness and versatility of the new LMC-SDS loss." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.745, + 0.388, + 0.76 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.794 + ], + "angle": 0, + "content": "Methods that use image diffusion models as priors for tasks other than image synthesis can be broadly categorized into two lines of work." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "The first class of methods leverage image diffusion networks to solve specific tasks by relying on the diffusion sampling process. We discuss here a number of representative methods and refer the reader to [6, 19] for detailed discussions." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.114, + 0.732, + 0.128 + ], + "angle": 0, + "content": "LMC-SDS" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.358 + ], + "angle": 0, + "content": "DDRM [14] tackles linear inverse problems such as deblurring, super-resolution, colorization, or inpainting by defining a modified variational posterior distribution conditioned on the observation signal, which can then be used for sampling. DPS [4] focuses on general inverse problems and proposes to augment the score prediction at each denoising step with an additional term that models the likelihood of the observations. MCG [5] builds on top and improves the fidelity of generation by introducing additional manifold constraints. SDEdit [25] uses diffusion models for image editing by first adding noise to the input image and then denoising it using the text-to-image diffusion prior. RePaint [24] uses unconditional image diffusion models as a prior for solving image inpainting. All these methods use pre-trained diffusion models. In contrast, methods like Palette [33] or Imagic [15] require training a diffusion model from scratch or fine-tuning a pretrained one, respectively. This line of work exclusively deals with image editing problems and is typically restricted to the resolution of the diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.359, + 0.789, + 0.722 + ], + "angle": 0, + "content": "More related to our method is the second category: This body of work uses pre-trained diffusion models within general purpose loss functions in iterative optimization settings. Score Jacobian Chaining [37] propagates the score of a text-to-image diffusion model through a differentiable renderer to supervise the generation of 3D models. DreamFusion [29] takes an orthogonal approach and proposes the Score Distillation Sampling loss. SDS is a modification of the diffusion training objective that does not require expensive backpropagation through the diffusion model and encourages the alignment of the generated image with a conditioning signal. DDS [8] proposes an improved version of SDS, specifically designed for image editing and aims to reduce the artifacts introduced by noisy SDS gradients. In similar spirit and in concurrent work, NFSD [13] aims to reduce noise in the gradients by adding negative conditioning. In contrast, our method neither requires a source prompt nor a negative prompt. SparseFusion [40] introduces a multi-step SDS loss that performs multiple denoising steps instead of one, which adds significant computational overhead. ProlificDreamer [38] proposes a generalization of SDS from point estimates to distributions, but unlike our approach requires computationally expensive fine-tuning of a diffusion model during optimization. Collaborative Score Distillation [16] generalizes SDS to multiple samples and is inspired by Stein variational gradient descent [22]. Category Score Distillation Sampling (C-SDS) focuses on 3D generation and replaces the noise estimation error in SDS with a difference between the noise predictions of a standard and a multi-view consistent diffusion model [41]. Unlike our method, C-SDS is only applicable to 3D generation tasks. In contrast, we propose a general purpose loss function with very little computational overhead." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.739, + 0.335, + 0.757 + ], + "angle": 0, + "content": "3 Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.842 + ], + "angle": 0, + "content": "An image diffusion model is a generative model that is trained to reverse the diffusion process that transforms a natural image into pure noise [10]. An image is formed by iteratively removing small amounts of Gaussian noise over a fixed variance schedule \\(\\bar{\\alpha}_t\\) until an image is created. Text-to-image models are additionally conditioned on text to steer the denoising process in a direction matching" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.128 + ], + "angle": 0, + "content": "Aldieck et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.205 + ], + "angle": 0, + "content": "the textural description. For a given image \\(\\mathbf{z}\\), its textual description \\(\\pmb{y}\\), a randomly sampled timestep \\(t\\sim \\mathcal{U}(0,1)\\), and random Gaussian noise \\(\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\), a denoising model \\(\\epsilon_{\\phi}\\) parameterized by \\(\\phi\\) can be trained by minimizing the diffusion loss" + }, + { + "type": "equation", + "bbox": [ + 0.391, + 0.206, + 0.786, + 0.226 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {d i f f}} = w (t) \\left\\| \\epsilon_ {\\phi} ^ {\\omega} (\\mathbf {z} _ {t}, \\mathbf {y}, t) - \\epsilon \\right\\| _ {2} ^ {2}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.232, + 0.784, + 0.308 + ], + "angle": 0, + "content": "where \\(\\mathbf{z}_t = \\sqrt{\\bar{\\alpha}_t}\\mathbf{z} + \\sqrt{1 - \\bar{\\alpha}_t}\\epsilon\\) refers to the noisy version of \\(\\mathbf{z}\\) and \\(w(t)\\) is a weighting function, omitted in the sequel. DreamFusion [29] showed that given a pre-trained model \\(\\epsilon_{\\phi}\\) the diffusion loss can be utilized for optimization problems. For an arbitrary differentiable rendering function returning \\(\\mathbf{z}\\) and parameterized by \\(\\pmb{\\theta}\\), the gradient of the denoising function w.r.t. \\(\\pmb{\\theta}\\) is given by" + }, + { + "type": "equation", + "bbox": [ + 0.347, + 0.32, + 0.786, + 0.352 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\boldsymbol {\\theta}} \\mathcal {L} _ {\\text {d i f f}} = \\left(\\boldsymbol {\\epsilon} _ {\\phi} ^ {\\omega} (\\mathbf {z} _ {t}, \\boldsymbol {y}, t) - \\boldsymbol {\\epsilon}\\right) \\frac {\\partial \\boldsymbol {\\epsilon} _ {\\phi} ^ {\\omega} (\\mathbf {z} , \\boldsymbol {y} , t)}{\\partial \\mathbf {z} _ {t}} \\frac {\\partial \\mathbf {z} _ {t}}{\\partial \\boldsymbol {\\theta}}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.362, + 0.784, + 0.392 + ], + "angle": 0, + "content": "In practice, the Jacobian term is omitted to avoid backpropagating through the denoising model and the gradient is approximated as" + }, + { + "type": "equation", + "bbox": [ + 0.386, + 0.404, + 0.786, + 0.433 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\boldsymbol {\\theta}} \\mathcal {L} _ {\\mathrm {S D S}} = \\left(\\boldsymbol {\\epsilon} _ {\\phi} ^ {\\omega} (\\mathbf {z} _ {t}, \\boldsymbol {y}, t) - \\boldsymbol {\\epsilon}\\right) \\frac {\\partial \\mathbf {z} _ {t}}{\\partial \\boldsymbol {\\theta}}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.442, + 0.784, + 0.472 + ], + "angle": 0, + "content": "resulting in the gradient of original SDS loss. We will now rewrite Eq. (3) and conduct an analysis of the derived components." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.473, + 0.784, + 0.517 + ], + "angle": 0, + "content": "Using classifier-free guidance [11], the predicted noise is the sum of the \\( \\pmb{y} \\)-conditioned and the unconditioned noise prediction weighted by the guidance weight \\( \\omega \\)" + }, + { + "type": "equation", + "bbox": [ + 0.344, + 0.519, + 0.785, + 0.535 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\epsilon} _ {\\phi} ^ {\\omega} \\left(\\mathbf {z} _ {t}, \\boldsymbol {y}, t\\right) = \\omega \\boldsymbol {\\epsilon} _ {\\phi} \\left(\\mathbf {z} _ {t}, \\boldsymbol {y}, t\\right) + (1 - \\omega) \\boldsymbol {\\epsilon} _ {\\phi} \\left(\\mathbf {z} _ {t}, t\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.541, + 0.392, + 0.554 + ], + "angle": 0, + "content": "which we can rewrite as" + }, + { + "type": "equation", + "bbox": [ + 0.323, + 0.569, + 0.785, + 0.586 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\epsilon} _ {\\phi} ^ {\\omega} \\left(\\mathbf {z} _ {t}, \\boldsymbol {y}, t\\right) = \\omega \\left(\\boldsymbol {\\epsilon} _ {\\phi} \\left(\\mathbf {z} _ {t}, \\boldsymbol {y}, t\\right) - \\boldsymbol {\\epsilon} _ {\\phi} \\left(\\mathbf {z} _ {t}, t\\right)\\right) + \\boldsymbol {\\epsilon} _ {\\phi} \\left(\\mathbf {z} _ {t}, t\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.596, + 0.648, + 0.612 + ], + "angle": 0, + "content": "By inserting Eq. (5) into Eq. (3) we can express \\(\\nabla_{\\pmb{\\theta}}\\mathcal{L}_{\\mathrm{SDS}}\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.393, + 0.625, + 0.785, + 0.64 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\boldsymbol {\\theta}} \\mathcal {L} _ {\\mathrm {S D S}} = \\nabla_ {\\boldsymbol {\\theta}} \\mathcal {L} _ {\\text {c o n d}} + \\nabla_ {\\boldsymbol {\\theta}} \\mathcal {L} _ {\\text {p r o j}} \\tag {6a}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.356, + 0.652, + 0.785, + 0.681 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\boldsymbol {\\theta}} \\mathcal {L} _ {\\text {c o n d}} = \\omega \\left(\\epsilon_ {\\phi} \\left(\\mathbf {z} _ {t}, \\boldsymbol {y}, t\\right) - \\epsilon_ {\\phi} \\left(\\mathbf {z} _ {t}, t\\right)\\right) \\frac {\\partial \\mathbf {z} _ {t}}{\\partial \\boldsymbol {\\theta}} \\tag {6b}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.396, + 0.687, + 0.785, + 0.715 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\boldsymbol {\\theta}} \\mathcal {L} _ {\\text {p r o j}} = \\left(\\epsilon_ {\\phi} \\left(\\mathbf {z} _ {t}, t\\right) - \\boldsymbol {\\epsilon}\\right) \\frac {\\partial \\mathbf {z} _ {t}}{\\partial \\boldsymbol {\\theta}}. \\tag {6c}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.784, + 0.84 + ], + "angle": 0, + "content": "The two loss components can be interpreted as follows: \\(\\mathcal{L}_{\\mathrm{cond}}\\) maximizes the agreement of the image with the text prompt by providing gradients towards images formed through conditioning on \\(\\pmb{y}\\). \\(\\mathcal{L}_{\\mathrm{proj}}\\) performs a single denoising step and provides gradients informing how the image \\(\\mathbf{z}\\) (and thereby parameters \\(\\pmb{\\theta}\\)) should change such that the image can be better denoised. As the denoising function was trained to minimize the expected squared error between the denoised and the original data, \\(\\nabla_{\\pmb{\\theta}}\\mathcal{L}_{\\mathrm{proj}}\\) can be understood as the gradient direction towards the manifold of natural images." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.114, + 0.732, + 0.128 + ], + "angle": 0, + "content": "LMC-SDS" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.436 + ], + "angle": 0, + "content": "DreamFusion [29] originally proposed to set \\(\\omega = 100\\). This means that, following our derivation, \\(\\mathcal{L}_{\\mathrm{cond}}\\) is weighted proportionally much higher than \\(\\mathcal{L}_{\\mathrm{proj}}\\). This can be problematic as large guidance weights have been identified to be an important cause of over-exposure in diffusion models [21]. In fact, DreamFusion produces detailed results but also has a tendency towards unrealistically saturated colors, see also Fig. 8. On the other hand, setting \\(\\omega\\) to lower values has been reported to result in very blurry images [8], a behavior that we also demonstrate in Sec. 5. Both behaviors, (1) the tendency to produce overly saturated results when using high guidance as well as (2) blurry results for low guidance, can be explained by looking at the components of \\(\\mathcal{L}_{\\mathrm{SDS}}\\) in isolation. To this end, we consider the case where the rendering function is a no-op and simply returns pixel values, i.e. \\(\\mathbf{z} = \\boldsymbol{\\theta}\\). In Fig. 1 (left), we visualize \\(\\nabla_{\\boldsymbol{\\theta}}\\mathcal{L}_{\\mathrm{cond}}\\) and \\(\\nabla_{\\boldsymbol{\\theta}}\\mathcal{L}_{\\mathrm{proj}}\\) for this case. \\(\\nabla_{\\boldsymbol{\\theta}}\\mathcal{L}_{\\mathrm{cond}}\\) returns the direction towards an image with altered colors. However, as \\(\\nabla_{\\boldsymbol{\\theta}}\\mathcal{L}_{\\mathrm{cond}}\\) is not anchored in the manifold of natural images, the gradient may eventually point away from that manifold. \\(\\nabla_{\\boldsymbol{\\theta}}\\mathcal{L}_{\\mathrm{cond}}\\) will always point towards changes that \"better\" respect the prompt and eventually produce saturated colors or other artifacts (cf. Sec. 5). \\(\\nabla_{\\boldsymbol{\\theta}}\\mathcal{L}_{\\mathrm{proj}}\\), on the other hand, incorrectly marks high-frequency detail originating from \\(\\mathbf{z}\\) for removal - the reason for blurriness. This behavior can be understood when rewriting \\(\\nabla_{\\boldsymbol{\\theta}}\\mathcal{L}_{\\mathrm{proj}}\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.351, + 0.448, + 0.786, + 0.481 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\pmb {\\theta}} \\mathcal {L} _ {\\mathrm {p r o j}} = \\left(\\frac {\\sqrt {\\alpha_ {t}}}{\\sqrt {1 - \\alpha_ {t}}} \\left(\\mathbf {z} - \\mathbf {z} _ {\\phi} (\\mathbf {z} _ {t}, t)\\right)\\right) \\frac {\\partial \\mathbf {z} _ {t}}{\\partial \\pmb {\\theta}}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.492, + 0.788, + 0.612 + ], + "angle": 0, + "content": "which is equivalent to comparing the predicted and true noise. By looking at Eq. (7), we can now see a possible explanation why \\(\\mathcal{L}_{\\mathrm{proj}}\\) might not be optimal: \\(\\hat{\\mathbf{z}}_t = \\mathbf{z}_\\phi (\\mathbf{z}_t,t)\\) is an approximation of \\(\\mathbf{z}\\) and most probably not a perfect replica of \\(\\mathbf{z}\\), even when \\(\\mathbf{z}\\) already lies on the natural image manifold, see Fig. 1 (left). This is especially the case for large values of \\(t\\), where the denoising model has to reconstruct the true image from almost pure noise. More formally, we can observe a frequency bias in \\(\\mathbf{z}_\\phi\\) dependent on \\(t\\) when comparing images \\(\\mathbf{z}\\) with their denoised counterparts \\(\\hat{\\mathbf{z}}_t\\), see Fig. 1 (right)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.788, + 0.718 + ], + "angle": 0, + "content": "We conclude the following: (1) The reason for oversaturated colors and other artifacts is due to \\(\\mathcal{L}_{\\mathrm{cond}}\\) being \"over-eager\" to maximize the agreement of the image with the conditioning variable. (2) \\(\\mathcal{L}_{\\mathrm{proj}}\\) provides deficient gradients ultimately removing high-frequency detail instead of \"correcting\" the gradient towards on-manifold solutions due to a frequency bias in \\(\\mathbf{z}_{\\phi}\\). Thus, \\(\\mathcal{L}_{\\mathrm{proj}}\\) appears to be a non-optimal choice and we provide an alternative in the sequel. But first we use our decomposition to discuss other approaches to \"fix\" the SDS loss." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "DDS [8] proposes to use the difference between two SDS losses for image editing, one computed on the current state w.r.t. the target prompt and one computed on the initial image w.r.t. a textual description of that image. By computing an SDS difference, DDS entirely removes \\(\\mathcal{L}_{\\mathrm{proj}}\\) and introduces a negative gradient pointing away from the initial state. However, the need for an initial image is limiting possible use-cases. In concurrent work, NFSD [13] aims to obtain better results by replacing \\(\\mathcal{L}_{\\mathrm{proj}}\\) with a \\(\\mathcal{L}_{\\mathrm{cond}}\\) variant using negative prompts for most timesteps, and a loss based on \\(\\epsilon\\) for small \\(t\\). While the root" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.128 + ], + "angle": 0, + "content": "Aldieck et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.345 + ], + "angle": 0, + "content": "for blurry results is removed by eliminating \\(\\mathcal{L}_{\\mathrm{proj}}\\), both DDS and NFSD do not anchor the optimization along the learned manifold. Other methods aim for a better noise prediction, where by improving \\(\\epsilon_{\\phi}(\\mathbf{z},t)\\), one may also improve \\(\\mathcal{L}_{\\mathrm{proj}}\\) (\\(\\mathcal{L}_{\\mathrm{proj}}\\) is the difference between the predicted and true noise). E.g. SparseFusion [40] introduces multi-step SDS. Hereby, instead of denoising in one step, up to 50 steps are used. While this may produce more informative gradients in \\(\\mathcal{L}_{\\mathrm{proj}}\\), this is also significantly more expensive. Through different means but towards the same goal, the concurrent work ProlificDreamer [38] learns a low-rank adaptation of the text-to-image model. While this approach is effective, it also adds the overhead of fine-tuning an additional diffusion model. Finally, some recent methods aim to alleviate artifacts by augmenting the original SDS formulation with advanced schedules for sampling \\(t\\) or with negative prompts [1,35]. In contrast, we aim for a generally applicable solution." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.355, + 0.331, + 0.371 + ], + "angle": 0, + "content": "4 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.375, + 0.787, + 0.422 + ], + "angle": 0, + "content": "We now present our solution to eliminating the frequency bias in \\(\\mathcal{L}_{\\mathrm{proj}}\\) that we identified earlier. In an attempt to better understand the behavior of the denoising model given \\(\\mathbf{z}\\) and \\(t\\), we model it as a two step process, as follows" + }, + { + "type": "equation", + "bbox": [ + 0.393, + 0.433, + 0.785, + 0.448 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} _ {\\phi} \\left(\\mathbf {z} _ {t}, t\\right) = b \\circ p \\left(\\mathbf {z} _ {t}, t\\right) = b \\left(\\tilde {\\mathbf {z}}, t\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.458, + 0.788, + 0.55 + ], + "angle": 0, + "content": "where \\( p \\) projects \\( \\mathbf{z} \\) onto the learned natural image manifold, \\( \\tilde{\\mathbf{z}} \\) is the projected image, and \\( \\tilde{\\mathbf{z}} = \\mathbf{z} \\) for images on the manifold. \\( b \\) is the error or frequency bias of the denoising model introduced by the current time step \\( t \\). In this two-step model we are only interested in \\( p \\) and would like to neglect any gradients coming from \\( b \\). As it is unfeasible to correct for \\( b \\), we propose to cancel out effects of \\( b \\) by approximating \\( b \\) with \\( \\hat{b} \\) and applying \\( \\hat{b} \\) on \\( \\mathbf{z} \\) before comparing \\( \\mathbf{z} \\) with \\( \\hat{\\mathbf{z}}_t \\):" + }, + { + "type": "equation", + "bbox": [ + 0.408, + 0.558, + 0.786, + 0.584 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\boldsymbol {\\theta}} \\mathcal {L} _ {\\mathrm {l m c}} = \\left(\\hat {b} (\\mathbf {z}, t) - \\hat {\\mathbf {z}} _ {t}\\right) \\frac {\\partial \\mathbf {z} _ {t}}{\\partial \\boldsymbol {\\theta}} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.589, + 0.788, + 0.715 + ], + "angle": 0, + "content": "In this formulation we omit both the Jacobian term from the denoising model (like in the original formulation, cf. Eq. (3)) as well as the Jacobian from \\(\\hat{b}\\). Also, we empirically found that by dropping the weighting term \\(\\frac{\\sqrt{\\alpha_t}}{\\sqrt{1 - \\alpha_t}}\\) we can further improve visual fidelity – which is similar to the common practice of predicting the original image instead of the noise when learning the denoising model. As demonstrated in Fig. 1 (left), given an adequate model \\(\\hat{b}\\), \\(\\nabla_{\\theta} \\mathcal{L}_{\\mathrm{Imc}}\\) focuses on global changes, whereas \\(\\nabla_{\\theta} \\mathcal{L}_{\\mathrm{proj}}\\) is dominated by high-frequent detail. We now explain how to obtain \\(\\hat{b}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.716, + 0.788, + 0.843 + ], + "angle": 0, + "content": "We propose to model \\(\\hat{b}\\) as a learnable function \\(\\hat{b}_{\\psi}(\\mathbf{z},t)\\) parameterized by \\(\\psi\\). We can observe \\(b\\) as a function of \\(t\\) by sampling triplets \\((\\mathbf{z},\\hat{\\mathbf{z}}_t,t)\\), where \\(\\mathbf{z}\\) are random natural images assumed to be approximately on manifold images, i.e. \\(\\mathbf{z} \\approx \\tilde{\\mathbf{z}}\\). Using these triplets we can learn \\(\\hat{b}_{\\psi}\\) by minimizing \\(\\| \\hat{\\mathbf{z}}_t - \\hat{b}_{\\mathbf{z},t}\\| _2^2\\), with \\(\\hat{b}_{\\mathbf{z},t} = \\hat{b}_{\\psi}(\\mathbf{z},t)\\). Modeled with a deterministic model, \\(\\hat{b}_{\\psi}\\) is expected to underfit the highly complex and probabilistic nature of \\(b\\). In fact, for a given pair \\((\\mathbf{z},t)\\) at best it can learn the mean over all possible \\(\\hat{\\mathbf{z}}_t\\). This is desired, because even after approximating \\(b(\\mathbf{z},t)\\) through \\(\\hat{b}_{\\psi}(\\mathbf{z},t)\\), we will be able to observe a" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.114, + 0.732, + 0.128 + ], + "angle": 0, + "content": "LMC-SDS" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.33 + ], + "angle": 0, + "content": "gradient in \\(\\mathcal{L}_{\\mathrm{lmc}}\\): a gradient towards a specific instance of \\(\\hat{\\mathbf{z}}_t\\). Further, \\(\\hat{b}_{\\psi}\\) only learns to \"blur\" images and neither generates new content nor is informed by the diffusion models' image manifold. Thus, it will perform similar operations for on-manifold and off-manifold images and \\(\\mathcal{L}_{\\mathrm{lmc}}\\) may correct non-manifold images. However, naively applying a \\(\\hat{b}_{\\psi}\\) in Eq. (9), learned that way, is causing an undesired effect, namely a drift in image statistics. Concretely we observed that images produced by \\(\\hat{b}_{\\psi}\\) tend to underestimate the true image dynamics, especially for large \\(t\\) where \\(\\hat{\\mathbf{z}}_t\\) may be vastly different to \\(\\mathbf{z}\\). By underestimating image dynamics, however, \\(\\nabla_{\\theta} \\mathcal{L}_{\\mathrm{lmc}}\\) contains a gradient to \"correct\" for that and we have found images obtained via optimization using this gradient to be overly saturated and contrasting. To this end, we make \\(\\hat{b}_{\\psi}\\) agnostic to the global image statistics by computing the training loss in a normalized space" + }, + { + "type": "equation", + "bbox": [ + 0.335, + 0.343, + 0.786, + 0.381 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {k}} = \\left\\| \\hat {\\mathbf {z}} _ {t} - \\left(\\frac {\\sigma \\left(\\hat {\\mathbf {z}} _ {t}\\right)}{\\sigma \\left(\\hat {b} _ {\\mathbf {z} , t}\\right)} \\left(\\hat {b} _ {\\mathbf {z}, t} - \\mu \\left(\\hat {b} _ {\\mathbf {z}, t}\\right)\\right) + \\mu \\left(\\hat {b} _ {\\mathbf {z}, t}\\right)\\right) \\right\\| _ {2} ^ {2}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.394, + 0.787, + 0.44 + ], + "angle": 0, + "content": "with standard deviation \\(\\sigma\\) and mean \\(\\mu\\). Predicting in a normalized space naturally requires to perform the same (normalisation) operation in Eq. (9), which we assume to be absorbed into \\(\\hat{b}\\) as follows." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.443, + 0.788, + 0.685 + ], + "angle": 0, + "content": "Increasing Sample Diversity. The original SDS formulation has been reported to have mode-seeking behaviour. Consequently, optimization results obtained via \\(\\mathcal{L}_{\\mathrm{SDS}}\\) have the tendency to look very similar, even when based on different random seeds. We hypothesize this is due to optimizing over many steps and hence averaging the effects of varying noise \\(\\epsilon\\). To prevent such averaging behaviour, we propose to fix \\(\\epsilon\\) over the course of optimization - which is in spirit similar to DDIM sampling [36]. Empirically, we found that optimizing with fixed \\(\\epsilon\\) is often unstable in the original SDS formulation. We hypothesize this is rooted in \\(\\mathcal{L}_{\\mathrm{proj}}\\) not providing proper gradients towards the image manifold and at the same time \\(\\mathcal{L}_{\\mathrm{cond}}\\) providing too strong gradients towards prompt agreement. The optimization state may leave the manifold of natural images and drift towards a state associated more to an \"adversarial example\" for the given prompt. While our loss formulation allows the use of a fixed \\(\\epsilon\\), probably due to both the usage of smaller \\(\\omega\\) and better manifold correctives, we observe that we can further robustify the formulation by only fixing \\(\\epsilon\\) in \\(\\mathcal{L}_{\\mathrm{cond}}\\), while continue to sample \\(\\epsilon\\) in \\(\\mathcal{L}_{\\mathrm{lmc}}\\). This results in high quality, yet diverse, results, cf. Sec. 5." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.687, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Implementation Details. We use an image diffusion model producing images of \\(128 \\times 128\\) pixel resolution and trained on internal data sources [34] as the pre-trained text-to-image diffusion model \\(\\epsilon_{\\phi}^{\\omega}\\) or \\(\\mathbf{z}_{\\phi}^{\\omega}\\), respectively. We model \\(\\hat{b}_{\\psi}\\) as a standard U-Net with four encoder layers with two Conv/ReLU/MaxPool blocks per layer, and skip connections to the decoder. The first layer contains 32 filters and the number of filters is doubled in each layer in the encoder and halved in the decoder. \\(\\hat{b}_{\\psi}\\) has been trained until convergence, using \\(\\mathcal{L}_{\\mathrm{k}}\\) as the only loss. For creating training triplets \\((\\mathbf{z},\\hat{\\mathbf{z}}_t,t)\\), we sampled random images \\(\\mathbf{z}\\) from OpenImages [18], random time-steps \\(t \\sim \\mathcal{U}(0.02,0.98)\\), and passed both into \\(\\mathbf{z}_{\\phi}\\) obtaining \\(\\hat{\\mathbf{z}}_t\\). We train \\(\\hat{b}_{\\psi}\\) once and keep it fixed for all experiments." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.117, + 0.229, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.127 + ], + "angle": 0, + "content": "Alldieck et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.787, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.258, + 0.788, + 0.342 + ], + "angle": 0, + "content": "Fig. 2: Optimization-based image synthesis results. We optimize an empty image to match a given prompt using our LMC-SDS, the original SDS loss, and \\(\\mathcal{L}_{\\mathrm{cond}}\\). SDS struggles to create detailed content when using low guidance \\(\\omega\\). High \\(\\omega\\) produces detailed results but colors may be oversaturated (chimpanzee face), fake detail may appear (2nd mouse tail), or artifacts emerge. \\(\\mathcal{L}_{\\mathrm{cond}}\\) is unstable to optimize and produces unrealistic colors. In contrast, our method produces detailed results with balanced colors." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.361, + 0.376, + 0.379 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.385, + 0.787, + 0.431 + ], + "angle": 0, + "content": "We evaluate our proposed LMC-SDS loss in a number of qualitative and quantitative experiments, including image synthesis, image editing, image-to-image translation network training, and 3D asset generation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.431, + 0.789, + 0.597 + ], + "angle": 0, + "content": "Set-up. If not noted otherwise, we scale the total loss by \\(\\frac{1}{\\omega}\\) to be able to ablate \\(\\omega\\) without changing the gradient magnitude. Fixed \\(\\epsilon\\) sampling is only used when explicitly stated. The pre-trained diffusion model we use in our experiments produces images of \\(128\\times 128\\mathrm{px}\\) resolution. We enable arbitrary higher resolutions in our image experiments through the following strategy: We downsample the optimization state to \\(128\\times 128\\mathrm{px}\\) for the first iterations. This ensures that the whole image is affected by the loss, but the signal may be blurry. Accounting for this, we sample random patches of arbitrary resolutions, from the image, for the remaining optimization steps. To prevent the loss from producing fake detail in sampled image patches, we lower the learning rate during this phase through Cosine decay." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.789, + 0.809 + ], + "angle": 0, + "content": "Baselines. This paper presents a generic novel loss function which uses an image diffusion model. To this end, we compare with the original SDS formulation and with related work focused on improving SDS. Specifically, we compare against DDS [8], NFSD [13], and multi-step SDS (MS-SDS) [40]. DDS, NFSD, and MS-SDS originally were proposed using a latent diffusion model [31] and operate in latent space. For fair comparisons, we adapted all methods to operate in pixel space and to use the same image diffusion model. Note that optimizing in the (ambient) pixel space is generally considered harder due to higher-dimensionality. Finally, we evaluate the effectiveness of our LMC-SDS loss for 3D asset generation using DreamFusion. Hereby we focus on the improvements of our loss and deliberately exclude recent improvements over DreamFusion rooted in other factors, e.g. novel-view synthesis networks or additional perceptual losses. While being important findings, those factors are orthogonal and complementary to our approach." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Metrics. We use the following metrics for evaluating our methods and baselines: We report Learned Perceptual Image Patch Similarity (LPIPS \\(\\downarrow\\)) [39] w.r.t." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.662, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "LMC-SDS" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.41, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.254, + 0.22, + 0.373, + 0.229 + ], + "angle": 0, + "content": "\"oil paiting of a landscape\"" + }, + { + "type": "image", + "bbox": [ + 0.412, + 0.145, + 0.597, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.468, + 0.22, + 0.535, + 0.228 + ], + "angle": 0, + "content": "\"a tiny dragon\"" + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.145, + 0.785, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.66, + 0.22, + 0.721, + 0.228 + ], + "angle": 0, + "content": "\"ahoneybee\"" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.229, + 0.409, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.265, + 0.304, + 0.359, + 0.312 + ], + "angle": 0, + "content": "\"bananas in a basket\"" + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.229, + 0.596, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.473, + 0.304, + 0.531, + 0.312 + ], + "angle": 0, + "content": "\"akangaroo\"" + }, + { + "type": "image", + "bbox": [ + 0.596, + 0.23, + 0.785, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.632, + 0.304, + 0.747, + 0.313 + ], + "angle": 0, + "content": "\"amarble bust of Caesar\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.316, + 0.784, + 0.343 + ], + "angle": 0, + "content": "Fig. 3: Examples of optimization-based image editing results. We show pairs of input images (left) and editing result (right)." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.349, + 0.331, + 0.437 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.257, + 0.439, + 0.294, + 0.45 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.332, + 0.349, + 0.557, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.361, + 0.439, + 0.415, + 0.449 + ], + "angle": 0, + "content": "seed = 1" + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.349, + 0.672, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.587, + 0.439, + 0.643, + 0.45 + ], + "angle": 0, + "content": "seed \\(= 3\\)" + }, + { + "type": "image", + "bbox": [ + 0.672, + 0.349, + 0.785, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.699, + 0.439, + 0.757, + 0.45 + ], + "angle": 0, + "content": "seed \\(= 4\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.453, + 0.785, + 0.495 + ], + "angle": 0, + "content": "Fig. 4: By fixing \\(\\epsilon\\) in \\(\\mathcal{L}_{\\mathrm{cond}}\\) we can obtain diverse editing results. We show four variants of optimization-based editing results of the input image under the given prompt \"a mountain during sunset\"." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.51, + 0.785, + 0.587 + ], + "angle": 0, + "content": "the source image, thus capturing how much an image was altered. Note that high LPIPS can be a result of either successful editing or image corruption. To disambiguate these, we additionally report the CLIP score \\((\\uparrow)\\) [9] w.r.t. the target prompt. As CLIP can be insensitive, we additionally report CLIP-R-Precision \\((\\uparrow)\\) [28], the retrieval accuracy using the CLIP model, where feasible." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.609, + 0.484, + 0.625 + ], + "angle": 0, + "content": "5.1 Image Synthesis & Editing" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.784, + 0.658 + ], + "angle": 0, + "content": "We first explore our LMC-SDS loss for optimization-based image synthesis and image editing." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.786, + 0.84 + ], + "angle": 0, + "content": "Synthesis. To be able to evaluate and ablate the original SDS loss and our LMC-SDS in isolation (without additional interfering losses), we propose the task of optimization-based image synthesis. Instead of using the denoising schedule of a diffusion model, we initialize an image with \\(50\\%\\) grey and optimize its pixel values to match a given prompt, following the patching strategy described earlier. While it is not expected that images obtained using this strategy are on par with those obtained via the generation process, the results allow to draw meaningful conclusions about the properties of the loss functions. We show results of our loss and SDS in Fig. 2. We additionally include results obtained by only using \\(\\mathcal{L}_{\\mathrm{cond}}\\), ablating the influence of our proposed \\(\\mathcal{L}_{\\mathrm{lmc}}\\). Optimizing solely with \\(\\mathcal{L}_{\\mathrm{cond}}\\) is unstable and the final images feature clearly visible defects, such as oversaturated colors or noisy patterns. The prompt is also not always well re" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.127 + ], + "angle": 0, + "content": "Aldieck et al." + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.145, + 0.455, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.467, + 0.143, + 0.782, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.24, + 0.788, + 0.352 + ], + "angle": 0, + "content": "Fig. 5: Quantitative results for optimization-based editing under varying \\(\\omega\\). Left: Our method results in the highest CLIP score over all baselines for all \\(\\omega\\). Right: We plot LPIPS over CLIP for further performance insights: DDS stays close to the original image (lowest LPIPS) by performing only small edits (low CLIP). SDS and MS-SDS respect the prompt better (higher CLIP), but corrupt the image (high LPIPS). NFSD corrupts the image less (lower LPIPS), but exhibits weak editing capabilities (low CLIP). Our method shows the strongest editing capabilities (highest CLIP), while staying close to the original structure." + }, + { + "type": "image", + "bbox": [ + 0.286, + 0.357, + 0.497, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.356, + 0.717, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.463, + 0.788, + 0.518 + ], + "angle": 0, + "content": "Fig. 6: Comparison of image statistics before (Original) and after editing with \\(\\omega = 5\\). We plot the average power spectra (left) and vertical derivative histograms (right). Ours, DDS, and NFSD preserve the image statistics well, while SDS-based methods introduce significant blur." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.535, + 0.788, + 0.655 + ], + "angle": 0, + "content": "spected as e.g. in the case of the gibbon. The original SDS formulation produces blurry and desaturated images when using low guidance \\(\\omega\\). High \\(\\omega\\) leads to better performance but artifacts are present, e.g. fake detail like the second mouse tail, oversaturated colors like in the chimpanzee's face, or blocky artifacts. In contrast, our method produces images of higher visual fidelity which respect the prompt well. Additionally, we empirically found that problems of the individual losses become more pronounced when optimizing longer (here we only optimize for 500 steps, cf. Sec. 5.2 for effects after 50K steps)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.658, + 0.788, + 0.794 + ], + "angle": 0, + "content": "Editing. Our loss can also be used for optimization-based image editing. We load existing natural images and optimize their pixel values directly. Optionally, we may constrain the optimization with an L2 loss w.r.t. the original pixel values. In Fig. 3 (more in Supp. Mat.) we show qualitative results of so edited images along with the corresponding target prompt. All results faithfully respect the target prompt while keeping the image structure intact. We can also vary the result by fixing \\(\\epsilon\\) over the course of the optimization, cf. Sec. 4. In Fig. 4, we show an input image along with four result variants. All results respect the target prompt well, while at the same time being significantly different from each other." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.842 + ], + "angle": 0, + "content": "For a numerical evaluation, we use an internal dataset of 1,000 images with text editing instructions in form of source and target prompts, similar to the dataset used in InstructPix2Pix [2]. We optimize images to match the target" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.662, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "LMC-SDS" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.329, + 0.788, + 0.385 + ], + "angle": 0, + "content": "Fig. 7: Results obtained with cats-to-others image translation networks trained with different losses. The network trained with SDS and NFSD fail to create detailed results. The DDS network produces rather uniform results, whereas the network trained with our LMC-SDS creates realistic and detailed images." + }, + { + "type": "table", + "bbox": [ + 0.306, + 0.396, + 0.695, + 0.482 + ], + "angle": 0, + "content": "
LPIPS ↓CLIP Score ↑CLIP-R-Precision ↑
DDS0.1430.2530.333
NFSD0.3320.2550.434
SDS0.4010.2560.355
SDS, ω = 1000.2010.2590.482
Ours0.1780.2610.592
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.485, + 0.788, + 0.54 + ], + "angle": 0, + "content": "Table 1: Numerical results of the cats-to-others network training experiment. Note that LPIPS is computed w.r.t. the source image, meaning a low score may mean the image has not been edited enough. This is the case for DDS, as evident from low R-Precision." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.789, + 0.842 + ], + "angle": 0, + "content": "prompt solely using our LMC-SDS and baselines DDS, NFSD, SDS, and MS-SDS. For MS-SDS we chose to use five step denoising. We also experimented with more steps, but didn't observe significantly different performance. We report numerical results with varying \\(\\omega\\) (without loss rescaling) in Fig. 5. We found DDS to be least effective and producing results closest to the original image. We hypothesize that the negative gradient w.r.t. to the original image may not always be helpful. This could also explain a behavior reported by original authors, where editing results on real images are \"inferior\" over those on synthetic images. A possible reason being, that since for synthetic images the original prompts are describing the image well, the negative gradients are more helpful in that case. SDS can effectively edit the image, but also corrupts the image. This is also evident from the altered image statistics, see Fig. 6. NFSD corrupts the image less, but exhibits weak editing capabilities. In our experiments, MS-SDS performs very similar to SDS, corrupting the image in similar manner, but also being slightly less effective for editing. MS-SDS produces clear and detailed denoised images as part of the loss computation. We hypothesize that since denoised images and the current optimization state may differ drastically, and as denoised images vary considerably at each step, the computed gradients are not always very meaningful. The original paper proposes MS-SDS" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "Alldieck et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "to be used with a view-conditioned diffusion model, where the denoised images probably vary much less and computed gradients may be more directed. Our loss respects the prompt best (highest CLIP), and finds a good balance between preserving the original structure and performing significant edits." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.788, + 0.313 + ], + "angle": 0, + "content": "To quantify the how much the result diversity can be increased through fixing \\(\\epsilon\\), we re-ran the experiment using LMC-SDS (\\(\\omega = 30\\)) ten times with different seeds, both using the standard strategy and with fixing \\(\\epsilon\\) in \\(\\mathcal{L}_{\\mathrm{cond}}\\). We then computed the LPIPS (lower is more similar) of nine runs w.r.t. to the 10th (reference) run. The results using the standard strategy are all very similar with a per-run average of \\(0.041 \\pm 0.007\\) w.r.t. the reference run. In contrast, our new strategy produces much more diverse results: \\(0.158 \\pm 0.018\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.331, + 0.642, + 0.346 + ], + "angle": 0, + "content": "5.2 Image-to-image Translation Network Training" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.351, + 0.789, + 0.76 + ], + "angle": 0, + "content": "Optimization-based editing is not always an optimal choice, e.g. when building interactive editing tools where execution speed matters. Further, optimization can be unstable or perform better or worse, depending on the input images or target prompts. To this end, we explore to what extend LMC-SDS can be used to train a reliable and fast image-to-image translation network. Given a known distribution of images, we train a network to translate images from that distribution to an unseen target distribution, only described by a target caption. Inspired by the authors of DDS [8], we train a cats-to-others network, with \"others\" being bears, dogs, lions, and squirrels. The network transforms an image of a cat into an image of one of the other species. In DDS, the authors fine-tune a pre-trained diffusion model into a single-step feedforward model using images created by that same diffusion model. In contrast, we train a standard U-Net from scratch and use a dataset of real images, a much more practical and versatile approach. Concretely, we train using cat images from AFHQ [3] and evaluate on \\(5\\%\\) held out test images. We use the target prompt \"a photo of a [class]\" with \\(\\omega = 15\\) for all methods and the source prompt \"a photo of a cat\" for DDS. Moreover, we stabilize the training using an L2 loss on the original image and use Cosine scheduling to let the network first focus on reconstructing the original image and then on edits, similarly to the experiment in DDS. The class is passed via a frequency encoded class id concatenated to each pixel. We report numerical results in Tab. 1 and show qualitative examples in Fig. 7. Additional analysis is provided in the Suppl. Mat. SDS performs similarly to previous experiments for this task. While the target class is evident in the resulting images, the results feature strong deficiencies. Again DDS struggles with significant edits and stays close to the source. In contrast, the network trained with our LMC-SDS loss produces realistic and convincing edits. This is also evident from the significantly higher CLIP-R-Precision among all network variants." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.776, + 0.504, + 0.79 + ], + "angle": 0, + "content": "5.3 Text-to-3D Asset Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "In our final experiment, we demonstrate how LMC-SDS can improve text-to-3D models in the style of DreamFusion. In Sec. 3, we discussed how large guidance weights \\(\\omega\\) may affect the quality of results. To provide further evidence for this" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.662, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "LMC-SDS" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.144, + 0.784, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.23, + 0.785, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.325, + 0.787, + 0.421 + ], + "angle": 0, + "content": "Fig. 8: The effect of increasing guidance weight \\(\\omega\\) in DreamFusion using the original SDS loss (even rows) and our proposed LMC-SDS (odd rows). Using larger \\(\\omega\\) saturates colors (chair, motorcycle, ship) and worsens the Janus problem (third wing and claw for the eagle). For low \\(\\omega\\), vanilla DreamFusion fails to produce detailed geometry (ship, eagle) or geometry at all (motorcycle), while our approach performs well for \\(\\omega = 20\\), often even for \\(\\omega = 10\\). Notice how our approach constantly produces more detailed results for all \\(\\omega\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.788, + 0.657 + ], + "angle": 0, + "content": "observation, we produce 3D assets with varying \\(\\omega\\) using the original DreamFusion formulation and a variant using our LMC-SDS. The results in Fig. 8 also show the trend of high \\(\\omega\\) having the tendency to produce oversaturated colors, albeit not as strongly as in the original paper, possibly due the use of a different image diffusion model. But we also observed an additional issue: the infamous Janus problem, the tendency to repeat detail in each view, becomes more severe with bigger \\(\\omega\\), as evident from the eagle example. On the other hand, when using low \\(\\omega\\) in the original SDS loss, we observe that DreamFusion struggles to create detailed structure or even structure at all. In contrast, when using LMC-SDS, DreamFusion successfully produces detailed results, even when using up to only one tenth of the originally proposed \\(\\omega\\). In Fig. 9 (more in Suppl. Mat.), we show additional side-by-side comparisons of the original SDS-based DreamFusion and our LMC-SDS version. Using our loss, the results are consistently sharper and feature far more detail. Finally in Fig. 10, we show that fixed \\(\\epsilon\\) sampling may also be used in 3D asset generation, to obtain diverse results." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.677, + 0.491, + 0.693 + ], + "angle": 0, + "content": "6 Discussion & Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.84 + ], + "angle": 0, + "content": "We proposed a novel loss formulation called LMC-SDS. LMC-SDS allows to effectively utilize text-to-image diffusion models as rich image priors in several optimization and network training tasks. We derived LMC-SDS by analyzing DreamFusion's SDS loss and by identifying an inherent problem in its formulation. Concretely, we expressed the SDS gradient as the sum of two terms. We identified that the second term \\(\\mathcal{L}_{\\mathrm{proj}}\\) provides deficient gradients and consequently was weighted down in DreamFusion. We argued that this term, however, should provide gradients towards the learned manifold of natural images associated to the diffusion model. We also provided evidence that the lack thereof" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.127 + ], + "angle": 0, + "content": "Aldieck et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.787, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.323, + 0.785, + 0.364 + ], + "angle": 0, + "content": "Fig. 9: Results of DreamFusion using the original SDS loss (right) and our proposed LMC-SDS (left) for \\(\\omega = 20\\). Our results are much sharper, contain more detail, and feature more realistic colors in all examples." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.37, + 0.78, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.423, + 0.785, + 0.465 + ], + "angle": 0, + "content": "Fig. 10: DreamFusion results using our LMC-SDS loss and fixed \\(\\epsilon\\) sampling for the prompts \"a mannequin wearing a [hoodie, hat]. Note the diversity in colors, poses, and mannequin styles." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.486, + 0.788, + 0.592 + ], + "angle": 0, + "content": "makes the optimization unstable, thus negatively impacting the fidelity of results. To this end, we proposed to model the denoising models' time step dependent frequency bias. This allowed us to factor it out and obtain the cleaner gradient \\(\\nabla_{\\theta}\\mathcal{L}_{\\mathrm{lmc}}\\). In extensive experiments, we have demonstrated the effectiveness and versatility of our novel LMC-SDS loss. Furthermore, we discussed how LMC-SDS differs from other SDS-style losses and demonstrated how its performance is superior both qualitatively and quantitatively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.593, + 0.788, + 0.759 + ], + "angle": 0, + "content": "Limitations. LMC-SDS finds its limits where either the signal from the diffusion model is too weak and ambiguous or when the optimization goes off the natural image manifold by too much. Concretely, LMC-SDS can only utilize whatever signal is coming from the diffusion model. This means, that when the diffusion model does not \"understand\" the prompt, LMC-SDS (and all other SDS-based approaches) will not be able to provide meaningful gradients. Further, whenever the current optimization state is too corrupted, LMC-SDS will not be able to correct for that and guide the optimization back to the manifold. Also, LMC-SDS sometimes overcompensates the frequency bias. Nevertheless, LMC-SDS is an important first step towards more stable and meaningful usage of the inherent image prior associated to image diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.759, + 0.788, + 0.835 + ], + "angle": 0, + "content": "We proposed LMC-SDS, a novel SDS-based loss formulation with cleaner gradients towards the image manifold. In the future, we would like to continue to investigate how the manifold corrective can be further improved. We also want to leverage our findings in concrete applications like text-to-3D or in image editing and stylization tasks." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "LMC-SDS" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.324, + 0.161 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.786, + 0.218 + ], + "angle": 0, + "content": "1. Armandpour, M., Zheng, H., Sadeghian, A., Sadeghian, A., Zhou, M.: Re-imagine the negative prompt algorithm: Transform 2d diffusion into 3d, alleviate janus problem and beyond. arXiv preprint arXiv:2304.04968 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.219, + 0.786, + 0.245 + ], + "angle": 0, + "content": "2. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.246, + 0.786, + 0.273 + ], + "angle": 0, + "content": "3. Choi, Y., Uh, Y., Yoo, J., Ha, J.W.: Stargan v2: Diverse image synthesis for multiple domains. In: IEEE Conf. Comput. Vis. Pattern Recog. (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.274, + 0.786, + 0.301 + ], + "angle": 0, + "content": "4. Chung, H., Kim, J., Mccann, M.T., Klasky, M.L., Ye, J.C.: Diffusion posterior sampling for general noisy inverse problems. In: Int. Conf. Learn. Represent. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.302, + 0.786, + 0.342 + ], + "angle": 0, + "content": "5. Chung, H., Sim, B., Ryu, D., Ye, J.C.: Improving diffusion models for inverse problems using manifold constraints. Adv. Neural Inform. Process. Syst. 35, 25683-25696 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.343, + 0.786, + 0.37 + ], + "angle": 0, + "content": "6. Croitoru, F.A., Hondru, V., Ionescu, R.T., Shah, M.: Diffusion models in vision: A survey. IEEE Trans. Pattern Anal. Mach. Intell. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.371, + 0.786, + 0.424 + ], + "angle": 0, + "content": "7. Gao, S., Liu, X., Zeng, B., Xu, S., Li, Y., Luo, X., Liu, J., Zhen, X., Zhang, B.: Implicit diffusion models for continuous super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10021-10030 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.426, + 0.786, + 0.453 + ], + "angle": 0, + "content": "8. Hertz, A., Aberman, K., Cohen-Or, D.: Delta denoising score. In: Int. Conf. Comput. Vis. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.454, + 0.786, + 0.481 + ], + "angle": 0, + "content": "9. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.482, + 0.786, + 0.508 + ], + "angle": 0, + "content": "0. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Adv. Neural Inform. Process. Syst. 33, 6840-6851 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.509, + 0.786, + 0.536 + ], + "angle": 0, + "content": "1. Ho, J., Salimans, T.: Classifier-free diffusion guidance. In: NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.537, + 0.786, + 0.578 + ], + "angle": 0, + "content": "2. Jain, A., Xie, A., Abbeel, P.: Vectorfusion: Text-to-svq by abstracting pixel-based diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1911-1920 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.579, + 0.786, + 0.605 + ], + "angle": 0, + "content": "3. Katzir, O., Patashnik, O., Cohen-Or, D., Lischinski, D.: Noise-free score distillation. arXiv preprint arXiv:2310.17590 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.606, + 0.786, + 0.633 + ], + "angle": 0, + "content": "4. Kawar, B., Elad, M., Ermon, S., Song, J.: Denoising diffusion restoration models. In: Adv. Neural Inform. Process. Syst. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.634, + 0.786, + 0.674 + ], + "angle": 0, + "content": "5. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 6007-6017 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.675, + 0.786, + 0.702 + ], + "angle": 0, + "content": "6. Kim, S., Lee, K., Choi, J.S., Jeong, J., Sohn, K., Shin, J.: Collaborative score distillation for consistent visual synthesis. arXiv preprint arXiv:2307.04787 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.703, + 0.786, + 0.743 + ], + "angle": 0, + "content": "7. Kolotouros, N., Alldieck, T., Zanfir, A., Bazavan, E.G., Fieraru, M., Sminchisescu, C.: Dreamhuman: Animatable 3d avatars from text. In: Adv. Neural Inform. Process. Syst. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.744, + 0.786, + 0.798 + ], + "angle": 0, + "content": "8. Kuznetsova, A., Rom, H., Alldrin, N., Uijlings, J., Krasin, I., Pont-Tuset, J., Kamali, S., Popov, S., Malloci, M., Kolesnikov, A., Duerig, T., Ferrari, V.: The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. Int. J. Comput. Vis. (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.786, + 0.84 + ], + "angle": 0, + "content": "9. Li, X., Ren, Y., Jin, X., Lan, C., Wang, X., Zeng, W., Wang, X., Chen, Z.: Diffusion models for image restoration and enhancement-a comprehensive survey. arXiv preprint arXiv:2308.09388 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.177, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.127 + ], + "angle": 0, + "content": "Aldieck et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "20. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3d: High-resolution text-to-3d content creation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 300–309 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.217 + ], + "angle": 0, + "content": "21. Lin, S., Liu, B., Li, J., Yang, X.: Common diffusion noise schedules and sample steps are flawed. arXiv preprint arXiv:2305.08891 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.244 + ], + "angle": 0, + "content": "22. Liu, Q., Wang, D.: Stein variational gradient descent: A general purpose bayesian inference algorithm. Advances in neural information processing systems 29 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.785, + 0.285 + ], + "angle": 0, + "content": "23. Liu, R., Wu, R., Van Hoorick, B., Tokmakov, P., Zakharov, S., Vondrick, C.: Zero-1-to-3: Zero-shot one image to 3d object. In: Int. Conf. Comput. Vis. pp. 9298-9309 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.327 + ], + "angle": 0, + "content": "24. Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Van Gool, L.: Repaint: Inpainting using denoising diffusion probabilistic models. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 11461-11471 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.369 + ], + "angle": 0, + "content": "25. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: SDEdit: Guided image synthesis and editing with stochastic differential equations. In: Int. Conf. Learn. Represent. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.37, + 0.785, + 0.411 + ], + "angle": 0, + "content": "26. Metzer, G., Richardson, E., Patashnik, O., Giryes, R., Cohen-Or, D.: Latent-nerf for shape-guided generation of 3d shapes and textures. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 12663-12673 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.452 + ], + "angle": 0, + "content": "27. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In: Int. Conf. on Mach. Learn. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.494 + ], + "angle": 0, + "content": "28. Park, D.H., Azadi, S., Liu, X., Darrell, T., Rohrbach, A.: Benchmark for compositional text-to-image synthesis. In: Conf. on Neural Inf. Proc. Systems Datasets and Benchmarks Track (Round 1) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.785, + 0.522 + ], + "angle": 0, + "content": "29. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. In: Int. Conf. Learn. Represent. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.785, + 0.563 + ], + "angle": 0, + "content": "30. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.785, + 0.605 + ], + "angle": 0, + "content": "31. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 10684-10695 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.785, + 0.646 + ], + "angle": 0, + "content": "32. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 22500-22510 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.688 + ], + "angle": 0, + "content": "33. Sahara, C., Chan, W., Chang, H., Lee, C., Ho, J., Salimans, T., Fleet, D., Norouzi, M.: Palette: Image-to-image diffusion models. In: ACM SIGGRAPH 2022 Conference Proceedings. pp. 1-10 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.785, + 0.742 + ], + "angle": 0, + "content": "34. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Adv. Neural Inform. Process. Syst. 35, 36479-36494 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.743, + 0.785, + 0.771 + ], + "angle": 0, + "content": "35. Shi, Y., Wang, P., Ye, J., Mai, L., Li, K., Yang, X.: Mvdream: Multi-view diffusion for 3d generation. arXiv:2308.16512 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.799 + ], + "angle": 0, + "content": "36. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. In: Int. Conf. Learn. Represent. (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "37. Wang, H., Du, X., Li, J., Yeh, R.A., Shakhnarovich, G.: Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 12619-12629 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.661, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "LMC-SDS" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "38. Wang, Z., Lu, C., Wang, Y., Bao, F., Li, C., Su, H., Zhu, J.: Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "39. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: IEEE Conf. Comput. Vis. Pattern Recog. (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.232, + 0.788, + 0.273 + ], + "angle": 0, + "content": "40. Zhou, Z., Tulsiani, S.: Sparsefusion: Distilling view-conditioned diffusion for 3d reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12588-12597 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.273, + 0.788, + 0.315 + ], + "angle": 0, + "content": "41. Zou, Z.X., Cheng, W., Cao, Y.P., Huang, S.S., Shan, Y., Zhang, S.H.: Sparse3d: Distilling multiview-consistent diffusion for object reconstruction from sparse views. arXiv preprint arXiv:2308.14078 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.315 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Score Distillation Sampling with Learned Manifold Corrective/71b1df1d-3510-4a6f-9a35-a2f80ebbbb53_origin.pdf b/2024/Score Distillation Sampling with Learned Manifold Corrective/71b1df1d-3510-4a6f-9a35-a2f80ebbbb53_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..82d3aa7b0ca28ab0a2418cf934f84073f50e62b8 --- /dev/null +++ b/2024/Score Distillation Sampling with Learned Manifold Corrective/71b1df1d-3510-4a6f-9a35-a2f80ebbbb53_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1181b8d7a7d1a57214b93f38bfd456dfee464f99d4cae92491a1fbfa1c9103ff +size 7993367 diff --git a/2024/Score Distillation Sampling with Learned Manifold Corrective/full.md b/2024/Score Distillation Sampling with Learned Manifold Corrective/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b22e1af855150ee315ac50af1da049fc59790570 --- /dev/null +++ b/2024/Score Distillation Sampling with Learned Manifold Corrective/full.md @@ -0,0 +1,288 @@ +# Score Distillation Sampling with Learned Manifold Corrective + +Thiemo Alldieck, Nikos Kolotouros, and Cristian Sminchisescu + +Google Research\* + +Abstract. Score Distillation Sampling (SDS) is a recent but already widely popular method that relies on an image diffusion model to control optimization problems using text prompts. aIn this paper, we conduct an in-depth analysis of the SDS loss function, identify an inherent problem with its formulation, and propose a surprisingly easy but effective fix. Specifically, we decompose the loss into different factors and isolate the component responsible for noisy gradients. In the original formulation, high text guidance is used to account for the noise, leading to unwanted side effects such as oversaturation or repeated detail. Instead, we train a shallow network mimicking the timestep-dependent frequency bias of the image diffusion model in order to effectively factor it out. We demonstrate the versatility and the effectiveness of our novel loss formulation through qualitative and quantitative experiments, including optimization-based image synthesis and editing, zero-shot image translation network training, and text-to-3D synthesis. + +# 1 Introduction + +Image diffusion models [10] have recently become the de facto standard for image generation. Especially text-to-image models [27,30,31,34] have emerged as powerful tools for high fidelity, diverse, image synthesis. By being controlled through natural language, these models are extremely easy to use and thus open up a wide range of creative applications, without the need for special training. Beyond image synthesis, image diffusion models have been successfully deployed in applications like image restoration, inpating, editing, super-resolution, or colorization [4,7,14,25], among others. Image diffusion models are typically obtained using very large training sets and thus inherently represent the distribution of natural images. While applications typically make use of the generation process of diffusion models, e.g. by inserting images into the process [24,25] or by altering the denoising function [4,5,14], relatively little research has been conducted on how diffusion models can be used as rich, general purpose image priors. Score Distillation Sampling (SDS) proposed in DreamFusion [29] is one of the few exceptions. The SDS loss formulation uses a pretrained text-to-image model [34] to measure how well a given text matches an observation. In DreamFusion this loss is being used to generate 3D assets from textual descriptions, an idea that was quickly adopted [17,20,23,26,32,37] and essentially established a new research direction: text-to-3D synthesis. While being mostly used in this context, the design + +![](images/92fd2c0d7eb99c8b112827956c346c8f9d50ae16387f90e9f92d2ffd741bb748.jpg) + +![](images/89828352ee95179d799d5a86f7af7c80ca6479f57dce5d81a46b1c5d5d1eb690.jpg) +Fig. 1: Left: Visualization of SDS and LMC-SDS gradients w.r.t. pixel values and estimated denoised images for the given image $\mathbf{z}$ , the prompt $\mathbf{y} = \text{"autumn"}$ , and $t = 0.5$ . We visualize the negative gradient, i.e. the direction of change. Right: Power spectra of denoised images $\hat{\mathbf{z}}_t$ for varying time-step $t$ compared to the power spectrum of natural images $\mathbf{z}$ . See Sec. 3 for details. + +![](images/318684887935c2babf608634a96a64b2ebee988815453627f685e267600e3a20.jpg) +Frequency + +of the SDS loss is by no means only tailored to text-to-3D applications. In fact, SDS is an image loss and can be used in much wider contexts [8,12]. However, in its original formulation the SDS loss may degrade the observation, be too eager to match the text prompt, or provide meaningless gradients which inject noise into the objective. In this paper, we conduct an extensive analysis of the SDS loss, identify an inherent problem with its formulation, propose a surprisingly easy but effective fix, and demonstrate its effectiveness in several applications, including optimization-based image synthesis and editing, zero-shot training of image translation networks, and text-to-3D synthesis. Concretely, our new loss formulation - Score Distillation Sampling with Learned Manifold Corrective, or LMC-SDS for short - aims to provide better gradients along the direction of the learned manifold of real images. We present evidence that gradients towards the learned manifold are extremely noisy in SDS and that high text guidance is needed to compensate for the noisy signal. Further, we show that high guidance - or lack thereof - is a possible explanation for the artifacts observed with SDS, cf. Sec. 3. In contrast, applications relying on our novel formulation benefit from meaningful manifold gradients, may use lower text guidance, and produce results of overall higher visual fidelity. + +In summary, our contributions are: (1) conducting an in-depth analysis of the SDS loss, (2) identifying an inherent problem with the formulation, and (3) proposing a novel LMC-SDS loss, providing cleaner gradients and eliminating the need for high guidance. (4) In extensive experiments, we demonstrate the effectiveness and versatility of the new LMC-SDS loss. + +# 2 Related Work + +Methods that use image diffusion models as priors for tasks other than image synthesis can be broadly categorized into two lines of work. + +The first class of methods leverage image diffusion networks to solve specific tasks by relying on the diffusion sampling process. We discuss here a number of representative methods and refer the reader to [6, 19] for detailed discussions. + +DDRM [14] tackles linear inverse problems such as deblurring, super-resolution, colorization, or inpainting by defining a modified variational posterior distribution conditioned on the observation signal, which can then be used for sampling. DPS [4] focuses on general inverse problems and proposes to augment the score prediction at each denoising step with an additional term that models the likelihood of the observations. MCG [5] builds on top and improves the fidelity of generation by introducing additional manifold constraints. SDEdit [25] uses diffusion models for image editing by first adding noise to the input image and then denoising it using the text-to-image diffusion prior. RePaint [24] uses unconditional image diffusion models as a prior for solving image inpainting. All these methods use pre-trained diffusion models. In contrast, methods like Palette [33] or Imagic [15] require training a diffusion model from scratch or fine-tuning a pretrained one, respectively. This line of work exclusively deals with image editing problems and is typically restricted to the resolution of the diffusion models. + +More related to our method is the second category: This body of work uses pre-trained diffusion models within general purpose loss functions in iterative optimization settings. Score Jacobian Chaining [37] propagates the score of a text-to-image diffusion model through a differentiable renderer to supervise the generation of 3D models. DreamFusion [29] takes an orthogonal approach and proposes the Score Distillation Sampling loss. SDS is a modification of the diffusion training objective that does not require expensive backpropagation through the diffusion model and encourages the alignment of the generated image with a conditioning signal. DDS [8] proposes an improved version of SDS, specifically designed for image editing and aims to reduce the artifacts introduced by noisy SDS gradients. In similar spirit and in concurrent work, NFSD [13] aims to reduce noise in the gradients by adding negative conditioning. In contrast, our method neither requires a source prompt nor a negative prompt. SparseFusion [40] introduces a multi-step SDS loss that performs multiple denoising steps instead of one, which adds significant computational overhead. ProlificDreamer [38] proposes a generalization of SDS from point estimates to distributions, but unlike our approach requires computationally expensive fine-tuning of a diffusion model during optimization. Collaborative Score Distillation [16] generalizes SDS to multiple samples and is inspired by Stein variational gradient descent [22]. Category Score Distillation Sampling (C-SDS) focuses on 3D generation and replaces the noise estimation error in SDS with a difference between the noise predictions of a standard and a multi-view consistent diffusion model [41]. Unlike our method, C-SDS is only applicable to 3D generation tasks. In contrast, we propose a general purpose loss function with very little computational overhead. + +# 3 Analysis + +An image diffusion model is a generative model that is trained to reverse the diffusion process that transforms a natural image into pure noise [10]. An image is formed by iteratively removing small amounts of Gaussian noise over a fixed variance schedule $\bar{\alpha}_t$ until an image is created. Text-to-image models are additionally conditioned on text to steer the denoising process in a direction matching + +the textural description. For a given image $\mathbf{z}$ , its textual description $\pmb{y}$ , a randomly sampled timestep $t\sim \mathcal{U}(0,1)$ , and random Gaussian noise $\epsilon \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ , a denoising model $\epsilon_{\phi}$ parameterized by $\phi$ can be trained by minimizing the diffusion loss + +$$ +\mathcal {L} _ {\text {d i f f}} = w (t) \left\| \epsilon_ {\phi} ^ {\omega} (\mathbf {z} _ {t}, \mathbf {y}, t) - \epsilon \right\| _ {2} ^ {2}, \tag {1} +$$ + +where $\mathbf{z}_t = \sqrt{\bar{\alpha}_t}\mathbf{z} + \sqrt{1 - \bar{\alpha}_t}\epsilon$ refers to the noisy version of $\mathbf{z}$ and $w(t)$ is a weighting function, omitted in the sequel. DreamFusion [29] showed that given a pre-trained model $\epsilon_{\phi}$ the diffusion loss can be utilized for optimization problems. For an arbitrary differentiable rendering function returning $\mathbf{z}$ and parameterized by $\pmb{\theta}$ , the gradient of the denoising function w.r.t. $\pmb{\theta}$ is given by + +$$ +\nabla_ {\boldsymbol {\theta}} \mathcal {L} _ {\text {d i f f}} = \left(\boldsymbol {\epsilon} _ {\phi} ^ {\omega} (\mathbf {z} _ {t}, \boldsymbol {y}, t) - \boldsymbol {\epsilon}\right) \frac {\partial \boldsymbol {\epsilon} _ {\phi} ^ {\omega} (\mathbf {z} , \boldsymbol {y} , t)}{\partial \mathbf {z} _ {t}} \frac {\partial \mathbf {z} _ {t}}{\partial \boldsymbol {\theta}}. \tag {2} +$$ + +In practice, the Jacobian term is omitted to avoid backpropagating through the denoising model and the gradient is approximated as + +$$ +\nabla_ {\boldsymbol {\theta}} \mathcal {L} _ {\mathrm {S D S}} = \left(\boldsymbol {\epsilon} _ {\phi} ^ {\omega} (\mathbf {z} _ {t}, \boldsymbol {y}, t) - \boldsymbol {\epsilon}\right) \frac {\partial \mathbf {z} _ {t}}{\partial \boldsymbol {\theta}}, \tag {3} +$$ + +resulting in the gradient of original SDS loss. We will now rewrite Eq. (3) and conduct an analysis of the derived components. + +Using classifier-free guidance [11], the predicted noise is the sum of the $\pmb{y}$ -conditioned and the unconditioned noise prediction weighted by the guidance weight $\omega$ + +$$ +\boldsymbol {\epsilon} _ {\phi} ^ {\omega} \left(\mathbf {z} _ {t}, \boldsymbol {y}, t\right) = \omega \boldsymbol {\epsilon} _ {\phi} \left(\mathbf {z} _ {t}, \boldsymbol {y}, t\right) + (1 - \omega) \boldsymbol {\epsilon} _ {\phi} \left(\mathbf {z} _ {t}, t\right), \tag {4} +$$ + +which we can rewrite as + +$$ +\boldsymbol {\epsilon} _ {\phi} ^ {\omega} \left(\mathbf {z} _ {t}, \boldsymbol {y}, t\right) = \omega \left(\boldsymbol {\epsilon} _ {\phi} \left(\mathbf {z} _ {t}, \boldsymbol {y}, t\right) - \boldsymbol {\epsilon} _ {\phi} \left(\mathbf {z} _ {t}, t\right)\right) + \boldsymbol {\epsilon} _ {\phi} \left(\mathbf {z} _ {t}, t\right). \tag {5} +$$ + +By inserting Eq. (5) into Eq. (3) we can express $\nabla_{\pmb{\theta}}\mathcal{L}_{\mathrm{SDS}}$ as + +$$ +\nabla_ {\boldsymbol {\theta}} \mathcal {L} _ {\mathrm {S D S}} = \nabla_ {\boldsymbol {\theta}} \mathcal {L} _ {\text {c o n d}} + \nabla_ {\boldsymbol {\theta}} \mathcal {L} _ {\text {p r o j}} \tag {6a} +$$ + +$$ +\nabla_ {\boldsymbol {\theta}} \mathcal {L} _ {\text {c o n d}} = \omega \left(\epsilon_ {\phi} \left(\mathbf {z} _ {t}, \boldsymbol {y}, t\right) - \epsilon_ {\phi} \left(\mathbf {z} _ {t}, t\right)\right) \frac {\partial \mathbf {z} _ {t}}{\partial \boldsymbol {\theta}} \tag {6b} +$$ + +$$ +\nabla_ {\boldsymbol {\theta}} \mathcal {L} _ {\text {p r o j}} = \left(\epsilon_ {\phi} \left(\mathbf {z} _ {t}, t\right) - \boldsymbol {\epsilon}\right) \frac {\partial \mathbf {z} _ {t}}{\partial \boldsymbol {\theta}}. \tag {6c} +$$ + +The two loss components can be interpreted as follows: $\mathcal{L}_{\mathrm{cond}}$ maximizes the agreement of the image with the text prompt by providing gradients towards images formed through conditioning on $\pmb{y}$ . $\mathcal{L}_{\mathrm{proj}}$ performs a single denoising step and provides gradients informing how the image $\mathbf{z}$ (and thereby parameters $\pmb{\theta}$ ) should change such that the image can be better denoised. As the denoising function was trained to minimize the expected squared error between the denoised and the original data, $\nabla_{\pmb{\theta}}\mathcal{L}_{\mathrm{proj}}$ can be understood as the gradient direction towards the manifold of natural images. + +DreamFusion [29] originally proposed to set $\omega = 100$ . This means that, following our derivation, $\mathcal{L}_{\mathrm{cond}}$ is weighted proportionally much higher than $\mathcal{L}_{\mathrm{proj}}$ . This can be problematic as large guidance weights have been identified to be an important cause of over-exposure in diffusion models [21]. In fact, DreamFusion produces detailed results but also has a tendency towards unrealistically saturated colors, see also Fig. 8. On the other hand, setting $\omega$ to lower values has been reported to result in very blurry images [8], a behavior that we also demonstrate in Sec. 5. Both behaviors, (1) the tendency to produce overly saturated results when using high guidance as well as (2) blurry results for low guidance, can be explained by looking at the components of $\mathcal{L}_{\mathrm{SDS}}$ in isolation. To this end, we consider the case where the rendering function is a no-op and simply returns pixel values, i.e. $\mathbf{z} = \boldsymbol{\theta}$ . In Fig. 1 (left), we visualize $\nabla_{\boldsymbol{\theta}}\mathcal{L}_{\mathrm{cond}}$ and $\nabla_{\boldsymbol{\theta}}\mathcal{L}_{\mathrm{proj}}$ for this case. $\nabla_{\boldsymbol{\theta}}\mathcal{L}_{\mathrm{cond}}$ returns the direction towards an image with altered colors. However, as $\nabla_{\boldsymbol{\theta}}\mathcal{L}_{\mathrm{cond}}$ is not anchored in the manifold of natural images, the gradient may eventually point away from that manifold. $\nabla_{\boldsymbol{\theta}}\mathcal{L}_{\mathrm{cond}}$ will always point towards changes that "better" respect the prompt and eventually produce saturated colors or other artifacts (cf. Sec. 5). $\nabla_{\boldsymbol{\theta}}\mathcal{L}_{\mathrm{proj}}$ , on the other hand, incorrectly marks high-frequency detail originating from $\mathbf{z}$ for removal - the reason for blurriness. This behavior can be understood when rewriting $\nabla_{\boldsymbol{\theta}}\mathcal{L}_{\mathrm{proj}}$ as + +$$ +\nabla_ {\pmb {\theta}} \mathcal {L} _ {\mathrm {p r o j}} = \left(\frac {\sqrt {\alpha_ {t}}}{\sqrt {1 - \alpha_ {t}}} \left(\mathbf {z} - \mathbf {z} _ {\phi} (\mathbf {z} _ {t}, t)\right)\right) \frac {\partial \mathbf {z} _ {t}}{\partial \pmb {\theta}}, \tag {7} +$$ + +which is equivalent to comparing the predicted and true noise. By looking at Eq. (7), we can now see a possible explanation why $\mathcal{L}_{\mathrm{proj}}$ might not be optimal: $\hat{\mathbf{z}}_t = \mathbf{z}_\phi (\mathbf{z}_t,t)$ is an approximation of $\mathbf{z}$ and most probably not a perfect replica of $\mathbf{z}$ , even when $\mathbf{z}$ already lies on the natural image manifold, see Fig. 1 (left). This is especially the case for large values of $t$ , where the denoising model has to reconstruct the true image from almost pure noise. More formally, we can observe a frequency bias in $\mathbf{z}_\phi$ dependent on $t$ when comparing images $\mathbf{z}$ with their denoised counterparts $\hat{\mathbf{z}}_t$ , see Fig. 1 (right). + +We conclude the following: (1) The reason for oversaturated colors and other artifacts is due to $\mathcal{L}_{\mathrm{cond}}$ being "over-eager" to maximize the agreement of the image with the conditioning variable. (2) $\mathcal{L}_{\mathrm{proj}}$ provides deficient gradients ultimately removing high-frequency detail instead of "correcting" the gradient towards on-manifold solutions due to a frequency bias in $\mathbf{z}_{\phi}$ . Thus, $\mathcal{L}_{\mathrm{proj}}$ appears to be a non-optimal choice and we provide an alternative in the sequel. But first we use our decomposition to discuss other approaches to "fix" the SDS loss. + +DDS [8] proposes to use the difference between two SDS losses for image editing, one computed on the current state w.r.t. the target prompt and one computed on the initial image w.r.t. a textual description of that image. By computing an SDS difference, DDS entirely removes $\mathcal{L}_{\mathrm{proj}}$ and introduces a negative gradient pointing away from the initial state. However, the need for an initial image is limiting possible use-cases. In concurrent work, NFSD [13] aims to obtain better results by replacing $\mathcal{L}_{\mathrm{proj}}$ with a $\mathcal{L}_{\mathrm{cond}}$ variant using negative prompts for most timesteps, and a loss based on $\epsilon$ for small $t$ . While the root + +for blurry results is removed by eliminating $\mathcal{L}_{\mathrm{proj}}$ , both DDS and NFSD do not anchor the optimization along the learned manifold. Other methods aim for a better noise prediction, where by improving $\epsilon_{\phi}(\mathbf{z},t)$ , one may also improve $\mathcal{L}_{\mathrm{proj}}$ ( $\mathcal{L}_{\mathrm{proj}}$ is the difference between the predicted and true noise). E.g. SparseFusion [40] introduces multi-step SDS. Hereby, instead of denoising in one step, up to 50 steps are used. While this may produce more informative gradients in $\mathcal{L}_{\mathrm{proj}}$ , this is also significantly more expensive. Through different means but towards the same goal, the concurrent work ProlificDreamer [38] learns a low-rank adaptation of the text-to-image model. While this approach is effective, it also adds the overhead of fine-tuning an additional diffusion model. Finally, some recent methods aim to alleviate artifacts by augmenting the original SDS formulation with advanced schedules for sampling $t$ or with negative prompts [1,35]. In contrast, we aim for a generally applicable solution. + +# 4 Method + +We now present our solution to eliminating the frequency bias in $\mathcal{L}_{\mathrm{proj}}$ that we identified earlier. In an attempt to better understand the behavior of the denoising model given $\mathbf{z}$ and $t$ , we model it as a two step process, as follows + +$$ +\mathbf {z} _ {\phi} \left(\mathbf {z} _ {t}, t\right) = b \circ p \left(\mathbf {z} _ {t}, t\right) = b \left(\tilde {\mathbf {z}}, t\right), \tag {8} +$$ + +where $p$ projects $\mathbf{z}$ onto the learned natural image manifold, $\tilde{\mathbf{z}}$ is the projected image, and $\tilde{\mathbf{z}} = \mathbf{z}$ for images on the manifold. $b$ is the error or frequency bias of the denoising model introduced by the current time step $t$ . In this two-step model we are only interested in $p$ and would like to neglect any gradients coming from $b$ . As it is unfeasible to correct for $b$ , we propose to cancel out effects of $b$ by approximating $b$ with $\hat{b}$ and applying $\hat{b}$ on $\mathbf{z}$ before comparing $\mathbf{z}$ with $\hat{\mathbf{z}}_t$ : + +$$ +\nabla_ {\boldsymbol {\theta}} \mathcal {L} _ {\mathrm {l m c}} = \left(\hat {b} (\mathbf {z}, t) - \hat {\mathbf {z}} _ {t}\right) \frac {\partial \mathbf {z} _ {t}}{\partial \boldsymbol {\theta}} \tag {9} +$$ + +In this formulation we omit both the Jacobian term from the denoising model (like in the original formulation, cf. Eq. (3)) as well as the Jacobian from $\hat{b}$ . Also, we empirically found that by dropping the weighting term $\frac{\sqrt{\alpha_t}}{\sqrt{1 - \alpha_t}}$ we can further improve visual fidelity – which is similar to the common practice of predicting the original image instead of the noise when learning the denoising model. As demonstrated in Fig. 1 (left), given an adequate model $\hat{b}$ , $\nabla_{\theta} \mathcal{L}_{\mathrm{Imc}}$ focuses on global changes, whereas $\nabla_{\theta} \mathcal{L}_{\mathrm{proj}}$ is dominated by high-frequent detail. We now explain how to obtain $\hat{b}$ . + +We propose to model $\hat{b}$ as a learnable function $\hat{b}_{\psi}(\mathbf{z},t)$ parameterized by $\psi$ . We can observe $b$ as a function of $t$ by sampling triplets $(\mathbf{z},\hat{\mathbf{z}}_t,t)$ , where $\mathbf{z}$ are random natural images assumed to be approximately on manifold images, i.e. $\mathbf{z} \approx \tilde{\mathbf{z}}$ . Using these triplets we can learn $\hat{b}_{\psi}$ by minimizing $\| \hat{\mathbf{z}}_t - \hat{b}_{\mathbf{z},t}\| _2^2$ , with $\hat{b}_{\mathbf{z},t} = \hat{b}_{\psi}(\mathbf{z},t)$ . Modeled with a deterministic model, $\hat{b}_{\psi}$ is expected to underfit the highly complex and probabilistic nature of $b$ . In fact, for a given pair $(\mathbf{z},t)$ at best it can learn the mean over all possible $\hat{\mathbf{z}}_t$ . This is desired, because even after approximating $b(\mathbf{z},t)$ through $\hat{b}_{\psi}(\mathbf{z},t)$ , we will be able to observe a + +gradient in $\mathcal{L}_{\mathrm{lmc}}$ : a gradient towards a specific instance of $\hat{\mathbf{z}}_t$ . Further, $\hat{b}_{\psi}$ only learns to "blur" images and neither generates new content nor is informed by the diffusion models' image manifold. Thus, it will perform similar operations for on-manifold and off-manifold images and $\mathcal{L}_{\mathrm{lmc}}$ may correct non-manifold images. However, naively applying a $\hat{b}_{\psi}$ in Eq. (9), learned that way, is causing an undesired effect, namely a drift in image statistics. Concretely we observed that images produced by $\hat{b}_{\psi}$ tend to underestimate the true image dynamics, especially for large $t$ where $\hat{\mathbf{z}}_t$ may be vastly different to $\mathbf{z}$ . By underestimating image dynamics, however, $\nabla_{\theta} \mathcal{L}_{\mathrm{lmc}}$ contains a gradient to "correct" for that and we have found images obtained via optimization using this gradient to be overly saturated and contrasting. To this end, we make $\hat{b}_{\psi}$ agnostic to the global image statistics by computing the training loss in a normalized space + +$$ +\mathcal {L} _ {\mathrm {k}} = \left\| \hat {\mathbf {z}} _ {t} - \left(\frac {\sigma \left(\hat {\mathbf {z}} _ {t}\right)}{\sigma \left(\hat {b} _ {\mathbf {z} , t}\right)} \left(\hat {b} _ {\mathbf {z}, t} - \mu \left(\hat {b} _ {\mathbf {z}, t}\right)\right) + \mu \left(\hat {b} _ {\mathbf {z}, t}\right)\right) \right\| _ {2} ^ {2}, \tag {10} +$$ + +with standard deviation $\sigma$ and mean $\mu$ . Predicting in a normalized space naturally requires to perform the same (normalisation) operation in Eq. (9), which we assume to be absorbed into $\hat{b}$ as follows. + +Increasing Sample Diversity. The original SDS formulation has been reported to have mode-seeking behaviour. Consequently, optimization results obtained via $\mathcal{L}_{\mathrm{SDS}}$ have the tendency to look very similar, even when based on different random seeds. We hypothesize this is due to optimizing over many steps and hence averaging the effects of varying noise $\epsilon$ . To prevent such averaging behaviour, we propose to fix $\epsilon$ over the course of optimization - which is in spirit similar to DDIM sampling [36]. Empirically, we found that optimizing with fixed $\epsilon$ is often unstable in the original SDS formulation. We hypothesize this is rooted in $\mathcal{L}_{\mathrm{proj}}$ not providing proper gradients towards the image manifold and at the same time $\mathcal{L}_{\mathrm{cond}}$ providing too strong gradients towards prompt agreement. The optimization state may leave the manifold of natural images and drift towards a state associated more to an "adversarial example" for the given prompt. While our loss formulation allows the use of a fixed $\epsilon$ , probably due to both the usage of smaller $\omega$ and better manifold correctives, we observe that we can further robustify the formulation by only fixing $\epsilon$ in $\mathcal{L}_{\mathrm{cond}}$ , while continue to sample $\epsilon$ in $\mathcal{L}_{\mathrm{lmc}}$ . This results in high quality, yet diverse, results, cf. Sec. 5. + +Implementation Details. We use an image diffusion model producing images of $128 \times 128$ pixel resolution and trained on internal data sources [34] as the pre-trained text-to-image diffusion model $\epsilon_{\phi}^{\omega}$ or $\mathbf{z}_{\phi}^{\omega}$ , respectively. We model $\hat{b}_{\psi}$ as a standard U-Net with four encoder layers with two Conv/ReLU/MaxPool blocks per layer, and skip connections to the decoder. The first layer contains 32 filters and the number of filters is doubled in each layer in the encoder and halved in the decoder. $\hat{b}_{\psi}$ has been trained until convergence, using $\mathcal{L}_{\mathrm{k}}$ as the only loss. For creating training triplets $(\mathbf{z},\hat{\mathbf{z}}_t,t)$ , we sampled random images $\mathbf{z}$ from OpenImages [18], random time-steps $t \sim \mathcal{U}(0.02,0.98)$ , and passed both into $\mathbf{z}_{\phi}$ obtaining $\hat{\mathbf{z}}_t$ . We train $\hat{b}_{\psi}$ once and keep it fixed for all experiments. + +![](images/f7dbffefb133f9df63588e53364cf03df4fc4cb598085ab761b4f1e89e6c5dc5.jpg) +Fig. 2: Optimization-based image synthesis results. We optimize an empty image to match a given prompt using our LMC-SDS, the original SDS loss, and $\mathcal{L}_{\mathrm{cond}}$ . SDS struggles to create detailed content when using low guidance $\omega$ . High $\omega$ produces detailed results but colors may be oversaturated (chimpanzee face), fake detail may appear (2nd mouse tail), or artifacts emerge. $\mathcal{L}_{\mathrm{cond}}$ is unstable to optimize and produces unrealistic colors. In contrast, our method produces detailed results with balanced colors. + +# 5 Experiments + +We evaluate our proposed LMC-SDS loss in a number of qualitative and quantitative experiments, including image synthesis, image editing, image-to-image translation network training, and 3D asset generation. + +Set-up. If not noted otherwise, we scale the total loss by $\frac{1}{\omega}$ to be able to ablate $\omega$ without changing the gradient magnitude. Fixed $\epsilon$ sampling is only used when explicitly stated. The pre-trained diffusion model we use in our experiments produces images of $128\times 128\mathrm{px}$ resolution. We enable arbitrary higher resolutions in our image experiments through the following strategy: We downsample the optimization state to $128\times 128\mathrm{px}$ for the first iterations. This ensures that the whole image is affected by the loss, but the signal may be blurry. Accounting for this, we sample random patches of arbitrary resolutions, from the image, for the remaining optimization steps. To prevent the loss from producing fake detail in sampled image patches, we lower the learning rate during this phase through Cosine decay. + +Baselines. This paper presents a generic novel loss function which uses an image diffusion model. To this end, we compare with the original SDS formulation and with related work focused on improving SDS. Specifically, we compare against DDS [8], NFSD [13], and multi-step SDS (MS-SDS) [40]. DDS, NFSD, and MS-SDS originally were proposed using a latent diffusion model [31] and operate in latent space. For fair comparisons, we adapted all methods to operate in pixel space and to use the same image diffusion model. Note that optimizing in the (ambient) pixel space is generally considered harder due to higher-dimensionality. Finally, we evaluate the effectiveness of our LMC-SDS loss for 3D asset generation using DreamFusion. Hereby we focus on the improvements of our loss and deliberately exclude recent improvements over DreamFusion rooted in other factors, e.g. novel-view synthesis networks or additional perceptual losses. While being important findings, those factors are orthogonal and complementary to our approach. + +Metrics. We use the following metrics for evaluating our methods and baselines: We report Learned Perceptual Image Patch Similarity (LPIPS $\downarrow$ ) [39] w.r.t. + +![](images/1d6092aeb8609f06ab3caf1cd0e7f721c37d0181de7d816454a7d2f4df65167a.jpg) + +![](images/d73dc3ff9c116bb001e1265c3faf4f3d5b9589d64196a0af7530469bc9af4424.jpg) +"a tiny dragon" + +![](images/2ded896864c0a31568640bd99067ac9f219526c503718d207b762403cac11a8f.jpg) +"ahoneybee" + +![](images/e6dc849adc76137711c827c0dca50a77f6e0c79ec8e2df46c4ef22fdcb51deaf.jpg) +"oil paiting of a landscape" +"bananas in a basket" + +![](images/27aaf75c6f56b59494840075b1a977a2614ccb209aba8ff75cd9e858bbd9215f.jpg) +"akangaroo" + +![](images/565bae8b01313f8243b2efe6281e6cf9bee7ee460f0b2e5ab8bf3bb04d3716f8.jpg) +"amarble bust of Caesar" + +![](images/c16e2a6890e18ac399300e3cd87ff18736a0b2490a9a871df13b132974434c06.jpg) +Fig. 3: Examples of optimization-based image editing results. We show pairs of input images (left) and editing result (right). +Input +Fig. 4: By fixing $\epsilon$ in $\mathcal{L}_{\mathrm{cond}}$ we can obtain diverse editing results. We show four variants of optimization-based editing results of the input image under the given prompt "a mountain during sunset". + +![](images/51ca10230060baa3e20691f7bddca0c52a8ad2028cb0be05056157e887e9459a.jpg) +seed = 1 + +![](images/f33e7d7377930a3685268d68b695aeb2c26820d595ed9b3eef2d161e21122c84.jpg) +seed $= 3$ + +![](images/383c261f0463ed4501e3ce7e2c83562b12642e98ced2904a7d81e92403a67d0b.jpg) +seed $= 4$ + +the source image, thus capturing how much an image was altered. Note that high LPIPS can be a result of either successful editing or image corruption. To disambiguate these, we additionally report the CLIP score $(\uparrow)$ [9] w.r.t. the target prompt. As CLIP can be insensitive, we additionally report CLIP-R-Precision $(\uparrow)$ [28], the retrieval accuracy using the CLIP model, where feasible. + +# 5.1 Image Synthesis & Editing + +We first explore our LMC-SDS loss for optimization-based image synthesis and image editing. + +Synthesis. To be able to evaluate and ablate the original SDS loss and our LMC-SDS in isolation (without additional interfering losses), we propose the task of optimization-based image synthesis. Instead of using the denoising schedule of a diffusion model, we initialize an image with $50\%$ grey and optimize its pixel values to match a given prompt, following the patching strategy described earlier. While it is not expected that images obtained using this strategy are on par with those obtained via the generation process, the results allow to draw meaningful conclusions about the properties of the loss functions. We show results of our loss and SDS in Fig. 2. We additionally include results obtained by only using $\mathcal{L}_{\mathrm{cond}}$ , ablating the influence of our proposed $\mathcal{L}_{\mathrm{lmc}}$ . Optimizing solely with $\mathcal{L}_{\mathrm{cond}}$ is unstable and the final images feature clearly visible defects, such as oversaturated colors or noisy patterns. The prompt is also not always well re + +![](images/bc0d8b288b67ae2f6b435606a31c9244b89f05562a29df3b2848945e444c1bb0.jpg) +Fig. 5: Quantitative results for optimization-based editing under varying $\omega$ . Left: Our method results in the highest CLIP score over all baselines for all $\omega$ . Right: We plot LPIPS over CLIP for further performance insights: DDS stays close to the original image (lowest LPIPS) by performing only small edits (low CLIP). SDS and MS-SDS respect the prompt better (higher CLIP), but corrupt the image (high LPIPS). NFSD corrupts the image less (lower LPIPS), but exhibits weak editing capabilities (low CLIP). Our method shows the strongest editing capabilities (highest CLIP), while staying close to the original structure. + +![](images/305999ff26a8456ced7cbf81e8ee6e5f8b7cf3e2b347c5e7bbdc631b0629c942.jpg) + +![](images/00911d2ae640ca2bf9d3a78c40ee89b5b2f08ac2d884acda9be80dc1c898df60.jpg) +Fig. 6: Comparison of image statistics before (Original) and after editing with $\omega = 5$ . We plot the average power spectra (left) and vertical derivative histograms (right). Ours, DDS, and NFSD preserve the image statistics well, while SDS-based methods introduce significant blur. + +![](images/af68a175d2317401efacf7ec6fce64dcaf4699a0c98d9443d0b39e0282850923.jpg) + +spected as e.g. in the case of the gibbon. The original SDS formulation produces blurry and desaturated images when using low guidance $\omega$ . High $\omega$ leads to better performance but artifacts are present, e.g. fake detail like the second mouse tail, oversaturated colors like in the chimpanzee's face, or blocky artifacts. In contrast, our method produces images of higher visual fidelity which respect the prompt well. Additionally, we empirically found that problems of the individual losses become more pronounced when optimizing longer (here we only optimize for 500 steps, cf. Sec. 5.2 for effects after 50K steps). + +Editing. Our loss can also be used for optimization-based image editing. We load existing natural images and optimize their pixel values directly. Optionally, we may constrain the optimization with an L2 loss w.r.t. the original pixel values. In Fig. 3 (more in Supp. Mat.) we show qualitative results of so edited images along with the corresponding target prompt. All results faithfully respect the target prompt while keeping the image structure intact. We can also vary the result by fixing $\epsilon$ over the course of the optimization, cf. Sec. 4. In Fig. 4, we show an input image along with four result variants. All results respect the target prompt well, while at the same time being significantly different from each other. + +For a numerical evaluation, we use an internal dataset of 1,000 images with text editing instructions in form of source and target prompts, similar to the dataset used in InstructPix2Pix [2]. We optimize images to match the target + +![](images/cc7936a0c758293b65df5dadd3d2a8ad515440d03153fac15ced5467624406d6.jpg) +Fig. 7: Results obtained with cats-to-others image translation networks trained with different losses. The network trained with SDS and NFSD fail to create detailed results. The DDS network produces rather uniform results, whereas the network trained with our LMC-SDS creates realistic and detailed images. + +
LPIPS ↓CLIP Score ↑CLIP-R-Precision ↑
DDS0.1430.2530.333
NFSD0.3320.2550.434
SDS0.4010.2560.355
SDS, ω = 1000.2010.2590.482
Ours0.1780.2610.592
+ +Table 1: Numerical results of the cats-to-others network training experiment. Note that LPIPS is computed w.r.t. the source image, meaning a low score may mean the image has not been edited enough. This is the case for DDS, as evident from low R-Precision. + +prompt solely using our LMC-SDS and baselines DDS, NFSD, SDS, and MS-SDS. For MS-SDS we chose to use five step denoising. We also experimented with more steps, but didn't observe significantly different performance. We report numerical results with varying $\omega$ (without loss rescaling) in Fig. 5. We found DDS to be least effective and producing results closest to the original image. We hypothesize that the negative gradient w.r.t. to the original image may not always be helpful. This could also explain a behavior reported by original authors, where editing results on real images are "inferior" over those on synthetic images. A possible reason being, that since for synthetic images the original prompts are describing the image well, the negative gradients are more helpful in that case. SDS can effectively edit the image, but also corrupts the image. This is also evident from the altered image statistics, see Fig. 6. NFSD corrupts the image less, but exhibits weak editing capabilities. In our experiments, MS-SDS performs very similar to SDS, corrupting the image in similar manner, but also being slightly less effective for editing. MS-SDS produces clear and detailed denoised images as part of the loss computation. We hypothesize that since denoised images and the current optimization state may differ drastically, and as denoised images vary considerably at each step, the computed gradients are not always very meaningful. The original paper proposes MS-SDS + +to be used with a view-conditioned diffusion model, where the denoised images probably vary much less and computed gradients may be more directed. Our loss respects the prompt best (highest CLIP), and finds a good balance between preserving the original structure and performing significant edits. + +To quantify the how much the result diversity can be increased through fixing $\epsilon$ , we re-ran the experiment using LMC-SDS ( $\omega = 30$ ) ten times with different seeds, both using the standard strategy and with fixing $\epsilon$ in $\mathcal{L}_{\mathrm{cond}}$ . We then computed the LPIPS (lower is more similar) of nine runs w.r.t. to the 10th (reference) run. The results using the standard strategy are all very similar with a per-run average of $0.041 \pm 0.007$ w.r.t. the reference run. In contrast, our new strategy produces much more diverse results: $0.158 \pm 0.018$ . + +# 5.2 Image-to-image Translation Network Training + +Optimization-based editing is not always an optimal choice, e.g. when building interactive editing tools where execution speed matters. Further, optimization can be unstable or perform better or worse, depending on the input images or target prompts. To this end, we explore to what extend LMC-SDS can be used to train a reliable and fast image-to-image translation network. Given a known distribution of images, we train a network to translate images from that distribution to an unseen target distribution, only described by a target caption. Inspired by the authors of DDS [8], we train a cats-to-others network, with "others" being bears, dogs, lions, and squirrels. The network transforms an image of a cat into an image of one of the other species. In DDS, the authors fine-tune a pre-trained diffusion model into a single-step feedforward model using images created by that same diffusion model. In contrast, we train a standard U-Net from scratch and use a dataset of real images, a much more practical and versatile approach. Concretely, we train using cat images from AFHQ [3] and evaluate on $5\%$ held out test images. We use the target prompt "a photo of a [class]" with $\omega = 15$ for all methods and the source prompt "a photo of a cat" for DDS. Moreover, we stabilize the training using an L2 loss on the original image and use Cosine scheduling to let the network first focus on reconstructing the original image and then on edits, similarly to the experiment in DDS. The class is passed via a frequency encoded class id concatenated to each pixel. We report numerical results in Tab. 1 and show qualitative examples in Fig. 7. Additional analysis is provided in the Suppl. Mat. SDS performs similarly to previous experiments for this task. While the target class is evident in the resulting images, the results feature strong deficiencies. Again DDS struggles with significant edits and stays close to the source. In contrast, the network trained with our LMC-SDS loss produces realistic and convincing edits. This is also evident from the significantly higher CLIP-R-Precision among all network variants. + +# 5.3 Text-to-3D Asset Generation + +In our final experiment, we demonstrate how LMC-SDS can improve text-to-3D models in the style of DreamFusion. In Sec. 3, we discussed how large guidance weights $\omega$ may affect the quality of results. To provide further evidence for this + +![](images/519d9e2b6fbed0e4548a1383d1ae7c738c897efadf3dd49418c64d405d424df8.jpg) + +![](images/489a39cdbd337c00259befbf8d224feb8653f9132242bd102f2cfada4ab6b047.jpg) +Fig. 8: The effect of increasing guidance weight $\omega$ in DreamFusion using the original SDS loss (even rows) and our proposed LMC-SDS (odd rows). Using larger $\omega$ saturates colors (chair, motorcycle, ship) and worsens the Janus problem (third wing and claw for the eagle). For low $\omega$ , vanilla DreamFusion fails to produce detailed geometry (ship, eagle) or geometry at all (motorcycle), while our approach performs well for $\omega = 20$ , often even for $\omega = 10$ . Notice how our approach constantly produces more detailed results for all $\omega$ . + +observation, we produce 3D assets with varying $\omega$ using the original DreamFusion formulation and a variant using our LMC-SDS. The results in Fig. 8 also show the trend of high $\omega$ having the tendency to produce oversaturated colors, albeit not as strongly as in the original paper, possibly due the use of a different image diffusion model. But we also observed an additional issue: the infamous Janus problem, the tendency to repeat detail in each view, becomes more severe with bigger $\omega$ , as evident from the eagle example. On the other hand, when using low $\omega$ in the original SDS loss, we observe that DreamFusion struggles to create detailed structure or even structure at all. In contrast, when using LMC-SDS, DreamFusion successfully produces detailed results, even when using up to only one tenth of the originally proposed $\omega$ . In Fig. 9 (more in Suppl. Mat.), we show additional side-by-side comparisons of the original SDS-based DreamFusion and our LMC-SDS version. Using our loss, the results are consistently sharper and feature far more detail. Finally in Fig. 10, we show that fixed $\epsilon$ sampling may also be used in 3D asset generation, to obtain diverse results. + +# 6 Discussion & Conclusion + +We proposed a novel loss formulation called LMC-SDS. LMC-SDS allows to effectively utilize text-to-image diffusion models as rich image priors in several optimization and network training tasks. We derived LMC-SDS by analyzing DreamFusion's SDS loss and by identifying an inherent problem in its formulation. Concretely, we expressed the SDS gradient as the sum of two terms. We identified that the second term $\mathcal{L}_{\mathrm{proj}}$ provides deficient gradients and consequently was weighted down in DreamFusion. We argued that this term, however, should provide gradients towards the learned manifold of natural images associated to the diffusion model. We also provided evidence that the lack thereof + +![](images/c1d0b6406303682d3d553725f0151343131c5d6569271a7909e8064cee2e251c.jpg) +Fig. 9: Results of DreamFusion using the original SDS loss (right) and our proposed LMC-SDS (left) for $\omega = 20$ . Our results are much sharper, contain more detail, and feature more realistic colors in all examples. + +![](images/da2d69e7c59d505d9c97f752e6ff124b46b4a62ff470de2e627258c1af0cd2eb.jpg) +Fig. 10: DreamFusion results using our LMC-SDS loss and fixed $\epsilon$ sampling for the prompts "a mannequin wearing a [hoodie, hat]. Note the diversity in colors, poses, and mannequin styles. + +makes the optimization unstable, thus negatively impacting the fidelity of results. To this end, we proposed to model the denoising models' time step dependent frequency bias. This allowed us to factor it out and obtain the cleaner gradient $\nabla_{\theta}\mathcal{L}_{\mathrm{lmc}}$ . In extensive experiments, we have demonstrated the effectiveness and versatility of our novel LMC-SDS loss. Furthermore, we discussed how LMC-SDS differs from other SDS-style losses and demonstrated how its performance is superior both qualitatively and quantitatively. + +Limitations. LMC-SDS finds its limits where either the signal from the diffusion model is too weak and ambiguous or when the optimization goes off the natural image manifold by too much. Concretely, LMC-SDS can only utilize whatever signal is coming from the diffusion model. This means, that when the diffusion model does not "understand" the prompt, LMC-SDS (and all other SDS-based approaches) will not be able to provide meaningful gradients. Further, whenever the current optimization state is too corrupted, LMC-SDS will not be able to correct for that and guide the optimization back to the manifold. Also, LMC-SDS sometimes overcompensates the frequency bias. Nevertheless, LMC-SDS is an important first step towards more stable and meaningful usage of the inherent image prior associated to image diffusion models. + +We proposed LMC-SDS, a novel SDS-based loss formulation with cleaner gradients towards the image manifold. In the future, we would like to continue to investigate how the manifold corrective can be further improved. We also want to leverage our findings in concrete applications like text-to-3D or in image editing and stylization tasks. + +# References + +1. Armandpour, M., Zheng, H., Sadeghian, A., Sadeghian, A., Zhou, M.: Re-imagine the negative prompt algorithm: Transform 2d diffusion into 3d, alleviate janus problem and beyond. arXiv preprint arXiv:2304.04968 (2023) +2. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023) +3. Choi, Y., Uh, Y., Yoo, J., Ha, J.W.: Stargan v2: Diverse image synthesis for multiple domains. In: IEEE Conf. Comput. Vis. Pattern Recog. (2020) +4. Chung, H., Kim, J., Mccann, M.T., Klasky, M.L., Ye, J.C.: Diffusion posterior sampling for general noisy inverse problems. In: Int. Conf. Learn. Represent. (2022) +5. Chung, H., Sim, B., Ryu, D., Ye, J.C.: Improving diffusion models for inverse problems using manifold constraints. Adv. Neural Inform. Process. Syst. 35, 25683-25696 (2022) +6. Croitoru, F.A., Hondru, V., Ionescu, R.T., Shah, M.: Diffusion models in vision: A survey. IEEE Trans. Pattern Anal. Mach. Intell. (2023) +7. Gao, S., Liu, X., Zeng, B., Xu, S., Li, Y., Luo, X., Liu, J., Zhen, X., Zhang, B.: Implicit diffusion models for continuous super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10021-10030 (2023) +8. Hertz, A., Aberman, K., Cohen-Or, D.: Delta denoising score. In: Int. Conf. Comput. Vis. (2023) +9. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718 (2021) +0. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Adv. Neural Inform. Process. Syst. 33, 6840-6851 (2020) +1. Ho, J., Salimans, T.: Classifier-free diffusion guidance. In: NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications (2021) +2. Jain, A., Xie, A., Abbeel, P.: Vectorfusion: Text-to-svq by abstracting pixel-based diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1911-1920 (2023) +3. Katzir, O., Patashnik, O., Cohen-Or, D., Lischinski, D.: Noise-free score distillation. arXiv preprint arXiv:2310.17590 (2023) +4. Kawar, B., Elad, M., Ermon, S., Song, J.: Denoising diffusion restoration models. In: Adv. Neural Inform. Process. Syst. (2022) +5. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 6007-6017 (2023) +6. Kim, S., Lee, K., Choi, J.S., Jeong, J., Sohn, K., Shin, J.: Collaborative score distillation for consistent visual synthesis. arXiv preprint arXiv:2307.04787 (2023) +7. Kolotouros, N., Alldieck, T., Zanfir, A., Bazavan, E.G., Fieraru, M., Sminchisescu, C.: Dreamhuman: Animatable 3d avatars from text. In: Adv. Neural Inform. Process. Syst. (2023) +8. Kuznetsova, A., Rom, H., Alldrin, N., Uijlings, J., Krasin, I., Pont-Tuset, J., Kamali, S., Popov, S., Malloci, M., Kolesnikov, A., Duerig, T., Ferrari, V.: The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. Int. J. Comput. Vis. (2020) +9. Li, X., Ren, Y., Jin, X., Lan, C., Wang, X., Zeng, W., Wang, X., Chen, Z.: Diffusion models for image restoration and enhancement-a comprehensive survey. arXiv preprint arXiv:2308.09388 (2023) + +20. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3d: High-resolution text-to-3d content creation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 300–309 (2023) +21. Lin, S., Liu, B., Li, J., Yang, X.: Common diffusion noise schedules and sample steps are flawed. arXiv preprint arXiv:2305.08891 (2023) +22. Liu, Q., Wang, D.: Stein variational gradient descent: A general purpose bayesian inference algorithm. Advances in neural information processing systems 29 (2016) +23. Liu, R., Wu, R., Van Hoorick, B., Tokmakov, P., Zakharov, S., Vondrick, C.: Zero-1-to-3: Zero-shot one image to 3d object. In: Int. Conf. Comput. Vis. pp. 9298-9309 (2023) +24. Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Van Gool, L.: Repaint: Inpainting using denoising diffusion probabilistic models. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 11461-11471 (2022) +25. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: SDEdit: Guided image synthesis and editing with stochastic differential equations. In: Int. Conf. Learn. Represent. (2022) +26. Metzer, G., Richardson, E., Patashnik, O., Giryes, R., Cohen-Or, D.: Latent-nerf for shape-guided generation of 3d shapes and textures. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 12663-12673 (2023) +27. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In: Int. Conf. on Mach. Learn. (2022) +28. Park, D.H., Azadi, S., Liu, X., Darrell, T., Rohrbach, A.: Benchmark for compositional text-to-image synthesis. In: Conf. on Neural Inf. Proc. Systems Datasets and Benchmarks Track (Round 1) (2021) +29. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. In: Int. Conf. Learn. Represent. (2022) +30. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022) +31. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 10684-10695 (2022) +32. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 22500-22510 (2023) +33. Sahara, C., Chan, W., Chang, H., Lee, C., Ho, J., Salimans, T., Fleet, D., Norouzi, M.: Palette: Image-to-image diffusion models. In: ACM SIGGRAPH 2022 Conference Proceedings. pp. 1-10 (2022) +34. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Adv. Neural Inform. Process. Syst. 35, 36479-36494 (2022) +35. Shi, Y., Wang, P., Ye, J., Mai, L., Li, K., Yang, X.: Mvdream: Multi-view diffusion for 3d generation. arXiv:2308.16512 (2023) +36. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. In: Int. Conf. Learn. Represent. (2020) +37. Wang, H., Du, X., Li, J., Yeh, R.A., Shakhnarovich, G.: Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 12619-12629 (2023) + +38. Wang, Z., Lu, C., Wang, Y., Bao, F., Li, C., Su, H., Zhu, J.: Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213 (2023) +39. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: IEEE Conf. Comput. Vis. Pattern Recog. (2018) +40. Zhou, Z., Tulsiani, S.: Sparsefusion: Distilling view-conditioned diffusion for 3d reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12588-12597 (2023) +41. Zou, Z.X., Cheng, W., Cao, Y.P., Huang, S.S., Shan, Y., Zhang, S.H.: Sparse3d: Distilling multiview-consistent diffusion for object reconstruction from sparse views. arXiv preprint arXiv:2308.14078 (2023) \ No newline at end of file diff --git a/2024/Score Distillation Sampling with Learned Manifold Corrective/images.zip b/2024/Score Distillation Sampling with Learned Manifold Corrective/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9565c74cfc9b4b9328bb9ef0729280ad97c1d45e --- /dev/null +++ b/2024/Score Distillation Sampling with Learned Manifold Corrective/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8567cc10ac940353a1588e69b6c9c9becd98e1ac4da820e8d78125f2fd2b419a +size 562258 diff --git a/2024/Score Distillation Sampling with Learned Manifold Corrective/layout.json b/2024/Score Distillation Sampling with Learned Manifold Corrective/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7aec2d1bdca5251805b90170ad9e0142c6228325 --- /dev/null +++ b/2024/Score Distillation Sampling with Learned Manifold Corrective/layout.json @@ -0,0 +1,10941 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 193, + 111, + 420, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 193, + 111, + 420, + 146 + ], + "spans": [ + { + "bbox": [ + 193, + 111, + 420, + 146 + ], + "type": "text", + "content": "Score Distillation Sampling with Learned Manifold Corrective" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 159, + 168, + 455, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 168, + 455, + 180 + ], + "spans": [ + { + "bbox": [ + 159, + 168, + 455, + 180 + ], + "type": "text", + "content": "Thiemo Alldieck, Nikos Kolotouros, and Cristian Sminchisescu" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 270, + 190, + 343, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 270, + 190, + 343, + 201 + ], + "spans": [ + { + "bbox": [ + 270, + 190, + 343, + 201 + ], + "type": "text", + "content": "Google Research\\*" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 220, + 455, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 220, + 455, + 384 + ], + "spans": [ + { + "bbox": [ + 160, + 220, + 455, + 384 + ], + "type": "text", + "content": "Abstract. Score Distillation Sampling (SDS) is a recent but already widely popular method that relies on an image diffusion model to control optimization problems using text prompts. aIn this paper, we conduct an in-depth analysis of the SDS loss function, identify an inherent problem with its formulation, and propose a surprisingly easy but effective fix. Specifically, we decompose the loss into different factors and isolate the component responsible for noisy gradients. In the original formulation, high text guidance is used to account for the noise, leading to unwanted side effects such as oversaturation or repeated detail. Instead, we train a shallow network mimicking the timestep-dependent frequency bias of the image diffusion model in order to effectively factor it out. We demonstrate the versatility and the effectiveness of our novel loss formulation through qualitative and quantitative experiments, including optimization-based image synthesis and editing, zero-shot image translation network training, and text-to-3D synthesis." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 400, + 230, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 400, + 230, + 412 + ], + "spans": [ + { + "bbox": [ + 133, + 400, + 230, + 412 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 418, + 482, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 482, + 648 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 482, + 648 + ], + "type": "text", + "content": "Image diffusion models [10] have recently become the de facto standard for image generation. Especially text-to-image models [27,30,31,34] have emerged as powerful tools for high fidelity, diverse, image synthesis. By being controlled through natural language, these models are extremely easy to use and thus open up a wide range of creative applications, without the need for special training. Beyond image synthesis, image diffusion models have been successfully deployed in applications like image restoration, inpating, editing, super-resolution, or colorization [4,7,14,25], among others. Image diffusion models are typically obtained using very large training sets and thus inherently represent the distribution of natural images. While applications typically make use of the generation process of diffusion models, e.g. by inserting images into the process [24,25] or by altering the denoising function [4,5,14], relatively little research has been conducted on how diffusion models can be used as rich, general purpose image priors. Score Distillation Sampling (SDS) proposed in DreamFusion [29] is one of the few exceptions. The SDS loss formulation uses a pretrained text-to-image model [34] to measure how well a given text matches an observation. In DreamFusion this loss is being used to generate 3D assets from textual descriptions, an idea that was quickly adopted [17,20,23,26,32,37] and essentially established a new research direction: text-to-3D synthesis. While being mostly used in this context, the design" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 654, + 248, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 654, + 248, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 654, + 248, + 665 + ], + "type": "text", + "content": "* now at Google DeepMind" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 294, + 168 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 294, + 168 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 294, + 168 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 294, + 168 + ], + "type": "image", + "image_path": "92fd2c0d7eb99c8b112827956c346c8f9d50ae16387f90e9f92d2ffd741bb748.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 168, + 294, + 236 + ], + "blocks": [ + { + "bbox": [ + 133, + 168, + 294, + 236 + ], + "lines": [ + { + "bbox": [ + 133, + 168, + 294, + 236 + ], + "spans": [ + { + "bbox": [ + 133, + 168, + 294, + 236 + ], + "type": "image", + "image_path": "89828352ee95179d799d5a86f7af7c80ca6479f57dce5d81a46b1c5d5d1eb690.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 239, + 482, + 294 + ], + "lines": [ + { + "bbox": [ + 130, + 239, + 482, + 294 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 482, + 294 + ], + "type": "text", + "content": "Fig. 1: Left: Visualization of SDS and LMC-SDS gradients w.r.t. pixel values and estimated denoised images for the given image " + }, + { + "bbox": [ + 130, + 239, + 482, + 294 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 130, + 239, + 482, + 294 + ], + "type": "text", + "content": ", the prompt " + }, + { + "bbox": [ + 130, + 239, + 482, + 294 + ], + "type": "inline_equation", + "content": "\\mathbf{y} = \\text{\"autumn\"}" + }, + { + "bbox": [ + 130, + 239, + 482, + 294 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 239, + 482, + 294 + ], + "type": "inline_equation", + "content": "t = 0.5" + }, + { + "bbox": [ + 130, + 239, + 482, + 294 + ], + "type": "text", + "content": ". We visualize the negative gradient, i.e. the direction of change. Right: Power spectra of denoised images " + }, + { + "bbox": [ + 130, + 239, + 482, + 294 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_t" + }, + { + "bbox": [ + 130, + 239, + 482, + 294 + ], + "type": "text", + "content": " for varying time-step " + }, + { + "bbox": [ + 130, + 239, + 482, + 294 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 239, + 482, + 294 + ], + "type": "text", + "content": " compared to the power spectrum of natural images " + }, + { + "bbox": [ + 130, + 239, + 482, + 294 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 130, + 239, + 482, + 294 + ], + "type": "text", + "content": ". See Sec. 3 for details." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 302, + 118, + 482, + 224 + ], + "blocks": [ + { + "bbox": [ + 302, + 118, + 482, + 224 + ], + "lines": [ + { + "bbox": [ + 302, + 118, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 302, + 118, + 482, + 224 + ], + "type": "image", + "image_path": "318684887935c2babf608634a96a64b2ebee988815453627f685e267600e3a20.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 394, + 227, + 429, + 236 + ], + "lines": [ + { + "bbox": [ + 394, + 227, + 429, + 236 + ], + "spans": [ + { + "bbox": [ + 394, + 227, + 429, + 236 + ], + "type": "text", + "content": "Frequency" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 304, + 482, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 482, + 520 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 482, + 520 + ], + "type": "text", + "content": "of the SDS loss is by no means only tailored to text-to-3D applications. In fact, SDS is an image loss and can be used in much wider contexts [8,12]. However, in its original formulation the SDS loss may degrade the observation, be too eager to match the text prompt, or provide meaningless gradients which inject noise into the objective. In this paper, we conduct an extensive analysis of the SDS loss, identify an inherent problem with its formulation, propose a surprisingly easy but effective fix, and demonstrate its effectiveness in several applications, including optimization-based image synthesis and editing, zero-shot training of image translation networks, and text-to-3D synthesis. Concretely, our new loss formulation - Score Distillation Sampling with Learned Manifold Corrective, or LMC-SDS for short - aims to provide better gradients along the direction of the learned manifold of real images. We present evidence that gradients towards the learned manifold are extremely noisy in SDS and that high text guidance is needed to compensate for the noisy signal. Further, we show that high guidance - or lack thereof - is a possible explanation for the artifacts observed with SDS, cf. Sec. 3. In contrast, applications relying on our novel formulation benefit from meaningful manifold gradients, may use lower text guidance, and produce results of overall higher visual fidelity." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 521, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 581 + ], + "type": "text", + "content": "In summary, our contributions are: (1) conducting an in-depth analysis of the SDS loss, (2) identifying an inherent problem with the formulation, and (3) proposing a novel LMC-SDS loss, providing cleaner gradients and eliminating the need for high guidance. (4) In extensive experiments, we demonstrate the effectiveness and versatility of the new LMC-SDS loss." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 590, + 237, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 590, + 237, + 601 + ], + "spans": [ + { + "bbox": [ + 131, + 590, + 237, + 601 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 605, + 481, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 628 + ], + "type": "text", + "content": "Methods that use image diffusion models as priors for tasks other than image synthesis can be broadly categorized into two lines of work." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "The first class of methods leverage image diffusion networks to solve specific tasks by relying on the diffusion sampling process. We discuss here a number of representative methods and refer the reader to [6, 19] for detailed discussions." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "Alldieck et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 283 + ], + "type": "text", + "content": "DDRM [14] tackles linear inverse problems such as deblurring, super-resolution, colorization, or inpainting by defining a modified variational posterior distribution conditioned on the observation signal, which can then be used for sampling. DPS [4] focuses on general inverse problems and proposes to augment the score prediction at each denoising step with an additional term that models the likelihood of the observations. MCG [5] builds on top and improves the fidelity of generation by introducing additional manifold constraints. SDEdit [25] uses diffusion models for image editing by first adding noise to the input image and then denoising it using the text-to-image diffusion prior. RePaint [24] uses unconditional image diffusion models as a prior for solving image inpainting. All these methods use pre-trained diffusion models. In contrast, methods like Palette [33] or Imagic [15] require training a diffusion model from scratch or fine-tuning a pretrained one, respectively. This line of work exclusively deals with image editing problems and is typically restricted to the resolution of the diffusion models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 284, + 482, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 284, + 482, + 571 + ], + "spans": [ + { + "bbox": [ + 130, + 284, + 482, + 571 + ], + "type": "text", + "content": "More related to our method is the second category: This body of work uses pre-trained diffusion models within general purpose loss functions in iterative optimization settings. Score Jacobian Chaining [37] propagates the score of a text-to-image diffusion model through a differentiable renderer to supervise the generation of 3D models. DreamFusion [29] takes an orthogonal approach and proposes the Score Distillation Sampling loss. SDS is a modification of the diffusion training objective that does not require expensive backpropagation through the diffusion model and encourages the alignment of the generated image with a conditioning signal. DDS [8] proposes an improved version of SDS, specifically designed for image editing and aims to reduce the artifacts introduced by noisy SDS gradients. In similar spirit and in concurrent work, NFSD [13] aims to reduce noise in the gradients by adding negative conditioning. In contrast, our method neither requires a source prompt nor a negative prompt. SparseFusion [40] introduces a multi-step SDS loss that performs multiple denoising steps instead of one, which adds significant computational overhead. ProlificDreamer [38] proposes a generalization of SDS from point estimates to distributions, but unlike our approach requires computationally expensive fine-tuning of a diffusion model during optimization. Collaborative Score Distillation [16] generalizes SDS to multiple samples and is inspired by Stein variational gradient descent [22]. Category Score Distillation Sampling (C-SDS) focuses on 3D generation and replaces the noise estimation error in SDS with a difference between the noise predictions of a standard and a multi-view consistent diffusion model [41]. Unlike our method, C-SDS is only applicable to 3D generation tasks. In contrast, we propose a general purpose loss function with very little computational overhead." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 585, + 205, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 585, + 205, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 585, + 205, + 599 + ], + "type": "text", + "content": "3 Analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "An image diffusion model is a generative model that is trained to reverse the diffusion process that transforms a natural image into pure noise [10]. An image is formed by iteratively removing small amounts of Gaussian noise over a fixed variance schedule " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " until an image is created. Text-to-image models are additionally conditioned on text to steer the denoising process in a direction matching" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 90, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 90, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 403, + 90, + 447, + 101 + ], + "type": "text", + "content": "LMC-SDS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 162 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 162 + ], + "type": "text", + "content": "the textural description. For a given image " + }, + { + "bbox": [ + 130, + 116, + 479, + 162 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 130, + 116, + 479, + 162 + ], + "type": "text", + "content": ", its textual description " + }, + { + "bbox": [ + 130, + 116, + 479, + 162 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 130, + 116, + 479, + 162 + ], + "type": "text", + "content": ", a randomly sampled timestep " + }, + { + "bbox": [ + 130, + 116, + 479, + 162 + ], + "type": "inline_equation", + "content": "t\\sim \\mathcal{U}(0,1)" + }, + { + "bbox": [ + 130, + 116, + 479, + 162 + ], + "type": "text", + "content": ", and random Gaussian noise " + }, + { + "bbox": [ + 130, + 116, + 479, + 162 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 130, + 116, + 479, + 162 + ], + "type": "text", + "content": ", a denoising model " + }, + { + "bbox": [ + 130, + 116, + 479, + 162 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}" + }, + { + "bbox": [ + 130, + 116, + 479, + 162 + ], + "type": "text", + "content": " parameterized by " + }, + { + "bbox": [ + 130, + 116, + 479, + 162 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 116, + 479, + 162 + ], + "type": "text", + "content": " can be trained by minimizing the diffusion loss" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 239, + 163, + 481, + 178 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 163, + 481, + 178 + ], + "spans": [ + { + "bbox": [ + 239, + 163, + 481, + 178 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {d i f f}} = w (t) \\left\\| \\epsilon_ {\\phi} ^ {\\omega} (\\mathbf {z} _ {t}, \\mathbf {y}, t) - \\epsilon \\right\\| _ {2} ^ {2}, \\tag {1}", + "image_path": "62b25c0305aaa350684ba8f8d611b2ebafba39b3c1a3bf6b42333f10f1f241d7.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 183, + 479, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 183, + 479, + 243 + ], + "spans": [ + { + "bbox": [ + 131, + 183, + 479, + 243 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 183, + 479, + 243 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t = \\sqrt{\\bar{\\alpha}_t}\\mathbf{z} + \\sqrt{1 - \\bar{\\alpha}_t}\\epsilon" + }, + { + "bbox": [ + 131, + 183, + 479, + 243 + ], + "type": "text", + "content": " refers to the noisy version of " + }, + { + "bbox": [ + 131, + 183, + 479, + 243 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 131, + 183, + 479, + 243 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 183, + 479, + 243 + ], + "type": "inline_equation", + "content": "w(t)" + }, + { + "bbox": [ + 131, + 183, + 479, + 243 + ], + "type": "text", + "content": " is a weighting function, omitted in the sequel. DreamFusion [29] showed that given a pre-trained model " + }, + { + "bbox": [ + 131, + 183, + 479, + 243 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}" + }, + { + "bbox": [ + 131, + 183, + 479, + 243 + ], + "type": "text", + "content": " the diffusion loss can be utilized for optimization problems. For an arbitrary differentiable rendering function returning " + }, + { + "bbox": [ + 131, + 183, + 479, + 243 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 131, + 183, + 479, + 243 + ], + "type": "text", + "content": " and parameterized by " + }, + { + "bbox": [ + 131, + 183, + 479, + 243 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}" + }, + { + "bbox": [ + 131, + 183, + 479, + 243 + ], + "type": "text", + "content": ", the gradient of the denoising function w.r.t. " + }, + { + "bbox": [ + 131, + 183, + 479, + 243 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}" + }, + { + "bbox": [ + 131, + 183, + 479, + 243 + ], + "type": "text", + "content": " is given by" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 212, + 253, + 481, + 278 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 253, + 481, + 278 + ], + "spans": [ + { + "bbox": [ + 212, + 253, + 481, + 278 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\boldsymbol {\\theta}} \\mathcal {L} _ {\\text {d i f f}} = \\left(\\boldsymbol {\\epsilon} _ {\\phi} ^ {\\omega} (\\mathbf {z} _ {t}, \\boldsymbol {y}, t) - \\boldsymbol {\\epsilon}\\right) \\frac {\\partial \\boldsymbol {\\epsilon} _ {\\phi} ^ {\\omega} (\\mathbf {z} , \\boldsymbol {y} , t)}{\\partial \\mathbf {z} _ {t}} \\frac {\\partial \\mathbf {z} _ {t}}{\\partial \\boldsymbol {\\theta}}. \\tag {2}", + "image_path": "9abb215294d23a23b7cbcc2b12c2cb6ab5e3a089040354b63e0d6a35dadb473d.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 286, + 479, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 286, + 479, + 310 + ], + "spans": [ + { + "bbox": [ + 131, + 286, + 479, + 310 + ], + "type": "text", + "content": "In practice, the Jacobian term is omitted to avoid backpropagating through the denoising model and the gradient is approximated as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 236, + 319, + 481, + 342 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 319, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 236, + 319, + 481, + 342 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\boldsymbol {\\theta}} \\mathcal {L} _ {\\mathrm {S D S}} = \\left(\\boldsymbol {\\epsilon} _ {\\phi} ^ {\\omega} (\\mathbf {z} _ {t}, \\boldsymbol {y}, t) - \\boldsymbol {\\epsilon}\\right) \\frac {\\partial \\mathbf {z} _ {t}}{\\partial \\boldsymbol {\\theta}}, \\tag {3}", + "image_path": "da9b7a209f2a52359dc6622f27afc90123e0991310fbc93a1e7bf9438f95cc91.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 350, + 479, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 479, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 479, + 373 + ], + "type": "text", + "content": "resulting in the gradient of original SDS loss. We will now rewrite Eq. (3) and conduct an analysis of the derived components." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 374, + 479, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 374, + 479, + 409 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 479, + 409 + ], + "type": "text", + "content": "Using classifier-free guidance [11], the predicted noise is the sum of the " + }, + { + "bbox": [ + 130, + 374, + 479, + 409 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 130, + 374, + 479, + 409 + ], + "type": "text", + "content": "-conditioned and the unconditioned noise prediction weighted by the guidance weight " + }, + { + "bbox": [ + 130, + 374, + 479, + 409 + ], + "type": "inline_equation", + "content": "\\omega" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 210, + 411, + 480, + 423 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 210, + 411, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 210, + 411, + 480, + 423 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\epsilon} _ {\\phi} ^ {\\omega} \\left(\\mathbf {z} _ {t}, \\boldsymbol {y}, t\\right) = \\omega \\boldsymbol {\\epsilon} _ {\\phi} \\left(\\mathbf {z} _ {t}, \\boldsymbol {y}, t\\right) + (1 - \\omega) \\boldsymbol {\\epsilon} _ {\\phi} \\left(\\mathbf {z} _ {t}, t\\right), \\tag {4}", + "image_path": "07f112a7097206f6e8e3da835459b12a05bdd617e5db1ccf5505d07c61516a09.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 428, + 239, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 428, + 239, + 438 + ], + "spans": [ + { + "bbox": [ + 131, + 428, + 239, + 438 + ], + "type": "text", + "content": "which we can rewrite as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 197, + 450, + 480, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 450, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 197, + 450, + 480, + 464 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\epsilon} _ {\\phi} ^ {\\omega} \\left(\\mathbf {z} _ {t}, \\boldsymbol {y}, t\\right) = \\omega \\left(\\boldsymbol {\\epsilon} _ {\\phi} \\left(\\mathbf {z} _ {t}, \\boldsymbol {y}, t\\right) - \\boldsymbol {\\epsilon} _ {\\phi} \\left(\\mathbf {z} _ {t}, t\\right)\\right) + \\boldsymbol {\\epsilon} _ {\\phi} \\left(\\mathbf {z} _ {t}, t\\right). \\tag {5}", + "image_path": "bea98d24da83857f80707d08e0f903c8bef7d8f836a6ba612ce7ea9e12c1c628.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 472, + 396, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 472, + 396, + 484 + ], + "spans": [ + { + "bbox": [ + 131, + 472, + 396, + 484 + ], + "type": "text", + "content": "By inserting Eq. (5) into Eq. (3) we can express " + }, + { + "bbox": [ + 131, + 472, + 396, + 484 + ], + "type": "inline_equation", + "content": "\\nabla_{\\pmb{\\theta}}\\mathcal{L}_{\\mathrm{SDS}}" + }, + { + "bbox": [ + 131, + 472, + 396, + 484 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 240, + 495, + 480, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 240, + 495, + 480, + 506 + ], + "spans": [ + { + "bbox": [ + 240, + 495, + 480, + 506 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\boldsymbol {\\theta}} \\mathcal {L} _ {\\mathrm {S D S}} = \\nabla_ {\\boldsymbol {\\theta}} \\mathcal {L} _ {\\text {c o n d}} + \\nabla_ {\\boldsymbol {\\theta}} \\mathcal {L} _ {\\text {p r o j}} \\tag {6a}", + "image_path": "b8d364983e5bfec9605c03f3583c8f6364493b08fc2cc8a47061f6d018773069.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 217, + 516, + 480, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 516, + 480, + 539 + ], + "spans": [ + { + "bbox": [ + 217, + 516, + 480, + 539 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\boldsymbol {\\theta}} \\mathcal {L} _ {\\text {c o n d}} = \\omega \\left(\\epsilon_ {\\phi} \\left(\\mathbf {z} _ {t}, \\boldsymbol {y}, t\\right) - \\epsilon_ {\\phi} \\left(\\mathbf {z} _ {t}, t\\right)\\right) \\frac {\\partial \\mathbf {z} _ {t}}{\\partial \\boldsymbol {\\theta}} \\tag {6b}", + "image_path": "db89a7b9733d009fb91e6b916b3f8e8cca10cb5dd17bad6636ea49d0bc25aad6.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 242, + 544, + 480, + 566 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 544, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 242, + 544, + 480, + 566 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\boldsymbol {\\theta}} \\mathcal {L} _ {\\text {p r o j}} = \\left(\\epsilon_ {\\phi} \\left(\\mathbf {z} _ {t}, t\\right) - \\boldsymbol {\\epsilon}\\right) \\frac {\\partial \\mathbf {z} _ {t}}{\\partial \\boldsymbol {\\theta}}. \\tag {6c}", + "image_path": "4e0886ebaedc55c868f52a7df72a4701a1d56c3d959067186a7252f5b619b835.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "text", + "content": "The two loss components can be interpreted as follows: " + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cond}}" + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "text", + "content": " maximizes the agreement of the image with the text prompt by providing gradients towards images formed through conditioning on " + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{proj}}" + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "text", + "content": " performs a single denoising step and provides gradients informing how the image " + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "text", + "content": " (and thereby parameters " + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}" + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "text", + "content": ") should change such that the image can be better denoised. As the denoising function was trained to minimize the expected squared error between the denoised and the original data, " + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\nabla_{\\pmb{\\theta}}\\mathcal{L}_{\\mathrm{proj}}" + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "text", + "content": " can be understood as the gradient direction towards the manifold of natural images." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "type": "text", + "content": "Aldieck et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 345 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 345 + ], + "type": "text", + "content": "DreamFusion [29] originally proposed to set " + }, + { + "bbox": [ + 130, + 115, + 482, + 345 + ], + "type": "inline_equation", + "content": "\\omega = 100" + }, + { + "bbox": [ + 130, + 115, + 482, + 345 + ], + "type": "text", + "content": ". This means that, following our derivation, " + }, + { + "bbox": [ + 130, + 115, + 482, + 345 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cond}}" + }, + { + "bbox": [ + 130, + 115, + 482, + 345 + ], + "type": "text", + "content": " is weighted proportionally much higher than " + }, + { + "bbox": [ + 130, + 115, + 482, + 345 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{proj}}" + }, + { + "bbox": [ + 130, + 115, + 482, + 345 + ], + "type": "text", + "content": ". This can be problematic as large guidance weights have been identified to be an important cause of over-exposure in diffusion models [21]. In fact, DreamFusion produces detailed results but also has a tendency towards unrealistically saturated colors, see also Fig. 8. On the other hand, setting " + }, + { + "bbox": [ + 130, + 115, + 482, + 345 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 130, + 115, + 482, + 345 + ], + "type": "text", + "content": " to lower values has been reported to result in very blurry images [8], a behavior that we also demonstrate in Sec. 5. Both behaviors, (1) the tendency to produce overly saturated results when using high guidance as well as (2) blurry results for low guidance, can be explained by looking at the components of " + }, + { + "bbox": [ + 130, + 115, + 482, + 345 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{SDS}}" + }, + { + "bbox": [ + 130, + 115, + 482, + 345 + ], + "type": "text", + "content": " in isolation. To this end, we consider the case where the rendering function is a no-op and simply returns pixel values, i.e. " + }, + { + "bbox": [ + 130, + 115, + 482, + 345 + ], + "type": "inline_equation", + "content": "\\mathbf{z} = \\boldsymbol{\\theta}" + }, + { + "bbox": [ + 130, + 115, + 482, + 345 + ], + "type": "text", + "content": ". In Fig. 1 (left), we visualize " + }, + { + "bbox": [ + 130, + 115, + 482, + 345 + ], + "type": "inline_equation", + "content": "\\nabla_{\\boldsymbol{\\theta}}\\mathcal{L}_{\\mathrm{cond}}" + }, + { + "bbox": [ + 130, + 115, + 482, + 345 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 115, + 482, + 345 + ], + "type": "inline_equation", + "content": "\\nabla_{\\boldsymbol{\\theta}}\\mathcal{L}_{\\mathrm{proj}}" + }, + { + "bbox": [ + 130, + 115, + 482, + 345 + ], + "type": "text", + "content": " for this case. " + }, + { + "bbox": [ + 130, + 115, + 482, + 345 + ], + "type": "inline_equation", + "content": "\\nabla_{\\boldsymbol{\\theta}}\\mathcal{L}_{\\mathrm{cond}}" + }, + { + "bbox": [ + 130, + 115, + 482, + 345 + ], + "type": "text", + "content": " returns the direction towards an image with altered colors. However, as " + }, + { + "bbox": [ + 130, + 115, + 482, + 345 + ], + "type": "inline_equation", + "content": "\\nabla_{\\boldsymbol{\\theta}}\\mathcal{L}_{\\mathrm{cond}}" + }, + { + "bbox": [ + 130, + 115, + 482, + 345 + ], + "type": "text", + "content": " is not anchored in the manifold of natural images, the gradient may eventually point away from that manifold. " + }, + { + "bbox": [ + 130, + 115, + 482, + 345 + ], + "type": "inline_equation", + "content": "\\nabla_{\\boldsymbol{\\theta}}\\mathcal{L}_{\\mathrm{cond}}" + }, + { + "bbox": [ + 130, + 115, + 482, + 345 + ], + "type": "text", + "content": " will always point towards changes that \"better\" respect the prompt and eventually produce saturated colors or other artifacts (cf. Sec. 5). " + }, + { + "bbox": [ + 130, + 115, + 482, + 345 + ], + "type": "inline_equation", + "content": "\\nabla_{\\boldsymbol{\\theta}}\\mathcal{L}_{\\mathrm{proj}}" + }, + { + "bbox": [ + 130, + 115, + 482, + 345 + ], + "type": "text", + "content": ", on the other hand, incorrectly marks high-frequency detail originating from " + }, + { + "bbox": [ + 130, + 115, + 482, + 345 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 130, + 115, + 482, + 345 + ], + "type": "text", + "content": " for removal - the reason for blurriness. This behavior can be understood when rewriting " + }, + { + "bbox": [ + 130, + 115, + 482, + 345 + ], + "type": "inline_equation", + "content": "\\nabla_{\\boldsymbol{\\theta}}\\mathcal{L}_{\\mathrm{proj}}" + }, + { + "bbox": [ + 130, + 115, + 482, + 345 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 214, + 354, + 481, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 354, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 214, + 354, + 481, + 380 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\pmb {\\theta}} \\mathcal {L} _ {\\mathrm {p r o j}} = \\left(\\frac {\\sqrt {\\alpha_ {t}}}{\\sqrt {1 - \\alpha_ {t}}} \\left(\\mathbf {z} - \\mathbf {z} _ {\\phi} (\\mathbf {z} _ {t}, t)\\right)\\right) \\frac {\\partial \\mathbf {z} _ {t}}{\\partial \\pmb {\\theta}}, \\tag {7}", + "image_path": "fbf06502e0cc2c13f3e06bed2e33efaa2ecb6f6f4dea1ea649e606c1c74402be.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 389, + 482, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 389, + 482, + 484 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 482, + 484 + ], + "type": "text", + "content": "which is equivalent to comparing the predicted and true noise. By looking at Eq. (7), we can now see a possible explanation why " + }, + { + "bbox": [ + 130, + 389, + 482, + 484 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{proj}}" + }, + { + "bbox": [ + 130, + 389, + 482, + 484 + ], + "type": "text", + "content": " might not be optimal: " + }, + { + "bbox": [ + 130, + 389, + 482, + 484 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_t = \\mathbf{z}_\\phi (\\mathbf{z}_t,t)" + }, + { + "bbox": [ + 130, + 389, + 482, + 484 + ], + "type": "text", + "content": " is an approximation of " + }, + { + "bbox": [ + 130, + 389, + 482, + 484 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 130, + 389, + 482, + 484 + ], + "type": "text", + "content": " and most probably not a perfect replica of " + }, + { + "bbox": [ + 130, + 389, + 482, + 484 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 130, + 389, + 482, + 484 + ], + "type": "text", + "content": ", even when " + }, + { + "bbox": [ + 130, + 389, + 482, + 484 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 130, + 389, + 482, + 484 + ], + "type": "text", + "content": " already lies on the natural image manifold, see Fig. 1 (left). This is especially the case for large values of " + }, + { + "bbox": [ + 130, + 389, + 482, + 484 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 389, + 482, + 484 + ], + "type": "text", + "content": ", where the denoising model has to reconstruct the true image from almost pure noise. More formally, we can observe a frequency bias in " + }, + { + "bbox": [ + 130, + 389, + 482, + 484 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_\\phi" + }, + { + "bbox": [ + 130, + 389, + 482, + 484 + ], + "type": "text", + "content": " dependent on " + }, + { + "bbox": [ + 130, + 389, + 482, + 484 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 389, + 482, + 484 + ], + "type": "text", + "content": " when comparing images " + }, + { + "bbox": [ + 130, + 389, + 482, + 484 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 130, + 389, + 482, + 484 + ], + "type": "text", + "content": " with their denoised counterparts " + }, + { + "bbox": [ + 130, + 389, + 482, + 484 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_t" + }, + { + "bbox": [ + 130, + 389, + 482, + 484 + ], + "type": "text", + "content": ", see Fig. 1 (right)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 485, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 482, + 568 + ], + "type": "text", + "content": "We conclude the following: (1) The reason for oversaturated colors and other artifacts is due to " + }, + { + "bbox": [ + 130, + 485, + 482, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cond}}" + }, + { + "bbox": [ + 130, + 485, + 482, + 568 + ], + "type": "text", + "content": " being \"over-eager\" to maximize the agreement of the image with the conditioning variable. (2) " + }, + { + "bbox": [ + 130, + 485, + 482, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{proj}}" + }, + { + "bbox": [ + 130, + 485, + 482, + 568 + ], + "type": "text", + "content": " provides deficient gradients ultimately removing high-frequency detail instead of \"correcting\" the gradient towards on-manifold solutions due to a frequency bias in " + }, + { + "bbox": [ + 130, + 485, + 482, + 568 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{\\phi}" + }, + { + "bbox": [ + 130, + 485, + 482, + 568 + ], + "type": "text", + "content": ". Thus, " + }, + { + "bbox": [ + 130, + 485, + 482, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{proj}}" + }, + { + "bbox": [ + 130, + 485, + 482, + 568 + ], + "type": "text", + "content": " appears to be a non-optimal choice and we provide an alternative in the sequel. But first we use our decomposition to discuss other approaches to \"fix\" the SDS loss." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "DDS [8] proposes to use the difference between two SDS losses for image editing, one computed on the current state w.r.t. the target prompt and one computed on the initial image w.r.t. a textual description of that image. By computing an SDS difference, DDS entirely removes " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{proj}}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " and introduces a negative gradient pointing away from the initial state. However, the need for an initial image is limiting possible use-cases. In concurrent work, NFSD [13] aims to obtain better results by replacing " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{proj}}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " with a " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cond}}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " variant using negative prompts for most timesteps, and a loss based on " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " for small " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": ". While the root" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 90, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 90, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 403, + 90, + 447, + 101 + ], + "type": "text", + "content": "LMC-SDS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "text", + "content": "for blurry results is removed by eliminating " + }, + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{proj}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "text", + "content": ", both DDS and NFSD do not anchor the optimization along the learned manifold. Other methods aim for a better noise prediction, where by improving " + }, + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}(\\mathbf{z},t)" + }, + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "text", + "content": ", one may also improve " + }, + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{proj}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{proj}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "text", + "content": " is the difference between the predicted and true noise). E.g. SparseFusion [40] introduces multi-step SDS. Hereby, instead of denoising in one step, up to 50 steps are used. While this may produce more informative gradients in " + }, + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{proj}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "text", + "content": ", this is also significantly more expensive. Through different means but towards the same goal, the concurrent work ProlificDreamer [38] learns a low-rank adaptation of the text-to-image model. While this approach is effective, it also adds the overhead of fine-tuning an additional diffusion model. Finally, some recent methods aim to alleviate artifacts by augmenting the original SDS formulation with advanced schedules for sampling " + }, + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "text", + "content": " or with negative prompts [1,35]. In contrast, we aim for a generally applicable solution." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 281, + 202, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 202, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 202, + 293 + ], + "type": "text", + "content": "4 Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 297, + 481, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 481, + 334 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 481, + 334 + ], + "type": "text", + "content": "We now present our solution to eliminating the frequency bias in " + }, + { + "bbox": [ + 130, + 297, + 481, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{proj}}" + }, + { + "bbox": [ + 130, + 297, + 481, + 334 + ], + "type": "text", + "content": " that we identified earlier. In an attempt to better understand the behavior of the denoising model given " + }, + { + "bbox": [ + 130, + 297, + 481, + 334 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 130, + 297, + 481, + 334 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 297, + 481, + 334 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 297, + 481, + 334 + ], + "type": "text", + "content": ", we model it as a two step process, as follows" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 240, + 342, + 480, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 240, + 342, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 240, + 342, + 480, + 354 + ], + "type": "interline_equation", + "content": "\\mathbf {z} _ {\\phi} \\left(\\mathbf {z} _ {t}, t\\right) = b \\circ p \\left(\\mathbf {z} _ {t}, t\\right) = b \\left(\\tilde {\\mathbf {z}}, t\\right), \\tag {8}", + "image_path": "75892a438928e7c0dec62bb4361e15d4d645fabbb327452f97c8d6ca04386eb1.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 362, + 482, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 362, + 482, + 435 + ], + "spans": [ + { + "bbox": [ + 130, + 362, + 482, + 435 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 362, + 482, + 435 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 362, + 482, + 435 + ], + "type": "text", + "content": " projects " + }, + { + "bbox": [ + 130, + 362, + 482, + 435 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 130, + 362, + 482, + 435 + ], + "type": "text", + "content": " onto the learned natural image manifold, " + }, + { + "bbox": [ + 130, + 362, + 482, + 435 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{z}}" + }, + { + "bbox": [ + 130, + 362, + 482, + 435 + ], + "type": "text", + "content": " is the projected image, and " + }, + { + "bbox": [ + 130, + 362, + 482, + 435 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{z}} = \\mathbf{z}" + }, + { + "bbox": [ + 130, + 362, + 482, + 435 + ], + "type": "text", + "content": " for images on the manifold. " + }, + { + "bbox": [ + 130, + 362, + 482, + 435 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 130, + 362, + 482, + 435 + ], + "type": "text", + "content": " is the error or frequency bias of the denoising model introduced by the current time step " + }, + { + "bbox": [ + 130, + 362, + 482, + 435 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 362, + 482, + 435 + ], + "type": "text", + "content": ". In this two-step model we are only interested in " + }, + { + "bbox": [ + 130, + 362, + 482, + 435 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 362, + 482, + 435 + ], + "type": "text", + "content": " and would like to neglect any gradients coming from " + }, + { + "bbox": [ + 130, + 362, + 482, + 435 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 130, + 362, + 482, + 435 + ], + "type": "text", + "content": ". As it is unfeasible to correct for " + }, + { + "bbox": [ + 130, + 362, + 482, + 435 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 130, + 362, + 482, + 435 + ], + "type": "text", + "content": ", we propose to cancel out effects of " + }, + { + "bbox": [ + 130, + 362, + 482, + 435 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 130, + 362, + 482, + 435 + ], + "type": "text", + "content": " by approximating " + }, + { + "bbox": [ + 130, + 362, + 482, + 435 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 130, + 362, + 482, + 435 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 362, + 482, + 435 + ], + "type": "inline_equation", + "content": "\\hat{b}" + }, + { + "bbox": [ + 130, + 362, + 482, + 435 + ], + "type": "text", + "content": " and applying " + }, + { + "bbox": [ + 130, + 362, + 482, + 435 + ], + "type": "inline_equation", + "content": "\\hat{b}" + }, + { + "bbox": [ + 130, + 362, + 482, + 435 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 130, + 362, + 482, + 435 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 130, + 362, + 482, + 435 + ], + "type": "text", + "content": " before comparing " + }, + { + "bbox": [ + 130, + 362, + 482, + 435 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 130, + 362, + 482, + 435 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 362, + 482, + 435 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_t" + }, + { + "bbox": [ + 130, + 362, + 482, + 435 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 249, + 441, + 481, + 462 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 441, + 481, + 462 + ], + "spans": [ + { + "bbox": [ + 249, + 441, + 481, + 462 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\boldsymbol {\\theta}} \\mathcal {L} _ {\\mathrm {l m c}} = \\left(\\hat {b} (\\mathbf {z}, t) - \\hat {\\mathbf {z}} _ {t}\\right) \\frac {\\partial \\mathbf {z} _ {t}}{\\partial \\boldsymbol {\\theta}} \\tag {9}", + "image_path": "63501c8fbd0e670418686c78945bceff6f94c6e2ec391ceb050fd101635746cc.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 466, + 482, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 466, + 482, + 566 + ], + "spans": [ + { + "bbox": [ + 130, + 466, + 482, + 566 + ], + "type": "text", + "content": "In this formulation we omit both the Jacobian term from the denoising model (like in the original formulation, cf. Eq. (3)) as well as the Jacobian from " + }, + { + "bbox": [ + 130, + 466, + 482, + 566 + ], + "type": "inline_equation", + "content": "\\hat{b}" + }, + { + "bbox": [ + 130, + 466, + 482, + 566 + ], + "type": "text", + "content": ". Also, we empirically found that by dropping the weighting term " + }, + { + "bbox": [ + 130, + 466, + 482, + 566 + ], + "type": "inline_equation", + "content": "\\frac{\\sqrt{\\alpha_t}}{\\sqrt{1 - \\alpha_t}}" + }, + { + "bbox": [ + 130, + 466, + 482, + 566 + ], + "type": "text", + "content": " we can further improve visual fidelity – which is similar to the common practice of predicting the original image instead of the noise when learning the denoising model. As demonstrated in Fig. 1 (left), given an adequate model " + }, + { + "bbox": [ + 130, + 466, + 482, + 566 + ], + "type": "inline_equation", + "content": "\\hat{b}" + }, + { + "bbox": [ + 130, + 466, + 482, + 566 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 466, + 482, + 566 + ], + "type": "inline_equation", + "content": "\\nabla_{\\theta} \\mathcal{L}_{\\mathrm{Imc}}" + }, + { + "bbox": [ + 130, + 466, + 482, + 566 + ], + "type": "text", + "content": " focuses on global changes, whereas " + }, + { + "bbox": [ + 130, + 466, + 482, + 566 + ], + "type": "inline_equation", + "content": "\\nabla_{\\theta} \\mathcal{L}_{\\mathrm{proj}}" + }, + { + "bbox": [ + 130, + 466, + 482, + 566 + ], + "type": "text", + "content": " is dominated by high-frequent detail. We now explain how to obtain " + }, + { + "bbox": [ + 130, + 466, + 482, + 566 + ], + "type": "inline_equation", + "content": "\\hat{b}" + }, + { + "bbox": [ + 130, + 466, + 482, + 566 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 567, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 567, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 567, + 482, + 667 + ], + "type": "text", + "content": "We propose to model " + }, + { + "bbox": [ + 130, + 567, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\hat{b}" + }, + { + "bbox": [ + 130, + 567, + 482, + 667 + ], + "type": "text", + "content": " as a learnable function " + }, + { + "bbox": [ + 130, + 567, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\hat{b}_{\\psi}(\\mathbf{z},t)" + }, + { + "bbox": [ + 130, + 567, + 482, + 667 + ], + "type": "text", + "content": " parameterized by " + }, + { + "bbox": [ + 130, + 567, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 130, + 567, + 482, + 667 + ], + "type": "text", + "content": ". We can observe " + }, + { + "bbox": [ + 130, + 567, + 482, + 667 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 130, + 567, + 482, + 667 + ], + "type": "text", + "content": " as a function of " + }, + { + "bbox": [ + 130, + 567, + 482, + 667 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 567, + 482, + 667 + ], + "type": "text", + "content": " by sampling triplets " + }, + { + "bbox": [ + 130, + 567, + 482, + 667 + ], + "type": "inline_equation", + "content": "(\\mathbf{z},\\hat{\\mathbf{z}}_t,t)" + }, + { + "bbox": [ + 130, + 567, + 482, + 667 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 567, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 130, + 567, + 482, + 667 + ], + "type": "text", + "content": " are random natural images assumed to be approximately on manifold images, i.e. " + }, + { + "bbox": [ + 130, + 567, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{z} \\approx \\tilde{\\mathbf{z}}" + }, + { + "bbox": [ + 130, + 567, + 482, + 667 + ], + "type": "text", + "content": ". Using these triplets we can learn " + }, + { + "bbox": [ + 130, + 567, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\hat{b}_{\\psi}" + }, + { + "bbox": [ + 130, + 567, + 482, + 667 + ], + "type": "text", + "content": " by minimizing " + }, + { + "bbox": [ + 130, + 567, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\| \\hat{\\mathbf{z}}_t - \\hat{b}_{\\mathbf{z},t}\\| _2^2" + }, + { + "bbox": [ + 130, + 567, + 482, + 667 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 130, + 567, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\hat{b}_{\\mathbf{z},t} = \\hat{b}_{\\psi}(\\mathbf{z},t)" + }, + { + "bbox": [ + 130, + 567, + 482, + 667 + ], + "type": "text", + "content": ". Modeled with a deterministic model, " + }, + { + "bbox": [ + 130, + 567, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\hat{b}_{\\psi}" + }, + { + "bbox": [ + 130, + 567, + 482, + 667 + ], + "type": "text", + "content": " is expected to underfit the highly complex and probabilistic nature of " + }, + { + "bbox": [ + 130, + 567, + 482, + 667 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 130, + 567, + 482, + 667 + ], + "type": "text", + "content": ". In fact, for a given pair " + }, + { + "bbox": [ + 130, + 567, + 482, + 667 + ], + "type": "inline_equation", + "content": "(\\mathbf{z},t)" + }, + { + "bbox": [ + 130, + 567, + 482, + 667 + ], + "type": "text", + "content": " at best it can learn the mean over all possible " + }, + { + "bbox": [ + 130, + 567, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_t" + }, + { + "bbox": [ + 130, + 567, + 482, + 667 + ], + "type": "text", + "content": ". This is desired, because even after approximating " + }, + { + "bbox": [ + 130, + 567, + 482, + 667 + ], + "type": "inline_equation", + "content": "b(\\mathbf{z},t)" + }, + { + "bbox": [ + 130, + 567, + 482, + 667 + ], + "type": "text", + "content": " through " + }, + { + "bbox": [ + 130, + 567, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\hat{b}_{\\psi}(\\mathbf{z},t)" + }, + { + "bbox": [ + 130, + 567, + 482, + 667 + ], + "type": "text", + "content": ", we will be able to observe a" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "type": "text", + "content": "Aldieck et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 261 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 261 + ], + "type": "text", + "content": "gradient in " + }, + { + "bbox": [ + 130, + 115, + 482, + 261 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{lmc}}" + }, + { + "bbox": [ + 130, + 115, + 482, + 261 + ], + "type": "text", + "content": ": a gradient towards a specific instance of " + }, + { + "bbox": [ + 130, + 115, + 482, + 261 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_t" + }, + { + "bbox": [ + 130, + 115, + 482, + 261 + ], + "type": "text", + "content": ". Further, " + }, + { + "bbox": [ + 130, + 115, + 482, + 261 + ], + "type": "inline_equation", + "content": "\\hat{b}_{\\psi}" + }, + { + "bbox": [ + 130, + 115, + 482, + 261 + ], + "type": "text", + "content": " only learns to \"blur\" images and neither generates new content nor is informed by the diffusion models' image manifold. Thus, it will perform similar operations for on-manifold and off-manifold images and " + }, + { + "bbox": [ + 130, + 115, + 482, + 261 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{lmc}}" + }, + { + "bbox": [ + 130, + 115, + 482, + 261 + ], + "type": "text", + "content": " may correct non-manifold images. However, naively applying a " + }, + { + "bbox": [ + 130, + 115, + 482, + 261 + ], + "type": "inline_equation", + "content": "\\hat{b}_{\\psi}" + }, + { + "bbox": [ + 130, + 115, + 482, + 261 + ], + "type": "text", + "content": " in Eq. (9), learned that way, is causing an undesired effect, namely a drift in image statistics. Concretely we observed that images produced by " + }, + { + "bbox": [ + 130, + 115, + 482, + 261 + ], + "type": "inline_equation", + "content": "\\hat{b}_{\\psi}" + }, + { + "bbox": [ + 130, + 115, + 482, + 261 + ], + "type": "text", + "content": " tend to underestimate the true image dynamics, especially for large " + }, + { + "bbox": [ + 130, + 115, + 482, + 261 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 115, + 482, + 261 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 115, + 482, + 261 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_t" + }, + { + "bbox": [ + 130, + 115, + 482, + 261 + ], + "type": "text", + "content": " may be vastly different to " + }, + { + "bbox": [ + 130, + 115, + 482, + 261 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 130, + 115, + 482, + 261 + ], + "type": "text", + "content": ". By underestimating image dynamics, however, " + }, + { + "bbox": [ + 130, + 115, + 482, + 261 + ], + "type": "inline_equation", + "content": "\\nabla_{\\theta} \\mathcal{L}_{\\mathrm{lmc}}" + }, + { + "bbox": [ + 130, + 115, + 482, + 261 + ], + "type": "text", + "content": " contains a gradient to \"correct\" for that and we have found images obtained via optimization using this gradient to be overly saturated and contrasting. To this end, we make " + }, + { + "bbox": [ + 130, + 115, + 482, + 261 + ], + "type": "inline_equation", + "content": "\\hat{b}_{\\psi}" + }, + { + "bbox": [ + 130, + 115, + 482, + 261 + ], + "type": "text", + "content": " agnostic to the global image statistics by computing the training loss in a normalized space" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 205, + 271, + 481, + 301 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 271, + 481, + 301 + ], + "spans": [ + { + "bbox": [ + 205, + 271, + 481, + 301 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {k}} = \\left\\| \\hat {\\mathbf {z}} _ {t} - \\left(\\frac {\\sigma \\left(\\hat {\\mathbf {z}} _ {t}\\right)}{\\sigma \\left(\\hat {b} _ {\\mathbf {z} , t}\\right)} \\left(\\hat {b} _ {\\mathbf {z}, t} - \\mu \\left(\\hat {b} _ {\\mathbf {z}, t}\\right)\\right) + \\mu \\left(\\hat {b} _ {\\mathbf {z}, t}\\right)\\right) \\right\\| _ {2} ^ {2}, \\tag {10}", + "image_path": "baee1cc63bf04facd7ec5065e02d7e78514fe1ddc1ded39cb4cd3eb8bec6f29a.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 312, + 481, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 481, + 348 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 481, + 348 + ], + "type": "text", + "content": "with standard deviation " + }, + { + "bbox": [ + 130, + 312, + 481, + 348 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 312, + 481, + 348 + ], + "type": "text", + "content": " and mean " + }, + { + "bbox": [ + 130, + 312, + 481, + 348 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 130, + 312, + 481, + 348 + ], + "type": "text", + "content": ". Predicting in a normalized space naturally requires to perform the same (normalisation) operation in Eq. (9), which we assume to be absorbed into " + }, + { + "bbox": [ + 130, + 312, + 481, + 348 + ], + "type": "inline_equation", + "content": "\\hat{b}" + }, + { + "bbox": [ + 130, + 312, + 481, + 348 + ], + "type": "text", + "content": " as follows." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 350, + 482, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 482, + 542 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 482, + 542 + ], + "type": "text", + "content": "Increasing Sample Diversity. The original SDS formulation has been reported to have mode-seeking behaviour. Consequently, optimization results obtained via " + }, + { + "bbox": [ + 130, + 350, + 482, + 542 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{SDS}}" + }, + { + "bbox": [ + 130, + 350, + 482, + 542 + ], + "type": "text", + "content": " have the tendency to look very similar, even when based on different random seeds. We hypothesize this is due to optimizing over many steps and hence averaging the effects of varying noise " + }, + { + "bbox": [ + 130, + 350, + 482, + 542 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 350, + 482, + 542 + ], + "type": "text", + "content": ". To prevent such averaging behaviour, we propose to fix " + }, + { + "bbox": [ + 130, + 350, + 482, + 542 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 350, + 482, + 542 + ], + "type": "text", + "content": " over the course of optimization - which is in spirit similar to DDIM sampling [36]. Empirically, we found that optimizing with fixed " + }, + { + "bbox": [ + 130, + 350, + 482, + 542 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 350, + 482, + 542 + ], + "type": "text", + "content": " is often unstable in the original SDS formulation. We hypothesize this is rooted in " + }, + { + "bbox": [ + 130, + 350, + 482, + 542 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{proj}}" + }, + { + "bbox": [ + 130, + 350, + 482, + 542 + ], + "type": "text", + "content": " not providing proper gradients towards the image manifold and at the same time " + }, + { + "bbox": [ + 130, + 350, + 482, + 542 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cond}}" + }, + { + "bbox": [ + 130, + 350, + 482, + 542 + ], + "type": "text", + "content": " providing too strong gradients towards prompt agreement. The optimization state may leave the manifold of natural images and drift towards a state associated more to an \"adversarial example\" for the given prompt. While our loss formulation allows the use of a fixed " + }, + { + "bbox": [ + 130, + 350, + 482, + 542 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 350, + 482, + 542 + ], + "type": "text", + "content": ", probably due to both the usage of smaller " + }, + { + "bbox": [ + 130, + 350, + 482, + 542 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 130, + 350, + 482, + 542 + ], + "type": "text", + "content": " and better manifold correctives, we observe that we can further robustify the formulation by only fixing " + }, + { + "bbox": [ + 130, + 350, + 482, + 542 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 350, + 482, + 542 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 130, + 350, + 482, + 542 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cond}}" + }, + { + "bbox": [ + 130, + 350, + 482, + 542 + ], + "type": "text", + "content": ", while continue to sample " + }, + { + "bbox": [ + 130, + 350, + 482, + 542 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 350, + 482, + 542 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 130, + 350, + 482, + 542 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{lmc}}" + }, + { + "bbox": [ + 130, + 350, + 482, + 542 + ], + "type": "text", + "content": ". This results in high quality, yet diverse, results, cf. Sec. 5." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 544, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 482, + 666 + ], + "type": "text", + "content": "Implementation Details. We use an image diffusion model producing images of " + }, + { + "bbox": [ + 130, + 544, + 482, + 666 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 130, + 544, + 482, + 666 + ], + "type": "text", + "content": " pixel resolution and trained on internal data sources [34] as the pre-trained text-to-image diffusion model " + }, + { + "bbox": [ + 130, + 544, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}^{\\omega}" + }, + { + "bbox": [ + 130, + 544, + 482, + 666 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 130, + 544, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{\\phi}^{\\omega}" + }, + { + "bbox": [ + 130, + 544, + 482, + 666 + ], + "type": "text", + "content": ", respectively. We model " + }, + { + "bbox": [ + 130, + 544, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{b}_{\\psi}" + }, + { + "bbox": [ + 130, + 544, + 482, + 666 + ], + "type": "text", + "content": " as a standard U-Net with four encoder layers with two Conv/ReLU/MaxPool blocks per layer, and skip connections to the decoder. The first layer contains 32 filters and the number of filters is doubled in each layer in the encoder and halved in the decoder. " + }, + { + "bbox": [ + 130, + 544, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{b}_{\\psi}" + }, + { + "bbox": [ + 130, + 544, + 482, + 666 + ], + "type": "text", + "content": " has been trained until convergence, using " + }, + { + "bbox": [ + 130, + 544, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{k}}" + }, + { + "bbox": [ + 130, + 544, + 482, + 666 + ], + "type": "text", + "content": " as the only loss. For creating training triplets " + }, + { + "bbox": [ + 130, + 544, + 482, + 666 + ], + "type": "inline_equation", + "content": "(\\mathbf{z},\\hat{\\mathbf{z}}_t,t)" + }, + { + "bbox": [ + 130, + 544, + 482, + 666 + ], + "type": "text", + "content": ", we sampled random images " + }, + { + "bbox": [ + 130, + 544, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 130, + 544, + 482, + 666 + ], + "type": "text", + "content": " from OpenImages [18], random time-steps " + }, + { + "bbox": [ + 130, + 544, + 482, + 666 + ], + "type": "inline_equation", + "content": "t \\sim \\mathcal{U}(0.02,0.98)" + }, + { + "bbox": [ + 130, + 544, + 482, + 666 + ], + "type": "text", + "content": ", and passed both into " + }, + { + "bbox": [ + 130, + 544, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{\\phi}" + }, + { + "bbox": [ + 130, + 544, + 482, + 666 + ], + "type": "text", + "content": " obtaining " + }, + { + "bbox": [ + 130, + 544, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_t" + }, + { + "bbox": [ + 130, + 544, + 482, + 666 + ], + "type": "text", + "content": ". We train " + }, + { + "bbox": [ + 130, + 544, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{b}_{\\psi}" + }, + { + "bbox": [ + 130, + 544, + 482, + 666 + ], + "type": "text", + "content": " once and keep it fixed for all experiments." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 90, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 90, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 403, + 90, + 447, + 101 + ], + "type": "text", + "content": "LMC-SDS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 202 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 202 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 202 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 202 + ], + "type": "image", + "image_path": "f7dbffefb133f9df63588e53364cf03df4fc4cb598085ab761b4f1e89e6c5dc5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 204, + 482, + 270 + ], + "lines": [ + { + "bbox": [ + 130, + 204, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 204, + 482, + 270 + ], + "type": "text", + "content": "Fig. 2: Optimization-based image synthesis results. We optimize an empty image to match a given prompt using our LMC-SDS, the original SDS loss, and " + }, + { + "bbox": [ + 130, + 204, + 482, + 270 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cond}}" + }, + { + "bbox": [ + 130, + 204, + 482, + 270 + ], + "type": "text", + "content": ". SDS struggles to create detailed content when using low guidance " + }, + { + "bbox": [ + 130, + 204, + 482, + 270 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 130, + 204, + 482, + 270 + ], + "type": "text", + "content": ". High " + }, + { + "bbox": [ + 130, + 204, + 482, + 270 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 130, + 204, + 482, + 270 + ], + "type": "text", + "content": " produces detailed results but colors may be oversaturated (chimpanzee face), fake detail may appear (2nd mouse tail), or artifacts emerge. " + }, + { + "bbox": [ + 130, + 204, + 482, + 270 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cond}}" + }, + { + "bbox": [ + 130, + 204, + 482, + 270 + ], + "type": "text", + "content": " is unstable to optimize and produces unrealistic colors. In contrast, our method produces detailed results with balanced colors." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 285, + 230, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 230, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 230, + 300 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 304, + 481, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 481, + 341 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 481, + 341 + ], + "type": "text", + "content": "We evaluate our proposed LMC-SDS loss in a number of qualitative and quantitative experiments, including image synthesis, image editing, image-to-image translation network training, and 3D asset generation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 341, + 482, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 341, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 341, + 482, + 472 + ], + "type": "text", + "content": "Set-up. If not noted otherwise, we scale the total loss by " + }, + { + "bbox": [ + 130, + 341, + 482, + 472 + ], + "type": "inline_equation", + "content": "\\frac{1}{\\omega}" + }, + { + "bbox": [ + 130, + 341, + 482, + 472 + ], + "type": "text", + "content": " to be able to ablate " + }, + { + "bbox": [ + 130, + 341, + 482, + 472 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 130, + 341, + 482, + 472 + ], + "type": "text", + "content": " without changing the gradient magnitude. Fixed " + }, + { + "bbox": [ + 130, + 341, + 482, + 472 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 341, + 482, + 472 + ], + "type": "text", + "content": " sampling is only used when explicitly stated. The pre-trained diffusion model we use in our experiments produces images of " + }, + { + "bbox": [ + 130, + 341, + 482, + 472 + ], + "type": "inline_equation", + "content": "128\\times 128\\mathrm{px}" + }, + { + "bbox": [ + 130, + 341, + 482, + 472 + ], + "type": "text", + "content": " resolution. We enable arbitrary higher resolutions in our image experiments through the following strategy: We downsample the optimization state to " + }, + { + "bbox": [ + 130, + 341, + 482, + 472 + ], + "type": "inline_equation", + "content": "128\\times 128\\mathrm{px}" + }, + { + "bbox": [ + 130, + 341, + 482, + 472 + ], + "type": "text", + "content": " for the first iterations. This ensures that the whole image is affected by the loss, but the signal may be blurry. Accounting for this, we sample random patches of arbitrary resolutions, from the image, for the remaining optimization steps. To prevent the loss from producing fake detail in sampled image patches, we lower the learning rate during this phase through Cosine decay." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 474, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 640 + ], + "type": "text", + "content": "Baselines. This paper presents a generic novel loss function which uses an image diffusion model. To this end, we compare with the original SDS formulation and with related work focused on improving SDS. Specifically, we compare against DDS [8], NFSD [13], and multi-step SDS (MS-SDS) [40]. DDS, NFSD, and MS-SDS originally were proposed using a latent diffusion model [31] and operate in latent space. For fair comparisons, we adapted all methods to operate in pixel space and to use the same image diffusion model. Note that optimizing in the (ambient) pixel space is generally considered harder due to higher-dimensionality. Finally, we evaluate the effectiveness of our LMC-SDS loss for 3D asset generation using DreamFusion. Hereby we focus on the improvements of our loss and deliberately exclude recent improvements over DreamFusion rooted in other factors, e.g. novel-view synthesis networks or additional perceptual losses. While being important findings, those factors are orthogonal and complementary to our approach." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Metrics. We use the following metrics for evaluating our methods and baselines: We report Learned Perceptual Image Patch Similarity (LPIPS " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": ") [39] w.r.t." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "Alldieck et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 250, + 173 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 250, + 173 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 250, + 173 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 250, + 173 + ], + "type": "image", + "image_path": "1d6092aeb8609f06ab3caf1cd0e7f721c37d0181de7d816454a7d2f4df65167a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 252, + 114, + 365, + 173 + ], + "blocks": [ + { + "bbox": [ + 252, + 114, + 365, + 173 + ], + "lines": [ + { + "bbox": [ + 252, + 114, + 365, + 173 + ], + "spans": [ + { + "bbox": [ + 252, + 114, + 365, + 173 + ], + "type": "image", + "image_path": "d73dc3ff9c116bb001e1265c3faf4f3d5b9589d64196a0af7530469bc9af4424.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 286, + 174, + 327, + 180 + ], + "lines": [ + { + "bbox": [ + 286, + 174, + 327, + 180 + ], + "spans": [ + { + "bbox": [ + 286, + 174, + 327, + 180 + ], + "type": "text", + "content": "\"a tiny dragon\"" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 365, + 114, + 480, + 173 + ], + "blocks": [ + { + "bbox": [ + 365, + 114, + 480, + 173 + ], + "lines": [ + { + "bbox": [ + 365, + 114, + 480, + 173 + ], + "spans": [ + { + "bbox": [ + 365, + 114, + 480, + 173 + ], + "type": "image", + "image_path": "2ded896864c0a31568640bd99067ac9f219526c503718d207b762403cac11a8f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 174, + 441, + 180 + ], + "lines": [ + { + "bbox": [ + 403, + 174, + 441, + 180 + ], + "spans": [ + { + "bbox": [ + 403, + 174, + 441, + 180 + ], + "type": "text", + "content": "\"ahoneybee\"" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 133, + 181, + 250, + 239 + ], + "blocks": [ + { + "bbox": [ + 155, + 174, + 228, + 181 + ], + "lines": [ + { + "bbox": [ + 155, + 174, + 228, + 181 + ], + "spans": [ + { + "bbox": [ + 155, + 174, + 228, + 181 + ], + "type": "text", + "content": "\"oil paiting of a landscape\"" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 133, + 181, + 250, + 239 + ], + "lines": [ + { + "bbox": [ + 133, + 181, + 250, + 239 + ], + "spans": [ + { + "bbox": [ + 133, + 181, + 250, + 239 + ], + "type": "image", + "image_path": "e6dc849adc76137711c827c0dca50a77f6e0c79ec8e2df46c4ef22fdcb51deaf.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 162, + 240, + 219, + 247 + ], + "lines": [ + { + "bbox": [ + 162, + 240, + 219, + 247 + ], + "spans": [ + { + "bbox": [ + 162, + 240, + 219, + 247 + ], + "type": "text", + "content": "\"bananas in a basket\"" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 250, + 181, + 364, + 239 + ], + "blocks": [ + { + "bbox": [ + 250, + 181, + 364, + 239 + ], + "lines": [ + { + "bbox": [ + 250, + 181, + 364, + 239 + ], + "spans": [ + { + "bbox": [ + 250, + 181, + 364, + 239 + ], + "type": "image", + "image_path": "27aaf75c6f56b59494840075b1a977a2614ccb209aba8ff75cd9e858bbd9215f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 289, + 240, + 324, + 247 + ], + "lines": [ + { + "bbox": [ + 289, + 240, + 324, + 247 + ], + "spans": [ + { + "bbox": [ + 289, + 240, + 324, + 247 + ], + "type": "text", + "content": "\"akangaroo\"" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 364, + 182, + 480, + 239 + ], + "blocks": [ + { + "bbox": [ + 364, + 182, + 480, + 239 + ], + "lines": [ + { + "bbox": [ + 364, + 182, + 480, + 239 + ], + "spans": [ + { + "bbox": [ + 364, + 182, + 480, + 239 + ], + "type": "image", + "image_path": "565bae8b01313f8243b2efe6281e6cf9bee7ee460f0b2e5ab8bf3bb04d3716f8.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 386, + 240, + 457, + 247 + ], + "lines": [ + { + "bbox": [ + 386, + 240, + 457, + 247 + ], + "spans": [ + { + "bbox": [ + 386, + 240, + 457, + 247 + ], + "type": "text", + "content": "\"amarble bust of Caesar\"" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 133, + 276, + 202, + 346 + ], + "blocks": [ + { + "bbox": [ + 132, + 250, + 479, + 271 + ], + "lines": [ + { + "bbox": [ + 132, + 250, + 479, + 271 + ], + "spans": [ + { + "bbox": [ + 132, + 250, + 479, + 271 + ], + "type": "text", + "content": "Fig. 3: Examples of optimization-based image editing results. We show pairs of input images (left) and editing result (right)." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 133, + 276, + 202, + 346 + ], + "lines": [ + { + "bbox": [ + 133, + 276, + 202, + 346 + ], + "spans": [ + { + "bbox": [ + 133, + 276, + 202, + 346 + ], + "type": "image", + "image_path": "c16e2a6890e18ac399300e3cd87ff18736a0b2490a9a871df13b132974434c06.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 157, + 347, + 179, + 356 + ], + "lines": [ + { + "bbox": [ + 157, + 347, + 179, + 356 + ], + "spans": [ + { + "bbox": [ + 157, + 347, + 179, + 356 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 358, + 480, + 392 + ], + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 392 + ], + "type": "text", + "content": "Fig. 4: By fixing " + }, + { + "bbox": [ + 132, + 358, + 480, + 392 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 132, + 358, + 480, + 392 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 132, + 358, + 480, + 392 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cond}}" + }, + { + "bbox": [ + 132, + 358, + 480, + 392 + ], + "type": "text", + "content": " we can obtain diverse editing results. We show four variants of optimization-based editing results of the input image under the given prompt \"a mountain during sunset\"." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 203, + 276, + 340, + 346 + ], + "blocks": [ + { + "bbox": [ + 203, + 276, + 340, + 346 + ], + "lines": [ + { + "bbox": [ + 203, + 276, + 340, + 346 + ], + "spans": [ + { + "bbox": [ + 203, + 276, + 340, + 346 + ], + "type": "image", + "image_path": "51ca10230060baa3e20691f7bddca0c52a8ad2028cb0be05056157e887e9459a.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 347, + 253, + 355 + ], + "lines": [ + { + "bbox": [ + 220, + 347, + 253, + 355 + ], + "spans": [ + { + "bbox": [ + 220, + 347, + 253, + 355 + ], + "type": "text", + "content": "seed = 1" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 340, + 276, + 411, + 346 + ], + "blocks": [ + { + "bbox": [ + 340, + 276, + 411, + 346 + ], + "lines": [ + { + "bbox": [ + 340, + 276, + 411, + 346 + ], + "spans": [ + { + "bbox": [ + 340, + 276, + 411, + 346 + ], + "type": "image", + "image_path": "f33e7d7377930a3685268d68b695aeb2c26820d595ed9b3eef2d161e21122c84.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 359, + 347, + 393, + 356 + ], + "lines": [ + { + "bbox": [ + 359, + 347, + 393, + 356 + ], + "spans": [ + { + "bbox": [ + 359, + 347, + 393, + 356 + ], + "type": "text", + "content": "seed " + }, + { + "bbox": [ + 359, + 347, + 393, + 356 + ], + "type": "inline_equation", + "content": "= 3" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 411, + 276, + 480, + 346 + ], + "blocks": [ + { + "bbox": [ + 411, + 276, + 480, + 346 + ], + "lines": [ + { + "bbox": [ + 411, + 276, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 411, + 276, + 480, + 346 + ], + "type": "image", + "image_path": "383c261f0463ed4501e3ce7e2c83562b12642e98ced2904a7d81e92403a67d0b.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 427, + 347, + 463, + 356 + ], + "lines": [ + { + "bbox": [ + 427, + 347, + 463, + 356 + ], + "spans": [ + { + "bbox": [ + 427, + 347, + 463, + 356 + ], + "type": "text", + "content": "seed " + }, + { + "bbox": [ + 427, + 347, + 463, + 356 + ], + "type": "inline_equation", + "content": "= 4" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "bbox": [ + 130, + 403, + 480, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 480, + 464 + ], + "type": "text", + "content": "the source image, thus capturing how much an image was altered. Note that high LPIPS can be a result of either successful editing or image corruption. To disambiguate these, we additionally report the CLIP score " + }, + { + "bbox": [ + 130, + 403, + 480, + 464 + ], + "type": "inline_equation", + "content": "(\\uparrow)" + }, + { + "bbox": [ + 130, + 403, + 480, + 464 + ], + "type": "text", + "content": " [9] w.r.t. the target prompt. As CLIP can be insensitive, we additionally report CLIP-R-Precision " + }, + { + "bbox": [ + 130, + 403, + 480, + 464 + ], + "type": "inline_equation", + "content": "(\\uparrow)" + }, + { + "bbox": [ + 130, + 403, + 480, + 464 + ], + "type": "text", + "content": " [28], the retrieval accuracy using the CLIP model, where feasible." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 132, + 482, + 296, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 482, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 132, + 482, + 296, + 495 + ], + "type": "text", + "content": "5.1 Image Synthesis & Editing" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 130, + 497, + 479, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 479, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 479, + 521 + ], + "type": "text", + "content": "We first explore our LMC-SDS loss for optimization-based image synthesis and image editing." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "type": "text", + "content": "Synthesis. To be able to evaluate and ablate the original SDS loss and our LMC-SDS in isolation (without additional interfering losses), we propose the task of optimization-based image synthesis. Instead of using the denoising schedule of a diffusion model, we initialize an image with " + }, + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "type": "text", + "content": " grey and optimize its pixel values to match a given prompt, following the patching strategy described earlier. While it is not expected that images obtained using this strategy are on par with those obtained via the generation process, the results allow to draw meaningful conclusions about the properties of the loss functions. We show results of our loss and SDS in Fig. 2. We additionally include results obtained by only using " + }, + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cond}}" + }, + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "type": "text", + "content": ", ablating the influence of our proposed " + }, + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{lmc}}" + }, + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "type": "text", + "content": ". Optimizing solely with " + }, + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cond}}" + }, + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "type": "text", + "content": " is unstable and the final images feature clearly visible defects, such as oversaturated colors or noisy patterns. The prompt is also not always well re" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "type": "text", + "content": "LMC-SDS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 278, + 187 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 278, + 187 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 278, + 187 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 278, + 187 + ], + "type": "image", + "image_path": "bc0d8b288b67ae2f6b435606a31c9244b89f05562a29df3b2848945e444c1bb0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 190, + 482, + 278 + ], + "lines": [ + { + "bbox": [ + 130, + 190, + 482, + 278 + ], + "spans": [ + { + "bbox": [ + 130, + 190, + 482, + 278 + ], + "type": "text", + "content": "Fig. 5: Quantitative results for optimization-based editing under varying " + }, + { + "bbox": [ + 130, + 190, + 482, + 278 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 130, + 190, + 482, + 278 + ], + "type": "text", + "content": ". Left: Our method results in the highest CLIP score over all baselines for all " + }, + { + "bbox": [ + 130, + 190, + 482, + 278 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 130, + 190, + 482, + 278 + ], + "type": "text", + "content": ". Right: We plot LPIPS over CLIP for further performance insights: DDS stays close to the original image (lowest LPIPS) by performing only small edits (low CLIP). SDS and MS-SDS respect the prompt better (higher CLIP), but corrupt the image (high LPIPS). NFSD corrupts the image less (lower LPIPS), but exhibits weak editing capabilities (low CLIP). Our method shows the strongest editing capabilities (highest CLIP), while staying close to the original structure." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 285, + 113, + 478, + 188 + ], + "blocks": [ + { + "bbox": [ + 285, + 113, + 478, + 188 + ], + "lines": [ + { + "bbox": [ + 285, + 113, + 478, + 188 + ], + "spans": [ + { + "bbox": [ + 285, + 113, + 478, + 188 + ], + "type": "image", + "image_path": "305999ff26a8456ced7cbf81e8ee6e5f8b7cf3e2b347c5e7bbdc631b0629c942.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 175, + 282, + 304, + 364 + ], + "blocks": [ + { + "bbox": [ + 175, + 282, + 304, + 364 + ], + "lines": [ + { + "bbox": [ + 175, + 282, + 304, + 364 + ], + "spans": [ + { + "bbox": [ + 175, + 282, + 304, + 364 + ], + "type": "image", + "image_path": "00911d2ae640ca2bf9d3a78c40ee89b5b2f08ac2d884acda9be80dc1c898df60.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 366, + 482, + 410 + ], + "lines": [ + { + "bbox": [ + 130, + 366, + 482, + 410 + ], + "spans": [ + { + "bbox": [ + 130, + 366, + 482, + 410 + ], + "type": "text", + "content": "Fig. 6: Comparison of image statistics before (Original) and after editing with " + }, + { + "bbox": [ + 130, + 366, + 482, + 410 + ], + "type": "inline_equation", + "content": "\\omega = 5" + }, + { + "bbox": [ + 130, + 366, + 482, + 410 + ], + "type": "text", + "content": ". We plot the average power spectra (left) and vertical derivative histograms (right). Ours, DDS, and NFSD preserve the image statistics well, while SDS-based methods introduce significant blur." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 307, + 281, + 438, + 365 + ], + "blocks": [ + { + "bbox": [ + 307, + 281, + 438, + 365 + ], + "lines": [ + { + "bbox": [ + 307, + 281, + 438, + 365 + ], + "spans": [ + { + "bbox": [ + 307, + 281, + 438, + 365 + ], + "type": "image", + "image_path": "af68a175d2317401efacf7ec6fce64dcaf4699a0c98d9443d0b39e0282850923.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 423, + 482, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 482, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 482, + 518 + ], + "type": "text", + "content": "spected as e.g. in the case of the gibbon. The original SDS formulation produces blurry and desaturated images when using low guidance " + }, + { + "bbox": [ + 130, + 423, + 482, + 518 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 130, + 423, + 482, + 518 + ], + "type": "text", + "content": ". High " + }, + { + "bbox": [ + 130, + 423, + 482, + 518 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 130, + 423, + 482, + 518 + ], + "type": "text", + "content": " leads to better performance but artifacts are present, e.g. fake detail like the second mouse tail, oversaturated colors like in the chimpanzee's face, or blocky artifacts. In contrast, our method produces images of higher visual fidelity which respect the prompt well. Additionally, we empirically found that problems of the individual losses become more pronounced when optimizing longer (here we only optimize for 500 steps, cf. Sec. 5.2 for effects after 50K steps)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "type": "text", + "content": "Editing. Our loss can also be used for optimization-based image editing. We load existing natural images and optimize their pixel values directly. Optionally, we may constrain the optimization with an L2 loss w.r.t. the original pixel values. In Fig. 3 (more in Supp. Mat.) we show qualitative results of so edited images along with the corresponding target prompt. All results faithfully respect the target prompt while keeping the image structure intact. We can also vary the result by fixing " + }, + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "type": "text", + "content": " over the course of the optimization, cf. Sec. 4. In Fig. 4, we show an input image along with four result variants. All results respect the target prompt well, while at the same time being significantly different from each other." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "For a numerical evaluation, we use an internal dataset of 1,000 images with text editing instructions in form of source and target prompts, similar to the dataset used in InstructPix2Pix [2]. We optimize images to match the target" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "Aldieck et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 482, + 258 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 258 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 258 + ], + "type": "image", + "image_path": "cc7936a0c758293b65df5dadd3d2a8ad515440d03153fac15ced5467624406d6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 260, + 482, + 304 + ], + "lines": [ + { + "bbox": [ + 130, + 260, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 260, + 482, + 304 + ], + "type": "text", + "content": "Fig. 7: Results obtained with cats-to-others image translation networks trained with different losses. The network trained with SDS and NFSD fail to create detailed results. The DDS network produces rather uniform results, whereas the network trained with our LMC-SDS creates realistic and detailed images." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 187, + 313, + 425, + 381 + ], + "blocks": [ + { + "bbox": [ + 187, + 313, + 425, + 381 + ], + "lines": [ + { + "bbox": [ + 187, + 313, + 425, + 381 + ], + "spans": [ + { + "bbox": [ + 187, + 313, + 425, + 381 + ], + "type": "table", + "html": "
LPIPS ↓CLIP Score ↑CLIP-R-Precision ↑
DDS0.1430.2530.333
NFSD0.3320.2550.434
SDS0.4010.2560.355
SDS, ω = 1000.2010.2590.482
Ours0.1780.2610.592
", + "image_path": "74dabe669053e10d97d71989757fa2bf781a92ddafc4ae2a0991c419a994ad11.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 384, + 482, + 427 + ], + "lines": [ + { + "bbox": [ + 130, + 384, + 482, + 427 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 482, + 427 + ], + "type": "text", + "content": "Table 1: Numerical results of the cats-to-others network training experiment. Note that LPIPS is computed w.r.t. the source image, meaning a low score may mean the image has not been edited enough. This is the case for DDS, as evident from low R-Precision." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "type": "text", + "content": "prompt solely using our LMC-SDS and baselines DDS, NFSD, SDS, and MS-SDS. For MS-SDS we chose to use five step denoising. We also experimented with more steps, but didn't observe significantly different performance. We report numerical results with varying " + }, + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "type": "text", + "content": " (without loss rescaling) in Fig. 5. We found DDS to be least effective and producing results closest to the original image. We hypothesize that the negative gradient w.r.t. to the original image may not always be helpful. This could also explain a behavior reported by original authors, where editing results on real images are \"inferior\" over those on synthetic images. A possible reason being, that since for synthetic images the original prompts are describing the image well, the negative gradients are more helpful in that case. SDS can effectively edit the image, but also corrupts the image. This is also evident from the altered image statistics, see Fig. 6. NFSD corrupts the image less, but exhibits weak editing capabilities. In our experiments, MS-SDS performs very similar to SDS, corrupting the image in similar manner, but also being slightly less effective for editing. MS-SDS produces clear and detailed denoised images as part of the loss computation. We hypothesize that since denoised images and the current optimization state may differ drastically, and as denoised images vary considerably at each step, the computed gradients are not always very meaningful. The original paper proposes MS-SDS" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "type": "text", + "content": "LMC-SDS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "to be used with a view-conditioned diffusion model, where the denoised images probably vary much less and computed gradients may be more directed. Our loss respects the prompt best (highest CLIP), and finds a good balance between preserving the original structure and performing significant edits." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 482, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 482, + 247 + ], + "type": "text", + "content": "To quantify the how much the result diversity can be increased through fixing " + }, + { + "bbox": [ + 130, + 164, + 482, + 247 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 164, + 482, + 247 + ], + "type": "text", + "content": ", we re-ran the experiment using LMC-SDS (" + }, + { + "bbox": [ + 130, + 164, + 482, + 247 + ], + "type": "inline_equation", + "content": "\\omega = 30" + }, + { + "bbox": [ + 130, + 164, + 482, + 247 + ], + "type": "text", + "content": ") ten times with different seeds, both using the standard strategy and with fixing " + }, + { + "bbox": [ + 130, + 164, + 482, + 247 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 164, + 482, + 247 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 130, + 164, + 482, + 247 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cond}}" + }, + { + "bbox": [ + 130, + 164, + 482, + 247 + ], + "type": "text", + "content": ". We then computed the LPIPS (lower is more similar) of nine runs w.r.t. to the 10th (reference) run. The results using the standard strategy are all very similar with a per-run average of " + }, + { + "bbox": [ + 130, + 164, + 482, + 247 + ], + "type": "inline_equation", + "content": "0.041 \\pm 0.007" + }, + { + "bbox": [ + 130, + 164, + 482, + 247 + ], + "type": "text", + "content": " w.r.t. the reference run. In contrast, our new strategy produces much more diverse results: " + }, + { + "bbox": [ + 130, + 164, + 482, + 247 + ], + "type": "inline_equation", + "content": "0.158 \\pm 0.018" + }, + { + "bbox": [ + 130, + 164, + 482, + 247 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 262, + 392, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 262, + 392, + 274 + ], + "spans": [ + { + "bbox": [ + 131, + 262, + 392, + 274 + ], + "type": "text", + "content": "5.2 Image-to-image Translation Network Training" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 277, + 482, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 482, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 482, + 601 + ], + "type": "text", + "content": "Optimization-based editing is not always an optimal choice, e.g. when building interactive editing tools where execution speed matters. Further, optimization can be unstable or perform better or worse, depending on the input images or target prompts. To this end, we explore to what extend LMC-SDS can be used to train a reliable and fast image-to-image translation network. Given a known distribution of images, we train a network to translate images from that distribution to an unseen target distribution, only described by a target caption. Inspired by the authors of DDS [8], we train a cats-to-others network, with \"others\" being bears, dogs, lions, and squirrels. The network transforms an image of a cat into an image of one of the other species. In DDS, the authors fine-tune a pre-trained diffusion model into a single-step feedforward model using images created by that same diffusion model. In contrast, we train a standard U-Net from scratch and use a dataset of real images, a much more practical and versatile approach. Concretely, we train using cat images from AFHQ [3] and evaluate on " + }, + { + "bbox": [ + 130, + 277, + 482, + 601 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 130, + 277, + 482, + 601 + ], + "type": "text", + "content": " held out test images. We use the target prompt \"a photo of a [class]\" with " + }, + { + "bbox": [ + 130, + 277, + 482, + 601 + ], + "type": "inline_equation", + "content": "\\omega = 15" + }, + { + "bbox": [ + 130, + 277, + 482, + 601 + ], + "type": "text", + "content": " for all methods and the source prompt \"a photo of a cat\" for DDS. Moreover, we stabilize the training using an L2 loss on the original image and use Cosine scheduling to let the network first focus on reconstructing the original image and then on edits, similarly to the experiment in DDS. The class is passed via a frequency encoded class id concatenated to each pixel. We report numerical results in Tab. 1 and show qualitative examples in Fig. 7. Additional analysis is provided in the Suppl. Mat. SDS performs similarly to previous experiments for this task. While the target class is evident in the resulting images, the results feature strong deficiencies. Again DDS struggles with significant edits and stays close to the source. In contrast, the network trained with our LMC-SDS loss produces realistic and convincing edits. This is also evident from the significantly higher CLIP-R-Precision among all network variants." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 614, + 308, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 614, + 308, + 625 + ], + "spans": [ + { + "bbox": [ + 131, + 614, + 308, + 625 + ], + "type": "text", + "content": "5.3 Text-to-3D Asset Generation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "In our final experiment, we demonstrate how LMC-SDS can improve text-to-3D models in the style of DreamFusion. In Sec. 3, we discussed how large guidance weights " + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": " may affect the quality of results. To provide further evidence for this" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "Alldieck et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 130, + 114, + 479, + 179 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 479, + 179 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 479, + 179 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 479, + 179 + ], + "type": "image", + "image_path": "519d9e2b6fbed0e4548a1383d1ae7c738c897efadf3dd49418c64d405d424df8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 130, + 182, + 480, + 256 + ], + "blocks": [ + { + "bbox": [ + 130, + 182, + 480, + 256 + ], + "lines": [ + { + "bbox": [ + 130, + 182, + 480, + 256 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 480, + 256 + ], + "type": "image", + "image_path": "489a39cdbd337c00259befbf8d224feb8653f9132242bd102f2cfada4ab6b047.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 257, + 481, + 333 + ], + "lines": [ + { + "bbox": [ + 130, + 257, + 481, + 333 + ], + "spans": [ + { + "bbox": [ + 130, + 257, + 481, + 333 + ], + "type": "text", + "content": "Fig. 8: The effect of increasing guidance weight " + }, + { + "bbox": [ + 130, + 257, + 481, + 333 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 130, + 257, + 481, + 333 + ], + "type": "text", + "content": " in DreamFusion using the original SDS loss (even rows) and our proposed LMC-SDS (odd rows). Using larger " + }, + { + "bbox": [ + 130, + 257, + 481, + 333 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 130, + 257, + 481, + 333 + ], + "type": "text", + "content": " saturates colors (chair, motorcycle, ship) and worsens the Janus problem (third wing and claw for the eagle). For low " + }, + { + "bbox": [ + 130, + 257, + 481, + 333 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 130, + 257, + 481, + 333 + ], + "type": "text", + "content": ", vanilla DreamFusion fails to produce detailed geometry (ship, eagle) or geometry at all (motorcycle), while our approach performs well for " + }, + { + "bbox": [ + 130, + 257, + 481, + 333 + ], + "type": "inline_equation", + "content": "\\omega = 20" + }, + { + "bbox": [ + 130, + 257, + 481, + 333 + ], + "type": "text", + "content": ", often even for " + }, + { + "bbox": [ + 130, + 257, + 481, + 333 + ], + "type": "inline_equation", + "content": "\\omega = 10" + }, + { + "bbox": [ + 130, + 257, + 481, + 333 + ], + "type": "text", + "content": ". Notice how our approach constantly produces more detailed results for all " + }, + { + "bbox": [ + 130, + 257, + 481, + 333 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 130, + 257, + 481, + 333 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 340, + 482, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 482, + 520 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 482, + 520 + ], + "type": "text", + "content": "observation, we produce 3D assets with varying " + }, + { + "bbox": [ + 130, + 340, + 482, + 520 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 130, + 340, + 482, + 520 + ], + "type": "text", + "content": " using the original DreamFusion formulation and a variant using our LMC-SDS. The results in Fig. 8 also show the trend of high " + }, + { + "bbox": [ + 130, + 340, + 482, + 520 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 130, + 340, + 482, + 520 + ], + "type": "text", + "content": " having the tendency to produce oversaturated colors, albeit not as strongly as in the original paper, possibly due the use of a different image diffusion model. But we also observed an additional issue: the infamous Janus problem, the tendency to repeat detail in each view, becomes more severe with bigger " + }, + { + "bbox": [ + 130, + 340, + 482, + 520 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 130, + 340, + 482, + 520 + ], + "type": "text", + "content": ", as evident from the eagle example. On the other hand, when using low " + }, + { + "bbox": [ + 130, + 340, + 482, + 520 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 130, + 340, + 482, + 520 + ], + "type": "text", + "content": " in the original SDS loss, we observe that DreamFusion struggles to create detailed structure or even structure at all. In contrast, when using LMC-SDS, DreamFusion successfully produces detailed results, even when using up to only one tenth of the originally proposed " + }, + { + "bbox": [ + 130, + 340, + 482, + 520 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 130, + 340, + 482, + 520 + ], + "type": "text", + "content": ". In Fig. 9 (more in Suppl. Mat.), we show additional side-by-side comparisons of the original SDS-based DreamFusion and our LMC-SDS version. Using our loss, the results are consistently sharper and feature far more detail. Finally in Fig. 10, we show that fixed " + }, + { + "bbox": [ + 130, + 340, + 482, + 520 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 340, + 482, + 520 + ], + "type": "text", + "content": " sampling may also be used in 3D asset generation, to obtain diverse results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 536, + 300, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 300, + 548 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 300, + 548 + ], + "type": "text", + "content": "6 Discussion & Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "type": "text", + "content": "We proposed a novel loss formulation called LMC-SDS. LMC-SDS allows to effectively utilize text-to-image diffusion models as rich image priors in several optimization and network training tasks. We derived LMC-SDS by analyzing DreamFusion's SDS loss and by identifying an inherent problem in its formulation. Concretely, we expressed the SDS gradient as the sum of two terms. We identified that the second term " + }, + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{proj}}" + }, + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "type": "text", + "content": " provides deficient gradients and consequently was weighted down in DreamFusion. We argued that this term, however, should provide gradients towards the learned manifold of natural images associated to the diffusion model. We also provided evidence that the lack thereof" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "type": "text", + "content": "LMC-SDS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 255 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 255 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 255 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 255 + ], + "type": "image", + "image_path": "c1d0b6406303682d3d553725f0151343131c5d6569271a7909e8064cee2e251c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 255, + 480, + 288 + ], + "lines": [ + { + "bbox": [ + 130, + 255, + 480, + 288 + ], + "spans": [ + { + "bbox": [ + 130, + 255, + 480, + 288 + ], + "type": "text", + "content": "Fig. 9: Results of DreamFusion using the original SDS loss (right) and our proposed LMC-SDS (left) for " + }, + { + "bbox": [ + 130, + 255, + 480, + 288 + ], + "type": "inline_equation", + "content": "\\omega = 20" + }, + { + "bbox": [ + 130, + 255, + 480, + 288 + ], + "type": "text", + "content": ". Our results are much sharper, contain more detail, and feature more realistic colors in all examples." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 293, + 477, + 334 + ], + "blocks": [ + { + "bbox": [ + 133, + 293, + 477, + 334 + ], + "lines": [ + { + "bbox": [ + 133, + 293, + 477, + 334 + ], + "spans": [ + { + "bbox": [ + 133, + 293, + 477, + 334 + ], + "type": "image", + "image_path": "da2d69e7c59d505d9c97f752e6ff124b46b4a62ff470de2e627258c1af0cd2eb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 335, + 480, + 368 + ], + "lines": [ + { + "bbox": [ + 130, + 335, + 480, + 368 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 480, + 368 + ], + "type": "text", + "content": "Fig. 10: DreamFusion results using our LMC-SDS loss and fixed " + }, + { + "bbox": [ + 130, + 335, + 480, + 368 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 335, + 480, + 368 + ], + "type": "text", + "content": " sampling for the prompts \"a mannequin wearing a [hoodie, hat]. Note the diversity in colors, poses, and mannequin styles." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 384, + 482, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 482, + 468 + ], + "type": "text", + "content": "makes the optimization unstable, thus negatively impacting the fidelity of results. To this end, we proposed to model the denoising models' time step dependent frequency bias. This allowed us to factor it out and obtain the cleaner gradient " + }, + { + "bbox": [ + 130, + 384, + 482, + 468 + ], + "type": "inline_equation", + "content": "\\nabla_{\\theta}\\mathcal{L}_{\\mathrm{lmc}}" + }, + { + "bbox": [ + 130, + 384, + 482, + 468 + ], + "type": "text", + "content": ". In extensive experiments, we have demonstrated the effectiveness and versatility of our novel LMC-SDS loss. Furthermore, we discussed how LMC-SDS differs from other SDS-style losses and demonstrated how its performance is superior both qualitatively and quantitatively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 469, + 482, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 469, + 482, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 469, + 482, + 601 + ], + "type": "text", + "content": "Limitations. LMC-SDS finds its limits where either the signal from the diffusion model is too weak and ambiguous or when the optimization goes off the natural image manifold by too much. Concretely, LMC-SDS can only utilize whatever signal is coming from the diffusion model. This means, that when the diffusion model does not \"understand\" the prompt, LMC-SDS (and all other SDS-based approaches) will not be able to provide meaningful gradients. Further, whenever the current optimization state is too corrupted, LMC-SDS will not be able to correct for that and guide the optimization back to the manifold. Also, LMC-SDS sometimes overcompensates the frequency bias. Nevertheless, LMC-SDS is an important first step towards more stable and meaningful usage of the inherent image prior associated to image diffusion models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 601, + 482, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 601, + 482, + 661 + ], + "spans": [ + { + "bbox": [ + 130, + 601, + 482, + 661 + ], + "type": "text", + "content": "We proposed LMC-SDS, a novel SDS-based loss formulation with cleaner gradients towards the image manifold. In the future, we would like to continue to investigate how the manifold corrective can be further improved. We also want to leverage our findings in concrete applications like text-to-3D or in image editing and stylization tasks." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "Aldieck et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 138, + 140, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 481, + 172 + ], + "type": "text", + "content": "1. Armandpour, M., Zheng, H., Sadeghian, A., Sadeghian, A., Zhou, M.: Re-imagine the negative prompt algorithm: Transform 2d diffusion into 3d, alleviate janus problem and beyond. arXiv preprint arXiv:2304.04968 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 173, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 173, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 138, + 173, + 481, + 194 + ], + "type": "text", + "content": "2. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 194, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 194, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 138, + 194, + 481, + 216 + ], + "type": "text", + "content": "3. Choi, Y., Uh, Y., Yoo, J., Ha, J.W.: Stargan v2: Diverse image synthesis for multiple domains. In: IEEE Conf. Comput. Vis. Pattern Recog. (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 217, + 481, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 217, + 481, + 238 + ], + "spans": [ + { + "bbox": [ + 138, + 217, + 481, + 238 + ], + "type": "text", + "content": "4. Chung, H., Kim, J., Mccann, M.T., Klasky, M.L., Ye, J.C.: Diffusion posterior sampling for general noisy inverse problems. In: Int. Conf. Learn. Represent. (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 239, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 239, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 138, + 239, + 481, + 270 + ], + "type": "text", + "content": "5. Chung, H., Sim, B., Ryu, D., Ye, J.C.: Improving diffusion models for inverse problems using manifold constraints. Adv. Neural Inform. Process. Syst. 35, 25683-25696 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 271, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 271, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 138, + 271, + 481, + 293 + ], + "type": "text", + "content": "6. Croitoru, F.A., Hondru, V., Ionescu, R.T., Shah, M.: Diffusion models in vision: A survey. IEEE Trans. Pattern Anal. Mach. Intell. (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 293, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 293, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 138, + 293, + 481, + 335 + ], + "type": "text", + "content": "7. Gao, S., Liu, X., Zeng, B., Xu, S., Li, Y., Luo, X., Liu, J., Zhen, X., Zhang, B.: Implicit diffusion models for continuous super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10021-10030 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 337, + 481, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 337, + 481, + 358 + ], + "spans": [ + { + "bbox": [ + 138, + 337, + 481, + 358 + ], + "type": "text", + "content": "8. Hertz, A., Aberman, K., Cohen-Or, D.: Delta denoising score. In: Int. Conf. Comput. Vis. (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 359, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 359, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 138, + 359, + 481, + 380 + ], + "type": "text", + "content": "9. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 381, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 381, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 138, + 381, + 481, + 402 + ], + "type": "text", + "content": "0. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Adv. Neural Inform. Process. Syst. 33, 6840-6851 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 403, + 481, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 403, + 481, + 424 + ], + "spans": [ + { + "bbox": [ + 138, + 403, + 481, + 424 + ], + "type": "text", + "content": "1. Ho, J., Salimans, T.: Classifier-free diffusion guidance. In: NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 425, + 481, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 425, + 481, + 457 + ], + "spans": [ + { + "bbox": [ + 138, + 425, + 481, + 457 + ], + "type": "text", + "content": "2. Jain, A., Xie, A., Abbeel, P.: Vectorfusion: Text-to-svq by abstracting pixel-based diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1911-1920 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 458, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 458, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 138, + 458, + 481, + 479 + ], + "type": "text", + "content": "3. Katzir, O., Patashnik, O., Cohen-Or, D., Lischinski, D.: Noise-free score distillation. arXiv preprint arXiv:2310.17590 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 479, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 479, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 138, + 479, + 481, + 501 + ], + "type": "text", + "content": "4. Kawar, B., Elad, M., Ermon, S., Song, J.: Denoising diffusion restoration models. In: Adv. Neural Inform. Process. Syst. (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 502, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 502, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 502, + 481, + 533 + ], + "type": "text", + "content": "5. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 6007-6017 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 534, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 534, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 534, + 481, + 555 + ], + "type": "text", + "content": "6. Kim, S., Lee, K., Choi, J.S., Jeong, J., Sohn, K., Shin, J.: Collaborative score distillation for consistent visual synthesis. arXiv preprint arXiv:2307.04787 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 556, + 481, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 556, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 138, + 556, + 481, + 588 + ], + "type": "text", + "content": "7. Kolotouros, N., Alldieck, T., Zanfir, A., Bazavan, E.G., Fieraru, M., Sminchisescu, C.: Dreamhuman: Animatable 3d avatars from text. In: Adv. Neural Inform. Process. Syst. (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 589, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 589, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 589, + 481, + 632 + ], + "type": "text", + "content": "8. Kuznetsova, A., Rom, H., Alldrin, N., Uijlings, J., Krasin, I., Pont-Tuset, J., Kamali, S., Popov, S., Malloci, M., Kolesnikov, A., Duerig, T., Ferrari, V.: The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. Int. J. Comput. Vis. (2020)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "text", + "content": "9. Li, X., Ren, Y., Jin, X., Lan, C., Wang, X., Zeng, W., Wang, X., Chen, Z.: Diffusion models for image restoration and enhancement-a comprehensive survey. arXiv preprint arXiv:2308.09388 (2023)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "text", + "content": "LMC-SDS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "20. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3d: High-resolution text-to-3d content creation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 300–309 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 171 + ], + "type": "text", + "content": "21. Lin, S., Liu, B., Li, J., Yang, X.: Common diffusion noise schedules and sample steps are flawed. arXiv preprint arXiv:2305.08891 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "type": "text", + "content": "22. Liu, Q., Wang, D.: Stein variational gradient descent: A general purpose bayesian inference algorithm. Advances in neural information processing systems 29 (2016)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 225 + ], + "type": "text", + "content": "23. Liu, R., Wu, R., Van Hoorick, B., Tokmakov, P., Zakharov, S., Vondrick, C.: Zero-1-to-3: Zero-shot one image to 3d object. In: Int. Conf. Comput. Vis. pp. 9298-9309 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "text", + "content": "24. Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Van Gool, L.: Repaint: Inpainting using denoising diffusion probabilistic models. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 11461-11471 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 480, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 292 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 292 + ], + "type": "text", + "content": "25. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: SDEdit: Guided image synthesis and editing with stochastic differential equations. In: Int. Conf. Learn. Represent. (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "text", + "content": "26. Metzer, G., Richardson, E., Patashnik, O., Giryes, R., Cohen-Or, D.: Latent-nerf for shape-guided generation of 3d shapes and textures. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 12663-12673 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "type": "text", + "content": "27. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In: Int. Conf. on Mach. Learn. (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 358, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 391 + ], + "type": "text", + "content": "28. Park, D.H., Azadi, S., Liu, X., Darrell, T., Rohrbach, A.: Benchmark for compositional text-to-image synthesis. In: Conf. on Neural Inf. Proc. Systems Datasets and Benchmarks Track (Round 1) (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "type": "text", + "content": "29. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. In: Int. Conf. Learn. Represent. (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 414, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 480, + 445 + ], + "type": "text", + "content": "30. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 447, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 480, + 479 + ], + "type": "text", + "content": "31. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 10684-10695 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "type": "text", + "content": "32. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 22500-22510 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "text", + "content": "33. Sahara, C., Chan, W., Chang, H., Lee, C., Ho, J., Salimans, T., Fleet, D., Norouzi, M.: Palette: Image-to-image diffusion models. In: ACM SIGGRAPH 2022 Conference Proceedings. pp. 1-10 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 545, + 480, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 480, + 587 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 480, + 587 + ], + "type": "text", + "content": "34. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Adv. Neural Inform. Process. Syst. 35, 36479-36494 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 588, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 588, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 480, + 610 + ], + "type": "text", + "content": "35. Shi, Y., Wang, P., Ye, J., Mai, L., Li, K., Yang, X.: Mvdream: Multi-view diffusion for 3d generation. arXiv:2308.16512 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "text", + "content": "36. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. In: Int. Conf. Learn. Represent. (2020)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "37. Wang, H., Du, X., Li, J., Yeh, R.A., Shakhnarovich, G.: Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 12619-12629 (2023)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "Aldieck et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 249 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "38. Wang, Z., Lu, C., Wang, Y., Bao, F., Li, C., Su, H., Zhu, J.: Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "text", + "content": "39. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: IEEE Conf. Comput. Vis. Pattern Recog. (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 482, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 482, + 216 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 482, + 216 + ], + "type": "text", + "content": "40. Zhou, Z., Tulsiani, S.: Sparsefusion: Distilling view-conditioned diffusion for 3d reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12588-12597 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 216, + 482, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 216, + 482, + 249 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 482, + 249 + ], + "type": "text", + "content": "41. Zou, Z.X., Cheng, W., Cao, Y.P., Huang, S.S., Shan, Y., Zhang, S.H.: Sparse3d: Distilling multiview-consistent diffusion for object reconstruction from sparse views. arXiv preprint arXiv:2308.14078 (2023)" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 404, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 404, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 404, + 91, + 447, + 100 + ], + "type": "text", + "content": "LMC-SDS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/ScribblePrompt_ Fast and Flexible Interactive Segmentation for Any Biomedical Image/97fd2117-c4b3-47cc-8dc3-f7ab8ca17648_content_list.json b/2024/ScribblePrompt_ Fast and Flexible Interactive Segmentation for Any Biomedical Image/97fd2117-c4b3-47cc-8dc3-f7ab8ca17648_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..795112462f11b78abf5feae3d4acb98d131350be --- /dev/null +++ b/2024/ScribblePrompt_ Fast and Flexible Interactive Segmentation for Any Biomedical Image/97fd2117-c4b3-47cc-8dc3-f7ab8ca17648_content_list.json @@ -0,0 +1,2490 @@ +[ + { + "type": "text", + "text": "ScribblePrompt: Fast and Flexible Interactive Segmentation for Any Biomedical Image", + "text_level": 1, + "bbox": [ + 240, + 140, + 763, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hallee E. Wong $^{1,2}$ , Marianne Rakic $^{1,2}$ , John Guttag $^{1}$ , and Adrian V. Dalca $^{1,2,3}$", + "bbox": [ + 274, + 210, + 727, + 242 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 MIT CSAIL, Cambridge, MA, USA \n2 Martins Center, Massachusetts General Hospital, Charlestown, MA, USA \n3 Harvard Medical School, Boston, MA, USA \n{hallel, mrukic, guttag, adalca}@mit.edu", + "bbox": [ + 243, + 253, + 754, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Biomedical image segmentation is a crucial part of both scientific research and clinical care. With enough labelled data, deep learning models can be trained to accurately automate specific biomedical image segmentation tasks. However, manually segmenting images to create training data is highly labor intensive and requires domain expertise. We present ScribblePrompt, a flexible neural network based interactive segmentation tool for biomedical imaging that enables human annotators to segment previously unseen structures using scribbles, clicks, and bounding boxes. Through rigorous quantitative experiments, we demonstrate that given comparable amounts of interaction, ScribblePrompt produces more accurate segmentations than previous methods on datasets unseen during training. In a user study with domain experts, ScribblePrompt reduced annotation time by $28\\%$ while improving Dice by $15\\%$ compared to the next best method. ScribblePrompt's success rests on a set of careful design decisions. These include a training strategy that incorporates both a highly diverse set of images and tasks, novel algorithms for simulated user interactions and labels, and a network that enables fast inference. We showcase ScribblePrompt in an interactive demo, provide code, and release a dataset of scribble annotations at https://scribbleprompt.csail.mit.edu", + "bbox": [ + 261, + 335, + 738, + 613 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Interactive Segmentation $\\cdot$ Biomedical Imaging $\\cdot$ Scribbles", + "bbox": [ + 261, + 626, + 736, + 640 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 661, + 375, + 676 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Biomedical image segmentation is an essential step in a wide range of biomedical research and clinical care pipelines. Deep learning has become the predominant method to automate existing segmentation tasks [43, 78, 92, 106]. Biomedical researchers and clinicians often encounter novel segmentation tasks involving either new regions of interest [119] or new image modalities [77,97]. Unfortunately, supervised training of accurate models for new domains requires diverse images with careful annotations by skilled experts.", + "bbox": [ + 212, + 688, + 785, + 794 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Most widely used interactive segmentation systems for biomedical imaging provide minimal, intensity-based, algorithmic assistance [118]. Despite a growing literature, learning-based systems are less widely used in practice, perhaps", + "bbox": [ + 212, + 795, + 785, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ab2a2cc597ac03ffede42436570130f5d75b3c0c560ce4b4d14c3225dabb1cb7.jpg", + "image_caption": [ + "Step 1" + ], + "image_footnote": [], + "bbox": [ + 218, + 154, + 287, + 209 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0ac61d155e098a22eb0ae8e997599603d95e09b732829390570cc097b4c7492f.jpg", + "image_caption": [ + "Step 2" + ], + "image_footnote": [], + "bbox": [ + 289, + 154, + 357, + 209 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a9016d7deb2cfa31379d1f933b4d180a89a4b1622c9286b10f94aa82d44d1668.jpg", + "image_caption": [ + "Step 3" + ], + "image_footnote": [], + "bbox": [ + 359, + 154, + 426, + 209 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/24a6138a080e8ccd473dd682cfeeefe45d4cc88a6c12a2f401143874bfcca03b.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 429, + 154, + 500, + 209 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2312047fc6d80b06fdcb4340e676ff1e78f67d7bb6a341a2722a33f00a12168a.jpg", + "image_caption": [ + "Step 1" + ], + "image_footnote": [], + "bbox": [ + 501, + 154, + 568, + 209 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9552129b489b034cf5e7e178f3b4cba83f8b8e43cd1c979ac1c80fe1143dbb69.jpg", + "image_caption": [ + "Step 2" + ], + "image_footnote": [], + "bbox": [ + 570, + 154, + 638, + 209 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/28d9093f08efce348f0b1855184a0cf61b42d36ec5c953a6820d223e8766a5c7.jpg", + "image_caption": [ + "Step 3" + ], + "image_footnote": [], + "bbox": [ + 638, + 154, + 709, + 209 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d4349865b1db0bd39b571fe9e818c8d814f9b6f77b6044e25ae1494740a5ba2f.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 710, + 154, + 782, + 209 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/697261cfefe8c33764db424f08fe24120cbb58f3a778970487dcf73e44b30bab.jpg", + "image_caption": [ + "Fig. 1: ScribblePrompt enables rapid iterative interactive segmentation of unseen tasks using bounding boxes, clicks, and scribbles. We show predictions from ScribblePrompt with iterative interaction steps on examples from datasets unseen during training. At each step, we visualize positive scribble and click inputs in green, negative scribble and click inputs in red, bounding box inputs in yellow, and the predicted segmentation in blue. Scribble thickness is enlarged for visual clarity. See Supplementary Material for more examples." + ], + "image_footnote": [], + "bbox": [ + 218, + 210, + 292, + 263 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/bad9e3e9b10dfb4f7f8e31e5c54b7d379a03a82597db2034c49ae538bfcfa64b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 294, + 210, + 361, + 263 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/35083b4c22cb5980a2ac19243b6ca7f5f39a1a1dee79373523a411b4ba0bf012.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 362, + 210, + 428, + 263 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/dd6305b1025ea0780e39f7fd7038aa1dc28f2fe9dbc210af503f1fa720ef0bf2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 429, + 210, + 500, + 263 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/56a8c466e30b316a0623ae88158886254c05eed26dea0283f90a63e8ac606546.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 210, + 568, + 263 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/60b534ecd8ee3b27f4b6037d885c698de56b2cae4777a65d37f36718123a0965.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 570, + 210, + 638, + 263 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b980f866bee65adbd815cbb0c10574c542e061fa5e67d7cc7f789a140699f31b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 210, + 709, + 263 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b663fb298dc9eec1c53be49c8be425b85f38b1a0b789db3f0d9e3b57feb3b85d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 210, + 782, + 263 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "because they are generally focused on specific tasks or modalities, which limits their applicability [68, 83, 95, 109, 110, 128]. Recent vision foundation models target broad applicability, but require fine-tuning to the medical domain [47, 112, 113, 129]. Most interactive segmentation models require specific interactions, like carefully placed clicks, making them easy to develop but difficult to use in practice. The result of these shortcomings is that existing learning-based models are not widely used in practice.", + "bbox": [ + 212, + 402, + 784, + 507 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we present ScribblePrompt, a general model for interactive biomedical image segmentation that", + "bbox": [ + 214, + 508, + 784, + 537 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. enables users to rapidly and accurately accomplish any biomedical image segmentation task, outperforming state-of-the-art models, particularly for unseen labels and image types;", + "2. is flexible to different annotation styles, including bounding boxes, clicks, and scribbles,", + "3. is computationally efficient, enabling fast inference, even on a single CPU." + ], + "bbox": [ + 220, + 546, + 784, + 635 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To achieve this, we focused on design decisions aimed at ease of use and realistic interactions. To train our model, we started by gathering a large corpus of biomedical imaging datasets. We then designed task augmentation through synthesis strategies to encourage generalization to unseen tasks. Finally, we designed new interaction simulation strategies for training. The model itself uses an architecture optimized for efficient inference.", + "bbox": [ + 212, + 643, + 784, + 733 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To evaluate ScribblePrompt, we compared its performance to that of existing interactive segmentation methods on datasets unseen during training. These experiments involved manually-collected scribbles, simulated interactions (Fig. 1), and a user study in which experienced annotators were asked to segment images. In the user study, ScribblePrompt reduced annotation time by $28\\%$ while increasing Dice by $15\\%$ compared to the next best method, the Segment Anything Model (SAM) [47]. Given similar amounts of interaction, ScribblePrompt", + "bbox": [ + 212, + 734, + 785, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "H.E. Wong et al.", + "bbox": [ + 271, + 114, + 385, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "achieved consistently better Dice scores than other methods. We release an interactive tool, code, model weights, and our dataset of manually-collected scribbles at https://scribbleprompt.csail.mit.edu.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 215, + 395, + 233 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Interactive Biomedical Image Segmentation. Early research into interactive segmentation of biomedical images focused on traditional intensity-based methods [13,21,33,107,118]. Recent deep-learning based techniques [68,83,109, 110] use human interaction to improve the accuracy of segmentation tasks where labelled data exists, but fully-automatic segmentation methods fail to produce accurate segmentations. This approach assumes that the model will be used to perform the same segmentation task(s) it was trained on, leading to domain-specific solutions [6,62,80,89,110,111]. A few methods generalize to new, but similar, classes [95] and modalities [68,109] to those seen in training. In contrast, ScribblePrompt is designed to help annotators segment a wide range of biomedical images and (potentially unseen) tasks at inference time.", + "bbox": [ + 212, + 250, + 787, + 415 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Foundation Models. Recent vision foundation models employ prompting to enable generalization to new tasks. These models, trained on large collections of (natural) image data, segment potentially-unseen structures specified by spatial prompts [47, 129], text [47, 87, 129], or examples [15, 22, 112, 113]. Some of these models can be used for interactive segmentation with spatial prompts.", + "bbox": [ + 212, + 416, + 787, + 491 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fine-tuning models initially developed using natural images is often unhelpful for specific tasks in the biomedical imaging domain, compared to training from scratch [88]. The Segment Anything Model (SAM) [47], a foundation model trained on natural images, performs well at biomedical image segmentation tasks with clear boundaries (e.g. organs in abdominal CT), but performs poorly on tasks and modalities involving more subtle delineations (e.g. deep structures in Brain MRI) [36, 41, 76, 100]. Several recent or concurrent papers fine-tune SAM for specific biomedical tasks or modalities [40, 46, 59, 81, 108, 120]. A few fine-tune SAM to segment medical images from multiple-modalities with bounding boxes [69], clicks [115] or both [19]. We show in our experiments that these models don't perform well on many unseen tasks and are limiting in their required interactions.", + "bbox": [ + 212, + 492, + 787, + 671 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Several natural image methods simulate iterative user interactions and condition the model on previous predictions to train a single network to make the initial prediction and perform refinement [14, 27, 47, 61, 102, 103]. This approach is appropriate for natural images where there is ample data and typically one modality. However, such methods have not previously been developed for the biomedical imaging domain, which is more fragmented because of specialized regions of interest and different modalities. We use iterative ideas in ScribblePrompt, but we facilitate iterative interactive segmentation of unseen biomedical imaging tasks.", + "bbox": [ + 212, + 672, + 787, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "User Interaction. Users can prompt interactive segmentation systems in many different ways, such as bounding boxes [9, 47, 89, 94, 109, 117], scribbles", + "bbox": [ + 214, + 809, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "ScribblePrompt", + "bbox": [ + 624, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "[18, 110, 111], or clicks [9, 61, 68, 103, 110]. Few interactive segmentation models incorporate multiple types of inputs. Some works have explored training with more specialized types of clicks [122], such as extreme points [72, 93], interior margin points [68], and center clicks [103, 116], to maximize the information per interaction. Specialized interactions often require more user time, and models trained only with such interactions are less robust to deviations from the interaction protocol [9, 103, 117]. In contrast, we focus on a user-friendly model that performs well under a variety of interaction scenarios.", + "bbox": [ + 212, + 146, + 787, + 267 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Scribbles are an intuitive form of interaction, however few interactive segmentation works employ them, because acquiring realistic scribbles for training and evaluation is challenging. Previous deep-learning models trained for interactive scribble-based segmentation either relied on collecting large datasets of manually drawn scribbles [53, 57, 128] or simplistic approaches such as sampling random points [5, 7, 110] or Bezier curves [2]. One work has explored simulating more complex scribbles for interactive segmentation of natural images [18]. We build on these concepts to create a new scribble simulation engine for training and evaluation.", + "bbox": [ + 212, + 268, + 787, + 404 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Scribble-supervised learning methods use scribble annotations as supervision to train automatic segmentation models for predicting segmentation given only an input image [32,56,57,66,121]. However, these methods require the manual scribble-annotation of many training images and retraining for each new task. In contrast, ScribblePrompt is trained on a large corpus of datasets once and then can be used to perform new segmentation tasks at inference time without retraining, using scribbles as input.", + "bbox": [ + 212, + 406, + 787, + 512 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 ScribblePrompt Approach", + "text_level": 1, + "bbox": [ + 215, + 542, + 504, + 559 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We present an interactive segmentation method that can generalize to new biomedical imaging modalities and regions of interest, while staying focused on practical usability. We describe the problem formulation, and present the important aspects of the ScribblePrompt framework: (i) simulation of realistic interactions during training, (ii) augmentation with synthetic labels during training to encourage generalization, and (iii) an efficient architecture for fast inference. We show an overview of training in Fig. 2 and 4.", + "bbox": [ + 212, + 580, + 787, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Problem Formulation", + "text_level": 1, + "bbox": [ + 215, + 715, + 439, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let $t$ be a segmentation task consisting of image and segmentation pairs, $\\{(x^t,y^t)_j\\}_{j = 1}^N$ . At step $i$ , given an image $x^{t}$ , a set of user interactions $u_{i}$ , and previous prediction $\\hat{y}_{i - 1}^{t}$ , we learn function $f_{\\theta}(x^{t},u_{i},\\hat{y}_{i - 1}^{t})$ with parameters $\\theta$ that produces a segmentation $\\hat{y}_i$ . The set of interactions $u_{i}$ , which may include positive or negative scribbles, positive or negative clicks, and bounding boxes, is provided by a user who has access to the image $x^{t}$ and previous prediction $\\hat{y}_{i - 1}^{t}$ .", + "bbox": [ + 212, + 750, + 787, + 842 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "H.E. Wong et al.", + "bbox": [ + 271, + 114, + 385, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We minimize the difference between the true segmentation $y^{t}$ and each of the $k$ iterative predictions $\\hat{y}_1, \\dots, \\hat{y}_k$ ,", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (\\theta ; \\mathcal {T}) = \\mathbb {E} _ {t \\in \\mathcal {T}} \\left[ \\mathbb {E} _ {(x ^ {t}, y ^ {t}) \\in t} \\left[ \\sum_ {i = 1} ^ {k} \\mathcal {L} _ {\\text {S e g}} \\left(y ^ {t}, f _ {\\theta} \\left(x ^ {t}, u _ {i}, \\hat {y} _ {i - 1} ^ {t}\\right)\\right) \\right] \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 284, + 184, + 785, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{Seg}$ is a supervised segmentation loss.", + "bbox": [ + 212, + 233, + 544, + 248 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During training, we sample a task $t$ from training task collection $\\mathcal{T}$ , from which we sample an image $x^t$ , and segmentation map $y^{t}$ . We simulate a possible set of interactions $u_{i}$ based on $y^{t}$ , and predict $\\hat{y}_i^t = f_\\theta (x^t,u_i,\\hat{y}_{i - 1}^t)$ . We simulate the next set of interactions $u_{i + 1}$ based on the error region $\\varepsilon_i^t = y^t -\\hat{y}_i^t$ , and repeat for $k$ iterations. In the following sections, we describe the core aspects of the framework: strategies for simulating $u_{i}$ , collecting $\\mathcal{T}$ , sampling $(x^{t},y^{t})$ , building $f_{\\theta}$ , and optimizing $\\mathcal{L}(\\theta ;\\mathcal{T})$ .", + "bbox": [ + 212, + 250, + 784, + 354 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f67e2ca5adc2a5a4162dc1ba591ce497f51e904eb348184d2b18087b5a5cbb82.jpg", + "image_caption": [ + "Fig. 2: Training. We simulate $k$ consecutive steps of interactive segmentation. Given an image segmentation pair $(x^t, y^t)$ , we first simulate a set of initial interactions $u_1$ , which may contain bounding boxes, clicks, and/or scribbles. We predict segmentation $\\hat{y}_1^t \\coloneqq f_\\theta(x^t, u_1, \\hat{y}_0^t)$ where the initial prediction $\\hat{y}_0^t$ is set to zeros. In the second step, we simulate corrections using the error region $\\varepsilon_1^t$ between the previous prediction $\\hat{y}_1^t$ and ground truth $y^t$ , and add them to the set of initial interactions $u_1$ to get $u_2$ . We predict segmentation $\\hat{y}_2^t \\coloneqq f_\\theta(x^t, u_2, \\hat{y}_1^t)$ and repeat to produce a series of predictions, $\\hat{y}_1^t, \\dots, \\hat{y}_k^t$ . We learn $\\theta$ to minimize $\\sum_{i=1}^{k} \\mathcal{L}_{seg}(y^t, \\hat{y}_i^t)$ , the sum of losses between the target segmentation $y^t$ and iterative predictions $\\hat{y}_1^t, \\dots, \\hat{y}_k^t$ ." + ], + "image_footnote": [], + "bbox": [ + 217, + 378, + 787, + 554 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Prompt Simulation", + "text_level": 1, + "bbox": [ + 214, + 741, + 419, + 756 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To enable a practical, easy-to-use model, we encourage robustness to different types of user interactions $u_{i}$ . We introduce algorithms for simulating scribbles, clicks, and bounding box inputs during training. Each scribble and click strategy (Fig. 3) can be applied to a ground truth segmentation label $y^{t}$ to simulate positive interactions, or to the background $1 - y^{t}$ to simulate negative interactions.", + "bbox": [ + 212, + 763, + 782, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "ScribblePrompt", + "bbox": [ + 624, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/073c31a57efb31e5feb13a8f48c98990c006c7d081dbef55624b903f19eed638.jpg", + "image_caption": [ + "Fig. 3: Simulated scribbles and clicks. Positive interactions (green) are simulated on the segmentation label $y^{t}$ (blue), while negative interactions (red) are simulated on the background $1 - y^{t}$ . Scribble thickness is enlarged for visual clarity." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 784, + 229 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We simulate positive and negative correction scribbles or clicks by applying the same strategies to the false negative region $\\varepsilon_{i}^{t} > 0$ and false positive region $\\varepsilon_{i}^{t} < 0$ for error region $\\varepsilon_{i}^{t}$ .", + "bbox": [ + 212, + 315, + 782, + 359 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Scribbles. Given binary mask $y \\in \\{0,1\\}^{h \\times w}$ , we simulate a scribble mask $s \\in [0,1]^{h \\times w}$ by first generating clean scribbles, and then corrupting them to account for user behavior and variability. We illustrate this process in Supplementary Material Sec. B.1. To generate the clean scribbles, we start with one of these strategies:", + "bbox": [ + 212, + 361, + 784, + 436 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(i) Line Scribbles: We draw random lines by connecting two end points sampled from $\\{(u,v)|y_{uv} = 1\\}$ .", + "(ii) Centerline Scribbles: We simulate scribbles in the center of label $y$ using a thinning algorithm [123] that reduces the label to a 1-pixel wide skeleton.", + "(iii) Contour Scribbles: We simulate a rough contour of the desired segmentation within the boundaries of the mask. We first blur the mask to reduce the size of the label such that $\\tilde{y} = \\min(y, y \\circ G_k)$ , where $G_k$ is a Gaussian blur kernel. Then we apply a threshold $\\tilde{y} < h$ sampled in some intensity range $h \\sim U[\\tilde{y}_{\\text{min}}, \\tilde{y}_{\\text{max}}]$ and extract a contour inside the boundary of the mask." + ], + "bbox": [ + 210, + 449, + 784, + 587 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To limit the size and complexity of the centerline and contour scribbles, we apply a random mask that breaks the scribbles into smaller parts. We generate the random mask by sampling a smooth noise image $p$ , where each pixel is sampled independently from $\\mathcal{N}(\\mu_p, \\sigma_p^2)$ , applying Gaussian blur, and then thresholding it at $\\mu_p$ . We warp the resulting scribble mask $s$ using a random deformation field $\\phi$ to vary the scribble shape and thickness. We ensure the resulting scribble is consistent with mask $y$ , by multiplying the warped scribble mask $s \\circ \\phi$ by $y$ .", + "bbox": [ + 212, + 598, + 784, + 704 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Clicks. Given a binary mask $y \\in \\{0,1\\}^{h \\times w}$ , we simulate $n \\sim U[n_{min}, n_{max}]$ clicks at a time using one of three strategies, illustrated in Fig. 3:", + "bbox": [ + 212, + 705, + 782, + 736 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(i) Random clicks: We randomly sample clicks from all pixels in the given region $\\{(u,v)|y_{uv} = 1\\}$ .", + "(ii) Center clicks: We sample clicks from the set of points at the center of each disconnected component of the label. First, we create a multi-label mask $m \\in \\{1, \\dots, C\\}^{h \\times w}$ identifying the $C$ components of label $y$ . We identify the center of each component using the euclidean distance transform [103, 116]." + ], + "bbox": [ + 212, + 747, + 784, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "H.E. Wong et al.", + "bbox": [ + 271, + 114, + 385, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(iii) Interior border region clicks: We sample clicks from a border region inside the boundary of the mask. We first blur the mask $y$ to reduce the size of the label such that $\\tilde{y} = \\min(y, y \\circ G_k)$ where $G_k$ is a Gaussian blur kernel. We then sample click coordinates from $\\{(u, v) | \\tilde{y}_{uv} \\in [a, b]\\}$ , where $a, b \\sim U[\\tilde{y}_{\\text{min}}, \\tilde{y}_{\\text{max}})$ are thresholds sampled in some intensity range.", + "bbox": [ + 209, + 145, + 787, + 223 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Since users are inclined to spread out their clicks, we impose a minimum separation of a few pixels between random clicks and border region clicks.", + "bbox": [ + 212, + 227, + 785, + 257 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Bounding Boxes. We compute the minimum bounding box that encloses the label $y^{t}$ , and enlarge each dimension by $r \\sim U[0,20]$ pixels to account for human variability.", + "bbox": [ + 212, + 257, + 785, + 301 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Iterative Training. During the first step $(i = 1)$ , we sample the combination of interactions and the number of initial positive and negative interactions $n_{pos}, n_{neg} \\sim U[n_{min}, n_{max}]$ . The initial interactions $u_{1}$ are simulated using the ground truth label $y^{t}$ . In subsequent steps, correction scribbles or clicks are sampled from the error region $\\varepsilon_{i-1}^{t}$ between the last prediction $\\hat{y}_{i-1}^{t}$ and the ground truth $y^{t}$ . Since a user can make multiple corrections in each step, we sample $n_{cor} \\sim U[n_{min}, n_{max}]$ corrections (scribbles or clicks) per step.", + "bbox": [ + 212, + 301, + 785, + 409 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Data", + "text_level": 1, + "bbox": [ + 215, + 426, + 303, + 440 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We build on large dataset gathering efforts like MegaMedical [15,90] to compile a collection of 77 open-access biomedical imaging datasets for training and evaluation, covering over 54k scans, 16 image types, and 711 labels. The collection includes a diverse array of biomedical domains, such as eyes [39,55,70,85,105], thorax [86,96,98,99,101], spine [64,98,114,126], cells [16,17,28,29,63,127], skin [20], abdominal [11,12,37,44,45,51,52,54,60,67,71,86,91,101,104], neck [48,79,82,84], brain [3,8,31,38,49,50,73-75,101], bones [34,98,114], teeth [1,65] and lesions [4,101,124,125].", + "bbox": [ + 212, + 446, + 785, + 568 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We define a 2D segmentation task as a combination of dataset, axis (for 3D modalities), and label. For datasets with multiple segmentation labels, we consider each label separately as a binary segmentation task and for 3D modalities we use the slice with maximum label area and the middle slice from each volume. We provide more details on the data in Supplementary Material Sec. C.", + "bbox": [ + 212, + 568, + 785, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Task Diversity. During training we sample image, segmentation pairs hierarchically - by dataset and modality, axis, and then label - to balance training on datasets of different sizes. To increase the diversity of segmentation tasks, we apply data augmentation [15] to both the input image and sampled segmentation prior to simulating the user interactions.", + "bbox": [ + 212, + 643, + 785, + 718 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Synthetic Labels. To limit overfitting to specific segmentation tasks, we introduce a mechanism to generate synthetic labels (Fig. 4). During training, for a given sample $(x_0,y_0)$ , with probability $p_{synth}$ , we replace $y$ with a synthetic label $y_{synth}$ . Given an image $x_0$ , we generate a synthetic label $y_{synth}$ by applying a superpixel algorithm [25] with scale parameter $\\lambda \\sim U[1,\\lambda_{max}]$ to partition the image into a multi-label mask of $k$ superpixels $z \\in \\{1,\\dots,k\\}^{n\\times n}$ , and then randomly select a superpixel $y_{synth} = \\mathbb{1}(z = c)$ . We conduct experiments varying $p_{synth}$ and show examples in Supplementary Material Sec. B.3.", + "bbox": [ + 212, + 719, + 785, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "ScribblePrompt", + "bbox": [ + 624, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f9a18ea5e998c44d1f23a8916c05b45ba89e9ccff0d65123a7d44b80b8497cb8.jpg", + "image_caption": [ + "Fig. 4: Task sampling and augmentation. During training, we sample an image and segmentation pair $(x_0, y_0)$ . With probability $p_{\\text{synth}}$ , we replace $y_0$ with a synthetic label $y_{\\text{synth}}$ . We generate $y_{\\text{synth}}$ by applying a superpixel algorithm to the image $x_0$ to generate a map $z$ of potential synthetic labels (superpixels), and then sampling one label. Finally, we apply random data augmentations to get $(x^t, y^t)$ ." + ], + "image_footnote": [], + "bbox": [ + 215, + 142, + 787, + 238 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Network", + "text_level": 1, + "bbox": [ + 215, + 347, + 334, + 361 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Motivated by producing a practical tool, we primarily demonstrate ScribblePrompt using an efficient fully-convolutional architecture similar to a UNet [92]. We also demonstrate ScribblePrompt using a vision transformer architecture [47].", + "bbox": [ + 212, + 372, + 797, + 419 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ScribblePrompt-UNet: We use an 8 layer CNN following a decoder-encoder structure similar to the popular UNet architecture [92] without Batch Norm. Each convolutional layer has 192 features and uses PReLU activation [35].", + "bbox": [ + 212, + 419, + 790, + 464 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ScribblePrompt-SAM: We take the smallest SAM model (ViT-b) [47] and fine-tune the decoder. SAM takes bounding box and click inputs as lists of $(x,y)$ coordinates, and the low-resolution logits of the previous prediction $\\hat{y}_{i-1}$ . To adapt SAM for scribbles, we consider each scribbled pixel as a click.", + "bbox": [ + 212, + 464, + 789, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Loss. We minimize eq. (1) where $\\mathcal{L}_{Seg}$ is a linear combination of soft Dice Loss [24] and Focal Loss [58]. In preliminary experiments (Supplementary Material Sec. B.2), training with Soft Dice Loss alone or a linear combination of Soft Dice Loss and Binary-Cross Entropy Loss resulted in slightly lower Dice scores.", + "bbox": [ + 212, + 523, + 787, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experimental Setup", + "text_level": 1, + "bbox": [ + 215, + 609, + 444, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We compare ScribblePrompt-UNet and ScribblePrompt-SAM to previous methods through experiments with manual scribbles, simulated interactions, and a user study with experienced annotators. Lastly, we report on inference runtime and ablation experiments.", + "bbox": [ + 212, + 642, + 787, + 703 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Data. We use 65 (out of 77) datasets during training. We partition another nine datasets, ACDC [10], BUID [4], BTCV Cervix [52], DRIVE [105], HipXRay [34], PanDental [1], SCD [86], SpineWeb [126], and WBC [127], each into validation (used in model selection, but not training) and test (used only for final evaluation). Three additional datasets, TotalSegmentator [114], SCR [30], and COBRE [3,23,26], were used only for final evaluation.", + "bbox": [ + 212, + 703, + 787, + 794 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We report final results on 12 sets of data (the test splits of the $9 + 3$ evaluation datasets). The evaluation datasets were not seen during training, and cover 608 tasks and 8 modalities, including unseen image types and unseen labels. We", + "bbox": [ + 212, + 794, + 787, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "H.E. Wong et al.", + "bbox": [ + 271, + 114, + 385, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "selected these 12 evaluation datasets to cover a variety of modalities (MRI, CT, ultrasound, fundus photography, microscopy) and anatomical regions of interest (brain, teeth, bones, abdominal organs, muscles, heart, thorax, cells), including both healthy anatomy and lesions.", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Training. During training, we simulated five steps of interactive segmentation per example and set the maximum number of interactions per step to three. We set the minimum number of initial negative prompts to zero, and the minimum initial positive and correction prompts to one. During the first step, we sample the combination of prompt types ( clicks, scribbles, bounding boxes) and then sample the number of positive and negative interactions for each type. In each subsequent steps, we simulate either scribble corrections or click corrections with equal probability. We selected these values based on what we believe to be reasonable interactions for a user to perform.", + "bbox": [ + 212, + 207, + 784, + 340 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Baselines. We compare to existing generalist methods for interactive segmentation, with a focus on methods developed for biomedical imaging.", + "bbox": [ + 212, + 342, + 782, + 371 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "SAM [47]: We evaluate the smallest (ViT-b) and largest (ViT-h) versions of the Segment Anything Model (SAM) trained on natural images. SAM takes bounding boxes, clicks, and the logits of the previous prediction as input.", + "bbox": [ + 212, + 372, + 782, + 417 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "SAM-Med2D [19]: SAM-Med2D is a SAM ViT-b model with additional adapter layers in the image encoder. SAM-Med2D was fine-tuned, using bounding boxes and iterative clicks as input, on a collection of biomedical imaging datasets containing 4.6M images and 19.7M segmentations. Following [19], we evaluate SAM-Med2D both with and without adapter layers.", + "bbox": [ + 212, + 417, + 782, + 492 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "MedSAM [69]: MedSAM is a SAM ViT-B model fine-tuned with bounding box prompts on a collection of biomedical imaging datasets containing 1.5M image segmentation pairs. We evaluate MedSAM with bounding boxes only, because we found it to perform poorly when given point or mask prompts.", + "bbox": [ + 212, + 493, + 782, + 553 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "MIDepSeg [68]: MIDepSeg is an interactive segmentation framework designed to generalize to unseen tasks on medical images. MIDepSeg takes interior margin points (positive clicks) as initial inputs, crops the image based on those points, and uses a CNN to make an initial prediction. Given additional clicks, the prediction can be refined using conditional random fields. We evaluate the pre-trained model, which was developed on placenta T2 MRI.", + "bbox": [ + 212, + 554, + 782, + 643 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 Evaluation", + "text_level": 1, + "bbox": [ + 214, + 664, + 356, + 681 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We evaluate all methods using both manual and simulated interactions, with a focus on scribbles. Simulated interactions enable us to test on a large volume of images and tasks. However, simulations do not always match user behavior. We address this by (1) collecting a diverse dataset of manual scribbles for evaluation, and (2) conducting an interactive user study.", + "bbox": [ + 212, + 693, + 782, + 768 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.1 Manual Scribbles", + "text_level": 1, + "bbox": [ + 214, + 787, + 406, + 801 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Setup. We evaluate on two datasets of manual scribbles. First, we collected a dataset (MedScribble) of manual scribbles from three annotators for seven", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "ScribblePrompt", + "bbox": [ + 624, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "segmentation tasks from unseen datasets, with a total of 31 examples. The annotators were shown five training examples with the ground truth segmentation per task and instructed to draw positive and negative scribbles to indicate the region of interest on new images. We provide more details on MedScribble and release the data in Supplementary Material Sec. E.1. Second, we used 380 slices from the ACDC dataset [10], which has scribble annotations for three labels and background.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In this evaluation, we make a prediction given a set of positive and negative scribbles, simulating a non-iterative scenario where the annotator draws several scribbles before running inference. We evaluate accuracy using Dice score [24] and 95th percentile Hausdorff Distance [42]. Since MIDepSeg's initial prediction network only takes positive inputs, we report results on both MIDepseg's initial prediction from the positive scribbles, and after applying MIDepSeg's refinement procedure using the negative scribbles as corrections. For SAM and variants fine-tuning SAM on clicks, we convert each scribble to a series of points. For MedSAM, we fit a bounding box to the positive scribbles.", + "bbox": [ + 212, + 252, + 787, + 388 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Results. ScribblePrompt-UNet and ScribblePrompt-SAM produce the most accurate segmentations in a single step of manual scribbles on both our manual scribbles dataset and the ACDC scribbles dataset (Tab. 1). We show example predictions in Fig. 5 and additional comparisons to scribble-supervised learning in Supplementary Material Sec. E.3.", + "bbox": [ + 212, + 388, + 787, + 463 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "SAM and SAM-Med2D do not generalize well to scribble inputs (which they were not trained for). MedSAM has better predictions than other SAM baselines using the SAM architecture, however it is not able to make use of the negative scribbles and thus often misses segmentations with holes in them (Fig. 5). The initial predictions from MIDepSeg's network are poor, but improve after applying the refinement procedure.", + "bbox": [ + 212, + 465, + 787, + 555 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/619a8b4fd09eb94904b1ce3f8397b5c1328e96440ec84151823e05e565c31d5c.jpg", + "table_caption": [ + "Table 1: Manual scribbles. Mean Dice and HD95 with $95\\%$ CI of segmentations predicted from manually-collected scribbles. Best and second best are highlighted." + ], + "table_footnote": [], + "table_body": "
MedScribble (n=93)ACDC (n=1,140)
Model↑ Dice Score↓ HD95↑ Dice Score↓ HD95
SAM (ViT-b)0.40 ± 0.0532.58 ± 3.800.20 ± 0.01108.39 ± 0.85
SAM (ViT-h)0.56 ± 0.0514.61 ± 3.180.42 ± 0.0259.78 ± 2.02
SAM-Med2D w.o. adapter0.52 ± 0.0530.34 ± 3.290.16 ± 0.01107.25 ± 1.12
SAM-Med2D w/ adapter0.30 ± 0.0619.04 ± 3.630.17 ± 0.0241.01 ± 3.04
MIDepSeg0.68 ± 0.046.94 ± 1.060.58 ± 0.015.78 ± 0.26
MIDepSeg w/ refinement0.81 ± 0.033.10 ± 0.670.73 ± 0.012.71 ± 0.18
MedSAM (box)0.70 ± 0.047.54 ± 1.350.70 ± 0.023.53 ± 0.28
ScribblePrompt-UNet0.84 ± 0.022.92 ± 0.880.84 ± 0.011.80 ± 0.11
ScribblePrompt-SAM0.87 ± 0.022.61 ± 0.840.77 ± 0.014.17 ± 0.32
", + "bbox": [ + 225, + 625, + 782, + 808 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "H.E. Wong et al.", + "bbox": [ + 271, + 114, + 383, + 128 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/3292108fb40a947b0416d1f0645e8d081c5ec96003e686f8297a372cfa1ab0f7.jpg", + "image_caption": [ + "Fig. 5: Example predictions. $\\mathrm{SP} =$ ScribblePrompt. Top: predictions after one step of manual scribbles. Bottom: predictions after five steps of simulated interactions (one center click followed by one center correction click per step)." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 285, + 263 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/2fcc093f3f1f6718de6b6abc8a7df8529e46cb0325fd4a8fa07ab84a9a343ae0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 289, + 145, + 356, + 263 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/f1e5ae55bbdcac26410ed052ab558e615b3d94cc7dcdac3370823f4c9108ebb8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 143, + 426, + 263 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/ef871048b2e18ae987d1724e68b5795e6b0afca13110736b76f522527d6f33a7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 431, + 145, + 500, + 263 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/7c93f90af4aa2ef573d2d05f948e5dc1311c8cf2cb8357acdb0b24b68156111d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 143, + 571, + 263 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/8d22d536c9e4881fcd9efd47dfef9a64ce92f5ad9a957240b4d522d3a265a793.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 576, + 143, + 642, + 263 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/caa0c9f793687c0de64da82217ca4cf3039054582eb7b0ae7980687e58730dc4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 145, + 712, + 263 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/2b0fdeff4d3a48c74be1b984389f29f44605ad234aa7f51cc193c3080612cd29.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 717, + 143, + 784, + 263 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.2 Simulated Interactions", + "text_level": 1, + "bbox": [ + 215, + 345, + 449, + 359 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Setup. We use our interaction simulator to evaluate the performance of ScribblePrompt and baselines with iterative scribbles across the 12 evaluation sets.", + "bbox": [ + 214, + 371, + 782, + 401 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "For scribble-focused prompting procedures, we use:", + "bbox": [ + 238, + 402, + 609, + 417 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Line Scribbles: Three positive line scribbles and three negative line scribbles to start, followed by one correction line scribble at each step.", + "- Centerline Scribbles: One positive and one negative centerline scribble to start, followed by one correction centerline scribble at each step.", + "- Contour Scribbles: One positive and one negative contour scribble to start, followed by one correction contour scribble at each step." + ], + "bbox": [ + 215, + 426, + 784, + 516 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We limit each scribble to cover a maximum of $w$ pixels, where $w$ is the maximum dimension of the image. For centerline and contour scribbles, a scribble might contain multiple disconnected components.", + "bbox": [ + 214, + 526, + 782, + 571 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "For click-focused prompting procedures, we use:", + "bbox": [ + 238, + 571, + 584, + 585 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Center Clicks: One positive click in the center of the largest component to start, followed by one (positive or negative) correction click per step in the center of the largest component of the error region.", + "- Random Clicks: One random positive click to start, followed by one (positive or negative) correction click per step, randomly sampled from the error region.", + "- Random Warm Start: Three random positive clicks and three random negative clicks to start, followed by one (positive or negative) correction click per step in the center of the largest component of the error region." + ], + "bbox": [ + 215, + 595, + 784, + 717 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "For each example and interaction procedure, we simulate a series of iterative interactions with five random seeds. Since MIDepSeg cannot make accurate predictions with only a few clicks, we exclude it from simulations that start with a single click. For MedSAM, we show results for a bounding box prompt.", + "bbox": [ + 214, + 727, + 782, + 787 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Results. Fig. 6 shows that both versions of ScribblePrompt outperform baseline methods for all simulated interaction procedures at all numbers of interactions.", + "bbox": [ + 214, + 809, + 782, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "ScribblePrompt", + "bbox": [ + 624, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/c30bd115037a18ea99ff9074afc0a7ffa858010d9cca4e27e2d0456919419029.jpg", + "image_caption": [ + "Fig. 6: Simulated clicks and scribbles. We simulate interactions following three scribble protocols and three click protocols. We show more results and example predictions, with similar trends, in Supplementary Material Sec. F.2." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 362, + 257 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/9d2bcb66fd3998c3665e198dbd1bfdba0347eddebf39b489768bc0591ec7325c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 143, + 501, + 257 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/f779e3bebd9b9bf6da2030411ffa0e05b1892f22d6a3fe7f55ee4f785d66ba42.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 143, + 643, + 257 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/54ea89099ce6f11cea8c1c5fb30574d71e868d67ad1ea77a2ed082a87ddd6349.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 143, + 785, + 257 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We show examples with simulated interactions in Fig. 1 and Fig. 5 (bottom). We show results with similar trends using bounding boxes and by dataset with fully-supervised baselines in Supplementary Material Sec. F.", + "bbox": [ + 212, + 335, + 782, + 383 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Discussion. Comparing the simulated scribble and click results (Fig. 6) highlights the efficiency of scribble-based interactions. Both ScribblePrompt models reach an average Dice above 0.8 in two scribble steps. It takes five steps of carefully-placed center clicks or eight random clicks to reach the same average Dice score with ScribblePrompt-SAM.", + "bbox": [ + 212, + 402, + 784, + 477 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Although SAM-Med2D was trained on a large biomedical imaging collection, it does not perform as well as SAM at refining its predictions on held-out evaluation sets. MedSAM only uses bounding box prompts and is not able to refine predictions, limiting its performance and usability.", + "bbox": [ + 212, + 479, + 784, + 540 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.3 User Study", + "text_level": 1, + "bbox": [ + 214, + 559, + 356, + 574 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To assess the practical utility of ScribblePrompt, we conducted a user study with experienced annotators. We compare ScribblePrompt-UNet to SAM (ViT-b). We selected ScribblePrompt-UNet for the user study because it had similar performance and lower latency compared to ScribblePrompt-SAM. We selected SAM (ViT-b) because it had the highest Dice score in our experiments with clicks and lower latency than the next closest baseline, SAM (ViT-h).", + "bbox": [ + 212, + 583, + 784, + 672 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Setup. Study participants ( $n = 16$ ) were neuroimaging researchers at an academic hospital. They used each model to segment a series of nine test images from different evaluation datasets. For each segmentation task, participants were shown the target segmentation and were asked to prompt the model until the predicted segmentation closely matched the target, or they could no longer improve the prediction. We provided participants with the target segmentation to disentangle the cognitive process of identifying the region of interest from prompting the model to achieve the desired segmentation. We provide additional details and visualizations in Supplementary Material Sec. G.", + "bbox": [ + 212, + 674, + 785, + 808 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Results. Participants produced more accurate segmentations (0.84 vs. 0.73 Dice; $p = 0.001$ using a paired t-test) using ScribblePrompt (Tab. 2). Partici-", + "bbox": [ + 214, + 809, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "H.E. Wong et al.", + "bbox": [ + 271, + 114, + 383, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/70b90ceead4ab70e81fc32d8d05a635d9b15914b5c022bc6db6fa29d9406fcd3.jpg", + "table_caption": [ + "Table 2: User study. Mean ± std. for Dice score, HD95 and total time per segmentation. Median (and max) number of times the participants refreshed the prediction." + ], + "table_footnote": [], + "table_body": "
Dice Score (↑)HD95 (↓)Sec./task (↓)Iter./task (↓)
SAM (ViT-b)0.73 ± 0.3016.1 ± 37.1131.7 ± 168.56 (47)
ScribblePrompt-UNet0.84 ± 0.134.5 ± 9.194.2 ± 89.03 (19)
", + "bbox": [ + 253, + 184, + 753, + 242 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "pants spent $\\approx 1.5$ minutes per segmentation on average using ScribblePrompt, compared to over 2 minutes per segmentation with SAM ( $p = 0.02$ using a paired t-test). While using ScribblePrompt, participants updated the prediction fewer times before being satisfied with the segmentation.", + "bbox": [ + 212, + 268, + 782, + 330 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Upon completion, 15 out of 16 participants reported they preferred using ScribblePrompt to SAM and one participant had no preference. All participants reported it was easier to achieve the target segmentation using ScribblePrompt compared to SAM. $93.8\\%$ of participants reported ScribblePrompt was better than SAM at refining its predictions in response to scribbles. $87.5\\%$ of participants preferred using ScribblePrompt over SAM for clicks.", + "bbox": [ + 212, + 330, + 784, + 419 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Discussion. The most common factors that influenced participant preference for ScribblePrompt were 1) being able get accurate predictions from multiple types of prompts, including scribbles and 2) responsiveness to their corrections, which enabled fine-grained control over the next prediction. For some tasks, such as retinal vein segmentation, SAM was unable to make accurate predictions even with many corrections. As a result, there was more variability in Dice score and time per segmentation, for SAM than ScribblePrompt.", + "bbox": [ + 212, + 421, + 784, + 526 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.4 Inference Runtime", + "text_level": 1, + "bbox": [ + 215, + 549, + 415, + 561 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Setup. We evaluate computational efficiency by measuring inference time on a single CPU for one prediction with a scribble input covering 128 pixels. Performance on a CPU reflects practical utility. Because of a variety of barriers such as protected health information, users may not be able to send their data to a server for computation, and must rely on local, most often CPU-only, computing.", + "bbox": [ + 212, + 573, + 784, + 648 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Results. On a single CPU, ScribblePrompt-UNet requires $0.27 \\pm 0.04$ sec per prediction, enabling the model to be used even in low-resource environments. The next most accurate model (Fig. 6), SAM ViT-h, requires over 2 minutes per prediction on a CPU ( $130.79 \\pm 7.96$ sec). SAM ViT-b takes $13.59 \\pm 0.77$ seconds per prediction. We report results for all baselines and on GPU hardware with similar trends in Supplementary Material Sec. H.", + "bbox": [ + 212, + 648, + 784, + 739 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "6 Ablations", + "text_level": 1, + "bbox": [ + 215, + 762, + 346, + 777 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We conduct two ablations of important design decisions: (1) synthetic label inputs used during training, and (2) types of prompts simulated during training. We report results on the validation splits of nine datasets unseen during training.", + "bbox": [ + 212, + 794, + 782, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "ScribblePrompt", + "bbox": [ + 624, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/12d889d1146f83e3028f91524b33199dfd5b2f5e9b6906932f7b908a2280af33.jpg", + "image_caption": [ + "(a) Probability of synthetic labels." + ], + "image_footnote": [], + "bbox": [ + 217, + 147, + 441, + 224 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/2a30fcd60f9694902e6ee937ad81de14cbe0f8544de94ab59c8db3407cade487.jpg", + "image_caption": [ + "(b) Interactions during training.", + "Fig. 7: Ablations. We report Dice after five steps of simulated interactions following three inference-time interaction procedures. Error bars show $95\\%$ CI. (a) shows mean change in Dice relative to ScribblePrompt-UNet trained without any synthetic labels." + ], + "image_footnote": [], + "bbox": [ + 459, + 143, + 779, + 224 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6.1 Synthetic Labels", + "text_level": 1, + "bbox": [ + 214, + 321, + 401, + 335 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Fig. 7a shows the effect of varying the probability of sampling a synthetic label during training for ScribblePrompt-UNet. Training with both real and synthetic labels improves generalization to new datasets, compared to training with only real labels. Using $p_{\\text{synth}} = 0.5$ results in the highest Dice. Training with only synthetic labels results in worse Dice scores. We show similar results for ScribblePrompt-SAM and other interactions in Supplementary Material Sec. I.1.", + "bbox": [ + 212, + 343, + 787, + 434 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6.2 Prompt Types", + "text_level": 1, + "bbox": [ + 214, + 454, + 383, + 469 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Setup. We evaluate ScribblePrompt-UNet models trained with different combinations of prompts, compared to the complete ScribblePrompt-UNet:", + "bbox": [ + 212, + 476, + 785, + 506 + ], + "page_idx": 13 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- ScribblePrompt-Unet (scribbles) trained on boxes and scribbles.", + "- ScribblePrompt-UNet (clicks) trained on boxes and clicks.", + "- ScribblePrompt-Unet (random clicks) trained on boxes and random clicks." + ], + "bbox": [ + 214, + 512, + 782, + 570 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Results. Fig. 7b shows ScribblePrompt-UNet trained with scribbles, clicks and bounding boxes predicts segmentations more accurately than do ablated versions of ScribblePrompt-UNet. We show results for other interaction procedures with similar trends in Supplementary Material Sec. I.2.", + "bbox": [ + 212, + 578, + 787, + 640 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "7 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 660, + 359, + 676 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We present ScribblePrompt, a practical framework for interactive segmentation that enables users to segment diverse medical images with scribbles, clicks, and bounding boxes. We introduce methods for simulating realistic user interactions and generating synthetic labels. These methods enable us to train models that generalize to unseen segmentation tasks and datasets. ScribblePrompt is more accurate than existing baselines, and ScribblePrompt-UNet is computationally efficient, even on a CPU. Our user study shows that nearly all users prefer ScribblePrompt and achieve segmentations with $15\\%$ higher Dice with less effort than the next most accurate baseline. ScribblePrompt promises to significantly reduce the burden of manual segmentation in biomedical imaging.", + "bbox": [ + 212, + 689, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "H.E. Wong et al.", + "bbox": [ + 271, + 114, + 387, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported in part by funding from Quanta Computer Inc., the Eric and Wendy Schmidt Center at the Broad Institute of MIT and Harvard and the Wistron Corporation. Research reported in this paper was supported by the National Institute of Biomedical Imaging and Bioengineering of the National Institutes of Health under award number R01EB033773. Much of the computation resources required for this research was performed on computational hardware generously provided by the Massachusetts Life Sciences Center.", + "bbox": [ + 212, + 174, + 787, + 282 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 301, + 321, + 316 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Abdi, A.H., Kasaei, S., Mehdizadeh, M.: Automatic segmentation of mandible in panoramic x-ray. Journal of Medical Imaging 2(4), 044003 (2015)", + "2. Agustsson, E., Uijlings, J.R., Ferrari, V.: Interactive Full Image Segmentation by Considering All Regions Jointly. In: 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11614-11623. IEEE, Long Beach, CA, USA (Jun 2019). https://doi.org/10.1109/CVPR.2019.01189, https://ieeexplore.ieee.org/document/8953526/", + "3. Aine, C.J., Bockholt, H.J., Bustillo, J.R., Canive, J.M., Caprihan, A., Gasparovic, C., Hanlon, F.M., Houck, J.M., Jung, R.E., Lauriello, J., Liu, J., Mayer, A.R., Perrone-Bizzozero, N.I., Posse, S., Stephen, J.M., Turner, J.A., Clark, V.P., Calhoun, V.D.: Multimodal Neuroimaging in Schizophrenia: Description and Dissemination. Neuroinformatics 15(4), 343-364 (Oct 2017). https://doi.org/10.1007/s12021-017-9338-9, http://link.springer.com/10.1007/s12021-017-9338-9", + "4. Al-Dhabyani, W., Gomaa, M., Khaled, H., Fahmy, A.: Dataset of breast ultrasound images. Data in Brief 28, 104863 (2020). https://doi.org/https://doi.org/10.1016/j.dib.2019.104863, https://www.sciencedirect.com/science/article/pii/S2352340919312181", + "5. Asad, M., Fidon, L., Vercauteren, T.: ECONet: Efficient Convolutional Online Likelihood Network for Scribble-based Interactive Segmentation. In: International Conference on Medical Imaging with Deep Learning. pp. 35-47. PMLR (2022), http://arxiv.org/abs/2201.04584, arXiv:2201.04584 [cs, eess]", + "6. Atzeni, A., Peter, L., Robinson, E., Blackburn, E., Althonayan, J., Alexander, D.C., Iglesias, J.E.: Deep active learning for suggestive segmentation of biomedical image stacks via optimisation of Dice scores and traced boundary length. Medical Image Analysis 81 (2022)", + "7. Bai, J., Wu, X.: Error-Tolerant Scribbles Based Interactive Image Segmentation. In: 2014 IEEE Conference on Computer Vision and Pattern Recognition. pp. 392-399. IEEE, Columbus, OH, USA (Jun 2014). https://doi.org/10.1109/CVPR.2014.57, https://ieeexplore.ieee.org/document/6909451", + "8. Baid, U., Ghodasara, S., Mohan, S., Bilello, M., Calabrese, E., Colak, E., Farahani, K., Kalpathy-Cramer, J., Kitamura, F.C., Pati, S., et al.: The rsna-asnrmiccai brats 2021 benchmark on brain tumor segmentation and radiogenomic classification. arXiv preprint arXiv:2107.02314 (2021)", + "9. Benenson, R., Popov, S., Ferrari, V.: Large-scale interactive object segmentation with human annotators. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11700-11709 (2019)" + ], + "bbox": [ + 230, + 330, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "ScribblePrompt", + "bbox": [ + 624, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "10. Bernard, O., Lalande, A., Zotti, C., Cervenansky, F., Yang, X., Heng, P.A., Cetin, I., Lekadir, K., Camara, O., Ballester, M.A.G., et al.: Deep learning techniques for automatic mri cardiac multi-structures segmentation and diagnosis: is the problem solved? IEEE transactions on medical imaging 37(11), 2514-2525 (2018)", + "11. Bilic, P., Christ, P.F., Vorontsov, E., Chlebus, G., Chen, H., Dou, Q., Fu, C.W., Han, X., Heng, P.A., Hesser, J., et al.: The liver tumor segmentation benchmark (lits). arXiv preprint arXiv:1901.04056 (2019)", + "12. Bloch, N., Madabhushi, A., Huisman, H., Freymann, J., Kirby, J., Grauer, M., Enquobahrie, A., Jaffe, C., Clarke, L., Farahani, K.: Nci-isbi 2013 challenge: automated segmentation of prostate structures. The Cancer Imaging Archive 370(6), 5 (2015)", + "13. Boykov, Y., Jolly, M.P.: Interactive graph cuts for optimal boundary & region segmentation of objects in N-D images. In: Proceedings Eighth IEEE International Conference on Computer Vision. ICCV 2001. vol. 1, pp. 105-112. IEEE Comput. Soc, Vancouver, BC, Canada (2001). https://doi.org/10.1109/ICCV.2001.937505, http://ieeexplore.ieee.org/document/937505/", + "14. Bredell, G., Tanner, C., Konukoglu, E.: Iterative interaction training for segmentation editing networks. In: Machine Learning in Medical Imaging: 9th International Workshop, MLMI 2018, Held in Conjunction with MICCAI 2018, Granada, Spain, September 16, 2018, Proceedings 9. pp. 363-370. Springer (2018)", + "15. Butoi*, V.I., Ortiz*, J.J.G., Ma, T., Sabuncu, M.R., Guttag, J., Dalca, A.V.: Universeg: Universal medical image segmentation. In: ICCV (2023)", + "16. Caicedo, J.C., Goodman, A., Karhohs, K.W., Cimini, B.A., Ackerman, J., Haghighi, M., Heng, C., Becker, T., Doan, M., McQuin, C., Rohhan, M., Singh, S., Carpenter, A.E.: Nucleus segmentation across imaging experiments: the 2018 Data Science Bowl. Nature Methods 16(12), 1247-1253 (Dec 2019). https://doi.org/10.1038/s41592-019-0612-7, https://doi.org/10.1038/s41592-019-0612-7", + "17. Cardona, A., Saalfeld, S., Preibisch, S., Schmid, B., Cheng, A., Pulokas, J., Tomancak, P., Hartenstein, V.: An integrated micro-and macroarchitectural analysis of the drosophila brain by computer-assisted serial section electron microscopy. PLoS biology 8(10), e1000502 (2010)", + "18. Chen, X., Cheung, Y.S.J., Lim, S.N., Zhao, H.: ScribbleSeg: Scribble-based Interactive Image Segmentation (Mar 2023), http://arxiv.org/abs/2303.11320, arXiv:2303.11320 [cs]", + "19. Cheng, J., Ye, J., Deng, Z., Chen, J., Li, T., Wang, H., Su, Y., Huang, Z., Chen, J., Jiang, L., Sun, H., He, J., Zhang, S., Zhu, M., Qiao, Y.: SAM-Med2D (Aug 2023), http://arxiv.org/abs/2308.16184, arXiv:2308.16184 [cs]", + "20. Codella, N.C.F., Gutman, D.A., Celebi, M.E., Helba, B., Marchetti, M.A., Dusza, S.W., Kalloo, A., Liopyris, K., Mishra, N.K., Kittler, H., Halpern, A.: Skin lesion analysis toward melanoma detection: A challenge at the 2017 international symposium on biomedical imaging (isbi), hosted by the international skin imaging collaboration (ISIC). CoRR abs/1710.05006 (2017), http://arxiv.org/abs/1710.05006", + "21. Criminisi, A., Sharp, T., Blake, A.: GeoS: Geodesic Image Segmentation. In: Forsyth, D., Torr, P., Zisserman, A. (eds.) Computer Vision - ECCV 2008, vol. 5302, pp. 99-112. Springer Berlin Heidelberg, Berlin, Heidelberg (2008). https://doi.org/10.1007/978-3-540-88682-2_9, http://link.springer.com/10.1007/978-3-540-88682-2_9, series Title: Lecture Notes in Computer Science" + ], + "bbox": [ + 225, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "H.E. Wong et al.", + "bbox": [ + 271, + 114, + 383, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "22. Czolbe, S., Dalca, A.V.: Neuralizer: General neuroimage analysis without retraining. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6217-6230 (2023)", + "23. Dalca, A.V., Guttag, J., Sabuncu, M.R.: Anatomical priors in convolutional networks for unsupervised biomedical segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 9290-9299 (2018)", + "24. Dice, L.R.: Measures of the amount of ecologic association between species. Ecology 26(3), 297-302 (1945)", + "25. Felzenszwalb, P.F., Huttenlocher, D.P.: Efficient graph-based image segmentation. International journal of computer vision 59, 167-181 (2004)", + "26. Fischl, B.: Freesurfer. Neuroimage 62(2), 774-781 (2012)", + "27. Forte, M., Price, B., Cohen, S., Xu, N., Pitie, F.: Getting to $99\\%$ Accuracy in Interactive Segmentation (Mar 2020), http://arxiv.org/abs/2003.07932, arXiv:2003.07932 [cs]", + "28. Gamper, J., Koohbanani, N., Benes, K., Graham, S., Jahanifar, M., Khurram, S., Azam, A., Hewitt, K., Rajpoot, N.: Pannuke dataset extension, insights and baselines. arxiv. 2020 doi: 10.48550. ARXIV (2003)", + "29. Gerhard, S., Funke, J., Martel, J., Cardona, A., Fetter, R.: Segmented anisotropic ssTEM dataset of neural tissue (11 2013). https://doi.org/10.6084/m9.figshare.856713.v1, https://figshare.com/articles/dataset/Segmented_anisotropic_ssTEM_dataset_of_neural_tissue/856713", + "30. van Ginneken, B., Stegmann, M.B., Loog, M.: Segmentation of anatomical structures in chest radiographs using supervised methods: a comparative study on a public database. Medical Image Analysis 10(1), 19-40 (2006). https://doi.org/https://doi.org/10.1016/j.media.2005.02.002, https://www.sciencedirect.com/science/article/pii/S1361841505000368", + "31. Gollub, R.L., Shoemaker, J.M., King, M.D., White, T., Ehrlich, S., Sponheim, S.R., Clark, V.P., Turner, J.A., Mueller, B.A., Magnotta, V., et al.: The mcic collection: a shared repository of multi-modal, multi-site brain image data from a clinical investigation of schizophrenia. Neuroinformatics 11, 367-388 (2013)", + "32. Gotkowski, K., Lüth, C., Jäger, P.F., Ziegler, S., Krämer, L., Denner, S., Xiao, S., Disch, N., Maier-Hein, K.H., Isensee, F.: Embarrassingly simple scribble supervision for 3d medical segmentation. arXiv preprint arXiv:2403.12834 (2024)", + "33. Grady, L.: Random Walks for Image Segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence 28(11), 1768-1783 (Nov 2006). https://doi.org/10.1109/TPAMI.2006.233, http://ieeexplore.ieee.org/document/1704833/", + "34. Gut, D.: X-ray images of the hip joints 1 (Jul 2021). https://doi.org/10.17632/zm6bxzhmfz.1, https://data.mendeley.com/datasets/zm6bxzhmfz/1, publisher: Mendeley Data", + "35. He, K., Zhang, X., Ren, S., Sun, J.: Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In: Proceedings of the IEEE international conference on computer vision. pp. 1026-1034 (2015)", + "36. He, S., Bao, R., Li, J., Stout, J., Bjornerud, A., Grant, P.E., Ou, Y.: Computer- vision benchmark segment-anything model (sam) in medical images: Accuracy in 12 datasets (2023)", + "37. Heller, N., Isensee, F., Maier-Hein, K.H., Hou, X., Xie, C., Li, F., Nan, Y., Mu, G., Lin, Z., Han, M., et al.: The state of the art in kidney and kidney tumor segmentation in contrast-enhanced ct imaging: Results of the kits19 challenge. Medical Image Analysis p. 101821 (2020)" + ], + "bbox": [ + 225, + 146, + 784, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "ScribblePrompt", + "bbox": [ + 624, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "38. Hernandez Petzsche, M.R., de la Rosa, E., Hanning, U., Wiest, R., Valenzuela, W., Reyes, M., Meyer, M., Liew, S.L., Kofler, F., Ezhov, I., et al.: Isles 2022: A multi-center magnetic resonance imaging stroke lesion segmentation dataset. Scientific data 9(1), 762 (2022)", + "39. Hoover, A., Kouznetsova, V., Goldbaum, M.: Locating blood vessels in retinal images by piecewise threshold probing of a matched filter response. IEEE Transactions on Medical imaging 19(3), 203-210 (2000)", + "40. Hu, X., Xu, X., Shi, Y.: How to efficiently adapt large segmentation model(sam) to medical images (2023)", + "41. Huang, Y., Yang, X., Liu, L., Zhou, H., Chang, A., Zhou, X., Chen, R., Yu, J., Chen, J., Chen, C., Chi, H., Hu, X., Fan, D.P., Dong, F., Ni, D.: Segment anything model for medical images? (2023)", + "42. Huttenlocher, D.P., Klanderman, G.A., Rucklidge, W.J.: Comparing images using the hausdorff distance. IEEE Transactions on pattern analysis and machine intelligence 15(9), 850-863 (1993)", + "43. Isensee, F., Jaeger, P.F., Kohl, S.A.A., Petersen, J., Maier-Hein, K.H.: nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nature Methods 18(2), 203-211 (Feb 2021). https://doi.org/10.1038/s41592-020-01008-z, http://www.nature.com/articles/s41592-020-01008-z", + "44. Ji, Y., Bai, H., Yang, J., Ge, C., Zhu, Y., Zhang, R., Li, Z., Zhang, L., Ma, W., Wan, X., et al.: Amos: A large-scale abdominal multi-organ benchmark for versatile medical image segmentation. arXiv preprint arXiv:2206.08023 (2022)", + "45. Kavur, A.E., Gezer, N.S., Baris, M., Aslan, S., Conze, P.H., Groza, V., Pham, D.D., Chatterjee, S., Ernst, P., Ozkan, S., Baydar, B., Lachinov, D., Han, S., Pauli, J., Isensee, F., Perkonigg, M., Sathish, R., Rajan, R., Sheet, D., Dovletov, G., Speck, O., Nurnberger, A., Maier-Hein, K.H., Bozdagi Akar, G., Ünal, G., Dicle, O., Selver, M.A.: CHAOS Challenge - combined (CT-MR) healthy abdominal organ segmentation. Medical Image Analysis 69, 101950 (2021). https://doi.org/https://doi.org/10.1016/j.media.2020.101950, https://www.sciencedirect.com/science/article/pii/S1361841520303145", + "46. Kim, S., Oh, H.J., Min, S., Jeong, W.K.: Evaluation and improvement of segment anything model for interactive histopathology image segmentation (2023)", + "47. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. In: ICCV (2023)", + "48. Krönke, M., Eilers, C., Dimova, D., Köhler, M., Buschner, G., Schweiger, L., Konstantinidou, L., Makowski, M., Nagarajah, J., Navab, N., et al.: Tracked 3d ultrasound and deep neural network-based thyroid segmentation reduce interobserver variability in thyroid volumetry. Plos one 17(7), e0268550 (2022)", + "49. Kuijf, H.J., Biesbroek, J.M., De Bresser, J., Heinen, R., Andermatt, S., Bento, M., Berseth, M., Belyaev, M., Cardoso, M.J., Casamitjana, A., et al.: Standardized assessment of automatic segmentation of white matter hyperintensities and results of the wmh segmentation challenge. IEEE transactions on medical imaging 38(11), 2556-2568 (2019)", + "50. Kuklisova-Murgasova, M., Aljabar, P., Srinivasan, L., Counsell, S.J., Doria, V., Serag, A., Gousias, I.S., Boardman, J.P., Rutherford, M.A., Edwards, A.D., et al.: A dynamic 4d probabilistic atlas of the developing brain. NeuroImage 54(4), 2750-2763 (2011)", + "51. Lambert, Z., Petitjean, C., Dubray, B., Kuan, S.: Segthor: segmentation of thoracic organs at risk in ct images. In: 2020 Tenth International Conference on Image Processing Theory, Tools and Applications (IPTA). pp. 1-6. IEEE (2020)" + ], + "bbox": [ + 225, + 147, + 784, + 840 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "H.E. Wong et al.", + "bbox": [ + 271, + 114, + 383, + 128 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "52. Landman, B., Xu, Z., Igelsias, J., Styner, M., Langerak, T., Klein, A.: Miccai multi-atlas labeling beyond the cranial vault-workshop and challenge. In: Proc. MICCAI Multi-Atlas Labeling Beyond Cranial Vault Workshop Challenge. vol. 5, p. 12 (2015)", + "53. Lee, H., Jeong, W.K.: Scribble2label: Scribble-supervised cell segmentation via self-generating pseudo-labels with consistency. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part I 23. pp. 14-23. Springer (2020)", + "54. Lemaître, G., Martí, R., Freixenet, J., Vilanova, J.C., Walker, P.M., Meriaudeau, F.: Computer-aided detection and diagnosis for prostate cancer based on mono and multi-parametric mri: a review. Computers in biology and medicine 60, 8-31 (2015)", + "55. Li, M., Zhang, Y., Ji, Z., Xie, K., Yuan, S., Liu, Q., Chen, Q.: Ipn-v2 and octa-500: Methodology and dataset for retinal image segmentation. arXiv preprint arXiv:2012.07261 (2020)", + "56. Li, Z., Zheng, Y., Luo, X., Shan, D., Hong, Q.: Scribblevc: Scribble-supervised medical image segmentation with vision-class embedding. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 3384-3393 (2023)", + "57. Lin, D., Dai, J., Jia, J., He, K., Sun, J.: Scribblesup: Scribble-supervised convolutional networks for semantic segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3159-3167 (2016)", + "58. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proceedings of the IEEE international conference on computer vision. pp. 2980-2988 (2017)", + "59. Lin, X., Xiang, Y., Zhang, L., Yang, X., Yan, Z., Yu, L.: Samus: Adapting segment anything model for clinically-friendly and generalizable ultrasound image segmentation (2023)", + "60. Litjens, G., Toth, R., van de Ven, W., Hoeks, C., Kerkstra, S., van Ginneken, B., Vincent, G., Guillard, G., Birbeck, N., Zhang, J., et al.: Evaluation of prostate segmentation algorithms for mri: the promise12 challenge. Medical image analysis 18(2), 359-373 (2014)", + "61. Liu, Q., Xu, Z., Bertasius, G., Niethammer, M.: Simpleclick: Interactive image segmentation with simple vision transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22290-22300 (2023)", + "62. Liu, Q., Xu, Z., Jiao, Y., Niethammer, M.: iSegFormer: Interactive Segmentation via Transformers with Application to 3D Knee MR Images. In: Medical Image Computing and Computer Assisted Intervention (MICCAI) (2022)", + "63. Ljosa, V., Sokolnicki, K.L., Carpenter, A.E.: Annotated high-throughput microscopy image sets for validation. Nature methods 9(7), 637-637 (2012)", + "64. Löffler, M.T., Sekuboyina, A., Jacob, A., Grau, A.L., Scharr, A., El Husseinini, M., Kallweit, M., Zimmer, C., Baum, T., Kirschke, J.S.: A vertebral segmentation dataset with fracture grading. Radiology: Artificial Intelligence 2(4), e190138 (2020)", + "65. in the Loop, H.: Teeth segmentation dataset, https://humansintheloop.org/resources/datasets/teeth-segmentation-dataset/", + "66. Luo, X., Hu, M., Liao, W., Zhai, S., Song, T., Wang, G., Zhang, S.: Scribblesupervised medical image segmentation via dual-branch network and dynamically mixed pseudo labels supervision. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 528-538. Springer (2022)" + ], + "bbox": [ + 225, + 146, + 784, + 839 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "ScribblePrompt", + "bbox": [ + 624, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "67. Luo, X., Liao, W., Xiao, J., Song, T., Zhang, X., Li, K., Wang, G., Zhang, S.: Word: Revisiting organs segmentation in the whole abdominal region. arXiv preprint arXiv:2111.02403 (2021)", + "68. Luo, X., Wang, G., Song, T., Zhang, J., Aertsen, M., Deprest, J., Ourselin, S., Vercauteren, T., Zhang, S.: MIDeepSeg: Minimally Interactive Segmentation of Unseen Objects from Medical Images Using Deep Learning. Medical Image Analysis 72, 102102 (2021)", + "69. Ma, J., He, Y., Li, F., Han, L., You, C., Wang, B.: Segment anything in medical images. Nature Communications 15, 1-9 (2024)", + "70. Ma, Y., Hao, H., Xie, J., Fu, H., Zhang, J., Yang, J., Wang, Z., Liu, J., Zheng, Y., Zhao, Y.: Rose: a retinal oct-angiography vessel segmentation dataset and new model. IEEE Transactions on Medical Imaging 40(3), 928–939 (2021). https://doi.org/10.1109/TMI.2020.3042802", + "71. Macdonald, J.A., Zhu, Z., Konkel, B., Mazurowski, M., Wiggins, W., Bashir, M.: Duke liver dataset (MRI) v2 (Apr 2023). https://doi.org/10.5281/zenodo.7774566, https://doi.org/10.5281/zenodo.7774566", + "72. Maninis, K.K., Caelles, S., Pont-Tuset, J., Van Gool, L.: Deep Extreme Cut: From Extreme Points to Object Segmentation (2018), arXiv:1711.09081", + "73. Marcus, D.S., Wang, T.H., Parker, J., Csernansky, J.G., Morris, J.C., Buckner, R.L.: Open access series of imaging studies (oasis): cross-sectional mri data in young, middle aged, nondemented, and demented older adults. Journal of cognitive neuroscience 19(9), 1498-1507 (2007)", + "74. Marek, K., Jennings, D., Lasch, S., Siderowf, A., Tanner, C., Simuni, T., Coffey, C., Kieburtz, K., Flagg, E., Chowdhury, S., et al.: The parkinson progression marker initiative (ppmi). Progress in neurobiology 95(4), 629-635 (2011)", + "75. Mazurowski, M.A., Clark, K., Czarnek, N.M., Shamsesfandabadi, P., Peters, K.B., Saha, A.: Radiogenomics of lower-grade glioma: algorithmically-assessed tumor shape is associated with tumor genomic subtypes and patient outcomes in a multi-institutional study with the cancer genome atlas data. Journal of neuro-oncology 133, 27-35 (2017)", + "76. Mazurowski, M.A., Dong, H., Gu, H., Yang, J., Konz, N., Zhang, Y.: Segment anything model for medical image analysis: An experimental study. Medical Image Analysis 89, 102918 (Oct 2023). https://doi.org/10.1016/j.media.2023.102918, https://www.sciencedirect.com/science/article/pii/S1361841523001780", + "77. Men, J., Huang, Y., Solanki, J., Zeng, X., Alex, A., Jerwick, J., Zhang, Z., Tanzi, R.E., Li, A., Zhou, C.: Optical coherence tomography for brain imaging and developmental biology. IEEE Journal of Selected Topics in Quantum Electronics 22(4), 120-132 (2015)", + "78. Milletari, F., Navab, N., Ahmadi, S.A.: V-net: Fully convolutional neural networks for volumetric medical image segmentation. Fourth International Conference on 3D Vision (3DV) pp. 565-571 (2016)", + "79. Montoya, A., Hasnin, kaggle446, shirzad, Cukierski, W., yffud: Ultrasound nerve segmentation (2016), https://kaggle.com/competitions/ultrasound-nerve-segmentation", + "80. Ooi, A.Z.H., Embong, Z., Abd Hamid, A.I., Zainon, R., Wang, S.L., Ng, T.F., Hamzah, R.A., Teoh, S.S., Ibrahim, H.: Interactive Blood Vessel Segmentation from Retinal Fundus Image Based on Canny Edge Detector. Sensors 21(19), 6380 (Sep 2021)" + ], + "bbox": [ + 222, + 146, + 785, + 839 + ], + "page_idx": 19 + }, + { + "type": "page_number", + "text": "20", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 19 + }, + { + "type": "header", + "text": "H.E. Wong et al.", + "bbox": [ + 271, + 114, + 383, + 128 + ], + "page_idx": 19 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "81. Paranjape, J.N., Nair, N.G., Sikder, S., Vedula, S.S., Patel, V.M.: Adaptivesam: Towards efficient tuning of sam for surgical scene segmentation. arXiv preprint arXiv:2308.03726 (2023)", + "82. Pedraza, L., Vargas, C., Narvaez, F., Durán, O., Muñoz, E., Romero, E.: An open access thyroid ultrasound image database. In: Romero, E., Lepore, N. (eds.) 10th international symposium on medical information processing and analysis. vol. 9287, p. 92870W. SPIE / International Society for Optics and Photonics (2015). https://doi.org/10.1117/12.2073532, https://doi.org/10.1117/12.2073532", + "83. Philbrick, K.A., Weston, A.D., Akkus, Z., Kline, T.L., Korfiatis, P., Sakinis, T., Kostandy, P., Boonrod, A., Zeinoddini, A., Takahashi, N., et al.: Ril-contour: a medical imaging dataset annotation tool for and with deep learning. Journal of digital imaging 32, 571-581 (2019)", + "84. Podobnik, G., Strojan, P., Peterlin, P., Ibragimov, B., Vrtovec, T.: HaN-Seg: The head and neck organ-at-risk CT and MR segmentation dataset. Medical Physics 50(3), 1917-1927 (2023). https://doi.org/https://doi.org/10.1002/mp.16197, https://aapm.onlinelibrary.wiley.com/doi/abs/10.1002/mp.16197, tex.eprint: https://aapm.onlinelibrary.wiley.com/doi/pdf/10.1002/mp.16197", + "85. Porwal, P., Pachade, S., Kamble, R., Kokare, M., Deshmukh, G., Sahasrabuddhe, V., Meriaudeau, F.: Indian diabetic retinopathy image dataset (idrid) (2018). https://doi.org/10.21227/H25W98, https://dx.doi.org/10.21227/H25W98", + "86. Radau, P., Lu, Y., Connelly, K., Paul, G., Dick, A., Wright, G.: Evaluation framework for algorithms segmenting short axis cardiac mri. The MIDAS Journal-Cardiac MR Left Ventricle Segmentation Challenge 49 (2009)", + "87. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)", + "88. Raghu, M., Zhang, C., Kleinberg, J., Bengio, S.: Transfusion: Understanding transfer learning for medical imaging. Advances in neural information processing systems 32 (2019)", + "89. Rajchl, M., Lee, M.C.H., Oktay, O., Kamnitsas, K., Passerat-Palmbach, J., Bai, W., Damodaram, M., Rutherford, M.A., Hajnal, J.V., Kainz, B., Rueckert, D.: DeepCut: Object Segmentation From Bounding Box Annotations Using Convolutional Neural Networks. IEEE Transactions on Medical Imaging 36(2), 674-683 (2017)", + "90. Rakic, M., Wong, H.E., Ortiz, J.J.G., Cimini, B., Guttag, J.V., Dalca, A.V.: Tyche: Stochastic in-context learning for medical image segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2024)", + "91. Rister, B., Yi, D., Shivakumar, K., Nobashi, T., Rubin, D.L.: CT-ORG, a new dataset for multiple organ segmentation in computed tomography. Scientific Data 7(1), 381 (Nov 2020). https://doi.org/10.1038/s41597-020-00715-8, https://www.nature.com/articles/s41597-020-00715-8", + "92. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015)", + "93. Roth, H.R., Yang, D., Xu, Z., Wang, X., Xu, D.: Going to Extremes: Weakly Supervised Medical Image Segmentation. Machine Learning and Knowledge Ex" + ], + "bbox": [ + 225, + 146, + 784, + 839 + ], + "page_idx": 20 + }, + { + "type": "header", + "text": "ScribblePrompt", + "bbox": [ + 624, + 114, + 730, + 128 + ], + "page_idx": 20 + }, + { + "type": "page_number", + "text": "21", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 20 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "traction 3(2), 507-524 (Jun 2021). https://doi.org/10.3390/make3020026, https://www.mdpi.com/2504-4990/3/2/26", + "94. Rother, C., Kolmogorov, V., Blake, A.: \"GrabCut\" — Interactive Foreground Extraction using Iterated Graph Cuts. ACM Transactions on Graphics 23, 309-314 (2004)", + "95. Sakinis, T., Miletari, F., Roth, H., Korfiatis, P., Kostandy, P., Philbrick, K., Akkus, Z., Xu, Z., Xu, D., Erickson, B.J.: Interactive segmentation of medical images through fully convolutional neural networks (2019), arXiv:1903.08205", + "96. Saporta, A., Gui, X., Agrawal, A., Pareek, A., Truong, S., Nguyen, C., Ngo, V.D., Seekins, J., Blankenberg, F.G., Ng, A., et al.: Deep learning saliency maps do not accurately highlight diagnostically relevant regions for medical image interpretation. MedRxiv (2021)", + "97. Sati, P., George, I.C., Shea, C.D., Gaitan, M.I., Reich, D.S.: Flair*: a combined mr contrast technique for visualizing white matter lesions and parenchymal veins. Radiology 265(3), 926-932 (2012)", + "98. Seibold, C., Reif, S., Sarfraz, S., Fink, M.A., Mayer, V., Sellner, J., Kim, M.S., Maier-Hein, K.H., Kleesiek, J., Stiefelhagen, R.: Detailed annotations of chest x-rays via ct projection for report understanding. In: Proceedings of the 33th British Machine Vision Conference (BMVC) (2022)", + "99. Setio, A.A.A., Traverso, A., De Bel, T., Berens, M.S., Van Den Bogaard, C., Cerello, P., Chen, H., Dou, Q., Fantacci, M.E., Geurts, B., et al.: Validation, comparison, and combination of algorithms for automatic detection of pulmonary nodules in computed tomography images: the luna16 challenge. Medical image analysis 42, 1-13 (2017)", + "100. Shi, P., Qiu, J., Abaxi, S.M.D., Wei, H., Lo, F.P.W., Yuan, W.: Generalist vision foundation models for medical imaging: A case study of segment anything model on zero-shot medical segmentation. Diagnostics 13(11), 1947 (2023)", + "101. Simpson, A.L., Antonelli, M., Bakas, S., Bilello, M., Farahani, K., Van Ginneken, B., Kopp-Schneider, A., Landman, B.A., Litjens, G., Menze, B., et al.: A large annotated medical image dataset for the development and evaluation of segmentation algorithms. arXiv preprint arXiv:1902.09063 (2019)", + "102. Sofiuk, K., Petrov, I., Barinova, O., Konushin, A.: F-BRS: Rethinking Backpropagating Refinement for Interactive Segmentation. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8620-8629. IEEE, Seattle, WA, USA (Jun 2020). https://doi.org/10.1109/CVPR42600.2020.00865, https://ieeexplore.ieee.org/document/9156403/", + "103. Sofiuk, K., Petrov, I.A., Konushin, A.: Reviving Iterative Training with Mask Guidance for Interactive Segmentation (Feb 2021), http://arxiv.org/abs/2102.06583, arXiv:2102.06583 [cs]", + "104. Song, Y., Zheng, J., Lei, L., Ni, Z., Zhao, B., Hu, Y.: CT2US: Cross-modal transfer learning for kidney segmentation in ultrasound images with synthesized data. Ultrasonics 122, 106706 (2022). https://doi.org/https://doi.org/10.1016/j.ultras.2022.106706, https://www.sciencedirect.com/science/article/pii/S0041624X22000191", + "105. Staal, J., Abramoff, M.D., Niemeijer, M., Viergever, M.A., Van Ginneken, B.: Ridge-based vessel segmentation in color images of the retina. IEEE transactions on medical imaging 23(4), 501-509 (2004)", + "106. Tang, Y., Yang, D., Li, W., Roth, H.R., Landman, B., Xu, D., Nath, V., Hatamizadeh, A.: Self-Supervised Pre-Training of Swin Transformers for 3D Medical Image Analysis. In: 2022 IEEE/CVF Conference on Computer Vision" + ], + "bbox": [ + 218, + 147, + 785, + 839 + ], + "page_idx": 21 + }, + { + "type": "page_number", + "text": "22", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 21 + }, + { + "type": "header", + "text": "H.E. Wong et al.", + "bbox": [ + 271, + 114, + 383, + 128 + ], + "page_idx": 21 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and Pattern Recognition (CVPR). pp. 20698-20708. IEEE, New Orleans, LA, USA (Jun 2022). https://doi.org/10.1109/CVPR52688.2022.02007, https://ieeexplore.ieee.org/document/9879123/", + "107. Vezhnevets, V., Konouchine, V.: \"GrowCut\" - Interactive Multi-Label N-D Image Segmentation By Cellular Automata. proc. of Graphicon 1(4), 150-156 (2005)", + "108. Wang, C., Chen, X., Ning, H., Li, S.: Sam-octa: A fine-tuning strategy for applying foundation model to octa image segmentation tasks (2023)", + "109. Wang, G., Li, W., Zuluaga, M.A., Pratt, R., Patel, P.A., Aertsen, M., Doel, T., David, A.L., Deprest, J., Ourselin, S., Vercauteren, T.: Interactive Medical Image Segmentation Using Deep Learning With Image-Specific Fine Tuning. IEEE Transactions on Medical Imaging 37(7), 1562-1573 (2018)", + "110. Wang, G., Zuluaga, M.A., Li, W., Pratt, R., Patel, P.A., Aertsen, M., Doel, T., David, A.L., Deprest, J., Ourselin, S., Vercauteren, T.: DeepIGeoS: A Deep Interactive Geodesic Framework for Medical Image Segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence 41(7), 1559-1572 (2019)", + "111. Wang, G., Zuluaga, M.A., Pratt, R., Aertsen, M., Doel, T., Klusmann, M., David, A.L., Deprest, J., Vercauteren, T., Ourselin, S.: Slic-Seg: A minimally interactive segmentation of the placenta from sparse and motion-corrupted fetal MRI in multiple views. Medical Image Analysis 34, 137-147 (Dec 2016). https://doi.org/10.1016/j.media.2016.04.009", + "112. Wang, X., Wang, W., Cao, Y., Shen, C., Huang, T.: Images speak in images: A generalist painter for in-context visual learning. arXiv preprint arXiv:2212.02499 (2022)", + "113. Wang, X., Zhang, X., Cao, Y., Wang, W., Shen, C., Huang, T.: SegGPT: Segmenting Everything In Context (Apr 2023), http://arxiv.org/abs/2304.03284, arXiv:2304.03284 [cs]", + "114. Wasserthal, J., Breit, H.C., Meyer, M.T., Pradella, M., Hinck, D., Sauter, A.W., Heye, T., Boll, D.T., Cyriac, J., Yang, S., et al.: Totalsegmentator: Robust segmentation of 104 anatomic structures in ct images. Radiology: Artificial Intelligence 5(5) (2023)", + "115. Wu, J., Fu, R., Fang, H., Liu, Y., Wang, Z., Xu, Y., Jin, Y., Arbel, T.: Medical SAM Adapter: Adapting Segment Anything Model for Medical Image Segmentation (Apr 2023), http://arxiv.org/abs/2304.12620, arXiv:2304.12620 [cs]", + "116. Xu, N., Price, B., Cohen, S., Yang, J., Huang, T.: Deep Interactive Object Selection. In: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 373-381. IEEE, Las Vegas, NV, USA (Jun 2016). https://doi.org/10.1109/CVPR.2016.47, http://ieeexplore.ieee.org/document/7780416/", + "117. Xu, N., Price, B., Cohen, S., Yang, J., Huang, T.: Deep GrabCut for Object Selection. arXiv (2017), arXiv:1707.00243", + "118. Yushkevich, P.A., Piven, J., Hazlett, H.C., Smith, R.G., Ho, S., Gee, J.C., Gering, G.: User-guided 3d active contour segmentation of anatomical structures: significantly improved efficiency and reliability. Neuroimage 31(3), 1116-1128 (2006)", + "119. Zhang, J., Ding, X., Hu, D., Jiang, Y.: Semantic segmentation of Covid-19 lesions with a multiscale dilated convolutional network. Scientific Reports 12(1), 1847 (2022)", + "120. Zhang, K., Liu, D.: Customized segment anything model for medical image segmentation (2023)", + "121. Zhang, K., Zhuang, X.: Cycle mix: A holistic strategy for medical image segmentation from scribble supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11656-11665 (2022)" + ], + "bbox": [ + 217, + 146, + 785, + 840 + ], + "page_idx": 22 + }, + { + "type": "header", + "text": "ScribblePrompt", + "bbox": [ + 624, + 114, + 730, + 128 + ], + "page_idx": 22 + }, + { + "type": "page_number", + "text": "23", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 22 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "122. Zhang, S., Liew, J.H., Wei, Y., Wei, S., Zhao, Y.: Interactive Object Segmentation With Inside-Outside Guidance. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 12231-12241. IEEE (2020)", + "123. Zhang, T.Y., Suen, C.Y.: A fast parallel algorithm for thinning digital patterns. Commun. ACM 27(3), 236-239 (mar 1984). https://doi.org/10.1145/357994.358023, https://doi.org/10.1145/357994.358023", + "124. Zhang, Y., Xian, M., Cheng, H.D., Shareef, B., Ding, J., Xu, F., Huang, K., Zhang, B., Ning, C., Wang, Y.: Busis: A benchmark for breast ultrasound image segmentation. In: Healthcare. vol. 10, p. 729. MDPI (2022)", + "125. Zhao, Q., Lyu, S., Bai, W., Cai, L., Liu, B., Wu, M., Sang, X., Yang, M., Chen, L.: A multi-modality ovarian tumor ultrasound image dataset for unsupervised cross-domain semantic segmentation. CoRR abs/2207.06799 (2022)", + "126. Zheng, G., Chu, C., Belavy, D.L., Ibragimov, B., Korez, R., Vrtovec, T., Hutt, H., Everson, R., Meakin, J., Andrade, I.L., et al.: Evaluation and comparison of 3d intervertebral disc localization and segmentation methods for 3d t2 mr data: A grand challenge. Medical image analysis 35, 327-344 (2017)", + "127. Zheng, X., Wang, Y., Wang, G., Liu, J.: Fast and robust segmentation of white blood cell images by self-supervised learning. Micron 107, 55-71 (2018). https://doi.org/https://doi.org/10.1016/j.micron.2018.01.010, https://www.sciencedirect.com/science/article/pii/S0968432817303037", + "128. Zhou, T., Li, L., Bredell, G., Li, J., Unkelbach, J., Konukoglu, E.: Volumetric memory network for interactive medical image segmentation. Medical Image Analysis 83, 102599 (Jan 2023). https://doi.org/10.1016/j.media.2022.102599, https://linkinghub.elsevier.com/retrieve/pii/S1361841522002316", + "129. Zou, X., Yang, J., Zhang, H., Li, F., Li, L., Wang, J., Wang, L., Gao, J., Lee, Y.J.: Segment everything everywhere all at once. Advances in Neural Information Processing Systems 36 (2024)" + ], + "bbox": [ + 215, + 146, + 785, + 521 + ], + "page_idx": 23 + }, + { + "type": "page_number", + "text": "24", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 23 + }, + { + "type": "header", + "text": "H.E. Wong et al.", + "bbox": [ + 271, + 114, + 383, + 128 + ], + "page_idx": 23 + } +] \ No newline at end of file diff --git a/2024/ScribblePrompt_ Fast and Flexible Interactive Segmentation for Any Biomedical Image/97fd2117-c4b3-47cc-8dc3-f7ab8ca17648_model.json b/2024/ScribblePrompt_ Fast and Flexible Interactive Segmentation for Any Biomedical Image/97fd2117-c4b3-47cc-8dc3-f7ab8ca17648_model.json new file mode 100644 index 0000000000000000000000000000000000000000..cf02f15250e9fa11ba1fcaeb07455b61eb4ea7a3 --- /dev/null +++ b/2024/ScribblePrompt_ Fast and Flexible Interactive Segmentation for Any Biomedical Image/97fd2117-c4b3-47cc-8dc3-f7ab8ca17648_model.json @@ -0,0 +1,4065 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.241, + 0.141, + 0.764, + 0.187 + ], + "angle": 0, + "content": "ScribblePrompt: Fast and Flexible Interactive Segmentation for Any Biomedical Image" + }, + { + "type": "text", + "bbox": [ + 0.275, + 0.212, + 0.728, + 0.243 + ], + "angle": 0, + "content": "Hallee E. Wong\\(^{1,2}\\), Marianne Rakic\\(^{1,2}\\), John Guttag\\(^{1}\\), and Adrian V. Dalca\\(^{1,2,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.254, + 0.756, + 0.311 + ], + "angle": 0, + "content": "1 MIT CSAIL, Cambridge, MA, USA \n2 Martins Center, Massachusetts General Hospital, Charlestown, MA, USA \n3 Harvard Medical School, Boston, MA, USA \n{hallel, mrukic, guttag, adalca}@mit.edu" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.337, + 0.74, + 0.614 + ], + "angle": 0, + "content": "Abstract. Biomedical image segmentation is a crucial part of both scientific research and clinical care. With enough labelled data, deep learning models can be trained to accurately automate specific biomedical image segmentation tasks. However, manually segmenting images to create training data is highly labor intensive and requires domain expertise. We present ScribblePrompt, a flexible neural network based interactive segmentation tool for biomedical imaging that enables human annotators to segment previously unseen structures using scribbles, clicks, and bounding boxes. Through rigorous quantitative experiments, we demonstrate that given comparable amounts of interaction, ScribblePrompt produces more accurate segmentations than previous methods on datasets unseen during training. In a user study with domain experts, ScribblePrompt reduced annotation time by \\(28\\%\\) while improving Dice by \\(15\\%\\) compared to the next best method. ScribblePrompt's success rests on a set of careful design decisions. These include a training strategy that incorporates both a highly diverse set of images and tasks, novel algorithms for simulated user interactions and labels, and a network that enables fast inference. We showcase ScribblePrompt in an interactive demo, provide code, and release a dataset of scribble annotations at https://scribbleprompt.csail.mit.edu" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.627, + 0.737, + 0.641 + ], + "angle": 0, + "content": "Keywords: Interactive Segmentation \\(\\cdot\\) Biomedical Imaging \\(\\cdot\\) Scribbles" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.662, + 0.377, + 0.677 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.787, + 0.795 + ], + "angle": 0, + "content": "Biomedical image segmentation is an essential step in a wide range of biomedical research and clinical care pipelines. Deep learning has become the predominant method to automate existing segmentation tasks [43, 78, 92, 106]. Biomedical researchers and clinicians often encounter novel segmentation tasks involving either new regions of interest [119] or new image modalities [77,97]. Unfortunately, supervised training of accurate models for new domains requires diverse images with careful annotations by skilled experts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Most widely used interactive segmentation systems for biomedical imaging provide minimal, intensity-based, algorithmic assistance [118]. Despite a growing literature, learning-based systems are less widely used in practice, perhaps" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.386, + 0.129 + ], + "angle": 0, + "content": "H.E. Wong et al." + }, + { + "type": "image_caption", + "bbox": [ + 0.242, + 0.145, + 0.274, + 0.153 + ], + "angle": 0, + "content": "Step 1" + }, + { + "type": "image_caption", + "bbox": [ + 0.308, + 0.145, + 0.342, + 0.153 + ], + "angle": 0, + "content": "Step 2" + }, + { + "type": "image_caption", + "bbox": [ + 0.357, + 0.145, + 0.411, + 0.153 + ], + "angle": 0, + "content": "Step 3" + }, + { + "type": "image_caption", + "bbox": [ + 0.43, + 0.145, + 0.499, + 0.153 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.145, + 0.553, + 0.153 + ], + "angle": 0, + "content": "Step 1" + }, + { + "type": "image_caption", + "bbox": [ + 0.588, + 0.145, + 0.622, + 0.153 + ], + "angle": 0, + "content": "Step 2" + }, + { + "type": "image_caption", + "bbox": [ + 0.659, + 0.145, + 0.692, + 0.153 + ], + "angle": 0, + "content": "Step 3" + }, + { + "type": "image_caption", + "bbox": [ + 0.713, + 0.145, + 0.783, + 0.153 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.155, + 0.288, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.155, + 0.359, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.155, + 0.428, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.43, + 0.155, + 0.5, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.155, + 0.57, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.155, + 0.639, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.155, + 0.71, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.155, + 0.783, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.211, + 0.294, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.295, + 0.211, + 0.362, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.363, + 0.211, + 0.429, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.43, + 0.211, + 0.5, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.211, + 0.57, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.211, + 0.639, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.211, + 0.71, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.211, + 0.783, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.277, + 0.788, + 0.374 + ], + "angle": 0, + "content": "Fig. 1: ScribblePrompt enables rapid iterative interactive segmentation of unseen tasks using bounding boxes, clicks, and scribbles. We show predictions from ScribblePrompt with iterative interaction steps on examples from datasets unseen during training. At each step, we visualize positive scribble and click inputs in green, negative scribble and click inputs in red, bounding box inputs in yellow, and the predicted segmentation in blue. Scribble thickness is enlarged for visual clarity. See Supplementary Material for more examples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.403, + 0.785, + 0.508 + ], + "angle": 0, + "content": "because they are generally focused on specific tasks or modalities, which limits their applicability [68, 83, 95, 109, 110, 128]. Recent vision foundation models target broad applicability, but require fine-tuning to the medical domain [47, 112, 113, 129]. Most interactive segmentation models require specific interactions, like carefully placed clicks, making them easy to develop but difficult to use in practice. The result of these shortcomings is that existing learning-based models are not widely used in practice." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.509, + 0.785, + 0.539 + ], + "angle": 0, + "content": "In this work, we present ScribblePrompt, a general model for interactive biomedical image segmentation that" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.547, + 0.785, + 0.59 + ], + "angle": 0, + "content": "1. enables users to rapidly and accurately accomplish any biomedical image segmentation task, outperforming state-of-the-art models, particularly for unseen labels and image types;" + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.591, + 0.785, + 0.62 + ], + "angle": 0, + "content": "2. is flexible to different annotation styles, including bounding boxes, clicks, and scribbles," + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.621, + 0.776, + 0.636 + ], + "angle": 0, + "content": "3. is computationally efficient, enabling fast inference, even on a single CPU." + }, + { + "type": "list", + "bbox": [ + 0.222, + 0.547, + 0.785, + 0.636 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.785, + 0.734 + ], + "angle": 0, + "content": "To achieve this, we focused on design decisions aimed at ease of use and realistic interactions. To train our model, we started by gathering a large corpus of biomedical imaging datasets. We then designed task augmentation through synthesis strategies to encourage generalization to unseen tasks. Finally, we designed new interaction simulation strategies for training. The model itself uses an architecture optimized for efficient inference." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.786, + 0.84 + ], + "angle": 0, + "content": "To evaluate ScribblePrompt, we compared its performance to that of existing interactive segmentation methods on datasets unseen during training. These experiments involved manually-collected scribbles, simulated interactions (Fig. 1), and a user study in which experienced annotators were asked to segment images. In the user study, ScribblePrompt reduced annotation time by \\(28\\%\\) while increasing Dice by \\(15\\%\\) compared to the next best method, the Segment Anything Model (SAM) [47]. Given similar amounts of interaction, ScribblePrompt" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.625, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "ScribblePrompt" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "achieved consistently better Dice scores than other methods. We release an interactive tool, code, model weights, and our dataset of manually-collected scribbles at https://scribbleprompt.csail.mit.edu." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.217, + 0.396, + 0.234 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.25, + 0.788, + 0.416 + ], + "angle": 0, + "content": "Interactive Biomedical Image Segmentation. Early research into interactive segmentation of biomedical images focused on traditional intensity-based methods [13,21,33,107,118]. Recent deep-learning based techniques [68,83,109, 110] use human interaction to improve the accuracy of segmentation tasks where labelled data exists, but fully-automatic segmentation methods fail to produce accurate segmentations. This approach assumes that the model will be used to perform the same segmentation task(s) it was trained on, leading to domain-specific solutions [6,62,80,89,110,111]. A few methods generalize to new, but similar, classes [95] and modalities [68,109] to those seen in training. In contrast, ScribblePrompt is designed to help annotators segment a wide range of biomedical images and (potentially unseen) tasks at inference time." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.417, + 0.788, + 0.492 + ], + "angle": 0, + "content": "Foundation Models. Recent vision foundation models employ prompting to enable generalization to new tasks. These models, trained on large collections of (natural) image data, segment potentially-unseen structures specified by spatial prompts [47, 129], text [47, 87, 129], or examples [15, 22, 112, 113]. Some of these models can be used for interactive segmentation with spatial prompts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.788, + 0.672 + ], + "angle": 0, + "content": "Fine-tuning models initially developed using natural images is often unhelpful for specific tasks in the biomedical imaging domain, compared to training from scratch [88]. The Segment Anything Model (SAM) [47], a foundation model trained on natural images, performs well at biomedical image segmentation tasks with clear boundaries (e.g. organs in abdominal CT), but performs poorly on tasks and modalities involving more subtle delineations (e.g. deep structures in Brain MRI) [36, 41, 76, 100]. Several recent or concurrent papers fine-tune SAM for specific biomedical tasks or modalities [40, 46, 59, 81, 108, 120]. A few fine-tune SAM to segment medical images from multiple-modalities with bounding boxes [69], clicks [115] or both [19]. We show in our experiments that these models don't perform well on many unseen tasks and are limiting in their required interactions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.809 + ], + "angle": 0, + "content": "Several natural image methods simulate iterative user interactions and condition the model on previous predictions to train a single network to make the initial prediction and perform refinement [14, 27, 47, 61, 102, 103]. This approach is appropriate for natural images where there is ample data and typically one modality. However, such methods have not previously been developed for the biomedical imaging domain, which is more fragmented because of specialized regions of interest and different modalities. We use iterative ideas in ScribblePrompt, but we facilitate iterative interactive segmentation of unseen biomedical imaging tasks." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "User Interaction. Users can prompt interactive segmentation systems in many different ways, such as bounding boxes [9, 47, 89, 94, 109, 117], scribbles" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.387, + 0.129 + ], + "angle": 0, + "content": "H.E. Wong et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.268 + ], + "angle": 0, + "content": "[18, 110, 111], or clicks [9, 61, 68, 103, 110]. Few interactive segmentation models incorporate multiple types of inputs. Some works have explored training with more specialized types of clicks [122], such as extreme points [72, 93], interior margin points [68], and center clicks [103, 116], to maximize the information per interaction. Specialized interactions often require more user time, and models trained only with such interactions are less robust to deviations from the interaction protocol [9, 103, 117]. In contrast, we focus on a user-friendly model that performs well under a variety of interaction scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.269, + 0.789, + 0.405 + ], + "angle": 0, + "content": "Scribbles are an intuitive form of interaction, however few interactive segmentation works employ them, because acquiring realistic scribbles for training and evaluation is challenging. Previous deep-learning models trained for interactive scribble-based segmentation either relied on collecting large datasets of manually drawn scribbles [53, 57, 128] or simplistic approaches such as sampling random points [5, 7, 110] or Bezier curves [2]. One work has explored simulating more complex scribbles for interactive segmentation of natural images [18]. We build on these concepts to create a new scribble simulation engine for training and evaluation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.407, + 0.789, + 0.513 + ], + "angle": 0, + "content": "Scribble-supervised learning methods use scribble annotations as supervision to train automatic segmentation models for predicting segmentation given only an input image [32,56,57,66,121]. However, these methods require the manual scribble-annotation of many training images and retraining for each new task. In contrast, ScribblePrompt is trained on a large corpus of datasets once and then can be used to perform new segmentation tasks at inference time without retraining, using scribbles as input." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.543, + 0.506, + 0.56 + ], + "angle": 0, + "content": "3 ScribblePrompt Approach" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.582, + 0.788, + 0.688 + ], + "angle": 0, + "content": "We present an interactive segmentation method that can generalize to new biomedical imaging modalities and regions of interest, while staying focused on practical usability. We describe the problem formulation, and present the important aspects of the ScribblePrompt framework: (i) simulation of realistic interactions during training, (ii) augmentation with synthetic labels during training to encourage generalization, and (iii) an efficient architecture for fast inference. We show an overview of training in Fig. 2 and 4." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.717, + 0.44, + 0.731 + ], + "angle": 0, + "content": "3.1 Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Let \\( t \\) be a segmentation task consisting of image and segmentation pairs, \\( \\{(x^t,y^t)_j\\}_{j = 1}^N \\). At step \\( i \\), given an image \\( x^{t} \\), a set of user interactions \\( u_{i} \\), and previous prediction \\( \\hat{y}_{i - 1}^{t} \\), we learn function \\( f_{\\theta}(x^{t},u_{i},\\hat{y}_{i - 1}^{t}) \\) with parameters \\( \\theta \\) that produces a segmentation \\( \\hat{y}_i \\). The set of interactions \\( u_{i} \\), which may include positive or negative scribbles, positive or negative clicks, and bounding boxes, is provided by a user who has access to the image \\( x^{t} \\) and previous prediction \\( \\hat{y}_{i - 1}^{t} \\)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.625, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "ScribblePrompt" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "We minimize the difference between the true segmentation \\( y^{t} \\) and each of the \\( k \\) iterative predictions \\( \\hat{y}_1, \\dots, \\hat{y}_k \\)," + }, + { + "type": "equation", + "bbox": [ + 0.285, + 0.185, + 0.786, + 0.228 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (\\theta ; \\mathcal {T}) = \\mathbb {E} _ {t \\in \\mathcal {T}} \\left[ \\mathbb {E} _ {(x ^ {t}, y ^ {t}) \\in t} \\left[ \\sum_ {i = 1} ^ {k} \\mathcal {L} _ {\\text {S e g}} \\left(y ^ {t}, f _ {\\theta} \\left(x ^ {t}, u _ {i}, \\hat {y} _ {i - 1} ^ {t}\\right)\\right) \\right] \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.234, + 0.545, + 0.249 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{Seg}\\) is a supervised segmentation loss." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.25, + 0.785, + 0.356 + ], + "angle": 0, + "content": "During training, we sample a task \\(t\\) from training task collection \\(\\mathcal{T}\\), from which we sample an image \\(x^t\\), and segmentation map \\(y^{t}\\). We simulate a possible set of interactions \\(u_{i}\\) based on \\(y^{t}\\), and predict \\(\\hat{y}_i^t = f_\\theta (x^t,u_i,\\hat{y}_{i - 1}^t)\\). We simulate the next set of interactions \\(u_{i + 1}\\) based on the error region \\(\\varepsilon_i^t = y^t -\\hat{y}_i^t\\), and repeat for \\(k\\) iterations. In the following sections, we describe the core aspects of the framework: strategies for simulating \\(u_{i}\\), collecting \\(\\mathcal{T}\\), sampling \\((x^{t},y^{t})\\), building \\(f_{\\theta}\\), and optimizing \\(\\mathcal{L}(\\theta ;\\mathcal{T})\\)." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.379, + 0.789, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.568, + 0.788, + 0.695 + ], + "angle": 0, + "content": "Fig. 2: Training. We simulate \\( k \\) consecutive steps of interactive segmentation. Given an image segmentation pair \\( (x^t, y^t) \\), we first simulate a set of initial interactions \\( u_1 \\), which may contain bounding boxes, clicks, and/or scribbles. We predict segmentation \\( \\hat{y}_1^t \\coloneqq f_\\theta(x^t, u_1, \\hat{y}_0^t) \\) where the initial prediction \\( \\hat{y}_0^t \\) is set to zeros. In the second step, we simulate corrections using the error region \\( \\varepsilon_1^t \\) between the previous prediction \\( \\hat{y}_1^t \\) and ground truth \\( y^t \\), and add them to the set of initial interactions \\( u_1 \\) to get \\( u_2 \\). We predict segmentation \\( \\hat{y}_2^t \\coloneqq f_\\theta(x^t, u_2, \\hat{y}_1^t) \\) and repeat to produce a series of predictions, \\( \\hat{y}_1^t, \\dots, \\hat{y}_k^t \\). We learn \\( \\theta \\) to minimize \\( \\sum_{i=1}^{k} \\mathcal{L}_{seg}(y^t, \\hat{y}_i^t) \\), the sum of losses between the target segmentation \\( y^t \\) and iterative predictions \\( \\hat{y}_1^t, \\dots, \\hat{y}_k^t \\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.742, + 0.421, + 0.757 + ], + "angle": 0, + "content": "3.2 Prompt Simulation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.784, + 0.84 + ], + "angle": 0, + "content": "To enable a practical, easy-to-use model, we encourage robustness to different types of user interactions \\( u_{i} \\). We introduce algorithms for simulating scribbles, clicks, and bounding box inputs during training. Each scribble and click strategy (Fig. 3) can be applied to a ground truth segmentation label \\( y^{t} \\) to simulate positive interactions, or to the background \\( 1 - y^{t} \\) to simulate negative interactions." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.386, + 0.129 + ], + "angle": 0, + "content": "H.E. Wong et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.145, + 0.785, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.242, + 0.788, + 0.286 + ], + "angle": 0, + "content": "Fig. 3: Simulated scribbles and clicks. Positive interactions (green) are simulated on the segmentation label \\( y^{t} \\) (blue), while negative interactions (red) are simulated on the background \\( 1 - y^{t} \\). Scribble thickness is enlarged for visual clarity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.316, + 0.784, + 0.361 + ], + "angle": 0, + "content": "We simulate positive and negative correction scribbles or clicks by applying the same strategies to the false negative region \\(\\varepsilon_{i}^{t} > 0\\) and false positive region \\(\\varepsilon_{i}^{t} < 0\\) for error region \\(\\varepsilon_{i}^{t}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.362, + 0.785, + 0.437 + ], + "angle": 0, + "content": "Scribbles. Given binary mask \\( y \\in \\{0,1\\}^{h \\times w} \\), we simulate a scribble mask \\( s \\in [0,1]^{h \\times w} \\) by first generating clean scribbles, and then corrupting them to account for user behavior and variability. We illustrate this process in Supplementary Material Sec. B.1. To generate the clean scribbles, we start with one of these strategies:" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.45, + 0.784, + 0.48 + ], + "angle": 0, + "content": "(i) Line Scribbles: We draw random lines by connecting two end points sampled from \\(\\{(u,v)|y_{uv} = 1\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.213, + 0.481, + 0.785, + 0.51 + ], + "angle": 0, + "content": "(ii) Centerline Scribbles: We simulate scribbles in the center of label \\( y \\) using a thinning algorithm [123] that reduces the label to a 1-pixel wide skeleton." + }, + { + "type": "text", + "bbox": [ + 0.211, + 0.512, + 0.785, + 0.588 + ], + "angle": 0, + "content": "(iii) Contour Scribbles: We simulate a rough contour of the desired segmentation within the boundaries of the mask. We first blur the mask to reduce the size of the label such that \\(\\tilde{y} = \\min(y, y \\circ G_k)\\), where \\(G_k\\) is a Gaussian blur kernel. Then we apply a threshold \\(\\tilde{y} < h\\) sampled in some intensity range \\(h \\sim U[\\tilde{y}_{\\text{min}}, \\tilde{y}_{\\text{max}}]\\) and extract a contour inside the boundary of the mask." + }, + { + "type": "list", + "bbox": [ + 0.211, + 0.45, + 0.785, + 0.588 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.785, + 0.705 + ], + "angle": 0, + "content": "To limit the size and complexity of the centerline and contour scribbles, we apply a random mask that breaks the scribbles into smaller parts. We generate the random mask by sampling a smooth noise image \\( p \\), where each pixel is sampled independently from \\( \\mathcal{N}(\\mu_p, \\sigma_p^2) \\), applying Gaussian blur, and then thresholding it at \\( \\mu_p \\). We warp the resulting scribble mask \\( s \\) using a random deformation field \\( \\phi \\) to vary the scribble shape and thickness. We ensure the resulting scribble is consistent with mask \\( y \\), by multiplying the warped scribble mask \\( s \\circ \\phi \\) by \\( y \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.706, + 0.784, + 0.737 + ], + "angle": 0, + "content": "Clicks. Given a binary mask \\( y \\in \\{0,1\\}^{h \\times w} \\), we simulate \\( n \\sim U[n_{min}, n_{max}] \\) clicks at a time using one of three strategies, illustrated in Fig. 3:" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.748, + 0.785, + 0.779 + ], + "angle": 0, + "content": "(i) Random clicks: We randomly sample clicks from all pixels in the given region \\(\\{(u,v)|y_{uv} = 1\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.785, + 0.84 + ], + "angle": 0, + "content": "(ii) Center clicks: We sample clicks from the set of points at the center of each disconnected component of the label. First, we create a multi-label mask \\( m \\in \\{1, \\dots, C\\}^{h \\times w} \\) identifying the \\( C \\) components of label \\( y \\). We identify the center of each component using the euclidean distance transform [103, 116]." + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.748, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.625, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "ScribblePrompt" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.21, + 0.146, + 0.788, + 0.224 + ], + "angle": 0, + "content": "(iii) Interior border region clicks: We sample clicks from a border region inside the boundary of the mask. We first blur the mask \\(y\\) to reduce the size of the label such that \\(\\tilde{y} = \\min(y, y \\circ G_k)\\) where \\(G_k\\) is a Gaussian blur kernel. We then sample click coordinates from \\(\\{(u, v) | \\tilde{y}_{uv} \\in [a, b]\\}\\), where \\(a, b \\sim U[\\tilde{y}_{\\text{min}}, \\tilde{y}_{\\text{max}})\\) are thresholds sampled in some intensity range." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.228, + 0.786, + 0.258 + ], + "angle": 0, + "content": "Since users are inclined to spread out their clicks, we impose a minimum separation of a few pixels between random clicks and border region clicks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.258, + 0.786, + 0.302 + ], + "angle": 0, + "content": "Bounding Boxes. We compute the minimum bounding box that encloses the label \\( y^{t} \\), and enlarge each dimension by \\( r \\sim U[0,20] \\) pixels to account for human variability." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.303, + 0.787, + 0.41 + ], + "angle": 0, + "content": "Iterative Training. During the first step \\((i = 1)\\), we sample the combination of interactions and the number of initial positive and negative interactions \\(n_{pos}, n_{neg} \\sim U[n_{min}, n_{max}]\\). The initial interactions \\(u_{1}\\) are simulated using the ground truth label \\(y^{t}\\). In subsequent steps, correction scribbles or clicks are sampled from the error region \\(\\varepsilon_{i-1}^{t}\\) between the last prediction \\(\\hat{y}_{i-1}^{t}\\) and the ground truth \\(y^{t}\\). Since a user can make multiple corrections in each step, we sample \\(n_{cor} \\sim U[n_{min}, n_{max}]\\) corrections (scribbles or clicks) per step." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.427, + 0.304, + 0.441 + ], + "angle": 0, + "content": "3.3 Data" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.447, + 0.787, + 0.569 + ], + "angle": 0, + "content": "We build on large dataset gathering efforts like MegaMedical [15,90] to compile a collection of 77 open-access biomedical imaging datasets for training and evaluation, covering over 54k scans, 16 image types, and 711 labels. The collection includes a diverse array of biomedical domains, such as eyes [39,55,70,85,105], thorax [86,96,98,99,101], spine [64,98,114,126], cells [16,17,28,29,63,127], skin [20], abdominal [11,12,37,44,45,51,52,54,60,67,71,86,91,101,104], neck [48,79,82,84], brain [3,8,31,38,49,50,73-75,101], bones [34,98,114], teeth [1,65] and lesions [4,101,124,125]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.787, + 0.644 + ], + "angle": 0, + "content": "We define a 2D segmentation task as a combination of dataset, axis (for 3D modalities), and label. For datasets with multiple segmentation labels, we consider each label separately as a binary segmentation task and for 3D modalities we use the slice with maximum label area and the middle slice from each volume. We provide more details on the data in Supplementary Material Sec. C." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.787, + 0.719 + ], + "angle": 0, + "content": "Task Diversity. During training we sample image, segmentation pairs hierarchically - by dataset and modality, axis, and then label - to balance training on datasets of different sizes. To increase the diversity of segmentation tasks, we apply data augmentation [15] to both the input image and sampled segmentation prior to simulating the user interactions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Synthetic Labels. To limit overfitting to specific segmentation tasks, we introduce a mechanism to generate synthetic labels (Fig. 4). During training, for a given sample \\((x_0,y_0)\\), with probability \\(p_{synth}\\), we replace \\(y\\) with a synthetic label \\(y_{synth}\\). Given an image \\(x_0\\), we generate a synthetic label \\(y_{synth}\\) by applying a superpixel algorithm [25] with scale parameter \\(\\lambda \\sim U[1,\\lambda_{max}]\\) to partition the image into a multi-label mask of \\(k\\) superpixels \\(z \\in \\{1,\\dots,k\\}^{n\\times n}\\), and then randomly select a superpixel \\(y_{synth} = \\mathbb{1}(z = c)\\). We conduct experiments varying \\(p_{synth}\\) and show examples in Supplementary Material Sec. B.3." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.386, + 0.129 + ], + "angle": 0, + "content": "H.E. Wong et al." + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.143, + 0.788, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.247, + 0.79, + 0.318 + ], + "angle": 0, + "content": "Fig. 4: Task sampling and augmentation. During training, we sample an image and segmentation pair \\((x_0, y_0)\\). With probability \\(p_{\\text{synth}}\\), we replace \\(y_0\\) with a synthetic label \\(y_{\\text{synth}}\\). We generate \\(y_{\\text{synth}}\\) by applying a superpixel algorithm to the image \\(x_0\\) to generate a map \\(z\\) of potential synthetic labels (superpixels), and then sampling one label. Finally, we apply random data augmentations to get \\((x^t, y^t)\\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.348, + 0.335, + 0.362 + ], + "angle": 0, + "content": "3.4 Network" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.373, + 0.799, + 0.42 + ], + "angle": 0, + "content": "Motivated by producing a practical tool, we primarily demonstrate ScribblePrompt using an efficient fully-convolutional architecture similar to a UNet [92]. We also demonstrate ScribblePrompt using a vision transformer architecture [47]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.42, + 0.792, + 0.465 + ], + "angle": 0, + "content": "ScribblePrompt-UNet: We use an 8 layer CNN following a decoder-encoder structure similar to the popular UNet architecture [92] without Batch Norm. Each convolutional layer has 192 features and uses PReLU activation [35]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.465, + 0.79, + 0.525 + ], + "angle": 0, + "content": "ScribblePrompt-SAM: We take the smallest SAM model (ViT-b) [47] and fine-tune the decoder. SAM takes bounding box and click inputs as lists of \\((x,y)\\) coordinates, and the low-resolution logits of the previous prediction \\(\\hat{y}_{i-1}\\). To adapt SAM for scribbles, we consider each scribbled pixel as a click." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.525, + 0.788, + 0.587 + ], + "angle": 0, + "content": "Loss. We minimize eq. (1) where \\(\\mathcal{L}_{Seg}\\) is a linear combination of soft Dice Loss [24] and Focal Loss [58]. In preliminary experiments (Supplementary Material Sec. B.2), training with Soft Dice Loss alone or a linear combination of Soft Dice Loss and Binary-Cross Entropy Loss resulted in slightly lower Dice scores." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.611, + 0.446, + 0.629 + ], + "angle": 0, + "content": "4 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.788, + 0.704 + ], + "angle": 0, + "content": "We compare ScribblePrompt-UNet and ScribblePrompt-SAM to previous methods through experiments with manual scribbles, simulated interactions, and a user study with experienced annotators. Lastly, we report on inference runtime and ablation experiments." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.795 + ], + "angle": 0, + "content": "Data. We use 65 (out of 77) datasets during training. We partition another nine datasets, ACDC [10], BUID [4], BTCV Cervix [52], DRIVE [105], HipXRay [34], PanDental [1], SCD [86], SpineWeb [126], and WBC [127], each into validation (used in model selection, but not training) and test (used only for final evaluation). Three additional datasets, TotalSegmentator [114], SCR [30], and COBRE [3,23,26], were used only for final evaluation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.84 + ], + "angle": 0, + "content": "We report final results on 12 sets of data (the test splits of the \\(9 + 3\\) evaluation datasets). The evaluation datasets were not seen during training, and cover 608 tasks and 8 modalities, including unseen image types and unseen labels. We" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.625, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "ScribblePrompt" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "selected these 12 evaluation datasets to cover a variety of modalities (MRI, CT, ultrasound, fundus photography, microscopy) and anatomical regions of interest (brain, teeth, bones, abdominal organs, muscles, heart, thorax, cells), including both healthy anatomy and lesions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.785, + 0.342 + ], + "angle": 0, + "content": "Training. During training, we simulated five steps of interactive segmentation per example and set the maximum number of interactions per step to three. We set the minimum number of initial negative prompts to zero, and the minimum initial positive and correction prompts to one. During the first step, we sample the combination of prompt types ( clicks, scribbles, bounding boxes) and then sample the number of positive and negative interactions for each type. In each subsequent steps, we simulate either scribble corrections or click corrections with equal probability. We selected these values based on what we believe to be reasonable interactions for a user to perform." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.343, + 0.784, + 0.372 + ], + "angle": 0, + "content": "Baselines. We compare to existing generalist methods for interactive segmentation, with a focus on methods developed for biomedical imaging." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.373, + 0.784, + 0.418 + ], + "angle": 0, + "content": "SAM [47]: We evaluate the smallest (ViT-b) and largest (ViT-h) versions of the Segment Anything Model (SAM) trained on natural images. SAM takes bounding boxes, clicks, and the logits of the previous prediction as input." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.419, + 0.784, + 0.493 + ], + "angle": 0, + "content": "SAM-Med2D [19]: SAM-Med2D is a SAM ViT-b model with additional adapter layers in the image encoder. SAM-Med2D was fine-tuned, using bounding boxes and iterative clicks as input, on a collection of biomedical imaging datasets containing 4.6M images and 19.7M segmentations. Following [19], we evaluate SAM-Med2D both with and without adapter layers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.494, + 0.784, + 0.554 + ], + "angle": 0, + "content": "MedSAM [69]: MedSAM is a SAM ViT-B model fine-tuned with bounding box prompts on a collection of biomedical imaging datasets containing 1.5M image segmentation pairs. We evaluate MedSAM with bounding boxes only, because we found it to perform poorly when given point or mask prompts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.555, + 0.784, + 0.645 + ], + "angle": 0, + "content": "MIDepSeg [68]: MIDepSeg is an interactive segmentation framework designed to generalize to unseen tasks on medical images. MIDepSeg takes interior margin points (positive clicks) as initial inputs, crops the image based on those points, and uses a CNN to make an initial prediction. Given additional clicks, the prediction can be refined using conditional random fields. We evaluate the pre-trained model, which was developed on placenta T2 MRI." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.665, + 0.357, + 0.682 + ], + "angle": 0, + "content": "5 Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.694, + 0.784, + 0.769 + ], + "angle": 0, + "content": "We evaluate all methods using both manual and simulated interactions, with a focus on scribbles. Simulated interactions enable us to test on a large volume of images and tasks. However, simulations do not always match user behavior. We address this by (1) collecting a diverse dataset of manual scribbles for evaluation, and (2) conducting an interactive user study." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.788, + 0.407, + 0.803 + ], + "angle": 0, + "content": "5.1 Manual Scribbles" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Setup. We evaluate on two datasets of manual scribbles. First, we collected a dataset (MedScribble) of manual scribbles from three annotators for seven" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.385, + 0.129 + ], + "angle": 0, + "content": "H.E. Wong et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "segmentation tasks from unseen datasets, with a total of 31 examples. The annotators were shown five training examples with the ground truth segmentation per task and instructed to draw positive and negative scribbles to indicate the region of interest on new images. We provide more details on MedScribble and release the data in Supplementary Material Sec. E.1. Second, we used 380 slices from the ACDC dataset [10], which has scribble annotations for three labels and background." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.788, + 0.389 + ], + "angle": 0, + "content": "In this evaluation, we make a prediction given a set of positive and negative scribbles, simulating a non-iterative scenario where the annotator draws several scribbles before running inference. We evaluate accuracy using Dice score [24] and 95th percentile Hausdorff Distance [42]. Since MIDepSeg's initial prediction network only takes positive inputs, we report results on both MIDepseg's initial prediction from the positive scribbles, and after applying MIDepSeg's refinement procedure using the negative scribbles as corrections. For SAM and variants fine-tuning SAM on clicks, we convert each scribble to a series of points. For MedSAM, we fit a bounding box to the positive scribbles." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.39, + 0.788, + 0.464 + ], + "angle": 0, + "content": "Results. ScribblePrompt-UNet and ScribblePrompt-SAM produce the most accurate segmentations in a single step of manual scribbles on both our manual scribbles dataset and the ACDC scribbles dataset (Tab. 1). We show example predictions in Fig. 5 and additional comparisons to scribble-supervised learning in Supplementary Material Sec. E.3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.466, + 0.788, + 0.556 + ], + "angle": 0, + "content": "SAM and SAM-Med2D do not generalize well to scribble inputs (which they were not trained for). MedSAM has better predictions than other SAM baselines using the SAM architecture, however it is not able to make use of the negative scribbles and thus often misses segmentations with holes in them (Fig. 5). The initial predictions from MIDepSeg's network are poor, but improve after applying the refinement procedure." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.584, + 0.788, + 0.613 + ], + "angle": 0, + "content": "Table 1: Manual scribbles. Mean Dice and HD95 with \\(95\\%\\) CI of segmentations predicted from manually-collected scribbles. Best and second best are highlighted." + }, + { + "type": "table", + "bbox": [ + 0.226, + 0.625, + 0.784, + 0.809 + ], + "angle": 0, + "content": "
MedScribble (n=93)ACDC (n=1,140)
Model↑ Dice Score↓ HD95↑ Dice Score↓ HD95
SAM (ViT-b)0.40 ± 0.0532.58 ± 3.800.20 ± 0.01108.39 ± 0.85
SAM (ViT-h)0.56 ± 0.0514.61 ± 3.180.42 ± 0.0259.78 ± 2.02
SAM-Med2D w.o. adapter0.52 ± 0.0530.34 ± 3.290.16 ± 0.01107.25 ± 1.12
SAM-Med2D w/ adapter0.30 ± 0.0619.04 ± 3.630.17 ± 0.0241.01 ± 3.04
MIDepSeg0.68 ± 0.046.94 ± 1.060.58 ± 0.015.78 ± 0.26
MIDepSeg w/ refinement0.81 ± 0.033.10 ± 0.670.73 ± 0.012.71 ± 0.18
MedSAM (box)0.70 ± 0.047.54 ± 1.350.70 ± 0.023.53 ± 0.28
ScribblePrompt-UNet0.84 ± 0.022.92 ± 0.880.84 ± 0.011.80 ± 0.11
ScribblePrompt-SAM0.87 ± 0.022.61 ± 0.840.77 ± 0.014.17 ± 0.32
" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.625, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "ScribblePrompt" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.144, + 0.287, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.146, + 0.357, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.145, + 0.428, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.432, + 0.146, + 0.5, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.145, + 0.573, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.578, + 0.145, + 0.643, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.646, + 0.146, + 0.714, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.718, + 0.145, + 0.785, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.275, + 0.787, + 0.318 + ], + "angle": 0, + "content": "Fig. 5: Example predictions. \\(\\mathrm{SP} =\\) ScribblePrompt. Top: predictions after one step of manual scribbles. Bottom: predictions after five steps of simulated interactions (one center click followed by one center correction click per step)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.347, + 0.45, + 0.361 + ], + "angle": 0, + "content": "5.2 Simulated Interactions" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.372, + 0.784, + 0.402 + ], + "angle": 0, + "content": "Setup. We use our interaction simulator to evaluate the performance of ScribblePrompt and baselines with iterative scribbles across the 12 evaluation sets." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.403, + 0.61, + 0.418 + ], + "angle": 0, + "content": "For scribble-focused prompting procedures, we use:" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.427, + 0.785, + 0.456 + ], + "angle": 0, + "content": "- Line Scribbles: Three positive line scribbles and three negative line scribbles to start, followed by one correction line scribble at each step." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.457, + 0.785, + 0.486 + ], + "angle": 0, + "content": "- Centerline Scribbles: One positive and one negative centerline scribble to start, followed by one correction centerline scribble at each step." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.488, + 0.785, + 0.517 + ], + "angle": 0, + "content": "- Contour Scribbles: One positive and one negative contour scribble to start, followed by one correction contour scribble at each step." + }, + { + "type": "list", + "bbox": [ + 0.216, + 0.427, + 0.785, + 0.517 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.527, + 0.784, + 0.572 + ], + "angle": 0, + "content": "We limit each scribble to cover a maximum of \\( w \\) pixels, where \\( w \\) is the maximum dimension of the image. For centerline and contour scribbles, a scribble might contain multiple disconnected components." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.573, + 0.586, + 0.587 + ], + "angle": 0, + "content": "For click-focused prompting procedures, we use:" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.597, + 0.785, + 0.642 + ], + "angle": 0, + "content": "- Center Clicks: One positive click in the center of the largest component to start, followed by one (positive or negative) correction click per step in the center of the largest component of the error region." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.643, + 0.785, + 0.672 + ], + "angle": 0, + "content": "- Random Clicks: One random positive click to start, followed by one (positive or negative) correction click per step, randomly sampled from the error region." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.673, + 0.785, + 0.718 + ], + "angle": 0, + "content": "- Random Warm Start: Three random positive clicks and three random negative clicks to start, followed by one (positive or negative) correction click per step in the center of the largest component of the error region." + }, + { + "type": "list", + "bbox": [ + 0.216, + 0.597, + 0.785, + 0.718 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.728, + 0.784, + 0.789 + ], + "angle": 0, + "content": "For each example and interaction procedure, we simulate a series of iterative interactions with five random seeds. Since MIDepSeg cannot make accurate predictions with only a few clicks, we exclude it from simulations that start with a single click. For MedSAM, we show results for a bounding box prompt." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Results. Fig. 6 shows that both versions of ScribblePrompt outperform baseline methods for all simulated interaction procedures at all numbers of interactions." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.385, + 0.129 + ], + "angle": 0, + "content": "H.E. Wong et al." + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.144, + 0.363, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.366, + 0.144, + 0.503, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.144, + 0.645, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.144, + 0.787, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.267, + 0.788, + 0.31 + ], + "angle": 0, + "content": "Fig. 6: Simulated clicks and scribbles. We simulate interactions following three scribble protocols and three click protocols. We show more results and example predictions, with similar trends, in Supplementary Material Sec. F.2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.337, + 0.784, + 0.385 + ], + "angle": 0, + "content": "We show examples with simulated interactions in Fig. 1 and Fig. 5 (bottom). We show results with similar trends using bounding boxes and by dataset with fully-supervised baselines in Supplementary Material Sec. F." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.403, + 0.785, + 0.478 + ], + "angle": 0, + "content": "Discussion. Comparing the simulated scribble and click results (Fig. 6) highlights the efficiency of scribble-based interactions. Both ScribblePrompt models reach an average Dice above 0.8 in two scribble steps. It takes five steps of carefully-placed center clicks or eight random clicks to reach the same average Dice score with ScribblePrompt-SAM." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.48, + 0.785, + 0.541 + ], + "angle": 0, + "content": "Although SAM-Med2D was trained on a large biomedical imaging collection, it does not perform as well as SAM at refining its predictions on held-out evaluation sets. MedSAM only uses bounding box prompts and is not able to refine predictions, limiting its performance and usability." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.56, + 0.357, + 0.575 + ], + "angle": 0, + "content": "5.3 User Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.785, + 0.674 + ], + "angle": 0, + "content": "To assess the practical utility of ScribblePrompt, we conducted a user study with experienced annotators. We compare ScribblePrompt-UNet to SAM (ViT-b). We selected ScribblePrompt-UNet for the user study because it had similar performance and lower latency compared to ScribblePrompt-SAM. We selected SAM (ViT-b) because it had the highest Dice score in our experiments with clicks and lower latency than the next closest baseline, SAM (ViT-h)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.786, + 0.809 + ], + "angle": 0, + "content": "Setup. Study participants (\\(n = 16\\)) were neuroimaging researchers at an academic hospital. They used each model to segment a series of nine test images from different evaluation datasets. For each segmentation task, participants were shown the target segmentation and were asked to prompt the model until the predicted segmentation closely matched the target, or they could no longer improve the prediction. We provided participants with the target segmentation to disentangle the cognitive process of identifying the region of interest from prompting the model to achieve the desired segmentation. We provide additional details and visualizations in Supplementary Material Sec. G." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.786, + 0.842 + ], + "angle": 0, + "content": "Results. Participants produced more accurate segmentations (0.84 vs. 0.73 Dice; \\( p = 0.001 \\) using a paired t-test) using ScribblePrompt (Tab. 2). Partici-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.625, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "ScribblePrompt" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.785, + 0.174 + ], + "angle": 0, + "content": "Table 2: User study. Mean ± std. for Dice score, HD95 and total time per segmentation. Median (and max) number of times the participants refreshed the prediction." + }, + { + "type": "table", + "bbox": [ + 0.254, + 0.185, + 0.754, + 0.243 + ], + "angle": 0, + "content": "
Dice Score (↑)HD95 (↓)Sec./task (↓)Iter./task (↓)
SAM (ViT-b)0.73 ± 0.3016.1 ± 37.1131.7 ± 168.56 (47)
ScribblePrompt-UNet0.84 ± 0.134.5 ± 9.194.2 ± 89.03 (19)
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.27, + 0.784, + 0.331 + ], + "angle": 0, + "content": "pants spent \\(\\approx 1.5\\) minutes per segmentation on average using ScribblePrompt, compared to over 2 minutes per segmentation with SAM (\\(p = 0.02\\) using a paired t-test). While using ScribblePrompt, participants updated the prediction fewer times before being satisfied with the segmentation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.331, + 0.785, + 0.42 + ], + "angle": 0, + "content": "Upon completion, 15 out of 16 participants reported they preferred using ScribblePrompt to SAM and one participant had no preference. All participants reported it was easier to achieve the target segmentation using ScribblePrompt compared to SAM. \\(93.8\\%\\) of participants reported ScribblePrompt was better than SAM at refining its predictions in response to scribbles. \\(87.5\\%\\) of participants preferred using ScribblePrompt over SAM for clicks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.422, + 0.785, + 0.527 + ], + "angle": 0, + "content": "Discussion. The most common factors that influenced participant preference for ScribblePrompt were 1) being able get accurate predictions from multiple types of prompts, including scribbles and 2) responsiveness to their corrections, which enabled fine-grained control over the next prediction. For some tasks, such as retinal vein segmentation, SAM was unable to make accurate predictions even with many corrections. As a result, there was more variability in Dice score and time per segmentation, for SAM than ScribblePrompt." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.55, + 0.416, + 0.563 + ], + "angle": 0, + "content": "5.4 Inference Runtime" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.574, + 0.785, + 0.649 + ], + "angle": 0, + "content": "Setup. We evaluate computational efficiency by measuring inference time on a single CPU for one prediction with a scribble input covering 128 pixels. Performance on a CPU reflects practical utility. Because of a variety of barriers such as protected health information, users may not be able to send their data to a server for computation, and must rely on local, most often CPU-only, computing." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.65, + 0.785, + 0.74 + ], + "angle": 0, + "content": "Results. On a single CPU, ScribblePrompt-UNet requires \\(0.27 \\pm 0.04\\) sec per prediction, enabling the model to be used even in low-resource environments. The next most accurate model (Fig. 6), SAM ViT-h, requires over 2 minutes per prediction on a CPU (\\(130.79 \\pm 7.96\\) sec). SAM ViT-b takes \\(13.59 \\pm 0.77\\) seconds per prediction. We report results for all baselines and on GPU hardware with similar trends in Supplementary Material Sec. H." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.763, + 0.347, + 0.779 + ], + "angle": 0, + "content": "6 Ablations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.784, + 0.84 + ], + "angle": 0, + "content": "We conduct two ablations of important design decisions: (1) synthetic label inputs used during training, and (2) types of prompts simulated during training. We report results on the validation splits of nine datasets unseen during training." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.388, + 0.129 + ], + "angle": 0, + "content": "H.E. Wong et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.148, + 0.442, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.219, + 0.228, + 0.446, + 0.241 + ], + "angle": 0, + "content": "(a) Probability of synthetic labels." + }, + { + "type": "image", + "bbox": [ + 0.46, + 0.145, + 0.78, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.229, + 0.729, + 0.242 + ], + "angle": 0, + "content": "(b) Interactions during training." + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.252, + 0.788, + 0.294 + ], + "angle": 0, + "content": "Fig. 7: Ablations. We report Dice after five steps of simulated interactions following three inference-time interaction procedures. Error bars show \\(95\\%\\) CI. (a) shows mean change in Dice relative to ScribblePrompt-UNet trained without any synthetic labels." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.322, + 0.402, + 0.337 + ], + "angle": 0, + "content": "6.1 Synthetic Labels" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.344, + 0.788, + 0.435 + ], + "angle": 0, + "content": "Fig. 7a shows the effect of varying the probability of sampling a synthetic label during training for ScribblePrompt-UNet. Training with both real and synthetic labels improves generalization to new datasets, compared to training with only real labels. Using \\( p_{\\text{synth}} = 0.5 \\) results in the highest Dice. Training with only synthetic labels results in worse Dice scores. We show similar results for ScribblePrompt-SAM and other interactions in Supplementary Material Sec. I.1." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.455, + 0.384, + 0.47 + ], + "angle": 0, + "content": "6.2 Prompt Types" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.477, + 0.786, + 0.507 + ], + "angle": 0, + "content": "Setup. We evaluate ScribblePrompt-UNet models trained with different combinations of prompts, compared to the complete ScribblePrompt-UNet:" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.513, + 0.736, + 0.527 + ], + "angle": 0, + "content": "- ScribblePrompt-Unet (scribbles) trained on boxes and scribbles." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.528, + 0.686, + 0.542 + ], + "angle": 0, + "content": "- ScribblePrompt-UNet (clicks) trained on boxes and clicks." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.543, + 0.784, + 0.571 + ], + "angle": 0, + "content": "- ScribblePrompt-Unet (random clicks) trained on boxes and random clicks." + }, + { + "type": "list", + "bbox": [ + 0.215, + 0.513, + 0.784, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.579, + 0.788, + 0.641 + ], + "angle": 0, + "content": "Results. Fig. 7b shows ScribblePrompt-UNet trained with scribbles, clicks and bounding boxes predicts segmentations more accurately than do ablated versions of ScribblePrompt-UNet. We show results for other interaction procedures with similar trends in Supplementary Material Sec. I.2." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.661, + 0.36, + 0.677 + ], + "angle": 0, + "content": "7 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.789, + 0.841 + ], + "angle": 0, + "content": "We present ScribblePrompt, a practical framework for interactive segmentation that enables users to segment diverse medical images with scribbles, clicks, and bounding boxes. We introduce methods for simulating realistic user interactions and generating synthetic labels. These methods enable us to train models that generalize to unseen segmentation tasks and datasets. ScribblePrompt is more accurate than existing baselines, and ScribblePrompt-UNet is computationally efficient, even on a CPU. Our user study shows that nearly all users prefer ScribblePrompt and achieve segmentations with \\(15\\%\\) higher Dice with less effort than the next most accurate baseline. ScribblePrompt promises to significantly reduce the burden of manual segmentation in biomedical imaging." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.625, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "ScribblePrompt" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.283 + ], + "angle": 0, + "content": "This work was supported in part by funding from Quanta Computer Inc., the Eric and Wendy Schmidt Center at the Broad Institute of MIT and Harvard and the Wistron Corporation. Research reported in this paper was supported by the National Institute of Biomedical Imaging and Bioengineering of the National Institutes of Health under award number R01EB033773. Much of the computation resources required for this research was performed on computational hardware generously provided by the Massachusetts Life Sciences Center." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.302, + 0.323, + 0.318 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.231, + 0.332, + 0.784, + 0.359 + ], + "angle": 0, + "content": "1. Abdi, A.H., Kasaei, S., Mehdizadeh, M.: Automatic segmentation of mandible in panoramic x-ray. Journal of Medical Imaging 2(4), 044003 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.231, + 0.36, + 0.785, + 0.427 + ], + "angle": 0, + "content": "2. Agustsson, E., Uijlings, J.R., Ferrari, V.: Interactive Full Image Segmentation by Considering All Regions Jointly. In: 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11614-11623. IEEE, Long Beach, CA, USA (Jun 2019). https://doi.org/10.1109/CVPR.2019.01189, https://ieeexplore.ieee.org/document/8953526/" + }, + { + "type": "ref_text", + "bbox": [ + 0.231, + 0.429, + 0.785, + 0.522 + ], + "angle": 0, + "content": "3. Aine, C.J., Bockholt, H.J., Bustillo, J.R., Canive, J.M., Caprihan, A., Gasparovic, C., Hanlon, F.M., Houck, J.M., Jung, R.E., Lauriello, J., Liu, J., Mayer, A.R., Perrone-Bizzozero, N.I., Posse, S., Stephen, J.M., Turner, J.A., Clark, V.P., Calhoun, V.D.: Multimodal Neuroimaging in Schizophrenia: Description and Dissemination. Neuroinformatics 15(4), 343-364 (Oct 2017). https://doi.org/10.1007/s12021-017-9338-9, http://link.springer.com/10.1007/s12021-017-9338-9" + }, + { + "type": "ref_text", + "bbox": [ + 0.231, + 0.525, + 0.785, + 0.579 + ], + "angle": 0, + "content": "4. Al-Dhabyani, W., Gomaa, M., Khaled, H., Fahmy, A.: Dataset of breast ultrasound images. Data in Brief 28, 104863 (2020). https://doi.org/https://doi.org/10.1016/j.dib.2019.104863, https://www.sciencedirect.com/science/article/pii/S2352340919312181" + }, + { + "type": "ref_text", + "bbox": [ + 0.231, + 0.58, + 0.785, + 0.634 + ], + "angle": 0, + "content": "5. Asad, M., Fidon, L., Vercauteren, T.: ECONet: Efficient Convolutional Online Likelihood Network for Scribble-based Interactive Segmentation. In: International Conference on Medical Imaging with Deep Learning. pp. 35-47. PMLR (2022), http://arxiv.org/abs/2201.04584, arXiv:2201.04584 [cs, eess]" + }, + { + "type": "ref_text", + "bbox": [ + 0.231, + 0.635, + 0.785, + 0.689 + ], + "angle": 0, + "content": "6. Atzeni, A., Peter, L., Robinson, E., Blackburn, E., Althonayan, J., Alexander, D.C., Iglesias, J.E.: Deep active learning for suggestive segmentation of biomedical image stacks via optimisation of Dice scores and traced boundary length. Medical Image Analysis 81 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.231, + 0.689, + 0.785, + 0.743 + ], + "angle": 0, + "content": "7. Bai, J., Wu, X.: Error-Tolerant Scribbles Based Interactive Image Segmentation. In: 2014 IEEE Conference on Computer Vision and Pattern Recognition. pp. 392-399. IEEE, Columbus, OH, USA (Jun 2014). https://doi.org/10.1109/CVPR.2014.57, https://ieeexplore.ieee.org/document/6909451" + }, + { + "type": "ref_text", + "bbox": [ + 0.231, + 0.744, + 0.785, + 0.799 + ], + "angle": 0, + "content": "8. Baid, U., Ghodasara, S., Mohan, S., Bilello, M., Calabrese, E., Colak, E., Farahani, K., Kalpathy-Cramer, J., Kitamura, F.C., Pati, S., et al.: The rsna-asnrmiccai brats 2021 benchmark on brain tumor segmentation and radiogenomic classification. arXiv preprint arXiv:2107.02314 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.231, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "9. Benenson, R., Popov, S., Ferrari, V.: Large-scale interactive object segmentation with human annotators. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11700-11709 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.231, + 0.332, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.385, + 0.129 + ], + "angle": 0, + "content": "H.E. Wong et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.147, + 0.787, + 0.204 + ], + "angle": 0, + "content": "10. Bernard, O., Lalande, A., Zotti, C., Cervenansky, F., Yang, X., Heng, P.A., Cetin, I., Lekadir, K., Camara, O., Ballester, M.A.G., et al.: Deep learning techniques for automatic mri cardiac multi-structures segmentation and diagnosis: is the problem solved? IEEE transactions on medical imaging 37(11), 2514-2525 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.205, + 0.787, + 0.245 + ], + "angle": 0, + "content": "11. Bilic, P., Christ, P.F., Vorontsov, E., Chlebus, G., Chen, H., Dou, Q., Fu, C.W., Han, X., Heng, P.A., Hesser, J., et al.: The liver tumor segmentation benchmark (lits). arXiv preprint arXiv:1901.04056 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.246, + 0.787, + 0.3 + ], + "angle": 0, + "content": "12. Bloch, N., Madabhushi, A., Huisman, H., Freymann, J., Kirby, J., Grauer, M., Enquobahrie, A., Jaffe, C., Clarke, L., Farahani, K.: Nci-isbi 2013 challenge: automated segmentation of prostate structures. The Cancer Imaging Archive 370(6), 5 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.301, + 0.787, + 0.37 + ], + "angle": 0, + "content": "13. Boykov, Y., Jolly, M.P.: Interactive graph cuts for optimal boundary & region segmentation of objects in N-D images. In: Proceedings Eighth IEEE International Conference on Computer Vision. ICCV 2001. vol. 1, pp. 105-112. IEEE Comput. Soc, Vancouver, BC, Canada (2001). https://doi.org/10.1109/ICCV.2001.937505, http://ieeexplore.ieee.org/document/937505/" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.371, + 0.787, + 0.425 + ], + "angle": 0, + "content": "14. Bredell, G., Tanner, C., Konukoglu, E.: Iterative interaction training for segmentation editing networks. In: Machine Learning in Medical Imaging: 9th International Workshop, MLMI 2018, Held in Conjunction with MICCAI 2018, Granada, Spain, September 16, 2018, Proceedings 9. pp. 363-370. Springer (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.228, + 0.426, + 0.787, + 0.452 + ], + "angle": 0, + "content": "15. Butoi*, V.I., Ortiz*, J.J.G., Ma, T., Sabuncu, M.R., Guttag, J., Dalca, A.V.: Universeg: Universal medical image segmentation. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.453, + 0.787, + 0.535 + ], + "angle": 0, + "content": "16. Caicedo, J.C., Goodman, A., Karhohs, K.W., Cimini, B.A., Ackerman, J., Haghighi, M., Heng, C., Becker, T., Doan, M., McQuin, C., Rohhan, M., Singh, S., Carpenter, A.E.: Nucleus segmentation across imaging experiments: the 2018 Data Science Bowl. Nature Methods 16(12), 1247-1253 (Dec 2019). https://doi.org/10.1038/s41592-019-0612-7, https://doi.org/10.1038/s41592-019-0612-7" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.536, + 0.787, + 0.592 + ], + "angle": 0, + "content": "17. Cardona, A., Saalfeld, S., Preibisch, S., Schmid, B., Cheng, A., Pulokas, J., Tomancak, P., Hartenstein, V.: An integrated micro-and macroarchitectural analysis of the drosophila brain by computer-assisted serial section electron microscopy. PLoS biology 8(10), e1000502 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.592, + 0.787, + 0.633 + ], + "angle": 0, + "content": "18. Chen, X., Cheung, Y.S.J., Lim, S.N., Zhao, H.: ScribbleSeg: Scribble-based Interactive Image Segmentation (Mar 2023), http://arxiv.org/abs/2303.11320, arXiv:2303.11320 [cs]" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.634, + 0.787, + 0.675 + ], + "angle": 0, + "content": "19. Cheng, J., Ye, J., Deng, Z., Chen, J., Li, T., Wang, H., Su, Y., Huang, Z., Chen, J., Jiang, L., Sun, H., He, J., Zhang, S., Zhu, M., Qiao, Y.: SAM-Med2D (Aug 2023), http://arxiv.org/abs/2308.16184, arXiv:2308.16184 [cs]" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.675, + 0.787, + 0.757 + ], + "angle": 0, + "content": "20. Codella, N.C.F., Gutman, D.A., Celebi, M.E., Helba, B., Marchetti, M.A., Dusza, S.W., Kalloo, A., Liopyris, K., Mishra, N.K., Kittler, H., Halpern, A.: Skin lesion analysis toward melanoma detection: A challenge at the 2017 international symposium on biomedical imaging (isbi), hosted by the international skin imaging collaboration (ISIC). CoRR abs/1710.05006 (2017), http://arxiv.org/abs/1710.05006" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.758, + 0.787, + 0.84 + ], + "angle": 0, + "content": "21. Criminisi, A., Sharp, T., Blake, A.: GeoS: Geodesic Image Segmentation. In: Forsyth, D., Torr, P., Zisserman, A. (eds.) Computer Vision - ECCV 2008, vol. 5302, pp. 99-112. Springer Berlin Heidelberg, Berlin, Heidelberg (2008). https://doi.org/10.1007/978-3-540-88682-2_9, http://link.springer.com/10.1007/978-3-540-88682-2_9, series Title: Lecture Notes in Computer Science" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.625, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "ScribblePrompt" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "22. Czolbe, S., Dalca, A.V.: Neuralizer: General neuroimage analysis without retraining. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6217-6230 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.19, + 0.785, + 0.232 + ], + "angle": 0, + "content": "23. Dalca, A.V., Guttag, J., Sabuncu, M.R.: Anatomical priors in convolutional networks for unsupervised biomedical segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 9290-9299 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.232, + 0.785, + 0.259 + ], + "angle": 0, + "content": "24. Dice, L.R.: Measures of the amount of ecologic association between species. Ecology 26(3), 297-302 (1945)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.26, + 0.785, + 0.286 + ], + "angle": 0, + "content": "25. Felzenszwalb, P.F., Huttenlocher, D.P.: Efficient graph-based image segmentation. International journal of computer vision 59, 167-181 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.287, + 0.631, + 0.301 + ], + "angle": 0, + "content": "26. Fischl, B.: Freesurfer. Neuroimage 62(2), 774-781 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.301, + 0.785, + 0.342 + ], + "angle": 0, + "content": "27. Forte, M., Price, B., Cohen, S., Xu, N., Pitie, F.: Getting to \\(99\\%\\) Accuracy in Interactive Segmentation (Mar 2020), http://arxiv.org/abs/2003.07932, arXiv:2003.07932 [cs]" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.343, + 0.785, + 0.383 + ], + "angle": 0, + "content": "28. Gamper, J., Koohbanani, N., Benes, K., Graham, S., Jahanifar, M., Khurram, S., Azam, A., Hewitt, K., Rajpoot, N.: Pannuke dataset extension, insights and baselines. arxiv. 2020 doi: 10.48550. ARXIV (2003)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.384, + 0.785, + 0.438 + ], + "angle": 0, + "content": "29. Gerhard, S., Funke, J., Martel, J., Cardona, A., Fetter, R.: Segmented anisotropic ssTEM dataset of neural tissue (11 2013). https://doi.org/10.6084/m9.figshare.856713.v1, https://figshare.com/articles/dataset/Segmented_anisotropic_ssTEM_dataset_of_neural_tissue/856713" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.439, + 0.785, + 0.508 + ], + "angle": 0, + "content": "30. van Ginneken, B., Stegmann, M.B., Loog, M.: Segmentation of anatomical structures in chest radiographs using supervised methods: a comparative study on a public database. Medical Image Analysis 10(1), 19-40 (2006). https://doi.org/https://doi.org/10.1016/j.media.2005.02.002, https://www.sciencedirect.com/science/article/pii/S1361841505000368" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.509, + 0.785, + 0.563 + ], + "angle": 0, + "content": "31. Gollub, R.L., Shoemaker, J.M., King, M.D., White, T., Ehrlich, S., Sponheim, S.R., Clark, V.P., Turner, J.A., Mueller, B.A., Magnotta, V., et al.: The mcic collection: a shared repository of multi-modal, multi-site brain image data from a clinical investigation of schizophrenia. Neuroinformatics 11, 367-388 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.565, + 0.785, + 0.605 + ], + "angle": 0, + "content": "32. Gotkowski, K., Lüth, C., Jäger, P.F., Ziegler, S., Krämer, L., Denner, S., Xiao, S., Disch, N., Maier-Hein, K.H., Isensee, F.: Embarrassingly simple scribble supervision for 3d medical segmentation. arXiv preprint arXiv:2403.12834 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.606, + 0.785, + 0.659 + ], + "angle": 0, + "content": "33. Grady, L.: Random Walks for Image Segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence 28(11), 1768-1783 (Nov 2006). https://doi.org/10.1109/TPAMI.2006.233, http://ieeexplore.ieee.org/document/1704833/" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.66, + 0.785, + 0.702 + ], + "angle": 0, + "content": "34. Gut, D.: X-ray images of the hip joints 1 (Jul 2021). https://doi.org/10.17632/zm6bxzhmfz.1, https://data.mendeley.com/datasets/zm6bxzhmfz/1, publisher: Mendeley Data" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.702, + 0.785, + 0.743 + ], + "angle": 0, + "content": "35. He, K., Zhang, X., Ren, S., Sun, J.: Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In: Proceedings of the IEEE international conference on computer vision. pp. 1026-1034 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.744, + 0.785, + 0.785 + ], + "angle": 0, + "content": "36. He, S., Bao, R., Li, J., Stout, J., Bjornerud, A., Grant, P.E., Ou, Y.: Computer- vision benchmark segment-anything model (sam) in medical images: Accuracy in 12 datasets (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.785, + 0.84 + ], + "angle": 0, + "content": "37. Heller, N., Isensee, F., Maier-Hein, K.H., Hou, X., Xie, C., Li, F., Nan, Y., Mu, G., Lin, Z., Han, M., et al.: The state of the art in kidney and kidney tumor segmentation in contrast-enhanced ct imaging: Results of the kits19 challenge. Medical Image Analysis p. 101821 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.385, + 0.129 + ], + "angle": 0, + "content": "H.E. Wong et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.148, + 0.784, + 0.203 + ], + "angle": 0, + "content": "38. Hernandez Petzsche, M.R., de la Rosa, E., Hanning, U., Wiest, R., Valenzuela, W., Reyes, M., Meyer, M., Liew, S.L., Kofler, F., Ezhov, I., et al.: Isles 2022: A multi-center magnetic resonance imaging stroke lesion segmentation dataset. Scientific data 9(1), 762 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.203, + 0.785, + 0.243 + ], + "angle": 0, + "content": "39. Hoover, A., Kouznetsova, V., Goldbaum, M.: Locating blood vessels in retinal images by piecewise threshold probing of a matched filter response. IEEE Transactions on Medical imaging 19(3), 203-210 (2000)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.243, + 0.785, + 0.27 + ], + "angle": 0, + "content": "40. Hu, X., Xu, X., Shi, Y.: How to efficiently adapt large segmentation model(sam) to medical images (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.27, + 0.785, + 0.311 + ], + "angle": 0, + "content": "41. Huang, Y., Yang, X., Liu, L., Zhou, H., Chang, A., Zhou, X., Chen, R., Yu, J., Chen, J., Chen, C., Chi, H., Hu, X., Fan, D.P., Dong, F., Ni, D.: Segment anything model for medical images? (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.311, + 0.785, + 0.351 + ], + "angle": 0, + "content": "42. Huttenlocher, D.P., Klanderman, G.A., Rucklidge, W.J.: Comparing images using the hausdorff distance. IEEE Transactions on pattern analysis and machine intelligence 15(9), 850-863 (1993)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.351, + 0.785, + 0.405 + ], + "angle": 0, + "content": "43. Isensee, F., Jaeger, P.F., Kohl, S.A.A., Petersen, J., Maier-Hein, K.H.: nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nature Methods 18(2), 203-211 (Feb 2021). https://doi.org/10.1038/s41592-020-01008-z, http://www.nature.com/articles/s41592-020-01008-z" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.405, + 0.785, + 0.446 + ], + "angle": 0, + "content": "44. Ji, Y., Bai, H., Yang, J., Ge, C., Zhu, Y., Zhang, R., Li, Z., Zhang, L., Ma, W., Wan, X., et al.: Amos: A large-scale abdominal multi-organ benchmark for versatile medical image segmentation. arXiv preprint arXiv:2206.08023 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.446, + 0.785, + 0.555 + ], + "angle": 0, + "content": "45. Kavur, A.E., Gezer, N.S., Baris, M., Aslan, S., Conze, P.H., Groza, V., Pham, D.D., Chatterjee, S., Ernst, P., Ozkan, S., Baydar, B., Lachinov, D., Han, S., Pauli, J., Isensee, F., Perkonigg, M., Sathish, R., Rajan, R., Sheet, D., Dovletov, G., Speck, O., Nurnberger, A., Maier-Hein, K.H., Bozdagi Akar, G., Ünal, G., Dicle, O., Selver, M.A.: CHAOS Challenge - combined (CT-MR) healthy abdominal organ segmentation. Medical Image Analysis 69, 101950 (2021). https://doi.org/https://doi.org/10.1016/j.media.2020.101950, https://www.sciencedirect.com/science/article/pii/S1361841520303145" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.555, + 0.785, + 0.583 + ], + "angle": 0, + "content": "46. Kim, S., Oh, H.J., Min, S., Jeong, W.K.: Evaluation and improvement of segment anything model for interactive histopathology image segmentation (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.583, + 0.785, + 0.622 + ], + "angle": 0, + "content": "47. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.622, + 0.785, + 0.677 + ], + "angle": 0, + "content": "48. Krönke, M., Eilers, C., Dimova, D., Köhler, M., Buschner, G., Schweiger, L., Konstantinidou, L., Makowski, M., Nagarajah, J., Navab, N., et al.: Tracked 3d ultrasound and deep neural network-based thyroid segmentation reduce interobserver variability in thyroid volumetry. Plos one 17(7), e0268550 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.677, + 0.785, + 0.745 + ], + "angle": 0, + "content": "49. Kuijf, H.J., Biesbroek, J.M., De Bresser, J., Heinen, R., Andermatt, S., Bento, M., Berseth, M., Belyaev, M., Cardoso, M.J., Casamitjana, A., et al.: Standardized assessment of automatic segmentation of white matter hyperintensities and results of the wmh segmentation challenge. IEEE transactions on medical imaging 38(11), 2556-2568 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.745, + 0.785, + 0.799 + ], + "angle": 0, + "content": "50. Kuklisova-Murgasova, M., Aljabar, P., Srinivasan, L., Counsell, S.J., Doria, V., Serag, A., Gousias, I.S., Boardman, J.P., Rutherford, M.A., Edwards, A.D., et al.: A dynamic 4d probabilistic atlas of the developing brain. NeuroImage 54(4), 2750-2763 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.785, + 0.841 + ], + "angle": 0, + "content": "51. Lambert, Z., Petitjean, C., Dubray, B., Kuan, S.: Segthor: segmentation of thoracic organs at risk in ct images. In: 2020 Tenth International Conference on Image Processing Theory, Tools and Applications (IPTA). pp. 1-6. IEEE (2020)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.148, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.625, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "ScribblePrompt" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.204 + ], + "angle": 0, + "content": "52. Landman, B., Xu, Z., Igelsias, J., Styner, M., Langerak, T., Klein, A.: Miccai multi-atlas labeling beyond the cranial vault-workshop and challenge. In: Proc. MICCAI Multi-Atlas Labeling Beyond Cranial Vault Workshop Challenge. vol. 5, p. 12 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.205, + 0.785, + 0.26 + ], + "angle": 0, + "content": "53. Lee, H., Jeong, W.K.: Scribble2label: Scribble-supervised cell segmentation via self-generating pseudo-labels with consistency. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part I 23. pp. 14-23. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.261, + 0.785, + 0.316 + ], + "angle": 0, + "content": "54. Lemaître, G., Martí, R., Freixenet, J., Vilanova, J.C., Walker, P.M., Meriaudeau, F.: Computer-aided detection and diagnosis for prostate cancer based on mono and multi-parametric mri: a review. Computers in biology and medicine 60, 8-31 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.317, + 0.785, + 0.358 + ], + "angle": 0, + "content": "55. Li, M., Zhang, Y., Ji, Z., Xie, K., Yuan, S., Liu, Q., Chen, Q.: Ipn-v2 and octa-500: Methodology and dataset for retinal image segmentation. arXiv preprint arXiv:2012.07261 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.36, + 0.785, + 0.402 + ], + "angle": 0, + "content": "56. Li, Z., Zheng, Y., Luo, X., Shan, D., Hong, Q.: Scribblevc: Scribble-supervised medical image segmentation with vision-class embedding. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 3384-3393 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.402, + 0.785, + 0.444 + ], + "angle": 0, + "content": "57. Lin, D., Dai, J., Jia, J., He, K., Sun, J.: Scribblesup: Scribble-supervised convolutional networks for semantic segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3159-3167 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.445, + 0.785, + 0.487 + ], + "angle": 0, + "content": "58. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proceedings of the IEEE international conference on computer vision. pp. 2980-2988 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.488, + 0.785, + 0.529 + ], + "angle": 0, + "content": "59. Lin, X., Xiang, Y., Zhang, L., Yang, X., Yan, Z., Yu, L.: Samus: Adapting segment anything model for clinically-friendly and generalizable ultrasound image segmentation (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.53, + 0.785, + 0.585 + ], + "angle": 0, + "content": "60. Litjens, G., Toth, R., van de Ven, W., Hoeks, C., Kerkstra, S., van Ginneken, B., Vincent, G., Guillard, G., Birbeck, N., Zhang, J., et al.: Evaluation of prostate segmentation algorithms for mri: the promise12 challenge. Medical image analysis 18(2), 359-373 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.587, + 0.785, + 0.628 + ], + "angle": 0, + "content": "61. Liu, Q., Xu, Z., Bertasius, G., Niethammer, M.: Simpleclick: Interactive image segmentation with simple vision transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22290-22300 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.629, + 0.785, + 0.671 + ], + "angle": 0, + "content": "62. Liu, Q., Xu, Z., Jiao, Y., Niethammer, M.: iSegFormer: Interactive Segmentation via Transformers with Application to 3D Knee MR Images. In: Medical Image Computing and Computer Assisted Intervention (MICCAI) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.672, + 0.785, + 0.699 + ], + "angle": 0, + "content": "63. Ljosa, V., Sokolnicki, K.L., Carpenter, A.E.: Annotated high-throughput microscopy image sets for validation. Nature methods 9(7), 637-637 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.7, + 0.785, + 0.755 + ], + "angle": 0, + "content": "64. Löffler, M.T., Sekuboyina, A., Jacob, A., Grau, A.L., Scharr, A., El Husseinini, M., Kallweit, M., Zimmer, C., Baum, T., Kirschke, J.S.: A vertebral segmentation dataset with fracture grading. Radiology: Artificial Intelligence 2(4), e190138 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.757, + 0.785, + 0.784 + ], + "angle": 0, + "content": "65. in the Loop, H.: Teeth segmentation dataset, https://humansintheloop.org/resources/datasets/teeth-segmentation-dataset/" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.785, + 0.84 + ], + "angle": 0, + "content": "66. Luo, X., Hu, M., Liao, W., Zhai, S., Song, T., Wang, G., Zhang, S.: Scribblesupervised medical image segmentation via dual-branch network and dynamically mixed pseudo labels supervision. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 528-538. Springer (2022)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "20" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.385, + 0.129 + ], + "angle": 0, + "content": "H.E. Wong et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "67. Luo, X., Liao, W., Xiao, J., Song, T., Zhang, X., Li, K., Wang, G., Zhang, S.: Word: Revisiting organs segmentation in the whole abdominal region. arXiv preprint arXiv:2111.02403 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.191, + 0.787, + 0.245 + ], + "angle": 0, + "content": "68. Luo, X., Wang, G., Song, T., Zhang, J., Aertsen, M., Deprest, J., Ourselin, S., Vercauteren, T., Zhang, S.: MIDeepSeg: Minimally Interactive Segmentation of Unseen Objects from Medical Images Using Deep Learning. Medical Image Analysis 72, 102102 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.247, + 0.786, + 0.274 + ], + "angle": 0, + "content": "69. Ma, J., He, Y., Li, F., Han, L., You, C., Wang, B.: Segment anything in medical images. Nature Communications 15, 1-9 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.276, + 0.786, + 0.331 + ], + "angle": 0, + "content": "70. Ma, Y., Hao, H., Xie, J., Fu, H., Zhang, J., Yang, J., Wang, Z., Liu, J., Zheng, Y., Zhao, Y.: Rose: a retinal oct-angiography vessel segmentation dataset and new model. IEEE Transactions on Medical Imaging 40(3), 928–939 (2021). https://doi.org/10.1109/TMI.2020.3042802" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.333, + 0.786, + 0.374 + ], + "angle": 0, + "content": "71. Macdonald, J.A., Zhu, Z., Konkel, B., Mazurowski, M., Wiggins, W., Bashir, M.: Duke liver dataset (MRI) v2 (Apr 2023). https://doi.org/10.5281/zenodo.7774566, https://doi.org/10.5281/zenodo.7774566" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.375, + 0.786, + 0.403 + ], + "angle": 0, + "content": "72. Maninis, K.K., Caelles, S., Pont-Tuset, J., Van Gool, L.: Deep Extreme Cut: From Extreme Points to Object Segmentation (2018), arXiv:1711.09081" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.404, + 0.786, + 0.459 + ], + "angle": 0, + "content": "73. Marcus, D.S., Wang, T.H., Parker, J., Csernansky, J.G., Morris, J.C., Buckner, R.L.: Open access series of imaging studies (oasis): cross-sectional mri data in young, middle aged, nondemented, and demented older adults. Journal of cognitive neuroscience 19(9), 1498-1507 (2007)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.46, + 0.786, + 0.502 + ], + "angle": 0, + "content": "74. Marek, K., Jennings, D., Lasch, S., Siderowf, A., Tanner, C., Simuni, T., Coffey, C., Kieburtz, K., Flagg, E., Chowdhury, S., et al.: The parkinson progression marker initiative (ppmi). Progress in neurobiology 95(4), 629-635 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.503, + 0.786, + 0.572 + ], + "angle": 0, + "content": "75. Mazurowski, M.A., Clark, K., Czarnek, N.M., Shamsesfandabadi, P., Peters, K.B., Saha, A.: Radiogenomics of lower-grade glioma: algorithmically-assessed tumor shape is associated with tumor genomic subtypes and patient outcomes in a multi-institutional study with the cancer genome atlas data. Journal of neuro-oncology 133, 27-35 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.573, + 0.786, + 0.641 + ], + "angle": 0, + "content": "76. Mazurowski, M.A., Dong, H., Gu, H., Yang, J., Konz, N., Zhang, Y.: Segment anything model for medical image analysis: An experimental study. Medical Image Analysis 89, 102918 (Oct 2023). https://doi.org/10.1016/j.media.2023.102918, https://www.sciencedirect.com/science/article/pii/S1361841523001780" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.643, + 0.786, + 0.699 + ], + "angle": 0, + "content": "77. Men, J., Huang, Y., Solanki, J., Zeng, X., Alex, A., Jerwick, J., Zhang, Z., Tanzi, R.E., Li, A., Zhou, C.: Optical coherence tomography for brain imaging and developmental biology. IEEE Journal of Selected Topics in Quantum Electronics 22(4), 120-132 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.7, + 0.786, + 0.741 + ], + "angle": 0, + "content": "78. Milletari, F., Navab, N., Ahmadi, S.A.: V-net: Fully convolutional neural networks for volumetric medical image segmentation. Fourth International Conference on 3D Vision (3DV) pp. 565-571 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.742, + 0.786, + 0.784 + ], + "angle": 0, + "content": "79. Montoya, A., Hasnin, kaggle446, shirzad, Cukierski, W., yffud: Ultrasound nerve segmentation (2016), https://kaggle.com/competitions/ultrasound-nerve-segmentation" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.786, + 0.84 + ], + "angle": 0, + "content": "80. Ooi, A.Z.H., Embong, Z., Abd Hamid, A.I., Zainon, R., Wang, S.L., Ng, T.F., Hamzah, R.A., Teoh, S.S., Ibrahim, H.: Interactive Blood Vessel Segmentation from Retinal Fundus Image Based on Canny Edge Detector. Sensors 21(19), 6380 (Sep 2021)" + }, + { + "type": "list", + "bbox": [ + 0.223, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.625, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "ScribblePrompt" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "21" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "81. Paranjape, J.N., Nair, N.G., Sikder, S., Vedula, S.S., Patel, V.M.: Adaptivesam: Towards efficient tuning of sam for surgical scene segmentation. arXiv preprint arXiv:2308.03726 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.19, + 0.785, + 0.272 + ], + "angle": 0, + "content": "82. Pedraza, L., Vargas, C., Narvaez, F., Durán, O., Muñoz, E., Romero, E.: An open access thyroid ultrasound image database. In: Romero, E., Lepore, N. (eds.) 10th international symposium on medical information processing and analysis. vol. 9287, p. 92870W. SPIE / International Society for Optics and Photonics (2015). https://doi.org/10.1117/12.2073532, https://doi.org/10.1117/12.2073532" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.273, + 0.785, + 0.327 + ], + "angle": 0, + "content": "83. Philbrick, K.A., Weston, A.D., Akkus, Z., Kline, T.L., Korfiatis, P., Sakinis, T., Kostandy, P., Boonrod, A., Zeinoddini, A., Takahashi, N., et al.: Ril-contour: a medical imaging dataset annotation tool for and with deep learning. Journal of digital imaging 32, 571-581 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.328, + 0.785, + 0.398 + ], + "angle": 0, + "content": "84. Podobnik, G., Strojan, P., Peterlin, P., Ibragimov, B., Vrtovec, T.: HaN-Seg: The head and neck organ-at-risk CT and MR segmentation dataset. Medical Physics 50(3), 1917-1927 (2023). https://doi.org/https://doi.org/10.1002/mp.16197, https://aapm.onlinelibrary.wiley.com/doi/abs/10.1002/mp.16197, tex.eprint: https://aapm.onlinelibrary.wiley.com/doi/pdf/10.1002/mp.16197" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.398, + 0.785, + 0.439 + ], + "angle": 0, + "content": "85. Porwal, P., Pachade, S., Kamble, R., Kokare, M., Deshmukh, G., Sahasrabuddhe, V., Meriaudeau, F.: Indian diabetic retinopathy image dataset (idrid) (2018). https://doi.org/10.21227/H25W98, https://dx.doi.org/10.21227/H25W98" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.439, + 0.785, + 0.481 + ], + "angle": 0, + "content": "86. Radau, P., Lu, Y., Connelly, K., Paul, G., Dick, A., Wright, G.: Evaluation framework for algorithms segmenting short axis cardiac mri. The MIDAS Journal-Cardiac MR Left Ventricle Segmentation Challenge 49 (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.481, + 0.785, + 0.536 + ], + "angle": 0, + "content": "87. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.536, + 0.785, + 0.577 + ], + "angle": 0, + "content": "88. Raghu, M., Zhang, C., Kleinberg, J., Bengio, S.: Transfusion: Understanding transfer learning for medical imaging. Advances in neural information processing systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.577, + 0.785, + 0.647 + ], + "angle": 0, + "content": "89. Rajchl, M., Lee, M.C.H., Oktay, O., Kamnitsas, K., Passerat-Palmbach, J., Bai, W., Damodaram, M., Rutherford, M.A., Hajnal, J.V., Kainz, B., Rueckert, D.: DeepCut: Object Segmentation From Bounding Box Annotations Using Convolutional Neural Networks. IEEE Transactions on Medical Imaging 36(2), 674-683 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.647, + 0.785, + 0.702 + ], + "angle": 0, + "content": "90. Rakic, M., Wong, H.E., Ortiz, J.J.G., Cimini, B., Guttag, J.V., Dalca, A.V.: Tyche: Stochastic in-context learning for medical image segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.702, + 0.785, + 0.757 + ], + "angle": 0, + "content": "91. Rister, B., Yi, D., Shivakumar, K., Nobashi, T., Rubin, D.L.: CT-ORG, a new dataset for multiple organ segmentation in computed tomography. Scientific Data 7(1), 381 (Nov 2020). https://doi.org/10.1038/s41597-020-00715-8, https://www.nature.com/articles/s41597-020-00715-8" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.757, + 0.785, + 0.813 + ], + "angle": 0, + "content": "92. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "93. Roth, H.R., Yang, D., Xu, Z., Wang, X., Xu, D.: Going to Extremes: Weakly Supervised Medical Image Segmentation. Machine Learning and Knowledge Ex" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "22" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.385, + 0.129 + ], + "angle": 0, + "content": "H.E. Wong et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.251, + 0.148, + 0.786, + 0.175 + ], + "angle": 0, + "content": "traction 3(2), 507-524 (Jun 2021). https://doi.org/10.3390/make3020026, https://www.mdpi.com/2504-4990/3/2/26" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.176, + 0.787, + 0.217 + ], + "angle": 0, + "content": "94. Rother, C., Kolmogorov, V., Blake, A.: \"GrabCut\" — Interactive Foreground Extraction using Iterated Graph Cuts. ACM Transactions on Graphics 23, 309-314 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.218, + 0.787, + 0.259 + ], + "angle": 0, + "content": "95. Sakinis, T., Miletari, F., Roth, H., Korfiatis, P., Kostandy, P., Philbrick, K., Akkus, Z., Xu, Z., Xu, D., Erickson, B.J.: Interactive segmentation of medical images through fully convolutional neural networks (2019), arXiv:1903.08205" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.26, + 0.787, + 0.314 + ], + "angle": 0, + "content": "96. Saporta, A., Gui, X., Agrawal, A., Pareek, A., Truong, S., Nguyen, C., Ngo, V.D., Seekins, J., Blankenberg, F.G., Ng, A., et al.: Deep learning saliency maps do not accurately highlight diagnostically relevant regions for medical image interpretation. MedRxiv (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.315, + 0.787, + 0.355 + ], + "angle": 0, + "content": "97. Sati, P., George, I.C., Shea, C.D., Gaitan, M.I., Reich, D.S.: Flair*: a combined mr contrast technique for visualizing white matter lesions and parenchymal veins. Radiology 265(3), 926-932 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.356, + 0.787, + 0.412 + ], + "angle": 0, + "content": "98. Seibold, C., Reif, S., Sarfraz, S., Fink, M.A., Mayer, V., Sellner, J., Kim, M.S., Maier-Hein, K.H., Kleesiek, J., Stiefelhagen, R.: Detailed annotations of chest x-rays via ct projection for report understanding. In: Proceedings of the 33th British Machine Vision Conference (BMVC) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.412, + 0.787, + 0.48 + ], + "angle": 0, + "content": "99. Setio, A.A.A., Traverso, A., De Bel, T., Berens, M.S., Van Den Bogaard, C., Cerello, P., Chen, H., Dou, Q., Fantacci, M.E., Geurts, B., et al.: Validation, comparison, and combination of algorithms for automatic detection of pulmonary nodules in computed tomography images: the luna16 challenge. Medical image analysis 42, 1-13 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.481, + 0.786, + 0.522 + ], + "angle": 0, + "content": "100. Shi, P., Qiu, J., Abaxi, S.M.D., Wei, H., Lo, F.P.W., Yuan, W.: Generalist vision foundation models for medical imaging: A case study of segment anything model on zero-shot medical segmentation. Diagnostics 13(11), 1947 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.523, + 0.786, + 0.578 + ], + "angle": 0, + "content": "101. Simpson, A.L., Antonelli, M., Bakas, S., Bilello, M., Farahani, K., Van Ginneken, B., Kopp-Schneider, A., Landman, B.A., Litjens, G., Menze, B., et al.: A large annotated medical image dataset for the development and evaluation of segmentation algorithms. arXiv preprint arXiv:1902.09063 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.578, + 0.786, + 0.647 + ], + "angle": 0, + "content": "102. Sofiuk, K., Petrov, I., Barinova, O., Konushin, A.: F-BRS: Rethinking Backpropagating Refinement for Interactive Segmentation. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8620-8629. IEEE, Seattle, WA, USA (Jun 2020). https://doi.org/10.1109/CVPR42600.2020.00865, https://ieeexplore.ieee.org/document/9156403/" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.647, + 0.786, + 0.688 + ], + "angle": 0, + "content": "103. Sofiuk, K., Petrov, I.A., Konushin, A.: Reviving Iterative Training with Mask Guidance for Interactive Segmentation (Feb 2021), http://arxiv.org/abs/2102.06583, arXiv:2102.06583 [cs]" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.689, + 0.787, + 0.757 + ], + "angle": 0, + "content": "104. Song, Y., Zheng, J., Lei, L., Ni, Z., Zhao, B., Hu, Y.: CT2US: Cross-modal transfer learning for kidney segmentation in ultrasound images with synthesized data. Ultrasonics 122, 106706 (2022). https://doi.org/https://doi.org/10.1016/j.ultras.2022.106706, https://www.sciencedirect.com/science/article/pii/S0041624X22000191" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.758, + 0.787, + 0.799 + ], + "angle": 0, + "content": "105. Staal, J., Abramoff, M.D., Niemeijer, M., Viergever, M.A., Van Ginneken, B.: Ridge-based vessel segmentation in color images of the retina. IEEE transactions on medical imaging 23(4), 501-509 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "106. Tang, Y., Yang, D., Li, W., Roth, H.R., Landman, B., Xu, D., Nath, V., Hatamizadeh, A.: Self-Supervised Pre-Training of Swin Transformers for 3D Medical Image Analysis. In: 2022 IEEE/CVF Conference on Computer Vision" + }, + { + "type": "list", + "bbox": [ + 0.219, + 0.148, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.625, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "ScribblePrompt" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "23" + }, + { + "type": "ref_text", + "bbox": [ + 0.249, + 0.147, + 0.786, + 0.189 + ], + "angle": 0, + "content": "and Pattern Recognition (CVPR). pp. 20698-20708. IEEE, New Orleans, LA, USA (Jun 2022). https://doi.org/10.1109/CVPR52688.2022.02007, https://ieeexplore.ieee.org/document/9879123/" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.218 + ], + "angle": 0, + "content": "107. Vezhnevets, V., Konouchine, V.: \"GrowCut\" - Interactive Multi-Label N-D Image Segmentation By Cellular Automata. proc. of Graphicon 1(4), 150-156 (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.787, + 0.245 + ], + "angle": 0, + "content": "108. Wang, C., Chen, X., Ning, H., Li, S.: Sam-octa: A fine-tuning strategy for applying foundation model to octa image segmentation tasks (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.245, + 0.787, + 0.3 + ], + "angle": 0, + "content": "109. Wang, G., Li, W., Zuluaga, M.A., Pratt, R., Patel, P.A., Aertsen, M., Doel, T., David, A.L., Deprest, J., Ourselin, S., Vercauteren, T.: Interactive Medical Image Segmentation Using Deep Learning With Image-Specific Fine Tuning. IEEE Transactions on Medical Imaging 37(7), 1562-1573 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.3, + 0.787, + 0.357 + ], + "angle": 0, + "content": "110. Wang, G., Zuluaga, M.A., Li, W., Pratt, R., Patel, P.A., Aertsen, M., Doel, T., David, A.L., Deprest, J., Ourselin, S., Vercauteren, T.: DeepIGeoS: A Deep Interactive Geodesic Framework for Medical Image Segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence 41(7), 1559-1572 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.356, + 0.787, + 0.424 + ], + "angle": 0, + "content": "111. Wang, G., Zuluaga, M.A., Pratt, R., Aertsen, M., Doel, T., Klusmann, M., David, A.L., Deprest, J., Vercauteren, T., Ourselin, S.: Slic-Seg: A minimally interactive segmentation of the placenta from sparse and motion-corrupted fetal MRI in multiple views. Medical Image Analysis 34, 137-147 (Dec 2016). https://doi.org/10.1016/j.media.2016.04.009" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.424, + 0.787, + 0.467 + ], + "angle": 0, + "content": "112. Wang, X., Wang, W., Cao, Y., Shen, C., Huang, T.: Images speak in images: A generalist painter for in-context visual learning. arXiv preprint arXiv:2212.02499 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.466, + 0.787, + 0.509 + ], + "angle": 0, + "content": "113. Wang, X., Zhang, X., Cao, Y., Wang, W., Shen, C., Huang, T.: SegGPT: Segmenting Everything In Context (Apr 2023), http://arxiv.org/abs/2304.03284, arXiv:2304.03284 [cs]" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.508, + 0.787, + 0.564 + ], + "angle": 0, + "content": "114. Wasserthal, J., Breit, H.C., Meyer, M.T., Pradella, M., Hinck, D., Sauter, A.W., Heye, T., Boll, D.T., Cyriac, J., Yang, S., et al.: Totalsegmentator: Robust segmentation of 104 anatomic structures in ct images. Radiology: Artificial Intelligence 5(5) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.564, + 0.787, + 0.607 + ], + "angle": 0, + "content": "115. Wu, J., Fu, R., Fang, H., Liu, Y., Wang, Z., Xu, Y., Jin, Y., Arbel, T.: Medical SAM Adapter: Adapting Segment Anything Model for Medical Image Segmentation (Apr 2023), http://arxiv.org/abs/2304.12620, arXiv:2304.12620 [cs]" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.606, + 0.787, + 0.661 + ], + "angle": 0, + "content": "116. Xu, N., Price, B., Cohen, S., Yang, J., Huang, T.: Deep Interactive Object Selection. In: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 373-381. IEEE, Las Vegas, NV, USA (Jun 2016). https://doi.org/10.1109/CVPR.2016.47, http://ieeexplore.ieee.org/document/7780416/" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.661, + 0.787, + 0.688 + ], + "angle": 0, + "content": "117. Xu, N., Price, B., Cohen, S., Yang, J., Huang, T.: Deep GrabCut for Object Selection. arXiv (2017), arXiv:1707.00243" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.688, + 0.787, + 0.731 + ], + "angle": 0, + "content": "118. Yushkevich, P.A., Piven, J., Hazlett, H.C., Smith, R.G., Ho, S., Gee, J.C., Gering, G.: User-guided 3d active contour segmentation of anatomical structures: significantly improved efficiency and reliability. Neuroimage 31(3), 1116-1128 (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.73, + 0.787, + 0.772 + ], + "angle": 0, + "content": "119. Zhang, J., Ding, X., Hu, D., Jiang, Y.: Semantic segmentation of Covid-19 lesions with a multiscale dilated convolutional network. Scientific Reports 12(1), 1847 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.772, + 0.787, + 0.799 + ], + "angle": 0, + "content": "120. Zhang, K., Liu, D.: Customized segment anything model for medical image segmentation (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.799, + 0.787, + 0.842 + ], + "angle": 0, + "content": "121. Zhang, K., Zhuang, X.: Cycle mix: A holistic strategy for medical image segmentation from scribble supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11656-11665 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.842 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "24" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.385, + 0.129 + ], + "angle": 0, + "content": "H.E. Wong et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "122. Zhang, S., Liew, J.H., Wei, Y., Wei, S., Zhao, Y.: Interactive Object Segmentation With Inside-Outside Guidance. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 12231-12241. IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.231 + ], + "angle": 0, + "content": "123. Zhang, T.Y., Suen, C.Y.: A fast parallel algorithm for thinning digital patterns. Commun. ACM 27(3), 236-239 (mar 1984). https://doi.org/10.1145/357994.358023, https://doi.org/10.1145/357994.358023" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.231, + 0.787, + 0.272 + ], + "angle": 0, + "content": "124. Zhang, Y., Xian, M., Cheng, H.D., Shareef, B., Ding, J., Xu, F., Huang, K., Zhang, B., Ning, C., Wang, Y.: Busis: A benchmark for breast ultrasound image segmentation. In: Healthcare. vol. 10, p. 729. MDPI (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.272, + 0.787, + 0.313 + ], + "angle": 0, + "content": "125. Zhao, Q., Lyu, S., Bai, W., Cai, L., Liu, B., Wu, M., Sang, X., Yang, M., Chen, L.: A multi-modality ovarian tumor ultrasound image dataset for unsupervised cross-domain semantic segmentation. CoRR abs/2207.06799 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.314, + 0.787, + 0.37 + ], + "angle": 0, + "content": "126. Zheng, G., Chu, C., Belavy, D.L., Ibragimov, B., Korez, R., Vrtovec, T., Hutt, H., Everson, R., Meakin, J., Andrade, I.L., et al.: Evaluation and comparison of 3d intervertebral disc localization and segmentation methods for 3d t2 mr data: A grand challenge. Medical image analysis 35, 327-344 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.37, + 0.787, + 0.424 + ], + "angle": 0, + "content": "127. Zheng, X., Wang, Y., Wang, G., Liu, J.: Fast and robust segmentation of white blood cell images by self-supervised learning. Micron 107, 55-71 (2018). https://doi.org/https://doi.org/10.1016/j.micron.2018.01.010, https://www.sciencedirect.com/science/article/pii/S0968432817303037" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.424, + 0.787, + 0.48 + ], + "angle": 0, + "content": "128. Zhou, T., Li, L., Bredell, G., Li, J., Unkelbach, J., Konukoglu, E.: Volumetric memory network for interactive medical image segmentation. Medical Image Analysis 83, 102599 (Jan 2023). https://doi.org/10.1016/j.media.2022.102599, https://linkinghub.elsevier.com/retrieve/pii/S1361841522002316" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.48, + 0.787, + 0.522 + ], + "angle": 0, + "content": "129. Zou, X., Yang, J., Zhang, H., Li, F., Li, L., Wang, J., Wang, L., Gao, J., Lee, Y.J.: Segment everything everywhere all at once. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.522 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/ScribblePrompt_ Fast and Flexible Interactive Segmentation for Any Biomedical Image/97fd2117-c4b3-47cc-8dc3-f7ab8ca17648_origin.pdf b/2024/ScribblePrompt_ Fast and Flexible Interactive Segmentation for Any Biomedical Image/97fd2117-c4b3-47cc-8dc3-f7ab8ca17648_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..54f2ebc48b29988189829673e3fab379f63198cf --- /dev/null +++ b/2024/ScribblePrompt_ Fast and Flexible Interactive Segmentation for Any Biomedical Image/97fd2117-c4b3-47cc-8dc3-f7ab8ca17648_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:446248620d6ae9258b043ea777296d14bdd2743b9fe428baf43ebe5b826d11d6 +size 2616590 diff --git a/2024/ScribblePrompt_ Fast and Flexible Interactive Segmentation for Any Biomedical Image/full.md b/2024/ScribblePrompt_ Fast and Flexible Interactive Segmentation for Any Biomedical Image/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9001902f38f1d4d67970615f035bafeec00cc9c1 --- /dev/null +++ b/2024/ScribblePrompt_ Fast and Flexible Interactive Segmentation for Any Biomedical Image/full.md @@ -0,0 +1,461 @@ +# ScribblePrompt: Fast and Flexible Interactive Segmentation for Any Biomedical Image + +Hallee E. Wong $^{1,2}$ , Marianne Rakic $^{1,2}$ , John Guttag $^{1}$ , and Adrian V. Dalca $^{1,2,3}$ + +1 MIT CSAIL, Cambridge, MA, USA +2 Martins Center, Massachusetts General Hospital, Charlestown, MA, USA +3 Harvard Medical School, Boston, MA, USA +{hallel, mrukic, guttag, adalca}@mit.edu + +Abstract. Biomedical image segmentation is a crucial part of both scientific research and clinical care. With enough labelled data, deep learning models can be trained to accurately automate specific biomedical image segmentation tasks. However, manually segmenting images to create training data is highly labor intensive and requires domain expertise. We present ScribblePrompt, a flexible neural network based interactive segmentation tool for biomedical imaging that enables human annotators to segment previously unseen structures using scribbles, clicks, and bounding boxes. Through rigorous quantitative experiments, we demonstrate that given comparable amounts of interaction, ScribblePrompt produces more accurate segmentations than previous methods on datasets unseen during training. In a user study with domain experts, ScribblePrompt reduced annotation time by $28\%$ while improving Dice by $15\%$ compared to the next best method. ScribblePrompt's success rests on a set of careful design decisions. These include a training strategy that incorporates both a highly diverse set of images and tasks, novel algorithms for simulated user interactions and labels, and a network that enables fast inference. We showcase ScribblePrompt in an interactive demo, provide code, and release a dataset of scribble annotations at https://scribbleprompt.csail.mit.edu + +Keywords: Interactive Segmentation $\cdot$ Biomedical Imaging $\cdot$ Scribbles + +# 1 Introduction + +Biomedical image segmentation is an essential step in a wide range of biomedical research and clinical care pipelines. Deep learning has become the predominant method to automate existing segmentation tasks [43, 78, 92, 106]. Biomedical researchers and clinicians often encounter novel segmentation tasks involving either new regions of interest [119] or new image modalities [77,97]. Unfortunately, supervised training of accurate models for new domains requires diverse images with careful annotations by skilled experts. + +Most widely used interactive segmentation systems for biomedical imaging provide minimal, intensity-based, algorithmic assistance [118]. Despite a growing literature, learning-based systems are less widely used in practice, perhaps + +![](images/ab2a2cc597ac03ffede42436570130f5d75b3c0c560ce4b4d14c3225dabb1cb7.jpg) +Step 1 + +![](images/0ac61d155e098a22eb0ae8e997599603d95e09b732829390570cc097b4c7492f.jpg) +Step 2 + +![](images/a9016d7deb2cfa31379d1f933b4d180a89a4b1622c9286b10f94aa82d44d1668.jpg) +Step 3 + +![](images/24a6138a080e8ccd473dd682cfeeefe45d4cc88a6c12a2f401143874bfcca03b.jpg) +Ground Truth + +![](images/2312047fc6d80b06fdcb4340e676ff1e78f67d7bb6a341a2722a33f00a12168a.jpg) +Step 1 + +![](images/9552129b489b034cf5e7e178f3b4cba83f8b8e43cd1c979ac1c80fe1143dbb69.jpg) +Step 2 + +![](images/28d9093f08efce348f0b1855184a0cf61b42d36ec5c953a6820d223e8766a5c7.jpg) +Step 3 + +![](images/d4349865b1db0bd39b571fe9e818c8d814f9b6f77b6044e25ae1494740a5ba2f.jpg) +Ground Truth + +![](images/697261cfefe8c33764db424f08fe24120cbb58f3a778970487dcf73e44b30bab.jpg) +Fig. 1: ScribblePrompt enables rapid iterative interactive segmentation of unseen tasks using bounding boxes, clicks, and scribbles. We show predictions from ScribblePrompt with iterative interaction steps on examples from datasets unseen during training. At each step, we visualize positive scribble and click inputs in green, negative scribble and click inputs in red, bounding box inputs in yellow, and the predicted segmentation in blue. Scribble thickness is enlarged for visual clarity. See Supplementary Material for more examples. + +![](images/bad9e3e9b10dfb4f7f8e31e5c54b7d379a03a82597db2034c49ae538bfcfa64b.jpg) + +![](images/35083b4c22cb5980a2ac19243b6ca7f5f39a1a1dee79373523a411b4ba0bf012.jpg) + +![](images/dd6305b1025ea0780e39f7fd7038aa1dc28f2fe9dbc210af503f1fa720ef0bf2.jpg) + +![](images/56a8c466e30b316a0623ae88158886254c05eed26dea0283f90a63e8ac606546.jpg) + +![](images/60b534ecd8ee3b27f4b6037d885c698de56b2cae4777a65d37f36718123a0965.jpg) + +![](images/b980f866bee65adbd815cbb0c10574c542e061fa5e67d7cc7f789a140699f31b.jpg) + +![](images/b663fb298dc9eec1c53be49c8be425b85f38b1a0b789db3f0d9e3b57feb3b85d.jpg) + +because they are generally focused on specific tasks or modalities, which limits their applicability [68, 83, 95, 109, 110, 128]. Recent vision foundation models target broad applicability, but require fine-tuning to the medical domain [47, 112, 113, 129]. Most interactive segmentation models require specific interactions, like carefully placed clicks, making them easy to develop but difficult to use in practice. The result of these shortcomings is that existing learning-based models are not widely used in practice. + +In this work, we present ScribblePrompt, a general model for interactive biomedical image segmentation that + +1. enables users to rapidly and accurately accomplish any biomedical image segmentation task, outperforming state-of-the-art models, particularly for unseen labels and image types; +2. is flexible to different annotation styles, including bounding boxes, clicks, and scribbles, +3. is computationally efficient, enabling fast inference, even on a single CPU. + +To achieve this, we focused on design decisions aimed at ease of use and realistic interactions. To train our model, we started by gathering a large corpus of biomedical imaging datasets. We then designed task augmentation through synthesis strategies to encourage generalization to unseen tasks. Finally, we designed new interaction simulation strategies for training. The model itself uses an architecture optimized for efficient inference. + +To evaluate ScribblePrompt, we compared its performance to that of existing interactive segmentation methods on datasets unseen during training. These experiments involved manually-collected scribbles, simulated interactions (Fig. 1), and a user study in which experienced annotators were asked to segment images. In the user study, ScribblePrompt reduced annotation time by $28\%$ while increasing Dice by $15\%$ compared to the next best method, the Segment Anything Model (SAM) [47]. Given similar amounts of interaction, ScribblePrompt + +achieved consistently better Dice scores than other methods. We release an interactive tool, code, model weights, and our dataset of manually-collected scribbles at https://scribbleprompt.csail.mit.edu. + +# 2 Related Works + +Interactive Biomedical Image Segmentation. Early research into interactive segmentation of biomedical images focused on traditional intensity-based methods [13,21,33,107,118]. Recent deep-learning based techniques [68,83,109, 110] use human interaction to improve the accuracy of segmentation tasks where labelled data exists, but fully-automatic segmentation methods fail to produce accurate segmentations. This approach assumes that the model will be used to perform the same segmentation task(s) it was trained on, leading to domain-specific solutions [6,62,80,89,110,111]. A few methods generalize to new, but similar, classes [95] and modalities [68,109] to those seen in training. In contrast, ScribblePrompt is designed to help annotators segment a wide range of biomedical images and (potentially unseen) tasks at inference time. + +Foundation Models. Recent vision foundation models employ prompting to enable generalization to new tasks. These models, trained on large collections of (natural) image data, segment potentially-unseen structures specified by spatial prompts [47, 129], text [47, 87, 129], or examples [15, 22, 112, 113]. Some of these models can be used for interactive segmentation with spatial prompts. + +Fine-tuning models initially developed using natural images is often unhelpful for specific tasks in the biomedical imaging domain, compared to training from scratch [88]. The Segment Anything Model (SAM) [47], a foundation model trained on natural images, performs well at biomedical image segmentation tasks with clear boundaries (e.g. organs in abdominal CT), but performs poorly on tasks and modalities involving more subtle delineations (e.g. deep structures in Brain MRI) [36, 41, 76, 100]. Several recent or concurrent papers fine-tune SAM for specific biomedical tasks or modalities [40, 46, 59, 81, 108, 120]. A few fine-tune SAM to segment medical images from multiple-modalities with bounding boxes [69], clicks [115] or both [19]. We show in our experiments that these models don't perform well on many unseen tasks and are limiting in their required interactions. + +Several natural image methods simulate iterative user interactions and condition the model on previous predictions to train a single network to make the initial prediction and perform refinement [14, 27, 47, 61, 102, 103]. This approach is appropriate for natural images where there is ample data and typically one modality. However, such methods have not previously been developed for the biomedical imaging domain, which is more fragmented because of specialized regions of interest and different modalities. We use iterative ideas in ScribblePrompt, but we facilitate iterative interactive segmentation of unseen biomedical imaging tasks. + +User Interaction. Users can prompt interactive segmentation systems in many different ways, such as bounding boxes [9, 47, 89, 94, 109, 117], scribbles + +[18, 110, 111], or clicks [9, 61, 68, 103, 110]. Few interactive segmentation models incorporate multiple types of inputs. Some works have explored training with more specialized types of clicks [122], such as extreme points [72, 93], interior margin points [68], and center clicks [103, 116], to maximize the information per interaction. Specialized interactions often require more user time, and models trained only with such interactions are less robust to deviations from the interaction protocol [9, 103, 117]. In contrast, we focus on a user-friendly model that performs well under a variety of interaction scenarios. + +Scribbles are an intuitive form of interaction, however few interactive segmentation works employ them, because acquiring realistic scribbles for training and evaluation is challenging. Previous deep-learning models trained for interactive scribble-based segmentation either relied on collecting large datasets of manually drawn scribbles [53, 57, 128] or simplistic approaches such as sampling random points [5, 7, 110] or Bezier curves [2]. One work has explored simulating more complex scribbles for interactive segmentation of natural images [18]. We build on these concepts to create a new scribble simulation engine for training and evaluation. + +Scribble-supervised learning methods use scribble annotations as supervision to train automatic segmentation models for predicting segmentation given only an input image [32,56,57,66,121]. However, these methods require the manual scribble-annotation of many training images and retraining for each new task. In contrast, ScribblePrompt is trained on a large corpus of datasets once and then can be used to perform new segmentation tasks at inference time without retraining, using scribbles as input. + +# 3 ScribblePrompt Approach + +We present an interactive segmentation method that can generalize to new biomedical imaging modalities and regions of interest, while staying focused on practical usability. We describe the problem formulation, and present the important aspects of the ScribblePrompt framework: (i) simulation of realistic interactions during training, (ii) augmentation with synthetic labels during training to encourage generalization, and (iii) an efficient architecture for fast inference. We show an overview of training in Fig. 2 and 4. + +# 3.1 Problem Formulation + +Let $t$ be a segmentation task consisting of image and segmentation pairs, $\{(x^t,y^t)_j\}_{j = 1}^N$ . At step $i$ , given an image $x^{t}$ , a set of user interactions $u_{i}$ , and previous prediction $\hat{y}_{i - 1}^{t}$ , we learn function $f_{\theta}(x^{t},u_{i},\hat{y}_{i - 1}^{t})$ with parameters $\theta$ that produces a segmentation $\hat{y}_i$ . The set of interactions $u_{i}$ , which may include positive or negative scribbles, positive or negative clicks, and bounding boxes, is provided by a user who has access to the image $x^{t}$ and previous prediction $\hat{y}_{i - 1}^{t}$ . + +We minimize the difference between the true segmentation $y^{t}$ and each of the $k$ iterative predictions $\hat{y}_1, \dots, \hat{y}_k$ , + +$$ +\mathcal {L} (\theta ; \mathcal {T}) = \mathbb {E} _ {t \in \mathcal {T}} \left[ \mathbb {E} _ {(x ^ {t}, y ^ {t}) \in t} \left[ \sum_ {i = 1} ^ {k} \mathcal {L} _ {\text {S e g}} \left(y ^ {t}, f _ {\theta} \left(x ^ {t}, u _ {i}, \hat {y} _ {i - 1} ^ {t}\right)\right) \right] \right], \tag {1} +$$ + +where $\mathcal{L}_{Seg}$ is a supervised segmentation loss. + +During training, we sample a task $t$ from training task collection $\mathcal{T}$ , from which we sample an image $x^t$ , and segmentation map $y^{t}$ . We simulate a possible set of interactions $u_{i}$ based on $y^{t}$ , and predict $\hat{y}_i^t = f_\theta (x^t,u_i,\hat{y}_{i - 1}^t)$ . We simulate the next set of interactions $u_{i + 1}$ based on the error region $\varepsilon_i^t = y^t -\hat{y}_i^t$ , and repeat for $k$ iterations. In the following sections, we describe the core aspects of the framework: strategies for simulating $u_{i}$ , collecting $\mathcal{T}$ , sampling $(x^{t},y^{t})$ , building $f_{\theta}$ , and optimizing $\mathcal{L}(\theta ;\mathcal{T})$ . + +![](images/f67e2ca5adc2a5a4162dc1ba591ce497f51e904eb348184d2b18087b5a5cbb82.jpg) +Fig. 2: Training. We simulate $k$ consecutive steps of interactive segmentation. Given an image segmentation pair $(x^t, y^t)$ , we first simulate a set of initial interactions $u_1$ , which may contain bounding boxes, clicks, and/or scribbles. We predict segmentation $\hat{y}_1^t \coloneqq f_\theta(x^t, u_1, \hat{y}_0^t)$ where the initial prediction $\hat{y}_0^t$ is set to zeros. In the second step, we simulate corrections using the error region $\varepsilon_1^t$ between the previous prediction $\hat{y}_1^t$ and ground truth $y^t$ , and add them to the set of initial interactions $u_1$ to get $u_2$ . We predict segmentation $\hat{y}_2^t \coloneqq f_\theta(x^t, u_2, \hat{y}_1^t)$ and repeat to produce a series of predictions, $\hat{y}_1^t, \dots, \hat{y}_k^t$ . We learn $\theta$ to minimize $\sum_{i=1}^{k} \mathcal{L}_{seg}(y^t, \hat{y}_i^t)$ , the sum of losses between the target segmentation $y^t$ and iterative predictions $\hat{y}_1^t, \dots, \hat{y}_k^t$ . + +# 3.2 Prompt Simulation + +To enable a practical, easy-to-use model, we encourage robustness to different types of user interactions $u_{i}$ . We introduce algorithms for simulating scribbles, clicks, and bounding box inputs during training. Each scribble and click strategy (Fig. 3) can be applied to a ground truth segmentation label $y^{t}$ to simulate positive interactions, or to the background $1 - y^{t}$ to simulate negative interactions. + +![](images/073c31a57efb31e5feb13a8f48c98990c006c7d081dbef55624b903f19eed638.jpg) +Fig. 3: Simulated scribbles and clicks. Positive interactions (green) are simulated on the segmentation label $y^{t}$ (blue), while negative interactions (red) are simulated on the background $1 - y^{t}$ . Scribble thickness is enlarged for visual clarity. + +We simulate positive and negative correction scribbles or clicks by applying the same strategies to the false negative region $\varepsilon_{i}^{t} > 0$ and false positive region $\varepsilon_{i}^{t} < 0$ for error region $\varepsilon_{i}^{t}$ . + +Scribbles. Given binary mask $y \in \{0,1\}^{h \times w}$ , we simulate a scribble mask $s \in [0,1]^{h \times w}$ by first generating clean scribbles, and then corrupting them to account for user behavior and variability. We illustrate this process in Supplementary Material Sec. B.1. To generate the clean scribbles, we start with one of these strategies: + +(i) Line Scribbles: We draw random lines by connecting two end points sampled from $\{(u,v)|y_{uv} = 1\}$ . +(ii) Centerline Scribbles: We simulate scribbles in the center of label $y$ using a thinning algorithm [123] that reduces the label to a 1-pixel wide skeleton. +(iii) Contour Scribbles: We simulate a rough contour of the desired segmentation within the boundaries of the mask. We first blur the mask to reduce the size of the label such that $\tilde{y} = \min(y, y \circ G_k)$ , where $G_k$ is a Gaussian blur kernel. Then we apply a threshold $\tilde{y} < h$ sampled in some intensity range $h \sim U[\tilde{y}_{\text{min}}, \tilde{y}_{\text{max}}]$ and extract a contour inside the boundary of the mask. + +To limit the size and complexity of the centerline and contour scribbles, we apply a random mask that breaks the scribbles into smaller parts. We generate the random mask by sampling a smooth noise image $p$ , where each pixel is sampled independently from $\mathcal{N}(\mu_p, \sigma_p^2)$ , applying Gaussian blur, and then thresholding it at $\mu_p$ . We warp the resulting scribble mask $s$ using a random deformation field $\phi$ to vary the scribble shape and thickness. We ensure the resulting scribble is consistent with mask $y$ , by multiplying the warped scribble mask $s \circ \phi$ by $y$ . + +Clicks. Given a binary mask $y \in \{0,1\}^{h \times w}$ , we simulate $n \sim U[n_{min}, n_{max}]$ clicks at a time using one of three strategies, illustrated in Fig. 3: + +(i) Random clicks: We randomly sample clicks from all pixels in the given region $\{(u,v)|y_{uv} = 1\}$ . +(ii) Center clicks: We sample clicks from the set of points at the center of each disconnected component of the label. First, we create a multi-label mask $m \in \{1, \dots, C\}^{h \times w}$ identifying the $C$ components of label $y$ . We identify the center of each component using the euclidean distance transform [103, 116]. + +(iii) Interior border region clicks: We sample clicks from a border region inside the boundary of the mask. We first blur the mask $y$ to reduce the size of the label such that $\tilde{y} = \min(y, y \circ G_k)$ where $G_k$ is a Gaussian blur kernel. We then sample click coordinates from $\{(u, v) | \tilde{y}_{uv} \in [a, b]\}$ , where $a, b \sim U[\tilde{y}_{\text{min}}, \tilde{y}_{\text{max}})$ are thresholds sampled in some intensity range. + +Since users are inclined to spread out their clicks, we impose a minimum separation of a few pixels between random clicks and border region clicks. + +Bounding Boxes. We compute the minimum bounding box that encloses the label $y^{t}$ , and enlarge each dimension by $r \sim U[0,20]$ pixels to account for human variability. + +Iterative Training. During the first step $(i = 1)$ , we sample the combination of interactions and the number of initial positive and negative interactions $n_{pos}, n_{neg} \sim U[n_{min}, n_{max}]$ . The initial interactions $u_{1}$ are simulated using the ground truth label $y^{t}$ . In subsequent steps, correction scribbles or clicks are sampled from the error region $\varepsilon_{i-1}^{t}$ between the last prediction $\hat{y}_{i-1}^{t}$ and the ground truth $y^{t}$ . Since a user can make multiple corrections in each step, we sample $n_{cor} \sim U[n_{min}, n_{max}]$ corrections (scribbles or clicks) per step. + +# 3.3 Data + +We build on large dataset gathering efforts like MegaMedical [15,90] to compile a collection of 77 open-access biomedical imaging datasets for training and evaluation, covering over 54k scans, 16 image types, and 711 labels. The collection includes a diverse array of biomedical domains, such as eyes [39,55,70,85,105], thorax [86,96,98,99,101], spine [64,98,114,126], cells [16,17,28,29,63,127], skin [20], abdominal [11,12,37,44,45,51,52,54,60,67,71,86,91,101,104], neck [48,79,82,84], brain [3,8,31,38,49,50,73-75,101], bones [34,98,114], teeth [1,65] and lesions [4,101,124,125]. + +We define a 2D segmentation task as a combination of dataset, axis (for 3D modalities), and label. For datasets with multiple segmentation labels, we consider each label separately as a binary segmentation task and for 3D modalities we use the slice with maximum label area and the middle slice from each volume. We provide more details on the data in Supplementary Material Sec. C. + +Task Diversity. During training we sample image, segmentation pairs hierarchically - by dataset and modality, axis, and then label - to balance training on datasets of different sizes. To increase the diversity of segmentation tasks, we apply data augmentation [15] to both the input image and sampled segmentation prior to simulating the user interactions. + +Synthetic Labels. To limit overfitting to specific segmentation tasks, we introduce a mechanism to generate synthetic labels (Fig. 4). During training, for a given sample $(x_0,y_0)$ , with probability $p_{synth}$ , we replace $y$ with a synthetic label $y_{synth}$ . Given an image $x_0$ , we generate a synthetic label $y_{synth}$ by applying a superpixel algorithm [25] with scale parameter $\lambda \sim U[1,\lambda_{max}]$ to partition the image into a multi-label mask of $k$ superpixels $z \in \{1,\dots,k\}^{n\times n}$ , and then randomly select a superpixel $y_{synth} = \mathbb{1}(z = c)$ . We conduct experiments varying $p_{synth}$ and show examples in Supplementary Material Sec. B.3. + +![](images/f9a18ea5e998c44d1f23a8916c05b45ba89e9ccff0d65123a7d44b80b8497cb8.jpg) +Fig. 4: Task sampling and augmentation. During training, we sample an image and segmentation pair $(x_0, y_0)$ . With probability $p_{\text{synth}}$ , we replace $y_0$ with a synthetic label $y_{\text{synth}}$ . We generate $y_{\text{synth}}$ by applying a superpixel algorithm to the image $x_0$ to generate a map $z$ of potential synthetic labels (superpixels), and then sampling one label. Finally, we apply random data augmentations to get $(x^t, y^t)$ . + +# 3.4 Network + +Motivated by producing a practical tool, we primarily demonstrate ScribblePrompt using an efficient fully-convolutional architecture similar to a UNet [92]. We also demonstrate ScribblePrompt using a vision transformer architecture [47]. + +ScribblePrompt-UNet: We use an 8 layer CNN following a decoder-encoder structure similar to the popular UNet architecture [92] without Batch Norm. Each convolutional layer has 192 features and uses PReLU activation [35]. + +ScribblePrompt-SAM: We take the smallest SAM model (ViT-b) [47] and fine-tune the decoder. SAM takes bounding box and click inputs as lists of $(x,y)$ coordinates, and the low-resolution logits of the previous prediction $\hat{y}_{i-1}$ . To adapt SAM for scribbles, we consider each scribbled pixel as a click. + +Loss. We minimize eq. (1) where $\mathcal{L}_{Seg}$ is a linear combination of soft Dice Loss [24] and Focal Loss [58]. In preliminary experiments (Supplementary Material Sec. B.2), training with Soft Dice Loss alone or a linear combination of Soft Dice Loss and Binary-Cross Entropy Loss resulted in slightly lower Dice scores. + +# 4 Experimental Setup + +We compare ScribblePrompt-UNet and ScribblePrompt-SAM to previous methods through experiments with manual scribbles, simulated interactions, and a user study with experienced annotators. Lastly, we report on inference runtime and ablation experiments. + +Data. We use 65 (out of 77) datasets during training. We partition another nine datasets, ACDC [10], BUID [4], BTCV Cervix [52], DRIVE [105], HipXRay [34], PanDental [1], SCD [86], SpineWeb [126], and WBC [127], each into validation (used in model selection, but not training) and test (used only for final evaluation). Three additional datasets, TotalSegmentator [114], SCR [30], and COBRE [3,23,26], were used only for final evaluation. + +We report final results on 12 sets of data (the test splits of the $9 + 3$ evaluation datasets). The evaluation datasets were not seen during training, and cover 608 tasks and 8 modalities, including unseen image types and unseen labels. We + +selected these 12 evaluation datasets to cover a variety of modalities (MRI, CT, ultrasound, fundus photography, microscopy) and anatomical regions of interest (brain, teeth, bones, abdominal organs, muscles, heart, thorax, cells), including both healthy anatomy and lesions. + +Training. During training, we simulated five steps of interactive segmentation per example and set the maximum number of interactions per step to three. We set the minimum number of initial negative prompts to zero, and the minimum initial positive and correction prompts to one. During the first step, we sample the combination of prompt types ( clicks, scribbles, bounding boxes) and then sample the number of positive and negative interactions for each type. In each subsequent steps, we simulate either scribble corrections or click corrections with equal probability. We selected these values based on what we believe to be reasonable interactions for a user to perform. + +Baselines. We compare to existing generalist methods for interactive segmentation, with a focus on methods developed for biomedical imaging. + +SAM [47]: We evaluate the smallest (ViT-b) and largest (ViT-h) versions of the Segment Anything Model (SAM) trained on natural images. SAM takes bounding boxes, clicks, and the logits of the previous prediction as input. + +SAM-Med2D [19]: SAM-Med2D is a SAM ViT-b model with additional adapter layers in the image encoder. SAM-Med2D was fine-tuned, using bounding boxes and iterative clicks as input, on a collection of biomedical imaging datasets containing 4.6M images and 19.7M segmentations. Following [19], we evaluate SAM-Med2D both with and without adapter layers. + +MedSAM [69]: MedSAM is a SAM ViT-B model fine-tuned with bounding box prompts on a collection of biomedical imaging datasets containing 1.5M image segmentation pairs. We evaluate MedSAM with bounding boxes only, because we found it to perform poorly when given point or mask prompts. + +MIDepSeg [68]: MIDepSeg is an interactive segmentation framework designed to generalize to unseen tasks on medical images. MIDepSeg takes interior margin points (positive clicks) as initial inputs, crops the image based on those points, and uses a CNN to make an initial prediction. Given additional clicks, the prediction can be refined using conditional random fields. We evaluate the pre-trained model, which was developed on placenta T2 MRI. + +# 5 Evaluation + +We evaluate all methods using both manual and simulated interactions, with a focus on scribbles. Simulated interactions enable us to test on a large volume of images and tasks. However, simulations do not always match user behavior. We address this by (1) collecting a diverse dataset of manual scribbles for evaluation, and (2) conducting an interactive user study. + +# 5.1 Manual Scribbles + +Setup. We evaluate on two datasets of manual scribbles. First, we collected a dataset (MedScribble) of manual scribbles from three annotators for seven + +segmentation tasks from unseen datasets, with a total of 31 examples. The annotators were shown five training examples with the ground truth segmentation per task and instructed to draw positive and negative scribbles to indicate the region of interest on new images. We provide more details on MedScribble and release the data in Supplementary Material Sec. E.1. Second, we used 380 slices from the ACDC dataset [10], which has scribble annotations for three labels and background. + +In this evaluation, we make a prediction given a set of positive and negative scribbles, simulating a non-iterative scenario where the annotator draws several scribbles before running inference. We evaluate accuracy using Dice score [24] and 95th percentile Hausdorff Distance [42]. Since MIDepSeg's initial prediction network only takes positive inputs, we report results on both MIDepseg's initial prediction from the positive scribbles, and after applying MIDepSeg's refinement procedure using the negative scribbles as corrections. For SAM and variants fine-tuning SAM on clicks, we convert each scribble to a series of points. For MedSAM, we fit a bounding box to the positive scribbles. + +Results. ScribblePrompt-UNet and ScribblePrompt-SAM produce the most accurate segmentations in a single step of manual scribbles on both our manual scribbles dataset and the ACDC scribbles dataset (Tab. 1). We show example predictions in Fig. 5 and additional comparisons to scribble-supervised learning in Supplementary Material Sec. E.3. + +SAM and SAM-Med2D do not generalize well to scribble inputs (which they were not trained for). MedSAM has better predictions than other SAM baselines using the SAM architecture, however it is not able to make use of the negative scribbles and thus often misses segmentations with holes in them (Fig. 5). The initial predictions from MIDepSeg's network are poor, but improve after applying the refinement procedure. + +Table 1: Manual scribbles. Mean Dice and HD95 with $95\%$ CI of segmentations predicted from manually-collected scribbles. Best and second best are highlighted. + +
MedScribble (n=93)ACDC (n=1,140)
Model↑ Dice Score↓ HD95↑ Dice Score↓ HD95
SAM (ViT-b)0.40 ± 0.0532.58 ± 3.800.20 ± 0.01108.39 ± 0.85
SAM (ViT-h)0.56 ± 0.0514.61 ± 3.180.42 ± 0.0259.78 ± 2.02
SAM-Med2D w.o. adapter0.52 ± 0.0530.34 ± 3.290.16 ± 0.01107.25 ± 1.12
SAM-Med2D w/ adapter0.30 ± 0.0619.04 ± 3.630.17 ± 0.0241.01 ± 3.04
MIDepSeg0.68 ± 0.046.94 ± 1.060.58 ± 0.015.78 ± 0.26
MIDepSeg w/ refinement0.81 ± 0.033.10 ± 0.670.73 ± 0.012.71 ± 0.18
MedSAM (box)0.70 ± 0.047.54 ± 1.350.70 ± 0.023.53 ± 0.28
ScribblePrompt-UNet0.84 ± 0.022.92 ± 0.880.84 ± 0.011.80 ± 0.11
ScribblePrompt-SAM0.87 ± 0.022.61 ± 0.840.77 ± 0.014.17 ± 0.32
+ +![](images/3292108fb40a947b0416d1f0645e8d081c5ec96003e686f8297a372cfa1ab0f7.jpg) +Fig. 5: Example predictions. $\mathrm{SP} =$ ScribblePrompt. Top: predictions after one step of manual scribbles. Bottom: predictions after five steps of simulated interactions (one center click followed by one center correction click per step). + +![](images/2fcc093f3f1f6718de6b6abc8a7df8529e46cb0325fd4a8fa07ab84a9a343ae0.jpg) + +![](images/f1e5ae55bbdcac26410ed052ab558e615b3d94cc7dcdac3370823f4c9108ebb8.jpg) + +![](images/ef871048b2e18ae987d1724e68b5795e6b0afca13110736b76f522527d6f33a7.jpg) + +![](images/7c93f90af4aa2ef573d2d05f948e5dc1311c8cf2cb8357acdb0b24b68156111d.jpg) + +![](images/8d22d536c9e4881fcd9efd47dfef9a64ce92f5ad9a957240b4d522d3a265a793.jpg) + +![](images/caa0c9f793687c0de64da82217ca4cf3039054582eb7b0ae7980687e58730dc4.jpg) + +![](images/2b0fdeff4d3a48c74be1b984389f29f44605ad234aa7f51cc193c3080612cd29.jpg) + +# 5.2 Simulated Interactions + +Setup. We use our interaction simulator to evaluate the performance of ScribblePrompt and baselines with iterative scribbles across the 12 evaluation sets. + +For scribble-focused prompting procedures, we use: + +- Line Scribbles: Three positive line scribbles and three negative line scribbles to start, followed by one correction line scribble at each step. +- Centerline Scribbles: One positive and one negative centerline scribble to start, followed by one correction centerline scribble at each step. +- Contour Scribbles: One positive and one negative contour scribble to start, followed by one correction contour scribble at each step. + +We limit each scribble to cover a maximum of $w$ pixels, where $w$ is the maximum dimension of the image. For centerline and contour scribbles, a scribble might contain multiple disconnected components. + +For click-focused prompting procedures, we use: + +- Center Clicks: One positive click in the center of the largest component to start, followed by one (positive or negative) correction click per step in the center of the largest component of the error region. +- Random Clicks: One random positive click to start, followed by one (positive or negative) correction click per step, randomly sampled from the error region. +- Random Warm Start: Three random positive clicks and three random negative clicks to start, followed by one (positive or negative) correction click per step in the center of the largest component of the error region. + +For each example and interaction procedure, we simulate a series of iterative interactions with five random seeds. Since MIDepSeg cannot make accurate predictions with only a few clicks, we exclude it from simulations that start with a single click. For MedSAM, we show results for a bounding box prompt. + +Results. Fig. 6 shows that both versions of ScribblePrompt outperform baseline methods for all simulated interaction procedures at all numbers of interactions. + +![](images/c30bd115037a18ea99ff9074afc0a7ffa858010d9cca4e27e2d0456919419029.jpg) +Fig. 6: Simulated clicks and scribbles. We simulate interactions following three scribble protocols and three click protocols. We show more results and example predictions, with similar trends, in Supplementary Material Sec. F.2. + +![](images/9d2bcb66fd3998c3665e198dbd1bfdba0347eddebf39b489768bc0591ec7325c.jpg) + +![](images/f779e3bebd9b9bf6da2030411ffa0e05b1892f22d6a3fe7f55ee4f785d66ba42.jpg) + +![](images/54ea89099ce6f11cea8c1c5fb30574d71e868d67ad1ea77a2ed082a87ddd6349.jpg) + +We show examples with simulated interactions in Fig. 1 and Fig. 5 (bottom). We show results with similar trends using bounding boxes and by dataset with fully-supervised baselines in Supplementary Material Sec. F. + +Discussion. Comparing the simulated scribble and click results (Fig. 6) highlights the efficiency of scribble-based interactions. Both ScribblePrompt models reach an average Dice above 0.8 in two scribble steps. It takes five steps of carefully-placed center clicks or eight random clicks to reach the same average Dice score with ScribblePrompt-SAM. + +Although SAM-Med2D was trained on a large biomedical imaging collection, it does not perform as well as SAM at refining its predictions on held-out evaluation sets. MedSAM only uses bounding box prompts and is not able to refine predictions, limiting its performance and usability. + +# 5.3 User Study + +To assess the practical utility of ScribblePrompt, we conducted a user study with experienced annotators. We compare ScribblePrompt-UNet to SAM (ViT-b). We selected ScribblePrompt-UNet for the user study because it had similar performance and lower latency compared to ScribblePrompt-SAM. We selected SAM (ViT-b) because it had the highest Dice score in our experiments with clicks and lower latency than the next closest baseline, SAM (ViT-h). + +Setup. Study participants ( $n = 16$ ) were neuroimaging researchers at an academic hospital. They used each model to segment a series of nine test images from different evaluation datasets. For each segmentation task, participants were shown the target segmentation and were asked to prompt the model until the predicted segmentation closely matched the target, or they could no longer improve the prediction. We provided participants with the target segmentation to disentangle the cognitive process of identifying the region of interest from prompting the model to achieve the desired segmentation. We provide additional details and visualizations in Supplementary Material Sec. G. + +Results. Participants produced more accurate segmentations (0.84 vs. 0.73 Dice; $p = 0.001$ using a paired t-test) using ScribblePrompt (Tab. 2). Partici- + +Table 2: User study. Mean ± std. for Dice score, HD95 and total time per segmentation. Median (and max) number of times the participants refreshed the prediction. + +
Dice Score (↑)HD95 (↓)Sec./task (↓)Iter./task (↓)
SAM (ViT-b)0.73 ± 0.3016.1 ± 37.1131.7 ± 168.56 (47)
ScribblePrompt-UNet0.84 ± 0.134.5 ± 9.194.2 ± 89.03 (19)
+ +pants spent $\approx 1.5$ minutes per segmentation on average using ScribblePrompt, compared to over 2 minutes per segmentation with SAM ( $p = 0.02$ using a paired t-test). While using ScribblePrompt, participants updated the prediction fewer times before being satisfied with the segmentation. + +Upon completion, 15 out of 16 participants reported they preferred using ScribblePrompt to SAM and one participant had no preference. All participants reported it was easier to achieve the target segmentation using ScribblePrompt compared to SAM. $93.8\%$ of participants reported ScribblePrompt was better than SAM at refining its predictions in response to scribbles. $87.5\%$ of participants preferred using ScribblePrompt over SAM for clicks. + +Discussion. The most common factors that influenced participant preference for ScribblePrompt were 1) being able get accurate predictions from multiple types of prompts, including scribbles and 2) responsiveness to their corrections, which enabled fine-grained control over the next prediction. For some tasks, such as retinal vein segmentation, SAM was unable to make accurate predictions even with many corrections. As a result, there was more variability in Dice score and time per segmentation, for SAM than ScribblePrompt. + +# 5.4 Inference Runtime + +Setup. We evaluate computational efficiency by measuring inference time on a single CPU for one prediction with a scribble input covering 128 pixels. Performance on a CPU reflects practical utility. Because of a variety of barriers such as protected health information, users may not be able to send their data to a server for computation, and must rely on local, most often CPU-only, computing. + +Results. On a single CPU, ScribblePrompt-UNet requires $0.27 \pm 0.04$ sec per prediction, enabling the model to be used even in low-resource environments. The next most accurate model (Fig. 6), SAM ViT-h, requires over 2 minutes per prediction on a CPU ( $130.79 \pm 7.96$ sec). SAM ViT-b takes $13.59 \pm 0.77$ seconds per prediction. We report results for all baselines and on GPU hardware with similar trends in Supplementary Material Sec. H. + +# 6 Ablations + +We conduct two ablations of important design decisions: (1) synthetic label inputs used during training, and (2) types of prompts simulated during training. We report results on the validation splits of nine datasets unseen during training. + +![](images/12d889d1146f83e3028f91524b33199dfd5b2f5e9b6906932f7b908a2280af33.jpg) +(a) Probability of synthetic labels. + +![](images/2a30fcd60f9694902e6ee937ad81de14cbe0f8544de94ab59c8db3407cade487.jpg) +(b) Interactions during training. +Fig. 7: Ablations. We report Dice after five steps of simulated interactions following three inference-time interaction procedures. Error bars show $95\%$ CI. (a) shows mean change in Dice relative to ScribblePrompt-UNet trained without any synthetic labels. + +# 6.1 Synthetic Labels + +Fig. 7a shows the effect of varying the probability of sampling a synthetic label during training for ScribblePrompt-UNet. Training with both real and synthetic labels improves generalization to new datasets, compared to training with only real labels. Using $p_{\text{synth}} = 0.5$ results in the highest Dice. Training with only synthetic labels results in worse Dice scores. We show similar results for ScribblePrompt-SAM and other interactions in Supplementary Material Sec. I.1. + +# 6.2 Prompt Types + +Setup. We evaluate ScribblePrompt-UNet models trained with different combinations of prompts, compared to the complete ScribblePrompt-UNet: + +- ScribblePrompt-Unet (scribbles) trained on boxes and scribbles. +- ScribblePrompt-UNet (clicks) trained on boxes and clicks. +- ScribblePrompt-Unet (random clicks) trained on boxes and random clicks. + +Results. Fig. 7b shows ScribblePrompt-UNet trained with scribbles, clicks and bounding boxes predicts segmentations more accurately than do ablated versions of ScribblePrompt-UNet. We show results for other interaction procedures with similar trends in Supplementary Material Sec. I.2. + +# 7 Conclusion + +We present ScribblePrompt, a practical framework for interactive segmentation that enables users to segment diverse medical images with scribbles, clicks, and bounding boxes. We introduce methods for simulating realistic user interactions and generating synthetic labels. These methods enable us to train models that generalize to unseen segmentation tasks and datasets. ScribblePrompt is more accurate than existing baselines, and ScribblePrompt-UNet is computationally efficient, even on a CPU. Our user study shows that nearly all users prefer ScribblePrompt and achieve segmentations with $15\%$ higher Dice with less effort than the next most accurate baseline. ScribblePrompt promises to significantly reduce the burden of manual segmentation in biomedical imaging. + +# Acknowledgements + +This work was supported in part by funding from Quanta Computer Inc., the Eric and Wendy Schmidt Center at the Broad Institute of MIT and Harvard and the Wistron Corporation. Research reported in this paper was supported by the National Institute of Biomedical Imaging and Bioengineering of the National Institutes of Health under award number R01EB033773. Much of the computation resources required for this research was performed on computational hardware generously provided by the Massachusetts Life Sciences Center. + +# References + +1. Abdi, A.H., Kasaei, S., Mehdizadeh, M.: Automatic segmentation of mandible in panoramic x-ray. Journal of Medical Imaging 2(4), 044003 (2015) +2. Agustsson, E., Uijlings, J.R., Ferrari, V.: Interactive Full Image Segmentation by Considering All Regions Jointly. In: 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11614-11623. IEEE, Long Beach, CA, USA (Jun 2019). https://doi.org/10.1109/CVPR.2019.01189, https://ieeexplore.ieee.org/document/8953526/ +3. Aine, C.J., Bockholt, H.J., Bustillo, J.R., Canive, J.M., Caprihan, A., Gasparovic, C., Hanlon, F.M., Houck, J.M., Jung, R.E., Lauriello, J., Liu, J., Mayer, A.R., Perrone-Bizzozero, N.I., Posse, S., Stephen, J.M., Turner, J.A., Clark, V.P., Calhoun, V.D.: Multimodal Neuroimaging in Schizophrenia: Description and Dissemination. Neuroinformatics 15(4), 343-364 (Oct 2017). https://doi.org/10.1007/s12021-017-9338-9, http://link.springer.com/10.1007/s12021-017-9338-9 +4. Al-Dhabyani, W., Gomaa, M., Khaled, H., Fahmy, A.: Dataset of breast ultrasound images. Data in Brief 28, 104863 (2020). https://doi.org/https://doi.org/10.1016/j.dib.2019.104863, https://www.sciencedirect.com/science/article/pii/S2352340919312181 +5. Asad, M., Fidon, L., Vercauteren, T.: ECONet: Efficient Convolutional Online Likelihood Network for Scribble-based Interactive Segmentation. In: International Conference on Medical Imaging with Deep Learning. pp. 35-47. PMLR (2022), http://arxiv.org/abs/2201.04584, arXiv:2201.04584 [cs, eess] +6. Atzeni, A., Peter, L., Robinson, E., Blackburn, E., Althonayan, J., Alexander, D.C., Iglesias, J.E.: Deep active learning for suggestive segmentation of biomedical image stacks via optimisation of Dice scores and traced boundary length. Medical Image Analysis 81 (2022) +7. Bai, J., Wu, X.: Error-Tolerant Scribbles Based Interactive Image Segmentation. In: 2014 IEEE Conference on Computer Vision and Pattern Recognition. pp. 392-399. IEEE, Columbus, OH, USA (Jun 2014). https://doi.org/10.1109/CVPR.2014.57, https://ieeexplore.ieee.org/document/6909451 +8. Baid, U., Ghodasara, S., Mohan, S., Bilello, M., Calabrese, E., Colak, E., Farahani, K., Kalpathy-Cramer, J., Kitamura, F.C., Pati, S., et al.: The rsna-asnrmiccai brats 2021 benchmark on brain tumor segmentation and radiogenomic classification. arXiv preprint arXiv:2107.02314 (2021) +9. Benenson, R., Popov, S., Ferrari, V.: Large-scale interactive object segmentation with human annotators. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11700-11709 (2019) + +10. Bernard, O., Lalande, A., Zotti, C., Cervenansky, F., Yang, X., Heng, P.A., Cetin, I., Lekadir, K., Camara, O., Ballester, M.A.G., et al.: Deep learning techniques for automatic mri cardiac multi-structures segmentation and diagnosis: is the problem solved? IEEE transactions on medical imaging 37(11), 2514-2525 (2018) +11. Bilic, P., Christ, P.F., Vorontsov, E., Chlebus, G., Chen, H., Dou, Q., Fu, C.W., Han, X., Heng, P.A., Hesser, J., et al.: The liver tumor segmentation benchmark (lits). arXiv preprint arXiv:1901.04056 (2019) +12. Bloch, N., Madabhushi, A., Huisman, H., Freymann, J., Kirby, J., Grauer, M., Enquobahrie, A., Jaffe, C., Clarke, L., Farahani, K.: Nci-isbi 2013 challenge: automated segmentation of prostate structures. The Cancer Imaging Archive 370(6), 5 (2015) +13. Boykov, Y., Jolly, M.P.: Interactive graph cuts for optimal boundary & region segmentation of objects in N-D images. In: Proceedings Eighth IEEE International Conference on Computer Vision. ICCV 2001. vol. 1, pp. 105-112. IEEE Comput. Soc, Vancouver, BC, Canada (2001). https://doi.org/10.1109/ICCV.2001.937505, http://ieeexplore.ieee.org/document/937505/ +14. Bredell, G., Tanner, C., Konukoglu, E.: Iterative interaction training for segmentation editing networks. In: Machine Learning in Medical Imaging: 9th International Workshop, MLMI 2018, Held in Conjunction with MICCAI 2018, Granada, Spain, September 16, 2018, Proceedings 9. pp. 363-370. Springer (2018) +15. Butoi*, V.I., Ortiz*, J.J.G., Ma, T., Sabuncu, M.R., Guttag, J., Dalca, A.V.: Universeg: Universal medical image segmentation. In: ICCV (2023) +16. Caicedo, J.C., Goodman, A., Karhohs, K.W., Cimini, B.A., Ackerman, J., Haghighi, M., Heng, C., Becker, T., Doan, M., McQuin, C., Rohhan, M., Singh, S., Carpenter, A.E.: Nucleus segmentation across imaging experiments: the 2018 Data Science Bowl. Nature Methods 16(12), 1247-1253 (Dec 2019). https://doi.org/10.1038/s41592-019-0612-7, https://doi.org/10.1038/s41592-019-0612-7 +17. Cardona, A., Saalfeld, S., Preibisch, S., Schmid, B., Cheng, A., Pulokas, J., Tomancak, P., Hartenstein, V.: An integrated micro-and macroarchitectural analysis of the drosophila brain by computer-assisted serial section electron microscopy. PLoS biology 8(10), e1000502 (2010) +18. Chen, X., Cheung, Y.S.J., Lim, S.N., Zhao, H.: ScribbleSeg: Scribble-based Interactive Image Segmentation (Mar 2023), http://arxiv.org/abs/2303.11320, arXiv:2303.11320 [cs] +19. Cheng, J., Ye, J., Deng, Z., Chen, J., Li, T., Wang, H., Su, Y., Huang, Z., Chen, J., Jiang, L., Sun, H., He, J., Zhang, S., Zhu, M., Qiao, Y.: SAM-Med2D (Aug 2023), http://arxiv.org/abs/2308.16184, arXiv:2308.16184 [cs] +20. Codella, N.C.F., Gutman, D.A., Celebi, M.E., Helba, B., Marchetti, M.A., Dusza, S.W., Kalloo, A., Liopyris, K., Mishra, N.K., Kittler, H., Halpern, A.: Skin lesion analysis toward melanoma detection: A challenge at the 2017 international symposium on biomedical imaging (isbi), hosted by the international skin imaging collaboration (ISIC). CoRR abs/1710.05006 (2017), http://arxiv.org/abs/1710.05006 +21. Criminisi, A., Sharp, T., Blake, A.: GeoS: Geodesic Image Segmentation. In: Forsyth, D., Torr, P., Zisserman, A. (eds.) Computer Vision - ECCV 2008, vol. 5302, pp. 99-112. Springer Berlin Heidelberg, Berlin, Heidelberg (2008). https://doi.org/10.1007/978-3-540-88682-2_9, http://link.springer.com/10.1007/978-3-540-88682-2_9, series Title: Lecture Notes in Computer Science + +22. Czolbe, S., Dalca, A.V.: Neuralizer: General neuroimage analysis without retraining. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6217-6230 (2023) +23. Dalca, A.V., Guttag, J., Sabuncu, M.R.: Anatomical priors in convolutional networks for unsupervised biomedical segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 9290-9299 (2018) +24. Dice, L.R.: Measures of the amount of ecologic association between species. Ecology 26(3), 297-302 (1945) +25. Felzenszwalb, P.F., Huttenlocher, D.P.: Efficient graph-based image segmentation. International journal of computer vision 59, 167-181 (2004) +26. Fischl, B.: Freesurfer. Neuroimage 62(2), 774-781 (2012) +27. Forte, M., Price, B., Cohen, S., Xu, N., Pitie, F.: Getting to $99\%$ Accuracy in Interactive Segmentation (Mar 2020), http://arxiv.org/abs/2003.07932, arXiv:2003.07932 [cs] +28. Gamper, J., Koohbanani, N., Benes, K., Graham, S., Jahanifar, M., Khurram, S., Azam, A., Hewitt, K., Rajpoot, N.: Pannuke dataset extension, insights and baselines. arxiv. 2020 doi: 10.48550. ARXIV (2003) +29. Gerhard, S., Funke, J., Martel, J., Cardona, A., Fetter, R.: Segmented anisotropic ssTEM dataset of neural tissue (11 2013). https://doi.org/10.6084/m9.figshare.856713.v1, https://figshare.com/articles/dataset/Segmented_anisotropic_ssTEM_dataset_of_neural_tissue/856713 +30. van Ginneken, B., Stegmann, M.B., Loog, M.: Segmentation of anatomical structures in chest radiographs using supervised methods: a comparative study on a public database. Medical Image Analysis 10(1), 19-40 (2006). https://doi.org/https://doi.org/10.1016/j.media.2005.02.002, https://www.sciencedirect.com/science/article/pii/S1361841505000368 +31. Gollub, R.L., Shoemaker, J.M., King, M.D., White, T., Ehrlich, S., Sponheim, S.R., Clark, V.P., Turner, J.A., Mueller, B.A., Magnotta, V., et al.: The mcic collection: a shared repository of multi-modal, multi-site brain image data from a clinical investigation of schizophrenia. Neuroinformatics 11, 367-388 (2013) +32. Gotkowski, K., Lüth, C., Jäger, P.F., Ziegler, S., Krämer, L., Denner, S., Xiao, S., Disch, N., Maier-Hein, K.H., Isensee, F.: Embarrassingly simple scribble supervision for 3d medical segmentation. arXiv preprint arXiv:2403.12834 (2024) +33. Grady, L.: Random Walks for Image Segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence 28(11), 1768-1783 (Nov 2006). https://doi.org/10.1109/TPAMI.2006.233, http://ieeexplore.ieee.org/document/1704833/ +34. Gut, D.: X-ray images of the hip joints 1 (Jul 2021). https://doi.org/10.17632/zm6bxzhmfz.1, https://data.mendeley.com/datasets/zm6bxzhmfz/1, publisher: Mendeley Data +35. He, K., Zhang, X., Ren, S., Sun, J.: Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In: Proceedings of the IEEE international conference on computer vision. pp. 1026-1034 (2015) +36. He, S., Bao, R., Li, J., Stout, J., Bjornerud, A., Grant, P.E., Ou, Y.: Computer- vision benchmark segment-anything model (sam) in medical images: Accuracy in 12 datasets (2023) +37. Heller, N., Isensee, F., Maier-Hein, K.H., Hou, X., Xie, C., Li, F., Nan, Y., Mu, G., Lin, Z., Han, M., et al.: The state of the art in kidney and kidney tumor segmentation in contrast-enhanced ct imaging: Results of the kits19 challenge. Medical Image Analysis p. 101821 (2020) + +38. Hernandez Petzsche, M.R., de la Rosa, E., Hanning, U., Wiest, R., Valenzuela, W., Reyes, M., Meyer, M., Liew, S.L., Kofler, F., Ezhov, I., et al.: Isles 2022: A multi-center magnetic resonance imaging stroke lesion segmentation dataset. Scientific data 9(1), 762 (2022) +39. Hoover, A., Kouznetsova, V., Goldbaum, M.: Locating blood vessels in retinal images by piecewise threshold probing of a matched filter response. IEEE Transactions on Medical imaging 19(3), 203-210 (2000) +40. Hu, X., Xu, X., Shi, Y.: How to efficiently adapt large segmentation model(sam) to medical images (2023) +41. Huang, Y., Yang, X., Liu, L., Zhou, H., Chang, A., Zhou, X., Chen, R., Yu, J., Chen, J., Chen, C., Chi, H., Hu, X., Fan, D.P., Dong, F., Ni, D.: Segment anything model for medical images? (2023) +42. Huttenlocher, D.P., Klanderman, G.A., Rucklidge, W.J.: Comparing images using the hausdorff distance. IEEE Transactions on pattern analysis and machine intelligence 15(9), 850-863 (1993) +43. Isensee, F., Jaeger, P.F., Kohl, S.A.A., Petersen, J., Maier-Hein, K.H.: nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nature Methods 18(2), 203-211 (Feb 2021). https://doi.org/10.1038/s41592-020-01008-z, http://www.nature.com/articles/s41592-020-01008-z +44. Ji, Y., Bai, H., Yang, J., Ge, C., Zhu, Y., Zhang, R., Li, Z., Zhang, L., Ma, W., Wan, X., et al.: Amos: A large-scale abdominal multi-organ benchmark for versatile medical image segmentation. arXiv preprint arXiv:2206.08023 (2022) +45. Kavur, A.E., Gezer, N.S., Baris, M., Aslan, S., Conze, P.H., Groza, V., Pham, D.D., Chatterjee, S., Ernst, P., Ozkan, S., Baydar, B., Lachinov, D., Han, S., Pauli, J., Isensee, F., Perkonigg, M., Sathish, R., Rajan, R., Sheet, D., Dovletov, G., Speck, O., Nurnberger, A., Maier-Hein, K.H., Bozdagi Akar, G., Ünal, G., Dicle, O., Selver, M.A.: CHAOS Challenge - combined (CT-MR) healthy abdominal organ segmentation. Medical Image Analysis 69, 101950 (2021). https://doi.org/https://doi.org/10.1016/j.media.2020.101950, https://www.sciencedirect.com/science/article/pii/S1361841520303145 +46. Kim, S., Oh, H.J., Min, S., Jeong, W.K.: Evaluation and improvement of segment anything model for interactive histopathology image segmentation (2023) +47. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. In: ICCV (2023) +48. Krönke, M., Eilers, C., Dimova, D., Köhler, M., Buschner, G., Schweiger, L., Konstantinidou, L., Makowski, M., Nagarajah, J., Navab, N., et al.: Tracked 3d ultrasound and deep neural network-based thyroid segmentation reduce interobserver variability in thyroid volumetry. Plos one 17(7), e0268550 (2022) +49. Kuijf, H.J., Biesbroek, J.M., De Bresser, J., Heinen, R., Andermatt, S., Bento, M., Berseth, M., Belyaev, M., Cardoso, M.J., Casamitjana, A., et al.: Standardized assessment of automatic segmentation of white matter hyperintensities and results of the wmh segmentation challenge. IEEE transactions on medical imaging 38(11), 2556-2568 (2019) +50. Kuklisova-Murgasova, M., Aljabar, P., Srinivasan, L., Counsell, S.J., Doria, V., Serag, A., Gousias, I.S., Boardman, J.P., Rutherford, M.A., Edwards, A.D., et al.: A dynamic 4d probabilistic atlas of the developing brain. NeuroImage 54(4), 2750-2763 (2011) +51. Lambert, Z., Petitjean, C., Dubray, B., Kuan, S.: Segthor: segmentation of thoracic organs at risk in ct images. In: 2020 Tenth International Conference on Image Processing Theory, Tools and Applications (IPTA). pp. 1-6. IEEE (2020) + +52. Landman, B., Xu, Z., Igelsias, J., Styner, M., Langerak, T., Klein, A.: Miccai multi-atlas labeling beyond the cranial vault-workshop and challenge. In: Proc. MICCAI Multi-Atlas Labeling Beyond Cranial Vault Workshop Challenge. vol. 5, p. 12 (2015) +53. Lee, H., Jeong, W.K.: Scribble2label: Scribble-supervised cell segmentation via self-generating pseudo-labels with consistency. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part I 23. pp. 14-23. Springer (2020) +54. Lemaître, G., Martí, R., Freixenet, J., Vilanova, J.C., Walker, P.M., Meriaudeau, F.: Computer-aided detection and diagnosis for prostate cancer based on mono and multi-parametric mri: a review. Computers in biology and medicine 60, 8-31 (2015) +55. Li, M., Zhang, Y., Ji, Z., Xie, K., Yuan, S., Liu, Q., Chen, Q.: Ipn-v2 and octa-500: Methodology and dataset for retinal image segmentation. arXiv preprint arXiv:2012.07261 (2020) +56. Li, Z., Zheng, Y., Luo, X., Shan, D., Hong, Q.: Scribblevc: Scribble-supervised medical image segmentation with vision-class embedding. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 3384-3393 (2023) +57. Lin, D., Dai, J., Jia, J., He, K., Sun, J.: Scribblesup: Scribble-supervised convolutional networks for semantic segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3159-3167 (2016) +58. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proceedings of the IEEE international conference on computer vision. pp. 2980-2988 (2017) +59. Lin, X., Xiang, Y., Zhang, L., Yang, X., Yan, Z., Yu, L.: Samus: Adapting segment anything model for clinically-friendly and generalizable ultrasound image segmentation (2023) +60. Litjens, G., Toth, R., van de Ven, W., Hoeks, C., Kerkstra, S., van Ginneken, B., Vincent, G., Guillard, G., Birbeck, N., Zhang, J., et al.: Evaluation of prostate segmentation algorithms for mri: the promise12 challenge. Medical image analysis 18(2), 359-373 (2014) +61. Liu, Q., Xu, Z., Bertasius, G., Niethammer, M.: Simpleclick: Interactive image segmentation with simple vision transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22290-22300 (2023) +62. Liu, Q., Xu, Z., Jiao, Y., Niethammer, M.: iSegFormer: Interactive Segmentation via Transformers with Application to 3D Knee MR Images. In: Medical Image Computing and Computer Assisted Intervention (MICCAI) (2022) +63. Ljosa, V., Sokolnicki, K.L., Carpenter, A.E.: Annotated high-throughput microscopy image sets for validation. Nature methods 9(7), 637-637 (2012) +64. Löffler, M.T., Sekuboyina, A., Jacob, A., Grau, A.L., Scharr, A., El Husseinini, M., Kallweit, M., Zimmer, C., Baum, T., Kirschke, J.S.: A vertebral segmentation dataset with fracture grading. Radiology: Artificial Intelligence 2(4), e190138 (2020) +65. in the Loop, H.: Teeth segmentation dataset, https://humansintheloop.org/resources/datasets/teeth-segmentation-dataset/ +66. Luo, X., Hu, M., Liao, W., Zhai, S., Song, T., Wang, G., Zhang, S.: Scribblesupervised medical image segmentation via dual-branch network and dynamically mixed pseudo labels supervision. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 528-538. Springer (2022) + +67. Luo, X., Liao, W., Xiao, J., Song, T., Zhang, X., Li, K., Wang, G., Zhang, S.: Word: Revisiting organs segmentation in the whole abdominal region. arXiv preprint arXiv:2111.02403 (2021) +68. Luo, X., Wang, G., Song, T., Zhang, J., Aertsen, M., Deprest, J., Ourselin, S., Vercauteren, T., Zhang, S.: MIDeepSeg: Minimally Interactive Segmentation of Unseen Objects from Medical Images Using Deep Learning. Medical Image Analysis 72, 102102 (2021) +69. Ma, J., He, Y., Li, F., Han, L., You, C., Wang, B.: Segment anything in medical images. Nature Communications 15, 1-9 (2024) +70. Ma, Y., Hao, H., Xie, J., Fu, H., Zhang, J., Yang, J., Wang, Z., Liu, J., Zheng, Y., Zhao, Y.: Rose: a retinal oct-angiography vessel segmentation dataset and new model. IEEE Transactions on Medical Imaging 40(3), 928–939 (2021). https://doi.org/10.1109/TMI.2020.3042802 +71. Macdonald, J.A., Zhu, Z., Konkel, B., Mazurowski, M., Wiggins, W., Bashir, M.: Duke liver dataset (MRI) v2 (Apr 2023). https://doi.org/10.5281/zenodo.7774566, https://doi.org/10.5281/zenodo.7774566 +72. Maninis, K.K., Caelles, S., Pont-Tuset, J., Van Gool, L.: Deep Extreme Cut: From Extreme Points to Object Segmentation (2018), arXiv:1711.09081 +73. Marcus, D.S., Wang, T.H., Parker, J., Csernansky, J.G., Morris, J.C., Buckner, R.L.: Open access series of imaging studies (oasis): cross-sectional mri data in young, middle aged, nondemented, and demented older adults. Journal of cognitive neuroscience 19(9), 1498-1507 (2007) +74. Marek, K., Jennings, D., Lasch, S., Siderowf, A., Tanner, C., Simuni, T., Coffey, C., Kieburtz, K., Flagg, E., Chowdhury, S., et al.: The parkinson progression marker initiative (ppmi). Progress in neurobiology 95(4), 629-635 (2011) +75. Mazurowski, M.A., Clark, K., Czarnek, N.M., Shamsesfandabadi, P., Peters, K.B., Saha, A.: Radiogenomics of lower-grade glioma: algorithmically-assessed tumor shape is associated with tumor genomic subtypes and patient outcomes in a multi-institutional study with the cancer genome atlas data. Journal of neuro-oncology 133, 27-35 (2017) +76. Mazurowski, M.A., Dong, H., Gu, H., Yang, J., Konz, N., Zhang, Y.: Segment anything model for medical image analysis: An experimental study. Medical Image Analysis 89, 102918 (Oct 2023). https://doi.org/10.1016/j.media.2023.102918, https://www.sciencedirect.com/science/article/pii/S1361841523001780 +77. Men, J., Huang, Y., Solanki, J., Zeng, X., Alex, A., Jerwick, J., Zhang, Z., Tanzi, R.E., Li, A., Zhou, C.: Optical coherence tomography for brain imaging and developmental biology. IEEE Journal of Selected Topics in Quantum Electronics 22(4), 120-132 (2015) +78. Milletari, F., Navab, N., Ahmadi, S.A.: V-net: Fully convolutional neural networks for volumetric medical image segmentation. Fourth International Conference on 3D Vision (3DV) pp. 565-571 (2016) +79. Montoya, A., Hasnin, kaggle446, shirzad, Cukierski, W., yffud: Ultrasound nerve segmentation (2016), https://kaggle.com/competitions/ultrasound-nerve-segmentation +80. Ooi, A.Z.H., Embong, Z., Abd Hamid, A.I., Zainon, R., Wang, S.L., Ng, T.F., Hamzah, R.A., Teoh, S.S., Ibrahim, H.: Interactive Blood Vessel Segmentation from Retinal Fundus Image Based on Canny Edge Detector. Sensors 21(19), 6380 (Sep 2021) + +81. Paranjape, J.N., Nair, N.G., Sikder, S., Vedula, S.S., Patel, V.M.: Adaptivesam: Towards efficient tuning of sam for surgical scene segmentation. arXiv preprint arXiv:2308.03726 (2023) +82. Pedraza, L., Vargas, C., Narvaez, F., Durán, O., Muñoz, E., Romero, E.: An open access thyroid ultrasound image database. In: Romero, E., Lepore, N. (eds.) 10th international symposium on medical information processing and analysis. vol. 9287, p. 92870W. SPIE / International Society for Optics and Photonics (2015). https://doi.org/10.1117/12.2073532, https://doi.org/10.1117/12.2073532 +83. Philbrick, K.A., Weston, A.D., Akkus, Z., Kline, T.L., Korfiatis, P., Sakinis, T., Kostandy, P., Boonrod, A., Zeinoddini, A., Takahashi, N., et al.: Ril-contour: a medical imaging dataset annotation tool for and with deep learning. Journal of digital imaging 32, 571-581 (2019) +84. Podobnik, G., Strojan, P., Peterlin, P., Ibragimov, B., Vrtovec, T.: HaN-Seg: The head and neck organ-at-risk CT and MR segmentation dataset. Medical Physics 50(3), 1917-1927 (2023). https://doi.org/https://doi.org/10.1002/mp.16197, https://aapm.onlinelibrary.wiley.com/doi/abs/10.1002/mp.16197, tex.eprint: https://aapm.onlinelibrary.wiley.com/doi/pdf/10.1002/mp.16197 +85. Porwal, P., Pachade, S., Kamble, R., Kokare, M., Deshmukh, G., Sahasrabuddhe, V., Meriaudeau, F.: Indian diabetic retinopathy image dataset (idrid) (2018). https://doi.org/10.21227/H25W98, https://dx.doi.org/10.21227/H25W98 +86. Radau, P., Lu, Y., Connelly, K., Paul, G., Dick, A., Wright, G.: Evaluation framework for algorithms segmenting short axis cardiac mri. The MIDAS Journal-Cardiac MR Left Ventricle Segmentation Challenge 49 (2009) +87. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) +88. Raghu, M., Zhang, C., Kleinberg, J., Bengio, S.: Transfusion: Understanding transfer learning for medical imaging. Advances in neural information processing systems 32 (2019) +89. Rajchl, M., Lee, M.C.H., Oktay, O., Kamnitsas, K., Passerat-Palmbach, J., Bai, W., Damodaram, M., Rutherford, M.A., Hajnal, J.V., Kainz, B., Rueckert, D.: DeepCut: Object Segmentation From Bounding Box Annotations Using Convolutional Neural Networks. IEEE Transactions on Medical Imaging 36(2), 674-683 (2017) +90. Rakic, M., Wong, H.E., Ortiz, J.J.G., Cimini, B., Guttag, J.V., Dalca, A.V.: Tyche: Stochastic in-context learning for medical image segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2024) +91. Rister, B., Yi, D., Shivakumar, K., Nobashi, T., Rubin, D.L.: CT-ORG, a new dataset for multiple organ segmentation in computed tomography. Scientific Data 7(1), 381 (Nov 2020). https://doi.org/10.1038/s41597-020-00715-8, https://www.nature.com/articles/s41597-020-00715-8 +92. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015) +93. Roth, H.R., Yang, D., Xu, Z., Wang, X., Xu, D.: Going to Extremes: Weakly Supervised Medical Image Segmentation. Machine Learning and Knowledge Ex + +traction 3(2), 507-524 (Jun 2021). https://doi.org/10.3390/make3020026, https://www.mdpi.com/2504-4990/3/2/26 +94. Rother, C., Kolmogorov, V., Blake, A.: "GrabCut" — Interactive Foreground Extraction using Iterated Graph Cuts. ACM Transactions on Graphics 23, 309-314 (2004) +95. Sakinis, T., Miletari, F., Roth, H., Korfiatis, P., Kostandy, P., Philbrick, K., Akkus, Z., Xu, Z., Xu, D., Erickson, B.J.: Interactive segmentation of medical images through fully convolutional neural networks (2019), arXiv:1903.08205 +96. Saporta, A., Gui, X., Agrawal, A., Pareek, A., Truong, S., Nguyen, C., Ngo, V.D., Seekins, J., Blankenberg, F.G., Ng, A., et al.: Deep learning saliency maps do not accurately highlight diagnostically relevant regions for medical image interpretation. MedRxiv (2021) +97. Sati, P., George, I.C., Shea, C.D., Gaitan, M.I., Reich, D.S.: Flair*: a combined mr contrast technique for visualizing white matter lesions and parenchymal veins. Radiology 265(3), 926-932 (2012) +98. Seibold, C., Reif, S., Sarfraz, S., Fink, M.A., Mayer, V., Sellner, J., Kim, M.S., Maier-Hein, K.H., Kleesiek, J., Stiefelhagen, R.: Detailed annotations of chest x-rays via ct projection for report understanding. In: Proceedings of the 33th British Machine Vision Conference (BMVC) (2022) +99. Setio, A.A.A., Traverso, A., De Bel, T., Berens, M.S., Van Den Bogaard, C., Cerello, P., Chen, H., Dou, Q., Fantacci, M.E., Geurts, B., et al.: Validation, comparison, and combination of algorithms for automatic detection of pulmonary nodules in computed tomography images: the luna16 challenge. Medical image analysis 42, 1-13 (2017) +100. Shi, P., Qiu, J., Abaxi, S.M.D., Wei, H., Lo, F.P.W., Yuan, W.: Generalist vision foundation models for medical imaging: A case study of segment anything model on zero-shot medical segmentation. Diagnostics 13(11), 1947 (2023) +101. Simpson, A.L., Antonelli, M., Bakas, S., Bilello, M., Farahani, K., Van Ginneken, B., Kopp-Schneider, A., Landman, B.A., Litjens, G., Menze, B., et al.: A large annotated medical image dataset for the development and evaluation of segmentation algorithms. arXiv preprint arXiv:1902.09063 (2019) +102. Sofiuk, K., Petrov, I., Barinova, O., Konushin, A.: F-BRS: Rethinking Backpropagating Refinement for Interactive Segmentation. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8620-8629. IEEE, Seattle, WA, USA (Jun 2020). https://doi.org/10.1109/CVPR42600.2020.00865, https://ieeexplore.ieee.org/document/9156403/ +103. Sofiuk, K., Petrov, I.A., Konushin, A.: Reviving Iterative Training with Mask Guidance for Interactive Segmentation (Feb 2021), http://arxiv.org/abs/2102.06583, arXiv:2102.06583 [cs] +104. Song, Y., Zheng, J., Lei, L., Ni, Z., Zhao, B., Hu, Y.: CT2US: Cross-modal transfer learning for kidney segmentation in ultrasound images with synthesized data. Ultrasonics 122, 106706 (2022). https://doi.org/https://doi.org/10.1016/j.ultras.2022.106706, https://www.sciencedirect.com/science/article/pii/S0041624X22000191 +105. Staal, J., Abramoff, M.D., Niemeijer, M., Viergever, M.A., Van Ginneken, B.: Ridge-based vessel segmentation in color images of the retina. IEEE transactions on medical imaging 23(4), 501-509 (2004) +106. Tang, Y., Yang, D., Li, W., Roth, H.R., Landman, B., Xu, D., Nath, V., Hatamizadeh, A.: Self-Supervised Pre-Training of Swin Transformers for 3D Medical Image Analysis. In: 2022 IEEE/CVF Conference on Computer Vision + +and Pattern Recognition (CVPR). pp. 20698-20708. IEEE, New Orleans, LA, USA (Jun 2022). https://doi.org/10.1109/CVPR52688.2022.02007, https://ieeexplore.ieee.org/document/9879123/ +107. Vezhnevets, V., Konouchine, V.: "GrowCut" - Interactive Multi-Label N-D Image Segmentation By Cellular Automata. proc. of Graphicon 1(4), 150-156 (2005) +108. Wang, C., Chen, X., Ning, H., Li, S.: Sam-octa: A fine-tuning strategy for applying foundation model to octa image segmentation tasks (2023) +109. Wang, G., Li, W., Zuluaga, M.A., Pratt, R., Patel, P.A., Aertsen, M., Doel, T., David, A.L., Deprest, J., Ourselin, S., Vercauteren, T.: Interactive Medical Image Segmentation Using Deep Learning With Image-Specific Fine Tuning. IEEE Transactions on Medical Imaging 37(7), 1562-1573 (2018) +110. Wang, G., Zuluaga, M.A., Li, W., Pratt, R., Patel, P.A., Aertsen, M., Doel, T., David, A.L., Deprest, J., Ourselin, S., Vercauteren, T.: DeepIGeoS: A Deep Interactive Geodesic Framework for Medical Image Segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence 41(7), 1559-1572 (2019) +111. Wang, G., Zuluaga, M.A., Pratt, R., Aertsen, M., Doel, T., Klusmann, M., David, A.L., Deprest, J., Vercauteren, T., Ourselin, S.: Slic-Seg: A minimally interactive segmentation of the placenta from sparse and motion-corrupted fetal MRI in multiple views. Medical Image Analysis 34, 137-147 (Dec 2016). https://doi.org/10.1016/j.media.2016.04.009 +112. Wang, X., Wang, W., Cao, Y., Shen, C., Huang, T.: Images speak in images: A generalist painter for in-context visual learning. arXiv preprint arXiv:2212.02499 (2022) +113. Wang, X., Zhang, X., Cao, Y., Wang, W., Shen, C., Huang, T.: SegGPT: Segmenting Everything In Context (Apr 2023), http://arxiv.org/abs/2304.03284, arXiv:2304.03284 [cs] +114. Wasserthal, J., Breit, H.C., Meyer, M.T., Pradella, M., Hinck, D., Sauter, A.W., Heye, T., Boll, D.T., Cyriac, J., Yang, S., et al.: Totalsegmentator: Robust segmentation of 104 anatomic structures in ct images. Radiology: Artificial Intelligence 5(5) (2023) +115. Wu, J., Fu, R., Fang, H., Liu, Y., Wang, Z., Xu, Y., Jin, Y., Arbel, T.: Medical SAM Adapter: Adapting Segment Anything Model for Medical Image Segmentation (Apr 2023), http://arxiv.org/abs/2304.12620, arXiv:2304.12620 [cs] +116. Xu, N., Price, B., Cohen, S., Yang, J., Huang, T.: Deep Interactive Object Selection. In: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 373-381. IEEE, Las Vegas, NV, USA (Jun 2016). https://doi.org/10.1109/CVPR.2016.47, http://ieeexplore.ieee.org/document/7780416/ +117. Xu, N., Price, B., Cohen, S., Yang, J., Huang, T.: Deep GrabCut for Object Selection. arXiv (2017), arXiv:1707.00243 +118. Yushkevich, P.A., Piven, J., Hazlett, H.C., Smith, R.G., Ho, S., Gee, J.C., Gering, G.: User-guided 3d active contour segmentation of anatomical structures: significantly improved efficiency and reliability. Neuroimage 31(3), 1116-1128 (2006) +119. Zhang, J., Ding, X., Hu, D., Jiang, Y.: Semantic segmentation of Covid-19 lesions with a multiscale dilated convolutional network. Scientific Reports 12(1), 1847 (2022) +120. Zhang, K., Liu, D.: Customized segment anything model for medical image segmentation (2023) +121. Zhang, K., Zhuang, X.: Cycle mix: A holistic strategy for medical image segmentation from scribble supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11656-11665 (2022) + +122. Zhang, S., Liew, J.H., Wei, Y., Wei, S., Zhao, Y.: Interactive Object Segmentation With Inside-Outside Guidance. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 12231-12241. IEEE (2020) +123. Zhang, T.Y., Suen, C.Y.: A fast parallel algorithm for thinning digital patterns. Commun. ACM 27(3), 236-239 (mar 1984). https://doi.org/10.1145/357994.358023, https://doi.org/10.1145/357994.358023 +124. Zhang, Y., Xian, M., Cheng, H.D., Shareef, B., Ding, J., Xu, F., Huang, K., Zhang, B., Ning, C., Wang, Y.: Busis: A benchmark for breast ultrasound image segmentation. In: Healthcare. vol. 10, p. 729. MDPI (2022) +125. Zhao, Q., Lyu, S., Bai, W., Cai, L., Liu, B., Wu, M., Sang, X., Yang, M., Chen, L.: A multi-modality ovarian tumor ultrasound image dataset for unsupervised cross-domain semantic segmentation. CoRR abs/2207.06799 (2022) +126. Zheng, G., Chu, C., Belavy, D.L., Ibragimov, B., Korez, R., Vrtovec, T., Hutt, H., Everson, R., Meakin, J., Andrade, I.L., et al.: Evaluation and comparison of 3d intervertebral disc localization and segmentation methods for 3d t2 mr data: A grand challenge. Medical image analysis 35, 327-344 (2017) +127. Zheng, X., Wang, Y., Wang, G., Liu, J.: Fast and robust segmentation of white blood cell images by self-supervised learning. Micron 107, 55-71 (2018). https://doi.org/https://doi.org/10.1016/j.micron.2018.01.010, https://www.sciencedirect.com/science/article/pii/S0968432817303037 +128. Zhou, T., Li, L., Bredell, G., Li, J., Unkelbach, J., Konukoglu, E.: Volumetric memory network for interactive medical image segmentation. Medical Image Analysis 83, 102599 (Jan 2023). https://doi.org/10.1016/j.media.2022.102599, https://linkinghub.elsevier.com/retrieve/pii/S1361841522002316 +129. Zou, X., Yang, J., Zhang, H., Li, F., Li, L., Wang, J., Wang, L., Gao, J., Lee, Y.J.: Segment everything everywhere all at once. Advances in Neural Information Processing Systems 36 (2024) \ No newline at end of file diff --git a/2024/ScribblePrompt_ Fast and Flexible Interactive Segmentation for Any Biomedical Image/images.zip b/2024/ScribblePrompt_ Fast and Flexible Interactive Segmentation for Any Biomedical Image/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d145090de5f3d6e434fdd74c775ef87f96653234 --- /dev/null +++ b/2024/ScribblePrompt_ Fast and Flexible Interactive Segmentation for Any Biomedical Image/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3e186673d24284412519c79a30991edca598aa42b3ae214830c3ca50acbce5e0 +size 375590 diff --git a/2024/ScribblePrompt_ Fast and Flexible Interactive Segmentation for Any Biomedical Image/layout.json b/2024/ScribblePrompt_ Fast and Flexible Interactive Segmentation for Any Biomedical Image/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..17b24299199de3412539fe1d747b4a3f5c52fe6f --- /dev/null +++ b/2024/ScribblePrompt_ Fast and Flexible Interactive Segmentation for Any Biomedical Image/layout.json @@ -0,0 +1,15263 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 147, + 111, + 467, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 111, + 467, + 148 + ], + "spans": [ + { + "bbox": [ + 147, + 111, + 467, + 148 + ], + "type": "text", + "content": "ScribblePrompt: Fast and Flexible Interactive Segmentation for Any Biomedical Image" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 168, + 167, + 445, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 167, + 445, + 192 + ], + "spans": [ + { + "bbox": [ + 168, + 167, + 445, + 192 + ], + "type": "text", + "content": "Hallee E. Wong" + }, + { + "bbox": [ + 168, + 167, + 445, + 192 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 168, + 167, + 445, + 192 + ], + "type": "text", + "content": ", Marianne Rakic" + }, + { + "bbox": [ + 168, + 167, + 445, + 192 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 168, + 167, + 445, + 192 + ], + "type": "text", + "content": ", John Guttag" + }, + { + "bbox": [ + 168, + 167, + 445, + 192 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 168, + 167, + 445, + 192 + ], + "type": "text", + "content": ", and Adrian V. Dalca" + }, + { + "bbox": [ + 168, + 167, + 445, + 192 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 149, + 201, + 462, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 201, + 462, + 246 + ], + "spans": [ + { + "bbox": [ + 149, + 201, + 462, + 246 + ], + "type": "text", + "content": "1 MIT CSAIL, Cambridge, MA, USA \n2 Martins Center, Massachusetts General Hospital, Charlestown, MA, USA \n3 Harvard Medical School, Boston, MA, USA \n{hallel, mrukic, guttag, adalca}@mit.edu" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 266, + 452, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 266, + 452, + 486 + ], + "spans": [ + { + "bbox": [ + 160, + 266, + 452, + 486 + ], + "type": "text", + "content": "Abstract. Biomedical image segmentation is a crucial part of both scientific research and clinical care. With enough labelled data, deep learning models can be trained to accurately automate specific biomedical image segmentation tasks. However, manually segmenting images to create training data is highly labor intensive and requires domain expertise. We present ScribblePrompt, a flexible neural network based interactive segmentation tool for biomedical imaging that enables human annotators to segment previously unseen structures using scribbles, clicks, and bounding boxes. Through rigorous quantitative experiments, we demonstrate that given comparable amounts of interaction, ScribblePrompt produces more accurate segmentations than previous methods on datasets unseen during training. In a user study with domain experts, ScribblePrompt reduced annotation time by " + }, + { + "bbox": [ + 160, + 266, + 452, + 486 + ], + "type": "inline_equation", + "content": "28\\%" + }, + { + "bbox": [ + 160, + 266, + 452, + 486 + ], + "type": "text", + "content": " while improving Dice by " + }, + { + "bbox": [ + 160, + 266, + 452, + 486 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 160, + 266, + 452, + 486 + ], + "type": "text", + "content": " compared to the next best method. ScribblePrompt's success rests on a set of careful design decisions. These include a training strategy that incorporates both a highly diverse set of images and tasks, novel algorithms for simulated user interactions and labels, and a network that enables fast inference. We showcase ScribblePrompt in an interactive demo, provide code, and release a dataset of scribble annotations at https://scribbleprompt.csail.mit.edu" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 496, + 451, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 496, + 451, + 507 + ], + "spans": [ + { + "bbox": [ + 160, + 496, + 451, + 507 + ], + "type": "text", + "content": "Keywords: Interactive Segmentation " + }, + { + "bbox": [ + 160, + 496, + 451, + 507 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 496, + 451, + 507 + ], + "type": "text", + "content": " Biomedical Imaging " + }, + { + "bbox": [ + 160, + 496, + 451, + 507 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 496, + 451, + 507 + ], + "type": "text", + "content": " Scribbles" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 524, + 230, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 524, + 230, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 524, + 230, + 536 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 545, + 481, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 481, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 481, + 629 + ], + "type": "text", + "content": "Biomedical image segmentation is an essential step in a wide range of biomedical research and clinical care pipelines. Deep learning has become the predominant method to automate existing segmentation tasks [43, 78, 92, 106]. Biomedical researchers and clinicians often encounter novel segmentation tasks involving either new regions of interest [119] or new image modalities [77,97]. Unfortunately, supervised training of accurate models for new domains requires diverse images with careful annotations by skilled experts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 630, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 481, + 666 + ], + "type": "text", + "content": "Most widely used interactive segmentation systems for biomedical imaging provide minimal, intensity-based, algorithmic assistance [118]. Despite a growing literature, learning-based systems are less widely used in practice, perhaps" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 122, + 176, + 166 + ], + "blocks": [ + { + "bbox": [ + 148, + 114, + 167, + 121 + ], + "lines": [ + { + "bbox": [ + 148, + 114, + 167, + 121 + ], + "spans": [ + { + "bbox": [ + 148, + 114, + 167, + 121 + ], + "type": "text", + "content": "Step 1" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 134, + 122, + 176, + 166 + ], + "lines": [ + { + "bbox": [ + 134, + 122, + 176, + 166 + ], + "spans": [ + { + "bbox": [ + 134, + 122, + 176, + 166 + ], + "type": "image", + "image_path": "ab2a2cc597ac03ffede42436570130f5d75b3c0c560ce4b4d14c3225dabb1cb7.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 177, + 122, + 219, + 166 + ], + "blocks": [ + { + "bbox": [ + 188, + 114, + 209, + 121 + ], + "lines": [ + { + "bbox": [ + 188, + 114, + 209, + 121 + ], + "spans": [ + { + "bbox": [ + 188, + 114, + 209, + 121 + ], + "type": "text", + "content": "Step 2" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 177, + 122, + 219, + 166 + ], + "lines": [ + { + "bbox": [ + 177, + 122, + 219, + 166 + ], + "spans": [ + { + "bbox": [ + 177, + 122, + 219, + 166 + ], + "type": "image", + "image_path": "0ac61d155e098a22eb0ae8e997599603d95e09b732829390570cc097b4c7492f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 220, + 122, + 261, + 166 + ], + "blocks": [ + { + "bbox": [ + 218, + 114, + 251, + 121 + ], + "lines": [ + { + "bbox": [ + 218, + 114, + 251, + 121 + ], + "spans": [ + { + "bbox": [ + 218, + 114, + 251, + 121 + ], + "type": "text", + "content": "Step 3" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 220, + 122, + 261, + 166 + ], + "lines": [ + { + "bbox": [ + 220, + 122, + 261, + 166 + ], + "spans": [ + { + "bbox": [ + 220, + 122, + 261, + 166 + ], + "type": "image", + "image_path": "a9016d7deb2cfa31379d1f933b4d180a89a4b1622c9286b10f94aa82d44d1668.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 263, + 122, + 306, + 166 + ], + "blocks": [ + { + "bbox": [ + 263, + 114, + 305, + 121 + ], + "lines": [ + { + "bbox": [ + 263, + 114, + 305, + 121 + ], + "spans": [ + { + "bbox": [ + 263, + 114, + 305, + 121 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 263, + 122, + 306, + 166 + ], + "lines": [ + { + "bbox": [ + 263, + 122, + 306, + 166 + ], + "spans": [ + { + "bbox": [ + 263, + 122, + 306, + 166 + ], + "type": "image", + "image_path": "24a6138a080e8ccd473dd682cfeeefe45d4cc88a6c12a2f401143874bfcca03b.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 307, + 122, + 348, + 166 + ], + "blocks": [ + { + "bbox": [ + 318, + 114, + 338, + 121 + ], + "lines": [ + { + "bbox": [ + 318, + 114, + 338, + 121 + ], + "spans": [ + { + "bbox": [ + 318, + 114, + 338, + 121 + ], + "type": "text", + "content": "Step 1" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 122, + 348, + 166 + ], + "lines": [ + { + "bbox": [ + 307, + 122, + 348, + 166 + ], + "spans": [ + { + "bbox": [ + 307, + 122, + 348, + 166 + ], + "type": "image", + "image_path": "2312047fc6d80b06fdcb4340e676ff1e78f67d7bb6a341a2722a33f00a12168a.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 349, + 122, + 391, + 166 + ], + "blocks": [ + { + "bbox": [ + 359, + 114, + 380, + 121 + ], + "lines": [ + { + "bbox": [ + 359, + 114, + 380, + 121 + ], + "spans": [ + { + "bbox": [ + 359, + 114, + 380, + 121 + ], + "type": "text", + "content": "Step 2" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 349, + 122, + 391, + 166 + ], + "lines": [ + { + "bbox": [ + 349, + 122, + 391, + 166 + ], + "spans": [ + { + "bbox": [ + 349, + 122, + 391, + 166 + ], + "type": "image", + "image_path": "9552129b489b034cf5e7e178f3b4cba83f8b8e43cd1c979ac1c80fe1143dbb69.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 391, + 122, + 434, + 166 + ], + "blocks": [ + { + "bbox": [ + 403, + 114, + 423, + 121 + ], + "lines": [ + { + "bbox": [ + 403, + 114, + 423, + 121 + ], + "spans": [ + { + "bbox": [ + 403, + 114, + 423, + 121 + ], + "type": "text", + "content": "Step 3" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 391, + 122, + 434, + 166 + ], + "lines": [ + { + "bbox": [ + 391, + 122, + 434, + 166 + ], + "spans": [ + { + "bbox": [ + 391, + 122, + 434, + 166 + ], + "type": "image", + "image_path": "28d9093f08efce348f0b1855184a0cf61b42d36ec5c953a6820d223e8766a5c7.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 435, + 122, + 479, + 166 + ], + "blocks": [ + { + "bbox": [ + 436, + 114, + 479, + 121 + ], + "lines": [ + { + "bbox": [ + 436, + 114, + 479, + 121 + ], + "spans": [ + { + "bbox": [ + 436, + 114, + 479, + 121 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 435, + 122, + 479, + 166 + ], + "lines": [ + { + "bbox": [ + 435, + 122, + 479, + 166 + ], + "spans": [ + { + "bbox": [ + 435, + 122, + 479, + 166 + ], + "type": "image", + "image_path": "d4349865b1db0bd39b571fe9e818c8d814f9b6f77b6044e25ae1494740a5ba2f.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 134, + 167, + 179, + 209 + ], + "blocks": [ + { + "bbox": [ + 134, + 167, + 179, + 209 + ], + "lines": [ + { + "bbox": [ + 134, + 167, + 179, + 209 + ], + "spans": [ + { + "bbox": [ + 134, + 167, + 179, + 209 + ], + "type": "image", + "image_path": "697261cfefe8c33764db424f08fe24120cbb58f3a778970487dcf73e44b30bab.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 219, + 482, + 296 + ], + "lines": [ + { + "bbox": [ + 131, + 219, + 482, + 296 + ], + "spans": [ + { + "bbox": [ + 131, + 219, + 482, + 296 + ], + "type": "text", + "content": "Fig. 1: ScribblePrompt enables rapid iterative interactive segmentation of unseen tasks using bounding boxes, clicks, and scribbles. We show predictions from ScribblePrompt with iterative interaction steps on examples from datasets unseen during training. At each step, we visualize positive scribble and click inputs in green, negative scribble and click inputs in red, bounding box inputs in yellow, and the predicted segmentation in blue. Scribble thickness is enlarged for visual clarity. See Supplementary Material for more examples." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 180, + 167, + 221, + 209 + ], + "blocks": [ + { + "bbox": [ + 180, + 167, + 221, + 209 + ], + "lines": [ + { + "bbox": [ + 180, + 167, + 221, + 209 + ], + "spans": [ + { + "bbox": [ + 180, + 167, + 221, + 209 + ], + "type": "image", + "image_path": "bad9e3e9b10dfb4f7f8e31e5c54b7d379a03a82597db2034c49ae538bfcfa64b.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 222, + 167, + 262, + 209 + ], + "blocks": [ + { + "bbox": [ + 222, + 167, + 262, + 209 + ], + "lines": [ + { + "bbox": [ + 222, + 167, + 262, + 209 + ], + "spans": [ + { + "bbox": [ + 222, + 167, + 262, + 209 + ], + "type": "image", + "image_path": "35083b4c22cb5980a2ac19243b6ca7f5f39a1a1dee79373523a411b4ba0bf012.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 263, + 167, + 306, + 209 + ], + "blocks": [ + { + "bbox": [ + 263, + 167, + 306, + 209 + ], + "lines": [ + { + "bbox": [ + 263, + 167, + 306, + 209 + ], + "spans": [ + { + "bbox": [ + 263, + 167, + 306, + 209 + ], + "type": "image", + "image_path": "dd6305b1025ea0780e39f7fd7038aa1dc28f2fe9dbc210af503f1fa720ef0bf2.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 307, + 167, + 348, + 209 + ], + "blocks": [ + { + "bbox": [ + 307, + 167, + 348, + 209 + ], + "lines": [ + { + "bbox": [ + 307, + 167, + 348, + 209 + ], + "spans": [ + { + "bbox": [ + 307, + 167, + 348, + 209 + ], + "type": "image", + "image_path": "56a8c466e30b316a0623ae88158886254c05eed26dea0283f90a63e8ac606546.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 349, + 167, + 391, + 209 + ], + "blocks": [ + { + "bbox": [ + 349, + 167, + 391, + 209 + ], + "lines": [ + { + "bbox": [ + 349, + 167, + 391, + 209 + ], + "spans": [ + { + "bbox": [ + 349, + 167, + 391, + 209 + ], + "type": "image", + "image_path": "60b534ecd8ee3b27f4b6037d885c698de56b2cae4777a65d37f36718123a0965.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 391, + 167, + 434, + 209 + ], + "blocks": [ + { + "bbox": [ + 391, + 167, + 434, + 209 + ], + "lines": [ + { + "bbox": [ + 391, + 167, + 434, + 209 + ], + "spans": [ + { + "bbox": [ + 391, + 167, + 434, + 209 + ], + "type": "image", + "image_path": "b980f866bee65adbd815cbb0c10574c542e061fa5e67d7cc7f789a140699f31b.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 435, + 167, + 479, + 209 + ], + "blocks": [ + { + "bbox": [ + 435, + 167, + 479, + 209 + ], + "lines": [ + { + "bbox": [ + 435, + 167, + 479, + 209 + ], + "spans": [ + { + "bbox": [ + 435, + 167, + 479, + 209 + ], + "type": "image", + "image_path": "b663fb298dc9eec1c53be49c8be425b85f38b1a0b789db3f0d9e3b57feb3b85d.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "bbox": [ + 130, + 319, + 480, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 480, + 402 + ], + "type": "text", + "content": "because they are generally focused on specific tasks or modalities, which limits their applicability [68, 83, 95, 109, 110, 128]. Recent vision foundation models target broad applicability, but require fine-tuning to the medical domain [47, 112, 113, 129]. Most interactive segmentation models require specific interactions, like carefully placed clicks, making them easy to develop but difficult to use in practice. The result of these shortcomings is that existing learning-based models are not widely used in practice." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 131, + 403, + 480, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 403, + 480, + 426 + ], + "spans": [ + { + "bbox": [ + 131, + 403, + 480, + 426 + ], + "type": "text", + "content": "In this work, we present ScribblePrompt, a general model for interactive biomedical image segmentation that" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 135, + 433, + 480, + 503 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 136, + 433, + 480, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 433, + 480, + 467 + ], + "spans": [ + { + "bbox": [ + 136, + 433, + 480, + 467 + ], + "type": "text", + "content": "1. enables users to rapidly and accurately accomplish any biomedical image segmentation task, outperforming state-of-the-art models, particularly for unseen labels and image types;" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 135, + 468, + 480, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 468, + 480, + 491 + ], + "spans": [ + { + "bbox": [ + 135, + 468, + 480, + 491 + ], + "type": "text", + "content": "2. is flexible to different annotation styles, including bounding boxes, clicks, and scribbles," + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 135, + 491, + 474, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 491, + 474, + 503 + ], + "spans": [ + { + "bbox": [ + 135, + 491, + 474, + 503 + ], + "type": "text", + "content": "3. is computationally efficient, enabling fast inference, even on a single CPU." + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 510, + 480, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 480, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 480, + 581 + ], + "type": "text", + "content": "To achieve this, we focused on design decisions aimed at ease of use and realistic interactions. To train our model, we started by gathering a large corpus of biomedical imaging datasets. We then designed task augmentation through synthesis strategies to encourage generalization to unseen tasks. Finally, we designed new interaction simulation strategies for training. The model itself uses an architecture optimized for efficient inference." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": "To evaluate ScribblePrompt, we compared its performance to that of existing interactive segmentation methods on datasets unseen during training. These experiments involved manually-collected scribbles, simulated interactions (Fig. 1), and a user study in which experienced annotators were asked to segment images. In the user study, ScribblePrompt reduced annotation time by " + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "inline_equation", + "content": "28\\%" + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": " while increasing Dice by " + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": " compared to the next best method, the Segment Anything Model (SAM) [47]. Given similar amounts of interaction, ScribblePrompt" + } + ] + } + ], + "index": 34 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 236, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 236, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 236, + 102 + ], + "type": "text", + "content": "H.E. Wong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "achieved consistently better Dice scores than other methods. We release an interactive tool, code, model weights, and our dataset of manually-collected scribbles at https://scribbleprompt.csail.mit.edu." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 171, + 242, + 185 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 171, + 242, + 185 + ], + "spans": [ + { + "bbox": [ + 132, + 171, + 242, + 185 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 198, + 482, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 198, + 482, + 329 + ], + "spans": [ + { + "bbox": [ + 130, + 198, + 482, + 329 + ], + "type": "text", + "content": "Interactive Biomedical Image Segmentation. Early research into interactive segmentation of biomedical images focused on traditional intensity-based methods [13,21,33,107,118]. Recent deep-learning based techniques [68,83,109, 110] use human interaction to improve the accuracy of segmentation tasks where labelled data exists, but fully-automatic segmentation methods fail to produce accurate segmentations. This approach assumes that the model will be used to perform the same segmentation task(s) it was trained on, leading to domain-specific solutions [6,62,80,89,110,111]. A few methods generalize to new, but similar, classes [95] and modalities [68,109] to those seen in training. In contrast, ScribblePrompt is designed to help annotators segment a wide range of biomedical images and (potentially unseen) tasks at inference time." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 330, + 482, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 330, + 482, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 330, + 482, + 389 + ], + "type": "text", + "content": "Foundation Models. Recent vision foundation models employ prompting to enable generalization to new tasks. These models, trained on large collections of (natural) image data, segment potentially-unseen structures specified by spatial prompts [47, 129], text [47, 87, 129], or examples [15, 22, 112, 113]. Some of these models can be used for interactive segmentation with spatial prompts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 390, + 482, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 482, + 532 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 482, + 532 + ], + "type": "text", + "content": "Fine-tuning models initially developed using natural images is often unhelpful for specific tasks in the biomedical imaging domain, compared to training from scratch [88]. The Segment Anything Model (SAM) [47], a foundation model trained on natural images, performs well at biomedical image segmentation tasks with clear boundaries (e.g. organs in abdominal CT), but performs poorly on tasks and modalities involving more subtle delineations (e.g. deep structures in Brain MRI) [36, 41, 76, 100]. Several recent or concurrent papers fine-tune SAM for specific biomedical tasks or modalities [40, 46, 59, 81, 108, 120]. A few fine-tune SAM to segment medical images from multiple-modalities with bounding boxes [69], clicks [115] or both [19]. We show in our experiments that these models don't perform well on many unseen tasks and are limiting in their required interactions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 640 + ], + "type": "text", + "content": "Several natural image methods simulate iterative user interactions and condition the model on previous predictions to train a single network to make the initial prediction and perform refinement [14, 27, 47, 61, 102, 103]. This approach is appropriate for natural images where there is ample data and typically one modality. However, such methods have not previously been developed for the biomedical imaging domain, which is more fragmented because of specialized regions of interest and different modalities. We use iterative ideas in ScribblePrompt, but we facilitate iterative interactive segmentation of unseen biomedical imaging tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": "User Interaction. Users can prompt interactive segmentation systems in many different ways, such as bounding boxes [9, 47, 89, 94, 109, 117], scribbles" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "type": "text", + "content": "ScribblePrompt" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": "[18, 110, 111], or clicks [9, 61, 68, 103, 110]. Few interactive segmentation models incorporate multiple types of inputs. Some works have explored training with more specialized types of clicks [122], such as extreme points [72, 93], interior margin points [68], and center clicks [103, 116], to maximize the information per interaction. Specialized interactions often require more user time, and models trained only with such interactions are less robust to deviations from the interaction protocol [9, 103, 117]. In contrast, we focus on a user-friendly model that performs well under a variety of interaction scenarios." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 213, + 482, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 482, + 320 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 482, + 320 + ], + "type": "text", + "content": "Scribbles are an intuitive form of interaction, however few interactive segmentation works employ them, because acquiring realistic scribbles for training and evaluation is challenging. Previous deep-learning models trained for interactive scribble-based segmentation either relied on collecting large datasets of manually drawn scribbles [53, 57, 128] or simplistic approaches such as sampling random points [5, 7, 110] or Bezier curves [2]. One work has explored simulating more complex scribbles for interactive segmentation of natural images [18]. We build on these concepts to create a new scribble simulation engine for training and evaluation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 322, + 482, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 322, + 482, + 406 + ], + "spans": [ + { + "bbox": [ + 130, + 322, + 482, + 406 + ], + "type": "text", + "content": "Scribble-supervised learning methods use scribble annotations as supervision to train automatic segmentation models for predicting segmentation given only an input image [32,56,57,66,121]. However, these methods require the manual scribble-annotation of many training images and retraining for each new task. In contrast, ScribblePrompt is trained on a large corpus of datasets once and then can be used to perform new segmentation tasks at inference time without retraining, using scribbles as input." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 430, + 309, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 309, + 443 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 309, + 443 + ], + "type": "text", + "content": "3 ScribblePrompt Approach" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 460, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 482, + 544 + ], + "type": "text", + "content": "We present an interactive segmentation method that can generalize to new biomedical imaging modalities and regions of interest, while staying focused on practical usability. We describe the problem formulation, and present the important aspects of the ScribblePrompt framework: (i) simulation of realistic interactions during training, (ii) augmentation with synthetic labels during training to encourage generalization, and (iii) an efficient architecture for fast inference. We show an overview of training in Fig. 2 and 4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 567, + 269, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 269, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 269, + 578 + ], + "type": "text", + "content": "3.1 Problem Formulation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": " be a segmentation task consisting of image and segmentation pairs, " + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\{(x^t,y^t)_j\\}_{j = 1}^N" + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": ". At step " + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": ", given an image " + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "inline_equation", + "content": "x^{t}" + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": ", a set of user interactions " + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "inline_equation", + "content": "u_{i}" + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": ", and previous prediction " + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\hat{y}_{i - 1}^{t}" + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": ", we learn function " + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "inline_equation", + "content": "f_{\\theta}(x^{t},u_{i},\\hat{y}_{i - 1}^{t})" + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": " with parameters " + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": " that produces a segmentation " + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\hat{y}_i" + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": ". The set of interactions " + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "inline_equation", + "content": "u_{i}" + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": ", which may include positive or negative scribbles, positive or negative clicks, and bounding boxes, is provided by a user who has access to the image " + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "inline_equation", + "content": "x^{t}" + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": " and previous prediction " + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\hat{y}_{i - 1}^{t}" + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 236, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 236, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 236, + 102 + ], + "type": "text", + "content": "H.E. Wong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "We minimize the difference between the true segmentation " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "y^{t}" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " and each of the " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " iterative predictions " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\hat{y}_1, \\dots, \\hat{y}_k" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 174, + 146, + 481, + 180 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 146, + 481, + 180 + ], + "spans": [ + { + "bbox": [ + 174, + 146, + 481, + 180 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (\\theta ; \\mathcal {T}) = \\mathbb {E} _ {t \\in \\mathcal {T}} \\left[ \\mathbb {E} _ {(x ^ {t}, y ^ {t}) \\in t} \\left[ \\sum_ {i = 1} ^ {k} \\mathcal {L} _ {\\text {S e g}} \\left(y ^ {t}, f _ {\\theta} \\left(x ^ {t}, u _ {i}, \\hat {y} _ {i - 1} ^ {t}\\right)\\right) \\right] \\right], \\tag {1}", + "image_path": "af97a829f4ca0d2c1efa742c630e6a5899f2cd11859b2d89b688832887bfaddc.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 185, + 333, + 197 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 185, + 333, + 197 + ], + "spans": [ + { + "bbox": [ + 130, + 185, + 333, + 197 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 185, + 333, + 197 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Seg}" + }, + { + "bbox": [ + 130, + 185, + 333, + 197 + ], + "type": "text", + "content": " is a supervised segmentation loss." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 198, + 480, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 198, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 198, + 480, + 281 + ], + "type": "text", + "content": "During training, we sample a task " + }, + { + "bbox": [ + 130, + 198, + 480, + 281 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 198, + 480, + 281 + ], + "type": "text", + "content": " from training task collection " + }, + { + "bbox": [ + 130, + 198, + 480, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 130, + 198, + 480, + 281 + ], + "type": "text", + "content": ", from which we sample an image " + }, + { + "bbox": [ + 130, + 198, + 480, + 281 + ], + "type": "inline_equation", + "content": "x^t" + }, + { + "bbox": [ + 130, + 198, + 480, + 281 + ], + "type": "text", + "content": ", and segmentation map " + }, + { + "bbox": [ + 130, + 198, + 480, + 281 + ], + "type": "inline_equation", + "content": "y^{t}" + }, + { + "bbox": [ + 130, + 198, + 480, + 281 + ], + "type": "text", + "content": ". We simulate a possible set of interactions " + }, + { + "bbox": [ + 130, + 198, + 480, + 281 + ], + "type": "inline_equation", + "content": "u_{i}" + }, + { + "bbox": [ + 130, + 198, + 480, + 281 + ], + "type": "text", + "content": " based on " + }, + { + "bbox": [ + 130, + 198, + 480, + 281 + ], + "type": "inline_equation", + "content": "y^{t}" + }, + { + "bbox": [ + 130, + 198, + 480, + 281 + ], + "type": "text", + "content": ", and predict " + }, + { + "bbox": [ + 130, + 198, + 480, + 281 + ], + "type": "inline_equation", + "content": "\\hat{y}_i^t = f_\\theta (x^t,u_i,\\hat{y}_{i - 1}^t)" + }, + { + "bbox": [ + 130, + 198, + 480, + 281 + ], + "type": "text", + "content": ". We simulate the next set of interactions " + }, + { + "bbox": [ + 130, + 198, + 480, + 281 + ], + "type": "inline_equation", + "content": "u_{i + 1}" + }, + { + "bbox": [ + 130, + 198, + 480, + 281 + ], + "type": "text", + "content": " based on the error region " + }, + { + "bbox": [ + 130, + 198, + 480, + 281 + ], + "type": "inline_equation", + "content": "\\varepsilon_i^t = y^t -\\hat{y}_i^t" + }, + { + "bbox": [ + 130, + 198, + 480, + 281 + ], + "type": "text", + "content": ", and repeat for " + }, + { + "bbox": [ + 130, + 198, + 480, + 281 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 198, + 480, + 281 + ], + "type": "text", + "content": " iterations. In the following sections, we describe the core aspects of the framework: strategies for simulating " + }, + { + "bbox": [ + 130, + 198, + 480, + 281 + ], + "type": "inline_equation", + "content": "u_{i}" + }, + { + "bbox": [ + 130, + 198, + 480, + 281 + ], + "type": "text", + "content": ", collecting " + }, + { + "bbox": [ + 130, + 198, + 480, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 130, + 198, + 480, + 281 + ], + "type": "text", + "content": ", sampling " + }, + { + "bbox": [ + 130, + 198, + 480, + 281 + ], + "type": "inline_equation", + "content": "(x^{t},y^{t})" + }, + { + "bbox": [ + 130, + 198, + 480, + 281 + ], + "type": "text", + "content": ", building " + }, + { + "bbox": [ + 130, + 198, + 480, + 281 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 130, + 198, + 480, + 281 + ], + "type": "text", + "content": ", and optimizing " + }, + { + "bbox": [ + 130, + 198, + 480, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\theta ;\\mathcal{T})" + }, + { + "bbox": [ + 130, + 198, + 480, + 281 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 133, + 300, + 482, + 439 + ], + "blocks": [ + { + "bbox": [ + 133, + 300, + 482, + 439 + ], + "lines": [ + { + "bbox": [ + 133, + 300, + 482, + 439 + ], + "spans": [ + { + "bbox": [ + 133, + 300, + 482, + 439 + ], + "type": "image", + "image_path": "f67e2ca5adc2a5a4162dc1ba591ce497f51e904eb348184d2b18087b5a5cbb82.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 449, + 482, + 550 + ], + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 550 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 550 + ], + "type": "text", + "content": "Fig. 2: Training. We simulate " + }, + { + "bbox": [ + 130, + 449, + 482, + 550 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 449, + 482, + 550 + ], + "type": "text", + "content": " consecutive steps of interactive segmentation. Given an image segmentation pair " + }, + { + "bbox": [ + 130, + 449, + 482, + 550 + ], + "type": "inline_equation", + "content": "(x^t, y^t)" + }, + { + "bbox": [ + 130, + 449, + 482, + 550 + ], + "type": "text", + "content": ", we first simulate a set of initial interactions " + }, + { + "bbox": [ + 130, + 449, + 482, + 550 + ], + "type": "inline_equation", + "content": "u_1" + }, + { + "bbox": [ + 130, + 449, + 482, + 550 + ], + "type": "text", + "content": ", which may contain bounding boxes, clicks, and/or scribbles. We predict segmentation " + }, + { + "bbox": [ + 130, + 449, + 482, + 550 + ], + "type": "inline_equation", + "content": "\\hat{y}_1^t \\coloneqq f_\\theta(x^t, u_1, \\hat{y}_0^t)" + }, + { + "bbox": [ + 130, + 449, + 482, + 550 + ], + "type": "text", + "content": " where the initial prediction " + }, + { + "bbox": [ + 130, + 449, + 482, + 550 + ], + "type": "inline_equation", + "content": "\\hat{y}_0^t" + }, + { + "bbox": [ + 130, + 449, + 482, + 550 + ], + "type": "text", + "content": " is set to zeros. In the second step, we simulate corrections using the error region " + }, + { + "bbox": [ + 130, + 449, + 482, + 550 + ], + "type": "inline_equation", + "content": "\\varepsilon_1^t" + }, + { + "bbox": [ + 130, + 449, + 482, + 550 + ], + "type": "text", + "content": " between the previous prediction " + }, + { + "bbox": [ + 130, + 449, + 482, + 550 + ], + "type": "inline_equation", + "content": "\\hat{y}_1^t" + }, + { + "bbox": [ + 130, + 449, + 482, + 550 + ], + "type": "text", + "content": " and ground truth " + }, + { + "bbox": [ + 130, + 449, + 482, + 550 + ], + "type": "inline_equation", + "content": "y^t" + }, + { + "bbox": [ + 130, + 449, + 482, + 550 + ], + "type": "text", + "content": ", and add them to the set of initial interactions " + }, + { + "bbox": [ + 130, + 449, + 482, + 550 + ], + "type": "inline_equation", + "content": "u_1" + }, + { + "bbox": [ + 130, + 449, + 482, + 550 + ], + "type": "text", + "content": " to get " + }, + { + "bbox": [ + 130, + 449, + 482, + 550 + ], + "type": "inline_equation", + "content": "u_2" + }, + { + "bbox": [ + 130, + 449, + 482, + 550 + ], + "type": "text", + "content": ". We predict segmentation " + }, + { + "bbox": [ + 130, + 449, + 482, + 550 + ], + "type": "inline_equation", + "content": "\\hat{y}_2^t \\coloneqq f_\\theta(x^t, u_2, \\hat{y}_1^t)" + }, + { + "bbox": [ + 130, + 449, + 482, + 550 + ], + "type": "text", + "content": " and repeat to produce a series of predictions, " + }, + { + "bbox": [ + 130, + 449, + 482, + 550 + ], + "type": "inline_equation", + "content": "\\hat{y}_1^t, \\dots, \\hat{y}_k^t" + }, + { + "bbox": [ + 130, + 449, + 482, + 550 + ], + "type": "text", + "content": ". We learn " + }, + { + "bbox": [ + 130, + 449, + 482, + 550 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 449, + 482, + 550 + ], + "type": "text", + "content": " to minimize " + }, + { + "bbox": [ + 130, + 449, + 482, + 550 + ], + "type": "inline_equation", + "content": "\\sum_{i=1}^{k} \\mathcal{L}_{seg}(y^t, \\hat{y}_i^t)" + }, + { + "bbox": [ + 130, + 449, + 482, + 550 + ], + "type": "text", + "content": ", the sum of losses between the target segmentation " + }, + { + "bbox": [ + 130, + 449, + 482, + 550 + ], + "type": "inline_equation", + "content": "y^t" + }, + { + "bbox": [ + 130, + 449, + 482, + 550 + ], + "type": "text", + "content": " and iterative predictions " + }, + { + "bbox": [ + 130, + 449, + 482, + 550 + ], + "type": "inline_equation", + "content": "\\hat{y}_1^t, \\dots, \\hat{y}_k^t" + }, + { + "bbox": [ + 130, + 449, + 482, + 550 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 587, + 257, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 587, + 257, + 599 + ], + "spans": [ + { + "bbox": [ + 131, + 587, + 257, + 599 + ], + "type": "text", + "content": "3.2 Prompt Simulation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "type": "text", + "content": "To enable a practical, easy-to-use model, we encourage robustness to different types of user interactions " + }, + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "type": "inline_equation", + "content": "u_{i}" + }, + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "type": "text", + "content": ". We introduce algorithms for simulating scribbles, clicks, and bounding box inputs during training. Each scribble and click strategy (Fig. 3) can be applied to a ground truth segmentation label " + }, + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "type": "inline_equation", + "content": "y^{t}" + }, + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "type": "text", + "content": " to simulate positive interactions, or to the background " + }, + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "type": "inline_equation", + "content": "1 - y^{t}" + }, + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "type": "text", + "content": " to simulate negative interactions." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "type": "text", + "content": "ScribblePrompt" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 480, + 182 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 480, + 182 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 480, + 182 + ], + "type": "image", + "image_path": "073c31a57efb31e5feb13a8f48c98990c006c7d081dbef55624b903f19eed638.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 191, + 482, + 226 + ], + "lines": [ + { + "bbox": [ + 130, + 191, + 482, + 226 + ], + "spans": [ + { + "bbox": [ + 130, + 191, + 482, + 226 + ], + "type": "text", + "content": "Fig. 3: Simulated scribbles and clicks. Positive interactions (green) are simulated on the segmentation label " + }, + { + "bbox": [ + 130, + 191, + 482, + 226 + ], + "type": "inline_equation", + "content": "y^{t}" + }, + { + "bbox": [ + 130, + 191, + 482, + 226 + ], + "type": "text", + "content": " (blue), while negative interactions (red) are simulated on the background " + }, + { + "bbox": [ + 130, + 191, + 482, + 226 + ], + "type": "inline_equation", + "content": "1 - y^{t}" + }, + { + "bbox": [ + 130, + 191, + 482, + 226 + ], + "type": "text", + "content": ". Scribble thickness is enlarged for visual clarity." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 250, + 479, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 250, + 479, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 250, + 479, + 285 + ], + "type": "text", + "content": "We simulate positive and negative correction scribbles or clicks by applying the same strategies to the false negative region " + }, + { + "bbox": [ + 130, + 250, + 479, + 285 + ], + "type": "inline_equation", + "content": "\\varepsilon_{i}^{t} > 0" + }, + { + "bbox": [ + 130, + 250, + 479, + 285 + ], + "type": "text", + "content": " and false positive region " + }, + { + "bbox": [ + 130, + 250, + 479, + 285 + ], + "type": "inline_equation", + "content": "\\varepsilon_{i}^{t} < 0" + }, + { + "bbox": [ + 130, + 250, + 479, + 285 + ], + "type": "text", + "content": " for error region " + }, + { + "bbox": [ + 130, + 250, + 479, + 285 + ], + "type": "inline_equation", + "content": "\\varepsilon_{i}^{t}" + }, + { + "bbox": [ + 130, + 250, + 479, + 285 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 286, + 480, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 286, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 286, + 480, + 346 + ], + "type": "text", + "content": "Scribbles. Given binary mask " + }, + { + "bbox": [ + 130, + 286, + 480, + 346 + ], + "type": "inline_equation", + "content": "y \\in \\{0,1\\}^{h \\times w}" + }, + { + "bbox": [ + 130, + 286, + 480, + 346 + ], + "type": "text", + "content": ", we simulate a scribble mask " + }, + { + "bbox": [ + 130, + 286, + 480, + 346 + ], + "type": "inline_equation", + "content": "s \\in [0,1]^{h \\times w}" + }, + { + "bbox": [ + 130, + 286, + 480, + 346 + ], + "type": "text", + "content": " by first generating clean scribbles, and then corrupting them to account for user behavior and variability. We illustrate this process in Supplementary Material Sec. B.1. To generate the clean scribbles, we start with one of these strategies:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 129, + 356, + 480, + 465 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 132, + 356, + 479, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 356, + 479, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 356, + 479, + 380 + ], + "type": "text", + "content": "(i) Line Scribbles: We draw random lines by connecting two end points sampled from " + }, + { + "bbox": [ + 132, + 356, + 479, + 380 + ], + "type": "inline_equation", + "content": "\\{(u,v)|y_{uv} = 1\\}" + }, + { + "bbox": [ + 132, + 356, + 479, + 380 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 380, + 480, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 480, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 480, + 403 + ], + "type": "text", + "content": "(ii) Centerline Scribbles: We simulate scribbles in the center of label " + }, + { + "bbox": [ + 130, + 380, + 480, + 403 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 380, + 480, + 403 + ], + "type": "text", + "content": " using a thinning algorithm [123] that reduces the label to a 1-pixel wide skeleton." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 129, + 405, + 480, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 405, + 480, + 465 + ], + "spans": [ + { + "bbox": [ + 129, + 405, + 480, + 465 + ], + "type": "text", + "content": "(iii) Contour Scribbles: We simulate a rough contour of the desired segmentation within the boundaries of the mask. We first blur the mask to reduce the size of the label such that " + }, + { + "bbox": [ + 129, + 405, + 480, + 465 + ], + "type": "inline_equation", + "content": "\\tilde{y} = \\min(y, y \\circ G_k)" + }, + { + "bbox": [ + 129, + 405, + 480, + 465 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 129, + 405, + 480, + 465 + ], + "type": "inline_equation", + "content": "G_k" + }, + { + "bbox": [ + 129, + 405, + 480, + 465 + ], + "type": "text", + "content": " is a Gaussian blur kernel. Then we apply a threshold " + }, + { + "bbox": [ + 129, + 405, + 480, + 465 + ], + "type": "inline_equation", + "content": "\\tilde{y} < h" + }, + { + "bbox": [ + 129, + 405, + 480, + 465 + ], + "type": "text", + "content": " sampled in some intensity range " + }, + { + "bbox": [ + 129, + 405, + 480, + 465 + ], + "type": "inline_equation", + "content": "h \\sim U[\\tilde{y}_{\\text{min}}, \\tilde{y}_{\\text{max}}]" + }, + { + "bbox": [ + 129, + 405, + 480, + 465 + ], + "type": "text", + "content": " and extract a contour inside the boundary of the mask." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 474, + 480, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 480, + 558 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 480, + 558 + ], + "type": "text", + "content": "To limit the size and complexity of the centerline and contour scribbles, we apply a random mask that breaks the scribbles into smaller parts. We generate the random mask by sampling a smooth noise image " + }, + { + "bbox": [ + 130, + 474, + 480, + 558 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 474, + 480, + 558 + ], + "type": "text", + "content": ", where each pixel is sampled independently from " + }, + { + "bbox": [ + 130, + 474, + 480, + 558 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mu_p, \\sigma_p^2)" + }, + { + "bbox": [ + 130, + 474, + 480, + 558 + ], + "type": "text", + "content": ", applying Gaussian blur, and then thresholding it at " + }, + { + "bbox": [ + 130, + 474, + 480, + 558 + ], + "type": "inline_equation", + "content": "\\mu_p" + }, + { + "bbox": [ + 130, + 474, + 480, + 558 + ], + "type": "text", + "content": ". We warp the resulting scribble mask " + }, + { + "bbox": [ + 130, + 474, + 480, + 558 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 474, + 480, + 558 + ], + "type": "text", + "content": " using a random deformation field " + }, + { + "bbox": [ + 130, + 474, + 480, + 558 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 474, + 480, + 558 + ], + "type": "text", + "content": " to vary the scribble shape and thickness. We ensure the resulting scribble is consistent with mask " + }, + { + "bbox": [ + 130, + 474, + 480, + 558 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 474, + 480, + 558 + ], + "type": "text", + "content": ", by multiplying the warped scribble mask " + }, + { + "bbox": [ + 130, + 474, + 480, + 558 + ], + "type": "inline_equation", + "content": "s \\circ \\phi" + }, + { + "bbox": [ + 130, + 474, + 480, + 558 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 130, + 474, + 480, + 558 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 474, + 480, + 558 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 559, + 479, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 559, + 479, + 583 + ], + "spans": [ + { + "bbox": [ + 130, + 559, + 479, + 583 + ], + "type": "text", + "content": "Clicks. Given a binary mask " + }, + { + "bbox": [ + 130, + 559, + 479, + 583 + ], + "type": "inline_equation", + "content": "y \\in \\{0,1\\}^{h \\times w}" + }, + { + "bbox": [ + 130, + 559, + 479, + 583 + ], + "type": "text", + "content": ", we simulate " + }, + { + "bbox": [ + 130, + 559, + 479, + 583 + ], + "type": "inline_equation", + "content": "n \\sim U[n_{min}, n_{max}]" + }, + { + "bbox": [ + 130, + 559, + 479, + 583 + ], + "type": "text", + "content": " clicks at a time using one of three strategies, illustrated in Fig. 3:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 592, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 132, + 592, + 480, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 592, + 480, + 616 + ], + "spans": [ + { + "bbox": [ + 132, + 592, + 480, + 616 + ], + "type": "text", + "content": "(i) Random clicks: We randomly sample clicks from all pixels in the given region " + }, + { + "bbox": [ + 132, + 592, + 480, + 616 + ], + "type": "inline_equation", + "content": "\\{(u,v)|y_{uv} = 1\\}" + }, + { + "bbox": [ + 132, + 592, + 480, + 616 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "type": "text", + "content": "(ii) Center clicks: We sample clicks from the set of points at the center of each disconnected component of the label. First, we create a multi-label mask " + }, + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "type": "inline_equation", + "content": "m \\in \\{1, \\dots, C\\}^{h \\times w}" + }, + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "type": "text", + "content": " identifying the " + }, + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "type": "text", + "content": " components of label " + }, + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "type": "text", + "content": ". We identify the center of each component using the euclidean distance transform [103, 116]." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 236, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 236, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 236, + 102 + ], + "type": "text", + "content": "H.E. Wong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 128, + 115, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 115, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 128, + 115, + 482, + 177 + ], + "type": "text", + "content": "(iii) Interior border region clicks: We sample clicks from a border region inside the boundary of the mask. We first blur the mask " + }, + { + "bbox": [ + 128, + 115, + 482, + 177 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 128, + 115, + 482, + 177 + ], + "type": "text", + "content": " to reduce the size of the label such that " + }, + { + "bbox": [ + 128, + 115, + 482, + 177 + ], + "type": "inline_equation", + "content": "\\tilde{y} = \\min(y, y \\circ G_k)" + }, + { + "bbox": [ + 128, + 115, + 482, + 177 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 128, + 115, + 482, + 177 + ], + "type": "inline_equation", + "content": "G_k" + }, + { + "bbox": [ + 128, + 115, + 482, + 177 + ], + "type": "text", + "content": " is a Gaussian blur kernel. We then sample click coordinates from " + }, + { + "bbox": [ + 128, + 115, + 482, + 177 + ], + "type": "inline_equation", + "content": "\\{(u, v) | \\tilde{y}_{uv} \\in [a, b]\\}" + }, + { + "bbox": [ + 128, + 115, + 482, + 177 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 128, + 115, + 482, + 177 + ], + "type": "inline_equation", + "content": "a, b \\sim U[\\tilde{y}_{\\text{min}}, \\tilde{y}_{\\text{max}})" + }, + { + "bbox": [ + 128, + 115, + 482, + 177 + ], + "type": "text", + "content": " are thresholds sampled in some intensity range." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 180, + 481, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 180, + 481, + 204 + ], + "spans": [ + { + "bbox": [ + 130, + 180, + 481, + 204 + ], + "type": "text", + "content": "Since users are inclined to spread out their clicks, we impose a minimum separation of a few pixels between random clicks and border region clicks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 204, + 481, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 204, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 130, + 204, + 481, + 239 + ], + "type": "text", + "content": "Bounding Boxes. We compute the minimum bounding box that encloses the label " + }, + { + "bbox": [ + 130, + 204, + 481, + 239 + ], + "type": "inline_equation", + "content": "y^{t}" + }, + { + "bbox": [ + 130, + 204, + 481, + 239 + ], + "type": "text", + "content": ", and enlarge each dimension by " + }, + { + "bbox": [ + 130, + 204, + 481, + 239 + ], + "type": "inline_equation", + "content": "r \\sim U[0,20]" + }, + { + "bbox": [ + 130, + 204, + 481, + 239 + ], + "type": "text", + "content": " pixels to account for human variability." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 239, + 481, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 239, + 481, + 324 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 481, + 324 + ], + "type": "text", + "content": "Iterative Training. During the first step " + }, + { + "bbox": [ + 130, + 239, + 481, + 324 + ], + "type": "inline_equation", + "content": "(i = 1)" + }, + { + "bbox": [ + 130, + 239, + 481, + 324 + ], + "type": "text", + "content": ", we sample the combination of interactions and the number of initial positive and negative interactions " + }, + { + "bbox": [ + 130, + 239, + 481, + 324 + ], + "type": "inline_equation", + "content": "n_{pos}, n_{neg} \\sim U[n_{min}, n_{max}]" + }, + { + "bbox": [ + 130, + 239, + 481, + 324 + ], + "type": "text", + "content": ". The initial interactions " + }, + { + "bbox": [ + 130, + 239, + 481, + 324 + ], + "type": "inline_equation", + "content": "u_{1}" + }, + { + "bbox": [ + 130, + 239, + 481, + 324 + ], + "type": "text", + "content": " are simulated using the ground truth label " + }, + { + "bbox": [ + 130, + 239, + 481, + 324 + ], + "type": "inline_equation", + "content": "y^{t}" + }, + { + "bbox": [ + 130, + 239, + 481, + 324 + ], + "type": "text", + "content": ". In subsequent steps, correction scribbles or clicks are sampled from the error region " + }, + { + "bbox": [ + 130, + 239, + 481, + 324 + ], + "type": "inline_equation", + "content": "\\varepsilon_{i-1}^{t}" + }, + { + "bbox": [ + 130, + 239, + 481, + 324 + ], + "type": "text", + "content": " between the last prediction " + }, + { + "bbox": [ + 130, + 239, + 481, + 324 + ], + "type": "inline_equation", + "content": "\\hat{y}_{i-1}^{t}" + }, + { + "bbox": [ + 130, + 239, + 481, + 324 + ], + "type": "text", + "content": " and the ground truth " + }, + { + "bbox": [ + 130, + 239, + 481, + 324 + ], + "type": "inline_equation", + "content": "y^{t}" + }, + { + "bbox": [ + 130, + 239, + 481, + 324 + ], + "type": "text", + "content": ". Since a user can make multiple corrections in each step, we sample " + }, + { + "bbox": [ + 130, + 239, + 481, + 324 + ], + "type": "inline_equation", + "content": "n_{cor} \\sim U[n_{min}, n_{max}]" + }, + { + "bbox": [ + 130, + 239, + 481, + 324 + ], + "type": "text", + "content": " corrections (scribbles or clicks) per step." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 338, + 186, + 349 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 338, + 186, + 349 + ], + "spans": [ + { + "bbox": [ + 132, + 338, + 186, + 349 + ], + "type": "text", + "content": "3.3 Data" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 354, + 481, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 481, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 481, + 450 + ], + "type": "text", + "content": "We build on large dataset gathering efforts like MegaMedical [15,90] to compile a collection of 77 open-access biomedical imaging datasets for training and evaluation, covering over 54k scans, 16 image types, and 711 labels. The collection includes a diverse array of biomedical domains, such as eyes [39,55,70,85,105], thorax [86,96,98,99,101], spine [64,98,114,126], cells [16,17,28,29,63,127], skin [20], abdominal [11,12,37,44,45,51,52,54,60,67,71,86,91,101,104], neck [48,79,82,84], brain [3,8,31,38,49,50,73-75,101], bones [34,98,114], teeth [1,65] and lesions [4,101,124,125]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 450, + 481, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 481, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 481, + 510 + ], + "type": "text", + "content": "We define a 2D segmentation task as a combination of dataset, axis (for 3D modalities), and label. For datasets with multiple segmentation labels, we consider each label separately as a binary segmentation task and for 3D modalities we use the slice with maximum label area and the middle slice from each volume. We provide more details on the data in Supplementary Material Sec. C." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 510, + 481, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 481, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 481, + 569 + ], + "type": "text", + "content": "Task Diversity. During training we sample image, segmentation pairs hierarchically - by dataset and modality, axis, and then label - to balance training on datasets of different sizes. To increase the diversity of segmentation tasks, we apply data augmentation [15] to both the input image and sampled segmentation prior to simulating the user interactions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "text", + "content": "Synthetic Labels. To limit overfitting to specific segmentation tasks, we introduce a mechanism to generate synthetic labels (Fig. 4). During training, for a given sample " + }, + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "inline_equation", + "content": "(x_0,y_0)" + }, + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "text", + "content": ", with probability " + }, + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "inline_equation", + "content": "p_{synth}" + }, + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "text", + "content": ", we replace " + }, + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "text", + "content": " with a synthetic label " + }, + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "inline_equation", + "content": "y_{synth}" + }, + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "text", + "content": ". Given an image " + }, + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "text", + "content": ", we generate a synthetic label " + }, + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "inline_equation", + "content": "y_{synth}" + }, + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "text", + "content": " by applying a superpixel algorithm [25] with scale parameter " + }, + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\lambda \\sim U[1,\\lambda_{max}]" + }, + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "text", + "content": " to partition the image into a multi-label mask of " + }, + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "text", + "content": " superpixels " + }, + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "inline_equation", + "content": "z \\in \\{1,\\dots,k\\}^{n\\times n}" + }, + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "text", + "content": ", and then randomly select a superpixel " + }, + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "inline_equation", + "content": "y_{synth} = \\mathbb{1}(z = c)" + }, + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "text", + "content": ". We conduct experiments varying " + }, + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "inline_equation", + "content": "p_{synth}" + }, + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "text", + "content": " and show examples in Supplementary Material Sec. B.3." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "type": "text", + "content": "ScribblePrompt" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 113, + 482, + 189 + ], + "blocks": [ + { + "bbox": [ + 132, + 113, + 482, + 189 + ], + "lines": [ + { + "bbox": [ + 132, + 113, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 132, + 113, + 482, + 189 + ], + "type": "image", + "image_path": "f9a18ea5e998c44d1f23a8916c05b45ba89e9ccff0d65123a7d44b80b8497cb8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 195, + 483, + 251 + ], + "lines": [ + { + "bbox": [ + 130, + 195, + 483, + 251 + ], + "spans": [ + { + "bbox": [ + 130, + 195, + 483, + 251 + ], + "type": "text", + "content": "Fig. 4: Task sampling and augmentation. During training, we sample an image and segmentation pair " + }, + { + "bbox": [ + 130, + 195, + 483, + 251 + ], + "type": "inline_equation", + "content": "(x_0, y_0)" + }, + { + "bbox": [ + 130, + 195, + 483, + 251 + ], + "type": "text", + "content": ". With probability " + }, + { + "bbox": [ + 130, + 195, + 483, + 251 + ], + "type": "inline_equation", + "content": "p_{\\text{synth}}" + }, + { + "bbox": [ + 130, + 195, + 483, + 251 + ], + "type": "text", + "content": ", we replace " + }, + { + "bbox": [ + 130, + 195, + 483, + 251 + ], + "type": "inline_equation", + "content": "y_0" + }, + { + "bbox": [ + 130, + 195, + 483, + 251 + ], + "type": "text", + "content": " with a synthetic label " + }, + { + "bbox": [ + 130, + 195, + 483, + 251 + ], + "type": "inline_equation", + "content": "y_{\\text{synth}}" + }, + { + "bbox": [ + 130, + 195, + 483, + 251 + ], + "type": "text", + "content": ". We generate " + }, + { + "bbox": [ + 130, + 195, + 483, + 251 + ], + "type": "inline_equation", + "content": "y_{\\text{synth}}" + }, + { + "bbox": [ + 130, + 195, + 483, + 251 + ], + "type": "text", + "content": " by applying a superpixel algorithm to the image " + }, + { + "bbox": [ + 130, + 195, + 483, + 251 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 130, + 195, + 483, + 251 + ], + "type": "text", + "content": " to generate a map " + }, + { + "bbox": [ + 130, + 195, + 483, + 251 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 195, + 483, + 251 + ], + "type": "text", + "content": " of potential synthetic labels (superpixels), and then sampling one label. Finally, we apply random data augmentations to get " + }, + { + "bbox": [ + 130, + 195, + 483, + 251 + ], + "type": "inline_equation", + "content": "(x^t, y^t)" + }, + { + "bbox": [ + 130, + 195, + 483, + 251 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 275, + 205, + 286 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 275, + 205, + 286 + ], + "spans": [ + { + "bbox": [ + 132, + 275, + 205, + 286 + ], + "type": "text", + "content": "3.4 Network" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 295, + 488, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 295, + 488, + 332 + ], + "spans": [ + { + "bbox": [ + 130, + 295, + 488, + 332 + ], + "type": "text", + "content": "Motivated by producing a practical tool, we primarily demonstrate ScribblePrompt using an efficient fully-convolutional architecture similar to a UNet [92]. We also demonstrate ScribblePrompt using a vision transformer architecture [47]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 332, + 484, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 332, + 484, + 368 + ], + "spans": [ + { + "bbox": [ + 130, + 332, + 484, + 368 + ], + "type": "text", + "content": "ScribblePrompt-UNet: We use an 8 layer CNN following a decoder-encoder structure similar to the popular UNet architecture [92] without Batch Norm. Each convolutional layer has 192 features and uses PReLU activation [35]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 368, + 483, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 368, + 483, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 368, + 483, + 415 + ], + "type": "text", + "content": "ScribblePrompt-SAM: We take the smallest SAM model (ViT-b) [47] and fine-tune the decoder. SAM takes bounding box and click inputs as lists of " + }, + { + "bbox": [ + 130, + 368, + 483, + 415 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 130, + 368, + 483, + 415 + ], + "type": "text", + "content": " coordinates, and the low-resolution logits of the previous prediction " + }, + { + "bbox": [ + 130, + 368, + 483, + 415 + ], + "type": "inline_equation", + "content": "\\hat{y}_{i-1}" + }, + { + "bbox": [ + 130, + 368, + 483, + 415 + ], + "type": "text", + "content": ". To adapt SAM for scribbles, we consider each scribbled pixel as a click." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 415, + 482, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 482, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 482, + 464 + ], + "type": "text", + "content": "Loss. We minimize eq. (1) where " + }, + { + "bbox": [ + 130, + 415, + 482, + 464 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Seg}" + }, + { + "bbox": [ + 130, + 415, + 482, + 464 + ], + "type": "text", + "content": " is a linear combination of soft Dice Loss [24] and Focal Loss [58]. In preliminary experiments (Supplementary Material Sec. B.2), training with Soft Dice Loss alone or a linear combination of Soft Dice Loss and Binary-Cross Entropy Loss resulted in slightly lower Dice scores." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 483, + 272, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 272, + 498 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 272, + 498 + ], + "type": "text", + "content": "4 Experimental Setup" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 509, + 482, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 482, + 557 + ], + "type": "text", + "content": "We compare ScribblePrompt-UNet and ScribblePrompt-SAM to previous methods through experiments with manual scribbles, simulated interactions, and a user study with experienced annotators. Lastly, we report on inference runtime and ablation experiments." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 557, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 629 + ], + "type": "text", + "content": "Data. We use 65 (out of 77) datasets during training. We partition another nine datasets, ACDC [10], BUID [4], BTCV Cervix [52], DRIVE [105], HipXRay [34], PanDental [1], SCD [86], SpineWeb [126], and WBC [127], each into validation (used in model selection, but not training) and test (used only for final evaluation). Three additional datasets, TotalSegmentator [114], SCR [30], and COBRE [3,23,26], were used only for final evaluation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": "We report final results on 12 sets of data (the test splits of the " + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "inline_equation", + "content": "9 + 3" + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": " evaluation datasets). The evaluation datasets were not seen during training, and cover 608 tasks and 8 modalities, including unseen image types and unseen labels. We" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 236, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 236, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 236, + 102 + ], + "type": "text", + "content": "H.E. Wong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "selected these 12 evaluation datasets to cover a variety of modalities (MRI, CT, ultrasound, fundus photography, microscopy) and anatomical regions of interest (brain, teeth, bones, abdominal organs, muscles, heart, thorax, cells), including both healthy anatomy and lesions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 480, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 480, + 270 + ], + "type": "text", + "content": "Training. During training, we simulated five steps of interactive segmentation per example and set the maximum number of interactions per step to three. We set the minimum number of initial negative prompts to zero, and the minimum initial positive and correction prompts to one. During the first step, we sample the combination of prompt types ( clicks, scribbles, bounding boxes) and then sample the number of positive and negative interactions for each type. In each subsequent steps, we simulate either scribble corrections or click corrections with equal probability. We selected these values based on what we believe to be reasonable interactions for a user to perform." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 271, + 479, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 271, + 479, + 294 + ], + "spans": [ + { + "bbox": [ + 130, + 271, + 479, + 294 + ], + "type": "text", + "content": "Baselines. We compare to existing generalist methods for interactive segmentation, with a focus on methods developed for biomedical imaging." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 295, + 479, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 295, + 479, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 295, + 479, + 331 + ], + "type": "text", + "content": "SAM [47]: We evaluate the smallest (ViT-b) and largest (ViT-h) versions of the Segment Anything Model (SAM) trained on natural images. SAM takes bounding boxes, clicks, and the logits of the previous prediction as input." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 331, + 479, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 479, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 479, + 390 + ], + "type": "text", + "content": "SAM-Med2D [19]: SAM-Med2D is a SAM ViT-b model with additional adapter layers in the image encoder. SAM-Med2D was fine-tuned, using bounding boxes and iterative clicks as input, on a collection of biomedical imaging datasets containing 4.6M images and 19.7M segmentations. Following [19], we evaluate SAM-Med2D both with and without adapter layers." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 391, + 479, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 391, + 479, + 438 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 479, + 438 + ], + "type": "text", + "content": "MedSAM [69]: MedSAM is a SAM ViT-B model fine-tuned with bounding box prompts on a collection of biomedical imaging datasets containing 1.5M image segmentation pairs. We evaluate MedSAM with bounding boxes only, because we found it to perform poorly when given point or mask prompts." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 439, + 479, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 439, + 479, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 439, + 479, + 510 + ], + "type": "text", + "content": "MIDepSeg [68]: MIDepSeg is an interactive segmentation framework designed to generalize to unseen tasks on medical images. MIDepSeg takes interior margin points (positive clicks) as initial inputs, crops the image based on those points, and uses a CNN to make an initial prediction. Given additional clicks, the prediction can be refined using conditional random fields. We evaluate the pre-trained model, which was developed on placenta T2 MRI." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 526, + 218, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 526, + 218, + 540 + ], + "spans": [ + { + "bbox": [ + 131, + 526, + 218, + 540 + ], + "type": "text", + "content": "5 Evaluation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 549, + 479, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 549, + 479, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 549, + 479, + 609 + ], + "type": "text", + "content": "We evaluate all methods using both manual and simulated interactions, with a focus on scribbles. Simulated interactions enable us to test on a large volume of images and tasks. However, simulations do not always match user behavior. We address this by (1) collecting a diverse dataset of manual scribbles for evaluation, and (2) conducting an interactive user study." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 624, + 249, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 624, + 249, + 635 + ], + "spans": [ + { + "bbox": [ + 131, + 624, + 249, + 635 + ], + "type": "text", + "content": "5.1 Manual Scribbles" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "Setup. We evaluate on two datasets of manual scribbles. First, we collected a dataset (MedScribble) of manual scribbles from three annotators for seven" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "type": "text", + "content": "ScribblePrompt" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "segmentation tasks from unseen datasets, with a total of 31 examples. The annotators were shown five training examples with the ground truth segmentation per task and instructed to draw positive and negative scribbles to indicate the region of interest on new images. We provide more details on MedScribble and release the data in Supplementary Material Sec. E.1. Second, we used 380 slices from the ACDC dataset [10], which has scribble annotations for three labels and background." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 482, + 308 + ], + "type": "text", + "content": "In this evaluation, we make a prediction given a set of positive and negative scribbles, simulating a non-iterative scenario where the annotator draws several scribbles before running inference. We evaluate accuracy using Dice score [24] and 95th percentile Hausdorff Distance [42]. Since MIDepSeg's initial prediction network only takes positive inputs, we report results on both MIDepseg's initial prediction from the positive scribbles, and after applying MIDepSeg's refinement procedure using the negative scribbles as corrections. For SAM and variants fine-tuning SAM on clicks, we convert each scribble to a series of points. For MedSAM, we fit a bounding box to the positive scribbles." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 308, + 482, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 482, + 367 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 482, + 367 + ], + "type": "text", + "content": "Results. ScribblePrompt-UNet and ScribblePrompt-SAM produce the most accurate segmentations in a single step of manual scribbles on both our manual scribbles dataset and the ACDC scribbles dataset (Tab. 1). We show example predictions in Fig. 5 and additional comparisons to scribble-supervised learning in Supplementary Material Sec. E.3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 369, + 482, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 482, + 440 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 482, + 440 + ], + "type": "text", + "content": "SAM and SAM-Med2D do not generalize well to scribble inputs (which they were not trained for). MedSAM has better predictions than other SAM baselines using the SAM architecture, however it is not able to make use of the negative scribbles and thus often misses segmentations with holes in them (Fig. 5). The initial predictions from MIDepSeg's network are poor, but improve after applying the refinement procedure." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 138, + 495, + 479, + 640 + ], + "blocks": [ + { + "bbox": [ + 132, + 462, + 482, + 485 + ], + "lines": [ + { + "bbox": [ + 132, + 462, + 482, + 485 + ], + "spans": [ + { + "bbox": [ + 132, + 462, + 482, + 485 + ], + "type": "text", + "content": "Table 1: Manual scribbles. Mean Dice and HD95 with " + }, + { + "bbox": [ + 132, + 462, + 482, + 485 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 132, + 462, + 482, + 485 + ], + "type": "text", + "content": " CI of segmentations predicted from manually-collected scribbles. Best and second best are highlighted." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 495, + 479, + 640 + ], + "lines": [ + { + "bbox": [ + 138, + 495, + 479, + 640 + ], + "spans": [ + { + "bbox": [ + 138, + 495, + 479, + 640 + ], + "type": "table", + "html": "
MedScribble (n=93)ACDC (n=1,140)
Model↑ Dice Score↓ HD95↑ Dice Score↓ HD95
SAM (ViT-b)0.40 ± 0.0532.58 ± 3.800.20 ± 0.01108.39 ± 0.85
SAM (ViT-h)0.56 ± 0.0514.61 ± 3.180.42 ± 0.0259.78 ± 2.02
SAM-Med2D w.o. adapter0.52 ± 0.0530.34 ± 3.290.16 ± 0.01107.25 ± 1.12
SAM-Med2D w/ adapter0.30 ± 0.0619.04 ± 3.630.17 ± 0.0241.01 ± 3.04
MIDepSeg0.68 ± 0.046.94 ± 1.060.58 ± 0.015.78 ± 0.26
MIDepSeg w/ refinement0.81 ± 0.033.10 ± 0.670.73 ± 0.012.71 ± 0.18
MedSAM (box)0.70 ± 0.047.54 ± 1.350.70 ± 0.023.53 ± 0.28
ScribblePrompt-UNet0.84 ± 0.022.92 ± 0.880.84 ± 0.011.80 ± 0.11
ScribblePrompt-SAM0.87 ± 0.022.61 ± 0.840.77 ± 0.014.17 ± 0.32
", + "image_path": "619a8b4fd09eb94904b1ce3f8397b5c1328e96440ec84151823e05e565c31d5c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 235, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 235, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 235, + 102 + ], + "type": "text", + "content": "H.E. Wong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 175, + 209 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 175, + 209 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 175, + 209 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 175, + 209 + ], + "type": "image", + "image_path": "3292108fb40a947b0416d1f0645e8d081c5ec96003e686f8297a372cfa1ab0f7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 217, + 481, + 251 + ], + "lines": [ + { + "bbox": [ + 131, + 217, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 131, + 217, + 481, + 251 + ], + "type": "text", + "content": "Fig. 5: Example predictions. " + }, + { + "bbox": [ + 131, + 217, + 481, + 251 + ], + "type": "inline_equation", + "content": "\\mathrm{SP} =" + }, + { + "bbox": [ + 131, + 217, + 481, + 251 + ], + "type": "text", + "content": " ScribblePrompt. Top: predictions after one step of manual scribbles. Bottom: predictions after five steps of simulated interactions (one center click followed by one center correction click per step)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 177, + 115, + 218, + 209 + ], + "blocks": [ + { + "bbox": [ + 177, + 115, + 218, + 209 + ], + "lines": [ + { + "bbox": [ + 177, + 115, + 218, + 209 + ], + "spans": [ + { + "bbox": [ + 177, + 115, + 218, + 209 + ], + "type": "image", + "image_path": "2fcc093f3f1f6718de6b6abc8a7df8529e46cb0325fd4a8fa07ab84a9a343ae0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 220, + 114, + 261, + 209 + ], + "blocks": [ + { + "bbox": [ + 220, + 114, + 261, + 209 + ], + "lines": [ + { + "bbox": [ + 220, + 114, + 261, + 209 + ], + "spans": [ + { + "bbox": [ + 220, + 114, + 261, + 209 + ], + "type": "image", + "image_path": "f1e5ae55bbdcac26410ed052ab558e615b3d94cc7dcdac3370823f4c9108ebb8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 264, + 115, + 306, + 209 + ], + "blocks": [ + { + "bbox": [ + 264, + 115, + 306, + 209 + ], + "lines": [ + { + "bbox": [ + 264, + 115, + 306, + 209 + ], + "spans": [ + { + "bbox": [ + 264, + 115, + 306, + 209 + ], + "type": "image", + "image_path": "ef871048b2e18ae987d1724e68b5795e6b0afca13110736b76f522527d6f33a7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 309, + 114, + 350, + 209 + ], + "blocks": [ + { + "bbox": [ + 309, + 114, + 350, + 209 + ], + "lines": [ + { + "bbox": [ + 309, + 114, + 350, + 209 + ], + "spans": [ + { + "bbox": [ + 309, + 114, + 350, + 209 + ], + "type": "image", + "image_path": "7c93f90af4aa2ef573d2d05f948e5dc1311c8cf2cb8357acdb0b24b68156111d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 353, + 114, + 393, + 209 + ], + "blocks": [ + { + "bbox": [ + 353, + 114, + 393, + 209 + ], + "lines": [ + { + "bbox": [ + 353, + 114, + 393, + 209 + ], + "spans": [ + { + "bbox": [ + 353, + 114, + 393, + 209 + ], + "type": "image", + "image_path": "8d22d536c9e4881fcd9efd47dfef9a64ce92f5ad9a957240b4d522d3a265a793.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 395, + 115, + 436, + 209 + ], + "blocks": [ + { + "bbox": [ + 395, + 115, + 436, + 209 + ], + "lines": [ + { + "bbox": [ + 395, + 115, + 436, + 209 + ], + "spans": [ + { + "bbox": [ + 395, + 115, + 436, + 209 + ], + "type": "image", + "image_path": "caa0c9f793687c0de64da82217ca4cf3039054582eb7b0ae7980687e58730dc4.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 439, + 114, + 480, + 209 + ], + "blocks": [ + { + "bbox": [ + 439, + 114, + 480, + 209 + ], + "lines": [ + { + "bbox": [ + 439, + 114, + 480, + 209 + ], + "spans": [ + { + "bbox": [ + 439, + 114, + 480, + 209 + ], + "type": "image", + "image_path": "2b0fdeff4d3a48c74be1b984389f29f44605ad234aa7f51cc193c3080612cd29.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 274, + 275, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 275, + 285 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 275, + 285 + ], + "type": "text", + "content": "5.2 Simulated Interactions" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 294, + 479, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 294, + 479, + 318 + ], + "spans": [ + { + "bbox": [ + 131, + 294, + 479, + 318 + ], + "type": "text", + "content": "Setup. We use our interaction simulator to evaluate the performance of ScribblePrompt and baselines with iterative scribbles across the 12 evaluation sets." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 146, + 319, + 373, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 319, + 373, + 331 + ], + "spans": [ + { + "bbox": [ + 146, + 319, + 373, + 331 + ], + "type": "text", + "content": "For scribble-focused prompting procedures, we use:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 338, + 480, + 409 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 132, + 338, + 480, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 338, + 480, + 361 + ], + "spans": [ + { + "bbox": [ + 132, + 338, + 480, + 361 + ], + "type": "text", + "content": "- Line Scribbles: Three positive line scribbles and three negative line scribbles to start, followed by one correction line scribble at each step." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 361, + 480, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 361, + 480, + 384 + ], + "spans": [ + { + "bbox": [ + 132, + 361, + 480, + 384 + ], + "type": "text", + "content": "- Centerline Scribbles: One positive and one negative centerline scribble to start, followed by one correction centerline scribble at each step." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 386, + 480, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 480, + 409 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 480, + 409 + ], + "type": "text", + "content": "- Contour Scribbles: One positive and one negative contour scribble to start, followed by one correction contour scribble at each step." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 131, + 417, + 479, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 417, + 479, + 453 + ], + "spans": [ + { + "bbox": [ + 131, + 417, + 479, + 453 + ], + "type": "text", + "content": "We limit each scribble to cover a maximum of " + }, + { + "bbox": [ + 131, + 417, + 479, + 453 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 131, + 417, + 479, + 453 + ], + "type": "text", + "content": " pixels, where " + }, + { + "bbox": [ + 131, + 417, + 479, + 453 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 131, + 417, + 479, + 453 + ], + "type": "text", + "content": " is the maximum dimension of the image. For centerline and contour scribbles, a scribble might contain multiple disconnected components." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 146, + 453, + 358, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 453, + 358, + 464 + ], + "spans": [ + { + "bbox": [ + 146, + 453, + 358, + 464 + ], + "type": "text", + "content": "For click-focused prompting procedures, we use:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 472, + 480, + 568 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 132, + 472, + 480, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 480, + 508 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 480, + 508 + ], + "type": "text", + "content": "- Center Clicks: One positive click in the center of the largest component to start, followed by one (positive or negative) correction click per step in the center of the largest component of the error region." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 509, + 480, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 509, + 480, + 532 + ], + "spans": [ + { + "bbox": [ + 132, + 509, + 480, + 532 + ], + "type": "text", + "content": "- Random Clicks: One random positive click to start, followed by one (positive or negative) correction click per step, randomly sampled from the error region." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 533, + 480, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 480, + 568 + ], + "type": "text", + "content": "- Random Warm Start: Three random positive clicks and three random negative clicks to start, followed by one (positive or negative) correction click per step in the center of the largest component of the error region." + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 131, + 576, + 479, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 576, + 479, + 624 + ], + "spans": [ + { + "bbox": [ + 131, + 576, + 479, + 624 + ], + "type": "text", + "content": "For each example and interaction procedure, we simulate a series of iterative interactions with five random seeds. Since MIDepSeg cannot make accurate predictions with only a few clicks, we exclude it from simulations that start with a single click. For MedSAM, we show results for a bounding box prompt." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "content": "Results. Fig. 6 shows that both versions of ScribblePrompt outperform baseline methods for all simulated interaction procedures at all numbers of interactions." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "type": "text", + "content": "ScribblePrompt" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 222, + 204 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 222, + 204 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 222, + 204 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 222, + 204 + ], + "type": "image", + "image_path": "c30bd115037a18ea99ff9074afc0a7ffa858010d9cca4e27e2d0456919419029.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 211, + 482, + 245 + ], + "lines": [ + { + "bbox": [ + 130, + 211, + 482, + 245 + ], + "spans": [ + { + "bbox": [ + 130, + 211, + 482, + 245 + ], + "type": "text", + "content": "Fig. 6: Simulated clicks and scribbles. We simulate interactions following three scribble protocols and three click protocols. We show more results and example predictions, with similar trends, in Supplementary Material Sec. F.2." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 223, + 114, + 307, + 204 + ], + "blocks": [ + { + "bbox": [ + 223, + 114, + 307, + 204 + ], + "lines": [ + { + "bbox": [ + 223, + 114, + 307, + 204 + ], + "spans": [ + { + "bbox": [ + 223, + 114, + 307, + 204 + ], + "type": "image", + "image_path": "9d2bcb66fd3998c3665e198dbd1bfdba0347eddebf39b489768bc0591ec7325c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 309, + 114, + 394, + 204 + ], + "blocks": [ + { + "bbox": [ + 309, + 114, + 394, + 204 + ], + "lines": [ + { + "bbox": [ + 309, + 114, + 394, + 204 + ], + "spans": [ + { + "bbox": [ + 309, + 114, + 394, + 204 + ], + "type": "image", + "image_path": "f779e3bebd9b9bf6da2030411ffa0e05b1892f22d6a3fe7f55ee4f785d66ba42.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 395, + 114, + 481, + 204 + ], + "blocks": [ + { + "bbox": [ + 395, + 114, + 481, + 204 + ], + "lines": [ + { + "bbox": [ + 395, + 114, + 481, + 204 + ], + "spans": [ + { + "bbox": [ + 395, + 114, + 481, + 204 + ], + "type": "image", + "image_path": "54ea89099ce6f11cea8c1c5fb30574d71e868d67ad1ea77a2ed082a87ddd6349.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 266, + 479, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 266, + 479, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 479, + 304 + ], + "type": "text", + "content": "We show examples with simulated interactions in Fig. 1 and Fig. 5 (bottom). We show results with similar trends using bounding boxes and by dataset with fully-supervised baselines in Supplementary Material Sec. F." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 319, + 480, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 480, + 378 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 480, + 378 + ], + "type": "text", + "content": "Discussion. Comparing the simulated scribble and click results (Fig. 6) highlights the efficiency of scribble-based interactions. Both ScribblePrompt models reach an average Dice above 0.8 in two scribble steps. It takes five steps of carefully-placed center clicks or eight random clicks to reach the same average Dice score with ScribblePrompt-SAM." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 380, + 480, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 480, + 428 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 480, + 428 + ], + "type": "text", + "content": "Although SAM-Med2D was trained on a large biomedical imaging collection, it does not perform as well as SAM at refining its predictions on held-out evaluation sets. MedSAM only uses bounding box prompts and is not able to refine predictions, limiting its performance and usability." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 443, + 218, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 443, + 218, + 455 + ], + "spans": [ + { + "bbox": [ + 131, + 443, + 218, + 455 + ], + "type": "text", + "content": "5.3 User Study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 462, + 480, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 480, + 533 + ], + "type": "text", + "content": "To assess the practical utility of ScribblePrompt, we conducted a user study with experienced annotators. We compare ScribblePrompt-UNet to SAM (ViT-b). We selected ScribblePrompt-UNet for the user study because it had similar performance and lower latency compared to ScribblePrompt-SAM. We selected SAM (ViT-b) because it had the highest Dice score in our experiments with clicks and lower latency than the next closest baseline, SAM (ViT-h)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 534, + 481, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 481, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 481, + 640 + ], + "type": "text", + "content": "Setup. Study participants (" + }, + { + "bbox": [ + 130, + 534, + 481, + 640 + ], + "type": "inline_equation", + "content": "n = 16" + }, + { + "bbox": [ + 130, + 534, + 481, + 640 + ], + "type": "text", + "content": ") were neuroimaging researchers at an academic hospital. They used each model to segment a series of nine test images from different evaluation datasets. For each segmentation task, participants were shown the target segmentation and were asked to prompt the model until the predicted segmentation closely matched the target, or they could no longer improve the prediction. We provided participants with the target segmentation to disentangle the cognitive process of identifying the region of interest from prompting the model to achieve the desired segmentation. We provide additional details and visualizations in Supplementary Material Sec. G." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": "Results. Participants produced more accurate segmentations (0.84 vs. 0.73 Dice; " + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "p = 0.001" + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": " using a paired t-test) using ScribblePrompt (Tab. 2). Partici-" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 235, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 235, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 235, + 102 + ], + "type": "text", + "content": "H.E. Wong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 155, + 146, + 461, + 192 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 2: User study. Mean ± std. for Dice score, HD95 and total time per segmentation. Median (and max) number of times the participants refreshed the prediction." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 155, + 146, + 461, + 192 + ], + "lines": [ + { + "bbox": [ + 155, + 146, + 461, + 192 + ], + "spans": [ + { + "bbox": [ + 155, + 146, + 461, + 192 + ], + "type": "table", + "html": "
Dice Score (↑)HD95 (↓)Sec./task (↓)Iter./task (↓)
SAM (ViT-b)0.73 ± 0.3016.1 ± 37.1131.7 ± 168.56 (47)
ScribblePrompt-UNet0.84 ± 0.134.5 ± 9.194.2 ± 89.03 (19)
", + "image_path": "70b90ceead4ab70e81fc32d8d05a635d9b15914b5c022bc6db6fa29d9406fcd3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 213, + 479, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 479, + 262 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 479, + 262 + ], + "type": "text", + "content": "pants spent " + }, + { + "bbox": [ + 130, + 213, + 479, + 262 + ], + "type": "inline_equation", + "content": "\\approx 1.5" + }, + { + "bbox": [ + 130, + 213, + 479, + 262 + ], + "type": "text", + "content": " minutes per segmentation on average using ScribblePrompt, compared to over 2 minutes per segmentation with SAM (" + }, + { + "bbox": [ + 130, + 213, + 479, + 262 + ], + "type": "inline_equation", + "content": "p = 0.02" + }, + { + "bbox": [ + 130, + 213, + 479, + 262 + ], + "type": "text", + "content": " using a paired t-test). While using ScribblePrompt, participants updated the prediction fewer times before being satisfied with the segmentation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 262, + 480, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 262, + 480, + 332 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 480, + 332 + ], + "type": "text", + "content": "Upon completion, 15 out of 16 participants reported they preferred using ScribblePrompt to SAM and one participant had no preference. All participants reported it was easier to achieve the target segmentation using ScribblePrompt compared to SAM. " + }, + { + "bbox": [ + 130, + 262, + 480, + 332 + ], + "type": "inline_equation", + "content": "93.8\\%" + }, + { + "bbox": [ + 130, + 262, + 480, + 332 + ], + "type": "text", + "content": " of participants reported ScribblePrompt was better than SAM at refining its predictions in response to scribbles. " + }, + { + "bbox": [ + 130, + 262, + 480, + 332 + ], + "type": "inline_equation", + "content": "87.5\\%" + }, + { + "bbox": [ + 130, + 262, + 480, + 332 + ], + "type": "text", + "content": " of participants preferred using ScribblePrompt over SAM for clicks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 334, + 480, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 334, + 480, + 417 + ], + "spans": [ + { + "bbox": [ + 130, + 334, + 480, + 417 + ], + "type": "text", + "content": "Discussion. The most common factors that influenced participant preference for ScribblePrompt were 1) being able get accurate predictions from multiple types of prompts, including scribbles and 2) responsiveness to their corrections, which enabled fine-grained control over the next prediction. For some tasks, such as retinal vein segmentation, SAM was unable to make accurate predictions even with many corrections. As a result, there was more variability in Dice score and time per segmentation, for SAM than ScribblePrompt." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 435, + 254, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 254, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 254, + 445 + ], + "type": "text", + "content": "5.4 Inference Runtime" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 454, + 480, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 454, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 130, + 454, + 480, + 514 + ], + "type": "text", + "content": "Setup. We evaluate computational efficiency by measuring inference time on a single CPU for one prediction with a scribble input covering 128 pixels. Performance on a CPU reflects practical utility. Because of a variety of barriers such as protected health information, users may not be able to send their data to a server for computation, and must rely on local, most often CPU-only, computing." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 514, + 480, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 514, + 480, + 586 + ], + "spans": [ + { + "bbox": [ + 130, + 514, + 480, + 586 + ], + "type": "text", + "content": "Results. On a single CPU, ScribblePrompt-UNet requires " + }, + { + "bbox": [ + 130, + 514, + 480, + 586 + ], + "type": "inline_equation", + "content": "0.27 \\pm 0.04" + }, + { + "bbox": [ + 130, + 514, + 480, + 586 + ], + "type": "text", + "content": " sec per prediction, enabling the model to be used even in low-resource environments. The next most accurate model (Fig. 6), SAM ViT-h, requires over 2 minutes per prediction on a CPU (" + }, + { + "bbox": [ + 130, + 514, + 480, + 586 + ], + "type": "inline_equation", + "content": "130.79 \\pm 7.96" + }, + { + "bbox": [ + 130, + 514, + 480, + 586 + ], + "type": "text", + "content": " sec). SAM ViT-b takes " + }, + { + "bbox": [ + 130, + 514, + 480, + 586 + ], + "type": "inline_equation", + "content": "13.59 \\pm 0.77" + }, + { + "bbox": [ + 130, + 514, + 480, + 586 + ], + "type": "text", + "content": " seconds per prediction. We report results for all baselines and on GPU hardware with similar trends in Supplementary Material Sec. H." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 604, + 212, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 604, + 212, + 616 + ], + "spans": [ + { + "bbox": [ + 132, + 604, + 212, + 616 + ], + "type": "text", + "content": "6 Ablations" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "content": "We conduct two ablations of important design decisions: (1) synthetic label inputs used during training, and (2) types of prompts simulated during training. We report results on the validation splits of nine datasets unseen during training." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "type": "text", + "content": "ScribblePrompt" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 117, + 270, + 178 + ], + "blocks": [ + { + "bbox": [ + 133, + 117, + 270, + 178 + ], + "lines": [ + { + "bbox": [ + 133, + 117, + 270, + 178 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 270, + 178 + ], + "type": "image", + "image_path": "12d889d1146f83e3028f91524b33199dfd5b2f5e9b6906932f7b908a2280af33.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 134, + 180, + 272, + 190 + ], + "lines": [ + { + "bbox": [ + 134, + 180, + 272, + 190 + ], + "spans": [ + { + "bbox": [ + 134, + 180, + 272, + 190 + ], + "type": "text", + "content": "(a) Probability of synthetic labels." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 281, + 114, + 477, + 178 + ], + "blocks": [ + { + "bbox": [ + 281, + 114, + 477, + 178 + ], + "lines": [ + { + "bbox": [ + 281, + 114, + 477, + 178 + ], + "spans": [ + { + "bbox": [ + 281, + 114, + 477, + 178 + ], + "type": "image", + "image_path": "2a30fcd60f9694902e6ee937ad81de14cbe0f8544de94ab59c8db3407cade487.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 181, + 446, + 191 + ], + "lines": [ + { + "bbox": [ + 317, + 181, + 446, + 191 + ], + "spans": [ + { + "bbox": [ + 317, + 181, + 446, + 191 + ], + "type": "text", + "content": "(b) Interactions during training." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 199, + 482, + 232 + ], + "lines": [ + { + "bbox": [ + 130, + 199, + 482, + 232 + ], + "spans": [ + { + "bbox": [ + 130, + 199, + 482, + 232 + ], + "type": "text", + "content": "Fig. 7: Ablations. We report Dice after five steps of simulated interactions following three inference-time interaction procedures. Error bars show " + }, + { + "bbox": [ + 130, + 199, + 482, + 232 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 130, + 199, + 482, + 232 + ], + "type": "text", + "content": " CI. (a) shows mean change in Dice relative to ScribblePrompt-UNet trained without any synthetic labels." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 255, + 246, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 255, + 246, + 266 + ], + "spans": [ + { + "bbox": [ + 131, + 255, + 246, + 266 + ], + "type": "text", + "content": "6.1 Synthetic Labels" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 272, + 482, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 272, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 482, + 344 + ], + "type": "text", + "content": "Fig. 7a shows the effect of varying the probability of sampling a synthetic label during training for ScribblePrompt-UNet. Training with both real and synthetic labels improves generalization to new datasets, compared to training with only real labels. Using " + }, + { + "bbox": [ + 130, + 272, + 482, + 344 + ], + "type": "inline_equation", + "content": "p_{\\text{synth}} = 0.5" + }, + { + "bbox": [ + 130, + 272, + 482, + 344 + ], + "type": "text", + "content": " results in the highest Dice. Training with only synthetic labels results in worse Dice scores. We show similar results for ScribblePrompt-SAM and other interactions in Supplementary Material Sec. I.1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 360, + 235, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 360, + 235, + 372 + ], + "spans": [ + { + "bbox": [ + 131, + 360, + 235, + 372 + ], + "type": "text", + "content": "6.2 Prompt Types" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 377, + 481, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 377, + 481, + 401 + ], + "spans": [ + { + "bbox": [ + 130, + 377, + 481, + 401 + ], + "type": "text", + "content": "Setup. We evaluate ScribblePrompt-UNet models trained with different combinations of prompts, compared to the complete ScribblePrompt-UNet:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 406, + 479, + 452 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 131, + 406, + 450, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 406, + 450, + 417 + ], + "spans": [ + { + "bbox": [ + 131, + 406, + 450, + 417 + ], + "type": "text", + "content": "- ScribblePrompt-Unet (scribbles) trained on boxes and scribbles." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 418, + 419, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 418, + 419, + 429 + ], + "spans": [ + { + "bbox": [ + 131, + 418, + 419, + 429 + ], + "type": "text", + "content": "- ScribblePrompt-UNet (clicks) trained on boxes and clicks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 430, + 479, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 430, + 479, + 452 + ], + "spans": [ + { + "bbox": [ + 131, + 430, + 479, + 452 + ], + "type": "text", + "content": "- ScribblePrompt-Unet (random clicks) trained on boxes and random clicks." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 458, + 482, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 458, + 482, + 507 + ], + "spans": [ + { + "bbox": [ + 130, + 458, + 482, + 507 + ], + "type": "text", + "content": "Results. Fig. 7b shows ScribblePrompt-UNet trained with scribbles, clicks and bounding boxes predicts segmentations more accurately than do ablated versions of ScribblePrompt-UNet. We show results for other interaction procedures with similar trends in Supplementary Material Sec. I.2." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 131, + 523, + 220, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 523, + 220, + 536 + ], + "spans": [ + { + "bbox": [ + 131, + 523, + 220, + 536 + ], + "type": "text", + "content": "7 Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": "We present ScribblePrompt, a practical framework for interactive segmentation that enables users to segment diverse medical images with scribbles, clicks, and bounding boxes. We introduce methods for simulating realistic user interactions and generating synthetic labels. These methods enable us to train models that generalize to unseen segmentation tasks and datasets. ScribblePrompt is more accurate than existing baselines, and ScribblePrompt-UNet is computationally efficient, even on a CPU. Our user study shows that nearly all users prefer ScribblePrompt and achieve segmentations with " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": " higher Dice with less effort than the next most accurate baseline. ScribblePrompt promises to significantly reduce the burden of manual segmentation in biomedical imaging." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "type": "text", + "content": "H.E. Wong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 224 + ], + "type": "text", + "content": "This work was supported in part by funding from Quanta Computer Inc., the Eric and Wendy Schmidt Center at the Broad Institute of MIT and Harvard and the Wistron Corporation. Research reported in this paper was supported by the National Institute of Biomedical Imaging and Bioengineering of the National Institutes of Health under award number R01EB033773. Much of the computation resources required for this research was performed on computational hardware generously provided by the Massachusetts Life Sciences Center." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 239, + 197, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 239, + 197, + 251 + ], + "spans": [ + { + "bbox": [ + 133, + 239, + 197, + 251 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 141, + 262, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 141, + 262, + 479, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 262, + 479, + 284 + ], + "spans": [ + { + "bbox": [ + 141, + 262, + 479, + 284 + ], + "type": "text", + "content": "1. Abdi, A.H., Kasaei, S., Mehdizadeh, M.: Automatic segmentation of mandible in panoramic x-ray. Journal of Medical Imaging 2(4), 044003 (2015)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 285, + 480, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 285, + 480, + 338 + ], + "spans": [ + { + "bbox": [ + 141, + 285, + 480, + 338 + ], + "type": "text", + "content": "2. Agustsson, E., Uijlings, J.R., Ferrari, V.: Interactive Full Image Segmentation by Considering All Regions Jointly. In: 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11614-11623. IEEE, Long Beach, CA, USA (Jun 2019). https://doi.org/10.1109/CVPR.2019.01189, https://ieeexplore.ieee.org/document/8953526/" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 339, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 339, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 141, + 339, + 480, + 413 + ], + "type": "text", + "content": "3. Aine, C.J., Bockholt, H.J., Bustillo, J.R., Canive, J.M., Caprihan, A., Gasparovic, C., Hanlon, F.M., Houck, J.M., Jung, R.E., Lauriello, J., Liu, J., Mayer, A.R., Perrone-Bizzozero, N.I., Posse, S., Stephen, J.M., Turner, J.A., Clark, V.P., Calhoun, V.D.: Multimodal Neuroimaging in Schizophrenia: Description and Dissemination. Neuroinformatics 15(4), 343-364 (Oct 2017). https://doi.org/10.1007/s12021-017-9338-9, http://link.springer.com/10.1007/s12021-017-9338-9" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 415, + 480, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 415, + 480, + 458 + ], + "spans": [ + { + "bbox": [ + 141, + 415, + 480, + 458 + ], + "type": "text", + "content": "4. Al-Dhabyani, W., Gomaa, M., Khaled, H., Fahmy, A.: Dataset of breast ultrasound images. Data in Brief 28, 104863 (2020). https://doi.org/https://doi.org/10.1016/j.dib.2019.104863, https://www.sciencedirect.com/science/article/pii/S2352340919312181" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 459, + 480, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 459, + 480, + 502 + ], + "spans": [ + { + "bbox": [ + 141, + 459, + 480, + 502 + ], + "type": "text", + "content": "5. Asad, M., Fidon, L., Vercauteren, T.: ECONet: Efficient Convolutional Online Likelihood Network for Scribble-based Interactive Segmentation. In: International Conference on Medical Imaging with Deep Learning. pp. 35-47. PMLR (2022), http://arxiv.org/abs/2201.04584, arXiv:2201.04584 [cs, eess]" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 502, + 480, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 502, + 480, + 545 + ], + "spans": [ + { + "bbox": [ + 141, + 502, + 480, + 545 + ], + "type": "text", + "content": "6. Atzeni, A., Peter, L., Robinson, E., Blackburn, E., Althonayan, J., Alexander, D.C., Iglesias, J.E.: Deep active learning for suggestive segmentation of biomedical image stacks via optimisation of Dice scores and traced boundary length. Medical Image Analysis 81 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 545, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 545, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 141, + 545, + 480, + 588 + ], + "type": "text", + "content": "7. Bai, J., Wu, X.: Error-Tolerant Scribbles Based Interactive Image Segmentation. In: 2014 IEEE Conference on Computer Vision and Pattern Recognition. pp. 392-399. IEEE, Columbus, OH, USA (Jun 2014). https://doi.org/10.1109/CVPR.2014.57, https://ieeexplore.ieee.org/document/6909451" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 589, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 589, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 141, + 589, + 480, + 632 + ], + "type": "text", + "content": "8. Baid, U., Ghodasara, S., Mohan, S., Bilello, M., Calabrese, E., Colak, E., Farahani, K., Kalpathy-Cramer, J., Kitamura, F.C., Pati, S., et al.: The rsna-asnrmiccai brats 2021 benchmark on brain tumor segmentation and radiogenomic classification. arXiv preprint arXiv:2107.02314 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 141, + 632, + 480, + 665 + ], + "type": "text", + "content": "9. Benenson, R., Popov, S., Ferrari, V.: Large-scale interactive object segmentation with human annotators. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11700-11709 (2019)" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "type": "text", + "content": "ScribblePrompt" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 138, + 116, + 481, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 116, + 481, + 161 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 481, + 161 + ], + "type": "text", + "content": "10. Bernard, O., Lalande, A., Zotti, C., Cervenansky, F., Yang, X., Heng, P.A., Cetin, I., Lekadir, K., Camara, O., Ballester, M.A.G., et al.: Deep learning techniques for automatic mri cardiac multi-structures segmentation and diagnosis: is the problem solved? IEEE transactions on medical imaging 37(11), 2514-2525 (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 162, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 162, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 138, + 162, + 481, + 194 + ], + "type": "text", + "content": "11. Bilic, P., Christ, P.F., Vorontsov, E., Chlebus, G., Chen, H., Dou, Q., Fu, C.W., Han, X., Heng, P.A., Hesser, J., et al.: The liver tumor segmentation benchmark (lits). arXiv preprint arXiv:1901.04056 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 194, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 194, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 138, + 194, + 481, + 237 + ], + "type": "text", + "content": "12. Bloch, N., Madabhushi, A., Huisman, H., Freymann, J., Kirby, J., Grauer, M., Enquobahrie, A., Jaffe, C., Clarke, L., Farahani, K.: Nci-isbi 2013 challenge: automated segmentation of prostate structures. The Cancer Imaging Archive 370(6), 5 (2015)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 238, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 238, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 138, + 238, + 481, + 293 + ], + "type": "text", + "content": "13. Boykov, Y., Jolly, M.P.: Interactive graph cuts for optimal boundary & region segmentation of objects in N-D images. In: Proceedings Eighth IEEE International Conference on Computer Vision. ICCV 2001. vol. 1, pp. 105-112. IEEE Comput. Soc, Vancouver, BC, Canada (2001). https://doi.org/10.1109/ICCV.2001.937505, http://ieeexplore.ieee.org/document/937505/" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 293, + 481, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 293, + 481, + 336 + ], + "spans": [ + { + "bbox": [ + 138, + 293, + 481, + 336 + ], + "type": "text", + "content": "14. Bredell, G., Tanner, C., Konukoglu, E.: Iterative interaction training for segmentation editing networks. In: Machine Learning in Medical Imaging: 9th International Workshop, MLMI 2018, Held in Conjunction with MICCAI 2018, Granada, Spain, September 16, 2018, Proceedings 9. pp. 363-370. Springer (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 139, + 337, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 337, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 139, + 337, + 481, + 357 + ], + "type": "text", + "content": "15. Butoi*, V.I., Ortiz*, J.J.G., Ma, T., Sabuncu, M.R., Guttag, J., Dalca, A.V.: Universeg: Universal medical image segmentation. In: ICCV (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 358, + 481, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 358, + 481, + 423 + ], + "spans": [ + { + "bbox": [ + 138, + 358, + 481, + 423 + ], + "type": "text", + "content": "16. Caicedo, J.C., Goodman, A., Karhohs, K.W., Cimini, B.A., Ackerman, J., Haghighi, M., Heng, C., Becker, T., Doan, M., McQuin, C., Rohhan, M., Singh, S., Carpenter, A.E.: Nucleus segmentation across imaging experiments: the 2018 Data Science Bowl. Nature Methods 16(12), 1247-1253 (Dec 2019). https://doi.org/10.1038/s41592-019-0612-7, https://doi.org/10.1038/s41592-019-0612-7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 424, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 424, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 424, + 481, + 468 + ], + "type": "text", + "content": "17. Cardona, A., Saalfeld, S., Preibisch, S., Schmid, B., Cheng, A., Pulokas, J., Tomancak, P., Hartenstein, V.: An integrated micro-and macroarchitectural analysis of the drosophila brain by computer-assisted serial section electron microscopy. PLoS biology 8(10), e1000502 (2010)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 468, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 468, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 138, + 468, + 481, + 501 + ], + "type": "text", + "content": "18. Chen, X., Cheung, Y.S.J., Lim, S.N., Zhao, H.: ScribbleSeg: Scribble-based Interactive Image Segmentation (Mar 2023), http://arxiv.org/abs/2303.11320, arXiv:2303.11320 [cs]" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 502, + 481, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 502, + 481, + 534 + ], + "spans": [ + { + "bbox": [ + 138, + 502, + 481, + 534 + ], + "type": "text", + "content": "19. Cheng, J., Ye, J., Deng, Z., Chen, J., Li, T., Wang, H., Su, Y., Huang, Z., Chen, J., Jiang, L., Sun, H., He, J., Zhang, S., Zhu, M., Qiao, Y.: SAM-Med2D (Aug 2023), http://arxiv.org/abs/2308.16184, arXiv:2308.16184 [cs]" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 534, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 534, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 138, + 534, + 481, + 599 + ], + "type": "text", + "content": "20. Codella, N.C.F., Gutman, D.A., Celebi, M.E., Helba, B., Marchetti, M.A., Dusza, S.W., Kalloo, A., Liopyris, K., Mishra, N.K., Kittler, H., Halpern, A.: Skin lesion analysis toward melanoma detection: A challenge at the 2017 international symposium on biomedical imaging (isbi), hosted by the international skin imaging collaboration (ISIC). CoRR abs/1710.05006 (2017), http://arxiv.org/abs/1710.05006" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 600, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 600, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 600, + 481, + 665 + ], + "type": "text", + "content": "21. Criminisi, A., Sharp, T., Blake, A.: GeoS: Geodesic Image Segmentation. In: Forsyth, D., Torr, P., Zisserman, A. (eds.) Computer Vision - ECCV 2008, vol. 5302, pp. 99-112. Springer Berlin Heidelberg, Berlin, Heidelberg (2008). https://doi.org/10.1007/978-3-540-88682-2_9, http://link.springer.com/10.1007/978-3-540-88682-2_9, series Title: Lecture Notes in Computer Science" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 235, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 235, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 235, + 102 + ], + "type": "text", + "content": "H.E. Wong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 480, + 149 + ], + "type": "text", + "content": "22. Czolbe, S., Dalca, A.V.: Neuralizer: General neuroimage analysis without retraining. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6217-6230 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 150, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 150, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 138, + 150, + 480, + 183 + ], + "type": "text", + "content": "23. Dalca, A.V., Guttag, J., Sabuncu, M.R.: Anatomical priors in convolutional networks for unsupervised biomedical segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 9290-9299 (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 183, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 183, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 138, + 183, + 480, + 205 + ], + "type": "text", + "content": "24. Dice, L.R.: Measures of the amount of ecologic association between species. Ecology 26(3), 297-302 (1945)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 205, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 205, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 138, + 205, + 480, + 226 + ], + "type": "text", + "content": "25. Felzenszwalb, P.F., Huttenlocher, D.P.: Efficient graph-based image segmentation. International journal of computer vision 59, 167-181 (2004)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 227, + 386, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 227, + 386, + 238 + ], + "spans": [ + { + "bbox": [ + 138, + 227, + 386, + 238 + ], + "type": "text", + "content": "26. Fischl, B.: Freesurfer. Neuroimage 62(2), 774-781 (2012)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 238, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 238, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 138, + 238, + 480, + 270 + ], + "type": "text", + "content": "27. Forte, M., Price, B., Cohen, S., Xu, N., Pitie, F.: Getting to " + }, + { + "bbox": [ + 138, + 238, + 480, + 270 + ], + "type": "inline_equation", + "content": "99\\%" + }, + { + "bbox": [ + 138, + 238, + 480, + 270 + ], + "type": "text", + "content": " Accuracy in Interactive Segmentation (Mar 2020), http://arxiv.org/abs/2003.07932, arXiv:2003.07932 [cs]" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 271, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 271, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 138, + 271, + 480, + 303 + ], + "type": "text", + "content": "28. Gamper, J., Koohbanani, N., Benes, K., Graham, S., Jahanifar, M., Khurram, S., Azam, A., Hewitt, K., Rajpoot, N.: Pannuke dataset extension, insights and baselines. arxiv. 2020 doi: 10.48550. ARXIV (2003)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 304, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 304, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 138, + 304, + 480, + 346 + ], + "type": "text", + "content": "29. Gerhard, S., Funke, J., Martel, J., Cardona, A., Fetter, R.: Segmented anisotropic ssTEM dataset of neural tissue (11 2013). https://doi.org/10.6084/m9.figshare.856713.v1, https://figshare.com/articles/dataset/Segmented_anisotropic_ssTEM_dataset_of_neural_tissue/856713" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 347, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 347, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 138, + 347, + 480, + 402 + ], + "type": "text", + "content": "30. van Ginneken, B., Stegmann, M.B., Loog, M.: Segmentation of anatomical structures in chest radiographs using supervised methods: a comparative study on a public database. Medical Image Analysis 10(1), 19-40 (2006). https://doi.org/https://doi.org/10.1016/j.media.2005.02.002, https://www.sciencedirect.com/science/article/pii/S1361841505000368" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 403, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 403, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 138, + 403, + 480, + 445 + ], + "type": "text", + "content": "31. Gollub, R.L., Shoemaker, J.M., King, M.D., White, T., Ehrlich, S., Sponheim, S.R., Clark, V.P., Turner, J.A., Mueller, B.A., Magnotta, V., et al.: The mcic collection: a shared repository of multi-modal, multi-site brain image data from a clinical investigation of schizophrenia. Neuroinformatics 11, 367-388 (2013)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 447, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 447, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 138, + 447, + 480, + 479 + ], + "type": "text", + "content": "32. Gotkowski, K., Lüth, C., Jäger, P.F., Ziegler, S., Krämer, L., Denner, S., Xiao, S., Disch, N., Maier-Hein, K.H., Isensee, F.: Embarrassingly simple scribble supervision for 3d medical segmentation. arXiv preprint arXiv:2403.12834 (2024)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 479, + 480, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 479, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 138, + 479, + 480, + 521 + ], + "type": "text", + "content": "33. Grady, L.: Random Walks for Image Segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence 28(11), 1768-1783 (Nov 2006). https://doi.org/10.1109/TPAMI.2006.233, http://ieeexplore.ieee.org/document/1704833/" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 522, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 522, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 522, + 480, + 555 + ], + "type": "text", + "content": "34. Gut, D.: X-ray images of the hip joints 1 (Jul 2021). https://doi.org/10.17632/zm6bxzhmfz.1, https://data.mendeley.com/datasets/zm6bxzhmfz/1, publisher: Mendeley Data" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 555, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 480, + 588 + ], + "type": "text", + "content": "35. He, K., Zhang, X., Ren, S., Sun, J.: Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In: Proceedings of the IEEE international conference on computer vision. pp. 1026-1034 (2015)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 589, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 589, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 138, + 589, + 480, + 621 + ], + "type": "text", + "content": "36. He, S., Bao, R., Li, J., Stout, J., Bjornerud, A., Grant, P.E., Ou, Y.: Computer- vision benchmark segment-anything model (sam) in medical images: Accuracy in 12 datasets (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 621, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 480, + 665 + ], + "type": "text", + "content": "37. Heller, N., Isensee, F., Maier-Hein, K.H., Hou, X., Xie, C., Li, F., Nan, Y., Mu, G., Lin, Z., Han, M., et al.: The state of the art in kidney and kidney tumor segmentation in contrast-enhanced ct imaging: Results of the kits19 challenge. Medical Image Analysis p. 101821 (2020)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "type": "text", + "content": "ScribblePrompt" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 117, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 138, + 117, + 479, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 117, + 479, + 160 + ], + "spans": [ + { + "bbox": [ + 138, + 117, + 479, + 160 + ], + "type": "text", + "content": "38. Hernandez Petzsche, M.R., de la Rosa, E., Hanning, U., Wiest, R., Valenzuela, W., Reyes, M., Meyer, M., Liew, S.L., Kofler, F., Ezhov, I., et al.: Isles 2022: A multi-center magnetic resonance imaging stroke lesion segmentation dataset. Scientific data 9(1), 762 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 160, + 480, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 160, + 480, + 192 + ], + "spans": [ + { + "bbox": [ + 138, + 160, + 480, + 192 + ], + "type": "text", + "content": "39. Hoover, A., Kouznetsova, V., Goldbaum, M.: Locating blood vessels in retinal images by piecewise threshold probing of a matched filter response. IEEE Transactions on Medical imaging 19(3), 203-210 (2000)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 192, + 480, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 192, + 480, + 213 + ], + "spans": [ + { + "bbox": [ + 138, + 192, + 480, + 213 + ], + "type": "text", + "content": "40. Hu, X., Xu, X., Shi, Y.: How to efficiently adapt large segmentation model(sam) to medical images (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 213, + 480, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 213, + 480, + 246 + ], + "spans": [ + { + "bbox": [ + 138, + 213, + 480, + 246 + ], + "type": "text", + "content": "41. Huang, Y., Yang, X., Liu, L., Zhou, H., Chang, A., Zhou, X., Chen, R., Yu, J., Chen, J., Chen, C., Chi, H., Hu, X., Fan, D.P., Dong, F., Ni, D.: Segment anything model for medical images? (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 246, + 480, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 246, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 138, + 246, + 480, + 277 + ], + "type": "text", + "content": "42. Huttenlocher, D.P., Klanderman, G.A., Rucklidge, W.J.: Comparing images using the hausdorff distance. IEEE Transactions on pattern analysis and machine intelligence 15(9), 850-863 (1993)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 277, + 480, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 277, + 480, + 320 + ], + "spans": [ + { + "bbox": [ + 138, + 277, + 480, + 320 + ], + "type": "text", + "content": "43. Isensee, F., Jaeger, P.F., Kohl, S.A.A., Petersen, J., Maier-Hein, K.H.: nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nature Methods 18(2), 203-211 (Feb 2021). https://doi.org/10.1038/s41592-020-01008-z, http://www.nature.com/articles/s41592-020-01008-z" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 320, + 480, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 320, + 480, + 353 + ], + "spans": [ + { + "bbox": [ + 138, + 320, + 480, + 353 + ], + "type": "text", + "content": "44. Ji, Y., Bai, H., Yang, J., Ge, C., Zhu, Y., Zhang, R., Li, Z., Zhang, L., Ma, W., Wan, X., et al.: Amos: A large-scale abdominal multi-organ benchmark for versatile medical image segmentation. arXiv preprint arXiv:2206.08023 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 353, + 480, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 353, + 480, + 439 + ], + "spans": [ + { + "bbox": [ + 138, + 353, + 480, + 439 + ], + "type": "text", + "content": "45. Kavur, A.E., Gezer, N.S., Baris, M., Aslan, S., Conze, P.H., Groza, V., Pham, D.D., Chatterjee, S., Ernst, P., Ozkan, S., Baydar, B., Lachinov, D., Han, S., Pauli, J., Isensee, F., Perkonigg, M., Sathish, R., Rajan, R., Sheet, D., Dovletov, G., Speck, O., Nurnberger, A., Maier-Hein, K.H., Bozdagi Akar, G., Ünal, G., Dicle, O., Selver, M.A.: CHAOS Challenge - combined (CT-MR) healthy abdominal organ segmentation. Medical Image Analysis 69, 101950 (2021). https://doi.org/https://doi.org/10.1016/j.media.2020.101950, https://www.sciencedirect.com/science/article/pii/S1361841520303145" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 439, + 480, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 439, + 480, + 461 + ], + "spans": [ + { + "bbox": [ + 138, + 439, + 480, + 461 + ], + "type": "text", + "content": "46. Kim, S., Oh, H.J., Min, S., Jeong, W.K.: Evaluation and improvement of segment anything model for interactive histopathology image segmentation (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 461, + 480, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 461, + 480, + 492 + ], + "spans": [ + { + "bbox": [ + 138, + 461, + 480, + 492 + ], + "type": "text", + "content": "47. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. In: ICCV (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 492, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 492, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 138, + 492, + 480, + 536 + ], + "type": "text", + "content": "48. Krönke, M., Eilers, C., Dimova, D., Köhler, M., Buschner, G., Schweiger, L., Konstantinidou, L., Makowski, M., Nagarajah, J., Navab, N., et al.: Tracked 3d ultrasound and deep neural network-based thyroid segmentation reduce interobserver variability in thyroid volumetry. Plos one 17(7), e0268550 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 536, + 480, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 536, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 138, + 536, + 480, + 590 + ], + "type": "text", + "content": "49. Kuijf, H.J., Biesbroek, J.M., De Bresser, J., Heinen, R., Andermatt, S., Bento, M., Berseth, M., Belyaev, M., Cardoso, M.J., Casamitjana, A., et al.: Standardized assessment of automatic segmentation of white matter hyperintensities and results of the wmh segmentation challenge. IEEE transactions on medical imaging 38(11), 2556-2568 (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 590, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 590, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 590, + 480, + 632 + ], + "type": "text", + "content": "50. Kuklisova-Murgasova, M., Aljabar, P., Srinivasan, L., Counsell, S.J., Doria, V., Serag, A., Gousias, I.S., Boardman, J.P., Rutherford, M.A., Edwards, A.D., et al.: A dynamic 4d probabilistic atlas of the developing brain. NeuroImage 54(4), 2750-2763 (2011)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 632, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 480, + 666 + ], + "type": "text", + "content": "51. Lambert, Z., Petitjean, C., Dubray, B., Kuan, S.: Segthor: segmentation of thoracic organs at risk in ct images. In: 2020 Tenth International Conference on Image Processing Theory, Tools and Applications (IPTA). pp. 1-6. IEEE (2020)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 235, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 235, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 235, + 102 + ], + "type": "text", + "content": "H.E. Wong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 116, + 480, + 161 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 480, + 161 + ], + "type": "text", + "content": "52. Landman, B., Xu, Z., Igelsias, J., Styner, M., Langerak, T., Klein, A.: Miccai multi-atlas labeling beyond the cranial vault-workshop and challenge. In: Proc. MICCAI Multi-Atlas Labeling Beyond Cranial Vault Workshop Challenge. vol. 5, p. 12 (2015)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 162, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 162, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 138, + 162, + 480, + 205 + ], + "type": "text", + "content": "53. Lee, H., Jeong, W.K.: Scribble2label: Scribble-supervised cell segmentation via self-generating pseudo-labels with consistency. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part I 23. pp. 14-23. Springer (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 206, + 480, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 206, + 480, + 250 + ], + "spans": [ + { + "bbox": [ + 138, + 206, + 480, + 250 + ], + "type": "text", + "content": "54. Lemaître, G., Martí, R., Freixenet, J., Vilanova, J.C., Walker, P.M., Meriaudeau, F.: Computer-aided detection and diagnosis for prostate cancer based on mono and multi-parametric mri: a review. Computers in biology and medicine 60, 8-31 (2015)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 251, + 480, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 251, + 480, + 283 + ], + "spans": [ + { + "bbox": [ + 138, + 251, + 480, + 283 + ], + "type": "text", + "content": "55. Li, M., Zhang, Y., Ji, Z., Xie, K., Yuan, S., Liu, Q., Chen, Q.: Ipn-v2 and octa-500: Methodology and dataset for retinal image segmentation. arXiv preprint arXiv:2012.07261 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 285, + 480, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 285, + 480, + 318 + ], + "spans": [ + { + "bbox": [ + 138, + 285, + 480, + 318 + ], + "type": "text", + "content": "56. Li, Z., Zheng, Y., Luo, X., Shan, D., Hong, Q.: Scribblevc: Scribble-supervised medical image segmentation with vision-class embedding. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 3384-3393 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 318, + 480, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 318, + 480, + 351 + ], + "spans": [ + { + "bbox": [ + 138, + 318, + 480, + 351 + ], + "type": "text", + "content": "57. Lin, D., Dai, J., Jia, J., He, K., Sun, J.: Scribblesup: Scribble-supervised convolutional networks for semantic segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3159-3167 (2016)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 352, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 352, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 138, + 352, + 480, + 385 + ], + "type": "text", + "content": "58. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proceedings of the IEEE international conference on computer vision. pp. 2980-2988 (2017)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 386, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 386, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 138, + 386, + 480, + 418 + ], + "type": "text", + "content": "59. Lin, X., Xiang, Y., Zhang, L., Yang, X., Yan, Z., Yu, L.: Samus: Adapting segment anything model for clinically-friendly and generalizable ultrasound image segmentation (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 419, + 480, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 419, + 480, + 463 + ], + "spans": [ + { + "bbox": [ + 138, + 419, + 480, + 463 + ], + "type": "text", + "content": "60. Litjens, G., Toth, R., van de Ven, W., Hoeks, C., Kerkstra, S., van Ginneken, B., Vincent, G., Guillard, G., Birbeck, N., Zhang, J., et al.: Evaluation of prostate segmentation algorithms for mri: the promise12 challenge. Medical image analysis 18(2), 359-373 (2014)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 464, + 480, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 464, + 480, + 497 + ], + "spans": [ + { + "bbox": [ + 138, + 464, + 480, + 497 + ], + "type": "text", + "content": "61. Liu, Q., Xu, Z., Bertasius, G., Niethammer, M.: Simpleclick: Interactive image segmentation with simple vision transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22290-22300 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 498, + 480, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 498, + 480, + 531 + ], + "spans": [ + { + "bbox": [ + 138, + 498, + 480, + 531 + ], + "type": "text", + "content": "62. Liu, Q., Xu, Z., Jiao, Y., Niethammer, M.: iSegFormer: Interactive Segmentation via Transformers with Application to 3D Knee MR Images. In: Medical Image Computing and Computer Assisted Intervention (MICCAI) (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 532, + 480, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 532, + 480, + 553 + ], + "spans": [ + { + "bbox": [ + 138, + 532, + 480, + 553 + ], + "type": "text", + "content": "63. Ljosa, V., Sokolnicki, K.L., Carpenter, A.E.: Annotated high-throughput microscopy image sets for validation. Nature methods 9(7), 637-637 (2012)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 554, + 480, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 554, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 138, + 554, + 480, + 597 + ], + "type": "text", + "content": "64. Löffler, M.T., Sekuboyina, A., Jacob, A., Grau, A.L., Scharr, A., El Husseinini, M., Kallweit, M., Zimmer, C., Baum, T., Kirschke, J.S.: A vertebral segmentation dataset with fracture grading. Radiology: Artificial Intelligence 2(4), e190138 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 599, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 599, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 138, + 599, + 480, + 620 + ], + "type": "text", + "content": "65. in the Loop, H.: Teeth segmentation dataset, https://humansintheloop.org/resources/datasets/teeth-segmentation-dataset/" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 621, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 480, + 665 + ], + "type": "text", + "content": "66. Luo, X., Hu, M., Liao, W., Zhai, S., Song, T., Wang, G., Zhang, S.: Scribblesupervised medical image segmentation via dual-branch network and dynamically mixed pseudo labels supervision. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 528-538. Springer (2022)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "type": "text", + "content": "ScribblePrompt" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + }, + { + "para_blocks": [ + { + "bbox": [ + 136, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 136, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 136, + 116, + 481, + 149 + ], + "type": "text", + "content": "67. Luo, X., Liao, W., Xiao, J., Song, T., Zhang, X., Li, K., Wang, G., Zhang, S.: Word: Revisiting organs segmentation in the whole abdominal region. arXiv preprint arXiv:2111.02403 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 151, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 151, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 138, + 151, + 481, + 194 + ], + "type": "text", + "content": "68. Luo, X., Wang, G., Song, T., Zhang, J., Aertsen, M., Deprest, J., Ourselin, S., Vercauteren, T., Zhang, S.: MIDeepSeg: Minimally Interactive Segmentation of Unseen Objects from Medical Images Using Deep Learning. Medical Image Analysis 72, 102102 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 195, + 481, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 195, + 481, + 217 + ], + "spans": [ + { + "bbox": [ + 138, + 195, + 481, + 217 + ], + "type": "text", + "content": "69. Ma, J., He, Y., Li, F., Han, L., You, C., Wang, B.: Segment anything in medical images. Nature Communications 15, 1-9 (2024)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 218, + 481, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 218, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 138, + 218, + 481, + 262 + ], + "type": "text", + "content": "70. Ma, Y., Hao, H., Xie, J., Fu, H., Zhang, J., Yang, J., Wang, Z., Liu, J., Zheng, Y., Zhao, Y.: Rose: a retinal oct-angiography vessel segmentation dataset and new model. IEEE Transactions on Medical Imaging 40(3), 928–939 (2021). https://doi.org/10.1109/TMI.2020.3042802" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 263, + 481, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 263, + 481, + 296 + ], + "spans": [ + { + "bbox": [ + 138, + 263, + 481, + 296 + ], + "type": "text", + "content": "71. Macdonald, J.A., Zhu, Z., Konkel, B., Mazurowski, M., Wiggins, W., Bashir, M.: Duke liver dataset (MRI) v2 (Apr 2023). https://doi.org/10.5281/zenodo.7774566, https://doi.org/10.5281/zenodo.7774566" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 297, + 481, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 297, + 481, + 319 + ], + "spans": [ + { + "bbox": [ + 138, + 297, + 481, + 319 + ], + "type": "text", + "content": "72. Maninis, K.K., Caelles, S., Pont-Tuset, J., Van Gool, L.: Deep Extreme Cut: From Extreme Points to Object Segmentation (2018), arXiv:1711.09081" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 319, + 481, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 319, + 481, + 363 + ], + "spans": [ + { + "bbox": [ + 138, + 319, + 481, + 363 + ], + "type": "text", + "content": "73. Marcus, D.S., Wang, T.H., Parker, J., Csernansky, J.G., Morris, J.C., Buckner, R.L.: Open access series of imaging studies (oasis): cross-sectional mri data in young, middle aged, nondemented, and demented older adults. Journal of cognitive neuroscience 19(9), 1498-1507 (2007)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 364, + 481, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 364, + 481, + 397 + ], + "spans": [ + { + "bbox": [ + 138, + 364, + 481, + 397 + ], + "type": "text", + "content": "74. Marek, K., Jennings, D., Lasch, S., Siderowf, A., Tanner, C., Simuni, T., Coffey, C., Kieburtz, K., Flagg, E., Chowdhury, S., et al.: The parkinson progression marker initiative (ppmi). Progress in neurobiology 95(4), 629-635 (2011)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 398, + 481, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 398, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 138, + 398, + 481, + 453 + ], + "type": "text", + "content": "75. Mazurowski, M.A., Clark, K., Czarnek, N.M., Shamsesfandabadi, P., Peters, K.B., Saha, A.: Radiogenomics of lower-grade glioma: algorithmically-assessed tumor shape is associated with tumor genomic subtypes and patient outcomes in a multi-institutional study with the cancer genome atlas data. Journal of neuro-oncology 133, 27-35 (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 453, + 481, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 453, + 481, + 507 + ], + "spans": [ + { + "bbox": [ + 138, + 453, + 481, + 507 + ], + "type": "text", + "content": "76. Mazurowski, M.A., Dong, H., Gu, H., Yang, J., Konz, N., Zhang, Y.: Segment anything model for medical image analysis: An experimental study. Medical Image Analysis 89, 102918 (Oct 2023). https://doi.org/10.1016/j.media.2023.102918, https://www.sciencedirect.com/science/article/pii/S1361841523001780" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 509, + 481, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 509, + 481, + 553 + ], + "spans": [ + { + "bbox": [ + 138, + 509, + 481, + 553 + ], + "type": "text", + "content": "77. Men, J., Huang, Y., Solanki, J., Zeng, X., Alex, A., Jerwick, J., Zhang, Z., Tanzi, R.E., Li, A., Zhou, C.: Optical coherence tomography for brain imaging and developmental biology. IEEE Journal of Selected Topics in Quantum Electronics 22(4), 120-132 (2015)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 554, + 481, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 554, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 138, + 554, + 481, + 586 + ], + "type": "text", + "content": "78. Milletari, F., Navab, N., Ahmadi, S.A.: V-net: Fully convolutional neural networks for volumetric medical image segmentation. Fourth International Conference on 3D Vision (3DV) pp. 565-571 (2016)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 587, + 481, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 587, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 138, + 587, + 481, + 620 + ], + "type": "text", + "content": "79. Montoya, A., Hasnin, kaggle446, shirzad, Cukierski, W., yffud: Ultrasound nerve segmentation (2016), https://kaggle.com/competitions/ultrasound-nerve-segmentation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 621, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 481, + 665 + ], + "type": "text", + "content": "80. Ooi, A.Z.H., Embong, Z., Abd Hamid, A.I., Zainon, R., Wang, S.L., Ng, T.F., Hamzah, R.A., Teoh, S.S., Ibrahim, H.: Interactive Blood Vessel Segmentation from Retinal Fundus Image Based on Canny Edge Detector. Sensors 21(19), 6380 (Sep 2021)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "20" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 235, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 235, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 235, + 102 + ], + "type": "text", + "content": "H.E. Wong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 19 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 480, + 149 + ], + "type": "text", + "content": "81. Paranjape, J.N., Nair, N.G., Sikder, S., Vedula, S.S., Patel, V.M.: Adaptivesam: Towards efficient tuning of sam for surgical scene segmentation. arXiv preprint arXiv:2308.03726 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 150, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 150, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 138, + 150, + 480, + 215 + ], + "type": "text", + "content": "82. Pedraza, L., Vargas, C., Narvaez, F., Durán, O., Muñoz, E., Romero, E.: An open access thyroid ultrasound image database. In: Romero, E., Lepore, N. (eds.) 10th international symposium on medical information processing and analysis. vol. 9287, p. 92870W. SPIE / International Society for Optics and Photonics (2015). https://doi.org/10.1117/12.2073532, https://doi.org/10.1117/12.2073532" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 216, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 216, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 138, + 216, + 480, + 258 + ], + "type": "text", + "content": "83. Philbrick, K.A., Weston, A.D., Akkus, Z., Kline, T.L., Korfiatis, P., Sakinis, T., Kostandy, P., Boonrod, A., Zeinoddini, A., Takahashi, N., et al.: Ril-contour: a medical imaging dataset annotation tool for and with deep learning. Journal of digital imaging 32, 571-581 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 259, + 480, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 259, + 480, + 315 + ], + "spans": [ + { + "bbox": [ + 138, + 259, + 480, + 315 + ], + "type": "text", + "content": "84. Podobnik, G., Strojan, P., Peterlin, P., Ibragimov, B., Vrtovec, T.: HaN-Seg: The head and neck organ-at-risk CT and MR segmentation dataset. Medical Physics 50(3), 1917-1927 (2023). https://doi.org/https://doi.org/10.1002/mp.16197, https://aapm.onlinelibrary.wiley.com/doi/abs/10.1002/mp.16197, tex.eprint: https://aapm.onlinelibrary.wiley.com/doi/pdf/10.1002/mp.16197" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 315, + 480, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 315, + 480, + 347 + ], + "spans": [ + { + "bbox": [ + 138, + 315, + 480, + 347 + ], + "type": "text", + "content": "85. Porwal, P., Pachade, S., Kamble, R., Kokare, M., Deshmukh, G., Sahasrabuddhe, V., Meriaudeau, F.: Indian diabetic retinopathy image dataset (idrid) (2018). https://doi.org/10.21227/H25W98, https://dx.doi.org/10.21227/H25W98" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 347, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 347, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 138, + 347, + 480, + 380 + ], + "type": "text", + "content": "86. Radau, P., Lu, Y., Connelly, K., Paul, G., Dick, A., Wright, G.: Evaluation framework for algorithms segmenting short axis cardiac mri. The MIDAS Journal-Cardiac MR Left Ventricle Segmentation Challenge 49 (2009)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 380, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 380, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 138, + 380, + 480, + 424 + ], + "type": "text", + "content": "87. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 424, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 424, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 138, + 424, + 480, + 456 + ], + "type": "text", + "content": "88. Raghu, M., Zhang, C., Kleinberg, J., Bengio, S.: Transfusion: Understanding transfer learning for medical imaging. Advances in neural information processing systems 32 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 456, + 480, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 456, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 138, + 456, + 480, + 512 + ], + "type": "text", + "content": "89. Rajchl, M., Lee, M.C.H., Oktay, O., Kamnitsas, K., Passerat-Palmbach, J., Bai, W., Damodaram, M., Rutherford, M.A., Hajnal, J.V., Kainz, B., Rueckert, D.: DeepCut: Object Segmentation From Bounding Box Annotations Using Convolutional Neural Networks. IEEE Transactions on Medical Imaging 36(2), 674-683 (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 512, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 512, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 512, + 480, + 555 + ], + "type": "text", + "content": "90. Rakic, M., Wong, H.E., Ortiz, J.J.G., Cimini, B., Guttag, J.V., Dalca, A.V.: Tyche: Stochastic in-context learning for medical image segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2024)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 555, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 480, + 599 + ], + "type": "text", + "content": "91. Rister, B., Yi, D., Shivakumar, K., Nobashi, T., Rubin, D.L.: CT-ORG, a new dataset for multiple organ segmentation in computed tomography. Scientific Data 7(1), 381 (Nov 2020). https://doi.org/10.1038/s41597-020-00715-8, https://www.nature.com/articles/s41597-020-00715-8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 599, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 599, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 599, + 480, + 643 + ], + "type": "text", + "content": "92. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "93. Roth, H.R., Yang, D., Xu, Z., Wang, X., Xu, D.: Going to Extremes: Weakly Supervised Medical Image Segmentation. Machine Learning and Knowledge Ex" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "type": "text", + "content": "ScribblePrompt" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "21" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 20 + }, + { + "para_blocks": [ + { + "bbox": [ + 134, + 117, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 153, + 117, + 481, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 117, + 481, + 138 + ], + "spans": [ + { + "bbox": [ + 153, + 117, + 481, + 138 + ], + "type": "text", + "content": "traction 3(2), 507-524 (Jun 2021). https://doi.org/10.3390/make3020026, https://www.mdpi.com/2504-4990/3/2/26" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 139, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 139, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 137, + 139, + 481, + 171 + ], + "type": "text", + "content": "94. Rother, C., Kolmogorov, V., Blake, A.: \"GrabCut\" — Interactive Foreground Extraction using Iterated Graph Cuts. ACM Transactions on Graphics 23, 309-314 (2004)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 137, + 172, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 172, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 137, + 172, + 481, + 205 + ], + "type": "text", + "content": "95. Sakinis, T., Miletari, F., Roth, H., Korfiatis, P., Kostandy, P., Philbrick, K., Akkus, Z., Xu, Z., Xu, D., Erickson, B.J.: Interactive segmentation of medical images through fully convolutional neural networks (2019), arXiv:1903.08205" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 205, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 205, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 137, + 205, + 481, + 248 + ], + "type": "text", + "content": "96. Saporta, A., Gui, X., Agrawal, A., Pareek, A., Truong, S., Nguyen, C., Ngo, V.D., Seekins, J., Blankenberg, F.G., Ng, A., et al.: Deep learning saliency maps do not accurately highlight diagnostically relevant regions for medical image interpretation. MedRxiv (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 249, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 249, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 138, + 249, + 481, + 281 + ], + "type": "text", + "content": "97. Sati, P., George, I.C., Shea, C.D., Gaitan, M.I., Reich, D.S.: Flair*: a combined mr contrast technique for visualizing white matter lesions and parenchymal veins. Radiology 265(3), 926-932 (2012)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 281, + 481, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 281, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 138, + 281, + 481, + 326 + ], + "type": "text", + "content": "98. Seibold, C., Reif, S., Sarfraz, S., Fink, M.A., Mayer, V., Sellner, J., Kim, M.S., Maier-Hein, K.H., Kleesiek, J., Stiefelhagen, R.: Detailed annotations of chest x-rays via ct projection for report understanding. In: Proceedings of the 33th British Machine Vision Conference (BMVC) (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 326, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 326, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 138, + 326, + 481, + 380 + ], + "type": "text", + "content": "99. Setio, A.A.A., Traverso, A., De Bel, T., Berens, M.S., Van Den Bogaard, C., Cerello, P., Chen, H., Dou, Q., Fantacci, M.E., Geurts, B., et al.: Validation, comparison, and combination of algorithms for automatic detection of pulmonary nodules in computed tomography images: the luna16 challenge. Medical image analysis 42, 1-13 (2017)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 134, + 380, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 380, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 134, + 380, + 481, + 413 + ], + "type": "text", + "content": "100. Shi, P., Qiu, J., Abaxi, S.M.D., Wei, H., Lo, F.P.W., Yuan, W.: Generalist vision foundation models for medical imaging: A case study of segment anything model on zero-shot medical segmentation. Diagnostics 13(11), 1947 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 134, + 414, + 481, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 414, + 481, + 457 + ], + "spans": [ + { + "bbox": [ + 134, + 414, + 481, + 457 + ], + "type": "text", + "content": "101. Simpson, A.L., Antonelli, M., Bakas, S., Bilello, M., Farahani, K., Van Ginneken, B., Kopp-Schneider, A., Landman, B.A., Litjens, G., Menze, B., et al.: A large annotated medical image dataset for the development and evaluation of segmentation algorithms. arXiv preprint arXiv:1902.09063 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 134, + 457, + 481, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 457, + 481, + 512 + ], + "spans": [ + { + "bbox": [ + 134, + 457, + 481, + 512 + ], + "type": "text", + "content": "102. Sofiuk, K., Petrov, I., Barinova, O., Konushin, A.: F-BRS: Rethinking Backpropagating Refinement for Interactive Segmentation. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8620-8629. IEEE, Seattle, WA, USA (Jun 2020). https://doi.org/10.1109/CVPR42600.2020.00865, https://ieeexplore.ieee.org/document/9156403/" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 134, + 512, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 512, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 134, + 512, + 481, + 544 + ], + "type": "text", + "content": "103. Sofiuk, K., Petrov, I.A., Konushin, A.: Reviving Iterative Training with Mask Guidance for Interactive Segmentation (Feb 2021), http://arxiv.org/abs/2102.06583, arXiv:2102.06583 [cs]" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 545, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 545, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 134, + 545, + 481, + 599 + ], + "type": "text", + "content": "104. Song, Y., Zheng, J., Lei, L., Ni, Z., Zhao, B., Hu, Y.: CT2US: Cross-modal transfer learning for kidney segmentation in ultrasound images with synthesized data. Ultrasonics 122, 106706 (2022). https://doi.org/https://doi.org/10.1016/j.ultras.2022.106706, https://www.sciencedirect.com/science/article/pii/S0041624X22000191" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 600, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 600, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 600, + 481, + 632 + ], + "type": "text", + "content": "105. Staal, J., Abramoff, M.D., Niemeijer, M., Viergever, M.A., Van Ginneken, B.: Ridge-based vessel segmentation in color images of the retina. IEEE transactions on medical imaging 23(4), 501-509 (2004)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 632, + 481, + 665 + ], + "type": "text", + "content": "106. Tang, Y., Yang, D., Li, W., Roth, H.R., Landman, B., Xu, D., Nath, V., Hatamizadeh, A.: Self-Supervised Pre-Training of Swin Transformers for 3D Medical Image Analysis. In: 2022 IEEE/CVF Conference on Computer Vision" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "22" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 235, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 235, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 235, + 102 + ], + "type": "text", + "content": "H.E. Wong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 21 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 152, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 152, + 116, + 481, + 149 + ], + "type": "text", + "content": "and Pattern Recognition (CVPR). pp. 20698-20708. IEEE, New Orleans, LA, USA (Jun 2022). https://doi.org/10.1109/CVPR52688.2022.02007, https://ieeexplore.ieee.org/document/9879123/" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 172 + ], + "type": "text", + "content": "107. Vezhnevets, V., Konouchine, V.: \"GrowCut\" - Interactive Multi-Label N-D Image Segmentation By Cellular Automata. proc. of Graphicon 1(4), 150-156 (2005)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 481, + 194 + ], + "type": "text", + "content": "108. Wang, C., Chen, X., Ning, H., Li, S.: Sam-octa: A fine-tuning strategy for applying foundation model to octa image segmentation tasks (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 194, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 194, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 134, + 194, + 481, + 237 + ], + "type": "text", + "content": "109. Wang, G., Li, W., Zuluaga, M.A., Pratt, R., Patel, P.A., Aertsen, M., Doel, T., David, A.L., Deprest, J., Ourselin, S., Vercauteren, T.: Interactive Medical Image Segmentation Using Deep Learning With Image-Specific Fine Tuning. IEEE Transactions on Medical Imaging 37(7), 1562-1573 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 237, + 481, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 237, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 133, + 237, + 481, + 282 + ], + "type": "text", + "content": "110. Wang, G., Zuluaga, M.A., Li, W., Pratt, R., Patel, P.A., Aertsen, M., Doel, T., David, A.L., Deprest, J., Ourselin, S., Vercauteren, T.: DeepIGeoS: A Deep Interactive Geodesic Framework for Medical Image Segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence 41(7), 1559-1572 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 281, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 281, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 133, + 281, + 481, + 335 + ], + "type": "text", + "content": "111. Wang, G., Zuluaga, M.A., Pratt, R., Aertsen, M., Doel, T., Klusmann, M., David, A.L., Deprest, J., Vercauteren, T., Ourselin, S.: Slic-Seg: A minimally interactive segmentation of the placenta from sparse and motion-corrupted fetal MRI in multiple views. Medical Image Analysis 34, 137-147 (Dec 2016). https://doi.org/10.1016/j.media.2016.04.009" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 335, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 335, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 133, + 335, + 481, + 369 + ], + "type": "text", + "content": "112. Wang, X., Wang, W., Cao, Y., Shen, C., Huang, T.: Images speak in images: A generalist painter for in-context visual learning. arXiv preprint arXiv:2212.02499 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 369, + 481, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 369, + 481, + 403 + ], + "spans": [ + { + "bbox": [ + 133, + 369, + 481, + 403 + ], + "type": "text", + "content": "113. Wang, X., Zhang, X., Cao, Y., Wang, W., Shen, C., Huang, T.: SegGPT: Segmenting Everything In Context (Apr 2023), http://arxiv.org/abs/2304.03284, arXiv:2304.03284 [cs]" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 402, + 481, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 402, + 481, + 446 + ], + "spans": [ + { + "bbox": [ + 133, + 402, + 481, + 446 + ], + "type": "text", + "content": "114. Wasserthal, J., Breit, H.C., Meyer, M.T., Pradella, M., Hinck, D., Sauter, A.W., Heye, T., Boll, D.T., Cyriac, J., Yang, S., et al.: Totalsegmentator: Robust segmentation of 104 anatomic structures in ct images. Radiology: Artificial Intelligence 5(5) (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 446, + 481, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 446, + 481, + 480 + ], + "spans": [ + { + "bbox": [ + 133, + 446, + 481, + 480 + ], + "type": "text", + "content": "115. Wu, J., Fu, R., Fang, H., Liu, Y., Wang, Z., Xu, Y., Jin, Y., Arbel, T.: Medical SAM Adapter: Adapting Segment Anything Model for Medical Image Segmentation (Apr 2023), http://arxiv.org/abs/2304.12620, arXiv:2304.12620 [cs]" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 479, + 481, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 479, + 481, + 523 + ], + "spans": [ + { + "bbox": [ + 133, + 479, + 481, + 523 + ], + "type": "text", + "content": "116. Xu, N., Price, B., Cohen, S., Yang, J., Huang, T.: Deep Interactive Object Selection. In: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 373-381. IEEE, Las Vegas, NV, USA (Jun 2016). https://doi.org/10.1109/CVPR.2016.47, http://ieeexplore.ieee.org/document/7780416/" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 523, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 523, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 133, + 523, + 481, + 544 + ], + "type": "text", + "content": "117. Xu, N., Price, B., Cohen, S., Yang, J., Huang, T.: Deep GrabCut for Object Selection. arXiv (2017), arXiv:1707.00243" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 544, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 544, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 133, + 544, + 481, + 578 + ], + "type": "text", + "content": "118. Yushkevich, P.A., Piven, J., Hazlett, H.C., Smith, R.G., Ho, S., Gee, J.C., Gering, G.: User-guided 3d active contour segmentation of anatomical structures: significantly improved efficiency and reliability. Neuroimage 31(3), 1116-1128 (2006)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 578, + 481, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 578, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 133, + 578, + 481, + 611 + ], + "type": "text", + "content": "119. Zhang, J., Ding, X., Hu, D., Jiang, Y.: Semantic segmentation of Covid-19 lesions with a multiscale dilated convolutional network. Scientific Reports 12(1), 1847 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 611, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 611, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 133, + 611, + 481, + 632 + ], + "type": "text", + "content": "120. Zhang, K., Liu, D.: Customized segment anything model for medical image segmentation (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 632, + 481, + 666 + ], + "type": "text", + "content": "121. Zhang, K., Zhuang, X.: Cycle mix: A holistic strategy for medical image segmentation from scribble supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11656-11665 (2022)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 382, + 91, + 447, + 102 + ], + "type": "text", + "content": "ScribblePrompt" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "23" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 22 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 413 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "122. Zhang, S., Liew, J.H., Wei, Y., Wei, S., Zhao, Y.: Interactive Object Segmentation With Inside-Outside Guidance. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 12231-12241. IEEE (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 182 + ], + "type": "text", + "content": "123. Zhang, T.Y., Suen, C.Y.: A fast parallel algorithm for thinning digital patterns. Commun. ACM 27(3), 236-239 (mar 1984). https://doi.org/10.1145/357994.358023, https://doi.org/10.1145/357994.358023" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 481, + 215 + ], + "type": "text", + "content": "124. Zhang, Y., Xian, M., Cheng, H.D., Shareef, B., Ding, J., Xu, F., Huang, K., Zhang, B., Ning, C., Wang, Y.: Busis: A benchmark for breast ultrasound image segmentation. In: Healthcare. vol. 10, p. 729. MDPI (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 215, + 481, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 215, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 215, + 481, + 247 + ], + "type": "text", + "content": "125. Zhao, Q., Lyu, S., Bai, W., Cai, L., Liu, B., Wu, M., Sang, X., Yang, M., Chen, L.: A multi-modality ovarian tumor ultrasound image dataset for unsupervised cross-domain semantic segmentation. CoRR abs/2207.06799 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 248, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 248, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 248, + 481, + 293 + ], + "type": "text", + "content": "126. Zheng, G., Chu, C., Belavy, D.L., Ibragimov, B., Korez, R., Vrtovec, T., Hutt, H., Everson, R., Meakin, J., Andrade, I.L., et al.: Evaluation and comparison of 3d intervertebral disc localization and segmentation methods for 3d t2 mr data: A grand challenge. Medical image analysis 35, 327-344 (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 335 + ], + "type": "text", + "content": "127. Zheng, X., Wang, Y., Wang, G., Liu, J.: Fast and robust segmentation of white blood cell images by self-supervised learning. Micron 107, 55-71 (2018). https://doi.org/https://doi.org/10.1016/j.micron.2018.01.010, https://www.sciencedirect.com/science/article/pii/S0968432817303037" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 335, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 335, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 335, + 481, + 380 + ], + "type": "text", + "content": "128. Zhou, T., Li, L., Bredell, G., Li, J., Unkelbach, J., Konukoglu, E.: Volumetric memory network for interactive medical image segmentation. Medical Image Analysis 83, 102599 (Jan 2023). https://doi.org/10.1016/j.media.2022.102599, https://linkinghub.elsevier.com/retrieve/pii/S1361841522002316" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "text", + "content": "129. Zou, X., Yang, J., Zhang, H., Li, F., Li, L., Wang, J., Wang, L., Gao, J., Lee, Y.J.: Segment everything everywhere all at once. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "24" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 235, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 235, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 235, + 102 + ], + "type": "text", + "content": "H.E. Wong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 23 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/SeA_ Semantic Adversarial Augmentation for Last Layer Features from Unsupervised Representation Learning/ca46bbef-cd36-48f1-b01d-aebb511e39a1_content_list.json b/2024/SeA_ Semantic Adversarial Augmentation for Last Layer Features from Unsupervised Representation Learning/ca46bbef-cd36-48f1-b01d-aebb511e39a1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a501a136711c8d24dd9fc32039e1ca784006b494 --- /dev/null +++ b/2024/SeA_ Semantic Adversarial Augmentation for Last Layer Features from Unsupervised Representation Learning/ca46bbef-cd36-48f1-b01d-aebb511e39a1_content_list.json @@ -0,0 +1,1874 @@ +[ + { + "type": "text", + "text": "SeA: Semantic Adversarial Augmentation for Last Layer Features from Unsupervised Representation Learning", + "text_level": 1, + "bbox": [ + 243, + 140, + 759, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qi Qian $^{1}$ , Yuanhong $\\mathrm{Xu}^{2}$ , and Juhua $\\mathrm{Hu}^{3}$", + "bbox": [ + 328, + 234, + 673, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Alibaba Group, Bellevue, WA 98004, USA", + "bbox": [ + 351, + 261, + 648, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Alibaba Group, Hangzhou, China", + "bbox": [ + 374, + 277, + 620, + 289 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 School of Engineering and Technology,", + "bbox": [ + 364, + 290, + 635, + 304 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "University of Washington, Tacoma, WA 98402, USA", + "bbox": [ + 326, + 305, + 674, + 318 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{qi.qian, yuanhong.xuyh}@alibaba-inc.com, juhuah@uw.edu", + "bbox": [ + 289, + 320, + 712, + 333 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Deep features extracted from certain layers of a pre-trained deep model show superior performance over the conventional hand-crafted features. Compared with fine-tuning or linear probing that can explore diverse augmentations, e.g., random crop/flipping, in the original input space, the appropriate augmentations for learning with fixed deep features are more challenging and have been less investigated, which degenerates the performance. To unleash the potential of fixed deep features, we propose a novel semantic adversarial augmentation (SeA) in the feature space for optimization. Concretely, the adversarial direction implied by the gradient will be projected to a subspace spanned by other examples to preserve the semantic information. Then, deep features will be perturbed with the semantic direction, and augmented features will be applied to learn the classifier. Experiments are conducted on 11 benchmark downstream classification tasks with 4 popular pre-trained models. Our method is $2\\%$ better than the deep features without SeA on average. Moreover, compared to the expensive fine-tuning that is expected to give good performance, SeA shows a comparable performance on 6 out of 11 tasks, demonstrating the effectiveness of our proposal in addition to its efficiency.", + "bbox": [ + 261, + 372, + 743, + 635 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Semantic augmentation $\\cdot$ Deep features $\\cdot$ Unsupervised representation learning $\\cdot$ Self-supervised learning", + "bbox": [ + 261, + 648, + 740, + 678 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 705, + 375, + 722 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep learning can be partially considered as a representation learning method that aims to extract features from raw data directly. By obtaining appropriate representations, a simple linear model can achieve state-of-the-art performance on challenging tasks, e.g., classification [16], object detection [30], etc. After the success of deep learning [21], researchers investigate the representations learned", + "bbox": [ + 212, + 738, + 787, + 815 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author", + "bbox": [ + 217, + 824, + 382, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6fe0bae98de123087cdf64e4d44152274d04a431f794049f1b69f26c35b1985c.jpg", + "image_caption": [ + "Fig. 1: Illustration of semantic adversarial augmentation (SeA). The red solid and empty circles denote the original data and its augmentation, respectively. Left: Conventional adversarial augmentation can perturb with arbitrary direction (e.g., augmentation may appear the same as the original); Right: SeA augments examples with semantic directions spanned by features from real data (e.g., different sectors show different subspaces for augmentation, where we can get more semantic meaningful augmentation)." + ], + "image_footnote": [], + "bbox": [ + 297, + 146, + 702, + 282 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "from a large-scale data set, i.e., ImageNet [31]. Surprisingly, the deep features extracted from a certain layer of a pre-trained model can outperform handcrafted features on various downstream tasks [9], demonstrating the efficacy of the data-dependent representation learning mechanism implied by deep learning.", + "bbox": [ + 212, + 428, + 787, + 489 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unlike hand-crafted features, representations obtained by deep learning are highly data/task-dependent as illustrated in [37]. When optimizing the task from the source domain, the neural network focuses on exploring the patterns related to the specific training task, while ignoring diverse information that has potential for different tasks. Hence, different representations can be learned from different training tasks even with the same training data from the source domain. Conventional methods learn representations with the labels of examples, which capture the knowledge only for the given labels and limit the information in deep features of each example.", + "bbox": [ + 212, + 489, + 787, + 626 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To mitigate the problem, fine-tuning parameters of the whole network becomes prevalent for various downstream tasks [18]. On one hand, fine-tuning can benefit from the prior knowledge in the pre-trained representations. On the other hand, it can further optimize representations with diverse augmentations on the target task. Consequently, fine-tuning works better than learning with fixed representations on downstream tasks.", + "bbox": [ + 212, + 627, + 787, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "An important advantage of fine-tuning over learning with fixed deep features is the additional information from semantic augmentations. Obtaining effective augmentations in image space is convenient with semantic operators such as random crop, flipping, etc. In contrast, obtaining appropriate augmentations in the feature space for deep features becomes challenging, due to the lack of semantic operations. To leverage input space augmentations, a linear classifier can be learned with a frozen backbone by linear probing [6] that generates augmented images for optimization at each iteration.", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Q. Qian et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While fine-tuning and linear probing show promising performance, learning with fixed features is still attractive due to its good properties for real-world applications. First, it only applies deep models to extract features for each original example once, which is applicable for limited computational resources. On the contrary, fine-tuning has the full forward and backward pass and linear probing has the forward pass for each augmented example, which is more expensive for optimization. Second, the same models can be reused for different downstream tasks to extract deep features, while fine-tuning will adjust all parameters in pre-trained models. Thereafter, it has to keep a specific deep model for each task, which becomes intractable for handling hundreds of downstream tasks simultaneously. Finally, given fixed features, learning a linear classifier can be formulated as a convex problem that has the global optimum with a theoretical guarantee [2], while preserving the knowledge from the pre-trained model in features. It requires much less tuning efforts than fine-tuning that has to be tuned carefully to avoid the collapse of pre-trained parameters and catastrophic forgetting [22]. Therefore, we focus on fixed deep features extracted after the last pooling layer (i.e., inputs for the last fully-connected layer) in this work.", + "bbox": [ + 217, + 146, + 785, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Although some existing works [32,35] consider augmentation in intermediate layers to help optimize the whole network, little efforts were devoted to the deep features after the last pooling layer. Moreover, the existing feature space augmentation method [35] is proposed to complement the input space augmentation. It relies on other augmentation techniques and cannot work well solely.", + "bbox": [ + 212, + 402, + 782, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To tackle the problem, in this work, we propose a novel semantic adversarial augmentation strategy in the feature space of fixed deep features. Concretely, the gradient of each example is computed at first, and then a semantic direction can be observed by projecting the gradient to the subspace spanned by real data. Finally, the features of examples are augmented according to the obtained semantic direction for learning the linear classifier. Fig. 1 illustrates the proposed augmentation method that obtains effective augmentations with only fixed features. Moreover, to further mitigate the information loss in deep features from supervised pre-trained models, those from pre-trained models with different unsupervised pretext tasks are also investigated with an ensemble strategy. Our main contributions can be summarized as follows.", + "bbox": [ + 212, + 479, + 785, + 645 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We empirically demonstrate the current performance gap between fine-tuning and learning with fixed deep features. Those deep features extracted from 4 representative pre-trained models are evaluated on 11 downstream tasks.", + "- To improve the performance of deep features, we propose a semantic adversarial augmentation method to obtain appropriate augmentations tailored for fixed features. In addition, a smoothed hinge loss is investigated to demonstrate the augmentation direction explicitly.", + "- Our proposed SeA gains $2\\%$ accuracy on average over 11 downstream tasks compared to the baseline using deep features without augmentations. Moreover, compared with fine-tuning that is expected to give good performance, our recipe for learning with deep features can achieve comparable performance on 6 out of 11 tasks with way less computational cost, which shows" + ], + "bbox": [ + 225, + 657, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Semantic Adversarial Augmentation for Last Layer Features", + "bbox": [ + 328, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the potential of self-supervised pre-trained deep features and demonstrates both the efficiency and effectiveness of our proposal.", + "bbox": [ + 241, + 146, + 784, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 198, + 387, + 214 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.1 Deep Features", + "text_level": 1, + "bbox": [ + 215, + 228, + 380, + 243 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Modern deep neural networks consist of multiple layers to exact representations from raw materials. After that, a simple linear model encoded by a fully-connected layer can be attached on top of the representations for classification. After pre-training on a large-scale data set, the obtained neural network can be considered as a feature extractor and a new fully-connected layer can be learned for target tasks.", + "bbox": [ + 212, + 252, + 785, + 340 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "When the full label information of data is available, the learning objective is explicit and a supervised representation learning can be conducted by optimizing a conventional classification task, which is equivalent to optimizing the triplet loss as in distance metric learning [28]. The obtained deep features can outperform hand-crafted features on different applications, e.g., classification [9], distance metric learning [27], etc. Besides, some work aimed to obtain robust representations for different downstream tasks within the framework of classification [26].", + "bbox": [ + 212, + 342, + 785, + 462 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Without any label information, the learning objective varies in unsupervised learning. First, each instance can be considered as an individual class and the representations can be learned by instance discrimination [5, 15]. In addition, clustering can be applied to capture the relationship between different instances. A coarse-grained classification task defined on clusters can be leveraged to optimize representations [3, 29]. Finally, other pretext tasks beyond classification also demonstrate the effective representation after pre-training [13]. Compared with supervised representation learning, the different objectives in unsupervised learning can learn various semantic information even from the same data set. In this work, we will systematically study representations from different pre-trained models and illustrate the performance gap compared to fine-tuning the whole neural network.", + "bbox": [ + 212, + 463, + 787, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Augmentation in Feature Space", + "text_level": 1, + "bbox": [ + 215, + 665, + 522, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Due to the over-parameterization property of deep neural networks, augmentation is essential for training deep models to avoid over-fitting [21]. Given an image, a perturbed copy can be observed by standard image operations such as random crop, flipping, color jitter, etc. [16]. The augmented images have a large variance while preserving the semantic information of the original image, i.e., an image of a cat is still a cat after augmentation, which helps train effective models with the additional information.", + "bbox": [ + 212, + 689, + 785, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, augmentation in feature space becomes more challenging due to the lack of semantic preserving operators. Some methods consider the augmentation of features from intermediate layers to help train the whole deep neural", + "bbox": [ + 212, + 795, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Q. Qian et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "networks in different applications [4, 32, 33, 35]. For example, [33] aims to learn feature augmentation with a feature generator for unsupervised domain adaptation. [4] tries to augment intermediate feature maps with adversarial feature moments in batch normalization [17] for efficient training. Manifold mixup [32] proposes to apply the original mixup [39] to features from multiple layers for augmentation, which also includes the layer investigated in this work. However, their study shows that the gain of mixup is mainly from the original image space and feature space of early stages, while that from the feature space of the last layer is negligible, which is consistent with our observation in the ablation study. ISDA [35] considers the semantic augmentation in feature space but it has to obtain the semantic direction with input space augmentations, which is complementary to input space augmentations but cannot work as the sole augmentation for deep features well. On the contrary, we propose SeA to project the adversarial direction with features of real data points. Moreover, our proposal is tailored for fixed features and is different from existing works that are for optimizing the whole network.", + "bbox": [ + 212, + 146, + 787, + 388 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Semantic Adversarial Augmentation", + "text_level": 1, + "bbox": [ + 215, + 412, + 602, + 429 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We start the analysis with a standard classification framework. Let $\\{x_{i},y_{i}\\}_{i = 1}^{n}$ denote the training data set and $\\{f_k\\}_{k = 1}^K$ contains a set of $K$ pre-trained models, which can be pre-trained with different learning objectives on different data sets. In this work, we directly use the representations extracted from pre-trained deep models and learn a simple linear model on top of it. Representations from different deep models are concatenated as the final representation for each example. Given $x_{i}$ , the representation will be extracted as", + "bbox": [ + 212, + 444, + 787, + 551 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {i} = \\left[ f _ {1} \\left(x _ {i}\\right), \\dots , f _ {K} \\left(x _ {i}\\right) \\right] \\in \\mathbb {R} ^ {d} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 390, + 561, + 785, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $d = \\sum_{k}d_{k}$ and $d_{k}$ is the dimension of the representation from the $k$ -th pre-trained deep model.", + "bbox": [ + 212, + 590, + 785, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With the above fixed deep features, a classification model can be learned by minimizing the empirical risk with the appropriate regularization as", + "bbox": [ + 212, + 622, + 785, + 652 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\mathbf {w}} \\sum_ {i} \\ell (\\mathbf {x} _ {i}, y _ {i}; \\mathbf {w}) + \\frac {\\tau}{2} \\| \\mathbf {w} \\| _ {F} ^ {2} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 393, + 662, + 785, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tau$ is the weight for $L_{2}$ regularization and $\\mathbf{w}$ denotes parameters of the classification model, which is a linear classifier with the prediction probability of the $i$ -th example on the $j$ -th class as", + "bbox": [ + 212, + 705, + 785, + 751 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np _ {i, j} = \\frac {\\exp \\left(\\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {j}\\right)}{\\sum_ {c = 1} ^ {C} \\exp \\left(\\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {c}\\right)} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 410, + 762, + 785, + 800 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$C$ is the number of classes and $\\ell$ is a loss function for learning that will be discussed later.", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Semantic Adversarial Augmentation for Last Layer Features", + "bbox": [ + 328, + 114, + 732, + 130 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Semantic Direction", + "text_level": 1, + "bbox": [ + 215, + 146, + 419, + 159 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In a standard fine-tuning pipeline, an image can be augmented by multiple random perturbations. When only deep features are available, it is hard to adopt existing techniques to generate appropriate augmented examples, since the semantic direction in the feature space is hard to capture. Manifold mixup [32] shows that randomly augmenting features from the last output layers has little improvement compared to the augmentation in input image space. To facilitate the performance of deep features, we investigate the semantic adversarial direction for augmentation.", + "bbox": [ + 212, + 170, + 787, + 290 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "First, for the $i$ -th example and its representation $\\mathbf{x}_i$ , we consider its adversarial direction that can be obtained by maximizing the loss function [12]", + "bbox": [ + 214, + 291, + 785, + 321 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {\\mathbf {x}: \\left\\| \\mathbf {x} - \\mathbf {x} _ {i} \\right\\| _ {2} \\leq \\gamma} \\ell \\left(\\mathbf {x} _ {i}, y _ {i}; \\mathbf {w}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 416, + 332, + 785, + 356 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Note that the gradient indicates the ideal direction for the adversarial perturbation, and a standard adversarial example can be obtained by gradient ascent as", + "bbox": [ + 214, + 364, + 785, + 407 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {x}} _ {i} = \\mathbf {x} _ {i} + \\eta \\nabla_ {\\mathbf {x} _ {i}} \\ell \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 437, + 421, + 784, + 436 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "However, the gradient direction may not be semantically informative. [12] shows that the loss can be substantially increased with the adversarial example generated from the gradient, while the appearance is almost the same as the original image as illustrated in Fig. 1. Therefore, the conventional adversarial learning method helps improve the robustness to the adversarial attack but may be inappropriate for regularizing the generic learning task that requires additional information from diverse images.", + "bbox": [ + 212, + 446, + 787, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To obtain the semantic direction, we propose to project the gradient direction to the subspace spanned by real data points. Concretely, let $\\{\\mathbf{x}_j\\}_{j=1}^b$ denote a mini-batch of data and $\\mathbf{g}_i = \\nabla_{\\mathbf{x}_i}\\ell$ indicate the gradient of $\\mathbf{x}_i$ . For the $i$ -th image, a semantic adversarial direction $\\hat{\\mathbf{g}}_i$ consists of representations from data as", + "bbox": [ + 214, + 553, + 787, + 613 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {g}} _ {i} = \\sum_ {j: j \\neq i} q _ {j} \\mathbf {x} _ {j}; \\quad \\mathbf {q} ^ {*} = \\arg \\min _ {\\mathbf {q} \\in \\Delta} \\mathcal {D} (\\hat {\\mathbf {g}} _ {i}, \\mathbf {g} _ {i}) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 354, + 623, + 785, + 655 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{D}(\\cdot, \\cdot)$ is a distance function and $\\varDelta$ is a simplex as $\\varDelta=\\{\\mathbf{q}\\in\\mathbb{R}^{b-1}|\\sum_{j}q_{j}=1;\\forall j,q_{j}\\geq 0\\}$ . With the projection, we aim to find an adversarial direction in a subspace consisting of original data points as illustrated in Fig. 1.", + "bbox": [ + 214, + 666, + 787, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The last challenge is to obtain $\\mathbf{q}$ efficiently with a distance function. By adopting the squared Euclidean distance, which is a standard distance measurement, the optimization problem for $\\mathbf{q}$ can be written as", + "bbox": [ + 212, + 712, + 785, + 758 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\mathbf {q} \\in \\Delta} \\| \\sum_ {j: j \\neq i} q _ {j} \\mathbf {x} _ {j} - \\mathbf {g} _ {i} \\| _ {2} ^ {2} - \\alpha H (\\mathbf {q}) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 768, + 785, + 800 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $H(\\mathbf{q})$ is the entropy of $\\mathbf{q}$ that helps improve the robustness to different batches. When normalizing features such that $\\| \\mathbf{x}_j\\| _2 = \\| \\mathbf{g}_i\\| _2 = 1$ to obtain", + "bbox": [ + 214, + 809, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Q. Qian et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 5 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Semantic Adversarial Augmentation (SeA) for Given Features" + ], + "code_body": "1: Input: Dataset $\\{x_i,y_i\\}$ , pre-trained models $\\{f_k\\}$ , iterations $T$ , linear model $\\mathbf{w}$ , $\\alpha$ , $\\tau$ , $\\eta$ \n2: Extract deep features by $\\{f_k\\}$ for all examples as $\\{\\mathbf{x}_i\\}$ \n3: for $t = 1,\\dots,T$ do \n4: Receive a mini-batch of examples $\\{\\mathbf{x}_i,y_i\\}_{i=1}^b$ \n5: Obtain augmented examples $\\{\\tilde{\\mathbf{x}}_i,y_i\\}$ as in Eqn. 11 \n6: Optimize $\\mathbf{w}$ by SGD: $\\mathbf{w} = \\mathbf{w} - \\eta_w\\left(\\frac{1}{b}\\sum_i^b\\nabla_{\\mathbf{w}}\\ell(\\tilde{\\mathbf{x}}_i,y_i;\\mathbf{w}) + \\tau \\mathbf{w}\\right)$ \n7: end for \n8: return $\\mathbf{w}$", + "bbox": [ + 222, + 166, + 785, + 292 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the direction without the influence from the magnitude, the problem can be upper-bounded as", + "bbox": [ + 212, + 321, + 782, + 352 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Proposition 1 With unit length variables $\\{\\mathbf{x}_j\\}$ and $\\mathbf{g}_i$ , we have", + "bbox": [ + 214, + 359, + 683, + 378 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\| \\sum_ {j: j \\neq i} q _ {j} \\mathbf {x} _ {j} - \\mathbf {g} _ {i} \\right\\| _ {2} ^ {2} \\leq 2 - 2 \\sum_ {j} q _ {j} \\mathbf {x} _ {j} ^ {\\top} \\mathbf {g} _ {i} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 361, + 386, + 785, + 419 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The detailed proof can be found in the supplementary.", + "bbox": [ + 214, + 428, + 609, + 443 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "By rearranging the terms, we can maximize the lower-bound of the original problem as", + "bbox": [ + 212, + 443, + 782, + 472 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {\\mathbf {q} \\in \\Delta} \\sum_ {j: j \\neq i} q _ {j} \\mathbf {x} _ {j} ^ {\\top} \\mathbf {g} _ {i} + \\alpha H (\\mathbf {q}) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 400, + 482, + 785, + 516 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "According to the K.K.T. condition [2], $\\mathbf{q}$ has a closed-form solution.", + "bbox": [ + 238, + 525, + 728, + 541 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Proposition 2 The problem in Eqn. 9 has the optimal solution as", + "bbox": [ + 214, + 549, + 699, + 564 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nq _ {j} = \\frac {\\exp \\left(\\mathbf {x} _ {j} ^ {\\top} \\mathbf {g} _ {i} / \\alpha\\right)}{Z}; \\quad Z = \\sum_ {k: k \\neq i} \\exp \\left(\\mathbf {x} _ {k} ^ {\\top} \\mathbf {g} _ {i} / \\alpha\\right) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 333, + 571, + 785, + 612 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Given the semantic adversarial direction, the target example with semantic perturbation can be obtained as", + "bbox": [ + 212, + 622, + 782, + 652 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {x}} _ {i} = \\Pi \\left(\\mathbf {x} _ {i} + \\eta \\Pi \\left(\\hat {\\mathbf {g}} _ {i}\\right)\\right) \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 416, + 662, + 785, + 680 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\Pi(\\cdot)$ normalizes the vector to the unit length if required and $\\eta$ denotes the step size for augmentation. Compared with the adversarial perturbation in Eqn. 5, the augmentation in Eqn. 11 projects the gradient to the direction consisting of real data points, which can capture the semantic information in the feature space effectively. Alg. 1 summarizes the proposed method.", + "bbox": [ + 212, + 689, + 785, + 765 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Illustration of Adversarial Direction", + "text_level": 1, + "bbox": [ + 214, + 785, + 558, + 800 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this subsection, we will generalize the cross entropy loss to help illustrate the proposed augmentation strategy.", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Semantic Adversarial Augmentation for Last Layer Features", + "bbox": [ + 328, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The standard multi-class hinge loss [8] for convex optimization can be written as", + "bbox": [ + 215, + 146, + 782, + 175 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\ell (\\mathbf {x} _ {i}, y _ {i}; \\mathbf {w}) = \\max \\left\\{0, \\delta + \\max _ {c \\neq y _ {i}} \\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {c} - \\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {y _ {i}} \\right\\} \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 331, + 188, + 784, + 213 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\delta$ is a pre-defined margin.", + "bbox": [ + 215, + 224, + 449, + 239 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Unlike convex optimization, cross-entropy loss is prevalent in deep learning, which is a smooth function and can help accelerate the convergence [2]. Therefore, we propose to obtain a smoothed hinge loss by introducing a distribution over logits from different classes $\\mathbf{p} \\in \\Delta$ where $\\Delta$ is the simplex.", + "bbox": [ + 214, + 239, + 782, + 299 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "First, the original hinge loss is equivalent to", + "bbox": [ + 238, + 301, + 560, + 316 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\ell \\left(\\mathbf {x} _ {i}, y _ {i}; \\mathbf {w}\\right) = \\max _ {\\mathbf {p} \\in \\Delta} p _ {y _ {i}} \\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {y _ {i}} + \\sum_ {c \\neq y _ {i}} p _ {c} \\left(\\delta + \\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {c}\\right) - \\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {y _ {i}} \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 290, + 328, + 782, + 359 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "According to the analysis for cross entropy loss [28], the loss can be smoothed by adding an entropy regularization for $\\mathbf{p}$ as", + "bbox": [ + 215, + 372, + 782, + 402 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\ell \\left(\\mathbf {x} _ {i}, y _ {i}; \\mathbf {w}\\right) = \\max _ {\\mathbf {p} \\in \\Delta} p _ {y _ {i}} \\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {y _ {i}} + \\sum_ {c \\neq y _ {i}} p _ {c} (\\delta + \\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {c}) + \\lambda H (\\mathbf {p}) - \\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {y _ {i}} \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 246, + 415, + 782, + 446 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $H(\\mathbf{p})$ denotes the entropy of $\\mathbf{p}$ and $\\lambda$ is the coefficient.", + "bbox": [ + 215, + 459, + 656, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Proposition 3 The loss function in Eqn. 14 is equivalent to", + "bbox": [ + 215, + 488, + 656, + 503 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\ell (\\mathbf {x} _ {i}, y _ {i}; \\mathbf {w}) = - \\lambda \\log \\frac {\\exp \\left(\\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {y _ {i}} / \\lambda\\right)}{\\exp \\left(\\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {y _ {i}} / \\lambda\\right) + \\sum_ {c \\neq y _ {i}} \\exp \\left(\\left(\\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {c} + \\delta\\right) / \\lambda\\right)} \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 259, + 515, + 782, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Remark It is obvious that the popular cross entropy loss is a special case of Eqn. 15 by letting $\\delta = 0$ and $\\lambda = 1$ . Our analysis connects the hinge loss in conventional methods to the popular loss function in deep learning.", + "bbox": [ + 214, + 563, + 782, + 607 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, since deep features are fixed, we can illustrate the gradient direction with the proposed smoothed loss function explicitly.", + "bbox": [ + 215, + 608, + 782, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Proposition 4 Given the loss function in Eqn. 15, the gradient of $\\mathbf{x}_i$ is", + "bbox": [ + 215, + 648, + 740, + 665 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\mathbf {x} _ {i}} \\ell (\\mathbf {x} _ {i}, y _ {i}; \\mathbf {w}) = \\sum_ {c = 1} ^ {C} p _ {c} \\mathbf {w} _ {c} - \\mathbf {w} _ {y _ {i}} \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 676, + 782, + 718 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{r} w h e r e p _ {c} = \\left\\{ \\begin{array}{l l} \\frac {\\exp ((\\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {c} + \\delta) / \\lambda)}{Z} c \\neq y _ {i} \\\\ \\frac {\\exp (\\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {c} / \\lambda)}{Z} c = y _ {i} \\end{array} a n d Z = \\exp (\\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {y _ {i}} / \\lambda) + \\sum_ {c \\neq y _ {i}} \\exp ((\\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {c} + \\delta) / \\lambda) \\leqslant 1, \\right. \\\\ \\delta) / \\lambda). \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 729, + 799, + 785 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Remark The gradient direction in the feature space indicates that an adversarial perturbation should be close to the primary directions of other classes, while far away from the direction of the corresponding class.", + "bbox": [ + 215, + 794, + 782, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Q. Qian et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 143, + 375, + 162 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To demonstrate our proposed method, we adopt state-of-the-art and widely used pre-trained models to extract deep features for evaluation. It should be noted that only two main architectures, including ResNet-50 [16] and vision transformer (ViT) [10], are prevalently applied in self-supervised learning [6,13,14,29]. Since ViT shows a worse classification performance than ResNet-50 with the frozen backbone as demonstrated in [14], we will focus on ResNet-50 with different public pre-trained parameters in the experiment. Specifically, one supervised pre-trained ResNet-50 and three self-supervised pre-trained ResNet-50 are applied for feature extraction. All of these models are pre-trained on ImageNet-1K [31] with different learning objectives. The details of different models can be found in the supplementary. The accuracy of the supervised model and that of linear probing for unsupervised models are summarized in Tab. 1.", + "bbox": [ + 212, + 180, + 787, + 363 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/efd01e69004db328cd85a6505de852edaf04f3f4ef9eb59a62ed2d5dd49af3fd.jpg", + "table_caption": [ + "Table 1: Performance of 4 pre-trained models on ImageNet." + ], + "table_footnote": [], + "table_body": "
Model#Pre-training EpochsObjectiveAcc%
Supervised [16] (S)901,000-class classification76.6
MoCo-v3 [6] (M)1,000instance discrimination74.6
BYOL [13] (B)1,000regression74.3
CoKe [29] (C)1,000cluster discrimination74.9
", + "bbox": [ + 256, + 415, + 746, + 491 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "11 diverse downstream data sets are applied for evaluation, including Aircraft [23], Caltech101 [11], Stanford Cars [19], CIFAR-10 [20], CIFAR-100 [20], CUB200-2011 (Birds) [34], Describable Textures Dataset (DTD) [7], Flowers [24], Food101 [1], Oxford-IIIT Pet (Pets) [25], and Sun397 [36]. We follow the evaluation protocol in [13]. Concretely, all models search hyper-parameters on the provided/generated validation set in each downstream task, and the standard metric on the provided test set is reported. Besides, mean per-class accuracy is reported on Aircraft, Caltech101, Flowers, and Pets, while Top-1 accuracy is utilized for other data sets. More details can be found in [13].", + "bbox": [ + 212, + 521, + 787, + 657 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Each model is fine-tuned with SGD by 100 epochs for sufficient training. The batch size is 256 and the momentum is 0.9. The learning rate is searched in a range of 7 logarithmically-spaced values between $10^{-4}$ and $10^{-1}$ . Weight decay is optional, for which if it is applied, the value will be searched with the same setting between $10^{-6}$ and $10^{-3}$ . The standard augmentation, i.e., random crop, random horizontal flipping, is applied as in most existing fine-tuning pipelines.", + "bbox": [ + 212, + 657, + 787, + 750 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For a fair comparison, we also apply SGD to learn the linear classifier with fixed deep features using the same batch size and momentum. Unlike fine-tuning, we have the constant learning rate for our method, which is searched in $\\{2^i\\}_{i = -2}^{i = 3}$ while weight decay is searched in $\\{0,10^{-6},10^{-5},10^{-4}\\}$ . For other parameters, we have $\\alpha$ , $\\lambda$ , $\\delta$ and $\\eta$ searched in $\\{0.01,0.02,0.05\\}$ , $\\{0.05,0.1,1\\}$ , $\\{0,1\\}$ , and $\\{0,0.2,0.4,0.8\\}$ , respectively.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Semantic Adversarial Augmentation for Last Layer Features", + "bbox": [ + 328, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For each image, we extract the output of the final pooling layer as deep features. Features from an individual model are normalized to the unit length. If multiple models are exploited, features from different models are concatenated as in Eqn. 1 and the combined feature vector is further normalized to the unit length. The only augmentation for training with fixed features is the proposed semantic adversarial augmentation. To approximate the direction of the gradient for each example, a mini-batch of data is leveraged to optimize the problem in Eqn. 9. The projection for normalization is also applied to the augmented examples.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 309, + 388, + 324 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Before the experiments on downstream tasks, we first investigate the effect of parameters in the proposed augmentation. The ablation study is conducted on CIFAR-100 with features concatenated from 4 pre-trained models and the accuracy is reported for comparison.", + "bbox": [ + 212, + 339, + 785, + 400 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Effect of Step Size for Augmentation First, we evaluate the step size $\\eta$ in the semantic adversarial augmentation. For SeA, we fix $\\lambda = 0.1$ , $\\delta = 0$ , $\\alpha = 0.01$ , and the learning rate as 1, according to the performance on the validation set. The weight of the augmentation $\\eta$ varies in $\\{0, 0.2, 0.4, 0.6, 0.8, 1.0\\}$ and the performance is summarized in Fig. 2-Fig. 3.", + "bbox": [ + 212, + 426, + 785, + 503 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/8af8176113b60714f697ac9f6830a26ba36edbd654a5dfd3770e0015a38a3082.jpg", + "image_caption": [ + "Fig. 2: Training accuracy Fig. 3: Test accuracy of Fig. 4: Test accuracy of various different $\\eta$ . different $\\eta$ ." + ], + "image_footnote": [], + "bbox": [ + 220, + 529, + 401, + 633 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/507bbac29dc0984305db8db92063b388e0db13e17504b0689ef7b587ab9354be.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 401, + 529, + 580, + 633 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/657555cd25279a74e4a27c3fa7884071d1d0a047c676f86fc08d378ca235c4b0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 591, + 532, + 769, + 632 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "When $\\eta = 0$ , there is no augmentation applied for training. Hence, we can observe that the training accuracy exceeds $92\\%$ after only 20 epochs. Due to the over-fitting on the training set, the test accuracy of this baseline is only $81.1\\%$ on CIFAR-100. By gradually increasing the step size for the augmentation, the training accuracy is decreased as expected, while the test accuracy increases. It demonstrates that the proposed semantic adversarial augmentation strategy can effectively mitigate the over-fitting problem for learning with fixed deep features. By setting an appropriate step size, the test accuracy can be improved to $82.9\\%$ , which helps shrink the gap to fine-tuning.", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Q. Qian et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Effect of Directions for Augmentation Different directions can be adopted for generating augmentation. Besides the proposed semantic adversarial aug-. mentations (SeA), four variants are included in the comparison.", + "bbox": [ + 212, + 146, + 787, + 191 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Adv: the original gradient direction without projection as in Eqn. 5.", + "- SeA_Neg: similar to the gradient direction in Proposition 4 but getting rid of the direction from the target class and keeping the direction of $\\sum_{c}^{C}p_{c}\\mathbf{w}_{c}$ .", + "- Rand: a uniformly random direction within subspace spanned by a minibatch of data, i.e., generating $\\mathbf{q}$ in Eqn. 6 as a random vector.", + "- Manifold mixup: a random semantic direction indicated by an example [32]." + ], + "bbox": [ + 223, + 203, + 782, + 296 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The step size for Adv is very sensitive and is searched in $\\{0.005, 0.01, 0.02, 0.05\\}$ , while the step size for others is searched in $\\{0.2, 0.4, 0.8, 1.0\\}$ .", + "bbox": [ + 212, + 306, + 795, + 337 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Fig. 4 shows the comparison on CIFAR-100. First, we can observe that most augmentation directions except the random one can improve the performance over the baseline without augmentation. It confirms that augmentation is important for better generalization even for learning from fixed representations. Then, the method with the random augmentation direction performs similarly to the baseline, which shows that it is challenging to find an effective direction for augmentation in feature space. However, with the original gradient direction directly as in Adv, the training loss will be significantly increased even with a small step size and only a step size that is much smaller than SeA can obtain an applicable model. This is because that the gradient direction only aims to maximize the loss while ignoring the data distribution in the feature space. By projecting the gradient to a data-dependent subspace, SeA and SeA_Neg can achieve a better performance than the baseline without a sophisticated setup for the step size. Moreover, by removing the direction from the target class, the accuracy is degenerated by about $1\\%$ . The comparison demonstrates that keeping the completed gradient direction is essential for obtaining appropriate augmentation. Finally, a random semantic direction indicated by examples as in mixup is worse than SeA with a clear margin.", + "bbox": [ + 212, + 338, + 787, + 609 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Comparison with Linear Probing Besides learning with fixed features, linear probing is another way to obtain the linear classifier with a frozen backbone. Compared with our method, it can obtain appropriate augmentations in image space. However, it has to forward each augmentation for learning and the cost is much higher than us.", + "bbox": [ + 212, + 633, + 787, + 708 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "To illustrate the effectiveness of the proposed augmentation method, we compare the performance between linear probing and our proposal in Tab. 2. For a fair comparison, all methods optimize features from a single model of MoCo. It shows that the proposed augmentation can outperform the augmentation in the input image space for learning the target classifier.", + "bbox": [ + 212, + 710, + 787, + 785 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Running Time Finally, we compare the running time of fine-tuning, linear probing, and SeA using fixed deep features in Tab. 3. Note that with features", + "bbox": [ + 212, + 809, + 787, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Semantic Adversarial Augmentation for Last Layer Features", + "bbox": [ + 328, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/8f4de505a39518388e7b05f4404e9692c789e153bea9dd487b7b547d71f0b758.jpg", + "table_caption": [ + "Table 2: Comparison of accuracy (%) Table 3: Comparison of running time. SeA on CIFAR-100. All methods utilize a utilize deep features from 4 models while single model from MoCo. other methods optimize only a single model." + ], + "table_footnote": [], + "table_body": "
baselinelinear probingSeA
77.677.879.0
Methodsfine-tuninglinear probing
Time (s)14,3834,437
", + "bbox": [ + 263, + 199, + 761, + 229 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "from 4 models, the dimension of the feature vector is 8, 192. However, optimizing the combined features is still much more efficient than fine-tuning a single model or linear probing with 2,048 final features. All methods train the corresponding model by 100 epochs and the running time is measured on a single V100 GPU. Fine-tuning costs 14,383 (sec) for the entire training. In contrast, SeA only costs 40 (sec) to obtain the optimal linear model. The efficiency of SeA using deep features implies its applicability for limited-resource scenarios. With efficient optimization, we can repeat experiments of SeA by multiple times. The standard deviation is only $0.03\\%$ on CIFAR-100 over 5 runs, which shows the stability of convex optimization. Although linear probing eliminates the cost of the backward pass, that of the forward pass at each iteration is still expensive. Considering that linear probing has a similar running cost as fine-tuning but with worse performance, it will not be included in the comparison over other downstream tasks. More ablation experiments can be found in the supplementary.", + "bbox": [ + 212, + 260, + 787, + 470 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.2 Comparison on Downstream Tasks", + "text_level": 1, + "bbox": [ + 215, + 494, + 547, + 510 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Now we evaluate features from different pre-trained models and the ensemble to demonstrate the effectiveness of SeA on downstream tasks. The comparison is summarized in Tab. 4. The linear model with fixed features from different models is denoted by the initial of the corresponding pre-trained model, i.e., S, M, B, and C stand for Supervised, MoCo, BYOL, and CoKe, respectively.", + "bbox": [ + 212, + 522, + 785, + 597 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "First, the performance of fine-tuning supervised and unsupervised models varies in different domains. For domains closely related to ImageNet (e.g., Birds, Flowers, Pets), the supervised pre-trained model shows better performance than the unsupervised ones, while unsupervised models surpass the supervised counterpart when the domain gap between pre-training and fine-tuning is large. The phenomenon indicates that representations from supervised/unsupervised models capture different patterns and suggests that combining the information from multiple models may handle downstream tasks better.", + "bbox": [ + 212, + 598, + 785, + 717 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Second, when having the fixed representation from only a single model with SeA, the performance is already close to fine-tuning on certain data sets, e.g., Caltech, DTD, and SUN. However, the information learned from a single model is biased toward the specific pre-training task. Without fine-tuning, the fixed representation only preserves limited patterns from the source domain, which is insufficient for different target domains.", + "bbox": [ + 212, + 719, + 785, + 808 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To mitigate the limitation of individual pre-training tasks, we consider combining the representations from multiple models, which can be done by a simple", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Q. Qian et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/311e31e13434adf83efec17da2d61fe189e6f19b2133a7d12f91bb64e7747baa.jpg", + "table_caption": [ + "Table 4: Comparison of accuracy (%) on downstream tasks. The best performance within each group is underlined while the global one is in bold." + ], + "table_footnote": [], + "table_body": "
MethodsAircraftBirdsCaltechCarsCIFAR10CIFAR100DTDFlowersFoodPetsSUN
Fine-tuning:
Supervised87.382.694.691.997.887.474.797.788.493.663.6
MoCo-v387.581.594.891.798.387.476.396.488.292.265.4
BYOL88.282.094.091.498.087.576.896.888.591.763.4
CoKe89.080.694.492.198.388.378.696.788.591.964.7
Fixed deep features with SeA:
S37.466.592.645.889.871.372.688.167.892.359.2
M61.571.495.469.793.479.076.494.674.990.764.2
B60.066.494.264.291.976.575.895.374.089.761.9
C57.867.094.363.493.078.577.895.676.191.065.3
S+M62.076.396.170.494.480.477.196.377.893.166.3
S+B60.773.795.767.593.479.177.996.377.493.465.5
S+C58.673.795.666.093.580.379.196.278.693.666.9
M+B65.772.595.772.794.680.878.296.277.691.965.5
M+C65.272.295.773.194.681.477.996.578.392.266.6
B+C64.070.195.270.193.980.778.896.478.391.166.4
S+M+B66.276.596.474.294.981.878.696.979.593.367.0
S+M+C65.876.796.373.195.181.978.896.979.993.667.7
S+B+C63.775.496.271.494.482.179.297.079.893.867.7
M+B+C67.273.295.974.795.282.378.996.879.492.067.2
S+M+B+C67.476.896.375.495.282.979.897.280.893.868.2
", + "bbox": [ + 220, + 184, + 785, + 470 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "feature concatenation operator. Compared with features from a single model, the accuracy of all downstream tasks is improved. When having fixed representations from all of 4 models, a linear classifier learned with SeA can outperform standard fine-tuning on 4 tasks (i.e., Caltech, DTD, Pets, and SUN) and achieve comparable performance on 2 tasks (i.e., CIFAR-10 and Flowers). It illustrates that collecting representations from multiple pre-training tasks can obtain complementary patterns that help mitigate the gap to different downstream tasks.", + "bbox": [ + 212, + 503, + 784, + 609 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Moreover, it can be observed that \"S+C\" performs better than \"S+M\" on DTD, while it is worse on Birds. The ensemble of all models, i.e., \"S+M+B+C\", shows the best performance on most tasks. Since the models in evaluation are pre-trained with different learning objectives, we find that the learned deep features complement each other. Therefore, more can be better. Note that ensemble can also be applied to fine-tuning. However, the computational cost will significantly increase depending on the number of models, which becomes unaffordable for real applications. In addition, its ensemble can still be worse than deep features. For example, an ensemble of fine-tuning on DTD achieves 78.9% that is about 1% worse than SeA.", + "bbox": [ + 212, + 611, + 785, + 762 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Although the result of SeA is promising, fixed features perform noticeably worse than fine-tuning on two tasks, e.g., Aircraft and Cars. It is because that the gap between the source domain and the target domain is too large. For example, ImageNet contains limited generic images for cars, while the target task is more fine-grained and consists of cars with different make, models, and", + "bbox": [ + 212, + 763, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Semantic Adversarial Augmentation for Last Layer Features", + "bbox": [ + 328, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "years. In this case, a larger model pre-trained on a larger data set covering more diverse domains (e.g., JFT-3B [38]) may help improve the performance of SeA.", + "bbox": [ + 215, + 146, + 784, + 176 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Comparison on Ensemble of Features Then, we illustrate the effectiveness of SeA by comparing it to the baseline without SeA and existing feature-space augmentation methods in Tab. 5, which includes the semantic augmentation method, i.e., ISDA [35] and a robust adversarial augmentation method, i.e., ME-ADA [40]. The experiment is conducted with the ensembled features from 4 models for all methods. The hyper-parameters in ISDA and ME-ADA are searched for the desired result.", + "bbox": [ + 215, + 196, + 785, + 301 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/356a98555fdc1a08f7e4e9a6ed6e377c23f85723c849f703f6bb1cce3a992ae6.jpg", + "table_caption": [ + "Table 5: Comparison of SeA and benchmark methods on $\\mathrm{S + M + B + C}$ . \"Gain\" denotes the improvement over the Baseline." + ], + "table_footnote": [], + "table_body": "
MethodsAircraftBirdsCaltechCarsCIFAR10CIFAR100DTDFlowersFoodPetsSUNAvg
Baseline63.673.495.469.294.881.179.496.279.593.166.981.1
ISDA63.373.695.969.494.981.379.695.679.693.267.081.2
ME-ADA64.474.995.770.995.081.679.895.880.393.267.481.7
SeA67.476.896.375.495.282.979.897.280.893.868.283.1
Gain3.83.40.96.20.41.80.41.01.30.71.32.0
", + "bbox": [ + 220, + 364, + 785, + 444 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "First, compared with the baseline, SeA improves the performance of deep features over all data sets, and the average gain is up to $2\\%$ as denoted by \"Gain\". The consistent improvement shows that augmentation is important for learning with fixed deep features and the proposed augmentation helps mitigate the overfitting problem effectively. Second, while both SeA and ISDA focus on semantic augmentation, ISDA relies on input space augmentations to obtain the semantic directions, while our method leverages features from other examples. Therefore, when input space augmentations are unavailable for deep features, ISDA cannot beat the baseline and it shows that our method is more appropriate for learning without input space augmentations. In addition, ME-ADA improves the vanilla adversarial augmentation by mitigating its sensitivity to the gradient direction. However, the optimization only considers the features from the target example. By incorporating the information from other examples in the same batch, the proposed method can outperform ME-ADA with a margin of $1.4\\%$ on average.", + "bbox": [ + 215, + 473, + 785, + 685 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 705, + 357, + 720 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "With the development of unsupervised representation learning, features with diverse information can be obtained by learning with different pretext tasks. Hence, we aim to investigate state-of-the-art performance of pre-trained deep features in this work. By introducing the novel semantic adversarial augmentation, we can show that learning with fixed features can achieve comparable performance to fine-tuning the whole network, but with way less cost. Investigating the efficacy of SeA in pre-training and fine-tuning can be our future work.", + "bbox": [ + 215, + 733, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Q. Qian et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bossard, L., Guillaumeun, M., Gool, L.V.: Food-101 - mining discriminative components with random forests. In: Fleet, D.J., Pajdla, T., Schiele, B., Tuytelaars, T. (eds.) ECCV. Lecture Notes in Computer Science, vol. 8694, pp. 446-461. Springer (2014)", + "2. Boyd, S., Boyd, S.P., Vandenberghe, L.: Convex optimization. Cambridge university press (2004)", + "3. Caron, M., Misra, I., Mairal, J., Goyal, P., Bojanowski, P., Joulin, A.: Unsupervised learning of visual features by contrasting cluster assignments. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) NeurIPS (2020)", + "4. Chen, T., Cheng, Y., Gan, Z., Wang, J., Wang, L., Liu, J., Wang, Z.: Adversarial feature augmentation and normalization for visual recognition. Trans. Mach. Learn. Res. 2022 (2022)", + "5. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.E.: A simple framework for contrastive learning of visual representations. In: ICML. Proceedings of Machine Learning Research, vol. 119, pp. 1597-1607. PMLR (2020)", + "6. Chen, X., Xie, S., He, K.: An empirical study of training self-supervised vision transformers. In: ICCV. pp. 9620-9629. IEEE (2021)", + "7. Cimpoi, M., Maji, S., Kokkinos, I., Mohamed, S., Vedaldi, A.: Describing textures in the wild. In: CVPR. pp. 3606-3613 (2014)", + "8. Crammer, K., Singer, Y.: On the algorithmic implementation of multiclass kernel-based vector machines. J. Mach. Learn. Res. 2, 265-292 (2001)", + "9. Donahue, J., Jia, Y., Vinyals, O., Hoffman, J., Zhang, N., Tzeng, E., Darrell, T.: Decaf: A deep convolutional activation feature for generic visual recognition. In: ICML. JMLR Workshop and Conference Proceedings, vol. 32, pp. 647-655. JMLR.org (2014)", + "0. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR. OpenReview.net (2021)", + "1. Fei-Fei, L., Fergus, R., Perona, P.: Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In: CVPR workshop. pp. 178-178. IEEE (2004)", + "2. Goodfellow, I.J., Shlens, J., Szegedy, C.: Explaining and harnessing adversarial examples. In: Bengio, Y., LeCun, Y. (eds.) ICLR (2015)", + "3. Grill, J., Strub, F., Altché, F., Tallec, C., Richemond, P.H., Buchatskaya, E., Doersch, C., Pires, B.A., Guo, Z., Azar, M.G., Piot, B., Kavukcuoglu, K., Munos, R., Valko, M.: Bootstrap your own latent - A new approach to self-supervised learning. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) NeurIPS (2020)", + "4. He, K., Chen, X., Xie, S., Li, Y., Dollar, P., Girshick, R.B.: Masked autoencoders are scalable vision learners. In: CVPR. pp. 15979-15988. IEEE (2022)", + "5. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.B.: Momentum contrast for unsupervised visual representation learning. In: CVPR. pp. 9726-9735. Computer Vision Foundation / IEEE (2020)", + "6. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778. IEEE Computer Society (2016)", + "7. Ioffe, S., Szegedy, C.: Batch normalization: Accelerating deep network training by reducing internal covariate shift. In: Bach, F.R., Blei, D.M. (eds.) ICML. JMLR Workshop and Conference Proceedings, vol. 37, pp. 448-456. JMLR.org (2015)" + ], + "bbox": [ + 225, + 172, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Semantic Adversarial Augmentation for Last Layer Features", + "bbox": [ + 328, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Kornblith, S., Shlens, J., Le, Q.V.: Do better imagenet models transfer better? In: CVPR. pp. 2661-2671. Computer Vision Foundation / IEEE (2019)", + "19. Krause, J., Stark, M., Deng, J., Fei-Fei, L.: 3d object representations for fine-grained categorization. In: ICCV workshop. pp. 554-561 (2013)", + "20. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)", + "21. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. In: Bartlett, P.L., Pereira, F.C.N., Burges, C.J.C., Bottou, L., Weinberger, K.Q. (eds.) NeurIPS. pp. 1106-1114 (2012)", + "22. Liu, Z., Xu, Y., Xu, Y., Qian, Q., Li, H., Ji, X., Chan, A.B., Jin, R.: Improved fine-tuning by better leveraging pre-training data. In: NeurIPS (2022)", + "23. Maji, S., Rahtu, E., Kannala, J., Blaschko, M., Vedaldi, A.: Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151 (2013)", + "24. Nilsback, M., Zisserman, A.: Automated flower classification over a large number of classes. In: ICVGIP. pp. 722-729. IEEE Computer Society (2008)", + "25. Parkhi, O.M., Vedaldi, A., Zisserman, A., Jawahar, C.: Cats and dogs. In: CVPR. pp. 3498-3505. IEEE (2012)", + "26. Qian, Q., Hu, J., Li, H.: Hierarchically robust representation learning. In: CVPR. pp. 7334-7342. Computer Vision Foundation / IEEE (2020)", + "27. Qian, Q., Jin, R., Zhu, S., Lin, Y.: Fine-grained visual categorization via multi-stage metric learning. In: CVPR. pp. 3716-3724. IEEE Computer Society (2015)", + "28. Qian, Q., Shang, L., Sun, B., Hu, J., Li, H., Jin, R.: Softtriple loss: Deep metric learning without triplet sampling. In: ICCV. pp. 6449-6457. IEEE (2019)", + "29. Qian, Q., Xu, Y., Hu, J., Li, H., Jin, R.: Unsupervised visual representation learning by online constrained k-means. In: CVPR. pp. 16619-16628. IEEE (2022)", + "30. Ren, S., He, K., Girshick, R.B., Sun, J.: Faster R-CNN: towards real-time object detection with region proposal networks. IEEE Trans. Pattern Anal. Mach. Intell. 39(6), 1137-1149 (2017)", + "31. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M.S., Berg, A.C., Fei-Fei, L.: Imagenet large scale visual recognition challenge. Int. J. Comput. Vis. 115(3), 211-252 (2015)", + "32. Verma, V., Lamb, A., Beckham, C., Najafi, A., Mitliagkas, I., Lopez-Paz, D., Bengio, Y.: Manifold mixup: Better representations by interpolating hidden states. In: Chaudhuri, K., Salakhutdinov, R. (eds.) ICML. Proceedings of Machine Learning Research, vol. 97, pp. 6438-6447. PMLR (2019)", + "33. Volpi, R., Morerio, P., Savarese, S., Murino, V.: Adversarial feature augmentation for unsupervised domain adaptation. In: CVPR. pp. 5495-5504. Computer Vision Foundation / IEEE Computer Society (2018)", + "34. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The caltech-ucsd birds-200-2011 dataset (2011)", + "35. Wang, Y., Huang, G., Song, S., Pan, X., Xia, Y., Wu, C.: Regularizing deep networks with semantic data augmentation. IEEE Trans. Pattern Anal. Mach. Intell. 44(7), 3733-3748 (2022)", + "36. Xiao, J., Hays, J., Ehinger, K.A., Oliva, A., Torralba, A.: SUN database: Large-scale scene recognition from abbey to zoo. In: CVPR. pp. 3485-3492. IEEE Computer Society (2010)", + "37. Xu, Y., Qian, Q., Li, H., Jin, R., Hu, J.: Weakly supervised representation learning with coarse labels. In: ICCV. pp. 10573-10581. IEEE (2021)", + "38. Zhai, X., Kolesnikov, A., Houlsby, N., Beyer, L.: Scaling vision transformers. In: CVPR. pp. 1204-1213. IEEE (2022)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Q. Qian et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "39. Zhang, H., Cissé, M., Dauphin, Y.N., Lopez-Paz, D.: mixup: Beyond empirical risk minimization. In: ICLR. OpenReview.net (2018)", + "40. Zhao, L., Liu, T., Peng, X., Metaxas, D.N.: Maximum-entropy adversarial data augmentation for improved generalization and robustness. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) NeurIPS (2020)" + ], + "bbox": [ + 215, + 146, + 784, + 217 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Semantic Adversarial Augmentation for Last Layer Features", + "bbox": [ + 328, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/SeA_ Semantic Adversarial Augmentation for Last Layer Features from Unsupervised Representation Learning/ca46bbef-cd36-48f1-b01d-aebb511e39a1_model.json b/2024/SeA_ Semantic Adversarial Augmentation for Last Layer Features from Unsupervised Representation Learning/ca46bbef-cd36-48f1-b01d-aebb511e39a1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..76c04055c49c9c40ae65a94b5acc8411d6f3ffb2 --- /dev/null +++ b/2024/SeA_ Semantic Adversarial Augmentation for Last Layer Features from Unsupervised Representation Learning/ca46bbef-cd36-48f1-b01d-aebb511e39a1_model.json @@ -0,0 +1,2379 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.245, + 0.141, + 0.761, + 0.21 + ], + "angle": 0, + "content": "SeA: Semantic Adversarial Augmentation for Last Layer Features from Unsupervised Representation Learning" + }, + { + "type": "text", + "bbox": [ + 0.329, + 0.235, + 0.674, + 0.253 + ], + "angle": 0, + "content": "Qi Qian\\(^{1}\\), Yuanhong \\(\\mathrm{Xu}^{2}\\), and Juhua \\(\\mathrm{Hu}^{3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.352, + 0.262, + 0.65, + 0.277 + ], + "angle": 0, + "content": "1 Alibaba Group, Bellevue, WA 98004, USA" + }, + { + "type": "text", + "bbox": [ + 0.375, + 0.278, + 0.621, + 0.29 + ], + "angle": 0, + "content": "2 Alibaba Group, Hangzhou, China" + }, + { + "type": "text", + "bbox": [ + 0.365, + 0.291, + 0.637, + 0.305 + ], + "angle": 0, + "content": "3 School of Engineering and Technology," + }, + { + "type": "text", + "bbox": [ + 0.328, + 0.306, + 0.676, + 0.319 + ], + "angle": 0, + "content": "University of Washington, Tacoma, WA 98402, USA" + }, + { + "type": "text", + "bbox": [ + 0.29, + 0.321, + 0.713, + 0.334 + ], + "angle": 0, + "content": "{qi.qian, yuanhong.xuyh}@alibaba-inc.com, juhuah@uw.edu" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.373, + 0.744, + 0.636 + ], + "angle": 0, + "content": "Abstract. Deep features extracted from certain layers of a pre-trained deep model show superior performance over the conventional hand-crafted features. Compared with fine-tuning or linear probing that can explore diverse augmentations, e.g., random crop/flipping, in the original input space, the appropriate augmentations for learning with fixed deep features are more challenging and have been less investigated, which degenerates the performance. To unleash the potential of fixed deep features, we propose a novel semantic adversarial augmentation (SeA) in the feature space for optimization. Concretely, the adversarial direction implied by the gradient will be projected to a subspace spanned by other examples to preserve the semantic information. Then, deep features will be perturbed with the semantic direction, and augmented features will be applied to learn the classifier. Experiments are conducted on 11 benchmark downstream classification tasks with 4 popular pre-trained models. Our method is \\(2\\%\\) better than the deep features without SeA on average. Moreover, compared to the expensive fine-tuning that is expected to give good performance, SeA shows a comparable performance on 6 out of 11 tasks, demonstrating the effectiveness of our proposal in addition to its efficiency." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.65, + 0.741, + 0.679 + ], + "angle": 0, + "content": "Keywords: Semantic augmentation \\(\\cdot\\) Deep features \\(\\cdot\\) Unsupervised representation learning \\(\\cdot\\) Self-supervised learning" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.706, + 0.376, + 0.723 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.739, + 0.788, + 0.816 + ], + "angle": 0, + "content": "Deep learning can be partially considered as a representation learning method that aims to extract features from raw data directly. By obtaining appropriate representations, a simple linear model can achieve state-of-the-art performance on challenging tasks, e.g., classification [16], object detection [30], etc. After the success of deep learning [21], researchers investigate the representations learned" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.383, + 0.84 + ], + "angle": 0, + "content": "* Corresponding author" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "Q. Qian et al." + }, + { + "type": "image", + "bbox": [ + 0.299, + 0.147, + 0.704, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.298, + 0.788, + 0.397 + ], + "angle": 0, + "content": "Fig. 1: Illustration of semantic adversarial augmentation (SeA). The red solid and empty circles denote the original data and its augmentation, respectively. Left: Conventional adversarial augmentation can perturb with arbitrary direction (e.g., augmentation may appear the same as the original); Right: SeA augments examples with semantic directions spanned by features from real data (e.g., different sectors show different subspaces for augmentation, where we can get more semantic meaningful augmentation)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.429, + 0.788, + 0.49 + ], + "angle": 0, + "content": "from a large-scale data set, i.e., ImageNet [31]. Surprisingly, the deep features extracted from a certain layer of a pre-trained model can outperform handcrafted features on various downstream tasks [9], demonstrating the efficacy of the data-dependent representation learning mechanism implied by deep learning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.491, + 0.788, + 0.627 + ], + "angle": 0, + "content": "Unlike hand-crafted features, representations obtained by deep learning are highly data/task-dependent as illustrated in [37]. When optimizing the task from the source domain, the neural network focuses on exploring the patterns related to the specific training task, while ignoring diverse information that has potential for different tasks. Hence, different representations can be learned from different training tasks even with the same training data from the source domain. Conventional methods learn representations with the labels of examples, which capture the knowledge only for the given labels and limit the information in deep features of each example." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.788, + 0.718 + ], + "angle": 0, + "content": "To mitigate the problem, fine-tuning parameters of the whole network becomes prevalent for various downstream tasks [18]. On one hand, fine-tuning can benefit from the prior knowledge in the pre-trained representations. On the other hand, it can further optimize representations with diverse augmentations on the target task. Consequently, fine-tuning works better than learning with fixed representations on downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "An important advantage of fine-tuning over learning with fixed deep features is the additional information from semantic augmentations. Obtaining effective augmentations in image space is convenient with semantic operators such as random crop, flipping, etc. In contrast, obtaining appropriate augmentations in the feature space for deep features becomes challenging, due to the lack of semantic operations. To leverage input space augmentations, a linear classifier can be learned with a frozen backbone by linear probing [6] that generates augmented images for optimization at each iteration." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.33, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Semantic Adversarial Augmentation for Last Layer Features" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.147, + 0.786, + 0.403 + ], + "angle": 0, + "content": "While fine-tuning and linear probing show promising performance, learning with fixed features is still attractive due to its good properties for real-world applications. First, it only applies deep models to extract features for each original example once, which is applicable for limited computational resources. On the contrary, fine-tuning has the full forward and backward pass and linear probing has the forward pass for each augmented example, which is more expensive for optimization. Second, the same models can be reused for different downstream tasks to extract deep features, while fine-tuning will adjust all parameters in pre-trained models. Thereafter, it has to keep a specific deep model for each task, which becomes intractable for handling hundreds of downstream tasks simultaneously. Finally, given fixed features, learning a linear classifier can be formulated as a convex problem that has the global optimum with a theoretical guarantee [2], while preserving the knowledge from the pre-trained model in features. It requires much less tuning efforts than fine-tuning that has to be tuned carefully to avoid the collapse of pre-trained parameters and catastrophic forgetting [22]. Therefore, we focus on fixed deep features extracted after the last pooling layer (i.e., inputs for the last fully-connected layer) in this work." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.404, + 0.784, + 0.48 + ], + "angle": 0, + "content": "Although some existing works [32,35] consider augmentation in intermediate layers to help optimize the whole network, little efforts were devoted to the deep features after the last pooling layer. Moreover, the existing feature space augmentation method [35] is proposed to complement the input space augmentation. It relies on other augmentation techniques and cannot work well solely." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.48, + 0.786, + 0.646 + ], + "angle": 0, + "content": "To tackle the problem, in this work, we propose a novel semantic adversarial augmentation strategy in the feature space of fixed deep features. Concretely, the gradient of each example is computed at first, and then a semantic direction can be observed by projecting the gradient to the subspace spanned by real data. Finally, the features of examples are augmented according to the obtained semantic direction for learning the linear classifier. Fig. 1 illustrates the proposed augmentation method that obtains effective augmentations with only fixed features. Moreover, to further mitigate the information loss in deep features from supervised pre-trained models, those from pre-trained models with different unsupervised pretext tasks are also investigated with an ensemble strategy. Our main contributions can be summarized as follows." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.658, + 0.787, + 0.703 + ], + "angle": 0, + "content": "- We empirically demonstrate the current performance gap between fine-tuning and learning with fixed deep features. Those deep features extracted from 4 representative pre-trained models are evaluated on 11 downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.705, + 0.787, + 0.763 + ], + "angle": 0, + "content": "- To improve the performance of deep features, we propose a semantic adversarial augmentation method to obtain appropriate augmentations tailored for fixed features. In addition, a smoothed hinge loss is investigated to demonstrate the augmentation direction explicitly." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.765, + 0.787, + 0.84 + ], + "angle": 0, + "content": "- Our proposed SeA gains \\(2\\%\\) accuracy on average over 11 downstream tasks compared to the baseline using deep features without augmentations. Moreover, compared with fine-tuning that is expected to give good performance, our recipe for learning with deep features can achieve comparable performance on 6 out of 11 tasks with way less computational cost, which shows" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.658, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "Q. Qian et al." + }, + { + "type": "text", + "bbox": [ + 0.243, + 0.147, + 0.785, + 0.178 + ], + "angle": 0, + "content": "the potential of self-supervised pre-trained deep features and demonstrates both the efficiency and effectiveness of our proposal." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.199, + 0.388, + 0.215 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.229, + 0.382, + 0.244 + ], + "angle": 0, + "content": "2.1 Deep Features" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.786, + 0.341 + ], + "angle": 0, + "content": "Modern deep neural networks consist of multiple layers to exact representations from raw materials. After that, a simple linear model encoded by a fully-connected layer can be attached on top of the representations for classification. After pre-training on a large-scale data set, the obtained neural network can be considered as a feature extractor and a new fully-connected layer can be learned for target tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.343, + 0.787, + 0.463 + ], + "angle": 0, + "content": "When the full label information of data is available, the learning objective is explicit and a supervised representation learning can be conducted by optimizing a conventional classification task, which is equivalent to optimizing the triplet loss as in distance metric learning [28]. The obtained deep features can outperform hand-crafted features on different applications, e.g., classification [9], distance metric learning [27], etc. Besides, some work aimed to obtain robust representations for different downstream tasks within the framework of classification [26]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.464, + 0.788, + 0.645 + ], + "angle": 0, + "content": "Without any label information, the learning objective varies in unsupervised learning. First, each instance can be considered as an individual class and the representations can be learned by instance discrimination [5, 15]. In addition, clustering can be applied to capture the relationship between different instances. A coarse-grained classification task defined on clusters can be leveraged to optimize representations [3, 29]. Finally, other pretext tasks beyond classification also demonstrate the effective representation after pre-training [13]. Compared with supervised representation learning, the different objectives in unsupervised learning can learn various semantic information even from the same data set. In this work, we will systematically study representations from different pre-trained models and illustrate the performance gap compared to fine-tuning the whole neural network." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.666, + 0.524, + 0.683 + ], + "angle": 0, + "content": "2.2 Augmentation in Feature Space" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.786, + 0.794 + ], + "angle": 0, + "content": "Due to the over-parameterization property of deep neural networks, augmentation is essential for training deep models to avoid over-fitting [21]. Given an image, a perturbed copy can be observed by standard image operations such as random crop, flipping, color jitter, etc. [16]. The augmented images have a large variance while preserving the semantic information of the original image, i.e., an image of a cat is still a cat after augmentation, which helps train effective models with the additional information." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.787, + 0.84 + ], + "angle": 0, + "content": "However, augmentation in feature space becomes more challenging due to the lack of semantic preserving operators. Some methods consider the augmentation of features from intermediate layers to help train the whole deep neural" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.33, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Semantic Adversarial Augmentation for Last Layer Features" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.389 + ], + "angle": 0, + "content": "networks in different applications [4, 32, 33, 35]. For example, [33] aims to learn feature augmentation with a feature generator for unsupervised domain adaptation. [4] tries to augment intermediate feature maps with adversarial feature moments in batch normalization [17] for efficient training. Manifold mixup [32] proposes to apply the original mixup [39] to features from multiple layers for augmentation, which also includes the layer investigated in this work. However, their study shows that the gain of mixup is mainly from the original image space and feature space of early stages, while that from the feature space of the last layer is negligible, which is consistent with our observation in the ablation study. ISDA [35] considers the semantic augmentation in feature space but it has to obtain the semantic direction with input space augmentations, which is complementary to input space augmentations but cannot work as the sole augmentation for deep features well. On the contrary, we propose SeA to project the adversarial direction with features of real data points. Moreover, our proposal is tailored for fixed features and is different from existing works that are for optimizing the whole network." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.413, + 0.603, + 0.43 + ], + "angle": 0, + "content": "3 Semantic Adversarial Augmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.445, + 0.788, + 0.552 + ], + "angle": 0, + "content": "We start the analysis with a standard classification framework. Let \\(\\{x_{i},y_{i}\\}_{i = 1}^{n}\\) denote the training data set and \\(\\{f_k\\}_{k = 1}^K\\) contains a set of \\(K\\) pre-trained models, which can be pre-trained with different learning objectives on different data sets. In this work, we directly use the representations extracted from pre-trained deep models and learn a simple linear model on top of it. Representations from different deep models are concatenated as the final representation for each example. Given \\(x_{i}\\), the representation will be extracted as" + }, + { + "type": "equation", + "bbox": [ + 0.391, + 0.563, + 0.786, + 0.581 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {i} = \\left[ f _ {1} \\left(x _ {i}\\right), \\dots , f _ {K} \\left(x _ {i}\\right) \\right] \\in \\mathbb {R} ^ {d} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.592, + 0.787, + 0.621 + ], + "angle": 0, + "content": "where \\( d = \\sum_{k}d_{k} \\) and \\( d_{k} \\) is the dimension of the representation from the \\( k \\)-th pre-trained deep model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.623, + 0.787, + 0.653 + ], + "angle": 0, + "content": "With the above fixed deep features, a classification model can be learned by minimizing the empirical risk with the appropriate regularization as" + }, + { + "type": "equation", + "bbox": [ + 0.394, + 0.664, + 0.786, + 0.697 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\mathbf {w}} \\sum_ {i} \\ell (\\mathbf {x} _ {i}, y _ {i}; \\mathbf {w}) + \\frac {\\tau}{2} \\| \\mathbf {w} \\| _ {F} ^ {2} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.707, + 0.787, + 0.752 + ], + "angle": 0, + "content": "where \\(\\tau\\) is the weight for \\(L_{2}\\) regularization and \\(\\mathbf{w}\\) denotes parameters of the classification model, which is a linear classifier with the prediction probability of the \\(i\\)-th example on the \\(j\\)-th class as" + }, + { + "type": "equation", + "bbox": [ + 0.411, + 0.763, + 0.786, + 0.801 + ], + "angle": 0, + "content": "\\[\np _ {i, j} = \\frac {\\exp \\left(\\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {j}\\right)}{\\sum_ {c = 1} ^ {C} \\exp \\left(\\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {c}\\right)} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "\\(C\\) is the number of classes and \\(\\ell\\) is a loss function for learning that will be discussed later." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.129 + ], + "angle": 0, + "content": "Q. Qian et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.421, + 0.16 + ], + "angle": 0, + "content": "3.1 Semantic Direction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.171, + 0.788, + 0.291 + ], + "angle": 0, + "content": "In a standard fine-tuning pipeline, an image can be augmented by multiple random perturbations. When only deep features are available, it is hard to adopt existing techniques to generate appropriate augmented examples, since the semantic direction in the feature space is hard to capture. Manifold mixup [32] shows that randomly augmenting features from the last output layers has little improvement compared to the augmentation in input image space. To facilitate the performance of deep features, we investigate the semantic adversarial direction for augmentation." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.292, + 0.787, + 0.323 + ], + "angle": 0, + "content": "First, for the \\(i\\)-th example and its representation \\(\\mathbf{x}_i\\), we consider its adversarial direction that can be obtained by maximizing the loss function [12]" + }, + { + "type": "equation", + "bbox": [ + 0.417, + 0.333, + 0.786, + 0.357 + ], + "angle": 0, + "content": "\\[\n\\max _ {\\mathbf {x}: \\left\\| \\mathbf {x} - \\mathbf {x} _ {i} \\right\\| _ {2} \\leq \\gamma} \\ell \\left(\\mathbf {x} _ {i}, y _ {i}; \\mathbf {w}\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.366, + 0.787, + 0.409 + ], + "angle": 0, + "content": "Note that the gradient indicates the ideal direction for the adversarial perturbation, and a standard adversarial example can be obtained by gradient ascent as" + }, + { + "type": "equation", + "bbox": [ + 0.439, + 0.422, + 0.785, + 0.438 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {x}} _ {i} = \\mathbf {x} _ {i} + \\eta \\nabla_ {\\mathbf {x} _ {i}} \\ell \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.788, + 0.553 + ], + "angle": 0, + "content": "However, the gradient direction may not be semantically informative. [12] shows that the loss can be substantially increased with the adversarial example generated from the gradient, while the appearance is almost the same as the original image as illustrated in Fig. 1. Therefore, the conventional adversarial learning method helps improve the robustness to the adversarial attack but may be inappropriate for regularizing the generic learning task that requires additional information from diverse images." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.554, + 0.788, + 0.614 + ], + "angle": 0, + "content": "To obtain the semantic direction, we propose to project the gradient direction to the subspace spanned by real data points. Concretely, let \\(\\{\\mathbf{x}_j\\}_{j=1}^b\\) denote a mini-batch of data and \\(\\mathbf{g}_i = \\nabla_{\\mathbf{x}_i}\\ell\\) indicate the gradient of \\(\\mathbf{x}_i\\). For the \\(i\\)-th image, a semantic adversarial direction \\(\\hat{\\mathbf{g}}_i\\) consists of representations from data as" + }, + { + "type": "equation", + "bbox": [ + 0.355, + 0.624, + 0.786, + 0.656 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {g}} _ {i} = \\sum_ {j: j \\neq i} q _ {j} \\mathbf {x} _ {j}; \\quad \\mathbf {q} ^ {*} = \\arg \\min _ {\\mathbf {q} \\in \\Delta} \\mathcal {D} (\\hat {\\mathbf {g}} _ {i}, \\mathbf {g} _ {i}) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.667, + 0.788, + 0.712 + ], + "angle": 0, + "content": "where \\(\\mathcal{D}(\\cdot, \\cdot)\\) is a distance function and \\(\\varDelta\\) is a simplex as \\(\\varDelta=\\{\\mathbf{q}\\in\\mathbb{R}^{b-1}|\\sum_{j}q_{j}=1;\\forall j,q_{j}\\geq 0\\}\\). With the projection, we aim to find an adversarial direction in a subspace consisting of original data points as illustrated in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.713, + 0.787, + 0.759 + ], + "angle": 0, + "content": "The last challenge is to obtain \\(\\mathbf{q}\\) efficiently with a distance function. By adopting the squared Euclidean distance, which is a standard distance measurement, the optimization problem for \\(\\mathbf{q}\\) can be written as" + }, + { + "type": "equation", + "bbox": [ + 0.384, + 0.769, + 0.786, + 0.801 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\mathbf {q} \\in \\Delta} \\| \\sum_ {j: j \\neq i} q _ {j} \\mathbf {x} _ {j} - \\mathbf {g} _ {i} \\| _ {2} ^ {2} - \\alpha H (\\mathbf {q}) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "where \\( H(\\mathbf{q}) \\) is the entropy of \\( \\mathbf{q} \\) that helps improve the robustness to different batches. When normalizing features such that \\( \\| \\mathbf{x}_j\\| _2 = \\| \\mathbf{g}_i\\| _2 = 1 \\) to obtain" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.33, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Semantic Adversarial Augmentation for Last Layer Features" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "code_caption", + "bbox": [ + 0.218, + 0.147, + 0.774, + 0.163 + ], + "angle": 0, + "content": "Algorithm 1 Semantic Adversarial Augmentation (SeA) for Given Features" + }, + { + "type": "algorithm", + "bbox": [ + 0.223, + 0.167, + 0.787, + 0.294 + ], + "angle": 0, + "content": "1: Input: Dataset \\(\\{x_i,y_i\\}\\), pre-trained models \\(\\{f_k\\}\\), iterations \\(T\\), linear model \\(\\mathbf{w}\\), \\(\\alpha\\), \\(\\tau\\), \\(\\eta\\) \n2: Extract deep features by \\(\\{f_k\\}\\) for all examples as \\(\\{\\mathbf{x}_i\\}\\) \n3: for \\(t = 1,\\dots,T\\) do \n4: Receive a mini-batch of examples \\(\\{\\mathbf{x}_i,y_i\\}_{i=1}^b\\) \n5: Obtain augmented examples \\(\\{\\tilde{\\mathbf{x}}_i,y_i\\}\\) as in Eqn. 11 \n6: Optimize \\(\\mathbf{w}\\) by SGD: \\(\\mathbf{w} = \\mathbf{w} - \\eta_w\\left(\\frac{1}{b}\\sum_i^b\\nabla_{\\mathbf{w}}\\ell(\\tilde{\\mathbf{x}}_i,y_i;\\mathbf{w}) + \\tau \\mathbf{w}\\right)\\) \n7: end for \n8: return \\(\\mathbf{w}\\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.323, + 0.784, + 0.353 + ], + "angle": 0, + "content": "the direction without the influence from the magnitude, the problem can be upper-bounded as" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.361, + 0.684, + 0.379 + ], + "angle": 0, + "content": "Proposition 1 With unit length variables \\(\\{\\mathbf{x}_j\\}\\) and \\(\\mathbf{g}_i\\), we have" + }, + { + "type": "equation", + "bbox": [ + 0.362, + 0.387, + 0.786, + 0.42 + ], + "angle": 0, + "content": "\\[\n\\left\\| \\sum_ {j: j \\neq i} q _ {j} \\mathbf {x} _ {j} - \\mathbf {g} _ {i} \\right\\| _ {2} ^ {2} \\leq 2 - 2 \\sum_ {j} q _ {j} \\mathbf {x} _ {j} ^ {\\top} \\mathbf {g} _ {i} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.429, + 0.61, + 0.444 + ], + "angle": 0, + "content": "The detailed proof can be found in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.444, + 0.784, + 0.473 + ], + "angle": 0, + "content": "By rearranging the terms, we can maximize the lower-bound of the original problem as" + }, + { + "type": "equation", + "bbox": [ + 0.401, + 0.483, + 0.786, + 0.517 + ], + "angle": 0, + "content": "\\[\n\\max _ {\\mathbf {q} \\in \\Delta} \\sum_ {j: j \\neq i} q _ {j} \\mathbf {x} _ {j} ^ {\\top} \\mathbf {g} _ {i} + \\alpha H (\\mathbf {q}) \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.526, + 0.729, + 0.542 + ], + "angle": 0, + "content": "According to the K.K.T. condition [2], \\(\\mathbf{q}\\) has a closed-form solution." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.55, + 0.7, + 0.565 + ], + "angle": 0, + "content": "Proposition 2 The problem in Eqn. 9 has the optimal solution as" + }, + { + "type": "equation", + "bbox": [ + 0.334, + 0.573, + 0.786, + 0.613 + ], + "angle": 0, + "content": "\\[\nq _ {j} = \\frac {\\exp \\left(\\mathbf {x} _ {j} ^ {\\top} \\mathbf {g} _ {i} / \\alpha\\right)}{Z}; \\quad Z = \\sum_ {k: k \\neq i} \\exp \\left(\\mathbf {x} _ {k} ^ {\\top} \\mathbf {g} _ {i} / \\alpha\\right) \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.623, + 0.784, + 0.654 + ], + "angle": 0, + "content": "Given the semantic adversarial direction, the target example with semantic perturbation can be obtained as" + }, + { + "type": "equation", + "bbox": [ + 0.417, + 0.664, + 0.786, + 0.681 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {x}} _ {i} = \\Pi \\left(\\mathbf {x} _ {i} + \\eta \\Pi \\left(\\hat {\\mathbf {g}} _ {i}\\right)\\right) \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.786, + 0.766 + ], + "angle": 0, + "content": "where \\(\\Pi(\\cdot)\\) normalizes the vector to the unit length if required and \\(\\eta\\) denotes the step size for augmentation. Compared with the adversarial perturbation in Eqn. 5, the augmentation in Eqn. 11 projects the gradient to the direction consisting of real data points, which can capture the semantic information in the feature space effectively. Alg. 1 summarizes the proposed method." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.786, + 0.559, + 0.801 + ], + "angle": 0, + "content": "3.2 Illustration of Adversarial Direction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "In this subsection, we will generalize the cross entropy loss to help illustrate the proposed augmentation strategy." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.13 + ], + "angle": 0, + "content": "Q. Qian et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.784, + 0.176 + ], + "angle": 0, + "content": "The standard multi-class hinge loss [8] for convex optimization can be written as" + }, + { + "type": "equation", + "bbox": [ + 0.333, + 0.189, + 0.785, + 0.214 + ], + "angle": 0, + "content": "\\[\n\\ell (\\mathbf {x} _ {i}, y _ {i}; \\mathbf {w}) = \\max \\left\\{0, \\delta + \\max _ {c \\neq y _ {i}} \\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {c} - \\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {y _ {i}} \\right\\} \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.225, + 0.45, + 0.24 + ], + "angle": 0, + "content": "where \\(\\delta\\) is a pre-defined margin." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.24, + 0.784, + 0.3 + ], + "angle": 0, + "content": "Unlike convex optimization, cross-entropy loss is prevalent in deep learning, which is a smooth function and can help accelerate the convergence [2]. Therefore, we propose to obtain a smoothed hinge loss by introducing a distribution over logits from different classes \\(\\mathbf{p} \\in \\Delta\\) where \\(\\Delta\\) is the simplex." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.302, + 0.561, + 0.317 + ], + "angle": 0, + "content": "First, the original hinge loss is equivalent to" + }, + { + "type": "equation", + "bbox": [ + 0.291, + 0.329, + 0.784, + 0.361 + ], + "angle": 0, + "content": "\\[\n\\ell \\left(\\mathbf {x} _ {i}, y _ {i}; \\mathbf {w}\\right) = \\max _ {\\mathbf {p} \\in \\Delta} p _ {y _ {i}} \\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {y _ {i}} + \\sum_ {c \\neq y _ {i}} p _ {c} \\left(\\delta + \\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {c}\\right) - \\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {y _ {i}} \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.373, + 0.784, + 0.403 + ], + "angle": 0, + "content": "According to the analysis for cross entropy loss [28], the loss can be smoothed by adding an entropy regularization for \\(\\mathbf{p}\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.248, + 0.416, + 0.784, + 0.448 + ], + "angle": 0, + "content": "\\[\n\\ell \\left(\\mathbf {x} _ {i}, y _ {i}; \\mathbf {w}\\right) = \\max _ {\\mathbf {p} \\in \\Delta} p _ {y _ {i}} \\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {y _ {i}} + \\sum_ {c \\neq y _ {i}} p _ {c} (\\delta + \\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {c}) + \\lambda H (\\mathbf {p}) - \\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {y _ {i}} \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.46, + 0.657, + 0.476 + ], + "angle": 0, + "content": "where \\(H(\\mathbf{p})\\) denotes the entropy of \\(\\mathbf{p}\\) and \\(\\lambda\\) is the coefficient." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.489, + 0.657, + 0.505 + ], + "angle": 0, + "content": "Proposition 3 The loss function in Eqn. 14 is equivalent to" + }, + { + "type": "equation", + "bbox": [ + 0.26, + 0.516, + 0.784, + 0.553 + ], + "angle": 0, + "content": "\\[\n\\ell (\\mathbf {x} _ {i}, y _ {i}; \\mathbf {w}) = - \\lambda \\log \\frac {\\exp \\left(\\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {y _ {i}} / \\lambda\\right)}{\\exp \\left(\\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {y _ {i}} / \\lambda\\right) + \\sum_ {c \\neq y _ {i}} \\exp \\left(\\left(\\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {c} + \\delta\\right) / \\lambda\\right)} \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.564, + 0.784, + 0.608 + ], + "angle": 0, + "content": "Remark It is obvious that the popular cross entropy loss is a special case of Eqn. 15 by letting \\(\\delta = 0\\) and \\(\\lambda = 1\\). Our analysis connects the hinge loss in conventional methods to the popular loss function in deep learning." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.609, + 0.784, + 0.639 + ], + "angle": 0, + "content": "Finally, since deep features are fixed, we can illustrate the gradient direction with the proposed smoothed loss function explicitly." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.65, + 0.741, + 0.666 + ], + "angle": 0, + "content": "Proposition 4 Given the loss function in Eqn. 15, the gradient of \\(\\mathbf{x}_i\\) is" + }, + { + "type": "equation", + "bbox": [ + 0.381, + 0.678, + 0.784, + 0.719 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\mathbf {x} _ {i}} \\ell (\\mathbf {x} _ {i}, y _ {i}; \\mathbf {w}) = \\sum_ {c = 1} ^ {C} p _ {c} \\mathbf {w} _ {c} - \\mathbf {w} _ {y _ {i}} \\tag {16}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.216, + 0.73, + 0.8, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{r} w h e r e p _ {c} = \\left\\{ \\begin{array}{l l} \\frac {\\exp ((\\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {c} + \\delta) / \\lambda)}{Z} c \\neq y _ {i} \\\\ \\frac {\\exp (\\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {c} / \\lambda)}{Z} c = y _ {i} \\end{array} a n d Z = \\exp (\\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {y _ {i}} / \\lambda) + \\sum_ {c \\neq y _ {i}} \\exp ((\\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {c} + \\delta) / \\lambda) \\leqslant 1, \\right. \\\\ \\delta) / \\lambda). \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.795, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Remark The gradient direction in the feature space indicates that an adversarial perturbation should be close to the primary directions of other classes, while far away from the direction of the corresponding class." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.33, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Semantic Adversarial Augmentation for Last Layer Features" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.377, + 0.163 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.181, + 0.788, + 0.364 + ], + "angle": 0, + "content": "To demonstrate our proposed method, we adopt state-of-the-art and widely used pre-trained models to extract deep features for evaluation. It should be noted that only two main architectures, including ResNet-50 [16] and vision transformer (ViT) [10], are prevalently applied in self-supervised learning [6,13,14,29]. Since ViT shows a worse classification performance than ResNet-50 with the frozen backbone as demonstrated in [14], we will focus on ResNet-50 with different public pre-trained parameters in the experiment. Specifically, one supervised pre-trained ResNet-50 and three self-supervised pre-trained ResNet-50 are applied for feature extraction. All of these models are pre-trained on ImageNet-1K [31] with different learning objectives. The details of different models can be found in the supplementary. The accuracy of the supervised model and that of linear probing for unsupervised models are summarized in Tab. 1." + }, + { + "type": "table_caption", + "bbox": [ + 0.299, + 0.39, + 0.704, + 0.404 + ], + "angle": 0, + "content": "Table 1: Performance of 4 pre-trained models on ImageNet." + }, + { + "type": "table", + "bbox": [ + 0.257, + 0.416, + 0.747, + 0.492 + ], + "angle": 0, + "content": "
Model#Pre-training EpochsObjectiveAcc%
Supervised [16] (S)901,000-class classification76.6
MoCo-v3 [6] (M)1,000instance discrimination74.6
BYOL [13] (B)1,000regression74.3
CoKe [29] (C)1,000cluster discrimination74.9
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.522, + 0.788, + 0.658 + ], + "angle": 0, + "content": "11 diverse downstream data sets are applied for evaluation, including Aircraft [23], Caltech101 [11], Stanford Cars [19], CIFAR-10 [20], CIFAR-100 [20], CUB200-2011 (Birds) [34], Describable Textures Dataset (DTD) [7], Flowers [24], Food101 [1], Oxford-IIIT Pet (Pets) [25], and Sun397 [36]. We follow the evaluation protocol in [13]. Concretely, all models search hyper-parameters on the provided/generated validation set in each downstream task, and the standard metric on the provided test set is reported. Besides, mean per-class accuracy is reported on Aircraft, Caltech101, Flowers, and Pets, while Top-1 accuracy is utilized for other data sets. More details can be found in [13]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.75 + ], + "angle": 0, + "content": "Each model is fine-tuned with SGD by 100 epochs for sufficient training. The batch size is 256 and the momentum is 0.9. The learning rate is searched in a range of 7 logarithmically-spaced values between \\(10^{-4}\\) and \\(10^{-1}\\). Weight decay is optional, for which if it is applied, the value will be searched with the same setting between \\(10^{-6}\\) and \\(10^{-3}\\). The standard augmentation, i.e., random crop, random horizontal flipping, is applied as in most existing fine-tuning pipelines." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "For a fair comparison, we also apply SGD to learn the linear classifier with fixed deep features using the same batch size and momentum. Unlike fine-tuning, we have the constant learning rate for our method, which is searched in \\(\\{2^i\\}_{i = -2}^{i = 3}\\) while weight decay is searched in \\(\\{0,10^{-6},10^{-5},10^{-4}\\}\\). For other parameters, we have \\(\\alpha\\), \\(\\lambda\\), \\(\\delta\\) and \\(\\eta\\) searched in \\(\\{0.01,0.02,0.05\\}\\), \\(\\{0.05,0.1,1\\}\\), \\(\\{0,1\\}\\), and \\(\\{0,0.2,0.4,0.8\\}\\), respectively." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "Q. Qian et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.283 + ], + "angle": 0, + "content": "For each image, we extract the output of the final pooling layer as deep features. Features from an individual model are normalized to the unit length. If multiple models are exploited, features from different models are concatenated as in Eqn. 1 and the combined feature vector is further normalized to the unit length. The only augmentation for training with fixed features is the proposed semantic adversarial augmentation. To approximate the direction of the gradient for each example, a mini-batch of data is leveraged to optimize the problem in Eqn. 9. The projection for normalization is also applied to the augmented examples." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.31, + 0.39, + 0.325 + ], + "angle": 0, + "content": "4.1 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.34, + 0.787, + 0.401 + ], + "angle": 0, + "content": "Before the experiments on downstream tasks, we first investigate the effect of parameters in the proposed augmentation. The ablation study is conducted on CIFAR-100 with features concatenated from 4 pre-trained models and the accuracy is reported for comparison." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.428, + 0.787, + 0.504 + ], + "angle": 0, + "content": "Effect of Step Size for Augmentation First, we evaluate the step size \\(\\eta\\) in the semantic adversarial augmentation. For SeA, we fix \\(\\lambda = 0.1\\), \\(\\delta = 0\\), \\(\\alpha = 0.01\\), and the learning rate as 1, according to the performance on the validation set. The weight of the augmentation \\(\\eta\\) varies in \\(\\{0, 0.2, 0.4, 0.6, 0.8, 1.0\\}\\) and the performance is summarized in Fig. 2-Fig. 3." + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.53, + 0.402, + 0.635 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.403, + 0.53, + 0.581, + 0.634 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.533, + 0.771, + 0.633 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.644, + 0.784, + 0.673 + ], + "angle": 0, + "content": "Fig. 2: Training accuracy Fig. 3: Test accuracy of Fig. 4: Test accuracy of various different \\(\\eta\\). different \\(\\eta\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.841 + ], + "angle": 0, + "content": "When \\(\\eta = 0\\), there is no augmentation applied for training. Hence, we can observe that the training accuracy exceeds \\(92\\%\\) after only 20 epochs. Due to the over-fitting on the training set, the test accuracy of this baseline is only \\(81.1\\%\\) on CIFAR-100. By gradually increasing the step size for the augmentation, the training accuracy is decreased as expected, while the test accuracy increases. It demonstrates that the proposed semantic adversarial augmentation strategy can effectively mitigate the over-fitting problem for learning with fixed deep features. By setting an appropriate step size, the test accuracy can be improved to \\(82.9\\%\\), which helps shrink the gap to fine-tuning." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.33, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Semantic Adversarial Augmentation for Last Layer Features" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.193 + ], + "angle": 0, + "content": "Effect of Directions for Augmentation Different directions can be adopted for generating augmentation. Besides the proposed semantic adversarial aug-. mentations (SeA), four variants are included in the comparison." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.204, + 0.733, + 0.218 + ], + "angle": 0, + "content": "- Adv: the original gradient direction without projection as in Eqn. 5." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.219, + 0.784, + 0.251 + ], + "angle": 0, + "content": "- SeA_Neg: similar to the gradient direction in Proposition 4 but getting rid of the direction from the target class and keeping the direction of \\(\\sum_{c}^{C}p_{c}\\mathbf{w}_{c}\\)." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.253, + 0.784, + 0.281 + ], + "angle": 0, + "content": "- Rand: a uniformly random direction within subspace spanned by a minibatch of data, i.e., generating \\(\\mathbf{q}\\) in Eqn. 6 as a random vector." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.282, + 0.784, + 0.297 + ], + "angle": 0, + "content": "- Manifold mixup: a random semantic direction indicated by an example [32]." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.204, + 0.784, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.308, + 0.797, + 0.338 + ], + "angle": 0, + "content": "The step size for Adv is very sensitive and is searched in \\(\\{0.005, 0.01, 0.02, 0.05\\}\\), while the step size for others is searched in \\(\\{0.2, 0.4, 0.8, 1.0\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.339, + 0.789, + 0.611 + ], + "angle": 0, + "content": "Fig. 4 shows the comparison on CIFAR-100. First, we can observe that most augmentation directions except the random one can improve the performance over the baseline without augmentation. It confirms that augmentation is important for better generalization even for learning from fixed representations. Then, the method with the random augmentation direction performs similarly to the baseline, which shows that it is challenging to find an effective direction for augmentation in feature space. However, with the original gradient direction directly as in Adv, the training loss will be significantly increased even with a small step size and only a step size that is much smaller than SeA can obtain an applicable model. This is because that the gradient direction only aims to maximize the loss while ignoring the data distribution in the feature space. By projecting the gradient to a data-dependent subspace, SeA and SeA_Neg can achieve a better performance than the baseline without a sophisticated setup for the step size. Moreover, by removing the direction from the target class, the accuracy is degenerated by about \\(1\\%\\). The comparison demonstrates that keeping the completed gradient direction is essential for obtaining appropriate augmentation. Finally, a random semantic direction indicated by examples as in mixup is worse than SeA with a clear margin." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.634, + 0.788, + 0.709 + ], + "angle": 0, + "content": "Comparison with Linear Probing Besides learning with fixed features, linear probing is another way to obtain the linear classifier with a frozen backbone. Compared with our method, it can obtain appropriate augmentations in image space. However, it has to forward each augmentation for learning and the cost is much higher than us." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.711, + 0.788, + 0.786 + ], + "angle": 0, + "content": "To illustrate the effectiveness of the proposed augmentation method, we compare the performance between linear probing and our proposal in Tab. 2. For a fair comparison, all methods optimize features from a single model of MoCo. It shows that the proposed augmentation can outperform the augmentation in the input image space for learning the target classifier." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Running Time Finally, we compare the running time of fine-tuning, linear probing, and SeA using fixed deep features in Tab. 3. Note that with features" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "Q. Qian et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.23, + 0.145, + 0.773, + 0.187 + ], + "angle": 0, + "content": "Table 2: Comparison of accuracy (%) Table 3: Comparison of running time. SeA on CIFAR-100. All methods utilize a utilize deep features from 4 models while single model from MoCo. other methods optimize only a single model." + }, + { + "type": "table", + "bbox": [ + 0.264, + 0.2, + 0.763, + 0.231 + ], + "angle": 0, + "content": "
baselinelinear probingSeA
77.677.879.0
Methodsfine-tuninglinear probing
Time (s)14,3834,437
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.261, + 0.788, + 0.472 + ], + "angle": 0, + "content": "from 4 models, the dimension of the feature vector is 8, 192. However, optimizing the combined features is still much more efficient than fine-tuning a single model or linear probing with 2,048 final features. All methods train the corresponding model by 100 epochs and the running time is measured on a single V100 GPU. Fine-tuning costs 14,383 (sec) for the entire training. In contrast, SeA only costs 40 (sec) to obtain the optimal linear model. The efficiency of SeA using deep features implies its applicability for limited-resource scenarios. With efficient optimization, we can repeat experiments of SeA by multiple times. The standard deviation is only \\(0.03\\%\\) on CIFAR-100 over 5 runs, which shows the stability of convex optimization. Although linear probing eliminates the cost of the backward pass, that of the forward pass at each iteration is still expensive. Considering that linear probing has a similar running cost as fine-tuning but with worse performance, it will not be included in the comparison over other downstream tasks. More ablation experiments can be found in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.496, + 0.548, + 0.511 + ], + "angle": 0, + "content": "4.2 Comparison on Downstream Tasks" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.787, + 0.598 + ], + "angle": 0, + "content": "Now we evaluate features from different pre-trained models and the ensemble to demonstrate the effectiveness of SeA on downstream tasks. The comparison is summarized in Tab. 4. The linear model with fixed features from different models is denoted by the initial of the corresponding pre-trained model, i.e., S, M, B, and C stand for Supervised, MoCo, BYOL, and CoKe, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.787, + 0.718 + ], + "angle": 0, + "content": "First, the performance of fine-tuning supervised and unsupervised models varies in different domains. For domains closely related to ImageNet (e.g., Birds, Flowers, Pets), the supervised pre-trained model shows better performance than the unsupervised ones, while unsupervised models surpass the supervised counterpart when the domain gap between pre-training and fine-tuning is large. The phenomenon indicates that representations from supervised/unsupervised models capture different patterns and suggests that combining the information from multiple models may handle downstream tasks better." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.787, + 0.809 + ], + "angle": 0, + "content": "Second, when having the fixed representation from only a single model with SeA, the performance is already close to fine-tuning on certain data sets, e.g., Caltech, DTD, and SUN. However, the information learned from a single model is biased toward the specific pre-training task. Without fine-tuning, the fixed representation only preserves limited patterns from the source domain, which is insufficient for different target domains." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "To mitigate the limitation of individual pre-training tasks, we consider combining the representations from multiple models, which can be done by a simple" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.33, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Semantic Adversarial Augmentation for Last Layer Features" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.785, + 0.174 + ], + "angle": 0, + "content": "Table 4: Comparison of accuracy (%) on downstream tasks. The best performance within each group is underlined while the global one is in bold." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.185, + 0.786, + 0.471 + ], + "angle": 0, + "content": "
MethodsAircraftBirdsCaltechCarsCIFAR10CIFAR100DTDFlowersFoodPetsSUN
Fine-tuning:
Supervised87.382.694.691.997.887.474.797.788.493.663.6
MoCo-v387.581.594.891.798.387.476.396.488.292.265.4
BYOL88.282.094.091.498.087.576.896.888.591.763.4
CoKe89.080.694.492.198.388.378.696.788.591.964.7
Fixed deep features with SeA:
S37.466.592.645.889.871.372.688.167.892.359.2
M61.571.495.469.793.479.076.494.674.990.764.2
B60.066.494.264.291.976.575.895.374.089.761.9
C57.867.094.363.493.078.577.895.676.191.065.3
S+M62.076.396.170.494.480.477.196.377.893.166.3
S+B60.773.795.767.593.479.177.996.377.493.465.5
S+C58.673.795.666.093.580.379.196.278.693.666.9
M+B65.772.595.772.794.680.878.296.277.691.965.5
M+C65.272.295.773.194.681.477.996.578.392.266.6
B+C64.070.195.270.193.980.778.896.478.391.166.4
S+M+B66.276.596.474.294.981.878.696.979.593.367.0
S+M+C65.876.796.373.195.181.978.896.979.993.667.7
S+B+C63.775.496.271.494.482.179.297.079.893.867.7
M+B+C67.273.295.974.795.282.378.996.879.492.067.2
S+M+B+C67.476.896.375.495.282.979.897.280.893.868.2
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.505, + 0.785, + 0.61 + ], + "angle": 0, + "content": "feature concatenation operator. Compared with features from a single model, the accuracy of all downstream tasks is improved. When having fixed representations from all of 4 models, a linear classifier learned with SeA can outperform standard fine-tuning on 4 tasks (i.e., Caltech, DTD, Pets, and SUN) and achieve comparable performance on 2 tasks (i.e., CIFAR-10 and Flowers). It illustrates that collecting representations from multiple pre-training tasks can obtain complementary patterns that help mitigate the gap to different downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.612, + 0.787, + 0.763 + ], + "angle": 0, + "content": "Moreover, it can be observed that \"S+C\" performs better than \"S+M\" on DTD, while it is worse on Birds. The ensemble of all models, i.e., \"S+M+B+C\", shows the best performance on most tasks. Since the models in evaluation are pre-trained with different learning objectives, we find that the learned deep features complement each other. Therefore, more can be better. Note that ensemble can also be applied to fine-tuning. However, the computational cost will significantly increase depending on the number of models, which becomes unaffordable for real applications. In addition, its ensemble can still be worse than deep features. For example, an ensemble of fine-tuning on DTD achieves 78.9% that is about 1% worse than SeA." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Although the result of SeA is promising, fixed features perform noticeably worse than fine-tuning on two tasks, e.g., Aircraft and Cars. It is because that the gap between the source domain and the target domain is too large. For example, ImageNet contains limited generic images for cars, while the target task is more fine-grained and consists of cars with different make, models, and" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "Q. Qian et al." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.178 + ], + "angle": 0, + "content": "years. In this case, a larger model pre-trained on a larger data set covering more diverse domains (e.g., JFT-3B [38]) may help improve the performance of SeA." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.197, + 0.787, + 0.302 + ], + "angle": 0, + "content": "Comparison on Ensemble of Features Then, we illustrate the effectiveness of SeA by comparing it to the baseline without SeA and existing feature-space augmentation methods in Tab. 5, which includes the semantic augmentation method, i.e., ISDA [35] and a robust adversarial augmentation method, i.e., ME-ADA [40]. The experiment is conducted with the ensembled features from 4 models for all methods. The hyper-parameters in ISDA and ME-ADA are searched for the desired result." + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.325, + 0.787, + 0.352 + ], + "angle": 0, + "content": "Table 5: Comparison of SeA and benchmark methods on \\(\\mathrm{S + M + B + C}\\). \"Gain\" denotes the improvement over the Baseline." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.365, + 0.787, + 0.445 + ], + "angle": 0, + "content": "
MethodsAircraftBirdsCaltechCarsCIFAR10CIFAR100DTDFlowersFoodPetsSUNAvg
Baseline63.673.495.469.294.881.179.496.279.593.166.981.1
ISDA63.373.695.969.494.981.379.695.679.693.267.081.2
ME-ADA64.474.995.770.995.081.679.895.880.393.267.481.7
SeA67.476.896.375.495.282.979.897.280.893.868.283.1
Gain3.83.40.96.20.41.80.41.01.30.71.32.0
" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.474, + 0.787, + 0.686 + ], + "angle": 0, + "content": "First, compared with the baseline, SeA improves the performance of deep features over all data sets, and the average gain is up to \\(2\\%\\) as denoted by \"Gain\". The consistent improvement shows that augmentation is important for learning with fixed deep features and the proposed augmentation helps mitigate the overfitting problem effectively. Second, while both SeA and ISDA focus on semantic augmentation, ISDA relies on input space augmentations to obtain the semantic directions, while our method leverages features from other examples. Therefore, when input space augmentations are unavailable for deep features, ISDA cannot beat the baseline and it shows that our method is more appropriate for learning without input space augmentations. In addition, ME-ADA improves the vanilla adversarial augmentation by mitigating its sensitivity to the gradient direction. However, the optimization only considers the features from the target example. By incorporating the information from other examples in the same batch, the proposed method can outperform ME-ADA with a margin of \\(1.4\\%\\) on average." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.706, + 0.358, + 0.722 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.734, + 0.787, + 0.84 + ], + "angle": 0, + "content": "With the development of unsupervised representation learning, features with diverse information can be obtained by learning with different pretext tasks. Hence, we aim to investigate state-of-the-art performance of pre-trained deep features in this work. By introducing the novel semantic adversarial augmentation, we can show that learning with fixed features can achieve comparable performance to fine-tuning the whole network, but with way less cost. Investigating the efficacy of SeA in pre-training and fine-tuning can be our future work." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.33, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Semantic Adversarial Augmentation for Last Layer Features" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.173, + 0.786, + 0.228 + ], + "angle": 0, + "content": "1. Bossard, L., Guillaumeun, M., Gool, L.V.: Food-101 - mining discriminative components with random forests. In: Fleet, D.J., Pajdla, T., Schiele, B., Tuytelaars, T. (eds.) ECCV. Lecture Notes in Computer Science, vol. 8694, pp. 446-461. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.229, + 0.786, + 0.255 + ], + "angle": 0, + "content": "2. Boyd, S., Boyd, S.P., Vandenberghe, L.: Convex optimization. Cambridge university press (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.256, + 0.786, + 0.297 + ], + "angle": 0, + "content": "3. Caron, M., Misra, I., Mairal, J., Goyal, P., Bojanowski, P., Joulin, A.: Unsupervised learning of visual features by contrasting cluster assignments. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.297, + 0.786, + 0.337 + ], + "angle": 0, + "content": "4. Chen, T., Cheng, Y., Gan, Z., Wang, J., Wang, L., Liu, J., Wang, Z.: Adversarial feature augmentation and normalization for visual recognition. Trans. Mach. Learn. Res. 2022 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.337, + 0.786, + 0.378 + ], + "angle": 0, + "content": "5. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.E.: A simple framework for contrastive learning of visual representations. In: ICML. Proceedings of Machine Learning Research, vol. 119, pp. 1597-1607. PMLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.379, + 0.786, + 0.404 + ], + "angle": 0, + "content": "6. Chen, X., Xie, S., He, K.: An empirical study of training self-supervised vision transformers. In: ICCV. pp. 9620-9629. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.405, + 0.786, + 0.432 + ], + "angle": 0, + "content": "7. Cimpoi, M., Maji, S., Kokkinos, I., Mohamed, S., Vedaldi, A.: Describing textures in the wild. In: CVPR. pp. 3606-3613 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.433, + 0.786, + 0.459 + ], + "angle": 0, + "content": "8. Crammer, K., Singer, Y.: On the algorithmic implementation of multiclass kernel-based vector machines. J. Mach. Learn. Res. 2, 265-292 (2001)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.459, + 0.786, + 0.513 + ], + "angle": 0, + "content": "9. Donahue, J., Jia, Y., Vinyals, O., Hoffman, J., Zhang, N., Tzeng, E., Darrell, T.: Decaf: A deep convolutional activation feature for generic visual recognition. In: ICML. JMLR Workshop and Conference Proceedings, vol. 32, pp. 647-655. JMLR.org (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.513, + 0.786, + 0.568 + ], + "angle": 0, + "content": "0. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR. OpenReview.net (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.568, + 0.786, + 0.609 + ], + "angle": 0, + "content": "1. Fei-Fei, L., Fergus, R., Perona, P.: Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In: CVPR workshop. pp. 178-178. IEEE (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.609, + 0.786, + 0.637 + ], + "angle": 0, + "content": "2. Goodfellow, I.J., Shlens, J., Szegedy, C.: Explaining and harnessing adversarial examples. In: Bengio, Y., LeCun, Y. (eds.) ICLR (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.637, + 0.786, + 0.704 + ], + "angle": 0, + "content": "3. Grill, J., Strub, F., Altché, F., Tallec, C., Richemond, P.H., Buchatskaya, E., Doersch, C., Pires, B.A., Guo, Z., Azar, M.G., Piot, B., Kavukcuoglu, K., Munos, R., Valko, M.: Bootstrap your own latent - A new approach to self-supervised learning. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.705, + 0.786, + 0.732 + ], + "angle": 0, + "content": "4. He, K., Chen, X., Xie, S., Li, Y., Dollar, P., Girshick, R.B.: Masked autoencoders are scalable vision learners. In: CVPR. pp. 15979-15988. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.732, + 0.786, + 0.772 + ], + "angle": 0, + "content": "5. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.B.: Momentum contrast for unsupervised visual representation learning. In: CVPR. pp. 9726-9735. Computer Vision Foundation / IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.786, + 0.8 + ], + "angle": 0, + "content": "6. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778. IEEE Computer Society (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.786, + 0.84 + ], + "angle": 0, + "content": "7. Ioffe, S., Szegedy, C.: Batch normalization: Accelerating deep network training by reducing internal covariate shift. In: Bach, F.R., Blei, D.M. (eds.) ICML. JMLR Workshop and Conference Proceedings, vol. 37, pp. 448-456. JMLR.org (2015)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.173, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "Q. Qian et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "18. Kornblith, S., Shlens, J., Le, Q.V.: Do better imagenet models transfer better? In: CVPR. pp. 2661-2671. Computer Vision Foundation / IEEE (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.785, + 0.203 + ], + "angle": 0, + "content": "19. Krause, J., Stark, M., Deng, J., Fei-Fei, L.: 3d object representations for fine-grained categorization. In: ICCV workshop. pp. 554-561 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.785, + 0.231 + ], + "angle": 0, + "content": "20. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.272 + ], + "angle": 0, + "content": "21. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. In: Bartlett, P.L., Pereira, F.C.N., Burges, C.J.C., Bottou, L., Weinberger, K.Q. (eds.) NeurIPS. pp. 1106-1114 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.3 + ], + "angle": 0, + "content": "22. Liu, Z., Xu, Y., Xu, Y., Qian, Q., Li, H., Ji, X., Chan, A.B., Jin, R.: Improved fine-tuning by better leveraging pre-training data. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.301, + 0.785, + 0.327 + ], + "angle": 0, + "content": "23. Maji, S., Rahtu, E., Kannala, J., Blaschko, M., Vedaldi, A.: Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.355 + ], + "angle": 0, + "content": "24. Nilsback, M., Zisserman, A.: Automated flower classification over a large number of classes. In: ICVGIP. pp. 722-729. IEEE Computer Society (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.785, + 0.383 + ], + "angle": 0, + "content": "25. Parkhi, O.M., Vedaldi, A., Zisserman, A., Jawahar, C.: Cats and dogs. In: CVPR. pp. 3498-3505. IEEE (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.411 + ], + "angle": 0, + "content": "26. Qian, Q., Hu, J., Li, H.: Hierarchically robust representation learning. In: CVPR. pp. 7334-7342. Computer Vision Foundation / IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.438 + ], + "angle": 0, + "content": "27. Qian, Q., Jin, R., Zhu, S., Lin, Y.: Fine-grained visual categorization via multi-stage metric learning. In: CVPR. pp. 3716-3724. IEEE Computer Society (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.44, + 0.785, + 0.466 + ], + "angle": 0, + "content": "28. Qian, Q., Shang, L., Sun, B., Hu, J., Li, H., Jin, R.: Softtriple loss: Deep metric learning without triplet sampling. In: ICCV. pp. 6449-6457. IEEE (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.493 + ], + "angle": 0, + "content": "29. Qian, Q., Xu, Y., Hu, J., Li, H., Jin, R.: Unsupervised visual representation learning by online constrained k-means. In: CVPR. pp. 16619-16628. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.785, + 0.535 + ], + "angle": 0, + "content": "30. Ren, S., He, K., Girshick, R.B., Sun, J.: Faster R-CNN: towards real-time object detection with region proposal networks. IEEE Trans. Pattern Anal. Mach. Intell. 39(6), 1137-1149 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.537, + 0.785, + 0.577 + ], + "angle": 0, + "content": "31. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M.S., Berg, A.C., Fei-Fei, L.: Imagenet large scale visual recognition challenge. Int. J. Comput. Vis. 115(3), 211-252 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.579, + 0.785, + 0.632 + ], + "angle": 0, + "content": "32. Verma, V., Lamb, A., Beckham, C., Najafi, A., Mitliagkas, I., Lopez-Paz, D., Bengio, Y.: Manifold mixup: Better representations by interpolating hidden states. In: Chaudhuri, K., Salakhutdinov, R. (eds.) ICML. Proceedings of Machine Learning Research, vol. 97, pp. 6438-6447. PMLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.634, + 0.785, + 0.673 + ], + "angle": 0, + "content": "33. Volpi, R., Morerio, P., Savarese, S., Murino, V.: Adversarial feature augmentation for unsupervised domain adaptation. In: CVPR. pp. 5495-5504. Computer Vision Foundation / IEEE Computer Society (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.785, + 0.701 + ], + "angle": 0, + "content": "34. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The caltech-ucsd birds-200-2011 dataset (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.785, + 0.742 + ], + "angle": 0, + "content": "35. Wang, Y., Huang, G., Song, S., Pan, X., Xia, Y., Wu, C.: Regularizing deep networks with semantic data augmentation. IEEE Trans. Pattern Anal. Mach. Intell. 44(7), 3733-3748 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.785, + 0.785 + ], + "angle": 0, + "content": "36. Xiao, J., Hays, J., Ehinger, K.A., Oliva, A., Torralba, A.: SUN database: Large-scale scene recognition from abbey to zoo. In: CVPR. pp. 3485-3492. IEEE Computer Society (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.785, + 0.812 + ], + "angle": 0, + "content": "37. Xu, Y., Qian, Q., Li, H., Jin, R., Hu, J.: Weakly supervised representation learning with coarse labels. In: ICCV. pp. 10573-10581. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "38. Zhai, X., Kolesnikov, A., Houlsby, N., Beyer, L.: Scaling vision transformers. In: CVPR. pp. 1204-1213. IEEE (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.33, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Semantic Adversarial Augmentation for Last Layer Features" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "39. Zhang, H., Cissé, M., Dauphin, Y.N., Lopez-Paz, D.: mixup: Beyond empirical risk minimization. In: ICLR. OpenReview.net (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.785, + 0.218 + ], + "angle": 0, + "content": "40. Zhao, L., Liu, T., Peng, X., Metaxas, D.N.: Maximum-entropy adversarial data augmentation for improved generalization and robustness. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) NeurIPS (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.218 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/SeA_ Semantic Adversarial Augmentation for Last Layer Features from Unsupervised Representation Learning/ca46bbef-cd36-48f1-b01d-aebb511e39a1_origin.pdf b/2024/SeA_ Semantic Adversarial Augmentation for Last Layer Features from Unsupervised Representation Learning/ca46bbef-cd36-48f1-b01d-aebb511e39a1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..835ea63576650a98cce5fd19605f97b013ac10e1 --- /dev/null +++ b/2024/SeA_ Semantic Adversarial Augmentation for Last Layer Features from Unsupervised Representation Learning/ca46bbef-cd36-48f1-b01d-aebb511e39a1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dddf77e733c4d4be8666b3c15097e4fe2498dc7bc12023f22c6f380abf127b4e +size 959688 diff --git a/2024/SeA_ Semantic Adversarial Augmentation for Last Layer Features from Unsupervised Representation Learning/full.md b/2024/SeA_ Semantic Adversarial Augmentation for Last Layer Features from Unsupervised Representation Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..28be782764541070d38e3118d9495ceb9d24405b --- /dev/null +++ b/2024/SeA_ Semantic Adversarial Augmentation for Last Layer Features from Unsupervised Representation Learning/full.md @@ -0,0 +1,345 @@ +# SeA: Semantic Adversarial Augmentation for Last Layer Features from Unsupervised Representation Learning + +Qi Qian $^{1}$ , Yuanhong $\mathrm{Xu}^{2}$ , and Juhua $\mathrm{Hu}^{3}$ + +1 Alibaba Group, Bellevue, WA 98004, USA + +2 Alibaba Group, Hangzhou, China + +3 School of Engineering and Technology, + +University of Washington, Tacoma, WA 98402, USA + +{qi.qian, yuanhong.xuyh}@alibaba-inc.com, juhuah@uw.edu + +Abstract. Deep features extracted from certain layers of a pre-trained deep model show superior performance over the conventional hand-crafted features. Compared with fine-tuning or linear probing that can explore diverse augmentations, e.g., random crop/flipping, in the original input space, the appropriate augmentations for learning with fixed deep features are more challenging and have been less investigated, which degenerates the performance. To unleash the potential of fixed deep features, we propose a novel semantic adversarial augmentation (SeA) in the feature space for optimization. Concretely, the adversarial direction implied by the gradient will be projected to a subspace spanned by other examples to preserve the semantic information. Then, deep features will be perturbed with the semantic direction, and augmented features will be applied to learn the classifier. Experiments are conducted on 11 benchmark downstream classification tasks with 4 popular pre-trained models. Our method is $2\%$ better than the deep features without SeA on average. Moreover, compared to the expensive fine-tuning that is expected to give good performance, SeA shows a comparable performance on 6 out of 11 tasks, demonstrating the effectiveness of our proposal in addition to its efficiency. + +Keywords: Semantic augmentation $\cdot$ Deep features $\cdot$ Unsupervised representation learning $\cdot$ Self-supervised learning + +# 1 Introduction + +Deep learning can be partially considered as a representation learning method that aims to extract features from raw data directly. By obtaining appropriate representations, a simple linear model can achieve state-of-the-art performance on challenging tasks, e.g., classification [16], object detection [30], etc. After the success of deep learning [21], researchers investigate the representations learned + +![](images/6fe0bae98de123087cdf64e4d44152274d04a431f794049f1b69f26c35b1985c.jpg) +Fig. 1: Illustration of semantic adversarial augmentation (SeA). The red solid and empty circles denote the original data and its augmentation, respectively. Left: Conventional adversarial augmentation can perturb with arbitrary direction (e.g., augmentation may appear the same as the original); Right: SeA augments examples with semantic directions spanned by features from real data (e.g., different sectors show different subspaces for augmentation, where we can get more semantic meaningful augmentation). + +from a large-scale data set, i.e., ImageNet [31]. Surprisingly, the deep features extracted from a certain layer of a pre-trained model can outperform handcrafted features on various downstream tasks [9], demonstrating the efficacy of the data-dependent representation learning mechanism implied by deep learning. + +Unlike hand-crafted features, representations obtained by deep learning are highly data/task-dependent as illustrated in [37]. When optimizing the task from the source domain, the neural network focuses on exploring the patterns related to the specific training task, while ignoring diverse information that has potential for different tasks. Hence, different representations can be learned from different training tasks even with the same training data from the source domain. Conventional methods learn representations with the labels of examples, which capture the knowledge only for the given labels and limit the information in deep features of each example. + +To mitigate the problem, fine-tuning parameters of the whole network becomes prevalent for various downstream tasks [18]. On one hand, fine-tuning can benefit from the prior knowledge in the pre-trained representations. On the other hand, it can further optimize representations with diverse augmentations on the target task. Consequently, fine-tuning works better than learning with fixed representations on downstream tasks. + +An important advantage of fine-tuning over learning with fixed deep features is the additional information from semantic augmentations. Obtaining effective augmentations in image space is convenient with semantic operators such as random crop, flipping, etc. In contrast, obtaining appropriate augmentations in the feature space for deep features becomes challenging, due to the lack of semantic operations. To leverage input space augmentations, a linear classifier can be learned with a frozen backbone by linear probing [6] that generates augmented images for optimization at each iteration. + +While fine-tuning and linear probing show promising performance, learning with fixed features is still attractive due to its good properties for real-world applications. First, it only applies deep models to extract features for each original example once, which is applicable for limited computational resources. On the contrary, fine-tuning has the full forward and backward pass and linear probing has the forward pass for each augmented example, which is more expensive for optimization. Second, the same models can be reused for different downstream tasks to extract deep features, while fine-tuning will adjust all parameters in pre-trained models. Thereafter, it has to keep a specific deep model for each task, which becomes intractable for handling hundreds of downstream tasks simultaneously. Finally, given fixed features, learning a linear classifier can be formulated as a convex problem that has the global optimum with a theoretical guarantee [2], while preserving the knowledge from the pre-trained model in features. It requires much less tuning efforts than fine-tuning that has to be tuned carefully to avoid the collapse of pre-trained parameters and catastrophic forgetting [22]. Therefore, we focus on fixed deep features extracted after the last pooling layer (i.e., inputs for the last fully-connected layer) in this work. + +Although some existing works [32,35] consider augmentation in intermediate layers to help optimize the whole network, little efforts were devoted to the deep features after the last pooling layer. Moreover, the existing feature space augmentation method [35] is proposed to complement the input space augmentation. It relies on other augmentation techniques and cannot work well solely. + +To tackle the problem, in this work, we propose a novel semantic adversarial augmentation strategy in the feature space of fixed deep features. Concretely, the gradient of each example is computed at first, and then a semantic direction can be observed by projecting the gradient to the subspace spanned by real data. Finally, the features of examples are augmented according to the obtained semantic direction for learning the linear classifier. Fig. 1 illustrates the proposed augmentation method that obtains effective augmentations with only fixed features. Moreover, to further mitigate the information loss in deep features from supervised pre-trained models, those from pre-trained models with different unsupervised pretext tasks are also investigated with an ensemble strategy. Our main contributions can be summarized as follows. + +- We empirically demonstrate the current performance gap between fine-tuning and learning with fixed deep features. Those deep features extracted from 4 representative pre-trained models are evaluated on 11 downstream tasks. +- To improve the performance of deep features, we propose a semantic adversarial augmentation method to obtain appropriate augmentations tailored for fixed features. In addition, a smoothed hinge loss is investigated to demonstrate the augmentation direction explicitly. +- Our proposed SeA gains $2\%$ accuracy on average over 11 downstream tasks compared to the baseline using deep features without augmentations. Moreover, compared with fine-tuning that is expected to give good performance, our recipe for learning with deep features can achieve comparable performance on 6 out of 11 tasks with way less computational cost, which shows + +the potential of self-supervised pre-trained deep features and demonstrates both the efficiency and effectiveness of our proposal. + +# 2 Related Work + +# 2.1 Deep Features + +Modern deep neural networks consist of multiple layers to exact representations from raw materials. After that, a simple linear model encoded by a fully-connected layer can be attached on top of the representations for classification. After pre-training on a large-scale data set, the obtained neural network can be considered as a feature extractor and a new fully-connected layer can be learned for target tasks. + +When the full label information of data is available, the learning objective is explicit and a supervised representation learning can be conducted by optimizing a conventional classification task, which is equivalent to optimizing the triplet loss as in distance metric learning [28]. The obtained deep features can outperform hand-crafted features on different applications, e.g., classification [9], distance metric learning [27], etc. Besides, some work aimed to obtain robust representations for different downstream tasks within the framework of classification [26]. + +Without any label information, the learning objective varies in unsupervised learning. First, each instance can be considered as an individual class and the representations can be learned by instance discrimination [5, 15]. In addition, clustering can be applied to capture the relationship between different instances. A coarse-grained classification task defined on clusters can be leveraged to optimize representations [3, 29]. Finally, other pretext tasks beyond classification also demonstrate the effective representation after pre-training [13]. Compared with supervised representation learning, the different objectives in unsupervised learning can learn various semantic information even from the same data set. In this work, we will systematically study representations from different pre-trained models and illustrate the performance gap compared to fine-tuning the whole neural network. + +# 2.2 Augmentation in Feature Space + +Due to the over-parameterization property of deep neural networks, augmentation is essential for training deep models to avoid over-fitting [21]. Given an image, a perturbed copy can be observed by standard image operations such as random crop, flipping, color jitter, etc. [16]. The augmented images have a large variance while preserving the semantic information of the original image, i.e., an image of a cat is still a cat after augmentation, which helps train effective models with the additional information. + +However, augmentation in feature space becomes more challenging due to the lack of semantic preserving operators. Some methods consider the augmentation of features from intermediate layers to help train the whole deep neural + +networks in different applications [4, 32, 33, 35]. For example, [33] aims to learn feature augmentation with a feature generator for unsupervised domain adaptation. [4] tries to augment intermediate feature maps with adversarial feature moments in batch normalization [17] for efficient training. Manifold mixup [32] proposes to apply the original mixup [39] to features from multiple layers for augmentation, which also includes the layer investigated in this work. However, their study shows that the gain of mixup is mainly from the original image space and feature space of early stages, while that from the feature space of the last layer is negligible, which is consistent with our observation in the ablation study. ISDA [35] considers the semantic augmentation in feature space but it has to obtain the semantic direction with input space augmentations, which is complementary to input space augmentations but cannot work as the sole augmentation for deep features well. On the contrary, we propose SeA to project the adversarial direction with features of real data points. Moreover, our proposal is tailored for fixed features and is different from existing works that are for optimizing the whole network. + +# 3 Semantic Adversarial Augmentation + +We start the analysis with a standard classification framework. Let $\{x_{i},y_{i}\}_{i = 1}^{n}$ denote the training data set and $\{f_k\}_{k = 1}^K$ contains a set of $K$ pre-trained models, which can be pre-trained with different learning objectives on different data sets. In this work, we directly use the representations extracted from pre-trained deep models and learn a simple linear model on top of it. Representations from different deep models are concatenated as the final representation for each example. Given $x_{i}$ , the representation will be extracted as + +$$ +\mathbf {x} _ {i} = \left[ f _ {1} \left(x _ {i}\right), \dots , f _ {K} \left(x _ {i}\right) \right] \in \mathbb {R} ^ {d} \tag {1} +$$ + +where $d = \sum_{k}d_{k}$ and $d_{k}$ is the dimension of the representation from the $k$ -th pre-trained deep model. + +With the above fixed deep features, a classification model can be learned by minimizing the empirical risk with the appropriate regularization as + +$$ +\min _ {\mathbf {w}} \sum_ {i} \ell (\mathbf {x} _ {i}, y _ {i}; \mathbf {w}) + \frac {\tau}{2} \| \mathbf {w} \| _ {F} ^ {2} \tag {2} +$$ + +where $\tau$ is the weight for $L_{2}$ regularization and $\mathbf{w}$ denotes parameters of the classification model, which is a linear classifier with the prediction probability of the $i$ -th example on the $j$ -th class as + +$$ +p _ {i, j} = \frac {\exp \left(\mathbf {x} _ {i} ^ {\top} \mathbf {w} _ {j}\right)}{\sum_ {c = 1} ^ {C} \exp \left(\mathbf {x} _ {i} ^ {\top} \mathbf {w} _ {c}\right)} \tag {3} +$$ + +$C$ is the number of classes and $\ell$ is a loss function for learning that will be discussed later. + +# 3.1 Semantic Direction + +In a standard fine-tuning pipeline, an image can be augmented by multiple random perturbations. When only deep features are available, it is hard to adopt existing techniques to generate appropriate augmented examples, since the semantic direction in the feature space is hard to capture. Manifold mixup [32] shows that randomly augmenting features from the last output layers has little improvement compared to the augmentation in input image space. To facilitate the performance of deep features, we investigate the semantic adversarial direction for augmentation. + +First, for the $i$ -th example and its representation $\mathbf{x}_i$ , we consider its adversarial direction that can be obtained by maximizing the loss function [12] + +$$ +\max _ {\mathbf {x}: \left\| \mathbf {x} - \mathbf {x} _ {i} \right\| _ {2} \leq \gamma} \ell \left(\mathbf {x} _ {i}, y _ {i}; \mathbf {w}\right) \tag {4} +$$ + +Note that the gradient indicates the ideal direction for the adversarial perturbation, and a standard adversarial example can be obtained by gradient ascent as + +$$ +\hat {\mathbf {x}} _ {i} = \mathbf {x} _ {i} + \eta \nabla_ {\mathbf {x} _ {i}} \ell \tag {5} +$$ + +However, the gradient direction may not be semantically informative. [12] shows that the loss can be substantially increased with the adversarial example generated from the gradient, while the appearance is almost the same as the original image as illustrated in Fig. 1. Therefore, the conventional adversarial learning method helps improve the robustness to the adversarial attack but may be inappropriate for regularizing the generic learning task that requires additional information from diverse images. + +To obtain the semantic direction, we propose to project the gradient direction to the subspace spanned by real data points. Concretely, let $\{\mathbf{x}_j\}_{j=1}^b$ denote a mini-batch of data and $\mathbf{g}_i = \nabla_{\mathbf{x}_i}\ell$ indicate the gradient of $\mathbf{x}_i$ . For the $i$ -th image, a semantic adversarial direction $\hat{\mathbf{g}}_i$ consists of representations from data as + +$$ +\hat {\mathbf {g}} _ {i} = \sum_ {j: j \neq i} q _ {j} \mathbf {x} _ {j}; \quad \mathbf {q} ^ {*} = \arg \min _ {\mathbf {q} \in \Delta} \mathcal {D} (\hat {\mathbf {g}} _ {i}, \mathbf {g} _ {i}) \tag {6} +$$ + +where $\mathcal{D}(\cdot, \cdot)$ is a distance function and $\varDelta$ is a simplex as $\varDelta=\{\mathbf{q}\in\mathbb{R}^{b-1}|\sum_{j}q_{j}=1;\forall j,q_{j}\geq 0\}$ . With the projection, we aim to find an adversarial direction in a subspace consisting of original data points as illustrated in Fig. 1. + +The last challenge is to obtain $\mathbf{q}$ efficiently with a distance function. By adopting the squared Euclidean distance, which is a standard distance measurement, the optimization problem for $\mathbf{q}$ can be written as + +$$ +\min _ {\mathbf {q} \in \Delta} \| \sum_ {j: j \neq i} q _ {j} \mathbf {x} _ {j} - \mathbf {g} _ {i} \| _ {2} ^ {2} - \alpha H (\mathbf {q}) \tag {7} +$$ + +where $H(\mathbf{q})$ is the entropy of $\mathbf{q}$ that helps improve the robustness to different batches. When normalizing features such that $\| \mathbf{x}_j\| _2 = \| \mathbf{g}_i\| _2 = 1$ to obtain + +Algorithm 1 Semantic Adversarial Augmentation (SeA) for Given Features +1: Input: Dataset $\{x_i,y_i\}$ , pre-trained models $\{f_k\}$ , iterations $T$ , linear model $\mathbf{w}$ , $\alpha$ , $\tau$ , $\eta$ +2: Extract deep features by $\{f_k\}$ for all examples as $\{\mathbf{x}_i\}$ +3: for $t = 1,\dots,T$ do +4: Receive a mini-batch of examples $\{\mathbf{x}_i,y_i\}_{i=1}^b$ +5: Obtain augmented examples $\{\tilde{\mathbf{x}}_i,y_i\}$ as in Eqn. 11 +6: Optimize $\mathbf{w}$ by SGD: $\mathbf{w} = \mathbf{w} - \eta_w\left(\frac{1}{b}\sum_i^b\nabla_{\mathbf{w}}\ell(\tilde{\mathbf{x}}_i,y_i;\mathbf{w}) + \tau \mathbf{w}\right)$ +7: end for +8: return $\mathbf{w}$ + +the direction without the influence from the magnitude, the problem can be upper-bounded as + +Proposition 1 With unit length variables $\{\mathbf{x}_j\}$ and $\mathbf{g}_i$ , we have + +$$ +\left\| \sum_ {j: j \neq i} q _ {j} \mathbf {x} _ {j} - \mathbf {g} _ {i} \right\| _ {2} ^ {2} \leq 2 - 2 \sum_ {j} q _ {j} \mathbf {x} _ {j} ^ {\top} \mathbf {g} _ {i} \tag {8} +$$ + +The detailed proof can be found in the supplementary. + +By rearranging the terms, we can maximize the lower-bound of the original problem as + +$$ +\max _ {\mathbf {q} \in \Delta} \sum_ {j: j \neq i} q _ {j} \mathbf {x} _ {j} ^ {\top} \mathbf {g} _ {i} + \alpha H (\mathbf {q}) \tag {9} +$$ + +According to the K.K.T. condition [2], $\mathbf{q}$ has a closed-form solution. + +Proposition 2 The problem in Eqn. 9 has the optimal solution as + +$$ +q _ {j} = \frac {\exp \left(\mathbf {x} _ {j} ^ {\top} \mathbf {g} _ {i} / \alpha\right)}{Z}; \quad Z = \sum_ {k: k \neq i} \exp \left(\mathbf {x} _ {k} ^ {\top} \mathbf {g} _ {i} / \alpha\right) \tag {10} +$$ + +Given the semantic adversarial direction, the target example with semantic perturbation can be obtained as + +$$ +\tilde {\mathbf {x}} _ {i} = \Pi \left(\mathbf {x} _ {i} + \eta \Pi \left(\hat {\mathbf {g}} _ {i}\right)\right) \tag {11} +$$ + +where $\Pi(\cdot)$ normalizes the vector to the unit length if required and $\eta$ denotes the step size for augmentation. Compared with the adversarial perturbation in Eqn. 5, the augmentation in Eqn. 11 projects the gradient to the direction consisting of real data points, which can capture the semantic information in the feature space effectively. Alg. 1 summarizes the proposed method. + +# 3.2 Illustration of Adversarial Direction + +In this subsection, we will generalize the cross entropy loss to help illustrate the proposed augmentation strategy. + +The standard multi-class hinge loss [8] for convex optimization can be written as + +$$ +\ell (\mathbf {x} _ {i}, y _ {i}; \mathbf {w}) = \max \left\{0, \delta + \max _ {c \neq y _ {i}} \mathbf {x} _ {i} ^ {\top} \mathbf {w} _ {c} - \mathbf {x} _ {i} ^ {\top} \mathbf {w} _ {y _ {i}} \right\} \tag {12} +$$ + +where $\delta$ is a pre-defined margin. + +Unlike convex optimization, cross-entropy loss is prevalent in deep learning, which is a smooth function and can help accelerate the convergence [2]. Therefore, we propose to obtain a smoothed hinge loss by introducing a distribution over logits from different classes $\mathbf{p} \in \Delta$ where $\Delta$ is the simplex. + +First, the original hinge loss is equivalent to + +$$ +\ell \left(\mathbf {x} _ {i}, y _ {i}; \mathbf {w}\right) = \max _ {\mathbf {p} \in \Delta} p _ {y _ {i}} \mathbf {x} _ {i} ^ {\top} \mathbf {w} _ {y _ {i}} + \sum_ {c \neq y _ {i}} p _ {c} \left(\delta + \mathbf {x} _ {i} ^ {\top} \mathbf {w} _ {c}\right) - \mathbf {x} _ {i} ^ {\top} \mathbf {w} _ {y _ {i}} \tag {13} +$$ + +According to the analysis for cross entropy loss [28], the loss can be smoothed by adding an entropy regularization for $\mathbf{p}$ as + +$$ +\ell \left(\mathbf {x} _ {i}, y _ {i}; \mathbf {w}\right) = \max _ {\mathbf {p} \in \Delta} p _ {y _ {i}} \mathbf {x} _ {i} ^ {\top} \mathbf {w} _ {y _ {i}} + \sum_ {c \neq y _ {i}} p _ {c} (\delta + \mathbf {x} _ {i} ^ {\top} \mathbf {w} _ {c}) + \lambda H (\mathbf {p}) - \mathbf {x} _ {i} ^ {\top} \mathbf {w} _ {y _ {i}} \tag {14} +$$ + +where $H(\mathbf{p})$ denotes the entropy of $\mathbf{p}$ and $\lambda$ is the coefficient. + +Proposition 3 The loss function in Eqn. 14 is equivalent to + +$$ +\ell (\mathbf {x} _ {i}, y _ {i}; \mathbf {w}) = - \lambda \log \frac {\exp \left(\mathbf {x} _ {i} ^ {\top} \mathbf {w} _ {y _ {i}} / \lambda\right)}{\exp \left(\mathbf {x} _ {i} ^ {\top} \mathbf {w} _ {y _ {i}} / \lambda\right) + \sum_ {c \neq y _ {i}} \exp \left(\left(\mathbf {x} _ {i} ^ {\top} \mathbf {w} _ {c} + \delta\right) / \lambda\right)} \tag {15} +$$ + +Remark It is obvious that the popular cross entropy loss is a special case of Eqn. 15 by letting $\delta = 0$ and $\lambda = 1$ . Our analysis connects the hinge loss in conventional methods to the popular loss function in deep learning. + +Finally, since deep features are fixed, we can illustrate the gradient direction with the proposed smoothed loss function explicitly. + +Proposition 4 Given the loss function in Eqn. 15, the gradient of $\mathbf{x}_i$ is + +$$ +\nabla_ {\mathbf {x} _ {i}} \ell (\mathbf {x} _ {i}, y _ {i}; \mathbf {w}) = \sum_ {c = 1} ^ {C} p _ {c} \mathbf {w} _ {c} - \mathbf {w} _ {y _ {i}} \tag {16} +$$ + +$$ +\begin{array}{r} w h e r e p _ {c} = \left\{ \begin{array}{l l} \frac {\exp ((\mathbf {x} _ {i} ^ {\top} \mathbf {w} _ {c} + \delta) / \lambda)}{Z} c \neq y _ {i} \\ \frac {\exp (\mathbf {x} _ {i} ^ {\top} \mathbf {w} _ {c} / \lambda)}{Z} c = y _ {i} \end{array} a n d Z = \exp (\mathbf {x} _ {i} ^ {\top} \mathbf {w} _ {y _ {i}} / \lambda) + \sum_ {c \neq y _ {i}} \exp ((\mathbf {x} _ {i} ^ {\top} \mathbf {w} _ {c} + \delta) / \lambda) \leqslant 1, \right. \\ \delta) / \lambda). \end{array} +$$ + +Remark The gradient direction in the feature space indicates that an adversarial perturbation should be close to the primary directions of other classes, while far away from the direction of the corresponding class. + +# 4 Experiments + +To demonstrate our proposed method, we adopt state-of-the-art and widely used pre-trained models to extract deep features for evaluation. It should be noted that only two main architectures, including ResNet-50 [16] and vision transformer (ViT) [10], are prevalently applied in self-supervised learning [6,13,14,29]. Since ViT shows a worse classification performance than ResNet-50 with the frozen backbone as demonstrated in [14], we will focus on ResNet-50 with different public pre-trained parameters in the experiment. Specifically, one supervised pre-trained ResNet-50 and three self-supervised pre-trained ResNet-50 are applied for feature extraction. All of these models are pre-trained on ImageNet-1K [31] with different learning objectives. The details of different models can be found in the supplementary. The accuracy of the supervised model and that of linear probing for unsupervised models are summarized in Tab. 1. + +Table 1: Performance of 4 pre-trained models on ImageNet. + +
Model#Pre-training EpochsObjectiveAcc%
Supervised [16] (S)901,000-class classification76.6
MoCo-v3 [6] (M)1,000instance discrimination74.6
BYOL [13] (B)1,000regression74.3
CoKe [29] (C)1,000cluster discrimination74.9
+ +11 diverse downstream data sets are applied for evaluation, including Aircraft [23], Caltech101 [11], Stanford Cars [19], CIFAR-10 [20], CIFAR-100 [20], CUB200-2011 (Birds) [34], Describable Textures Dataset (DTD) [7], Flowers [24], Food101 [1], Oxford-IIIT Pet (Pets) [25], and Sun397 [36]. We follow the evaluation protocol in [13]. Concretely, all models search hyper-parameters on the provided/generated validation set in each downstream task, and the standard metric on the provided test set is reported. Besides, mean per-class accuracy is reported on Aircraft, Caltech101, Flowers, and Pets, while Top-1 accuracy is utilized for other data sets. More details can be found in [13]. + +Each model is fine-tuned with SGD by 100 epochs for sufficient training. The batch size is 256 and the momentum is 0.9. The learning rate is searched in a range of 7 logarithmically-spaced values between $10^{-4}$ and $10^{-1}$ . Weight decay is optional, for which if it is applied, the value will be searched with the same setting between $10^{-6}$ and $10^{-3}$ . The standard augmentation, i.e., random crop, random horizontal flipping, is applied as in most existing fine-tuning pipelines. + +For a fair comparison, we also apply SGD to learn the linear classifier with fixed deep features using the same batch size and momentum. Unlike fine-tuning, we have the constant learning rate for our method, which is searched in $\{2^i\}_{i = -2}^{i = 3}$ while weight decay is searched in $\{0,10^{-6},10^{-5},10^{-4}\}$ . For other parameters, we have $\alpha$ , $\lambda$ , $\delta$ and $\eta$ searched in $\{0.01,0.02,0.05\}$ , $\{0.05,0.1,1\}$ , $\{0,1\}$ , and $\{0,0.2,0.4,0.8\}$ , respectively. + +For each image, we extract the output of the final pooling layer as deep features. Features from an individual model are normalized to the unit length. If multiple models are exploited, features from different models are concatenated as in Eqn. 1 and the combined feature vector is further normalized to the unit length. The only augmentation for training with fixed features is the proposed semantic adversarial augmentation. To approximate the direction of the gradient for each example, a mini-batch of data is leveraged to optimize the problem in Eqn. 9. The projection for normalization is also applied to the augmented examples. + +# 4.1 Ablation Study + +Before the experiments on downstream tasks, we first investigate the effect of parameters in the proposed augmentation. The ablation study is conducted on CIFAR-100 with features concatenated from 4 pre-trained models and the accuracy is reported for comparison. + +Effect of Step Size for Augmentation First, we evaluate the step size $\eta$ in the semantic adversarial augmentation. For SeA, we fix $\lambda = 0.1$ , $\delta = 0$ , $\alpha = 0.01$ , and the learning rate as 1, according to the performance on the validation set. The weight of the augmentation $\eta$ varies in $\{0, 0.2, 0.4, 0.6, 0.8, 1.0\}$ and the performance is summarized in Fig. 2-Fig. 3. + +![](images/8af8176113b60714f697ac9f6830a26ba36edbd654a5dfd3770e0015a38a3082.jpg) +Fig. 2: Training accuracy Fig. 3: Test accuracy of Fig. 4: Test accuracy of various different $\eta$ . different $\eta$ . + +![](images/507bbac29dc0984305db8db92063b388e0db13e17504b0689ef7b587ab9354be.jpg) + +![](images/657555cd25279a74e4a27c3fa7884071d1d0a047c676f86fc08d378ca235c4b0.jpg) + +When $\eta = 0$ , there is no augmentation applied for training. Hence, we can observe that the training accuracy exceeds $92\%$ after only 20 epochs. Due to the over-fitting on the training set, the test accuracy of this baseline is only $81.1\%$ on CIFAR-100. By gradually increasing the step size for the augmentation, the training accuracy is decreased as expected, while the test accuracy increases. It demonstrates that the proposed semantic adversarial augmentation strategy can effectively mitigate the over-fitting problem for learning with fixed deep features. By setting an appropriate step size, the test accuracy can be improved to $82.9\%$ , which helps shrink the gap to fine-tuning. + +Effect of Directions for Augmentation Different directions can be adopted for generating augmentation. Besides the proposed semantic adversarial aug-. mentations (SeA), four variants are included in the comparison. + +- Adv: the original gradient direction without projection as in Eqn. 5. +- SeA_Neg: similar to the gradient direction in Proposition 4 but getting rid of the direction from the target class and keeping the direction of $\sum_{c}^{C}p_{c}\mathbf{w}_{c}$ . +- Rand: a uniformly random direction within subspace spanned by a minibatch of data, i.e., generating $\mathbf{q}$ in Eqn. 6 as a random vector. +- Manifold mixup: a random semantic direction indicated by an example [32]. + +The step size for Adv is very sensitive and is searched in $\{0.005, 0.01, 0.02, 0.05\}$ , while the step size for others is searched in $\{0.2, 0.4, 0.8, 1.0\}$ . + +Fig. 4 shows the comparison on CIFAR-100. First, we can observe that most augmentation directions except the random one can improve the performance over the baseline without augmentation. It confirms that augmentation is important for better generalization even for learning from fixed representations. Then, the method with the random augmentation direction performs similarly to the baseline, which shows that it is challenging to find an effective direction for augmentation in feature space. However, with the original gradient direction directly as in Adv, the training loss will be significantly increased even with a small step size and only a step size that is much smaller than SeA can obtain an applicable model. This is because that the gradient direction only aims to maximize the loss while ignoring the data distribution in the feature space. By projecting the gradient to a data-dependent subspace, SeA and SeA_Neg can achieve a better performance than the baseline without a sophisticated setup for the step size. Moreover, by removing the direction from the target class, the accuracy is degenerated by about $1\%$ . The comparison demonstrates that keeping the completed gradient direction is essential for obtaining appropriate augmentation. Finally, a random semantic direction indicated by examples as in mixup is worse than SeA with a clear margin. + +Comparison with Linear Probing Besides learning with fixed features, linear probing is another way to obtain the linear classifier with a frozen backbone. Compared with our method, it can obtain appropriate augmentations in image space. However, it has to forward each augmentation for learning and the cost is much higher than us. + +To illustrate the effectiveness of the proposed augmentation method, we compare the performance between linear probing and our proposal in Tab. 2. For a fair comparison, all methods optimize features from a single model of MoCo. It shows that the proposed augmentation can outperform the augmentation in the input image space for learning the target classifier. + +Running Time Finally, we compare the running time of fine-tuning, linear probing, and SeA using fixed deep features in Tab. 3. Note that with features + +Table 2: Comparison of accuracy (%) Table 3: Comparison of running time. SeA on CIFAR-100. All methods utilize a utilize deep features from 4 models while single model from MoCo. other methods optimize only a single model. + +
baselinelinear probingSeA
77.677.879.0
Methodsfine-tuninglinear probing
Time (s)14,3834,437
+ +from 4 models, the dimension of the feature vector is 8, 192. However, optimizing the combined features is still much more efficient than fine-tuning a single model or linear probing with 2,048 final features. All methods train the corresponding model by 100 epochs and the running time is measured on a single V100 GPU. Fine-tuning costs 14,383 (sec) for the entire training. In contrast, SeA only costs 40 (sec) to obtain the optimal linear model. The efficiency of SeA using deep features implies its applicability for limited-resource scenarios. With efficient optimization, we can repeat experiments of SeA by multiple times. The standard deviation is only $0.03\%$ on CIFAR-100 over 5 runs, which shows the stability of convex optimization. Although linear probing eliminates the cost of the backward pass, that of the forward pass at each iteration is still expensive. Considering that linear probing has a similar running cost as fine-tuning but with worse performance, it will not be included in the comparison over other downstream tasks. More ablation experiments can be found in the supplementary. + +# 4.2 Comparison on Downstream Tasks + +Now we evaluate features from different pre-trained models and the ensemble to demonstrate the effectiveness of SeA on downstream tasks. The comparison is summarized in Tab. 4. The linear model with fixed features from different models is denoted by the initial of the corresponding pre-trained model, i.e., S, M, B, and C stand for Supervised, MoCo, BYOL, and CoKe, respectively. + +First, the performance of fine-tuning supervised and unsupervised models varies in different domains. For domains closely related to ImageNet (e.g., Birds, Flowers, Pets), the supervised pre-trained model shows better performance than the unsupervised ones, while unsupervised models surpass the supervised counterpart when the domain gap between pre-training and fine-tuning is large. The phenomenon indicates that representations from supervised/unsupervised models capture different patterns and suggests that combining the information from multiple models may handle downstream tasks better. + +Second, when having the fixed representation from only a single model with SeA, the performance is already close to fine-tuning on certain data sets, e.g., Caltech, DTD, and SUN. However, the information learned from a single model is biased toward the specific pre-training task. Without fine-tuning, the fixed representation only preserves limited patterns from the source domain, which is insufficient for different target domains. + +To mitigate the limitation of individual pre-training tasks, we consider combining the representations from multiple models, which can be done by a simple + +Table 4: Comparison of accuracy (%) on downstream tasks. The best performance within each group is underlined while the global one is in bold. + +
MethodsAircraftBirdsCaltechCarsCIFAR10CIFAR100DTDFlowersFoodPetsSUN
Fine-tuning:
Supervised87.382.694.691.997.887.474.797.788.493.663.6
MoCo-v387.581.594.891.798.387.476.396.488.292.265.4
BYOL88.282.094.091.498.087.576.896.888.591.763.4
CoKe89.080.694.492.198.388.378.696.788.591.964.7
Fixed deep features with SeA:
S37.466.592.645.889.871.372.688.167.892.359.2
M61.571.495.469.793.479.076.494.674.990.764.2
B60.066.494.264.291.976.575.895.374.089.761.9
C57.867.094.363.493.078.577.895.676.191.065.3
S+M62.076.396.170.494.480.477.196.377.893.166.3
S+B60.773.795.767.593.479.177.996.377.493.465.5
S+C58.673.795.666.093.580.379.196.278.693.666.9
M+B65.772.595.772.794.680.878.296.277.691.965.5
M+C65.272.295.773.194.681.477.996.578.392.266.6
B+C64.070.195.270.193.980.778.896.478.391.166.4
S+M+B66.276.596.474.294.981.878.696.979.593.367.0
S+M+C65.876.796.373.195.181.978.896.979.993.667.7
S+B+C63.775.496.271.494.482.179.297.079.893.867.7
M+B+C67.273.295.974.795.282.378.996.879.492.067.2
S+M+B+C67.476.896.375.495.282.979.897.280.893.868.2
+ +feature concatenation operator. Compared with features from a single model, the accuracy of all downstream tasks is improved. When having fixed representations from all of 4 models, a linear classifier learned with SeA can outperform standard fine-tuning on 4 tasks (i.e., Caltech, DTD, Pets, and SUN) and achieve comparable performance on 2 tasks (i.e., CIFAR-10 and Flowers). It illustrates that collecting representations from multiple pre-training tasks can obtain complementary patterns that help mitigate the gap to different downstream tasks. + +Moreover, it can be observed that "S+C" performs better than "S+M" on DTD, while it is worse on Birds. The ensemble of all models, i.e., "S+M+B+C", shows the best performance on most tasks. Since the models in evaluation are pre-trained with different learning objectives, we find that the learned deep features complement each other. Therefore, more can be better. Note that ensemble can also be applied to fine-tuning. However, the computational cost will significantly increase depending on the number of models, which becomes unaffordable for real applications. In addition, its ensemble can still be worse than deep features. For example, an ensemble of fine-tuning on DTD achieves 78.9% that is about 1% worse than SeA. + +Although the result of SeA is promising, fixed features perform noticeably worse than fine-tuning on two tasks, e.g., Aircraft and Cars. It is because that the gap between the source domain and the target domain is too large. For example, ImageNet contains limited generic images for cars, while the target task is more fine-grained and consists of cars with different make, models, and + +years. In this case, a larger model pre-trained on a larger data set covering more diverse domains (e.g., JFT-3B [38]) may help improve the performance of SeA. + +Comparison on Ensemble of Features Then, we illustrate the effectiveness of SeA by comparing it to the baseline without SeA and existing feature-space augmentation methods in Tab. 5, which includes the semantic augmentation method, i.e., ISDA [35] and a robust adversarial augmentation method, i.e., ME-ADA [40]. The experiment is conducted with the ensembled features from 4 models for all methods. The hyper-parameters in ISDA and ME-ADA are searched for the desired result. + +Table 5: Comparison of SeA and benchmark methods on $\mathrm{S + M + B + C}$ . "Gain" denotes the improvement over the Baseline. + +
MethodsAircraftBirdsCaltechCarsCIFAR10CIFAR100DTDFlowersFoodPetsSUNAvg
Baseline63.673.495.469.294.881.179.496.279.593.166.981.1
ISDA63.373.695.969.494.981.379.695.679.693.267.081.2
ME-ADA64.474.995.770.995.081.679.895.880.393.267.481.7
SeA67.476.896.375.495.282.979.897.280.893.868.283.1
Gain3.83.40.96.20.41.80.41.01.30.71.32.0
+ +First, compared with the baseline, SeA improves the performance of deep features over all data sets, and the average gain is up to $2\%$ as denoted by "Gain". The consistent improvement shows that augmentation is important for learning with fixed deep features and the proposed augmentation helps mitigate the overfitting problem effectively. Second, while both SeA and ISDA focus on semantic augmentation, ISDA relies on input space augmentations to obtain the semantic directions, while our method leverages features from other examples. Therefore, when input space augmentations are unavailable for deep features, ISDA cannot beat the baseline and it shows that our method is more appropriate for learning without input space augmentations. In addition, ME-ADA improves the vanilla adversarial augmentation by mitigating its sensitivity to the gradient direction. However, the optimization only considers the features from the target example. By incorporating the information from other examples in the same batch, the proposed method can outperform ME-ADA with a margin of $1.4\%$ on average. + +# 5 Conclusion + +With the development of unsupervised representation learning, features with diverse information can be obtained by learning with different pretext tasks. Hence, we aim to investigate state-of-the-art performance of pre-trained deep features in this work. By introducing the novel semantic adversarial augmentation, we can show that learning with fixed features can achieve comparable performance to fine-tuning the whole network, but with way less cost. Investigating the efficacy of SeA in pre-training and fine-tuning can be our future work. + +# References + +1. Bossard, L., Guillaumeun, M., Gool, L.V.: Food-101 - mining discriminative components with random forests. In: Fleet, D.J., Pajdla, T., Schiele, B., Tuytelaars, T. (eds.) ECCV. Lecture Notes in Computer Science, vol. 8694, pp. 446-461. Springer (2014) +2. Boyd, S., Boyd, S.P., Vandenberghe, L.: Convex optimization. Cambridge university press (2004) +3. Caron, M., Misra, I., Mairal, J., Goyal, P., Bojanowski, P., Joulin, A.: Unsupervised learning of visual features by contrasting cluster assignments. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) NeurIPS (2020) +4. Chen, T., Cheng, Y., Gan, Z., Wang, J., Wang, L., Liu, J., Wang, Z.: Adversarial feature augmentation and normalization for visual recognition. Trans. Mach. Learn. Res. 2022 (2022) +5. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.E.: A simple framework for contrastive learning of visual representations. In: ICML. Proceedings of Machine Learning Research, vol. 119, pp. 1597-1607. PMLR (2020) +6. Chen, X., Xie, S., He, K.: An empirical study of training self-supervised vision transformers. In: ICCV. pp. 9620-9629. IEEE (2021) +7. Cimpoi, M., Maji, S., Kokkinos, I., Mohamed, S., Vedaldi, A.: Describing textures in the wild. In: CVPR. pp. 3606-3613 (2014) +8. Crammer, K., Singer, Y.: On the algorithmic implementation of multiclass kernel-based vector machines. J. Mach. Learn. Res. 2, 265-292 (2001) +9. Donahue, J., Jia, Y., Vinyals, O., Hoffman, J., Zhang, N., Tzeng, E., Darrell, T.: Decaf: A deep convolutional activation feature for generic visual recognition. In: ICML. JMLR Workshop and Conference Proceedings, vol. 32, pp. 647-655. JMLR.org (2014) +0. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR. OpenReview.net (2021) +1. Fei-Fei, L., Fergus, R., Perona, P.: Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In: CVPR workshop. pp. 178-178. IEEE (2004) +2. Goodfellow, I.J., Shlens, J., Szegedy, C.: Explaining and harnessing adversarial examples. In: Bengio, Y., LeCun, Y. (eds.) ICLR (2015) +3. Grill, J., Strub, F., Altché, F., Tallec, C., Richemond, P.H., Buchatskaya, E., Doersch, C., Pires, B.A., Guo, Z., Azar, M.G., Piot, B., Kavukcuoglu, K., Munos, R., Valko, M.: Bootstrap your own latent - A new approach to self-supervised learning. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) NeurIPS (2020) +4. He, K., Chen, X., Xie, S., Li, Y., Dollar, P., Girshick, R.B.: Masked autoencoders are scalable vision learners. In: CVPR. pp. 15979-15988. IEEE (2022) +5. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.B.: Momentum contrast for unsupervised visual representation learning. In: CVPR. pp. 9726-9735. Computer Vision Foundation / IEEE (2020) +6. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778. IEEE Computer Society (2016) +7. Ioffe, S., Szegedy, C.: Batch normalization: Accelerating deep network training by reducing internal covariate shift. In: Bach, F.R., Blei, D.M. (eds.) ICML. JMLR Workshop and Conference Proceedings, vol. 37, pp. 448-456. JMLR.org (2015) + +18. Kornblith, S., Shlens, J., Le, Q.V.: Do better imagenet models transfer better? In: CVPR. pp. 2661-2671. Computer Vision Foundation / IEEE (2019) +19. Krause, J., Stark, M., Deng, J., Fei-Fei, L.: 3d object representations for fine-grained categorization. In: ICCV workshop. pp. 554-561 (2013) +20. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009) +21. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. In: Bartlett, P.L., Pereira, F.C.N., Burges, C.J.C., Bottou, L., Weinberger, K.Q. (eds.) NeurIPS. pp. 1106-1114 (2012) +22. Liu, Z., Xu, Y., Xu, Y., Qian, Q., Li, H., Ji, X., Chan, A.B., Jin, R.: Improved fine-tuning by better leveraging pre-training data. In: NeurIPS (2022) +23. Maji, S., Rahtu, E., Kannala, J., Blaschko, M., Vedaldi, A.: Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151 (2013) +24. Nilsback, M., Zisserman, A.: Automated flower classification over a large number of classes. In: ICVGIP. pp. 722-729. IEEE Computer Society (2008) +25. Parkhi, O.M., Vedaldi, A., Zisserman, A., Jawahar, C.: Cats and dogs. In: CVPR. pp. 3498-3505. IEEE (2012) +26. Qian, Q., Hu, J., Li, H.: Hierarchically robust representation learning. In: CVPR. pp. 7334-7342. Computer Vision Foundation / IEEE (2020) +27. Qian, Q., Jin, R., Zhu, S., Lin, Y.: Fine-grained visual categorization via multi-stage metric learning. In: CVPR. pp. 3716-3724. IEEE Computer Society (2015) +28. Qian, Q., Shang, L., Sun, B., Hu, J., Li, H., Jin, R.: Softtriple loss: Deep metric learning without triplet sampling. In: ICCV. pp. 6449-6457. IEEE (2019) +29. Qian, Q., Xu, Y., Hu, J., Li, H., Jin, R.: Unsupervised visual representation learning by online constrained k-means. In: CVPR. pp. 16619-16628. IEEE (2022) +30. Ren, S., He, K., Girshick, R.B., Sun, J.: Faster R-CNN: towards real-time object detection with region proposal networks. IEEE Trans. Pattern Anal. Mach. Intell. 39(6), 1137-1149 (2017) +31. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M.S., Berg, A.C., Fei-Fei, L.: Imagenet large scale visual recognition challenge. Int. J. Comput. Vis. 115(3), 211-252 (2015) +32. Verma, V., Lamb, A., Beckham, C., Najafi, A., Mitliagkas, I., Lopez-Paz, D., Bengio, Y.: Manifold mixup: Better representations by interpolating hidden states. In: Chaudhuri, K., Salakhutdinov, R. (eds.) ICML. Proceedings of Machine Learning Research, vol. 97, pp. 6438-6447. PMLR (2019) +33. Volpi, R., Morerio, P., Savarese, S., Murino, V.: Adversarial feature augmentation for unsupervised domain adaptation. In: CVPR. pp. 5495-5504. Computer Vision Foundation / IEEE Computer Society (2018) +34. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The caltech-ucsd birds-200-2011 dataset (2011) +35. Wang, Y., Huang, G., Song, S., Pan, X., Xia, Y., Wu, C.: Regularizing deep networks with semantic data augmentation. IEEE Trans. Pattern Anal. Mach. Intell. 44(7), 3733-3748 (2022) +36. Xiao, J., Hays, J., Ehinger, K.A., Oliva, A., Torralba, A.: SUN database: Large-scale scene recognition from abbey to zoo. In: CVPR. pp. 3485-3492. IEEE Computer Society (2010) +37. Xu, Y., Qian, Q., Li, H., Jin, R., Hu, J.: Weakly supervised representation learning with coarse labels. In: ICCV. pp. 10573-10581. IEEE (2021) +38. Zhai, X., Kolesnikov, A., Houlsby, N., Beyer, L.: Scaling vision transformers. In: CVPR. pp. 1204-1213. IEEE (2022) + +39. Zhang, H., Cissé, M., Dauphin, Y.N., Lopez-Paz, D.: mixup: Beyond empirical risk minimization. In: ICLR. OpenReview.net (2018) +40. Zhao, L., Liu, T., Peng, X., Metaxas, D.N.: Maximum-entropy adversarial data augmentation for improved generalization and robustness. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) NeurIPS (2020) \ No newline at end of file diff --git a/2024/SeA_ Semantic Adversarial Augmentation for Last Layer Features from Unsupervised Representation Learning/images.zip b/2024/SeA_ Semantic Adversarial Augmentation for Last Layer Features from Unsupervised Representation Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3de1700e3d65c04bf5dcb523a166de9f46376e22 --- /dev/null +++ b/2024/SeA_ Semantic Adversarial Augmentation for Last Layer Features from Unsupervised Representation Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:93bd3afd46736b0628e43462793db8e969e3dd84f295105a96ce0022fbce5276 +size 387010 diff --git a/2024/SeA_ Semantic Adversarial Augmentation for Last Layer Features from Unsupervised Representation Learning/layout.json b/2024/SeA_ Semantic Adversarial Augmentation for Last Layer Features from Unsupervised Representation Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..badcf0373f24e56d68b501c86d25d0327d2ba37c --- /dev/null +++ b/2024/SeA_ Semantic Adversarial Augmentation for Last Layer Features from Unsupervised Representation Learning/layout.json @@ -0,0 +1,9267 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 149, + 111, + 465, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 111, + 465, + 166 + ], + "spans": [ + { + "bbox": [ + 149, + 111, + 465, + 166 + ], + "type": "text", + "content": "SeA: Semantic Adversarial Augmentation for Last Layer Features from Unsupervised Representation Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 201, + 186, + 412, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 186, + 412, + 200 + ], + "spans": [ + { + "bbox": [ + 201, + 186, + 412, + 200 + ], + "type": "text", + "content": "Qi Qian" + }, + { + "bbox": [ + 201, + 186, + 412, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 201, + 186, + 412, + 200 + ], + "type": "text", + "content": ", Yuanhong " + }, + { + "bbox": [ + 201, + 186, + 412, + 200 + ], + "type": "inline_equation", + "content": "\\mathrm{Xu}^{2}" + }, + { + "bbox": [ + 201, + 186, + 412, + 200 + ], + "type": "text", + "content": ", and Juhua " + }, + { + "bbox": [ + 201, + 186, + 412, + 200 + ], + "type": "inline_equation", + "content": "\\mathrm{Hu}^{3}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 215, + 207, + 397, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 215, + 207, + 397, + 219 + ], + "spans": [ + { + "bbox": [ + 215, + 207, + 397, + 219 + ], + "type": "text", + "content": "1 Alibaba Group, Bellevue, WA 98004, USA" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 229, + 220, + 380, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 229, + 220, + 380, + 229 + ], + "spans": [ + { + "bbox": [ + 229, + 220, + 380, + 229 + ], + "type": "text", + "content": "2 Alibaba Group, Hangzhou, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 223, + 230, + 389, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 230, + 389, + 241 + ], + "spans": [ + { + "bbox": [ + 223, + 230, + 389, + 241 + ], + "type": "text", + "content": "3 School of Engineering and Technology," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 200, + 242, + 413, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 242, + 413, + 252 + ], + "spans": [ + { + "bbox": [ + 200, + 242, + 413, + 252 + ], + "type": "text", + "content": "University of Washington, Tacoma, WA 98402, USA" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 177, + 254, + 436, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 177, + 254, + 436, + 264 + ], + "spans": [ + { + "bbox": [ + 177, + 254, + 436, + 264 + ], + "type": "text", + "content": "{qi.qian, yuanhong.xuyh}@alibaba-inc.com, juhuah@uw.edu" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 160, + 295, + 455, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 295, + 455, + 503 + ], + "spans": [ + { + "bbox": [ + 160, + 295, + 455, + 503 + ], + "type": "text", + "content": "Abstract. Deep features extracted from certain layers of a pre-trained deep model show superior performance over the conventional hand-crafted features. Compared with fine-tuning or linear probing that can explore diverse augmentations, e.g., random crop/flipping, in the original input space, the appropriate augmentations for learning with fixed deep features are more challenging and have been less investigated, which degenerates the performance. To unleash the potential of fixed deep features, we propose a novel semantic adversarial augmentation (SeA) in the feature space for optimization. Concretely, the adversarial direction implied by the gradient will be projected to a subspace spanned by other examples to preserve the semantic information. Then, deep features will be perturbed with the semantic direction, and augmented features will be applied to learn the classifier. Experiments are conducted on 11 benchmark downstream classification tasks with 4 popular pre-trained models. Our method is " + }, + { + "bbox": [ + 160, + 295, + 455, + 503 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 160, + 295, + 455, + 503 + ], + "type": "text", + "content": " better than the deep features without SeA on average. Moreover, compared to the expensive fine-tuning that is expected to give good performance, SeA shows a comparable performance on 6 out of 11 tasks, demonstrating the effectiveness of our proposal in addition to its efficiency." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 160, + 514, + 453, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 514, + 453, + 537 + ], + "spans": [ + { + "bbox": [ + 160, + 514, + 453, + 537 + ], + "type": "text", + "content": "Keywords: Semantic augmentation " + }, + { + "bbox": [ + 160, + 514, + 453, + 537 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 514, + 453, + 537 + ], + "type": "text", + "content": " Deep features " + }, + { + "bbox": [ + 160, + 514, + 453, + 537 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 514, + 453, + 537 + ], + "type": "text", + "content": " Unsupervised representation learning " + }, + { + "bbox": [ + 160, + 514, + 453, + 537 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 514, + 453, + 537 + ], + "type": "text", + "content": " Self-supervised learning" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 559, + 230, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 559, + 230, + 572 + ], + "spans": [ + { + "bbox": [ + 132, + 559, + 230, + 572 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 585, + 482, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 585, + 482, + 646 + ], + "spans": [ + { + "bbox": [ + 130, + 585, + 482, + 646 + ], + "type": "text", + "content": "Deep learning can be partially considered as a representation learning method that aims to extract features from raw data directly. By obtaining appropriate representations, a simple linear model can achieve state-of-the-art performance on challenging tasks, e.g., classification [16], object detection [30], etc. After the success of deep learning [21], researchers investigate the representations learned" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 234, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 234, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 234, + 665 + ], + "type": "text", + "content": "* Corresponding author" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 182, + 116, + 430, + 224 + ], + "blocks": [ + { + "bbox": [ + 182, + 116, + 430, + 224 + ], + "lines": [ + { + "bbox": [ + 182, + 116, + 430, + 224 + ], + "spans": [ + { + "bbox": [ + 182, + 116, + 430, + 224 + ], + "type": "image", + "image_path": "6fe0bae98de123087cdf64e4d44152274d04a431f794049f1b69f26c35b1985c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 236, + 482, + 314 + ], + "lines": [ + { + "bbox": [ + 130, + 236, + 482, + 314 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 482, + 314 + ], + "type": "text", + "content": "Fig. 1: Illustration of semantic adversarial augmentation (SeA). The red solid and empty circles denote the original data and its augmentation, respectively. Left: Conventional adversarial augmentation can perturb with arbitrary direction (e.g., augmentation may appear the same as the original); Right: SeA augments examples with semantic directions spanned by features from real data (e.g., different sectors show different subspaces for augmentation, where we can get more semantic meaningful augmentation)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 339, + 482, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 339, + 482, + 388 + ], + "spans": [ + { + "bbox": [ + 130, + 339, + 482, + 388 + ], + "type": "text", + "content": "from a large-scale data set, i.e., ImageNet [31]. Surprisingly, the deep features extracted from a certain layer of a pre-trained model can outperform handcrafted features on various downstream tasks [9], demonstrating the efficacy of the data-dependent representation learning mechanism implied by deep learning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 388, + 482, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 482, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 482, + 496 + ], + "type": "text", + "content": "Unlike hand-crafted features, representations obtained by deep learning are highly data/task-dependent as illustrated in [37]. When optimizing the task from the source domain, the neural network focuses on exploring the patterns related to the specific training task, while ignoring diverse information that has potential for different tasks. Hence, different representations can be learned from different training tasks even with the same training data from the source domain. Conventional methods learn representations with the labels of examples, which capture the knowledge only for the given labels and limit the information in deep features of each example." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 497, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 482, + 568 + ], + "type": "text", + "content": "To mitigate the problem, fine-tuning parameters of the whole network becomes prevalent for various downstream tasks [18]. On one hand, fine-tuning can benefit from the prior knowledge in the pre-trained representations. On the other hand, it can further optimize representations with diverse augmentations on the target task. Consequently, fine-tuning works better than learning with fixed representations on downstream tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "An important advantage of fine-tuning over learning with fixed deep features is the additional information from semantic augmentations. Obtaining effective augmentations in image space is convenient with semantic operators such as random crop, flipping, etc. In contrast, obtaining appropriate augmentations in the feature space for deep features becomes challenging, due to the lack of semantic operations. To leverage input space augmentations, a linear classifier can be learned with a frozen backbone by linear probing [6] that generates augmented images for optimization at each iteration." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "Q. Qian et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 319 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 319 + ], + "type": "text", + "content": "While fine-tuning and linear probing show promising performance, learning with fixed features is still attractive due to its good properties for real-world applications. First, it only applies deep models to extract features for each original example once, which is applicable for limited computational resources. On the contrary, fine-tuning has the full forward and backward pass and linear probing has the forward pass for each augmented example, which is more expensive for optimization. Second, the same models can be reused for different downstream tasks to extract deep features, while fine-tuning will adjust all parameters in pre-trained models. Thereafter, it has to keep a specific deep model for each task, which becomes intractable for handling hundreds of downstream tasks simultaneously. Finally, given fixed features, learning a linear classifier can be formulated as a convex problem that has the global optimum with a theoretical guarantee [2], while preserving the knowledge from the pre-trained model in features. It requires much less tuning efforts than fine-tuning that has to be tuned carefully to avoid the collapse of pre-trained parameters and catastrophic forgetting [22]. Therefore, we focus on fixed deep features extracted after the last pooling layer (i.e., inputs for the last fully-connected layer) in this work." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 319, + 479, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 479, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 479, + 380 + ], + "type": "text", + "content": "Although some existing works [32,35] consider augmentation in intermediate layers to help optimize the whole network, little efforts were devoted to the deep features after the last pooling layer. Moreover, the existing feature space augmentation method [35] is proposed to complement the input space augmentation. It relies on other augmentation techniques and cannot work well solely." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 380, + 481, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 481, + 511 + ], + "type": "text", + "content": "To tackle the problem, in this work, we propose a novel semantic adversarial augmentation strategy in the feature space of fixed deep features. Concretely, the gradient of each example is computed at first, and then a semantic direction can be observed by projecting the gradient to the subspace spanned by real data. Finally, the features of examples are augmented according to the obtained semantic direction for learning the linear classifier. Fig. 1 illustrates the proposed augmentation method that obtains effective augmentations with only fixed features. Moreover, to further mitigate the information loss in deep features from supervised pre-trained models, those from pre-trained models with different unsupervised pretext tasks are also investigated with an ensemble strategy. Our main contributions can be summarized as follows." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 521, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 521, + 481, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 521, + 481, + 556 + ], + "spans": [ + { + "bbox": [ + 138, + 521, + 481, + 556 + ], + "type": "text", + "content": "- We empirically demonstrate the current performance gap between fine-tuning and learning with fixed deep features. Those deep features extracted from 4 representative pre-trained models are evaluated on 11 downstream tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 558, + 481, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 558, + 481, + 604 + ], + "spans": [ + { + "bbox": [ + 138, + 558, + 481, + 604 + ], + "type": "text", + "content": "- To improve the performance of deep features, we propose a semantic adversarial augmentation method to obtain appropriate augmentations tailored for fixed features. In addition, a smoothed hinge loss is investigated to demonstrate the augmentation direction explicitly." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 605, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 605, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 605, + 481, + 665 + ], + "type": "text", + "content": "- Our proposed SeA gains " + }, + { + "bbox": [ + 138, + 605, + 481, + 665 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 138, + 605, + 481, + 665 + ], + "type": "text", + "content": " accuracy on average over 11 downstream tasks compared to the baseline using deep features without augmentations. Moreover, compared with fine-tuning that is expected to give good performance, our recipe for learning with deep features can achieve comparable performance on 6 out of 11 tasks with way less computational cost, which shows" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "text", + "content": "Semantic Adversarial Augmentation for Last Layer Features" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 148, + 116, + 480, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 116, + 480, + 140 + ], + "spans": [ + { + "bbox": [ + 148, + 116, + 480, + 140 + ], + "type": "text", + "content": "the potential of self-supervised pre-trained deep features and demonstrates both the efficiency and effectiveness of our proposal." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 157, + 237, + 170 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 157, + 237, + 170 + ], + "spans": [ + { + "bbox": [ + 132, + 157, + 237, + 170 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 181, + 233, + 193 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 181, + 233, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 181, + 233, + 193 + ], + "type": "text", + "content": "2.1 Deep Features" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 200, + 481, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 481, + 270 + ], + "type": "text", + "content": "Modern deep neural networks consist of multiple layers to exact representations from raw materials. After that, a simple linear model encoded by a fully-connected layer can be attached on top of the representations for classification. After pre-training on a large-scale data set, the obtained neural network can be considered as a feature extractor and a new fully-connected layer can be learned for target tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 271, + 481, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 271, + 481, + 366 + ], + "spans": [ + { + "bbox": [ + 130, + 271, + 481, + 366 + ], + "type": "text", + "content": "When the full label information of data is available, the learning objective is explicit and a supervised representation learning can be conducted by optimizing a conventional classification task, which is equivalent to optimizing the triplet loss as in distance metric learning [28]. The obtained deep features can outperform hand-crafted features on different applications, e.g., classification [9], distance metric learning [27], etc. Besides, some work aimed to obtain robust representations for different downstream tasks within the framework of classification [26]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 367, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 367, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 367, + 482, + 510 + ], + "type": "text", + "content": "Without any label information, the learning objective varies in unsupervised learning. First, each instance can be considered as an individual class and the representations can be learned by instance discrimination [5, 15]. In addition, clustering can be applied to capture the relationship between different instances. A coarse-grained classification task defined on clusters can be leveraged to optimize representations [3, 29]. Finally, other pretext tasks beyond classification also demonstrate the effective representation after pre-training [13]. Compared with supervised representation learning, the different objectives in unsupervised learning can learn various semantic information even from the same data set. In this work, we will systematically study representations from different pre-trained models and illustrate the performance gap compared to fine-tuning the whole neural network." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 527, + 320, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 527, + 320, + 540 + ], + "spans": [ + { + "bbox": [ + 132, + 527, + 320, + 540 + ], + "type": "text", + "content": "2.2 Augmentation in Feature Space" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 546, + 481, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 481, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 481, + 628 + ], + "type": "text", + "content": "Due to the over-parameterization property of deep neural networks, augmentation is essential for training deep models to avoid over-fitting [21]. Given an image, a perturbed copy can be observed by standard image operations such as random crop, flipping, color jitter, etc. [16]. The augmented images have a large variance while preserving the semantic information of the original image, i.e., an image of a cat is still a cat after augmentation, which helps train effective models with the additional information." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "text", + "content": "However, augmentation in feature space becomes more challenging due to the lack of semantic preserving operators. Some methods consider the augmentation of features from intermediate layers to help train the whole deep neural" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "Q. Qian et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "content": "networks in different applications [4, 32, 33, 35]. For example, [33] aims to learn feature augmentation with a feature generator for unsupervised domain adaptation. [4] tries to augment intermediate feature maps with adversarial feature moments in batch normalization [17] for efficient training. Manifold mixup [32] proposes to apply the original mixup [39] to features from multiple layers for augmentation, which also includes the layer investigated in this work. However, their study shows that the gain of mixup is mainly from the original image space and feature space of early stages, while that from the feature space of the last layer is negligible, which is consistent with our observation in the ablation study. ISDA [35] considers the semantic augmentation in feature space but it has to obtain the semantic direction with input space augmentations, which is complementary to input space augmentations but cannot work as the sole augmentation for deep features well. On the contrary, we propose SeA to project the adversarial direction with features of real data points. Moreover, our proposal is tailored for fixed features and is different from existing works that are for optimizing the whole network." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 327, + 369, + 340 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 327, + 369, + 340 + ], + "spans": [ + { + "bbox": [ + 132, + 327, + 369, + 340 + ], + "type": "text", + "content": "3 Semantic Adversarial Augmentation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 352, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 352, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 352, + 482, + 437 + ], + "type": "text", + "content": "We start the analysis with a standard classification framework. Let " + }, + { + "bbox": [ + 130, + 352, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\{x_{i},y_{i}\\}_{i = 1}^{n}" + }, + { + "bbox": [ + 130, + 352, + 482, + 437 + ], + "type": "text", + "content": " denote the training data set and " + }, + { + "bbox": [ + 130, + 352, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\{f_k\\}_{k = 1}^K" + }, + { + "bbox": [ + 130, + 352, + 482, + 437 + ], + "type": "text", + "content": " contains a set of " + }, + { + "bbox": [ + 130, + 352, + 482, + 437 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 352, + 482, + 437 + ], + "type": "text", + "content": " pre-trained models, which can be pre-trained with different learning objectives on different data sets. In this work, we directly use the representations extracted from pre-trained deep models and learn a simple linear model on top of it. Representations from different deep models are concatenated as the final representation for each example. Given " + }, + { + "bbox": [ + 130, + 352, + 482, + 437 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 130, + 352, + 482, + 437 + ], + "type": "text", + "content": ", the representation will be extracted as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 239, + 445, + 481, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 445, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 239, + 445, + 481, + 460 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {i} = \\left[ f _ {1} \\left(x _ {i}\\right), \\dots , f _ {K} \\left(x _ {i}\\right) \\right] \\in \\mathbb {R} ^ {d} \\tag {1}", + "image_path": "bc9454af7900cd6dedfc704d9f1f4a598a4fd7fe2ccb6c8a78f66481f11de5e4.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 468, + 481, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 481, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 481, + 491 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 468, + 481, + 491 + ], + "type": "inline_equation", + "content": "d = \\sum_{k}d_{k}" + }, + { + "bbox": [ + 130, + 468, + 481, + 491 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 468, + 481, + 491 + ], + "type": "inline_equation", + "content": "d_{k}" + }, + { + "bbox": [ + 130, + 468, + 481, + 491 + ], + "type": "text", + "content": " is the dimension of the representation from the " + }, + { + "bbox": [ + 130, + 468, + 481, + 491 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 468, + 481, + 491 + ], + "type": "text", + "content": "-th pre-trained deep model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 493, + 481, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 493, + 481, + 517 + ], + "spans": [ + { + "bbox": [ + 130, + 493, + 481, + 517 + ], + "type": "text", + "content": "With the above fixed deep features, a classification model can be learned by minimizing the empirical risk with the appropriate regularization as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 241, + 525, + 481, + 552 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 525, + 481, + 552 + ], + "spans": [ + { + "bbox": [ + 241, + 525, + 481, + 552 + ], + "type": "interline_equation", + "content": "\\min _ {\\mathbf {w}} \\sum_ {i} \\ell (\\mathbf {x} _ {i}, y _ {i}; \\mathbf {w}) + \\frac {\\tau}{2} \\| \\mathbf {w} \\| _ {F} ^ {2} \\tag {2}", + "image_path": "bb91157d0fc4ccec1b21b2676678efb6a90dfbfb16e655aabe3d4756e35f4c45.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 559, + 481, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 559, + 481, + 595 + ], + "spans": [ + { + "bbox": [ + 130, + 559, + 481, + 595 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 559, + 481, + 595 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 559, + 481, + 595 + ], + "type": "text", + "content": " is the weight for " + }, + { + "bbox": [ + 130, + 559, + 481, + 595 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 130, + 559, + 481, + 595 + ], + "type": "text", + "content": " regularization and " + }, + { + "bbox": [ + 130, + 559, + 481, + 595 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 130, + 559, + 481, + 595 + ], + "type": "text", + "content": " denotes parameters of the classification model, which is a linear classifier with the prediction probability of the " + }, + { + "bbox": [ + 130, + 559, + 481, + 595 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 559, + 481, + 595 + ], + "type": "text", + "content": "-th example on the " + }, + { + "bbox": [ + 130, + 559, + 481, + 595 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 559, + 481, + 595 + ], + "type": "text", + "content": "-th class as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 251, + 604, + 481, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 604, + 481, + 634 + ], + "spans": [ + { + "bbox": [ + 251, + 604, + 481, + 634 + ], + "type": "interline_equation", + "content": "p _ {i, j} = \\frac {\\exp \\left(\\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {j}\\right)}{\\sum_ {c = 1} ^ {C} \\exp \\left(\\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {c}\\right)} \\tag {3}", + "image_path": "121cb61c4de623e3f4d7020b6eb2a332135618863abb12af68b8f7780a4f0b09.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": " is the number of classes and " + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": " is a loss function for learning that will be discussed later." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 201, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 201, + 91, + 448, + 103 + ], + "type": "text", + "content": "Semantic Adversarial Augmentation for Last Layer Features" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 257, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 257, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 257, + 126 + ], + "type": "text", + "content": "3.1 Semantic Direction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 135, + 482, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 135, + 482, + 230 + ], + "spans": [ + { + "bbox": [ + 130, + 135, + 482, + 230 + ], + "type": "text", + "content": "In a standard fine-tuning pipeline, an image can be augmented by multiple random perturbations. When only deep features are available, it is hard to adopt existing techniques to generate appropriate augmented examples, since the semantic direction in the feature space is hard to capture. Manifold mixup [32] shows that randomly augmenting features from the last output layers has little improvement compared to the augmentation in input image space. To facilitate the performance of deep features, we investigate the semantic adversarial direction for augmentation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 231, + 481, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 231, + 481, + 255 + ], + "spans": [ + { + "bbox": [ + 131, + 231, + 481, + 255 + ], + "type": "text", + "content": "First, for the " + }, + { + "bbox": [ + 131, + 231, + 481, + 255 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 231, + 481, + 255 + ], + "type": "text", + "content": "-th example and its representation " + }, + { + "bbox": [ + 131, + 231, + 481, + 255 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 131, + 231, + 481, + 255 + ], + "type": "text", + "content": ", we consider its adversarial direction that can be obtained by maximizing the loss function [12]" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 255, + 263, + 481, + 282 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 263, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 255, + 263, + 481, + 282 + ], + "type": "interline_equation", + "content": "\\max _ {\\mathbf {x}: \\left\\| \\mathbf {x} - \\mathbf {x} _ {i} \\right\\| _ {2} \\leq \\gamma} \\ell \\left(\\mathbf {x} _ {i}, y _ {i}; \\mathbf {w}\\right) \\tag {4}", + "image_path": "efb43755af883db56f9410045b1294c73eb26c4f8a23db1c06e92af889c1acfe.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 289, + 481, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 289, + 481, + 323 + ], + "spans": [ + { + "bbox": [ + 131, + 289, + 481, + 323 + ], + "type": "text", + "content": "Note that the gradient indicates the ideal direction for the adversarial perturbation, and a standard adversarial example can be obtained by gradient ascent as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 268, + 334, + 480, + 346 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 268, + 334, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 268, + 334, + 480, + 346 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {x}} _ {i} = \\mathbf {x} _ {i} + \\eta \\nabla_ {\\mathbf {x} _ {i}} \\ell \\tag {5}", + "image_path": "1dc4c98d6773277a6f2565d3a8582c41321e0370bd3a67321fbadad4d91b715d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 354, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 437 + ], + "type": "text", + "content": "However, the gradient direction may not be semantically informative. [12] shows that the loss can be substantially increased with the adversarial example generated from the gradient, while the appearance is almost the same as the original image as illustrated in Fig. 1. Therefore, the conventional adversarial learning method helps improve the robustness to the adversarial attack but may be inappropriate for regularizing the generic learning task that requires additional information from diverse images." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 438, + 482, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 438, + 482, + 486 + ], + "spans": [ + { + "bbox": [ + 131, + 438, + 482, + 486 + ], + "type": "text", + "content": "To obtain the semantic direction, we propose to project the gradient direction to the subspace spanned by real data points. Concretely, let " + }, + { + "bbox": [ + 131, + 438, + 482, + 486 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_j\\}_{j=1}^b" + }, + { + "bbox": [ + 131, + 438, + 482, + 486 + ], + "type": "text", + "content": " denote a mini-batch of data and " + }, + { + "bbox": [ + 131, + 438, + 482, + 486 + ], + "type": "inline_equation", + "content": "\\mathbf{g}_i = \\nabla_{\\mathbf{x}_i}\\ell" + }, + { + "bbox": [ + 131, + 438, + 482, + 486 + ], + "type": "text", + "content": " indicate the gradient of " + }, + { + "bbox": [ + 131, + 438, + 482, + 486 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 131, + 438, + 482, + 486 + ], + "type": "text", + "content": ". For the " + }, + { + "bbox": [ + 131, + 438, + 482, + 486 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 438, + 482, + 486 + ], + "type": "text", + "content": "-th image, a semantic adversarial direction " + }, + { + "bbox": [ + 131, + 438, + 482, + 486 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{g}}_i" + }, + { + "bbox": [ + 131, + 438, + 482, + 486 + ], + "type": "text", + "content": " consists of representations from data as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 217, + 494, + 481, + 519 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 494, + 481, + 519 + ], + "spans": [ + { + "bbox": [ + 217, + 494, + 481, + 519 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {g}} _ {i} = \\sum_ {j: j \\neq i} q _ {j} \\mathbf {x} _ {j}; \\quad \\mathbf {q} ^ {*} = \\arg \\min _ {\\mathbf {q} \\in \\Delta} \\mathcal {D} (\\hat {\\mathbf {g}} _ {i}, \\mathbf {g} _ {i}) \\tag {6}", + "image_path": "c88de891de158065da8cad4e268dfb207b84afafa0bfd3098c381310d79141ee.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 528, + 482, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 528, + 482, + 563 + ], + "spans": [ + { + "bbox": [ + 131, + 528, + 482, + 563 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 528, + 482, + 563 + ], + "type": "inline_equation", + "content": "\\mathcal{D}(\\cdot, \\cdot)" + }, + { + "bbox": [ + 131, + 528, + 482, + 563 + ], + "type": "text", + "content": " is a distance function and " + }, + { + "bbox": [ + 131, + 528, + 482, + 563 + ], + "type": "inline_equation", + "content": "\\varDelta" + }, + { + "bbox": [ + 131, + 528, + 482, + 563 + ], + "type": "text", + "content": " is a simplex as " + }, + { + "bbox": [ + 131, + 528, + 482, + 563 + ], + "type": "inline_equation", + "content": "\\varDelta=\\{\\mathbf{q}\\in\\mathbb{R}^{b-1}|\\sum_{j}q_{j}=1;\\forall j,q_{j}\\geq 0\\}" + }, + { + "bbox": [ + 131, + 528, + 482, + 563 + ], + "type": "text", + "content": ". With the projection, we aim to find an adversarial direction in a subspace consisting of original data points as illustrated in Fig. 1." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 564, + 481, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 564, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 564, + 481, + 601 + ], + "type": "text", + "content": "The last challenge is to obtain " + }, + { + "bbox": [ + 130, + 564, + 481, + 601 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 130, + 564, + 481, + 601 + ], + "type": "text", + "content": " efficiently with a distance function. By adopting the squared Euclidean distance, which is a standard distance measurement, the optimization problem for " + }, + { + "bbox": [ + 130, + 564, + 481, + 601 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 130, + 564, + 481, + 601 + ], + "type": "text", + "content": " can be written as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 235, + 609, + 481, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 609, + 481, + 634 + ], + "spans": [ + { + "bbox": [ + 235, + 609, + 481, + 634 + ], + "type": "interline_equation", + "content": "\\min _ {\\mathbf {q} \\in \\Delta} \\| \\sum_ {j: j \\neq i} q _ {j} \\mathbf {x} _ {j} - \\mathbf {g} _ {i} \\| _ {2} ^ {2} - \\alpha H (\\mathbf {q}) \\tag {7}", + "image_path": "ec8da37533cb8af6eefc732fd64560c742132e81a5acd2d0df33985fe3a33a97.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "H(\\mathbf{q})" + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": " is the entropy of " + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": " that helps improve the robustness to different batches. When normalizing features such that " + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\| \\mathbf{x}_j\\| _2 = \\| \\mathbf{g}_i\\| _2 = 1" + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": " to obtain" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "Q. Qian et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 136, + 132, + 481, + 232 + ], + "blocks": [ + { + "bbox": [ + 133, + 116, + 473, + 129 + ], + "lines": [ + { + "bbox": [ + 133, + 116, + 473, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 473, + 129 + ], + "type": "text", + "content": "Algorithm 1 Semantic Adversarial Augmentation (SeA) for Given Features" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 136, + 132, + 481, + 232 + ], + "lines": [ + { + "bbox": [ + 136, + 132, + 481, + 232 + ], + "spans": [ + { + "bbox": [ + 136, + 132, + 481, + 232 + ], + "type": "text", + "content": "1: Input: Dataset " + }, + { + "bbox": [ + 136, + 132, + 481, + 232 + ], + "type": "inline_equation", + "content": "\\{x_i,y_i\\}" + }, + { + "bbox": [ + 136, + 132, + 481, + 232 + ], + "type": "text", + "content": ", pre-trained models " + }, + { + "bbox": [ + 136, + 132, + 481, + 232 + ], + "type": "inline_equation", + "content": "\\{f_k\\}" + }, + { + "bbox": [ + 136, + 132, + 481, + 232 + ], + "type": "text", + "content": ", iterations " + }, + { + "bbox": [ + 136, + 132, + 481, + 232 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 136, + 132, + 481, + 232 + ], + "type": "text", + "content": ", linear model " + }, + { + "bbox": [ + 136, + 132, + 481, + 232 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 136, + 132, + 481, + 232 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 136, + 132, + 481, + 232 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 136, + 132, + 481, + 232 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 136, + 132, + 481, + 232 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 136, + 132, + 481, + 232 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 136, + 132, + 481, + 232 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 136, + 132, + 481, + 232 + ], + "type": "text", + "content": " \n2: Extract deep features by " + }, + { + "bbox": [ + 136, + 132, + 481, + 232 + ], + "type": "inline_equation", + "content": "\\{f_k\\}" + }, + { + "bbox": [ + 136, + 132, + 481, + 232 + ], + "type": "text", + "content": " for all examples as " + }, + { + "bbox": [ + 136, + 132, + 481, + 232 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_i\\}" + }, + { + "bbox": [ + 136, + 132, + 481, + 232 + ], + "type": "text", + "content": " \n3: for " + }, + { + "bbox": [ + 136, + 132, + 481, + 232 + ], + "type": "inline_equation", + "content": "t = 1,\\dots,T" + }, + { + "bbox": [ + 136, + 132, + 481, + 232 + ], + "type": "text", + "content": " do \n4: Receive a mini-batch of examples " + }, + { + "bbox": [ + 136, + 132, + 481, + 232 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_i,y_i\\}_{i=1}^b" + }, + { + "bbox": [ + 136, + 132, + 481, + 232 + ], + "type": "text", + "content": " \n5: Obtain augmented examples " + }, + { + "bbox": [ + 136, + 132, + 481, + 232 + ], + "type": "inline_equation", + "content": "\\{\\tilde{\\mathbf{x}}_i,y_i\\}" + }, + { + "bbox": [ + 136, + 132, + 481, + 232 + ], + "type": "text", + "content": " as in Eqn. 11 \n6: Optimize " + }, + { + "bbox": [ + 136, + 132, + 481, + 232 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 136, + 132, + 481, + 232 + ], + "type": "text", + "content": " by SGD: " + }, + { + "bbox": [ + 136, + 132, + 481, + 232 + ], + "type": "inline_equation", + "content": "\\mathbf{w} = \\mathbf{w} - \\eta_w\\left(\\frac{1}{b}\\sum_i^b\\nabla_{\\mathbf{w}}\\ell(\\tilde{\\mathbf{x}}_i,y_i;\\mathbf{w}) + \\tau \\mathbf{w}\\right)" + }, + { + "bbox": [ + 136, + 132, + 481, + 232 + ], + "type": "text", + "content": " \n7: end for \n8: return " + }, + { + "bbox": [ + 136, + 132, + 481, + 232 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "algorithm" + }, + { + "bbox": [ + 130, + 255, + 479, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 255, + 479, + 279 + ], + "spans": [ + { + "bbox": [ + 130, + 255, + 479, + 279 + ], + "type": "text", + "content": "the direction without the influence from the magnitude, the problem can be upper-bounded as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 285, + 418, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 285, + 418, + 300 + ], + "spans": [ + { + "bbox": [ + 131, + 285, + 418, + 300 + ], + "type": "text", + "content": "Proposition 1 With unit length variables " + }, + { + "bbox": [ + 131, + 285, + 418, + 300 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_j\\}" + }, + { + "bbox": [ + 131, + 285, + 418, + 300 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 285, + 418, + 300 + ], + "type": "inline_equation", + "content": "\\mathbf{g}_i" + }, + { + "bbox": [ + 131, + 285, + 418, + 300 + ], + "type": "text", + "content": ", we have" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 221, + 306, + 481, + 332 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 221, + 306, + 481, + 332 + ], + "spans": [ + { + "bbox": [ + 221, + 306, + 481, + 332 + ], + "type": "interline_equation", + "content": "\\left\\| \\sum_ {j: j \\neq i} q _ {j} \\mathbf {x} _ {j} - \\mathbf {g} _ {i} \\right\\| _ {2} ^ {2} \\leq 2 - 2 \\sum_ {j} q _ {j} \\mathbf {x} _ {j} ^ {\\top} \\mathbf {g} _ {i} \\tag {8}", + "image_path": "b46906784beef23599124ba9038b5dba3f68ef9e70c7563b0f51c37c287d4e4b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 339, + 373, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 339, + 373, + 351 + ], + "spans": [ + { + "bbox": [ + 131, + 339, + 373, + 351 + ], + "type": "text", + "content": "The detailed proof can be found in the supplementary." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 351, + 479, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 351, + 479, + 374 + ], + "spans": [ + { + "bbox": [ + 130, + 351, + 479, + 374 + ], + "type": "text", + "content": "By rearranging the terms, we can maximize the lower-bound of the original problem as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 245, + 382, + 481, + 409 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 245, + 382, + 481, + 409 + ], + "spans": [ + { + "bbox": [ + 245, + 382, + 481, + 409 + ], + "type": "interline_equation", + "content": "\\max _ {\\mathbf {q} \\in \\Delta} \\sum_ {j: j \\neq i} q _ {j} \\mathbf {x} _ {j} ^ {\\top} \\mathbf {g} _ {i} + \\alpha H (\\mathbf {q}) \\tag {9}", + "image_path": "e9d7ba38fa34486c475404b8ef9328704ecc3e3c8aab86eb60b601156fe99f7b.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 146, + 416, + 446, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 416, + 446, + 429 + ], + "spans": [ + { + "bbox": [ + 146, + 416, + 446, + 429 + ], + "type": "text", + "content": "According to the K.K.T. condition [2], " + }, + { + "bbox": [ + 146, + 416, + 446, + 429 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 146, + 416, + 446, + 429 + ], + "type": "text", + "content": " has a closed-form solution." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 435, + 428, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 435, + 428, + 447 + ], + "spans": [ + { + "bbox": [ + 131, + 435, + 428, + 447 + ], + "type": "text", + "content": "Proposition 2 The problem in Eqn. 9 has the optimal solution as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 204, + 453, + 481, + 485 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 204, + 453, + 481, + 485 + ], + "spans": [ + { + "bbox": [ + 204, + 453, + 481, + 485 + ], + "type": "interline_equation", + "content": "q _ {j} = \\frac {\\exp \\left(\\mathbf {x} _ {j} ^ {\\top} \\mathbf {g} _ {i} / \\alpha\\right)}{Z}; \\quad Z = \\sum_ {k: k \\neq i} \\exp \\left(\\mathbf {x} _ {k} ^ {\\top} \\mathbf {g} _ {i} / \\alpha\\right) \\tag {10}", + "image_path": "df27cb2dc1c04cb18af89c37c53fe626146e01c92bc168332c30831ab2ce1a8f.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 493, + 479, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 493, + 479, + 517 + ], + "spans": [ + { + "bbox": [ + 130, + 493, + 479, + 517 + ], + "type": "text", + "content": "Given the semantic adversarial direction, the target example with semantic perturbation can be obtained as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 255, + 525, + 481, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 525, + 481, + 539 + ], + "spans": [ + { + "bbox": [ + 255, + 525, + 481, + 539 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {x}} _ {i} = \\Pi \\left(\\mathbf {x} _ {i} + \\eta \\Pi \\left(\\hat {\\mathbf {g}} _ {i}\\right)\\right) \\tag {11}", + "image_path": "8cf32c601879bcfb393962751cf5f3e4094454fb652fa77e6fdb6aa5fbf4b575.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 546, + 481, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 481, + 606 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 481, + 606 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 546, + 481, + 606 + ], + "type": "inline_equation", + "content": "\\Pi(\\cdot)" + }, + { + "bbox": [ + 130, + 546, + 481, + 606 + ], + "type": "text", + "content": " normalizes the vector to the unit length if required and " + }, + { + "bbox": [ + 130, + 546, + 481, + 606 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 130, + 546, + 481, + 606 + ], + "type": "text", + "content": " denotes the step size for augmentation. Compared with the adversarial perturbation in Eqn. 5, the augmentation in Eqn. 11 projects the gradient to the direction consisting of real data points, which can capture the semantic information in the feature space effectively. Alg. 1 summarizes the proposed method." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 131, + 622, + 342, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 622, + 342, + 634 + ], + "spans": [ + { + "bbox": [ + 131, + 622, + 342, + 634 + ], + "type": "text", + "content": "3.2 Illustration of Adversarial Direction" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "In this subsection, we will generalize the cross entropy loss to help illustrate the proposed augmentation strategy." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 201, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 201, + 91, + 448, + 102 + ], + "type": "text", + "content": "Semantic Adversarial Augmentation for Last Layer Features" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 479, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 479, + 139 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 479, + 139 + ], + "type": "text", + "content": "The standard multi-class hinge loss [8] for convex optimization can be written as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 203, + 149, + 480, + 169 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 149, + 480, + 169 + ], + "spans": [ + { + "bbox": [ + 203, + 149, + 480, + 169 + ], + "type": "interline_equation", + "content": "\\ell (\\mathbf {x} _ {i}, y _ {i}; \\mathbf {w}) = \\max \\left\\{0, \\delta + \\max _ {c \\neq y _ {i}} \\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {c} - \\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {y _ {i}} \\right\\} \\tag {12}", + "image_path": "2653880f5f67be5c77d5fa65df1bd705eb5bb3cb2a95d3cdf0dbebe38faaf8b3.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 178, + 275, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 178, + 275, + 190 + ], + "spans": [ + { + "bbox": [ + 132, + 178, + 275, + 190 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 178, + 275, + 190 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 132, + 178, + 275, + 190 + ], + "type": "text", + "content": " is a pre-defined margin." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 190, + 479, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 190, + 479, + 237 + ], + "spans": [ + { + "bbox": [ + 131, + 190, + 479, + 237 + ], + "type": "text", + "content": "Unlike convex optimization, cross-entropy loss is prevalent in deep learning, which is a smooth function and can help accelerate the convergence [2]. Therefore, we propose to obtain a smoothed hinge loss by introducing a distribution over logits from different classes " + }, + { + "bbox": [ + 131, + 190, + 479, + 237 + ], + "type": "inline_equation", + "content": "\\mathbf{p} \\in \\Delta" + }, + { + "bbox": [ + 131, + 190, + 479, + 237 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 131, + 190, + 479, + 237 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 131, + 190, + 479, + 237 + ], + "type": "text", + "content": " is the simplex." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 146, + 239, + 343, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 239, + 343, + 251 + ], + "spans": [ + { + "bbox": [ + 146, + 239, + 343, + 251 + ], + "type": "text", + "content": "First, the original hinge loss is equivalent to" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 178, + 260, + 479, + 285 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 260, + 479, + 285 + ], + "spans": [ + { + "bbox": [ + 178, + 260, + 479, + 285 + ], + "type": "interline_equation", + "content": "\\ell \\left(\\mathbf {x} _ {i}, y _ {i}; \\mathbf {w}\\right) = \\max _ {\\mathbf {p} \\in \\Delta} p _ {y _ {i}} \\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {y _ {i}} + \\sum_ {c \\neq y _ {i}} p _ {c} \\left(\\delta + \\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {c}\\right) - \\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {y _ {i}} \\tag {13}", + "image_path": "a3f29822113f7782ca7b91b13eedd733345189f7da7fb911466d7620472aa0b2.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 295, + 479, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 295, + 479, + 319 + ], + "spans": [ + { + "bbox": [ + 132, + 295, + 479, + 319 + ], + "type": "text", + "content": "According to the analysis for cross entropy loss [28], the loss can be smoothed by adding an entropy regularization for " + }, + { + "bbox": [ + 132, + 295, + 479, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 132, + 295, + 479, + 319 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 151, + 329, + 479, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 329, + 479, + 354 + ], + "spans": [ + { + "bbox": [ + 151, + 329, + 479, + 354 + ], + "type": "interline_equation", + "content": "\\ell \\left(\\mathbf {x} _ {i}, y _ {i}; \\mathbf {w}\\right) = \\max _ {\\mathbf {p} \\in \\Delta} p _ {y _ {i}} \\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {y _ {i}} + \\sum_ {c \\neq y _ {i}} p _ {c} (\\delta + \\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {c}) + \\lambda H (\\mathbf {p}) - \\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {y _ {i}} \\tag {14}", + "image_path": "53e27558f87cc83628a3be1380ead185b5d9411cd2a0a8c035670d5dbf22bf3c.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 364, + 402, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 402, + 376 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 402, + 376 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 364, + 402, + 376 + ], + "type": "inline_equation", + "content": "H(\\mathbf{p})" + }, + { + "bbox": [ + 132, + 364, + 402, + 376 + ], + "type": "text", + "content": " denotes the entropy of " + }, + { + "bbox": [ + 132, + 364, + 402, + 376 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 132, + 364, + 402, + 376 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 364, + 402, + 376 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 132, + 364, + 402, + 376 + ], + "type": "text", + "content": " is the coefficient." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 387, + 402, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 387, + 402, + 399 + ], + "spans": [ + { + "bbox": [ + 132, + 387, + 402, + 399 + ], + "type": "text", + "content": "Proposition 3 The loss function in Eqn. 14 is equivalent to" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 159, + 408, + 479, + 437 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 408, + 479, + 437 + ], + "spans": [ + { + "bbox": [ + 159, + 408, + 479, + 437 + ], + "type": "interline_equation", + "content": "\\ell (\\mathbf {x} _ {i}, y _ {i}; \\mathbf {w}) = - \\lambda \\log \\frac {\\exp \\left(\\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {y _ {i}} / \\lambda\\right)}{\\exp \\left(\\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {y _ {i}} / \\lambda\\right) + \\sum_ {c \\neq y _ {i}} \\exp \\left(\\left(\\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {c} + \\delta\\right) / \\lambda\\right)} \\tag {15}", + "image_path": "4bc2612eab1081933f588a108af3658a8b4844adebfbe3c37f698669793e5cc3.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 446, + 479, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 446, + 479, + 481 + ], + "spans": [ + { + "bbox": [ + 131, + 446, + 479, + 481 + ], + "type": "text", + "content": "Remark It is obvious that the popular cross entropy loss is a special case of Eqn. 15 by letting " + }, + { + "bbox": [ + 131, + 446, + 479, + 481 + ], + "type": "inline_equation", + "content": "\\delta = 0" + }, + { + "bbox": [ + 131, + 446, + 479, + 481 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 446, + 479, + 481 + ], + "type": "inline_equation", + "content": "\\lambda = 1" + }, + { + "bbox": [ + 131, + 446, + 479, + 481 + ], + "type": "text", + "content": ". Our analysis connects the hinge loss in conventional methods to the popular loss function in deep learning." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 482, + 479, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 482, + 479, + 506 + ], + "spans": [ + { + "bbox": [ + 132, + 482, + 479, + 506 + ], + "type": "text", + "content": "Finally, since deep features are fixed, we can illustrate the gradient direction with the proposed smoothed loss function explicitly." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 514, + 453, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 453, + 527 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 453, + 527 + ], + "type": "text", + "content": "Proposition 4 Given the loss function in Eqn. 15, the gradient of " + }, + { + "bbox": [ + 132, + 514, + 453, + 527 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 132, + 514, + 453, + 527 + ], + "type": "text", + "content": " is" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 233, + 536, + 479, + 569 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 536, + 479, + 569 + ], + "spans": [ + { + "bbox": [ + 233, + 536, + 479, + 569 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\mathbf {x} _ {i}} \\ell (\\mathbf {x} _ {i}, y _ {i}; \\mathbf {w}) = \\sum_ {c = 1} ^ {C} p _ {c} \\mathbf {w} _ {c} - \\mathbf {w} _ {y _ {i}} \\tag {16}", + "image_path": "ee00b0387e818de3136b1a11f7ddfd3a5cfa3a632944a896d013d4964aacb782.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 578, + 489, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 489, + 622 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 489, + 622 + ], + "type": "interline_equation", + "content": "\\begin{array}{r} w h e r e p _ {c} = \\left\\{ \\begin{array}{l l} \\frac {\\exp ((\\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {c} + \\delta) / \\lambda)}{Z} c \\neq y _ {i} \\\\ \\frac {\\exp (\\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {c} / \\lambda)}{Z} c = y _ {i} \\end{array} a n d Z = \\exp (\\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {y _ {i}} / \\lambda) + \\sum_ {c \\neq y _ {i}} \\exp ((\\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {c} + \\delta) / \\lambda) \\leqslant 1, \\right. \\\\ \\delta) / \\lambda). \\end{array}", + "image_path": "76994aa66e8404e818081ea9f6dff9832ddfff3dae4db465c79c4454af36b0b0.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 629, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 629, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 629, + 479, + 665 + ], + "type": "text", + "content": "Remark The gradient direction in the feature space indicates that an adversarial perturbation should be close to the primary directions of other classes, while far away from the direction of the corresponding class." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "Q. Qian et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 143, + 482, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 143, + 482, + 288 + ], + "spans": [ + { + "bbox": [ + 130, + 143, + 482, + 288 + ], + "type": "text", + "content": "To demonstrate our proposed method, we adopt state-of-the-art and widely used pre-trained models to extract deep features for evaluation. It should be noted that only two main architectures, including ResNet-50 [16] and vision transformer (ViT) [10], are prevalently applied in self-supervised learning [6,13,14,29]. Since ViT shows a worse classification performance than ResNet-50 with the frozen backbone as demonstrated in [14], we will focus on ResNet-50 with different public pre-trained parameters in the experiment. Specifically, one supervised pre-trained ResNet-50 and three self-supervised pre-trained ResNet-50 are applied for feature extraction. All of these models are pre-trained on ImageNet-1K [31] with different learning objectives. The details of different models can be found in the supplementary. The accuracy of the supervised model and that of linear probing for unsupervised models are summarized in Tab. 1." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 157, + 329, + 457, + 389 + ], + "blocks": [ + { + "bbox": [ + 182, + 308, + 430, + 319 + ], + "lines": [ + { + "bbox": [ + 182, + 308, + 430, + 319 + ], + "spans": [ + { + "bbox": [ + 182, + 308, + 430, + 319 + ], + "type": "text", + "content": "Table 1: Performance of 4 pre-trained models on ImageNet." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 157, + 329, + 457, + 389 + ], + "lines": [ + { + "bbox": [ + 157, + 329, + 457, + 389 + ], + "spans": [ + { + "bbox": [ + 157, + 329, + 457, + 389 + ], + "type": "table", + "html": "
Model#Pre-training EpochsObjectiveAcc%
Supervised [16] (S)901,000-class classification76.6
MoCo-v3 [6] (M)1,000instance discrimination74.6
BYOL [13] (B)1,000regression74.3
CoKe [29] (C)1,000cluster discrimination74.9
", + "image_path": "efd01e69004db328cd85a6505de852edaf04f3f4ef9eb59a62ed2d5dd49af3fd.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 413, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 482, + 521 + ], + "type": "text", + "content": "11 diverse downstream data sets are applied for evaluation, including Aircraft [23], Caltech101 [11], Stanford Cars [19], CIFAR-10 [20], CIFAR-100 [20], CUB200-2011 (Birds) [34], Describable Textures Dataset (DTD) [7], Flowers [24], Food101 [1], Oxford-IIIT Pet (Pets) [25], and Sun397 [36]. We follow the evaluation protocol in [13]. Concretely, all models search hyper-parameters on the provided/generated validation set in each downstream task, and the standard metric on the provided test set is reported. Besides, mean per-class accuracy is reported on Aircraft, Caltech101, Flowers, and Pets, while Top-1 accuracy is utilized for other data sets. More details can be found in [13]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 521, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 594 + ], + "type": "text", + "content": "Each model is fine-tuned with SGD by 100 epochs for sufficient training. The batch size is 256 and the momentum is 0.9. The learning rate is searched in a range of 7 logarithmically-spaced values between " + }, + { + "bbox": [ + 130, + 521, + 482, + 594 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 130, + 521, + 482, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 521, + 482, + 594 + ], + "type": "inline_equation", + "content": "10^{-1}" + }, + { + "bbox": [ + 130, + 521, + 482, + 594 + ], + "type": "text", + "content": ". Weight decay is optional, for which if it is applied, the value will be searched with the same setting between " + }, + { + "bbox": [ + 130, + 521, + 482, + 594 + ], + "type": "inline_equation", + "content": "10^{-6}" + }, + { + "bbox": [ + 130, + 521, + 482, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 521, + 482, + 594 + ], + "type": "inline_equation", + "content": "10^{-3}" + }, + { + "bbox": [ + 130, + 521, + 482, + 594 + ], + "type": "text", + "content": ". The standard augmentation, i.e., random crop, random horizontal flipping, is applied as in most existing fine-tuning pipelines." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "For a fair comparison, we also apply SGD to learn the linear classifier with fixed deep features using the same batch size and momentum. Unlike fine-tuning, we have the constant learning rate for our method, which is searched in " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\{2^i\\}_{i = -2}^{i = 3}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " while weight decay is searched in " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\{0,10^{-6},10^{-5},10^{-4}\\}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ". For other parameters, we have " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " searched in " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\{0.01,0.02,0.05\\}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\{0.05,0.1,1\\}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\{0,1\\}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\{0,0.2,0.4,0.8\\}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "text", + "content": "Semantic Adversarial Augmentation for Last Layer Features" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "For each image, we extract the output of the final pooling layer as deep features. Features from an individual model are normalized to the unit length. If multiple models are exploited, features from different models are concatenated as in Eqn. 1 and the combined feature vector is further normalized to the unit length. The only augmentation for training with fixed features is the proposed semantic adversarial augmentation. To approximate the direction of the gradient for each example, a mini-batch of data is leveraged to optimize the problem in Eqn. 9. The projection for normalization is also applied to the augmented examples." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 245, + 238, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 245, + 238, + 257 + ], + "spans": [ + { + "bbox": [ + 131, + 245, + 238, + 257 + ], + "type": "text", + "content": "4.1 Ablation Study" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 269, + 481, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 269, + 481, + 317 + ], + "spans": [ + { + "bbox": [ + 130, + 269, + 481, + 317 + ], + "type": "text", + "content": "Before the experiments on downstream tasks, we first investigate the effect of parameters in the proposed augmentation. The ablation study is conducted on CIFAR-100 with features concatenated from 4 pre-trained models and the accuracy is reported for comparison." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 338, + 481, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 338, + 481, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 481, + 399 + ], + "type": "text", + "content": "Effect of Step Size for Augmentation First, we evaluate the step size " + }, + { + "bbox": [ + 130, + 338, + 481, + 399 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 130, + 338, + 481, + 399 + ], + "type": "text", + "content": " in the semantic adversarial augmentation. For SeA, we fix " + }, + { + "bbox": [ + 130, + 338, + 481, + 399 + ], + "type": "inline_equation", + "content": "\\lambda = 0.1" + }, + { + "bbox": [ + 130, + 338, + 481, + 399 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 338, + 481, + 399 + ], + "type": "inline_equation", + "content": "\\delta = 0" + }, + { + "bbox": [ + 130, + 338, + 481, + 399 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 338, + 481, + 399 + ], + "type": "inline_equation", + "content": "\\alpha = 0.01" + }, + { + "bbox": [ + 130, + 338, + 481, + 399 + ], + "type": "text", + "content": ", and the learning rate as 1, according to the performance on the validation set. The weight of the augmentation " + }, + { + "bbox": [ + 130, + 338, + 481, + 399 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 130, + 338, + 481, + 399 + ], + "type": "text", + "content": " varies in " + }, + { + "bbox": [ + 130, + 338, + 481, + 399 + ], + "type": "inline_equation", + "content": "\\{0, 0.2, 0.4, 0.6, 0.8, 1.0\\}" + }, + { + "bbox": [ + 130, + 338, + 481, + 399 + ], + "type": "text", + "content": " and the performance is summarized in Fig. 2-Fig. 3." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 135, + 419, + 246, + 502 + ], + "blocks": [ + { + "bbox": [ + 135, + 419, + 246, + 502 + ], + "lines": [ + { + "bbox": [ + 135, + 419, + 246, + 502 + ], + "spans": [ + { + "bbox": [ + 135, + 419, + 246, + 502 + ], + "type": "image", + "image_path": "8af8176113b60714f697ac9f6830a26ba36edbd654a5dfd3770e0015a38a3082.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 133, + 510, + 479, + 533 + ], + "lines": [ + { + "bbox": [ + 133, + 510, + 479, + 533 + ], + "spans": [ + { + "bbox": [ + 133, + 510, + 479, + 533 + ], + "type": "text", + "content": "Fig. 2: Training accuracy Fig. 3: Test accuracy of Fig. 4: Test accuracy of various different " + }, + { + "bbox": [ + 133, + 510, + 479, + 533 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 133, + 510, + 479, + 533 + ], + "type": "text", + "content": ". different " + }, + { + "bbox": [ + 133, + 510, + 479, + 533 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 133, + 510, + 479, + 533 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 246, + 419, + 355, + 502 + ], + "blocks": [ + { + "bbox": [ + 246, + 419, + 355, + 502 + ], + "lines": [ + { + "bbox": [ + 246, + 419, + 355, + 502 + ], + "spans": [ + { + "bbox": [ + 246, + 419, + 355, + 502 + ], + "type": "image", + "image_path": "507bbac29dc0984305db8db92063b388e0db13e17504b0689ef7b587ab9354be.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 362, + 422, + 471, + 501 + ], + "blocks": [ + { + "bbox": [ + 362, + 422, + 471, + 501 + ], + "lines": [ + { + "bbox": [ + 362, + 422, + 471, + 501 + ], + "spans": [ + { + "bbox": [ + 362, + 422, + 471, + 501 + ], + "type": "image", + "image_path": "657555cd25279a74e4a27c3fa7884071d1d0a047c676f86fc08d378ca235c4b0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "When " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\eta = 0" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": ", there is no augmentation applied for training. Hence, we can observe that the training accuracy exceeds " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "92\\%" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " after only 20 epochs. Due to the over-fitting on the training set, the test accuracy of this baseline is only " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "81.1\\%" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " on CIFAR-100. By gradually increasing the step size for the augmentation, the training accuracy is decreased as expected, while the test accuracy increases. It demonstrates that the proposed semantic adversarial augmentation strategy can effectively mitigate the over-fitting problem for learning with fixed deep features. By setting an appropriate step size, the test accuracy can be improved to " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "82.9\\%" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": ", which helps shrink the gap to fine-tuning." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "Q. Qian et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": "Effect of Directions for Augmentation Different directions can be adopted for generating augmentation. Besides the proposed semantic adversarial aug-. mentations (SeA), four variants are included in the comparison." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 161, + 479, + 235 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 138, + 161, + 448, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 161, + 448, + 172 + ], + "spans": [ + { + "bbox": [ + 138, + 161, + 448, + 172 + ], + "type": "text", + "content": "- Adv: the original gradient direction without projection as in Eqn. 5." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 173, + 479, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 173, + 479, + 198 + ], + "spans": [ + { + "bbox": [ + 138, + 173, + 479, + 198 + ], + "type": "text", + "content": "- SeA_Neg: similar to the gradient direction in Proposition 4 but getting rid of the direction from the target class and keeping the direction of " + }, + { + "bbox": [ + 138, + 173, + 479, + 198 + ], + "type": "inline_equation", + "content": "\\sum_{c}^{C}p_{c}\\mathbf{w}_{c}" + }, + { + "bbox": [ + 138, + 173, + 479, + 198 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 200, + 479, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 200, + 479, + 222 + ], + "spans": [ + { + "bbox": [ + 137, + 200, + 479, + 222 + ], + "type": "text", + "content": "- Rand: a uniformly random direction within subspace spanned by a minibatch of data, i.e., generating " + }, + { + "bbox": [ + 137, + 200, + 479, + 222 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 137, + 200, + 479, + 222 + ], + "type": "text", + "content": " in Eqn. 6 as a random vector." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 223, + 479, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 223, + 479, + 235 + ], + "spans": [ + { + "bbox": [ + 137, + 223, + 479, + 235 + ], + "type": "text", + "content": "- Manifold mixup: a random semantic direction indicated by an example [32]." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 243, + 487, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 243, + 487, + 267 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 487, + 267 + ], + "type": "text", + "content": "The step size for Adv is very sensitive and is searched in " + }, + { + "bbox": [ + 130, + 243, + 487, + 267 + ], + "type": "inline_equation", + "content": "\\{0.005, 0.01, 0.02, 0.05\\}" + }, + { + "bbox": [ + 130, + 243, + 487, + 267 + ], + "type": "text", + "content": ", while the step size for others is searched in " + }, + { + "bbox": [ + 130, + 243, + 487, + 267 + ], + "type": "inline_equation", + "content": "\\{0.2, 0.4, 0.8, 1.0\\}" + }, + { + "bbox": [ + 130, + 243, + 487, + 267 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 268, + 482, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 268, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 268, + 482, + 483 + ], + "type": "text", + "content": "Fig. 4 shows the comparison on CIFAR-100. First, we can observe that most augmentation directions except the random one can improve the performance over the baseline without augmentation. It confirms that augmentation is important for better generalization even for learning from fixed representations. Then, the method with the random augmentation direction performs similarly to the baseline, which shows that it is challenging to find an effective direction for augmentation in feature space. However, with the original gradient direction directly as in Adv, the training loss will be significantly increased even with a small step size and only a step size that is much smaller than SeA can obtain an applicable model. This is because that the gradient direction only aims to maximize the loss while ignoring the data distribution in the feature space. By projecting the gradient to a data-dependent subspace, SeA and SeA_Neg can achieve a better performance than the baseline without a sophisticated setup for the step size. Moreover, by removing the direction from the target class, the accuracy is degenerated by about " + }, + { + "bbox": [ + 130, + 268, + 482, + 483 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 268, + 482, + 483 + ], + "type": "text", + "content": ". The comparison demonstrates that keeping the completed gradient direction is essential for obtaining appropriate augmentation. Finally, a random semantic direction indicated by examples as in mixup is worse than SeA with a clear margin." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 502, + 482, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 502, + 482, + 561 + ], + "spans": [ + { + "bbox": [ + 130, + 502, + 482, + 561 + ], + "type": "text", + "content": "Comparison with Linear Probing Besides learning with fixed features, linear probing is another way to obtain the linear classifier with a frozen backbone. Compared with our method, it can obtain appropriate augmentations in image space. However, it has to forward each augmentation for learning and the cost is much higher than us." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 563, + 482, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 563, + 482, + 622 + ], + "spans": [ + { + "bbox": [ + 130, + 563, + 482, + 622 + ], + "type": "text", + "content": "To illustrate the effectiveness of the proposed augmentation method, we compare the performance between linear probing and our proposal in Tab. 2. For a fair comparison, all methods optimize features from a single model of MoCo. It shows that the proposed augmentation can outperform the augmentation in the input image space for learning the target classifier." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "content": "Running Time Finally, we compare the running time of fine-tuning, linear probing, and SeA using fixed deep features in Tab. 3. Note that with features" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "text", + "content": "Semantic Adversarial Augmentation for Last Layer Features" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 161, + 158, + 466, + 182 + ], + "blocks": [ + { + "bbox": [ + 140, + 114, + 473, + 148 + ], + "lines": [ + { + "bbox": [ + 140, + 114, + 473, + 148 + ], + "spans": [ + { + "bbox": [ + 140, + 114, + 473, + 148 + ], + "type": "text", + "content": "Table 2: Comparison of accuracy (%) Table 3: Comparison of running time. SeA on CIFAR-100. All methods utilize a utilize deep features from 4 models while single model from MoCo. other methods optimize only a single model." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 161, + 158, + 466, + 182 + ], + "lines": [ + { + "bbox": [ + 161, + 158, + 466, + 182 + ], + "spans": [ + { + "bbox": [ + 161, + 158, + 466, + 182 + ], + "type": "table", + "html": "
baselinelinear probingSeA
77.677.879.0
Methodsfine-tuninglinear probing
Time (s)14,3834,437
", + "image_path": "8f4de505a39518388e7b05f4404e9692c789e153bea9dd487b7b547d71f0b758.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 206, + 482, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 206, + 482, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 206, + 482, + 373 + ], + "type": "text", + "content": "from 4 models, the dimension of the feature vector is 8, 192. However, optimizing the combined features is still much more efficient than fine-tuning a single model or linear probing with 2,048 final features. All methods train the corresponding model by 100 epochs and the running time is measured on a single V100 GPU. Fine-tuning costs 14,383 (sec) for the entire training. In contrast, SeA only costs 40 (sec) to obtain the optimal linear model. The efficiency of SeA using deep features implies its applicability for limited-resource scenarios. With efficient optimization, we can repeat experiments of SeA by multiple times. The standard deviation is only " + }, + { + "bbox": [ + 130, + 206, + 482, + 373 + ], + "type": "inline_equation", + "content": "0.03\\%" + }, + { + "bbox": [ + 130, + 206, + 482, + 373 + ], + "type": "text", + "content": " on CIFAR-100 over 5 runs, which shows the stability of convex optimization. Although linear probing eliminates the cost of the backward pass, that of the forward pass at each iteration is still expensive. Considering that linear probing has a similar running cost as fine-tuning but with worse performance, it will not be included in the comparison over other downstream tasks. More ablation experiments can be found in the supplementary." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 392, + 335, + 404 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 335, + 404 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 335, + 404 + ], + "type": "text", + "content": "4.2 Comparison on Downstream Tasks" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 414, + 481, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 481, + 473 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 481, + 473 + ], + "type": "text", + "content": "Now we evaluate features from different pre-trained models and the ensemble to demonstrate the effectiveness of SeA on downstream tasks. The comparison is summarized in Tab. 4. The linear model with fixed features from different models is denoted by the initial of the corresponding pre-trained model, i.e., S, M, B, and C stand for Supervised, MoCo, BYOL, and CoKe, respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 474, + 481, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 481, + 568 + ], + "type": "text", + "content": "First, the performance of fine-tuning supervised and unsupervised models varies in different domains. For domains closely related to ImageNet (e.g., Birds, Flowers, Pets), the supervised pre-trained model shows better performance than the unsupervised ones, while unsupervised models surpass the supervised counterpart when the domain gap between pre-training and fine-tuning is large. The phenomenon indicates that representations from supervised/unsupervised models capture different patterns and suggests that combining the information from multiple models may handle downstream tasks better." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 481, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 481, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 481, + 640 + ], + "type": "text", + "content": "Second, when having the fixed representation from only a single model with SeA, the performance is already close to fine-tuning on certain data sets, e.g., Caltech, DTD, and SUN. However, the information learned from a single model is biased toward the specific pre-training task. Without fine-tuning, the fixed representation only preserves limited patterns from the source domain, which is insufficient for different target domains." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "To mitigate the limitation of individual pre-training tasks, we consider combining the representations from multiple models, which can be done by a simple" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "Q. Qian et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 146, + 481, + 373 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 4: Comparison of accuracy (%) on downstream tasks. The best performance within each group is underlined while the global one is in bold." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 146, + 481, + 373 + ], + "lines": [ + { + "bbox": [ + 135, + 146, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 135, + 146, + 481, + 373 + ], + "type": "table", + "html": "
MethodsAircraftBirdsCaltechCarsCIFAR10CIFAR100DTDFlowersFoodPetsSUN
Fine-tuning:
Supervised87.382.694.691.997.887.474.797.788.493.663.6
MoCo-v387.581.594.891.798.387.476.396.488.292.265.4
BYOL88.282.094.091.498.087.576.896.888.591.763.4
CoKe89.080.694.492.198.388.378.696.788.591.964.7
Fixed deep features with SeA:
S37.466.592.645.889.871.372.688.167.892.359.2
M61.571.495.469.793.479.076.494.674.990.764.2
B60.066.494.264.291.976.575.895.374.089.761.9
C57.867.094.363.493.078.577.895.676.191.065.3
S+M62.076.396.170.494.480.477.196.377.893.166.3
S+B60.773.795.767.593.479.177.996.377.493.465.5
S+C58.673.795.666.093.580.379.196.278.693.666.9
M+B65.772.595.772.794.680.878.296.277.691.965.5
M+C65.272.295.773.194.681.477.996.578.392.266.6
B+C64.070.195.270.193.980.778.896.478.391.166.4
S+M+B66.276.596.474.294.981.878.696.979.593.367.0
S+M+C65.876.796.373.195.181.978.896.979.993.667.7
S+B+C63.775.496.271.494.482.179.297.079.893.867.7
M+B+C67.273.295.974.795.282.378.996.879.492.067.2
S+M+B+C67.476.896.375.495.282.979.897.280.893.868.2
", + "image_path": "311e31e13434adf83efec17da2d61fe189e6f19b2133a7d12f91bb64e7747baa.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 399, + 480, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 480, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 480, + 483 + ], + "type": "text", + "content": "feature concatenation operator. Compared with features from a single model, the accuracy of all downstream tasks is improved. When having fixed representations from all of 4 models, a linear classifier learned with SeA can outperform standard fine-tuning on 4 tasks (i.e., Caltech, DTD, Pets, and SUN) and achieve comparable performance on 2 tasks (i.e., CIFAR-10 and Flowers). It illustrates that collecting representations from multiple pre-training tasks can obtain complementary patterns that help mitigate the gap to different downstream tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 484, + 481, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 484, + 481, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 484, + 481, + 604 + ], + "type": "text", + "content": "Moreover, it can be observed that \"S+C\" performs better than \"S+M\" on DTD, while it is worse on Birds. The ensemble of all models, i.e., \"S+M+B+C\", shows the best performance on most tasks. Since the models in evaluation are pre-trained with different learning objectives, we find that the learned deep features complement each other. Therefore, more can be better. Note that ensemble can also be applied to fine-tuning. However, the computational cost will significantly increase depending on the number of models, which becomes unaffordable for real applications. In addition, its ensemble can still be worse than deep features. For example, an ensemble of fine-tuning on DTD achieves 78.9% that is about 1% worse than SeA." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": "Although the result of SeA is promising, fixed features perform noticeably worse than fine-tuning on two tasks, e.g., Aircraft and Cars. It is because that the gap between the source domain and the target domain is too large. For example, ImageNet contains limited generic images for cars, while the target task is more fine-grained and consists of cars with different make, models, and" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "text", + "content": "Semantic Adversarial Augmentation for Last Layer Features" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "content": "years. In this case, a larger model pre-trained on a larger data set covering more diverse domains (e.g., JFT-3B [38]) may help improve the performance of SeA." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 156, + 481, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 156, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 156, + 481, + 239 + ], + "type": "text", + "content": "Comparison on Ensemble of Features Then, we illustrate the effectiveness of SeA by comparing it to the baseline without SeA and existing feature-space augmentation methods in Tab. 5, which includes the semantic augmentation method, i.e., ISDA [35] and a robust adversarial augmentation method, i.e., ME-ADA [40]. The experiment is conducted with the ensembled features from 4 models for all methods. The hyper-parameters in ISDA and ME-ADA are searched for the desired result." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 135, + 289, + 481, + 352 + ], + "blocks": [ + { + "bbox": [ + 132, + 257, + 481, + 278 + ], + "lines": [ + { + "bbox": [ + 132, + 257, + 481, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 481, + 278 + ], + "type": "text", + "content": "Table 5: Comparison of SeA and benchmark methods on " + }, + { + "bbox": [ + 132, + 257, + 481, + 278 + ], + "type": "inline_equation", + "content": "\\mathrm{S + M + B + C}" + }, + { + "bbox": [ + 132, + 257, + 481, + 278 + ], + "type": "text", + "content": ". \"Gain\" denotes the improvement over the Baseline." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 289, + 481, + 352 + ], + "lines": [ + { + "bbox": [ + 135, + 289, + 481, + 352 + ], + "spans": [ + { + "bbox": [ + 135, + 289, + 481, + 352 + ], + "type": "table", + "html": "
MethodsAircraftBirdsCaltechCarsCIFAR10CIFAR100DTDFlowersFoodPetsSUNAvg
Baseline63.673.495.469.294.881.179.496.279.593.166.981.1
ISDA63.373.695.969.494.981.379.695.679.693.267.081.2
ME-ADA64.474.995.770.995.081.679.895.880.393.267.481.7
SeA67.476.896.375.495.282.979.897.280.893.868.283.1
Gain3.83.40.96.20.41.80.41.01.30.71.32.0
", + "image_path": "356a98555fdc1a08f7e4e9a6ed6e377c23f85723c849f703f6bb1cce3a992ae6.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 375, + 481, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 375, + 481, + 543 + ], + "spans": [ + { + "bbox": [ + 132, + 375, + 481, + 543 + ], + "type": "text", + "content": "First, compared with the baseline, SeA improves the performance of deep features over all data sets, and the average gain is up to " + }, + { + "bbox": [ + 132, + 375, + 481, + 543 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 132, + 375, + 481, + 543 + ], + "type": "text", + "content": " as denoted by \"Gain\". The consistent improvement shows that augmentation is important for learning with fixed deep features and the proposed augmentation helps mitigate the overfitting problem effectively. Second, while both SeA and ISDA focus on semantic augmentation, ISDA relies on input space augmentations to obtain the semantic directions, while our method leverages features from other examples. Therefore, when input space augmentations are unavailable for deep features, ISDA cannot beat the baseline and it shows that our method is more appropriate for learning without input space augmentations. In addition, ME-ADA improves the vanilla adversarial augmentation by mitigating its sensitivity to the gradient direction. However, the optimization only considers the features from the target example. By incorporating the information from other examples in the same batch, the proposed method can outperform ME-ADA with a margin of " + }, + { + "bbox": [ + 132, + 375, + 481, + 543 + ], + "type": "inline_equation", + "content": "1.4\\%" + }, + { + "bbox": [ + 132, + 375, + 481, + 543 + ], + "type": "text", + "content": " on average." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 559, + 219, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 559, + 219, + 571 + ], + "spans": [ + { + "bbox": [ + 132, + 559, + 219, + 571 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 581, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 581, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 581, + 481, + 665 + ], + "type": "text", + "content": "With the development of unsupervised representation learning, features with diverse information can be obtained by learning with different pretext tasks. Hence, we aim to investigate state-of-the-art performance of pre-trained deep features in this work. By introducing the novel semantic adversarial augmentation, we can show that learning with fixed features can achieve comparable performance to fine-tuning the whole network, but with way less cost. Investigating the efficacy of SeA in pre-training and fine-tuning can be our future work." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "Q. Qian et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 137, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 137, + 481, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 137, + 481, + 180 + ], + "spans": [ + { + "bbox": [ + 138, + 137, + 481, + 180 + ], + "type": "text", + "content": "1. Bossard, L., Guillaumeun, M., Gool, L.V.: Food-101 - mining discriminative components with random forests. In: Fleet, D.J., Pajdla, T., Schiele, B., Tuytelaars, T. (eds.) ECCV. Lecture Notes in Computer Science, vol. 8694, pp. 446-461. Springer (2014)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 181, + 481, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 181, + 481, + 201 + ], + "spans": [ + { + "bbox": [ + 138, + 181, + 481, + 201 + ], + "type": "text", + "content": "2. Boyd, S., Boyd, S.P., Vandenberghe, L.: Convex optimization. Cambridge university press (2004)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 202, + 481, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 202, + 481, + 235 + ], + "spans": [ + { + "bbox": [ + 138, + 202, + 481, + 235 + ], + "type": "text", + "content": "3. Caron, M., Misra, I., Mairal, J., Goyal, P., Bojanowski, P., Joulin, A.: Unsupervised learning of visual features by contrasting cluster assignments. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) NeurIPS (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 235, + 481, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 235, + 481, + 266 + ], + "spans": [ + { + "bbox": [ + 138, + 235, + 481, + 266 + ], + "type": "text", + "content": "4. Chen, T., Cheng, Y., Gan, Z., Wang, J., Wang, L., Liu, J., Wang, Z.: Adversarial feature augmentation and normalization for visual recognition. Trans. Mach. Learn. Res. 2022 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 266, + 481, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 266, + 481, + 299 + ], + "spans": [ + { + "bbox": [ + 138, + 266, + 481, + 299 + ], + "type": "text", + "content": "5. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.E.: A simple framework for contrastive learning of visual representations. In: ICML. Proceedings of Machine Learning Research, vol. 119, pp. 1597-1607. PMLR (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 300, + 481, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 300, + 481, + 319 + ], + "spans": [ + { + "bbox": [ + 138, + 300, + 481, + 319 + ], + "type": "text", + "content": "6. Chen, X., Xie, S., He, K.: An empirical study of training self-supervised vision transformers. In: ICCV. pp. 9620-9629. IEEE (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 320, + 481, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 320, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 320, + 481, + 342 + ], + "type": "text", + "content": "7. Cimpoi, M., Maji, S., Kokkinos, I., Mohamed, S., Vedaldi, A.: Describing textures in the wild. In: CVPR. pp. 3606-3613 (2014)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 342, + 481, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 342, + 481, + 363 + ], + "spans": [ + { + "bbox": [ + 138, + 342, + 481, + 363 + ], + "type": "text", + "content": "8. Crammer, K., Singer, Y.: On the algorithmic implementation of multiclass kernel-based vector machines. J. Mach. Learn. Res. 2, 265-292 (2001)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 363, + 481, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 363, + 481, + 406 + ], + "spans": [ + { + "bbox": [ + 138, + 363, + 481, + 406 + ], + "type": "text", + "content": "9. Donahue, J., Jia, Y., Vinyals, O., Hoffman, J., Zhang, N., Tzeng, E., Darrell, T.: Decaf: A deep convolutional activation feature for generic visual recognition. In: ICML. JMLR Workshop and Conference Proceedings, vol. 32, pp. 647-655. JMLR.org (2014)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 406, + 481, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 406, + 481, + 449 + ], + "spans": [ + { + "bbox": [ + 138, + 406, + 481, + 449 + ], + "type": "text", + "content": "0. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR. OpenReview.net (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 449, + 481, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 449, + 481, + 482 + ], + "spans": [ + { + "bbox": [ + 138, + 449, + 481, + 482 + ], + "type": "text", + "content": "1. Fei-Fei, L., Fergus, R., Perona, P.: Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In: CVPR workshop. pp. 178-178. IEEE (2004)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 482, + 481, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 482, + 481, + 504 + ], + "spans": [ + { + "bbox": [ + 138, + 482, + 481, + 504 + ], + "type": "text", + "content": "2. Goodfellow, I.J., Shlens, J., Szegedy, C.: Explaining and harnessing adversarial examples. In: Bengio, Y., LeCun, Y. (eds.) ICLR (2015)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 504, + 481, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 504, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 138, + 504, + 481, + 557 + ], + "type": "text", + "content": "3. Grill, J., Strub, F., Altché, F., Tallec, C., Richemond, P.H., Buchatskaya, E., Doersch, C., Pires, B.A., Guo, Z., Azar, M.G., Piot, B., Kavukcuoglu, K., Munos, R., Valko, M.: Bootstrap your own latent - A new approach to self-supervised learning. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) NeurIPS (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 558, + 481, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 558, + 481, + 579 + ], + "spans": [ + { + "bbox": [ + 138, + 558, + 481, + 579 + ], + "type": "text", + "content": "4. He, K., Chen, X., Xie, S., Li, Y., Dollar, P., Girshick, R.B.: Masked autoencoders are scalable vision learners. In: CVPR. pp. 15979-15988. IEEE (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 579, + 481, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 579, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 138, + 579, + 481, + 611 + ], + "type": "text", + "content": "5. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.B.: Momentum contrast for unsupervised visual representation learning. In: CVPR. pp. 9726-9735. Computer Vision Foundation / IEEE (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 611, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 481, + 633 + ], + "type": "text", + "content": "6. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778. IEEE Computer Society (2016)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "type": "text", + "content": "7. Ioffe, S., Szegedy, C.: Batch normalization: Accelerating deep network training by reducing internal covariate shift. In: Bach, F.R., Blei, D.M. (eds.) ICML. JMLR Workshop and Conference Proceedings, vol. 37, pp. 448-456. JMLR.org (2015)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "text", + "content": "Semantic Adversarial Augmentation for Last Layer Features" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "18. Kornblith, S., Shlens, J., Le, Q.V.: Do better imagenet models transfer better? In: CVPR. pp. 2661-2671. Computer Vision Foundation / IEEE (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "type": "text", + "content": "19. Krause, J., Stark, M., Deng, J., Fei-Fei, L.: 3d object representations for fine-grained categorization. In: ICCV workshop. pp. 554-561 (2013)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "type": "text", + "content": "20. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "type": "text", + "content": "21. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. In: Bartlett, P.L., Pereira, F.C.N., Burges, C.J.C., Bottou, L., Weinberger, K.Q. (eds.) NeurIPS. pp. 1106-1114 (2012)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "type": "text", + "content": "22. Liu, Z., Xu, Y., Xu, Y., Qian, Q., Li, H., Ji, X., Chan, A.B., Jin, R.: Improved fine-tuning by better leveraging pre-training data. In: NeurIPS (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 238, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 480, + 258 + ], + "type": "text", + "content": "23. Maji, S., Rahtu, E., Kannala, J., Blaschko, M., Vedaldi, A.: Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151 (2013)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "text", + "content": "24. Nilsback, M., Zisserman, A.: Automated flower classification over a large number of classes. In: ICVGIP. pp. 722-729. IEEE Computer Society (2008)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 282, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 480, + 303 + ], + "type": "text", + "content": "25. Parkhi, O.M., Vedaldi, A., Zisserman, A., Jawahar, C.: Cats and dogs. In: CVPR. pp. 3498-3505. IEEE (2012)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "text", + "content": "26. Qian, Q., Hu, J., Li, H.: Hierarchically robust representation learning. In: CVPR. pp. 7334-7342. Computer Vision Foundation / IEEE (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 326, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 346 + ], + "type": "text", + "content": "27. Qian, Q., Jin, R., Zhu, S., Lin, Y.: Fine-grained visual categorization via multi-stage metric learning. In: CVPR. pp. 3716-3724. IEEE Computer Society (2015)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 348, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 348, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 348, + 480, + 369 + ], + "type": "text", + "content": "28. Qian, Q., Shang, L., Sun, B., Hu, J., Li, H., Jin, R.: Softtriple loss: Deep metric learning without triplet sampling. In: ICCV. pp. 6449-6457. IEEE (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 369, + 480, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 390 + ], + "type": "text", + "content": "29. Qian, Q., Xu, Y., Hu, J., Li, H., Jin, R.: Unsupervised visual representation learning by online constrained k-means. In: CVPR. pp. 16619-16628. IEEE (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 392, + 480, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 423 + ], + "type": "text", + "content": "30. Ren, S., He, K., Girshick, R.B., Sun, J.: Faster R-CNN: towards real-time object detection with region proposal networks. IEEE Trans. Pattern Anal. Mach. Intell. 39(6), 1137-1149 (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 425, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 425, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 425, + 480, + 456 + ], + "type": "text", + "content": "31. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M.S., Berg, A.C., Fei-Fei, L.: Imagenet large scale visual recognition challenge. Int. J. Comput. Vis. 115(3), 211-252 (2015)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 458, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 458, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 458, + 480, + 500 + ], + "type": "text", + "content": "32. Verma, V., Lamb, A., Beckham, C., Najafi, A., Mitliagkas, I., Lopez-Paz, D., Bengio, Y.: Manifold mixup: Better representations by interpolating hidden states. In: Chaudhuri, K., Salakhutdinov, R. (eds.) ICML. Proceedings of Machine Learning Research, vol. 97, pp. 6438-6447. PMLR (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 502, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 480, + 533 + ], + "type": "text", + "content": "33. Volpi, R., Morerio, P., Savarese, S., Murino, V.: Adversarial feature augmentation for unsupervised domain adaptation. In: CVPR. pp. 5495-5504. Computer Vision Foundation / IEEE Computer Society (2018)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "type": "text", + "content": "34. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The caltech-ucsd birds-200-2011 dataset (2011)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 555, + 480, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 480, + 587 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 480, + 587 + ], + "type": "text", + "content": "35. Wang, Y., Huang, G., Song, S., Pan, X., Xia, Y., Wu, C.: Regularizing deep networks with semantic data augmentation. IEEE Trans. Pattern Anal. Mach. Intell. 44(7), 3733-3748 (2022)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 589, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 480, + 621 + ], + "type": "text", + "content": "36. Xiao, J., Hays, J., Ehinger, K.A., Oliva, A., Torralba, A.: SUN database: Large-scale scene recognition from abbey to zoo. In: CVPR. pp. 3485-3492. IEEE Computer Society (2010)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "text", + "content": "37. Xu, Y., Qian, Q., Li, H., Jin, R., Hu, J.: Weakly supervised representation learning with coarse labels. In: ICCV. pp. 10573-10581. IEEE (2021)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "38. Zhai, X., Kolesnikov, A., Houlsby, N., Beyer, L.: Scaling vision transformers. In: CVPR. pp. 1204-1213. IEEE (2022)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "Q. Qian et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 172 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "39. Zhang, H., Cissé, M., Dauphin, Y.N., Lopez-Paz, D.: mixup: Beyond empirical risk minimization. In: ICLR. OpenReview.net (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 480, + 172 + ], + "type": "text", + "content": "40. Zhao, L., Liu, T., Peng, X., Metaxas, D.N.: Maximum-entropy adversarial data augmentation for improved generalization and robustness. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) NeurIPS (2020)" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "text", + "content": "Semantic Adversarial Augmentation for Last Layer Features" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/SeFlow_ A Self-Supervised Scene Flow Method in Autonomous Driving/e7db8302-f2d3-4bc6-8b61-d5de1eb958ee_content_list.json b/2024/SeFlow_ A Self-Supervised Scene Flow Method in Autonomous Driving/e7db8302-f2d3-4bc6-8b61-d5de1eb958ee_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4fa7d0753900349857887241b1b0033c708fe035 --- /dev/null +++ b/2024/SeFlow_ A Self-Supervised Scene Flow Method in Autonomous Driving/e7db8302-f2d3-4bc6-8b61-d5de1eb958ee_content_list.json @@ -0,0 +1,1790 @@ +[ + { + "type": "text", + "text": "SeFlow: A Self-Supervised Scene Flow Method in Autonomous Driving", + "text_level": 1, + "bbox": [ + 282, + 140, + 722, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qingwen Zhang $^{1}$ , Yi Yang $^{1,2}$ , Peizheng Li $^{3,4}$ , Olov Andersson $^{1}$ , and Patric Jensfelt $^{1}$", + "bbox": [ + 321, + 210, + 679, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ RPL, KTH Royal Institute of Technology, Stockholm, Sweden \n $^{2}$ Scania CV AB, Södertälje, Sweden", + "bbox": [ + 287, + 253, + 715, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 University of Tübingen, Tübingen, Germany", + "bbox": [ + 344, + 282, + 656, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 Mercedes-Benz AG, Sindelfingen, Germany {qingwen,yiya,olovand,patric}@kth.se,", + "bbox": [ + 348, + 296, + 648, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "carol-yi.yang@scania.com, peizheng.li@mercedes-benz.com", + "bbox": [ + 292, + 325, + 710, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Scene flow estimation predicts the 3D motion at each point in successive LiDAR scans. This detailed, point-level, information can help autonomous vehicles to accurately predict and understand dynamic changes in their surroundings. Current state-of-the-art methods require annotated data to train scene flow networks and the expense of labeling inherently limits their scalability. Self-supervised approaches can overcome the above limitations, yet face two principal challenges that hinder optimal performance: point distribution imbalance and disregard for object-level motion constraints. In this paper, we propose SeFlow, a self-supervised method that integrates efficient dynamic classification into a learning-based scene flow pipeline. We demonstrate that classifying static and dynamic points helps design targeted objective functions for different motion patterns. We also emphasize the importance of internal cluster consistency and correct object point association to refine the scene flow estimation, in particular on object details. Our real-time capable method achieves state-of-the-art performance on the self-supervised scene flow task on Argoverse 2 and Waymo datasets. The code is open-sourced at https://github.com/KTH-RPL/SeFlow.", + "bbox": [ + 259, + 373, + 743, + 623 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: 3D scene flow, self-supervised, autonomous driving, large-scale point cloud", + "bbox": [ + 259, + 635, + 740, + 664 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 688, + 375, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Scene flow [30] captures the 3D velocity at every point in a point cloud. These detailed 3D flow estimates can enhance downstream tasks in autonomous driving, such as detection, segmentation, tracking, and occupancy flow estimation [20]. A common paradigm for addressing the scene flow problem is supervised learning by utilizing annotated LiDAR data [11, 38]. However, expensive labeling inherently limits the scalability of supervised learning methods.", + "bbox": [ + 212, + 719, + 787, + 809 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Given the difficulty and expense of labeling scene flow ground truth, we instead focus on self-supervised scene flow approaches. Existing self-supervised", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b734f2cbf7a5cfbf504f0c2e94d336bb71f4313459c39f351b024c3efa560eee.jpg", + "image_caption": [ + "Fig. 1: LiDAR scene flow estimation using our SeFlow method on Argoverse 2. The predicted scene flow for each point is color-coded based on direction. The white indicates static points whose flow is zero. More saturated colors indicate higher velocities. (a) Camera view for visualization purposes only. (b),(c) are zoomed-in views showing the baseline from ZeroFlow [29] as well as SeFlow (ours). When predicting the flow of a large and long vehicle, the baseline predicts a portion of the flow as 0, whereas our estimates are consistent. In addition, the baseline tends to ignore small-scale objects, e.g., pedestrians, while our method can better handle such small and slow-moving objects." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 782, + 356 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "methods can be divided into two categories: knowledge distillation and data exploration. Knowledge distillation methods [2,25,29] typically rely on a 'teacher' method to generate pseudo flow labels. These pseudo labels are then used to supervise student models. Data exploration methods [3, 14, 15, 19, 31], on the other hand, directly utilize the predicted flow to project the input points and find nearest neighbors in the next frame to establish constraints for training.", + "bbox": [ + 212, + 522, + 785, + 612 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A key challenge for self-supervised methods is that the vast majority of the points are static (about $86\\%$ of points are background [7, 34]). This data imbalance often leads to overly conservative scene flow predictions. One common way to counter this is to use large amounts of training data as in ZeroFlow [29]. ZeroFlow experimentally shows that more data helps to improve the dynamic flow accuracy. In this paper, we investigate ways to improve data efficiency and take inspiration from supervised methods [11, 38]. We address the data imbalance by first classifying points as dynamic or static based on traditional ray casting when integrating frames [9]. This allows us to constrain the corresponding motion patterns by proposing novel loss functions.", + "bbox": [ + 212, + 613, + 787, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another shortcoming of existing self-supervised scene flow methods is that they disregard object-level motion constraints. The flow should be consistent, i.e., all points in a rigid object should have similar flows. A clear case of inconsistent flow can be seen on the big truck in Fig. 1, where ZeroFlow [29] predicts an absence of flow in certain sections. This inconsistency is caused by incorrect", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Q Zhang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "object point associations which also can be observed in small-scale objects, as shown in the case of the pedestrian. To account for object-level motion constraints, we propose to cluster the dynamic points into object candidates and encourage consistent flow and correct associations for the points in each cluster. This reduces the fragmentation of the flow estimate for points inside the same object.", + "bbox": [ + 212, + 146, + 782, + 234 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overall, our method improves the estimated flow accuracy and addresses issues in previous methods as shown in Fig. 1(c). We propose an efficient and effective self-supervised scene flow method that integrates traditional classification and learning-based strategies. Our approach is open-source at https://github.com/KTH-RPL/SeFlow. Our primary contributions include:", + "bbox": [ + 212, + 237, + 782, + 311 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose SeFlow, a novel method that integrates a dynamic classification method in formulating efficient self-supervision objectives.", + "- We further construct loss functions to learn dynamic flow estimation in imbalanced data and ensure consistent object-level flow, mitigating the effects of correspondence errors.", + "- We show that SeFlow achieves state-of-the-art results on the self-supervised scene flow task in Argoverse 2 and Waymo datasets and even outperforms all but one supervised method on the leaderboard." + ], + "bbox": [ + 225, + 323, + 784, + 444 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 469, + 385, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we explore existing works in scene flow estimation. We also discuss current traditional frameworks capable of classifying dynamic points, highlighting their relevance and application in the context of scene flow tasks.", + "bbox": [ + 212, + 502, + 782, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Scene Flow Estimation", + "text_level": 1, + "bbox": [ + 215, + 571, + 450, + 585 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Scene flow estimation in autonomous driving is slightly different from flow estimation in object registration. Methods in object registration [8, 13, 16, 24, 32, 33] focus on relatively small-scale point cloud data like synthetic datasets Shapenet [5] and FlyingThing3D [17]. These methods scale poorly with the number of points. When applied to point cloud data for autonomous driving, they require downsampling to 8,192 points or less [11, 18]. In recent datasets like Argoverse 2 [34] and Waymo [11], the number of points in one full frame is around $80\\mathrm{k - }177\\mathrm{k}$ . Methods that can handle such large-scale point cloud data as input usually use a voxel-based pipeline [11, 38]. However, expensive labeling of ground truth flow limits the scalability of these supervised methods, especially in autonomous driving where we have continuously increasing amounts of potential training data.", + "bbox": [ + 212, + 598, + 784, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To train models without labeled data, recent methods propose self-supervised losses. Many commonly used losses, such as Chamfer distance, are based on the nearest neighbor distance between two point cloud inputs [3,12,14,15,19,27,36]. However, a major limitation of nearest neighbor based losses is that only part", + "bbox": [ + 212, + 779, + 784, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "SeFlow: A Self-Supervised Scene Flow Method in Autonomous Driving", + "bbox": [ + 258, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "of the points on dynamic objects provides good supervision due to incorrect associations. This is especially apparent when big trucks are moving fast in autonomous driving scenarios (see Fig. 1 and Fig. 3). Mittal et al. [19] define a self-supervised loss by tracking a patch forward and backward in time to form a cycle while penalizing the errors through cycle consistency and feature similarity. Baur et al. [3] propose three loss functions with k-NN loss, rigid cycle consistency inspired by Mittal, and artificial labels based on the distance between points in the original and estimated point clouds. In the following we describe NSFP [14], FastNSF [15], and ZeroFlow [29] in more detail. These are the publically available methods that perform best on the Argoverse 2 self-supervised scene flow challenge and are therefore used as our baselines.", + "bbox": [ + 212, + 146, + 787, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "NSFP [14] is an optimization-based method. For each pair of consecutive input point clouds, NSFP iteratively learns new weights for an MLP network to predict the flow using the Chamfer distance loss. However, thousands of iterations are needed and their runtimes extend from 26 to 35 seconds per frame, which fails to meet the real-time requirements of autonomous driving. FastNSF [15] improves the efficiency by voxelizing the point cloud first and then converting it to a distance transform map for faster neighbor calculation. This reduces the runtime to 0.5 seconds, at the expense of increased error.", + "bbox": [ + 212, + 311, + 787, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ZeroFlow [29] adopts a semi-supervised strategy. Pseudo-flow labels are created offline using NSFP [14], and a FastFlow3D [11] model is used as a student for knowledge distillation. This setup allows the student model to perform real-time inference. In this case, the teacher network needs significant resources (around 3.6 GPU months [29]) to label the entire training data. The final accuracy is also influenced by the performance of the teacher.", + "bbox": [ + 212, + 431, + 787, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To increase data efficiency and address the limitations of the above methods, we propose to integrate efficient ray-casting-based dynamic awareness mapping into our pipeline. The core idea is to classify which points move and cluster these points into objects for which we can estimate group-level motion statistics and define better self-supervised loss functions.", + "bbox": [ + 212, + 523, + 787, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Dynamic Awareness in Mapping", + "text_level": 1, + "bbox": [ + 214, + 619, + 529, + 636 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the field of Simultaneous Localization and Mapping (SLAM), there is a growing interest in dynamic awareness [9,22,23,35,37]. This interest stems from the fact that points associated with moving objects in a scene can significantly impact localization and planning performance. Existing dynamic awareness methods in mapping [9,23,37] often utilize ray casting techniques to binary classify points as either dynamic or static.", + "bbox": [ + 212, + 643, + 787, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The scene flow task, on the other hand, aims to predict the specific 3D flow at each point, a goal that extends beyond the mere categorization of dynamic and static points. In mapping, a point is considered dynamic if it moves once within a scene (even if it becomes static later), whereas, in scene flow tasks, a point is deemed dynamic if it moves faster than a certain velocity threshold in the current frame. Despite these differences, insights from the mapping field are invaluable. By integrating information over time, these frameworks develop a", + "bbox": [ + 212, + 734, + 787, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Q Zhang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "more comprehensive understanding of the entire scene. Such scene level dynamic awareness can inform and enhance our exploration of point cloud data in scene flow tasks.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Problem Statement", + "text_level": 1, + "bbox": [ + 215, + 215, + 439, + 229 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This work addresses the problem of real-time scene flow estimation in autonomous driving. Given two consecutive input point clouds, $\\mathcal{P}_t$ and $\\mathcal{P}_{t + 1}$ , along with the ego movement $\\mathbf{T}_{t,t + 1} \\in SE(3)$ , the goal is to predict the motion vector as flow $\\hat{\\mathcal{F}}_{t,t + 1}(p) = (x,y,z)^T$ for each point $p \\in \\mathcal{P}_t$ .", + "bbox": [ + 212, + 246, + 785, + 309 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The objective is to minimize the End Point Error (EPE) which represents the difference between the predicted flow and the ground truth flow, as expressed by the following equation:", + "bbox": [ + 212, + 309, + 785, + 354 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min \\underbrace {\\frac {1}{| \\mathcal {P} _ {t} |} \\sum_ {p \\in \\mathcal {P} _ {t}} \\left\\| \\hat {\\mathcal {F}} (p) - \\mathcal {F} _ {g t} (p) \\right\\| _ {2}} _ {\\text {E P E}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 364, + 785, + 421 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $|\\mathcal{P}_t|$ denotes the cardinality of (i.e., the number of points in) $\\mathcal{P}_t$ . For consistency, in the subsequent presentation, capital symbols correspond to sets, while the lowercase symbols represent variables of specific points.", + "bbox": [ + 212, + 431, + 784, + 478 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 Method", + "text_level": 1, + "bbox": [ + 215, + 500, + 330, + 515 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1 Input and Output", + "text_level": 1, + "bbox": [ + 215, + 532, + 410, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The first step to process the input data, as is commonly done, is to remove ground points from $\\mathcal{P}_t$ and $\\mathcal{P}_{t + 1}$ . This is typically done using HD maps [26,34] or ground segmentation techniques [10].", + "bbox": [ + 212, + 556, + 782, + 602 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The estimated flow $\\tilde{\\mathcal{F}}$ from $\\mathcal{P}_t$ to $\\mathcal{P}_{t + 1}$ can be decomposed into two parts as follows:", + "bbox": [ + 212, + 602, + 782, + 630 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {F}} = \\mathcal {F} _ {e g o} + \\Delta \\hat {\\mathcal {F}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 441, + 631, + 784, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{F}_{ego}$ is the flow resulting from the ego vehicle's motion which can be directly obtained from $\\mathbf{T}_{t,t + 1}$ , and $\\varDelta\\hat{\\mathcal{F}}$ is our network output.", + "bbox": [ + 212, + 654, + 782, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2 Model Backbone", + "text_level": 1, + "bbox": [ + 215, + 709, + 400, + 723 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To enable real-time computation and estimation of scene flow across a large point set, voxelization is considered a practical encoding strategy in the model backbone. However, the reduction in resolution often leads to a poorer distinction of point features within the same voxel. With this in mind, we use DeFlow [38] as the architectural basis in this paper. DeFlow integrates GRU [6] with iterative refinement in the decoder design. More specifically, the GRU module maintains voxel features as its hidden state and selectively forgets or updates", + "bbox": [ + 212, + 734, + 784, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "SeFlow: A Self-Supervised Scene Flow Method in Autonomous Driving", + "bbox": [ + 258, + 114, + 732, + 130 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4b64c4f073be133b622aa2a3dc9f8f5b242e5240fa56a253c64412b99e17f8c6.jpg", + "image_caption": [ + "Training Stage Sec. 4.4 Self-supervised Loss (including Sec. 4.3 Dynamic Classification)" + ], + "image_footnote": [], + "bbox": [ + 225, + 146, + 776, + 223 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9175ccddfa2f4ff0fb904a45fc31a3f448aae906454462b96f0708da665eea7b.jpg", + "image_caption": [ + "Fig. 2: SeFlow Architecture. Top: With two consecutive point clouds as inputs, our model predicts the estimated flows of all points. Bottom: Conceptual visualization of the Chamfer loss and the three proposed training losses. With the original input $\\mathcal{P}_t$ (2 static points for the building, and 2 dynamic points for the car) plus the estimated flow $\\hat{\\mathcal{F}}_t$ , we can calculate the error between estimated $\\hat{\\mathcal{P}}_{t+1}$ and the next frame point cloud $\\mathcal{P}_{t+1}$ ( $\\mathcal{L}_{\\mathrm{cham}}$ ). The second part is $\\mathcal{L}_{\\mathrm{dcham}}$ that only calculates the distance error between dynamic points. The third loss says that the estimated flows of static points should be zero. Finally, we assume that the flow at points from the same cluster should be consistent, and mitigate underestimation by using the proposed upper bound on the flow." + ], + "image_footnote": [], + "bbox": [ + 228, + 253, + 772, + 385 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the information of the hidden state during each iteration according to the input point features. After multiple iterations, the optimized voxel features are concatenated with the original point features to obtain the final individual point features. Benefiting from this design, we improve the inference efficiency compared to the commonly used backbone FastFlow3D [11] without sacrificing the accuracy of scene flow estimation in coarse resolution settings.", + "bbox": [ + 212, + 568, + 787, + 659 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3 Dynamic Classification of Points", + "text_level": 1, + "bbox": [ + 214, + 686, + 529, + 702 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To facilitate dynamic classification of points during the training stage, we incorporate the DUFOMap framework, a mapping-based dynamic awareness approach [9]. The key insight of this is that points observed inside a region that at one time has been observed as empty must be dynamic. Built on this insight, DUFOMap utilizes ray-casting to classify dynamic points at the sensor rate on the CPU. The result is two disjoint sets, $\\mathcal{P}_{t,d}$ and $\\mathcal{P}_{t,s}$ , where $\\mathcal{P}_{t,d}$ is the set of dynamic points that have moved once inside a scene (even if they later become static) and $\\mathcal{P}_{t,s}$ is the set of static points that did not move at any time.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Q Zhang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Note that the dynamic-static classification is separated from the inference and training, offering our method good flexibility.", + "bbox": [ + 215, + 146, + 782, + 176 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4 Self-supervised Loss", + "text_level": 1, + "bbox": [ + 215, + 195, + 426, + 210 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As discussed, self-supervised learning with only Chamfer distance loss is vulnerable to problems with data imbalance and incorrect associations. We therefore construct three additional loss functions, illustrated in Fig. 2 and described in turn below, to mitigate problems with data imbalance as well as encourage consistent object-level flow.", + "bbox": [ + 212, + 215, + 782, + 292 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Chamfer Distance The Chamfer distance, a common self-supervised loss in existing work [3, 14, 15, 19], has the following definition:", + "bbox": [ + 214, + 310, + 782, + 340 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c h a m}} = \\frac {1}{| \\hat {\\mathcal {P}} _ {t + 1} |} \\sum_ {p \\in \\hat {\\mathcal {P}} _ {t + 1}} \\mathrm {S} (p, \\mathcal {P} _ {t + 1}) + \\frac {1}{| \\mathcal {P} _ {t + 1} |} \\sum_ {p \\in \\mathcal {P} _ {t + 1}} \\mathrm {S} (p, \\hat {\\mathcal {P}} _ {t + 1}) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 300, + 347, + 784, + 385 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {S} (p, \\mathcal {P} _ {o}) = \\min _ {p _ {i} \\in \\mathcal {P} _ {o}} \\| p - p _ {i} \\| _ {2} ^ {2}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 285, + 390, + 784, + 412 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where, $\\hat{\\mathcal{P}}_{t + 1} = \\mathcal{P}_t + \\hat{\\mathcal{F}}_t$ , $\\mathrm{S}(\\cdot) = \\mathrm{D}(\\cdot)^2$ , $\\mathrm{D}(p,\\mathcal{P}_o)$ denotes the distance between point $p$ and its nearest neighbor in $\\mathcal{P}_o$ .", + "bbox": [ + 212, + 420, + 782, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The Chamfer distance is proposed to calculate a similarity between two point clouds. However, when using it directly as a supervised signal for scene flow in autonomous driving, there are two issues we note. First, the number of static points is often much larger than that of dynamic points and averaging $S(\\cdot)$ over all points leads the training to favor static points, i.e., zero flow estimation. Second, due to erroneous correspondence assumptions, only part of the flows within a dynamic object can be estimated correctly as shown in Fig. 3, where the estimated flows of points in the overlap area are zero.", + "bbox": [ + 212, + 452, + 784, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To solve these issues, we contribute the following constraints using the dynamic classification information prior to supervising the network.", + "bbox": [ + 214, + 573, + 782, + 603 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Dynamic Chamfer Distance The imbalance in the number of dynamic and static points presents a significant challenge in scene flow estimation, as highlighted in previous works [11,29,38]. Supervised methods, provided with ground truth labels for point velocity or classification, can weight their loss functions to account for this imbalance. For self-supervised learning, we introduce a dynamic Chamfer distance. In this context, 'dynamic' points are identified based on the output from Sec. 4.3. Specifically, this loss, $\\mathcal{L}_{\\mathrm{dcham}}$ , is only computed on points classified as dynamic in two consecutive point clouds. By exclusively considering dynamic points, this loss captures the nuances of motion in point cloud data. This is defined as:", + "bbox": [ + 214, + 621, + 784, + 771 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {d c h a m}} = \\frac {1}{| \\hat {\\mathcal {P}} _ {t + 1 , d} |} \\sum_ {p \\in \\hat {\\mathcal {P}} _ {t + 1, d}} \\mathrm {S} (p, \\mathcal {P} _ {t + 1, d}) + \\frac {1}{| \\mathcal {P} _ {t + 1 , d} |} \\sum_ {p \\in \\mathcal {P} _ {t + 1, d}} \\mathrm {S} (p, \\hat {\\mathcal {P}} _ {t + 1, d}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 245, + 776, + 784, + 815 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\hat{\\mathcal{P}}_{t + 1,d} = \\mathcal{P}_{t,d} + \\hat{\\mathcal{F}}_{t,d}$", + "bbox": [ + 215, + 823, + 413, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "SeFlow: A Self-Supervised Scene Flow Method in Autonomous Driving", + "bbox": [ + 258, + 114, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/afbec55c88188da0e4b650e4654f135a33a82413da23f76200ca23077c41c692.jpg", + "image_caption": [ + "(a) Chamfer distance" + ], + "image_footnote": [], + "bbox": [ + 225, + 141, + 517, + 243 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a08235b6b1185d0e2165915e5d62cf503a554d62f64af9db437ddd1c8914492a.jpg", + "image_caption": [ + "(b) Point flow magnitude", + "Fig. 3: Simple visualization of the shortcomings of using Chamfer distance as a supervisory signal for flow value estimation. The denser color on points in (b) represents higher flow values and white means the point's flow is zero. (a) illustrates how to calculate loss based on Chamfer distance, and (b) shows that the flow results, based on the nearest neighbor principle, can lead to zero flow estimation for the middle of the object." + ], + "image_footnote": [], + "bbox": [ + 527, + 142, + 769, + 247 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Static Flow The standard Chamfer distance rests on a very strong assumption: nearest neighbor matching can find perfect one-to-one correspondences of all points between two frames. However, due to varying observations, the number of points in the two frames differs, leading to inconsistencies and potential mismatches. To deal with the possible mismatches in the static areas, we add a constraint to encourage the model to estimate zero flow for static points. This loss is defined as follows:", + "bbox": [ + 212, + 385, + 787, + 491 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s t a t i c}} = \\frac {1}{| \\mathcal {P} _ {t , s} |} \\sum_ {p \\in \\mathcal {P} _ {t, s}} \\| \\Delta \\hat {\\mathcal {F}} (p) \\| _ {2} ^ {2}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 497, + 785, + 535 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\varDelta\\hat{\\mathcal{F}}(p)$ represents the network output flow of point $p$ .", + "bbox": [ + 215, + 545, + 650, + 564 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Dynamic Cluster Flow For dynamic fields, the inconsistencies and correspondence errors are even more complicated. For instance, we observe that nearest neighbor matching often leads to erroneous results, e.g., on large objects undergoing translation, which is very common for moving vehicles in urban street environments as shown in Fig. 1. More specifically, Figure 3 illustrates the issue on a simplified car example. In this case, the flow is parallel to the object's surface, which means that nearest neighbor matching does not provide good supervision in the interior of the surface ( $\\mathrm{D} \\approx 0$ for the points in the overlap area). This mismatch results in an incorrect local optimum that remains unresolved during the training process as it minimizes the losses based on Chamfer Distance.", + "bbox": [ + 212, + 582, + 787, + 747 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Therefore, we suggest that the motion of all parts within an object (a cluster) should be approximately homogeneous over a short time interval. We use the HDBSCAN [4,21] clustering algorithm to identify different moving objects. This clustering is only applied to dynamic points, thus reducing the computational overhead.", + "bbox": [ + 212, + 750, + 787, + 823 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {C} _ {t, d} = \\operatorname {H D B S C A N} \\left(\\mathcal {P} _ {t, d}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 408, + 824, + 785, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Q Zhang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Directly constraining the variance of the flow distribution inside a cluster would be a straightforward idea, but does not guarantee the correctness of the obtained mean value after convergence. As we know that nearest-neighbour point correspondence can considerably underestimate flow on geometrically featureless object surfaces, we here instead propose to use an upper bound on the object motion as the supervisory signal. We derive this upper bound, per object cluster, by taking the maximum inter-frame distance of all point correspondences within the cluster. Specifically, we exploit information from the original input point cloud data, i.e., $\\mathcal{P}_t$ and $\\mathcal{P}_{t + 1}$ . After the dynamic classification and clustering process, we find the index of the point $p_k$ in cluster $c_{i} \\in \\mathcal{C}_{t,d}$ with the largest distance to its nearest neighbor point in $\\mathcal{P}_{t + 1,d}$ , i.e.,", + "bbox": [ + 212, + 146, + 787, + 313 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\kappa_ {i} = \\underset {k} {\\arg \\max } \\left\\{\\mathrm {D} \\left(p _ {k}, \\mathcal {P} _ {t + 1, d}\\right) \\mid p _ {k} \\in \\mathcal {P} _ {c _ {i}} \\right\\}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 366, + 325, + 785, + 348 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We calculate the upper bound, $\\tilde{f}_{c_i}$ on the flow for cluster $c_{i}$ as", + "bbox": [ + 214, + 361, + 666, + 378 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {f} _ {c _ {i}} = p _ {\\kappa_ {i}} ^ {\\prime} - p _ {\\kappa_ {i}}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 446, + 388, + 784, + 407 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $p_{\\kappa_i}'$ is the nearest neighbor of point $p_{\\kappa_i}$ in $\\mathcal{P}_{t+1,d}$ . We use this to drive the estimated flows of cluster $c_i$ towards $\\tilde{f}_{c_i}$ as follows:", + "bbox": [ + 214, + 417, + 785, + 450 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c _ {i}} = \\sum_ {p _ {j} \\in \\mathcal {P} _ {c _ {i}}} | | \\hat {f} _ {p _ {j}} - \\tilde {f} _ {c _ {i}} | | _ {2} ^ {2}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 416, + 464, + 784, + 496 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {d c l s}} = \\frac {1}{| \\mathcal {P} _ {t , d} |} \\sum_ {c _ {i} \\in \\mathcal {C} _ {t, d}} \\mathcal {L} _ {c _ {i}}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 501, + 784, + 536 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In conclusion, the final SeFlow loss incorporates all four losses introduced above,", + "bbox": [ + 212, + 547, + 782, + 577 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {c h a m}} + \\mathcal {L} _ {\\text {s t a t i c}} + \\mathcal {L} _ {\\text {d c h a m}} + \\mathcal {L} _ {\\text {d c l s}}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 354, + 579, + 784, + 594 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The effect of each loss will be evaluated in the ablation study in Sec. 5.3.", + "bbox": [ + 214, + 601, + 740, + 616 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 Experiment", + "text_level": 1, + "bbox": [ + 214, + 638, + 366, + 656 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this section, we first outline evaluation details and then present quantitative comparisons with state-of-the-art methods on two benchmark datasets. A series of ablation studies are presented to better evaluate the individual components of our approach. Finally, we conclude with qualitative results on Argoverse 2 and discuss limitations of the approach.", + "bbox": [ + 212, + 671, + 787, + 748 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.1 Dataset and Metric", + "text_level": 1, + "bbox": [ + 214, + 768, + 423, + 784 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We briefly introduce the dataset and metrics for evaluation in the following section. More implementation details, such as hyperparameters for training and dataset descriptions, can be found in the supplementary material.", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "SeFlow: A Self-Supervised Scene Flow Method in Autonomous Driving", + "bbox": [ + 258, + 114, + 732, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/2628bf735d7f6e7d5996efed1edb5475ea5707794ee42fa3d0211c5e7a309d42.jpg", + "table_caption": [ + "Table 1: Comparisons on Argoverse 2 test set from the online leaderboard [1]. Upper groups are supervised methods, lower groups are self-supervised methods. Our method achieves state-of-art performance in the self-supervised scene flow task. $\\dagger$ means methods can run in real-time (10 Hz) onboard." + ], + "table_footnote": [], + "table_body": "
MethodRun Time per frame [ms]EPE ↓
3-wayFDFSBS
FastFlow3D†[11]34 ± 50.07820.20720.02530.0020
DeFlow†[38]48 ± 40.05340.13400.02320.0029
FastNSF [15]507 ± 3120.16570.35400.04060.1025
NSFP [14]32,060 ± 10,1120.06850.15030.03020.0248
ZeroFlow†[29]34 ± 50.08140.21090.02540.0080
SeFlow (Ours)†48 ± 40.06280.15250.03210.0038
", + "bbox": [ + 256, + 212, + 741, + 369 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Dataset We evaluate our approach on two large-scale autonomous driving datasets: Argoverse 2 [34] and Waymo [26]. Ground removal is performed using HD maps for both datasets as described in [34]. Waymo datasets contain 798 training and 202 validation scenes respectively. We focus our description here on Argoverse 2, which provides official and public scene flow challenges [1,28], with the Sensor and LiDAR datasets. The Sensor dataset encompasses 700 training and 150 validation scenes. Each scene is approximately 15 seconds long in $10\\mathrm{Hz}$ with complete annotations for evaluation. The evaluation for another 150 test scenes can be accessed indirectly by submitting a solution to the leaderboard. The LiDAR dataset contains 20,000 scenes without any annotation and is only used as extra data in Sec. 5.3.", + "bbox": [ + 215, + 395, + 787, + 561 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Metric The benchmark follows existing works [1,7,29,38] and uses the three-way End Point Error. End Point Error (EPE), as defined in Eq. (1), measures the L2 norm of the discrepancy between the predicted and actual flow vectors, expressed in meters. The EPE three-way (3-way) is defined as the unweighted average EPE across three disjoint sets of points: Foreground Dynamic (FD), Foreground Static (FS), and Background Static (BS). The definition of 'dynamic' is as follows: If the flow at a point exceeds the threshold (the public leaderboard setting is $0.05\\mathrm{m}$ ), the point is defined as dynamic. Given the $10\\mathrm{Hz}$ sensor frequency, this threshold corresponds to a speed of $0.5\\mathrm{m/s}$ . All evaluations are limited to points that are within a $100\\mathrm{m} \\times 100\\mathrm{m}$ box centered on the ego vehicle.", + "bbox": [ + 215, + 582, + 787, + 734 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.2 Quantitative Results", + "text_level": 1, + "bbox": [ + 217, + 753, + 431, + 770 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We evaluate the performance of our method SeFlow and compare it with the currently best-performing methods on the Argoverse 2 test set and Waymo validation set. In Tab. 1 and Tab. 2, the upper group, FastFlow3D and DeFlow, are supervised methods trained with ground truth flow. Compared to the supervised", + "bbox": [ + 215, + 777, + 785, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Q Zhang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/03e0445014b8ec2137144fd9371dfdeca109786ca81c3a44b406aa5229f579cb.jpg", + "table_caption": [ + "Table 2: Comparisons on Waymo Open dataset validation set. Upper groups are supervised methods, lower groups are self-supervised methods. Our method achieves state-of-art performance in the self-supervised scene flow task. $\\dagger$ means methods can run in real-time (10 Hz) onboard." + ], + "table_footnote": [], + "table_body": "
MethodRun Time per frame [ms]EPE ↓
3-wayFDFSBS
FastFlow3D†[11]27 ± 60.07820.19540.02460.0152
DeFlow†[38]42 ± 40.04460.09800.02590.0098
FastNSF [14]593 ± 3080.15790.30120.01460.0403
NSFP [15]76,163 ± 32,2560.10050.17120.10810.0221
ZeroFlow†[29]27 ± 60.09210.21620.01530.0241
SeFlow (Ours)†42 ± 40.05980.15060.01810.0106
", + "bbox": [ + 256, + 212, + 741, + 369 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "methods, we note that SeFlow outperforms FastFlow3D and approaches the level of DeFlow in terms of EPE 3-way. This comparative analysis demonstrates the great potential of self-supervised learning in scene flow estimation and validates the effectiveness of our approach.", + "bbox": [ + 212, + 397, + 784, + 457 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In the self-supervised category, our SeFlow method achieves state-of-the-art performance on both datasets. While FastNSF and NSFP are optimization-based methods that do not rely on pre-trained weights for subsequent estimations, their inference times are not conducive to real-time requirements. NSFP, with the second best result in Tab. 1, takes approximately 30 seconds to predict a single frame, which is impractical for real-time autonomous driving applications. FastNSF, on the other hand, enhances efficiency through voxelization for quicker Chamfer distance calculation. Although significantly faster than NSFP, the performance of FastNSFP is the worst among all methods evaluated.", + "bbox": [ + 212, + 458, + 787, + 593 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "ZeroFlow is capable of estimating scene flow in real-time, with an accuracy similar to NSFP (slightly better on Waymo and slightly worse on Argoverse 2). SeFlow stands out not only for achieving the best result, but also for drastically reducing the run time to 50 milliseconds, which is more than two orders of magnitude faster than NSFP while providing more accurate flow estimates. SeFlow's state-of-the-art performance demonstrated in Tab. 1 and Tab. 2 underscores the effectiveness of our novel self-supervised method in scene flow estimation.", + "bbox": [ + 212, + 594, + 787, + 700 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.3 Ablation study", + "text_level": 1, + "bbox": [ + 214, + 723, + 387, + 739 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In this section, we delve into two key aspects of our SeFlow pipeline. Firstly, we examine the impact of different loss terms on the accuracy of our flow prediction results. This analysis aims to demonstrate the necessity and effectiveness of the loss components we propose. Secondly, we explore how the size of the training dataset, especially in the case of limited training data, affects the outcomes of our self-supervised training process.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "SeFlow: A Self-Supervised Scene Flow Method in Autonomous Driving", + "bbox": [ + 258, + 113, + 730, + 130 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/674d1867891e4548232e22dc697832e7559c5aa5eb168b845c9b0c32031c6081.jpg", + "table_caption": [ + "Table 3: Ablation study of loss terms. Results are evaluated on the Argoverse 2 validation set with 20 training epochs." + ], + "table_footnote": [], + "table_body": "
Exp. IdLchamLdchamLstaticLdclsEPE ↓
3-wayFDFSBS
1✓0.09620.2030.0520.033
2✓✓0.09160.1810.0590.035
3✓✓✓0.07790.2200.0120.002
4✓✓✓✓0.06430.1600.0290.004
", + "bbox": [ + 264, + 184, + 727, + 305 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Loss Terms The advantages of each loss design are evident in Tab. 3 evaluated on the Argoverse 2 validation set with 20 training epochs. Instead of relying solely on the chamfer loss $\\mathcal{L}_{\\mathrm{cham}}$ , we investigate how incorporating the additional three losses ( $\\mathcal{L}_{\\mathrm{static}}$ , $\\mathcal{L}_{\\mathrm{dcham}}$ , $\\mathcal{L}_{\\mathrm{dcls}}$ ) can boost the performance.", + "bbox": [ + 212, + 333, + 785, + 393 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "After adding the dynamic Chamfer loss $\\mathcal{L}_{\\mathrm{dcham}}$ , experiment 2 shows a decrease in flow estimation error of about 0.22 (10%) for foreground dynamics (FD), while the static errors (FS, BS) remain essentially the same. Experiment 3 then incorporates $\\mathcal{L}_{\\mathrm{static}}$ constraint, which significantly reduces the foreground and background static flow estimation errors, 80% for FS and 94% for BS. However, adding $\\mathcal{L}_{\\mathrm{static}}$ also increases the foreground dynamic error. We attribute this to the difficulty of estimating the flow of moving, geometrically featureless, objects as mentioned previously, which would be reinforced by $\\mathcal{L}_{\\mathrm{static}}$ . Even so, considering the two static components in the EPE 3-way metric, the EPE 3-way would still benefit considerably from this loss (15% error reduction). Finally, in experiment 4, we incorporate the $\\mathcal{L}_{\\mathrm{dcls}}$ . This results in a notable decrease in overall EPE 3-way by 33% compared to solely using the chamfer loss (experiment 1). The above experiments illustrate that our method is not a simple stack of losses, but a complementary holistic design.", + "bbox": [ + 212, + 393, + 787, + 604 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Training Dataset Size In the context of robotics and autonomous driving, there are situations where the number of accessible frames is limited. This section evaluates the performance of SeFlow given different amounts of training data.", + "bbox": [ + 212, + 627, + 785, + 672 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In this experiment, we denote the entire Argoverse 2 Sensor training set as $100\\%$ . Extra unlabeled data can be retrieved from the LiDAR dataset for self-supervised methods. To assess the impact of dataset size, we conduct the same 50 epochs of training for methods using $10\\%$ , $20\\%$ , $50\\%$ , and $100\\%$ of the total data in Tab. 4 and Fig. 4 and evaluate the resulting models.", + "bbox": [ + 212, + 674, + 785, + 748 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "From Tab. 4, we can observe that with only $20\\%$ or $50\\%$ training data, our SeFlow can already outperform ZeroFlow which uses $100\\%$ or $200\\%$ data. Figure 4 illustrates even more intuitively that our SeFlow can easily outperform both existing supervised (FastFlow3D) and unsupervised (ZeroFlow) approaches with the same amount of data. We attribute the demonstrated data efficiency of our method to the well-designed loss functions, which integrate a dynamic", + "bbox": [ + 212, + 750, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Q Zhang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "awareness mapping method into our framework and enable better scene-level comprehension.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/93ca902642cded9633b6a7fc72aa7a77b6a59c360b87c651b5bb385571aed17e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetEPE ↓
3-wayFDFSBS
10%0.0940.2340.0400.006
20%0.0780.1970.0320.004
50%0.0660.1670.0280.004
100%0.0590.1470.0260.004
ZF 100%0.0880.2310.0220.011
ZF 200%0.0760.1980.0180.011
", + "bbox": [ + 225, + 200, + 498, + 371 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 4: Ablation study of dataset size compared with Zeroflow (ZF), another self-supervised learning method. Results are evaluated on the Argoverse 2 validation set with 50 training epochs. The total dataset $(100\\%)$ is $107\\mathrm{k}$ frames. $200\\%$ data means all of the Sensor dataset plus an equal amount of the LiDAR dataset (214k frames).", + "bbox": [ + 212, + 380, + 503, + 492 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/f331b8fda3d355540cfbc7e315602ca7e50ca3bea6c3270cdc7fc383d3899b8b.jpg", + "image_caption": [ + "Fig.4: The relationship between flow estimation error and training dataset size, scaled in $\\log_{10}$ . Methods with \\* are supervised by ground truth labels. SeFlow uses less data but gets comparable results compared to FastFlow3D and ZeroFlow." + ], + "image_footnote": [], + "bbox": [ + 526, + 200, + 782, + 385 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.4 Qualitative Results", + "text_level": 1, + "bbox": [ + 214, + 523, + 419, + 537 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Figure 5 presents a qualitative flow estimation result on a sequence of scenes in the Argoverse 2 validation set. The first three columns showcase the same scene at different timestamps, and the last column shows a zoomed-in view of the scene.", + "bbox": [ + 212, + 551, + 787, + 611 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In the first column featuring a vehicle partially occluded by a traffic pole, SeFlow demonstrates superior flow estimation capabilities compared to ZeroFlow. Sometimes, SeFlow can even detect flow overlooked by the ground truth annotations. Argoverse 2 derives ground truth flow from manual instance level labels, and as a result, the flow of points outside the bounding box may be ignored. This issue is particularly pronounced in smaller objects, and an example of this can be seen in the fourth column where we zoom in on a case. In this example, a pedestrian is pushing a shopping cart across the road. By comparing the first and the third column, we can observe that both objects are moving. However, the ground truth labels zero flow (white) for the shopping cart. SeFlow, on the other hand, successfully predicts consistent flow even at this small scale. In comparison, the baseline method ZeroFlow suffers at flow prediction of small objects and regards both the pedestrian and the shopping cart as static (white).", + "bbox": [ + 212, + 612, + 787, + 809 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Although our method shows superior scene flow estimation compared to other baseline methods, it also has some limitations. Based on the zoom-in view in the", + "bbox": [ + 212, + 809, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "SeFlow: A Self-Supervised Scene Flow Method in Autonomous Driving", + "bbox": [ + 258, + 113, + 732, + 130 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/fa1e6b21bceb26a2908cdc77e14dd6cee8f7d946660a5657001c89dc85546bfc.jpg", + "image_caption": [ + "Fig. 5: Qualitative results from Argoverse 2 validation set. The top row displays the ground truth flow, the middle row presents the SeFlow result, and the bottom row showcases another self-supervised method ZeroFlow [29] result. Different color indicates different directions and more saturated color means larger flow estimation. Ego motion is compensated for a clearer view." + ], + "image_footnote": [], + "bbox": [ + 212, + 143, + 787, + 372 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "fourth column in Fig. 5, we can find some purple flow labels at the top right corner in the first row (Ground Truth). Both SeFlow (Ours) and ZeroFlow fail to predict any flow for these points. One possible reason is that the point cloud of distant objects is sparse and easily ignored by both the voxelization process and clustering algorithms. This makes it difficult to estimate the flow of distant objects using only two consecutive frames. Based on this, multi-modality or time-consistent networks would be a further direction for future research.", + "bbox": [ + 212, + 483, + 789, + 590 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusions", + "text_level": 1, + "bbox": [ + 215, + 619, + 370, + 635 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we proposed SeFlow, an efficient and effective self-supervised training method for scene flow estimation in autonomous driving with large-scale point clouds as input. Our primary contributions include a learning method that integrates static and dynamic awareness to construct self-supervision objectives. We further identify problems with the correspondence assumptions of Chamfer-based loss functions commonly used for self-supervised learning, and mitigate these with a constraint based on the upper bound of object motion. Our experimental results underscore the efficacy of our approach.", + "bbox": [ + 212, + 656, + 787, + 777 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Future research may concentrate on integrating multi-modality (e.g., cameras and radar) within the SeFlow framework for higher flow estimation accuracy or embedding temporal coherence concepts within our pipeline. Additionally, the optimization of multi-loss learning weights warrants further exploration.", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Q Zhang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 215, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Thanks to RPL member: Li Ling helps review this manuscript. Thanks to Kyle Vedder (the author of ZeroFlow), who kindly opened his code including pretrained weights, and discussed their result with us which helped this work a lot. This work was partially supported by the Wallenberg AI, Autonomous Systems and Software Program (WASP) funded by the Knut and Alice Wallenberg Foundation and Prosense (2020-02963) funded by Vinnova. The computations were enabled by the supercomputing resource Berzelius provided by National Supercomputer Centre at Linköping University and the Knut and Alice Wallenberg Foundation, Sweden.", + "bbox": [ + 212, + 174, + 787, + 310 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 330, + 321, + 347 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. 2, A.: Argoverse 2 scene flow online leaderboard. https://eval.ai/web/challenges/challenge-page/2010/leaderboard/4759 (2024 Mar 4th)", + "2. Aleotti, F., Poggi, M., Tosi, F., Mattoccia, S.: Learning end-to-end scene flow by distilling single tasks knowledge. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 10435-10442 (2020)", + "3. Baur, S.A., Emmerichs, D.J., Moosmann, F., Pinggera, P., Ommer, B., Geiger, A.: Slim: Self-supervised lidar scene flow and motion segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13126-13136 (2021)", + "4. Campello, R.J., Moulavi, D., Sander, J.: Density-based clustering based on hierarchical density estimates. In: Pacific-Asia conference on knowledge discovery and data mining. pp. 160-172. Springer (2013)", + "5. Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., et al.: Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012 (2015)", + "6. Cho, K., Van Merrienboer, B., Gulcehre, C., Bahdanau, D., Bougares, F., Schwenk, H., Bengio, Y.: Learning phrase representations using rnn encoder-decoder for statistical machine translation. arXiv preprint arXiv:1406.1078 (2014)", + "7. Chodosh, N., Ramanan, D., Lucey, S.: Re-evaluating lidar scene flow for autonomous driving. arXiv preprint arXiv:2304.02150 (2023)", + "8. Deng, D., Zakhor, A.: Rsf: Optimizing rigid scene flow from 3d point clouds without labels. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1277-1286 (2023)", + "9. Duberg, D., Zhang, Q., Jia, M., Jensfelt, P.: DUFOMap: Efficient dynamic awareness mapping. IEEE Robotics and Automation Letters 9(6), 5038-5045 (2024). https://doi.org/10.1109/LRA.2024.3387658", + "10. Himmelsbach, M., Hundelshausen, F.V., Wuensche, H.J.: Fast segmentation of 3d point clouds for ground vehicles. In: Intelligent Vehicles Symposium (IV), 2010 IEEE. pp. 560-565. IEEE (2010)", + "11. Jund, P., Sweeney, C., Abdo, N., Chen, Z., Shlens, J.: Scalable scene flow from point clouds in the real world. IEEE Robotics and Automation Letters 7(2), 1589-1596 (2021)", + "12. Kittenplon, Y., Eldar, Y.C., Raviv, D.: Flowstep3d: Model unrolling for self-supervised scene flow estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4114-4123 (2021)" + ], + "bbox": [ + 218, + 359, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "SeFlow: A Self-Supervised Scene Flow Method in Autonomous Driving", + "bbox": [ + 258, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "13. Lang, I., Aiger, D., Cole, F., Avidan, S., Rubinstein, M.: Scoop: Self-supervised correspondence and optimization-based scene flow. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5281-5290 (2023)", + "14. Li, X., Kaesemodel Pontes, J., Lucey, S.: Neural scene flow prior. Advances in Neural Information Processing Systems 34, 7838-7851 (2021)", + "15. Li, X., Zheng, J., Ferroni, F., Pontes, J.K., Lucey, S.: Fast neural scene flow. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9878-9890 (2023)", + "16. Liu, J., Wang, G., Ye, W., Jiang, C., Han, J., Liu, Z., Zhang, G., Du, D., Wang, H.: Difflow3d: Toward robust uncertainty-aware scene flow estimation with diffusion model. arXiv preprint arXiv:2311.17456 (2023)", + "17. Mayer, N., Ilg, E., Hausser, P., Fischer, P., Cremers, D., Dosovitskiy, A., Brox, T.: A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4040-4048 (2016)", + "18. Menze, M., Geiger, A.: Object scene flow for autonomous vehicles. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3061-3070 (2015)", + "19. Mittal, H., Okorn, B., Held, D.: Just go with the flow: Self-supervised scene flow estimation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11177-11185 (2020)", + "20. Najibi, M., Ji, J., Zhou, Y., Qi, C.R., Yan, X., Ettinger, S., Anguelov, D.: Motion inspired unsupervised perception and prediction in autonomous driving. In: European Conference on Computer Vision. pp. 424-443. Springer (2022)", + "21. Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A., Cournapeau, D., Brucher, M., Perrot, M., Duchesnay, E.: Scikit-learn: Machine learning in Python. Journal of Machine Learning Research 12, 2825–2830 (2011)", + "22. Pfreundschuh, P., Hendrikx, H.F., Reijgwart, V., Dubé, R., Siegwart, R., Cramar-iuc, A.: Dynamic object aware lidar slam based on automatic generation of training data. In: 2021 IEEE International Conference on Robotics and Automation (ICRA). pp. 11641-11647. IEEE (2021)", + "23. Schmid, L., Andersson, O., Sulser, A., Pfreundschuh, P., Siegwart, R.: Dynablox: Real-time detection of diverse dynamic objects in complex environments. IEEE Robotics and Automation Letters (RA-L) 8(10), 6259 - 6266 (2023). https://doi.org/10.1109/LRA.2023.3305239", + "24. Shen, Y., Hui, L., Xie, J., Yang, J.: Self-supervised 3d scene flow estimation guided by superpoints. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5271-5280 (2023)", + "25. Song, J., Lee, S.J.: Knowledge distillation of multi-scale dense prediction transformer for self-supervised depth estimation. Scientific Reports (18939) (2023)", + "26. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., et al.: Scalability in perception for autonomous driving: Waymo open dataset. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2446-2454 (2020)", + "27. Tishchenko, I., Lombardi, S., Oswald, M.R., Pollefeys, M.: Self-supervised learning of non-rigid residual flow and ego-motion. In: 2020 international conference on 3D vision (3DV). pp. 150-159. IEEE (2020)", + "28. Vedder, K., Khatri, I., Peri, N., Chodosh, N., Liu, Y., Hays, J.: Av2 2024 scene flow challenge announcement. https://wwwargonse.org/sceneflow.html (2024 Feb 25th)" + ], + "bbox": [ + 217, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Q Zhang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "29. Vedder, K., Peri, N., Chodosh, N., Khatri, I., Eaton, E., Jayaraman, D., Ramanan, Y.L.D., Hays, J.: ZeroFlow: Fast Zero Label Scene Flow via Distillation. International Conference on Learning Representations (ICLR) (2024)", + "30. Vedula, S., Rander, P., Collins, R., Kanade, T.: Three-dimensional scene flow. IEEE transactions on pattern analysis and machine intelligence 27(3), 475-480 (2005)", + "31. Vidanapathirana, K., Chng, S.F., Li, X., Lucey, S.: Multi-body neural scene flow. In: 2024 International Conference on 3D Vision (3DV). pp. 126-136. IEEE (2024)", + "32. Wang, Z., Wei, Y., Rao, Y., Zhou, J., Lu, J.: 3d point-voxel correlation fields for scene flow estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)", + "33. Wei, Y., Wang, Z., Rao, Y., Lu, J., Zhou, J.: PV-RAFT: Point-Voxel Correlation Fields for Scene Flow Estimation of Point Clouds. In: CVPR (2021)", + "34. Wilson, B., Qi, W., Agarwal, T., Lambert, J., Singh, J., et al.: Argoverse 2: Next generation datasets for self-driving perception and forecasting. In: Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks (NeurIPS Datasets and Benchmarks 2021) (2021)", + "35. Wu, H., Li, Y., Xu, W., Kong, F., Zhang, F.: Moving event detection from lidar point streams. Nature Communications 15(1), 345 (2024)", + "36. Wu, W., Wang, Z.Y., Li, Z., Liu, W., Fuxin, L.: Pointpwc-net: Cost volume on point clouds for self-supervised scene flow estimation. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16. pp. 88-107. Springer (2020)", + "37. Zhang, Q., Duberg, D., Geng, R., Jia, M., Wang, L., Jensfelt, P.: A dynamic points removal benchmark in point cloud maps. In: IEEE 26th International Conference on Intelligent Transportation Systems (ITSC). pp. 608-614 (2023)", + "38. Zhang, Q., Yang, Y., Fang, H., Geng, R., Jensfelt, P.: Deflow: Decoder of scene flow network in autonomous driving. arXiv preprint arXiv:2401.16122 (2024)" + ], + "bbox": [ + 215, + 146, + 785, + 535 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "SeFlow: A Self-Supervised Scene Flow Method in Autonomous Driving", + "bbox": [ + 259, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/SeFlow_ A Self-Supervised Scene Flow Method in Autonomous Driving/e7db8302-f2d3-4bc6-8b61-d5de1eb958ee_model.json b/2024/SeFlow_ A Self-Supervised Scene Flow Method in Autonomous Driving/e7db8302-f2d3-4bc6-8b61-d5de1eb958ee_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d25580601f347b9b41c0df99b95a0239564e28e1 --- /dev/null +++ b/2024/SeFlow_ A Self-Supervised Scene Flow Method in Autonomous Driving/e7db8302-f2d3-4bc6-8b61-d5de1eb958ee_model.json @@ -0,0 +1,2269 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.284, + 0.141, + 0.723, + 0.188 + ], + "angle": 0, + "content": "SeFlow: A Self-Supervised Scene Flow Method in Autonomous Driving" + }, + { + "type": "text", + "bbox": [ + 0.322, + 0.212, + 0.681, + 0.244 + ], + "angle": 0, + "content": "Qingwen Zhang\\(^{1}\\), Yi Yang\\(^{1,2}\\), Peizheng Li\\(^{3,4}\\), Olov Andersson\\(^{1}\\), and Patric Jensfelt\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.288, + 0.255, + 0.717, + 0.282 + ], + "angle": 0, + "content": "\\(^{1}\\) RPL, KTH Royal Institute of Technology, Stockholm, Sweden \n\\(^{2}\\) Scania CV AB, Södertälje, Sweden" + }, + { + "type": "text", + "bbox": [ + 0.345, + 0.283, + 0.657, + 0.298 + ], + "angle": 0, + "content": "3 University of Tübingen, Tübingen, Germany" + }, + { + "type": "text", + "bbox": [ + 0.349, + 0.297, + 0.65, + 0.326 + ], + "angle": 0, + "content": "4 Mercedes-Benz AG, Sindelfingen, Germany {qingwen,yiya,olovand,patric}@kth.se," + }, + { + "type": "text", + "bbox": [ + 0.293, + 0.327, + 0.712, + 0.339 + ], + "angle": 0, + "content": "carol-yi.yang@scania.com, peizheng.li@mercedes-benz.com" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.374, + 0.744, + 0.624 + ], + "angle": 0, + "content": "Abstract. Scene flow estimation predicts the 3D motion at each point in successive LiDAR scans. This detailed, point-level, information can help autonomous vehicles to accurately predict and understand dynamic changes in their surroundings. Current state-of-the-art methods require annotated data to train scene flow networks and the expense of labeling inherently limits their scalability. Self-supervised approaches can overcome the above limitations, yet face two principal challenges that hinder optimal performance: point distribution imbalance and disregard for object-level motion constraints. In this paper, we propose SeFlow, a self-supervised method that integrates efficient dynamic classification into a learning-based scene flow pipeline. We demonstrate that classifying static and dynamic points helps design targeted objective functions for different motion patterns. We also emphasize the importance of internal cluster consistency and correct object point association to refine the scene flow estimation, in particular on object details. Our real-time capable method achieves state-of-the-art performance on the self-supervised scene flow task on Argoverse 2 and Waymo datasets. The code is open-sourced at https://github.com/KTH-RPL/SeFlow." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.636, + 0.741, + 0.665 + ], + "angle": 0, + "content": "Keywords: 3D scene flow, self-supervised, autonomous driving, large-scale point cloud" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.689, + 0.377, + 0.705 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.81 + ], + "angle": 0, + "content": "Scene flow [30] captures the 3D velocity at every point in a point cloud. These detailed 3D flow estimates can enhance downstream tasks in autonomous driving, such as detection, segmentation, tracking, and occupancy flow estimation [20]. A common paradigm for addressing the scene flow problem is supervised learning by utilizing annotated LiDAR data [11, 38]. However, expensive labeling inherently limits the scalability of supervised learning methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Given the difficulty and expense of labeling scene flow ground truth, we instead focus on self-supervised scene flow approaches. Existing self-supervised" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Q Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.784, + 0.357 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.367, + 0.788, + 0.492 + ], + "angle": 0, + "content": "Fig. 1: LiDAR scene flow estimation using our SeFlow method on Argoverse 2. The predicted scene flow for each point is color-coded based on direction. The white indicates static points whose flow is zero. More saturated colors indicate higher velocities. (a) Camera view for visualization purposes only. (b),(c) are zoomed-in views showing the baseline from ZeroFlow [29] as well as SeFlow (ours). When predicting the flow of a large and long vehicle, the baseline predicts a portion of the flow as 0, whereas our estimates are consistent. In addition, the baseline tends to ignore small-scale objects, e.g., pedestrians, while our method can better handle such small and slow-moving objects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.787, + 0.613 + ], + "angle": 0, + "content": "methods can be divided into two categories: knowledge distillation and data exploration. Knowledge distillation methods [2,25,29] typically rely on a 'teacher' method to generate pseudo flow labels. These pseudo labels are then used to supervise student models. Data exploration methods [3, 14, 15, 19, 31], on the other hand, directly utilize the predicted flow to project the input points and find nearest neighbors in the next frame to establish constraints for training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.765 + ], + "angle": 0, + "content": "A key challenge for self-supervised methods is that the vast majority of the points are static (about \\(86\\%\\) of points are background [7, 34]). This data imbalance often leads to overly conservative scene flow predictions. One common way to counter this is to use large amounts of training data as in ZeroFlow [29]. ZeroFlow experimentally shows that more data helps to improve the dynamic flow accuracy. In this paper, we investigate ways to improve data efficiency and take inspiration from supervised methods [11, 38]. We address the data imbalance by first classifying points as dynamic or static based on traditional ray casting when integrating frames [9]. This allows us to constrain the corresponding motion patterns by proposing novel loss functions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Another shortcoming of existing self-supervised scene flow methods is that they disregard object-level motion constraints. The flow should be consistent, i.e., all points in a rigid object should have similar flows. A clear case of inconsistent flow can be seen on the big truck in Fig. 1, where ZeroFlow [29] predicts an absence of flow in certain sections. This inconsistency is caused by incorrect" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.259, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SeFlow: A Self-Supervised Scene Flow Method in Autonomous Driving" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.783, + 0.236 + ], + "angle": 0, + "content": "object point associations which also can be observed in small-scale objects, as shown in the case of the pedestrian. To account for object-level motion constraints, we propose to cluster the dynamic points into object candidates and encourage consistent flow and correct associations for the points in each cluster. This reduces the fragmentation of the flow estimate for points inside the same object." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.784, + 0.313 + ], + "angle": 0, + "content": "Overall, our method improves the estimated flow accuracy and addresses issues in previous methods as shown in Fig. 1(c). We propose an efficient and effective self-supervised scene flow method that integrates traditional classification and learning-based strategies. Our approach is open-source at https://github.com/KTH-RPL/SeFlow. Our primary contributions include:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.324, + 0.784, + 0.354 + ], + "angle": 0, + "content": "- We propose SeFlow, a novel method that integrates a dynamic classification method in formulating efficient self-supervision objectives." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.355, + 0.785, + 0.399 + ], + "angle": 0, + "content": "- We further construct loss functions to learn dynamic flow estimation in imbalanced data and ensure consistent object-level flow, mitigating the effects of correspondence errors." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.4, + 0.785, + 0.445 + ], + "angle": 0, + "content": "- We show that SeFlow achieves state-of-the-art results on the self-supervised scene flow task in Argoverse 2 and Waymo datasets and even outperforms all but one supervised method on the leaderboard." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.324, + 0.785, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.47, + 0.387, + 0.485 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.503, + 0.784, + 0.549 + ], + "angle": 0, + "content": "In this section, we explore existing works in scene flow estimation. We also discuss current traditional frameworks capable of classifying dynamic points, highlighting their relevance and application in the context of scene flow tasks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.572, + 0.451, + 0.586 + ], + "angle": 0, + "content": "2.1 Scene Flow Estimation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.785, + 0.779 + ], + "angle": 0, + "content": "Scene flow estimation in autonomous driving is slightly different from flow estimation in object registration. Methods in object registration [8, 13, 16, 24, 32, 33] focus on relatively small-scale point cloud data like synthetic datasets Shapenet [5] and FlyingThing3D [17]. These methods scale poorly with the number of points. When applied to point cloud data for autonomous driving, they require downsampling to 8,192 points or less [11, 18]. In recent datasets like Argoverse 2 [34] and Waymo [11], the number of points in one full frame is around \\(80\\mathrm{k - }177\\mathrm{k}\\). Methods that can handle such large-scale point cloud data as input usually use a voxel-based pipeline [11, 38]. However, expensive labeling of ground truth flow limits the scalability of these supervised methods, especially in autonomous driving where we have continuously increasing amounts of potential training data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.785, + 0.84 + ], + "angle": 0, + "content": "To train models without labeled data, recent methods propose self-supervised losses. Many commonly used losses, such as Chamfer distance, are based on the nearest neighbor distance between two point cloud inputs [3,12,14,15,19,27,36]. However, a major limitation of nearest neighbor based losses is that only part" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Q Zhang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.312 + ], + "angle": 0, + "content": "of the points on dynamic objects provides good supervision due to incorrect associations. This is especially apparent when big trucks are moving fast in autonomous driving scenarios (see Fig. 1 and Fig. 3). Mittal et al. [19] define a self-supervised loss by tracking a patch forward and backward in time to form a cycle while penalizing the errors through cycle consistency and feature similarity. Baur et al. [3] propose three loss functions with k-NN loss, rigid cycle consistency inspired by Mittal, and artificial labels based on the distance between points in the original and estimated point clouds. In the following we describe NSFP [14], FastNSF [15], and ZeroFlow [29] in more detail. These are the publically available methods that perform best on the Argoverse 2 self-supervised scene flow challenge and are therefore used as our baselines." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.313, + 0.788, + 0.432 + ], + "angle": 0, + "content": "NSFP [14] is an optimization-based method. For each pair of consecutive input point clouds, NSFP iteratively learns new weights for an MLP network to predict the flow using the Chamfer distance loss. However, thousands of iterations are needed and their runtimes extend from 26 to 35 seconds per frame, which fails to meet the real-time requirements of autonomous driving. FastNSF [15] improves the efficiency by voxelizing the point cloud first and then converting it to a distance transform map for faster neighbor calculation. This reduces the runtime to 0.5 seconds, at the expense of increased error." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.433, + 0.788, + 0.524 + ], + "angle": 0, + "content": "ZeroFlow [29] adopts a semi-supervised strategy. Pseudo-flow labels are created offline using NSFP [14], and a FastFlow3D [11] model is used as a student for knowledge distillation. This setup allows the student model to perform real-time inference. In this case, the teacher network needs significant resources (around 3.6 GPU months [29]) to label the entire training data. The final accuracy is also influenced by the performance of the teacher." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.524, + 0.788, + 0.6 + ], + "angle": 0, + "content": "To increase data efficiency and address the limitations of the above methods, we propose to integrate efficient ray-casting-based dynamic awareness mapping into our pipeline. The core idea is to classify which points move and cluster these points into objects for which we can estimate group-level motion statistics and define better self-supervised loss functions." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.621, + 0.53, + 0.637 + ], + "angle": 0, + "content": "2.2 Dynamic Awareness in Mapping" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.734 + ], + "angle": 0, + "content": "In the field of Simultaneous Localization and Mapping (SLAM), there is a growing interest in dynamic awareness [9,22,23,35,37]. This interest stems from the fact that points associated with moving objects in a scene can significantly impact localization and planning performance. Existing dynamic awareness methods in mapping [9,23,37] often utilize ray casting techniques to binary classify points as either dynamic or static." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.84 + ], + "angle": 0, + "content": "The scene flow task, on the other hand, aims to predict the specific 3D flow at each point, a goal that extends beyond the mere categorization of dynamic and static points. In mapping, a point is considered dynamic if it moves once within a scene (even if it becomes static later), whereas, in scene flow tasks, a point is deemed dynamic if it moves faster than a certain velocity threshold in the current frame. Despite these differences, insights from the mapping field are invaluable. By integrating information over time, these frameworks develop a" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.259, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "SeFlow: A Self-Supervised Scene Flow Method in Autonomous Driving" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "more comprehensive understanding of the entire scene. Such scene level dynamic awareness can inform and enhance our exploration of point cloud data in scene flow tasks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.216, + 0.441, + 0.231 + ], + "angle": 0, + "content": "3 Problem Statement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.247, + 0.787, + 0.31 + ], + "angle": 0, + "content": "This work addresses the problem of real-time scene flow estimation in autonomous driving. Given two consecutive input point clouds, \\(\\mathcal{P}_t\\) and \\(\\mathcal{P}_{t + 1}\\), along with the ego movement \\(\\mathbf{T}_{t,t + 1} \\in SE(3)\\), the goal is to predict the motion vector as flow \\(\\hat{\\mathcal{F}}_{t,t + 1}(p) = (x,y,z)^T\\) for each point \\(p \\in \\mathcal{P}_t\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.31, + 0.786, + 0.356 + ], + "angle": 0, + "content": "The objective is to minimize the End Point Error (EPE) which represents the difference between the predicted flow and the ground truth flow, as expressed by the following equation:" + }, + { + "type": "equation", + "bbox": [ + 0.389, + 0.366, + 0.786, + 0.422 + ], + "angle": 0, + "content": "\\[\n\\min \\underbrace {\\frac {1}{| \\mathcal {P} _ {t} |} \\sum_ {p \\in \\mathcal {P} _ {t}} \\left\\| \\hat {\\mathcal {F}} (p) - \\mathcal {F} _ {g t} (p) \\right\\| _ {2}} _ {\\text {E P E}}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.432, + 0.785, + 0.479 + ], + "angle": 0, + "content": "where \\( |\\mathcal{P}_t| \\) denotes the cardinality of (i.e., the number of points in) \\( \\mathcal{P}_t \\). For consistency, in the subsequent presentation, capital symbols correspond to sets, while the lowercase symbols represent variables of specific points." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.501, + 0.331, + 0.516 + ], + "angle": 0, + "content": "4 Method" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.533, + 0.411, + 0.548 + ], + "angle": 0, + "content": "4.1 Input and Output" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.557, + 0.784, + 0.603 + ], + "angle": 0, + "content": "The first step to process the input data, as is commonly done, is to remove ground points from \\(\\mathcal{P}_t\\) and \\(\\mathcal{P}_{t + 1}\\). This is typically done using HD maps [26,34] or ground segmentation techniques [10]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.603, + 0.784, + 0.631 + ], + "angle": 0, + "content": "The estimated flow \\(\\tilde{\\mathcal{F}}\\) from \\(\\mathcal{P}_t\\) to \\(\\mathcal{P}_{t + 1}\\) can be decomposed into two parts as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.442, + 0.632, + 0.785, + 0.649 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {F}} = \\mathcal {F} _ {e g o} + \\Delta \\hat {\\mathcal {F}}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.655, + 0.784, + 0.688 + ], + "angle": 0, + "content": "where \\(\\mathcal{F}_{ego}\\) is the flow resulting from the ego vehicle's motion which can be directly obtained from \\(\\mathbf{T}_{t,t + 1}\\), and \\(\\varDelta\\hat{\\mathcal{F}}\\) is our network output." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.71, + 0.401, + 0.724 + ], + "angle": 0, + "content": "4.2 Model Backbone" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.785, + 0.841 + ], + "angle": 0, + "content": "To enable real-time computation and estimation of scene flow across a large point set, voxelization is considered a practical encoding strategy in the model backbone. However, the reduction in resolution often leads to a poorer distinction of point features within the same voxel. With this in mind, we use DeFlow [38] as the architectural basis in this paper. DeFlow integrates GRU [6] with iterative refinement in the decoder design. More specifically, the GRU module maintains voxel features as its hidden state and selectively forgets or updates" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Q Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.147, + 0.777, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.234, + 0.23, + 0.662, + 0.243 + ], + "angle": 0, + "content": "Training Stage Sec. 4.4 Self-supervised Loss (including Sec. 4.3 Dynamic Classification)" + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.254, + 0.774, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.397, + 0.788, + 0.535 + ], + "angle": 0, + "content": "Fig. 2: SeFlow Architecture. Top: With two consecutive point clouds as inputs, our model predicts the estimated flows of all points. Bottom: Conceptual visualization of the Chamfer loss and the three proposed training losses. With the original input \\(\\mathcal{P}_t\\) (2 static points for the building, and 2 dynamic points for the car) plus the estimated flow \\(\\hat{\\mathcal{F}}_t\\), we can calculate the error between estimated \\(\\hat{\\mathcal{P}}_{t+1}\\) and the next frame point cloud \\(\\mathcal{P}_{t+1}\\) (\\(\\mathcal{L}_{\\mathrm{cham}}\\)). The second part is \\(\\mathcal{L}_{\\mathrm{dcham}}\\) that only calculates the distance error between dynamic points. The third loss says that the estimated flows of static points should be zero. Finally, we assume that the flow at points from the same cluster should be consistent, and mitigate underestimation by using the proposed upper bound on the flow." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.788, + 0.66 + ], + "angle": 0, + "content": "the information of the hidden state during each iteration according to the input point features. After multiple iterations, the optimized voxel features are concatenated with the original point features to obtain the final individual point features. Benefiting from this design, we improve the inference efficiency compared to the commonly used backbone FastFlow3D [11] without sacrificing the accuracy of scene flow estimation in coarse resolution settings." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.688, + 0.53, + 0.703 + ], + "angle": 0, + "content": "4.3 Dynamic Classification of Points" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.842 + ], + "angle": 0, + "content": "To facilitate dynamic classification of points during the training stage, we incorporate the DUFOMap framework, a mapping-based dynamic awareness approach [9]. The key insight of this is that points observed inside a region that at one time has been observed as empty must be dynamic. Built on this insight, DUFOMap utilizes ray-casting to classify dynamic points at the sensor rate on the CPU. The result is two disjoint sets, \\(\\mathcal{P}_{t,d}\\) and \\(\\mathcal{P}_{t,s}\\), where \\(\\mathcal{P}_{t,d}\\) is the set of dynamic points that have moved once inside a scene (even if they later become static) and \\(\\mathcal{P}_{t,s}\\) is the set of static points that did not move at any time." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.259, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "SeFlow: A Self-Supervised Scene Flow Method in Autonomous Driving" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "Note that the dynamic-static classification is separated from the inference and training, offering our method good flexibility." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.196, + 0.428, + 0.211 + ], + "angle": 0, + "content": "4.4 Self-supervised Loss" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.217, + 0.784, + 0.293 + ], + "angle": 0, + "content": "As discussed, self-supervised learning with only Chamfer distance loss is vulnerable to problems with data imbalance and incorrect associations. We therefore construct three additional loss functions, illustrated in Fig. 2 and described in turn below, to mitigate problems with data imbalance as well as encourage consistent object-level flow." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.311, + 0.784, + 0.341 + ], + "angle": 0, + "content": "Chamfer Distance The Chamfer distance, a common self-supervised loss in existing work [3, 14, 15, 19], has the following definition:" + }, + { + "type": "equation", + "bbox": [ + 0.301, + 0.348, + 0.785, + 0.386 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c h a m}} = \\frac {1}{| \\hat {\\mathcal {P}} _ {t + 1} |} \\sum_ {p \\in \\hat {\\mathcal {P}} _ {t + 1}} \\mathrm {S} (p, \\mathcal {P} _ {t + 1}) + \\frac {1}{| \\mathcal {P} _ {t + 1} |} \\sum_ {p \\in \\mathcal {P} _ {t + 1}} \\mathrm {S} (p, \\hat {\\mathcal {P}} _ {t + 1}) \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.286, + 0.391, + 0.785, + 0.414 + ], + "angle": 0, + "content": "\\[\n\\mathrm {S} (p, \\mathcal {P} _ {o}) = \\min _ {p _ {i} \\in \\mathcal {P} _ {o}} \\| p - p _ {i} \\| _ {2} ^ {2}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.421, + 0.784, + 0.452 + ], + "angle": 0, + "content": "where, \\(\\hat{\\mathcal{P}}_{t + 1} = \\mathcal{P}_t + \\hat{\\mathcal{F}}_t\\), \\(\\mathrm{S}(\\cdot) = \\mathrm{D}(\\cdot)^2\\), \\(\\mathrm{D}(p,\\mathcal{P}_o)\\) denotes the distance between point \\(p\\) and its nearest neighbor in \\(\\mathcal{P}_o\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.453, + 0.785, + 0.573 + ], + "angle": 0, + "content": "The Chamfer distance is proposed to calculate a similarity between two point clouds. However, when using it directly as a supervised signal for scene flow in autonomous driving, there are two issues we note. First, the number of static points is often much larger than that of dynamic points and averaging \\( S(\\cdot) \\) over all points leads the training to favor static points, i.e., zero flow estimation. Second, due to erroneous correspondence assumptions, only part of the flows within a dynamic object can be estimated correctly as shown in Fig. 3, where the estimated flows of points in the overlap area are zero." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.574, + 0.784, + 0.604 + ], + "angle": 0, + "content": "To solve these issues, we contribute the following constraints using the dynamic classification information prior to supervising the network." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.622, + 0.785, + 0.772 + ], + "angle": 0, + "content": "Dynamic Chamfer Distance The imbalance in the number of dynamic and static points presents a significant challenge in scene flow estimation, as highlighted in previous works [11,29,38]. Supervised methods, provided with ground truth labels for point velocity or classification, can weight their loss functions to account for this imbalance. For self-supervised learning, we introduce a dynamic Chamfer distance. In this context, 'dynamic' points are identified based on the output from Sec. 4.3. Specifically, this loss, \\(\\mathcal{L}_{\\mathrm{dcham}}\\), is only computed on points classified as dynamic in two consecutive point clouds. By exclusively considering dynamic points, this loss captures the nuances of motion in point cloud data. This is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.246, + 0.777, + 0.785, + 0.816 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {d c h a m}} = \\frac {1}{| \\hat {\\mathcal {P}} _ {t + 1 , d} |} \\sum_ {p \\in \\hat {\\mathcal {P}} _ {t + 1, d}} \\mathrm {S} (p, \\mathcal {P} _ {t + 1, d}) + \\frac {1}{| \\mathcal {P} _ {t + 1 , d} |} \\sum_ {p \\in \\mathcal {P} _ {t + 1, d}} \\mathrm {S} (p, \\hat {\\mathcal {P}} _ {t + 1, d}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.824, + 0.414, + 0.842 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathcal{P}}_{t + 1,d} = \\mathcal{P}_{t,d} + \\hat{\\mathcal{F}}_{t,d}\\)" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Q Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.142, + 0.518, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.296, + 0.249, + 0.42, + 0.261 + ], + "angle": 0, + "content": "(a) Chamfer distance" + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.143, + 0.77, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.571, + 0.249, + 0.72, + 0.262 + ], + "angle": 0, + "content": "(b) Point flow magnitude" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.274, + 0.788, + 0.358 + ], + "angle": 0, + "content": "Fig. 3: Simple visualization of the shortcomings of using Chamfer distance as a supervisory signal for flow value estimation. The denser color on points in (b) represents higher flow values and white means the point's flow is zero. (a) illustrates how to calculate loss based on Chamfer distance, and (b) shows that the flow results, based on the nearest neighbor principle, can lead to zero flow estimation for the middle of the object." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.386, + 0.789, + 0.492 + ], + "angle": 0, + "content": "Static Flow The standard Chamfer distance rests on a very strong assumption: nearest neighbor matching can find perfect one-to-one correspondences of all points between two frames. However, due to varying observations, the number of points in the two frames differs, leading to inconsistencies and potential mismatches. To deal with the possible mismatches in the static areas, we add a constraint to encourage the model to estimate zero flow for static points. This loss is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.384, + 0.498, + 0.787, + 0.536 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s t a t i c}} = \\frac {1}{| \\mathcal {P} _ {t , s} |} \\sum_ {p \\in \\mathcal {P} _ {t, s}} \\| \\Delta \\hat {\\mathcal {F}} (p) \\| _ {2} ^ {2}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.546, + 0.651, + 0.565 + ], + "angle": 0, + "content": "where \\(\\varDelta\\hat{\\mathcal{F}}(p)\\) represents the network output flow of point \\(p\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.789, + 0.748 + ], + "angle": 0, + "content": "Dynamic Cluster Flow For dynamic fields, the inconsistencies and correspondence errors are even more complicated. For instance, we observe that nearest neighbor matching often leads to erroneous results, e.g., on large objects undergoing translation, which is very common for moving vehicles in urban street environments as shown in Fig. 1. More specifically, Figure 3 illustrates the issue on a simplified car example. In this case, the flow is parallel to the object's surface, which means that nearest neighbor matching does not provide good supervision in the interior of the surface (\\( \\mathrm{D} \\approx 0 \\) for the points in the overlap area). This mismatch results in an incorrect local optimum that remains unresolved during the training process as it minimizes the losses based on Chamfer Distance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.824 + ], + "angle": 0, + "content": "Therefore, we suggest that the motion of all parts within an object (a cluster) should be approximately homogeneous over a short time interval. We use the HDBSCAN [4,21] clustering algorithm to identify different moving objects. This clustering is only applied to dynamic points, thus reducing the computational overhead." + }, + { + "type": "equation", + "bbox": [ + 0.41, + 0.825, + 0.787, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\mathcal {C} _ {t, d} = \\operatorname {H D B S C A N} \\left(\\mathcal {P} _ {t, d}\\right). \\tag {7}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.259, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "SeFlow: A Self-Supervised Scene Flow Method in Autonomous Driving" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.314 + ], + "angle": 0, + "content": "Directly constraining the variance of the flow distribution inside a cluster would be a straightforward idea, but does not guarantee the correctness of the obtained mean value after convergence. As we know that nearest-neighbour point correspondence can considerably underestimate flow on geometrically featureless object surfaces, we here instead propose to use an upper bound on the object motion as the supervisory signal. We derive this upper bound, per object cluster, by taking the maximum inter-frame distance of all point correspondences within the cluster. Specifically, we exploit information from the original input point cloud data, i.e., \\(\\mathcal{P}_t\\) and \\(\\mathcal{P}_{t + 1}\\). After the dynamic classification and clustering process, we find the index of the point \\(p_k\\) in cluster \\(c_{i} \\in \\mathcal{C}_{t,d}\\) with the largest distance to its nearest neighbor point in \\(\\mathcal{P}_{t + 1,d}\\), i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.367, + 0.326, + 0.786, + 0.349 + ], + "angle": 0, + "content": "\\[\n\\kappa_ {i} = \\underset {k} {\\arg \\max } \\left\\{\\mathrm {D} \\left(p _ {k}, \\mathcal {P} _ {t + 1, d}\\right) \\mid p _ {k} \\in \\mathcal {P} _ {c _ {i}} \\right\\}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.362, + 0.668, + 0.38 + ], + "angle": 0, + "content": "We calculate the upper bound, \\(\\tilde{f}_{c_i}\\) on the flow for cluster \\(c_{i}\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.447, + 0.39, + 0.785, + 0.408 + ], + "angle": 0, + "content": "\\[\n\\tilde {f} _ {c _ {i}} = p _ {\\kappa_ {i}} ^ {\\prime} - p _ {\\kappa_ {i}}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.419, + 0.787, + 0.452 + ], + "angle": 0, + "content": "where \\( p_{\\kappa_i}' \\) is the nearest neighbor of point \\( p_{\\kappa_i} \\) in \\( \\mathcal{P}_{t+1,d} \\). We use this to drive the estimated flows of cluster \\( c_i \\) towards \\( \\tilde{f}_{c_i} \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.418, + 0.465, + 0.785, + 0.497 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c _ {i}} = \\sum_ {p _ {j} \\in \\mathcal {P} _ {c _ {i}}} | | \\hat {f} _ {p _ {j}} - \\tilde {f} _ {c _ {i}} | | _ {2} ^ {2}, \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.407, + 0.502, + 0.785, + 0.537 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {d c l s}} = \\frac {1}{| \\mathcal {P} _ {t , d} |} \\sum_ {c _ {i} \\in \\mathcal {C} _ {t, d}} \\mathcal {L} _ {c _ {i}}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.549, + 0.784, + 0.578 + ], + "angle": 0, + "content": "In conclusion, the final SeFlow loss incorporates all four losses introduced above," + }, + { + "type": "equation", + "bbox": [ + 0.356, + 0.58, + 0.785, + 0.595 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {c h a m}} + \\mathcal {L} _ {\\text {s t a t i c}} + \\mathcal {L} _ {\\text {d c h a m}} + \\mathcal {L} _ {\\text {d c l s}}. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.602, + 0.741, + 0.617 + ], + "angle": 0, + "content": "The effect of each loss will be evaluated in the ablation study in Sec. 5.3." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.64, + 0.367, + 0.657 + ], + "angle": 0, + "content": "5 Experiment" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.672, + 0.788, + 0.749 + ], + "angle": 0, + "content": "In this section, we first outline evaluation details and then present quantitative comparisons with state-of-the-art methods on two benchmark datasets. A series of ablation studies are presented to better evaluate the individual components of our approach. Finally, we conclude with qualitative results on Argoverse 2 and discuss limitations of the approach." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.77, + 0.424, + 0.785 + ], + "angle": 0, + "content": "5.1 Dataset and Metric" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "We briefly introduce the dataset and metrics for evaluation in the following section. More implementation details, such as hyperparameters for training and dataset descriptions, can be found in the supplementary material." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Q Zhang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.144, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 1: Comparisons on Argoverse 2 test set from the online leaderboard [1]. Upper groups are supervised methods, lower groups are self-supervised methods. Our method achieves state-of-art performance in the self-supervised scene flow task. \\(\\dagger\\) means methods can run in real-time (10 Hz) onboard." + }, + { + "type": "table", + "bbox": [ + 0.258, + 0.213, + 0.742, + 0.37 + ], + "angle": 0, + "content": "
MethodRun Time per frame [ms]EPE ↓
3-wayFDFSBS
FastFlow3D†[11]34 ± 50.07820.20720.02530.0020
DeFlow†[38]48 ± 40.05340.13400.02320.0029
FastNSF [15]507 ± 3120.16570.35400.04060.1025
NSFP [14]32,060 ± 10,1120.06850.15030.03020.0248
ZeroFlow†[29]34 ± 50.08140.21090.02540.0080
SeFlow (Ours)†48 ± 40.06280.15250.03210.0038
" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.396, + 0.788, + 0.562 + ], + "angle": 0, + "content": "Dataset We evaluate our approach on two large-scale autonomous driving datasets: Argoverse 2 [34] and Waymo [26]. Ground removal is performed using HD maps for both datasets as described in [34]. Waymo datasets contain 798 training and 202 validation scenes respectively. We focus our description here on Argoverse 2, which provides official and public scene flow challenges [1,28], with the Sensor and LiDAR datasets. The Sensor dataset encompasses 700 training and 150 validation scenes. Each scene is approximately 15 seconds long in \\(10\\mathrm{Hz}\\) with complete annotations for evaluation. The evaluation for another 150 test scenes can be accessed indirectly by submitting a solution to the leaderboard. The LiDAR dataset contains 20,000 scenes without any annotation and is only used as extra data in Sec. 5.3." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.583, + 0.788, + 0.735 + ], + "angle": 0, + "content": "Metric The benchmark follows existing works [1,7,29,38] and uses the three-way End Point Error. End Point Error (EPE), as defined in Eq. (1), measures the L2 norm of the discrepancy between the predicted and actual flow vectors, expressed in meters. The EPE three-way (3-way) is defined as the unweighted average EPE across three disjoint sets of points: Foreground Dynamic (FD), Foreground Static (FS), and Background Static (BS). The definition of 'dynamic' is as follows: If the flow at a point exceeds the threshold (the public leaderboard setting is \\(0.05\\mathrm{m}\\)), the point is defined as dynamic. Given the \\(10\\mathrm{Hz}\\) sensor frequency, this threshold corresponds to a speed of \\(0.5\\mathrm{m/s}\\). All evaluations are limited to points that are within a \\(100\\mathrm{m} \\times 100\\mathrm{m}\\) box centered on the ego vehicle." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.755, + 0.433, + 0.771 + ], + "angle": 0, + "content": "5.2 Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.779, + 0.787, + 0.841 + ], + "angle": 0, + "content": "We evaluate the performance of our method SeFlow and compare it with the currently best-performing methods on the Argoverse 2 test set and Waymo validation set. In Tab. 1 and Tab. 2, the upper group, FastFlow3D and DeFlow, are supervised methods trained with ground truth flow. Compared to the supervised" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.259, + 0.114, + 0.732, + 0.131 + ], + "angle": 0, + "content": "SeFlow: A Self-Supervised Scene Flow Method in Autonomous Driving" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.203 + ], + "angle": 0, + "content": "Table 2: Comparisons on Waymo Open dataset validation set. Upper groups are supervised methods, lower groups are self-supervised methods. Our method achieves state-of-art performance in the self-supervised scene flow task. \\(\\dagger\\) means methods can run in real-time (10 Hz) onboard." + }, + { + "type": "table", + "bbox": [ + 0.258, + 0.213, + 0.742, + 0.37 + ], + "angle": 0, + "content": "
MethodRun Time per frame [ms]EPE ↓
3-wayFDFSBS
FastFlow3D†[11]27 ± 60.07820.19540.02460.0152
DeFlow†[38]42 ± 40.04460.09800.02590.0098
FastNSF [14]593 ± 3080.15790.30120.01460.0403
NSFP [15]76,163 ± 32,2560.10050.17120.10810.0221
ZeroFlow†[29]27 ± 60.09210.21620.01530.0241
SeFlow (Ours)†42 ± 40.05980.15060.01810.0106
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.398, + 0.785, + 0.458 + ], + "angle": 0, + "content": "methods, we note that SeFlow outperforms FastFlow3D and approaches the level of DeFlow in terms of EPE 3-way. This comparative analysis demonstrates the great potential of self-supervised learning in scene flow estimation and validates the effectiveness of our approach." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.459, + 0.788, + 0.594 + ], + "angle": 0, + "content": "In the self-supervised category, our SeFlow method achieves state-of-the-art performance on both datasets. While FastNSF and NSFP are optimization-based methods that do not rely on pre-trained weights for subsequent estimations, their inference times are not conducive to real-time requirements. NSFP, with the second best result in Tab. 1, takes approximately 30 seconds to predict a single frame, which is impractical for real-time autonomous driving applications. FastNSF, on the other hand, enhances efficiency through voxelization for quicker Chamfer distance calculation. Although significantly faster than NSFP, the performance of FastNSFP is the worst among all methods evaluated." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.595, + 0.788, + 0.701 + ], + "angle": 0, + "content": "ZeroFlow is capable of estimating scene flow in real-time, with an accuracy similar to NSFP (slightly better on Waymo and slightly worse on Argoverse 2). SeFlow stands out not only for achieving the best result, but also for drastically reducing the run time to 50 milliseconds, which is more than two orders of magnitude faster than NSFP while providing more accurate flow estimates. SeFlow's state-of-the-art performance demonstrated in Tab. 1 and Tab. 2 underscores the effectiveness of our novel self-supervised method in scene flow estimation." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.724, + 0.388, + 0.74 + ], + "angle": 0, + "content": "5.3 Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this section, we delve into two key aspects of our SeFlow pipeline. Firstly, we examine the impact of different loss terms on the accuracy of our flow prediction results. This analysis aims to demonstrate the necessity and effectiveness of the loss components we propose. Secondly, we explore how the size of the training dataset, especially in the case of limited training data, affects the outcomes of our self-supervised training process." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Q Zhang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.787, + 0.174 + ], + "angle": 0, + "content": "Table 3: Ablation study of loss terms. Results are evaluated on the Argoverse 2 validation set with 20 training epochs." + }, + { + "type": "table", + "bbox": [ + 0.266, + 0.185, + 0.728, + 0.306 + ], + "angle": 0, + "content": "
Exp. IdLchamLdchamLstaticLdclsEPE ↓
3-wayFDFSBS
1✓0.09620.2030.0520.033
2✓✓0.09160.1810.0590.035
3✓✓✓0.07790.2200.0120.002
4✓✓✓✓0.06430.1600.0290.004
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.334, + 0.786, + 0.394 + ], + "angle": 0, + "content": "Loss Terms The advantages of each loss design are evident in Tab. 3 evaluated on the Argoverse 2 validation set with 20 training epochs. Instead of relying solely on the chamfer loss \\(\\mathcal{L}_{\\mathrm{cham}}\\), we investigate how incorporating the additional three losses (\\(\\mathcal{L}_{\\mathrm{static}}\\), \\(\\mathcal{L}_{\\mathrm{dcham}}\\), \\(\\mathcal{L}_{\\mathrm{dcls}}\\)) can boost the performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.395, + 0.788, + 0.606 + ], + "angle": 0, + "content": "After adding the dynamic Chamfer loss \\(\\mathcal{L}_{\\mathrm{dcham}}\\), experiment 2 shows a decrease in flow estimation error of about 0.22 (10%) for foreground dynamics (FD), while the static errors (FS, BS) remain essentially the same. Experiment 3 then incorporates \\(\\mathcal{L}_{\\mathrm{static}}\\) constraint, which significantly reduces the foreground and background static flow estimation errors, 80% for FS and 94% for BS. However, adding \\(\\mathcal{L}_{\\mathrm{static}}\\) also increases the foreground dynamic error. We attribute this to the difficulty of estimating the flow of moving, geometrically featureless, objects as mentioned previously, which would be reinforced by \\(\\mathcal{L}_{\\mathrm{static}}\\). Even so, considering the two static components in the EPE 3-way metric, the EPE 3-way would still benefit considerably from this loss (15% error reduction). Finally, in experiment 4, we incorporate the \\(\\mathcal{L}_{\\mathrm{dcls}}\\). This results in a notable decrease in overall EPE 3-way by 33% compared to solely using the chamfer loss (experiment 1). The above experiments illustrate that our method is not a simple stack of losses, but a complementary holistic design." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.787, + 0.674 + ], + "angle": 0, + "content": "Training Dataset Size In the context of robotics and autonomous driving, there are situations where the number of accessible frames is limited. This section evaluates the performance of SeFlow given different amounts of training data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.787, + 0.749 + ], + "angle": 0, + "content": "In this experiment, we denote the entire Argoverse 2 Sensor training set as \\(100\\%\\). Extra unlabeled data can be retrieved from the LiDAR dataset for self-supervised methods. To assess the impact of dataset size, we conduct the same 50 epochs of training for methods using \\(10\\%\\), \\(20\\%\\), \\(50\\%\\), and \\(100\\%\\) of the total data in Tab. 4 and Fig. 4 and evaluate the resulting models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.787, + 0.84 + ], + "angle": 0, + "content": "From Tab. 4, we can observe that with only \\(20\\%\\) or \\(50\\%\\) training data, our SeFlow can already outperform ZeroFlow which uses \\(100\\%\\) or \\(200\\%\\) data. Figure 4 illustrates even more intuitively that our SeFlow can easily outperform both existing supervised (FastFlow3D) and unsupervised (ZeroFlow) approaches with the same amount of data. We attribute the demonstrated data efficiency of our method to the well-designed loss functions, which integrate a dynamic" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.259, + 0.114, + 0.733, + 0.131 + ], + "angle": 0, + "content": "SeFlow: A Self-Supervised Scene Flow Method in Autonomous Driving" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "awareness mapping method into our framework and enable better scene-level comprehension." + }, + { + "type": "table", + "bbox": [ + 0.226, + 0.201, + 0.499, + 0.372 + ], + "angle": 0, + "content": "
DatasetEPE ↓
3-wayFDFSBS
10%0.0940.2340.0400.006
20%0.0780.1970.0320.004
50%0.0660.1670.0280.004
100%0.0590.1470.0260.004
ZF 100%0.0880.2310.0220.011
ZF 200%0.0760.1980.0180.011
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.381, + 0.504, + 0.493 + ], + "angle": 0, + "content": "Table 4: Ablation study of dataset size compared with Zeroflow (ZF), another self-supervised learning method. Results are evaluated on the Argoverse 2 validation set with 50 training epochs. The total dataset \\((100\\%)\\) is \\(107\\mathrm{k}\\) frames. \\(200\\%\\) data means all of the Sensor dataset plus an equal amount of the LiDAR dataset (214k frames)." + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.201, + 0.784, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.52, + 0.395, + 0.789, + 0.492 + ], + "angle": 0, + "content": "Fig.4: The relationship between flow estimation error and training dataset size, scaled in \\(\\log_{10}\\). Methods with \\* are supervised by ground truth labels. SeFlow uses less data but gets comparable results compared to FastFlow3D and ZeroFlow." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.524, + 0.421, + 0.539 + ], + "angle": 0, + "content": "5.4 Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.552, + 0.788, + 0.612 + ], + "angle": 0, + "content": "Figure 5 presents a qualitative flow estimation result on a sequence of scenes in the Argoverse 2 validation set. The first three columns showcase the same scene at different timestamps, and the last column shows a zoomed-in view of the scene." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.788, + 0.81 + ], + "angle": 0, + "content": "In the first column featuring a vehicle partially occluded by a traffic pole, SeFlow demonstrates superior flow estimation capabilities compared to ZeroFlow. Sometimes, SeFlow can even detect flow overlooked by the ground truth annotations. Argoverse 2 derives ground truth flow from manual instance level labels, and as a result, the flow of points outside the bounding box may be ignored. This issue is particularly pronounced in smaller objects, and an example of this can be seen in the fourth column where we zoom in on a case. In this example, a pedestrian is pushing a shopping cart across the road. By comparing the first and the third column, we can observe that both objects are moving. However, the ground truth labels zero flow (white) for the shopping cart. SeFlow, on the other hand, successfully predicts consistent flow even at this small scale. In comparison, the baseline method ZeroFlow suffers at flow prediction of small objects and regards both the pedestrian and the shopping cart as static (white)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Although our method shows superior scene flow estimation compared to other baseline methods, it also has some limitations. Based on the zoom-in view in the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Q Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.144, + 0.788, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.381, + 0.79, + 0.452 + ], + "angle": 0, + "content": "Fig. 5: Qualitative results from Argoverse 2 validation set. The top row displays the ground truth flow, the middle row presents the SeFlow result, and the bottom row showcases another self-supervised method ZeroFlow [29] result. Different color indicates different directions and more saturated color means larger flow estimation. Ego motion is compensated for a clearer view." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.484, + 0.79, + 0.591 + ], + "angle": 0, + "content": "fourth column in Fig. 5, we can find some purple flow labels at the top right corner in the first row (Ground Truth). Both SeFlow (Ours) and ZeroFlow fail to predict any flow for these points. One possible reason is that the point cloud of distant objects is sparse and easily ignored by both the voxelization process and clustering algorithms. This makes it difficult to estimate the flow of distant objects using only two consecutive frames. Based on this, multi-modality or time-consistent networks would be a further direction for future research." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.62, + 0.371, + 0.636 + ], + "angle": 0, + "content": "6 Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.657, + 0.788, + 0.779 + ], + "angle": 0, + "content": "In this paper, we proposed SeFlow, an efficient and effective self-supervised training method for scene flow estimation in autonomous driving with large-scale point clouds as input. Our primary contributions include a learning method that integrates static and dynamic awareness to construct self-supervision objectives. We further identify problems with the correspondence assumptions of Chamfer-based loss functions commonly used for self-supervised learning, and mitigate these with a constraint based on the upper bound of object motion. Our experimental results underscore the efficacy of our approach." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Future research may concentrate on integrating multi-modality (e.g., cameras and radar) within the SeFlow framework for higher flow estimation accuracy or embedding temporal coherence concepts within our pipeline. Additionally, the optimization of multi-loss learning weights warrants further exploration." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.259, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SeFlow: A Self-Supervised Scene Flow Method in Autonomous Driving" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.311 + ], + "angle": 0, + "content": "Thanks to RPL member: Li Ling helps review this manuscript. Thanks to Kyle Vedder (the author of ZeroFlow), who kindly opened his code including pretrained weights, and discussed their result with us which helped this work a lot. This work was partially supported by the Wallenberg AI, Autonomous Systems and Software Program (WASP) funded by the Knut and Alice Wallenberg Foundation and Prosense (2020-02963) funded by Vinnova. The computations were enabled by the supercomputing resource Berzelius provided by National Supercomputer Centre at Linköping University and the Knut and Alice Wallenberg Foundation, Sweden." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.332, + 0.323, + 0.348 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.361, + 0.786, + 0.388 + ], + "angle": 0, + "content": "1. 2, A.: Argoverse 2 scene flow online leaderboard. https://eval.ai/web/challenges/challenge-page/2010/leaderboard/4759 (2024 Mar 4th)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.39, + 0.786, + 0.43 + ], + "angle": 0, + "content": "2. Aleotti, F., Poggi, M., Tosi, F., Mattoccia, S.: Learning end-to-end scene flow by distilling single tasks knowledge. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 10435-10442 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.431, + 0.786, + 0.484 + ], + "angle": 0, + "content": "3. Baur, S.A., Emmerichs, D.J., Moosmann, F., Pinggera, P., Ommer, B., Geiger, A.: Slim: Self-supervised lidar scene flow and motion segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13126-13136 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.486, + 0.786, + 0.525 + ], + "angle": 0, + "content": "4. Campello, R.J., Moulavi, D., Sander, J.: Density-based clustering based on hierarchical density estimates. In: Pacific-Asia conference on knowledge discovery and data mining. pp. 160-172. Springer (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.527, + 0.786, + 0.567 + ], + "angle": 0, + "content": "5. Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., et al.: Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.568, + 0.786, + 0.608 + ], + "angle": 0, + "content": "6. Cho, K., Van Merrienboer, B., Gulcehre, C., Bahdanau, D., Bougares, F., Schwenk, H., Bengio, Y.: Learning phrase representations using rnn encoder-decoder for statistical machine translation. arXiv preprint arXiv:1406.1078 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.61, + 0.786, + 0.636 + ], + "angle": 0, + "content": "7. Chodosh, N., Ramanan, D., Lucey, S.: Re-evaluating lidar scene flow for autonomous driving. arXiv preprint arXiv:2304.02150 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.637, + 0.786, + 0.676 + ], + "angle": 0, + "content": "8. Deng, D., Zakhor, A.: Rsf: Optimizing rigid scene flow from 3d point clouds without labels. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1277-1286 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.677, + 0.786, + 0.717 + ], + "angle": 0, + "content": "9. Duberg, D., Zhang, Q., Jia, M., Jensfelt, P.: DUFOMap: Efficient dynamic awareness mapping. IEEE Robotics and Automation Letters 9(6), 5038-5045 (2024). https://doi.org/10.1109/LRA.2024.3387658" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.718, + 0.786, + 0.758 + ], + "angle": 0, + "content": "10. Himmelsbach, M., Hundelshausen, F.V., Wuensche, H.J.: Fast segmentation of 3d point clouds for ground vehicles. In: Intelligent Vehicles Symposium (IV), 2010 IEEE. pp. 560-565. IEEE (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.759, + 0.786, + 0.799 + ], + "angle": 0, + "content": "11. Jund, P., Sweeney, C., Abdo, N., Chen, Z., Shlens, J.: Scalable scene flow from point clouds in the real world. IEEE Robotics and Automation Letters 7(2), 1589-1596 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.8, + 0.786, + 0.84 + ], + "angle": 0, + "content": "12. Kittenplon, Y., Eldar, Y.C., Raviv, D.: Flowstep3d: Model unrolling for self-supervised scene flow estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4114-4123 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.219, + 0.361, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Q Zhang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "13. Lang, I., Aiger, D., Cole, F., Avidan, S., Rubinstein, M.: Scoop: Self-supervised correspondence and optimization-based scene flow. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5281-5290 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.217 + ], + "angle": 0, + "content": "14. Li, X., Kaesemodel Pontes, J., Lucey, S.: Neural scene flow prior. Advances in Neural Information Processing Systems 34, 7838-7851 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.217, + 0.785, + 0.257 + ], + "angle": 0, + "content": "15. Li, X., Zheng, J., Ferroni, F., Pontes, J.K., Lucey, S.: Fast neural scene flow. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9878-9890 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.257, + 0.785, + 0.298 + ], + "angle": 0, + "content": "16. Liu, J., Wang, G., Ye, W., Jiang, C., Han, J., Liu, Z., Zhang, G., Du, D., Wang, H.: Difflow3d: Toward robust uncertainty-aware scene flow estimation with diffusion model. arXiv preprint arXiv:2311.17456 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.298, + 0.785, + 0.352 + ], + "angle": 0, + "content": "17. Mayer, N., Ilg, E., Hausser, P., Fischer, P., Cremers, D., Dosovitskiy, A., Brox, T.: A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4040-4048 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.352, + 0.785, + 0.393 + ], + "angle": 0, + "content": "18. Menze, M., Geiger, A.: Object scene flow for autonomous vehicles. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3061-3070 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.393, + 0.785, + 0.433 + ], + "angle": 0, + "content": "19. Mittal, H., Okorn, B., Held, D.: Just go with the flow: Self-supervised scene flow estimation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11177-11185 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.433, + 0.785, + 0.473 + ], + "angle": 0, + "content": "20. Najibi, M., Ji, J., Zhou, Y., Qi, C.R., Yan, X., Ettinger, S., Anguelov, D.: Motion inspired unsupervised perception and prediction in autonomous driving. In: European Conference on Computer Vision. pp. 424-443. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.473, + 0.785, + 0.528 + ], + "angle": 0, + "content": "21. Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A., Cournapeau, D., Brucher, M., Perrot, M., Duchesnay, E.: Scikit-learn: Machine learning in Python. Journal of Machine Learning Research 12, 2825–2830 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.528, + 0.785, + 0.583 + ], + "angle": 0, + "content": "22. Pfreundschuh, P., Hendrikx, H.F., Reijgwart, V., Dubé, R., Siegwart, R., Cramar-iuc, A.: Dynamic object aware lidar slam based on automatic generation of training data. In: 2021 IEEE International Conference on Robotics and Automation (ICRA). pp. 11641-11647. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.583, + 0.785, + 0.637 + ], + "angle": 0, + "content": "23. Schmid, L., Andersson, O., Sulser, A., Pfreundschuh, P., Siegwart, R.: Dynablox: Real-time detection of diverse dynamic objects in complex environments. IEEE Robotics and Automation Letters (RA-L) 8(10), 6259 - 6266 (2023). https://doi.org/10.1109/LRA.2023.3305239" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.637, + 0.785, + 0.678 + ], + "angle": 0, + "content": "24. Shen, Y., Hui, L., Xie, J., Yang, J.: Self-supervised 3d scene flow estimation guided by superpoints. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5271-5280 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.679, + 0.785, + 0.704 + ], + "angle": 0, + "content": "25. Song, J., Lee, S.J.: Knowledge distillation of multi-scale dense prediction transformer for self-supervised depth estimation. Scientific Reports (18939) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.704, + 0.785, + 0.759 + ], + "angle": 0, + "content": "26. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., et al.: Scalability in perception for autonomous driving: Waymo open dataset. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2446-2454 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.759, + 0.785, + 0.8 + ], + "angle": 0, + "content": "27. Tishchenko, I., Lombardi, S., Oswald, M.R., Pollefeys, M.: Self-supervised learning of non-rigid residual flow and ego-motion. In: 2020 international conference on 3D vision (3DV). pp. 150-159. IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "28. Vedder, K., Khatri, I., Peri, N., Chodosh, N., Liu, Y., Hays, J.: Av2 2024 scene flow challenge announcement. https://wwwargonse.org/sceneflow.html (2024 Feb 25th)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.26, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SeFlow: A Self-Supervised Scene Flow Method in Autonomous Driving" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "29. Vedder, K., Peri, N., Chodosh, N., Khatri, I., Eaton, E., Jayaraman, D., Ramanan, Y.L.D., Hays, J.: ZeroFlow: Fast Zero Label Scene Flow via Distillation. International Conference on Learning Representations (ICLR) (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "30. Vedula, S., Rander, P., Collins, R., Kanade, T.: Three-dimensional scene flow. IEEE transactions on pattern analysis and machine intelligence 27(3), 475-480 (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.259 + ], + "angle": 0, + "content": "31. Vidanapathirana, K., Chng, S.F., Li, X., Lucey, S.: Multi-body neural scene flow. In: 2024 International Conference on 3D Vision (3DV). pp. 126-136. IEEE (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.785, + 0.3 + ], + "angle": 0, + "content": "32. Wang, Z., Wei, Y., Rao, Y., Zhou, J., Lu, J.: 3d point-voxel correlation fields for scene flow estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.785, + 0.327 + ], + "angle": 0, + "content": "33. Wei, Y., Wang, Z., Rao, Y., Lu, J., Zhou, J.: PV-RAFT: Point-Voxel Correlation Fields for Scene Flow Estimation of Point Clouds. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.383 + ], + "angle": 0, + "content": "34. Wilson, B., Qi, W., Agarwal, T., Lambert, J., Singh, J., et al.: Argoverse 2: Next generation datasets for self-driving perception and forecasting. In: Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks (NeurIPS Datasets and Benchmarks 2021) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.411 + ], + "angle": 0, + "content": "35. Wu, H., Li, Y., Xu, W., Kong, F., Zhang, F.: Moving event detection from lidar point streams. Nature Communications 15(1), 345 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.785, + 0.467 + ], + "angle": 0, + "content": "36. Wu, W., Wang, Z.Y., Li, Z., Liu, W., Fuxin, L.: Pointpwc-net: Cost volume on point clouds for self-supervised scene flow estimation. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16. pp. 88-107. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.508 + ], + "angle": 0, + "content": "37. Zhang, Q., Duberg, D., Geng, R., Jia, M., Wang, L., Jensfelt, P.: A dynamic points removal benchmark in point cloud maps. In: IEEE 26th International Conference on Intelligent Transportation Systems (ITSC). pp. 608-614 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.508, + 0.785, + 0.536 + ], + "angle": 0, + "content": "38. Zhang, Q., Yang, Y., Fang, H., Geng, R., Jensfelt, P.: Deflow: Decoder of scene flow network in autonomous driving. arXiv preprint arXiv:2401.16122 (2024)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.536 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/SeFlow_ A Self-Supervised Scene Flow Method in Autonomous Driving/e7db8302-f2d3-4bc6-8b61-d5de1eb958ee_origin.pdf b/2024/SeFlow_ A Self-Supervised Scene Flow Method in Autonomous Driving/e7db8302-f2d3-4bc6-8b61-d5de1eb958ee_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..51d9f845a1f78658630c6915f3be07c8f3e169a1 --- /dev/null +++ b/2024/SeFlow_ A Self-Supervised Scene Flow Method in Autonomous Driving/e7db8302-f2d3-4bc6-8b61-d5de1eb958ee_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9a6d9311cde66d9fec5634528dcd7e5c6eb8bd8615c1d81a9daaf825aaea4144 +size 4149077 diff --git a/2024/SeFlow_ A Self-Supervised Scene Flow Method in Autonomous Driving/full.md b/2024/SeFlow_ A Self-Supervised Scene Flow Method in Autonomous Driving/full.md new file mode 100644 index 0000000000000000000000000000000000000000..bee74da627477c572f6a9400d1297818977bc3ca --- /dev/null +++ b/2024/SeFlow_ A Self-Supervised Scene Flow Method in Autonomous Driving/full.md @@ -0,0 +1,313 @@ +# SeFlow: A Self-Supervised Scene Flow Method in Autonomous Driving + +Qingwen Zhang $^{1}$ , Yi Yang $^{1,2}$ , Peizheng Li $^{3,4}$ , Olov Andersson $^{1}$ , and Patric Jensfelt $^{1}$ + +$^{1}$ RPL, KTH Royal Institute of Technology, Stockholm, Sweden + $^{2}$ Scania CV AB, Södertälje, Sweden + +3 University of Tübingen, Tübingen, Germany + +4 Mercedes-Benz AG, Sindelfingen, Germany {qingwen,yiya,olovand,patric}@kth.se, + +carol-yi.yang@scania.com, peizheng.li@mercedes-benz.com + +Abstract. Scene flow estimation predicts the 3D motion at each point in successive LiDAR scans. This detailed, point-level, information can help autonomous vehicles to accurately predict and understand dynamic changes in their surroundings. Current state-of-the-art methods require annotated data to train scene flow networks and the expense of labeling inherently limits their scalability. Self-supervised approaches can overcome the above limitations, yet face two principal challenges that hinder optimal performance: point distribution imbalance and disregard for object-level motion constraints. In this paper, we propose SeFlow, a self-supervised method that integrates efficient dynamic classification into a learning-based scene flow pipeline. We demonstrate that classifying static and dynamic points helps design targeted objective functions for different motion patterns. We also emphasize the importance of internal cluster consistency and correct object point association to refine the scene flow estimation, in particular on object details. Our real-time capable method achieves state-of-the-art performance on the self-supervised scene flow task on Argoverse 2 and Waymo datasets. The code is open-sourced at https://github.com/KTH-RPL/SeFlow. + +Keywords: 3D scene flow, self-supervised, autonomous driving, large-scale point cloud + +# 1 Introduction + +Scene flow [30] captures the 3D velocity at every point in a point cloud. These detailed 3D flow estimates can enhance downstream tasks in autonomous driving, such as detection, segmentation, tracking, and occupancy flow estimation [20]. A common paradigm for addressing the scene flow problem is supervised learning by utilizing annotated LiDAR data [11, 38]. However, expensive labeling inherently limits the scalability of supervised learning methods. + +Given the difficulty and expense of labeling scene flow ground truth, we instead focus on self-supervised scene flow approaches. Existing self-supervised + +![](images/b734f2cbf7a5cfbf504f0c2e94d336bb71f4313459c39f351b024c3efa560eee.jpg) +Fig. 1: LiDAR scene flow estimation using our SeFlow method on Argoverse 2. The predicted scene flow for each point is color-coded based on direction. The white indicates static points whose flow is zero. More saturated colors indicate higher velocities. (a) Camera view for visualization purposes only. (b),(c) are zoomed-in views showing the baseline from ZeroFlow [29] as well as SeFlow (ours). When predicting the flow of a large and long vehicle, the baseline predicts a portion of the flow as 0, whereas our estimates are consistent. In addition, the baseline tends to ignore small-scale objects, e.g., pedestrians, while our method can better handle such small and slow-moving objects. + +methods can be divided into two categories: knowledge distillation and data exploration. Knowledge distillation methods [2,25,29] typically rely on a 'teacher' method to generate pseudo flow labels. These pseudo labels are then used to supervise student models. Data exploration methods [3, 14, 15, 19, 31], on the other hand, directly utilize the predicted flow to project the input points and find nearest neighbors in the next frame to establish constraints for training. + +A key challenge for self-supervised methods is that the vast majority of the points are static (about $86\%$ of points are background [7, 34]). This data imbalance often leads to overly conservative scene flow predictions. One common way to counter this is to use large amounts of training data as in ZeroFlow [29]. ZeroFlow experimentally shows that more data helps to improve the dynamic flow accuracy. In this paper, we investigate ways to improve data efficiency and take inspiration from supervised methods [11, 38]. We address the data imbalance by first classifying points as dynamic or static based on traditional ray casting when integrating frames [9]. This allows us to constrain the corresponding motion patterns by proposing novel loss functions. + +Another shortcoming of existing self-supervised scene flow methods is that they disregard object-level motion constraints. The flow should be consistent, i.e., all points in a rigid object should have similar flows. A clear case of inconsistent flow can be seen on the big truck in Fig. 1, where ZeroFlow [29] predicts an absence of flow in certain sections. This inconsistency is caused by incorrect + +object point associations which also can be observed in small-scale objects, as shown in the case of the pedestrian. To account for object-level motion constraints, we propose to cluster the dynamic points into object candidates and encourage consistent flow and correct associations for the points in each cluster. This reduces the fragmentation of the flow estimate for points inside the same object. + +Overall, our method improves the estimated flow accuracy and addresses issues in previous methods as shown in Fig. 1(c). We propose an efficient and effective self-supervised scene flow method that integrates traditional classification and learning-based strategies. Our approach is open-source at https://github.com/KTH-RPL/SeFlow. Our primary contributions include: + +- We propose SeFlow, a novel method that integrates a dynamic classification method in formulating efficient self-supervision objectives. +- We further construct loss functions to learn dynamic flow estimation in imbalanced data and ensure consistent object-level flow, mitigating the effects of correspondence errors. +- We show that SeFlow achieves state-of-the-art results on the self-supervised scene flow task in Argoverse 2 and Waymo datasets and even outperforms all but one supervised method on the leaderboard. + +# 2 Related Work + +In this section, we explore existing works in scene flow estimation. We also discuss current traditional frameworks capable of classifying dynamic points, highlighting their relevance and application in the context of scene flow tasks. + +# 2.1 Scene Flow Estimation + +Scene flow estimation in autonomous driving is slightly different from flow estimation in object registration. Methods in object registration [8, 13, 16, 24, 32, 33] focus on relatively small-scale point cloud data like synthetic datasets Shapenet [5] and FlyingThing3D [17]. These methods scale poorly with the number of points. When applied to point cloud data for autonomous driving, they require downsampling to 8,192 points or less [11, 18]. In recent datasets like Argoverse 2 [34] and Waymo [11], the number of points in one full frame is around $80\mathrm{k - }177\mathrm{k}$ . Methods that can handle such large-scale point cloud data as input usually use a voxel-based pipeline [11, 38]. However, expensive labeling of ground truth flow limits the scalability of these supervised methods, especially in autonomous driving where we have continuously increasing amounts of potential training data. + +To train models without labeled data, recent methods propose self-supervised losses. Many commonly used losses, such as Chamfer distance, are based on the nearest neighbor distance between two point cloud inputs [3,12,14,15,19,27,36]. However, a major limitation of nearest neighbor based losses is that only part + +of the points on dynamic objects provides good supervision due to incorrect associations. This is especially apparent when big trucks are moving fast in autonomous driving scenarios (see Fig. 1 and Fig. 3). Mittal et al. [19] define a self-supervised loss by tracking a patch forward and backward in time to form a cycle while penalizing the errors through cycle consistency and feature similarity. Baur et al. [3] propose three loss functions with k-NN loss, rigid cycle consistency inspired by Mittal, and artificial labels based on the distance between points in the original and estimated point clouds. In the following we describe NSFP [14], FastNSF [15], and ZeroFlow [29] in more detail. These are the publically available methods that perform best on the Argoverse 2 self-supervised scene flow challenge and are therefore used as our baselines. + +NSFP [14] is an optimization-based method. For each pair of consecutive input point clouds, NSFP iteratively learns new weights for an MLP network to predict the flow using the Chamfer distance loss. However, thousands of iterations are needed and their runtimes extend from 26 to 35 seconds per frame, which fails to meet the real-time requirements of autonomous driving. FastNSF [15] improves the efficiency by voxelizing the point cloud first and then converting it to a distance transform map for faster neighbor calculation. This reduces the runtime to 0.5 seconds, at the expense of increased error. + +ZeroFlow [29] adopts a semi-supervised strategy. Pseudo-flow labels are created offline using NSFP [14], and a FastFlow3D [11] model is used as a student for knowledge distillation. This setup allows the student model to perform real-time inference. In this case, the teacher network needs significant resources (around 3.6 GPU months [29]) to label the entire training data. The final accuracy is also influenced by the performance of the teacher. + +To increase data efficiency and address the limitations of the above methods, we propose to integrate efficient ray-casting-based dynamic awareness mapping into our pipeline. The core idea is to classify which points move and cluster these points into objects for which we can estimate group-level motion statistics and define better self-supervised loss functions. + +# 2.2 Dynamic Awareness in Mapping + +In the field of Simultaneous Localization and Mapping (SLAM), there is a growing interest in dynamic awareness [9,22,23,35,37]. This interest stems from the fact that points associated with moving objects in a scene can significantly impact localization and planning performance. Existing dynamic awareness methods in mapping [9,23,37] often utilize ray casting techniques to binary classify points as either dynamic or static. + +The scene flow task, on the other hand, aims to predict the specific 3D flow at each point, a goal that extends beyond the mere categorization of dynamic and static points. In mapping, a point is considered dynamic if it moves once within a scene (even if it becomes static later), whereas, in scene flow tasks, a point is deemed dynamic if it moves faster than a certain velocity threshold in the current frame. Despite these differences, insights from the mapping field are invaluable. By integrating information over time, these frameworks develop a + +more comprehensive understanding of the entire scene. Such scene level dynamic awareness can inform and enhance our exploration of point cloud data in scene flow tasks. + +# 3 Problem Statement + +This work addresses the problem of real-time scene flow estimation in autonomous driving. Given two consecutive input point clouds, $\mathcal{P}_t$ and $\mathcal{P}_{t + 1}$ , along with the ego movement $\mathbf{T}_{t,t + 1} \in SE(3)$ , the goal is to predict the motion vector as flow $\hat{\mathcal{F}}_{t,t + 1}(p) = (x,y,z)^T$ for each point $p \in \mathcal{P}_t$ . + +The objective is to minimize the End Point Error (EPE) which represents the difference between the predicted flow and the ground truth flow, as expressed by the following equation: + +$$ +\min \underbrace {\frac {1}{| \mathcal {P} _ {t} |} \sum_ {p \in \mathcal {P} _ {t}} \left\| \hat {\mathcal {F}} (p) - \mathcal {F} _ {g t} (p) \right\| _ {2}} _ {\text {E P E}}, \tag {1} +$$ + +where $|\mathcal{P}_t|$ denotes the cardinality of (i.e., the number of points in) $\mathcal{P}_t$ . For consistency, in the subsequent presentation, capital symbols correspond to sets, while the lowercase symbols represent variables of specific points. + +# 4 Method + +# 4.1 Input and Output + +The first step to process the input data, as is commonly done, is to remove ground points from $\mathcal{P}_t$ and $\mathcal{P}_{t + 1}$ . This is typically done using HD maps [26,34] or ground segmentation techniques [10]. + +The estimated flow $\tilde{\mathcal{F}}$ from $\mathcal{P}_t$ to $\mathcal{P}_{t + 1}$ can be decomposed into two parts as follows: + +$$ +\hat {\mathcal {F}} = \mathcal {F} _ {e g o} + \Delta \hat {\mathcal {F}}, \tag {2} +$$ + +where $\mathcal{F}_{ego}$ is the flow resulting from the ego vehicle's motion which can be directly obtained from $\mathbf{T}_{t,t + 1}$ , and $\varDelta\hat{\mathcal{F}}$ is our network output. + +# 4.2 Model Backbone + +To enable real-time computation and estimation of scene flow across a large point set, voxelization is considered a practical encoding strategy in the model backbone. However, the reduction in resolution often leads to a poorer distinction of point features within the same voxel. With this in mind, we use DeFlow [38] as the architectural basis in this paper. DeFlow integrates GRU [6] with iterative refinement in the decoder design. More specifically, the GRU module maintains voxel features as its hidden state and selectively forgets or updates + +![](images/4b64c4f073be133b622aa2a3dc9f8f5b242e5240fa56a253c64412b99e17f8c6.jpg) +Training Stage Sec. 4.4 Self-supervised Loss (including Sec. 4.3 Dynamic Classification) + +![](images/9175ccddfa2f4ff0fb904a45fc31a3f448aae906454462b96f0708da665eea7b.jpg) +Fig. 2: SeFlow Architecture. Top: With two consecutive point clouds as inputs, our model predicts the estimated flows of all points. Bottom: Conceptual visualization of the Chamfer loss and the three proposed training losses. With the original input $\mathcal{P}_t$ (2 static points for the building, and 2 dynamic points for the car) plus the estimated flow $\hat{\mathcal{F}}_t$ , we can calculate the error between estimated $\hat{\mathcal{P}}_{t+1}$ and the next frame point cloud $\mathcal{P}_{t+1}$ ( $\mathcal{L}_{\mathrm{cham}}$ ). The second part is $\mathcal{L}_{\mathrm{dcham}}$ that only calculates the distance error between dynamic points. The third loss says that the estimated flows of static points should be zero. Finally, we assume that the flow at points from the same cluster should be consistent, and mitigate underestimation by using the proposed upper bound on the flow. + +the information of the hidden state during each iteration according to the input point features. After multiple iterations, the optimized voxel features are concatenated with the original point features to obtain the final individual point features. Benefiting from this design, we improve the inference efficiency compared to the commonly used backbone FastFlow3D [11] without sacrificing the accuracy of scene flow estimation in coarse resolution settings. + +# 4.3 Dynamic Classification of Points + +To facilitate dynamic classification of points during the training stage, we incorporate the DUFOMap framework, a mapping-based dynamic awareness approach [9]. The key insight of this is that points observed inside a region that at one time has been observed as empty must be dynamic. Built on this insight, DUFOMap utilizes ray-casting to classify dynamic points at the sensor rate on the CPU. The result is two disjoint sets, $\mathcal{P}_{t,d}$ and $\mathcal{P}_{t,s}$ , where $\mathcal{P}_{t,d}$ is the set of dynamic points that have moved once inside a scene (even if they later become static) and $\mathcal{P}_{t,s}$ is the set of static points that did not move at any time. + +Note that the dynamic-static classification is separated from the inference and training, offering our method good flexibility. + +# 4.4 Self-supervised Loss + +As discussed, self-supervised learning with only Chamfer distance loss is vulnerable to problems with data imbalance and incorrect associations. We therefore construct three additional loss functions, illustrated in Fig. 2 and described in turn below, to mitigate problems with data imbalance as well as encourage consistent object-level flow. + +Chamfer Distance The Chamfer distance, a common self-supervised loss in existing work [3, 14, 15, 19], has the following definition: + +$$ +\mathcal {L} _ {\text {c h a m}} = \frac {1}{| \hat {\mathcal {P}} _ {t + 1} |} \sum_ {p \in \hat {\mathcal {P}} _ {t + 1}} \mathrm {S} (p, \mathcal {P} _ {t + 1}) + \frac {1}{| \mathcal {P} _ {t + 1} |} \sum_ {p \in \mathcal {P} _ {t + 1}} \mathrm {S} (p, \hat {\mathcal {P}} _ {t + 1}) \tag {3} +$$ + +$$ +\mathrm {S} (p, \mathcal {P} _ {o}) = \min _ {p _ {i} \in \mathcal {P} _ {o}} \| p - p _ {i} \| _ {2} ^ {2}, \tag {4} +$$ + +where, $\hat{\mathcal{P}}_{t + 1} = \mathcal{P}_t + \hat{\mathcal{F}}_t$ , $\mathrm{S}(\cdot) = \mathrm{D}(\cdot)^2$ , $\mathrm{D}(p,\mathcal{P}_o)$ denotes the distance between point $p$ and its nearest neighbor in $\mathcal{P}_o$ . + +The Chamfer distance is proposed to calculate a similarity between two point clouds. However, when using it directly as a supervised signal for scene flow in autonomous driving, there are two issues we note. First, the number of static points is often much larger than that of dynamic points and averaging $S(\cdot)$ over all points leads the training to favor static points, i.e., zero flow estimation. Second, due to erroneous correspondence assumptions, only part of the flows within a dynamic object can be estimated correctly as shown in Fig. 3, where the estimated flows of points in the overlap area are zero. + +To solve these issues, we contribute the following constraints using the dynamic classification information prior to supervising the network. + +Dynamic Chamfer Distance The imbalance in the number of dynamic and static points presents a significant challenge in scene flow estimation, as highlighted in previous works [11,29,38]. Supervised methods, provided with ground truth labels for point velocity or classification, can weight their loss functions to account for this imbalance. For self-supervised learning, we introduce a dynamic Chamfer distance. In this context, 'dynamic' points are identified based on the output from Sec. 4.3. Specifically, this loss, $\mathcal{L}_{\mathrm{dcham}}$ , is only computed on points classified as dynamic in two consecutive point clouds. By exclusively considering dynamic points, this loss captures the nuances of motion in point cloud data. This is defined as: + +$$ +\mathcal {L} _ {\mathrm {d c h a m}} = \frac {1}{| \hat {\mathcal {P}} _ {t + 1 , d} |} \sum_ {p \in \hat {\mathcal {P}} _ {t + 1, d}} \mathrm {S} (p, \mathcal {P} _ {t + 1, d}) + \frac {1}{| \mathcal {P} _ {t + 1 , d} |} \sum_ {p \in \mathcal {P} _ {t + 1, d}} \mathrm {S} (p, \hat {\mathcal {P}} _ {t + 1, d}), \tag {5} +$$ + +where $\hat{\mathcal{P}}_{t + 1,d} = \mathcal{P}_{t,d} + \hat{\mathcal{F}}_{t,d}$ + +![](images/afbec55c88188da0e4b650e4654f135a33a82413da23f76200ca23077c41c692.jpg) +(a) Chamfer distance + +![](images/a08235b6b1185d0e2165915e5d62cf503a554d62f64af9db437ddd1c8914492a.jpg) +(b) Point flow magnitude +Fig. 3: Simple visualization of the shortcomings of using Chamfer distance as a supervisory signal for flow value estimation. The denser color on points in (b) represents higher flow values and white means the point's flow is zero. (a) illustrates how to calculate loss based on Chamfer distance, and (b) shows that the flow results, based on the nearest neighbor principle, can lead to zero flow estimation for the middle of the object. + +Static Flow The standard Chamfer distance rests on a very strong assumption: nearest neighbor matching can find perfect one-to-one correspondences of all points between two frames. However, due to varying observations, the number of points in the two frames differs, leading to inconsistencies and potential mismatches. To deal with the possible mismatches in the static areas, we add a constraint to encourage the model to estimate zero flow for static points. This loss is defined as follows: + +$$ +\mathcal {L} _ {\text {s t a t i c}} = \frac {1}{| \mathcal {P} _ {t , s} |} \sum_ {p \in \mathcal {P} _ {t, s}} \| \Delta \hat {\mathcal {F}} (p) \| _ {2} ^ {2}, \tag {6} +$$ + +where $\varDelta\hat{\mathcal{F}}(p)$ represents the network output flow of point $p$ . + +Dynamic Cluster Flow For dynamic fields, the inconsistencies and correspondence errors are even more complicated. For instance, we observe that nearest neighbor matching often leads to erroneous results, e.g., on large objects undergoing translation, which is very common for moving vehicles in urban street environments as shown in Fig. 1. More specifically, Figure 3 illustrates the issue on a simplified car example. In this case, the flow is parallel to the object's surface, which means that nearest neighbor matching does not provide good supervision in the interior of the surface ( $\mathrm{D} \approx 0$ for the points in the overlap area). This mismatch results in an incorrect local optimum that remains unresolved during the training process as it minimizes the losses based on Chamfer Distance. + +Therefore, we suggest that the motion of all parts within an object (a cluster) should be approximately homogeneous over a short time interval. We use the HDBSCAN [4,21] clustering algorithm to identify different moving objects. This clustering is only applied to dynamic points, thus reducing the computational overhead. + +$$ +\mathcal {C} _ {t, d} = \operatorname {H D B S C A N} \left(\mathcal {P} _ {t, d}\right). \tag {7} +$$ + +Directly constraining the variance of the flow distribution inside a cluster would be a straightforward idea, but does not guarantee the correctness of the obtained mean value after convergence. As we know that nearest-neighbour point correspondence can considerably underestimate flow on geometrically featureless object surfaces, we here instead propose to use an upper bound on the object motion as the supervisory signal. We derive this upper bound, per object cluster, by taking the maximum inter-frame distance of all point correspondences within the cluster. Specifically, we exploit information from the original input point cloud data, i.e., $\mathcal{P}_t$ and $\mathcal{P}_{t + 1}$ . After the dynamic classification and clustering process, we find the index of the point $p_k$ in cluster $c_{i} \in \mathcal{C}_{t,d}$ with the largest distance to its nearest neighbor point in $\mathcal{P}_{t + 1,d}$ , i.e., + +$$ +\kappa_ {i} = \underset {k} {\arg \max } \left\{\mathrm {D} \left(p _ {k}, \mathcal {P} _ {t + 1, d}\right) \mid p _ {k} \in \mathcal {P} _ {c _ {i}} \right\}. \tag {8} +$$ + +We calculate the upper bound, $\tilde{f}_{c_i}$ on the flow for cluster $c_{i}$ as + +$$ +\tilde {f} _ {c _ {i}} = p _ {\kappa_ {i}} ^ {\prime} - p _ {\kappa_ {i}}, \tag {9} +$$ + +where $p_{\kappa_i}'$ is the nearest neighbor of point $p_{\kappa_i}$ in $\mathcal{P}_{t+1,d}$ . We use this to drive the estimated flows of cluster $c_i$ towards $\tilde{f}_{c_i}$ as follows: + +$$ +\mathcal {L} _ {c _ {i}} = \sum_ {p _ {j} \in \mathcal {P} _ {c _ {i}}} | | \hat {f} _ {p _ {j}} - \tilde {f} _ {c _ {i}} | | _ {2} ^ {2}, \tag {10} +$$ + +$$ +\mathcal {L} _ {\mathrm {d c l s}} = \frac {1}{| \mathcal {P} _ {t , d} |} \sum_ {c _ {i} \in \mathcal {C} _ {t, d}} \mathcal {L} _ {c _ {i}}. \tag {11} +$$ + +In conclusion, the final SeFlow loss incorporates all four losses introduced above, + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\text {c h a m}} + \mathcal {L} _ {\text {s t a t i c}} + \mathcal {L} _ {\text {d c h a m}} + \mathcal {L} _ {\text {d c l s}}. \tag {12} +$$ + +The effect of each loss will be evaluated in the ablation study in Sec. 5.3. + +# 5 Experiment + +In this section, we first outline evaluation details and then present quantitative comparisons with state-of-the-art methods on two benchmark datasets. A series of ablation studies are presented to better evaluate the individual components of our approach. Finally, we conclude with qualitative results on Argoverse 2 and discuss limitations of the approach. + +# 5.1 Dataset and Metric + +We briefly introduce the dataset and metrics for evaluation in the following section. More implementation details, such as hyperparameters for training and dataset descriptions, can be found in the supplementary material. + +Table 1: Comparisons on Argoverse 2 test set from the online leaderboard [1]. Upper groups are supervised methods, lower groups are self-supervised methods. Our method achieves state-of-art performance in the self-supervised scene flow task. $\dagger$ means methods can run in real-time (10 Hz) onboard. + +
MethodRun Time per frame [ms]EPE ↓
3-wayFDFSBS
FastFlow3D†[11]34 ± 50.07820.20720.02530.0020
DeFlow†[38]48 ± 40.05340.13400.02320.0029
FastNSF [15]507 ± 3120.16570.35400.04060.1025
NSFP [14]32,060 ± 10,1120.06850.15030.03020.0248
ZeroFlow†[29]34 ± 50.08140.21090.02540.0080
SeFlow (Ours)†48 ± 40.06280.15250.03210.0038
+ +Dataset We evaluate our approach on two large-scale autonomous driving datasets: Argoverse 2 [34] and Waymo [26]. Ground removal is performed using HD maps for both datasets as described in [34]. Waymo datasets contain 798 training and 202 validation scenes respectively. We focus our description here on Argoverse 2, which provides official and public scene flow challenges [1,28], with the Sensor and LiDAR datasets. The Sensor dataset encompasses 700 training and 150 validation scenes. Each scene is approximately 15 seconds long in $10\mathrm{Hz}$ with complete annotations for evaluation. The evaluation for another 150 test scenes can be accessed indirectly by submitting a solution to the leaderboard. The LiDAR dataset contains 20,000 scenes without any annotation and is only used as extra data in Sec. 5.3. + +Metric The benchmark follows existing works [1,7,29,38] and uses the three-way End Point Error. End Point Error (EPE), as defined in Eq. (1), measures the L2 norm of the discrepancy between the predicted and actual flow vectors, expressed in meters. The EPE three-way (3-way) is defined as the unweighted average EPE across three disjoint sets of points: Foreground Dynamic (FD), Foreground Static (FS), and Background Static (BS). The definition of 'dynamic' is as follows: If the flow at a point exceeds the threshold (the public leaderboard setting is $0.05\mathrm{m}$ ), the point is defined as dynamic. Given the $10\mathrm{Hz}$ sensor frequency, this threshold corresponds to a speed of $0.5\mathrm{m/s}$ . All evaluations are limited to points that are within a $100\mathrm{m} \times 100\mathrm{m}$ box centered on the ego vehicle. + +# 5.2 Quantitative Results + +We evaluate the performance of our method SeFlow and compare it with the currently best-performing methods on the Argoverse 2 test set and Waymo validation set. In Tab. 1 and Tab. 2, the upper group, FastFlow3D and DeFlow, are supervised methods trained with ground truth flow. Compared to the supervised + +Table 2: Comparisons on Waymo Open dataset validation set. Upper groups are supervised methods, lower groups are self-supervised methods. Our method achieves state-of-art performance in the self-supervised scene flow task. $\dagger$ means methods can run in real-time (10 Hz) onboard. + +
MethodRun Time per frame [ms]EPE ↓
3-wayFDFSBS
FastFlow3D†[11]27 ± 60.07820.19540.02460.0152
DeFlow†[38]42 ± 40.04460.09800.02590.0098
FastNSF [14]593 ± 3080.15790.30120.01460.0403
NSFP [15]76,163 ± 32,2560.10050.17120.10810.0221
ZeroFlow†[29]27 ± 60.09210.21620.01530.0241
SeFlow (Ours)†42 ± 40.05980.15060.01810.0106
+ +methods, we note that SeFlow outperforms FastFlow3D and approaches the level of DeFlow in terms of EPE 3-way. This comparative analysis demonstrates the great potential of self-supervised learning in scene flow estimation and validates the effectiveness of our approach. + +In the self-supervised category, our SeFlow method achieves state-of-the-art performance on both datasets. While FastNSF and NSFP are optimization-based methods that do not rely on pre-trained weights for subsequent estimations, their inference times are not conducive to real-time requirements. NSFP, with the second best result in Tab. 1, takes approximately 30 seconds to predict a single frame, which is impractical for real-time autonomous driving applications. FastNSF, on the other hand, enhances efficiency through voxelization for quicker Chamfer distance calculation. Although significantly faster than NSFP, the performance of FastNSFP is the worst among all methods evaluated. + +ZeroFlow is capable of estimating scene flow in real-time, with an accuracy similar to NSFP (slightly better on Waymo and slightly worse on Argoverse 2). SeFlow stands out not only for achieving the best result, but also for drastically reducing the run time to 50 milliseconds, which is more than two orders of magnitude faster than NSFP while providing more accurate flow estimates. SeFlow's state-of-the-art performance demonstrated in Tab. 1 and Tab. 2 underscores the effectiveness of our novel self-supervised method in scene flow estimation. + +# 5.3 Ablation study + +In this section, we delve into two key aspects of our SeFlow pipeline. Firstly, we examine the impact of different loss terms on the accuracy of our flow prediction results. This analysis aims to demonstrate the necessity and effectiveness of the loss components we propose. Secondly, we explore how the size of the training dataset, especially in the case of limited training data, affects the outcomes of our self-supervised training process. + +Table 3: Ablation study of loss terms. Results are evaluated on the Argoverse 2 validation set with 20 training epochs. + +
Exp. IdLchamLdchamLstaticLdclsEPE ↓
3-wayFDFSBS
1✓0.09620.2030.0520.033
2✓✓0.09160.1810.0590.035
3✓✓✓0.07790.2200.0120.002
4✓✓✓✓0.06430.1600.0290.004
+ +Loss Terms The advantages of each loss design are evident in Tab. 3 evaluated on the Argoverse 2 validation set with 20 training epochs. Instead of relying solely on the chamfer loss $\mathcal{L}_{\mathrm{cham}}$ , we investigate how incorporating the additional three losses ( $\mathcal{L}_{\mathrm{static}}$ , $\mathcal{L}_{\mathrm{dcham}}$ , $\mathcal{L}_{\mathrm{dcls}}$ ) can boost the performance. + +After adding the dynamic Chamfer loss $\mathcal{L}_{\mathrm{dcham}}$ , experiment 2 shows a decrease in flow estimation error of about 0.22 (10%) for foreground dynamics (FD), while the static errors (FS, BS) remain essentially the same. Experiment 3 then incorporates $\mathcal{L}_{\mathrm{static}}$ constraint, which significantly reduces the foreground and background static flow estimation errors, 80% for FS and 94% for BS. However, adding $\mathcal{L}_{\mathrm{static}}$ also increases the foreground dynamic error. We attribute this to the difficulty of estimating the flow of moving, geometrically featureless, objects as mentioned previously, which would be reinforced by $\mathcal{L}_{\mathrm{static}}$ . Even so, considering the two static components in the EPE 3-way metric, the EPE 3-way would still benefit considerably from this loss (15% error reduction). Finally, in experiment 4, we incorporate the $\mathcal{L}_{\mathrm{dcls}}$ . This results in a notable decrease in overall EPE 3-way by 33% compared to solely using the chamfer loss (experiment 1). The above experiments illustrate that our method is not a simple stack of losses, but a complementary holistic design. + +Training Dataset Size In the context of robotics and autonomous driving, there are situations where the number of accessible frames is limited. This section evaluates the performance of SeFlow given different amounts of training data. + +In this experiment, we denote the entire Argoverse 2 Sensor training set as $100\%$ . Extra unlabeled data can be retrieved from the LiDAR dataset for self-supervised methods. To assess the impact of dataset size, we conduct the same 50 epochs of training for methods using $10\%$ , $20\%$ , $50\%$ , and $100\%$ of the total data in Tab. 4 and Fig. 4 and evaluate the resulting models. + +From Tab. 4, we can observe that with only $20\%$ or $50\%$ training data, our SeFlow can already outperform ZeroFlow which uses $100\%$ or $200\%$ data. Figure 4 illustrates even more intuitively that our SeFlow can easily outperform both existing supervised (FastFlow3D) and unsupervised (ZeroFlow) approaches with the same amount of data. We attribute the demonstrated data efficiency of our method to the well-designed loss functions, which integrate a dynamic + +awareness mapping method into our framework and enable better scene-level comprehension. + +
DatasetEPE ↓
3-wayFDFSBS
10%0.0940.2340.0400.006
20%0.0780.1970.0320.004
50%0.0660.1670.0280.004
100%0.0590.1470.0260.004
ZF 100%0.0880.2310.0220.011
ZF 200%0.0760.1980.0180.011
+ +Table 4: Ablation study of dataset size compared with Zeroflow (ZF), another self-supervised learning method. Results are evaluated on the Argoverse 2 validation set with 50 training epochs. The total dataset $(100\%)$ is $107\mathrm{k}$ frames. $200\%$ data means all of the Sensor dataset plus an equal amount of the LiDAR dataset (214k frames). + +![](images/f331b8fda3d355540cfbc7e315602ca7e50ca3bea6c3270cdc7fc383d3899b8b.jpg) +Fig.4: The relationship between flow estimation error and training dataset size, scaled in $\log_{10}$ . Methods with \* are supervised by ground truth labels. SeFlow uses less data but gets comparable results compared to FastFlow3D and ZeroFlow. + +# 5.4 Qualitative Results + +Figure 5 presents a qualitative flow estimation result on a sequence of scenes in the Argoverse 2 validation set. The first three columns showcase the same scene at different timestamps, and the last column shows a zoomed-in view of the scene. + +In the first column featuring a vehicle partially occluded by a traffic pole, SeFlow demonstrates superior flow estimation capabilities compared to ZeroFlow. Sometimes, SeFlow can even detect flow overlooked by the ground truth annotations. Argoverse 2 derives ground truth flow from manual instance level labels, and as a result, the flow of points outside the bounding box may be ignored. This issue is particularly pronounced in smaller objects, and an example of this can be seen in the fourth column where we zoom in on a case. In this example, a pedestrian is pushing a shopping cart across the road. By comparing the first and the third column, we can observe that both objects are moving. However, the ground truth labels zero flow (white) for the shopping cart. SeFlow, on the other hand, successfully predicts consistent flow even at this small scale. In comparison, the baseline method ZeroFlow suffers at flow prediction of small objects and regards both the pedestrian and the shopping cart as static (white). + +Although our method shows superior scene flow estimation compared to other baseline methods, it also has some limitations. Based on the zoom-in view in the + +![](images/fa1e6b21bceb26a2908cdc77e14dd6cee8f7d946660a5657001c89dc85546bfc.jpg) +Fig. 5: Qualitative results from Argoverse 2 validation set. The top row displays the ground truth flow, the middle row presents the SeFlow result, and the bottom row showcases another self-supervised method ZeroFlow [29] result. Different color indicates different directions and more saturated color means larger flow estimation. Ego motion is compensated for a clearer view. + +fourth column in Fig. 5, we can find some purple flow labels at the top right corner in the first row (Ground Truth). Both SeFlow (Ours) and ZeroFlow fail to predict any flow for these points. One possible reason is that the point cloud of distant objects is sparse and easily ignored by both the voxelization process and clustering algorithms. This makes it difficult to estimate the flow of distant objects using only two consecutive frames. Based on this, multi-modality or time-consistent networks would be a further direction for future research. + +# 6 Conclusions + +In this paper, we proposed SeFlow, an efficient and effective self-supervised training method for scene flow estimation in autonomous driving with large-scale point clouds as input. Our primary contributions include a learning method that integrates static and dynamic awareness to construct self-supervision objectives. We further identify problems with the correspondence assumptions of Chamfer-based loss functions commonly used for self-supervised learning, and mitigate these with a constraint based on the upper bound of object motion. Our experimental results underscore the efficacy of our approach. + +Future research may concentrate on integrating multi-modality (e.g., cameras and radar) within the SeFlow framework for higher flow estimation accuracy or embedding temporal coherence concepts within our pipeline. Additionally, the optimization of multi-loss learning weights warrants further exploration. + +# Acknowledgement + +Thanks to RPL member: Li Ling helps review this manuscript. Thanks to Kyle Vedder (the author of ZeroFlow), who kindly opened his code including pretrained weights, and discussed their result with us which helped this work a lot. This work was partially supported by the Wallenberg AI, Autonomous Systems and Software Program (WASP) funded by the Knut and Alice Wallenberg Foundation and Prosense (2020-02963) funded by Vinnova. The computations were enabled by the supercomputing resource Berzelius provided by National Supercomputer Centre at Linköping University and the Knut and Alice Wallenberg Foundation, Sweden. + +# References + +1. 2, A.: Argoverse 2 scene flow online leaderboard. https://eval.ai/web/challenges/challenge-page/2010/leaderboard/4759 (2024 Mar 4th) +2. Aleotti, F., Poggi, M., Tosi, F., Mattoccia, S.: Learning end-to-end scene flow by distilling single tasks knowledge. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 10435-10442 (2020) +3. Baur, S.A., Emmerichs, D.J., Moosmann, F., Pinggera, P., Ommer, B., Geiger, A.: Slim: Self-supervised lidar scene flow and motion segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13126-13136 (2021) +4. Campello, R.J., Moulavi, D., Sander, J.: Density-based clustering based on hierarchical density estimates. In: Pacific-Asia conference on knowledge discovery and data mining. pp. 160-172. Springer (2013) +5. Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., et al.: Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012 (2015) +6. Cho, K., Van Merrienboer, B., Gulcehre, C., Bahdanau, D., Bougares, F., Schwenk, H., Bengio, Y.: Learning phrase representations using rnn encoder-decoder for statistical machine translation. arXiv preprint arXiv:1406.1078 (2014) +7. Chodosh, N., Ramanan, D., Lucey, S.: Re-evaluating lidar scene flow for autonomous driving. arXiv preprint arXiv:2304.02150 (2023) +8. Deng, D., Zakhor, A.: Rsf: Optimizing rigid scene flow from 3d point clouds without labels. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1277-1286 (2023) +9. Duberg, D., Zhang, Q., Jia, M., Jensfelt, P.: DUFOMap: Efficient dynamic awareness mapping. IEEE Robotics and Automation Letters 9(6), 5038-5045 (2024). https://doi.org/10.1109/LRA.2024.3387658 +10. Himmelsbach, M., Hundelshausen, F.V., Wuensche, H.J.: Fast segmentation of 3d point clouds for ground vehicles. In: Intelligent Vehicles Symposium (IV), 2010 IEEE. pp. 560-565. IEEE (2010) +11. Jund, P., Sweeney, C., Abdo, N., Chen, Z., Shlens, J.: Scalable scene flow from point clouds in the real world. IEEE Robotics and Automation Letters 7(2), 1589-1596 (2021) +12. Kittenplon, Y., Eldar, Y.C., Raviv, D.: Flowstep3d: Model unrolling for self-supervised scene flow estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4114-4123 (2021) + +13. Lang, I., Aiger, D., Cole, F., Avidan, S., Rubinstein, M.: Scoop: Self-supervised correspondence and optimization-based scene flow. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5281-5290 (2023) +14. Li, X., Kaesemodel Pontes, J., Lucey, S.: Neural scene flow prior. Advances in Neural Information Processing Systems 34, 7838-7851 (2021) +15. Li, X., Zheng, J., Ferroni, F., Pontes, J.K., Lucey, S.: Fast neural scene flow. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9878-9890 (2023) +16. Liu, J., Wang, G., Ye, W., Jiang, C., Han, J., Liu, Z., Zhang, G., Du, D., Wang, H.: Difflow3d: Toward robust uncertainty-aware scene flow estimation with diffusion model. arXiv preprint arXiv:2311.17456 (2023) +17. Mayer, N., Ilg, E., Hausser, P., Fischer, P., Cremers, D., Dosovitskiy, A., Brox, T.: A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4040-4048 (2016) +18. Menze, M., Geiger, A.: Object scene flow for autonomous vehicles. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3061-3070 (2015) +19. Mittal, H., Okorn, B., Held, D.: Just go with the flow: Self-supervised scene flow estimation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11177-11185 (2020) +20. Najibi, M., Ji, J., Zhou, Y., Qi, C.R., Yan, X., Ettinger, S., Anguelov, D.: Motion inspired unsupervised perception and prediction in autonomous driving. In: European Conference on Computer Vision. pp. 424-443. Springer (2022) +21. Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A., Cournapeau, D., Brucher, M., Perrot, M., Duchesnay, E.: Scikit-learn: Machine learning in Python. Journal of Machine Learning Research 12, 2825–2830 (2011) +22. Pfreundschuh, P., Hendrikx, H.F., Reijgwart, V., Dubé, R., Siegwart, R., Cramar-iuc, A.: Dynamic object aware lidar slam based on automatic generation of training data. In: 2021 IEEE International Conference on Robotics and Automation (ICRA). pp. 11641-11647. IEEE (2021) +23. Schmid, L., Andersson, O., Sulser, A., Pfreundschuh, P., Siegwart, R.: Dynablox: Real-time detection of diverse dynamic objects in complex environments. IEEE Robotics and Automation Letters (RA-L) 8(10), 6259 - 6266 (2023). https://doi.org/10.1109/LRA.2023.3305239 +24. Shen, Y., Hui, L., Xie, J., Yang, J.: Self-supervised 3d scene flow estimation guided by superpoints. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5271-5280 (2023) +25. Song, J., Lee, S.J.: Knowledge distillation of multi-scale dense prediction transformer for self-supervised depth estimation. Scientific Reports (18939) (2023) +26. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., et al.: Scalability in perception for autonomous driving: Waymo open dataset. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2446-2454 (2020) +27. Tishchenko, I., Lombardi, S., Oswald, M.R., Pollefeys, M.: Self-supervised learning of non-rigid residual flow and ego-motion. In: 2020 international conference on 3D vision (3DV). pp. 150-159. IEEE (2020) +28. Vedder, K., Khatri, I., Peri, N., Chodosh, N., Liu, Y., Hays, J.: Av2 2024 scene flow challenge announcement. https://wwwargonse.org/sceneflow.html (2024 Feb 25th) + +29. Vedder, K., Peri, N., Chodosh, N., Khatri, I., Eaton, E., Jayaraman, D., Ramanan, Y.L.D., Hays, J.: ZeroFlow: Fast Zero Label Scene Flow via Distillation. International Conference on Learning Representations (ICLR) (2024) +30. Vedula, S., Rander, P., Collins, R., Kanade, T.: Three-dimensional scene flow. IEEE transactions on pattern analysis and machine intelligence 27(3), 475-480 (2005) +31. Vidanapathirana, K., Chng, S.F., Li, X., Lucey, S.: Multi-body neural scene flow. In: 2024 International Conference on 3D Vision (3DV). pp. 126-136. IEEE (2024) +32. Wang, Z., Wei, Y., Rao, Y., Zhou, J., Lu, J.: 3d point-voxel correlation fields for scene flow estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023) +33. Wei, Y., Wang, Z., Rao, Y., Lu, J., Zhou, J.: PV-RAFT: Point-Voxel Correlation Fields for Scene Flow Estimation of Point Clouds. In: CVPR (2021) +34. Wilson, B., Qi, W., Agarwal, T., Lambert, J., Singh, J., et al.: Argoverse 2: Next generation datasets for self-driving perception and forecasting. In: Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks (NeurIPS Datasets and Benchmarks 2021) (2021) +35. Wu, H., Li, Y., Xu, W., Kong, F., Zhang, F.: Moving event detection from lidar point streams. Nature Communications 15(1), 345 (2024) +36. Wu, W., Wang, Z.Y., Li, Z., Liu, W., Fuxin, L.: Pointpwc-net: Cost volume on point clouds for self-supervised scene flow estimation. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16. pp. 88-107. Springer (2020) +37. Zhang, Q., Duberg, D., Geng, R., Jia, M., Wang, L., Jensfelt, P.: A dynamic points removal benchmark in point cloud maps. In: IEEE 26th International Conference on Intelligent Transportation Systems (ITSC). pp. 608-614 (2023) +38. Zhang, Q., Yang, Y., Fang, H., Geng, R., Jensfelt, P.: Deflow: Decoder of scene flow network in autonomous driving. arXiv preprint arXiv:2401.16122 (2024) \ No newline at end of file diff --git a/2024/SeFlow_ A Self-Supervised Scene Flow Method in Autonomous Driving/images.zip b/2024/SeFlow_ A Self-Supervised Scene Flow Method in Autonomous Driving/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..07f5657168935cc2e250d81c86f478b0be089625 --- /dev/null +++ b/2024/SeFlow_ A Self-Supervised Scene Flow Method in Autonomous Driving/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:450f54d34a7fc6d5fce878354ed6ce6dfda69ae6a74f8316387e00506441273a +size 538028 diff --git a/2024/SeFlow_ A Self-Supervised Scene Flow Method in Autonomous Driving/layout.json b/2024/SeFlow_ A Self-Supervised Scene Flow Method in Autonomous Driving/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8529786f91d5a7fd6ce4999ae99665bcad1c8e8d --- /dev/null +++ b/2024/SeFlow_ A Self-Supervised Scene Flow Method in Autonomous Driving/layout.json @@ -0,0 +1,8653 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 173, + 111, + 442, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 111, + 442, + 148 + ], + "spans": [ + { + "bbox": [ + 173, + 111, + 442, + 148 + ], + "type": "text", + "content": "SeFlow: A Self-Supervised Scene Flow Method in Autonomous Driving" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 197, + 167, + 416, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 167, + 416, + 193 + ], + "spans": [ + { + "bbox": [ + 197, + 167, + 416, + 193 + ], + "type": "text", + "content": "Qingwen Zhang" + }, + { + "bbox": [ + 197, + 167, + 416, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 197, + 167, + 416, + 193 + ], + "type": "text", + "content": ", Yi Yang" + }, + { + "bbox": [ + 197, + 167, + 416, + 193 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 197, + 167, + 416, + 193 + ], + "type": "text", + "content": ", Peizheng Li" + }, + { + "bbox": [ + 197, + 167, + 416, + 193 + ], + "type": "inline_equation", + "content": "^{3,4}" + }, + { + "bbox": [ + 197, + 167, + 416, + 193 + ], + "type": "text", + "content": ", Olov Andersson" + }, + { + "bbox": [ + 197, + 167, + 416, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 197, + 167, + 416, + 193 + ], + "type": "text", + "content": ", and Patric Jensfelt" + }, + { + "bbox": [ + 197, + 167, + 416, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 176, + 201, + 438, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 201, + 438, + 223 + ], + "spans": [ + { + "bbox": [ + 176, + 201, + 438, + 223 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 176, + 201, + 438, + 223 + ], + "type": "text", + "content": " RPL, KTH Royal Institute of Technology, Stockholm, Sweden \n" + }, + { + "bbox": [ + 176, + 201, + 438, + 223 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 176, + 201, + 438, + 223 + ], + "type": "text", + "content": " Scania CV AB, Södertälje, Sweden" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 211, + 224, + 402, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 224, + 402, + 236 + ], + "spans": [ + { + "bbox": [ + 211, + 224, + 402, + 236 + ], + "type": "text", + "content": "3 University of Tübingen, Tübingen, Germany" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 213, + 235, + 397, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 213, + 235, + 397, + 258 + ], + "spans": [ + { + "bbox": [ + 213, + 235, + 397, + 258 + ], + "type": "text", + "content": "4 Mercedes-Benz AG, Sindelfingen, Germany {qingwen,yiya,olovand,patric}@kth.se," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 179, + 258, + 435, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 258, + 435, + 268 + ], + "spans": [ + { + "bbox": [ + 179, + 258, + 435, + 268 + ], + "type": "text", + "content": "carol-yi.yang@scania.com, peizheng.li@mercedes-benz.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 296, + 455, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 296, + 455, + 494 + ], + "spans": [ + { + "bbox": [ + 159, + 296, + 455, + 494 + ], + "type": "text", + "content": "Abstract. Scene flow estimation predicts the 3D motion at each point in successive LiDAR scans. This detailed, point-level, information can help autonomous vehicles to accurately predict and understand dynamic changes in their surroundings. Current state-of-the-art methods require annotated data to train scene flow networks and the expense of labeling inherently limits their scalability. Self-supervised approaches can overcome the above limitations, yet face two principal challenges that hinder optimal performance: point distribution imbalance and disregard for object-level motion constraints. In this paper, we propose SeFlow, a self-supervised method that integrates efficient dynamic classification into a learning-based scene flow pipeline. We demonstrate that classifying static and dynamic points helps design targeted objective functions for different motion patterns. We also emphasize the importance of internal cluster consistency and correct object point association to refine the scene flow estimation, in particular on object details. Our real-time capable method achieves state-of-the-art performance on the self-supervised scene flow task on Argoverse 2 and Waymo datasets. The code is open-sourced at https://github.com/KTH-RPL/SeFlow." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 159, + 503, + 453, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 503, + 453, + 526 + ], + "spans": [ + { + "bbox": [ + 159, + 503, + 453, + 526 + ], + "type": "text", + "content": "Keywords: 3D scene flow, self-supervised, autonomous driving, large-scale point cloud" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 545, + 230, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 230, + 558 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 230, + 558 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 570, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 641 + ], + "type": "text", + "content": "Scene flow [30] captures the 3D velocity at every point in a point cloud. These detailed 3D flow estimates can enhance downstream tasks in autonomous driving, such as detection, segmentation, tracking, and occupancy flow estimation [20]. A common paradigm for addressing the scene flow problem is supervised learning by utilizing annotated LiDAR data [11, 38]. However, expensive labeling inherently limits the scalability of supervised learning methods." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Given the difficulty and expense of labeling scene flow ground truth, we instead focus on self-supervised scene flow approaches. Existing self-supervised" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 479, + 282 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 479, + 282 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 479, + 282 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 479, + 282 + ], + "type": "image", + "image_path": "b734f2cbf7a5cfbf504f0c2e94d336bb71f4313459c39f351b024c3efa560eee.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 290, + 482, + 389 + ], + "lines": [ + { + "bbox": [ + 130, + 290, + 482, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 290, + 482, + 389 + ], + "type": "text", + "content": "Fig. 1: LiDAR scene flow estimation using our SeFlow method on Argoverse 2. The predicted scene flow for each point is color-coded based on direction. The white indicates static points whose flow is zero. More saturated colors indicate higher velocities. (a) Camera view for visualization purposes only. (b),(c) are zoomed-in views showing the baseline from ZeroFlow [29] as well as SeFlow (ours). When predicting the flow of a large and long vehicle, the baseline predicts a portion of the flow as 0, whereas our estimates are consistent. In addition, the baseline tends to ignore small-scale objects, e.g., pedestrians, while our method can better handle such small and slow-moving objects." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 414, + 481, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 481, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 481, + 485 + ], + "type": "text", + "content": "methods can be divided into two categories: knowledge distillation and data exploration. Knowledge distillation methods [2,25,29] typically rely on a 'teacher' method to generate pseudo flow labels. These pseudo labels are then used to supervise student models. Data exploration methods [3, 14, 15, 19, 31], on the other hand, directly utilize the predicted flow to project the input points and find nearest neighbors in the next frame to establish constraints for training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 486, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 605 + ], + "type": "text", + "content": "A key challenge for self-supervised methods is that the vast majority of the points are static (about " + }, + { + "bbox": [ + 130, + 486, + 482, + 605 + ], + "type": "inline_equation", + "content": "86\\%" + }, + { + "bbox": [ + 130, + 486, + 482, + 605 + ], + "type": "text", + "content": " of points are background [7, 34]). This data imbalance often leads to overly conservative scene flow predictions. One common way to counter this is to use large amounts of training data as in ZeroFlow [29]. ZeroFlow experimentally shows that more data helps to improve the dynamic flow accuracy. In this paper, we investigate ways to improve data efficiency and take inspiration from supervised methods [11, 38]. We address the data imbalance by first classifying points as dynamic or static based on traditional ray casting when integrating frames [9]. This allows us to constrain the corresponding motion patterns by proposing novel loss functions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "Another shortcoming of existing self-supervised scene flow methods is that they disregard object-level motion constraints. The flow should be consistent, i.e., all points in a rigid object should have similar flows. A clear case of inconsistent flow can be seen on the big truck in Fig. 1, where ZeroFlow [29] predicts an absence of flow in certain sections. This inconsistency is caused by incorrect" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Q Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "content": "object point associations which also can be observed in small-scale objects, as shown in the case of the pedestrian. To account for object-level motion constraints, we propose to cluster the dynamic points into object candidates and encourage consistent flow and correct associations for the points in each cluster. This reduces the fragmentation of the flow estimate for points inside the same object." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 479, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 479, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 479, + 247 + ], + "type": "text", + "content": "Overall, our method improves the estimated flow accuracy and addresses issues in previous methods as shown in Fig. 1(c). We propose an efficient and effective self-supervised scene flow method that integrates traditional classification and learning-based strategies. Our approach is open-source at https://github.com/KTH-RPL/SeFlow. Our primary contributions include:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 256, + 480, + 352 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 138, + 256, + 479, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 256, + 479, + 280 + ], + "spans": [ + { + "bbox": [ + 138, + 256, + 479, + 280 + ], + "type": "text", + "content": "- We propose SeFlow, a novel method that integrates a dynamic classification method in formulating efficient self-supervision objectives." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 281, + 480, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 281, + 480, + 316 + ], + "spans": [ + { + "bbox": [ + 138, + 281, + 480, + 316 + ], + "type": "text", + "content": "- We further construct loss functions to learn dynamic flow estimation in imbalanced data and ensure consistent object-level flow, mitigating the effects of correspondence errors." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 316, + 480, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 316, + 480, + 352 + ], + "spans": [ + { + "bbox": [ + 138, + 316, + 480, + 352 + ], + "type": "text", + "content": "- We show that SeFlow achieves state-of-the-art results on the self-supervised scene flow task in Argoverse 2 and Waymo datasets and even outperforms all but one supervised method on the leaderboard." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 372, + 236, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 372, + 236, + 384 + ], + "spans": [ + { + "bbox": [ + 132, + 372, + 236, + 384 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 398, + 479, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 398, + 479, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 398, + 479, + 434 + ], + "type": "text", + "content": "In this section, we explore existing works in scene flow estimation. We also discuss current traditional frameworks capable of classifying dynamic points, highlighting their relevance and application in the context of scene flow tasks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 453, + 276, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 276, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 276, + 464 + ], + "type": "text", + "content": "2.1 Scene Flow Estimation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 474, + 480, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 480, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 480, + 616 + ], + "type": "text", + "content": "Scene flow estimation in autonomous driving is slightly different from flow estimation in object registration. Methods in object registration [8, 13, 16, 24, 32, 33] focus on relatively small-scale point cloud data like synthetic datasets Shapenet [5] and FlyingThing3D [17]. These methods scale poorly with the number of points. When applied to point cloud data for autonomous driving, they require downsampling to 8,192 points or less [11, 18]. In recent datasets like Argoverse 2 [34] and Waymo [11], the number of points in one full frame is around " + }, + { + "bbox": [ + 130, + 474, + 480, + 616 + ], + "type": "inline_equation", + "content": "80\\mathrm{k - }177\\mathrm{k}" + }, + { + "bbox": [ + 130, + 474, + 480, + 616 + ], + "type": "text", + "content": ". Methods that can handle such large-scale point cloud data as input usually use a voxel-based pipeline [11, 38]. However, expensive labeling of ground truth flow limits the scalability of these supervised methods, especially in autonomous driving where we have continuously increasing amounts of potential training data." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "type": "text", + "content": "To train models without labeled data, recent methods propose self-supervised losses. Many commonly used losses, such as Chamfer distance, are based on the nearest neighbor distance between two point cloud inputs [3,12,14,15,19,27,36]. However, a major limitation of nearest neighbor based losses is that only part" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 158, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 158, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 158, + 91, + 447, + 102 + ], + "type": "text", + "content": "SeFlow: A Self-Supervised Scene Flow Method in Autonomous Driving" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": "of the points on dynamic objects provides good supervision due to incorrect associations. This is especially apparent when big trucks are moving fast in autonomous driving scenarios (see Fig. 1 and Fig. 3). Mittal et al. [19] define a self-supervised loss by tracking a patch forward and backward in time to form a cycle while penalizing the errors through cycle consistency and feature similarity. Baur et al. [3] propose three loss functions with k-NN loss, rigid cycle consistency inspired by Mittal, and artificial labels based on the distance between points in the original and estimated point clouds. In the following we describe NSFP [14], FastNSF [15], and ZeroFlow [29] in more detail. These are the publically available methods that perform best on the Argoverse 2 self-supervised scene flow challenge and are therefore used as our baselines." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 247, + 482, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 342 + ], + "type": "text", + "content": "NSFP [14] is an optimization-based method. For each pair of consecutive input point clouds, NSFP iteratively learns new weights for an MLP network to predict the flow using the Chamfer distance loss. However, thousands of iterations are needed and their runtimes extend from 26 to 35 seconds per frame, which fails to meet the real-time requirements of autonomous driving. FastNSF [15] improves the efficiency by voxelizing the point cloud first and then converting it to a distance transform map for faster neighbor calculation. This reduces the runtime to 0.5 seconds, at the expense of increased error." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 342, + 482, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 415 + ], + "type": "text", + "content": "ZeroFlow [29] adopts a semi-supervised strategy. Pseudo-flow labels are created offline using NSFP [14], and a FastFlow3D [11] model is used as a student for knowledge distillation. This setup allows the student model to perform real-time inference. In this case, the teacher network needs significant resources (around 3.6 GPU months [29]) to label the entire training data. The final accuracy is also influenced by the performance of the teacher." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 415, + 482, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 482, + 475 + ], + "type": "text", + "content": "To increase data efficiency and address the limitations of the above methods, we propose to integrate efficient ray-casting-based dynamic awareness mapping into our pipeline. The core idea is to classify which points move and cluster these points into objects for which we can estimate group-level motion statistics and define better self-supervised loss functions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 491, + 324, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 491, + 324, + 504 + ], + "spans": [ + { + "bbox": [ + 131, + 491, + 324, + 504 + ], + "type": "text", + "content": "2.2 Dynamic Awareness in Mapping" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 510, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 581 + ], + "type": "text", + "content": "In the field of Simultaneous Localization and Mapping (SLAM), there is a growing interest in dynamic awareness [9,22,23,35,37]. This interest stems from the fact that points associated with moving objects in a scene can significantly impact localization and planning performance. Existing dynamic awareness methods in mapping [9,23,37] often utilize ray casting techniques to binary classify points as either dynamic or static." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": "The scene flow task, on the other hand, aims to predict the specific 3D flow at each point, a goal that extends beyond the mere categorization of dynamic and static points. In mapping, a point is considered dynamic if it moves once within a scene (even if it becomes static later), whereas, in scene flow tasks, a point is deemed dynamic if it moves faster than a certain velocity threshold in the current frame. Despite these differences, insights from the mapping field are invaluable. By integrating information over time, these frameworks develop a" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Q Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "more comprehensive understanding of the entire scene. Such scene level dynamic awareness can inform and enhance our exploration of point cloud data in scene flow tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 171, + 269, + 182 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 171, + 269, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 171, + 269, + 182 + ], + "type": "text", + "content": "3 Problem Statement" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 195, + 481, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 195, + 481, + 245 + ], + "spans": [ + { + "bbox": [ + 130, + 195, + 481, + 245 + ], + "type": "text", + "content": "This work addresses the problem of real-time scene flow estimation in autonomous driving. Given two consecutive input point clouds, " + }, + { + "bbox": [ + 130, + 195, + 481, + 245 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_t" + }, + { + "bbox": [ + 130, + 195, + 481, + 245 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 195, + 481, + 245 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{t + 1}" + }, + { + "bbox": [ + 130, + 195, + 481, + 245 + ], + "type": "text", + "content": ", along with the ego movement " + }, + { + "bbox": [ + 130, + 195, + 481, + 245 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{t,t + 1} \\in SE(3)" + }, + { + "bbox": [ + 130, + 195, + 481, + 245 + ], + "type": "text", + "content": ", the goal is to predict the motion vector as flow " + }, + { + "bbox": [ + 130, + 195, + 481, + 245 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{F}}_{t,t + 1}(p) = (x,y,z)^T" + }, + { + "bbox": [ + 130, + 195, + 481, + 245 + ], + "type": "text", + "content": " for each point " + }, + { + "bbox": [ + 130, + 195, + 481, + 245 + ], + "type": "inline_equation", + "content": "p \\in \\mathcal{P}_t" + }, + { + "bbox": [ + 130, + 195, + 481, + 245 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 245, + 481, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 245, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 245, + 481, + 281 + ], + "type": "text", + "content": "The objective is to minimize the End Point Error (EPE) which represents the difference between the predicted flow and the ground truth flow, as expressed by the following equation:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 238, + 289, + 481, + 334 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 289, + 481, + 334 + ], + "spans": [ + { + "bbox": [ + 238, + 289, + 481, + 334 + ], + "type": "interline_equation", + "content": "\\min \\underbrace {\\frac {1}{| \\mathcal {P} _ {t} |} \\sum_ {p \\in \\mathcal {P} _ {t}} \\left\\| \\hat {\\mathcal {F}} (p) - \\mathcal {F} _ {g t} (p) \\right\\| _ {2}} _ {\\text {E P E}}, \\tag {1}", + "image_path": "b86850b25245f634113ee47d1c725d1ad725ebc8041fe7fddaec9a277d0fd7b5.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 342, + 480, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 480, + 379 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 480, + 379 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 342, + 480, + 379 + ], + "type": "inline_equation", + "content": "|\\mathcal{P}_t|" + }, + { + "bbox": [ + 130, + 342, + 480, + 379 + ], + "type": "text", + "content": " denotes the cardinality of (i.e., the number of points in) " + }, + { + "bbox": [ + 130, + 342, + 480, + 379 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_t" + }, + { + "bbox": [ + 130, + 342, + 480, + 379 + ], + "type": "text", + "content": ". For consistency, in the subsequent presentation, capital symbols correspond to sets, while the lowercase symbols represent variables of specific points." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 396, + 202, + 408 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 202, + 408 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 202, + 408 + ], + "type": "text", + "content": "4 Method" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 422, + 251, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 422, + 251, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 422, + 251, + 434 + ], + "type": "text", + "content": "4.1 Input and Output" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 441, + 479, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 441, + 479, + 477 + ], + "spans": [ + { + "bbox": [ + 130, + 441, + 479, + 477 + ], + "type": "text", + "content": "The first step to process the input data, as is commonly done, is to remove ground points from " + }, + { + "bbox": [ + 130, + 441, + 479, + 477 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_t" + }, + { + "bbox": [ + 130, + 441, + 479, + 477 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 441, + 479, + 477 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{t + 1}" + }, + { + "bbox": [ + 130, + 441, + 479, + 477 + ], + "type": "text", + "content": ". This is typically done using HD maps [26,34] or ground segmentation techniques [10]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 477, + 479, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 477, + 479, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 477, + 479, + 499 + ], + "type": "text", + "content": "The estimated flow " + }, + { + "bbox": [ + 130, + 477, + 479, + 499 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{F}}" + }, + { + "bbox": [ + 130, + 477, + 479, + 499 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 130, + 477, + 479, + 499 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_t" + }, + { + "bbox": [ + 130, + 477, + 479, + 499 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 477, + 479, + 499 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{t + 1}" + }, + { + "bbox": [ + 130, + 477, + 479, + 499 + ], + "type": "text", + "content": " can be decomposed into two parts as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 270, + 500, + 480, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 270, + 500, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 270, + 500, + 480, + 514 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {F}} = \\mathcal {F} _ {e g o} + \\Delta \\hat {\\mathcal {F}}, \\tag {2}", + "image_path": "1c1484b1460b53e5f0df2171b54e99badef534c90510140a36c43405a6f7d590.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 518, + 479, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 518, + 479, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 518, + 479, + 544 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 518, + 479, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{ego}" + }, + { + "bbox": [ + 130, + 518, + 479, + 544 + ], + "type": "text", + "content": " is the flow resulting from the ego vehicle's motion which can be directly obtained from " + }, + { + "bbox": [ + 130, + 518, + 479, + 544 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{t,t + 1}" + }, + { + "bbox": [ + 130, + 518, + 479, + 544 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 518, + 479, + 544 + ], + "type": "inline_equation", + "content": "\\varDelta\\hat{\\mathcal{F}}" + }, + { + "bbox": [ + 130, + 518, + 479, + 544 + ], + "type": "text", + "content": " is our network output." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 562, + 245, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 562, + 245, + 573 + ], + "spans": [ + { + "bbox": [ + 132, + 562, + 245, + 573 + ], + "type": "text", + "content": "4.2 Model Backbone" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 582, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 480, + 666 + ], + "type": "text", + "content": "To enable real-time computation and estimation of scene flow across a large point set, voxelization is considered a practical encoding strategy in the model backbone. However, the reduction in resolution often leads to a poorer distinction of point features within the same voxel. With this in mind, we use DeFlow [38] as the architectural basis in this paper. DeFlow integrates GRU [6] with iterative refinement in the decoder design. More specifically, the GRU module maintains voxel features as its hidden state and selectively forgets or updates" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 158, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 158, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 158, + 91, + 448, + 103 + ], + "type": "text", + "content": "SeFlow: A Self-Supervised Scene Flow Method in Autonomous Driving" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 116, + 475, + 177 + ], + "blocks": [ + { + "bbox": [ + 138, + 116, + 475, + 177 + ], + "lines": [ + { + "bbox": [ + 138, + 116, + 475, + 177 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 475, + 177 + ], + "type": "image", + "image_path": "4b64c4f073be133b622aa2a3dc9f8f5b242e5240fa56a253c64412b99e17f8c6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 143, + 182, + 405, + 192 + ], + "lines": [ + { + "bbox": [ + 143, + 182, + 405, + 192 + ], + "spans": [ + { + "bbox": [ + 143, + 182, + 405, + 192 + ], + "type": "text", + "content": "Training Stage Sec. 4.4 Self-supervised Loss (including Sec. 4.3 Dynamic Classification)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 140, + 201, + 473, + 305 + ], + "blocks": [ + { + "bbox": [ + 140, + 201, + 473, + 305 + ], + "lines": [ + { + "bbox": [ + 140, + 201, + 473, + 305 + ], + "spans": [ + { + "bbox": [ + 140, + 201, + 473, + 305 + ], + "type": "image", + "image_path": "9175ccddfa2f4ff0fb904a45fc31a3f448aae906454462b96f0708da665eea7b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 314, + 482, + 423 + ], + "lines": [ + { + "bbox": [ + 130, + 314, + 482, + 423 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 482, + 423 + ], + "type": "text", + "content": "Fig. 2: SeFlow Architecture. Top: With two consecutive point clouds as inputs, our model predicts the estimated flows of all points. Bottom: Conceptual visualization of the Chamfer loss and the three proposed training losses. With the original input " + }, + { + "bbox": [ + 130, + 314, + 482, + 423 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_t" + }, + { + "bbox": [ + 130, + 314, + 482, + 423 + ], + "type": "text", + "content": " (2 static points for the building, and 2 dynamic points for the car) plus the estimated flow " + }, + { + "bbox": [ + 130, + 314, + 482, + 423 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{F}}_t" + }, + { + "bbox": [ + 130, + 314, + 482, + 423 + ], + "type": "text", + "content": ", we can calculate the error between estimated " + }, + { + "bbox": [ + 130, + 314, + 482, + 423 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{P}}_{t+1}" + }, + { + "bbox": [ + 130, + 314, + 482, + 423 + ], + "type": "text", + "content": " and the next frame point cloud " + }, + { + "bbox": [ + 130, + 314, + 482, + 423 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{t+1}" + }, + { + "bbox": [ + 130, + 314, + 482, + 423 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 314, + 482, + 423 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cham}}" + }, + { + "bbox": [ + 130, + 314, + 482, + 423 + ], + "type": "text", + "content": "). The second part is " + }, + { + "bbox": [ + 130, + 314, + 482, + 423 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{dcham}}" + }, + { + "bbox": [ + 130, + 314, + 482, + 423 + ], + "type": "text", + "content": " that only calculates the distance error between dynamic points. The third loss says that the estimated flows of static points should be zero. Finally, we assume that the flow at points from the same cluster should be consistent, and mitigate underestimation by using the proposed upper bound on the flow." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 450, + 482, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 522 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 522 + ], + "type": "text", + "content": "the information of the hidden state during each iteration according to the input point features. After multiple iterations, the optimized voxel features are concatenated with the original point features to obtain the final individual point features. Benefiting from this design, we improve the inference efficiency compared to the commonly used backbone FastFlow3D [11] without sacrificing the accuracy of scene flow estimation in coarse resolution settings." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 544, + 324, + 556 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 544, + 324, + 556 + ], + "spans": [ + { + "bbox": [ + 131, + 544, + 324, + 556 + ], + "type": "text", + "content": "4.3 Dynamic Classification of Points" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "To facilitate dynamic classification of points during the training stage, we incorporate the DUFOMap framework, a mapping-based dynamic awareness approach [9]. The key insight of this is that points observed inside a region that at one time has been observed as empty must be dynamic. Built on this insight, DUFOMap utilizes ray-casting to classify dynamic points at the sensor rate on the CPU. The result is two disjoint sets, " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{t,d}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{t,s}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{t,d}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " is the set of dynamic points that have moved once inside a scene (even if they later become static) and " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{t,s}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " is the set of static points that did not move at any time." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Q Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "type": "text", + "content": "Note that the dynamic-static classification is separated from the inference and training, offering our method good flexibility." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 155, + 261, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 155, + 261, + 167 + ], + "spans": [ + { + "bbox": [ + 132, + 155, + 261, + 167 + ], + "type": "text", + "content": "4.4 Self-supervised Loss" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 171, + 479, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 171, + 479, + 232 + ], + "spans": [ + { + "bbox": [ + 130, + 171, + 479, + 232 + ], + "type": "text", + "content": "As discussed, self-supervised learning with only Chamfer distance loss is vulnerable to problems with data imbalance and incorrect associations. We therefore construct three additional loss functions, illustrated in Fig. 2 and described in turn below, to mitigate problems with data imbalance as well as encourage consistent object-level flow." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 246, + 479, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 246, + 479, + 270 + ], + "spans": [ + { + "bbox": [ + 131, + 246, + 479, + 270 + ], + "type": "text", + "content": "Chamfer Distance The Chamfer distance, a common self-supervised loss in existing work [3, 14, 15, 19], has the following definition:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 184, + 275, + 480, + 305 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 275, + 480, + 305 + ], + "spans": [ + { + "bbox": [ + 184, + 275, + 480, + 305 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c h a m}} = \\frac {1}{| \\hat {\\mathcal {P}} _ {t + 1} |} \\sum_ {p \\in \\hat {\\mathcal {P}} _ {t + 1}} \\mathrm {S} (p, \\mathcal {P} _ {t + 1}) + \\frac {1}{| \\mathcal {P} _ {t + 1} |} \\sum_ {p \\in \\mathcal {P} _ {t + 1}} \\mathrm {S} (p, \\hat {\\mathcal {P}} _ {t + 1}) \\tag {3}", + "image_path": "f3b05e2af82a073500ae0df35cdcc526d90263e0626e6c3a718e1dac9952872d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 175, + 309, + 480, + 327 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 309, + 480, + 327 + ], + "spans": [ + { + "bbox": [ + 175, + 309, + 480, + 327 + ], + "type": "interline_equation", + "content": "\\mathrm {S} (p, \\mathcal {P} _ {o}) = \\min _ {p _ {i} \\in \\mathcal {P} _ {o}} \\| p - p _ {i} \\| _ {2} ^ {2}, \\tag {4}", + "image_path": "2a59c0e3931f00f50dfc99fae4a4ca850325350df1bb7673975cda02d485077b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 333, + 479, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 333, + 479, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 333, + 479, + 357 + ], + "type": "text", + "content": "where, " + }, + { + "bbox": [ + 130, + 333, + 479, + 357 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{P}}_{t + 1} = \\mathcal{P}_t + \\hat{\\mathcal{F}}_t" + }, + { + "bbox": [ + 130, + 333, + 479, + 357 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 333, + 479, + 357 + ], + "type": "inline_equation", + "content": "\\mathrm{S}(\\cdot) = \\mathrm{D}(\\cdot)^2" + }, + { + "bbox": [ + 130, + 333, + 479, + 357 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 333, + 479, + 357 + ], + "type": "inline_equation", + "content": "\\mathrm{D}(p,\\mathcal{P}_o)" + }, + { + "bbox": [ + 130, + 333, + 479, + 357 + ], + "type": "text", + "content": " denotes the distance between point " + }, + { + "bbox": [ + 130, + 333, + 479, + 357 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 333, + 479, + 357 + ], + "type": "text", + "content": " and its nearest neighbor in " + }, + { + "bbox": [ + 130, + 333, + 479, + 357 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_o" + }, + { + "bbox": [ + 130, + 333, + 479, + 357 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 358, + 480, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 358, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 358, + 480, + 453 + ], + "type": "text", + "content": "The Chamfer distance is proposed to calculate a similarity between two point clouds. However, when using it directly as a supervised signal for scene flow in autonomous driving, there are two issues we note. First, the number of static points is often much larger than that of dynamic points and averaging " + }, + { + "bbox": [ + 130, + 358, + 480, + 453 + ], + "type": "inline_equation", + "content": "S(\\cdot)" + }, + { + "bbox": [ + 130, + 358, + 480, + 453 + ], + "type": "text", + "content": " over all points leads the training to favor static points, i.e., zero flow estimation. Second, due to erroneous correspondence assumptions, only part of the flows within a dynamic object can be estimated correctly as shown in Fig. 3, where the estimated flows of points in the overlap area are zero." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 454, + 479, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 454, + 479, + 478 + ], + "spans": [ + { + "bbox": [ + 131, + 454, + 479, + 478 + ], + "type": "text", + "content": "To solve these issues, we contribute the following constraints using the dynamic classification information prior to supervising the network." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 492, + 480, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 492, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 131, + 492, + 480, + 611 + ], + "type": "text", + "content": "Dynamic Chamfer Distance The imbalance in the number of dynamic and static points presents a significant challenge in scene flow estimation, as highlighted in previous works [11,29,38]. Supervised methods, provided with ground truth labels for point velocity or classification, can weight their loss functions to account for this imbalance. For self-supervised learning, we introduce a dynamic Chamfer distance. In this context, 'dynamic' points are identified based on the output from Sec. 4.3. Specifically, this loss, " + }, + { + "bbox": [ + 131, + 492, + 480, + 611 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{dcham}}" + }, + { + "bbox": [ + 131, + 492, + 480, + 611 + ], + "type": "text", + "content": ", is only computed on points classified as dynamic in two consecutive point clouds. By exclusively considering dynamic points, this loss captures the nuances of motion in point cloud data. This is defined as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 150, + 615, + 480, + 646 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 615, + 480, + 646 + ], + "spans": [ + { + "bbox": [ + 150, + 615, + 480, + 646 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {d c h a m}} = \\frac {1}{| \\hat {\\mathcal {P}} _ {t + 1 , d} |} \\sum_ {p \\in \\hat {\\mathcal {P}} _ {t + 1, d}} \\mathrm {S} (p, \\mathcal {P} _ {t + 1, d}) + \\frac {1}{| \\mathcal {P} _ {t + 1 , d} |} \\sum_ {p \\in \\mathcal {P} _ {t + 1, d}} \\mathrm {S} (p, \\hat {\\mathcal {P}} _ {t + 1, d}), \\tag {5}", + "image_path": "f0a1cfd4be2ba60fdf7994f6dd7db5dfc7d0771df49eec525dbb5af6a99c8a00.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 652, + 253, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 652, + 253, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 652, + 253, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 652, + 253, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{P}}_{t + 1,d} = \\mathcal{P}_{t,d} + \\hat{\\mathcal{F}}_{t,d}" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 158, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 158, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 158, + 91, + 448, + 103 + ], + "type": "text", + "content": "SeFlow: A Self-Supervised Scene Flow Method in Autonomous Driving" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 112, + 317, + 193 + ], + "blocks": [ + { + "bbox": [ + 138, + 112, + 317, + 193 + ], + "lines": [ + { + "bbox": [ + 138, + 112, + 317, + 193 + ], + "spans": [ + { + "bbox": [ + 138, + 112, + 317, + 193 + ], + "type": "image", + "image_path": "afbec55c88188da0e4b650e4654f135a33a82413da23f76200ca23077c41c692.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 181, + 197, + 257, + 206 + ], + "lines": [ + { + "bbox": [ + 181, + 197, + 257, + 206 + ], + "spans": [ + { + "bbox": [ + 181, + 197, + 257, + 206 + ], + "type": "text", + "content": "(a) Chamfer distance" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 323, + 113, + 471, + 196 + ], + "blocks": [ + { + "bbox": [ + 323, + 113, + 471, + 196 + ], + "lines": [ + { + "bbox": [ + 323, + 113, + 471, + 196 + ], + "spans": [ + { + "bbox": [ + 323, + 113, + 471, + 196 + ], + "type": "image", + "image_path": "a08235b6b1185d0e2165915e5d62cf503a554d62f64af9db437ddd1c8914492a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 349, + 197, + 440, + 207 + ], + "lines": [ + { + "bbox": [ + 349, + 197, + 440, + 207 + ], + "spans": [ + { + "bbox": [ + 349, + 197, + 440, + 207 + ], + "type": "text", + "content": "(b) Point flow magnitude" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 217, + 482, + 283 + ], + "lines": [ + { + "bbox": [ + 130, + 217, + 482, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 217, + 482, + 283 + ], + "type": "text", + "content": "Fig. 3: Simple visualization of the shortcomings of using Chamfer distance as a supervisory signal for flow value estimation. The denser color on points in (b) represents higher flow values and white means the point's flow is zero. (a) illustrates how to calculate loss based on Chamfer distance, and (b) shows that the flow results, based on the nearest neighbor principle, can lead to zero flow estimation for the middle of the object." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 305, + 482, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 305, + 482, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 305, + 482, + 389 + ], + "type": "text", + "content": "Static Flow The standard Chamfer distance rests on a very strong assumption: nearest neighbor matching can find perfect one-to-one correspondences of all points between two frames. However, due to varying observations, the number of points in the two frames differs, leading to inconsistencies and potential mismatches. To deal with the possible mismatches in the static areas, we add a constraint to encourage the model to estimate zero flow for static points. This loss is defined as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 235, + 394, + 481, + 424 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 394, + 481, + 424 + ], + "spans": [ + { + "bbox": [ + 235, + 394, + 481, + 424 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s t a t i c}} = \\frac {1}{| \\mathcal {P} _ {t , s} |} \\sum_ {p \\in \\mathcal {P} _ {t, s}} \\| \\Delta \\hat {\\mathcal {F}} (p) \\| _ {2} ^ {2}, \\tag {6}", + "image_path": "0f916866fe8b5b7404bfa7962844b27ab47eafd5274474f620b6eaa601698af9.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 432, + 398, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 432, + 398, + 447 + ], + "spans": [ + { + "bbox": [ + 132, + 432, + 398, + 447 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 432, + 398, + 447 + ], + "type": "inline_equation", + "content": "\\varDelta\\hat{\\mathcal{F}}(p)" + }, + { + "bbox": [ + 132, + 432, + 398, + 447 + ], + "type": "text", + "content": " represents the network output flow of point " + }, + { + "bbox": [ + 132, + 432, + 398, + 447 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 132, + 432, + 398, + 447 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 461, + 482, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 592 + ], + "type": "text", + "content": "Dynamic Cluster Flow For dynamic fields, the inconsistencies and correspondence errors are even more complicated. For instance, we observe that nearest neighbor matching often leads to erroneous results, e.g., on large objects undergoing translation, which is very common for moving vehicles in urban street environments as shown in Fig. 1. More specifically, Figure 3 illustrates the issue on a simplified car example. In this case, the flow is parallel to the object's surface, which means that nearest neighbor matching does not provide good supervision in the interior of the surface (" + }, + { + "bbox": [ + 130, + 461, + 482, + 592 + ], + "type": "inline_equation", + "content": "\\mathrm{D} \\approx 0" + }, + { + "bbox": [ + 130, + 461, + 482, + 592 + ], + "type": "text", + "content": " for the points in the overlap area). This mismatch results in an incorrect local optimum that remains unresolved during the training process as it minimizes the losses based on Chamfer Distance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 594, + 482, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 652 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 652 + ], + "type": "text", + "content": "Therefore, we suggest that the motion of all parts within an object (a cluster) should be approximately homogeneous over a short time interval. We use the HDBSCAN [4,21] clustering algorithm to identify different moving objects. This clustering is only applied to dynamic points, thus reducing the computational overhead." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 250, + 653, + 481, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 653, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 250, + 653, + 481, + 667 + ], + "type": "interline_equation", + "content": "\\mathcal {C} _ {t, d} = \\operatorname {H D B S C A N} \\left(\\mathcal {P} _ {t, d}\\right). \\tag {7}", + "image_path": "f02547856f60913e734dcc9ed2621accad2bf5bf5b0d04841db6ae0b8f064157.jpg" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Q Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": "Directly constraining the variance of the flow distribution inside a cluster would be a straightforward idea, but does not guarantee the correctness of the obtained mean value after convergence. As we know that nearest-neighbour point correspondence can considerably underestimate flow on geometrically featureless object surfaces, we here instead propose to use an upper bound on the object motion as the supervisory signal. We derive this upper bound, per object cluster, by taking the maximum inter-frame distance of all point correspondences within the cluster. Specifically, we exploit information from the original input point cloud data, i.e., " + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_t" + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{t + 1}" + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": ". After the dynamic classification and clustering process, we find the index of the point " + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "inline_equation", + "content": "p_k" + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": " in cluster " + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "inline_equation", + "content": "c_{i} \\in \\mathcal{C}_{t,d}" + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": " with the largest distance to its nearest neighbor point in " + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{t + 1,d}" + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": ", i.e.," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 224, + 258, + 481, + 276 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 258, + 481, + 276 + ], + "spans": [ + { + "bbox": [ + 224, + 258, + 481, + 276 + ], + "type": "interline_equation", + "content": "\\kappa_ {i} = \\underset {k} {\\arg \\max } \\left\\{\\mathrm {D} \\left(p _ {k}, \\mathcal {P} _ {t + 1, d}\\right) \\mid p _ {k} \\in \\mathcal {P} _ {c _ {i}} \\right\\}. \\tag {8}", + "image_path": "cf7313015090051629f4c8763fc682e18d9b45a0faf2399fad59aa1fe7d118da.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 286, + 408, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 286, + 408, + 300 + ], + "spans": [ + { + "bbox": [ + 131, + 286, + 408, + 300 + ], + "type": "text", + "content": "We calculate the upper bound, " + }, + { + "bbox": [ + 131, + 286, + 408, + 300 + ], + "type": "inline_equation", + "content": "\\tilde{f}_{c_i}" + }, + { + "bbox": [ + 131, + 286, + 408, + 300 + ], + "type": "text", + "content": " on the flow for cluster " + }, + { + "bbox": [ + 131, + 286, + 408, + 300 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 131, + 286, + 408, + 300 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 273, + 308, + 480, + 323 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 273, + 308, + 480, + 323 + ], + "spans": [ + { + "bbox": [ + 273, + 308, + 480, + 323 + ], + "type": "interline_equation", + "content": "\\tilde {f} _ {c _ {i}} = p _ {\\kappa_ {i}} ^ {\\prime} - p _ {\\kappa_ {i}}, \\tag {9}", + "image_path": "a1661359dc9735583a3efe1f78f6eea81a4d650929849e980209137f55439777.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 331, + 481, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 331, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 131, + 331, + 481, + 357 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 331, + 481, + 357 + ], + "type": "inline_equation", + "content": "p_{\\kappa_i}'" + }, + { + "bbox": [ + 131, + 331, + 481, + 357 + ], + "type": "text", + "content": " is the nearest neighbor of point " + }, + { + "bbox": [ + 131, + 331, + 481, + 357 + ], + "type": "inline_equation", + "content": "p_{\\kappa_i}" + }, + { + "bbox": [ + 131, + 331, + 481, + 357 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 131, + 331, + 481, + 357 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{t+1,d}" + }, + { + "bbox": [ + 131, + 331, + 481, + 357 + ], + "type": "text", + "content": ". We use this to drive the estimated flows of cluster " + }, + { + "bbox": [ + 131, + 331, + 481, + 357 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 131, + 331, + 481, + 357 + ], + "type": "text", + "content": " towards " + }, + { + "bbox": [ + 131, + 331, + 481, + 357 + ], + "type": "inline_equation", + "content": "\\tilde{f}_{c_i}" + }, + { + "bbox": [ + 131, + 331, + 481, + 357 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 255, + 368, + 480, + 393 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 368, + 480, + 393 + ], + "spans": [ + { + "bbox": [ + 255, + 368, + 480, + 393 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c _ {i}} = \\sum_ {p _ {j} \\in \\mathcal {P} _ {c _ {i}}} | | \\hat {f} _ {p _ {j}} - \\tilde {f} _ {c _ {i}} | | _ {2} ^ {2}, \\tag {10}", + "image_path": "98a9768994359bb7db8085779fb46a98cd445aefc480a7ad4e2659296977361c.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 249, + 397, + 480, + 425 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 397, + 480, + 425 + ], + "spans": [ + { + "bbox": [ + 249, + 397, + 480, + 425 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {d c l s}} = \\frac {1}{| \\mathcal {P} _ {t , d} |} \\sum_ {c _ {i} \\in \\mathcal {C} _ {t, d}} \\mathcal {L} _ {c _ {i}}. \\tag {11}", + "image_path": "30439f172b032c1ca4d6aeff0d6d42d8d462dfd32f96cdfb8c2d2178b0810821.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 434, + 479, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 479, + 457 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 479, + 457 + ], + "type": "text", + "content": "In conclusion, the final SeFlow loss incorporates all four losses introduced above," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 217, + 459, + 480, + 471 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 459, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 217, + 459, + 480, + 471 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {c h a m}} + \\mathcal {L} _ {\\text {s t a t i c}} + \\mathcal {L} _ {\\text {d c h a m}} + \\mathcal {L} _ {\\text {d c l s}}. \\tag {12}", + "image_path": "77ff7815e64f148cc675b15e050a075957bfa593d61e4a957322a2fcee52e96a.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 476, + 453, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 476, + 453, + 488 + ], + "spans": [ + { + "bbox": [ + 131, + 476, + 453, + 488 + ], + "type": "text", + "content": "The effect of each loss will be evaluated in the ablation study in Sec. 5.3." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 506, + 224, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 506, + 224, + 520 + ], + "spans": [ + { + "bbox": [ + 131, + 506, + 224, + 520 + ], + "type": "text", + "content": "5 Experiment" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 532, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 532, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 532, + 482, + 593 + ], + "type": "text", + "content": "In this section, we first outline evaluation details and then present quantitative comparisons with state-of-the-art methods on two benchmark datasets. A series of ablation studies are presented to better evaluate the individual components of our approach. Finally, we conclude with qualitative results on Argoverse 2 and discuss limitations of the approach." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 609, + 259, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 609, + 259, + 621 + ], + "spans": [ + { + "bbox": [ + 131, + 609, + 259, + 621 + ], + "type": "text", + "content": "5.1 Dataset and Metric" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "We briefly introduce the dataset and metrics for evaluation in the following section. More implementation details, such as hyperparameters for training and dataset descriptions, can be found in the supplementary material." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 158, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 158, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 158, + 91, + 448, + 103 + ], + "type": "text", + "content": "SeFlow: A Self-Supervised Scene Flow Method in Autonomous Driving" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 157, + 168, + 454, + 293 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 1: Comparisons on Argoverse 2 test set from the online leaderboard [1]. Upper groups are supervised methods, lower groups are self-supervised methods. Our method achieves state-of-art performance in the self-supervised scene flow task. " + }, + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "text", + "content": " means methods can run in real-time (10 Hz) onboard." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 157, + 168, + 454, + 293 + ], + "lines": [ + { + "bbox": [ + 157, + 168, + 454, + 293 + ], + "spans": [ + { + "bbox": [ + 157, + 168, + 454, + 293 + ], + "type": "table", + "html": "
MethodRun Time per frame [ms]EPE ↓
3-wayFDFSBS
FastFlow3D†[11]34 ± 50.07820.20720.02530.0020
DeFlow†[38]48 ± 40.05340.13400.02320.0029
FastNSF [15]507 ± 3120.16570.35400.04060.1025
NSFP [14]32,060 ± 10,1120.06850.15030.03020.0248
ZeroFlow†[29]34 ± 50.08140.21090.02540.0080
SeFlow (Ours)†48 ± 40.06280.15250.03210.0038
", + "image_path": "2628bf735d7f6e7d5996efed1edb5475ea5707794ee42fa3d0211c5e7a309d42.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 313, + 482, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 313, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 313, + 482, + 445 + ], + "type": "text", + "content": "Dataset We evaluate our approach on two large-scale autonomous driving datasets: Argoverse 2 [34] and Waymo [26]. Ground removal is performed using HD maps for both datasets as described in [34]. Waymo datasets contain 798 training and 202 validation scenes respectively. We focus our description here on Argoverse 2, which provides official and public scene flow challenges [1,28], with the Sensor and LiDAR datasets. The Sensor dataset encompasses 700 training and 150 validation scenes. Each scene is approximately 15 seconds long in " + }, + { + "bbox": [ + 132, + 313, + 482, + 445 + ], + "type": "inline_equation", + "content": "10\\mathrm{Hz}" + }, + { + "bbox": [ + 132, + 313, + 482, + 445 + ], + "type": "text", + "content": " with complete annotations for evaluation. The evaluation for another 150 test scenes can be accessed indirectly by submitting a solution to the leaderboard. The LiDAR dataset contains 20,000 scenes without any annotation and is only used as extra data in Sec. 5.3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 461, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 461, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 132, + 461, + 482, + 582 + ], + "type": "text", + "content": "Metric The benchmark follows existing works [1,7,29,38] and uses the three-way End Point Error. End Point Error (EPE), as defined in Eq. (1), measures the L2 norm of the discrepancy between the predicted and actual flow vectors, expressed in meters. The EPE three-way (3-way) is defined as the unweighted average EPE across three disjoint sets of points: Foreground Dynamic (FD), Foreground Static (FS), and Background Static (BS). The definition of 'dynamic' is as follows: If the flow at a point exceeds the threshold (the public leaderboard setting is " + }, + { + "bbox": [ + 132, + 461, + 482, + 582 + ], + "type": "inline_equation", + "content": "0.05\\mathrm{m}" + }, + { + "bbox": [ + 132, + 461, + 482, + 582 + ], + "type": "text", + "content": "), the point is defined as dynamic. Given the " + }, + { + "bbox": [ + 132, + 461, + 482, + 582 + ], + "type": "inline_equation", + "content": "10\\mathrm{Hz}" + }, + { + "bbox": [ + 132, + 461, + 482, + 582 + ], + "type": "text", + "content": " sensor frequency, this threshold corresponds to a speed of " + }, + { + "bbox": [ + 132, + 461, + 482, + 582 + ], + "type": "inline_equation", + "content": "0.5\\mathrm{m/s}" + }, + { + "bbox": [ + 132, + 461, + 482, + 582 + ], + "type": "text", + "content": ". All evaluations are limited to points that are within a " + }, + { + "bbox": [ + 132, + 461, + 482, + 582 + ], + "type": "inline_equation", + "content": "100\\mathrm{m} \\times 100\\mathrm{m}" + }, + { + "bbox": [ + 132, + 461, + 482, + 582 + ], + "type": "text", + "content": " box centered on the ego vehicle." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 597, + 264, + 610 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 597, + 264, + 610 + ], + "spans": [ + { + "bbox": [ + 133, + 597, + 264, + 610 + ], + "type": "text", + "content": "5.2 Quantitative Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 616, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 616, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 616, + 481, + 666 + ], + "type": "text", + "content": "We evaluate the performance of our method SeFlow and compare it with the currently best-performing methods on the Argoverse 2 test set and Waymo validation set. In Tab. 1 and Tab. 2, the upper group, FastFlow3D and DeFlow, are supervised methods trained with ground truth flow. Compared to the supervised" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Q Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 157, + 168, + 454, + 293 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "text", + "content": "Table 2: Comparisons on Waymo Open dataset validation set. Upper groups are supervised methods, lower groups are self-supervised methods. Our method achieves state-of-art performance in the self-supervised scene flow task. " + }, + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "text", + "content": " means methods can run in real-time (10 Hz) onboard." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 157, + 168, + 454, + 293 + ], + "lines": [ + { + "bbox": [ + 157, + 168, + 454, + 293 + ], + "spans": [ + { + "bbox": [ + 157, + 168, + 454, + 293 + ], + "type": "table", + "html": "
MethodRun Time per frame [ms]EPE ↓
3-wayFDFSBS
FastFlow3D†[11]27 ± 60.07820.19540.02460.0152
DeFlow†[38]42 ± 40.04460.09800.02590.0098
FastNSF [14]593 ± 3080.15790.30120.01460.0403
NSFP [15]76,163 ± 32,2560.10050.17120.10810.0221
ZeroFlow†[29]27 ± 60.09210.21620.01530.0241
SeFlow (Ours)†42 ± 40.05980.15060.01810.0106
", + "image_path": "03e0445014b8ec2137144fd9371dfdeca109786ca81c3a44b406aa5229f579cb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 315, + 480, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 315, + 480, + 362 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 480, + 362 + ], + "type": "text", + "content": "methods, we note that SeFlow outperforms FastFlow3D and approaches the level of DeFlow in terms of EPE 3-way. This comparative analysis demonstrates the great potential of self-supervised learning in scene flow estimation and validates the effectiveness of our approach." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 363, + 482, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 363, + 482, + 470 + ], + "spans": [ + { + "bbox": [ + 130, + 363, + 482, + 470 + ], + "type": "text", + "content": "In the self-supervised category, our SeFlow method achieves state-of-the-art performance on both datasets. While FastNSF and NSFP are optimization-based methods that do not rely on pre-trained weights for subsequent estimations, their inference times are not conducive to real-time requirements. NSFP, with the second best result in Tab. 1, takes approximately 30 seconds to predict a single frame, which is impractical for real-time autonomous driving applications. FastNSF, on the other hand, enhances efficiency through voxelization for quicker Chamfer distance calculation. Although significantly faster than NSFP, the performance of FastNSFP is the worst among all methods evaluated." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 471, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 471, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 471, + 482, + 555 + ], + "type": "text", + "content": "ZeroFlow is capable of estimating scene flow in real-time, with an accuracy similar to NSFP (slightly better on Waymo and slightly worse on Argoverse 2). SeFlow stands out not only for achieving the best result, but also for drastically reducing the run time to 50 milliseconds, which is more than two orders of magnitude faster than NSFP while providing more accurate flow estimates. SeFlow's state-of-the-art performance demonstrated in Tab. 1 and Tab. 2 underscores the effectiveness of our novel self-supervised method in scene flow estimation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 573, + 237, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 573, + 237, + 586 + ], + "spans": [ + { + "bbox": [ + 131, + 573, + 237, + 586 + ], + "type": "text", + "content": "5.3 Ablation study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "In this section, we delve into two key aspects of our SeFlow pipeline. Firstly, we examine the impact of different loss terms on the accuracy of our flow prediction results. This analysis aims to demonstrate the necessity and effectiveness of the loss components we propose. Secondly, we explore how the size of the training dataset, especially in the case of limited training data, affects the outcomes of our self-supervised training process." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 158, + 90, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 158, + 90, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 158, + 90, + 447, + 103 + ], + "type": "text", + "content": "SeFlow: A Self-Supervised Scene Flow Method in Autonomous Driving" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 162, + 146, + 445, + 242 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "type": "text", + "content": "Table 3: Ablation study of loss terms. Results are evaluated on the Argoverse 2 validation set with 20 training epochs." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 162, + 146, + 445, + 242 + ], + "lines": [ + { + "bbox": [ + 162, + 146, + 445, + 242 + ], + "spans": [ + { + "bbox": [ + 162, + 146, + 445, + 242 + ], + "type": "table", + "html": "
Exp. IdLchamLdchamLstaticLdclsEPE ↓
3-wayFDFSBS
1✓0.09620.2030.0520.033
2✓✓0.09160.1810.0590.035
3✓✓✓0.07790.2200.0120.002
4✓✓✓✓0.06430.1600.0290.004
", + "image_path": "674d1867891e4548232e22dc697832e7559c5aa5eb168b845c9b0c32031c6081.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 264, + 481, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 264, + 481, + 312 + ], + "spans": [ + { + "bbox": [ + 130, + 264, + 481, + 312 + ], + "type": "text", + "content": "Loss Terms The advantages of each loss design are evident in Tab. 3 evaluated on the Argoverse 2 validation set with 20 training epochs. Instead of relying solely on the chamfer loss " + }, + { + "bbox": [ + 130, + 264, + 481, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cham}}" + }, + { + "bbox": [ + 130, + 264, + 481, + 312 + ], + "type": "text", + "content": ", we investigate how incorporating the additional three losses (" + }, + { + "bbox": [ + 130, + 264, + 481, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{static}}" + }, + { + "bbox": [ + 130, + 264, + 481, + 312 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 264, + 481, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{dcham}}" + }, + { + "bbox": [ + 130, + 264, + 481, + 312 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 264, + 481, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{dcls}}" + }, + { + "bbox": [ + 130, + 264, + 481, + 312 + ], + "type": "text", + "content": ") can boost the performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 312, + 482, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 482, + 479 + ], + "type": "text", + "content": "After adding the dynamic Chamfer loss " + }, + { + "bbox": [ + 130, + 312, + 482, + 479 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{dcham}}" + }, + { + "bbox": [ + 130, + 312, + 482, + 479 + ], + "type": "text", + "content": ", experiment 2 shows a decrease in flow estimation error of about 0.22 (10%) for foreground dynamics (FD), while the static errors (FS, BS) remain essentially the same. Experiment 3 then incorporates " + }, + { + "bbox": [ + 130, + 312, + 482, + 479 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{static}}" + }, + { + "bbox": [ + 130, + 312, + 482, + 479 + ], + "type": "text", + "content": " constraint, which significantly reduces the foreground and background static flow estimation errors, 80% for FS and 94% for BS. However, adding " + }, + { + "bbox": [ + 130, + 312, + 482, + 479 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{static}}" + }, + { + "bbox": [ + 130, + 312, + 482, + 479 + ], + "type": "text", + "content": " also increases the foreground dynamic error. We attribute this to the difficulty of estimating the flow of moving, geometrically featureless, objects as mentioned previously, which would be reinforced by " + }, + { + "bbox": [ + 130, + 312, + 482, + 479 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{static}}" + }, + { + "bbox": [ + 130, + 312, + 482, + 479 + ], + "type": "text", + "content": ". Even so, considering the two static components in the EPE 3-way metric, the EPE 3-way would still benefit considerably from this loss (15% error reduction). Finally, in experiment 4, we incorporate the " + }, + { + "bbox": [ + 130, + 312, + 482, + 479 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{dcls}}" + }, + { + "bbox": [ + 130, + 312, + 482, + 479 + ], + "type": "text", + "content": ". This results in a notable decrease in overall EPE 3-way by 33% compared to solely using the chamfer loss (experiment 1). The above experiments illustrate that our method is not a simple stack of losses, but a complementary holistic design." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 497, + 481, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 481, + 533 + ], + "type": "text", + "content": "Training Dataset Size In the context of robotics and autonomous driving, there are situations where the number of accessible frames is limited. This section evaluates the performance of SeFlow given different amounts of training data." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 534, + 481, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 481, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 481, + 593 + ], + "type": "text", + "content": "In this experiment, we denote the entire Argoverse 2 Sensor training set as " + }, + { + "bbox": [ + 130, + 534, + 481, + 593 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 130, + 534, + 481, + 593 + ], + "type": "text", + "content": ". Extra unlabeled data can be retrieved from the LiDAR dataset for self-supervised methods. To assess the impact of dataset size, we conduct the same 50 epochs of training for methods using " + }, + { + "bbox": [ + 130, + 534, + 481, + 593 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 534, + 481, + 593 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 534, + 481, + 593 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 534, + 481, + 593 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 534, + 481, + 593 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 534, + 481, + 593 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 534, + 481, + 593 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 130, + 534, + 481, + 593 + ], + "type": "text", + "content": " of the total data in Tab. 4 and Fig. 4 and evaluate the resulting models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": "From Tab. 4, we can observe that with only " + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": " training data, our SeFlow can already outperform ZeroFlow which uses " + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "inline_equation", + "content": "200\\%" + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": " data. Figure 4 illustrates even more intuitively that our SeFlow can easily outperform both existing supervised (FastFlow3D) and unsupervised (ZeroFlow) approaches with the same amount of data. We attribute the demonstrated data efficiency of our method to the well-designed loss functions, which integrate a dynamic" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Q Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "awareness mapping method into our framework and enable better scene-level comprehension." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 138, + 159, + 305, + 294 + ], + "blocks": [ + { + "bbox": [ + 138, + 159, + 305, + 294 + ], + "lines": [ + { + "bbox": [ + 138, + 159, + 305, + 294 + ], + "spans": [ + { + "bbox": [ + 138, + 159, + 305, + 294 + ], + "type": "table", + "html": "
DatasetEPE ↓
3-wayFDFSBS
10%0.0940.2340.0400.006
20%0.0780.1970.0320.004
50%0.0660.1670.0280.004
100%0.0590.1470.0260.004
ZF 100%0.0880.2310.0220.011
ZF 200%0.0760.1980.0180.011
", + "image_path": "93ca902642cded9633b6a7fc72aa7a77b6a59c360b87c651b5bb385571aed17e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 301, + 308, + 390 + ], + "lines": [ + { + "bbox": [ + 130, + 301, + 308, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 301, + 308, + 390 + ], + "type": "text", + "content": "Table 4: Ablation study of dataset size compared with Zeroflow (ZF), another self-supervised learning method. Results are evaluated on the Argoverse 2 validation set with 50 training epochs. The total dataset " + }, + { + "bbox": [ + 130, + 301, + 308, + 390 + ], + "type": "inline_equation", + "content": "(100\\%)" + }, + { + "bbox": [ + 130, + 301, + 308, + 390 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 130, + 301, + 308, + 390 + ], + "type": "inline_equation", + "content": "107\\mathrm{k}" + }, + { + "bbox": [ + 130, + 301, + 308, + 390 + ], + "type": "text", + "content": " frames. " + }, + { + "bbox": [ + 130, + 301, + 308, + 390 + ], + "type": "inline_equation", + "content": "200\\%" + }, + { + "bbox": [ + 130, + 301, + 308, + 390 + ], + "type": "text", + "content": " data means all of the Sensor dataset plus an equal amount of the LiDAR dataset (214k frames)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 322, + 159, + 479, + 305 + ], + "blocks": [ + { + "bbox": [ + 322, + 159, + 479, + 305 + ], + "lines": [ + { + "bbox": [ + 322, + 159, + 479, + 305 + ], + "spans": [ + { + "bbox": [ + 322, + 159, + 479, + 305 + ], + "type": "image", + "image_path": "f331b8fda3d355540cfbc7e315602ca7e50ca3bea6c3270cdc7fc383d3899b8b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 312, + 482, + 389 + ], + "lines": [ + { + "bbox": [ + 318, + 312, + 482, + 389 + ], + "spans": [ + { + "bbox": [ + 318, + 312, + 482, + 389 + ], + "type": "text", + "content": "Fig.4: The relationship between flow estimation error and training dataset size, scaled in " + }, + { + "bbox": [ + 318, + 312, + 482, + 389 + ], + "type": "inline_equation", + "content": "\\log_{10}" + }, + { + "bbox": [ + 318, + 312, + 482, + 389 + ], + "type": "text", + "content": ". Methods with \\* are supervised by ground truth labels. SeFlow uses less data but gets comparable results compared to FastFlow3D and ZeroFlow." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 415, + 257, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 415, + 257, + 426 + ], + "spans": [ + { + "bbox": [ + 131, + 415, + 257, + 426 + ], + "type": "text", + "content": "5.4 Qualitative Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 437, + 482, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 484 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 484 + ], + "type": "text", + "content": "Figure 5 presents a qualitative flow estimation result on a sequence of scenes in the Argoverse 2 validation set. The first three columns showcase the same scene at different timestamps, and the last column shows a zoomed-in view of the scene." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 485, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 482, + 641 + ], + "type": "text", + "content": "In the first column featuring a vehicle partially occluded by a traffic pole, SeFlow demonstrates superior flow estimation capabilities compared to ZeroFlow. Sometimes, SeFlow can even detect flow overlooked by the ground truth annotations. Argoverse 2 derives ground truth flow from manual instance level labels, and as a result, the flow of points outside the bounding box may be ignored. This issue is particularly pronounced in smaller objects, and an example of this can be seen in the fourth column where we zoom in on a case. In this example, a pedestrian is pushing a shopping cart across the road. By comparing the first and the third column, we can observe that both objects are moving. However, the ground truth labels zero flow (white) for the shopping cart. SeFlow, on the other hand, successfully predicts consistent flow even at this small scale. In comparison, the baseline method ZeroFlow suffers at flow prediction of small objects and regards both the pedestrian and the shopping cart as static (white)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "content": "Although our method shows superior scene flow estimation compared to other baseline methods, it also has some limitations. Based on the zoom-in view in the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 158, + 90, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 158, + 90, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 158, + 90, + 448, + 103 + ], + "type": "text", + "content": "SeFlow: A Self-Supervised Scene Flow Method in Autonomous Driving" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 130, + 114, + 482, + 295 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 295 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 295 + ], + "type": "image", + "image_path": "fa1e6b21bceb26a2908cdc77e14dd6cee8f7d946660a5657001c89dc85546bfc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 301, + 483, + 357 + ], + "lines": [ + { + "bbox": [ + 130, + 301, + 483, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 301, + 483, + 357 + ], + "type": "text", + "content": "Fig. 5: Qualitative results from Argoverse 2 validation set. The top row displays the ground truth flow, the middle row presents the SeFlow result, and the bottom row showcases another self-supervised method ZeroFlow [29] result. Different color indicates different directions and more saturated color means larger flow estimation. Ego motion is compensated for a clearer view." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 383, + 483, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 383, + 483, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 383, + 483, + 468 + ], + "type": "text", + "content": "fourth column in Fig. 5, we can find some purple flow labels at the top right corner in the first row (Ground Truth). Both SeFlow (Ours) and ZeroFlow fail to predict any flow for these points. One possible reason is that the point cloud of distant objects is sparse and easily ignored by both the voxelization process and clustering algorithms. This makes it difficult to estimate the flow of distant objects using only two consecutive frames. Based on this, multi-modality or time-consistent networks would be a further direction for future research." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 491, + 227, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 227, + 503 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 227, + 503 + ], + "type": "text", + "content": "6 Conclusions" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 520, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 520, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 520, + 482, + 616 + ], + "type": "text", + "content": "In this paper, we proposed SeFlow, an efficient and effective self-supervised training method for scene flow estimation in autonomous driving with large-scale point clouds as input. Our primary contributions include a learning method that integrates static and dynamic awareness to construct self-supervision objectives. We further identify problems with the correspondence assumptions of Chamfer-based loss functions commonly used for self-supervised learning, and mitigate these with a constraint based on the upper bound of object motion. Our experimental results underscore the efficacy of our approach." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Future research may concentrate on integrating multi-modality (e.g., cameras and radar) within the SeFlow framework for higher flow estimation accuracy or embedding temporal coherence concepts within our pipeline. Additionally, the optimization of multi-loss learning weights warrants further exploration." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Q Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 246 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 246 + ], + "type": "text", + "content": "Thanks to RPL member: Li Ling helps review this manuscript. Thanks to Kyle Vedder (the author of ZeroFlow), who kindly opened his code including pretrained weights, and discussed their result with us which helped this work a lot. This work was partially supported by the Wallenberg AI, Autonomous Systems and Software Program (WASP) funded by the Knut and Alice Wallenberg Foundation and Prosense (2020-02963) funded by Vinnova. The computations were enabled by the supercomputing resource Berzelius provided by National Supercomputer Centre at Linköping University and the Knut and Alice Wallenberg Foundation, Sweden." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 262, + 197, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 262, + 197, + 275 + ], + "spans": [ + { + "bbox": [ + 132, + 262, + 197, + 275 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 285, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 138, + 285, + 481, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 285, + 481, + 307 + ], + "spans": [ + { + "bbox": [ + 138, + 285, + 481, + 307 + ], + "type": "text", + "content": "1. 2, A.: Argoverse 2 scene flow online leaderboard. https://eval.ai/web/challenges/challenge-page/2010/leaderboard/4759 (2024 Mar 4th)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 308, + 481, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 308, + 481, + 340 + ], + "spans": [ + { + "bbox": [ + 138, + 308, + 481, + 340 + ], + "type": "text", + "content": "2. Aleotti, F., Poggi, M., Tosi, F., Mattoccia, S.: Learning end-to-end scene flow by distilling single tasks knowledge. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 10435-10442 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 341, + 481, + 383 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 341, + 481, + 383 + ], + "spans": [ + { + "bbox": [ + 138, + 341, + 481, + 383 + ], + "type": "text", + "content": "3. Baur, S.A., Emmerichs, D.J., Moosmann, F., Pinggera, P., Ommer, B., Geiger, A.: Slim: Self-supervised lidar scene flow and motion segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13126-13136 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 384, + 481, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 384, + 481, + 415 + ], + "spans": [ + { + "bbox": [ + 138, + 384, + 481, + 415 + ], + "type": "text", + "content": "4. Campello, R.J., Moulavi, D., Sander, J.: Density-based clustering based on hierarchical density estimates. In: Pacific-Asia conference on knowledge discovery and data mining. pp. 160-172. Springer (2013)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 417, + 481, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 417, + 481, + 449 + ], + "spans": [ + { + "bbox": [ + 138, + 417, + 481, + 449 + ], + "type": "text", + "content": "5. Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., et al.: Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012 (2015)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 449, + 481, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 449, + 481, + 481 + ], + "spans": [ + { + "bbox": [ + 138, + 449, + 481, + 481 + ], + "type": "text", + "content": "6. Cho, K., Van Merrienboer, B., Gulcehre, C., Bahdanau, D., Bougares, F., Schwenk, H., Bengio, Y.: Learning phrase representations using rnn encoder-decoder for statistical machine translation. arXiv preprint arXiv:1406.1078 (2014)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 483, + 481, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 483, + 481, + 503 + ], + "spans": [ + { + "bbox": [ + 138, + 483, + 481, + 503 + ], + "type": "text", + "content": "7. Chodosh, N., Ramanan, D., Lucey, S.: Re-evaluating lidar scene flow for autonomous driving. arXiv preprint arXiv:2304.02150 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 504, + 481, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 504, + 481, + 535 + ], + "spans": [ + { + "bbox": [ + 138, + 504, + 481, + 535 + ], + "type": "text", + "content": "8. Deng, D., Zakhor, A.: Rsf: Optimizing rigid scene flow from 3d point clouds without labels. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1277-1286 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 536, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 536, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 536, + 481, + 567 + ], + "type": "text", + "content": "9. Duberg, D., Zhang, Q., Jia, M., Jensfelt, P.: DUFOMap: Efficient dynamic awareness mapping. IEEE Robotics and Automation Letters 9(6), 5038-5045 (2024). https://doi.org/10.1109/LRA.2024.3387658" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 568, + 481, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 568, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 134, + 568, + 481, + 600 + ], + "type": "text", + "content": "10. Himmelsbach, M., Hundelshausen, F.V., Wuensche, H.J.: Fast segmentation of 3d point clouds for ground vehicles. In: Intelligent Vehicles Symposium (IV), 2010 IEEE. pp. 560-565. IEEE (2010)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 601, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 601, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 601, + 481, + 632 + ], + "type": "text", + "content": "11. Jund, P., Sweeney, C., Abdo, N., Chen, Z., Shlens, J.: Scalable scene flow from point clouds in the real world. IEEE Robotics and Automation Letters 7(2), 1589-1596 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 633, + 481, + 665 + ], + "type": "text", + "content": "12. Kittenplon, Y., Eldar, Y.C., Raviv, D.: Flowstep3d: Model unrolling for self-supervised scene flow estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4114-4123 (2021)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 158, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 158, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 158, + 91, + 447, + 102 + ], + "type": "text", + "content": "SeFlow: A Self-Supervised Scene Flow Method in Autonomous Driving" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "13. Lang, I., Aiger, D., Cole, F., Avidan, S., Rubinstein, M.: Scoop: Self-supervised correspondence and optimization-based scene flow. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5281-5290 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 171 + ], + "type": "text", + "content": "14. Li, X., Kaesemodel Pontes, J., Lucey, S.: Neural scene flow prior. Advances in Neural Information Processing Systems 34, 7838-7851 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 171, + 480, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 171, + 480, + 203 + ], + "spans": [ + { + "bbox": [ + 133, + 171, + 480, + 203 + ], + "type": "text", + "content": "15. Li, X., Zheng, J., Ferroni, F., Pontes, J.K., Lucey, S.: Fast neural scene flow. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9878-9890 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 203, + 480, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 203, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 133, + 203, + 480, + 236 + ], + "type": "text", + "content": "16. Liu, J., Wang, G., Ye, W., Jiang, C., Han, J., Liu, Z., Zhang, G., Du, D., Wang, H.: Difflow3d: Toward robust uncertainty-aware scene flow estimation with diffusion model. arXiv preprint arXiv:2311.17456 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 236, + 480, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 236, + 480, + 278 + ], + "spans": [ + { + "bbox": [ + 133, + 236, + 480, + 278 + ], + "type": "text", + "content": "17. Mayer, N., Ilg, E., Hausser, P., Fischer, P., Cremers, D., Dosovitskiy, A., Brox, T.: A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4040-4048 (2016)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 278, + 480, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 278, + 480, + 311 + ], + "spans": [ + { + "bbox": [ + 133, + 278, + 480, + 311 + ], + "type": "text", + "content": "18. Menze, M., Geiger, A.: Object scene flow for autonomous vehicles. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3061-3070 (2015)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 311, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 311, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 133, + 311, + 480, + 342 + ], + "type": "text", + "content": "19. Mittal, H., Okorn, B., Held, D.: Just go with the flow: Self-supervised scene flow estimation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11177-11185 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 342, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 342, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 133, + 342, + 480, + 374 + ], + "type": "text", + "content": "20. Najibi, M., Ji, J., Zhou, Y., Qi, C.R., Yan, X., Ettinger, S., Anguelov, D.: Motion inspired unsupervised perception and prediction in autonomous driving. In: European Conference on Computer Vision. pp. 424-443. Springer (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 374, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 374, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 133, + 374, + 480, + 418 + ], + "type": "text", + "content": "21. Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A., Cournapeau, D., Brucher, M., Perrot, M., Duchesnay, E.: Scikit-learn: Machine learning in Python. Journal of Machine Learning Research 12, 2825–2830 (2011)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 418, + 480, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 418, + 480, + 461 + ], + "spans": [ + { + "bbox": [ + 133, + 418, + 480, + 461 + ], + "type": "text", + "content": "22. Pfreundschuh, P., Hendrikx, H.F., Reijgwart, V., Dubé, R., Siegwart, R., Cramar-iuc, A.: Dynamic object aware lidar slam based on automatic generation of training data. In: 2021 IEEE International Conference on Robotics and Automation (ICRA). pp. 11641-11647. IEEE (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 461, + 480, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 461, + 480, + 504 + ], + "spans": [ + { + "bbox": [ + 133, + 461, + 480, + 504 + ], + "type": "text", + "content": "23. Schmid, L., Andersson, O., Sulser, A., Pfreundschuh, P., Siegwart, R.: Dynablox: Real-time detection of diverse dynamic objects in complex environments. IEEE Robotics and Automation Letters (RA-L) 8(10), 6259 - 6266 (2023). https://doi.org/10.1109/LRA.2023.3305239" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 504, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 504, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 133, + 504, + 480, + 536 + ], + "type": "text", + "content": "24. Shen, Y., Hui, L., Xie, J., Yang, J.: Self-supervised 3d scene flow estimation guided by superpoints. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5271-5280 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 537, + 480, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 537, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 133, + 537, + 480, + 557 + ], + "type": "text", + "content": "25. Song, J., Lee, S.J.: Knowledge distillation of multi-scale dense prediction transformer for self-supervised depth estimation. Scientific Reports (18939) (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 557, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 557, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 133, + 557, + 480, + 601 + ], + "type": "text", + "content": "26. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., et al.: Scalability in perception for autonomous driving: Waymo open dataset. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2446-2454 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 601, + 480, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 601, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 133, + 601, + 480, + 633 + ], + "type": "text", + "content": "27. Tishchenko, I., Lombardi, S., Oswald, M.R., Pollefeys, M.: Self-supervised learning of non-rigid residual flow and ego-motion. In: 2020 international conference on 3D vision (3DV). pp. 150-159. IEEE (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 633, + 480, + 665 + ], + "type": "text", + "content": "28. Vedder, K., Khatri, I., Peri, N., Chodosh, N., Liu, Y., Hays, J.: Av2 2024 scene flow challenge announcement. https://wwwargonse.org/sceneflow.html (2024 Feb 25th)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Q Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 424 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "29. Vedder, K., Peri, N., Chodosh, N., Khatri, I., Eaton, E., Jayaraman, D., Ramanan, Y.L.D., Hays, J.: ZeroFlow: Fast Zero Label Scene Flow via Distillation. International Conference on Learning Representations (ICLR) (2024)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "type": "text", + "content": "30. Vedula, S., Rander, P., Collins, R., Kanade, T.: Three-dimensional scene flow. IEEE transactions on pattern analysis and machine intelligence 27(3), 475-480 (2005)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "type": "text", + "content": "31. Vidanapathirana, K., Chng, S.F., Li, X., Lucey, S.: Multi-body neural scene flow. In: 2024 International Conference on 3D Vision (3DV). pp. 126-136. IEEE (2024)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "text", + "content": "32. Wang, Z., Wei, Y., Rao, Y., Zhou, J., Lu, J.: 3d point-voxel correlation fields for scene flow estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 480, + 258 + ], + "type": "text", + "content": "33. Wei, Y., Wang, Z., Rao, Y., Lu, J., Zhou, J.: PV-RAFT: Point-Voxel Correlation Fields for Scene Flow Estimation of Point Clouds. In: CVPR (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 303 + ], + "type": "text", + "content": "34. Wilson, B., Qi, W., Agarwal, T., Lambert, J., Singh, J., et al.: Argoverse 2: Next generation datasets for self-driving perception and forecasting. In: Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks (NeurIPS Datasets and Benchmarks 2021) (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "text", + "content": "35. Wu, H., Li, Y., Xu, W., Kong, F., Zhang, F.: Moving event detection from lidar point streams. Nature Communications 15(1), 345 (2024)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 325, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 480, + 369 + ], + "type": "text", + "content": "36. Wu, W., Wang, Z.Y., Li, Z., Liu, W., Fuxin, L.: Pointpwc-net: Cost volume on point clouds for self-supervised scene flow estimation. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16. pp. 88-107. Springer (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "text", + "content": "37. Zhang, Q., Duberg, D., Geng, R., Jia, M., Wang, L., Jensfelt, P.: A dynamic points removal benchmark in point cloud maps. In: IEEE 26th International Conference on Intelligent Transportation Systems (ITSC). pp. 608-614 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 402, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 480, + 424 + ], + "type": "text", + "content": "38. Zhang, Q., Yang, Y., Fang, H., Geng, R., Jensfelt, P.: Deflow: Decoder of scene flow network in autonomous driving. arXiv preprint arXiv:2401.16122 (2024)" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 159, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 159, + 91, + 447, + 102 + ], + "type": "text", + "content": "SeFlow: A Self-Supervised Scene Flow Method in Autonomous Driving" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/See and Think_ Embodied Agent in Virtual Environment/0ba9ae38-8422-4b1f-90c0-ee0bf0ad307a_content_list.json b/2024/See and Think_ Embodied Agent in Virtual Environment/0ba9ae38-8422-4b1f-90c0-ee0bf0ad307a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..83859af33b5766ee618b0cac29e1ab6aad29c9eb --- /dev/null +++ b/2024/See and Think_ Embodied Agent in Virtual Environment/0ba9ae38-8422-4b1f-90c0-ee0bf0ad307a_content_list.json @@ -0,0 +1,2278 @@ +[ + { + "type": "text", + "text": "See and Think: Embodied Agent in Virtual Environment", + "text_level": 1, + "bbox": [ + 254, + 140, + 750, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhonghan Zhao $^{1, \\spadesuit}$ , Wenhao Chai $^{2, \\spadesuit}$ , Xuan Wang $^{1, \\spadesuit}$ , Boyi Li $^{1}$ , Shengyu Hao $^{1}$ , Shidong Cao $^{1}$ , Tian Ye $^{3}$ , and Gaoang Wang $^{1, \\dagger}$", + "bbox": [ + 238, + 210, + 761, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Zhejiang University", + "bbox": [ + 426, + 253, + 575, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 University of Washington", + "bbox": [ + 408, + 268, + 593, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Hong Kong University of Science and Technology (GZ)", + "bbox": [ + 310, + 282, + 691, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Large language models (LLMs) have achieved impressive progress on several open-world tasks. Recently, using LLMs to build embodied agents has been a hotspot. This paper proposes STEVE, a comprehensive and visionary embodied agent in the Minecraft virtual environment. STEVE comprises three key components: vision perception, language instruction, and code action. Vision perception involves interpreting visual information in the environment, which is then integrated into the LLMs component with agent state and task instruction. Language instruction is responsible for iterative reasoning and decomposing complex tasks into manageable guidelines. Code action generates executable skill actions based on retrieval in skill database, enabling the agent to interact effectively within the Minecraft environment. We also collect STEVE-21K dataset, which includes $600+$ vision-environment pairs, 20K knowledge question-answering pairs, and $200+$ skill-code pairs. We conduct continuous block search, knowledge question and answering, and tech tree mastery to evaluate the performance. Extensive experiments show that STEVE achieves at most $1.5 \\times$ faster unlocking key tech trees and $2.5 \\times$ quicker in block search tasks.", + "bbox": [ + 259, + 330, + 743, + 580 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Open-world embodied agent $\\cdot$ Multimodal pre-training $\\cdot$ Large language model", + "bbox": [ + 259, + 593, + 740, + 621 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 681, + 375, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Designing agents that demonstrate intelligent behavior and adaptability in open-world settings has been a longstanding and significant challenge in the field of artificial intelligence [13,14,26,46,70]. However, recent progress in the development of large language models (LLMs) [5,55] has exhibited their potential as versatile, general-purpose assistants. Recent innovations in agent design [57,59,69,77]", + "bbox": [ + 212, + 712, + 787, + 789 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\spadesuit$ Equal contribution: zhaozhonghan@zju.edu.cn, wchai@uw.edu, xu-anw@zju.edu.cn.", + "bbox": [ + 230, + 797, + 785, + 825 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author: gaoangwang@intl.zju.edu.cn.", + "bbox": [ + 250, + 825, + 604, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8e20506f446eeac2546e04d44592fb05bd49ec21a9a6b5405d67407105162973.jpg", + "image_caption": [ + "Fig. 1: STEVE integrates the three parts: Vision Perception, Language Instruction, and Code Action, supported closely by our proposed STEVE-21K. It demonstrates commendable performance on Continuous Block Search, Knowledge QA, and Tech Tree Mastery." + ], + "image_footnote": [], + "bbox": [ + 232, + 143, + 774, + 260 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "have effectively harnessed these advanced LLMs, tapping into their extensive world knowledge and reasoning abilities. This development has paved the way for agents, that are autonomously driven, to formulate and implement strategies and actions across a diverse array of skills and tasks in diverse open-world environments.", + "bbox": [ + 212, + 364, + 787, + 439 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In many open-world settings, like Minecraft, contemporary agents predominantly use LLMs for their textual interactions. However, this reliance on text for communication poses considerable limitations in their interactions within these worlds including low-level regrading cases [24, 56, 65, 66]. Minecraft, with its expansive and interactive sandbox environment [17, 20], demands a variety of skills from agents, ranging from crafting basic items to executing complex tasks. Yet, agents driven by LLMs often generate unpredictable outputs. The effectiveness of their interactions is largely contingent on meticulously crafted prompts [23], designed to align the LLM's understanding with the environmental context and the intended objectives. This process of prompt engineering is not only laborious but also fails to meet the goal of fostering autonomous, self-directed agents. Furthermore, textual communication has its limitations in naturally conveying certain concepts of the world, like crafting recipes, which are often more effectively communicated through vision.", + "bbox": [ + 212, + 443, + 787, + 654 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Players have the distinct capability to assimilate and convey information using both visual and textual channels, significantly enhancing our interactions with the world around us. Yet, the integration of LLM-based agents with multimodal inputs in open-ended environments remains an under-explored area. STEVE in STEVE-Series [71-73], is named after the protagonist of the game \"Minecraft.\" It is our proposed framework to build an embodied agent based on the vision model and LLMs within an open world, as illustrated in Fig. 1. STEVE harnesses a vision model to perceive its surroundings visually, coupled with an LLM to strategize and plan actions. This model represents a leap forward in agent design, combining these two input modes, vision, and text, to offer a more nuanced and comprehensive understanding of the environment and practical and executable skills.", + "bbox": [ + 212, + 657, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Z. Zhao et al.", + "bbox": [ + 271, + 114, + 362, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our key contributions are outlined as follows:", + "bbox": [ + 238, + 146, + 570, + 161 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose STEVE, an embodied agent in virtual environment, consists of vision perception, language instruction, and code action, achieving $1.5 \\times$ faster unlocking of key tech trees and is $2.3 \\times$ quicker in block search tasks compared to previous state-of-the-art methods.", + "- We present STEVE-7B/13B, a series of large language model obtained by fine-tuning with Minecraft knowledge question-answering pairs from Llama2-7B/13B.", + "- We collect STEVE-21K dataset, including $600+$ vision-environment pairs, 20K knowledge question-answering pairs, and $200+$ skill-code pairs, for justifying the effective performance of STEVE." + ], + "bbox": [ + 227, + 164, + 785, + 315 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 339, + 395, + 356 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Intelligent Agent in Minecraft", + "text_level": 1, + "bbox": [ + 215, + 372, + 509, + 387 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As an open-ended sandbox game, Minecraft has always been an ideal setting for testing the performance of intelligent agents [21, 25]. The agents are required to autonomously perform various tasks in Minecraft, such as chopping trees, crafting tools, and mining diamonds. At the beginning, much of the works focus on exploring reinforcement learning [32, 33, 42, 50] or imitation learning [2, 3], without satisfactory performance. VPT [3] and MineDojo [17] collect internet-scale datasets for their model pre-training. More specifically, VPT offers the exciting possibility of directly learning to act during video pre-training and using these learned behavioral priors as extremely effective exploration priors for reinforcement learning. Yet, recent works found that the pre-trained LLMs could serve as a strong \"mind\" that provides planning ability to the agents. Voyager [57] leverages GPT-4 [44] as both a high-level planner and a low-level action code generator. Plan4MC [69] proposes a skill graph pre-generated by the LLMs. DEPS [59], an interactive planning method based on LLMs, addresses multi-step reasoning issue in open-world planning. GITM [77] develops a set of structured actions and leverages LLMs to generate action plans for the agents to execute, achieving impressive results in various tasks.", + "bbox": [ + 212, + 397, + 787, + 654 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 Embodied Multimodal Model", + "text_level": 1, + "bbox": [ + 215, + 676, + 504, + 691 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Embodied agent operates within various environments by synthesizing sensory perceptions and physical actions supported by computational intelligence. This synthesis enables the agent to undertake a variety of tasks, achieving specific objectives. Its key areas of application are diverse, including Navigation [6, 16, 27, 43, 61, 68], Embodied Question Answering [10, 11, 67], Active Visual Tracking [37, 38, 74, 75], and Visual Exploration [7, 12, 35]. The field is evolving rapidly with the development of Large Language Models (LLMs) [51] and Multimodal LLMs (MLLMs) [1, 9, 18, 19, 28–30, 34, 39, 41, 54, 58, 64, 76], integrating multiple modalities for more effective processing. A prime example of this innovation is", + "bbox": [ + 212, + 704, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "STEVE", + "bbox": [ + 676, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1b252c5e77571c428318903d9dd735b64f6b4dec4c2328c8baa1f1769e6e9e2f.jpg", + "image_caption": [ + "Fig. 2: STEVE framework. The Vision Perception part takes images or videos, encodes them into tokens, and combines them with Agent State and Task tokens as input. The STEVE-13B in the Language Instruction part is used for automatic reasoning and task decomposition, and it calls the Skill Database in the form of the Query to output code as action." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 782, + 256 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "PaLM-E [15], a sophisticated multimodal model with 562B parameters, adept at a broad spectrum of embodied tasks and demonstrating exceptional capabilities in visual reasoning.", + "bbox": [ + 212, + 366, + 787, + 412 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Large Language Model with Equipped Tools", + "text_level": 1, + "bbox": [ + 214, + 436, + 627, + 454 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While Large Language Models (LLMs) demonstrate impressive skill in tackling novel tasks via prompt-based instructions, they often face challenges in areas where simpler models or tools excel, like mathematical calculations or identifying palindromes. However, LLMs' potential is significantly expanded when integrated with other modality-specific models, such as those for vision or audio, enabling multi-modal capabilities [4,36]. Innovations like Toolformer [47] demonstrate LLMs' self-learning to utilize tools through finetuning with extensive API call samples. Visual ChatGPT [62] extends this by integrating various visual foundation models, facilitating interactive user experiences with ChatGPT. Similarly, HuggingGPT [48] presents a framework that harnesses LLMs to link diverse models from Hugging Face for task resolution. AutoGPT [49] is an open-source application that broadens GPT-4's capabilities with internet access, memory management, and plug-ins. The recent introduction of MovieChat [52,53] brings a memory mechanism to MLLM, enhancing its performance in video understanding tasks. Furthermore, LLMs can be used for goal planning, analogous to language translation [63]. This evolving landscape suggests that tool-equipped LLMs could forge a new paradigm in AI solution design.", + "bbox": [ + 212, + 465, + 787, + 723 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method: STEVE", + "text_level": 1, + "bbox": [ + 214, + 747, + 416, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Overview", + "text_level": 1, + "bbox": [ + 214, + 782, + 341, + 796 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Figure 2, STEVE $\\mathcal{F}$ is an LLM-based multi-modal autonomous system for embodied agents in Minecraft, which can use visual status $X^v$ in the", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Z. Zhao et al.", + "bbox": [ + 271, + 114, + 362, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "form of images or videos and agent state $X^s$ to manage and execute complex task $X^t$ for executive code action $\\mathbf{a}^c$ :", + "bbox": [ + 212, + 146, + 782, + 175 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {a} ^ {c} = \\mathcal {F} \\left(X ^ {v}, X ^ {s}, X ^ {t}\\right) = \\mathcal {A} ^ {c} \\left(\\mathcal {I} ^ {l} \\left(\\mathcal {P} ^ {v} \\left(X ^ {v}, X ^ {s}, X ^ {t}\\right)\\right)\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 331, + 191, + 785, + 208 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where STEVE $\\mathcal{F}$ integrates Visual Perception $\\mathcal{P}^v$ into the Language Instruction $\\mathcal{I}^l$ with large language model (LLM) and combines them with Code Action $\\mathcal{A}^c$ , a skill retrieval method to execute actions, details are demonstrated as follows.", + "bbox": [ + 212, + 213, + 782, + 258 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Vision Perception $\\mathcal{P}^v$ (Section 3.2); a vision encoder to interpret visual information of the environment, such as blocks and entities and a text tokenizer for agent state $s$ and task $t$ .", + "- Language Instruction $\\mathcal{I}^l$ (Section 3.3): a powerful language model system fine-tuned specifically for Minecraft content using LLaMA2-13B [18]. This model enables adaptive interaction for iterative reasoning and step-by-step decomposition.", + "- Code Action $\\mathcal{A}^c$ (Section 3.4): a skill retrieval method based on our skill-code database." + ], + "bbox": [ + 225, + 271, + 782, + 404 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Vision Perception $\\mathcal{P}^v$", + "text_level": 1, + "bbox": [ + 214, + 428, + 437, + 444 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The vision perception part includes a vision encoder $E^v$ and a text tokenizer $T$ , which converts the visual status $X^v$ , agent state $X^s$ , and task $X^t$ to the text-space tokenizer representation $\\mathbf{Y} = \\{Y^v, Y^s, Y^t\\}$ :", + "bbox": [ + 212, + 453, + 782, + 500 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Y} _ {i} = \\mathcal {P} ^ {v} \\left(X _ {i} ^ {v}, X _ {i} ^ {s}, X _ {i} ^ {t}\\right), \\mathcal {P} ^ {v} = \\left\\{E ^ {v}, T \\right\\}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 370, + 513, + 784, + 530 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the vision encoder $E$ , the visual branch of EfficientFormer [31], encodes visual status $X_{i}^{v}$ at each step $i$ into visual tokens $Y^{v} = \\{y_{1}^{v}, y_{2}^{v}, \\dots, y_{n}^{v}\\} \\in \\mathbb{R}^{n \\times d}$ , where $n$ denotes the number of visual tokens $y^{v}$ and $d$ is the dimensionality of each token. The text tokenizer converts agent state $X_{i}^{s}$ and task $X_{i}^{t}$ at each step $i$ in the form of text into textual tokens $Y^{s}$ and $Y^{t}$ . Note that visual tokens are amalgamated with textual tokens representing the agent's current state (e.g., health, inventory, etc.) and the task description. This is accomplished using a tokenizer that maps state variables and task parameters to a tokenized form. The resultant unified token set serves as a comprehensive representation of the current situational context.", + "bbox": [ + 212, + 536, + 787, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3 Language Instruction $\\mathcal{I}^l$", + "text_level": 1, + "bbox": [ + 214, + 709, + 460, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Language Instruction $\\mathcal{I}^l = \\{\\mathcal{P}_l, \\mathcal{C}_r, \\mathcal{C}_u, \\mathcal{D}_s\\}$ , which consists of Planner $\\mathcal{P}_l$ , Critic $\\mathcal{C}_r$ , Curriculum $\\mathcal{C}_u$ , and Descriptor $\\mathcal{D}_s$ , i.e., four independent LLM-based agents with different functions. They formulate high-level task guidelines, refine strategies through feedback, facilitate continuous learning and adaptation through a curriculum of complex tasks, and finally decompose strategic guidelines into the executable low-level textual action step $\\mathbf{a}^s$ , $\\{\\mathbf{a}^s \\sim \\mathcal{I}^l(\\cdot|\\mathbf{Y}_i)\\}_{i=0}^N$ towards efficient completion of the task.", + "bbox": [ + 212, + 733, + 782, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "STEVE", + "bbox": [ + 676, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that the Planner $\\mathcal{P}_l$ , Critic $\\mathcal{C}_r$ , Curriculum $\\mathcal{C}_u$ and Descriptor $\\mathcal{D}_s$ are based on STEVE-7B/13B, a powerful language model derived from LLaMA-2-13B [18], fine-tuned specifically on Minecraft-related content from the STEVE-21K. This model's expertise covers a broad spectrum of game-specific knowledge areas on worlds, entities, player mechanics, survival, and even game practical experience. Finally, they have the performance of their different functions:", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Planner $\\mathcal{P}_l$ : formulating comprehensive guidelines and executive plans that align with the overarching objectives of the task.", + "- Critic $\\mathcal{C}_r$ : evaluating the planner decisions, providing feedback that can refine strategies.", + "- Curriculum $\\mathcal{C}_u$ : facilitating continuous learning and adaptation for action agents by engaging with a series of progressively complex tasks.", + "- Descriptor $\\mathcal{D}_s$ : distilling the extensive data into a concise summary, making it more manageable and interpretable." + ], + "bbox": [ + 225, + 250, + 782, + 369 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Iterative reasoning. The STEVE-13B receives a stream of tokens that encode the current visual scene, the agent's state, and the task's textual description. STEVE-13B interprets this rich context to undertake complex reasoning. The model initiates the reasoning process by constructing a series of high-level strategies that outline the pathway to task completion. Considering all gameplay elements, the reasoning mechanism is akin to an experienced player who can visualize the end game and chart a course to victory. This approach ensures the plans are reactive and proactive, allowing the agent to anticipate and mitigate future challenges. However, most strategies are high-level and abstract therefore, they often require step-by-step decomposition to derive executable guidelines.", + "bbox": [ + 212, + 386, + 787, + 537 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Decomposition. The decomposition process makes the complex strategies breakdown into simple, low-level guidelines that can be directly mapped to actions in Minecraft. It is similar to how a high-level command like \"build a shelter\" is divided into actionable instructions like \"collect wood\", \"craft planks\", and \"place blocks\". The granular steps are structured to provide explicit instructions that the game engine can readily interpret. This systematic breakdown from high-level reasoning to low-level actionable steps is the hallmark of the STEVE system, enabling the embodied agent to interact with the Minecraft environment in a meaningful and goal-oriented manner. Through this intricate process of reasoning and decomposition, STEVE embodies the cognitive capabilities required for sophisticated task management in virtual environments.", + "bbox": [ + 212, + 551, + 787, + 720 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Curriculum learning with memory. We draw inspiration from the lifelong learning strategy utilized in many reinforcement learning problems [57] in both closed-world and open-world settings. We start by creating a set of tasks that serve as a curriculum for agents to explore the environment. During this process, STEVE generates plans, interacts with the surroundings, learns from mistakes, and stores all these experiences in memory. Next, we evaluate STEVE on various tasks after this learning stage. Consequently, STEVE can produce better plans with its", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Z. Zhao et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "memory teaming up with the planning experiences. We use this as the default setting for all tasks in our experiments.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Continuous learning with summarization. We've observed that the learning process, where the memory is being filled, can continue throughout the gameplay. The agent can gradually acquire more skills as the gameplay progresses and more experiences are gained. However, as the memory gets larger, it becomes difficult to understand the game's situation and interact with the game slowly. To tackle these challenges, we have implemented the Chain of Summarization method [40]. By finding better references, we can improve our ability to handle challenging tasks, such as \"Diamond Tool\" in the tech tree, including obtaining materials and creating diamond tools. This will lead to a higher success rate, as shown in Section 5.3. Additionally, curriculum learning with memory allows for in-context lifelong learning without needing gradient updates.", + "bbox": [ + 212, + 196, + 787, + 363 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Code Action $\\mathcal{A}^c$", + "text_level": 1, + "bbox": [ + 214, + 388, + 395, + 402 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The code action part $\\mathcal{A}^c$ is the execution phase, where the STEVE system converts planned, decomposed guidelines into concrete actions within the Minecraft environment. This process leverages a specialized skill database that pairs code snippets with their descriptions and relevant metadata, encoded as vectors $\\mathbf{v}^{\\mathbf{s}}$ for efficient retrieval. The transition from language instruction to executable code is achieved through the retrieval process $\\mathcal{R}$ :", + "bbox": [ + 212, + 419, + 787, + 510 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {R} (\\mathbf {q}, \\mathbf {v}) = \\sigma (\\mathbf {q}, \\mathbf {v}), \\mathbf {q} = E ^ {q} \\left(\\mathbf {a} ^ {\\mathbf {s}}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 392, + 527, + 784, + 545 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $E^q$ and $\\sigma$ are query encoding and cosine similarity matching. Each low-level textual action step $\\mathbf{a}^s$ derived from the Language Instruction is encoded into a query $\\mathbf{q}$ . This encoding captures the essence of the action to be performed in a format that the skill database can understand. Once the queries $\\mathbf{q}$ are encoded, the system computes similarities between the query vectors and the code snippets vectors $\\mathbf{v}$ stored in the database to determine which skill best matches the required action.", + "bbox": [ + 212, + 554, + 787, + 659 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.5 Training", + "text_level": 1, + "bbox": [ + 214, + 686, + 333, + 702 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To reduce the training overhead to a certain extent, we adopt a two-stage training method, the first stage being a warm-up on the question-answering pairs of STEVE-21K to ensure a certain degree of instruction capability. The second stage is simulated in the environment, and the Expert LLM $\\mathcal{E}_p$ integrated with GPT-4 [44] generates instructions for the same situation to modify our model.", + "bbox": [ + 212, + 715, + 787, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In training, we use either a single-round conversation $\\{X^Q,X^A\\}$ or a multiround conversation $\\{X_i^Q,X_i^A\\}_{i\\leq N}$ , where $N$ is the total number of rounds in the conversation. We use the negative log-likelihood objective over the prediction", + "bbox": [ + 212, + 792, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "STEVE", + "bbox": [ + 676, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tokens with training and finetuning:", + "bbox": [ + 215, + 146, + 478, + 161 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (\\theta) = - \\sum_ {j = 1} ^ {L} \\log \\mathcal {F} _ {\\theta} \\left(Y _ {j} \\mid X ^ {v}, \\hat {Y} _ {1: j - 1}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 375, + 175, + 785, + 217 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $Y$ and $\\hat{Y}$ refer to the non-vision input and target token sequences, $\\theta$ represents the model parameters, and $L$ represents the length of the target sequence. The vision input $X^v$ varies depending on the input step. We only consider the answer tokens $X_A$ when computing the loss to ensure that the model focuses on generating coherent responses.", + "bbox": [ + 212, + 231, + 782, + 306 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Offline warm-up. In the first stage, we finetune the STEVE-7B/13B only on the single-round question-answering pairs of STEVE-21K to ensure a certain degree of instruction capability, as shown in Tab. 3.", + "bbox": [ + 212, + 306, + 782, + 352 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Online finetuning. In the second stage, we simulate the warmed-up STEVE-7B/13B in Minecraft, and both train the vision encoder $E^v$ and finetune the STEVE-7B/13B simultaneously.", + "bbox": [ + 212, + 353, + 784, + 398 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "It is necessary to obtain the environment information to train the vision encoder, denoted as $l = R_{T}(X^{v})$ , where $l$ is the label of seen blocks and entities, $R_{T}$ is the Ray Tracing method [60] to eliminate all content outside the agent's perspective screen. We train the vision encoder $E^{v}$ once we have obtained the environment information. To finetune the warmed-up STEVE-7B/13B in simulation, we give both the STEVE-7B/13B and the Expert LLM the same input, and we consider the output of the Expert LLM as the ground truth label.", + "bbox": [ + 212, + 398, + 784, + 503 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "After 5,000 simulations, we collect all the context information from successful runs, including sequences of vision and question-answering chat flow as the Vision-Environment part of STEVE-21K, as mentioned in Section 4.", + "bbox": [ + 212, + 505, + 785, + 550 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Dataset: STEVE-21K", + "text_level": 1, + "bbox": [ + 215, + 577, + 460, + 593 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Fig. 3, STEVE-21K has been meticulously compiled, featuring a diverse collection of Vision-Environment pairs, Question-Answering pairs, and a Skill-Code database.", + "bbox": [ + 212, + 612, + 782, + 655 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Vision-Environment pairs contain 600 pairs of first-person perspective videos from Minecraft gameplay across six different terrains (including forest, desert, coastal, etc.), along with corresponding environmental block entities in the field of vision and context information for each timestamp. Additionally, all pairs are oriented around executing the skill-related task supported by Skill-Code part mentioned in Section 4. We employ the STEVE-7B/13B model to enable robots to plan and execute actions autonomously based on tasks defined by human supervisors. We record the video of the agent operation, the environment information, and all the corresponding chatflow.", + "bbox": [ + 212, + 657, + 784, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Question-Answering pairs contain 20K question-answering pairs from the Minecraft-Wiki and Reddit corpus across six data types partly sourced from [17]. The pairs are organized into instruction, input, and output triplets and used to", + "bbox": [ + 212, + 795, + 785, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Z. Zhao et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4f0778d76d1ad6c52cc0e0aedd4767906814279c72cb3fab3da16fa83f00fe92.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 143, + 671, + 243 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Skill-Code", + "text_level": 1, + "bbox": [ + 687, + 148, + 746, + 160 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Human Player", + "bbox": [ + 687, + 166, + 761, + 174 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/3204dda95c0702225d88f8d15a4b7828f65ac2d6679bd4ed5068a39a667c6f58.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 715, + 186, + 740, + 205 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "prompt", + "bbox": [ + 741, + 210, + 769, + 217 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/1e1670b491385802b28f5fe1e2fd9c1f605e4fa7f9fed088c8ab601eccfeb9b3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 717, + 226, + 740, + 243 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "GPT", + "bbox": [ + 746, + 229, + 761, + 237 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "output", + "bbox": [ + 717, + 248, + 767, + 258 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "code snippets", + "bbox": [ + 707, + 268, + 758, + 276 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/fec95c1c56cf7cfcb11205fcee473edbe1b0bd7a1c8b3d30f066c14f173805e6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 715, + 289, + 741, + 299 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "execute", + "bbox": [ + 741, + 285, + 769, + 292 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/02bd1387363039c3fd997487d4a4b920c9dd4bda4fc10bbffd4b6f2b5933a415.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 715, + 300, + 743, + 320 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "check and revise", + "bbox": [ + 696, + 325, + 764, + 333 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Question-Answering", + "text_level": 1, + "bbox": [ + 233, + 253, + 344, + 263 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/1f57fa5bba76ddfef6a911c46b7ce601b83e13e345e685ec8366aa261fa44900.jpg", + "image_caption": [ + "Wikipedia" + ], + "image_footnote": [], + "bbox": [ + 233, + 287, + 305, + 323 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The crafting menu in Minecraft, showing the crafting origins of a clone.", + "bbox": [ + 233, + 323, + 300, + 329 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "showing the eating recipe or a special taste as well as some other blocks and", + "bbox": [ + 233, + 329, + 299, + 335 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/60b81a3852d8af7061f20dd0b31c16210869d84a7930ee35b04a45fb03a288ee.jpg", + "image_caption": [ + "Reddit" + ], + "image_footnote": [], + "bbox": [ + 320, + 289, + 385, + 339 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/a54b823eedd10842692d2794e3523dd8248c7e2eced360f5ce1571ae17dfd7c9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 406, + 263, + 429, + 282 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "ChatGPT", + "bbox": [ + 431, + 268, + 470, + 277 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/2831b8b1e4fd69d911915dbd4407eae73d885f7845e3e2f70795cbf433d150d3.jpg", + "image_caption": [ + "Fig.3: STEVE-21K collection pipeline. In the Vision-Environment section, STEVE-13B plays the game according to specified tasks defined by the human player, collecting visual information through prismarine-viewer and capturing environmental information from the screen using Ray Tracing [60]. Note that the language instruction task is also performed during the collection phase. We simultaneously record and save the chat flow from the reasoning and decomposition stages. In the Question-Answering section, we obtain information from the Minecraft-Wiki and Reddit forums and use GPT-3.5 to clean the data into single-round QA pairs. In the Skill-Code section, we use GPT-3.5 combined with the human player's code to synthesize code snippets and then check and revise them in the game environment." + ], + "image_footnote": [], + "bbox": [ + 403, + 340, + 416, + 345 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Clean", + "bbox": [ + 429, + 289, + 478, + 299 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Construct", + "bbox": [ + 447, + 304, + 483, + 311 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Check", + "bbox": [ + 452, + 323, + 478, + 329 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Single-round QA pairs", + "bbox": [ + 560, + 267, + 643, + 275 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. World & Entities", + "2. Player Mechanics & Survival", + "3. Knowledge & Discovery", + "4. Resources & Crafting", + "5. Tools & Utilities", + "6. Miscellaneous" + ], + "bbox": [ + 539, + 280, + 661, + 330 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "train STEVE-13B. The GPT-3.5 [5] is employed to derive meaningful single-round question-answer pairs, and LoRA [22] is incorporated during the fin-tuning process for efficient resource allocation.", + "bbox": [ + 212, + 525, + 784, + 570 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Skill-Code pairs contain 210 skill execution scripts with descriptions, covering 8 skill types including collecting, crafting, exploration etc. The code part is collected by manual coding. We use GPT-3.5 [5] to describe all codes and utilize langchain vecordb to give all pairs a database vector.", + "bbox": [ + 212, + 570, + 785, + 631 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 214, + 652, + 374, + 669 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 214, + 681, + 426, + 696 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We train STEVE-7B/13B, which is finetuned from LLaMA-2 [18] with the Question-Answering pair in STEVE-21K dataset for warm-up and simulation context data from successful runs. We use LoRA [22] for finetuning process. Note that the process is to adjust STEVE-13B to work on correct simulation knowledge while remaining adapted to visual perception. In the text part, we set all temperatures to 0 except for the task proposal, which uses 0.9 to encourage task diversity. The vision unit is based on EfficientFormerV2-S0 [31], which is trained on the Vision-Environment part of our STEVE-21K dataset. Our simulation environment is built on top of MineDojo [17] and leverages Mineflayer [45].", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "STEVE", + "bbox": [ + 676, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/5d30bf45360efd026ed4a76d83aeb386e7564905498c795635ad07fe0ea13126.jpg", + "image_caption": [ + "Fig. 4: Example of Vision-Environment pairs. It represents the data format of the Vision-Environment pairs in our STEVE-21K dataset: including visual signals, environmental information, Chat Log, Context QA pairs, and planning in actual tasks." + ], + "image_footnote": [], + "bbox": [ + 236, + 146, + 769, + 467 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We use GPT-4-0613 for all GPT-4 models used in Voyager [57] and code generation tasks.", + "bbox": [ + 212, + 545, + 787, + 575 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.2Baselines", + "text_level": 1, + "bbox": [ + 215, + 595, + 339, + 609 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "As no vision-based LLM-driven agents are immediately operable in Minecraft, we selected several algorithms as baselines that extract information from a system's backend, differing significantly from real-world applications.", + "bbox": [ + 212, + 618, + 787, + 666 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "AutoGPT [49] is an NLP automation tool that decomposes a high-level goal into executable subgoals in MineDojo, aligning with our experimental framework. Our setup, AutoGPT, powered by GPT-4 [44], processes agent states, environment feedback, and execution errors to manage subgoal execution.", + "bbox": [ + 212, + 676, + 787, + 739 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Voyager [57] relies solely on textual grounding for perception and features a long-term procedural memory with a hierarchical library of code-based procedures, allowing the integration of simple skills into complex behaviors. Proficient in environment exploration and tech tree mastery, Voyager uses GPT-4 [44] for processing background text in embodied agents, with a lesser emphasis on visual perception.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Z. Zhao et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/ac1d1fd6c6392197cc1ecae407139e6f255a7c8c499915b10d7c9a862bad374d.jpg", + "table_caption": [ + "Table 2: Comparison on tech tree mastery task. The values presented are in fractions, representing successful trials out of three attempts. A score of $0/3$ signifies the method's inability to progress within the tech tree after a maximum of 160 prompting iterations. The reported numbers denote the average iterations across three trials. Lower iteration values indicate higher efficiency of the respective method." + ], + "table_footnote": [], + "table_body": "
MethodWooden ToolStone ToolIron ToolDiamond Tool
AutoGPT [49]92 ± 72 (3/3)94 ± 72 (3/3)135 ± 103 (3/3)N/A (0/3)
Voyager [57]6 ± 2 (3/3)11 ± 2 (3/3)21 ± 7 (3/3)102 (1/3)
STEVE4 ± 1 (3/3)8 ± 1 (3/3)15 ± 2 (3/3)106 ± 12 (3/3)
", + "bbox": [ + 222, + 226, + 781, + 309 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/07be74d97082cd27390d15e42a08bd14c3302aa14b1273954adad3044baa5e09.jpg", + "table_caption": [ + "Table 3: Quantitative comparison on knowledge question and answering task. Questions, model-generated responses, and ground truth inputs are evaluated in GPT-4 [5], Claude-2 [8] and human blind rating rated on a scale of 0 to 10; The scores above are the average of them. Higher scores indicate greater alignment of the generated answers with the ground truth. Wld., Ent., Mech., Surv., Know., Disc., Res., Craft., Tools, Util., Miscell. stand for World, Entities, Player Mechanics, Survival, Knowledge, Discovery, Resources, Crafting, Tools, Utilities and Miscellaneous." + ], + "table_footnote": [], + "table_body": "
MethodWld. & Ent.Mech. & Surv.Know. & Disc.Res. & Craft.Tl. & Util.Miscell.Overall
Llama2-7B6.446.686.586.426.806.966.56
Llama2-13B6.936.956.776.776.986.646.89
STEVE-7B7.997.887.847.957.937.827.94
STEVE-13B8.148.138.038.158.127.728.12
GPT-48.068.078.077.928.098.218.04
", + "bbox": [ + 220, + 436, + 781, + 529 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.3 Evaluation Results", + "text_level": 1, + "bbox": [ + 215, + 561, + 418, + 575 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Continuous block search. As shown in Tab. 1, we experiment with block-searching tasks to assess the agent's exploratory capabilities and proficiency in locating specified blocks. Diamond blocks are placed at every 16-block interval across the land map. The agent's objective is to identify as many blocks as possible within the fewest iterations, which indicates the method's efficiency.", + "bbox": [ + 212, + 625, + 495, + 773 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "As shown in Fig. 5, enriching information through visual perception significantly enhances the efficiency of search and exploration tasks, leading to more effective world exploration.", + "bbox": [ + 212, + 773, + 785, + 819 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/cb37e65ab50b47f4cf8c67bc3cd1a0e979e14341b0a2f1939d0ab5f70c0c39c9.jpg", + "table_caption": [ + "Table 1: Comparison on continues block search task. # Iters stand for average iterations to find 10 diamonds (max 100). # Blocks stand for average diamonds found in 100 iterations." + ], + "table_footnote": [], + "table_body": "
Method# Iters (↓) # Blocks (↑)
AutoGPT [49]N/A 7
Voyager [57]35 26
STEVE14 67
", + "bbox": [ + 500, + 699, + 787, + 773 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "STEVE", + "bbox": [ + 676, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/c3ffbd171cdc51eefbe4bfce50cd9a15ff960f3abcf3684c6d35aa8c13925cc9.jpg", + "image_caption": [ + "Fig.5: Schematic of the continuous block search task. We capture an asynchronous segment with each method 30 iterations from the experiments for illustration. The reason we choose diamond blocks is that they are not naturally occurring in the given context, making them easily distinguishable from other blocks." + ], + "image_footnote": [], + "bbox": [ + 274, + 143, + 727, + 470 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Knowledge question and answering. Using a validation dataset, we established a question-answering database to evaluate our model's performance on Minecraft-related queries. Responses from each model are rated blindly by GPT-4, Claude2, and human participants based on accuracy, relevance, and detail. Responses are initially checked for accuracy, and the comprehensive evaluation yields a score ranging from 0 to 10, with higher scores indicating superior performance.", + "bbox": [ + 212, + 582, + 787, + 674 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "As detailed in Tab. 3, we tested the instructional capabilities of various LLM models using the STEVE-21k dataset, divided into 18,622 training and 1,000 testing samples. The STEVE-7B and STEVE-13B models surpassed LLaMA2 in all metrics, with STEVE-13B scoring the highest at 8.12, demonstrating its exceptional understanding of Minecraft-related queries. This indicates that STEVE models, optimized specifically for Minecraft content, perform better in knowledge-intensive tasks.", + "bbox": [ + 212, + 681, + 787, + 787 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Our results confirm that larger, domain-specifically tuned models like STEVE-13B outperform broader models like GPT-4, underscoring the benefit of specialized fine-tuning for domain-specific applications.", + "bbox": [ + 212, + 794, + 789, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Z. Zhao et al.", + "bbox": [ + 271, + 114, + 362, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/e91ccdc799cd7cd93fb52b9144b461282a4db2a37877cbcce4f3d97d2dfbe61a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 223, + 146, + 357, + 205 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/f95ac2622f513174bd27eba737e9765cde210570ec63862fa79bdf9bab14841d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 146, + 500, + 205 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/fc15d8b43a6c6a219405b1d97201503701202651a2f7d734e028fe522ba6484d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 146, + 640, + 205 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/826c2d76a5225d8c324005a52ac9cc6da89c423df693190dee59d07b3648129f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 146, + 782, + 205 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/e8070962c9f4a31a7677295ef27366233d4975df507973f088419d805fa02cbb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 223, + 207, + 357, + 229 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/3e28f748d3988cb1973af8c1df6eec5e33e4da8bd1276a7727972f2a9c4f86f2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 207, + 500, + 229 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/2324a70c27db609ef3d1411c5025133c864e153dfe60253efcb06a9a55ab6a70.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 208, + 640, + 229 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/087df738ce9447c33074cd6826cefb2d7fc946fda7f4b7ca8604f99ae1bf4fbb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 208, + 782, + 229 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/ae46503e694aa75d149810b04b5403c9befffb96e2a4e08c74e63134dee5111c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 222, + 229, + 357, + 276 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/f74fe55769e23c2d01404e6b275ba8a4ca5a81d9da5594903fdbce86fce42076.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 229, + 500, + 286 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/617f7a9b756b06ff25e43a44bf5948fce18fa933315ddf8f864628d7b0fd925b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 229, + 640, + 285 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/e609be31ebf1b0024d4118fdfa22b57c753d09ccbc627e5d3d163aa183cfb842.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 229, + 782, + 286 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/c5b040852d0693814ed8f16ca4a63c4e0d6452a82e506ae37a75449c814516e1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 222, + 286, + 357, + 330 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/05047e68461b01b191892fc16bd6bc59e36da72c351c8660b722cb198b3583e1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 286, + 500, + 330 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/96b586df6b7667fa9e0203dd7313c6411d9b45c5380c63ca50ecf4cb29f4e256.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 286, + 640, + 330 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/630b50962ab2d34f205431369fa48fea883ec201012cebf11e9bea5d2d0acbcd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 286, + 782, + 330 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/508641d0d956ba9a170be8b6d37737357a9c553eeabafeeb8f70f6333d062b04.jpg", + "image_caption": [ + "Fig. 6: Qualitative comparison on knowledge question and answering tasks. Green (Red) highlights the correct or good (wrong or bad) answer. Blue indicates the suboptimal answer. Grey indicates the meaningless answer." + ], + "image_footnote": [], + "bbox": [ + 222, + 330, + 357, + 387 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/c30134ec6506fd05cc9e003840ad135900d7545f3bf00ee52d3a61549895794d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 330, + 500, + 387 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/e55b903770769d2184e50c956562fadd07a462f2d95589666552e3622843d302.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 330, + 607, + 387 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/daed7dce2e57dd123eb53e9fbaf4f7032951c523ddb4f63c08a49b6a26a41f3f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 624, + 330, + 782, + 387 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Tech tree mastery. As shown in Tab. 2, we experiment on the Minecraft tech tree mastery to test the agent's ability to craft and use a hierarchy of tools. Progressing through this tree (wooden tool $\\rightarrow$ stone tool $\\rightarrow$ iron tool $\\rightarrow$ diamond tool) requires the agent to master systematic and compositional skills. As to the wooden, stone, and iron levels of the tech tree, STEVE achieves remarkable efficiency: $23 \\times$ , $11.8 \\times$ , and $9 \\times$ faster than AutoGPT [49], and $1.5 \\times$ , $1.4 \\times$ , and $1.3 \\times$ faster than Voyager [57]. STEVE has achieved the diamond level, as shown in Tab. 2. Its performance slightly lags behind Voyager [57], which also uses GPT4 [44] for critical inference. However, STEVE is more cost-effective, starting with lower initial performance. It includes a vision unit, prioritizing visual data over background information, offering distinct advantages. Additionally, we observed a decrease in performance when using a basic skill database.", + "bbox": [ + 212, + 482, + 787, + 662 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.4 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 688, + 388, + 704 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "To understand the impact of different components on the performance of our system, we conducted ablation studies focusing on the tech tree mastery task in Minecraft. The results, as shown in Tab. 4, provide insights into the effectiveness of the vision unit and compare our STEVE model with the STEVE GPT-4 version (with the same vision unit as ours). Note that the w/o vision unit setup is that the environmental perception encompasses data on blocks within an 8x8 area surrounding the agent, including the front, back, left, and right directions. The following observations are made:", + "bbox": [ + 212, + 715, + 787, + 835 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "STEVE", + "bbox": [ + 676, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/e2bde9aae14746f4b2def2f7f9bc2ec18a9126fc7f6f2c5f1868c45e87ae463e.jpg", + "table_caption": [ + "Table 4: Ablation studies for the tech tree mastery. STEVE (Ours) is the STEVE-13B version. The $0/3$ score means the method can't progress beyond 160 iterations in the tech tree." + ], + "table_footnote": [], + "table_body": "
MethodWooden ToolStone ToolIron ToolDiamond Tool
w/o vision unit11 ± 5 (3/3)27 ± 5 (3/3)46 ± 11 (3/3)158 (1/3)
STEVE (GPT-4)6 ± 2 (3/3)10 ± 1 (3/3)14 ± 3 (3/3)89 ± 9 (3/3)
STEVE (Ours)4 ± 1 (3/3)8 ± 1 (3/3)15 ± 2 (3/3)106 ± 12 (3/3)
", + "bbox": [ + 246, + 198, + 750, + 275 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Vision unit is critical. The omission of the vision unit markedly affects the system's performance, especially in more advanced tasks. While it successfully crafts Wooden, Stone, and Iron Tools, it is challenged with Diamond Tools. This outcome underscores the vital importance of visual information in accomplishing complex tasks.", + "bbox": [ + 212, + 306, + 787, + 381 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Comparison with GPT-4. As our vision encoder directly encodes into text space, it can be easily replaced with any language model. For instance, the GPT-4 we compared exhibits consistent success across all categories and secures a flawless success rate. Interestingly, the STEVE-13B version excels in simpler tasks such as crafting wooden and stone tools. Moreover, it requires fewer iterations than methods without the vision part, underscoring its superior efficiency.", + "bbox": [ + 212, + 404, + 787, + 494 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5.5 Case Study", + "text_level": 1, + "bbox": [ + 214, + 525, + 356, + 541 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "As shown in Fig. 6, we perform an extensive case study comparison of GPT-4, LLaMA2-13B, and our method STEVE-13B. Each model maintains the same information and question inputs to compare feedback under different environmental information. Our STEVE overall achieves the best results, surpassing GPT-4 and showing significant improvement compared to the original LLaMA. Especially in parts involving numerical calculations, such as the leftmost image, STEVE accurately tracks food values to restore hunger levels.", + "bbox": [ + 212, + 558, + 787, + 665 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 694, + 359, + 710 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "STEVE enhances multi-modal learning by combining visual encoder and LLM-based agents. It has three functions: vision perception, language instruction, and code action, allowing it to understand, predict, and act in virtual environments. We provide a straightforward approach to creating a robust, multimodal, autonomous, embodied agent using an open-source language model with a small number of parameters. Additionally, we provide a comprehensive dataset STEVE-21K for sustainable community development that can be verified.", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Z. Zhao et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work is supported by the Zhejiang Provincial Natural Science Foundation of China (No. LZ24F030005) and the National Natural Science Foundation of China (No. 62106219).", + "bbox": [ + 212, + 178, + 787, + 224 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 248, + 321, + 263 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems 35, 23716-23736 (2022)", + "2. Amiranashvili, A., Dorka, N., Burgard, W., Koltun, V., Brox, T.: Scaling imitation learning in apacheft. arXiv preprint arXiv:2007.02701 (2020)", + "3. Baker, B., Akkaya, I., Zhokov, P., Huizinga, J., Tang, J., Ecoffet, A., Houghton, B., Sampedro, R., Clune, J.: Video pretraining (vpt): Learning to act by watching unlabeled online videos. Advances in Neural Information Processing Systems 35, 24639-24654 (2022)", + "4. Chai, W., Wang, G.: Deep vision multimodal learning: Methodology, benchmark, and trend. Applied Sciences 12(13), 6588 (2022)", + "5. Introducing chatgpt (2022), https://openai.com/blog/chatgpt", + "6. Chen, S., Guhur, P.L., Schmid, C., Laptev, I.: History aware multimodal transformer for vision-and-language navigation. Advances in neural information processing systems 34, 5834-5847 (2021)", + "7. Chen, T., Gupta, S., Gupta, A.: Learning exploration policies for navigation. In: International Conference on Learning Representations (2018)", + "8. Talk to claude (2023), https://claude.ai", + "9. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500 (2023)", + "0. Das, A., Datta, S., Gkioxari, G., Lee, S., Parikh, D., Batra, D.: Embodied question answering. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1-10 (2018)", + "1. Datta, S., Dharur, S., Cartillier, V., Desai, R., Khanna, M., Batra, D., Parikh, D.: Episodic memory question answering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19119-19128 (2022)", + "2. Dean, V., Tulsiani, S., Gupta, A.: See, hear, explore: Curiosity via audio-visual association. Advances in neural information processing systems 33, 14961-14972 (2020)", + "3. Deng, J., Chai, W., Guo, J., Huang, Q., Hu, W., Hwang, J.N., Wang, G.: Citygen: Infinite and controllable 3d city layout generation. arXiv preprint arXiv:2312.01508 (2023)", + "4. Deng, J., Chai, W., Huang, J., Zhao, Z., Huang, Q., Gao, M., Guo, J., Hao, S., Hu, W., Hwang, J.N., et al.: Citycraft: A real craftsfer for 3d city generation. arXiv preprint arXiv:2406.04983 (2024)", + "5. Driess, D., Xia, F., Sajjadi, M.S., Lynch, C., Chowdhery, A., Ichter, B., Wahid, A., Tompson, J., Vuong, Q., Yu, T., et al.: Palm-e: An embodied multimodal language model. arXiv preprint arXiv:2303.03378 (2023)" + ], + "bbox": [ + 225, + 281, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "STEVE", + "bbox": [ + 676, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Du, H., Yu, X., Zheng, L.: Vtnet: Visual transformer network for object goal navigation. In: International Conference on Learning Representations (2020)", + "17. Fan, L., Wang, G., Jiang, Y., Mandlekar, A., Yang, Y., Zhu, H., Tang, A., Huang, D.A., Zhu, Y., Anandkumar, A.: Minedojo: Building open-ended embodied agents with internet-scale knowledge. Advances in Neural Information Processing Systems 35, 18343-18362 (2022)", + "18. Gao, P., Han, J., Zhang, R., Lin, Z., Geng, S., Zhou, A., Zhang, W., Lu, P., He, C., Yue, X., et al.: Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010 (2023)", + "19. Gong, T., Lyu, C., Zhang, S., Wang, Y., Zheng, M., Zhao, Q., Liu, K., Zhang, W., Luo, P., Chen, K.: Multimodal-gpt: A vision and language model for dialogue with humans. arXiv preprint arXiv:2305.04790 (2023)", + "20. Guss, W.H., Houghton, B., Topin, N., Wang, P., Codel, C., Veloso, M., Salakhutdinov, R.: Minerl: A large-scale dataset of apache demonstrations. arXiv preprint arXiv:1907.13440 (2019)", + "21. Hofmann, K.: Minecraft as ai playground and laboratory. In: Proceedings of the annual symposium on computer-human interaction in play. pp. 1-1 (2019)", + "22. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021)", + "23. Huang, W., Xia, F., Xiao, T., Chan, H., Liang, J., Florence, P., Zeng, A., Tompson, J., Mordatch, I., Chebotar, Y., et al.: Inner monologue: Embodied reasoning through planning with language models. arXiv preprint arXiv:2207.05608 (2022)", + "24. Jiang, J., Ye, T., Bai, J., Chen, S., Chai, W., Jun, S., Liu, Y., Chen, E.: Five a+ network: You only need 9k parameters for underwater image enhancement. arXiv preprint arXiv:2305.08824 (2023)", + "25. Johnson, M., Hofmann, K., Hutton, T., Bignell, D.: The malmo platform for artificial intelligence experimentation. In: Ijcai. pp. 4246-4247 (2016)", + "26. Kolve, E., Mottaghi, R., Han, W., VanderBilt, E., Weihs, L., Herrasti, A., Deitke, M., Ehsani, K., Gordon, D., Zhu, Y., et al.: Ai2-thor: An interactive 3d environment for visual ai. arXiv preprint arXiv:1712.05474 (2017)", + "27. Kwon, O., Park, J., Oh, S.: Rendering neural radiance map for visual navigation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9099-9108 (2023)", + "28. Li, B., Zhang, Y., Chen, L., Wang, J., Yang, J., Liu, Z.: Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726 (2023)", + "29. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023)", + "30. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022)", + "31. Li, Y., Yuan, G., Wen, Y., Hu, J., Evangelidis, G., Tulyakov, S., Wang, Y., Ren, J.: Efficientformer: Vision transformers at mobilenet speed. Advances in Neural Information Processing Systems 35, 12934-12949 (2022)", + "32. Lifshitz, S., Paster, K., Chan, H., Ba, J., McIlraith, S.: Steve-1: A generative model for text-to-behavior in apache (abridged version). In: NeurIPS 2023 Workshop on Goal-Conditioned Reinforcement Learning (2023)", + "33. Lin, Z., Li, J., Shi, J., Ye, D., Fu, Q., Yang, W.: Juewu-mc: Playing minecraft with sample-efficient hierarchical reinforcement learning. arXiv preprint arXiv:2112.04907 (2021)" + ], + "bbox": [ + 215, + 147, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Z. Zhao et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "34. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv preprint arXiv:2304.08485 (2023)", + "35. Liu, S., Okatani, T.: Symmetry-aware neural architecture for embodied visual exploration. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 17221-17230. IEEE (2022)", + "36. Lu, P., Peng, B., Cheng, H., Galley, M., Chang, K.W., Wu, Y.N., Zhu, S.C., Gao, J.: Chameleon: Plug-and-play compositional reasoning with large language models. arXiv preprint arXiv:2304.09842 (2023)", + "37. Luo, W., Sun, P., Zhong, F., Liu, W., Zhang, T., Wang, Y.: End-to-end active object tracking via reinforcement learning. In: International conference on machine learning. pp. 3286-3295. PMLR (2018)", + "38. Luo, W., Sun, P., Zhong, F., Liu, W., Zhang, T., Wang, Y.: End-to-end active object tracking and its real-world deployment via reinforcement learning. IEEE transactions on pattern analysis and machine intelligence 42(6), 1317-1332 (2019)", + "39. Lyu, C., Wu, M., Wang, L., Huang, X., Liu, B., Du, Z., Shi, S., Tu, Z.: Macaw-llm: Multi-modal language modeling with image, audio, video, and text integration. arXiv preprint arXiv:2306.09093 (2023)", + "40. Ma, W., Mi, Q., Yan, X., Wu, Y., Lin, R., Zhang, H., Wang, J.: Large language models play starcraft ii: Benchmarks and a chain of summarization approach. arXiv preprint arXiv:2312.11865 (2023)", + "41. Maaz, M., Rasheed, H., Khan, S., Khan, F.S.: Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424 (2023)", + "42. Mao, H., Wang, C., Hao, X., Mao, Y., Lu, Y., Wu, C., Hao, J., Li, D., Tang, P.: Seihai: A sample-efficient hierarchical ai for the minerl competition. In: Distributed Artificial Intelligence: Third International Conference, DAI 2021, Shanghai, China, December 17-18, 2021, Proceedings 3. pp. 38-51. Springer (2022)", + "43. Moudgil, A., Majumdar, A., Agrawal, H., Lee, S., Batra, D.: Soat: A scene-and object-aware transformer for vision-and-language navigation. Advances in Neural Information Processing Systems 34, 7357-7367 (2021)", + "44. OpenAI: Gpt-4 technical report. arXiv preprint arXiv: Arxiv-2303.08774 (2023)", + "45. PrismarineJS: Prismarinejs/mineflayer: Create apache bots with a powerful, stable, and high level javascript api. (2013), https://github.com/PrismarineJS/mineflayer/tree/master", + "46. Savva, M., Kadian, A., Maksymets, O., Zhao, Y., Wijmans, E., Jain, B., Straub, J., Liu, J., Koltun, V., Malik, J., et al.: Habitat: A platform for embodied ai research. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9339-9347 (2019)", + "47. Schick, T., Dwivedi-Yu, J., Dessi, R., Raileanu, R., Lomeli, M., Zettlemoyer, L., Cancedda, N., Scialom, T.: Toolformer: Language models can teach themselves to use tools. arXiv preprint arXiv:2302.04761 (2023)", + "48. Shen, Y., Song, K., Tan, X., Li, D., Lu, W., Zhuang, Y.: Hugginggpt: Solving ai tasks with chatgpt and its friends in huggingface. arXiv preprint arXiv:2303.17580 (2023)", + "49. Significant-Gravitas: Auto-gpt. https://github.com/Significant-Gravitas/Auto-GPT (2023)", + "50. Skrynnik, A., Staroverov, A., Aitygulov, E., Aksenov, K., Davydov, V., Panov, A.I.: Hierarchical deep q-network from imperfect demonstrations in minecraft. Cognitive Systems Research 65, 74-78 (2021)" + ], + "bbox": [ + 212, + 146, + 784, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "STEVE", + "bbox": [ + 676, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "51. Song, C.H., Wu, J., Washington, C., Sadler, B.M., Chao, W.L., Su, Y.: Llm-planner: Few-shot grounded planning for embodied agents with large language models. arXiv preprint arXiv:2212.04088 (2022)", + "52. Song, E., Chai, W., Wang, G., Zhang, Y., Zhou, H., Wu, F., Guo, X., Ye, T., Lu, Y., Hwang, J.N., et al.: Moviechat: From dense token to sparse memory for long video understanding. arXiv preprint arXiv:2307.16449 (2023)", + "53. Song, E., Chai, W., Ye, T., Hwang, J.N., Li, X., Wang, G.: Moviechat+: Question-aware sparse memory for long video question answering. arXiv preprint arXiv:2404.17176 (2024)", + "54. Su, Y., Lan, T., Li, H., Xu, J., Wang, Y., Cai, D.: Pandagpt: One model to instruction-follow them all. arXiv preprint arXiv:2305.16355 (2023)", + "55. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)", + "56. Vasluianu, F.A., Seizinger, T., Zhou, Z., Wu, Z., Chen, C., Timofte, R., Dong, W., Zhou, H., Tian, Y., Chen, J., et al.: Ntire 2024 image shadow removal challenge report. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6547-6570 (2024)", + "57. Wang, G., Xie, Y., Jiang, Y., Mandlekar, A., Xiao, C., Zhu, Y., Fan, L., Anandkumar, A.: Voyager: An open-ended embodied agent with large language models. arXiv preprint arXiv:2305.16291 (2023)", + "58. Wang, W., Chen, Z., Chen, X., Wu, J., Zhu, X., Zeng, G., Luo, P., Lu, T., Zhou, J., Qiao, Y., et al.: Visionllm: Large language model is also an open-ended decoder for vision-centric tasks. arXiv preprint arXiv:2305.11175 (2023)", + "59. Wang, Z., Cai, S., Liu, A., Ma, X., Liang, Y.: Describe, explain, plan and select: Interactive planning with large language models enables open-world multi-task agents. arXiv preprint arXiv:2302.01560 (2023)", + "60. Whitted, T.: An improved illumination model for shaded display. In: ACM SIGGRAPH 2005 Courses, pp. 4-es (2005)", + "61. Wijmans, E., Kadian, A., Morcos, A., Lee, S., Essa, I., Parikh, D., Savva, M., Batra, D.: Dd-ppo: Learning near-perfect pointgoal navigators from 2.5 billion frames. In: International Conference on Learning Representations (2019)", + "62. Wu, C., Yin, S., Qi, W., Wang, X., Tang, Z., Duan, N.: Visual chatgpt: Talking, drawing and editing with visual foundation models. arXiv preprint arXiv:2303.04671 (2023)", + "63. Xie, Y., Yu, C., Zhu, T., Bai, J., Gong, Z., Soh, H.: Translating natural language to planning goals with large-language models. arXiv preprint arXiv:2302.05128 (2023)", + "64. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178 (2023)", + "65. Ye, T., Chen, S., Liu, Y., Chai, W., Bai, J., Zou, W., Zhang, Y., Jiang, M., Chen, E., Xue, C.: Sequential affinity learning for video restoration. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 4147-4156 (2023)", + "66. Ye, T., Zhang, Y., Jiang, M., Chen, L., Liu, Y., Chen, S., Chen, E.: Perceiving and modeling density for image dehazing. In: European conference on computer vision. pp. 130-145. Springer (2022)", + "67. Yu, L., Chen, X., Gkioxari, G., Bansal, M., Berg, T.L., Batra, D.: Multi-target embodied question answering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6309-6318 (2019)", + "68. Yu, Y., Huang, W., Sun, F., Chen, C., Wang, Y., Liu, X.: Sound adversarial audiovisual navigation. In: International Conference on Learning Representations (2021)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Z. Zhao et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "69. Yuan, H., Zhang, C., Wang, H., Xie, F., Cai, P., Dong, H., Lu, Z.: Plan4mc: Skill reinforcement learning and planning for open-world apache tasks. arXiv preprint arXiv:2303.16563 (2023)", + "70. Zhao, Z., Chai, W., Hao, S., Hu, W., Wang, G., Cao, S., Song, M., Hwang, J.N., Wang, G.: A survey of deep learning in sports applications: Perception, comprehension, and decision. arXiv preprint arXiv:2307.03353 (2023)", + "71. Zhao, Z., Chai, W., Wang, X., Ma, K., Chen, K., Guo, D., Ye, T., Zhang, Y., Wang, H., Wang, G.: Steve series: Step-by-step construction of agent systems in northwest China. arXiv preprint arXiv:2406.11247 (2024)", + "72. Zhao, Z., Chen, K., Guo, D., Chai, W., Ye, T., Zhang, Y., Wang, G.: Hierarchical auto-organizing system for open-ended multi-agent navigation. arXiv preprint arXiv:2403.08282 (2024)", + "73. Zhao, Z., Ma, K., Chai, W., Wang, X., Chen, K., Guo, D., Zhang, Y., Wang, H., Wang, G.: Do we really need a complex agent system? distill embodied agent into a single model. arXiv preprint arXiv:2404.04619 (2024)", + "74. Zhong, F., Sun, P., Luo, W., Yan, T., Wang, Y.: Ad-vat+: An asymmetric dueling mechanism for learning and understanding visual active tracking. IEEE transactions on pattern analysis and machine intelligence 43(5), 1467-1482 (2019)", + "75. Zhong, F., Sun, P., Luo, W., Yan, T., Wang, Y.: Towards distraction-robust active visual tracking. In: International Conference on Machine Learning. pp. 12782-12792. PMLR (2021)", + "76. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023)", + "77. Zhu, X., Chen, Y., Tian, H., Tao, C., Su, W., Yang, C., Huang, G., Li, B., Lu, L., Wang, X., et al.: Ghost in the minecraft: Generally capable agents for open-world enviroments via large language models with text-based knowledge and memory. arXiv preprint arXiv:2305.17144 (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 534 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "STEVE", + "bbox": [ + 676, + 114, + 730, + 126 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/See and Think_ Embodied Agent in Virtual Environment/0ba9ae38-8422-4b1f-90c0-ee0bf0ad307a_model.json b/2024/See and Think_ Embodied Agent in Virtual Environment/0ba9ae38-8422-4b1f-90c0-ee0bf0ad307a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a25ce7d63e3f8103e1b5ab32f1e6f6b9e733a5e5 --- /dev/null +++ b/2024/See and Think_ Embodied Agent in Virtual Environment/0ba9ae38-8422-4b1f-90c0-ee0bf0ad307a_model.json @@ -0,0 +1,3230 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.255, + 0.141, + 0.75, + 0.185 + ], + "angle": 0, + "content": "See and Think: Embodied Agent in Virtual Environment" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.212, + 0.763, + 0.245 + ], + "angle": 0, + "content": "Zhonghan Zhao\\(^{1, \\spadesuit}\\), Wenhao Chai\\(^{2, \\spadesuit}\\), Xuan Wang\\(^{1, \\spadesuit}\\), Boyi Li\\(^{1}\\), Shengyu Hao\\(^{1}\\), Shidong Cao\\(^{1}\\), Tian Ye\\(^{3}\\), and Gaoang Wang\\(^{1, \\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.428, + 0.255, + 0.576, + 0.269 + ], + "angle": 0, + "content": "1 Zhejiang University" + }, + { + "type": "text", + "bbox": [ + 0.409, + 0.27, + 0.594, + 0.283 + ], + "angle": 0, + "content": "2 University of Washington" + }, + { + "type": "text", + "bbox": [ + 0.312, + 0.283, + 0.692, + 0.298 + ], + "angle": 0, + "content": "\\(^{3}\\) Hong Kong University of Science and Technology (GZ)" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.331, + 0.744, + 0.581 + ], + "angle": 0, + "content": "Abstract. Large language models (LLMs) have achieved impressive progress on several open-world tasks. Recently, using LLMs to build embodied agents has been a hotspot. This paper proposes STEVE, a comprehensive and visionary embodied agent in the Minecraft virtual environment. STEVE comprises three key components: vision perception, language instruction, and code action. Vision perception involves interpreting visual information in the environment, which is then integrated into the LLMs component with agent state and task instruction. Language instruction is responsible for iterative reasoning and decomposing complex tasks into manageable guidelines. Code action generates executable skill actions based on retrieval in skill database, enabling the agent to interact effectively within the Minecraft environment. We also collect STEVE-21K dataset, which includes \\(600+\\) vision-environment pairs, 20K knowledge question-answering pairs, and \\(200+\\) skill-code pairs. We conduct continuous block search, knowledge question and answering, and tech tree mastery to evaluate the performance. Extensive experiments show that STEVE achieves at most \\(1.5 \\times\\) faster unlocking key tech trees and \\(2.5 \\times\\) quicker in block search tasks." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.594, + 0.741, + 0.622 + ], + "angle": 0, + "content": "Keywords: Open-world embodied agent \\(\\cdot\\) Multimodal pre-training \\(\\cdot\\) Large language model" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.683, + 0.377, + 0.699 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.713, + 0.788, + 0.79 + ], + "angle": 0, + "content": "Designing agents that demonstrate intelligent behavior and adaptability in open-world settings has been a longstanding and significant challenge in the field of artificial intelligence [13,14,26,46,70]. However, recent progress in the development of large language models (LLMs) [5,55] has exhibited their potential as versatile, general-purpose assistants. Recent innovations in agent design [57,59,69,77]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.798, + 0.787, + 0.826 + ], + "angle": 0, + "content": "\\(\\spadesuit\\) Equal contribution: zhaozhonghan@zju.edu.cn, wchai@uw.edu, xu-anw@zju.edu.cn." + }, + { + "type": "page_footnote", + "bbox": [ + 0.251, + 0.826, + 0.606, + 0.84 + ], + "angle": 0, + "content": "† Corresponding author: gaoangwang@intl.zju.edu.cn." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.127 + ], + "angle": 0, + "content": "Z. Zhao et al." + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.145, + 0.775, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.271, + 0.788, + 0.327 + ], + "angle": 0, + "content": "Fig. 1: STEVE integrates the three parts: Vision Perception, Language Instruction, and Code Action, supported closely by our proposed STEVE-21K. It demonstrates commendable performance on Continuous Block Search, Knowledge QA, and Tech Tree Mastery." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.365, + 0.788, + 0.44 + ], + "angle": 0, + "content": "have effectively harnessed these advanced LLMs, tapping into their extensive world knowledge and reasoning abilities. This development has paved the way for agents, that are autonomously driven, to formulate and implement strategies and actions across a diverse array of skills and tasks in diverse open-world environments." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.444, + 0.788, + 0.655 + ], + "angle": 0, + "content": "In many open-world settings, like Minecraft, contemporary agents predominantly use LLMs for their textual interactions. However, this reliance on text for communication poses considerable limitations in their interactions within these worlds including low-level regrading cases [24, 56, 65, 66]. Minecraft, with its expansive and interactive sandbox environment [17, 20], demands a variety of skills from agents, ranging from crafting basic items to executing complex tasks. Yet, agents driven by LLMs often generate unpredictable outputs. The effectiveness of their interactions is largely contingent on meticulously crafted prompts [23], designed to align the LLM's understanding with the environmental context and the intended objectives. This process of prompt engineering is not only laborious but also fails to meet the goal of fostering autonomous, self-directed agents. Furthermore, textual communication has its limitations in naturally conveying certain concepts of the world, like crafting recipes, which are often more effectively communicated through vision." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Players have the distinct capability to assimilate and convey information using both visual and textual channels, significantly enhancing our interactions with the world around us. Yet, the integration of LLM-based agents with multimodal inputs in open-ended environments remains an under-explored area. STEVE in STEVE-Series [71-73], is named after the protagonist of the game \"Minecraft.\" It is our proposed framework to build an embodied agent based on the vision model and LLMs within an open world, as illustrated in Fig. 1. STEVE harnesses a vision model to perceive its surroundings visually, coupled with an LLM to strategize and plan actions. This model represents a leap forward in agent design, combining these two input modes, vision, and text, to offer a more nuanced and comprehensive understanding of the environment and practical and executable skills." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.677, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "STEVE" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.147, + 0.571, + 0.162 + ], + "angle": 0, + "content": "Our key contributions are outlined as follows:" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.165, + 0.786, + 0.224 + ], + "angle": 0, + "content": "- We propose STEVE, an embodied agent in virtual environment, consists of vision perception, language instruction, and code action, achieving \\(1.5 \\times\\) faster unlocking of key tech trees and is \\(2.3 \\times\\) quicker in block search tasks compared to previous state-of-the-art methods." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.226, + 0.784, + 0.269 + ], + "angle": 0, + "content": "- We present STEVE-7B/13B, a series of large language model obtained by fine-tuning with Minecraft knowledge question-answering pairs from Llama2-7B/13B." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.271, + 0.784, + 0.316 + ], + "angle": 0, + "content": "- We collect STEVE-21K dataset, including \\(600+\\) vision-environment pairs, 20K knowledge question-answering pairs, and \\(200+\\) skill-code pairs, for justifying the effective performance of STEVE." + }, + { + "type": "list", + "bbox": [ + 0.228, + 0.165, + 0.786, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.34, + 0.396, + 0.357 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.373, + 0.51, + 0.388 + ], + "angle": 0, + "content": "2.1 Intelligent Agent in Minecraft" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.398, + 0.788, + 0.655 + ], + "angle": 0, + "content": "As an open-ended sandbox game, Minecraft has always been an ideal setting for testing the performance of intelligent agents [21, 25]. The agents are required to autonomously perform various tasks in Minecraft, such as chopping trees, crafting tools, and mining diamonds. At the beginning, much of the works focus on exploring reinforcement learning [32, 33, 42, 50] or imitation learning [2, 3], without satisfactory performance. VPT [3] and MineDojo [17] collect internet-scale datasets for their model pre-training. More specifically, VPT offers the exciting possibility of directly learning to act during video pre-training and using these learned behavioral priors as extremely effective exploration priors for reinforcement learning. Yet, recent works found that the pre-trained LLMs could serve as a strong \"mind\" that provides planning ability to the agents. Voyager [57] leverages GPT-4 [44] as both a high-level planner and a low-level action code generator. Plan4MC [69] proposes a skill graph pre-generated by the LLMs. DEPS [59], an interactive planning method based on LLMs, addresses multi-step reasoning issue in open-world planning. GITM [77] develops a set of structured actions and leverages LLMs to generate action plans for the agents to execute, achieving impressive results in various tasks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.678, + 0.506, + 0.693 + ], + "angle": 0, + "content": "2.2 Embodied Multimodal Model" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Embodied agent operates within various environments by synthesizing sensory perceptions and physical actions supported by computational intelligence. This synthesis enables the agent to undertake a variety of tasks, achieving specific objectives. Its key areas of application are diverse, including Navigation [6, 16, 27, 43, 61, 68], Embodied Question Answering [10, 11, 67], Active Visual Tracking [37, 38, 74, 75], and Visual Exploration [7, 12, 35]. The field is evolving rapidly with the development of Large Language Models (LLMs) [51] and Multimodal LLMs (MLLMs) [1, 9, 18, 19, 28–30, 34, 39, 41, 54, 58, 64, 76], integrating multiple modalities for more effective processing. A prime example of this innovation is" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.127 + ], + "angle": 0, + "content": "Z. Zhao et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.144, + 0.784, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.267, + 0.788, + 0.336 + ], + "angle": 0, + "content": "Fig. 2: STEVE framework. The Vision Perception part takes images or videos, encodes them into tokens, and combines them with Agent State and Task tokens as input. The STEVE-13B in the Language Instruction part is used for automatic reasoning and task decomposition, and it calls the Skill Database in the form of the Query to output code as action." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.367, + 0.788, + 0.414 + ], + "angle": 0, + "content": "PaLM-E [15], a sophisticated multimodal model with 562B parameters, adept at a broad spectrum of embodied tasks and demonstrating exceptional capabilities in visual reasoning." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.438, + 0.629, + 0.455 + ], + "angle": 0, + "content": "2.3 Large Language Model with Equipped Tools" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.466, + 0.789, + 0.724 + ], + "angle": 0, + "content": "While Large Language Models (LLMs) demonstrate impressive skill in tackling novel tasks via prompt-based instructions, they often face challenges in areas where simpler models or tools excel, like mathematical calculations or identifying palindromes. However, LLMs' potential is significantly expanded when integrated with other modality-specific models, such as those for vision or audio, enabling multi-modal capabilities [4,36]. Innovations like Toolformer [47] demonstrate LLMs' self-learning to utilize tools through finetuning with extensive API call samples. Visual ChatGPT [62] extends this by integrating various visual foundation models, facilitating interactive user experiences with ChatGPT. Similarly, HuggingGPT [48] presents a framework that harnesses LLMs to link diverse models from Hugging Face for task resolution. AutoGPT [49] is an open-source application that broadens GPT-4's capabilities with internet access, memory management, and plug-ins. The recent introduction of MovieChat [52,53] brings a memory mechanism to MLLM, enhancing its performance in video understanding tasks. Furthermore, LLMs can be used for goal planning, analogous to language translation [63]. This evolving landscape suggests that tool-equipped LLMs could forge a new paradigm in AI solution design." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.748, + 0.418, + 0.765 + ], + "angle": 0, + "content": "3 Method: STEVE" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.783, + 0.342, + 0.797 + ], + "angle": 0, + "content": "3.1 Overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "As shown in Figure 2, STEVE \\(\\mathcal{F}\\) is an LLM-based multi-modal autonomous system for embodied agents in Minecraft, which can use visual status \\(X^v\\) in the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.677, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "STEVE" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.176 + ], + "angle": 0, + "content": "form of images or videos and agent state \\( X^s \\) to manage and execute complex task \\( X^t \\) for executive code action \\( \\mathbf{a}^c \\):" + }, + { + "type": "equation", + "bbox": [ + 0.333, + 0.192, + 0.786, + 0.209 + ], + "angle": 0, + "content": "\\[\n\\mathbf {a} ^ {c} = \\mathcal {F} \\left(X ^ {v}, X ^ {s}, X ^ {t}\\right) = \\mathcal {A} ^ {c} \\left(\\mathcal {I} ^ {l} \\left(\\mathcal {P} ^ {v} \\left(X ^ {v}, X ^ {s}, X ^ {t}\\right)\\right)\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.214, + 0.784, + 0.26 + ], + "angle": 0, + "content": "where STEVE \\(\\mathcal{F}\\) integrates Visual Perception \\(\\mathcal{P}^v\\) into the Language Instruction \\(\\mathcal{I}^l\\) with large language model (LLM) and combines them with Code Action \\(\\mathcal{A}^c\\), a skill retrieval method to execute actions, details are demonstrated as follows." + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.272, + 0.784, + 0.315 + ], + "angle": 0, + "content": "- Vision Perception \\(\\mathcal{P}^v\\) (Section 3.2); a vision encoder to interpret visual information of the environment, such as blocks and entities and a text tokenizer for agent state \\(s\\) and task \\(t\\)." + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.317, + 0.784, + 0.375 + ], + "angle": 0, + "content": "- Language Instruction \\(\\mathcal{I}^l\\) (Section 3.3): a powerful language model system fine-tuned specifically for Minecraft content using LLaMA2-13B [18]. This model enables adaptive interaction for iterative reasoning and step-by-step decomposition." + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.377, + 0.784, + 0.405 + ], + "angle": 0, + "content": "- Code Action \\(\\mathcal{A}^c\\) (Section 3.4): a skill retrieval method based on our skill-code database." + }, + { + "type": "list", + "bbox": [ + 0.227, + 0.272, + 0.784, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.429, + 0.438, + 0.445 + ], + "angle": 0, + "content": "3.2 Vision Perception \\(\\mathcal{P}^v\\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.454, + 0.784, + 0.501 + ], + "angle": 0, + "content": "The vision perception part includes a vision encoder \\( E^v \\) and a text tokenizer \\( T \\), which converts the visual status \\( X^v \\), agent state \\( X^s \\), and task \\( X^t \\) to the text-space tokenizer representation \\( \\mathbf{Y} = \\{Y^v, Y^s, Y^t\\} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.371, + 0.515, + 0.785, + 0.531 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Y} _ {i} = \\mathcal {P} ^ {v} \\left(X _ {i} ^ {v}, X _ {i} ^ {s}, X _ {i} ^ {t}\\right), \\mathcal {P} ^ {v} = \\left\\{E ^ {v}, T \\right\\}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.537, + 0.788, + 0.688 + ], + "angle": 0, + "content": "where the vision encoder \\( E \\), the visual branch of EfficientFormer [31], encodes visual status \\( X_{i}^{v} \\) at each step \\( i \\) into visual tokens \\( Y^{v} = \\{y_{1}^{v}, y_{2}^{v}, \\dots, y_{n}^{v}\\} \\in \\mathbb{R}^{n \\times d} \\), where \\( n \\) denotes the number of visual tokens \\( y^{v} \\) and \\( d \\) is the dimensionality of each token. The text tokenizer converts agent state \\( X_{i}^{s} \\) and task \\( X_{i}^{t} \\) at each step \\( i \\) in the form of text into textual tokens \\( Y^{s} \\) and \\( Y^{t} \\). Note that visual tokens are amalgamated with textual tokens representing the agent's current state (e.g., health, inventory, etc.) and the task description. This is accomplished using a tokenizer that maps state variables and task parameters to a tokenized form. The resultant unified token set serves as a comprehensive representation of the current situational context." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.71, + 0.462, + 0.726 + ], + "angle": 0, + "content": "3.3 Language Instruction \\(\\mathcal{I}^l\\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.784, + 0.84 + ], + "angle": 0, + "content": "The Language Instruction \\(\\mathcal{I}^l = \\{\\mathcal{P}_l, \\mathcal{C}_r, \\mathcal{C}_u, \\mathcal{D}_s\\}\\), which consists of Planner \\(\\mathcal{P}_l\\), Critic \\(\\mathcal{C}_r\\), Curriculum \\(\\mathcal{C}_u\\), and Descriptor \\(\\mathcal{D}_s\\), i.e., four independent LLM-based agents with different functions. They formulate high-level task guidelines, refine strategies through feedback, facilitate continuous learning and adaptation through a curriculum of complex tasks, and finally decompose strategic guidelines into the executable low-level textual action step \\(\\mathbf{a}^s\\), \\(\\{\\mathbf{a}^s \\sim \\mathcal{I}^l(\\cdot|\\mathbf{Y}_i)\\}_{i=0}^N\\) towards efficient completion of the task." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "Z. Zhao et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "Note that the Planner \\(\\mathcal{P}_l\\), Critic \\(\\mathcal{C}_r\\), Curriculum \\(\\mathcal{C}_u\\) and Descriptor \\(\\mathcal{D}_s\\) are based on STEVE-7B/13B, a powerful language model derived from LLaMA-2-13B [18], fine-tuned specifically on Minecraft-related content from the STEVE-21K. This model's expertise covers a broad spectrum of game-specific knowledge areas on worlds, entities, player mechanics, survival, and even game practical experience. Finally, they have the performance of their different functions:" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.251, + 0.784, + 0.279 + ], + "angle": 0, + "content": "- Planner \\(\\mathcal{P}_l\\): formulating comprehensive guidelines and executive plans that align with the overarching objectives of the task." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.281, + 0.784, + 0.31 + ], + "angle": 0, + "content": "- Critic \\(\\mathcal{C}_r\\): evaluating the planner decisions, providing feedback that can refine strategies." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.311, + 0.784, + 0.34 + ], + "angle": 0, + "content": "- Curriculum \\(\\mathcal{C}_u\\): facilitating continuous learning and adaptation for action agents by engaging with a series of progressively complex tasks." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.342, + 0.784, + 0.371 + ], + "angle": 0, + "content": "- Descriptor \\( \\mathcal{D}_s \\): distilling the extensive data into a concise summary, making it more manageable and interpretable." + }, + { + "type": "list", + "bbox": [ + 0.227, + 0.251, + 0.784, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.387, + 0.788, + 0.538 + ], + "angle": 0, + "content": "Iterative reasoning. The STEVE-13B receives a stream of tokens that encode the current visual scene, the agent's state, and the task's textual description. STEVE-13B interprets this rich context to undertake complex reasoning. The model initiates the reasoning process by constructing a series of high-level strategies that outline the pathway to task completion. Considering all gameplay elements, the reasoning mechanism is akin to an experienced player who can visualize the end game and chart a course to victory. This approach ensures the plans are reactive and proactive, allowing the agent to anticipate and mitigate future challenges. However, most strategies are high-level and abstract therefore, they often require step-by-step decomposition to derive executable guidelines." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.788, + 0.721 + ], + "angle": 0, + "content": "Decomposition. The decomposition process makes the complex strategies breakdown into simple, low-level guidelines that can be directly mapped to actions in Minecraft. It is similar to how a high-level command like \"build a shelter\" is divided into actionable instructions like \"collect wood\", \"craft planks\", and \"place blocks\". The granular steps are structured to provide explicit instructions that the game engine can readily interpret. This systematic breakdown from high-level reasoning to low-level actionable steps is the hallmark of the STEVE system, enabling the embodied agent to interact with the Minecraft environment in a meaningful and goal-oriented manner. Through this intricate process of reasoning and decomposition, STEVE embodies the cognitive capabilities required for sophisticated task management in virtual environments." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Curriculum learning with memory. We draw inspiration from the lifelong learning strategy utilized in many reinforcement learning problems [57] in both closed-world and open-world settings. We start by creating a set of tasks that serve as a curriculum for agents to explore the environment. During this process, STEVE generates plans, interacts with the surroundings, learns from mistakes, and stores all these experiences in memory. Next, we evaluate STEVE on various tasks after this learning stage. Consequently, STEVE can produce better plans with its" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.677, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "STEVE" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "memory teaming up with the planning experiences. We use this as the default setting for all tasks in our experiments." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.197, + 0.788, + 0.364 + ], + "angle": 0, + "content": "Continuous learning with summarization. We've observed that the learning process, where the memory is being filled, can continue throughout the gameplay. The agent can gradually acquire more skills as the gameplay progresses and more experiences are gained. However, as the memory gets larger, it becomes difficult to understand the game's situation and interact with the game slowly. To tackle these challenges, we have implemented the Chain of Summarization method [40]. By finding better references, we can improve our ability to handle challenging tasks, such as \"Diamond Tool\" in the tech tree, including obtaining materials and creating diamond tools. This will lead to a higher success rate, as shown in Section 5.3. Additionally, curriculum learning with memory allows for in-context lifelong learning without needing gradient updates." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.39, + 0.396, + 0.404 + ], + "angle": 0, + "content": "3.4 Code Action \\(\\mathcal{A}^c\\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.42, + 0.788, + 0.511 + ], + "angle": 0, + "content": "The code action part \\(\\mathcal{A}^c\\) is the execution phase, where the STEVE system converts planned, decomposed guidelines into concrete actions within the Minecraft environment. This process leverages a specialized skill database that pairs code snippets with their descriptions and relevant metadata, encoded as vectors \\(\\mathbf{v}^{\\mathbf{s}}\\) for efficient retrieval. The transition from language instruction to executable code is achieved through the retrieval process \\(\\mathcal{R}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.393, + 0.529, + 0.785, + 0.546 + ], + "angle": 0, + "content": "\\[\n\\mathcal {R} (\\mathbf {q}, \\mathbf {v}) = \\sigma (\\mathbf {q}, \\mathbf {v}), \\mathbf {q} = E ^ {q} \\left(\\mathbf {a} ^ {\\mathbf {s}}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.555, + 0.788, + 0.66 + ], + "angle": 0, + "content": "where \\( E^q \\) and \\( \\sigma \\) are query encoding and cosine similarity matching. Each low-level textual action step \\( \\mathbf{a}^s \\) derived from the Language Instruction is encoded into a query \\( \\mathbf{q} \\). This encoding captures the essence of the action to be performed in a format that the skill database can understand. Once the queries \\( \\mathbf{q} \\) are encoded, the system computes similarities between the query vectors and the code snippets vectors \\( \\mathbf{v} \\) stored in the database to determine which skill best matches the required action." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.688, + 0.334, + 0.703 + ], + "angle": 0, + "content": "3.5 Training" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.717, + 0.788, + 0.793 + ], + "angle": 0, + "content": "To reduce the training overhead to a certain extent, we adopt a two-stage training method, the first stage being a warm-up on the question-answering pairs of STEVE-21K to ensure a certain degree of instruction capability. The second stage is simulated in the environment, and the Expert LLM \\(\\mathcal{E}_p\\) integrated with GPT-4 [44] generates instructions for the same situation to modify our model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.794, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In training, we use either a single-round conversation \\(\\{X^Q,X^A\\}\\) or a multiround conversation \\(\\{X_i^Q,X_i^A\\}_{i\\leq N}\\), where \\(N\\) is the total number of rounds in the conversation. We use the negative log-likelihood objective over the prediction" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "Z. Zhao et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.48, + 0.162 + ], + "angle": 0, + "content": "tokens with training and finetuning:" + }, + { + "type": "equation", + "bbox": [ + 0.377, + 0.176, + 0.787, + 0.218 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (\\theta) = - \\sum_ {j = 1} ^ {L} \\log \\mathcal {F} _ {\\theta} \\left(Y _ {j} \\mid X ^ {v}, \\hat {Y} _ {1: j - 1}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.232, + 0.784, + 0.307 + ], + "angle": 0, + "content": "where \\( Y \\) and \\( \\hat{Y} \\) refer to the non-vision input and target token sequences, \\( \\theta \\) represents the model parameters, and \\( L \\) represents the length of the target sequence. The vision input \\( X^v \\) varies depending on the input step. We only consider the answer tokens \\( X_A \\) when computing the loss to ensure that the model focuses on generating coherent responses." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.308, + 0.784, + 0.353 + ], + "angle": 0, + "content": "Offline warm-up. In the first stage, we finetune the STEVE-7B/13B only on the single-round question-answering pairs of STEVE-21K to ensure a certain degree of instruction capability, as shown in Tab. 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.354, + 0.785, + 0.399 + ], + "angle": 0, + "content": "Online finetuning. In the second stage, we simulate the warmed-up STEVE-7B/13B in Minecraft, and both train the vision encoder \\( E^v \\) and finetune the STEVE-7B/13B simultaneously." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.4, + 0.785, + 0.505 + ], + "angle": 0, + "content": "It is necessary to obtain the environment information to train the vision encoder, denoted as \\( l = R_{T}(X^{v}) \\), where \\( l \\) is the label of seen blocks and entities, \\( R_{T} \\) is the Ray Tracing method [60] to eliminate all content outside the agent's perspective screen. We train the vision encoder \\( E^{v} \\) once we have obtained the environment information. To finetune the warmed-up STEVE-7B/13B in simulation, we give both the STEVE-7B/13B and the Expert LLM the same input, and we consider the output of the Expert LLM as the ground truth label." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.506, + 0.787, + 0.551 + ], + "angle": 0, + "content": "After 5,000 simulations, we collect all the context information from successful runs, including sequences of vision and question-answering chat flow as the Vision-Environment part of STEVE-21K, as mentioned in Section 4." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.578, + 0.462, + 0.594 + ], + "angle": 0, + "content": "4 Dataset: STEVE-21K" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.784, + 0.656 + ], + "angle": 0, + "content": "As shown in Fig. 3, STEVE-21K has been meticulously compiled, featuring a diverse collection of Vision-Environment pairs, Question-Answering pairs, and a Skill-Code database." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.785, + 0.794 + ], + "angle": 0, + "content": "Vision-Environment pairs contain 600 pairs of first-person perspective videos from Minecraft gameplay across six different terrains (including forest, desert, coastal, etc.), along with corresponding environmental block entities in the field of vision and context information for each timestamp. Additionally, all pairs are oriented around executing the skill-related task supported by Skill-Code part mentioned in Section 4. We employ the STEVE-7B/13B model to enable robots to plan and execute actions autonomously based on tasks defined by human supervisors. We record the video of the agent operation, the environment information, and all the corresponding chatflow." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Question-Answering pairs contain 20K question-answering pairs from the Minecraft-Wiki and Reddit corpus across six data types partly sourced from [17]. The pairs are organized into instruction, input, and output triplets and used to" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.678, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "STEVE" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.145, + 0.673, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.689, + 0.149, + 0.747, + 0.161 + ], + "angle": 0, + "content": "Skill-Code" + }, + { + "type": "text", + "bbox": [ + 0.689, + 0.167, + 0.762, + 0.175 + ], + "angle": 0, + "content": "Human Player" + }, + { + "type": "image", + "bbox": [ + 0.717, + 0.188, + 0.741, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.743, + 0.211, + 0.771, + 0.218 + ], + "angle": 0, + "content": "prompt" + }, + { + "type": "image", + "bbox": [ + 0.718, + 0.227, + 0.741, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.747, + 0.23, + 0.762, + 0.238 + ], + "angle": 0, + "content": "GPT" + }, + { + "type": "text", + "bbox": [ + 0.718, + 0.249, + 0.769, + 0.259 + ], + "angle": 0, + "content": "output" + }, + { + "type": "text", + "bbox": [ + 0.708, + 0.269, + 0.759, + 0.277 + ], + "angle": 0, + "content": "code snippets" + }, + { + "type": "image", + "bbox": [ + 0.717, + 0.29, + 0.743, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.743, + 0.286, + 0.771, + 0.293 + ], + "angle": 0, + "content": "execute" + }, + { + "type": "image", + "bbox": [ + 0.716, + 0.301, + 0.744, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.697, + 0.326, + 0.765, + 0.334 + ], + "angle": 0, + "content": "check and revise" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.255, + 0.345, + 0.265 + ], + "angle": 0, + "content": "Question-Answering" + }, + { + "type": "image_caption", + "bbox": [ + 0.249, + 0.277, + 0.292, + 0.285 + ], + "angle": 0, + "content": "Wikipedia" + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.289, + 0.307, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.324, + 0.301, + 0.33 + ], + "angle": 0, + "content": "The crafting menu in Minecraft, showing the crafting origins of a clone." + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.33, + 0.3, + 0.337 + ], + "angle": 0, + "content": "showing the eating recipe or a special taste as well as some other blocks and" + }, + { + "type": "image_caption", + "bbox": [ + 0.342, + 0.277, + 0.369, + 0.285 + ], + "angle": 0, + "content": "Reddit" + }, + { + "type": "image", + "bbox": [ + 0.321, + 0.29, + 0.386, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.407, + 0.265, + 0.43, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.432, + 0.27, + 0.472, + 0.279 + ], + "angle": 0, + "content": "ChatGPT" + }, + { + "type": "image", + "bbox": [ + 0.404, + 0.341, + 0.418, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.43, + 0.29, + 0.479, + 0.3 + ], + "angle": 0, + "content": "Clean" + }, + { + "type": "text", + "bbox": [ + 0.448, + 0.305, + 0.484, + 0.312 + ], + "angle": 0, + "content": "Construct" + }, + { + "type": "text", + "bbox": [ + 0.453, + 0.324, + 0.479, + 0.33 + ], + "angle": 0, + "content": "Check" + }, + { + "type": "text", + "bbox": [ + 0.561, + 0.268, + 0.645, + 0.276 + ], + "angle": 0, + "content": "Single-round QA pairs" + }, + { + "type": "text", + "bbox": [ + 0.54, + 0.281, + 0.618, + 0.289 + ], + "angle": 0, + "content": "1. World & Entities" + }, + { + "type": "text", + "bbox": [ + 0.54, + 0.289, + 0.663, + 0.297 + ], + "angle": 0, + "content": "2. Player Mechanics & Survival" + }, + { + "type": "text", + "bbox": [ + 0.54, + 0.297, + 0.645, + 0.305 + ], + "angle": 0, + "content": "3. Knowledge & Discovery" + }, + { + "type": "text", + "bbox": [ + 0.54, + 0.305, + 0.635, + 0.313 + ], + "angle": 0, + "content": "4. Resources & Crafting" + }, + { + "type": "text", + "bbox": [ + 0.54, + 0.313, + 0.616, + 0.321 + ], + "angle": 0, + "content": "5. Tools & Utilities" + }, + { + "type": "text", + "bbox": [ + 0.54, + 0.321, + 0.609, + 0.331 + ], + "angle": 0, + "content": "6. Miscellaneous" + }, + { + "type": "list", + "bbox": [ + 0.54, + 0.281, + 0.663, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.36, + 0.788, + 0.498 + ], + "angle": 0, + "content": "Fig.3: STEVE-21K collection pipeline. In the Vision-Environment section, STEVE-13B plays the game according to specified tasks defined by the human player, collecting visual information through prismarine-viewer and capturing environmental information from the screen using Ray Tracing [60]. Note that the language instruction task is also performed during the collection phase. We simultaneously record and save the chat flow from the reasoning and decomposition stages. In the Question-Answering section, we obtain information from the Minecraft-Wiki and Reddit forums and use GPT-3.5 to clean the data into single-round QA pairs. In the Skill-Code section, we use GPT-3.5 combined with the human player's code to synthesize code snippets and then check and revise them in the game environment." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.526, + 0.785, + 0.571 + ], + "angle": 0, + "content": "train STEVE-13B. The GPT-3.5 [5] is employed to derive meaningful single-round question-answer pairs, and LoRA [22] is incorporated during the fin-tuning process for efficient resource allocation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.571, + 0.787, + 0.632 + ], + "angle": 0, + "content": "Skill-Code pairs contain 210 skill execution scripts with descriptions, covering 8 skill types including collecting, crafting, exploration etc. The code part is collected by manual coding. We use GPT-3.5 [5] to describe all codes and utilize langchain vecordb to give all pairs a database vector." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.653, + 0.375, + 0.67 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.682, + 0.427, + 0.698 + ], + "angle": 0, + "content": "5.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.842 + ], + "angle": 0, + "content": "We train STEVE-7B/13B, which is finetuned from LLaMA-2 [18] with the Question-Answering pair in STEVE-21K dataset for warm-up and simulation context data from successful runs. We use LoRA [22] for finetuning process. Note that the process is to adjust STEVE-13B to work on correct simulation knowledge while remaining adapted to visual perception. In the text part, we set all temperatures to 0 except for the task proposal, which uses 0.9 to encourage task diversity. The vision unit is based on EfficientFormerV2-S0 [31], which is trained on the Vision-Environment part of our STEVE-21K dataset. Our simulation environment is built on top of MineDojo [17] and leverages Mineflayer [45]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "Z. Zhao et al." + }, + { + "type": "image", + "bbox": [ + 0.237, + 0.147, + 0.77, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.477, + 0.789, + 0.52 + ], + "angle": 0, + "content": "Fig. 4: Example of Vision-Environment pairs. It represents the data format of the Vision-Environment pairs in our STEVE-21K dataset: including visual signals, environmental information, Chat Log, Context QA pairs, and planning in actual tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.546, + 0.788, + 0.577 + ], + "angle": 0, + "content": "We use GPT-4-0613 for all GPT-4 models used in Voyager [57] and code generation tasks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.597, + 0.341, + 0.611 + ], + "angle": 0, + "content": "5.2Baselines" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.619, + 0.788, + 0.667 + ], + "angle": 0, + "content": "As no vision-based LLM-driven agents are immediately operable in Minecraft, we selected several algorithms as baselines that extract information from a system's backend, differing significantly from real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.677, + 0.788, + 0.74 + ], + "angle": 0, + "content": "AutoGPT [49] is an NLP automation tool that decomposes a high-level goal into executable subgoals in MineDojo, aligning with our experimental framework. Our setup, AutoGPT, powered by GPT-4 [44], processes agent states, environment feedback, and execution errors to manage subgoal execution." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Voyager [57] relies solely on textual grounding for perception and features a long-term procedural memory with a hierarchical library of code-based procedures, allowing the integration of simple skills into complex behaviors. Proficient in environment exploration and tech tree mastery, Voyager uses GPT-4 [44] for processing background text in embodied agents, with a lesser emphasis on visual perception." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.677, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "STEVE" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 2: Comparison on tech tree mastery task. The values presented are in fractions, representing successful trials out of three attempts. A score of \\(0/3\\) signifies the method's inability to progress within the tech tree after a maximum of 160 prompting iterations. The reported numbers denote the average iterations across three trials. Lower iteration values indicate higher efficiency of the respective method." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.227, + 0.782, + 0.31 + ], + "angle": 0, + "content": "
MethodWooden ToolStone ToolIron ToolDiamond Tool
AutoGPT [49]92 ± 72 (3/3)94 ± 72 (3/3)135 ± 103 (3/3)N/A (0/3)
Voyager [57]6 ± 2 (3/3)11 ± 2 (3/3)21 ± 7 (3/3)102 (1/3)
STEVE4 ± 1 (3/3)8 ± 1 (3/3)15 ± 2 (3/3)106 ± 12 (3/3)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.326, + 0.788, + 0.424 + ], + "angle": 0, + "content": "Table 3: Quantitative comparison on knowledge question and answering task. Questions, model-generated responses, and ground truth inputs are evaluated in GPT-4 [5], Claude-2 [8] and human blind rating rated on a scale of 0 to 10; The scores above are the average of them. Higher scores indicate greater alignment of the generated answers with the ground truth. Wld., Ent., Mech., Surv., Know., Disc., Res., Craft., Tools, Util., Miscell. stand for World, Entities, Player Mechanics, Survival, Knowledge, Discovery, Resources, Crafting, Tools, Utilities and Miscellaneous." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.437, + 0.782, + 0.53 + ], + "angle": 0, + "content": "
MethodWld. & Ent.Mech. & Surv.Know. & Disc.Res. & Craft.Tl. & Util.Miscell.Overall
Llama2-7B6.446.686.586.426.806.966.56
Llama2-13B6.936.956.776.776.986.646.89
STEVE-7B7.997.887.847.957.937.827.94
STEVE-13B8.148.138.038.158.127.728.12
GPT-48.068.078.077.928.098.218.04
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.562, + 0.419, + 0.576 + ], + "angle": 0, + "content": "5.3 Evaluation Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.625, + 0.496, + 0.775 + ], + "angle": 0, + "content": "Continuous block search. As shown in Tab. 1, we experiment with block-searching tasks to assess the agent's exploratory capabilities and proficiency in locating specified blocks. Diamond blocks are placed at every 16-block interval across the land map. The agent's objective is to identify as many blocks as possible within the fewest iterations, which indicates the method's efficiency." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.775, + 0.787, + 0.82 + ], + "angle": 0, + "content": "As shown in Fig. 5, enriching information through visual perception significantly enhances the efficiency of search and exploration tasks, leading to more effective world exploration." + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.629, + 0.788, + 0.698 + ], + "angle": 0, + "content": "Table 1: Comparison on continues block search task. # Iters stand for average iterations to find 10 diamonds (max 100). # Blocks stand for average diamonds found in 100 iterations." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.7, + 0.789, + 0.775 + ], + "angle": 0, + "content": "
Method# Iters (↓) # Blocks (↑)
AutoGPT [49]N/A 7
Voyager [57]35 26
STEVE14 67
" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.127 + ], + "angle": 0, + "content": "Z. Zhao et al." + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.145, + 0.728, + 0.471 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.482, + 0.788, + 0.54 + ], + "angle": 0, + "content": "Fig.5: Schematic of the continuous block search task. We capture an asynchronous segment with each method 30 iterations from the experiments for illustration. The reason we choose diamond blocks is that they are not naturally occurring in the given context, making them easily distinguishable from other blocks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.788, + 0.675 + ], + "angle": 0, + "content": "Knowledge question and answering. Using a validation dataset, we established a question-answering database to evaluate our model's performance on Minecraft-related queries. Responses from each model are rated blindly by GPT-4, Claude2, and human participants based on accuracy, relevance, and detail. Responses are initially checked for accuracy, and the comprehensive evaluation yields a score ranging from 0 to 10, with higher scores indicating superior performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.682, + 0.788, + 0.788 + ], + "angle": 0, + "content": "As detailed in Tab. 3, we tested the instructional capabilities of various LLM models using the STEVE-21k dataset, divided into 18,622 training and 1,000 testing samples. The STEVE-7B and STEVE-13B models surpassed LLaMA2 in all metrics, with STEVE-13B scoring the highest at 8.12, demonstrating its exceptional understanding of Minecraft-related queries. This indicates that STEVE models, optimized specifically for Minecraft content, perform better in knowledge-intensive tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.79, + 0.84 + ], + "angle": 0, + "content": "Our results confirm that larger, domain-specifically tuned models like STEVE-13B outperform broader models like GPT-4, underscoring the benefit of specialized fine-tuning for domain-specific applications." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.678, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "STEVE" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.147, + 0.359, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.147, + 0.501, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.147, + 0.642, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.147, + 0.784, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.208, + 0.359, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.208, + 0.501, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.209, + 0.642, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.209, + 0.784, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.231, + 0.359, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.231, + 0.501, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.231, + 0.642, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.231, + 0.784, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.287, + 0.359, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.287, + 0.501, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.287, + 0.642, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.287, + 0.784, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.332, + 0.359, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.332, + 0.501, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.332, + 0.609, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.625, + 0.332, + 0.784, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.405, + 0.788, + 0.452 + ], + "angle": 0, + "content": "Fig. 6: Qualitative comparison on knowledge question and answering tasks. Green (Red) highlights the correct or good (wrong or bad) answer. Blue indicates the suboptimal answer. Grey indicates the meaningless answer." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.483, + 0.789, + 0.664 + ], + "angle": 0, + "content": "Tech tree mastery. As shown in Tab. 2, we experiment on the Minecraft tech tree mastery to test the agent's ability to craft and use a hierarchy of tools. Progressing through this tree (wooden tool \\(\\rightarrow\\) stone tool \\(\\rightarrow\\) iron tool \\(\\rightarrow\\) diamond tool) requires the agent to master systematic and compositional skills. As to the wooden, stone, and iron levels of the tech tree, STEVE achieves remarkable efficiency: \\(23 \\times\\), \\(11.8 \\times\\), and \\(9 \\times\\) faster than AutoGPT [49], and \\(1.5 \\times\\), \\(1.4 \\times\\), and \\(1.3 \\times\\) faster than Voyager [57]. STEVE has achieved the diamond level, as shown in Tab. 2. Its performance slightly lags behind Voyager [57], which also uses GPT4 [44] for critical inference. However, STEVE is more cost-effective, starting with lower initial performance. It includes a vision unit, prioritizing visual data over background information, offering distinct advantages. Additionally, we observed a decrease in performance when using a basic skill database." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.689, + 0.389, + 0.705 + ], + "angle": 0, + "content": "5.4 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.717, + 0.788, + 0.837 + ], + "angle": 0, + "content": "To understand the impact of different components on the performance of our system, we conducted ablation studies focusing on the tech tree mastery task in Minecraft. The results, as shown in Tab. 4, provide insights into the effectiveness of the vision unit and compare our STEVE model with the STEVE GPT-4 version (with the same vision unit as ours). Note that the w/o vision unit setup is that the environmental perception encompasses data on blocks within an 8x8 area surrounding the agent, including the front, back, left, and right directions. The following observations are made:" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "Z. Zhao et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 4: Ablation studies for the tech tree mastery. STEVE (Ours) is the STEVE-13B version. The \\(0/3\\) score means the method can't progress beyond 160 iterations in the tech tree." + }, + { + "type": "table", + "bbox": [ + 0.248, + 0.199, + 0.751, + 0.276 + ], + "angle": 0, + "content": "
MethodWooden ToolStone ToolIron ToolDiamond Tool
w/o vision unit11 ± 5 (3/3)27 ± 5 (3/3)46 ± 11 (3/3)158 (1/3)
STEVE (GPT-4)6 ± 2 (3/3)10 ± 1 (3/3)14 ± 3 (3/3)89 ± 9 (3/3)
STEVE (Ours)4 ± 1 (3/3)8 ± 1 (3/3)15 ± 2 (3/3)106 ± 12 (3/3)
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.307, + 0.788, + 0.382 + ], + "angle": 0, + "content": "Vision unit is critical. The omission of the vision unit markedly affects the system's performance, especially in more advanced tasks. While it successfully crafts Wooden, Stone, and Iron Tools, it is challenged with Diamond Tools. This outcome underscores the vital importance of visual information in accomplishing complex tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.405, + 0.788, + 0.496 + ], + "angle": 0, + "content": "Comparison with GPT-4. As our vision encoder directly encodes into text space, it can be easily replaced with any language model. For instance, the GPT-4 we compared exhibits consistent success across all categories and secures a flawless success rate. Interestingly, the STEVE-13B version excels in simpler tasks such as crafting wooden and stone tools. Moreover, it requires fewer iterations than methods without the vision part, underscoring its superior efficiency." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.526, + 0.357, + 0.542 + ], + "angle": 0, + "content": "5.5 Case Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.559, + 0.788, + 0.666 + ], + "angle": 0, + "content": "As shown in Fig. 6, we perform an extensive case study comparison of GPT-4, LLaMA2-13B, and our method STEVE-13B. Each model maintains the same information and question inputs to compare feedback under different environmental information. Our STEVE overall achieves the best results, surpassing GPT-4 and showing significant improvement compared to the original LLaMA. Especially in parts involving numerical calculations, such as the leftmost image, STEVE accurately tracks food values to restore hunger levels." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.695, + 0.36, + 0.712 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "STEVE enhances multi-modal learning by combining visual encoder and LLM-based agents. It has three functions: vision perception, language instruction, and code action, allowing it to understand, predict, and act in virtual environments. We provide a straightforward approach to creating a robust, multimodal, autonomous, embodied agent using an open-source language model with a small number of parameters. Additionally, we provide a comprehensive dataset STEVE-21K for sustainable community development that can be verified." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.677, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "STEVE" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.788, + 0.225 + ], + "angle": 0, + "content": "This work is supported by the Zhejiang Provincial Natural Science Foundation of China (No. LZ24F030005) and the National Natural Science Foundation of China (No. 62106219)." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.249, + 0.323, + 0.265 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.282, + 0.786, + 0.337 + ], + "angle": 0, + "content": "1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems 35, 23716-23736 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.339, + 0.786, + 0.365 + ], + "angle": 0, + "content": "2. Amiranashvili, A., Dorka, N., Burgard, W., Koltun, V., Brox, T.: Scaling imitation learning in apacheft. arXiv preprint arXiv:2007.02701 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.367, + 0.786, + 0.42 + ], + "angle": 0, + "content": "3. Baker, B., Akkaya, I., Zhokov, P., Huizinga, J., Tang, J., Ecoffet, A., Houghton, B., Sampedro, R., Clune, J.: Video pretraining (vpt): Learning to act by watching unlabeled online videos. Advances in Neural Information Processing Systems 35, 24639-24654 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.422, + 0.786, + 0.449 + ], + "angle": 0, + "content": "4. Chai, W., Wang, G.: Deep vision multimodal learning: Methodology, benchmark, and trend. Applied Sciences 12(13), 6588 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.451, + 0.674, + 0.464 + ], + "angle": 0, + "content": "5. Introducing chatgpt (2022), https://openai.com/blog/chatgpt" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.465, + 0.786, + 0.505 + ], + "angle": 0, + "content": "6. Chen, S., Guhur, P.L., Schmid, C., Laptev, I.: History aware multimodal transformer for vision-and-language navigation. Advances in neural information processing systems 34, 5834-5847 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.507, + 0.786, + 0.533 + ], + "angle": 0, + "content": "7. Chen, T., Gupta, S., Gupta, A.: Learning exploration policies for navigation. In: International Conference on Learning Representations (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.535, + 0.529, + 0.547 + ], + "angle": 0, + "content": "8. Talk to claude (2023), https://claude.ai" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.549, + 0.786, + 0.589 + ], + "angle": 0, + "content": "9. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.591, + 0.786, + 0.631 + ], + "angle": 0, + "content": "0. Das, A., Datta, S., Gkioxari, G., Lee, S., Parikh, D., Batra, D.: Embodied question answering. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1-10 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.633, + 0.786, + 0.673 + ], + "angle": 0, + "content": "1. Datta, S., Dharur, S., Cartillier, V., Desai, R., Khanna, M., Batra, D., Parikh, D.: Episodic memory question answering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19119-19128 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.675, + 0.786, + 0.714 + ], + "angle": 0, + "content": "2. Dean, V., Tulsiani, S., Gupta, A.: See, hear, explore: Curiosity via audio-visual association. Advances in neural information processing systems 33, 14961-14972 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.716, + 0.786, + 0.755 + ], + "angle": 0, + "content": "3. Deng, J., Chai, W., Guo, J., Huang, Q., Hu, W., Hwang, J.N., Wang, G.: Citygen: Infinite and controllable 3d city layout generation. arXiv preprint arXiv:2312.01508 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.758, + 0.786, + 0.798 + ], + "angle": 0, + "content": "4. Deng, J., Chai, W., Huang, J., Zhao, Z., Huang, Q., Gao, M., Guo, J., Hao, S., Hu, W., Hwang, J.N., et al.: Citycraft: A real craftsfer for 3d city generation. arXiv preprint arXiv:2406.04983 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.786, + 0.84 + ], + "angle": 0, + "content": "5. Driess, D., Xia, F., Sajjadi, M.S., Lynch, C., Chowdhery, A., Ichter, B., Wahid, A., Tompson, J., Vuong, Q., Yu, T., et al.: Palm-e: An embodied multimodal language model. arXiv preprint arXiv:2303.03378 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.282, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "Z. Zhao et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "16. Du, H., Yu, X., Zheng, L.: Vtnet: Visual transformer network for object goal navigation. In: International Conference on Learning Representations (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.787, + 0.229 + ], + "angle": 0, + "content": "17. Fan, L., Wang, G., Jiang, Y., Mandlekar, A., Yang, Y., Zhu, H., Tang, A., Huang, D.A., Zhu, Y., Anandkumar, A.: Minedojo: Building open-ended embodied agents with internet-scale knowledge. Advances in Neural Information Processing Systems 35, 18343-18362 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.231, + 0.786, + 0.271 + ], + "angle": 0, + "content": "18. Gao, P., Han, J., Zhang, R., Lin, Z., Geng, S., Zhou, A., Zhang, W., Lu, P., He, C., Yue, X., et al.: Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.272, + 0.787, + 0.311 + ], + "angle": 0, + "content": "19. Gong, T., Lyu, C., Zhang, S., Wang, Y., Zheng, M., Zhao, Q., Liu, K., Zhang, W., Luo, P., Chen, K.: Multimodal-gpt: A vision and language model for dialogue with humans. arXiv preprint arXiv:2305.04790 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.312, + 0.787, + 0.352 + ], + "angle": 0, + "content": "20. Guss, W.H., Houghton, B., Topin, N., Wang, P., Codel, C., Veloso, M., Salakhutdinov, R.: Minerl: A large-scale dataset of apache demonstrations. arXiv preprint arXiv:1907.13440 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.353, + 0.787, + 0.379 + ], + "angle": 0, + "content": "21. Hofmann, K.: Minecraft as ai playground and laboratory. In: Proceedings of the annual symposium on computer-human interaction in play. pp. 1-1 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.787, + 0.419 + ], + "angle": 0, + "content": "22. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.42, + 0.787, + 0.46 + ], + "angle": 0, + "content": "23. Huang, W., Xia, F., Xiao, T., Chan, H., Liang, J., Florence, P., Zeng, A., Tompson, J., Mordatch, I., Chebotar, Y., et al.: Inner monologue: Embodied reasoning through planning with language models. arXiv preprint arXiv:2207.05608 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.461, + 0.787, + 0.501 + ], + "angle": 0, + "content": "24. Jiang, J., Ye, T., Bai, J., Chen, S., Chai, W., Jun, S., Liu, Y., Chen, E.: Five a+ network: You only need 9k parameters for underwater image enhancement. arXiv preprint arXiv:2305.08824 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.502, + 0.787, + 0.528 + ], + "angle": 0, + "content": "25. Johnson, M., Hofmann, K., Hutton, T., Bignell, D.: The malmo platform for artificial intelligence experimentation. In: Ijcai. pp. 4246-4247 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.529, + 0.787, + 0.569 + ], + "angle": 0, + "content": "26. Kolve, E., Mottaghi, R., Han, W., VanderBilt, E., Weihs, L., Herrasti, A., Deitke, M., Ehsani, K., Gordon, D., Zhu, Y., et al.: Ai2-thor: An interactive 3d environment for visual ai. arXiv preprint arXiv:1712.05474 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.569, + 0.787, + 0.609 + ], + "angle": 0, + "content": "27. Kwon, O., Park, J., Oh, S.: Rendering neural radiance map for visual navigation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9099-9108 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.61, + 0.787, + 0.637 + ], + "angle": 0, + "content": "28. Li, B., Zhang, Y., Chen, L., Wang, J., Yang, J., Liu, Z.: Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.637, + 0.787, + 0.677 + ], + "angle": 0, + "content": "29. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.678, + 0.787, + 0.718 + ], + "angle": 0, + "content": "30. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.719, + 0.787, + 0.759 + ], + "angle": 0, + "content": "31. Li, Y., Yuan, G., Wen, Y., Hu, J., Evangelidis, G., Tulyakov, S., Wang, Y., Ren, J.: Efficientformer: Vision transformers at mobilenet speed. Advances in Neural Information Processing Systems 35, 12934-12949 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.787, + 0.8 + ], + "angle": 0, + "content": "32. Lifshitz, S., Paster, K., Chan, H., Ba, J., McIlraith, S.: Steve-1: A generative model for text-to-behavior in apache (abridged version). In: NeurIPS 2023 Workshop on Goal-Conditioned Reinforcement Learning (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.787, + 0.84 + ], + "angle": 0, + "content": "33. Lin, Z., Li, J., Shi, J., Ye, D., Fu, Q., Yang, W.: Juewu-mc: Playing minecraft with sample-efficient hierarchical reinforcement learning. arXiv preprint arXiv:2112.04907 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.677, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "STEVE" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "34. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv preprint arXiv:2304.08485 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.218 + ], + "angle": 0, + "content": "35. Liu, S., Okatani, T.: Symmetry-aware neural architecture for embodied visual exploration. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 17221-17230. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.219, + 0.785, + 0.261 + ], + "angle": 0, + "content": "36. Lu, P., Peng, B., Cheng, H., Galley, M., Chang, K.W., Wu, Y.N., Zhu, S.C., Gao, J.: Chameleon: Plug-and-play compositional reasoning with large language models. arXiv preprint arXiv:2304.09842 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.262, + 0.785, + 0.303 + ], + "angle": 0, + "content": "37. Luo, W., Sun, P., Zhong, F., Liu, W., Zhang, T., Wang, Y.: End-to-end active object tracking via reinforcement learning. In: International conference on machine learning. pp. 3286-3295. PMLR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.304, + 0.785, + 0.346 + ], + "angle": 0, + "content": "38. Luo, W., Sun, P., Zhong, F., Liu, W., Zhang, T., Wang, Y.: End-to-end active object tracking and its real-world deployment via reinforcement learning. IEEE transactions on pattern analysis and machine intelligence 42(6), 1317-1332 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.346, + 0.785, + 0.388 + ], + "angle": 0, + "content": "39. Lyu, C., Wu, M., Wang, L., Huang, X., Liu, B., Du, Z., Shi, S., Tu, Z.: Macaw-llm: Multi-modal language modeling with image, audio, video, and text integration. arXiv preprint arXiv:2306.09093 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.389, + 0.785, + 0.43 + ], + "angle": 0, + "content": "40. Ma, W., Mi, Q., Yan, X., Wu, Y., Lin, R., Zhang, H., Wang, J.: Large language models play starcraft ii: Benchmarks and a chain of summarization approach. arXiv preprint arXiv:2312.11865 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.431, + 0.785, + 0.472 + ], + "angle": 0, + "content": "41. Maaz, M., Rasheed, H., Khan, S., Khan, F.S.: Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.785, + 0.529 + ], + "angle": 0, + "content": "42. Mao, H., Wang, C., Hao, X., Mao, Y., Lu, Y., Wu, C., Hao, J., Li, D., Tang, P.: Seihai: A sample-efficient hierarchical ai for the minerl competition. In: Distributed Artificial Intelligence: Third International Conference, DAI 2021, Shanghai, China, December 17-18, 2021, Proceedings 3. pp. 38-51. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.53, + 0.785, + 0.572 + ], + "angle": 0, + "content": "43. Moudgil, A., Majumdar, A., Agrawal, H., Lee, S., Batra, D.: Soat: A scene-and object-aware transformer for vision-and-language navigation. Advances in Neural Information Processing Systems 34, 7357-7367 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.573, + 0.772, + 0.586 + ], + "angle": 0, + "content": "44. OpenAI: Gpt-4 technical report. arXiv preprint arXiv: Arxiv-2303.08774 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.587, + 0.785, + 0.628 + ], + "angle": 0, + "content": "45. PrismarineJS: Prismarinejs/mineflayer: Create apache bots with a powerful, stable, and high level javascript api. (2013), https://github.com/PrismarineJS/mineflayer/tree/master" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.63, + 0.785, + 0.685 + ], + "angle": 0, + "content": "46. Savva, M., Kadian, A., Maksymets, O., Zhao, Y., Wijmans, E., Jain, B., Straub, J., Liu, J., Koltun, V., Malik, J., et al.: Habitat: A platform for embodied ai research. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9339-9347 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.686, + 0.785, + 0.727 + ], + "angle": 0, + "content": "47. Schick, T., Dwivedi-Yu, J., Dessi, R., Raileanu, R., Lomeli, M., Zettlemoyer, L., Cancedda, N., Scialom, T.: Toolformer: Language models can teach themselves to use tools. arXiv preprint arXiv:2302.04761 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.728, + 0.785, + 0.769 + ], + "angle": 0, + "content": "48. Shen, Y., Song, K., Tan, X., Li, D., Lu, W., Zhuang, Y.: Hugginggpt: Solving ai tasks with chatgpt and its friends in huggingface. arXiv preprint arXiv:2303.17580 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.77, + 0.785, + 0.798 + ], + "angle": 0, + "content": "49. Significant-Gravitas: Auto-gpt. https://github.com/Significant-Gravitas/Auto-GPT (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.841 + ], + "angle": 0, + "content": "50. Skrynnik, A., Staroverov, A., Aitygulov, E., Aksenov, K., Davydov, V., Panov, A.I.: Hierarchical deep q-network from imperfect demonstrations in minecraft. Cognitive Systems Research 65, 74-78 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "Z. Zhao et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "51. Song, C.H., Wu, J., Washington, C., Sadler, B.M., Chao, W.L., Su, Y.: Llm-planner: Few-shot grounded planning for embodied agents with large language models. arXiv preprint arXiv:2212.04088 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.189, + 0.785, + 0.23 + ], + "angle": 0, + "content": "52. Song, E., Chai, W., Wang, G., Zhang, Y., Zhou, H., Wu, F., Guo, X., Ye, T., Lu, Y., Hwang, J.N., et al.: Moviechat: From dense token to sparse memory for long video understanding. arXiv preprint arXiv:2307.16449 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.23, + 0.785, + 0.271 + ], + "angle": 0, + "content": "53. Song, E., Chai, W., Ye, T., Hwang, J.N., Li, X., Wang, G.: Moviechat+: Question-aware sparse memory for long video question answering. arXiv preprint arXiv:2404.17176 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.271, + 0.785, + 0.298 + ], + "angle": 0, + "content": "54. Su, Y., Lan, T., Li, H., Xu, J., Wang, Y., Cai, D.: Pandagpt: One model to instruction-follow them all. arXiv preprint arXiv:2305.16355 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.298, + 0.785, + 0.338 + ], + "angle": 0, + "content": "55. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.338, + 0.785, + 0.393 + ], + "angle": 0, + "content": "56. Vasluianu, F.A., Seizinger, T., Zhou, Z., Wu, Z., Chen, C., Timofte, R., Dong, W., Zhou, H., Tian, Y., Chen, J., et al.: Ntire 2024 image shadow removal challenge report. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6547-6570 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.393, + 0.785, + 0.434 + ], + "angle": 0, + "content": "57. Wang, G., Xie, Y., Jiang, Y., Mandlekar, A., Xiao, C., Zhu, Y., Fan, L., Anandkumar, A.: Voyager: An open-ended embodied agent with large language models. arXiv preprint arXiv:2305.16291 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.434, + 0.785, + 0.474 + ], + "angle": 0, + "content": "58. Wang, W., Chen, Z., Chen, X., Wu, J., Zhu, X., Zeng, G., Luo, P., Lu, T., Zhou, J., Qiao, Y., et al.: Visionllm: Large language model is also an open-ended decoder for vision-centric tasks. arXiv preprint arXiv:2305.11175 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.474, + 0.785, + 0.515 + ], + "angle": 0, + "content": "59. Wang, Z., Cai, S., Liu, A., Ma, X., Liang, Y.: Describe, explain, plan and select: Interactive planning with large language models enables open-world multi-task agents. arXiv preprint arXiv:2302.01560 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.515, + 0.785, + 0.542 + ], + "angle": 0, + "content": "60. Whitted, T.: An improved illumination model for shaded display. In: ACM SIGGRAPH 2005 Courses, pp. 4-es (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.542, + 0.785, + 0.582 + ], + "angle": 0, + "content": "61. Wijmans, E., Kadian, A., Morcos, A., Lee, S., Essa, I., Parikh, D., Savva, M., Batra, D.: Dd-ppo: Learning near-perfect pointgoal navigators from 2.5 billion frames. In: International Conference on Learning Representations (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.582, + 0.785, + 0.623 + ], + "angle": 0, + "content": "62. Wu, C., Yin, S., Qi, W., Wang, X., Tang, Z., Duan, N.: Visual chatgpt: Talking, drawing and editing with visual foundation models. arXiv preprint arXiv:2303.04671 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.623, + 0.785, + 0.651 + ], + "angle": 0, + "content": "63. Xie, Y., Yu, C., Zhu, T., Bai, J., Gong, Z., Soh, H.: Translating natural language to planning goals with large-language models. arXiv preprint arXiv:2302.05128 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.651, + 0.785, + 0.691 + ], + "angle": 0, + "content": "64. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.691, + 0.785, + 0.732 + ], + "angle": 0, + "content": "65. Ye, T., Chen, S., Liu, Y., Chai, W., Bai, J., Zou, W., Zhang, Y., Jiang, M., Chen, E., Xue, C.: Sequential affinity learning for video restoration. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 4147-4156 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.732, + 0.785, + 0.772 + ], + "angle": 0, + "content": "66. Ye, T., Zhang, Y., Jiang, M., Chen, L., Liu, Y., Chen, S., Chen, E.: Perceiving and modeling density for image dehazing. In: European conference on computer vision. pp. 130-145. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.813 + ], + "angle": 0, + "content": "67. Yu, L., Chen, X., Gkioxari, G., Bansal, M., Berg, T.L., Batra, D.: Multi-target embodied question answering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6309-6318 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "68. Yu, Y., Huang, W., Sun, F., Chen, C., Wang, Y., Liu, X.: Sound adversarial audiovisual navigation. In: International Conference on Learning Representations (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.677, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "STEVE" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "69. Yuan, H., Zhang, C., Wang, H., Xie, F., Cai, P., Dong, H., Lu, Z.: Plan4mc: Skill reinforcement learning and planning for open-world apache tasks. arXiv preprint arXiv:2303.16563 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.788, + 0.231 + ], + "angle": 0, + "content": "70. Zhao, Z., Chai, W., Hao, S., Hu, W., Wang, G., Cao, S., Song, M., Hwang, J.N., Wang, G.: A survey of deep learning in sports applications: Perception, comprehension, and decision. arXiv preprint arXiv:2307.03353 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.231, + 0.788, + 0.272 + ], + "angle": 0, + "content": "71. Zhao, Z., Chai, W., Wang, X., Ma, K., Chen, K., Guo, D., Ye, T., Zhang, Y., Wang, H., Wang, G.: Steve series: Step-by-step construction of agent systems in northwest China. arXiv preprint arXiv:2406.11247 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.272, + 0.788, + 0.313 + ], + "angle": 0, + "content": "72. Zhao, Z., Chen, K., Guo, D., Chai, W., Ye, T., Zhang, Y., Wang, G.: Hierarchical auto-organizing system for open-ended multi-agent navigation. arXiv preprint arXiv:2403.08282 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.314, + 0.788, + 0.355 + ], + "angle": 0, + "content": "73. Zhao, Z., Ma, K., Chai, W., Wang, X., Chen, K., Guo, D., Zhang, Y., Wang, H., Wang, G.: Do we really need a complex agent system? distill embodied agent into a single model. arXiv preprint arXiv:2404.04619 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.356, + 0.788, + 0.397 + ], + "angle": 0, + "content": "74. Zhong, F., Sun, P., Luo, W., Yan, T., Wang, Y.: Ad-vat+: An asymmetric dueling mechanism for learning and understanding visual active tracking. IEEE transactions on pattern analysis and machine intelligence 43(5), 1467-1482 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.397, + 0.788, + 0.438 + ], + "angle": 0, + "content": "75. Zhong, F., Sun, P., Luo, W., Yan, T., Wang, Y.: Towards distraction-robust active visual tracking. In: International Conference on Machine Learning. pp. 12782-12792. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.439, + 0.788, + 0.48 + ], + "angle": 0, + "content": "76. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.48, + 0.788, + 0.535 + ], + "angle": 0, + "content": "77. Zhu, X., Chen, Y., Tian, H., Tao, C., Su, W., Yang, C., Huang, G., Li, B., Lu, L., Wang, X., et al.: Ghost in the minecraft: Generally capable agents for open-world enviroments via large language models with text-based knowledge and memory. arXiv preprint arXiv:2305.17144 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.535 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/See and Think_ Embodied Agent in Virtual Environment/0ba9ae38-8422-4b1f-90c0-ee0bf0ad307a_origin.pdf b/2024/See and Think_ Embodied Agent in Virtual Environment/0ba9ae38-8422-4b1f-90c0-ee0bf0ad307a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..09e78cc45b462e35e4dfab11c3fdce59894756f1 --- /dev/null +++ b/2024/See and Think_ Embodied Agent in Virtual Environment/0ba9ae38-8422-4b1f-90c0-ee0bf0ad307a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68f8cc2085d0b74ce63965e183b647d8701e53ba19153b338681419b3e40b90e +size 2674300 diff --git a/2024/See and Think_ Embodied Agent in Virtual Environment/full.md b/2024/See and Think_ Embodied Agent in Virtual Environment/full.md new file mode 100644 index 0000000000000000000000000000000000000000..87fab28c925d4b45c0710c94cd0baf750ed46dce --- /dev/null +++ b/2024/See and Think_ Embodied Agent in Virtual Environment/full.md @@ -0,0 +1,401 @@ +# See and Think: Embodied Agent in Virtual Environment + +Zhonghan Zhao $^{1, \spadesuit}$ , Wenhao Chai $^{2, \spadesuit}$ , Xuan Wang $^{1, \spadesuit}$ , Boyi Li $^{1}$ , Shengyu Hao $^{1}$ , Shidong Cao $^{1}$ , Tian Ye $^{3}$ , and Gaoang Wang $^{1, \dagger}$ + +1 Zhejiang University + +2 University of Washington + +$^{3}$ Hong Kong University of Science and Technology (GZ) + +Abstract. Large language models (LLMs) have achieved impressive progress on several open-world tasks. Recently, using LLMs to build embodied agents has been a hotspot. This paper proposes STEVE, a comprehensive and visionary embodied agent in the Minecraft virtual environment. STEVE comprises three key components: vision perception, language instruction, and code action. Vision perception involves interpreting visual information in the environment, which is then integrated into the LLMs component with agent state and task instruction. Language instruction is responsible for iterative reasoning and decomposing complex tasks into manageable guidelines. Code action generates executable skill actions based on retrieval in skill database, enabling the agent to interact effectively within the Minecraft environment. We also collect STEVE-21K dataset, which includes $600+$ vision-environment pairs, 20K knowledge question-answering pairs, and $200+$ skill-code pairs. We conduct continuous block search, knowledge question and answering, and tech tree mastery to evaluate the performance. Extensive experiments show that STEVE achieves at most $1.5 \times$ faster unlocking key tech trees and $2.5 \times$ quicker in block search tasks. + +Keywords: Open-world embodied agent $\cdot$ Multimodal pre-training $\cdot$ Large language model + +# 1 Introduction + +Designing agents that demonstrate intelligent behavior and adaptability in open-world settings has been a longstanding and significant challenge in the field of artificial intelligence [13,14,26,46,70]. However, recent progress in the development of large language models (LLMs) [5,55] has exhibited their potential as versatile, general-purpose assistants. Recent innovations in agent design [57,59,69,77] + +![](images/8e20506f446eeac2546e04d44592fb05bd49ec21a9a6b5405d67407105162973.jpg) +Fig. 1: STEVE integrates the three parts: Vision Perception, Language Instruction, and Code Action, supported closely by our proposed STEVE-21K. It demonstrates commendable performance on Continuous Block Search, Knowledge QA, and Tech Tree Mastery. + +have effectively harnessed these advanced LLMs, tapping into their extensive world knowledge and reasoning abilities. This development has paved the way for agents, that are autonomously driven, to formulate and implement strategies and actions across a diverse array of skills and tasks in diverse open-world environments. + +In many open-world settings, like Minecraft, contemporary agents predominantly use LLMs for their textual interactions. However, this reliance on text for communication poses considerable limitations in their interactions within these worlds including low-level regrading cases [24, 56, 65, 66]. Minecraft, with its expansive and interactive sandbox environment [17, 20], demands a variety of skills from agents, ranging from crafting basic items to executing complex tasks. Yet, agents driven by LLMs often generate unpredictable outputs. The effectiveness of their interactions is largely contingent on meticulously crafted prompts [23], designed to align the LLM's understanding with the environmental context and the intended objectives. This process of prompt engineering is not only laborious but also fails to meet the goal of fostering autonomous, self-directed agents. Furthermore, textual communication has its limitations in naturally conveying certain concepts of the world, like crafting recipes, which are often more effectively communicated through vision. + +Players have the distinct capability to assimilate and convey information using both visual and textual channels, significantly enhancing our interactions with the world around us. Yet, the integration of LLM-based agents with multimodal inputs in open-ended environments remains an under-explored area. STEVE in STEVE-Series [71-73], is named after the protagonist of the game "Minecraft." It is our proposed framework to build an embodied agent based on the vision model and LLMs within an open world, as illustrated in Fig. 1. STEVE harnesses a vision model to perceive its surroundings visually, coupled with an LLM to strategize and plan actions. This model represents a leap forward in agent design, combining these two input modes, vision, and text, to offer a more nuanced and comprehensive understanding of the environment and practical and executable skills. + +Our key contributions are outlined as follows: + +- We propose STEVE, an embodied agent in virtual environment, consists of vision perception, language instruction, and code action, achieving $1.5 \times$ faster unlocking of key tech trees and is $2.3 \times$ quicker in block search tasks compared to previous state-of-the-art methods. +- We present STEVE-7B/13B, a series of large language model obtained by fine-tuning with Minecraft knowledge question-answering pairs from Llama2-7B/13B. +- We collect STEVE-21K dataset, including $600+$ vision-environment pairs, 20K knowledge question-answering pairs, and $200+$ skill-code pairs, for justifying the effective performance of STEVE. + +# 2 Related Works + +# 2.1 Intelligent Agent in Minecraft + +As an open-ended sandbox game, Minecraft has always been an ideal setting for testing the performance of intelligent agents [21, 25]. The agents are required to autonomously perform various tasks in Minecraft, such as chopping trees, crafting tools, and mining diamonds. At the beginning, much of the works focus on exploring reinforcement learning [32, 33, 42, 50] or imitation learning [2, 3], without satisfactory performance. VPT [3] and MineDojo [17] collect internet-scale datasets for their model pre-training. More specifically, VPT offers the exciting possibility of directly learning to act during video pre-training and using these learned behavioral priors as extremely effective exploration priors for reinforcement learning. Yet, recent works found that the pre-trained LLMs could serve as a strong "mind" that provides planning ability to the agents. Voyager [57] leverages GPT-4 [44] as both a high-level planner and a low-level action code generator. Plan4MC [69] proposes a skill graph pre-generated by the LLMs. DEPS [59], an interactive planning method based on LLMs, addresses multi-step reasoning issue in open-world planning. GITM [77] develops a set of structured actions and leverages LLMs to generate action plans for the agents to execute, achieving impressive results in various tasks. + +# 2.2 Embodied Multimodal Model + +Embodied agent operates within various environments by synthesizing sensory perceptions and physical actions supported by computational intelligence. This synthesis enables the agent to undertake a variety of tasks, achieving specific objectives. Its key areas of application are diverse, including Navigation [6, 16, 27, 43, 61, 68], Embodied Question Answering [10, 11, 67], Active Visual Tracking [37, 38, 74, 75], and Visual Exploration [7, 12, 35]. The field is evolving rapidly with the development of Large Language Models (LLMs) [51] and Multimodal LLMs (MLLMs) [1, 9, 18, 19, 28–30, 34, 39, 41, 54, 58, 64, 76], integrating multiple modalities for more effective processing. A prime example of this innovation is + +![](images/1b252c5e77571c428318903d9dd735b64f6b4dec4c2328c8baa1f1769e6e9e2f.jpg) +Fig. 2: STEVE framework. The Vision Perception part takes images or videos, encodes them into tokens, and combines them with Agent State and Task tokens as input. The STEVE-13B in the Language Instruction part is used for automatic reasoning and task decomposition, and it calls the Skill Database in the form of the Query to output code as action. + +PaLM-E [15], a sophisticated multimodal model with 562B parameters, adept at a broad spectrum of embodied tasks and demonstrating exceptional capabilities in visual reasoning. + +# 2.3 Large Language Model with Equipped Tools + +While Large Language Models (LLMs) demonstrate impressive skill in tackling novel tasks via prompt-based instructions, they often face challenges in areas where simpler models or tools excel, like mathematical calculations or identifying palindromes. However, LLMs' potential is significantly expanded when integrated with other modality-specific models, such as those for vision or audio, enabling multi-modal capabilities [4,36]. Innovations like Toolformer [47] demonstrate LLMs' self-learning to utilize tools through finetuning with extensive API call samples. Visual ChatGPT [62] extends this by integrating various visual foundation models, facilitating interactive user experiences with ChatGPT. Similarly, HuggingGPT [48] presents a framework that harnesses LLMs to link diverse models from Hugging Face for task resolution. AutoGPT [49] is an open-source application that broadens GPT-4's capabilities with internet access, memory management, and plug-ins. The recent introduction of MovieChat [52,53] brings a memory mechanism to MLLM, enhancing its performance in video understanding tasks. Furthermore, LLMs can be used for goal planning, analogous to language translation [63]. This evolving landscape suggests that tool-equipped LLMs could forge a new paradigm in AI solution design. + +# 3 Method: STEVE + +# 3.1 Overview + +As shown in Figure 2, STEVE $\mathcal{F}$ is an LLM-based multi-modal autonomous system for embodied agents in Minecraft, which can use visual status $X^v$ in the + +form of images or videos and agent state $X^s$ to manage and execute complex task $X^t$ for executive code action $\mathbf{a}^c$ : + +$$ +\mathbf {a} ^ {c} = \mathcal {F} \left(X ^ {v}, X ^ {s}, X ^ {t}\right) = \mathcal {A} ^ {c} \left(\mathcal {I} ^ {l} \left(\mathcal {P} ^ {v} \left(X ^ {v}, X ^ {s}, X ^ {t}\right)\right)\right) \tag {1} +$$ + +where STEVE $\mathcal{F}$ integrates Visual Perception $\mathcal{P}^v$ into the Language Instruction $\mathcal{I}^l$ with large language model (LLM) and combines them with Code Action $\mathcal{A}^c$ , a skill retrieval method to execute actions, details are demonstrated as follows. + +- Vision Perception $\mathcal{P}^v$ (Section 3.2); a vision encoder to interpret visual information of the environment, such as blocks and entities and a text tokenizer for agent state $s$ and task $t$ . +- Language Instruction $\mathcal{I}^l$ (Section 3.3): a powerful language model system fine-tuned specifically for Minecraft content using LLaMA2-13B [18]. This model enables adaptive interaction for iterative reasoning and step-by-step decomposition. +- Code Action $\mathcal{A}^c$ (Section 3.4): a skill retrieval method based on our skill-code database. + +# 3.2 Vision Perception $\mathcal{P}^v$ + +The vision perception part includes a vision encoder $E^v$ and a text tokenizer $T$ , which converts the visual status $X^v$ , agent state $X^s$ , and task $X^t$ to the text-space tokenizer representation $\mathbf{Y} = \{Y^v, Y^s, Y^t\}$ : + +$$ +\mathbf {Y} _ {i} = \mathcal {P} ^ {v} \left(X _ {i} ^ {v}, X _ {i} ^ {s}, X _ {i} ^ {t}\right), \mathcal {P} ^ {v} = \left\{E ^ {v}, T \right\}, \tag {2} +$$ + +where the vision encoder $E$ , the visual branch of EfficientFormer [31], encodes visual status $X_{i}^{v}$ at each step $i$ into visual tokens $Y^{v} = \{y_{1}^{v}, y_{2}^{v}, \dots, y_{n}^{v}\} \in \mathbb{R}^{n \times d}$ , where $n$ denotes the number of visual tokens $y^{v}$ and $d$ is the dimensionality of each token. The text tokenizer converts agent state $X_{i}^{s}$ and task $X_{i}^{t}$ at each step $i$ in the form of text into textual tokens $Y^{s}$ and $Y^{t}$ . Note that visual tokens are amalgamated with textual tokens representing the agent's current state (e.g., health, inventory, etc.) and the task description. This is accomplished using a tokenizer that maps state variables and task parameters to a tokenized form. The resultant unified token set serves as a comprehensive representation of the current situational context. + +# 3.3 Language Instruction $\mathcal{I}^l$ + +The Language Instruction $\mathcal{I}^l = \{\mathcal{P}_l, \mathcal{C}_r, \mathcal{C}_u, \mathcal{D}_s\}$ , which consists of Planner $\mathcal{P}_l$ , Critic $\mathcal{C}_r$ , Curriculum $\mathcal{C}_u$ , and Descriptor $\mathcal{D}_s$ , i.e., four independent LLM-based agents with different functions. They formulate high-level task guidelines, refine strategies through feedback, facilitate continuous learning and adaptation through a curriculum of complex tasks, and finally decompose strategic guidelines into the executable low-level textual action step $\mathbf{a}^s$ , $\{\mathbf{a}^s \sim \mathcal{I}^l(\cdot|\mathbf{Y}_i)\}_{i=0}^N$ towards efficient completion of the task. + +Note that the Planner $\mathcal{P}_l$ , Critic $\mathcal{C}_r$ , Curriculum $\mathcal{C}_u$ and Descriptor $\mathcal{D}_s$ are based on STEVE-7B/13B, a powerful language model derived from LLaMA-2-13B [18], fine-tuned specifically on Minecraft-related content from the STEVE-21K. This model's expertise covers a broad spectrum of game-specific knowledge areas on worlds, entities, player mechanics, survival, and even game practical experience. Finally, they have the performance of their different functions: + +- Planner $\mathcal{P}_l$ : formulating comprehensive guidelines and executive plans that align with the overarching objectives of the task. +- Critic $\mathcal{C}_r$ : evaluating the planner decisions, providing feedback that can refine strategies. +- Curriculum $\mathcal{C}_u$ : facilitating continuous learning and adaptation for action agents by engaging with a series of progressively complex tasks. +- Descriptor $\mathcal{D}_s$ : distilling the extensive data into a concise summary, making it more manageable and interpretable. + +Iterative reasoning. The STEVE-13B receives a stream of tokens that encode the current visual scene, the agent's state, and the task's textual description. STEVE-13B interprets this rich context to undertake complex reasoning. The model initiates the reasoning process by constructing a series of high-level strategies that outline the pathway to task completion. Considering all gameplay elements, the reasoning mechanism is akin to an experienced player who can visualize the end game and chart a course to victory. This approach ensures the plans are reactive and proactive, allowing the agent to anticipate and mitigate future challenges. However, most strategies are high-level and abstract therefore, they often require step-by-step decomposition to derive executable guidelines. + +Decomposition. The decomposition process makes the complex strategies breakdown into simple, low-level guidelines that can be directly mapped to actions in Minecraft. It is similar to how a high-level command like "build a shelter" is divided into actionable instructions like "collect wood", "craft planks", and "place blocks". The granular steps are structured to provide explicit instructions that the game engine can readily interpret. This systematic breakdown from high-level reasoning to low-level actionable steps is the hallmark of the STEVE system, enabling the embodied agent to interact with the Minecraft environment in a meaningful and goal-oriented manner. Through this intricate process of reasoning and decomposition, STEVE embodies the cognitive capabilities required for sophisticated task management in virtual environments. + +Curriculum learning with memory. We draw inspiration from the lifelong learning strategy utilized in many reinforcement learning problems [57] in both closed-world and open-world settings. We start by creating a set of tasks that serve as a curriculum for agents to explore the environment. During this process, STEVE generates plans, interacts with the surroundings, learns from mistakes, and stores all these experiences in memory. Next, we evaluate STEVE on various tasks after this learning stage. Consequently, STEVE can produce better plans with its + +memory teaming up with the planning experiences. We use this as the default setting for all tasks in our experiments. + +Continuous learning with summarization. We've observed that the learning process, where the memory is being filled, can continue throughout the gameplay. The agent can gradually acquire more skills as the gameplay progresses and more experiences are gained. However, as the memory gets larger, it becomes difficult to understand the game's situation and interact with the game slowly. To tackle these challenges, we have implemented the Chain of Summarization method [40]. By finding better references, we can improve our ability to handle challenging tasks, such as "Diamond Tool" in the tech tree, including obtaining materials and creating diamond tools. This will lead to a higher success rate, as shown in Section 5.3. Additionally, curriculum learning with memory allows for in-context lifelong learning without needing gradient updates. + +# 3.4 Code Action $\mathcal{A}^c$ + +The code action part $\mathcal{A}^c$ is the execution phase, where the STEVE system converts planned, decomposed guidelines into concrete actions within the Minecraft environment. This process leverages a specialized skill database that pairs code snippets with their descriptions and relevant metadata, encoded as vectors $\mathbf{v}^{\mathbf{s}}$ for efficient retrieval. The transition from language instruction to executable code is achieved through the retrieval process $\mathcal{R}$ : + +$$ +\mathcal {R} (\mathbf {q}, \mathbf {v}) = \sigma (\mathbf {q}, \mathbf {v}), \mathbf {q} = E ^ {q} \left(\mathbf {a} ^ {\mathbf {s}}\right), \tag {3} +$$ + +where $E^q$ and $\sigma$ are query encoding and cosine similarity matching. Each low-level textual action step $\mathbf{a}^s$ derived from the Language Instruction is encoded into a query $\mathbf{q}$ . This encoding captures the essence of the action to be performed in a format that the skill database can understand. Once the queries $\mathbf{q}$ are encoded, the system computes similarities between the query vectors and the code snippets vectors $\mathbf{v}$ stored in the database to determine which skill best matches the required action. + +# 3.5 Training + +To reduce the training overhead to a certain extent, we adopt a two-stage training method, the first stage being a warm-up on the question-answering pairs of STEVE-21K to ensure a certain degree of instruction capability. The second stage is simulated in the environment, and the Expert LLM $\mathcal{E}_p$ integrated with GPT-4 [44] generates instructions for the same situation to modify our model. + +In training, we use either a single-round conversation $\{X^Q,X^A\}$ or a multiround conversation $\{X_i^Q,X_i^A\}_{i\leq N}$ , where $N$ is the total number of rounds in the conversation. We use the negative log-likelihood objective over the prediction + +tokens with training and finetuning: + +$$ +\mathcal {L} (\theta) = - \sum_ {j = 1} ^ {L} \log \mathcal {F} _ {\theta} \left(Y _ {j} \mid X ^ {v}, \hat {Y} _ {1: j - 1}\right), \tag {4} +$$ + +where $Y$ and $\hat{Y}$ refer to the non-vision input and target token sequences, $\theta$ represents the model parameters, and $L$ represents the length of the target sequence. The vision input $X^v$ varies depending on the input step. We only consider the answer tokens $X_A$ when computing the loss to ensure that the model focuses on generating coherent responses. + +Offline warm-up. In the first stage, we finetune the STEVE-7B/13B only on the single-round question-answering pairs of STEVE-21K to ensure a certain degree of instruction capability, as shown in Tab. 3. + +Online finetuning. In the second stage, we simulate the warmed-up STEVE-7B/13B in Minecraft, and both train the vision encoder $E^v$ and finetune the STEVE-7B/13B simultaneously. + +It is necessary to obtain the environment information to train the vision encoder, denoted as $l = R_{T}(X^{v})$ , where $l$ is the label of seen blocks and entities, $R_{T}$ is the Ray Tracing method [60] to eliminate all content outside the agent's perspective screen. We train the vision encoder $E^{v}$ once we have obtained the environment information. To finetune the warmed-up STEVE-7B/13B in simulation, we give both the STEVE-7B/13B and the Expert LLM the same input, and we consider the output of the Expert LLM as the ground truth label. + +After 5,000 simulations, we collect all the context information from successful runs, including sequences of vision and question-answering chat flow as the Vision-Environment part of STEVE-21K, as mentioned in Section 4. + +# 4 Dataset: STEVE-21K + +As shown in Fig. 3, STEVE-21K has been meticulously compiled, featuring a diverse collection of Vision-Environment pairs, Question-Answering pairs, and a Skill-Code database. + +Vision-Environment pairs contain 600 pairs of first-person perspective videos from Minecraft gameplay across six different terrains (including forest, desert, coastal, etc.), along with corresponding environmental block entities in the field of vision and context information for each timestamp. Additionally, all pairs are oriented around executing the skill-related task supported by Skill-Code part mentioned in Section 4. We employ the STEVE-7B/13B model to enable robots to plan and execute actions autonomously based on tasks defined by human supervisors. We record the video of the agent operation, the environment information, and all the corresponding chatflow. + +Question-Answering pairs contain 20K question-answering pairs from the Minecraft-Wiki and Reddit corpus across six data types partly sourced from [17]. The pairs are organized into instruction, input, and output triplets and used to + +![](images/4f0778d76d1ad6c52cc0e0aedd4767906814279c72cb3fab3da16fa83f00fe92.jpg) + +# Skill-Code + +Human Player + +![](images/3204dda95c0702225d88f8d15a4b7828f65ac2d6679bd4ed5068a39a667c6f58.jpg) + +prompt + +![](images/1e1670b491385802b28f5fe1e2fd9c1f605e4fa7f9fed088c8ab601eccfeb9b3.jpg) + +GPT + +output + +code snippets + +![](images/fec95c1c56cf7cfcb11205fcee473edbe1b0bd7a1c8b3d30f066c14f173805e6.jpg) + +execute + +![](images/02bd1387363039c3fd997487d4a4b920c9dd4bda4fc10bbffd4b6f2b5933a415.jpg) + +check and revise + +# Question-Answering + +![](images/1f57fa5bba76ddfef6a911c46b7ce601b83e13e345e685ec8366aa261fa44900.jpg) +Wikipedia + +The crafting menu in Minecraft, showing the crafting origins of a clone. + +showing the eating recipe or a special taste as well as some other blocks and + +![](images/60b81a3852d8af7061f20dd0b31c16210869d84a7930ee35b04a45fb03a288ee.jpg) +Reddit + +![](images/a54b823eedd10842692d2794e3523dd8248c7e2eced360f5ce1571ae17dfd7c9.jpg) + +ChatGPT + +![](images/2831b8b1e4fd69d911915dbd4407eae73d885f7845e3e2f70795cbf433d150d3.jpg) +Fig.3: STEVE-21K collection pipeline. In the Vision-Environment section, STEVE-13B plays the game according to specified tasks defined by the human player, collecting visual information through prismarine-viewer and capturing environmental information from the screen using Ray Tracing [60]. Note that the language instruction task is also performed during the collection phase. We simultaneously record and save the chat flow from the reasoning and decomposition stages. In the Question-Answering section, we obtain information from the Minecraft-Wiki and Reddit forums and use GPT-3.5 to clean the data into single-round QA pairs. In the Skill-Code section, we use GPT-3.5 combined with the human player's code to synthesize code snippets and then check and revise them in the game environment. + +Clean + +Construct + +Check + +Single-round QA pairs + +1. World & Entities +2. Player Mechanics & Survival +3. Knowledge & Discovery +4. Resources & Crafting +5. Tools & Utilities +6. Miscellaneous + +train STEVE-13B. The GPT-3.5 [5] is employed to derive meaningful single-round question-answer pairs, and LoRA [22] is incorporated during the fin-tuning process for efficient resource allocation. + +Skill-Code pairs contain 210 skill execution scripts with descriptions, covering 8 skill types including collecting, crafting, exploration etc. The code part is collected by manual coding. We use GPT-3.5 [5] to describe all codes and utilize langchain vecordb to give all pairs a database vector. + +# 5 Experiments + +# 5.1 Experimental Setup + +We train STEVE-7B/13B, which is finetuned from LLaMA-2 [18] with the Question-Answering pair in STEVE-21K dataset for warm-up and simulation context data from successful runs. We use LoRA [22] for finetuning process. Note that the process is to adjust STEVE-13B to work on correct simulation knowledge while remaining adapted to visual perception. In the text part, we set all temperatures to 0 except for the task proposal, which uses 0.9 to encourage task diversity. The vision unit is based on EfficientFormerV2-S0 [31], which is trained on the Vision-Environment part of our STEVE-21K dataset. Our simulation environment is built on top of MineDojo [17] and leverages Mineflayer [45]. + +![](images/5d30bf45360efd026ed4a76d83aeb386e7564905498c795635ad07fe0ea13126.jpg) +Fig. 4: Example of Vision-Environment pairs. It represents the data format of the Vision-Environment pairs in our STEVE-21K dataset: including visual signals, environmental information, Chat Log, Context QA pairs, and planning in actual tasks. + +We use GPT-4-0613 for all GPT-4 models used in Voyager [57] and code generation tasks. + +# 5.2Baselines + +As no vision-based LLM-driven agents are immediately operable in Minecraft, we selected several algorithms as baselines that extract information from a system's backend, differing significantly from real-world applications. + +AutoGPT [49] is an NLP automation tool that decomposes a high-level goal into executable subgoals in MineDojo, aligning with our experimental framework. Our setup, AutoGPT, powered by GPT-4 [44], processes agent states, environment feedback, and execution errors to manage subgoal execution. + +Voyager [57] relies solely on textual grounding for perception and features a long-term procedural memory with a hierarchical library of code-based procedures, allowing the integration of simple skills into complex behaviors. Proficient in environment exploration and tech tree mastery, Voyager uses GPT-4 [44] for processing background text in embodied agents, with a lesser emphasis on visual perception. + +Table 2: Comparison on tech tree mastery task. The values presented are in fractions, representing successful trials out of three attempts. A score of $0/3$ signifies the method's inability to progress within the tech tree after a maximum of 160 prompting iterations. The reported numbers denote the average iterations across three trials. Lower iteration values indicate higher efficiency of the respective method. + +
MethodWooden ToolStone ToolIron ToolDiamond Tool
AutoGPT [49]92 ± 72 (3/3)94 ± 72 (3/3)135 ± 103 (3/3)N/A (0/3)
Voyager [57]6 ± 2 (3/3)11 ± 2 (3/3)21 ± 7 (3/3)102 (1/3)
STEVE4 ± 1 (3/3)8 ± 1 (3/3)15 ± 2 (3/3)106 ± 12 (3/3)
+ +Table 3: Quantitative comparison on knowledge question and answering task. Questions, model-generated responses, and ground truth inputs are evaluated in GPT-4 [5], Claude-2 [8] and human blind rating rated on a scale of 0 to 10; The scores above are the average of them. Higher scores indicate greater alignment of the generated answers with the ground truth. Wld., Ent., Mech., Surv., Know., Disc., Res., Craft., Tools, Util., Miscell. stand for World, Entities, Player Mechanics, Survival, Knowledge, Discovery, Resources, Crafting, Tools, Utilities and Miscellaneous. + +
MethodWld. & Ent.Mech. & Surv.Know. & Disc.Res. & Craft.Tl. & Util.Miscell.Overall
Llama2-7B6.446.686.586.426.806.966.56
Llama2-13B6.936.956.776.776.986.646.89
STEVE-7B7.997.887.847.957.937.827.94
STEVE-13B8.148.138.038.158.127.728.12
GPT-48.068.078.077.928.098.218.04
+ +# 5.3 Evaluation Results + +Continuous block search. As shown in Tab. 1, we experiment with block-searching tasks to assess the agent's exploratory capabilities and proficiency in locating specified blocks. Diamond blocks are placed at every 16-block interval across the land map. The agent's objective is to identify as many blocks as possible within the fewest iterations, which indicates the method's efficiency. + +As shown in Fig. 5, enriching information through visual perception significantly enhances the efficiency of search and exploration tasks, leading to more effective world exploration. + +Table 1: Comparison on continues block search task. # Iters stand for average iterations to find 10 diamonds (max 100). # Blocks stand for average diamonds found in 100 iterations. + +
Method# Iters (↓) # Blocks (↑)
AutoGPT [49]N/A 7
Voyager [57]35 26
STEVE14 67
+ +![](images/c3ffbd171cdc51eefbe4bfce50cd9a15ff960f3abcf3684c6d35aa8c13925cc9.jpg) +Fig.5: Schematic of the continuous block search task. We capture an asynchronous segment with each method 30 iterations from the experiments for illustration. The reason we choose diamond blocks is that they are not naturally occurring in the given context, making them easily distinguishable from other blocks. + +Knowledge question and answering. Using a validation dataset, we established a question-answering database to evaluate our model's performance on Minecraft-related queries. Responses from each model are rated blindly by GPT-4, Claude2, and human participants based on accuracy, relevance, and detail. Responses are initially checked for accuracy, and the comprehensive evaluation yields a score ranging from 0 to 10, with higher scores indicating superior performance. + +As detailed in Tab. 3, we tested the instructional capabilities of various LLM models using the STEVE-21k dataset, divided into 18,622 training and 1,000 testing samples. The STEVE-7B and STEVE-13B models surpassed LLaMA2 in all metrics, with STEVE-13B scoring the highest at 8.12, demonstrating its exceptional understanding of Minecraft-related queries. This indicates that STEVE models, optimized specifically for Minecraft content, perform better in knowledge-intensive tasks. + +Our results confirm that larger, domain-specifically tuned models like STEVE-13B outperform broader models like GPT-4, underscoring the benefit of specialized fine-tuning for domain-specific applications. + +![](images/e91ccdc799cd7cd93fb52b9144b461282a4db2a37877cbcce4f3d97d2dfbe61a.jpg) + +![](images/f95ac2622f513174bd27eba737e9765cde210570ec63862fa79bdf9bab14841d.jpg) + +![](images/fc15d8b43a6c6a219405b1d97201503701202651a2f7d734e028fe522ba6484d.jpg) + +![](images/826c2d76a5225d8c324005a52ac9cc6da89c423df693190dee59d07b3648129f.jpg) + +![](images/e8070962c9f4a31a7677295ef27366233d4975df507973f088419d805fa02cbb.jpg) + +![](images/3e28f748d3988cb1973af8c1df6eec5e33e4da8bd1276a7727972f2a9c4f86f2.jpg) + +![](images/2324a70c27db609ef3d1411c5025133c864e153dfe60253efcb06a9a55ab6a70.jpg) + +![](images/087df738ce9447c33074cd6826cefb2d7fc946fda7f4b7ca8604f99ae1bf4fbb.jpg) + +![](images/ae46503e694aa75d149810b04b5403c9befffb96e2a4e08c74e63134dee5111c.jpg) + +![](images/f74fe55769e23c2d01404e6b275ba8a4ca5a81d9da5594903fdbce86fce42076.jpg) + +![](images/617f7a9b756b06ff25e43a44bf5948fce18fa933315ddf8f864628d7b0fd925b.jpg) + +![](images/e609be31ebf1b0024d4118fdfa22b57c753d09ccbc627e5d3d163aa183cfb842.jpg) + +![](images/c5b040852d0693814ed8f16ca4a63c4e0d6452a82e506ae37a75449c814516e1.jpg) + +![](images/05047e68461b01b191892fc16bd6bc59e36da72c351c8660b722cb198b3583e1.jpg) + +![](images/96b586df6b7667fa9e0203dd7313c6411d9b45c5380c63ca50ecf4cb29f4e256.jpg) + +![](images/630b50962ab2d34f205431369fa48fea883ec201012cebf11e9bea5d2d0acbcd.jpg) + +![](images/508641d0d956ba9a170be8b6d37737357a9c553eeabafeeb8f70f6333d062b04.jpg) +Fig. 6: Qualitative comparison on knowledge question and answering tasks. Green (Red) highlights the correct or good (wrong or bad) answer. Blue indicates the suboptimal answer. Grey indicates the meaningless answer. + +![](images/c30134ec6506fd05cc9e003840ad135900d7545f3bf00ee52d3a61549895794d.jpg) + +![](images/e55b903770769d2184e50c956562fadd07a462f2d95589666552e3622843d302.jpg) + +![](images/daed7dce2e57dd123eb53e9fbaf4f7032951c523ddb4f63c08a49b6a26a41f3f.jpg) + +Tech tree mastery. As shown in Tab. 2, we experiment on the Minecraft tech tree mastery to test the agent's ability to craft and use a hierarchy of tools. Progressing through this tree (wooden tool $\rightarrow$ stone tool $\rightarrow$ iron tool $\rightarrow$ diamond tool) requires the agent to master systematic and compositional skills. As to the wooden, stone, and iron levels of the tech tree, STEVE achieves remarkable efficiency: $23 \times$ , $11.8 \times$ , and $9 \times$ faster than AutoGPT [49], and $1.5 \times$ , $1.4 \times$ , and $1.3 \times$ faster than Voyager [57]. STEVE has achieved the diamond level, as shown in Tab. 2. Its performance slightly lags behind Voyager [57], which also uses GPT4 [44] for critical inference. However, STEVE is more cost-effective, starting with lower initial performance. It includes a vision unit, prioritizing visual data over background information, offering distinct advantages. Additionally, we observed a decrease in performance when using a basic skill database. + +# 5.4 Ablation Study + +To understand the impact of different components on the performance of our system, we conducted ablation studies focusing on the tech tree mastery task in Minecraft. The results, as shown in Tab. 4, provide insights into the effectiveness of the vision unit and compare our STEVE model with the STEVE GPT-4 version (with the same vision unit as ours). Note that the w/o vision unit setup is that the environmental perception encompasses data on blocks within an 8x8 area surrounding the agent, including the front, back, left, and right directions. The following observations are made: + +Table 4: Ablation studies for the tech tree mastery. STEVE (Ours) is the STEVE-13B version. The $0/3$ score means the method can't progress beyond 160 iterations in the tech tree. + +
MethodWooden ToolStone ToolIron ToolDiamond Tool
w/o vision unit11 ± 5 (3/3)27 ± 5 (3/3)46 ± 11 (3/3)158 (1/3)
STEVE (GPT-4)6 ± 2 (3/3)10 ± 1 (3/3)14 ± 3 (3/3)89 ± 9 (3/3)
STEVE (Ours)4 ± 1 (3/3)8 ± 1 (3/3)15 ± 2 (3/3)106 ± 12 (3/3)
+ +Vision unit is critical. The omission of the vision unit markedly affects the system's performance, especially in more advanced tasks. While it successfully crafts Wooden, Stone, and Iron Tools, it is challenged with Diamond Tools. This outcome underscores the vital importance of visual information in accomplishing complex tasks. + +Comparison with GPT-4. As our vision encoder directly encodes into text space, it can be easily replaced with any language model. For instance, the GPT-4 we compared exhibits consistent success across all categories and secures a flawless success rate. Interestingly, the STEVE-13B version excels in simpler tasks such as crafting wooden and stone tools. Moreover, it requires fewer iterations than methods without the vision part, underscoring its superior efficiency. + +# 5.5 Case Study + +As shown in Fig. 6, we perform an extensive case study comparison of GPT-4, LLaMA2-13B, and our method STEVE-13B. Each model maintains the same information and question inputs to compare feedback under different environmental information. Our STEVE overall achieves the best results, surpassing GPT-4 and showing significant improvement compared to the original LLaMA. Especially in parts involving numerical calculations, such as the leftmost image, STEVE accurately tracks food values to restore hunger levels. + +# 6 Conclusion + +STEVE enhances multi-modal learning by combining visual encoder and LLM-based agents. It has three functions: vision perception, language instruction, and code action, allowing it to understand, predict, and act in virtual environments. We provide a straightforward approach to creating a robust, multimodal, autonomous, embodied agent using an open-source language model with a small number of parameters. Additionally, we provide a comprehensive dataset STEVE-21K for sustainable community development that can be verified. + +# Acknowledgement + +This work is supported by the Zhejiang Provincial Natural Science Foundation of China (No. LZ24F030005) and the National Natural Science Foundation of China (No. 62106219). + +# References + +1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems 35, 23716-23736 (2022) +2. Amiranashvili, A., Dorka, N., Burgard, W., Koltun, V., Brox, T.: Scaling imitation learning in apacheft. arXiv preprint arXiv:2007.02701 (2020) +3. Baker, B., Akkaya, I., Zhokov, P., Huizinga, J., Tang, J., Ecoffet, A., Houghton, B., Sampedro, R., Clune, J.: Video pretraining (vpt): Learning to act by watching unlabeled online videos. Advances in Neural Information Processing Systems 35, 24639-24654 (2022) +4. Chai, W., Wang, G.: Deep vision multimodal learning: Methodology, benchmark, and trend. Applied Sciences 12(13), 6588 (2022) +5. Introducing chatgpt (2022), https://openai.com/blog/chatgpt +6. Chen, S., Guhur, P.L., Schmid, C., Laptev, I.: History aware multimodal transformer for vision-and-language navigation. Advances in neural information processing systems 34, 5834-5847 (2021) +7. Chen, T., Gupta, S., Gupta, A.: Learning exploration policies for navigation. In: International Conference on Learning Representations (2018) +8. Talk to claude (2023), https://claude.ai +9. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500 (2023) +0. Das, A., Datta, S., Gkioxari, G., Lee, S., Parikh, D., Batra, D.: Embodied question answering. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1-10 (2018) +1. Datta, S., Dharur, S., Cartillier, V., Desai, R., Khanna, M., Batra, D., Parikh, D.: Episodic memory question answering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19119-19128 (2022) +2. Dean, V., Tulsiani, S., Gupta, A.: See, hear, explore: Curiosity via audio-visual association. Advances in neural information processing systems 33, 14961-14972 (2020) +3. Deng, J., Chai, W., Guo, J., Huang, Q., Hu, W., Hwang, J.N., Wang, G.: Citygen: Infinite and controllable 3d city layout generation. arXiv preprint arXiv:2312.01508 (2023) +4. Deng, J., Chai, W., Huang, J., Zhao, Z., Huang, Q., Gao, M., Guo, J., Hao, S., Hu, W., Hwang, J.N., et al.: Citycraft: A real craftsfer for 3d city generation. arXiv preprint arXiv:2406.04983 (2024) +5. Driess, D., Xia, F., Sajjadi, M.S., Lynch, C., Chowdhery, A., Ichter, B., Wahid, A., Tompson, J., Vuong, Q., Yu, T., et al.: Palm-e: An embodied multimodal language model. arXiv preprint arXiv:2303.03378 (2023) + +16. Du, H., Yu, X., Zheng, L.: Vtnet: Visual transformer network for object goal navigation. In: International Conference on Learning Representations (2020) +17. Fan, L., Wang, G., Jiang, Y., Mandlekar, A., Yang, Y., Zhu, H., Tang, A., Huang, D.A., Zhu, Y., Anandkumar, A.: Minedojo: Building open-ended embodied agents with internet-scale knowledge. Advances in Neural Information Processing Systems 35, 18343-18362 (2022) +18. Gao, P., Han, J., Zhang, R., Lin, Z., Geng, S., Zhou, A., Zhang, W., Lu, P., He, C., Yue, X., et al.: Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010 (2023) +19. Gong, T., Lyu, C., Zhang, S., Wang, Y., Zheng, M., Zhao, Q., Liu, K., Zhang, W., Luo, P., Chen, K.: Multimodal-gpt: A vision and language model for dialogue with humans. arXiv preprint arXiv:2305.04790 (2023) +20. Guss, W.H., Houghton, B., Topin, N., Wang, P., Codel, C., Veloso, M., Salakhutdinov, R.: Minerl: A large-scale dataset of apache demonstrations. arXiv preprint arXiv:1907.13440 (2019) +21. Hofmann, K.: Minecraft as ai playground and laboratory. In: Proceedings of the annual symposium on computer-human interaction in play. pp. 1-1 (2019) +22. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021) +23. Huang, W., Xia, F., Xiao, T., Chan, H., Liang, J., Florence, P., Zeng, A., Tompson, J., Mordatch, I., Chebotar, Y., et al.: Inner monologue: Embodied reasoning through planning with language models. arXiv preprint arXiv:2207.05608 (2022) +24. Jiang, J., Ye, T., Bai, J., Chen, S., Chai, W., Jun, S., Liu, Y., Chen, E.: Five a+ network: You only need 9k parameters for underwater image enhancement. arXiv preprint arXiv:2305.08824 (2023) +25. Johnson, M., Hofmann, K., Hutton, T., Bignell, D.: The malmo platform for artificial intelligence experimentation. In: Ijcai. pp. 4246-4247 (2016) +26. Kolve, E., Mottaghi, R., Han, W., VanderBilt, E., Weihs, L., Herrasti, A., Deitke, M., Ehsani, K., Gordon, D., Zhu, Y., et al.: Ai2-thor: An interactive 3d environment for visual ai. arXiv preprint arXiv:1712.05474 (2017) +27. Kwon, O., Park, J., Oh, S.: Rendering neural radiance map for visual navigation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9099-9108 (2023) +28. Li, B., Zhang, Y., Chen, L., Wang, J., Yang, J., Liu, Z.: Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726 (2023) +29. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023) +30. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022) +31. Li, Y., Yuan, G., Wen, Y., Hu, J., Evangelidis, G., Tulyakov, S., Wang, Y., Ren, J.: Efficientformer: Vision transformers at mobilenet speed. Advances in Neural Information Processing Systems 35, 12934-12949 (2022) +32. Lifshitz, S., Paster, K., Chan, H., Ba, J., McIlraith, S.: Steve-1: A generative model for text-to-behavior in apache (abridged version). In: NeurIPS 2023 Workshop on Goal-Conditioned Reinforcement Learning (2023) +33. Lin, Z., Li, J., Shi, J., Ye, D., Fu, Q., Yang, W.: Juewu-mc: Playing minecraft with sample-efficient hierarchical reinforcement learning. arXiv preprint arXiv:2112.04907 (2021) + +34. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv preprint arXiv:2304.08485 (2023) +35. Liu, S., Okatani, T.: Symmetry-aware neural architecture for embodied visual exploration. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 17221-17230. IEEE (2022) +36. Lu, P., Peng, B., Cheng, H., Galley, M., Chang, K.W., Wu, Y.N., Zhu, S.C., Gao, J.: Chameleon: Plug-and-play compositional reasoning with large language models. arXiv preprint arXiv:2304.09842 (2023) +37. Luo, W., Sun, P., Zhong, F., Liu, W., Zhang, T., Wang, Y.: End-to-end active object tracking via reinforcement learning. In: International conference on machine learning. pp. 3286-3295. PMLR (2018) +38. Luo, W., Sun, P., Zhong, F., Liu, W., Zhang, T., Wang, Y.: End-to-end active object tracking and its real-world deployment via reinforcement learning. IEEE transactions on pattern analysis and machine intelligence 42(6), 1317-1332 (2019) +39. Lyu, C., Wu, M., Wang, L., Huang, X., Liu, B., Du, Z., Shi, S., Tu, Z.: Macaw-llm: Multi-modal language modeling with image, audio, video, and text integration. arXiv preprint arXiv:2306.09093 (2023) +40. Ma, W., Mi, Q., Yan, X., Wu, Y., Lin, R., Zhang, H., Wang, J.: Large language models play starcraft ii: Benchmarks and a chain of summarization approach. arXiv preprint arXiv:2312.11865 (2023) +41. Maaz, M., Rasheed, H., Khan, S., Khan, F.S.: Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424 (2023) +42. Mao, H., Wang, C., Hao, X., Mao, Y., Lu, Y., Wu, C., Hao, J., Li, D., Tang, P.: Seihai: A sample-efficient hierarchical ai for the minerl competition. In: Distributed Artificial Intelligence: Third International Conference, DAI 2021, Shanghai, China, December 17-18, 2021, Proceedings 3. pp. 38-51. Springer (2022) +43. Moudgil, A., Majumdar, A., Agrawal, H., Lee, S., Batra, D.: Soat: A scene-and object-aware transformer for vision-and-language navigation. Advances in Neural Information Processing Systems 34, 7357-7367 (2021) +44. OpenAI: Gpt-4 technical report. arXiv preprint arXiv: Arxiv-2303.08774 (2023) +45. PrismarineJS: Prismarinejs/mineflayer: Create apache bots with a powerful, stable, and high level javascript api. (2013), https://github.com/PrismarineJS/mineflayer/tree/master +46. Savva, M., Kadian, A., Maksymets, O., Zhao, Y., Wijmans, E., Jain, B., Straub, J., Liu, J., Koltun, V., Malik, J., et al.: Habitat: A platform for embodied ai research. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9339-9347 (2019) +47. Schick, T., Dwivedi-Yu, J., Dessi, R., Raileanu, R., Lomeli, M., Zettlemoyer, L., Cancedda, N., Scialom, T.: Toolformer: Language models can teach themselves to use tools. arXiv preprint arXiv:2302.04761 (2023) +48. Shen, Y., Song, K., Tan, X., Li, D., Lu, W., Zhuang, Y.: Hugginggpt: Solving ai tasks with chatgpt and its friends in huggingface. arXiv preprint arXiv:2303.17580 (2023) +49. Significant-Gravitas: Auto-gpt. https://github.com/Significant-Gravitas/Auto-GPT (2023) +50. Skrynnik, A., Staroverov, A., Aitygulov, E., Aksenov, K., Davydov, V., Panov, A.I.: Hierarchical deep q-network from imperfect demonstrations in minecraft. Cognitive Systems Research 65, 74-78 (2021) + +51. Song, C.H., Wu, J., Washington, C., Sadler, B.M., Chao, W.L., Su, Y.: Llm-planner: Few-shot grounded planning for embodied agents with large language models. arXiv preprint arXiv:2212.04088 (2022) +52. Song, E., Chai, W., Wang, G., Zhang, Y., Zhou, H., Wu, F., Guo, X., Ye, T., Lu, Y., Hwang, J.N., et al.: Moviechat: From dense token to sparse memory for long video understanding. arXiv preprint arXiv:2307.16449 (2023) +53. Song, E., Chai, W., Ye, T., Hwang, J.N., Li, X., Wang, G.: Moviechat+: Question-aware sparse memory for long video question answering. arXiv preprint arXiv:2404.17176 (2024) +54. Su, Y., Lan, T., Li, H., Xu, J., Wang, Y., Cai, D.: Pandagpt: One model to instruction-follow them all. arXiv preprint arXiv:2305.16355 (2023) +55. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023) +56. Vasluianu, F.A., Seizinger, T., Zhou, Z., Wu, Z., Chen, C., Timofte, R., Dong, W., Zhou, H., Tian, Y., Chen, J., et al.: Ntire 2024 image shadow removal challenge report. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6547-6570 (2024) +57. Wang, G., Xie, Y., Jiang, Y., Mandlekar, A., Xiao, C., Zhu, Y., Fan, L., Anandkumar, A.: Voyager: An open-ended embodied agent with large language models. arXiv preprint arXiv:2305.16291 (2023) +58. Wang, W., Chen, Z., Chen, X., Wu, J., Zhu, X., Zeng, G., Luo, P., Lu, T., Zhou, J., Qiao, Y., et al.: Visionllm: Large language model is also an open-ended decoder for vision-centric tasks. arXiv preprint arXiv:2305.11175 (2023) +59. Wang, Z., Cai, S., Liu, A., Ma, X., Liang, Y.: Describe, explain, plan and select: Interactive planning with large language models enables open-world multi-task agents. arXiv preprint arXiv:2302.01560 (2023) +60. Whitted, T.: An improved illumination model for shaded display. In: ACM SIGGRAPH 2005 Courses, pp. 4-es (2005) +61. Wijmans, E., Kadian, A., Morcos, A., Lee, S., Essa, I., Parikh, D., Savva, M., Batra, D.: Dd-ppo: Learning near-perfect pointgoal navigators from 2.5 billion frames. In: International Conference on Learning Representations (2019) +62. Wu, C., Yin, S., Qi, W., Wang, X., Tang, Z., Duan, N.: Visual chatgpt: Talking, drawing and editing with visual foundation models. arXiv preprint arXiv:2303.04671 (2023) +63. Xie, Y., Yu, C., Zhu, T., Bai, J., Gong, Z., Soh, H.: Translating natural language to planning goals with large-language models. arXiv preprint arXiv:2302.05128 (2023) +64. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178 (2023) +65. Ye, T., Chen, S., Liu, Y., Chai, W., Bai, J., Zou, W., Zhang, Y., Jiang, M., Chen, E., Xue, C.: Sequential affinity learning for video restoration. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 4147-4156 (2023) +66. Ye, T., Zhang, Y., Jiang, M., Chen, L., Liu, Y., Chen, S., Chen, E.: Perceiving and modeling density for image dehazing. In: European conference on computer vision. pp. 130-145. Springer (2022) +67. Yu, L., Chen, X., Gkioxari, G., Bansal, M., Berg, T.L., Batra, D.: Multi-target embodied question answering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6309-6318 (2019) +68. Yu, Y., Huang, W., Sun, F., Chen, C., Wang, Y., Liu, X.: Sound adversarial audiovisual navigation. In: International Conference on Learning Representations (2021) + +69. Yuan, H., Zhang, C., Wang, H., Xie, F., Cai, P., Dong, H., Lu, Z.: Plan4mc: Skill reinforcement learning and planning for open-world apache tasks. arXiv preprint arXiv:2303.16563 (2023) +70. Zhao, Z., Chai, W., Hao, S., Hu, W., Wang, G., Cao, S., Song, M., Hwang, J.N., Wang, G.: A survey of deep learning in sports applications: Perception, comprehension, and decision. arXiv preprint arXiv:2307.03353 (2023) +71. Zhao, Z., Chai, W., Wang, X., Ma, K., Chen, K., Guo, D., Ye, T., Zhang, Y., Wang, H., Wang, G.: Steve series: Step-by-step construction of agent systems in northwest China. arXiv preprint arXiv:2406.11247 (2024) +72. Zhao, Z., Chen, K., Guo, D., Chai, W., Ye, T., Zhang, Y., Wang, G.: Hierarchical auto-organizing system for open-ended multi-agent navigation. arXiv preprint arXiv:2403.08282 (2024) +73. Zhao, Z., Ma, K., Chai, W., Wang, X., Chen, K., Guo, D., Zhang, Y., Wang, H., Wang, G.: Do we really need a complex agent system? distill embodied agent into a single model. arXiv preprint arXiv:2404.04619 (2024) +74. Zhong, F., Sun, P., Luo, W., Yan, T., Wang, Y.: Ad-vat+: An asymmetric dueling mechanism for learning and understanding visual active tracking. IEEE transactions on pattern analysis and machine intelligence 43(5), 1467-1482 (2019) +75. Zhong, F., Sun, P., Luo, W., Yan, T., Wang, Y.: Towards distraction-robust active visual tracking. In: International Conference on Machine Learning. pp. 12782-12792. PMLR (2021) +76. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023) +77. Zhu, X., Chen, Y., Tian, H., Tao, C., Su, W., Yang, C., Huang, G., Li, B., Lu, L., Wang, X., et al.: Ghost in the minecraft: Generally capable agents for open-world enviroments via large language models with text-based knowledge and memory. arXiv preprint arXiv:2305.17144 (2023) \ No newline at end of file diff --git a/2024/See and Think_ Embodied Agent in Virtual Environment/images.zip b/2024/See and Think_ Embodied Agent in Virtual Environment/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d5906a26335e7241eec6e7d6a3da632367f206e1 --- /dev/null +++ b/2024/See and Think_ Embodied Agent in Virtual Environment/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5b02ceb873ea65c12f6b0b8dbe5754415b23deb67bc3fb26a1d17e771013bdf9 +size 497107 diff --git a/2024/See and Think_ Embodied Agent in Virtual Environment/layout.json b/2024/See and Think_ Embodied Agent in Virtual Environment/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..bc52743e0670830d34a0a4e35c34afb31d12d1c5 --- /dev/null +++ b/2024/See and Think_ Embodied Agent in Virtual Environment/layout.json @@ -0,0 +1,12106 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 156, + 111, + 459, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 111, + 459, + 146 + ], + "spans": [ + { + "bbox": [ + 156, + 111, + 459, + 146 + ], + "type": "text", + "content": "See and Think: Embodied Agent in Virtual Environment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 167, + 466, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 167, + 466, + 194 + ], + "spans": [ + { + "bbox": [ + 146, + 167, + 466, + 194 + ], + "type": "text", + "content": "Zhonghan Zhao" + }, + { + "bbox": [ + 146, + 167, + 466, + 194 + ], + "type": "inline_equation", + "content": "^{1, \\spadesuit}" + }, + { + "bbox": [ + 146, + 167, + 466, + 194 + ], + "type": "text", + "content": ", Wenhao Chai" + }, + { + "bbox": [ + 146, + 167, + 466, + 194 + ], + "type": "inline_equation", + "content": "^{2, \\spadesuit}" + }, + { + "bbox": [ + 146, + 167, + 466, + 194 + ], + "type": "text", + "content": ", Xuan Wang" + }, + { + "bbox": [ + 146, + 167, + 466, + 194 + ], + "type": "inline_equation", + "content": "^{1, \\spadesuit}" + }, + { + "bbox": [ + 146, + 167, + 466, + 194 + ], + "type": "text", + "content": ", Boyi Li" + }, + { + "bbox": [ + 146, + 167, + 466, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 146, + 167, + 466, + 194 + ], + "type": "text", + "content": ", Shengyu Hao" + }, + { + "bbox": [ + 146, + 167, + 466, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 146, + 167, + 466, + 194 + ], + "type": "text", + "content": ", Shidong Cao" + }, + { + "bbox": [ + 146, + 167, + 466, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 146, + 167, + 466, + 194 + ], + "type": "text", + "content": ", Tian Ye" + }, + { + "bbox": [ + 146, + 167, + 466, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 146, + 167, + 466, + 194 + ], + "type": "text", + "content": ", and Gaoang Wang" + }, + { + "bbox": [ + 146, + 167, + 466, + 194 + ], + "type": "inline_equation", + "content": "^{1, \\dagger}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 261, + 201, + 352, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 201, + 352, + 213 + ], + "spans": [ + { + "bbox": [ + 261, + 201, + 352, + 213 + ], + "type": "text", + "content": "1 Zhejiang University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 250, + 213, + 363, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 213, + 363, + 224 + ], + "spans": [ + { + "bbox": [ + 250, + 213, + 363, + 224 + ], + "type": "text", + "content": "2 University of Washington" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 190, + 224, + 423, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 224, + 423, + 236 + ], + "spans": [ + { + "bbox": [ + 190, + 224, + 423, + 236 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 190, + 224, + 423, + 236 + ], + "type": "text", + "content": " Hong Kong University of Science and Technology (GZ)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 262, + 455, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 262, + 455, + 460 + ], + "spans": [ + { + "bbox": [ + 159, + 262, + 455, + 460 + ], + "type": "text", + "content": "Abstract. Large language models (LLMs) have achieved impressive progress on several open-world tasks. Recently, using LLMs to build embodied agents has been a hotspot. This paper proposes STEVE, a comprehensive and visionary embodied agent in the Minecraft virtual environment. STEVE comprises three key components: vision perception, language instruction, and code action. Vision perception involves interpreting visual information in the environment, which is then integrated into the LLMs component with agent state and task instruction. Language instruction is responsible for iterative reasoning and decomposing complex tasks into manageable guidelines. Code action generates executable skill actions based on retrieval in skill database, enabling the agent to interact effectively within the Minecraft environment. We also collect STEVE-21K dataset, which includes " + }, + { + "bbox": [ + 159, + 262, + 455, + 460 + ], + "type": "inline_equation", + "content": "600+" + }, + { + "bbox": [ + 159, + 262, + 455, + 460 + ], + "type": "text", + "content": " vision-environment pairs, 20K knowledge question-answering pairs, and " + }, + { + "bbox": [ + 159, + 262, + 455, + 460 + ], + "type": "inline_equation", + "content": "200+" + }, + { + "bbox": [ + 159, + 262, + 455, + 460 + ], + "type": "text", + "content": " skill-code pairs. We conduct continuous block search, knowledge question and answering, and tech tree mastery to evaluate the performance. Extensive experiments show that STEVE achieves at most " + }, + { + "bbox": [ + 159, + 262, + 455, + 460 + ], + "type": "inline_equation", + "content": "1.5 \\times" + }, + { + "bbox": [ + 159, + 262, + 455, + 460 + ], + "type": "text", + "content": " faster unlocking key tech trees and " + }, + { + "bbox": [ + 159, + 262, + 455, + 460 + ], + "type": "inline_equation", + "content": "2.5 \\times" + }, + { + "bbox": [ + 159, + 262, + 455, + 460 + ], + "type": "text", + "content": " quicker in block search tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 470, + 453, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 470, + 453, + 492 + ], + "spans": [ + { + "bbox": [ + 159, + 470, + 453, + 492 + ], + "type": "text", + "content": "Keywords: Open-world embodied agent " + }, + { + "bbox": [ + 159, + 470, + 453, + 492 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 470, + 453, + 492 + ], + "type": "text", + "content": " Multimodal pre-training " + }, + { + "bbox": [ + 159, + 470, + 453, + 492 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 470, + 453, + 492 + ], + "type": "text", + "content": " Large language model" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 540, + 230, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 540, + 230, + 553 + ], + "spans": [ + { + "bbox": [ + 132, + 540, + 230, + 553 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 564, + 482, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 564, + 482, + 625 + ], + "spans": [ + { + "bbox": [ + 130, + 564, + 482, + 625 + ], + "type": "text", + "content": "Designing agents that demonstrate intelligent behavior and adaptability in open-world settings has been a longstanding and significant challenge in the field of artificial intelligence [13,14,26,46,70]. However, recent progress in the development of large language models (LLMs) [5,55] has exhibited their potential as versatile, general-purpose assistants. Recent innovations in agent design [57,59,69,77]" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 632, + 481, + 654 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 632, + 481, + 654 + ], + "spans": [ + { + "bbox": [ + 141, + 632, + 481, + 654 + ], + "type": "inline_equation", + "content": "\\spadesuit" + }, + { + "bbox": [ + 141, + 632, + 481, + 654 + ], + "type": "text", + "content": " Equal contribution: zhaozhonghan@zju.edu.cn, wchai@uw.edu, xu-anw@zju.edu.cn." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 153, + 654, + 370, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 654, + 370, + 665 + ], + "spans": [ + { + "bbox": [ + 153, + 654, + 370, + 665 + ], + "type": "text", + "content": "† Corresponding author: gaoangwang@intl.zju.edu.cn." + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 142, + 114, + 474, + 206 + ], + "blocks": [ + { + "bbox": [ + 142, + 114, + 474, + 206 + ], + "lines": [ + { + "bbox": [ + 142, + 114, + 474, + 206 + ], + "spans": [ + { + "bbox": [ + 142, + 114, + 474, + 206 + ], + "type": "image", + "image_path": "8e20506f446eeac2546e04d44592fb05bd49ec21a9a6b5405d67407105162973.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 214, + 482, + 258 + ], + "lines": [ + { + "bbox": [ + 130, + 214, + 482, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 214, + 482, + 258 + ], + "type": "text", + "content": "Fig. 1: STEVE integrates the three parts: Vision Perception, Language Instruction, and Code Action, supported closely by our proposed STEVE-21K. It demonstrates commendable performance on Continuous Block Search, Knowledge QA, and Tech Tree Mastery." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 289, + 482, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 348 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 348 + ], + "type": "text", + "content": "have effectively harnessed these advanced LLMs, tapping into their extensive world knowledge and reasoning abilities. This development has paved the way for agents, that are autonomously driven, to formulate and implement strategies and actions across a diverse array of skills and tasks in diverse open-world environments." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 351, + 482, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 351, + 482, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 351, + 482, + 518 + ], + "type": "text", + "content": "In many open-world settings, like Minecraft, contemporary agents predominantly use LLMs for their textual interactions. However, this reliance on text for communication poses considerable limitations in their interactions within these worlds including low-level regrading cases [24, 56, 65, 66]. Minecraft, with its expansive and interactive sandbox environment [17, 20], demands a variety of skills from agents, ranging from crafting basic items to executing complex tasks. Yet, agents driven by LLMs often generate unpredictable outputs. The effectiveness of their interactions is largely contingent on meticulously crafted prompts [23], designed to align the LLM's understanding with the environmental context and the intended objectives. This process of prompt engineering is not only laborious but also fails to meet the goal of fostering autonomous, self-directed agents. Furthermore, textual communication has its limitations in naturally conveying certain concepts of the world, like crafting recipes, which are often more effectively communicated through vision." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "text", + "content": "Players have the distinct capability to assimilate and convey information using both visual and textual channels, significantly enhancing our interactions with the world around us. Yet, the integration of LLM-based agents with multimodal inputs in open-ended environments remains an under-explored area. STEVE in STEVE-Series [71-73], is named after the protagonist of the game \"Minecraft.\" It is our proposed framework to build an embodied agent based on the vision model and LLMs within an open world, as illustrated in Fig. 1. STEVE harnesses a vision model to perceive its surroundings visually, coupled with an LLM to strategize and plan actions. This model represents a leap forward in agent design, combining these two input modes, vision, and text, to offer a more nuanced and comprehensive understanding of the environment and practical and executable skills." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "text", + "content": "Z. Zhao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 146, + 116, + 349, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 116, + 349, + 128 + ], + "spans": [ + { + "bbox": [ + 146, + 116, + 349, + 128 + ], + "type": "text", + "content": "Our key contributions are outlined as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 139, + 130, + 481, + 250 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 139, + 130, + 481, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 130, + 481, + 177 + ], + "spans": [ + { + "bbox": [ + 139, + 130, + 481, + 177 + ], + "type": "text", + "content": "- We propose STEVE, an embodied agent in virtual environment, consists of vision perception, language instruction, and code action, achieving " + }, + { + "bbox": [ + 139, + 130, + 481, + 177 + ], + "type": "inline_equation", + "content": "1.5 \\times" + }, + { + "bbox": [ + 139, + 130, + 481, + 177 + ], + "type": "text", + "content": " faster unlocking of key tech trees and is " + }, + { + "bbox": [ + 139, + 130, + 481, + 177 + ], + "type": "inline_equation", + "content": "2.3 \\times" + }, + { + "bbox": [ + 139, + 130, + 481, + 177 + ], + "type": "text", + "content": " quicker in block search tasks compared to previous state-of-the-art methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 139, + 178, + 479, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 178, + 479, + 213 + ], + "spans": [ + { + "bbox": [ + 139, + 178, + 479, + 213 + ], + "type": "text", + "content": "- We present STEVE-7B/13B, a series of large language model obtained by fine-tuning with Minecraft knowledge question-answering pairs from Llama2-7B/13B." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 139, + 214, + 479, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 214, + 479, + 250 + ], + "spans": [ + { + "bbox": [ + 139, + 214, + 479, + 250 + ], + "type": "text", + "content": "- We collect STEVE-21K dataset, including " + }, + { + "bbox": [ + 139, + 214, + 479, + 250 + ], + "type": "inline_equation", + "content": "600+" + }, + { + "bbox": [ + 139, + 214, + 479, + 250 + ], + "type": "text", + "content": " vision-environment pairs, 20K knowledge question-answering pairs, and " + }, + { + "bbox": [ + 139, + 214, + 479, + 250 + ], + "type": "inline_equation", + "content": "200+" + }, + { + "bbox": [ + 139, + 214, + 479, + 250 + ], + "type": "text", + "content": " skill-code pairs, for justifying the effective performance of STEVE." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 269, + 242, + 282 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 269, + 242, + 282 + ], + "spans": [ + { + "bbox": [ + 132, + 269, + 242, + 282 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 295, + 312, + 307 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 295, + 312, + 307 + ], + "spans": [ + { + "bbox": [ + 132, + 295, + 312, + 307 + ], + "type": "text", + "content": "2.1 Intelligent Agent in Minecraft" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 315, + 482, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 315, + 482, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 482, + 518 + ], + "type": "text", + "content": "As an open-ended sandbox game, Minecraft has always been an ideal setting for testing the performance of intelligent agents [21, 25]. The agents are required to autonomously perform various tasks in Minecraft, such as chopping trees, crafting tools, and mining diamonds. At the beginning, much of the works focus on exploring reinforcement learning [32, 33, 42, 50] or imitation learning [2, 3], without satisfactory performance. VPT [3] and MineDojo [17] collect internet-scale datasets for their model pre-training. More specifically, VPT offers the exciting possibility of directly learning to act during video pre-training and using these learned behavioral priors as extremely effective exploration priors for reinforcement learning. Yet, recent works found that the pre-trained LLMs could serve as a strong \"mind\" that provides planning ability to the agents. Voyager [57] leverages GPT-4 [44] as both a high-level planner and a low-level action code generator. Plan4MC [69] proposes a skill graph pre-generated by the LLMs. DEPS [59], an interactive planning method based on LLMs, addresses multi-step reasoning issue in open-world planning. GITM [77] develops a set of structured actions and leverages LLMs to generate action plans for the agents to execute, achieving impressive results in various tasks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 536, + 309, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 309, + 548 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 309, + 548 + ], + "type": "text", + "content": "2.2 Embodied Multimodal Model" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "type": "text", + "content": "Embodied agent operates within various environments by synthesizing sensory perceptions and physical actions supported by computational intelligence. This synthesis enables the agent to undertake a variety of tasks, achieving specific objectives. Its key areas of application are diverse, including Navigation [6, 16, 27, 43, 61, 68], Embodied Question Answering [10, 11, 67], Active Visual Tracking [37, 38, 74, 75], and Visual Exploration [7, 12, 35]. The field is evolving rapidly with the development of Large Language Models (LLMs) [51] and Multimodal LLMs (MLLMs) [1, 9, 18, 19, 28–30, 34, 39, 41, 54, 58, 64, 76], integrating multiple modalities for more effective processing. A prime example of this innovation is" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "type": "text", + "content": "STEVE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 479, + 203 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 479, + 203 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 479, + 203 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 479, + 203 + ], + "type": "image", + "image_path": "1b252c5e77571c428318903d9dd735b64f6b4dec4c2328c8baa1f1769e6e9e2f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 211, + 482, + 266 + ], + "lines": [ + { + "bbox": [ + 130, + 211, + 482, + 266 + ], + "spans": [ + { + "bbox": [ + 130, + 211, + 482, + 266 + ], + "type": "text", + "content": "Fig. 2: STEVE framework. The Vision Perception part takes images or videos, encodes them into tokens, and combines them with Agent State and Task tokens as input. The STEVE-13B in the Language Instruction part is used for automatic reasoning and task decomposition, and it calls the Skill Database in the form of the Query to output code as action." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 290, + 482, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 290, + 482, + 327 + ], + "spans": [ + { + "bbox": [ + 130, + 290, + 482, + 327 + ], + "type": "text", + "content": "PaLM-E [15], a sophisticated multimodal model with 562B parameters, adept at a broad spectrum of embodied tasks and demonstrating exceptional capabilities in visual reasoning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 346, + 384, + 360 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 346, + 384, + 360 + ], + "spans": [ + { + "bbox": [ + 131, + 346, + 384, + 360 + ], + "type": "text", + "content": "2.3 Large Language Model with Equipped Tools" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 369, + 482, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 482, + 573 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 482, + 573 + ], + "type": "text", + "content": "While Large Language Models (LLMs) demonstrate impressive skill in tackling novel tasks via prompt-based instructions, they often face challenges in areas where simpler models or tools excel, like mathematical calculations or identifying palindromes. However, LLMs' potential is significantly expanded when integrated with other modality-specific models, such as those for vision or audio, enabling multi-modal capabilities [4,36]. Innovations like Toolformer [47] demonstrate LLMs' self-learning to utilize tools through finetuning with extensive API call samples. Visual ChatGPT [62] extends this by integrating various visual foundation models, facilitating interactive user experiences with ChatGPT. Similarly, HuggingGPT [48] presents a framework that harnesses LLMs to link diverse models from Hugging Face for task resolution. AutoGPT [49] is an open-source application that broadens GPT-4's capabilities with internet access, memory management, and plug-ins. The recent introduction of MovieChat [52,53] brings a memory mechanism to MLLM, enhancing its performance in video understanding tasks. Furthermore, LLMs can be used for goal planning, analogous to language translation [63]. This evolving landscape suggests that tool-equipped LLMs could forge a new paradigm in AI solution design." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 592, + 255, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 592, + 255, + 605 + ], + "spans": [ + { + "bbox": [ + 131, + 592, + 255, + 605 + ], + "type": "text", + "content": "3 Method: STEVE" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 620, + 209, + 631 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 620, + 209, + 631 + ], + "spans": [ + { + "bbox": [ + 131, + 620, + 209, + 631 + ], + "type": "text", + "content": "3.1 Overview" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "As shown in Figure 2, STEVE " + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": " is an LLM-based multi-modal autonomous system for embodied agents in Minecraft, which can use visual status " + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "inline_equation", + "content": "X^v" + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": " in the" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "text", + "content": "Z. Zhao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "content": "form of images or videos and agent state " + }, + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "inline_equation", + "content": "X^s" + }, + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "content": " to manage and execute complex task " + }, + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "inline_equation", + "content": "X^t" + }, + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "content": " for executive code action " + }, + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "inline_equation", + "content": "\\mathbf{a}^c" + }, + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 203, + 152, + 481, + 165 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 152, + 481, + 165 + ], + "spans": [ + { + "bbox": [ + 203, + 152, + 481, + 165 + ], + "type": "interline_equation", + "content": "\\mathbf {a} ^ {c} = \\mathcal {F} \\left(X ^ {v}, X ^ {s}, X ^ {t}\\right) = \\mathcal {A} ^ {c} \\left(\\mathcal {I} ^ {l} \\left(\\mathcal {P} ^ {v} \\left(X ^ {v}, X ^ {s}, X ^ {t}\\right)\\right)\\right) \\tag {1}", + "image_path": "1511b31a72bd993d2d2685a4c3a7dec0b16f834779fc7f971171dd541cd528f7.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 169, + 479, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 169, + 479, + 205 + ], + "spans": [ + { + "bbox": [ + 130, + 169, + 479, + 205 + ], + "type": "text", + "content": "where STEVE " + }, + { + "bbox": [ + 130, + 169, + 479, + 205 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 130, + 169, + 479, + 205 + ], + "type": "text", + "content": " integrates Visual Perception " + }, + { + "bbox": [ + 130, + 169, + 479, + 205 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^v" + }, + { + "bbox": [ + 130, + 169, + 479, + 205 + ], + "type": "text", + "content": " into the Language Instruction " + }, + { + "bbox": [ + 130, + 169, + 479, + 205 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^l" + }, + { + "bbox": [ + 130, + 169, + 479, + 205 + ], + "type": "text", + "content": " with large language model (LLM) and combines them with Code Action " + }, + { + "bbox": [ + 130, + 169, + 479, + 205 + ], + "type": "inline_equation", + "content": "\\mathcal{A}^c" + }, + { + "bbox": [ + 130, + 169, + 479, + 205 + ], + "type": "text", + "content": ", a skill retrieval method to execute actions, details are demonstrated as follows." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 215, + 479, + 320 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 215, + 479, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 215, + 479, + 249 + ], + "spans": [ + { + "bbox": [ + 138, + 215, + 479, + 249 + ], + "type": "text", + "content": "- Vision Perception " + }, + { + "bbox": [ + 138, + 215, + 479, + 249 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^v" + }, + { + "bbox": [ + 138, + 215, + 479, + 249 + ], + "type": "text", + "content": " (Section 3.2); a vision encoder to interpret visual information of the environment, such as blocks and entities and a text tokenizer for agent state " + }, + { + "bbox": [ + 138, + 215, + 479, + 249 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 138, + 215, + 479, + 249 + ], + "type": "text", + "content": " and task " + }, + { + "bbox": [ + 138, + 215, + 479, + 249 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 138, + 215, + 479, + 249 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 251, + 479, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 251, + 479, + 297 + ], + "spans": [ + { + "bbox": [ + 138, + 251, + 479, + 297 + ], + "type": "text", + "content": "- Language Instruction " + }, + { + "bbox": [ + 138, + 251, + 479, + 297 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^l" + }, + { + "bbox": [ + 138, + 251, + 479, + 297 + ], + "type": "text", + "content": " (Section 3.3): a powerful language model system fine-tuned specifically for Minecraft content using LLaMA2-13B [18]. This model enables adaptive interaction for iterative reasoning and step-by-step decomposition." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 298, + 479, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 298, + 479, + 320 + ], + "spans": [ + { + "bbox": [ + 138, + 298, + 479, + 320 + ], + "type": "text", + "content": "- Code Action " + }, + { + "bbox": [ + 138, + 298, + 479, + 320 + ], + "type": "inline_equation", + "content": "\\mathcal{A}^c" + }, + { + "bbox": [ + 138, + 298, + 479, + 320 + ], + "type": "text", + "content": " (Section 3.4): a skill retrieval method based on our skill-code database." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 131, + 339, + 268, + 352 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 339, + 268, + 352 + ], + "spans": [ + { + "bbox": [ + 131, + 339, + 268, + 352 + ], + "type": "text", + "content": "3.2 Vision Perception " + }, + { + "bbox": [ + 131, + 339, + 268, + 352 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^v" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 359, + 479, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 359, + 479, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 359, + 479, + 396 + ], + "type": "text", + "content": "The vision perception part includes a vision encoder " + }, + { + "bbox": [ + 130, + 359, + 479, + 396 + ], + "type": "inline_equation", + "content": "E^v" + }, + { + "bbox": [ + 130, + 359, + 479, + 396 + ], + "type": "text", + "content": " and a text tokenizer " + }, + { + "bbox": [ + 130, + 359, + 479, + 396 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 359, + 479, + 396 + ], + "type": "text", + "content": ", which converts the visual status " + }, + { + "bbox": [ + 130, + 359, + 479, + 396 + ], + "type": "inline_equation", + "content": "X^v" + }, + { + "bbox": [ + 130, + 359, + 479, + 396 + ], + "type": "text", + "content": ", agent state " + }, + { + "bbox": [ + 130, + 359, + 479, + 396 + ], + "type": "inline_equation", + "content": "X^s" + }, + { + "bbox": [ + 130, + 359, + 479, + 396 + ], + "type": "text", + "content": ", and task " + }, + { + "bbox": [ + 130, + 359, + 479, + 396 + ], + "type": "inline_equation", + "content": "X^t" + }, + { + "bbox": [ + 130, + 359, + 479, + 396 + ], + "type": "text", + "content": " to the text-space tokenizer representation " + }, + { + "bbox": [ + 130, + 359, + 479, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{Y} = \\{Y^v, Y^s, Y^t\\}" + }, + { + "bbox": [ + 130, + 359, + 479, + 396 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 227, + 407, + 480, + 420 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 407, + 480, + 420 + ], + "spans": [ + { + "bbox": [ + 227, + 407, + 480, + 420 + ], + "type": "interline_equation", + "content": "\\mathbf {Y} _ {i} = \\mathcal {P} ^ {v} \\left(X _ {i} ^ {v}, X _ {i} ^ {s}, X _ {i} ^ {t}\\right), \\mathcal {P} ^ {v} = \\left\\{E ^ {v}, T \\right\\}, \\tag {2}", + "image_path": "ccf99bfc9c534e07edcbadbcd5ad7fa8fa3579d32e5c41825888f2153b1e18a3.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "text", + "content": "where the vision encoder " + }, + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "text", + "content": ", the visual branch of EfficientFormer [31], encodes visual status " + }, + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "inline_equation", + "content": "X_{i}^{v}" + }, + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "text", + "content": " at each step " + }, + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "text", + "content": " into visual tokens " + }, + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "inline_equation", + "content": "Y^{v} = \\{y_{1}^{v}, y_{2}^{v}, \\dots, y_{n}^{v}\\} \\in \\mathbb{R}^{n \\times d}" + }, + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "text", + "content": " denotes the number of visual tokens " + }, + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "inline_equation", + "content": "y^{v}" + }, + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "text", + "content": " is the dimensionality of each token. The text tokenizer converts agent state " + }, + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "inline_equation", + "content": "X_{i}^{s}" + }, + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "text", + "content": " and task " + }, + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "inline_equation", + "content": "X_{i}^{t}" + }, + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "text", + "content": " at each step " + }, + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "text", + "content": " in the form of text into textual tokens " + }, + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "inline_equation", + "content": "Y^{s}" + }, + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "inline_equation", + "content": "Y^{t}" + }, + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "text", + "content": ". Note that visual tokens are amalgamated with textual tokens representing the agent's current state (e.g., health, inventory, etc.) and the task description. This is accomplished using a tokenizer that maps state variables and task parameters to a tokenized form. The resultant unified token set serves as a comprehensive representation of the current situational context." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 562, + 282, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 562, + 282, + 574 + ], + "spans": [ + { + "bbox": [ + 131, + 562, + 282, + 574 + ], + "type": "text", + "content": "3.3 Language Instruction " + }, + { + "bbox": [ + 131, + 562, + 282, + 574 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^l" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 581, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 479, + 665 + ], + "type": "text", + "content": "The Language Instruction " + }, + { + "bbox": [ + 130, + 581, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^l = \\{\\mathcal{P}_l, \\mathcal{C}_r, \\mathcal{C}_u, \\mathcal{D}_s\\}" + }, + { + "bbox": [ + 130, + 581, + 479, + 665 + ], + "type": "text", + "content": ", which consists of Planner " + }, + { + "bbox": [ + 130, + 581, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_l" + }, + { + "bbox": [ + 130, + 581, + 479, + 665 + ], + "type": "text", + "content": ", Critic " + }, + { + "bbox": [ + 130, + 581, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_r" + }, + { + "bbox": [ + 130, + 581, + 479, + 665 + ], + "type": "text", + "content": ", Curriculum " + }, + { + "bbox": [ + 130, + 581, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_u" + }, + { + "bbox": [ + 130, + 581, + 479, + 665 + ], + "type": "text", + "content": ", and Descriptor " + }, + { + "bbox": [ + 130, + 581, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_s" + }, + { + "bbox": [ + 130, + 581, + 479, + 665 + ], + "type": "text", + "content": ", i.e., four independent LLM-based agents with different functions. They formulate high-level task guidelines, refine strategies through feedback, facilitate continuous learning and adaptation through a curriculum of complex tasks, and finally decompose strategic guidelines into the executable low-level textual action step " + }, + { + "bbox": [ + 130, + 581, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{a}^s" + }, + { + "bbox": [ + 130, + 581, + 479, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 581, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{a}^s \\sim \\mathcal{I}^l(\\cdot|\\mathbf{Y}_i)\\}_{i=0}^N" + }, + { + "bbox": [ + 130, + 581, + 479, + 665 + ], + "type": "text", + "content": " towards efficient completion of the task." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "type": "text", + "content": "STEVE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "Note that the Planner " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_l" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": ", Critic " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_r" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": ", Curriculum " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_u" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": " and Descriptor " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_s" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": " are based on STEVE-7B/13B, a powerful language model derived from LLaMA-2-13B [18], fine-tuned specifically on Minecraft-related content from the STEVE-21K. This model's expertise covers a broad spectrum of game-specific knowledge areas on worlds, entities, player mechanics, survival, and even game practical experience. Finally, they have the performance of their different functions:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 198, + 479, + 293 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 138, + 198, + 479, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 198, + 479, + 220 + ], + "spans": [ + { + "bbox": [ + 138, + 198, + 479, + 220 + ], + "type": "text", + "content": "- Planner " + }, + { + "bbox": [ + 138, + 198, + 479, + 220 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_l" + }, + { + "bbox": [ + 138, + 198, + 479, + 220 + ], + "type": "text", + "content": ": formulating comprehensive guidelines and executive plans that align with the overarching objectives of the task." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 139, + 222, + 479, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 222, + 479, + 245 + ], + "spans": [ + { + "bbox": [ + 139, + 222, + 479, + 245 + ], + "type": "text", + "content": "- Critic " + }, + { + "bbox": [ + 139, + 222, + 479, + 245 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_r" + }, + { + "bbox": [ + 139, + 222, + 479, + 245 + ], + "type": "text", + "content": ": evaluating the planner decisions, providing feedback that can refine strategies." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 139, + 246, + 479, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 246, + 479, + 269 + ], + "spans": [ + { + "bbox": [ + 139, + 246, + 479, + 269 + ], + "type": "text", + "content": "- Curriculum " + }, + { + "bbox": [ + 139, + 246, + 479, + 269 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_u" + }, + { + "bbox": [ + 139, + 246, + 479, + 269 + ], + "type": "text", + "content": ": facilitating continuous learning and adaptation for action agents by engaging with a series of progressively complex tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 139, + 270, + 479, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 270, + 479, + 293 + ], + "spans": [ + { + "bbox": [ + 139, + 270, + 479, + 293 + ], + "type": "text", + "content": "- Descriptor " + }, + { + "bbox": [ + 139, + 270, + 479, + 293 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_s" + }, + { + "bbox": [ + 139, + 270, + 479, + 293 + ], + "type": "text", + "content": ": distilling the extensive data into a concise summary, making it more manageable and interpretable." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 306, + 482, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 306, + 482, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 306, + 482, + 426 + ], + "type": "text", + "content": "Iterative reasoning. The STEVE-13B receives a stream of tokens that encode the current visual scene, the agent's state, and the task's textual description. STEVE-13B interprets this rich context to undertake complex reasoning. The model initiates the reasoning process by constructing a series of high-level strategies that outline the pathway to task completion. Considering all gameplay elements, the reasoning mechanism is akin to an experienced player who can visualize the end game and chart a course to victory. This approach ensures the plans are reactive and proactive, allowing the agent to anticipate and mitigate future challenges. However, most strategies are high-level and abstract therefore, they often require step-by-step decomposition to derive executable guidelines." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 437, + 482, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 571 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 571 + ], + "type": "text", + "content": "Decomposition. The decomposition process makes the complex strategies breakdown into simple, low-level guidelines that can be directly mapped to actions in Minecraft. It is similar to how a high-level command like \"build a shelter\" is divided into actionable instructions like \"collect wood\", \"craft planks\", and \"place blocks\". The granular steps are structured to provide explicit instructions that the game engine can readily interpret. This systematic breakdown from high-level reasoning to low-level actionable steps is the hallmark of the STEVE system, enabling the embodied agent to interact with the Minecraft environment in a meaningful and goal-oriented manner. Through this intricate process of reasoning and decomposition, STEVE embodies the cognitive capabilities required for sophisticated task management in virtual environments." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Curriculum learning with memory. We draw inspiration from the lifelong learning strategy utilized in many reinforcement learning problems [57] in both closed-world and open-world settings. We start by creating a set of tasks that serve as a curriculum for agents to explore the environment. During this process, STEVE generates plans, interacts with the surroundings, learns from mistakes, and stores all these experiences in memory. Next, we evaluate STEVE on various tasks after this learning stage. Consequently, STEVE can produce better plans with its" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "Z. Zhao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "memory teaming up with the planning experiences. We use this as the default setting for all tasks in our experiments." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 156, + 482, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 156, + 482, + 288 + ], + "spans": [ + { + "bbox": [ + 130, + 156, + 482, + 288 + ], + "type": "text", + "content": "Continuous learning with summarization. We've observed that the learning process, where the memory is being filled, can continue throughout the gameplay. The agent can gradually acquire more skills as the gameplay progresses and more experiences are gained. However, as the memory gets larger, it becomes difficult to understand the game's situation and interact with the game slowly. To tackle these challenges, we have implemented the Chain of Summarization method [40]. By finding better references, we can improve our ability to handle challenging tasks, such as \"Diamond Tool\" in the tech tree, including obtaining materials and creating diamond tools. This will lead to a higher success rate, as shown in Section 5.3. Additionally, curriculum learning with memory allows for in-context lifelong learning without needing gradient updates." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 308, + 242, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 308, + 242, + 319 + ], + "spans": [ + { + "bbox": [ + 131, + 308, + 242, + 319 + ], + "type": "text", + "content": "3.4 Code Action " + }, + { + "bbox": [ + 131, + 308, + 242, + 319 + ], + "type": "inline_equation", + "content": "\\mathcal{A}^c" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 332, + 482, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 332, + 482, + 404 + ], + "spans": [ + { + "bbox": [ + 130, + 332, + 482, + 404 + ], + "type": "text", + "content": "The code action part " + }, + { + "bbox": [ + 130, + 332, + 482, + 404 + ], + "type": "inline_equation", + "content": "\\mathcal{A}^c" + }, + { + "bbox": [ + 130, + 332, + 482, + 404 + ], + "type": "text", + "content": " is the execution phase, where the STEVE system converts planned, decomposed guidelines into concrete actions within the Minecraft environment. This process leverages a specialized skill database that pairs code snippets with their descriptions and relevant metadata, encoded as vectors " + }, + { + "bbox": [ + 130, + 332, + 482, + 404 + ], + "type": "inline_equation", + "content": "\\mathbf{v}^{\\mathbf{s}}" + }, + { + "bbox": [ + 130, + 332, + 482, + 404 + ], + "type": "text", + "content": " for efficient retrieval. The transition from language instruction to executable code is achieved through the retrieval process " + }, + { + "bbox": [ + 130, + 332, + 482, + 404 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 130, + 332, + 482, + 404 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 240, + 418, + 480, + 432 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 240, + 418, + 480, + 432 + ], + "spans": [ + { + "bbox": [ + 240, + 418, + 480, + 432 + ], + "type": "interline_equation", + "content": "\\mathcal {R} (\\mathbf {q}, \\mathbf {v}) = \\sigma (\\mathbf {q}, \\mathbf {v}), \\mathbf {q} = E ^ {q} \\left(\\mathbf {a} ^ {\\mathbf {s}}\\right), \\tag {3}", + "image_path": "f4731743255b89a316ef5ca150b93fed3a0f437425356e759735034c74a04cfd.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 439, + 482, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 439, + 482, + 522 + ], + "spans": [ + { + "bbox": [ + 130, + 439, + 482, + 522 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 439, + 482, + 522 + ], + "type": "inline_equation", + "content": "E^q" + }, + { + "bbox": [ + 130, + 439, + 482, + 522 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 439, + 482, + 522 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 439, + 482, + 522 + ], + "type": "text", + "content": " are query encoding and cosine similarity matching. Each low-level textual action step " + }, + { + "bbox": [ + 130, + 439, + 482, + 522 + ], + "type": "inline_equation", + "content": "\\mathbf{a}^s" + }, + { + "bbox": [ + 130, + 439, + 482, + 522 + ], + "type": "text", + "content": " derived from the Language Instruction is encoded into a query " + }, + { + "bbox": [ + 130, + 439, + 482, + 522 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 130, + 439, + 482, + 522 + ], + "type": "text", + "content": ". This encoding captures the essence of the action to be performed in a format that the skill database can understand. Once the queries " + }, + { + "bbox": [ + 130, + 439, + 482, + 522 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 130, + 439, + 482, + 522 + ], + "type": "text", + "content": " are encoded, the system computes similarities between the query vectors and the code snippets vectors " + }, + { + "bbox": [ + 130, + 439, + 482, + 522 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 130, + 439, + 482, + 522 + ], + "type": "text", + "content": " stored in the database to determine which skill best matches the required action." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 544, + 204, + 556 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 544, + 204, + 556 + ], + "spans": [ + { + "bbox": [ + 131, + 544, + 204, + 556 + ], + "type": "text", + "content": "3.5 Training" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 567, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 567, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 567, + 482, + 628 + ], + "type": "text", + "content": "To reduce the training overhead to a certain extent, we adopt a two-stage training method, the first stage being a warm-up on the question-answering pairs of STEVE-21K to ensure a certain degree of instruction capability. The second stage is simulated in the environment, and the Expert LLM " + }, + { + "bbox": [ + 130, + 567, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_p" + }, + { + "bbox": [ + 130, + 567, + 482, + 628 + ], + "type": "text", + "content": " integrated with GPT-4 [44] generates instructions for the same situation to modify our model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "text", + "content": "In training, we use either a single-round conversation " + }, + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\{X^Q,X^A\\}" + }, + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "text", + "content": " or a multiround conversation " + }, + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\{X_i^Q,X_i^A\\}_{i\\leq N}" + }, + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "text", + "content": " is the total number of rounds in the conversation. We use the negative log-likelihood objective over the prediction" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "type": "text", + "content": "STEVE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 293, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 293, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 293, + 128 + ], + "type": "text", + "content": "tokens with training and finetuning:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 230, + 139, + 481, + 172 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 139, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 230, + 139, + 481, + 172 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (\\theta) = - \\sum_ {j = 1} ^ {L} \\log \\mathcal {F} _ {\\theta} \\left(Y _ {j} \\mid X ^ {v}, \\hat {Y} _ {1: j - 1}\\right), \\tag {4}", + "image_path": "069cb222e59e4de52b839c8c765a95b99f9a5fcacae88f42a55675d9c4e1905d.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 479, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 479, + 243 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 479, + 243 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 183, + 479, + 243 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 130, + 183, + 479, + 243 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 183, + 479, + 243 + ], + "type": "inline_equation", + "content": "\\hat{Y}" + }, + { + "bbox": [ + 130, + 183, + 479, + 243 + ], + "type": "text", + "content": " refer to the non-vision input and target token sequences, " + }, + { + "bbox": [ + 130, + 183, + 479, + 243 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 183, + 479, + 243 + ], + "type": "text", + "content": " represents the model parameters, and " + }, + { + "bbox": [ + 130, + 183, + 479, + 243 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 183, + 479, + 243 + ], + "type": "text", + "content": " represents the length of the target sequence. The vision input " + }, + { + "bbox": [ + 130, + 183, + 479, + 243 + ], + "type": "inline_equation", + "content": "X^v" + }, + { + "bbox": [ + 130, + 183, + 479, + 243 + ], + "type": "text", + "content": " varies depending on the input step. We only consider the answer tokens " + }, + { + "bbox": [ + 130, + 183, + 479, + 243 + ], + "type": "inline_equation", + "content": "X_A" + }, + { + "bbox": [ + 130, + 183, + 479, + 243 + ], + "type": "text", + "content": " when computing the loss to ensure that the model focuses on generating coherent responses." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 243, + 479, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 243, + 479, + 279 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 479, + 279 + ], + "type": "text", + "content": "Offline warm-up. In the first stage, we finetune the STEVE-7B/13B only on the single-round question-answering pairs of STEVE-21K to ensure a certain degree of instruction capability, as shown in Tab. 3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 280, + 480, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 280, + 480, + 316 + ], + "spans": [ + { + "bbox": [ + 130, + 280, + 480, + 316 + ], + "type": "text", + "content": "Online finetuning. In the second stage, we simulate the warmed-up STEVE-7B/13B in Minecraft, and both train the vision encoder " + }, + { + "bbox": [ + 130, + 280, + 480, + 316 + ], + "type": "inline_equation", + "content": "E^v" + }, + { + "bbox": [ + 130, + 280, + 480, + 316 + ], + "type": "text", + "content": " and finetune the STEVE-7B/13B simultaneously." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 316, + 480, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 480, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 480, + 399 + ], + "type": "text", + "content": "It is necessary to obtain the environment information to train the vision encoder, denoted as " + }, + { + "bbox": [ + 130, + 316, + 480, + 399 + ], + "type": "inline_equation", + "content": "l = R_{T}(X^{v})" + }, + { + "bbox": [ + 130, + 316, + 480, + 399 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 316, + 480, + 399 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 316, + 480, + 399 + ], + "type": "text", + "content": " is the label of seen blocks and entities, " + }, + { + "bbox": [ + 130, + 316, + 480, + 399 + ], + "type": "inline_equation", + "content": "R_{T}" + }, + { + "bbox": [ + 130, + 316, + 480, + 399 + ], + "type": "text", + "content": " is the Ray Tracing method [60] to eliminate all content outside the agent's perspective screen. We train the vision encoder " + }, + { + "bbox": [ + 130, + 316, + 480, + 399 + ], + "type": "inline_equation", + "content": "E^{v}" + }, + { + "bbox": [ + 130, + 316, + 480, + 399 + ], + "type": "text", + "content": " once we have obtained the environment information. To finetune the warmed-up STEVE-7B/13B in simulation, we give both the STEVE-7B/13B and the Expert LLM the same input, and we consider the output of the Expert LLM as the ground truth label." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 400, + 481, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 400, + 481, + 436 + ], + "spans": [ + { + "bbox": [ + 130, + 400, + 481, + 436 + ], + "type": "text", + "content": "After 5,000 simulations, we collect all the context information from successful runs, including sequences of vision and question-answering chat flow as the Vision-Environment part of STEVE-21K, as mentioned in Section 4." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 457, + 282, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 282, + 470 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 282, + 470 + ], + "type": "text", + "content": "4 Dataset: STEVE-21K" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 485, + 479, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 479, + 519 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 479, + 519 + ], + "type": "text", + "content": "As shown in Fig. 3, STEVE-21K has been meticulously compiled, featuring a diverse collection of Vision-Environment pairs, Question-Answering pairs, and a Skill-Code database." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 521, + 480, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 480, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 480, + 628 + ], + "type": "text", + "content": "Vision-Environment pairs contain 600 pairs of first-person perspective videos from Minecraft gameplay across six different terrains (including forest, desert, coastal, etc.), along with corresponding environmental block entities in the field of vision and context information for each timestamp. Additionally, all pairs are oriented around executing the skill-related task supported by Skill-Code part mentioned in Section 4. We employ the STEVE-7B/13B model to enable robots to plan and execute actions autonomously based on tasks defined by human supervisors. We record the video of the agent operation, the environment information, and all the corresponding chatflow." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "text", + "content": "Question-Answering pairs contain 20K question-answering pairs from the Minecraft-Wiki and Reddit corpus across six data types partly sourced from [17]. The pairs are organized into instruction, input, and output triplets and used to" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "Z. Zhao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 411, + 193 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 411, + 193 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 411, + 193 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 411, + 193 + ], + "type": "image", + "image_path": "4f0778d76d1ad6c52cc0e0aedd4767906814279c72cb3fab3da16fa83f00fe92.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 421, + 118, + 457, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 421, + 118, + 457, + 127 + ], + "spans": [ + { + "bbox": [ + 421, + 118, + 457, + 127 + ], + "type": "text", + "content": "Skill-Code" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 421, + 132, + 466, + 138 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 421, + 132, + 466, + 138 + ], + "spans": [ + { + "bbox": [ + 421, + 132, + 466, + 138 + ], + "type": "text", + "content": "Human Player" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 438, + 148, + 453, + 163 + ], + "blocks": [ + { + "bbox": [ + 438, + 148, + 453, + 163 + ], + "lines": [ + { + "bbox": [ + 438, + 148, + 453, + 163 + ], + "spans": [ + { + "bbox": [ + 438, + 148, + 453, + 163 + ], + "type": "image", + "image_path": "3204dda95c0702225d88f8d15a4b7828f65ac2d6679bd4ed5068a39a667c6f58.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 454, + 167, + 471, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 454, + 167, + 471, + 172 + ], + "spans": [ + { + "bbox": [ + 454, + 167, + 471, + 172 + ], + "type": "text", + "content": "prompt" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 439, + 179, + 453, + 193 + ], + "blocks": [ + { + "bbox": [ + 439, + 179, + 453, + 193 + ], + "lines": [ + { + "bbox": [ + 439, + 179, + 453, + 193 + ], + "spans": [ + { + "bbox": [ + 439, + 179, + 453, + 193 + ], + "type": "image", + "image_path": "1e1670b491385802b28f5fe1e2fd9c1f605e4fa7f9fed088c8ab601eccfeb9b3.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 457, + 182, + 466, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 457, + 182, + 466, + 188 + ], + "spans": [ + { + "bbox": [ + 457, + 182, + 466, + 188 + ], + "type": "text", + "content": "GPT" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 439, + 197, + 470, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 439, + 197, + 470, + 205 + ], + "spans": [ + { + "bbox": [ + 439, + 197, + 470, + 205 + ], + "type": "text", + "content": "output" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 433, + 213, + 464, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 433, + 213, + 464, + 219 + ], + "spans": [ + { + "bbox": [ + 433, + 213, + 464, + 219 + ], + "type": "text", + "content": "code snippets" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 438, + 229, + 454, + 237 + ], + "blocks": [ + { + "bbox": [ + 438, + 229, + 454, + 237 + ], + "lines": [ + { + "bbox": [ + 438, + 229, + 454, + 237 + ], + "spans": [ + { + "bbox": [ + 438, + 229, + 454, + 237 + ], + "type": "image", + "image_path": "fec95c1c56cf7cfcb11205fcee473edbe1b0bd7a1c8b3d30f066c14f173805e6.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 454, + 226, + 471, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 454, + 226, + 471, + 232 + ], + "spans": [ + { + "bbox": [ + 454, + 226, + 471, + 232 + ], + "type": "text", + "content": "execute" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 438, + 238, + 455, + 254 + ], + "blocks": [ + { + "bbox": [ + 438, + 238, + 455, + 254 + ], + "lines": [ + { + "bbox": [ + 438, + 238, + 455, + 254 + ], + "spans": [ + { + "bbox": [ + 438, + 238, + 455, + 254 + ], + "type": "image", + "image_path": "02bd1387363039c3fd997487d4a4b920c9dd4bda4fc10bbffd4b6f2b5933a415.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 426, + 258, + 468, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 426, + 258, + 468, + 264 + ], + "spans": [ + { + "bbox": [ + 426, + 258, + 468, + 264 + ], + "type": "text", + "content": "check and revise" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 143, + 201, + 211, + 209 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 201, + 211, + 209 + ], + "spans": [ + { + "bbox": [ + 143, + 201, + 211, + 209 + ], + "type": "text", + "content": "Question-Answering" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 143, + 228, + 187, + 256 + ], + "blocks": [ + { + "bbox": [ + 152, + 219, + 178, + 225 + ], + "lines": [ + { + "bbox": [ + 152, + 219, + 178, + 225 + ], + "spans": [ + { + "bbox": [ + 152, + 219, + 178, + 225 + ], + "type": "text", + "content": "Wikipedia" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 143, + 228, + 187, + 256 + ], + "lines": [ + { + "bbox": [ + 143, + 228, + 187, + 256 + ], + "spans": [ + { + "bbox": [ + 143, + 228, + 187, + 256 + ], + "type": "image", + "image_path": "1f57fa5bba76ddfef6a911c46b7ce601b83e13e345e685ec8366aa261fa44900.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 143, + 256, + 184, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 256, + 184, + 261 + ], + "spans": [ + { + "bbox": [ + 143, + 256, + 184, + 261 + ], + "type": "text", + "content": "The crafting menu in Minecraft, showing the crafting origins of a clone." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 143, + 261, + 183, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 261, + 183, + 266 + ], + "spans": [ + { + "bbox": [ + 143, + 261, + 183, + 266 + ], + "type": "text", + "content": "showing the eating recipe or a special taste as well as some other blocks and" + } + ] + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 196, + 229, + 236, + 269 + ], + "blocks": [ + { + "bbox": [ + 209, + 219, + 225, + 225 + ], + "lines": [ + { + "bbox": [ + 209, + 219, + 225, + 225 + ], + "spans": [ + { + "bbox": [ + 209, + 219, + 225, + 225 + ], + "type": "text", + "content": "Reddit" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 196, + 229, + 236, + 269 + ], + "lines": [ + { + "bbox": [ + 196, + 229, + 236, + 269 + ], + "spans": [ + { + "bbox": [ + 196, + 229, + 236, + 269 + ], + "type": "image", + "image_path": "60b81a3852d8af7061f20dd0b31c16210869d84a7930ee35b04a45fb03a288ee.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 249, + 209, + 263, + 224 + ], + "blocks": [ + { + "bbox": [ + 249, + 209, + 263, + 224 + ], + "lines": [ + { + "bbox": [ + 249, + 209, + 263, + 224 + ], + "spans": [ + { + "bbox": [ + 249, + 209, + 263, + 224 + ], + "type": "image", + "image_path": "a54b823eedd10842692d2794e3523dd8248c7e2eced360f5ce1571ae17dfd7c9.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "bbox": [ + 264, + 213, + 288, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 213, + 288, + 220 + ], + "spans": [ + { + "bbox": [ + 264, + 213, + 288, + 220 + ], + "type": "text", + "content": "ChatGPT" + } + ] + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 247, + 270, + 255, + 274 + ], + "blocks": [ + { + "bbox": [ + 247, + 270, + 255, + 274 + ], + "lines": [ + { + "bbox": [ + 247, + 270, + 255, + 274 + ], + "spans": [ + { + "bbox": [ + 247, + 270, + 255, + 274 + ], + "type": "image", + "image_path": "2831b8b1e4fd69d911915dbd4407eae73d885f7845e3e2f70795cbf433d150d3.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 285, + 482, + 394 + ], + "lines": [ + { + "bbox": [ + 130, + 285, + 482, + 394 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 482, + 394 + ], + "type": "text", + "content": "Fig.3: STEVE-21K collection pipeline. In the Vision-Environment section, STEVE-13B plays the game according to specified tasks defined by the human player, collecting visual information through prismarine-viewer and capturing environmental information from the screen using Ray Tracing [60]. Note that the language instruction task is also performed during the collection phase. We simultaneously record and save the chat flow from the reasoning and decomposition stages. In the Question-Answering section, we obtain information from the Minecraft-Wiki and Reddit forums and use GPT-3.5 to clean the data into single-round QA pairs. In the Skill-Code section, we use GPT-3.5 combined with the human player's code to synthesize code snippets and then check and revise them in the game environment." + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "bbox": [ + 263, + 229, + 293, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 263, + 229, + 293, + 237 + ], + "spans": [ + { + "bbox": [ + 263, + 229, + 293, + 237 + ], + "type": "text", + "content": "Clean" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 274, + 241, + 296, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 274, + 241, + 296, + 247 + ], + "spans": [ + { + "bbox": [ + 274, + 241, + 296, + 247 + ], + "type": "text", + "content": "Construct" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 277, + 256, + 293, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 256, + 293, + 261 + ], + "spans": [ + { + "bbox": [ + 277, + 256, + 293, + 261 + ], + "type": "text", + "content": "Check" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 343, + 212, + 394, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 212, + 394, + 218 + ], + "spans": [ + { + "bbox": [ + 343, + 212, + 394, + 218 + ], + "type": "text", + "content": "Single-round QA pairs" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 330, + 222, + 405, + 262 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 330, + 222, + 378, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 222, + 378, + 228 + ], + "spans": [ + { + "bbox": [ + 330, + 222, + 378, + 228 + ], + "type": "text", + "content": "1. World & Entities" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 330, + 228, + 405, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 228, + 405, + 235 + ], + "spans": [ + { + "bbox": [ + 330, + 228, + 405, + 235 + ], + "type": "text", + "content": "2. Player Mechanics & Survival" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 330, + 235, + 394, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 235, + 394, + 241 + ], + "spans": [ + { + "bbox": [ + 330, + 235, + 394, + 241 + ], + "type": "text", + "content": "3. Knowledge & Discovery" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 330, + 241, + 388, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 241, + 388, + 247 + ], + "spans": [ + { + "bbox": [ + 330, + 241, + 388, + 247 + ], + "type": "text", + "content": "4. Resources & Crafting" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 330, + 247, + 376, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 247, + 376, + 254 + ], + "spans": [ + { + "bbox": [ + 330, + 247, + 376, + 254 + ], + "type": "text", + "content": "5. Tools & Utilities" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 330, + 254, + 372, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 254, + 372, + 262 + ], + "spans": [ + { + "bbox": [ + 330, + 254, + 372, + 262 + ], + "type": "text", + "content": "6. Miscellaneous" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 416, + 480, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 416, + 480, + 452 + ], + "spans": [ + { + "bbox": [ + 130, + 416, + 480, + 452 + ], + "type": "text", + "content": "train STEVE-13B. The GPT-3.5 [5] is employed to derive meaningful single-round question-answer pairs, and LoRA [22] is incorporated during the fin-tuning process for efficient resource allocation." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 130, + 452, + 481, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 452, + 481, + 500 + ], + "spans": [ + { + "bbox": [ + 130, + 452, + 481, + 500 + ], + "type": "text", + "content": "Skill-Code pairs contain 210 skill execution scripts with descriptions, covering 8 skill types including collecting, crafting, exploration etc. The code part is collected by manual coding. We use GPT-3.5 [5] to describe all codes and utilize langchain vecordb to give all pairs a database vector." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 131, + 517, + 229, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 517, + 229, + 530 + ], + "spans": [ + { + "bbox": [ + 131, + 517, + 229, + 530 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 131, + 540, + 261, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 540, + 261, + 552 + ], + "spans": [ + { + "bbox": [ + 131, + 540, + 261, + 552 + ], + "type": "text", + "content": "5.1 Experimental Setup" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "We train STEVE-7B/13B, which is finetuned from LLaMA-2 [18] with the Question-Answering pair in STEVE-21K dataset for warm-up and simulation context data from successful runs. We use LoRA [22] for finetuning process. Note that the process is to adjust STEVE-13B to work on correct simulation knowledge while remaining adapted to visual perception. In the text part, we set all temperatures to 0 except for the task proposal, which uses 0.9 to encourage task diversity. The vision unit is based on EfficientFormerV2-S0 [31], which is trained on the Vision-Environment part of our STEVE-21K dataset. Our simulation environment is built on top of MineDojo [17] and leverages Mineflayer [45]." + } + ] + } + ], + "index": 41 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "type": "text", + "content": "STEVE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 145, + 116, + 471, + 370 + ], + "blocks": [ + { + "bbox": [ + 145, + 116, + 471, + 370 + ], + "lines": [ + { + "bbox": [ + 145, + 116, + 471, + 370 + ], + "spans": [ + { + "bbox": [ + 145, + 116, + 471, + 370 + ], + "type": "image", + "image_path": "5d30bf45360efd026ed4a76d83aeb386e7564905498c795635ad07fe0ea13126.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 377, + 482, + 411 + ], + "lines": [ + { + "bbox": [ + 130, + 377, + 482, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 377, + 482, + 411 + ], + "type": "text", + "content": "Fig. 4: Example of Vision-Environment pairs. It represents the data format of the Vision-Environment pairs in our STEVE-21K dataset: including visual signals, environmental information, Chat Log, Context QA pairs, and planning in actual tasks." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 432, + 482, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 432, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 432, + 482, + 456 + ], + "type": "text", + "content": "We use GPT-4-0613 for all GPT-4 models used in Voyager [57] and code generation tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 472, + 208, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 208, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 208, + 483 + ], + "type": "text", + "content": "5.2Baselines" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 490, + 482, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 490, + 482, + 528 + ], + "spans": [ + { + "bbox": [ + 130, + 490, + 482, + 528 + ], + "type": "text", + "content": "As no vision-based LLM-driven agents are immediately operable in Minecraft, we selected several algorithms as baselines that extract information from a system's backend, differing significantly from real-world applications." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 536, + 482, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 536, + 482, + 586 + ], + "spans": [ + { + "bbox": [ + 130, + 536, + 482, + 586 + ], + "type": "text", + "content": "AutoGPT [49] is an NLP automation tool that decomposes a high-level goal into executable subgoals in MineDojo, aligning with our experimental framework. Our setup, AutoGPT, powered by GPT-4 [44], processes agent states, environment feedback, and execution errors to manage subgoal execution." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Voyager [57] relies solely on textual grounding for perception and features a long-term procedural memory with a hierarchical library of code-based procedures, allowing the integration of simple skills into complex behaviors. Proficient in environment exploration and tech tree mastery, Voyager uses GPT-4 [44] for processing background text in embodied agents, with a lesser emphasis on visual perception." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "Z. Zhao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 179, + 478, + 245 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 2: Comparison on tech tree mastery task. The values presented are in fractions, representing successful trials out of three attempts. A score of " + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "0/3" + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": " signifies the method's inability to progress within the tech tree after a maximum of 160 prompting iterations. The reported numbers denote the average iterations across three trials. Lower iteration values indicate higher efficiency of the respective method." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 179, + 478, + 245 + ], + "lines": [ + { + "bbox": [ + 136, + 179, + 478, + 245 + ], + "spans": [ + { + "bbox": [ + 136, + 179, + 478, + 245 + ], + "type": "table", + "html": "
MethodWooden ToolStone ToolIron ToolDiamond Tool
AutoGPT [49]92 ± 72 (3/3)94 ± 72 (3/3)135 ± 103 (3/3)N/A (0/3)
Voyager [57]6 ± 2 (3/3)11 ± 2 (3/3)21 ± 7 (3/3)102 (1/3)
STEVE4 ± 1 (3/3)8 ± 1 (3/3)15 ± 2 (3/3)106 ± 12 (3/3)
", + "image_path": "ac1d1fd6c6392197cc1ecae407139e6f255a7c8c499915b10d7c9a862bad374d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 135, + 346, + 478, + 419 + ], + "blocks": [ + { + "bbox": [ + 130, + 258, + 482, + 335 + ], + "lines": [ + { + "bbox": [ + 130, + 258, + 482, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 482, + 335 + ], + "type": "text", + "content": "Table 3: Quantitative comparison on knowledge question and answering task. Questions, model-generated responses, and ground truth inputs are evaluated in GPT-4 [5], Claude-2 [8] and human blind rating rated on a scale of 0 to 10; The scores above are the average of them. Higher scores indicate greater alignment of the generated answers with the ground truth. Wld., Ent., Mech., Surv., Know., Disc., Res., Craft., Tools, Util., Miscell. stand for World, Entities, Player Mechanics, Survival, Knowledge, Discovery, Resources, Crafting, Tools, Utilities and Miscellaneous." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 346, + 478, + 419 + ], + "lines": [ + { + "bbox": [ + 135, + 346, + 478, + 419 + ], + "spans": [ + { + "bbox": [ + 135, + 346, + 478, + 419 + ], + "type": "table", + "html": "
MethodWld. & Ent.Mech. & Surv.Know. & Disc.Res. & Craft.Tl. & Util.Miscell.Overall
Llama2-7B6.446.686.586.426.806.966.56
Llama2-13B6.936.956.776.776.986.646.89
STEVE-7B7.997.887.847.957.937.827.94
STEVE-13B8.148.138.038.158.127.728.12
GPT-48.068.078.077.928.098.218.04
", + "image_path": "07be74d97082cd27390d15e42a08bd14c3302aa14b1273954adad3044baa5e09.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 445, + 256, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 445, + 256, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 445, + 256, + 456 + ], + "type": "text", + "content": "5.3 Evaluation Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 495, + 303, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 303, + 613 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 303, + 613 + ], + "type": "text", + "content": "Continuous block search. As shown in Tab. 1, we experiment with block-searching tasks to assess the agent's exploratory capabilities and proficiency in locating specified blocks. Diamond blocks are placed at every 16-block interval across the land map. The agent's objective is to identify as many blocks as possible within the fewest iterations, which indicates the method's efficiency." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 613, + 481, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 613, + 481, + 649 + ], + "spans": [ + { + "bbox": [ + 130, + 613, + 481, + 649 + ], + "type": "text", + "content": "As shown in Fig. 5, enriching information through visual perception significantly enhances the efficiency of search and exploration tasks, leading to more effective world exploration." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 306, + 554, + 482, + 613 + ], + "blocks": [ + { + "bbox": [ + 304, + 498, + 482, + 552 + ], + "lines": [ + { + "bbox": [ + 304, + 498, + 482, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 482, + 552 + ], + "type": "text", + "content": "Table 1: Comparison on continues block search task. # Iters stand for average iterations to find 10 diamonds (max 100). # Blocks stand for average diamonds found in 100 iterations." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 554, + 482, + 613 + ], + "lines": [ + { + "bbox": [ + 306, + 554, + 482, + 613 + ], + "spans": [ + { + "bbox": [ + 306, + 554, + 482, + 613 + ], + "type": "table", + "html": "
Method# Iters (↓) # Blocks (↑)
AutoGPT [49]N/A 7
Voyager [57]35 26
STEVE14 67
", + "image_path": "cb37e65ab50b47f4cf8c67bc3cd1a0e979e14341b0a2f1939d0ab5f70c0c39c9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "type": "text", + "content": "STEVE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 168, + 114, + 445, + 373 + ], + "blocks": [ + { + "bbox": [ + 168, + 114, + 445, + 373 + ], + "lines": [ + { + "bbox": [ + 168, + 114, + 445, + 373 + ], + "spans": [ + { + "bbox": [ + 168, + 114, + 445, + 373 + ], + "type": "image", + "image_path": "c3ffbd171cdc51eefbe4bfce50cd9a15ff960f3abcf3684c6d35aa8c13925cc9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 381, + 482, + 427 + ], + "lines": [ + { + "bbox": [ + 130, + 381, + 482, + 427 + ], + "spans": [ + { + "bbox": [ + 130, + 381, + 482, + 427 + ], + "type": "text", + "content": "Fig.5: Schematic of the continuous block search task. We capture an asynchronous segment with each method 30 iterations from the experiments for illustration. The reason we choose diamond blocks is that they are not naturally occurring in the given context, making them easily distinguishable from other blocks." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 461, + 482, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 534 + ], + "type": "text", + "content": "Knowledge question and answering. Using a validation dataset, we established a question-answering database to evaluate our model's performance on Minecraft-related queries. Responses from each model are rated blindly by GPT-4, Claude2, and human participants based on accuracy, relevance, and detail. Responses are initially checked for accuracy, and the comprehensive evaluation yields a score ranging from 0 to 10, with higher scores indicating superior performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 540, + 482, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 540, + 482, + 624 + ], + "spans": [ + { + "bbox": [ + 130, + 540, + 482, + 624 + ], + "type": "text", + "content": "As detailed in Tab. 3, we tested the instructional capabilities of various LLM models using the STEVE-21k dataset, divided into 18,622 training and 1,000 testing samples. The STEVE-7B and STEVE-13B models surpassed LLaMA2 in all metrics, with STEVE-13B scoring the highest at 8.12, demonstrating its exceptional understanding of Minecraft-related queries. This indicates that STEVE models, optimized specifically for Minecraft content, perform better in knowledge-intensive tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 629, + 483, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 483, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 483, + 665 + ], + "type": "text", + "content": "Our results confirm that larger, domain-specifically tuned models like STEVE-13B outperform broader models like GPT-4, underscoring the benefit of specialized fine-tuning for domain-specific applications." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "text", + "content": "Z. Zhao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 116, + 219, + 163 + ], + "blocks": [ + { + "bbox": [ + 137, + 116, + 219, + 163 + ], + "lines": [ + { + "bbox": [ + 137, + 116, + 219, + 163 + ], + "spans": [ + { + "bbox": [ + 137, + 116, + 219, + 163 + ], + "type": "image", + "image_path": "e91ccdc799cd7cd93fb52b9144b461282a4db2a37877cbcce4f3d97d2dfbe61a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 225, + 116, + 306, + 163 + ], + "blocks": [ + { + "bbox": [ + 225, + 116, + 306, + 163 + ], + "lines": [ + { + "bbox": [ + 225, + 116, + 306, + 163 + ], + "spans": [ + { + "bbox": [ + 225, + 116, + 306, + 163 + ], + "type": "image", + "image_path": "f95ac2622f513174bd27eba737e9765cde210570ec63862fa79bdf9bab14841d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 312, + 116, + 392, + 163 + ], + "blocks": [ + { + "bbox": [ + 312, + 116, + 392, + 163 + ], + "lines": [ + { + "bbox": [ + 312, + 116, + 392, + 163 + ], + "spans": [ + { + "bbox": [ + 312, + 116, + 392, + 163 + ], + "type": "image", + "image_path": "fc15d8b43a6c6a219405b1d97201503701202651a2f7d734e028fe522ba6484d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 399, + 116, + 479, + 163 + ], + "blocks": [ + { + "bbox": [ + 399, + 116, + 479, + 163 + ], + "lines": [ + { + "bbox": [ + 399, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 399, + 116, + 479, + 163 + ], + "type": "image", + "image_path": "826c2d76a5225d8c324005a52ac9cc6da89c423df693190dee59d07b3648129f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 137, + 164, + 219, + 182 + ], + "blocks": [ + { + "bbox": [ + 137, + 164, + 219, + 182 + ], + "lines": [ + { + "bbox": [ + 137, + 164, + 219, + 182 + ], + "spans": [ + { + "bbox": [ + 137, + 164, + 219, + 182 + ], + "type": "image", + "image_path": "e8070962c9f4a31a7677295ef27366233d4975df507973f088419d805fa02cbb.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 225, + 164, + 306, + 182 + ], + "blocks": [ + { + "bbox": [ + 225, + 164, + 306, + 182 + ], + "lines": [ + { + "bbox": [ + 225, + 164, + 306, + 182 + ], + "spans": [ + { + "bbox": [ + 225, + 164, + 306, + 182 + ], + "type": "image", + "image_path": "3e28f748d3988cb1973af8c1df6eec5e33e4da8bd1276a7727972f2a9c4f86f2.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 312, + 165, + 392, + 182 + ], + "blocks": [ + { + "bbox": [ + 312, + 165, + 392, + 182 + ], + "lines": [ + { + "bbox": [ + 312, + 165, + 392, + 182 + ], + "spans": [ + { + "bbox": [ + 312, + 165, + 392, + 182 + ], + "type": "image", + "image_path": "2324a70c27db609ef3d1411c5025133c864e153dfe60253efcb06a9a55ab6a70.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 399, + 165, + 479, + 182 + ], + "blocks": [ + { + "bbox": [ + 399, + 165, + 479, + 182 + ], + "lines": [ + { + "bbox": [ + 399, + 165, + 479, + 182 + ], + "spans": [ + { + "bbox": [ + 399, + 165, + 479, + 182 + ], + "type": "image", + "image_path": "087df738ce9447c33074cd6826cefb2d7fc946fda7f4b7ca8604f99ae1bf4fbb.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 136, + 182, + 219, + 219 + ], + "blocks": [ + { + "bbox": [ + 136, + 182, + 219, + 219 + ], + "lines": [ + { + "bbox": [ + 136, + 182, + 219, + 219 + ], + "spans": [ + { + "bbox": [ + 136, + 182, + 219, + 219 + ], + "type": "image", + "image_path": "ae46503e694aa75d149810b04b5403c9befffb96e2a4e08c74e63134dee5111c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 225, + 182, + 306, + 227 + ], + "blocks": [ + { + "bbox": [ + 225, + 182, + 306, + 227 + ], + "lines": [ + { + "bbox": [ + 225, + 182, + 306, + 227 + ], + "spans": [ + { + "bbox": [ + 225, + 182, + 306, + 227 + ], + "type": "image", + "image_path": "f74fe55769e23c2d01404e6b275ba8a4ca5a81d9da5594903fdbce86fce42076.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 312, + 182, + 392, + 226 + ], + "blocks": [ + { + "bbox": [ + 312, + 182, + 392, + 226 + ], + "lines": [ + { + "bbox": [ + 312, + 182, + 392, + 226 + ], + "spans": [ + { + "bbox": [ + 312, + 182, + 392, + 226 + ], + "type": "image", + "image_path": "617f7a9b756b06ff25e43a44bf5948fce18fa933315ddf8f864628d7b0fd925b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 399, + 182, + 479, + 227 + ], + "blocks": [ + { + "bbox": [ + 399, + 182, + 479, + 227 + ], + "lines": [ + { + "bbox": [ + 399, + 182, + 479, + 227 + ], + "spans": [ + { + "bbox": [ + 399, + 182, + 479, + 227 + ], + "type": "image", + "image_path": "e609be31ebf1b0024d4118fdfa22b57c753d09ccbc627e5d3d163aa183cfb842.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 136, + 227, + 219, + 262 + ], + "blocks": [ + { + "bbox": [ + 136, + 227, + 219, + 262 + ], + "lines": [ + { + "bbox": [ + 136, + 227, + 219, + 262 + ], + "spans": [ + { + "bbox": [ + 136, + 227, + 219, + 262 + ], + "type": "image", + "image_path": "c5b040852d0693814ed8f16ca4a63c4e0d6452a82e506ae37a75449c814516e1.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 225, + 227, + 306, + 262 + ], + "blocks": [ + { + "bbox": [ + 225, + 227, + 306, + 262 + ], + "lines": [ + { + "bbox": [ + 225, + 227, + 306, + 262 + ], + "spans": [ + { + "bbox": [ + 225, + 227, + 306, + 262 + ], + "type": "image", + "image_path": "05047e68461b01b191892fc16bd6bc59e36da72c351c8660b722cb198b3583e1.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 312, + 227, + 392, + 262 + ], + "blocks": [ + { + "bbox": [ + 312, + 227, + 392, + 262 + ], + "lines": [ + { + "bbox": [ + 312, + 227, + 392, + 262 + ], + "spans": [ + { + "bbox": [ + 312, + 227, + 392, + 262 + ], + "type": "image", + "image_path": "96b586df6b7667fa9e0203dd7313c6411d9b45c5380c63ca50ecf4cb29f4e256.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 399, + 227, + 479, + 262 + ], + "blocks": [ + { + "bbox": [ + 399, + 227, + 479, + 262 + ], + "lines": [ + { + "bbox": [ + 399, + 227, + 479, + 262 + ], + "spans": [ + { + "bbox": [ + 399, + 227, + 479, + 262 + ], + "type": "image", + "image_path": "630b50962ab2d34f205431369fa48fea883ec201012cebf11e9bea5d2d0acbcd.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 136, + 262, + 219, + 307 + ], + "blocks": [ + { + "bbox": [ + 136, + 262, + 219, + 307 + ], + "lines": [ + { + "bbox": [ + 136, + 262, + 219, + 307 + ], + "spans": [ + { + "bbox": [ + 136, + 262, + 219, + 307 + ], + "type": "image", + "image_path": "508641d0d956ba9a170be8b6d37737357a9c553eeabafeeb8f70f6333d062b04.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 320, + 482, + 357 + ], + "lines": [ + { + "bbox": [ + 132, + 320, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 320, + 482, + 357 + ], + "type": "text", + "content": "Fig. 6: Qualitative comparison on knowledge question and answering tasks. Green (Red) highlights the correct or good (wrong or bad) answer. Blue indicates the suboptimal answer. Grey indicates the meaningless answer." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 225, + 262, + 306, + 307 + ], + "blocks": [ + { + "bbox": [ + 225, + 262, + 306, + 307 + ], + "lines": [ + { + "bbox": [ + 225, + 262, + 306, + 307 + ], + "spans": [ + { + "bbox": [ + 225, + 262, + 306, + 307 + ], + "type": "image", + "image_path": "c30134ec6506fd05cc9e003840ad135900d7545f3bf00ee52d3a61549895794d.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 312, + 262, + 372, + 307 + ], + "blocks": [ + { + "bbox": [ + 312, + 262, + 372, + 307 + ], + "lines": [ + { + "bbox": [ + 312, + 262, + 372, + 307 + ], + "spans": [ + { + "bbox": [ + 312, + 262, + 372, + 307 + ], + "type": "image", + "image_path": "e55b903770769d2184e50c956562fadd07a462f2d95589666552e3622843d302.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 382, + 262, + 479, + 307 + ], + "blocks": [ + { + "bbox": [ + 382, + 262, + 479, + 307 + ], + "lines": [ + { + "bbox": [ + 382, + 262, + 479, + 307 + ], + "spans": [ + { + "bbox": [ + 382, + 262, + 479, + 307 + ], + "type": "image", + "image_path": "daed7dce2e57dd123eb53e9fbaf4f7032951c523ddb4f63c08a49b6a26a41f3f.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 130, + 382, + 482, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 382, + 482, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 382, + 482, + 525 + ], + "type": "text", + "content": "Tech tree mastery. As shown in Tab. 2, we experiment on the Minecraft tech tree mastery to test the agent's ability to craft and use a hierarchy of tools. Progressing through this tree (wooden tool " + }, + { + "bbox": [ + 130, + 382, + 482, + 525 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 382, + 482, + 525 + ], + "type": "text", + "content": " stone tool " + }, + { + "bbox": [ + 130, + 382, + 482, + 525 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 382, + 482, + 525 + ], + "type": "text", + "content": " iron tool " + }, + { + "bbox": [ + 130, + 382, + 482, + 525 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 382, + 482, + 525 + ], + "type": "text", + "content": " diamond tool) requires the agent to master systematic and compositional skills. As to the wooden, stone, and iron levels of the tech tree, STEVE achieves remarkable efficiency: " + }, + { + "bbox": [ + 130, + 382, + 482, + 525 + ], + "type": "inline_equation", + "content": "23 \\times" + }, + { + "bbox": [ + 130, + 382, + 482, + 525 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 382, + 482, + 525 + ], + "type": "inline_equation", + "content": "11.8 \\times" + }, + { + "bbox": [ + 130, + 382, + 482, + 525 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 382, + 482, + 525 + ], + "type": "inline_equation", + "content": "9 \\times" + }, + { + "bbox": [ + 130, + 382, + 482, + 525 + ], + "type": "text", + "content": " faster than AutoGPT [49], and " + }, + { + "bbox": [ + 130, + 382, + 482, + 525 + ], + "type": "inline_equation", + "content": "1.5 \\times" + }, + { + "bbox": [ + 130, + 382, + 482, + 525 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 382, + 482, + 525 + ], + "type": "inline_equation", + "content": "1.4 \\times" + }, + { + "bbox": [ + 130, + 382, + 482, + 525 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 382, + 482, + 525 + ], + "type": "inline_equation", + "content": "1.3 \\times" + }, + { + "bbox": [ + 130, + 382, + 482, + 525 + ], + "type": "text", + "content": " faster than Voyager [57]. STEVE has achieved the diamond level, as shown in Tab. 2. Its performance slightly lags behind Voyager [57], which also uses GPT4 [44] for critical inference. However, STEVE is more cost-effective, starting with lower initial performance. It includes a vision unit, prioritizing visual data over background information, offering distinct advantages. Additionally, we observed a decrease in performance when using a basic skill database." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 132, + 545, + 238, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 238, + 558 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 238, + 558 + ], + "type": "text", + "content": "5.4 Ablation Study" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 130, + 567, + 482, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 567, + 482, + 662 + ], + "spans": [ + { + "bbox": [ + 130, + 567, + 482, + 662 + ], + "type": "text", + "content": "To understand the impact of different components on the performance of our system, we conducted ablation studies focusing on the tech tree mastery task in Minecraft. The results, as shown in Tab. 4, provide insights into the effectiveness of the vision unit and compare our STEVE model with the STEVE GPT-4 version (with the same vision unit as ours). Note that the w/o vision unit setup is that the environmental perception encompasses data on blocks within an 8x8 area surrounding the agent, including the front, back, left, and right directions. The following observations are made:" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "type": "text", + "content": "STEVE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 151, + 157, + 459, + 218 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 4: Ablation studies for the tech tree mastery. STEVE (Ours) is the STEVE-13B version. The " + }, + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "0/3" + }, + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "text", + "content": " score means the method can't progress beyond 160 iterations in the tech tree." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 151, + 157, + 459, + 218 + ], + "lines": [ + { + "bbox": [ + 151, + 157, + 459, + 218 + ], + "spans": [ + { + "bbox": [ + 151, + 157, + 459, + 218 + ], + "type": "table", + "html": "
MethodWooden ToolStone ToolIron ToolDiamond Tool
w/o vision unit11 ± 5 (3/3)27 ± 5 (3/3)46 ± 11 (3/3)158 (1/3)
STEVE (GPT-4)6 ± 2 (3/3)10 ± 1 (3/3)14 ± 3 (3/3)89 ± 9 (3/3)
STEVE (Ours)4 ± 1 (3/3)8 ± 1 (3/3)15 ± 2 (3/3)106 ± 12 (3/3)
", + "image_path": "e2bde9aae14746f4b2def2f7f9bc2ec18a9126fc7f6f2c5f1868c45e87ae463e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 243, + 482, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 243, + 482, + 302 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 482, + 302 + ], + "type": "text", + "content": "Vision unit is critical. The omission of the vision unit markedly affects the system's performance, especially in more advanced tasks. While it successfully crafts Wooden, Stone, and Iron Tools, it is challenged with Diamond Tools. This outcome underscores the vital importance of visual information in accomplishing complex tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 320, + 482, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 320, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 320, + 482, + 392 + ], + "type": "text", + "content": "Comparison with GPT-4. As our vision encoder directly encodes into text space, it can be easily replaced with any language model. For instance, the GPT-4 we compared exhibits consistent success across all categories and secures a flawless success rate. Interestingly, the STEVE-13B version excels in simpler tasks such as crafting wooden and stone tools. Moreover, it requires fewer iterations than methods without the vision part, underscoring its superior efficiency." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 416, + 218, + 429 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 416, + 218, + 429 + ], + "spans": [ + { + "bbox": [ + 131, + 416, + 218, + 429 + ], + "type": "text", + "content": "5.5 Case Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 442, + 482, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 442, + 482, + 527 + ], + "spans": [ + { + "bbox": [ + 130, + 442, + 482, + 527 + ], + "type": "text", + "content": "As shown in Fig. 6, we perform an extensive case study comparison of GPT-4, LLaMA2-13B, and our method STEVE-13B. Each model maintains the same information and question inputs to compare feedback under different environmental information. Our STEVE overall achieves the best results, surpassing GPT-4 and showing significant improvement compared to the original LLaMA. Especially in parts involving numerical calculations, such as the leftmost image, STEVE accurately tracks food values to restore hunger levels." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 550, + 220, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 550, + 220, + 563 + ], + "spans": [ + { + "bbox": [ + 131, + 550, + 220, + 563 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "STEVE enhances multi-modal learning by combining visual encoder and LLM-based agents. It has three functions: vision perception, language instruction, and code action, allowing it to understand, predict, and act in virtual environments. We provide a straightforward approach to creating a robust, multimodal, autonomous, embodied agent using an open-source language model with a small number of parameters. Additionally, we provide a comprehensive dataset STEVE-21K for sustainable community development that can be verified." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "Z. Zhao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 482, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 482, + 178 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 482, + 178 + ], + "type": "text", + "content": "This work is supported by the Zhejiang Provincial Natural Science Foundation of China (No. LZ24F030005) and the National Natural Science Foundation of China (No. 62106219)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 197, + 197, + 209 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 197, + 197, + 209 + ], + "spans": [ + { + "bbox": [ + 133, + 197, + 197, + 209 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 223, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 223, + 481, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 223, + 481, + 266 + ], + "spans": [ + { + "bbox": [ + 138, + 223, + 481, + 266 + ], + "type": "text", + "content": "1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems 35, 23716-23736 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 268, + 481, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 268, + 481, + 289 + ], + "spans": [ + { + "bbox": [ + 138, + 268, + 481, + 289 + ], + "type": "text", + "content": "2. Amiranashvili, A., Dorka, N., Burgard, W., Koltun, V., Brox, T.: Scaling imitation learning in apacheft. arXiv preprint arXiv:2007.02701 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 290, + 481, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 290, + 481, + 332 + ], + "spans": [ + { + "bbox": [ + 138, + 290, + 481, + 332 + ], + "type": "text", + "content": "3. Baker, B., Akkaya, I., Zhokov, P., Huizinga, J., Tang, J., Ecoffet, A., Houghton, B., Sampedro, R., Clune, J.: Video pretraining (vpt): Learning to act by watching unlabeled online videos. Advances in Neural Information Processing Systems 35, 24639-24654 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 334, + 481, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 334, + 481, + 355 + ], + "spans": [ + { + "bbox": [ + 138, + 334, + 481, + 355 + ], + "type": "text", + "content": "4. Chai, W., Wang, G.: Deep vision multimodal learning: Methodology, benchmark, and trend. Applied Sciences 12(13), 6588 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 357, + 412, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 357, + 412, + 367 + ], + "spans": [ + { + "bbox": [ + 138, + 357, + 412, + 367 + ], + "type": "text", + "content": "5. Introducing chatgpt (2022), https://openai.com/blog/chatgpt" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 368, + 481, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 368, + 481, + 399 + ], + "spans": [ + { + "bbox": [ + 138, + 368, + 481, + 399 + ], + "type": "text", + "content": "6. Chen, S., Guhur, P.L., Schmid, C., Laptev, I.: History aware multimodal transformer for vision-and-language navigation. Advances in neural information processing systems 34, 5834-5847 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 401, + 481, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 401, + 481, + 422 + ], + "spans": [ + { + "bbox": [ + 138, + 401, + 481, + 422 + ], + "type": "text", + "content": "7. Chen, T., Gupta, S., Gupta, A.: Learning exploration policies for navigation. In: International Conference on Learning Representations (2018)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 423, + 323, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 423, + 323, + 433 + ], + "spans": [ + { + "bbox": [ + 138, + 423, + 323, + 433 + ], + "type": "text", + "content": "8. Talk to claude (2023), https://claude.ai" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 434, + 481, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 434, + 481, + 466 + ], + "spans": [ + { + "bbox": [ + 138, + 434, + 481, + 466 + ], + "type": "text", + "content": "9. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 468, + 481, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 468, + 481, + 499 + ], + "spans": [ + { + "bbox": [ + 138, + 468, + 481, + 499 + ], + "type": "text", + "content": "0. Das, A., Datta, S., Gkioxari, G., Lee, S., Parikh, D., Batra, D.: Embodied question answering. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1-10 (2018)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 501, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 501, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 501, + 481, + 533 + ], + "type": "text", + "content": "1. Datta, S., Dharur, S., Cartillier, V., Desai, R., Khanna, M., Batra, D., Parikh, D.: Episodic memory question answering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19119-19128 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 534, + 481, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 534, + 481, + 565 + ], + "spans": [ + { + "bbox": [ + 138, + 534, + 481, + 565 + ], + "type": "text", + "content": "2. Dean, V., Tulsiani, S., Gupta, A.: See, hear, explore: Curiosity via audio-visual association. Advances in neural information processing systems 33, 14961-14972 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 567, + 481, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 481, + 597 + ], + "type": "text", + "content": "3. Deng, J., Chai, W., Guo, J., Huang, Q., Hu, W., Hwang, J.N., Wang, G.: Citygen: Infinite and controllable 3d city layout generation. arXiv preprint arXiv:2312.01508 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 600, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 600, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 600, + 481, + 632 + ], + "type": "text", + "content": "4. Deng, J., Chai, W., Huang, J., Zhao, Z., Huang, Q., Gao, M., Guo, J., Hao, S., Hu, W., Hwang, J.N., et al.: Citycraft: A real craftsfer for 3d city generation. arXiv preprint arXiv:2406.04983 (2024)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "type": "text", + "content": "5. Driess, D., Xia, F., Sajjadi, M.S., Lynch, C., Chowdhery, A., Ichter, B., Wahid, A., Tompson, J., Vuong, Q., Yu, T., et al.: Palm-e: An embodied multimodal language model. arXiv preprint arXiv:2303.03378 (2023)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "type": "text", + "content": "STEVE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "16. Du, H., Yu, X., Zheng, L.: Vtnet: Visual transformer network for object goal navigation. In: International Conference on Learning Representations (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 481, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 481, + 181 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 481, + 181 + ], + "type": "text", + "content": "17. Fan, L., Wang, G., Jiang, Y., Mandlekar, A., Yang, Y., Zhu, H., Tang, A., Huang, D.A., Zhu, Y., Anandkumar, A.: Minedojo: Building open-ended embodied agents with internet-scale knowledge. Advances in Neural Information Processing Systems 35, 18343-18362 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 481, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 481, + 214 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 481, + 214 + ], + "type": "text", + "content": "18. Gao, P., Han, J., Zhang, R., Lin, Z., Geng, S., Zhou, A., Zhang, W., Lu, P., He, C., Yue, X., et al.: Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 215, + 481, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 215, + 481, + 246 + ], + "spans": [ + { + "bbox": [ + 132, + 215, + 481, + 246 + ], + "type": "text", + "content": "19. Gong, T., Lyu, C., Zhang, S., Wang, Y., Zheng, M., Zhao, Q., Liu, K., Zhang, W., Luo, P., Chen, K.: Multimodal-gpt: A vision and language model for dialogue with humans. arXiv preprint arXiv:2305.04790 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 247, + 481, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 247, + 481, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 247, + 481, + 278 + ], + "type": "text", + "content": "20. Guss, W.H., Houghton, B., Topin, N., Wang, P., Codel, C., Veloso, M., Salakhutdinov, R.: Minerl: A large-scale dataset of apache demonstrations. arXiv preprint arXiv:1907.13440 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 279, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 279, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 279, + 481, + 300 + ], + "type": "text", + "content": "21. Hofmann, K.: Minecraft as ai playground and laboratory. In: Proceedings of the annual symposium on computer-human interaction in play. pp. 1-1 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 300, + 481, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 481, + 331 + ], + "type": "text", + "content": "22. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 332, + 481, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 481, + 364 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 481, + 364 + ], + "type": "text", + "content": "23. Huang, W., Xia, F., Xiao, T., Chan, H., Liang, J., Florence, P., Zeng, A., Tompson, J., Mordatch, I., Chebotar, Y., et al.: Inner monologue: Embodied reasoning through planning with language models. arXiv preprint arXiv:2207.05608 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 365, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 481, + 396 + ], + "type": "text", + "content": "24. Jiang, J., Ye, T., Bai, J., Chen, S., Chai, W., Jun, S., Liu, Y., Chen, E.: Five a+ network: You only need 9k parameters for underwater image enhancement. arXiv preprint arXiv:2305.08824 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 397, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 481, + 418 + ], + "type": "text", + "content": "25. Johnson, M., Hofmann, K., Hutton, T., Bignell, D.: The malmo platform for artificial intelligence experimentation. In: Ijcai. pp. 4246-4247 (2016)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 418, + 481, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 481, + 450 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 481, + 450 + ], + "type": "text", + "content": "26. Kolve, E., Mottaghi, R., Han, W., VanderBilt, E., Weihs, L., Herrasti, A., Deitke, M., Ehsani, K., Gordon, D., Zhu, Y., et al.: Ai2-thor: An interactive 3d environment for visual ai. arXiv preprint arXiv:1712.05474 (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 450, + 481, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 481, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 481, + 482 + ], + "type": "text", + "content": "27. Kwon, O., Park, J., Oh, S.: Rendering neural radiance map for visual navigation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9099-9108 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 483, + 481, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 481, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 481, + 504 + ], + "type": "text", + "content": "28. Li, B., Zhang, Y., Chen, L., Wang, J., Yang, J., Liu, Z.: Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 504, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 481, + 536 + ], + "type": "text", + "content": "29. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 536, + 481, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 481, + 568 + ], + "type": "text", + "content": "30. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 569, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 569, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 569, + 481, + 601 + ], + "type": "text", + "content": "31. Li, Y., Yuan, G., Wen, Y., Hu, J., Evangelidis, G., Tulyakov, S., Wang, Y., Ren, J.: Efficientformer: Vision transformers at mobilenet speed. Advances in Neural Information Processing Systems 35, 12934-12949 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "type": "text", + "content": "32. Lifshitz, S., Paster, K., Chan, H., Ba, J., McIlraith, S.: Steve-1: A generative model for text-to-behavior in apache (abridged version). In: NeurIPS 2023 Workshop on Goal-Conditioned Reinforcement Learning (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "type": "text", + "content": "33. Lin, Z., Li, J., Shi, J., Ye, D., Fu, Q., Yang, W.: Juewu-mc: Playing minecraft with sample-efficient hierarchical reinforcement learning. arXiv preprint arXiv:2112.04907 (2021)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "Z. Zhao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "type": "text", + "content": "34. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv preprint arXiv:2304.08485 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 172 + ], + "type": "text", + "content": "35. Liu, S., Okatani, T.: Symmetry-aware neural architecture for embodied visual exploration. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 17221-17230. IEEE (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 173, + 480, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 173, + 480, + 206 + ], + "spans": [ + { + "bbox": [ + 132, + 173, + 480, + 206 + ], + "type": "text", + "content": "36. Lu, P., Peng, B., Cheng, H., Galley, M., Chang, K.W., Wu, Y.N., Zhu, S.C., Gao, J.: Chameleon: Plug-and-play compositional reasoning with large language models. arXiv preprint arXiv:2304.09842 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 207, + 480, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 207, + 480, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 207, + 480, + 239 + ], + "type": "text", + "content": "37. Luo, W., Sun, P., Zhong, F., Liu, W., Zhang, T., Wang, Y.: End-to-end active object tracking via reinforcement learning. In: International conference on machine learning. pp. 3286-3295. PMLR (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 240, + 480, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 240, + 480, + 274 + ], + "spans": [ + { + "bbox": [ + 132, + 240, + 480, + 274 + ], + "type": "text", + "content": "38. Luo, W., Sun, P., Zhong, F., Liu, W., Zhang, T., Wang, Y.: End-to-end active object tracking and its real-world deployment via reinforcement learning. IEEE transactions on pattern analysis and machine intelligence 42(6), 1317-1332 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 274, + 480, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 480, + 307 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 480, + 307 + ], + "type": "text", + "content": "39. Lyu, C., Wu, M., Wang, L., Huang, X., Liu, B., Du, Z., Shi, S., Tu, Z.: Macaw-llm: Multi-modal language modeling with image, audio, video, and text integration. arXiv preprint arXiv:2306.09093 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 308, + 480, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 480, + 340 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 480, + 340 + ], + "type": "text", + "content": "40. Ma, W., Mi, Q., Yan, X., Wu, Y., Lin, R., Zhang, H., Wang, J.: Large language models play starcraft ii: Benchmarks and a chain of summarization approach. arXiv preprint arXiv:2312.11865 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 341, + 480, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 341, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 341, + 480, + 373 + ], + "type": "text", + "content": "41. Maaz, M., Rasheed, H., Khan, S., Khan, F.S.: Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 374, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 480, + 418 + ], + "type": "text", + "content": "42. Mao, H., Wang, C., Hao, X., Mao, Y., Lu, Y., Wu, C., Hao, J., Li, D., Tang, P.: Seihai: A sample-efficient hierarchical ai for the minerl competition. In: Distributed Artificial Intelligence: Third International Conference, DAI 2021, Shanghai, China, December 17-18, 2021, Proceedings 3. pp. 38-51. Springer (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 419, + 480, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 419, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 419, + 480, + 453 + ], + "type": "text", + "content": "43. Moudgil, A., Majumdar, A., Agrawal, H., Lee, S., Batra, D.: Soat: A scene-and object-aware transformer for vision-and-language navigation. Advances in Neural Information Processing Systems 34, 7357-7367 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 453, + 472, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 472, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 472, + 464 + ], + "type": "text", + "content": "44. OpenAI: Gpt-4 technical report. arXiv preprint arXiv: Arxiv-2303.08774 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 464, + 480, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 480, + 497 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 480, + 497 + ], + "type": "text", + "content": "45. PrismarineJS: Prismarinejs/mineflayer: Create apache bots with a powerful, stable, and high level javascript api. (2013), https://github.com/PrismarineJS/mineflayer/tree/master" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 498, + 480, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 480, + 542 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 480, + 542 + ], + "type": "text", + "content": "46. Savva, M., Kadian, A., Maksymets, O., Zhao, Y., Wijmans, E., Jain, B., Straub, J., Liu, J., Koltun, V., Malik, J., et al.: Habitat: A platform for embodied ai research. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9339-9347 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 543, + 480, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 543, + 480, + 575 + ], + "spans": [ + { + "bbox": [ + 132, + 543, + 480, + 575 + ], + "type": "text", + "content": "47. Schick, T., Dwivedi-Yu, J., Dessi, R., Raileanu, R., Lomeli, M., Zettlemoyer, L., Cancedda, N., Scialom, T.: Toolformer: Language models can teach themselves to use tools. arXiv preprint arXiv:2302.04761 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 576, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 576, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 576, + 480, + 609 + ], + "type": "text", + "content": "48. Shen, Y., Song, K., Tan, X., Li, D., Lu, W., Zhuang, Y.: Hugginggpt: Solving ai tasks with chatgpt and its friends in huggingface. arXiv preprint arXiv:2303.17580 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 609, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 480, + 632 + ], + "type": "text", + "content": "49. Significant-Gravitas: Auto-gpt. https://github.com/Significant-Gravitas/Auto-GPT (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "text", + "content": "50. Skrynnik, A., Staroverov, A., Aitygulov, E., Aksenov, K., Davydov, V., Panov, A.I.: Hierarchical deep q-network from imperfect demonstrations in minecraft. Cognitive Systems Research 65, 74-78 (2021)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "type": "text", + "content": "STEVE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "51. Song, C.H., Wu, J., Washington, C., Sadler, B.M., Chao, W.L., Su, Y.: Llm-planner: Few-shot grounded planning for embodied agents with large language models. arXiv preprint arXiv:2212.04088 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 149, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 480, + 182 + ], + "type": "text", + "content": "52. Song, E., Chai, W., Wang, G., Zhang, Y., Zhou, H., Wu, F., Guo, X., Ye, T., Lu, Y., Hwang, J.N., et al.: Moviechat: From dense token to sparse memory for long video understanding. arXiv preprint arXiv:2307.16449 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 480, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 480, + 214 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 480, + 214 + ], + "type": "text", + "content": "53. Song, E., Chai, W., Ye, T., Hwang, J.N., Li, X., Wang, G.: Moviechat+: Question-aware sparse memory for long video question answering. arXiv preprint arXiv:2404.17176 (2024)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 214, + 480, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 214, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 214, + 480, + 236 + ], + "type": "text", + "content": "54. Su, Y., Lan, T., Li, H., Xu, J., Wang, Y., Cai, D.: Pandagpt: One model to instruction-follow them all. arXiv preprint arXiv:2305.16355 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 236, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 480, + 267 + ], + "type": "text", + "content": "55. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 267, + 480, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 267, + 480, + 311 + ], + "spans": [ + { + "bbox": [ + 132, + 267, + 480, + 311 + ], + "type": "text", + "content": "56. Vasluianu, F.A., Seizinger, T., Zhou, Z., Wu, Z., Chen, C., Timofte, R., Dong, W., Zhou, H., Tian, Y., Chen, J., et al.: Ntire 2024 image shadow removal challenge report. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6547-6570 (2024)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 311, + 480, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 311, + 480, + 343 + ], + "spans": [ + { + "bbox": [ + 132, + 311, + 480, + 343 + ], + "type": "text", + "content": "57. Wang, G., Xie, Y., Jiang, Y., Mandlekar, A., Xiao, C., Zhu, Y., Fan, L., Anandkumar, A.: Voyager: An open-ended embodied agent with large language models. arXiv preprint arXiv:2305.16291 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 343, + 480, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 480, + 375 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 480, + 375 + ], + "type": "text", + "content": "58. Wang, W., Chen, Z., Chen, X., Wu, J., Zhu, X., Zeng, G., Luo, P., Lu, T., Zhou, J., Qiao, Y., et al.: Visionllm: Large language model is also an open-ended decoder for vision-centric tasks. arXiv preprint arXiv:2305.11175 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 375, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 375, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 375, + 480, + 407 + ], + "type": "text", + "content": "59. Wang, Z., Cai, S., Liu, A., Ma, X., Liang, Y.: Describe, explain, plan and select: Interactive planning with large language models enables open-world multi-task agents. arXiv preprint arXiv:2302.01560 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 407, + 480, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 480, + 429 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 480, + 429 + ], + "type": "text", + "content": "60. Whitted, T.: An improved illumination model for shaded display. In: ACM SIGGRAPH 2005 Courses, pp. 4-es (2005)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 429, + 480, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 429, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 480, + 460 + ], + "type": "text", + "content": "61. Wijmans, E., Kadian, A., Morcos, A., Lee, S., Essa, I., Parikh, D., Savva, M., Batra, D.: Dd-ppo: Learning near-perfect pointgoal navigators from 2.5 billion frames. In: International Conference on Learning Representations (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 460, + 480, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 460, + 480, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 480, + 493 + ], + "type": "text", + "content": "62. Wu, C., Yin, S., Qi, W., Wang, X., Tang, Z., Duan, N.: Visual chatgpt: Talking, drawing and editing with visual foundation models. arXiv preprint arXiv:2303.04671 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 493, + 480, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 493, + 480, + 515 + ], + "spans": [ + { + "bbox": [ + 132, + 493, + 480, + 515 + ], + "type": "text", + "content": "63. Xie, Y., Yu, C., Zhu, T., Bai, J., Gong, Z., Soh, H.: Translating natural language to planning goals with large-language models. arXiv preprint arXiv:2302.05128 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 515, + 480, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 515, + 480, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 515, + 480, + 547 + ], + "type": "text", + "content": "64. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 547, + 480, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 547, + 480, + 579 + ], + "spans": [ + { + "bbox": [ + 132, + 547, + 480, + 579 + ], + "type": "text", + "content": "65. Ye, T., Chen, S., Liu, Y., Chai, W., Bai, J., Zou, W., Zhang, Y., Jiang, M., Chen, E., Xue, C.: Sequential affinity learning for video restoration. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 4147-4156 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 579, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 579, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 579, + 480, + 611 + ], + "type": "text", + "content": "66. Ye, T., Zhang, Y., Jiang, M., Chen, L., Liu, Y., Chen, S., Chen, E.: Perceiving and modeling density for image dehazing. In: European conference on computer vision. pp. 130-145. Springer (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "text", + "content": "67. Yu, L., Chen, X., Gkioxari, G., Bansal, M., Berg, T.L., Batra, D.: Multi-target embodied question answering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6309-6318 (2019)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "68. Yu, Y., Huang, W., Sun, F., Chen, C., Wang, Y., Liu, X.: Sound adversarial audiovisual navigation. In: International Conference on Learning Representations (2021)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "Z. Zhao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 423 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "69. Yuan, H., Zhang, C., Wang, H., Xie, F., Cai, P., Dong, H., Lu, Z.: Plan4mc: Skill reinforcement learning and planning for open-world apache tasks. arXiv preprint arXiv:2303.16563 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "type": "text", + "content": "70. Zhao, Z., Chai, W., Hao, S., Hu, W., Wang, G., Cao, S., Song, M., Hwang, J.N., Wang, G.: A survey of deep learning in sports applications: Perception, comprehension, and decision. arXiv preprint arXiv:2307.03353 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "type": "text", + "content": "71. Zhao, Z., Chai, W., Wang, X., Ma, K., Chen, K., Guo, D., Ye, T., Zhang, Y., Wang, H., Wang, G.: Steve series: Step-by-step construction of agent systems in northwest China. arXiv preprint arXiv:2406.11247 (2024)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 215, + 482, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 215, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 215, + 482, + 247 + ], + "type": "text", + "content": "72. Zhao, Z., Chen, K., Guo, D., Chai, W., Ye, T., Zhang, Y., Wang, G.: Hierarchical auto-organizing system for open-ended multi-agent navigation. arXiv preprint arXiv:2403.08282 (2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 248, + 482, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 482, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 482, + 281 + ], + "type": "text", + "content": "73. Zhao, Z., Ma, K., Chai, W., Wang, X., Chen, K., Guo, D., Zhang, Y., Wang, H., Wang, G.: Do we really need a complex agent system? distill embodied agent into a single model. arXiv preprint arXiv:2404.04619 (2024)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 281, + 482, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 314 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 314 + ], + "type": "text", + "content": "74. Zhong, F., Sun, P., Luo, W., Yan, T., Wang, Y.: Ad-vat+: An asymmetric dueling mechanism for learning and understanding visual active tracking. IEEE transactions on pattern analysis and machine intelligence 43(5), 1467-1482 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 314, + 482, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 314, + 482, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 482, + 346 + ], + "type": "text", + "content": "75. Zhong, F., Sun, P., Luo, W., Yan, T., Wang, Y.: Towards distraction-robust active visual tracking. In: International Conference on Machine Learning. pp. 12782-12792. PMLR (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 347, + 482, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 347, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 482, + 380 + ], + "type": "text", + "content": "76. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 380, + 482, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 423 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 423 + ], + "type": "text", + "content": "77. Zhu, X., Chen, Y., Tian, H., Tao, C., Su, W., Yang, C., Huang, G., Li, B., Lu, L., Wang, X., et al.: Ghost in the minecraft: Generally capable agents for open-world enviroments via large language models with text-based knowledge and memory. arXiv preprint arXiv:2305.17144 (2023)" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "type": "text", + "content": "STEVE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Seeing Faces in Things_ A Model and Dataset for Pareidolia/78844de0-7820-498e-be36-c38ded4cc8e3_content_list.json b/2024/Seeing Faces in Things_ A Model and Dataset for Pareidolia/78844de0-7820-498e-be36-c38ded4cc8e3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..853b6e440cfcd974a161d78ceb835f040fde2faf --- /dev/null +++ b/2024/Seeing Faces in Things_ A Model and Dataset for Pareidolia/78844de0-7820-498e-be36-c38ded4cc8e3_content_list.json @@ -0,0 +1,2030 @@ +[ + { + "type": "text", + "text": "Seeing Faces in Things: A Model and Dataset for Pareidolia", + "text_level": 1, + "bbox": [ + 193, + 48, + 803, + 104 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mark Hamilton $^{1,2}$ , Simon Stent $^{3}$ , Vasha DuTell $^{1}$ , Anne Harrington $^{1}$ , Jennifer Corbett $^{1}$ , Ruth Rosenholtz $^{4}$ , and William T. Freeman $^{1}$", + "bbox": [ + 101, + 139, + 895, + 180 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 MIT, 2 Microsoft, 3 Toyota Research Institute, 4 NVIDIA", + "bbox": [ + 198, + 193, + 796, + 213 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. The human visual system is well-tuned to detect faces of all shapes and sizes. While this brings obvious survival advantages, such as a better chance of spotting unknown predators in the bush, it also leads to spurious face detections. \"Face pareidolia\" describes the perception of face-like structure among otherwise random stimuli: seeing faces in coffee stains or clouds in the sky. In this paper, we study face pareidolia from a computer vision perspective. We present an image dataset of \"Faces in Things\", consisting of five thousand web images with human-annotated pareidolic faces. Using this dataset, we examine the extent to which a state-of-the-art human face detector exhibits pareidolia, and find a significant behavioral gap between humans and machines. We find that the evolutionary need for humans to detect animal faces, as well as human faces, may explain some of this gap. Finally, we propose a simple statistical model of pareidolia in images. Through studies on human subjects and our pareidolic face detectors we confirm a key prediction of our model regarding what image conditions are most likely to induce pareidolia. Dataset and Website: https://aka.ms/faces-in-things", + "bbox": [ + 147, + 258, + 849, + 562 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Pareidolia $\\cdot$ Face Detection $\\cdot$ Human Psychophysics", + "bbox": [ + 142, + 580, + 779, + 598 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 77, + 630, + 309, + 650 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hamlet: Do you see yonder cloud that's almost in the shape of a camel?", + "bbox": [ + 72, + 674, + 794, + 692 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Polonius: By the Mass and 'tis, like a camel indeed.", + "bbox": [ + 72, + 692, + 598, + 708 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hamlet: Methinks it is a weasel.", + "bbox": [ + 72, + 710, + 404, + 726 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Polonius: It is back'd like a weasel.", + "bbox": [ + 72, + 728, + 433, + 743 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hamlet: Or like a whale.", + "bbox": [ + 72, + 746, + 331, + 760 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Polonius: Very like a whale.", + "bbox": [ + 72, + 764, + 360, + 780 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "— Hamlet, Act III, Scene ii, William Shakespeare", + "bbox": [ + 421, + 791, + 920, + 811 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pareidolia is a type of visual \"apophenia\", which refers to the perception of patterns in random data. This occurs frequently in human perception as we look at clouds, mountain skylines, and burnt toast. Pareidolia is even described in an exchange in Hamlet [43]. When it was first described, pareidolia was seen as an early symptom of psychosis [7, 44]. Today we know pareidolia is common among healthy humans [45] and infants [23]. It is also not confined to humans:", + "bbox": [ + 72, + 830, + 920, + 949 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f78aab5b9190439d9daef9925f38018d7e676cab28517fbb93a0561c8b304a7e.jpg", + "image_caption": [ + "Fig. 1: You print out an exciting new computer vision paper to review, but as you sit down at your desk to start reading you knock over your coffee cup. At first, you are annoyed, but then, you laugh! The sight of the stain induces \"pareidolia\" in your brain: rather than an unsightly blemish, you see a happy face. In this paper we explore the phenomenon of face pareidolia: Why don't we see faces all the time? Why do we see them at all when they are clearly so different from human faces? Can a better understanding of face pareidolia help computer vision-based face detection?" + ], + "image_footnote": [], + "bbox": [ + 79, + 69, + 401, + 256 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "rhesus macaques, for example, have been shown to spend more time fixating on pareidolic than non-pareidolic images, in a manner similar to humans [49].", + "bbox": [ + 77, + 323, + 917, + 362 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As an intriguing phenomenon of our visual system, pareidolia presents many opportunities in the study of the visual perception of both humans and machines. It offers a controlled setting in which to study object detection: we can present random signals to the visual system and study what detections arise. Do computer vision detectors exhibit similar misidentifications, and if not, why not? Why don't humans see pareidolic effects everywhere, in any textured region?", + "bbox": [ + 77, + 364, + 917, + 479 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To help answer these questions, we introduce an annotated dataset of five thousand pareidolic face images, called \"Faces in Things\". With this dataset, we examine whether modern computer vision face detection systems, trained to robustly detect human faces, exhibit pareidolia. We show that a state-of-the-art neural network trained on the popular WIDER FACE detection benchmark [60] fails to detect pareidolic faces well, even when detection thresholds are relaxed. By fine-tuning the same model on the Faces in Things training data we create a simple and strong baseline for the task of pareidolic face detection, which shows that significantly higher machine pareidolic performance is within reach.", + "bbox": [ + 77, + 481, + 917, + 653 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Next, we explore how we might bridge this gap to supervised—or ultimately, human—performance, without access to pareidolic training data? Could pareidolia appear in a face detector in a more natural way? The Faces in Things dataset provides a clean testbed to explore these questions in machines. We test a variety of different interventions ranging from image augmentation techniques to additional sources of training data. We find one possible mechanism that accounts for roughly half of the performance gap: when models are fine-tuned to detect animal faces, pareidolic face detection is significantly improved. This suggests that face pareidolia may arise in part from a more general, evolutionary need to detect diverse faces in the natural environment.", + "bbox": [ + 77, + 656, + 917, + 848 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Finally, we consider why pareidolic faces are not all around us, and why certain textures seem to cause the effect more often. We propose two simple mathematical models, a simple Gaussian process model, and a second deep feature-based model, that capture important features of pareidolia. In particular, we show how these simple models both predict a \"Goldilocks\" zone, where", + "bbox": [ + 77, + 850, + 917, + 947 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 77, + 17, + 94, + 29 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "M. Hamilton, S. Stent, V. DuTell, et al.", + "bbox": [ + 157, + 14, + 549, + 32 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "conditions are ideal to induce pareidolia. We confirm the existence of this zone with experiments on both human subjects and face detection models.", + "bbox": [ + 77, + 55, + 917, + 95 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Through the contributions of our open-source dataset, models, and experimental findings, we bring the study of the intriguing phenomenon of face pareidolia to the computer vision community.", + "bbox": [ + 77, + 95, + 917, + 154 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 79, + 185, + 331, + 206 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Face detection. One of the most famous early examples of face detection was the Viola-Jones face detector [52, 53]. This detector used binary Haar features through simple-to-compute integral images and achieved greater precision and efficiency than early neural-network-based detectors [40, 41] and other feature-based methods [37, 59]. Following the deep learning breakthroughs of the 2010s, methods transitioned from hand-crafted features to learned features, and convolutional neural networks (CNNs) achieved close to human levels of performance on ever-larger datasets [8, 27-29, 32, 36, 48, 63]. For a broader survey of face detection methods, we refer the interested reader to [36, 62]. In our work, we use the recent RetinaFace model [8] as a strong face detection baseline.", + "bbox": [ + 79, + 227, + 917, + 421 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Neuroscience of face pareidolia. The face is a highly unique stimulus for the human visual system [26, 50]: we find faces easy to spot and difficult to ignore. Face detection can occur in both noise and highly degraded images [5]. Prior work shows that face detection occurs in a dedicated brain region, the Fusiform Face Area [33]. But exactly what constitutes a face for the visual cortex and what are the mechanisms underlying pareidolia? A recent study into the temporal dynamics of neuro-imaging data during pareidolic face viewing showed results consistent with \"a broadly-tuned face detection mechanism that privileges sensitivity over selectivity\" [56]. Pareidolic faces do more than give the impression of the presence of faces: [46] show that they can trigger an additional face-specific attentional process, consuming more time and processing power than similar non-pareidolic stimuli, and even enhancing the detection of face-pareidolic objects [47]. Analyses in [30] revealed a network of neurons in the brain specialized to detect face pareidolia. Their results suggested that face processing has a strong top-down component whereby sensory input with even the slightest suggestion of a face can result in the interpretation of a face. Such top-down information might be supportive of some form of inverse rendering as a cognitive mechanism to explain the remarkable robustness of human perception of faces in degraded viewing conditions [11]. While our dataset allows the study of several types of face detection models, we focus our study on feedforward neural networks which are known to yield close to human performance on challenging \"in-the-wild\" datasets [60].", + "bbox": [ + 79, + 422, + 917, + 850 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Face pareidolia in computer vision. Face detection and face recognition have been core topics in computer vision for many decades, but the study of face pareidolia—and its deep relationship with visual object representation learning—has been relatively overlooked. Face pareidolia has some similarities with the problem of cross-modal recognition or cross-depiction: recognizing the", + "bbox": [ + 79, + 850, + 917, + 947 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Seeing Faces in Things: A Model and Dataset for Pareidolia", + "bbox": [ + 249, + 14, + 837, + 34 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 900, + 16, + 917, + 30 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2c6c9ce1d33b960aefc866ea82de667f059dbd27dbdb4c0d8e0b9156c43988f6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 79, + 53, + 215, + 144 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c9440c481f248626894a3857bc3bf0ee7db197b841718070524d8008aab6f59b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 53, + 358, + 144 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ec394d7d13921bcbb2ed32ccb1ec3c17291cde68d790a96fa8c15d4fb55deccf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 360, + 53, + 496, + 144 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/cd5d70c018eb7e659accd8d19f5b9ecac68f750deabeece273d5cca2d104779e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 53, + 636, + 144 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c44a94591afc74fcf11bfec2abbc9543e14fd928dc8f0ecc397f6ba4951b73ee.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 641, + 53, + 774, + 144 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b466120961e6f4f3bf79ad33314b316fb42121b985be010676e5233f7c3b1bb4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 779, + 53, + 917, + 144 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7a91bc71b6a0fc47fc2abd7dcda24d2e70a8bfa4938b916428c04c41d8cc0517.jpg", + "image_caption": [ + "Fig. 2: Examples of face pareidolia from our \"Faces in Things\" dataset. Faces in Things consists of five thousand images annotated with bounding boxes (shown here), and facial attributes such as perceived emotion, gender, and intentionality." + ], + "image_footnote": [], + "bbox": [ + 79, + 146, + 215, + 237 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fed12bb90d7d5dc7cceb66d1c79e1e7b72a84cd2e1673a83f345cd74d3f8452a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 220, + 146, + 355, + 237 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/24d54f5511d082e81699aa5cac7fb17905cf784474aa737d64672ef0f43ddbc0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 360, + 146, + 496, + 237 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fc5bbf07dc400fbf7da6e37e7eaf3cf1f7e598befabf815b81fe9a7bec6efe6f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 498, + 146, + 636, + 237 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/aead20ea089a04b0a828202de810ef18ce8578e7bffddbbb44df0efa6b9b9a14.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 641, + 146, + 774, + 237 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/bcccada3f18eba5d5d9bca8a5b7272e56e23c81dd6af447257c2d4b9074cf87a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 779, + 146, + 917, + 237 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "same objects across different modalities irrespective of how the object is visually depicted. This has been explored particularly in the context of detecting faces, people and objects across modalities such as photography, different art movements, cartoons and sketches [4, 15, 35, 58]. The importance of capturing spatial relationships for robust cross-modal detection has also been highlighted [4]. The work of Castrejon et al. [6] showed how, when learning cross-modal scene representations with neural networks, units would emerge in the shared representation that tended to activate on consistent concepts, independently of the modality. This tendency was used by Abbas & Chalup [1], who found that mid-level units learned during human face detection could generalize to detect semantically similar facial key-points in pareidolic images, showing promise for pareidolia to emerge. However, the authors only evaluated the method qualitatively over a small test set of ten images. One route to a larger dataset may be through pareidolic image generation, which shows promise but does not yet produce convincingly natural images [12]. Curating a larger dataset, \"Totally-Looks-Like\" [39] explored the perceptual judgment of image similarity between humans and CNNs, using images which had been paired by humans as visually similar but semantically disparate. They found that visual representations extracted from CNNs such as ResNet [18] perform poorly in terms of reproducing the matching selected by humans. Though this dataset is of similar size to ours (6k samples), it is not specifically tailed towards face pareidolia and offers no bounding box or key-point annotations. Other datasets such as COCO-Periph [17] have been used to show that object detection behavior in CNNs and transformers diverge from human perception in peripheral vision.", + "bbox": [ + 72, + 344, + 920, + 813 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In summary, there has yet to be a computational model of how or why pareidolia might arise proposed in the literature. Moreover, despite the abundance of face detection datasets [36], there is no large-scale dataset to directly support the study of face pareidolia. A large-scale pareidolia dataset would help the community explore the mechanisms underlying pareidolia, which may in turn help us to understand and harness human visual attention (which is drawn towards face-like objects), reduce pareidolic false positives in face detectors, help design-", + "bbox": [ + 72, + 813, + 920, + 949 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 79, + 17, + 94, + 29 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "M. Hamilton, S. Stent, V. DuTell, et al.", + "bbox": [ + 157, + 14, + 549, + 30 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/381d138c54bb93c8a47c365609893c23b74f1943a15df88948761c921386993f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 234, + 50, + 380, + 175 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e92b075219e7351279b284eefc97d74a999f84de426d3a414611312bbe9aedb2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 435, + 50, + 581, + 177 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f2db8752b921deb79b0ebbf52724d0a97394ad614af6d8f6e2d4209ff46db6ef.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 624, + 50, + 835, + 173 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/667e0c78a2a5bc39739edcca9f8919f314703945c93c5eb2d9f84d54ec00e464.jpg", + "image_caption": [ + "Fig. 3: Attributes of the Faces in Things Dataset. We find that $31\\%$ of faces are considered challenging to spot; faces are largely $(31\\%)$ judged as happy; approximately half $(47\\%)$ are judged as accidental rather than by design; animals and humans are seen in roughly equal numbers; and we observe a slight bias $(16\\%$ vs $3\\%)$ towards male over female faces, similar to biases observed in prior studies [54,55]." + ], + "image_footnote": [], + "bbox": [ + 159, + 188, + 612, + 313 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/40f4656c5a7e936915bb292a257874ae8eff05271cf043ea0e84bb926ab8aaec.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 634, + 188, + 782, + 313 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ers avoid or create pareidolia, improve pareidolic animation, and create systems that better understand how humans perceive the world.", + "bbox": [ + 72, + 456, + 920, + 495 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Faces in Things Dataset", + "text_level": 1, + "bbox": [ + 75, + 520, + 474, + 544 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address this gap, we begin by sampling candidate pareidolic images from the LAION-5B dataset [42]. This dataset consists of 5.85 billion CLIP-filtered image-text pairs, of which $40\\%$ of captions contain English. We use CLIP retrieval [2] to build a raw image set based on text queries including \"pareidolia\", \"faces in things\", \"accidental faces\", and \"[object] looks like a face\". We download images, check for duplicates, then downsample to $512 \\times 512$ pixels while preserving the aspect ratio with white-space padding. We used the VGG Image Annotation tool [10] to manually annotate images, removing samples that contain the faces of actual humans or animals. Some examples of annotated images are shown in Fig. 2. Our annotations include the bounding boxes of pareidolic faces and basic facial attributes as summarized in Fig. 3. Though beyond the scope of the current paper, we note that these attributes could be useful for other future studies. We divide the dataset at random into training ( $70\\%$ ) and testing ( $30\\%$ ) sets. We refer to this as the 'Pareidolic' dataset.", + "bbox": [ + 72, + 556, + 922, + 829 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 75, + 853, + 314, + 878 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We use the following additional datasets. Fig. 4 shows the average faces within our dataset (Pareidolic) and the WIDER FACE (Human), and AnimalWeb (Animal) datasets.", + "bbox": [ + 75, + 889, + 922, + 949 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Seeing Faces in Things: A Model and Dataset for Pareidolia", + "bbox": [ + 246, + 14, + 840, + 34 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 900, + 16, + 917, + 30 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b355225945630caf5e4ad7588fe38a29252a3fcf514c6357df5ccd14d6e4b8fc.jpg", + "image_caption": [ + "Pareidolic" + ], + "image_footnote": [], + "bbox": [ + 244, + 52, + 411, + 200 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/10f44b727bc30bb1bf68f955c155619e737a361d6ecd5925a342806226f81cf7.jpg", + "image_caption": [ + "Human", + "Fig. 4: The Appearance of an Average Pareidolic Face. Per-channel histogram-qualified average images for registered pareidolic faces (our Faces in Things dataset), human faces (samples from the WIDER FACE dataset [60]), and animal faces (AnimalWeb [24]). The average pareidolic face, while less distinct than human or animal, has surprisingly clear eye, nose, and mouth features, and vertical symmetry." + ], + "image_footnote": [], + "bbox": [ + 416, + 53, + 581, + 201 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7dd3ed7904da2e4d6f9ab13b2c34ef64a2a0af5870e9bfc29a295a399db27a26.jpg", + "image_caption": [ + "Animal" + ], + "image_footnote": [], + "bbox": [ + 585, + 53, + 750, + 201 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "WIDER FACE [60] is a popular face detection benchmark dataset with 32,203 images and 393,703 faces. It contains a high degree of variability in scale, pose, makeup, lighting, emotion, and occlusion, organized across 61 event classes. We use the provided $40\\% / 10\\% / 50\\%$ splits for training, validation, and testing. We refer to this as the 'Human' dataset.", + "bbox": [ + 72, + 364, + 917, + 460 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "AnimalWeb [24] is a collection of 22,451 faces from 334 diverse species and 21 animal orders across biological taxonomy. These faces are captured 'in-the-wild' and are consistently annotated with 9 landmarks on key facial features. We convert these landmarks to bounding boxes, by finding the tightest box that captures the points and expanding this box's width and height by $15\\%$ . We refer to this as the 'Animal' dataset.", + "bbox": [ + 72, + 461, + 917, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "WIDER FACE Corruptions. To measure whether pareidolia could arise from common data augmentations we corrupt the WIDER FACE images using the level 3 strength of the corruptions used in both the COCO-C [34] and ImageNet-C [19] datasets. We also include a Sobel filtering corruption [13] which has been shown to reduce a model's dependence on texture information [22].", + "bbox": [ + 72, + 578, + 917, + 676 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Models and Training. We use RetinaFace [8] which achieves state-of-the-art performance on WIDER FACE easy and medium subsets and is the third-best face detector on the hard subset, missing the top model by less than a percentage point of Average Precision (AP). We perform experiments using both their MobileNet [21] and ResNet50 [18] backbones and use the Pytorch_Retinaface [3] repository to ensure the same experimental conditions, dataset characteristics, and preprocessing. We use pre-trained models provided by this repository and fine-tune them for 10 epochs with the AdamW optimizer [31] using a learning rate of $10^{-4}$ and a weight decay of $5 \\times 10^{-4}$ . We verify that fine-tuning using this strategy on the original WIDER FACE training dataset does not hurt model performance. When fine-tuning on Faces in Things (Pareidolia), Animal-Web (Animal), WIDER FACE (Human), and Corruption datasets we randomly replace images in the original WIDER FACE stream of training data with data", + "bbox": [ + 72, + 695, + 920, + 949 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 77, + 16, + 94, + 30 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "M. Hamilton, S. Stent, V. DuTell, et al.", + "bbox": [ + 157, + 14, + 549, + 32 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/81f85d3b6f3d075405bb4d35b624458249040e69a5854a209fa4a5101ea3ab53.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 79, + 53, + 244, + 162 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b6e3b2594a68761b6673c849d4c1975eb99a0542499d534b102b81e649772fd8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 246, + 53, + 411, + 162 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dba4bb603318a3e173a6c694e1a84b057af610916e425e7bd92f26313777fe15.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 416, + 53, + 581, + 162 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b991c14e2c09afd338714c7569dee4ed4023fb87f5fc9a0ae988fa7acdc42040.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 585, + 53, + 748, + 162 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5f30a7a53706be0d06c02f993d9322ef078d14d64c01fab4ad2916d28a743973.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 753, + 53, + 917, + 162 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f2e63abf3778b1c4d5365f2663025ca0a06d5639eee031bd9394a71b45e455b0.jpg", + "image_caption": [ + "Fig. 5: Qualitative Analysis of Transfer Experiments. On a sample of held-out test images, we visualize the confident $(p > 10\\%)$ detections of our ground truth (red), our model fine-tuned on human faces (blue), and our model fine-tuned on animal faces (green). It is evident from these and Table 1 that fine-tuning on animal faces significantly boosts the model's ability to detect pareidolic faces." + ], + "image_footnote": [], + "bbox": [ + 79, + 162, + 244, + 273 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/43e441d080954bf4c40ec580f72b97edf49c45ac9482bd3c390b2715b0fcfefb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 246, + 162, + 411, + 273 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3190ad5285dbdeef8ff7432865185177621797c123820dce659fa94d29773329.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 416, + 162, + 581, + 273 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/87aa715e559cbb15179ab6405864f508fc833193c7f0fde8e05a9a4de33a01a5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 585, + 162, + 748, + 273 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1daac9a1541026579175baac1a136e0c0741356d8c58c2c9e31b0847635024ad.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 753, + 162, + 917, + 273 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "from the target dataset $90\\%$ of the time. This allows the network to learn the new task without catastrophic forgetting [25]. These changes to the optimizer, learning rate, number of epochs, and stream of training data are the only changes we make to the training paradigm of [8]. Figures in this work use the MobileNet architecture of RetinaNet unless specified otherwise. AP evaluation computations share the same setting and parameters as [8].", + "bbox": [ + 72, + 419, + 917, + 536 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1 Does a SOTA Face Detector Exhibit Pareidolia?", + "text_level": 1, + "bbox": [ + 75, + 567, + 731, + 588 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We measure the Average Precision (AP) of the MobileNet and ResNet50 RetinaNet architectures on the Faces in Things dataset. The first row of Table 1 shows results for existing pre-trained models, and the second row shows those for models fine-tuned on the original WIDER FACE training data. These act as control groups to ensure our transfer learning procedure does not interfere with our measurement of the effects of other interventions. Though these models exhibit pareidolia to a small extent, they fall far short of a model fine-tuned to detect pareidolic faces. Fig. 5 also depicts some of these predictions with blue boxes. On the whole, the models trained only on human faces are largely silent across the Faces in Things dataset.", + "bbox": [ + 72, + 604, + 920, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2 How Might Pareidolia Emerge?", + "text_level": 1, + "bbox": [ + 75, + 834, + 527, + 855 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The WIDER FACE dataset is known for its diversity of lighting, pose, makeup, emotion, and scale of faces. This fact, coupled with the results of Section 4.1 begs the question: What else do models need to experience pareidolia as humans do? The Faces in Things dataset provides a clean and robust setting to explore", + "bbox": [ + 72, + 869, + 917, + 949 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Seeing Faces in Things: A Model and Dataset for Pareidolia", + "bbox": [ + 246, + 14, + 840, + 34 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 903, + 16, + 920, + 30 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3d5c50beae2e3d1a197a94b6d4c0bfc70531368543c1922368e170a4649826ae.jpg", + "image_caption": [ + "Fig. 6: Measuring the effect of several training interventions on pareidolic face detection The left plot shows that fine-tuning RetinaNet on animals improves pareidolic face detection more than any other intervention. Conversely, the right plot shows that pareidolic fine-tuning improves animal face detection performance." + ], + "image_footnote": [], + "bbox": [ + 84, + 45, + 489, + 248 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a364ae7066474aaa2c442a6555a8edad35f41f1e3f6b43191897d9b923c9118c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 498, + 45, + 910, + 248 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the development of pareidolia in algorithms. Unlike in humans, where it is impossible to causally intervene on their facial training data, we can easily modify an algorithm's training data. This makes it possible to explore whether one can induce pareidolia in algorithms using specific stimuli.", + "bbox": [ + 72, + 393, + 917, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To this end, we investigate whether a variety of training data interventions can induce pareidolia in algorithms. In particular, we measure the effect of adding several data augmentations from the COCO-C [34] and ImageNet-C [19] datasets and explore a Sobel filtering augmentation which reduces models' dependence on texture. Additionally, we also measure the effect of adding animal faces to the training data. Animal faces show a far greater breadth of variation in coloration, structure, and appearance than human faces. Recognizing animal faces provides many evolutionary advantages including gaze detection during hunting and avoiding onlooking predators. The generality required to detect this wide space of faces could yield a greater number of \"false positives\" that lead to the sensation of pareidolia. Indeed, some recent studies provide some corroborating evidence for this hypothesis. Firstly, Rhesus Monkeys exhibit pareidolia [49] showing this effect does not only occur in humans. Secondly, the experience of pareidolia is a rapid cognitive process and not a \"late re-interpretation\" of input signals [16, 56] which the authors conclude is evidence that pareidolia could be linked to the need to quickly react to predators.", + "bbox": [ + 72, + 476, + 920, + 788 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We plot the change to pareidolic face detection performance as a function of each training intervention in the left panel of Figure 6. Of the different corruption interventions, we find that Sobel filtering, motion blur, Gaussian noise, and fog tend to slightly improve pareidolic face detection while most other corruptions do not improve pareidolic face detection performance. Most strikingly, the addition of animal faces to the training data roughly doubles the algorithms' ability to detect pareidolic faces compared to the control group, closing around half of the gap between a human-trained model and a pareidolia trained model.", + "bbox": [ + 72, + 793, + 920, + 949 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 79, + 17, + 94, + 29 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "M. Hamilton, S. Stent, V. DuTell, et al.", + "bbox": [ + 157, + 14, + 549, + 32 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1e4b8f123970d85c99c74bcc70d8e16b82737453ed47dba71648a20a9f8e6e7f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
FinetuningAP
MobileNetResNet50
None7.9%2.8%
Human (Control)9.8%3.6%
Animal16.7%15.4%
Pareidolia33.9%27.1%
Animal + Pareidolia36.4%31.7%
", + "bbox": [ + 283, + 50, + 707, + 200 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 1: Effect of Fine Tuning on Pareidolic Face Detection. Our results show that WIDER FACE-trained RetinaFace models do not detect many pareidolic images. Fine-tuning these models on animal faces approximately doubles pareidolic face detection rates. Interestingly, adding animal faces alongside pareidolic faces (30%/70% split respectively) can improve performance over fine-tuning on pareidolic faces alone.", + "bbox": [ + 77, + 204, + 917, + 295 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Reciprocally, the right-hand plot of Figure 6 shows that fine-tuning on pareidolic images yields the greatest improvement in animal face detection. We further explore this phenomenon in Table 1, where we show that this effect occurs across both MobileNet and ResNet50 architectures. Finally, we also show that adding a small number of animal faces (30% animal 70% pareidolic) can improve pareidolic face detection performance over pareidolic images alone.", + "bbox": [ + 77, + 326, + 917, + 443 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To understand this effect better, Fig. 7 visualizes the inner representations of this model across the three datasets (Human = WIDER FACE, Animal = AnimalWeb, Pareidolic = Faces in Things). Specifically, we extract multi-scale features from the animal and pareidolia fine-tuned RetinaNet shared feature layer before the application of the classification and regression heads. We average pool these features across the bounding box for each face and visualize them with t-SNE [20]. This figure shows that RetinaNet's representations of animal and pareidolic faces tend to cluster together and are distinct from its representations of human faces. This lends evidence to the relative similarity of pareidolic and animal faces compared to human faces. We also reiterate that we filtered the Faces in Things dataset to avoid images of real animals.", + "bbox": [ + 77, + 445, + 917, + 656 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 Modeling Pareidolia", + "text_level": 1, + "bbox": [ + 79, + 682, + 414, + 705 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Though many prior works have measured pareidolia, there has yet to be a simple mathematical model that describes the high-level structure of this phenomenon. In this section we provide two simple formal models of pareidolia and show that they both exhibit a testable prediction: the existence of a peak in pareidolic face detection as a function of an image's complexity. Section 5.4 presents experimental evidence of this \"pareidolic peak\" in both humans and machines.", + "bbox": [ + 79, + 720, + 917, + 837 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.1 Gaussian Model of Pareidolia", + "text_level": 1, + "bbox": [ + 79, + 861, + 501, + 879 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "A model of pareidolia needs to describe two processes: (1) the random process that generates candidate images, and (2) the face detection process which determines when an image is pareidolic. We begin with a simple Gaussian model for", + "bbox": [ + 79, + 889, + 917, + 947 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Seeing Faces in Things: A Model and Dataset for Pareidolia", + "bbox": [ + 249, + 14, + 837, + 34 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 900, + 16, + 917, + 30 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/3ed09d02fe6255fd7f0a9ace6c40ba6d1a19d1d8f20ab1df02d7d8862da99be6.jpg", + "image_caption": [ + "Fig. 7: Visualizing RetinaNet Representations across Datasets. Animal+Pareidolia fine-tuned RetinaNet representations tend to group animal and pareidolic faces together. This lends evidence to the hypothesis that the perception of animal and pareidolic faces are linked. (To highlight the commonality of pareidolic animal detection we note the similarity of these points to a frog.)" + ], + "image_footnote": [], + "bbox": [ + 254, + 58, + 736, + 300 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "each. We model the image generation process as a sum of independent normal modes, each contributing a zero-mean Gaussian of a specified variance, multiplied by the mode image $y_{i}$ . For example, as in [9, 51], these modes could be the principal components of a mean-subtracted image dataset. In this setting the generated image, $y$ , is a weighted sum of the normal modes:", + "bbox": [ + 72, + 448, + 917, + 547 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {y} = \\sum_ {i} n _ {i} \\boldsymbol {y} _ {i} \\quad \\text {w h e r e}, \\quad n _ {i} \\sim N (0, \\sigma_ {i}) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 290, + 570, + 917, + 611 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To model our face detection process we capture the intuition of matching an image to a template image and note that this can be generalized to distributions of template images. In particular, the target pareidolic image is represented as a vector, $\\pmb{a}$ , of statistically independent target coefficients, $a_{i}$ , for each mode. The probability that this mode contributes towards the face detection, $P(a_{i})$ , is the probability of detecting the pareidolic value, $a_{i}$ , at the $i$ th mode. We assume a Gaussian detection process: $P(a_{i}|y_{i}) \\sim N(a_{i}, \\gamma_{i})$ . Because each mode's coefficient is a zero-mean Gaussian distribution, $P(y_{i}) \\sim N(0, \\sigma)$ , we have:", + "bbox": [ + 72, + 634, + 920, + 793 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} P \\left(a _ {i}\\right) = \\int_ {y _ {i}} P \\left(a _ {i}, y _ {i}\\right) \\mathrm {d} y _ {i} (2) \\\\ = \\int_ {y _ {i}} P \\left(a _ {i} \\mid y _ {i}\\right) P \\left(y _ {i}\\right) \\mathrm {d} y _ {i} (3) \\\\ = \\frac {1}{2 \\pi \\gamma_ {i} \\sigma_ {i}} \\int_ {y _ {i}} e ^ {- \\frac {(y _ {i} - a _ {i}) ^ {2}}{2 \\gamma_ {i} ^ {2}}} e ^ {- \\frac {y _ {i} ^ {2}}{2 \\sigma^ {2}}} \\mathrm {d} y _ {i} (4) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 292, + 814, + 917, + 954 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 79, + 16, + 104, + 30 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "M. Hamilton, S. Stent, V. DuTell, et al.", + "bbox": [ + 157, + 14, + 549, + 32 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/899529a899e718c0b32f4bf85ae3787bee3b1aa6decae8eafedf09ba9775a7c4.jpg", + "image_caption": [ + "Fig. 8: Illustration of the proposed Gaussian model for pareidolia with three example generating distributions. To make pareidolia likely, the generating distribution needs a proper distribution of spatial frequencies. A process with too few spatial frequencies (left) is likely to only generate weak face-like details (\"face-ness\": low). In contrast, with too many frequencies (right), faces can be modeled with exquisite detail (\"face-ness\": high), but the likelihood of drawing any particular desired combination become vanishingly small. The most likely pareidolic images form when the generating distribution has the right spectrum (middle), enabling reasonable faces to emerge with reasonable likelihood. In other words, this model predicts that pareidolic faces will match the low frequencies of faces but differ in the higher frequency details." + ], + "image_footnote": [], + "bbox": [ + 164, + 43, + 832, + 362 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Note that $\\sigma_i^2$ is the variance of the random process generating the pareidolia, while $\\gamma_i^2$ is the variance of the likelihood term — how far a mode is allowed to vary from the target mode value before it stops looking like the target image, $\\pmb{a}$ .", + "bbox": [ + 75, + 588, + 917, + 647 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We can complete the square in Eq. 4 to write the product of Gaussians in $P(a_{i})$ as a single Gaussian. Integrating that Gaussian over all possible observations $y_{i}$ gives the probability of finding the pareidolic value $a_{i}$ from mode $i$ :", + "bbox": [ + 75, + 647, + 917, + 705 + ], + "page_idx": 10 + }, + { + "type": "equation", + "text": "\n$$\nP (a _ {i}) = \\frac {1}{\\sqrt {2 \\pi (\\gamma_ {i} ^ {2} + \\sigma_ {i} ^ {2})}} e ^ {- \\frac {a _ {i} ^ {2}}{2 (\\sigma_ {i} ^ {2} + \\gamma_ {i} ^ {2})}} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 317, + 721, + 917, + 773 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.2 Predicting Peak Pareidolia", + "text_level": 1, + "bbox": [ + 75, + 814, + 474, + 834 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "For a given mode's detection variance, $\\gamma_i^2$ , and target mode coefficient, $a_i$ , Eq. 4 allows us to find the optimal mode variance to generate pareidolia, i.e., to maximize $P(a_i)$ . Unfortunately, we seldom have the flexibility to design a random process one mode at a time. But we may have the option to select between image generation processes that have different numbers of modes, $M$ . Since each mode", + "bbox": [ + 72, + 850, + 920, + 949 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Seeing Faces in Things: A Model and Dataset for Pareidolia", + "bbox": [ + 246, + 14, + 840, + 34 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 891, + 16, + 915, + 30 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/3278e6b04f301679f9177bab9716292e2ccee0e6299185ecbe679e8961c01c25.jpg", + "image_caption": [ + "Fig. 9: Probability of pareidolia (Eq. 6) in the Gaussian model $(\\sigma = 10)$ across images with different spatial frequency distributions (Sec. 5.4). This assumes spatial frequencies are uncorrelated and thus underestimates the probability of pareidolia, however peak pareidolia is still present." + ], + "image_footnote": [], + "bbox": [ + 84, + 50, + 479, + 227 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/56d8b68dea877bd85cd69a98be77bd3e6d8a994945f7650a41100c8e0bcec897.jpg", + "image_caption": [ + "Fig. 10: Probability of pareidolia under the feature-based example of Eq. (8) as a function of the rate of feature detection, $\\lambda_{i} = \\lambda$ for all $i$ , within the random images. Note the low probability of pareidolia for both feature-free $(\\lambda \\to 0)$ and feature-rich $(\\lambda \\gg 0)$ random images." + ], + "image_footnote": [], + "bbox": [ + 520, + 50, + 910, + 203 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "is independent, the probability of a pareidolic detection of the target object template is the product of detecting the target coefficient for each of the M modes:", + "bbox": [ + 72, + 401, + 917, + 440 + ], + "page_idx": 11 + }, + { + "type": "equation", + "text": "\n$$\nP (a) = \\prod_ {i} ^ {M} P \\left(a _ {i}\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 404, + 458, + 917, + 510 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We plot some predictions of our Gaussian model in Fig. 9 for a target template with a $\\frac{1}{f}$ power spectrum (standard deviation of each mode inversely proportional to mode number) on noise images of varying complexity. We note the existence of a peak in pareidolic detection probability for random image generation processes with a mid-range number of spatial modes, as measured by the width of Gaussian that modulates power in Fourier space. Too few modes in the random generation process, and no image will ever have enough complexity to render the target well. Too many modes and pareidolia becomes unlikely because so many modes need to match the desired target values. Each added mode multiplies the pareidolia probability by another small factor. In between, there is what we call peak pareidolia. As the detection value, $\\sigma$ , becomes more stringent (smaller) the peak pareidolia value occurs at a larger number of modes and becomes less probable. We illustrate this effect in Figure 8", + "bbox": [ + 72, + 521, + 917, + 778 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.3 Higher-Level Feature Model of Pareidolia", + "text_level": 1, + "bbox": [ + 75, + 813, + 648, + 832 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The Gaussian model for pareidolia above lays out important aspects of pareidolia, but relies on a naive model of object detection, the squared distance from a template image. We assume that a more realistic model of human perception would incorporate higher-level features and introduce a still simple, yet more realistic, feature-based model.", + "bbox": [ + 72, + 850, + 917, + 947 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 79, + 16, + 104, + 30 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "M. Hamilton, S. Stent, V. DuTell, et al.", + "bbox": [ + 157, + 14, + 549, + 30 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We assume that the detection of an object requires particular features to be detected in certain spatial regions, e.g. an eye in the top left and right, a nose in the center, and a mouth in the bottom. Such an approach has been used in computer vision object detection algorithms, e.g. [14, 57, 61]. Any given object template has some number of regions, $R_{i}$ , indexed by $i$ , within which a given feature, $F_{i}$ , must be detected. The other features, $F_{j\\neq i}$ , should not be detected in region $i$ . For a given random image where we hope to detect pareidolia, we assume that feature existence is a spatial Poisson process. In this process, the probability of $n$ feature instances for any given feature $i$ over some area $B_{i}$ is", + "bbox": [ + 72, + 56, + 922, + 232 + ], + "page_idx": 12 + }, + { + "type": "equation", + "text": "\n$$\nP (n _ {i}) = \\frac {\\left(\\lambda_ {i} | B _ {i} |\\right) ^ {n _ {i}}}{n _ {i} !} e ^ {- \\lambda_ {i} | B _ {i} |} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 358, + 245, + 917, + 287 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "To detect a pareidolic instance of the object template, we must detect one feature of the correct type, $F_{i}$ , in each region $i$ of the face template, and zero features of the wrong type, $F_{j\\neq i}$ in each region $i$ . Assuming independence of the feature detections, and for simplicity setting all the feature detection rates to be the same, $\\lambda_{i} = \\lambda$ , and all the template areas to be the same, $B_{i} = 1$ , we have for the probability, $P(O)$ , of pareidolic detection of object $O$ :", + "bbox": [ + 72, + 295, + 922, + 413 + ], + "page_idx": 12 + }, + { + "type": "equation", + "text": "\n$$\nP (O) = \\prod_ {i} ^ {M} \\lambda^ {1} e ^ {- \\lambda (M - 1)} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 429, + 917, + 479 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "For the case of $M = 4$ , a simple detection model for two eyes, a nose, and a mouth, we have $P(o) = \\lambda^4 e^{-16\\lambda}$ , which is plotted in Fig. 10. In this feature-based object detection model, we also find the existence of \"peak pareidolia\". Again, it is governed by a parameter describing the complexity of the random image, in this case a Poisson process rate parameter, $\\lambda$ , that governs the probability of a feature detection per unit area. For too low a rate, the model doesn't generate enough features to satisfy the object template, for too high a rate, the probability of seeing only the right features in just the right places becomes very small. In between is the most probable rate for pareidolia.", + "bbox": [ + 72, + 495, + 922, + 671 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.4 Measuring the Pareidolic Peak in Humans and Machines", + "text_level": 1, + "bbox": [ + 75, + 700, + 837, + 721 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Both mathematical models of Section 5 predict the existence of a peak of pareidolic face detection as a function of image complexity. We show the existence of this pareidolic peak in both humans and machines. In particular, we perform a psychophysics experiment where human subjects view noise images of varying complexity and report how many pareidolic faces they saw in each image, from zero to nine. Campbell [5] demonstrated that a 12x12 array of random, binary squares is sufficient to evoke human and animal faces. We generate noise images of varying complexity by randomly sampling Fourier coefficients and modulating these coefficients with a zero mean $\\sigma^2$ variance Gaussian in Fourier space. We show some samples of these images on the x-axis of Fig. 11. Intuitively, the Gaussian envelope in frequency space filters out most frequencies higher than $\\sigma$", + "bbox": [ + 72, + 734, + 922, + 949 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Seeing Faces in Things: A Model and Dataset for Pareidolia", + "bbox": [ + 246, + 14, + 837, + 34 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 891, + 16, + 917, + 30 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/5c27903441e7a250a8d902412ab58a269b2f1c382c1d60fca216919c893ba5e7.jpg", + "image_caption": [ + "Fig. 11: Measuring Peak Pareidolia. Left: Subjects were asked how many faces they see in each noise image. We plot the average number of faces detected as a function of noise frequency (examples on x-axis), the mean over all subjects and its $95\\%$ confidence interval in red. Right: average number of faces detected by our fine-tuned models. This reveals the \"peak pareidolia\" effect predicted in Section 5 across humans and machines." + ], + "image_footnote": [], + "bbox": [ + 79, + 40, + 489, + 229 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/fa61f090fc589435100f26d5052696fcc4096df385ab7978f92b8a49e1628b2f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 498, + 40, + 912, + 229 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "after applying an inverse Fourier transform. We detail our image stimuli creation method in the Supplement.", + "bbox": [ + 72, + 372, + 917, + 411 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We find that humans exhibit the model-expected peak pareidolia, with a maximum number of faces detected at a frequency filter width of 16 (Fig. 11, left). The existence of a pareidolic peak at or near this filter width is consistent among all subjects even for those that reported fewer faces overall. Although response time did decrease slightly at higher frequencies, it did not fall off completely at the highest frequency levels, indicating that fewer reported faces were not the result of subjects \"giving up\" on the task. We provide additional details and analysis of this experiment in the Supplement.", + "bbox": [ + 72, + 414, + 917, + 569 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Finally, we evaluate our fine-tuned models from Section 4 on the same images to test whether machines also exhibit peak pareidolia (Fig. 11, right). In particular, we showed the models 5,000 sampled noise images of varying frequency levels and counted the number of face detections they make with confidence $>10\\%$ . We find the same characteristic \"pareidolic peak\" where models detect the most faces in medium-complexity images.", + "bbox": [ + 72, + 573, + 917, + 689 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 75, + 736, + 285, + 756 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We have taken initial steps towards the mathematical modelling of pareidolia and build a richly annotated dataset of images for face pareidolia. We showed through experiments on modern face detectors that detecting animal faces may partly explain the emergence of pareidolia in a complex vision system. The Faces in Things dataset can help the community address other questions about how and why pareidolic behavior emerges, a hallmark of humans' robust recognition system. We hope that our findings and dataset will spark further study of pareidolia and its potential use to improve computer vision systems.", + "bbox": [ + 72, + 791, + 920, + 949 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 79, + 17, + 104, + 30 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "M. Hamilton, S. Stent, V. DuTell, et al.", + "bbox": [ + 157, + 14, + 549, + 32 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 79, + 56, + 297, + 74 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "We would like to thank Karen Hamilton for her hundreds of hours of annotations for the Faces in Things dataset. We also thank Abhishek Dutta who created the VGG Image Annotation tool, for his kind generosity to support our project.", + "bbox": [ + 79, + 82, + 917, + 139 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "We would like to thank the Microsoft Research Grand Central Resources team for their gracious help performing the experiments in this work. Special thanks to Oleg Losinets and Lifeng Li for their consistent, gracious, and timely help, debugging, and expertise. Without them, none of the experiments could have been run.", + "bbox": [ + 79, + 141, + 917, + 235 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This material is based upon work supported by the National Science Foundation Graduate Research Fellowship under Grant No. 2021323067. Any opinion, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of their employers, or the National Science Foundation.", + "bbox": [ + 79, + 237, + 917, + 333 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work is supported by the National Science Foundation under Cooperative Agreement PHY-2019786 (The NSF AI Institute for Artificial Intelligence and Fundamental Interactions, http://iaifi.org/) and the CSAIL MEnTorEd Opportunities in Research (METEOR) Fellowship.", + "bbox": [ + 79, + 334, + 917, + 411 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Research was sponsored by the United States Air Force Research Laboratory and the United States Air Force Artificial Intelligence Accelerator and was accomplished under Cooperative Agreement Number FA8750-192-1000. The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the United States Air Force or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for Government purposes notwithstanding any copyright notation herein.", + "bbox": [ + 79, + 414, + 917, + 567 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "The authors acknowledge the MIT SuperCloud [38] and Lincoln Laboratory Supercomputing Center for providing HPC resources that have contributed to the research results reported within this paper.", + "bbox": [ + 79, + 567, + 917, + 624 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 79, + 652, + 234, + 671 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Abbas, A., Chalup, S.: From Face Recognition to Facial Pareidolia: Analysing Hidden Neuron Activations in CNNs for Cross-Depiction Recognition. In: 2019 International Joint Conference on Neural Networks (IJCNN). pp. 1–8 (Jul 2019). https://doi.org/10.1109/IJCNN.2019.8852013, iSSN: 2161-4393", + "2. Beaumont, R.: Clip retrieval: Easily compute clip embeddings and build a clip retrieval system with them. https://github.com/rom1504/clip-retrieval (2022)", + "3. biubug6: Retinaface in pytorch (Nov 2021), https://github.com/biubug6/Pytorch_Retinaface", + "4. Cai, H., Wu, Q., Corradi, T., Hall, P.: The Cross-Depiction Problem: Computer Vision Algorithms for Recognising Objects in Artwork and in Photographs. arXiv:1505.00110 [cs] (May 2015), http://arxiv.org/abs/1505.00110, arXiv: 1505.00110", + "5. Campbell, F.: How much of the information falling on the retina reaches the visual cortex and how much is stored in the visual memory. Pattern recognition mechanisms 54, 83-94 (1983)" + ], + "bbox": [ + 87, + 684, + 917, + 947 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Seeing Faces in Things: A Model and Dataset for Pareidolia", + "bbox": [ + 246, + 14, + 837, + 34 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 891, + 16, + 917, + 30 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "6. Castrejon, L., Aytar, Y., Vondrick, C., Pirsiavash, H., Torralba, A.: Learning Aligned Cross-Modal Representations from Weakly Aligned Data. In: CVPR (Jun 2016)", + "7. Conrad, K.: Die beginnende Schizophrenie. Versuch einer Gestaltanalyse des Wahns (1958)", + "8. Deng, J., Guo, J., Zhou, Y., Yu, J., Kotsia, I., Zafeiriou, S.: RetinaFace: Single-stage Dense Face Localisation in the Wild. CVPR (2020)", + "9. Duda, R., Hart, P., Stork, D.: Pattern Classification. Wiley (2012)", + "10. Dutta, A., Zisserman, A.: The VIA annotation software for images, audio and video. In: Proceedings of the 27th ACM International Conference on Multimedia. MM '19, ACM, New York, NY, USA (2019). https://doi.org/10.1145/3343031.3350535, https://doi.org/10.1145/3343031.3350535", + "11. Egger, B., Siegel, M.H., Arora, R., Soltani, A.A., Yildirim, I., Tenenbaum, J.: Inverse rendering best explains face perception under extreme illuminations. In: CogSci (2020)", + "12. Endo, Y., Asanuma, R., Shimojo, S., Akashi, T.: Systematic face pareidolia generation method using cycle-consistent adversarial networks. IEEE Transactions on Electrical and Electronic Engineering 19(4), 535-541 (2024)", + "13. Farid, H., Simoncelli, E.P.: Optimally rotation-equivariant directional derivative kernels. In: International Conference on Computer Analysis of Images and Patterns. pp. 207-214. Springer (1997)", + "14. Felzenszwalb, P., Girshick, R., McAllester, D., Ramanan, D.: Object detection with discriminatively trained part based models. IEEE TPAMI 32(9) (2010)", + "15. Ginosar, S., Haas, D., Brown, T., Malik, J.: Detecting people in cubist art. In: ECCV Workshops (2014)", + "16. Hadjikhani, N., Kveraga, K., Naik, P., Ahlfors, S.P.: Early (n170) activation of face-specific cortex by face-like objects. Neuroreport 20(4), 403 (2009)", + "17. Harrington, A., DuTell, V., Hamilton, M., Tewari, A., Stent, S., Freeman, W.T., Rosenholtz, R.: Coco-periph: Bridging the gap between human and machine perception in the periphery. In: The Twelfth International Conference on Learning Representations", + "18. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016)", + "19. Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. Proceedings of the International Conference on Learning Representations (2019)", + "20. Hinton, G.E., Roweis, S.: Stochastic neighbor embedding. Advances in neural information processing systems 15 (2002)", + "21. Howard, A.G., Zhu, M., Chen, B., Kalenichenko, D., Wang, W., Weyand, T., Andreetto, M., Adam, H.: Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861 (2017)", + "22. Ji, X., Henriques, J.F., Vedaldi, A.: Invariant information clustering for unsupervised image classification and segmentation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9865-9874 (2019)", + "23. Kato, M., Mugitani, R.: Pareidolia in Infants. PLoS ONE 10(2) (Feb 2015). https://doi.org/10.1371/journal.pone.0118539, https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4331561/", + "24. Khan, M.H., McDonagh, J., Khan, S., Shahabuddin, M., Arora, A., Khan, F.S., Shao, L., Tzimiropoulos, G.: AnimalWeb: A large-scale hierarchical dataset of annotated animal faces. In: CVPR (2020)" + ], + "bbox": [ + 79, + 56, + 917, + 947 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 79, + 16, + 104, + 30 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "M. Hamilton, S. Stent, V. DuTell, et al.", + "bbox": [ + 157, + 14, + 552, + 32 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "25. Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A.A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., et al.: Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences 114(13), 3521-3526 (2017)", + "26. Leopold, D.A., Rhodes, G.: A comparative view of face perception. Journal of Comparative Psychology 124(3), 233 (2010)", + "27. Li, H., Lin, Z., Shen, X., Brandt, J., Hua, G.: A convolutional neural network cascade for face detection. In: CVPR (2015)", + "28. Li, J., Wang, Y., Wang, C., Tai, Y., Qian, J., Yang, J., Wang, C., Li, J., Huang, F.: DSFD: dual shot face detector. In: CVPR (2019)", + "29. Liao, S., Jain, A.K., Li, S.Z.: A Fast and Accurate Unconstrained Face Detector. PAMI 38(2), 211-223 (Feb 2016). https://doi.org/10.1109/TPAMI.2015.2448075", + "30. Liu, J., Li, J., Feng, L., Li, L., Tian, J., Lee, K.: Seeing Jesus in toast: Neural and behavioral correlates of face pareidolia. Cortex 53, 60-77 (Apr 2014). https://doi.org/10.1016/j.cortex.2014.01.013, http://www.sciencedirect.com/science/article/pii/S0010945214000288", + "31. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)", + "32. Mathias, M., Benenson, R., Pedersoli, M., Van Gool, L.: Face detection without bells and whistles. In: ECCV (2014)", + "33. Mcgugin, R., Gatenby, C., Gore, J., Gauthier, I.: High-resolution imaging of expertise reveals reliable object selectivity in the FFA related to perceptual performance. Proceedings of the National Academy of Sciences of the United States of America 109, 17063-8 (Oct 2012). https://doi.org/10.1073/pnas.1116333109", + "34. Michaelis, C., Mitzkus, B., Geirhos, R., Rusak, E., Bringmann, O., Ecker, A.S., Bethge, M., Brendel, W.: Benchmarking robustness in object detection: Autonomous driving when winter is coming. arXiv preprint arXiv:1907.07484 (2019)", + "35. Mishra, A., Nandan Rai, S., Mishra, A., Jawahar, C.V.: Iit-cfw: A benchmark database of cartoon faces in the wild. In: VASE ECCV Workshops (2016)", + "36. Ranjan, R., Sankar, S., Bansal, A., Bodla, N., Chen, J.C., Patel, V., Castillo, C., Chellappa, R.: Deep Learning for Understanding Faces: Machines May Be Just as Good, or Better, than Humans. IEEE Signal Processing Magazine 35, 66-83 (Jan 2018). https://doi.org/10.1109/MSP.2017.2764116", + "37. Rein-Lien Hsu, Abdel-Mottaleb, M., Jain, A.: Face detection in color images. PAMI 24(5), 696-706 (May 2002). https://doi.org/10.1109/34.1000242, http://ieeexplore.ieee.org/document/1000242/", + "38. Reuther, A., Kepner, J., Byun, C., Samsi, S., Arcand, W., Bestor, D., Bergeron, B., Gadepally, V., Houle, M., Hubbell, M., Jones, M., Klein, A., Milechin, L., Mullen, J., Prout, A., Rosa, A., Yee, C., Michaleas, P.: Interactive supercomputing on 40,000 cores for machine learning and data analysis. In: 2018 IEEE High Performance extreme Computing Conference (HPEC). pp. 1-6. IEEE (2018)", + "39. Rosenfeld, A., Solbach, M.D., Tsotsos, J.K.: Totally looks like-how humans compare, compared to machines. In: ACCV (2018)", + "40. Rowley, H.A., Baluja, S., Kanade, T.: Human Face Detection in Visual Scenes. In: Touretzky, D.S., Mozer, M.C., Hasselmo, M.E. (eds.) NeurIPS (1996), http://papers.nips.cc/paper/1168-human-face-detection-in-visual-scenes.pdf", + "41. Rowley, H.A., Baluja, S., Kanade, T.: Neural network-based face detection. PAMI 20(1), 23-38 (1998)" + ], + "bbox": [ + 75, + 56, + 917, + 947 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Seeing Faces in Things: A Model and Dataset for Pareidolia", + "bbox": [ + 246, + 14, + 840, + 34 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 891, + 16, + 917, + 30 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "42. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C.W., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., Schramowski, P., Kundurthy, S.R., Crowson, K., Schmidt, L., Kaczmarczyk, R., Jitsev, J.: LAION5b: An open large-scale dataset for training next generation image-text models. In: NeurIPS Datasets and Benchmarks Track (2022), https://openreview.net/forum?id=M3Y74vmsMcY", + "43. Shakespeare, W.: The Tragedy of Hamlet, Prince of Denmark. The Folio Society (1954)", + "44. Sims, A.: Symptoms in the mind: An introduction to descriptive psychopathology. Bailliere Tindall Publishers (1988)", + "45. Summerfield, C., Egner, T., Mangels, J., Hirsch, J.: Mistaking a house for a face: neural correlates of misperception in healthy humans. Cerebral cortex 16(4), 500-508 (2006)", + "46. Takahashi, K., Watanabe, K.: Gaze Cueing by Pareidolia Faces. i-Perception 4(8), 490-492 (Dec 2013). https://doi.org/10.1068/i0617sas, https://doi.org/10.1068/i0617sas", + "47. Takahashi, K., Watanabe, K.: Seeing Objects as Faces Enhances Object Detection. i-Perception 6(5) (2015). https://doi.org/10.1177/2041669515606007, https://doi.org/10.1177/2041669515606007", + "48. Tang, X., Du, D.K., He, Z., Liu, J.: PyramidBox: A Context-Assisted Single Shot Face Detector. In: Ferrari, V., Hebert, M., Sminchisescu, C., Weiss, Y. (eds.) ECCV (2018)", + "49. Taubert, J., Wardle, S.G., Flessert, M., Leopold, D.A., Ungerleider, L.G.: Face Pareidolia in the Rhesus Monkey. Current Biology 27(16), 2505-2509 (Aug 2017). https://doi.org/10.1016/j.cub.2017.06.075, http://www.sciencedirect.com/science/article/pii/S0960982217308126", + "50. Tsao, D.Y., Livingstone, M.S.: Mechanisms of face perception. Annu. Rev. Neurosci. 31, 411-437 (2008)", + "51. Turk, M., Pentland, A.: Eigenfaces for recognition. J. of Cognitive Neuroscience 3(1) (1991)", + "52. Viola, P., Jones, M.: Rapid object detection using a boosted cascade of simple features. In: CVPR (2001)", + "53. Viola, P., Jones, M.J.: Robust real-time face detection. IJCV 57(2), 137-154 (2004)", + "54. Wardle, S.G., Ewing, L., Malcolm, G.L., Paranjape, S., Baker, C.I.: Children perceive illusory faces in objects as male more often than female. Cognition 235 (2023). https://doi.org/https://doi.org/10.1016/j.cognition.2023.105398, https://www.sciencedirect.com/science/article/pii/S001002772300032X", + "55. Wardle, S.G., Paranjape, S., Taubert, J., Baker, C.I.: Illusory faces are more likely to be perceived as male than female. Proceedings of the National Academy of Sciences 119(5) (2022). https://doi.org/10.1073/pnas.2117413119, https://www.pnas.org/doi/abs/10.1073/pnas.2117413119", + "56. Wardle, S.G., Taubert, J., Teichmann, L., Baker, C.I.: Rapid and dynamic processing of face pareidolia in the human brain. Nature communications 11(1), 4518 (2020)", + "57. Weber, M., Welling, M., Perona, P.: Unsupervised learning of models for recognition. In: ECCV (2000)", + "58. Westlake, N., Cai, H., Hall, P.: Detecting People in Artwork with CNNs. In: ECCV Workshops (2016)", + "59. Yang, M.H., Kriegman, D., Ahuja, N.: Detecting faces in images: a survey. PAMI 24(1), 34-58 (Jan 2002). https://doi.org/10.1109/34.982883" + ], + "bbox": [ + 79, + 56, + 917, + 947 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 79, + 16, + 104, + 30 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "M. Hamilton, S. Stent, V. DuTell, et al.", + "bbox": [ + 157, + 14, + 549, + 32 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "60. Yang, S., Luo, P., Loy, C.C., Tang, X.: WIDER FACE: A Face Detection Benchmark. In: CVPR (2016)", + "61. Yuille, A.L.: Deformable templates for face recognition. Journal of Cognitive Neuroscience 3(1), 59-70 (1991)", + "62. Zafeiriou, S., Zhang, C., Zhang, Z.: A survey on face detection in the wild: Past, present and future. Computer Vision and Image Understanding 138, 1-24 (Sep 2015). https://doi.org/10.1016/j.cviu.2015.03.015, https://linkinghub.elsevier.com/retrieve/pii/S1077314215000727", + "63. Zhang, S., Zhu, X., Lei, Z., Shi, H., Wang, X., Li, S.Z.: S3fd: Single shot scale-invariant face detector. In: ICCV (2017)" + ], + "bbox": [ + 75, + 56, + 920, + 235 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "Seeing Faces in Things: A Model and Dataset for Pareidolia", + "bbox": [ + 246, + 14, + 837, + 34 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 891, + 16, + 917, + 30 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/Seeing Faces in Things_ A Model and Dataset for Pareidolia/78844de0-7820-498e-be36-c38ded4cc8e3_model.json b/2024/Seeing Faces in Things_ A Model and Dataset for Pareidolia/78844de0-7820-498e-be36-c38ded4cc8e3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4876329b60be78c4935cedce277a1a2f0c596a19 --- /dev/null +++ b/2024/Seeing Faces in Things_ A Model and Dataset for Pareidolia/78844de0-7820-498e-be36-c38ded4cc8e3_model.json @@ -0,0 +1,2713 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.195, + 0.05, + 0.805, + 0.105 + ], + "angle": 0, + "content": "Seeing Faces in Things: A Model and Dataset for Pareidolia" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.14, + 0.896, + 0.181 + ], + "angle": 0, + "content": "Mark Hamilton\\(^{1,2}\\), Simon Stent\\(^{3}\\), Vasha DuTell\\(^{1}\\), Anne Harrington\\(^{1}\\), Jennifer Corbett\\(^{1}\\), Ruth Rosenholtz\\(^{4}\\), and William T. Freeman\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.2, + 0.194, + 0.797, + 0.214 + ], + "angle": 0, + "content": "1 MIT, 2 Microsoft, 3 Toyota Research Institute, 4 NVIDIA" + }, + { + "type": "text", + "bbox": [ + 0.149, + 0.26, + 0.851, + 0.563 + ], + "angle": 0, + "content": "Abstract. The human visual system is well-tuned to detect faces of all shapes and sizes. While this brings obvious survival advantages, such as a better chance of spotting unknown predators in the bush, it also leads to spurious face detections. \"Face pareidolia\" describes the perception of face-like structure among otherwise random stimuli: seeing faces in coffee stains or clouds in the sky. In this paper, we study face pareidolia from a computer vision perspective. We present an image dataset of \"Faces in Things\", consisting of five thousand web images with human-annotated pareidolic faces. Using this dataset, we examine the extent to which a state-of-the-art human face detector exhibits pareidolia, and find a significant behavioral gap between humans and machines. We find that the evolutionary need for humans to detect animal faces, as well as human faces, may explain some of this gap. Finally, we propose a simple statistical model of pareidolia in images. Through studies on human subjects and our pareidolic face detectors we confirm a key prediction of our model regarding what image conditions are most likely to induce pareidolia. Dataset and Website: https://aka.ms/faces-in-things" + }, + { + "type": "text", + "bbox": [ + 0.145, + 0.581, + 0.782, + 0.599 + ], + "angle": 0, + "content": "Keywords: Pareidolia \\(\\cdot\\) Face Detection \\(\\cdot\\) Human Psychophysics" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.632, + 0.311, + 0.652 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.675, + 0.796, + 0.693 + ], + "angle": 0, + "content": "Hamlet: Do you see yonder cloud that's almost in the shape of a camel?" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.694, + 0.599, + 0.71 + ], + "angle": 0, + "content": "Polonius: By the Mass and 'tis, like a camel indeed." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.712, + 0.406, + 0.727 + ], + "angle": 0, + "content": "Hamlet: Methinks it is a weasel." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.729, + 0.434, + 0.744 + ], + "angle": 0, + "content": "Polonius: It is back'd like a weasel." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.747, + 0.332, + 0.762 + ], + "angle": 0, + "content": "Hamlet: Or like a whale." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.765, + 0.363, + 0.781 + ], + "angle": 0, + "content": "Polonius: Very like a whale." + }, + { + "type": "text", + "bbox": [ + 0.423, + 0.793, + 0.921, + 0.812 + ], + "angle": 0, + "content": "— Hamlet, Act III, Scene ii, William Shakespeare" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.832, + 0.922, + 0.95 + ], + "angle": 0, + "content": "Pareidolia is a type of visual \"apophenia\", which refers to the perception of patterns in random data. This occurs frequently in human perception as we look at clouds, mountain skylines, and burnt toast. Pareidolia is even described in an exchange in Hamlet [43]. When it was first described, pareidolia was seen as an early symptom of psychosis [7, 44]. Today we know pareidolia is common among healthy humans [45] and infants [23]. It is also not confined to humans:" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.079, + 0.018, + 0.096, + 0.03 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.159, + 0.016, + 0.55, + 0.033 + ], + "angle": 0, + "content": "M. Hamilton, S. Stent, V. DuTell, et al." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.071, + 0.403, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.41, + 0.054, + 0.923, + 0.268 + ], + "angle": 0, + "content": "Fig. 1: You print out an exciting new computer vision paper to review, but as you sit down at your desk to start reading you knock over your coffee cup. At first, you are annoyed, but then, you laugh! The sight of the stain induces \"pareidolia\" in your brain: rather than an unsightly blemish, you see a happy face. In this paper we explore the phenomenon of face pareidolia: Why don't we see faces all the time? Why do we see them at all when they are clearly so different from human faces? Can a better understanding of face pareidolia help computer vision-based face detection?" + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.324, + 0.918, + 0.364 + ], + "angle": 0, + "content": "rhesus macaques, for example, have been shown to spend more time fixating on pareidolic than non-pareidolic images, in a manner similar to humans [49]." + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.365, + 0.92, + 0.481 + ], + "angle": 0, + "content": "As an intriguing phenomenon of our visual system, pareidolia presents many opportunities in the study of the visual perception of both humans and machines. It offers a controlled setting in which to study object detection: we can present random signals to the visual system and study what detections arise. Do computer vision detectors exhibit similar misidentifications, and if not, why not? Why don't humans see pareidolic effects everywhere, in any textured region?" + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.482, + 0.92, + 0.655 + ], + "angle": 0, + "content": "To help answer these questions, we introduce an annotated dataset of five thousand pareidolic face images, called \"Faces in Things\". With this dataset, we examine whether modern computer vision face detection systems, trained to robustly detect human faces, exhibit pareidolia. We show that a state-of-the-art neural network trained on the popular WIDER FACE detection benchmark [60] fails to detect pareidolic faces well, even when detection thresholds are relaxed. By fine-tuning the same model on the Faces in Things training data we create a simple and strong baseline for the task of pareidolic face detection, which shows that significantly higher machine pareidolic performance is within reach." + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.657, + 0.92, + 0.85 + ], + "angle": 0, + "content": "Next, we explore how we might bridge this gap to supervised—or ultimately, human—performance, without access to pareidolic training data? Could pareidolia appear in a face detector in a more natural way? The Faces in Things dataset provides a clean testbed to explore these questions in machines. We test a variety of different interventions ranging from image augmentation techniques to additional sources of training data. We find one possible mechanism that accounts for roughly half of the performance gap: when models are fine-tuned to detect animal faces, pareidolic face detection is significantly improved. This suggests that face pareidolia may arise in part from a more general, evolutionary need to detect diverse faces in the natural environment." + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.852, + 0.92, + 0.948 + ], + "angle": 0, + "content": "Finally, we consider why pareidolic faces are not all around us, and why certain textures seem to cause the effect more often. We propose two simple mathematical models, a simple Gaussian process model, and a second deep feature-based model, that capture important features of pareidolia. In particular, we show how these simple models both predict a \"Goldilocks\" zone, where" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.25, + 0.015, + 0.84, + 0.035 + ], + "angle": 0, + "content": "Seeing Faces in Things: A Model and Dataset for Pareidolia" + }, + { + "type": "page_number", + "bbox": [ + 0.903, + 0.017, + 0.918, + 0.031 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.056, + 0.918, + 0.096 + ], + "angle": 0, + "content": "conditions are ideal to induce pareidolia. We confirm the existence of this zone with experiments on both human subjects and face detection models." + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.096, + 0.918, + 0.155 + ], + "angle": 0, + "content": "Through the contributions of our open-source dataset, models, and experimental findings, we bring the study of the intriguing phenomenon of face pareidolia to the computer vision community." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.186, + 0.332, + 0.208 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.228, + 0.918, + 0.422 + ], + "angle": 0, + "content": "Face detection. One of the most famous early examples of face detection was the Viola-Jones face detector [52, 53]. This detector used binary Haar features through simple-to-compute integral images and achieved greater precision and efficiency than early neural-network-based detectors [40, 41] and other feature-based methods [37, 59]. Following the deep learning breakthroughs of the 2010s, methods transitioned from hand-crafted features to learned features, and convolutional neural networks (CNNs) achieved close to human levels of performance on ever-larger datasets [8, 27-29, 32, 36, 48, 63]. For a broader survey of face detection methods, we refer the interested reader to [36, 62]. In our work, we use the recent RetinaFace model [8] as a strong face detection baseline." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.424, + 0.918, + 0.851 + ], + "angle": 0, + "content": "Neuroscience of face pareidolia. The face is a highly unique stimulus for the human visual system [26, 50]: we find faces easy to spot and difficult to ignore. Face detection can occur in both noise and highly degraded images [5]. Prior work shows that face detection occurs in a dedicated brain region, the Fusiform Face Area [33]. But exactly what constitutes a face for the visual cortex and what are the mechanisms underlying pareidolia? A recent study into the temporal dynamics of neuro-imaging data during pareidolic face viewing showed results consistent with \"a broadly-tuned face detection mechanism that privileges sensitivity over selectivity\" [56]. Pareidolic faces do more than give the impression of the presence of faces: [46] show that they can trigger an additional face-specific attentional process, consuming more time and processing power than similar non-pareidolic stimuli, and even enhancing the detection of face-pareidolic objects [47]. Analyses in [30] revealed a network of neurons in the brain specialized to detect face pareidolia. Their results suggested that face processing has a strong top-down component whereby sensory input with even the slightest suggestion of a face can result in the interpretation of a face. Such top-down information might be supportive of some form of inverse rendering as a cognitive mechanism to explain the remarkable robustness of human perception of faces in degraded viewing conditions [11]. While our dataset allows the study of several types of face detection models, we focus our study on feedforward neural networks which are known to yield close to human performance on challenging \"in-the-wild\" datasets [60]." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.852, + 0.918, + 0.949 + ], + "angle": 0, + "content": "Face pareidolia in computer vision. Face detection and face recognition have been core topics in computer vision for many decades, but the study of face pareidolia—and its deep relationship with visual object representation learning—has been relatively overlooked. Face pareidolia has some similarities with the problem of cross-modal recognition or cross-depiction: recognizing the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.018, + 0.095, + 0.03 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.159, + 0.016, + 0.55, + 0.032 + ], + "angle": 0, + "content": "M. Hamilton, S. Stent, V. DuTell, et al." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.054, + 0.217, + 0.146 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.054, + 0.359, + 0.146 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.054, + 0.497, + 0.146 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.054, + 0.638, + 0.146 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.054, + 0.777, + 0.146 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.782, + 0.054, + 0.918, + 0.146 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.147, + 0.217, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.147, + 0.358, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.147, + 0.497, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.147, + 0.638, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.147, + 0.777, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.782, + 0.147, + 0.918, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.253, + 0.921, + 0.307 + ], + "angle": 0, + "content": "Fig. 2: Examples of face pareidolia from our \"Faces in Things\" dataset. Faces in Things consists of five thousand images annotated with bounding boxes (shown here), and facial attributes such as perceived emotion, gender, and intentionality." + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.346, + 0.921, + 0.814 + ], + "angle": 0, + "content": "same objects across different modalities irrespective of how the object is visually depicted. This has been explored particularly in the context of detecting faces, people and objects across modalities such as photography, different art movements, cartoons and sketches [4, 15, 35, 58]. The importance of capturing spatial relationships for robust cross-modal detection has also been highlighted [4]. The work of Castrejon et al. [6] showed how, when learning cross-modal scene representations with neural networks, units would emerge in the shared representation that tended to activate on consistent concepts, independently of the modality. This tendency was used by Abbas & Chalup [1], who found that mid-level units learned during human face detection could generalize to detect semantically similar facial key-points in pareidolic images, showing promise for pareidolia to emerge. However, the authors only evaluated the method qualitatively over a small test set of ten images. One route to a larger dataset may be through pareidolic image generation, which shows promise but does not yet produce convincingly natural images [12]. Curating a larger dataset, \"Totally-Looks-Like\" [39] explored the perceptual judgment of image similarity between humans and CNNs, using images which had been paired by humans as visually similar but semantically disparate. They found that visual representations extracted from CNNs such as ResNet [18] perform poorly in terms of reproducing the matching selected by humans. Though this dataset is of similar size to ours (6k samples), it is not specifically tailed towards face pareidolia and offers no bounding box or key-point annotations. Other datasets such as COCO-Periph [17] have been used to show that object detection behavior in CNNs and transformers diverge from human perception in peripheral vision." + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.814, + 0.922, + 0.95 + ], + "angle": 0, + "content": "In summary, there has yet to be a computational model of how or why pareidolia might arise proposed in the literature. Moreover, despite the abundance of face detection datasets [36], there is no large-scale dataset to directly support the study of face pareidolia. A large-scale pareidolia dataset would help the community explore the mechanisms underlying pareidolia, which may in turn help us to understand and harness human visual attention (which is drawn towards face-like objects), reduce pareidolic false positives in face detectors, help design-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.248, + 0.015, + 0.842, + 0.035 + ], + "angle": 0, + "content": "Seeing Faces in Things: A Model and Dataset for Pareidolia" + }, + { + "type": "page_number", + "bbox": [ + 0.903, + 0.017, + 0.919, + 0.031 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.051, + 0.382, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.438, + 0.051, + 0.583, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.051, + 0.837, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.162, + 0.19, + 0.613, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.19, + 0.783, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.333, + 0.923, + 0.424 + ], + "angle": 0, + "content": "Fig. 3: Attributes of the Faces in Things Dataset. We find that \\(31\\%\\) of faces are considered challenging to spot; faces are largely \\((31\\%)\\) judged as happy; approximately half \\((47\\%)\\) are judged as accidental rather than by design; animals and humans are seen in roughly equal numbers; and we observe a slight bias \\((16\\%\\) vs \\(3\\%)\\) towards male over female faces, similar to biases observed in prior studies [54,55]." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.457, + 0.921, + 0.496 + ], + "angle": 0, + "content": "ers avoid or create pareidolia, improve pareidolic animation, and create systems that better understand how humans perceive the world." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.521, + 0.475, + 0.546 + ], + "angle": 0, + "content": "3 Faces in Things Dataset" + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.557, + 0.923, + 0.83 + ], + "angle": 0, + "content": "To address this gap, we begin by sampling candidate pareidolic images from the LAION-5B dataset [42]. This dataset consists of 5.85 billion CLIP-filtered image-text pairs, of which \\(40\\%\\) of captions contain English. We use CLIP retrieval [2] to build a raw image set based on text queries including \"pareidolia\", \"faces in things\", \"accidental faces\", and \"[object] looks like a face\". We download images, check for duplicates, then downsample to \\(512 \\times 512\\) pixels while preserving the aspect ratio with white-space padding. We used the VGG Image Annotation tool [10] to manually annotate images, removing samples that contain the faces of actual humans or animals. Some examples of annotated images are shown in Fig. 2. Our annotations include the bounding boxes of pareidolic faces and basic facial attributes as summarized in Fig. 3. Though beyond the scope of the current paper, we note that these attributes could be useful for other future studies. We divide the dataset at random into training (\\(70\\%\\)) and testing (\\(30\\%\\)) sets. We refer to this as the 'Pareidolic' dataset." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.855, + 0.315, + 0.879 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.891, + 0.923, + 0.951 + ], + "angle": 0, + "content": "Datasets. We use the following additional datasets. Fig. 4 shows the average faces within our dataset (Pareidolic) and the WIDER FACE (Human), and AnimalWeb (Animal) datasets." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.079, + 0.017, + 0.095, + 0.031 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.159, + 0.015, + 0.55, + 0.033 + ], + "angle": 0, + "content": "M. Hamilton, S. Stent, V. DuTell, et al." + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.053, + 0.412, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.271, + 0.203, + 0.386, + 0.221 + ], + "angle": 0, + "content": "Pareidolic" + }, + { + "type": "image", + "bbox": [ + 0.417, + 0.054, + 0.582, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.457, + 0.204, + 0.542, + 0.221 + ], + "angle": 0, + "content": "Human" + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.054, + 0.753, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.626, + 0.204, + 0.712, + 0.221 + ], + "angle": 0, + "content": "Animal" + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.236, + 0.921, + 0.326 + ], + "angle": 0, + "content": "Fig. 4: The Appearance of an Average Pareidolic Face. Per-channel histogram-qualified average images for registered pareidolic faces (our Faces in Things dataset), human faces (samples from the WIDER FACE dataset [60]), and animal faces (AnimalWeb [24]). The average pareidolic face, while less distinct than human or animal, has surprisingly clear eye, nose, and mouth features, and vertical symmetry." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.365, + 0.919, + 0.461 + ], + "angle": 0, + "content": "WIDER FACE [60] is a popular face detection benchmark dataset with 32,203 images and 393,703 faces. It contains a high degree of variability in scale, pose, makeup, lighting, emotion, and occlusion, organized across 61 event classes. We use the provided \\(40\\% / 10\\% / 50\\%\\) splits for training, validation, and testing. We refer to this as the 'Human' dataset." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.462, + 0.92, + 0.577 + ], + "angle": 0, + "content": "AnimalWeb [24] is a collection of 22,451 faces from 334 diverse species and 21 animal orders across biological taxonomy. These faces are captured 'in-the-wild' and are consistently annotated with 9 landmarks on key facial features. We convert these landmarks to bounding boxes, by finding the tightest box that captures the points and expanding this box's width and height by \\(15\\%\\). We refer to this as the 'Animal' dataset." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.58, + 0.92, + 0.678 + ], + "angle": 0, + "content": "WIDER FACE Corruptions. To measure whether pareidolia could arise from common data augmentations we corrupt the WIDER FACE images using the level 3 strength of the corruptions used in both the COCO-C [34] and ImageNet-C [19] datasets. We also include a Sobel filtering corruption [13] which has been shown to reduce a model's dependence on texture information [22]." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.696, + 0.922, + 0.95 + ], + "angle": 0, + "content": "Models and Training. We use RetinaFace [8] which achieves state-of-the-art performance on WIDER FACE easy and medium subsets and is the third-best face detector on the hard subset, missing the top model by less than a percentage point of Average Precision (AP). We perform experiments using both their MobileNet [21] and ResNet50 [18] backbones and use the Pytorch_Retinaface [3] repository to ensure the same experimental conditions, dataset characteristics, and preprocessing. We use pre-trained models provided by this repository and fine-tune them for 10 epochs with the AdamW optimizer [31] using a learning rate of \\(10^{-4}\\) and a weight decay of \\(5 \\times 10^{-4}\\). We verify that fine-tuning using this strategy on the original WIDER FACE training dataset does not hurt model performance. When fine-tuning on Faces in Things (Pareidolia), Animal-Web (Animal), WIDER FACE (Human), and Corruption datasets we randomly replace images in the original WIDER FACE stream of training data with data" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.248, + 0.015, + 0.842, + 0.035 + ], + "angle": 0, + "content": "Seeing Faces in Things: A Model and Dataset for Pareidolia" + }, + { + "type": "page_number", + "bbox": [ + 0.904, + 0.017, + 0.921, + 0.031 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.054, + 0.245, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.249, + 0.054, + 0.412, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.418, + 0.054, + 0.582, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.054, + 0.75, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.054, + 0.919, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.164, + 0.245, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.249, + 0.164, + 0.412, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.418, + 0.164, + 0.582, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.164, + 0.75, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.164, + 0.919, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.289, + 0.924, + 0.381 + ], + "angle": 0, + "content": "Fig. 5: Qualitative Analysis of Transfer Experiments. On a sample of held-out test images, we visualize the confident \\((p > 10\\%)\\) detections of our ground truth (red), our model fine-tuned on human faces (blue), and our model fine-tuned on animal faces (green). It is evident from these and Table 1 that fine-tuning on animal faces significantly boosts the model's ability to detect pareidolic faces." + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.42, + 0.92, + 0.538 + ], + "angle": 0, + "content": "from the target dataset \\(90\\%\\) of the time. This allows the network to learn the new task without catastrophic forgetting [25]. These changes to the optimizer, learning rate, number of epochs, and stream of training data are the only changes we make to the training paradigm of [8]. Figures in this work use the MobileNet architecture of RetinaNet unless specified otherwise. AP evaluation computations share the same setting and parameters as [8]." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.569, + 0.733, + 0.589 + ], + "angle": 0, + "content": "4.1 Does a SOTA Face Detector Exhibit Pareidolia?" + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.606, + 0.922, + 0.803 + ], + "angle": 0, + "content": "We measure the Average Precision (AP) of the MobileNet and ResNet50 RetinaNet architectures on the Faces in Things dataset. The first row of Table 1 shows results for existing pre-trained models, and the second row shows those for models fine-tuned on the original WIDER FACE training data. These act as control groups to ensure our transfer learning procedure does not interfere with our measurement of the effects of other interventions. Though these models exhibit pareidolia to a small extent, they fall far short of a model fine-tuned to detect pareidolic faces. Fig. 5 also depicts some of these predictions with blue boxes. On the whole, the models trained only on human faces are largely silent across the Faces in Things dataset." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.835, + 0.53, + 0.856 + ], + "angle": 0, + "content": "4.2 How Might Pareidolia Emerge?" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.871, + 0.92, + 0.951 + ], + "angle": 0, + "content": "The WIDER FACE dataset is known for its diversity of lighting, pose, makeup, emotion, and scale of faces. This fact, coupled with the results of Section 4.1 begs the question: What else do models need to experience pareidolia as humans do? The Faces in Things dataset provides a clean and robust setting to explore" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.018, + 0.095, + 0.03 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.159, + 0.016, + 0.55, + 0.033 + ], + "angle": 0, + "content": "M. Hamilton, S. Stent, V. DuTell, et al." + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.046, + 0.491, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.046, + 0.911, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.275, + 0.921, + 0.349 + ], + "angle": 0, + "content": "Fig. 6: Measuring the effect of several training interventions on pareidolic face detection The left plot shows that fine-tuning RetinaNet on animals improves pareidolic face detection more than any other intervention. Conversely, the right plot shows that pareidolic fine-tuning improves animal face detection performance." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.395, + 0.919, + 0.475 + ], + "angle": 0, + "content": "the development of pareidolia in algorithms. Unlike in humans, where it is impossible to causally intervene on their facial training data, we can easily modify an algorithm's training data. This makes it possible to explore whether one can induce pareidolia in algorithms using specific stimuli." + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.478, + 0.921, + 0.79 + ], + "angle": 0, + "content": "To this end, we investigate whether a variety of training data interventions can induce pareidolia in algorithms. In particular, we measure the effect of adding several data augmentations from the COCO-C [34] and ImageNet-C [19] datasets and explore a Sobel filtering augmentation which reduces models' dependence on texture. Additionally, we also measure the effect of adding animal faces to the training data. Animal faces show a far greater breadth of variation in coloration, structure, and appearance than human faces. Recognizing animal faces provides many evolutionary advantages including gaze detection during hunting and avoiding onlooking predators. The generality required to detect this wide space of faces could yield a greater number of \"false positives\" that lead to the sensation of pareidolia. Indeed, some recent studies provide some corroborating evidence for this hypothesis. Firstly, Rhesus Monkeys exhibit pareidolia [49] showing this effect does not only occur in humans. Secondly, the experience of pareidolia is a rapid cognitive process and not a \"late re-interpretation\" of input signals [16, 56] which the authors conclude is evidence that pareidolia could be linked to the need to quickly react to predators." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.794, + 0.922, + 0.95 + ], + "angle": 0, + "content": "We plot the change to pareidolic face detection performance as a function of each training intervention in the left panel of Figure 6. Of the different corruption interventions, we find that Sobel filtering, motion blur, Gaussian noise, and fog tend to slightly improve pareidolic face detection while most other corruptions do not improve pareidolic face detection performance. Most strikingly, the addition of animal faces to the training data roughly doubles the algorithms' ability to detect pareidolic faces compared to the control group, closing around half of the gap between a human-trained model and a pareidolia trained model." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.25, + 0.015, + 0.84, + 0.035 + ], + "angle": 0, + "content": "Seeing Faces in Things: A Model and Dataset for Pareidolia" + }, + { + "type": "page_number", + "bbox": [ + 0.903, + 0.017, + 0.918, + 0.031 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table", + "bbox": [ + 0.285, + 0.052, + 0.709, + 0.2 + ], + "angle": 0, + "content": "
FinetuningAP
MobileNetResNet50
None7.9%2.8%
Human (Control)9.8%3.6%
Animal16.7%15.4%
Pareidolia33.9%27.1%
Animal + Pareidolia36.4%31.7%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.079, + 0.206, + 0.92, + 0.296 + ], + "angle": 0, + "content": "Table 1: Effect of Fine Tuning on Pareidolic Face Detection. Our results show that WIDER FACE-trained RetinaFace models do not detect many pareidolic images. Fine-tuning these models on animal faces approximately doubles pareidolic face detection rates. Interestingly, adding animal faces alongside pareidolic faces (30%/70% split respectively) can improve performance over fine-tuning on pareidolic faces alone." + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.328, + 0.918, + 0.445 + ], + "angle": 0, + "content": "Reciprocally, the right-hand plot of Figure 6 shows that fine-tuning on pareidolic images yields the greatest improvement in animal face detection. We further explore this phenomenon in Table 1, where we show that this effect occurs across both MobileNet and ResNet50 architectures. Finally, we also show that adding a small number of animal faces (30% animal 70% pareidolic) can improve pareidolic face detection performance over pareidolic images alone." + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.447, + 0.918, + 0.658 + ], + "angle": 0, + "content": "To understand this effect better, Fig. 7 visualizes the inner representations of this model across the three datasets (Human = WIDER FACE, Animal = AnimalWeb, Pareidolic = Faces in Things). Specifically, we extract multi-scale features from the animal and pareidolia fine-tuned RetinaNet shared feature layer before the application of the classification and regression heads. We average pool these features across the bounding box for each face and visualize them with t-SNE [20]. This figure shows that RetinaNet's representations of animal and pareidolic faces tend to cluster together and are distinct from its representations of human faces. This lends evidence to the relative similarity of pareidolic and animal faces compared to human faces. We also reiterate that we filtered the Faces in Things dataset to avoid images of real animals." + }, + { + "type": "title", + "bbox": [ + 0.081, + 0.684, + 0.416, + 0.707 + ], + "angle": 0, + "content": "5 Modeling Pareidolia" + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.721, + 0.918, + 0.838 + ], + "angle": 0, + "content": "Though many prior works have measured pareidolia, there has yet to be a simple mathematical model that describes the high-level structure of this phenomenon. In this section we provide two simple formal models of pareidolia and show that they both exhibit a testable prediction: the existence of a peak in pareidolic face detection as a function of an image's complexity. Section 5.4 presents experimental evidence of this \"pareidolic peak\" in both humans and machines." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.862, + 0.502, + 0.881 + ], + "angle": 0, + "content": "5.1 Gaussian Model of Pareidolia" + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.89, + 0.918, + 0.948 + ], + "angle": 0, + "content": "A model of pareidolia needs to describe two processes: (1) the random process that generates candidate images, and (2) the face detection process which determines when an image is pareidolic. We begin with a simple Gaussian model for" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.017, + 0.105, + 0.031 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.159, + 0.016, + 0.55, + 0.033 + ], + "angle": 0, + "content": "M. Hamilton, S. Stent, V. DuTell, et al." + }, + { + "type": "image", + "bbox": [ + 0.256, + 0.059, + 0.737, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.311, + 0.924, + 0.404 + ], + "angle": 0, + "content": "Fig. 7: Visualizing RetinaNet Representations across Datasets. Animal+Pareidolia fine-tuned RetinaNet representations tend to group animal and pareidolic faces together. This lends evidence to the hypothesis that the perception of animal and pareidolic faces are linked. (To highlight the commonality of pareidolic animal detection we note the similarity of these points to a frog.)" + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.45, + 0.92, + 0.548 + ], + "angle": 0, + "content": "each. We model the image generation process as a sum of independent normal modes, each contributing a zero-mean Gaussian of a specified variance, multiplied by the mode image \\( y_{i} \\). For example, as in [9, 51], these modes could be the principal components of a mean-subtracted image dataset. In this setting the generated image, \\( y \\), is a weighted sum of the normal modes:" + }, + { + "type": "equation", + "bbox": [ + 0.291, + 0.572, + 0.92, + 0.612 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {y} = \\sum_ {i} n _ {i} \\boldsymbol {y} _ {i} \\quad \\text {w h e r e}, \\quad n _ {i} \\sim N (0, \\sigma_ {i}) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.635, + 0.922, + 0.794 + ], + "angle": 0, + "content": "To model our face detection process we capture the intuition of matching an image to a template image and note that this can be generalized to distributions of template images. In particular, the target pareidolic image is represented as a vector, \\(\\pmb{a}\\), of statistically independent target coefficients, \\(a_{i}\\), for each mode. The probability that this mode contributes towards the face detection, \\(P(a_{i})\\), is the probability of detecting the pareidolic value, \\(a_{i}\\), at the \\(i\\)th mode. We assume a Gaussian detection process: \\(P(a_{i}|y_{i}) \\sim N(a_{i}, \\gamma_{i})\\). Because each mode's coefficient is a zero-mean Gaussian distribution, \\(P(y_{i}) \\sim N(0, \\sigma)\\), we have:" + }, + { + "type": "equation", + "bbox": [ + 0.295, + 0.816, + 0.919, + 0.955 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} P \\left(a _ {i}\\right) = \\int_ {y _ {i}} P \\left(a _ {i}, y _ {i}\\right) \\mathrm {d} y _ {i} (2) \\\\ = \\int_ {y _ {i}} P \\left(a _ {i} \\mid y _ {i}\\right) P \\left(y _ {i}\\right) \\mathrm {d} y _ {i} (3) \\\\ = \\frac {1}{2 \\pi \\gamma_ {i} \\sigma_ {i}} \\int_ {y _ {i}} e ^ {- \\frac {(y _ {i} - a _ {i}) ^ {2}}{2 \\gamma_ {i} ^ {2}}} e ^ {- \\frac {y _ {i} ^ {2}}{2 \\sigma^ {2}}} \\mathrm {d} y _ {i} (4) \\\\ \\end{array}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.248, + 0.015, + 0.842, + 0.035 + ], + "angle": 0, + "content": "Seeing Faces in Things: A Model and Dataset for Pareidolia" + }, + { + "type": "page_number", + "bbox": [ + 0.893, + 0.017, + 0.917, + 0.031 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.167, + 0.045, + 0.835, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.369, + 0.922, + 0.549 + ], + "angle": 0, + "content": "Fig. 8: Illustration of the proposed Gaussian model for pareidolia with three example generating distributions. To make pareidolia likely, the generating distribution needs a proper distribution of spatial frequencies. A process with too few spatial frequencies (left) is likely to only generate weak face-like details (\"face-ness\": low). In contrast, with too many frequencies (right), faces can be modeled with exquisite detail (\"face-ness\": high), but the likelihood of drawing any particular desired combination become vanishingly small. The most likely pareidolic images form when the generating distribution has the right spectrum (middle), enabling reasonable faces to emerge with reasonable likelihood. In other words, this model predicts that pareidolic faces will match the low frequencies of faces but differ in the higher frequency details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.589, + 0.92, + 0.648 + ], + "angle": 0, + "content": "Note that \\(\\sigma_i^2\\) is the variance of the random process generating the pareidolia, while \\(\\gamma_i^2\\) is the variance of the likelihood term — how far a mode is allowed to vary from the target mode value before it stops looking like the target image, \\(\\pmb{a}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.648, + 0.92, + 0.707 + ], + "angle": 0, + "content": "We can complete the square in Eq. 4 to write the product of Gaussians in \\( P(a_{i}) \\) as a single Gaussian. Integrating that Gaussian over all possible observations \\( y_{i} \\) gives the probability of finding the pareidolic value \\( a_{i} \\) from mode \\( i \\):" + }, + { + "type": "equation", + "bbox": [ + 0.319, + 0.722, + 0.919, + 0.774 + ], + "angle": 0, + "content": "\\[\nP (a _ {i}) = \\frac {1}{\\sqrt {2 \\pi (\\gamma_ {i} ^ {2} + \\sigma_ {i} ^ {2})}} e ^ {- \\frac {a _ {i} ^ {2}}{2 (\\sigma_ {i} ^ {2} + \\gamma_ {i} ^ {2})}} \\tag {5}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.815, + 0.475, + 0.835 + ], + "angle": 0, + "content": "5.2 Predicting Peak Pareidolia" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.852, + 0.922, + 0.95 + ], + "angle": 0, + "content": "For a given mode's detection variance, \\(\\gamma_i^2\\), and target mode coefficient, \\(a_i\\), Eq. 4 allows us to find the optimal mode variance to generate pareidolia, i.e., to maximize \\(P(a_i)\\). Unfortunately, we seldom have the flexibility to design a random process one mode at a time. But we may have the option to select between image generation processes that have different numbers of modes, \\(M\\). Since each mode" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.017, + 0.105, + 0.031 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.159, + 0.016, + 0.55, + 0.032 + ], + "angle": 0, + "content": "M. Hamilton, S. Stent, V. DuTell, et al." + }, + { + "type": "image", + "bbox": [ + 0.087, + 0.052, + 0.481, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.235, + 0.487, + 0.36 + ], + "angle": 0, + "content": "Fig. 9: Probability of pareidolia (Eq. 6) in the Gaussian model \\((\\sigma = 10)\\) across images with different spatial frequency distributions (Sec. 5.4). This assumes spatial frequencies are uncorrelated and thus underestimates the probability of pareidolia, however peak pareidolia is still present." + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.051, + 0.912, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.51, + 0.224, + 0.92, + 0.351 + ], + "angle": 0, + "content": "Fig. 10: Probability of pareidolia under the feature-based example of Eq. (8) as a function of the rate of feature detection, \\(\\lambda_{i} = \\lambda\\) for all \\(i\\), within the random images. Note the low probability of pareidolia for both feature-free \\((\\lambda \\to 0)\\) and feature-rich \\((\\lambda \\gg 0)\\) random images." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.402, + 0.918, + 0.441 + ], + "angle": 0, + "content": "is independent, the probability of a pareidolic detection of the target object template is the product of detecting the target coefficient for each of the M modes:" + }, + { + "type": "equation", + "bbox": [ + 0.406, + 0.459, + 0.918, + 0.511 + ], + "angle": 0, + "content": "\\[\nP (a) = \\prod_ {i} ^ {M} P \\left(a _ {i}\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.523, + 0.92, + 0.78 + ], + "angle": 0, + "content": "We plot some predictions of our Gaussian model in Fig. 9 for a target template with a \\(\\frac{1}{f}\\) power spectrum (standard deviation of each mode inversely proportional to mode number) on noise images of varying complexity. We note the existence of a peak in pareidolic detection probability for random image generation processes with a mid-range number of spatial modes, as measured by the width of Gaussian that modulates power in Fourier space. Too few modes in the random generation process, and no image will ever have enough complexity to render the target well. Too many modes and pareidolia becomes unlikely because so many modes need to match the desired target values. Each added mode multiplies the pareidolia probability by another small factor. In between, there is what we call peak pareidolia. As the detection value, \\(\\sigma\\), becomes more stringent (smaller) the peak pareidolia value occurs at a larger number of modes and becomes less probable. We illustrate this effect in Figure 8" + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.814, + 0.651, + 0.833 + ], + "angle": 0, + "content": "5.3 Higher-Level Feature Model of Pareidolia" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.852, + 0.92, + 0.949 + ], + "angle": 0, + "content": "The Gaussian model for pareidolia above lays out important aspects of pareidolia, but relies on a naive model of object detection, the squared distance from a template image. We assume that a more realistic model of human perception would incorporate higher-level features and introduce a still simple, yet more realistic, feature-based model." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.247, + 0.015, + 0.84, + 0.035 + ], + "angle": 0, + "content": "Seeing Faces in Things: A Model and Dataset for Pareidolia" + }, + { + "type": "page_number", + "bbox": [ + 0.893, + 0.017, + 0.92, + 0.031 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.057, + 0.923, + 0.234 + ], + "angle": 0, + "content": "We assume that the detection of an object requires particular features to be detected in certain spatial regions, e.g. an eye in the top left and right, a nose in the center, and a mouth in the bottom. Such an approach has been used in computer vision object detection algorithms, e.g. [14, 57, 61]. Any given object template has some number of regions, \\( R_{i} \\), indexed by \\( i \\), within which a given feature, \\( F_{i} \\), must be detected. The other features, \\( F_{j\\neq i} \\), should not be detected in region \\( i \\). For a given random image where we hope to detect pareidolia, we assume that feature existence is a spatial Poisson process. In this process, the probability of \\( n \\) feature instances for any given feature \\( i \\) over some area \\( B_{i} \\) is" + }, + { + "type": "equation", + "bbox": [ + 0.359, + 0.247, + 0.92, + 0.288 + ], + "angle": 0, + "content": "\\[\nP (n _ {i}) = \\frac {\\left(\\lambda_ {i} | B _ {i} |\\right) ^ {n _ {i}}}{n _ {i} !} e ^ {- \\lambda_ {i} | B _ {i} |} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.296, + 0.924, + 0.414 + ], + "angle": 0, + "content": "To detect a pareidolic instance of the object template, we must detect one feature of the correct type, \\( F_{i} \\), in each region \\( i \\) of the face template, and zero features of the wrong type, \\( F_{j\\neq i} \\) in each region \\( i \\). Assuming independence of the feature detections, and for simplicity setting all the feature detection rates to be the same, \\( \\lambda_{i} = \\lambda \\), and all the template areas to be the same, \\( B_{i} = 1 \\), we have for the probability, \\( P(O) \\), of pareidolic detection of object \\( O \\):" + }, + { + "type": "equation", + "bbox": [ + 0.375, + 0.43, + 0.919, + 0.481 + ], + "angle": 0, + "content": "\\[\nP (O) = \\prod_ {i} ^ {M} \\lambda^ {1} e ^ {- \\lambda (M - 1)} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.496, + 0.924, + 0.673 + ], + "angle": 0, + "content": "For the case of \\( M = 4 \\), a simple detection model for two eyes, a nose, and a mouth, we have \\( P(o) = \\lambda^4 e^{-16\\lambda} \\), which is plotted in Fig. 10. In this feature-based object detection model, we also find the existence of \"peak pareidolia\". Again, it is governed by a parameter describing the complexity of the random image, in this case a Poisson process rate parameter, \\( \\lambda \\), that governs the probability of a feature detection per unit area. For too low a rate, the model doesn't generate enough features to satisfy the object template, for too high a rate, the probability of seeing only the right features in just the right places becomes very small. In between is the most probable rate for pareidolia." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.702, + 0.838, + 0.722 + ], + "angle": 0, + "content": "5.4 Measuring the Pareidolic Peak in Humans and Machines" + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.735, + 0.924, + 0.95 + ], + "angle": 0, + "content": "Both mathematical models of Section 5 predict the existence of a peak of pareidolic face detection as a function of image complexity. We show the existence of this pareidolic peak in both humans and machines. In particular, we perform a psychophysics experiment where human subjects view noise images of varying complexity and report how many pareidolic faces they saw in each image, from zero to nine. Campbell [5] demonstrated that a 12x12 array of random, binary squares is sufficient to evoke human and animal faces. We generate noise images of varying complexity by randomly sampling Fourier coefficients and modulating these coefficients with a zero mean \\(\\sigma^2\\) variance Gaussian in Fourier space. We show some samples of these images on the x-axis of Fig. 11. Intuitively, the Gaussian envelope in frequency space filters out most frequencies higher than \\(\\sigma\\)" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.018, + 0.105, + 0.031 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.159, + 0.016, + 0.55, + 0.033 + ], + "angle": 0, + "content": "M. Hamilton, S. Stent, V. DuTell, et al." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.042, + 0.49, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.041, + 0.915, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.238, + 0.921, + 0.327 + ], + "angle": 0, + "content": "Fig. 11: Measuring Peak Pareidolia. Left: Subjects were asked how many faces they see in each noise image. We plot the average number of faces detected as a function of noise frequency (examples on x-axis), the mean over all subjects and its \\(95\\%\\) confidence interval in red. Right: average number of faces detected by our fine-tuned models. This reveals the \"peak pareidolia\" effect predicted in Section 5 across humans and machines." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.373, + 0.918, + 0.412 + ], + "angle": 0, + "content": "after applying an inverse Fourier transform. We detail our image stimuli creation method in the Supplement." + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.415, + 0.92, + 0.57 + ], + "angle": 0, + "content": "We find that humans exhibit the model-expected peak pareidolia, with a maximum number of faces detected at a frequency filter width of 16 (Fig. 11, left). The existence of a pareidolic peak at or near this filter width is consistent among all subjects even for those that reported fewer faces overall. Although response time did decrease slightly at higher frequencies, it did not fall off completely at the highest frequency levels, indicating that fewer reported faces were not the result of subjects \"giving up\" on the task. We provide additional details and analysis of this experiment in the Supplement." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.574, + 0.92, + 0.691 + ], + "angle": 0, + "content": "Finally, we evaluate our fine-tuned models from Section 4 on the same images to test whether machines also exhibit peak pareidolia (Fig. 11, right). In particular, we showed the models 5,000 sampled noise images of varying frequency levels and counted the number of face detections they make with confidence \\(>10\\%\\). We find the same characteristic \"pareidolic peak\" where models detect the most faces in medium-complexity images." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.737, + 0.288, + 0.757 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.793, + 0.922, + 0.95 + ], + "angle": 0, + "content": "We have taken initial steps towards the mathematical modelling of pareidolia and build a richly annotated dataset of images for face pareidolia. We showed through experiments on modern face detectors that detecting animal faces may partly explain the emergence of pareidolia in a complex vision system. The Faces in Things dataset can help the community address other questions about how and why pareidolic behavior emerges, a hallmark of humans' robust recognition system. We hope that our findings and dataset will spark further study of pareidolia and its potential use to improve computer vision systems." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.248, + 0.015, + 0.84, + 0.035 + ], + "angle": 0, + "content": "Seeing Faces in Things: A Model and Dataset for Pareidolia" + }, + { + "type": "page_number", + "bbox": [ + 0.893, + 0.017, + 0.918, + 0.031 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.081, + 0.057, + 0.3, + 0.075 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.083, + 0.918, + 0.141 + ], + "angle": 0, + "content": "We would like to thank Karen Hamilton for her hundreds of hours of annotations for the Faces in Things dataset. We also thank Abhishek Dutta who created the VGG Image Annotation tool, for his kind generosity to support our project." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.143, + 0.918, + 0.237 + ], + "angle": 0, + "content": "We would like to thank the Microsoft Research Grand Central Resources team for their gracious help performing the experiments in this work. Special thanks to Oleg Losinets and Lifeng Li for their consistent, gracious, and timely help, debugging, and expertise. Without them, none of the experiments could have been run." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.239, + 0.918, + 0.334 + ], + "angle": 0, + "content": "This material is based upon work supported by the National Science Foundation Graduate Research Fellowship under Grant No. 2021323067. Any opinion, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of their employers, or the National Science Foundation." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.336, + 0.918, + 0.413 + ], + "angle": 0, + "content": "This work is supported by the National Science Foundation under Cooperative Agreement PHY-2019786 (The NSF AI Institute for Artificial Intelligence and Fundamental Interactions, http://iaifi.org/) and the CSAIL MEnTorEd Opportunities in Research (METEOR) Fellowship." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.415, + 0.918, + 0.568 + ], + "angle": 0, + "content": "Research was sponsored by the United States Air Force Research Laboratory and the United States Air Force Artificial Intelligence Accelerator and was accomplished under Cooperative Agreement Number FA8750-192-1000. The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the United States Air Force or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for Government purposes notwithstanding any copyright notation herein." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.569, + 0.918, + 0.626 + ], + "angle": 0, + "content": "The authors acknowledge the MIT SuperCloud [38] and Lincoln Laboratory Supercomputing Center for providing HPC resources that have contributed to the research results reported within this paper." + }, + { + "type": "title", + "bbox": [ + 0.081, + 0.653, + 0.235, + 0.672 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.686, + 0.918, + 0.757 + ], + "angle": 0, + "content": "1. Abbas, A., Chalup, S.: From Face Recognition to Facial Pareidolia: Analysing Hidden Neuron Activations in CNNs for Cross-Depiction Recognition. In: 2019 International Joint Conference on Neural Networks (IJCNN). pp. 1–8 (Jul 2019). https://doi.org/10.1109/IJCNN.2019.8852013, iSSN: 2161-4393" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.758, + 0.918, + 0.791 + ], + "angle": 0, + "content": "2. Beaumont, R.: Clip retrieval: Easily compute clip embeddings and build a clip retrieval system with them. https://github.com/rom1504/clip-retrieval (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.792, + 0.918, + 0.825 + ], + "angle": 0, + "content": "3. biubug6: Retinaface in pytorch (Nov 2021), https://github.com/biubug6/Pytorch_Retinaface" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.826, + 0.918, + 0.894 + ], + "angle": 0, + "content": "4. Cai, H., Wu, Q., Corradi, T., Hall, P.: The Cross-Depiction Problem: Computer Vision Algorithms for Recognising Objects in Artwork and in Photographs. arXiv:1505.00110 [cs] (May 2015), http://arxiv.org/abs/1505.00110, arXiv: 1505.00110" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.895, + 0.918, + 0.948 + ], + "angle": 0, + "content": "5. Campbell, F.: How much of the information falling on the retina reaches the visual cortex and how much is stored in the visual memory. Pattern recognition mechanisms 54, 83-94 (1983)" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.686, + 0.918, + 0.948 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.017, + 0.106, + 0.032 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.159, + 0.016, + 0.554, + 0.034 + ], + "angle": 0, + "content": "M. Hamilton, S. Stent, V. DuTell, et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.057, + 0.92, + 0.111 + ], + "angle": 0, + "content": "6. Castrejon, L., Aytar, Y., Vondrick, C., Pirsiavash, H., Torralba, A.: Learning Aligned Cross-Modal Representations from Weakly Aligned Data. In: CVPR (Jun 2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.112, + 0.92, + 0.147 + ], + "angle": 0, + "content": "7. Conrad, K.: Die beginnende Schizophrenie. Versuch einer Gestaltanalyse des Wahns (1958)" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.147, + 0.918, + 0.182 + ], + "angle": 0, + "content": "8. Deng, J., Guo, J., Zhou, Y., Yu, J., Kotsia, I., Zafeiriou, S.: RetinaFace: Single-stage Dense Face Localisation in the Wild. CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.183, + 0.773, + 0.2 + ], + "angle": 0, + "content": "9. Duda, R., Hart, P., Stork, D.: Pattern Classification. Wiley (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.201, + 0.92, + 0.272 + ], + "angle": 0, + "content": "10. Dutta, A., Zisserman, A.: The VIA annotation software for images, audio and video. In: Proceedings of the 27th ACM International Conference on Multimedia. MM '19, ACM, New York, NY, USA (2019). https://doi.org/10.1145/3343031.3350535, https://doi.org/10.1145/3343031.3350535" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.272, + 0.918, + 0.325 + ], + "angle": 0, + "content": "11. Egger, B., Siegel, M.H., Arora, R., Soltani, A.A., Yildirim, I., Tenenbaum, J.: Inverse rendering best explains face perception under extreme illuminations. In: CogSci (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.326, + 0.918, + 0.379 + ], + "angle": 0, + "content": "12. Endo, Y., Asanuma, R., Shimojo, S., Akashi, T.: Systematic face pareidolia generation method using cycle-consistent adversarial networks. IEEE Transactions on Electrical and Electronic Engineering 19(4), 535-541 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.379, + 0.918, + 0.432 + ], + "angle": 0, + "content": "13. Farid, H., Simoncelli, E.P.: Optimally rotation-equivariant directional derivative kernels. In: International Conference on Computer Analysis of Images and Patterns. pp. 207-214. Springer (1997)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.433, + 0.918, + 0.468 + ], + "angle": 0, + "content": "14. Felzenszwalb, P., Girshick, R., McAllester, D., Ramanan, D.: Object detection with discriminatively trained part based models. IEEE TPAMI 32(9) (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.918, + 0.503 + ], + "angle": 0, + "content": "15. Ginosar, S., Haas, D., Brown, T., Malik, J.: Detecting people in cubist art. In: ECCV Workshops (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.504, + 0.92, + 0.539 + ], + "angle": 0, + "content": "16. Hadjikhani, N., Kveraga, K., Naik, P., Ahlfors, S.P.: Early (n170) activation of face-specific cortex by face-like objects. Neuroreport 20(4), 403 (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.92, + 0.609 + ], + "angle": 0, + "content": "17. Harrington, A., DuTell, V., Hamilton, M., Tewari, A., Stent, S., Freeman, W.T., Rosenholtz, R.: Coco-periph: Bridging the gap between human and machine perception in the periphery. In: The Twelfth International Conference on Learning Representations" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.61, + 0.918, + 0.646 + ], + "angle": 0, + "content": "18. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.918, + 0.699 + ], + "angle": 0, + "content": "19. Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. Proceedings of the International Conference on Learning Representations (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.7, + 0.918, + 0.735 + ], + "angle": 0, + "content": "20. Hinton, G.E., Roweis, S.: Stochastic neighbor embedding. Advances in neural information processing systems 15 (2002)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.736, + 0.918, + 0.789 + ], + "angle": 0, + "content": "21. Howard, A.G., Zhu, M., Chen, B., Kalenichenko, D., Wang, W., Weyand, T., Andreetto, M., Adam, H.: Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.918, + 0.842 + ], + "angle": 0, + "content": "22. Ji, X., Henriques, J.F., Vedaldi, A.: Invariant information clustering for unsupervised image classification and segmentation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9865-9874 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.842, + 0.918, + 0.895 + ], + "angle": 0, + "content": "23. Kato, M., Mugitani, R.: Pareidolia in Infants. PLoS ONE 10(2) (Feb 2015). https://doi.org/10.1371/journal.pone.0118539, https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4331561/" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.896, + 0.918, + 0.949 + ], + "angle": 0, + "content": "24. Khan, M.H., McDonagh, J., Khan, S., Shahabuddin, M., Arora, A., Khan, F.S., Shao, L., Tzimiropoulos, G.: AnimalWeb: A large-scale hierarchical dataset of annotated animal faces. In: CVPR (2020)" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.057, + 0.92, + 0.949 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.248, + 0.015, + 0.842, + 0.035 + ], + "angle": 0, + "content": "Seeing Faces in Things: A Model and Dataset for Pareidolia" + }, + { + "type": "page_number", + "bbox": [ + 0.893, + 0.017, + 0.92, + 0.031 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.077, + 0.057, + 0.92, + 0.129 + ], + "angle": 0, + "content": "25. Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A.A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., et al.: Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences 114(13), 3521-3526 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.077, + 0.131, + 0.92, + 0.166 + ], + "angle": 0, + "content": "26. Leopold, D.A., Rhodes, G.: A comparative view of face perception. Journal of Comparative Psychology 124(3), 233 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.168, + 0.918, + 0.202 + ], + "angle": 0, + "content": "27. Li, H., Lin, Z., Shen, X., Brandt, J., Hua, G.: A convolutional neural network cascade for face detection. In: CVPR (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.204, + 0.918, + 0.239 + ], + "angle": 0, + "content": "28. Li, J., Wang, Y., Wang, C., Tai, Y., Qian, J., Yang, J., Wang, C., Li, J., Huang, F.: DSFD: dual shot face detector. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.241, + 0.918, + 0.292 + ], + "angle": 0, + "content": "29. Liao, S., Jain, A.K., Li, S.Z.: A Fast and Accurate Unconstrained Face Detector. PAMI 38(2), 211-223 (Feb 2016). https://doi.org/10.1109/TPAMI.2015.2448075" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.296, + 0.918, + 0.366 + ], + "angle": 0, + "content": "30. Liu, J., Li, J., Feng, L., Li, L., Tian, J., Lee, K.: Seeing Jesus in toast: Neural and behavioral correlates of face pareidolia. Cortex 53, 60-77 (Apr 2014). https://doi.org/10.1016/j.cortex.2014.01.013, http://www.sciencedirect.com/science/article/pii/S0010945214000288" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.368, + 0.918, + 0.403 + ], + "angle": 0, + "content": "31. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.405, + 0.918, + 0.44 + ], + "angle": 0, + "content": "32. Mathias, M., Benenson, R., Pedersoli, M., Van Gool, L.: Face detection without bells and whistles. In: ECCV (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.442, + 0.918, + 0.512 + ], + "angle": 0, + "content": "33. Mcgugin, R., Gatenby, C., Gore, J., Gauthier, I.: High-resolution imaging of expertise reveals reliable object selectivity in the FFA related to perceptual performance. Proceedings of the National Academy of Sciences of the United States of America 109, 17063-8 (Oct 2012). https://doi.org/10.1073/pnas.1116333109" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.514, + 0.918, + 0.568 + ], + "angle": 0, + "content": "34. Michaelis, C., Mitzkus, B., Geirhos, R., Rusak, E., Bringmann, O., Ecker, A.S., Bethge, M., Brendel, W.: Benchmarking robustness in object detection: Autonomous driving when winter is coming. arXiv preprint arXiv:1907.07484 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.569, + 0.918, + 0.604 + ], + "angle": 0, + "content": "35. Mishra, A., Nandan Rai, S., Mishra, A., Jawahar, C.V.: Iit-cfw: A benchmark database of cartoon faces in the wild. In: VASE ECCV Workshops (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.605, + 0.918, + 0.675 + ], + "angle": 0, + "content": "36. Ranjan, R., Sankar, S., Bansal, A., Bodla, N., Chen, J.C., Patel, V., Castillo, C., Chellappa, R.: Deep Learning for Understanding Faces: Machines May Be Just as Good, or Better, than Humans. IEEE Signal Processing Magazine 35, 66-83 (Jan 2018). https://doi.org/10.1109/MSP.2017.2764116" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.678, + 0.918, + 0.731 + ], + "angle": 0, + "content": "37. Rein-Lien Hsu, Abdel-Mottaleb, M., Jain, A.: Face detection in color images. PAMI 24(5), 696-706 (May 2002). https://doi.org/10.1109/34.1000242, http://ieeexplore.ieee.org/document/1000242/" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.732, + 0.918, + 0.821 + ], + "angle": 0, + "content": "38. Reuther, A., Kepner, J., Byun, C., Samsi, S., Arcand, W., Bestor, D., Bergeron, B., Gadepally, V., Houle, M., Hubbell, M., Jones, M., Klein, A., Milechin, L., Mullen, J., Prout, A., Rosa, A., Yee, C., Michaleas, P.: Interactive supercomputing on 40,000 cores for machine learning and data analysis. In: 2018 IEEE High Performance extreme Computing Conference (HPEC). pp. 1-6. IEEE (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.823, + 0.918, + 0.858 + ], + "angle": 0, + "content": "39. Rosenfeld, A., Solbach, M.D., Tsotsos, J.K.: Totally looks like-how humans compare, compared to machines. In: ACCV (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.86, + 0.918, + 0.913 + ], + "angle": 0, + "content": "40. Rowley, H.A., Baluja, S., Kanade, T.: Human Face Detection in Visual Scenes. In: Touretzky, D.S., Mozer, M.C., Hasselmo, M.E. (eds.) NeurIPS (1996), http://papers.nips.cc/paper/1168-human-face-detection-in-visual-scenes.pdf" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.914, + 0.918, + 0.949 + ], + "angle": 0, + "content": "41. Rowley, H.A., Baluja, S., Kanade, T.: Neural network-based face detection. PAMI 20(1), 23-38 (1998)" + }, + { + "type": "list", + "bbox": [ + 0.077, + 0.057, + 0.92, + 0.949 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.017, + 0.105, + 0.031 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.159, + 0.016, + 0.55, + 0.033 + ], + "angle": 0, + "content": "M. Hamilton, S. Stent, V. DuTell, et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.057, + 0.918, + 0.163 + ], + "angle": 0, + "content": "42. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C.W., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., Schramowski, P., Kundurthy, S.R., Crowson, K., Schmidt, L., Kaczmarczyk, R., Jitsev, J.: LAION5b: An open large-scale dataset for training next generation image-text models. In: NeurIPS Datasets and Benchmarks Track (2022), https://openreview.net/forum?id=M3Y74vmsMcY" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.165, + 0.918, + 0.2 + ], + "angle": 0, + "content": "43. Shakespeare, W.: The Tragedy of Hamlet, Prince of Denmark. The Folio Society (1954)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.202, + 0.917, + 0.236 + ], + "angle": 0, + "content": "44. Sims, A.: Symptoms in the mind: An introduction to descriptive psychopathology. Bailliere Tindall Publishers (1988)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.237, + 0.918, + 0.288 + ], + "angle": 0, + "content": "45. Summerfield, C., Egner, T., Mangels, J., Hirsch, J.: Mistaking a house for a face: neural correlates of misperception in healthy humans. Cerebral cortex 16(4), 500-508 (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.29, + 0.918, + 0.342 + ], + "angle": 0, + "content": "46. Takahashi, K., Watanabe, K.: Gaze Cueing by Pareidolia Faces. i-Perception 4(8), 490-492 (Dec 2013). https://doi.org/10.1068/i0617sas, https://doi.org/10.1068/i0617sas" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.343, + 0.917, + 0.395 + ], + "angle": 0, + "content": "47. Takahashi, K., Watanabe, K.: Seeing Objects as Faces Enhances Object Detection. i-Perception 6(5) (2015). https://doi.org/10.1177/2041669515606007, https://doi.org/10.1177/2041669515606007" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.397, + 0.918, + 0.45 + ], + "angle": 0, + "content": "48. Tang, X., Du, D.K., He, Z., Liu, J.: PyramidBox: A Context-Assisted Single Shot Face Detector. In: Ferrari, V., Hebert, M., Sminchisescu, C., Weiss, Y. (eds.) ECCV (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.451, + 0.918, + 0.521 + ], + "angle": 0, + "content": "49. Taubert, J., Wardle, S.G., Flessert, M., Leopold, D.A., Ungerleider, L.G.: Face Pareidolia in the Rhesus Monkey. Current Biology 27(16), 2505-2509 (Aug 2017). https://doi.org/10.1016/j.cub.2017.06.075, http://www.sciencedirect.com/science/article/pii/S0960982217308126" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.522, + 0.918, + 0.556 + ], + "angle": 0, + "content": "50. Tsao, D.Y., Livingstone, M.S.: Mechanisms of face perception. Annu. Rev. Neurosci. 31, 411-437 (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.558, + 0.918, + 0.592 + ], + "angle": 0, + "content": "51. Turk, M., Pentland, A.: Eigenfaces for recognition. J. of Cognitive Neuroscience 3(1) (1991)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.594, + 0.918, + 0.627 + ], + "angle": 0, + "content": "52. Viola, P., Jones, M.: Rapid object detection using a boosted cascade of simple features. In: CVPR (2001)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.629, + 0.918, + 0.646 + ], + "angle": 0, + "content": "53. Viola, P., Jones, M.J.: Robust real-time face detection. IJCV 57(2), 137-154 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.918, + 0.717 + ], + "angle": 0, + "content": "54. Wardle, S.G., Ewing, L., Malcolm, G.L., Paranjape, S., Baker, C.I.: Children perceive illusory faces in objects as male more often than female. Cognition 235 (2023). https://doi.org/https://doi.org/10.1016/j.cognition.2023.105398, https://www.sciencedirect.com/science/article/pii/S001002772300032X" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.918, + 0.788 + ], + "angle": 0, + "content": "55. Wardle, S.G., Paranjape, S., Taubert, J., Baker, C.I.: Illusory faces are more likely to be perceived as male than female. Proceedings of the National Academy of Sciences 119(5) (2022). https://doi.org/10.1073/pnas.2117413119, https://www.pnas.org/doi/abs/10.1073/pnas.2117413119" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.918, + 0.842 + ], + "angle": 0, + "content": "56. Wardle, S.G., Taubert, J., Teichmann, L., Baker, C.I.: Rapid and dynamic processing of face pareidolia in the human brain. Nature communications 11(1), 4518 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.843, + 0.918, + 0.877 + ], + "angle": 0, + "content": "57. Weber, M., Welling, M., Perona, P.: Unsupervised learning of models for recognition. In: ECCV (2000)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.878, + 0.918, + 0.913 + ], + "angle": 0, + "content": "58. Westlake, N., Cai, H., Hall, P.: Detecting People in Artwork with CNNs. In: ECCV Workshops (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.914, + 0.918, + 0.948 + ], + "angle": 0, + "content": "59. Yang, M.H., Kriegman, D., Ahuja, N.: Detecting faces in images: a survey. PAMI 24(1), 34-58 (Jan 2002). https://doi.org/10.1109/34.982883" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.057, + 0.918, + 0.948 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.248, + 0.016, + 0.84, + 0.035 + ], + "angle": 0, + "content": "Seeing Faces in Things: A Model and Dataset for Pareidolia" + }, + { + "type": "page_number", + "bbox": [ + 0.893, + 0.017, + 0.919, + 0.031 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.077, + 0.057, + 0.92, + 0.093 + ], + "angle": 0, + "content": "60. Yang, S., Luo, P., Loy, C.C., Tang, X.: WIDER FACE: A Face Detection Benchmark. In: CVPR (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.077, + 0.094, + 0.92, + 0.129 + ], + "angle": 0, + "content": "61. Yuille, A.L.: Deformable templates for face recognition. Journal of Cognitive Neuroscience 3(1), 59-70 (1991)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.13, + 0.921, + 0.2 + ], + "angle": 0, + "content": "62. Zafeiriou, S., Zhang, C., Zhang, Z.: A survey on face detection in the wild: Past, present and future. Computer Vision and Image Understanding 138, 1-24 (Sep 2015). https://doi.org/10.1016/j.cviu.2015.03.015, https://linkinghub.elsevier.com/retrieve/pii/S1077314215000727" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.201, + 0.92, + 0.237 + ], + "angle": 0, + "content": "63. Zhang, S., Zhu, X., Lei, Z., Shi, H., Wang, X., Li, S.Z.: S3fd: Single shot scale-invariant face detector. In: ICCV (2017)" + }, + { + "type": "list", + "bbox": [ + 0.077, + 0.057, + 0.921, + 0.237 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Seeing Faces in Things_ A Model and Dataset for Pareidolia/78844de0-7820-498e-be36-c38ded4cc8e3_origin.pdf b/2024/Seeing Faces in Things_ A Model and Dataset for Pareidolia/78844de0-7820-498e-be36-c38ded4cc8e3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5699db1ab3d5fe795fa2ed02729602bd90cd1fef --- /dev/null +++ b/2024/Seeing Faces in Things_ A Model and Dataset for Pareidolia/78844de0-7820-498e-be36-c38ded4cc8e3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ce313ba3a414ff23812a04aee5c504cdd0f24235e62aa75b1de01b755b3f4df1 +size 8766583 diff --git a/2024/Seeing Faces in Things_ A Model and Dataset for Pareidolia/full.md b/2024/Seeing Faces in Things_ A Model and Dataset for Pareidolia/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e7af62b18b0a55d7b4a6e62971a9f05fc47c8048 --- /dev/null +++ b/2024/Seeing Faces in Things_ A Model and Dataset for Pareidolia/full.md @@ -0,0 +1,343 @@ +# Seeing Faces in Things: A Model and Dataset for Pareidolia + +Mark Hamilton $^{1,2}$ , Simon Stent $^{3}$ , Vasha DuTell $^{1}$ , Anne Harrington $^{1}$ , Jennifer Corbett $^{1}$ , Ruth Rosenholtz $^{4}$ , and William T. Freeman $^{1}$ + +1 MIT, 2 Microsoft, 3 Toyota Research Institute, 4 NVIDIA + +Abstract. The human visual system is well-tuned to detect faces of all shapes and sizes. While this brings obvious survival advantages, such as a better chance of spotting unknown predators in the bush, it also leads to spurious face detections. "Face pareidolia" describes the perception of face-like structure among otherwise random stimuli: seeing faces in coffee stains or clouds in the sky. In this paper, we study face pareidolia from a computer vision perspective. We present an image dataset of "Faces in Things", consisting of five thousand web images with human-annotated pareidolic faces. Using this dataset, we examine the extent to which a state-of-the-art human face detector exhibits pareidolia, and find a significant behavioral gap between humans and machines. We find that the evolutionary need for humans to detect animal faces, as well as human faces, may explain some of this gap. Finally, we propose a simple statistical model of pareidolia in images. Through studies on human subjects and our pareidolic face detectors we confirm a key prediction of our model regarding what image conditions are most likely to induce pareidolia. Dataset and Website: https://aka.ms/faces-in-things + +Keywords: Pareidolia $\cdot$ Face Detection $\cdot$ Human Psychophysics + +# 1 Introduction + +Hamlet: Do you see yonder cloud that's almost in the shape of a camel? + +Polonius: By the Mass and 'tis, like a camel indeed. + +Hamlet: Methinks it is a weasel. + +Polonius: It is back'd like a weasel. + +Hamlet: Or like a whale. + +Polonius: Very like a whale. + +— Hamlet, Act III, Scene ii, William Shakespeare + +Pareidolia is a type of visual "apophenia", which refers to the perception of patterns in random data. This occurs frequently in human perception as we look at clouds, mountain skylines, and burnt toast. Pareidolia is even described in an exchange in Hamlet [43]. When it was first described, pareidolia was seen as an early symptom of psychosis [7, 44]. Today we know pareidolia is common among healthy humans [45] and infants [23]. It is also not confined to humans: + +![](images/f78aab5b9190439d9daef9925f38018d7e676cab28517fbb93a0561c8b304a7e.jpg) +Fig. 1: You print out an exciting new computer vision paper to review, but as you sit down at your desk to start reading you knock over your coffee cup. At first, you are annoyed, but then, you laugh! The sight of the stain induces "pareidolia" in your brain: rather than an unsightly blemish, you see a happy face. In this paper we explore the phenomenon of face pareidolia: Why don't we see faces all the time? Why do we see them at all when they are clearly so different from human faces? Can a better understanding of face pareidolia help computer vision-based face detection? + +rhesus macaques, for example, have been shown to spend more time fixating on pareidolic than non-pareidolic images, in a manner similar to humans [49]. + +As an intriguing phenomenon of our visual system, pareidolia presents many opportunities in the study of the visual perception of both humans and machines. It offers a controlled setting in which to study object detection: we can present random signals to the visual system and study what detections arise. Do computer vision detectors exhibit similar misidentifications, and if not, why not? Why don't humans see pareidolic effects everywhere, in any textured region? + +To help answer these questions, we introduce an annotated dataset of five thousand pareidolic face images, called "Faces in Things". With this dataset, we examine whether modern computer vision face detection systems, trained to robustly detect human faces, exhibit pareidolia. We show that a state-of-the-art neural network trained on the popular WIDER FACE detection benchmark [60] fails to detect pareidolic faces well, even when detection thresholds are relaxed. By fine-tuning the same model on the Faces in Things training data we create a simple and strong baseline for the task of pareidolic face detection, which shows that significantly higher machine pareidolic performance is within reach. + +Next, we explore how we might bridge this gap to supervised—or ultimately, human—performance, without access to pareidolic training data? Could pareidolia appear in a face detector in a more natural way? The Faces in Things dataset provides a clean testbed to explore these questions in machines. We test a variety of different interventions ranging from image augmentation techniques to additional sources of training data. We find one possible mechanism that accounts for roughly half of the performance gap: when models are fine-tuned to detect animal faces, pareidolic face detection is significantly improved. This suggests that face pareidolia may arise in part from a more general, evolutionary need to detect diverse faces in the natural environment. + +Finally, we consider why pareidolic faces are not all around us, and why certain textures seem to cause the effect more often. We propose two simple mathematical models, a simple Gaussian process model, and a second deep feature-based model, that capture important features of pareidolia. In particular, we show how these simple models both predict a "Goldilocks" zone, where + +conditions are ideal to induce pareidolia. We confirm the existence of this zone with experiments on both human subjects and face detection models. + +Through the contributions of our open-source dataset, models, and experimental findings, we bring the study of the intriguing phenomenon of face pareidolia to the computer vision community. + +# 2 Related Work + +Face detection. One of the most famous early examples of face detection was the Viola-Jones face detector [52, 53]. This detector used binary Haar features through simple-to-compute integral images and achieved greater precision and efficiency than early neural-network-based detectors [40, 41] and other feature-based methods [37, 59]. Following the deep learning breakthroughs of the 2010s, methods transitioned from hand-crafted features to learned features, and convolutional neural networks (CNNs) achieved close to human levels of performance on ever-larger datasets [8, 27-29, 32, 36, 48, 63]. For a broader survey of face detection methods, we refer the interested reader to [36, 62]. In our work, we use the recent RetinaFace model [8] as a strong face detection baseline. + +Neuroscience of face pareidolia. The face is a highly unique stimulus for the human visual system [26, 50]: we find faces easy to spot and difficult to ignore. Face detection can occur in both noise and highly degraded images [5]. Prior work shows that face detection occurs in a dedicated brain region, the Fusiform Face Area [33]. But exactly what constitutes a face for the visual cortex and what are the mechanisms underlying pareidolia? A recent study into the temporal dynamics of neuro-imaging data during pareidolic face viewing showed results consistent with "a broadly-tuned face detection mechanism that privileges sensitivity over selectivity" [56]. Pareidolic faces do more than give the impression of the presence of faces: [46] show that they can trigger an additional face-specific attentional process, consuming more time and processing power than similar non-pareidolic stimuli, and even enhancing the detection of face-pareidolic objects [47]. Analyses in [30] revealed a network of neurons in the brain specialized to detect face pareidolia. Their results suggested that face processing has a strong top-down component whereby sensory input with even the slightest suggestion of a face can result in the interpretation of a face. Such top-down information might be supportive of some form of inverse rendering as a cognitive mechanism to explain the remarkable robustness of human perception of faces in degraded viewing conditions [11]. While our dataset allows the study of several types of face detection models, we focus our study on feedforward neural networks which are known to yield close to human performance on challenging "in-the-wild" datasets [60]. + +Face pareidolia in computer vision. Face detection and face recognition have been core topics in computer vision for many decades, but the study of face pareidolia—and its deep relationship with visual object representation learning—has been relatively overlooked. Face pareidolia has some similarities with the problem of cross-modal recognition or cross-depiction: recognizing the + +![](images/2c6c9ce1d33b960aefc866ea82de667f059dbd27dbdb4c0d8e0b9156c43988f6.jpg) + +![](images/c9440c481f248626894a3857bc3bf0ee7db197b841718070524d8008aab6f59b.jpg) + +![](images/ec394d7d13921bcbb2ed32ccb1ec3c17291cde68d790a96fa8c15d4fb55deccf.jpg) + +![](images/cd5d70c018eb7e659accd8d19f5b9ecac68f750deabeece273d5cca2d104779e.jpg) + +![](images/c44a94591afc74fcf11bfec2abbc9543e14fd928dc8f0ecc397f6ba4951b73ee.jpg) + +![](images/b466120961e6f4f3bf79ad33314b316fb42121b985be010676e5233f7c3b1bb4.jpg) + +![](images/7a91bc71b6a0fc47fc2abd7dcda24d2e70a8bfa4938b916428c04c41d8cc0517.jpg) +Fig. 2: Examples of face pareidolia from our "Faces in Things" dataset. Faces in Things consists of five thousand images annotated with bounding boxes (shown here), and facial attributes such as perceived emotion, gender, and intentionality. + +![](images/fed12bb90d7d5dc7cceb66d1c79e1e7b72a84cd2e1673a83f345cd74d3f8452a.jpg) + +![](images/24d54f5511d082e81699aa5cac7fb17905cf784474aa737d64672ef0f43ddbc0.jpg) + +![](images/fc5bbf07dc400fbf7da6e37e7eaf3cf1f7e598befabf815b81fe9a7bec6efe6f.jpg) + +![](images/aead20ea089a04b0a828202de810ef18ce8578e7bffddbbb44df0efa6b9b9a14.jpg) + +![](images/bcccada3f18eba5d5d9bca8a5b7272e56e23c81dd6af447257c2d4b9074cf87a.jpg) + +same objects across different modalities irrespective of how the object is visually depicted. This has been explored particularly in the context of detecting faces, people and objects across modalities such as photography, different art movements, cartoons and sketches [4, 15, 35, 58]. The importance of capturing spatial relationships for robust cross-modal detection has also been highlighted [4]. The work of Castrejon et al. [6] showed how, when learning cross-modal scene representations with neural networks, units would emerge in the shared representation that tended to activate on consistent concepts, independently of the modality. This tendency was used by Abbas & Chalup [1], who found that mid-level units learned during human face detection could generalize to detect semantically similar facial key-points in pareidolic images, showing promise for pareidolia to emerge. However, the authors only evaluated the method qualitatively over a small test set of ten images. One route to a larger dataset may be through pareidolic image generation, which shows promise but does not yet produce convincingly natural images [12]. Curating a larger dataset, "Totally-Looks-Like" [39] explored the perceptual judgment of image similarity between humans and CNNs, using images which had been paired by humans as visually similar but semantically disparate. They found that visual representations extracted from CNNs such as ResNet [18] perform poorly in terms of reproducing the matching selected by humans. Though this dataset is of similar size to ours (6k samples), it is not specifically tailed towards face pareidolia and offers no bounding box or key-point annotations. Other datasets such as COCO-Periph [17] have been used to show that object detection behavior in CNNs and transformers diverge from human perception in peripheral vision. + +In summary, there has yet to be a computational model of how or why pareidolia might arise proposed in the literature. Moreover, despite the abundance of face detection datasets [36], there is no large-scale dataset to directly support the study of face pareidolia. A large-scale pareidolia dataset would help the community explore the mechanisms underlying pareidolia, which may in turn help us to understand and harness human visual attention (which is drawn towards face-like objects), reduce pareidolic false positives in face detectors, help design- + +![](images/381d138c54bb93c8a47c365609893c23b74f1943a15df88948761c921386993f.jpg) + +![](images/e92b075219e7351279b284eefc97d74a999f84de426d3a414611312bbe9aedb2.jpg) + +![](images/f2db8752b921deb79b0ebbf52724d0a97394ad614af6d8f6e2d4209ff46db6ef.jpg) + +![](images/667e0c78a2a5bc39739edcca9f8919f314703945c93c5eb2d9f84d54ec00e464.jpg) +Fig. 3: Attributes of the Faces in Things Dataset. We find that $31\%$ of faces are considered challenging to spot; faces are largely $(31\%)$ judged as happy; approximately half $(47\%)$ are judged as accidental rather than by design; animals and humans are seen in roughly equal numbers; and we observe a slight bias $(16\%$ vs $3\%)$ towards male over female faces, similar to biases observed in prior studies [54,55]. + +![](images/40f4656c5a7e936915bb292a257874ae8eff05271cf043ea0e84bb926ab8aaec.jpg) + +ers avoid or create pareidolia, improve pareidolic animation, and create systems that better understand how humans perceive the world. + +# 3 Faces in Things Dataset + +To address this gap, we begin by sampling candidate pareidolic images from the LAION-5B dataset [42]. This dataset consists of 5.85 billion CLIP-filtered image-text pairs, of which $40\%$ of captions contain English. We use CLIP retrieval [2] to build a raw image set based on text queries including "pareidolia", "faces in things", "accidental faces", and "[object] looks like a face". We download images, check for duplicates, then downsample to $512 \times 512$ pixels while preserving the aspect ratio with white-space padding. We used the VGG Image Annotation tool [10] to manually annotate images, removing samples that contain the faces of actual humans or animals. Some examples of annotated images are shown in Fig. 2. Our annotations include the bounding boxes of pareidolic faces and basic facial attributes as summarized in Fig. 3. Though beyond the scope of the current paper, we note that these attributes could be useful for other future studies. We divide the dataset at random into training ( $70\%$ ) and testing ( $30\%$ ) sets. We refer to this as the 'Pareidolic' dataset. + +# 4 Experiments + +Datasets. We use the following additional datasets. Fig. 4 shows the average faces within our dataset (Pareidolic) and the WIDER FACE (Human), and AnimalWeb (Animal) datasets. + +![](images/b355225945630caf5e4ad7588fe38a29252a3fcf514c6357df5ccd14d6e4b8fc.jpg) +Pareidolic + +![](images/10f44b727bc30bb1bf68f955c155619e737a361d6ecd5925a342806226f81cf7.jpg) +Human +Fig. 4: The Appearance of an Average Pareidolic Face. Per-channel histogram-qualified average images for registered pareidolic faces (our Faces in Things dataset), human faces (samples from the WIDER FACE dataset [60]), and animal faces (AnimalWeb [24]). The average pareidolic face, while less distinct than human or animal, has surprisingly clear eye, nose, and mouth features, and vertical symmetry. + +![](images/7dd3ed7904da2e4d6f9ab13b2c34ef64a2a0af5870e9bfc29a295a399db27a26.jpg) +Animal + +WIDER FACE [60] is a popular face detection benchmark dataset with 32,203 images and 393,703 faces. It contains a high degree of variability in scale, pose, makeup, lighting, emotion, and occlusion, organized across 61 event classes. We use the provided $40\% / 10\% / 50\%$ splits for training, validation, and testing. We refer to this as the 'Human' dataset. + +AnimalWeb [24] is a collection of 22,451 faces from 334 diverse species and 21 animal orders across biological taxonomy. These faces are captured 'in-the-wild' and are consistently annotated with 9 landmarks on key facial features. We convert these landmarks to bounding boxes, by finding the tightest box that captures the points and expanding this box's width and height by $15\%$ . We refer to this as the 'Animal' dataset. + +WIDER FACE Corruptions. To measure whether pareidolia could arise from common data augmentations we corrupt the WIDER FACE images using the level 3 strength of the corruptions used in both the COCO-C [34] and ImageNet-C [19] datasets. We also include a Sobel filtering corruption [13] which has been shown to reduce a model's dependence on texture information [22]. + +Models and Training. We use RetinaFace [8] which achieves state-of-the-art performance on WIDER FACE easy and medium subsets and is the third-best face detector on the hard subset, missing the top model by less than a percentage point of Average Precision (AP). We perform experiments using both their MobileNet [21] and ResNet50 [18] backbones and use the Pytorch_Retinaface [3] repository to ensure the same experimental conditions, dataset characteristics, and preprocessing. We use pre-trained models provided by this repository and fine-tune them for 10 epochs with the AdamW optimizer [31] using a learning rate of $10^{-4}$ and a weight decay of $5 \times 10^{-4}$ . We verify that fine-tuning using this strategy on the original WIDER FACE training dataset does not hurt model performance. When fine-tuning on Faces in Things (Pareidolia), Animal-Web (Animal), WIDER FACE (Human), and Corruption datasets we randomly replace images in the original WIDER FACE stream of training data with data + +![](images/81f85d3b6f3d075405bb4d35b624458249040e69a5854a209fa4a5101ea3ab53.jpg) + +![](images/b6e3b2594a68761b6673c849d4c1975eb99a0542499d534b102b81e649772fd8.jpg) + +![](images/dba4bb603318a3e173a6c694e1a84b057af610916e425e7bd92f26313777fe15.jpg) + +![](images/b991c14e2c09afd338714c7569dee4ed4023fb87f5fc9a0ae988fa7acdc42040.jpg) + +![](images/5f30a7a53706be0d06c02f993d9322ef078d14d64c01fab4ad2916d28a743973.jpg) + +![](images/f2e63abf3778b1c4d5365f2663025ca0a06d5639eee031bd9394a71b45e455b0.jpg) +Fig. 5: Qualitative Analysis of Transfer Experiments. On a sample of held-out test images, we visualize the confident $(p > 10\%)$ detections of our ground truth (red), our model fine-tuned on human faces (blue), and our model fine-tuned on animal faces (green). It is evident from these and Table 1 that fine-tuning on animal faces significantly boosts the model's ability to detect pareidolic faces. + +![](images/43e441d080954bf4c40ec580f72b97edf49c45ac9482bd3c390b2715b0fcfefb.jpg) + +![](images/3190ad5285dbdeef8ff7432865185177621797c123820dce659fa94d29773329.jpg) + +![](images/87aa715e559cbb15179ab6405864f508fc833193c7f0fde8e05a9a4de33a01a5.jpg) + +![](images/1daac9a1541026579175baac1a136e0c0741356d8c58c2c9e31b0847635024ad.jpg) + +from the target dataset $90\%$ of the time. This allows the network to learn the new task without catastrophic forgetting [25]. These changes to the optimizer, learning rate, number of epochs, and stream of training data are the only changes we make to the training paradigm of [8]. Figures in this work use the MobileNet architecture of RetinaNet unless specified otherwise. AP evaluation computations share the same setting and parameters as [8]. + +# 4.1 Does a SOTA Face Detector Exhibit Pareidolia? + +We measure the Average Precision (AP) of the MobileNet and ResNet50 RetinaNet architectures on the Faces in Things dataset. The first row of Table 1 shows results for existing pre-trained models, and the second row shows those for models fine-tuned on the original WIDER FACE training data. These act as control groups to ensure our transfer learning procedure does not interfere with our measurement of the effects of other interventions. Though these models exhibit pareidolia to a small extent, they fall far short of a model fine-tuned to detect pareidolic faces. Fig. 5 also depicts some of these predictions with blue boxes. On the whole, the models trained only on human faces are largely silent across the Faces in Things dataset. + +# 4.2 How Might Pareidolia Emerge? + +The WIDER FACE dataset is known for its diversity of lighting, pose, makeup, emotion, and scale of faces. This fact, coupled with the results of Section 4.1 begs the question: What else do models need to experience pareidolia as humans do? The Faces in Things dataset provides a clean and robust setting to explore + +![](images/3d5c50beae2e3d1a197a94b6d4c0bfc70531368543c1922368e170a4649826ae.jpg) +Fig. 6: Measuring the effect of several training interventions on pareidolic face detection The left plot shows that fine-tuning RetinaNet on animals improves pareidolic face detection more than any other intervention. Conversely, the right plot shows that pareidolic fine-tuning improves animal face detection performance. + +![](images/a364ae7066474aaa2c442a6555a8edad35f41f1e3f6b43191897d9b923c9118c.jpg) + +the development of pareidolia in algorithms. Unlike in humans, where it is impossible to causally intervene on their facial training data, we can easily modify an algorithm's training data. This makes it possible to explore whether one can induce pareidolia in algorithms using specific stimuli. + +To this end, we investigate whether a variety of training data interventions can induce pareidolia in algorithms. In particular, we measure the effect of adding several data augmentations from the COCO-C [34] and ImageNet-C [19] datasets and explore a Sobel filtering augmentation which reduces models' dependence on texture. Additionally, we also measure the effect of adding animal faces to the training data. Animal faces show a far greater breadth of variation in coloration, structure, and appearance than human faces. Recognizing animal faces provides many evolutionary advantages including gaze detection during hunting and avoiding onlooking predators. The generality required to detect this wide space of faces could yield a greater number of "false positives" that lead to the sensation of pareidolia. Indeed, some recent studies provide some corroborating evidence for this hypothesis. Firstly, Rhesus Monkeys exhibit pareidolia [49] showing this effect does not only occur in humans. Secondly, the experience of pareidolia is a rapid cognitive process and not a "late re-interpretation" of input signals [16, 56] which the authors conclude is evidence that pareidolia could be linked to the need to quickly react to predators. + +We plot the change to pareidolic face detection performance as a function of each training intervention in the left panel of Figure 6. Of the different corruption interventions, we find that Sobel filtering, motion blur, Gaussian noise, and fog tend to slightly improve pareidolic face detection while most other corruptions do not improve pareidolic face detection performance. Most strikingly, the addition of animal faces to the training data roughly doubles the algorithms' ability to detect pareidolic faces compared to the control group, closing around half of the gap between a human-trained model and a pareidolia trained model. + +
FinetuningAP
MobileNetResNet50
None7.9%2.8%
Human (Control)9.8%3.6%
Animal16.7%15.4%
Pareidolia33.9%27.1%
Animal + Pareidolia36.4%31.7%
+ +Table 1: Effect of Fine Tuning on Pareidolic Face Detection. Our results show that WIDER FACE-trained RetinaFace models do not detect many pareidolic images. Fine-tuning these models on animal faces approximately doubles pareidolic face detection rates. Interestingly, adding animal faces alongside pareidolic faces (30%/70% split respectively) can improve performance over fine-tuning on pareidolic faces alone. + +Reciprocally, the right-hand plot of Figure 6 shows that fine-tuning on pareidolic images yields the greatest improvement in animal face detection. We further explore this phenomenon in Table 1, where we show that this effect occurs across both MobileNet and ResNet50 architectures. Finally, we also show that adding a small number of animal faces (30% animal 70% pareidolic) can improve pareidolic face detection performance over pareidolic images alone. + +To understand this effect better, Fig. 7 visualizes the inner representations of this model across the three datasets (Human = WIDER FACE, Animal = AnimalWeb, Pareidolic = Faces in Things). Specifically, we extract multi-scale features from the animal and pareidolia fine-tuned RetinaNet shared feature layer before the application of the classification and regression heads. We average pool these features across the bounding box for each face and visualize them with t-SNE [20]. This figure shows that RetinaNet's representations of animal and pareidolic faces tend to cluster together and are distinct from its representations of human faces. This lends evidence to the relative similarity of pareidolic and animal faces compared to human faces. We also reiterate that we filtered the Faces in Things dataset to avoid images of real animals. + +# 5 Modeling Pareidolia + +Though many prior works have measured pareidolia, there has yet to be a simple mathematical model that describes the high-level structure of this phenomenon. In this section we provide two simple formal models of pareidolia and show that they both exhibit a testable prediction: the existence of a peak in pareidolic face detection as a function of an image's complexity. Section 5.4 presents experimental evidence of this "pareidolic peak" in both humans and machines. + +# 5.1 Gaussian Model of Pareidolia + +A model of pareidolia needs to describe two processes: (1) the random process that generates candidate images, and (2) the face detection process which determines when an image is pareidolic. We begin with a simple Gaussian model for + +![](images/3ed09d02fe6255fd7f0a9ace6c40ba6d1a19d1d8f20ab1df02d7d8862da99be6.jpg) +Fig. 7: Visualizing RetinaNet Representations across Datasets. Animal+Pareidolia fine-tuned RetinaNet representations tend to group animal and pareidolic faces together. This lends evidence to the hypothesis that the perception of animal and pareidolic faces are linked. (To highlight the commonality of pareidolic animal detection we note the similarity of these points to a frog.) + +each. We model the image generation process as a sum of independent normal modes, each contributing a zero-mean Gaussian of a specified variance, multiplied by the mode image $y_{i}$ . For example, as in [9, 51], these modes could be the principal components of a mean-subtracted image dataset. In this setting the generated image, $y$ , is a weighted sum of the normal modes: + +$$ +\boldsymbol {y} = \sum_ {i} n _ {i} \boldsymbol {y} _ {i} \quad \text {w h e r e}, \quad n _ {i} \sim N (0, \sigma_ {i}) \tag {1} +$$ + +To model our face detection process we capture the intuition of matching an image to a template image and note that this can be generalized to distributions of template images. In particular, the target pareidolic image is represented as a vector, $\pmb{a}$ , of statistically independent target coefficients, $a_{i}$ , for each mode. The probability that this mode contributes towards the face detection, $P(a_{i})$ , is the probability of detecting the pareidolic value, $a_{i}$ , at the $i$ th mode. We assume a Gaussian detection process: $P(a_{i}|y_{i}) \sim N(a_{i}, \gamma_{i})$ . Because each mode's coefficient is a zero-mean Gaussian distribution, $P(y_{i}) \sim N(0, \sigma)$ , we have: + +$$ +\begin{array}{l} P \left(a _ {i}\right) = \int_ {y _ {i}} P \left(a _ {i}, y _ {i}\right) \mathrm {d} y _ {i} (2) \\ = \int_ {y _ {i}} P \left(a _ {i} \mid y _ {i}\right) P \left(y _ {i}\right) \mathrm {d} y _ {i} (3) \\ = \frac {1}{2 \pi \gamma_ {i} \sigma_ {i}} \int_ {y _ {i}} e ^ {- \frac {(y _ {i} - a _ {i}) ^ {2}}{2 \gamma_ {i} ^ {2}}} e ^ {- \frac {y _ {i} ^ {2}}{2 \sigma^ {2}}} \mathrm {d} y _ {i} (4) \\ \end{array} +$$ + +![](images/899529a899e718c0b32f4bf85ae3787bee3b1aa6decae8eafedf09ba9775a7c4.jpg) +Fig. 8: Illustration of the proposed Gaussian model for pareidolia with three example generating distributions. To make pareidolia likely, the generating distribution needs a proper distribution of spatial frequencies. A process with too few spatial frequencies (left) is likely to only generate weak face-like details ("face-ness": low). In contrast, with too many frequencies (right), faces can be modeled with exquisite detail ("face-ness": high), but the likelihood of drawing any particular desired combination become vanishingly small. The most likely pareidolic images form when the generating distribution has the right spectrum (middle), enabling reasonable faces to emerge with reasonable likelihood. In other words, this model predicts that pareidolic faces will match the low frequencies of faces but differ in the higher frequency details. + +Note that $\sigma_i^2$ is the variance of the random process generating the pareidolia, while $\gamma_i^2$ is the variance of the likelihood term — how far a mode is allowed to vary from the target mode value before it stops looking like the target image, $\pmb{a}$ . + +We can complete the square in Eq. 4 to write the product of Gaussians in $P(a_{i})$ as a single Gaussian. Integrating that Gaussian over all possible observations $y_{i}$ gives the probability of finding the pareidolic value $a_{i}$ from mode $i$ : + +$$ +P (a _ {i}) = \frac {1}{\sqrt {2 \pi (\gamma_ {i} ^ {2} + \sigma_ {i} ^ {2})}} e ^ {- \frac {a _ {i} ^ {2}}{2 (\sigma_ {i} ^ {2} + \gamma_ {i} ^ {2})}} \tag {5} +$$ + +# 5.2 Predicting Peak Pareidolia + +For a given mode's detection variance, $\gamma_i^2$ , and target mode coefficient, $a_i$ , Eq. 4 allows us to find the optimal mode variance to generate pareidolia, i.e., to maximize $P(a_i)$ . Unfortunately, we seldom have the flexibility to design a random process one mode at a time. But we may have the option to select between image generation processes that have different numbers of modes, $M$ . Since each mode + +![](images/3278e6b04f301679f9177bab9716292e2ccee0e6299185ecbe679e8961c01c25.jpg) +Fig. 9: Probability of pareidolia (Eq. 6) in the Gaussian model $(\sigma = 10)$ across images with different spatial frequency distributions (Sec. 5.4). This assumes spatial frequencies are uncorrelated and thus underestimates the probability of pareidolia, however peak pareidolia is still present. + +![](images/56d8b68dea877bd85cd69a98be77bd3e6d8a994945f7650a41100c8e0bcec897.jpg) +Fig. 10: Probability of pareidolia under the feature-based example of Eq. (8) as a function of the rate of feature detection, $\lambda_{i} = \lambda$ for all $i$ , within the random images. Note the low probability of pareidolia for both feature-free $(\lambda \to 0)$ and feature-rich $(\lambda \gg 0)$ random images. + +is independent, the probability of a pareidolic detection of the target object template is the product of detecting the target coefficient for each of the M modes: + +$$ +P (a) = \prod_ {i} ^ {M} P \left(a _ {i}\right) \tag {6} +$$ + +We plot some predictions of our Gaussian model in Fig. 9 for a target template with a $\frac{1}{f}$ power spectrum (standard deviation of each mode inversely proportional to mode number) on noise images of varying complexity. We note the existence of a peak in pareidolic detection probability for random image generation processes with a mid-range number of spatial modes, as measured by the width of Gaussian that modulates power in Fourier space. Too few modes in the random generation process, and no image will ever have enough complexity to render the target well. Too many modes and pareidolia becomes unlikely because so many modes need to match the desired target values. Each added mode multiplies the pareidolia probability by another small factor. In between, there is what we call peak pareidolia. As the detection value, $\sigma$ , becomes more stringent (smaller) the peak pareidolia value occurs at a larger number of modes and becomes less probable. We illustrate this effect in Figure 8 + +# 5.3 Higher-Level Feature Model of Pareidolia + +The Gaussian model for pareidolia above lays out important aspects of pareidolia, but relies on a naive model of object detection, the squared distance from a template image. We assume that a more realistic model of human perception would incorporate higher-level features and introduce a still simple, yet more realistic, feature-based model. + +We assume that the detection of an object requires particular features to be detected in certain spatial regions, e.g. an eye in the top left and right, a nose in the center, and a mouth in the bottom. Such an approach has been used in computer vision object detection algorithms, e.g. [14, 57, 61]. Any given object template has some number of regions, $R_{i}$ , indexed by $i$ , within which a given feature, $F_{i}$ , must be detected. The other features, $F_{j\neq i}$ , should not be detected in region $i$ . For a given random image where we hope to detect pareidolia, we assume that feature existence is a spatial Poisson process. In this process, the probability of $n$ feature instances for any given feature $i$ over some area $B_{i}$ is + +$$ +P (n _ {i}) = \frac {\left(\lambda_ {i} | B _ {i} |\right) ^ {n _ {i}}}{n _ {i} !} e ^ {- \lambda_ {i} | B _ {i} |} \tag {7} +$$ + +To detect a pareidolic instance of the object template, we must detect one feature of the correct type, $F_{i}$ , in each region $i$ of the face template, and zero features of the wrong type, $F_{j\neq i}$ in each region $i$ . Assuming independence of the feature detections, and for simplicity setting all the feature detection rates to be the same, $\lambda_{i} = \lambda$ , and all the template areas to be the same, $B_{i} = 1$ , we have for the probability, $P(O)$ , of pareidolic detection of object $O$ : + +$$ +P (O) = \prod_ {i} ^ {M} \lambda^ {1} e ^ {- \lambda (M - 1)} \tag {8} +$$ + +For the case of $M = 4$ , a simple detection model for two eyes, a nose, and a mouth, we have $P(o) = \lambda^4 e^{-16\lambda}$ , which is plotted in Fig. 10. In this feature-based object detection model, we also find the existence of "peak pareidolia". Again, it is governed by a parameter describing the complexity of the random image, in this case a Poisson process rate parameter, $\lambda$ , that governs the probability of a feature detection per unit area. For too low a rate, the model doesn't generate enough features to satisfy the object template, for too high a rate, the probability of seeing only the right features in just the right places becomes very small. In between is the most probable rate for pareidolia. + +# 5.4 Measuring the Pareidolic Peak in Humans and Machines + +Both mathematical models of Section 5 predict the existence of a peak of pareidolic face detection as a function of image complexity. We show the existence of this pareidolic peak in both humans and machines. In particular, we perform a psychophysics experiment where human subjects view noise images of varying complexity and report how many pareidolic faces they saw in each image, from zero to nine. Campbell [5] demonstrated that a 12x12 array of random, binary squares is sufficient to evoke human and animal faces. We generate noise images of varying complexity by randomly sampling Fourier coefficients and modulating these coefficients with a zero mean $\sigma^2$ variance Gaussian in Fourier space. We show some samples of these images on the x-axis of Fig. 11. Intuitively, the Gaussian envelope in frequency space filters out most frequencies higher than $\sigma$ + +![](images/5c27903441e7a250a8d902412ab58a269b2f1c382c1d60fca216919c893ba5e7.jpg) +Fig. 11: Measuring Peak Pareidolia. Left: Subjects were asked how many faces they see in each noise image. We plot the average number of faces detected as a function of noise frequency (examples on x-axis), the mean over all subjects and its $95\%$ confidence interval in red. Right: average number of faces detected by our fine-tuned models. This reveals the "peak pareidolia" effect predicted in Section 5 across humans and machines. + +![](images/fa61f090fc589435100f26d5052696fcc4096df385ab7978f92b8a49e1628b2f.jpg) + +after applying an inverse Fourier transform. We detail our image stimuli creation method in the Supplement. + +We find that humans exhibit the model-expected peak pareidolia, with a maximum number of faces detected at a frequency filter width of 16 (Fig. 11, left). The existence of a pareidolic peak at or near this filter width is consistent among all subjects even for those that reported fewer faces overall. Although response time did decrease slightly at higher frequencies, it did not fall off completely at the highest frequency levels, indicating that fewer reported faces were not the result of subjects "giving up" on the task. We provide additional details and analysis of this experiment in the Supplement. + +Finally, we evaluate our fine-tuned models from Section 4 on the same images to test whether machines also exhibit peak pareidolia (Fig. 11, right). In particular, we showed the models 5,000 sampled noise images of varying frequency levels and counted the number of face detections they make with confidence $>10\%$ . We find the same characteristic "pareidolic peak" where models detect the most faces in medium-complexity images. + +# 6 Conclusion + +We have taken initial steps towards the mathematical modelling of pareidolia and build a richly annotated dataset of images for face pareidolia. We showed through experiments on modern face detectors that detecting animal faces may partly explain the emergence of pareidolia in a complex vision system. The Faces in Things dataset can help the community address other questions about how and why pareidolic behavior emerges, a hallmark of humans' robust recognition system. We hope that our findings and dataset will spark further study of pareidolia and its potential use to improve computer vision systems. + +# Acknowledgments + +We would like to thank Karen Hamilton for her hundreds of hours of annotations for the Faces in Things dataset. We also thank Abhishek Dutta who created the VGG Image Annotation tool, for his kind generosity to support our project. + +We would like to thank the Microsoft Research Grand Central Resources team for their gracious help performing the experiments in this work. Special thanks to Oleg Losinets and Lifeng Li for their consistent, gracious, and timely help, debugging, and expertise. Without them, none of the experiments could have been run. + +This material is based upon work supported by the National Science Foundation Graduate Research Fellowship under Grant No. 2021323067. Any opinion, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of their employers, or the National Science Foundation. + +This work is supported by the National Science Foundation under Cooperative Agreement PHY-2019786 (The NSF AI Institute for Artificial Intelligence and Fundamental Interactions, http://iaifi.org/) and the CSAIL MEnTorEd Opportunities in Research (METEOR) Fellowship. + +Research was sponsored by the United States Air Force Research Laboratory and the United States Air Force Artificial Intelligence Accelerator and was accomplished under Cooperative Agreement Number FA8750-192-1000. The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the United States Air Force or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for Government purposes notwithstanding any copyright notation herein. + +The authors acknowledge the MIT SuperCloud [38] and Lincoln Laboratory Supercomputing Center for providing HPC resources that have contributed to the research results reported within this paper. + +# References + +1. Abbas, A., Chalup, S.: From Face Recognition to Facial Pareidolia: Analysing Hidden Neuron Activations in CNNs for Cross-Depiction Recognition. In: 2019 International Joint Conference on Neural Networks (IJCNN). pp. 1–8 (Jul 2019). https://doi.org/10.1109/IJCNN.2019.8852013, iSSN: 2161-4393 +2. Beaumont, R.: Clip retrieval: Easily compute clip embeddings and build a clip retrieval system with them. https://github.com/rom1504/clip-retrieval (2022) +3. biubug6: Retinaface in pytorch (Nov 2021), https://github.com/biubug6/Pytorch_Retinaface +4. Cai, H., Wu, Q., Corradi, T., Hall, P.: The Cross-Depiction Problem: Computer Vision Algorithms for Recognising Objects in Artwork and in Photographs. arXiv:1505.00110 [cs] (May 2015), http://arxiv.org/abs/1505.00110, arXiv: 1505.00110 +5. Campbell, F.: How much of the information falling on the retina reaches the visual cortex and how much is stored in the visual memory. Pattern recognition mechanisms 54, 83-94 (1983) + +6. Castrejon, L., Aytar, Y., Vondrick, C., Pirsiavash, H., Torralba, A.: Learning Aligned Cross-Modal Representations from Weakly Aligned Data. In: CVPR (Jun 2016) +7. Conrad, K.: Die beginnende Schizophrenie. Versuch einer Gestaltanalyse des Wahns (1958) +8. Deng, J., Guo, J., Zhou, Y., Yu, J., Kotsia, I., Zafeiriou, S.: RetinaFace: Single-stage Dense Face Localisation in the Wild. CVPR (2020) +9. Duda, R., Hart, P., Stork, D.: Pattern Classification. Wiley (2012) +10. Dutta, A., Zisserman, A.: The VIA annotation software for images, audio and video. In: Proceedings of the 27th ACM International Conference on Multimedia. MM '19, ACM, New York, NY, USA (2019). https://doi.org/10.1145/3343031.3350535, https://doi.org/10.1145/3343031.3350535 +11. Egger, B., Siegel, M.H., Arora, R., Soltani, A.A., Yildirim, I., Tenenbaum, J.: Inverse rendering best explains face perception under extreme illuminations. In: CogSci (2020) +12. Endo, Y., Asanuma, R., Shimojo, S., Akashi, T.: Systematic face pareidolia generation method using cycle-consistent adversarial networks. IEEE Transactions on Electrical and Electronic Engineering 19(4), 535-541 (2024) +13. Farid, H., Simoncelli, E.P.: Optimally rotation-equivariant directional derivative kernels. In: International Conference on Computer Analysis of Images and Patterns. pp. 207-214. Springer (1997) +14. Felzenszwalb, P., Girshick, R., McAllester, D., Ramanan, D.: Object detection with discriminatively trained part based models. IEEE TPAMI 32(9) (2010) +15. Ginosar, S., Haas, D., Brown, T., Malik, J.: Detecting people in cubist art. In: ECCV Workshops (2014) +16. Hadjikhani, N., Kveraga, K., Naik, P., Ahlfors, S.P.: Early (n170) activation of face-specific cortex by face-like objects. Neuroreport 20(4), 403 (2009) +17. Harrington, A., DuTell, V., Hamilton, M., Tewari, A., Stent, S., Freeman, W.T., Rosenholtz, R.: Coco-periph: Bridging the gap between human and machine perception in the periphery. In: The Twelfth International Conference on Learning Representations +18. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016) +19. Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. Proceedings of the International Conference on Learning Representations (2019) +20. Hinton, G.E., Roweis, S.: Stochastic neighbor embedding. Advances in neural information processing systems 15 (2002) +21. Howard, A.G., Zhu, M., Chen, B., Kalenichenko, D., Wang, W., Weyand, T., Andreetto, M., Adam, H.: Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861 (2017) +22. Ji, X., Henriques, J.F., Vedaldi, A.: Invariant information clustering for unsupervised image classification and segmentation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9865-9874 (2019) +23. Kato, M., Mugitani, R.: Pareidolia in Infants. PLoS ONE 10(2) (Feb 2015). https://doi.org/10.1371/journal.pone.0118539, https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4331561/ +24. Khan, M.H., McDonagh, J., Khan, S., Shahabuddin, M., Arora, A., Khan, F.S., Shao, L., Tzimiropoulos, G.: AnimalWeb: A large-scale hierarchical dataset of annotated animal faces. In: CVPR (2020) + +25. Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A.A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., et al.: Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences 114(13), 3521-3526 (2017) +26. Leopold, D.A., Rhodes, G.: A comparative view of face perception. Journal of Comparative Psychology 124(3), 233 (2010) +27. Li, H., Lin, Z., Shen, X., Brandt, J., Hua, G.: A convolutional neural network cascade for face detection. In: CVPR (2015) +28. Li, J., Wang, Y., Wang, C., Tai, Y., Qian, J., Yang, J., Wang, C., Li, J., Huang, F.: DSFD: dual shot face detector. In: CVPR (2019) +29. Liao, S., Jain, A.K., Li, S.Z.: A Fast and Accurate Unconstrained Face Detector. PAMI 38(2), 211-223 (Feb 2016). https://doi.org/10.1109/TPAMI.2015.2448075 +30. Liu, J., Li, J., Feng, L., Li, L., Tian, J., Lee, K.: Seeing Jesus in toast: Neural and behavioral correlates of face pareidolia. Cortex 53, 60-77 (Apr 2014). https://doi.org/10.1016/j.cortex.2014.01.013, http://www.sciencedirect.com/science/article/pii/S0010945214000288 +31. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017) +32. Mathias, M., Benenson, R., Pedersoli, M., Van Gool, L.: Face detection without bells and whistles. In: ECCV (2014) +33. Mcgugin, R., Gatenby, C., Gore, J., Gauthier, I.: High-resolution imaging of expertise reveals reliable object selectivity in the FFA related to perceptual performance. Proceedings of the National Academy of Sciences of the United States of America 109, 17063-8 (Oct 2012). https://doi.org/10.1073/pnas.1116333109 +34. Michaelis, C., Mitzkus, B., Geirhos, R., Rusak, E., Bringmann, O., Ecker, A.S., Bethge, M., Brendel, W.: Benchmarking robustness in object detection: Autonomous driving when winter is coming. arXiv preprint arXiv:1907.07484 (2019) +35. Mishra, A., Nandan Rai, S., Mishra, A., Jawahar, C.V.: Iit-cfw: A benchmark database of cartoon faces in the wild. In: VASE ECCV Workshops (2016) +36. Ranjan, R., Sankar, S., Bansal, A., Bodla, N., Chen, J.C., Patel, V., Castillo, C., Chellappa, R.: Deep Learning for Understanding Faces: Machines May Be Just as Good, or Better, than Humans. IEEE Signal Processing Magazine 35, 66-83 (Jan 2018). https://doi.org/10.1109/MSP.2017.2764116 +37. Rein-Lien Hsu, Abdel-Mottaleb, M., Jain, A.: Face detection in color images. PAMI 24(5), 696-706 (May 2002). https://doi.org/10.1109/34.1000242, http://ieeexplore.ieee.org/document/1000242/ +38. Reuther, A., Kepner, J., Byun, C., Samsi, S., Arcand, W., Bestor, D., Bergeron, B., Gadepally, V., Houle, M., Hubbell, M., Jones, M., Klein, A., Milechin, L., Mullen, J., Prout, A., Rosa, A., Yee, C., Michaleas, P.: Interactive supercomputing on 40,000 cores for machine learning and data analysis. In: 2018 IEEE High Performance extreme Computing Conference (HPEC). pp. 1-6. IEEE (2018) +39. Rosenfeld, A., Solbach, M.D., Tsotsos, J.K.: Totally looks like-how humans compare, compared to machines. In: ACCV (2018) +40. Rowley, H.A., Baluja, S., Kanade, T.: Human Face Detection in Visual Scenes. In: Touretzky, D.S., Mozer, M.C., Hasselmo, M.E. (eds.) NeurIPS (1996), http://papers.nips.cc/paper/1168-human-face-detection-in-visual-scenes.pdf +41. Rowley, H.A., Baluja, S., Kanade, T.: Neural network-based face detection. PAMI 20(1), 23-38 (1998) + +42. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C.W., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., Schramowski, P., Kundurthy, S.R., Crowson, K., Schmidt, L., Kaczmarczyk, R., Jitsev, J.: LAION5b: An open large-scale dataset for training next generation image-text models. In: NeurIPS Datasets and Benchmarks Track (2022), https://openreview.net/forum?id=M3Y74vmsMcY +43. Shakespeare, W.: The Tragedy of Hamlet, Prince of Denmark. The Folio Society (1954) +44. Sims, A.: Symptoms in the mind: An introduction to descriptive psychopathology. Bailliere Tindall Publishers (1988) +45. Summerfield, C., Egner, T., Mangels, J., Hirsch, J.: Mistaking a house for a face: neural correlates of misperception in healthy humans. Cerebral cortex 16(4), 500-508 (2006) +46. Takahashi, K., Watanabe, K.: Gaze Cueing by Pareidolia Faces. i-Perception 4(8), 490-492 (Dec 2013). https://doi.org/10.1068/i0617sas, https://doi.org/10.1068/i0617sas +47. Takahashi, K., Watanabe, K.: Seeing Objects as Faces Enhances Object Detection. i-Perception 6(5) (2015). https://doi.org/10.1177/2041669515606007, https://doi.org/10.1177/2041669515606007 +48. Tang, X., Du, D.K., He, Z., Liu, J.: PyramidBox: A Context-Assisted Single Shot Face Detector. In: Ferrari, V., Hebert, M., Sminchisescu, C., Weiss, Y. (eds.) ECCV (2018) +49. Taubert, J., Wardle, S.G., Flessert, M., Leopold, D.A., Ungerleider, L.G.: Face Pareidolia in the Rhesus Monkey. Current Biology 27(16), 2505-2509 (Aug 2017). https://doi.org/10.1016/j.cub.2017.06.075, http://www.sciencedirect.com/science/article/pii/S0960982217308126 +50. Tsao, D.Y., Livingstone, M.S.: Mechanisms of face perception. Annu. Rev. Neurosci. 31, 411-437 (2008) +51. Turk, M., Pentland, A.: Eigenfaces for recognition. J. of Cognitive Neuroscience 3(1) (1991) +52. Viola, P., Jones, M.: Rapid object detection using a boosted cascade of simple features. In: CVPR (2001) +53. Viola, P., Jones, M.J.: Robust real-time face detection. IJCV 57(2), 137-154 (2004) +54. Wardle, S.G., Ewing, L., Malcolm, G.L., Paranjape, S., Baker, C.I.: Children perceive illusory faces in objects as male more often than female. Cognition 235 (2023). https://doi.org/https://doi.org/10.1016/j.cognition.2023.105398, https://www.sciencedirect.com/science/article/pii/S001002772300032X +55. Wardle, S.G., Paranjape, S., Taubert, J., Baker, C.I.: Illusory faces are more likely to be perceived as male than female. Proceedings of the National Academy of Sciences 119(5) (2022). https://doi.org/10.1073/pnas.2117413119, https://www.pnas.org/doi/abs/10.1073/pnas.2117413119 +56. Wardle, S.G., Taubert, J., Teichmann, L., Baker, C.I.: Rapid and dynamic processing of face pareidolia in the human brain. Nature communications 11(1), 4518 (2020) +57. Weber, M., Welling, M., Perona, P.: Unsupervised learning of models for recognition. In: ECCV (2000) +58. Westlake, N., Cai, H., Hall, P.: Detecting People in Artwork with CNNs. In: ECCV Workshops (2016) +59. Yang, M.H., Kriegman, D., Ahuja, N.: Detecting faces in images: a survey. PAMI 24(1), 34-58 (Jan 2002). https://doi.org/10.1109/34.982883 + +60. Yang, S., Luo, P., Loy, C.C., Tang, X.: WIDER FACE: A Face Detection Benchmark. In: CVPR (2016) +61. Yuille, A.L.: Deformable templates for face recognition. Journal of Cognitive Neuroscience 3(1), 59-70 (1991) +62. Zafeiriou, S., Zhang, C., Zhang, Z.: A survey on face detection in the wild: Past, present and future. Computer Vision and Image Understanding 138, 1-24 (Sep 2015). https://doi.org/10.1016/j.cviu.2015.03.015, https://linkinghub.elsevier.com/retrieve/pii/S1077314215000727 +63. Zhang, S., Zhu, X., Lei, Z., Shi, H., Wang, X., Li, S.Z.: S3fd: Single shot scale-invariant face detector. In: ICCV (2017) \ No newline at end of file diff --git a/2024/Seeing Faces in Things_ A Model and Dataset for Pareidolia/images.zip b/2024/Seeing Faces in Things_ A Model and Dataset for Pareidolia/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..831ecd0e59993b053638677e402f0968e5d5c837 --- /dev/null +++ b/2024/Seeing Faces in Things_ A Model and Dataset for Pareidolia/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0cdbf8404459c8e2b66e1fbc46ca36c552b84639fcd7a7cd1838dd0c1ff06e94 +size 457612 diff --git a/2024/Seeing Faces in Things_ A Model and Dataset for Pareidolia/layout.json b/2024/Seeing Faces in Things_ A Model and Dataset for Pareidolia/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..61de0605d211a0318171162fefedd8c6bf8b1a28 --- /dev/null +++ b/2024/Seeing Faces in Things_ A Model and Dataset for Pareidolia/layout.json @@ -0,0 +1,10085 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 80, + 30, + 332, + 64 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 30, + 332, + 64 + ], + "spans": [ + { + "bbox": [ + 80, + 30, + 332, + 64 + ], + "type": "text", + "content": "Seeing Faces in Things: A Model and Dataset for Pareidolia" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 42, + 86, + 370, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 42, + 86, + 370, + 111 + ], + "spans": [ + { + "bbox": [ + 42, + 86, + 370, + 111 + ], + "type": "text", + "content": "Mark Hamilton" + }, + { + "bbox": [ + 42, + 86, + 370, + 111 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 42, + 86, + 370, + 111 + ], + "type": "text", + "content": ", Simon Stent" + }, + { + "bbox": [ + 42, + 86, + 370, + 111 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 42, + 86, + 370, + 111 + ], + "type": "text", + "content": ", Vasha DuTell" + }, + { + "bbox": [ + 42, + 86, + 370, + 111 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 42, + 86, + 370, + 111 + ], + "type": "text", + "content": ", Anne Harrington" + }, + { + "bbox": [ + 42, + 86, + 370, + 111 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 42, + 86, + 370, + 111 + ], + "type": "text", + "content": ", Jennifer Corbett" + }, + { + "bbox": [ + 42, + 86, + 370, + 111 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 42, + 86, + 370, + 111 + ], + "type": "text", + "content": ", Ruth Rosenholtz" + }, + { + "bbox": [ + 42, + 86, + 370, + 111 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 42, + 86, + 370, + 111 + ], + "type": "text", + "content": ", and William T. Freeman" + }, + { + "bbox": [ + 42, + 86, + 370, + 111 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 82, + 119, + 329, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 119, + 329, + 131 + ], + "spans": [ + { + "bbox": [ + 82, + 119, + 329, + 131 + ], + "type": "text", + "content": "1 MIT, 2 Microsoft, 3 Toyota Research Institute, 4 NVIDIA" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 159, + 351, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 159, + 351, + 346 + ], + "spans": [ + { + "bbox": [ + 61, + 159, + 351, + 346 + ], + "type": "text", + "content": "Abstract. The human visual system is well-tuned to detect faces of all shapes and sizes. While this brings obvious survival advantages, such as a better chance of spotting unknown predators in the bush, it also leads to spurious face detections. \"Face pareidolia\" describes the perception of face-like structure among otherwise random stimuli: seeing faces in coffee stains or clouds in the sky. In this paper, we study face pareidolia from a computer vision perspective. We present an image dataset of \"Faces in Things\", consisting of five thousand web images with human-annotated pareidolic faces. Using this dataset, we examine the extent to which a state-of-the-art human face detector exhibits pareidolia, and find a significant behavioral gap between humans and machines. We find that the evolutionary need for humans to detect animal faces, as well as human faces, may explain some of this gap. Finally, we propose a simple statistical model of pareidolia in images. Through studies on human subjects and our pareidolic face detectors we confirm a key prediction of our model regarding what image conditions are most likely to induce pareidolia. Dataset and Website: https://aka.ms/faces-in-things" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 357, + 322, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 357, + 322, + 368 + ], + "spans": [ + { + "bbox": [ + 59, + 357, + 322, + 368 + ], + "type": "text", + "content": "Keywords: Pareidolia " + }, + { + "bbox": [ + 59, + 357, + 322, + 368 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 59, + 357, + 322, + 368 + ], + "type": "text", + "content": " Face Detection " + }, + { + "bbox": [ + 59, + 357, + 322, + 368 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 59, + 357, + 322, + 368 + ], + "type": "text", + "content": " Human Psychophysics" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 32, + 388, + 128, + 400 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 388, + 128, + 400 + ], + "spans": [ + { + "bbox": [ + 32, + 388, + 128, + 400 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 30, + 415, + 328, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 415, + 328, + 426 + ], + "spans": [ + { + "bbox": [ + 30, + 415, + 328, + 426 + ], + "type": "text", + "content": "Hamlet: Do you see yonder cloud that's almost in the shape of a camel?" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 30, + 426, + 247, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 426, + 247, + 436 + ], + "spans": [ + { + "bbox": [ + 30, + 426, + 247, + 436 + ], + "type": "text", + "content": "Polonius: By the Mass and 'tis, like a camel indeed." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 30, + 437, + 167, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 437, + 167, + 447 + ], + "spans": [ + { + "bbox": [ + 30, + 437, + 167, + 447 + ], + "type": "text", + "content": "Hamlet: Methinks it is a weasel." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 30, + 448, + 179, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 448, + 179, + 457 + ], + "spans": [ + { + "bbox": [ + 30, + 448, + 179, + 457 + ], + "type": "text", + "content": "Polonius: It is back'd like a weasel." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 30, + 459, + 137, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 459, + 137, + 468 + ], + "spans": [ + { + "bbox": [ + 30, + 459, + 137, + 468 + ], + "type": "text", + "content": "Hamlet: Or like a whale." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 30, + 470, + 149, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 470, + 149, + 480 + ], + "spans": [ + { + "bbox": [ + 30, + 470, + 149, + 480 + ], + "type": "text", + "content": "Polonius: Very like a whale." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 174, + 487, + 380, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 487, + 380, + 499 + ], + "spans": [ + { + "bbox": [ + 174, + 487, + 380, + 499 + ], + "type": "text", + "content": "— Hamlet, Act III, Scene ii, William Shakespeare" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 30, + 511, + 380, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 511, + 380, + 584 + ], + "spans": [ + { + "bbox": [ + 30, + 511, + 380, + 584 + ], + "type": "text", + "content": "Pareidolia is a type of visual \"apophenia\", which refers to the perception of patterns in random data. This occurs frequently in human perception as we look at clouds, mountain skylines, and burnt toast. Pareidolia is even described in an exchange in Hamlet [43]. When it was first described, pareidolia was seen as an early symptom of psychosis [7, 44]. Today we know pareidolia is common among healthy humans [45] and infants [23]. It is also not confined to humans:" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [], + "page_size": [ + 413, + 615 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 33, + 43, + 166, + 158 + ], + "blocks": [ + { + "bbox": [ + 33, + 43, + 166, + 158 + ], + "lines": [ + { + "bbox": [ + 33, + 43, + 166, + 158 + ], + "spans": [ + { + "bbox": [ + 33, + 43, + 166, + 158 + ], + "type": "image", + "image_path": "f78aab5b9190439d9daef9925f38018d7e676cab28517fbb93a0561c8b304a7e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 169, + 33, + 381, + 164 + ], + "lines": [ + { + "bbox": [ + 169, + 33, + 381, + 164 + ], + "spans": [ + { + "bbox": [ + 169, + 33, + 381, + 164 + ], + "type": "text", + "content": "Fig. 1: You print out an exciting new computer vision paper to review, but as you sit down at your desk to start reading you knock over your coffee cup. At first, you are annoyed, but then, you laugh! The sight of the stain induces \"pareidolia\" in your brain: rather than an unsightly blemish, you see a happy face. In this paper we explore the phenomenon of face pareidolia: Why don't we see faces all the time? Why do we see them at all when they are clearly so different from human faces? Can a better understanding of face pareidolia help computer vision-based face detection?" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 32, + 199, + 379, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 199, + 379, + 223 + ], + "spans": [ + { + "bbox": [ + 32, + 199, + 379, + 223 + ], + "type": "text", + "content": "rhesus macaques, for example, have been shown to spend more time fixating on pareidolic than non-pareidolic images, in a manner similar to humans [49]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 32, + 224, + 379, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 224, + 379, + 295 + ], + "spans": [ + { + "bbox": [ + 32, + 224, + 379, + 295 + ], + "type": "text", + "content": "As an intriguing phenomenon of our visual system, pareidolia presents many opportunities in the study of the visual perception of both humans and machines. It offers a controlled setting in which to study object detection: we can present random signals to the visual system and study what detections arise. Do computer vision detectors exhibit similar misidentifications, and if not, why not? Why don't humans see pareidolic effects everywhere, in any textured region?" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 32, + 296, + 379, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 296, + 379, + 402 + ], + "spans": [ + { + "bbox": [ + 32, + 296, + 379, + 402 + ], + "type": "text", + "content": "To help answer these questions, we introduce an annotated dataset of five thousand pareidolic face images, called \"Faces in Things\". With this dataset, we examine whether modern computer vision face detection systems, trained to robustly detect human faces, exhibit pareidolia. We show that a state-of-the-art neural network trained on the popular WIDER FACE detection benchmark [60] fails to detect pareidolic faces well, even when detection thresholds are relaxed. By fine-tuning the same model on the Faces in Things training data we create a simple and strong baseline for the task of pareidolic face detection, which shows that significantly higher machine pareidolic performance is within reach." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 32, + 404, + 379, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 404, + 379, + 522 + ], + "spans": [ + { + "bbox": [ + 32, + 404, + 379, + 522 + ], + "type": "text", + "content": "Next, we explore how we might bridge this gap to supervised—or ultimately, human—performance, without access to pareidolic training data? Could pareidolia appear in a face detector in a more natural way? The Faces in Things dataset provides a clean testbed to explore these questions in machines. We test a variety of different interventions ranging from image augmentation techniques to additional sources of training data. We find one possible mechanism that accounts for roughly half of the performance gap: when models are fine-tuned to detect animal faces, pareidolic face detection is significantly improved. This suggests that face pareidolia may arise in part from a more general, evolutionary need to detect diverse faces in the natural environment." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 32, + 523, + 379, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 523, + 379, + 583 + ], + "spans": [ + { + "bbox": [ + 32, + 523, + 379, + 583 + ], + "type": "text", + "content": "Finally, we consider why pareidolic faces are not all around us, and why certain textures seem to cause the effect more often. We propose two simple mathematical models, a simple Gaussian process model, and a second deep feature-based model, that capture important features of pareidolia. In particular, we show how these simple models both predict a \"Goldilocks\" zone, where" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 32, + 11, + 39, + 18 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 11, + 39, + 18 + ], + "spans": [ + { + "bbox": [ + 32, + 11, + 39, + 18 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 9, + 227, + 20 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 9, + 227, + 20 + ], + "spans": [ + { + "bbox": [ + 65, + 9, + 227, + 20 + ], + "type": "text", + "content": "M. Hamilton, S. Stent, V. DuTell, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 32, + 34, + 379, + 59 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 34, + 379, + 59 + ], + "spans": [ + { + "bbox": [ + 32, + 34, + 379, + 59 + ], + "type": "text", + "content": "conditions are ideal to induce pareidolia. We confirm the existence of this zone with experiments on both human subjects and face detection models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 32, + 59, + 379, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 59, + 379, + 95 + ], + "spans": [ + { + "bbox": [ + 32, + 59, + 379, + 95 + ], + "type": "text", + "content": "Through the contributions of our open-source dataset, models, and experimental findings, we bring the study of the intriguing phenomenon of face pareidolia to the computer vision community." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 33, + 114, + 137, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 114, + 137, + 127 + ], + "spans": [ + { + "bbox": [ + 33, + 114, + 137, + 127 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 33, + 140, + 379, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 140, + 379, + 259 + ], + "spans": [ + { + "bbox": [ + 33, + 140, + 379, + 259 + ], + "type": "text", + "content": "Face detection. One of the most famous early examples of face detection was the Viola-Jones face detector [52, 53]. This detector used binary Haar features through simple-to-compute integral images and achieved greater precision and efficiency than early neural-network-based detectors [40, 41] and other feature-based methods [37, 59]. Following the deep learning breakthroughs of the 2010s, methods transitioned from hand-crafted features to learned features, and convolutional neural networks (CNNs) achieved close to human levels of performance on ever-larger datasets [8, 27-29, 32, 36, 48, 63]. For a broader survey of face detection methods, we refer the interested reader to [36, 62]. In our work, we use the recent RetinaFace model [8] as a strong face detection baseline." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 33, + 260, + 379, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 260, + 379, + 523 + ], + "spans": [ + { + "bbox": [ + 33, + 260, + 379, + 523 + ], + "type": "text", + "content": "Neuroscience of face pareidolia. The face is a highly unique stimulus for the human visual system [26, 50]: we find faces easy to spot and difficult to ignore. Face detection can occur in both noise and highly degraded images [5]. Prior work shows that face detection occurs in a dedicated brain region, the Fusiform Face Area [33]. But exactly what constitutes a face for the visual cortex and what are the mechanisms underlying pareidolia? A recent study into the temporal dynamics of neuro-imaging data during pareidolic face viewing showed results consistent with \"a broadly-tuned face detection mechanism that privileges sensitivity over selectivity\" [56]. Pareidolic faces do more than give the impression of the presence of faces: [46] show that they can trigger an additional face-specific attentional process, consuming more time and processing power than similar non-pareidolic stimuli, and even enhancing the detection of face-pareidolic objects [47]. Analyses in [30] revealed a network of neurons in the brain specialized to detect face pareidolia. Their results suggested that face processing has a strong top-down component whereby sensory input with even the slightest suggestion of a face can result in the interpretation of a face. Such top-down information might be supportive of some form of inverse rendering as a cognitive mechanism to explain the remarkable robustness of human perception of faces in degraded viewing conditions [11]. While our dataset allows the study of several types of face detection models, we focus our study on feedforward neural networks which are known to yield close to human performance on challenging \"in-the-wild\" datasets [60]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 33, + 523, + 379, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 523, + 379, + 583 + ], + "spans": [ + { + "bbox": [ + 33, + 523, + 379, + 583 + ], + "type": "text", + "content": "Face pareidolia in computer vision. Face detection and face recognition have been core topics in computer vision for many decades, but the study of face pareidolia—and its deep relationship with visual object representation learning—has been relatively overlooked. Face pareidolia has some similarities with the problem of cross-modal recognition or cross-depiction: recognizing the" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 103, + 9, + 346, + 21 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 9, + 346, + 21 + ], + "spans": [ + { + "bbox": [ + 103, + 9, + 346, + 21 + ], + "type": "text", + "content": "Seeing Faces in Things: A Model and Dataset for Pareidolia" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 372, + 10, + 379, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 10, + 379, + 19 + ], + "spans": [ + { + "bbox": [ + 372, + 10, + 379, + 19 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 33, + 33, + 89, + 89 + ], + "blocks": [ + { + "bbox": [ + 33, + 33, + 89, + 89 + ], + "lines": [ + { + "bbox": [ + 33, + 33, + 89, + 89 + ], + "spans": [ + { + "bbox": [ + 33, + 33, + 89, + 89 + ], + "type": "image", + "image_path": "2c6c9ce1d33b960aefc866ea82de667f059dbd27dbdb4c0d8e0b9156c43988f6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 90, + 33, + 148, + 89 + ], + "blocks": [ + { + "bbox": [ + 90, + 33, + 148, + 89 + ], + "lines": [ + { + "bbox": [ + 90, + 33, + 148, + 89 + ], + "spans": [ + { + "bbox": [ + 90, + 33, + 148, + 89 + ], + "type": "image", + "image_path": "c9440c481f248626894a3857bc3bf0ee7db197b841718070524d8008aab6f59b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 149, + 33, + 205, + 89 + ], + "blocks": [ + { + "bbox": [ + 149, + 33, + 205, + 89 + ], + "lines": [ + { + "bbox": [ + 149, + 33, + 205, + 89 + ], + "spans": [ + { + "bbox": [ + 149, + 33, + 205, + 89 + ], + "type": "image", + "image_path": "ec394d7d13921bcbb2ed32ccb1ec3c17291cde68d790a96fa8c15d4fb55deccf.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 207, + 33, + 263, + 89 + ], + "blocks": [ + { + "bbox": [ + 207, + 33, + 263, + 89 + ], + "lines": [ + { + "bbox": [ + 207, + 33, + 263, + 89 + ], + "spans": [ + { + "bbox": [ + 207, + 33, + 263, + 89 + ], + "type": "image", + "image_path": "cd5d70c018eb7e659accd8d19f5b9ecac68f750deabeece273d5cca2d104779e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 265, + 33, + 320, + 89 + ], + "blocks": [ + { + "bbox": [ + 265, + 33, + 320, + 89 + ], + "lines": [ + { + "bbox": [ + 265, + 33, + 320, + 89 + ], + "spans": [ + { + "bbox": [ + 265, + 33, + 320, + 89 + ], + "type": "image", + "image_path": "c44a94591afc74fcf11bfec2abbc9543e14fd928dc8f0ecc397f6ba4951b73ee.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 322, + 33, + 379, + 89 + ], + "blocks": [ + { + "bbox": [ + 322, + 33, + 379, + 89 + ], + "lines": [ + { + "bbox": [ + 322, + 33, + 379, + 89 + ], + "spans": [ + { + "bbox": [ + 322, + 33, + 379, + 89 + ], + "type": "image", + "image_path": "b466120961e6f4f3bf79ad33314b316fb42121b985be010676e5233f7c3b1bb4.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 33, + 90, + 89, + 146 + ], + "blocks": [ + { + "bbox": [ + 33, + 90, + 89, + 146 + ], + "lines": [ + { + "bbox": [ + 33, + 90, + 89, + 146 + ], + "spans": [ + { + "bbox": [ + 33, + 90, + 89, + 146 + ], + "type": "image", + "image_path": "7a91bc71b6a0fc47fc2abd7dcda24d2e70a8bfa4938b916428c04c41d8cc0517.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 30, + 155, + 380, + 188 + ], + "lines": [ + { + "bbox": [ + 30, + 155, + 380, + 188 + ], + "spans": [ + { + "bbox": [ + 30, + 155, + 380, + 188 + ], + "type": "text", + "content": "Fig. 2: Examples of face pareidolia from our \"Faces in Things\" dataset. Faces in Things consists of five thousand images annotated with bounding boxes (shown here), and facial attributes such as perceived emotion, gender, and intentionality." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 91, + 90, + 147, + 146 + ], + "blocks": [ + { + "bbox": [ + 91, + 90, + 147, + 146 + ], + "lines": [ + { + "bbox": [ + 91, + 90, + 147, + 146 + ], + "spans": [ + { + "bbox": [ + 91, + 90, + 147, + 146 + ], + "type": "image", + "image_path": "fed12bb90d7d5dc7cceb66d1c79e1e7b72a84cd2e1673a83f345cd74d3f8452a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 149, + 90, + 205, + 146 + ], + "blocks": [ + { + "bbox": [ + 149, + 90, + 205, + 146 + ], + "lines": [ + { + "bbox": [ + 149, + 90, + 205, + 146 + ], + "spans": [ + { + "bbox": [ + 149, + 90, + 205, + 146 + ], + "type": "image", + "image_path": "24d54f5511d082e81699aa5cac7fb17905cf784474aa737d64672ef0f43ddbc0.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 206, + 90, + 263, + 146 + ], + "blocks": [ + { + "bbox": [ + 206, + 90, + 263, + 146 + ], + "lines": [ + { + "bbox": [ + 206, + 90, + 263, + 146 + ], + "spans": [ + { + "bbox": [ + 206, + 90, + 263, + 146 + ], + "type": "image", + "image_path": "fc5bbf07dc400fbf7da6e37e7eaf3cf1f7e598befabf815b81fe9a7bec6efe6f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 265, + 90, + 320, + 146 + ], + "blocks": [ + { + "bbox": [ + 265, + 90, + 320, + 146 + ], + "lines": [ + { + "bbox": [ + 265, + 90, + 320, + 146 + ], + "spans": [ + { + "bbox": [ + 265, + 90, + 320, + 146 + ], + "type": "image", + "image_path": "aead20ea089a04b0a828202de810ef18ce8578e7bffddbbb44df0efa6b9b9a14.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 322, + 90, + 379, + 146 + ], + "blocks": [ + { + "bbox": [ + 322, + 90, + 379, + 146 + ], + "lines": [ + { + "bbox": [ + 322, + 90, + 379, + 146 + ], + "spans": [ + { + "bbox": [ + 322, + 90, + 379, + 146 + ], + "type": "image", + "image_path": "bcccada3f18eba5d5d9bca8a5b7272e56e23c81dd6af447257c2d4b9074cf87a.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 30, + 212, + 380, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 212, + 380, + 500 + ], + "spans": [ + { + "bbox": [ + 30, + 212, + 380, + 500 + ], + "type": "text", + "content": "same objects across different modalities irrespective of how the object is visually depicted. This has been explored particularly in the context of detecting faces, people and objects across modalities such as photography, different art movements, cartoons and sketches [4, 15, 35, 58]. The importance of capturing spatial relationships for robust cross-modal detection has also been highlighted [4]. The work of Castrejon et al. [6] showed how, when learning cross-modal scene representations with neural networks, units would emerge in the shared representation that tended to activate on consistent concepts, independently of the modality. This tendency was used by Abbas & Chalup [1], who found that mid-level units learned during human face detection could generalize to detect semantically similar facial key-points in pareidolic images, showing promise for pareidolia to emerge. However, the authors only evaluated the method qualitatively over a small test set of ten images. One route to a larger dataset may be through pareidolic image generation, which shows promise but does not yet produce convincingly natural images [12]. Curating a larger dataset, \"Totally-Looks-Like\" [39] explored the perceptual judgment of image similarity between humans and CNNs, using images which had been paired by humans as visually similar but semantically disparate. They found that visual representations extracted from CNNs such as ResNet [18] perform poorly in terms of reproducing the matching selected by humans. Though this dataset is of similar size to ours (6k samples), it is not specifically tailed towards face pareidolia and offers no bounding box or key-point annotations. Other datasets such as COCO-Periph [17] have been used to show that object detection behavior in CNNs and transformers diverge from human perception in peripheral vision." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 30, + 500, + 380, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 500, + 380, + 584 + ], + "spans": [ + { + "bbox": [ + 30, + 500, + 380, + 584 + ], + "type": "text", + "content": "In summary, there has yet to be a computational model of how or why pareidolia might arise proposed in the literature. Moreover, despite the abundance of face detection datasets [36], there is no large-scale dataset to directly support the study of face pareidolia. A large-scale pareidolia dataset would help the community explore the mechanisms underlying pareidolia, which may in turn help us to understand and harness human visual attention (which is drawn towards face-like objects), reduce pareidolic false positives in face detectors, help design-" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 11, + 39, + 18 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 11, + 39, + 18 + ], + "spans": [ + { + "bbox": [ + 33, + 11, + 39, + 18 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 9, + 227, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 9, + 227, + 19 + ], + "spans": [ + { + "bbox": [ + 65, + 9, + 227, + 19 + ], + "type": "text", + "content": "M. Hamilton, S. Stent, V. DuTell, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 97, + 31, + 157, + 108 + ], + "blocks": [ + { + "bbox": [ + 97, + 31, + 157, + 108 + ], + "lines": [ + { + "bbox": [ + 97, + 31, + 157, + 108 + ], + "spans": [ + { + "bbox": [ + 97, + 31, + 157, + 108 + ], + "type": "image", + "image_path": "381d138c54bb93c8a47c365609893c23b74f1943a15df88948761c921386993f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 180, + 31, + 240, + 109 + ], + "blocks": [ + { + "bbox": [ + 180, + 31, + 240, + 109 + ], + "lines": [ + { + "bbox": [ + 180, + 31, + 240, + 109 + ], + "spans": [ + { + "bbox": [ + 180, + 31, + 240, + 109 + ], + "type": "image", + "image_path": "e92b075219e7351279b284eefc97d74a999f84de426d3a414611312bbe9aedb2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 258, + 31, + 345, + 107 + ], + "blocks": [ + { + "bbox": [ + 258, + 31, + 345, + 107 + ], + "lines": [ + { + "bbox": [ + 258, + 31, + 345, + 107 + ], + "spans": [ + { + "bbox": [ + 258, + 31, + 345, + 107 + ], + "type": "image", + "image_path": "f2db8752b921deb79b0ebbf52724d0a97394ad614af6d8f6e2d4209ff46db6ef.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 66, + 116, + 253, + 193 + ], + "blocks": [ + { + "bbox": [ + 66, + 116, + 253, + 193 + ], + "lines": [ + { + "bbox": [ + 66, + 116, + 253, + 193 + ], + "spans": [ + { + "bbox": [ + 66, + 116, + 253, + 193 + ], + "type": "image", + "image_path": "667e0c78a2a5bc39739edcca9f8919f314703945c93c5eb2d9f84d54ec00e464.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 30, + 204, + 381, + 260 + ], + "lines": [ + { + "bbox": [ + 30, + 204, + 381, + 260 + ], + "spans": [ + { + "bbox": [ + 30, + 204, + 381, + 260 + ], + "type": "text", + "content": "Fig. 3: Attributes of the Faces in Things Dataset. We find that " + }, + { + "bbox": [ + 30, + 204, + 381, + 260 + ], + "type": "inline_equation", + "content": "31\\%" + }, + { + "bbox": [ + 30, + 204, + 381, + 260 + ], + "type": "text", + "content": " of faces are considered challenging to spot; faces are largely " + }, + { + "bbox": [ + 30, + 204, + 381, + 260 + ], + "type": "inline_equation", + "content": "(31\\%)" + }, + { + "bbox": [ + 30, + 204, + 381, + 260 + ], + "type": "text", + "content": " judged as happy; approximately half " + }, + { + "bbox": [ + 30, + 204, + 381, + 260 + ], + "type": "inline_equation", + "content": "(47\\%)" + }, + { + "bbox": [ + 30, + 204, + 381, + 260 + ], + "type": "text", + "content": " are judged as accidental rather than by design; animals and humans are seen in roughly equal numbers; and we observe a slight bias " + }, + { + "bbox": [ + 30, + 204, + 381, + 260 + ], + "type": "inline_equation", + "content": "(16\\%" + }, + { + "bbox": [ + 30, + 204, + 381, + 260 + ], + "type": "text", + "content": " vs " + }, + { + "bbox": [ + 30, + 204, + 381, + 260 + ], + "type": "inline_equation", + "content": "3\\%)" + }, + { + "bbox": [ + 30, + 204, + 381, + 260 + ], + "type": "text", + "content": " towards male over female faces, similar to biases observed in prior studies [54,55]." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 262, + 116, + 323, + 193 + ], + "blocks": [ + { + "bbox": [ + 262, + 116, + 323, + 193 + ], + "lines": [ + { + "bbox": [ + 262, + 116, + 323, + 193 + ], + "spans": [ + { + "bbox": [ + 262, + 116, + 323, + 193 + ], + "type": "image", + "image_path": "40f4656c5a7e936915bb292a257874ae8eff05271cf043ea0e84bb926ab8aaec.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 30, + 281, + 380, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 281, + 380, + 305 + ], + "spans": [ + { + "bbox": [ + 30, + 281, + 380, + 305 + ], + "type": "text", + "content": "ers avoid or create pareidolia, improve pareidolic animation, and create systems that better understand how humans perceive the world." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 31, + 320, + 196, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 320, + 196, + 335 + ], + "spans": [ + { + "bbox": [ + 31, + 320, + 196, + 335 + ], + "type": "text", + "content": "3 Faces in Things Dataset" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 30, + 342, + 381, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 342, + 381, + 510 + ], + "spans": [ + { + "bbox": [ + 30, + 342, + 381, + 510 + ], + "type": "text", + "content": "To address this gap, we begin by sampling candidate pareidolic images from the LAION-5B dataset [42]. This dataset consists of 5.85 billion CLIP-filtered image-text pairs, of which " + }, + { + "bbox": [ + 30, + 342, + 381, + 510 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 30, + 342, + 381, + 510 + ], + "type": "text", + "content": " of captions contain English. We use CLIP retrieval [2] to build a raw image set based on text queries including \"pareidolia\", \"faces in things\", \"accidental faces\", and \"[object] looks like a face\". We download images, check for duplicates, then downsample to " + }, + { + "bbox": [ + 30, + 342, + 381, + 510 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 30, + 342, + 381, + 510 + ], + "type": "text", + "content": " pixels while preserving the aspect ratio with white-space padding. We used the VGG Image Annotation tool [10] to manually annotate images, removing samples that contain the faces of actual humans or animals. Some examples of annotated images are shown in Fig. 2. Our annotations include the bounding boxes of pareidolic faces and basic facial attributes as summarized in Fig. 3. Though beyond the scope of the current paper, we note that these attributes could be useful for other future studies. We divide the dataset at random into training (" + }, + { + "bbox": [ + 30, + 342, + 381, + 510 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 30, + 342, + 381, + 510 + ], + "type": "text", + "content": ") and testing (" + }, + { + "bbox": [ + 30, + 342, + 381, + 510 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 30, + 342, + 381, + 510 + ], + "type": "text", + "content": ") sets. We refer to this as the 'Pareidolic' dataset." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 31, + 525, + 130, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 525, + 130, + 540 + ], + "spans": [ + { + "bbox": [ + 31, + 525, + 130, + 540 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 31, + 547, + 381, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 547, + 381, + 584 + ], + "spans": [ + { + "bbox": [ + 31, + 547, + 381, + 584 + ], + "type": "text", + "content": "Datasets. We use the following additional datasets. Fig. 4 shows the average faces within our dataset (Pareidolic) and the WIDER FACE (Human), and AnimalWeb (Animal) datasets." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 102, + 9, + 347, + 21 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 9, + 347, + 21 + ], + "spans": [ + { + "bbox": [ + 102, + 9, + 347, + 21 + ], + "type": "text", + "content": "Seeing Faces in Things: A Model and Dataset for Pareidolia" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 372, + 10, + 379, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 10, + 379, + 19 + ], + "spans": [ + { + "bbox": [ + 372, + 10, + 379, + 19 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 101, + 32, + 170, + 123 + ], + "blocks": [ + { + "bbox": [ + 101, + 32, + 170, + 123 + ], + "lines": [ + { + "bbox": [ + 101, + 32, + 170, + 123 + ], + "spans": [ + { + "bbox": [ + 101, + 32, + 170, + 123 + ], + "type": "image", + "image_path": "b355225945630caf5e4ad7588fe38a29252a3fcf514c6357df5ccd14d6e4b8fc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 111, + 124, + 159, + 135 + ], + "lines": [ + { + "bbox": [ + 111, + 124, + 159, + 135 + ], + "spans": [ + { + "bbox": [ + 111, + 124, + 159, + 135 + ], + "type": "text", + "content": "Pareidolic" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 172, + 33, + 240, + 124 + ], + "blocks": [ + { + "bbox": [ + 172, + 33, + 240, + 124 + ], + "lines": [ + { + "bbox": [ + 172, + 33, + 240, + 124 + ], + "spans": [ + { + "bbox": [ + 172, + 33, + 240, + 124 + ], + "type": "image", + "image_path": "10f44b727bc30bb1bf68f955c155619e737a361d6ecd5925a342806226f81cf7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 188, + 125, + 223, + 135 + ], + "lines": [ + { + "bbox": [ + 188, + 125, + 223, + 135 + ], + "spans": [ + { + "bbox": [ + 188, + 125, + 223, + 135 + ], + "type": "text", + "content": "Human" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 30, + 145, + 380, + 200 + ], + "lines": [ + { + "bbox": [ + 30, + 145, + 380, + 200 + ], + "spans": [ + { + "bbox": [ + 30, + 145, + 380, + 200 + ], + "type": "text", + "content": "Fig. 4: The Appearance of an Average Pareidolic Face. Per-channel histogram-qualified average images for registered pareidolic faces (our Faces in Things dataset), human faces (samples from the WIDER FACE dataset [60]), and animal faces (AnimalWeb [24]). The average pareidolic face, while less distinct than human or animal, has surprisingly clear eye, nose, and mouth features, and vertical symmetry." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 242, + 33, + 310, + 124 + ], + "blocks": [ + { + "bbox": [ + 242, + 33, + 310, + 124 + ], + "lines": [ + { + "bbox": [ + 242, + 33, + 310, + 124 + ], + "spans": [ + { + "bbox": [ + 242, + 33, + 310, + 124 + ], + "type": "image", + "image_path": "7dd3ed7904da2e4d6f9ab13b2c34ef64a2a0af5870e9bfc29a295a399db27a26.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 258, + 125, + 294, + 135 + ], + "lines": [ + { + "bbox": [ + 258, + 125, + 294, + 135 + ], + "spans": [ + { + "bbox": [ + 258, + 125, + 294, + 135 + ], + "type": "text", + "content": "Animal" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 30, + 224, + 379, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 224, + 379, + 283 + ], + "spans": [ + { + "bbox": [ + 30, + 224, + 379, + 283 + ], + "type": "text", + "content": "WIDER FACE [60] is a popular face detection benchmark dataset with 32,203 images and 393,703 faces. It contains a high degree of variability in scale, pose, makeup, lighting, emotion, and occlusion, organized across 61 event classes. We use the provided " + }, + { + "bbox": [ + 30, + 224, + 379, + 283 + ], + "type": "inline_equation", + "content": "40\\% / 10\\% / 50\\%" + }, + { + "bbox": [ + 30, + 224, + 379, + 283 + ], + "type": "text", + "content": " splits for training, validation, and testing. We refer to this as the 'Human' dataset." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 30, + 284, + 379, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 284, + 379, + 354 + ], + "spans": [ + { + "bbox": [ + 30, + 284, + 379, + 354 + ], + "type": "text", + "content": "AnimalWeb [24] is a collection of 22,451 faces from 334 diverse species and 21 animal orders across biological taxonomy. These faces are captured 'in-the-wild' and are consistently annotated with 9 landmarks on key facial features. We convert these landmarks to bounding boxes, by finding the tightest box that captures the points and expanding this box's width and height by " + }, + { + "bbox": [ + 30, + 284, + 379, + 354 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 30, + 284, + 379, + 354 + ], + "type": "text", + "content": ". We refer to this as the 'Animal' dataset." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 30, + 356, + 379, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 356, + 379, + 416 + ], + "spans": [ + { + "bbox": [ + 30, + 356, + 379, + 416 + ], + "type": "text", + "content": "WIDER FACE Corruptions. To measure whether pareidolia could arise from common data augmentations we corrupt the WIDER FACE images using the level 3 strength of the corruptions used in both the COCO-C [34] and ImageNet-C [19] datasets. We also include a Sobel filtering corruption [13] which has been shown to reduce a model's dependence on texture information [22]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 30, + 428, + 380, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 428, + 380, + 584 + ], + "spans": [ + { + "bbox": [ + 30, + 428, + 380, + 584 + ], + "type": "text", + "content": "Models and Training. We use RetinaFace [8] which achieves state-of-the-art performance on WIDER FACE easy and medium subsets and is the third-best face detector on the hard subset, missing the top model by less than a percentage point of Average Precision (AP). We perform experiments using both their MobileNet [21] and ResNet50 [18] backbones and use the Pytorch_Retinaface [3] repository to ensure the same experimental conditions, dataset characteristics, and preprocessing. We use pre-trained models provided by this repository and fine-tune them for 10 epochs with the AdamW optimizer [31] using a learning rate of " + }, + { + "bbox": [ + 30, + 428, + 380, + 584 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 30, + 428, + 380, + 584 + ], + "type": "text", + "content": " and a weight decay of " + }, + { + "bbox": [ + 30, + 428, + 380, + 584 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-4}" + }, + { + "bbox": [ + 30, + 428, + 380, + 584 + ], + "type": "text", + "content": ". We verify that fine-tuning using this strategy on the original WIDER FACE training dataset does not hurt model performance. When fine-tuning on Faces in Things (Pareidolia), Animal-Web (Animal), WIDER FACE (Human), and Corruption datasets we randomly replace images in the original WIDER FACE stream of training data with data" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 32, + 10, + 39, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 10, + 39, + 19 + ], + "spans": [ + { + "bbox": [ + 32, + 10, + 39, + 19 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 9, + 227, + 20 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 9, + 227, + 20 + ], + "spans": [ + { + "bbox": [ + 65, + 9, + 227, + 20 + ], + "type": "text", + "content": "M. Hamilton, S. Stent, V. DuTell, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 33, + 33, + 101, + 100 + ], + "blocks": [ + { + "bbox": [ + 33, + 33, + 101, + 100 + ], + "lines": [ + { + "bbox": [ + 33, + 33, + 101, + 100 + ], + "spans": [ + { + "bbox": [ + 33, + 33, + 101, + 100 + ], + "type": "image", + "image_path": "81f85d3b6f3d075405bb4d35b624458249040e69a5854a209fa4a5101ea3ab53.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 102, + 33, + 170, + 100 + ], + "blocks": [ + { + "bbox": [ + 102, + 33, + 170, + 100 + ], + "lines": [ + { + "bbox": [ + 102, + 33, + 170, + 100 + ], + "spans": [ + { + "bbox": [ + 102, + 33, + 170, + 100 + ], + "type": "image", + "image_path": "b6e3b2594a68761b6673c849d4c1975eb99a0542499d534b102b81e649772fd8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 172, + 33, + 240, + 100 + ], + "blocks": [ + { + "bbox": [ + 172, + 33, + 240, + 100 + ], + "lines": [ + { + "bbox": [ + 172, + 33, + 240, + 100 + ], + "spans": [ + { + "bbox": [ + 172, + 33, + 240, + 100 + ], + "type": "image", + "image_path": "dba4bb603318a3e173a6c694e1a84b057af610916e425e7bd92f26313777fe15.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 242, + 33, + 309, + 100 + ], + "blocks": [ + { + "bbox": [ + 242, + 33, + 309, + 100 + ], + "lines": [ + { + "bbox": [ + 242, + 33, + 309, + 100 + ], + "spans": [ + { + "bbox": [ + 242, + 33, + 309, + 100 + ], + "type": "image", + "image_path": "b991c14e2c09afd338714c7569dee4ed4023fb87f5fc9a0ae988fa7acdc42040.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 311, + 33, + 379, + 100 + ], + "blocks": [ + { + "bbox": [ + 311, + 33, + 379, + 100 + ], + "lines": [ + { + "bbox": [ + 311, + 33, + 379, + 100 + ], + "spans": [ + { + "bbox": [ + 311, + 33, + 379, + 100 + ], + "type": "image", + "image_path": "5f30a7a53706be0d06c02f993d9322ef078d14d64c01fab4ad2916d28a743973.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 33, + 100, + 101, + 168 + ], + "blocks": [ + { + "bbox": [ + 33, + 100, + 101, + 168 + ], + "lines": [ + { + "bbox": [ + 33, + 100, + 101, + 168 + ], + "spans": [ + { + "bbox": [ + 33, + 100, + 101, + 168 + ], + "type": "image", + "image_path": "f2e63abf3778b1c4d5365f2663025ca0a06d5639eee031bd9394a71b45e455b0.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 30, + 177, + 381, + 234 + ], + "lines": [ + { + "bbox": [ + 30, + 177, + 381, + 234 + ], + "spans": [ + { + "bbox": [ + 30, + 177, + 381, + 234 + ], + "type": "text", + "content": "Fig. 5: Qualitative Analysis of Transfer Experiments. On a sample of held-out test images, we visualize the confident " + }, + { + "bbox": [ + 30, + 177, + 381, + 234 + ], + "type": "inline_equation", + "content": "(p > 10\\%)" + }, + { + "bbox": [ + 30, + 177, + 381, + 234 + ], + "type": "text", + "content": " detections of our ground truth (red), our model fine-tuned on human faces (blue), and our model fine-tuned on animal faces (green). It is evident from these and Table 1 that fine-tuning on animal faces significantly boosts the model's ability to detect pareidolic faces." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 102, + 100, + 170, + 168 + ], + "blocks": [ + { + "bbox": [ + 102, + 100, + 170, + 168 + ], + "lines": [ + { + "bbox": [ + 102, + 100, + 170, + 168 + ], + "spans": [ + { + "bbox": [ + 102, + 100, + 170, + 168 + ], + "type": "image", + "image_path": "43e441d080954bf4c40ec580f72b97edf49c45ac9482bd3c390b2715b0fcfefb.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 172, + 100, + 240, + 168 + ], + "blocks": [ + { + "bbox": [ + 172, + 100, + 240, + 168 + ], + "lines": [ + { + "bbox": [ + 172, + 100, + 240, + 168 + ], + "spans": [ + { + "bbox": [ + 172, + 100, + 240, + 168 + ], + "type": "image", + "image_path": "3190ad5285dbdeef8ff7432865185177621797c123820dce659fa94d29773329.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 242, + 100, + 309, + 168 + ], + "blocks": [ + { + "bbox": [ + 242, + 100, + 309, + 168 + ], + "lines": [ + { + "bbox": [ + 242, + 100, + 309, + 168 + ], + "spans": [ + { + "bbox": [ + 242, + 100, + 309, + 168 + ], + "type": "image", + "image_path": "87aa715e559cbb15179ab6405864f508fc833193c7f0fde8e05a9a4de33a01a5.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 311, + 100, + 379, + 168 + ], + "blocks": [ + { + "bbox": [ + 311, + 100, + 379, + 168 + ], + "lines": [ + { + "bbox": [ + 311, + 100, + 379, + 168 + ], + "spans": [ + { + "bbox": [ + 311, + 100, + 379, + 168 + ], + "type": "image", + "image_path": "1daac9a1541026579175baac1a136e0c0741356d8c58c2c9e31b0847635024ad.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 30, + 258, + 379, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 258, + 379, + 330 + ], + "spans": [ + { + "bbox": [ + 30, + 258, + 379, + 330 + ], + "type": "text", + "content": "from the target dataset " + }, + { + "bbox": [ + 30, + 258, + 379, + 330 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 30, + 258, + 379, + 330 + ], + "type": "text", + "content": " of the time. This allows the network to learn the new task without catastrophic forgetting [25]. These changes to the optimizer, learning rate, number of epochs, and stream of training data are the only changes we make to the training paradigm of [8]. Figures in this work use the MobileNet architecture of RetinaNet unless specified otherwise. AP evaluation computations share the same setting and parameters as [8]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 31, + 349, + 302, + 362 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 349, + 302, + 362 + ], + "spans": [ + { + "bbox": [ + 31, + 349, + 302, + 362 + ], + "type": "text", + "content": "4.1 Does a SOTA Face Detector Exhibit Pareidolia?" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 30, + 372, + 380, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 372, + 380, + 493 + ], + "spans": [ + { + "bbox": [ + 30, + 372, + 380, + 493 + ], + "type": "text", + "content": "We measure the Average Precision (AP) of the MobileNet and ResNet50 RetinaNet architectures on the Faces in Things dataset. The first row of Table 1 shows results for existing pre-trained models, and the second row shows those for models fine-tuned on the original WIDER FACE training data. These act as control groups to ensure our transfer learning procedure does not interfere with our measurement of the effects of other interventions. Though these models exhibit pareidolia to a small extent, they fall far short of a model fine-tuned to detect pareidolic faces. Fig. 5 also depicts some of these predictions with blue boxes. On the whole, the models trained only on human faces are largely silent across the Faces in Things dataset." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 31, + 513, + 218, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 513, + 218, + 526 + ], + "spans": [ + { + "bbox": [ + 31, + 513, + 218, + 526 + ], + "type": "text", + "content": "4.2 How Might Pareidolia Emerge?" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 30, + 535, + 379, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 535, + 379, + 584 + ], + "spans": [ + { + "bbox": [ + 30, + 535, + 379, + 584 + ], + "type": "text", + "content": "The WIDER FACE dataset is known for its diversity of lighting, pose, makeup, emotion, and scale of faces. This fact, coupled with the results of Section 4.1 begs the question: What else do models need to experience pareidolia as humans do? The Faces in Things dataset provides a clean and robust setting to explore" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 102, + 9, + 347, + 21 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 9, + 347, + 21 + ], + "spans": [ + { + "bbox": [ + 102, + 9, + 347, + 21 + ], + "type": "text", + "content": "Seeing Faces in Things: A Model and Dataset for Pareidolia" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 373, + 10, + 380, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 10, + 380, + 19 + ], + "spans": [ + { + "bbox": [ + 373, + 10, + 380, + 19 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 35, + 28, + 202, + 153 + ], + "blocks": [ + { + "bbox": [ + 35, + 28, + 202, + 153 + ], + "lines": [ + { + "bbox": [ + 35, + 28, + 202, + 153 + ], + "spans": [ + { + "bbox": [ + 35, + 28, + 202, + 153 + ], + "type": "image", + "image_path": "3d5c50beae2e3d1a197a94b6d4c0bfc70531368543c1922368e170a4649826ae.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 30, + 169, + 380, + 214 + ], + "lines": [ + { + "bbox": [ + 30, + 169, + 380, + 214 + ], + "spans": [ + { + "bbox": [ + 30, + 169, + 380, + 214 + ], + "type": "text", + "content": "Fig. 6: Measuring the effect of several training interventions on pareidolic face detection The left plot shows that fine-tuning RetinaNet on animals improves pareidolic face detection more than any other intervention. Conversely, the right plot shows that pareidolic fine-tuning improves animal face detection performance." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 206, + 28, + 376, + 153 + ], + "blocks": [ + { + "bbox": [ + 206, + 28, + 376, + 153 + ], + "lines": [ + { + "bbox": [ + 206, + 28, + 376, + 153 + ], + "spans": [ + { + "bbox": [ + 206, + 28, + 376, + 153 + ], + "type": "image", + "image_path": "a364ae7066474aaa2c442a6555a8edad35f41f1e3f6b43191897d9b923c9118c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 30, + 242, + 379, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 242, + 379, + 292 + ], + "spans": [ + { + "bbox": [ + 30, + 242, + 379, + 292 + ], + "type": "text", + "content": "the development of pareidolia in algorithms. Unlike in humans, where it is impossible to causally intervene on their facial training data, we can easily modify an algorithm's training data. This makes it possible to explore whether one can induce pareidolia in algorithms using specific stimuli." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 30, + 293, + 380, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 293, + 380, + 485 + ], + "spans": [ + { + "bbox": [ + 30, + 293, + 380, + 485 + ], + "type": "text", + "content": "To this end, we investigate whether a variety of training data interventions can induce pareidolia in algorithms. In particular, we measure the effect of adding several data augmentations from the COCO-C [34] and ImageNet-C [19] datasets and explore a Sobel filtering augmentation which reduces models' dependence on texture. Additionally, we also measure the effect of adding animal faces to the training data. Animal faces show a far greater breadth of variation in coloration, structure, and appearance than human faces. Recognizing animal faces provides many evolutionary advantages including gaze detection during hunting and avoiding onlooking predators. The generality required to detect this wide space of faces could yield a greater number of \"false positives\" that lead to the sensation of pareidolia. Indeed, some recent studies provide some corroborating evidence for this hypothesis. Firstly, Rhesus Monkeys exhibit pareidolia [49] showing this effect does not only occur in humans. Secondly, the experience of pareidolia is a rapid cognitive process and not a \"late re-interpretation\" of input signals [16, 56] which the authors conclude is evidence that pareidolia could be linked to the need to quickly react to predators." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 30, + 488, + 380, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 488, + 380, + 584 + ], + "spans": [ + { + "bbox": [ + 30, + 488, + 380, + 584 + ], + "type": "text", + "content": "We plot the change to pareidolic face detection performance as a function of each training intervention in the left panel of Figure 6. Of the different corruption interventions, we find that Sobel filtering, motion blur, Gaussian noise, and fog tend to slightly improve pareidolic face detection while most other corruptions do not improve pareidolic face detection performance. Most strikingly, the addition of animal faces to the training data roughly doubles the algorithms' ability to detect pareidolic faces compared to the control group, closing around half of the gap between a human-trained model and a pareidolia trained model." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 11, + 39, + 18 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 11, + 39, + 18 + ], + "spans": [ + { + "bbox": [ + 33, + 11, + 39, + 18 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 9, + 227, + 20 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 9, + 227, + 20 + ], + "spans": [ + { + "bbox": [ + 65, + 9, + 227, + 20 + ], + "type": "text", + "content": "M. Hamilton, S. Stent, V. DuTell, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 117, + 31, + 292, + 123 + ], + "blocks": [ + { + "bbox": [ + 117, + 31, + 292, + 123 + ], + "lines": [ + { + "bbox": [ + 117, + 31, + 292, + 123 + ], + "spans": [ + { + "bbox": [ + 117, + 31, + 292, + 123 + ], + "type": "table", + "html": "
FinetuningAP
MobileNetResNet50
None7.9%2.8%
Human (Control)9.8%3.6%
Animal16.7%15.4%
Pareidolia33.9%27.1%
Animal + Pareidolia36.4%31.7%
", + "image_path": "1e4b8f123970d85c99c74bcc70d8e16b82737453ed47dba71648a20a9f8e6e7f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 32, + 126, + 379, + 182 + ], + "lines": [ + { + "bbox": [ + 32, + 126, + 379, + 182 + ], + "spans": [ + { + "bbox": [ + 32, + 126, + 379, + 182 + ], + "type": "text", + "content": "Table 1: Effect of Fine Tuning on Pareidolic Face Detection. Our results show that WIDER FACE-trained RetinaFace models do not detect many pareidolic images. Fine-tuning these models on animal faces approximately doubles pareidolic face detection rates. Interestingly, adding animal faces alongside pareidolic faces (30%/70% split respectively) can improve performance over fine-tuning on pareidolic faces alone." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 32, + 201, + 379, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 201, + 379, + 273 + ], + "spans": [ + { + "bbox": [ + 32, + 201, + 379, + 273 + ], + "type": "text", + "content": "Reciprocally, the right-hand plot of Figure 6 shows that fine-tuning on pareidolic images yields the greatest improvement in animal face detection. We further explore this phenomenon in Table 1, where we show that this effect occurs across both MobileNet and ResNet50 architectures. Finally, we also show that adding a small number of animal faces (30% animal 70% pareidolic) can improve pareidolic face detection performance over pareidolic images alone." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 32, + 274, + 379, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 274, + 379, + 404 + ], + "spans": [ + { + "bbox": [ + 32, + 274, + 379, + 404 + ], + "type": "text", + "content": "To understand this effect better, Fig. 7 visualizes the inner representations of this model across the three datasets (Human = WIDER FACE, Animal = AnimalWeb, Pareidolic = Faces in Things). Specifically, we extract multi-scale features from the animal and pareidolia fine-tuned RetinaNet shared feature layer before the application of the classification and regression heads. We average pool these features across the bounding box for each face and visualize them with t-SNE [20]. This figure shows that RetinaNet's representations of animal and pareidolic faces tend to cluster together and are distinct from its representations of human faces. This lends evidence to the relative similarity of pareidolic and animal faces compared to human faces. We also reiterate that we filtered the Faces in Things dataset to avoid images of real animals." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 33, + 420, + 171, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 420, + 171, + 434 + ], + "spans": [ + { + "bbox": [ + 33, + 420, + 171, + 434 + ], + "type": "text", + "content": "5 Modeling Pareidolia" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 33, + 443, + 379, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 443, + 379, + 515 + ], + "spans": [ + { + "bbox": [ + 33, + 443, + 379, + 515 + ], + "type": "text", + "content": "Though many prior works have measured pareidolia, there has yet to be a simple mathematical model that describes the high-level structure of this phenomenon. In this section we provide two simple formal models of pareidolia and show that they both exhibit a testable prediction: the existence of a peak in pareidolic face detection as a function of an image's complexity. Section 5.4 presents experimental evidence of this \"pareidolic peak\" in both humans and machines." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 33, + 530, + 207, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 530, + 207, + 541 + ], + "spans": [ + { + "bbox": [ + 33, + 530, + 207, + 541 + ], + "type": "text", + "content": "5.1 Gaussian Model of Pareidolia" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 33, + 547, + 379, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 547, + 379, + 583 + ], + "spans": [ + { + "bbox": [ + 33, + 547, + 379, + 583 + ], + "type": "text", + "content": "A model of pareidolia needs to describe two processes: (1) the random process that generates candidate images, and (2) the face detection process which determines when an image is pareidolic. We begin with a simple Gaussian model for" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 103, + 9, + 346, + 21 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 9, + 346, + 21 + ], + "spans": [ + { + "bbox": [ + 103, + 9, + 346, + 21 + ], + "type": "text", + "content": "Seeing Faces in Things: A Model and Dataset for Pareidolia" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 372, + 10, + 379, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 10, + 379, + 19 + ], + "spans": [ + { + "bbox": [ + 372, + 10, + 379, + 19 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 105, + 36, + 304, + 185 + ], + "blocks": [ + { + "bbox": [ + 105, + 36, + 304, + 185 + ], + "lines": [ + { + "bbox": [ + 105, + 36, + 304, + 185 + ], + "spans": [ + { + "bbox": [ + 105, + 36, + 304, + 185 + ], + "type": "image", + "image_path": "3ed09d02fe6255fd7f0a9ace6c40ba6d1a19d1d8f20ab1df02d7d8862da99be6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 30, + 191, + 381, + 248 + ], + "lines": [ + { + "bbox": [ + 30, + 191, + 381, + 248 + ], + "spans": [ + { + "bbox": [ + 30, + 191, + 381, + 248 + ], + "type": "text", + "content": "Fig. 7: Visualizing RetinaNet Representations across Datasets. Animal+Pareidolia fine-tuned RetinaNet representations tend to group animal and pareidolic faces together. This lends evidence to the hypothesis that the perception of animal and pareidolic faces are linked. (To highlight the commonality of pareidolic animal detection we note the similarity of these points to a frog.)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 30, + 276, + 379, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 276, + 379, + 337 + ], + "spans": [ + { + "bbox": [ + 30, + 276, + 379, + 337 + ], + "type": "text", + "content": "each. We model the image generation process as a sum of independent normal modes, each contributing a zero-mean Gaussian of a specified variance, multiplied by the mode image " + }, + { + "bbox": [ + 30, + 276, + 379, + 337 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 30, + 276, + 379, + 337 + ], + "type": "text", + "content": ". For example, as in [9, 51], these modes could be the principal components of a mean-subtracted image dataset. In this setting the generated image, " + }, + { + "bbox": [ + 30, + 276, + 379, + 337 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 30, + 276, + 379, + 337 + ], + "type": "text", + "content": ", is a weighted sum of the normal modes:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 120, + 351, + 379, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 351, + 379, + 376 + ], + "spans": [ + { + "bbox": [ + 120, + 351, + 379, + 376 + ], + "type": "interline_equation", + "content": "\\boldsymbol {y} = \\sum_ {i} n _ {i} \\boldsymbol {y} _ {i} \\quad \\text {w h e r e}, \\quad n _ {i} \\sim N (0, \\sigma_ {i}) \\tag {1}", + "image_path": "2bfcdb2b6423a481b13e64a35d37ddf5c20fe05e21e197f33f3572025dad568a.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 30, + 390, + 380, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 390, + 380, + 488 + ], + "spans": [ + { + "bbox": [ + 30, + 390, + 380, + 488 + ], + "type": "text", + "content": "To model our face detection process we capture the intuition of matching an image to a template image and note that this can be generalized to distributions of template images. In particular, the target pareidolic image is represented as a vector, " + }, + { + "bbox": [ + 30, + 390, + 380, + 488 + ], + "type": "inline_equation", + "content": "\\pmb{a}" + }, + { + "bbox": [ + 30, + 390, + 380, + 488 + ], + "type": "text", + "content": ", of statistically independent target coefficients, " + }, + { + "bbox": [ + 30, + 390, + 380, + 488 + ], + "type": "inline_equation", + "content": "a_{i}" + }, + { + "bbox": [ + 30, + 390, + 380, + 488 + ], + "type": "text", + "content": ", for each mode. The probability that this mode contributes towards the face detection, " + }, + { + "bbox": [ + 30, + 390, + 380, + 488 + ], + "type": "inline_equation", + "content": "P(a_{i})" + }, + { + "bbox": [ + 30, + 390, + 380, + 488 + ], + "type": "text", + "content": ", is the probability of detecting the pareidolic value, " + }, + { + "bbox": [ + 30, + 390, + 380, + 488 + ], + "type": "inline_equation", + "content": "a_{i}" + }, + { + "bbox": [ + 30, + 390, + 380, + 488 + ], + "type": "text", + "content": ", at the " + }, + { + "bbox": [ + 30, + 390, + 380, + 488 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 30, + 390, + 380, + 488 + ], + "type": "text", + "content": "th mode. We assume a Gaussian detection process: " + }, + { + "bbox": [ + 30, + 390, + 380, + 488 + ], + "type": "inline_equation", + "content": "P(a_{i}|y_{i}) \\sim N(a_{i}, \\gamma_{i})" + }, + { + "bbox": [ + 30, + 390, + 380, + 488 + ], + "type": "text", + "content": ". Because each mode's coefficient is a zero-mean Gaussian distribution, " + }, + { + "bbox": [ + 30, + 390, + 380, + 488 + ], + "type": "inline_equation", + "content": "P(y_{i}) \\sim N(0, \\sigma)" + }, + { + "bbox": [ + 30, + 390, + 380, + 488 + ], + "type": "text", + "content": ", we have:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 121, + 501, + 379, + 587 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 501, + 379, + 587 + ], + "spans": [ + { + "bbox": [ + 121, + 501, + 379, + 587 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} P \\left(a _ {i}\\right) = \\int_ {y _ {i}} P \\left(a _ {i}, y _ {i}\\right) \\mathrm {d} y _ {i} (2) \\\\ = \\int_ {y _ {i}} P \\left(a _ {i} \\mid y _ {i}\\right) P \\left(y _ {i}\\right) \\mathrm {d} y _ {i} (3) \\\\ = \\frac {1}{2 \\pi \\gamma_ {i} \\sigma_ {i}} \\int_ {y _ {i}} e ^ {- \\frac {(y _ {i} - a _ {i}) ^ {2}}{2 \\gamma_ {i} ^ {2}}} e ^ {- \\frac {y _ {i} ^ {2}}{2 \\sigma^ {2}}} \\mathrm {d} y _ {i} (4) \\\\ \\end{array}", + "image_path": "f20c820c79a5f7533e12cd1331dfcaf9219c1382c27b38bf36e9dba21c47a128.jpg" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "spans": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 9, + 227, + 20 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 9, + 227, + 20 + ], + "spans": [ + { + "bbox": [ + 65, + 9, + 227, + 20 + ], + "type": "text", + "content": "M. Hamilton, S. Stent, V. DuTell, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 68, + 27, + 344, + 223 + ], + "blocks": [ + { + "bbox": [ + 68, + 27, + 344, + 223 + ], + "lines": [ + { + "bbox": [ + 68, + 27, + 344, + 223 + ], + "spans": [ + { + "bbox": [ + 68, + 27, + 344, + 223 + ], + "type": "image", + "image_path": "899529a899e718c0b32f4bf85ae3787bee3b1aa6decae8eafedf09ba9775a7c4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 31, + 226, + 380, + 337 + ], + "lines": [ + { + "bbox": [ + 31, + 226, + 380, + 337 + ], + "spans": [ + { + "bbox": [ + 31, + 226, + 380, + 337 + ], + "type": "text", + "content": "Fig. 8: Illustration of the proposed Gaussian model for pareidolia with three example generating distributions. To make pareidolia likely, the generating distribution needs a proper distribution of spatial frequencies. A process with too few spatial frequencies (left) is likely to only generate weak face-like details (\"face-ness\": low). In contrast, with too many frequencies (right), faces can be modeled with exquisite detail (\"face-ness\": high), but the likelihood of drawing any particular desired combination become vanishingly small. The most likely pareidolic images form when the generating distribution has the right spectrum (middle), enabling reasonable faces to emerge with reasonable likelihood. In other words, this model predicts that pareidolic faces will match the low frequencies of faces but differ in the higher frequency details." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 31, + 362, + 379, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 362, + 379, + 398 + ], + "spans": [ + { + "bbox": [ + 31, + 362, + 379, + 398 + ], + "type": "text", + "content": "Note that " + }, + { + "bbox": [ + 31, + 362, + 379, + 398 + ], + "type": "inline_equation", + "content": "\\sigma_i^2" + }, + { + "bbox": [ + 31, + 362, + 379, + 398 + ], + "type": "text", + "content": " is the variance of the random process generating the pareidolia, while " + }, + { + "bbox": [ + 31, + 362, + 379, + 398 + ], + "type": "inline_equation", + "content": "\\gamma_i^2" + }, + { + "bbox": [ + 31, + 362, + 379, + 398 + ], + "type": "text", + "content": " is the variance of the likelihood term — how far a mode is allowed to vary from the target mode value before it stops looking like the target image, " + }, + { + "bbox": [ + 31, + 362, + 379, + 398 + ], + "type": "inline_equation", + "content": "\\pmb{a}" + }, + { + "bbox": [ + 31, + 362, + 379, + 398 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 31, + 398, + 379, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 398, + 379, + 434 + ], + "spans": [ + { + "bbox": [ + 31, + 398, + 379, + 434 + ], + "type": "text", + "content": "We can complete the square in Eq. 4 to write the product of Gaussians in " + }, + { + "bbox": [ + 31, + 398, + 379, + 434 + ], + "type": "inline_equation", + "content": "P(a_{i})" + }, + { + "bbox": [ + 31, + 398, + 379, + 434 + ], + "type": "text", + "content": " as a single Gaussian. Integrating that Gaussian over all possible observations " + }, + { + "bbox": [ + 31, + 398, + 379, + 434 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 31, + 398, + 379, + 434 + ], + "type": "text", + "content": " gives the probability of finding the pareidolic value " + }, + { + "bbox": [ + 31, + 398, + 379, + 434 + ], + "type": "inline_equation", + "content": "a_{i}" + }, + { + "bbox": [ + 31, + 398, + 379, + 434 + ], + "type": "text", + "content": " from mode " + }, + { + "bbox": [ + 31, + 398, + 379, + 434 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 31, + 398, + 379, + 434 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 444, + 379, + 476 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 444, + 379, + 476 + ], + "spans": [ + { + "bbox": [ + 131, + 444, + 379, + 476 + ], + "type": "interline_equation", + "content": "P (a _ {i}) = \\frac {1}{\\sqrt {2 \\pi (\\gamma_ {i} ^ {2} + \\sigma_ {i} ^ {2})}} e ^ {- \\frac {a _ {i} ^ {2}}{2 (\\sigma_ {i} ^ {2} + \\gamma_ {i} ^ {2})}} \\tag {5}", + "image_path": "8ca738fb2a831f95cd7fd43d4b7f5e229dce28b402c6f6cde821d354b8d67fd4.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 31, + 501, + 196, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 501, + 196, + 513 + ], + "spans": [ + { + "bbox": [ + 31, + 501, + 196, + 513 + ], + "type": "text", + "content": "5.2 Predicting Peak Pareidolia" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 30, + 523, + 380, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 523, + 380, + 584 + ], + "spans": [ + { + "bbox": [ + 30, + 523, + 380, + 584 + ], + "type": "text", + "content": "For a given mode's detection variance, " + }, + { + "bbox": [ + 30, + 523, + 380, + 584 + ], + "type": "inline_equation", + "content": "\\gamma_i^2" + }, + { + "bbox": [ + 30, + 523, + 380, + 584 + ], + "type": "text", + "content": ", and target mode coefficient, " + }, + { + "bbox": [ + 30, + 523, + 380, + 584 + ], + "type": "inline_equation", + "content": "a_i" + }, + { + "bbox": [ + 30, + 523, + 380, + 584 + ], + "type": "text", + "content": ", Eq. 4 allows us to find the optimal mode variance to generate pareidolia, i.e., to maximize " + }, + { + "bbox": [ + 30, + 523, + 380, + 584 + ], + "type": "inline_equation", + "content": "P(a_i)" + }, + { + "bbox": [ + 30, + 523, + 380, + 584 + ], + "type": "text", + "content": ". Unfortunately, we seldom have the flexibility to design a random process one mode at a time. But we may have the option to select between image generation processes that have different numbers of modes, " + }, + { + "bbox": [ + 30, + 523, + 380, + 584 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 30, + 523, + 380, + 584 + ], + "type": "text", + "content": ". Since each mode" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 102, + 9, + 347, + 21 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 9, + 347, + 21 + ], + "spans": [ + { + "bbox": [ + 102, + 9, + 347, + 21 + ], + "type": "text", + "content": "Seeing Faces in Things: A Model and Dataset for Pareidolia" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 368, + 10, + 378, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 10, + 378, + 19 + ], + "spans": [ + { + "bbox": [ + 368, + 10, + 378, + 19 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 35, + 31, + 198, + 140 + ], + "blocks": [ + { + "bbox": [ + 35, + 31, + 198, + 140 + ], + "lines": [ + { + "bbox": [ + 35, + 31, + 198, + 140 + ], + "spans": [ + { + "bbox": [ + 35, + 31, + 198, + 140 + ], + "type": "image", + "image_path": "3278e6b04f301679f9177bab9716292e2ccee0e6299185ecbe679e8961c01c25.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 31, + 144, + 201, + 221 + ], + "lines": [ + { + "bbox": [ + 31, + 144, + 201, + 221 + ], + "spans": [ + { + "bbox": [ + 31, + 144, + 201, + 221 + ], + "type": "text", + "content": "Fig. 9: Probability of pareidolia (Eq. 6) in the Gaussian model " + }, + { + "bbox": [ + 31, + 144, + 201, + 221 + ], + "type": "inline_equation", + "content": "(\\sigma = 10)" + }, + { + "bbox": [ + 31, + 144, + 201, + 221 + ], + "type": "text", + "content": " across images with different spatial frequency distributions (Sec. 5.4). This assumes spatial frequencies are uncorrelated and thus underestimates the probability of pareidolia, however peak pareidolia is still present." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 215, + 31, + 376, + 125 + ], + "blocks": [ + { + "bbox": [ + 215, + 31, + 376, + 125 + ], + "lines": [ + { + "bbox": [ + 215, + 31, + 376, + 125 + ], + "spans": [ + { + "bbox": [ + 215, + 31, + 376, + 125 + ], + "type": "image", + "image_path": "56d8b68dea877bd85cd69a98be77bd3e6d8a994945f7650a41100c8e0bcec897.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 210, + 137, + 379, + 215 + ], + "lines": [ + { + "bbox": [ + 210, + 137, + 379, + 215 + ], + "spans": [ + { + "bbox": [ + 210, + 137, + 379, + 215 + ], + "type": "text", + "content": "Fig. 10: Probability of pareidolia under the feature-based example of Eq. (8) as a function of the rate of feature detection, " + }, + { + "bbox": [ + 210, + 137, + 379, + 215 + ], + "type": "inline_equation", + "content": "\\lambda_{i} = \\lambda" + }, + { + "bbox": [ + 210, + 137, + 379, + 215 + ], + "type": "text", + "content": " for all " + }, + { + "bbox": [ + 210, + 137, + 379, + 215 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 210, + 137, + 379, + 215 + ], + "type": "text", + "content": ", within the random images. Note the low probability of pareidolia for both feature-free " + }, + { + "bbox": [ + 210, + 137, + 379, + 215 + ], + "type": "inline_equation", + "content": "(\\lambda \\to 0)" + }, + { + "bbox": [ + 210, + 137, + 379, + 215 + ], + "type": "text", + "content": " and feature-rich " + }, + { + "bbox": [ + 210, + 137, + 379, + 215 + ], + "type": "inline_equation", + "content": "(\\lambda \\gg 0)" + }, + { + "bbox": [ + 210, + 137, + 379, + 215 + ], + "type": "text", + "content": " random images." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 30, + 247, + 379, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 247, + 379, + 271 + ], + "spans": [ + { + "bbox": [ + 30, + 247, + 379, + 271 + ], + "type": "text", + "content": "is independent, the probability of a pareidolic detection of the target object template is the product of detecting the target coefficient for each of the M modes:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 167, + 282, + 379, + 314 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 282, + 379, + 314 + ], + "spans": [ + { + "bbox": [ + 167, + 282, + 379, + 314 + ], + "type": "interline_equation", + "content": "P (a) = \\prod_ {i} ^ {M} P \\left(a _ {i}\\right) \\tag {6}", + "image_path": "8866f8d9766115dd4652550a8736866213ebfbaba623e8f5d39dfcc1a8ac4709.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 30, + 321, + 379, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 321, + 379, + 479 + ], + "spans": [ + { + "bbox": [ + 30, + 321, + 379, + 479 + ], + "type": "text", + "content": "We plot some predictions of our Gaussian model in Fig. 9 for a target template with a " + }, + { + "bbox": [ + 30, + 321, + 379, + 479 + ], + "type": "inline_equation", + "content": "\\frac{1}{f}" + }, + { + "bbox": [ + 30, + 321, + 379, + 479 + ], + "type": "text", + "content": " power spectrum (standard deviation of each mode inversely proportional to mode number) on noise images of varying complexity. We note the existence of a peak in pareidolic detection probability for random image generation processes with a mid-range number of spatial modes, as measured by the width of Gaussian that modulates power in Fourier space. Too few modes in the random generation process, and no image will ever have enough complexity to render the target well. Too many modes and pareidolia becomes unlikely because so many modes need to match the desired target values. Each added mode multiplies the pareidolia probability by another small factor. In between, there is what we call peak pareidolia. As the detection value, " + }, + { + "bbox": [ + 30, + 321, + 379, + 479 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 30, + 321, + 379, + 479 + ], + "type": "text", + "content": ", becomes more stringent (smaller) the peak pareidolia value occurs at a larger number of modes and becomes less probable. We illustrate this effect in Figure 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 31, + 500, + 268, + 512 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 500, + 268, + 512 + ], + "spans": [ + { + "bbox": [ + 31, + 500, + 268, + 512 + ], + "type": "text", + "content": "5.3 Higher-Level Feature Model of Pareidolia" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 30, + 523, + 379, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 523, + 379, + 583 + ], + "spans": [ + { + "bbox": [ + 30, + 523, + 379, + 583 + ], + "type": "text", + "content": "The Gaussian model for pareidolia above lays out important aspects of pareidolia, but relies on a naive model of object detection, the squared distance from a template image. We assume that a more realistic model of human perception would incorporate higher-level features and introduce a still simple, yet more realistic, feature-based model." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "spans": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 9, + 227, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 9, + 227, + 19 + ], + "spans": [ + { + "bbox": [ + 65, + 9, + 227, + 19 + ], + "type": "text", + "content": "M. Hamilton, S. Stent, V. DuTell, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 30, + 35, + 381, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 35, + 381, + 143 + ], + "spans": [ + { + "bbox": [ + 30, + 35, + 381, + 143 + ], + "type": "text", + "content": "We assume that the detection of an object requires particular features to be detected in certain spatial regions, e.g. an eye in the top left and right, a nose in the center, and a mouth in the bottom. Such an approach has been used in computer vision object detection algorithms, e.g. [14, 57, 61]. Any given object template has some number of regions, " + }, + { + "bbox": [ + 30, + 35, + 381, + 143 + ], + "type": "inline_equation", + "content": "R_{i}" + }, + { + "bbox": [ + 30, + 35, + 381, + 143 + ], + "type": "text", + "content": ", indexed by " + }, + { + "bbox": [ + 30, + 35, + 381, + 143 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 30, + 35, + 381, + 143 + ], + "type": "text", + "content": ", within which a given feature, " + }, + { + "bbox": [ + 30, + 35, + 381, + 143 + ], + "type": "inline_equation", + "content": "F_{i}" + }, + { + "bbox": [ + 30, + 35, + 381, + 143 + ], + "type": "text", + "content": ", must be detected. The other features, " + }, + { + "bbox": [ + 30, + 35, + 381, + 143 + ], + "type": "inline_equation", + "content": "F_{j\\neq i}" + }, + { + "bbox": [ + 30, + 35, + 381, + 143 + ], + "type": "text", + "content": ", should not be detected in region " + }, + { + "bbox": [ + 30, + 35, + 381, + 143 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 30, + 35, + 381, + 143 + ], + "type": "text", + "content": ". For a given random image where we hope to detect pareidolia, we assume that feature existence is a spatial Poisson process. In this process, the probability of " + }, + { + "bbox": [ + 30, + 35, + 381, + 143 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 30, + 35, + 381, + 143 + ], + "type": "text", + "content": " feature instances for any given feature " + }, + { + "bbox": [ + 30, + 35, + 381, + 143 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 30, + 35, + 381, + 143 + ], + "type": "text", + "content": " over some area " + }, + { + "bbox": [ + 30, + 35, + 381, + 143 + ], + "type": "inline_equation", + "content": "B_{i}" + }, + { + "bbox": [ + 30, + 35, + 381, + 143 + ], + "type": "text", + "content": " is" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 148, + 151, + 379, + 177 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 151, + 379, + 177 + ], + "spans": [ + { + "bbox": [ + 148, + 151, + 379, + 177 + ], + "type": "interline_equation", + "content": "P (n _ {i}) = \\frac {\\left(\\lambda_ {i} | B _ {i} |\\right) ^ {n _ {i}}}{n _ {i} !} e ^ {- \\lambda_ {i} | B _ {i} |} \\tag {7}", + "image_path": "d7349249e605415e4cc7a12e824aa81eb600e640ba43aef5fbc85120454f7a01.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 30, + 182, + 381, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 182, + 381, + 254 + ], + "spans": [ + { + "bbox": [ + 30, + 182, + 381, + 254 + ], + "type": "text", + "content": "To detect a pareidolic instance of the object template, we must detect one feature of the correct type, " + }, + { + "bbox": [ + 30, + 182, + 381, + 254 + ], + "type": "inline_equation", + "content": "F_{i}" + }, + { + "bbox": [ + 30, + 182, + 381, + 254 + ], + "type": "text", + "content": ", in each region " + }, + { + "bbox": [ + 30, + 182, + 381, + 254 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 30, + 182, + 381, + 254 + ], + "type": "text", + "content": " of the face template, and zero features of the wrong type, " + }, + { + "bbox": [ + 30, + 182, + 381, + 254 + ], + "type": "inline_equation", + "content": "F_{j\\neq i}" + }, + { + "bbox": [ + 30, + 182, + 381, + 254 + ], + "type": "text", + "content": " in each region " + }, + { + "bbox": [ + 30, + 182, + 381, + 254 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 30, + 182, + 381, + 254 + ], + "type": "text", + "content": ". Assuming independence of the feature detections, and for simplicity setting all the feature detection rates to be the same, " + }, + { + "bbox": [ + 30, + 182, + 381, + 254 + ], + "type": "inline_equation", + "content": "\\lambda_{i} = \\lambda" + }, + { + "bbox": [ + 30, + 182, + 381, + 254 + ], + "type": "text", + "content": ", and all the template areas to be the same, " + }, + { + "bbox": [ + 30, + 182, + 381, + 254 + ], + "type": "inline_equation", + "content": "B_{i} = 1" + }, + { + "bbox": [ + 30, + 182, + 381, + 254 + ], + "type": "text", + "content": ", we have for the probability, " + }, + { + "bbox": [ + 30, + 182, + 381, + 254 + ], + "type": "inline_equation", + "content": "P(O)" + }, + { + "bbox": [ + 30, + 182, + 381, + 254 + ], + "type": "text", + "content": ", of pareidolic detection of object " + }, + { + "bbox": [ + 30, + 182, + 381, + 254 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 30, + 182, + 381, + 254 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 154, + 264, + 379, + 295 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 264, + 379, + 295 + ], + "spans": [ + { + "bbox": [ + 154, + 264, + 379, + 295 + ], + "type": "interline_equation", + "content": "P (O) = \\prod_ {i} ^ {M} \\lambda^ {1} e ^ {- \\lambda (M - 1)} \\tag {8}", + "image_path": "3e56321bf79a0afcc85431a4815a2a558d3b346892189919d84cfcb12e42f79d.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 30, + 305, + 381, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 305, + 381, + 413 + ], + "spans": [ + { + "bbox": [ + 30, + 305, + 381, + 413 + ], + "type": "text", + "content": "For the case of " + }, + { + "bbox": [ + 30, + 305, + 381, + 413 + ], + "type": "inline_equation", + "content": "M = 4" + }, + { + "bbox": [ + 30, + 305, + 381, + 413 + ], + "type": "text", + "content": ", a simple detection model for two eyes, a nose, and a mouth, we have " + }, + { + "bbox": [ + 30, + 305, + 381, + 413 + ], + "type": "inline_equation", + "content": "P(o) = \\lambda^4 e^{-16\\lambda}" + }, + { + "bbox": [ + 30, + 305, + 381, + 413 + ], + "type": "text", + "content": ", which is plotted in Fig. 10. In this feature-based object detection model, we also find the existence of \"peak pareidolia\". Again, it is governed by a parameter describing the complexity of the random image, in this case a Poisson process rate parameter, " + }, + { + "bbox": [ + 30, + 305, + 381, + 413 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 30, + 305, + 381, + 413 + ], + "type": "text", + "content": ", that governs the probability of a feature detection per unit area. For too low a rate, the model doesn't generate enough features to satisfy the object template, for too high a rate, the probability of seeing only the right features in just the right places becomes very small. In between is the most probable rate for pareidolia." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 31, + 431, + 346, + 444 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 431, + 346, + 444 + ], + "spans": [ + { + "bbox": [ + 31, + 431, + 346, + 444 + ], + "type": "text", + "content": "5.4 Measuring the Pareidolic Peak in Humans and Machines" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 30, + 452, + 381, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 452, + 381, + 584 + ], + "spans": [ + { + "bbox": [ + 30, + 452, + 381, + 584 + ], + "type": "text", + "content": "Both mathematical models of Section 5 predict the existence of a peak of pareidolic face detection as a function of image complexity. We show the existence of this pareidolic peak in both humans and machines. In particular, we perform a psychophysics experiment where human subjects view noise images of varying complexity and report how many pareidolic faces they saw in each image, from zero to nine. Campbell [5] demonstrated that a 12x12 array of random, binary squares is sufficient to evoke human and animal faces. We generate noise images of varying complexity by randomly sampling Fourier coefficients and modulating these coefficients with a zero mean " + }, + { + "bbox": [ + 30, + 452, + 381, + 584 + ], + "type": "inline_equation", + "content": "\\sigma^2" + }, + { + "bbox": [ + 30, + 452, + 381, + 584 + ], + "type": "text", + "content": " variance Gaussian in Fourier space. We show some samples of these images on the x-axis of Fig. 11. Intuitively, the Gaussian envelope in frequency space filters out most frequencies higher than " + }, + { + "bbox": [ + 30, + 452, + 381, + 584 + ], + "type": "inline_equation", + "content": "\\sigma" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 102, + 9, + 346, + 21 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 9, + 346, + 21 + ], + "spans": [ + { + "bbox": [ + 102, + 9, + 346, + 21 + ], + "type": "text", + "content": "Seeing Faces in Things: A Model and Dataset for Pareidolia" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 368, + 10, + 379, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 10, + 379, + 19 + ], + "spans": [ + { + "bbox": [ + 368, + 10, + 379, + 19 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 33, + 25, + 202, + 141 + ], + "blocks": [ + { + "bbox": [ + 33, + 25, + 202, + 141 + ], + "lines": [ + { + "bbox": [ + 33, + 25, + 202, + 141 + ], + "spans": [ + { + "bbox": [ + 33, + 25, + 202, + 141 + ], + "type": "image", + "image_path": "5c27903441e7a250a8d902412ab58a269b2f1c382c1d60fca216919c893ba5e7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 30, + 146, + 380, + 201 + ], + "lines": [ + { + "bbox": [ + 30, + 146, + 380, + 201 + ], + "spans": [ + { + "bbox": [ + 30, + 146, + 380, + 201 + ], + "type": "text", + "content": "Fig. 11: Measuring Peak Pareidolia. Left: Subjects were asked how many faces they see in each noise image. We plot the average number of faces detected as a function of noise frequency (examples on x-axis), the mean over all subjects and its " + }, + { + "bbox": [ + 30, + 146, + 380, + 201 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 30, + 146, + 380, + 201 + ], + "type": "text", + "content": " confidence interval in red. Right: average number of faces detected by our fine-tuned models. This reveals the \"peak pareidolia\" effect predicted in Section 5 across humans and machines." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 206, + 25, + 377, + 141 + ], + "blocks": [ + { + "bbox": [ + 206, + 25, + 377, + 141 + ], + "lines": [ + { + "bbox": [ + 206, + 25, + 377, + 141 + ], + "spans": [ + { + "bbox": [ + 206, + 25, + 377, + 141 + ], + "type": "image", + "image_path": "fa61f090fc589435100f26d5052696fcc4096df385ab7978f92b8a49e1628b2f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 30, + 229, + 379, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 229, + 379, + 253 + ], + "spans": [ + { + "bbox": [ + 30, + 229, + 379, + 253 + ], + "type": "text", + "content": "after applying an inverse Fourier transform. We detail our image stimuli creation method in the Supplement." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 30, + 255, + 379, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 255, + 379, + 350 + ], + "spans": [ + { + "bbox": [ + 30, + 255, + 379, + 350 + ], + "type": "text", + "content": "We find that humans exhibit the model-expected peak pareidolia, with a maximum number of faces detected at a frequency filter width of 16 (Fig. 11, left). The existence of a pareidolic peak at or near this filter width is consistent among all subjects even for those that reported fewer faces overall. Although response time did decrease slightly at higher frequencies, it did not fall off completely at the highest frequency levels, indicating that fewer reported faces were not the result of subjects \"giving up\" on the task. We provide additional details and analysis of this experiment in the Supplement." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 30, + 353, + 379, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 353, + 379, + 424 + ], + "spans": [ + { + "bbox": [ + 30, + 353, + 379, + 424 + ], + "type": "text", + "content": "Finally, we evaluate our fine-tuned models from Section 4 on the same images to test whether machines also exhibit peak pareidolia (Fig. 11, right). In particular, we showed the models 5,000 sampled noise images of varying frequency levels and counted the number of face detections they make with confidence " + }, + { + "bbox": [ + 30, + 353, + 379, + 424 + ], + "type": "inline_equation", + "content": ">10\\%" + }, + { + "bbox": [ + 30, + 353, + 379, + 424 + ], + "type": "text", + "content": ". We find the same characteristic \"pareidolic peak\" where models detect the most faces in medium-complexity images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 31, + 453, + 118, + 465 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 453, + 118, + 465 + ], + "spans": [ + { + "bbox": [ + 31, + 453, + 118, + 465 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 30, + 487, + 380, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 487, + 380, + 584 + ], + "spans": [ + { + "bbox": [ + 30, + 487, + 380, + 584 + ], + "type": "text", + "content": "We have taken initial steps towards the mathematical modelling of pareidolia and build a richly annotated dataset of images for face pareidolia. We showed through experiments on modern face detectors that detecting animal faces may partly explain the emergence of pareidolia in a complex vision system. The Faces in Things dataset can help the community address other questions about how and why pareidolic behavior emerges, a hallmark of humans' robust recognition system. We hope that our findings and dataset will spark further study of pareidolia and its potential use to improve computer vision systems." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 11, + 43, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 11, + 43, + 19 + ], + "spans": [ + { + "bbox": [ + 33, + 11, + 43, + 19 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 9, + 227, + 20 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 9, + 227, + 20 + ], + "spans": [ + { + "bbox": [ + 65, + 9, + 227, + 20 + ], + "type": "text", + "content": "M. Hamilton, S. Stent, V. DuTell, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 33, + 35, + 123, + 46 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 35, + 123, + 46 + ], + "spans": [ + { + "bbox": [ + 33, + 35, + 123, + 46 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 33, + 51, + 379, + 86 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 51, + 379, + 86 + ], + "spans": [ + { + "bbox": [ + 33, + 51, + 379, + 86 + ], + "type": "text", + "content": "We would like to thank Karen Hamilton for her hundreds of hours of annotations for the Faces in Things dataset. We also thank Abhishek Dutta who created the VGG Image Annotation tool, for his kind generosity to support our project." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 33, + 87, + 379, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 87, + 379, + 145 + ], + "spans": [ + { + "bbox": [ + 33, + 87, + 379, + 145 + ], + "type": "text", + "content": "We would like to thank the Microsoft Research Grand Central Resources team for their gracious help performing the experiments in this work. Special thanks to Oleg Losinets and Lifeng Li for their consistent, gracious, and timely help, debugging, and expertise. Without them, none of the experiments could have been run." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 33, + 146, + 379, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 146, + 379, + 205 + ], + "spans": [ + { + "bbox": [ + 33, + 146, + 379, + 205 + ], + "type": "text", + "content": "This material is based upon work supported by the National Science Foundation Graduate Research Fellowship under Grant No. 2021323067. Any opinion, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of their employers, or the National Science Foundation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 33, + 206, + 379, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 206, + 379, + 253 + ], + "spans": [ + { + "bbox": [ + 33, + 206, + 379, + 253 + ], + "type": "text", + "content": "This work is supported by the National Science Foundation under Cooperative Agreement PHY-2019786 (The NSF AI Institute for Artificial Intelligence and Fundamental Interactions, http://iaifi.org/) and the CSAIL MEnTorEd Opportunities in Research (METEOR) Fellowship." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 33, + 255, + 379, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 255, + 379, + 349 + ], + "spans": [ + { + "bbox": [ + 33, + 255, + 379, + 349 + ], + "type": "text", + "content": "Research was sponsored by the United States Air Force Research Laboratory and the United States Air Force Artificial Intelligence Accelerator and was accomplished under Cooperative Agreement Number FA8750-192-1000. The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the United States Air Force or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for Government purposes notwithstanding any copyright notation herein." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 33, + 349, + 379, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 349, + 379, + 384 + ], + "spans": [ + { + "bbox": [ + 33, + 349, + 379, + 384 + ], + "type": "text", + "content": "The authors acknowledge the MIT SuperCloud [38] and Lincoln Laboratory Supercomputing Center for providing HPC resources that have contributed to the research results reported within this paper." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 33, + 401, + 97, + 413 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 401, + 97, + 413 + ], + "spans": [ + { + "bbox": [ + 33, + 401, + 97, + 413 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 36, + 421, + 379, + 583 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 36, + 421, + 379, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 36, + 421, + 379, + 465 + ], + "spans": [ + { + "bbox": [ + 36, + 421, + 379, + 465 + ], + "type": "text", + "content": "1. Abbas, A., Chalup, S.: From Face Recognition to Facial Pareidolia: Analysing Hidden Neuron Activations in CNNs for Cross-Depiction Recognition. In: 2019 International Joint Conference on Neural Networks (IJCNN). pp. 1–8 (Jul 2019). https://doi.org/10.1109/IJCNN.2019.8852013, iSSN: 2161-4393" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 36, + 466, + 379, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 36, + 466, + 379, + 486 + ], + "spans": [ + { + "bbox": [ + 36, + 466, + 379, + 486 + ], + "type": "text", + "content": "2. Beaumont, R.: Clip retrieval: Easily compute clip embeddings and build a clip retrieval system with them. https://github.com/rom1504/clip-retrieval (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 36, + 487, + 379, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 36, + 487, + 379, + 507 + ], + "spans": [ + { + "bbox": [ + 36, + 487, + 379, + 507 + ], + "type": "text", + "content": "3. biubug6: Retinaface in pytorch (Nov 2021), https://github.com/biubug6/Pytorch_Retinaface" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 36, + 507, + 379, + 549 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 36, + 507, + 379, + 549 + ], + "spans": [ + { + "bbox": [ + 36, + 507, + 379, + 549 + ], + "type": "text", + "content": "4. Cai, H., Wu, Q., Corradi, T., Hall, P.: The Cross-Depiction Problem: Computer Vision Algorithms for Recognising Objects in Artwork and in Photographs. arXiv:1505.00110 [cs] (May 2015), http://arxiv.org/abs/1505.00110, arXiv: 1505.00110" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 36, + 550, + 379, + 583 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 36, + 550, + 379, + 583 + ], + "spans": [ + { + "bbox": [ + 36, + 550, + 379, + 583 + ], + "type": "text", + "content": "5. Campbell, F.: How much of the information falling on the retina reaches the visual cortex and how much is stored in the visual memory. Pattern recognition mechanisms 54, 83-94 (1983)" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 102, + 9, + 346, + 21 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 9, + 346, + 21 + ], + "spans": [ + { + "bbox": [ + 102, + 9, + 346, + 21 + ], + "type": "text", + "content": "Seeing Faces in Things: A Model and Dataset for Pareidolia" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 368, + 10, + 379, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 10, + 379, + 19 + ], + "spans": [ + { + "bbox": [ + 368, + 10, + 379, + 19 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 33, + 35, + 379, + 583 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 36, + 35, + 379, + 68 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 36, + 35, + 379, + 68 + ], + "spans": [ + { + "bbox": [ + 36, + 35, + 379, + 68 + ], + "type": "text", + "content": "6. Castrejon, L., Aytar, Y., Vondrick, C., Pirsiavash, H., Torralba, A.: Learning Aligned Cross-Modal Representations from Weakly Aligned Data. In: CVPR (Jun 2016)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 36, + 68, + 379, + 90 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 36, + 68, + 379, + 90 + ], + "spans": [ + { + "bbox": [ + 36, + 68, + 379, + 90 + ], + "type": "text", + "content": "7. Conrad, K.: Die beginnende Schizophrenie. Versuch einer Gestaltanalyse des Wahns (1958)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 36, + 90, + 379, + 111 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 36, + 90, + 379, + 111 + ], + "spans": [ + { + "bbox": [ + 36, + 90, + 379, + 111 + ], + "type": "text", + "content": "8. Deng, J., Guo, J., Zhou, Y., Yu, J., Kotsia, I., Zafeiriou, S.: RetinaFace: Single-stage Dense Face Localisation in the Wild. CVPR (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 36, + 112, + 319, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 36, + 112, + 319, + 123 + ], + "spans": [ + { + "bbox": [ + 36, + 112, + 319, + 123 + ], + "type": "text", + "content": "9. Duda, R., Hart, P., Stork, D.: Pattern Classification. Wiley (2012)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 33, + 123, + 379, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 123, + 379, + 167 + ], + "spans": [ + { + "bbox": [ + 33, + 123, + 379, + 167 + ], + "type": "text", + "content": "10. Dutta, A., Zisserman, A.: The VIA annotation software for images, audio and video. In: Proceedings of the 27th ACM International Conference on Multimedia. MM '19, ACM, New York, NY, USA (2019). https://doi.org/10.1145/3343031.3350535, https://doi.org/10.1145/3343031.3350535" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 33, + 167, + 379, + 199 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 167, + 379, + 199 + ], + "spans": [ + { + "bbox": [ + 33, + 167, + 379, + 199 + ], + "type": "text", + "content": "11. Egger, B., Siegel, M.H., Arora, R., Soltani, A.A., Yildirim, I., Tenenbaum, J.: Inverse rendering best explains face perception under extreme illuminations. In: CogSci (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 33, + 200, + 379, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 200, + 379, + 233 + ], + "spans": [ + { + "bbox": [ + 33, + 200, + 379, + 233 + ], + "type": "text", + "content": "12. Endo, Y., Asanuma, R., Shimojo, S., Akashi, T.: Systematic face pareidolia generation method using cycle-consistent adversarial networks. IEEE Transactions on Electrical and Electronic Engineering 19(4), 535-541 (2024)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 33, + 233, + 379, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 233, + 379, + 265 + ], + "spans": [ + { + "bbox": [ + 33, + 233, + 379, + 265 + ], + "type": "text", + "content": "13. Farid, H., Simoncelli, E.P.: Optimally rotation-equivariant directional derivative kernels. In: International Conference on Computer Analysis of Images and Patterns. pp. 207-214. Springer (1997)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 33, + 266, + 379, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 266, + 379, + 287 + ], + "spans": [ + { + "bbox": [ + 33, + 266, + 379, + 287 + ], + "type": "text", + "content": "14. Felzenszwalb, P., Girshick, R., McAllester, D., Ramanan, D.: Object detection with discriminatively trained part based models. IEEE TPAMI 32(9) (2010)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 33, + 288, + 379, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 288, + 379, + 309 + ], + "spans": [ + { + "bbox": [ + 33, + 288, + 379, + 309 + ], + "type": "text", + "content": "15. Ginosar, S., Haas, D., Brown, T., Malik, J.: Detecting people in cubist art. In: ECCV Workshops (2014)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 33, + 309, + 379, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 309, + 379, + 331 + ], + "spans": [ + { + "bbox": [ + 33, + 309, + 379, + 331 + ], + "type": "text", + "content": "16. Hadjikhani, N., Kveraga, K., Naik, P., Ahlfors, S.P.: Early (n170) activation of face-specific cortex by face-like objects. Neuroreport 20(4), 403 (2009)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 33, + 331, + 379, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 331, + 379, + 374 + ], + "spans": [ + { + "bbox": [ + 33, + 331, + 379, + 374 + ], + "type": "text", + "content": "17. Harrington, A., DuTell, V., Hamilton, M., Tewari, A., Stent, S., Freeman, W.T., Rosenholtz, R.: Coco-periph: Bridging the gap between human and machine perception in the periphery. In: The Twelfth International Conference on Learning Representations" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 33, + 375, + 379, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 375, + 379, + 397 + ], + "spans": [ + { + "bbox": [ + 33, + 375, + 379, + 397 + ], + "type": "text", + "content": "18. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 33, + 397, + 379, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 397, + 379, + 429 + ], + "spans": [ + { + "bbox": [ + 33, + 397, + 379, + 429 + ], + "type": "text", + "content": "19. Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. Proceedings of the International Conference on Learning Representations (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 33, + 430, + 379, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 430, + 379, + 452 + ], + "spans": [ + { + "bbox": [ + 33, + 430, + 379, + 452 + ], + "type": "text", + "content": "20. Hinton, G.E., Roweis, S.: Stochastic neighbor embedding. Advances in neural information processing systems 15 (2002)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 33, + 452, + 379, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 452, + 379, + 485 + ], + "spans": [ + { + "bbox": [ + 33, + 452, + 379, + 485 + ], + "type": "text", + "content": "21. Howard, A.G., Zhu, M., Chen, B., Kalenichenko, D., Wang, W., Weyand, T., Andreetto, M., Adam, H.: Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861 (2017)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 33, + 485, + 379, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 485, + 379, + 517 + ], + "spans": [ + { + "bbox": [ + 33, + 485, + 379, + 517 + ], + "type": "text", + "content": "22. Ji, X., Henriques, J.F., Vedaldi, A.: Invariant information clustering for unsupervised image classification and segmentation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9865-9874 (2019)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 33, + 517, + 379, + 550 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 517, + 379, + 550 + ], + "spans": [ + { + "bbox": [ + 33, + 517, + 379, + 550 + ], + "type": "text", + "content": "23. Kato, M., Mugitani, R.: Pareidolia in Infants. PLoS ONE 10(2) (Feb 2015). https://doi.org/10.1371/journal.pone.0118539, https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4331561/" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 33, + 551, + 379, + 583 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 551, + 379, + 583 + ], + "spans": [ + { + "bbox": [ + 33, + 551, + 379, + 583 + ], + "type": "text", + "content": "24. Khan, M.H., McDonagh, J., Khan, S., Shahabuddin, M., Arora, A., Khan, F.S., Shao, L., Tzimiropoulos, G.: AnimalWeb: A large-scale hierarchical dataset of annotated animal faces. In: CVPR (2020)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "spans": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 9, + 228, + 20 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 9, + 228, + 20 + ], + "spans": [ + { + "bbox": [ + 65, + 9, + 228, + 20 + ], + "type": "text", + "content": "M. Hamilton, S. Stent, V. DuTell, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 31, + 35, + 379, + 583 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 31, + 35, + 379, + 79 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 35, + 379, + 79 + ], + "spans": [ + { + "bbox": [ + 31, + 35, + 379, + 79 + ], + "type": "text", + "content": "25. Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A.A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., et al.: Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences 114(13), 3521-3526 (2017)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 31, + 80, + 379, + 102 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 80, + 379, + 102 + ], + "spans": [ + { + "bbox": [ + 31, + 80, + 379, + 102 + ], + "type": "text", + "content": "26. Leopold, D.A., Rhodes, G.: A comparative view of face perception. Journal of Comparative Psychology 124(3), 233 (2010)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 32, + 103, + 379, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 103, + 379, + 124 + ], + "spans": [ + { + "bbox": [ + 32, + 103, + 379, + 124 + ], + "type": "text", + "content": "27. Li, H., Lin, Z., Shen, X., Brandt, J., Hua, G.: A convolutional neural network cascade for face detection. In: CVPR (2015)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 32, + 125, + 379, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 125, + 379, + 146 + ], + "spans": [ + { + "bbox": [ + 32, + 125, + 379, + 146 + ], + "type": "text", + "content": "28. Li, J., Wang, Y., Wang, C., Tai, Y., Qian, J., Yang, J., Wang, C., Li, J., Huang, F.: DSFD: dual shot face detector. In: CVPR (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 32, + 148, + 379, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 148, + 379, + 179 + ], + "spans": [ + { + "bbox": [ + 32, + 148, + 379, + 179 + ], + "type": "text", + "content": "29. Liao, S., Jain, A.K., Li, S.Z.: A Fast and Accurate Unconstrained Face Detector. PAMI 38(2), 211-223 (Feb 2016). https://doi.org/10.1109/TPAMI.2015.2448075" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 32, + 182, + 379, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 182, + 379, + 225 + ], + "spans": [ + { + "bbox": [ + 32, + 182, + 379, + 225 + ], + "type": "text", + "content": "30. Liu, J., Li, J., Feng, L., Li, L., Tian, J., Lee, K.: Seeing Jesus in toast: Neural and behavioral correlates of face pareidolia. Cortex 53, 60-77 (Apr 2014). https://doi.org/10.1016/j.cortex.2014.01.013, http://www.sciencedirect.com/science/article/pii/S0010945214000288" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 32, + 226, + 379, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 226, + 379, + 247 + ], + "spans": [ + { + "bbox": [ + 32, + 226, + 379, + 247 + ], + "type": "text", + "content": "31. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 32, + 249, + 379, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 249, + 379, + 270 + ], + "spans": [ + { + "bbox": [ + 32, + 249, + 379, + 270 + ], + "type": "text", + "content": "32. Mathias, M., Benenson, R., Pedersoli, M., Van Gool, L.: Face detection without bells and whistles. In: ECCV (2014)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 32, + 271, + 379, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 271, + 379, + 314 + ], + "spans": [ + { + "bbox": [ + 32, + 271, + 379, + 314 + ], + "type": "text", + "content": "33. Mcgugin, R., Gatenby, C., Gore, J., Gauthier, I.: High-resolution imaging of expertise reveals reliable object selectivity in the FFA related to perceptual performance. Proceedings of the National Academy of Sciences of the United States of America 109, 17063-8 (Oct 2012). https://doi.org/10.1073/pnas.1116333109" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 32, + 316, + 379, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 316, + 379, + 349 + ], + "spans": [ + { + "bbox": [ + 32, + 316, + 379, + 349 + ], + "type": "text", + "content": "34. Michaelis, C., Mitzkus, B., Geirhos, R., Rusak, E., Bringmann, O., Ecker, A.S., Bethge, M., Brendel, W.: Benchmarking robustness in object detection: Autonomous driving when winter is coming. arXiv preprint arXiv:1907.07484 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 32, + 349, + 379, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 349, + 379, + 371 + ], + "spans": [ + { + "bbox": [ + 32, + 349, + 379, + 371 + ], + "type": "text", + "content": "35. Mishra, A., Nandan Rai, S., Mishra, A., Jawahar, C.V.: Iit-cfw: A benchmark database of cartoon faces in the wild. In: VASE ECCV Workshops (2016)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 32, + 372, + 379, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 372, + 379, + 415 + ], + "spans": [ + { + "bbox": [ + 32, + 372, + 379, + 415 + ], + "type": "text", + "content": "36. Ranjan, R., Sankar, S., Bansal, A., Bodla, N., Chen, J.C., Patel, V., Castillo, C., Chellappa, R.: Deep Learning for Understanding Faces: Machines May Be Just as Good, or Better, than Humans. IEEE Signal Processing Magazine 35, 66-83 (Jan 2018). https://doi.org/10.1109/MSP.2017.2764116" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 32, + 416, + 379, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 416, + 379, + 449 + ], + "spans": [ + { + "bbox": [ + 32, + 416, + 379, + 449 + ], + "type": "text", + "content": "37. Rein-Lien Hsu, Abdel-Mottaleb, M., Jain, A.: Face detection in color images. PAMI 24(5), 696-706 (May 2002). https://doi.org/10.1109/34.1000242, http://ieeexplore.ieee.org/document/1000242/" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 32, + 450, + 379, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 450, + 379, + 504 + ], + "spans": [ + { + "bbox": [ + 32, + 450, + 379, + 504 + ], + "type": "text", + "content": "38. Reuther, A., Kepner, J., Byun, C., Samsi, S., Arcand, W., Bestor, D., Bergeron, B., Gadepally, V., Houle, M., Hubbell, M., Jones, M., Klein, A., Milechin, L., Mullen, J., Prout, A., Rosa, A., Yee, C., Michaleas, P.: Interactive supercomputing on 40,000 cores for machine learning and data analysis. In: 2018 IEEE High Performance extreme Computing Conference (HPEC). pp. 1-6. IEEE (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 32, + 506, + 379, + 527 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 506, + 379, + 527 + ], + "spans": [ + { + "bbox": [ + 32, + 506, + 379, + 527 + ], + "type": "text", + "content": "39. Rosenfeld, A., Solbach, M.D., Tsotsos, J.K.: Totally looks like-how humans compare, compared to machines. In: ACCV (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 32, + 528, + 379, + 561 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 528, + 379, + 561 + ], + "spans": [ + { + "bbox": [ + 32, + 528, + 379, + 561 + ], + "type": "text", + "content": "40. Rowley, H.A., Baluja, S., Kanade, T.: Human Face Detection in Visual Scenes. In: Touretzky, D.S., Mozer, M.C., Hasselmo, M.E. (eds.) NeurIPS (1996), http://papers.nips.cc/paper/1168-human-face-detection-in-visual-scenes.pdf" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 32, + 562, + 379, + 583 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 562, + 379, + 583 + ], + "spans": [ + { + "bbox": [ + 32, + 562, + 379, + 583 + ], + "type": "text", + "content": "41. Rowley, H.A., Baluja, S., Kanade, T.: Neural network-based face detection. PAMI 20(1), 23-38 (1998)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 102, + 9, + 347, + 21 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 9, + 347, + 21 + ], + "spans": [ + { + "bbox": [ + 102, + 9, + 347, + 21 + ], + "type": "text", + "content": "Seeing Faces in Things: A Model and Dataset for Pareidolia" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 368, + 10, + 379, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 10, + 379, + 19 + ], + "spans": [ + { + "bbox": [ + 368, + 10, + 379, + 19 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 33, + 35, + 379, + 583 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 33, + 35, + 379, + 100 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 35, + 379, + 100 + ], + "spans": [ + { + "bbox": [ + 33, + 35, + 379, + 100 + ], + "type": "text", + "content": "42. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C.W., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., Schramowski, P., Kundurthy, S.R., Crowson, K., Schmidt, L., Kaczmarczyk, R., Jitsev, J.: LAION5b: An open large-scale dataset for training next generation image-text models. In: NeurIPS Datasets and Benchmarks Track (2022), https://openreview.net/forum?id=M3Y74vmsMcY" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 33, + 101, + 379, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 101, + 379, + 123 + ], + "spans": [ + { + "bbox": [ + 33, + 101, + 379, + 123 + ], + "type": "text", + "content": "43. Shakespeare, W.: The Tragedy of Hamlet, Prince of Denmark. The Folio Society (1954)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 33, + 124, + 378, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 124, + 378, + 145 + ], + "spans": [ + { + "bbox": [ + 33, + 124, + 378, + 145 + ], + "type": "text", + "content": "44. Sims, A.: Symptoms in the mind: An introduction to descriptive psychopathology. Bailliere Tindall Publishers (1988)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 33, + 145, + 379, + 177 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 145, + 379, + 177 + ], + "spans": [ + { + "bbox": [ + 33, + 145, + 379, + 177 + ], + "type": "text", + "content": "45. Summerfield, C., Egner, T., Mangels, J., Hirsch, J.: Mistaking a house for a face: neural correlates of misperception in healthy humans. Cerebral cortex 16(4), 500-508 (2006)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 33, + 178, + 379, + 210 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 178, + 379, + 210 + ], + "spans": [ + { + "bbox": [ + 33, + 178, + 379, + 210 + ], + "type": "text", + "content": "46. Takahashi, K., Watanabe, K.: Gaze Cueing by Pareidolia Faces. i-Perception 4(8), 490-492 (Dec 2013). https://doi.org/10.1068/i0617sas, https://doi.org/10.1068/i0617sas" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 33, + 210, + 378, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 210, + 378, + 242 + ], + "spans": [ + { + "bbox": [ + 33, + 210, + 378, + 242 + ], + "type": "text", + "content": "47. Takahashi, K., Watanabe, K.: Seeing Objects as Faces Enhances Object Detection. i-Perception 6(5) (2015). https://doi.org/10.1177/2041669515606007, https://doi.org/10.1177/2041669515606007" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 33, + 244, + 379, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 244, + 379, + 276 + ], + "spans": [ + { + "bbox": [ + 33, + 244, + 379, + 276 + ], + "type": "text", + "content": "48. Tang, X., Du, D.K., He, Z., Liu, J.: PyramidBox: A Context-Assisted Single Shot Face Detector. In: Ferrari, V., Hebert, M., Sminchisescu, C., Weiss, Y. (eds.) ECCV (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 33, + 277, + 379, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 277, + 379, + 320 + ], + "spans": [ + { + "bbox": [ + 33, + 277, + 379, + 320 + ], + "type": "text", + "content": "49. Taubert, J., Wardle, S.G., Flessert, M., Leopold, D.A., Ungerleider, L.G.: Face Pareidolia in the Rhesus Monkey. Current Biology 27(16), 2505-2509 (Aug 2017). https://doi.org/10.1016/j.cub.2017.06.075, http://www.sciencedirect.com/science/article/pii/S0960982217308126" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 33, + 321, + 379, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 321, + 379, + 341 + ], + "spans": [ + { + "bbox": [ + 33, + 321, + 379, + 341 + ], + "type": "text", + "content": "50. Tsao, D.Y., Livingstone, M.S.: Mechanisms of face perception. Annu. Rev. Neurosci. 31, 411-437 (2008)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 33, + 343, + 379, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 343, + 379, + 364 + ], + "spans": [ + { + "bbox": [ + 33, + 343, + 379, + 364 + ], + "type": "text", + "content": "51. Turk, M., Pentland, A.: Eigenfaces for recognition. J. of Cognitive Neuroscience 3(1) (1991)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 33, + 365, + 379, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 365, + 379, + 385 + ], + "spans": [ + { + "bbox": [ + 33, + 365, + 379, + 385 + ], + "type": "text", + "content": "52. Viola, P., Jones, M.: Rapid object detection using a boosted cascade of simple features. In: CVPR (2001)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 33, + 386, + 379, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 386, + 379, + 397 + ], + "spans": [ + { + "bbox": [ + 33, + 386, + 379, + 397 + ], + "type": "text", + "content": "53. Viola, P., Jones, M.J.: Robust real-time face detection. IJCV 57(2), 137-154 (2004)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 33, + 397, + 379, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 397, + 379, + 440 + ], + "spans": [ + { + "bbox": [ + 33, + 397, + 379, + 440 + ], + "type": "text", + "content": "54. Wardle, S.G., Ewing, L., Malcolm, G.L., Paranjape, S., Baker, C.I.: Children perceive illusory faces in objects as male more often than female. Cognition 235 (2023). https://doi.org/https://doi.org/10.1016/j.cognition.2023.105398, https://www.sciencedirect.com/science/article/pii/S001002772300032X" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 33, + 441, + 379, + 484 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 441, + 379, + 484 + ], + "spans": [ + { + "bbox": [ + 33, + 441, + 379, + 484 + ], + "type": "text", + "content": "55. Wardle, S.G., Paranjape, S., Taubert, J., Baker, C.I.: Illusory faces are more likely to be perceived as male than female. Proceedings of the National Academy of Sciences 119(5) (2022). https://doi.org/10.1073/pnas.2117413119, https://www.pnas.org/doi/abs/10.1073/pnas.2117413119" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 33, + 485, + 379, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 485, + 379, + 517 + ], + "spans": [ + { + "bbox": [ + 33, + 485, + 379, + 517 + ], + "type": "text", + "content": "56. Wardle, S.G., Taubert, J., Teichmann, L., Baker, C.I.: Rapid and dynamic processing of face pareidolia in the human brain. Nature communications 11(1), 4518 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 33, + 518, + 379, + 539 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 518, + 379, + 539 + ], + "spans": [ + { + "bbox": [ + 33, + 518, + 379, + 539 + ], + "type": "text", + "content": "57. Weber, M., Welling, M., Perona, P.: Unsupervised learning of models for recognition. In: ECCV (2000)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 33, + 539, + 379, + 561 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 539, + 379, + 561 + ], + "spans": [ + { + "bbox": [ + 33, + 539, + 379, + 561 + ], + "type": "text", + "content": "58. Westlake, N., Cai, H., Hall, P.: Detecting People in Artwork with CNNs. In: ECCV Workshops (2016)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 33, + 562, + 379, + 583 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 562, + 379, + 583 + ], + "spans": [ + { + "bbox": [ + 33, + 562, + 379, + 583 + ], + "type": "text", + "content": "59. Yang, M.H., Kriegman, D., Ahuja, N.: Detecting faces in images: a survey. PAMI 24(1), 34-58 (Jan 2002). https://doi.org/10.1109/34.982883" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "spans": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 9, + 227, + 20 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 9, + 227, + 20 + ], + "spans": [ + { + "bbox": [ + 65, + 9, + 227, + 20 + ], + "type": "text", + "content": "M. Hamilton, S. Stent, V. DuTell, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 31, + 35, + 380, + 145 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 31, + 35, + 379, + 57 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 35, + 379, + 57 + ], + "spans": [ + { + "bbox": [ + 31, + 35, + 379, + 57 + ], + "type": "text", + "content": "60. Yang, S., Luo, P., Loy, C.C., Tang, X.: WIDER FACE: A Face Detection Benchmark. In: CVPR (2016)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 31, + 57, + 379, + 79 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 57, + 379, + 79 + ], + "spans": [ + { + "bbox": [ + 31, + 57, + 379, + 79 + ], + "type": "text", + "content": "61. Yuille, A.L.: Deformable templates for face recognition. Journal of Cognitive Neuroscience 3(1), 59-70 (1991)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 32, + 79, + 380, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 79, + 380, + 123 + ], + "spans": [ + { + "bbox": [ + 32, + 79, + 380, + 123 + ], + "type": "text", + "content": "62. Zafeiriou, S., Zhang, C., Zhang, Z.: A survey on face detection in the wild: Past, present and future. Computer Vision and Image Understanding 138, 1-24 (Sep 2015). https://doi.org/10.1016/j.cviu.2015.03.015, https://linkinghub.elsevier.com/retrieve/pii/S1077314215000727" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 32, + 123, + 379, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 123, + 379, + 145 + ], + "spans": [ + { + "bbox": [ + 32, + 123, + 379, + 145 + ], + "type": "text", + "content": "63. Zhang, S., Zhu, X., Lei, Z., Shi, H., Wang, X., Li, S.Z.: S3fd: Single shot scale-invariant face detector. In: ICCV (2017)" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 102, + 9, + 346, + 21 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 9, + 346, + 21 + ], + "spans": [ + { + "bbox": [ + 102, + 9, + 346, + 21 + ], + "type": "text", + "content": "Seeing Faces in Things: A Model and Dataset for Pareidolia" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 368, + 10, + 379, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 10, + 379, + 19 + ], + "spans": [ + { + "bbox": [ + 368, + 10, + 379, + 19 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Seeing the Unseen_ A Frequency Prompt Guided Transformer for Image Restoration/73dcd519-18ab-4cac-87c9-f1d814570e98_content_list.json b/2024/Seeing the Unseen_ A Frequency Prompt Guided Transformer for Image Restoration/73dcd519-18ab-4cac-87c9-f1d814570e98_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d5bca83b2133f1c8b191e4c5516c4b8c994d8138 --- /dev/null +++ b/2024/Seeing the Unseen_ A Frequency Prompt Guided Transformer for Image Restoration/73dcd519-18ab-4cac-87c9-f1d814570e98_content_list.json @@ -0,0 +1,2216 @@ +[ + { + "type": "text", + "text": "Seeing the Unseen: A Frequency Prompt Guided Transformer for Image Restoration", + "text_level": 1, + "bbox": [ + 222, + 140, + 781, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shihao Zhou $^{1,2}$ , Jinshan Pan $^{3}$ , Jinglei Shi $^{1*}$ , Duosheng Chen $^{1}$ , Lishen Qu $^{1}$ , and Jufeng Yang $^{1,2}$", + "bbox": [ + 251, + 210, + 750, + 244 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1 VCIP & TMCC & DISSec, College of Computer Science, Nankai University", + "$^{2}$ Nankai International Advanced Research Institute (SHENZHEN·FUTIAN)" + ], + "bbox": [ + 240, + 253, + 759, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 School of Computer Science and Engineering, Nanjing University of Science and Technology", + "bbox": [ + 225, + 282, + 772, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zhoushihao96,duoshengchen,lishenqu}@mail.nankai.edu.cn, \n{jinglei.shi,yangjufeng}@nankai.edu.cn,SDLuran@gmail.com", + "bbox": [ + 284, + 311, + 715, + 339 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. How to explore useful features from images as prompts to guide the deep image restoration models is an effective way to solve image restoration. In contrast to mining spatial relations within images as prompt, which leads to characteristics of different frequencies being neglected and further remaining subtle or undetectable artifacts in the restored image, we develop a Frequency Prompting image restoration method, dubbed FPro, which can effectively provide prompt components from a frequency perspective to guild the restoration model address these differences. Specifically, we first decompose input features into separate frequency parts via dynamically learned filters, where we introduce a gating mechanism for suppressing the less informative elements within the kernels. To propagate useful frequency information as prompt, we then propose a dual prompt block, consisting of a low-frequency prompt modulator (LPM) and a high-frequency prompt modulator (HPM), to handle signals from different bands respectively. Each modulator contains a generation process to incorporate prompting components into the extracted frequency maps, and a modulation part that modifies the prompt feature with the guidance of the decoder features. Experimental results on several popular datasets have demonstrated the favorable performance of our pipeline against SOTA methods on 5 image restoration tasks, including deraining, deraindrop, demoiréing, deblurring, and dehazing. The source code is available at https://github.com/joshyZhou/FPro.", + "bbox": [ + 259, + 369, + 743, + 678 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Image Restoration $\\cdot$ Frequency Prompt $\\cdot$ Transformer", + "bbox": [ + 261, + 688, + 709, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 727, + 377, + 742 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Capturing images in unsatisfactory environments, e.g., rain, haze, usually leads to low-quality ones that accordingly affect the application of downstream tasks in practice. Thus, developing an effective image restoration method to restore clear images from degraded ones is an important task.", + "bbox": [ + 212, + 756, + 787, + 818 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author.", + "bbox": [ + 217, + 824, + 385, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Significant progress has been made due to kinds of the deep learning models [4,5,81], and these deep learning-based approaches become predominant ones as they achieve better performance than the conventional hand-crafted prior-based approaches [3,21,31,44,84].", + "bbox": [ + 212, + 146, + 782, + 207 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing methods, e.g., [33, 72, 81] achieve promising performances in kinds of image restoration tasks. However, these learning-based methods intend to learn a mapping function between degraded images and clear ones, where the characteristics of the specific degradation are less considered. For example, rain streaks tend to obscure the background partially, whereas raindrops typically result in a more pronounced regional occlusion. Accordingly, these models are hindered from generating better results.", + "bbox": [ + 212, + 210, + 784, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "More recently, prompt-learning based methods [50,67,68] serve as an alternative approach to encode useful content of specific degradation for modulating the network, and make a clear performance boost for image restoration. However, we notice that these methods [50,68] pay attention to mining spatial correlations to provide degradation information, whereas the task-specific frequency cues are less studied. Indeed, since various forms of degradation exhibit distinct impacts on image content, they affect information from different frequency bands. Hence, it is crucial to develop an efficient prompt mechanism that explores useful prompts from a frequency perspective for identifying specific characteristics of diverse degradation, which can boost the model to effectively restore images with finer details and non-local structures of the scenes.", + "bbox": [ + 212, + 319, + 785, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This paper proposes a Frequency Prompting image restoration method, dubbed FPro, to modulate the network by encoding degradation-specific frequency cues as prompts. As mentioned above, existing prompt strategies [50,68] focus on mining spatial relations as useful prompts. In this way, differences between the restored image and the real one within frequency domain [25] are ignored, which remain subtle or undetectable artifacts in the spatial domain. Instead, our FPro aims to enjoy benefits from the capability of prompt learning in different frequency bands at multi-scale resolutions to recover clean images.", + "bbox": [ + 212, + 489, + 785, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We present two designs to enhance FPro for general image restoration: 1). We first decouple input features into separate low-/high-frequency parts using a gated dynamic decoupler, as signals in different frequency bands encode image patterns from distinct views, i.e., local details and global structures. To this end, a gating mechanism is introduced to help learn the enhanced low-pass filters by suppressing the less informative elements within the kernel, which are then employed to generate low-frequency maps. Meanwhile, the corresponding high-pass filter is obtained by subtracting the low-pass filter from the identity kernel, for generating high-frequency maps. 2). We propose a Dual Prompt Block (DPB), which consists of two modulators, i.e., the Low-frequency Prompt Modulator (LPM) and the High-frequency Prompt Modulator (HPM), to handle low- and high-frequency information respectively. Each modulator includes (a) a generation part that incorporates prompting components into the extracted frequency maps, which is supposed to help distinguish various elements within features, such as rain patterns in the context of deraining; and (b) a modu", + "bbox": [ + 212, + 613, + 785, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Zhou et al.", + "bbox": [ + 271, + 114, + 344, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "lation part that modifies the prompt feature with the guidance of the feature in the restoration process. In terms of functionality, LPM enhances the low-frequency characteristics through a gating mechanism in the Fourier domain before injecting the prompting components, which is proven equivalent to dynamic large-kernel depth-wise convolution in the spatial domain while computationally efficient, and then encodes low-frequency interactions via global cross-attention. As a complement, HPM applies a locally-enhanced gating mechanism to obtain useful high-frequency signals, and then encodes high-frequency interactions via local cross-attention.", + "bbox": [ + 212, + 146, + 787, + 281 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We summarize our main contributions as follows:", + "bbox": [ + 240, + 282, + 594, + 297 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose FPro, which benefits from prompting learning of frequency components for general image restoration. Instead of mining spatial relations as in previous methods, we explore frequency maps to encode specific degradation information as prompts to guide the image restoration model for restoring finer details and the global structure of the scenes.", + "- We decouple input features into different frequency bands using learnable low-pass filters, and propose a dual prompt block, which is composed of low-frequency prompt modulator (LPM) and high-frequency prompt modulator (HPM), to explore both details and structures for better restoration.", + "- Experimental results on 5 restoration tasks: deraining, deraindrop, demoiréing, deblurring and dehazing, show that FPro achieves favorable performance against state-of-the-art methods." + ], + "bbox": [ + 225, + 306, + 790, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 508, + 387, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image Restoration. The goal of the image restoration task is to recover high-quality images from the degraded ones. Going beyond conventional prior-based solutions [3, 21], this community has witnessed the great success of a body of learning-based approaches [34, 48, 79]. Despite the promising results obtained by various CNN-based architectures [9, 35, 58], the main concern for methods of this kind is that they pose a limited receptive field problem of the basic convolution operation. This means that the feature map contains less global context (corresponding to low-frequency characteristics in an image), and the final prediction can get stuck in this limitation. This drawback has motivated the increased interest in exploring components to capture desired global cues, like attention mechanisms [10, 45, 60], where better restoration performance can be achieved. For instance, MIRNet [82] proposes a dual attention unit to capture contextual information in dual dimensions. NLSN [41] employs a self-attention mechanism to collect global correlation information for super-resolution.", + "bbox": [ + 212, + 537, + 787, + 748 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Transformer-based Restoration. In recent years, the idea of using Transformer architecture [64] to address various computer vision-related tasks has been popular. Thanks to their discriminative feature representation capability, they not only earn advantages in solving high-level vision tasks [13, 14, 71], but also are extended to low-level image restoration tasks [27, 85, 88]. Unfortunately, as the complexity of vanilla self-attention is quadratic to the image size,", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "A Frequency Prompt Guided Transformer for Image Restoration", + "bbox": [ + 300, + 114, + 730, + 130 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "this mechanism suffers from non-trivial computational costs in handling high-resolution input. To address this, some attempts have been made to explore efficient transformer architectures [4,72,87]. Specifically, SwinIR [33] introduces a window-based self-attention scheme to improve efficiency. Restormer [81] adopts channel-wise self-attention to reduce the computational costs. The majority of these works have offered reliable solutions to recover clean images, however, some works [11,49] realized that the low-pass filter nature of self-attention, which could lose the high-frequency information, such as textures and edges. Even though these models have achieved superior performance, few high-frequency details can be leveraged to implement image restoration, limiting better recovery as a result. Visual Prompt Learning. Prompt learning, which emerged recently in the NLP field [1], has resulted in rapid advancements in its adaptation to computer vision tasks [17,24,26]. Contrary to high-level vision problems, motivated by high effectiveness, some works also consider seeking proper prompts for the low-level pipelines [40, 74, 78].", + "bbox": [ + 212, + 146, + 787, + 372 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The aim of this paper is not to prompt models for addressing the ALL-in-One problem (in fact, the previous works of [32,40,50] have addressed this nicely by designing various degradation prompt modules). However, our approach is relevant to recent studies [67,68] exploring degradation-specific information for better image restoration results. In contrast to these attempts that generate raw degradation features with a pre-trained model, we prompt the restoration models from a frequency perspective.", + "bbox": [ + 212, + 373, + 787, + 479 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Proposed Method", + "text_level": 1, + "bbox": [ + 215, + 500, + 426, + 517 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Overall Pipeline", + "text_level": 1, + "bbox": [ + 215, + 530, + 397, + 545 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As depicted in Fig. 1, the overview of our proposed FPro contains the upper restoration branch, like existing works [33,81], and the bottom prompt branch to extract informative frequency maps and then modulate them as prompts. Restoration Branch. Given a degraded image $\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}$ as input, FPro first employs a convolution layer to obtain shallow feature $\\mathbf{F}_s \\in \\mathbb{R}^{H \\times W \\times C}$ ; where $H \\times W$ represents the spatial dimension and $C$ is the number of channel. Next, the shallow feature passes through the upper $N_1$ -level encoder-decoder restoration branch to extract deep feature $\\mathbf{F}_d \\in \\mathbb{R}^{H \\times W \\times C}$ . Early layers in Transformer-based models focus on aggregate local patterns [76], whereas the self-attention module acts as a low-pass filter and tends to dilute high-frequency local details [49]. To alleviate the two contradictory factors, we remove the attention mechanism within the encoder of the restoration branch. Specifically, each level of the encoder includes $N_2$ feed-forward network (FFN) [81] and the paired convolution layer for down-sampling. The encoder features are then fused with the ones in the decoder through skip connections by $1 \\times 1$ convolution. For the decoder part, each level is composed of $N_2$ pairs of FFN and multi-head self-attention mechanisms (MSA) [81], along with the convolution layer for up-sampling. Finally, a $3 \\times 3$ convolution layer is employed to deep feature $\\mathbf{F}_d$ for generating residual image $\\mathbf{R} \\in \\mathbb{R}^{H \\times W \\times 3}$ . The restored image $\\hat{\\mathbf{I}}$ is estimated by: $\\hat{\\mathbf{I}} = \\mathbf{I} + \\mathbf{R}$ .", + "bbox": [ + 212, + 553, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Zhou et al.", + "bbox": [ + 271, + 114, + 344, + 126 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1f959115b233ea26db4867f087af8516eb152ff14c61502f2a3eedd4b992396a.jpg", + "image_caption": [ + "Fig. 1: Overview of the proposed FPro. Except for the common upper restoration branch, which is similar to existing methods [33,81], FPro contains another bottom prompt branch to extract informative features from a frequency perspective. Specifically, the primary components of the prompt branch in this framework are the gated dynamic decoupler (GDD) and dual prompt block (DPB). The GDD is employed to decompose the low-frequency components and corresponding high-frequency characteristics from the input features. Then these frequency-specific features are further processed in DPB, i.e., the high-frequency prompt modulator (HPM) and the low-frequency prompt modulator (LPM), which generates representative frequency prompt to facilitate the clear image reconstruction." + ], + "image_footnote": [], + "bbox": [ + 222, + 143, + 781, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Prompt Branch. In this branch, we take as input the shallow feature $\\mathbf{F}_s$ to generate useful frequency prompts, which are further leveraged to facilitate the latent clear image reconstruction. To achieve this goal, we first decompose the input feature into different frequency bands using a gated dynamic decouple (GDD) (see Section 3.2). After that, low-/high-frequency maps are injected with prompt components to distinguish informative elements according to specific tasks, and then modulated as different prompts (i.e., $\\mathbf{F}_{hi}^{out}$ and $\\mathbf{F}_{low}^{out}$ ) to interact with the decoder features by $1\\times 1$ convolution (see Section 3.3). Next, we present the modules of the prompt branch.", + "bbox": [ + 212, + 486, + 787, + 622 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Gated Dynamic Decoupler", + "text_level": 1, + "bbox": [ + 215, + 646, + 480, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Each type of degradation affects image content in different ways. For instance, rain streaks partially occlude the background while raindrops often cause much greater obstruction, which corresponds to touch high-/low-frequency bands respectively. To handle these differences, as shown in Fig. 2, we decompose the input features into separate frequency parts based on gated and dynamically learned filters. The key ingredient is to introduce a gating mechanism to help generate the gated learnable low-pass filter and the corresponding high-pass filter, which are then employed to obtain low- and high-frequency maps, respectively. These filters are dynamically learned for each spatial location and channel group to balance computation burden and feature diversity. Specifically, given the input shallow feature map $\\mathbf{F}_s\\in \\mathbb{R}^{H\\times W\\times C}$ , we firstly predict the low-pass", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "A Frequency Prompt Guided Transformer for Image Restoration", + "bbox": [ + 300, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/bde03dd66a0f3c102343539db53440dd6486179893e19e00d528aea0a9410f04.jpg", + "image_caption": [ + "Fig. 2: Illustrations of the Gated Dynamic Decoupler." + ], + "image_footnote": [], + "bbox": [ + 220, + 142, + 781, + 291 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "filter for each feature channel group, which can be formulated as:", + "bbox": [ + 214, + 343, + 686, + 358 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {F}} _ {s} = \\operatorname {C o n v} _ {1 \\times 1} (\\operatorname {G A P} (\\mathbf {F} _ {s})),\n$$\n", + "text_format": "latex", + "bbox": [ + 397, + 366, + 589, + 385 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {F}} _ {s} = \\hat {\\mathbf {F}} _ {s} \\odot \\phi (\\operatorname {C o n v} _ {1 \\times 1} (\\hat {\\mathbf {F}} _ {s})), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 397, + 388, + 785, + 407 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} ^ {l} = \\mathrm {S o f t m a x} (\\mathcal {B} (\\tilde {\\mathbf {F}} _ {s}))\n$$\n", + "text_format": "latex", + "bbox": [ + 401, + 407, + 558, + 428 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{F}^l\\in \\mathbb{R}^{g\\times k^2\\times 1\\times 1}$ , $g$ is the number of channel groups and $k^2$ corresponds to the kernel size of the learned filter; $\\mathrm{GAP}(\\cdot)$ and $\\mathrm{Conv}_{1\\times 1}(\\cdot)$ are global average pooling layer and convolution operation with the filter size of $1\\times 1$ , respectively; $\\phi (\\cdot)$ denotes sigmoid activation, $\\odot$ refers to the element-wise product, and $\\mathcal{B}(\\cdot)$ means Batch Normalization. Particularly, $\\mathrm{Softmax}(\\cdot)$ is a softmax layer, which ensures the generated filters are low-pass [89]. Then, we apply these learned filters to each group input feature $\\mathbf{F}_i\\in \\mathbb{R}^{H\\times W\\times C_i}$ to obtain low-frequency components:", + "bbox": [ + 214, + 436, + 787, + 546 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {i, c, h, w} ^ {l o} = \\sum_ {p, q} \\mathbf {F} _ {i, p, q} ^ {L} \\mathbf {F} _ {i, c, h + p, w + q}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 560, + 785, + 593 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{F}^L\\in \\mathbb{R}^{g\\times k\\times k}$ is the reshaped filter, $i$ denotes the group index, $C_i = \\frac{C}{g}$ refers to number of the group channel, $c$ means the index of a channel, $h$ and $w$ are spatial coordinates, $p,q\\in \\{-1,0,1\\}$ point to the surrounding locations.", + "bbox": [ + 212, + 597, + 785, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Meanwhile, we invert this process by subtracting the low-pass filter from the identity kernel to attain the high-pass filter, which is employed to generate the corresponding high-frequency components $\\mathbf{F}_{hi}$ .", + "bbox": [ + 212, + 645, + 785, + 690 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Dual Prompt Block", + "text_level": 1, + "bbox": [ + 214, + 709, + 424, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Considering that the extracted features, i.e., low-/high-frequency maps, encode image patterns from distinct views (local detail and main structure of the image). We design the Dual Prompt Block that includes two components, i.e., High-frequency Prompt Modulator (HPM) and Low-frequency Prompt Modulator (LPM), to deal with these feature maps, respectively.", + "bbox": [ + 212, + 732, + 787, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "High-frequency Prompt Modulator. Given the two input feature maps, including the $l$ -level feature $\\mathbf{F}_l \\in \\mathbb{R}^{\\hat{H} \\times \\hat{W} \\times \\hat{C}}$ and high-frequency feature $\\mathbf{F}_{hi} \\in$", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Zhou et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8304b4ce1b3aa19dc61980f1ddb181234139e82114cd6e96adf9522a75660b31.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 145, + 782, + 267 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/922e68f04880e8da8fae2aa4e54cb778321805861a6af23fc0d475e194619642.jpg", + "image_caption": [ + "Fig. 3: Illustrations of the proposed components. (a) High-frequency Prompt Modulator (HPM); (b) Low-frequency Prompt Modulator (LPM)." + ], + "image_footnote": [], + "bbox": [ + 217, + 270, + 782, + 393 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$\\mathbb{R}^{H\\times W\\times C'}$ , we first resize $\\mathbf{F}_{hi}$ and obtain $\\tilde{\\mathbf{F}}_{hi}\\in \\mathbb{R}^{\\hat{H}\\times \\hat{W}\\times C'}$ . Towards highlighting high-frequency characteristics, we employ a gating mechanism to adaptively determine the useful frequency information:", + "bbox": [ + 214, + 457, + 782, + 503 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {F}} _ {h i} = \\tilde {\\mathbf {F}} _ {h i} \\odot \\sigma (\\mathrm {D C o n v} _ {3 \\times 3} (\\tilde {\\mathbf {F}} _ {h i})), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 511, + 784, + 530 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{F}}_{hi} \\in \\mathbb{R}^{\\hat{H} \\times \\hat{W} \\times C'}$ is the processed feature, $\\mathrm{DConv}_{3 \\times 3}(\\cdot)$ denotes a depth-wise convolution operation with the filter size of $3 \\times 3$ , and $\\sigma(\\cdot)$ is the GELU activation function [22]. Then, we leverage the learnable high-frequency prompt components $\\mathbf{P}_{hi} \\in \\mathbb{R}^{\\hat{H} \\times \\hat{W} \\times C'}$ to make adjustments to the input features, which aims to help distinguish various elements, such as rain patterns and streaks of different orientations and magnitudes in the context of deraining:", + "bbox": [ + 214, + 537, + 787, + 633 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {h i} ^ {\\text {p r o m p t}} = \\hat {\\mathbf {F}} _ {h i} \\odot \\mathbf {P} _ {h i}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 421, + 638, + 784, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathbf{F}_{hi}^{prompt} \\in \\mathbb{R}^{\\hat{H} \\times \\hat{W} \\times C'}$ is the obtained high-frequency feature prompt.", + "bbox": [ + 214, + 666, + 751, + 684 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Next, we modify the high-frequency prompt $\\mathbf{F}_{hi}^{prompt}$ according to the input feature $\\mathbf{F}_l$ . To be specific, we utilize a depth-wise convolution operator, which acts as a high-pass filter [49], to enhance the high-frequency sources in the input $\\mathbf{F}_{hi}^{prompt}$ . Then, we generate query $(\\mathbf{Q}_{hi})$ projection from $\\mathbf{F}_l$ , key $(\\mathbf{K}_{hi})$ and value $(\\mathbf{V}_{hi})$ projections from the processed feature map $\\hat{\\mathbf{F}}_{hi}^{prompt} = \\mathrm{DConv}_{3\\times 3}(\\mathbf{F}_{hi}^{prompt})$ , respectively. Meanwhile, as the high-frequency information usually corresponds to image details and is a local feature, it could be redundant to calculate global attention. Therefore, before leveraging the linear layer to obtain the matrices of $\\mathbf{Q}_{hi}$ , $\\mathbf{K}_{hi}$ , and $\\mathbf{V}_{hi}$ , the local window self-attention mechanism is adopted to save computational complexity and capture fine-grained high", + "bbox": [ + 212, + 685, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "A Frequency Prompt Guided Transformer for Image Restoration", + "bbox": [ + 300, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "frequencies, which yields $\\mathbf{Q}_{hi} = W_{p}^{Q_{hi}}\\cdot R(\\mathbf{F}_{l})$ , $\\mathbf{K}_{hi} = W_{p}^{K_{hi}}\\cdot R(\\hat{\\mathbf{F}}_{hi}^{prompt})$ , $\\mathbf{V}_{hi} = W_{p}^{V_{hi}}\\cdot R(\\hat{\\mathbf{F}}_{hi}^{prompt})$ . Where $W_{p}^{(\\cdot)}$ represents the projection matrices, and $R(\\cdot)$ denotes the window partition strategy [39]. Generally, we have $\\mathbf{Q}_{hi}\\in \\mathbb{R}^{\\frac{\\hat{H}\\hat{W}}{M^2}\\times M^2\\times\\hat{C}}$ , $\\mathbf{K}_{hi}\\in \\mathbb{R}^{\\frac{\\hat{H}\\hat{W}}{M^2}\\times \\hat{C}\\times M^2}$ , and $\\mathbf{V}_{hi}\\in \\mathbb{R}^{\\frac{\\hat{H}\\hat{W}}{M^2}\\times M^2\\times\\hat{C}}$ , where $M^2$ is the size of split windows. The attention matrix is thus calculated to tune the high-frequency prompt:", + "bbox": [ + 215, + 143, + 787, + 238 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {h i} ^ {o u t} = \\mathbf {V} _ {h i} \\cdot \\mathrm {S o f t m a x} (\\mathbf {K} _ {h i} \\cdot \\mathbf {Q} _ {h i} / \\sqrt {d}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 251, + 785, + 270 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathbf{F}_{hi}^{out}\\in \\mathbb{R}^{\\hat{H}\\times \\hat{W}\\times \\hat{C}}$ is the output feature map of the high-frequency prompt modulation branch; $d$ is the query/ key dimension, following [33].", + "bbox": [ + 212, + 277, + 782, + 310 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Low-frequency Prompt Modulator. Given the two input feature maps, including the $l$ -level feature $\\mathbf{F}_l \\in \\mathbb{R}^{\\hat{H} \\times \\hat{W} \\times \\hat{C}}$ and low-frequency feature $\\mathbf{F}_{lo} \\in \\mathbb{R}^{H \\times W \\times C'}$ , we first resize $\\mathbf{F}_{lo}$ and obtain $\\tilde{\\mathbf{F}}_{lo} \\in \\mathbb{R}^{\\hat{H} \\times \\hat{W} \\times C'}$ . Towards handling low-frequency signals effectively, we project $\\tilde{\\mathbf{F}}_{lo}$ into the frequency domain via the fast Fourier transform (FFT). Then, a gating mechanism is adopted to control the useful low-frequency components flow forward:", + "bbox": [ + 212, + 310, + 784, + 404 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {F}} _ {l o} = \\mathcal {F} (\\tilde {\\mathbf {F}} _ {l o}) \\odot \\sigma (\\operatorname {C o n v} _ {1 \\times 1} (\\mathcal {F} (\\tilde {\\mathbf {F}} _ {l o}))), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 415, + 785, + 434 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{F}}_{lo}\\in \\mathbb{R}^{\\hat{H}\\times (\\frac{\\hat{W}}{2} +1)\\times 2C^{\\prime}}$ is the processed feature, $\\mathcal{F}(\\cdot)$ represents the FFT. Next, we calibrate the input features by injecting learnable low-frequency prompt components $\\mathbf{P}_{lo}\\in \\mathbb{R}^{\\hat{H}\\times (\\frac{\\hat{W}}{2} +1)\\times 2C^{\\prime}}$ , which is then transformed to spatial domain:", + "bbox": [ + 212, + 446, + 782, + 498 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {l o} ^ {\\text {p r o m p t}} = \\mathcal {F} ^ {- 1} \\left(\\hat {\\mathbf {F}} _ {l o} \\odot \\mathbf {P} _ {l o}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 403, + 512, + 785, + 531 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathbf{F}_{lo}^{prompt} \\in \\mathbb{R}^{\\hat{H} \\times \\hat{W} \\times C'}$ is the generated low-frequency feature prompt, and $\\mathcal{F}^{-1}(\\cdot)$ denotes the inverse FFT.", + "bbox": [ + 212, + 537, + 782, + 570 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Noted, we perform the feature transformation in the Fourier domain for efficient global information interaction. The convolution theorem [46, 55] indicates the Hadamard product of two signals in the Fourier domain equals to implement the Fourier transform of a convolution of these two signals in the original spatial domain. Base on this insight, we can combine Eq.(6) and Eq.(7):", + "bbox": [ + 212, + 571, + 784, + 647 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {F} _ {l o} ^ {\\text {p r o m p t}} = \\mathcal {F} ^ {- 1} \\left(\\mathcal {F} (\\tilde {\\mathbf {F}} _ {l o}) \\odot \\sigma \\left(\\operatorname {C o n v} _ {1 \\times 1} \\left(\\mathcal {F} (\\tilde {\\mathbf {F}} _ {l o})\\right)\\right) \\odot \\mathbf {P} _ {l o}\\right) \\\\ = \\mathcal {F} ^ {- 1} \\left(\\mathcal {F} \\left(\\tilde {\\mathbf {F}} _ {l o} * \\mathcal {F} ^ {- 1} \\left(\\sigma \\left(\\operatorname {C o n v} _ {1 \\times 1} \\left(\\mathcal {F} \\left(\\tilde {\\mathbf {F}} _ {l o}\\right)\\right)\\right) \\odot \\mathbf {P} _ {l o}\\right)\\right)\\right) \\tag {8} \\\\ = \\tilde {\\mathbf {F}} _ {l o} * \\mathcal {F} ^ {- 1} \\left(\\sigma \\left(\\operatorname {C o n v} _ {1 \\times 1} \\left(\\mathcal {F} (\\tilde {\\mathbf {F}} _ {l o})\\right)\\right) \\odot \\mathbf {P} _ {l o}\\right) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 287, + 659, + 784, + 720 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\circledast$ is the convolution operation. Since $\\mathcal{F}^{-1}(\\sigma (\\mathrm{Conv}_{1\\times 1}(\\mathcal{F}(\\tilde{\\mathbf{F}}_{lo})))\\odot \\mathbf{P}_{lo})$ is a tensor that shares the same shape with $\\tilde{\\mathbf{F}}_{lo}$ , it can be served as a dynamic depth-wise convolution kernel as large as $\\tilde{\\mathbf{F}}_{lo}$ in spatial domain while introducing less model complexity.", + "bbox": [ + 212, + 732, + 782, + 794 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Subsequently, we further modulate the low-frequency visual prompt $\\mathbf{F}_{lo}^{prompt}$ with the guidance of the input feature $\\mathbf{F}_l$ . Specifically, we adopt an adaptive average pooling operator, which serves as a low-pass filter [65], to enhance the", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Zhou et al.", + "bbox": [ + 271, + 114, + 346, + 127 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6461215bd5e6a770e95d644a147b0fa7e7d2956ec771ab5b8871fbcf219b827e.jpg", + "table_caption": [ + "Table 1: Quantitative comparison on the SPAD dataset [70] for image deraining." + ], + "table_footnote": [], + "table_body": "
MethodDDN [16]PReNet [56]RCDNet [69]MPRNet [83]SPAIR [51]Uformer-S [72]SCD-Former [18]IDT [75]Restormer [81]DRSformer [7]FPro (Ours)
PSNR ↑36.1640.1643.3643.6444.1046.1346.8947.3447.9848.5348.99
SSIM ↑0.94630.98160.98310.98440.98720.99130.99410.99290.99210.99240.9936
", + "bbox": [ + 227, + 171, + 774, + 224 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/94b349f7baad60840f9d62e6b58938b7e6ba03eaf265a9037451ad4720e6e0f1.jpg", + "table_caption": [ + "Table 2: Quantitative comparison on the AGAN-Data [52] for image deraindrop." + ], + "table_footnote": [], + "table_body": "
MethodEigen's [15]Pix2pix [23]Uformer-S [72]WeatherDiff128 [47]TransWeather [63]DuRN [38]RaindropAttn [54]AttentiveGAN [52]IDT [75]Restormer [81]FPro (Ours)
PSNR ↑21.3128.0229.4229.6630.1731.2431.3731.5931.6331.6831.96
SSIM ↑0.7570.8550.9060.9230.9160.9260.9180.9170.9360.9340.937
", + "bbox": [ + 222, + 263, + 777, + 313 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "low-frequency content in the input $\\mathbf{F}_{lo}^{prompt}$ . After that, we generate query $(\\mathbf{Q}_{lo})$ projection from reshaped $\\mathbf{F}_l$ , key $(\\mathbf{K}_{lo})$ and value $(\\mathbf{V}_{lo})$ projections from the average-pooled feature $\\hat{\\mathbf{F}}_{lo}^{prompt} = \\mathrm{AAP}(\\mathbf{F}_{lo}^{prompt})$ , respectively. Here, $\\mathrm{AAP}(\\cdot)$ means the adaptive average pooling operation. Then, we employ $1\\times 1$ convolutions to yield $\\mathbf{Q}_{lo} = W_p^{Q_{lo}}\\mathbf{F}_l$ , $\\mathbf{K}_{lo} = W_p^{K_{lo}}\\hat{\\mathbf{F}}_{lo}^{prompt}$ , and $\\mathbf{V}_{lo} = W_p^{V_{lo}}\\hat{\\mathbf{F}}_{lo}^{prompt}$ . Here $W_{p}^{(\\cdot)}$ is the $1\\times 1$ convolution. Next, we calculate the dot-product of query and key projections, which generates the transposed-attention $\\mathbf{A}\\in \\mathbb{R}^{\\hat{H}\\hat{W}\\times 1}$ . Overall, the process of modulating the low-frequency prompt is defined as:", + "bbox": [ + 212, + 340, + 787, + 470 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {l o} ^ {o u t} = \\mathbf {V} _ {l o} \\cdot \\operatorname {S o f t m a x} (\\mathbf {K} _ {l o} \\cdot \\mathbf {Q} _ {l o} / \\alpha), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 479, + 785, + 498 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\mathbf{F}_{lo}^{out} \\in \\mathbb{R}^{\\hat{H} \\times \\hat{W} \\times C'}$ is the output feature map of the low-frequency prompt modulation branch; $\\mathbf{Q}_{lo} \\in \\mathbb{R}^{\\hat{H}\\hat{W} \\times C'}$ , $\\mathbf{K}_{lo} \\in \\mathbb{R}^{C' \\times 1}$ , and $\\mathbf{V}_{lo} \\in \\mathbb{R}^{1 \\times C'}$ are the input matrices; $\\alpha$ is the learnable scaling parameter.", + "bbox": [ + 212, + 511, + 784, + 560 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For both low/high-frequency modulators, we perform the attention map calculation several times in parallel, and these results are then concatenated for multi-head self-attention (MSA) [64].", + "bbox": [ + 212, + 561, + 785, + 607 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 630, + 375, + 648 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Experimental settings", + "text_level": 1, + "bbox": [ + 215, + 662, + 442, + 679 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Metrics. We adopt commonly used peak signal-to-noise ratio (PSNR) [73] and structural similarity (SSIM) metrics to evaluate restored images. Meanwhile, perceptual metric NIQE [42] is employed as a non-reference metric. Following previous works [69, 72], PSNR/SSIM computations are implemented on the Y channel in the YCbCr space for the image deraining task, while calculated in the RGB color space for other restoration tasks. In the reported tables, the best and second-best scores are highlighted and underlined, respectively.", + "bbox": [ + 212, + 688, + 787, + 794 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Implementation Details. FPro contains $N_{1} = 3$ levels encoder-decoder, where the encode and decoder share the same $N_{2} = [2,3,6]$ blocks. We set embedding dimensions $C$ as 48, and the attention heads as [2,4,8]. The expanding channel", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "A Frequency Prompt Guided Transformer for Image Restoration", + "bbox": [ + 300, + 114, + 730, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/cb8c158c2627eb421fa411a46f7657705d58e20d1e2fb0997b2fb008d51d22f1.jpg", + "table_caption": [ + "Table 3: Quantitative comparison on the TIP-2018 [61] for image demoiré." + ], + "table_footnote": [], + "table_body": "
MethodAMNet [80]DMCNN [61]UNet [59]WDNet [36]MopNet [19]TAPE-Net [37]FHD2eNet [20]MBCNN [86]Uformer-S [72]Wang et al. [66]FPro (Ours)
PSNR ↑25.4726.1026.4927.1227.4827.5227.7928.4028.6328.8729.25
SSIM ↑0.8330.8440.8640.8540.8610.8660.8670.8710.8720.8940.879
", + "bbox": [ + 220, + 170, + 785, + 227 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/494cb6ae4c5d792e32c9176fbc38757df7fed86b79f2888e12002246ae893eb7.jpg", + "table_caption": [ + "Table 4: Quantitative comparison on the SOTS [29] for image dehazing." + ], + "table_footnote": [], + "table_body": "
MethodAODNet [28]MSCNN [57]DehazeNet [2]EPDN [53]FDGAN [12]AirNet [30]Restormer [81]PromptIR [50]FPro (Ours)
PSNR ↑20.2922.0622.4622.5723.1523.1830.8731.3132.85
SSIM ↑0.8770.9080.8510.8630.9210.9000.9690.9730.983
", + "bbox": [ + 228, + 268, + 774, + 334 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "capacity factor in FFN is 3. The default split window size in HPM is set as $M = 8$ . The pixel-unshuffle and pixel-shuffle are employed for downsampling and upsampling. We use the AdamW optimizer with an initial learning rate of $3 \\times 10^{-4}$ , which is gradually reduced to $1 \\times 10^{-6}$ using cosine annealing, to train FPro. Additionally, we adopt the widely used loss function [68] to constrain the network training.", + "bbox": [ + 212, + 362, + 787, + 454 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Main Results", + "text_level": 1, + "bbox": [ + 215, + 479, + 372, + 492 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Rain Streak Removal. The proposed FPro is compared with general restoration approaches [51,72,81,83] as well as with task-specific methods [7,16,18,56,69, 75]. Tab. 1 shows that FPro makes superior performance over current methods for real image deraining on SPAD [70]. Compared to the previous best approach DRSformer [7], FPro achieves a $0.46\\mathrm{dB}$ performance boost. In addition, FPro obtains 2.1 dB PSNR improvement when compared to the recent model SCD-Former [18]. Fig. 4 provides a visual deraining example, where FPro successfully removes the rain degradation while maintaining the structural information.", + "bbox": [ + 212, + 507, + 787, + 628 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Raindrop Removal. For image deraindrop, we compare FPro with some existing deraindrop methods, including Eigen's [15], Pix2pix [23], TransWeather [63], Uformer [72], WeatherDiff128 [47], DuRN [38], RaindropAttn [54], AttentiveGAN [52], IDT [75], and Restormer [81]. We report the quantitative results on AGAN-Data [52] in Tab. 2. Our FPro obtains the best performance against all considered methods in terms of both PSNR/SSIM scores. FPro makes a 0.28 dB PSNR performance boost over Restormer [81], and 2.3 dB over the recent", + "bbox": [ + 212, + 628, + 787, + 736 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/e137bb23d466f951816bff5f6533476ebe6652864beee47d56c987eeca1cc09d.jpg", + "image_caption": [ + "Fig. 4: Qualitative comparisons on the SPAD [70] for real image deraining." + ], + "image_footnote": [], + "bbox": [ + 223, + 762, + 315, + 810 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/67ecf5d624dc5ba28359dbef46e6a72737710bcb3768954392dd474b3db599ba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 318, + 762, + 408, + 810 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/369d8fd845dcb1c32152af2d7f33a40a62da76979e6540fe4579b8f1a7d79da6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 410, + 762, + 500, + 810 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/483381e9ea8a98f27a7551e3a41af775f8a0abb22be36be18a631065219c1e9c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 762, + 591, + 810 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/8d1163c8fd5fe63bbf97b9f2413a0b181c40c42e5dc6cd0675f38b3f0f325b36.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 594, + 762, + 686, + 810 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/907a70c72d658cd9a11cb62750eb0f88e8575bb447bae85dfacafe02945462a8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 687, + 762, + 777, + 810 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Zhou et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/512b4edc48053fff22afab0915d1fac941980e4d2641aa936269f659d5f6d02d.jpg", + "table_caption": [ + "Table 5: Quantitative comparison on the GoPro [43] for image deblurring." + ], + "table_footnote": [], + "table_body": "
MethodCODE [85]IPT [4]MPRNet [83]MIMO [9]HINet [6]MAXIM [62]Restormer [81]PromptRestorer [68]FPro (Ours)
PSNR ↑31.9432.5832.6632.6832.7132.8632.9233.0633.05
SSIM ↑0.928-0.9590.9590.9590.9610.9610.9620.961
", + "bbox": [ + 228, + 171, + 774, + 237 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/dc49230e5812839ddc4f2f5d85d7304305fc5fcf28221b3d3e36456ee9ec5dc9.jpg", + "table_caption": [ + "Table 6: Effectiveness of GDD." + ], + "table_footnote": [], + "table_body": "
ModelsPSNRSSIM
(a) w/o prompt branch48.400.9920
(b) Multi DC [8]48.520.9926
(c) Multi GDD48.910.9934
(d) Single GDD48.990.9936
", + "bbox": [ + 233, + 277, + 488, + 364 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/01a30543c1dfd7a84cbe05a8a15b6460787b21a5faf5ab392275b91a2e37d5d8.jpg", + "table_caption": [ + "Table 7: Ablation study of DPB." + ], + "table_footnote": [], + "table_body": "
ModelsPSNRSSIM
(a) w/o prompt branch48.400.9920
(b) w/o HPM48.770.9931
(c) w/o LPM48.890.9933
(d) Full48.990.9936
", + "bbox": [ + 519, + 277, + 774, + 364 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "method WeatherDiff $_{128}$ [47]. Fig. 5 shows the visual comparisons, where FPro generates a result with finer details.", + "bbox": [ + 212, + 393, + 784, + 422 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Moiré pattern Removal. We perform image demoiring experiments on TIP-2018 [61] dataset, and compare FPro with ten demoiring methods, including AMNet [80], DMCNN [61], UNet [59], WDNet [36], MopNet [19], TAPE-Net [37], $\\mathrm{FHD}^2\\mathrm{eNet}$ [20], MBCNN [86], Uformer-S [72], and Wang et al. [66]. In Tab. 3, FPro yields a 0.38 performance boost against the previous best method Wang et al. [66], and outperforms the recent model TAPE-Net [37] by 1.73 dB.", + "bbox": [ + 212, + 424, + 785, + 515 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Haze Removal. We perform image dehazing experiments on SOTS [29] benchmark. We compare FPro with eight representative approaches, including AOD-Net [28], MSCNN [57], DehazeNet [2], EPDN [53], FDGAN [12], AirNet [30], Restormer [81], and PromptIR [50]. As shown in Tab. 4, FPro achieves the best scores among all considered methods. Compared to the recent prompt-based method PromptIR [50], FPro makes a substantial performance gain of $1.54\\mathrm{dB}$ .", + "bbox": [ + 212, + 515, + 787, + 606 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Motion Blur Removal. We evaluate image deblurring performance on the GoPro dataset [43]. For synthetic deblurring, we compare FPro with eight representative models: CODE [85], IPT [4], MPRNet [83], MIMO [9], HINet [6], MAXIM [62], Restormer [81] and PromptRestorer [68]. Tab. 5 shows that our FPro achieves competitive performance against recent prompt-based PromptRestorer [68], where we cost half FLOPs. Meanwhile, compared to the recent method CODE [85], FPro obtains a 1.11 dB gain on PSNR while using lower FLOPs.", + "bbox": [ + 212, + 607, + 787, + 727 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/40a4d6f437ca090cac087e76357ab59c2495a4b0ef498302a190cabb0afe0d5d.jpg", + "image_caption": [ + "Raindrop", + "Fig. 5: Qualitative comparisons on the AGAN-Data [52] for image deraindrop." + ], + "image_footnote": [], + "bbox": [ + 223, + 752, + 316, + 800 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/a90c5281145716aba555285ff88823fbc5f98eb2a0ed910b8fc97ea83014f18f.jpg", + "image_caption": [ + "Reference" + ], + "image_footnote": [], + "bbox": [ + 316, + 752, + 406, + 799 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/33f3ac49e2f3876c24abd9196ee934784fecc7cf6e7a5aaa0014ab908c824ff8.jpg", + "image_caption": [ + "RaindropAttn [54]" + ], + "image_footnote": [], + "bbox": [ + 410, + 752, + 500, + 799 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/9a84dd51f770445bd79e7c20fbd51d31af229295d548a65e15cf3d1657f25ae5.jpg", + "image_caption": [ + "Uformer [72]" + ], + "image_footnote": [], + "bbox": [ + 504, + 752, + 593, + 799 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/6e2a4fab8a559a41699158da474d935fa5baea64089aee7ba4222fec0e4e972b.jpg", + "image_caption": [ + "Restormer [81]" + ], + "image_footnote": [], + "bbox": [ + 596, + 752, + 684, + 799 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/0e3a319dcef59ea91c1fb62d49003746c19193733ef0cb88b1ee53b840a9a2c5.jpg", + "image_caption": [ + "FPro" + ], + "image_footnote": [], + "bbox": [ + 689, + 752, + 777, + 799 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "A Frequency Prompt Guided Transformer for Image Restoration", + "bbox": [ + 300, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Analysis and Discussion", + "text_level": 1, + "bbox": [ + 215, + 146, + 460, + 161 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "For the ablation studies, we study different models for rain streak removal on SPAD [70] with $256 \\times 256$ patches for 300K iterations. Testing is conducted on SPAD testing dataset [70].", + "bbox": [ + 212, + 175, + 782, + 219 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Effectiveness of Gated Dynamic Decoupler. To demonstrate the effectiveness of the Gated Dynamic Decoupler, we conduct experiments on different model variants in Tab. 6. Compared to the model equipped with Multiple Dynamic Convolution [8] (DC) for separating different frequency parts (b), directly replacing it with GDD (c) results in a performance gain of $0.39\\mathrm{dB}$ in terms of PSNR. Meanwhile, instead of injecting GDD into each DPB (c) to employ multiple decouplers, we attempt to share one GDD module to divide the low-/high frequency information (d), which slightly reduces the complexity $(0.02\\mathrm{M})$ of the whole framework and brings a $0.08\\mathrm{dB}$ performance boost.", + "bbox": [ + 212, + 220, + 784, + 356 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Effectiveness of Dual Propmt Block. To investigate the developed DPB, ablation studies are performed in Tab. 7. Disabling HPM or LPM leads to a clear drop of $0.22\\mathrm{dB}$ and $0.1\\mathrm{dB}$ , respectively. These experimental results indicate that both HPM and LPM play a positive role in restoring high-quality images. Moreover, we present visualizations to better show the effect of DPB. As shown in Fig. 7, we visualize the generated low-/high-frequency feature maps from each branch along with the analysis in the Fourier domain, where the low-frequency prompt feature encodes information such as structures while the high-frequency prompt one focuses on information such as edges and texture. Meantime, we provide visual comparisons in Fig. 8 to show the effectiveness of the proposed HPM/LPM.", + "bbox": [ + 212, + 358, + 784, + 523 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Perceptual Quality Assessment. To test the perceptual quality of the proposed FPro, following [7], we randomly choose 20 rainy images under real-world scenes from Internet-Data [70] to perform the evaluation. As shown in Tab. 9, compared to other considered methods, FPro achieves a lower NIQE score, which means the generated results contain clearer content and better perceptual quality. Through qualitative comparison in Fig. 6, FPro obtains a visually pleasant result against other models, indicating that it handles unseen degradation well.", + "bbox": [ + 212, + 525, + 785, + 630 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Model Efficiency. We provide the comparison of performance (PSNR), complexity (FLOPs and Parameters), and latency (Run-times) for image deraining. FLOPs and Runtimes are measured when input with the size of $256 \\times 256$ , and PSNR scores are tested on SPAD [70]. As shown in Tab. 8, though FPro achieves better performance in terms of PSNR metric, it has less model complexity than", + "bbox": [ + 212, + 631, + 784, + 705 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/622a81c1edf129dc614b540540abb783051853d29afe715c83c1d61d78cb430b.jpg", + "image_caption": [ + "Rainy" + ], + "image_footnote": [], + "bbox": [ + 222, + 734, + 313, + 785 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/8c08f6bd07e0097d3f98e6939644e6ac2d07d80b2ffa0185fc9ede4c8b6c71af.jpg", + "image_caption": [ + "Restormer [81]", + "Fig.6: Qualitative comparisons with state-of-the-art methods on Internet-Data [70] for real rain removal. (Zoom in for a better view.)" + ], + "image_footnote": [], + "bbox": [ + 316, + 734, + 410, + 786 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/a6831f37ff6f2669140cf55022036189844a67d5b00590ae30b3eec40bb333d9.jpg", + "image_caption": [ + "Uformer [72]" + ], + "image_footnote": [], + "bbox": [ + 413, + 734, + 504, + 786 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/d53d1fe432d337afdbae48eb39480b493db010fea9c8f69b48ce6a179c8c5b6e.jpg", + "image_caption": [ + "IDT [75]" + ], + "image_footnote": [], + "bbox": [ + 506, + 734, + 599, + 786 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/b158043eeccd22d548cb11f36f0adfa8091e44abbe93557bbbdb84de76dd7d25.jpg", + "image_caption": [ + "DRSformer [7]" + ], + "image_footnote": [], + "bbox": [ + 602, + 734, + 694, + 786 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/a7dfed3a940587bed00847112344ffa4e57d3b510543ac28aa2e58dccc9184e7.jpg", + "image_caption": [ + "FPro" + ], + "image_footnote": [], + "bbox": [ + 697, + 734, + 789, + 786 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Zhou et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/57af3348afa766a14815e89fc8b9036059a92527ad826c2bb86fc75c11e4c203.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 222, + 143, + 308, + 210 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/4f8ba90a5cc279e9f070880d31ab5c65b9b352eba6311dd87553d0ac737e4126.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 313, + 143, + 400, + 210 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/de544173de0447c58b119a3f87ce1ae4c5a24e4f0a8ae7bfe3263e4ac76fe934.jpg", + "image_caption": [ + "(a) Low-frequency", + "Fig. 7: Feature analysis. We visualize the features from the LPM branch (a), and the HPM one (b). In the right-bottom, we show the results of the average features over the channel dimension in the Fourier domain. (Zoom in for a better view.)" + ], + "image_footnote": [], + "bbox": [ + 222, + 217, + 308, + 282 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/34b08a68a9c0f83fc545ebf8119f603bca049f0c2a9551d81f0151e411e8392c.jpg", + "image_caption": [ + "(b) High-frequency" + ], + "image_footnote": [], + "bbox": [ + 313, + 218, + 398, + 282 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/277c1ccfb0ce356cd9d8e8b066edd0183de4b28225b9fac6cd2208bfed7e9bc2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 429, + 143, + 513, + 209 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/5436f3a955f6ebb47d0473590704236ec63268e5ef087bb88242c63610295d7b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 145, + 599, + 209 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/d793089ce4bc8969d46fe3be9242a946ad5fb13c142acef0bb52d34b958e4086.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 602, + 145, + 684, + 209 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/1b268fdb2ef23b77f007d88c9bc908dbf45bea54b68fd6da089976e86818d1c4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 687, + 145, + 772, + 209 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/4fa86cc22f5a89a9583bdb53ebd09039774265983adf3bbec43d06b54f39f78a.jpg", + "image_caption": [ + "(a) w/o LPM", + "(e) w/LPM", + "Fig. 8: Effect of DBP. Columns 1 and 3 show low-pass and high-pass filtered results, while columns 2 and 4 show the difference (Diff.) between processed results with corresponding filtered ground-truth. Compared with (a), FPro w/ LPM (e) performs better in capturing information such as structures, resulting in fewer erroneous predictions (f). Compared with (c), FPro w/ HPM (g) restores clear edges and shapes, which indicates it enjoys the benefits from the high-frequency information prompt. (Zoom in for a better view.)" + ], + "image_footnote": [], + "bbox": [ + 429, + 218, + 514, + 282 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/7802302ae02ba88c4b3b86499871b0cc829366fa220a0bb25e81f93f43072b1d.jpg", + "image_caption": [ + "(b) Diff.", + "(f) Diff." + ], + "image_footnote": [], + "bbox": [ + 516, + 218, + 601, + 282 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/979bdc46fd9a2a949b09d9d5fb811cc58e9517113ea47f9e4071e51b82a7378c.jpg", + "image_caption": [ + "(c) w/o HPM", + "(g) w/ HPM" + ], + "image_footnote": [], + "bbox": [ + 602, + 218, + 684, + 282 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/51a0aa5820758f9790116a3335d368233db2b35330b57d05dd8041426464facd.jpg", + "image_caption": [ + "(d) Diff.", + "(h) Diff." + ], + "image_footnote": [], + "bbox": [ + 687, + 218, + 772, + 282 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Restormer [81] and DRSformer [7]. Compared to other CNN-/Transformer-based approaches, FPro still has a less or comparable model complexity.", + "bbox": [ + 212, + 494, + 784, + 526 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Comparisons with Alternatives to FPro.", + "text_level": 1, + "bbox": [ + 214, + 537, + 542, + 551 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "To further demonstrate the superiority of FPro, we compare it with recent prompt-based methods that mine spatial relations as prompts, including PromptIR [50] and PromptRestorer [68]. As shown in Tab. 10, following PromptIR [50], we train and validate FPro on Rain100L [77]. We achieve a substantial gain of 2.16 dB over PromptIR, and a 0.16 dB performance boost against PromptRestore", + "bbox": [ + 212, + 551, + 544, + 688 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/6aa40f3548898e6f3f59c5cf1a4b273d734022d3d4fece14d4b63e89ea163ca2.jpg", + "table_caption": [ + "Table 10: Comparisons with alternatives to FPro." + ], + "table_footnote": [], + "table_body": "
ModelsParamsFLOPsPSNR
PromptIR [50]35.617337.04
PromptRestorer [68]24.418639.04
FPro22.38239.20
", + "bbox": [ + 560, + 609, + 785, + 670 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/bc419c273db9d6d2d9dc6926a74d85cfeab91a90587771ff8ebe2fc8655e050a.jpg", + "table_caption": [ + "Table 8: Model efficiency analysis on SPAD [70]." + ], + "table_footnote": [], + "table_body": "
MethodMPRNet [83]SwinIR [33]Uformer-S [72]Restormer [81]IDT [75]DRSformer [7]FPro
FLOPs/G175.8238.043.9174.761.9242.981.9
Parameters/M20.111.520.626.116.433.722.3
Run-times/s0.031.830.120.140.280.080.08
PSNR/dB43.6444.9746.1347.9847.3448.5348.99
", + "bbox": [ + 220, + 763, + 789, + 835 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "A Frequency Prompt Guided Transformer for Image Restoration", + "bbox": [ + 300, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/5a660243d3365e9b481b0cd826e993dcde553dc6252e88c3ca2a8e219807045e.jpg", + "table_caption": [ + "Table 9: Quantitative results (NIQE) for real-world image deraining." + ], + "table_footnote": [], + "table_body": "
Methods|Input Uformer-S [72] Restormer [81] IDT [75] DRSformer [7] FPro
NIQE ↓5.80125.69715.66315.60855.59425.2999
", + "bbox": [ + 222, + 171, + 785, + 219 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/9cc4a6674949be79df7209c8c806c8782f88704b18e459bda2c9125addbf31f4.jpg", + "image_caption": [ + "(a) Input" + ], + "image_footnote": [], + "bbox": [ + 243, + 237, + 500, + 349 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/c05e9955531fb6c91afd56c7ccacd582731c6cab10d72208127accbb082de0e9.jpg", + "image_caption": [ + "(b) FPro", + "Fig. 9: Examples of erroneous reconstruction are shown, where heavy degradation in nighttime real-world scenes leads to typical failures of FPro." + ], + "image_footnote": [], + "bbox": [ + 501, + 238, + 759, + 349 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 436, + 359, + 452 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we explore the benefits of prompt learning from a frequency perspective for the task of image restoration. First, when dynamic decoupling the input features with a gating mechanism to select representative elements, we obtain the related frequency components with regard to the specific degradation removal task. Then, we propose modulating the low-/high-frequency signals with separate branches, which concern the intrinsic characteristics of feature maps from different frequency bands. With these modules, our proposed FPro surpasses previous state-of-the-art methods in several image restoration tasks, while performing competitively in terms of computational cost.", + "bbox": [ + 212, + 482, + 787, + 619 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations. There remain many avenues for further improvements. For instance, one could achieve better performance by addressing failure cases shown in Fig. 9, where FPro meets challenges in dealing with heavy degradation in the nighttime real-world scene. Intuitively, collecting a large-scale real-world dataset is a potential direction for improvements.", + "bbox": [ + 212, + 621, + 787, + 696 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 215, + 733, + 403, + 752 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This work was supported by Natural Science Foundation of Tianjin, China (No. 20JCJQJC00020), the National Natural Science Foundation of China (Nos. U22B2049, 62302240), Fundamental Research Funds for the Central Universities, and Supercomputing Center of Nankai University (NKSC).", + "bbox": [ + 214, + 779, + 785, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Zhou et al.", + "bbox": [ + 271, + 114, + 346, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Brown, T.B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., Ziegler, D.M., Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Chess, B., Clark, J., Berner, C., McCandlish, S., Radford, A., Sutskever, I., Amodei, D.: Language models are few-shot learners. In: NeurIPS (2020)", + "2. Cai, B., Xu, X., Jia, K., Qing, C., Tao, D.: Dehazenet: An end-to-end system for single image haze removal. TIP (2016)", + "3. Chantas, G., Galatsanos, N.P., Molina, R., Katsaggelos, A.K.: Variational bayesian image restoration with a product of spatially weighted total variation image priors. TIP (2009)", + "4. Chen, H., Wang, Y., Guo, T., Xu, C., Deng, Y., Liu, Z., Ma, S., Xu, C., Xu, C., Gao, W.: Pre-trained image processing transformer. In: CVPR (2021)", + "5. Chen, L., Chu, X., Zhang, X., Sun, J.: Simple baselines for image restoration. In: ECCV (2022)", + "6. Chen, L., Lu, X., Zhang, J., Chu, X., Chen, C.: Hinet: Half instance normalization network for image restoration. In: CVPR Workshops (2021)", + "7. Chen, X., Li, H., Li, M., Pan, J.: Learning a sparse transformer network for effective image deraining. In: CVPR (2023)", + "8. Chen, Y., Dai, X., Liu, M., Chen, D., Yuan, L., Liu, Z.: Dynamic convolution: Attention over convolution kernels. In: CVPR (2020)", + "9. Cho, S.J., Ji, S.W., Hong, J.P., Jung, S.W., Ko, S.J.: Rethinking coarse-to-fine approach in single image deblurring. In: ICCV (2021)", + "0. Deng, X., Dragotti, P.L.: Deep convolutional neural network for multi-modal image restoration and fusion. TPAMI (2021)", + "1. Dong, J., Pan, J., Yang, Z., Tang, J.: Multi-scale residual low-pass filter network for image deblurring. In: ICCV (2023)", + "2. Dong, Y., Liu, Y., Zhang, H., Chen, S., Qiao, Y.: Fd-gan: Generative adversarial networks with fusion-discriminator for single image dehazing. In: AAAI (2020)", + "13. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2021)", + "14. d'Ascoli, S., Touvron, H., Leavitt, M.L., Morcos, A.S., Biroli, G., Sagun, L.: Convit: Improving vision transformers with soft convolutional inductive biases. In: ICML (2021)", + "15. Eigen, D., Krishnan, D., Fergus, R.: Restoring an image taken through a window covered with dirt or rain. In: ICCV (2013)", + "16. Fu, X., Huang, J., Zeng, D., Huang, Y., Ding, X., Paisley, J.: Removing rain from single images via a deep detail network. In: CVPR (2017)", + "17. Gan, Y., Bai, Y., Lou, Y., Ma, X., Zhang, R., Shi, N., Luo, L.: Decorate the newcomers: Visual domain prompt for continual test time adaptation. In: AAAI (2023)", + "18. Guo, Y., Xiao, X., Chang, Y., Deng, S., Yan, L.: From sky to the ground: A large-scale benchmark and simple baseline towards real rain removal. In: ICCV (2023)", + "19. He, B., Wang, C., Shi, B., Duan, L.: Mop moiré patterns using mopnet. In: ICCV (2019)" + ], + "bbox": [ + 218, + 178, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "A Frequency Prompt Guided Transformer for Image Restoration", + "bbox": [ + 300, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "20. He, B., Wang, C., Shi, B., Duan, L.: Fhde $^2$ net: Full high definition demoireing network. In: ECCV (2020)", + "21. He, K., Sun, J., Tang, X.: Single image haze removal using dark channel prior. TPAMI (2010)", + "22. Hendrycks, D., Gimpel, K.: Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415 (2016)", + "23. Isola, P., Zhu, J.Y., Zhou, T., Efros, A.A.: Image-to-image translation with conditional adversarial networks. In: CVPR (2017)", + "24. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: ECCV (2022)", + "25. Jiang, L., Dai, B., Wu, W., Loy, C.C.: Focal frequency loss for image reconstruction and synthesis. In: CVPR (2021)", + "26. Khattak, M.U., Rasheed, H., Maaz, M., Khan, S., Khan, F.S.: Maple: Multi-modal prompt learning. In: CVPR (2023)", + "27. Kong, L., Dong, J., Ge, J., Li, M., Pan, J.: Efficient frequency domain-based transformers for high-quality image deblurring. In: CVPR (2023)", + "28. Li, B., Peng, X., Wang, Z., Xu, J., Feng, D.: Aod-net: All-in-one dehazing network. In: ICCV (2017)", + "29. Li, B., Ren, W., Fu, D., Tao, D., Feng, D., Zeng, W., Wang, Z.: Benchmarking single-image dehazing and beyond. TIP (2018)", + "30. Li, B., Liu, X., Hu, P., Wu, Z., Lv, J., Peng, X.: All-in-one image restoration for unknown corruption. In: CVPR (2022)", + "31. Li, Y., Tan, R.T., Guo, X., Lu, J., Brown, M.S.: Rain streak removal using layer priors. In: CVPR (2016)", + "32. Li, Z., Lei, Y., Ma, C., Zhang, J., Shan, H.: Prompt-in-prompt learning for universal image restoration. arXiv preprint arXiv:2312.05038 (2023)", + "33. Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: Swinir: Image restoration using swin transformer. In: ICCV Workshops (2021)", + "34. Liu, D., Wen, B., Fan, Y., Loy, C.C., Huang, T.S.: Non-local recurrent network for image restoration. In: NeurIPS (2018)", + "35. Liu, J., Yan, M., Zeng, T.: Surface-aware blind image deblurring. TPAMI (2021)", + "36. Liu, L., Liu, J., Yuan, S., Slabaugh, G., Leonardis, A., Zhou, W., Tian, Q.: Wavelet-based dual-branch network for image demoiring. In: ECCV (2020)", + "37. Liu, L., Xie, L., Zhang, X., Yuan, S., Chen, X., Zhou, W., Li, H., Tian, Q.: Tape: Task-agnostic prior embedding for image restoration. In: ECCV (2022)", + "38. Liu, X., Suganuma, M., Sun, Z., Okatani, T.: Dual residual networks leveraging the potential of paired operations for image restoration. In: CVPR (2019)", + "39. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: ICCV (2021)", + "40. Ma, J., Cheng, T., Wang, G., Zhang, Q., Wang, X., Zhang, L.: Prores: Exploring degradation-aware visual prompt for universal image restoration. arXiv preprint arXiv:2306.13653 (2023)", + "41. Mei, Y., Fan, Y., Zhou, Y.: Image super-resolution with non-local sparse attention. In: CVPR (2021)", + "42. Mittal, A., Soundararajan, R., Bovik, A.C.: Making a \"completely blind\" image quality analyzer. IEEE SPL (2012)", + "43. Nah, S., Hyun Kim, T., Mu Lee, K.: Deep multi-scale convolutional neural network for dynamic scene deblurring. In: CVPR (2017)", + "44. Narasimhan, S.G., Nayar, S.K.: Contrast restoration of weather degraded images. TPAMI (2003)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Zhou et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "45. Niu, B., Wen, W., Ren, W., Zhang, X., Yang, L., Wang, S., Zhang, K., Cao, X., Shen, H.: Single image super-resolution via a holistic attention network. In: ECCV (2020)", + "46. Oppenheim, A.: Discrete-time signal processing. Prentice-Hall google schola (1999)", + "47. Ozdenizci, O., Legenstein, R.: Restoring vision in adverse weather conditions with patch-based denoising diffusion models. TPAMI (2023)", + "48. Pan, X., Zhan, X., Dai, B., Lin, D., Loy, C.C., Luo, P.: Exploiting deep generative prior for versatile image restoration and manipulation. TPAMI (2022)", + "49. Park, N., Kim, S.: How do vision transformers work? In: ICLR (2022)", + "50. Potlapalli, V., Zamir, S.W., Khan, S., Khan, F.S.: Prompt: Prompting for all-in-one blind image restoration. In: NeurIPS (2023)", + "51. Purohit, K., Suin, M., Rajagopalan, A., Boddeti, V.N.: Spatially-adaptive image restoration using distortion-guided networks. In: ICCV (2021)", + "52. Qian, R., Tan, R.T., Yang, W., Su, J., Liu, J.: Attentive generative adversarial network for raindrop removal from a single image. In: CVPR (2018)", + "53. Qu, Y., Chen, Y., Huang, J., Xie, Y.: Enhanced pix2pix dehazing network. In: CVPR (2019)", + "54. Quan, Y., Deng, S., Chen, Y., Ji, H.: Deep learning for seeing through window with raindrops. In: ICCV (2019)", + "55. Rabiner, L.R., Gold, B.: Theory and application of digital signal processing. Englewood Cliffs: Prentice-Hall (1975)", + "56. Ren, D., Zuo, W., Hu, Q., Zhu, P., Meng, D.: Progressive image deraining networks: A better and simpler baseline. In: CVPR (2019)", + "57. Ren, W., Liu, S., Zhang, H., Pan, J., Cao, X., Yang, M.H.: Single image dehazing via multi-scale convolutional neural networks. In: ECCV (2016)", + "58. Ren, W., Zhang, J., Pan, J., Liu, S., Ren, J.S., Du, J., Cao, X., Yang, M.H.: Deblurring dynamic scenes via spatially varying recurrent neural networks. TPAMI (2022)", + "59. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: MICCAI (2015)", + "60. Song, X., Zhou, D., Li, W., Dai, Y., Shen, Z., Zhang, L., Li, H.: Tusr-net: Triple unfolding single image dehazing with self-regularization and dual feature to pixel attention. TIP (2023)", + "61. Sun, Y., Yu, Y., Wang, W.: Moiré photo restoration using multiresolution convolutional neural networks. TIP (2018)", + "62. Tu, Z., Talebi, H., Zhang, H., Yang, F., Milanfar, P., Bovik, A., Li, Y.: Maxim: Multi-axis mlp for image processing. In: CVPR (2022)", + "63. Valanarasu, J.M.J., Yasarla, R., Patel, V.M.: Transweather: Transformer-based restoration of images degraded by adverse weather conditions. In: CVPR (2022)", + "64. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017)", + "65. Voigtman, E., Winefordner, J.D.: Low-pass filters for signal averaging. Review of Scientific Instruments (1986)", + "66. Wang, C., He, B., Wu, S., Wan, R., Shi, B., Duan, L.Y.: Coarse-to-fine disentangling demoiréing framework for recaptured screen images. TPAMI (2023)", + "67. Wang, C., Pan, J., Lin, W., Dong, J., Wu, X.M.: Selfpromer: Self-prompt dehazing transformers with depth-consistency. arXiv preprint arXiv:2303.07033 (2023)", + "68. Wang, C., Pan, J., Wang, W., Dong, J., Wang, M., Ju, Y., Chen, J.: Promptstorer: A prompting image restoration method with degradation perception. In: NeurIPS (2023)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "A Frequency Prompt Guided Transformer for Image Restoration", + "bbox": [ + 300, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "69. Wang, H., Xie, Q., Zhao, Q., Meng, D.: A model-driven deep neural network for single image rain removal. In: CVPR (2020)", + "70. Wang, T., Yang, X., Xu, K., Chen, S., Zhang, Q., Lau, R.W.: Spatial attentive single-image deraining with a high quality real rain dataset. In: CVPR (2019)", + "71. Wang, W., Xie, E., Li, X., Fan, D.P., Song, K., Liang, D., Lu, T., Luo, P., Shao, L.: Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In: ICCV (2021)", + "72. Wang, Z., Cun, X., Bao, J., Zhou, W., Liu, J., Li, H.: Uformer: A general u-shaped transformer for image restoration. In: CVPR (2022)", + "73. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. TIP (2004)", + "74. Wu, R., Yang, T., Sun, L., Zhang, Z., Li, S., Zhang, L.: Seesr: Towards semantics-aware real-world image super-resolution. arXiv preprint arXiv:2311.16518 (2023)", + "75. Xiao, J., Fu, X., Liu, A., Wu, F., Zha, Z.J.: Image de-raining transformer. TPAMI (2022)", + "76. Xiao, T., Singh, M., Mintun, E., Darrell, T., Dollar, P., Girshick, R.: Early convolutions help transformers see better. In: NeurIPS (2021)", + "77. Yang, W., Tan, R.T., Feng, J., Guo, Z., Yan, S., Liu, J.: Joint rain detection and removal from a single image with contextualized deep networks. TPAMI (2019)", + "78. Yu, F., Gu, J., Li, Z., Hu, J., Kong, X., Wang, X., He, J., Qiao, Y., Dong, C.: Scaling up to excellence: Practicing model scaling for photo-realistic image restoration in the wild. arXiv preprint arXiv:2401.13627 (2024)", + "79. Yu, K., Wang, X., Dong, C., Tang, X., Loy, C.C.: Path-restore: Learning network path selection for image restoration. TPAMI (2022)", + "80. Yue, H., Mao, Y., Liang, L., Xu, H., Hou, C., Yang, J.: Recaptured screen image demoiring. TCSVT (2021)", + "81. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: CVPR (2022)", + "82. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Learning enriched features for real image restoration and enhancement. In: ECCV (2020)", + "83. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Multi-stage progressive image restoration. In: CVPR (2021)", + "84. Zhang, K., Li, Y., Zuo, W., Zhang, L., Van Gool, L., Timofte, R.: Plug-and-play image restoration with deep denoiser prior. TPAMI (2021)", + "85. Zhao, H., Gou, Y., Li, B., Peng, D., Lv, J., Peng, X.: Comprehensive and delicate: An efficient transformer for image restoration. In: CVPR (2023)", + "86. Zheng, B., Yuan, S., Slabaugh, G., Leonardis, A.: Image demoiring with learnable bandpass filters. In: CVPR (2020)", + "87. Zheng, C., Zhang, Y., Gu, J., Zhang, Y., Kong, L., Yuan, X.: Cross aggregation transformer for image restoration. In: NeurIPS (2022)", + "88. Zhou, S., Chen, D., Pan, J., Shi, J., Yang, J.: Adapt or perish: Adaptive sparse transformer with attentive feature refinement for image restoration. In: CVPR (2024)", + "89. Zou, X., Xiao, F., Yu, Z., Li, Y., Lee, Y.J.: Delving deeper into anti-aliasing in convnets. IJCV (2023)" + ], + "bbox": [ + 215, + 146, + 784, + 784 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Zhou et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Seeing the Unseen_ A Frequency Prompt Guided Transformer for Image Restoration/73dcd519-18ab-4cac-87c9-f1d814570e98_model.json b/2024/Seeing the Unseen_ A Frequency Prompt Guided Transformer for Image Restoration/73dcd519-18ab-4cac-87c9-f1d814570e98_model.json new file mode 100644 index 0000000000000000000000000000000000000000..edb9d339a492fea19fdd61181eae05a58cf2eca3 --- /dev/null +++ b/2024/Seeing the Unseen_ A Frequency Prompt Guided Transformer for Image Restoration/73dcd519-18ab-4cac-87c9-f1d814570e98_model.json @@ -0,0 +1,3448 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.223, + 0.141, + 0.782, + 0.187 + ], + "angle": 0, + "content": "Seeing the Unseen: A Frequency Prompt Guided Transformer for Image Restoration" + }, + { + "type": "text", + "bbox": [ + 0.253, + 0.212, + 0.751, + 0.245 + ], + "angle": 0, + "content": "Shihao Zhou\\(^{1,2}\\), Jinshan Pan\\(^{3}\\), Jinglei Shi\\(^{1*}\\), Duosheng Chen\\(^{1}\\), Lishen Qu\\(^{1}\\), and Jufeng Yang\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.255, + 0.761, + 0.269 + ], + "angle": 0, + "content": "1 VCIP & TMCC & DISSec, College of Computer Science, Nankai University" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.269, + 0.761, + 0.284 + ], + "angle": 0, + "content": "\\(^{2}\\) Nankai International Advanced Research Institute (SHENZHEN·FUTIAN)" + }, + { + "type": "list", + "bbox": [ + 0.241, + 0.255, + 0.761, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.283, + 0.774, + 0.311 + ], + "angle": 0, + "content": "3 School of Computer Science and Engineering, Nanjing University of Science and Technology" + }, + { + "type": "text", + "bbox": [ + 0.285, + 0.312, + 0.717, + 0.34 + ], + "angle": 0, + "content": "{zhoushihao96,duoshengchen,lishenqu}@mail.nankai.edu.cn, \n{jinglei.shi,yangjufeng}@nankai.edu.cn,SDLuran@gmail.com" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.371, + 0.744, + 0.679 + ], + "angle": 0, + "content": "Abstract. How to explore useful features from images as prompts to guide the deep image restoration models is an effective way to solve image restoration. In contrast to mining spatial relations within images as prompt, which leads to characteristics of different frequencies being neglected and further remaining subtle or undetectable artifacts in the restored image, we develop a Frequency Prompting image restoration method, dubbed FPro, which can effectively provide prompt components from a frequency perspective to guild the restoration model address these differences. Specifically, we first decompose input features into separate frequency parts via dynamically learned filters, where we introduce a gating mechanism for suppressing the less informative elements within the kernels. To propagate useful frequency information as prompt, we then propose a dual prompt block, consisting of a low-frequency prompt modulator (LPM) and a high-frequency prompt modulator (HPM), to handle signals from different bands respectively. Each modulator contains a generation process to incorporate prompting components into the extracted frequency maps, and a modulation part that modifies the prompt feature with the guidance of the decoder features. Experimental results on several popular datasets have demonstrated the favorable performance of our pipeline against SOTA methods on 5 image restoration tasks, including deraining, deraindrop, demoiréing, deblurring, and dehazing. The source code is available at https://github.com/joshyZhou/FPro." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.689, + 0.71, + 0.705 + ], + "angle": 0, + "content": "Keywords: Image Restoration \\(\\cdot\\) Frequency Prompt \\(\\cdot\\) Transformer" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.728, + 0.378, + 0.743 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.757, + 0.788, + 0.819 + ], + "angle": 0, + "content": "Capturing images in unsatisfactory environments, e.g., rain, haze, usually leads to low-quality ones that accordingly affect the application of downstream tasks in practice. Thus, developing an effective image restoration method to restore clear images from degraded ones is an important task." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.386, + 0.841 + ], + "angle": 0, + "content": "* Corresponding author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.346, + 0.127 + ], + "angle": 0, + "content": "Zhou et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.208 + ], + "angle": 0, + "content": "Significant progress has been made due to kinds of the deep learning models [4,5,81], and these deep learning-based approaches become predominant ones as they achieve better performance than the conventional hand-crafted prior-based approaches [3,21,31,44,84]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.211, + 0.785, + 0.317 + ], + "angle": 0, + "content": "Existing methods, e.g., [33, 72, 81] achieve promising performances in kinds of image restoration tasks. However, these learning-based methods intend to learn a mapping function between degraded images and clear ones, where the characteristics of the specific degradation are less considered. For example, rain streaks tend to obscure the background partially, whereas raindrops typically result in a more pronounced regional occlusion. Accordingly, these models are hindered from generating better results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.32, + 0.787, + 0.485 + ], + "angle": 0, + "content": "More recently, prompt-learning based methods [50,67,68] serve as an alternative approach to encode useful content of specific degradation for modulating the network, and make a clear performance boost for image restoration. However, we notice that these methods [50,68] pay attention to mining spatial correlations to provide degradation information, whereas the task-specific frequency cues are less studied. Indeed, since various forms of degradation exhibit distinct impacts on image content, they affect information from different frequency bands. Hence, it is crucial to develop an efficient prompt mechanism that explores useful prompts from a frequency perspective for identifying specific characteristics of diverse degradation, which can boost the model to effectively restore images with finer details and non-local structures of the scenes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.49, + 0.787, + 0.61 + ], + "angle": 0, + "content": "This paper proposes a Frequency Prompting image restoration method, dubbed FPro, to modulate the network by encoding degradation-specific frequency cues as prompts. As mentioned above, existing prompt strategies [50,68] focus on mining spatial relations as useful prompts. In this way, differences between the restored image and the real one within frequency domain [25] are ignored, which remain subtle or undetectable artifacts in the spatial domain. Instead, our FPro aims to enjoy benefits from the capability of prompt learning in different frequency bands at multi-scale resolutions to recover clean images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.787, + 0.84 + ], + "angle": 0, + "content": "We present two designs to enhance FPro for general image restoration: 1). We first decouple input features into separate low-/high-frequency parts using a gated dynamic decoupler, as signals in different frequency bands encode image patterns from distinct views, i.e., local details and global structures. To this end, a gating mechanism is introduced to help learn the enhanced low-pass filters by suppressing the less informative elements within the kernel, which are then employed to generate low-frequency maps. Meanwhile, the corresponding high-pass filter is obtained by subtracting the low-pass filter from the identity kernel, for generating high-frequency maps. 2). We propose a Dual Prompt Block (DPB), which consists of two modulators, i.e., the Low-frequency Prompt Modulator (LPM) and the High-frequency Prompt Modulator (HPM), to handle low- and high-frequency information respectively. Each modulator includes (a) a generation part that incorporates prompting components into the extracted frequency maps, which is supposed to help distinguish various elements within features, such as rain patterns in the context of deraining; and (b) a modu" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.301, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "A Frequency Prompt Guided Transformer for Image Restoration" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.282 + ], + "angle": 0, + "content": "lation part that modifies the prompt feature with the guidance of the feature in the restoration process. In terms of functionality, LPM enhances the low-frequency characteristics through a gating mechanism in the Fourier domain before injecting the prompting components, which is proven equivalent to dynamic large-kernel depth-wise convolution in the spatial domain while computationally efficient, and then encodes low-frequency interactions via global cross-attention. As a complement, HPM applies a locally-enhanced gating mechanism to obtain useful high-frequency signals, and then encodes high-frequency interactions via local cross-attention." + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.283, + 0.596, + 0.298 + ], + "angle": 0, + "content": "We summarize our main contributions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.307, + 0.784, + 0.38 + ], + "angle": 0, + "content": "- We propose FPro, which benefits from prompting learning of frequency components for general image restoration. Instead of mining spatial relations as in previous methods, we explore frequency maps to encode specific degradation information as prompts to guide the image restoration model for restoring finer details and the global structure of the scenes." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.382, + 0.784, + 0.439 + ], + "angle": 0, + "content": "- We decouple input features into different frequency bands using learnable low-pass filters, and propose a dual prompt block, which is composed of low-frequency prompt modulator (LPM) and high-frequency prompt modulator (HPM), to explore both details and structures for better restoration." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.443, + 0.792, + 0.485 + ], + "angle": 0, + "content": "- Experimental results on 5 restoration tasks: deraining, deraindrop, demoiréing, deblurring and dehazing, show that FPro achieves favorable performance against state-of-the-art methods." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.307, + 0.792, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.509, + 0.388, + 0.525 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.788, + 0.749 + ], + "angle": 0, + "content": "Image Restoration. The goal of the image restoration task is to recover high-quality images from the degraded ones. Going beyond conventional prior-based solutions [3, 21], this community has witnessed the great success of a body of learning-based approaches [34, 48, 79]. Despite the promising results obtained by various CNN-based architectures [9, 35, 58], the main concern for methods of this kind is that they pose a limited receptive field problem of the basic convolution operation. This means that the feature map contains less global context (corresponding to low-frequency characteristics in an image), and the final prediction can get stuck in this limitation. This drawback has motivated the increased interest in exploring components to capture desired global cues, like attention mechanisms [10, 45, 60], where better restoration performance can be achieved. For instance, MIRNet [82] proposes a dual attention unit to capture contextual information in dual dimensions. NLSN [41] employs a self-attention mechanism to collect global correlation information for super-resolution." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Transformer-based Restoration. In recent years, the idea of using Transformer architecture [64] to address various computer vision-related tasks has been popular. Thanks to their discriminative feature representation capability, they not only earn advantages in solving high-level vision tasks [13, 14, 71], but also are extended to low-level image restoration tasks [27, 85, 88]. Unfortunately, as the complexity of vanilla self-attention is quadratic to the image size," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.346, + 0.127 + ], + "angle": 0, + "content": "Zhou et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.373 + ], + "angle": 0, + "content": "this mechanism suffers from non-trivial computational costs in handling high-resolution input. To address this, some attempts have been made to explore efficient transformer architectures [4,72,87]. Specifically, SwinIR [33] introduces a window-based self-attention scheme to improve efficiency. Restormer [81] adopts channel-wise self-attention to reduce the computational costs. The majority of these works have offered reliable solutions to recover clean images, however, some works [11,49] realized that the low-pass filter nature of self-attention, which could lose the high-frequency information, such as textures and edges. Even though these models have achieved superior performance, few high-frequency details can be leveraged to implement image restoration, limiting better recovery as a result. Visual Prompt Learning. Prompt learning, which emerged recently in the NLP field [1], has resulted in rapid advancements in its adaptation to computer vision tasks [17,24,26]. Contrary to high-level vision problems, motivated by high effectiveness, some works also consider seeking proper prompts for the low-level pipelines [40, 74, 78]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.374, + 0.788, + 0.48 + ], + "angle": 0, + "content": "The aim of this paper is not to prompt models for addressing the ALL-in-One problem (in fact, the previous works of [32,40,50] have addressed this nicely by designing various degradation prompt modules). However, our approach is relevant to recent studies [67,68] exploring degradation-specific information for better image restoration results. In contrast to these attempts that generate raw degradation features with a pre-trained model, we prompt the restoration models from a frequency perspective." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.501, + 0.427, + 0.518 + ], + "angle": 0, + "content": "3 Proposed Method" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.531, + 0.398, + 0.546 + ], + "angle": 0, + "content": "3.1 Overall Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.554, + 0.788, + 0.841 + ], + "angle": 0, + "content": "As depicted in Fig. 1, the overview of our proposed FPro contains the upper restoration branch, like existing works [33,81], and the bottom prompt branch to extract informative frequency maps and then modulate them as prompts. Restoration Branch. Given a degraded image \\(\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}\\) as input, FPro first employs a convolution layer to obtain shallow feature \\(\\mathbf{F}_s \\in \\mathbb{R}^{H \\times W \\times C}\\); where \\(H \\times W\\) represents the spatial dimension and \\(C\\) is the number of channel. Next, the shallow feature passes through the upper \\(N_1\\)-level encoder-decoder restoration branch to extract deep feature \\(\\mathbf{F}_d \\in \\mathbb{R}^{H \\times W \\times C}\\). Early layers in Transformer-based models focus on aggregate local patterns [76], whereas the self-attention module acts as a low-pass filter and tends to dilute high-frequency local details [49]. To alleviate the two contradictory factors, we remove the attention mechanism within the encoder of the restoration branch. Specifically, each level of the encoder includes \\(N_2\\) feed-forward network (FFN) [81] and the paired convolution layer for down-sampling. The encoder features are then fused with the ones in the decoder through skip connections by \\(1 \\times 1\\) convolution. For the decoder part, each level is composed of \\(N_2\\) pairs of FFN and multi-head self-attention mechanisms (MSA) [81], along with the convolution layer for up-sampling. Finally, a \\(3 \\times 3\\) convolution layer is employed to deep feature \\(\\mathbf{F}_d\\) for generating residual image \\(\\mathbf{R} \\in \\mathbb{R}^{H \\times W \\times 3}\\). The restored image \\(\\hat{\\mathbf{I}}\\) is estimated by: \\(\\hat{\\mathbf{I}} = \\mathbf{I} + \\mathbf{R}\\)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.301, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "A Frequency Prompt Guided Transformer for Image Restoration" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.144, + 0.782, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.317, + 0.788, + 0.456 + ], + "angle": 0, + "content": "Fig. 1: Overview of the proposed FPro. Except for the common upper restoration branch, which is similar to existing methods [33,81], FPro contains another bottom prompt branch to extract informative features from a frequency perspective. Specifically, the primary components of the prompt branch in this framework are the gated dynamic decoupler (GDD) and dual prompt block (DPB). The GDD is employed to decompose the low-frequency components and corresponding high-frequency characteristics from the input features. Then these frequency-specific features are further processed in DPB, i.e., the high-frequency prompt modulator (HPM) and the low-frequency prompt modulator (LPM), which generates representative frequency prompt to facilitate the clear image reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.487, + 0.788, + 0.623 + ], + "angle": 0, + "content": "Prompt Branch. In this branch, we take as input the shallow feature \\(\\mathbf{F}_s\\) to generate useful frequency prompts, which are further leveraged to facilitate the latent clear image reconstruction. To achieve this goal, we first decompose the input feature into different frequency bands using a gated dynamic decouple (GDD) (see Section 3.2). After that, low-/high-frequency maps are injected with prompt components to distinguish informative elements according to specific tasks, and then modulated as different prompts (i.e., \\(\\mathbf{F}_{hi}^{out}\\) and \\(\\mathbf{F}_{low}^{out}\\)) to interact with the decoder features by \\(1\\times 1\\) convolution (see Section 3.3). Next, we present the modules of the prompt branch." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.647, + 0.482, + 0.663 + ], + "angle": 0, + "content": "3.2 Gated Dynamic Decoupler" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Each type of degradation affects image content in different ways. For instance, rain streaks partially occlude the background while raindrops often cause much greater obstruction, which corresponds to touch high-/low-frequency bands respectively. To handle these differences, as shown in Fig. 2, we decompose the input features into separate frequency parts based on gated and dynamically learned filters. The key ingredient is to introduce a gating mechanism to help generate the gated learnable low-pass filter and the corresponding high-pass filter, which are then employed to obtain low- and high-frequency maps, respectively. These filters are dynamically learned for each spatial location and channel group to balance computation burden and feature diversity. Specifically, given the input shallow feature map \\(\\mathbf{F}_s\\in \\mathbb{R}^{H\\times W\\times C}\\), we firstly predict the low-pass" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.127 + ], + "angle": 0, + "content": "Zhou et al." + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.143, + 0.782, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.319, + 0.302, + 0.682, + 0.316 + ], + "angle": 0, + "content": "Fig. 2: Illustrations of the Gated Dynamic Decoupler." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.344, + 0.687, + 0.359 + ], + "angle": 0, + "content": "filter for each feature channel group, which can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.398, + 0.367, + 0.59, + 0.386 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {F}} _ {s} = \\operatorname {C o n v} _ {1 \\times 1} (\\operatorname {G A P} (\\mathbf {F} _ {s})),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.398, + 0.389, + 0.786, + 0.408 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {F}} _ {s} = \\hat {\\mathbf {F}} _ {s} \\odot \\phi (\\operatorname {C o n v} _ {1 \\times 1} (\\hat {\\mathbf {F}} _ {s})), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.402, + 0.409, + 0.559, + 0.429 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} ^ {l} = \\mathrm {S o f t m a x} (\\mathcal {B} (\\tilde {\\mathbf {F}} _ {s}))\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.438, + 0.788, + 0.547 + ], + "angle": 0, + "content": "where \\(\\mathbf{F}^l\\in \\mathbb{R}^{g\\times k^2\\times 1\\times 1}\\), \\(g\\) is the number of channel groups and \\(k^2\\) corresponds to the kernel size of the learned filter; \\(\\mathrm{GAP}(\\cdot)\\) and \\(\\mathrm{Conv}_{1\\times 1}(\\cdot)\\) are global average pooling layer and convolution operation with the filter size of \\(1\\times 1\\), respectively; \\(\\phi (\\cdot)\\) denotes sigmoid activation, \\(\\odot\\) refers to the element-wise product, and \\(\\mathcal{B}(\\cdot)\\) means Batch Normalization. Particularly, \\(\\mathrm{Softmax}(\\cdot)\\) is a softmax layer, which ensures the generated filters are low-pass [89]. Then, we apply these learned filters to each group input feature \\(\\mathbf{F}_i\\in \\mathbb{R}^{H\\times W\\times C_i}\\) to obtain low-frequency components:" + }, + { + "type": "equation", + "bbox": [ + 0.384, + 0.561, + 0.786, + 0.594 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {i, c, h, w} ^ {l o} = \\sum_ {p, q} \\mathbf {F} _ {i, p, q} ^ {L} \\mathbf {F} _ {i, c, h + p, w + q}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.787, + 0.645 + ], + "angle": 0, + "content": "where \\(\\mathbf{F}^L\\in \\mathbb{R}^{g\\times k\\times k}\\) is the reshaped filter, \\(i\\) denotes the group index, \\(C_i = \\frac{C}{g}\\) refers to number of the group channel, \\(c\\) means the index of a channel, \\(h\\) and \\(w\\) are spatial coordinates, \\(p,q\\in \\{-1,0,1\\}\\) point to the surrounding locations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.646, + 0.787, + 0.691 + ], + "angle": 0, + "content": "Meanwhile, we invert this process by subtracting the low-pass filter from the identity kernel to attain the high-pass filter, which is employed to generate the corresponding high-frequency components \\(\\mathbf{F}_{hi}\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.71, + 0.425, + 0.726 + ], + "angle": 0, + "content": "3.3 Dual Prompt Block" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.733, + 0.788, + 0.809 + ], + "angle": 0, + "content": "Considering that the extracted features, i.e., low-/high-frequency maps, encode image patterns from distinct views (local detail and main structure of the image). We design the Dual Prompt Block that includes two components, i.e., High-frequency Prompt Modulator (HPM) and Low-frequency Prompt Modulator (LPM), to deal with these feature maps, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "High-frequency Prompt Modulator. Given the two input feature maps, including the \\(l\\)-level feature \\(\\mathbf{F}_l \\in \\mathbb{R}^{\\hat{H} \\times \\hat{W} \\times \\hat{C}}\\) and high-frequency feature \\(\\mathbf{F}_{hi} \\in\\)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.301, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "A Frequency Prompt Guided Transformer for Image Restoration" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.146, + 0.784, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.271, + 0.784, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.404, + 0.785, + 0.434 + ], + "angle": 0, + "content": "Fig. 3: Illustrations of the proposed components. (a) High-frequency Prompt Modulator (HPM); (b) Low-frequency Prompt Modulator (LPM)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.458, + 0.784, + 0.505 + ], + "angle": 0, + "content": "\\(\\mathbb{R}^{H\\times W\\times C'}\\), we first resize \\(\\mathbf{F}_{hi}\\) and obtain \\(\\tilde{\\mathbf{F}}_{hi}\\in \\mathbb{R}^{\\hat{H}\\times \\hat{W}\\times C'}\\). Towards highlighting high-frequency characteristics, we employ a gating mechanism to adaptively determine the useful frequency information:" + }, + { + "type": "equation", + "bbox": [ + 0.382, + 0.512, + 0.785, + 0.531 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {F}} _ {h i} = \\tilde {\\mathbf {F}} _ {h i} \\odot \\sigma (\\mathrm {D C o n v} _ {3 \\times 3} (\\tilde {\\mathbf {F}} _ {h i})), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.539, + 0.788, + 0.634 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{F}}_{hi} \\in \\mathbb{R}^{\\hat{H} \\times \\hat{W} \\times C'}\\) is the processed feature, \\(\\mathrm{DConv}_{3 \\times 3}(\\cdot)\\) denotes a depth-wise convolution operation with the filter size of \\(3 \\times 3\\), and \\(\\sigma(\\cdot)\\) is the GELU activation function [22]. Then, we leverage the learnable high-frequency prompt components \\(\\mathbf{P}_{hi} \\in \\mathbb{R}^{\\hat{H} \\times \\hat{W} \\times C'}\\) to make adjustments to the input features, which aims to help distinguish various elements, such as rain patterns and streaks of different orientations and magnitudes in the context of deraining:" + }, + { + "type": "equation", + "bbox": [ + 0.422, + 0.64, + 0.785, + 0.659 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {h i} ^ {\\text {p r o m p t}} = \\hat {\\mathbf {F}} _ {h i} \\odot \\mathbf {P} _ {h i}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.667, + 0.753, + 0.685 + ], + "angle": 0, + "content": "where \\(\\mathbf{F}_{hi}^{prompt} \\in \\mathbb{R}^{\\hat{H} \\times \\hat{W} \\times C'}\\) is the obtained high-frequency feature prompt." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.686, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Next, we modify the high-frequency prompt \\(\\mathbf{F}_{hi}^{prompt}\\) according to the input feature \\(\\mathbf{F}_l\\). To be specific, we utilize a depth-wise convolution operator, which acts as a high-pass filter [49], to enhance the high-frequency sources in the input \\(\\mathbf{F}_{hi}^{prompt}\\). Then, we generate query \\((\\mathbf{Q}_{hi})\\) projection from \\(\\mathbf{F}_l\\), key \\((\\mathbf{K}_{hi})\\) and value \\((\\mathbf{V}_{hi})\\) projections from the processed feature map \\(\\hat{\\mathbf{F}}_{hi}^{prompt} = \\mathrm{DConv}_{3\\times 3}(\\mathbf{F}_{hi}^{prompt})\\), respectively. Meanwhile, as the high-frequency information usually corresponds to image details and is a local feature, it could be redundant to calculate global attention. Therefore, before leveraging the linear layer to obtain the matrices of \\(\\mathbf{Q}_{hi}\\), \\(\\mathbf{K}_{hi}\\), and \\(\\mathbf{V}_{hi}\\), the local window self-attention mechanism is adopted to save computational complexity and capture fine-grained high" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.128 + ], + "angle": 0, + "content": "Zhou et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.239 + ], + "angle": 0, + "content": "frequencies, which yields \\(\\mathbf{Q}_{hi} = W_{p}^{Q_{hi}}\\cdot R(\\mathbf{F}_{l})\\), \\(\\mathbf{K}_{hi} = W_{p}^{K_{hi}}\\cdot R(\\hat{\\mathbf{F}}_{hi}^{prompt})\\), \\(\\mathbf{V}_{hi} = W_{p}^{V_{hi}}\\cdot R(\\hat{\\mathbf{F}}_{hi}^{prompt})\\). Where \\(W_{p}^{(\\cdot)}\\) represents the projection matrices, and \\(R(\\cdot)\\) denotes the window partition strategy [39]. Generally, we have \\(\\mathbf{Q}_{hi}\\in \\mathbb{R}^{\\frac{\\hat{H}\\hat{W}}{M^2}\\times M^2\\times\\hat{C}}\\), \\(\\mathbf{K}_{hi}\\in \\mathbb{R}^{\\frac{\\hat{H}\\hat{W}}{M^2}\\times \\hat{C}\\times M^2}\\), and \\(\\mathbf{V}_{hi}\\in \\mathbb{R}^{\\frac{\\hat{H}\\hat{W}}{M^2}\\times M^2\\times\\hat{C}}\\), where \\(M^2\\) is the size of split windows. The attention matrix is thus calculated to tune the high-frequency prompt:" + }, + { + "type": "equation", + "bbox": [ + 0.368, + 0.252, + 0.786, + 0.271 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {h i} ^ {o u t} = \\mathbf {V} _ {h i} \\cdot \\mathrm {S o f t m a x} (\\mathbf {K} _ {h i} \\cdot \\mathbf {Q} _ {h i} / \\sqrt {d}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.278, + 0.784, + 0.311 + ], + "angle": 0, + "content": "where \\(\\mathbf{F}_{hi}^{out}\\in \\mathbb{R}^{\\hat{H}\\times \\hat{W}\\times \\hat{C}}\\) is the output feature map of the high-frequency prompt modulation branch; \\(d\\) is the query/ key dimension, following [33]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.311, + 0.785, + 0.405 + ], + "angle": 0, + "content": "Low-frequency Prompt Modulator. Given the two input feature maps, including the \\(l\\)-level feature \\(\\mathbf{F}_l \\in \\mathbb{R}^{\\hat{H} \\times \\hat{W} \\times \\hat{C}}\\) and low-frequency feature \\(\\mathbf{F}_{lo} \\in \\mathbb{R}^{H \\times W \\times C'}\\), we first resize \\(\\mathbf{F}_{lo}\\) and obtain \\(\\tilde{\\mathbf{F}}_{lo} \\in \\mathbb{R}^{\\hat{H} \\times \\hat{W} \\times C'}\\). Towards handling low-frequency signals effectively, we project \\(\\tilde{\\mathbf{F}}_{lo}\\) into the frequency domain via the fast Fourier transform (FFT). Then, a gating mechanism is adopted to control the useful low-frequency components flow forward:" + }, + { + "type": "equation", + "bbox": [ + 0.368, + 0.416, + 0.786, + 0.435 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {F}} _ {l o} = \\mathcal {F} (\\tilde {\\mathbf {F}} _ {l o}) \\odot \\sigma (\\operatorname {C o n v} _ {1 \\times 1} (\\mathcal {F} (\\tilde {\\mathbf {F}} _ {l o}))), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.784, + 0.499 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{F}}_{lo}\\in \\mathbb{R}^{\\hat{H}\\times (\\frac{\\hat{W}}{2} +1)\\times 2C^{\\prime}}\\) is the processed feature, \\(\\mathcal{F}(\\cdot)\\) represents the FFT. Next, we calibrate the input features by injecting learnable low-frequency prompt components \\(\\mathbf{P}_{lo}\\in \\mathbb{R}^{\\hat{H}\\times (\\frac{\\hat{W}}{2} +1)\\times 2C^{\\prime}}\\), which is then transformed to spatial domain:" + }, + { + "type": "equation", + "bbox": [ + 0.405, + 0.513, + 0.786, + 0.532 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {l o} ^ {\\text {p r o m p t}} = \\mathcal {F} ^ {- 1} \\left(\\hat {\\mathbf {F}} _ {l o} \\odot \\mathbf {P} _ {l o}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.539, + 0.784, + 0.571 + ], + "angle": 0, + "content": "where \\(\\mathbf{F}_{lo}^{prompt} \\in \\mathbb{R}^{\\hat{H} \\times \\hat{W} \\times C'}\\) is the generated low-frequency feature prompt, and \\(\\mathcal{F}^{-1}(\\cdot)\\) denotes the inverse FFT." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.572, + 0.785, + 0.648 + ], + "angle": 0, + "content": "Noted, we perform the feature transformation in the Fourier domain for efficient global information interaction. The convolution theorem [46, 55] indicates the Hadamard product of two signals in the Fourier domain equals to implement the Fourier transform of a convolution of these two signals in the original spatial domain. Base on this insight, we can combine Eq.(6) and Eq.(7):" + }, + { + "type": "equation", + "bbox": [ + 0.289, + 0.66, + 0.785, + 0.721 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {F} _ {l o} ^ {\\text {p r o m p t}} = \\mathcal {F} ^ {- 1} \\left(\\mathcal {F} (\\tilde {\\mathbf {F}} _ {l o}) \\odot \\sigma \\left(\\operatorname {C o n v} _ {1 \\times 1} \\left(\\mathcal {F} (\\tilde {\\mathbf {F}} _ {l o})\\right)\\right) \\odot \\mathbf {P} _ {l o}\\right) \\\\ = \\mathcal {F} ^ {- 1} \\left(\\mathcal {F} \\left(\\tilde {\\mathbf {F}} _ {l o} * \\mathcal {F} ^ {- 1} \\left(\\sigma \\left(\\operatorname {C o n v} _ {1 \\times 1} \\left(\\mathcal {F} \\left(\\tilde {\\mathbf {F}} _ {l o}\\right)\\right)\\right) \\odot \\mathbf {P} _ {l o}\\right)\\right)\\right) \\tag {8} \\\\ = \\tilde {\\mathbf {F}} _ {l o} * \\mathcal {F} ^ {- 1} \\left(\\sigma \\left(\\operatorname {C o n v} _ {1 \\times 1} \\left(\\mathcal {F} (\\tilde {\\mathbf {F}} _ {l o})\\right)\\right) \\odot \\mathbf {P} _ {l o}\\right) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.733, + 0.784, + 0.795 + ], + "angle": 0, + "content": "where \\( \\circledast \\) is the convolution operation. Since \\( \\mathcal{F}^{-1}(\\sigma (\\mathrm{Conv}_{1\\times 1}(\\mathcal{F}(\\tilde{\\mathbf{F}}_{lo})))\\odot \\mathbf{P}_{lo}) \\) is a tensor that shares the same shape with \\( \\tilde{\\mathbf{F}}_{lo} \\), it can be served as a dynamic depth-wise convolution kernel as large as \\( \\tilde{\\mathbf{F}}_{lo} \\) in spatial domain while introducing less model complexity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.786, + 0.842 + ], + "angle": 0, + "content": "Subsequently, we further modulate the low-frequency visual prompt \\(\\mathbf{F}_{lo}^{prompt}\\) with the guidance of the input feature \\(\\mathbf{F}_l\\). Specifically, we adopt an adaptive average pooling operator, which serves as a low-pass filter [65], to enhance the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.301, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "A Frequency Prompt Guided Transformer for Image Restoration" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.229, + 0.145, + 0.772, + 0.16 + ], + "angle": 0, + "content": "Table 1: Quantitative comparison on the SPAD dataset [70] for image deraining." + }, + { + "type": "table", + "bbox": [ + 0.228, + 0.172, + 0.776, + 0.226 + ], + "angle": 0, + "content": "
MethodDDN [16]PReNet [56]RCDNet [69]MPRNet [83]SPAIR [51]Uformer-S [72]SCD-Former [18]IDT [75]Restormer [81]DRSformer [7]FPro (Ours)
PSNR ↑36.1640.1643.3643.6444.1046.1346.8947.3447.9848.5348.99
SSIM ↑0.94630.98160.98310.98440.98720.99130.99410.99290.99210.99240.9936
" + }, + { + "type": "table_caption", + "bbox": [ + 0.229, + 0.238, + 0.772, + 0.253 + ], + "angle": 0, + "content": "Table 2: Quantitative comparison on the AGAN-Data [52] for image deraindrop." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.265, + 0.779, + 0.314 + ], + "angle": 0, + "content": "
MethodEigen's [15]Pix2pix [23]Uformer-S [72]WeatherDiff128 [47]TransWeather [63]DuRN [38]RaindropAttn [54]AttentiveGAN [52]IDT [75]Restormer [81]FPro (Ours)
PSNR ↑21.3128.0229.4229.6630.1731.2431.3731.5931.6331.6831.96
SSIM ↑0.7570.8550.9060.9230.9160.9260.9180.9170.9360.9340.937
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.341, + 0.788, + 0.472 + ], + "angle": 0, + "content": "low-frequency content in the input \\(\\mathbf{F}_{lo}^{prompt}\\). After that, we generate query \\((\\mathbf{Q}_{lo})\\) projection from reshaped \\(\\mathbf{F}_l\\), key \\((\\mathbf{K}_{lo})\\) and value \\((\\mathbf{V}_{lo})\\) projections from the average-pooled feature \\(\\hat{\\mathbf{F}}_{lo}^{prompt} = \\mathrm{AAP}(\\mathbf{F}_{lo}^{prompt})\\), respectively. Here, \\(\\mathrm{AAP}(\\cdot)\\) means the adaptive average pooling operation. Then, we employ \\(1\\times 1\\) convolutions to yield \\(\\mathbf{Q}_{lo} = W_p^{Q_{lo}}\\mathbf{F}_l\\), \\(\\mathbf{K}_{lo} = W_p^{K_{lo}}\\hat{\\mathbf{F}}_{lo}^{prompt}\\), and \\(\\mathbf{V}_{lo} = W_p^{V_{lo}}\\hat{\\mathbf{F}}_{lo}^{prompt}\\). Here \\(W_{p}^{(\\cdot)}\\) is the \\(1\\times 1\\) convolution. Next, we calculate the dot-product of query and key projections, which generates the transposed-attention \\(\\mathbf{A}\\in \\mathbb{R}^{\\hat{H}\\hat{W}\\times 1}\\). Overall, the process of modulating the low-frequency prompt is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.374, + 0.48, + 0.786, + 0.499 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {l o} ^ {o u t} = \\mathbf {V} _ {l o} \\cdot \\operatorname {S o f t m a x} (\\mathbf {K} _ {l o} \\cdot \\mathbf {Q} _ {l o} / \\alpha), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.512, + 0.785, + 0.561 + ], + "angle": 0, + "content": "where \\(\\mathbf{F}_{lo}^{out} \\in \\mathbb{R}^{\\hat{H} \\times \\hat{W} \\times C'}\\) is the output feature map of the low-frequency prompt modulation branch; \\(\\mathbf{Q}_{lo} \\in \\mathbb{R}^{\\hat{H}\\hat{W} \\times C'}\\), \\(\\mathbf{K}_{lo} \\in \\mathbb{R}^{C' \\times 1}\\), and \\(\\mathbf{V}_{lo} \\in \\mathbb{R}^{1 \\times C'}\\) are the input matrices; \\(\\alpha\\) is the learnable scaling parameter." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.562, + 0.787, + 0.608 + ], + "angle": 0, + "content": "For both low/high-frequency modulators, we perform the attention map calculation several times in parallel, and these results are then concatenated for multi-head self-attention (MSA) [64]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.631, + 0.376, + 0.649 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.663, + 0.444, + 0.68 + ], + "angle": 0, + "content": "4.1 Experimental settings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.795 + ], + "angle": 0, + "content": "Metrics. We adopt commonly used peak signal-to-noise ratio (PSNR) [73] and structural similarity (SSIM) metrics to evaluate restored images. Meanwhile, perceptual metric NIQE [42] is employed as a non-reference metric. Following previous works [69, 72], PSNR/SSIM computations are implemented on the Y channel in the YCbCr space for the image deraining task, while calculated in the RGB color space for other restoration tasks. In the reported tables, the best and second-best scores are highlighted and underlined, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Implementation Details. FPro contains \\( N_{1} = 3 \\) levels encoder-decoder, where the encode and decoder share the same \\( N_{2} = [2,3,6] \\) blocks. We set embedding dimensions \\( C \\) as 48, and the attention heads as [2,4,8]. The expanding channel" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.127 + ], + "angle": 0, + "content": "Zhou et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.249, + 0.145, + 0.753, + 0.16 + ], + "angle": 0, + "content": "Table 3: Quantitative comparison on the TIP-2018 [61] for image demoiré." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.171, + 0.787, + 0.228 + ], + "angle": 0, + "content": "
MethodAMNet [80]DMCNN [61]UNet [59]WDNet [36]MopNet [19]TAPE-Net [37]FHD2eNet [20]MBCNN [86]Uformer-S [72]Wang et al. [66]FPro (Ours)
PSNR ↑25.4726.1026.4927.1227.4827.5227.7928.4028.6328.8729.25
SSIM ↑0.8330.8440.8640.8540.8610.8660.8670.8710.8720.8940.879
" + }, + { + "type": "table_caption", + "bbox": [ + 0.257, + 0.242, + 0.744, + 0.258 + ], + "angle": 0, + "content": "Table 4: Quantitative comparison on the SOTS [29] for image dehazing." + }, + { + "type": "table", + "bbox": [ + 0.229, + 0.269, + 0.776, + 0.335 + ], + "angle": 0, + "content": "
MethodAODNet [28]MSCNN [57]DehazeNet [2]EPDN [53]FDGAN [12]AirNet [30]Restormer [81]PromptIR [50]FPro (Ours)
PSNR ↑20.2922.0622.4622.5723.1523.1830.8731.3132.85
SSIM ↑0.8770.9080.8510.8630.9210.9000.9690.9730.983
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.363, + 0.788, + 0.455 + ], + "angle": 0, + "content": "capacity factor in FFN is 3. The default split window size in HPM is set as \\( M = 8 \\). The pixel-unshuffle and pixel-shuffle are employed for downsampling and upsampling. We use the AdamW optimizer with an initial learning rate of \\( 3 \\times 10^{-4} \\), which is gradually reduced to \\( 1 \\times 10^{-6} \\) using cosine annealing, to train FPro. Additionally, we adopt the widely used loss function [68] to constrain the network training." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.48, + 0.373, + 0.493 + ], + "angle": 0, + "content": "4.2 Main Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.629 + ], + "angle": 0, + "content": "Rain Streak Removal. The proposed FPro is compared with general restoration approaches [51,72,81,83] as well as with task-specific methods [7,16,18,56,69, 75]. Tab. 1 shows that FPro makes superior performance over current methods for real image deraining on SPAD [70]. Compared to the previous best approach DRSformer [7], FPro achieves a \\(0.46\\mathrm{dB}\\) performance boost. In addition, FPro obtains 2.1 dB PSNR improvement when compared to the recent model SCD-Former [18]. Fig. 4 provides a visual deraining example, where FPro successfully removes the rain degradation while maintaining the structural information." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.788, + 0.737 + ], + "angle": 0, + "content": "Raindrop Removal. For image deraindrop, we compare FPro with some existing deraindrop methods, including Eigen's [15], Pix2pix [23], TransWeather [63], Uformer [72], WeatherDiff128 [47], DuRN [38], RaindropAttn [54], AttentiveGAN [52], IDT [75], and Restormer [81]. We report the quantitative results on AGAN-Data [52] in Tab. 2. Our FPro obtains the best performance against all considered methods in terms of both PSNR/SSIM scores. FPro makes a 0.28 dB PSNR performance boost over Restormer [81], and 2.3 dB over the recent" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.763, + 0.316, + 0.811 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.319, + 0.763, + 0.409, + 0.811 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.763, + 0.501, + 0.811 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.763, + 0.593, + 0.811 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.595, + 0.763, + 0.687, + 0.811 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.688, + 0.763, + 0.779, + 0.811 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.249, + 0.822, + 0.751, + 0.837 + ], + "angle": 0, + "content": "Fig. 4: Qualitative comparisons on the SPAD [70] for real image deraining." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.301, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "A Frequency Prompt Guided Transformer for Image Restoration" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.25, + 0.145, + 0.751, + 0.16 + ], + "angle": 0, + "content": "Table 5: Quantitative comparison on the GoPro [43] for image deblurring." + }, + { + "type": "table", + "bbox": [ + 0.229, + 0.172, + 0.776, + 0.238 + ], + "angle": 0, + "content": "
MethodCODE [85]IPT [4]MPRNet [83]MIMO [9]HINet [6]MAXIM [62]Restormer [81]PromptRestorer [68]FPro (Ours)
PSNR ↑31.9432.5832.6632.6832.7132.8632.9233.0633.05
SSIM ↑0.928-0.9590.9590.9590.9610.9610.9620.961
" + }, + { + "type": "table_caption", + "bbox": [ + 0.252, + 0.251, + 0.467, + 0.264 + ], + "angle": 0, + "content": "Table 6: Effectiveness of GDD." + }, + { + "type": "table", + "bbox": [ + 0.234, + 0.278, + 0.489, + 0.365 + ], + "angle": 0, + "content": "
ModelsPSNRSSIM
(a) w/o prompt branch48.400.9920
(b) Multi DC [8]48.520.9926
(c) Multi GDD48.910.9934
(d) Single GDD48.990.9936
" + }, + { + "type": "table_caption", + "bbox": [ + 0.53, + 0.251, + 0.757, + 0.265 + ], + "angle": 0, + "content": "Table 7: Ablation study of DPB." + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.278, + 0.776, + 0.365 + ], + "angle": 0, + "content": "
ModelsPSNRSSIM
(a) w/o prompt branch48.400.9920
(b) w/o HPM48.770.9931
(c) w/o LPM48.890.9933
(d) Full48.990.9936
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.394, + 0.785, + 0.424 + ], + "angle": 0, + "content": "method WeatherDiff\\(_{128}\\) [47]. Fig. 5 shows the visual comparisons, where FPro generates a result with finer details." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.425, + 0.787, + 0.516 + ], + "angle": 0, + "content": "Moiré pattern Removal. We perform image demoiring experiments on TIP-2018 [61] dataset, and compare FPro with ten demoiring methods, including AMNet [80], DMCNN [61], UNet [59], WDNet [36], MopNet [19], TAPE-Net [37], \\(\\mathrm{FHD}^2\\mathrm{eNet}\\) [20], MBCNN [86], Uformer-S [72], and Wang et al. [66]. In Tab. 3, FPro yields a 0.38 performance boost against the previous best method Wang et al. [66], and outperforms the recent model TAPE-Net [37] by 1.73 dB." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.516, + 0.788, + 0.607 + ], + "angle": 0, + "content": "Haze Removal. We perform image dehazing experiments on SOTS [29] benchmark. We compare FPro with eight representative approaches, including AOD-Net [28], MSCNN [57], DehazeNet [2], EPDN [53], FDGAN [12], AirNet [30], Restormer [81], and PromptIR [50]. As shown in Tab. 4, FPro achieves the best scores among all considered methods. Compared to the recent prompt-based method PromptIR [50], FPro makes a substantial performance gain of \\(1.54\\mathrm{dB}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.608, + 0.788, + 0.728 + ], + "angle": 0, + "content": "Motion Blur Removal. We evaluate image deblurring performance on the GoPro dataset [43]. For synthetic deblurring, we compare FPro with eight representative models: CODE [85], IPT [4], MPRNet [83], MIMO [9], HINet [6], MAXIM [62], Restormer [81] and PromptRestorer [68]. Tab. 5 shows that our FPro achieves competitive performance against recent prompt-based PromptRestorer [68], where we cost half FLOPs. Meanwhile, compared to the recent method CODE [85], FPro obtains a 1.11 dB gain on PSNR while using lower FLOPs." + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.753, + 0.317, + 0.801 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.245, + 0.801, + 0.294, + 0.81 + ], + "angle": 0, + "content": "Raindrop" + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.753, + 0.407, + 0.8 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.338, + 0.801, + 0.386, + 0.81 + ], + "angle": 0, + "content": "Reference" + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.753, + 0.5, + 0.8 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.411, + 0.801, + 0.5, + 0.81 + ], + "angle": 0, + "content": "RaindropAttn [54]" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.753, + 0.594, + 0.8 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.801, + 0.58, + 0.81 + ], + "angle": 0, + "content": "Uformer [72]" + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.753, + 0.686, + 0.8 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.605, + 0.801, + 0.678, + 0.81 + ], + "angle": 0, + "content": "Restormer [81]" + }, + { + "type": "image", + "bbox": [ + 0.69, + 0.753, + 0.779, + 0.8 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.721, + 0.801, + 0.747, + 0.81 + ], + "angle": 0, + "content": "FPro" + }, + { + "type": "image_caption", + "bbox": [ + 0.238, + 0.822, + 0.763, + 0.837 + ], + "angle": 0, + "content": "Fig. 5: Qualitative comparisons on the AGAN-Data [52] for image deraindrop." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.127 + ], + "angle": 0, + "content": "Zhou et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.461, + 0.162 + ], + "angle": 0, + "content": "4.3 Analysis and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.784, + 0.22 + ], + "angle": 0, + "content": "For the ablation studies, we study different models for rain streak removal on SPAD [70] with \\(256 \\times 256\\) patches for 300K iterations. Testing is conducted on SPAD testing dataset [70]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.222, + 0.785, + 0.357 + ], + "angle": 0, + "content": "Effectiveness of Gated Dynamic Decoupler. To demonstrate the effectiveness of the Gated Dynamic Decoupler, we conduct experiments on different model variants in Tab. 6. Compared to the model equipped with Multiple Dynamic Convolution [8] (DC) for separating different frequency parts (b), directly replacing it with GDD (c) results in a performance gain of \\(0.39\\mathrm{dB}\\) in terms of PSNR. Meanwhile, instead of injecting GDD into each DPB (c) to employ multiple decouplers, we attempt to share one GDD module to divide the low-/high frequency information (d), which slightly reduces the complexity \\((0.02\\mathrm{M})\\) of the whole framework and brings a \\(0.08\\mathrm{dB}\\) performance boost." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.359, + 0.785, + 0.524 + ], + "angle": 0, + "content": "Effectiveness of Dual Propmt Block. To investigate the developed DPB, ablation studies are performed in Tab. 7. Disabling HPM or LPM leads to a clear drop of \\(0.22\\mathrm{dB}\\) and \\(0.1\\mathrm{dB}\\), respectively. These experimental results indicate that both HPM and LPM play a positive role in restoring high-quality images. Moreover, we present visualizations to better show the effect of DPB. As shown in Fig. 7, we visualize the generated low-/high-frequency feature maps from each branch along with the analysis in the Fourier domain, where the low-frequency prompt feature encodes information such as structures while the high-frequency prompt one focuses on information such as edges and texture. Meantime, we provide visual comparisons in Fig. 8 to show the effectiveness of the proposed HPM/LPM." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.526, + 0.787, + 0.631 + ], + "angle": 0, + "content": "Perceptual Quality Assessment. To test the perceptual quality of the proposed FPro, following [7], we randomly choose 20 rainy images under real-world scenes from Internet-Data [70] to perform the evaluation. As shown in Tab. 9, compared to other considered methods, FPro achieves a lower NIQE score, which means the generated results contain clearer content and better perceptual quality. Through qualitative comparison in Fig. 6, FPro obtains a visually pleasant result against other models, indicating that it handles unseen degradation well." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.632, + 0.785, + 0.707 + ], + "angle": 0, + "content": "Model Efficiency. We provide the comparison of performance (PSNR), complexity (FLOPs and Parameters), and latency (Run-times) for image deraining. FLOPs and Runtimes are measured when input with the size of \\(256 \\times 256\\), and PSNR scores are tested on SPAD [70]. As shown in Tab. 8, though FPro achieves better performance in terms of PSNR metric, it has less model complexity than" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.735, + 0.315, + 0.786 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.254, + 0.788, + 0.284, + 0.796 + ], + "angle": 0, + "content": "Rainy" + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.736, + 0.411, + 0.787 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.327, + 0.788, + 0.4, + 0.796 + ], + "angle": 0, + "content": "Restormer [81]" + }, + { + "type": "image", + "bbox": [ + 0.414, + 0.736, + 0.506, + 0.787 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.428, + 0.788, + 0.49, + 0.796 + ], + "angle": 0, + "content": "Uformer [72]" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.736, + 0.601, + 0.787 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.532, + 0.788, + 0.575, + 0.796 + ], + "angle": 0, + "content": "IDT [75]" + }, + { + "type": "image", + "bbox": [ + 0.603, + 0.736, + 0.695, + 0.787 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.614, + 0.788, + 0.685, + 0.796 + ], + "angle": 0, + "content": "DRSformer [7]" + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.736, + 0.79, + 0.787 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.731, + 0.788, + 0.757, + 0.796 + ], + "angle": 0, + "content": "FPro" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.808, + 0.784, + 0.836 + ], + "angle": 0, + "content": "Fig.6: Qualitative comparisons with state-of-the-art methods on Internet-Data [70] for real rain removal. (Zoom in for a better view.)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.301, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "A Frequency Prompt Guided Transformer for Image Restoration" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.144, + 0.31, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.144, + 0.401, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.218, + 0.309, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.284, + 0.311, + 0.293 + ], + "angle": 0, + "content": "(a) Low-frequency" + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.219, + 0.4, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.314, + 0.284, + 0.403, + 0.293 + ], + "angle": 0, + "content": "(b) High-frequency" + }, + { + "type": "image", + "bbox": [ + 0.43, + 0.145, + 0.514, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.439, + 0.211, + 0.505, + 0.219 + ], + "angle": 0, + "content": "(a) w/o LPM" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.146, + 0.601, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.538, + 0.211, + 0.577, + 0.219 + ], + "angle": 0, + "content": "(b) Diff." + }, + { + "type": "image", + "bbox": [ + 0.603, + 0.146, + 0.686, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.609, + 0.211, + 0.681, + 0.219 + ], + "angle": 0, + "content": "(c) w/o HPM" + }, + { + "type": "image", + "bbox": [ + 0.689, + 0.146, + 0.773, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.71, + 0.211, + 0.749, + 0.219 + ], + "angle": 0, + "content": "(d) Diff." + }, + { + "type": "image", + "bbox": [ + 0.43, + 0.219, + 0.515, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.443, + 0.284, + 0.502, + 0.292 + ], + "angle": 0, + "content": "(e) w/LPM" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.219, + 0.602, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.539, + 0.284, + 0.577, + 0.292 + ], + "angle": 0, + "content": "(f) Diff." + }, + { + "type": "image", + "bbox": [ + 0.603, + 0.219, + 0.686, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.614, + 0.284, + 0.675, + 0.292 + ], + "angle": 0, + "content": "(g) w/ HPM" + }, + { + "type": "image", + "bbox": [ + 0.689, + 0.219, + 0.773, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.71, + 0.284, + 0.75, + 0.292 + ], + "angle": 0, + "content": "(h) Diff." + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.305, + 0.403, + 0.443 + ], + "angle": 0, + "content": "Fig. 7: Feature analysis. We visualize the features from the LPM branch (a), and the HPM one (b). In the right-bottom, we show the results of the average features over the channel dimension in the Fourier domain. (Zoom in for a better view.)" + }, + { + "type": "image_caption", + "bbox": [ + 0.416, + 0.304, + 0.785, + 0.443 + ], + "angle": 0, + "content": "Fig. 8: Effect of DBP. Columns 1 and 3 show low-pass and high-pass filtered results, while columns 2 and 4 show the difference (Diff.) between processed results with corresponding filtered ground-truth. Compared with (a), FPro w/ LPM (e) performs better in capturing information such as structures, resulting in fewer erroneous predictions (f). Compared with (c), FPro w/ HPM (g) restores clear edges and shapes, which indicates it enjoys the benefits from the high-frequency information prompt. (Zoom in for a better view.)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.495, + 0.785, + 0.527 + ], + "angle": 0, + "content": "Restormer [81] and DRSformer [7]. Compared to other CNN-/Transformer-based approaches, FPro still has a less or comparable model complexity." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.538, + 0.544, + 0.552 + ], + "angle": 0, + "content": "Comparisons with Alternatives to FPro." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.545, + 0.689 + ], + "angle": 0, + "content": "To further demonstrate the superiority of FPro, we compare it with recent prompt-based methods that mine spatial relations as prompts, including PromptIR [50] and PromptRestorer [68]. As shown in Tab. 10, following PromptIR [50], we train and validate FPro on Rain100L [77]. We achieve a substantial gain of 2.16 dB over PromptIR, and a 0.16 dB performance boost against PromptRestore" + }, + { + "type": "table_caption", + "bbox": [ + 0.555, + 0.581, + 0.785, + 0.608 + ], + "angle": 0, + "content": "Table 10: Comparisons with alternatives to FPro." + }, + { + "type": "table", + "bbox": [ + 0.561, + 0.61, + 0.787, + 0.671 + ], + "angle": 0, + "content": "
ModelsParamsFLOPsPSNR
PromptIR [50]35.617337.04
PromptRestorer [68]24.418639.04
FPro22.38239.20
" + }, + { + "type": "table_caption", + "bbox": [ + 0.335, + 0.737, + 0.667, + 0.753 + ], + "angle": 0, + "content": "Table 8: Model efficiency analysis on SPAD [70]." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.764, + 0.79, + 0.837 + ], + "angle": 0, + "content": "
MethodMPRNet [83]SwinIR [33]Uformer-S [72]Restormer [81]IDT [75]DRSformer [7]FPro
FLOPs/G175.8238.043.9174.761.9242.981.9
Parameters/M20.111.520.626.116.433.722.3
Run-times/s0.031.830.120.140.280.080.08
PSNR/dB43.6444.9746.1347.9847.3448.5348.99
" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.128 + ], + "angle": 0, + "content": "Zhou et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.269, + 0.145, + 0.733, + 0.16 + ], + "angle": 0, + "content": "Table 9: Quantitative results (NIQE) for real-world image deraining." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.172, + 0.787, + 0.22 + ], + "angle": 0, + "content": "
Methods|Input Uformer-S [72] Restormer [81] IDT [75] DRSformer [7] FPro
NIQE ↓5.80125.69715.66315.60855.59425.2999
" + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.238, + 0.501, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.346, + 0.352, + 0.4, + 0.364 + ], + "angle": 0, + "content": "(a) Input" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.239, + 0.761, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.606, + 0.352, + 0.658, + 0.363 + ], + "angle": 0, + "content": "(b) FPro" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.375, + 0.787, + 0.404 + ], + "angle": 0, + "content": "Fig. 9: Examples of erroneous reconstruction are shown, where heavy degradation in nighttime real-world scenes leads to typical failures of FPro." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.437, + 0.36, + 0.453 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.483, + 0.788, + 0.62 + ], + "angle": 0, + "content": "In this work, we explore the benefits of prompt learning from a frequency perspective for the task of image restoration. First, when dynamic decoupling the input features with a gating mechanism to select representative elements, we obtain the related frequency components with regard to the specific degradation removal task. Then, we propose modulating the low-/high-frequency signals with separate branches, which concern the intrinsic characteristics of feature maps from different frequency bands. With these modules, our proposed FPro surpasses previous state-of-the-art methods in several image restoration tasks, while performing competitively in terms of computational cost." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.622, + 0.788, + 0.698 + ], + "angle": 0, + "content": "Limitations. There remain many avenues for further improvements. For instance, one could achieve better performance by addressing failure cases shown in Fig. 9, where FPro meets challenges in dealing with heavy degradation in the nighttime real-world scene. Intuitively, collecting a large-scale real-world dataset is a potential direction for improvements." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.734, + 0.405, + 0.753 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.78, + 0.787, + 0.842 + ], + "angle": 0, + "content": "This work was supported by Natural Science Foundation of Tianjin, China (No. 20JCJQJC00020), the National Natural Science Foundation of China (Nos. U22B2049, 62302240), Fundamental Research Funds for the Central Universities, and Supercomputing Center of Nankai University (NKSC)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.301, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "A Frequency Prompt Guided Transformer for Image Restoration" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.179, + 0.786, + 0.262 + ], + "angle": 0, + "content": "1. Brown, T.B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., Ziegler, D.M., Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Chess, B., Clark, J., Berner, C., McCandlish, S., Radford, A., Sutskever, I., Amodei, D.: Language models are few-shot learners. In: NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.264, + 0.786, + 0.29 + ], + "angle": 0, + "content": "2. Cai, B., Xu, X., Jia, K., Qing, C., Tao, D.: Dehazenet: An end-to-end system for single image haze removal. TIP (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.292, + 0.786, + 0.333 + ], + "angle": 0, + "content": "3. Chantas, G., Galatsanos, N.P., Molina, R., Katsaggelos, A.K.: Variational bayesian image restoration with a product of spatially weighted total variation image priors. TIP (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.335, + 0.786, + 0.362 + ], + "angle": 0, + "content": "4. Chen, H., Wang, Y., Guo, T., Xu, C., Deng, Y., Liu, Z., Ma, S., Xu, C., Xu, C., Gao, W.: Pre-trained image processing transformer. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.363, + 0.786, + 0.389 + ], + "angle": 0, + "content": "5. Chen, L., Chu, X., Zhang, X., Sun, J.: Simple baselines for image restoration. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.391, + 0.786, + 0.418 + ], + "angle": 0, + "content": "6. Chen, L., Lu, X., Zhang, J., Chu, X., Chen, C.: Hinet: Half instance normalization network for image restoration. In: CVPR Workshops (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.419, + 0.786, + 0.446 + ], + "angle": 0, + "content": "7. Chen, X., Li, H., Li, M., Pan, J.: Learning a sparse transformer network for effective image deraining. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.447, + 0.786, + 0.474 + ], + "angle": 0, + "content": "8. Chen, Y., Dai, X., Liu, M., Chen, D., Yuan, L., Liu, Z.: Dynamic convolution: Attention over convolution kernels. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.476, + 0.786, + 0.503 + ], + "angle": 0, + "content": "9. Cho, S.J., Ji, S.W., Hong, J.P., Jung, S.W., Ko, S.J.: Rethinking coarse-to-fine approach in single image deblurring. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.504, + 0.786, + 0.531 + ], + "angle": 0, + "content": "0. Deng, X., Dragotti, P.L.: Deep convolutional neural network for multi-modal image restoration and fusion. TPAMI (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.532, + 0.786, + 0.559 + ], + "angle": 0, + "content": "1. Dong, J., Pan, J., Yang, Z., Tang, J.: Multi-scale residual low-pass filter network for image deblurring. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.561, + 0.786, + 0.587 + ], + "angle": 0, + "content": "2. Dong, Y., Liu, Y., Zhang, H., Chen, S., Qiao, Y.: Fd-gan: Generative adversarial networks with fusion-discriminator for single image dehazing. In: AAAI (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.589, + 0.786, + 0.643 + ], + "angle": 0, + "content": "13. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.645, + 0.786, + 0.685 + ], + "angle": 0, + "content": "14. d'Ascoli, S., Touvron, H., Leavitt, M.L., Morcos, A.S., Biroli, G., Sagun, L.: Convit: Improving vision transformers with soft convolutional inductive biases. In: ICML (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.687, + 0.786, + 0.713 + ], + "angle": 0, + "content": "15. Eigen, D., Krishnan, D., Fergus, R.: Restoring an image taken through a window covered with dirt or rain. In: ICCV (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.715, + 0.786, + 0.742 + ], + "angle": 0, + "content": "16. Fu, X., Huang, J., Zeng, D., Huang, Y., Ding, X., Paisley, J.: Removing rain from single images via a deep detail network. In: CVPR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.743, + 0.786, + 0.783 + ], + "angle": 0, + "content": "17. Gan, Y., Bai, Y., Lou, Y., Ma, X., Zhang, R., Shi, N., Luo, L.: Decorate the newcomers: Visual domain prompt for continual test time adaptation. In: AAAI (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.785, + 0.786, + 0.812 + ], + "angle": 0, + "content": "18. Guo, Y., Xiao, X., Chang, Y., Deng, S., Yan, L.: From sky to the ground: A large-scale benchmark and simple baseline towards real rain removal. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.813, + 0.786, + 0.84 + ], + "angle": 0, + "content": "19. He, B., Wang, C., Shi, B., Duan, L.: Mop moiré patterns using mopnet. In: ICCV (2019)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.179, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.127 + ], + "angle": 0, + "content": "Zhou et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.786, + 0.175 + ], + "angle": 0, + "content": "20. He, B., Wang, C., Shi, B., Duan, L.: Fhde\\(^2\\)net: Full high definition demoireing network. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.786, + 0.203 + ], + "angle": 0, + "content": "21. He, K., Sun, J., Tang, X.: Single image haze removal using dark channel prior. TPAMI (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.786, + 0.231 + ], + "angle": 0, + "content": "22. Hendrycks, D., Gimpel, K.: Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.786, + 0.259 + ], + "angle": 0, + "content": "23. Isola, P., Zhu, J.Y., Zhou, T., Efros, A.A.: Image-to-image translation with conditional adversarial networks. In: CVPR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.786, + 0.286 + ], + "angle": 0, + "content": "24. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.786, + 0.314 + ], + "angle": 0, + "content": "25. Jiang, L., Dai, B., Wu, W., Loy, C.C.: Focal frequency loss for image reconstruction and synthesis. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.786, + 0.342 + ], + "angle": 0, + "content": "26. Khattak, M.U., Rasheed, H., Maaz, M., Khan, S., Khan, F.S.: Maple: Multi-modal prompt learning. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.786, + 0.37 + ], + "angle": 0, + "content": "27. Kong, L., Dong, J., Ge, J., Li, M., Pan, J.: Efficient frequency domain-based transformers for high-quality image deblurring. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.786, + 0.397 + ], + "angle": 0, + "content": "28. Li, B., Peng, X., Wang, Z., Xu, J., Feng, D.: Aod-net: All-in-one dehazing network. In: ICCV (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.786, + 0.424 + ], + "angle": 0, + "content": "29. Li, B., Ren, W., Fu, D., Tao, D., Feng, D., Zeng, W., Wang, Z.: Benchmarking single-image dehazing and beyond. TIP (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.786, + 0.452 + ], + "angle": 0, + "content": "30. Li, B., Liu, X., Hu, P., Wu, Z., Lv, J., Peng, X.: All-in-one image restoration for unknown corruption. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.786, + 0.479 + ], + "angle": 0, + "content": "31. Li, Y., Tan, R.T., Guo, X., Lu, J., Brown, M.S.: Rain streak removal using layer priors. In: CVPR (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.48, + 0.786, + 0.508 + ], + "angle": 0, + "content": "32. Li, Z., Lei, Y., Ma, C., Zhang, J., Shan, H.: Prompt-in-prompt learning for universal image restoration. arXiv preprint arXiv:2312.05038 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.786, + 0.536 + ], + "angle": 0, + "content": "33. Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: Swinir: Image restoration using swin transformer. In: ICCV Workshops (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.537, + 0.786, + 0.563 + ], + "angle": 0, + "content": "34. Liu, D., Wen, B., Fan, Y., Loy, C.C., Huang, T.S.: Non-local recurrent network for image restoration. In: NeurIPS (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.786, + 0.577 + ], + "angle": 0, + "content": "35. Liu, J., Yan, M., Zeng, T.: Surface-aware blind image deblurring. TPAMI (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.786, + 0.604 + ], + "angle": 0, + "content": "36. Liu, L., Liu, J., Yuan, S., Slabaugh, G., Leonardis, A., Zhou, W., Tian, Q.: Wavelet-based dual-branch network for image demoiring. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.786, + 0.633 + ], + "angle": 0, + "content": "37. Liu, L., Xie, L., Zhang, X., Yuan, S., Chen, X., Zhou, W., Li, H., Tian, Q.: Tape: Task-agnostic prior embedding for image restoration. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.634, + 0.786, + 0.66 + ], + "angle": 0, + "content": "38. Liu, X., Suganuma, M., Sun, Z., Okatani, T.: Dual residual networks leveraging the potential of paired operations for image restoration. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.786, + 0.688 + ], + "angle": 0, + "content": "39. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.786, + 0.729 + ], + "angle": 0, + "content": "40. Ma, J., Cheng, T., Wang, G., Zhang, Q., Wang, X., Zhang, L.: Prores: Exploring degradation-aware visual prompt for universal image restoration. arXiv preprint arXiv:2306.13653 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.786, + 0.756 + ], + "angle": 0, + "content": "41. Mei, Y., Fan, Y., Zhou, Y.: Image super-resolution with non-local sparse attention. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.786, + 0.785 + ], + "angle": 0, + "content": "42. Mittal, A., Soundararajan, R., Bovik, A.C.: Making a \"completely blind\" image quality analyzer. IEEE SPL (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.786, + 0.813 + ], + "angle": 0, + "content": "43. Nah, S., Hyun Kim, T., Mu Lee, K.: Deep multi-scale convolutional neural network for dynamic scene deblurring. In: CVPR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.814, + 0.786, + 0.84 + ], + "angle": 0, + "content": "44. Narasimhan, S.G., Nayar, S.K.: Contrast restoration of weather degraded images. TPAMI (2003)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.301, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "A Frequency Prompt Guided Transformer for Image Restoration" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "45. Niu, B., Wen, W., Ren, W., Zhang, X., Yang, L., Wang, S., Zhang, K., Cao, X., Shen, H.: Single image super-resolution via a holistic attention network. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.203 + ], + "angle": 0, + "content": "46. Oppenheim, A.: Discrete-time signal processing. Prentice-Hall google schola (1999)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.785, + 0.231 + ], + "angle": 0, + "content": "47. Ozdenizci, O., Legenstein, R.: Restoring vision in adverse weather conditions with patch-based denoising diffusion models. TPAMI (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.259 + ], + "angle": 0, + "content": "48. Pan, X., Zhan, X., Dai, B., Lin, D., Loy, C.C., Luo, P.: Exploiting deep generative prior for versatile image restoration and manipulation. TPAMI (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.26, + 0.709, + 0.272 + ], + "angle": 0, + "content": "49. Park, N., Kim, S.: How do vision transformers work? In: ICLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.299 + ], + "angle": 0, + "content": "50. Potlapalli, V., Zamir, S.W., Khan, S., Khan, F.S.: Prompt: Prompting for all-in-one blind image restoration. In: NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.785, + 0.327 + ], + "angle": 0, + "content": "51. Purohit, K., Suin, M., Rajagopalan, A., Boddeti, V.N.: Spatially-adaptive image restoration using distortion-guided networks. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.355 + ], + "angle": 0, + "content": "52. Qian, R., Tan, R.T., Yang, W., Su, J., Liu, J.: Attentive generative adversarial network for raindrop removal from a single image. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.785, + 0.383 + ], + "angle": 0, + "content": "53. Qu, Y., Chen, Y., Huang, J., Xie, Y.: Enhanced pix2pix dehazing network. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.411 + ], + "angle": 0, + "content": "54. Quan, Y., Deng, S., Chen, Y., Ji, H.: Deep learning for seeing through window with raindrops. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.438 + ], + "angle": 0, + "content": "55. Rabiner, L.R., Gold, B.: Theory and application of digital signal processing. Englewood Cliffs: Prentice-Hall (1975)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.466 + ], + "angle": 0, + "content": "56. Ren, D., Zuo, W., Hu, Q., Zhu, P., Meng, D.: Progressive image deraining networks: A better and simpler baseline. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.493 + ], + "angle": 0, + "content": "57. Ren, W., Liu, S., Zhang, H., Pan, J., Cao, X., Yang, M.H.: Single image dehazing via multi-scale convolutional neural networks. In: ECCV (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.494, + 0.785, + 0.535 + ], + "angle": 0, + "content": "58. Ren, W., Zhang, J., Pan, J., Liu, S., Ren, J.S., Du, J., Cao, X., Yang, M.H.: Deblurring dynamic scenes via spatially varying recurrent neural networks. TPAMI (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.785, + 0.563 + ], + "angle": 0, + "content": "59. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: MICCAI (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.785, + 0.604 + ], + "angle": 0, + "content": "60. Song, X., Zhou, D., Li, W., Dai, Y., Shen, Z., Zhang, L., Li, H.: Tusr-net: Triple unfolding single image dehazing with self-regularization and dual feature to pixel attention. TIP (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.605, + 0.785, + 0.632 + ], + "angle": 0, + "content": "61. Sun, Y., Yu, Y., Wang, W.: Moiré photo restoration using multiresolution convolutional neural networks. TIP (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.66 + ], + "angle": 0, + "content": "62. Tu, Z., Talebi, H., Zhang, H., Yang, F., Milanfar, P., Bovik, A., Li, Y.: Maxim: Multi-axis mlp for image processing. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.66, + 0.785, + 0.688 + ], + "angle": 0, + "content": "63. Valanarasu, J.M.J., Yasarla, R., Patel, V.M.: Transweather: Transformer-based restoration of images degraded by adverse weather conditions. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.785, + 0.715 + ], + "angle": 0, + "content": "64. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.785, + 0.743 + ], + "angle": 0, + "content": "65. Voigtman, E., Winefordner, J.D.: Low-pass filters for signal averaging. Review of Scientific Instruments (1986)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.785, + 0.771 + ], + "angle": 0, + "content": "66. Wang, C., He, B., Wu, S., Wan, R., Shi, B., Duan, L.Y.: Coarse-to-fine disentangling demoiréing framework for recaptured screen images. TPAMI (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.798 + ], + "angle": 0, + "content": "67. Wang, C., Pan, J., Lin, W., Dong, J., Wu, X.M.: Selfpromer: Self-prompt dehazing transformers with depth-consistency. arXiv preprint arXiv:2303.07033 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "68. Wang, C., Pan, J., Wang, W., Dong, J., Wang, M., Ju, Y., Chen, J.: Promptstorer: A prompting image restoration method with degradation perception. In: NeurIPS (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.127 + ], + "angle": 0, + "content": "Zhou et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "69. Wang, H., Xie, Q., Zhao, Q., Meng, D.: A model-driven deep neural network for single image rain removal. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.785, + 0.203 + ], + "angle": 0, + "content": "70. Wang, T., Yang, X., Xu, K., Chen, S., Zhang, Q., Lau, R.W.: Spatial attentive single-image deraining with a high quality real rain dataset. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.785, + 0.244 + ], + "angle": 0, + "content": "71. Wang, W., Xie, E., Li, X., Fan, D.P., Song, K., Liang, D., Lu, T., Luo, P., Shao, L.: Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.785, + 0.272 + ], + "angle": 0, + "content": "72. Wang, Z., Cun, X., Bao, J., Zhou, W., Liu, J., Li, H.: Uformer: A general u-shaped transformer for image restoration. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.3 + ], + "angle": 0, + "content": "73. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. TIP (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.785, + 0.328 + ], + "angle": 0, + "content": "74. Wu, R., Yang, T., Sun, L., Zhang, Z., Li, S., Zhang, L.: Seesr: Towards semantics-aware real-world image super-resolution. arXiv preprint arXiv:2311.16518 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.355 + ], + "angle": 0, + "content": "75. Xiao, J., Fu, X., Liu, A., Wu, F., Zha, Z.J.: Image de-raining transformer. TPAMI (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.785, + 0.383 + ], + "angle": 0, + "content": "76. Xiao, T., Singh, M., Mintun, E., Darrell, T., Dollar, P., Girshick, R.: Early convolutions help transformers see better. In: NeurIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.383, + 0.785, + 0.411 + ], + "angle": 0, + "content": "77. Yang, W., Tan, R.T., Feng, J., Guo, Z., Yan, S., Liu, J.: Joint rain detection and removal from a single image with contextualized deep networks. TPAMI (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.785, + 0.452 + ], + "angle": 0, + "content": "78. Yu, F., Gu, J., Li, Z., Hu, J., Kong, X., Wang, X., He, J., Qiao, Y., Dong, C.: Scaling up to excellence: Practicing model scaling for photo-realistic image restoration in the wild. arXiv preprint arXiv:2401.13627 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.479 + ], + "angle": 0, + "content": "79. Yu, K., Wang, X., Dong, C., Tang, X., Loy, C.C.: Path-restore: Learning network path selection for image restoration. TPAMI (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.48, + 0.785, + 0.508 + ], + "angle": 0, + "content": "80. Yue, H., Mao, Y., Liang, L., Xu, H., Hou, C., Yang, J.: Recaptured screen image demoiring. TCSVT (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.508, + 0.785, + 0.536 + ], + "angle": 0, + "content": "81. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.785, + 0.577 + ], + "angle": 0, + "content": "82. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Learning enriched features for real image restoration and enhancement. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.577, + 0.785, + 0.605 + ], + "angle": 0, + "content": "83. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Multi-stage progressive image restoration. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.605, + 0.785, + 0.633 + ], + "angle": 0, + "content": "84. Zhang, K., Li, Y., Zuo, W., Zhang, L., Van Gool, L., Timofte, R.: Plug-and-play image restoration with deep denoiser prior. TPAMI (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.66 + ], + "angle": 0, + "content": "85. Zhao, H., Gou, Y., Li, B., Peng, D., Lv, J., Peng, X.: Comprehensive and delicate: An efficient transformer for image restoration. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.66, + 0.785, + 0.688 + ], + "angle": 0, + "content": "86. Zheng, B., Yuan, S., Slabaugh, G., Leonardis, A.: Image demoiring with learnable bandpass filters. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.688, + 0.785, + 0.716 + ], + "angle": 0, + "content": "87. Zheng, C., Zhang, Y., Gu, J., Zhang, Y., Kong, L., Yuan, X.: Cross aggregation transformer for image restoration. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.785, + 0.757 + ], + "angle": 0, + "content": "88. Zhou, S., Chen, D., Pan, J., Shi, J., Yang, J.: Adapt or perish: Adaptive sparse transformer with attentive feature refinement for image restoration. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.757, + 0.785, + 0.785 + ], + "angle": 0, + "content": "89. Zou, X., Xiao, F., Yu, Z., Li, Y., Lee, Y.J.: Delving deeper into anti-aliasing in convnets. IJCV (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.785 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Seeing the Unseen_ A Frequency Prompt Guided Transformer for Image Restoration/73dcd519-18ab-4cac-87c9-f1d814570e98_origin.pdf b/2024/Seeing the Unseen_ A Frequency Prompt Guided Transformer for Image Restoration/73dcd519-18ab-4cac-87c9-f1d814570e98_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f8859f88aa91b4d95485e6eff7f61814a64768fc --- /dev/null +++ b/2024/Seeing the Unseen_ A Frequency Prompt Guided Transformer for Image Restoration/73dcd519-18ab-4cac-87c9-f1d814570e98_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d7e803eed3e3f2a662ef64d204941b711eefb4b2b3e665b6d720d684a0be8e85 +size 6167076 diff --git a/2024/Seeing the Unseen_ A Frequency Prompt Guided Transformer for Image Restoration/full.md b/2024/Seeing the Unseen_ A Frequency Prompt Guided Transformer for Image Restoration/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7b02609098ef748cc95f459bc5da7f793ff136ca --- /dev/null +++ b/2024/Seeing the Unseen_ A Frequency Prompt Guided Transformer for Image Restoration/full.md @@ -0,0 +1,439 @@ +# Seeing the Unseen: A Frequency Prompt Guided Transformer for Image Restoration + +Shihao Zhou $^{1,2}$ , Jinshan Pan $^{3}$ , Jinglei Shi $^{1*}$ , Duosheng Chen $^{1}$ , Lishen Qu $^{1}$ , and Jufeng Yang $^{1,2}$ + +1 VCIP & TMCC & DISSec, College of Computer Science, Nankai University +$^{2}$ Nankai International Advanced Research Institute (SHENZHEN·FUTIAN) + +3 School of Computer Science and Engineering, Nanjing University of Science and Technology + +{zhoushihao96,duoshengchen,lishenqu}@mail.nankai.edu.cn, +{jinglei.shi,yangjufeng}@nankai.edu.cn,SDLuran@gmail.com + +Abstract. How to explore useful features from images as prompts to guide the deep image restoration models is an effective way to solve image restoration. In contrast to mining spatial relations within images as prompt, which leads to characteristics of different frequencies being neglected and further remaining subtle or undetectable artifacts in the restored image, we develop a Frequency Prompting image restoration method, dubbed FPro, which can effectively provide prompt components from a frequency perspective to guild the restoration model address these differences. Specifically, we first decompose input features into separate frequency parts via dynamically learned filters, where we introduce a gating mechanism for suppressing the less informative elements within the kernels. To propagate useful frequency information as prompt, we then propose a dual prompt block, consisting of a low-frequency prompt modulator (LPM) and a high-frequency prompt modulator (HPM), to handle signals from different bands respectively. Each modulator contains a generation process to incorporate prompting components into the extracted frequency maps, and a modulation part that modifies the prompt feature with the guidance of the decoder features. Experimental results on several popular datasets have demonstrated the favorable performance of our pipeline against SOTA methods on 5 image restoration tasks, including deraining, deraindrop, demoiréing, deblurring, and dehazing. The source code is available at https://github.com/joshyZhou/FPro. + +Keywords: Image Restoration $\cdot$ Frequency Prompt $\cdot$ Transformer + +# 1 Introduction + +Capturing images in unsatisfactory environments, e.g., rain, haze, usually leads to low-quality ones that accordingly affect the application of downstream tasks in practice. Thus, developing an effective image restoration method to restore clear images from degraded ones is an important task. + +Significant progress has been made due to kinds of the deep learning models [4,5,81], and these deep learning-based approaches become predominant ones as they achieve better performance than the conventional hand-crafted prior-based approaches [3,21,31,44,84]. + +Existing methods, e.g., [33, 72, 81] achieve promising performances in kinds of image restoration tasks. However, these learning-based methods intend to learn a mapping function between degraded images and clear ones, where the characteristics of the specific degradation are less considered. For example, rain streaks tend to obscure the background partially, whereas raindrops typically result in a more pronounced regional occlusion. Accordingly, these models are hindered from generating better results. + +More recently, prompt-learning based methods [50,67,68] serve as an alternative approach to encode useful content of specific degradation for modulating the network, and make a clear performance boost for image restoration. However, we notice that these methods [50,68] pay attention to mining spatial correlations to provide degradation information, whereas the task-specific frequency cues are less studied. Indeed, since various forms of degradation exhibit distinct impacts on image content, they affect information from different frequency bands. Hence, it is crucial to develop an efficient prompt mechanism that explores useful prompts from a frequency perspective for identifying specific characteristics of diverse degradation, which can boost the model to effectively restore images with finer details and non-local structures of the scenes. + +This paper proposes a Frequency Prompting image restoration method, dubbed FPro, to modulate the network by encoding degradation-specific frequency cues as prompts. As mentioned above, existing prompt strategies [50,68] focus on mining spatial relations as useful prompts. In this way, differences between the restored image and the real one within frequency domain [25] are ignored, which remain subtle or undetectable artifacts in the spatial domain. Instead, our FPro aims to enjoy benefits from the capability of prompt learning in different frequency bands at multi-scale resolutions to recover clean images. + +We present two designs to enhance FPro for general image restoration: 1). We first decouple input features into separate low-/high-frequency parts using a gated dynamic decoupler, as signals in different frequency bands encode image patterns from distinct views, i.e., local details and global structures. To this end, a gating mechanism is introduced to help learn the enhanced low-pass filters by suppressing the less informative elements within the kernel, which are then employed to generate low-frequency maps. Meanwhile, the corresponding high-pass filter is obtained by subtracting the low-pass filter from the identity kernel, for generating high-frequency maps. 2). We propose a Dual Prompt Block (DPB), which consists of two modulators, i.e., the Low-frequency Prompt Modulator (LPM) and the High-frequency Prompt Modulator (HPM), to handle low- and high-frequency information respectively. Each modulator includes (a) a generation part that incorporates prompting components into the extracted frequency maps, which is supposed to help distinguish various elements within features, such as rain patterns in the context of deraining; and (b) a modu + +lation part that modifies the prompt feature with the guidance of the feature in the restoration process. In terms of functionality, LPM enhances the low-frequency characteristics through a gating mechanism in the Fourier domain before injecting the prompting components, which is proven equivalent to dynamic large-kernel depth-wise convolution in the spatial domain while computationally efficient, and then encodes low-frequency interactions via global cross-attention. As a complement, HPM applies a locally-enhanced gating mechanism to obtain useful high-frequency signals, and then encodes high-frequency interactions via local cross-attention. + +We summarize our main contributions as follows: + +- We propose FPro, which benefits from prompting learning of frequency components for general image restoration. Instead of mining spatial relations as in previous methods, we explore frequency maps to encode specific degradation information as prompts to guide the image restoration model for restoring finer details and the global structure of the scenes. +- We decouple input features into different frequency bands using learnable low-pass filters, and propose a dual prompt block, which is composed of low-frequency prompt modulator (LPM) and high-frequency prompt modulator (HPM), to explore both details and structures for better restoration. +- Experimental results on 5 restoration tasks: deraining, deraindrop, demoiréing, deblurring and dehazing, show that FPro achieves favorable performance against state-of-the-art methods. + +# 2 Related Work + +Image Restoration. The goal of the image restoration task is to recover high-quality images from the degraded ones. Going beyond conventional prior-based solutions [3, 21], this community has witnessed the great success of a body of learning-based approaches [34, 48, 79]. Despite the promising results obtained by various CNN-based architectures [9, 35, 58], the main concern for methods of this kind is that they pose a limited receptive field problem of the basic convolution operation. This means that the feature map contains less global context (corresponding to low-frequency characteristics in an image), and the final prediction can get stuck in this limitation. This drawback has motivated the increased interest in exploring components to capture desired global cues, like attention mechanisms [10, 45, 60], where better restoration performance can be achieved. For instance, MIRNet [82] proposes a dual attention unit to capture contextual information in dual dimensions. NLSN [41] employs a self-attention mechanism to collect global correlation information for super-resolution. + +Transformer-based Restoration. In recent years, the idea of using Transformer architecture [64] to address various computer vision-related tasks has been popular. Thanks to their discriminative feature representation capability, they not only earn advantages in solving high-level vision tasks [13, 14, 71], but also are extended to low-level image restoration tasks [27, 85, 88]. Unfortunately, as the complexity of vanilla self-attention is quadratic to the image size, + +this mechanism suffers from non-trivial computational costs in handling high-resolution input. To address this, some attempts have been made to explore efficient transformer architectures [4,72,87]. Specifically, SwinIR [33] introduces a window-based self-attention scheme to improve efficiency. Restormer [81] adopts channel-wise self-attention to reduce the computational costs. The majority of these works have offered reliable solutions to recover clean images, however, some works [11,49] realized that the low-pass filter nature of self-attention, which could lose the high-frequency information, such as textures and edges. Even though these models have achieved superior performance, few high-frequency details can be leveraged to implement image restoration, limiting better recovery as a result. Visual Prompt Learning. Prompt learning, which emerged recently in the NLP field [1], has resulted in rapid advancements in its adaptation to computer vision tasks [17,24,26]. Contrary to high-level vision problems, motivated by high effectiveness, some works also consider seeking proper prompts for the low-level pipelines [40, 74, 78]. + +The aim of this paper is not to prompt models for addressing the ALL-in-One problem (in fact, the previous works of [32,40,50] have addressed this nicely by designing various degradation prompt modules). However, our approach is relevant to recent studies [67,68] exploring degradation-specific information for better image restoration results. In contrast to these attempts that generate raw degradation features with a pre-trained model, we prompt the restoration models from a frequency perspective. + +# 3 Proposed Method + +# 3.1 Overall Pipeline + +As depicted in Fig. 1, the overview of our proposed FPro contains the upper restoration branch, like existing works [33,81], and the bottom prompt branch to extract informative frequency maps and then modulate them as prompts. Restoration Branch. Given a degraded image $\mathbf{I} \in \mathbb{R}^{H \times W \times 3}$ as input, FPro first employs a convolution layer to obtain shallow feature $\mathbf{F}_s \in \mathbb{R}^{H \times W \times C}$ ; where $H \times W$ represents the spatial dimension and $C$ is the number of channel. Next, the shallow feature passes through the upper $N_1$ -level encoder-decoder restoration branch to extract deep feature $\mathbf{F}_d \in \mathbb{R}^{H \times W \times C}$ . Early layers in Transformer-based models focus on aggregate local patterns [76], whereas the self-attention module acts as a low-pass filter and tends to dilute high-frequency local details [49]. To alleviate the two contradictory factors, we remove the attention mechanism within the encoder of the restoration branch. Specifically, each level of the encoder includes $N_2$ feed-forward network (FFN) [81] and the paired convolution layer for down-sampling. The encoder features are then fused with the ones in the decoder through skip connections by $1 \times 1$ convolution. For the decoder part, each level is composed of $N_2$ pairs of FFN and multi-head self-attention mechanisms (MSA) [81], along with the convolution layer for up-sampling. Finally, a $3 \times 3$ convolution layer is employed to deep feature $\mathbf{F}_d$ for generating residual image $\mathbf{R} \in \mathbb{R}^{H \times W \times 3}$ . The restored image $\hat{\mathbf{I}}$ is estimated by: $\hat{\mathbf{I}} = \mathbf{I} + \mathbf{R}$ . + +![](images/1f959115b233ea26db4867f087af8516eb152ff14c61502f2a3eedd4b992396a.jpg) +Fig. 1: Overview of the proposed FPro. Except for the common upper restoration branch, which is similar to existing methods [33,81], FPro contains another bottom prompt branch to extract informative features from a frequency perspective. Specifically, the primary components of the prompt branch in this framework are the gated dynamic decoupler (GDD) and dual prompt block (DPB). The GDD is employed to decompose the low-frequency components and corresponding high-frequency characteristics from the input features. Then these frequency-specific features are further processed in DPB, i.e., the high-frequency prompt modulator (HPM) and the low-frequency prompt modulator (LPM), which generates representative frequency prompt to facilitate the clear image reconstruction. + +Prompt Branch. In this branch, we take as input the shallow feature $\mathbf{F}_s$ to generate useful frequency prompts, which are further leveraged to facilitate the latent clear image reconstruction. To achieve this goal, we first decompose the input feature into different frequency bands using a gated dynamic decouple (GDD) (see Section 3.2). After that, low-/high-frequency maps are injected with prompt components to distinguish informative elements according to specific tasks, and then modulated as different prompts (i.e., $\mathbf{F}_{hi}^{out}$ and $\mathbf{F}_{low}^{out}$ ) to interact with the decoder features by $1\times 1$ convolution (see Section 3.3). Next, we present the modules of the prompt branch. + +# 3.2 Gated Dynamic Decoupler + +Each type of degradation affects image content in different ways. For instance, rain streaks partially occlude the background while raindrops often cause much greater obstruction, which corresponds to touch high-/low-frequency bands respectively. To handle these differences, as shown in Fig. 2, we decompose the input features into separate frequency parts based on gated and dynamically learned filters. The key ingredient is to introduce a gating mechanism to help generate the gated learnable low-pass filter and the corresponding high-pass filter, which are then employed to obtain low- and high-frequency maps, respectively. These filters are dynamically learned for each spatial location and channel group to balance computation burden and feature diversity. Specifically, given the input shallow feature map $\mathbf{F}_s\in \mathbb{R}^{H\times W\times C}$ , we firstly predict the low-pass + +![](images/bde03dd66a0f3c102343539db53440dd6486179893e19e00d528aea0a9410f04.jpg) +Fig. 2: Illustrations of the Gated Dynamic Decoupler. + +filter for each feature channel group, which can be formulated as: + +$$ +\hat {\mathbf {F}} _ {s} = \operatorname {C o n v} _ {1 \times 1} (\operatorname {G A P} (\mathbf {F} _ {s})), +$$ + +$$ +\tilde {\mathbf {F}} _ {s} = \hat {\mathbf {F}} _ {s} \odot \phi (\operatorname {C o n v} _ {1 \times 1} (\hat {\mathbf {F}} _ {s})), \tag {1} +$$ + +$$ +\mathbf {F} ^ {l} = \mathrm {S o f t m a x} (\mathcal {B} (\tilde {\mathbf {F}} _ {s})) +$$ + +where $\mathbf{F}^l\in \mathbb{R}^{g\times k^2\times 1\times 1}$ , $g$ is the number of channel groups and $k^2$ corresponds to the kernel size of the learned filter; $\mathrm{GAP}(\cdot)$ and $\mathrm{Conv}_{1\times 1}(\cdot)$ are global average pooling layer and convolution operation with the filter size of $1\times 1$ , respectively; $\phi (\cdot)$ denotes sigmoid activation, $\odot$ refers to the element-wise product, and $\mathcal{B}(\cdot)$ means Batch Normalization. Particularly, $\mathrm{Softmax}(\cdot)$ is a softmax layer, which ensures the generated filters are low-pass [89]. Then, we apply these learned filters to each group input feature $\mathbf{F}_i\in \mathbb{R}^{H\times W\times C_i}$ to obtain low-frequency components: + +$$ +\mathbf {F} _ {i, c, h, w} ^ {l o} = \sum_ {p, q} \mathbf {F} _ {i, p, q} ^ {L} \mathbf {F} _ {i, c, h + p, w + q}, \tag {2} +$$ + +where $\mathbf{F}^L\in \mathbb{R}^{g\times k\times k}$ is the reshaped filter, $i$ denotes the group index, $C_i = \frac{C}{g}$ refers to number of the group channel, $c$ means the index of a channel, $h$ and $w$ are spatial coordinates, $p,q\in \{-1,0,1\}$ point to the surrounding locations. + +Meanwhile, we invert this process by subtracting the low-pass filter from the identity kernel to attain the high-pass filter, which is employed to generate the corresponding high-frequency components $\mathbf{F}_{hi}$ . + +# 3.3 Dual Prompt Block + +Considering that the extracted features, i.e., low-/high-frequency maps, encode image patterns from distinct views (local detail and main structure of the image). We design the Dual Prompt Block that includes two components, i.e., High-frequency Prompt Modulator (HPM) and Low-frequency Prompt Modulator (LPM), to deal with these feature maps, respectively. + +High-frequency Prompt Modulator. Given the two input feature maps, including the $l$ -level feature $\mathbf{F}_l \in \mathbb{R}^{\hat{H} \times \hat{W} \times \hat{C}}$ and high-frequency feature $\mathbf{F}_{hi} \in$ + +![](images/8304b4ce1b3aa19dc61980f1ddb181234139e82114cd6e96adf9522a75660b31.jpg) + +![](images/922e68f04880e8da8fae2aa4e54cb778321805861a6af23fc0d475e194619642.jpg) +Fig. 3: Illustrations of the proposed components. (a) High-frequency Prompt Modulator (HPM); (b) Low-frequency Prompt Modulator (LPM). + +$\mathbb{R}^{H\times W\times C'}$ , we first resize $\mathbf{F}_{hi}$ and obtain $\tilde{\mathbf{F}}_{hi}\in \mathbb{R}^{\hat{H}\times \hat{W}\times C'}$ . Towards highlighting high-frequency characteristics, we employ a gating mechanism to adaptively determine the useful frequency information: + +$$ +\hat {\mathbf {F}} _ {h i} = \tilde {\mathbf {F}} _ {h i} \odot \sigma (\mathrm {D C o n v} _ {3 \times 3} (\tilde {\mathbf {F}} _ {h i})), \tag {3} +$$ + +where $\hat{\mathbf{F}}_{hi} \in \mathbb{R}^{\hat{H} \times \hat{W} \times C'}$ is the processed feature, $\mathrm{DConv}_{3 \times 3}(\cdot)$ denotes a depth-wise convolution operation with the filter size of $3 \times 3$ , and $\sigma(\cdot)$ is the GELU activation function [22]. Then, we leverage the learnable high-frequency prompt components $\mathbf{P}_{hi} \in \mathbb{R}^{\hat{H} \times \hat{W} \times C'}$ to make adjustments to the input features, which aims to help distinguish various elements, such as rain patterns and streaks of different orientations and magnitudes in the context of deraining: + +$$ +\mathbf {F} _ {h i} ^ {\text {p r o m p t}} = \hat {\mathbf {F}} _ {h i} \odot \mathbf {P} _ {h i}, \tag {4} +$$ + +where $\mathbf{F}_{hi}^{prompt} \in \mathbb{R}^{\hat{H} \times \hat{W} \times C'}$ is the obtained high-frequency feature prompt. + +Next, we modify the high-frequency prompt $\mathbf{F}_{hi}^{prompt}$ according to the input feature $\mathbf{F}_l$ . To be specific, we utilize a depth-wise convolution operator, which acts as a high-pass filter [49], to enhance the high-frequency sources in the input $\mathbf{F}_{hi}^{prompt}$ . Then, we generate query $(\mathbf{Q}_{hi})$ projection from $\mathbf{F}_l$ , key $(\mathbf{K}_{hi})$ and value $(\mathbf{V}_{hi})$ projections from the processed feature map $\hat{\mathbf{F}}_{hi}^{prompt} = \mathrm{DConv}_{3\times 3}(\mathbf{F}_{hi}^{prompt})$ , respectively. Meanwhile, as the high-frequency information usually corresponds to image details and is a local feature, it could be redundant to calculate global attention. Therefore, before leveraging the linear layer to obtain the matrices of $\mathbf{Q}_{hi}$ , $\mathbf{K}_{hi}$ , and $\mathbf{V}_{hi}$ , the local window self-attention mechanism is adopted to save computational complexity and capture fine-grained high + +frequencies, which yields $\mathbf{Q}_{hi} = W_{p}^{Q_{hi}}\cdot R(\mathbf{F}_{l})$ , $\mathbf{K}_{hi} = W_{p}^{K_{hi}}\cdot R(\hat{\mathbf{F}}_{hi}^{prompt})$ , $\mathbf{V}_{hi} = W_{p}^{V_{hi}}\cdot R(\hat{\mathbf{F}}_{hi}^{prompt})$ . Where $W_{p}^{(\cdot)}$ represents the projection matrices, and $R(\cdot)$ denotes the window partition strategy [39]. Generally, we have $\mathbf{Q}_{hi}\in \mathbb{R}^{\frac{\hat{H}\hat{W}}{M^2}\times M^2\times\hat{C}}$ , $\mathbf{K}_{hi}\in \mathbb{R}^{\frac{\hat{H}\hat{W}}{M^2}\times \hat{C}\times M^2}$ , and $\mathbf{V}_{hi}\in \mathbb{R}^{\frac{\hat{H}\hat{W}}{M^2}\times M^2\times\hat{C}}$ , where $M^2$ is the size of split windows. The attention matrix is thus calculated to tune the high-frequency prompt: + +$$ +\mathbf {F} _ {h i} ^ {o u t} = \mathbf {V} _ {h i} \cdot \mathrm {S o f t m a x} (\mathbf {K} _ {h i} \cdot \mathbf {Q} _ {h i} / \sqrt {d}), \tag {5} +$$ + +where $\mathbf{F}_{hi}^{out}\in \mathbb{R}^{\hat{H}\times \hat{W}\times \hat{C}}$ is the output feature map of the high-frequency prompt modulation branch; $d$ is the query/ key dimension, following [33]. + +Low-frequency Prompt Modulator. Given the two input feature maps, including the $l$ -level feature $\mathbf{F}_l \in \mathbb{R}^{\hat{H} \times \hat{W} \times \hat{C}}$ and low-frequency feature $\mathbf{F}_{lo} \in \mathbb{R}^{H \times W \times C'}$ , we first resize $\mathbf{F}_{lo}$ and obtain $\tilde{\mathbf{F}}_{lo} \in \mathbb{R}^{\hat{H} \times \hat{W} \times C'}$ . Towards handling low-frequency signals effectively, we project $\tilde{\mathbf{F}}_{lo}$ into the frequency domain via the fast Fourier transform (FFT). Then, a gating mechanism is adopted to control the useful low-frequency components flow forward: + +$$ +\hat {\mathbf {F}} _ {l o} = \mathcal {F} (\tilde {\mathbf {F}} _ {l o}) \odot \sigma (\operatorname {C o n v} _ {1 \times 1} (\mathcal {F} (\tilde {\mathbf {F}} _ {l o}))), \tag {6} +$$ + +where $\hat{\mathbf{F}}_{lo}\in \mathbb{R}^{\hat{H}\times (\frac{\hat{W}}{2} +1)\times 2C^{\prime}}$ is the processed feature, $\mathcal{F}(\cdot)$ represents the FFT. Next, we calibrate the input features by injecting learnable low-frequency prompt components $\mathbf{P}_{lo}\in \mathbb{R}^{\hat{H}\times (\frac{\hat{W}}{2} +1)\times 2C^{\prime}}$ , which is then transformed to spatial domain: + +$$ +\mathbf {F} _ {l o} ^ {\text {p r o m p t}} = \mathcal {F} ^ {- 1} \left(\hat {\mathbf {F}} _ {l o} \odot \mathbf {P} _ {l o}\right), \tag {7} +$$ + +where $\mathbf{F}_{lo}^{prompt} \in \mathbb{R}^{\hat{H} \times \hat{W} \times C'}$ is the generated low-frequency feature prompt, and $\mathcal{F}^{-1}(\cdot)$ denotes the inverse FFT. + +Noted, we perform the feature transformation in the Fourier domain for efficient global information interaction. The convolution theorem [46, 55] indicates the Hadamard product of two signals in the Fourier domain equals to implement the Fourier transform of a convolution of these two signals in the original spatial domain. Base on this insight, we can combine Eq.(6) and Eq.(7): + +$$ +\begin{array}{l} \mathbf {F} _ {l o} ^ {\text {p r o m p t}} = \mathcal {F} ^ {- 1} \left(\mathcal {F} (\tilde {\mathbf {F}} _ {l o}) \odot \sigma \left(\operatorname {C o n v} _ {1 \times 1} \left(\mathcal {F} (\tilde {\mathbf {F}} _ {l o})\right)\right) \odot \mathbf {P} _ {l o}\right) \\ = \mathcal {F} ^ {- 1} \left(\mathcal {F} \left(\tilde {\mathbf {F}} _ {l o} * \mathcal {F} ^ {- 1} \left(\sigma \left(\operatorname {C o n v} _ {1 \times 1} \left(\mathcal {F} \left(\tilde {\mathbf {F}} _ {l o}\right)\right)\right) \odot \mathbf {P} _ {l o}\right)\right)\right) \tag {8} \\ = \tilde {\mathbf {F}} _ {l o} * \mathcal {F} ^ {- 1} \left(\sigma \left(\operatorname {C o n v} _ {1 \times 1} \left(\mathcal {F} (\tilde {\mathbf {F}} _ {l o})\right)\right) \odot \mathbf {P} _ {l o}\right) \\ \end{array} +$$ + +where $\circledast$ is the convolution operation. Since $\mathcal{F}^{-1}(\sigma (\mathrm{Conv}_{1\times 1}(\mathcal{F}(\tilde{\mathbf{F}}_{lo})))\odot \mathbf{P}_{lo})$ is a tensor that shares the same shape with $\tilde{\mathbf{F}}_{lo}$ , it can be served as a dynamic depth-wise convolution kernel as large as $\tilde{\mathbf{F}}_{lo}$ in spatial domain while introducing less model complexity. + +Subsequently, we further modulate the low-frequency visual prompt $\mathbf{F}_{lo}^{prompt}$ with the guidance of the input feature $\mathbf{F}_l$ . Specifically, we adopt an adaptive average pooling operator, which serves as a low-pass filter [65], to enhance the + +Table 1: Quantitative comparison on the SPAD dataset [70] for image deraining. + +
MethodDDN [16]PReNet [56]RCDNet [69]MPRNet [83]SPAIR [51]Uformer-S [72]SCD-Former [18]IDT [75]Restormer [81]DRSformer [7]FPro (Ours)
PSNR ↑36.1640.1643.3643.6444.1046.1346.8947.3447.9848.5348.99
SSIM ↑0.94630.98160.98310.98440.98720.99130.99410.99290.99210.99240.9936
+ +Table 2: Quantitative comparison on the AGAN-Data [52] for image deraindrop. + +
MethodEigen's [15]Pix2pix [23]Uformer-S [72]WeatherDiff128 [47]TransWeather [63]DuRN [38]RaindropAttn [54]AttentiveGAN [52]IDT [75]Restormer [81]FPro (Ours)
PSNR ↑21.3128.0229.4229.6630.1731.2431.3731.5931.6331.6831.96
SSIM ↑0.7570.8550.9060.9230.9160.9260.9180.9170.9360.9340.937
+ +low-frequency content in the input $\mathbf{F}_{lo}^{prompt}$ . After that, we generate query $(\mathbf{Q}_{lo})$ projection from reshaped $\mathbf{F}_l$ , key $(\mathbf{K}_{lo})$ and value $(\mathbf{V}_{lo})$ projections from the average-pooled feature $\hat{\mathbf{F}}_{lo}^{prompt} = \mathrm{AAP}(\mathbf{F}_{lo}^{prompt})$ , respectively. Here, $\mathrm{AAP}(\cdot)$ means the adaptive average pooling operation. Then, we employ $1\times 1$ convolutions to yield $\mathbf{Q}_{lo} = W_p^{Q_{lo}}\mathbf{F}_l$ , $\mathbf{K}_{lo} = W_p^{K_{lo}}\hat{\mathbf{F}}_{lo}^{prompt}$ , and $\mathbf{V}_{lo} = W_p^{V_{lo}}\hat{\mathbf{F}}_{lo}^{prompt}$ . Here $W_{p}^{(\cdot)}$ is the $1\times 1$ convolution. Next, we calculate the dot-product of query and key projections, which generates the transposed-attention $\mathbf{A}\in \mathbb{R}^{\hat{H}\hat{W}\times 1}$ . Overall, the process of modulating the low-frequency prompt is defined as: + +$$ +\mathbf {F} _ {l o} ^ {o u t} = \mathbf {V} _ {l o} \cdot \operatorname {S o f t m a x} (\mathbf {K} _ {l o} \cdot \mathbf {Q} _ {l o} / \alpha), \tag {9} +$$ + +where $\mathbf{F}_{lo}^{out} \in \mathbb{R}^{\hat{H} \times \hat{W} \times C'}$ is the output feature map of the low-frequency prompt modulation branch; $\mathbf{Q}_{lo} \in \mathbb{R}^{\hat{H}\hat{W} \times C'}$ , $\mathbf{K}_{lo} \in \mathbb{R}^{C' \times 1}$ , and $\mathbf{V}_{lo} \in \mathbb{R}^{1 \times C'}$ are the input matrices; $\alpha$ is the learnable scaling parameter. + +For both low/high-frequency modulators, we perform the attention map calculation several times in parallel, and these results are then concatenated for multi-head self-attention (MSA) [64]. + +# 4 Experiments + +# 4.1 Experimental settings + +Metrics. We adopt commonly used peak signal-to-noise ratio (PSNR) [73] and structural similarity (SSIM) metrics to evaluate restored images. Meanwhile, perceptual metric NIQE [42] is employed as a non-reference metric. Following previous works [69, 72], PSNR/SSIM computations are implemented on the Y channel in the YCbCr space for the image deraining task, while calculated in the RGB color space for other restoration tasks. In the reported tables, the best and second-best scores are highlighted and underlined, respectively. + +Implementation Details. FPro contains $N_{1} = 3$ levels encoder-decoder, where the encode and decoder share the same $N_{2} = [2,3,6]$ blocks. We set embedding dimensions $C$ as 48, and the attention heads as [2,4,8]. The expanding channel + +Table 3: Quantitative comparison on the TIP-2018 [61] for image demoiré. + +
MethodAMNet [80]DMCNN [61]UNet [59]WDNet [36]MopNet [19]TAPE-Net [37]FHD2eNet [20]MBCNN [86]Uformer-S [72]Wang et al. [66]FPro (Ours)
PSNR ↑25.4726.1026.4927.1227.4827.5227.7928.4028.6328.8729.25
SSIM ↑0.8330.8440.8640.8540.8610.8660.8670.8710.8720.8940.879
+ +Table 4: Quantitative comparison on the SOTS [29] for image dehazing. + +
MethodAODNet [28]MSCNN [57]DehazeNet [2]EPDN [53]FDGAN [12]AirNet [30]Restormer [81]PromptIR [50]FPro (Ours)
PSNR ↑20.2922.0622.4622.5723.1523.1830.8731.3132.85
SSIM ↑0.8770.9080.8510.8630.9210.9000.9690.9730.983
+ +capacity factor in FFN is 3. The default split window size in HPM is set as $M = 8$ . The pixel-unshuffle and pixel-shuffle are employed for downsampling and upsampling. We use the AdamW optimizer with an initial learning rate of $3 \times 10^{-4}$ , which is gradually reduced to $1 \times 10^{-6}$ using cosine annealing, to train FPro. Additionally, we adopt the widely used loss function [68] to constrain the network training. + +# 4.2 Main Results + +Rain Streak Removal. The proposed FPro is compared with general restoration approaches [51,72,81,83] as well as with task-specific methods [7,16,18,56,69, 75]. Tab. 1 shows that FPro makes superior performance over current methods for real image deraining on SPAD [70]. Compared to the previous best approach DRSformer [7], FPro achieves a $0.46\mathrm{dB}$ performance boost. In addition, FPro obtains 2.1 dB PSNR improvement when compared to the recent model SCD-Former [18]. Fig. 4 provides a visual deraining example, where FPro successfully removes the rain degradation while maintaining the structural information. + +Raindrop Removal. For image deraindrop, we compare FPro with some existing deraindrop methods, including Eigen's [15], Pix2pix [23], TransWeather [63], Uformer [72], WeatherDiff128 [47], DuRN [38], RaindropAttn [54], AttentiveGAN [52], IDT [75], and Restormer [81]. We report the quantitative results on AGAN-Data [52] in Tab. 2. Our FPro obtains the best performance against all considered methods in terms of both PSNR/SSIM scores. FPro makes a 0.28 dB PSNR performance boost over Restormer [81], and 2.3 dB over the recent + +![](images/e137bb23d466f951816bff5f6533476ebe6652864beee47d56c987eeca1cc09d.jpg) +Fig. 4: Qualitative comparisons on the SPAD [70] for real image deraining. + +![](images/67ecf5d624dc5ba28359dbef46e6a72737710bcb3768954392dd474b3db599ba.jpg) + +![](images/369d8fd845dcb1c32152af2d7f33a40a62da76979e6540fe4579b8f1a7d79da6.jpg) + +![](images/483381e9ea8a98f27a7551e3a41af775f8a0abb22be36be18a631065219c1e9c.jpg) + +![](images/8d1163c8fd5fe63bbf97b9f2413a0b181c40c42e5dc6cd0675f38b3f0f325b36.jpg) + +![](images/907a70c72d658cd9a11cb62750eb0f88e8575bb447bae85dfacafe02945462a8.jpg) + +Table 5: Quantitative comparison on the GoPro [43] for image deblurring. + +
MethodCODE [85]IPT [4]MPRNet [83]MIMO [9]HINet [6]MAXIM [62]Restormer [81]PromptRestorer [68]FPro (Ours)
PSNR ↑31.9432.5832.6632.6832.7132.8632.9233.0633.05
SSIM ↑0.928-0.9590.9590.9590.9610.9610.9620.961
+ +Table 6: Effectiveness of GDD. + +
ModelsPSNRSSIM
(a) w/o prompt branch48.400.9920
(b) Multi DC [8]48.520.9926
(c) Multi GDD48.910.9934
(d) Single GDD48.990.9936
+ +Table 7: Ablation study of DPB. + +
ModelsPSNRSSIM
(a) w/o prompt branch48.400.9920
(b) w/o HPM48.770.9931
(c) w/o LPM48.890.9933
(d) Full48.990.9936
+ +method WeatherDiff $_{128}$ [47]. Fig. 5 shows the visual comparisons, where FPro generates a result with finer details. + +Moiré pattern Removal. We perform image demoiring experiments on TIP-2018 [61] dataset, and compare FPro with ten demoiring methods, including AMNet [80], DMCNN [61], UNet [59], WDNet [36], MopNet [19], TAPE-Net [37], $\mathrm{FHD}^2\mathrm{eNet}$ [20], MBCNN [86], Uformer-S [72], and Wang et al. [66]. In Tab. 3, FPro yields a 0.38 performance boost against the previous best method Wang et al. [66], and outperforms the recent model TAPE-Net [37] by 1.73 dB. + +Haze Removal. We perform image dehazing experiments on SOTS [29] benchmark. We compare FPro with eight representative approaches, including AOD-Net [28], MSCNN [57], DehazeNet [2], EPDN [53], FDGAN [12], AirNet [30], Restormer [81], and PromptIR [50]. As shown in Tab. 4, FPro achieves the best scores among all considered methods. Compared to the recent prompt-based method PromptIR [50], FPro makes a substantial performance gain of $1.54\mathrm{dB}$ . + +Motion Blur Removal. We evaluate image deblurring performance on the GoPro dataset [43]. For synthetic deblurring, we compare FPro with eight representative models: CODE [85], IPT [4], MPRNet [83], MIMO [9], HINet [6], MAXIM [62], Restormer [81] and PromptRestorer [68]. Tab. 5 shows that our FPro achieves competitive performance against recent prompt-based PromptRestorer [68], where we cost half FLOPs. Meanwhile, compared to the recent method CODE [85], FPro obtains a 1.11 dB gain on PSNR while using lower FLOPs. + +![](images/40a4d6f437ca090cac087e76357ab59c2495a4b0ef498302a190cabb0afe0d5d.jpg) +Raindrop +Fig. 5: Qualitative comparisons on the AGAN-Data [52] for image deraindrop. + +![](images/a90c5281145716aba555285ff88823fbc5f98eb2a0ed910b8fc97ea83014f18f.jpg) +Reference + +![](images/33f3ac49e2f3876c24abd9196ee934784fecc7cf6e7a5aaa0014ab908c824ff8.jpg) +RaindropAttn [54] + +![](images/9a84dd51f770445bd79e7c20fbd51d31af229295d548a65e15cf3d1657f25ae5.jpg) +Uformer [72] + +![](images/6e2a4fab8a559a41699158da474d935fa5baea64089aee7ba4222fec0e4e972b.jpg) +Restormer [81] + +![](images/0e3a319dcef59ea91c1fb62d49003746c19193733ef0cb88b1ee53b840a9a2c5.jpg) +FPro + +# 4.3 Analysis and Discussion + +For the ablation studies, we study different models for rain streak removal on SPAD [70] with $256 \times 256$ patches for 300K iterations. Testing is conducted on SPAD testing dataset [70]. + +Effectiveness of Gated Dynamic Decoupler. To demonstrate the effectiveness of the Gated Dynamic Decoupler, we conduct experiments on different model variants in Tab. 6. Compared to the model equipped with Multiple Dynamic Convolution [8] (DC) for separating different frequency parts (b), directly replacing it with GDD (c) results in a performance gain of $0.39\mathrm{dB}$ in terms of PSNR. Meanwhile, instead of injecting GDD into each DPB (c) to employ multiple decouplers, we attempt to share one GDD module to divide the low-/high frequency information (d), which slightly reduces the complexity $(0.02\mathrm{M})$ of the whole framework and brings a $0.08\mathrm{dB}$ performance boost. + +Effectiveness of Dual Propmt Block. To investigate the developed DPB, ablation studies are performed in Tab. 7. Disabling HPM or LPM leads to a clear drop of $0.22\mathrm{dB}$ and $0.1\mathrm{dB}$ , respectively. These experimental results indicate that both HPM and LPM play a positive role in restoring high-quality images. Moreover, we present visualizations to better show the effect of DPB. As shown in Fig. 7, we visualize the generated low-/high-frequency feature maps from each branch along with the analysis in the Fourier domain, where the low-frequency prompt feature encodes information such as structures while the high-frequency prompt one focuses on information such as edges and texture. Meantime, we provide visual comparisons in Fig. 8 to show the effectiveness of the proposed HPM/LPM. + +Perceptual Quality Assessment. To test the perceptual quality of the proposed FPro, following [7], we randomly choose 20 rainy images under real-world scenes from Internet-Data [70] to perform the evaluation. As shown in Tab. 9, compared to other considered methods, FPro achieves a lower NIQE score, which means the generated results contain clearer content and better perceptual quality. Through qualitative comparison in Fig. 6, FPro obtains a visually pleasant result against other models, indicating that it handles unseen degradation well. + +Model Efficiency. We provide the comparison of performance (PSNR), complexity (FLOPs and Parameters), and latency (Run-times) for image deraining. FLOPs and Runtimes are measured when input with the size of $256 \times 256$ , and PSNR scores are tested on SPAD [70]. As shown in Tab. 8, though FPro achieves better performance in terms of PSNR metric, it has less model complexity than + +![](images/622a81c1edf129dc614b540540abb783051853d29afe715c83c1d61d78cb430b.jpg) +Rainy + +![](images/8c08f6bd07e0097d3f98e6939644e6ac2d07d80b2ffa0185fc9ede4c8b6c71af.jpg) +Restormer [81] +Fig.6: Qualitative comparisons with state-of-the-art methods on Internet-Data [70] for real rain removal. (Zoom in for a better view.) + +![](images/a6831f37ff6f2669140cf55022036189844a67d5b00590ae30b3eec40bb333d9.jpg) +Uformer [72] + +![](images/d53d1fe432d337afdbae48eb39480b493db010fea9c8f69b48ce6a179c8c5b6e.jpg) +IDT [75] + +![](images/b158043eeccd22d548cb11f36f0adfa8091e44abbe93557bbbdb84de76dd7d25.jpg) +DRSformer [7] + +![](images/a7dfed3a940587bed00847112344ffa4e57d3b510543ac28aa2e58dccc9184e7.jpg) +FPro + +![](images/57af3348afa766a14815e89fc8b9036059a92527ad826c2bb86fc75c11e4c203.jpg) + +![](images/4f8ba90a5cc279e9f070880d31ab5c65b9b352eba6311dd87553d0ac737e4126.jpg) + +![](images/de544173de0447c58b119a3f87ce1ae4c5a24e4f0a8ae7bfe3263e4ac76fe934.jpg) +(a) Low-frequency +Fig. 7: Feature analysis. We visualize the features from the LPM branch (a), and the HPM one (b). In the right-bottom, we show the results of the average features over the channel dimension in the Fourier domain. (Zoom in for a better view.) + +![](images/34b08a68a9c0f83fc545ebf8119f603bca049f0c2a9551d81f0151e411e8392c.jpg) +(b) High-frequency + +![](images/277c1ccfb0ce356cd9d8e8b066edd0183de4b28225b9fac6cd2208bfed7e9bc2.jpg) + +![](images/5436f3a955f6ebb47d0473590704236ec63268e5ef087bb88242c63610295d7b.jpg) + +![](images/d793089ce4bc8969d46fe3be9242a946ad5fb13c142acef0bb52d34b958e4086.jpg) + +![](images/1b268fdb2ef23b77f007d88c9bc908dbf45bea54b68fd6da089976e86818d1c4.jpg) + +![](images/4fa86cc22f5a89a9583bdb53ebd09039774265983adf3bbec43d06b54f39f78a.jpg) +(a) w/o LPM +(e) w/LPM +Fig. 8: Effect of DBP. Columns 1 and 3 show low-pass and high-pass filtered results, while columns 2 and 4 show the difference (Diff.) between processed results with corresponding filtered ground-truth. Compared with (a), FPro w/ LPM (e) performs better in capturing information such as structures, resulting in fewer erroneous predictions (f). Compared with (c), FPro w/ HPM (g) restores clear edges and shapes, which indicates it enjoys the benefits from the high-frequency information prompt. (Zoom in for a better view.) + +![](images/7802302ae02ba88c4b3b86499871b0cc829366fa220a0bb25e81f93f43072b1d.jpg) +(b) Diff. +(f) Diff. + +![](images/979bdc46fd9a2a949b09d9d5fb811cc58e9517113ea47f9e4071e51b82a7378c.jpg) +(c) w/o HPM +(g) w/ HPM + +![](images/51a0aa5820758f9790116a3335d368233db2b35330b57d05dd8041426464facd.jpg) +(d) Diff. +(h) Diff. + +Restormer [81] and DRSformer [7]. Compared to other CNN-/Transformer-based approaches, FPro still has a less or comparable model complexity. + +# Comparisons with Alternatives to FPro. + +To further demonstrate the superiority of FPro, we compare it with recent prompt-based methods that mine spatial relations as prompts, including PromptIR [50] and PromptRestorer [68]. As shown in Tab. 10, following PromptIR [50], we train and validate FPro on Rain100L [77]. We achieve a substantial gain of 2.16 dB over PromptIR, and a 0.16 dB performance boost against PromptRestore + +Table 10: Comparisons with alternatives to FPro. + +
ModelsParamsFLOPsPSNR
PromptIR [50]35.617337.04
PromptRestorer [68]24.418639.04
FPro22.38239.20
+ +Table 8: Model efficiency analysis on SPAD [70]. + +
MethodMPRNet [83]SwinIR [33]Uformer-S [72]Restormer [81]IDT [75]DRSformer [7]FPro
FLOPs/G175.8238.043.9174.761.9242.981.9
Parameters/M20.111.520.626.116.433.722.3
Run-times/s0.031.830.120.140.280.080.08
PSNR/dB43.6444.9746.1347.9847.3448.5348.99
+ +Table 9: Quantitative results (NIQE) for real-world image deraining. + +
Methods|Input Uformer-S [72] Restormer [81] IDT [75] DRSformer [7] FPro
NIQE ↓5.80125.69715.66315.60855.59425.2999
+ +![](images/9cc4a6674949be79df7209c8c806c8782f88704b18e459bda2c9125addbf31f4.jpg) +(a) Input + +![](images/c05e9955531fb6c91afd56c7ccacd582731c6cab10d72208127accbb082de0e9.jpg) +(b) FPro +Fig. 9: Examples of erroneous reconstruction are shown, where heavy degradation in nighttime real-world scenes leads to typical failures of FPro. + +# 5 Conclusion + +In this work, we explore the benefits of prompt learning from a frequency perspective for the task of image restoration. First, when dynamic decoupling the input features with a gating mechanism to select representative elements, we obtain the related frequency components with regard to the specific degradation removal task. Then, we propose modulating the low-/high-frequency signals with separate branches, which concern the intrinsic characteristics of feature maps from different frequency bands. With these modules, our proposed FPro surpasses previous state-of-the-art methods in several image restoration tasks, while performing competitively in terms of computational cost. + +Limitations. There remain many avenues for further improvements. For instance, one could achieve better performance by addressing failure cases shown in Fig. 9, where FPro meets challenges in dealing with heavy degradation in the nighttime real-world scene. Intuitively, collecting a large-scale real-world dataset is a potential direction for improvements. + +# Acknowledgements + +This work was supported by Natural Science Foundation of Tianjin, China (No. 20JCJQJC00020), the National Natural Science Foundation of China (Nos. U22B2049, 62302240), Fundamental Research Funds for the Central Universities, and Supercomputing Center of Nankai University (NKSC). + +# References + +1. Brown, T.B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., Ziegler, D.M., Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Chess, B., Clark, J., Berner, C., McCandlish, S., Radford, A., Sutskever, I., Amodei, D.: Language models are few-shot learners. In: NeurIPS (2020) +2. Cai, B., Xu, X., Jia, K., Qing, C., Tao, D.: Dehazenet: An end-to-end system for single image haze removal. TIP (2016) +3. Chantas, G., Galatsanos, N.P., Molina, R., Katsaggelos, A.K.: Variational bayesian image restoration with a product of spatially weighted total variation image priors. TIP (2009) +4. Chen, H., Wang, Y., Guo, T., Xu, C., Deng, Y., Liu, Z., Ma, S., Xu, C., Xu, C., Gao, W.: Pre-trained image processing transformer. In: CVPR (2021) +5. Chen, L., Chu, X., Zhang, X., Sun, J.: Simple baselines for image restoration. In: ECCV (2022) +6. Chen, L., Lu, X., Zhang, J., Chu, X., Chen, C.: Hinet: Half instance normalization network for image restoration. In: CVPR Workshops (2021) +7. Chen, X., Li, H., Li, M., Pan, J.: Learning a sparse transformer network for effective image deraining. In: CVPR (2023) +8. Chen, Y., Dai, X., Liu, M., Chen, D., Yuan, L., Liu, Z.: Dynamic convolution: Attention over convolution kernels. In: CVPR (2020) +9. Cho, S.J., Ji, S.W., Hong, J.P., Jung, S.W., Ko, S.J.: Rethinking coarse-to-fine approach in single image deblurring. In: ICCV (2021) +0. Deng, X., Dragotti, P.L.: Deep convolutional neural network for multi-modal image restoration and fusion. TPAMI (2021) +1. Dong, J., Pan, J., Yang, Z., Tang, J.: Multi-scale residual low-pass filter network for image deblurring. In: ICCV (2023) +2. Dong, Y., Liu, Y., Zhang, H., Chen, S., Qiao, Y.: Fd-gan: Generative adversarial networks with fusion-discriminator for single image dehazing. In: AAAI (2020) +13. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2021) +14. d'Ascoli, S., Touvron, H., Leavitt, M.L., Morcos, A.S., Biroli, G., Sagun, L.: Convit: Improving vision transformers with soft convolutional inductive biases. In: ICML (2021) +15. Eigen, D., Krishnan, D., Fergus, R.: Restoring an image taken through a window covered with dirt or rain. In: ICCV (2013) +16. Fu, X., Huang, J., Zeng, D., Huang, Y., Ding, X., Paisley, J.: Removing rain from single images via a deep detail network. In: CVPR (2017) +17. Gan, Y., Bai, Y., Lou, Y., Ma, X., Zhang, R., Shi, N., Luo, L.: Decorate the newcomers: Visual domain prompt for continual test time adaptation. In: AAAI (2023) +18. Guo, Y., Xiao, X., Chang, Y., Deng, S., Yan, L.: From sky to the ground: A large-scale benchmark and simple baseline towards real rain removal. In: ICCV (2023) +19. He, B., Wang, C., Shi, B., Duan, L.: Mop moiré patterns using mopnet. In: ICCV (2019) + +20. He, B., Wang, C., Shi, B., Duan, L.: Fhde $^2$ net: Full high definition demoireing network. In: ECCV (2020) +21. He, K., Sun, J., Tang, X.: Single image haze removal using dark channel prior. TPAMI (2010) +22. Hendrycks, D., Gimpel, K.: Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415 (2016) +23. Isola, P., Zhu, J.Y., Zhou, T., Efros, A.A.: Image-to-image translation with conditional adversarial networks. In: CVPR (2017) +24. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: ECCV (2022) +25. Jiang, L., Dai, B., Wu, W., Loy, C.C.: Focal frequency loss for image reconstruction and synthesis. In: CVPR (2021) +26. Khattak, M.U., Rasheed, H., Maaz, M., Khan, S., Khan, F.S.: Maple: Multi-modal prompt learning. In: CVPR (2023) +27. Kong, L., Dong, J., Ge, J., Li, M., Pan, J.: Efficient frequency domain-based transformers for high-quality image deblurring. In: CVPR (2023) +28. Li, B., Peng, X., Wang, Z., Xu, J., Feng, D.: Aod-net: All-in-one dehazing network. In: ICCV (2017) +29. Li, B., Ren, W., Fu, D., Tao, D., Feng, D., Zeng, W., Wang, Z.: Benchmarking single-image dehazing and beyond. TIP (2018) +30. Li, B., Liu, X., Hu, P., Wu, Z., Lv, J., Peng, X.: All-in-one image restoration for unknown corruption. In: CVPR (2022) +31. Li, Y., Tan, R.T., Guo, X., Lu, J., Brown, M.S.: Rain streak removal using layer priors. In: CVPR (2016) +32. Li, Z., Lei, Y., Ma, C., Zhang, J., Shan, H.: Prompt-in-prompt learning for universal image restoration. arXiv preprint arXiv:2312.05038 (2023) +33. Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: Swinir: Image restoration using swin transformer. In: ICCV Workshops (2021) +34. Liu, D., Wen, B., Fan, Y., Loy, C.C., Huang, T.S.: Non-local recurrent network for image restoration. In: NeurIPS (2018) +35. Liu, J., Yan, M., Zeng, T.: Surface-aware blind image deblurring. TPAMI (2021) +36. Liu, L., Liu, J., Yuan, S., Slabaugh, G., Leonardis, A., Zhou, W., Tian, Q.: Wavelet-based dual-branch network for image demoiring. In: ECCV (2020) +37. Liu, L., Xie, L., Zhang, X., Yuan, S., Chen, X., Zhou, W., Li, H., Tian, Q.: Tape: Task-agnostic prior embedding for image restoration. In: ECCV (2022) +38. Liu, X., Suganuma, M., Sun, Z., Okatani, T.: Dual residual networks leveraging the potential of paired operations for image restoration. In: CVPR (2019) +39. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: ICCV (2021) +40. Ma, J., Cheng, T., Wang, G., Zhang, Q., Wang, X., Zhang, L.: Prores: Exploring degradation-aware visual prompt for universal image restoration. arXiv preprint arXiv:2306.13653 (2023) +41. Mei, Y., Fan, Y., Zhou, Y.: Image super-resolution with non-local sparse attention. In: CVPR (2021) +42. Mittal, A., Soundararajan, R., Bovik, A.C.: Making a "completely blind" image quality analyzer. IEEE SPL (2012) +43. Nah, S., Hyun Kim, T., Mu Lee, K.: Deep multi-scale convolutional neural network for dynamic scene deblurring. In: CVPR (2017) +44. Narasimhan, S.G., Nayar, S.K.: Contrast restoration of weather degraded images. TPAMI (2003) + +45. Niu, B., Wen, W., Ren, W., Zhang, X., Yang, L., Wang, S., Zhang, K., Cao, X., Shen, H.: Single image super-resolution via a holistic attention network. In: ECCV (2020) +46. Oppenheim, A.: Discrete-time signal processing. Prentice-Hall google schola (1999) +47. Ozdenizci, O., Legenstein, R.: Restoring vision in adverse weather conditions with patch-based denoising diffusion models. TPAMI (2023) +48. Pan, X., Zhan, X., Dai, B., Lin, D., Loy, C.C., Luo, P.: Exploiting deep generative prior for versatile image restoration and manipulation. TPAMI (2022) +49. Park, N., Kim, S.: How do vision transformers work? In: ICLR (2022) +50. Potlapalli, V., Zamir, S.W., Khan, S., Khan, F.S.: Prompt: Prompting for all-in-one blind image restoration. In: NeurIPS (2023) +51. Purohit, K., Suin, M., Rajagopalan, A., Boddeti, V.N.: Spatially-adaptive image restoration using distortion-guided networks. In: ICCV (2021) +52. Qian, R., Tan, R.T., Yang, W., Su, J., Liu, J.: Attentive generative adversarial network for raindrop removal from a single image. In: CVPR (2018) +53. Qu, Y., Chen, Y., Huang, J., Xie, Y.: Enhanced pix2pix dehazing network. In: CVPR (2019) +54. Quan, Y., Deng, S., Chen, Y., Ji, H.: Deep learning for seeing through window with raindrops. In: ICCV (2019) +55. Rabiner, L.R., Gold, B.: Theory and application of digital signal processing. Englewood Cliffs: Prentice-Hall (1975) +56. Ren, D., Zuo, W., Hu, Q., Zhu, P., Meng, D.: Progressive image deraining networks: A better and simpler baseline. In: CVPR (2019) +57. Ren, W., Liu, S., Zhang, H., Pan, J., Cao, X., Yang, M.H.: Single image dehazing via multi-scale convolutional neural networks. In: ECCV (2016) +58. Ren, W., Zhang, J., Pan, J., Liu, S., Ren, J.S., Du, J., Cao, X., Yang, M.H.: Deblurring dynamic scenes via spatially varying recurrent neural networks. TPAMI (2022) +59. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: MICCAI (2015) +60. Song, X., Zhou, D., Li, W., Dai, Y., Shen, Z., Zhang, L., Li, H.: Tusr-net: Triple unfolding single image dehazing with self-regularization and dual feature to pixel attention. TIP (2023) +61. Sun, Y., Yu, Y., Wang, W.: Moiré photo restoration using multiresolution convolutional neural networks. TIP (2018) +62. Tu, Z., Talebi, H., Zhang, H., Yang, F., Milanfar, P., Bovik, A., Li, Y.: Maxim: Multi-axis mlp for image processing. In: CVPR (2022) +63. Valanarasu, J.M.J., Yasarla, R., Patel, V.M.: Transweather: Transformer-based restoration of images degraded by adverse weather conditions. In: CVPR (2022) +64. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017) +65. Voigtman, E., Winefordner, J.D.: Low-pass filters for signal averaging. Review of Scientific Instruments (1986) +66. Wang, C., He, B., Wu, S., Wan, R., Shi, B., Duan, L.Y.: Coarse-to-fine disentangling demoiréing framework for recaptured screen images. TPAMI (2023) +67. Wang, C., Pan, J., Lin, W., Dong, J., Wu, X.M.: Selfpromer: Self-prompt dehazing transformers with depth-consistency. arXiv preprint arXiv:2303.07033 (2023) +68. Wang, C., Pan, J., Wang, W., Dong, J., Wang, M., Ju, Y., Chen, J.: Promptstorer: A prompting image restoration method with degradation perception. In: NeurIPS (2023) + +69. Wang, H., Xie, Q., Zhao, Q., Meng, D.: A model-driven deep neural network for single image rain removal. In: CVPR (2020) +70. Wang, T., Yang, X., Xu, K., Chen, S., Zhang, Q., Lau, R.W.: Spatial attentive single-image deraining with a high quality real rain dataset. In: CVPR (2019) +71. Wang, W., Xie, E., Li, X., Fan, D.P., Song, K., Liang, D., Lu, T., Luo, P., Shao, L.: Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In: ICCV (2021) +72. Wang, Z., Cun, X., Bao, J., Zhou, W., Liu, J., Li, H.: Uformer: A general u-shaped transformer for image restoration. In: CVPR (2022) +73. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. TIP (2004) +74. Wu, R., Yang, T., Sun, L., Zhang, Z., Li, S., Zhang, L.: Seesr: Towards semantics-aware real-world image super-resolution. arXiv preprint arXiv:2311.16518 (2023) +75. Xiao, J., Fu, X., Liu, A., Wu, F., Zha, Z.J.: Image de-raining transformer. TPAMI (2022) +76. Xiao, T., Singh, M., Mintun, E., Darrell, T., Dollar, P., Girshick, R.: Early convolutions help transformers see better. In: NeurIPS (2021) +77. Yang, W., Tan, R.T., Feng, J., Guo, Z., Yan, S., Liu, J.: Joint rain detection and removal from a single image with contextualized deep networks. TPAMI (2019) +78. Yu, F., Gu, J., Li, Z., Hu, J., Kong, X., Wang, X., He, J., Qiao, Y., Dong, C.: Scaling up to excellence: Practicing model scaling for photo-realistic image restoration in the wild. arXiv preprint arXiv:2401.13627 (2024) +79. Yu, K., Wang, X., Dong, C., Tang, X., Loy, C.C.: Path-restore: Learning network path selection for image restoration. TPAMI (2022) +80. Yue, H., Mao, Y., Liang, L., Xu, H., Hou, C., Yang, J.: Recaptured screen image demoiring. TCSVT (2021) +81. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: CVPR (2022) +82. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Learning enriched features for real image restoration and enhancement. In: ECCV (2020) +83. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Multi-stage progressive image restoration. In: CVPR (2021) +84. Zhang, K., Li, Y., Zuo, W., Zhang, L., Van Gool, L., Timofte, R.: Plug-and-play image restoration with deep denoiser prior. TPAMI (2021) +85. Zhao, H., Gou, Y., Li, B., Peng, D., Lv, J., Peng, X.: Comprehensive and delicate: An efficient transformer for image restoration. In: CVPR (2023) +86. Zheng, B., Yuan, S., Slabaugh, G., Leonardis, A.: Image demoiring with learnable bandpass filters. In: CVPR (2020) +87. Zheng, C., Zhang, Y., Gu, J., Zhang, Y., Kong, L., Yuan, X.: Cross aggregation transformer for image restoration. In: NeurIPS (2022) +88. Zhou, S., Chen, D., Pan, J., Shi, J., Yang, J.: Adapt or perish: Adaptive sparse transformer with attentive feature refinement for image restoration. In: CVPR (2024) +89. Zou, X., Xiao, F., Yu, Z., Li, Y., Lee, Y.J.: Delving deeper into anti-aliasing in convnets. IJCV (2023) \ No newline at end of file diff --git a/2024/Seeing the Unseen_ A Frequency Prompt Guided Transformer for Image Restoration/images.zip b/2024/Seeing the Unseen_ A Frequency Prompt Guided Transformer for Image Restoration/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..53f950626f5317aa18f49bdf9a421d5a57a2546b --- /dev/null +++ b/2024/Seeing the Unseen_ A Frequency Prompt Guided Transformer for Image Restoration/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7eccfbb72245a6cd0e7c9998095874faadeb68af999877f574715040326c4a85 +size 603604 diff --git a/2024/Seeing the Unseen_ A Frequency Prompt Guided Transformer for Image Restoration/layout.json b/2024/Seeing the Unseen_ A Frequency Prompt Guided Transformer for Image Restoration/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..254ac8d3bd07656202d4d30988759f8845544aa5 --- /dev/null +++ b/2024/Seeing the Unseen_ A Frequency Prompt Guided Transformer for Image Restoration/layout.json @@ -0,0 +1,13553 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 136, + 111, + 478, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 111, + 478, + 148 + ], + "spans": [ + { + "bbox": [ + 136, + 111, + 478, + 148 + ], + "type": "text", + "content": "Seeing the Unseen: A Frequency Prompt Guided Transformer for Image Restoration" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 154, + 167, + 459, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 167, + 459, + 194 + ], + "spans": [ + { + "bbox": [ + 154, + 167, + 459, + 194 + ], + "type": "text", + "content": "Shihao Zhou" + }, + { + "bbox": [ + 154, + 167, + 459, + 194 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 154, + 167, + 459, + 194 + ], + "type": "text", + "content": ", Jinshan Pan" + }, + { + "bbox": [ + 154, + 167, + 459, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 154, + 167, + 459, + 194 + ], + "type": "text", + "content": ", Jinglei Shi" + }, + { + "bbox": [ + 154, + 167, + 459, + 194 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 154, + 167, + 459, + 194 + ], + "type": "text", + "content": ", Duosheng Chen" + }, + { + "bbox": [ + 154, + 167, + 459, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 154, + 167, + 459, + 194 + ], + "type": "text", + "content": ", Lishen Qu" + }, + { + "bbox": [ + 154, + 167, + 459, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 154, + 167, + 459, + 194 + ], + "type": "text", + "content": ", and Jufeng Yang" + }, + { + "bbox": [ + 154, + 167, + 459, + 194 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 147, + 201, + 465, + 224 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 147, + 201, + 465, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 201, + 465, + 213 + ], + "spans": [ + { + "bbox": [ + 147, + 201, + 465, + 213 + ], + "type": "text", + "content": "1 VCIP & TMCC & DISSec, College of Computer Science, Nankai University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 147, + 213, + 465, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 213, + 465, + 224 + ], + "spans": [ + { + "bbox": [ + 147, + 213, + 465, + 224 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 147, + 213, + 465, + 224 + ], + "type": "text", + "content": " Nankai International Advanced Research Institute (SHENZHEN·FUTIAN)" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 138, + 224, + 473, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 224, + 473, + 246 + ], + "spans": [ + { + "bbox": [ + 138, + 224, + 473, + 246 + ], + "type": "text", + "content": "3 School of Computer Science and Engineering, Nanjing University of Science and Technology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 174, + 247, + 438, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 247, + 438, + 269 + ], + "spans": [ + { + "bbox": [ + 174, + 247, + 438, + 269 + ], + "type": "text", + "content": "{zhoushihao96,duoshengchen,lishenqu}@mail.nankai.edu.cn, \n{jinglei.shi,yangjufeng}@nankai.edu.cn,SDLuran@gmail.com" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 159, + 293, + 455, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 293, + 455, + 537 + ], + "spans": [ + { + "bbox": [ + 159, + 293, + 455, + 537 + ], + "type": "text", + "content": "Abstract. How to explore useful features from images as prompts to guide the deep image restoration models is an effective way to solve image restoration. In contrast to mining spatial relations within images as prompt, which leads to characteristics of different frequencies being neglected and further remaining subtle or undetectable artifacts in the restored image, we develop a Frequency Prompting image restoration method, dubbed FPro, which can effectively provide prompt components from a frequency perspective to guild the restoration model address these differences. Specifically, we first decompose input features into separate frequency parts via dynamically learned filters, where we introduce a gating mechanism for suppressing the less informative elements within the kernels. To propagate useful frequency information as prompt, we then propose a dual prompt block, consisting of a low-frequency prompt modulator (LPM) and a high-frequency prompt modulator (HPM), to handle signals from different bands respectively. Each modulator contains a generation process to incorporate prompting components into the extracted frequency maps, and a modulation part that modifies the prompt feature with the guidance of the decoder features. Experimental results on several popular datasets have demonstrated the favorable performance of our pipeline against SOTA methods on 5 image restoration tasks, including deraining, deraindrop, demoiréing, deblurring, and dehazing. The source code is available at https://github.com/joshyZhou/FPro." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 160, + 545, + 434, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 545, + 434, + 558 + ], + "spans": [ + { + "bbox": [ + 160, + 545, + 434, + 558 + ], + "type": "text", + "content": "Keywords: Image Restoration " + }, + { + "bbox": [ + 160, + 545, + 434, + 558 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 545, + 434, + 558 + ], + "type": "text", + "content": " Frequency Prompt " + }, + { + "bbox": [ + 160, + 545, + 434, + 558 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 545, + 434, + 558 + ], + "type": "text", + "content": " Transformer" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 576, + 231, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 576, + 231, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 576, + 231, + 588 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 599, + 482, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 599, + 482, + 648 + ], + "spans": [ + { + "bbox": [ + 130, + 599, + 482, + 648 + ], + "type": "text", + "content": "Capturing images in unsatisfactory environments, e.g., rain, haze, usually leads to low-quality ones that accordingly affect the application of downstream tasks in practice. Thus, developing an effective image restoration method to restore clear images from degraded ones is an important task." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 236, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 236, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 236, + 666 + ], + "type": "text", + "content": "* Corresponding author." + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 164 + ], + "type": "text", + "content": "Significant progress has been made due to kinds of the deep learning models [4,5,81], and these deep learning-based approaches become predominant ones as they achieve better performance than the conventional hand-crafted prior-based approaches [3,21,31,44,84]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 167, + 480, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 167, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 130, + 167, + 480, + 251 + ], + "type": "text", + "content": "Existing methods, e.g., [33, 72, 81] achieve promising performances in kinds of image restoration tasks. However, these learning-based methods intend to learn a mapping function between degraded images and clear ones, where the characteristics of the specific degradation are less considered. For example, rain streaks tend to obscure the background partially, whereas raindrops typically result in a more pronounced regional occlusion. Accordingly, these models are hindered from generating better results." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 253, + 481, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 253, + 481, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 253, + 481, + 384 + ], + "type": "text", + "content": "More recently, prompt-learning based methods [50,67,68] serve as an alternative approach to encode useful content of specific degradation for modulating the network, and make a clear performance boost for image restoration. However, we notice that these methods [50,68] pay attention to mining spatial correlations to provide degradation information, whereas the task-specific frequency cues are less studied. Indeed, since various forms of degradation exhibit distinct impacts on image content, they affect information from different frequency bands. Hence, it is crucial to develop an efficient prompt mechanism that explores useful prompts from a frequency perspective for identifying specific characteristics of diverse degradation, which can boost the model to effectively restore images with finer details and non-local structures of the scenes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 388, + 481, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 481, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 481, + 483 + ], + "type": "text", + "content": "This paper proposes a Frequency Prompting image restoration method, dubbed FPro, to modulate the network by encoding degradation-specific frequency cues as prompts. As mentioned above, existing prompt strategies [50,68] focus on mining spatial relations as useful prompts. In this way, differences between the restored image and the real one within frequency domain [25] are ignored, which remain subtle or undetectable artifacts in the spatial domain. Instead, our FPro aims to enjoy benefits from the capability of prompt learning in different frequency bands at multi-scale resolutions to recover clean images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 486, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 481, + 665 + ], + "type": "text", + "content": "We present two designs to enhance FPro for general image restoration: 1). We first decouple input features into separate low-/high-frequency parts using a gated dynamic decoupler, as signals in different frequency bands encode image patterns from distinct views, i.e., local details and global structures. To this end, a gating mechanism is introduced to help learn the enhanced low-pass filters by suppressing the less informative elements within the kernel, which are then employed to generate low-frequency maps. Meanwhile, the corresponding high-pass filter is obtained by subtracting the low-pass filter from the identity kernel, for generating high-frequency maps. 2). We propose a Dual Prompt Block (DPB), which consists of two modulators, i.e., the Low-frequency Prompt Modulator (LPM) and the High-frequency Prompt Modulator (HPM), to handle low- and high-frequency information respectively. Each modulator includes (a) a generation part that incorporates prompting components into the extracted frequency maps, which is supposed to help distinguish various elements within features, such as rain patterns in the context of deraining; and (b) a modu" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "type": "text", + "content": "Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 223 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 223 + ], + "type": "text", + "content": "lation part that modifies the prompt feature with the guidance of the feature in the restoration process. In terms of functionality, LPM enhances the low-frequency characteristics through a gating mechanism in the Fourier domain before injecting the prompting components, which is proven equivalent to dynamic large-kernel depth-wise convolution in the spatial domain while computationally efficient, and then encodes low-frequency interactions via global cross-attention. As a complement, HPM applies a locally-enhanced gating mechanism to obtain useful high-frequency signals, and then encodes high-frequency interactions via local cross-attention." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 147, + 224, + 364, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 224, + 364, + 236 + ], + "spans": [ + { + "bbox": [ + 147, + 224, + 364, + 236 + ], + "type": "text", + "content": "We summarize our main contributions as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 243, + 484, + 384 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 138, + 243, + 479, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 243, + 479, + 300 + ], + "spans": [ + { + "bbox": [ + 138, + 243, + 479, + 300 + ], + "type": "text", + "content": "- We propose FPro, which benefits from prompting learning of frequency components for general image restoration. Instead of mining spatial relations as in previous methods, we explore frequency maps to encode specific degradation information as prompts to guide the image restoration model for restoring finer details and the global structure of the scenes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 302, + 479, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 302, + 479, + 347 + ], + "spans": [ + { + "bbox": [ + 138, + 302, + 479, + 347 + ], + "type": "text", + "content": "- We decouple input features into different frequency bands using learnable low-pass filters, and propose a dual prompt block, which is composed of low-frequency prompt modulator (LPM) and high-frequency prompt modulator (HPM), to explore both details and structures for better restoration." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 350, + 484, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 350, + 484, + 384 + ], + "spans": [ + { + "bbox": [ + 138, + 350, + 484, + 384 + ], + "type": "text", + "content": "- Experimental results on 5 restoration tasks: deraining, deraindrop, demoiréing, deblurring and dehazing, show that FPro achieves favorable performance against state-of-the-art methods." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 403, + 237, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 237, + 415 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 237, + 415 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 426, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 593 + ], + "type": "text", + "content": "Image Restoration. The goal of the image restoration task is to recover high-quality images from the degraded ones. Going beyond conventional prior-based solutions [3, 21], this community has witnessed the great success of a body of learning-based approaches [34, 48, 79]. Despite the promising results obtained by various CNN-based architectures [9, 35, 58], the main concern for methods of this kind is that they pose a limited receptive field problem of the basic convolution operation. This means that the feature map contains less global context (corresponding to low-frequency characteristics in an image), and the final prediction can get stuck in this limitation. This drawback has motivated the increased interest in exploring components to capture desired global cues, like attention mechanisms [10, 45, 60], where better restoration performance can be achieved. For instance, MIRNet [82] proposes a dual attention unit to capture contextual information in dual dimensions. NLSN [41] employs a self-attention mechanism to collect global correlation information for super-resolution." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Transformer-based Restoration. In recent years, the idea of using Transformer architecture [64] to address various computer vision-related tasks has been popular. Thanks to their discriminative feature representation capability, they not only earn advantages in solving high-level vision tasks [13, 14, 71], but also are extended to low-level image restoration tasks [27, 85, 88]. Unfortunately, as the complexity of vanilla self-attention is quadratic to the image size," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 184, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 184, + 91, + 447, + 103 + ], + "type": "text", + "content": "A Frequency Prompt Guided Transformer for Image Restoration" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 295 + ], + "type": "text", + "content": "this mechanism suffers from non-trivial computational costs in handling high-resolution input. To address this, some attempts have been made to explore efficient transformer architectures [4,72,87]. Specifically, SwinIR [33] introduces a window-based self-attention scheme to improve efficiency. Restormer [81] adopts channel-wise self-attention to reduce the computational costs. The majority of these works have offered reliable solutions to recover clean images, however, some works [11,49] realized that the low-pass filter nature of self-attention, which could lose the high-frequency information, such as textures and edges. Even though these models have achieved superior performance, few high-frequency details can be leveraged to implement image restoration, limiting better recovery as a result. Visual Prompt Learning. Prompt learning, which emerged recently in the NLP field [1], has resulted in rapid advancements in its adaptation to computer vision tasks [17,24,26]. Contrary to high-level vision problems, motivated by high effectiveness, some works also consider seeking proper prompts for the low-level pipelines [40, 74, 78]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 296, + 482, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 296, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 482, + 380 + ], + "type": "text", + "content": "The aim of this paper is not to prompt models for addressing the ALL-in-One problem (in fact, the previous works of [32,40,50] have addressed this nicely by designing various degradation prompt modules). However, our approach is relevant to recent studies [67,68] exploring degradation-specific information for better image restoration results. In contrast to these attempts that generate raw degradation features with a pre-trained model, we prompt the restoration models from a frequency perspective." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 396, + 261, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 261, + 410 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 261, + 410 + ], + "type": "text", + "content": "3 Proposed Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 420, + 243, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 420, + 243, + 432 + ], + "spans": [ + { + "bbox": [ + 132, + 420, + 243, + 432 + ], + "type": "text", + "content": "3.1 Overall Pipeline" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 438, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 482, + 666 + ], + "type": "text", + "content": "As depicted in Fig. 1, the overview of our proposed FPro contains the upper restoration branch, like existing works [33,81], and the bottom prompt branch to extract informative frequency maps and then modulate them as prompts. Restoration Branch. Given a degraded image " + }, + { + "bbox": [ + 130, + 438, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 130, + 438, + 482, + 666 + ], + "type": "text", + "content": " as input, FPro first employs a convolution layer to obtain shallow feature " + }, + { + "bbox": [ + 130, + 438, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_s \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 130, + 438, + 482, + 666 + ], + "type": "text", + "content": "; where " + }, + { + "bbox": [ + 130, + 438, + 482, + 666 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 130, + 438, + 482, + 666 + ], + "type": "text", + "content": " represents the spatial dimension and " + }, + { + "bbox": [ + 130, + 438, + 482, + 666 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 438, + 482, + 666 + ], + "type": "text", + "content": " is the number of channel. Next, the shallow feature passes through the upper " + }, + { + "bbox": [ + 130, + 438, + 482, + 666 + ], + "type": "inline_equation", + "content": "N_1" + }, + { + "bbox": [ + 130, + 438, + 482, + 666 + ], + "type": "text", + "content": "-level encoder-decoder restoration branch to extract deep feature " + }, + { + "bbox": [ + 130, + 438, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_d \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 130, + 438, + 482, + 666 + ], + "type": "text", + "content": ". Early layers in Transformer-based models focus on aggregate local patterns [76], whereas the self-attention module acts as a low-pass filter and tends to dilute high-frequency local details [49]. To alleviate the two contradictory factors, we remove the attention mechanism within the encoder of the restoration branch. Specifically, each level of the encoder includes " + }, + { + "bbox": [ + 130, + 438, + 482, + 666 + ], + "type": "inline_equation", + "content": "N_2" + }, + { + "bbox": [ + 130, + 438, + 482, + 666 + ], + "type": "text", + "content": " feed-forward network (FFN) [81] and the paired convolution layer for down-sampling. The encoder features are then fused with the ones in the decoder through skip connections by " + }, + { + "bbox": [ + 130, + 438, + 482, + 666 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 130, + 438, + 482, + 666 + ], + "type": "text", + "content": " convolution. For the decoder part, each level is composed of " + }, + { + "bbox": [ + 130, + 438, + 482, + 666 + ], + "type": "inline_equation", + "content": "N_2" + }, + { + "bbox": [ + 130, + 438, + 482, + 666 + ], + "type": "text", + "content": " pairs of FFN and multi-head self-attention mechanisms (MSA) [81], along with the convolution layer for up-sampling. Finally, a " + }, + { + "bbox": [ + 130, + 438, + 482, + 666 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 438, + 482, + 666 + ], + "type": "text", + "content": " convolution layer is employed to deep feature " + }, + { + "bbox": [ + 130, + 438, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_d" + }, + { + "bbox": [ + 130, + 438, + 482, + 666 + ], + "type": "text", + "content": " for generating residual image " + }, + { + "bbox": [ + 130, + 438, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{R} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 130, + 438, + 482, + 666 + ], + "type": "text", + "content": ". The restored image " + }, + { + "bbox": [ + 130, + 438, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}" + }, + { + "bbox": [ + 130, + 438, + 482, + 666 + ], + "type": "text", + "content": " is estimated by: " + }, + { + "bbox": [ + 130, + 438, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}} = \\mathbf{I} + \\mathbf{R}" + }, + { + "bbox": [ + 130, + 438, + 482, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "type": "text", + "content": "Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 478, + 243 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 478, + 243 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 478, + 243 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 478, + 243 + ], + "type": "image", + "image_path": "1f959115b233ea26db4867f087af8516eb152ff14c61502f2a3eedd4b992396a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 251, + 482, + 361 + ], + "lines": [ + { + "bbox": [ + 130, + 251, + 482, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 482, + 361 + ], + "type": "text", + "content": "Fig. 1: Overview of the proposed FPro. Except for the common upper restoration branch, which is similar to existing methods [33,81], FPro contains another bottom prompt branch to extract informative features from a frequency perspective. Specifically, the primary components of the prompt branch in this framework are the gated dynamic decoupler (GDD) and dual prompt block (DPB). The GDD is employed to decompose the low-frequency components and corresponding high-frequency characteristics from the input features. Then these frequency-specific features are further processed in DPB, i.e., the high-frequency prompt modulator (HPM) and the low-frequency prompt modulator (LPM), which generates representative frequency prompt to facilitate the clear image reconstruction." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 385, + 482, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 385, + 482, + 493 + ], + "spans": [ + { + "bbox": [ + 130, + 385, + 482, + 493 + ], + "type": "text", + "content": "Prompt Branch. In this branch, we take as input the shallow feature " + }, + { + "bbox": [ + 130, + 385, + 482, + 493 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_s" + }, + { + "bbox": [ + 130, + 385, + 482, + 493 + ], + "type": "text", + "content": " to generate useful frequency prompts, which are further leveraged to facilitate the latent clear image reconstruction. To achieve this goal, we first decompose the input feature into different frequency bands using a gated dynamic decouple (GDD) (see Section 3.2). After that, low-/high-frequency maps are injected with prompt components to distinguish informative elements according to specific tasks, and then modulated as different prompts (i.e., " + }, + { + "bbox": [ + 130, + 385, + 482, + 493 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{hi}^{out}" + }, + { + "bbox": [ + 130, + 385, + 482, + 493 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 385, + 482, + 493 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{low}^{out}" + }, + { + "bbox": [ + 130, + 385, + 482, + 493 + ], + "type": "text", + "content": ") to interact with the decoder features by " + }, + { + "bbox": [ + 130, + 385, + 482, + 493 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 130, + 385, + 482, + 493 + ], + "type": "text", + "content": " convolution (see Section 3.3). Next, we present the modules of the prompt branch." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 512, + 294, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 294, + 525 + ], + "type": "text", + "content": "3.2 Gated Dynamic Decoupler" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Each type of degradation affects image content in different ways. For instance, rain streaks partially occlude the background while raindrops often cause much greater obstruction, which corresponds to touch high-/low-frequency bands respectively. To handle these differences, as shown in Fig. 2, we decompose the input features into separate frequency parts based on gated and dynamically learned filters. The key ingredient is to introduce a gating mechanism to help generate the gated learnable low-pass filter and the corresponding high-pass filter, which are then employed to obtain low- and high-frequency maps, respectively. These filters are dynamically learned for each spatial location and channel group to balance computation burden and feature diversity. Specifically, given the input shallow feature map " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_s\\in \\mathbb{R}^{H\\times W\\times C}" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": ", we firstly predict the low-pass" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 184, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 184, + 91, + 447, + 102 + ], + "type": "text", + "content": "A Frequency Prompt Guided Transformer for Image Restoration" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 113, + 478, + 231 + ], + "blocks": [ + { + "bbox": [ + 135, + 113, + 478, + 231 + ], + "lines": [ + { + "bbox": [ + 135, + 113, + 478, + 231 + ], + "spans": [ + { + "bbox": [ + 135, + 113, + 478, + 231 + ], + "type": "image", + "image_path": "bde03dd66a0f3c102343539db53440dd6486179893e19e00d528aea0a9410f04.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 195, + 239, + 417, + 250 + ], + "lines": [ + { + "bbox": [ + 195, + 239, + 417, + 250 + ], + "spans": [ + { + "bbox": [ + 195, + 239, + 417, + 250 + ], + "type": "text", + "content": "Fig. 2: Illustrations of the Gated Dynamic Decoupler." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 272, + 420, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 272, + 420, + 284 + ], + "spans": [ + { + "bbox": [ + 131, + 272, + 420, + 284 + ], + "type": "text", + "content": "filter for each feature channel group, which can be formulated as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 243, + 290, + 361, + 305 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 243, + 290, + 361, + 305 + ], + "spans": [ + { + "bbox": [ + 243, + 290, + 361, + 305 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {F}} _ {s} = \\operatorname {C o n v} _ {1 \\times 1} (\\operatorname {G A P} (\\mathbf {F} _ {s})),", + "image_path": "90d454ca64f38dec83895ba0699ba994b8d177268a5968816866f4ec9a6416b9.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 243, + 308, + 481, + 323 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 243, + 308, + 481, + 323 + ], + "spans": [ + { + "bbox": [ + 243, + 308, + 481, + 323 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {F}} _ {s} = \\hat {\\mathbf {F}} _ {s} \\odot \\phi (\\operatorname {C o n v} _ {1 \\times 1} (\\hat {\\mathbf {F}} _ {s})), \\tag {1}", + "image_path": "75f16c247083fc9d4ae843fca2ad513337c51a19270ef3d20a17251f77c2f40c.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 246, + 323, + 342, + 339 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 323, + 342, + 339 + ], + "spans": [ + { + "bbox": [ + 246, + 323, + 342, + 339 + ], + "type": "interline_equation", + "content": "\\mathbf {F} ^ {l} = \\mathrm {S o f t m a x} (\\mathcal {B} (\\tilde {\\mathbf {F}} _ {s}))", + "image_path": "665379946f512a91397cdec6c91dd3c27d21e3a11dc290b5c9897e28b56ce771.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 346, + 482, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 346, + 482, + 433 + ], + "spans": [ + { + "bbox": [ + 131, + 346, + 482, + 433 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 346, + 482, + 433 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^l\\in \\mathbb{R}^{g\\times k^2\\times 1\\times 1}" + }, + { + "bbox": [ + 131, + 346, + 482, + 433 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 346, + 482, + 433 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 131, + 346, + 482, + 433 + ], + "type": "text", + "content": " is the number of channel groups and " + }, + { + "bbox": [ + 131, + 346, + 482, + 433 + ], + "type": "inline_equation", + "content": "k^2" + }, + { + "bbox": [ + 131, + 346, + 482, + 433 + ], + "type": "text", + "content": " corresponds to the kernel size of the learned filter; " + }, + { + "bbox": [ + 131, + 346, + 482, + 433 + ], + "type": "inline_equation", + "content": "\\mathrm{GAP}(\\cdot)" + }, + { + "bbox": [ + 131, + 346, + 482, + 433 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 346, + 482, + 433 + ], + "type": "inline_equation", + "content": "\\mathrm{Conv}_{1\\times 1}(\\cdot)" + }, + { + "bbox": [ + 131, + 346, + 482, + 433 + ], + "type": "text", + "content": " are global average pooling layer and convolution operation with the filter size of " + }, + { + "bbox": [ + 131, + 346, + 482, + 433 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 131, + 346, + 482, + 433 + ], + "type": "text", + "content": ", respectively; " + }, + { + "bbox": [ + 131, + 346, + 482, + 433 + ], + "type": "inline_equation", + "content": "\\phi (\\cdot)" + }, + { + "bbox": [ + 131, + 346, + 482, + 433 + ], + "type": "text", + "content": " denotes sigmoid activation, " + }, + { + "bbox": [ + 131, + 346, + 482, + 433 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 131, + 346, + 482, + 433 + ], + "type": "text", + "content": " refers to the element-wise product, and " + }, + { + "bbox": [ + 131, + 346, + 482, + 433 + ], + "type": "inline_equation", + "content": "\\mathcal{B}(\\cdot)" + }, + { + "bbox": [ + 131, + 346, + 482, + 433 + ], + "type": "text", + "content": " means Batch Normalization. Particularly, " + }, + { + "bbox": [ + 131, + 346, + 482, + 433 + ], + "type": "inline_equation", + "content": "\\mathrm{Softmax}(\\cdot)" + }, + { + "bbox": [ + 131, + 346, + 482, + 433 + ], + "type": "text", + "content": " is a softmax layer, which ensures the generated filters are low-pass [89]. Then, we apply these learned filters to each group input feature " + }, + { + "bbox": [ + 131, + 346, + 482, + 433 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i\\in \\mathbb{R}^{H\\times W\\times C_i}" + }, + { + "bbox": [ + 131, + 346, + 482, + 433 + ], + "type": "text", + "content": " to obtain low-frequency components:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 235, + 444, + 481, + 470 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 444, + 481, + 470 + ], + "spans": [ + { + "bbox": [ + 235, + 444, + 481, + 470 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {i, c, h, w} ^ {l o} = \\sum_ {p, q} \\mathbf {F} _ {i, p, q} ^ {L} \\mathbf {F} _ {i, c, h + p, w + q}, \\tag {2}", + "image_path": "0056639c828541e08d4876ac0d9188d74add36005c507eb4e76220eaa3aa3212.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 473, + 481, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 481, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 481, + 510 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 473, + 481, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^L\\in \\mathbb{R}^{g\\times k\\times k}" + }, + { + "bbox": [ + 130, + 473, + 481, + 510 + ], + "type": "text", + "content": " is the reshaped filter, " + }, + { + "bbox": [ + 130, + 473, + 481, + 510 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 473, + 481, + 510 + ], + "type": "text", + "content": " denotes the group index, " + }, + { + "bbox": [ + 130, + 473, + 481, + 510 + ], + "type": "inline_equation", + "content": "C_i = \\frac{C}{g}" + }, + { + "bbox": [ + 130, + 473, + 481, + 510 + ], + "type": "text", + "content": " refers to number of the group channel, " + }, + { + "bbox": [ + 130, + 473, + 481, + 510 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 473, + 481, + 510 + ], + "type": "text", + "content": " means the index of a channel, " + }, + { + "bbox": [ + 130, + 473, + 481, + 510 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 130, + 473, + 481, + 510 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 473, + 481, + 510 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 130, + 473, + 481, + 510 + ], + "type": "text", + "content": " are spatial coordinates, " + }, + { + "bbox": [ + 130, + 473, + 481, + 510 + ], + "type": "inline_equation", + "content": "p,q\\in \\{-1,0,1\\}" + }, + { + "bbox": [ + 130, + 473, + 481, + 510 + ], + "type": "text", + "content": " point to the surrounding locations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 511, + 481, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 511, + 481, + 547 + ], + "spans": [ + { + "bbox": [ + 130, + 511, + 481, + 547 + ], + "type": "text", + "content": "Meanwhile, we invert this process by subtracting the low-pass filter from the identity kernel to attain the high-pass filter, which is employed to generate the corresponding high-frequency components " + }, + { + "bbox": [ + 130, + 511, + 481, + 547 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{hi}" + }, + { + "bbox": [ + 130, + 511, + 481, + 547 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 562, + 260, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 562, + 260, + 574 + ], + "spans": [ + { + "bbox": [ + 131, + 562, + 260, + 574 + ], + "type": "text", + "content": "3.3 Dual Prompt Block" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 580, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 580, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 580, + 482, + 640 + ], + "type": "text", + "content": "Considering that the extracted features, i.e., low-/high-frequency maps, encode image patterns from distinct views (local detail and main structure of the image). We design the Dual Prompt Block that includes two components, i.e., High-frequency Prompt Modulator (HPM) and Low-frequency Prompt Modulator (LPM), to deal with these feature maps, respectively." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "High-frequency Prompt Modulator. Given the two input feature maps, including the " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "-level feature " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_l \\in \\mathbb{R}^{\\hat{H} \\times \\hat{W} \\times \\hat{C}}" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": " and high-frequency feature " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{hi} \\in" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 115, + 479, + 212 + ], + "blocks": [ + { + "bbox": [ + 134, + 115, + 479, + 212 + ], + "lines": [ + { + "bbox": [ + 134, + 115, + 479, + 212 + ], + "spans": [ + { + "bbox": [ + 134, + 115, + 479, + 212 + ], + "type": "image", + "image_path": "8304b4ce1b3aa19dc61980f1ddb181234139e82114cd6e96adf9522a75660b31.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 214, + 479, + 312 + ], + "blocks": [ + { + "bbox": [ + 133, + 214, + 479, + 312 + ], + "lines": [ + { + "bbox": [ + 133, + 214, + 479, + 312 + ], + "spans": [ + { + "bbox": [ + 133, + 214, + 479, + 312 + ], + "type": "image", + "image_path": "922e68f04880e8da8fae2aa4e54cb778321805861a6af23fc0d475e194619642.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 319, + 480, + 343 + ], + "lines": [ + { + "bbox": [ + 131, + 319, + 480, + 343 + ], + "spans": [ + { + "bbox": [ + 131, + 319, + 480, + 343 + ], + "type": "text", + "content": "Fig. 3: Illustrations of the proposed components. (a) High-frequency Prompt Modulator (HPM); (b) Low-frequency Prompt Modulator (LPM)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 362, + 479, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 362, + 479, + 399 + ], + "spans": [ + { + "bbox": [ + 131, + 362, + 479, + 399 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{H\\times W\\times C'}" + }, + { + "bbox": [ + 131, + 362, + 479, + 399 + ], + "type": "text", + "content": ", we first resize " + }, + { + "bbox": [ + 131, + 362, + 479, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{hi}" + }, + { + "bbox": [ + 131, + 362, + 479, + 399 + ], + "type": "text", + "content": " and obtain " + }, + { + "bbox": [ + 131, + 362, + 479, + 399 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{F}}_{hi}\\in \\mathbb{R}^{\\hat{H}\\times \\hat{W}\\times C'}" + }, + { + "bbox": [ + 131, + 362, + 479, + 399 + ], + "type": "text", + "content": ". Towards highlighting high-frequency characteristics, we employ a gating mechanism to adaptively determine the useful frequency information:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 233, + 405, + 480, + 420 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 405, + 480, + 420 + ], + "spans": [ + { + "bbox": [ + 233, + 405, + 480, + 420 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {F}} _ {h i} = \\tilde {\\mathbf {F}} _ {h i} \\odot \\sigma (\\mathrm {D C o n v} _ {3 \\times 3} (\\tilde {\\mathbf {F}} _ {h i})), \\tag {3}", + "image_path": "13929d69b449f1f299195ba324a51a28ac344304511a210982ecc8b128ef634d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 426, + 482, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 426, + 482, + 502 + ], + "spans": [ + { + "bbox": [ + 131, + 426, + 482, + 502 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 426, + 482, + 502 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_{hi} \\in \\mathbb{R}^{\\hat{H} \\times \\hat{W} \\times C'}" + }, + { + "bbox": [ + 131, + 426, + 482, + 502 + ], + "type": "text", + "content": " is the processed feature, " + }, + { + "bbox": [ + 131, + 426, + 482, + 502 + ], + "type": "inline_equation", + "content": "\\mathrm{DConv}_{3 \\times 3}(\\cdot)" + }, + { + "bbox": [ + 131, + 426, + 482, + 502 + ], + "type": "text", + "content": " denotes a depth-wise convolution operation with the filter size of " + }, + { + "bbox": [ + 131, + 426, + 482, + 502 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 131, + 426, + 482, + 502 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 426, + 482, + 502 + ], + "type": "inline_equation", + "content": "\\sigma(\\cdot)" + }, + { + "bbox": [ + 131, + 426, + 482, + 502 + ], + "type": "text", + "content": " is the GELU activation function [22]. Then, we leverage the learnable high-frequency prompt components " + }, + { + "bbox": [ + 131, + 426, + 482, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{hi} \\in \\mathbb{R}^{\\hat{H} \\times \\hat{W} \\times C'}" + }, + { + "bbox": [ + 131, + 426, + 482, + 502 + ], + "type": "text", + "content": " to make adjustments to the input features, which aims to help distinguish various elements, such as rain patterns and streaks of different orientations and magnitudes in the context of deraining:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 258, + 506, + 480, + 521 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 506, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 258, + 506, + 480, + 521 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {h i} ^ {\\text {p r o m p t}} = \\hat {\\mathbf {F}} _ {h i} \\odot \\mathbf {P} _ {h i}, \\tag {4}", + "image_path": "2eb9751a4fd52f7c4be255241c0c3777134e59649ac4b036155c3e044468f107.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 528, + 460, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 528, + 460, + 542 + ], + "spans": [ + { + "bbox": [ + 131, + 528, + 460, + 542 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 528, + 460, + 542 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{hi}^{prompt} \\in \\mathbb{R}^{\\hat{H} \\times \\hat{W} \\times C'}" + }, + { + "bbox": [ + 131, + 528, + 460, + 542 + ], + "type": "text", + "content": " is the obtained high-frequency feature prompt." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 543, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 543, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 543, + 482, + 666 + ], + "type": "text", + "content": "Next, we modify the high-frequency prompt " + }, + { + "bbox": [ + 130, + 543, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{hi}^{prompt}" + }, + { + "bbox": [ + 130, + 543, + 482, + 666 + ], + "type": "text", + "content": " according to the input feature " + }, + { + "bbox": [ + 130, + 543, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_l" + }, + { + "bbox": [ + 130, + 543, + 482, + 666 + ], + "type": "text", + "content": ". To be specific, we utilize a depth-wise convolution operator, which acts as a high-pass filter [49], to enhance the high-frequency sources in the input " + }, + { + "bbox": [ + 130, + 543, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{hi}^{prompt}" + }, + { + "bbox": [ + 130, + 543, + 482, + 666 + ], + "type": "text", + "content": ". Then, we generate query " + }, + { + "bbox": [ + 130, + 543, + 482, + 666 + ], + "type": "inline_equation", + "content": "(\\mathbf{Q}_{hi})" + }, + { + "bbox": [ + 130, + 543, + 482, + 666 + ], + "type": "text", + "content": " projection from " + }, + { + "bbox": [ + 130, + 543, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_l" + }, + { + "bbox": [ + 130, + 543, + 482, + 666 + ], + "type": "text", + "content": ", key " + }, + { + "bbox": [ + 130, + 543, + 482, + 666 + ], + "type": "inline_equation", + "content": "(\\mathbf{K}_{hi})" + }, + { + "bbox": [ + 130, + 543, + 482, + 666 + ], + "type": "text", + "content": " and value " + }, + { + "bbox": [ + 130, + 543, + 482, + 666 + ], + "type": "inline_equation", + "content": "(\\mathbf{V}_{hi})" + }, + { + "bbox": [ + 130, + 543, + 482, + 666 + ], + "type": "text", + "content": " projections from the processed feature map " + }, + { + "bbox": [ + 130, + 543, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_{hi}^{prompt} = \\mathrm{DConv}_{3\\times 3}(\\mathbf{F}_{hi}^{prompt})" + }, + { + "bbox": [ + 130, + 543, + 482, + 666 + ], + "type": "text", + "content": ", respectively. Meanwhile, as the high-frequency information usually corresponds to image details and is a local feature, it could be redundant to calculate global attention. Therefore, before leveraging the linear layer to obtain the matrices of " + }, + { + "bbox": [ + 130, + 543, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{hi}" + }, + { + "bbox": [ + 130, + 543, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 543, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_{hi}" + }, + { + "bbox": [ + 130, + 543, + 482, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 543, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{hi}" + }, + { + "bbox": [ + 130, + 543, + 482, + 666 + ], + "type": "text", + "content": ", the local window self-attention mechanism is adopted to save computational complexity and capture fine-grained high" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 184, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 184, + 91, + 447, + 102 + ], + "type": "text", + "content": "A Frequency Prompt Guided Transformer for Image Restoration" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 189 + ], + "type": "text", + "content": "frequencies, which yields " + }, + { + "bbox": [ + 132, + 114, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{hi} = W_{p}^{Q_{hi}}\\cdot R(\\mathbf{F}_{l})" + }, + { + "bbox": [ + 132, + 114, + 482, + 189 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 114, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_{hi} = W_{p}^{K_{hi}}\\cdot R(\\hat{\\mathbf{F}}_{hi}^{prompt})" + }, + { + "bbox": [ + 132, + 114, + 482, + 189 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 114, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{hi} = W_{p}^{V_{hi}}\\cdot R(\\hat{\\mathbf{F}}_{hi}^{prompt})" + }, + { + "bbox": [ + 132, + 114, + 482, + 189 + ], + "type": "text", + "content": ". Where " + }, + { + "bbox": [ + 132, + 114, + 482, + 189 + ], + "type": "inline_equation", + "content": "W_{p}^{(\\cdot)}" + }, + { + "bbox": [ + 132, + 114, + 482, + 189 + ], + "type": "text", + "content": " represents the projection matrices, and " + }, + { + "bbox": [ + 132, + 114, + 482, + 189 + ], + "type": "inline_equation", + "content": "R(\\cdot)" + }, + { + "bbox": [ + 132, + 114, + 482, + 189 + ], + "type": "text", + "content": " denotes the window partition strategy [39]. Generally, we have " + }, + { + "bbox": [ + 132, + 114, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{hi}\\in \\mathbb{R}^{\\frac{\\hat{H}\\hat{W}}{M^2}\\times M^2\\times\\hat{C}}" + }, + { + "bbox": [ + 132, + 114, + 482, + 189 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 114, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_{hi}\\in \\mathbb{R}^{\\frac{\\hat{H}\\hat{W}}{M^2}\\times \\hat{C}\\times M^2}" + }, + { + "bbox": [ + 132, + 114, + 482, + 189 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 132, + 114, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{hi}\\in \\mathbb{R}^{\\frac{\\hat{H}\\hat{W}}{M^2}\\times M^2\\times\\hat{C}}" + }, + { + "bbox": [ + 132, + 114, + 482, + 189 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 132, + 114, + 482, + 189 + ], + "type": "inline_equation", + "content": "M^2" + }, + { + "bbox": [ + 132, + 114, + 482, + 189 + ], + "type": "text", + "content": " is the size of split windows. The attention matrix is thus calculated to tune the high-frequency prompt:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 225, + 199, + 481, + 214 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 199, + 481, + 214 + ], + "spans": [ + { + "bbox": [ + 225, + 199, + 481, + 214 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {h i} ^ {o u t} = \\mathbf {V} _ {h i} \\cdot \\mathrm {S o f t m a x} (\\mathbf {K} _ {h i} \\cdot \\mathbf {Q} _ {h i} / \\sqrt {d}), \\tag {5}", + "image_path": "81bc2a7a3847df9d9c98fda37ffaa2f15f529cda6447cfb7e0d233b82034abca.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 220, + 479, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 220, + 479, + 246 + ], + "spans": [ + { + "bbox": [ + 130, + 220, + 479, + 246 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 220, + 479, + 246 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{hi}^{out}\\in \\mathbb{R}^{\\hat{H}\\times \\hat{W}\\times \\hat{C}}" + }, + { + "bbox": [ + 130, + 220, + 479, + 246 + ], + "type": "text", + "content": " is the output feature map of the high-frequency prompt modulation branch; " + }, + { + "bbox": [ + 130, + 220, + 479, + 246 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 220, + 479, + 246 + ], + "type": "text", + "content": " is the query/ key dimension, following [33]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 246, + 480, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 246, + 480, + 320 + ], + "spans": [ + { + "bbox": [ + 130, + 246, + 480, + 320 + ], + "type": "text", + "content": "Low-frequency Prompt Modulator. Given the two input feature maps, including the " + }, + { + "bbox": [ + 130, + 246, + 480, + 320 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 246, + 480, + 320 + ], + "type": "text", + "content": "-level feature " + }, + { + "bbox": [ + 130, + 246, + 480, + 320 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_l \\in \\mathbb{R}^{\\hat{H} \\times \\hat{W} \\times \\hat{C}}" + }, + { + "bbox": [ + 130, + 246, + 480, + 320 + ], + "type": "text", + "content": " and low-frequency feature " + }, + { + "bbox": [ + 130, + 246, + 480, + 320 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{lo} \\in \\mathbb{R}^{H \\times W \\times C'}" + }, + { + "bbox": [ + 130, + 246, + 480, + 320 + ], + "type": "text", + "content": ", we first resize " + }, + { + "bbox": [ + 130, + 246, + 480, + 320 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{lo}" + }, + { + "bbox": [ + 130, + 246, + 480, + 320 + ], + "type": "text", + "content": " and obtain " + }, + { + "bbox": [ + 130, + 246, + 480, + 320 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{F}}_{lo} \\in \\mathbb{R}^{\\hat{H} \\times \\hat{W} \\times C'}" + }, + { + "bbox": [ + 130, + 246, + 480, + 320 + ], + "type": "text", + "content": ". Towards handling low-frequency signals effectively, we project " + }, + { + "bbox": [ + 130, + 246, + 480, + 320 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{F}}_{lo}" + }, + { + "bbox": [ + 130, + 246, + 480, + 320 + ], + "type": "text", + "content": " into the frequency domain via the fast Fourier transform (FFT). Then, a gating mechanism is adopted to control the useful low-frequency components flow forward:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 225, + 329, + 481, + 344 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 329, + 481, + 344 + ], + "spans": [ + { + "bbox": [ + 225, + 329, + 481, + 344 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {F}} _ {l o} = \\mathcal {F} (\\tilde {\\mathbf {F}} _ {l o}) \\odot \\sigma (\\operatorname {C o n v} _ {1 \\times 1} (\\mathcal {F} (\\tilde {\\mathbf {F}} _ {l o}))), \\tag {6}", + "image_path": "f2d8b3f74fb038abdaa551a14faaea2b1db2f31787cac2544821c34e948b5651.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 354, + 479, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 479, + 395 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 479, + 395 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 354, + 479, + 395 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_{lo}\\in \\mathbb{R}^{\\hat{H}\\times (\\frac{\\hat{W}}{2} +1)\\times 2C^{\\prime}}" + }, + { + "bbox": [ + 130, + 354, + 479, + 395 + ], + "type": "text", + "content": " is the processed feature, " + }, + { + "bbox": [ + 130, + 354, + 479, + 395 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\cdot)" + }, + { + "bbox": [ + 130, + 354, + 479, + 395 + ], + "type": "text", + "content": " represents the FFT. Next, we calibrate the input features by injecting learnable low-frequency prompt components " + }, + { + "bbox": [ + 130, + 354, + 479, + 395 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{lo}\\in \\mathbb{R}^{\\hat{H}\\times (\\frac{\\hat{W}}{2} +1)\\times 2C^{\\prime}}" + }, + { + "bbox": [ + 130, + 354, + 479, + 395 + ], + "type": "text", + "content": ", which is then transformed to spatial domain:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 247, + 406, + 481, + 421 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 406, + 481, + 421 + ], + "spans": [ + { + "bbox": [ + 247, + 406, + 481, + 421 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {l o} ^ {\\text {p r o m p t}} = \\mathcal {F} ^ {- 1} \\left(\\hat {\\mathbf {F}} _ {l o} \\odot \\mathbf {P} _ {l o}\\right), \\tag {7}", + "image_path": "5a745948cf9aab219978bb504542ebbfa5bdb043a4d6e65c70f05983f5fdf953.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 426, + 479, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 479, + 452 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 479, + 452 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 426, + 479, + 452 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{lo}^{prompt} \\in \\mathbb{R}^{\\hat{H} \\times \\hat{W} \\times C'}" + }, + { + "bbox": [ + 130, + 426, + 479, + 452 + ], + "type": "text", + "content": " is the generated low-frequency feature prompt, and " + }, + { + "bbox": [ + 130, + 426, + 479, + 452 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{-1}(\\cdot)" + }, + { + "bbox": [ + 130, + 426, + 479, + 452 + ], + "type": "text", + "content": " denotes the inverse FFT." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 453, + 480, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 453, + 480, + 513 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 480, + 513 + ], + "type": "text", + "content": "Noted, we perform the feature transformation in the Fourier domain for efficient global information interaction. The convolution theorem [46, 55] indicates the Hadamard product of two signals in the Fourier domain equals to implement the Fourier transform of a convolution of these two signals in the original spatial domain. Base on this insight, we can combine Eq.(6) and Eq.(7):" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 176, + 522, + 480, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 522, + 480, + 571 + ], + "spans": [ + { + "bbox": [ + 176, + 522, + 480, + 571 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {F} _ {l o} ^ {\\text {p r o m p t}} = \\mathcal {F} ^ {- 1} \\left(\\mathcal {F} (\\tilde {\\mathbf {F}} _ {l o}) \\odot \\sigma \\left(\\operatorname {C o n v} _ {1 \\times 1} \\left(\\mathcal {F} (\\tilde {\\mathbf {F}} _ {l o})\\right)\\right) \\odot \\mathbf {P} _ {l o}\\right) \\\\ = \\mathcal {F} ^ {- 1} \\left(\\mathcal {F} \\left(\\tilde {\\mathbf {F}} _ {l o} * \\mathcal {F} ^ {- 1} \\left(\\sigma \\left(\\operatorname {C o n v} _ {1 \\times 1} \\left(\\mathcal {F} \\left(\\tilde {\\mathbf {F}} _ {l o}\\right)\\right)\\right) \\odot \\mathbf {P} _ {l o}\\right)\\right)\\right) \\tag {8} \\\\ = \\tilde {\\mathbf {F}} _ {l o} * \\mathcal {F} ^ {- 1} \\left(\\sigma \\left(\\operatorname {C o n v} _ {1 \\times 1} \\left(\\mathcal {F} (\\tilde {\\mathbf {F}} _ {l o})\\right)\\right) \\odot \\mathbf {P} _ {l o}\\right) \\\\ \\end{array}", + "image_path": "27f71d39f3d189cb79c65494d50b0989cd34ad4e54f703eaa6d89997f16d6743.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 580, + 479, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 580, + 479, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 580, + 479, + 629 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 580, + 479, + 629 + ], + "type": "inline_equation", + "content": "\\circledast" + }, + { + "bbox": [ + 130, + 580, + 479, + 629 + ], + "type": "text", + "content": " is the convolution operation. Since " + }, + { + "bbox": [ + 130, + 580, + 479, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{-1}(\\sigma (\\mathrm{Conv}_{1\\times 1}(\\mathcal{F}(\\tilde{\\mathbf{F}}_{lo})))\\odot \\mathbf{P}_{lo})" + }, + { + "bbox": [ + 130, + 580, + 479, + 629 + ], + "type": "text", + "content": " is a tensor that shares the same shape with " + }, + { + "bbox": [ + 130, + 580, + 479, + 629 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{F}}_{lo}" + }, + { + "bbox": [ + 130, + 580, + 479, + 629 + ], + "type": "text", + "content": ", it can be served as a dynamic depth-wise convolution kernel as large as " + }, + { + "bbox": [ + 130, + 580, + 479, + 629 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{F}}_{lo}" + }, + { + "bbox": [ + 130, + 580, + 479, + 629 + ], + "type": "text", + "content": " in spatial domain while introducing less model complexity." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "Subsequently, we further modulate the low-frequency visual prompt " + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{lo}^{prompt}" + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": " with the guidance of the input feature " + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_l" + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": ". Specifically, we adopt an adaptive average pooling operator, which serves as a low-pass filter [65], to enhance the" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "text", + "content": "Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 139, + 136, + 474, + 178 + ], + "blocks": [ + { + "bbox": [ + 140, + 114, + 472, + 126 + ], + "lines": [ + { + "bbox": [ + 140, + 114, + 472, + 126 + ], + "spans": [ + { + "bbox": [ + 140, + 114, + 472, + 126 + ], + "type": "text", + "content": "Table 1: Quantitative comparison on the SPAD dataset [70] for image deraining." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 139, + 136, + 474, + 178 + ], + "lines": [ + { + "bbox": [ + 139, + 136, + 474, + 178 + ], + "spans": [ + { + "bbox": [ + 139, + 136, + 474, + 178 + ], + "type": "table", + "html": "
MethodDDN [16]PReNet [56]RCDNet [69]MPRNet [83]SPAIR [51]Uformer-S [72]SCD-Former [18]IDT [75]Restormer [81]DRSformer [7]FPro (Ours)
PSNR ↑36.1640.1643.3643.6444.1046.1346.8947.3447.9848.5348.99
SSIM ↑0.94630.98160.98310.98440.98720.99130.99410.99290.99210.99240.9936
", + "image_path": "6461215bd5e6a770e95d644a147b0fa7e7d2956ec771ab5b8871fbcf219b827e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 136, + 209, + 476, + 248 + ], + "blocks": [ + { + "bbox": [ + 140, + 188, + 472, + 200 + ], + "lines": [ + { + "bbox": [ + 140, + 188, + 472, + 200 + ], + "spans": [ + { + "bbox": [ + 140, + 188, + 472, + 200 + ], + "type": "text", + "content": "Table 2: Quantitative comparison on the AGAN-Data [52] for image deraindrop." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 209, + 476, + 248 + ], + "lines": [ + { + "bbox": [ + 136, + 209, + 476, + 248 + ], + "spans": [ + { + "bbox": [ + 136, + 209, + 476, + 248 + ], + "type": "table", + "html": "
MethodEigen's [15]Pix2pix [23]Uformer-S [72]WeatherDiff128 [47]TransWeather [63]DuRN [38]RaindropAttn [54]AttentiveGAN [52]IDT [75]Restormer [81]FPro (Ours)
PSNR ↑21.3128.0229.4229.6630.1731.2431.3731.5931.6331.6831.96
SSIM ↑0.7570.8550.9060.9230.9160.9260.9180.9170.9360.9340.937
", + "image_path": "94b349f7baad60840f9d62e6b58938b7e6ba03eaf265a9037451ad4720e6e0f1.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 270, + 482, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 373 + ], + "type": "text", + "content": "low-frequency content in the input " + }, + { + "bbox": [ + 130, + 270, + 482, + 373 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{lo}^{prompt}" + }, + { + "bbox": [ + 130, + 270, + 482, + 373 + ], + "type": "text", + "content": ". After that, we generate query " + }, + { + "bbox": [ + 130, + 270, + 482, + 373 + ], + "type": "inline_equation", + "content": "(\\mathbf{Q}_{lo})" + }, + { + "bbox": [ + 130, + 270, + 482, + 373 + ], + "type": "text", + "content": " projection from reshaped " + }, + { + "bbox": [ + 130, + 270, + 482, + 373 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_l" + }, + { + "bbox": [ + 130, + 270, + 482, + 373 + ], + "type": "text", + "content": ", key " + }, + { + "bbox": [ + 130, + 270, + 482, + 373 + ], + "type": "inline_equation", + "content": "(\\mathbf{K}_{lo})" + }, + { + "bbox": [ + 130, + 270, + 482, + 373 + ], + "type": "text", + "content": " and value " + }, + { + "bbox": [ + 130, + 270, + 482, + 373 + ], + "type": "inline_equation", + "content": "(\\mathbf{V}_{lo})" + }, + { + "bbox": [ + 130, + 270, + 482, + 373 + ], + "type": "text", + "content": " projections from the average-pooled feature " + }, + { + "bbox": [ + 130, + 270, + 482, + 373 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_{lo}^{prompt} = \\mathrm{AAP}(\\mathbf{F}_{lo}^{prompt})" + }, + { + "bbox": [ + 130, + 270, + 482, + 373 + ], + "type": "text", + "content": ", respectively. Here, " + }, + { + "bbox": [ + 130, + 270, + 482, + 373 + ], + "type": "inline_equation", + "content": "\\mathrm{AAP}(\\cdot)" + }, + { + "bbox": [ + 130, + 270, + 482, + 373 + ], + "type": "text", + "content": " means the adaptive average pooling operation. Then, we employ " + }, + { + "bbox": [ + 130, + 270, + 482, + 373 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 130, + 270, + 482, + 373 + ], + "type": "text", + "content": " convolutions to yield " + }, + { + "bbox": [ + 130, + 270, + 482, + 373 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{lo} = W_p^{Q_{lo}}\\mathbf{F}_l" + }, + { + "bbox": [ + 130, + 270, + 482, + 373 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 270, + 482, + 373 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_{lo} = W_p^{K_{lo}}\\hat{\\mathbf{F}}_{lo}^{prompt}" + }, + { + "bbox": [ + 130, + 270, + 482, + 373 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 270, + 482, + 373 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{lo} = W_p^{V_{lo}}\\hat{\\mathbf{F}}_{lo}^{prompt}" + }, + { + "bbox": [ + 130, + 270, + 482, + 373 + ], + "type": "text", + "content": ". Here " + }, + { + "bbox": [ + 130, + 270, + 482, + 373 + ], + "type": "inline_equation", + "content": "W_{p}^{(\\cdot)}" + }, + { + "bbox": [ + 130, + 270, + 482, + 373 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 130, + 270, + 482, + 373 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 130, + 270, + 482, + 373 + ], + "type": "text", + "content": " convolution. Next, we calculate the dot-product of query and key projections, which generates the transposed-attention " + }, + { + "bbox": [ + 130, + 270, + 482, + 373 + ], + "type": "inline_equation", + "content": "\\mathbf{A}\\in \\mathbb{R}^{\\hat{H}\\hat{W}\\times 1}" + }, + { + "bbox": [ + 130, + 270, + 482, + 373 + ], + "type": "text", + "content": ". Overall, the process of modulating the low-frequency prompt is defined as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 228, + 380, + 481, + 395 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 380, + 481, + 395 + ], + "spans": [ + { + "bbox": [ + 228, + 380, + 481, + 395 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {l o} ^ {o u t} = \\mathbf {V} _ {l o} \\cdot \\operatorname {S o f t m a x} (\\mathbf {K} _ {l o} \\cdot \\mathbf {Q} _ {l o} / \\alpha), \\tag {9}", + "image_path": "65c6ed5e4b751e42b2a120b13322af8dbc83fa6f58a2e5b4c10ff79ef2748ac3.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 405, + 480, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 405, + 480, + 444 + ], + "spans": [ + { + "bbox": [ + 130, + 405, + 480, + 444 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 405, + 480, + 444 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{lo}^{out} \\in \\mathbb{R}^{\\hat{H} \\times \\hat{W} \\times C'}" + }, + { + "bbox": [ + 130, + 405, + 480, + 444 + ], + "type": "text", + "content": " is the output feature map of the low-frequency prompt modulation branch; " + }, + { + "bbox": [ + 130, + 405, + 480, + 444 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{lo} \\in \\mathbb{R}^{\\hat{H}\\hat{W} \\times C'}" + }, + { + "bbox": [ + 130, + 405, + 480, + 444 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 405, + 480, + 444 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_{lo} \\in \\mathbb{R}^{C' \\times 1}" + }, + { + "bbox": [ + 130, + 405, + 480, + 444 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 405, + 480, + 444 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{lo} \\in \\mathbb{R}^{1 \\times C'}" + }, + { + "bbox": [ + 130, + 405, + 480, + 444 + ], + "type": "text", + "content": " are the input matrices; " + }, + { + "bbox": [ + 130, + 405, + 480, + 444 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 405, + 480, + 444 + ], + "type": "text", + "content": " is the learnable scaling parameter." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 445, + 481, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 445, + 481, + 481 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 481, + 481 + ], + "type": "text", + "content": "For both low/high-frequency modulators, we perform the attention map calculation several times in parallel, and these results are then concatenated for multi-head self-attention (MSA) [64]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 499, + 230, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 499, + 230, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 499, + 230, + 514 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 525, + 271, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 271, + 538 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 271, + 538 + ], + "type": "text", + "content": "4.1 Experimental settings" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 545, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 629 + ], + "type": "text", + "content": "Metrics. We adopt commonly used peak signal-to-noise ratio (PSNR) [73] and structural similarity (SSIM) metrics to evaluate restored images. Meanwhile, perceptual metric NIQE [42] is employed as a non-reference metric. Following previous works [69, 72], PSNR/SSIM computations are implemented on the Y channel in the YCbCr space for the image deraining task, while calculated in the RGB color space for other restoration tasks. In the reported tables, the best and second-best scores are highlighted and underlined, respectively." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "Implementation Details. FPro contains " + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "inline_equation", + "content": "N_{1} = 3" + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": " levels encoder-decoder, where the encode and decoder share the same " + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "inline_equation", + "content": "N_{2} = [2,3,6]" + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": " blocks. We set embedding dimensions " + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": " as 48, and the attention heads as [2,4,8]. The expanding channel" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 184, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 184, + 91, + 447, + 103 + ], + "type": "text", + "content": "A Frequency Prompt Guided Transformer for Image Restoration" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 135, + 481, + 180 + ], + "blocks": [ + { + "bbox": [ + 152, + 114, + 460, + 126 + ], + "lines": [ + { + "bbox": [ + 152, + 114, + 460, + 126 + ], + "spans": [ + { + "bbox": [ + 152, + 114, + 460, + 126 + ], + "type": "text", + "content": "Table 3: Quantitative comparison on the TIP-2018 [61] for image demoiré." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 135, + 481, + 180 + ], + "lines": [ + { + "bbox": [ + 135, + 135, + 481, + 180 + ], + "spans": [ + { + "bbox": [ + 135, + 135, + 481, + 180 + ], + "type": "table", + "html": "
MethodAMNet [80]DMCNN [61]UNet [59]WDNet [36]MopNet [19]TAPE-Net [37]FHD2eNet [20]MBCNN [86]Uformer-S [72]Wang et al. [66]FPro (Ours)
PSNR ↑25.4726.1026.4927.1227.4827.5227.7928.4028.6328.8729.25
SSIM ↑0.8330.8440.8640.8540.8610.8660.8670.8710.8720.8940.879
", + "image_path": "cb8c158c2627eb421fa411a46f7657705d58e20d1e2fb0997b2fb008d51d22f1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 140, + 213, + 474, + 265 + ], + "blocks": [ + { + "bbox": [ + 157, + 191, + 455, + 204 + ], + "lines": [ + { + "bbox": [ + 157, + 191, + 455, + 204 + ], + "spans": [ + { + "bbox": [ + 157, + 191, + 455, + 204 + ], + "type": "text", + "content": "Table 4: Quantitative comparison on the SOTS [29] for image dehazing." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 140, + 213, + 474, + 265 + ], + "lines": [ + { + "bbox": [ + 140, + 213, + 474, + 265 + ], + "spans": [ + { + "bbox": [ + 140, + 213, + 474, + 265 + ], + "type": "table", + "html": "
MethodAODNet [28]MSCNN [57]DehazeNet [2]EPDN [53]FDGAN [12]AirNet [30]Restormer [81]PromptIR [50]FPro (Ours)
PSNR ↑20.2922.0622.4622.5723.1523.1830.8731.3132.85
SSIM ↑0.8770.9080.8510.8630.9210.9000.9690.9730.983
", + "image_path": "494cb6ae4c5d792e32c9176fbc38757df7fed86b79f2888e12002246ae893eb7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 287, + 482, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 287, + 482, + 360 + ], + "spans": [ + { + "bbox": [ + 130, + 287, + 482, + 360 + ], + "type": "text", + "content": "capacity factor in FFN is 3. The default split window size in HPM is set as " + }, + { + "bbox": [ + 130, + 287, + 482, + 360 + ], + "type": "inline_equation", + "content": "M = 8" + }, + { + "bbox": [ + 130, + 287, + 482, + 360 + ], + "type": "text", + "content": ". The pixel-unshuffle and pixel-shuffle are employed for downsampling and upsampling. We use the AdamW optimizer with an initial learning rate of " + }, + { + "bbox": [ + 130, + 287, + 482, + 360 + ], + "type": "inline_equation", + "content": "3 \\times 10^{-4}" + }, + { + "bbox": [ + 130, + 287, + 482, + 360 + ], + "type": "text", + "content": ", which is gradually reduced to " + }, + { + "bbox": [ + 130, + 287, + 482, + 360 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-6}" + }, + { + "bbox": [ + 130, + 287, + 482, + 360 + ], + "type": "text", + "content": " using cosine annealing, to train FPro. Additionally, we adopt the widely used loss function [68] to constrain the network training." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 380, + 228, + 390 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 228, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 228, + 390 + ], + "type": "text", + "content": "4.2 Main Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 402, + 482, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 498 + ], + "type": "text", + "content": "Rain Streak Removal. The proposed FPro is compared with general restoration approaches [51,72,81,83] as well as with task-specific methods [7,16,18,56,69, 75]. Tab. 1 shows that FPro makes superior performance over current methods for real image deraining on SPAD [70]. Compared to the previous best approach DRSformer [7], FPro achieves a " + }, + { + "bbox": [ + 130, + 402, + 482, + 498 + ], + "type": "inline_equation", + "content": "0.46\\mathrm{dB}" + }, + { + "bbox": [ + 130, + 402, + 482, + 498 + ], + "type": "text", + "content": " performance boost. In addition, FPro obtains 2.1 dB PSNR improvement when compared to the recent model SCD-Former [18]. Fig. 4 provides a visual deraining example, where FPro successfully removes the rain degradation while maintaining the structural information." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 498, + 482, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 583 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 583 + ], + "type": "text", + "content": "Raindrop Removal. For image deraindrop, we compare FPro with some existing deraindrop methods, including Eigen's [15], Pix2pix [23], TransWeather [63], Uformer [72], WeatherDiff128 [47], DuRN [38], RaindropAttn [54], AttentiveGAN [52], IDT [75], and Restormer [81]. We report the quantitative results on AGAN-Data [52] in Tab. 2. Our FPro obtains the best performance against all considered methods in terms of both PSNR/SSIM scores. FPro makes a 0.28 dB PSNR performance boost over Restormer [81], and 2.3 dB over the recent" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 137, + 604, + 193, + 642 + ], + "blocks": [ + { + "bbox": [ + 137, + 604, + 193, + 642 + ], + "lines": [ + { + "bbox": [ + 137, + 604, + 193, + 642 + ], + "spans": [ + { + "bbox": [ + 137, + 604, + 193, + 642 + ], + "type": "image", + "image_path": "e137bb23d466f951816bff5f6533476ebe6652864beee47d56c987eeca1cc09d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 152, + 651, + 459, + 662 + ], + "lines": [ + { + "bbox": [ + 152, + 651, + 459, + 662 + ], + "spans": [ + { + "bbox": [ + 152, + 651, + 459, + 662 + ], + "type": "text", + "content": "Fig. 4: Qualitative comparisons on the SPAD [70] for real image deraining." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 195, + 604, + 250, + 642 + ], + "blocks": [ + { + "bbox": [ + 195, + 604, + 250, + 642 + ], + "lines": [ + { + "bbox": [ + 195, + 604, + 250, + 642 + ], + "spans": [ + { + "bbox": [ + 195, + 604, + 250, + 642 + ], + "type": "image", + "image_path": "67ecf5d624dc5ba28359dbef46e6a72737710bcb3768954392dd474b3db599ba.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 251, + 604, + 306, + 642 + ], + "blocks": [ + { + "bbox": [ + 251, + 604, + 306, + 642 + ], + "lines": [ + { + "bbox": [ + 251, + 604, + 306, + 642 + ], + "spans": [ + { + "bbox": [ + 251, + 604, + 306, + 642 + ], + "type": "image", + "image_path": "369d8fd845dcb1c32152af2d7f33a40a62da76979e6540fe4579b8f1a7d79da6.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 307, + 604, + 362, + 642 + ], + "blocks": [ + { + "bbox": [ + 307, + 604, + 362, + 642 + ], + "lines": [ + { + "bbox": [ + 307, + 604, + 362, + 642 + ], + "spans": [ + { + "bbox": [ + 307, + 604, + 362, + 642 + ], + "type": "image", + "image_path": "483381e9ea8a98f27a7551e3a41af775f8a0abb22be36be18a631065219c1e9c.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 364, + 604, + 420, + 642 + ], + "blocks": [ + { + "bbox": [ + 364, + 604, + 420, + 642 + ], + "lines": [ + { + "bbox": [ + 364, + 604, + 420, + 642 + ], + "spans": [ + { + "bbox": [ + 364, + 604, + 420, + 642 + ], + "type": "image", + "image_path": "8d1163c8fd5fe63bbf97b9f2413a0b181c40c42e5dc6cd0675f38b3f0f325b36.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 421, + 604, + 476, + 642 + ], + "blocks": [ + { + "bbox": [ + 421, + 604, + 476, + 642 + ], + "lines": [ + { + "bbox": [ + 421, + 604, + 476, + 642 + ], + "spans": [ + { + "bbox": [ + 421, + 604, + 476, + 642 + ], + "type": "image", + "image_path": "907a70c72d658cd9a11cb62750eb0f88e8575bb447bae85dfacafe02945462a8.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 140, + 136, + 474, + 188 + ], + "blocks": [ + { + "bbox": [ + 153, + 114, + 459, + 126 + ], + "lines": [ + { + "bbox": [ + 153, + 114, + 459, + 126 + ], + "spans": [ + { + "bbox": [ + 153, + 114, + 459, + 126 + ], + "type": "text", + "content": "Table 5: Quantitative comparison on the GoPro [43] for image deblurring." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 140, + 136, + 474, + 188 + ], + "lines": [ + { + "bbox": [ + 140, + 136, + 474, + 188 + ], + "spans": [ + { + "bbox": [ + 140, + 136, + 474, + 188 + ], + "type": "table", + "html": "
MethodCODE [85]IPT [4]MPRNet [83]MIMO [9]HINet [6]MAXIM [62]Restormer [81]PromptRestorer [68]FPro (Ours)
PSNR ↑31.9432.5832.6632.6832.7132.8632.9233.0633.05
SSIM ↑0.928-0.9590.9590.9590.9610.9610.9620.961
", + "image_path": "512b4edc48053fff22afab0915d1fac941980e4d2641aa936269f659d5f6d02d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 143, + 220, + 299, + 289 + ], + "blocks": [ + { + "bbox": [ + 154, + 198, + 285, + 209 + ], + "lines": [ + { + "bbox": [ + 154, + 198, + 285, + 209 + ], + "spans": [ + { + "bbox": [ + 154, + 198, + 285, + 209 + ], + "type": "text", + "content": "Table 6: Effectiveness of GDD." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 143, + 220, + 299, + 289 + ], + "lines": [ + { + "bbox": [ + 143, + 220, + 299, + 289 + ], + "spans": [ + { + "bbox": [ + 143, + 220, + 299, + 289 + ], + "type": "table", + "html": "
ModelsPSNRSSIM
(a) w/o prompt branch48.400.9920
(b) Multi DC [8]48.520.9926
(c) Multi GDD48.910.9934
(d) Single GDD48.990.9936
", + "image_path": "dc49230e5812839ddc4f2f5d85d7304305fc5fcf28221b3d3e36456ee9ec5dc9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 318, + 220, + 474, + 289 + ], + "blocks": [ + { + "bbox": [ + 324, + 198, + 463, + 209 + ], + "lines": [ + { + "bbox": [ + 324, + 198, + 463, + 209 + ], + "spans": [ + { + "bbox": [ + 324, + 198, + 463, + 209 + ], + "type": "text", + "content": "Table 7: Ablation study of DPB." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 220, + 474, + 289 + ], + "lines": [ + { + "bbox": [ + 318, + 220, + 474, + 289 + ], + "spans": [ + { + "bbox": [ + 318, + 220, + 474, + 289 + ], + "type": "table", + "html": "
ModelsPSNRSSIM
(a) w/o prompt branch48.400.9920
(b) w/o HPM48.770.9931
(c) w/o LPM48.890.9933
(d) Full48.990.9936
", + "image_path": "01a30543c1dfd7a84cbe05a8a15b6460787b21a5faf5ab392275b91a2e37d5d8.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 312, + 480, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 480, + 335 + ], + "type": "text", + "content": "method WeatherDiff" + }, + { + "bbox": [ + 130, + 312, + 480, + 335 + ], + "type": "inline_equation", + "content": "_{128}" + }, + { + "bbox": [ + 130, + 312, + 480, + 335 + ], + "type": "text", + "content": " [47]. Fig. 5 shows the visual comparisons, where FPro generates a result with finer details." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 336, + 481, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 336, + 481, + 408 + ], + "spans": [ + { + "bbox": [ + 130, + 336, + 481, + 408 + ], + "type": "text", + "content": "Moiré pattern Removal. We perform image demoiring experiments on TIP-2018 [61] dataset, and compare FPro with ten demoiring methods, including AMNet [80], DMCNN [61], UNet [59], WDNet [36], MopNet [19], TAPE-Net [37], " + }, + { + "bbox": [ + 130, + 336, + 481, + 408 + ], + "type": "inline_equation", + "content": "\\mathrm{FHD}^2\\mathrm{eNet}" + }, + { + "bbox": [ + 130, + 336, + 481, + 408 + ], + "type": "text", + "content": " [20], MBCNN [86], Uformer-S [72], and Wang et al. [66]. In Tab. 3, FPro yields a 0.38 performance boost against the previous best method Wang et al. [66], and outperforms the recent model TAPE-Net [37] by 1.73 dB." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 408, + 482, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 408, + 482, + 480 + ], + "spans": [ + { + "bbox": [ + 130, + 408, + 482, + 480 + ], + "type": "text", + "content": "Haze Removal. We perform image dehazing experiments on SOTS [29] benchmark. We compare FPro with eight representative approaches, including AOD-Net [28], MSCNN [57], DehazeNet [2], EPDN [53], FDGAN [12], AirNet [30], Restormer [81], and PromptIR [50]. As shown in Tab. 4, FPro achieves the best scores among all considered methods. Compared to the recent prompt-based method PromptIR [50], FPro makes a substantial performance gain of " + }, + { + "bbox": [ + 130, + 408, + 482, + 480 + ], + "type": "inline_equation", + "content": "1.54\\mathrm{dB}" + }, + { + "bbox": [ + 130, + 408, + 482, + 480 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 481, + 482, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 481, + 482, + 576 + ], + "spans": [ + { + "bbox": [ + 130, + 481, + 482, + 576 + ], + "type": "text", + "content": "Motion Blur Removal. We evaluate image deblurring performance on the GoPro dataset [43]. For synthetic deblurring, we compare FPro with eight representative models: CODE [85], IPT [4], MPRNet [83], MIMO [9], HINet [6], MAXIM [62], Restormer [81] and PromptRestorer [68]. Tab. 5 shows that our FPro achieves competitive performance against recent prompt-based PromptRestorer [68], where we cost half FLOPs. Meanwhile, compared to the recent method CODE [85], FPro obtains a 1.11 dB gain on PSNR while using lower FLOPs." + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 137, + 596, + 194, + 634 + ], + "blocks": [ + { + "bbox": [ + 137, + 596, + 194, + 634 + ], + "lines": [ + { + "bbox": [ + 137, + 596, + 194, + 634 + ], + "spans": [ + { + "bbox": [ + 137, + 596, + 194, + 634 + ], + "type": "image", + "image_path": "40a4d6f437ca090cac087e76357ab59c2495a4b0ef498302a190cabb0afe0d5d.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 149, + 634, + 179, + 641 + ], + "lines": [ + { + "bbox": [ + 149, + 634, + 179, + 641 + ], + "spans": [ + { + "bbox": [ + 149, + 634, + 179, + 641 + ], + "type": "text", + "content": "Raindrop" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 145, + 651, + 466, + 662 + ], + "lines": [ + { + "bbox": [ + 145, + 651, + 466, + 662 + ], + "spans": [ + { + "bbox": [ + 145, + 651, + 466, + 662 + ], + "type": "text", + "content": "Fig. 5: Qualitative comparisons on the AGAN-Data [52] for image deraindrop." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 194, + 596, + 249, + 633 + ], + "blocks": [ + { + "bbox": [ + 194, + 596, + 249, + 633 + ], + "lines": [ + { + "bbox": [ + 194, + 596, + 249, + 633 + ], + "spans": [ + { + "bbox": [ + 194, + 596, + 249, + 633 + ], + "type": "image", + "image_path": "a90c5281145716aba555285ff88823fbc5f98eb2a0ed910b8fc97ea83014f18f.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 206, + 634, + 236, + 641 + ], + "lines": [ + { + "bbox": [ + 206, + 634, + 236, + 641 + ], + "spans": [ + { + "bbox": [ + 206, + 634, + 236, + 641 + ], + "type": "text", + "content": "Reference" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 251, + 596, + 306, + 633 + ], + "blocks": [ + { + "bbox": [ + 251, + 596, + 306, + 633 + ], + "lines": [ + { + "bbox": [ + 251, + 596, + 306, + 633 + ], + "spans": [ + { + "bbox": [ + 251, + 596, + 306, + 633 + ], + "type": "image", + "image_path": "33f3ac49e2f3876c24abd9196ee934784fecc7cf6e7a5aaa0014ab908c824ff8.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 251, + 634, + 306, + 641 + ], + "lines": [ + { + "bbox": [ + 251, + 634, + 306, + 641 + ], + "spans": [ + { + "bbox": [ + 251, + 634, + 306, + 641 + ], + "type": "text", + "content": "RaindropAttn [54]" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 309, + 596, + 363, + 633 + ], + "blocks": [ + { + "bbox": [ + 309, + 596, + 363, + 633 + ], + "lines": [ + { + "bbox": [ + 309, + 596, + 363, + 633 + ], + "spans": [ + { + "bbox": [ + 309, + 596, + 363, + 633 + ], + "type": "image", + "image_path": "9a84dd51f770445bd79e7c20fbd51d31af229295d548a65e15cf3d1657f25ae5.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 316, + 634, + 354, + 641 + ], + "lines": [ + { + "bbox": [ + 316, + 634, + 354, + 641 + ], + "spans": [ + { + "bbox": [ + 316, + 634, + 354, + 641 + ], + "type": "text", + "content": "Uformer [72]" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 365, + 596, + 419, + 633 + ], + "blocks": [ + { + "bbox": [ + 365, + 596, + 419, + 633 + ], + "lines": [ + { + "bbox": [ + 365, + 596, + 419, + 633 + ], + "spans": [ + { + "bbox": [ + 365, + 596, + 419, + 633 + ], + "type": "image", + "image_path": "6e2a4fab8a559a41699158da474d935fa5baea64089aee7ba4222fec0e4e972b.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 370, + 634, + 414, + 641 + ], + "lines": [ + { + "bbox": [ + 370, + 634, + 414, + 641 + ], + "spans": [ + { + "bbox": [ + 370, + 634, + 414, + 641 + ], + "type": "text", + "content": "Restormer [81]" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 422, + 596, + 476, + 633 + ], + "blocks": [ + { + "bbox": [ + 422, + 596, + 476, + 633 + ], + "lines": [ + { + "bbox": [ + 422, + 596, + 476, + 633 + ], + "spans": [ + { + "bbox": [ + 422, + 596, + 476, + 633 + ], + "type": "image", + "image_path": "0e3a319dcef59ea91c1fb62d49003746c19193733ef0cb88b1ee53b840a9a2c5.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 441, + 634, + 457, + 641 + ], + "lines": [ + { + "bbox": [ + 441, + 634, + 457, + 641 + ], + "spans": [ + { + "bbox": [ + 441, + 634, + 457, + 641 + ], + "type": "text", + "content": "FPro" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 184, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 184, + 91, + 447, + 102 + ], + "type": "text", + "content": "A Frequency Prompt Guided Transformer for Image Restoration" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 282, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 282, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 282, + 128 + ], + "type": "text", + "content": "4.3 Analysis and Discussion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 479, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 479, + 174 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 479, + 174 + ], + "type": "text", + "content": "For the ablation studies, we study different models for rain streak removal on SPAD [70] with " + }, + { + "bbox": [ + 130, + 139, + 479, + 174 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 130, + 139, + 479, + 174 + ], + "type": "text", + "content": " patches for 300K iterations. Testing is conducted on SPAD testing dataset [70]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 175, + 480, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 175, + 480, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 175, + 480, + 282 + ], + "type": "text", + "content": "Effectiveness of Gated Dynamic Decoupler. To demonstrate the effectiveness of the Gated Dynamic Decoupler, we conduct experiments on different model variants in Tab. 6. Compared to the model equipped with Multiple Dynamic Convolution [8] (DC) for separating different frequency parts (b), directly replacing it with GDD (c) results in a performance gain of " + }, + { + "bbox": [ + 130, + 175, + 480, + 282 + ], + "type": "inline_equation", + "content": "0.39\\mathrm{dB}" + }, + { + "bbox": [ + 130, + 175, + 480, + 282 + ], + "type": "text", + "content": " in terms of PSNR. Meanwhile, instead of injecting GDD into each DPB (c) to employ multiple decouplers, we attempt to share one GDD module to divide the low-/high frequency information (d), which slightly reduces the complexity " + }, + { + "bbox": [ + 130, + 175, + 480, + 282 + ], + "type": "inline_equation", + "content": "(0.02\\mathrm{M})" + }, + { + "bbox": [ + 130, + 175, + 480, + 282 + ], + "type": "text", + "content": " of the whole framework and brings a " + }, + { + "bbox": [ + 130, + 175, + 480, + 282 + ], + "type": "inline_equation", + "content": "0.08\\mathrm{dB}" + }, + { + "bbox": [ + 130, + 175, + 480, + 282 + ], + "type": "text", + "content": " performance boost." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 284, + 480, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 284, + 480, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 284, + 480, + 415 + ], + "type": "text", + "content": "Effectiveness of Dual Propmt Block. To investigate the developed DPB, ablation studies are performed in Tab. 7. Disabling HPM or LPM leads to a clear drop of " + }, + { + "bbox": [ + 130, + 284, + 480, + 415 + ], + "type": "inline_equation", + "content": "0.22\\mathrm{dB}" + }, + { + "bbox": [ + 130, + 284, + 480, + 415 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 284, + 480, + 415 + ], + "type": "inline_equation", + "content": "0.1\\mathrm{dB}" + }, + { + "bbox": [ + 130, + 284, + 480, + 415 + ], + "type": "text", + "content": ", respectively. These experimental results indicate that both HPM and LPM play a positive role in restoring high-quality images. Moreover, we present visualizations to better show the effect of DPB. As shown in Fig. 7, we visualize the generated low-/high-frequency feature maps from each branch along with the analysis in the Fourier domain, where the low-frequency prompt feature encodes information such as structures while the high-frequency prompt one focuses on information such as edges and texture. Meantime, we provide visual comparisons in Fig. 8 to show the effectiveness of the proposed HPM/LPM." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 416, + 481, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 416, + 481, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 416, + 481, + 499 + ], + "type": "text", + "content": "Perceptual Quality Assessment. To test the perceptual quality of the proposed FPro, following [7], we randomly choose 20 rainy images under real-world scenes from Internet-Data [70] to perform the evaluation. As shown in Tab. 9, compared to other considered methods, FPro achieves a lower NIQE score, which means the generated results contain clearer content and better perceptual quality. Through qualitative comparison in Fig. 6, FPro obtains a visually pleasant result against other models, indicating that it handles unseen degradation well." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 500, + 480, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 500, + 480, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 500, + 480, + 559 + ], + "type": "text", + "content": "Model Efficiency. We provide the comparison of performance (PSNR), complexity (FLOPs and Parameters), and latency (Run-times) for image deraining. FLOPs and Runtimes are measured when input with the size of " + }, + { + "bbox": [ + 130, + 500, + 480, + 559 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 130, + 500, + 480, + 559 + ], + "type": "text", + "content": ", and PSNR scores are tested on SPAD [70]. As shown in Tab. 8, though FPro achieves better performance in terms of PSNR metric, it has less model complexity than" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 136, + 582, + 192, + 622 + ], + "blocks": [ + { + "bbox": [ + 136, + 582, + 192, + 622 + ], + "lines": [ + { + "bbox": [ + 136, + 582, + 192, + 622 + ], + "spans": [ + { + "bbox": [ + 136, + 582, + 192, + 622 + ], + "type": "image", + "image_path": "622a81c1edf129dc614b540540abb783051853d29afe715c83c1d61d78cb430b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 155, + 624, + 173, + 630 + ], + "lines": [ + { + "bbox": [ + 155, + 624, + 173, + 630 + ], + "spans": [ + { + "bbox": [ + 155, + 624, + 173, + 630 + ], + "type": "text", + "content": "Rainy" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 194, + 582, + 251, + 623 + ], + "blocks": [ + { + "bbox": [ + 194, + 582, + 251, + 623 + ], + "lines": [ + { + "bbox": [ + 194, + 582, + 251, + 623 + ], + "spans": [ + { + "bbox": [ + 194, + 582, + 251, + 623 + ], + "type": "image", + "image_path": "8c08f6bd07e0097d3f98e6939644e6ac2d07d80b2ffa0185fc9ede4c8b6c71af.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 200, + 624, + 244, + 630 + ], + "lines": [ + { + "bbox": [ + 200, + 624, + 244, + 630 + ], + "spans": [ + { + "bbox": [ + 200, + 624, + 244, + 630 + ], + "type": "text", + "content": "Restormer [81]" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 639, + 479, + 662 + ], + "lines": [ + { + "bbox": [ + 130, + 639, + 479, + 662 + ], + "spans": [ + { + "bbox": [ + 130, + 639, + 479, + 662 + ], + "type": "text", + "content": "Fig.6: Qualitative comparisons with state-of-the-art methods on Internet-Data [70] for real rain removal. (Zoom in for a better view.)" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 253, + 582, + 309, + 623 + ], + "blocks": [ + { + "bbox": [ + 253, + 582, + 309, + 623 + ], + "lines": [ + { + "bbox": [ + 253, + 582, + 309, + 623 + ], + "spans": [ + { + "bbox": [ + 253, + 582, + 309, + 623 + ], + "type": "image", + "image_path": "a6831f37ff6f2669140cf55022036189844a67d5b00590ae30b3eec40bb333d9.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 261, + 624, + 299, + 630 + ], + "lines": [ + { + "bbox": [ + 261, + 624, + 299, + 630 + ], + "spans": [ + { + "bbox": [ + 261, + 624, + 299, + 630 + ], + "type": "text", + "content": "Uformer [72]" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 310, + 582, + 367, + 623 + ], + "blocks": [ + { + "bbox": [ + 310, + 582, + 367, + 623 + ], + "lines": [ + { + "bbox": [ + 310, + 582, + 367, + 623 + ], + "spans": [ + { + "bbox": [ + 310, + 582, + 367, + 623 + ], + "type": "image", + "image_path": "d53d1fe432d337afdbae48eb39480b493db010fea9c8f69b48ce6a179c8c5b6e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 325, + 624, + 351, + 630 + ], + "lines": [ + { + "bbox": [ + 325, + 624, + 351, + 630 + ], + "spans": [ + { + "bbox": [ + 325, + 624, + 351, + 630 + ], + "type": "text", + "content": "IDT [75]" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 369, + 582, + 425, + 623 + ], + "blocks": [ + { + "bbox": [ + 369, + 582, + 425, + 623 + ], + "lines": [ + { + "bbox": [ + 369, + 582, + 425, + 623 + ], + "spans": [ + { + "bbox": [ + 369, + 582, + 425, + 623 + ], + "type": "image", + "image_path": "b158043eeccd22d548cb11f36f0adfa8091e44abbe93557bbbdb84de76dd7d25.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 375, + 624, + 419, + 630 + ], + "lines": [ + { + "bbox": [ + 375, + 624, + 419, + 630 + ], + "spans": [ + { + "bbox": [ + 375, + 624, + 419, + 630 + ], + "type": "text", + "content": "DRSformer [7]" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 427, + 582, + 483, + 623 + ], + "blocks": [ + { + "bbox": [ + 427, + 582, + 483, + 623 + ], + "lines": [ + { + "bbox": [ + 427, + 582, + 483, + 623 + ], + "spans": [ + { + "bbox": [ + 427, + 582, + 483, + 623 + ], + "type": "image", + "image_path": "a7dfed3a940587bed00847112344ffa4e57d3b510543ac28aa2e58dccc9184e7.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 447, + 624, + 463, + 630 + ], + "lines": [ + { + "bbox": [ + 447, + 624, + 463, + 630 + ], + "spans": [ + { + "bbox": [ + 447, + 624, + 463, + 630 + ], + "type": "text", + "content": "FPro" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 189, + 167 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 189, + 167 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 189, + 167 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 189, + 167 + ], + "type": "image", + "image_path": "57af3348afa766a14815e89fc8b9036059a92527ad826c2bb86fc75c11e4c203.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 192, + 114, + 245, + 167 + ], + "blocks": [ + { + "bbox": [ + 192, + 114, + 245, + 167 + ], + "lines": [ + { + "bbox": [ + 192, + 114, + 245, + 167 + ], + "spans": [ + { + "bbox": [ + 192, + 114, + 245, + 167 + ], + "type": "image", + "image_path": "4f8ba90a5cc279e9f070880d31ab5c65b9b352eba6311dd87553d0ac737e4126.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 136, + 172, + 189, + 224 + ], + "blocks": [ + { + "bbox": [ + 136, + 172, + 189, + 224 + ], + "lines": [ + { + "bbox": [ + 136, + 172, + 189, + 224 + ], + "spans": [ + { + "bbox": [ + 136, + 172, + 189, + 224 + ], + "type": "image", + "image_path": "de544173de0447c58b119a3f87ce1ae4c5a24e4f0a8ae7bfe3263e4ac76fe934.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 136, + 224, + 190, + 232 + ], + "lines": [ + { + "bbox": [ + 136, + 224, + 190, + 232 + ], + "spans": [ + { + "bbox": [ + 136, + 224, + 190, + 232 + ], + "type": "text", + "content": "(a) Low-frequency" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 241, + 246, + 350 + ], + "lines": [ + { + "bbox": [ + 132, + 241, + 246, + 350 + ], + "spans": [ + { + "bbox": [ + 132, + 241, + 246, + 350 + ], + "type": "text", + "content": "Fig. 7: Feature analysis. We visualize the features from the LPM branch (a), and the HPM one (b). In the right-bottom, we show the results of the average features over the channel dimension in the Fourier domain. (Zoom in for a better view.)" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 192, + 173, + 244, + 224 + ], + "blocks": [ + { + "bbox": [ + 192, + 173, + 244, + 224 + ], + "lines": [ + { + "bbox": [ + 192, + 173, + 244, + 224 + ], + "spans": [ + { + "bbox": [ + 192, + 173, + 244, + 224 + ], + "type": "image", + "image_path": "34b08a68a9c0f83fc545ebf8119f603bca049f0c2a9551d81f0151e411e8392c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 192, + 224, + 246, + 232 + ], + "lines": [ + { + "bbox": [ + 192, + 224, + 246, + 232 + ], + "spans": [ + { + "bbox": [ + 192, + 224, + 246, + 232 + ], + "type": "text", + "content": "(b) High-frequency" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 263, + 114, + 314, + 166 + ], + "blocks": [ + { + "bbox": [ + 263, + 114, + 314, + 166 + ], + "lines": [ + { + "bbox": [ + 263, + 114, + 314, + 166 + ], + "spans": [ + { + "bbox": [ + 263, + 114, + 314, + 166 + ], + "type": "image", + "image_path": "277c1ccfb0ce356cd9d8e8b066edd0183de4b28225b9fac6cd2208bfed7e9bc2.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 115, + 367, + 166 + ], + "blocks": [ + { + "bbox": [ + 316, + 115, + 367, + 166 + ], + "lines": [ + { + "bbox": [ + 316, + 115, + 367, + 166 + ], + "spans": [ + { + "bbox": [ + 316, + 115, + 367, + 166 + ], + "type": "image", + "image_path": "5436f3a955f6ebb47d0473590704236ec63268e5ef087bb88242c63610295d7b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 369, + 115, + 419, + 166 + ], + "blocks": [ + { + "bbox": [ + 369, + 115, + 419, + 166 + ], + "lines": [ + { + "bbox": [ + 369, + 115, + 419, + 166 + ], + "spans": [ + { + "bbox": [ + 369, + 115, + 419, + 166 + ], + "type": "image", + "image_path": "d793089ce4bc8969d46fe3be9242a946ad5fb13c142acef0bb52d34b958e4086.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 421, + 115, + 473, + 166 + ], + "blocks": [ + { + "bbox": [ + 421, + 115, + 473, + 166 + ], + "lines": [ + { + "bbox": [ + 421, + 115, + 473, + 166 + ], + "spans": [ + { + "bbox": [ + 421, + 115, + 473, + 166 + ], + "type": "image", + "image_path": "1b268fdb2ef23b77f007d88c9bc908dbf45bea54b68fd6da089976e86818d1c4.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 263, + 173, + 315, + 224 + ], + "blocks": [ + { + "bbox": [ + 268, + 167, + 309, + 173 + ], + "lines": [ + { + "bbox": [ + 268, + 167, + 309, + 173 + ], + "spans": [ + { + "bbox": [ + 268, + 167, + 309, + 173 + ], + "type": "text", + "content": "(a) w/o LPM" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 263, + 173, + 315, + 224 + ], + "lines": [ + { + "bbox": [ + 263, + 173, + 315, + 224 + ], + "spans": [ + { + "bbox": [ + 263, + 173, + 315, + 224 + ], + "type": "image", + "image_path": "4fa86cc22f5a89a9583bdb53ebd09039774265983adf3bbec43d06b54f39f78a.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 271, + 224, + 307, + 231 + ], + "lines": [ + { + "bbox": [ + 271, + 224, + 307, + 231 + ], + "spans": [ + { + "bbox": [ + 271, + 224, + 307, + 231 + ], + "type": "text", + "content": "(e) w/LPM" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 254, + 240, + 480, + 350 + ], + "lines": [ + { + "bbox": [ + 254, + 240, + 480, + 350 + ], + "spans": [ + { + "bbox": [ + 254, + 240, + 480, + 350 + ], + "type": "text", + "content": "Fig. 8: Effect of DBP. Columns 1 and 3 show low-pass and high-pass filtered results, while columns 2 and 4 show the difference (Diff.) between processed results with corresponding filtered ground-truth. Compared with (a), FPro w/ LPM (e) performs better in capturing information such as structures, resulting in fewer erroneous predictions (f). Compared with (c), FPro w/ HPM (g) restores clear edges and shapes, which indicates it enjoys the benefits from the high-frequency information prompt. (Zoom in for a better view.)" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 316, + 173, + 368, + 224 + ], + "blocks": [ + { + "bbox": [ + 329, + 167, + 353, + 173 + ], + "lines": [ + { + "bbox": [ + 329, + 167, + 353, + 173 + ], + "spans": [ + { + "bbox": [ + 329, + 167, + 353, + 173 + ], + "type": "text", + "content": "(b) Diff." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 316, + 173, + 368, + 224 + ], + "lines": [ + { + "bbox": [ + 316, + 173, + 368, + 224 + ], + "spans": [ + { + "bbox": [ + 316, + 173, + 368, + 224 + ], + "type": "image", + "image_path": "7802302ae02ba88c4b3b86499871b0cc829366fa220a0bb25e81f93f43072b1d.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 224, + 353, + 231 + ], + "lines": [ + { + "bbox": [ + 329, + 224, + 353, + 231 + ], + "spans": [ + { + "bbox": [ + 329, + 224, + 353, + 231 + ], + "type": "text", + "content": "(f) Diff." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 369, + 173, + 419, + 224 + ], + "blocks": [ + { + "bbox": [ + 372, + 167, + 416, + 173 + ], + "lines": [ + { + "bbox": [ + 372, + 167, + 416, + 173 + ], + "spans": [ + { + "bbox": [ + 372, + 167, + 416, + 173 + ], + "type": "text", + "content": "(c) w/o HPM" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 369, + 173, + 419, + 224 + ], + "lines": [ + { + "bbox": [ + 369, + 173, + 419, + 224 + ], + "spans": [ + { + "bbox": [ + 369, + 173, + 419, + 224 + ], + "type": "image", + "image_path": "979bdc46fd9a2a949b09d9d5fb811cc58e9517113ea47f9e4071e51b82a7378c.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 375, + 224, + 413, + 231 + ], + "lines": [ + { + "bbox": [ + 375, + 224, + 413, + 231 + ], + "spans": [ + { + "bbox": [ + 375, + 224, + 413, + 231 + ], + "type": "text", + "content": "(g) w/ HPM" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 421, + 173, + 473, + 224 + ], + "blocks": [ + { + "bbox": [ + 434, + 167, + 458, + 173 + ], + "lines": [ + { + "bbox": [ + 434, + 167, + 458, + 173 + ], + "spans": [ + { + "bbox": [ + 434, + 167, + 458, + 173 + ], + "type": "text", + "content": "(d) Diff." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 421, + 173, + 473, + 224 + ], + "lines": [ + { + "bbox": [ + 421, + 173, + 473, + 224 + ], + "spans": [ + { + "bbox": [ + 421, + 173, + 473, + 224 + ], + "type": "image", + "image_path": "51a0aa5820758f9790116a3335d368233db2b35330b57d05dd8041426464facd.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 434, + 224, + 459, + 231 + ], + "lines": [ + { + "bbox": [ + 434, + 224, + 459, + 231 + ], + "spans": [ + { + "bbox": [ + 434, + 224, + 459, + 231 + ], + "type": "text", + "content": "(h) Diff." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 130, + 392, + 480, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 480, + 417 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 480, + 417 + ], + "type": "text", + "content": "Restormer [81] and DRSformer [7]. Compared to other CNN-/Transformer-based approaches, FPro still has a less or comparable model complexity." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 131, + 426, + 332, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 426, + 332, + 437 + ], + "spans": [ + { + "bbox": [ + 131, + 426, + 332, + 437 + ], + "type": "text", + "content": "Comparisons with Alternatives to FPro." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 130, + 437, + 333, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 333, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 333, + 545 + ], + "type": "text", + "content": "To further demonstrate the superiority of FPro, we compare it with recent prompt-based methods that mine spatial relations as prompts, including PromptIR [50] and PromptRestorer [68]. As shown in Tab. 10, following PromptIR [50], we train and validate FPro on Rain100L [77]. We achieve a substantial gain of 2.16 dB over PromptIR, and a 0.16 dB performance boost against PromptRestore" + } + ] + } + ], + "index": 28 + }, + { + "type": "table", + "bbox": [ + 343, + 483, + 481, + 531 + ], + "blocks": [ + { + "bbox": [ + 339, + 460, + 480, + 481 + ], + "lines": [ + { + "bbox": [ + 339, + 460, + 480, + 481 + ], + "spans": [ + { + "bbox": [ + 339, + 460, + 480, + 481 + ], + "type": "text", + "content": "Table 10: Comparisons with alternatives to FPro." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 343, + 483, + 481, + 531 + ], + "lines": [ + { + "bbox": [ + 343, + 483, + 481, + 531 + ], + "spans": [ + { + "bbox": [ + 343, + 483, + 481, + 531 + ], + "type": "table", + "html": "
ModelsParamsFLOPsPSNR
PromptIR [50]35.617337.04
PromptRestorer [68]24.418639.04
FPro22.38239.20
", + "image_path": "6aa40f3548898e6f3f59c5cf1a4b273d734022d3d4fece14d4b63e89ea163ca2.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "table_body" + } + ], + "index": 30 + }, + { + "type": "table", + "bbox": [ + 135, + 605, + 483, + 662 + ], + "blocks": [ + { + "bbox": [ + 205, + 583, + 408, + 596 + ], + "lines": [ + { + "bbox": [ + 205, + 583, + 408, + 596 + ], + "spans": [ + { + "bbox": [ + 205, + 583, + 408, + 596 + ], + "type": "text", + "content": "Table 8: Model efficiency analysis on SPAD [70]." + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 605, + 483, + 662 + ], + "lines": [ + { + "bbox": [ + 135, + 605, + 483, + 662 + ], + "spans": [ + { + "bbox": [ + 135, + 605, + 483, + 662 + ], + "type": "table", + "html": "
MethodMPRNet [83]SwinIR [33]Uformer-S [72]Restormer [81]IDT [75]DRSformer [7]FPro
FLOPs/G175.8238.043.9174.761.9242.981.9
Parameters/M20.111.520.626.116.433.722.3
Run-times/s0.031.830.120.140.280.080.08
PSNR/dB43.6444.9746.1347.9847.3448.5348.99
", + "image_path": "bc419c273db9d6d2d9dc6926a74d85cfeab91a90587771ff8ebe2fc8655e050a.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "table_body" + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 184, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 184, + 91, + 447, + 102 + ], + "type": "text", + "content": "A Frequency Prompt Guided Transformer for Image Restoration" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 136, + 481, + 174 + ], + "blocks": [ + { + "bbox": [ + 164, + 114, + 448, + 126 + ], + "lines": [ + { + "bbox": [ + 164, + 114, + 448, + 126 + ], + "spans": [ + { + "bbox": [ + 164, + 114, + 448, + 126 + ], + "type": "text", + "content": "Table 9: Quantitative results (NIQE) for real-world image deraining." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 136, + 481, + 174 + ], + "lines": [ + { + "bbox": [ + 136, + 136, + 481, + 174 + ], + "spans": [ + { + "bbox": [ + 136, + 136, + 481, + 174 + ], + "type": "table", + "html": "
Methods|Input Uformer-S [72] Restormer [81] IDT [75] DRSformer [7] FPro
NIQE ↓5.80125.69715.66315.60855.59425.2999
", + "image_path": "5a660243d3365e9b481b0cd826e993dcde553dc6252e88c3ca2a8e219807045e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 149, + 188, + 306, + 277 + ], + "blocks": [ + { + "bbox": [ + 149, + 188, + 306, + 277 + ], + "lines": [ + { + "bbox": [ + 149, + 188, + 306, + 277 + ], + "spans": [ + { + "bbox": [ + 149, + 188, + 306, + 277 + ], + "type": "image", + "image_path": "9cc4a6674949be79df7209c8c806c8782f88704b18e459bda2c9125addbf31f4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 211, + 278, + 244, + 288 + ], + "lines": [ + { + "bbox": [ + 211, + 278, + 244, + 288 + ], + "spans": [ + { + "bbox": [ + 211, + 278, + 244, + 288 + ], + "type": "text", + "content": "(a) Input" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 189, + 465, + 277 + ], + "blocks": [ + { + "bbox": [ + 307, + 189, + 465, + 277 + ], + "lines": [ + { + "bbox": [ + 307, + 189, + 465, + 277 + ], + "spans": [ + { + "bbox": [ + 307, + 189, + 465, + 277 + ], + "type": "image", + "image_path": "c05e9955531fb6c91afd56c7ccacd582731c6cab10d72208127accbb082de0e9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 370, + 278, + 402, + 287 + ], + "lines": [ + { + "bbox": [ + 370, + 278, + 402, + 287 + ], + "spans": [ + { + "bbox": [ + 370, + 278, + 402, + 287 + ], + "type": "text", + "content": "(b) FPro" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 297, + 481, + 319 + ], + "lines": [ + { + "bbox": [ + 131, + 297, + 481, + 319 + ], + "spans": [ + { + "bbox": [ + 131, + 297, + 481, + 319 + ], + "type": "text", + "content": "Fig. 9: Examples of erroneous reconstruction are shown, where heavy degradation in nighttime real-world scenes leads to typical failures of FPro." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 346, + 220, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 346, + 220, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 346, + 220, + 358 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 382, + 482, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 382, + 482, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 382, + 482, + 491 + ], + "type": "text", + "content": "In this work, we explore the benefits of prompt learning from a frequency perspective for the task of image restoration. First, when dynamic decoupling the input features with a gating mechanism to select representative elements, we obtain the related frequency components with regard to the specific degradation removal task. Then, we propose modulating the low-/high-frequency signals with separate branches, which concern the intrinsic characteristics of feature maps from different frequency bands. With these modules, our proposed FPro surpasses previous state-of-the-art methods in several image restoration tasks, while performing competitively in terms of computational cost." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 492, + 482, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 492, + 482, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 492, + 482, + 552 + ], + "type": "text", + "content": "Limitations. There remain many avenues for further improvements. For instance, one could achieve better performance by addressing failure cases shown in Fig. 9, where FPro meets challenges in dealing with heavy degradation in the nighttime real-world scene. Intuitively, collecting a large-scale real-world dataset is a potential direction for improvements." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 581, + 247, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 581, + 247, + 596 + ], + "spans": [ + { + "bbox": [ + 132, + 581, + 247, + 596 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 617, + 481, + 666 + ], + "type": "text", + "content": "This work was supported by Natural Science Foundation of Tianjin, China (No. 20JCJQJC00020), the National Natural Science Foundation of China (Nos. U22B2049, 62302240), Fundamental Research Funds for the Central Universities, and Supercomputing Center of Nankai University (NKSC)." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "text", + "content": "Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 141, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 138, + 141, + 481, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 141, + 481, + 207 + ], + "spans": [ + { + "bbox": [ + 138, + 141, + 481, + 207 + ], + "type": "text", + "content": "1. Brown, T.B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., Ziegler, D.M., Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Chess, B., Clark, J., Berner, C., McCandlish, S., Radford, A., Sutskever, I., Amodei, D.: Language models are few-shot learners. In: NeurIPS (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 209, + 481, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 209, + 481, + 229 + ], + "spans": [ + { + "bbox": [ + 138, + 209, + 481, + 229 + ], + "type": "text", + "content": "2. Cai, B., Xu, X., Jia, K., Qing, C., Tao, D.: Dehazenet: An end-to-end system for single image haze removal. TIP (2016)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 231, + 481, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 231, + 481, + 263 + ], + "spans": [ + { + "bbox": [ + 138, + 231, + 481, + 263 + ], + "type": "text", + "content": "3. Chantas, G., Galatsanos, N.P., Molina, R., Katsaggelos, A.K.: Variational bayesian image restoration with a product of spatially weighted total variation image priors. TIP (2009)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 265, + 481, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 265, + 481, + 286 + ], + "spans": [ + { + "bbox": [ + 138, + 265, + 481, + 286 + ], + "type": "text", + "content": "4. Chen, H., Wang, Y., Guo, T., Xu, C., Deng, Y., Liu, Z., Ma, S., Xu, C., Xu, C., Gao, W.: Pre-trained image processing transformer. In: CVPR (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 287, + 481, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 287, + 481, + 308 + ], + "spans": [ + { + "bbox": [ + 138, + 287, + 481, + 308 + ], + "type": "text", + "content": "5. Chen, L., Chu, X., Zhang, X., Sun, J.: Simple baselines for image restoration. In: ECCV (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 309, + 481, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 309, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 138, + 309, + 481, + 331 + ], + "type": "text", + "content": "6. Chen, L., Lu, X., Zhang, J., Chu, X., Chen, C.: Hinet: Half instance normalization network for image restoration. In: CVPR Workshops (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 331, + 481, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 331, + 481, + 353 + ], + "spans": [ + { + "bbox": [ + 138, + 331, + 481, + 353 + ], + "type": "text", + "content": "7. Chen, X., Li, H., Li, M., Pan, J.: Learning a sparse transformer network for effective image deraining. In: CVPR (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 354, + 481, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 354, + 481, + 375 + ], + "spans": [ + { + "bbox": [ + 138, + 354, + 481, + 375 + ], + "type": "text", + "content": "8. Chen, Y., Dai, X., Liu, M., Chen, D., Yuan, L., Liu, Z.: Dynamic convolution: Attention over convolution kernels. In: CVPR (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 376, + 481, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 376, + 481, + 398 + ], + "spans": [ + { + "bbox": [ + 138, + 376, + 481, + 398 + ], + "type": "text", + "content": "9. Cho, S.J., Ji, S.W., Hong, J.P., Jung, S.W., Ko, S.J.: Rethinking coarse-to-fine approach in single image deblurring. In: ICCV (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 399, + 481, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 399, + 481, + 420 + ], + "spans": [ + { + "bbox": [ + 138, + 399, + 481, + 420 + ], + "type": "text", + "content": "0. Deng, X., Dragotti, P.L.: Deep convolutional neural network for multi-modal image restoration and fusion. TPAMI (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 421, + 481, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 421, + 481, + 442 + ], + "spans": [ + { + "bbox": [ + 138, + 421, + 481, + 442 + ], + "type": "text", + "content": "1. Dong, J., Pan, J., Yang, Z., Tang, J.: Multi-scale residual low-pass filter network for image deblurring. In: ICCV (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 444, + 481, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 444, + 481, + 464 + ], + "spans": [ + { + "bbox": [ + 138, + 444, + 481, + 464 + ], + "type": "text", + "content": "2. Dong, Y., Liu, Y., Zhang, H., Chen, S., Qiao, Y.: Fd-gan: Generative adversarial networks with fusion-discriminator for single image dehazing. In: AAAI (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 466, + 481, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 466, + 481, + 509 + ], + "spans": [ + { + "bbox": [ + 134, + 466, + 481, + 509 + ], + "type": "text", + "content": "13. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 510, + 481, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 510, + 481, + 542 + ], + "spans": [ + { + "bbox": [ + 134, + 510, + 481, + 542 + ], + "type": "text", + "content": "14. d'Ascoli, S., Touvron, H., Leavitt, M.L., Morcos, A.S., Biroli, G., Sagun, L.: Convit: Improving vision transformers with soft convolutional inductive biases. In: ICML (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 544, + 481, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 544, + 481, + 564 + ], + "spans": [ + { + "bbox": [ + 134, + 544, + 481, + 564 + ], + "type": "text", + "content": "15. Eigen, D., Krishnan, D., Fergus, R.: Restoring an image taken through a window covered with dirt or rain. In: ICCV (2013)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 566, + 481, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 566, + 481, + 587 + ], + "spans": [ + { + "bbox": [ + 134, + 566, + 481, + 587 + ], + "type": "text", + "content": "16. Fu, X., Huang, J., Zeng, D., Huang, Y., Ding, X., Paisley, J.: Removing rain from single images via a deep detail network. In: CVPR (2017)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 134, + 588, + 481, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 588, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 134, + 588, + 481, + 620 + ], + "type": "text", + "content": "17. Gan, Y., Bai, Y., Lou, Y., Ma, X., Zhang, R., Shi, N., Luo, L.: Decorate the newcomers: Visual domain prompt for continual test time adaptation. In: AAAI (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 134, + 621, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 621, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 134, + 621, + 481, + 643 + ], + "type": "text", + "content": "18. Guo, Y., Xiao, X., Chang, Y., Deng, S., Yan, L.: From sky to the ground: A large-scale benchmark and simple baseline towards real rain removal. In: ICCV (2023)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 134, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 643, + 481, + 665 + ], + "type": "text", + "content": "19. He, B., Wang, C., Shi, B., Duan, L.: Mop moiré patterns using mopnet. In: ICCV (2019)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 184, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 184, + 91, + 447, + 102 + ], + "type": "text", + "content": "A Frequency Prompt Guided Transformer for Image Restoration" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 481, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 481, + 138 + ], + "type": "text", + "content": "20. He, B., Wang, C., Shi, B., Duan, L.: Fhde" + }, + { + "bbox": [ + 132, + 116, + 481, + 138 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 132, + 116, + 481, + 138 + ], + "type": "text", + "content": "net: Full high definition demoireing network. In: ECCV (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 481, + 160 + ], + "type": "text", + "content": "21. He, K., Sun, J., Tang, X.: Single image haze removal using dark channel prior. TPAMI (2010)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 161, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 481, + 182 + ], + "type": "text", + "content": "22. Hendrycks, D., Gimpel, K.: Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415 (2016)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "type": "text", + "content": "23. Isola, P., Zhu, J.Y., Zhou, T., Efros, A.A.: Image-to-image translation with conditional adversarial networks. In: CVPR (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 205, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 226 + ], + "type": "text", + "content": "24. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: ECCV (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "type": "text", + "content": "25. Jiang, L., Dai, B., Wu, W., Loy, C.C.: Focal frequency loss for image reconstruction and synthesis. In: CVPR (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "type": "text", + "content": "26. Khattak, M.U., Rasheed, H., Maaz, M., Khan, S., Khan, F.S.: Maple: Multi-modal prompt learning. In: CVPR (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 271, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 481, + 293 + ], + "type": "text", + "content": "27. Kong, L., Dong, J., Ge, J., Li, M., Pan, J.: Efficient frequency domain-based transformers for high-quality image deblurring. In: CVPR (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 293, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 314 + ], + "type": "text", + "content": "28. Li, B., Peng, X., Wang, Z., Xu, J., Feng, D.: Aod-net: All-in-one dehazing network. In: ICCV (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 315, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 335 + ], + "type": "text", + "content": "29. Li, B., Ren, W., Fu, D., Tao, D., Feng, D., Zeng, W., Wang, Z.: Benchmarking single-image dehazing and beyond. TIP (2018)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 336, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 481, + 357 + ], + "type": "text", + "content": "30. Li, B., Liu, X., Hu, P., Wu, Z., Lv, J., Peng, X.: All-in-one image restoration for unknown corruption. In: CVPR (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 358, + 481, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 379 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 379 + ], + "type": "text", + "content": "31. Li, Y., Tan, R.T., Guo, X., Lu, J., Brown, M.S.: Rain streak removal using layer priors. In: CVPR (2016)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 380, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 402 + ], + "type": "text", + "content": "32. Li, Z., Lei, Y., Ma, C., Zhang, J., Shan, H.: Prompt-in-prompt learning for universal image restoration. arXiv preprint arXiv:2312.05038 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 403, + 481, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 481, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 481, + 424 + ], + "type": "text", + "content": "33. Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: Swinir: Image restoration using swin transformer. In: ICCV Workshops (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 425, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 425, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 425, + 481, + 445 + ], + "type": "text", + "content": "34. Liu, D., Wen, B., Fan, Y., Loy, C.C., Huang, T.S.: Non-local recurrent network for image restoration. In: NeurIPS (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 447, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 481, + 456 + ], + "type": "text", + "content": "35. Liu, J., Yan, M., Zeng, T.: Surface-aware blind image deblurring. TPAMI (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 457, + 481, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 481, + 478 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 481, + 478 + ], + "type": "text", + "content": "36. Liu, L., Liu, J., Yuan, S., Slabaugh, G., Leonardis, A., Zhou, W., Tian, Q.: Wavelet-based dual-branch network for image demoiring. In: ECCV (2020)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 479, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 501 + ], + "type": "text", + "content": "37. Liu, L., Xie, L., Zhang, X., Yuan, S., Chen, X., Zhou, W., Li, H., Tian, Q.: Tape: Task-agnostic prior embedding for image restoration. In: ECCV (2022)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 502, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 481, + 522 + ], + "type": "text", + "content": "38. Liu, X., Suganuma, M., Sun, Z., Okatani, T.: Dual residual networks leveraging the potential of paired operations for image restoration. In: CVPR (2019)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 523, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 481, + 544 + ], + "type": "text", + "content": "39. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: ICCV (2021)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 545, + 481, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 481, + 577 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 481, + 577 + ], + "type": "text", + "content": "40. Ma, J., Cheng, T., Wang, G., Zhang, Q., Wang, X., Zhang, L.: Prores: Exploring degradation-aware visual prompt for universal image restoration. arXiv preprint arXiv:2306.13653 (2023)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 578, + 481, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 481, + 598 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 481, + 598 + ], + "type": "text", + "content": "41. Mei, Y., Fan, Y., Zhou, Y.: Image super-resolution with non-local sparse attention. In: CVPR (2021)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 132, + 600, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 481, + 621 + ], + "type": "text", + "content": "42. Mittal, A., Soundararajan, R., Bovik, A.C.: Making a \"completely blind\" image quality analyzer. IEEE SPL (2012)" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 132, + 622, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 481, + 643 + ], + "type": "text", + "content": "43. Nah, S., Hyun Kim, T., Mu Lee, K.: Deep multi-scale convolutional neural network for dynamic scene deblurring. In: CVPR (2017)" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "type": "text", + "content": "44. Narasimhan, S.G., Nayar, S.K.: Contrast restoration of weather degraded images. TPAMI (2003)" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "45. Niu, B., Wen, W., Ren, W., Zhang, X., Yang, L., Wang, S., Zhang, K., Cao, X., Shen, H.: Single image super-resolution via a holistic attention network. In: ECCV (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 160 + ], + "type": "text", + "content": "46. Oppenheim, A.: Discrete-time signal processing. Prentice-Hall google schola (1999)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 161, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 480, + 182 + ], + "type": "text", + "content": "47. Ozdenizci, O., Legenstein, R.: Restoring vision in adverse weather conditions with patch-based denoising diffusion models. TPAMI (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "type": "text", + "content": "48. Pan, X., Zhan, X., Dai, B., Lin, D., Loy, C.C., Luo, P.: Exploiting deep generative prior for versatile image restoration and manipulation. TPAMI (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 205, + 433, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 205, + 433, + 215 + ], + "spans": [ + { + "bbox": [ + 133, + 205, + 433, + 215 + ], + "type": "text", + "content": "49. Park, N., Kim, S.: How do vision transformers work? In: ICLR (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 216, + 480, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 236 + ], + "type": "text", + "content": "50. Potlapalli, V., Zamir, S.W., Khan, S., Khan, F.S.: Prompt: Prompting for all-in-one blind image restoration. In: NeurIPS (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 237, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 480, + 258 + ], + "type": "text", + "content": "51. Purohit, K., Suin, M., Rajagopalan, A., Boddeti, V.N.: Spatially-adaptive image restoration using distortion-guided networks. In: ICCV (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "text", + "content": "52. Qian, R., Tan, R.T., Yang, W., Su, J., Liu, J.: Attentive generative adversarial network for raindrop removal from a single image. In: CVPR (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 281, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 480, + 303 + ], + "type": "text", + "content": "53. Qu, Y., Chen, Y., Huang, J., Xie, Y.: Enhanced pix2pix dehazing network. In: CVPR (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "text", + "content": "54. Quan, Y., Deng, S., Chen, Y., Ji, H.: Deep learning for seeing through window with raindrops. In: ICCV (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 326, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 346 + ], + "type": "text", + "content": "55. Rabiner, L.R., Gold, B.: Theory and application of digital signal processing. Englewood Cliffs: Prentice-Hall (1975)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "type": "text", + "content": "56. Ren, D., Zuo, W., Hu, Q., Zhu, P., Meng, D.: Progressive image deraining networks: A better and simpler baseline. In: CVPR (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 369, + 480, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 390 + ], + "type": "text", + "content": "57. Ren, W., Liu, S., Zhang, H., Pan, J., Cao, X., Yang, M.H.: Single image dehazing via multi-scale convolutional neural networks. In: ECCV (2016)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 391, + 480, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 391, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 391, + 480, + 423 + ], + "type": "text", + "content": "58. Ren, W., Zhang, J., Pan, J., Liu, S., Ren, J.S., Du, J., Cao, X., Yang, M.H.: Deblurring dynamic scenes via spatially varying recurrent neural networks. TPAMI (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 424, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 480, + 445 + ], + "type": "text", + "content": "59. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: MICCAI (2015)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 446, + 480, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 480, + 478 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 480, + 478 + ], + "type": "text", + "content": "60. Song, X., Zhou, D., Li, W., Dai, Y., Shen, Z., Zhang, L., Li, H.: Tusr-net: Triple unfolding single image dehazing with self-regularization and dual feature to pixel attention. TIP (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 479, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 500 + ], + "type": "text", + "content": "61. Sun, Y., Yu, Y., Wang, W.: Moiré photo restoration using multiresolution convolutional neural networks. TIP (2018)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 501, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 522 + ], + "type": "text", + "content": "62. Tu, Z., Talebi, H., Zhang, H., Yang, F., Milanfar, P., Bovik, A., Li, Y.: Maxim: Multi-axis mlp for image processing. In: CVPR (2022)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 522, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 480, + 544 + ], + "type": "text", + "content": "63. Valanarasu, J.M.J., Yasarla, R., Patel, V.M.: Transweather: Transformer-based restoration of images degraded by adverse weather conditions. In: CVPR (2022)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 545, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 480, + 566 + ], + "type": "text", + "content": "64. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "type": "text", + "content": "65. Voigtman, E., Winefordner, J.D.: Low-pass filters for signal averaging. Review of Scientific Instruments (1986)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 589, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 480, + 610 + ], + "type": "text", + "content": "66. Wang, C., He, B., Wu, S., Wan, R., Shi, B., Duan, L.Y.: Coarse-to-fine disentangling demoiréing framework for recaptured screen images. TPAMI (2023)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "text", + "content": "67. Wang, C., Pan, J., Lin, W., Dong, J., Wu, X.M.: Selfpromer: Self-prompt dehazing transformers with depth-consistency. arXiv preprint arXiv:2303.07033 (2023)" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "68. Wang, C., Pan, J., Wang, W., Dong, J., Wang, M., Ju, Y., Chen, J.: Promptstorer: A prompting image restoration method with degradation perception. In: NeurIPS (2023)" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 184, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 184, + 91, + 447, + 102 + ], + "type": "text", + "content": "A Frequency Prompt Guided Transformer for Image Restoration" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 621 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "69. Wang, H., Xie, Q., Zhao, Q., Meng, D.: A model-driven deep neural network for single image rain removal. In: CVPR (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 480, + 160 + ], + "type": "text", + "content": "70. Wang, T., Yang, X., Xu, K., Chen, S., Zhang, Q., Lau, R.W.: Spatial attentive single-image deraining with a high quality real rain dataset. In: CVPR (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 161, + 480, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 480, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 480, + 193 + ], + "type": "text", + "content": "71. Wang, W., Xie, E., Li, X., Fan, D.P., Song, K., Liang, D., Lu, T., Luo, P., Shao, L.: Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In: ICCV (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 215 + ], + "type": "text", + "content": "72. Wang, Z., Cun, X., Bao, J., Zhou, W., Liu, J., Li, H.: Uformer: A general u-shaped transformer for image restoration. In: CVPR (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "type": "text", + "content": "73. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. TIP (2004)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 237, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 480, + 259 + ], + "type": "text", + "content": "74. Wu, R., Yang, T., Sun, L., Zhang, Z., Li, S., Zhang, L.: Seesr: Towards semantics-aware real-world image super-resolution. arXiv preprint arXiv:2311.16518 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "text", + "content": "75. Xiao, J., Fu, X., Liu, A., Wu, F., Zha, Z.J.: Image de-raining transformer. TPAMI (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 281, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 480, + 303 + ], + "type": "text", + "content": "76. Xiao, T., Singh, M., Mintun, E., Darrell, T., Dollar, P., Girshick, R.: Early convolutions help transformers see better. In: NeurIPS (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 303, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 303, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 303, + 480, + 325 + ], + "type": "text", + "content": "77. Yang, W., Tan, R.T., Feng, J., Guo, Z., Yan, S., Liu, J.: Joint rain detection and removal from a single image with contextualized deep networks. TPAMI (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 325, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 480, + 357 + ], + "type": "text", + "content": "78. Yu, F., Gu, J., Li, Z., Hu, J., Kong, X., Wang, X., He, J., Qiao, Y., Dong, C.: Scaling up to excellence: Practicing model scaling for photo-realistic image restoration in the wild. arXiv preprint arXiv:2401.13627 (2024)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "type": "text", + "content": "79. Yu, K., Wang, X., Dong, C., Tang, X., Loy, C.C.: Path-restore: Learning network path selection for image restoration. TPAMI (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "text", + "content": "80. Yue, H., Mao, Y., Liang, L., Xu, H., Hou, C., Yang, J.: Recaptured screen image demoiring. TCSVT (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 402, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 480, + 424 + ], + "type": "text", + "content": "81. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: CVPR (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "type": "text", + "content": "82. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Learning enriched features for real image restoration and enhancement. In: ECCV (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 456, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 456, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 456, + 480, + 479 + ], + "type": "text", + "content": "83. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Multi-stage progressive image restoration. In: CVPR (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "type": "text", + "content": "84. Zhang, K., Li, Y., Zuo, W., Zhang, L., Van Gool, L., Timofte, R.: Plug-and-play image restoration with deep denoiser prior. TPAMI (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 501, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 522 + ], + "type": "text", + "content": "85. Zhao, H., Gou, Y., Li, B., Peng, D., Lv, J., Peng, X.: Comprehensive and delicate: An efficient transformer for image restoration. In: CVPR (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 522, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 480, + 544 + ], + "type": "text", + "content": "86. Zheng, B., Yuan, S., Slabaugh, G., Leonardis, A.: Image demoiring with learnable bandpass filters. In: CVPR (2020)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 544, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 480, + 567 + ], + "type": "text", + "content": "87. Zheng, C., Zhang, Y., Gu, J., Zhang, Y., Kong, L., Yuan, X.: Cross aggregation transformer for image restoration. In: NeurIPS (2022)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "type": "text", + "content": "88. Zhou, S., Chen, D., Pan, J., Shi, J., Yang, J.: Adapt or perish: Adaptive sparse transformer with attentive feature refinement for image restoration. In: CVPR (2024)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 599, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 480, + 621 + ], + "type": "text", + "content": "89. Zou, X., Xiao, F., Yu, Z., Li, Y., Lee, Y.J.: Delving deeper into anti-aliasing in convnets. IJCV (2023)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/SegGen_ Supercharging Segmentation Models with Text2Mask and Mask2Img Synthesis/acac7544-25ea-4a13-aba6-d24e96031cd3_content_list.json b/2024/SegGen_ Supercharging Segmentation Models with Text2Mask and Mask2Img Synthesis/acac7544-25ea-4a13-aba6-d24e96031cd3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..53686448132d034538a18c249ec3759ba00170e9 --- /dev/null +++ b/2024/SegGen_ Supercharging Segmentation Models with Text2Mask and Mask2Img Synthesis/acac7544-25ea-4a13-aba6-d24e96031cd3_content_list.json @@ -0,0 +1,1466 @@ +[ + { + "type": "text", + "text": "SegGen: Supercharging Segmentation Models with Text2Mask and Mask2Img Synthesis", + "text_level": 1, + "bbox": [ + 243, + 140, + 759, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hanrong Ye $^{1}$ , Jason Kuen $^{2}$ , Qing Liu $^{2}$ , Zhe Lin $^{2}$ , Brian Price $^{2}$ , Dan Xu $^{1}$", + "bbox": [ + 240, + 210, + 759, + 227 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1 CSE, HKUST", + "2 Adobe Research" + ], + "bbox": [ + 439, + 238, + 562, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. We present SegGen, a new data generation approach that pushes the performance boundaries of state-of-the-art image segmentation models. One major bottleneck of previous data synthesis methods for segmentation is the design of \"segmentation labeler module\", which is used to synthesize segmentation masks for images [48]. The segmentation labeler modules, which are segmentation models by themselves, bound the performance of downstream segmentation models trained on the synthetic masks. These methods encounter a \"chicken or egg dilemma\" and thus fail to outperform existing segmentation models. To address this issue, we propose a novel method that reverses the traditional data generation process: we first (i) generate highly diverse segmentation masks that match real-world distribution from text prompts, and then (ii) synthesize realistic images conditioned on the segmentation masks. In this way, we avoid the need for any segmentation labeler module. SegGen integrates two data generation strategies, namely MaskSyn and ImgSyn, to largely improve data diversity in synthetic masks and images. Notably, the high quality of our synthetic data enables our method to outperform the previous data synthesis method [48] by $+25.2$ mIoU on ADE20K when trained with pure synthetic data. On the highly competitive ADE20K and COCO benchmarks, our data generation method markedly improves the performance of state-of-the-art segmentation models in semantic segmentation, panoptic segmentation, and instance segmentation. Moreover, experiments show that training with our synthetic data makes the segmentation models more robust towards unseen data domains, including real-world and AI-generated images.", + "bbox": [ + 261, + 306, + 738, + 652 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Image Segmentation $\\cdot$ Image Generation $\\cdot$ Deep Learning", + "bbox": [ + 261, + 667, + 730, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 731, + 374, + 747 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image segmentation explores the identification of objects in visual inputs at the pixel level. Based on the different emphases on category and instance membership information, researchers have divided image segmentation into several tasks [5,25,35,40]. For example, semantic segmentation studies pixel-level understanding of object categories, instance segmentation focuses on instance grouping", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2abb841d4bbb9857413b2985a699cd0e8646bc00e92b8d8c9523bda968ee9011.jpg", + "image_caption": [ + "Fig. 1: Effectiveness of SegGen on various data domains: Through training with synthetic data generated by the proposed SegGen, we significantly boost the performance of state-of-the-art segmentation model Mask2Former [7] on evaluation benchmarks including ADE20K [60] and COCO [33], whilst making it more robust towards challenging images from other domains (the three columns on the left are from PASCAL [10]; the three on the right are synthesized by image generation model Kandinsky 2 [12]). SegGen outperforms the previous best data generation method (DiffuMask [48]) by a huge margin when models are trained on pure synthetic data. \"mIoU*\" is the average IoU metric defined by [48] which focuses on three common classes." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 787, + 305 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "of pixels, while panoptic segmentation considers both. For all these segmentation tasks, obtaining high-quality annotation is challenging as every individual pixel requires human labeling, and a single image can contain millions of pixels. Therefore, compared to other public datasets like ImageNet-21K (with around 14M images), the prevailing human-annotated segmentation datasets are notably smaller. For example, ADE20K dataset [61] contains about 20K images in its training split, while COCO [33] has around 118K training images. Although there has been significant development in the structure of segmentation models [8,29,52,55,57], the limited size of training data hinders further performance enhancements and results in inadequate generalization ability to handle images from unfamiliar domains, such as those from other scenes or synthesized by generative models as shown in the second row of Fig. 1.", + "bbox": [ + 212, + 443, + 787, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inspired by the recent success of image generation [9, 23], researchers start to explore using generative models to enhance image segmentation [1]. A representative direction of this research focuses on synthesizing segmentation training data in a cost-effective manner [26]. In related methods, the synthetic segmentation masks are obtained from some manually designed \"segmentation labeler modules\". The segmentation labeler modules are essentially segmentation models by themselves. They are either small-scale segmentation networks as in DatasetGAN [56] and Grounded Diffusion [32], or post-processing methods on the image features as in DiffuMask [48]. The performance of the downstream segmentation models is constrained by the quality of the synthetic masks they are trained on, which in turn relies on the capabilities of those segmentation labeler modules. This is a \"chicken or egg dilemma\" and the performance bottleneck is the segmentation labeler module. Therefore, while they achieve encouraging success in settings with limited training data, their methods fail to deliver notable en", + "bbox": [ + 212, + 628, + 789, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Hanrong Ye et al.", + "bbox": [ + 271, + 114, + 390, + 128 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/53e2a0790b39b289b0717afba00ada1f43d88652c137dca938dd01b9f8c52f99.jpg", + "image_caption": [ + "Fig. 2: Comparison with previous data generation methods for segmentation: (a) Earlier methods [48, 56] rely on segmentation labeler modules to produce segmentation masks for synthetic images. However, the performance of these downstream segmentation models, trained on synthetic data, is bounded by the capacity of the segmentation labeler modules. The segmentation labeler is a major bottleneck for the quality of the generated data. (b) We design a reverse pipeline: we first create diverse new masks from text prompts via a proposed Text2Mask generation model and then synthesize images conditioned on the segmentation masks. This methodology avoids any usage of segmentation labeler networks, resulting in significantly improved quality of our synthesis data." + ], + "image_footnote": [], + "bbox": [ + 223, + 145, + 787, + 223 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "hancements for the state-of-the-art models on the most popular segmentation benchmarks in settings using the complete training sets.", + "bbox": [ + 212, + 380, + 784, + 411 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "On the other hand, a handful of works [54, 58] have demonstrated that high-quality synthetic images can be generated by conditioning text-to-image models on dense input maps (e.g., segmentation masks or canny edge). Moreover, these dense-input conditional models provide a strong alignment between the dense input label maps and generated images. These two qualities motivate us to leverage the powerful capabilities of such models for effective segmentation data generation. To this end, we propose a novel segmentation data generation method, coined as SegGen, for generating high-quality segmentation training data. We reverse the traditional data generation pipeline for segmentation: we first synthesize high-quality segmentation masks from a proposed text-to-mask generation model, and then generate images based on the segmentation masks using a mask-to-image generation model. In this new framework, we avoid using any segmentation labeler modules, which results in significantly improved data quality. Fig. 2 shows the difference between our method and previous data generation methods for segmentation [48, 56].", + "bbox": [ + 212, + 412, + 787, + 641 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As each training sample of image segmentation consists of two components: segmentation masks and the corresponding image, we develop two novel data generation approaches, emphasizing improvements in two distinct aspects of data diversity: (i) segmentation masks and (ii) images. The first data generation approach, named MaskSyn, centers around the generation of new segmentation masks. It learns a text-to-mask (Text2Mask) generation model to produce completely new segmentation masks given text prompts. Then, it learns a mask-to-image (Mask2Img) generation model to synthesize images that align with the synthetic segmentation masks. The second data generation approach, named ImgSyn, utilizes the above-mentioned Mask2Img model to synthesize new images given human-annotated segmentation masks. With MaskSyn and ImgSyn, we can readily generate a vast array of diverse and high-quality synthetic training data. The combined synthetic data is used to train segmenta", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "SegGen", + "bbox": [ + 678, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tion models in conjunction with real training samples from human-annotated datasets. Experiments show that our synthetic data can significantly boost the performance of the image segmentation models on challenging benchmarks including ADE20K semantic segmentation, COCO panoptic segmentation, and COCO instance segmentation, achieving new state-of-the-art performances without using extra human-annotated data. Notably, SegGen remarkably outperforms the previous segmentation data generation method [48] by $+25.2$ mIoU when trained with pure synthetic data on ADE20K. Our method boosts the mIoU of Mask2Former by $+2.7$ (R50) and $+1.3$ (Swin-L) on the ADE20K semantic segmentation benchmark. Furthermore, segmentation models trained on our synthetic data exhibit a remarkably stronger ability to generalize across unfamiliar image domains, including real images from other distributions and machine-generated images.", + "bbox": [ + 212, + 146, + 787, + 343 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We summarize the contribution of this work in three points: (i) We propose a revolutionary generation framework that reverses the traditional segmentation data generation pipeline and solves the \"chicken or egg dilemma\". The new framework enables us to produce high-quality segmentation training data at scale, thus enabling the training of more powerful image segmentation models. (ii) We introduce two effective generative models, one for text-to-mask generation and the other for mask-to-image generation. Based on these models, we propose two novel segmentation training data generation approaches, namely MaskSyn and ImgSyn. They significantly improve the data diversity, with MaskSyn focusing on new segmentation masks and ImgSyn on new images. (iii) SegGen successfully improves the performance of the leading-edge segmentation models across the highly competitive benchmarks on ADE20K and COCO. Moreover, SegGen enhances the generalization ability of segmentation models towards unseen image domains. Rigorous experiments, including ablation study and peer comparison, strongly suggest the effectiveness of the proposed method.", + "bbox": [ + 212, + 343, + 787, + 570 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 594, + 387, + 611 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Generation for Segmentation Image segmentation is one of the most studied visual perception problems [30]. In recent years, there has been a surge in efforts to harness the capabilities of generative models for segmentation tasks. These efforts can be broadly classified into three categories based on their methodologies: (i) Extracting visual features from generative models for segmentation [1, 38, 50, 59]. These methods harvest the strong representation ability of diffusion models trained on large-scale datasets but are limited in the precision of predicted masks. (ii) Formulating segmentation tasks directly as generative models [6, 20, 45]. Although these methods propose exciting new model architectures, they usually consume higher computational costs while showing unimproved performance compared with conventional segmentation models. (iii) Synthesizing segmentation training data using generative models [26, 32, 37, 47-49, 56]. While these methods have showcased commendable results against their respective baselines especially when the training data is highly limited, they have", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Hanrong Ye et al.", + "bbox": [ + 271, + 114, + 390, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "yet to exhibit notably superior performance on the most rigorous benchmarks including ADE20K (150 categories) [61] and COCO (133 categories) [33] under fully-supervised setting. A primary concern with these techniques is the subpar quality of the generated segmentation masks. This stems from their dependence on the bottleneck segmentation labeler modules during the synthetic mask generation process. Therefore, a concurrent work [51] abandons the generation of segmentation masks and directly generates images based on existing masks, which is highly limited in mask diversity. In contrast, our data generation workflow reverses the traditional data generation pipeline: we first generate segmentation masks from text, and then synthesize images based on segmentation masks. Our method avoids the need for any segmentation labeler which results in largely improved data quality. Our SegGen can strongly enhance the existing state-of-the-art segmentation models on different challenging benchmarks.", + "bbox": [ + 212, + 146, + 787, + 343 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Conditional Image Synthesis Within the realm of conditional image synthesis, Generative Adversarial Networks [2, 13, 21, 23, 43], Variational Autoencoders [24], and Diffusion Models [9, 17, 42, 44] have been at the forefront. Recently, the open-source research community has shown a burgeoning interest in the latent diffusion-based Stable Diffusion (SD) series for text-to-image synthesis [41]. The most recent iteration, SDXL, introduced by [39], expands the model capacity, yielding significantly enhanced results. Therefore, we build SegGen upon SDXL. Regarding mask-to-image generation models [18, 31, 53], ControlNet [54] and T2I-Adapter [36] suggest freezing the parameters of the SD model and introducing a set of more compact, learnable modules. This approach enables image generation conditioned on the given segmentation masks. We adopt the structure of ControlNet in our mask-to-image generation model. To the best of our knowledge, there have been no prior endeavors on text-to-mask generation.", + "bbox": [ + 212, + 343, + 787, + 541 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 560, + 330, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Overview", + "text_level": 1, + "bbox": [ + 215, + 590, + 339, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SegGen is designed to synthesize high-quality training samples for improving the performance of segmentation models. The overall workflow is shown on the left of Fig. 3. We first train SegGen with human-annotated training samples from public datasets. After training, we use SegGen to produce new segmentation training samples at scale. The generated training samples are incorporated into the training process of segmentation models to enhance the model performance.", + "bbox": [ + 212, + 614, + 787, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Models in SegGen", + "text_level": 1, + "bbox": [ + 215, + 724, + 413, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 3, we first utilize a captioner model to extract captions of the real training images as text prompts from the target dataset. The text prompts will condition the data generation process. Then, two conditional generative models are introduced: a text-to-mask (Text2Mask) generation model and a mask-to-image (Mask2Img) generation model. Both generative models are built upon the SDXL model [39] which provides top-notch image generation quality.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "SegGen", + "bbox": [ + 678, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9a4bc0c5cf53f18d4e5ccdec7be3fb1a83f078c3c5a3da5f075adf4afb65aa36.jpg", + "image_caption": [ + "Fig.3: Illustration of the workflow of our proposed SegGen: We introduce two generative models: a text-to-mask (Text2Mask) generation model and a mask-to-image (Mask2Img) generation model, based on which we design two approaches for synthesizing segmentation training samples: MaskSyn and ImgSyn. (a) MaskSyn focuses on generating new segmentation masks. It first extracts the caption of the real image as a text prompt and uses it to generate new masks with the Text2Mask model. Then, the new masks and text prompt are fed into the Mask2Img model to produce the corresponding new images. (b) ImgSyn focuses on the synthesis of new images. It inputs human-labeled masks and text prompts into the Mask2Img model." + ], + "image_footnote": [], + "bbox": [ + 217, + 148, + 787, + 267 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Captioner Model To obtain image captions of existing training samples, we employ BLIP2-FlanT5xxl [28] model, which is a state-of-the-art vision-language model. We feed the text prompt \"Question: What are shown in the photo? Answer:\" and image as input to the model. The responses serve as text prompts to condition the following generation process.", + "bbox": [ + 212, + 409, + 787, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Text2Mask Model We design a Text2Mask model in SegGen for generating diverse segmentation masks based on given text prompts. To leverage the generation capacity of text-to-image generation models pre-trained on large-scale datasets, we encode the segmentation masks (the pixel values are category IDs) as three-channel RGB-like color maps, where one color represents a certain category. From our experiments, the color map reconstructed by VAE [41] in SDXL appears almost indistinguishable from the original input as shown in the supplementary materials. Therefore, we can directly fine-tune the text-to-image SDXL-base model with [text, segmentation color map] training pairs, which are from the public image segmentation dataset (e.g. ADE20K). During sampling, our Text2Mask model can generate diverse color maps conditioned on text prompts, which are subsequently converted into segmentation masks. Formally, suppose the input text prompt is $\\mathbf{T}$ , the target height and width are $H$ and $W$ , the synthesized color map is $\\mathbf{C}_{\\mathrm{syn}} \\in \\mathbb{W}^{H \\times W \\times 3}$ , and the synthesized segmentation map (with $N$ masks) is $\\mathbf{M}_{\\mathrm{syn}} \\in \\mathbb{W}^{H \\times W \\times N}$ , the generation process is:", + "bbox": [ + 212, + 486, + 787, + 714 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {C} _ {\\mathrm {s y n}} = \\operatorname {T e x t 2 M a s k} (\\mathbf {T}), \\\\ \\mathbf {M} _ {\\text {s y n}} = f _ {\\text {c o l o r} \\rightarrow \\text {m a s k}} \\left(\\mathbf {C} _ {\\text {s y n}}\\right), \\tag {1} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 728, + 784, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $f_{\\mathrm{color} \\to \\mathrm{mask}}: \\mathbb{W}^{H \\times W \\times 3} \\to \\mathbb{W}^{H \\times W \\times N}$ is the function that projects the color maps to segmentation masks. Specifically, for each pixel on the color maps, we identify its nearest color (in Euclidean space) in the aforementioned lookup table, and assign the corresponding class to the pixel in the segmentation masks.", + "bbox": [ + 212, + 777, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Hanrong Ye et al.", + "bbox": [ + 271, + 114, + 392, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c9017728c00b05353ee46f2c44a7598961464a543b8f2a9149d487578b592409.jpg", + "image_caption": [ + "Fig. 4: Generated samples by MaskSyn on ADE20K: The third row overlays the mask and the image together to demonstrate the alignment between them. The generated segmentation masks and images demonstrate high perceptual quality and excellent alignment (see more samples in supplementary materials)." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 284 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Mask2Img Model The goal of the Mask2Img model is to synthesize new images that align well with the given segmentation masks and text prompts. Specifically we adopt the structure of ControlNet [54]: we freeze the pre-trained weights of the SDXL-base model and train an additional side network for mask-conditioned image generation. It simultaneously keeps the generalization ability of the pretrained diffusion model and provides excellent controllable generation ability. The Mask2Img model is trained with the [text, segmentation color map, image] triplets gathered from the training splits of the target datasets. We denote the input segmentation map as $\\mathbf{M} \\in \\mathbb{W}^{H \\times W \\times N}$ , the color map as $\\mathbf{C} \\in \\mathbb{W}^{H \\times W \\times 3}$ , the synthetic image as $\\mathbf{I}_{\\mathrm{syn}} \\in \\mathbb{W}^{H \\times W \\times 3}$ , and the generation process is:", + "bbox": [ + 212, + 364, + 787, + 517 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {C} = f _ {\\text {m a s k} \\rightarrow \\text {c o l o r}} (\\mathbf {M}), \\\\ \\mathbf {I} _ {\\text {s y n}} = \\operatorname {M a s k 2 I m g} (\\mathbf {T}, \\mathbf {C}), \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 413, + 530, + 784, + 564 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $f_{\\mathrm{mask}\\rightarrow \\mathrm{color}}:\\mathbb{W}^{H\\times W\\times N}\\to \\mathbb{W}^{H\\times W\\times 3}$ is the function to convert the segmentation masks into a color map. For semantic segmentation, the value of each pixel on the segmentation mask corresponds to a category ID, allowing us to convert the masks directly into an RGB color map using a pre-defined lookup table. For panoptic and instance segmentation, after mapping the category IDs to color maps, it is essential to outline each segment with a special edge color on the color map. This ensures the model recognizes the specific instance it belongs to. The segmentation map $\\mathbf{M}$ can be human-annotated or synthetic (i.e., $\\mathbf{M}_{\\mathrm{syn}}$ from Text2Mask Model as shown in Eq. 1).", + "bbox": [ + 212, + 577, + 787, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Data Generation", + "text_level": 1, + "bbox": [ + 214, + 738, + 401, + 752 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "With the aforementioned generative models, SegGen proposes two approaches for synthesizing new segmentation training samples: MaskSyn and ImgSyn. MaskSyn focuses on enhancing the diversity of synthetic segmentation masks, whereas ImgSyn concentrates on diversifying synthetic images. These generation approaches are illustrated on the right of Fig. 3.", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "SegGen", + "bbox": [ + 678, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3c600986a67bc54a0876c08e61aed94f78788a7b34d9b9135ab8376737e72cc6.jpg", + "image_caption": [ + "Fig. 5: Generated samples by ImgSyn on ADE20K: The generated images exhibit remarkable realism and align well with the human-labeled masks and text prompts (see more samples in supplementary materials)." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 241 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/200a30a9c1257688104fa7d6aeaad1e3dba5099ac0f91d4f95da3e7689242c02.jpg", + "image_caption": [ + "Fig. 6: Zoom-in comparison of real images and our ImgSyn images: As highlighted in the circles, our synthetic images align better with the human-labeled masks in many cases because of the inaccuracies in annotations. The left 4 columns are from ADE20K and the right 4 from COCO." + ], + "image_footnote": [], + "bbox": [ + 217, + 292, + 785, + 404 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "MaskSyn MaskSyn starts with a real training sample pair [image, segmentation masks] from a human-annotated segmentation dataset. It first extracts the caption of the image with the image captioner model. The caption serves as a text prompt and is used to generate a set of diverse new segmentation masks with the Text2Mask model following Eq. 1. Subsequently, the new segmentation masks and the corresponding text prompt are fed into the Mask2Img model to generate a new image that aligns well with its mask. As such, each training sample crafted by MaskSyn includes both a novel segmentation mask and a new image. MaskSyn effectively increases the data diversity in segmentation masks to a great extent. Some generated samples are shown in Fig. 4.", + "bbox": [ + 212, + 473, + 787, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ImgSyn Different from MaskSyn, ImgSyn focuses on increasing the data diversity of images based on human-annotated segmentation masks. For each real training sample pair [image, segmentation mask], it takes the human-annotated segmentation mask and the text prompt extracted from the image as input, and generates a set of varied images that align well with the human-annotated mask. In this way, new training samples consisting of human-labeled masks and new synthetic images are generated. ImgSyn can also be viewed as a kind of data augmentation method that enhances the data diversity on the image side. Our experiments reveal a remarkably high alignment between the synthetic images and their respective segmentation masks. Some synthesized results are showcased in Fig. 5. Our machine-generated synthetic images achieve a better mask-image alignment than real images in many cases, as shown in Fig. 6. This phenomenon occurs because human annotations tend to be imperfect due to the high difficulty of annotating segmentation masks.", + "bbox": [ + 212, + 628, + 789, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Hanrong Ye et al.", + "bbox": [ + 271, + 114, + 392, + 128 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9803c9b23853486064edb37743d11a1cb0f86c60c04ebe75624fe3c378b5461f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodVenueBackboneCrop SizeIterationsmIoU (s.s.)mIoU (m.s.)
MaskFormer [8]NeurIPS 2021R50512160k44.546.7
Mask DINO [27]CVPR 2023R50512160k48.7-
OneFormer [19]CVPR 2023R50512160k47.3-
Mask2Former [7]CVPR 2022R50512160k47.249.2
—w/ SegGen (ours)-R50512160k49.9 (+2.7)51.4 (+2.2)
MaskFormer [8]NeurIPS 2021Swin-L640160k54.155.6
Mask DINO [27]CVPR 2023Swin-L640160k56.6-
OneFormer [19]CVPR 2023Swin-L640160k57.057.7
Mask2Former [7]CVPR 2022Swin-L640160k56.157.3
—w/ SegGen (ours)-Swin-L640160k57.4 (+1.3)58.7 (+1.4)
", + "bbox": [ + 232, + 133, + 769, + 284 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 1: Semantic segmentation on ADE20K val (150 categories): Synthetic data is used as a data augmentation. With the same model structures and training recipes, our SegGen boosts the performance of Mask2Former by a large margin and establishes a new SOTA performance without extra real data under both single-scale (s.s.) and multi-scale (m.s.) test settings.", + "bbox": [ + 212, + 285, + 784, + 353 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.4 Training Segmentation Models with Synthetic Data", + "text_level": 1, + "bbox": [ + 214, + 363, + 686, + 378 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The final goal of this work is to improve the performance of current segmentation models with the synthetically generated training data. Therefore, we use both synthetic data produced by SegGen and training data from existing datasets in the training process of segmentation models.", + "bbox": [ + 212, + 385, + 784, + 445 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We empirically investigate two training strategies utilizing the synthetic data: (i) Synthetic data augmentation strategy. The synthetic data is used for random data augmentation. In every iteration in the training process, each real training sample is replaced by the synthetic training sample with a probability $p_{\\mathrm{aug}}$ . (ii) Synthetic data pre-training strategy. It comprises two training stages: pretraining and fine-tuning. In the pre-training stage, we pre-train the segmentation models on synthetic data, so that they learn good weights that are transferable and favorable for fine-tuning. In the subsequent fine-tuning stage, the segmentation models are trained with solely human-annotated data.", + "bbox": [ + 212, + 446, + 785, + 582 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 603, + 375, + 619 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To accurately evaluate the effectiveness of our data generation method in improving segmentation performance, we adopt mainstream segmentation models and commonly used evaluation benchmarks for several typical segmentation tasks. The experiments are conducted mostly under the fully-supervised learning setting, meaning all human-annotated training samples from the evaluated datasets are used alongside our synthetic data. To guarantee an unbiased assessment of the impact of our synthetic data, we keep the architectures of the segmentation models and training protocols consistent with their respective original implementations throughout our studies.", + "bbox": [ + 212, + 631, + 787, + 767 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Implementation Details", + "text_level": 1, + "bbox": [ + 214, + 787, + 455, + 801 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Segmentation Datasets and Evaluation We conduct experiments on three image segmentation benchmarks following the main experimental settings of", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "SegGen", + "bbox": [ + 678, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Mask2Former [7]: ADE20K semantic segmentation [60], COCO panoptic segmentation, and COCO instance segmentation [33]. Our evaluation uses all 150 classes for ADE20K and 133 classes for COCO. We use all the images from the training splits in the training of segmentation models. For semantic segmentation, we show the mean Intersection-over-Union metric (mIoU). For instance segmentation, the average precision (AP) is used. For panoptic segmentation, we report panoptic quality (PQ), \"thing\" instance segmentation $\\mathrm{AP}_{\\mathrm{pan}}^{\\mathrm{Th}}$ , and semantic segmentation $\\mathrm{mIoU}_{\\mathrm{pan}}$ .", + "bbox": [ + 212, + 145, + 787, + 267 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Segmentation Models We adopt Mask2Former [7], a recently-proposed prevalent transformer model, as the default segmentation model. Two typical backbones, i.e., R50 [16] and Swin-L [34], are studied. We keep the official implementation and training hyper-parameters of the segmentation models unchanged. For more training details please refer to their paper. We also conduct experiments on Mask DINO [27], a detection-aided segmentation model, and HRNet W48 [46], a representative fully-convolutional model. More implementation details are introduced in the supplementary materials.", + "bbox": [ + 212, + 267, + 787, + 387 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Data Generation The Text2Mask and Mask2Img models are both based on the SDXL-base model [39]. They are trained on the training splits of the target segmentation datasets (i.e. ADE20K or COCO) separately for 30,000 iterations using a learning rate of $10^{-5}$ . AdamW optimizer is employed, and the models are trained at a resolution of 768. Pre-trained weights from SDXL-base [39] are utilized. Random flipping data augmentation is used. During sampling, the default EDM sampler [22] is used. The sampling steps are 200 in the Text2Mask model and 40 in the Mask2Img model. During data sampling, for each training sample in the ADE20K semantic segmentation dataset, we produce 10 synthetic mask-image pairs using MaskSyn, resulting in 202,100 training samples. Additionally, we synthesize 50 images based on each human-labeled mask with ImgSyn, leading to a total of 1,010,500 samples. On COCO, we solely use ImgSyn for data synthesis. By generating 10 synthetic images conditioned on each human-labeled panoptic segmentation mask via ImgSyn, our synthetic set amounts to 1,182,870 synthetic samples, which are used in the training of both panoptic and instance segmentation models.", + "bbox": [ + 212, + 388, + 787, + 628 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Main Results", + "text_level": 1, + "bbox": [ + 215, + 650, + 372, + 664 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "ADE20K Semantic Segmentation We employ the synthetic data augmentation strategy with $p_{\\mathrm{aug}} = 60\\%$ for Mask2Former, and show the results in Table 1. SegGen significantly boosts the mIoU of Mask2Former R50 by +2.7 for single-scale inference and +2.2 for multi-scale inference, achieving 49.9 and 51.4 correspondingly. The Swin-L variant is also largely improved from 56.1/57.3 (single-scale/multi-scale) to 57.4/58.7 (+1.3/+1.4). Notably, our data synthesis method helps Mask2Former surpass the newer methods such as Mask DINO and One-Former [19], while establishing new SOTA results for R50 and Swin-L settings without using additional human-annotated data.", + "bbox": [ + 212, + 672, + 789, + 808 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "COCO Panoptic Segmentation On COCO we adopt the synthetic data pretraining strategy to utilize our synthetic data. Specifically, we pre-train the mod-", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Hanrong Ye et al.", + "bbox": [ + 271, + 114, + 390, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/3b3794b082d636e53e627d00a44c617533c9be569a29cdb280b3b085a3cc6fb2.jpg", + "table_caption": [], + "table_footnote": [ + "Table 2: Panoptic segmentation on COCO panoptic val2017 (133 categories): Synthetic data is used for pre-training. Our SegGen significantly improves the performance compared with real data pre-training baselines, and establishes new SOTA performance without extra real data." + ], + "table_body": "
MethodBackbone QueriesEpochsPQ\\( PQ^{Th} \\)\\( PQ^{St} \\)\\( AP_{pan}^{Th} \\)\\( mIoU_{pan} \\)
DETR [3]R50100500+2543.448.236.331.1-
MaskFormer [8]R5010030046.551.039.833.057.8
Mask2Former [7]R501005051.957.743.041.761.7
Mask DINO [27]R503005053.059.143.943.3-
Mask2Former [7]R5010050+5052.057.943.442.061.0
— w/SegGen (ours)R5010050+5052.7 (+0.7)58.8 (+0.9)43.6 (+0.2)43.1 (+1.1)62.6 (+1.6)
Mask DINO [27]R5030050+5053.459.344.444.260.5
— w/SegGen (ours)R5030050+5054.0 (+0.6)60.2 (+0.9)44.7 (+0.3)45.4 (+1.2)61.5 (+1.0)
MaskFormer [8]Swin-L10030052.758.544.040.164.8
OneFormer [19]Swin-L15010057.964.448.049.067.4
Mask2Former [7]Swin-L20010057.864.248.148.667.4
Mask DINOSwin-L3005058.365.148.050.6-
Mask2Former [7]Swin-L200100+10057.364.346.848.066.2
— w/SegGen (ours)Swin-L200100+10058.0 (+0.7)64.5 (+0.2)48.1 (+1.3)48.8 (+0.8)67.4 (+1.2)
Mask DINO [27]Swin-L30050+5058.665.448.350.467.0
— w/SegGen (ours)Swin-L30050+5059.3 (+0.7)65.9 (+0.5)49.3 (+1.0)51.1 (+0.7)68.1 (+1.1)
", + "bbox": [ + 218, + 130, + 782, + 318 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/40c149afd602b58615e14563ec5f4c51e77302485dcc7cec11d7eae6f21cb13d.jpg", + "image_caption": [ + "Fig.7: ADE20K mIoU of Mask2Former R50 with different augmentation probability $p_{\\mathrm{aug}}$ ." + ], + "image_footnote": [], + "bbox": [ + 217, + 378, + 375, + 454 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/f3f8b83eb7106c77ba68509af2cf5a40779e37be6b68068d8485d7691bc073c2.jpg", + "image_caption": [ + "Fig.8: ADE20K mIoU of Mask2Former R50 using different numbers of synthetic samples." + ], + "image_footnote": [], + "bbox": [ + 383, + 378, + 542, + 454 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/898e04b33595341fd958a6d25c84b442a4cea32b94e13090245d04d42e98ff92.jpg", + "table_caption": [], + "table_footnote": [ + "Table 4: Comparison with DiffuMask [48] on ADE20K: All models trained purely on synthetic data. mIoU* is defined in [48]." + ], + "table_body": "
MethodmIoU*
Pure Real Data83.3
DiffuMask (ICCV 2023)57.0
MaskSyn+ImgSyn (ours)82.2(+25.2)
MaskSyn (ours)73.2
ImgSyn (ours)80.0
", + "bbox": [ + 553, + 378, + 777, + 450 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "els purely on our synthetic data and fine-tune the models on real data. The results are demonstrated in Table 2. For a fair comparison, we compare with models pretrained on purely real data with the same training settings. Compared with the real data pre-trained baselines, SegGen consistently brings performance gains on all metrics with both R50 and Swin-L backbones, achieving new state-of-the-art performance without extra real data. Notably, for Mask DINO with Swin-L backbone, PQ is improved by $+0.7$ , \"thing\" instance segmentation $\\mathrm{AP}_{\\mathrm{pan}}^{\\mathrm{Th}}$ is increased by $+0.7$ , and semantic segmentation $\\mathrm{mIoU}_{\\mathrm{pan}}$ is boosted by $+1.1$ . These noteworthy findings demonstrate that training with our synthetic data systematically improves performance across various segmentation tasks. We observe similar significant improvement on Mask2Former, further showing the effectiveness of our synthetic data on different models.", + "bbox": [ + 212, + 520, + 787, + 700 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "COCO Instance Segmentation The performance on instance segmentation, both before and after incorporating our SegGen synthetic data, is detailed in Table 3. Compared to the baselines pre-trained on real data, our technique is superior across all metrics. Specifically, there is a solid $+0.5$ improvement in AP with the R50 backbone, and a $+0.8$ increase when using the Swin-L backbone. Additional assessment on the LVIS [14]-annotated COCO validation set, whose annotations are more accurate than the original version, provides additional evidence of the effectiveness of SegGen, as it enhances the average precision (AP) by $+1.2$ when used with the larger Swin-L backbone.", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "SegGen", + "bbox": [ + 678, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/e515cf0a7239893925791bd18643546f4c2685abff5677f9f1d77b606abce828.jpg", + "table_caption": [], + "table_footnote": [ + "Table 3: Instance segmentation on COCO val2017 (80 categories): Synthetic data is used for pre-training. The gray rows correspond to the evaluations conducted on LVIS [14] which offers instance masks of higher quality. Our SegGen significantly improves the performance of instance segmentation for various sizes of objects compared with pre-training on real data." + ], + "table_body": "
MethodBackbone QueriesEpochsAP\\( AP^S \\)\\( AP^M \\)\\( AP^L \\)
Mask R-CNN [15]R50anchors40042.523.845.060.0
HTC [4]R50anchors3639.722.642.250.6
QueryInst [11]R503003640.623.442.552.8
MaskFormer [8]R5010030034.016.437.854.2
Mask2Former [7]R501005043.723.447.264.8
Mask2Former [7]R501005047.126.655.472.7
Mask2Former [7]R5010050+5044.123.447.766.1
— w/ SegGen (ours)R5010050+5044.6 (+0.5)24.0 (+0.6)48.2 (+0.5)66.2 (+0.1)
Mask2Former [7]R5010050+5047.927.856.373.5
— w/ SegGen (ours)R5010050+5048.4 (+0.5)28.3 (+0.6)57.4 (+1.1)74.1 (+0.6)
Swin-HTC++ [34]Swin-Lanchors7249.531.052.467.2
QueryInst [11]Swin-L3005048.930.852.668.3
Oneformer [19]Swin-L15010048.9---
Mask2Former [7]Swin-L20010050.129.953.972.1
Mask2Former [7]Swin-L20010054.335.763.379.1
Mask2Former [7]Swin-L200100+10049.529.253.870.5
— w/ SegGen (ours)Swin-L200100+10050.3 (+0.8)31.2 (+2.0)54.3 (+0.5)72.2 (+1.7)
Mask2Former [7]Swin-L200100+10053.635.262.677.8
— w/ SegGen (ours)Swin-L200100+10054.8 (+1.2)35.9 (+0.8)64.2 (+1.6)79.2 (+1.5)
", + "bbox": [ + 218, + 143, + 782, + 364 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Comparison with Segmentation Data Generation Method In Table 4, we compare various versions of our SegGen with DiffuMask [48], which is a recently-proposed (ICCV2023) segmentation data generation approach. All methods employ the Mask2Former R50 model and are purely trained on synthetic samples. We adhere to the evaluation setting designed by DiffuMask where the IoU for three common classes are examined. Our method markedly surpasses DiffuMask by $+25.2$ mIoU and demonstrates performance comparable to training purely with real data, highlighting the unprecedented quality of our synthetic data.", + "bbox": [ + 212, + 441, + 787, + 561 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Generalization Ability on Unseen Domains We visualize the segmentation results of Mask2Former R50 on images from PASCAL val dataset [10] in Fig. 1 and Fig. 9, comparing models trained both with and without our synthetic data. The models are trained on COCO or ADE20K. When trained using our extensively varied synthetic data, the segmentation model demonstrates notably improved performance on the unseen domain. We further study the segmentation results on AI-generated images, which are synthesized using different image generation models, in the supplementary materials.", + "bbox": [ + 212, + 564, + 787, + 685 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Visual Analysis of Generated Samples by MaskSyn We randomly select some samples generated by MaskSyn in Fig. 4 (more in supplementary materials). The generative models are trained on ADE20K. Given the text prompts, our Text2Mask model can generate diverse segmentation masks, and the Mask2Img model produces realistic synthetic images with a good alignment with the masks and text prompts.", + "bbox": [ + 212, + 686, + 787, + 777 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Visual Analysis of Generated Samples by ImgSyn We showcase the outputs by ImgSyn in Fig. 5 (more in supplementary materials). The generated images are in good agreement with the text prompts and human-labeled segmentation masks. Furthermore, we compare the mask-image alignment of the", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Hanrong Ye et al.", + "bbox": [ + 271, + 114, + 390, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/4fc732a14710873ab1877fa725b0cfb655db2a34ddf98116257f16a1079b58a4.jpg", + "image_caption": [ + "Fig. 9: Generalization Ability on unseen domains: Segmentation outputs on unseen domain (images from PASCAL [10]). The models are trained on ADE20K. Training with our SegGen demonstrates enhanced robustness towards unfamiliar domains." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 279 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/4ced0171a20b9f2f477d4712d991c9dd9a7255dd1e66fdfbe01891e4199b35eb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodIterations Real SamplesmIoUMethodIterations Real SamplesmIoU
Baseline160K2021047.2Baseline100K2021044.4
- w/ SegGen MaskSyn160K2021048.5- w/ SegGen MaskSyn100K2021044.8
- w/ SegGen ImgSyn160K2021049.3- w/ SegGen ImgSyn100K2021045.4
- w/ SegGen ImgSyn + MaskSyn160K2021049.9- w/ SegGen ImgSyn + MaskSyn100K2021045.6
Baseline160K100023.7Baseline100K100023.6
- w/ SegGen MaskSyn160K100026.5- w/ SegGen MaskSyn100K100026.0
- w/ SegGen ImgSyn160K100024.1- w/ SegGen ImgSyn100K100023.9
- w/ SegGen ImgSyn + MaskSyn160K100027.1- w/ SegGen ImgSyn + MaskSyn100K100026.9
", + "bbox": [ + 217, + 321, + 779, + 412 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 5: Ablation Study of MaskSyn and ImgSyn on ADE20K: Both MaskSyn and ImgSyn notably enhance the performance of Mask2Former and HRNet, while MaskSyn has a more pronounced impact when there are fewer real samples available.", + "bbox": [ + 215, + 414, + 782, + 455 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "real training samples and our synthetic training samples in Fig. 6. Our synthetic samples exhibit superior alignment quality in many cases, compared to human annotations which are often imperfect on the boundaries.", + "bbox": [ + 215, + 465, + 782, + 510 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.3 Quantitative Analysis", + "text_level": 1, + "bbox": [ + 215, + 537, + 439, + 553 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effectiveness of MaskSyn and ImgSyn We evaluate the data generated by MaskSyn and ImgSyn separately and jointly, by examining their impacts on the performance of segmentation models, as shown in Table 5. The experiments are conducted on ADE20K with Mask2Former R50 and HRNet W48. We find that the training samples generated by both MaskSyn and ImgSyn bring significant performance gains compared with the baseline. When combining these two types of synthetic data together, the segmentation model can achieve the best performance. We delve deeper into a scenario where only 1,000 real training samples are available in the training of both generative models and segmentation models. It is observed that MaskSyn substantially enhances the segmentation results in situations with limited real data. This could be attributed to the ability of MaskSyn to amplify data diversity by creating entirely new segmentation masks and images. Moreover, the significant performance improvements achieved by our SegGen, trained on merely 1,000 real samples, demonstrate the data efficiency and robustness of our proposal.", + "bbox": [ + 212, + 566, + 787, + 792 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Influence of Synthetic Data Augmentation Probability $p_{\\mathrm{aug}}$ on ADE20K We investigate the effect of $p_{\\mathrm{aug}}$ in Fig. 7. When $p_{\\mathrm{aug}} = 0$ , it implies training using only real data, whereas $p_{\\mathrm{aug}} = 100\\%$ means training entirely with synthetic", + "bbox": [ + 212, + 794, + 792, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "SegGen", + "bbox": [ + 678, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/1e15dcfbd1b7f0031c4fe4d0fce82b13a40dd934c2771ba5fcf21e302b3319da.jpg", + "table_caption": [], + "table_footnote": [ + "(a) ADE20K Semantic Segmentation: Using synthetic data as random data augmentation obtains the best performance on ADE20K." + ], + "table_body": "
MethodIterationsmIoU
Syn. Pre-Train160K+160K47.2
Syn. Data Aug160K49.9
", + "bbox": [ + 228, + 143, + 485, + 184 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/1474d5246ce5edd087c72eb9a827a2a186e5d75bfb8b975d41d981a734d0cb2e.jpg", + "table_caption": [], + "table_footnote": [ + "(b) COCO Panoptic Segmentation and Instance Segmentation: Using synthetic data for pre-training achieves better performance on COCO dataset." + ], + "table_body": "
MethodEpochsPQAP
Syn. Pre-Training50+5052.744.6
Syn. Data Aug5051.343.2
", + "bbox": [ + 506, + 143, + 771, + 183 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Table 6: Different training strategies with synthetic data using Mask2Former R50.", + "bbox": [ + 220, + 228, + 777, + 242 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "data. The best results are achieved at $p_{\\mathrm{aug}} = 60\\%$ . It is noteworthy that training exclusively on our synthetic data results in an impressive $44.8~\\mathrm{mIoU}$", + "bbox": [ + 212, + 247, + 784, + 276 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Influence of Synthetic Data Size To evaluate the impact of synthetic data volume, we use different numbers of synthetic training samples produced by ImgSyn in training and report the results in Fig. 8. $p_{\\mathrm{aug}}$ is fixed at $60\\%$ . We notice a substantial improvement in performance when increasing the synthetic data quantity from 20K to 200K, underscoring the importance of collecting a significantly larger synthetic dataset to avoid overfitting. The performance appears to plateau after 600K. More studies on the scale of synthetic data are presented in the supplementary materials.", + "bbox": [ + 212, + 277, + 785, + 398 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Influence of Training Strategies We conduct an ablation study using different training strategies with synthetic data on ADE20K (Table 6a) and COCO (Table 6b). We find that using the synthetic data augmentation strategy achieves better performance on ADE20K semantic segmentation, While on COCO panoptic and instance segmentation, the synthetic pre-training strategy works better. We believe that due to the limited size of the ADE20K training set, which contains merely $\\sim 20\\mathrm{k}$ images, segmentation models easily overfit the scarcely-available training images and segmentation layouts. Utilizing synthetic data (including synthetic masks and images) for data augmentation can help mitigate such an overfitting problem. On the other hand, for COCO, given its larger training set size (around 100K images), overfitting is less of a concern. Hence, it is sufficient to provide the COCO models with good initial weights pre-trained on our synthetic data. Meanwhile, previous work [14] finds that the annotation accuracy of COCO is significantly more biased than ADE20K, a finding that aligns with our visualization results depicted in Fig. 6. A domain gap exists between the COCO data and our high-quality synthetic data. Hence, employing our method for pre-training and incorporating real data for fine-tuning on COCO is a more favorable approach.", + "bbox": [ + 212, + 398, + 787, + 672 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 689, + 359, + 705 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We present SegGen, a highly-effective data synthesis method for image segmentation. SegGen introduces a revolutionary segmentation data generation framework, and proposes two data generation approaches, namely MaskSyn and ImgSyn, to synthesize segmentation masks and images, with the help of the designed text-to-mask and mask-to-image generation models. Our method not only strongly enhances the model performance on semantic, panoptic, and instance segmentation benchmarks, but also substantially improves segmentation generalization ability in unseen data distribution.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Hanrong Ye et al.", + "bbox": [ + 271, + 114, + 390, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 217, + 143, + 393, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This research is supported in part by the Early Career Scheme of the Research Grants Council (RGC) of the Hong Kong SAR under grant No. 26202321, SAIL Research Project, HKUST-Zeekr Collaborative Research Fund, HKUST-WeBank Joint Lab Project, and Tencent Rhino-Bird Focused Research Program.", + "bbox": [ + 212, + 178, + 787, + 239 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 262, + 323, + 277 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Baranchuk, D., Rubachev, I., Voynov, A., Khrulkov, V., Babenko, A.: Label-efficient semantic segmentation with diffusion models. In: ICLR (2022)", + "2. Brock, A., Donahue, J., Simonyan, K.: Large scale gan training for high fidelity natural image synthesis. In: ICLR (2019)", + "3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: ECCV (2020)", + "4. Chen, K., Pang, J., Wang, J., Xiong, Y., Li, X., Sun, S., Feng, W., Liu, Z., Shi, J., Ouyang, W., et al.: Hybrid task cascade for instance segmentation. In: CVPR (2019)", + "5. Chen, L.C., Papandreou, G., Kokkinos, I., Murphy, K., Yuille, A.L.: Semantic image segmentation with deep convolutional nets and fully connected crfs. In: ICLR (2015)", + "6. Chen, T., Li, L., Saxena, S., Hinton, G., Fleet, D.J.: A generalist framework for panoptic segmentation of images and videos. arXiv preprint arXiv:2210.06366 (2022)", + "7. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: CVPR (2022)", + "8. Cheng, B., Schwing, A.G., Kirillov, A.: Per-pixel classification is not all you need for semantic segmentation. In: NeurIPS (2021)", + "9. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. In: NeruIPS (2021)", + "0. Everingham, M., Eslami, S.A., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The PASCAL visual object classes challenge: A retrospective. IJCV (2015)", + "1. Fang, Y., Yang, S., Wang, X., Li, Y., Fang, C., Shan, Y., Feng, B., Liu, W.: Instances as queries. In: ICCV (2021)", + "2. Forever, A.: Kandinsky-2. GitHub repository (2023), https://github.com/ai-forever/Kandinsky-2", + "3. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. CACM (2020)", + "4. Gupta, A., Dollar, P., Girshick, R.: LVIS: A dataset for large vocabulary instance segmentation. In: ICCV (2019)", + "5. He, K., Gkioxari, G., Dolkar, P., Girshick, R.: Mask R-CNN. In: ICCV (2017)", + "6. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016)", + "7. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeruIPS (2020)", + "8. Huang, L., Chen, D., Liu, Y., Yujun, S., Zhao, D., Jingren, Z.: Composer: Creative and controllable image synthesis with composable conditions. arXiv preprint arxiv:2302.09778 (2023)" + ], + "bbox": [ + 225, + 295, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "SegGen", + "bbox": [ + 678, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "19. Jain, J., Li, J., Chiu, M.T., Hassani, A., Orlov, N., Shi, H.: Oneformer: One transformer to rule universal image segmentation. In: CVPR (2023)", + "20. Ji, Y., Chen, Z., Xie, E., Hong, L., Liu, X., Liu, Z., Lu, T., Li, Z., Luo, P.: Ddp: Diffusion model for dense visual prediction. arXiv preprint arXiv:2303.17559 (2023)", + "21. Kang, M., Zhu, J.Y., Zhang, R., Park, J., Shechtman, E., Paris, S., Park, T.: Scaling up gans for text-to-image synthesis. In: CVPR (2023)", + "22. Karras, T., Aittala, M., Aila, T., Laine, S.: Elucidating the design space of diffusion-based generative models. In: NeruIPS (2022)", + "23. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: CVPR (2019)", + "24. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013)", + "25. Kirillov, A., He, K., Girshick, R., Rother, C., Dlár, P.: Panoptic segmentation. In: CVPR (2019)", + "26. Li, D., Ling, H., Kim, S.W., Kreis, K., Fidler, S., Torralba, A.: Bigdatasetgan: Synthesizing imagenet with pixel-wise annotations. In: CVPR (2022)", + "27. Li, F., Zhang, H., Xu, H., Liu, S., Zhang, L., Ni, L.M., Shum, H.Y.: Mask dino: Towards a unified transformer-based framework for object detection and segmentation. In: CVPR (2023)", + "28. Li, J., Li, D., Savarese, S., Hoi, S.: BLIP-2: bootstrapping language-image pretraining with frozen image encoders and large language models. In: ICML (2023)", + "29. Li, X., Zhong, Z., Wu, J., Yang, Y., Lin, Z., Liu, H.: Expectation-maximization attention networks for semantic segmentation. In: ICCV (2019)", + "30. Li, X., Ding, H., Zhang, W., Yuan, H., Pang, J., Cheng, G., Chen, K., Liu, Z., Loy, C.C.: Transformer-based visual segmentation: A survey. arXiv preprint arXiv:2304.09854 (2023)", + "31. Li, Y., Liu, H., Wu, Q., Mu, F., Yang, J., Gao, J., Li, C., Lee, Y.J.: Gligen: Open-set grounded text-to-image generation. In: CVPR (2023)", + "32. Li, Z., Zhou, Q., Zhang, X., Zhang, Y., Wang, Y., Xie, W.: Open-vocabulary object segmentation with diffusion models. In: ICCV (2023)", + "33. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólár, P., Zitnick, C.L.: Microsoft COCO: Common objects in context. In: ECCV (2014)", + "34. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. arXiv:2103.14030 (2021)", + "35. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: CVPR (2015)", + "36. Mou, C., Wang, X., Xie, L., Wu, Y., Zhang, J., Qi, Z., Shan, Y., Qie, X.: T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453 (2023)", + "37. Nguyen, Q.N.T.V.A.T.K.: Dataset diffusion: Diffusion-based synthetic dataset generation for pixel-level semantic segmentation. arXiv preprint arXiv:2309.14303 (2023)", + "38. PNVR, K., Singh, B., Ghosh, P., Siddiquie, B., Jacobs, D.: Ld-znet: A latent diffusion approach for text-based image segmentation. In: ICCV (2023)", + "39. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952 (2023)", + "40. Qi, L., Kuen, J., Wang, Y., Gu, J., Zhao, H., Torr, P., Lin, Z., Jia, J.: Open world entity segmentation. TPAMI (2022)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Hanrong Ye et al.", + "bbox": [ + 271, + 114, + 390, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "41. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)", + "42. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. In: NeruIPS (2022)", + "43. Sauer, A., Karras, T., Laine, S., Geiger, A., Aila, T.: StyleGAN-T: Unlocking the power of GANs for fast large-scale text-to-image synthesis. In: ICML (2023)", + "44. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: ICML (2015)", + "45. Wang, H., Cao, J., Anwer, R.M., Xie, J., Khan, F.S., Pang, Y.: Dformer: Diffusion-guided transformer for universal image segmentation. arXiv preprint arXiv:2306.03437 (2023)", + "46. Wang, J., Sun, K., Cheng, T., Jiang, B., Deng, C., Zhao, Y., Liu, D., Mu, Y., Tan, M., Wang, X., Liu, W., Xiao, B.: Deep high-resolution representation learning for visual recognition. TPAMI (2019)", + "47. Wu, W., Zhao, Y., Chen, H., Gu, Y., Zhao, R., He, Y., Zhou, H., Shou, M.Z., Shen, C.: Datasetsd: Synthesizing data with perception annotations using diffusion models. arXiv preprint arXiv:2308.06160 (2023)", + "48. Wu, W., Zhao, Y., Shou, M.Z., Zhou, H., Shen, C.: Diffumask: Synthesizing images with pixel-level annotations for semantic segmentation using diffusion models. In: ICCV (2023)", + "49. Xie, J., Li, W., Li, X., Liu, Z., Ong, Y.S., Loy, C.C.: Mosaic fusion: Diffusion models as data augmenters for large vocabulary instance segmentation. arXiv preprint arXiv:2309.13042 (2023)", + "50. Xu, J., Liu, S., Vahdat, A., Byeon, W., Wang, X., De Mello, S.: Open-vocabulary panoptic segmentation with text-to-image diffusion models. In: CVPR (2023)", + "51. Yang, L., Xu, X., Kang, B., Shi, Y., Zhao, H.: Freemask: Synthetic images with dense annotations make stronger segmentation models. In: NeurIPS (2023)", + "52. Ye, H., Xu, D.: Taskprompter: Spatial-channel multi-task prompting for dense scene understanding. In: ICLR (2023)", + "53. Zeng, Y., Lin, Z., Zhang, J., Liu, Q., Collomosse, J., Kuen, J., Patel, V.M.: Scenecomposer: Any-level semantic image synthesis. In: CVPR (2023)", + "54. Zhang, L., Anyi, R., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: ICCV (2023)", + "55. Zhang, W., Pang, J., Chen, K., Loy, C.C.: K-net: Towards unified image segmentation. In: NeurIPS (2021)", + "56. Zhang, Y., Ling, H., Gao, J., Yin, K., Lafleche, J.F., Barriuso, A., Torralba, A., Fidler, S.: Datasetgan: Efficient labeled data factory with minimal human effort. In: CVPR (2021)", + "57. Zhao, H., Shi, J., Qi, X., Wang, X., Jia, J.: Pyramid scene parsing network. In: CVPR (2017)", + "58. Zhao, S., Chen, D., Chen, Y.C., Bao, J., Hao, S., Yuan, L., Wong, K.Y.K.: Uncontrolnet: All-in-one control to text-to-image diffusion models. NeurIPS (2023)", + "59. Zhao, W., Rao, Y., Liu, Z., Liu, B., Zhou, J., Lu, J.: Unleashing text-to-image diffusion models for visual perception. In: ICCV (2023)", + "60. Zhou, B., Zhao, H., Puig, X., Fidler, S., Barriuso, A., Torralba, A.: Scene parsing challenge 2016. http://sceneParsing.csail.mit.edu/index_challenge.htm (2016)", + "61. Zhou, B., Zhao, H., Puig, X., Fidler, S., Barriuso, A., Torralba, A.: Scene parsing through ADE20K dataset. In: CVPR (2017)" + ], + "bbox": [ + 212, + 146, + 787, + 829 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "SegGen", + "bbox": [ + 678, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/SegGen_ Supercharging Segmentation Models with Text2Mask and Mask2Img Synthesis/acac7544-25ea-4a13-aba6-d24e96031cd3_model.json b/2024/SegGen_ Supercharging Segmentation Models with Text2Mask and Mask2Img Synthesis/acac7544-25ea-4a13-aba6-d24e96031cd3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c1582299cb6826c4790757841fb6a6f32435d27c --- /dev/null +++ b/2024/SegGen_ Supercharging Segmentation Models with Text2Mask and Mask2Img Synthesis/acac7544-25ea-4a13-aba6-d24e96031cd3_model.json @@ -0,0 +1,2192 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.244, + 0.141, + 0.761, + 0.187 + ], + "angle": 0, + "content": "SegGen: Supercharging Segmentation Models with Text2Mask and Mask2Img Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.212, + 0.761, + 0.228 + ], + "angle": 0, + "content": "Hanrong Ye\\(^{1}\\), Jason Kuen\\(^{2}\\), Qing Liu\\(^{2}\\), Zhe Lin\\(^{2}\\), Brian Price\\(^{2}\\), Dan Xu\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.443, + 0.239, + 0.558, + 0.253 + ], + "angle": 0, + "content": "1 CSE, HKUST" + }, + { + "type": "text", + "bbox": [ + 0.441, + 0.254, + 0.563, + 0.267 + ], + "angle": 0, + "content": "2 Adobe Research" + }, + { + "type": "list", + "bbox": [ + 0.441, + 0.239, + 0.563, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.308, + 0.74, + 0.653 + ], + "angle": 0, + "content": "Abstract. We present SegGen, a new data generation approach that pushes the performance boundaries of state-of-the-art image segmentation models. One major bottleneck of previous data synthesis methods for segmentation is the design of \"segmentation labeler module\", which is used to synthesize segmentation masks for images [48]. The segmentation labeler modules, which are segmentation models by themselves, bound the performance of downstream segmentation models trained on the synthetic masks. These methods encounter a \"chicken or egg dilemma\" and thus fail to outperform existing segmentation models. To address this issue, we propose a novel method that reverses the traditional data generation process: we first (i) generate highly diverse segmentation masks that match real-world distribution from text prompts, and then (ii) synthesize realistic images conditioned on the segmentation masks. In this way, we avoid the need for any segmentation labeler module. SegGen integrates two data generation strategies, namely MaskSyn and ImgSyn, to largely improve data diversity in synthetic masks and images. Notably, the high quality of our synthetic data enables our method to outperform the previous data synthesis method [48] by \\(+25.2\\) mIoU on ADE20K when trained with pure synthetic data. On the highly competitive ADE20K and COCO benchmarks, our data generation method markedly improves the performance of state-of-the-art segmentation models in semantic segmentation, panoptic segmentation, and instance segmentation. Moreover, experiments show that training with our synthetic data makes the segmentation models more robust towards unseen data domains, including real-world and AI-generated images." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.668, + 0.731, + 0.682 + ], + "angle": 0, + "content": "Keywords: Image Segmentation \\(\\cdot\\) Image Generation \\(\\cdot\\) Deep Learning" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.732, + 0.375, + 0.748 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Image segmentation explores the identification of objects in visual inputs at the pixel level. Based on the different emphases on category and instance membership information, researchers have divided image segmentation into several tasks [5,25,35,40]. For example, semantic segmentation studies pixel-level understanding of object categories, instance segmentation focuses on instance grouping" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.392, + 0.129 + ], + "angle": 0, + "content": "Hanrong Ye et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.788, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.306, + 0.789, + 0.432 + ], + "angle": 0, + "content": "Fig. 1: Effectiveness of SegGen on various data domains: Through training with synthetic data generated by the proposed SegGen, we significantly boost the performance of state-of-the-art segmentation model Mask2Former [7] on evaluation benchmarks including ADE20K [60] and COCO [33], whilst making it more robust towards challenging images from other domains (the three columns on the left are from PASCAL [10]; the three on the right are synthesized by image generation model Kandinsky 2 [12]). SegGen outperforms the previous best data generation method (DiffuMask [48]) by a huge margin when models are trained on pure synthetic data. \"mIoU*\" is the average IoU metric defined by [48] which focuses on three common classes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.444, + 0.789, + 0.626 + ], + "angle": 0, + "content": "of pixels, while panoptic segmentation considers both. For all these segmentation tasks, obtaining high-quality annotation is challenging as every individual pixel requires human labeling, and a single image can contain millions of pixels. Therefore, compared to other public datasets like ImageNet-21K (with around 14M images), the prevailing human-annotated segmentation datasets are notably smaller. For example, ADE20K dataset [61] contains about 20K images in its training split, while COCO [33] has around 118K training images. Although there has been significant development in the structure of segmentation models [8,29,52,55,57], the limited size of training data hinders further performance enhancements and results in inadequate generalization ability to handle images from unfamiliar domains, such as those from other scenes or synthesized by generative models as shown in the second row of Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.79, + 0.842 + ], + "angle": 0, + "content": "Inspired by the recent success of image generation [9, 23], researchers start to explore using generative models to enhance image segmentation [1]. A representative direction of this research focuses on synthesizing segmentation training data in a cost-effective manner [26]. In related methods, the synthetic segmentation masks are obtained from some manually designed \"segmentation labeler modules\". The segmentation labeler modules are essentially segmentation models by themselves. They are either small-scale segmentation networks as in DatasetGAN [56] and Grounded Diffusion [32], or post-processing methods on the image features as in DiffuMask [48]. The performance of the downstream segmentation models is constrained by the quality of the synthetic masks they are trained on, which in turn relies on the capabilities of those segmentation labeler modules. This is a \"chicken or egg dilemma\" and the performance bottleneck is the segmentation labeler module. Therefore, while they achieve encouraging success in settings with limited training data, their methods fail to deliver notable en" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.679, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SegGen" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.146, + 0.788, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.225, + 0.788, + 0.364 + ], + "angle": 0, + "content": "Fig. 2: Comparison with previous data generation methods for segmentation: (a) Earlier methods [48, 56] rely on segmentation labeler modules to produce segmentation masks for synthetic images. However, the performance of these downstream segmentation models, trained on synthetic data, is bounded by the capacity of the segmentation labeler modules. The segmentation labeler is a major bottleneck for the quality of the generated data. (b) We design a reverse pipeline: we first create diverse new masks from text prompts via a proposed Text2Mask generation model and then synthesize images conditioned on the segmentation masks. This methodology avoids any usage of segmentation labeler networks, resulting in significantly improved quality of our synthesis data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.381, + 0.785, + 0.412 + ], + "angle": 0, + "content": "hancements for the state-of-the-art models on the most popular segmentation benchmarks in settings using the complete training sets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.414, + 0.788, + 0.642 + ], + "angle": 0, + "content": "On the other hand, a handful of works [54, 58] have demonstrated that high-quality synthetic images can be generated by conditioning text-to-image models on dense input maps (e.g., segmentation masks or canny edge). Moreover, these dense-input conditional models provide a strong alignment between the dense input label maps and generated images. These two qualities motivate us to leverage the powerful capabilities of such models for effective segmentation data generation. To this end, we propose a novel segmentation data generation method, coined as SegGen, for generating high-quality segmentation training data. We reverse the traditional data generation pipeline for segmentation: we first synthesize high-quality segmentation masks from a proposed text-to-mask generation model, and then generate images based on the segmentation masks using a mask-to-image generation model. In this new framework, we avoid using any segmentation labeler modules, which results in significantly improved data quality. Fig. 2 shows the difference between our method and previous data generation methods for segmentation [48, 56]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.789, + 0.841 + ], + "angle": 0, + "content": "As each training sample of image segmentation consists of two components: segmentation masks and the corresponding image, we develop two novel data generation approaches, emphasizing improvements in two distinct aspects of data diversity: (i) segmentation masks and (ii) images. The first data generation approach, named MaskSyn, centers around the generation of new segmentation masks. It learns a text-to-mask (Text2Mask) generation model to produce completely new segmentation masks given text prompts. Then, it learns a mask-to-image (Mask2Img) generation model to synthesize images that align with the synthetic segmentation masks. The second data generation approach, named ImgSyn, utilizes the above-mentioned Mask2Img model to synthesize new images given human-annotated segmentation masks. With MaskSyn and ImgSyn, we can readily generate a vast array of diverse and high-quality synthetic training data. The combined synthetic data is used to train segmenta" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.392, + 0.129 + ], + "angle": 0, + "content": "Hanrong Ye et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.344 + ], + "angle": 0, + "content": "tion models in conjunction with real training samples from human-annotated datasets. Experiments show that our synthetic data can significantly boost the performance of the image segmentation models on challenging benchmarks including ADE20K semantic segmentation, COCO panoptic segmentation, and COCO instance segmentation, achieving new state-of-the-art performances without using extra human-annotated data. Notably, SegGen remarkably outperforms the previous segmentation data generation method [48] by \\(+25.2\\) mIoU when trained with pure synthetic data on ADE20K. Our method boosts the mIoU of Mask2Former by \\(+2.7\\) (R50) and \\(+1.3\\) (Swin-L) on the ADE20K semantic segmentation benchmark. Furthermore, segmentation models trained on our synthetic data exhibit a remarkably stronger ability to generalize across unfamiliar image domains, including real images from other distributions and machine-generated images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.344, + 0.788, + 0.571 + ], + "angle": 0, + "content": "We summarize the contribution of this work in three points: (i) We propose a revolutionary generation framework that reverses the traditional segmentation data generation pipeline and solves the \"chicken or egg dilemma\". The new framework enables us to produce high-quality segmentation training data at scale, thus enabling the training of more powerful image segmentation models. (ii) We introduce two effective generative models, one for text-to-mask generation and the other for mask-to-image generation. Based on these models, we propose two novel segmentation training data generation approaches, namely MaskSyn and ImgSyn. They significantly improve the data diversity, with MaskSyn focusing on new segmentation masks and ImgSyn on new images. (iii) SegGen successfully improves the performance of the leading-edge segmentation models across the highly competitive benchmarks on ADE20K and COCO. Moreover, SegGen enhances the generalization ability of segmentation models towards unseen image domains. Rigorous experiments, including ablation study and peer comparison, strongly suggest the effectiveness of the proposed method." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.595, + 0.388, + 0.612 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Generation for Segmentation Image segmentation is one of the most studied visual perception problems [30]. In recent years, there has been a surge in efforts to harness the capabilities of generative models for segmentation tasks. These efforts can be broadly classified into three categories based on their methodologies: (i) Extracting visual features from generative models for segmentation [1, 38, 50, 59]. These methods harvest the strong representation ability of diffusion models trained on large-scale datasets but are limited in the precision of predicted masks. (ii) Formulating segmentation tasks directly as generative models [6, 20, 45]. Although these methods propose exciting new model architectures, they usually consume higher computational costs while showing unimproved performance compared with conventional segmentation models. (iii) Synthesizing segmentation training data using generative models [26, 32, 37, 47-49, 56]. While these methods have showcased commendable results against their respective baselines especially when the training data is highly limited, they have" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.679, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SegGen" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.344 + ], + "angle": 0, + "content": "yet to exhibit notably superior performance on the most rigorous benchmarks including ADE20K (150 categories) [61] and COCO (133 categories) [33] under fully-supervised setting. A primary concern with these techniques is the subpar quality of the generated segmentation masks. This stems from their dependence on the bottleneck segmentation labeler modules during the synthetic mask generation process. Therefore, a concurrent work [51] abandons the generation of segmentation masks and directly generates images based on existing masks, which is highly limited in mask diversity. In contrast, our data generation workflow reverses the traditional data generation pipeline: we first generate segmentation masks from text, and then synthesize images based on segmentation masks. Our method avoids the need for any segmentation labeler which results in largely improved data quality. Our SegGen can strongly enhance the existing state-of-the-art segmentation models on different challenging benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.344, + 0.788, + 0.542 + ], + "angle": 0, + "content": "Conditional Image Synthesis Within the realm of conditional image synthesis, Generative Adversarial Networks [2, 13, 21, 23, 43], Variational Autoencoders [24], and Diffusion Models [9, 17, 42, 44] have been at the forefront. Recently, the open-source research community has shown a burgeoning interest in the latent diffusion-based Stable Diffusion (SD) series for text-to-image synthesis [41]. The most recent iteration, SDXL, introduced by [39], expands the model capacity, yielding significantly enhanced results. Therefore, we build SegGen upon SDXL. Regarding mask-to-image generation models [18, 31, 53], ControlNet [54] and T2I-Adapter [36] suggest freezing the parameters of the SD model and introducing a set of more compact, learnable modules. This approach enables image generation conditioned on the given segmentation masks. We adopt the structure of ControlNet in our mask-to-image generation model. To the best of our knowledge, there have been no prior endeavors on text-to-mask generation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.561, + 0.331, + 0.577 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.592, + 0.341, + 0.606 + ], + "angle": 0, + "content": "3.1 Overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.615, + 0.788, + 0.706 + ], + "angle": 0, + "content": "SegGen is designed to synthesize high-quality training samples for improving the performance of segmentation models. The overall workflow is shown on the left of Fig. 3. We first train SegGen with human-annotated training samples from public datasets. After training, we use SegGen to produce new segmentation training samples at scale. The generated training samples are incorporated into the training process of segmentation models to enhance the model performance." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.726, + 0.414, + 0.741 + ], + "angle": 0, + "content": "3.2 Models in SegGen" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.842 + ], + "angle": 0, + "content": "As shown in Fig. 3, we first utilize a captioner model to extract captions of the real training images as text prompts from the target dataset. The text prompts will condition the data generation process. Then, two conditional generative models are introduced: a text-to-mask (Text2Mask) generation model and a mask-to-image (Mask2Img) generation model. Both generative models are built upon the SDXL model [39] which provides top-notch image generation quality." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.129 + ], + "angle": 0, + "content": "Hanrong Ye et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.149, + 0.789, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.271, + 0.788, + 0.397 + ], + "angle": 0, + "content": "Fig.3: Illustration of the workflow of our proposed SegGen: We introduce two generative models: a text-to-mask (Text2Mask) generation model and a mask-to-image (Mask2Img) generation model, based on which we design two approaches for synthesizing segmentation training samples: MaskSyn and ImgSyn. (a) MaskSyn focuses on generating new segmentation masks. It first extracts the caption of the real image as a text prompt and uses it to generate new masks with the Text2Mask model. Then, the new masks and text prompt are fed into the Mask2Img model to produce the corresponding new images. (b) ImgSyn focuses on the synthesis of new images. It inputs human-labeled masks and text prompts into the Mask2Img model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.41, + 0.788, + 0.485 + ], + "angle": 0, + "content": "Captioner Model To obtain image captions of existing training samples, we employ BLIP2-FlanT5xxl [28] model, which is a state-of-the-art vision-language model. We feed the text prompt \"Question: What are shown in the photo? Answer:\" and image as input to the model. The responses serve as text prompts to condition the following generation process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.487, + 0.789, + 0.715 + ], + "angle": 0, + "content": "Text2Mask Model We design a Text2Mask model in SegGen for generating diverse segmentation masks based on given text prompts. To leverage the generation capacity of text-to-image generation models pre-trained on large-scale datasets, we encode the segmentation masks (the pixel values are category IDs) as three-channel RGB-like color maps, where one color represents a certain category. From our experiments, the color map reconstructed by VAE [41] in SDXL appears almost indistinguishable from the original input as shown in the supplementary materials. Therefore, we can directly fine-tune the text-to-image SDXL-base model with [text, segmentation color map] training pairs, which are from the public image segmentation dataset (e.g. ADE20K). During sampling, our Text2Mask model can generate diverse color maps conditioned on text prompts, which are subsequently converted into segmentation masks. Formally, suppose the input text prompt is \\(\\mathbf{T}\\), the target height and width are \\(H\\) and \\(W\\), the synthesized color map is \\(\\mathbf{C}_{\\mathrm{syn}} \\in \\mathbb{W}^{H \\times W \\times 3}\\), and the synthesized segmentation map (with \\(N\\) masks) is \\(\\mathbf{M}_{\\mathrm{syn}} \\in \\mathbb{W}^{H \\times W \\times N}\\), the generation process is:" + }, + { + "type": "equation", + "bbox": [ + 0.407, + 0.729, + 0.785, + 0.764 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {C} _ {\\mathrm {s y n}} = \\operatorname {T e x t 2 M a s k} (\\mathbf {T}), \\\\ \\mathbf {M} _ {\\text {s y n}} = f _ {\\text {c o l o r} \\rightarrow \\text {m a s k}} \\left(\\mathbf {C} _ {\\text {s y n}}\\right), \\tag {1} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.779, + 0.788, + 0.841 + ], + "angle": 0, + "content": "where \\( f_{\\mathrm{color} \\to \\mathrm{mask}}: \\mathbb{W}^{H \\times W \\times 3} \\to \\mathbb{W}^{H \\times W \\times N} \\) is the function that projects the color maps to segmentation masks. Specifically, for each pixel on the color maps, we identify its nearest color (in Euclidean space) in the aforementioned lookup table, and assign the corresponding class to the pixel in the segmentation masks." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.679, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SegGen" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.787, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.287, + 0.788, + 0.342 + ], + "angle": 0, + "content": "Fig. 4: Generated samples by MaskSyn on ADE20K: The third row overlays the mask and the image together to demonstrate the alignment between them. The generated segmentation masks and images demonstrate high perceptual quality and excellent alignment (see more samples in supplementary materials)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.366, + 0.789, + 0.518 + ], + "angle": 0, + "content": "Mask2Img Model The goal of the Mask2Img model is to synthesize new images that align well with the given segmentation masks and text prompts. Specifically we adopt the structure of ControlNet [54]: we freeze the pre-trained weights of the SDXL-base model and train an additional side network for mask-conditioned image generation. It simultaneously keeps the generalization ability of the pretrained diffusion model and provides excellent controllable generation ability. The Mask2Img model is trained with the [text, segmentation color map, image] triplets gathered from the training splits of the target datasets. We denote the input segmentation map as \\(\\mathbf{M} \\in \\mathbb{W}^{H \\times W \\times N}\\), the color map as \\(\\mathbf{C} \\in \\mathbb{W}^{H \\times W \\times 3}\\), the synthetic image as \\(\\mathbf{I}_{\\mathrm{syn}} \\in \\mathbb{W}^{H \\times W \\times 3}\\), and the generation process is:" + }, + { + "type": "equation", + "bbox": [ + 0.415, + 0.531, + 0.785, + 0.565 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {C} = f _ {\\text {m a s k} \\rightarrow \\text {c o l o r}} (\\mathbf {M}), \\\\ \\mathbf {I} _ {\\text {s y n}} = \\operatorname {M a s k 2 I m g} (\\mathbf {T}, \\mathbf {C}), \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.578, + 0.788, + 0.716 + ], + "angle": 0, + "content": "where \\( f_{\\mathrm{mask}\\rightarrow \\mathrm{color}}:\\mathbb{W}^{H\\times W\\times N}\\to \\mathbb{W}^{H\\times W\\times 3} \\) is the function to convert the segmentation masks into a color map. For semantic segmentation, the value of each pixel on the segmentation mask corresponds to a category ID, allowing us to convert the masks directly into an RGB color map using a pre-defined lookup table. For panoptic and instance segmentation, after mapping the category IDs to color maps, it is essential to outline each segment with a special edge color on the color map. This ensures the model recognizes the specific instance it belongs to. The segmentation map \\( \\mathbf{M} \\) can be human-annotated or synthetic (i.e., \\( \\mathbf{M}_{\\mathrm{syn}} \\) from Text2Mask Model as shown in Eq. 1)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.739, + 0.403, + 0.753 + ], + "angle": 0, + "content": "3.3 Data Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "With the aforementioned generative models, SegGen proposes two approaches for synthesizing new segmentation training samples: MaskSyn and ImgSyn. MaskSyn focuses on enhancing the diversity of synthetic segmentation masks, whereas ImgSyn concentrates on diversifying synthetic images. These generation approaches are illustrated on the right of Fig. 3." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.129 + ], + "angle": 0, + "content": "Hanrong Ye et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.787, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.242, + 0.788, + 0.285 + ], + "angle": 0, + "content": "Fig. 5: Generated samples by ImgSyn on ADE20K: The generated images exhibit remarkable realism and align well with the human-labeled masks and text prompts (see more samples in supplementary materials)." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.293, + 0.787, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.406, + 0.788, + 0.462 + ], + "angle": 0, + "content": "Fig. 6: Zoom-in comparison of real images and our ImgSyn images: As highlighted in the circles, our synthetic images align better with the human-labeled masks in many cases because of the inaccuracies in annotations. The left 4 columns are from ADE20K and the right 4 from COCO." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.474, + 0.789, + 0.626 + ], + "angle": 0, + "content": "MaskSyn MaskSyn starts with a real training sample pair [image, segmentation masks] from a human-annotated segmentation dataset. It first extracts the caption of the image with the image captioner model. The caption serves as a text prompt and is used to generate a set of diverse new segmentation masks with the Text2Mask model following Eq. 1. Subsequently, the new segmentation masks and the corresponding text prompt are fed into the Mask2Img model to generate a new image that aligns well with its mask. As such, each training sample crafted by MaskSyn includes both a novel segmentation mask and a new image. MaskSyn effectively increases the data diversity in segmentation masks to a great extent. Some generated samples are shown in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.79, + 0.841 + ], + "angle": 0, + "content": "ImgSyn Different from MaskSyn, ImgSyn focuses on increasing the data diversity of images based on human-annotated segmentation masks. For each real training sample pair [image, segmentation mask], it takes the human-annotated segmentation mask and the text prompt extracted from the image as input, and generates a set of varied images that align well with the human-annotated mask. In this way, new training samples consisting of human-labeled masks and new synthetic images are generated. ImgSyn can also be viewed as a kind of data augmentation method that enhances the data diversity on the image side. Our experiments reveal a remarkably high alignment between the synthetic images and their respective segmentation masks. Some synthesized results are showcased in Fig. 5. Our machine-generated synthetic images achieve a better mask-image alignment than real images in many cases, as shown in Fig. 6. This phenomenon occurs because human annotations tend to be imperfect due to the high difficulty of annotating segmentation masks." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.679, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SegGen" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table", + "bbox": [ + 0.233, + 0.135, + 0.77, + 0.285 + ], + "angle": 0, + "content": "
MethodVenueBackboneCrop SizeIterationsmIoU (s.s.)mIoU (m.s.)
MaskFormer [8]NeurIPS 2021R50512160k44.546.7
Mask DINO [27]CVPR 2023R50512160k48.7-
OneFormer [19]CVPR 2023R50512160k47.3-
Mask2Former [7]CVPR 2022R50512160k47.249.2
—w/ SegGen (ours)-R50512160k49.9 (+2.7)51.4 (+2.2)
MaskFormer [8]NeurIPS 2021Swin-L640160k54.155.6
Mask DINO [27]CVPR 2023Swin-L640160k56.6-
OneFormer [19]CVPR 2023Swin-L640160k57.057.7
Mask2Former [7]CVPR 2022Swin-L640160k56.157.3
—w/ SegGen (ours)-Swin-L640160k57.4 (+1.3)58.7 (+1.4)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.286, + 0.785, + 0.354 + ], + "angle": 0, + "content": "Table 1: Semantic segmentation on ADE20K val (150 categories): Synthetic data is used as a data augmentation. With the same model structures and training recipes, our SegGen boosts the performance of Mask2Former by a large margin and establishes a new SOTA performance without extra real data under both single-scale (s.s.) and multi-scale (m.s.) test settings." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.364, + 0.687, + 0.379 + ], + "angle": 0, + "content": "3.4 Training Segmentation Models with Synthetic Data" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.386, + 0.785, + 0.446 + ], + "angle": 0, + "content": "The final goal of this work is to improve the performance of current segmentation models with the synthetically generated training data. Therefore, we use both synthetic data produced by SegGen and training data from existing datasets in the training process of segmentation models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.447, + 0.787, + 0.583 + ], + "angle": 0, + "content": "We empirically investigate two training strategies utilizing the synthetic data: (i) Synthetic data augmentation strategy. The synthetic data is used for random data augmentation. In every iteration in the training process, each real training sample is replaced by the synthetic training sample with a probability \\( p_{\\mathrm{aug}} \\). (ii) Synthetic data pre-training strategy. It comprises two training stages: pretraining and fine-tuning. In the pre-training stage, we pre-train the segmentation models on synthetic data, so that they learn good weights that are transferable and favorable for fine-tuning. In the subsequent fine-tuning stage, the segmentation models are trained with solely human-annotated data." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.604, + 0.376, + 0.621 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.632, + 0.788, + 0.768 + ], + "angle": 0, + "content": "To accurately evaluate the effectiveness of our data generation method in improving segmentation performance, we adopt mainstream segmentation models and commonly used evaluation benchmarks for several typical segmentation tasks. The experiments are conducted mostly under the fully-supervised learning setting, meaning all human-annotated training samples from the evaluated datasets are used alongside our synthetic data. To guarantee an unbiased assessment of the impact of our synthetic data, we keep the architectures of the segmentation models and training protocols consistent with their respective original implementations throughout our studies." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.788, + 0.457, + 0.803 + ], + "angle": 0, + "content": "4.1 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Segmentation Datasets and Evaluation We conduct experiments on three image segmentation benchmarks following the main experimental settings of" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.392, + 0.129 + ], + "angle": 0, + "content": "Hanrong Ye et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.268 + ], + "angle": 0, + "content": "Mask2Former [7]: ADE20K semantic segmentation [60], COCO panoptic segmentation, and COCO instance segmentation [33]. Our evaluation uses all 150 classes for ADE20K and 133 classes for COCO. We use all the images from the training splits in the training of segmentation models. For semantic segmentation, we show the mean Intersection-over-Union metric (mIoU). For instance segmentation, the average precision (AP) is used. For panoptic segmentation, we report panoptic quality (PQ), \"thing\" instance segmentation \\(\\mathrm{AP}_{\\mathrm{pan}}^{\\mathrm{Th}}\\), and semantic segmentation \\(\\mathrm{mIoU}_{\\mathrm{pan}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.268, + 0.788, + 0.388 + ], + "angle": 0, + "content": "Segmentation Models We adopt Mask2Former [7], a recently-proposed prevalent transformer model, as the default segmentation model. Two typical backbones, i.e., R50 [16] and Swin-L [34], are studied. We keep the official implementation and training hyper-parameters of the segmentation models unchanged. For more training details please refer to their paper. We also conduct experiments on Mask DINO [27], a detection-aided segmentation model, and HRNet W48 [46], a representative fully-convolutional model. More implementation details are introduced in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.389, + 0.788, + 0.63 + ], + "angle": 0, + "content": "Data Generation The Text2Mask and Mask2Img models are both based on the SDXL-base model [39]. They are trained on the training splits of the target segmentation datasets (i.e. ADE20K or COCO) separately for 30,000 iterations using a learning rate of \\(10^{-5}\\). AdamW optimizer is employed, and the models are trained at a resolution of 768. Pre-trained weights from SDXL-base [39] are utilized. Random flipping data augmentation is used. During sampling, the default EDM sampler [22] is used. The sampling steps are 200 in the Text2Mask model and 40 in the Mask2Img model. During data sampling, for each training sample in the ADE20K semantic segmentation dataset, we produce 10 synthetic mask-image pairs using MaskSyn, resulting in 202,100 training samples. Additionally, we synthesize 50 images based on each human-labeled mask with ImgSyn, leading to a total of 1,010,500 samples. On COCO, we solely use ImgSyn for data synthesis. By generating 10 synthetic images conditioned on each human-labeled panoptic segmentation mask via ImgSyn, our synthetic set amounts to 1,182,870 synthetic samples, which are used in the training of both panoptic and instance segmentation models." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.651, + 0.373, + 0.665 + ], + "angle": 0, + "content": "4.2 Main Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.79, + 0.809 + ], + "angle": 0, + "content": "ADE20K Semantic Segmentation We employ the synthetic data augmentation strategy with \\( p_{\\mathrm{aug}} = 60\\% \\) for Mask2Former, and show the results in Table 1. SegGen significantly boosts the mIoU of Mask2Former R50 by +2.7 for single-scale inference and +2.2 for multi-scale inference, achieving 49.9 and 51.4 correspondingly. The Swin-L variant is also largely improved from 56.1/57.3 (single-scale/multi-scale) to 57.4/58.7 (+1.3/+1.4). Notably, our data synthesis method helps Mask2Former surpass the newer methods such as Mask DINO and One-Former [19], while establishing new SOTA results for R50 and Swin-L settings without using additional human-annotated data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "COCO Panoptic Segmentation On COCO we adopt the synthetic data pretraining strategy to utilize our synthetic data. Specifically, we pre-train the mod-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.679, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SegGen" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.131, + 0.784, + 0.319 + ], + "angle": 0, + "content": "
MethodBackbone QueriesEpochsPQ\\( PQ^{Th} \\)\\( PQ^{St} \\)\\( AP_{pan}^{Th} \\)\\( mIoU_{pan} \\)
DETR [3]R50100500+2543.448.236.331.1-
MaskFormer [8]R5010030046.551.039.833.057.8
Mask2Former [7]R501005051.957.743.041.761.7
Mask DINO [27]R503005053.059.143.943.3-
Mask2Former [7]R5010050+5052.057.943.442.061.0
— w/SegGen (ours)R5010050+5052.7 (+0.7)58.8 (+0.9)43.6 (+0.2)43.1 (+1.1)62.6 (+1.6)
Mask DINO [27]R5030050+5053.459.344.444.260.5
— w/SegGen (ours)R5030050+5054.0 (+0.6)60.2 (+0.9)44.7 (+0.3)45.4 (+1.2)61.5 (+1.0)
MaskFormer [8]Swin-L10030052.758.544.040.164.8
OneFormer [19]Swin-L15010057.964.448.049.067.4
Mask2Former [7]Swin-L20010057.864.248.148.667.4
Mask DINOSwin-L3005058.365.148.050.6-
Mask2Former [7]Swin-L200100+10057.364.346.848.066.2
— w/SegGen (ours)Swin-L200100+10058.0 (+0.7)64.5 (+0.2)48.1 (+1.3)48.8 (+0.8)67.4 (+1.2)
Mask DINO [27]Swin-L30050+5058.665.448.350.467.0
— w/SegGen (ours)Swin-L30050+5059.3 (+0.7)65.9 (+0.5)49.3 (+1.0)51.1 (+0.7)68.1 (+1.1)
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.216, + 0.319, + 0.784, + 0.373 + ], + "angle": 0, + "content": "Table 2: Panoptic segmentation on COCO panoptic val2017 (133 categories): Synthetic data is used for pre-training. Our SegGen significantly improves the performance compared with real data pre-training baselines, and establishes new SOTA performance without extra real data." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.38, + 0.376, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.457, + 0.379, + 0.513 + ], + "angle": 0, + "content": "Fig.7: ADE20K mIoU of Mask2Former R50 with different augmentation probability \\( p_{\\mathrm{aug}} \\)." + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.38, + 0.544, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.383, + 0.457, + 0.547, + 0.512 + ], + "angle": 0, + "content": "Fig.8: ADE20K mIoU of Mask2Former R50 using different numbers of synthetic samples." + }, + { + "type": "table", + "bbox": [ + 0.554, + 0.379, + 0.779, + 0.451 + ], + "angle": 0, + "content": "
MethodmIoU*
Pure Real Data83.3
DiffuMask (ICCV 2023)57.0
MaskSyn+ImgSyn (ours)82.2(+25.2)
MaskSyn (ours)73.2
ImgSyn (ours)80.0
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.553, + 0.452, + 0.782, + 0.507 + ], + "angle": 0, + "content": "Table 4: Comparison with DiffuMask [48] on ADE20K: All models trained purely on synthetic data. mIoU* is defined in [48]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.521, + 0.788, + 0.702 + ], + "angle": 0, + "content": "els purely on our synthetic data and fine-tune the models on real data. The results are demonstrated in Table 2. For a fair comparison, we compare with models pretrained on purely real data with the same training settings. Compared with the real data pre-trained baselines, SegGen consistently brings performance gains on all metrics with both R50 and Swin-L backbones, achieving new state-of-the-art performance without extra real data. Notably, for Mask DINO with Swin-L backbone, PQ is improved by \\(+0.7\\), \"thing\" instance segmentation \\(\\mathrm{AP}_{\\mathrm{pan}}^{\\mathrm{Th}}\\) is increased by \\(+0.7\\), and semantic segmentation \\(\\mathrm{mIoU}_{\\mathrm{pan}}\\) is boosted by \\(+1.1\\). These noteworthy findings demonstrate that training with our synthetic data systematically improves performance across various segmentation tasks. We observe similar significant improvement on Mask2Former, further showing the effectiveness of our synthetic data on different models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.841 + ], + "angle": 0, + "content": "COCO Instance Segmentation The performance on instance segmentation, both before and after incorporating our SegGen synthetic data, is detailed in Table 3. Compared to the baselines pre-trained on real data, our technique is superior across all metrics. Specifically, there is a solid \\(+0.5\\) improvement in AP with the R50 backbone, and a \\(+0.8\\) increase when using the Swin-L backbone. Additional assessment on the LVIS [14]-annotated COCO validation set, whose annotations are more accurate than the original version, provides additional evidence of the effectiveness of SegGen, as it enhances the average precision (AP) by \\(+1.2\\) when used with the larger Swin-L backbone." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.392, + 0.129 + ], + "angle": 0, + "content": "Hanrong Ye et al." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.144, + 0.784, + 0.365 + ], + "angle": 0, + "content": "
MethodBackbone QueriesEpochsAP\\( AP^S \\)\\( AP^M \\)\\( AP^L \\)
Mask R-CNN [15]R50anchors40042.523.845.060.0
HTC [4]R50anchors3639.722.642.250.6
QueryInst [11]R503003640.623.442.552.8
MaskFormer [8]R5010030034.016.437.854.2
Mask2Former [7]R501005043.723.447.264.8
Mask2Former [7]R501005047.126.655.472.7
Mask2Former [7]R5010050+5044.123.447.766.1
— w/ SegGen (ours)R5010050+5044.6 (+0.5)24.0 (+0.6)48.2 (+0.5)66.2 (+0.1)
Mask2Former [7]R5010050+5047.927.856.373.5
— w/ SegGen (ours)R5010050+5048.4 (+0.5)28.3 (+0.6)57.4 (+1.1)74.1 (+0.6)
Swin-HTC++ [34]Swin-Lanchors7249.531.052.467.2
QueryInst [11]Swin-L3005048.930.852.668.3
Oneformer [19]Swin-L15010048.9---
Mask2Former [7]Swin-L20010050.129.953.972.1
Mask2Former [7]Swin-L20010054.335.763.379.1
Mask2Former [7]Swin-L200100+10049.529.253.870.5
— w/ SegGen (ours)Swin-L200100+10050.3 (+0.8)31.2 (+2.0)54.3 (+0.5)72.2 (+1.7)
Mask2Former [7]Swin-L200100+10053.635.262.677.8
— w/ SegGen (ours)Swin-L200100+10054.8 (+1.2)35.9 (+0.8)64.2 (+1.6)79.2 (+1.5)
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.214, + 0.365, + 0.788, + 0.435 + ], + "angle": 0, + "content": "Table 3: Instance segmentation on COCO val2017 (80 categories): Synthetic data is used for pre-training. The gray rows correspond to the evaluations conducted on LVIS [14] which offers instance masks of higher quality. Our SegGen significantly improves the performance of instance segmentation for various sizes of objects compared with pre-training on real data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.442, + 0.788, + 0.563 + ], + "angle": 0, + "content": "Comparison with Segmentation Data Generation Method In Table 4, we compare various versions of our SegGen with DiffuMask [48], which is a recently-proposed (ICCV2023) segmentation data generation approach. All methods employ the Mask2Former R50 model and are purely trained on synthetic samples. We adhere to the evaluation setting designed by DiffuMask where the IoU for three common classes are examined. Our method markedly surpasses DiffuMask by \\(+25.2\\) mIoU and demonstrates performance comparable to training purely with real data, highlighting the unprecedented quality of our synthetic data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.565, + 0.788, + 0.686 + ], + "angle": 0, + "content": "Generalization Ability on Unseen Domains We visualize the segmentation results of Mask2Former R50 on images from PASCAL val dataset [10] in Fig. 1 and Fig. 9, comparing models trained both with and without our synthetic data. The models are trained on COCO or ADE20K. When trained using our extensively varied synthetic data, the segmentation model demonstrates notably improved performance on the unseen domain. We further study the segmentation results on AI-generated images, which are synthesized using different image generation models, in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.687, + 0.788, + 0.779 + ], + "angle": 0, + "content": "Visual Analysis of Generated Samples by MaskSyn We randomly select some samples generated by MaskSyn in Fig. 4 (more in supplementary materials). The generative models are trained on ADE20K. Given the text prompts, our Text2Mask model can generate diverse segmentation masks, and the Mask2Img model produces realistic synthetic images with a good alignment with the masks and text prompts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Visual Analysis of Generated Samples by ImgSyn We showcase the outputs by ImgSyn in Fig. 5 (more in supplementary materials). The generated images are in good agreement with the text prompts and human-labeled segmentation masks. Furthermore, we compare the mask-image alignment of the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.679, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SegGen" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.787, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.28, + 0.784, + 0.32 + ], + "angle": 0, + "content": "Fig. 9: Generalization Ability on unseen domains: Segmentation outputs on unseen domain (images from PASCAL [10]). The models are trained on ADE20K. Training with our SegGen demonstrates enhanced robustness towards unfamiliar domains." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.323, + 0.78, + 0.414 + ], + "angle": 0, + "content": "
MethodIterations Real SamplesmIoUMethodIterations Real SamplesmIoU
Baseline160K2021047.2Baseline100K2021044.4
- w/ SegGen MaskSyn160K2021048.5- w/ SegGen MaskSyn100K2021044.8
- w/ SegGen ImgSyn160K2021049.3- w/ SegGen ImgSyn100K2021045.4
- w/ SegGen ImgSyn + MaskSyn160K2021049.9- w/ SegGen ImgSyn + MaskSyn100K2021045.6
Baseline160K100023.7Baseline100K100023.6
- w/ SegGen MaskSyn160K100026.5- w/ SegGen MaskSyn100K100026.0
- w/ SegGen ImgSyn160K100024.1- w/ SegGen ImgSyn100K100023.9
- w/ SegGen ImgSyn + MaskSyn160K100027.1- w/ SegGen ImgSyn + MaskSyn100K100026.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.415, + 0.784, + 0.457 + ], + "angle": 0, + "content": "Table 5: Ablation Study of MaskSyn and ImgSyn on ADE20K: Both MaskSyn and ImgSyn notably enhance the performance of Mask2Former and HRNet, while MaskSyn has a more pronounced impact when there are fewer real samples available." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.466, + 0.784, + 0.511 + ], + "angle": 0, + "content": "real training samples and our synthetic training samples in Fig. 6. Our synthetic samples exhibit superior alignment quality in many cases, compared to human annotations which are often imperfect on the boundaries." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.538, + 0.441, + 0.554 + ], + "angle": 0, + "content": "4.3 Quantitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.788, + 0.794 + ], + "angle": 0, + "content": "Effectiveness of MaskSyn and ImgSyn We evaluate the data generated by MaskSyn and ImgSyn separately and jointly, by examining their impacts on the performance of segmentation models, as shown in Table 5. The experiments are conducted on ADE20K with Mask2Former R50 and HRNet W48. We find that the training samples generated by both MaskSyn and ImgSyn bring significant performance gains compared with the baseline. When combining these two types of synthetic data together, the segmentation model can achieve the best performance. We delve deeper into a scenario where only 1,000 real training samples are available in the training of both generative models and segmentation models. It is observed that MaskSyn substantially enhances the segmentation results in situations with limited real data. This could be attributed to the ability of MaskSyn to amplify data diversity by creating entirely new segmentation masks and images. Moreover, the significant performance improvements achieved by our SegGen, trained on merely 1,000 real samples, demonstrate the data efficiency and robustness of our proposal." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.794, + 0.842 + ], + "angle": 0, + "content": "Influence of Synthetic Data Augmentation Probability \\( p_{\\mathrm{aug}} \\) on ADE20K We investigate the effect of \\( p_{\\mathrm{aug}} \\) in Fig. 7. When \\( p_{\\mathrm{aug}} = 0 \\), it implies training using only real data, whereas \\( p_{\\mathrm{aug}} = 100\\% \\) means training entirely with synthetic" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.392, + 0.129 + ], + "angle": 0, + "content": "Hanrong Ye et al." + }, + { + "type": "table", + "bbox": [ + 0.23, + 0.145, + 0.486, + 0.185 + ], + "angle": 0, + "content": "
MethodIterationsmIoU
Syn. Pre-Train160K+160K47.2
Syn. Data Aug160K49.9
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.217, + 0.186, + 0.497, + 0.218 + ], + "angle": 0, + "content": "(a) ADE20K Semantic Segmentation: Using synthetic data as random data augmentation obtains the best performance on ADE20K." + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.145, + 0.772, + 0.184 + ], + "angle": 0, + "content": "
MethodEpochsPQAP
Syn. Pre-Training50+5052.744.6
Syn. Data Aug5051.343.2
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.5, + 0.186, + 0.782, + 0.227 + ], + "angle": 0, + "content": "(b) COCO Panoptic Segmentation and Instance Segmentation: Using synthetic data for pre-training achieves better performance on COCO dataset." + }, + { + "type": "table_caption", + "bbox": [ + 0.221, + 0.229, + 0.779, + 0.243 + ], + "angle": 0, + "content": "Table 6: Different training strategies with synthetic data using Mask2Former R50." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.248, + 0.785, + 0.277 + ], + "angle": 0, + "content": "data. The best results are achieved at \\( p_{\\mathrm{aug}} = 60\\% \\). It is noteworthy that training exclusively on our synthetic data results in an impressive \\( 44.8~\\mathrm{mIoU} \\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.278, + 0.787, + 0.399 + ], + "angle": 0, + "content": "Influence of Synthetic Data Size To evaluate the impact of synthetic data volume, we use different numbers of synthetic training samples produced by ImgSyn in training and report the results in Fig. 8. \\( p_{\\mathrm{aug}} \\) is fixed at \\( 60\\% \\). We notice a substantial improvement in performance when increasing the synthetic data quantity from 20K to 200K, underscoring the importance of collecting a significantly larger synthetic dataset to avoid overfitting. The performance appears to plateau after 600K. More studies on the scale of synthetic data are presented in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.399, + 0.789, + 0.673 + ], + "angle": 0, + "content": "Influence of Training Strategies We conduct an ablation study using different training strategies with synthetic data on ADE20K (Table 6a) and COCO (Table 6b). We find that using the synthetic data augmentation strategy achieves better performance on ADE20K semantic segmentation, While on COCO panoptic and instance segmentation, the synthetic pre-training strategy works better. We believe that due to the limited size of the ADE20K training set, which contains merely \\(\\sim 20\\mathrm{k}\\) images, segmentation models easily overfit the scarcely-available training images and segmentation layouts. Utilizing synthetic data (including synthetic masks and images) for data augmentation can help mitigate such an overfitting problem. On the other hand, for COCO, given its larger training set size (around 100K images), overfitting is less of a concern. Hence, it is sufficient to provide the COCO models with good initial weights pre-trained on our synthetic data. Meanwhile, previous work [14] finds that the annotation accuracy of COCO is significantly more biased than ADE20K, a finding that aligns with our visualization results depicted in Fig. 6. A domain gap exists between the COCO data and our high-quality synthetic data. Hence, employing our method for pre-training and incorporating real data for fine-tuning on COCO is a more favorable approach." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.69, + 0.36, + 0.707 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We present SegGen, a highly-effective data synthesis method for image segmentation. SegGen introduces a revolutionary segmentation data generation framework, and proposes two data generation approaches, namely MaskSyn and ImgSyn, to synthesize segmentation masks and images, with the help of the designed text-to-mask and mask-to-image generation models. Our method not only strongly enhances the model performance on semantic, panoptic, and instance segmentation benchmarks, but also substantially improves segmentation generalization ability in unseen data distribution." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.679, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SegGen" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.395, + 0.163 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.788, + 0.24 + ], + "angle": 0, + "content": "This research is supported in part by the Early Career Scheme of the Research Grants Council (RGC) of the Hong Kong SAR under grant No. 26202321, SAIL Research Project, HKUST-Zeekr Collaborative Research Fund, HKUST-WeBank Joint Lab Project, and Tencent Rhino-Bird Focused Research Program." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.263, + 0.324, + 0.279 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.296, + 0.785, + 0.323 + ], + "angle": 0, + "content": "1. Baranchuk, D., Rubachev, I., Voynov, A., Khrulkov, V., Babenko, A.: Label-efficient semantic segmentation with diffusion models. In: ICLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.325, + 0.785, + 0.351 + ], + "angle": 0, + "content": "2. Brock, A., Donahue, J., Simonyan, K.: Large scale gan training for high fidelity natural image synthesis. In: ICLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.353, + 0.785, + 0.379 + ], + "angle": 0, + "content": "3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.381, + 0.785, + 0.42 + ], + "angle": 0, + "content": "4. Chen, K., Pang, J., Wang, J., Xiong, Y., Li, X., Sun, S., Feng, W., Liu, Z., Shi, J., Ouyang, W., et al.: Hybrid task cascade for instance segmentation. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.422, + 0.785, + 0.462 + ], + "angle": 0, + "content": "5. Chen, L.C., Papandreou, G., Kokkinos, I., Murphy, K., Yuille, A.L.: Semantic image segmentation with deep convolutional nets and fully connected crfs. In: ICLR (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.464, + 0.785, + 0.504 + ], + "angle": 0, + "content": "6. Chen, T., Li, L., Saxena, S., Hinton, G., Fleet, D.J.: A generalist framework for panoptic segmentation of images and videos. arXiv preprint arXiv:2210.06366 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.506, + 0.785, + 0.533 + ], + "angle": 0, + "content": "7. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.535, + 0.785, + 0.56 + ], + "angle": 0, + "content": "8. Cheng, B., Schwing, A.G., Kirillov, A.: Per-pixel classification is not all you need for semantic segmentation. In: NeurIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.562, + 0.785, + 0.588 + ], + "angle": 0, + "content": "9. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. In: NeruIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.59, + 0.785, + 0.616 + ], + "angle": 0, + "content": "0. Everingham, M., Eslami, S.A., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The PASCAL visual object classes challenge: A retrospective. IJCV (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.618, + 0.785, + 0.644 + ], + "angle": 0, + "content": "1. Fang, Y., Yang, S., Wang, X., Li, Y., Fang, C., Shan, Y., Feng, B., Liu, W.: Instances as queries. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.646, + 0.785, + 0.672 + ], + "angle": 0, + "content": "2. Forever, A.: Kandinsky-2. GitHub repository (2023), https://github.com/ai-forever/Kandinsky-2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.674, + 0.785, + 0.7 + ], + "angle": 0, + "content": "3. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. CACM (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.702, + 0.785, + 0.728 + ], + "angle": 0, + "content": "4. Gupta, A., Dollar, P., Girshick, R.: LVIS: A dataset for large vocabulary instance segmentation. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.729, + 0.759, + 0.743 + ], + "angle": 0, + "content": "5. He, K., Gkioxari, G., Dolkar, P., Girshick, R.: Mask R-CNN. In: ICCV (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.745, + 0.785, + 0.77 + ], + "angle": 0, + "content": "6. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.785, + 0.798 + ], + "angle": 0, + "content": "7. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeruIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "8. Huang, L., Chen, D., Liu, Y., Yujun, S., Zhao, D., Jingren, Z.: Composer: Creative and controllable image synthesis with composable conditions. arXiv preprint arxiv:2302.09778 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.296, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.392, + 0.129 + ], + "angle": 0, + "content": "Hanrong Ye et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "19. Jain, J., Li, J., Chiu, M.T., Hassani, A., Orlov, N., Shi, H.: Oneformer: One transformer to rule universal image segmentation. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.203 + ], + "angle": 0, + "content": "20. Ji, Y., Chen, Z., Xie, E., Hong, L., Liu, X., Liu, Z., Lu, T., Li, Z., Luo, P.: Ddp: Diffusion model for dense visual prediction. arXiv preprint arXiv:2303.17559 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.785, + 0.231 + ], + "angle": 0, + "content": "21. Kang, M., Zhu, J.Y., Zhang, R., Park, J., Shechtman, E., Paris, S., Park, T.: Scaling up gans for text-to-image synthesis. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.259 + ], + "angle": 0, + "content": "22. Karras, T., Aittala, M., Aila, T., Laine, S.: Elucidating the design space of diffusion-based generative models. In: NeruIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.785, + 0.285 + ], + "angle": 0, + "content": "23. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.314 + ], + "angle": 0, + "content": "24. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.342 + ], + "angle": 0, + "content": "25. Kirillov, A., He, K., Girshick, R., Rother, C., Dlár, P.: Panoptic segmentation. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.785, + 0.37 + ], + "angle": 0, + "content": "26. Li, D., Ling, H., Kim, S.W., Kreis, K., Fidler, S., Torralba, A.: Bigdatasetgan: Synthesizing imagenet with pixel-wise annotations. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.41 + ], + "angle": 0, + "content": "27. Li, F., Zhang, H., Xu, H., Liu, S., Zhang, L., Ni, L.M., Shum, H.Y.: Mask dino: Towards a unified transformer-based framework for object detection and segmentation. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.438 + ], + "angle": 0, + "content": "28. Li, J., Li, D., Savarese, S., Hoi, S.: BLIP-2: bootstrapping language-image pretraining with frozen image encoders and large language models. In: ICML (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.44, + 0.785, + 0.466 + ], + "angle": 0, + "content": "29. Li, X., Zhong, Z., Wu, J., Yang, Y., Lin, Z., Liu, H.: Expectation-maximization attention networks for semantic segmentation. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.508 + ], + "angle": 0, + "content": "30. Li, X., Ding, H., Zhang, W., Yuan, H., Pang, J., Cheng, G., Chen, K., Liu, Z., Loy, C.C.: Transformer-based visual segmentation: A survey. arXiv preprint arXiv:2304.09854 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.785, + 0.536 + ], + "angle": 0, + "content": "31. Li, Y., Liu, H., Wu, Q., Mu, F., Yang, J., Gao, J., Li, C., Lee, Y.J.: Gligen: Open-set grounded text-to-image generation. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.537, + 0.785, + 0.563 + ], + "angle": 0, + "content": "32. Li, Z., Zhou, Q., Zhang, X., Zhang, Y., Wang, Y., Xie, W.: Open-vocabulary object segmentation with diffusion models. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.785, + 0.591 + ], + "angle": 0, + "content": "33. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólár, P., Zitnick, C.L.: Microsoft COCO: Common objects in context. In: ECCV (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.785, + 0.632 + ], + "angle": 0, + "content": "34. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. arXiv:2103.14030 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.66 + ], + "angle": 0, + "content": "35. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: CVPR (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.785, + 0.702 + ], + "angle": 0, + "content": "36. Mou, C., Wang, X., Xie, L., Wu, Y., Zhang, J., Qi, Z., Shan, Y., Qie, X.: T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.703, + 0.785, + 0.743 + ], + "angle": 0, + "content": "37. Nguyen, Q.N.T.V.A.T.K.: Dataset diffusion: Diffusion-based synthetic dataset generation for pixel-level semantic segmentation. arXiv preprint arXiv:2309.14303 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.785, + 0.771 + ], + "angle": 0, + "content": "38. PNVR, K., Singh, B., Ghosh, P., Siddiquie, B., Jacobs, D.: Ld-znet: A latent diffusion approach for text-based image segmentation. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.812 + ], + "angle": 0, + "content": "39. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "40. Qi, L., Kuen, J., Wang, Y., Gu, J., Zhao, H., Torr, P., Lin, Z., Jia, J.: Open world entity segmentation. TPAMI (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.679, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SegGen" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.175 + ], + "angle": 0, + "content": "41. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.217 + ], + "angle": 0, + "content": "42. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. In: NeruIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.217, + 0.788, + 0.243 + ], + "angle": 0, + "content": "43. Sauer, A., Karras, T., Laine, S., Geiger, A., Aila, T.: StyleGAN-T: Unlocking the power of GANs for fast large-scale text-to-image synthesis. In: ICML (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.243, + 0.788, + 0.271 + ], + "angle": 0, + "content": "44. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: ICML (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.271, + 0.788, + 0.311 + ], + "angle": 0, + "content": "45. Wang, H., Cao, J., Anwer, R.M., Xie, J., Khan, F.S., Pang, Y.: Dformer: Diffusion-guided transformer for universal image segmentation. arXiv preprint arXiv:2306.03437 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.311, + 0.788, + 0.353 + ], + "angle": 0, + "content": "46. Wang, J., Sun, K., Cheng, T., Jiang, B., Deng, C., Zhao, Y., Liu, D., Mu, Y., Tan, M., Wang, X., Liu, W., Xiao, B.: Deep high-resolution representation learning for visual recognition. TPAMI (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.353, + 0.788, + 0.393 + ], + "angle": 0, + "content": "47. Wu, W., Zhao, Y., Chen, H., Gu, Y., Zhao, R., He, Y., Zhou, H., Shou, M.Z., Shen, C.: Datasetsd: Synthesizing data with perception annotations using diffusion models. arXiv preprint arXiv:2308.06160 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.393, + 0.788, + 0.435 + ], + "angle": 0, + "content": "48. Wu, W., Zhao, Y., Shou, M.Z., Zhou, H., Shen, C.: Diffumask: Synthesizing images with pixel-level annotations for semantic segmentation using diffusion models. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.435, + 0.788, + 0.475 + ], + "angle": 0, + "content": "49. Xie, J., Li, W., Li, X., Liu, Z., Ong, Y.S., Loy, C.C.: Mosaic fusion: Diffusion models as data augmenters for large vocabulary instance segmentation. arXiv preprint arXiv:2309.13042 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.476, + 0.788, + 0.503 + ], + "angle": 0, + "content": "50. Xu, J., Liu, S., Vahdat, A., Byeon, W., Wang, X., De Mello, S.: Open-vocabulary panoptic segmentation with text-to-image diffusion models. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.503, + 0.788, + 0.53 + ], + "angle": 0, + "content": "51. Yang, L., Xu, X., Kang, B., Shi, Y., Zhao, H.: Freemask: Synthetic images with dense annotations make stronger segmentation models. In: NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.53, + 0.788, + 0.557 + ], + "angle": 0, + "content": "52. Ye, H., Xu, D.: Taskprompter: Spatial-channel multi-task prompting for dense scene understanding. In: ICLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.557, + 0.788, + 0.584 + ], + "angle": 0, + "content": "53. Zeng, Y., Lin, Z., Zhang, J., Liu, Q., Collomosse, J., Kuen, J., Patel, V.M.: Scenecomposer: Any-level semantic image synthesis. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.584, + 0.788, + 0.612 + ], + "angle": 0, + "content": "54. Zhang, L., Anyi, R., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.612, + 0.788, + 0.639 + ], + "angle": 0, + "content": "55. Zhang, W., Pang, J., Chen, K., Loy, C.C.: K-net: Towards unified image segmentation. In: NeurIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.639, + 0.788, + 0.68 + ], + "angle": 0, + "content": "56. Zhang, Y., Ling, H., Gao, J., Yin, K., Lafleche, J.F., Barriuso, A., Torralba, A., Fidler, S.: Datasetgan: Efficient labeled data factory with minimal human effort. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.68, + 0.788, + 0.707 + ], + "angle": 0, + "content": "57. Zhao, H., Shi, J., Qi, X., Wang, X., Jia, J.: Pyramid scene parsing network. In: CVPR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.707, + 0.788, + 0.734 + ], + "angle": 0, + "content": "58. Zhao, S., Chen, D., Chen, Y.C., Bao, J., Hao, S., Yuan, L., Wong, K.Y.K.: Uncontrolnet: All-in-one control to text-to-image diffusion models. NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.762 + ], + "angle": 0, + "content": "59. Zhao, W., Rao, Y., Liu, Z., Liu, B., Zhou, J., Lu, J.: Unleashing text-to-image diffusion models for visual perception. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.762, + 0.788, + 0.802 + ], + "angle": 0, + "content": "60. Zhou, B., Zhao, H., Puig, X., Fidler, S., Barriuso, A., Torralba, A.: Scene parsing challenge 2016. http://sceneParsing.csail.mit.edu/index_challenge.htm (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.802, + 0.788, + 0.83 + ], + "angle": 0, + "content": "61. Zhou, B., Zhao, H., Puig, X., Fidler, S., Barriuso, A., Torralba, A.: Scene parsing through ADE20K dataset. In: CVPR (2017)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.83 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/SegGen_ Supercharging Segmentation Models with Text2Mask and Mask2Img Synthesis/acac7544-25ea-4a13-aba6-d24e96031cd3_origin.pdf b/2024/SegGen_ Supercharging Segmentation Models with Text2Mask and Mask2Img Synthesis/acac7544-25ea-4a13-aba6-d24e96031cd3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5852ae5ed10b41c157371b031bbd939c0c5511e3 --- /dev/null +++ b/2024/SegGen_ Supercharging Segmentation Models with Text2Mask and Mask2Img Synthesis/acac7544-25ea-4a13-aba6-d24e96031cd3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:08c801592befa2f228bdabccd89e8086382e5ea8f335d956f96eaaaf7ce69281 +size 7044383 diff --git a/2024/SegGen_ Supercharging Segmentation Models with Text2Mask and Mask2Img Synthesis/full.md b/2024/SegGen_ Supercharging Segmentation Models with Text2Mask and Mask2Img Synthesis/full.md new file mode 100644 index 0000000000000000000000000000000000000000..485e55858fc188aff26a641e8fb419122e8cdbd0 --- /dev/null +++ b/2024/SegGen_ Supercharging Segmentation Models with Text2Mask and Mask2Img Synthesis/full.md @@ -0,0 +1,255 @@ +# SegGen: Supercharging Segmentation Models with Text2Mask and Mask2Img Synthesis + +Hanrong Ye $^{1}$ , Jason Kuen $^{2}$ , Qing Liu $^{2}$ , Zhe Lin $^{2}$ , Brian Price $^{2}$ , Dan Xu $^{1}$ + +1 CSE, HKUST +2 Adobe Research + +Abstract. We present SegGen, a new data generation approach that pushes the performance boundaries of state-of-the-art image segmentation models. One major bottleneck of previous data synthesis methods for segmentation is the design of "segmentation labeler module", which is used to synthesize segmentation masks for images [48]. The segmentation labeler modules, which are segmentation models by themselves, bound the performance of downstream segmentation models trained on the synthetic masks. These methods encounter a "chicken or egg dilemma" and thus fail to outperform existing segmentation models. To address this issue, we propose a novel method that reverses the traditional data generation process: we first (i) generate highly diverse segmentation masks that match real-world distribution from text prompts, and then (ii) synthesize realistic images conditioned on the segmentation masks. In this way, we avoid the need for any segmentation labeler module. SegGen integrates two data generation strategies, namely MaskSyn and ImgSyn, to largely improve data diversity in synthetic masks and images. Notably, the high quality of our synthetic data enables our method to outperform the previous data synthesis method [48] by $+25.2$ mIoU on ADE20K when trained with pure synthetic data. On the highly competitive ADE20K and COCO benchmarks, our data generation method markedly improves the performance of state-of-the-art segmentation models in semantic segmentation, panoptic segmentation, and instance segmentation. Moreover, experiments show that training with our synthetic data makes the segmentation models more robust towards unseen data domains, including real-world and AI-generated images. + +Keywords: Image Segmentation $\cdot$ Image Generation $\cdot$ Deep Learning + +# 1 Introduction + +Image segmentation explores the identification of objects in visual inputs at the pixel level. Based on the different emphases on category and instance membership information, researchers have divided image segmentation into several tasks [5,25,35,40]. For example, semantic segmentation studies pixel-level understanding of object categories, instance segmentation focuses on instance grouping + +![](images/2abb841d4bbb9857413b2985a699cd0e8646bc00e92b8d8c9523bda968ee9011.jpg) +Fig. 1: Effectiveness of SegGen on various data domains: Through training with synthetic data generated by the proposed SegGen, we significantly boost the performance of state-of-the-art segmentation model Mask2Former [7] on evaluation benchmarks including ADE20K [60] and COCO [33], whilst making it more robust towards challenging images from other domains (the three columns on the left are from PASCAL [10]; the three on the right are synthesized by image generation model Kandinsky 2 [12]). SegGen outperforms the previous best data generation method (DiffuMask [48]) by a huge margin when models are trained on pure synthetic data. "mIoU*" is the average IoU metric defined by [48] which focuses on three common classes. + +of pixels, while panoptic segmentation considers both. For all these segmentation tasks, obtaining high-quality annotation is challenging as every individual pixel requires human labeling, and a single image can contain millions of pixels. Therefore, compared to other public datasets like ImageNet-21K (with around 14M images), the prevailing human-annotated segmentation datasets are notably smaller. For example, ADE20K dataset [61] contains about 20K images in its training split, while COCO [33] has around 118K training images. Although there has been significant development in the structure of segmentation models [8,29,52,55,57], the limited size of training data hinders further performance enhancements and results in inadequate generalization ability to handle images from unfamiliar domains, such as those from other scenes or synthesized by generative models as shown in the second row of Fig. 1. + +Inspired by the recent success of image generation [9, 23], researchers start to explore using generative models to enhance image segmentation [1]. A representative direction of this research focuses on synthesizing segmentation training data in a cost-effective manner [26]. In related methods, the synthetic segmentation masks are obtained from some manually designed "segmentation labeler modules". The segmentation labeler modules are essentially segmentation models by themselves. They are either small-scale segmentation networks as in DatasetGAN [56] and Grounded Diffusion [32], or post-processing methods on the image features as in DiffuMask [48]. The performance of the downstream segmentation models is constrained by the quality of the synthetic masks they are trained on, which in turn relies on the capabilities of those segmentation labeler modules. This is a "chicken or egg dilemma" and the performance bottleneck is the segmentation labeler module. Therefore, while they achieve encouraging success in settings with limited training data, their methods fail to deliver notable en + +![](images/53e2a0790b39b289b0717afba00ada1f43d88652c137dca938dd01b9f8c52f99.jpg) +Fig. 2: Comparison with previous data generation methods for segmentation: (a) Earlier methods [48, 56] rely on segmentation labeler modules to produce segmentation masks for synthetic images. However, the performance of these downstream segmentation models, trained on synthetic data, is bounded by the capacity of the segmentation labeler modules. The segmentation labeler is a major bottleneck for the quality of the generated data. (b) We design a reverse pipeline: we first create diverse new masks from text prompts via a proposed Text2Mask generation model and then synthesize images conditioned on the segmentation masks. This methodology avoids any usage of segmentation labeler networks, resulting in significantly improved quality of our synthesis data. + +hancements for the state-of-the-art models on the most popular segmentation benchmarks in settings using the complete training sets. + +On the other hand, a handful of works [54, 58] have demonstrated that high-quality synthetic images can be generated by conditioning text-to-image models on dense input maps (e.g., segmentation masks or canny edge). Moreover, these dense-input conditional models provide a strong alignment between the dense input label maps and generated images. These two qualities motivate us to leverage the powerful capabilities of such models for effective segmentation data generation. To this end, we propose a novel segmentation data generation method, coined as SegGen, for generating high-quality segmentation training data. We reverse the traditional data generation pipeline for segmentation: we first synthesize high-quality segmentation masks from a proposed text-to-mask generation model, and then generate images based on the segmentation masks using a mask-to-image generation model. In this new framework, we avoid using any segmentation labeler modules, which results in significantly improved data quality. Fig. 2 shows the difference between our method and previous data generation methods for segmentation [48, 56]. + +As each training sample of image segmentation consists of two components: segmentation masks and the corresponding image, we develop two novel data generation approaches, emphasizing improvements in two distinct aspects of data diversity: (i) segmentation masks and (ii) images. The first data generation approach, named MaskSyn, centers around the generation of new segmentation masks. It learns a text-to-mask (Text2Mask) generation model to produce completely new segmentation masks given text prompts. Then, it learns a mask-to-image (Mask2Img) generation model to synthesize images that align with the synthetic segmentation masks. The second data generation approach, named ImgSyn, utilizes the above-mentioned Mask2Img model to synthesize new images given human-annotated segmentation masks. With MaskSyn and ImgSyn, we can readily generate a vast array of diverse and high-quality synthetic training data. The combined synthetic data is used to train segmenta + +tion models in conjunction with real training samples from human-annotated datasets. Experiments show that our synthetic data can significantly boost the performance of the image segmentation models on challenging benchmarks including ADE20K semantic segmentation, COCO panoptic segmentation, and COCO instance segmentation, achieving new state-of-the-art performances without using extra human-annotated data. Notably, SegGen remarkably outperforms the previous segmentation data generation method [48] by $+25.2$ mIoU when trained with pure synthetic data on ADE20K. Our method boosts the mIoU of Mask2Former by $+2.7$ (R50) and $+1.3$ (Swin-L) on the ADE20K semantic segmentation benchmark. Furthermore, segmentation models trained on our synthetic data exhibit a remarkably stronger ability to generalize across unfamiliar image domains, including real images from other distributions and machine-generated images. + +We summarize the contribution of this work in three points: (i) We propose a revolutionary generation framework that reverses the traditional segmentation data generation pipeline and solves the "chicken or egg dilemma". The new framework enables us to produce high-quality segmentation training data at scale, thus enabling the training of more powerful image segmentation models. (ii) We introduce two effective generative models, one for text-to-mask generation and the other for mask-to-image generation. Based on these models, we propose two novel segmentation training data generation approaches, namely MaskSyn and ImgSyn. They significantly improve the data diversity, with MaskSyn focusing on new segmentation masks and ImgSyn on new images. (iii) SegGen successfully improves the performance of the leading-edge segmentation models across the highly competitive benchmarks on ADE20K and COCO. Moreover, SegGen enhances the generalization ability of segmentation models towards unseen image domains. Rigorous experiments, including ablation study and peer comparison, strongly suggest the effectiveness of the proposed method. + +# 2 Related Work + +Generation for Segmentation Image segmentation is one of the most studied visual perception problems [30]. In recent years, there has been a surge in efforts to harness the capabilities of generative models for segmentation tasks. These efforts can be broadly classified into three categories based on their methodologies: (i) Extracting visual features from generative models for segmentation [1, 38, 50, 59]. These methods harvest the strong representation ability of diffusion models trained on large-scale datasets but are limited in the precision of predicted masks. (ii) Formulating segmentation tasks directly as generative models [6, 20, 45]. Although these methods propose exciting new model architectures, they usually consume higher computational costs while showing unimproved performance compared with conventional segmentation models. (iii) Synthesizing segmentation training data using generative models [26, 32, 37, 47-49, 56]. While these methods have showcased commendable results against their respective baselines especially when the training data is highly limited, they have + +yet to exhibit notably superior performance on the most rigorous benchmarks including ADE20K (150 categories) [61] and COCO (133 categories) [33] under fully-supervised setting. A primary concern with these techniques is the subpar quality of the generated segmentation masks. This stems from their dependence on the bottleneck segmentation labeler modules during the synthetic mask generation process. Therefore, a concurrent work [51] abandons the generation of segmentation masks and directly generates images based on existing masks, which is highly limited in mask diversity. In contrast, our data generation workflow reverses the traditional data generation pipeline: we first generate segmentation masks from text, and then synthesize images based on segmentation masks. Our method avoids the need for any segmentation labeler which results in largely improved data quality. Our SegGen can strongly enhance the existing state-of-the-art segmentation models on different challenging benchmarks. + +Conditional Image Synthesis Within the realm of conditional image synthesis, Generative Adversarial Networks [2, 13, 21, 23, 43], Variational Autoencoders [24], and Diffusion Models [9, 17, 42, 44] have been at the forefront. Recently, the open-source research community has shown a burgeoning interest in the latent diffusion-based Stable Diffusion (SD) series for text-to-image synthesis [41]. The most recent iteration, SDXL, introduced by [39], expands the model capacity, yielding significantly enhanced results. Therefore, we build SegGen upon SDXL. Regarding mask-to-image generation models [18, 31, 53], ControlNet [54] and T2I-Adapter [36] suggest freezing the parameters of the SD model and introducing a set of more compact, learnable modules. This approach enables image generation conditioned on the given segmentation masks. We adopt the structure of ControlNet in our mask-to-image generation model. To the best of our knowledge, there have been no prior endeavors on text-to-mask generation. + +# 3 Method + +# 3.1 Overview + +SegGen is designed to synthesize high-quality training samples for improving the performance of segmentation models. The overall workflow is shown on the left of Fig. 3. We first train SegGen with human-annotated training samples from public datasets. After training, we use SegGen to produce new segmentation training samples at scale. The generated training samples are incorporated into the training process of segmentation models to enhance the model performance. + +# 3.2 Models in SegGen + +As shown in Fig. 3, we first utilize a captioner model to extract captions of the real training images as text prompts from the target dataset. The text prompts will condition the data generation process. Then, two conditional generative models are introduced: a text-to-mask (Text2Mask) generation model and a mask-to-image (Mask2Img) generation model. Both generative models are built upon the SDXL model [39] which provides top-notch image generation quality. + +![](images/9a4bc0c5cf53f18d4e5ccdec7be3fb1a83f078c3c5a3da5f075adf4afb65aa36.jpg) +Fig.3: Illustration of the workflow of our proposed SegGen: We introduce two generative models: a text-to-mask (Text2Mask) generation model and a mask-to-image (Mask2Img) generation model, based on which we design two approaches for synthesizing segmentation training samples: MaskSyn and ImgSyn. (a) MaskSyn focuses on generating new segmentation masks. It first extracts the caption of the real image as a text prompt and uses it to generate new masks with the Text2Mask model. Then, the new masks and text prompt are fed into the Mask2Img model to produce the corresponding new images. (b) ImgSyn focuses on the synthesis of new images. It inputs human-labeled masks and text prompts into the Mask2Img model. + +Captioner Model To obtain image captions of existing training samples, we employ BLIP2-FlanT5xxl [28] model, which is a state-of-the-art vision-language model. We feed the text prompt "Question: What are shown in the photo? Answer:" and image as input to the model. The responses serve as text prompts to condition the following generation process. + +Text2Mask Model We design a Text2Mask model in SegGen for generating diverse segmentation masks based on given text prompts. To leverage the generation capacity of text-to-image generation models pre-trained on large-scale datasets, we encode the segmentation masks (the pixel values are category IDs) as three-channel RGB-like color maps, where one color represents a certain category. From our experiments, the color map reconstructed by VAE [41] in SDXL appears almost indistinguishable from the original input as shown in the supplementary materials. Therefore, we can directly fine-tune the text-to-image SDXL-base model with [text, segmentation color map] training pairs, which are from the public image segmentation dataset (e.g. ADE20K). During sampling, our Text2Mask model can generate diverse color maps conditioned on text prompts, which are subsequently converted into segmentation masks. Formally, suppose the input text prompt is $\mathbf{T}$ , the target height and width are $H$ and $W$ , the synthesized color map is $\mathbf{C}_{\mathrm{syn}} \in \mathbb{W}^{H \times W \times 3}$ , and the synthesized segmentation map (with $N$ masks) is $\mathbf{M}_{\mathrm{syn}} \in \mathbb{W}^{H \times W \times N}$ , the generation process is: + +$$ +\begin{array}{l} \mathbf {C} _ {\mathrm {s y n}} = \operatorname {T e x t 2 M a s k} (\mathbf {T}), \\ \mathbf {M} _ {\text {s y n}} = f _ {\text {c o l o r} \rightarrow \text {m a s k}} \left(\mathbf {C} _ {\text {s y n}}\right), \tag {1} \\ \end{array} +$$ + +where $f_{\mathrm{color} \to \mathrm{mask}}: \mathbb{W}^{H \times W \times 3} \to \mathbb{W}^{H \times W \times N}$ is the function that projects the color maps to segmentation masks. Specifically, for each pixel on the color maps, we identify its nearest color (in Euclidean space) in the aforementioned lookup table, and assign the corresponding class to the pixel in the segmentation masks. + +![](images/c9017728c00b05353ee46f2c44a7598961464a543b8f2a9149d487578b592409.jpg) +Fig. 4: Generated samples by MaskSyn on ADE20K: The third row overlays the mask and the image together to demonstrate the alignment between them. The generated segmentation masks and images demonstrate high perceptual quality and excellent alignment (see more samples in supplementary materials). + +Mask2Img Model The goal of the Mask2Img model is to synthesize new images that align well with the given segmentation masks and text prompts. Specifically we adopt the structure of ControlNet [54]: we freeze the pre-trained weights of the SDXL-base model and train an additional side network for mask-conditioned image generation. It simultaneously keeps the generalization ability of the pretrained diffusion model and provides excellent controllable generation ability. The Mask2Img model is trained with the [text, segmentation color map, image] triplets gathered from the training splits of the target datasets. We denote the input segmentation map as $\mathbf{M} \in \mathbb{W}^{H \times W \times N}$ , the color map as $\mathbf{C} \in \mathbb{W}^{H \times W \times 3}$ , the synthetic image as $\mathbf{I}_{\mathrm{syn}} \in \mathbb{W}^{H \times W \times 3}$ , and the generation process is: + +$$ +\begin{array}{l} \mathbf {C} = f _ {\text {m a s k} \rightarrow \text {c o l o r}} (\mathbf {M}), \\ \mathbf {I} _ {\text {s y n}} = \operatorname {M a s k 2 I m g} (\mathbf {T}, \mathbf {C}), \tag {2} \\ \end{array} +$$ + +where $f_{\mathrm{mask}\rightarrow \mathrm{color}}:\mathbb{W}^{H\times W\times N}\to \mathbb{W}^{H\times W\times 3}$ is the function to convert the segmentation masks into a color map. For semantic segmentation, the value of each pixel on the segmentation mask corresponds to a category ID, allowing us to convert the masks directly into an RGB color map using a pre-defined lookup table. For panoptic and instance segmentation, after mapping the category IDs to color maps, it is essential to outline each segment with a special edge color on the color map. This ensures the model recognizes the specific instance it belongs to. The segmentation map $\mathbf{M}$ can be human-annotated or synthetic (i.e., $\mathbf{M}_{\mathrm{syn}}$ from Text2Mask Model as shown in Eq. 1). + +# 3.3 Data Generation + +With the aforementioned generative models, SegGen proposes two approaches for synthesizing new segmentation training samples: MaskSyn and ImgSyn. MaskSyn focuses on enhancing the diversity of synthetic segmentation masks, whereas ImgSyn concentrates on diversifying synthetic images. These generation approaches are illustrated on the right of Fig. 3. + +![](images/3c600986a67bc54a0876c08e61aed94f78788a7b34d9b9135ab8376737e72cc6.jpg) +Fig. 5: Generated samples by ImgSyn on ADE20K: The generated images exhibit remarkable realism and align well with the human-labeled masks and text prompts (see more samples in supplementary materials). + +![](images/200a30a9c1257688104fa7d6aeaad1e3dba5099ac0f91d4f95da3e7689242c02.jpg) +Fig. 6: Zoom-in comparison of real images and our ImgSyn images: As highlighted in the circles, our synthetic images align better with the human-labeled masks in many cases because of the inaccuracies in annotations. The left 4 columns are from ADE20K and the right 4 from COCO. + +MaskSyn MaskSyn starts with a real training sample pair [image, segmentation masks] from a human-annotated segmentation dataset. It first extracts the caption of the image with the image captioner model. The caption serves as a text prompt and is used to generate a set of diverse new segmentation masks with the Text2Mask model following Eq. 1. Subsequently, the new segmentation masks and the corresponding text prompt are fed into the Mask2Img model to generate a new image that aligns well with its mask. As such, each training sample crafted by MaskSyn includes both a novel segmentation mask and a new image. MaskSyn effectively increases the data diversity in segmentation masks to a great extent. Some generated samples are shown in Fig. 4. + +ImgSyn Different from MaskSyn, ImgSyn focuses on increasing the data diversity of images based on human-annotated segmentation masks. For each real training sample pair [image, segmentation mask], it takes the human-annotated segmentation mask and the text prompt extracted from the image as input, and generates a set of varied images that align well with the human-annotated mask. In this way, new training samples consisting of human-labeled masks and new synthetic images are generated. ImgSyn can also be viewed as a kind of data augmentation method that enhances the data diversity on the image side. Our experiments reveal a remarkably high alignment between the synthetic images and their respective segmentation masks. Some synthesized results are showcased in Fig. 5. Our machine-generated synthetic images achieve a better mask-image alignment than real images in many cases, as shown in Fig. 6. This phenomenon occurs because human annotations tend to be imperfect due to the high difficulty of annotating segmentation masks. + +
MethodVenueBackboneCrop SizeIterationsmIoU (s.s.)mIoU (m.s.)
MaskFormer [8]NeurIPS 2021R50512160k44.546.7
Mask DINO [27]CVPR 2023R50512160k48.7-
OneFormer [19]CVPR 2023R50512160k47.3-
Mask2Former [7]CVPR 2022R50512160k47.249.2
—w/ SegGen (ours)-R50512160k49.9 (+2.7)51.4 (+2.2)
MaskFormer [8]NeurIPS 2021Swin-L640160k54.155.6
Mask DINO [27]CVPR 2023Swin-L640160k56.6-
OneFormer [19]CVPR 2023Swin-L640160k57.057.7
Mask2Former [7]CVPR 2022Swin-L640160k56.157.3
—w/ SegGen (ours)-Swin-L640160k57.4 (+1.3)58.7 (+1.4)
+ +Table 1: Semantic segmentation on ADE20K val (150 categories): Synthetic data is used as a data augmentation. With the same model structures and training recipes, our SegGen boosts the performance of Mask2Former by a large margin and establishes a new SOTA performance without extra real data under both single-scale (s.s.) and multi-scale (m.s.) test settings. + +# 3.4 Training Segmentation Models with Synthetic Data + +The final goal of this work is to improve the performance of current segmentation models with the synthetically generated training data. Therefore, we use both synthetic data produced by SegGen and training data from existing datasets in the training process of segmentation models. + +We empirically investigate two training strategies utilizing the synthetic data: (i) Synthetic data augmentation strategy. The synthetic data is used for random data augmentation. In every iteration in the training process, each real training sample is replaced by the synthetic training sample with a probability $p_{\mathrm{aug}}$ . (ii) Synthetic data pre-training strategy. It comprises two training stages: pretraining and fine-tuning. In the pre-training stage, we pre-train the segmentation models on synthetic data, so that they learn good weights that are transferable and favorable for fine-tuning. In the subsequent fine-tuning stage, the segmentation models are trained with solely human-annotated data. + +# 4 Experiments + +To accurately evaluate the effectiveness of our data generation method in improving segmentation performance, we adopt mainstream segmentation models and commonly used evaluation benchmarks for several typical segmentation tasks. The experiments are conducted mostly under the fully-supervised learning setting, meaning all human-annotated training samples from the evaluated datasets are used alongside our synthetic data. To guarantee an unbiased assessment of the impact of our synthetic data, we keep the architectures of the segmentation models and training protocols consistent with their respective original implementations throughout our studies. + +# 4.1 Implementation Details + +Segmentation Datasets and Evaluation We conduct experiments on three image segmentation benchmarks following the main experimental settings of + +Mask2Former [7]: ADE20K semantic segmentation [60], COCO panoptic segmentation, and COCO instance segmentation [33]. Our evaluation uses all 150 classes for ADE20K and 133 classes for COCO. We use all the images from the training splits in the training of segmentation models. For semantic segmentation, we show the mean Intersection-over-Union metric (mIoU). For instance segmentation, the average precision (AP) is used. For panoptic segmentation, we report panoptic quality (PQ), "thing" instance segmentation $\mathrm{AP}_{\mathrm{pan}}^{\mathrm{Th}}$ , and semantic segmentation $\mathrm{mIoU}_{\mathrm{pan}}$ . + +Segmentation Models We adopt Mask2Former [7], a recently-proposed prevalent transformer model, as the default segmentation model. Two typical backbones, i.e., R50 [16] and Swin-L [34], are studied. We keep the official implementation and training hyper-parameters of the segmentation models unchanged. For more training details please refer to their paper. We also conduct experiments on Mask DINO [27], a detection-aided segmentation model, and HRNet W48 [46], a representative fully-convolutional model. More implementation details are introduced in the supplementary materials. + +Data Generation The Text2Mask and Mask2Img models are both based on the SDXL-base model [39]. They are trained on the training splits of the target segmentation datasets (i.e. ADE20K or COCO) separately for 30,000 iterations using a learning rate of $10^{-5}$ . AdamW optimizer is employed, and the models are trained at a resolution of 768. Pre-trained weights from SDXL-base [39] are utilized. Random flipping data augmentation is used. During sampling, the default EDM sampler [22] is used. The sampling steps are 200 in the Text2Mask model and 40 in the Mask2Img model. During data sampling, for each training sample in the ADE20K semantic segmentation dataset, we produce 10 synthetic mask-image pairs using MaskSyn, resulting in 202,100 training samples. Additionally, we synthesize 50 images based on each human-labeled mask with ImgSyn, leading to a total of 1,010,500 samples. On COCO, we solely use ImgSyn for data synthesis. By generating 10 synthetic images conditioned on each human-labeled panoptic segmentation mask via ImgSyn, our synthetic set amounts to 1,182,870 synthetic samples, which are used in the training of both panoptic and instance segmentation models. + +# 4.2 Main Results + +ADE20K Semantic Segmentation We employ the synthetic data augmentation strategy with $p_{\mathrm{aug}} = 60\%$ for Mask2Former, and show the results in Table 1. SegGen significantly boosts the mIoU of Mask2Former R50 by +2.7 for single-scale inference and +2.2 for multi-scale inference, achieving 49.9 and 51.4 correspondingly. The Swin-L variant is also largely improved from 56.1/57.3 (single-scale/multi-scale) to 57.4/58.7 (+1.3/+1.4). Notably, our data synthesis method helps Mask2Former surpass the newer methods such as Mask DINO and One-Former [19], while establishing new SOTA results for R50 and Swin-L settings without using additional human-annotated data. + +COCO Panoptic Segmentation On COCO we adopt the synthetic data pretraining strategy to utilize our synthetic data. Specifically, we pre-train the mod- + +
MethodBackbone QueriesEpochsPQ\( PQ^{Th} \)\( PQ^{St} \)\( AP_{pan}^{Th} \)\( mIoU_{pan} \)
DETR [3]R50100500+2543.448.236.331.1-
MaskFormer [8]R5010030046.551.039.833.057.8
Mask2Former [7]R501005051.957.743.041.761.7
Mask DINO [27]R503005053.059.143.943.3-
Mask2Former [7]R5010050+5052.057.943.442.061.0
— w/SegGen (ours)R5010050+5052.7 (+0.7)58.8 (+0.9)43.6 (+0.2)43.1 (+1.1)62.6 (+1.6)
Mask DINO [27]R5030050+5053.459.344.444.260.5
— w/SegGen (ours)R5030050+5054.0 (+0.6)60.2 (+0.9)44.7 (+0.3)45.4 (+1.2)61.5 (+1.0)
MaskFormer [8]Swin-L10030052.758.544.040.164.8
OneFormer [19]Swin-L15010057.964.448.049.067.4
Mask2Former [7]Swin-L20010057.864.248.148.667.4
Mask DINOSwin-L3005058.365.148.050.6-
Mask2Former [7]Swin-L200100+10057.364.346.848.066.2
— w/SegGen (ours)Swin-L200100+10058.0 (+0.7)64.5 (+0.2)48.1 (+1.3)48.8 (+0.8)67.4 (+1.2)
Mask DINO [27]Swin-L30050+5058.665.448.350.467.0
— w/SegGen (ours)Swin-L30050+5059.3 (+0.7)65.9 (+0.5)49.3 (+1.0)51.1 (+0.7)68.1 (+1.1)
+ +Table 2: Panoptic segmentation on COCO panoptic val2017 (133 categories): Synthetic data is used for pre-training. Our SegGen significantly improves the performance compared with real data pre-training baselines, and establishes new SOTA performance without extra real data. + +![](images/40c149afd602b58615e14563ec5f4c51e77302485dcc7cec11d7eae6f21cb13d.jpg) +Fig.7: ADE20K mIoU of Mask2Former R50 with different augmentation probability $p_{\mathrm{aug}}$ . + +![](images/f3f8b83eb7106c77ba68509af2cf5a40779e37be6b68068d8485d7691bc073c2.jpg) +Fig.8: ADE20K mIoU of Mask2Former R50 using different numbers of synthetic samples. + +
MethodmIoU*
Pure Real Data83.3
DiffuMask (ICCV 2023)57.0
MaskSyn+ImgSyn (ours)82.2(+25.2)
MaskSyn (ours)73.2
ImgSyn (ours)80.0
+ +Table 4: Comparison with DiffuMask [48] on ADE20K: All models trained purely on synthetic data. mIoU* is defined in [48]. + +els purely on our synthetic data and fine-tune the models on real data. The results are demonstrated in Table 2. For a fair comparison, we compare with models pretrained on purely real data with the same training settings. Compared with the real data pre-trained baselines, SegGen consistently brings performance gains on all metrics with both R50 and Swin-L backbones, achieving new state-of-the-art performance without extra real data. Notably, for Mask DINO with Swin-L backbone, PQ is improved by $+0.7$ , "thing" instance segmentation $\mathrm{AP}_{\mathrm{pan}}^{\mathrm{Th}}$ is increased by $+0.7$ , and semantic segmentation $\mathrm{mIoU}_{\mathrm{pan}}$ is boosted by $+1.1$ . These noteworthy findings demonstrate that training with our synthetic data systematically improves performance across various segmentation tasks. We observe similar significant improvement on Mask2Former, further showing the effectiveness of our synthetic data on different models. + +COCO Instance Segmentation The performance on instance segmentation, both before and after incorporating our SegGen synthetic data, is detailed in Table 3. Compared to the baselines pre-trained on real data, our technique is superior across all metrics. Specifically, there is a solid $+0.5$ improvement in AP with the R50 backbone, and a $+0.8$ increase when using the Swin-L backbone. Additional assessment on the LVIS [14]-annotated COCO validation set, whose annotations are more accurate than the original version, provides additional evidence of the effectiveness of SegGen, as it enhances the average precision (AP) by $+1.2$ when used with the larger Swin-L backbone. + +
MethodBackbone QueriesEpochsAP\( AP^S \)\( AP^M \)\( AP^L \)
Mask R-CNN [15]R50anchors40042.523.845.060.0
HTC [4]R50anchors3639.722.642.250.6
QueryInst [11]R503003640.623.442.552.8
MaskFormer [8]R5010030034.016.437.854.2
Mask2Former [7]R501005043.723.447.264.8
Mask2Former [7]R501005047.126.655.472.7
Mask2Former [7]R5010050+5044.123.447.766.1
— w/ SegGen (ours)R5010050+5044.6 (+0.5)24.0 (+0.6)48.2 (+0.5)66.2 (+0.1)
Mask2Former [7]R5010050+5047.927.856.373.5
— w/ SegGen (ours)R5010050+5048.4 (+0.5)28.3 (+0.6)57.4 (+1.1)74.1 (+0.6)
Swin-HTC++ [34]Swin-Lanchors7249.531.052.467.2
QueryInst [11]Swin-L3005048.930.852.668.3
Oneformer [19]Swin-L15010048.9---
Mask2Former [7]Swin-L20010050.129.953.972.1
Mask2Former [7]Swin-L20010054.335.763.379.1
Mask2Former [7]Swin-L200100+10049.529.253.870.5
— w/ SegGen (ours)Swin-L200100+10050.3 (+0.8)31.2 (+2.0)54.3 (+0.5)72.2 (+1.7)
Mask2Former [7]Swin-L200100+10053.635.262.677.8
— w/ SegGen (ours)Swin-L200100+10054.8 (+1.2)35.9 (+0.8)64.2 (+1.6)79.2 (+1.5)
+ +Table 3: Instance segmentation on COCO val2017 (80 categories): Synthetic data is used for pre-training. The gray rows correspond to the evaluations conducted on LVIS [14] which offers instance masks of higher quality. Our SegGen significantly improves the performance of instance segmentation for various sizes of objects compared with pre-training on real data. + +Comparison with Segmentation Data Generation Method In Table 4, we compare various versions of our SegGen with DiffuMask [48], which is a recently-proposed (ICCV2023) segmentation data generation approach. All methods employ the Mask2Former R50 model and are purely trained on synthetic samples. We adhere to the evaluation setting designed by DiffuMask where the IoU for three common classes are examined. Our method markedly surpasses DiffuMask by $+25.2$ mIoU and demonstrates performance comparable to training purely with real data, highlighting the unprecedented quality of our synthetic data. + +Generalization Ability on Unseen Domains We visualize the segmentation results of Mask2Former R50 on images from PASCAL val dataset [10] in Fig. 1 and Fig. 9, comparing models trained both with and without our synthetic data. The models are trained on COCO or ADE20K. When trained using our extensively varied synthetic data, the segmentation model demonstrates notably improved performance on the unseen domain. We further study the segmentation results on AI-generated images, which are synthesized using different image generation models, in the supplementary materials. + +Visual Analysis of Generated Samples by MaskSyn We randomly select some samples generated by MaskSyn in Fig. 4 (more in supplementary materials). The generative models are trained on ADE20K. Given the text prompts, our Text2Mask model can generate diverse segmentation masks, and the Mask2Img model produces realistic synthetic images with a good alignment with the masks and text prompts. + +Visual Analysis of Generated Samples by ImgSyn We showcase the outputs by ImgSyn in Fig. 5 (more in supplementary materials). The generated images are in good agreement with the text prompts and human-labeled segmentation masks. Furthermore, we compare the mask-image alignment of the + +![](images/4fc732a14710873ab1877fa725b0cfb655db2a34ddf98116257f16a1079b58a4.jpg) +Fig. 9: Generalization Ability on unseen domains: Segmentation outputs on unseen domain (images from PASCAL [10]). The models are trained on ADE20K. Training with our SegGen demonstrates enhanced robustness towards unfamiliar domains. + +
MethodIterations Real SamplesmIoUMethodIterations Real SamplesmIoU
Baseline160K2021047.2Baseline100K2021044.4
- w/ SegGen MaskSyn160K2021048.5- w/ SegGen MaskSyn100K2021044.8
- w/ SegGen ImgSyn160K2021049.3- w/ SegGen ImgSyn100K2021045.4
- w/ SegGen ImgSyn + MaskSyn160K2021049.9- w/ SegGen ImgSyn + MaskSyn100K2021045.6
Baseline160K100023.7Baseline100K100023.6
- w/ SegGen MaskSyn160K100026.5- w/ SegGen MaskSyn100K100026.0
- w/ SegGen ImgSyn160K100024.1- w/ SegGen ImgSyn100K100023.9
- w/ SegGen ImgSyn + MaskSyn160K100027.1- w/ SegGen ImgSyn + MaskSyn100K100026.9
+ +Table 5: Ablation Study of MaskSyn and ImgSyn on ADE20K: Both MaskSyn and ImgSyn notably enhance the performance of Mask2Former and HRNet, while MaskSyn has a more pronounced impact when there are fewer real samples available. + +real training samples and our synthetic training samples in Fig. 6. Our synthetic samples exhibit superior alignment quality in many cases, compared to human annotations which are often imperfect on the boundaries. + +# 4.3 Quantitative Analysis + +Effectiveness of MaskSyn and ImgSyn We evaluate the data generated by MaskSyn and ImgSyn separately and jointly, by examining their impacts on the performance of segmentation models, as shown in Table 5. The experiments are conducted on ADE20K with Mask2Former R50 and HRNet W48. We find that the training samples generated by both MaskSyn and ImgSyn bring significant performance gains compared with the baseline. When combining these two types of synthetic data together, the segmentation model can achieve the best performance. We delve deeper into a scenario where only 1,000 real training samples are available in the training of both generative models and segmentation models. It is observed that MaskSyn substantially enhances the segmentation results in situations with limited real data. This could be attributed to the ability of MaskSyn to amplify data diversity by creating entirely new segmentation masks and images. Moreover, the significant performance improvements achieved by our SegGen, trained on merely 1,000 real samples, demonstrate the data efficiency and robustness of our proposal. + +Influence of Synthetic Data Augmentation Probability $p_{\mathrm{aug}}$ on ADE20K We investigate the effect of $p_{\mathrm{aug}}$ in Fig. 7. When $p_{\mathrm{aug}} = 0$ , it implies training using only real data, whereas $p_{\mathrm{aug}} = 100\%$ means training entirely with synthetic + +
MethodIterationsmIoU
Syn. Pre-Train160K+160K47.2
Syn. Data Aug160K49.9
+ +(a) ADE20K Semantic Segmentation: Using synthetic data as random data augmentation obtains the best performance on ADE20K. + +
MethodEpochsPQAP
Syn. Pre-Training50+5052.744.6
Syn. Data Aug5051.343.2
+ +(b) COCO Panoptic Segmentation and Instance Segmentation: Using synthetic data for pre-training achieves better performance on COCO dataset. + +Table 6: Different training strategies with synthetic data using Mask2Former R50. + +data. The best results are achieved at $p_{\mathrm{aug}} = 60\%$ . It is noteworthy that training exclusively on our synthetic data results in an impressive $44.8~\mathrm{mIoU}$ + +Influence of Synthetic Data Size To evaluate the impact of synthetic data volume, we use different numbers of synthetic training samples produced by ImgSyn in training and report the results in Fig. 8. $p_{\mathrm{aug}}$ is fixed at $60\%$ . We notice a substantial improvement in performance when increasing the synthetic data quantity from 20K to 200K, underscoring the importance of collecting a significantly larger synthetic dataset to avoid overfitting. The performance appears to plateau after 600K. More studies on the scale of synthetic data are presented in the supplementary materials. + +Influence of Training Strategies We conduct an ablation study using different training strategies with synthetic data on ADE20K (Table 6a) and COCO (Table 6b). We find that using the synthetic data augmentation strategy achieves better performance on ADE20K semantic segmentation, While on COCO panoptic and instance segmentation, the synthetic pre-training strategy works better. We believe that due to the limited size of the ADE20K training set, which contains merely $\sim 20\mathrm{k}$ images, segmentation models easily overfit the scarcely-available training images and segmentation layouts. Utilizing synthetic data (including synthetic masks and images) for data augmentation can help mitigate such an overfitting problem. On the other hand, for COCO, given its larger training set size (around 100K images), overfitting is less of a concern. Hence, it is sufficient to provide the COCO models with good initial weights pre-trained on our synthetic data. Meanwhile, previous work [14] finds that the annotation accuracy of COCO is significantly more biased than ADE20K, a finding that aligns with our visualization results depicted in Fig. 6. A domain gap exists between the COCO data and our high-quality synthetic data. Hence, employing our method for pre-training and incorporating real data for fine-tuning on COCO is a more favorable approach. + +# 5 Conclusion + +We present SegGen, a highly-effective data synthesis method for image segmentation. SegGen introduces a revolutionary segmentation data generation framework, and proposes two data generation approaches, namely MaskSyn and ImgSyn, to synthesize segmentation masks and images, with the help of the designed text-to-mask and mask-to-image generation models. Our method not only strongly enhances the model performance on semantic, panoptic, and instance segmentation benchmarks, but also substantially improves segmentation generalization ability in unseen data distribution. + +# Acknowledgement + +This research is supported in part by the Early Career Scheme of the Research Grants Council (RGC) of the Hong Kong SAR under grant No. 26202321, SAIL Research Project, HKUST-Zeekr Collaborative Research Fund, HKUST-WeBank Joint Lab Project, and Tencent Rhino-Bird Focused Research Program. + +# References + +1. Baranchuk, D., Rubachev, I., Voynov, A., Khrulkov, V., Babenko, A.: Label-efficient semantic segmentation with diffusion models. In: ICLR (2022) +2. Brock, A., Donahue, J., Simonyan, K.: Large scale gan training for high fidelity natural image synthesis. In: ICLR (2019) +3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: ECCV (2020) +4. Chen, K., Pang, J., Wang, J., Xiong, Y., Li, X., Sun, S., Feng, W., Liu, Z., Shi, J., Ouyang, W., et al.: Hybrid task cascade for instance segmentation. In: CVPR (2019) +5. Chen, L.C., Papandreou, G., Kokkinos, I., Murphy, K., Yuille, A.L.: Semantic image segmentation with deep convolutional nets and fully connected crfs. In: ICLR (2015) +6. Chen, T., Li, L., Saxena, S., Hinton, G., Fleet, D.J.: A generalist framework for panoptic segmentation of images and videos. arXiv preprint arXiv:2210.06366 (2022) +7. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: CVPR (2022) +8. Cheng, B., Schwing, A.G., Kirillov, A.: Per-pixel classification is not all you need for semantic segmentation. In: NeurIPS (2021) +9. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. In: NeruIPS (2021) +0. Everingham, M., Eslami, S.A., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The PASCAL visual object classes challenge: A retrospective. IJCV (2015) +1. Fang, Y., Yang, S., Wang, X., Li, Y., Fang, C., Shan, Y., Feng, B., Liu, W.: Instances as queries. In: ICCV (2021) +2. Forever, A.: Kandinsky-2. GitHub repository (2023), https://github.com/ai-forever/Kandinsky-2 +3. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. CACM (2020) +4. Gupta, A., Dollar, P., Girshick, R.: LVIS: A dataset for large vocabulary instance segmentation. In: ICCV (2019) +5. He, K., Gkioxari, G., Dolkar, P., Girshick, R.: Mask R-CNN. In: ICCV (2017) +6. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016) +7. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeruIPS (2020) +8. Huang, L., Chen, D., Liu, Y., Yujun, S., Zhao, D., Jingren, Z.: Composer: Creative and controllable image synthesis with composable conditions. arXiv preprint arxiv:2302.09778 (2023) + +19. Jain, J., Li, J., Chiu, M.T., Hassani, A., Orlov, N., Shi, H.: Oneformer: One transformer to rule universal image segmentation. In: CVPR (2023) +20. Ji, Y., Chen, Z., Xie, E., Hong, L., Liu, X., Liu, Z., Lu, T., Li, Z., Luo, P.: Ddp: Diffusion model for dense visual prediction. arXiv preprint arXiv:2303.17559 (2023) +21. Kang, M., Zhu, J.Y., Zhang, R., Park, J., Shechtman, E., Paris, S., Park, T.: Scaling up gans for text-to-image synthesis. In: CVPR (2023) +22. Karras, T., Aittala, M., Aila, T., Laine, S.: Elucidating the design space of diffusion-based generative models. In: NeruIPS (2022) +23. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: CVPR (2019) +24. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013) +25. Kirillov, A., He, K., Girshick, R., Rother, C., Dlár, P.: Panoptic segmentation. In: CVPR (2019) +26. Li, D., Ling, H., Kim, S.W., Kreis, K., Fidler, S., Torralba, A.: Bigdatasetgan: Synthesizing imagenet with pixel-wise annotations. In: CVPR (2022) +27. Li, F., Zhang, H., Xu, H., Liu, S., Zhang, L., Ni, L.M., Shum, H.Y.: Mask dino: Towards a unified transformer-based framework for object detection and segmentation. In: CVPR (2023) +28. Li, J., Li, D., Savarese, S., Hoi, S.: BLIP-2: bootstrapping language-image pretraining with frozen image encoders and large language models. In: ICML (2023) +29. Li, X., Zhong, Z., Wu, J., Yang, Y., Lin, Z., Liu, H.: Expectation-maximization attention networks for semantic segmentation. In: ICCV (2019) +30. Li, X., Ding, H., Zhang, W., Yuan, H., Pang, J., Cheng, G., Chen, K., Liu, Z., Loy, C.C.: Transformer-based visual segmentation: A survey. arXiv preprint arXiv:2304.09854 (2023) +31. Li, Y., Liu, H., Wu, Q., Mu, F., Yang, J., Gao, J., Li, C., Lee, Y.J.: Gligen: Open-set grounded text-to-image generation. In: CVPR (2023) +32. Li, Z., Zhou, Q., Zhang, X., Zhang, Y., Wang, Y., Xie, W.: Open-vocabulary object segmentation with diffusion models. In: ICCV (2023) +33. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólár, P., Zitnick, C.L.: Microsoft COCO: Common objects in context. In: ECCV (2014) +34. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. arXiv:2103.14030 (2021) +35. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: CVPR (2015) +36. Mou, C., Wang, X., Xie, L., Wu, Y., Zhang, J., Qi, Z., Shan, Y., Qie, X.: T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453 (2023) +37. Nguyen, Q.N.T.V.A.T.K.: Dataset diffusion: Diffusion-based synthetic dataset generation for pixel-level semantic segmentation. arXiv preprint arXiv:2309.14303 (2023) +38. PNVR, K., Singh, B., Ghosh, P., Siddiquie, B., Jacobs, D.: Ld-znet: A latent diffusion approach for text-based image segmentation. In: ICCV (2023) +39. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952 (2023) +40. Qi, L., Kuen, J., Wang, Y., Gu, J., Zhao, H., Torr, P., Lin, Z., Jia, J.: Open world entity segmentation. TPAMI (2022) + +41. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022) +42. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. In: NeruIPS (2022) +43. Sauer, A., Karras, T., Laine, S., Geiger, A., Aila, T.: StyleGAN-T: Unlocking the power of GANs for fast large-scale text-to-image synthesis. In: ICML (2023) +44. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: ICML (2015) +45. Wang, H., Cao, J., Anwer, R.M., Xie, J., Khan, F.S., Pang, Y.: Dformer: Diffusion-guided transformer for universal image segmentation. arXiv preprint arXiv:2306.03437 (2023) +46. Wang, J., Sun, K., Cheng, T., Jiang, B., Deng, C., Zhao, Y., Liu, D., Mu, Y., Tan, M., Wang, X., Liu, W., Xiao, B.: Deep high-resolution representation learning for visual recognition. TPAMI (2019) +47. Wu, W., Zhao, Y., Chen, H., Gu, Y., Zhao, R., He, Y., Zhou, H., Shou, M.Z., Shen, C.: Datasetsd: Synthesizing data with perception annotations using diffusion models. arXiv preprint arXiv:2308.06160 (2023) +48. Wu, W., Zhao, Y., Shou, M.Z., Zhou, H., Shen, C.: Diffumask: Synthesizing images with pixel-level annotations for semantic segmentation using diffusion models. In: ICCV (2023) +49. Xie, J., Li, W., Li, X., Liu, Z., Ong, Y.S., Loy, C.C.: Mosaic fusion: Diffusion models as data augmenters for large vocabulary instance segmentation. arXiv preprint arXiv:2309.13042 (2023) +50. Xu, J., Liu, S., Vahdat, A., Byeon, W., Wang, X., De Mello, S.: Open-vocabulary panoptic segmentation with text-to-image diffusion models. In: CVPR (2023) +51. Yang, L., Xu, X., Kang, B., Shi, Y., Zhao, H.: Freemask: Synthetic images with dense annotations make stronger segmentation models. In: NeurIPS (2023) +52. Ye, H., Xu, D.: Taskprompter: Spatial-channel multi-task prompting for dense scene understanding. In: ICLR (2023) +53. Zeng, Y., Lin, Z., Zhang, J., Liu, Q., Collomosse, J., Kuen, J., Patel, V.M.: Scenecomposer: Any-level semantic image synthesis. In: CVPR (2023) +54. Zhang, L., Anyi, R., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: ICCV (2023) +55. Zhang, W., Pang, J., Chen, K., Loy, C.C.: K-net: Towards unified image segmentation. In: NeurIPS (2021) +56. Zhang, Y., Ling, H., Gao, J., Yin, K., Lafleche, J.F., Barriuso, A., Torralba, A., Fidler, S.: Datasetgan: Efficient labeled data factory with minimal human effort. In: CVPR (2021) +57. Zhao, H., Shi, J., Qi, X., Wang, X., Jia, J.: Pyramid scene parsing network. In: CVPR (2017) +58. Zhao, S., Chen, D., Chen, Y.C., Bao, J., Hao, S., Yuan, L., Wong, K.Y.K.: Uncontrolnet: All-in-one control to text-to-image diffusion models. NeurIPS (2023) +59. Zhao, W., Rao, Y., Liu, Z., Liu, B., Zhou, J., Lu, J.: Unleashing text-to-image diffusion models for visual perception. In: ICCV (2023) +60. Zhou, B., Zhao, H., Puig, X., Fidler, S., Barriuso, A., Torralba, A.: Scene parsing challenge 2016. http://sceneParsing.csail.mit.edu/index_challenge.htm (2016) +61. Zhou, B., Zhao, H., Puig, X., Fidler, S., Barriuso, A., Torralba, A.: Scene parsing through ADE20K dataset. In: CVPR (2017) \ No newline at end of file diff --git a/2024/SegGen_ Supercharging Segmentation Models with Text2Mask and Mask2Img Synthesis/images.zip b/2024/SegGen_ Supercharging Segmentation Models with Text2Mask and Mask2Img Synthesis/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6a4a9d79fdc855cfd6bfe781fbeec9aa3b4407f1 --- /dev/null +++ b/2024/SegGen_ Supercharging Segmentation Models with Text2Mask and Mask2Img Synthesis/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:12d3eb18f5de1c9566006ebc97c83496d6f834943a4980117332968e99c1f6fe +size 683734 diff --git a/2024/SegGen_ Supercharging Segmentation Models with Text2Mask and Mask2Img Synthesis/layout.json b/2024/SegGen_ Supercharging Segmentation Models with Text2Mask and Mask2Img Synthesis/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b16dcace448bed84f49f08b8fa3f3ee474f2e790 --- /dev/null +++ b/2024/SegGen_ Supercharging Segmentation Models with Text2Mask and Mask2Img Synthesis/layout.json @@ -0,0 +1,7745 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 149, + 111, + 465, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 111, + 465, + 148 + ], + "spans": [ + { + "bbox": [ + 149, + 111, + 465, + 148 + ], + "type": "text", + "content": "SegGen: Supercharging Segmentation Models with Text2Mask and Mask2Img Synthesis" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 147, + 167, + 465, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 167, + 465, + 180 + ], + "spans": [ + { + "bbox": [ + 147, + 167, + 465, + 180 + ], + "type": "text", + "content": "Hanrong Ye" + }, + { + "bbox": [ + 147, + 167, + 465, + 180 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 147, + 167, + 465, + 180 + ], + "type": "text", + "content": ", Jason Kuen" + }, + { + "bbox": [ + 147, + 167, + 465, + 180 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 147, + 167, + 465, + 180 + ], + "type": "text", + "content": ", Qing Liu" + }, + { + "bbox": [ + 147, + 167, + 465, + 180 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 147, + 167, + 465, + 180 + ], + "type": "text", + "content": ", Zhe Lin" + }, + { + "bbox": [ + 147, + 167, + 465, + 180 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 147, + 167, + 465, + 180 + ], + "type": "text", + "content": ", Brian Price" + }, + { + "bbox": [ + 147, + 167, + 465, + 180 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 147, + 167, + 465, + 180 + ], + "type": "text", + "content": ", Dan Xu" + }, + { + "bbox": [ + 147, + 167, + 465, + 180 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 269, + 189, + 344, + 211 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 271, + 189, + 341, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 271, + 189, + 341, + 200 + ], + "spans": [ + { + "bbox": [ + 271, + 189, + 341, + 200 + ], + "type": "text", + "content": "1 CSE, HKUST" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 269, + 201, + 344, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 269, + 201, + 344, + 211 + ], + "spans": [ + { + "bbox": [ + 269, + 201, + 344, + 211 + ], + "type": "text", + "content": "2 Adobe Research" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 160, + 243, + 452, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 243, + 452, + 517 + ], + "spans": [ + { + "bbox": [ + 160, + 243, + 452, + 517 + ], + "type": "text", + "content": "Abstract. We present SegGen, a new data generation approach that pushes the performance boundaries of state-of-the-art image segmentation models. One major bottleneck of previous data synthesis methods for segmentation is the design of \"segmentation labeler module\", which is used to synthesize segmentation masks for images [48]. The segmentation labeler modules, which are segmentation models by themselves, bound the performance of downstream segmentation models trained on the synthetic masks. These methods encounter a \"chicken or egg dilemma\" and thus fail to outperform existing segmentation models. To address this issue, we propose a novel method that reverses the traditional data generation process: we first (i) generate highly diverse segmentation masks that match real-world distribution from text prompts, and then (ii) synthesize realistic images conditioned on the segmentation masks. In this way, we avoid the need for any segmentation labeler module. SegGen integrates two data generation strategies, namely MaskSyn and ImgSyn, to largely improve data diversity in synthetic masks and images. Notably, the high quality of our synthetic data enables our method to outperform the previous data synthesis method [48] by " + }, + { + "bbox": [ + 160, + 243, + 452, + 517 + ], + "type": "inline_equation", + "content": "+25.2" + }, + { + "bbox": [ + 160, + 243, + 452, + 517 + ], + "type": "text", + "content": " mIoU on ADE20K when trained with pure synthetic data. On the highly competitive ADE20K and COCO benchmarks, our data generation method markedly improves the performance of state-of-the-art segmentation models in semantic segmentation, panoptic segmentation, and instance segmentation. Moreover, experiments show that training with our synthetic data makes the segmentation models more robust towards unseen data domains, including real-world and AI-generated images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 160, + 529, + 447, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 529, + 447, + 540 + ], + "spans": [ + { + "bbox": [ + 160, + 529, + 447, + 540 + ], + "type": "text", + "content": "Keywords: Image Segmentation " + }, + { + "bbox": [ + 160, + 529, + 447, + 540 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 529, + 447, + 540 + ], + "type": "text", + "content": " Image Generation " + }, + { + "bbox": [ + 160, + 529, + 447, + 540 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 529, + 447, + 540 + ], + "type": "text", + "content": " Deep Learning" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 579, + 229, + 592 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 579, + 229, + 592 + ], + "spans": [ + { + "bbox": [ + 133, + 579, + 229, + 592 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Image segmentation explores the identification of objects in visual inputs at the pixel level. Based on the different emphases on category and instance membership information, researchers have divided image segmentation into several tasks [5,25,35,40]. For example, semantic segmentation studies pixel-level understanding of object categories, instance segmentation focuses on instance grouping" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 482, + 242 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 482, + 242 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 482, + 242 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 482, + 242 + ], + "type": "image", + "image_path": "2abb841d4bbb9857413b2985a699cd0e8646bc00e92b8d8c9523bda968ee9011.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 242, + 482, + 342 + ], + "lines": [ + { + "bbox": [ + 130, + 242, + 482, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 242, + 482, + 342 + ], + "type": "text", + "content": "Fig. 1: Effectiveness of SegGen on various data domains: Through training with synthetic data generated by the proposed SegGen, we significantly boost the performance of state-of-the-art segmentation model Mask2Former [7] on evaluation benchmarks including ADE20K [60] and COCO [33], whilst making it more robust towards challenging images from other domains (the three columns on the left are from PASCAL [10]; the three on the right are synthesized by image generation model Kandinsky 2 [12]). SegGen outperforms the previous best data generation method (DiffuMask [48]) by a huge margin when models are trained on pure synthetic data. \"mIoU*\" is the average IoU metric defined by [48] which focuses on three common classes." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 351, + 482, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 351, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 351, + 482, + 495 + ], + "type": "text", + "content": "of pixels, while panoptic segmentation considers both. For all these segmentation tasks, obtaining high-quality annotation is challenging as every individual pixel requires human labeling, and a single image can contain millions of pixels. Therefore, compared to other public datasets like ImageNet-21K (with around 14M images), the prevailing human-annotated segmentation datasets are notably smaller. For example, ADE20K dataset [61] contains about 20K images in its training split, while COCO [33] has around 118K training images. Although there has been significant development in the structure of segmentation models [8,29,52,55,57], the limited size of training data hinders further performance enhancements and results in inadequate generalization ability to handle images from unfamiliar domains, such as those from other scenes or synthesized by generative models as shown in the second row of Fig. 1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 498, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 483, + 666 + ], + "type": "text", + "content": "Inspired by the recent success of image generation [9, 23], researchers start to explore using generative models to enhance image segmentation [1]. A representative direction of this research focuses on synthesizing segmentation training data in a cost-effective manner [26]. In related methods, the synthetic segmentation masks are obtained from some manually designed \"segmentation labeler modules\". The segmentation labeler modules are essentially segmentation models by themselves. They are either small-scale segmentation networks as in DatasetGAN [56] and Grounded Diffusion [32], or post-processing methods on the image features as in DiffuMask [48]. The performance of the downstream segmentation models is constrained by the quality of the synthetic masks they are trained on, which in turn relies on the capabilities of those segmentation labeler modules. This is a \"chicken or egg dilemma\" and the performance bottleneck is the segmentation labeler module. Therefore, while they achieve encouraging success in settings with limited training data, their methods fail to deliver notable en" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 239, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 239, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 239, + 102 + ], + "type": "text", + "content": "Hanrong Ye et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 115, + 482, + 177 + ], + "blocks": [ + { + "bbox": [ + 137, + 115, + 482, + 177 + ], + "lines": [ + { + "bbox": [ + 137, + 115, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 137, + 115, + 482, + 177 + ], + "type": "image", + "image_path": "53e2a0790b39b289b0717afba00ada1f43d88652c137dca938dd01b9f8c52f99.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 178, + 482, + 288 + ], + "lines": [ + { + "bbox": [ + 130, + 178, + 482, + 288 + ], + "spans": [ + { + "bbox": [ + 130, + 178, + 482, + 288 + ], + "type": "text", + "content": "Fig. 2: Comparison with previous data generation methods for segmentation: (a) Earlier methods [48, 56] rely on segmentation labeler modules to produce segmentation masks for synthetic images. However, the performance of these downstream segmentation models, trained on synthetic data, is bounded by the capacity of the segmentation labeler modules. The segmentation labeler is a major bottleneck for the quality of the generated data. (b) We design a reverse pipeline: we first create diverse new masks from text prompts via a proposed Text2Mask generation model and then synthesize images conditioned on the segmentation masks. This methodology avoids any usage of segmentation labeler networks, resulting in significantly improved quality of our synthesis data." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 301, + 480, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 301, + 480, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 301, + 480, + 326 + ], + "type": "text", + "content": "hancements for the state-of-the-art models on the most popular segmentation benchmarks in settings using the complete training sets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 327, + 482, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 482, + 508 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 482, + 508 + ], + "type": "text", + "content": "On the other hand, a handful of works [54, 58] have demonstrated that high-quality synthetic images can be generated by conditioning text-to-image models on dense input maps (e.g., segmentation masks or canny edge). Moreover, these dense-input conditional models provide a strong alignment between the dense input label maps and generated images. These two qualities motivate us to leverage the powerful capabilities of such models for effective segmentation data generation. To this end, we propose a novel segmentation data generation method, coined as SegGen, for generating high-quality segmentation training data. We reverse the traditional data generation pipeline for segmentation: we first synthesize high-quality segmentation masks from a proposed text-to-mask generation model, and then generate images based on the segmentation masks using a mask-to-image generation model. In this new framework, we avoid using any segmentation labeler modules, which results in significantly improved data quality. Fig. 2 shows the difference between our method and previous data generation methods for segmentation [48, 56]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "As each training sample of image segmentation consists of two components: segmentation masks and the corresponding image, we develop two novel data generation approaches, emphasizing improvements in two distinct aspects of data diversity: (i) segmentation masks and (ii) images. The first data generation approach, named MaskSyn, centers around the generation of new segmentation masks. It learns a text-to-mask (Text2Mask) generation model to produce completely new segmentation masks given text prompts. Then, it learns a mask-to-image (Mask2Img) generation model to synthesize images that align with the synthetic segmentation masks. The second data generation approach, named ImgSyn, utilizes the above-mentioned Mask2Img model to synthesize new images given human-annotated segmentation masks. With MaskSyn and ImgSyn, we can readily generate a vast array of diverse and high-quality synthetic training data. The combined synthetic data is used to train segmenta" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 415, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 415, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 415, + 91, + 447, + 102 + ], + "type": "text", + "content": "SegGen" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": "tion models in conjunction with real training samples from human-annotated datasets. Experiments show that our synthetic data can significantly boost the performance of the image segmentation models on challenging benchmarks including ADE20K semantic segmentation, COCO panoptic segmentation, and COCO instance segmentation, achieving new state-of-the-art performances without using extra human-annotated data. Notably, SegGen remarkably outperforms the previous segmentation data generation method [48] by " + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "inline_equation", + "content": "+25.2" + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": " mIoU when trained with pure synthetic data on ADE20K. Our method boosts the mIoU of Mask2Former by " + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "inline_equation", + "content": "+2.7" + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": " (R50) and " + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "inline_equation", + "content": "+1.3" + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": " (Swin-L) on the ADE20K semantic segmentation benchmark. Furthermore, segmentation models trained on our synthetic data exhibit a remarkably stronger ability to generalize across unfamiliar image domains, including real images from other distributions and machine-generated images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 272, + 482, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 272, + 482, + 452 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 482, + 452 + ], + "type": "text", + "content": "We summarize the contribution of this work in three points: (i) We propose a revolutionary generation framework that reverses the traditional segmentation data generation pipeline and solves the \"chicken or egg dilemma\". The new framework enables us to produce high-quality segmentation training data at scale, thus enabling the training of more powerful image segmentation models. (ii) We introduce two effective generative models, one for text-to-mask generation and the other for mask-to-image generation. Based on these models, we propose two novel segmentation training data generation approaches, namely MaskSyn and ImgSyn. They significantly improve the data diversity, with MaskSyn focusing on new segmentation masks and ImgSyn on new images. (iii) SegGen successfully improves the performance of the leading-edge segmentation models across the highly competitive benchmarks on ADE20K and COCO. Moreover, SegGen enhances the generalization ability of segmentation models towards unseen image domains. Rigorous experiments, including ablation study and peer comparison, strongly suggest the effectiveness of the proposed method." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 471, + 237, + 484 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 471, + 237, + 484 + ], + "spans": [ + { + "bbox": [ + 132, + 471, + 237, + 484 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "Generation for Segmentation Image segmentation is one of the most studied visual perception problems [30]. In recent years, there has been a surge in efforts to harness the capabilities of generative models for segmentation tasks. These efforts can be broadly classified into three categories based on their methodologies: (i) Extracting visual features from generative models for segmentation [1, 38, 50, 59]. These methods harvest the strong representation ability of diffusion models trained on large-scale datasets but are limited in the precision of predicted masks. (ii) Formulating segmentation tasks directly as generative models [6, 20, 45]. Although these methods propose exciting new model architectures, they usually consume higher computational costs while showing unimproved performance compared with conventional segmentation models. (iii) Synthesizing segmentation training data using generative models [26, 32, 37, 47-49, 56]. While these methods have showcased commendable results against their respective baselines especially when the training data is highly limited, they have" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 239, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 239, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 239, + 102 + ], + "type": "text", + "content": "Hanrong Ye et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": "yet to exhibit notably superior performance on the most rigorous benchmarks including ADE20K (150 categories) [61] and COCO (133 categories) [33] under fully-supervised setting. A primary concern with these techniques is the subpar quality of the generated segmentation masks. This stems from their dependence on the bottleneck segmentation labeler modules during the synthetic mask generation process. Therefore, a concurrent work [51] abandons the generation of segmentation masks and directly generates images based on existing masks, which is highly limited in mask diversity. In contrast, our data generation workflow reverses the traditional data generation pipeline: we first generate segmentation masks from text, and then synthesize images based on segmentation masks. Our method avoids the need for any segmentation labeler which results in largely improved data quality. Our SegGen can strongly enhance the existing state-of-the-art segmentation models on different challenging benchmarks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 272, + 482, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 272, + 482, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 482, + 429 + ], + "type": "text", + "content": "Conditional Image Synthesis Within the realm of conditional image synthesis, Generative Adversarial Networks [2, 13, 21, 23, 43], Variational Autoencoders [24], and Diffusion Models [9, 17, 42, 44] have been at the forefront. Recently, the open-source research community has shown a burgeoning interest in the latent diffusion-based Stable Diffusion (SD) series for text-to-image synthesis [41]. The most recent iteration, SDXL, introduced by [39], expands the model capacity, yielding significantly enhanced results. Therefore, we build SegGen upon SDXL. Regarding mask-to-image generation models [18, 31, 53], ControlNet [54] and T2I-Adapter [36] suggest freezing the parameters of the SD model and introducing a set of more compact, learnable modules. This approach enables image generation conditioned on the given segmentation masks. We adopt the structure of ControlNet in our mask-to-image generation model. To the best of our knowledge, there have been no prior endeavors on text-to-mask generation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 444, + 202, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 444, + 202, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 444, + 202, + 456 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 468, + 208, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 208, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 208, + 479 + ], + "type": "text", + "content": "3.1 Overview" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 487, + 482, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 482, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 482, + 559 + ], + "type": "text", + "content": "SegGen is designed to synthesize high-quality training samples for improving the performance of segmentation models. The overall workflow is shown on the left of Fig. 3. We first train SegGen with human-annotated training samples from public datasets. After training, we use SegGen to produce new segmentation training samples at scale. The generated training samples are incorporated into the training process of segmentation models to enhance the model performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 574, + 253, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 574, + 253, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 574, + 253, + 586 + ], + "type": "text", + "content": "3.2 Models in SegGen" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "As shown in Fig. 3, we first utilize a captioner model to extract captions of the real training images as text prompts from the target dataset. The text prompts will condition the data generation process. Then, two conditional generative models are introduced: a text-to-mask (Text2Mask) generation model and a mask-to-image (Mask2Img) generation model. Both generative models are built upon the SDXL model [39] which provides top-notch image generation quality." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 415, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 415, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 415, + 91, + 447, + 102 + ], + "type": "text", + "content": "SegGen" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 118, + 482, + 212 + ], + "blocks": [ + { + "bbox": [ + 133, + 118, + 482, + 212 + ], + "lines": [ + { + "bbox": [ + 133, + 118, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 133, + 118, + 482, + 212 + ], + "type": "image", + "image_path": "9a4bc0c5cf53f18d4e5ccdec7be3fb1a83f078c3c5a3da5f075adf4afb65aa36.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 214, + 482, + 314 + ], + "lines": [ + { + "bbox": [ + 130, + 214, + 482, + 314 + ], + "spans": [ + { + "bbox": [ + 130, + 214, + 482, + 314 + ], + "type": "text", + "content": "Fig.3: Illustration of the workflow of our proposed SegGen: We introduce two generative models: a text-to-mask (Text2Mask) generation model and a mask-to-image (Mask2Img) generation model, based on which we design two approaches for synthesizing segmentation training samples: MaskSyn and ImgSyn. (a) MaskSyn focuses on generating new segmentation masks. It first extracts the caption of the real image as a text prompt and uses it to generate new masks with the Text2Mask model. Then, the new masks and text prompt are fed into the Mask2Img model to produce the corresponding new images. (b) ImgSyn focuses on the synthesis of new images. It inputs human-labeled masks and text prompts into the Mask2Img model." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 324, + 482, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 324, + 482, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 324, + 482, + 384 + ], + "type": "text", + "content": "Captioner Model To obtain image captions of existing training samples, we employ BLIP2-FlanT5xxl [28] model, which is a state-of-the-art vision-language model. We feed the text prompt \"Question: What are shown in the photo? Answer:\" and image as input to the model. The responses serve as text prompts to condition the following generation process." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 385, + 482, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 385, + 482, + 566 + ], + "spans": [ + { + "bbox": [ + 130, + 385, + 482, + 566 + ], + "type": "text", + "content": "Text2Mask Model We design a Text2Mask model in SegGen for generating diverse segmentation masks based on given text prompts. To leverage the generation capacity of text-to-image generation models pre-trained on large-scale datasets, we encode the segmentation masks (the pixel values are category IDs) as three-channel RGB-like color maps, where one color represents a certain category. From our experiments, the color map reconstructed by VAE [41] in SDXL appears almost indistinguishable from the original input as shown in the supplementary materials. Therefore, we can directly fine-tune the text-to-image SDXL-base model with [text, segmentation color map] training pairs, which are from the public image segmentation dataset (e.g. ADE20K). During sampling, our Text2Mask model can generate diverse color maps conditioned on text prompts, which are subsequently converted into segmentation masks. Formally, suppose the input text prompt is " + }, + { + "bbox": [ + 130, + 385, + 482, + 566 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 130, + 385, + 482, + 566 + ], + "type": "text", + "content": ", the target height and width are " + }, + { + "bbox": [ + 130, + 385, + 482, + 566 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 130, + 385, + 482, + 566 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 385, + 482, + 566 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 385, + 482, + 566 + ], + "type": "text", + "content": ", the synthesized color map is " + }, + { + "bbox": [ + 130, + 385, + 482, + 566 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{\\mathrm{syn}} \\in \\mathbb{W}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 130, + 385, + 482, + 566 + ], + "type": "text", + "content": ", and the synthesized segmentation map (with " + }, + { + "bbox": [ + 130, + 385, + 482, + 566 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 385, + 482, + 566 + ], + "type": "text", + "content": " masks) is " + }, + { + "bbox": [ + 130, + 385, + 482, + 566 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathrm{syn}} \\in \\mathbb{W}^{H \\times W \\times N}" + }, + { + "bbox": [ + 130, + 385, + 482, + 566 + ], + "type": "text", + "content": ", the generation process is:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 249, + 577, + 480, + 605 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 577, + 480, + 605 + ], + "spans": [ + { + "bbox": [ + 249, + 577, + 480, + 605 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {C} _ {\\mathrm {s y n}} = \\operatorname {T e x t 2 M a s k} (\\mathbf {T}), \\\\ \\mathbf {M} _ {\\text {s y n}} = f _ {\\text {c o l o r} \\rightarrow \\text {m a s k}} \\left(\\mathbf {C} _ {\\text {s y n}}\\right), \\tag {1} \\\\ \\end{array}", + "image_path": "6b70d9abf1bfa70717f923cc8babc62bcf67a0f1a16af1a1953dbc9e4e779861.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{color} \\to \\mathrm{mask}}: \\mathbb{W}^{H \\times W \\times 3} \\to \\mathbb{W}^{H \\times W \\times N}" + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "content": " is the function that projects the color maps to segmentation masks. Specifically, for each pixel on the color maps, we identify its nearest color (in Euclidean space) in the aforementioned lookup table, and assign the corresponding class to the pixel in the segmentation masks." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "type": "text", + "content": "Hanrong Ye et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 225 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 225 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 225 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 225 + ], + "type": "image", + "image_path": "c9017728c00b05353ee46f2c44a7598961464a543b8f2a9149d487578b592409.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 227, + 482, + 270 + ], + "lines": [ + { + "bbox": [ + 130, + 227, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 227, + 482, + 270 + ], + "type": "text", + "content": "Fig. 4: Generated samples by MaskSyn on ADE20K: The third row overlays the mask and the image together to demonstrate the alignment between them. The generated segmentation masks and images demonstrate high perceptual quality and excellent alignment (see more samples in supplementary materials)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 289, + 482, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 410 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 410 + ], + "type": "text", + "content": "Mask2Img Model The goal of the Mask2Img model is to synthesize new images that align well with the given segmentation masks and text prompts. Specifically we adopt the structure of ControlNet [54]: we freeze the pre-trained weights of the SDXL-base model and train an additional side network for mask-conditioned image generation. It simultaneously keeps the generalization ability of the pretrained diffusion model and provides excellent controllable generation ability. The Mask2Img model is trained with the [text, segmentation color map, image] triplets gathered from the training splits of the target datasets. We denote the input segmentation map as " + }, + { + "bbox": [ + 130, + 289, + 482, + 410 + ], + "type": "inline_equation", + "content": "\\mathbf{M} \\in \\mathbb{W}^{H \\times W \\times N}" + }, + { + "bbox": [ + 130, + 289, + 482, + 410 + ], + "type": "text", + "content": ", the color map as " + }, + { + "bbox": [ + 130, + 289, + 482, + 410 + ], + "type": "inline_equation", + "content": "\\mathbf{C} \\in \\mathbb{W}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 130, + 289, + 482, + 410 + ], + "type": "text", + "content": ", the synthetic image as " + }, + { + "bbox": [ + 130, + 289, + 482, + 410 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{syn}} \\in \\mathbb{W}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 130, + 289, + 482, + 410 + ], + "type": "text", + "content": ", and the generation process is:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 253, + 420, + 480, + 447 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 420, + 480, + 447 + ], + "spans": [ + { + "bbox": [ + 253, + 420, + 480, + 447 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {C} = f _ {\\text {m a s k} \\rightarrow \\text {c o l o r}} (\\mathbf {M}), \\\\ \\mathbf {I} _ {\\text {s y n}} = \\operatorname {M a s k 2 I m g} (\\mathbf {T}, \\mathbf {C}), \\tag {2} \\\\ \\end{array}", + "image_path": "51afd7f568b0c61e7d3ec88a80008cbb9f19888775e64d49d77e5697cd6b6e17.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 457, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 457, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 457, + 482, + 567 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 457, + 482, + 567 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{mask}\\rightarrow \\mathrm{color}}:\\mathbb{W}^{H\\times W\\times N}\\to \\mathbb{W}^{H\\times W\\times 3}" + }, + { + "bbox": [ + 130, + 457, + 482, + 567 + ], + "type": "text", + "content": " is the function to convert the segmentation masks into a color map. For semantic segmentation, the value of each pixel on the segmentation mask corresponds to a category ID, allowing us to convert the masks directly into an RGB color map using a pre-defined lookup table. For panoptic and instance segmentation, after mapping the category IDs to color maps, it is essential to outline each segment with a special edge color on the color map. This ensures the model recognizes the specific instance it belongs to. The segmentation map " + }, + { + "bbox": [ + 130, + 457, + 482, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 130, + 457, + 482, + 567 + ], + "type": "text", + "content": " can be human-annotated or synthetic (i.e., " + }, + { + "bbox": [ + 130, + 457, + 482, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathrm{syn}}" + }, + { + "bbox": [ + 130, + 457, + 482, + 567 + ], + "type": "text", + "content": " from Text2Mask Model as shown in Eq. 1)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 585, + 246, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 585, + 246, + 596 + ], + "spans": [ + { + "bbox": [ + 131, + 585, + 246, + 596 + ], + "type": "text", + "content": "3.3 Data Generation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "With the aforementioned generative models, SegGen proposes two approaches for synthesizing new segmentation training samples: MaskSyn and ImgSyn. MaskSyn focuses on enhancing the diversity of synthetic segmentation masks, whereas ImgSyn concentrates on diversifying synthetic images. These generation approaches are illustrated on the right of Fig. 3." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 415, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 415, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 415, + 91, + 447, + 102 + ], + "type": "text", + "content": "SegGen" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 191 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 191 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 191 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 191 + ], + "type": "image", + "image_path": "3c600986a67bc54a0876c08e61aed94f78788a7b34d9b9135ab8376737e72cc6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 191, + 482, + 225 + ], + "lines": [ + { + "bbox": [ + 132, + 191, + 482, + 225 + ], + "spans": [ + { + "bbox": [ + 132, + 191, + 482, + 225 + ], + "type": "text", + "content": "Fig. 5: Generated samples by ImgSyn on ADE20K: The generated images exhibit remarkable realism and align well with the human-labeled masks and text prompts (see more samples in supplementary materials)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 232, + 481, + 320 + ], + "blocks": [ + { + "bbox": [ + 133, + 232, + 481, + 320 + ], + "lines": [ + { + "bbox": [ + 133, + 232, + 481, + 320 + ], + "spans": [ + { + "bbox": [ + 133, + 232, + 481, + 320 + ], + "type": "image", + "image_path": "200a30a9c1257688104fa7d6aeaad1e3dba5099ac0f91d4f95da3e7689242c02.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 321, + 482, + 365 + ], + "lines": [ + { + "bbox": [ + 132, + 321, + 482, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 321, + 482, + 365 + ], + "type": "text", + "content": "Fig. 6: Zoom-in comparison of real images and our ImgSyn images: As highlighted in the circles, our synthetic images align better with the human-labeled masks in many cases because of the inaccuracies in annotations. The left 4 columns are from ADE20K and the right 4 from COCO." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 375, + 482, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 375, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 375, + 482, + 495 + ], + "type": "text", + "content": "MaskSyn MaskSyn starts with a real training sample pair [image, segmentation masks] from a human-annotated segmentation dataset. It first extracts the caption of the image with the image captioner model. The caption serves as a text prompt and is used to generate a set of diverse new segmentation masks with the Text2Mask model following Eq. 1. Subsequently, the new segmentation masks and the corresponding text prompt are fed into the Mask2Img model to generate a new image that aligns well with its mask. As such, each training sample crafted by MaskSyn includes both a novel segmentation mask and a new image. MaskSyn effectively increases the data diversity in segmentation masks to a great extent. Some generated samples are shown in Fig. 4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 498, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 483, + 666 + ], + "type": "text", + "content": "ImgSyn Different from MaskSyn, ImgSyn focuses on increasing the data diversity of images based on human-annotated segmentation masks. For each real training sample pair [image, segmentation mask], it takes the human-annotated segmentation mask and the text prompt extracted from the image as input, and generates a set of varied images that align well with the human-annotated mask. In this way, new training samples consisting of human-labeled masks and new synthetic images are generated. ImgSyn can also be viewed as a kind of data augmentation method that enhances the data diversity on the image side. Our experiments reveal a remarkably high alignment between the synthetic images and their respective segmentation masks. Some synthesized results are showcased in Fig. 5. Our machine-generated synthetic images achieve a better mask-image alignment than real images in many cases, as shown in Fig. 6. This phenomenon occurs because human annotations tend to be imperfect due to the high difficulty of annotating segmentation masks." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "type": "text", + "content": "Hanrong Ye et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 142, + 106, + 471, + 225 + ], + "blocks": [ + { + "bbox": [ + 142, + 106, + 471, + 225 + ], + "lines": [ + { + "bbox": [ + 142, + 106, + 471, + 225 + ], + "spans": [ + { + "bbox": [ + 142, + 106, + 471, + 225 + ], + "type": "table", + "html": "
MethodVenueBackboneCrop SizeIterationsmIoU (s.s.)mIoU (m.s.)
MaskFormer [8]NeurIPS 2021R50512160k44.546.7
Mask DINO [27]CVPR 2023R50512160k48.7-
OneFormer [19]CVPR 2023R50512160k47.3-
Mask2Former [7]CVPR 2022R50512160k47.249.2
—w/ SegGen (ours)-R50512160k49.9 (+2.7)51.4 (+2.2)
MaskFormer [8]NeurIPS 2021Swin-L640160k54.155.6
Mask DINO [27]CVPR 2023Swin-L640160k56.6-
OneFormer [19]CVPR 2023Swin-L640160k57.057.7
Mask2Former [7]CVPR 2022Swin-L640160k56.157.3
—w/ SegGen (ours)-Swin-L640160k57.4 (+1.3)58.7 (+1.4)
", + "image_path": "9803c9b23853486064edb37743d11a1cb0f86c60c04ebe75624fe3c378b5461f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 226, + 480, + 280 + ], + "lines": [ + { + "bbox": [ + 130, + 226, + 480, + 280 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 480, + 280 + ], + "type": "text", + "content": "Table 1: Semantic segmentation on ADE20K val (150 categories): Synthetic data is used as a data augmentation. With the same model structures and training recipes, our SegGen boosts the performance of Mask2Former by a large margin and establishes a new SOTA performance without extra real data under both single-scale (s.s.) and multi-scale (m.s.) test settings." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 131, + 288, + 420, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 288, + 420, + 300 + ], + "spans": [ + { + "bbox": [ + 131, + 288, + 420, + 300 + ], + "type": "text", + "content": "3.4 Training Segmentation Models with Synthetic Data" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 305, + 480, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 305, + 480, + 353 + ], + "spans": [ + { + "bbox": [ + 130, + 305, + 480, + 353 + ], + "type": "text", + "content": "The final goal of this work is to improve the performance of current segmentation models with the synthetically generated training data. Therefore, we use both synthetic data produced by SegGen and training data from existing datasets in the training process of segmentation models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 354, + 481, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 481, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 481, + 461 + ], + "type": "text", + "content": "We empirically investigate two training strategies utilizing the synthetic data: (i) Synthetic data augmentation strategy. The synthetic data is used for random data augmentation. In every iteration in the training process, each real training sample is replaced by the synthetic training sample with a probability " + }, + { + "bbox": [ + 130, + 354, + 481, + 461 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{aug}}" + }, + { + "bbox": [ + 130, + 354, + 481, + 461 + ], + "type": "text", + "content": ". (ii) Synthetic data pre-training strategy. It comprises two training stages: pretraining and fine-tuning. In the pre-training stage, we pre-train the segmentation models on synthetic data, so that they learn good weights that are transferable and favorable for fine-tuning. In the subsequent fine-tuning stage, the segmentation models are trained with solely human-annotated data." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 478, + 230, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 478, + 230, + 491 + ], + "spans": [ + { + "bbox": [ + 131, + 478, + 230, + 491 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 500, + 482, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 500, + 482, + 608 + ], + "spans": [ + { + "bbox": [ + 130, + 500, + 482, + 608 + ], + "type": "text", + "content": "To accurately evaluate the effectiveness of our data generation method in improving segmentation performance, we adopt mainstream segmentation models and commonly used evaluation benchmarks for several typical segmentation tasks. The experiments are conducted mostly under the fully-supervised learning setting, meaning all human-annotated training samples from the evaluated datasets are used alongside our synthetic data. To guarantee an unbiased assessment of the impact of our synthetic data, we keep the architectures of the segmentation models and training protocols consistent with their respective original implementations throughout our studies." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 624, + 279, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 624, + 279, + 635 + ], + "spans": [ + { + "bbox": [ + 131, + 624, + 279, + 635 + ], + "type": "text", + "content": "4.1 Implementation Details" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Segmentation Datasets and Evaluation We conduct experiments on three image segmentation benchmarks following the main experimental settings of" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 415, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 415, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 415, + 91, + 447, + 102 + ], + "type": "text", + "content": "SegGen" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 212 + ], + "type": "text", + "content": "Mask2Former [7]: ADE20K semantic segmentation [60], COCO panoptic segmentation, and COCO instance segmentation [33]. Our evaluation uses all 150 classes for ADE20K and 133 classes for COCO. We use all the images from the training splits in the training of segmentation models. For semantic segmentation, we show the mean Intersection-over-Union metric (mIoU). For instance segmentation, the average precision (AP) is used. For panoptic segmentation, we report panoptic quality (PQ), \"thing\" instance segmentation " + }, + { + "bbox": [ + 130, + 115, + 482, + 212 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{\\mathrm{pan}}^{\\mathrm{Th}}" + }, + { + "bbox": [ + 130, + 115, + 482, + 212 + ], + "type": "text", + "content": ", and semantic segmentation " + }, + { + "bbox": [ + 130, + 115, + 482, + 212 + ], + "type": "inline_equation", + "content": "\\mathrm{mIoU}_{\\mathrm{pan}}" + }, + { + "bbox": [ + 130, + 115, + 482, + 212 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 212, + 482, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 212, + 482, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 482, + 307 + ], + "type": "text", + "content": "Segmentation Models We adopt Mask2Former [7], a recently-proposed prevalent transformer model, as the default segmentation model. Two typical backbones, i.e., R50 [16] and Swin-L [34], are studied. We keep the official implementation and training hyper-parameters of the segmentation models unchanged. For more training details please refer to their paper. We also conduct experiments on Mask DINO [27], a detection-aided segmentation model, and HRNet W48 [46], a representative fully-convolutional model. More implementation details are introduced in the supplementary materials." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 308, + 482, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 482, + 498 + ], + "type": "text", + "content": "Data Generation The Text2Mask and Mask2Img models are both based on the SDXL-base model [39]. They are trained on the training splits of the target segmentation datasets (i.e. ADE20K or COCO) separately for 30,000 iterations using a learning rate of " + }, + { + "bbox": [ + 130, + 308, + 482, + 498 + ], + "type": "inline_equation", + "content": "10^{-5}" + }, + { + "bbox": [ + 130, + 308, + 482, + 498 + ], + "type": "text", + "content": ". AdamW optimizer is employed, and the models are trained at a resolution of 768. Pre-trained weights from SDXL-base [39] are utilized. Random flipping data augmentation is used. During sampling, the default EDM sampler [22] is used. The sampling steps are 200 in the Text2Mask model and 40 in the Mask2Img model. During data sampling, for each training sample in the ADE20K semantic segmentation dataset, we produce 10 synthetic mask-image pairs using MaskSyn, resulting in 202,100 training samples. Additionally, we synthesize 50 images based on each human-labeled mask with ImgSyn, leading to a total of 1,010,500 samples. On COCO, we solely use ImgSyn for data synthesis. By generating 10 synthetic images conditioned on each human-labeled panoptic segmentation mask via ImgSyn, our synthetic set amounts to 1,182,870 synthetic samples, which are used in the training of both panoptic and instance segmentation models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 515, + 228, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 515, + 228, + 526 + ], + "spans": [ + { + "bbox": [ + 132, + 515, + 228, + 526 + ], + "type": "text", + "content": "4.2 Main Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 533, + 483, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 483, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 483, + 640 + ], + "type": "text", + "content": "ADE20K Semantic Segmentation We employ the synthetic data augmentation strategy with " + }, + { + "bbox": [ + 130, + 533, + 483, + 640 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{aug}} = 60\\%" + }, + { + "bbox": [ + 130, + 533, + 483, + 640 + ], + "type": "text", + "content": " for Mask2Former, and show the results in Table 1. SegGen significantly boosts the mIoU of Mask2Former R50 by +2.7 for single-scale inference and +2.2 for multi-scale inference, achieving 49.9 and 51.4 correspondingly. The Swin-L variant is also largely improved from 56.1/57.3 (single-scale/multi-scale) to 57.4/58.7 (+1.3/+1.4). Notably, our data synthesis method helps Mask2Former surpass the newer methods such as Mask DINO and One-Former [19], while establishing new SOTA results for R50 and Swin-L settings without using additional human-annotated data." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "COCO Panoptic Segmentation On COCO we adopt the synthetic data pretraining strategy to utilize our synthetic data. Specifically, we pre-train the mod-" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 239, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 239, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 239, + 102 + ], + "type": "text", + "content": "Hanrong Ye et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 103, + 479, + 252 + ], + "blocks": [ + { + "bbox": [ + 134, + 103, + 479, + 252 + ], + "lines": [ + { + "bbox": [ + 134, + 103, + 479, + 252 + ], + "spans": [ + { + "bbox": [ + 134, + 103, + 479, + 252 + ], + "type": "table", + "html": "
MethodBackbone QueriesEpochsPQ\\( PQ^{Th} \\)\\( PQ^{St} \\)\\( AP_{pan}^{Th} \\)\\( mIoU_{pan} \\)
DETR [3]R50100500+2543.448.236.331.1-
MaskFormer [8]R5010030046.551.039.833.057.8
Mask2Former [7]R501005051.957.743.041.761.7
Mask DINO [27]R503005053.059.143.943.3-
Mask2Former [7]R5010050+5052.057.943.442.061.0
— w/SegGen (ours)R5010050+5052.7 (+0.7)58.8 (+0.9)43.6 (+0.2)43.1 (+1.1)62.6 (+1.6)
Mask DINO [27]R5030050+5053.459.344.444.260.5
— w/SegGen (ours)R5030050+5054.0 (+0.6)60.2 (+0.9)44.7 (+0.3)45.4 (+1.2)61.5 (+1.0)
MaskFormer [8]Swin-L10030052.758.544.040.164.8
OneFormer [19]Swin-L15010057.964.448.049.067.4
Mask2Former [7]Swin-L20010057.864.248.148.667.4
Mask DINOSwin-L3005058.365.148.050.6-
Mask2Former [7]Swin-L200100+10057.364.346.848.066.2
— w/SegGen (ours)Swin-L200100+10058.0 (+0.7)64.5 (+0.2)48.1 (+1.3)48.8 (+0.8)67.4 (+1.2)
Mask DINO [27]Swin-L30050+5058.665.448.350.467.0
— w/SegGen (ours)Swin-L30050+5059.3 (+0.7)65.9 (+0.5)49.3 (+1.0)51.1 (+0.7)68.1 (+1.1)
", + "image_path": "3b3794b082d636e53e627d00a44c617533c9be569a29cdb280b3b085a3cc6fb2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 132, + 252, + 479, + 295 + ], + "lines": [ + { + "bbox": [ + 132, + 252, + 479, + 295 + ], + "spans": [ + { + "bbox": [ + 132, + 252, + 479, + 295 + ], + "type": "text", + "content": "Table 2: Panoptic segmentation on COCO panoptic val2017 (133 categories): Synthetic data is used for pre-training. Our SegGen significantly improves the performance compared with real data pre-training baselines, and establishes new SOTA performance without extra real data." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 300, + 230, + 360 + ], + "blocks": [ + { + "bbox": [ + 133, + 300, + 230, + 360 + ], + "lines": [ + { + "bbox": [ + 133, + 300, + 230, + 360 + ], + "spans": [ + { + "bbox": [ + 133, + 300, + 230, + 360 + ], + "type": "image", + "image_path": "40c149afd602b58615e14563ec5f4c51e77302485dcc7cec11d7eae6f21cb13d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 361, + 231, + 406 + ], + "lines": [ + { + "bbox": [ + 132, + 361, + 231, + 406 + ], + "spans": [ + { + "bbox": [ + 132, + 361, + 231, + 406 + ], + "type": "text", + "content": "Fig.7: ADE20K mIoU of Mask2Former R50 with different augmentation probability " + }, + { + "bbox": [ + 132, + 361, + 231, + 406 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{aug}}" + }, + { + "bbox": [ + 132, + 361, + 231, + 406 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 235, + 300, + 332, + 360 + ], + "blocks": [ + { + "bbox": [ + 235, + 300, + 332, + 360 + ], + "lines": [ + { + "bbox": [ + 235, + 300, + 332, + 360 + ], + "spans": [ + { + "bbox": [ + 235, + 300, + 332, + 360 + ], + "type": "image", + "image_path": "f3f8b83eb7106c77ba68509af2cf5a40779e37be6b68068d8485d7691bc073c2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 234, + 361, + 334, + 405 + ], + "lines": [ + { + "bbox": [ + 234, + 361, + 334, + 405 + ], + "spans": [ + { + "bbox": [ + 234, + 361, + 334, + 405 + ], + "type": "text", + "content": "Fig.8: ADE20K mIoU of Mask2Former R50 using different numbers of synthetic samples." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 339, + 300, + 476, + 357 + ], + "blocks": [ + { + "bbox": [ + 339, + 300, + 476, + 357 + ], + "lines": [ + { + "bbox": [ + 339, + 300, + 476, + 357 + ], + "spans": [ + { + "bbox": [ + 339, + 300, + 476, + 357 + ], + "type": "table", + "html": "
MethodmIoU*
Pure Real Data83.3
DiffuMask (ICCV 2023)57.0
MaskSyn+ImgSyn (ours)82.2(+25.2)
MaskSyn (ours)73.2
ImgSyn (ours)80.0
", + "image_path": "898e04b33595341fd958a6d25c84b442a4cea32b94e13090245d04d42e98ff92.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 338, + 357, + 478, + 401 + ], + "lines": [ + { + "bbox": [ + 338, + 357, + 478, + 401 + ], + "spans": [ + { + "bbox": [ + 338, + 357, + 478, + 401 + ], + "type": "text", + "content": "Table 4: Comparison with DiffuMask [48] on ADE20K: All models trained purely on synthetic data. mIoU* is defined in [48]." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 412, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 412, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 412, + 482, + 555 + ], + "type": "text", + "content": "els purely on our synthetic data and fine-tune the models on real data. The results are demonstrated in Table 2. For a fair comparison, we compare with models pretrained on purely real data with the same training settings. Compared with the real data pre-trained baselines, SegGen consistently brings performance gains on all metrics with both R50 and Swin-L backbones, achieving new state-of-the-art performance without extra real data. Notably, for Mask DINO with Swin-L backbone, PQ is improved by " + }, + { + "bbox": [ + 130, + 412, + 482, + 555 + ], + "type": "inline_equation", + "content": "+0.7" + }, + { + "bbox": [ + 130, + 412, + 482, + 555 + ], + "type": "text", + "content": ", \"thing\" instance segmentation " + }, + { + "bbox": [ + 130, + 412, + 482, + 555 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{\\mathrm{pan}}^{\\mathrm{Th}}" + }, + { + "bbox": [ + 130, + 412, + 482, + 555 + ], + "type": "text", + "content": " is increased by " + }, + { + "bbox": [ + 130, + 412, + 482, + 555 + ], + "type": "inline_equation", + "content": "+0.7" + }, + { + "bbox": [ + 130, + 412, + 482, + 555 + ], + "type": "text", + "content": ", and semantic segmentation " + }, + { + "bbox": [ + 130, + 412, + 482, + 555 + ], + "type": "inline_equation", + "content": "\\mathrm{mIoU}_{\\mathrm{pan}}" + }, + { + "bbox": [ + 130, + 412, + 482, + 555 + ], + "type": "text", + "content": " is boosted by " + }, + { + "bbox": [ + 130, + 412, + 482, + 555 + ], + "type": "inline_equation", + "content": "+1.1" + }, + { + "bbox": [ + 130, + 412, + 482, + 555 + ], + "type": "text", + "content": ". These noteworthy findings demonstrate that training with our synthetic data systematically improves performance across various segmentation tasks. We observe similar significant improvement on Mask2Former, further showing the effectiveness of our synthetic data on different models." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "COCO Instance Segmentation The performance on instance segmentation, both before and after incorporating our SegGen synthetic data, is detailed in Table 3. Compared to the baselines pre-trained on real data, our technique is superior across all metrics. Specifically, there is a solid " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "+0.5" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " improvement in AP with the R50 backbone, and a " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "+0.8" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " increase when using the Swin-L backbone. Additional assessment on the LVIS [14]-annotated COCO validation set, whose annotations are more accurate than the original version, provides additional evidence of the effectiveness of SegGen, as it enhances the average precision (AP) by " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "+1.2" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " when used with the larger Swin-L backbone." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "type": "text", + "content": "SegGen" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 114, + 479, + 289 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 479, + 289 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 479, + 289 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 479, + 289 + ], + "type": "table", + "html": "
MethodBackbone QueriesEpochsAP\\( AP^S \\)\\( AP^M \\)\\( AP^L \\)
Mask R-CNN [15]R50anchors40042.523.845.060.0
HTC [4]R50anchors3639.722.642.250.6
QueryInst [11]R503003640.623.442.552.8
MaskFormer [8]R5010030034.016.437.854.2
Mask2Former [7]R501005043.723.447.264.8
Mask2Former [7]R501005047.126.655.472.7
Mask2Former [7]R5010050+5044.123.447.766.1
— w/ SegGen (ours)R5010050+5044.6 (+0.5)24.0 (+0.6)48.2 (+0.5)66.2 (+0.1)
Mask2Former [7]R5010050+5047.927.856.373.5
— w/ SegGen (ours)R5010050+5048.4 (+0.5)28.3 (+0.6)57.4 (+1.1)74.1 (+0.6)
Swin-HTC++ [34]Swin-Lanchors7249.531.052.467.2
QueryInst [11]Swin-L3005048.930.852.668.3
Oneformer [19]Swin-L15010048.9---
Mask2Former [7]Swin-L20010050.129.953.972.1
Mask2Former [7]Swin-L20010054.335.763.379.1
Mask2Former [7]Swin-L200100+10049.529.253.870.5
— w/ SegGen (ours)Swin-L200100+10050.3 (+0.8)31.2 (+2.0)54.3 (+0.5)72.2 (+1.7)
Mask2Former [7]Swin-L200100+10053.635.262.677.8
— w/ SegGen (ours)Swin-L200100+10054.8 (+1.2)35.9 (+0.8)64.2 (+1.6)79.2 (+1.5)
", + "image_path": "e515cf0a7239893925791bd18643546f4c2685abff5677f9f1d77b606abce828.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 130, + 289, + 482, + 344 + ], + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 344 + ], + "type": "text", + "content": "Table 3: Instance segmentation on COCO val2017 (80 categories): Synthetic data is used for pre-training. The gray rows correspond to the evaluations conducted on LVIS [14] which offers instance masks of higher quality. Our SegGen significantly improves the performance of instance segmentation for various sizes of objects compared with pre-training on real data." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 350, + 482, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 482, + 445 + ], + "type": "text", + "content": "Comparison with Segmentation Data Generation Method In Table 4, we compare various versions of our SegGen with DiffuMask [48], which is a recently-proposed (ICCV2023) segmentation data generation approach. All methods employ the Mask2Former R50 model and are purely trained on synthetic samples. We adhere to the evaluation setting designed by DiffuMask where the IoU for three common classes are examined. Our method markedly surpasses DiffuMask by " + }, + { + "bbox": [ + 130, + 350, + 482, + 445 + ], + "type": "inline_equation", + "content": "+25.2" + }, + { + "bbox": [ + 130, + 350, + 482, + 445 + ], + "type": "text", + "content": " mIoU and demonstrates performance comparable to training purely with real data, highlighting the unprecedented quality of our synthetic data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 447, + 482, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 447, + 482, + 543 + ], + "spans": [ + { + "bbox": [ + 130, + 447, + 482, + 543 + ], + "type": "text", + "content": "Generalization Ability on Unseen Domains We visualize the segmentation results of Mask2Former R50 on images from PASCAL val dataset [10] in Fig. 1 and Fig. 9, comparing models trained both with and without our synthetic data. The models are trained on COCO or ADE20K. When trained using our extensively varied synthetic data, the segmentation model demonstrates notably improved performance on the unseen domain. We further study the segmentation results on AI-generated images, which are synthesized using different image generation models, in the supplementary materials." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 544, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 482, + 616 + ], + "type": "text", + "content": "Visual Analysis of Generated Samples by MaskSyn We randomly select some samples generated by MaskSyn in Fig. 4 (more in supplementary materials). The generative models are trained on ADE20K. Given the text prompts, our Text2Mask model can generate diverse segmentation masks, and the Mask2Img model produces realistic synthetic images with a good alignment with the masks and text prompts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Visual Analysis of Generated Samples by ImgSyn We showcase the outputs by ImgSyn in Fig. 5 (more in supplementary materials). The generated images are in good agreement with the text prompts and human-labeled segmentation masks. Furthermore, we compare the mask-image alignment of the" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 239, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 239, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 239, + 102 + ], + "type": "text", + "content": "Hanrong Ye et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 221 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 221 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 221 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 221 + ], + "type": "image", + "image_path": "4fc732a14710873ab1877fa725b0cfb655db2a34ddf98116257f16a1079b58a4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 221, + 479, + 253 + ], + "lines": [ + { + "bbox": [ + 132, + 221, + 479, + 253 + ], + "spans": [ + { + "bbox": [ + 132, + 221, + 479, + 253 + ], + "type": "text", + "content": "Fig. 9: Generalization Ability on unseen domains: Segmentation outputs on unseen domain (images from PASCAL [10]). The models are trained on ADE20K. Training with our SegGen demonstrates enhanced robustness towards unfamiliar domains." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 133, + 255, + 477, + 327 + ], + "blocks": [ + { + "bbox": [ + 133, + 255, + 477, + 327 + ], + "lines": [ + { + "bbox": [ + 133, + 255, + 477, + 327 + ], + "spans": [ + { + "bbox": [ + 133, + 255, + 477, + 327 + ], + "type": "table", + "html": "
MethodIterations Real SamplesmIoUMethodIterations Real SamplesmIoU
Baseline160K2021047.2Baseline100K2021044.4
- w/ SegGen MaskSyn160K2021048.5- w/ SegGen MaskSyn100K2021044.8
- w/ SegGen ImgSyn160K2021049.3- w/ SegGen ImgSyn100K2021045.4
- w/ SegGen ImgSyn + MaskSyn160K2021049.9- w/ SegGen ImgSyn + MaskSyn100K2021045.6
Baseline160K100023.7Baseline100K100023.6
- w/ SegGen MaskSyn160K100026.5- w/ SegGen MaskSyn100K100026.0
- w/ SegGen ImgSyn160K100024.1- w/ SegGen ImgSyn100K100023.9
- w/ SegGen ImgSyn + MaskSyn160K100027.1- w/ SegGen ImgSyn + MaskSyn100K100026.9
", + "image_path": "4ced0171a20b9f2f477d4712d991c9dd9a7255dd1e66fdfbe01891e4199b35eb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 328, + 479, + 361 + ], + "lines": [ + { + "bbox": [ + 132, + 328, + 479, + 361 + ], + "spans": [ + { + "bbox": [ + 132, + 328, + 479, + 361 + ], + "type": "text", + "content": "Table 5: Ablation Study of MaskSyn and ImgSyn on ADE20K: Both MaskSyn and ImgSyn notably enhance the performance of Mask2Former and HRNet, while MaskSyn has a more pronounced impact when there are fewer real samples available." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 132, + 369, + 479, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 479, + 404 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 479, + 404 + ], + "type": "text", + "content": "real training samples and our synthetic training samples in Fig. 6. Our synthetic samples exhibit superior alignment quality in many cases, compared to human annotations which are often imperfect on the boundaries." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 426, + 269, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 426, + 269, + 438 + ], + "spans": [ + { + "bbox": [ + 132, + 426, + 269, + 438 + ], + "type": "text", + "content": "4.3 Quantitative Analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 449, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 628 + ], + "type": "text", + "content": "Effectiveness of MaskSyn and ImgSyn We evaluate the data generated by MaskSyn and ImgSyn separately and jointly, by examining their impacts on the performance of segmentation models, as shown in Table 5. The experiments are conducted on ADE20K with Mask2Former R50 and HRNet W48. We find that the training samples generated by both MaskSyn and ImgSyn bring significant performance gains compared with the baseline. When combining these two types of synthetic data together, the segmentation model can achieve the best performance. We delve deeper into a scenario where only 1,000 real training samples are available in the training of both generative models and segmentation models. It is observed that MaskSyn substantially enhances the segmentation results in situations with limited real data. This could be attributed to the ability of MaskSyn to amplify data diversity by creating entirely new segmentation masks and images. Moreover, the significant performance improvements achieved by our SegGen, trained on merely 1,000 real samples, demonstrate the data efficiency and robustness of our proposal." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 485, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 485, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 485, + 666 + ], + "type": "text", + "content": "Influence of Synthetic Data Augmentation Probability " + }, + { + "bbox": [ + 130, + 629, + 485, + 666 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{aug}}" + }, + { + "bbox": [ + 130, + 629, + 485, + 666 + ], + "type": "text", + "content": " on ADE20K We investigate the effect of " + }, + { + "bbox": [ + 130, + 629, + 485, + 666 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{aug}}" + }, + { + "bbox": [ + 130, + 629, + 485, + 666 + ], + "type": "text", + "content": " in Fig. 7. When " + }, + { + "bbox": [ + 130, + 629, + 485, + 666 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{aug}} = 0" + }, + { + "bbox": [ + 130, + 629, + 485, + 666 + ], + "type": "text", + "content": ", it implies training using only real data, whereas " + }, + { + "bbox": [ + 130, + 629, + 485, + 666 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{aug}} = 100\\%" + }, + { + "bbox": [ + 130, + 629, + 485, + 666 + ], + "type": "text", + "content": " means training entirely with synthetic" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 415, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 415, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 415, + 91, + 447, + 102 + ], + "type": "text", + "content": "SegGen" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 140, + 114, + 297, + 146 + ], + "blocks": [ + { + "bbox": [ + 140, + 114, + 297, + 146 + ], + "lines": [ + { + "bbox": [ + 140, + 114, + 297, + 146 + ], + "spans": [ + { + "bbox": [ + 140, + 114, + 297, + 146 + ], + "type": "table", + "html": "
MethodIterationsmIoU
Syn. Pre-Train160K+160K47.2
Syn. Data Aug160K49.9
", + "image_path": "1e15dcfbd1b7f0031c4fe4d0fce82b13a40dd934c2771ba5fcf21e302b3319da.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 132, + 147, + 304, + 172 + ], + "lines": [ + { + "bbox": [ + 132, + 147, + 304, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 147, + 304, + 172 + ], + "type": "text", + "content": "(a) ADE20K Semantic Segmentation: Using synthetic data as random data augmentation obtains the best performance on ADE20K." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 310, + 114, + 472, + 145 + ], + "blocks": [ + { + "bbox": [ + 310, + 114, + 472, + 145 + ], + "lines": [ + { + "bbox": [ + 310, + 114, + 472, + 145 + ], + "spans": [ + { + "bbox": [ + 310, + 114, + 472, + 145 + ], + "type": "table", + "html": "
MethodEpochsPQAP
Syn. Pre-Training50+5052.744.6
Syn. Data Aug5051.343.2
", + "image_path": "1474d5246ce5edd087c72eb9a827a2a186e5d75bfb8b975d41d981a734d0cb2e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 306, + 147, + 478, + 179 + ], + "lines": [ + { + "bbox": [ + 306, + 147, + 478, + 179 + ], + "spans": [ + { + "bbox": [ + 306, + 147, + 478, + 179 + ], + "type": "text", + "content": "(b) COCO Panoptic Segmentation and Instance Segmentation: Using synthetic data for pre-training achieves better performance on COCO dataset." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 4 + }, + { + "bbox": [ + 135, + 181, + 476, + 192 + ], + "lines": [ + { + "bbox": [ + 135, + 181, + 476, + 192 + ], + "spans": [ + { + "bbox": [ + 135, + 181, + 476, + 192 + ], + "type": "text", + "content": "Table 6: Different training strategies with synthetic data using Mask2Former R50." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 196, + 480, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 196, + 480, + 219 + ], + "spans": [ + { + "bbox": [ + 130, + 196, + 480, + 219 + ], + "type": "text", + "content": "data. The best results are achieved at " + }, + { + "bbox": [ + 130, + 196, + 480, + 219 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{aug}} = 60\\%" + }, + { + "bbox": [ + 130, + 196, + 480, + 219 + ], + "type": "text", + "content": ". It is noteworthy that training exclusively on our synthetic data results in an impressive " + }, + { + "bbox": [ + 130, + 196, + 480, + 219 + ], + "type": "inline_equation", + "content": "44.8~\\mathrm{mIoU}" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 220, + 481, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 220, + 481, + 316 + ], + "spans": [ + { + "bbox": [ + 130, + 220, + 481, + 316 + ], + "type": "text", + "content": "Influence of Synthetic Data Size To evaluate the impact of synthetic data volume, we use different numbers of synthetic training samples produced by ImgSyn in training and report the results in Fig. 8. " + }, + { + "bbox": [ + 130, + 220, + 481, + 316 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{aug}}" + }, + { + "bbox": [ + 130, + 220, + 481, + 316 + ], + "type": "text", + "content": " is fixed at " + }, + { + "bbox": [ + 130, + 220, + 481, + 316 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 130, + 220, + 481, + 316 + ], + "type": "text", + "content": ". We notice a substantial improvement in performance when increasing the synthetic data quantity from 20K to 200K, underscoring the importance of collecting a significantly larger synthetic dataset to avoid overfitting. The performance appears to plateau after 600K. More studies on the scale of synthetic data are presented in the supplementary materials." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 316, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 482, + 533 + ], + "type": "text", + "content": "Influence of Training Strategies We conduct an ablation study using different training strategies with synthetic data on ADE20K (Table 6a) and COCO (Table 6b). We find that using the synthetic data augmentation strategy achieves better performance on ADE20K semantic segmentation, While on COCO panoptic and instance segmentation, the synthetic pre-training strategy works better. We believe that due to the limited size of the ADE20K training set, which contains merely " + }, + { + "bbox": [ + 130, + 316, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\sim 20\\mathrm{k}" + }, + { + "bbox": [ + 130, + 316, + 482, + 533 + ], + "type": "text", + "content": " images, segmentation models easily overfit the scarcely-available training images and segmentation layouts. Utilizing synthetic data (including synthetic masks and images) for data augmentation can help mitigate such an overfitting problem. On the other hand, for COCO, given its larger training set size (around 100K images), overfitting is less of a concern. Hence, it is sufficient to provide the COCO models with good initial weights pre-trained on our synthetic data. Meanwhile, previous work [14] finds that the annotation accuracy of COCO is significantly more biased than ADE20K, a finding that aligns with our visualization results depicted in Fig. 6. A domain gap exists between the COCO data and our high-quality synthetic data. Hence, employing our method for pre-training and incorporating real data for fine-tuning on COCO is a more favorable approach." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 546, + 220, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 546, + 220, + 559 + ], + "spans": [ + { + "bbox": [ + 132, + 546, + 220, + 559 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "We present SegGen, a highly-effective data synthesis method for image segmentation. SegGen introduces a revolutionary segmentation data generation framework, and proposes two data generation approaches, namely MaskSyn and ImgSyn, to synthesize segmentation masks and images, with the help of the designed text-to-mask and mask-to-image generation models. Our method not only strongly enhances the model performance on semantic, panoptic, and instance segmentation benchmarks, but also substantially improves segmentation generalization ability in unseen data distribution." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 239, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 239, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 239, + 102 + ], + "type": "text", + "content": "Hanrong Ye et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 241, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 241, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 241, + 129 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 482, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 482, + 190 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 482, + 190 + ], + "type": "text", + "content": "This research is supported in part by the Early Career Scheme of the Research Grants Council (RGC) of the Hong Kong SAR under grant No. 26202321, SAIL Research Project, HKUST-Zeekr Collaborative Research Fund, HKUST-WeBank Joint Lab Project, and Tencent Rhino-Bird Focused Research Program." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 208, + 198, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 208, + 198, + 220 + ], + "spans": [ + { + "bbox": [ + 132, + 208, + 198, + 220 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 234, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 138, + 234, + 480, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 234, + 480, + 255 + ], + "spans": [ + { + "bbox": [ + 138, + 234, + 480, + 255 + ], + "type": "text", + "content": "1. Baranchuk, D., Rubachev, I., Voynov, A., Khrulkov, V., Babenko, A.: Label-efficient semantic segmentation with diffusion models. In: ICLR (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 257, + 480, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 257, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 138, + 257, + 480, + 277 + ], + "type": "text", + "content": "2. Brock, A., Donahue, J., Simonyan, K.: Large scale gan training for high fidelity natural image synthesis. In: ICLR (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 279, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 279, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 138, + 279, + 480, + 300 + ], + "type": "text", + "content": "3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: ECCV (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 301, + 480, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 301, + 480, + 332 + ], + "spans": [ + { + "bbox": [ + 138, + 301, + 480, + 332 + ], + "type": "text", + "content": "4. Chen, K., Pang, J., Wang, J., Xiong, Y., Li, X., Sun, S., Feng, W., Liu, Z., Shi, J., Ouyang, W., et al.: Hybrid task cascade for instance segmentation. In: CVPR (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 334, + 480, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 334, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 138, + 334, + 480, + 365 + ], + "type": "text", + "content": "5. Chen, L.C., Papandreou, G., Kokkinos, I., Murphy, K., Yuille, A.L.: Semantic image segmentation with deep convolutional nets and fully connected crfs. In: ICLR (2015)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 367, + 480, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 367, + 480, + 399 + ], + "spans": [ + { + "bbox": [ + 138, + 367, + 480, + 399 + ], + "type": "text", + "content": "6. Chen, T., Li, L., Saxena, S., Hinton, G., Fleet, D.J.: A generalist framework for panoptic segmentation of images and videos. arXiv preprint arXiv:2210.06366 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 400, + 480, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 400, + 480, + 422 + ], + "spans": [ + { + "bbox": [ + 138, + 400, + 480, + 422 + ], + "type": "text", + "content": "7. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: CVPR (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 423, + 480, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 423, + 480, + 443 + ], + "spans": [ + { + "bbox": [ + 138, + 423, + 480, + 443 + ], + "type": "text", + "content": "8. Cheng, B., Schwing, A.G., Kirillov, A.: Per-pixel classification is not all you need for semantic segmentation. In: NeurIPS (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 445, + 480, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 445, + 480, + 465 + ], + "spans": [ + { + "bbox": [ + 138, + 445, + 480, + 465 + ], + "type": "text", + "content": "9. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. In: NeruIPS (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 467, + 480, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 467, + 480, + 487 + ], + "spans": [ + { + "bbox": [ + 138, + 467, + 480, + 487 + ], + "type": "text", + "content": "0. Everingham, M., Eslami, S.A., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The PASCAL visual object classes challenge: A retrospective. IJCV (2015)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 489, + 480, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 489, + 480, + 510 + ], + "spans": [ + { + "bbox": [ + 138, + 489, + 480, + 510 + ], + "type": "text", + "content": "1. Fang, Y., Yang, S., Wang, X., Li, Y., Fang, C., Shan, Y., Feng, B., Liu, W.: Instances as queries. In: ICCV (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 511, + 480, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 511, + 480, + 532 + ], + "spans": [ + { + "bbox": [ + 138, + 511, + 480, + 532 + ], + "type": "text", + "content": "2. Forever, A.: Kandinsky-2. GitHub repository (2023), https://github.com/ai-forever/Kandinsky-2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 533, + 480, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 533, + 480, + 554 + ], + "spans": [ + { + "bbox": [ + 138, + 533, + 480, + 554 + ], + "type": "text", + "content": "3. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. CACM (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 555, + 480, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 480, + 576 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 480, + 576 + ], + "type": "text", + "content": "4. Gupta, A., Dollar, P., Girshick, R.: LVIS: A dataset for large vocabulary instance segmentation. In: ICCV (2019)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 577, + 464, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 577, + 464, + 588 + ], + "spans": [ + { + "bbox": [ + 138, + 577, + 464, + 588 + ], + "type": "text", + "content": "5. He, K., Gkioxari, G., Dolkar, P., Girshick, R.: Mask R-CNN. In: ICCV (2017)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 590, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 590, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 138, + 590, + 480, + 609 + ], + "type": "text", + "content": "6. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "type": "text", + "content": "7. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeruIPS (2020)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "text", + "content": "8. Huang, L., Chen, D., Liu, Y., Yujun, S., Zhao, D., Jingren, Z.: Composer: Creative and controllable image synthesis with composable conditions. arXiv preprint arxiv:2302.09778 (2023)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 415, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 415, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 415, + 91, + 447, + 102 + ], + "type": "text", + "content": "SegGen" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "19. Jain, J., Li, J., Chiu, M.T., Hassani, A., Orlov, N., Shi, H.: Oneformer: One transformer to rule universal image segmentation. In: CVPR (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "type": "text", + "content": "20. Ji, Y., Chen, Z., Xie, E., Hong, L., Liu, X., Liu, Z., Lu, T., Li, Z., Luo, P.: Ddp: Diffusion model for dense visual prediction. arXiv preprint arXiv:2303.17559 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "type": "text", + "content": "21. Kang, M., Zhu, J.Y., Zhang, R., Park, J., Shechtman, E., Paris, S., Park, T.: Scaling up gans for text-to-image synthesis. In: CVPR (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "type": "text", + "content": "22. Karras, T., Aittala, M., Aila, T., Laine, S.: Elucidating the design space of diffusion-based generative models. In: NeruIPS (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 205, + 480, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 225 + ], + "type": "text", + "content": "23. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: CVPR (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "text", + "content": "24. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "text", + "content": "25. Kirillov, A., He, K., Girshick, R., Rother, C., Dlár, P.: Panoptic segmentation. In: CVPR (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 271, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 480, + 293 + ], + "type": "text", + "content": "26. Li, D., Ling, H., Kim, S.W., Kreis, K., Fidler, S., Torralba, A.: Bigdatasetgan: Synthesizing imagenet with pixel-wise annotations. In: CVPR (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 293, + 480, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 324 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 324 + ], + "type": "text", + "content": "27. Li, F., Zhang, H., Xu, H., Liu, S., Zhang, L., Ni, L.M., Shum, H.Y.: Mask dino: Towards a unified transformer-based framework for object detection and segmentation. In: CVPR (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 326, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 346 + ], + "type": "text", + "content": "28. Li, J., Li, D., Savarese, S., Hoi, S.: BLIP-2: bootstrapping language-image pretraining with frozen image encoders and large language models. In: ICML (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 348, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 348, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 348, + 480, + 369 + ], + "type": "text", + "content": "29. Li, X., Zhong, Z., Wu, J., Yang, Y., Lin, Z., Liu, H.: Expectation-maximization attention networks for semantic segmentation. In: ICCV (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "text", + "content": "30. Li, X., Ding, H., Zhang, W., Yuan, H., Pang, J., Cheng, G., Chen, K., Liu, Z., Loy, C.C.: Transformer-based visual segmentation: A survey. arXiv preprint arXiv:2304.09854 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "type": "text", + "content": "31. Li, Y., Liu, H., Wu, Q., Mu, F., Yang, J., Gao, J., Li, C., Lee, Y.J.: Gligen: Open-set grounded text-to-image generation. In: CVPR (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 425, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 425, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 425, + 480, + 445 + ], + "type": "text", + "content": "32. Li, Z., Zhou, Q., Zhang, X., Zhang, Y., Wang, Y., Xie, W.: Open-vocabulary object segmentation with diffusion models. In: ICCV (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 447, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 480, + 468 + ], + "type": "text", + "content": "33. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólár, P., Zitnick, C.L.: Microsoft COCO: Common objects in context. In: ECCV (2014)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "type": "text", + "content": "34. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. arXiv:2103.14030 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 501, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 522 + ], + "type": "text", + "content": "35. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: CVPR (2015)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 523, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 480, + 555 + ], + "type": "text", + "content": "36. Mou, C., Wang, X., Xie, L., Wu, Y., Zhang, J., Qi, Z., Shan, Y., Qie, X.: T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453 (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 556, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 556, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 556, + 480, + 588 + ], + "type": "text", + "content": "37. Nguyen, Q.N.T.V.A.T.K.: Dataset diffusion: Diffusion-based synthetic dataset generation for pixel-level semantic segmentation. arXiv preprint arXiv:2309.14303 (2023)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 589, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 480, + 610 + ], + "type": "text", + "content": "38. PNVR, K., Singh, B., Ghosh, P., Siddiquie, B., Jacobs, D.: Ld-znet: A latent diffusion approach for text-based image segmentation. In: ICCV (2023)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "text", + "content": "39. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952 (2023)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "40. Qi, L., Kuen, J., Wang, Y., Gu, J., Zhao, H., Torr, P., Lin, Z., Jia, J.: Open world entity segmentation. TPAMI (2022)" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 239, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 239, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 239, + 102 + ], + "type": "text", + "content": "Hanrong Ye et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 657 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "text", + "content": "41. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 171 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 171 + ], + "type": "text", + "content": "42. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. In: NeruIPS (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 171, + 482, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 171, + 482, + 192 + ], + "spans": [ + { + "bbox": [ + 130, + 171, + 482, + 192 + ], + "type": "text", + "content": "43. Sauer, A., Karras, T., Laine, S., Geiger, A., Aila, T.: StyleGAN-T: Unlocking the power of GANs for fast large-scale text-to-image synthesis. In: ICML (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 192, + 482, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 192, + 482, + 214 + ], + "spans": [ + { + "bbox": [ + 130, + 192, + 482, + 214 + ], + "type": "text", + "content": "44. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: ICML (2015)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 214, + 482, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 214, + 482, + 246 + ], + "spans": [ + { + "bbox": [ + 130, + 214, + 482, + 246 + ], + "type": "text", + "content": "45. Wang, H., Cao, J., Anwer, R.M., Xie, J., Khan, F.S., Pang, Y.: Dformer: Diffusion-guided transformer for universal image segmentation. arXiv preprint arXiv:2306.03437 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 246, + 482, + 279 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 246, + 482, + 279 + ], + "spans": [ + { + "bbox": [ + 130, + 246, + 482, + 279 + ], + "type": "text", + "content": "46. Wang, J., Sun, K., Cheng, T., Jiang, B., Deng, C., Zhao, Y., Liu, D., Mu, Y., Tan, M., Wang, X., Liu, W., Xiao, B.: Deep high-resolution representation learning for visual recognition. TPAMI (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 279, + 482, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 279, + 482, + 311 + ], + "spans": [ + { + "bbox": [ + 130, + 279, + 482, + 311 + ], + "type": "text", + "content": "47. Wu, W., Zhao, Y., Chen, H., Gu, Y., Zhao, R., He, Y., Zhou, H., Shou, M.Z., Shen, C.: Datasetsd: Synthesizing data with perception annotations using diffusion models. arXiv preprint arXiv:2308.06160 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 311, + 482, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 311, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 311, + 482, + 344 + ], + "type": "text", + "content": "48. Wu, W., Zhao, Y., Shou, M.Z., Zhou, H., Shen, C.: Diffumask: Synthesizing images with pixel-level annotations for semantic segmentation using diffusion models. In: ICCV (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 344, + 482, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 344, + 482, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 344, + 482, + 376 + ], + "type": "text", + "content": "49. Xie, J., Li, W., Li, X., Liu, Z., Ong, Y.S., Loy, C.C.: Mosaic fusion: Diffusion models as data augmenters for large vocabulary instance segmentation. arXiv preprint arXiv:2309.13042 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 376, + 482, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 398 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 398 + ], + "type": "text", + "content": "50. Xu, J., Liu, S., Vahdat, A., Byeon, W., Wang, X., De Mello, S.: Open-vocabulary panoptic segmentation with text-to-image diffusion models. In: CVPR (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 398, + 482, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 398, + 482, + 419 + ], + "spans": [ + { + "bbox": [ + 130, + 398, + 482, + 419 + ], + "type": "text", + "content": "51. Yang, L., Xu, X., Kang, B., Shi, Y., Zhao, H.: Freemask: Synthetic images with dense annotations make stronger segmentation models. In: NeurIPS (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 419, + 482, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 419, + 482, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 419, + 482, + 441 + ], + "type": "text", + "content": "52. Ye, H., Xu, D.: Taskprompter: Spatial-channel multi-task prompting for dense scene understanding. In: ICLR (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 441, + 482, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 441, + 482, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 441, + 482, + 462 + ], + "type": "text", + "content": "53. Zeng, Y., Lin, Z., Zhang, J., Liu, Q., Collomosse, J., Kuen, J., Patel, V.M.: Scenecomposer: Any-level semantic image synthesis. In: CVPR (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 462, + 482, + 484 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 484 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 484 + ], + "type": "text", + "content": "54. Zhang, L., Anyi, R., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: ICCV (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 484, + 482, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 484, + 482, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 484, + 482, + 506 + ], + "type": "text", + "content": "55. Zhang, W., Pang, J., Chen, K., Loy, C.C.: K-net: Towards unified image segmentation. In: NeurIPS (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 506, + 482, + 538 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 506, + 482, + 538 + ], + "spans": [ + { + "bbox": [ + 130, + 506, + 482, + 538 + ], + "type": "text", + "content": "56. Zhang, Y., Ling, H., Gao, J., Yin, K., Lafleche, J.F., Barriuso, A., Torralba, A., Fidler, S.: Datasetgan: Efficient labeled data factory with minimal human effort. In: CVPR (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 538, + 482, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 538, + 482, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 538, + 482, + 559 + ], + "type": "text", + "content": "57. Zhao, H., Shi, J., Qi, X., Wang, X., Jia, J.: Pyramid scene parsing network. In: CVPR (2017)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 559, + 482, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 559, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 559, + 482, + 581 + ], + "type": "text", + "content": "58. Zhao, S., Chen, D., Chen, Y.C., Bao, J., Hao, S., Yuan, L., Wong, K.Y.K.: Uncontrolnet: All-in-one control to text-to-image diffusion models. NeurIPS (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 581, + 482, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 603 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 603 + ], + "type": "text", + "content": "59. Zhao, W., Rao, Y., Liu, Z., Liu, B., Zhou, J., Lu, J.: Unleashing text-to-image diffusion models for visual perception. In: ICCV (2023)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 603, + 482, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 603, + 482, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 603, + 482, + 635 + ], + "type": "text", + "content": "60. Zhou, B., Zhao, H., Puig, X., Fidler, S., Barriuso, A., Torralba, A.: Scene parsing challenge 2016. http://sceneParsing.csail.mit.edu/index_challenge.htm (2016)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 130, + 635, + 482, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 635, + 482, + 657 + ], + "spans": [ + { + "bbox": [ + 130, + 635, + 482, + 657 + ], + "type": "text", + "content": "61. Zhou, B., Zhao, H., Puig, X., Fidler, S., Barriuso, A., Torralba, A.: Scene parsing through ADE20K dataset. In: CVPR (2017)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 415, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 415, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 415, + 91, + 447, + 102 + ], + "type": "text", + "content": "SegGen" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/SegPoint_ Segment Any Point Cloud via Large Language Model/d283f64a-a017-47de-9be3-9206675b2c4f_content_list.json b/2024/SegPoint_ Segment Any Point Cloud via Large Language Model/d283f64a-a017-47de-9be3-9206675b2c4f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7a19c993e674354d5419c88f251531795dd13cf5 --- /dev/null +++ b/2024/SegPoint_ Segment Any Point Cloud via Large Language Model/d283f64a-a017-47de-9be3-9206675b2c4f_content_list.json @@ -0,0 +1,1813 @@ +[ + { + "type": "text", + "text": "SegPoint: Segment Any Point Cloud via Large Language Model", + "text_level": 1, + "bbox": [ + 302, + 154, + 759, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shuting $\\mathrm{He}^{1}$ , Henghui Ding2, Xudong Jiang1, and Bihan $\\mathrm{Wen}^{1}$", + "bbox": [ + 241, + 223, + 759, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Nanyang Technological University", + "bbox": [ + 359, + 251, + 640, + 265 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Institute of Big Data, Fudan University", + "bbox": [ + 359, + 265, + 640, + 280 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{heshuting555, henghui.ding}@gmail.com, {exdjiang, bihan.wen}@ntu.edu.sg", + "bbox": [ + 223, + 280, + 781, + 294 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://heshuting555.github.io/SegPoint", + "bbox": [ + 349, + 294, + 651, + 308 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Despite significant progress in 3D point cloud segmentation, existing methods primarily address specific tasks and depend on explicit instructions to identify targets, lacking the capability to infer and understand implicit user intentions in a unified framework. In this work, we propose a model, called SegPoint, that leverages the reasoning capabilities of a multi-modal Large Language Model (LLM) to produce point-wise segmentation masks across a diverse range of tasks: 1) 3D instruction segmentation, 2) 3D referring segmentation, 3) 3D semantic segmentation, and 4) 3D open-vocabulary semantic segmentation. To advance 3D instruction research, we introduce a new benchmark, Instruct3D, designed to evaluate segmentation performance from complex and implicit instructional texts, featuring 2,565 point cloud-instruction pairs. Our experimental results demonstrate that SegPoint achieves competitive performance on established benchmarks such as ScanRefer for referring segmentation and ScanNet for semantic segmentation, while delivering outstanding outcomes on the Instruct3D dataset. To our knowledge, SegPoint is the first model to address these varied segmentation tasks within a single framework, achieving satisfactory performance.", + "bbox": [ + 259, + 342, + 743, + 593 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Instruct3D dataset $\\cdot$ Unified framework $\\cdot$ 3D point cloud segmentation $\\cdot$ Large language model", + "bbox": [ + 259, + 604, + 740, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 679, + 375, + 695 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D point cloud segmentation, a critical challenge in the 3D vision community, aims to interpret and classify each point in a point cloud to understand its semantic properties [27, 45, 46, 54, 56, 80]. This longstanding issue has spurred significant advancements across various fields, including robotics, autonomous driving, virtual reality, etc. This challenge has evolved into a series of specialized tasks, each targeting a specific segmentation aspect. Overall, tasks cover basic semantic and instance segmentation [3, 7, 53, 74], as well as more practical tasks", + "bbox": [ + 212, + 710, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "#", + "bbox": [ + 241, + 143, + 287, + 176 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Corresponding author", + "bbox": [ + 228, + 825, + 401, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/83c779324b564b4662467dc72345b67eba7c11560d8f0f52d90769a1aef47e55.jpg", + "image_caption": [ + "Fig. 1: Example of functionality in SegPoint. SegPoint can complete various point cloud tasks in a unified framework by leveraging task-specific prompts, including 1) 3D instruction segmentation, 2) 3D referring segmentation, 3) 3D semantic segmentation, and 4) 3D open-vocabulary semantic segmentation." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 782, + 459 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "such as referring segmentation [1, 4, 24, 33, 79, 84], which segments points based on explicit textual descriptions, and open-vocabulary segmentation [12, 40, 42, 57, 65, 77] designed for the dynamic and complex nature of real-world.", + "bbox": [ + 212, + 547, + 787, + 595 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite significant progress achieved within the 3D community toward accurately segmenting objects through specifically designed models, each model is typically developed to tackle one specific segmentation task, leading to inefficiencies and a lack of versatility for real-world application. Furthermore, previous perception approaches heavily depend on predefined categories or explicit expressions for language understanding. Such approaches fall short in interpreting and acting on implicit instructions often found in human language, a critical gap that hinders the development of truly intelligent next-generation perception systems. This brings a pivotal question: Is it possible to design a unified model capable of comprehensively addressing all aforementioned 3D tasks with human-like instructions? The exploration of this question not only challenges the current paradigms of 3D point cloud segmentation but also opens the door to groundbreaking advancements in robotic perception and interaction.", + "bbox": [ + 212, + 595, + 787, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose a model called SegPoint, leveraging the Large Language Model's (LLM) advanced ability to reason and comprehend user instructions. To enhance 3D scene comprehension, we integrate a Geometric Enhancer", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "S. He et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Module that extracts local semantics from point clouds, seamlessly incorporating this geometric insight into the feature extraction process. Furthermore, a Geometric-guided Feature Propagation is designed to utilize semantic flows derived from geometric priors and the LLM's hidden embeddings, facilitating the generation of fine-grained, high-quality features for accurate dense prediction tasks. Unlike previous attempts at 2D field [8, 28, 50], we do not depend on additional costly pre-trained segmentation models like SAM [26].", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Moreover, we introduce a benchmark named Instruct3D, designed to advance research in the field of segmentation driven by implicit and complex instructions. Understanding these nuanced instructions necessitates reasoning abilities and extensive knowledge of the world. It includes a total of 2,565 diverse pairs of instructions and point clouds for tuning and evaluation. Our comprehensive experiments demonstrate the benchmark's utility in evaluating the model's capability of segmentation based on human-like instructions.", + "bbox": [ + 212, + 252, + 787, + 358 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Taking advantage of a multi-modal LLM and task-specific prompts, SegPoint is capable of generating segmentation masks for a wide range of tasks in a unified model: 1) 3D instruction segmentation, 2) 3D referring segmentation, 3) 3D semantic segmentation, and 4) 3D open-vocabulary semantic segmentation, as depicted in Fig. 1. SegPoint achieves competitive results on established benchmarks like ScanRefer [4] for referring segmentation and ScanNet [7] for semantic segmentation while showing remarkable performance on the Instruct3D dataset.", + "bbox": [ + 212, + 358, + 787, + 464 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, our main contributions are as follows:", + "bbox": [ + 238, + 465, + 607, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- We propose SegPoint, the first 3D segmentation model that can comprehend human intentions and address multiple segmentation tasks within one framework by harnessing the Large Language Model's reasoning capabilities.", + "bbox": [ + 225, + 491, + 784, + 536 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- We present a Geometric Enhancer Module that integrates comprehensive scene information into the process of 3D scene understanding. Besides, A Geometric-guided Feature Propagation is designed to achieve accurate and fine-grained segmentation. These two modules supplement the missing local information and grasp fine-grained features for dense prediction tasks.", + "bbox": [ + 225, + 537, + 784, + 612 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- We introduce a new task called 3D instruction segmentation and construct a new dataset Instruct3D, which necessitates a model's self-reasoning to interpret implicit instructions for segmenting the target object.", + "bbox": [ + 225, + 614, + 784, + 660 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- Our experimental findings reveal that SegPoint not only competes strongly in 3D semantic, referring, and open-vocabulary semantic segmentation but also excels in 3D instruction segmentation, showcasing its versatility and effectiveness across a spectrum of segmentation challenges.", + "bbox": [ + 225, + 662, + 784, + 722 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 214, + 748, + 387, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Multi-modal Large Language Model", + "text_level": 1, + "bbox": [ + 214, + 782, + 560, + 797 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inspired by the exceptional reasoning abilities of Large Language Models, researchers are delving into transferring these capabilities into the vision realm [9,", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "SegPoint", + "bbox": [ + 669, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 127 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "28, 36, 55], developing multi-modal LLMs. Flamingo [2], BLIP-2 [31], mPLUGOWL [73], Otter [30], LLaVA [37] and MiniGPT-4 [83] first construct image-text feature alignment followed by instruction tuning and achieve superior performance. Recent studies have increasingly concentrated on integrating foundational models with tasks that demand a refined understanding at the region or pixel level. VisionLLM [61] introduces a versatile interaction interface for various vision-centric tasks via instruction tuning, albeit without fully leveraging LLMs for intricate reasoning tasks. Kosmos-2 [43] has built substantial data of grounded image-text pairs, thereby embedding grounding capabilities into LLMs. DetGPT [44] links the fixed multi-modal LLM with an open-vocabulary detector, facilitating user instruction-based detection tasks. This growing interest has spurred further innovations, including GPT4RoI [78], LLaVA-grounding [76], Ferret [75], LISA [28], GlaMM [50], PixelLM [52], Sphinx [32], each contributing to the evolving landscape of instruction-based, multi-modal understanding.", + "bbox": [ + 212, + 145, + 787, + 357 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Building on advancements of multi-modal large language models in 2D image domain, the field is witnessing a seamless transition into 3D spaces. PointLLM [70] harnesses the prowess of large language models and trains with 3D point clouds following the paradigm of LLaVA [37]. 3D-LLM [20] utilizes 2D foundation models to encode multi-view images of 3D point clouds. Point-Bind [15] aligns point clouds with Image-Bind [14] and leverages ImageBind-LLM to reason multimodality input without 3D-instruction data training. GPT4Point [47] pioneers in facilitating a unified approach towards 3D object understanding and generation, setting a new standard for versatility. However, these models primarily concentrate on scene-level insights, often overlooking the intricate details at the region, or point level. Contrasting with existing models, our research focuses on two pivotal goals: 1) efficiently inject segmentation capabilities into multi-modal LLMs to conduct point-level understanding and 2) design a unified framework for 3D point cloud segmentation via the reasoning ability of LLMs.", + "bbox": [ + 212, + 358, + 789, + 569 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 3D Point Cloud Segmentation", + "text_level": 1, + "bbox": [ + 215, + 589, + 509, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3D point cloud segmentation, a crucial task in computer vision, can be categorized into 3D semantic, instance, and panoptic segmentation. 3D semantic segmentation [3,7,19,45,53,74,80] assigns each point in a 3D space to specific, predefined classes. In contrast, 3D instance segmentation [25,27,54] goes a step further by classifying each point and differentiating between distinct objects of the same class. 3D panoptic segmentation [68,82] aims to group 3D points according to their semantics and identities. Lately, more practical tasks have emerged, such as 3D referring segmentation [1,4,17,23,67,79], which extends referring expression segmentation [8,10,11,33-35] to 3D and segments a target instance based on explicit linguistic descriptions, and 3D open-vocabulary segmentation [12,40,42,57], designed to identify and segment unseen objects beyond a fixed set of known categories.", + "bbox": [ + 212, + 613, + 787, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Despite the significant progress made by transformer-based models: Mask3D [54], SPFormer [56], MAFT [29], and OneFormer3D [27] in basic 3D segmentation tasks, their application in real-world scenarios requiring human language", + "bbox": [ + 214, + 795, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "S. He et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/59e9162900bb37310cabfb6aa72651bec476a3448dd6d46a0ab4266b49ffe291.jpg", + "image_caption": [ + "Fig. 2: The pipeline of SegPoint. Given input point cloud and text query, the multimodal LLM $\\mathcal{F}$ generates text output. Geometric Enhancer Module $\\mathcal{G}$ injects geometric information into Point Encoder $\\mathcal{E}$ and obtains point features $\\hat{\\pmb{f}}_{\\text {point }}$ . Per-point embeddings $\\pmb{f}_{\\mathcal{P}}$ derived from Geometric-guided Feature Propagation $\\mathcal{P}$ multiplied with the embedding associated with the token yield the final segmentation masks." + ], + "image_footnote": [], + "bbox": [ + 222, + 148, + 782, + 324 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "interaction remains constrained. Besides, Seal [38] aims to segment any point cloud through distilling vision foundation models, while it doesn't use language as cues. TGNN [22] is the first work to tackle referring segmentation problem that proposes aggregating textual features by considering the neighboring local structure of each instance but it heavily depends on explicit expressions or predefined categories for language understanding. Furthermore, the development of models tailored to specific segmentation tasks restricts their versatility and applicability in diverse real-world scenarios. Therefore, it is imperative to develop more intelligent interaction ways and a unified model for 3D point segmentation.", + "bbox": [ + 212, + 417, + 787, + 554 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Approach", + "text_level": 1, + "bbox": [ + 215, + 575, + 346, + 593 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Architecture Overview", + "text_level": 1, + "bbox": [ + 215, + 608, + 449, + 622 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overall architecture of SegPoint is presented in Fig. 2. SegPoint mainly comprises four parts: $i$ ) a pre-trained point encoder $\\mathcal{E}$ tailored for aligning with textual data; $ii$ ) a large language model $\\mathcal{F}$ endowed with advanced reasoning capabilities; $iii$ ) a Geometric Enhancer Module $\\mathcal{G}$ responsible for extracting geometric representation from input point clouds and infusing these priors into the point encoder; and $iv$ ) a Geometric-guided Feature Propagation $\\mathcal{P}$ which is key to achieving precise mask generation. The collaboration between the Geometric Enhancer Module and Geometric-guided Feature Propagation is crucial, as it equips LLMs with the ability to generate masks effectively in various scenarios.", + "bbox": [ + 212, + 633, + 787, + 770 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Vanilla Baseline", + "text_level": 1, + "bbox": [ + 215, + 784, + 393, + 797 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The input of the framework is the text instructions $\\pmb{i}_{\\text{txt}}$ and point cloud $\\pmb{i}_{\\text{point}} \\in \\mathbb{R}^{N \\times (3 + F)}$ . Specifically, a point cloud scene, comprising $N$ points, each includes", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "SegPoint", + "bbox": [ + 669, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3D coordinates $\\in \\mathbb{R}^3$ and an auxiliary feature vector $\\in \\mathbb{R}^F$ (e.g., color). The point cloud $i_{point}$ is fed into the point encoder $\\mathcal{E}$ , which extracts point features $f_{point} \\in \\mathbb{R}^{N_1 \\times D}$ , where $N_1 \\ll N$ , $D$ is the feature dimension. Concurrently, the text instruction $i_{txt}$ undergoes tokenization via $\\mathcal{F}_{tokenize}$ . These prepared inputs are then fed into the Large Language Model $\\mathcal{F}$ , resulting in a textual response $\\pmb{y}$ . The above process can be formulated as:", + "bbox": [ + 212, + 145, + 787, + 238 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {f} _ {\\text {p o i n t}} = \\mathcal {E} (\\boldsymbol {i} _ {\\text {p o i n t}}), \\quad \\boldsymbol {f} _ {\\text {t x t}} = \\mathcal {F} _ {\\text {t o k e n i z e}} (\\boldsymbol {i} _ {\\text {t x t}}), \\quad \\boldsymbol {y} = \\mathcal {F} (\\boldsymbol {f} _ {\\text {p o i n t}}, \\boldsymbol {f} _ {\\text {t x t}}). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 269, + 251, + 785, + 268 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Building on the approach introduced by LISA [28], SegPoint enhances the segmentation capabilities of Large Language Models (LLMs) by expanding their vocabulary with a new special token, $<\\mathrm{SEG}>.$ This modification enables the model to recognize and predict the $<\\mathrm{SEG}>$ token within the output sequence as a signal to identify segmentation targets. Upon detecting a $<\\mathrm{SEG}>$ token, the corresponding output sequence belonging to $<\\mathrm{SEG}>$ token is extracted and processed through an MLP layer $\\gamma$ , generating mask embeddings $h_{seg}$ . The final step involves computing each binary mask prediction $\\pmb{m} \\in \\mathbb{R}^{N}$ by performing a dot product between the mask embeddings $h_{seg}$ and the upsampled per-point embeddings derived from the point features $\\pmb{f}_{point}$ . The formulation of the aforementioned process is given by:", + "bbox": [ + 212, + 279, + 787, + 445 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {h} _ {\\text {s e g}} = \\gamma (\\boldsymbol {y} _ {[ \\text {s e g} ]}), \\quad \\boldsymbol {m} = \\boldsymbol {h} _ {\\text {s e g}} \\otimes U p S. (\\boldsymbol {f} _ {\\text {p o i n t}}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 341, + 458, + 784, + 476 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $UpS$ denotes the upsampling operation following PointNet++ [46] on $f_{point}$ . The vanilla baseline represents an initial attempt to bridge the gap between LLMs' text comprehension and point cloud segmentation tasks. It encounters two primary issues. Firstly, the point encoder is trained on a scene-level dataset for classification achieving alignment between text and point clouds, not specifically trained for dense prediction tasks. Besides, the point encoder's first layer employs Farthest Point Sampling (FPS) [46] to reduce the point cloud to $N_1$ points, risking the loss of details vital for accurate dense predictions. Secondly, the operation of directly upsampling from $N_1$ to $N$ points to obtain per-point embeddings is prone to losing structural information and introducing a notable degree of noise, undermining the model's efficacy in segmentation tasks.", + "bbox": [ + 212, + 486, + 787, + 652 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Geometric Enhancer Module", + "text_level": 1, + "bbox": [ + 215, + 676, + 501, + 691 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To adapt the pre-trained point encoder for dense prediction tasks while maintaining its superior scene recognition capability, our objective is to harness the geometric information across the entire scene to guide the further feature learning process. Drawing inspiration from recent advancements in 2D computer vision, where studies [5,41,62,64] demonstrate that convolutions enhance transformers' ability to capture local spatial information, we introduce the Geometric Enhancer Module (GEM). This module is specifically designed to grasp the local geometric contexts within point clouds while enabling the preservation of the point encoder's foundational architecture and information integrity.", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "S. He et al.", + "bbox": [ + 271, + 114, + 346, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/084d27946842b10c23a8f76c1674d672fa1b8c092118cfb564c8430982cad1fe.jpg", + "image_caption": [ + "Fig. 3: Architecture of the proposed (b) Geometric Enhance Module (GEM) and (c) Geometric-guided Feature Propagation (GFP) interaction with (a) Point Encoder." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 782, + 352 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Fig. 3, Geometric Enhancer Module $\\mathcal{G}$ is composed of three blocks, each featuring a KPCConv [58] layer followed by BN and ReLU activation. The architecture is similar to the 2D convolutional stem [16]. We utilize KPCConv instead of vanilla convolution or linear layer here to facilitate grasping the local geometric information effectively. The resultant geometric feature, represented by $\\pmb{g}_f\\in \\mathbb{R}^{N\\times D}$ , contains the features across all points, thereby supplementing the missing local information. This $\\pmb{g}_f$ is then leveraged to infuse geometric insights into the point encoder's features via a cross-attention mechanism, the above process can be expressed as:", + "bbox": [ + 212, + 406, + 787, + 542 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {g} _ {f} = \\mathcal {G} (\\boldsymbol {i} _ {\\text {p o i n t}}), \\quad \\hat {\\boldsymbol {f}} _ {i} = \\boldsymbol {f} _ {i} + g _ {i} \\cdot \\operatorname {s o f t m a x} \\left(\\frac {\\boldsymbol {f} _ {i} \\boldsymbol {g} _ {f} ^ {T}}{\\sqrt {D}}\\right) \\boldsymbol {g} _ {f}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 308, + 547, + 784, + 589 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\pmb{f}_i$ represents the feature from the $i$ -th block of the point encoder and $l$ consecutive transformer layers are regarded as one block for the convenience of explanation. To fine-tune the integration of geometric information, we introduce a learnable gating factor $g_i$ that modulates the balance between the attention layer's output and the input feature $\\pmb{f}_i$ . This gating factor is initially set to zero, to ensure that the incorporation of geometric data does not abruptly alter $\\pmb{f}$ feature distribution. Such an approach allows for the preservation and effective utilization of the point encoder's pre-trained weights. Upon processing through the Geometric Enhancer Module (GEM), the modified output of the point encoder, LLM are formulated as:", + "bbox": [ + 212, + 595, + 787, + 744 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\pmb {f}} _ {p o i n t} = \\mathcal {E} (\\pmb {i} _ {p o i n t}, \\pmb {g} _ {f}), \\quad \\hat {\\pmb {y}} = \\mathcal {F} (\\hat {\\pmb {f}} _ {p o i n t}, \\pmb {f} _ {t x t}), \\quad \\hat {\\pmb {h}} _ {s e g} = \\gamma (\\hat {\\pmb {y}} _ {[ s e g ]}). \\qquad (4)\n$$\n", + "text_format": "latex", + "bbox": [ + 263, + 752, + 785, + 772 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Geometric-guided Feature Propagation", + "text_level": 1, + "bbox": [ + 214, + 787, + 584, + 803 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Addressing the challenge of upsampling point clouds from a sparse set of $N_{1}$ points to a denser set of $N$ points is crucial, as direct upsampling inevitably intro-", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "SegPoint", + "bbox": [ + 669, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "duces noise and results in information loss, leading to sub-optimal performance in segmentation tasks. To mitigate these issues, we introduce Geometric-guided Feature Propagation designed to generate high-quality per-point embeddings. Geometric features $\\pmb{g}_f$ , which carry comprehensive point information, serving as a \"gold message\" for enhancing the upsampling process. By integrating these geometric features, we aim to significantly improve the quality and accuracy of the generated dense per-point embeddings.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As illustrated in Fig. 3, we begin by upsampling the higher-layer features $\\pmb{f}_3, \\pmb{f}_4$ from a smaller set of points $N_1$ to larger sets $N_3, N_2$ , employing Point-Net++'s [46] propagation techniques. This step yields features $\\pmb{f}_3' \\in \\mathbb{R}^{N_3 \\times D}$ , and $\\pmb{f}_4' \\in \\mathbb{R}^{N_2 \\times D}$ . Subsequently, we perform downsampling on the geometric features $\\pmb{g}_f$ from the original number of points $N$ to reduced counts $N_2, N_3$ , respectively, utilizing the Farthest Point Sampling (FPS) technique. In this process, we directly obtain the features of the sampled points without performing additional $k$ -nearest neighbor ( $k$ -NN) and pooling operations to simplify the computation and produce features $\\pmb{f}_{g,1} \\in \\mathbb{R}^{N_3 \\times D}$ , and $\\pmb{f}_{g,2} \\in \\mathbb{R}^{N_2 \\times D}$ .", + "bbox": [ + 212, + 252, + 787, + 388 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the next phase, we integrate the up- and downsampled features, processing them through fully connected layers and ReLU activation to update the feature vectors $\\tilde{\\pmb{f}}_3\\in \\mathbb{R}^{N_3\\times D}$ , and $\\tilde{\\pmb{f}}_4\\in \\mathbb{R}^{N_2\\times D}$ . Note that the last layer feature $\\pmb{f}_{5}$ bypasses this step. Instead, we concatenate it with $\\hat{\\pmb{h}}_{point}$ from the LLM output to form $\\tilde{\\pmb{f}}_5$ to perceive multi-modal information from LLM.", + "bbox": [ + 212, + 388, + 787, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, to enable information exchange across different point densities, we propose attentive propagation. Take the propagation from $\\tilde{\\pmb{f}}_5$ to $\\tilde{\\pmb{f}}_4$ as example. Here, $\\tilde{\\pmb{f}}_4\\in \\mathbb{R}^{N_2\\times D}$ acts as a set of local centers. For each local center within $\\tilde{\\pmb{f}}_4$ , we identify its neighboring points from $\\tilde{\\pmb{f}}_5$ using the $k$ -NN algorithm, resulting in $\\pmb{f}_{54}\\in \\mathbb{R}^{N_2\\times k\\times D}$ . Then, employing cross-attention mechanism, where $\\tilde{\\pmb{f}}_4$ serves as query and $\\pmb{f}_{54}$ as both key and value, facilitates information flow across different point densities and effectively extract relevant details into the query points.", + "bbox": [ + 212, + 467, + 787, + 571 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\tilde {\\boldsymbol {f}}} _ {4} = \\tilde {\\boldsymbol {f}} _ {4} + \\operatorname {s o f t m a x} \\left(\\frac {\\tilde {\\boldsymbol {f}} _ {4} \\boldsymbol {f} _ {5 4} ^ {T}}{\\sqrt {D}}\\right) \\boldsymbol {f} _ {5 4}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 580, + 785, + 621 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Leveraging geometric-guided feature propagation enables us to produce high-quality per-point embeddings denoted as $\\pmb{f}_{\\mathcal{P}}$ , laying the foundation for generating precise segmentation masks expressed as follows:", + "bbox": [ + 212, + 628, + 785, + 674 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {f} _ {\\mathcal {P}} = \\mathcal {P} \\left(\\hat {\\boldsymbol {f}} _ {\\text {p o i n t}}, \\boldsymbol {g} _ {f}\\right), \\quad \\hat {\\boldsymbol {m}} = \\hat {\\boldsymbol {h}} _ {\\text {s e g}} \\otimes \\boldsymbol {f} _ {\\mathcal {P}}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 362, + 684, + 785, + 702 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.5 Training Objectives", + "text_level": 1, + "bbox": [ + 214, + 720, + 426, + 736 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our model is trained end-to-end leveraging the text classification loss and the segmentation mask loss:", + "bbox": [ + 212, + 743, + 785, + 773 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda_ {t x t} \\mathcal {L} _ {t x t} + \\lambda_ {b c e} \\mathcal {L} _ {b c e} + \\lambda_ {d i c e} \\mathcal {L} _ {d i c e}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 370, + 785, + 785, + 800 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{txt}$ denotes the auto-regressive cross-entropy loss targeting text generation accuracy, segmentation mask loss includes both binary cross-entropy (BCE)", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "S. He et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "loss $\\mathcal{L}_{bce}$ and DICE loss $\\mathcal{L}_{dice}$ , aims at refining segmentation quality. The weights $\\lambda_{txt}$ , $\\lambda_{bce}$ and $\\lambda_{dice}$ are utilized to balance the different loss items. The model's training is guided by the ground-truth labels $\\pmb{y}_{txt}$ for text and $\\pmb{M}$ for masks.", + "bbox": [ + 212, + 146, + 787, + 193 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.6 Instruct3D Dataset Collection", + "text_level": 1, + "bbox": [ + 214, + 213, + 511, + 227 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Although 3D instruction segmentation and 3D referring segmentation [1,4,79] are both language-based segmentation, 3D referring segmentation guides segmentation with explicit target object names, e.g., \"chair\", lacking more complicated reasoning instructions, e.g., \"Where to sit in the room?\". Besides, they also fall short in offering multi-target question-answer pairs with target descriptions directly connected to multiple segmentation masks, which cannot meet a common requirement in real-world scenarios, like \"How to play computer games\".", + "bbox": [ + 212, + 237, + 785, + 343 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To enhance the assessment and analysis of instruction segmentation capabilities, we have developed a benchmark, referred to Instruct3D. This benchmark incorporates 280 scenes specifically selected for instruction segmentation tuning and evaluation, sourced from the recently introduced ScanNet++ [74] dataset. Each scene comes with approximately 10 different segmentation instructions, resulting in 2,565 instruction-point cloud pairings. This dataset is then divided into two splits: train, and val, containing 2,052, and 513 question-answer pairs, respectively. Our dataset is uniquely designed to encompass both multi-target and zero-target scenarios, addressing the real-world requirement of identifying multiple objects in response to text queries and accounting for situations where objects mentioned in the text may not be present in the paired point cloud. Besides, we take into account the characteristics of 3D scenes and incorporate diverse locations and view descriptions e.g., \"something that is used for sitting while working at a desk. It is the one facing the window.\" The model needs to have not only reasoning capabilities but also the ability to perceive views and directions in 3D scenes. These designs underscore the dataset's practical value.", + "bbox": [ + 212, + 343, + 787, + 585 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 607, + 375, + 625 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Datasets and Evaluation Metrics", + "text_level": 1, + "bbox": [ + 214, + 638, + 532, + 652 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Datasets. Our training data is composed of two types of datasets: (1) semantic segmentation dataset including ScanNet200 [53], and S3DIS [3]; (2) referring segmentation dataset consisting of ScanRefer [4], ReferIt3D [1](including Sr3D and Nr3D), and Multi3DRefer [79]. We design task-specific prompts to facilitate the joint training of various tasks within a unified framework.", + "bbox": [ + 212, + 657, + 785, + 734 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Semantic Segmentation Dataset. We use two strategies to generate templates. 1) segment the specific category: \"USER: Can you segment the {category} category in this point cloud? ASSISTANT: {category} ,\" where category is the random chosen category, and denotes the placeholder for tokens of point cloud patches. 2) segment all the categories: \"USER: Can you segment all the semantic masks in this point cloud and output separate masks for each category in the alphabetical order of the", + "bbox": [ + 212, + 734, + 785, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "SegPoint", + "bbox": [ + 669, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "categories? ASSISTANT: {category} , {category} , ...” To simplify the output and ensure it has only one possible answer, we add the constraints “in the alphabetical order of the categories”. To avoid generating class names not in the dataset, we incorporate category names in a dataset into the prompts during training and inference.", + "bbox": [ + 212, + 146, + 782, + 219 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Referring Segmentation Dataset. We use template prompts: \"USER: Can you segment the object {description} in this point cloud? ASSISTANT: {category} ,\" where {description} is the given explicit description from referring segmentation dataset. It is worth noting that during training, we also use other templates to generate the QA data to ensure data diversity. We add {category} in front of to unify the output format so that when outputting semantic masks, the output category name is the label it predicts.", + "bbox": [ + 212, + 220, + 782, + 325 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation Metrics. We follow most previous works on 3D segmentation [22, 27, 54] to adopt mIoU as primary metric. mIoU is defined by the average of all per-point cloud scene Intersection-over-Unions (IoUs). Besides, we employ accuracy (Acc) as a metric to evaluate whether the model accurately identifies targets with which the predictions have an IoU greater than 0.5.", + "bbox": [ + 212, + 327, + 782, + 401 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Implementation Details", + "text_level": 1, + "bbox": [ + 214, + 411, + 455, + 425 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In our experiments, unless specified otherwise, we employ the LLaMA2-7B model [59] as the large language model $\\mathcal{F}$ and Uni3D [81] as the point cloud processing backbone $\\mathcal{E}$ . The training stage leverages the deepspeed [51] engine for efficiency, with the AdamW [39] optimizer guiding the learning process. The learning rate and weight decay are set to 0.0003 and 0, respectively, enhanced by a WarmupDecayLR learning rate scheduler that initiates with 100 warmup iterations. The projection layer $\\gamma$ utilizes an MLP with channel sizes of [256, 4096, 4096]. We set balancing weight $\\lambda_{txt\\_gen}$ , $\\lambda_{bce}$ , and $\\lambda_{dice}$ to 1.0, 2.0, 2.0, respectively. The experiments utilize a total batch size of 16, distributed across 4 NVIDIA 80G A100 GPUs, and span 5,000 iterations, culminating in a training period of approximately 3 days. During training, we use all mentioned datasets in Sec. 4.1 for joint training by leveraging task-specific prompts. For evaluation on a specific dataset, we finetune the trained model on the corresponding dataset.", + "bbox": [ + 212, + 431, + 784, + 627 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Results on Instruct3D", + "text_level": 1, + "bbox": [ + 214, + 637, + 444, + 651 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The instruction segmentation results, as detailed in Table 1, underscore a significant advancement: where existing methodologies fall short, our model demonstrates exceptional prowess, achieving a more than $15\\%$ improvement in mIoU for tasks requiring intricate reasoning. Unlike conventional referring segmentation tasks, instruction segmentation demands not just identification but also understanding, necessitating the model's reasoning capabilities and access to world knowledge. Existing approaches, confined to explicit references, struggle with implicit queries due to their lack of understanding, which further underscores the task's inherent challenges. In contrast, our model leverages LLMs to bridge this gap, demonstrating superior performance by comprehending and interpreting the queries accurately. Moreover, SegPoint configuration substantially outperforms $\\mathrm{SegPoint}^{\\dagger}$ , highlighting the critical role of our designed Geometric", + "bbox": [ + 212, + 657, + 782, + 838 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "S. He et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/37b1abd7acfbd466834013166a7736ff01b6b3d21873657aae79a4a5a03cdaa3.jpg", + "table_caption": [ + "Table 1: 3D instruction segmentation benchmark results on Instruct3D val split evaluated by Acc and mIoU. † denotes our vanilla baseline removing geometric enhancer module and geometric-guided feature propagation. * represents adding an auxiliary mask head through our implementation." + ], + "table_footnote": [], + "table_body": "
StageMethodReferenceAccmIoU
TwoScanRefer [4][ECCV'20]12.06.9
TwoReferIt3D [1][ECCV'20]11.76.4
TwoM3DRef-CLIP [79][ICCV'23]18.112.8
SingleTGNN [22][AAAI'21]12.97.1
SingleBUTD-DETR [23]*[ECCV'22]16.310.9
SingleEDA [67]*[CVPR'23]16.612.1
SingleSegPoint†[ECCV'24]21.816.1
SingleSegPoint[ECCV'24]31.627.5
", + "bbox": [ + 235, + 202, + 767, + 308 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/63eba471c420636cbd6ca462e4011389cd2a3b32b286114f8059ac0967ac0b41.jpg", + "table_caption": [ + "Table 2: 3D Semantic segmentation benchmark results on S3DIS [3], ScanNet [7], and ScanNet200 [53]. We evaluate on the Area 5 of S3DIS and validation split of ScanNet and ScanNet200." + ], + "table_footnote": [], + "table_body": "
MethodReferenceScanNetScanNet200S3DIS
PointNet++ [46][NeurIPS'17]53.5--
MinkUNet [6][CVPR'19]72.225.065.4
PTv1 [80][ICCV'21]70.627.870.4
PTv2 [66][NeurIPS'22]75.430.271.6
PointNeXt [48][NeurIPS'22]71.5-70.5
OctFormer [60][SIGGRAPH'23]75.732.6-
Swin3D [72][ArXiv]75.5-72.5
SegPoint[ECCV'24]74.135.372.4
", + "bbox": [ + 228, + 363, + 771, + 469 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Enhancer Module and Geometric-guided Feature Propagation components. Notably, even in its baseline form, $\\mathrm{SegPoint}^{\\dagger}$ surpasses all competing methods, validating the effectiveness and rationale behind our pipeline design.", + "bbox": [ + 212, + 478, + 784, + 523 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Besides, different from traditional two-stage approaches that first generate mask proposals using a pre-trained segmentor like Mask3D [54] and then apply language-aware networks for selection, SegPoint directly tackles the task, bypassing the need for preliminary mask proposals, enhancing its efficiency.", + "bbox": [ + 212, + 523, + 785, + 585 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.4 Results on Semantic Segmentation", + "text_level": 1, + "bbox": [ + 214, + 590, + 547, + 606 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 2 present SegPoint's performance on semantic segmentation, delivering competitive results across diverse datasets. Our model uses a simple yet effective answer format, category $<\\mathrm{SEG}>$ , to use category name as predicted labels, achieving particularly stronger performance in datasets with various categories like ScanNet200 [53], where it surpasses SOTA methods by $2.1\\%$ mIoU. To ensure fair comparisons, we fine-tune our model on each semantic segmentation dataset to accommodate the varying class category definitions.", + "bbox": [ + 212, + 612, + 787, + 719 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.5 Results on Referring Segmentation", + "text_level": 1, + "bbox": [ + 214, + 724, + 552, + 739 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 3 presents results on referring segmentation datasets. SegPoint showcases outstanding performance in both single-target (e.g., ScanRefer [4], Nr3D [1]) and multi-target and zero-target contexts within the Multi3DRefer [79] dataset. For multi-targets, we aggregate masks into a single ground truth, and for zero-target, we use an empty mask, indicated by \"ASSISTANT: There is no mask.\" SegPoint significantly surpasses other approaches, achieving $2.5\\%$ mIoU increase. The evaluation process of two-stage method M3DRef-CLIP is similar to Sec.4.3.", + "bbox": [ + 212, + 746, + 787, + 852 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "SegPoint", + "bbox": [ + 669, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/63931d6d9e89f271d82b1bbea49b2ed6de9200a61082e695f1523572d4f736e0.jpg", + "table_caption": [ + "Table 3: 3D referring segmentation benchmark results on ScanRefer [4], Nr3D [1], and Multi3Drefer [79] evaluated by mIoU. * represents adding an auxiliary mask head through our implementation." + ], + "table_footnote": [], + "table_body": "
StageMethodReferenceScanReferNr3DMulti3DRefer
TwoM3DRef-CLIP [79][ICCV'23]35.727.032.6
Two3D-STMN [63][AAAI'24]39.5--
SingleTGNN [22][AAAI'21]27.8--
SingleBUTD-DETR [23]*[ECCV'22]35.427.526.2
SingleEDA [67]*[CVPR'23]36.229.328.9
SingleX-RefSeg3D [49][AAAI'24]29.9--
SingleRefMask3D [18][ACMMM'24]44.8--
SingleSegPoint[ECCV'24]41.732.236.1
", + "bbox": [ + 217, + 188, + 787, + 292 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/a64748139057ad34afa4b22616d4fdc85278df4840627e1944ba0c26a79e3bdb.jpg", + "table_caption": [ + "Table 4: 3D open-vocabulary semantic segmentation benchmark results on val split of ScanNet++ [74]." + ], + "table_footnote": [], + "table_body": "
TypeMethodReferenceScanNet++
SupervisedPointNet [45][CVPR'17]7.0
PointNet++ [46][NeurIPS'17]15.0
MinkUNet [6][CVPR'19]28.0
KPCConv [58][ICCV'19]30.0
Open-VocabularyOpenScene [42][CVPR'23]12.8
PLA [12][CVPR'23]14.2
RegionPLC [71][CVPR'24]14.9
UnifiedSegPoint[ECCV'24]19.3
", + "bbox": [ + 222, + 337, + 772, + 441 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.6 Results on Open-vocabulary Semantic Segmentation", + "text_level": 1, + "bbox": [ + 215, + 453, + 692, + 468 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 4 shows our method's open-vocabulary segmentation performance which is directly evaluated on ScanNet++ [74] following the setting in prevalent methodologies in the 2D domain [13,69]. It demonstrates our superiority over both existing open-vocabulary techniques and even several supervised approaches, showing our model's robust generalization capabilities. It effectively aligns and interprets categories with visual scenes, underscoring the reasoning prowess of large language models. A notable issue is the potential misalignment between output categories of SegPoint and val split category names. To address this, we employ GPT-4 to match its most similar category names in the val split.", + "bbox": [ + 212, + 477, + 787, + 613 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.7 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 619, + 388, + 635 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We conduct extensive experiments to verify the effectiveness of our proposed components in Table 5 (a) on both Instruct3D and ScanRefer [4] dataset. We established a vanilla baseline as described in Sec. 3.2 following the paradigm of LISA [28], which cannot achieve satisfactory performance and only obtain $16.1\\% / 30.3 \\, \\mathrm{mIoU}$ on Instruct3D/ScanRefer, respectively. Further analysis, both qualitative and quantitative, of our proposed components reveals that their integration substantially outperforms the baseline.", + "bbox": [ + 212, + 643, + 787, + 747 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Geometric Enhancer Module Integrating the Geometric Enhancer Module (GEM) into our point encoder results in a notable $5.3\\% / 5.5\\%$ mIoU improvement on Instruct3D/ScanRefer, effectively addressing compatibility issues with dense prediction tasks. An ablation study, shown in Table 5 (b), shows that our improvement is not due to an increase in parameters. Our approach outperforms traditional full fine-tuning, LoRA [21] strategies, and the addition of", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "S. He et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/8854cbebaaf3c9caeb7c11ebdc6e5adec24b15f2c6a0d4939192d2f59d702967.jpg", + "table_caption": [ + "Table 5: Ablation studies on Instruct3D and ScanRefer." + ], + "table_footnote": [ + "(a) Effect of different main components." + ], + "table_body": "
IndexComponentsInstruct3D\nmIoUScanRefer\nmIoU
GEMGFP
0XX16.130.3
1✓X21.435.8
2X✓23.238.1
3✓✓27.541.7
", + "bbox": [ + 233, + 170, + 480, + 243 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/c5e35d614a8a56dabd59d11e08ed0282068b3f70159ce735681869801aff6142.jpg", + "table_caption": [], + "table_footnote": [ + "(b) Effect of different tuning methods." + ], + "table_body": "
MethodInstruct3DScanRefer
Full tuning26.239.1
LoRA [21]24.838.5
MLP23.136.8
GEM27.541.7
", + "bbox": [ + 537, + 175, + 746, + 238 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/51e47c303ec60ce760ff9fb71819deb3f1c4098c4be0ff84843e4d7a33669bc8.jpg", + "image_caption": [ + "(a) Scene" + ], + "image_footnote": [], + "bbox": [ + 220, + 266, + 385, + 335 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/4964324b8e48caa54aaba3259917eb79329a6315d1f68c79b16146e456d1fd6f.jpg", + "image_caption": [ + "(b) Baseline", + "Fig. 4: (Best viewed in color) We visualize the feature responses between a given point (in red) and other points in the scene from per-point embeddings $\\pmb{f}_{\\mathcal{P}}$ for the baseline and our SegPoint, respectively. The color changes from yellow to red, indicating increasing feature similarity." + ], + "image_footnote": [], + "bbox": [ + 416, + 270, + 584, + 335 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/3fa27e0764c2c5533d83c3f73c921b9fb735de1198169d78d22860929c027981.jpg", + "image_caption": [ + "(c) Ours" + ], + "image_footnote": [], + "bbox": [ + 617, + 270, + 782, + 335 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "MLP layers for feature adapter, underscoring its effectiveness in embedding 3D domain-specific knowledge into point encoder. Unlike adapter techniques common in language and 2D image processing, our GEM is designed to address the unique challenges in dense prediction tasks.", + "bbox": [ + 212, + 426, + 784, + 486 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Geometric-guided Feature Propagation Introducing Geometric-guided Feature Propagation (GFP) results in a substantial improvement over the baseline, as shown in Table 5 (a) (index 2). This underscores our method's capability to minimize information loss and reduce noise during the upsampling phase, leading to higher-quality per-point embeddings.", + "bbox": [ + 212, + 484, + 784, + 560 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.8 Qualitative Visualization", + "text_level": 1, + "bbox": [ + 214, + 570, + 465, + 585 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Fig. 4 qualitatively illustrates the feature responses between a given point and others in a scene. From left to right, it presents the original scene, baseline method, and our SegPoint, respectively, with warmer colors indicating closer feature relationships. The baseline method predominantly highlights spatially proximate points, often missing important but distant features, such as chair cushions. In contrast, SegPoint leverages global information, allowing for the identification of both near and distant relevant features, such as cushions and casters. This demonstrates our model's superior ability to capture global context and recognize intricate structures within the scene.", + "bbox": [ + 212, + 594, + 787, + 729 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "As shown in Fig. 5, we provide some typical qualitative results from SegPoint on Instruct3D. Given an implicit instruction, e.g., \"the object that is used for dispensing water during bathing\", SegPoint successfully infers the role of a shower. In other scenes, the instructions include queries requiring extensive knowledge of the world and complex reasoning, like \"What appliance is used for heating or cooking food quickly using electromagnetic radiation.\" SegPoint can still segment the microwave oven very well, which shows that SegPoint can make good use of the reasoning ability of LLM and provide high-quality masks.", + "bbox": [ + 212, + 731, + 787, + 852 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "SegPoint", + "bbox": [ + 669, + 114, + 730, + 127 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/e476decbca80d6647d5c0d0f8dc189a75d2ac5cba270fa39298ad535ccb4c3bf.jpg", + "image_caption": [ + "Fig. 5: Qualitative results from val split of Instruct3D. SegPoint understand the human instruction and accurately segment the target object. We omitted the \"please output segmentation mask\" in the sentence for simplicity." + ], + "image_footnote": [], + "bbox": [ + 218, + 146, + 785, + 508 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 566, + 359, + 583 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we introduce SegPoint, an effective model supported by LLM for point-level reasoning and segmentation. Benefiting from the proposed Geometric Enhancer Module and Geometric-guided Feature Propagation, SegPoint is adept at solving a variety of segmentation tasks in a unified framework. Additionally, we construct a comprehensive Instruct3D benchmark to bolster research area in segmentation via implicit and complex instructions, introducing more challenges to promote it closer to real-world applications. Through thorough experiments, SegPoint achieves promising results across multiple benchmarks.", + "bbox": [ + 212, + 597, + 787, + 719 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations Although SegPoint demonstrates notable success in tasks driven by text prompts, its current framework cannot process non-textual prompts, such as boxes and points. Future developments will explore the adoption of a prompt encoder, inspired by SAM model [26], to extend support for these formats.", + "bbox": [ + 212, + 723, + 787, + 782 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements We thank the anonymous reviewers for their constructive suggestions. Following their advice, we have incorporated diverse location and view descriptions into our Instruct3D. This work was partially supported by the National Research Foundation Singapore Competitive Research Program (CRP29-2022-0003).", + "bbox": [ + 212, + 784, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "S. He et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 323, + 160 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Achlioptas, P., Abdelreheem, A., Xia, F., Elhoseiny, M., Guibas, L.: Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In: ECCV (2020)", + "2. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. In: NeurIPS (2022)", + "3. Armeni, I., Sener, O., Zamir, A.R., Jiang, H., Brilakis, I., Fischer, M., Savarese, S.: 3d semantic parsing of large-scale indoor spaces. In: CVPR (2016)", + "4. Chen, D.Z., Chang, A.X., Nießner, M.: Scanrefer: 3d object localization in rgb-d scans using natural language. In: ECCV (2020)", + "5. Chen, Z., Duan, Y., Wang, W., He, J., Lu, T., Dai, J., Qiao, Y.: Vision transformer adapter for dense predictions. In: ICLR (2023)", + "6. Choy, C., Gwak, J., Savarese, S.: 4d spatio-temporal convnets: Minkowski convolutional neural networks. In: CVPR (2019)", + "7. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: CVPR (2017)", + "8. Ding, H., Liu, C., He, S., Jiang, X., Loy, C.C.: MeViS: A large-scale benchmark for video segmentation with motion expressions. In: ICCV (2023)", + "9. Ding, H., Liu, C., He, S., Jiang, X., Torr, P.H., Bai, S.: MOSE: A new dataset for video object segmentation in complex scenes. In: ICCV (2023)", + "10. Ding, H., Liu, C., Wang, S., Jiang, X.: Vision-language transformer and query generation for referring segmentation. In: ICCV (2021)", + "1. Ding, H., Liu, C., Wang, S., Jiang, X.: VLT: Vision-language transformer and query generation for referring segmentation. IEEE TPAMI (2023)", + "2. Ding, R., Yang, J., Xue, C., Zhang, W., Bai, S., Qi, X.: Pla: Language-driven open-vocabulary 3d scene understanding. In: CVPR (2023)", + "3. Ding, Z., Wang, J., Tu, Z.: Open-vocabulary panoptic segmentation with maskclip. In: ICLR (2023)", + "4. Girdhar, R., El-Nouby, A., Liu, Z., Singh, M., Alwala, K.V., Joulin, A., Misra, I.: Imagebind: One embedding space to bind them all. In: CVPR (2023)", + "5. Guo, Z., Zhang, R., Zhu, X., Tang, Y., Ma, X., Han, J., Chen, K., Gao, P., Li, X., Li, H., et al.: Point-bind & point-llm: Aligning point cloud with multi-modality for 3d understanding, generation, and instruction following. arXiv preprint arXiv:2309.00615 (2023)", + "6. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016)", + "7. He, S., Ding, H.: Decoupling static and hierarchical motion perception for referring video segmentation. In: CVPR (2024)", + "8. He, S., Ding, H.: RefMask3D: Language-guided transformer for 3d referring segmentation. In: ACM MM (2024)", + "9. He, S., Jiang, X., Jiang, W., Ding, H.: Prototype adaption and projection for few-and zero-shot 3d point cloud semantic segmentation. IEEE TIP (2023)", + "20. Hong, Y., Zhen, H., Chen, P., Zheng, S., Du, Y., Chen, Z., Gan, C.: 3d-llm: Injecting the 3d world into large language models. In: NeurIPS (2023)", + "21. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. In: ICLR (2022)", + "22. Huang, P.H., Lee, H.H., Chen, H.T., Liu, T.L.: Text-guided graph neural networks for referring 3d instance segmentation. In: AAAI (2021)" + ], + "bbox": [ + 225, + 176, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "SegPoint", + "bbox": [ + 669, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "23. Jain, A., Gkanatsios, N., Mediratta, I., Fragkiadaki, K.: Bottom up top down detection transformers for language grounding in images and point clouds. In: ECCV (2022)", + "24. Jia, B., Chen, Y., Yu, H., Wang, Y., Niu, X., Liu, T., Li, Q., Huang, S.: Sceneverse: Scaling 3d vision-language learning for grounded scene understanding. In: ECCV (2024)", + "25. Jiang, L., Zhao, H., Shi, S., Liu, S., Fu, C.W., Jia, J.: Pointgroup: Dual-set point grouping for 3d instance segmentation. In: CVPR (2020)", + "26. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: ICCV (2023)", + "27. Kolodiazhnyi, M., Vorontsova, A., Konushin, A., Rukhovich, D.: Oneformer3d: One transformer for unified point cloud segmentation. In: CVPR (2024)", + "28. Lai, X., Tian, Z., Chen, Y., Li, Y., Yuan, Y., Liu, S., Jia, J.: Lisa: Reasoning segmentation via large language model. In: CVPR (2024)", + "29. Lai, X., Yuan, Y., Chu, R., Chen, Y., Hu, H., Jia, J.: Mask-attention-free transformer for 3d instance segmentation. In: ICCV (2023)", + "30. Li, B., Zhang, Y., Chen, L., Wang, J., Yang, J., Liu, Z.: Otter: A multi-modal model with in-context instruction tuning. arXiv:2305.03726 (2023)", + "31. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. In: NeurIPS (2023)", + "32. Lin, Z., Liu, C., Zhang, R., Gao, P., Qiu, L., Xiao, H., Qiu, H., Lin, C., Shao, W., Chen, K., et al.: Sphinx: The joint mixing of weights, tasks, and visual embeddings for multi-modal large language models. arXiv preprint arXiv:2311.07575 (2023)", + "33. Liu, C., Ding, H., Jiang, X.: GRES: Generalized referring expression segmentation. In: CVPR (2023)", + "34. Liu, C., Ding, H., Zhang, Y., Jiang, X.: Multi-modal mutual attention and iterative interaction for referring image segmentation. IEEE TIP (2023)", + "35. Liu, C., Jiang, X., Ding, H.: Instance-specific feature propagation for referring segmentation. IEEE TMM (2022)", + "36. Liu, C., Jiang, X., Ding, H.: Primitivenet: decomposing the global constraints for referring segmentation. Visual Intelligence 2(1), 16 (2024)", + "37. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: NeurIPS (2023)", + "38. Liu, Y., Kong, L., Cen, J., Chen, R., Zhang, W., Pan, L., Chen, K., Liu, Z.: Segment any point cloud sequences by distilling vision foundation models. In: NeurIPS (2023)", + "39. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: ICLR (2019)", + "40. Nguyen, P.D., Ngo, T.D., Gan, C., Kalogerakis, E., Tran, A., Pham, C., Nguyen, K.: Open3dis: Open-vocabulary 3d instance segmentation with 2d mask guidance. In: CVPR (2024)", + "41. Park, N., Kim, S.: How do vision transformers work? arXiv preprint arXiv:2202.06709 (2022)", + "42. Peng, S., Genova, K., Jiang, C., Tagliasacchi, A., Pollefeys, M., Funkhouser, T., et al.: Openscene: 3d scene understanding with open vocabularies. In: CVPR (2023)", + "43. Peng, Z., Wang, W., Dong, L., Hao, Y., Huang, S., Ma, S., Wei, F.: Kosmos-2: Grounding multimodal large language models to the world. In: ICLR (2024)", + "44. Pi, R., Gao, J., Diao, S., Pan, R., Dong, H., Zhang, J., Yao, L., Han, J., Xu, H., Zhang, L.K.T.: Detgpt: Detect what you need via reasoning. In: EMNLP (2023)", + "45. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: CVPR (2017)", + "46. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In: NeurIPS (2017)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "S. He et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "47. Qi, Z., Fang, Y., Sun, Z., Wu, X., Wu, T., Wang, J., Lin, D., Zhao, H.: Gpt4point: A unified framework for point-language understanding and generation. In: CVPR (2024)", + "48. Qian, G., Li, Y., Peng, H., Mai, J., Hammoud, H., Elhoseiny, M., Ghanem, B.: Pointnext: Revisiting pointnet++ with improved training and scaling strategies. In: NeurIPS (2022)", + "49. Qian, Z., Ma, Y., Ji, J., Sun, X.: X-refseg3d: Enhancing referring 3d instance segmentation via structured cross-modal graph neural networks. In: AAAI (2024)", + "50. Rasheed, H., Maaz, M., Shaji, S., Shaker, A., Khan, S., Cholakkal, H., Anwer, R.M., Xing, E., Yang, M.H., Khan, F.S.: Glamm: Pixel grounding large multimodal model. In: CVPR (2024)", + "51. Rasley, J., Rajbhandari, S., Ruwase, O., He, Y.: Deepspeed: System optimizations enable training deep learning models with over 100 billion parameters. In: KDD (2020)", + "52. Ren, Z., Huang, Z., Wei, Y., Zhao, Y., Fu, D., Feng, J., Jin, X.: Pixelm: Pixel reasoning with large multimodal model. In: CVPR (2024)", + "53. Rozenberszki, D., Litany, O., Dai, A.: Language-grounded indoor 3d semantic segmentation in the wild. In: ECCV (2022)", + "54. Schult, J., Engelmann, F., Hermans, A., Litany, O., Tang, S., Leibe, B.: Mask3d: Mask transformer for 3d semantic instance segmentation. In: ICRA (2023)", + "55. Shuai, X., Ding, H., Ma, X., Tu, R., Jiang, Y.G., Tao, D.: A survey of multimodal-guided image editing with text-to-image diffusion models. arXiv preprint arXiv:2406.14555 (2024)", + "56. Sun, J., Qing, C., Tan, J., Xu, X.: Superpoint transformer for 3d scene instance segmentation. In: AAAI (2023)", + "57. Takmaz, A., Fedele, E., Sumner, R.W., Pollefeys, M., Tombari, F., Engelmann, F.: Openmask3d: Open-vocabulary 3d instance segmentation. In: NeurIPS (2023)", + "58. Thomas, H., Qi, C.R., Deschaud, J.E., Marcotegui, B., Goulette, F., Guibas, L.J.: Kpconv: Flexible and deformable convolution for point clouds. In: ICCV (2019)", + "59. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)", + "60. Wang, P.S.: Octformer: Octree-based transformers for 3d point clouds. In: SIGGRAPH (2023)", + "61. Wang, W., Chen, Z., Chen, X., Wu, J., Zhu, X., Zeng, G., Luo, P., Lu, T., Zhou, J., Qiao, Y., et al.: Visionllm: Large language model is also an open-ended decoder for vision-centric tasks. In: NeurIPS (2023)", + "62. Wang, W., Xie, E., Li, X., Fan, D.P., Song, K., Liang, D., Lu, T., Luo, P., Shao, L.: Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media (2022)", + "63. Wu, C., Ma, Y., Chen, Q., Wang, H., Luo, G., Ji, J., Sun, X.: 3d-stmn: Dependency-driven superpoint-text matching network for end-to-end 3d referring expression segmentation. In: AAAI (2024)", + "64. Wu, H., Xiao, B., Codella, N., Liu, M., Dai, X., Yuan, L., Zhang, L.: Cvt: Introducing convolutions to vision transformers. In: ICCV (2021)", + "65. Wu, J., Li, X., Xu, S., Yuan, H., Ding, H., Yang, Y., Li, X., Zhang, J., Tong, Y., Jiang, X., Ghanem, B., Tao, D.: Towards open vocabulary learning: A survey. IEEE TPAMI (2024)", + "66. Wu, X., Lao, Y., Jiang, L., Liu, X., Zhao, H.: Point transformer v2: Grouped vector attention and partition-based pooling. In: NeurIPS (2022)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "SegPoint", + "bbox": [ + 669, + 114, + 730, + 127 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "67. Wu, Y., Cheng, X., Zhang, R., Cheng, Z., Zhang, J.: Eda: Explicit text-decoupling and dense alignment for 3d visual grounding. In: CVPR (2023)", + "68. Xiao, Z., Zhang, W., Wang, T., Loy, C.C., Lin, D., Pang, J.: Position-guided point cloud panoptic segmentation transformer. arXiv preprint arXiv:2303.13509 (2023)", + "69. Xu, J., Liu, S., Vahdat, A., Byeon, W., Wang, X., De Mello, S.: Open-vocabulary panoptic segmentation with text-to-image diffusion models. In: CVPR (2023)", + "70. Xu, R., Wang, X., Wang, T., Chen, Y., Pang, J., Lin, D.: Pointllm: Empowering large language models to understand point clouds. In: ECCV (2024)", + "71. Yang, J., Ding, R., Deng, W., Wang, Z., Qi, X.: Regionplc: Regional point-language contrastive learning for open-world 3d scene understanding. In: CVPR (2024)", + "72. Yang, Y.Q., Guo, Y.X., Xiong, J.Y., Liu, Y., Pan, H., Wang, P.S., Tong, X., Guo, B.: Swin3d: A pretrained transformer backbone for 3d indoor scene understanding. arXiv preprint arXiv:2304.06906 (2023)", + "73. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: MPLug-owl: Modularization empowers large language models with multimodality. arXiv:2304.14178 (2023)", + "74. Yeshwanth, C., Liu, Y.C., Nießner, M., Dai, A.: Scannet++: A high-fidelity dataset of 3d indoor scenes. In: ICCV (2023)", + "75. You, H., Zhang, H., Gan, Z., Du, X., Zhang, B., Wang, Z., Cao, L., Chang, S.F., Yang, Y.: Ferret: Refer and ground anything anywhere at any granularity. In: ICLR (2024)", + "76. Zhang, H., Li, H., Li, F., Ren, T., Zou, X., Liu, S., Huang, S., Gao, J., Zhang, L., Li, C., et al.: Llava-grounding: Grounded visual chat with large multimodal models. arXiv preprint arXiv:2312.02949 (2023)", + "77. Zhang, H., Ding, H.: Prototypical matching and open set rejection for zero-shot semantic segmentation. In: ICCV (2021)", + "78. Zhang, S., Sun, P., Chen, S., Xiao, M., Shao, W., Zhang, W., Chen, K., Luo, P.: Gpt4roi: Instruction tuning large language model on region-of-interest. arXiv:2307.03601 (2023)", + "79. Zhang, Y., Gong, Z., Chang, A.X.: Multi3defer: Grounding text description to multiple 3d objects. In: ICCV (2023)", + "80. Zhao, H., Jiang, L., Jia, J., Torr, P.H., Koltun, V.: Point transformer. In: ICCV (2021)", + "81. Zhou, J., Wang, J., Ma, B., Liu, Y.S., Huang, T., Wang, X.: Uni3d: Exploring unified 3d representation at scale. In: ICLR (2024)", + "82. Zhou, Z., Zhang, Y., Foroosh, H.: Panoptic-polarnet: Proposal-free lidar point cloud panoptic segmentation. In: CVPR (2021)", + "83. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. In: ICLR (2024)", + "84. Zhu, Z., Ma, X., Chen, Y., Deng, Z., Huang, S., Li, Q.: 3d-vista: Pre-trained transformer for 3d vision and text alignment. In: ICCV (2023)" + ], + "bbox": [ + 215, + 147, + 784, + 714 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "S. He et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/SegPoint_ Segment Any Point Cloud via Large Language Model/d283f64a-a017-47de-9be3-9206675b2c4f_model.json b/2024/SegPoint_ Segment Any Point Cloud via Large Language Model/d283f64a-a017-47de-9be3-9206675b2c4f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2bfce24df50d8b6236f45d9a73b354f09dcff1ba --- /dev/null +++ b/2024/SegPoint_ Segment Any Point Cloud via Large Language Model/d283f64a-a017-47de-9be3-9206675b2c4f_model.json @@ -0,0 +1,2755 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.243, + 0.145, + 0.288, + 0.178 + ], + "angle": 0, + "content": "#" + }, + { + "type": "title", + "bbox": [ + 0.303, + 0.155, + 0.761, + 0.199 + ], + "angle": 0, + "content": "SegPoint: Segment Any Point Cloud via Large Language Model" + }, + { + "type": "text", + "bbox": [ + 0.243, + 0.224, + 0.761, + 0.241 + ], + "angle": 0, + "content": "Shuting \\(\\mathrm{He}^{1}\\) , Henghui Ding2, Xudong Jiang1, and Bihan \\(\\mathrm{Wen}^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.361, + 0.252, + 0.642, + 0.266 + ], + "angle": 0, + "content": "1 Nanyang Technological University" + }, + { + "type": "text", + "bbox": [ + 0.361, + 0.266, + 0.642, + 0.281 + ], + "angle": 0, + "content": "\\(^{2}\\) Institute of Big Data, Fudan University" + }, + { + "type": "text", + "bbox": [ + 0.224, + 0.281, + 0.782, + 0.295 + ], + "angle": 0, + "content": "{heshuting555, henghui.ding}@gmail.com, {exdjiang, bihan.wen}@ntu.edu.sg" + }, + { + "type": "text", + "bbox": [ + 0.35, + 0.295, + 0.653, + 0.309 + ], + "angle": 0, + "content": "https://heshuting555.github.io/SegPoint" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.343, + 0.744, + 0.594 + ], + "angle": 0, + "content": "Abstract. Despite significant progress in 3D point cloud segmentation, existing methods primarily address specific tasks and depend on explicit instructions to identify targets, lacking the capability to infer and understand implicit user intentions in a unified framework. In this work, we propose a model, called SegPoint, that leverages the reasoning capabilities of a multi-modal Large Language Model (LLM) to produce point-wise segmentation masks across a diverse range of tasks: 1) 3D instruction segmentation, 2) 3D referring segmentation, 3) 3D semantic segmentation, and 4) 3D open-vocabulary semantic segmentation. To advance 3D instruction research, we introduce a new benchmark, Instruct3D, designed to evaluate segmentation performance from complex and implicit instructional texts, featuring 2,565 point cloud-instruction pairs. Our experimental results demonstrate that SegPoint achieves competitive performance on established benchmarks such as ScanRefer for referring segmentation and ScanNet for semantic segmentation, while delivering outstanding outcomes on the Instruct3D dataset. To our knowledge, SegPoint is the first model to address these varied segmentation tasks within a single framework, achieving satisfactory performance." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.606, + 0.741, + 0.634 + ], + "angle": 0, + "content": "Keywords: Instruct3D dataset \\(\\cdot\\) Unified framework \\(\\cdot\\) 3D point cloud segmentation \\(\\cdot\\) Large language model" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.68, + 0.377, + 0.696 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.711, + 0.788, + 0.817 + ], + "angle": 0, + "content": "3D point cloud segmentation, a critical challenge in the 3D vision community, aims to interpret and classify each point in a point cloud to understand its semantic properties [27, 45, 46, 54, 56, 80]. This longstanding issue has spurred significant advancements across various fields, including robotics, autonomous driving, virtual reality, etc. This challenge has evolved into a series of specialized tasks, each targeting a specific segmentation aspect. Overall, tasks cover basic semantic and instance segmentation [3, 7, 53, 74], as well as more practical tasks" + }, + { + "type": "page_footnote", + "bbox": [ + 0.23, + 0.826, + 0.402, + 0.841 + ], + "angle": 0, + "content": "Corresponding author" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.35, + 0.128 + ], + "angle": 0, + "content": "S. He et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.145, + 0.784, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.473, + 0.788, + 0.53 + ], + "angle": 0, + "content": "Fig. 1: Example of functionality in SegPoint. SegPoint can complete various point cloud tasks in a unified framework by leveraging task-specific prompts, including 1) 3D instruction segmentation, 2) 3D referring segmentation, 3) 3D semantic segmentation, and 4) 3D open-vocabulary semantic segmentation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.549, + 0.788, + 0.596 + ], + "angle": 0, + "content": "such as referring segmentation [1, 4, 24, 33, 79, 84], which segments points based on explicit textual descriptions, and open-vocabulary segmentation [12, 40, 42, 57, 65, 77] designed for the dynamic and complex nature of real-world." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.597, + 0.789, + 0.794 + ], + "angle": 0, + "content": "Despite significant progress achieved within the 3D community toward accurately segmenting objects through specifically designed models, each model is typically developed to tackle one specific segmentation task, leading to inefficiencies and a lack of versatility for real-world application. Furthermore, previous perception approaches heavily depend on predefined categories or explicit expressions for language understanding. Such approaches fall short in interpreting and acting on implicit instructions often found in human language, a critical gap that hinders the development of truly intelligent next-generation perception systems. This brings a pivotal question: Is it possible to design a unified model capable of comprehensively addressing all aforementioned 3D tasks with human-like instructions? The exploration of this question not only challenges the current paradigms of 3D point cloud segmentation but also opens the door to groundbreaking advancements in robotic perception and interaction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.789, + 0.842 + ], + "angle": 0, + "content": "In this work, we propose a model called SegPoint, leveraging the Large Language Model's (LLM) advanced ability to reason and comprehend user instructions. To enhance 3D scene comprehension, we integrate a Geometric Enhancer" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.67, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SegPoint" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.128 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "Module that extracts local semantics from point clouds, seamlessly incorporating this geometric insight into the feature extraction process. Furthermore, a Geometric-guided Feature Propagation is designed to utilize semantic flows derived from geometric priors and the LLM's hidden embeddings, facilitating the generation of fine-grained, high-quality features for accurate dense prediction tasks. Unlike previous attempts at 2D field [8, 28, 50], we do not depend on additional costly pre-trained segmentation models like SAM [26]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.788, + 0.359 + ], + "angle": 0, + "content": "Moreover, we introduce a benchmark named Instruct3D, designed to advance research in the field of segmentation driven by implicit and complex instructions. Understanding these nuanced instructions necessitates reasoning abilities and extensive knowledge of the world. It includes a total of 2,565 diverse pairs of instructions and point clouds for tuning and evaluation. Our comprehensive experiments demonstrate the benchmark's utility in evaluating the model's capability of segmentation based on human-like instructions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.359, + 0.788, + 0.465 + ], + "angle": 0, + "content": "Taking advantage of a multi-modal LLM and task-specific prompts, SegPoint is capable of generating segmentation masks for a wide range of tasks in a unified model: 1) 3D instruction segmentation, 2) 3D referring segmentation, 3) 3D semantic segmentation, and 4) 3D open-vocabulary semantic segmentation, as depicted in Fig. 1. SegPoint achieves competitive results on established benchmarks like ScanRefer [4] for referring segmentation and ScanNet [7] for semantic segmentation while showing remarkable performance on the Instruct3D dataset." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.466, + 0.608, + 0.48 + ], + "angle": 0, + "content": "In summary, our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.492, + 0.785, + 0.537 + ], + "angle": 0, + "content": "- We propose SegPoint, the first 3D segmentation model that can comprehend human intentions and address multiple segmentation tasks within one framework by harnessing the Large Language Model's reasoning capabilities." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.539, + 0.785, + 0.613 + ], + "angle": 0, + "content": "- We present a Geometric Enhancer Module that integrates comprehensive scene information into the process of 3D scene understanding. Besides, A Geometric-guided Feature Propagation is designed to achieve accurate and fine-grained segmentation. These two modules supplement the missing local information and grasp fine-grained features for dense prediction tasks." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.616, + 0.785, + 0.661 + ], + "angle": 0, + "content": "- We introduce a new task called 3D instruction segmentation and construct a new dataset Instruct3D, which necessitates a model's self-reasoning to interpret implicit instructions for segmenting the target object." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.664, + 0.785, + 0.723 + ], + "angle": 0, + "content": "- Our experimental findings reveal that SegPoint not only competes strongly in 3D semantic, referring, and open-vocabulary semantic segmentation but also excels in 3D instruction segmentation, showcasing its versatility and effectiveness across a spectrum of segmentation challenges." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.749, + 0.388, + 0.765 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.783, + 0.561, + 0.799 + ], + "angle": 0, + "content": "2.1 Multi-modal Large Language Model" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Inspired by the exceptional reasoning abilities of Large Language Models, researchers are delving into transferring these capabilities into the vision realm [9," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.348, + 0.127 + ], + "angle": 0, + "content": "S. He et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.358 + ], + "angle": 0, + "content": "28, 36, 55], developing multi-modal LLMs. Flamingo [2], BLIP-2 [31], mPLUGOWL [73], Otter [30], LLaVA [37] and MiniGPT-4 [83] first construct image-text feature alignment followed by instruction tuning and achieve superior performance. Recent studies have increasingly concentrated on integrating foundational models with tasks that demand a refined understanding at the region or pixel level. VisionLLM [61] introduces a versatile interaction interface for various vision-centric tasks via instruction tuning, albeit without fully leveraging LLMs for intricate reasoning tasks. Kosmos-2 [43] has built substantial data of grounded image-text pairs, thereby embedding grounding capabilities into LLMs. DetGPT [44] links the fixed multi-modal LLM with an open-vocabulary detector, facilitating user instruction-based detection tasks. This growing interest has spurred further innovations, including GPT4RoI [78], LLaVA-grounding [76], Ferret [75], LISA [28], GlaMM [50], PixelLM [52], Sphinx [32], each contributing to the evolving landscape of instruction-based, multi-modal understanding." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.359, + 0.79, + 0.57 + ], + "angle": 0, + "content": "Building on advancements of multi-modal large language models in 2D image domain, the field is witnessing a seamless transition into 3D spaces. PointLLM [70] harnesses the prowess of large language models and trains with 3D point clouds following the paradigm of LLaVA [37]. 3D-LLM [20] utilizes 2D foundation models to encode multi-view images of 3D point clouds. Point-Bind [15] aligns point clouds with Image-Bind [14] and leverages ImageBind-LLM to reason multimodality input without 3D-instruction data training. GPT4Point [47] pioneers in facilitating a unified approach towards 3D object understanding and generation, setting a new standard for versatility. However, these models primarily concentrate on scene-level insights, often overlooking the intricate details at the region, or point level. Contrasting with existing models, our research focuses on two pivotal goals: 1) efficiently inject segmentation capabilities into multi-modal LLMs to conduct point-level understanding and 2) design a unified framework for 3D point cloud segmentation via the reasoning ability of LLMs." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.59, + 0.51, + 0.606 + ], + "angle": 0, + "content": "2.2 3D Point Cloud Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.794 + ], + "angle": 0, + "content": "3D point cloud segmentation, a crucial task in computer vision, can be categorized into 3D semantic, instance, and panoptic segmentation. 3D semantic segmentation [3,7,19,45,53,74,80] assigns each point in a 3D space to specific, predefined classes. In contrast, 3D instance segmentation [25,27,54] goes a step further by classifying each point and differentiating between distinct objects of the same class. 3D panoptic segmentation [68,82] aims to group 3D points according to their semantics and identities. Lately, more practical tasks have emerged, such as 3D referring segmentation [1,4,17,23,67,79], which extends referring expression segmentation [8,10,11,33-35] to 3D and segments a target instance based on explicit linguistic descriptions, and 3D open-vocabulary segmentation [12,40,42,57], designed to identify and segment unseen objects beyond a fixed set of known categories." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.796, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Despite the significant progress made by transformer-based models: Mask3D [54], SPFormer [56], MAFT [29], and OneFormer3D [27] in basic 3D segmentation tasks, their application in real-world scenarios requiring human language" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.67, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SegPoint" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.149, + 0.784, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.332, + 0.788, + 0.404 + ], + "angle": 0, + "content": "Fig. 2: The pipeline of SegPoint. Given input point cloud and text query, the multimodal LLM \\(\\mathcal{F}\\) generates text output. Geometric Enhancer Module \\(\\mathcal{G}\\) injects geometric information into Point Encoder \\(\\mathcal{E}\\) and obtains point features \\(\\hat{\\pmb{f}}_{\\text {point }}\\). Per-point embeddings \\(\\pmb{f}_{\\mathcal{P}}\\) derived from Geometric-guided Feature Propagation \\(\\mathcal{P}\\) multiplied with the embedding associated with the token yield the final segmentation masks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.418, + 0.789, + 0.555 + ], + "angle": 0, + "content": "interaction remains constrained. Besides, Seal [38] aims to segment any point cloud through distilling vision foundation models, while it doesn't use language as cues. TGNN [22] is the first work to tackle referring segmentation problem that proposes aggregating textual features by considering the neighboring local structure of each instance but it heavily depends on explicit expressions or predefined categories for language understanding. Furthermore, the development of models tailored to specific segmentation tasks restricts their versatility and applicability in diverse real-world scenarios. Therefore, it is imperative to develop more intelligent interaction ways and a unified model for 3D point segmentation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.577, + 0.347, + 0.594 + ], + "angle": 0, + "content": "3 Approach" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.609, + 0.45, + 0.623 + ], + "angle": 0, + "content": "3.1 Architecture Overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.634, + 0.788, + 0.771 + ], + "angle": 0, + "content": "The overall architecture of SegPoint is presented in Fig. 2. SegPoint mainly comprises four parts: \\(i\\) ) a pre-trained point encoder \\(\\mathcal{E}\\) tailored for aligning with textual data; \\(ii\\) ) a large language model \\(\\mathcal{F}\\) endowed with advanced reasoning capabilities; \\(iii\\) ) a Geometric Enhancer Module \\(\\mathcal{G}\\) responsible for extracting geometric representation from input point clouds and infusing these priors into the point encoder; and \\(iv\\) ) a Geometric-guided Feature Propagation \\(\\mathcal{P}\\) which is key to achieving precise mask generation. The collaboration between the Geometric Enhancer Module and Geometric-guided Feature Propagation is crucial, as it equips LLMs with the ability to generate masks effectively in various scenarios." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.785, + 0.395, + 0.799 + ], + "angle": 0, + "content": "3.2 Vanilla Baseline" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "The input of the framework is the text instructions \\( \\pmb{i}_{\\text{txt}} \\) and point cloud \\( \\pmb{i}_{\\text{point}} \\in \\mathbb{R}^{N \\times (3 + F)} \\). Specifically, a point cloud scene, comprising \\( N \\) points, each includes" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.348, + 0.128 + ], + "angle": 0, + "content": "S. He et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.239 + ], + "angle": 0, + "content": "3D coordinates \\(\\in \\mathbb{R}^3\\) and an auxiliary feature vector \\(\\in \\mathbb{R}^F\\) (e.g., color). The point cloud \\(i_{point}\\) is fed into the point encoder \\(\\mathcal{E}\\), which extracts point features \\(f_{point} \\in \\mathbb{R}^{N_1 \\times D}\\), where \\(N_1 \\ll N\\), \\(D\\) is the feature dimension. Concurrently, the text instruction \\(i_{txt}\\) undergoes tokenization via \\(\\mathcal{F}_{tokenize}\\). These prepared inputs are then fed into the Large Language Model \\(\\mathcal{F}\\), resulting in a textual response \\(\\pmb{y}\\). The above process can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.27, + 0.252, + 0.786, + 0.269 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {f} _ {\\text {p o i n t}} = \\mathcal {E} (\\boldsymbol {i} _ {\\text {p o i n t}}), \\quad \\boldsymbol {f} _ {\\text {t x t}} = \\mathcal {F} _ {\\text {t o k e n i z e}} (\\boldsymbol {i} _ {\\text {t x t}}), \\quad \\boldsymbol {y} = \\mathcal {F} (\\boldsymbol {f} _ {\\text {p o i n t}}, \\boldsymbol {f} _ {\\text {t x t}}). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.28, + 0.789, + 0.446 + ], + "angle": 0, + "content": "Building on the approach introduced by LISA [28], SegPoint enhances the segmentation capabilities of Large Language Models (LLMs) by expanding their vocabulary with a new special token, \\(<\\mathrm{SEG}>.\\) This modification enables the model to recognize and predict the \\(<\\mathrm{SEG}>\\) token within the output sequence as a signal to identify segmentation targets. Upon detecting a \\(<\\mathrm{SEG}>\\) token, the corresponding output sequence belonging to \\(<\\mathrm{SEG}>\\) token is extracted and processed through an MLP layer \\(\\gamma\\), generating mask embeddings \\(h_{seg}\\). The final step involves computing each binary mask prediction \\(\\pmb{m} \\in \\mathbb{R}^{N}\\) by performing a dot product between the mask embeddings \\(h_{seg}\\) and the upsampled per-point embeddings derived from the point features \\(\\pmb{f}_{point}\\). The formulation of the aforementioned process is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.342, + 0.459, + 0.785, + 0.477 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {h} _ {\\text {s e g}} = \\gamma (\\boldsymbol {y} _ {[ \\text {s e g} ]}), \\quad \\boldsymbol {m} = \\boldsymbol {h} _ {\\text {s e g}} \\otimes U p S. (\\boldsymbol {f} _ {\\text {p o i n t}}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.487, + 0.788, + 0.653 + ], + "angle": 0, + "content": "where \\( UpS \\) denotes the upsampling operation following PointNet++ [46] on \\( f_{point} \\). The vanilla baseline represents an initial attempt to bridge the gap between LLMs' text comprehension and point cloud segmentation tasks. It encounters two primary issues. Firstly, the point encoder is trained on a scene-level dataset for classification achieving alignment between text and point clouds, not specifically trained for dense prediction tasks. Besides, the point encoder's first layer employs Farthest Point Sampling (FPS) [46] to reduce the point cloud to \\( N_1 \\) points, risking the loss of details vital for accurate dense predictions. Secondly, the operation of directly upsampling from \\( N_1 \\) to \\( N \\) points to obtain per-point embeddings is prone to losing structural information and introducing a notable degree of noise, undermining the model's efficacy in segmentation tasks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.678, + 0.502, + 0.693 + ], + "angle": 0, + "content": "3.3 Geometric Enhancer Module" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.789, + 0.841 + ], + "angle": 0, + "content": "To adapt the pre-trained point encoder for dense prediction tasks while maintaining its superior scene recognition capability, our objective is to harness the geometric information across the entire scene to guide the further feature learning process. Drawing inspiration from recent advancements in 2D computer vision, where studies [5,41,62,64] demonstrate that convolutions enhance transformers' ability to capture local spatial information, we introduce the Geometric Enhancer Module (GEM). This module is specifically designed to grasp the local geometric contexts within point clouds while enabling the preservation of the point encoder's foundational architecture and information integrity." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.67, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SegPoint" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.784, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.361, + 0.784, + 0.391 + ], + "angle": 0, + "content": "Fig. 3: Architecture of the proposed (b) Geometric Enhance Module (GEM) and (c) Geometric-guided Feature Propagation (GFP) interaction with (a) Point Encoder." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.407, + 0.788, + 0.543 + ], + "angle": 0, + "content": "As shown in Fig. 3, Geometric Enhancer Module \\(\\mathcal{G}\\) is composed of three blocks, each featuring a KPCConv [58] layer followed by BN and ReLU activation. The architecture is similar to the 2D convolutional stem [16]. We utilize KPCConv instead of vanilla convolution or linear layer here to facilitate grasping the local geometric information effectively. The resultant geometric feature, represented by \\(\\pmb{g}_f\\in \\mathbb{R}^{N\\times D}\\), contains the features across all points, thereby supplementing the missing local information. This \\(\\pmb{g}_f\\) is then leveraged to infuse geometric insights into the point encoder's features via a cross-attention mechanism, the above process can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.31, + 0.549, + 0.785, + 0.59 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {g} _ {f} = \\mathcal {G} (\\boldsymbol {i} _ {\\text {p o i n t}}), \\quad \\hat {\\boldsymbol {f}} _ {i} = \\boldsymbol {f} _ {i} + g _ {i} \\cdot \\operatorname {s o f t m a x} \\left(\\frac {\\boldsymbol {f} _ {i} \\boldsymbol {g} _ {f} ^ {T}}{\\sqrt {D}}\\right) \\boldsymbol {g} _ {f}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.596, + 0.788, + 0.746 + ], + "angle": 0, + "content": "where \\( \\pmb{f}_i \\) represents the feature from the \\( i \\)-th block of the point encoder and \\( l \\) consecutive transformer layers are regarded as one block for the convenience of explanation. To fine-tune the integration of geometric information, we introduce a learnable gating factor \\( g_i \\) that modulates the balance between the attention layer's output and the input feature \\( \\pmb{f}_i \\). This gating factor is initially set to zero, to ensure that the incorporation of geometric data does not abruptly alter \\( \\pmb{f} \\) feature distribution. Such an approach allows for the preservation and effective utilization of the point encoder's pre-trained weights. Upon processing through the Geometric Enhancer Module (GEM), the modified output of the point encoder, LLM are formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.264, + 0.753, + 0.787, + 0.773 + ], + "angle": 0, + "content": "\\[\n\\hat {\\pmb {f}} _ {p o i n t} = \\mathcal {E} (\\pmb {i} _ {p o i n t}, \\pmb {g} _ {f}), \\quad \\hat {\\pmb {y}} = \\mathcal {F} (\\hat {\\pmb {f}} _ {p o i n t}, \\pmb {f} _ {t x t}), \\quad \\hat {\\pmb {h}} _ {s e g} = \\gamma (\\hat {\\pmb {y}} _ {[ s e g ]}). \\qquad (4)\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.789, + 0.585, + 0.804 + ], + "angle": 0, + "content": "3.4 Geometric-guided Feature Propagation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Addressing the challenge of upsampling point clouds from a sparse set of \\( N_{1} \\) points to a denser set of \\( N \\) points is crucial, as direct upsampling inevitably intro-" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.348, + 0.127 + ], + "angle": 0, + "content": "S. He et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "duces noise and results in information loss, leading to sub-optimal performance in segmentation tasks. To mitigate these issues, we introduce Geometric-guided Feature Propagation designed to generate high-quality per-point embeddings. Geometric features \\( \\pmb{g}_f \\), which carry comprehensive point information, serving as a \"gold message\" for enhancing the upsampling process. By integrating these geometric features, we aim to significantly improve the quality and accuracy of the generated dense per-point embeddings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.788, + 0.389 + ], + "angle": 0, + "content": "As illustrated in Fig. 3, we begin by upsampling the higher-layer features \\( \\pmb{f}_3, \\pmb{f}_4 \\) from a smaller set of points \\( N_1 \\) to larger sets \\( N_3, N_2 \\), employing Point-Net++'s [46] propagation techniques. This step yields features \\( \\pmb{f}_3' \\in \\mathbb{R}^{N_3 \\times D} \\), and \\( \\pmb{f}_4' \\in \\mathbb{R}^{N_2 \\times D} \\). Subsequently, we perform downsampling on the geometric features \\( \\pmb{g}_f \\) from the original number of points \\( N \\) to reduced counts \\( N_2, N_3 \\), respectively, utilizing the Farthest Point Sampling (FPS) technique. In this process, we directly obtain the features of the sampled points without performing additional \\( k \\)-nearest neighbor (\\( k \\)-NN) and pooling operations to simplify the computation and produce features \\( \\pmb{f}_{g,1} \\in \\mathbb{R}^{N_3 \\times D} \\), and \\( \\pmb{f}_{g,2} \\in \\mathbb{R}^{N_2 \\times D} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.39, + 0.788, + 0.467 + ], + "angle": 0, + "content": "In the next phase, we integrate the up- and downsampled features, processing them through fully connected layers and ReLU activation to update the feature vectors \\(\\tilde{\\pmb{f}}_3\\in \\mathbb{R}^{N_3\\times D}\\), and \\(\\tilde{\\pmb{f}}_4\\in \\mathbb{R}^{N_2\\times D}\\). Note that the last layer feature \\(\\pmb{f}_{5}\\) bypasses this step. Instead, we concatenate it with \\(\\hat{\\pmb{h}}_{point}\\) from the LLM output to form \\(\\tilde{\\pmb{f}}_5\\) to perceive multi-modal information from LLM." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.468, + 0.788, + 0.573 + ], + "angle": 0, + "content": "Finally, to enable information exchange across different point densities, we propose attentive propagation. Take the propagation from \\(\\tilde{\\pmb{f}}_5\\) to \\(\\tilde{\\pmb{f}}_4\\) as example. Here, \\(\\tilde{\\pmb{f}}_4\\in \\mathbb{R}^{N_2\\times D}\\) acts as a set of local centers. For each local center within \\(\\tilde{\\pmb{f}}_4\\), we identify its neighboring points from \\(\\tilde{\\pmb{f}}_5\\) using the \\(k\\)-NN algorithm, resulting in \\(\\pmb{f}_{54}\\in \\mathbb{R}^{N_2\\times k\\times D}\\). Then, employing cross-attention mechanism, where \\(\\tilde{\\pmb{f}}_4\\) serves as query and \\(\\pmb{f}_{54}\\) as both key and value, facilitates information flow across different point densities and effectively extract relevant details into the query points." + }, + { + "type": "equation", + "bbox": [ + 0.381, + 0.582, + 0.787, + 0.622 + ], + "angle": 0, + "content": "\\[\n\\hat {\\tilde {\\boldsymbol {f}}} _ {4} = \\tilde {\\boldsymbol {f}} _ {4} + \\operatorname {s o f t m a x} \\left(\\frac {\\tilde {\\boldsymbol {f}} _ {4} \\boldsymbol {f} _ {5 4} ^ {T}}{\\sqrt {D}}\\right) \\boldsymbol {f} _ {5 4}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.787, + 0.675 + ], + "angle": 0, + "content": "Leveraging geometric-guided feature propagation enables us to produce high-quality per-point embeddings denoted as \\( \\pmb{f}_{\\mathcal{P}} \\), laying the foundation for generating precise segmentation masks expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.364, + 0.685, + 0.787, + 0.703 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {f} _ {\\mathcal {P}} = \\mathcal {P} \\left(\\hat {\\boldsymbol {f}} _ {\\text {p o i n t}}, \\boldsymbol {g} _ {f}\\right), \\quad \\hat {\\boldsymbol {m}} = \\hat {\\boldsymbol {h}} _ {\\text {s e g}} \\otimes \\boldsymbol {f} _ {\\mathcal {P}}. \\tag {6}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.721, + 0.427, + 0.737 + ], + "angle": 0, + "content": "3.5 Training Objectives" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.744, + 0.787, + 0.774 + ], + "angle": 0, + "content": "Our model is trained end-to-end leveraging the text classification loss and the segmentation mask loss:" + }, + { + "type": "equation", + "bbox": [ + 0.371, + 0.786, + 0.787, + 0.801 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda_ {t x t} \\mathcal {L} _ {t x t} + \\lambda_ {b c e} \\mathcal {L} _ {b c e} + \\lambda_ {d i c e} \\mathcal {L} _ {d i c e}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{txt}\\) denotes the auto-regressive cross-entropy loss targeting text generation accuracy, segmentation mask loss includes both binary cross-entropy (BCE)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.67, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SegPoint" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.194 + ], + "angle": 0, + "content": "loss \\(\\mathcal{L}_{bce}\\) and DICE loss \\(\\mathcal{L}_{dice}\\), aims at refining segmentation quality. The weights \\(\\lambda_{txt}\\), \\(\\lambda_{bce}\\) and \\(\\lambda_{dice}\\) are utilized to balance the different loss items. The model's training is guided by the ground-truth labels \\(\\pmb{y}_{txt}\\) for text and \\(\\pmb{M}\\) for masks." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.214, + 0.513, + 0.228 + ], + "angle": 0, + "content": "3.6 Instruct3D Dataset Collection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.787, + 0.344 + ], + "angle": 0, + "content": "Although 3D instruction segmentation and 3D referring segmentation [1,4,79] are both language-based segmentation, 3D referring segmentation guides segmentation with explicit target object names, e.g., \"chair\", lacking more complicated reasoning instructions, e.g., \"Where to sit in the room?\". Besides, they also fall short in offering multi-target question-answer pairs with target descriptions directly connected to multiple segmentation masks, which cannot meet a common requirement in real-world scenarios, like \"How to play computer games\"." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.344, + 0.789, + 0.586 + ], + "angle": 0, + "content": "To enhance the assessment and analysis of instruction segmentation capabilities, we have developed a benchmark, referred to Instruct3D. This benchmark incorporates 280 scenes specifically selected for instruction segmentation tuning and evaluation, sourced from the recently introduced ScanNet++ [74] dataset. Each scene comes with approximately 10 different segmentation instructions, resulting in 2,565 instruction-point cloud pairings. This dataset is then divided into two splits: train, and val, containing 2,052, and 513 question-answer pairs, respectively. Our dataset is uniquely designed to encompass both multi-target and zero-target scenarios, addressing the real-world requirement of identifying multiple objects in response to text queries and accounting for situations where objects mentioned in the text may not be present in the paired point cloud. Besides, we take into account the characteristics of 3D scenes and incorporate diverse locations and view descriptions e.g., \"something that is used for sitting while working at a desk. It is the one facing the window.\" The model needs to have not only reasoning capabilities but also the ability to perceive views and directions in 3D scenes. These designs underscore the dataset's practical value." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.608, + 0.377, + 0.625 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.639, + 0.533, + 0.653 + ], + "angle": 0, + "content": "4.1 Datasets and Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.787, + 0.735 + ], + "angle": 0, + "content": "Datasets. Our training data is composed of two types of datasets: (1) semantic segmentation dataset including ScanNet200 [53], and S3DIS [3]; (2) referring segmentation dataset consisting of ScanRefer [4], ReferIt3D [1](including Sr3D and Nr3D), and Multi3DRefer [79]. We design task-specific prompts to facilitate the joint training of various tasks within a unified framework." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Semantic Segmentation Dataset. We use two strategies to generate templates. 1) segment the specific category: \"USER: Can you segment the {category} category in this point cloud? ASSISTANT: {category} ,\" where category is the random chosen category, and denotes the placeholder for tokens of point cloud patches. 2) segment all the categories: \"USER: Can you segment all the semantic masks in this point cloud and output separate masks for each category in the alphabetical order of the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.348, + 0.127 + ], + "angle": 0, + "content": "S. He et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.783, + 0.22 + ], + "angle": 0, + "content": "categories? ASSISTANT: {category} , {category} , ...” To simplify the output and ensure it has only one possible answer, we add the constraints “in the alphabetical order of the categories”. To avoid generating class names not in the dataset, we incorporate category names in a dataset into the prompts during training and inference." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.222, + 0.784, + 0.326 + ], + "angle": 0, + "content": "Referring Segmentation Dataset. We use template prompts: \"USER: Can you segment the object {description} in this point cloud? ASSISTANT: {category} ,\" where {description} is the given explicit description from referring segmentation dataset. It is worth noting that during training, we also use other templates to generate the QA data to ensure data diversity. We add {category} in front of to unify the output format so that when outputting semantic masks, the output category name is the label it predicts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.328, + 0.784, + 0.402 + ], + "angle": 0, + "content": "Evaluation Metrics. We follow most previous works on 3D segmentation [22, 27, 54] to adopt mIoU as primary metric. mIoU is defined by the average of all per-point cloud scene Intersection-over-Unions (IoUs). Besides, we employ accuracy (Acc) as a metric to evaluate whether the model accurately identifies targets with which the predictions have an IoU greater than 0.5." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.412, + 0.456, + 0.426 + ], + "angle": 0, + "content": "4.2 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.433, + 0.785, + 0.628 + ], + "angle": 0, + "content": "In our experiments, unless specified otherwise, we employ the LLaMA2-7B model [59] as the large language model \\(\\mathcal{F}\\) and Uni3D [81] as the point cloud processing backbone \\(\\mathcal{E}\\). The training stage leverages the deepspeed [51] engine for efficiency, with the AdamW [39] optimizer guiding the learning process. The learning rate and weight decay are set to 0.0003 and 0, respectively, enhanced by a WarmupDecayLR learning rate scheduler that initiates with 100 warmup iterations. The projection layer \\(\\gamma\\) utilizes an MLP with channel sizes of [256, 4096, 4096]. We set balancing weight \\(\\lambda_{txt\\_gen}\\), \\(\\lambda_{bce}\\), and \\(\\lambda_{dice}\\) to 1.0, 2.0, 2.0, respectively. The experiments utilize a total batch size of 16, distributed across 4 NVIDIA 80G A100 GPUs, and span 5,000 iterations, culminating in a training period of approximately 3 days. During training, we use all mentioned datasets in Sec. 4.1 for joint training by leveraging task-specific prompts. For evaluation on a specific dataset, we finetune the trained model on the corresponding dataset." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.638, + 0.446, + 0.652 + ], + "angle": 0, + "content": "4.3 Results on Instruct3D" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.784, + 0.839 + ], + "angle": 0, + "content": "The instruction segmentation results, as detailed in Table 1, underscore a significant advancement: where existing methodologies fall short, our model demonstrates exceptional prowess, achieving a more than \\(15\\%\\) improvement in mIoU for tasks requiring intricate reasoning. Unlike conventional referring segmentation tasks, instruction segmentation demands not just identification but also understanding, necessitating the model's reasoning capabilities and access to world knowledge. Existing approaches, confined to explicit references, struggle with implicit queries due to their lack of understanding, which further underscores the task's inherent challenges. In contrast, our model leverages LLMs to bridge this gap, demonstrating superior performance by comprehending and interpreting the queries accurately. Moreover, SegPoint configuration substantially outperforms \\(\\mathrm{SegPoint}^{\\dagger}\\), highlighting the critical role of our designed Geometric" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.67, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SegPoint" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.2 + ], + "angle": 0, + "content": "Table 1: 3D instruction segmentation benchmark results on Instruct3D val split evaluated by Acc and mIoU. † denotes our vanilla baseline removing geometric enhancer module and geometric-guided feature propagation. * represents adding an auxiliary mask head through our implementation." + }, + { + "type": "table", + "bbox": [ + 0.236, + 0.203, + 0.769, + 0.309 + ], + "angle": 0, + "content": "
StageMethodReferenceAccmIoU
TwoScanRefer [4][ECCV'20]12.06.9
TwoReferIt3D [1][ECCV'20]11.76.4
TwoM3DRef-CLIP [79][ICCV'23]18.112.8
SingleTGNN [22][AAAI'21]12.97.1
SingleBUTD-DETR [23]*[ECCV'22]16.310.9
SingleEDA [67]*[CVPR'23]16.612.1
SingleSegPoint†[ECCV'24]21.816.1
SingleSegPoint[ECCV'24]31.627.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.321, + 0.785, + 0.361 + ], + "angle": 0, + "content": "Table 2: 3D Semantic segmentation benchmark results on S3DIS [3], ScanNet [7], and ScanNet200 [53]. We evaluate on the Area 5 of S3DIS and validation split of ScanNet and ScanNet200." + }, + { + "type": "table", + "bbox": [ + 0.23, + 0.364, + 0.772, + 0.47 + ], + "angle": 0, + "content": "
MethodReferenceScanNetScanNet200S3DIS
PointNet++ [46][NeurIPS'17]53.5--
MinkUNet [6][CVPR'19]72.225.065.4
PTv1 [80][ICCV'21]70.627.870.4
PTv2 [66][NeurIPS'22]75.430.271.6
PointNeXt [48][NeurIPS'22]71.5-70.5
OctFormer [60][SIGGRAPH'23]75.732.6-
Swin3D [72][ArXiv]75.5-72.5
SegPoint[ECCV'24]74.135.372.4
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.479, + 0.785, + 0.525 + ], + "angle": 0, + "content": "Enhancer Module and Geometric-guided Feature Propagation components. Notably, even in its baseline form, \\(\\mathrm{SegPoint}^{\\dagger}\\) surpasses all competing methods, validating the effectiveness and rationale behind our pipeline design." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.525, + 0.787, + 0.587 + ], + "angle": 0, + "content": "Besides, different from traditional two-stage approaches that first generate mask proposals using a pre-trained segmentor like Mask3D [54] and then apply language-aware networks for selection, SegPoint directly tackles the task, bypassing the need for preliminary mask proposals, enhancing its efficiency." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.592, + 0.548, + 0.607 + ], + "angle": 0, + "content": "4.4 Results on Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.788, + 0.72 + ], + "angle": 0, + "content": "Table 2 present SegPoint's performance on semantic segmentation, delivering competitive results across diverse datasets. Our model uses a simple yet effective answer format, category \\(<\\mathrm{SEG}>\\), to use category name as predicted labels, achieving particularly stronger performance in datasets with various categories like ScanNet200 [53], where it surpasses SOTA methods by \\(2.1\\%\\) mIoU. To ensure fair comparisons, we fine-tune our model on each semantic segmentation dataset to accommodate the varying class category definitions." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.726, + 0.553, + 0.741 + ], + "angle": 0, + "content": "4.5 Results on Referring Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.747, + 0.788, + 0.853 + ], + "angle": 0, + "content": "Table 3 presents results on referring segmentation datasets. SegPoint showcases outstanding performance in both single-target (e.g., ScanRefer [4], Nr3D [1]) and multi-target and zero-target contexts within the Multi3DRefer [79] dataset. For multi-targets, we aggregate masks into a single ground truth, and for zero-target, we use an empty mask, indicated by \"ASSISTANT: There is no mask.\" SegPoint significantly surpasses other approaches, achieving \\(2.5\\%\\) mIoU increase. The evaluation process of two-stage method M3DRef-CLIP is similar to Sec.4.3." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.348, + 0.127 + ], + "angle": 0, + "content": "S. He et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 3: 3D referring segmentation benchmark results on ScanRefer [4], Nr3D [1], and Multi3Drefer [79] evaluated by mIoU. * represents adding an auxiliary mask head through our implementation." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.189, + 0.789, + 0.294 + ], + "angle": 0, + "content": "
StageMethodReferenceScanReferNr3DMulti3DRefer
TwoM3DRef-CLIP [79][ICCV'23]35.727.032.6
Two3D-STMN [63][AAAI'24]39.5--
SingleTGNN [22][AAAI'21]27.8--
SingleBUTD-DETR [23]*[ECCV'22]35.427.526.2
SingleEDA [67]*[CVPR'23]36.229.328.9
SingleX-RefSeg3D [49][AAAI'24]29.9--
SingleRefMask3D [18][ACMMM'24]44.8--
SingleSegPoint[ECCV'24]41.732.236.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.308, + 0.787, + 0.336 + ], + "angle": 0, + "content": "Table 4: 3D open-vocabulary semantic segmentation benchmark results on val split of ScanNet++ [74]." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.338, + 0.773, + 0.442 + ], + "angle": 0, + "content": "
TypeMethodReferenceScanNet++
SupervisedPointNet [45][CVPR'17]7.0
PointNet++ [46][NeurIPS'17]15.0
MinkUNet [6][CVPR'19]28.0
KPCConv [58][ICCV'19]30.0
Open-VocabularyOpenScene [42][CVPR'23]12.8
PLA [12][CVPR'23]14.2
RegionPLC [71][CVPR'24]14.9
UnifiedSegPoint[ECCV'24]19.3
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.454, + 0.694, + 0.469 + ], + "angle": 0, + "content": "4.6 Results on Open-vocabulary Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.478, + 0.788, + 0.614 + ], + "angle": 0, + "content": "Table 4 shows our method's open-vocabulary segmentation performance which is directly evaluated on ScanNet++ [74] following the setting in prevalent methodologies in the 2D domain [13,69]. It demonstrates our superiority over both existing open-vocabulary techniques and even several supervised approaches, showing our model's robust generalization capabilities. It effectively aligns and interprets categories with visual scenes, underscoring the reasoning prowess of large language models. A notable issue is the potential misalignment between output categories of SegPoint and val split category names. To address this, we employ GPT-4 to match its most similar category names in the val split." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.62, + 0.39, + 0.636 + ], + "angle": 0, + "content": "4.7 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.748 + ], + "angle": 0, + "content": "We conduct extensive experiments to verify the effectiveness of our proposed components in Table 5 (a) on both Instruct3D and ScanRefer [4] dataset. We established a vanilla baseline as described in Sec. 3.2 following the paradigm of LISA [28], which cannot achieve satisfactory performance and only obtain \\(16.1\\% / 30.3 \\, \\mathrm{mIoU}\\) on Instruct3D/ScanRefer, respectively. Further analysis, both qualitative and quantitative, of our proposed components reveals that their integration substantially outperforms the baseline." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Geometric Enhancer Module Integrating the Geometric Enhancer Module (GEM) into our point encoder results in a notable \\(5.3\\% / 5.5\\%\\) mIoU improvement on Instruct3D/ScanRefer, effectively addressing compatibility issues with dense prediction tasks. An ablation study, shown in Table 5 (b), shows that our improvement is not due to an increase in parameters. Our approach outperforms traditional full fine-tuning, LoRA [21] strategies, and the addition of" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.67, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "SegPoint" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.302, + 0.145, + 0.699, + 0.159 + ], + "angle": 0, + "content": "Table 5: Ablation studies on Instruct3D and ScanRefer." + }, + { + "type": "table", + "bbox": [ + 0.235, + 0.171, + 0.482, + 0.244 + ], + "angle": 0, + "content": "
IndexComponentsInstruct3D\nmIoUScanRefer\nmIoU
GEMGFP
0XX16.130.3
1✓X21.435.8
2X✓23.238.1
3✓✓27.541.7
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.245, + 0.246, + 0.478, + 0.257 + ], + "angle": 0, + "content": "(a) Effect of different main components." + }, + { + "type": "table", + "bbox": [ + 0.538, + 0.176, + 0.747, + 0.239 + ], + "angle": 0, + "content": "
MethodInstruct3DScanRefer
Full tuning26.239.1
LoRA [21]24.838.5
MLP23.136.8
GEM27.541.7
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.535, + 0.239, + 0.757, + 0.252 + ], + "angle": 0, + "content": "(b) Effect of different tuning methods." + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.267, + 0.387, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.291, + 0.337, + 0.333, + 0.345 + ], + "angle": 0, + "content": "(a) Scene" + }, + { + "type": "image", + "bbox": [ + 0.418, + 0.271, + 0.586, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.485, + 0.337, + 0.536, + 0.345 + ], + "angle": 0, + "content": "(b) Baseline" + }, + { + "type": "image", + "bbox": [ + 0.618, + 0.271, + 0.784, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.692, + 0.337, + 0.728, + 0.345 + ], + "angle": 0, + "content": "(c) Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.352, + 0.788, + 0.407 + ], + "angle": 0, + "content": "Fig. 4: (Best viewed in color) We visualize the feature responses between a given point (in red) and other points in the scene from per-point embeddings \\( \\pmb{f}_{\\mathcal{P}} \\) for the baseline and our SegPoint, respectively. The color changes from yellow to red, indicating increasing feature similarity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.428, + 0.785, + 0.487 + ], + "angle": 0, + "content": "MLP layers for feature adapter, underscoring its effectiveness in embedding 3D domain-specific knowledge into point encoder. Unlike adapter techniques common in language and 2D image processing, our GEM is designed to address the unique challenges in dense prediction tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.485, + 0.785, + 0.561 + ], + "angle": 0, + "content": "Geometric-guided Feature Propagation Introducing Geometric-guided Feature Propagation (GFP) results in a substantial improvement over the baseline, as shown in Table 5 (a) (index 2). This underscores our method's capability to minimize information loss and reduce noise during the upsampling phase, leading to higher-quality per-point embeddings." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.571, + 0.467, + 0.586 + ], + "angle": 0, + "content": "4.8 Qualitative Visualization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.595, + 0.788, + 0.731 + ], + "angle": 0, + "content": "Fig. 4 qualitatively illustrates the feature responses between a given point and others in a scene. From left to right, it presents the original scene, baseline method, and our SegPoint, respectively, with warmer colors indicating closer feature relationships. The baseline method predominantly highlights spatially proximate points, often missing important but distant features, such as chair cushions. In contrast, SegPoint leverages global information, allowing for the identification of both near and distant relevant features, such as cushions and casters. This demonstrates our model's superior ability to capture global context and recognize intricate structures within the scene." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.732, + 0.788, + 0.853 + ], + "angle": 0, + "content": "As shown in Fig. 5, we provide some typical qualitative results from SegPoint on Instruct3D. Given an implicit instruction, e.g., \"the object that is used for dispensing water during bathing\", SegPoint successfully infers the role of a shower. In other scenes, the instructions include queries requiring extensive knowledge of the world and complex reasoning, like \"What appliance is used for heating or cooking food quickly using electromagnetic radiation.\" SegPoint can still segment the microwave oven very well, which shows that SegPoint can make good use of the reasoning ability of LLM and provide high-quality masks." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.348, + 0.127 + ], + "angle": 0, + "content": "S. He et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.147, + 0.787, + 0.509 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.516, + 0.788, + 0.56 + ], + "angle": 0, + "content": "Fig. 5: Qualitative results from val split of Instruct3D. SegPoint understand the human instruction and accurately segment the target object. We omitted the \"please output segmentation mask\" in the sentence for simplicity." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.567, + 0.36, + 0.584 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.788, + 0.72 + ], + "angle": 0, + "content": "In this work, we introduce SegPoint, an effective model supported by LLM for point-level reasoning and segmentation. Benefiting from the proposed Geometric Enhancer Module and Geometric-guided Feature Propagation, SegPoint is adept at solving a variety of segmentation tasks in a unified framework. Additionally, we construct a comprehensive Instruct3D benchmark to bolster research area in segmentation via implicit and complex instructions, introducing more challenges to promote it closer to real-world applications. Through thorough experiments, SegPoint achieves promising results across multiple benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.724, + 0.788, + 0.784 + ], + "angle": 0, + "content": "Limitations Although SegPoint demonstrates notable success in tasks driven by text prompts, its current framework cannot process non-textual prompts, such as boxes and points. Future developments will explore the adoption of a prompt encoder, inspired by SAM model [26], to extend support for these formats." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.785, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Acknowledgements We thank the anonymous reviewers for their constructive suggestions. Following their advice, we have incorporated diverse location and view descriptions into our Instruct3D. This work was partially supported by the National Research Foundation Singapore Competitive Research Program (CRP29-2022-0003)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.67, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SegPoint" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.324, + 0.161 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.218 + ], + "angle": 0, + "content": "1. Achlioptas, P., Abdelreheem, A., Xia, F., Elhoseiny, M., Guibas, L.: Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.219, + 0.785, + 0.259 + ], + "angle": 0, + "content": "2. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.26, + 0.785, + 0.287 + ], + "angle": 0, + "content": "3. Armeni, I., Sener, O., Zamir, A.R., Jiang, H., Brilakis, I., Fischer, M., Savarese, S.: 3d semantic parsing of large-scale indoor spaces. In: CVPR (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.288, + 0.785, + 0.315 + ], + "angle": 0, + "content": "4. Chen, D.Z., Chang, A.X., Nießner, M.: Scanrefer: 3d object localization in rgb-d scans using natural language. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.316, + 0.785, + 0.343 + ], + "angle": 0, + "content": "5. Chen, Z., Duan, Y., Wang, W., He, J., Lu, T., Dai, J., Qiao, Y.: Vision transformer adapter for dense predictions. In: ICLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.343, + 0.785, + 0.37 + ], + "angle": 0, + "content": "6. Choy, C., Gwak, J., Savarese, S.: 4d spatio-temporal convnets: Minkowski convolutional neural networks. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.371, + 0.785, + 0.398 + ], + "angle": 0, + "content": "7. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: CVPR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.399, + 0.785, + 0.425 + ], + "angle": 0, + "content": "8. Ding, H., Liu, C., He, S., Jiang, X., Loy, C.C.: MeViS: A large-scale benchmark for video segmentation with motion expressions. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.426, + 0.785, + 0.453 + ], + "angle": 0, + "content": "9. Ding, H., Liu, C., He, S., Jiang, X., Torr, P.H., Bai, S.: MOSE: A new dataset for video object segmentation in complex scenes. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.454, + 0.785, + 0.481 + ], + "angle": 0, + "content": "10. Ding, H., Liu, C., Wang, S., Jiang, X.: Vision-language transformer and query generation for referring segmentation. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.482, + 0.785, + 0.509 + ], + "angle": 0, + "content": "1. Ding, H., Liu, C., Wang, S., Jiang, X.: VLT: Vision-language transformer and query generation for referring segmentation. IEEE TPAMI (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.509, + 0.785, + 0.536 + ], + "angle": 0, + "content": "2. Ding, R., Yang, J., Xue, C., Zhang, W., Bai, S., Qi, X.: Pla: Language-driven open-vocabulary 3d scene understanding. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.537, + 0.785, + 0.563 + ], + "angle": 0, + "content": "3. Ding, Z., Wang, J., Tu, Z.: Open-vocabulary panoptic segmentation with maskclip. In: ICLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.565, + 0.785, + 0.592 + ], + "angle": 0, + "content": "4. Girdhar, R., El-Nouby, A., Liu, Z., Singh, M., Alwala, K.V., Joulin, A., Misra, I.: Imagebind: One embedding space to bind them all. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.592, + 0.785, + 0.647 + ], + "angle": 0, + "content": "5. Guo, Z., Zhang, R., Zhu, X., Tang, Y., Ma, X., Han, J., Chen, K., Gao, P., Li, X., Li, H., et al.: Point-bind & point-llm: Aligning point cloud with multi-modality for 3d understanding, generation, and instruction following. arXiv preprint arXiv:2309.00615 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.647, + 0.785, + 0.674 + ], + "angle": 0, + "content": "6. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.675, + 0.785, + 0.702 + ], + "angle": 0, + "content": "7. He, S., Ding, H.: Decoupling static and hierarchical motion perception for referring video segmentation. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.703, + 0.785, + 0.73 + ], + "angle": 0, + "content": "8. He, S., Ding, H.: RefMask3D: Language-guided transformer for 3d referring segmentation. In: ACM MM (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.73, + 0.785, + 0.757 + ], + "angle": 0, + "content": "9. He, S., Jiang, X., Jiang, W., Ding, H.: Prototype adaption and projection for few-and zero-shot 3d point cloud semantic segmentation. IEEE TIP (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.758, + 0.785, + 0.785 + ], + "angle": 0, + "content": "20. Hong, Y., Zhen, H., Chen, P., Zheng, S., Du, Y., Chen, Z., Gan, C.: 3d-llm: Injecting the 3d world into large language models. In: NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.786, + 0.785, + 0.813 + ], + "angle": 0, + "content": "21. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. In: ICLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "22. Huang, P.H., Lee, H.H., Chen, H.T., Liu, T.L.: Text-guided graph neural networks for referring 3d instance segmentation. In: AAAI (2021)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.348, + 0.127 + ], + "angle": 0, + "content": "S. He et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.188 + ], + "angle": 0, + "content": "23. Jain, A., Gkanatsios, N., Mediratta, I., Fragkiadaki, K.: Bottom up top down detection transformers for language grounding in images and point clouds. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.189, + 0.785, + 0.229 + ], + "angle": 0, + "content": "24. Jia, B., Chen, Y., Yu, H., Wang, Y., Niu, X., Liu, T., Li, Q., Huang, S.: Sceneverse: Scaling 3d vision-language learning for grounded scene understanding. In: ECCV (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.231, + 0.785, + 0.257 + ], + "angle": 0, + "content": "25. Jiang, L., Zhao, H., Shi, S., Liu, S., Fu, C.W., Jia, J.: Pointgroup: Dual-set point grouping for 3d instance segmentation. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.258, + 0.785, + 0.284 + ], + "angle": 0, + "content": "26. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.285, + 0.785, + 0.311 + ], + "angle": 0, + "content": "27. Kolodiazhnyi, M., Vorontsova, A., Konushin, A., Rukhovich, D.: Oneformer3d: One transformer for unified point cloud segmentation. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.312, + 0.785, + 0.339 + ], + "angle": 0, + "content": "28. Lai, X., Tian, Z., Chen, Y., Li, Y., Yuan, Y., Liu, S., Jia, J.: Lisa: Reasoning segmentation via large language model. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.34, + 0.785, + 0.366 + ], + "angle": 0, + "content": "29. Lai, X., Yuan, Y., Chu, R., Chen, Y., Hu, H., Jia, J.: Mask-attention-free transformer for 3d instance segmentation. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.367, + 0.785, + 0.393 + ], + "angle": 0, + "content": "30. Li, B., Zhang, Y., Chen, L., Wang, J., Yang, J., Liu, Z.: Otter: A multi-modal model with in-context instruction tuning. arXiv:2305.03726 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.394, + 0.785, + 0.42 + ], + "angle": 0, + "content": "31. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. In: NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.421, + 0.785, + 0.461 + ], + "angle": 0, + "content": "32. Lin, Z., Liu, C., Zhang, R., Gao, P., Qiu, L., Xiao, H., Qiu, H., Lin, C., Shao, W., Chen, K., et al.: Sphinx: The joint mixing of weights, tasks, and visual embeddings for multi-modal large language models. arXiv preprint arXiv:2311.07575 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.462, + 0.785, + 0.488 + ], + "angle": 0, + "content": "33. Liu, C., Ding, H., Jiang, X.: GRES: Generalized referring expression segmentation. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.489, + 0.785, + 0.515 + ], + "angle": 0, + "content": "34. Liu, C., Ding, H., Zhang, Y., Jiang, X.: Multi-modal mutual attention and iterative interaction for referring image segmentation. IEEE TIP (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.516, + 0.785, + 0.542 + ], + "angle": 0, + "content": "35. Liu, C., Jiang, X., Ding, H.: Instance-specific feature propagation for referring segmentation. IEEE TMM (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.543, + 0.785, + 0.569 + ], + "angle": 0, + "content": "36. Liu, C., Jiang, X., Ding, H.: Primitivenet: decomposing the global constraints for referring segmentation. Visual Intelligence 2(1), 16 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.57, + 0.776, + 0.583 + ], + "angle": 0, + "content": "37. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.584, + 0.785, + 0.623 + ], + "angle": 0, + "content": "38. Liu, Y., Kong, L., Cen, J., Chen, R., Zhang, W., Pan, L., Chen, K., Liu, Z.: Segment any point cloud sequences by distilling vision foundation models. In: NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.624, + 0.785, + 0.637 + ], + "angle": 0, + "content": "39. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: ICLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.638, + 0.785, + 0.677 + ], + "angle": 0, + "content": "40. Nguyen, P.D., Ngo, T.D., Gan, C., Kalogerakis, E., Tran, A., Pham, C., Nguyen, K.: Open3dis: Open-vocabulary 3d instance segmentation with 2d mask guidance. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.678, + 0.785, + 0.704 + ], + "angle": 0, + "content": "41. Park, N., Kim, S.: How do vision transformers work? arXiv preprint arXiv:2202.06709 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.705, + 0.785, + 0.731 + ], + "angle": 0, + "content": "42. Peng, S., Genova, K., Jiang, C., Tagliasacchi, A., Pollefeys, M., Funkhouser, T., et al.: Openscene: 3d scene understanding with open vocabularies. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.732, + 0.785, + 0.759 + ], + "angle": 0, + "content": "43. Peng, Z., Wang, W., Dong, L., Hao, Y., Huang, S., Ma, S., Wei, F.: Kosmos-2: Grounding multimodal large language models to the world. In: ICLR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.76, + 0.785, + 0.785 + ], + "angle": 0, + "content": "44. Pi, R., Gao, J., Diao, S., Pan, R., Dong, H., Zhang, J., Yao, L., Han, J., Xu, H., Zhang, L.K.T.: Detgpt: Detect what you need via reasoning. In: EMNLP (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.785, + 0.813 + ], + "angle": 0, + "content": "45. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: CVPR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "46. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In: NeurIPS (2017)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.67, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "SegPoint" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "47. Qi, Z., Fang, Y., Sun, Z., Wu, X., Wu, T., Wang, J., Lin, D., Zhao, H.: Gpt4point: A unified framework for point-language understanding and generation. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "48. Qian, G., Li, Y., Peng, H., Mai, J., Hammoud, H., Elhoseiny, M., Ghanem, B.: Pointnext: Revisiting pointnet++ with improved training and scaling strategies. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.232, + 0.788, + 0.259 + ], + "angle": 0, + "content": "49. Qian, Z., Ma, Y., Ji, J., Sun, X.: X-refseg3d: Enhancing referring 3d instance segmentation via structured cross-modal graph neural networks. In: AAAI (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.26, + 0.788, + 0.3 + ], + "angle": 0, + "content": "50. Rasheed, H., Maaz, M., Shaji, S., Shaker, A., Khan, S., Cholakkal, H., Anwer, R.M., Xing, E., Yang, M.H., Khan, F.S.: Glamm: Pixel grounding large multimodal model. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.3, + 0.788, + 0.342 + ], + "angle": 0, + "content": "51. Rasley, J., Rajbhandari, S., Ruwase, O., He, Y.: Deepspeed: System optimizations enable training deep learning models with over 100 billion parameters. In: KDD (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.342, + 0.788, + 0.37 + ], + "angle": 0, + "content": "52. Ren, Z., Huang, Z., Wei, Y., Zhao, Y., Fu, D., Feng, J., Jin, X.: Pixelm: Pixel reasoning with large multimodal model. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.371, + 0.788, + 0.397 + ], + "angle": 0, + "content": "53. Rozenberszki, D., Litany, O., Dai, A.: Language-grounded indoor 3d semantic segmentation in the wild. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.398, + 0.788, + 0.424 + ], + "angle": 0, + "content": "54. Schult, J., Engelmann, F., Hermans, A., Litany, O., Tang, S., Leibe, B.: Mask3d: Mask transformer for 3d semantic instance segmentation. In: ICRA (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.425, + 0.788, + 0.467 + ], + "angle": 0, + "content": "55. Shuai, X., Ding, H., Ma, X., Tu, R., Jiang, Y.G., Tao, D.: A survey of multimodal-guided image editing with text-to-image diffusion models. arXiv preprint arXiv:2406.14555 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.467, + 0.788, + 0.494 + ], + "angle": 0, + "content": "56. Sun, J., Qing, C., Tan, J., Xu, X.: Superpoint transformer for 3d scene instance segmentation. In: AAAI (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.495, + 0.788, + 0.522 + ], + "angle": 0, + "content": "57. Takmaz, A., Fedele, E., Sumner, R.W., Pollefeys, M., Tombari, F., Engelmann, F.: Openmask3d: Open-vocabulary 3d instance segmentation. In: NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.522, + 0.788, + 0.55 + ], + "angle": 0, + "content": "58. Thomas, H., Qi, C.R., Deschaud, J.E., Marcotegui, B., Goulette, F., Guibas, L.J.: Kpconv: Flexible and deformable convolution for point clouds. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.55, + 0.788, + 0.591 + ], + "angle": 0, + "content": "59. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.591, + 0.788, + 0.619 + ], + "angle": 0, + "content": "60. Wang, P.S.: Octformer: Octree-based transformers for 3d point clouds. In: SIGGRAPH (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.619, + 0.788, + 0.66 + ], + "angle": 0, + "content": "61. Wang, W., Chen, Z., Chen, X., Wu, J., Zhu, X., Zeng, G., Luo, P., Lu, T., Zhou, J., Qiao, Y., et al.: Visionllm: Large language model is also an open-ended decoder for vision-centric tasks. In: NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.66, + 0.788, + 0.702 + ], + "angle": 0, + "content": "62. Wang, W., Xie, E., Li, X., Fan, D.P., Song, K., Liang, D., Lu, T., Luo, P., Shao, L.: Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.702, + 0.788, + 0.743 + ], + "angle": 0, + "content": "63. Wu, C., Ma, Y., Chen, Q., Wang, H., Luo, G., Ji, J., Sun, X.: 3d-stmn: Dependency-driven superpoint-text matching network for end-to-end 3d referring expression segmentation. In: AAAI (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.743, + 0.788, + 0.771 + ], + "angle": 0, + "content": "64. Wu, H., Xiao, B., Codella, N., Liu, M., Dai, X., Yuan, L., Zhang, L.: Cvt: Introducing convolutions to vision transformers. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.771, + 0.788, + 0.813 + ], + "angle": 0, + "content": "65. Wu, J., Li, X., Xu, S., Yuan, H., Ding, H., Yang, Y., Li, X., Zhang, J., Tong, Y., Jiang, X., Ghanem, B., Tao, D.: Towards open vocabulary learning: A survey. IEEE TPAMI (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.813, + 0.788, + 0.841 + ], + "angle": 0, + "content": "66. Wu, X., Lao, Y., Jiang, L., Liu, X., Zhao, H.: Point transformer v2: Grouped vector attention and partition-based pooling. In: NeurIPS (2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.348, + 0.127 + ], + "angle": 0, + "content": "S. He et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "67. Wu, Y., Cheng, X., Zhang, R., Cheng, Z., Zhang, J.: Eda: Explicit text-decoupling and dense alignment for 3d visual grounding. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.203 + ], + "angle": 0, + "content": "68. Xiao, Z., Zhang, W., Wang, T., Loy, C.C., Lin, D., Pang, J.: Position-guided point cloud panoptic segmentation transformer. arXiv preprint arXiv:2303.13509 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.785, + 0.231 + ], + "angle": 0, + "content": "69. Xu, J., Liu, S., Vahdat, A., Byeon, W., Wang, X., De Mello, S.: Open-vocabulary panoptic segmentation with text-to-image diffusion models. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.232, + 0.785, + 0.259 + ], + "angle": 0, + "content": "70. Xu, R., Wang, X., Wang, T., Chen, Y., Pang, J., Lin, D.: Pointllm: Empowering large language models to understand point clouds. In: ECCV (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.26, + 0.785, + 0.286 + ], + "angle": 0, + "content": "71. Yang, J., Ding, R., Deng, W., Wang, Z., Qi, X.: Regionplc: Regional point-language contrastive learning for open-world 3d scene understanding. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.287, + 0.785, + 0.327 + ], + "angle": 0, + "content": "72. Yang, Y.Q., Guo, Y.X., Xiong, J.Y., Liu, Y., Pan, H., Wang, P.S., Tong, X., Guo, B.: Swin3d: A pretrained transformer backbone for 3d indoor scene understanding. arXiv preprint arXiv:2304.06906 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.328, + 0.785, + 0.369 + ], + "angle": 0, + "content": "73. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: MPLug-owl: Modularization empowers large language models with multimodality. arXiv:2304.14178 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.37, + 0.785, + 0.397 + ], + "angle": 0, + "content": "74. Yeshwanth, C., Liu, Y.C., Nießner, M., Dai, A.: Scannet++: A high-fidelity dataset of 3d indoor scenes. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.398, + 0.785, + 0.438 + ], + "angle": 0, + "content": "75. You, H., Zhang, H., Gan, Z., Du, X., Zhang, B., Wang, Z., Cao, L., Chang, S.F., Yang, Y.: Ferret: Refer and ground anything anywhere at any granularity. In: ICLR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.439, + 0.785, + 0.481 + ], + "angle": 0, + "content": "76. Zhang, H., Li, H., Li, F., Ren, T., Zou, X., Liu, S., Huang, S., Gao, J., Zhang, L., Li, C., et al.: Llava-grounding: Grounded visual chat with large multimodal models. arXiv preprint arXiv:2312.02949 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.481, + 0.785, + 0.508 + ], + "angle": 0, + "content": "77. Zhang, H., Ding, H.: Prototypical matching and open set rejection for zero-shot semantic segmentation. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.509, + 0.785, + 0.55 + ], + "angle": 0, + "content": "78. Zhang, S., Sun, P., Chen, S., Xiao, M., Shao, W., Zhang, W., Chen, K., Luo, P.: Gpt4roi: Instruction tuning large language model on region-of-interest. arXiv:2307.03601 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.55, + 0.785, + 0.577 + ], + "angle": 0, + "content": "79. Zhang, Y., Gong, Z., Chang, A.X.: Multi3defer: Grounding text description to multiple 3d objects. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.578, + 0.785, + 0.604 + ], + "angle": 0, + "content": "80. Zhao, H., Jiang, L., Jia, J., Torr, P.H., Koltun, V.: Point transformer. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.606, + 0.785, + 0.632 + ], + "angle": 0, + "content": "81. Zhou, J., Wang, J., Ma, B., Liu, Y.S., Huang, T., Wang, X.: Uni3d: Exploring unified 3d representation at scale. In: ICLR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.633, + 0.785, + 0.66 + ], + "angle": 0, + "content": "82. Zhou, Z., Zhang, Y., Foroosh, H.: Panoptic-polarnet: Proposal-free lidar point cloud panoptic segmentation. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.661, + 0.785, + 0.688 + ], + "angle": 0, + "content": "83. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. In: ICLR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.689, + 0.785, + 0.715 + ], + "angle": 0, + "content": "84. Zhu, Z., Ma, X., Chen, Y., Deng, Z., Huang, S., Li, Q.: 3d-vista: Pre-trained transformer for 3d vision and text alignment. In: ICCV (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.715 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/SegPoint_ Segment Any Point Cloud via Large Language Model/d283f64a-a017-47de-9be3-9206675b2c4f_origin.pdf b/2024/SegPoint_ Segment Any Point Cloud via Large Language Model/d283f64a-a017-47de-9be3-9206675b2c4f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..06f950750e328f0f53ca5ce14951e62923623756 --- /dev/null +++ b/2024/SegPoint_ Segment Any Point Cloud via Large Language Model/d283f64a-a017-47de-9be3-9206675b2c4f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:97e596b58e7f1c1ef2949568b34bef3750f3ada355fde0cb9f250bee16b1c488 +size 9821230 diff --git a/2024/SegPoint_ Segment Any Point Cloud via Large Language Model/full.md b/2024/SegPoint_ Segment Any Point Cloud via Large Language Model/full.md new file mode 100644 index 0000000000000000000000000000000000000000..184a3002ffbcfbf3011c092a5d1bb1f7645748cc --- /dev/null +++ b/2024/SegPoint_ Segment Any Point Cloud via Large Language Model/full.md @@ -0,0 +1,341 @@ +# SegPoint: Segment Any Point Cloud via Large Language Model + +Shuting $\mathrm{He}^{1}$ , Henghui Ding2, Xudong Jiang1, and Bihan $\mathrm{Wen}^{1}$ + +1 Nanyang Technological University + +$^{2}$ Institute of Big Data, Fudan University + +{heshuting555, henghui.ding}@gmail.com, {exdjiang, bihan.wen}@ntu.edu.sg + +https://heshuting555.github.io/SegPoint + +Abstract. Despite significant progress in 3D point cloud segmentation, existing methods primarily address specific tasks and depend on explicit instructions to identify targets, lacking the capability to infer and understand implicit user intentions in a unified framework. In this work, we propose a model, called SegPoint, that leverages the reasoning capabilities of a multi-modal Large Language Model (LLM) to produce point-wise segmentation masks across a diverse range of tasks: 1) 3D instruction segmentation, 2) 3D referring segmentation, 3) 3D semantic segmentation, and 4) 3D open-vocabulary semantic segmentation. To advance 3D instruction research, we introduce a new benchmark, Instruct3D, designed to evaluate segmentation performance from complex and implicit instructional texts, featuring 2,565 point cloud-instruction pairs. Our experimental results demonstrate that SegPoint achieves competitive performance on established benchmarks such as ScanRefer for referring segmentation and ScanNet for semantic segmentation, while delivering outstanding outcomes on the Instruct3D dataset. To our knowledge, SegPoint is the first model to address these varied segmentation tasks within a single framework, achieving satisfactory performance. + +Keywords: Instruct3D dataset $\cdot$ Unified framework $\cdot$ 3D point cloud segmentation $\cdot$ Large language model + +# 1 Introduction + +3D point cloud segmentation, a critical challenge in the 3D vision community, aims to interpret and classify each point in a point cloud to understand its semantic properties [27, 45, 46, 54, 56, 80]. This longstanding issue has spurred significant advancements across various fields, including robotics, autonomous driving, virtual reality, etc. This challenge has evolved into a series of specialized tasks, each targeting a specific segmentation aspect. Overall, tasks cover basic semantic and instance segmentation [3, 7, 53, 74], as well as more practical tasks + +![](images/83c779324b564b4662467dc72345b67eba7c11560d8f0f52d90769a1aef47e55.jpg) +Fig. 1: Example of functionality in SegPoint. SegPoint can complete various point cloud tasks in a unified framework by leveraging task-specific prompts, including 1) 3D instruction segmentation, 2) 3D referring segmentation, 3) 3D semantic segmentation, and 4) 3D open-vocabulary semantic segmentation. + +such as referring segmentation [1, 4, 24, 33, 79, 84], which segments points based on explicit textual descriptions, and open-vocabulary segmentation [12, 40, 42, 57, 65, 77] designed for the dynamic and complex nature of real-world. + +Despite significant progress achieved within the 3D community toward accurately segmenting objects through specifically designed models, each model is typically developed to tackle one specific segmentation task, leading to inefficiencies and a lack of versatility for real-world application. Furthermore, previous perception approaches heavily depend on predefined categories or explicit expressions for language understanding. Such approaches fall short in interpreting and acting on implicit instructions often found in human language, a critical gap that hinders the development of truly intelligent next-generation perception systems. This brings a pivotal question: Is it possible to design a unified model capable of comprehensively addressing all aforementioned 3D tasks with human-like instructions? The exploration of this question not only challenges the current paradigms of 3D point cloud segmentation but also opens the door to groundbreaking advancements in robotic perception and interaction. + +In this work, we propose a model called SegPoint, leveraging the Large Language Model's (LLM) advanced ability to reason and comprehend user instructions. To enhance 3D scene comprehension, we integrate a Geometric Enhancer + +Module that extracts local semantics from point clouds, seamlessly incorporating this geometric insight into the feature extraction process. Furthermore, a Geometric-guided Feature Propagation is designed to utilize semantic flows derived from geometric priors and the LLM's hidden embeddings, facilitating the generation of fine-grained, high-quality features for accurate dense prediction tasks. Unlike previous attempts at 2D field [8, 28, 50], we do not depend on additional costly pre-trained segmentation models like SAM [26]. + +Moreover, we introduce a benchmark named Instruct3D, designed to advance research in the field of segmentation driven by implicit and complex instructions. Understanding these nuanced instructions necessitates reasoning abilities and extensive knowledge of the world. It includes a total of 2,565 diverse pairs of instructions and point clouds for tuning and evaluation. Our comprehensive experiments demonstrate the benchmark's utility in evaluating the model's capability of segmentation based on human-like instructions. + +Taking advantage of a multi-modal LLM and task-specific prompts, SegPoint is capable of generating segmentation masks for a wide range of tasks in a unified model: 1) 3D instruction segmentation, 2) 3D referring segmentation, 3) 3D semantic segmentation, and 4) 3D open-vocabulary semantic segmentation, as depicted in Fig. 1. SegPoint achieves competitive results on established benchmarks like ScanRefer [4] for referring segmentation and ScanNet [7] for semantic segmentation while showing remarkable performance on the Instruct3D dataset. + +In summary, our main contributions are as follows: + +- We propose SegPoint, the first 3D segmentation model that can comprehend human intentions and address multiple segmentation tasks within one framework by harnessing the Large Language Model's reasoning capabilities. + +- We present a Geometric Enhancer Module that integrates comprehensive scene information into the process of 3D scene understanding. Besides, A Geometric-guided Feature Propagation is designed to achieve accurate and fine-grained segmentation. These two modules supplement the missing local information and grasp fine-grained features for dense prediction tasks. + +- We introduce a new task called 3D instruction segmentation and construct a new dataset Instruct3D, which necessitates a model's self-reasoning to interpret implicit instructions for segmenting the target object. + +- Our experimental findings reveal that SegPoint not only competes strongly in 3D semantic, referring, and open-vocabulary semantic segmentation but also excels in 3D instruction segmentation, showcasing its versatility and effectiveness across a spectrum of segmentation challenges. + +# 2 Related Work + +# 2.1 Multi-modal Large Language Model + +Inspired by the exceptional reasoning abilities of Large Language Models, researchers are delving into transferring these capabilities into the vision realm [9, + +28, 36, 55], developing multi-modal LLMs. Flamingo [2], BLIP-2 [31], mPLUGOWL [73], Otter [30], LLaVA [37] and MiniGPT-4 [83] first construct image-text feature alignment followed by instruction tuning and achieve superior performance. Recent studies have increasingly concentrated on integrating foundational models with tasks that demand a refined understanding at the region or pixel level. VisionLLM [61] introduces a versatile interaction interface for various vision-centric tasks via instruction tuning, albeit without fully leveraging LLMs for intricate reasoning tasks. Kosmos-2 [43] has built substantial data of grounded image-text pairs, thereby embedding grounding capabilities into LLMs. DetGPT [44] links the fixed multi-modal LLM with an open-vocabulary detector, facilitating user instruction-based detection tasks. This growing interest has spurred further innovations, including GPT4RoI [78], LLaVA-grounding [76], Ferret [75], LISA [28], GlaMM [50], PixelLM [52], Sphinx [32], each contributing to the evolving landscape of instruction-based, multi-modal understanding. + +Building on advancements of multi-modal large language models in 2D image domain, the field is witnessing a seamless transition into 3D spaces. PointLLM [70] harnesses the prowess of large language models and trains with 3D point clouds following the paradigm of LLaVA [37]. 3D-LLM [20] utilizes 2D foundation models to encode multi-view images of 3D point clouds. Point-Bind [15] aligns point clouds with Image-Bind [14] and leverages ImageBind-LLM to reason multimodality input without 3D-instruction data training. GPT4Point [47] pioneers in facilitating a unified approach towards 3D object understanding and generation, setting a new standard for versatility. However, these models primarily concentrate on scene-level insights, often overlooking the intricate details at the region, or point level. Contrasting with existing models, our research focuses on two pivotal goals: 1) efficiently inject segmentation capabilities into multi-modal LLMs to conduct point-level understanding and 2) design a unified framework for 3D point cloud segmentation via the reasoning ability of LLMs. + +# 2.2 3D Point Cloud Segmentation + +3D point cloud segmentation, a crucial task in computer vision, can be categorized into 3D semantic, instance, and panoptic segmentation. 3D semantic segmentation [3,7,19,45,53,74,80] assigns each point in a 3D space to specific, predefined classes. In contrast, 3D instance segmentation [25,27,54] goes a step further by classifying each point and differentiating between distinct objects of the same class. 3D panoptic segmentation [68,82] aims to group 3D points according to their semantics and identities. Lately, more practical tasks have emerged, such as 3D referring segmentation [1,4,17,23,67,79], which extends referring expression segmentation [8,10,11,33-35] to 3D and segments a target instance based on explicit linguistic descriptions, and 3D open-vocabulary segmentation [12,40,42,57], designed to identify and segment unseen objects beyond a fixed set of known categories. + +Despite the significant progress made by transformer-based models: Mask3D [54], SPFormer [56], MAFT [29], and OneFormer3D [27] in basic 3D segmentation tasks, their application in real-world scenarios requiring human language + +![](images/59e9162900bb37310cabfb6aa72651bec476a3448dd6d46a0ab4266b49ffe291.jpg) +Fig. 2: The pipeline of SegPoint. Given input point cloud and text query, the multimodal LLM $\mathcal{F}$ generates text output. Geometric Enhancer Module $\mathcal{G}$ injects geometric information into Point Encoder $\mathcal{E}$ and obtains point features $\hat{\pmb{f}}_{\text {point }}$ . Per-point embeddings $\pmb{f}_{\mathcal{P}}$ derived from Geometric-guided Feature Propagation $\mathcal{P}$ multiplied with the embedding associated with the token yield the final segmentation masks. + +interaction remains constrained. Besides, Seal [38] aims to segment any point cloud through distilling vision foundation models, while it doesn't use language as cues. TGNN [22] is the first work to tackle referring segmentation problem that proposes aggregating textual features by considering the neighboring local structure of each instance but it heavily depends on explicit expressions or predefined categories for language understanding. Furthermore, the development of models tailored to specific segmentation tasks restricts their versatility and applicability in diverse real-world scenarios. Therefore, it is imperative to develop more intelligent interaction ways and a unified model for 3D point segmentation. + +# 3 Approach + +# 3.1 Architecture Overview + +The overall architecture of SegPoint is presented in Fig. 2. SegPoint mainly comprises four parts: $i$ ) a pre-trained point encoder $\mathcal{E}$ tailored for aligning with textual data; $ii$ ) a large language model $\mathcal{F}$ endowed with advanced reasoning capabilities; $iii$ ) a Geometric Enhancer Module $\mathcal{G}$ responsible for extracting geometric representation from input point clouds and infusing these priors into the point encoder; and $iv$ ) a Geometric-guided Feature Propagation $\mathcal{P}$ which is key to achieving precise mask generation. The collaboration between the Geometric Enhancer Module and Geometric-guided Feature Propagation is crucial, as it equips LLMs with the ability to generate masks effectively in various scenarios. + +# 3.2 Vanilla Baseline + +The input of the framework is the text instructions $\pmb{i}_{\text{txt}}$ and point cloud $\pmb{i}_{\text{point}} \in \mathbb{R}^{N \times (3 + F)}$ . Specifically, a point cloud scene, comprising $N$ points, each includes + +3D coordinates $\in \mathbb{R}^3$ and an auxiliary feature vector $\in \mathbb{R}^F$ (e.g., color). The point cloud $i_{point}$ is fed into the point encoder $\mathcal{E}$ , which extracts point features $f_{point} \in \mathbb{R}^{N_1 \times D}$ , where $N_1 \ll N$ , $D$ is the feature dimension. Concurrently, the text instruction $i_{txt}$ undergoes tokenization via $\mathcal{F}_{tokenize}$ . These prepared inputs are then fed into the Large Language Model $\mathcal{F}$ , resulting in a textual response $\pmb{y}$ . The above process can be formulated as: + +$$ +\boldsymbol {f} _ {\text {p o i n t}} = \mathcal {E} (\boldsymbol {i} _ {\text {p o i n t}}), \quad \boldsymbol {f} _ {\text {t x t}} = \mathcal {F} _ {\text {t o k e n i z e}} (\boldsymbol {i} _ {\text {t x t}}), \quad \boldsymbol {y} = \mathcal {F} (\boldsymbol {f} _ {\text {p o i n t}}, \boldsymbol {f} _ {\text {t x t}}). \tag {1} +$$ + +Building on the approach introduced by LISA [28], SegPoint enhances the segmentation capabilities of Large Language Models (LLMs) by expanding their vocabulary with a new special token, $<\mathrm{SEG}>.$ This modification enables the model to recognize and predict the $<\mathrm{SEG}>$ token within the output sequence as a signal to identify segmentation targets. Upon detecting a $<\mathrm{SEG}>$ token, the corresponding output sequence belonging to $<\mathrm{SEG}>$ token is extracted and processed through an MLP layer $\gamma$ , generating mask embeddings $h_{seg}$ . The final step involves computing each binary mask prediction $\pmb{m} \in \mathbb{R}^{N}$ by performing a dot product between the mask embeddings $h_{seg}$ and the upsampled per-point embeddings derived from the point features $\pmb{f}_{point}$ . The formulation of the aforementioned process is given by: + +$$ +\boldsymbol {h} _ {\text {s e g}} = \gamma (\boldsymbol {y} _ {[ \text {s e g} ]}), \quad \boldsymbol {m} = \boldsymbol {h} _ {\text {s e g}} \otimes U p S. (\boldsymbol {f} _ {\text {p o i n t}}), \tag {2} +$$ + +where $UpS$ denotes the upsampling operation following PointNet++ [46] on $f_{point}$ . The vanilla baseline represents an initial attempt to bridge the gap between LLMs' text comprehension and point cloud segmentation tasks. It encounters two primary issues. Firstly, the point encoder is trained on a scene-level dataset for classification achieving alignment between text and point clouds, not specifically trained for dense prediction tasks. Besides, the point encoder's first layer employs Farthest Point Sampling (FPS) [46] to reduce the point cloud to $N_1$ points, risking the loss of details vital for accurate dense predictions. Secondly, the operation of directly upsampling from $N_1$ to $N$ points to obtain per-point embeddings is prone to losing structural information and introducing a notable degree of noise, undermining the model's efficacy in segmentation tasks. + +# 3.3 Geometric Enhancer Module + +To adapt the pre-trained point encoder for dense prediction tasks while maintaining its superior scene recognition capability, our objective is to harness the geometric information across the entire scene to guide the further feature learning process. Drawing inspiration from recent advancements in 2D computer vision, where studies [5,41,62,64] demonstrate that convolutions enhance transformers' ability to capture local spatial information, we introduce the Geometric Enhancer Module (GEM). This module is specifically designed to grasp the local geometric contexts within point clouds while enabling the preservation of the point encoder's foundational architecture and information integrity. + +![](images/084d27946842b10c23a8f76c1674d672fa1b8c092118cfb564c8430982cad1fe.jpg) +Fig. 3: Architecture of the proposed (b) Geometric Enhance Module (GEM) and (c) Geometric-guided Feature Propagation (GFP) interaction with (a) Point Encoder. + +As shown in Fig. 3, Geometric Enhancer Module $\mathcal{G}$ is composed of three blocks, each featuring a KPCConv [58] layer followed by BN and ReLU activation. The architecture is similar to the 2D convolutional stem [16]. We utilize KPCConv instead of vanilla convolution or linear layer here to facilitate grasping the local geometric information effectively. The resultant geometric feature, represented by $\pmb{g}_f\in \mathbb{R}^{N\times D}$ , contains the features across all points, thereby supplementing the missing local information. This $\pmb{g}_f$ is then leveraged to infuse geometric insights into the point encoder's features via a cross-attention mechanism, the above process can be expressed as: + +$$ +\boldsymbol {g} _ {f} = \mathcal {G} (\boldsymbol {i} _ {\text {p o i n t}}), \quad \hat {\boldsymbol {f}} _ {i} = \boldsymbol {f} _ {i} + g _ {i} \cdot \operatorname {s o f t m a x} \left(\frac {\boldsymbol {f} _ {i} \boldsymbol {g} _ {f} ^ {T}}{\sqrt {D}}\right) \boldsymbol {g} _ {f}, \tag {3} +$$ + +where $\pmb{f}_i$ represents the feature from the $i$ -th block of the point encoder and $l$ consecutive transformer layers are regarded as one block for the convenience of explanation. To fine-tune the integration of geometric information, we introduce a learnable gating factor $g_i$ that modulates the balance between the attention layer's output and the input feature $\pmb{f}_i$ . This gating factor is initially set to zero, to ensure that the incorporation of geometric data does not abruptly alter $\pmb{f}$ feature distribution. Such an approach allows for the preservation and effective utilization of the point encoder's pre-trained weights. Upon processing through the Geometric Enhancer Module (GEM), the modified output of the point encoder, LLM are formulated as: + +$$ +\hat {\pmb {f}} _ {p o i n t} = \mathcal {E} (\pmb {i} _ {p o i n t}, \pmb {g} _ {f}), \quad \hat {\pmb {y}} = \mathcal {F} (\hat {\pmb {f}} _ {p o i n t}, \pmb {f} _ {t x t}), \quad \hat {\pmb {h}} _ {s e g} = \gamma (\hat {\pmb {y}} _ {[ s e g ]}). \qquad (4) +$$ + +# 3.4 Geometric-guided Feature Propagation + +Addressing the challenge of upsampling point clouds from a sparse set of $N_{1}$ points to a denser set of $N$ points is crucial, as direct upsampling inevitably intro- + +duces noise and results in information loss, leading to sub-optimal performance in segmentation tasks. To mitigate these issues, we introduce Geometric-guided Feature Propagation designed to generate high-quality per-point embeddings. Geometric features $\pmb{g}_f$ , which carry comprehensive point information, serving as a "gold message" for enhancing the upsampling process. By integrating these geometric features, we aim to significantly improve the quality and accuracy of the generated dense per-point embeddings. + +As illustrated in Fig. 3, we begin by upsampling the higher-layer features $\pmb{f}_3, \pmb{f}_4$ from a smaller set of points $N_1$ to larger sets $N_3, N_2$ , employing Point-Net++'s [46] propagation techniques. This step yields features $\pmb{f}_3' \in \mathbb{R}^{N_3 \times D}$ , and $\pmb{f}_4' \in \mathbb{R}^{N_2 \times D}$ . Subsequently, we perform downsampling on the geometric features $\pmb{g}_f$ from the original number of points $N$ to reduced counts $N_2, N_3$ , respectively, utilizing the Farthest Point Sampling (FPS) technique. In this process, we directly obtain the features of the sampled points without performing additional $k$ -nearest neighbor ( $k$ -NN) and pooling operations to simplify the computation and produce features $\pmb{f}_{g,1} \in \mathbb{R}^{N_3 \times D}$ , and $\pmb{f}_{g,2} \in \mathbb{R}^{N_2 \times D}$ . + +In the next phase, we integrate the up- and downsampled features, processing them through fully connected layers and ReLU activation to update the feature vectors $\tilde{\pmb{f}}_3\in \mathbb{R}^{N_3\times D}$ , and $\tilde{\pmb{f}}_4\in \mathbb{R}^{N_2\times D}$ . Note that the last layer feature $\pmb{f}_{5}$ bypasses this step. Instead, we concatenate it with $\hat{\pmb{h}}_{point}$ from the LLM output to form $\tilde{\pmb{f}}_5$ to perceive multi-modal information from LLM. + +Finally, to enable information exchange across different point densities, we propose attentive propagation. Take the propagation from $\tilde{\pmb{f}}_5$ to $\tilde{\pmb{f}}_4$ as example. Here, $\tilde{\pmb{f}}_4\in \mathbb{R}^{N_2\times D}$ acts as a set of local centers. For each local center within $\tilde{\pmb{f}}_4$ , we identify its neighboring points from $\tilde{\pmb{f}}_5$ using the $k$ -NN algorithm, resulting in $\pmb{f}_{54}\in \mathbb{R}^{N_2\times k\times D}$ . Then, employing cross-attention mechanism, where $\tilde{\pmb{f}}_4$ serves as query and $\pmb{f}_{54}$ as both key and value, facilitates information flow across different point densities and effectively extract relevant details into the query points. + +$$ +\hat {\tilde {\boldsymbol {f}}} _ {4} = \tilde {\boldsymbol {f}} _ {4} + \operatorname {s o f t m a x} \left(\frac {\tilde {\boldsymbol {f}} _ {4} \boldsymbol {f} _ {5 4} ^ {T}}{\sqrt {D}}\right) \boldsymbol {f} _ {5 4}. \tag {5} +$$ + +Leveraging geometric-guided feature propagation enables us to produce high-quality per-point embeddings denoted as $\pmb{f}_{\mathcal{P}}$ , laying the foundation for generating precise segmentation masks expressed as follows: + +$$ +\boldsymbol {f} _ {\mathcal {P}} = \mathcal {P} \left(\hat {\boldsymbol {f}} _ {\text {p o i n t}}, \boldsymbol {g} _ {f}\right), \quad \hat {\boldsymbol {m}} = \hat {\boldsymbol {h}} _ {\text {s e g}} \otimes \boldsymbol {f} _ {\mathcal {P}}. \tag {6} +$$ + +# 3.5 Training Objectives + +Our model is trained end-to-end leveraging the text classification loss and the segmentation mask loss: + +$$ +\mathcal {L} = \lambda_ {t x t} \mathcal {L} _ {t x t} + \lambda_ {b c e} \mathcal {L} _ {b c e} + \lambda_ {d i c e} \mathcal {L} _ {d i c e}, \tag {7} +$$ + +where $\mathcal{L}_{txt}$ denotes the auto-regressive cross-entropy loss targeting text generation accuracy, segmentation mask loss includes both binary cross-entropy (BCE) + +loss $\mathcal{L}_{bce}$ and DICE loss $\mathcal{L}_{dice}$ , aims at refining segmentation quality. The weights $\lambda_{txt}$ , $\lambda_{bce}$ and $\lambda_{dice}$ are utilized to balance the different loss items. The model's training is guided by the ground-truth labels $\pmb{y}_{txt}$ for text and $\pmb{M}$ for masks. + +# 3.6 Instruct3D Dataset Collection + +Although 3D instruction segmentation and 3D referring segmentation [1,4,79] are both language-based segmentation, 3D referring segmentation guides segmentation with explicit target object names, e.g., "chair", lacking more complicated reasoning instructions, e.g., "Where to sit in the room?". Besides, they also fall short in offering multi-target question-answer pairs with target descriptions directly connected to multiple segmentation masks, which cannot meet a common requirement in real-world scenarios, like "How to play computer games". + +To enhance the assessment and analysis of instruction segmentation capabilities, we have developed a benchmark, referred to Instruct3D. This benchmark incorporates 280 scenes specifically selected for instruction segmentation tuning and evaluation, sourced from the recently introduced ScanNet++ [74] dataset. Each scene comes with approximately 10 different segmentation instructions, resulting in 2,565 instruction-point cloud pairings. This dataset is then divided into two splits: train, and val, containing 2,052, and 513 question-answer pairs, respectively. Our dataset is uniquely designed to encompass both multi-target and zero-target scenarios, addressing the real-world requirement of identifying multiple objects in response to text queries and accounting for situations where objects mentioned in the text may not be present in the paired point cloud. Besides, we take into account the characteristics of 3D scenes and incorporate diverse locations and view descriptions e.g., "something that is used for sitting while working at a desk. It is the one facing the window." The model needs to have not only reasoning capabilities but also the ability to perceive views and directions in 3D scenes. These designs underscore the dataset's practical value. + +# 4 Experiments + +# 4.1 Datasets and Evaluation Metrics + +Datasets. Our training data is composed of two types of datasets: (1) semantic segmentation dataset including ScanNet200 [53], and S3DIS [3]; (2) referring segmentation dataset consisting of ScanRefer [4], ReferIt3D [1](including Sr3D and Nr3D), and Multi3DRefer [79]. We design task-specific prompts to facilitate the joint training of various tasks within a unified framework. + +Semantic Segmentation Dataset. We use two strategies to generate templates. 1) segment the specific category: "USER: Can you segment the {category} category in this point cloud? ASSISTANT: {category} ," where category is the random chosen category, and denotes the placeholder for tokens of point cloud patches. 2) segment all the categories: "USER: Can you segment all the semantic masks in this point cloud and output separate masks for each category in the alphabetical order of the + +categories? ASSISTANT: {category} , {category} , ...” To simplify the output and ensure it has only one possible answer, we add the constraints “in the alphabetical order of the categories”. To avoid generating class names not in the dataset, we incorporate category names in a dataset into the prompts during training and inference. + +Referring Segmentation Dataset. We use template prompts: "USER: Can you segment the object {description} in this point cloud? ASSISTANT: {category} ," where {description} is the given explicit description from referring segmentation dataset. It is worth noting that during training, we also use other templates to generate the QA data to ensure data diversity. We add {category} in front of to unify the output format so that when outputting semantic masks, the output category name is the label it predicts. + +Evaluation Metrics. We follow most previous works on 3D segmentation [22, 27, 54] to adopt mIoU as primary metric. mIoU is defined by the average of all per-point cloud scene Intersection-over-Unions (IoUs). Besides, we employ accuracy (Acc) as a metric to evaluate whether the model accurately identifies targets with which the predictions have an IoU greater than 0.5. + +# 4.2 Implementation Details + +In our experiments, unless specified otherwise, we employ the LLaMA2-7B model [59] as the large language model $\mathcal{F}$ and Uni3D [81] as the point cloud processing backbone $\mathcal{E}$ . The training stage leverages the deepspeed [51] engine for efficiency, with the AdamW [39] optimizer guiding the learning process. The learning rate and weight decay are set to 0.0003 and 0, respectively, enhanced by a WarmupDecayLR learning rate scheduler that initiates with 100 warmup iterations. The projection layer $\gamma$ utilizes an MLP with channel sizes of [256, 4096, 4096]. We set balancing weight $\lambda_{txt\_gen}$ , $\lambda_{bce}$ , and $\lambda_{dice}$ to 1.0, 2.0, 2.0, respectively. The experiments utilize a total batch size of 16, distributed across 4 NVIDIA 80G A100 GPUs, and span 5,000 iterations, culminating in a training period of approximately 3 days. During training, we use all mentioned datasets in Sec. 4.1 for joint training by leveraging task-specific prompts. For evaluation on a specific dataset, we finetune the trained model on the corresponding dataset. + +# 4.3 Results on Instruct3D + +The instruction segmentation results, as detailed in Table 1, underscore a significant advancement: where existing methodologies fall short, our model demonstrates exceptional prowess, achieving a more than $15\%$ improvement in mIoU for tasks requiring intricate reasoning. Unlike conventional referring segmentation tasks, instruction segmentation demands not just identification but also understanding, necessitating the model's reasoning capabilities and access to world knowledge. Existing approaches, confined to explicit references, struggle with implicit queries due to their lack of understanding, which further underscores the task's inherent challenges. In contrast, our model leverages LLMs to bridge this gap, demonstrating superior performance by comprehending and interpreting the queries accurately. Moreover, SegPoint configuration substantially outperforms $\mathrm{SegPoint}^{\dagger}$ , highlighting the critical role of our designed Geometric + +Table 1: 3D instruction segmentation benchmark results on Instruct3D val split evaluated by Acc and mIoU. † denotes our vanilla baseline removing geometric enhancer module and geometric-guided feature propagation. * represents adding an auxiliary mask head through our implementation. + +
StageMethodReferenceAccmIoU
TwoScanRefer [4][ECCV'20]12.06.9
TwoReferIt3D [1][ECCV'20]11.76.4
TwoM3DRef-CLIP [79][ICCV'23]18.112.8
SingleTGNN [22][AAAI'21]12.97.1
SingleBUTD-DETR [23]*[ECCV'22]16.310.9
SingleEDA [67]*[CVPR'23]16.612.1
SingleSegPoint†[ECCV'24]21.816.1
SingleSegPoint[ECCV'24]31.627.5
+ +Table 2: 3D Semantic segmentation benchmark results on S3DIS [3], ScanNet [7], and ScanNet200 [53]. We evaluate on the Area 5 of S3DIS and validation split of ScanNet and ScanNet200. + +
MethodReferenceScanNetScanNet200S3DIS
PointNet++ [46][NeurIPS'17]53.5--
MinkUNet [6][CVPR'19]72.225.065.4
PTv1 [80][ICCV'21]70.627.870.4
PTv2 [66][NeurIPS'22]75.430.271.6
PointNeXt [48][NeurIPS'22]71.5-70.5
OctFormer [60][SIGGRAPH'23]75.732.6-
Swin3D [72][ArXiv]75.5-72.5
SegPoint[ECCV'24]74.135.372.4
+ +Enhancer Module and Geometric-guided Feature Propagation components. Notably, even in its baseline form, $\mathrm{SegPoint}^{\dagger}$ surpasses all competing methods, validating the effectiveness and rationale behind our pipeline design. + +Besides, different from traditional two-stage approaches that first generate mask proposals using a pre-trained segmentor like Mask3D [54] and then apply language-aware networks for selection, SegPoint directly tackles the task, bypassing the need for preliminary mask proposals, enhancing its efficiency. + +# 4.4 Results on Semantic Segmentation + +Table 2 present SegPoint's performance on semantic segmentation, delivering competitive results across diverse datasets. Our model uses a simple yet effective answer format, category $<\mathrm{SEG}>$ , to use category name as predicted labels, achieving particularly stronger performance in datasets with various categories like ScanNet200 [53], where it surpasses SOTA methods by $2.1\%$ mIoU. To ensure fair comparisons, we fine-tune our model on each semantic segmentation dataset to accommodate the varying class category definitions. + +# 4.5 Results on Referring Segmentation + +Table 3 presents results on referring segmentation datasets. SegPoint showcases outstanding performance in both single-target (e.g., ScanRefer [4], Nr3D [1]) and multi-target and zero-target contexts within the Multi3DRefer [79] dataset. For multi-targets, we aggregate masks into a single ground truth, and for zero-target, we use an empty mask, indicated by "ASSISTANT: There is no mask." SegPoint significantly surpasses other approaches, achieving $2.5\%$ mIoU increase. The evaluation process of two-stage method M3DRef-CLIP is similar to Sec.4.3. + +Table 3: 3D referring segmentation benchmark results on ScanRefer [4], Nr3D [1], and Multi3Drefer [79] evaluated by mIoU. * represents adding an auxiliary mask head through our implementation. + +
StageMethodReferenceScanReferNr3DMulti3DRefer
TwoM3DRef-CLIP [79][ICCV'23]35.727.032.6
Two3D-STMN [63][AAAI'24]39.5--
SingleTGNN [22][AAAI'21]27.8--
SingleBUTD-DETR [23]*[ECCV'22]35.427.526.2
SingleEDA [67]*[CVPR'23]36.229.328.9
SingleX-RefSeg3D [49][AAAI'24]29.9--
SingleRefMask3D [18][ACMMM'24]44.8--
SingleSegPoint[ECCV'24]41.732.236.1
+ +Table 4: 3D open-vocabulary semantic segmentation benchmark results on val split of ScanNet++ [74]. + +
TypeMethodReferenceScanNet++
SupervisedPointNet [45][CVPR'17]7.0
PointNet++ [46][NeurIPS'17]15.0
MinkUNet [6][CVPR'19]28.0
KPCConv [58][ICCV'19]30.0
Open-VocabularyOpenScene [42][CVPR'23]12.8
PLA [12][CVPR'23]14.2
RegionPLC [71][CVPR'24]14.9
UnifiedSegPoint[ECCV'24]19.3
+ +# 4.6 Results on Open-vocabulary Semantic Segmentation + +Table 4 shows our method's open-vocabulary segmentation performance which is directly evaluated on ScanNet++ [74] following the setting in prevalent methodologies in the 2D domain [13,69]. It demonstrates our superiority over both existing open-vocabulary techniques and even several supervised approaches, showing our model's robust generalization capabilities. It effectively aligns and interprets categories with visual scenes, underscoring the reasoning prowess of large language models. A notable issue is the potential misalignment between output categories of SegPoint and val split category names. To address this, we employ GPT-4 to match its most similar category names in the val split. + +# 4.7 Ablation Study + +We conduct extensive experiments to verify the effectiveness of our proposed components in Table 5 (a) on both Instruct3D and ScanRefer [4] dataset. We established a vanilla baseline as described in Sec. 3.2 following the paradigm of LISA [28], which cannot achieve satisfactory performance and only obtain $16.1\% / 30.3 \, \mathrm{mIoU}$ on Instruct3D/ScanRefer, respectively. Further analysis, both qualitative and quantitative, of our proposed components reveals that their integration substantially outperforms the baseline. + +Geometric Enhancer Module Integrating the Geometric Enhancer Module (GEM) into our point encoder results in a notable $5.3\% / 5.5\%$ mIoU improvement on Instruct3D/ScanRefer, effectively addressing compatibility issues with dense prediction tasks. An ablation study, shown in Table 5 (b), shows that our improvement is not due to an increase in parameters. Our approach outperforms traditional full fine-tuning, LoRA [21] strategies, and the addition of + +Table 5: Ablation studies on Instruct3D and ScanRefer. + +
IndexComponentsInstruct3D +mIoUScanRefer +mIoU
GEMGFP
0XX16.130.3
1✓X21.435.8
2X✓23.238.1
3✓✓27.541.7
+ +(a) Effect of different main components. + +
MethodInstruct3DScanRefer
Full tuning26.239.1
LoRA [21]24.838.5
MLP23.136.8
GEM27.541.7
+ +(b) Effect of different tuning methods. + +![](images/51e47c303ec60ce760ff9fb71819deb3f1c4098c4be0ff84843e4d7a33669bc8.jpg) +(a) Scene + +![](images/4964324b8e48caa54aaba3259917eb79329a6315d1f68c79b16146e456d1fd6f.jpg) +(b) Baseline +Fig. 4: (Best viewed in color) We visualize the feature responses between a given point (in red) and other points in the scene from per-point embeddings $\pmb{f}_{\mathcal{P}}$ for the baseline and our SegPoint, respectively. The color changes from yellow to red, indicating increasing feature similarity. + +![](images/3fa27e0764c2c5533d83c3f73c921b9fb735de1198169d78d22860929c027981.jpg) +(c) Ours + +MLP layers for feature adapter, underscoring its effectiveness in embedding 3D domain-specific knowledge into point encoder. Unlike adapter techniques common in language and 2D image processing, our GEM is designed to address the unique challenges in dense prediction tasks. + +Geometric-guided Feature Propagation Introducing Geometric-guided Feature Propagation (GFP) results in a substantial improvement over the baseline, as shown in Table 5 (a) (index 2). This underscores our method's capability to minimize information loss and reduce noise during the upsampling phase, leading to higher-quality per-point embeddings. + +# 4.8 Qualitative Visualization + +Fig. 4 qualitatively illustrates the feature responses between a given point and others in a scene. From left to right, it presents the original scene, baseline method, and our SegPoint, respectively, with warmer colors indicating closer feature relationships. The baseline method predominantly highlights spatially proximate points, often missing important but distant features, such as chair cushions. In contrast, SegPoint leverages global information, allowing for the identification of both near and distant relevant features, such as cushions and casters. This demonstrates our model's superior ability to capture global context and recognize intricate structures within the scene. + +As shown in Fig. 5, we provide some typical qualitative results from SegPoint on Instruct3D. Given an implicit instruction, e.g., "the object that is used for dispensing water during bathing", SegPoint successfully infers the role of a shower. In other scenes, the instructions include queries requiring extensive knowledge of the world and complex reasoning, like "What appliance is used for heating or cooking food quickly using electromagnetic radiation." SegPoint can still segment the microwave oven very well, which shows that SegPoint can make good use of the reasoning ability of LLM and provide high-quality masks. + +![](images/e476decbca80d6647d5c0d0f8dc189a75d2ac5cba270fa39298ad535ccb4c3bf.jpg) +Fig. 5: Qualitative results from val split of Instruct3D. SegPoint understand the human instruction and accurately segment the target object. We omitted the "please output segmentation mask" in the sentence for simplicity. + +# 5 Conclusion + +In this work, we introduce SegPoint, an effective model supported by LLM for point-level reasoning and segmentation. Benefiting from the proposed Geometric Enhancer Module and Geometric-guided Feature Propagation, SegPoint is adept at solving a variety of segmentation tasks in a unified framework. Additionally, we construct a comprehensive Instruct3D benchmark to bolster research area in segmentation via implicit and complex instructions, introducing more challenges to promote it closer to real-world applications. Through thorough experiments, SegPoint achieves promising results across multiple benchmarks. + +Limitations Although SegPoint demonstrates notable success in tasks driven by text prompts, its current framework cannot process non-textual prompts, such as boxes and points. Future developments will explore the adoption of a prompt encoder, inspired by SAM model [26], to extend support for these formats. + +Acknowledgements We thank the anonymous reviewers for their constructive suggestions. Following their advice, we have incorporated diverse location and view descriptions into our Instruct3D. This work was partially supported by the National Research Foundation Singapore Competitive Research Program (CRP29-2022-0003). + +# References + +1. Achlioptas, P., Abdelreheem, A., Xia, F., Elhoseiny, M., Guibas, L.: Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In: ECCV (2020) +2. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. In: NeurIPS (2022) +3. Armeni, I., Sener, O., Zamir, A.R., Jiang, H., Brilakis, I., Fischer, M., Savarese, S.: 3d semantic parsing of large-scale indoor spaces. In: CVPR (2016) +4. Chen, D.Z., Chang, A.X., Nießner, M.: Scanrefer: 3d object localization in rgb-d scans using natural language. In: ECCV (2020) +5. Chen, Z., Duan, Y., Wang, W., He, J., Lu, T., Dai, J., Qiao, Y.: Vision transformer adapter for dense predictions. In: ICLR (2023) +6. Choy, C., Gwak, J., Savarese, S.: 4d spatio-temporal convnets: Minkowski convolutional neural networks. In: CVPR (2019) +7. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: CVPR (2017) +8. Ding, H., Liu, C., He, S., Jiang, X., Loy, C.C.: MeViS: A large-scale benchmark for video segmentation with motion expressions. In: ICCV (2023) +9. Ding, H., Liu, C., He, S., Jiang, X., Torr, P.H., Bai, S.: MOSE: A new dataset for video object segmentation in complex scenes. In: ICCV (2023) +10. Ding, H., Liu, C., Wang, S., Jiang, X.: Vision-language transformer and query generation for referring segmentation. In: ICCV (2021) +1. Ding, H., Liu, C., Wang, S., Jiang, X.: VLT: Vision-language transformer and query generation for referring segmentation. IEEE TPAMI (2023) +2. Ding, R., Yang, J., Xue, C., Zhang, W., Bai, S., Qi, X.: Pla: Language-driven open-vocabulary 3d scene understanding. In: CVPR (2023) +3. Ding, Z., Wang, J., Tu, Z.: Open-vocabulary panoptic segmentation with maskclip. In: ICLR (2023) +4. Girdhar, R., El-Nouby, A., Liu, Z., Singh, M., Alwala, K.V., Joulin, A., Misra, I.: Imagebind: One embedding space to bind them all. In: CVPR (2023) +5. Guo, Z., Zhang, R., Zhu, X., Tang, Y., Ma, X., Han, J., Chen, K., Gao, P., Li, X., Li, H., et al.: Point-bind & point-llm: Aligning point cloud with multi-modality for 3d understanding, generation, and instruction following. arXiv preprint arXiv:2309.00615 (2023) +6. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016) +7. He, S., Ding, H.: Decoupling static and hierarchical motion perception for referring video segmentation. In: CVPR (2024) +8. He, S., Ding, H.: RefMask3D: Language-guided transformer for 3d referring segmentation. In: ACM MM (2024) +9. He, S., Jiang, X., Jiang, W., Ding, H.: Prototype adaption and projection for few-and zero-shot 3d point cloud semantic segmentation. IEEE TIP (2023) +20. Hong, Y., Zhen, H., Chen, P., Zheng, S., Du, Y., Chen, Z., Gan, C.: 3d-llm: Injecting the 3d world into large language models. In: NeurIPS (2023) +21. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. In: ICLR (2022) +22. Huang, P.H., Lee, H.H., Chen, H.T., Liu, T.L.: Text-guided graph neural networks for referring 3d instance segmentation. In: AAAI (2021) + +23. Jain, A., Gkanatsios, N., Mediratta, I., Fragkiadaki, K.: Bottom up top down detection transformers for language grounding in images and point clouds. In: ECCV (2022) +24. Jia, B., Chen, Y., Yu, H., Wang, Y., Niu, X., Liu, T., Li, Q., Huang, S.: Sceneverse: Scaling 3d vision-language learning for grounded scene understanding. In: ECCV (2024) +25. Jiang, L., Zhao, H., Shi, S., Liu, S., Fu, C.W., Jia, J.: Pointgroup: Dual-set point grouping for 3d instance segmentation. In: CVPR (2020) +26. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: ICCV (2023) +27. Kolodiazhnyi, M., Vorontsova, A., Konushin, A., Rukhovich, D.: Oneformer3d: One transformer for unified point cloud segmentation. In: CVPR (2024) +28. Lai, X., Tian, Z., Chen, Y., Li, Y., Yuan, Y., Liu, S., Jia, J.: Lisa: Reasoning segmentation via large language model. In: CVPR (2024) +29. Lai, X., Yuan, Y., Chu, R., Chen, Y., Hu, H., Jia, J.: Mask-attention-free transformer for 3d instance segmentation. In: ICCV (2023) +30. Li, B., Zhang, Y., Chen, L., Wang, J., Yang, J., Liu, Z.: Otter: A multi-modal model with in-context instruction tuning. arXiv:2305.03726 (2023) +31. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. In: NeurIPS (2023) +32. Lin, Z., Liu, C., Zhang, R., Gao, P., Qiu, L., Xiao, H., Qiu, H., Lin, C., Shao, W., Chen, K., et al.: Sphinx: The joint mixing of weights, tasks, and visual embeddings for multi-modal large language models. arXiv preprint arXiv:2311.07575 (2023) +33. Liu, C., Ding, H., Jiang, X.: GRES: Generalized referring expression segmentation. In: CVPR (2023) +34. Liu, C., Ding, H., Zhang, Y., Jiang, X.: Multi-modal mutual attention and iterative interaction for referring image segmentation. IEEE TIP (2023) +35. Liu, C., Jiang, X., Ding, H.: Instance-specific feature propagation for referring segmentation. IEEE TMM (2022) +36. Liu, C., Jiang, X., Ding, H.: Primitivenet: decomposing the global constraints for referring segmentation. Visual Intelligence 2(1), 16 (2024) +37. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: NeurIPS (2023) +38. Liu, Y., Kong, L., Cen, J., Chen, R., Zhang, W., Pan, L., Chen, K., Liu, Z.: Segment any point cloud sequences by distilling vision foundation models. In: NeurIPS (2023) +39. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: ICLR (2019) +40. Nguyen, P.D., Ngo, T.D., Gan, C., Kalogerakis, E., Tran, A., Pham, C., Nguyen, K.: Open3dis: Open-vocabulary 3d instance segmentation with 2d mask guidance. In: CVPR (2024) +41. Park, N., Kim, S.: How do vision transformers work? arXiv preprint arXiv:2202.06709 (2022) +42. Peng, S., Genova, K., Jiang, C., Tagliasacchi, A., Pollefeys, M., Funkhouser, T., et al.: Openscene: 3d scene understanding with open vocabularies. In: CVPR (2023) +43. Peng, Z., Wang, W., Dong, L., Hao, Y., Huang, S., Ma, S., Wei, F.: Kosmos-2: Grounding multimodal large language models to the world. In: ICLR (2024) +44. Pi, R., Gao, J., Diao, S., Pan, R., Dong, H., Zhang, J., Yao, L., Han, J., Xu, H., Zhang, L.K.T.: Detgpt: Detect what you need via reasoning. In: EMNLP (2023) +45. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: CVPR (2017) +46. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In: NeurIPS (2017) + +47. Qi, Z., Fang, Y., Sun, Z., Wu, X., Wu, T., Wang, J., Lin, D., Zhao, H.: Gpt4point: A unified framework for point-language understanding and generation. In: CVPR (2024) +48. Qian, G., Li, Y., Peng, H., Mai, J., Hammoud, H., Elhoseiny, M., Ghanem, B.: Pointnext: Revisiting pointnet++ with improved training and scaling strategies. In: NeurIPS (2022) +49. Qian, Z., Ma, Y., Ji, J., Sun, X.: X-refseg3d: Enhancing referring 3d instance segmentation via structured cross-modal graph neural networks. In: AAAI (2024) +50. Rasheed, H., Maaz, M., Shaji, S., Shaker, A., Khan, S., Cholakkal, H., Anwer, R.M., Xing, E., Yang, M.H., Khan, F.S.: Glamm: Pixel grounding large multimodal model. In: CVPR (2024) +51. Rasley, J., Rajbhandari, S., Ruwase, O., He, Y.: Deepspeed: System optimizations enable training deep learning models with over 100 billion parameters. In: KDD (2020) +52. Ren, Z., Huang, Z., Wei, Y., Zhao, Y., Fu, D., Feng, J., Jin, X.: Pixelm: Pixel reasoning with large multimodal model. In: CVPR (2024) +53. Rozenberszki, D., Litany, O., Dai, A.: Language-grounded indoor 3d semantic segmentation in the wild. In: ECCV (2022) +54. Schult, J., Engelmann, F., Hermans, A., Litany, O., Tang, S., Leibe, B.: Mask3d: Mask transformer for 3d semantic instance segmentation. In: ICRA (2023) +55. Shuai, X., Ding, H., Ma, X., Tu, R., Jiang, Y.G., Tao, D.: A survey of multimodal-guided image editing with text-to-image diffusion models. arXiv preprint arXiv:2406.14555 (2024) +56. Sun, J., Qing, C., Tan, J., Xu, X.: Superpoint transformer for 3d scene instance segmentation. In: AAAI (2023) +57. Takmaz, A., Fedele, E., Sumner, R.W., Pollefeys, M., Tombari, F., Engelmann, F.: Openmask3d: Open-vocabulary 3d instance segmentation. In: NeurIPS (2023) +58. Thomas, H., Qi, C.R., Deschaud, J.E., Marcotegui, B., Goulette, F., Guibas, L.J.: Kpconv: Flexible and deformable convolution for point clouds. In: ICCV (2019) +59. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023) +60. Wang, P.S.: Octformer: Octree-based transformers for 3d point clouds. In: SIGGRAPH (2023) +61. Wang, W., Chen, Z., Chen, X., Wu, J., Zhu, X., Zeng, G., Luo, P., Lu, T., Zhou, J., Qiao, Y., et al.: Visionllm: Large language model is also an open-ended decoder for vision-centric tasks. In: NeurIPS (2023) +62. Wang, W., Xie, E., Li, X., Fan, D.P., Song, K., Liang, D., Lu, T., Luo, P., Shao, L.: Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media (2022) +63. Wu, C., Ma, Y., Chen, Q., Wang, H., Luo, G., Ji, J., Sun, X.: 3d-stmn: Dependency-driven superpoint-text matching network for end-to-end 3d referring expression segmentation. In: AAAI (2024) +64. Wu, H., Xiao, B., Codella, N., Liu, M., Dai, X., Yuan, L., Zhang, L.: Cvt: Introducing convolutions to vision transformers. In: ICCV (2021) +65. Wu, J., Li, X., Xu, S., Yuan, H., Ding, H., Yang, Y., Li, X., Zhang, J., Tong, Y., Jiang, X., Ghanem, B., Tao, D.: Towards open vocabulary learning: A survey. IEEE TPAMI (2024) +66. Wu, X., Lao, Y., Jiang, L., Liu, X., Zhao, H.: Point transformer v2: Grouped vector attention and partition-based pooling. In: NeurIPS (2022) + +67. Wu, Y., Cheng, X., Zhang, R., Cheng, Z., Zhang, J.: Eda: Explicit text-decoupling and dense alignment for 3d visual grounding. In: CVPR (2023) +68. Xiao, Z., Zhang, W., Wang, T., Loy, C.C., Lin, D., Pang, J.: Position-guided point cloud panoptic segmentation transformer. arXiv preprint arXiv:2303.13509 (2023) +69. Xu, J., Liu, S., Vahdat, A., Byeon, W., Wang, X., De Mello, S.: Open-vocabulary panoptic segmentation with text-to-image diffusion models. In: CVPR (2023) +70. Xu, R., Wang, X., Wang, T., Chen, Y., Pang, J., Lin, D.: Pointllm: Empowering large language models to understand point clouds. In: ECCV (2024) +71. Yang, J., Ding, R., Deng, W., Wang, Z., Qi, X.: Regionplc: Regional point-language contrastive learning for open-world 3d scene understanding. In: CVPR (2024) +72. Yang, Y.Q., Guo, Y.X., Xiong, J.Y., Liu, Y., Pan, H., Wang, P.S., Tong, X., Guo, B.: Swin3d: A pretrained transformer backbone for 3d indoor scene understanding. arXiv preprint arXiv:2304.06906 (2023) +73. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: MPLug-owl: Modularization empowers large language models with multimodality. arXiv:2304.14178 (2023) +74. Yeshwanth, C., Liu, Y.C., Nießner, M., Dai, A.: Scannet++: A high-fidelity dataset of 3d indoor scenes. In: ICCV (2023) +75. You, H., Zhang, H., Gan, Z., Du, X., Zhang, B., Wang, Z., Cao, L., Chang, S.F., Yang, Y.: Ferret: Refer and ground anything anywhere at any granularity. In: ICLR (2024) +76. Zhang, H., Li, H., Li, F., Ren, T., Zou, X., Liu, S., Huang, S., Gao, J., Zhang, L., Li, C., et al.: Llava-grounding: Grounded visual chat with large multimodal models. arXiv preprint arXiv:2312.02949 (2023) +77. Zhang, H., Ding, H.: Prototypical matching and open set rejection for zero-shot semantic segmentation. In: ICCV (2021) +78. Zhang, S., Sun, P., Chen, S., Xiao, M., Shao, W., Zhang, W., Chen, K., Luo, P.: Gpt4roi: Instruction tuning large language model on region-of-interest. arXiv:2307.03601 (2023) +79. Zhang, Y., Gong, Z., Chang, A.X.: Multi3defer: Grounding text description to multiple 3d objects. In: ICCV (2023) +80. Zhao, H., Jiang, L., Jia, J., Torr, P.H., Koltun, V.: Point transformer. In: ICCV (2021) +81. Zhou, J., Wang, J., Ma, B., Liu, Y.S., Huang, T., Wang, X.: Uni3d: Exploring unified 3d representation at scale. In: ICLR (2024) +82. Zhou, Z., Zhang, Y., Foroosh, H.: Panoptic-polarnet: Proposal-free lidar point cloud panoptic segmentation. In: CVPR (2021) +83. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. In: ICLR (2024) +84. Zhu, Z., Ma, X., Chen, Y., Deng, Z., Huang, S., Li, Q.: 3d-vista: Pre-trained transformer for 3d vision and text alignment. In: ICCV (2023) \ No newline at end of file diff --git a/2024/SegPoint_ Segment Any Point Cloud via Large Language Model/images.zip b/2024/SegPoint_ Segment Any Point Cloud via Large Language Model/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..cc8f976822d76d6ac3ed720735a2361cb4dd9dfe --- /dev/null +++ b/2024/SegPoint_ Segment Any Point Cloud via Large Language Model/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:12520e7ead9b2b8261b5f98d58d88db89c25840394ad1909d9a152bfb8b5ed14 +size 594793 diff --git a/2024/SegPoint_ Segment Any Point Cloud via Large Language Model/layout.json b/2024/SegPoint_ Segment Any Point Cloud via Large Language Model/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e4c3417021a27f364a7e1ae7541f4a21fb4802f0 --- /dev/null +++ b/2024/SegPoint_ Segment Any Point Cloud via Large Language Model/layout.json @@ -0,0 +1,10628 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 185, + 122, + 465, + 157 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 185, + 122, + 465, + 157 + ], + "spans": [ + { + "bbox": [ + 185, + 122, + 465, + 157 + ], + "type": "text", + "content": "SegPoint: Segment Any Point Cloud via Large Language Model" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 148, + 177, + 465, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 177, + 465, + 190 + ], + "spans": [ + { + "bbox": [ + 148, + 177, + 465, + 190 + ], + "type": "text", + "content": "Shuting " + }, + { + "bbox": [ + 148, + 177, + 465, + 190 + ], + "type": "inline_equation", + "content": "\\mathrm{He}^{1}" + }, + { + "bbox": [ + 148, + 177, + 465, + 190 + ], + "type": "text", + "content": " , Henghui Ding2, Xudong Jiang1, and Bihan " + }, + { + "bbox": [ + 148, + 177, + 465, + 190 + ], + "type": "inline_equation", + "content": "\\mathrm{Wen}^{1}" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 220, + 199, + 392, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 199, + 392, + 210 + ], + "spans": [ + { + "bbox": [ + 220, + 199, + 392, + 210 + ], + "type": "text", + "content": "1 Nanyang Technological University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 220, + 210, + 392, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 210, + 392, + 222 + ], + "spans": [ + { + "bbox": [ + 220, + 210, + 392, + 222 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 220, + 210, + 392, + 222 + ], + "type": "text", + "content": " Institute of Big Data, Fudan University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 222, + 478, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 222, + 478, + 233 + ], + "spans": [ + { + "bbox": [ + 137, + 222, + 478, + 233 + ], + "type": "text", + "content": "{heshuting555, henghui.ding}@gmail.com, {exdjiang, bihan.wen}@ntu.edu.sg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 214, + 233, + 399, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 233, + 399, + 244 + ], + "spans": [ + { + "bbox": [ + 214, + 233, + 399, + 244 + ], + "type": "text", + "content": "https://heshuting555.github.io/SegPoint" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 159, + 271, + 455, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 271, + 455, + 470 + ], + "spans": [ + { + "bbox": [ + 159, + 271, + 455, + 470 + ], + "type": "text", + "content": "Abstract. Despite significant progress in 3D point cloud segmentation, existing methods primarily address specific tasks and depend on explicit instructions to identify targets, lacking the capability to infer and understand implicit user intentions in a unified framework. In this work, we propose a model, called SegPoint, that leverages the reasoning capabilities of a multi-modal Large Language Model (LLM) to produce point-wise segmentation masks across a diverse range of tasks: 1) 3D instruction segmentation, 2) 3D referring segmentation, 3) 3D semantic segmentation, and 4) 3D open-vocabulary semantic segmentation. To advance 3D instruction research, we introduce a new benchmark, Instruct3D, designed to evaluate segmentation performance from complex and implicit instructional texts, featuring 2,565 point cloud-instruction pairs. Our experimental results demonstrate that SegPoint achieves competitive performance on established benchmarks such as ScanRefer for referring segmentation and ScanNet for semantic segmentation, while delivering outstanding outcomes on the Instruct3D dataset. To our knowledge, SegPoint is the first model to address these varied segmentation tasks within a single framework, achieving satisfactory performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 159, + 479, + 453, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 479, + 453, + 502 + ], + "spans": [ + { + "bbox": [ + 159, + 479, + 453, + 502 + ], + "type": "text", + "content": "Keywords: Instruct3D dataset " + }, + { + "bbox": [ + 159, + 479, + 453, + 502 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 479, + 453, + 502 + ], + "type": "text", + "content": " Unified framework " + }, + { + "bbox": [ + 159, + 479, + 453, + 502 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 479, + 453, + 502 + ], + "type": "text", + "content": " 3D point cloud segmentation " + }, + { + "bbox": [ + 159, + 479, + 453, + 502 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 479, + 453, + 502 + ], + "type": "text", + "content": " Large language model" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 538, + 230, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 538, + 230, + 551 + ], + "spans": [ + { + "bbox": [ + 132, + 538, + 230, + 551 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 563, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 563, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 563, + 482, + 647 + ], + "type": "text", + "content": "3D point cloud segmentation, a critical challenge in the 3D vision community, aims to interpret and classify each point in a point cloud to understand its semantic properties [27, 45, 46, 54, 56, 80]. This longstanding issue has spurred significant advancements across various fields, including robotics, autonomous driving, virtual reality, etc. This challenge has evolved into a series of specialized tasks, each targeting a specific segmentation aspect. Overall, tasks cover basic semantic and instance segmentation [3, 7, 53, 74], as well as more practical tasks" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 148, + 114, + 176, + 140 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 114, + 176, + 140 + ], + "spans": [ + { + "bbox": [ + 148, + 114, + 176, + 140 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 140, + 654, + 246, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 654, + 246, + 666 + ], + "spans": [ + { + "bbox": [ + 140, + 654, + 246, + 666 + ], + "type": "text", + "content": "Corresponding author" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 479, + 364 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 479, + 364 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 479, + 364 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 479, + 364 + ], + "type": "image", + "image_path": "83c779324b564b4662467dc72345b67eba7c11560d8f0f52d90769a1aef47e55.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 374, + 482, + 419 + ], + "lines": [ + { + "bbox": [ + 130, + 374, + 482, + 419 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 482, + 419 + ], + "type": "text", + "content": "Fig. 1: Example of functionality in SegPoint. SegPoint can complete various point cloud tasks in a unified framework by leveraging task-specific prompts, including 1) 3D instruction segmentation, 2) 3D referring segmentation, 3) 3D semantic segmentation, and 4) 3D open-vocabulary semantic segmentation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 434, + 482, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 482, + 472 + ], + "type": "text", + "content": "such as referring segmentation [1, 4, 24, 33, 79, 84], which segments points based on explicit textual descriptions, and open-vocabulary segmentation [12, 40, 42, 57, 65, 77] designed for the dynamic and complex nature of real-world." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 472, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 628 + ], + "type": "text", + "content": "Despite significant progress achieved within the 3D community toward accurately segmenting objects through specifically designed models, each model is typically developed to tackle one specific segmentation task, leading to inefficiencies and a lack of versatility for real-world application. Furthermore, previous perception approaches heavily depend on predefined categories or explicit expressions for language understanding. Such approaches fall short in interpreting and acting on implicit instructions often found in human language, a critical gap that hinders the development of truly intelligent next-generation perception systems. This brings a pivotal question: Is it possible to design a unified model capable of comprehensively addressing all aforementioned 3D tasks with human-like instructions? The exploration of this question not only challenges the current paradigms of 3D point cloud segmentation but also opens the door to groundbreaking advancements in robotic perception and interaction." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "In this work, we propose a model called SegPoint, leveraging the Large Language Model's (LLM) advanced ability to reason and comprehend user instructions. To enhance 3D scene comprehension, we integrate a Geometric Enhancer" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "S. He et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "Module that extracts local semantics from point clouds, seamlessly incorporating this geometric insight into the feature extraction process. Furthermore, a Geometric-guided Feature Propagation is designed to utilize semantic flows derived from geometric priors and the LLM's hidden embeddings, facilitating the generation of fine-grained, high-quality features for accurate dense prediction tasks. Unlike previous attempts at 2D field [8, 28, 50], we do not depend on additional costly pre-trained segmentation models like SAM [26]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 482, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 482, + 284 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 482, + 284 + ], + "type": "text", + "content": "Moreover, we introduce a benchmark named Instruct3D, designed to advance research in the field of segmentation driven by implicit and complex instructions. Understanding these nuanced instructions necessitates reasoning abilities and extensive knowledge of the world. It includes a total of 2,565 diverse pairs of instructions and point clouds for tuning and evaluation. Our comprehensive experiments demonstrate the benchmark's utility in evaluating the model's capability of segmentation based on human-like instructions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 284, + 482, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 284, + 482, + 368 + ], + "spans": [ + { + "bbox": [ + 130, + 284, + 482, + 368 + ], + "type": "text", + "content": "Taking advantage of a multi-modal LLM and task-specific prompts, SegPoint is capable of generating segmentation masks for a wide range of tasks in a unified model: 1) 3D instruction segmentation, 2) 3D referring segmentation, 3) 3D semantic segmentation, and 4) 3D open-vocabulary semantic segmentation, as depicted in Fig. 1. SegPoint achieves competitive results on established benchmarks like ScanRefer [4] for referring segmentation and ScanNet [7] for semantic segmentation while showing remarkable performance on the Instruct3D dataset." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 146, + 369, + 372, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 369, + 372, + 380 + ], + "spans": [ + { + "bbox": [ + 146, + 369, + 372, + 380 + ], + "type": "text", + "content": "In summary, our main contributions are as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 389, + 480, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 389, + 480, + 425 + ], + "spans": [ + { + "bbox": [ + 138, + 389, + 480, + 425 + ], + "type": "text", + "content": "- We propose SegPoint, the first 3D segmentation model that can comprehend human intentions and address multiple segmentation tasks within one framework by harnessing the Large Language Model's reasoning capabilities." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 426, + 480, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 426, + 480, + 485 + ], + "spans": [ + { + "bbox": [ + 138, + 426, + 480, + 485 + ], + "type": "text", + "content": "- We present a Geometric Enhancer Module that integrates comprehensive scene information into the process of 3D scene understanding. Besides, A Geometric-guided Feature Propagation is designed to achieve accurate and fine-grained segmentation. These two modules supplement the missing local information and grasp fine-grained features for dense prediction tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 487, + 480, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 487, + 480, + 523 + ], + "spans": [ + { + "bbox": [ + 138, + 487, + 480, + 523 + ], + "type": "text", + "content": "- We introduce a new task called 3D instruction segmentation and construct a new dataset Instruct3D, which necessitates a model's self-reasoning to interpret implicit instructions for segmenting the target object." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 525, + 480, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 525, + 480, + 572 + ], + "spans": [ + { + "bbox": [ + 138, + 525, + 480, + 572 + ], + "type": "text", + "content": "- Our experimental findings reveal that SegPoint not only competes strongly in 3D semantic, referring, and open-vocabulary semantic segmentation but also excels in 3D instruction segmentation, showcasing its versatility and effectiveness across a spectrum of segmentation challenges." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 593, + 237, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 593, + 237, + 605 + ], + "spans": [ + { + "bbox": [ + 131, + 593, + 237, + 605 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 620, + 343, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 620, + 343, + 632 + ], + "spans": [ + { + "bbox": [ + 131, + 620, + 343, + 632 + ], + "type": "text", + "content": "2.1 Multi-modal Large Language Model" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Inspired by the exceptional reasoning abilities of Large Language Models, researchers are delving into transferring these capabilities into the vision realm [9," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 410, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 410, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 410, + 91, + 447, + 102 + ], + "type": "text", + "content": "SegPoint" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 283 + ], + "type": "text", + "content": "28, 36, 55], developing multi-modal LLMs. Flamingo [2], BLIP-2 [31], mPLUGOWL [73], Otter [30], LLaVA [37] and MiniGPT-4 [83] first construct image-text feature alignment followed by instruction tuning and achieve superior performance. Recent studies have increasingly concentrated on integrating foundational models with tasks that demand a refined understanding at the region or pixel level. VisionLLM [61] introduces a versatile interaction interface for various vision-centric tasks via instruction tuning, albeit without fully leveraging LLMs for intricate reasoning tasks. Kosmos-2 [43] has built substantial data of grounded image-text pairs, thereby embedding grounding capabilities into LLMs. DetGPT [44] links the fixed multi-modal LLM with an open-vocabulary detector, facilitating user instruction-based detection tasks. This growing interest has spurred further innovations, including GPT4RoI [78], LLaVA-grounding [76], Ferret [75], LISA [28], GlaMM [50], PixelLM [52], Sphinx [32], each contributing to the evolving landscape of instruction-based, multi-modal understanding." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 284, + 483, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 284, + 483, + 451 + ], + "spans": [ + { + "bbox": [ + 130, + 284, + 483, + 451 + ], + "type": "text", + "content": "Building on advancements of multi-modal large language models in 2D image domain, the field is witnessing a seamless transition into 3D spaces. PointLLM [70] harnesses the prowess of large language models and trains with 3D point clouds following the paradigm of LLaVA [37]. 3D-LLM [20] utilizes 2D foundation models to encode multi-view images of 3D point clouds. Point-Bind [15] aligns point clouds with Image-Bind [14] and leverages ImageBind-LLM to reason multimodality input without 3D-instruction data training. GPT4Point [47] pioneers in facilitating a unified approach towards 3D object understanding and generation, setting a new standard for versatility. However, these models primarily concentrate on scene-level insights, often overlooking the intricate details at the region, or point level. Contrasting with existing models, our research focuses on two pivotal goals: 1) efficiently inject segmentation capabilities into multi-modal LLMs to conduct point-level understanding and 2) design a unified framework for 3D point cloud segmentation via the reasoning ability of LLMs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 467, + 312, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 467, + 312, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 467, + 312, + 479 + ], + "type": "text", + "content": "2.2 3D Point Cloud Segmentation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 486, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 628 + ], + "type": "text", + "content": "3D point cloud segmentation, a crucial task in computer vision, can be categorized into 3D semantic, instance, and panoptic segmentation. 3D semantic segmentation [3,7,19,45,53,74,80] assigns each point in a 3D space to specific, predefined classes. In contrast, 3D instance segmentation [25,27,54] goes a step further by classifying each point and differentiating between distinct objects of the same class. 3D panoptic segmentation [68,82] aims to group 3D points according to their semantics and identities. Lately, more practical tasks have emerged, such as 3D referring segmentation [1,4,17,23,67,79], which extends referring expression segmentation [8,10,11,33-35] to 3D and segments a target instance based on explicit linguistic descriptions, and 3D open-vocabulary segmentation [12,40,42,57], designed to identify and segment unseen objects beyond a fixed set of known categories." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 630, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 630, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 630, + 481, + 665 + ], + "type": "text", + "content": "Despite the significant progress made by transformer-based models: Mask3D [54], SPFormer [56], MAFT [29], and OneFormer3D [27] in basic 3D segmentation tasks, their application in real-world scenarios requiring human language" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "S. He et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 118, + 479, + 257 + ], + "blocks": [ + { + "bbox": [ + 136, + 118, + 479, + 257 + ], + "lines": [ + { + "bbox": [ + 136, + 118, + 479, + 257 + ], + "spans": [ + { + "bbox": [ + 136, + 118, + 479, + 257 + ], + "type": "image", + "image_path": "59e9162900bb37310cabfb6aa72651bec476a3448dd6d46a0ab4266b49ffe291.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 262, + 482, + 319 + ], + "lines": [ + { + "bbox": [ + 130, + 262, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 482, + 319 + ], + "type": "text", + "content": "Fig. 2: The pipeline of SegPoint. Given input point cloud and text query, the multimodal LLM " + }, + { + "bbox": [ + 130, + 262, + 482, + 319 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 130, + 262, + 482, + 319 + ], + "type": "text", + "content": " generates text output. Geometric Enhancer Module " + }, + { + "bbox": [ + 130, + 262, + 482, + 319 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 130, + 262, + 482, + 319 + ], + "type": "text", + "content": " injects geometric information into Point Encoder " + }, + { + "bbox": [ + 130, + 262, + 482, + 319 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 130, + 262, + 482, + 319 + ], + "type": "text", + "content": " and obtains point features " + }, + { + "bbox": [ + 130, + 262, + 482, + 319 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{f}}_{\\text {point }}" + }, + { + "bbox": [ + 130, + 262, + 482, + 319 + ], + "type": "text", + "content": ". Per-point embeddings " + }, + { + "bbox": [ + 130, + 262, + 482, + 319 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{\\mathcal{P}}" + }, + { + "bbox": [ + 130, + 262, + 482, + 319 + ], + "type": "text", + "content": " derived from Geometric-guided Feature Propagation " + }, + { + "bbox": [ + 130, + 262, + 482, + 319 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 130, + 262, + 482, + 319 + ], + "type": "text", + "content": " multiplied with the embedding associated with the token yield the final segmentation masks." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 331, + 482, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 482, + 439 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 482, + 439 + ], + "type": "text", + "content": "interaction remains constrained. Besides, Seal [38] aims to segment any point cloud through distilling vision foundation models, while it doesn't use language as cues. TGNN [22] is the first work to tackle referring segmentation problem that proposes aggregating textual features by considering the neighboring local structure of each instance but it heavily depends on explicit expressions or predefined categories for language understanding. Furthermore, the development of models tailored to specific segmentation tasks restricts their versatility and applicability in diverse real-world scenarios. Therefore, it is imperative to develop more intelligent interaction ways and a unified model for 3D point segmentation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 456, + 212, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 456, + 212, + 470 + ], + "spans": [ + { + "bbox": [ + 132, + 456, + 212, + 470 + ], + "type": "text", + "content": "3 Approach" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 482, + 275, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 482, + 275, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 482, + 275, + 493 + ], + "type": "text", + "content": "3.1 Architecture Overview" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 502, + 482, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 502, + 482, + 610 + ], + "spans": [ + { + "bbox": [ + 130, + 502, + 482, + 610 + ], + "type": "text", + "content": "The overall architecture of SegPoint is presented in Fig. 2. SegPoint mainly comprises four parts: " + }, + { + "bbox": [ + 130, + 502, + 482, + 610 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 502, + 482, + 610 + ], + "type": "text", + "content": " ) a pre-trained point encoder " + }, + { + "bbox": [ + 130, + 502, + 482, + 610 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 130, + 502, + 482, + 610 + ], + "type": "text", + "content": " tailored for aligning with textual data; " + }, + { + "bbox": [ + 130, + 502, + 482, + 610 + ], + "type": "inline_equation", + "content": "ii" + }, + { + "bbox": [ + 130, + 502, + 482, + 610 + ], + "type": "text", + "content": " ) a large language model " + }, + { + "bbox": [ + 130, + 502, + 482, + 610 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 130, + 502, + 482, + 610 + ], + "type": "text", + "content": " endowed with advanced reasoning capabilities; " + }, + { + "bbox": [ + 130, + 502, + 482, + 610 + ], + "type": "inline_equation", + "content": "iii" + }, + { + "bbox": [ + 130, + 502, + 482, + 610 + ], + "type": "text", + "content": " ) a Geometric Enhancer Module " + }, + { + "bbox": [ + 130, + 502, + 482, + 610 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 130, + 502, + 482, + 610 + ], + "type": "text", + "content": " responsible for extracting geometric representation from input point clouds and infusing these priors into the point encoder; and " + }, + { + "bbox": [ + 130, + 502, + 482, + 610 + ], + "type": "inline_equation", + "content": "iv" + }, + { + "bbox": [ + 130, + 502, + 482, + 610 + ], + "type": "text", + "content": " ) a Geometric-guided Feature Propagation " + }, + { + "bbox": [ + 130, + 502, + 482, + 610 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 130, + 502, + 482, + 610 + ], + "type": "text", + "content": " which is key to achieving precise mask generation. The collaboration between the Geometric Enhancer Module and Geometric-guided Feature Propagation is crucial, as it equips LLMs with the ability to generate masks effectively in various scenarios." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 621, + 241, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 241, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 241, + 632 + ], + "type": "text", + "content": "3.2 Vanilla Baseline" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "The input of the framework is the text instructions " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{i}_{\\text{txt}}" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": " and point cloud " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{i}_{\\text{point}} \\in \\mathbb{R}^{N \\times (3 + F)}" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": ". Specifically, a point cloud scene, comprising " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": " points, each includes" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 410, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 410, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 410, + 91, + 447, + 102 + ], + "type": "text", + "content": "SegPoint" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "text", + "content": "3D coordinates " + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "text", + "content": " and an auxiliary feature vector " + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\in \\mathbb{R}^F" + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "text", + "content": " (e.g., color). The point cloud " + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "inline_equation", + "content": "i_{point}" + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "text", + "content": " is fed into the point encoder " + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "text", + "content": ", which extracts point features " + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "inline_equation", + "content": "f_{point} \\in \\mathbb{R}^{N_1 \\times D}" + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "inline_equation", + "content": "N_1 \\ll N" + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "text", + "content": " is the feature dimension. Concurrently, the text instruction " + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "inline_equation", + "content": "i_{txt}" + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "text", + "content": " undergoes tokenization via " + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{tokenize}" + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "text", + "content": ". These prepared inputs are then fed into the Large Language Model " + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "text", + "content": ", resulting in a textual response " + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "text", + "content": ". The above process can be formulated as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 165, + 199, + 481, + 213 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 199, + 481, + 213 + ], + "spans": [ + { + "bbox": [ + 165, + 199, + 481, + 213 + ], + "type": "interline_equation", + "content": "\\boldsymbol {f} _ {\\text {p o i n t}} = \\mathcal {E} (\\boldsymbol {i} _ {\\text {p o i n t}}), \\quad \\boldsymbol {f} _ {\\text {t x t}} = \\mathcal {F} _ {\\text {t o k e n i z e}} (\\boldsymbol {i} _ {\\text {t x t}}), \\quad \\boldsymbol {y} = \\mathcal {F} (\\boldsymbol {f} _ {\\text {p o i n t}}, \\boldsymbol {f} _ {\\text {t x t}}). \\tag {1}", + "image_path": "31751f0ffd497c664c4b3b33b8e8749e98fc2d7d15b5f83db9fd432bb0dce3b4.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 221, + 482, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 221, + 482, + 353 + ], + "spans": [ + { + "bbox": [ + 130, + 221, + 482, + 353 + ], + "type": "text", + "content": "Building on the approach introduced by LISA [28], SegPoint enhances the segmentation capabilities of Large Language Models (LLMs) by expanding their vocabulary with a new special token, " + }, + { + "bbox": [ + 130, + 221, + 482, + 353 + ], + "type": "inline_equation", + "content": "<\\mathrm{SEG}>." + }, + { + "bbox": [ + 130, + 221, + 482, + 353 + ], + "type": "text", + "content": " This modification enables the model to recognize and predict the " + }, + { + "bbox": [ + 130, + 221, + 482, + 353 + ], + "type": "inline_equation", + "content": "<\\mathrm{SEG}>" + }, + { + "bbox": [ + 130, + 221, + 482, + 353 + ], + "type": "text", + "content": " token within the output sequence as a signal to identify segmentation targets. Upon detecting a " + }, + { + "bbox": [ + 130, + 221, + 482, + 353 + ], + "type": "inline_equation", + "content": "<\\mathrm{SEG}>" + }, + { + "bbox": [ + 130, + 221, + 482, + 353 + ], + "type": "text", + "content": " token, the corresponding output sequence belonging to " + }, + { + "bbox": [ + 130, + 221, + 482, + 353 + ], + "type": "inline_equation", + "content": "<\\mathrm{SEG}>" + }, + { + "bbox": [ + 130, + 221, + 482, + 353 + ], + "type": "text", + "content": " token is extracted and processed through an MLP layer " + }, + { + "bbox": [ + 130, + 221, + 482, + 353 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 221, + 482, + 353 + ], + "type": "text", + "content": ", generating mask embeddings " + }, + { + "bbox": [ + 130, + 221, + 482, + 353 + ], + "type": "inline_equation", + "content": "h_{seg}" + }, + { + "bbox": [ + 130, + 221, + 482, + 353 + ], + "type": "text", + "content": ". The final step involves computing each binary mask prediction " + }, + { + "bbox": [ + 130, + 221, + 482, + 353 + ], + "type": "inline_equation", + "content": "\\pmb{m} \\in \\mathbb{R}^{N}" + }, + { + "bbox": [ + 130, + 221, + 482, + 353 + ], + "type": "text", + "content": " by performing a dot product between the mask embeddings " + }, + { + "bbox": [ + 130, + 221, + 482, + 353 + ], + "type": "inline_equation", + "content": "h_{seg}" + }, + { + "bbox": [ + 130, + 221, + 482, + 353 + ], + "type": "text", + "content": " and the upsampled per-point embeddings derived from the point features " + }, + { + "bbox": [ + 130, + 221, + 482, + 353 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{point}" + }, + { + "bbox": [ + 130, + 221, + 482, + 353 + ], + "type": "text", + "content": ". The formulation of the aforementioned process is given by:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 209, + 363, + 480, + 377 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 363, + 480, + 377 + ], + "spans": [ + { + "bbox": [ + 209, + 363, + 480, + 377 + ], + "type": "interline_equation", + "content": "\\boldsymbol {h} _ {\\text {s e g}} = \\gamma (\\boldsymbol {y} _ {[ \\text {s e g} ]}), \\quad \\boldsymbol {m} = \\boldsymbol {h} _ {\\text {s e g}} \\otimes U p S. (\\boldsymbol {f} _ {\\text {p o i n t}}), \\tag {2}", + "image_path": "63e750a0c3d1bafe0939d9cf5bcae50918220fd0d6f6a7a6b67e1c32e6667198.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 385, + 482, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 385, + 482, + 517 + ], + "spans": [ + { + "bbox": [ + 130, + 385, + 482, + 517 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 385, + 482, + 517 + ], + "type": "inline_equation", + "content": "UpS" + }, + { + "bbox": [ + 130, + 385, + 482, + 517 + ], + "type": "text", + "content": " denotes the upsampling operation following PointNet++ [46] on " + }, + { + "bbox": [ + 130, + 385, + 482, + 517 + ], + "type": "inline_equation", + "content": "f_{point}" + }, + { + "bbox": [ + 130, + 385, + 482, + 517 + ], + "type": "text", + "content": ". The vanilla baseline represents an initial attempt to bridge the gap between LLMs' text comprehension and point cloud segmentation tasks. It encounters two primary issues. Firstly, the point encoder is trained on a scene-level dataset for classification achieving alignment between text and point clouds, not specifically trained for dense prediction tasks. Besides, the point encoder's first layer employs Farthest Point Sampling (FPS) [46] to reduce the point cloud to " + }, + { + "bbox": [ + 130, + 385, + 482, + 517 + ], + "type": "inline_equation", + "content": "N_1" + }, + { + "bbox": [ + 130, + 385, + 482, + 517 + ], + "type": "text", + "content": " points, risking the loss of details vital for accurate dense predictions. Secondly, the operation of directly upsampling from " + }, + { + "bbox": [ + 130, + 385, + 482, + 517 + ], + "type": "inline_equation", + "content": "N_1" + }, + { + "bbox": [ + 130, + 385, + 482, + 517 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 385, + 482, + 517 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 385, + 482, + 517 + ], + "type": "text", + "content": " points to obtain per-point embeddings is prone to losing structural information and introducing a notable degree of noise, undermining the model's efficacy in segmentation tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 536, + 307, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 307, + 548 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 307, + 548 + ], + "type": "text", + "content": "3.3 Geometric Enhancer Module" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "To adapt the pre-trained point encoder for dense prediction tasks while maintaining its superior scene recognition capability, our objective is to harness the geometric information across the entire scene to guide the further feature learning process. Drawing inspiration from recent advancements in 2D computer vision, where studies [5,41,62,64] demonstrate that convolutions enhance transformers' ability to capture local spatial information, we introduce the Geometric Enhancer Module (GEM). This module is specifically designed to grasp the local geometric contexts within point clouds while enabling the preservation of the point encoder's foundational architecture and information integrity." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "text", + "content": "S. He et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 479, + 279 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 479, + 279 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 479, + 279 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 479, + 279 + ], + "type": "image", + "image_path": "084d27946842b10c23a8f76c1674d672fa1b8c092118cfb564c8430982cad1fe.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 285, + 479, + 309 + ], + "lines": [ + { + "bbox": [ + 131, + 285, + 479, + 309 + ], + "spans": [ + { + "bbox": [ + 131, + 285, + 479, + 309 + ], + "type": "text", + "content": "Fig. 3: Architecture of the proposed (b) Geometric Enhance Module (GEM) and (c) Geometric-guided Feature Propagation (GFP) interaction with (a) Point Encoder." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 322, + 482, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 322, + 482, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 322, + 482, + 430 + ], + "type": "text", + "content": "As shown in Fig. 3, Geometric Enhancer Module " + }, + { + "bbox": [ + 130, + 322, + 482, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 130, + 322, + 482, + 430 + ], + "type": "text", + "content": " is composed of three blocks, each featuring a KPCConv [58] layer followed by BN and ReLU activation. The architecture is similar to the 2D convolutional stem [16]. We utilize KPCConv instead of vanilla convolution or linear layer here to facilitate grasping the local geometric information effectively. The resultant geometric feature, represented by " + }, + { + "bbox": [ + 130, + 322, + 482, + 430 + ], + "type": "inline_equation", + "content": "\\pmb{g}_f\\in \\mathbb{R}^{N\\times D}" + }, + { + "bbox": [ + 130, + 322, + 482, + 430 + ], + "type": "text", + "content": ", contains the features across all points, thereby supplementing the missing local information. This " + }, + { + "bbox": [ + 130, + 322, + 482, + 430 + ], + "type": "inline_equation", + "content": "\\pmb{g}_f" + }, + { + "bbox": [ + 130, + 322, + 482, + 430 + ], + "type": "text", + "content": " is then leveraged to infuse geometric insights into the point encoder's features via a cross-attention mechanism, the above process can be expressed as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 189, + 434, + 480, + 467 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 434, + 480, + 467 + ], + "spans": [ + { + "bbox": [ + 189, + 434, + 480, + 467 + ], + "type": "interline_equation", + "content": "\\boldsymbol {g} _ {f} = \\mathcal {G} (\\boldsymbol {i} _ {\\text {p o i n t}}), \\quad \\hat {\\boldsymbol {f}} _ {i} = \\boldsymbol {f} _ {i} + g _ {i} \\cdot \\operatorname {s o f t m a x} \\left(\\frac {\\boldsymbol {f} _ {i} \\boldsymbol {g} _ {f} ^ {T}}{\\sqrt {D}}\\right) \\boldsymbol {g} _ {f}, \\tag {3}", + "image_path": "18e082ce5b79fd82793f82eb1b3f6cd180c8bc6754cea044f9746fd201a9d31c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 472, + 482, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 590 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 472, + 482, + 590 + ], + "type": "inline_equation", + "content": "\\pmb{f}_i" + }, + { + "bbox": [ + 130, + 472, + 482, + 590 + ], + "type": "text", + "content": " represents the feature from the " + }, + { + "bbox": [ + 130, + 472, + 482, + 590 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 472, + 482, + 590 + ], + "type": "text", + "content": "-th block of the point encoder and " + }, + { + "bbox": [ + 130, + 472, + 482, + 590 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 472, + 482, + 590 + ], + "type": "text", + "content": " consecutive transformer layers are regarded as one block for the convenience of explanation. To fine-tune the integration of geometric information, we introduce a learnable gating factor " + }, + { + "bbox": [ + 130, + 472, + 482, + 590 + ], + "type": "inline_equation", + "content": "g_i" + }, + { + "bbox": [ + 130, + 472, + 482, + 590 + ], + "type": "text", + "content": " that modulates the balance between the attention layer's output and the input feature " + }, + { + "bbox": [ + 130, + 472, + 482, + 590 + ], + "type": "inline_equation", + "content": "\\pmb{f}_i" + }, + { + "bbox": [ + 130, + 472, + 482, + 590 + ], + "type": "text", + "content": ". This gating factor is initially set to zero, to ensure that the incorporation of geometric data does not abruptly alter " + }, + { + "bbox": [ + 130, + 472, + 482, + 590 + ], + "type": "inline_equation", + "content": "\\pmb{f}" + }, + { + "bbox": [ + 130, + 472, + 482, + 590 + ], + "type": "text", + "content": " feature distribution. Such an approach allows for the preservation and effective utilization of the point encoder's pre-trained weights. Upon processing through the Geometric Enhancer Module (GEM), the modified output of the point encoder, LLM are formulated as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 161, + 596, + 481, + 612 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 596, + 481, + 612 + ], + "spans": [ + { + "bbox": [ + 161, + 596, + 481, + 612 + ], + "type": "interline_equation", + "content": "\\hat {\\pmb {f}} _ {p o i n t} = \\mathcal {E} (\\pmb {i} _ {p o i n t}, \\pmb {g} _ {f}), \\quad \\hat {\\pmb {y}} = \\mathcal {F} (\\hat {\\pmb {f}} _ {p o i n t}, \\pmb {f} _ {t x t}), \\quad \\hat {\\pmb {h}} _ {s e g} = \\gamma (\\hat {\\pmb {y}} _ {[ s e g ]}). \\qquad (4)", + "image_path": "cc8b682d56db5c4c0537b472d638c36eb1369949c7fe723e1627ca7199a7f127.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 624, + 358, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 624, + 358, + 636 + ], + "spans": [ + { + "bbox": [ + 131, + 624, + 358, + 636 + ], + "type": "text", + "content": "3.4 Geometric-guided Feature Propagation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "Addressing the challenge of upsampling point clouds from a sparse set of " + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "inline_equation", + "content": "N_{1}" + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": " points to a denser set of " + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": " points is crucial, as direct upsampling inevitably intro-" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 410, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 410, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 410, + 91, + 447, + 102 + ], + "type": "text", + "content": "SegPoint" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "duces noise and results in information loss, leading to sub-optimal performance in segmentation tasks. To mitigate these issues, we introduce Geometric-guided Feature Propagation designed to generate high-quality per-point embeddings. Geometric features " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "\\pmb{g}_f" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": ", which carry comprehensive point information, serving as a \"gold message\" for enhancing the upsampling process. By integrating these geometric features, we aim to significantly improve the quality and accuracy of the generated dense per-point embeddings." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 482, + 308 + ], + "type": "text", + "content": "As illustrated in Fig. 3, we begin by upsampling the higher-layer features " + }, + { + "bbox": [ + 130, + 200, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\pmb{f}_3, \\pmb{f}_4" + }, + { + "bbox": [ + 130, + 200, + 482, + 308 + ], + "type": "text", + "content": " from a smaller set of points " + }, + { + "bbox": [ + 130, + 200, + 482, + 308 + ], + "type": "inline_equation", + "content": "N_1" + }, + { + "bbox": [ + 130, + 200, + 482, + 308 + ], + "type": "text", + "content": " to larger sets " + }, + { + "bbox": [ + 130, + 200, + 482, + 308 + ], + "type": "inline_equation", + "content": "N_3, N_2" + }, + { + "bbox": [ + 130, + 200, + 482, + 308 + ], + "type": "text", + "content": ", employing Point-Net++'s [46] propagation techniques. This step yields features " + }, + { + "bbox": [ + 130, + 200, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\pmb{f}_3' \\in \\mathbb{R}^{N_3 \\times D}" + }, + { + "bbox": [ + 130, + 200, + 482, + 308 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 200, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\pmb{f}_4' \\in \\mathbb{R}^{N_2 \\times D}" + }, + { + "bbox": [ + 130, + 200, + 482, + 308 + ], + "type": "text", + "content": ". Subsequently, we perform downsampling on the geometric features " + }, + { + "bbox": [ + 130, + 200, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\pmb{g}_f" + }, + { + "bbox": [ + 130, + 200, + 482, + 308 + ], + "type": "text", + "content": " from the original number of points " + }, + { + "bbox": [ + 130, + 200, + 482, + 308 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 200, + 482, + 308 + ], + "type": "text", + "content": " to reduced counts " + }, + { + "bbox": [ + 130, + 200, + 482, + 308 + ], + "type": "inline_equation", + "content": "N_2, N_3" + }, + { + "bbox": [ + 130, + 200, + 482, + 308 + ], + "type": "text", + "content": ", respectively, utilizing the Farthest Point Sampling (FPS) technique. In this process, we directly obtain the features of the sampled points without performing additional " + }, + { + "bbox": [ + 130, + 200, + 482, + 308 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 200, + 482, + 308 + ], + "type": "text", + "content": "-nearest neighbor (" + }, + { + "bbox": [ + 130, + 200, + 482, + 308 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 200, + 482, + 308 + ], + "type": "text", + "content": "-NN) and pooling operations to simplify the computation and produce features " + }, + { + "bbox": [ + 130, + 200, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{g,1} \\in \\mathbb{R}^{N_3 \\times D}" + }, + { + "bbox": [ + 130, + 200, + 482, + 308 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 200, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{g,2} \\in \\mathbb{R}^{N_2 \\times D}" + }, + { + "bbox": [ + 130, + 200, + 482, + 308 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 308, + 482, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 482, + 369 + ], + "type": "text", + "content": "In the next phase, we integrate the up- and downsampled features, processing them through fully connected layers and ReLU activation to update the feature vectors " + }, + { + "bbox": [ + 130, + 308, + 482, + 369 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{f}}_3\\in \\mathbb{R}^{N_3\\times D}" + }, + { + "bbox": [ + 130, + 308, + 482, + 369 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 308, + 482, + 369 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{f}}_4\\in \\mathbb{R}^{N_2\\times D}" + }, + { + "bbox": [ + 130, + 308, + 482, + 369 + ], + "type": "text", + "content": ". Note that the last layer feature " + }, + { + "bbox": [ + 130, + 308, + 482, + 369 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{5}" + }, + { + "bbox": [ + 130, + 308, + 482, + 369 + ], + "type": "text", + "content": " bypasses this step. Instead, we concatenate it with " + }, + { + "bbox": [ + 130, + 308, + 482, + 369 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{h}}_{point}" + }, + { + "bbox": [ + 130, + 308, + 482, + 369 + ], + "type": "text", + "content": " from the LLM output to form " + }, + { + "bbox": [ + 130, + 308, + 482, + 369 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{f}}_5" + }, + { + "bbox": [ + 130, + 308, + 482, + 369 + ], + "type": "text", + "content": " to perceive multi-modal information from LLM." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 370, + 482, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 370, + 482, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 370, + 482, + 453 + ], + "type": "text", + "content": "Finally, to enable information exchange across different point densities, we propose attentive propagation. Take the propagation from " + }, + { + "bbox": [ + 130, + 370, + 482, + 453 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{f}}_5" + }, + { + "bbox": [ + 130, + 370, + 482, + 453 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 370, + 482, + 453 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{f}}_4" + }, + { + "bbox": [ + 130, + 370, + 482, + 453 + ], + "type": "text", + "content": " as example. Here, " + }, + { + "bbox": [ + 130, + 370, + 482, + 453 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{f}}_4\\in \\mathbb{R}^{N_2\\times D}" + }, + { + "bbox": [ + 130, + 370, + 482, + 453 + ], + "type": "text", + "content": " acts as a set of local centers. For each local center within " + }, + { + "bbox": [ + 130, + 370, + 482, + 453 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{f}}_4" + }, + { + "bbox": [ + 130, + 370, + 482, + 453 + ], + "type": "text", + "content": ", we identify its neighboring points from " + }, + { + "bbox": [ + 130, + 370, + 482, + 453 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{f}}_5" + }, + { + "bbox": [ + 130, + 370, + 482, + 453 + ], + "type": "text", + "content": " using the " + }, + { + "bbox": [ + 130, + 370, + 482, + 453 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 370, + 482, + 453 + ], + "type": "text", + "content": "-NN algorithm, resulting in " + }, + { + "bbox": [ + 130, + 370, + 482, + 453 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{54}\\in \\mathbb{R}^{N_2\\times k\\times D}" + }, + { + "bbox": [ + 130, + 370, + 482, + 453 + ], + "type": "text", + "content": ". Then, employing cross-attention mechanism, where " + }, + { + "bbox": [ + 130, + 370, + 482, + 453 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{f}}_4" + }, + { + "bbox": [ + 130, + 370, + 482, + 453 + ], + "type": "text", + "content": " serves as query and " + }, + { + "bbox": [ + 130, + 370, + 482, + 453 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{54}" + }, + { + "bbox": [ + 130, + 370, + 482, + 453 + ], + "type": "text", + "content": " as both key and value, facilitates information flow across different point densities and effectively extract relevant details into the query points." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 233, + 460, + 481, + 492 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 460, + 481, + 492 + ], + "spans": [ + { + "bbox": [ + 233, + 460, + 481, + 492 + ], + "type": "interline_equation", + "content": "\\hat {\\tilde {\\boldsymbol {f}}} _ {4} = \\tilde {\\boldsymbol {f}} _ {4} + \\operatorname {s o f t m a x} \\left(\\frac {\\tilde {\\boldsymbol {f}} _ {4} \\boldsymbol {f} _ {5 4} ^ {T}}{\\sqrt {D}}\\right) \\boldsymbol {f} _ {5 4}. \\tag {5}", + "image_path": "0b547eadc6b3bec911f1150bff0dc6c7c7da08d3e705a5d362cc7e2c12b197ad.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 498, + 481, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 481, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 481, + 534 + ], + "type": "text", + "content": "Leveraging geometric-guided feature propagation enables us to produce high-quality per-point embeddings denoted as " + }, + { + "bbox": [ + 130, + 498, + 481, + 534 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{\\mathcal{P}}" + }, + { + "bbox": [ + 130, + 498, + 481, + 534 + ], + "type": "text", + "content": ", laying the foundation for generating precise segmentation masks expressed as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 222, + 542, + 481, + 556 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 542, + 481, + 556 + ], + "spans": [ + { + "bbox": [ + 222, + 542, + 481, + 556 + ], + "type": "interline_equation", + "content": "\\boldsymbol {f} _ {\\mathcal {P}} = \\mathcal {P} \\left(\\hat {\\boldsymbol {f}} _ {\\text {p o i n t}}, \\boldsymbol {g} _ {f}\\right), \\quad \\hat {\\boldsymbol {m}} = \\hat {\\boldsymbol {h}} _ {\\text {s e g}} \\otimes \\boldsymbol {f} _ {\\mathcal {P}}. \\tag {6}", + "image_path": "3fe22472747635a902c804c09d1af59304cb2ef7d1d0619e39a503e82d247dfc.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 571, + 261, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 571, + 261, + 583 + ], + "spans": [ + { + "bbox": [ + 131, + 571, + 261, + 583 + ], + "type": "text", + "content": "3.5 Training Objectives" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 589, + 481, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 589, + 481, + 613 + ], + "spans": [ + { + "bbox": [ + 130, + 589, + 481, + 613 + ], + "type": "text", + "content": "Our model is trained end-to-end leveraging the text classification loss and the segmentation mask loss:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 227, + 622, + 481, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 622, + 481, + 634 + ], + "spans": [ + { + "bbox": [ + 227, + 622, + 481, + 634 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda_ {t x t} \\mathcal {L} _ {t x t} + \\lambda_ {b c e} \\mathcal {L} _ {b c e} + \\lambda_ {d i c e} \\mathcal {L} _ {d i c e}, \\tag {7}", + "image_path": "c62026472b071efcc239315cbe923803fe85abb4ad807a676c718db22b0b726c.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{txt}" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": " denotes the auto-regressive cross-entropy loss targeting text generation accuracy, segmentation mask loss includes both binary cross-entropy (BCE)" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "S. He et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "type": "text", + "content": "loss " + }, + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{bce}" + }, + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "type": "text", + "content": " and DICE loss " + }, + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{dice}" + }, + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "type": "text", + "content": ", aims at refining segmentation quality. The weights " + }, + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "type": "inline_equation", + "content": "\\lambda_{txt}" + }, + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "type": "inline_equation", + "content": "\\lambda_{bce}" + }, + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "type": "inline_equation", + "content": "\\lambda_{dice}" + }, + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "type": "text", + "content": " are utilized to balance the different loss items. The model's training is guided by the ground-truth labels " + }, + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "type": "inline_equation", + "content": "\\pmb{y}_{txt}" + }, + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "type": "text", + "content": " for text and " + }, + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "type": "inline_equation", + "content": "\\pmb{M}" + }, + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "type": "text", + "content": " for masks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 169, + 313, + 180 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 169, + 313, + 180 + ], + "spans": [ + { + "bbox": [ + 131, + 169, + 313, + 180 + ], + "type": "text", + "content": "3.6 Instruct3D Dataset Collection" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 188, + 481, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 481, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 481, + 272 + ], + "type": "text", + "content": "Although 3D instruction segmentation and 3D referring segmentation [1,4,79] are both language-based segmentation, 3D referring segmentation guides segmentation with explicit target object names, e.g., \"chair\", lacking more complicated reasoning instructions, e.g., \"Where to sit in the room?\". Besides, they also fall short in offering multi-target question-answer pairs with target descriptions directly connected to multiple segmentation masks, which cannot meet a common requirement in real-world scenarios, like \"How to play computer games\"." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 272, + 482, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 272, + 482, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 482, + 464 + ], + "type": "text", + "content": "To enhance the assessment and analysis of instruction segmentation capabilities, we have developed a benchmark, referred to Instruct3D. This benchmark incorporates 280 scenes specifically selected for instruction segmentation tuning and evaluation, sourced from the recently introduced ScanNet++ [74] dataset. Each scene comes with approximately 10 different segmentation instructions, resulting in 2,565 instruction-point cloud pairings. This dataset is then divided into two splits: train, and val, containing 2,052, and 513 question-answer pairs, respectively. Our dataset is uniquely designed to encompass both multi-target and zero-target scenarios, addressing the real-world requirement of identifying multiple objects in response to text queries and accounting for situations where objects mentioned in the text may not be present in the paired point cloud. Besides, we take into account the characteristics of 3D scenes and incorporate diverse locations and view descriptions e.g., \"something that is used for sitting while working at a desk. It is the one facing the window.\" The model needs to have not only reasoning capabilities but also the ability to perceive views and directions in 3D scenes. These designs underscore the dataset's practical value." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 481, + 230, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 481, + 230, + 495 + ], + "spans": [ + { + "bbox": [ + 131, + 481, + 230, + 495 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 506, + 326, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 506, + 326, + 517 + ], + "spans": [ + { + "bbox": [ + 131, + 506, + 326, + 517 + ], + "type": "text", + "content": "4.1 Datasets and Evaluation Metrics" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 521, + 481, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 481, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 481, + 582 + ], + "type": "text", + "content": "Datasets. Our training data is composed of two types of datasets: (1) semantic segmentation dataset including ScanNet200 [53], and S3DIS [3]; (2) referring segmentation dataset consisting of ScanRefer [4], ReferIt3D [1](including Sr3D and Nr3D), and Multi3DRefer [79]. We design task-specific prompts to facilitate the joint training of various tasks within a unified framework." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": "Semantic Segmentation Dataset. We use two strategies to generate templates. 1) segment the specific category: \"USER: Can you segment the {category} category in this point cloud? ASSISTANT: {category} ,\" where category is the random chosen category, and denotes the placeholder for tokens of point cloud patches. 2) segment all the categories: \"USER: Can you segment all the semantic masks in this point cloud and output separate masks for each category in the alphabetical order of the" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 410, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 410, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 410, + 91, + 447, + 102 + ], + "type": "text", + "content": "SegPoint" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 174 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 174 + ], + "type": "text", + "content": "categories? ASSISTANT: {category} , {category} , ...” To simplify the output and ensure it has only one possible answer, we add the constraints “in the alphabetical order of the categories”. To avoid generating class names not in the dataset, we incorporate category names in a dataset into the prompts during training and inference." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 175, + 479, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 175, + 479, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 175, + 479, + 258 + ], + "type": "text", + "content": "Referring Segmentation Dataset. We use template prompts: \"USER: Can you segment the object {description} in this point cloud? ASSISTANT: {category} ,\" where {description} is the given explicit description from referring segmentation dataset. It is worth noting that during training, we also use other templates to generate the QA data to ensure data diversity. We add {category} in front of to unify the output format so that when outputting semantic masks, the output category name is the label it predicts." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 259, + 479, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 479, + 318 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 479, + 318 + ], + "type": "text", + "content": "Evaluation Metrics. We follow most previous works on 3D segmentation [22, 27, 54] to adopt mIoU as primary metric. mIoU is defined by the average of all per-point cloud scene Intersection-over-Unions (IoUs). Besides, we employ accuracy (Acc) as a metric to evaluate whether the model accurately identifies targets with which the predictions have an IoU greater than 0.5." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 326, + 279, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 326, + 279, + 337 + ], + "spans": [ + { + "bbox": [ + 131, + 326, + 279, + 337 + ], + "type": "text", + "content": "4.2 Implementation Details" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 342, + 480, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 480, + 497 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 480, + 497 + ], + "type": "text", + "content": "In our experiments, unless specified otherwise, we employ the LLaMA2-7B model [59] as the large language model " + }, + { + "bbox": [ + 130, + 342, + 480, + 497 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 130, + 342, + 480, + 497 + ], + "type": "text", + "content": " and Uni3D [81] as the point cloud processing backbone " + }, + { + "bbox": [ + 130, + 342, + 480, + 497 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 130, + 342, + 480, + 497 + ], + "type": "text", + "content": ". The training stage leverages the deepspeed [51] engine for efficiency, with the AdamW [39] optimizer guiding the learning process. The learning rate and weight decay are set to 0.0003 and 0, respectively, enhanced by a WarmupDecayLR learning rate scheduler that initiates with 100 warmup iterations. The projection layer " + }, + { + "bbox": [ + 130, + 342, + 480, + 497 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 342, + 480, + 497 + ], + "type": "text", + "content": " utilizes an MLP with channel sizes of [256, 4096, 4096]. We set balancing weight " + }, + { + "bbox": [ + 130, + 342, + 480, + 497 + ], + "type": "inline_equation", + "content": "\\lambda_{txt\\_gen}" + }, + { + "bbox": [ + 130, + 342, + 480, + 497 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 342, + 480, + 497 + ], + "type": "inline_equation", + "content": "\\lambda_{bce}" + }, + { + "bbox": [ + 130, + 342, + 480, + 497 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 342, + 480, + 497 + ], + "type": "inline_equation", + "content": "\\lambda_{dice}" + }, + { + "bbox": [ + 130, + 342, + 480, + 497 + ], + "type": "text", + "content": " to 1.0, 2.0, 2.0, respectively. The experiments utilize a total batch size of 16, distributed across 4 NVIDIA 80G A100 GPUs, and span 5,000 iterations, culminating in a training period of approximately 3 days. During training, we use all mentioned datasets in Sec. 4.1 for joint training by leveraging task-specific prompts. For evaluation on a specific dataset, we finetune the trained model on the corresponding dataset." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 505, + 272, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 505, + 272, + 516 + ], + "spans": [ + { + "bbox": [ + 131, + 505, + 272, + 516 + ], + "type": "text", + "content": "4.3 Results on Instruct3D" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 521, + 479, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 479, + 664 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 479, + 664 + ], + "type": "text", + "content": "The instruction segmentation results, as detailed in Table 1, underscore a significant advancement: where existing methodologies fall short, our model demonstrates exceptional prowess, achieving a more than " + }, + { + "bbox": [ + 130, + 521, + 479, + 664 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 130, + 521, + 479, + 664 + ], + "type": "text", + "content": " improvement in mIoU for tasks requiring intricate reasoning. Unlike conventional referring segmentation tasks, instruction segmentation demands not just identification but also understanding, necessitating the model's reasoning capabilities and access to world knowledge. Existing approaches, confined to explicit references, struggle with implicit queries due to their lack of understanding, which further underscores the task's inherent challenges. In contrast, our model leverages LLMs to bridge this gap, demonstrating superior performance by comprehending and interpreting the queries accurately. Moreover, SegPoint configuration substantially outperforms " + }, + { + "bbox": [ + 130, + 521, + 479, + 664 + ], + "type": "inline_equation", + "content": "\\mathrm{SegPoint}^{\\dagger}" + }, + { + "bbox": [ + 130, + 521, + 479, + 664 + ], + "type": "text", + "content": ", highlighting the critical role of our designed Geometric" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "S. He et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 144, + 160, + 470, + 244 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 158 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 158 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 158 + ], + "type": "text", + "content": "Table 1: 3D instruction segmentation benchmark results on Instruct3D val split evaluated by Acc and mIoU. † denotes our vanilla baseline removing geometric enhancer module and geometric-guided feature propagation. * represents adding an auxiliary mask head through our implementation." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 144, + 160, + 470, + 244 + ], + "lines": [ + { + "bbox": [ + 144, + 160, + 470, + 244 + ], + "spans": [ + { + "bbox": [ + 144, + 160, + 470, + 244 + ], + "type": "table", + "html": "
StageMethodReferenceAccmIoU
TwoScanRefer [4][ECCV'20]12.06.9
TwoReferIt3D [1][ECCV'20]11.76.4
TwoM3DRef-CLIP [79][ICCV'23]18.112.8
SingleTGNN [22][AAAI'21]12.97.1
SingleBUTD-DETR [23]*[ECCV'22]16.310.9
SingleEDA [67]*[CVPR'23]16.612.1
SingleSegPoint†[ECCV'24]21.816.1
SingleSegPoint[ECCV'24]31.627.5
", + "image_path": "37b1abd7acfbd466834013166a7736ff01b6b3d21873657aae79a4a5a03cdaa3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 140, + 288, + 472, + 372 + ], + "blocks": [ + { + "bbox": [ + 130, + 254, + 480, + 285 + ], + "lines": [ + { + "bbox": [ + 130, + 254, + 480, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 254, + 480, + 285 + ], + "type": "text", + "content": "Table 2: 3D Semantic segmentation benchmark results on S3DIS [3], ScanNet [7], and ScanNet200 [53]. We evaluate on the Area 5 of S3DIS and validation split of ScanNet and ScanNet200." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 140, + 288, + 472, + 372 + ], + "lines": [ + { + "bbox": [ + 140, + 288, + 472, + 372 + ], + "spans": [ + { + "bbox": [ + 140, + 288, + 472, + 372 + ], + "type": "table", + "html": "
MethodReferenceScanNetScanNet200S3DIS
PointNet++ [46][NeurIPS'17]53.5--
MinkUNet [6][CVPR'19]72.225.065.4
PTv1 [80][ICCV'21]70.627.870.4
PTv2 [66][NeurIPS'22]75.430.271.6
PointNeXt [48][NeurIPS'22]71.5-70.5
OctFormer [60][SIGGRAPH'23]75.732.6-
Swin3D [72][ArXiv]75.5-72.5
SegPoint[ECCV'24]74.135.372.4
", + "image_path": "63eba471c420636cbd6ca462e4011389cd2a3b32b286114f8059ac0967ac0b41.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 379, + 480, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 379, + 480, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 379, + 480, + 415 + ], + "type": "text", + "content": "Enhancer Module and Geometric-guided Feature Propagation components. Notably, even in its baseline form, " + }, + { + "bbox": [ + 130, + 379, + 480, + 415 + ], + "type": "inline_equation", + "content": "\\mathrm{SegPoint}^{\\dagger}" + }, + { + "bbox": [ + 130, + 379, + 480, + 415 + ], + "type": "text", + "content": " surpasses all competing methods, validating the effectiveness and rationale behind our pipeline design." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 415, + 481, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 481, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 481, + 464 + ], + "type": "text", + "content": "Besides, different from traditional two-stage approaches that first generate mask proposals using a pre-trained segmentor like Mask3D [54] and then apply language-aware networks for selection, SegPoint directly tackles the task, bypassing the need for preliminary mask proposals, enhancing its efficiency." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 468, + 335, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 468, + 335, + 480 + ], + "spans": [ + { + "bbox": [ + 131, + 468, + 335, + 480 + ], + "type": "text", + "content": "4.4 Results on Semantic Segmentation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 485, + 482, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 482, + 570 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 482, + 570 + ], + "type": "text", + "content": "Table 2 present SegPoint's performance on semantic segmentation, delivering competitive results across diverse datasets. Our model uses a simple yet effective answer format, category " + }, + { + "bbox": [ + 130, + 485, + 482, + 570 + ], + "type": "inline_equation", + "content": "<\\mathrm{SEG}>" + }, + { + "bbox": [ + 130, + 485, + 482, + 570 + ], + "type": "text", + "content": ", to use category name as predicted labels, achieving particularly stronger performance in datasets with various categories like ScanNet200 [53], where it surpasses SOTA methods by " + }, + { + "bbox": [ + 130, + 485, + 482, + 570 + ], + "type": "inline_equation", + "content": "2.1\\%" + }, + { + "bbox": [ + 130, + 485, + 482, + 570 + ], + "type": "text", + "content": " mIoU. To ensure fair comparisons, we fine-tune our model on each semantic segmentation dataset to accommodate the varying class category definitions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 574, + 338, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 574, + 338, + 586 + ], + "spans": [ + { + "bbox": [ + 131, + 574, + 338, + 586 + ], + "type": "text", + "content": "4.5 Results on Referring Segmentation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 591, + 482, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 591, + 482, + 675 + ], + "spans": [ + { + "bbox": [ + 130, + 591, + 482, + 675 + ], + "type": "text", + "content": "Table 3 presents results on referring segmentation datasets. SegPoint showcases outstanding performance in both single-target (e.g., ScanRefer [4], Nr3D [1]) and multi-target and zero-target contexts within the Multi3DRefer [79] dataset. For multi-targets, we aggregate masks into a single ground truth, and for zero-target, we use an empty mask, indicated by \"ASSISTANT: There is no mask.\" SegPoint significantly surpasses other approaches, achieving " + }, + { + "bbox": [ + 130, + 591, + 482, + 675 + ], + "type": "inline_equation", + "content": "2.5\\%" + }, + { + "bbox": [ + 130, + 591, + 482, + 675 + ], + "type": "text", + "content": " mIoU increase. The evaluation process of two-stage method M3DRef-CLIP is similar to Sec.4.3." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 410, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 410, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 410, + 91, + 447, + 102 + ], + "type": "text", + "content": "SegPoint" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 149, + 482, + 232 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 3: 3D referring segmentation benchmark results on ScanRefer [4], Nr3D [1], and Multi3Drefer [79] evaluated by mIoU. * represents adding an auxiliary mask head through our implementation." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 149, + 482, + 232 + ], + "lines": [ + { + "bbox": [ + 133, + 149, + 482, + 232 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 482, + 232 + ], + "type": "table", + "html": "
StageMethodReferenceScanReferNr3DMulti3DRefer
TwoM3DRef-CLIP [79][ICCV'23]35.727.032.6
Two3D-STMN [63][AAAI'24]39.5--
SingleTGNN [22][AAAI'21]27.8--
SingleBUTD-DETR [23]*[ECCV'22]35.427.526.2
SingleEDA [67]*[CVPR'23]36.229.328.9
SingleX-RefSeg3D [49][AAAI'24]29.9--
SingleRefMask3D [18][ACMMM'24]44.8--
SingleSegPoint[ECCV'24]41.732.236.1
", + "image_path": "63931d6d9e89f271d82b1bbea49b2ed6de9200a61082e695f1523572d4f736e0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 136, + 267, + 473, + 350 + ], + "blocks": [ + { + "bbox": [ + 132, + 243, + 481, + 266 + ], + "lines": [ + { + "bbox": [ + 132, + 243, + 481, + 266 + ], + "spans": [ + { + "bbox": [ + 132, + 243, + 481, + 266 + ], + "type": "text", + "content": "Table 4: 3D open-vocabulary semantic segmentation benchmark results on val split of ScanNet++ [74]." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 267, + 473, + 350 + ], + "lines": [ + { + "bbox": [ + 136, + 267, + 473, + 350 + ], + "spans": [ + { + "bbox": [ + 136, + 267, + 473, + 350 + ], + "type": "table", + "html": "
TypeMethodReferenceScanNet++
SupervisedPointNet [45][CVPR'17]7.0
PointNet++ [46][NeurIPS'17]15.0
MinkUNet [6][CVPR'19]28.0
KPCConv [58][ICCV'19]30.0
Open-VocabularyOpenScene [42][CVPR'23]12.8
PLA [12][CVPR'23]14.2
RegionPLC [71][CVPR'24]14.9
UnifiedSegPoint[ECCV'24]19.3
", + "image_path": "a64748139057ad34afa4b22616d4fdc85278df4840627e1944ba0c26a79e3bdb.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 359, + 424, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 359, + 424, + 371 + ], + "spans": [ + { + "bbox": [ + 132, + 359, + 424, + 371 + ], + "type": "text", + "content": "4.6 Results on Open-vocabulary Semantic Segmentation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 378, + 482, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 378, + 482, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 378, + 482, + 486 + ], + "type": "text", + "content": "Table 4 shows our method's open-vocabulary segmentation performance which is directly evaluated on ScanNet++ [74] following the setting in prevalent methodologies in the 2D domain [13,69]. It demonstrates our superiority over both existing open-vocabulary techniques and even several supervised approaches, showing our model's robust generalization capabilities. It effectively aligns and interprets categories with visual scenes, underscoring the reasoning prowess of large language models. A notable issue is the potential misalignment between output categories of SegPoint and val split category names. To address this, we employ GPT-4 to match its most similar category names in the val split." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 491, + 238, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 238, + 503 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 238, + 503 + ], + "type": "text", + "content": "4.7 Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 510, + 482, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 592 + ], + "type": "text", + "content": "We conduct extensive experiments to verify the effectiveness of our proposed components in Table 5 (a) on both Instruct3D and ScanRefer [4] dataset. We established a vanilla baseline as described in Sec. 3.2 following the paradigm of LISA [28], which cannot achieve satisfactory performance and only obtain " + }, + { + "bbox": [ + 130, + 510, + 482, + 592 + ], + "type": "inline_equation", + "content": "16.1\\% / 30.3 \\, \\mathrm{mIoU}" + }, + { + "bbox": [ + 130, + 510, + 482, + 592 + ], + "type": "text", + "content": " on Instruct3D/ScanRefer, respectively. Further analysis, both qualitative and quantitative, of our proposed components reveals that their integration substantially outperforms the baseline." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Geometric Enhancer Module Integrating the Geometric Enhancer Module (GEM) into our point encoder results in a notable " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "5.3\\% / 5.5\\%" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " mIoU improvement on Instruct3D/ScanRefer, effectively addressing compatibility issues with dense prediction tasks. An ablation study, shown in Table 5 (b), shows that our improvement is not due to an increase in parameters. Our approach outperforms traditional full fine-tuning, LoRA [21] strategies, and the addition of" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "S. He et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 143, + 135, + 294, + 193 + ], + "blocks": [ + { + "bbox": [ + 184, + 114, + 427, + 125 + ], + "lines": [ + { + "bbox": [ + 184, + 114, + 427, + 125 + ], + "spans": [ + { + "bbox": [ + 184, + 114, + 427, + 125 + ], + "type": "text", + "content": "Table 5: Ablation studies on Instruct3D and ScanRefer." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 143, + 135, + 294, + 193 + ], + "lines": [ + { + "bbox": [ + 143, + 135, + 294, + 193 + ], + "spans": [ + { + "bbox": [ + 143, + 135, + 294, + 193 + ], + "type": "table", + "html": "
IndexComponentsInstruct3D\nmIoUScanRefer\nmIoU
GEMGFP
0XX16.130.3
1✓X21.435.8
2X✓23.238.1
3✓✓27.541.7
", + "image_path": "8854cbebaaf3c9caeb7c11ebdc6e5adec24b15f2c6a0d4939192d2f59d702967.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 149, + 194, + 292, + 203 + ], + "lines": [ + { + "bbox": [ + 149, + 194, + 292, + 203 + ], + "spans": [ + { + "bbox": [ + 149, + 194, + 292, + 203 + ], + "type": "text", + "content": "(a) Effect of different main components." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 329, + 139, + 457, + 189 + ], + "blocks": [ + { + "bbox": [ + 329, + 139, + 457, + 189 + ], + "lines": [ + { + "bbox": [ + 329, + 139, + 457, + 189 + ], + "spans": [ + { + "bbox": [ + 329, + 139, + 457, + 189 + ], + "type": "table", + "html": "
MethodInstruct3DScanRefer
Full tuning26.239.1
LoRA [21]24.838.5
MLP23.136.8
GEM27.541.7
", + "image_path": "c5e35d614a8a56dabd59d11e08ed0282068b3f70159ce735681869801aff6142.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 327, + 189, + 463, + 199 + ], + "lines": [ + { + "bbox": [ + 327, + 189, + 463, + 199 + ], + "spans": [ + { + "bbox": [ + 327, + 189, + 463, + 199 + ], + "type": "text", + "content": "(b) Effect of different tuning methods." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 135, + 211, + 236, + 266 + ], + "blocks": [ + { + "bbox": [ + 135, + 211, + 236, + 266 + ], + "lines": [ + { + "bbox": [ + 135, + 211, + 236, + 266 + ], + "spans": [ + { + "bbox": [ + 135, + 211, + 236, + 266 + ], + "type": "image", + "image_path": "51e47c303ec60ce760ff9fb71819deb3f1c4098c4be0ff84843e4d7a33669bc8.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 178, + 266, + 203, + 273 + ], + "lines": [ + { + "bbox": [ + 178, + 266, + 203, + 273 + ], + "spans": [ + { + "bbox": [ + 178, + 266, + 203, + 273 + ], + "type": "text", + "content": "(a) Scene" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 255, + 214, + 358, + 266 + ], + "blocks": [ + { + "bbox": [ + 255, + 214, + 358, + 266 + ], + "lines": [ + { + "bbox": [ + 255, + 214, + 358, + 266 + ], + "spans": [ + { + "bbox": [ + 255, + 214, + 358, + 266 + ], + "type": "image", + "image_path": "4964324b8e48caa54aaba3259917eb79329a6315d1f68c79b16146e456d1fd6f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 296, + 266, + 328, + 273 + ], + "lines": [ + { + "bbox": [ + 296, + 266, + 328, + 273 + ], + "spans": [ + { + "bbox": [ + 296, + 266, + 328, + 273 + ], + "type": "text", + "content": "(b) Baseline" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 278, + 482, + 322 + ], + "lines": [ + { + "bbox": [ + 130, + 278, + 482, + 322 + ], + "spans": [ + { + "bbox": [ + 130, + 278, + 482, + 322 + ], + "type": "text", + "content": "Fig. 4: (Best viewed in color) We visualize the feature responses between a given point (in red) and other points in the scene from per-point embeddings " + }, + { + "bbox": [ + 130, + 278, + 482, + 322 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{\\mathcal{P}}" + }, + { + "bbox": [ + 130, + 278, + 482, + 322 + ], + "type": "text", + "content": " for the baseline and our SegPoint, respectively. The color changes from yellow to red, indicating increasing feature similarity." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 378, + 214, + 479, + 266 + ], + "blocks": [ + { + "bbox": [ + 378, + 214, + 479, + 266 + ], + "lines": [ + { + "bbox": [ + 378, + 214, + 479, + 266 + ], + "spans": [ + { + "bbox": [ + 378, + 214, + 479, + 266 + ], + "type": "image", + "image_path": "3fa27e0764c2c5533d83c3f73c921b9fb735de1198169d78d22860929c027981.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 423, + 266, + 445, + 273 + ], + "lines": [ + { + "bbox": [ + 423, + 266, + 445, + 273 + ], + "spans": [ + { + "bbox": [ + 423, + 266, + 445, + 273 + ], + "type": "text", + "content": "(c) Ours" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 338, + 480, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 338, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 480, + 385 + ], + "type": "text", + "content": "MLP layers for feature adapter, underscoring its effectiveness in embedding 3D domain-specific knowledge into point encoder. Unlike adapter techniques common in language and 2D image processing, our GEM is designed to address the unique challenges in dense prediction tasks." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 384, + 480, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 480, + 444 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 480, + 444 + ], + "type": "text", + "content": "Geometric-guided Feature Propagation Introducing Geometric-guided Feature Propagation (GFP) results in a substantial improvement over the baseline, as shown in Table 5 (a) (index 2). This underscores our method's capability to minimize information loss and reduce noise during the upsampling phase, leading to higher-quality per-point embeddings." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 131, + 452, + 285, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 452, + 285, + 464 + ], + "spans": [ + { + "bbox": [ + 131, + 452, + 285, + 464 + ], + "type": "text", + "content": "4.8 Qualitative Visualization" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 471, + 482, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 471, + 482, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 471, + 482, + 578 + ], + "type": "text", + "content": "Fig. 4 qualitatively illustrates the feature responses between a given point and others in a scene. From left to right, it presents the original scene, baseline method, and our SegPoint, respectively, with warmer colors indicating closer feature relationships. The baseline method predominantly highlights spatially proximate points, often missing important but distant features, such as chair cushions. In contrast, SegPoint leverages global information, allowing for the identification of both near and distant relevant features, such as cushions and casters. This demonstrates our model's superior ability to capture global context and recognize intricate structures within the scene." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 579, + 482, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 579, + 482, + 675 + ], + "spans": [ + { + "bbox": [ + 130, + 579, + 482, + 675 + ], + "type": "text", + "content": "As shown in Fig. 5, we provide some typical qualitative results from SegPoint on Instruct3D. Given an implicit instruction, e.g., \"the object that is used for dispensing water during bathing\", SegPoint successfully infers the role of a shower. In other scenes, the instructions include queries requiring extensive knowledge of the world and complex reasoning, like \"What appliance is used for heating or cooking food quickly using electromagnetic radiation.\" SegPoint can still segment the microwave oven very well, which shows that SegPoint can make good use of the reasoning ability of LLM and provide high-quality masks." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 410, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 410, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 410, + 91, + 447, + 101 + ], + "type": "text", + "content": "SegPoint" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 116, + 481, + 403 + ], + "blocks": [ + { + "bbox": [ + 134, + 116, + 481, + 403 + ], + "lines": [ + { + "bbox": [ + 134, + 116, + 481, + 403 + ], + "spans": [ + { + "bbox": [ + 134, + 116, + 481, + 403 + ], + "type": "image", + "image_path": "e476decbca80d6647d5c0d0f8dc189a75d2ac5cba270fa39298ad535ccb4c3bf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 408, + 482, + 443 + ], + "lines": [ + { + "bbox": [ + 130, + 408, + 482, + 443 + ], + "spans": [ + { + "bbox": [ + 130, + 408, + 482, + 443 + ], + "type": "text", + "content": "Fig. 5: Qualitative results from val split of Instruct3D. SegPoint understand the human instruction and accurately segment the target object. We omitted the \"please output segmentation mask\" in the sentence for simplicity." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 449, + 220, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 449, + 220, + 462 + ], + "spans": [ + { + "bbox": [ + 132, + 449, + 220, + 462 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 473, + 482, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 482, + 570 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 482, + 570 + ], + "type": "text", + "content": "In this work, we introduce SegPoint, an effective model supported by LLM for point-level reasoning and segmentation. Benefiting from the proposed Geometric Enhancer Module and Geometric-guided Feature Propagation, SegPoint is adept at solving a variety of segmentation tasks in a unified framework. Additionally, we construct a comprehensive Instruct3D benchmark to bolster research area in segmentation via implicit and complex instructions, introducing more challenges to promote it closer to real-world applications. Through thorough experiments, SegPoint achieves promising results across multiple benchmarks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 573, + 482, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 573, + 482, + 620 + ], + "spans": [ + { + "bbox": [ + 130, + 573, + 482, + 620 + ], + "type": "text", + "content": "Limitations Although SegPoint demonstrates notable success in tasks driven by text prompts, its current framework cannot process non-textual prompts, such as boxes and points. Future developments will explore the adoption of a prompt encoder, inspired by SAM model [26], to extend support for these formats." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 621, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 621, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 621, + 482, + 666 + ], + "type": "text", + "content": "Acknowledgements We thank the anonymous reviewers for their constructive suggestions. Following their advice, we have incorporated diverse location and view descriptions into our Instruct3D. This work was partially supported by the National Research Foundation Singapore Competitive Research Program (CRP29-2022-0003)." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "S. He et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "type": "text", + "content": "1. Achlioptas, P., Abdelreheem, A., Xia, F., Elhoseiny, M., Guibas, L.: Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In: ECCV (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 173, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 173, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 138, + 173, + 480, + 205 + ], + "type": "text", + "content": "2. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. In: NeurIPS (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 205, + 480, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 205, + 480, + 227 + ], + "spans": [ + { + "bbox": [ + 138, + 205, + 480, + 227 + ], + "type": "text", + "content": "3. Armeni, I., Sener, O., Zamir, A.R., Jiang, H., Brilakis, I., Fischer, M., Savarese, S.: 3d semantic parsing of large-scale indoor spaces. In: CVPR (2016)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 228, + 480, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 480, + 249 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 480, + 249 + ], + "type": "text", + "content": "4. Chen, D.Z., Chang, A.X., Nießner, M.: Scanrefer: 3d object localization in rgb-d scans using natural language. In: ECCV (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 250, + 480, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 250, + 480, + 271 + ], + "spans": [ + { + "bbox": [ + 138, + 250, + 480, + 271 + ], + "type": "text", + "content": "5. Chen, Z., Duan, Y., Wang, W., He, J., Lu, T., Dai, J., Qiao, Y.: Vision transformer adapter for dense predictions. In: ICLR (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 271, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 271, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 138, + 271, + 480, + 293 + ], + "type": "text", + "content": "6. Choy, C., Gwak, J., Savarese, S.: 4d spatio-temporal convnets: Minkowski convolutional neural networks. In: CVPR (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 293, + 480, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 293, + 480, + 315 + ], + "spans": [ + { + "bbox": [ + 138, + 293, + 480, + 315 + ], + "type": "text", + "content": "7. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: CVPR (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 316, + 480, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 316, + 480, + 336 + ], + "spans": [ + { + "bbox": [ + 138, + 316, + 480, + 336 + ], + "type": "text", + "content": "8. Ding, H., Liu, C., He, S., Jiang, X., Loy, C.C.: MeViS: A large-scale benchmark for video segmentation with motion expressions. In: ICCV (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 337, + 480, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 337, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 138, + 337, + 480, + 358 + ], + "type": "text", + "content": "9. Ding, H., Liu, C., He, S., Jiang, X., Torr, P.H., Bai, S.: MOSE: A new dataset for video object segmentation in complex scenes. In: ICCV (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 359, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 359, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 138, + 359, + 480, + 380 + ], + "type": "text", + "content": "10. Ding, H., Liu, C., Wang, S., Jiang, X.: Vision-language transformer and query generation for referring segmentation. In: ICCV (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 381, + 480, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 381, + 480, + 403 + ], + "spans": [ + { + "bbox": [ + 138, + 381, + 480, + 403 + ], + "type": "text", + "content": "1. Ding, H., Liu, C., Wang, S., Jiang, X.: VLT: Vision-language transformer and query generation for referring segmentation. IEEE TPAMI (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 403, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 403, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 138, + 403, + 480, + 424 + ], + "type": "text", + "content": "2. Ding, R., Yang, J., Xue, C., Zhang, W., Bai, S., Qi, X.: Pla: Language-driven open-vocabulary 3d scene understanding. In: CVPR (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 425, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 425, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 138, + 425, + 480, + 445 + ], + "type": "text", + "content": "3. Ding, Z., Wang, J., Tu, Z.: Open-vocabulary panoptic segmentation with maskclip. In: ICLR (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 447, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 447, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 447, + 480, + 468 + ], + "type": "text", + "content": "4. Girdhar, R., El-Nouby, A., Liu, Z., Singh, M., Alwala, K.V., Joulin, A., Misra, I.: Imagebind: One embedding space to bind them all. In: CVPR (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 468, + 480, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 468, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 138, + 468, + 480, + 512 + ], + "type": "text", + "content": "5. Guo, Z., Zhang, R., Zhu, X., Tang, Y., Ma, X., Han, J., Chen, K., Gao, P., Li, X., Li, H., et al.: Point-bind & point-llm: Aligning point cloud with multi-modality for 3d understanding, generation, and instruction following. arXiv preprint arXiv:2309.00615 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 512, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 512, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 512, + 480, + 533 + ], + "type": "text", + "content": "6. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 534, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 534, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 534, + 480, + 555 + ], + "type": "text", + "content": "7. He, S., Ding, H.: Decoupling static and hierarchical motion perception for referring video segmentation. In: CVPR (2024)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 556, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 556, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 556, + 480, + 578 + ], + "type": "text", + "content": "8. He, S., Ding, H.: RefMask3D: Language-guided transformer for 3d referring segmentation. In: ACM MM (2024)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 578, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 480, + 599 + ], + "type": "text", + "content": "9. He, S., Jiang, X., Jiang, W., Ding, H.: Prototype adaption and projection for few-and zero-shot 3d point cloud semantic segmentation. IEEE TIP (2023)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 138, + 600, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 600, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 138, + 600, + 480, + 621 + ], + "type": "text", + "content": "20. Hong, Y., Zhen, H., Chen, P., Zheng, S., Du, Y., Chen, Z., Gan, C.: 3d-llm: Injecting the 3d world into large language models. In: NeurIPS (2023)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "type": "text", + "content": "21. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. In: ICLR (2022)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "type": "text", + "content": "22. Huang, P.H., Lee, H.H., Chen, H.T., Liu, T.L.: Text-guided graph neural networks for referring 3d instance segmentation. In: AAAI (2021)" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 410, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 410, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 410, + 91, + 447, + 102 + ], + "type": "text", + "content": "SegPoint" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 148 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 148 + ], + "type": "text", + "content": "23. Jain, A., Gkanatsios, N., Mediratta, I., Fragkiadaki, K.: Bottom up top down detection transformers for language grounding in images and point clouds. In: ECCV (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 149, + 480, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 480, + 181 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 480, + 181 + ], + "type": "text", + "content": "24. Jia, B., Chen, Y., Yu, H., Wang, Y., Niu, X., Liu, T., Li, Q., Huang, S.: Sceneverse: Scaling 3d vision-language learning for grounded scene understanding. In: ECCV (2024)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 480, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 480, + 203 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 480, + 203 + ], + "type": "text", + "content": "25. Jiang, L., Zhao, H., Shi, S., Liu, S., Fu, C.W., Jia, J.: Pointgroup: Dual-set point grouping for 3d instance segmentation. In: CVPR (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 204, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 204, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 204, + 480, + 224 + ], + "type": "text", + "content": "26. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: ICCV (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 225, + 480, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 225, + 480, + 246 + ], + "spans": [ + { + "bbox": [ + 132, + 225, + 480, + 246 + ], + "type": "text", + "content": "27. Kolodiazhnyi, M., Vorontsova, A., Konushin, A., Rukhovich, D.: Oneformer3d: One transformer for unified point cloud segmentation. In: CVPR (2024)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 247, + 480, + 268 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 247, + 480, + 268 + ], + "spans": [ + { + "bbox": [ + 132, + 247, + 480, + 268 + ], + "type": "text", + "content": "28. Lai, X., Tian, Z., Chen, Y., Li, Y., Yuan, Y., Liu, S., Jia, J.: Lisa: Reasoning segmentation via large language model. In: CVPR (2024)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 269, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 269, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 269, + 480, + 289 + ], + "type": "text", + "content": "29. Lai, X., Yuan, Y., Chu, R., Chen, Y., Hu, H., Jia, J.: Mask-attention-free transformer for 3d instance segmentation. In: ICCV (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 290, + 480, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 290, + 480, + 311 + ], + "spans": [ + { + "bbox": [ + 132, + 290, + 480, + 311 + ], + "type": "text", + "content": "30. Li, B., Zhang, Y., Chen, L., Wang, J., Yang, J., Liu, Z.: Otter: A multi-modal model with in-context instruction tuning. arXiv:2305.03726 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 312, + 480, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 312, + 480, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 312, + 480, + 332 + ], + "type": "text", + "content": "31. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. In: NeurIPS (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 333, + 480, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 333, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 333, + 480, + 365 + ], + "type": "text", + "content": "32. Lin, Z., Liu, C., Zhang, R., Gao, P., Qiu, L., Xiao, H., Qiu, H., Lin, C., Shao, W., Chen, K., et al.: Sphinx: The joint mixing of weights, tasks, and visual embeddings for multi-modal large language models. arXiv preprint arXiv:2311.07575 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 365, + 480, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 480, + 386 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 480, + 386 + ], + "type": "text", + "content": "33. Liu, C., Ding, H., Jiang, X.: GRES: Generalized referring expression segmentation. In: CVPR (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 387, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 387, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 387, + 480, + 407 + ], + "type": "text", + "content": "34. Liu, C., Ding, H., Zhang, Y., Jiang, X.: Multi-modal mutual attention and iterative interaction for referring image segmentation. IEEE TIP (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 408, + 480, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 408, + 480, + 429 + ], + "spans": [ + { + "bbox": [ + 132, + 408, + 480, + 429 + ], + "type": "text", + "content": "35. Liu, C., Jiang, X., Ding, H.: Instance-specific feature propagation for referring segmentation. IEEE TMM (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 430, + 480, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 480, + 450 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 480, + 450 + ], + "type": "text", + "content": "36. Liu, C., Jiang, X., Ding, H.: Primitivenet: decomposing the global constraints for referring segmentation. Visual Intelligence 2(1), 16 (2024)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 451, + 474, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 451, + 474, + 461 + ], + "spans": [ + { + "bbox": [ + 132, + 451, + 474, + 461 + ], + "type": "text", + "content": "37. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: NeurIPS (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 462, + 480, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 462, + 480, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 462, + 480, + 493 + ], + "type": "text", + "content": "38. Liu, Y., Kong, L., Cen, J., Chen, R., Zhang, W., Pan, L., Chen, K., Liu, Z.: Segment any point cloud sequences by distilling vision foundation models. In: NeurIPS (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 494, + 480, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 494, + 480, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 494, + 480, + 504 + ], + "type": "text", + "content": "39. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: ICLR (2019)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 505, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 505, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 505, + 480, + 536 + ], + "type": "text", + "content": "40. Nguyen, P.D., Ngo, T.D., Gan, C., Kalogerakis, E., Tran, A., Pham, C., Nguyen, K.: Open3dis: Open-vocabulary 3d instance segmentation with 2d mask guidance. In: CVPR (2024)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 536, + 480, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 480, + 557 + ], + "type": "text", + "content": "41. Park, N., Kim, S.: How do vision transformers work? arXiv preprint arXiv:2202.06709 (2022)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 558, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 558, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 558, + 480, + 578 + ], + "type": "text", + "content": "42. Peng, S., Genova, K., Jiang, C., Tagliasacchi, A., Pollefeys, M., Funkhouser, T., et al.: Openscene: 3d scene understanding with open vocabularies. In: CVPR (2023)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 579, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 579, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 579, + 480, + 601 + ], + "type": "text", + "content": "43. Peng, Z., Wang, W., Dong, L., Hao, Y., Huang, S., Ma, S., Wei, F.: Kosmos-2: Grounding multimodal large language models to the world. In: ICLR (2024)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 601, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 480, + 621 + ], + "type": "text", + "content": "44. Pi, R., Gao, J., Diao, S., Pan, R., Dong, H., Zhang, J., Yao, L., Han, J., Xu, H., Zhang, L.K.T.: Detgpt: Detect what you need via reasoning. In: EMNLP (2023)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "text", + "content": "45. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: CVPR (2017)" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "type": "text", + "content": "46. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In: NeurIPS (2017)" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "S. He et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "47. Qi, Z., Fang, Y., Sun, Z., Wu, X., Wu, T., Wang, J., Lin, D., Zhao, H.: Gpt4point: A unified framework for point-language understanding and generation. In: CVPR (2024)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "text", + "content": "48. Qian, G., Li, Y., Peng, H., Mai, J., Hammoud, H., Elhoseiny, M., Ghanem, B.: Pointnext: Revisiting pointnet++ with improved training and scaling strategies. In: NeurIPS (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 482, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 482, + 205 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 482, + 205 + ], + "type": "text", + "content": "49. Qian, Z., Ma, Y., Ji, J., Sun, X.: X-refseg3d: Enhancing referring 3d instance segmentation via structured cross-modal graph neural networks. In: AAAI (2024)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 205, + 482, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 482, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 482, + 237 + ], + "type": "text", + "content": "50. Rasheed, H., Maaz, M., Shaji, S., Shaker, A., Khan, S., Cholakkal, H., Anwer, R.M., Xing, E., Yang, M.H., Khan, F.S.: Glamm: Pixel grounding large multimodal model. In: CVPR (2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 237, + 482, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 482, + 270 + ], + "type": "text", + "content": "51. Rasley, J., Rajbhandari, S., Ruwase, O., He, Y.: Deepspeed: System optimizations enable training deep learning models with over 100 billion parameters. In: KDD (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 270, + 482, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 293 + ], + "type": "text", + "content": "52. Ren, Z., Huang, Z., Wei, Y., Zhao, Y., Fu, D., Feng, J., Jin, X.: Pixelm: Pixel reasoning with large multimodal model. In: CVPR (2024)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 293, + 482, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 314 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 314 + ], + "type": "text", + "content": "53. Rozenberszki, D., Litany, O., Dai, A.: Language-grounded indoor 3d semantic segmentation in the wild. In: ECCV (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 315, + 482, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 315, + 482, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 482, + 335 + ], + "type": "text", + "content": "54. Schult, J., Engelmann, F., Hermans, A., Litany, O., Tang, S., Leibe, B.: Mask3d: Mask transformer for 3d semantic instance segmentation. In: ICRA (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 336, + 482, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 336, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 336, + 482, + 369 + ], + "type": "text", + "content": "55. Shuai, X., Ding, H., Ma, X., Tu, R., Jiang, Y.G., Tao, D.: A survey of multimodal-guided image editing with text-to-image diffusion models. arXiv preprint arXiv:2406.14555 (2024)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 369, + 482, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 482, + 391 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 482, + 391 + ], + "type": "text", + "content": "56. Sun, J., Qing, C., Tan, J., Xu, X.: Superpoint transformer for 3d scene instance segmentation. In: AAAI (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 392, + 482, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 482, + 413 + ], + "type": "text", + "content": "57. Takmaz, A., Fedele, E., Sumner, R.W., Pollefeys, M., Tombari, F., Engelmann, F.: Openmask3d: Open-vocabulary 3d instance segmentation. In: NeurIPS (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 413, + 482, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 482, + 435 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 482, + 435 + ], + "type": "text", + "content": "58. Thomas, H., Qi, C.R., Deschaud, J.E., Marcotegui, B., Goulette, F., Guibas, L.J.: Kpconv: Flexible and deformable convolution for point clouds. In: ICCV (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 435, + 482, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 435, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 435, + 482, + 468 + ], + "type": "text", + "content": "59. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 468, + 482, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 482, + 490 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 482, + 490 + ], + "type": "text", + "content": "60. Wang, P.S.: Octformer: Octree-based transformers for 3d point clouds. In: SIGGRAPH (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 490, + 482, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 490, + 482, + 522 + ], + "spans": [ + { + "bbox": [ + 130, + 490, + 482, + 522 + ], + "type": "text", + "content": "61. Wang, W., Chen, Z., Chen, X., Wu, J., Zhu, X., Zeng, G., Luo, P., Lu, T., Zhou, J., Qiao, Y., et al.: Visionllm: Large language model is also an open-ended decoder for vision-centric tasks. In: NeurIPS (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 522, + 482, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 482, + 555 + ], + "type": "text", + "content": "62. Wang, W., Xie, E., Li, X., Fan, D.P., Song, K., Liang, D., Lu, T., Luo, P., Shao, L.: Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 555, + 482, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 555, + 482, + 588 + ], + "spans": [ + { + "bbox": [ + 130, + 555, + 482, + 588 + ], + "type": "text", + "content": "63. Wu, C., Ma, Y., Chen, Q., Wang, H., Luo, G., Ji, J., Sun, X.: 3d-stmn: Dependency-driven superpoint-text matching network for end-to-end 3d referring expression segmentation. In: AAAI (2024)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 588, + 482, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 588, + 482, + 610 + ], + "spans": [ + { + "bbox": [ + 130, + 588, + 482, + 610 + ], + "type": "text", + "content": "64. Wu, H., Xiao, B., Codella, N., Liu, M., Dai, X., Yuan, L., Zhang, L.: Cvt: Introducing convolutions to vision transformers. In: ICCV (2021)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 610, + 482, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 610, + 482, + 643 + ], + "spans": [ + { + "bbox": [ + 130, + 610, + 482, + 643 + ], + "type": "text", + "content": "65. Wu, J., Li, X., Xu, S., Yuan, H., Ding, H., Yang, Y., Li, X., Zhang, J., Tong, Y., Jiang, X., Ghanem, B., Tao, D.: Towards open vocabulary learning: A survey. IEEE TPAMI (2024)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 643, + 482, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 643, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 643, + 482, + 666 + ], + "type": "text", + "content": "66. Wu, X., Lao, Y., Jiang, L., Liu, X., Zhao, H.: Point transformer v2: Grouped vector attention and partition-based pooling. In: NeurIPS (2022)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 410, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 410, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 410, + 91, + 447, + 101 + ], + "type": "text", + "content": "SegPoint" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 566 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "text", + "content": "67. Wu, Y., Cheng, X., Zhang, R., Cheng, Z., Zhang, J.: Eda: Explicit text-decoupling and dense alignment for 3d visual grounding. In: CVPR (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "type": "text", + "content": "68. Xiao, Z., Zhang, W., Wang, T., Loy, C.C., Lin, D., Pang, J.: Position-guided point cloud panoptic segmentation transformer. arXiv preprint arXiv:2303.13509 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "type": "text", + "content": "69. Xu, J., Liu, S., Vahdat, A., Byeon, W., Wang, X., De Mello, S.: Open-vocabulary panoptic segmentation with text-to-image diffusion models. In: CVPR (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 183, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 183, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 183, + 480, + 205 + ], + "type": "text", + "content": "70. Xu, R., Wang, X., Wang, T., Chen, Y., Pang, J., Lin, D.: Pointllm: Empowering large language models to understand point clouds. In: ECCV (2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 205, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 205, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 133, + 205, + 480, + 226 + ], + "type": "text", + "content": "71. Yang, J., Ding, R., Deng, W., Wang, Z., Qi, X.: Regionplc: Regional point-language contrastive learning for open-world 3d scene understanding. In: CVPR (2024)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 227, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 227, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 133, + 227, + 480, + 258 + ], + "type": "text", + "content": "72. Yang, Y.Q., Guo, Y.X., Xiong, J.Y., Liu, Y., Pan, H., Wang, P.S., Tong, X., Guo, B.: Swin3d: A pretrained transformer backbone for 3d indoor scene understanding. arXiv preprint arXiv:2304.06906 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 259, + 480, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 259, + 480, + 292 + ], + "spans": [ + { + "bbox": [ + 133, + 259, + 480, + 292 + ], + "type": "text", + "content": "73. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: MPLug-owl: Modularization empowers large language models with multimodality. arXiv:2304.14178 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 293, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 293, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 133, + 293, + 480, + 314 + ], + "type": "text", + "content": "74. Yeshwanth, C., Liu, Y.C., Nießner, M., Dai, A.: Scannet++: A high-fidelity dataset of 3d indoor scenes. In: ICCV (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 315, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 315, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 133, + 315, + 480, + 346 + ], + "type": "text", + "content": "75. You, H., Zhang, H., Gan, Z., Du, X., Zhang, B., Wang, Z., Cao, L., Chang, S.F., Yang, Y.: Ferret: Refer and ground anything anywhere at any granularity. In: ICLR (2024)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 347, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 347, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 133, + 347, + 480, + 380 + ], + "type": "text", + "content": "76. Zhang, H., Li, H., Li, F., Ren, T., Zou, X., Liu, S., Huang, S., Gao, J., Zhang, L., Li, C., et al.: Llava-grounding: Grounded visual chat with large multimodal models. arXiv preprint arXiv:2312.02949 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 380, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 380, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 133, + 380, + 480, + 402 + ], + "type": "text", + "content": "77. Zhang, H., Ding, H.: Prototypical matching and open set rejection for zero-shot semantic segmentation. In: ICCV (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 403, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 403, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 133, + 403, + 480, + 435 + ], + "type": "text", + "content": "78. Zhang, S., Sun, P., Chen, S., Xiao, M., Shao, W., Zhang, W., Chen, K., Luo, P.: Gpt4roi: Instruction tuning large language model on region-of-interest. arXiv:2307.03601 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 435, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 435, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 133, + 435, + 480, + 456 + ], + "type": "text", + "content": "79. Zhang, Y., Gong, Z., Chang, A.X.: Multi3defer: Grounding text description to multiple 3d objects. In: ICCV (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 457, + 480, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 457, + 480, + 478 + ], + "spans": [ + { + "bbox": [ + 133, + 457, + 480, + 478 + ], + "type": "text", + "content": "80. Zhao, H., Jiang, L., Jia, J., Torr, P.H., Koltun, V.: Point transformer. In: ICCV (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 479, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 479, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 133, + 479, + 480, + 500 + ], + "type": "text", + "content": "81. Zhou, J., Wang, J., Ma, B., Liu, Y.S., Huang, T., Wang, X.: Uni3d: Exploring unified 3d representation at scale. In: ICLR (2024)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 501, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 501, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 133, + 501, + 480, + 522 + ], + "type": "text", + "content": "82. Zhou, Z., Zhang, Y., Foroosh, H.: Panoptic-polarnet: Proposal-free lidar point cloud panoptic segmentation. In: CVPR (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 523, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 523, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 133, + 523, + 480, + 544 + ], + "type": "text", + "content": "83. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. In: ICLR (2024)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 133, + 545, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 545, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 133, + 545, + 480, + 566 + ], + "type": "text", + "content": "84. Zhu, Z., Ma, X., Chen, Y., Deng, Z., Huang, S., Li, Q.: 3d-vista: Pre-trained transformer for 3d vision and text alignment. In: ICCV (2023)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "S. He et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/SegVG_ Transferring Object Bounding Box to Segmentation for Visual Grounding/4d7de385-7f41-401e-b1cf-a9c2348c4d1c_content_list.json b/2024/SegVG_ Transferring Object Bounding Box to Segmentation for Visual Grounding/4d7de385-7f41-401e-b1cf-a9c2348c4d1c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..fedda3ecb2100b475ad78268c5d53fc0d29d6d4b --- /dev/null +++ b/2024/SegVG_ Transferring Object Bounding Box to Segmentation for Visual Grounding/4d7de385-7f41-401e-b1cf-a9c2348c4d1c_content_list.json @@ -0,0 +1,1658 @@ +[ + { + "type": "text", + "text": "SegVG: Transferring Object Bounding Box to Segmentation for Visual Grounding", + "text_level": 1, + "bbox": [ + 240, + 140, + 763, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Weitai Kang $^{1}$ , Gaowen Liu $^{2}$ , Mubarak Shah $^{3}$ , and Yan Yan $^{1}$", + "bbox": [ + 276, + 210, + 725, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Illinois Institute of Technology", + "bbox": [ + 390, + 238, + 609, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Cisco Research", + "bbox": [ + 442, + 252, + 558, + 265 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 University of Central Florida", + "bbox": [ + 395, + 267, + 604, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Different from Object Detection, Visual Grounding deals with detecting a bounding box for each text-image pair. This one box for each text-image data provides sparse supervision signals. Although previous works achieve impressive results, their passive utilization of annotation, i.e. the sole use of the box annotation as regression ground truth, results in a suboptimal performance. In this paper, we present SegVG, a novel method transfers the box-level annotation as Segmentation signals to provide an additional pixel-level supervision for Visual Grounding. Specifically, we propose the Multi-layer Multi-task Encoder-Decoder as the target grounding stage, where we learn a regression query and multiple segmentation queries to ground the target by regression and segmentation of the box in each decoding layer, respectively. This approach allows us to iteratively exploit the annotation as signals for both box-level regression and pixel-level segmentation. Moreover, as the backbones are typically initialized by pretrained parameters learned from unimodal tasks and the queries for both regression and segmentation are static learnable embeddings, a domain discrepancy remains among these three types of features, which impairs subsequent target grounding. To mitigate this discrepancy, we introduce the Triple Alignment module, where the query, text, and vision tokens are triangurally updated to share the same space by triple attention mechanism. Extensive experiments on five widely used datasets validate our state-of-the-art (SOTA) performance. Code is available at https://github.com/WeitaiKang/SegVG.", + "bbox": [ + 261, + 319, + 743, + 638 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Visual Grounding $\\cdot$ Detection $\\cdot$ Transformer", + "bbox": [ + 261, + 652, + 645, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 715, + 375, + 732 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual grounding [14,17,26,30,52] aims to localize a target object within an image based on a free-form natural language text expression. It is particularly important for numerous downstream multimodal reasoning systems, such as visual question answering [10,35,41] and image captioning [1,4,50]. Previous works can be broadly categorized into three distinct groups: two-stage methods [3,44,45,51], one-stage methods [47,48], and transformer-based ones [6,15,16,31,43,49]. Both", + "bbox": [ + 212, + 748, + 787, + 843 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/411ba3ae96ffde238132b8d517bf288127d3eb0f330fc4fe4597e893c10d5b6d.jpg", + "image_caption": [ + "Fig. 1: The comparison of visual grounding frameworks. The block with a dashed border indicates that the module may not necessarily exist. (a) Previous baseline method consists of two backbones and additional transformer layers for target grounding, where a regression query is supervised to regress the box. Current SOTA methods further employ a text-to-visual module to align the visual features with text features. (b) Our method incorporates segmentation queries, which utilizes the box annotation at the pixel-level to segment the target. Additionally, we propose the Triple Alignment module to eliminate the domain discrepancy of the query, text, and vision features." + ], + "image_footnote": [], + "bbox": [ + 292, + 143, + 715, + 338 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "two-stage and one-stage approaches use convolutional neural networks for candidate proposals and the selection of the best-matching candidate. Nonetheless, these approaches rely on intricate modules that employ manually-crafted techniques for performing language inference and multi-modal integration.", + "bbox": [ + 212, + 491, + 784, + 551 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inspired by the success of the transformer [7, 8], TransVG [6] proposes a transformer-based pipeline. As shown in Fig. 1.(a), this pipeline extracts vision and text features via DETR [2] and BERT [7], respectively. To ground the target, they use the transformer encoder to fuse multimodal features along with a learnable regression query and decode the query through an MLP. To enhance the final target grounding stage, subsequent studies continue with some text-to-visual modules in the early stage to modulate the vision features to align with the text features. For example, QRNet [49] proposes a query-modulated method for extracting language-aware vision features within the vision backbone. VLTVG [43] introduces a verification map to activate the vision features to align with the text features before multimodal fusion.", + "bbox": [ + 212, + 551, + 787, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite their advancements, the suboptimal annotation utilization, i.e., only using the box annotation as a regression annotation, limits their performance. As discussed in [37], Visual Grounding presents unique challenges compared to Object Detection due to its sparse supervision signals. Specifically, it provides only one box label for each text-image pair, while necessitating detection within a multimodal setting. Therefore, it is essential to fully exploit the box annotation, by treating it as a segmentation mask (pixels within the bounding box are assigned a value of 1, while pixels outside the bounding box are assigned as 0).", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "W. Kang et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we introduce SegVG (see Fig. 1.(b)), a novel method that leverages the pixel-level details within the box annotation as segmentation signals to offer additional fine-grained supervision for Visual Grounding. Specifically, we propose the Multi-layer Multi-task Encoder-Decoder as the target grounding stage, where we learn a regression query and multiple segmentation queries to ground the target by regression and segmentation of the box in each decoding layer, respectively. The confidence score derived from the segmentation can further serve as a Focal Loss [21] scaling factor to adaptively emphasize the other losses of challenging training samples. This approach allows us to iteratively exploit the annotation as signals for both box-level regression and pixel-level segmentation. Furthermore, the initial parameters for model backbones, typically derived from pretrained unimodal tasks, along with data-agnostic static embeddings used as queries for decoding, result in a domain discrepancy among different sources of feature, affecting the effectiveness of target grounding. To tackle this problem, we present the Triple Alignment module, where we harmonize the domain of query, text, and vision features by implementing a triangular update process through a triple attention mechanism. As a result, we ensure that all features adapt and integrate within the same multimodal space, thereby enhancing subsequent target grounding. Our contributions are as follows:", + "bbox": [ + 212, + 146, + 787, + 434 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose the Multi-layer Multi-task Encoder-Decoder to maximize the utilization of the box annotation, which introduces an additional segmentation format for pixel-level supervision in Visual Grounding.", + "- To eliminate the domain discrepancy among the query, text, and vision, we introduce the Triple Alignment to update these three types of features into a sharing domain, which facilitates the subsequent target grounding.", + "- We conduct extensive experiments on five widely used datasets to show the performance superiority of our proposed methods compared with previous state-of-the-art methods and further investigate the reliability benefits derived from the segmentation output in real applications.", + "- We will release code and checkpoints for future research development." + ], + "bbox": [ + 225, + 446, + 784, + 614 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 652, + 387, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Visual grounding methods can be roughly classified into three pipelines: two-stage methods, one-stage methods, and transformer-based methods.", + "bbox": [ + 212, + 688, + 782, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Two-stage methods Two-stage approaches [3,51] treat visual grounding as first generating candidate object proposals and then finding the best match to the text. In the first stage, an off-the-shelf detector processes the image and proposes regions that may contain the target. In the second stage, a ranking network calculates the similarity between candidate regions and processed text features, selecting the region with the highest similarity score as the final result. Training losses include binary classification loss [29] or maximum-margin ranking loss [51]. To better understand the text and cross-modality matching, MattNet [51] focuses", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Transferring Object Bounding Box to Segmentation for Visual Grounding", + "bbox": [ + 240, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/73df9bb07d93244cf5086901d4f2c7a651af179c31859a2f5a5ebe3fb84b9f16.jpg", + "image_caption": [ + "Fig. 2: SegVG: The upper figure includes the vision and text backbone. Our proposed Triple Alignment module is iteratively inserted into intermediate layers to eliminate domain discrepancy. The lower figure shows our Multi-layer Multi-task Encoder-Decoder, which adopts a transformer encoder-decoder to update multimodal features and ground the target. In this architecture, we make the best of the box annotation as a segmentation ground truth and integrate an additional segmentation task into Visual Grounding. Additionally, the segmentation output serves as a Focal Loss factor, allowing adaptive emphasis on challenging cases for the regression loss. $\\mathrm{M} = 6$ , $\\mathrm{R} = 6$ ." + ], + "image_footnote": [], + "bbox": [ + 241, + 145, + 763, + 381 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "on decomposing the text into subject, location, and relationship components. [3] introduces an expression-aware score for improved candidate region ranking.", + "bbox": [ + 212, + 535, + 782, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "One-stage methods One-stage approaches [47, 48] directly concatenate vision and text features in the channel dimension and rank confidence values for candidate regions proposed based on the concatenated multimodal features. For example, FAOA [48] predicts bounding boxes using a YOLOv3 detector [32] on the concatenated features. ReSC [47] further improves the ability to ground complex queries by introducing a recursive sub-query construction module.", + "bbox": [ + 212, + 566, + 784, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Transformer-based methods Transformer-based approach is first introduced by TransVG [6]. Unlike previous methods, TransVG concatenates the regression query (a learnable embedding), vision tokens, and text tokens and uses transformer encoders [38] to perform cross-modal fusion and target grounding. The query is then processed through an MLP to decode the box. Benefiting from the flexible structure of transformer modules in processing multimodal features, recent works continue adopting this pipeline and propose novelties regarding feature extraction. VLTVG [43] develops a visual-linguistic verification module before the target grounding stage to modulate the vision features with the relationship between vision and text features. QRNet [49] proposes a Query-modulated Refinement Network to mitigate the gap between features from the unimodal vision backbone and those needed for multi-modal reasoning.", + "bbox": [ + 212, + 657, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "W. Kang et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multi-task Visual Grounding Multi-task learning is extensively utilized in object detection and segmentation [2,11], often capitalizing on a shared backbone and task-specific heads. Expanding upon this idea, several studies [19, 25, 36] have proposed solutions to the Multi-task Visual Grounding problem. In this problem, they jointly tackle Referring Expression Comprehension (REC, also known as Visual Grounding) and Referring Expression Segmentation (RES), requiring both box annotations and segmentation annotations. It is important to note that, unlike those approaches, even though we incorporate segmentation losses in our method, we do not need segmentation annotations but only box annotations, focusing specifically on the Visual Grounding task.", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 215, + 320, + 380, + 339 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we present the components of our SegVG in the order of the data flow: starting with the backbones, followed by our proposed Triple Alignment, and finally our Multi-layer Multi-task Encoder-Decoder.", + "bbox": [ + 212, + 352, + 785, + 398 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Backbones", + "text_level": 1, + "bbox": [ + 215, + 420, + 349, + 434 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 2 (upper), our vision backbone consists of ResNet and transformer encoder from DETR [2], with parameters pretrained on Object Detection task using the MSCOCO dataset [22], excluding the validation and test sets of Visual Grounding dataset. The text backbone is the base model of BERT [7].", + "bbox": [ + 212, + 445, + 787, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Vision Backbone Given an input image $\\mathbf{I_0}$ $(\\mathbb{R}^{3\\times H_0\\times W_0})$ , we employ ResNet to generate a 2D feature map $\\mathbf{I}\\in \\mathbb{R}^{C\\times H\\times W}$ ( $C = 2048$ , $H = \\frac{H_0}{32}$ , $W = \\frac{W_0}{32}$ ). A 1x1 convolutional layer is then used to reduce the channel dimension of $\\mathbf{I}$ to $C_v = 256$ , resulting in $\\mathbf{I}'$ . We further flatten $\\mathbf{I}'$ into $\\mathbf{Z}_v\\in \\mathbb{R}^{C_v\\times N_v}$ ( $N_v = H\\times W$ ). Position embedding is then added to $\\mathbf{Z}_v$ to preserve sensitivity to the original 2D spatial locations. $\\mathbf{Z}_v$ is then iteratively processed through DETR's encoder layer (total 6 transformer layers) and the Triple Alignment to obtain the output $\\mathbf{Z}_v$ .", + "bbox": [ + 212, + 520, + 787, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Text Backbone Given a text, we initially utilize the BERT's embedding layer to convert it into $N_{t}$ language tokens with $C_t$ channel dimension. In alignment with [6], we preprocess a [CLS] token and append a [SEP] token to the beginning and end positions of the tokenized language, respectively. Following this, we iteratively input the language tokens into BERT's layers (total 12 transformer layers) and the Triple Alignment, generating language embedding $\\mathbf{Z}_t$ .", + "bbox": [ + 212, + 641, + 787, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Triple Alignment", + "text_level": 1, + "bbox": [ + 215, + 753, + 406, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given that the text and vision backbones are pretrained from unimodal tasks and the queries are data-agnostic, the subsequent target grounding stage faces the challenge of aligning these three types of features into the same space before performing multimodal fusion for target grounding. Additionally, considering", + "bbox": [ + 212, + 779, + 787, + 842 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Transferring Object Bounding Box to Segmentation for Visual Grounding", + "bbox": [ + 240, + 114, + 730, + 130 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "that the backbones usually contribute the majority of the overall parameters, solely using them for extracting unimodal features without incorporating multimodal alignment is sub-optimal. Therefore, an optimal solution is to address the domain discrepancy before moving on to the subsequent target grounding stage.", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Triple Alignment (Fig. 2 (upper)) utilizes an attention mechanism to perform triangular feature sampling, aiming to ensure domain consistency among the query, text, and vision features. The queries, $\\mathbf{Z}_o$ , are first initialized by N learnable embeddings, where one embedding is for the regression query and the rest of the embeddings are for multiple segmentation queries. The data flow is:", + "bbox": [ + 212, + 207, + 784, + 282 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Z} _ {v} ^ {i + 1} = \\operatorname {D E T R L a y e r} _ {i} \\left(\\mathbf {Z} _ {v} ^ {i}\\right), i \\in \\{0, 1,.., L - 1 \\} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 290, + 784, + 308 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Z} _ {t} ^ {i + 1} = \\operatorname {B E R T L a y e r} _ {\\mathcal {2 i + 1}} (\\operatorname {B E R T L a y e r} _ {\\mathcal {2 i}} (\\mathbf {Z} _ {t} ^ {i})), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 328, + 784, + 347 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ \\mathbf {Z} _ {o} ^ {\\prime}, \\mathbf {Z} _ {t} ^ {\\prime}, \\mathbf {Z} _ {v} ^ {\\prime} \\right] = \\operatorname {T r i - M H A} \\left(\\mathbf {Z} _ {o}, \\mathbf {Z} _ {t} ^ {i + 1}, \\mathbf {Z} _ {v} ^ {i + 1}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 359, + 363, + 784, + 382 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Z} _ {o} = \\mathbf {Z} _ {o} + \\mathbf {Z} _ {o} ^ {\\prime}, \\mathbf {Z} _ {t} ^ {i + 1} = \\mathbf {Z} _ {t} ^ {i + 1} + \\mathbf {Z} _ {t} ^ {\\prime}, \\mathbf {Z} _ {v} ^ {i + 1} = \\mathbf {Z} _ {v} ^ {i + 1} + \\mathbf {Z} _ {v} ^ {\\prime}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 316, + 398, + 784, + 417 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where L is the number of layers, BERTLayer is the layer of BERT and DETRLayer is the layer of DETR's encoder. The vision and text features are first encoded by Eq. 1 and Eq. 2. Subsequently, the three types of tokens (query, text, and vision) are updated by our Triple Multi-Head Attention Layer (Tri-MHA) using Eq. 3. The output tokens are merged back to their original branches respectively by Eq. 4. Within each head of the Triple Multi-Head Attention Layer (Tri-MHA), each type of the features simultaneously computes its updated representation by attending to both the others and itself:", + "bbox": [ + 212, + 420, + 782, + 541 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nS = \\left[ \\mathbf {Z} _ {o} W ^ {(o, S)}, \\mathbf {Z} _ {t} W ^ {(t, S)}, \\mathbf {Z} _ {v} W ^ {(v, S)} \\right], S \\in \\{Q, K, V \\}\n$$\n", + "text_format": "latex", + "bbox": [ + 321, + 549, + 681, + 566 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n[ \\mathbf {Z} _ {o}, \\mathbf {Z} _ {t}, \\mathbf {Z} _ {v} ] = \\operatorname {S o f t M a x} \\left(Q K ^ {T} / \\sqrt {d _ {k}}\\right) V, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 571, + 784, + 589 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Z} _ {e} ^ {\\prime} = \\mathbf {Z} _ {e} W ^ {e}, e \\in \\{o, t, v \\},\n$$\n", + "text_format": "latex", + "bbox": [ + 411, + 594, + 591, + 611 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\{W^{(e,S)}, W^e : e \\in \\{o, t, v\\}, S \\in \\{Q, K, V\\}\\}$ are trainable parameters. As a result, each of the output features is triangular sampling from all of the three types of features, which alleviates the domain discrepancy.", + "bbox": [ + 212, + 619, + 782, + 666 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Multi-layer Multi-task Encoder-Decoder", + "text_level": 1, + "bbox": [ + 214, + 685, + 594, + 699 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The Multi-layer Multi-task Encoder-Decoder serves as the target grounding stage, where we use a transformer encoder-decoder for cross-modal fusion and target grounding to perform a box regression task and a box segmentation task.", + "bbox": [ + 212, + 707, + 782, + 752 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Encoder As shown in Fig. 2 (lower left), given the aligned output text and vision features from the backbones, the encoder fuses the two modalities into the multimodal features by a stack of transformer layers. In each layer, the concatenated text and vision tokens go through the Multi-Head Self-Attention layer (MHSA) and the Feed Forward Network (FFN) with the residual connection.", + "bbox": [ + 212, + 763, + 782, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "W. Kang et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Decoder In each decoder layer, we aim to fully exploit the box annotation. We propose the bbox2seg paradigm to transform the box annotation into a segmentation mask, which classifies all pixels within the box as foreground (with a value of one) and those outside the box as background (with a value of zero). As shown in Fig. 2 (lower right), one regression query aims to regress the box, while the remaining segmentation queries aim to segment the box. Different segmentation queries are endowed with different learnable positional embeddings to enhance the robustness of each decoder layer, since the decoder layer, when confronted with various queries, is required to segment the same box. Following that, the queries pass through the Multi-Head Self-Attention layer to exchange information about the same target, prompting each other to better locate the target. Subsequently, the queries undergo the Multi-Head Cross-Attention layer and the Feed Forward Network, where multimodal features serve as the Key and Value to ground the target. Finally, a shared MLP across all decoder layers decodes the regression query into the box result, supervised by L1 loss and Giou loss [33]. Each segmentation query is repeated $N_{v}$ times and concatenated with visual tokens along the channel dimension. Another shared MLP decodes the concatenated feature into the segmentation mask, supervised by Focal loss [21] and Dice loss [27]. It is noteworthy that our segmentation paradigm shares the same semantic foundation as the regression paradigm, i.e., to distinguish bounding box, rather than instance segmentation. Therefore, incorporating non-object pixels in the segmented foreground does not introduce ambiguity to the model. We provide qualitative results 4.8 to demonstrate this feature. To alleviate multi-task optimization challenges, we freeze the backbones for the initial k epochs to stabilize the training process.", + "bbox": [ + 212, + 146, + 787, + 525 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Confidence score Since both the regression output and segmentation output share the same aim, we can additionally obtain the confidence score for the foreground by averaging values inside the ground truth box of the segmentation output to reflect the confidence of the regression output. In the training process, we can transform this confidence score as the Focal loss factor [21] to adaptively emphasize the other losses of challenging training samples. The final loss function of each decoder layer is formulated as follows:", + "bbox": [ + 212, + 535, + 787, + 641 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nL = \\lambda_ {1} c _ {f o c a l} L _ {1} + \\lambda_ {g i o u} c _ {f o c a l} L _ {g i o u} + \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 375, + 648, + 785, + 672 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\lambda_ {d i c e} c _ {f o c a l} L _ {d i c e} + \\lambda_ {f o c a l} L _ {f o c a l},\n$$\n", + "text_format": "latex", + "bbox": [ + 410, + 667, + 629, + 683 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\lambda_{1}$ , $\\lambda_{giou}$ , $\\lambda_{focal}$ and $\\lambda_{dice}$ are hyperparameters. $L_{1}$ is the L1 loss. $L_{giou}$ is the GIoU loss [33]. $L_{focal}$ is the Focal loss [21]. $L_{dice}$ is the Dice loss [27]. $c_{focal}$ is the above Focal loss factor averaged across all segmentation outputs.", + "bbox": [ + 212, + 689, + 787, + 734 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the real-world application perspective, the visual grounding task can be viewed as open-vocabulary object detection [54], where target objects lack predetermined categories. Therefore, previous transformer-based methods directly regress the box without confidence scores, since there is no candidate proposal or selection stage in transformer-based pipeline. However, confidence scores are valuable for enhancing the control or reliability of predictions by filtering out low-confidence predictions. This feature could benefit the future integration of visual", + "bbox": [ + 212, + 734, + 787, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Transferring Object Bounding Box to Segmentation for Visual Grounding", + "bbox": [ + 240, + 114, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/34fa8a3bb2ea6caafb8b3dee8b79473c88b80965f9984a34a2fb30ed605acac0.jpg", + "table_caption": [ + "Table 1: Comparisons with state-of-the-art methods on widely used datasets. We highlight the best and second best performance in red and blue, and bold our model." + ], + "table_footnote": [], + "table_body": "
ModelsBackboneRefCOCORefCOCO+RefCOCOgReferItGame test
valtestAtestBvaltestAtestBval-gval-utest-u
Two-stage:
CMN [12]VGG16-71.0365.77-54.3247.7657.47--28.33
VC [55]VGG16-73.3367.44-58.4053.1862.30--31.13
ParalAttn [57]VGG16-75.3165.52-61.3450.8658.03---
MAttNet [51]ResNet-10176.6581.1469.9965.3371.6256.02-66.5867.2729.04
Similarity Net [39]ResNet-101---------34.54
CITE [29]ResNet-101---------35.07
DDPN [53]ResNet-101---------63.00
LGRANs [40]VGG16-76.6066.40-64.0053.4061.78---
DGA [44]VGG16-78.4265.53-69.0751.99--63.28-
RvG-Tree [29]ResNet-10175.0678.6169.8563.5167.4556.66-66.9566.51-
NMTree [23]ResNet-10176.4181.2170.0966.4672.0257.5264.6265.8766.44-
Ref-NMS [3]ResNet-10180.7084.0076.0468.2573.6859.42-70.5570.62-
One-stage:
SSG [5]DarkNet-53-76.5167.50-62.1449.2747.4758.80-54.24
ZSGNet [34]ResNet-50---------58.63
FAOA [48]DarkNet-5372.5474.3568.5056.8160.2349.6056.1261.3360.3660.67
RCCF [20]DLA-34-81.0671.85-70.3556.32--65.7363.79
ReSC-Large [47]DarkNet-5377.6380.4572.3063.5968.3656.8163.1267.3067.2064.60
LBYL-Net [13]DarkNet-5379.6782.9174.1568.6473.3859.4962.70--67.47
Transformer-based:
TransVG [6]ResNet-10181.0282.7278.3564.8270.7056.9467.0268.6767.7370.73
QRNet [49]Swin-S84.0185.8582.3472.9476.1763.8171.8973.0372.5274.61
VLTVG [43]ResNet-10184.7787.2480.4974.1978.9365.1772.9876.0474.1871.98
SegVG (ours)ResNet-10186.8489.4683.0777.1882.6367.5976.0178.3577.4275.59
", + "bbox": [ + 251, + 184, + 754, + 459 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "grounding models into downstream multimodal reasoning systems or real-world applications. To meet the requirements of this feature, our approach incorporates a confidence score derived from the segmentation output during inference. Specifically, we calculate the model's confidence by averaging values greater than or equal to 0.35 (adopted from [42]) in the segmentation output of one segmentation query. Analyses 4.7 in the Experiment section demonstrate the faithfulness and benefits of incorporating this additional confidence score.", + "bbox": [ + 212, + 488, + 787, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 617, + 375, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Metric and Datasets", + "text_level": 1, + "bbox": [ + 215, + 648, + 431, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Metric A predicted bounding box is considered accurate if its Intersection over Union (IoU) with the ground-truth bounding box exceeds 0.5. In accordance with the established practices in preceding studies [6,43], we employ top-1 accuracy (measured in percentage) as the primary metric to assess our method.", + "bbox": [ + 215, + 672, + 785, + 736 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Datasets There are five standard benchmarks: RefCOCO [52], RefCOCO+ [52], RefCOCOg-g [26], RefCOCOg-umd [26], and ReferItGame [17]. Four of them (RefCOCO, RefCOCO+, and RefCOCOg-(g/umd)) are all derived from MSCOCO [22]. RefCOCO consists of 19,994 images and 142,210 referring texts, which is divided into four subsets: a training set with 120,624 texts, a validation set with 10,834 texts, and two test sets (testA and testB) containing 5,657", + "bbox": [ + 215, + 748, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "W. Kang et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and 5,095 texts, respectively. RefCOCO+ includes 19,992 images and 141,564 referring texts, which is partitioned into four subsets: a training set with 120,191 texts, a validation set with 10,758 texts, and two test sets (testA and testB) containing 5,726 and 4,889 texts, respectively. RefCOCOg contains 25,799 images and 95,010 longer texts. Two widely accepted splitting methods—RefCOCOg-g [26] and RefCOCOg-umd [28]—are employed for this dataset, and we perform experiments using both RefCOCOg-g (val-g) and RefCOCOg-umd (val-u and test-u) splitting conventions. The ReferItGame dataset, featuring 20,000 images from SAIAPR-12 [9], is divided into three segments: a training set with 54,127 texts, a validation set with 5,842 texts, and a testing set comprising 60,103 texts.", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Implementation", + "text_level": 1, + "bbox": [ + 215, + 321, + 395, + 338 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Our approach uses an input image size of $640 \\times 640$ and sets the maximum expression length at 40. When resizing images, we preserve the original aspect ratio. The longer edge is resized to 640, and the shorter edge is padded to 640 with the value of zero. Texts exceeding 38 tokens are truncated, reserving a start position and an end position of the characters for the [CLS] and [SEP] token, respectively. If the text is shorter, empty tokens are added after the [SEP] token to reach an input length of 40. Paddings for the input image are not tracked by masks, while empty tokens of text employ masks.", + "bbox": [ + 212, + 349, + 787, + 470 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We use the AdamW optimizer. The initial learning rate of 1e-5 is assigned to backbones, and 1e-4 to the rest parameters. Weight decay is set at 1e-4. The visual backbone is initialized with the DETR's backbone and encoder, while the language branch uses the basic BERT model. For the final results, our model is trained for 90 epochs, with the learning rate decreasing by a factor of 10 after 60 epochs. The k hyperparameter in Multi-layer Multi-task Encoder-Decoder is set to 10. We use a batch size of 64. For the ablation studies presented in Table 3, the models are trained for 60 epochs with k equal to 20, and the learning rate drops after 40 epochs. We set $\\lambda_{1} = 5$ , $\\lambda_{giou} = 2$ , $\\lambda_{focal} = 1$ and $\\lambda_{dice} = 1$ . We adhere to previous practices [6, 43] for data augmentation during training.", + "bbox": [ + 212, + 470, + 787, + 623 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3 Quantitative Results", + "text_level": 1, + "bbox": [ + 215, + 646, + 433, + 662 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We report the performance of our SegVG on all benchmark datasets. As presented in Table 1, our SegVG model demonstrates superiority across all of the datasets. This indicates the effectiveness and generalizability of our approach. It is worth noting that RefCOCO+ and RefCOCOg are relatively more challenging datasets, as RefCOCO+ does not include location terms in its language expressions, and RefCOCOg has longer language expressions compared to other datasets. Despite these challenges, our model exhibits significant improvements on these two difficult datasets. Specifically, on RefCOCO+, our model outperforms the previous SOTA models with $+2.99\\%$ , $+3.7\\%$ , and $+2.42\\%$ on the val, testA, and testB subsets, respectively. On RefCOCOg, our model also surpasses the previous SOTA models with $+3.03\\%$ , $+2.31\\%$ , and $+3.24\\%$ on the val-g,", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Transferring Object Bounding Box to Segmentation for Visual Grounding", + "bbox": [ + 240, + 114, + 730, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/2c7dac147682ba1d367a7ed158acdc81fe015f05c40e72a5a698159bd3d6b6f4.jpg", + "table_caption": [ + "Table 2: Compare transformer-based models on Parameter count and GFLOPS." + ], + "table_footnote": [], + "table_body": "
BackboneParameter count (M)GFLOPS (G)
TransVG [6]ResNet101141.5572.61
VLTVG [43]ResNet101141.6169.87
QRNet [49]Swin-S247.0680.12
SegVGResNet101155.2873.48
", + "bbox": [ + 279, + 170, + 723, + 234 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/790acc68e2eab7f52d69644d371b5adb0c6435d292cbe858d81764e0aecaa489.jpg", + "table_caption": [ + "Table 3: Ablation study on RefCOCOg test-u. Encoder and Decoder are the encoder and decoder of Multi-layer Multi-task Encoder-Decoder, respectively. MMDecoder represents Multi-layer and Multi-task supervision in the Decoder. N is the number of segmentation queries. Triple means Triple Alignment. Excluding Query in Triple (g) means using bidirectional alignment with concatenated text-vision tokens." + ], + "table_footnote": [], + "table_body": "
IdModelAcc(%)
(a)Backbones + Encoder + Decoder66.97
(b)Backbones + Triple + Encoder + Decoder75.75
(c)Backbones + Encoder + MMDecoder (N=1)72.61
(d)Backbones + Encoder + MMDecoder (N=5)76.21
(e)Backbones + Triple + Encoder + MMDecoder (N=5)77.29
(f)(e) w/o Encoder76.65(-0.64)
(g)(e) w/o Query in Triple76.37(-0.92)
(h)(e) w/o Focal loss76.83(-0.46)
", + "bbox": [ + 284, + 327, + 725, + 465 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "val-u, and test-u subsets, respectively. These results suggest that under the reinforcement of Triple Alignment and Multi-layer Multi-task Encoder-Decoder, the query, text, and vision tokens are triangularly updated to share the same space, and the model fully exploits the bounding box as fine-grained pixel-level supervision for comprehensive end-to-end learning.", + "bbox": [ + 212, + 492, + 787, + 566 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We also conduct a comparison of the number of parameters and GFLOPS across transformer-based models to evaluate computational costs. As depicted in Table 2, the computational cost of SegVG falls within a reasonable range.", + "bbox": [ + 212, + 566, + 789, + 613 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.4 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 633, + 390, + 648 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In this section, we aim to validate the efficacy of each proposed module. We conduct ablation studies on the RefCOCOg-umd test dataset. Specifically, we start by evaluating a basic structure, i.e., the backbones with encoder-decoder structure. After that, we systematically incorporate the Triple Alignment module into the backbones and introduce the Multi-layer Multi-task supervision into the decoder through the controlled variable approach. Meanwhile, we conduct additional ablation experiments on specific details, including assessing the efficacy of incorporating the encoder, introducing Query in the Triple Alignment, and introducing the Focal loss from the segmentation output to the other losses.", + "bbox": [ + 212, + 657, + 787, + 794 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "As shown in Table 3, we can draw the following conclusions when comparing the experimental results under controlled variables: 1) [(a) v.s. (b)]: Incorporating the Triple Alignment can effectively eliminate the domain discrepancy", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "W. Kang et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/dd8ace7895fd9ff62d8c47d167304e138e5dd13faafe5e76c634cf3079bbf960.jpg", + "table_caption": [ + "Table 4: Ablation study on RefCOCOg test-u regarding the number of segmentation queries. $\\Delta$ Hour and $\\Delta$ Acc are the additional time cost and accuracy improvement compared to ID(i) experiment, respectively." + ], + "table_footnote": [], + "table_body": "
IDnum_queryAccTime CostΔHour / ΔAcc (v.s. ID(i))
iN = 172.6116.32h-
iiN = 373.0218.97h6.46
iiiN = 576.2120.10h1.05
ivN = 774.3822.24h3.34
vN = 973.9124.07h5.96
", + "bbox": [ + 295, + 198, + 710, + 275 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "among the query, text, and vision features, thereby facilitating subsequent target grounding. 2) [(a) v.s. (c)]: Introducing Multi-layer Multi-task supervision can iteratively make the best of the annotations in the target grounding stage, thereby enhancing the learning of query representations. 3) [(c) v.s. (d)]: Increasing the number of segmentation queries can further improve the robustness of the decoder when provided with different queries and required to segment the same box. 4) [(a), (b), (d), and (e)]: Combining the Triple Alignment and Multi-layer Multi-task Encoder-Decoder can effectively enhance the overall performance, achieving the optimal result. 5) [(e) v.s. (f)]: Even if we include the Triple Alignment supporting multimodal communication, it remains necessary for the subsequent encoder to update the unimodal features generated by the backbones into multimodal features. 6) [(e) v.s. (g)]: It is necessary to involve queries in Triple Alignment to transfer the data-agnostic embedding into data-related queries. Otherwise, using only Bidirectional Alignment (BA) for text and vision tokens, similar to approaches like Deep Fusion in GLIP [18] and Grounding DINO [24], and WPA in CoupAlign [56], causes a noticeable decline $(-0.92\\%)$ . 7) [(d), (e) and (g)]: The slight gain of (g)(76.37%) over (d)(76.21%) stems from our MMDecoder's pixel-level signals, which already boosts BA in the Encoder. Thus, extra BA effort in the backbone is marginal, failing to solve the unaligned query issue. Instead, Tri-Align (e)(77.29%) can solve this issue, showing its novelty. Notably, (d)(76.21%), with a basic encoder-decoder, already achieves SOTA performance, emphasizing our bbox2seg paradigm's simplicity and effectiveness. 8) [(e) v.s. (h)]: The segmentation output can further derive the confidence score of the model's prediction, which is transformed into the Focal loss factor to adaptively scale the other losses to focus more attention on challenging cases.", + "bbox": [ + 212, + 304, + 787, + 681 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We further conduct a more detailed abaltion study about the improvement and corresponding cost of adding more segmentation queries of Tab.3(c)-(d). As shown in Table 4, the best performance is observed with five segmentation queries. Adding more than that increases the burden from pixel-level constraint without benefits, also increasing computational cost per accuracy improvement.", + "bbox": [ + 212, + 683, + 787, + 758 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.5 Triple Alignment Analysis", + "text_level": 1, + "bbox": [ + 215, + 782, + 478, + 797 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In addition to the improvement results in ablation study, we further deepen our understanding of the Triple Alignment by analyzing the attention behavior.", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Transferring Object Bounding Box to Segmentation for Visual Grounding", + "bbox": [ + 240, + 114, + 730, + 130 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/dd687706b187069271b3a5aeddfed66851fd7c9ab7745d2b81045bcab5d0871b.jpg", + "table_caption": [ + "Table 5: Attention values from query to text-referred visual region in Triple Alignment." + ], + "table_footnote": [], + "table_body": "
LayerRefCOCORefCOCO+RefCOCOgReferItGame test
valtestAtestBvaltestAtestBval-gval-utest-utest
layer217%18%16%19%21%18%20%18%17%25%
layer434%36%33%29%31%28%27%33%33%30%
layer661%63%60%55%58%50%55%59%59%38%
", + "bbox": [ + 279, + 170, + 717, + 231 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/1ebe678e72ac204077307d846409e5669211539792f79917bf61117aa2c64dbf.jpg", + "table_caption": [ + "Table 6: Comparison with alternative method using RIS to provide pseudo segmentation labels for supervision." + ], + "table_footnote": [], + "table_body": "
ReferItGameSegVG+RIS (LAVT)SegVG
val74.9176.85
test73.7675.06
", + "bbox": [ + 238, + 297, + 473, + 335 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/43c9c28742868b801d5c3dc12fcbb2b3f044ec8e40686f5ec7f2c5273e9dae18.jpg", + "table_caption": [ + "Table 7: AP50 using different segmentation queries for confidence score calculation on RefCOCOg test-u." + ], + "table_footnote": [], + "table_body": "
Seg Query1st2nd3rd4th5th
AP5084.6584.5784.7384.7884.63
", + "bbox": [ + 532, + 301, + 767, + 330 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Specifically, we calculate the sum of the attention values from the query to the text-referred visual region (target bbox) as a percentage of the total attention (which includes attention to the query, text, and visual tokens) to illustrate the extent of alignment across these three modalities in the second, fourth, and final layer of Triple Alignment. We average the percentage across all the attention heads and queries, and perform the analysis across all the datasets. As shown in Table 5, in all the datasets, the attention values increase as the layer progresses, indicating that Triple Alignment progressively aligns the query to comprehend the text and then focus on the referred visual region.", + "bbox": [ + 212, + 364, + 787, + 501 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.6 Comparison with alternative method", + "text_level": 1, + "bbox": [ + 215, + 523, + 566, + 539 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Given the development of Referring Expression Segmentation (RES), a natural alternative method would be using a RES method to generate pseudo segmentation labels to substitute our bbox2seg paradigm. Therefore, to mimic real-world scenarios, we use LAVT [46] trained on RefCOCO to obtain pseudo segmentation labels on ReferItGame. We follow the same training setting in ablation study to conduct the comparison on ReferItGame. As shown in Table 6, our SegVG outperforms the alternative method. This demonstrates that our bbox2seg paradigm is more effective than using a RES model to provide pseudo segmentation labels which might suffer from the errors from the RES model.", + "bbox": [ + 212, + 549, + 787, + 686 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.7 Confidence Score Analysis", + "text_level": 1, + "bbox": [ + 215, + 708, + 478, + 723 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Selection of segmentation query To show the effect of different selection of segmentation query for the calculation of confidence score, we calculate AP50 on RefCOCOg test-u. As shown in Table 7, the performance variations are slight among them. We use the first one to calculate confidence score for simplicity.", + "bbox": [ + 214, + 733, + 785, + 794 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Confidence Score Faithfulness To evaluate the faithfulness of our confidence score, i.e., whether a higher confidence score indicates better performance, we", + "bbox": [ + 214, + 809, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "W. Kang et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/60e14df3e9978e3090ace5b2d575b2626131b34f201cadfef4ea883ab81a9dc4.jpg", + "image_caption": [ + "Fig. 3: IoU and Accuracy of different confidence scores on RefCOCOg test-u." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 480, + 287 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/1a91b373bb385194d7a4bcc94bba2fb480539a1ac9fadf1589ff014082bc0c25.jpg", + "table_caption": [ + "Table 8: Performance of different confidence levels on RefCOCOg test-u." + ], + "table_footnote": [], + "table_body": "
ConfidenceIoUAcc(%)Proportion (%)
≥ 0.650.706777.41%100.00%
≥ 0.700.707277.48%99.89%
≥ 0.750.709177.69%99.44%
≥ 0.800.714378.37%97.98%
≥ 0.850.722679.29%95.04%
", + "bbox": [ + 529, + 213, + 785, + 297 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/c91bfc47273881b5edc3e6de06845807920d290f47f514cf1edd7a5cb420201f.jpg", + "image_caption": [ + "Fig. 4: Qualitative comparison between (c) (the first line) and (d) (the second line) of Table. 3. Red boxes are ground truth. Blue boxes are model predictions." + ], + "image_footnote": [], + "bbox": [ + 271, + 338, + 733, + 534 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "assess the relationship of our confidence score and model performance metrics (IoU and Accuracy) as shown in Fig 3. We sort the RefCOCOg-umd test set by confidence score, split it into five equal parts, and calculate each part's average score and performance. We observe a positive correlation between the performance metrics and our confidence score, which confirms its faithfulness.", + "bbox": [ + 212, + 614, + 787, + 689 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Confidence Score Application In real applications, the confidence score can be used to enhance the model's reliability. Specifically, we can apply different confidence thresholds to achieve different predictions, as shown in Table 8. First, we observe that accuracy increases with higher thresholds, indicating that adjusting the threshold can enhance the model's localization ability. Furthermore, the mean IoU also increases with the increasing threshold. Therefore, in downstream applications, such as using the model to provide pseudo-labels, we can increase the threshold to obtain a more accurate box. Since low-confidence outputs are excluded, the output proportion is slightly reduced, i.e., yields fewer outputs.", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Transferring Object Bounding Box to Segmentation for Visual Grounding", + "bbox": [ + 240, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/7adbdf7fdbc2d748611024734ad73347d1709ec5c78a791704222e5fa6d12ba1.jpg", + "image_caption": [ + "Fig. 5: Layer $i$ refers to the output of the $i$ -th layer of the decoder. The blue boxes represent the models' predictions, while the red boxes denote the ground truth. In the segmentation mask shown in the second column from the right, red indicates high confidence for the foreground. Note that VLTVG does not provide segmentation output." + ], + "image_footnote": [], + "bbox": [ + 245, + 142, + 759, + 323 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "4.8 Qualitative Results", + "text_level": 1, + "bbox": [ + 215, + 420, + 419, + 435 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We compare Tab. 3 (c) and (d) qualitatively to show the robustness enhancement. As shown in Fig. 4, adding segmentation queries improves the robustness to distinguish the target from the distractor, e.g. the case of two dogs.", + "bbox": [ + 212, + 446, + 784, + 492 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "As depicted in Fig. 5, we compare the box prediction quality by each decoding layer of SegVG with VLTVG [43] which also involves multi-layer supervision. As seen in the upper two rows of Fig. 5, VLTVG initially misses the target \"young man\" but improves its prediction gradually and finally makes the correct prediction. In contrast, due to the full exploitation of the annotations and the domain alignment in our Triple Alignment, SegVG successfully identifies the location of the target in the early decoding layer and consistently makes the correct prediction in each layer. Another example can be observed in the lower two rows of Fig. 5. In this image, due to the complex colors, VLTVG fails to locate the target \"plate\" and consistently repeats the same mistake. Instead, SegVG correctly detects the target, even in the first decoder layer. Additionally, we visualize the segmentation mask obtained by SegVG in Fig. 5, which accurately identifies the target box with high confidence. This behavior aligns with the box regression, demonstrating their shared objective, i.e., to distinguish the box.", + "bbox": [ + 212, + 493, + 787, + 705 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 729, + 359, + 746 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We propose SegVG, a visual grounding model, where we iteratively make the best of box annotations to involve pixel-level supervision and address the domain discrepancy among queries, text, and vision by Triple Alignment module. Experiments show the superior performance of SegVG. Furthermore, we explore the reliability benefits of our segmentation output in real-world applications.", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "W. Kang et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 145, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This research is supported by NSF IIS-2309073 and ECCS-2123521. This article solely reflects the opinions and conclusions of its authors and not the funding agencies.", + "bbox": [ + 212, + 178, + 787, + 223 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 248, + 321, + 263 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Anderson, P., He, X., Buehler, C., Teney, D., Johnson, M., Gould, S., Zhang, L.: Bottom-up and top-down attention for image captioning and visual question answering. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6077-6086 (2018)", + "2. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020)", + "3. Chen, L., Ma, W., Xiao, J., Zhang, H., Chang, S.F.: Ref-nms: Breaking proposal bottlenecks in two-stage referring expression grounding. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 1036-1044 (2021)", + "4. Chen, S., Jin, Q., Wang, P., Wu, Q.: Say as you wish: Fine-grained control of image caption generation with abstract scene graphs. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9962-9971 (2020)", + "5. Chen, X., Ma, L., Chen, J., Jie, Z., Liu, W., Luo, J.: Real-time referring expression comprehension by single-stage grounding network. arXiv preprint arXiv:1812.03426 (2018)", + "6. Deng, J., Yang, Z., Chen, T., Zhou, W., Li, H.: Transvg: End-to-end visual grounding with transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1769-1779 (2021)", + "7. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)", + "8. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)", + "9. Escalante, H.J., Hernández, C.A., Gonzalez, J.A., López-López, A., Montes, M., Morales, E.F., Sucar, L.E., Villasenor, L., Grubinger, M.: The segmented and annotated iapr tc-12 benchmark. Computer vision and image understanding 114(4), 419-428 (2010)", + "10. Gan, C., Li, Y., Li, H., Sun, C., Gong, B.: Vqs: Linking segmentations to questions and answers for supervised attention in vqa and question-focused semantic segmentation. In: Proceedings of the IEEE international conference on computer vision. pp. 1811-1820 (2017)", + "11. He, K., Gkioxari, G., Dollár, P., Girshick, R.: Mask r-cnn. In: Proceedings of the IEEE international conference on computer vision. pp. 2961-2969 (2017)", + "12. Hu, R., Rohrbach, M., Andreas, J., Darrell, T., Saenko, K.: Modeling relationships in referential expressions with compositional modular networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1115-1124 (2017)" + ], + "bbox": [ + 218, + 281, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Transferring Object Bounding Box to Segmentation for Visual Grounding", + "bbox": [ + 240, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "13. Huang, B., Lian, D., Luo, W., Gao, S.: Look before you leap: Learning landmark features for one-stage visual grounding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16888-16897 (2021)", + "14. Kang, W., Qu, M., Kini, J., Wei, Y., Shah, M., Yan, Y.: Intent3d: 3d object detection in rgb-d scans based on human intention. arXiv preprint arXiv:2405.18295 (2024)", + "15. Kang, W., Qu, M., Wei, Y., Yan, Y.: Actress: Active retraining for semi-supervised visual grounding. arXiv preprint arXiv:2407.03251 (2024)", + "16. Kang, W., Zhou, L., Wu, J., Sun, C., Yan, Y.: Visual grounding with attention-driven constraint balancing. arXiv preprint arXiv:2407.03243 (2024)", + "17. Kazemzadeh, S., Ordonez, V., Matten, M., Berg, T.: Referitgame: Referring to objects in photographs of natural scenes. In: Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP). pp. 787-798 (2014)", + "18. Li, L.H., Zhang, P., Zhang, H., Yang, J., Li, C., Zhong, Y., Wang, L., Yuan, L., Zhang, L., Hwang, J.N., et al.: Grounded language-image pre-training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10965-10975 (2022)", + "19. Li, M., Sigal, L.: Referring transformer: A one-step approach to multi-task visual grounding. Advances in neural information processing systems 34, 19652-19664 (2021)", + "20. Liao, Y., Liu, S., Li, G., Wang, F., Chen, Y., Qian, C., Li, B.: A real-time cross-modality correlation filtering method for referring expression comprehension. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10880-10889 (2020)", + "21. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proceedings of the IEEE international conference on computer vision. pp. 2980-2988 (2017)", + "22. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)", + "23. Liu, D., Zhang, H., Wu, F., Zha, Z.J.: Learning to assemble neural module tree networks for visual grounding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4673-4682 (2019)", + "24. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023)", + "25. Luo, G., Zhou, Y., Sun, X., Cao, L., Wu, C., Deng, C., Ji, R.: Multi-task collaborative network for joint referring expression comprehension and segmentation. In: Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition. pp. 10034-10043 (2020)", + "26. Mao, J., Huang, J., Toshev, A., Camburu, O., Yuille, A.L., Murphy, K.: Generation and comprehension of unambiguous object descriptions. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 11-20 (2016)", + "27. Milletari, F., Navab, N., Ahmadi, S.A.: V-net: Fully convolutional neural networks for volumetric medical image segmentation. In: 2016 fourth international conference on 3D vision (3DV). pp. 565-571. IEEE (2016)", + "28. Nagaraja, V.K., Morariu, V.I., Davis, L.S.: Modeling context between objects for referring expression understanding. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14. pp. 792-807. Springer (2016)" + ], + "bbox": [ + 217, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "W. Kang et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "29. Plummer, B.A., Kordas, P., Kiapour, M.H., Zheng, S., Piramuthu, R., Lazebnik, S.: Conditional image-text embedding networks. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 249-264 (2018)", + "30. Plummer, B.A., Wang, L., Cervantes, C.M., Caicedo, J.C., Hockenmaier, J., Lazebnik, S.: Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In: Proceedings of the IEEE international conference on computer vision. pp. 2641-2649 (2015)", + "31. Qu, M., Wu, Y., Liu, W., Gong, Q., Liang, X., Russakovsky, O., Zhao, Y., Wei, Y.: Siri: A simple selective retraining mechanism for transformer-based visual grounding. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXV. pp. 546-562. Springer (2022)", + "32. Redmon, J., Farhadi, A.: Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767 (2018)", + "33. Rezatofighi, H., Tsoi, N., Gwak, J., Sadeghian, A., Reid, I., Savarese, S.: Generalized intersection over union: A metric and a loss for bounding box regression. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 658-666 (2019)", + "34. Sadhu, A., Chen, K., Nevatia, R.: Zero-shot grounding of objects from natural language queries. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4694-4703 (2019)", + "35. Shang, Y., Cai, M., Xu, B., Lee, Y.J., Yan, Y.: Llava-prumerge: Adaptive token reduction for efficient large multimodal models. arXiv preprint arXiv:2403.15388 (2024)", + "36. Su, W., Miao, P., Dou, H., Wang, G., Qiao, L., Li, Z., Li, X.: Language adaptive weight generation for multi-task visual grounding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10857-10866 (2023)", + "37. Sun, J., Luo, G., Zhou, Y., Sun, X., Jiang, G., Wang, Z., Ji, R.: Refteacher: A strong baseline for semi-supervised referring expression comprehension. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19144-19154 (2023)", + "38. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)", + "39. Wang, L., Li, Y., Huang, J., Lazebnik, S.: Learning two-branch neural networks for image-text matching tasks. IEEE Transactions on Pattern Analysis and Machine Intelligence 41(2), 394-407 (2018)", + "40. Wang, P., Wu, Q., Cao, J., Shen, C., Gao, L., Hengel, A.v.d.: Neighbourhood watch: Referring expression comprehension via language-guided graph attention networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1960-1968 (2019)", + "41. Wang, X., Liu, Y., Shen, C., Ng, C.C., Luo, C., Jin, L., Chan, C.S., Hengel, A.v.d., Wang, L.: On the general value of evidence, and bilingual scene-text visual question answering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10126-10135 (2020)", + "42. Wang, Z., Lu, Y., Li, Q., Tao, X., Guo, Y., Gong, M., Liu, T.: Cris: Clip-driven referring image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11686-11695 (2022)", + "43. Yang, L., Xu, Y., Yuan, C., Liu, W., Li, B., Hu, W.: Improving visual grounding with visual-linguistic verification and iterative reasoning. In: Proceedings of the" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Transferring Object Bounding Box to Segmentation for Visual Grounding", + "bbox": [ + 240, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9499-9508 (2022)", + "44. Yang, S., Li, G., Yu, Y.: Dynamic graph attention for referring expression comprehension. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4644-4653 (2019)", + "45. Yang, S., Li, G., Yu, Y.: Graph-structured referring expression reasoning in the wild. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9952–9961 (2020)", + "46. Yang, Z., Wang, J., Tang, Y., Chen, K., Zhao, H., Torr, P.H.: Lavo: Language-aware vision transformer for referring image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18155-18165 (2022)", + "47. Yang, Z., Chen, T., Wang, L., Luo, J.: Improving one-stage visual grounding by recursive sub-query construction. In: European Conference on Computer Vision. pp. 387-404. Springer (2020)", + "48. Yang, Z., Gong, B., Wang, L., Huang, W., Yu, D., Luo, J.: A fast and accurate one-stage approach to visual grounding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4683-4693 (2019)", + "49. Ye, J., Tian, J., Yan, M., Yang, X., Wang, X., Zhang, J., He, L., Lin, X.: Shifting more attention to visual backbone: Query-modulated refinement networks for end-to-end visual grounding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15502-15512 (2022)", + "50. You, Q., Jin, H., Wang, Z., Fang, C., Luo, J.: Image captioning with semantic attention. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4651-4659 (2016)", + "51. Yu, L., Lin, Z., Shen, X., Yang, J., Lu, X., Bansal, M., Berg, T.L.: Matttnet: Modular attention network for referring expression comprehension. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1307-1315 (2018)", + "52. Yu, L., Poirson, P., Yang, S., Berg, A.C., Berg, T.L.: Modeling context in referring expressions. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14. pp. 69-85. Springer (2016)", + "53. Yu, Z., Yu, J., Xiang, C., Zhao, Z., Tian, Q., Tao, D.: Rethinking diversified and discriminative proposal generation for visual grounding. arXiv preprint arXiv:1805.03508 (2018)", + "54. Zareian, A., Rosa, K.D., Hu, D.H., Chang, S.F.: Open-vocabulary object detection using captions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14393-14402 (2021)", + "55. Zhang, H., Niu, Y., Chang, S.F.: Grounding referring expressions in images by variational context. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4158-4166 (2018)", + "56. Zhang, Z., Zhu, Y., Liu, J., Liang, X., Ke, W.: Coupling word-pixel with sentence-mask alignments for referring image segmentation. arXiv preprint arXiv:2212.01769 (2022)", + "57. Zhuang, B., Wu, Q., Shen, C., Reid, I., Van Den Hengel, A.: Parallel attention: A unified framework for visual object discovery through dialogs and queries. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4252-4261 (2018)" + ], + "bbox": [ + 217, + 146, + 785, + 825 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "W. Kang et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/SegVG_ Transferring Object Bounding Box to Segmentation for Visual Grounding/4d7de385-7f41-401e-b1cf-a9c2348c4d1c_model.json b/2024/SegVG_ Transferring Object Bounding Box to Segmentation for Visual Grounding/4d7de385-7f41-401e-b1cf-a9c2348c4d1c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..fd127a2004a18c16ceda8e13a137aa8cacc2e9e3 --- /dev/null +++ b/2024/SegVG_ Transferring Object Bounding Box to Segmentation for Visual Grounding/4d7de385-7f41-401e-b1cf-a9c2348c4d1c_model.json @@ -0,0 +1,2359 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.241, + 0.141, + 0.764, + 0.187 + ], + "angle": 0, + "content": "SegVG: Transferring Object Bounding Box to Segmentation for Visual Grounding" + }, + { + "type": "text", + "bbox": [ + 0.277, + 0.212, + 0.726, + 0.229 + ], + "angle": 0, + "content": "Weitai Kang\\(^{1}\\), Gaowen Liu\\(^{2}\\), Mubarak Shah\\(^{3}\\), and Yan Yan\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.391, + 0.239, + 0.611, + 0.253 + ], + "angle": 0, + "content": "1 Illinois Institute of Technology" + }, + { + "type": "text", + "bbox": [ + 0.444, + 0.253, + 0.56, + 0.266 + ], + "angle": 0, + "content": "2 Cisco Research" + }, + { + "type": "text", + "bbox": [ + 0.397, + 0.268, + 0.605, + 0.282 + ], + "angle": 0, + "content": "3 University of Central Florida" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.32, + 0.744, + 0.64 + ], + "angle": 0, + "content": "Abstract. Different from Object Detection, Visual Grounding deals with detecting a bounding box for each text-image pair. This one box for each text-image data provides sparse supervision signals. Although previous works achieve impressive results, their passive utilization of annotation, i.e. the sole use of the box annotation as regression ground truth, results in a suboptimal performance. In this paper, we present SegVG, a novel method transfers the box-level annotation as Segmentation signals to provide an additional pixel-level supervision for Visual Grounding. Specifically, we propose the Multi-layer Multi-task Encoder-Decoder as the target grounding stage, where we learn a regression query and multiple segmentation queries to ground the target by regression and segmentation of the box in each decoding layer, respectively. This approach allows us to iteratively exploit the annotation as signals for both box-level regression and pixel-level segmentation. Moreover, as the backbones are typically initialized by pretrained parameters learned from unimodal tasks and the queries for both regression and segmentation are static learnable embeddings, a domain discrepancy remains among these three types of features, which impairs subsequent target grounding. To mitigate this discrepancy, we introduce the Triple Alignment module, where the query, text, and vision tokens are triangurally updated to share the same space by triple attention mechanism. Extensive experiments on five widely used datasets validate our state-of-the-art (SOTA) performance. Code is available at https://github.com/WeitaiKang/SegVG." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.653, + 0.647, + 0.668 + ], + "angle": 0, + "content": "Keywords: Visual Grounding \\(\\cdot\\) Detection \\(\\cdot\\) Transformer" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.717, + 0.377, + 0.733 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.749, + 0.788, + 0.844 + ], + "angle": 0, + "content": "Visual grounding [14,17,26,30,52] aims to localize a target object within an image based on a free-form natural language text expression. It is particularly important for numerous downstream multimodal reasoning systems, such as visual question answering [10,35,41] and image captioning [1,4,50]. Previous works can be broadly categorized into three distinct groups: two-stage methods [3,44,45,51], one-stage methods [47,48], and transformer-based ones [6,15,16,31,43,49]. Both" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.129 + ], + "angle": 0, + "content": "W. Kang et al." + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.144, + 0.716, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.349, + 0.788, + 0.462 + ], + "angle": 0, + "content": "Fig. 1: The comparison of visual grounding frameworks. The block with a dashed border indicates that the module may not necessarily exist. (a) Previous baseline method consists of two backbones and additional transformer layers for target grounding, where a regression query is supervised to regress the box. Current SOTA methods further employ a text-to-visual module to align the visual features with text features. (b) Our method incorporates segmentation queries, which utilizes the box annotation at the pixel-level to segment the target. Additionally, we propose the Triple Alignment module to eliminate the domain discrepancy of the query, text, and vision features." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.492, + 0.785, + 0.552 + ], + "angle": 0, + "content": "two-stage and one-stage approaches use convolutional neural networks for candidate proposals and the selection of the best-matching candidate. Nonetheless, these approaches rely on intricate modules that employ manually-crafted techniques for performing language inference and multi-modal integration." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.788, + 0.718 + ], + "angle": 0, + "content": "Inspired by the success of the transformer [7, 8], TransVG [6] proposes a transformer-based pipeline. As shown in Fig. 1.(a), this pipeline extracts vision and text features via DETR [2] and BERT [7], respectively. To ground the target, they use the transformer encoder to fuse multimodal features along with a learnable regression query and decode the query through an MLP. To enhance the final target grounding stage, subsequent studies continue with some text-to-visual modules in the early stage to modulate the vision features to align with the text features. For example, QRNet [49] proposes a query-modulated method for extracting language-aware vision features within the vision backbone. VLTVG [43] introduces a verification map to activate the vision features to align with the text features before multimodal fusion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Despite their advancements, the suboptimal annotation utilization, i.e., only using the box annotation as a regression annotation, limits their performance. As discussed in [37], Visual Grounding presents unique challenges compared to Object Detection due to its sparse supervision signals. Specifically, it provides only one box label for each text-image pair, while necessitating detection within a multimodal setting. Therefore, it is essential to fully exploit the box annotation, by treating it as a segmentation mask (pixels within the bounding box are assigned a value of 1, while pixels outside the bounding box are assigned as 0)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.241, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Transferring Object Bounding Box to Segmentation for Visual Grounding" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.435 + ], + "angle": 0, + "content": "In this paper, we introduce SegVG (see Fig. 1.(b)), a novel method that leverages the pixel-level details within the box annotation as segmentation signals to offer additional fine-grained supervision for Visual Grounding. Specifically, we propose the Multi-layer Multi-task Encoder-Decoder as the target grounding stage, where we learn a regression query and multiple segmentation queries to ground the target by regression and segmentation of the box in each decoding layer, respectively. The confidence score derived from the segmentation can further serve as a Focal Loss [21] scaling factor to adaptively emphasize the other losses of challenging training samples. This approach allows us to iteratively exploit the annotation as signals for both box-level regression and pixel-level segmentation. Furthermore, the initial parameters for model backbones, typically derived from pretrained unimodal tasks, along with data-agnostic static embeddings used as queries for decoding, result in a domain discrepancy among different sources of feature, affecting the effectiveness of target grounding. To tackle this problem, we present the Triple Alignment module, where we harmonize the domain of query, text, and vision features by implementing a triangular update process through a triple attention mechanism. As a result, we ensure that all features adapt and integrate within the same multimodal space, thereby enhancing subsequent target grounding. Our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.447, + 0.784, + 0.491 + ], + "angle": 0, + "content": "- We propose the Multi-layer Multi-task Encoder-Decoder to maximize the utilization of the box annotation, which introduces an additional segmentation format for pixel-level supervision in Visual Grounding." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.493, + 0.785, + 0.537 + ], + "angle": 0, + "content": "- To eliminate the domain discrepancy among the query, text, and vision, we introduce the Triple Alignment to update these three types of features into a sharing domain, which facilitates the subsequent target grounding." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.539, + 0.785, + 0.599 + ], + "angle": 0, + "content": "- We conduct extensive experiments on five widely used datasets to show the performance superiority of our proposed methods compared with previous state-of-the-art methods and further investigate the reliability benefits derived from the segmentation output in real applications." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.6, + 0.744, + 0.615 + ], + "angle": 0, + "content": "- We will release code and checkpoints for future research development." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.447, + 0.785, + 0.615 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.653, + 0.388, + 0.669 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.784, + 0.718 + ], + "angle": 0, + "content": "Visual grounding methods can be roughly classified into three pipelines: two-stage methods, one-stage methods, and transformer-based methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Two-stage methods Two-stage approaches [3,51] treat visual grounding as first generating candidate object proposals and then finding the best match to the text. In the first stage, an off-the-shelf detector processes the image and proposes regions that may contain the target. In the second stage, a ranking network calculates the similarity between candidate regions and processed text features, selecting the region with the highest similarity score as the final result. Training losses include binary classification loss [29] or maximum-margin ranking loss [51]. To better understand the text and cross-modality matching, MattNet [51] focuses" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.129 + ], + "angle": 0, + "content": "W. Kang et al." + }, + { + "type": "image", + "bbox": [ + 0.242, + 0.146, + 0.764, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.393, + 0.788, + 0.506 + ], + "angle": 0, + "content": "Fig. 2: SegVG: The upper figure includes the vision and text backbone. Our proposed Triple Alignment module is iteratively inserted into intermediate layers to eliminate domain discrepancy. The lower figure shows our Multi-layer Multi-task Encoder-Decoder, which adopts a transformer encoder-decoder to update multimodal features and ground the target. In this architecture, we make the best of the box annotation as a segmentation ground truth and integrate an additional segmentation task into Visual Grounding. Additionally, the segmentation output serves as a Focal Loss factor, allowing adaptive emphasis on challenging cases for the regression loss. \\( \\mathrm{M} = 6 \\), \\( \\mathrm{R} = 6 \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.536, + 0.784, + 0.567 + ], + "angle": 0, + "content": "on decomposing the text into subject, location, and relationship components. [3] introduces an expression-aware score for improved candidate region ranking." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.785, + 0.658 + ], + "angle": 0, + "content": "One-stage methods One-stage approaches [47, 48] directly concatenate vision and text features in the channel dimension and rank confidence values for candidate regions proposed based on the concatenated multimodal features. For example, FAOA [48] predicts bounding boxes using a YOLOv3 detector [32] on the concatenated features. ReSC [47] further improves the ability to ground complex queries by introducing a recursive sub-query construction module." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Transformer-based methods Transformer-based approach is first introduced by TransVG [6]. Unlike previous methods, TransVG concatenates the regression query (a learnable embedding), vision tokens, and text tokens and uses transformer encoders [38] to perform cross-modal fusion and target grounding. The query is then processed through an MLP to decode the box. Benefiting from the flexible structure of transformer modules in processing multimodal features, recent works continue adopting this pipeline and propose novelties regarding feature extraction. VLTVG [43] develops a visual-linguistic verification module before the target grounding stage to modulate the vision features with the relationship between vision and text features. QRNet [49] proposes a Query-modulated Refinement Network to mitigate the gap between features from the unimodal vision backbone and those needed for multi-modal reasoning." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.241, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "Transferring Object Bounding Box to Segmentation for Visual Grounding" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.299 + ], + "angle": 0, + "content": "Multi-task Visual Grounding Multi-task learning is extensively utilized in object detection and segmentation [2,11], often capitalizing on a shared backbone and task-specific heads. Expanding upon this idea, several studies [19, 25, 36] have proposed solutions to the Multi-task Visual Grounding problem. In this problem, they jointly tackle Referring Expression Comprehension (REC, also known as Visual Grounding) and Referring Expression Segmentation (RES), requiring both box annotations and segmentation annotations. It is important to note that, unlike those approaches, even though we incorporate segmentation losses in our method, we do not need segmentation annotations but only box annotations, focusing specifically on the Visual Grounding task." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.321, + 0.381, + 0.34 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.353, + 0.787, + 0.399 + ], + "angle": 0, + "content": "In this section, we present the components of our SegVG in the order of the data flow: starting with the backbones, followed by our proposed Triple Alignment, and finally our Multi-layer Multi-task Encoder-Decoder." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.421, + 0.351, + 0.435 + ], + "angle": 0, + "content": "3.1 Backbones" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.446, + 0.788, + 0.509 + ], + "angle": 0, + "content": "As shown in Fig. 2 (upper), our vision backbone consists of ResNet and transformer encoder from DETR [2], with parameters pretrained on Object Detection task using the MSCOCO dataset [22], excluding the validation and test sets of Visual Grounding dataset. The text backbone is the base model of BERT [7]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.521, + 0.788, + 0.63 + ], + "angle": 0, + "content": "Vision Backbone Given an input image \\(\\mathbf{I_0}\\) \\((\\mathbb{R}^{3\\times H_0\\times W_0})\\), we employ ResNet to generate a 2D feature map \\(\\mathbf{I}\\in \\mathbb{R}^{C\\times H\\times W}\\) (\\(C = 2048\\), \\(H = \\frac{H_0}{32}\\), \\(W = \\frac{W_0}{32}\\)). A 1x1 convolutional layer is then used to reduce the channel dimension of \\(\\mathbf{I}\\) to \\(C_v = 256\\), resulting in \\(\\mathbf{I}'\\). We further flatten \\(\\mathbf{I}'\\) into \\(\\mathbf{Z}_v\\in \\mathbb{R}^{C_v\\times N_v}\\) (\\(N_v = H\\times W\\)). Position embedding is then added to \\(\\mathbf{Z}_v\\) to preserve sensitivity to the original 2D spatial locations. \\(\\mathbf{Z}_v\\) is then iteratively processed through DETR's encoder layer (total 6 transformer layers) and the Triple Alignment to obtain the output \\(\\mathbf{Z}_v\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.642, + 0.788, + 0.735 + ], + "angle": 0, + "content": "Text Backbone Given a text, we initially utilize the BERT's embedding layer to convert it into \\( N_{t} \\) language tokens with \\( C_t \\) channel dimension. In alignment with [6], we preprocess a [CLS] token and append a [SEP] token to the beginning and end positions of the tokenized language, respectively. Following this, we iteratively input the language tokens into BERT's layers (total 12 transformer layers) and the Triple Alignment, generating language embedding \\( \\mathbf{Z}_t \\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.755, + 0.407, + 0.772 + ], + "angle": 0, + "content": "3.2 Triple Alignment" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Given that the text and vision backbones are pretrained from unimodal tasks and the queries are data-agnostic, the subsequent target grounding stage faces the challenge of aligning these three types of features into the same space before performing multimodal fusion for target grounding. Additionally, considering" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.129 + ], + "angle": 0, + "content": "W. Kang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "that the backbones usually contribute the majority of the overall parameters, solely using them for extracting unimodal features without incorporating multimodal alignment is sub-optimal. Therefore, an optimal solution is to address the domain discrepancy before moving on to the subsequent target grounding stage." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.785, + 0.283 + ], + "angle": 0, + "content": "Triple Alignment (Fig. 2 (upper)) utilizes an attention mechanism to perform triangular feature sampling, aiming to ensure domain consistency among the query, text, and vision features. The queries, \\(\\mathbf{Z}_o\\), are first initialized by N learnable embeddings, where one embedding is for the regression query and the rest of the embeddings are for multiple segmentation queries. The data flow is:" + }, + { + "type": "equation", + "bbox": [ + 0.345, + 0.291, + 0.785, + 0.309 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Z} _ {v} ^ {i + 1} = \\operatorname {D E T R L a y e r} _ {i} \\left(\\mathbf {Z} _ {v} ^ {i}\\right), i \\in \\{0, 1,.., L - 1 \\} \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.346, + 0.329, + 0.785, + 0.348 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Z} _ {t} ^ {i + 1} = \\operatorname {B E R T L a y e r} _ {\\mathcal {2 i + 1}} (\\operatorname {B E R T L a y e r} _ {\\mathcal {2 i}} (\\mathbf {Z} _ {t} ^ {i})), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.36, + 0.364, + 0.785, + 0.383 + ], + "angle": 0, + "content": "\\[\n\\left[ \\mathbf {Z} _ {o} ^ {\\prime}, \\mathbf {Z} _ {t} ^ {\\prime}, \\mathbf {Z} _ {v} ^ {\\prime} \\right] = \\operatorname {T r i - M H A} \\left(\\mathbf {Z} _ {o}, \\mathbf {Z} _ {t} ^ {i + 1}, \\mathbf {Z} _ {v} ^ {i + 1}\\right), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.317, + 0.399, + 0.785, + 0.418 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Z} _ {o} = \\mathbf {Z} _ {o} + \\mathbf {Z} _ {o} ^ {\\prime}, \\mathbf {Z} _ {t} ^ {i + 1} = \\mathbf {Z} _ {t} ^ {i + 1} + \\mathbf {Z} _ {t} ^ {\\prime}, \\mathbf {Z} _ {v} ^ {i + 1} = \\mathbf {Z} _ {v} ^ {i + 1} + \\mathbf {Z} _ {v} ^ {\\prime}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.421, + 0.784, + 0.542 + ], + "angle": 0, + "content": "where L is the number of layers, BERTLayer is the layer of BERT and DETRLayer is the layer of DETR's encoder. The vision and text features are first encoded by Eq. 1 and Eq. 2. Subsequently, the three types of tokens (query, text, and vision) are updated by our Triple Multi-Head Attention Layer (Tri-MHA) using Eq. 3. The output tokens are merged back to their original branches respectively by Eq. 4. Within each head of the Triple Multi-Head Attention Layer (Tri-MHA), each type of the features simultaneously computes its updated representation by attending to both the others and itself:" + }, + { + "type": "equation", + "bbox": [ + 0.323, + 0.55, + 0.682, + 0.568 + ], + "angle": 0, + "content": "\\[\nS = \\left[ \\mathbf {Z} _ {o} W ^ {(o, S)}, \\mathbf {Z} _ {t} W ^ {(t, S)}, \\mathbf {Z} _ {v} W ^ {(v, S)} \\right], S \\in \\{Q, K, V \\}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.379, + 0.572, + 0.785, + 0.59 + ], + "angle": 0, + "content": "\\[\n[ \\mathbf {Z} _ {o}, \\mathbf {Z} _ {t}, \\mathbf {Z} _ {v} ] = \\operatorname {S o f t M a x} \\left(Q K ^ {T} / \\sqrt {d _ {k}}\\right) V, \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.412, + 0.595, + 0.593, + 0.612 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Z} _ {e} ^ {\\prime} = \\mathbf {Z} _ {e} W ^ {e}, e \\in \\{o, t, v \\},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.62, + 0.784, + 0.667 + ], + "angle": 0, + "content": "where \\(\\{W^{(e,S)}, W^e : e \\in \\{o, t, v\\}, S \\in \\{Q, K, V\\}\\}\\) are trainable parameters. As a result, each of the output features is triangular sampling from all of the three types of features, which alleviates the domain discrepancy." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.686, + 0.595, + 0.7 + ], + "angle": 0, + "content": "3.3 Multi-layer Multi-task Encoder-Decoder" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.708, + 0.784, + 0.753 + ], + "angle": 0, + "content": "The Multi-layer Multi-task Encoder-Decoder serves as the target grounding stage, where we use a transformer encoder-decoder for cross-modal fusion and target grounding to perform a box regression task and a box segmentation task." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Encoder As shown in Fig. 2 (lower left), given the aligned output text and vision features from the backbones, the encoder fuses the two modalities into the multimodal features by a stack of transformer layers. In each layer, the concatenated text and vision tokens go through the Multi-Head Self-Attention layer (MHSA) and the Feed Forward Network (FFN) with the residual connection." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.241, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Transferring Object Bounding Box to Segmentation for Visual Grounding" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.526 + ], + "angle": 0, + "content": "Decoder In each decoder layer, we aim to fully exploit the box annotation. We propose the bbox2seg paradigm to transform the box annotation into a segmentation mask, which classifies all pixels within the box as foreground (with a value of one) and those outside the box as background (with a value of zero). As shown in Fig. 2 (lower right), one regression query aims to regress the box, while the remaining segmentation queries aim to segment the box. Different segmentation queries are endowed with different learnable positional embeddings to enhance the robustness of each decoder layer, since the decoder layer, when confronted with various queries, is required to segment the same box. Following that, the queries pass through the Multi-Head Self-Attention layer to exchange information about the same target, prompting each other to better locate the target. Subsequently, the queries undergo the Multi-Head Cross-Attention layer and the Feed Forward Network, where multimodal features serve as the Key and Value to ground the target. Finally, a shared MLP across all decoder layers decodes the regression query into the box result, supervised by L1 loss and Giou loss [33]. Each segmentation query is repeated \\( N_{v} \\) times and concatenated with visual tokens along the channel dimension. Another shared MLP decodes the concatenated feature into the segmentation mask, supervised by Focal loss [21] and Dice loss [27]. It is noteworthy that our segmentation paradigm shares the same semantic foundation as the regression paradigm, i.e., to distinguish bounding box, rather than instance segmentation. Therefore, incorporating non-object pixels in the segmented foreground does not introduce ambiguity to the model. We provide qualitative results 4.8 to demonstrate this feature. To alleviate multi-task optimization challenges, we freeze the backbones for the initial k epochs to stabilize the training process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.536, + 0.788, + 0.642 + ], + "angle": 0, + "content": "Confidence score Since both the regression output and segmentation output share the same aim, we can additionally obtain the confidence score for the foreground by averaging values inside the ground truth box of the segmentation output to reflect the confidence of the regression output. In the training process, we can transform this confidence score as the Focal loss factor [21] to adaptively emphasize the other losses of challenging training samples. The final loss function of each decoder layer is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.377, + 0.649, + 0.786, + 0.674 + ], + "angle": 0, + "content": "\\[\nL = \\lambda_ {1} c _ {f o c a l} L _ {1} + \\lambda_ {g i o u} c _ {f o c a l} L _ {g i o u} + \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.411, + 0.669, + 0.63, + 0.684 + ], + "angle": 0, + "content": "\\[\n\\lambda_ {d i c e} c _ {f o c a l} L _ {d i c e} + \\lambda_ {f o c a l} L _ {f o c a l},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.788, + 0.735 + ], + "angle": 0, + "content": "where \\(\\lambda_{1}\\), \\(\\lambda_{giou}\\), \\(\\lambda_{focal}\\) and \\(\\lambda_{dice}\\) are hyperparameters. \\(L_{1}\\) is the L1 loss. \\(L_{giou}\\) is the GIoU loss [33]. \\(L_{focal}\\) is the Focal loss [21]. \\(L_{dice}\\) is the Dice loss [27]. \\(c_{focal}\\) is the above Focal loss factor averaged across all segmentation outputs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.84 + ], + "angle": 0, + "content": "In the real-world application perspective, the visual grounding task can be viewed as open-vocabulary object detection [54], where target objects lack predetermined categories. Therefore, previous transformer-based methods directly regress the box without confidence scores, since there is no candidate proposal or selection stage in transformer-based pipeline. However, confidence scores are valuable for enhancing the control or reliability of predictions by filtering out low-confidence predictions. This feature could benefit the future integration of visual" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.374, + 0.129 + ], + "angle": 0, + "content": "W. Kang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 1: Comparisons with state-of-the-art methods on widely used datasets. We highlight the best and second best performance in red and blue, and bold our model." + }, + { + "type": "table", + "bbox": [ + 0.252, + 0.185, + 0.756, + 0.46 + ], + "angle": 0, + "content": "
ModelsBackboneRefCOCORefCOCO+RefCOCOgReferItGame test
valtestAtestBvaltestAtestBval-gval-utest-u
Two-stage:
CMN [12]VGG16-71.0365.77-54.3247.7657.47--28.33
VC [55]VGG16-73.3367.44-58.4053.1862.30--31.13
ParalAttn [57]VGG16-75.3165.52-61.3450.8658.03---
MAttNet [51]ResNet-10176.6581.1469.9965.3371.6256.02-66.5867.2729.04
Similarity Net [39]ResNet-101---------34.54
CITE [29]ResNet-101---------35.07
DDPN [53]ResNet-101---------63.00
LGRANs [40]VGG16-76.6066.40-64.0053.4061.78---
DGA [44]VGG16-78.4265.53-69.0751.99--63.28-
RvG-Tree [29]ResNet-10175.0678.6169.8563.5167.4556.66-66.9566.51-
NMTree [23]ResNet-10176.4181.2170.0966.4672.0257.5264.6265.8766.44-
Ref-NMS [3]ResNet-10180.7084.0076.0468.2573.6859.42-70.5570.62-
One-stage:
SSG [5]DarkNet-53-76.5167.50-62.1449.2747.4758.80-54.24
ZSGNet [34]ResNet-50---------58.63
FAOA [48]DarkNet-5372.5474.3568.5056.8160.2349.6056.1261.3360.3660.67
RCCF [20]DLA-34-81.0671.85-70.3556.32--65.7363.79
ReSC-Large [47]DarkNet-5377.6380.4572.3063.5968.3656.8163.1267.3067.2064.60
LBYL-Net [13]DarkNet-5379.6782.9174.1568.6473.3859.4962.70--67.47
Transformer-based:
TransVG [6]ResNet-10181.0282.7278.3564.8270.7056.9467.0268.6767.7370.73
QRNet [49]Swin-S84.0185.8582.3472.9476.1763.8171.8973.0372.5274.61
VLTVG [43]ResNet-10184.7787.2480.4974.1978.9365.1772.9876.0474.1871.98
SegVG (ours)ResNet-10186.8489.4683.0777.1882.6367.5976.0178.3577.4275.59
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.489, + 0.788, + 0.596 + ], + "angle": 0, + "content": "grounding models into downstream multimodal reasoning systems or real-world applications. To meet the requirements of this feature, our approach incorporates a confidence score derived from the segmentation output during inference. Specifically, we calculate the model's confidence by averaging values greater than or equal to 0.35 (adopted from [42]) in the segmentation output of one segmentation query. Analyses 4.7 in the Experiment section demonstrate the faithfulness and benefits of incorporating this additional confidence score." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.618, + 0.377, + 0.635 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.649, + 0.433, + 0.663 + ], + "angle": 0, + "content": "4.1 Metric and Datasets" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.674, + 0.787, + 0.737 + ], + "angle": 0, + "content": "Metric A predicted bounding box is considered accurate if its Intersection over Union (IoU) with the ground-truth bounding box exceeds 0.5. In accordance with the established practices in preceding studies [6,43], we employ top-1 accuracy (measured in percentage) as the primary metric to assess our method." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.749, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Datasets There are five standard benchmarks: RefCOCO [52], RefCOCO+ [52], RefCOCOg-g [26], RefCOCOg-umd [26], and ReferItGame [17]. Four of them (RefCOCO, RefCOCO+, and RefCOCOg-(g/umd)) are all derived from MSCOCO [22]. RefCOCO consists of 19,994 images and 142,210 referring texts, which is divided into four subsets: a training set with 120,624 texts, a validation set with 10,834 texts, and two test sets (testA and testB) containing 5,657" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.241, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "Transferring Object Bounding Box to Segmentation for Visual Grounding" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.299 + ], + "angle": 0, + "content": "and 5,095 texts, respectively. RefCOCO+ includes 19,992 images and 141,564 referring texts, which is partitioned into four subsets: a training set with 120,191 texts, a validation set with 10,758 texts, and two test sets (testA and testB) containing 5,726 and 4,889 texts, respectively. RefCOCOg contains 25,799 images and 95,010 longer texts. Two widely accepted splitting methods—RefCOCOg-g [26] and RefCOCOg-umd [28]—are employed for this dataset, and we perform experiments using both RefCOCOg-g (val-g) and RefCOCOg-umd (val-u and test-u) splitting conventions. The ReferItGame dataset, featuring 20,000 images from SAIAPR-12 [9], is divided into three segments: a training set with 54,127 texts, a validation set with 5,842 texts, and a testing set comprising 60,103 texts." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.323, + 0.396, + 0.339 + ], + "angle": 0, + "content": "4.2 Implementation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.35, + 0.788, + 0.472 + ], + "angle": 0, + "content": "Our approach uses an input image size of \\(640 \\times 640\\) and sets the maximum expression length at 40. When resizing images, we preserve the original aspect ratio. The longer edge is resized to 640, and the shorter edge is padded to 640 with the value of zero. Texts exceeding 38 tokens are truncated, reserving a start position and an end position of the characters for the [CLS] and [SEP] token, respectively. If the text is shorter, empty tokens are added after the [SEP] token to reach an input length of 40. Paddings for the input image are not tracked by masks, while empty tokens of text employ masks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.472, + 0.789, + 0.624 + ], + "angle": 0, + "content": "We use the AdamW optimizer. The initial learning rate of 1e-5 is assigned to backbones, and 1e-4 to the rest parameters. Weight decay is set at 1e-4. The visual backbone is initialized with the DETR's backbone and encoder, while the language branch uses the basic BERT model. For the final results, our model is trained for 90 epochs, with the learning rate decreasing by a factor of 10 after 60 epochs. The k hyperparameter in Multi-layer Multi-task Encoder-Decoder is set to 10. We use a batch size of 64. For the ablation studies presented in Table 3, the models are trained for 60 epochs with k equal to 20, and the learning rate drops after 40 epochs. We set \\(\\lambda_{1} = 5\\), \\(\\lambda_{giou} = 2\\), \\(\\lambda_{focal} = 1\\) and \\(\\lambda_{dice} = 1\\). We adhere to previous practices [6, 43] for data augmentation during training." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.647, + 0.434, + 0.663 + ], + "angle": 0, + "content": "4.3 Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.842 + ], + "angle": 0, + "content": "We report the performance of our SegVG on all benchmark datasets. As presented in Table 1, our SegVG model demonstrates superiority across all of the datasets. This indicates the effectiveness and generalizability of our approach. It is worth noting that RefCOCO+ and RefCOCOg are relatively more challenging datasets, as RefCOCO+ does not include location terms in its language expressions, and RefCOCOg has longer language expressions compared to other datasets. Despite these challenges, our model exhibits significant improvements on these two difficult datasets. Specifically, on RefCOCO+, our model outperforms the previous SOTA models with \\(+2.99\\%\\), \\(+3.7\\%\\), and \\(+2.42\\%\\) on the val, testA, and testB subsets, respectively. On RefCOCOg, our model also surpasses the previous SOTA models with \\(+3.03\\%\\), \\(+2.31\\%\\), and \\(+3.24\\%\\) on the val-g," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.129 + ], + "angle": 0, + "content": "W. Kang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.23, + 0.145, + 0.772, + 0.16 + ], + "angle": 0, + "content": "Table 2: Compare transformer-based models on Parameter count and GFLOPS." + }, + { + "type": "table", + "bbox": [ + 0.28, + 0.171, + 0.724, + 0.235 + ], + "angle": 0, + "content": "
BackboneParameter count (M)GFLOPS (G)
TransVG [6]ResNet101141.5572.61
VLTVG [43]ResNet101141.6169.87
QRNet [49]Swin-S247.0680.12
SegVGResNet101155.2873.48
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.247, + 0.789, + 0.318 + ], + "angle": 0, + "content": "Table 3: Ablation study on RefCOCOg test-u. Encoder and Decoder are the encoder and decoder of Multi-layer Multi-task Encoder-Decoder, respectively. MMDecoder represents Multi-layer and Multi-task supervision in the Decoder. N is the number of segmentation queries. Triple means Triple Alignment. Excluding Query in Triple (g) means using bidirectional alignment with concatenated text-vision tokens." + }, + { + "type": "table", + "bbox": [ + 0.285, + 0.328, + 0.726, + 0.466 + ], + "angle": 0, + "content": "
IdModelAcc(%)
(a)Backbones + Encoder + Decoder66.97
(b)Backbones + Triple + Encoder + Decoder75.75
(c)Backbones + Encoder + MMDecoder (N=1)72.61
(d)Backbones + Encoder + MMDecoder (N=5)76.21
(e)Backbones + Triple + Encoder + MMDecoder (N=5)77.29
(f)(e) w/o Encoder76.65(-0.64)
(g)(e) w/o Query in Triple76.37(-0.92)
(h)(e) w/o Focal loss76.83(-0.46)
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.789, + 0.568 + ], + "angle": 0, + "content": "val-u, and test-u subsets, respectively. These results suggest that under the reinforcement of Triple Alignment and Multi-layer Multi-task Encoder-Decoder, the query, text, and vision tokens are triangularly updated to share the same space, and the model fully exploits the bounding box as fine-grained pixel-level supervision for comprehensive end-to-end learning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.79, + 0.614 + ], + "angle": 0, + "content": "We also conduct a comparison of the number of parameters and GFLOPS across transformer-based models to evaluate computational costs. As depicted in Table 2, the computational cost of SegVG falls within a reasonable range." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.634, + 0.391, + 0.65 + ], + "angle": 0, + "content": "4.4 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.789, + 0.795 + ], + "angle": 0, + "content": "In this section, we aim to validate the efficacy of each proposed module. We conduct ablation studies on the RefCOCOg-umd test dataset. Specifically, we start by evaluating a basic structure, i.e., the backbones with encoder-decoder structure. After that, we systematically incorporate the Triple Alignment module into the backbones and introduce the Multi-layer Multi-task supervision into the decoder through the controlled variable approach. Meanwhile, we conduct additional ablation experiments on specific details, including assessing the efficacy of incorporating the encoder, introducing Query in the Triple Alignment, and introducing the Focal loss from the segmentation output to the other losses." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.789, + 0.842 + ], + "angle": 0, + "content": "As shown in Table 3, we can draw the following conclusions when comparing the experimental results under controlled variables: 1) [(a) v.s. (b)]: Incorporating the Triple Alignment can effectively eliminate the domain discrepancy" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.241, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "Transferring Object Bounding Box to Segmentation for Visual Grounding" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.189 + ], + "angle": 0, + "content": "Table 4: Ablation study on RefCOCOg test-u regarding the number of segmentation queries. \\( \\Delta \\) Hour and \\( \\Delta \\) Acc are the additional time cost and accuracy improvement compared to ID(i) experiment, respectively." + }, + { + "type": "table", + "bbox": [ + 0.297, + 0.199, + 0.712, + 0.276 + ], + "angle": 0, + "content": "
IDnum_queryAccTime CostΔHour / ΔAcc (v.s. ID(i))
iN = 172.6116.32h-
iiN = 373.0218.97h6.46
iiiN = 576.2120.10h1.05
ivN = 774.3822.24h3.34
vN = 973.9124.07h5.96
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.305, + 0.789, + 0.683 + ], + "angle": 0, + "content": "among the query, text, and vision features, thereby facilitating subsequent target grounding. 2) [(a) v.s. (c)]: Introducing Multi-layer Multi-task supervision can iteratively make the best of the annotations in the target grounding stage, thereby enhancing the learning of query representations. 3) [(c) v.s. (d)]: Increasing the number of segmentation queries can further improve the robustness of the decoder when provided with different queries and required to segment the same box. 4) [(a), (b), (d), and (e)]: Combining the Triple Alignment and Multi-layer Multi-task Encoder-Decoder can effectively enhance the overall performance, achieving the optimal result. 5) [(e) v.s. (f)]: Even if we include the Triple Alignment supporting multimodal communication, it remains necessary for the subsequent encoder to update the unimodal features generated by the backbones into multimodal features. 6) [(e) v.s. (g)]: It is necessary to involve queries in Triple Alignment to transfer the data-agnostic embedding into data-related queries. Otherwise, using only Bidirectional Alignment (BA) for text and vision tokens, similar to approaches like Deep Fusion in GLIP [18] and Grounding DINO [24], and WPA in CoupAlign [56], causes a noticeable decline \\((-0.92\\%)\\). 7) [(d), (e) and (g)]: The slight gain of (g)(76.37%) over (d)(76.21%) stems from our MMDecoder's pixel-level signals, which already boosts BA in the Encoder. Thus, extra BA effort in the backbone is marginal, failing to solve the unaligned query issue. Instead, Tri-Align (e)(77.29%) can solve this issue, showing its novelty. Notably, (d)(76.21%), with a basic encoder-decoder, already achieves SOTA performance, emphasizing our bbox2seg paradigm's simplicity and effectiveness. 8) [(e) v.s. (h)]: The segmentation output can further derive the confidence score of the model's prediction, which is transformed into the Focal loss factor to adaptively scale the other losses to focus more attention on challenging cases." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.684, + 0.788, + 0.759 + ], + "angle": 0, + "content": "We further conduct a more detailed abaltion study about the improvement and corresponding cost of adding more segmentation queries of Tab.3(c)-(d). As shown in Table 4, the best performance is observed with five segmentation queries. Adding more than that increases the burden from pixel-level constraint without benefits, also increasing computational cost per accuracy improvement." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.783, + 0.48, + 0.799 + ], + "angle": 0, + "content": "4.5 Triple Alignment Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "In addition to the improvement results in ablation study, we further deepen our understanding of the Triple Alignment by analyzing the attention behavior." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.374, + 0.129 + ], + "angle": 0, + "content": "W. Kang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.784, + 0.16 + ], + "angle": 0, + "content": "Table 5: Attention values from query to text-referred visual region in Triple Alignment." + }, + { + "type": "table", + "bbox": [ + 0.281, + 0.171, + 0.718, + 0.232 + ], + "angle": 0, + "content": "
LayerRefCOCORefCOCO+RefCOCOgReferItGame test
valtestAtestBvaltestAtestBval-gval-utest-utest
layer217%18%16%19%21%18%20%18%17%25%
layer434%36%33%29%31%28%27%33%33%30%
layer661%63%60%55%58%50%55%59%59%38%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.244, + 0.493, + 0.286 + ], + "angle": 0, + "content": "Table 6: Comparison with alternative method using RIS to provide pseudo segmentation labels for supervision." + }, + { + "type": "table", + "bbox": [ + 0.24, + 0.298, + 0.475, + 0.337 + ], + "angle": 0, + "content": "
ReferItGameSegVG+RIS (LAVT)SegVG
val74.9176.85
test73.7675.06
" + }, + { + "type": "table_caption", + "bbox": [ + 0.51, + 0.25, + 0.787, + 0.291 + ], + "angle": 0, + "content": "Table 7: AP50 using different segmentation queries for confidence score calculation on RefCOCOg test-u." + }, + { + "type": "table", + "bbox": [ + 0.533, + 0.303, + 0.769, + 0.332 + ], + "angle": 0, + "content": "
Seg Query1st2nd3rd4th5th
AP5084.6584.5784.7384.7884.63
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.365, + 0.788, + 0.502 + ], + "angle": 0, + "content": "Specifically, we calculate the sum of the attention values from the query to the text-referred visual region (target bbox) as a percentage of the total attention (which includes attention to the query, text, and visual tokens) to illustrate the extent of alignment across these three modalities in the second, fourth, and final layer of Triple Alignment. We average the percentage across all the attention heads and queries, and perform the analysis across all the datasets. As shown in Table 5, in all the datasets, the attention values increase as the layer progresses, indicating that Triple Alignment progressively aligns the query to comprehend the text and then focus on the referred visual region." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.525, + 0.567, + 0.54 + ], + "angle": 0, + "content": "4.6 Comparison with alternative method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.55, + 0.788, + 0.687 + ], + "angle": 0, + "content": "Given the development of Referring Expression Segmentation (RES), a natural alternative method would be using a RES method to generate pseudo segmentation labels to substitute our bbox2seg paradigm. Therefore, to mimic real-world scenarios, we use LAVT [46] trained on RefCOCO to obtain pseudo segmentation labels on ReferItGame. We follow the same training setting in ablation study to conduct the comparison on ReferItGame. As shown in Table 6, our SegVG outperforms the alternative method. This demonstrates that our bbox2seg paradigm is more effective than using a RES model to provide pseudo segmentation labels which might suffer from the errors from the RES model." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.709, + 0.48, + 0.724 + ], + "angle": 0, + "content": "4.7 Confidence Score Analysis" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.734, + 0.787, + 0.795 + ], + "angle": 0, + "content": "Selection of segmentation query To show the effect of different selection of segmentation query for the calculation of confidence score, we calculate AP50 on RefCOCOg test-u. As shown in Table 7, the performance variations are slight among them. We use the first one to calculate confidence score for simplicity." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Confidence Score Faithfulness To evaluate the faithfulness of our confidence score, i.e., whether a higher confidence score indicates better performance, we" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.241, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Transferring Object Bounding Box to Segmentation for Visual Grounding" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.481, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.297, + 0.481, + 0.325 + ], + "angle": 0, + "content": "Fig. 3: IoU and Accuracy of different confidence scores on RefCOCOg test-u." + }, + { + "type": "table_caption", + "bbox": [ + 0.521, + 0.185, + 0.788, + 0.213 + ], + "angle": 0, + "content": "Table 8: Performance of different confidence levels on RefCOCOg test-u." + }, + { + "type": "table", + "bbox": [ + 0.53, + 0.214, + 0.787, + 0.298 + ], + "angle": 0, + "content": "
ConfidenceIoUAcc(%)Proportion (%)
≥ 0.650.706777.41%100.00%
≥ 0.700.707277.48%99.89%
≥ 0.750.709177.69%99.44%
≥ 0.800.714378.37%97.98%
≥ 0.850.722679.29%95.04%
" + }, + { + "type": "image", + "bbox": [ + 0.272, + 0.339, + 0.734, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.545, + 0.785, + 0.574 + ], + "angle": 0, + "content": "Fig. 4: Qualitative comparison between (c) (the first line) and (d) (the second line) of Table. 3. Red boxes are ground truth. Blue boxes are model predictions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.615, + 0.788, + 0.69 + ], + "angle": 0, + "content": "assess the relationship of our confidence score and model performance metrics (IoU and Accuracy) as shown in Fig 3. We sort the RefCOCOg-umd test set by confidence score, split it into five equal parts, and calculate each part's average score and performance. We observe a positive correlation between the performance metrics and our confidence score, which confirms its faithfulness." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Confidence Score Application In real applications, the confidence score can be used to enhance the model's reliability. Specifically, we can apply different confidence thresholds to achieve different predictions, as shown in Table 8. First, we observe that accuracy increases with higher thresholds, indicating that adjusting the threshold can enhance the model's localization ability. Furthermore, the mean IoU also increases with the increasing threshold. Therefore, in downstream applications, such as using the model to provide pseudo-labels, we can increase the threshold to obtain a more accurate box. Since low-confidence outputs are excluded, the output proportion is slightly reduced, i.e., yields fewer outputs." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.129 + ], + "angle": 0, + "content": "W. Kang et al." + }, + { + "type": "image", + "bbox": [ + 0.246, + 0.143, + 0.761, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.334, + 0.788, + 0.391 + ], + "angle": 0, + "content": "Fig. 5: Layer \\( i \\) refers to the output of the \\( i \\)-th layer of the decoder. The blue boxes represent the models' predictions, while the red boxes denote the ground truth. In the segmentation mask shown in the second column from the right, red indicates high confidence for the foreground. Note that VLTVG does not provide segmentation output." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.421, + 0.421, + 0.436 + ], + "angle": 0, + "content": "4.8 Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.785, + 0.493 + ], + "angle": 0, + "content": "We compare Tab. 3 (c) and (d) qualitatively to show the robustness enhancement. As shown in Fig. 4, adding segmentation queries improves the robustness to distinguish the target from the distractor, e.g. the case of two dogs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.494, + 0.788, + 0.706 + ], + "angle": 0, + "content": "As depicted in Fig. 5, we compare the box prediction quality by each decoding layer of SegVG with VLTVG [43] which also involves multi-layer supervision. As seen in the upper two rows of Fig. 5, VLTVG initially misses the target \"young man\" but improves its prediction gradually and finally makes the correct prediction. In contrast, due to the full exploitation of the annotations and the domain alignment in our Triple Alignment, SegVG successfully identifies the location of the target in the early decoding layer and consistently makes the correct prediction in each layer. Another example can be observed in the lower two rows of Fig. 5. In this image, due to the complex colors, VLTVG fails to locate the target \"plate\" and consistently repeats the same mistake. Instead, SegVG correctly detects the target, even in the first decoder layer. Additionally, we visualize the segmentation mask obtained by SegVG in Fig. 5, which accurately identifies the target box with high confidence. This behavior aligns with the box regression, demonstrating their shared objective, i.e., to distinguish the box." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.731, + 0.36, + 0.747 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We propose SegVG, a visual grounding model, where we iteratively make the best of box annotations to involve pixel-level supervision and address the domain discrepancy among queries, text, and vision by Triple Alignment module. Experiments show the superior performance of SegVG. Furthermore, we explore the reliability benefits of our segmentation output in real-world applications." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.241, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Transferring Object Bounding Box to Segmentation for Visual Grounding" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.146, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.788, + 0.224 + ], + "angle": 0, + "content": "This research is supported by NSF IIS-2309073 and ECCS-2123521. This article solely reflects the opinions and conclusions of its authors and not the funding agencies." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.249, + 0.323, + 0.265 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.282, + 0.785, + 0.338 + ], + "angle": 0, + "content": "1. Anderson, P., He, X., Buehler, C., Teney, D., Johnson, M., Gould, S., Zhang, L.: Bottom-up and top-down attention for image captioning and visual question answering. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6077-6086 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.339, + 0.785, + 0.38 + ], + "angle": 0, + "content": "2. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.381, + 0.785, + 0.422 + ], + "angle": 0, + "content": "3. Chen, L., Ma, W., Xiao, J., Zhang, H., Chang, S.F.: Ref-nms: Breaking proposal bottlenecks in two-stage referring expression grounding. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 1036-1044 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.423, + 0.785, + 0.464 + ], + "angle": 0, + "content": "4. Chen, S., Jin, Q., Wang, P., Wu, Q.: Say as you wish: Fine-grained control of image caption generation with abstract scene graphs. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9962-9971 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.465, + 0.785, + 0.505 + ], + "angle": 0, + "content": "5. Chen, X., Ma, L., Chen, J., Jie, Z., Liu, W., Luo, J.: Real-time referring expression comprehension by single-stage grounding network. arXiv preprint arXiv:1812.03426 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.507, + 0.785, + 0.548 + ], + "angle": 0, + "content": "6. Deng, J., Yang, Z., Chen, T., Zhou, W., Li, H.: Transvg: End-to-end visual grounding with transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1769-1779 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.549, + 0.785, + 0.589 + ], + "angle": 0, + "content": "7. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.59, + 0.785, + 0.645 + ], + "angle": 0, + "content": "8. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.647, + 0.785, + 0.701 + ], + "angle": 0, + "content": "9. Escalante, H.J., Hernández, C.A., Gonzalez, J.A., López-López, A., Montes, M., Morales, E.F., Sucar, L.E., Villasenor, L., Grubinger, M.: The segmented and annotated iapr tc-12 benchmark. Computer vision and image understanding 114(4), 419-428 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.702, + 0.785, + 0.757 + ], + "angle": 0, + "content": "10. Gan, C., Li, Y., Li, H., Sun, C., Gong, B.: Vqs: Linking segmentations to questions and answers for supervised attention in vqa and question-focused semantic segmentation. In: Proceedings of the IEEE international conference on computer vision. pp. 1811-1820 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.758, + 0.785, + 0.785 + ], + "angle": 0, + "content": "11. He, K., Gkioxari, G., Dollár, P., Girshick, R.: Mask r-cnn. In: Proceedings of the IEEE international conference on computer vision. pp. 2961-2969 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.786, + 0.785, + 0.84 + ], + "angle": 0, + "content": "12. Hu, R., Rohrbach, M., Andreas, J., Darrell, T., Saenko, K.: Modeling relationships in referential expressions with compositional modular networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1115-1124 (2017)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.282, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.129 + ], + "angle": 0, + "content": "W. Kang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "13. Huang, B., Lian, D., Luo, W., Gao, S.: Look before you leap: Learning landmark features for one-stage visual grounding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16888-16897 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.787, + 0.229 + ], + "angle": 0, + "content": "14. Kang, W., Qu, M., Kini, J., Wei, Y., Shah, M., Yan, Y.: Intent3d: 3d object detection in rgb-d scans based on human intention. arXiv preprint arXiv:2405.18295 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.229, + 0.787, + 0.257 + ], + "angle": 0, + "content": "15. Kang, W., Qu, M., Wei, Y., Yan, Y.: Actress: Active retraining for semi-supervised visual grounding. arXiv preprint arXiv:2407.03251 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.257, + 0.787, + 0.283 + ], + "angle": 0, + "content": "16. Kang, W., Zhou, L., Wu, J., Sun, C., Yan, Y.: Visual grounding with attention-driven constraint balancing. arXiv preprint arXiv:2407.03243 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.283, + 0.787, + 0.324 + ], + "angle": 0, + "content": "17. Kazemzadeh, S., Ordonez, V., Matten, M., Berg, T.: Referitgame: Referring to objects in photographs of natural scenes. In: Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP). pp. 787-798 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.324, + 0.787, + 0.378 + ], + "angle": 0, + "content": "18. Li, L.H., Zhang, P., Zhang, H., Yang, J., Li, C., Zhong, Y., Wang, L., Yuan, L., Zhang, L., Hwang, J.N., et al.: Grounded language-image pre-training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10965-10975 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.378, + 0.787, + 0.418 + ], + "angle": 0, + "content": "19. Li, M., Sigal, L.: Referring transformer: A one-step approach to multi-task visual grounding. Advances in neural information processing systems 34, 19652-19664 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.418, + 0.787, + 0.473 + ], + "angle": 0, + "content": "20. Liao, Y., Liu, S., Li, G., Wang, F., Chen, Y., Qian, C., Li, B.: A real-time cross-modality correlation filtering method for referring expression comprehension. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10880-10889 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.473, + 0.787, + 0.514 + ], + "angle": 0, + "content": "21. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proceedings of the IEEE international conference on computer vision. pp. 2980-2988 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.514, + 0.787, + 0.569 + ], + "angle": 0, + "content": "22. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.569, + 0.787, + 0.609 + ], + "angle": 0, + "content": "23. Liu, D., Zhang, H., Wu, F., Zha, Z.J.: Learning to assemble neural module tree networks for visual grounding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4673-4682 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.609, + 0.787, + 0.65 + ], + "angle": 0, + "content": "24. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.65, + 0.787, + 0.704 + ], + "angle": 0, + "content": "25. Luo, G., Zhou, Y., Sun, X., Cao, L., Wu, C., Deng, C., Ji, R.: Multi-task collaborative network for joint referring expression comprehension and segmentation. In: Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition. pp. 10034-10043 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.704, + 0.787, + 0.744 + ], + "angle": 0, + "content": "26. Mao, J., Huang, J., Toshev, A., Camburu, O., Yuille, A.L., Murphy, K.: Generation and comprehension of unambiguous object descriptions. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 11-20 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.744, + 0.787, + 0.786 + ], + "angle": 0, + "content": "27. Milletari, F., Navab, N., Ahmadi, S.A.: V-net: Fully convolutional neural networks for volumetric medical image segmentation. In: 2016 fourth international conference on 3D vision (3DV). pp. 565-571. IEEE (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.786, + 0.787, + 0.84 + ], + "angle": 0, + "content": "28. Nagaraja, V.K., Morariu, V.I., Davis, L.S.: Modeling context between objects for referring expression understanding. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14. pp. 792-807. Springer (2016)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.241, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Transferring Object Bounding Box to Segmentation for Visual Grounding" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "29. Plummer, B.A., Kordas, P., Kiapour, M.H., Zheng, S., Piramuthu, R., Lazebnik, S.: Conditional image-text embedding networks. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 249-264 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.245 + ], + "angle": 0, + "content": "30. Plummer, B.A., Wang, L., Cervantes, C.M., Caicedo, J.C., Hockenmaier, J., Lazebnik, S.: Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In: Proceedings of the IEEE international conference on computer vision. pp. 2641-2649 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.785, + 0.3 + ], + "angle": 0, + "content": "31. Qu, M., Wu, Y., Liu, W., Gong, Q., Liang, X., Russakovsky, O., Zhao, Y., Wei, Y.: Siri: A simple selective retraining mechanism for transformer-based visual grounding. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXV. pp. 546-562. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.301, + 0.785, + 0.327 + ], + "angle": 0, + "content": "32. Redmon, J., Farhadi, A.: Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.384 + ], + "angle": 0, + "content": "33. Rezatofighi, H., Tsoi, N., Gwak, J., Sadeghian, A., Reid, I., Savarese, S.: Generalized intersection over union: A metric and a loss for bounding box regression. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 658-666 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.424 + ], + "angle": 0, + "content": "34. Sadhu, A., Chen, K., Nevatia, R.: Zero-shot grounding of objects from natural language queries. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4694-4703 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.466 + ], + "angle": 0, + "content": "35. Shang, Y., Cai, M., Xu, B., Lee, Y.J., Yan, Y.: Llava-prumerge: Adaptive token reduction for efficient large multimodal models. arXiv preprint arXiv:2403.15388 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.522 + ], + "angle": 0, + "content": "36. Su, W., Miao, P., Dou, H., Wang, G., Qiao, L., Li, Z., Li, X.: Language adaptive weight generation for multi-task visual grounding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10857-10866 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.785, + 0.577 + ], + "angle": 0, + "content": "37. Sun, J., Luo, G., Zhou, Y., Sun, X., Jiang, G., Wang, Z., Ji, R.: Refteacher: A strong baseline for semi-supervised referring expression comprehension. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19144-19154 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.618 + ], + "angle": 0, + "content": "38. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.785, + 0.66 + ], + "angle": 0, + "content": "39. Wang, L., Li, Y., Huang, J., Lazebnik, S.: Learning two-branch neural networks for image-text matching tasks. IEEE Transactions on Pattern Analysis and Machine Intelligence 41(2), 394-407 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.785, + 0.715 + ], + "angle": 0, + "content": "40. Wang, P., Wu, Q., Cao, J., Shen, C., Gao, L., Hengel, A.v.d.: Neighbourhood watch: Referring expression comprehension via language-guided graph attention networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1960-1968 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.785, + 0.77 + ], + "angle": 0, + "content": "41. Wang, X., Liu, Y., Shen, C., Ng, C.C., Luo, C., Jin, L., Chan, C.S., Hengel, A.v.d., Wang, L.: On the general value of evidence, and bilingual scene-text visual question answering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10126-10135 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.813 + ], + "angle": 0, + "content": "42. Wang, Z., Lu, Y., Li, Q., Tao, X., Guo, Y., Gong, M., Liu, T.: Cris: Clip-driven referring image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11686-11695 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "43. Yang, L., Xu, Y., Yuan, C., Liu, W., Li, B., Hu, W.: Improving visual grounding with visual-linguistic verification and iterative reasoning. In: Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.129 + ], + "angle": 0, + "content": "W. Kang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.241, + 0.147, + 0.786, + 0.175 + ], + "angle": 0, + "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9499-9508 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.787, + 0.217 + ], + "angle": 0, + "content": "44. Yang, S., Li, G., Yu, Y.: Dynamic graph attention for referring expression comprehension. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4644-4653 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.787, + 0.259 + ], + "angle": 0, + "content": "45. Yang, S., Li, G., Yu, Y.: Graph-structured referring expression reasoning in the wild. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9952–9961 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.26, + 0.787, + 0.313 + ], + "angle": 0, + "content": "46. Yang, Z., Wang, J., Tang, Y., Chen, K., Zhao, H., Torr, P.H.: Lavo: Language-aware vision transformer for referring image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18155-18165 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.315, + 0.787, + 0.355 + ], + "angle": 0, + "content": "47. Yang, Z., Chen, T., Wang, L., Luo, J.: Improving one-stage visual grounding by recursive sub-query construction. In: European Conference on Computer Vision. pp. 387-404. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.357, + 0.787, + 0.397 + ], + "angle": 0, + "content": "48. Yang, Z., Gong, B., Wang, L., Huang, W., Yu, D., Luo, J.: A fast and accurate one-stage approach to visual grounding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4683-4693 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.398, + 0.787, + 0.453 + ], + "angle": 0, + "content": "49. Ye, J., Tian, J., Yan, M., Yang, X., Wang, X., Zhang, J., He, L., Lin, X.: Shifting more attention to visual backbone: Query-modulated refinement networks for end-to-end visual grounding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15502-15512 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.453, + 0.787, + 0.493 + ], + "angle": 0, + "content": "50. You, Q., Jin, H., Wang, Z., Fang, C., Luo, J.: Image captioning with semantic attention. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4651-4659 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.494, + 0.787, + 0.549 + ], + "angle": 0, + "content": "51. Yu, L., Lin, Z., Shen, X., Yang, J., Lu, X., Bansal, M., Berg, T.L.: Matttnet: Modular attention network for referring expression comprehension. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1307-1315 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.55, + 0.787, + 0.605 + ], + "angle": 0, + "content": "52. Yu, L., Poirson, P., Yang, S., Berg, A.C., Berg, T.L.: Modeling context in referring expressions. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14. pp. 69-85. Springer (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.606, + 0.787, + 0.646 + ], + "angle": 0, + "content": "53. Yu, Z., Yu, J., Xiang, C., Zhao, Z., Tian, Q., Tao, D.: Rethinking diversified and discriminative proposal generation for visual grounding. arXiv preprint arXiv:1805.03508 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.647, + 0.787, + 0.688 + ], + "angle": 0, + "content": "54. Zareian, A., Rosa, K.D., Hu, D.H., Chang, S.F.: Open-vocabulary object detection using captions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14393-14402 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.689, + 0.787, + 0.729 + ], + "angle": 0, + "content": "55. Zhang, H., Niu, Y., Chang, S.F.: Grounding referring expressions in images by variational context. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4158-4166 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.73, + 0.787, + 0.77 + ], + "angle": 0, + "content": "56. Zhang, Z., Zhu, Y., Liu, J., Liang, X., Ke, W.: Coupling word-pixel with sentence-mask alignments for referring image segmentation. arXiv preprint arXiv:2212.01769 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.771, + 0.787, + 0.827 + ], + "angle": 0, + "content": "57. Zhuang, B., Wu, Q., Shen, C., Reid, I., Van Den Hengel, A.: Parallel attention: A unified framework for visual object discovery through dialogs and queries. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4252-4261 (2018)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.827 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/SegVG_ Transferring Object Bounding Box to Segmentation for Visual Grounding/4d7de385-7f41-401e-b1cf-a9c2348c4d1c_origin.pdf b/2024/SegVG_ Transferring Object Bounding Box to Segmentation for Visual Grounding/4d7de385-7f41-401e-b1cf-a9c2348c4d1c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..572b362a8464229332b60efd2b2e69dffe0c2f8b --- /dev/null +++ b/2024/SegVG_ Transferring Object Bounding Box to Segmentation for Visual Grounding/4d7de385-7f41-401e-b1cf-a9c2348c4d1c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ae2c5cf12c0d035ebbf7fe9b9aa8bab56485a46b49290c17e12860b197cdd13c +size 3904983 diff --git a/2024/SegVG_ Transferring Object Bounding Box to Segmentation for Visual Grounding/full.md b/2024/SegVG_ Transferring Object Bounding Box to Segmentation for Visual Grounding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..241fe495ee9813457dc5889e4d94ca85426c7416 --- /dev/null +++ b/2024/SegVG_ Transferring Object Bounding Box to Segmentation for Visual Grounding/full.md @@ -0,0 +1,299 @@ +# SegVG: Transferring Object Bounding Box to Segmentation for Visual Grounding + +Weitai Kang $^{1}$ , Gaowen Liu $^{2}$ , Mubarak Shah $^{3}$ , and Yan Yan $^{1}$ + +1 Illinois Institute of Technology + +2 Cisco Research + +3 University of Central Florida + +Abstract. Different from Object Detection, Visual Grounding deals with detecting a bounding box for each text-image pair. This one box for each text-image data provides sparse supervision signals. Although previous works achieve impressive results, their passive utilization of annotation, i.e. the sole use of the box annotation as regression ground truth, results in a suboptimal performance. In this paper, we present SegVG, a novel method transfers the box-level annotation as Segmentation signals to provide an additional pixel-level supervision for Visual Grounding. Specifically, we propose the Multi-layer Multi-task Encoder-Decoder as the target grounding stage, where we learn a regression query and multiple segmentation queries to ground the target by regression and segmentation of the box in each decoding layer, respectively. This approach allows us to iteratively exploit the annotation as signals for both box-level regression and pixel-level segmentation. Moreover, as the backbones are typically initialized by pretrained parameters learned from unimodal tasks and the queries for both regression and segmentation are static learnable embeddings, a domain discrepancy remains among these three types of features, which impairs subsequent target grounding. To mitigate this discrepancy, we introduce the Triple Alignment module, where the query, text, and vision tokens are triangurally updated to share the same space by triple attention mechanism. Extensive experiments on five widely used datasets validate our state-of-the-art (SOTA) performance. Code is available at https://github.com/WeitaiKang/SegVG. + +Keywords: Visual Grounding $\cdot$ Detection $\cdot$ Transformer + +# 1 Introduction + +Visual grounding [14,17,26,30,52] aims to localize a target object within an image based on a free-form natural language text expression. It is particularly important for numerous downstream multimodal reasoning systems, such as visual question answering [10,35,41] and image captioning [1,4,50]. Previous works can be broadly categorized into three distinct groups: two-stage methods [3,44,45,51], one-stage methods [47,48], and transformer-based ones [6,15,16,31,43,49]. Both + +![](images/411ba3ae96ffde238132b8d517bf288127d3eb0f330fc4fe4597e893c10d5b6d.jpg) +Fig. 1: The comparison of visual grounding frameworks. The block with a dashed border indicates that the module may not necessarily exist. (a) Previous baseline method consists of two backbones and additional transformer layers for target grounding, where a regression query is supervised to regress the box. Current SOTA methods further employ a text-to-visual module to align the visual features with text features. (b) Our method incorporates segmentation queries, which utilizes the box annotation at the pixel-level to segment the target. Additionally, we propose the Triple Alignment module to eliminate the domain discrepancy of the query, text, and vision features. + +two-stage and one-stage approaches use convolutional neural networks for candidate proposals and the selection of the best-matching candidate. Nonetheless, these approaches rely on intricate modules that employ manually-crafted techniques for performing language inference and multi-modal integration. + +Inspired by the success of the transformer [7, 8], TransVG [6] proposes a transformer-based pipeline. As shown in Fig. 1.(a), this pipeline extracts vision and text features via DETR [2] and BERT [7], respectively. To ground the target, they use the transformer encoder to fuse multimodal features along with a learnable regression query and decode the query through an MLP. To enhance the final target grounding stage, subsequent studies continue with some text-to-visual modules in the early stage to modulate the vision features to align with the text features. For example, QRNet [49] proposes a query-modulated method for extracting language-aware vision features within the vision backbone. VLTVG [43] introduces a verification map to activate the vision features to align with the text features before multimodal fusion. + +Despite their advancements, the suboptimal annotation utilization, i.e., only using the box annotation as a regression annotation, limits their performance. As discussed in [37], Visual Grounding presents unique challenges compared to Object Detection due to its sparse supervision signals. Specifically, it provides only one box label for each text-image pair, while necessitating detection within a multimodal setting. Therefore, it is essential to fully exploit the box annotation, by treating it as a segmentation mask (pixels within the bounding box are assigned a value of 1, while pixels outside the bounding box are assigned as 0). + +In this paper, we introduce SegVG (see Fig. 1.(b)), a novel method that leverages the pixel-level details within the box annotation as segmentation signals to offer additional fine-grained supervision for Visual Grounding. Specifically, we propose the Multi-layer Multi-task Encoder-Decoder as the target grounding stage, where we learn a regression query and multiple segmentation queries to ground the target by regression and segmentation of the box in each decoding layer, respectively. The confidence score derived from the segmentation can further serve as a Focal Loss [21] scaling factor to adaptively emphasize the other losses of challenging training samples. This approach allows us to iteratively exploit the annotation as signals for both box-level regression and pixel-level segmentation. Furthermore, the initial parameters for model backbones, typically derived from pretrained unimodal tasks, along with data-agnostic static embeddings used as queries for decoding, result in a domain discrepancy among different sources of feature, affecting the effectiveness of target grounding. To tackle this problem, we present the Triple Alignment module, where we harmonize the domain of query, text, and vision features by implementing a triangular update process through a triple attention mechanism. As a result, we ensure that all features adapt and integrate within the same multimodal space, thereby enhancing subsequent target grounding. Our contributions are as follows: + +- We propose the Multi-layer Multi-task Encoder-Decoder to maximize the utilization of the box annotation, which introduces an additional segmentation format for pixel-level supervision in Visual Grounding. +- To eliminate the domain discrepancy among the query, text, and vision, we introduce the Triple Alignment to update these three types of features into a sharing domain, which facilitates the subsequent target grounding. +- We conduct extensive experiments on five widely used datasets to show the performance superiority of our proposed methods compared with previous state-of-the-art methods and further investigate the reliability benefits derived from the segmentation output in real applications. +- We will release code and checkpoints for future research development. + +# 2 Related Work + +Visual grounding methods can be roughly classified into three pipelines: two-stage methods, one-stage methods, and transformer-based methods. + +Two-stage methods Two-stage approaches [3,51] treat visual grounding as first generating candidate object proposals and then finding the best match to the text. In the first stage, an off-the-shelf detector processes the image and proposes regions that may contain the target. In the second stage, a ranking network calculates the similarity between candidate regions and processed text features, selecting the region with the highest similarity score as the final result. Training losses include binary classification loss [29] or maximum-margin ranking loss [51]. To better understand the text and cross-modality matching, MattNet [51] focuses + +![](images/73df9bb07d93244cf5086901d4f2c7a651af179c31859a2f5a5ebe3fb84b9f16.jpg) +Fig. 2: SegVG: The upper figure includes the vision and text backbone. Our proposed Triple Alignment module is iteratively inserted into intermediate layers to eliminate domain discrepancy. The lower figure shows our Multi-layer Multi-task Encoder-Decoder, which adopts a transformer encoder-decoder to update multimodal features and ground the target. In this architecture, we make the best of the box annotation as a segmentation ground truth and integrate an additional segmentation task into Visual Grounding. Additionally, the segmentation output serves as a Focal Loss factor, allowing adaptive emphasis on challenging cases for the regression loss. $\mathrm{M} = 6$ , $\mathrm{R} = 6$ . + +on decomposing the text into subject, location, and relationship components. [3] introduces an expression-aware score for improved candidate region ranking. + +One-stage methods One-stage approaches [47, 48] directly concatenate vision and text features in the channel dimension and rank confidence values for candidate regions proposed based on the concatenated multimodal features. For example, FAOA [48] predicts bounding boxes using a YOLOv3 detector [32] on the concatenated features. ReSC [47] further improves the ability to ground complex queries by introducing a recursive sub-query construction module. + +Transformer-based methods Transformer-based approach is first introduced by TransVG [6]. Unlike previous methods, TransVG concatenates the regression query (a learnable embedding), vision tokens, and text tokens and uses transformer encoders [38] to perform cross-modal fusion and target grounding. The query is then processed through an MLP to decode the box. Benefiting from the flexible structure of transformer modules in processing multimodal features, recent works continue adopting this pipeline and propose novelties regarding feature extraction. VLTVG [43] develops a visual-linguistic verification module before the target grounding stage to modulate the vision features with the relationship between vision and text features. QRNet [49] proposes a Query-modulated Refinement Network to mitigate the gap between features from the unimodal vision backbone and those needed for multi-modal reasoning. + +Multi-task Visual Grounding Multi-task learning is extensively utilized in object detection and segmentation [2,11], often capitalizing on a shared backbone and task-specific heads. Expanding upon this idea, several studies [19, 25, 36] have proposed solutions to the Multi-task Visual Grounding problem. In this problem, they jointly tackle Referring Expression Comprehension (REC, also known as Visual Grounding) and Referring Expression Segmentation (RES), requiring both box annotations and segmentation annotations. It is important to note that, unlike those approaches, even though we incorporate segmentation losses in our method, we do not need segmentation annotations but only box annotations, focusing specifically on the Visual Grounding task. + +# 3 Methodology + +In this section, we present the components of our SegVG in the order of the data flow: starting with the backbones, followed by our proposed Triple Alignment, and finally our Multi-layer Multi-task Encoder-Decoder. + +# 3.1 Backbones + +As shown in Fig. 2 (upper), our vision backbone consists of ResNet and transformer encoder from DETR [2], with parameters pretrained on Object Detection task using the MSCOCO dataset [22], excluding the validation and test sets of Visual Grounding dataset. The text backbone is the base model of BERT [7]. + +Vision Backbone Given an input image $\mathbf{I_0}$ $(\mathbb{R}^{3\times H_0\times W_0})$ , we employ ResNet to generate a 2D feature map $\mathbf{I}\in \mathbb{R}^{C\times H\times W}$ ( $C = 2048$ , $H = \frac{H_0}{32}$ , $W = \frac{W_0}{32}$ ). A 1x1 convolutional layer is then used to reduce the channel dimension of $\mathbf{I}$ to $C_v = 256$ , resulting in $\mathbf{I}'$ . We further flatten $\mathbf{I}'$ into $\mathbf{Z}_v\in \mathbb{R}^{C_v\times N_v}$ ( $N_v = H\times W$ ). Position embedding is then added to $\mathbf{Z}_v$ to preserve sensitivity to the original 2D spatial locations. $\mathbf{Z}_v$ is then iteratively processed through DETR's encoder layer (total 6 transformer layers) and the Triple Alignment to obtain the output $\mathbf{Z}_v$ . + +Text Backbone Given a text, we initially utilize the BERT's embedding layer to convert it into $N_{t}$ language tokens with $C_t$ channel dimension. In alignment with [6], we preprocess a [CLS] token and append a [SEP] token to the beginning and end positions of the tokenized language, respectively. Following this, we iteratively input the language tokens into BERT's layers (total 12 transformer layers) and the Triple Alignment, generating language embedding $\mathbf{Z}_t$ . + +# 3.2 Triple Alignment + +Given that the text and vision backbones are pretrained from unimodal tasks and the queries are data-agnostic, the subsequent target grounding stage faces the challenge of aligning these three types of features into the same space before performing multimodal fusion for target grounding. Additionally, considering + +that the backbones usually contribute the majority of the overall parameters, solely using them for extracting unimodal features without incorporating multimodal alignment is sub-optimal. Therefore, an optimal solution is to address the domain discrepancy before moving on to the subsequent target grounding stage. + +Triple Alignment (Fig. 2 (upper)) utilizes an attention mechanism to perform triangular feature sampling, aiming to ensure domain consistency among the query, text, and vision features. The queries, $\mathbf{Z}_o$ , are first initialized by N learnable embeddings, where one embedding is for the regression query and the rest of the embeddings are for multiple segmentation queries. The data flow is: + +$$ +\mathbf {Z} _ {v} ^ {i + 1} = \operatorname {D E T R L a y e r} _ {i} \left(\mathbf {Z} _ {v} ^ {i}\right), i \in \{0, 1,.., L - 1 \} \tag {1} +$$ + +$$ +\mathbf {Z} _ {t} ^ {i + 1} = \operatorname {B E R T L a y e r} _ {\mathcal {2 i + 1}} (\operatorname {B E R T L a y e r} _ {\mathcal {2 i}} (\mathbf {Z} _ {t} ^ {i})), \tag {2} +$$ + +$$ +\left[ \mathbf {Z} _ {o} ^ {\prime}, \mathbf {Z} _ {t} ^ {\prime}, \mathbf {Z} _ {v} ^ {\prime} \right] = \operatorname {T r i - M H A} \left(\mathbf {Z} _ {o}, \mathbf {Z} _ {t} ^ {i + 1}, \mathbf {Z} _ {v} ^ {i + 1}\right), \tag {3} +$$ + +$$ +\mathbf {Z} _ {o} = \mathbf {Z} _ {o} + \mathbf {Z} _ {o} ^ {\prime}, \mathbf {Z} _ {t} ^ {i + 1} = \mathbf {Z} _ {t} ^ {i + 1} + \mathbf {Z} _ {t} ^ {\prime}, \mathbf {Z} _ {v} ^ {i + 1} = \mathbf {Z} _ {v} ^ {i + 1} + \mathbf {Z} _ {v} ^ {\prime}, \tag {4} +$$ + +where L is the number of layers, BERTLayer is the layer of BERT and DETRLayer is the layer of DETR's encoder. The vision and text features are first encoded by Eq. 1 and Eq. 2. Subsequently, the three types of tokens (query, text, and vision) are updated by our Triple Multi-Head Attention Layer (Tri-MHA) using Eq. 3. The output tokens are merged back to their original branches respectively by Eq. 4. Within each head of the Triple Multi-Head Attention Layer (Tri-MHA), each type of the features simultaneously computes its updated representation by attending to both the others and itself: + +$$ +S = \left[ \mathbf {Z} _ {o} W ^ {(o, S)}, \mathbf {Z} _ {t} W ^ {(t, S)}, \mathbf {Z} _ {v} W ^ {(v, S)} \right], S \in \{Q, K, V \} +$$ + +$$ +[ \mathbf {Z} _ {o}, \mathbf {Z} _ {t}, \mathbf {Z} _ {v} ] = \operatorname {S o f t M a x} \left(Q K ^ {T} / \sqrt {d _ {k}}\right) V, \tag {5} +$$ + +$$ +\mathbf {Z} _ {e} ^ {\prime} = \mathbf {Z} _ {e} W ^ {e}, e \in \{o, t, v \}, +$$ + +where $\{W^{(e,S)}, W^e : e \in \{o, t, v\}, S \in \{Q, K, V\}\}$ are trainable parameters. As a result, each of the output features is triangular sampling from all of the three types of features, which alleviates the domain discrepancy. + +# 3.3 Multi-layer Multi-task Encoder-Decoder + +The Multi-layer Multi-task Encoder-Decoder serves as the target grounding stage, where we use a transformer encoder-decoder for cross-modal fusion and target grounding to perform a box regression task and a box segmentation task. + +Encoder As shown in Fig. 2 (lower left), given the aligned output text and vision features from the backbones, the encoder fuses the two modalities into the multimodal features by a stack of transformer layers. In each layer, the concatenated text and vision tokens go through the Multi-Head Self-Attention layer (MHSA) and the Feed Forward Network (FFN) with the residual connection. + +Decoder In each decoder layer, we aim to fully exploit the box annotation. We propose the bbox2seg paradigm to transform the box annotation into a segmentation mask, which classifies all pixels within the box as foreground (with a value of one) and those outside the box as background (with a value of zero). As shown in Fig. 2 (lower right), one regression query aims to regress the box, while the remaining segmentation queries aim to segment the box. Different segmentation queries are endowed with different learnable positional embeddings to enhance the robustness of each decoder layer, since the decoder layer, when confronted with various queries, is required to segment the same box. Following that, the queries pass through the Multi-Head Self-Attention layer to exchange information about the same target, prompting each other to better locate the target. Subsequently, the queries undergo the Multi-Head Cross-Attention layer and the Feed Forward Network, where multimodal features serve as the Key and Value to ground the target. Finally, a shared MLP across all decoder layers decodes the regression query into the box result, supervised by L1 loss and Giou loss [33]. Each segmentation query is repeated $N_{v}$ times and concatenated with visual tokens along the channel dimension. Another shared MLP decodes the concatenated feature into the segmentation mask, supervised by Focal loss [21] and Dice loss [27]. It is noteworthy that our segmentation paradigm shares the same semantic foundation as the regression paradigm, i.e., to distinguish bounding box, rather than instance segmentation. Therefore, incorporating non-object pixels in the segmented foreground does not introduce ambiguity to the model. We provide qualitative results 4.8 to demonstrate this feature. To alleviate multi-task optimization challenges, we freeze the backbones for the initial k epochs to stabilize the training process. + +Confidence score Since both the regression output and segmentation output share the same aim, we can additionally obtain the confidence score for the foreground by averaging values inside the ground truth box of the segmentation output to reflect the confidence of the regression output. In the training process, we can transform this confidence score as the Focal loss factor [21] to adaptively emphasize the other losses of challenging training samples. The final loss function of each decoder layer is formulated as follows: + +$$ +L = \lambda_ {1} c _ {f o c a l} L _ {1} + \lambda_ {g i o u} c _ {f o c a l} L _ {g i o u} + \tag {6} +$$ + +$$ +\lambda_ {d i c e} c _ {f o c a l} L _ {d i c e} + \lambda_ {f o c a l} L _ {f o c a l}, +$$ + +where $\lambda_{1}$ , $\lambda_{giou}$ , $\lambda_{focal}$ and $\lambda_{dice}$ are hyperparameters. $L_{1}$ is the L1 loss. $L_{giou}$ is the GIoU loss [33]. $L_{focal}$ is the Focal loss [21]. $L_{dice}$ is the Dice loss [27]. $c_{focal}$ is the above Focal loss factor averaged across all segmentation outputs. + +In the real-world application perspective, the visual grounding task can be viewed as open-vocabulary object detection [54], where target objects lack predetermined categories. Therefore, previous transformer-based methods directly regress the box without confidence scores, since there is no candidate proposal or selection stage in transformer-based pipeline. However, confidence scores are valuable for enhancing the control or reliability of predictions by filtering out low-confidence predictions. This feature could benefit the future integration of visual + +Table 1: Comparisons with state-of-the-art methods on widely used datasets. We highlight the best and second best performance in red and blue, and bold our model. + +
ModelsBackboneRefCOCORefCOCO+RefCOCOgReferItGame test
valtestAtestBvaltestAtestBval-gval-utest-u
Two-stage:
CMN [12]VGG16-71.0365.77-54.3247.7657.47--28.33
VC [55]VGG16-73.3367.44-58.4053.1862.30--31.13
ParalAttn [57]VGG16-75.3165.52-61.3450.8658.03---
MAttNet [51]ResNet-10176.6581.1469.9965.3371.6256.02-66.5867.2729.04
Similarity Net [39]ResNet-101---------34.54
CITE [29]ResNet-101---------35.07
DDPN [53]ResNet-101---------63.00
LGRANs [40]VGG16-76.6066.40-64.0053.4061.78---
DGA [44]VGG16-78.4265.53-69.0751.99--63.28-
RvG-Tree [29]ResNet-10175.0678.6169.8563.5167.4556.66-66.9566.51-
NMTree [23]ResNet-10176.4181.2170.0966.4672.0257.5264.6265.8766.44-
Ref-NMS [3]ResNet-10180.7084.0076.0468.2573.6859.42-70.5570.62-
One-stage:
SSG [5]DarkNet-53-76.5167.50-62.1449.2747.4758.80-54.24
ZSGNet [34]ResNet-50---------58.63
FAOA [48]DarkNet-5372.5474.3568.5056.8160.2349.6056.1261.3360.3660.67
RCCF [20]DLA-34-81.0671.85-70.3556.32--65.7363.79
ReSC-Large [47]DarkNet-5377.6380.4572.3063.5968.3656.8163.1267.3067.2064.60
LBYL-Net [13]DarkNet-5379.6782.9174.1568.6473.3859.4962.70--67.47
Transformer-based:
TransVG [6]ResNet-10181.0282.7278.3564.8270.7056.9467.0268.6767.7370.73
QRNet [49]Swin-S84.0185.8582.3472.9476.1763.8171.8973.0372.5274.61
VLTVG [43]ResNet-10184.7787.2480.4974.1978.9365.1772.9876.0474.1871.98
SegVG (ours)ResNet-10186.8489.4683.0777.1882.6367.5976.0178.3577.4275.59
+ +grounding models into downstream multimodal reasoning systems or real-world applications. To meet the requirements of this feature, our approach incorporates a confidence score derived from the segmentation output during inference. Specifically, we calculate the model's confidence by averaging values greater than or equal to 0.35 (adopted from [42]) in the segmentation output of one segmentation query. Analyses 4.7 in the Experiment section demonstrate the faithfulness and benefits of incorporating this additional confidence score. + +# 4 Experiments + +# 4.1 Metric and Datasets + +Metric A predicted bounding box is considered accurate if its Intersection over Union (IoU) with the ground-truth bounding box exceeds 0.5. In accordance with the established practices in preceding studies [6,43], we employ top-1 accuracy (measured in percentage) as the primary metric to assess our method. + +Datasets There are five standard benchmarks: RefCOCO [52], RefCOCO+ [52], RefCOCOg-g [26], RefCOCOg-umd [26], and ReferItGame [17]. Four of them (RefCOCO, RefCOCO+, and RefCOCOg-(g/umd)) are all derived from MSCOCO [22]. RefCOCO consists of 19,994 images and 142,210 referring texts, which is divided into four subsets: a training set with 120,624 texts, a validation set with 10,834 texts, and two test sets (testA and testB) containing 5,657 + +and 5,095 texts, respectively. RefCOCO+ includes 19,992 images and 141,564 referring texts, which is partitioned into four subsets: a training set with 120,191 texts, a validation set with 10,758 texts, and two test sets (testA and testB) containing 5,726 and 4,889 texts, respectively. RefCOCOg contains 25,799 images and 95,010 longer texts. Two widely accepted splitting methods—RefCOCOg-g [26] and RefCOCOg-umd [28]—are employed for this dataset, and we perform experiments using both RefCOCOg-g (val-g) and RefCOCOg-umd (val-u and test-u) splitting conventions. The ReferItGame dataset, featuring 20,000 images from SAIAPR-12 [9], is divided into three segments: a training set with 54,127 texts, a validation set with 5,842 texts, and a testing set comprising 60,103 texts. + +# 4.2 Implementation + +Our approach uses an input image size of $640 \times 640$ and sets the maximum expression length at 40. When resizing images, we preserve the original aspect ratio. The longer edge is resized to 640, and the shorter edge is padded to 640 with the value of zero. Texts exceeding 38 tokens are truncated, reserving a start position and an end position of the characters for the [CLS] and [SEP] token, respectively. If the text is shorter, empty tokens are added after the [SEP] token to reach an input length of 40. Paddings for the input image are not tracked by masks, while empty tokens of text employ masks. + +We use the AdamW optimizer. The initial learning rate of 1e-5 is assigned to backbones, and 1e-4 to the rest parameters. Weight decay is set at 1e-4. The visual backbone is initialized with the DETR's backbone and encoder, while the language branch uses the basic BERT model. For the final results, our model is trained for 90 epochs, with the learning rate decreasing by a factor of 10 after 60 epochs. The k hyperparameter in Multi-layer Multi-task Encoder-Decoder is set to 10. We use a batch size of 64. For the ablation studies presented in Table 3, the models are trained for 60 epochs with k equal to 20, and the learning rate drops after 40 epochs. We set $\lambda_{1} = 5$ , $\lambda_{giou} = 2$ , $\lambda_{focal} = 1$ and $\lambda_{dice} = 1$ . We adhere to previous practices [6, 43] for data augmentation during training. + +# 4.3 Quantitative Results + +We report the performance of our SegVG on all benchmark datasets. As presented in Table 1, our SegVG model demonstrates superiority across all of the datasets. This indicates the effectiveness and generalizability of our approach. It is worth noting that RefCOCO+ and RefCOCOg are relatively more challenging datasets, as RefCOCO+ does not include location terms in its language expressions, and RefCOCOg has longer language expressions compared to other datasets. Despite these challenges, our model exhibits significant improvements on these two difficult datasets. Specifically, on RefCOCO+, our model outperforms the previous SOTA models with $+2.99\%$ , $+3.7\%$ , and $+2.42\%$ on the val, testA, and testB subsets, respectively. On RefCOCOg, our model also surpasses the previous SOTA models with $+3.03\%$ , $+2.31\%$ , and $+3.24\%$ on the val-g, + +Table 2: Compare transformer-based models on Parameter count and GFLOPS. + +
BackboneParameter count (M)GFLOPS (G)
TransVG [6]ResNet101141.5572.61
VLTVG [43]ResNet101141.6169.87
QRNet [49]Swin-S247.0680.12
SegVGResNet101155.2873.48
+ +Table 3: Ablation study on RefCOCOg test-u. Encoder and Decoder are the encoder and decoder of Multi-layer Multi-task Encoder-Decoder, respectively. MMDecoder represents Multi-layer and Multi-task supervision in the Decoder. N is the number of segmentation queries. Triple means Triple Alignment. Excluding Query in Triple (g) means using bidirectional alignment with concatenated text-vision tokens. + +
IdModelAcc(%)
(a)Backbones + Encoder + Decoder66.97
(b)Backbones + Triple + Encoder + Decoder75.75
(c)Backbones + Encoder + MMDecoder (N=1)72.61
(d)Backbones + Encoder + MMDecoder (N=5)76.21
(e)Backbones + Triple + Encoder + MMDecoder (N=5)77.29
(f)(e) w/o Encoder76.65(-0.64)
(g)(e) w/o Query in Triple76.37(-0.92)
(h)(e) w/o Focal loss76.83(-0.46)
+ +val-u, and test-u subsets, respectively. These results suggest that under the reinforcement of Triple Alignment and Multi-layer Multi-task Encoder-Decoder, the query, text, and vision tokens are triangularly updated to share the same space, and the model fully exploits the bounding box as fine-grained pixel-level supervision for comprehensive end-to-end learning. + +We also conduct a comparison of the number of parameters and GFLOPS across transformer-based models to evaluate computational costs. As depicted in Table 2, the computational cost of SegVG falls within a reasonable range. + +# 4.4 Ablation Study + +In this section, we aim to validate the efficacy of each proposed module. We conduct ablation studies on the RefCOCOg-umd test dataset. Specifically, we start by evaluating a basic structure, i.e., the backbones with encoder-decoder structure. After that, we systematically incorporate the Triple Alignment module into the backbones and introduce the Multi-layer Multi-task supervision into the decoder through the controlled variable approach. Meanwhile, we conduct additional ablation experiments on specific details, including assessing the efficacy of incorporating the encoder, introducing Query in the Triple Alignment, and introducing the Focal loss from the segmentation output to the other losses. + +As shown in Table 3, we can draw the following conclusions when comparing the experimental results under controlled variables: 1) [(a) v.s. (b)]: Incorporating the Triple Alignment can effectively eliminate the domain discrepancy + +Table 4: Ablation study on RefCOCOg test-u regarding the number of segmentation queries. $\Delta$ Hour and $\Delta$ Acc are the additional time cost and accuracy improvement compared to ID(i) experiment, respectively. + +
IDnum_queryAccTime CostΔHour / ΔAcc (v.s. ID(i))
iN = 172.6116.32h-
iiN = 373.0218.97h6.46
iiiN = 576.2120.10h1.05
ivN = 774.3822.24h3.34
vN = 973.9124.07h5.96
+ +among the query, text, and vision features, thereby facilitating subsequent target grounding. 2) [(a) v.s. (c)]: Introducing Multi-layer Multi-task supervision can iteratively make the best of the annotations in the target grounding stage, thereby enhancing the learning of query representations. 3) [(c) v.s. (d)]: Increasing the number of segmentation queries can further improve the robustness of the decoder when provided with different queries and required to segment the same box. 4) [(a), (b), (d), and (e)]: Combining the Triple Alignment and Multi-layer Multi-task Encoder-Decoder can effectively enhance the overall performance, achieving the optimal result. 5) [(e) v.s. (f)]: Even if we include the Triple Alignment supporting multimodal communication, it remains necessary for the subsequent encoder to update the unimodal features generated by the backbones into multimodal features. 6) [(e) v.s. (g)]: It is necessary to involve queries in Triple Alignment to transfer the data-agnostic embedding into data-related queries. Otherwise, using only Bidirectional Alignment (BA) for text and vision tokens, similar to approaches like Deep Fusion in GLIP [18] and Grounding DINO [24], and WPA in CoupAlign [56], causes a noticeable decline $(-0.92\%)$ . 7) [(d), (e) and (g)]: The slight gain of (g)(76.37%) over (d)(76.21%) stems from our MMDecoder's pixel-level signals, which already boosts BA in the Encoder. Thus, extra BA effort in the backbone is marginal, failing to solve the unaligned query issue. Instead, Tri-Align (e)(77.29%) can solve this issue, showing its novelty. Notably, (d)(76.21%), with a basic encoder-decoder, already achieves SOTA performance, emphasizing our bbox2seg paradigm's simplicity and effectiveness. 8) [(e) v.s. (h)]: The segmentation output can further derive the confidence score of the model's prediction, which is transformed into the Focal loss factor to adaptively scale the other losses to focus more attention on challenging cases. + +We further conduct a more detailed abaltion study about the improvement and corresponding cost of adding more segmentation queries of Tab.3(c)-(d). As shown in Table 4, the best performance is observed with five segmentation queries. Adding more than that increases the burden from pixel-level constraint without benefits, also increasing computational cost per accuracy improvement. + +# 4.5 Triple Alignment Analysis + +In addition to the improvement results in ablation study, we further deepen our understanding of the Triple Alignment by analyzing the attention behavior. + +Table 5: Attention values from query to text-referred visual region in Triple Alignment. + +
LayerRefCOCORefCOCO+RefCOCOgReferItGame test
valtestAtestBvaltestAtestBval-gval-utest-utest
layer217%18%16%19%21%18%20%18%17%25%
layer434%36%33%29%31%28%27%33%33%30%
layer661%63%60%55%58%50%55%59%59%38%
+ +Table 6: Comparison with alternative method using RIS to provide pseudo segmentation labels for supervision. + +
ReferItGameSegVG+RIS (LAVT)SegVG
val74.9176.85
test73.7675.06
+ +Table 7: AP50 using different segmentation queries for confidence score calculation on RefCOCOg test-u. + +
Seg Query1st2nd3rd4th5th
AP5084.6584.5784.7384.7884.63
+ +Specifically, we calculate the sum of the attention values from the query to the text-referred visual region (target bbox) as a percentage of the total attention (which includes attention to the query, text, and visual tokens) to illustrate the extent of alignment across these three modalities in the second, fourth, and final layer of Triple Alignment. We average the percentage across all the attention heads and queries, and perform the analysis across all the datasets. As shown in Table 5, in all the datasets, the attention values increase as the layer progresses, indicating that Triple Alignment progressively aligns the query to comprehend the text and then focus on the referred visual region. + +# 4.6 Comparison with alternative method + +Given the development of Referring Expression Segmentation (RES), a natural alternative method would be using a RES method to generate pseudo segmentation labels to substitute our bbox2seg paradigm. Therefore, to mimic real-world scenarios, we use LAVT [46] trained on RefCOCO to obtain pseudo segmentation labels on ReferItGame. We follow the same training setting in ablation study to conduct the comparison on ReferItGame. As shown in Table 6, our SegVG outperforms the alternative method. This demonstrates that our bbox2seg paradigm is more effective than using a RES model to provide pseudo segmentation labels which might suffer from the errors from the RES model. + +# 4.7 Confidence Score Analysis + +Selection of segmentation query To show the effect of different selection of segmentation query for the calculation of confidence score, we calculate AP50 on RefCOCOg test-u. As shown in Table 7, the performance variations are slight among them. We use the first one to calculate confidence score for simplicity. + +Confidence Score Faithfulness To evaluate the faithfulness of our confidence score, i.e., whether a higher confidence score indicates better performance, we + +![](images/60e14df3e9978e3090ace5b2d575b2626131b34f201cadfef4ea883ab81a9dc4.jpg) +Fig. 3: IoU and Accuracy of different confidence scores on RefCOCOg test-u. + +Table 8: Performance of different confidence levels on RefCOCOg test-u. + +
ConfidenceIoUAcc(%)Proportion (%)
≥ 0.650.706777.41%100.00%
≥ 0.700.707277.48%99.89%
≥ 0.750.709177.69%99.44%
≥ 0.800.714378.37%97.98%
≥ 0.850.722679.29%95.04%
+ +![](images/c91bfc47273881b5edc3e6de06845807920d290f47f514cf1edd7a5cb420201f.jpg) +Fig. 4: Qualitative comparison between (c) (the first line) and (d) (the second line) of Table. 3. Red boxes are ground truth. Blue boxes are model predictions. + +assess the relationship of our confidence score and model performance metrics (IoU and Accuracy) as shown in Fig 3. We sort the RefCOCOg-umd test set by confidence score, split it into five equal parts, and calculate each part's average score and performance. We observe a positive correlation between the performance metrics and our confidence score, which confirms its faithfulness. + +Confidence Score Application In real applications, the confidence score can be used to enhance the model's reliability. Specifically, we can apply different confidence thresholds to achieve different predictions, as shown in Table 8. First, we observe that accuracy increases with higher thresholds, indicating that adjusting the threshold can enhance the model's localization ability. Furthermore, the mean IoU also increases with the increasing threshold. Therefore, in downstream applications, such as using the model to provide pseudo-labels, we can increase the threshold to obtain a more accurate box. Since low-confidence outputs are excluded, the output proportion is slightly reduced, i.e., yields fewer outputs. + +![](images/7adbdf7fdbc2d748611024734ad73347d1709ec5c78a791704222e5fa6d12ba1.jpg) +Fig. 5: Layer $i$ refers to the output of the $i$ -th layer of the decoder. The blue boxes represent the models' predictions, while the red boxes denote the ground truth. In the segmentation mask shown in the second column from the right, red indicates high confidence for the foreground. Note that VLTVG does not provide segmentation output. + +# 4.8 Qualitative Results + +We compare Tab. 3 (c) and (d) qualitatively to show the robustness enhancement. As shown in Fig. 4, adding segmentation queries improves the robustness to distinguish the target from the distractor, e.g. the case of two dogs. + +As depicted in Fig. 5, we compare the box prediction quality by each decoding layer of SegVG with VLTVG [43] which also involves multi-layer supervision. As seen in the upper two rows of Fig. 5, VLTVG initially misses the target "young man" but improves its prediction gradually and finally makes the correct prediction. In contrast, due to the full exploitation of the annotations and the domain alignment in our Triple Alignment, SegVG successfully identifies the location of the target in the early decoding layer and consistently makes the correct prediction in each layer. Another example can be observed in the lower two rows of Fig. 5. In this image, due to the complex colors, VLTVG fails to locate the target "plate" and consistently repeats the same mistake. Instead, SegVG correctly detects the target, even in the first decoder layer. Additionally, we visualize the segmentation mask obtained by SegVG in Fig. 5, which accurately identifies the target box with high confidence. This behavior aligns with the box regression, demonstrating their shared objective, i.e., to distinguish the box. + +# 5 Conclusion + +We propose SegVG, a visual grounding model, where we iteratively make the best of box annotations to involve pixel-level supervision and address the domain discrepancy among queries, text, and vision by Triple Alignment module. Experiments show the superior performance of SegVG. Furthermore, we explore the reliability benefits of our segmentation output in real-world applications. + +# Acknowledgements + +This research is supported by NSF IIS-2309073 and ECCS-2123521. This article solely reflects the opinions and conclusions of its authors and not the funding agencies. + +# References + +1. Anderson, P., He, X., Buehler, C., Teney, D., Johnson, M., Gould, S., Zhang, L.: Bottom-up and top-down attention for image captioning and visual question answering. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6077-6086 (2018) +2. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020) +3. Chen, L., Ma, W., Xiao, J., Zhang, H., Chang, S.F.: Ref-nms: Breaking proposal bottlenecks in two-stage referring expression grounding. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 1036-1044 (2021) +4. Chen, S., Jin, Q., Wang, P., Wu, Q.: Say as you wish: Fine-grained control of image caption generation with abstract scene graphs. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9962-9971 (2020) +5. Chen, X., Ma, L., Chen, J., Jie, Z., Liu, W., Luo, J.: Real-time referring expression comprehension by single-stage grounding network. arXiv preprint arXiv:1812.03426 (2018) +6. Deng, J., Yang, Z., Chen, T., Zhou, W., Li, H.: Transvg: End-to-end visual grounding with transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1769-1779 (2021) +7. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018) +8. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) +9. Escalante, H.J., Hernández, C.A., Gonzalez, J.A., López-López, A., Montes, M., Morales, E.F., Sucar, L.E., Villasenor, L., Grubinger, M.: The segmented and annotated iapr tc-12 benchmark. Computer vision and image understanding 114(4), 419-428 (2010) +10. Gan, C., Li, Y., Li, H., Sun, C., Gong, B.: Vqs: Linking segmentations to questions and answers for supervised attention in vqa and question-focused semantic segmentation. In: Proceedings of the IEEE international conference on computer vision. pp. 1811-1820 (2017) +11. He, K., Gkioxari, G., Dollár, P., Girshick, R.: Mask r-cnn. In: Proceedings of the IEEE international conference on computer vision. pp. 2961-2969 (2017) +12. Hu, R., Rohrbach, M., Andreas, J., Darrell, T., Saenko, K.: Modeling relationships in referential expressions with compositional modular networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1115-1124 (2017) + +13. Huang, B., Lian, D., Luo, W., Gao, S.: Look before you leap: Learning landmark features for one-stage visual grounding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16888-16897 (2021) +14. Kang, W., Qu, M., Kini, J., Wei, Y., Shah, M., Yan, Y.: Intent3d: 3d object detection in rgb-d scans based on human intention. arXiv preprint arXiv:2405.18295 (2024) +15. Kang, W., Qu, M., Wei, Y., Yan, Y.: Actress: Active retraining for semi-supervised visual grounding. arXiv preprint arXiv:2407.03251 (2024) +16. Kang, W., Zhou, L., Wu, J., Sun, C., Yan, Y.: Visual grounding with attention-driven constraint balancing. arXiv preprint arXiv:2407.03243 (2024) +17. Kazemzadeh, S., Ordonez, V., Matten, M., Berg, T.: Referitgame: Referring to objects in photographs of natural scenes. In: Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP). pp. 787-798 (2014) +18. Li, L.H., Zhang, P., Zhang, H., Yang, J., Li, C., Zhong, Y., Wang, L., Yuan, L., Zhang, L., Hwang, J.N., et al.: Grounded language-image pre-training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10965-10975 (2022) +19. Li, M., Sigal, L.: Referring transformer: A one-step approach to multi-task visual grounding. Advances in neural information processing systems 34, 19652-19664 (2021) +20. Liao, Y., Liu, S., Li, G., Wang, F., Chen, Y., Qian, C., Li, B.: A real-time cross-modality correlation filtering method for referring expression comprehension. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10880-10889 (2020) +21. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proceedings of the IEEE international conference on computer vision. pp. 2980-2988 (2017) +22. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014) +23. Liu, D., Zhang, H., Wu, F., Zha, Z.J.: Learning to assemble neural module tree networks for visual grounding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4673-4682 (2019) +24. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023) +25. Luo, G., Zhou, Y., Sun, X., Cao, L., Wu, C., Deng, C., Ji, R.: Multi-task collaborative network for joint referring expression comprehension and segmentation. In: Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition. pp. 10034-10043 (2020) +26. Mao, J., Huang, J., Toshev, A., Camburu, O., Yuille, A.L., Murphy, K.: Generation and comprehension of unambiguous object descriptions. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 11-20 (2016) +27. Milletari, F., Navab, N., Ahmadi, S.A.: V-net: Fully convolutional neural networks for volumetric medical image segmentation. In: 2016 fourth international conference on 3D vision (3DV). pp. 565-571. IEEE (2016) +28. Nagaraja, V.K., Morariu, V.I., Davis, L.S.: Modeling context between objects for referring expression understanding. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14. pp. 792-807. Springer (2016) + +29. Plummer, B.A., Kordas, P., Kiapour, M.H., Zheng, S., Piramuthu, R., Lazebnik, S.: Conditional image-text embedding networks. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 249-264 (2018) +30. Plummer, B.A., Wang, L., Cervantes, C.M., Caicedo, J.C., Hockenmaier, J., Lazebnik, S.: Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In: Proceedings of the IEEE international conference on computer vision. pp. 2641-2649 (2015) +31. Qu, M., Wu, Y., Liu, W., Gong, Q., Liang, X., Russakovsky, O., Zhao, Y., Wei, Y.: Siri: A simple selective retraining mechanism for transformer-based visual grounding. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXV. pp. 546-562. Springer (2022) +32. Redmon, J., Farhadi, A.: Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767 (2018) +33. Rezatofighi, H., Tsoi, N., Gwak, J., Sadeghian, A., Reid, I., Savarese, S.: Generalized intersection over union: A metric and a loss for bounding box regression. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 658-666 (2019) +34. Sadhu, A., Chen, K., Nevatia, R.: Zero-shot grounding of objects from natural language queries. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4694-4703 (2019) +35. Shang, Y., Cai, M., Xu, B., Lee, Y.J., Yan, Y.: Llava-prumerge: Adaptive token reduction for efficient large multimodal models. arXiv preprint arXiv:2403.15388 (2024) +36. Su, W., Miao, P., Dou, H., Wang, G., Qiao, L., Li, Z., Li, X.: Language adaptive weight generation for multi-task visual grounding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10857-10866 (2023) +37. Sun, J., Luo, G., Zhou, Y., Sun, X., Jiang, G., Wang, Z., Ji, R.: Refteacher: A strong baseline for semi-supervised referring expression comprehension. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19144-19154 (2023) +38. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017) +39. Wang, L., Li, Y., Huang, J., Lazebnik, S.: Learning two-branch neural networks for image-text matching tasks. IEEE Transactions on Pattern Analysis and Machine Intelligence 41(2), 394-407 (2018) +40. Wang, P., Wu, Q., Cao, J., Shen, C., Gao, L., Hengel, A.v.d.: Neighbourhood watch: Referring expression comprehension via language-guided graph attention networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1960-1968 (2019) +41. Wang, X., Liu, Y., Shen, C., Ng, C.C., Luo, C., Jin, L., Chan, C.S., Hengel, A.v.d., Wang, L.: On the general value of evidence, and bilingual scene-text visual question answering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10126-10135 (2020) +42. Wang, Z., Lu, Y., Li, Q., Tao, X., Guo, Y., Gong, M., Liu, T.: Cris: Clip-driven referring image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11686-11695 (2022) +43. Yang, L., Xu, Y., Yuan, C., Liu, W., Li, B., Hu, W.: Improving visual grounding with visual-linguistic verification and iterative reasoning. In: Proceedings of the + +IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9499-9508 (2022) +44. Yang, S., Li, G., Yu, Y.: Dynamic graph attention for referring expression comprehension. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4644-4653 (2019) +45. Yang, S., Li, G., Yu, Y.: Graph-structured referring expression reasoning in the wild. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9952–9961 (2020) +46. Yang, Z., Wang, J., Tang, Y., Chen, K., Zhao, H., Torr, P.H.: Lavo: Language-aware vision transformer for referring image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18155-18165 (2022) +47. Yang, Z., Chen, T., Wang, L., Luo, J.: Improving one-stage visual grounding by recursive sub-query construction. In: European Conference on Computer Vision. pp. 387-404. Springer (2020) +48. Yang, Z., Gong, B., Wang, L., Huang, W., Yu, D., Luo, J.: A fast and accurate one-stage approach to visual grounding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4683-4693 (2019) +49. Ye, J., Tian, J., Yan, M., Yang, X., Wang, X., Zhang, J., He, L., Lin, X.: Shifting more attention to visual backbone: Query-modulated refinement networks for end-to-end visual grounding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15502-15512 (2022) +50. You, Q., Jin, H., Wang, Z., Fang, C., Luo, J.: Image captioning with semantic attention. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4651-4659 (2016) +51. Yu, L., Lin, Z., Shen, X., Yang, J., Lu, X., Bansal, M., Berg, T.L.: Matttnet: Modular attention network for referring expression comprehension. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1307-1315 (2018) +52. Yu, L., Poirson, P., Yang, S., Berg, A.C., Berg, T.L.: Modeling context in referring expressions. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14. pp. 69-85. Springer (2016) +53. Yu, Z., Yu, J., Xiang, C., Zhao, Z., Tian, Q., Tao, D.: Rethinking diversified and discriminative proposal generation for visual grounding. arXiv preprint arXiv:1805.03508 (2018) +54. Zareian, A., Rosa, K.D., Hu, D.H., Chang, S.F.: Open-vocabulary object detection using captions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14393-14402 (2021) +55. Zhang, H., Niu, Y., Chang, S.F.: Grounding referring expressions in images by variational context. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4158-4166 (2018) +56. Zhang, Z., Zhu, Y., Liu, J., Liang, X., Ke, W.: Coupling word-pixel with sentence-mask alignments for referring image segmentation. arXiv preprint arXiv:2212.01769 (2022) +57. Zhuang, B., Wu, Q., Shen, C., Reid, I., Van Den Hengel, A.: Parallel attention: A unified framework for visual object discovery through dialogs and queries. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4252-4261 (2018) \ No newline at end of file diff --git a/2024/SegVG_ Transferring Object Bounding Box to Segmentation for Visual Grounding/images.zip b/2024/SegVG_ Transferring Object Bounding Box to Segmentation for Visual Grounding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2d417d7c6ebb5c6299318f603307a19c8fa16784 --- /dev/null +++ b/2024/SegVG_ Transferring Object Bounding Box to Segmentation for Visual Grounding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ee4da95e33436611e82e1a1f0e2d7911e18e82bbd14894071d856e21fe992b46 +size 596199 diff --git a/2024/SegVG_ Transferring Object Bounding Box to Segmentation for Visual Grounding/layout.json b/2024/SegVG_ Transferring Object Bounding Box to Segmentation for Visual Grounding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9a63bc1b934e8cfda1f4eed7eaf65708a3b0ede2 --- /dev/null +++ b/2024/SegVG_ Transferring Object Bounding Box to Segmentation for Visual Grounding/layout.json @@ -0,0 +1,8304 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 147, + 111, + 467, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 111, + 467, + 148 + ], + "spans": [ + { + "bbox": [ + 147, + 111, + 467, + 148 + ], + "type": "text", + "content": "SegVG: Transferring Object Bounding Box to Segmentation for Visual Grounding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 169, + 167, + 444, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 167, + 444, + 181 + ], + "spans": [ + { + "bbox": [ + 169, + 167, + 444, + 181 + ], + "type": "text", + "content": "Weitai Kang" + }, + { + "bbox": [ + 169, + 167, + 444, + 181 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 169, + 167, + 444, + 181 + ], + "type": "text", + "content": ", Gaowen Liu" + }, + { + "bbox": [ + 169, + 167, + 444, + 181 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 169, + 167, + 444, + 181 + ], + "type": "text", + "content": ", Mubarak Shah" + }, + { + "bbox": [ + 169, + 167, + 444, + 181 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 169, + 167, + 444, + 181 + ], + "type": "text", + "content": ", and Yan Yan" + }, + { + "bbox": [ + 169, + 167, + 444, + 181 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 239, + 189, + 373, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 189, + 373, + 200 + ], + "spans": [ + { + "bbox": [ + 239, + 189, + 373, + 200 + ], + "type": "text", + "content": "1 Illinois Institute of Technology" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 271, + 200, + 342, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 271, + 200, + 342, + 210 + ], + "spans": [ + { + "bbox": [ + 271, + 200, + 342, + 210 + ], + "type": "text", + "content": "2 Cisco Research" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 242, + 212, + 370, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 212, + 370, + 223 + ], + "spans": [ + { + "bbox": [ + 242, + 212, + 370, + 223 + ], + "type": "text", + "content": "3 University of Central Florida" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 253, + 455, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 253, + 455, + 506 + ], + "spans": [ + { + "bbox": [ + 160, + 253, + 455, + 506 + ], + "type": "text", + "content": "Abstract. Different from Object Detection, Visual Grounding deals with detecting a bounding box for each text-image pair. This one box for each text-image data provides sparse supervision signals. Although previous works achieve impressive results, their passive utilization of annotation, i.e. the sole use of the box annotation as regression ground truth, results in a suboptimal performance. In this paper, we present SegVG, a novel method transfers the box-level annotation as Segmentation signals to provide an additional pixel-level supervision for Visual Grounding. Specifically, we propose the Multi-layer Multi-task Encoder-Decoder as the target grounding stage, where we learn a regression query and multiple segmentation queries to ground the target by regression and segmentation of the box in each decoding layer, respectively. This approach allows us to iteratively exploit the annotation as signals for both box-level regression and pixel-level segmentation. Moreover, as the backbones are typically initialized by pretrained parameters learned from unimodal tasks and the queries for both regression and segmentation are static learnable embeddings, a domain discrepancy remains among these three types of features, which impairs subsequent target grounding. To mitigate this discrepancy, we introduce the Triple Alignment module, where the query, text, and vision tokens are triangurally updated to share the same space by triple attention mechanism. Extensive experiments on five widely used datasets validate our state-of-the-art (SOTA) performance. Code is available at https://github.com/WeitaiKang/SegVG." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 160, + 517, + 395, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 517, + 395, + 529 + ], + "spans": [ + { + "bbox": [ + 160, + 517, + 395, + 529 + ], + "type": "text", + "content": "Keywords: Visual Grounding " + }, + { + "bbox": [ + 160, + 517, + 395, + 529 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 517, + 395, + 529 + ], + "type": "text", + "content": " Detection " + }, + { + "bbox": [ + 160, + 517, + 395, + 529 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 517, + 395, + 529 + ], + "type": "text", + "content": " Transformer" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 567, + 230, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 230, + 580 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 230, + 580 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 593, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 593, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 593, + 482, + 668 + ], + "type": "text", + "content": "Visual grounding [14,17,26,30,52] aims to localize a target object within an image based on a free-form natural language text expression. It is particularly important for numerous downstream multimodal reasoning systems, such as visual question answering [10,35,41] and image captioning [1,4,50]. Previous works can be broadly categorized into three distinct groups: two-stage methods [3,44,45,51], one-stage methods [47,48], and transformer-based ones [6,15,16,31,43,49]. Both" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 179, + 114, + 438, + 268 + ], + "blocks": [ + { + "bbox": [ + 179, + 114, + 438, + 268 + ], + "lines": [ + { + "bbox": [ + 179, + 114, + 438, + 268 + ], + "spans": [ + { + "bbox": [ + 179, + 114, + 438, + 268 + ], + "type": "image", + "image_path": "411ba3ae96ffde238132b8d517bf288127d3eb0f330fc4fe4597e893c10d5b6d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 276, + 482, + 365 + ], + "lines": [ + { + "bbox": [ + 130, + 276, + 482, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 276, + 482, + 365 + ], + "type": "text", + "content": "Fig. 1: The comparison of visual grounding frameworks. The block with a dashed border indicates that the module may not necessarily exist. (a) Previous baseline method consists of two backbones and additional transformer layers for target grounding, where a regression query is supervised to regress the box. Current SOTA methods further employ a text-to-visual module to align the visual features with text features. (b) Our method incorporates segmentation queries, which utilizes the box annotation at the pixel-level to segment the target. Additionally, we propose the Triple Alignment module to eliminate the domain discrepancy of the query, text, and vision features." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 389, + 480, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 389, + 480, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 480, + 437 + ], + "type": "text", + "content": "two-stage and one-stage approaches use convolutional neural networks for candidate proposals and the selection of the best-matching candidate. Nonetheless, these approaches rely on intricate modules that employ manually-crafted techniques for performing language inference and multi-modal integration." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 437, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 568 + ], + "type": "text", + "content": "Inspired by the success of the transformer [7, 8], TransVG [6] proposes a transformer-based pipeline. As shown in Fig. 1.(a), this pipeline extracts vision and text features via DETR [2] and BERT [7], respectively. To ground the target, they use the transformer encoder to fuse multimodal features along with a learnable regression query and decode the query through an MLP. To enhance the final target grounding stage, subsequent studies continue with some text-to-visual modules in the early stage to modulate the vision features to align with the text features. For example, QRNet [49] proposes a query-modulated method for extracting language-aware vision features within the vision backbone. VLTVG [43] introduces a verification map to activate the vision features to align with the text features before multimodal fusion." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Despite their advancements, the suboptimal annotation utilization, i.e., only using the box annotation as a regression annotation, limits their performance. As discussed in [37], Visual Grounding presents unique challenges compared to Object Detection due to its sparse supervision signals. Specifically, it provides only one box label for each text-image pair, while necessitating detection within a multimodal setting. Therefore, it is essential to fully exploit the box annotation, by treating it as a segmentation mask (pixels within the bounding box are assigned a value of 1, while pixels outside the bounding box are assigned as 0)." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "W. Kang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 344 + ], + "type": "text", + "content": "In this paper, we introduce SegVG (see Fig. 1.(b)), a novel method that leverages the pixel-level details within the box annotation as segmentation signals to offer additional fine-grained supervision for Visual Grounding. Specifically, we propose the Multi-layer Multi-task Encoder-Decoder as the target grounding stage, where we learn a regression query and multiple segmentation queries to ground the target by regression and segmentation of the box in each decoding layer, respectively. The confidence score derived from the segmentation can further serve as a Focal Loss [21] scaling factor to adaptively emphasize the other losses of challenging training samples. This approach allows us to iteratively exploit the annotation as signals for both box-level regression and pixel-level segmentation. Furthermore, the initial parameters for model backbones, typically derived from pretrained unimodal tasks, along with data-agnostic static embeddings used as queries for decoding, result in a domain discrepancy among different sources of feature, affecting the effectiveness of target grounding. To tackle this problem, we present the Triple Alignment module, where we harmonize the domain of query, text, and vision features by implementing a triangular update process through a triple attention mechanism. As a result, we ensure that all features adapt and integrate within the same multimodal space, thereby enhancing subsequent target grounding. Our contributions are as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 354, + 480, + 487 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 138, + 354, + 479, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 354, + 479, + 388 + ], + "spans": [ + { + "bbox": [ + 138, + 354, + 479, + 388 + ], + "type": "text", + "content": "- We propose the Multi-layer Multi-task Encoder-Decoder to maximize the utilization of the box annotation, which introduces an additional segmentation format for pixel-level supervision in Visual Grounding." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 390, + 480, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 390, + 480, + 425 + ], + "spans": [ + { + "bbox": [ + 138, + 390, + 480, + 425 + ], + "type": "text", + "content": "- To eliminate the domain discrepancy among the query, text, and vision, we introduce the Triple Alignment to update these three types of features into a sharing domain, which facilitates the subsequent target grounding." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 426, + 480, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 426, + 480, + 474 + ], + "spans": [ + { + "bbox": [ + 138, + 426, + 480, + 474 + ], + "type": "text", + "content": "- We conduct extensive experiments on five widely used datasets to show the performance superiority of our proposed methods compared with previous state-of-the-art methods and further investigate the reliability benefits derived from the segmentation output in real applications." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 475, + 455, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 475, + 455, + 487 + ], + "spans": [ + { + "bbox": [ + 138, + 475, + 455, + 487 + ], + "type": "text", + "content": "- We will release code and checkpoints for future research development." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 517, + 237, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 517, + 237, + 529 + ], + "spans": [ + { + "bbox": [ + 132, + 517, + 237, + 529 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 545, + 479, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 479, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 479, + 568 + ], + "type": "text", + "content": "Visual grounding methods can be roughly classified into three pipelines: two-stage methods, one-stage methods, and transformer-based methods." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "Two-stage methods Two-stage approaches [3,51] treat visual grounding as first generating candidate object proposals and then finding the best match to the text. In the first stage, an off-the-shelf detector processes the image and proposes regions that may contain the target. In the second stage, a ranking network calculates the similarity between candidate regions and processed text features, selecting the region with the highest similarity score as the final result. Training losses include binary classification loss [29] or maximum-margin ranking loss [51]. To better understand the text and cross-modality matching, MattNet [51] focuses" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 147, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 147, + 91, + 447, + 102 + ], + "type": "text", + "content": "Transferring Object Bounding Box to Segmentation for Visual Grounding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 148, + 115, + 467, + 302 + ], + "blocks": [ + { + "bbox": [ + 148, + 115, + 467, + 302 + ], + "lines": [ + { + "bbox": [ + 148, + 115, + 467, + 302 + ], + "spans": [ + { + "bbox": [ + 148, + 115, + 467, + 302 + ], + "type": "image", + "image_path": "73df9bb07d93244cf5086901d4f2c7a651af179c31859a2f5a5ebe3fb84b9f16.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 311, + 482, + 400 + ], + "lines": [ + { + "bbox": [ + 130, + 311, + 482, + 400 + ], + "spans": [ + { + "bbox": [ + 130, + 311, + 482, + 400 + ], + "type": "text", + "content": "Fig. 2: SegVG: The upper figure includes the vision and text backbone. Our proposed Triple Alignment module is iteratively inserted into intermediate layers to eliminate domain discrepancy. The lower figure shows our Multi-layer Multi-task Encoder-Decoder, which adopts a transformer encoder-decoder to update multimodal features and ground the target. In this architecture, we make the best of the box annotation as a segmentation ground truth and integrate an additional segmentation task into Visual Grounding. Additionally, the segmentation output serves as a Focal Loss factor, allowing adaptive emphasis on challenging cases for the regression loss. " + }, + { + "bbox": [ + 130, + 311, + 482, + 400 + ], + "type": "inline_equation", + "content": "\\mathrm{M} = 6" + }, + { + "bbox": [ + 130, + 311, + 482, + 400 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 311, + 482, + 400 + ], + "type": "inline_equation", + "content": "\\mathrm{R} = 6" + }, + { + "bbox": [ + 130, + 311, + 482, + 400 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 424, + 479, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 424, + 479, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 424, + 479, + 449 + ], + "type": "text", + "content": "on decomposing the text into subject, location, and relationship components. [3] introduces an expression-aware score for improved candidate region ranking." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 449, + 480, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 480, + 521 + ], + "type": "text", + "content": "One-stage methods One-stage approaches [47, 48] directly concatenate vision and text features in the channel dimension and rank confidence values for candidate regions proposed based on the concatenated multimodal features. For example, FAOA [48] predicts bounding boxes using a YOLOv3 detector [32] on the concatenated features. ReSC [47] further improves the ability to ground complex queries by introducing a recursive sub-query construction module." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "type": "text", + "content": "Transformer-based methods Transformer-based approach is first introduced by TransVG [6]. Unlike previous methods, TransVG concatenates the regression query (a learnable embedding), vision tokens, and text tokens and uses transformer encoders [38] to perform cross-modal fusion and target grounding. The query is then processed through an MLP to decode the box. Benefiting from the flexible structure of transformer modules in processing multimodal features, recent works continue adopting this pipeline and propose novelties regarding feature extraction. VLTVG [43] develops a visual-linguistic verification module before the target grounding stage to modulate the vision features with the relationship between vision and text features. QRNet [49] proposes a Query-modulated Refinement Network to mitigate the gap between features from the unimodal vision backbone and those needed for multi-modal reasoning." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "W. Kang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "Multi-task Visual Grounding Multi-task learning is extensively utilized in object detection and segmentation [2,11], often capitalizing on a shared backbone and task-specific heads. Expanding upon this idea, several studies [19, 25, 36] have proposed solutions to the Multi-task Visual Grounding problem. In this problem, they jointly tackle Referring Expression Comprehension (REC, also known as Visual Grounding) and Referring Expression Segmentation (RES), requiring both box annotations and segmentation annotations. It is important to note that, unlike those approaches, even though we incorporate segmentation losses in our method, we do not need segmentation annotations but only box annotations, focusing specifically on the Visual Grounding task." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 254, + 233, + 269 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 254, + 233, + 269 + ], + "spans": [ + { + "bbox": [ + 132, + 254, + 233, + 269 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 279, + 481, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 279, + 481, + 316 + ], + "spans": [ + { + "bbox": [ + 130, + 279, + 481, + 316 + ], + "type": "text", + "content": "In this section, we present the components of our SegVG in the order of the data flow: starting with the backbones, followed by our proposed Triple Alignment, and finally our Multi-layer Multi-task Encoder-Decoder." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 333, + 214, + 344 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 333, + 214, + 344 + ], + "spans": [ + { + "bbox": [ + 132, + 333, + 214, + 344 + ], + "type": "text", + "content": "3.1 Backbones" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 353, + 482, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 353, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 353, + 482, + 403 + ], + "type": "text", + "content": "As shown in Fig. 2 (upper), our vision backbone consists of ResNet and transformer encoder from DETR [2], with parameters pretrained on Object Detection task using the MSCOCO dataset [22], excluding the validation and test sets of Visual Grounding dataset. The text backbone is the base model of BERT [7]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 412, + 482, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 412, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 412, + 482, + 498 + ], + "type": "text", + "content": "Vision Backbone Given an input image " + }, + { + "bbox": [ + 130, + 412, + 482, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{I_0}" + }, + { + "bbox": [ + 130, + 412, + 482, + 498 + ], + "type": "inline_equation", + "content": "(\\mathbb{R}^{3\\times H_0\\times W_0})" + }, + { + "bbox": [ + 130, + 412, + 482, + 498 + ], + "type": "text", + "content": ", we employ ResNet to generate a 2D feature map " + }, + { + "bbox": [ + 130, + 412, + 482, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{I}\\in \\mathbb{R}^{C\\times H\\times W}" + }, + { + "bbox": [ + 130, + 412, + 482, + 498 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 412, + 482, + 498 + ], + "type": "inline_equation", + "content": "C = 2048" + }, + { + "bbox": [ + 130, + 412, + 482, + 498 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 412, + 482, + 498 + ], + "type": "inline_equation", + "content": "H = \\frac{H_0}{32}" + }, + { + "bbox": [ + 130, + 412, + 482, + 498 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 412, + 482, + 498 + ], + "type": "inline_equation", + "content": "W = \\frac{W_0}{32}" + }, + { + "bbox": [ + 130, + 412, + 482, + 498 + ], + "type": "text", + "content": "). A 1x1 convolutional layer is then used to reduce the channel dimension of " + }, + { + "bbox": [ + 130, + 412, + 482, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 130, + 412, + 482, + 498 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 412, + 482, + 498 + ], + "type": "inline_equation", + "content": "C_v = 256" + }, + { + "bbox": [ + 130, + 412, + 482, + 498 + ], + "type": "text", + "content": ", resulting in " + }, + { + "bbox": [ + 130, + 412, + 482, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{I}'" + }, + { + "bbox": [ + 130, + 412, + 482, + 498 + ], + "type": "text", + "content": ". We further flatten " + }, + { + "bbox": [ + 130, + 412, + 482, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{I}'" + }, + { + "bbox": [ + 130, + 412, + 482, + 498 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 130, + 412, + 482, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_v\\in \\mathbb{R}^{C_v\\times N_v}" + }, + { + "bbox": [ + 130, + 412, + 482, + 498 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 412, + 482, + 498 + ], + "type": "inline_equation", + "content": "N_v = H\\times W" + }, + { + "bbox": [ + 130, + 412, + 482, + 498 + ], + "type": "text", + "content": "). Position embedding is then added to " + }, + { + "bbox": [ + 130, + 412, + 482, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_v" + }, + { + "bbox": [ + 130, + 412, + 482, + 498 + ], + "type": "text", + "content": " to preserve sensitivity to the original 2D spatial locations. " + }, + { + "bbox": [ + 130, + 412, + 482, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_v" + }, + { + "bbox": [ + 130, + 412, + 482, + 498 + ], + "type": "text", + "content": " is then iteratively processed through DETR's encoder layer (total 6 transformer layers) and the Triple Alignment to obtain the output " + }, + { + "bbox": [ + 130, + 412, + 482, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_v" + }, + { + "bbox": [ + 130, + 412, + 482, + 498 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 508, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 508, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 508, + 482, + 582 + ], + "type": "text", + "content": "Text Backbone Given a text, we initially utilize the BERT's embedding layer to convert it into " + }, + { + "bbox": [ + 130, + 508, + 482, + 582 + ], + "type": "inline_equation", + "content": "N_{t}" + }, + { + "bbox": [ + 130, + 508, + 482, + 582 + ], + "type": "text", + "content": " language tokens with " + }, + { + "bbox": [ + 130, + 508, + 482, + 582 + ], + "type": "inline_equation", + "content": "C_t" + }, + { + "bbox": [ + 130, + 508, + 482, + 582 + ], + "type": "text", + "content": " channel dimension. In alignment with [6], we preprocess a [CLS] token and append a [SEP] token to the beginning and end positions of the tokenized language, respectively. Following this, we iteratively input the language tokens into BERT's layers (total 12 transformer layers) and the Triple Alignment, generating language embedding " + }, + { + "bbox": [ + 130, + 508, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_t" + }, + { + "bbox": [ + 130, + 508, + 482, + 582 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 597, + 249, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 597, + 249, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 597, + 249, + 611 + ], + "type": "text", + "content": "3.2 Triple Alignment" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "content": "Given that the text and vision backbones are pretrained from unimodal tasks and the queries are data-agnostic, the subsequent target grounding stage faces the challenge of aligning these three types of features into the same space before performing multimodal fusion for target grounding. Additionally, considering" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 147, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 147, + 91, + 447, + 103 + ], + "type": "text", + "content": "Transferring Object Bounding Box to Segmentation for Visual Grounding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "that the backbones usually contribute the majority of the overall parameters, solely using them for extracting unimodal features without incorporating multimodal alignment is sub-optimal. Therefore, an optimal solution is to address the domain discrepancy before moving on to the subsequent target grounding stage." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 480, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 480, + 224 + ], + "type": "text", + "content": "Triple Alignment (Fig. 2 (upper)) utilizes an attention mechanism to perform triangular feature sampling, aiming to ensure domain consistency among the query, text, and vision features. The queries, " + }, + { + "bbox": [ + 130, + 164, + 480, + 224 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_o" + }, + { + "bbox": [ + 130, + 164, + 480, + 224 + ], + "type": "text", + "content": ", are first initialized by N learnable embeddings, where one embedding is for the regression query and the rest of the embeddings are for multiple segmentation queries. The data flow is:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 211, + 230, + 480, + 244 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 230, + 480, + 244 + ], + "spans": [ + { + "bbox": [ + 211, + 230, + 480, + 244 + ], + "type": "interline_equation", + "content": "\\mathbf {Z} _ {v} ^ {i + 1} = \\operatorname {D E T R L a y e r} _ {i} \\left(\\mathbf {Z} _ {v} ^ {i}\\right), i \\in \\{0, 1,.., L - 1 \\} \\tag {1}", + "image_path": "f99dcdd24d8214bf3b4a3630391a5e2fe1af26fd09048f0a131baa2a0ef39ca4.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 211, + 260, + 480, + 275 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 260, + 480, + 275 + ], + "spans": [ + { + "bbox": [ + 211, + 260, + 480, + 275 + ], + "type": "interline_equation", + "content": "\\mathbf {Z} _ {t} ^ {i + 1} = \\operatorname {B E R T L a y e r} _ {\\mathcal {2 i + 1}} (\\operatorname {B E R T L a y e r} _ {\\mathcal {2 i}} (\\mathbf {Z} _ {t} ^ {i})), \\tag {2}", + "image_path": "5b0f9ae92a28eef9df96928e8703ce159a5ec4f807d0e174af425c220b6c1b67.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 220, + 288, + 480, + 303 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 288, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 220, + 288, + 480, + 303 + ], + "type": "interline_equation", + "content": "\\left[ \\mathbf {Z} _ {o} ^ {\\prime}, \\mathbf {Z} _ {t} ^ {\\prime}, \\mathbf {Z} _ {v} ^ {\\prime} \\right] = \\operatorname {T r i - M H A} \\left(\\mathbf {Z} _ {o}, \\mathbf {Z} _ {t} ^ {i + 1}, \\mathbf {Z} _ {v} ^ {i + 1}\\right), \\tag {3}", + "image_path": "4b70fdb93d51c1c4d77f89be5a96ddefff25ff6871a1583ed66a1e69a58715b5.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 194, + 316, + 480, + 331 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 316, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 194, + 316, + 480, + 331 + ], + "type": "interline_equation", + "content": "\\mathbf {Z} _ {o} = \\mathbf {Z} _ {o} + \\mathbf {Z} _ {o} ^ {\\prime}, \\mathbf {Z} _ {t} ^ {i + 1} = \\mathbf {Z} _ {t} ^ {i + 1} + \\mathbf {Z} _ {t} ^ {\\prime}, \\mathbf {Z} _ {v} ^ {i + 1} = \\mathbf {Z} _ {v} ^ {i + 1} + \\mathbf {Z} _ {v} ^ {\\prime}, \\tag {4}", + "image_path": "8fca60374039456c5fc00f75f0c839256b425bd3ee2dd7ad8fff33e538f723b6.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 333, + 479, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 333, + 479, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 333, + 479, + 429 + ], + "type": "text", + "content": "where L is the number of layers, BERTLayer is the layer of BERT and DETRLayer is the layer of DETR's encoder. The vision and text features are first encoded by Eq. 1 and Eq. 2. Subsequently, the three types of tokens (query, text, and vision) are updated by our Triple Multi-Head Attention Layer (Tri-MHA) using Eq. 3. The output tokens are merged back to their original branches respectively by Eq. 4. Within each head of the Triple Multi-Head Attention Layer (Tri-MHA), each type of the features simultaneously computes its updated representation by attending to both the others and itself:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 197, + 435, + 417, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 435, + 417, + 449 + ], + "spans": [ + { + "bbox": [ + 197, + 435, + 417, + 449 + ], + "type": "interline_equation", + "content": "S = \\left[ \\mathbf {Z} _ {o} W ^ {(o, S)}, \\mathbf {Z} _ {t} W ^ {(t, S)}, \\mathbf {Z} _ {v} W ^ {(v, S)} \\right], S \\in \\{Q, K, V \\}", + "image_path": "3c5af39287bae86f9176619939e8aa1452470a05fd2cc91cb8a2182b84d05ef9.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 231, + 453, + 480, + 467 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 453, + 480, + 467 + ], + "spans": [ + { + "bbox": [ + 231, + 453, + 480, + 467 + ], + "type": "interline_equation", + "content": "[ \\mathbf {Z} _ {o}, \\mathbf {Z} _ {t}, \\mathbf {Z} _ {v} ] = \\operatorname {S o f t M a x} \\left(Q K ^ {T} / \\sqrt {d _ {k}}\\right) V, \\tag {5}", + "image_path": "d5e17cb18d76e0ce3e81ba8ceafdf544d578592d8976b3854a37143aedfadafc.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 252, + 471, + 362, + 484 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 471, + 362, + 484 + ], + "spans": [ + { + "bbox": [ + 252, + 471, + 362, + 484 + ], + "type": "interline_equation", + "content": "\\mathbf {Z} _ {e} ^ {\\prime} = \\mathbf {Z} _ {e} W ^ {e}, e \\in \\{o, t, v \\},", + "image_path": "42c499b16754d2f5291436edfecd012d3bd9f18c5b7da61081956c4c74ac10a5.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 491, + 479, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 491, + 479, + 528 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 479, + 528 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 491, + 479, + 528 + ], + "type": "inline_equation", + "content": "\\{W^{(e,S)}, W^e : e \\in \\{o, t, v\\}, S \\in \\{Q, K, V\\}\\}" + }, + { + "bbox": [ + 130, + 491, + 479, + 528 + ], + "type": "text", + "content": " are trainable parameters. As a result, each of the output features is triangular sampling from all of the three types of features, which alleviates the domain discrepancy." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 543, + 364, + 554 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 543, + 364, + 554 + ], + "spans": [ + { + "bbox": [ + 131, + 543, + 364, + 554 + ], + "type": "text", + "content": "3.3 Multi-layer Multi-task Encoder-Decoder" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 560, + 479, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 560, + 479, + 596 + ], + "spans": [ + { + "bbox": [ + 130, + 560, + 479, + 596 + ], + "type": "text", + "content": "The Multi-layer Multi-task Encoder-Decoder serves as the target grounding stage, where we use a transformer encoder-decoder for cross-modal fusion and target grounding to perform a box regression task and a box segmentation task." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "type": "text", + "content": "Encoder As shown in Fig. 2 (lower left), given the aligned output text and vision features from the backbones, the encoder fuses the two modalities into the multimodal features by a stack of transformer layers. In each layer, the concatenated text and vision tokens go through the Multi-Head Self-Attention layer (MHSA) and the Feed Forward Network (FFN) with the residual connection." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "W. Kang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 416 + ], + "type": "text", + "content": "Decoder In each decoder layer, we aim to fully exploit the box annotation. We propose the bbox2seg paradigm to transform the box annotation into a segmentation mask, which classifies all pixels within the box as foreground (with a value of one) and those outside the box as background (with a value of zero). As shown in Fig. 2 (lower right), one regression query aims to regress the box, while the remaining segmentation queries aim to segment the box. Different segmentation queries are endowed with different learnable positional embeddings to enhance the robustness of each decoder layer, since the decoder layer, when confronted with various queries, is required to segment the same box. Following that, the queries pass through the Multi-Head Self-Attention layer to exchange information about the same target, prompting each other to better locate the target. Subsequently, the queries undergo the Multi-Head Cross-Attention layer and the Feed Forward Network, where multimodal features serve as the Key and Value to ground the target. Finally, a shared MLP across all decoder layers decodes the regression query into the box result, supervised by L1 loss and Giou loss [33]. Each segmentation query is repeated " + }, + { + "bbox": [ + 130, + 116, + 482, + 416 + ], + "type": "inline_equation", + "content": "N_{v}" + }, + { + "bbox": [ + 130, + 116, + 482, + 416 + ], + "type": "text", + "content": " times and concatenated with visual tokens along the channel dimension. Another shared MLP decodes the concatenated feature into the segmentation mask, supervised by Focal loss [21] and Dice loss [27]. It is noteworthy that our segmentation paradigm shares the same semantic foundation as the regression paradigm, i.e., to distinguish bounding box, rather than instance segmentation. Therefore, incorporating non-object pixels in the segmented foreground does not introduce ambiguity to the model. We provide qualitative results 4.8 to demonstrate this feature. To alleviate multi-task optimization challenges, we freeze the backbones for the initial k epochs to stabilize the training process." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 424, + 482, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 424, + 482, + 508 + ], + "spans": [ + { + "bbox": [ + 130, + 424, + 482, + 508 + ], + "type": "text", + "content": "Confidence score Since both the regression output and segmentation output share the same aim, we can additionally obtain the confidence score for the foreground by averaging values inside the ground truth box of the segmentation output to reflect the confidence of the regression output. In the training process, we can transform this confidence score as the Focal loss factor [21] to adaptively emphasize the other losses of challenging training samples. The final loss function of each decoder layer is formulated as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 230, + 514, + 481, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 514, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 230, + 514, + 481, + 533 + ], + "type": "interline_equation", + "content": "L = \\lambda_ {1} c _ {f o c a l} L _ {1} + \\lambda_ {g i o u} c _ {f o c a l} L _ {g i o u} + \\tag {6}", + "image_path": "8b4c6efd744f222f78481980419262bbcbe4dc1c05acb07b3f93ae5a67875068.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 251, + 529, + 385, + 541 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 529, + 385, + 541 + ], + "spans": [ + { + "bbox": [ + 251, + 529, + 385, + 541 + ], + "type": "interline_equation", + "content": "\\lambda_ {d i c e} c _ {f o c a l} L _ {d i c e} + \\lambda_ {f o c a l} L _ {f o c a l},", + "image_path": "813243e29fd52678853c81fb7b3edc9603093107caec1a5e6a253e1bb86bd114.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 546, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 582 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 546, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 130, + 546, + 482, + 582 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 546, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\lambda_{giou}" + }, + { + "bbox": [ + 130, + 546, + 482, + 582 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 546, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\lambda_{focal}" + }, + { + "bbox": [ + 130, + 546, + 482, + 582 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 546, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\lambda_{dice}" + }, + { + "bbox": [ + 130, + 546, + 482, + 582 + ], + "type": "text", + "content": " are hyperparameters. " + }, + { + "bbox": [ + 130, + 546, + 482, + 582 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 130, + 546, + 482, + 582 + ], + "type": "text", + "content": " is the L1 loss. " + }, + { + "bbox": [ + 130, + 546, + 482, + 582 + ], + "type": "inline_equation", + "content": "L_{giou}" + }, + { + "bbox": [ + 130, + 546, + 482, + 582 + ], + "type": "text", + "content": " is the GIoU loss [33]. " + }, + { + "bbox": [ + 130, + 546, + 482, + 582 + ], + "type": "inline_equation", + "content": "L_{focal}" + }, + { + "bbox": [ + 130, + 546, + 482, + 582 + ], + "type": "text", + "content": " is the Focal loss [21]. " + }, + { + "bbox": [ + 130, + 546, + 482, + 582 + ], + "type": "inline_equation", + "content": "L_{dice}" + }, + { + "bbox": [ + 130, + 546, + 482, + 582 + ], + "type": "text", + "content": " is the Dice loss [27]. " + }, + { + "bbox": [ + 130, + 546, + 482, + 582 + ], + "type": "inline_equation", + "content": "c_{focal}" + }, + { + "bbox": [ + 130, + 546, + 482, + 582 + ], + "type": "text", + "content": " is the above Focal loss factor averaged across all segmentation outputs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": "In the real-world application perspective, the visual grounding task can be viewed as open-vocabulary object detection [54], where target objects lack predetermined categories. Therefore, previous transformer-based methods directly regress the box without confidence scores, since there is no candidate proposal or selection stage in transformer-based pipeline. However, confidence scores are valuable for enhancing the control or reliability of predictions by filtering out low-confidence predictions. This feature could benefit the future integration of visual" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 147, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 147, + 91, + 448, + 103 + ], + "type": "text", + "content": "Transferring Object Bounding Box to Segmentation for Visual Grounding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 154, + 146, + 462, + 364 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 1: Comparisons with state-of-the-art methods on widely used datasets. We highlight the best and second best performance in red and blue, and bold our model." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 154, + 146, + 462, + 364 + ], + "lines": [ + { + "bbox": [ + 154, + 146, + 462, + 364 + ], + "spans": [ + { + "bbox": [ + 154, + 146, + 462, + 364 + ], + "type": "table", + "html": "
ModelsBackboneRefCOCORefCOCO+RefCOCOgReferItGame test
valtestAtestBvaltestAtestBval-gval-utest-u
Two-stage:
CMN [12]VGG16-71.0365.77-54.3247.7657.47--28.33
VC [55]VGG16-73.3367.44-58.4053.1862.30--31.13
ParalAttn [57]VGG16-75.3165.52-61.3450.8658.03---
MAttNet [51]ResNet-10176.6581.1469.9965.3371.6256.02-66.5867.2729.04
Similarity Net [39]ResNet-101---------34.54
CITE [29]ResNet-101---------35.07
DDPN [53]ResNet-101---------63.00
LGRANs [40]VGG16-76.6066.40-64.0053.4061.78---
DGA [44]VGG16-78.4265.53-69.0751.99--63.28-
RvG-Tree [29]ResNet-10175.0678.6169.8563.5167.4556.66-66.9566.51-
NMTree [23]ResNet-10176.4181.2170.0966.4672.0257.5264.6265.8766.44-
Ref-NMS [3]ResNet-10180.7084.0076.0468.2573.6859.42-70.5570.62-
One-stage:
SSG [5]DarkNet-53-76.5167.50-62.1449.2747.4758.80-54.24
ZSGNet [34]ResNet-50---------58.63
FAOA [48]DarkNet-5372.5474.3568.5056.8160.2349.6056.1261.3360.3660.67
RCCF [20]DLA-34-81.0671.85-70.3556.32--65.7363.79
ReSC-Large [47]DarkNet-5377.6380.4572.3063.5968.3656.8163.1267.3067.2064.60
LBYL-Net [13]DarkNet-5379.6782.9174.1568.6473.3859.4962.70--67.47
Transformer-based:
TransVG [6]ResNet-10181.0282.7278.3564.8270.7056.9467.0268.6767.7370.73
QRNet [49]Swin-S84.0185.8582.3472.9476.1763.8171.8973.0372.5274.61
VLTVG [43]ResNet-10184.7787.2480.4974.1978.9365.1772.9876.0474.1871.98
SegVG (ours)ResNet-10186.8489.4683.0777.1882.6367.5976.0178.3577.4275.59
", + "image_path": "34fa8a3bb2ea6caafb8b3dee8b79473c88b80965f9984a34a2fb30ed605acac0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 387, + 482, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 387, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 387, + 482, + 472 + ], + "type": "text", + "content": "grounding models into downstream multimodal reasoning systems or real-world applications. To meet the requirements of this feature, our approach incorporates a confidence score derived from the segmentation output during inference. Specifically, we calculate the model's confidence by averaging values greater than or equal to 0.35 (adopted from [42]) in the segmentation output of one segmentation query. Analyses 4.7 in the Experiment section demonstrate the faithfulness and benefits of incorporating this additional confidence score." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 489, + 230, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 489, + 230, + 502 + ], + "spans": [ + { + "bbox": [ + 132, + 489, + 230, + 502 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 514, + 264, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 264, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 264, + 525 + ], + "type": "text", + "content": "4.1 Metric and Datasets" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 533, + 481, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 481, + 583 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 481, + 583 + ], + "type": "text", + "content": "Metric A predicted bounding box is considered accurate if its Intersection over Union (IoU) with the ground-truth bounding box exceeds 0.5. In accordance with the established practices in preceding studies [6,43], we employ top-1 accuracy (measured in percentage) as the primary metric to assess our method." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 593, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 593, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 593, + 482, + 666 + ], + "type": "text", + "content": "Datasets There are five standard benchmarks: RefCOCO [52], RefCOCO+ [52], RefCOCOg-g [26], RefCOCOg-umd [26], and ReferItGame [17]. Four of them (RefCOCO, RefCOCO+, and RefCOCOg-(g/umd)) are all derived from MSCOCO [22]. RefCOCO consists of 19,994 images and 142,210 referring texts, which is divided into four subsets: a training set with 120,624 texts, a validation set with 10,834 texts, and two test sets (testA and testB) containing 5,657" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "W. Kang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "and 5,095 texts, respectively. RefCOCO+ includes 19,992 images and 141,564 referring texts, which is partitioned into four subsets: a training set with 120,191 texts, a validation set with 10,758 texts, and two test sets (testA and testB) containing 5,726 and 4,889 texts, respectively. RefCOCOg contains 25,799 images and 95,010 longer texts. Two widely accepted splitting methods—RefCOCOg-g [26] and RefCOCOg-umd [28]—are employed for this dataset, and we perform experiments using both RefCOCOg-g (val-g) and RefCOCOg-umd (val-u and test-u) splitting conventions. The ReferItGame dataset, featuring 20,000 images from SAIAPR-12 [9], is divided into three segments: a training set with 54,127 texts, a validation set with 5,842 texts, and a testing set comprising 60,103 texts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 255, + 242, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 255, + 242, + 268 + ], + "spans": [ + { + "bbox": [ + 132, + 255, + 242, + 268 + ], + "type": "text", + "content": "4.2 Implementation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 277, + 482, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 482, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 482, + 373 + ], + "type": "text", + "content": "Our approach uses an input image size of " + }, + { + "bbox": [ + 130, + 277, + 482, + 373 + ], + "type": "inline_equation", + "content": "640 \\times 640" + }, + { + "bbox": [ + 130, + 277, + 482, + 373 + ], + "type": "text", + "content": " and sets the maximum expression length at 40. When resizing images, we preserve the original aspect ratio. The longer edge is resized to 640, and the shorter edge is padded to 640 with the value of zero. Texts exceeding 38 tokens are truncated, reserving a start position and an end position of the characters for the [CLS] and [SEP] token, respectively. If the text is shorter, empty tokens are added after the [SEP] token to reach an input length of 40. Paddings for the input image are not tracked by masks, while empty tokens of text employ masks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 373, + 482, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 373, + 482, + 494 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 482, + 494 + ], + "type": "text", + "content": "We use the AdamW optimizer. The initial learning rate of 1e-5 is assigned to backbones, and 1e-4 to the rest parameters. Weight decay is set at 1e-4. The visual backbone is initialized with the DETR's backbone and encoder, while the language branch uses the basic BERT model. For the final results, our model is trained for 90 epochs, with the learning rate decreasing by a factor of 10 after 60 epochs. The k hyperparameter in Multi-layer Multi-task Encoder-Decoder is set to 10. We use a batch size of 64. For the ablation studies presented in Table 3, the models are trained for 60 epochs with k equal to 20, and the learning rate drops after 40 epochs. We set " + }, + { + "bbox": [ + 130, + 373, + 482, + 494 + ], + "type": "inline_equation", + "content": "\\lambda_{1} = 5" + }, + { + "bbox": [ + 130, + 373, + 482, + 494 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 373, + 482, + 494 + ], + "type": "inline_equation", + "content": "\\lambda_{giou} = 2" + }, + { + "bbox": [ + 130, + 373, + 482, + 494 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 373, + 482, + 494 + ], + "type": "inline_equation", + "content": "\\lambda_{focal} = 1" + }, + { + "bbox": [ + 130, + 373, + 482, + 494 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 373, + 482, + 494 + ], + "type": "inline_equation", + "content": "\\lambda_{dice} = 1" + }, + { + "bbox": [ + 130, + 373, + 482, + 494 + ], + "type": "text", + "content": ". We adhere to previous practices [6, 43] for data augmentation during training." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 512, + 265, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 265, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 265, + 525 + ], + "type": "text", + "content": "4.3 Quantitative Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "We report the performance of our SegVG on all benchmark datasets. As presented in Table 1, our SegVG model demonstrates superiority across all of the datasets. This indicates the effectiveness and generalizability of our approach. It is worth noting that RefCOCO+ and RefCOCOg are relatively more challenging datasets, as RefCOCO+ does not include location terms in its language expressions, and RefCOCOg has longer language expressions compared to other datasets. Despite these challenges, our model exhibits significant improvements on these two difficult datasets. Specifically, on RefCOCO+, our model outperforms the previous SOTA models with " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "+2.99\\%" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "+3.7\\%" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "+2.42\\%" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " on the val, testA, and testB subsets, respectively. On RefCOCOg, our model also surpasses the previous SOTA models with " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "+3.03\\%" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "+2.31\\%" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "+3.24\\%" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " on the val-g," + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 147, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 147, + 91, + 447, + 103 + ], + "type": "text", + "content": "Transferring Object Bounding Box to Segmentation for Visual Grounding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 171, + 135, + 443, + 186 + ], + "blocks": [ + { + "bbox": [ + 140, + 114, + 472, + 126 + ], + "lines": [ + { + "bbox": [ + 140, + 114, + 472, + 126 + ], + "spans": [ + { + "bbox": [ + 140, + 114, + 472, + 126 + ], + "type": "text", + "content": "Table 2: Compare transformer-based models on Parameter count and GFLOPS." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 171, + 135, + 443, + 186 + ], + "lines": [ + { + "bbox": [ + 171, + 135, + 443, + 186 + ], + "spans": [ + { + "bbox": [ + 171, + 135, + 443, + 186 + ], + "type": "table", + "html": "
BackboneParameter count (M)GFLOPS (G)
TransVG [6]ResNet101141.5572.61
VLTVG [43]ResNet101141.6169.87
QRNet [49]Swin-S247.0680.12
SegVGResNet101155.2873.48
", + "image_path": "2c7dac147682ba1d367a7ed158acdc81fe015f05c40e72a5a698159bd3d6b6f4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 174, + 259, + 444, + 369 + ], + "blocks": [ + { + "bbox": [ + 130, + 195, + 482, + 251 + ], + "lines": [ + { + "bbox": [ + 130, + 195, + 482, + 251 + ], + "spans": [ + { + "bbox": [ + 130, + 195, + 482, + 251 + ], + "type": "text", + "content": "Table 3: Ablation study on RefCOCOg test-u. Encoder and Decoder are the encoder and decoder of Multi-layer Multi-task Encoder-Decoder, respectively. MMDecoder represents Multi-layer and Multi-task supervision in the Decoder. N is the number of segmentation queries. Triple means Triple Alignment. Excluding Query in Triple (g) means using bidirectional alignment with concatenated text-vision tokens." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 174, + 259, + 444, + 369 + ], + "lines": [ + { + "bbox": [ + 174, + 259, + 444, + 369 + ], + "spans": [ + { + "bbox": [ + 174, + 259, + 444, + 369 + ], + "type": "table", + "html": "
IdModelAcc(%)
(a)Backbones + Encoder + Decoder66.97
(b)Backbones + Triple + Encoder + Decoder75.75
(c)Backbones + Encoder + MMDecoder (N=1)72.61
(d)Backbones + Encoder + MMDecoder (N=5)76.21
(e)Backbones + Triple + Encoder + MMDecoder (N=5)77.29
(f)(e) w/o Encoder76.65(-0.64)
(g)(e) w/o Query in Triple76.37(-0.92)
(h)(e) w/o Focal loss76.83(-0.46)
", + "image_path": "790acc68e2eab7f52d69644d371b5adb0c6435d292cbe858d81764e0aecaa489.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 390, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 482, + 449 + ], + "type": "text", + "content": "val-u, and test-u subsets, respectively. These results suggest that under the reinforcement of Triple Alignment and Multi-layer Multi-task Encoder-Decoder, the query, text, and vision tokens are triangularly updated to share the same space, and the model fully exploits the bounding box as fine-grained pixel-level supervision for comprehensive end-to-end learning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 449, + 483, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 483, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 483, + 486 + ], + "type": "text", + "content": "We also conduct a comparison of the number of parameters and GFLOPS across transformer-based models to evaluate computational costs. As depicted in Table 2, the computational cost of SegVG falls within a reasonable range." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 502, + 239, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 502, + 239, + 514 + ], + "spans": [ + { + "bbox": [ + 131, + 502, + 239, + 514 + ], + "type": "text", + "content": "4.4 Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 521, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 629 + ], + "type": "text", + "content": "In this section, we aim to validate the efficacy of each proposed module. We conduct ablation studies on the RefCOCOg-umd test dataset. Specifically, we start by evaluating a basic structure, i.e., the backbones with encoder-decoder structure. After that, we systematically incorporate the Triple Alignment module into the backbones and introduce the Multi-layer Multi-task supervision into the decoder through the controlled variable approach. Meanwhile, we conduct additional ablation experiments on specific details, including assessing the efficacy of incorporating the encoder, introducing Query in the Triple Alignment, and introducing the Focal loss from the segmentation output to the other losses." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "As shown in Table 3, we can draw the following conclusions when comparing the experimental results under controlled variables: 1) [(a) v.s. (b)]: Incorporating the Triple Alignment can effectively eliminate the domain discrepancy" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "W. Kang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 181, + 157, + 435, + 218 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 149 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 149 + ], + "type": "text", + "content": "Table 4: Ablation study on RefCOCOg test-u regarding the number of segmentation queries. " + }, + { + "bbox": [ + 131, + 114, + 482, + 149 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 131, + 114, + 482, + 149 + ], + "type": "text", + "content": " Hour and " + }, + { + "bbox": [ + 131, + 114, + 482, + 149 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 131, + 114, + 482, + 149 + ], + "type": "text", + "content": " Acc are the additional time cost and accuracy improvement compared to ID(i) experiment, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 181, + 157, + 435, + 218 + ], + "lines": [ + { + "bbox": [ + 181, + 157, + 435, + 218 + ], + "spans": [ + { + "bbox": [ + 181, + 157, + 435, + 218 + ], + "type": "table", + "html": "
IDnum_queryAccTime CostΔHour / ΔAcc (v.s. ID(i))
iN = 172.6116.32h-
iiN = 373.0218.97h6.46
iiiN = 576.2120.10h1.05
ivN = 774.3822.24h3.34
vN = 973.9124.07h5.96
", + "image_path": "dd8ace7895fd9ff62d8c47d167304e138e5dd13faafe5e76c634cf3079bbf960.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 241, + 482, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 241, + 482, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 241, + 482, + 540 + ], + "type": "text", + "content": "among the query, text, and vision features, thereby facilitating subsequent target grounding. 2) [(a) v.s. (c)]: Introducing Multi-layer Multi-task supervision can iteratively make the best of the annotations in the target grounding stage, thereby enhancing the learning of query representations. 3) [(c) v.s. (d)]: Increasing the number of segmentation queries can further improve the robustness of the decoder when provided with different queries and required to segment the same box. 4) [(a), (b), (d), and (e)]: Combining the Triple Alignment and Multi-layer Multi-task Encoder-Decoder can effectively enhance the overall performance, achieving the optimal result. 5) [(e) v.s. (f)]: Even if we include the Triple Alignment supporting multimodal communication, it remains necessary for the subsequent encoder to update the unimodal features generated by the backbones into multimodal features. 6) [(e) v.s. (g)]: It is necessary to involve queries in Triple Alignment to transfer the data-agnostic embedding into data-related queries. Otherwise, using only Bidirectional Alignment (BA) for text and vision tokens, similar to approaches like Deep Fusion in GLIP [18] and Grounding DINO [24], and WPA in CoupAlign [56], causes a noticeable decline " + }, + { + "bbox": [ + 130, + 241, + 482, + 540 + ], + "type": "inline_equation", + "content": "(-0.92\\%)" + }, + { + "bbox": [ + 130, + 241, + 482, + 540 + ], + "type": "text", + "content": ". 7) [(d), (e) and (g)]: The slight gain of (g)(76.37%) over (d)(76.21%) stems from our MMDecoder's pixel-level signals, which already boosts BA in the Encoder. Thus, extra BA effort in the backbone is marginal, failing to solve the unaligned query issue. Instead, Tri-Align (e)(77.29%) can solve this issue, showing its novelty. Notably, (d)(76.21%), with a basic encoder-decoder, already achieves SOTA performance, emphasizing our bbox2seg paradigm's simplicity and effectiveness. 8) [(e) v.s. (h)]: The segmentation output can further derive the confidence score of the model's prediction, which is transformed into the Focal loss factor to adaptively scale the other losses to focus more attention on challenging cases." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 541, + 482, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 541, + 482, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 541, + 482, + 601 + ], + "type": "text", + "content": "We further conduct a more detailed abaltion study about the improvement and corresponding cost of adding more segmentation queries of Tab.3(c)-(d). As shown in Table 4, the best performance is observed with five segmentation queries. Adding more than that increases the burden from pixel-level constraint without benefits, also increasing computational cost per accuracy improvement." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 620, + 293, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 620, + 293, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 620, + 293, + 632 + ], + "type": "text", + "content": "4.5 Triple Alignment Analysis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "In addition to the improvement results in ablation study, we further deepen our understanding of the Triple Alignment by analyzing the attention behavior." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 147, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 147, + 91, + 447, + 103 + ], + "type": "text", + "content": "Transferring Object Bounding Box to Segmentation for Visual Grounding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 171, + 135, + 439, + 183 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 126 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 126 + ], + "type": "text", + "content": "Table 5: Attention values from query to text-referred visual region in Triple Alignment." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 171, + 135, + 439, + 183 + ], + "lines": [ + { + "bbox": [ + 171, + 135, + 439, + 183 + ], + "spans": [ + { + "bbox": [ + 171, + 135, + 439, + 183 + ], + "type": "table", + "html": "
LayerRefCOCORefCOCO+RefCOCOgReferItGame test
valtestAtestBvaltestAtestBval-gval-utest-utest
layer217%18%16%19%21%18%20%18%17%25%
layer434%36%33%29%31%28%27%33%33%30%
layer661%63%60%55%58%50%55%59%59%38%
", + "image_path": "dd687706b187069271b3a5aeddfed66851fd7c9ab7745d2b81045bcab5d0871b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 146, + 236, + 290, + 266 + ], + "blocks": [ + { + "bbox": [ + 132, + 193, + 301, + 226 + ], + "lines": [ + { + "bbox": [ + 132, + 193, + 301, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 193, + 301, + 226 + ], + "type": "text", + "content": "Table 6: Comparison with alternative method using RIS to provide pseudo segmentation labels for supervision." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 146, + 236, + 290, + 266 + ], + "lines": [ + { + "bbox": [ + 146, + 236, + 290, + 266 + ], + "spans": [ + { + "bbox": [ + 146, + 236, + 290, + 266 + ], + "type": "table", + "html": "
ReferItGameSegVG+RIS (LAVT)SegVG
val74.9176.85
test73.7675.06
", + "image_path": "1ebe678e72ac204077307d846409e5669211539792f79917bf61117aa2c64dbf.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 326, + 239, + 470, + 262 + ], + "blocks": [ + { + "bbox": [ + 312, + 198, + 481, + 230 + ], + "lines": [ + { + "bbox": [ + 312, + 198, + 481, + 230 + ], + "spans": [ + { + "bbox": [ + 312, + 198, + 481, + 230 + ], + "type": "text", + "content": "Table 7: AP50 using different segmentation queries for confidence score calculation on RefCOCOg test-u." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 326, + 239, + 470, + 262 + ], + "lines": [ + { + "bbox": [ + 326, + 239, + 470, + 262 + ], + "spans": [ + { + "bbox": [ + 326, + 239, + 470, + 262 + ], + "type": "table", + "html": "
Seg Query1st2nd3rd4th5th
AP5084.6584.5784.7384.7884.63
", + "image_path": "43c9c28742868b801d5c3dc12fcbb2b3f044ec8e40686f5ec7f2c5273e9dae18.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 289, + 482, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 397 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 397 + ], + "type": "text", + "content": "Specifically, we calculate the sum of the attention values from the query to the text-referred visual region (target bbox) as a percentage of the total attention (which includes attention to the query, text, and visual tokens) to illustrate the extent of alignment across these three modalities in the second, fourth, and final layer of Triple Alignment. We average the percentage across all the attention heads and queries, and perform the analysis across all the datasets. As shown in Table 5, in all the datasets, the attention values increase as the layer progresses, indicating that Triple Alignment progressively aligns the query to comprehend the text and then focus on the referred visual region." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 415, + 347, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 415, + 347, + 427 + ], + "spans": [ + { + "bbox": [ + 132, + 415, + 347, + 427 + ], + "type": "text", + "content": "4.6 Comparison with alternative method" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 435, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 435, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 435, + 482, + 544 + ], + "type": "text", + "content": "Given the development of Referring Expression Segmentation (RES), a natural alternative method would be using a RES method to generate pseudo segmentation labels to substitute our bbox2seg paradigm. Therefore, to mimic real-world scenarios, we use LAVT [46] trained on RefCOCO to obtain pseudo segmentation labels on ReferItGame. We follow the same training setting in ablation study to conduct the comparison on ReferItGame. As shown in Table 6, our SegVG outperforms the alternative method. This demonstrates that our bbox2seg paradigm is more effective than using a RES model to provide pseudo segmentation labels which might suffer from the errors from the RES model." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 561, + 293, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 561, + 293, + 573 + ], + "spans": [ + { + "bbox": [ + 132, + 561, + 293, + 573 + ], + "type": "text", + "content": "4.7 Confidence Score Analysis" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 581, + 481, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 581, + 481, + 629 + ], + "spans": [ + { + "bbox": [ + 131, + 581, + 481, + 629 + ], + "type": "text", + "content": "Selection of segmentation query To show the effect of different selection of segmentation query for the calculation of confidence score, we calculate AP50 on RefCOCOg test-u. As shown in Table 7, the performance variations are slight among them. We use the first one to calculate confidence score for simplicity." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": "Confidence Score Faithfulness To evaluate the faithfulness of our confidence score, i.e., whether a higher confidence score indicates better performance, we" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "W. Kang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 294, + 228 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 294, + 228 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 294, + 228 + ], + "type": "image", + "image_path": "60e14df3e9978e3090ace5b2d575b2626131b34f201cadfef4ea883ab81a9dc4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 235, + 294, + 257 + ], + "lines": [ + { + "bbox": [ + 132, + 235, + 294, + 257 + ], + "spans": [ + { + "bbox": [ + 132, + 235, + 294, + 257 + ], + "type": "text", + "content": "Fig. 3: IoU and Accuracy of different confidence scores on RefCOCOg test-u." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 324, + 169, + 481, + 236 + ], + "blocks": [ + { + "bbox": [ + 318, + 146, + 482, + 168 + ], + "lines": [ + { + "bbox": [ + 318, + 146, + 482, + 168 + ], + "spans": [ + { + "bbox": [ + 318, + 146, + 482, + 168 + ], + "type": "text", + "content": "Table 8: Performance of different confidence levels on RefCOCOg test-u." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 324, + 169, + 481, + 236 + ], + "lines": [ + { + "bbox": [ + 324, + 169, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 324, + 169, + 481, + 236 + ], + "type": "table", + "html": "
ConfidenceIoUAcc(%)Proportion (%)
≥ 0.650.706777.41%100.00%
≥ 0.700.707277.48%99.89%
≥ 0.750.709177.69%99.44%
≥ 0.800.714378.37%97.98%
≥ 0.850.722679.29%95.04%
", + "image_path": "1a91b373bb385194d7a4bcc94bba2fb480539a1ac9fadf1589ff014082bc0c25.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 166, + 268, + 449, + 423 + ], + "blocks": [ + { + "bbox": [ + 166, + 268, + 449, + 423 + ], + "lines": [ + { + "bbox": [ + 166, + 268, + 449, + 423 + ], + "spans": [ + { + "bbox": [ + 166, + 268, + 449, + 423 + ], + "type": "image", + "image_path": "c91bfc47273881b5edc3e6de06845807920d290f47f514cf1edd7a5cb420201f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 431, + 480, + 454 + ], + "lines": [ + { + "bbox": [ + 130, + 431, + 480, + 454 + ], + "spans": [ + { + "bbox": [ + 130, + 431, + 480, + 454 + ], + "type": "text", + "content": "Fig. 4: Qualitative comparison between (c) (the first line) and (d) (the second line) of Table. 3. Red boxes are ground truth. Blue boxes are model predictions." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 487, + 482, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 482, + 546 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 482, + 546 + ], + "type": "text", + "content": "assess the relationship of our confidence score and model performance metrics (IoU and Accuracy) as shown in Fig 3. We sort the RefCOCOg-umd test set by confidence score, split it into five equal parts, and calculate each part's average score and performance. We observe a positive correlation between the performance metrics and our confidence score, which confirms its faithfulness." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "Confidence Score Application In real applications, the confidence score can be used to enhance the model's reliability. Specifically, we can apply different confidence thresholds to achieve different predictions, as shown in Table 8. First, we observe that accuracy increases with higher thresholds, indicating that adjusting the threshold can enhance the model's localization ability. Furthermore, the mean IoU also increases with the increasing threshold. Therefore, in downstream applications, such as using the model to provide pseudo-labels, we can increase the threshold to obtain a more accurate box. Since low-confidence outputs are excluded, the output proportion is slightly reduced, i.e., yields fewer outputs." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 147, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 147, + 91, + 447, + 102 + ], + "type": "text", + "content": "Transferring Object Bounding Box to Segmentation for Visual Grounding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 150, + 113, + 465, + 256 + ], + "blocks": [ + { + "bbox": [ + 150, + 113, + 465, + 256 + ], + "lines": [ + { + "bbox": [ + 150, + 113, + 465, + 256 + ], + "spans": [ + { + "bbox": [ + 150, + 113, + 465, + 256 + ], + "type": "image", + "image_path": "7adbdf7fdbc2d748611024734ad73347d1709ec5c78a791704222e5fa6d12ba1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 264, + 482, + 309 + ], + "lines": [ + { + "bbox": [ + 130, + 264, + 482, + 309 + ], + "spans": [ + { + "bbox": [ + 130, + 264, + 482, + 309 + ], + "type": "text", + "content": "Fig. 5: Layer " + }, + { + "bbox": [ + 130, + 264, + 482, + 309 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 264, + 482, + 309 + ], + "type": "text", + "content": " refers to the output of the " + }, + { + "bbox": [ + 130, + 264, + 482, + 309 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 264, + 482, + 309 + ], + "type": "text", + "content": "-th layer of the decoder. The blue boxes represent the models' predictions, while the red boxes denote the ground truth. In the segmentation mask shown in the second column from the right, red indicates high confidence for the foreground. Note that VLTVG does not provide segmentation output." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 333, + 257, + 345 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 333, + 257, + 345 + ], + "spans": [ + { + "bbox": [ + 132, + 333, + 257, + 345 + ], + "type": "text", + "content": "4.8 Qualitative Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 354, + 480, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 480, + 390 + ], + "type": "text", + "content": "We compare Tab. 3 (c) and (d) qualitatively to show the robustness enhancement. As shown in Fig. 4, adding segmentation queries improves the robustness to distinguish the target from the distractor, e.g. the case of two dogs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 391, + 482, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 391, + 482, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 482, + 559 + ], + "type": "text", + "content": "As depicted in Fig. 5, we compare the box prediction quality by each decoding layer of SegVG with VLTVG [43] which also involves multi-layer supervision. As seen in the upper two rows of Fig. 5, VLTVG initially misses the target \"young man\" but improves its prediction gradually and finally makes the correct prediction. In contrast, due to the full exploitation of the annotations and the domain alignment in our Triple Alignment, SegVG successfully identifies the location of the target in the early decoding layer and consistently makes the correct prediction in each layer. Another example can be observed in the lower two rows of Fig. 5. In this image, due to the complex colors, VLTVG fails to locate the target \"plate\" and consistently repeats the same mistake. Instead, SegVG correctly detects the target, even in the first decoder layer. Additionally, we visualize the segmentation mask obtained by SegVG in Fig. 5, which accurately identifies the target box with high confidence. This behavior aligns with the box regression, demonstrating their shared objective, i.e., to distinguish the box." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 578, + 220, + 591 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 220, + 591 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 220, + 591 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "We propose SegVG, a visual grounding model, where we iteratively make the best of box annotations to involve pixel-level supervision and address the domain discrepancy among queries, text, and vision by Triple Alignment module. Experiments show the superior performance of SegVG. Furthermore, we explore the reliability benefits of our segmentation output in real-world applications." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "W. Kang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 115, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 115, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 482, + 177 + ], + "type": "text", + "content": "This research is supported by NSF IIS-2309073 and ECCS-2123521. This article solely reflects the opinions and conclusions of its authors and not the funding agencies." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 197, + 197, + 209 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 197, + 197, + 209 + ], + "spans": [ + { + "bbox": [ + 133, + 197, + 197, + 209 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 223, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 138, + 223, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 223, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 138, + 223, + 480, + 267 + ], + "type": "text", + "content": "1. Anderson, P., He, X., Buehler, C., Teney, D., Johnson, M., Gould, S., Zhang, L.: Bottom-up and top-down attention for image captioning and visual question answering. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6077-6086 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 268, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 268, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 138, + 268, + 480, + 300 + ], + "type": "text", + "content": "2. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 301, + 480, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 301, + 480, + 334 + ], + "spans": [ + { + "bbox": [ + 138, + 301, + 480, + 334 + ], + "type": "text", + "content": "3. Chen, L., Ma, W., Xiao, J., Zhang, H., Chang, S.F.: Ref-nms: Breaking proposal bottlenecks in two-stage referring expression grounding. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 1036-1044 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 335, + 480, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 335, + 480, + 367 + ], + "spans": [ + { + "bbox": [ + 138, + 335, + 480, + 367 + ], + "type": "text", + "content": "4. Chen, S., Jin, Q., Wang, P., Wu, Q.: Say as you wish: Fine-grained control of image caption generation with abstract scene graphs. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9962-9971 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 368, + 480, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 368, + 480, + 399 + ], + "spans": [ + { + "bbox": [ + 138, + 368, + 480, + 399 + ], + "type": "text", + "content": "5. Chen, X., Ma, L., Chen, J., Jie, Z., Liu, W., Luo, J.: Real-time referring expression comprehension by single-stage grounding network. arXiv preprint arXiv:1812.03426 (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 401, + 480, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 401, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 138, + 401, + 480, + 434 + ], + "type": "text", + "content": "6. Deng, J., Yang, Z., Chen, T., Zhou, W., Li, H.: Transvg: End-to-end visual grounding with transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1769-1779 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 434, + 480, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 434, + 480, + 466 + ], + "spans": [ + { + "bbox": [ + 138, + 434, + 480, + 466 + ], + "type": "text", + "content": "7. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 467, + 480, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 467, + 480, + 510 + ], + "spans": [ + { + "bbox": [ + 138, + 467, + 480, + 510 + ], + "type": "text", + "content": "8. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 512, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 512, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 512, + 480, + 555 + ], + "type": "text", + "content": "9. Escalante, H.J., Hernández, C.A., Gonzalez, J.A., López-López, A., Montes, M., Morales, E.F., Sucar, L.E., Villasenor, L., Grubinger, M.: The segmented and annotated iapr tc-12 benchmark. Computer vision and image understanding 114(4), 419-428 (2010)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 555, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 555, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 134, + 555, + 480, + 599 + ], + "type": "text", + "content": "10. Gan, C., Li, Y., Li, H., Sun, C., Gong, B.: Vqs: Linking segmentations to questions and answers for supervised attention in vqa and question-focused semantic segmentation. In: Proceedings of the IEEE international conference on computer vision. pp. 1811-1820 (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 600, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 600, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 134, + 600, + 480, + 621 + ], + "type": "text", + "content": "11. He, K., Gkioxari, G., Dollár, P., Girshick, R.: Mask r-cnn. In: Proceedings of the IEEE international conference on computer vision. pp. 2961-2969 (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 622, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 622, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 622, + 480, + 665 + ], + "type": "text", + "content": "12. Hu, R., Rohrbach, M., Andreas, J., Darrell, T., Saenko, K.: Modeling relationships in referential expressions with compositional modular networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1115-1124 (2017)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 147, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 147, + 91, + 447, + 102 + ], + "type": "text", + "content": "Transferring Object Bounding Box to Segmentation for Visual Grounding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "13. Huang, B., Lian, D., Luo, W., Gao, S.: Look before you leap: Learning landmark features for one-stage visual grounding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16888-16897 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 481, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 481, + 181 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 481, + 181 + ], + "type": "text", + "content": "14. Kang, W., Qu, M., Kini, J., Wei, Y., Shah, M., Yan, Y.: Intent3d: 3d object detection in rgb-d scans based on human intention. arXiv preprint arXiv:2405.18295 (2024)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 181, + 481, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 181, + 481, + 203 + ], + "spans": [ + { + "bbox": [ + 133, + 181, + 481, + 203 + ], + "type": "text", + "content": "15. Kang, W., Qu, M., Wei, Y., Yan, Y.: Actress: Active retraining for semi-supervised visual grounding. arXiv preprint arXiv:2407.03251 (2024)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 203, + 481, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 203, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 133, + 203, + 481, + 224 + ], + "type": "text", + "content": "16. Kang, W., Zhou, L., Wu, J., Sun, C., Yan, Y.: Visual grounding with attention-driven constraint balancing. arXiv preprint arXiv:2407.03243 (2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 224, + 481, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 224, + 481, + 256 + ], + "spans": [ + { + "bbox": [ + 133, + 224, + 481, + 256 + ], + "type": "text", + "content": "17. Kazemzadeh, S., Ordonez, V., Matten, M., Berg, T.: Referitgame: Referring to objects in photographs of natural scenes. In: Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP). pp. 787-798 (2014)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 256, + 481, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 256, + 481, + 299 + ], + "spans": [ + { + "bbox": [ + 133, + 256, + 481, + 299 + ], + "type": "text", + "content": "18. Li, L.H., Zhang, P., Zhang, H., Yang, J., Li, C., Zhong, Y., Wang, L., Yuan, L., Zhang, L., Hwang, J.N., et al.: Grounded language-image pre-training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10965-10975 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 299, + 481, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 299, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 133, + 299, + 481, + 331 + ], + "type": "text", + "content": "19. Li, M., Sigal, L.: Referring transformer: A one-step approach to multi-task visual grounding. Advances in neural information processing systems 34, 19652-19664 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 331, + 481, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 331, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 133, + 331, + 481, + 374 + ], + "type": "text", + "content": "20. Liao, Y., Liu, S., Li, G., Wang, F., Chen, Y., Qian, C., Li, B.: A real-time cross-modality correlation filtering method for referring expression comprehension. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10880-10889 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 374, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 374, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 133, + 374, + 481, + 407 + ], + "type": "text", + "content": "21. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proceedings of the IEEE international conference on computer vision. pp. 2980-2988 (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 407, + 481, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 407, + 481, + 450 + ], + "spans": [ + { + "bbox": [ + 133, + 407, + 481, + 450 + ], + "type": "text", + "content": "22. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 450, + 481, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 450, + 481, + 482 + ], + "spans": [ + { + "bbox": [ + 133, + 450, + 481, + 482 + ], + "type": "text", + "content": "23. Liu, D., Zhang, H., Wu, F., Zha, Z.J.: Learning to assemble neural module tree networks for visual grounding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4673-4682 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 482, + 481, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 482, + 481, + 514 + ], + "spans": [ + { + "bbox": [ + 133, + 482, + 481, + 514 + ], + "type": "text", + "content": "24. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 514, + 481, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 514, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 133, + 514, + 481, + 557 + ], + "type": "text", + "content": "25. Luo, G., Zhou, Y., Sun, X., Cao, L., Wu, C., Deng, C., Ji, R.: Multi-task collaborative network for joint referring expression comprehension and segmentation. In: Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition. pp. 10034-10043 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 557, + 481, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 557, + 481, + 589 + ], + "spans": [ + { + "bbox": [ + 133, + 557, + 481, + 589 + ], + "type": "text", + "content": "26. Mao, J., Huang, J., Toshev, A., Camburu, O., Yuille, A.L., Murphy, K.: Generation and comprehension of unambiguous object descriptions. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 11-20 (2016)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 589, + 481, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 589, + 481, + 622 + ], + "spans": [ + { + "bbox": [ + 133, + 589, + 481, + 622 + ], + "type": "text", + "content": "27. Milletari, F., Navab, N., Ahmadi, S.A.: V-net: Fully convolutional neural networks for volumetric medical image segmentation. In: 2016 fourth international conference on 3D vision (3DV). pp. 565-571. IEEE (2016)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 622, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 622, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 622, + 481, + 665 + ], + "type": "text", + "content": "28. Nagaraja, V.K., Morariu, V.I., Davis, L.S.: Modeling context between objects for referring expression understanding. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14. pp. 792-807. Springer (2016)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "W. Kang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "29. Plummer, B.A., Kordas, P., Kiapour, M.H., Zheng, S., Piramuthu, R., Lazebnik, S.: Conditional image-text embedding networks. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 249-264 (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 194 + ], + "type": "text", + "content": "30. Plummer, B.A., Wang, L., Cervantes, C.M., Caicedo, J.C., Hockenmaier, J., Lazebnik, S.: Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In: Proceedings of the IEEE international conference on computer vision. pp. 2641-2649 (2015)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 237 + ], + "type": "text", + "content": "31. Qu, M., Wu, Y., Liu, W., Gong, Q., Liang, X., Russakovsky, O., Zhao, Y., Wei, Y.: Siri: A simple selective retraining mechanism for transformer-based visual grounding. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXV. pp. 546-562. Springer (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 238, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 480, + 258 + ], + "type": "text", + "content": "32. Redmon, J., Farhadi, A.: Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 480, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 304 + ], + "type": "text", + "content": "33. Rezatofighi, H., Tsoi, N., Gwak, J., Sadeghian, A., Reid, I., Savarese, S.: Generalized intersection over union: A metric and a loss for bounding box regression. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 658-666 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "type": "text", + "content": "34. Sadhu, A., Chen, K., Nevatia, R.: Zero-shot grounding of objects from natural language queries. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4694-4703 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "text", + "content": "35. Shang, Y., Cai, M., Xu, B., Lee, Y.J., Yan, Y.: Llava-prumerge: Adaptive token reduction for efficient large multimodal models. arXiv preprint arXiv:2403.15388 (2024)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 369, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 413 + ], + "type": "text", + "content": "36. Su, W., Miao, P., Dou, H., Wang, G., Qiao, L., Li, Z., Li, X.: Language adaptive weight generation for multi-task visual grounding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10857-10866 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 414, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 480, + 456 + ], + "type": "text", + "content": "37. Sun, J., Luo, G., Zhou, Y., Sun, X., Jiang, G., Wang, Z., Ji, R.: Refteacher: A strong baseline for semi-supervised referring expression comprehension. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19144-19154 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 457, + 480, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 489 + ], + "type": "text", + "content": "38. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 490, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 480, + 522 + ], + "type": "text", + "content": "39. Wang, L., Li, Y., Huang, J., Lazebnik, S.: Learning two-branch neural networks for image-text matching tasks. IEEE Transactions on Pattern Analysis and Machine Intelligence 41(2), 394-407 (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 523, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 480, + 566 + ], + "type": "text", + "content": "40. Wang, P., Wu, Q., Cao, J., Shen, C., Gao, L., Hengel, A.v.d.: Neighbourhood watch: Referring expression comprehension via language-guided graph attention networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1960-1968 (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 567, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 609 + ], + "type": "text", + "content": "41. Wang, X., Liu, Y., Shen, C., Ng, C.C., Luo, C., Jin, L., Chan, C.S., Hengel, A.v.d., Wang, L.: On the general value of evidence, and bilingual scene-text visual question answering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10126-10135 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "text", + "content": "42. Wang, Z., Lu, Y., Li, Q., Tao, X., Guo, Y., Gong, M., Liu, T.: Cris: Clip-driven referring image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11686-11695 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "type": "text", + "content": "43. Yang, L., Xu, Y., Yuan, C., Liu, W., Li, B., Hu, W.: Improving visual grounding with visual-linguistic verification and iterative reasoning. In: Proceedings of the" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 147, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 147, + 91, + 447, + 102 + ], + "type": "text", + "content": "Transferring Object Bounding Box to Segmentation for Visual Grounding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 654 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 147, + 116, + 481, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 116, + 481, + 138 + ], + "spans": [ + { + "bbox": [ + 147, + 116, + 481, + 138 + ], + "type": "text", + "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9499-9508 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 481, + 171 + ], + "type": "text", + "content": "44. Yang, S., Li, G., Yu, Y.: Dynamic graph attention for referring expression comprehension. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4644-4653 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 481, + 205 + ], + "type": "text", + "content": "45. Yang, S., Li, G., Yu, Y.: Graph-structured referring expression reasoning in the wild. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9952–9961 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 205, + 481, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 205, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 134, + 205, + 481, + 247 + ], + "type": "text", + "content": "46. Yang, Z., Wang, J., Tang, Y., Chen, K., Zhao, H., Torr, P.H.: Lavo: Language-aware vision transformer for referring image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18155-18165 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 249, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 249, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 133, + 249, + 481, + 281 + ], + "type": "text", + "content": "47. Yang, Z., Chen, T., Wang, L., Luo, J.: Improving one-stage visual grounding by recursive sub-query construction. In: European Conference on Computer Vision. pp. 387-404. Springer (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 282, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 282, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 133, + 282, + 481, + 314 + ], + "type": "text", + "content": "48. Yang, Z., Gong, B., Wang, L., Huang, W., Yu, D., Luo, J.: A fast and accurate one-stage approach to visual grounding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4683-4693 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 315, + 481, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 315, + 481, + 358 + ], + "spans": [ + { + "bbox": [ + 133, + 315, + 481, + 358 + ], + "type": "text", + "content": "49. Ye, J., Tian, J., Yan, M., Yang, X., Wang, X., Zhang, J., He, L., Lin, X.: Shifting more attention to visual backbone: Query-modulated refinement networks for end-to-end visual grounding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15502-15512 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 358, + 481, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 358, + 481, + 390 + ], + "spans": [ + { + "bbox": [ + 133, + 358, + 481, + 390 + ], + "type": "text", + "content": "50. You, Q., Jin, H., Wang, Z., Fang, C., Luo, J.: Image captioning with semantic attention. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4651-4659 (2016)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 391, + 481, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 391, + 481, + 434 + ], + "spans": [ + { + "bbox": [ + 133, + 391, + 481, + 434 + ], + "type": "text", + "content": "51. Yu, L., Lin, Z., Shen, X., Yang, J., Lu, X., Bansal, M., Berg, T.L.: Matttnet: Modular attention network for referring expression comprehension. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1307-1315 (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 435, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 435, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 133, + 435, + 481, + 479 + ], + "type": "text", + "content": "52. Yu, L., Poirson, P., Yang, S., Berg, A.C., Berg, T.L.: Modeling context in referring expressions. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14. pp. 69-85. Springer (2016)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 479, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 479, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 133, + 479, + 481, + 511 + ], + "type": "text", + "content": "53. Yu, Z., Yu, J., Xiang, C., Zhao, Z., Tian, Q., Tao, D.: Rethinking diversified and discriminative proposal generation for visual grounding. arXiv preprint arXiv:1805.03508 (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 512, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 512, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 133, + 512, + 481, + 544 + ], + "type": "text", + "content": "54. Zareian, A., Rosa, K.D., Hu, D.H., Chang, S.F.: Open-vocabulary object detection using captions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14393-14402 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 545, + 481, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 545, + 481, + 577 + ], + "spans": [ + { + "bbox": [ + 133, + 545, + 481, + 577 + ], + "type": "text", + "content": "55. Zhang, H., Niu, Y., Chang, S.F.: Grounding referring expressions in images by variational context. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4158-4166 (2018)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 578, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 578, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 133, + 578, + 481, + 609 + ], + "type": "text", + "content": "56. Zhang, Z., Zhu, Y., Liu, J., Liang, X., Ke, W.: Coupling word-pixel with sentence-mask alignments for referring image segmentation. arXiv preprint arXiv:2212.01769 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 610, + 481, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 610, + 481, + 654 + ], + "spans": [ + { + "bbox": [ + 133, + 610, + 481, + 654 + ], + "type": "text", + "content": "57. Zhuang, B., Wu, Q., Shen, C., Reid, I., Van Den Hengel, A.: Parallel attention: A unified framework for visual object discovery through dialogs and queries. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4252-4261 (2018)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "W. Kang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Segment and Recognize Anything at Any Granularity/ec47fa22-1632-444f-b911-9d6b0be51867_content_list.json b/2024/Segment and Recognize Anything at Any Granularity/ec47fa22-1632-444f-b911-9d6b0be51867_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d204ddb0c1dd2a548fb451e3506f78a8e625c3f4 --- /dev/null +++ b/2024/Segment and Recognize Anything at Any Granularity/ec47fa22-1632-444f-b911-9d6b0be51867_content_list.json @@ -0,0 +1,1584 @@ +[ + { + "type": "text", + "text": "Segment and Recognize Anything at Any Granularity", + "text_level": 1, + "bbox": [ + 264, + 140, + 736, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Feng Li\\* Hao Zhang\\* Peize Sun, Xueyan Zou, Shilong Liu, Chunyuan Li Jianwei Yang1, Lei Zhang2, Jianfeng Gao", + "bbox": [ + 251, + 210, + 754, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "§ HKUST ‡ Microsoft Research, Redmond † IDEA # HKU § UW-Madison Tsinghua * Core Contributor 1. Project Lead 2. Equal Advisory Contribution", + "bbox": [ + 228, + 244, + 766, + 275 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. In this work, we introduce Semantic-SAM, an augmented image segmentation foundation for segmenting and recognizing anything at desired granularities. Compared to the foundational segmentation model SAM [31], our model has two unique advantages: (i) granularity-controllability in that the model can produce segmentation masks at any desired granularities, from objects to parts to both; (ii) semantic-awareness in that the model simultaneously predicts semantic labels for masks at different granularities. To enable multi-granularity capabilities, we propose a multi-choice learning scheme, where each click point generates a set of masks at multiple levels of granularity, corresponding to a set of ground-truth masks. To achieve semantic awareness, we consolidate multiple datasets of different levels of granularity and train our model using decoupled object- and part-based tasks to facilitate knowledge sharing and transfer among different tasks. To the best of our knowledge, this work is the first attempt to jointly train a model on SA-1B, instance-level, and part-level segmentation datasets. Experimental results and visualizations demonstrate that our model successfully achieves the desired goals. Furthermore, we show that multi-task training using the segmentation task defined on SA-1B and other segmentation tasks (e.g., panoptic and part segmentation) leads to performance gains on all segmentation tasks. In particular, we achieve a new state-of-the-art in COCO panoptic segmentation 60.2 PQ by adding SAM data.", + "bbox": [ + 259, + 311, + 743, + 617 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 642, + 375, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The universal and interactive AI systems that follow human intents have shown their potential in natural language processing [46, 47] and controllable image generation [52, 66]. However, such a universal system for pixel-level image understanding remains less explored. We argue that a foundational segmentation model should possess the following important properties: granularity-controllability and semantic-awareness. Regardless of the specific image domain or prompt context, the model is capable of acquiring a versatile representation, predicting segmentation masks in multi-granularity, and understanding the semantic meaning behind each segmented region.", + "bbox": [ + 212, + 672, + 797, + 809 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Previous works [31,58,70] have strived for these properties but only achieved part of the goals. The main obstacles impeding the progress of such a universal", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2b430727df2cf1fd4619fbf2c34f3a4d0a905b8a4dd535ff64f5d280268ff9e6.jpg", + "image_caption": [ + "Fig. 1: Dataset distribution in the semantics-granularity space.remove" + ], + "image_footnote": [], + "bbox": [ + 334, + 157, + 656, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "image segmentation model can be attributed to the limited flexibility of model design and the lack of training data.", + "bbox": [ + 212, + 343, + 784, + 375 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Model Design. The existing image segmentation model architectures are dominated by the single-input-single-output pipeline that discards any ambiguity. While this pipeline is prevalent in both anchor-based CNN architectures [24] and query-based Transformer architectures [4, 11], and has demonstrated remarkable performance in semantic, instance, and panoptic segmentation tasks [30, 39, 68], it inherently restricts the model to predict multi-granularity segmentation masks in an end-to-end manner. Although clustering postprocessing techniques [13] can produce multiple masks for a single object query, they are neither efficient nor effective solutions for a granularity-aware segmentation model.", + "bbox": [ + 225, + 388, + 787, + 537 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Training Data. Scaling up segmentation datasets that possess both semantic-awareness and granularity-awareness is a costly endeavor. Existing generic detection and segmentation datasets such as MSCOCO [39] and Objects365 [53] offer large amounts of data and rich semantic annotations, but only at the object level. On the other hand, part segmentation datasets such as Pascal Part [9], PartImageNet [23], and PACO [49] provide more fine-grained semantic annotations, but their data volumes are limited. Recently, SAM [31] has successfully scaled up the multi-granularity mask data to millions of images, but it does not include semantic annotations. In order to achieve the dual objectives of semantic-awareness and granularity-abundance, there is a pressing need to unify model design and training on various data formats to facilitate knowledge transfer. However, the inherent differences in semantics and granularity across different datasets pose a significant challenge to joint training efforts.", + "bbox": [ + 225, + 539, + 800, + 750 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we introduce Semantic-SAM, a universal image segmentation model designed to enable segmenting and recognizing objects at any desired granularity. Given one click from a user, our model addresses the spatial ambiguity by predicting masks in multiple granularities, accompanied by semantic labels at both the object and part levels. As shown in Figure 2, our model gener-", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Authors Suppressed Due to Excessive Length", + "bbox": [ + 271, + 114, + 573, + 128 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/67b2b578dfbc0c79c7dcf2547ccdab60c900c5d2500206ad4be73bc15d5afadb.jpg", + "image_caption": [ + "Fig. 2: Our model is capable of dealing with various segmentation tasks including open-set and interactive segmentation. (a) Our model can accomplish instance, semantic, panoptic segmentation, and part segmentation. (b) Our model is able to output multilevel semantics in different granularities. The yellow star point on the left-most image is the user click. (c) We connect our model with an inpainting model to perform multilevel inpainting. The prompts are \"Spider-Man\" and \"BMW car\", respectively. Note that only one click is needed to produce the results in (b) and (c), respectively." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 787, + 388 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ates multi-level segmentation masks ranging from the person head to the whole truck.", + "bbox": [ + 212, + 531, + 784, + 560 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The multi-granularity capability is achieved through a multi-choice learning design [22, 37] incorporated into the decoder architecture. Each click point is represented with multiple queries, each containing a different level of embedding. These queries are trained to learn from all available ground-truth masks representing different granularities. To establish a correspondence between multiple masks and ground-truths, we employ a many-to-many matching scheme to ensure that a single click point could generate high-quality masks in multiple granularities.", + "bbox": [ + 212, + 563, + 787, + 683 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To accomplish semantic-awareness with a generalized capability, we introduce a decoupled approach for objects and parts classification, to allow knowledge transfer among rich object and part segmentation datasets. By leveraging a shared text encoder to encode both objects and parts independently, we adapt the loss function based on the data type. For instance, generic segmentation data lacks part classification loss, whereas SAM data does not include classification loss.", + "bbox": [ + 212, + 686, + 787, + 790 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To enrich semantics and granularity within our model, we consolidate seven datasets on three types of granularities, including generic segmentation of MSCOCO [39], Objects365 [53], ADE20k [68], part segmentation of PASCAL Part [9], PACO [49],", + "bbox": [ + 212, + 794, + 839, + 842 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Segment and Recognize Anything at Any Granularity", + "bbox": [ + 372, + 114, + 732, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a7e5a3a04f70df4b12adceb061fea3b19c41956a8a1a6dae4b9d646b735ecc53.jpg", + "image_caption": [ + "Fig. 3: Semantic-SAM is a universal segmentation framework that can take multiple types of segmentation data including generic, part, and class-agnostic segmentation data. The vision encoder is used to extract image features. The mask decoder can do both generic segmentation and promptable segmentation with various types of prompts. For point and box, we input them via anchor boxes to the mask decoder. Since there is an ambiguity of granularity for a point input, we duplicate each point 6 times and give them different levels of embeddings. The output masks of point prompts match with multiple GT masks of different granularities." + ], + "image_footnote": [], + "bbox": [ + 223, + 146, + 777, + 306 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "PartImagenet [23], and SA-1B [31]. Their data formats are reorganized to match our training objectives accordingly. After joint training, our model obtains a strong performance across a variety of datasets. Notably, we find that learning from interactive segmentation could improve generic and part segmentation. For example, by jointly training SA-1B promptable segmentation and COCO panoptic segmentation, we achieve a gain of 2.2 box AP and a gain of 1.3 mask AP. In addition, through comprehensive experiments, we demonstrate that our model can cover more segmentation granularities than SAM, outperforming 3.4 IoU with one point click.", + "bbox": [ + 212, + 452, + 789, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Semantic-SAM", + "text_level": 1, + "bbox": [ + 215, + 613, + 398, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.1 Mode Design", + "text_level": 1, + "bbox": [ + 215, + 646, + 372, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our Semantic-SAM follows MaskDINO [33] to employ a query-based mask decoder. To produce semantic-aware and multi-granularity masks, we exploit two types of queries: (i) generic queries as used in MaskDINO and (ii) spatial queries including point and box for granularity-controllable interactive segmentation. The overall pipeline is shown in Figure 3.", + "bbox": [ + 212, + 672, + 787, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We represent both point and box prompts into anchor boxes as a unified format. In particular, we convert user click point $(x,y)$ into an anchor box $(x,y,w,h)$ with small width $w$ and height $h$ , so that the anchor box can closely approximate the point. To capture different granularities of masks, each click is first encoded to position prompt and combined with $K$ different content prompts, each of which is represented as a learnable embedding for a given", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Authors Suppressed Due to Excessive Length", + "bbox": [ + 271, + 114, + 573, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/293359e4ab4ffd0cf73dfbe792229bbfa860280f86b3bf5a3ad215afe67a0014.jpg", + "image_caption": [ + "Fig. 4: Decoupled object and part classification. Given the same decoder outputs (right), two set of embeddings are utilized for object and part classification separately." + ], + "image_footnote": [], + "bbox": [ + 318, + 157, + 676, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "granularity level. As a result, we feed into the decoder $K$ query embeddings $\\mathbf{Q} = (\\mathbf{q}_1,\\dots ,\\mathbf{q}_K)$ for a click represented as an anchor box $\\mathbf{b} = (x,y,w,h)$ , and the $i$ -th query is formulated as", + "bbox": [ + 214, + 364, + 782, + 411 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {q} _ {i} = \\mathbf {q} _ {i} ^ {\\text {l e v e l}} + \\mathbf {q} ^ {\\text {a n c h o r}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 418, + 421, + 785, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{q}_i^{\\mathrm{level}}$ is the content embedding for granularity level $i$ , and $\\mathbf{q}^{\\mathrm{anchor}}$ is a position embedding derived from the sine encoding of the anchor box. Here we empirically choose $K = 6$ , considering there are at most 6 levels of masks per user click for the majority of images in SA-1B [31]. As for the box prompt, we simply use a single content embedding, i.e., $K = 1$ .", + "bbox": [ + 212, + 450, + 782, + 525 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Assuming that the output image feature from the vision encoder is $\\mathbf{F}$ , the mask decoder takes a set of queries and produces the outputs:", + "bbox": [ + 214, + 526, + 782, + 556 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {O} = \\operatorname {D e f o r m D e c} (\\mathbf {Q}, \\mathbf {F}) \\text {w i t h} \\mathbf {O} = \\left(\\mathbf {o} _ {1}, \\dots , \\mathbf {o} _ {K}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 568, + 784, + 585 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathsf{DeformDec}(\\cdot, \\cdot, \\cdot)$ is a deformable decoder that takes anchor box queries, and image features as input to output queried features. $\\mathbf{o}_i$ is the model output for the $i$ -th input query $\\mathbf{q}_i$ . Each $\\mathbf{o}_i = (\\mathbf{c}_i, \\mathbf{m}_i)$ consists of the predicted semantic category $\\mathbf{c}_i$ and mask $\\mathbf{m}_i$ , which are used to construct the concept recognition loss and mask prediction loss, respectively.", + "bbox": [ + 214, + 595, + 782, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2.2 Training", + "text_level": 1, + "bbox": [ + 215, + 694, + 333, + 709 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Segment any granularity. To endow the model with a multi-granularity segmentation ability, we propose a many-to-many matching method during training. We found that SAM fails to provide good multi-level segmentation results with a single click. Since SAM uses many-to-one matching [31] during training, the three SAM-predicted masks for each click point only match with one GT mask. This means that points located in masks of small levels cannot predict large masks with high quality according to our observation. In contrast, we fully leverage the structures in both data and algorithm to enable multi-level mask prediction with", + "bbox": [ + 212, + 719, + 782, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Segment and Recognize Anything at Any Granularity", + "bbox": [ + 372, + 114, + 730, + 130 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7045eae0504bb1930ebf07c469707350b1a5efb3e1a5e12eae41c5a2693b747b.jpg", + "image_caption": [ + "Fig. 5: Comparisons of different interactive learning strategies. a) One-to-one: traditional interactive segmentation models that focus on object-level, i.e., SEEM [70]. b) Many-to-one: multi-choice learning for single granularity, i.e., SAM [31]. c) Many-to-many: our Semantic-SAM enforces the model to predict all the possible granularities of a single click for more diverse segmentation. d) As a result, our model is able to output multiple masks in richer granularities." + ], + "image_footnote": [], + "bbox": [ + 222, + 143, + 777, + 229 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "a single click. First, we reorganized the data by clustering multiple GT masks of different levels sharing the same click. To allow multiple predictions of the same click to match with the GT masks, we employ the Hungarian algorithm to enable the many-to-many matching. The similarity matrix and scores vary based on the availability of different segmentation data components.", + "bbox": [ + 212, + 359, + 787, + 436 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For box input and generic segmentation, we follow existing methods. Specifically, to generate a mask from an input box, we follow a similar idea as in denoising training (DN) [33]. We add noises to ground-truth boxes to simulate inaccurate box inputs from users, and these noised boxes serve as spatial prompts for the decoder. The model is trained to reconstruct the original boxes and masks given noised boxes. For the content part of box prompts, we adopt a learnable token as a general prompt. Note that this is the only difference from DN, as DN uses ground-truth label embedding as the content prompts. For generic segmentation, we follow the same pipeline as in Mask DINO [33].", + "bbox": [ + 212, + 438, + 787, + 574 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Recognize anything. We train with various types of data with different semantic annotations, in which some contain object-level annotations (COCO), some contain both object and part-level annotations (Pascal Part), and SA-1B has no semantic annotations but contains masks of all semantic levels. Note that a large number of part concepts are shared across different objects, for example, head for all animals, we aim to transfer the part concept knowledge across objects trained with only object-level annotations in our joint training. To address this discrepancy between semantic annotations and better transfer semantics of different granularity, we propose to decouple object and part recognition. As shown in Figure 4, we utilize a shared text encoder to encode objects and parts, which are used to perform object and part segmentation separately.", + "bbox": [ + 212, + 595, + 787, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "While all types of segmentation data share a unified format, the loss varies for different data types. In Table 1, we summarize the loss items to construct the training objective in Semantic-SAM. It is the part-level data that bridges the gap to recognize semantic concepts between part and object levels, and it is the use of SAM data that bridges the gap to segment masks at any granularity.", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Authors Suppressed Due to Excessive Length", + "bbox": [ + 271, + 114, + 573, + 128 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f79b5a79211bc59f30be04c0bf3cd36b07dfc4e13572015902575642fafead3c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DataRecognize AnythingSegment at Any Granularity
PartObjectBox Mask #GT in Matching
SAM dataXX✓✓Many
Object-level dataX✓✓✓One
Part-level data✓✓✓✓One
", + "bbox": [ + 263, + 143, + 738, + 224 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/cafa63027f2350a62d6c713ac37cac7a8722469d1b5780ee5e418af3d9dd6520.jpg", + "table_caption": [ + "Table 1: The loss items to construct the training objective in Semantic-SAM. The four loss items are part classification, object classification, box loss, and mask loss, respectively. The last column indicates the number of ground-truth masks in the matching." + ], + "table_footnote": [], + "table_body": "
TypeData#ImagesSemantic ConceptGranularity Level
PartObjectPartWhole
Class-agnosticSA-1B11MXX✓✓
Object-levelObjects3651.7MX365X✓
COCO110KX133X✓
ADE20K20KX150X✓
Part-levelPACO-LVIS45K20175✓✓
PartImageNet16K1311✓✓
Pascal Part5K3020✓✓
", + "bbox": [ + 289, + 313, + 715, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2: The statistics of training data used in Semantic-SAM.", + "bbox": [ + 287, + 452, + 712, + 465 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Discussion. As shown in Figure 5, Semantic-SAM differs from previous interactive segmentation models in two aspects. Firstly, our model is trained with multi-choice learning to output all the possible segmentation masks with one click point. Secondly, our output granularities are richer to generate diverse segmentation masks.", + "bbox": [ + 212, + 521, + 784, + 595 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "2.3 Data Unification", + "text_level": 1, + "bbox": [ + 215, + 648, + 398, + 662 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In order for multi-level semantics, we include seven datasets that contain different granularity-level masks. The datasets are SA-1B [31], COCO panoptic [39], ADE20k panoptic [68], PASCAL part [44], PACO [49], PartImageNet [23], and Objects365 [53]. Within them, COCO and ADE20k panoptic datasets contain object-level masks and class labels. PASCAL part, PACO, and PartImageNet contain part-level masks and class labels. SA-1B contains up to 6-level masks without labels, while Objects365 contains abundant class labels for object-level instances. The details of these datasets are shown in Table 2. We further visualize the data distribution of different data types in Figure 1.", + "bbox": [ + 212, + 703, + 785, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Segment and Recognize Anything at Any Granularity", + "bbox": [ + 372, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9c20188c2d8c6c7c8b9877bbe5bce2243acead9448360ccac632f2de66ab58f1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTypeTraining DataPQmIoUAPAPsAPmAPI box mask
boxmaskboxmaskboxmask
Mask2Former (T) [11]Close-setCOCO53.263.246.143.3-----
X-Decoder (T) [69]Open-setCOCO+VL52.662.443.641.3-----
OpenSeed (T) [65]Open-setCOCO+O36555.463.851.247.134.527.454.350.466.2 66.8
Semantic-SAM (T) (ours)Open-setCOCO54.663.250.146.134.427.153.249.466.1 66.1
Semantic-SAM (T) (ours)Open-setCOCO+SAM55.263.452.3(+2.2)47.4(+1.3)36.1(+1.7)28.3(+1.2)55.6(+2.4)50.7(+1.3)67.3 66.2
OpenSeed (L) [65]Open-setCOCO+O36559.568.658.253.2-----
Oneformer (H) [26]Open-setCOCO+O36560.068.8-52.0-----
Semantic-SAM (L) (ours)Open-setCOCO+SAM60.268.957.553.042.935.461.256.971.9 71.4
", + "bbox": [ + 217, + 143, + 784, + 256 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3: Results for Semantic-SAM and other panoptic segmentation models on COCO val2017. Our model is jointly trained on COCO [8] and SA-1B [31] (1/10 data) and directly evaluates COCO.add model param size, hq-sam comparison, more datasets (ADE, cityscapes and compare with SEEM), more impressive visualization", + "bbox": [ + 215, + 257, + 785, + 311 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3 Experiments", + "text_level": 1, + "bbox": [ + 217, + 359, + 374, + 378 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 217, + 409, + 426, + 425 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Implementation Details. In our experiments, we train on SA-1b [31] with no semantic annotation, COCO [39] with object-level annotation and Pascal Part [15] with part-level annotation. We implement our model based on Mask DINO [33]. Mask DINO is a unified detection and segmentation framework that simultaneously predicts box and mask. We follow [33] to use 300 latent queries and nine decoder layers for all segmentation tasks. For the visual backbone, we adopt pre-trained Swin-T/L [41] by default. For the language backbone, we adopt the pre-trained base model in UniCL [62]. As SA-1B [31] dominates the data, during training, we first train on only SA-1B data. Then, we add object and part-level data to jointly train the three types of data. During training, the image resolution is $1024 \\times 1024$ for all data. We use AdamW [43] as the optimizer. We use large-scale jittering for object and part-level data and did not use data augmentations for SA-1B data, as SA-1B images are abundant. We set the learning rate to 0.0001, which is decayed at 0.9 and 0.95 fractions of the total number of steps by 10.", + "bbox": [ + 215, + 450, + 787, + 678 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Evaluation. We mainly evaluate two datasets, including COCO val2017 and a subset of SA-1B [31] with 1000 images. For evaluation metrics, we evaluate PQ and AP for generic and part segmentation datasets. For single-granularity interactive segmentation, we report 1-click mIoU (denoted as 1-IoU) under Point (Max) and Point (Oracle) settings. Max denotes we select the output mask with the maximum confidence score. Oracle denotes we select the output mask with the max IoU by calculating the IoU between the prediction and target mask. For multi-granularity interactive segmentation, we report 1-IoU@All Granularity that matches all the possible ground-truth masks for a single click to the multi-granularity predictions and then calculate the average IoU of all granularities.", + "bbox": [ + 215, + 688, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 116, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Authors Suppressed Due to Excessive Length", + "bbox": [ + 271, + 114, + 573, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.2 Semantic Segmentation of Anything", + "text_level": 1, + "bbox": [ + 215, + 146, + 558, + 162 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Generic Segmentation. As shown in Table 3, to validate the compatibility of multi-granularity interactive segmentation and generic segmentation, we jointly train with SA-1B [31] (1/10 data) and COCO panoptic segmentation. The result indicates that interactive segmentation with SAM can significantly help the instance-level detection and segmentation with a performance improvement of $+2.2$ AP on the box and $+1.3$ AP on the mask. Notably, OpenSeed [65] and Semantic-SAM are both based on Mask DINO [33]. Our joint training with SA-1B even outperforms OpenSeed which is trained with Object365 [53]. In addition, adding SA-1B mainly improves small object detection (APs and APm), as there are a large number of small objects in SA-1B.", + "bbox": [ + 212, + 170, + 787, + 321 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/31444f2e9451806b9b84273ebcbeeac29b1bfb9ba05a507c177d38b8e93edb90.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTypeTraining DataAPAPsAPmAP1
boxmaskbox maskbox maskbox mask
VLPart (B) [55]Open-setPascal Part29.227.4- -- -- -
Semantic-SAM (T) (ours)Open-setPascal Part27.030.516.6 19.138.1 41.643.8 49.1
Semantic-SAM (T) (ours)Open-setPascal Part+SAM28.0(+1.0)31.4(+0.9)17.3 19.940.0 42.545.7 49.7
", + "bbox": [ + 222, + 359, + 781, + 433 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 4: Results for Semantic-SAM and other part segmentation models on Pascal Part. Our model is jointly trained on Pascal Part [15] and SA-1B [31] (1/10 data) and directly evaluates Pascal Part.", + "bbox": [ + 215, + 434, + 784, + 474 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Part Segmentation. We also validate the compatibility of joint training SA-1B (1/10 data) and part segmentation. As shown in Table 4, adding SA-1B brings a decent performance improvement on Pascal Part [15], +1.0 AP on the box and +0.9 AP on the mask. Importantly, Semantic-SAM (T) has better mask AP than VLPart (B) [55], showing the high mask quality of our model.", + "bbox": [ + 212, + 510, + 785, + 588 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Single-granularity Interactive Segmentation. In Table 5, we evaluate 1-IoU for SAM and our model on COCO val2017. Our model outperforms SAM under both Point (Max) and Point (Oracle) settings.", + "bbox": [ + 212, + 599, + 785, + 646 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Multi-granularity Interactive Segmentation. In Table 6, we compare SAM [31] and our model on the output granularities for a single click. We adopt a Hungarian Matching to match all the possible target masks with the predicted masks for the click and calculate the average IoU score. As SAM has only three prompts, we also sample two clicks from a single mask to produce six output masks for a fair comparison. Notably, SAM has been trained on this validation set while we do not.", + "bbox": [ + 212, + 657, + 787, + 765 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.3 Ablations", + "text_level": 1, + "bbox": [ + 215, + 785, + 341, + 800 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Match Strategy. As shown in Table 7, we compare different match strategies in our model. When using many-to-many matching to match all the possible", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Segment and Recognize Anything at Any Granularity", + "bbox": [ + 372, + 114, + 730, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/785195d79969182c9fb90a45a5b78ad9e91a7c61afd8415534b9639b578eff11.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCOCO
Point (Max) Point (Oracle)
1-IoU1-IoU
SAM (B)52.168.2
SAM (L)55.770.5
Semantic-SAM (T)54.573.8
Semantic-SAM (L)57.074.2
", + "bbox": [ + 341, + 143, + 661, + 243 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/89aabf222456e1a50bbc09f56252c8407cdbd162811d7bd093cd1870b5dfbb00.jpg", + "table_caption": [ + "Table 5: Comparison with previous models on point interactions. Both SAM [31] and our model are trained on only SA-1B and directly evaluate on COCO val2017 for fair comparison. Max denotes we select the output with the max confidence score prediction. Oracle denotes we select the output with the max IoU by calculating the IoU between the prediction and target mask." + ], + "table_footnote": [], + "table_body": "
MethodGranularity1-IoU@All Granularity
SAM (B)†375.6
SAM (L)†382.5
SAM (H)†383.5
Semantic-SAM(T)688.1
Semantic-SAM(L)689.0
", + "bbox": [ + 318, + 339, + 683, + 426 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 6: Mask granularity comparison between SAM and our model. We evaluated both on an excluded subset of 1000 images from SA-1B. For each click, we evaluate all the possible ground-truth masks to calculate the 1-IoU@All Granularity. SAM [31] and Semantic-SAM adopts three and six prompts for a single click of a mask, respectively. $\\dagger$ denotes that SAM has been trained on this validation subset while we do not.", + "bbox": [ + 212, + 426, + 787, + 494 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "ground-truth masks for each click, the 1-IoU@All Granularity performance is significantly improved. This validates our matching strategy is effective in learning complete granularities.", + "bbox": [ + 212, + 539, + 785, + 585 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Box Interactive Evaluation. We also evaluate the 1-IoU given boxes in Table 8. We achieve better performance compared with object-level interactive segmentation model SEEM [70] and multi-granularity model SAM [31].", + "bbox": [ + 212, + 603, + 785, + 650 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Increasing SA-1B Training Data. In Table 9, we show the performance improvement on COCO val 2017 when training with more SA-1B data. The performance is saturated after using more than $15\\%$ of the total data. It indicates that we do not need to train with the whole SA-1B data to get a good zero-shot performance.", + "bbox": [ + 212, + 665, + 785, + 742 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "3.4 Visualization", + "text_level": 1, + "bbox": [ + 214, + 766, + 370, + 780 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We conducted a comparative analysis between our model and SAM to showcase the superior performance of our approach in generating high-quality masks, as depicted in Figure 6. The following key properties are highlighted:", + "bbox": [ + 212, + 794, + 785, + 842 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Authors Suppressed Due to Excessive Length", + "bbox": [ + 271, + 114, + 573, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/643e94acfc69cd02f471d4c3432be36c4269ff490b87e5a94c478416742b3c7b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodMatch1-IoU
Semantic-SAM(T)m2o73.2
Semantic-SAM(T)m2m88.1
", + "bbox": [ + 222, + 150, + 498, + 207 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/8cf06dc13ad3f0ed87898931b2ddb7e085654ec8ddd8ec43cb058e622231d3fe.jpg", + "table_caption": [ + "Table 7: Effect of match strategy on segmentation performance." + ], + "table_footnote": [], + "table_body": "
MethodBox 1-IoU
SAM [31](B)50.7
SEEM [70](T)73.7
Semantic-SAM(T)76.1
", + "bbox": [ + 531, + 143, + 782, + 215 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/3f3b0ae86ad7e1bae8a940ad787837c6455e8be297baddd70117267f853518e5.jpg", + "table_caption": [ + "Table 8: Box 1-IoU evaluation on COCO val2017." + ], + "table_footnote": [], + "table_body": "
MethodData Portion of SA-1BCOCO
Point (Max) Point (Oracle)
1-IoU1-IoU
SAM (L)100%55.770.5
Semantic-SAM (L)3%55.273.5
Semantic-SAM (L)15%56.773.6
Semantic-SAM (L)30%55.773.7
Semantic-SAM (L)50%55.373.9
Semantic-SAM (L)100%57.074.2
", + "bbox": [ + 277, + 266, + 725, + 386 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 9: Models trained with different portions of SA-1B [31], and directly evaluated on COCO val2017.", + "bbox": [ + 215, + 386, + 784, + 412 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/848f5cf19c3d946c1951de6cd2a2d92e473ace4d8e93f04af2922ede1bc78bd4.jpg", + "image_caption": [ + "Fig. 6: Given the user clicks shown in the first column, odd rows and even rows show the generated masks by SAM and our model, respectively. We removed the duplicated masks from our model to simplify the visualizations. More results are in supplementary." + ], + "image_footnote": [], + "bbox": [ + 225, + 446, + 774, + 734 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Segment and Recognize Anything at Any Granularity", + "bbox": [ + 372, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/03de6788d13da82b10ef4c0202c426e1c65f5fc9e2ec21ce9e3e4f311b5aa2e8.jpg", + "image_caption": [ + "Fig. 7: Our model learns an emergent capability that each level prompt corresponds to a certain granularity of mask. Given a single point click (red star) at the first column, our model produces masks with increasing areas from left to right for different images. A coherent order of granularities are observed for the prompts 1-6 across different images." + ], + "image_footnote": [], + "bbox": [ + 225, + 143, + 777, + 296 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Multi-Level Masks. Our model excels in producing masks with a greater range of meaningful granularities. Unlike SAM, which typically outputs a maximum of three masks with occasional duplication, our model generates a more diverse set of masks at various levels, enhancing the overall flexibility and precision of mask outputs.", + "- Mask Qualities. Our model has demonstrated its ability to generate masks of higher quality compared to SAM. SAM occasionally produces masks with artifacts such as holes or islands, particularly for large regions when the click falls within an interior region. In contrast, our model consistently generates high-quality masks across all levels, offering accurate segmentation results.", + "- Comparison with SA-1B Ground-truth Granularity. We extend the granularity of SAM data by producing masks with more meaningful levels of detail. Our model extends the original number of granularity levels, providing finer distinctions and more precise segmentation results on SAM data.", + "- Comparison with SAM in segmenting everything. We compare with SAM to segment everything in one image with the same number of points. We output more granularities for the image." + ], + "bbox": [ + 212, + 429, + 787, + 724 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Query semantics. We also find that each point content prompt embedding learns to correspond to a fixed granularity. As shown in Fig. 7, when we visualize masks in a specific order of the corresponding content embeddings, the masks follow the order from small to large in each row consistently. This proves that each content embedding represents a semantic granularity level in our model.", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Authors Suppressed Due to Excessive Length", + "bbox": [ + 271, + 114, + 573, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/3c4bdeb0fbcaf61596925ba4b964b57a4b80ab627b14449f40b83eda4b7d440a.jpg", + "image_caption": [ + "Fig. 8: Compare SAM and our model to segment everything in one image." + ], + "image_footnote": [], + "bbox": [ + 243, + 143, + 754, + 246 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4 Related works", + "text_level": 1, + "bbox": [ + 215, + 280, + 392, + 296 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.1 Generic Segmentation", + "text_level": 1, + "bbox": [ + 215, + 309, + 444, + 324 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Segmenting visual concepts is well-documented within the expansive field of computer vision [16,17,45,71]. Broad segmentation methodologies comprise several subdivisions, such as instance segmentation, semantic segmentation, and panoptic segmentation [6,24,30], each catering to a unique semantic degree. For example, semantic segmentation's goal is to detect and assign a label to each pixel in an image according to its corresponding semantic class [7,11,42]. Conversely, instance segmentation seeks to cluster pixels associated with the same semantic class into distinct object instances [3,24,33]. Panoptic segmentation is the hybrid of these two tasks. Recently, Transformer-based methods [4,56] have contributed to significant progress in segmentation tasks [11, 26, 33, 38, 64]. Generic object detection and segmentation have led to the development of abundant datasets, such as MSCOCO [39], LVIS [21], Objects365 [53], PASCAL [15], CityScapes [12], ADE20k [68].", + "bbox": [ + 212, + 330, + 787, + 529 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.2 Part Segmentation", + "text_level": 1, + "bbox": [ + 215, + 546, + 418, + 561 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Beyond generic segmentation, part segmentation aims to more fine-grained visual understanding. Most early works were bottom-up methods by grouping superpixels into parts and then objects [1, 2, 20]. Later, based on high-performance object detection networks [24, 51], top-down methods were developed by firstly detecting an object and then parsing it to part segmentation [27, 34, 63]. To segment the scene in multi-granularity, part-aware panoptic segmentation [13] is introduced. PPS [13] establishes the baseline by assembling panoptic and part segmentation models. JPPF [25] simplifies the model by a shared image encoder for both panoptic segmentation and part segmentation. By representing things, stuffs, and parts as object queries, Panoptic-PartFormer [35] proposes a unified architecture based on Transformer. While part segmentation data is much more expensive than object detection and segmentation data, a number of public datasets are available. Datasets for specific domains include cars [54], birds [57], and fashion [29]. General objects include Pascal-Part [9], PartImageNet [23], ADE20K [67], Cityscapes-Panoptic-Parts [44], and PACO [49]. More recently, SAM [31] provides a large-scale multi-granularity class-agnostic segmentation dataset. Our work is jointly trained on these datasets and contributes to a multi-granularity segmentation model.", + "bbox": [ + 212, + 568, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Segment and Recognize Anything at Any Granularity", + "bbox": [ + 372, + 114, + 732, + 130 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 127 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.3 Open-Vocabulary Segmentation", + "text_level": 1, + "bbox": [ + 215, + 146, + 522, + 162 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "While generic segmentation and part segmentation have made remarkable progress, they can only segment the image in a close-set vocabulary. To expand the vocabulary size, recent works leverage the visual-semantic knowledge from large-scale foundation models like CLIP [48], ALIGN [28] and Diffusion models [60] to various segmentation tasks. LSeg [32], OpenSeg [18], GroupViT [59] achieves open-vocabulary semantic segmentation ability on ADE20K and PASCAL. Dense-CLIP [50] and MaskCLIP [14] achieve open-vocabulary instance and panoptic segmentation on COCO dataset. More recently, X-Decoder [69] proposes a unified approach to tackle various segmentation and vision-language tasks for open-vocabulary segmentation, OpenSeeD [65] proposes to use a large amount of detection data and a joint training method to improve segmentation. To segment open-vocabulary masks in part-level, VLPart [55] leverages three part segmentation datasets and learns from the dense correspondence [5] between base objects and novel objects. Our work unifies these tasks into one architecture and builds up open-vocabulary segmentation in multi-granularity.", + "bbox": [ + 212, + 176, + 797, + 402 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "4.4 Interactive Segmentation", + "text_level": 1, + "bbox": [ + 215, + 429, + 470, + 445 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Interactive segmentation refers to the process of separating objects by actively integrating user inputs. This enduring challenge has seen notable advancements [10, 19, 31, 36, 40, 61]. Previous works only focus on a small set of data or semantic-agnostic instance masks. Recently, SAM [31] enlarges the training data from 0.12M COCO images to 10M SAM fine-grained images. SEEM [70] enriches the modality to language and function to both generic and grounded segmentation with an impressive composition. However, they focus either enriching the mask granularity or enhancing the semantic understanding. Our work is the first that reconciles both in a single unified model.", + "bbox": [ + 212, + 459, + 789, + 595 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 622, + 359, + 638 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we have presented Semantic-SAM, which can segment and recognize anything at desired granularities. Apart from performing generic open-vocabulary segmentation, Semantic-SAM demonstrates the advantages of semantic awareness and granularity abundance. To achieve such advantages, we propose improvements on data, model, and training where we utilize datasets from multiple granularity and semantic levels, multi-choice learning for training, and a universal framework for modeling. Comprehensive experiments and visualizations have verified the semantic awareness and granularity abundance are achieved in our model. Further, Semantic-SAM is the first successful attempt to jointly train on SA-1B and other classic segmentation datasets, and shows that training with SA-1B improves other tasks such as panoptic and part segmentation.", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Authors Suppressed Due to Excessive Length", + "bbox": [ + 271, + 114, + 573, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Pablo Arbelaez, Michael Maire, Charless Fowlkes, and Jitendra Malik. Contour detection and hierarchical image segmentation. IEEE transactions on pattern analysis and machine intelligence, 33(5):898-916, 2010.", + "2. Pablo Arbeláez, Jordi Pont-Tuset, Jonathan T Barron, Ferran Marques, and Jitendra Malik. Multiscale combinatorial grouping. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 328-335, 2014.", + "3. Daniel Bolya, Chong Zhou, Fanyi Xiao, and Yong Jae Lee. Yolact: Real-time instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9157-9166, 2019.", + "4. Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision, pages 213-229. Springer, 2020.", + "5. Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021.", + "6. Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834-848, 2017.", + "7. Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587. 2017.", + "8. Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dolkar, and C. Lawrence Zitnick. Microsoft COCO captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015.", + "9. Xianjie Chen, Roozbeh Mottaghi, Xiaobai Liu, Sanja Fidler, Raquel Urtasun, and Alan Yuille. Detect what you can: Detecting and representing objects using holistic models and body parts, 2014.", + "10. Xi Chen, Zhiyan Zhao, Yilei Zhang, Manni Duan, Donglian Qi, and Hengshuang Zhao. Focalclick: towards practical interactive image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1300-1309, 2022.", + "11. Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290-1299, 2022.", + "12. Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213-3223, 2016.", + "13. Daan de Geus, Panagiotis Meletis, Chenyang Lu, Xiaoxiao Wen, and Gijs Dubbelman. Part-aware panoptic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5485-5494, 2021.", + "14. Zheng Ding, Jieke Wang, and Zhuowen Tu. Open-vocabulary panoptic segmentation with maskclip. arXiv preprint arXiv:2208.08984, 2022.", + "15. Mark Everingham and John Winn. The pascal visual object classes challenge 2012 (voc2012) development kit. Pattern Analysis, Statistical Modelling and Computational Learning, Tech. Rep, 8(5), 2011." + ], + "bbox": [ + 218, + 172, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Segment and Recognize Anything at Any Granularity", + "bbox": [ + 372, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Pedro F Felzenszwalb, Ross B Girshick, David McAllester, and Deva Ramanan. Object detection with discriminatively trained part-based models. IEEE transactions on pattern analysis and machine intelligence, 32(9):1627-1645, 2009.", + "17. King-Sun Fu and JK Mui. A survey on image segmentation. Pattern recognition, 13(1):3-16, 1981.", + "18. Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Open-vocabulary image segmentation. arXiv preprint arXiv:2112.12143, 2021.", + "19. Leo Grady. Random walks for image segmentation. IEEE transactions on pattern analysis and machine intelligence, 28(11):1768-1783, 2006.", + "20. Matthias Grundmann, Vivek Kwatra, Mei Han, and Irfan Essa. Efficient hierarchical graph-based video segmentation. In 2010 IEEE computer society conference on computer vision and pattern recognition, pages 2141-2148. IEEE, 2010.", + "21. Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5356-5364, 2019.", + "22. Abner Guzman-Rivera, Dhruv Batra, and Pushmeet Kohli. Multiple choice learning: Learning to produce multiple structured outputs. Advances in neural information processing systems, 25, 2012.", + "23. Ju He, Shuo Yang, Shaokang Yang, Adam Kortylewski, Xiaoding Yuan, Jie-Neng Chen, Shuai Liu, Cheng Yang, and Alan Yuille. Partimagenet: A large, high-quality dataset of parts. arXiv preprint arXiv:2112.00933, 2021.", + "24. Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017.", + "25. Sravan Kumar Jagadeesh, René Schuster, and Didier Stricker. Multi-task fusion for efficient panoptic-part segmentation. arXiv preprint arXiv:2212.07671, 2022.", + "26. Jitesh Jain, Jiachen Li, MangTik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. Oneformer: One transformer to rule universal image segmentation. arXiv preprint arXiv:2211.06220, 2022.", + "27. Ruyi Ji, Dawei Du, Libo Zhang, Longyin Wen, Yanjun Wu, Chen Zhao, Feiyue Huang, and Siwei Lyu. Learning semantic neural tree for human parsing. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIII 16, pages 205-221. Springer, 2020.", + "28. Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V Le, Yunhsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021.", + "29. Menglin Jia, Mengyun Shi, Mikhail Sirotenko, Yin Cui, Claire Cardie, Bharath Hariharan, Hartwig Adam, and Serge Belongie. Fashionpedia: Ontology, segmentation, and an attribute localization dataset. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 316-332. Springer, 2020.", + "30. Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dólar. Panoptic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9404-9413, 2019.", + "31. Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything, 2023.", + "32. Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and René Ranftl. Language-driven semantic segmentation. arXiv preprint arXiv:2201.03546, 2022.", + "33. Feng Li, Hao Zhang, Shilong Liu, Lei Zhang, Lionel M Ni, Heung-Yeung Shum, et al. Mask dino: Towards a unified transformer-based framework for object detection and segmentation. arXiv preprint arXiv:2206.02777, 2022." + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Authors Suppressed Due to Excessive Length", + "bbox": [ + 271, + 114, + 573, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "34. Qizhu Li, Anurag Arnab, and Philip HS Torr. Holistic, instance-level human parsing. arXiv preprint arXiv:1709.03612, 2017.", + "35. Xiangtai Li, Shilin Xu, Yibo Yang, Guangliang Cheng, Yunhai Tong, and Dacheng Tao. Panoptic-partformer: Learning a unified model for panoptic part segmentation. In European Conference on Computer Vision, pages 729-747. Springer, 2022.", + "36. Yin Li, Jian Sun, Chi-Keung Tang, and Heung-Yeung Shum. Lazy snapping. ACM Transactions on Graphics (ToG), 23(3):303-308, 2004.", + "37. Zhuwen Li, Qifeng Chen, and Vladlen Koltun. Interactive image segmentation with latent diversity. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 577-585, 2018.", + "38. Zhiqi Li, Wenhai Wang, Enze Xie, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, Ping Luo, and Tong Lu. Panoptic segformer: Delving deeper into panoptic segmentation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1280-1289, 2022.", + "39. Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014.", + "40. Qin Liu, Zhenlin Xu, Gedas Bertasius, and Marc Niethammer. Simpleclick: Interactive image segmentation with simple vision transformers. arXiv preprint arXiv:2210.11006, 2022.", + "41. Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021.", + "42. Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3431-3440, 2015.", + "43. Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017.", + "44. Panagiotis Meletis, Xiaoxiao Wen, Chenyang Lu, Daan de Geus, and Gijs Dubbelman. Cityscapes-panoptic-parts and pascal-panoptic-parts datasets for scene understanding. arXiv preprint arXiv:2004.07944, 2020.", + "45. Shervin Minaee, Yuri Y Boykov, Fatih Porikli, Antonio J Plaza, Nasser Kehtarnavaz, and Demetri Terzopoulos. Image segmentation using deep learning: A survey. IEEE transactions on pattern analysis and machine intelligence, 2021.", + "46. OpenAI. Chatgpt. https://openai.com/blog/chatgpt, 2022.", + "47. OpenAI. Gpt-4 technical report, 2023.", + "48. Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021.", + "49. Vignesh Ramanathan, Anmol Kalia, Vladan Petrovic, Yi Wen, Baixue Zheng, Baishan Guo, Rui Wang, Aaron Marquez, Rama Kovvuri, Abhishek Kadian, Amir Mousavi, Yiwen Song, Abhimanyu Dubey, and Dhruv Mahajan. PACO: Parts and attributes of common objects. In arXiv preprint arXiv:2301.01795, 2023.", + "50. Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18082-18091, 2022.", + "51. Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28:91-99." + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Segment and Recognize Anything at Any Granularity", + "bbox": [ + 372, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "52. Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022.", + "53. Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8430-8439, 2019.", + "54. Xibin Song, Peng Wang, Dingfu Zhou, Rui Zhu, Chenye Guan, Yuchao Dai, Hao Su, Hongdong Li, and Ruigang Yang. Apollocar3d: A large 3d car instance understanding benchmark for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5452-5462, 2019.", + "55. Peize Sun, Shoufa Chen, Chenchen Zhu, Fanyi Xiao, Ping Luo, Saining Xie, and Zhicheng Yan. Going denser with open-vocabulary part segmentation, 2023.", + "56. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017.", + "57. Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. technical report, 2011.", + "58. Xinlong Wang, Xiaosong Zhang, Yue Cao, Wen Wang, Chunhua Shen, and Tiejun Huang. Seggpt: Segmenting everything in context. arXiv preprint arXiv:2304.03284, 2023.", + "59. Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, and Xiaolong Wang. Groupvit: Semantic segmentation emerges from text supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18134-18144, 2022.", + "60. Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. arXiv preprint arXiv:2303.04803, 2023.", + "61. Ning Xu, Brian Price, Scott Cohen, Jimei Yang, and Thomas S Huang. Deep interactive object selection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 373-381, 2016.", + "62. Jianwei Yang, Chunyuan Li, Pengchuan Zhang, Bin Xiao, Ce Liu, Lu Yuan, and Jianfeng Gao. Unified contrastive learning in image-text-label space. In CVPR, 2022.", + "63. Lu Yang, Qing Song, Zhihui Wang, and Ming Jiang. Parsing r-cnn for instance-level human analysis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 364-373, 2019.", + "64. Hao Zhang, Feng Li, Huaizhe Xu, Shijia Huang, Shilong Liu, Lionel M Ni, and Lei Zhang. Mp-former: Mask-piloted transformer for image segmentation. arXiv preprint arXiv:2303.07336, 2023.", + "65. Hao Zhang, Feng Li, Xueyan Zou, Shilong Liu, Chunyuan Li, Jianfeng Gao, Jianwei Yang, and Lei Zhang. A simple framework for open-vocabulary segmentation and detection. arXiv preprint arXiv:2303.08131, 2023.", + "66. Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543, 2023.", + "67. Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017.", + "68. Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset, 2018." + ], + "bbox": [ + 215, + 147, + 785, + 838 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Authors Suppressed Due to Excessive Length", + "bbox": [ + 271, + 114, + 573, + 128 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "69. Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. arXiv preprint arXiv:2212.11270, 2022.", + "70. Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. arXiv preprint arXiv:2304.06718, 2023.", + "71. Zhengxia Zou, Zhenwei Shi, Yuhong Guo, and Jieping Ye. Object detection in 20 years: A survey. arXiv preprint arXiv:1905.05055, 2019." + ], + "bbox": [ + 215, + 146, + 787, + 253 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "Segment and Recognize Anything at Any Granularity", + "bbox": [ + 372, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/Segment and Recognize Anything at Any Granularity/ec47fa22-1632-444f-b911-9d6b0be51867_model.json b/2024/Segment and Recognize Anything at Any Granularity/ec47fa22-1632-444f-b911-9d6b0be51867_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3c2a6a3fef9cc5d11c6b44d0f1345a3852b51264 --- /dev/null +++ b/2024/Segment and Recognize Anything at Any Granularity/ec47fa22-1632-444f-b911-9d6b0be51867_model.json @@ -0,0 +1,2405 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.266, + 0.141, + 0.737, + 0.187 + ], + "angle": 0, + "content": "Segment and Recognize Anything at Any Granularity" + }, + { + "type": "text", + "bbox": [ + 0.252, + 0.212, + 0.756, + 0.244 + ], + "angle": 0, + "content": "Feng Li\\* Hao Zhang\\* Peize Sun, Xueyan Zou, Shilong Liu, Chunyuan Li Jianwei Yang1, Lei Zhang2, Jianfeng Gao" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.245, + 0.767, + 0.276 + ], + "angle": 0, + "content": "§ HKUST ‡ Microsoft Research, Redmond † IDEA # HKU § UW-Madison Tsinghua * Core Contributor 1. Project Lead 2. Equal Advisory Contribution" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.312, + 0.744, + 0.618 + ], + "angle": 0, + "content": "Abstract. In this work, we introduce Semantic-SAM, an augmented image segmentation foundation for segmenting and recognizing anything at desired granularities. Compared to the foundational segmentation model SAM [31], our model has two unique advantages: (i) granularity-controllability in that the model can produce segmentation masks at any desired granularities, from objects to parts to both; (ii) semantic-awareness in that the model simultaneously predicts semantic labels for masks at different granularities. To enable multi-granularity capabilities, we propose a multi-choice learning scheme, where each click point generates a set of masks at multiple levels of granularity, corresponding to a set of ground-truth masks. To achieve semantic awareness, we consolidate multiple datasets of different levels of granularity and train our model using decoupled object- and part-based tasks to facilitate knowledge sharing and transfer among different tasks. To the best of our knowledge, this work is the first attempt to jointly train a model on SA-1B, instance-level, and part-level segmentation datasets. Experimental results and visualizations demonstrate that our model successfully achieves the desired goals. Furthermore, we show that multi-task training using the segmentation task defined on SA-1B and other segmentation tasks (e.g., panoptic and part segmentation) leads to performance gains on all segmentation tasks. In particular, we achieve a new state-of-the-art in COCO panoptic segmentation 60.2 PQ by adding SAM data." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.643, + 0.377, + 0.658 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.799, + 0.81 + ], + "angle": 0, + "content": "The universal and interactive AI systems that follow human intents have shown their potential in natural language processing [46, 47] and controllable image generation [52, 66]. However, such a universal system for pixel-level image understanding remains less explored. We argue that a foundational segmentation model should possess the following important properties: granularity-controllability and semantic-awareness. Regardless of the specific image domain or prompt context, the model is capable of acquiring a versatile representation, predicting segmentation masks in multi-granularity, and understanding the semantic meaning behind each segmented region." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Previous works [31,58,70] have strived for these properties but only achieved part of the goals. The main obstacles impeding the progress of such a universal" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.575, + 0.13 + ], + "angle": 0, + "content": "Authors Suppressed Due to Excessive Length" + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.158, + 0.658, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.266, + 0.3, + 0.737, + 0.314 + ], + "angle": 0, + "content": "Fig. 1: Dataset distribution in the semantics-granularity space.remove" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.344, + 0.785, + 0.375 + ], + "angle": 0, + "content": "image segmentation model can be attributed to the limited flexibility of model design and the lack of training data." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.389, + 0.789, + 0.539 + ], + "angle": 0, + "content": "- Model Design. The existing image segmentation model architectures are dominated by the single-input-single-output pipeline that discards any ambiguity. While this pipeline is prevalent in both anchor-based CNN architectures [24] and query-based Transformer architectures [4, 11], and has demonstrated remarkable performance in semantic, instance, and panoptic segmentation tasks [30, 39, 68], it inherently restricts the model to predict multi-granularity segmentation masks in an end-to-end manner. Although clustering postprocessing techniques [13] can produce multiple masks for a single object query, they are neither efficient nor effective solutions for a granularity-aware segmentation model." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.54, + 0.802, + 0.751 + ], + "angle": 0, + "content": "- Training Data. Scaling up segmentation datasets that possess both semantic-awareness and granularity-awareness is a costly endeavor. Existing generic detection and segmentation datasets such as MSCOCO [39] and Objects365 [53] offer large amounts of data and rich semantic annotations, but only at the object level. On the other hand, part segmentation datasets such as Pascal Part [9], PartImageNet [23], and PACO [49] provide more fine-grained semantic annotations, but their data volumes are limited. Recently, SAM [31] has successfully scaled up the multi-granularity mask data to millions of images, but it does not include semantic annotations. In order to achieve the dual objectives of semantic-awareness and granularity-abundance, there is a pressing need to unify model design and training on various data formats to facilitate knowledge transfer. However, the inherent differences in semantics and granularity across different datasets pose a significant challenge to joint training efforts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this paper, we introduce Semantic-SAM, a universal image segmentation model designed to enable segmenting and recognizing objects at any desired granularity. Given one click from a user, our model addresses the spatial ambiguity by predicting masks in multiple granularities, accompanied by semantic labels at both the object and part levels. As shown in Figure 2, our model gener-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.374, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Segment and Recognize Anything at Any Granularity" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.144, + 0.788, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.4, + 0.788, + 0.498 + ], + "angle": 0, + "content": "Fig. 2: Our model is capable of dealing with various segmentation tasks including open-set and interactive segmentation. (a) Our model can accomplish instance, semantic, panoptic segmentation, and part segmentation. (b) Our model is able to output multilevel semantics in different granularities. The yellow star point on the left-most image is the user click. (c) We connect our model with an inpainting model to perform multilevel inpainting. The prompts are \"Spider-Man\" and \"BMW car\", respectively. Note that only one click is needed to produce the results in (b) and (c), respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.532, + 0.785, + 0.561 + ], + "angle": 0, + "content": "ates multi-level segmentation masks ranging from the person head to the whole truck." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.564, + 0.788, + 0.684 + ], + "angle": 0, + "content": "The multi-granularity capability is achieved through a multi-choice learning design [22, 37] incorporated into the decoder architecture. Each click point is represented with multiple queries, each containing a different level of embedding. These queries are trained to learn from all available ground-truth masks representing different granularities. To establish a correspondence between multiple masks and ground-truths, we employ a many-to-many matching scheme to ensure that a single click point could generate high-quality masks in multiple granularities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.687, + 0.788, + 0.791 + ], + "angle": 0, + "content": "To accomplish semantic-awareness with a generalized capability, we introduce a decoupled approach for objects and parts classification, to allow knowledge transfer among rich object and part segmentation datasets. By leveraging a shared text encoder to encode both objects and parts independently, we adapt the loss function based on the data type. For instance, generic segmentation data lacks part classification loss, whereas SAM data does not include classification loss." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.84, + 0.843 + ], + "angle": 0, + "content": "To enrich semantics and granularity within our model, we consolidate seven datasets on three types of granularities, including generic segmentation of MSCOCO [39], Objects365 [53], ADE20k [68], part segmentation of PASCAL Part [9], PACO [49]," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.575, + 0.13 + ], + "angle": 0, + "content": "Authors Suppressed Due to Excessive Length" + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.147, + 0.778, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.316, + 0.789, + 0.427 + ], + "angle": 0, + "content": "Fig. 3: Semantic-SAM is a universal segmentation framework that can take multiple types of segmentation data including generic, part, and class-agnostic segmentation data. The vision encoder is used to extract image features. The mask decoder can do both generic segmentation and promptable segmentation with various types of prompts. For point and box, we input them via anchor boxes to the mask decoder. Since there is an ambiguity of granularity for a point input, we duplicate each point 6 times and give them different levels of embeddings. The output masks of point prompts match with multiple GT masks of different granularities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.453, + 0.79, + 0.59 + ], + "angle": 0, + "content": "PartImagenet [23], and SA-1B [31]. Their data formats are reorganized to match our training objectives accordingly. After joint training, our model obtains a strong performance across a variety of datasets. Notably, we find that learning from interactive segmentation could improve generic and part segmentation. For example, by jointly training SA-1B promptable segmentation and COCO panoptic segmentation, we achieve a gain of 2.2 box AP and a gain of 1.3 mask AP. In addition, through comprehensive experiments, we demonstrate that our model can cover more segmentation granularities than SAM, outperforming 3.4 IoU with one point click." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.614, + 0.4, + 0.63 + ], + "angle": 0, + "content": "2 Semantic-SAM" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.647, + 0.373, + 0.663 + ], + "angle": 0, + "content": "2.1 Mode Design" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.749 + ], + "angle": 0, + "content": "Our Semantic-SAM follows MaskDINO [33] to employ a query-based mask decoder. To produce semantic-aware and multi-granularity masks, we exploit two types of queries: (i) generic queries as used in MaskDINO and (ii) spatial queries including point and box for granularity-controllable interactive segmentation. The overall pipeline is shown in Figure 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We represent both point and box prompts into anchor boxes as a unified format. In particular, we convert user click point \\((x,y)\\) into an anchor box \\((x,y,w,h)\\) with small width \\(w\\) and height \\(h\\), so that the anchor box can closely approximate the point. To capture different granularities of masks, each click is first encoded to position prompt and combined with \\(K\\) different content prompts, each of which is represented as a learnable embedding for a given" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.373, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "Segment and Recognize Anything at Any Granularity" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.319, + 0.158, + 0.677, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.308, + 0.785, + 0.339 + ], + "angle": 0, + "content": "Fig. 4: Decoupled object and part classification. Given the same decoder outputs (right), two set of embeddings are utilized for object and part classification separately." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.366, + 0.784, + 0.412 + ], + "angle": 0, + "content": "granularity level. As a result, we feed into the decoder \\(K\\) query embeddings \\(\\mathbf{Q} = (\\mathbf{q}_1,\\dots ,\\mathbf{q}_K)\\) for a click represented as an anchor box \\(\\mathbf{b} = (x,y,w,h)\\), and the \\(i\\)-th query is formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.419, + 0.422, + 0.786, + 0.442 + ], + "angle": 0, + "content": "\\[\n\\mathbf {q} _ {i} = \\mathbf {q} _ {i} ^ {\\text {l e v e l}} + \\mathbf {q} ^ {\\text {a n c h o r}}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.451, + 0.784, + 0.526 + ], + "angle": 0, + "content": "where \\(\\mathbf{q}_i^{\\mathrm{level}}\\) is the content embedding for granularity level \\(i\\), and \\(\\mathbf{q}^{\\mathrm{anchor}}\\) is a position embedding derived from the sine encoding of the anchor box. Here we empirically choose \\(K = 6\\), considering there are at most 6 levels of masks per user click for the majority of images in SA-1B [31]. As for the box prompt, we simply use a single content embedding, i.e., \\(K = 1\\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.527, + 0.784, + 0.557 + ], + "angle": 0, + "content": "Assuming that the output image feature from the vision encoder is \\(\\mathbf{F}\\), the mask decoder takes a set of queries and produces the outputs:" + }, + { + "type": "equation", + "bbox": [ + 0.327, + 0.569, + 0.785, + 0.587 + ], + "angle": 0, + "content": "\\[\n\\mathbf {O} = \\operatorname {D e f o r m D e c} (\\mathbf {Q}, \\mathbf {F}) \\text {w i t h} \\mathbf {O} = \\left(\\mathbf {o} _ {1}, \\dots , \\mathbf {o} _ {K}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.597, + 0.784, + 0.673 + ], + "angle": 0, + "content": "where \\(\\mathsf{DeformDec}(\\cdot, \\cdot, \\cdot)\\) is a deformable decoder that takes anchor box queries, and image features as input to output queried features. \\(\\mathbf{o}_i\\) is the model output for the \\(i\\)-th input query \\(\\mathbf{q}_i\\). Each \\(\\mathbf{o}_i = (\\mathbf{c}_i, \\mathbf{m}_i)\\) consists of the predicted semantic category \\(\\mathbf{c}_i\\) and mask \\(\\mathbf{m}_i\\), which are used to construct the concept recognition loss and mask prediction loss, respectively." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.695, + 0.334, + 0.71 + ], + "angle": 0, + "content": "2.2 Training" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Segment any granularity. To endow the model with a multi-granularity segmentation ability, we propose a many-to-many matching method during training. We found that SAM fails to provide good multi-level segmentation results with a single click. Since SAM uses many-to-one matching [31] during training, the three SAM-predicted masks for each click point only match with one GT mask. This means that points located in masks of small levels cannot predict large masks with high quality according to our observation. In contrast, we fully leverage the structures in both data and algorithm to enable multi-level mask prediction with" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.575, + 0.13 + ], + "angle": 0, + "content": "Authors Suppressed Due to Excessive Length" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.145, + 0.778, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.244, + 0.788, + 0.327 + ], + "angle": 0, + "content": "Fig. 5: Comparisons of different interactive learning strategies. a) One-to-one: traditional interactive segmentation models that focus on object-level, i.e., SEEM [70]. b) Many-to-one: multi-choice learning for single granularity, i.e., SAM [31]. c) Many-to-many: our Semantic-SAM enforces the model to predict all the possible granularities of a single click for more diverse segmentation. d) As a result, our model is able to output multiple masks in richer granularities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.361, + 0.788, + 0.437 + ], + "angle": 0, + "content": "a single click. First, we reorganized the data by clustering multiple GT masks of different levels sharing the same click. To allow multiple predictions of the same click to match with the GT masks, we employ the Hungarian algorithm to enable the many-to-many matching. The similarity matrix and scores vary based on the availability of different segmentation data components." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.439, + 0.788, + 0.575 + ], + "angle": 0, + "content": "For box input and generic segmentation, we follow existing methods. Specifically, to generate a mask from an input box, we follow a similar idea as in denoising training (DN) [33]. We add noises to ground-truth boxes to simulate inaccurate box inputs from users, and these noised boxes serve as spatial prompts for the decoder. The model is trained to reconstruct the original boxes and masks given noised boxes. For the content part of box prompts, we adopt a learnable token as a general prompt. Note that this is the only difference from DN, as DN uses ground-truth label embedding as the content prompts. For generic segmentation, we follow the same pipeline as in Mask DINO [33]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.597, + 0.788, + 0.764 + ], + "angle": 0, + "content": "Recognize anything. We train with various types of data with different semantic annotations, in which some contain object-level annotations (COCO), some contain both object and part-level annotations (Pascal Part), and SA-1B has no semantic annotations but contains masks of all semantic levels. Note that a large number of part concepts are shared across different objects, for example, head for all animals, we aim to transfer the part concept knowledge across objects trained with only object-level annotations in our joint training. To address this discrepancy between semantic annotations and better transfer semantics of different granularity, we propose to decouple object and part recognition. As shown in Figure 4, we utilize a shared text encoder to encode objects and parts, which are used to perform object and part segmentation separately." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "While all types of segmentation data share a unified format, the loss varies for different data types. In Table 1, we summarize the loss items to construct the training objective in Semantic-SAM. It is the part-level data that bridges the gap to recognize semantic concepts between part and object levels, and it is the use of SAM data that bridges the gap to segment masks at any granularity." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.373, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Segment and Recognize Anything at Any Granularity" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "table", + "bbox": [ + 0.264, + 0.144, + 0.739, + 0.226 + ], + "angle": 0, + "content": "
DataRecognize AnythingSegment at Any Granularity
PartObjectBox Mask #GT in Matching
SAM dataXX✓✓Many
Object-level dataX✓✓✓One
Part-level data✓✓✓✓One
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.226, + 0.784, + 0.268 + ], + "angle": 0, + "content": "Table 1: The loss items to construct the training objective in Semantic-SAM. The four loss items are part classification, object classification, box loss, and mask loss, respectively. The last column indicates the number of ground-truth masks in the matching." + }, + { + "type": "table", + "bbox": [ + 0.29, + 0.314, + 0.716, + 0.452 + ], + "angle": 0, + "content": "
TypeData#ImagesSemantic ConceptGranularity Level
PartObjectPartWhole
Class-agnosticSA-1B11MXX✓✓
Object-levelObjects3651.7MX365X✓
COCO110KX133X✓
ADE20K20KX150X✓
Part-levelPACO-LVIS45K20175✓✓
PartImageNet16K1311✓✓
Pascal Part5K3020✓✓
" + }, + { + "type": "table_caption", + "bbox": [ + 0.289, + 0.453, + 0.714, + 0.466 + ], + "angle": 0, + "content": "Table 2: The statistics of training data used in Semantic-SAM." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.522, + 0.785, + 0.596 + ], + "angle": 0, + "content": "Discussion. As shown in Figure 5, Semantic-SAM differs from previous interactive segmentation models in two aspects. Firstly, our model is trained with multi-choice learning to output all the possible segmentation masks with one click point. Secondly, our output granularities are richer to generate diverse segmentation masks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.65, + 0.399, + 0.663 + ], + "angle": 0, + "content": "2.3 Data Unification" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.787, + 0.841 + ], + "angle": 0, + "content": "In order for multi-level semantics, we include seven datasets that contain different granularity-level masks. The datasets are SA-1B [31], COCO panoptic [39], ADE20k panoptic [68], PASCAL part [44], PACO [49], PartImageNet [23], and Objects365 [53]. Within them, COCO and ADE20k panoptic datasets contain object-level masks and class labels. PASCAL part, PACO, and PartImageNet contain part-level masks and class labels. SA-1B contains up to 6-level masks without labels, while Objects365 contains abundant class labels for object-level instances. The details of these datasets are shown in Table 2. We further visualize the data distribution of different data types in Figure 1." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.117, + 0.229, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.575, + 0.13 + ], + "angle": 0, + "content": "Authors Suppressed Due to Excessive Length" + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.144, + 0.785, + 0.257 + ], + "angle": 0, + "content": "
MethodTypeTraining DataPQmIoUAPAPsAPmAPI box mask
boxmaskboxmaskboxmask
Mask2Former (T) [11]Close-setCOCO53.263.246.143.3-----
X-Decoder (T) [69]Open-setCOCO+VL52.662.443.641.3-----
OpenSeed (T) [65]Open-setCOCO+O36555.463.851.247.134.527.454.350.466.2 66.8
Semantic-SAM (T) (ours)Open-setCOCO54.663.250.146.134.427.153.249.466.1 66.1
Semantic-SAM (T) (ours)Open-setCOCO+SAM55.263.452.3(+2.2)47.4(+1.3)36.1(+1.7)28.3(+1.2)55.6(+2.4)50.7(+1.3)67.3 66.2
OpenSeed (L) [65]Open-setCOCO+O36559.568.658.253.2-----
Oneformer (H) [26]Open-setCOCO+O36560.068.8-52.0-----
Semantic-SAM (L) (ours)Open-setCOCO+SAM60.268.957.553.042.935.461.256.971.9 71.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.258, + 0.787, + 0.313 + ], + "angle": 0, + "content": "Table 3: Results for Semantic-SAM and other panoptic segmentation models on COCO val2017. Our model is jointly trained on COCO [8] and SA-1B [31] (1/10 data) and directly evaluates COCO.add model param size, hq-sam comparison, more datasets (ADE, cityscapes and compare with SEEM), more impressive visualization" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.361, + 0.375, + 0.379 + ], + "angle": 0, + "content": "3 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.41, + 0.427, + 0.426 + ], + "angle": 0, + "content": "3.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.452, + 0.788, + 0.679 + ], + "angle": 0, + "content": "Implementation Details. In our experiments, we train on SA-1b [31] with no semantic annotation, COCO [39] with object-level annotation and Pascal Part [15] with part-level annotation. We implement our model based on Mask DINO [33]. Mask DINO is a unified detection and segmentation framework that simultaneously predicts box and mask. We follow [33] to use 300 latent queries and nine decoder layers for all segmentation tasks. For the visual backbone, we adopt pre-trained Swin-T/L [41] by default. For the language backbone, we adopt the pre-trained base model in UniCL [62]. As SA-1B [31] dominates the data, during training, we first train on only SA-1B data. Then, we add object and part-level data to jointly train the three types of data. During training, the image resolution is \\(1024 \\times 1024\\) for all data. We use AdamW [43] as the optimizer. We use large-scale jittering for object and part-level data and did not use data augmentations for SA-1B data, as SA-1B images are abundant. We set the learning rate to 0.0001, which is decayed at 0.9 and 0.95 fractions of the total number of steps by 10." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Evaluation. We mainly evaluate two datasets, including COCO val2017 and a subset of SA-1B [31] with 1000 images. For evaluation metrics, we evaluate PQ and AP for generic and part segmentation datasets. For single-granularity interactive segmentation, we report 1-click mIoU (denoted as 1-IoU) under Point (Max) and Point (Oracle) settings. Max denotes we select the output mask with the maximum confidence score. Oracle denotes we select the output mask with the max IoU by calculating the IoU between the prediction and target mask. For multi-granularity interactive segmentation, we report 1-IoU@All Granularity that matches all the possible ground-truth masks for a single click to the multi-granularity predictions and then calculate the average IoU of all granularities." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.373, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "Segment and Recognize Anything at Any Granularity" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.56, + 0.163 + ], + "angle": 0, + "content": "3.2 Semantic Segmentation of Anything" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.171, + 0.788, + 0.323 + ], + "angle": 0, + "content": "Generic Segmentation. As shown in Table 3, to validate the compatibility of multi-granularity interactive segmentation and generic segmentation, we jointly train with SA-1B [31] (1/10 data) and COCO panoptic segmentation. The result indicates that interactive segmentation with SAM can significantly help the instance-level detection and segmentation with a performance improvement of \\(+2.2\\) AP on the box and \\(+1.3\\) AP on the mask. Notably, OpenSeed [65] and Semantic-SAM are both based on Mask DINO [33]. Our joint training with SA-1B even outperforms OpenSeed which is trained with Object365 [53]. In addition, adding SA-1B mainly improves small object detection (APs and APm), as there are a large number of small objects in SA-1B." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.361, + 0.782, + 0.434 + ], + "angle": 0, + "content": "
MethodTypeTraining DataAPAPsAPmAP1
boxmaskbox maskbox maskbox mask
VLPart (B) [55]Open-setPascal Part29.227.4- -- -- -
Semantic-SAM (T) (ours)Open-setPascal Part27.030.516.6 19.138.1 41.643.8 49.1
Semantic-SAM (T) (ours)Open-setPascal Part+SAM28.0(+1.0)31.4(+0.9)17.3 19.940.0 42.545.7 49.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.435, + 0.785, + 0.476 + ], + "angle": 0, + "content": "Table 4: Results for Semantic-SAM and other part segmentation models on Pascal Part. Our model is jointly trained on Pascal Part [15] and SA-1B [31] (1/10 data) and directly evaluates Pascal Part." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.511, + 0.787, + 0.589 + ], + "angle": 0, + "content": "Part Segmentation. We also validate the compatibility of joint training SA-1B (1/10 data) and part segmentation. As shown in Table 4, adding SA-1B brings a decent performance improvement on Pascal Part [15], +1.0 AP on the box and +0.9 AP on the mask. Importantly, Semantic-SAM (T) has better mask AP than VLPart (B) [55], showing the high mask quality of our model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.6, + 0.787, + 0.647 + ], + "angle": 0, + "content": "Single-granularity Interactive Segmentation. In Table 5, we evaluate 1-IoU for SAM and our model on COCO val2017. Our model outperforms SAM under both Point (Max) and Point (Oracle) settings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.766 + ], + "angle": 0, + "content": "Multi-granularity Interactive Segmentation. In Table 6, we compare SAM [31] and our model on the output granularities for a single click. We adopt a Hungarian Matching to match all the possible target masks with the predicted masks for the click and calculate the average IoU score. As SAM has only three prompts, we also sample two clicks from a single mask to produce six output masks for a fair comparison. Notably, SAM has been trained on this validation set while we do not." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.786, + 0.342, + 0.801 + ], + "angle": 0, + "content": "3.3 Ablations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Match Strategy. As shown in Table 7, we compare different match strategies in our model. When using many-to-many matching to match all the possible" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.575, + 0.129 + ], + "angle": 0, + "content": "Authors Suppressed Due to Excessive Length" + }, + { + "type": "table", + "bbox": [ + 0.342, + 0.144, + 0.662, + 0.244 + ], + "angle": 0, + "content": "
MethodCOCO
Point (Max) Point (Oracle)
1-IoU1-IoU
SAM (B)52.168.2
SAM (L)55.770.5
Semantic-SAM (T)54.573.8
Semantic-SAM (L)57.074.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.244, + 0.788, + 0.314 + ], + "angle": 0, + "content": "Table 5: Comparison with previous models on point interactions. Both SAM [31] and our model are trained on only SA-1B and directly evaluate on COCO val2017 for fair comparison. Max denotes we select the output with the max confidence score prediction. Oracle denotes we select the output with the max IoU by calculating the IoU between the prediction and target mask." + }, + { + "type": "table", + "bbox": [ + 0.319, + 0.34, + 0.684, + 0.428 + ], + "angle": 0, + "content": "
MethodGranularity1-IoU@All Granularity
SAM (B)†375.6
SAM (L)†382.5
SAM (H)†383.5
Semantic-SAM(T)688.1
Semantic-SAM(L)689.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.428, + 0.788, + 0.496 + ], + "angle": 0, + "content": "Table 6: Mask granularity comparison between SAM and our model. We evaluated both on an excluded subset of 1000 images from SA-1B. For each click, we evaluate all the possible ground-truth masks to calculate the 1-IoU@All Granularity. SAM [31] and Semantic-SAM adopts three and six prompts for a single click of a mask, respectively. \\(\\dagger\\) denotes that SAM has been trained on this validation subset while we do not." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.54, + 0.787, + 0.587 + ], + "angle": 0, + "content": "ground-truth masks for each click, the 1-IoU@All Granularity performance is significantly improved. This validates our matching strategy is effective in learning complete granularities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.604, + 0.787, + 0.651 + ], + "angle": 0, + "content": "Box Interactive Evaluation. We also evaluate the 1-IoU given boxes in Table 8. We achieve better performance compared with object-level interactive segmentation model SEEM [70] and multi-granularity model SAM [31]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.666, + 0.787, + 0.743 + ], + "angle": 0, + "content": "Increasing SA-1B Training Data. In Table 9, we show the performance improvement on COCO val 2017 when training with more SA-1B data. The performance is saturated after using more than \\(15\\%\\) of the total data. It indicates that we do not need to train with the whole SA-1B data to get a good zero-shot performance." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.767, + 0.371, + 0.781 + ], + "angle": 0, + "content": "3.4 Visualization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.843 + ], + "angle": 0, + "content": "We conducted a comparative analysis between our model and SAM to showcase the superior performance of our approach in generating high-quality masks, as depicted in Figure 6. The following key properties are highlighted:" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.373, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Segment and Recognize Anything at Any Granularity" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.151, + 0.499, + 0.208 + ], + "angle": 0, + "content": "
MethodMatch1-IoU
Semantic-SAM(T)m2o73.2
Semantic-SAM(T)m2m88.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.208, + 0.502, + 0.235 + ], + "angle": 0, + "content": "Table 7: Effect of match strategy on segmentation performance." + }, + { + "type": "table", + "bbox": [ + 0.532, + 0.144, + 0.783, + 0.216 + ], + "angle": 0, + "content": "
MethodBox 1-IoU
SAM [31](B)50.7
SEEM [70](T)73.7
Semantic-SAM(T)76.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.528, + 0.216, + 0.786, + 0.241 + ], + "angle": 0, + "content": "Table 8: Box 1-IoU evaluation on COCO val2017." + }, + { + "type": "table", + "bbox": [ + 0.278, + 0.267, + 0.726, + 0.387 + ], + "angle": 0, + "content": "
MethodData Portion of SA-1BCOCO
Point (Max) Point (Oracle)
1-IoU1-IoU
SAM (L)100%55.770.5
Semantic-SAM (L)3%55.273.5
Semantic-SAM (L)15%56.773.6
Semantic-SAM (L)30%55.773.7
Semantic-SAM (L)50%55.373.9
Semantic-SAM (L)100%57.074.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.387, + 0.785, + 0.414 + ], + "angle": 0, + "content": "Table 9: Models trained with different portions of SA-1B [31], and directly evaluated on COCO val2017." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.448, + 0.775, + 0.735 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.756, + 0.788, + 0.799 + ], + "angle": 0, + "content": "Fig. 6: Given the user clicks shown in the first column, odd rows and even rows show the generated masks by SAM and our model, respectively. We removed the duplicated masks from our model to simplify the visualizations. More results are in supplementary." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.575, + 0.129 + ], + "angle": 0, + "content": "Authors Suppressed Due to Excessive Length" + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.144, + 0.778, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.31, + 0.788, + 0.381 + ], + "angle": 0, + "content": "Fig. 7: Our model learns an emergent capability that each level prompt corresponds to a certain granularity of mask. Given a single point click (red star) at the first column, our model produces masks with increasing areas from left to right for different images. A coherent order of granularities are observed for the prompts 1-6 across different images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.788, + 0.507 + ], + "angle": 0, + "content": "- Multi-Level Masks. Our model excels in producing masks with a greater range of meaningful granularities. Unlike SAM, which typically outputs a maximum of three masks with occasional duplication, our model generates a more diverse set of masks at various levels, enhancing the overall flexibility and precision of mask outputs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.518, + 0.788, + 0.595 + ], + "angle": 0, + "content": "- Mask Qualities. Our model has demonstrated its ability to generate masks of higher quality compared to SAM. SAM occasionally produces masks with artifacts such as holes or islands, particularly for large regions when the click falls within an interior region. In contrast, our model consistently generates high-quality masks across all levels, offering accurate segmentation results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.606, + 0.788, + 0.669 + ], + "angle": 0, + "content": "- Comparison with SA-1B Ground-truth Granularity. We extend the granularity of SAM data by producing masks with more meaningful levels of detail. Our model extends the original number of granularity levels, providing finer distinctions and more precise segmentation results on SAM data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.679, + 0.788, + 0.726 + ], + "angle": 0, + "content": "- Comparison with SAM in segmenting everything. We compare with SAM to segment everything in one image with the same number of points. We output more granularities for the image." + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.43, + 0.788, + 0.726 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Query semantics. We also find that each point content prompt embedding learns to correspond to a fixed granularity. As shown in Fig. 7, when we visualize masks in a specific order of the corresponding content embeddings, the masks follow the order from small to large in each row consistently. This proves that each content embedding represents a semantic granularity level in our model." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.373, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Segment and Recognize Anything at Any Granularity" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.787, + 0.128 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.144, + 0.756, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.25, + 0.253, + 0.751, + 0.267 + ], + "angle": 0, + "content": "Fig. 8: Compare SAM and our model to segment everything in one image." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.281, + 0.393, + 0.297 + ], + "angle": 0, + "content": "4 Related works" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.31, + 0.445, + 0.325 + ], + "angle": 0, + "content": "4.1 Generic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.331, + 0.789, + 0.53 + ], + "angle": 0, + "content": "Segmenting visual concepts is well-documented within the expansive field of computer vision [16,17,45,71]. Broad segmentation methodologies comprise several subdivisions, such as instance segmentation, semantic segmentation, and panoptic segmentation [6,24,30], each catering to a unique semantic degree. For example, semantic segmentation's goal is to detect and assign a label to each pixel in an image according to its corresponding semantic class [7,11,42]. Conversely, instance segmentation seeks to cluster pixels associated with the same semantic class into distinct object instances [3,24,33]. Panoptic segmentation is the hybrid of these two tasks. Recently, Transformer-based methods [4,56] have contributed to significant progress in segmentation tasks [11, 26, 33, 38, 64]. Generic object detection and segmentation have led to the development of abundant datasets, such as MSCOCO [39], LVIS [21], Objects365 [53], PASCAL [15], CityScapes [12], ADE20k [68]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.547, + 0.419, + 0.563 + ], + "angle": 0, + "content": "4.2 Part Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Beyond generic segmentation, part segmentation aims to more fine-grained visual understanding. Most early works were bottom-up methods by grouping superpixels into parts and then objects [1, 2, 20]. Later, based on high-performance object detection networks [24, 51], top-down methods were developed by firstly detecting an object and then parsing it to part segmentation [27, 34, 63]. To segment the scene in multi-granularity, part-aware panoptic segmentation [13] is introduced. PPS [13] establishes the baseline by assembling panoptic and part segmentation models. JPPF [25] simplifies the model by a shared image encoder for both panoptic segmentation and part segmentation. By representing things, stuffs, and parts as object queries, Panoptic-PartFormer [35] proposes a unified architecture based on Transformer. While part segmentation data is much more expensive than object detection and segmentation data, a number of public datasets are available. Datasets for specific domains include cars [54], birds [57], and fashion [29]. General objects include Pascal-Part [9], PartImageNet [23], ADE20K [67], Cityscapes-Panoptic-Parts [44], and PACO [49]. More recently, SAM [31] provides a large-scale multi-granularity class-agnostic segmentation dataset. Our work is jointly trained on these datasets and contributes to a multi-granularity segmentation model." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.575, + 0.129 + ], + "angle": 0, + "content": "Authors Suppressed Due to Excessive Length" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.147, + 0.524, + 0.163 + ], + "angle": 0, + "content": "4.3 Open-Vocabulary Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.177, + 0.799, + 0.404 + ], + "angle": 0, + "content": "While generic segmentation and part segmentation have made remarkable progress, they can only segment the image in a close-set vocabulary. To expand the vocabulary size, recent works leverage the visual-semantic knowledge from large-scale foundation models like CLIP [48], ALIGN [28] and Diffusion models [60] to various segmentation tasks. LSeg [32], OpenSeg [18], GroupViT [59] achieves open-vocabulary semantic segmentation ability on ADE20K and PASCAL. Dense-CLIP [50] and MaskCLIP [14] achieve open-vocabulary instance and panoptic segmentation on COCO dataset. More recently, X-Decoder [69] proposes a unified approach to tackle various segmentation and vision-language tasks for open-vocabulary segmentation, OpenSeeD [65] proposes to use a large amount of detection data and a joint training method to improve segmentation. To segment open-vocabulary masks in part-level, VLPart [55] leverages three part segmentation datasets and learns from the dense correspondence [5] between base objects and novel objects. Our work unifies these tasks into one architecture and builds up open-vocabulary segmentation in multi-granularity." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.43, + 0.471, + 0.446 + ], + "angle": 0, + "content": "4.4 Interactive Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.46, + 0.79, + 0.596 + ], + "angle": 0, + "content": "Interactive segmentation refers to the process of separating objects by actively integrating user inputs. This enduring challenge has seen notable advancements [10, 19, 31, 36, 40, 61]. Previous works only focus on a small set of data or semantic-agnostic instance masks. Recently, SAM [31] enlarges the training data from 0.12M COCO images to 10M SAM fine-grained images. SEEM [70] enriches the modality to language and function to both generic and grounded segmentation with an impressive composition. However, they focus either enriching the mask granularity or enhancing the semantic understanding. Our work is the first that reconciles both in a single unified model." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.623, + 0.36, + 0.64 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.789, + 0.841 + ], + "angle": 0, + "content": "In this paper, we have presented Semantic-SAM, which can segment and recognize anything at desired granularities. Apart from performing generic open-vocabulary segmentation, Semantic-SAM demonstrates the advantages of semantic awareness and granularity abundance. To achieve such advantages, we propose improvements on data, model, and training where we utilize datasets from multiple granularity and semantic levels, multi-choice learning for training, and a universal framework for modeling. Comprehensive experiments and visualizations have verified the semantic awareness and granularity abundance are achieved in our model. Further, Semantic-SAM is the first successful attempt to jointly train on SA-1B and other classic segmentation datasets, and shows that training with SA-1B improves other tasks such as panoptic and part segmentation." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.374, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Segment and Recognize Anything at Any Granularity" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.174, + 0.786, + 0.214 + ], + "angle": 0, + "content": "1. Pablo Arbelaez, Michael Maire, Charless Fowlkes, and Jitendra Malik. Contour detection and hierarchical image segmentation. IEEE transactions on pattern analysis and machine intelligence, 33(5):898-916, 2010." + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.215, + 0.787, + 0.254 + ], + "angle": 0, + "content": "2. Pablo Arbeláez, Jordi Pont-Tuset, Jonathan T Barron, Ferran Marques, and Jitendra Malik. Multiscale combinatorial grouping. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 328-335, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.255, + 0.786, + 0.293 + ], + "angle": 0, + "content": "3. Daniel Bolya, Chong Zhou, Fanyi Xiao, and Yong Jae Lee. Yolact: Real-time instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9157-9166, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.294, + 0.786, + 0.333 + ], + "angle": 0, + "content": "4. Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision, pages 213-229. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.334, + 0.786, + 0.386 + ], + "angle": 0, + "content": "5. Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.387, + 0.786, + 0.44 + ], + "angle": 0, + "content": "6. Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834-848, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.441, + 0.786, + 0.479 + ], + "angle": 0, + "content": "7. Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587. 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.48, + 0.786, + 0.52 + ], + "angle": 0, + "content": "8. Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dolkar, and C. Lawrence Zitnick. Microsoft COCO captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.521, + 0.786, + 0.559 + ], + "angle": 0, + "content": "9. Xianjie Chen, Roozbeh Mottaghi, Xiaobai Liu, Sanja Fidler, Raquel Urtasun, and Alan Yuille. Detect what you can: Detecting and representing objects using holistic models and body parts, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.56, + 0.786, + 0.611 + ], + "angle": 0, + "content": "10. Xi Chen, Zhiyan Zhao, Yilei Zhang, Manni Duan, Donglian Qi, and Hengshuang Zhao. Focalclick: towards practical interactive image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1300-1309, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.612, + 0.786, + 0.665 + ], + "angle": 0, + "content": "11. Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290-1299, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.666, + 0.786, + 0.732 + ], + "angle": 0, + "content": "12. Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213-3223, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.733, + 0.786, + 0.773 + ], + "angle": 0, + "content": "13. Daan de Geus, Panagiotis Meletis, Chenyang Lu, Xiaoxiao Wen, and Gijs Dubbelman. Part-aware panoptic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5485-5494, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.774, + 0.786, + 0.799 + ], + "angle": 0, + "content": "14. Zheng Ding, Jieke Wang, and Zhuowen Tu. Open-vocabulary panoptic segmentation with maskclip. arXiv preprint arXiv:2208.08984, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.799, + 0.786, + 0.84 + ], + "angle": 0, + "content": "15. Mark Everingham and John Winn. The pascal visual object classes challenge 2012 (voc2012) development kit. Pattern Analysis, Statistical Modelling and Computational Learning, Tech. Rep, 8(5), 2011." + }, + { + "type": "list", + "bbox": [ + 0.219, + 0.174, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.574, + 0.129 + ], + "angle": 0, + "content": "Authors Suppressed Due to Excessive Length" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.188 + ], + "angle": 0, + "content": "16. Pedro F Felzenszwalb, Ross B Girshick, David McAllester, and Deva Ramanan. Object detection with discriminatively trained part-based models. IEEE transactions on pattern analysis and machine intelligence, 32(9):1627-1645, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.785, + 0.214 + ], + "angle": 0, + "content": "17. King-Sun Fu and JK Mui. A survey on image segmentation. Pattern recognition, 13(1):3-16, 1981." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.215, + 0.785, + 0.24 + ], + "angle": 0, + "content": "18. Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Open-vocabulary image segmentation. arXiv preprint arXiv:2112.12143, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.241, + 0.785, + 0.267 + ], + "angle": 0, + "content": "19. Leo Grady. Random walks for image segmentation. IEEE transactions on pattern analysis and machine intelligence, 28(11):1768-1783, 2006." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.267, + 0.785, + 0.307 + ], + "angle": 0, + "content": "20. Matthias Grundmann, Vivek Kwatra, Mei Han, and Irfan Essa. Efficient hierarchical graph-based video segmentation. In 2010 IEEE computer society conference on computer vision and pattern recognition, pages 2141-2148. IEEE, 2010." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.307, + 0.785, + 0.347 + ], + "angle": 0, + "content": "21. Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5356-5364, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.347, + 0.785, + 0.387 + ], + "angle": 0, + "content": "22. Abner Guzman-Rivera, Dhruv Batra, and Pushmeet Kohli. Multiple choice learning: Learning to produce multiple structured outputs. Advances in neural information processing systems, 25, 2012." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.387, + 0.785, + 0.426 + ], + "angle": 0, + "content": "23. Ju He, Shuo Yang, Shaokang Yang, Adam Kortylewski, Xiaoding Yuan, Jie-Neng Chen, Shuai Liu, Cheng Yang, and Alan Yuille. Partimagenet: A large, high-quality dataset of parts. arXiv preprint arXiv:2112.00933, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.426, + 0.785, + 0.466 + ], + "angle": 0, + "content": "24. Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.466, + 0.785, + 0.492 + ], + "angle": 0, + "content": "25. Sravan Kumar Jagadeesh, René Schuster, and Didier Stricker. Multi-task fusion for efficient panoptic-part segmentation. arXiv preprint arXiv:2212.07671, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.492, + 0.785, + 0.532 + ], + "angle": 0, + "content": "26. Jitesh Jain, Jiachen Li, MangTik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. Oneformer: One transformer to rule universal image segmentation. arXiv preprint arXiv:2211.06220, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.532, + 0.785, + 0.586 + ], + "angle": 0, + "content": "27. Ruyi Ji, Dawei Du, Libo Zhang, Longyin Wen, Yanjun Wu, Chen Zhao, Feiyue Huang, and Siwei Lyu. Learning semantic neural tree for human parsing. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIII 16, pages 205-221. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.586, + 0.785, + 0.625 + ], + "angle": 0, + "content": "28. Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V Le, Yunhsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.626, + 0.785, + 0.693 + ], + "angle": 0, + "content": "29. Menglin Jia, Mengyun Shi, Mikhail Sirotenko, Yin Cui, Claire Cardie, Bharath Hariharan, Hartwig Adam, and Serge Belongie. Fashionpedia: Ontology, segmentation, and an attribute localization dataset. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 316-332. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.693, + 0.785, + 0.733 + ], + "angle": 0, + "content": "30. Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dólar. Panoptic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9404-9413, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.733, + 0.785, + 0.772 + ], + "angle": 0, + "content": "31. Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.799 + ], + "angle": 0, + "content": "32. Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and René Ranftl. Language-driven semantic segmentation. arXiv preprint arXiv:2201.03546, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "33. Feng Li, Hao Zhang, Shilong Liu, Lei Zhang, Lionel M Ni, Heung-Yeung Shum, et al. Mask dino: Towards a unified transformer-based framework for object detection and segmentation. arXiv preprint arXiv:2206.02777, 2022." + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.374, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Segment and Recognize Anything at Any Granularity" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.174 + ], + "angle": 0, + "content": "34. Qizhu Li, Anurag Arnab, and Philip HS Torr. Holistic, instance-level human parsing. arXiv preprint arXiv:1709.03612, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.174, + 0.785, + 0.226 + ], + "angle": 0, + "content": "35. Xiangtai Li, Shilin Xu, Yibo Yang, Guangliang Cheng, Yunhai Tong, and Dacheng Tao. Panoptic-partformer: Learning a unified model for panoptic part segmentation. In European Conference on Computer Vision, pages 729-747. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.227, + 0.785, + 0.253 + ], + "angle": 0, + "content": "36. Yin Li, Jian Sun, Chi-Keung Tang, and Heung-Yeung Shum. Lazy snapping. ACM Transactions on Graphics (ToG), 23(3):303-308, 2004." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.253, + 0.785, + 0.292 + ], + "angle": 0, + "content": "37. Zhuwen Li, Qifeng Chen, and Vladlen Koltun. Interactive image segmentation with latent diversity. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 577-585, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.292, + 0.785, + 0.345 + ], + "angle": 0, + "content": "38. Zhiqi Li, Wenhai Wang, Enze Xie, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, Ping Luo, and Tong Lu. Panoptic segformer: Delving deeper into panoptic segmentation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1280-1289, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.345, + 0.785, + 0.383 + ], + "angle": 0, + "content": "39. Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.383, + 0.785, + 0.422 + ], + "angle": 0, + "content": "40. Qin Liu, Zhenlin Xu, Gedas Bertasius, and Marc Niethammer. Simpleclick: Interactive image segmentation with simple vision transformers. arXiv preprint arXiv:2210.11006, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.422, + 0.785, + 0.476 + ], + "angle": 0, + "content": "41. Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.476, + 0.785, + 0.515 + ], + "angle": 0, + "content": "42. Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3431-3440, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.515, + 0.785, + 0.54 + ], + "angle": 0, + "content": "43. Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.54, + 0.785, + 0.579 + ], + "angle": 0, + "content": "44. Panagiotis Meletis, Xiaoxiao Wen, Chenyang Lu, Daan de Geus, and Gijs Dubbelman. Cityscapes-panoptic-parts and pascal-panoptic-parts datasets for scene understanding. arXiv preprint arXiv:2004.07944, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.579, + 0.785, + 0.618 + ], + "angle": 0, + "content": "45. Shervin Minaee, Yuri Y Boykov, Fatih Porikli, Antonio J Plaza, Nasser Kehtarnavaz, and Demetri Terzopoulos. Image segmentation using deep learning: A survey. IEEE transactions on pattern analysis and machine intelligence, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.618, + 0.692, + 0.63 + ], + "angle": 0, + "content": "46. OpenAI. Chatgpt. https://openai.com/blog/chatgpt, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.63, + 0.501, + 0.641 + ], + "angle": 0, + "content": "47. OpenAI. Gpt-4 technical report, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.641, + 0.785, + 0.693 + ], + "angle": 0, + "content": "48. Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.693, + 0.785, + 0.747 + ], + "angle": 0, + "content": "49. Vignesh Ramanathan, Anmol Kalia, Vladan Petrovic, Yi Wen, Baixue Zheng, Baishan Guo, Rui Wang, Aaron Marquez, Rama Kovvuri, Abhishek Kadian, Amir Mousavi, Yiwen Song, Abhimanyu Dubey, and Dhruv Mahajan. PACO: Parts and attributes of common objects. In arXiv preprint arXiv:2301.01795, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.747, + 0.785, + 0.799 + ], + "angle": 0, + "content": "50. Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18082-18091, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "51. Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28:91-99." + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.574, + 0.129 + ], + "angle": 0, + "content": "Authors Suppressed Due to Excessive Length" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.787, + 0.202 + ], + "angle": 0, + "content": "52. Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.203, + 0.787, + 0.257 + ], + "angle": 0, + "content": "53. Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8430-8439, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.257, + 0.787, + 0.31 + ], + "angle": 0, + "content": "54. Xibin Song, Peng Wang, Dingfu Zhou, Rui Zhu, Chenye Guan, Yuchao Dai, Hao Su, Hongdong Li, and Ruigang Yang. Apollocar3d: A large 3d car instance understanding benchmark for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5452-5462, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.31, + 0.787, + 0.336 + ], + "angle": 0, + "content": "55. Peize Sun, Shoufa Chen, Chenchen Zhu, Fanyi Xiao, Ping Luo, Saining Xie, and Zhicheng Yan. Going denser with open-vocabulary part segmentation, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.336, + 0.787, + 0.376 + ], + "angle": 0, + "content": "56. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.376, + 0.787, + 0.401 + ], + "angle": 0, + "content": "57. Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. technical report, 2011." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.402, + 0.787, + 0.441 + ], + "angle": 0, + "content": "58. Xinlong Wang, Xiaosong Zhang, Yue Cao, Wen Wang, Chunhua Shen, and Tiejun Huang. Seggpt: Segmenting everything in context. arXiv preprint arXiv:2304.03284, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.441, + 0.787, + 0.495 + ], + "angle": 0, + "content": "59. Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, and Xiaolong Wang. Groupvit: Semantic segmentation emerges from text supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18134-18144, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.787, + 0.535 + ], + "angle": 0, + "content": "60. Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. arXiv preprint arXiv:2303.04803, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.535, + 0.787, + 0.575 + ], + "angle": 0, + "content": "61. Ning Xu, Brian Price, Scott Cohen, Jimei Yang, and Thomas S Huang. Deep interactive object selection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 373-381, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.575, + 0.787, + 0.613 + ], + "angle": 0, + "content": "62. Jianwei Yang, Chunyuan Li, Pengchuan Zhang, Bin Xiao, Ce Liu, Lu Yuan, and Jianfeng Gao. Unified contrastive learning in image-text-label space. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.614, + 0.787, + 0.654 + ], + "angle": 0, + "content": "63. Lu Yang, Qing Song, Zhihui Wang, and Ming Jiang. Parsing r-cnn for instance-level human analysis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 364-373, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.654, + 0.787, + 0.693 + ], + "angle": 0, + "content": "64. Hao Zhang, Feng Li, Huaizhe Xu, Shijia Huang, Shilong Liu, Lionel M Ni, and Lei Zhang. Mp-former: Mask-piloted transformer for image segmentation. arXiv preprint arXiv:2303.07336, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.693, + 0.787, + 0.734 + ], + "angle": 0, + "content": "65. Hao Zhang, Feng Li, Xueyan Zou, Shilong Liu, Chunyuan Li, Jianfeng Gao, Jianwei Yang, and Lei Zhang. A simple framework for open-vocabulary segmentation and detection. arXiv preprint arXiv:2303.08131, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.734, + 0.787, + 0.759 + ], + "angle": 0, + "content": "66. Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.787, + 0.8 + ], + "angle": 0, + "content": "67. Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.787, + 0.839 + ], + "angle": 0, + "content": "68. Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset, 2018." + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.839 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.373, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Segment and Recognize Anything at Any Granularity" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "69. Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. arXiv preprint arXiv:2212.11270, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.189, + 0.788, + 0.228 + ], + "angle": 0, + "content": "70. Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. arXiv preprint arXiv:2304.06718, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.228, + 0.787, + 0.255 + ], + "angle": 0, + "content": "71. Zhengxia Zou, Zhenwei Shi, Yuhong Guo, and Jieping Ye. Object detection in 20 years: A survey. arXiv preprint arXiv:1905.05055, 2019." + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.255 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Segment and Recognize Anything at Any Granularity/ec47fa22-1632-444f-b911-9d6b0be51867_origin.pdf b/2024/Segment and Recognize Anything at Any Granularity/ec47fa22-1632-444f-b911-9d6b0be51867_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ff8cd724eff919f91a40c0d7b0a608411cdb4419 --- /dev/null +++ b/2024/Segment and Recognize Anything at Any Granularity/ec47fa22-1632-444f-b911-9d6b0be51867_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0d5d45f6ee5b3e4063efb60e76096ceb7a686fb20723ad118bd7261be4ee7863 +size 28405236 diff --git a/2024/Segment and Recognize Anything at Any Granularity/full.md b/2024/Segment and Recognize Anything at Any Granularity/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d411d9e0e7a4b4dc37b3fddfd8703ae24e2f1d28 --- /dev/null +++ b/2024/Segment and Recognize Anything at Any Granularity/full.md @@ -0,0 +1,272 @@ +# Segment and Recognize Anything at Any Granularity + +Feng Li\* Hao Zhang\* Peize Sun, Xueyan Zou, Shilong Liu, Chunyuan Li Jianwei Yang1, Lei Zhang2, Jianfeng Gao + +§ HKUST ‡ Microsoft Research, Redmond † IDEA # HKU § UW-Madison Tsinghua * Core Contributor 1. Project Lead 2. Equal Advisory Contribution + +Abstract. In this work, we introduce Semantic-SAM, an augmented image segmentation foundation for segmenting and recognizing anything at desired granularities. Compared to the foundational segmentation model SAM [31], our model has two unique advantages: (i) granularity-controllability in that the model can produce segmentation masks at any desired granularities, from objects to parts to both; (ii) semantic-awareness in that the model simultaneously predicts semantic labels for masks at different granularities. To enable multi-granularity capabilities, we propose a multi-choice learning scheme, where each click point generates a set of masks at multiple levels of granularity, corresponding to a set of ground-truth masks. To achieve semantic awareness, we consolidate multiple datasets of different levels of granularity and train our model using decoupled object- and part-based tasks to facilitate knowledge sharing and transfer among different tasks. To the best of our knowledge, this work is the first attempt to jointly train a model on SA-1B, instance-level, and part-level segmentation datasets. Experimental results and visualizations demonstrate that our model successfully achieves the desired goals. Furthermore, we show that multi-task training using the segmentation task defined on SA-1B and other segmentation tasks (e.g., panoptic and part segmentation) leads to performance gains on all segmentation tasks. In particular, we achieve a new state-of-the-art in COCO panoptic segmentation 60.2 PQ by adding SAM data. + +# 1 Introduction + +The universal and interactive AI systems that follow human intents have shown their potential in natural language processing [46, 47] and controllable image generation [52, 66]. However, such a universal system for pixel-level image understanding remains less explored. We argue that a foundational segmentation model should possess the following important properties: granularity-controllability and semantic-awareness. Regardless of the specific image domain or prompt context, the model is capable of acquiring a versatile representation, predicting segmentation masks in multi-granularity, and understanding the semantic meaning behind each segmented region. + +Previous works [31,58,70] have strived for these properties but only achieved part of the goals. The main obstacles impeding the progress of such a universal + +![](images/2b430727df2cf1fd4619fbf2c34f3a4d0a905b8a4dd535ff64f5d280268ff9e6.jpg) +Fig. 1: Dataset distribution in the semantics-granularity space.remove + +image segmentation model can be attributed to the limited flexibility of model design and the lack of training data. + +- Model Design. The existing image segmentation model architectures are dominated by the single-input-single-output pipeline that discards any ambiguity. While this pipeline is prevalent in both anchor-based CNN architectures [24] and query-based Transformer architectures [4, 11], and has demonstrated remarkable performance in semantic, instance, and panoptic segmentation tasks [30, 39, 68], it inherently restricts the model to predict multi-granularity segmentation masks in an end-to-end manner. Although clustering postprocessing techniques [13] can produce multiple masks for a single object query, they are neither efficient nor effective solutions for a granularity-aware segmentation model. + +- Training Data. Scaling up segmentation datasets that possess both semantic-awareness and granularity-awareness is a costly endeavor. Existing generic detection and segmentation datasets such as MSCOCO [39] and Objects365 [53] offer large amounts of data and rich semantic annotations, but only at the object level. On the other hand, part segmentation datasets such as Pascal Part [9], PartImageNet [23], and PACO [49] provide more fine-grained semantic annotations, but their data volumes are limited. Recently, SAM [31] has successfully scaled up the multi-granularity mask data to millions of images, but it does not include semantic annotations. In order to achieve the dual objectives of semantic-awareness and granularity-abundance, there is a pressing need to unify model design and training on various data formats to facilitate knowledge transfer. However, the inherent differences in semantics and granularity across different datasets pose a significant challenge to joint training efforts. + +In this paper, we introduce Semantic-SAM, a universal image segmentation model designed to enable segmenting and recognizing objects at any desired granularity. Given one click from a user, our model addresses the spatial ambiguity by predicting masks in multiple granularities, accompanied by semantic labels at both the object and part levels. As shown in Figure 2, our model gener- + +![](images/67b2b578dfbc0c79c7dcf2547ccdab60c900c5d2500206ad4be73bc15d5afadb.jpg) +Fig. 2: Our model is capable of dealing with various segmentation tasks including open-set and interactive segmentation. (a) Our model can accomplish instance, semantic, panoptic segmentation, and part segmentation. (b) Our model is able to output multilevel semantics in different granularities. The yellow star point on the left-most image is the user click. (c) We connect our model with an inpainting model to perform multilevel inpainting. The prompts are "Spider-Man" and "BMW car", respectively. Note that only one click is needed to produce the results in (b) and (c), respectively. + +ates multi-level segmentation masks ranging from the person head to the whole truck. + +The multi-granularity capability is achieved through a multi-choice learning design [22, 37] incorporated into the decoder architecture. Each click point is represented with multiple queries, each containing a different level of embedding. These queries are trained to learn from all available ground-truth masks representing different granularities. To establish a correspondence between multiple masks and ground-truths, we employ a many-to-many matching scheme to ensure that a single click point could generate high-quality masks in multiple granularities. + +To accomplish semantic-awareness with a generalized capability, we introduce a decoupled approach for objects and parts classification, to allow knowledge transfer among rich object and part segmentation datasets. By leveraging a shared text encoder to encode both objects and parts independently, we adapt the loss function based on the data type. For instance, generic segmentation data lacks part classification loss, whereas SAM data does not include classification loss. + +To enrich semantics and granularity within our model, we consolidate seven datasets on three types of granularities, including generic segmentation of MSCOCO [39], Objects365 [53], ADE20k [68], part segmentation of PASCAL Part [9], PACO [49], + +![](images/a7e5a3a04f70df4b12adceb061fea3b19c41956a8a1a6dae4b9d646b735ecc53.jpg) +Fig. 3: Semantic-SAM is a universal segmentation framework that can take multiple types of segmentation data including generic, part, and class-agnostic segmentation data. The vision encoder is used to extract image features. The mask decoder can do both generic segmentation and promptable segmentation with various types of prompts. For point and box, we input them via anchor boxes to the mask decoder. Since there is an ambiguity of granularity for a point input, we duplicate each point 6 times and give them different levels of embeddings. The output masks of point prompts match with multiple GT masks of different granularities. + +PartImagenet [23], and SA-1B [31]. Their data formats are reorganized to match our training objectives accordingly. After joint training, our model obtains a strong performance across a variety of datasets. Notably, we find that learning from interactive segmentation could improve generic and part segmentation. For example, by jointly training SA-1B promptable segmentation and COCO panoptic segmentation, we achieve a gain of 2.2 box AP and a gain of 1.3 mask AP. In addition, through comprehensive experiments, we demonstrate that our model can cover more segmentation granularities than SAM, outperforming 3.4 IoU with one point click. + +# 2 Semantic-SAM + +# 2.1 Mode Design + +Our Semantic-SAM follows MaskDINO [33] to employ a query-based mask decoder. To produce semantic-aware and multi-granularity masks, we exploit two types of queries: (i) generic queries as used in MaskDINO and (ii) spatial queries including point and box for granularity-controllable interactive segmentation. The overall pipeline is shown in Figure 3. + +We represent both point and box prompts into anchor boxes as a unified format. In particular, we convert user click point $(x,y)$ into an anchor box $(x,y,w,h)$ with small width $w$ and height $h$ , so that the anchor box can closely approximate the point. To capture different granularities of masks, each click is first encoded to position prompt and combined with $K$ different content prompts, each of which is represented as a learnable embedding for a given + +![](images/293359e4ab4ffd0cf73dfbe792229bbfa860280f86b3bf5a3ad215afe67a0014.jpg) +Fig. 4: Decoupled object and part classification. Given the same decoder outputs (right), two set of embeddings are utilized for object and part classification separately. + +granularity level. As a result, we feed into the decoder $K$ query embeddings $\mathbf{Q} = (\mathbf{q}_1,\dots ,\mathbf{q}_K)$ for a click represented as an anchor box $\mathbf{b} = (x,y,w,h)$ , and the $i$ -th query is formulated as + +$$ +\mathbf {q} _ {i} = \mathbf {q} _ {i} ^ {\text {l e v e l}} + \mathbf {q} ^ {\text {a n c h o r}}, \tag {1} +$$ + +where $\mathbf{q}_i^{\mathrm{level}}$ is the content embedding for granularity level $i$ , and $\mathbf{q}^{\mathrm{anchor}}$ is a position embedding derived from the sine encoding of the anchor box. Here we empirically choose $K = 6$ , considering there are at most 6 levels of masks per user click for the majority of images in SA-1B [31]. As for the box prompt, we simply use a single content embedding, i.e., $K = 1$ . + +Assuming that the output image feature from the vision encoder is $\mathbf{F}$ , the mask decoder takes a set of queries and produces the outputs: + +$$ +\mathbf {O} = \operatorname {D e f o r m D e c} (\mathbf {Q}, \mathbf {F}) \text {w i t h} \mathbf {O} = \left(\mathbf {o} _ {1}, \dots , \mathbf {o} _ {K}\right), \tag {2} +$$ + +where $\mathsf{DeformDec}(\cdot, \cdot, \cdot)$ is a deformable decoder that takes anchor box queries, and image features as input to output queried features. $\mathbf{o}_i$ is the model output for the $i$ -th input query $\mathbf{q}_i$ . Each $\mathbf{o}_i = (\mathbf{c}_i, \mathbf{m}_i)$ consists of the predicted semantic category $\mathbf{c}_i$ and mask $\mathbf{m}_i$ , which are used to construct the concept recognition loss and mask prediction loss, respectively. + +# 2.2 Training + +Segment any granularity. To endow the model with a multi-granularity segmentation ability, we propose a many-to-many matching method during training. We found that SAM fails to provide good multi-level segmentation results with a single click. Since SAM uses many-to-one matching [31] during training, the three SAM-predicted masks for each click point only match with one GT mask. This means that points located in masks of small levels cannot predict large masks with high quality according to our observation. In contrast, we fully leverage the structures in both data and algorithm to enable multi-level mask prediction with + +![](images/7045eae0504bb1930ebf07c469707350b1a5efb3e1a5e12eae41c5a2693b747b.jpg) +Fig. 5: Comparisons of different interactive learning strategies. a) One-to-one: traditional interactive segmentation models that focus on object-level, i.e., SEEM [70]. b) Many-to-one: multi-choice learning for single granularity, i.e., SAM [31]. c) Many-to-many: our Semantic-SAM enforces the model to predict all the possible granularities of a single click for more diverse segmentation. d) As a result, our model is able to output multiple masks in richer granularities. + +a single click. First, we reorganized the data by clustering multiple GT masks of different levels sharing the same click. To allow multiple predictions of the same click to match with the GT masks, we employ the Hungarian algorithm to enable the many-to-many matching. The similarity matrix and scores vary based on the availability of different segmentation data components. + +For box input and generic segmentation, we follow existing methods. Specifically, to generate a mask from an input box, we follow a similar idea as in denoising training (DN) [33]. We add noises to ground-truth boxes to simulate inaccurate box inputs from users, and these noised boxes serve as spatial prompts for the decoder. The model is trained to reconstruct the original boxes and masks given noised boxes. For the content part of box prompts, we adopt a learnable token as a general prompt. Note that this is the only difference from DN, as DN uses ground-truth label embedding as the content prompts. For generic segmentation, we follow the same pipeline as in Mask DINO [33]. + +Recognize anything. We train with various types of data with different semantic annotations, in which some contain object-level annotations (COCO), some contain both object and part-level annotations (Pascal Part), and SA-1B has no semantic annotations but contains masks of all semantic levels. Note that a large number of part concepts are shared across different objects, for example, head for all animals, we aim to transfer the part concept knowledge across objects trained with only object-level annotations in our joint training. To address this discrepancy between semantic annotations and better transfer semantics of different granularity, we propose to decouple object and part recognition. As shown in Figure 4, we utilize a shared text encoder to encode objects and parts, which are used to perform object and part segmentation separately. + +While all types of segmentation data share a unified format, the loss varies for different data types. In Table 1, we summarize the loss items to construct the training objective in Semantic-SAM. It is the part-level data that bridges the gap to recognize semantic concepts between part and object levels, and it is the use of SAM data that bridges the gap to segment masks at any granularity. + +
DataRecognize AnythingSegment at Any Granularity
PartObjectBox Mask #GT in Matching
SAM dataXX✓✓Many
Object-level dataX✓✓✓One
Part-level data✓✓✓✓One
+ +Table 1: The loss items to construct the training objective in Semantic-SAM. The four loss items are part classification, object classification, box loss, and mask loss, respectively. The last column indicates the number of ground-truth masks in the matching. + +
TypeData#ImagesSemantic ConceptGranularity Level
PartObjectPartWhole
Class-agnosticSA-1B11MXX✓✓
Object-levelObjects3651.7MX365X✓
COCO110KX133X✓
ADE20K20KX150X✓
Part-levelPACO-LVIS45K20175✓✓
PartImageNet16K1311✓✓
Pascal Part5K3020✓✓
+ +Table 2: The statistics of training data used in Semantic-SAM. + +Discussion. As shown in Figure 5, Semantic-SAM differs from previous interactive segmentation models in two aspects. Firstly, our model is trained with multi-choice learning to output all the possible segmentation masks with one click point. Secondly, our output granularities are richer to generate diverse segmentation masks. + +# 2.3 Data Unification + +In order for multi-level semantics, we include seven datasets that contain different granularity-level masks. The datasets are SA-1B [31], COCO panoptic [39], ADE20k panoptic [68], PASCAL part [44], PACO [49], PartImageNet [23], and Objects365 [53]. Within them, COCO and ADE20k panoptic datasets contain object-level masks and class labels. PASCAL part, PACO, and PartImageNet contain part-level masks and class labels. SA-1B contains up to 6-level masks without labels, while Objects365 contains abundant class labels for object-level instances. The details of these datasets are shown in Table 2. We further visualize the data distribution of different data types in Figure 1. + +
MethodTypeTraining DataPQmIoUAPAPsAPmAPI box mask
boxmaskboxmaskboxmask
Mask2Former (T) [11]Close-setCOCO53.263.246.143.3-----
X-Decoder (T) [69]Open-setCOCO+VL52.662.443.641.3-----
OpenSeed (T) [65]Open-setCOCO+O36555.463.851.247.134.527.454.350.466.2 66.8
Semantic-SAM (T) (ours)Open-setCOCO54.663.250.146.134.427.153.249.466.1 66.1
Semantic-SAM (T) (ours)Open-setCOCO+SAM55.263.452.3(+2.2)47.4(+1.3)36.1(+1.7)28.3(+1.2)55.6(+2.4)50.7(+1.3)67.3 66.2
OpenSeed (L) [65]Open-setCOCO+O36559.568.658.253.2-----
Oneformer (H) [26]Open-setCOCO+O36560.068.8-52.0-----
Semantic-SAM (L) (ours)Open-setCOCO+SAM60.268.957.553.042.935.461.256.971.9 71.4
+ +Table 3: Results for Semantic-SAM and other panoptic segmentation models on COCO val2017. Our model is jointly trained on COCO [8] and SA-1B [31] (1/10 data) and directly evaluates COCO.add model param size, hq-sam comparison, more datasets (ADE, cityscapes and compare with SEEM), more impressive visualization + +# 3 Experiments + +# 3.1 Experimental Setup + +Implementation Details. In our experiments, we train on SA-1b [31] with no semantic annotation, COCO [39] with object-level annotation and Pascal Part [15] with part-level annotation. We implement our model based on Mask DINO [33]. Mask DINO is a unified detection and segmentation framework that simultaneously predicts box and mask. We follow [33] to use 300 latent queries and nine decoder layers for all segmentation tasks. For the visual backbone, we adopt pre-trained Swin-T/L [41] by default. For the language backbone, we adopt the pre-trained base model in UniCL [62]. As SA-1B [31] dominates the data, during training, we first train on only SA-1B data. Then, we add object and part-level data to jointly train the three types of data. During training, the image resolution is $1024 \times 1024$ for all data. We use AdamW [43] as the optimizer. We use large-scale jittering for object and part-level data and did not use data augmentations for SA-1B data, as SA-1B images are abundant. We set the learning rate to 0.0001, which is decayed at 0.9 and 0.95 fractions of the total number of steps by 10. + +Evaluation. We mainly evaluate two datasets, including COCO val2017 and a subset of SA-1B [31] with 1000 images. For evaluation metrics, we evaluate PQ and AP for generic and part segmentation datasets. For single-granularity interactive segmentation, we report 1-click mIoU (denoted as 1-IoU) under Point (Max) and Point (Oracle) settings. Max denotes we select the output mask with the maximum confidence score. Oracle denotes we select the output mask with the max IoU by calculating the IoU between the prediction and target mask. For multi-granularity interactive segmentation, we report 1-IoU@All Granularity that matches all the possible ground-truth masks for a single click to the multi-granularity predictions and then calculate the average IoU of all granularities. + +# 3.2 Semantic Segmentation of Anything + +Generic Segmentation. As shown in Table 3, to validate the compatibility of multi-granularity interactive segmentation and generic segmentation, we jointly train with SA-1B [31] (1/10 data) and COCO panoptic segmentation. The result indicates that interactive segmentation with SAM can significantly help the instance-level detection and segmentation with a performance improvement of $+2.2$ AP on the box and $+1.3$ AP on the mask. Notably, OpenSeed [65] and Semantic-SAM are both based on Mask DINO [33]. Our joint training with SA-1B even outperforms OpenSeed which is trained with Object365 [53]. In addition, adding SA-1B mainly improves small object detection (APs and APm), as there are a large number of small objects in SA-1B. + +
MethodTypeTraining DataAPAPsAPmAP1
boxmaskbox maskbox maskbox mask
VLPart (B) [55]Open-setPascal Part29.227.4- -- -- -
Semantic-SAM (T) (ours)Open-setPascal Part27.030.516.6 19.138.1 41.643.8 49.1
Semantic-SAM (T) (ours)Open-setPascal Part+SAM28.0(+1.0)31.4(+0.9)17.3 19.940.0 42.545.7 49.7
+ +Table 4: Results for Semantic-SAM and other part segmentation models on Pascal Part. Our model is jointly trained on Pascal Part [15] and SA-1B [31] (1/10 data) and directly evaluates Pascal Part. + +Part Segmentation. We also validate the compatibility of joint training SA-1B (1/10 data) and part segmentation. As shown in Table 4, adding SA-1B brings a decent performance improvement on Pascal Part [15], +1.0 AP on the box and +0.9 AP on the mask. Importantly, Semantic-SAM (T) has better mask AP than VLPart (B) [55], showing the high mask quality of our model. + +Single-granularity Interactive Segmentation. In Table 5, we evaluate 1-IoU for SAM and our model on COCO val2017. Our model outperforms SAM under both Point (Max) and Point (Oracle) settings. + +Multi-granularity Interactive Segmentation. In Table 6, we compare SAM [31] and our model on the output granularities for a single click. We adopt a Hungarian Matching to match all the possible target masks with the predicted masks for the click and calculate the average IoU score. As SAM has only three prompts, we also sample two clicks from a single mask to produce six output masks for a fair comparison. Notably, SAM has been trained on this validation set while we do not. + +# 3.3 Ablations + +Match Strategy. As shown in Table 7, we compare different match strategies in our model. When using many-to-many matching to match all the possible + +
MethodCOCO
Point (Max) Point (Oracle)
1-IoU1-IoU
SAM (B)52.168.2
SAM (L)55.770.5
Semantic-SAM (T)54.573.8
Semantic-SAM (L)57.074.2
+ +Table 5: Comparison with previous models on point interactions. Both SAM [31] and our model are trained on only SA-1B and directly evaluate on COCO val2017 for fair comparison. Max denotes we select the output with the max confidence score prediction. Oracle denotes we select the output with the max IoU by calculating the IoU between the prediction and target mask. + +
MethodGranularity1-IoU@All Granularity
SAM (B)†375.6
SAM (L)†382.5
SAM (H)†383.5
Semantic-SAM(T)688.1
Semantic-SAM(L)689.0
+ +Table 6: Mask granularity comparison between SAM and our model. We evaluated both on an excluded subset of 1000 images from SA-1B. For each click, we evaluate all the possible ground-truth masks to calculate the 1-IoU@All Granularity. SAM [31] and Semantic-SAM adopts three and six prompts for a single click of a mask, respectively. $\dagger$ denotes that SAM has been trained on this validation subset while we do not. + +ground-truth masks for each click, the 1-IoU@All Granularity performance is significantly improved. This validates our matching strategy is effective in learning complete granularities. + +Box Interactive Evaluation. We also evaluate the 1-IoU given boxes in Table 8. We achieve better performance compared with object-level interactive segmentation model SEEM [70] and multi-granularity model SAM [31]. + +Increasing SA-1B Training Data. In Table 9, we show the performance improvement on COCO val 2017 when training with more SA-1B data. The performance is saturated after using more than $15\%$ of the total data. It indicates that we do not need to train with the whole SA-1B data to get a good zero-shot performance. + +# 3.4 Visualization + +We conducted a comparative analysis between our model and SAM to showcase the superior performance of our approach in generating high-quality masks, as depicted in Figure 6. The following key properties are highlighted: + +
MethodMatch1-IoU
Semantic-SAM(T)m2o73.2
Semantic-SAM(T)m2m88.1
+ +Table 7: Effect of match strategy on segmentation performance. + +
MethodBox 1-IoU
SAM [31](B)50.7
SEEM [70](T)73.7
Semantic-SAM(T)76.1
+ +Table 8: Box 1-IoU evaluation on COCO val2017. + +
MethodData Portion of SA-1BCOCO
Point (Max) Point (Oracle)
1-IoU1-IoU
SAM (L)100%55.770.5
Semantic-SAM (L)3%55.273.5
Semantic-SAM (L)15%56.773.6
Semantic-SAM (L)30%55.773.7
Semantic-SAM (L)50%55.373.9
Semantic-SAM (L)100%57.074.2
+ +Table 9: Models trained with different portions of SA-1B [31], and directly evaluated on COCO val2017. + +![](images/848f5cf19c3d946c1951de6cd2a2d92e473ace4d8e93f04af2922ede1bc78bd4.jpg) +Fig. 6: Given the user clicks shown in the first column, odd rows and even rows show the generated masks by SAM and our model, respectively. We removed the duplicated masks from our model to simplify the visualizations. More results are in supplementary. + +![](images/03de6788d13da82b10ef4c0202c426e1c65f5fc9e2ec21ce9e3e4f311b5aa2e8.jpg) +Fig. 7: Our model learns an emergent capability that each level prompt corresponds to a certain granularity of mask. Given a single point click (red star) at the first column, our model produces masks with increasing areas from left to right for different images. A coherent order of granularities are observed for the prompts 1-6 across different images. + +- Multi-Level Masks. Our model excels in producing masks with a greater range of meaningful granularities. Unlike SAM, which typically outputs a maximum of three masks with occasional duplication, our model generates a more diverse set of masks at various levels, enhancing the overall flexibility and precision of mask outputs. +- Mask Qualities. Our model has demonstrated its ability to generate masks of higher quality compared to SAM. SAM occasionally produces masks with artifacts such as holes or islands, particularly for large regions when the click falls within an interior region. In contrast, our model consistently generates high-quality masks across all levels, offering accurate segmentation results. +- Comparison with SA-1B Ground-truth Granularity. We extend the granularity of SAM data by producing masks with more meaningful levels of detail. Our model extends the original number of granularity levels, providing finer distinctions and more precise segmentation results on SAM data. +- Comparison with SAM in segmenting everything. We compare with SAM to segment everything in one image with the same number of points. We output more granularities for the image. + +Query semantics. We also find that each point content prompt embedding learns to correspond to a fixed granularity. As shown in Fig. 7, when we visualize masks in a specific order of the corresponding content embeddings, the masks follow the order from small to large in each row consistently. This proves that each content embedding represents a semantic granularity level in our model. + +![](images/3c4bdeb0fbcaf61596925ba4b964b57a4b80ab627b14449f40b83eda4b7d440a.jpg) +Fig. 8: Compare SAM and our model to segment everything in one image. + +# 4 Related works + +# 4.1 Generic Segmentation + +Segmenting visual concepts is well-documented within the expansive field of computer vision [16,17,45,71]. Broad segmentation methodologies comprise several subdivisions, such as instance segmentation, semantic segmentation, and panoptic segmentation [6,24,30], each catering to a unique semantic degree. For example, semantic segmentation's goal is to detect and assign a label to each pixel in an image according to its corresponding semantic class [7,11,42]. Conversely, instance segmentation seeks to cluster pixels associated with the same semantic class into distinct object instances [3,24,33]. Panoptic segmentation is the hybrid of these two tasks. Recently, Transformer-based methods [4,56] have contributed to significant progress in segmentation tasks [11, 26, 33, 38, 64]. Generic object detection and segmentation have led to the development of abundant datasets, such as MSCOCO [39], LVIS [21], Objects365 [53], PASCAL [15], CityScapes [12], ADE20k [68]. + +# 4.2 Part Segmentation + +Beyond generic segmentation, part segmentation aims to more fine-grained visual understanding. Most early works were bottom-up methods by grouping superpixels into parts and then objects [1, 2, 20]. Later, based on high-performance object detection networks [24, 51], top-down methods were developed by firstly detecting an object and then parsing it to part segmentation [27, 34, 63]. To segment the scene in multi-granularity, part-aware panoptic segmentation [13] is introduced. PPS [13] establishes the baseline by assembling panoptic and part segmentation models. JPPF [25] simplifies the model by a shared image encoder for both panoptic segmentation and part segmentation. By representing things, stuffs, and parts as object queries, Panoptic-PartFormer [35] proposes a unified architecture based on Transformer. While part segmentation data is much more expensive than object detection and segmentation data, a number of public datasets are available. Datasets for specific domains include cars [54], birds [57], and fashion [29]. General objects include Pascal-Part [9], PartImageNet [23], ADE20K [67], Cityscapes-Panoptic-Parts [44], and PACO [49]. More recently, SAM [31] provides a large-scale multi-granularity class-agnostic segmentation dataset. Our work is jointly trained on these datasets and contributes to a multi-granularity segmentation model. + +# 4.3 Open-Vocabulary Segmentation + +While generic segmentation and part segmentation have made remarkable progress, they can only segment the image in a close-set vocabulary. To expand the vocabulary size, recent works leverage the visual-semantic knowledge from large-scale foundation models like CLIP [48], ALIGN [28] and Diffusion models [60] to various segmentation tasks. LSeg [32], OpenSeg [18], GroupViT [59] achieves open-vocabulary semantic segmentation ability on ADE20K and PASCAL. Dense-CLIP [50] and MaskCLIP [14] achieve open-vocabulary instance and panoptic segmentation on COCO dataset. More recently, X-Decoder [69] proposes a unified approach to tackle various segmentation and vision-language tasks for open-vocabulary segmentation, OpenSeeD [65] proposes to use a large amount of detection data and a joint training method to improve segmentation. To segment open-vocabulary masks in part-level, VLPart [55] leverages three part segmentation datasets and learns from the dense correspondence [5] between base objects and novel objects. Our work unifies these tasks into one architecture and builds up open-vocabulary segmentation in multi-granularity. + +# 4.4 Interactive Segmentation + +Interactive segmentation refers to the process of separating objects by actively integrating user inputs. This enduring challenge has seen notable advancements [10, 19, 31, 36, 40, 61]. Previous works only focus on a small set of data or semantic-agnostic instance masks. Recently, SAM [31] enlarges the training data from 0.12M COCO images to 10M SAM fine-grained images. SEEM [70] enriches the modality to language and function to both generic and grounded segmentation with an impressive composition. However, they focus either enriching the mask granularity or enhancing the semantic understanding. Our work is the first that reconciles both in a single unified model. + +# 5 Conclusion + +In this paper, we have presented Semantic-SAM, which can segment and recognize anything at desired granularities. Apart from performing generic open-vocabulary segmentation, Semantic-SAM demonstrates the advantages of semantic awareness and granularity abundance. To achieve such advantages, we propose improvements on data, model, and training where we utilize datasets from multiple granularity and semantic levels, multi-choice learning for training, and a universal framework for modeling. Comprehensive experiments and visualizations have verified the semantic awareness and granularity abundance are achieved in our model. Further, Semantic-SAM is the first successful attempt to jointly train on SA-1B and other classic segmentation datasets, and shows that training with SA-1B improves other tasks such as panoptic and part segmentation. + +# References + +1. Pablo Arbelaez, Michael Maire, Charless Fowlkes, and Jitendra Malik. Contour detection and hierarchical image segmentation. IEEE transactions on pattern analysis and machine intelligence, 33(5):898-916, 2010. +2. Pablo Arbeláez, Jordi Pont-Tuset, Jonathan T Barron, Ferran Marques, and Jitendra Malik. Multiscale combinatorial grouping. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 328-335, 2014. +3. Daniel Bolya, Chong Zhou, Fanyi Xiao, and Yong Jae Lee. Yolact: Real-time instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9157-9166, 2019. +4. Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision, pages 213-229. Springer, 2020. +5. Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. +6. Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834-848, 2017. +7. Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587. 2017. +8. Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dolkar, and C. Lawrence Zitnick. Microsoft COCO captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. +9. Xianjie Chen, Roozbeh Mottaghi, Xiaobai Liu, Sanja Fidler, Raquel Urtasun, and Alan Yuille. Detect what you can: Detecting and representing objects using holistic models and body parts, 2014. +10. Xi Chen, Zhiyan Zhao, Yilei Zhang, Manni Duan, Donglian Qi, and Hengshuang Zhao. Focalclick: towards practical interactive image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1300-1309, 2022. +11. Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290-1299, 2022. +12. Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213-3223, 2016. +13. Daan de Geus, Panagiotis Meletis, Chenyang Lu, Xiaoxiao Wen, and Gijs Dubbelman. Part-aware panoptic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5485-5494, 2021. +14. Zheng Ding, Jieke Wang, and Zhuowen Tu. Open-vocabulary panoptic segmentation with maskclip. arXiv preprint arXiv:2208.08984, 2022. +15. Mark Everingham and John Winn. The pascal visual object classes challenge 2012 (voc2012) development kit. Pattern Analysis, Statistical Modelling and Computational Learning, Tech. Rep, 8(5), 2011. + +16. Pedro F Felzenszwalb, Ross B Girshick, David McAllester, and Deva Ramanan. Object detection with discriminatively trained part-based models. IEEE transactions on pattern analysis and machine intelligence, 32(9):1627-1645, 2009. +17. King-Sun Fu and JK Mui. A survey on image segmentation. Pattern recognition, 13(1):3-16, 1981. +18. Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Open-vocabulary image segmentation. arXiv preprint arXiv:2112.12143, 2021. +19. Leo Grady. Random walks for image segmentation. IEEE transactions on pattern analysis and machine intelligence, 28(11):1768-1783, 2006. +20. Matthias Grundmann, Vivek Kwatra, Mei Han, and Irfan Essa. Efficient hierarchical graph-based video segmentation. In 2010 IEEE computer society conference on computer vision and pattern recognition, pages 2141-2148. IEEE, 2010. +21. Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5356-5364, 2019. +22. Abner Guzman-Rivera, Dhruv Batra, and Pushmeet Kohli. Multiple choice learning: Learning to produce multiple structured outputs. Advances in neural information processing systems, 25, 2012. +23. Ju He, Shuo Yang, Shaokang Yang, Adam Kortylewski, Xiaoding Yuan, Jie-Neng Chen, Shuai Liu, Cheng Yang, and Alan Yuille. Partimagenet: A large, high-quality dataset of parts. arXiv preprint arXiv:2112.00933, 2021. +24. Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. +25. Sravan Kumar Jagadeesh, René Schuster, and Didier Stricker. Multi-task fusion for efficient panoptic-part segmentation. arXiv preprint arXiv:2212.07671, 2022. +26. Jitesh Jain, Jiachen Li, MangTik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. Oneformer: One transformer to rule universal image segmentation. arXiv preprint arXiv:2211.06220, 2022. +27. Ruyi Ji, Dawei Du, Libo Zhang, Longyin Wen, Yanjun Wu, Chen Zhao, Feiyue Huang, and Siwei Lyu. Learning semantic neural tree for human parsing. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIII 16, pages 205-221. Springer, 2020. +28. Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V Le, Yunhsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. +29. Menglin Jia, Mengyun Shi, Mikhail Sirotenko, Yin Cui, Claire Cardie, Bharath Hariharan, Hartwig Adam, and Serge Belongie. Fashionpedia: Ontology, segmentation, and an attribute localization dataset. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 316-332. Springer, 2020. +30. Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dólar. Panoptic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9404-9413, 2019. +31. Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything, 2023. +32. Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and René Ranftl. Language-driven semantic segmentation. arXiv preprint arXiv:2201.03546, 2022. +33. Feng Li, Hao Zhang, Shilong Liu, Lei Zhang, Lionel M Ni, Heung-Yeung Shum, et al. Mask dino: Towards a unified transformer-based framework for object detection and segmentation. arXiv preprint arXiv:2206.02777, 2022. + +34. Qizhu Li, Anurag Arnab, and Philip HS Torr. Holistic, instance-level human parsing. arXiv preprint arXiv:1709.03612, 2017. +35. Xiangtai Li, Shilin Xu, Yibo Yang, Guangliang Cheng, Yunhai Tong, and Dacheng Tao. Panoptic-partformer: Learning a unified model for panoptic part segmentation. In European Conference on Computer Vision, pages 729-747. Springer, 2022. +36. Yin Li, Jian Sun, Chi-Keung Tang, and Heung-Yeung Shum. Lazy snapping. ACM Transactions on Graphics (ToG), 23(3):303-308, 2004. +37. Zhuwen Li, Qifeng Chen, and Vladlen Koltun. Interactive image segmentation with latent diversity. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 577-585, 2018. +38. Zhiqi Li, Wenhai Wang, Enze Xie, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, Ping Luo, and Tong Lu. Panoptic segformer: Delving deeper into panoptic segmentation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1280-1289, 2022. +39. Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. +40. Qin Liu, Zhenlin Xu, Gedas Bertasius, and Marc Niethammer. Simpleclick: Interactive image segmentation with simple vision transformers. arXiv preprint arXiv:2210.11006, 2022. +41. Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. +42. Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3431-3440, 2015. +43. Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. +44. Panagiotis Meletis, Xiaoxiao Wen, Chenyang Lu, Daan de Geus, and Gijs Dubbelman. Cityscapes-panoptic-parts and pascal-panoptic-parts datasets for scene understanding. arXiv preprint arXiv:2004.07944, 2020. +45. Shervin Minaee, Yuri Y Boykov, Fatih Porikli, Antonio J Plaza, Nasser Kehtarnavaz, and Demetri Terzopoulos. Image segmentation using deep learning: A survey. IEEE transactions on pattern analysis and machine intelligence, 2021. +46. OpenAI. Chatgpt. https://openai.com/blog/chatgpt, 2022. +47. OpenAI. Gpt-4 technical report, 2023. +48. Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. +49. Vignesh Ramanathan, Anmol Kalia, Vladan Petrovic, Yi Wen, Baixue Zheng, Baishan Guo, Rui Wang, Aaron Marquez, Rama Kovvuri, Abhishek Kadian, Amir Mousavi, Yiwen Song, Abhimanyu Dubey, and Dhruv Mahajan. PACO: Parts and attributes of common objects. In arXiv preprint arXiv:2301.01795, 2023. +50. Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18082-18091, 2022. +51. Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28:91-99. + +52. Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. +53. Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8430-8439, 2019. +54. Xibin Song, Peng Wang, Dingfu Zhou, Rui Zhu, Chenye Guan, Yuchao Dai, Hao Su, Hongdong Li, and Ruigang Yang. Apollocar3d: A large 3d car instance understanding benchmark for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5452-5462, 2019. +55. Peize Sun, Shoufa Chen, Chenchen Zhu, Fanyi Xiao, Ping Luo, Saining Xie, and Zhicheng Yan. Going denser with open-vocabulary part segmentation, 2023. +56. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. +57. Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. technical report, 2011. +58. Xinlong Wang, Xiaosong Zhang, Yue Cao, Wen Wang, Chunhua Shen, and Tiejun Huang. Seggpt: Segmenting everything in context. arXiv preprint arXiv:2304.03284, 2023. +59. Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, and Xiaolong Wang. Groupvit: Semantic segmentation emerges from text supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18134-18144, 2022. +60. Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. arXiv preprint arXiv:2303.04803, 2023. +61. Ning Xu, Brian Price, Scott Cohen, Jimei Yang, and Thomas S Huang. Deep interactive object selection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 373-381, 2016. +62. Jianwei Yang, Chunyuan Li, Pengchuan Zhang, Bin Xiao, Ce Liu, Lu Yuan, and Jianfeng Gao. Unified contrastive learning in image-text-label space. In CVPR, 2022. +63. Lu Yang, Qing Song, Zhihui Wang, and Ming Jiang. Parsing r-cnn for instance-level human analysis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 364-373, 2019. +64. Hao Zhang, Feng Li, Huaizhe Xu, Shijia Huang, Shilong Liu, Lionel M Ni, and Lei Zhang. Mp-former: Mask-piloted transformer for image segmentation. arXiv preprint arXiv:2303.07336, 2023. +65. Hao Zhang, Feng Li, Xueyan Zou, Shilong Liu, Chunyuan Li, Jianfeng Gao, Jianwei Yang, and Lei Zhang. A simple framework for open-vocabulary segmentation and detection. arXiv preprint arXiv:2303.08131, 2023. +66. Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543, 2023. +67. Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. +68. Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset, 2018. + +69. Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. arXiv preprint arXiv:2212.11270, 2022. +70. Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. arXiv preprint arXiv:2304.06718, 2023. +71. Zhengxia Zou, Zhenwei Shi, Yuhong Guo, and Jieping Ye. Object detection in 20 years: A survey. arXiv preprint arXiv:1905.05055, 2019. \ No newline at end of file diff --git a/2024/Segment and Recognize Anything at Any Granularity/images.zip b/2024/Segment and Recognize Anything at Any Granularity/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..003322270f11f168d248c639ab6d75bb637ed9fb --- /dev/null +++ b/2024/Segment and Recognize Anything at Any Granularity/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5000b7a1942dd8fa9e41dad355a08fa5d396bff62e408618531b1f55b9f6c7cc +size 714267 diff --git a/2024/Segment and Recognize Anything at Any Granularity/layout.json b/2024/Segment and Recognize Anything at Any Granularity/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d656c76e0e47c5609fcc49598fecdca8feddf105 --- /dev/null +++ b/2024/Segment and Recognize Anything at Any Granularity/layout.json @@ -0,0 +1,7950 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 162, + 111, + 451, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 111, + 451, + 148 + ], + "spans": [ + { + "bbox": [ + 162, + 111, + 451, + 148 + ], + "type": "text", + "content": "Segment and Recognize Anything at Any Granularity" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 154, + 167, + 462, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 167, + 462, + 193 + ], + "spans": [ + { + "bbox": [ + 154, + 167, + 462, + 193 + ], + "type": "text", + "content": "Feng Li\\* Hao Zhang\\* Peize Sun, Xueyan Zou, Shilong Liu, Chunyuan Li Jianwei Yang1, Lei Zhang2, Jianfeng Gao" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 140, + 194, + 469, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 194, + 469, + 218 + ], + "spans": [ + { + "bbox": [ + 140, + 194, + 469, + 218 + ], + "type": "text", + "content": "§ HKUST ‡ Microsoft Research, Redmond † IDEA # HKU § UW-Madison Tsinghua * Core Contributor 1. Project Lead 2. Equal Advisory Contribution" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 159, + 247, + 455, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 247, + 455, + 489 + ], + "spans": [ + { + "bbox": [ + 159, + 247, + 455, + 489 + ], + "type": "text", + "content": "Abstract. In this work, we introduce Semantic-SAM, an augmented image segmentation foundation for segmenting and recognizing anything at desired granularities. Compared to the foundational segmentation model SAM [31], our model has two unique advantages: (i) granularity-controllability in that the model can produce segmentation masks at any desired granularities, from objects to parts to both; (ii) semantic-awareness in that the model simultaneously predicts semantic labels for masks at different granularities. To enable multi-granularity capabilities, we propose a multi-choice learning scheme, where each click point generates a set of masks at multiple levels of granularity, corresponding to a set of ground-truth masks. To achieve semantic awareness, we consolidate multiple datasets of different levels of granularity and train our model using decoupled object- and part-based tasks to facilitate knowledge sharing and transfer among different tasks. To the best of our knowledge, this work is the first attempt to jointly train a model on SA-1B, instance-level, and part-level segmentation datasets. Experimental results and visualizations demonstrate that our model successfully achieves the desired goals. Furthermore, we show that multi-task training using the segmentation task defined on SA-1B and other segmentation tasks (e.g., panoptic and part segmentation) leads to performance gains on all segmentation tasks. In particular, we achieve a new state-of-the-art in COCO panoptic segmentation 60.2 PQ by adding SAM data." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 509, + 230, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 509, + 230, + 521 + ], + "spans": [ + { + "bbox": [ + 132, + 509, + 230, + 521 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 533, + 488, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 488, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 488, + 641 + ], + "type": "text", + "content": "The universal and interactive AI systems that follow human intents have shown their potential in natural language processing [46, 47] and controllable image generation [52, 66]. However, such a universal system for pixel-level image understanding remains less explored. We argue that a foundational segmentation model should possess the following important properties: granularity-controllability and semantic-awareness. Regardless of the specific image domain or prompt context, the model is capable of acquiring a versatile representation, predicting segmentation masks in multi-granularity, and understanding the semantic meaning behind each segmented region." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Previous works [31,58,70] have strived for these properties but only achieved part of the goals. The main obstacles impeding the progress of such a universal" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 205, + 125, + 402, + 228 + ], + "blocks": [ + { + "bbox": [ + 205, + 125, + 402, + 228 + ], + "lines": [ + { + "bbox": [ + 205, + 125, + 402, + 228 + ], + "spans": [ + { + "bbox": [ + 205, + 125, + 402, + 228 + ], + "type": "image", + "image_path": "2b430727df2cf1fd4619fbf2c34f3a4d0a905b8a4dd535ff64f5d280268ff9e6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 162, + 237, + 451, + 248 + ], + "lines": [ + { + "bbox": [ + 162, + 237, + 451, + 248 + ], + "spans": [ + { + "bbox": [ + 162, + 237, + 451, + 248 + ], + "type": "text", + "content": "Fig. 1: Dataset distribution in the semantics-granularity space.remove" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 272, + 480, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 272, + 480, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 480, + 297 + ], + "type": "text", + "content": "image segmentation model can be attributed to the limited flexibility of model design and the lack of training data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 308, + 482, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 308, + 482, + 426 + ], + "spans": [ + { + "bbox": [ + 138, + 308, + 482, + 426 + ], + "type": "text", + "content": "- Model Design. The existing image segmentation model architectures are dominated by the single-input-single-output pipeline that discards any ambiguity. While this pipeline is prevalent in both anchor-based CNN architectures [24] and query-based Transformer architectures [4, 11], and has demonstrated remarkable performance in semantic, instance, and panoptic segmentation tasks [30, 39, 68], it inherently restricts the model to predict multi-granularity segmentation masks in an end-to-end manner. Although clustering postprocessing techniques [13] can produce multiple masks for a single object query, they are neither efficient nor effective solutions for a granularity-aware segmentation model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 427, + 490, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 427, + 490, + 594 + ], + "spans": [ + { + "bbox": [ + 138, + 427, + 490, + 594 + ], + "type": "text", + "content": "- Training Data. Scaling up segmentation datasets that possess both semantic-awareness and granularity-awareness is a costly endeavor. Existing generic detection and segmentation datasets such as MSCOCO [39] and Objects365 [53] offer large amounts of data and rich semantic annotations, but only at the object level. On the other hand, part segmentation datasets such as Pascal Part [9], PartImageNet [23], and PACO [49] provide more fine-grained semantic annotations, but their data volumes are limited. Recently, SAM [31] has successfully scaled up the multi-granularity mask data to millions of images, but it does not include semantic annotations. In order to achieve the dual objectives of semantic-awareness and granularity-abundance, there is a pressing need to unify model design and training on various data formats to facilitate knowledge transfer. However, the inherent differences in semantics and granularity across different datasets pose a significant challenge to joint training efforts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we introduce Semantic-SAM, a universal image segmentation model designed to enable segmenting and recognizing objects at any desired granularity. Given one click from a user, our model addresses the spatial ambiguity by predicting masks in multiple granularities, accompanied by semantic labels at both the object and part levels. As shown in Figure 2, our model gener-" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 351, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 351, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 351, + 102 + ], + "type": "text", + "content": "Authors Suppressed Due to Excessive Length" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 482, + 308 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 482, + 308 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 482, + 308 + ], + "type": "image", + "image_path": "67b2b578dfbc0c79c7dcf2547ccdab60c900c5d2500206ad4be73bc15d5afadb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 316, + 482, + 394 + ], + "lines": [ + { + "bbox": [ + 130, + 316, + 482, + 394 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 482, + 394 + ], + "type": "text", + "content": "Fig. 2: Our model is capable of dealing with various segmentation tasks including open-set and interactive segmentation. (a) Our model can accomplish instance, semantic, panoptic segmentation, and part segmentation. (b) Our model is able to output multilevel semantics in different granularities. The yellow star point on the left-most image is the user click. (c) We connect our model with an inpainting model to perform multilevel inpainting. The prompts are \"Spider-Man\" and \"BMW car\", respectively. Note that only one click is needed to produce the results in (b) and (c), respectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 421, + 480, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 421, + 480, + 444 + ], + "spans": [ + { + "bbox": [ + 130, + 421, + 480, + 444 + ], + "type": "text", + "content": "ates multi-level segmentation masks ranging from the person head to the whole truck." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 446, + 482, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 446, + 482, + 541 + ], + "spans": [ + { + "bbox": [ + 130, + 446, + 482, + 541 + ], + "type": "text", + "content": "The multi-granularity capability is achieved through a multi-choice learning design [22, 37] incorporated into the decoder architecture. Each click point is represented with multiple queries, each containing a different level of embedding. These queries are trained to learn from all available ground-truth masks representing different granularities. To establish a correspondence between multiple masks and ground-truths, we employ a many-to-many matching scheme to ensure that a single click point could generate high-quality masks in multiple granularities." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 544, + 482, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 482, + 626 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 482, + 626 + ], + "type": "text", + "content": "To accomplish semantic-awareness with a generalized capability, we introduce a decoupled approach for objects and parts classification, to allow knowledge transfer among rich object and part segmentation datasets. By leveraging a shared text encoder to encode both objects and parts independently, we adapt the loss function based on the data type. For instance, generic segmentation data lacks part classification loss, whereas SAM data does not include classification loss." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 629, + 514, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 514, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 514, + 667 + ], + "type": "text", + "content": "To enrich semantics and granularity within our model, we consolidate seven datasets on three types of granularities, including generic segmentation of MSCOCO [39], Objects365 [53], ADE20k [68], part segmentation of PASCAL Part [9], PACO [49]," + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 228, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 228, + 91, + 448, + 102 + ], + "type": "text", + "content": "Segment and Recognize Anything at Any Granularity" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 116, + 476, + 243 + ], + "blocks": [ + { + "bbox": [ + 137, + 116, + 476, + 243 + ], + "lines": [ + { + "bbox": [ + 137, + 116, + 476, + 243 + ], + "spans": [ + { + "bbox": [ + 137, + 116, + 476, + 243 + ], + "type": "image", + "image_path": "a7e5a3a04f70df4b12adceb061fea3b19c41956a8a1a6dae4b9d646b735ecc53.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 250, + 482, + 338 + ], + "lines": [ + { + "bbox": [ + 130, + 250, + 482, + 338 + ], + "spans": [ + { + "bbox": [ + 130, + 250, + 482, + 338 + ], + "type": "text", + "content": "Fig. 3: Semantic-SAM is a universal segmentation framework that can take multiple types of segmentation data including generic, part, and class-agnostic segmentation data. The vision encoder is used to extract image features. The mask decoder can do both generic segmentation and promptable segmentation with various types of prompts. For point and box, we input them via anchor boxes to the mask decoder. Since there is an ambiguity of granularity for a point input, we duplicate each point 6 times and give them different levels of embeddings. The output masks of point prompts match with multiple GT masks of different granularities." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 358, + 483, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 358, + 483, + 467 + ], + "spans": [ + { + "bbox": [ + 130, + 358, + 483, + 467 + ], + "type": "text", + "content": "PartImagenet [23], and SA-1B [31]. Their data formats are reorganized to match our training objectives accordingly. After joint training, our model obtains a strong performance across a variety of datasets. Notably, we find that learning from interactive segmentation could improve generic and part segmentation. For example, by jointly training SA-1B promptable segmentation and COCO panoptic segmentation, we achieve a gain of 2.2 box AP and a gain of 1.3 mask AP. In addition, through comprehensive experiments, we demonstrate that our model can cover more segmentation granularities than SAM, outperforming 3.4 IoU with one point click." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 486, + 244, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 486, + 244, + 498 + ], + "spans": [ + { + "bbox": [ + 132, + 486, + 244, + 498 + ], + "type": "text", + "content": "2 Semantic-SAM" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 512, + 228, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 228, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 228, + 525 + ], + "type": "text", + "content": "2.1 Mode Design" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "type": "text", + "content": "Our Semantic-SAM follows MaskDINO [33] to employ a query-based mask decoder. To produce semantic-aware and multi-granularity masks, we exploit two types of queries: (i) generic queries as used in MaskDINO and (ii) spatial queries including point and box for granularity-controllable interactive segmentation. The overall pipeline is shown in Figure 3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "We represent both point and box prompts into anchor boxes as a unified format. In particular, we convert user click point " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " into an anchor box " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "(x,y,w,h)" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " with small width " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " and height " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ", so that the anchor box can closely approximate the point. To capture different granularities of masks, each click is first encoded to position prompt and combined with " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " different content prompts, each of which is represented as a learnable embedding for a given" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 351, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 351, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 351, + 102 + ], + "type": "text", + "content": "Authors Suppressed Due to Excessive Length" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 195, + 125, + 414, + 228 + ], + "blocks": [ + { + "bbox": [ + 195, + 125, + 414, + 228 + ], + "lines": [ + { + "bbox": [ + 195, + 125, + 414, + 228 + ], + "spans": [ + { + "bbox": [ + 195, + 125, + 414, + 228 + ], + "type": "image", + "image_path": "293359e4ab4ffd0cf73dfbe792229bbfa860280f86b3bf5a3ad215afe67a0014.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 243, + 480, + 268 + ], + "lines": [ + { + "bbox": [ + 132, + 243, + 480, + 268 + ], + "spans": [ + { + "bbox": [ + 132, + 243, + 480, + 268 + ], + "type": "text", + "content": "Fig. 4: Decoupled object and part classification. Given the same decoder outputs (right), two set of embeddings are utilized for object and part classification separately." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 289, + 479, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 289, + 479, + 326 + ], + "spans": [ + { + "bbox": [ + 131, + 289, + 479, + 326 + ], + "type": "text", + "content": "granularity level. As a result, we feed into the decoder " + }, + { + "bbox": [ + 131, + 289, + 479, + 326 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 131, + 289, + 479, + 326 + ], + "type": "text", + "content": " query embeddings " + }, + { + "bbox": [ + 131, + 289, + 479, + 326 + ], + "type": "inline_equation", + "content": "\\mathbf{Q} = (\\mathbf{q}_1,\\dots ,\\mathbf{q}_K)" + }, + { + "bbox": [ + 131, + 289, + 479, + 326 + ], + "type": "text", + "content": " for a click represented as an anchor box " + }, + { + "bbox": [ + 131, + 289, + 479, + 326 + ], + "type": "inline_equation", + "content": "\\mathbf{b} = (x,y,w,h)" + }, + { + "bbox": [ + 131, + 289, + 479, + 326 + ], + "type": "text", + "content": ", and the " + }, + { + "bbox": [ + 131, + 289, + 479, + 326 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 289, + 479, + 326 + ], + "type": "text", + "content": "-th query is formulated as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 256, + 334, + 481, + 350 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 334, + 481, + 350 + ], + "spans": [ + { + "bbox": [ + 256, + 334, + 481, + 350 + ], + "type": "interline_equation", + "content": "\\mathbf {q} _ {i} = \\mathbf {q} _ {i} ^ {\\text {l e v e l}} + \\mathbf {q} ^ {\\text {a n c h o r}}, \\tag {1}", + "image_path": "36ebe5863a67d701e6bd6680203de9dd5df7d98af4847d983d24850d86822ad6.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 357, + 479, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 357, + 479, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 357, + 479, + 416 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 357, + 479, + 416 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_i^{\\mathrm{level}}" + }, + { + "bbox": [ + 130, + 357, + 479, + 416 + ], + "type": "text", + "content": " is the content embedding for granularity level " + }, + { + "bbox": [ + 130, + 357, + 479, + 416 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 357, + 479, + 416 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 357, + 479, + 416 + ], + "type": "inline_equation", + "content": "\\mathbf{q}^{\\mathrm{anchor}}" + }, + { + "bbox": [ + 130, + 357, + 479, + 416 + ], + "type": "text", + "content": " is a position embedding derived from the sine encoding of the anchor box. Here we empirically choose " + }, + { + "bbox": [ + 130, + 357, + 479, + 416 + ], + "type": "inline_equation", + "content": "K = 6" + }, + { + "bbox": [ + 130, + 357, + 479, + 416 + ], + "type": "text", + "content": ", considering there are at most 6 levels of masks per user click for the majority of images in SA-1B [31]. As for the box prompt, we simply use a single content embedding, i.e., " + }, + { + "bbox": [ + 130, + 357, + 479, + 416 + ], + "type": "inline_equation", + "content": "K = 1" + }, + { + "bbox": [ + 130, + 357, + 479, + 416 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 417, + 479, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 417, + 479, + 441 + ], + "spans": [ + { + "bbox": [ + 131, + 417, + 479, + 441 + ], + "type": "text", + "content": "Assuming that the output image feature from the vision encoder is " + }, + { + "bbox": [ + 131, + 417, + 479, + 441 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 131, + 417, + 479, + 441 + ], + "type": "text", + "content": ", the mask decoder takes a set of queries and produces the outputs:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 200, + 450, + 480, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 450, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 200, + 450, + 480, + 464 + ], + "type": "interline_equation", + "content": "\\mathbf {O} = \\operatorname {D e f o r m D e c} (\\mathbf {Q}, \\mathbf {F}) \\text {w i t h} \\mathbf {O} = \\left(\\mathbf {o} _ {1}, \\dots , \\mathbf {o} _ {K}\\right), \\tag {2}", + "image_path": "d42f3d1f9265534be435c921e17a944587e2a521a470838c900c26757fb7024d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 472, + 479, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 472, + 479, + 533 + ], + "spans": [ + { + "bbox": [ + 131, + 472, + 479, + 533 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 472, + 479, + 533 + ], + "type": "inline_equation", + "content": "\\mathsf{DeformDec}(\\cdot, \\cdot, \\cdot)" + }, + { + "bbox": [ + 131, + 472, + 479, + 533 + ], + "type": "text", + "content": " is a deformable decoder that takes anchor box queries, and image features as input to output queried features. " + }, + { + "bbox": [ + 131, + 472, + 479, + 533 + ], + "type": "inline_equation", + "content": "\\mathbf{o}_i" + }, + { + "bbox": [ + 131, + 472, + 479, + 533 + ], + "type": "text", + "content": " is the model output for the " + }, + { + "bbox": [ + 131, + 472, + 479, + 533 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 472, + 479, + 533 + ], + "type": "text", + "content": "-th input query " + }, + { + "bbox": [ + 131, + 472, + 479, + 533 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_i" + }, + { + "bbox": [ + 131, + 472, + 479, + 533 + ], + "type": "text", + "content": ". Each " + }, + { + "bbox": [ + 131, + 472, + 479, + 533 + ], + "type": "inline_equation", + "content": "\\mathbf{o}_i = (\\mathbf{c}_i, \\mathbf{m}_i)" + }, + { + "bbox": [ + 131, + 472, + 479, + 533 + ], + "type": "text", + "content": " consists of the predicted semantic category " + }, + { + "bbox": [ + 131, + 472, + 479, + 533 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_i" + }, + { + "bbox": [ + 131, + 472, + 479, + 533 + ], + "type": "text", + "content": " and mask " + }, + { + "bbox": [ + 131, + 472, + 479, + 533 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_i" + }, + { + "bbox": [ + 131, + 472, + 479, + 533 + ], + "type": "text", + "content": ", which are used to construct the concept recognition loss and mask prediction loss, respectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 550, + 204, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 550, + 204, + 562 + ], + "spans": [ + { + "bbox": [ + 132, + 550, + 204, + 562 + ], + "type": "text", + "content": "2.2 Training" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "text", + "content": "Segment any granularity. To endow the model with a multi-granularity segmentation ability, we propose a many-to-many matching method during training. We found that SAM fails to provide good multi-level segmentation results with a single click. Since SAM uses many-to-one matching [31] during training, the three SAM-predicted masks for each click point only match with one GT mask. This means that points located in masks of small levels cannot predict large masks with high quality according to our observation. In contrast, we fully leverage the structures in both data and algorithm to enable multi-level mask prediction with" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 228, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 228, + 91, + 447, + 103 + ], + "type": "text", + "content": "Segment and Recognize Anything at Any Granularity" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 476, + 182 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 476, + 182 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 476, + 182 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 476, + 182 + ], + "type": "image", + "image_path": "7045eae0504bb1930ebf07c469707350b1a5efb3e1a5e12eae41c5a2693b747b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 193, + 482, + 258 + ], + "lines": [ + { + "bbox": [ + 130, + 193, + 482, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 193, + 482, + 258 + ], + "type": "text", + "content": "Fig. 5: Comparisons of different interactive learning strategies. a) One-to-one: traditional interactive segmentation models that focus on object-level, i.e., SEEM [70]. b) Many-to-one: multi-choice learning for single granularity, i.e., SAM [31]. c) Many-to-many: our Semantic-SAM enforces the model to predict all the possible granularities of a single click for more diverse segmentation. d) As a result, our model is able to output multiple masks in richer granularities." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 285, + 482, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 482, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 482, + 346 + ], + "type": "text", + "content": "a single click. First, we reorganized the data by clustering multiple GT masks of different levels sharing the same click. To allow multiple predictions of the same click to match with the GT masks, we employ the Hungarian algorithm to enable the many-to-many matching. The similarity matrix and scores vary based on the availability of different segmentation data components." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 347, + 482, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 347, + 482, + 455 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 482, + 455 + ], + "type": "text", + "content": "For box input and generic segmentation, we follow existing methods. Specifically, to generate a mask from an input box, we follow a similar idea as in denoising training (DN) [33]. We add noises to ground-truth boxes to simulate inaccurate box inputs from users, and these noised boxes serve as spatial prompts for the decoder. The model is trained to reconstruct the original boxes and masks given noised boxes. For the content part of box prompts, we adopt a learnable token as a general prompt. Note that this is the only difference from DN, as DN uses ground-truth label embedding as the content prompts. For generic segmentation, we follow the same pipeline as in Mask DINO [33]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 472, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 605 + ], + "type": "text", + "content": "Recognize anything. We train with various types of data with different semantic annotations, in which some contain object-level annotations (COCO), some contain both object and part-level annotations (Pascal Part), and SA-1B has no semantic annotations but contains masks of all semantic levels. Note that a large number of part concepts are shared across different objects, for example, head for all animals, we aim to transfer the part concept knowledge across objects trained with only object-level annotations in our joint training. To address this discrepancy between semantic annotations and better transfer semantics of different granularity, we propose to decouple object and part recognition. As shown in Figure 4, we utilize a shared text encoder to encode objects and parts, which are used to perform object and part segmentation separately." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "While all types of segmentation data share a unified format, the loss varies for different data types. In Table 1, we summarize the loss items to construct the training objective in Semantic-SAM. It is the part-level data that bridges the gap to recognize semantic concepts between part and object levels, and it is the use of SAM data that bridges the gap to segment masks at any granularity." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 351, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 351, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 351, + 102 + ], + "type": "text", + "content": "Authors Suppressed Due to Excessive Length" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 161, + 114, + 452, + 178 + ], + "blocks": [ + { + "bbox": [ + 161, + 114, + 452, + 178 + ], + "lines": [ + { + "bbox": [ + 161, + 114, + 452, + 178 + ], + "spans": [ + { + "bbox": [ + 161, + 114, + 452, + 178 + ], + "type": "table", + "html": "
DataRecognize AnythingSegment at Any Granularity
PartObjectBox Mask #GT in Matching
SAM dataXX✓✓Many
Object-level dataX✓✓✓One
Part-level data✓✓✓✓One
", + "image_path": "f79b5a79211bc59f30be04c0bf3cd36b07dfc4e13572015902575642fafead3c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 177, + 248, + 438, + 357 + ], + "blocks": [ + { + "bbox": [ + 130, + 178, + 479, + 212 + ], + "lines": [ + { + "bbox": [ + 130, + 178, + 479, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 178, + 479, + 212 + ], + "type": "text", + "content": "Table 1: The loss items to construct the training objective in Semantic-SAM. The four loss items are part classification, object classification, box loss, and mask loss, respectively. The last column indicates the number of ground-truth masks in the matching." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 177, + 248, + 438, + 357 + ], + "lines": [ + { + "bbox": [ + 177, + 248, + 438, + 357 + ], + "spans": [ + { + "bbox": [ + 177, + 248, + 438, + 357 + ], + "type": "table", + "html": "
TypeData#ImagesSemantic ConceptGranularity Level
PartObjectPartWhole
Class-agnosticSA-1B11MXX✓✓
Object-levelObjects3651.7MX365X✓
COCO110KX133X✓
ADE20K20KX150X✓
Part-levelPACO-LVIS45K20175✓✓
PartImageNet16K1311✓✓
Pascal Part5K3020✓✓
", + "image_path": "cafa63027f2350a62d6c713ac37cac7a8722469d1b5780ee5e418af3d9dd6520.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 176, + 358, + 436, + 369 + ], + "lines": [ + { + "bbox": [ + 176, + 358, + 436, + 369 + ], + "spans": [ + { + "bbox": [ + 176, + 358, + 436, + 369 + ], + "type": "text", + "content": "Table 2: The statistics of training data used in Semantic-SAM." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 413, + 480, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 480, + 472 + ], + "type": "text", + "content": "Discussion. As shown in Figure 5, Semantic-SAM differs from previous interactive segmentation models in two aspects. Firstly, our model is trained with multi-choice learning to output all the possible segmentation masks with one click point. Secondly, our output granularities are richer to generate diverse segmentation masks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 514, + 244, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 244, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 244, + 525 + ], + "type": "text", + "content": "2.3 Data Unification" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 557, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 481, + 666 + ], + "type": "text", + "content": "In order for multi-level semantics, we include seven datasets that contain different granularity-level masks. The datasets are SA-1B [31], COCO panoptic [39], ADE20k panoptic [68], PASCAL part [44], PACO [49], PartImageNet [23], and Objects365 [53]. Within them, COCO and ADE20k panoptic datasets contain object-level masks and class labels. PASCAL part, PACO, and PartImageNet contain part-level masks and class labels. SA-1B contains up to 6-level masks without labels, while Objects365 contains abundant class labels for object-level instances. The details of these datasets are shown in Table 2. We further visualize the data distribution of different data types in Figure 1." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 228, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 228, + 91, + 447, + 102 + ], + "type": "text", + "content": "Segment and Recognize Anything at Any Granularity" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 114, + 480, + 203 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 480, + 203 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 480, + 203 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 480, + 203 + ], + "type": "table", + "html": "
MethodTypeTraining DataPQmIoUAPAPsAPmAPI box mask
boxmaskboxmaskboxmask
Mask2Former (T) [11]Close-setCOCO53.263.246.143.3-----
X-Decoder (T) [69]Open-setCOCO+VL52.662.443.641.3-----
OpenSeed (T) [65]Open-setCOCO+O36555.463.851.247.134.527.454.350.466.2 66.8
Semantic-SAM (T) (ours)Open-setCOCO54.663.250.146.134.427.153.249.466.1 66.1
Semantic-SAM (T) (ours)Open-setCOCO+SAM55.263.452.3(+2.2)47.4(+1.3)36.1(+1.7)28.3(+1.2)55.6(+2.4)50.7(+1.3)67.3 66.2
OpenSeed (L) [65]Open-setCOCO+O36559.568.658.253.2-----
Oneformer (H) [26]Open-setCOCO+O36560.068.8-52.0-----
Semantic-SAM (L) (ours)Open-setCOCO+SAM60.268.957.553.042.935.461.256.971.9 71.4
", + "image_path": "9c20188c2d8c6c7c8b9877bbe5bce2243acead9448360ccac632f2de66ab58f1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 204, + 481, + 247 + ], + "lines": [ + { + "bbox": [ + 132, + 204, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 204, + 481, + 247 + ], + "type": "text", + "content": "Table 3: Results for Semantic-SAM and other panoptic segmentation models on COCO val2017. Our model is jointly trained on COCO [8] and SA-1B [31] (1/10 data) and directly evaluates COCO.add model param size, hq-sam comparison, more datasets (ADE, cityscapes and compare with SEEM), more impressive visualization" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 133, + 285, + 229, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 285, + 229, + 300 + ], + "spans": [ + { + "bbox": [ + 133, + 285, + 229, + 300 + ], + "type": "text", + "content": "3 Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 324, + 261, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 324, + 261, + 337 + ], + "spans": [ + { + "bbox": [ + 133, + 324, + 261, + 337 + ], + "type": "text", + "content": "3.1 Experimental Setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 357, + 482, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 357, + 482, + 537 + ], + "spans": [ + { + "bbox": [ + 132, + 357, + 482, + 537 + ], + "type": "text", + "content": "Implementation Details. In our experiments, we train on SA-1b [31] with no semantic annotation, COCO [39] with object-level annotation and Pascal Part [15] with part-level annotation. We implement our model based on Mask DINO [33]. Mask DINO is a unified detection and segmentation framework that simultaneously predicts box and mask. We follow [33] to use 300 latent queries and nine decoder layers for all segmentation tasks. For the visual backbone, we adopt pre-trained Swin-T/L [41] by default. For the language backbone, we adopt the pre-trained base model in UniCL [62]. As SA-1B [31] dominates the data, during training, we first train on only SA-1B data. Then, we add object and part-level data to jointly train the three types of data. During training, the image resolution is " + }, + { + "bbox": [ + 132, + 357, + 482, + 537 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 132, + 357, + 482, + 537 + ], + "type": "text", + "content": " for all data. We use AdamW [43] as the optimizer. We use large-scale jittering for object and part-level data and did not use data augmentations for SA-1B data, as SA-1B images are abundant. We set the learning rate to 0.0001, which is decayed at 0.9 and 0.95 fractions of the total number of steps by 10." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 482, + 666 + ], + "type": "text", + "content": "Evaluation. We mainly evaluate two datasets, including COCO val2017 and a subset of SA-1B [31] with 1000 images. For evaluation metrics, we evaluate PQ and AP for generic and part segmentation datasets. For single-granularity interactive segmentation, we report 1-click mIoU (denoted as 1-IoU) under Point (Max) and Point (Oracle) settings. Max denotes we select the output mask with the maximum confidence score. Oracle denotes we select the output mask with the max IoU by calculating the IoU between the prediction and target mask. For multi-granularity interactive segmentation, we report 1-IoU@All Granularity that matches all the possible ground-truth masks for a single click to the multi-granularity predictions and then calculate the average IoU of all granularities." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 351, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 351, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 351, + 102 + ], + "type": "text", + "content": "Authors Suppressed Due to Excessive Length" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 342, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 342, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 342, + 129 + ], + "type": "text", + "content": "3.2 Semantic Segmentation of Anything" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 135, + 482, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 135, + 482, + 255 + ], + "spans": [ + { + "bbox": [ + 130, + 135, + 482, + 255 + ], + "type": "text", + "content": "Generic Segmentation. As shown in Table 3, to validate the compatibility of multi-granularity interactive segmentation and generic segmentation, we jointly train with SA-1B [31] (1/10 data) and COCO panoptic segmentation. The result indicates that interactive segmentation with SAM can significantly help the instance-level detection and segmentation with a performance improvement of " + }, + { + "bbox": [ + 130, + 135, + 482, + 255 + ], + "type": "inline_equation", + "content": "+2.2" + }, + { + "bbox": [ + 130, + 135, + 482, + 255 + ], + "type": "text", + "content": " AP on the box and " + }, + { + "bbox": [ + 130, + 135, + 482, + 255 + ], + "type": "inline_equation", + "content": "+1.3" + }, + { + "bbox": [ + 130, + 135, + 482, + 255 + ], + "type": "text", + "content": " AP on the mask. Notably, OpenSeed [65] and Semantic-SAM are both based on Mask DINO [33]. Our joint training with SA-1B even outperforms OpenSeed which is trained with Object365 [53]. In addition, adding SA-1B mainly improves small object detection (APs and APm), as there are a large number of small objects in SA-1B." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 136, + 285, + 478, + 343 + ], + "blocks": [ + { + "bbox": [ + 136, + 285, + 478, + 343 + ], + "lines": [ + { + "bbox": [ + 136, + 285, + 478, + 343 + ], + "spans": [ + { + "bbox": [ + 136, + 285, + 478, + 343 + ], + "type": "table", + "html": "
MethodTypeTraining DataAPAPsAPmAP1
boxmaskbox maskbox maskbox mask
VLPart (B) [55]Open-setPascal Part29.227.4- -- -- -
Semantic-SAM (T) (ours)Open-setPascal Part27.030.516.6 19.138.1 41.643.8 49.1
Semantic-SAM (T) (ours)Open-setPascal Part+SAM28.0(+1.0)31.4(+0.9)17.3 19.940.0 42.545.7 49.7
", + "image_path": "31444f2e9451806b9b84273ebcbeeac29b1bfb9ba05a507c177d38b8e93edb90.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 344, + 480, + 376 + ], + "lines": [ + { + "bbox": [ + 132, + 344, + 480, + 376 + ], + "spans": [ + { + "bbox": [ + 132, + 344, + 480, + 376 + ], + "type": "text", + "content": "Table 4: Results for Semantic-SAM and other part segmentation models on Pascal Part. Our model is jointly trained on Pascal Part [15] and SA-1B [31] (1/10 data) and directly evaluates Pascal Part." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 404, + 481, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 404, + 481, + 466 + ], + "spans": [ + { + "bbox": [ + 130, + 404, + 481, + 466 + ], + "type": "text", + "content": "Part Segmentation. We also validate the compatibility of joint training SA-1B (1/10 data) and part segmentation. As shown in Table 4, adding SA-1B brings a decent performance improvement on Pascal Part [15], +1.0 AP on the box and +0.9 AP on the mask. Importantly, Semantic-SAM (T) has better mask AP than VLPart (B) [55], showing the high mask quality of our model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 475, + 481, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 475, + 481, + 512 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 481, + 512 + ], + "type": "text", + "content": "Single-granularity Interactive Segmentation. In Table 5, we evaluate 1-IoU for SAM and our model on COCO val2017. Our model outperforms SAM under both Point (Max) and Point (Oracle) settings." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 521, + 482, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 606 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 606 + ], + "type": "text", + "content": "Multi-granularity Interactive Segmentation. In Table 6, we compare SAM [31] and our model on the output granularities for a single click. We adopt a Hungarian Matching to match all the possible target masks with the predicted masks for the click and calculate the average IoU score. As SAM has only three prompts, we also sample two clicks from a single mask to produce six output masks for a fair comparison. Notably, SAM has been trained on this validation set while we do not." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 622, + 209, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 209, + 634 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 209, + 634 + ], + "type": "text", + "content": "3.3 Ablations" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Match Strategy. As shown in Table 7, we compare different match strategies in our model. When using many-to-many matching to match all the possible" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 228, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 228, + 91, + 447, + 103 + ], + "type": "text", + "content": "Segment and Recognize Anything at Any Granularity" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 209, + 114, + 405, + 193 + ], + "blocks": [ + { + "bbox": [ + 209, + 114, + 405, + 193 + ], + "lines": [ + { + "bbox": [ + 209, + 114, + 405, + 193 + ], + "spans": [ + { + "bbox": [ + 209, + 114, + 405, + 193 + ], + "type": "table", + "html": "
MethodCOCO
Point (Max) Point (Oracle)
1-IoU1-IoU
SAM (B)52.168.2
SAM (L)55.770.5
Semantic-SAM (T)54.573.8
Semantic-SAM (L)57.074.2
", + "image_path": "785195d79969182c9fb90a45a5b78ad9e91a7c61afd8415534b9639b578eff11.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 195, + 269, + 418, + 338 + ], + "blocks": [ + { + "bbox": [ + 130, + 193, + 482, + 248 + ], + "lines": [ + { + "bbox": [ + 130, + 193, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 193, + 482, + 248 + ], + "type": "text", + "content": "Table 5: Comparison with previous models on point interactions. Both SAM [31] and our model are trained on only SA-1B and directly evaluate on COCO val2017 for fair comparison. Max denotes we select the output with the max confidence score prediction. Oracle denotes we select the output with the max IoU by calculating the IoU between the prediction and target mask." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 195, + 269, + 418, + 338 + ], + "lines": [ + { + "bbox": [ + 195, + 269, + 418, + 338 + ], + "spans": [ + { + "bbox": [ + 195, + 269, + 418, + 338 + ], + "type": "table", + "html": "
MethodGranularity1-IoU@All Granularity
SAM (B)†375.6
SAM (L)†382.5
SAM (H)†383.5
Semantic-SAM(T)688.1
Semantic-SAM(L)689.0
", + "image_path": "89aabf222456e1a50bbc09f56252c8407cdbd162811d7bd093cd1870b5dfbb00.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 338, + 482, + 392 + ], + "lines": [ + { + "bbox": [ + 130, + 338, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 482, + 392 + ], + "type": "text", + "content": "Table 6: Mask granularity comparison between SAM and our model. We evaluated both on an excluded subset of 1000 images from SA-1B. For each click, we evaluate all the possible ground-truth masks to calculate the 1-IoU@All Granularity. SAM [31] and Semantic-SAM adopts three and six prompts for a single click of a mask, respectively. " + }, + { + "bbox": [ + 130, + 338, + 482, + 392 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 130, + 338, + 482, + 392 + ], + "type": "text", + "content": " denotes that SAM has been trained on this validation subset while we do not." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 427, + 481, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 427, + 481, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 427, + 481, + 464 + ], + "type": "text", + "content": "ground-truth masks for each click, the 1-IoU@All Granularity performance is significantly improved. This validates our matching strategy is effective in learning complete granularities." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 478, + 481, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 478, + 481, + 515 + ], + "spans": [ + { + "bbox": [ + 130, + 478, + 481, + 515 + ], + "type": "text", + "content": "Box Interactive Evaluation. We also evaluate the 1-IoU given boxes in Table 8. We achieve better performance compared with object-level interactive segmentation model SEEM [70] and multi-granularity model SAM [31]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 527, + 481, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 527, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 130, + 527, + 481, + 588 + ], + "type": "text", + "content": "Increasing SA-1B Training Data. In Table 9, we show the performance improvement on COCO val 2017 when training with more SA-1B data. The performance is saturated after using more than " + }, + { + "bbox": [ + 130, + 527, + 481, + 588 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 130, + 527, + 481, + 588 + ], + "type": "text", + "content": " of the total data. It indicates that we do not need to train with the whole SA-1B data to get a good zero-shot performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 607, + 227, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 607, + 227, + 618 + ], + "spans": [ + { + "bbox": [ + 131, + 607, + 227, + 618 + ], + "type": "text", + "content": "3.4 Visualization" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "text", + "content": "We conducted a comparative analysis between our model and SAM to showcase the superior performance of our approach in generating high-quality masks, as depicted in Figure 6. The following key properties are highlighted:" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 351, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 351, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 351, + 102 + ], + "type": "text", + "content": "Authors Suppressed Due to Excessive Length" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 119, + 305, + 164 + ], + "blocks": [ + { + "bbox": [ + 136, + 119, + 305, + 164 + ], + "lines": [ + { + "bbox": [ + 136, + 119, + 305, + 164 + ], + "spans": [ + { + "bbox": [ + 136, + 119, + 305, + 164 + ], + "type": "table", + "html": "
MethodMatch1-IoU
Semantic-SAM(T)m2o73.2
Semantic-SAM(T)m2m88.1
", + "image_path": "643e94acfc69cd02f471d4c3432be36c4269ff490b87e5a94c478416742b3c7b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 325, + 114, + 479, + 171 + ], + "blocks": [ + { + "bbox": [ + 132, + 164, + 307, + 186 + ], + "lines": [ + { + "bbox": [ + 132, + 164, + 307, + 186 + ], + "spans": [ + { + "bbox": [ + 132, + 164, + 307, + 186 + ], + "type": "text", + "content": "Table 7: Effect of match strategy on segmentation performance." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 325, + 114, + 479, + 171 + ], + "lines": [ + { + "bbox": [ + 325, + 114, + 479, + 171 + ], + "spans": [ + { + "bbox": [ + 325, + 114, + 479, + 171 + ], + "type": "table", + "html": "
MethodBox 1-IoU
SAM [31](B)50.7
SEEM [70](T)73.7
Semantic-SAM(T)76.1
", + "image_path": "8cf06dc13ad3f0ed87898931b2ddb7e085654ec8ddd8ec43cb058e622231d3fe.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 170, + 211, + 444, + 306 + ], + "blocks": [ + { + "bbox": [ + 323, + 171, + 481, + 190 + ], + "lines": [ + { + "bbox": [ + 323, + 171, + 481, + 190 + ], + "spans": [ + { + "bbox": [ + 323, + 171, + 481, + 190 + ], + "type": "text", + "content": "Table 8: Box 1-IoU evaluation on COCO val2017." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 170, + 211, + 444, + 306 + ], + "lines": [ + { + "bbox": [ + 170, + 211, + 444, + 306 + ], + "spans": [ + { + "bbox": [ + 170, + 211, + 444, + 306 + ], + "type": "table", + "html": "
MethodData Portion of SA-1BCOCO
Point (Max) Point (Oracle)
1-IoU1-IoU
SAM (L)100%55.770.5
Semantic-SAM (L)3%55.273.5
Semantic-SAM (L)15%56.773.6
Semantic-SAM (L)30%55.773.7
Semantic-SAM (L)50%55.373.9
Semantic-SAM (L)100%57.074.2
", + "image_path": "3f3b0ae86ad7e1bae8a940ad787837c6455e8be297baddd70117267f853518e5.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 306, + 480, + 327 + ], + "lines": [ + { + "bbox": [ + 132, + 306, + 480, + 327 + ], + "spans": [ + { + "bbox": [ + 132, + 306, + 480, + 327 + ], + "type": "text", + "content": "Table 9: Models trained with different portions of SA-1B [31], and directly evaluated on COCO val2017." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 138, + 354, + 474, + 582 + ], + "blocks": [ + { + "bbox": [ + 138, + 354, + 474, + 582 + ], + "lines": [ + { + "bbox": [ + 138, + 354, + 474, + 582 + ], + "spans": [ + { + "bbox": [ + 138, + 354, + 474, + 582 + ], + "type": "image", + "image_path": "848f5cf19c3d946c1951de6cd2a2d92e473ace4d8e93f04af2922ede1bc78bd4.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 598, + 482, + 632 + ], + "lines": [ + { + "bbox": [ + 132, + 598, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 598, + 482, + 632 + ], + "type": "text", + "content": "Fig. 6: Given the user clicks shown in the first column, odd rows and even rows show the generated masks by SAM and our model, respectively. We removed the duplicated masks from our model to simplify the visualizations. More results are in supplementary." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 228, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 228, + 91, + 447, + 102 + ], + "type": "text", + "content": "Segment and Recognize Anything at Any Granularity" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 114, + 476, + 235 + ], + "blocks": [ + { + "bbox": [ + 138, + 114, + 476, + 235 + ], + "lines": [ + { + "bbox": [ + 138, + 114, + 476, + 235 + ], + "spans": [ + { + "bbox": [ + 138, + 114, + 476, + 235 + ], + "type": "image", + "image_path": "03de6788d13da82b10ef4c0202c426e1c65f5fc9e2ec21ce9e3e4f311b5aa2e8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 245, + 482, + 301 + ], + "lines": [ + { + "bbox": [ + 130, + 245, + 482, + 301 + ], + "spans": [ + { + "bbox": [ + 130, + 245, + 482, + 301 + ], + "type": "text", + "content": "Fig. 7: Our model learns an emergent capability that each level prompt corresponds to a certain granularity of mask. Given a single point click (red star) at the first column, our model produces masks with increasing areas from left to right for different images. A coherent order of granularities are observed for the prompts 1-6 across different images." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 340, + 482, + 574 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 130, + 340, + 482, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 482, + 401 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 482, + 401 + ], + "type": "text", + "content": "- Multi-Level Masks. Our model excels in producing masks with a greater range of meaningful granularities. Unlike SAM, which typically outputs a maximum of three masks with occasional duplication, our model generates a more diverse set of masks at various levels, enhancing the overall flexibility and precision of mask outputs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 410, + 482, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 410, + 482, + 471 + ], + "spans": [ + { + "bbox": [ + 130, + 410, + 482, + 471 + ], + "type": "text", + "content": "- Mask Qualities. Our model has demonstrated its ability to generate masks of higher quality compared to SAM. SAM occasionally produces masks with artifacts such as holes or islands, particularly for large regions when the click falls within an interior region. In contrast, our model consistently generates high-quality masks across all levels, offering accurate segmentation results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 479, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 482, + 529 + ], + "type": "text", + "content": "- Comparison with SA-1B Ground-truth Granularity. We extend the granularity of SAM data by producing masks with more meaningful levels of detail. Our model extends the original number of granularity levels, providing finer distinctions and more precise segmentation results on SAM data." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 537, + 482, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 537, + 482, + 574 + ], + "spans": [ + { + "bbox": [ + 130, + 537, + 482, + 574 + ], + "type": "text", + "content": "- Comparison with SAM in segmenting everything. We compare with SAM to segment everything in one image with the same number of points. We output more granularities for the image." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Query semantics. We also find that each point content prompt embedding learns to correspond to a fixed granularity. As shown in Fig. 7, when we visualize masks in a specific order of the corresponding content embeddings, the masks follow the order from small to large in each row consistently. This proves that each content embedding represents a semantic granularity level in our model." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 351, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 351, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 351, + 102 + ], + "type": "text", + "content": "Authors Suppressed Due to Excessive Length" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 114, + 462, + 195 + ], + "blocks": [ + { + "bbox": [ + 149, + 114, + 462, + 195 + ], + "lines": [ + { + "bbox": [ + 149, + 114, + 462, + 195 + ], + "spans": [ + { + "bbox": [ + 149, + 114, + 462, + 195 + ], + "type": "image", + "image_path": "3c4bdeb0fbcaf61596925ba4b964b57a4b80ab627b14449f40b83eda4b7d440a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 200, + 459, + 211 + ], + "lines": [ + { + "bbox": [ + 153, + 200, + 459, + 211 + ], + "spans": [ + { + "bbox": [ + 153, + 200, + 459, + 211 + ], + "type": "text", + "content": "Fig. 8: Compare SAM and our model to segment everything in one image." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 222, + 240, + 235 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 222, + 240, + 235 + ], + "spans": [ + { + "bbox": [ + 132, + 222, + 240, + 235 + ], + "type": "text", + "content": "4 Related works" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 245, + 272, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 245, + 272, + 257 + ], + "spans": [ + { + "bbox": [ + 132, + 245, + 272, + 257 + ], + "type": "text", + "content": "4.1 Generic Segmentation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 262, + 482, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 262, + 482, + 419 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 482, + 419 + ], + "type": "text", + "content": "Segmenting visual concepts is well-documented within the expansive field of computer vision [16,17,45,71]. Broad segmentation methodologies comprise several subdivisions, such as instance segmentation, semantic segmentation, and panoptic segmentation [6,24,30], each catering to a unique semantic degree. For example, semantic segmentation's goal is to detect and assign a label to each pixel in an image according to its corresponding semantic class [7,11,42]. Conversely, instance segmentation seeks to cluster pixels associated with the same semantic class into distinct object instances [3,24,33]. Panoptic segmentation is the hybrid of these two tasks. Recently, Transformer-based methods [4,56] have contributed to significant progress in segmentation tasks [11, 26, 33, 38, 64]. Generic object detection and segmentation have led to the development of abundant datasets, such as MSCOCO [39], LVIS [21], Objects365 [53], PASCAL [15], CityScapes [12], ADE20k [68]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 433, + 256, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 433, + 256, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 433, + 256, + 445 + ], + "type": "text", + "content": "4.2 Part Segmentation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "content": "Beyond generic segmentation, part segmentation aims to more fine-grained visual understanding. Most early works were bottom-up methods by grouping superpixels into parts and then objects [1, 2, 20]. Later, based on high-performance object detection networks [24, 51], top-down methods were developed by firstly detecting an object and then parsing it to part segmentation [27, 34, 63]. To segment the scene in multi-granularity, part-aware panoptic segmentation [13] is introduced. PPS [13] establishes the baseline by assembling panoptic and part segmentation models. JPPF [25] simplifies the model by a shared image encoder for both panoptic segmentation and part segmentation. By representing things, stuffs, and parts as object queries, Panoptic-PartFormer [35] proposes a unified architecture based on Transformer. While part segmentation data is much more expensive than object detection and segmentation data, a number of public datasets are available. Datasets for specific domains include cars [54], birds [57], and fashion [29]. General objects include Pascal-Part [9], PartImageNet [23], ADE20K [67], Cityscapes-Panoptic-Parts [44], and PACO [49]. More recently, SAM [31] provides a large-scale multi-granularity class-agnostic segmentation dataset. Our work is jointly trained on these datasets and contributes to a multi-granularity segmentation model." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 228, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 228, + 91, + 448, + 103 + ], + "type": "text", + "content": "Segment and Recognize Anything at Any Granularity" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 101 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 101 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 320, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 320, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 320, + 129 + ], + "type": "text", + "content": "4.3 Open-Vocabulary Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 488, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 488, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 488, + 319 + ], + "type": "text", + "content": "While generic segmentation and part segmentation have made remarkable progress, they can only segment the image in a close-set vocabulary. To expand the vocabulary size, recent works leverage the visual-semantic knowledge from large-scale foundation models like CLIP [48], ALIGN [28] and Diffusion models [60] to various segmentation tasks. LSeg [32], OpenSeg [18], GroupViT [59] achieves open-vocabulary semantic segmentation ability on ADE20K and PASCAL. Dense-CLIP [50] and MaskCLIP [14] achieve open-vocabulary instance and panoptic segmentation on COCO dataset. More recently, X-Decoder [69] proposes a unified approach to tackle various segmentation and vision-language tasks for open-vocabulary segmentation, OpenSeeD [65] proposes to use a large amount of detection data and a joint training method to improve segmentation. To segment open-vocabulary masks in part-level, VLPart [55] leverages three part segmentation datasets and learns from the dense correspondence [5] between base objects and novel objects. Our work unifies these tasks into one architecture and builds up open-vocabulary segmentation in multi-granularity." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 340, + 288, + 353 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 340, + 288, + 353 + ], + "spans": [ + { + "bbox": [ + 132, + 340, + 288, + 353 + ], + "type": "text", + "content": "4.4 Interactive Segmentation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 364, + 483, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 364, + 483, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 364, + 483, + 472 + ], + "type": "text", + "content": "Interactive segmentation refers to the process of separating objects by actively integrating user inputs. This enduring challenge has seen notable advancements [10, 19, 31, 36, 40, 61]. Previous works only focus on a small set of data or semantic-agnostic instance masks. Recently, SAM [31] enlarges the training data from 0.12M COCO images to 10M SAM fine-grained images. SEEM [70] enriches the modality to language and function to both generic and grounded segmentation with an impressive composition. However, they focus either enriching the mask granularity or enhancing the semantic understanding. Our work is the first that reconciles both in a single unified model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 493, + 220, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 493, + 220, + 506 + ], + "spans": [ + { + "bbox": [ + 132, + 493, + 220, + 506 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we have presented Semantic-SAM, which can segment and recognize anything at desired granularities. Apart from performing generic open-vocabulary segmentation, Semantic-SAM demonstrates the advantages of semantic awareness and granularity abundance. To achieve such advantages, we propose improvements on data, model, and training where we utilize datasets from multiple granularity and semantic levels, multi-choice learning for training, and a universal framework for modeling. Comprehensive experiments and visualizations have verified the semantic awareness and granularity abundance are achieved in our model. Further, Semantic-SAM is the first successful attempt to jointly train on SA-1B and other classic segmentation datasets, and shows that training with SA-1B improves other tasks such as panoptic and part segmentation." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 351, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 351, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 351, + 102 + ], + "type": "text", + "content": "Authors Suppressed Due to Excessive Length" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 137, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 137, + 481, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 137, + 481, + 169 + ], + "spans": [ + { + "bbox": [ + 138, + 137, + 481, + 169 + ], + "type": "text", + "content": "1. Pablo Arbelaez, Michael Maire, Charless Fowlkes, and Jitendra Malik. Contour detection and hierarchical image segmentation. IEEE transactions on pattern analysis and machine intelligence, 33(5):898-916, 2010." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 170, + 481, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 170, + 481, + 201 + ], + "spans": [ + { + "bbox": [ + 138, + 170, + 481, + 201 + ], + "type": "text", + "content": "2. Pablo Arbeláez, Jordi Pont-Tuset, Jonathan T Barron, Ferran Marques, and Jitendra Malik. Multiscale combinatorial grouping. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 328-335, 2014." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 201, + 481, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 201, + 481, + 232 + ], + "spans": [ + { + "bbox": [ + 138, + 201, + 481, + 232 + ], + "type": "text", + "content": "3. Daniel Bolya, Chong Zhou, Fanyi Xiao, and Yong Jae Lee. Yolact: Real-time instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9157-9166, 2019." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 232, + 481, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 232, + 481, + 263 + ], + "spans": [ + { + "bbox": [ + 138, + 232, + 481, + 263 + ], + "type": "text", + "content": "4. Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision, pages 213-229. Springer, 2020." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 264, + 481, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 264, + 481, + 305 + ], + "spans": [ + { + "bbox": [ + 138, + 264, + 481, + 305 + ], + "type": "text", + "content": "5. Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 306, + 481, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 306, + 481, + 348 + ], + "spans": [ + { + "bbox": [ + 138, + 306, + 481, + 348 + ], + "type": "text", + "content": "6. Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834-848, 2017." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 349, + 481, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 349, + 481, + 379 + ], + "spans": [ + { + "bbox": [ + 138, + 349, + 481, + 379 + ], + "type": "text", + "content": "7. Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587. 2017." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 380, + 481, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 380, + 481, + 411 + ], + "spans": [ + { + "bbox": [ + 138, + 380, + 481, + 411 + ], + "type": "text", + "content": "8. Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dolkar, and C. Lawrence Zitnick. Microsoft COCO captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 412, + 481, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 412, + 481, + 442 + ], + "spans": [ + { + "bbox": [ + 138, + 412, + 481, + 442 + ], + "type": "text", + "content": "9. Xianjie Chen, Roozbeh Mottaghi, Xiaobai Liu, Sanja Fidler, Raquel Urtasun, and Alan Yuille. Detect what you can: Detecting and representing objects using holistic models and body parts, 2014." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 134, + 443, + 481, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 443, + 481, + 483 + ], + "spans": [ + { + "bbox": [ + 134, + 443, + 481, + 483 + ], + "type": "text", + "content": "10. Xi Chen, Zhiyan Zhao, Yilei Zhang, Manni Duan, Donglian Qi, and Hengshuang Zhao. Focalclick: towards practical interactive image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1300-1309, 2022." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 484, + 481, + 526 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 484, + 481, + 526 + ], + "spans": [ + { + "bbox": [ + 134, + 484, + 481, + 526 + ], + "type": "text", + "content": "11. Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290-1299, 2022." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 527, + 481, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 527, + 481, + 579 + ], + "spans": [ + { + "bbox": [ + 134, + 527, + 481, + 579 + ], + "type": "text", + "content": "12. Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213-3223, 2016." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 580, + 481, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 580, + 481, + 612 + ], + "spans": [ + { + "bbox": [ + 134, + 580, + 481, + 612 + ], + "type": "text", + "content": "13. Daan de Geus, Panagiotis Meletis, Chenyang Lu, Xiaoxiao Wen, and Gijs Dubbelman. Part-aware panoptic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5485-5494, 2021." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 613, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 613, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 613, + 481, + 632 + ], + "type": "text", + "content": "14. Zheng Ding, Jieke Wang, and Zhuowen Tu. Open-vocabulary panoptic segmentation with maskclip. arXiv preprint arXiv:2208.08984, 2022." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 632, + 481, + 665 + ], + "type": "text", + "content": "15. Mark Everingham and John Winn. The pascal visual object classes challenge 2012 (voc2012) development kit. Pattern Analysis, Statistical Modelling and Computational Learning, Tech. Rep, 8(5), 2011." + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 228, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 228, + 91, + 447, + 102 + ], + "type": "text", + "content": "Segment and Recognize Anything at Any Granularity" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 148 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 148 + ], + "type": "text", + "content": "16. Pedro F Felzenszwalb, Ross B Girshick, David McAllester, and Deva Ramanan. Object detection with discriminatively trained part-based models. IEEE transactions on pattern analysis and machine intelligence, 32(9):1627-1645, 2009." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 480, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 480, + 169 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 480, + 169 + ], + "type": "text", + "content": "17. King-Sun Fu and JK Mui. A survey on image segmentation. Pattern recognition, 13(1):3-16, 1981." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 170, + 480, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 170, + 480, + 190 + ], + "spans": [ + { + "bbox": [ + 133, + 170, + 480, + 190 + ], + "type": "text", + "content": "18. Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Open-vocabulary image segmentation. arXiv preprint arXiv:2112.12143, 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 190, + 480, + 211 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 190, + 480, + 211 + ], + "spans": [ + { + "bbox": [ + 133, + 190, + 480, + 211 + ], + "type": "text", + "content": "19. Leo Grady. Random walks for image segmentation. IEEE transactions on pattern analysis and machine intelligence, 28(11):1768-1783, 2006." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 211, + 480, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 211, + 480, + 243 + ], + "spans": [ + { + "bbox": [ + 132, + 211, + 480, + 243 + ], + "type": "text", + "content": "20. Matthias Grundmann, Vivek Kwatra, Mei Han, and Irfan Essa. Efficient hierarchical graph-based video segmentation. In 2010 IEEE computer society conference on computer vision and pattern recognition, pages 2141-2148. IEEE, 2010." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 243, + 480, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 243, + 480, + 274 + ], + "spans": [ + { + "bbox": [ + 132, + 243, + 480, + 274 + ], + "type": "text", + "content": "21. Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5356-5364, 2019." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 274, + 480, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 480, + 306 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 480, + 306 + ], + "type": "text", + "content": "22. Abner Guzman-Rivera, Dhruv Batra, and Pushmeet Kohli. Multiple choice learning: Learning to produce multiple structured outputs. Advances in neural information processing systems, 25, 2012." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 306, + 480, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 306, + 480, + 337 + ], + "spans": [ + { + "bbox": [ + 132, + 306, + 480, + 337 + ], + "type": "text", + "content": "23. Ju He, Shuo Yang, Shaokang Yang, Adam Kortylewski, Xiaoding Yuan, Jie-Neng Chen, Shuai Liu, Cheng Yang, and Alan Yuille. Partimagenet: A large, high-quality dataset of parts. arXiv preprint arXiv:2112.00933, 2021." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 337, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 337, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 337, + 480, + 369 + ], + "type": "text", + "content": "24. Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 369, + 480, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 389 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 389 + ], + "type": "text", + "content": "25. Sravan Kumar Jagadeesh, René Schuster, and Didier Stricker. Multi-task fusion for efficient panoptic-part segmentation. arXiv preprint arXiv:2212.07671, 2022." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 389, + 480, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 389, + 480, + 421 + ], + "spans": [ + { + "bbox": [ + 132, + 389, + 480, + 421 + ], + "type": "text", + "content": "26. Jitesh Jain, Jiachen Li, MangTik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. Oneformer: One transformer to rule universal image segmentation. arXiv preprint arXiv:2211.06220, 2022." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 421, + 480, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 421, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 421, + 480, + 464 + ], + "type": "text", + "content": "27. Ruyi Ji, Dawei Du, Libo Zhang, Longyin Wen, Yanjun Wu, Chen Zhao, Feiyue Huang, and Siwei Lyu. Learning semantic neural tree for human parsing. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIII 16, pages 205-221. Springer, 2020." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 464, + 480, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 480, + 495 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 480, + 495 + ], + "type": "text", + "content": "28. Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V Le, Yunhsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 495, + 480, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 495, + 480, + 548 + ], + "spans": [ + { + "bbox": [ + 132, + 495, + 480, + 548 + ], + "type": "text", + "content": "29. Menglin Jia, Mengyun Shi, Mikhail Sirotenko, Yin Cui, Claire Cardie, Bharath Hariharan, Hartwig Adam, and Serge Belongie. Fashionpedia: Ontology, segmentation, and an attribute localization dataset. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 316-332. Springer, 2020." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 548, + 480, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 548, + 480, + 580 + ], + "spans": [ + { + "bbox": [ + 132, + 548, + 480, + 580 + ], + "type": "text", + "content": "30. Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dólar. Panoptic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9404-9413, 2019." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 580, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 580, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 580, + 480, + 611 + ], + "type": "text", + "content": "31. Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything, 2023." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "text", + "content": "32. Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and René Ranftl. Language-driven semantic segmentation. arXiv preprint arXiv:2201.03546, 2022." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "33. Feng Li, Hao Zhang, Shilong Liu, Lei Zhang, Lionel M Ni, Heung-Yeung Shum, et al. Mask dino: Towards a unified transformer-based framework for object detection and segmentation. arXiv preprint arXiv:2206.02777, 2022." + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 351, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 351, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 351, + 102 + ], + "type": "text", + "content": "Authors Suppressed Due to Excessive Length" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 117, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 480, + 137 + ], + "type": "text", + "content": "34. Qizhu Li, Anurag Arnab, and Philip HS Torr. Holistic, instance-level human parsing. arXiv preprint arXiv:1709.03612, 2017." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 137, + 480, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 137, + 480, + 178 + ], + "spans": [ + { + "bbox": [ + 132, + 137, + 480, + 178 + ], + "type": "text", + "content": "35. Xiangtai Li, Shilin Xu, Yibo Yang, Guangliang Cheng, Yunhai Tong, and Dacheng Tao. Panoptic-partformer: Learning a unified model for panoptic part segmentation. In European Conference on Computer Vision, pages 729-747. Springer, 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 179, + 480, + 200 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 179, + 480, + 200 + ], + "spans": [ + { + "bbox": [ + 132, + 179, + 480, + 200 + ], + "type": "text", + "content": "36. Yin Li, Jian Sun, Chi-Keung Tang, and Heung-Yeung Shum. Lazy snapping. ACM Transactions on Graphics (ToG), 23(3):303-308, 2004." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 200, + 480, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 200, + 480, + 231 + ], + "spans": [ + { + "bbox": [ + 132, + 200, + 480, + 231 + ], + "type": "text", + "content": "37. Zhuwen Li, Qifeng Chen, and Vladlen Koltun. Interactive image segmentation with latent diversity. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 577-585, 2018." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 231, + 480, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 231, + 480, + 273 + ], + "spans": [ + { + "bbox": [ + 132, + 231, + 480, + 273 + ], + "type": "text", + "content": "38. Zhiqi Li, Wenhai Wang, Enze Xie, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, Ping Luo, and Tong Lu. Panoptic segformer: Delving deeper into panoptic segmentation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1280-1289, 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 273, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 273, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 273, + 480, + 303 + ], + "type": "text", + "content": "39. Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 303, + 480, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 303, + 480, + 334 + ], + "spans": [ + { + "bbox": [ + 132, + 303, + 480, + 334 + ], + "type": "text", + "content": "40. Qin Liu, Zhenlin Xu, Gedas Bertasius, and Marc Niethammer. Simpleclick: Interactive image segmentation with simple vision transformers. arXiv preprint arXiv:2210.11006, 2022." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 334, + 480, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 334, + 480, + 376 + ], + "spans": [ + { + "bbox": [ + 132, + 334, + 480, + 376 + ], + "type": "text", + "content": "41. Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 376, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 376, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 376, + 480, + 407 + ], + "type": "text", + "content": "42. Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3431-3440, 2015." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 407, + 480, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 480, + 427 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 480, + 427 + ], + "type": "text", + "content": "43. Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 427, + 480, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 427, + 480, + 458 + ], + "spans": [ + { + "bbox": [ + 132, + 427, + 480, + 458 + ], + "type": "text", + "content": "44. Panagiotis Meletis, Xiaoxiao Wen, Chenyang Lu, Daan de Geus, and Gijs Dubbelman. Cityscapes-panoptic-parts and pascal-panoptic-parts datasets for scene understanding. arXiv preprint arXiv:2004.07944, 2020." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 458, + 480, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 458, + 480, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 458, + 480, + 489 + ], + "type": "text", + "content": "45. Shervin Minaee, Yuri Y Boykov, Fatih Porikli, Antonio J Plaza, Nasser Kehtarnavaz, and Demetri Terzopoulos. Image segmentation using deep learning: A survey. IEEE transactions on pattern analysis and machine intelligence, 2021." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 489, + 423, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 489, + 423, + 498 + ], + "spans": [ + { + "bbox": [ + 132, + 489, + 423, + 498 + ], + "type": "text", + "content": "46. OpenAI. Chatgpt. https://openai.com/blog/chatgpt, 2022." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 498, + 306, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 306, + 507 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 306, + 507 + ], + "type": "text", + "content": "47. OpenAI. Gpt-4 technical report, 2023." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 507, + 480, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 507, + 480, + 548 + ], + "spans": [ + { + "bbox": [ + 132, + 507, + 480, + 548 + ], + "type": "text", + "content": "48. Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 548, + 480, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 548, + 480, + 591 + ], + "spans": [ + { + "bbox": [ + 132, + 548, + 480, + 591 + ], + "type": "text", + "content": "49. Vignesh Ramanathan, Anmol Kalia, Vladan Petrovic, Yi Wen, Baixue Zheng, Baishan Guo, Rui Wang, Aaron Marquez, Rama Kovvuri, Abhishek Kadian, Amir Mousavi, Yiwen Song, Abhimanyu Dubey, and Dhruv Mahajan. PACO: Parts and attributes of common objects. In arXiv preprint arXiv:2301.01795, 2023." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 591, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 591, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 591, + 480, + 632 + ], + "type": "text", + "content": "50. Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18082-18091, 2022." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "51. Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28:91-99." + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 228, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 228, + 91, + 447, + 102 + ], + "type": "text", + "content": "Segment and Recognize Anything at Any Granularity" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 664 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 117, + 481, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 481, + 159 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 481, + 159 + ], + "type": "text", + "content": "52. Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 160, + 481, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 481, + 203 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 481, + 203 + ], + "type": "text", + "content": "53. Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8430-8439, 2019." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 203, + 481, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 203, + 481, + 245 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 481, + 245 + ], + "type": "text", + "content": "54. Xibin Song, Peng Wang, Dingfu Zhou, Rui Zhu, Chenye Guan, Yuchao Dai, Hao Su, Hongdong Li, and Ruigang Yang. Apollocar3d: A large 3d car instance understanding benchmark for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5452-5462, 2019." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 245, + 481, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 245, + 481, + 266 + ], + "spans": [ + { + "bbox": [ + 132, + 245, + 481, + 266 + ], + "type": "text", + "content": "55. Peize Sun, Shoufa Chen, Chenchen Zhu, Fanyi Xiao, Ping Luo, Saining Xie, and Zhicheng Yan. Going denser with open-vocabulary part segmentation, 2023." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 266, + 481, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 266, + 481, + 297 + ], + "spans": [ + { + "bbox": [ + 132, + 266, + 481, + 297 + ], + "type": "text", + "content": "56. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 297, + 481, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 297, + 481, + 317 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 481, + 317 + ], + "type": "text", + "content": "57. Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. technical report, 2011." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 318, + 481, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 318, + 481, + 349 + ], + "spans": [ + { + "bbox": [ + 132, + 318, + 481, + 349 + ], + "type": "text", + "content": "58. Xinlong Wang, Xiaosong Zhang, Yue Cao, Wen Wang, Chunhua Shen, and Tiejun Huang. Seggpt: Segmenting everything in context. arXiv preprint arXiv:2304.03284, 2023." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 349, + 481, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 349, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 349, + 481, + 392 + ], + "type": "text", + "content": "59. Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, and Xiaolong Wang. Groupvit: Semantic segmentation emerges from text supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18134-18144, 2022." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 392, + 481, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 481, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 481, + 423 + ], + "type": "text", + "content": "60. Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. arXiv preprint arXiv:2303.04803, 2023." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 423, + 481, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 423, + 481, + 455 + ], + "spans": [ + { + "bbox": [ + 132, + 423, + 481, + 455 + ], + "type": "text", + "content": "61. Ning Xu, Brian Price, Scott Cohen, Jimei Yang, and Thomas S Huang. Deep interactive object selection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 373-381, 2016." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 455, + 481, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 455, + 481, + 485 + ], + "spans": [ + { + "bbox": [ + 132, + 455, + 481, + 485 + ], + "type": "text", + "content": "62. Jianwei Yang, Chunyuan Li, Pengchuan Zhang, Bin Xiao, Ce Liu, Lu Yuan, and Jianfeng Gao. Unified contrastive learning in image-text-label space. In CVPR, 2022." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 486, + 481, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 486, + 481, + 517 + ], + "spans": [ + { + "bbox": [ + 132, + 486, + 481, + 517 + ], + "type": "text", + "content": "63. Lu Yang, Qing Song, Zhihui Wang, and Ming Jiang. Parsing r-cnn for instance-level human analysis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 364-373, 2019." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 517, + 481, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 517, + 481, + 548 + ], + "spans": [ + { + "bbox": [ + 132, + 517, + 481, + 548 + ], + "type": "text", + "content": "64. Hao Zhang, Feng Li, Huaizhe Xu, Shijia Huang, Shilong Liu, Lionel M Ni, and Lei Zhang. Mp-former: Mask-piloted transformer for image segmentation. arXiv preprint arXiv:2303.07336, 2023." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 548, + 481, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 548, + 481, + 581 + ], + "spans": [ + { + "bbox": [ + 132, + 548, + 481, + 581 + ], + "type": "text", + "content": "65. Hao Zhang, Feng Li, Xueyan Zou, Shilong Liu, Chunyuan Li, Jianfeng Gao, Jianwei Yang, and Lei Zhang. A simple framework for open-vocabulary segmentation and detection. arXiv preprint arXiv:2303.08131, 2023." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 581, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 581, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 581, + 481, + 601 + ], + "type": "text", + "content": "66. Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543, 2023." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "type": "text", + "content": "67. Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 633, + 481, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 481, + 664 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 481, + 664 + ], + "type": "text", + "content": "68. Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset, 2018." + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 351, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 351, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 351, + 102 + ], + "type": "text", + "content": "Authors Suppressed Due to Excessive Length" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 201 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "text", + "content": "69. Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. arXiv preprint arXiv:2212.11270, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 149, + 482, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 482, + 180 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 482, + 180 + ], + "type": "text", + "content": "70. Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. arXiv preprint arXiv:2304.06718, 2023." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 180, + 481, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 180, + 481, + 201 + ], + "spans": [ + { + "bbox": [ + 132, + 180, + 481, + 201 + ], + "type": "text", + "content": "71. Zhengxia Zou, Zhenwei Shi, Yuhong Guo, and Jieping Ye. Object detection in 20 years: A survey. arXiv preprint arXiv:1905.05055, 2019." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 228, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 228, + 91, + 447, + 102 + ], + "type": "text", + "content": "Segment and Recognize Anything at Any Granularity" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Segment, Lift and Fit_ Automatic 3D Shape Labeling from 2D Prompts/478f2a06-4d04-42d7-815f-2c137ffb864d_content_list.json b/2024/Segment, Lift and Fit_ Automatic 3D Shape Labeling from 2D Prompts/478f2a06-4d04-42d7-815f-2c137ffb864d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..55f248a71c333bd444e41a05ba1bc9327d4a7a5c --- /dev/null +++ b/2024/Segment, Lift and Fit_ Automatic 3D Shape Labeling from 2D Prompts/478f2a06-4d04-42d7-815f-2c137ffb864d_content_list.json @@ -0,0 +1,1775 @@ +[ + { + "type": "text", + "text": "Segment, Lift and Fit: Automatic 3D Shape Labeling from 2D Prompts", + "text_level": 1, + "bbox": [ + 250, + 140, + 754, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jianhao Li $^{1}$ , Tianyu Sun $^{2}$ , Zhongdao Wang $^{3\\dagger}$ , Enze Xie $^{3}$ , Bailan Feng $^{3}$ , Hongbo Zhang $^{3}$ , Ze Yuan $^{1}$ , Ke Xu $^{1}$ , Jiaheng Liu $^{4\\dagger}$ , and Ping Luo $^{5}$", + "bbox": [ + 241, + 210, + 759, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Beihang University, Beijing, China", + "bbox": [ + 374, + 253, + 625, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Tsinghua University, Beijing, China", + "bbox": [ + 374, + 268, + 625, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Noah's Ark Lab, Huawei", + "bbox": [ + 410, + 282, + 589, + 296 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 Nanjing University", + "bbox": [ + 429, + 296, + 571, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "5 The University of Hong Kong", + "bbox": [ + 395, + 310, + 607, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{lijianhao, yuanze1024, kexu}@buaa.edu.cn", + "bbox": [ + 341, + 325, + 660, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{wangzhongdao, Johnny_ez, fengbailan, zhanghongbo888}@huawei.com", + "bbox": [ + 251, + 339, + 750, + 352 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sty21@mails.tsinghua.edu.cn, buaaljiaheng@gmail.com, pluo.lhi@gmail.com", + "bbox": [ + 220, + 353, + 769, + 366 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. This paper proposes an algorithm for automatically labeling 3D objects from 2D point or box prompts, especially focusing on applications in autonomous driving. Unlike previous arts, our auto-labeler predicts 3D shapes instead of bounding boxes and does not require training on a specific dataset. We propose a Segment, Lift, and Fit (SLF) paradigm to achieve this goal. Firstly, we segment high-quality instance masks from the prompts using the Segment Anything Model (SAM) and transform the remaining problem into predicting 3D shapes from given 2D masks. Due to the ill-posed nature of this problem, it presents a significant challenge as multiple 3D shapes can project into an identical mask. To tackle this issue, we then lift 2D masks to 3D forms and employ gradient descent to adjust their poses and shapes until the projections fit the masks and the surfaces conform to surrounding LiDAR points. Notably, since we do not train on a specific dataset, the SLF auto-labeler does not overfit to biased annotation patterns in the training set as other methods do. Thus, the generalization ability across different datasets improves. Experimental results on the KITTI dataset demonstrate that the SLF auto-labeler produces high-quality bounding box annotations, achieving an AP@0.5 IoU of nearly $90\\%$ . Detectors trained with the generated pseudo-labels perform nearly as well as those trained with actual ground-truth annotations. Furthermore, the SLF auto-labeler shows promising results in detailed shape predictions, providing a potential alternative for the occupancy annotation of dynamic objects.", + "bbox": [ + 259, + 392, + 740, + 710 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: 3D Auto-labelers $\\cdot$ Shape Optimization $\\cdot$ Training Free", + "bbox": [ + 261, + 724, + 715, + 738 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 758, + 375, + 776 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In modern robotics and autonomous driving, a significant amount of labeled data is required for supervised learning to understand the 3D scene [8,16,17,19,", + "bbox": [ + 212, + 787, + 787, + 818 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding authors: Zhongdao Wang; Jiaheng Liu.", + "bbox": [ + 230, + 824, + 612, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5e3f82b4b5a52d18b0696738ee5adf02a82ff4230af217d24bf50743cb51ac8d.jpg", + "image_caption": [ + "(a) Human Annotation" + ], + "image_footnote": [], + "bbox": [ + 217, + 146, + 457, + 258 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/30ab53df3ef06079ff2931a42e8d06cb898243baa547769543e4a9482ab4c689.jpg", + "image_caption": [ + "(b) Auto-labeler Annotation using SLF", + "Fig. 1: Comparison between human annotation and auto-labeler annotation using SLF." + ], + "image_footnote": [], + "bbox": [ + 478, + 143, + 784, + 263 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "[20], emphasizing dynamic objects such as vehicles and pedestrians. Therefore, there is a pressing need to improve the degree of automation for 3D labeling. The preferred labeling format is typically the 3D bounding box, which can be represented compactly with as few as 7-9 degrees of freedom. However, manually labeling a large number of 3D boxes remains a laborious and costly task, posing challenges for scaling up 3D object detectors as shown in Fig. 1(a).", + "bbox": [ + 212, + 335, + 782, + 428 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Meanwhile, as 3D perception models evolve, there is an increasing demand for more precise labeling granularity. While bounding boxes offer a compact representation, they are not optimal for 3D objects and are being replaced by finer representations. For instance, voxel occupancy [10,38,41] has emerged as a dominant format for 3D scenes in autonomous driving, as it unifies the representation of static road elements and dynamic objects. However, finer-grained labels further complicate the annotation pipeline, exacerbating the annotation efficiency problem and presenting a significant obstacle to scaling up 3D models.", + "bbox": [ + 212, + 429, + 784, + 550 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a novel approach for 3D auto-labeling. The proposed auto-labeler follows a Segment, Lift, and Fit (SLF) paradigm as shown in Fig. 1(b). Specifically, SLF utilizes the 2D points and bounding boxes as prompts and outputs the estimated 3D shapes and poses of target objects, thus supporting semi- or fully automatic labeling depending on whether the prompts come from humans or machines. To accomplish this, we leverage the Segment Anything Model (SAM) [12], a foundational vision model, to generate high-quality instance masks from the input prompts. Subsequently, we lift 2D instance masks into 3D forms with Signed Distance Functions (SDFs) [26], and then iteratively optimize the shapes and poses until their projections fit the mask and its surface conforms to the surrounding LiDAR points. Notably, the proposed SLF auto-labeler does not require supervised training on a specific dataset, and is therefore not susceptible to overfitting biased annotation patterns in the training dataset.", + "bbox": [ + 212, + 566, + 784, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Experimental results show that SLF can generate high-quality 3D labels. As our method outputs detailed shapes of objects, while existing benchmarks [1,7] only provide bounding box annotations, for a fair comparison with existing auto-labelers, we convert shapes to boxes and assess the label quality regarding box accuracy. When compared among unsupervised auto-labelers [23,39,44] that do", + "bbox": [ + 212, + 763, + 785, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "J. Li et al.", + "bbox": [ + 271, + 114, + 343, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "not train with 3D labels as ours, our method outperforms the best-performing FGR [39] auto-labeler in both settings of direct evaluation on pseudo labels and training object detectors with generated pseudo labels. Moreover, because SLF does not learn biased annotation patterns from the training data, it significantly outperforms state of art supervised model MTrans [15] on \"unseen\" datasets, demonstrating its superior generalization capacity. We also show qualitatively that the predicted shapes align well with the actual geometry of the target objects, underpinning a new possibility for voxel occupancy annotation [38,41]. Our contribution can be summarized as follows:", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose SLF, a 3D auto-labeler that improves existing methods by predicting detailed shapes and poses rather than just bounding boxes. The shape and pose are directly optimized using gradient descent without requiring a neural network trained on a small portion of labels, which enables a \"cold-start\" scenario for auto-labeling.", + "- We demonstrate that SLF generates superior labels to other unsupervised auto-labelers. Additionally, it shows good generalization ability across different datasets, surpassing supervised auto-labelers.", + "- The predicted shapes align well with the actual geometries of the objects, suggesting that SLF can serve as a viable alternative solution for annotating fine-grained occupancy of dynamic objects." + ], + "bbox": [ + 225, + 290, + 787, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 477, + 397, + 493 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Auto-labelers. Compared to supervised 3D object detectors [14,18,42,43], auto-labelers require fewer or no training labels. Supervised auto-labelers learn to annotate from limited samples with ground-truth 3D labels, often in the hundreds. For instance, WS3D [24] generates cylindrical object proposals on bird's eye view (BEV) maps using center clicks in BEV and fine-tunes a 3D detector pre-trained on a few human-annotated data. MTrans [15] employs a multi-task design of segmentation, point generation, and box regression to train a 3D detector that takes ground truth 2D boxes and LiDAR points as input. Unsupervised auto-labelers do not require 3D labels for training. CC [35] uses PointNet [30] to segment and regress 3D boxes from frustum points. SDFLabel [44] trains a DeepSDF [27] network supervised with image patch and CAD models of cars on a synthetic dataset, then uses RANSAC [6] to fit the predicted shape to frustum sub-points. VS3D [31] employs an unsupervised UPM module to generate 3D proposals and filter them using a pre-trained image model. Our SLF, is also an unsupervised approach that does not require training with 3D labels. SLF achieves finer annotations with weaker prompts compared with a closely related unsupervised and non-learning approach, FGR [39]. FGR requires amodal boxes as inputs, in which oracle occlusion information has already been considered. LPCG [28] finds that concerning label accuracy, the 3D location part in the label is preferred compared to other parts of labels in monocular 3D object detection and proposed training-free and high-cost label pipeline, which is similar to FGR and MTrans respectively. In contrast, SLF takes normal 2D", + "bbox": [ + 212, + 507, + 789, + 843 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Segment, Lift and Fit: Automatic 3D Shape Labeling from 2D Prompts", + "bbox": [ + 253, + 114, + 732, + 130 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "modal bounding boxes as input, which an off-the-shelf 2D detector can easily obtain. Furthermore, SLF can produce 3D object shapes rather than just bounding boxes.", + "bbox": [ + 212, + 146, + 782, + 190 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Shape Prior. To reconstruct the 3D geometry for instances from a specific semantic class, e.g., cars or pedestrians, the shape prior of the given class is often utilized to ease the estimation. For instance, McCraith et al. [23] uses a fixed car shape as the most straightforward shape prior. Autoshape [22] develops a shape-labeler for the car class, but it requires ground truth 3D bounding box and amodal instance mask which contains oracle occlusion information as input and it uses part-based deformable representation as shape prior which needs to label the key points and different part of CAD model. Another usage of shape prior is to encode it into a compact low-dimensional space with PCA [4,5,29,37] or neural network [13,44]. SDFLabel [44] pretrains a DeepSDF [27] model to predict the hidden representation of shape conditioned on cropped image patches. 3D-RCNN [13] regresses poses and shape parameters in a single forward pass. However, in 3D-RCNN shape representation is not differentiable, and the finite differences that are used for gradient approximation cause ambiguity during training. Upon the compact PCA representation, Engelmann et al. [4] and DirectShape [37] optimize the shape and pose concurrently with the stereo reconstruction of the object to refine the 3D detection result. SAMP [5] takes sequence of stereo depth maps and 3D trajectory to generate target's shape and refine its trajectory. Prisacariu et al. [29] requires sequence of instance masks and results of 3D tracker to optimize 3D model. Similar to previous PCA based method, our method adopts the PCA shape prior in a differentiable optimization process, however, SLF is training-free and only requires a single raw point cloud and RGB image with 2D prompts, which is more practical.", + "bbox": [ + 212, + 191, + 787, + 540 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Approach", + "text_level": 1, + "bbox": [ + 215, + 563, + 346, + 579 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 2(a), we aim to recover the 3D shape and pose of an interested object by asking a labeler to click one or more points or draw a box on the image as the prompt for the target object. A labeler could be a human in semiautomatic labeling and a machine algorithm such as a 2D object detector in case of full-automatic labeling. Fig. 2(a) shows a brief illustration. We propose a Segment, Lift, and Fit (SLF) paradigm to address this problem. We first segment a 2D mask of the target object using the input prompts, then lift it into a 3D model and iteratively optimize its shape and pose until it fits the 2D mask and surrounding LiDAR points. An overall illustration is shown in Fig. 2(b).", + "bbox": [ + 212, + 594, + 787, + 731 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Segment: From 2D Prompts to Instance Masks", + "text_level": 1, + "bbox": [ + 215, + 753, + 648, + 768 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The problem of auto-labeling 3D shapes from 2D prompts is a significant issue in autonomous driving because it would be highly beneficial to effectively utilize the already established robust 2D instance segmentation models for this purpose. This work employs the strong Segment Anything Model (SAM) [12] to transfer", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "J. Li et al.", + "bbox": [ + 271, + 114, + 343, + 126 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e4e36a39259e7796dca42e05a0704ea8c03dda066acfc7caa8da949a34313bdf.jpg", + "image_caption": [ + "(a) Definition & challenge" + ], + "image_footnote": [], + "bbox": [ + 218, + 145, + 400, + 284 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b4978b9de0169b630f9a0e8efb1bf8581ca0b749110a2477a8b876175954700b.jpg", + "image_caption": [ + "(b) Overall illustration of the proposed method" + ], + "image_footnote": [], + "bbox": [ + 406, + 146, + 566, + 282 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/13046f29a8c3bb633932bc7285a2df9e30b8e2375e6ca34107f801167761e2a0.jpg", + "image_caption": [ + "Fig. 2: (a) We aim to recover the 3D shape and pose of an interested object, given 2D box or point as prompts. This problem is highly ill-posed. (b) In our SLF, we propose to firstly segment the 2D mask of the target object, and then lift the mask to a 3D form and optimize over its shape and pose by gradient descent until the 3D object fits the mask and LiDAR points." + ], + "image_footnote": [], + "bbox": [ + 570, + 147, + 782, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the input point or box prompts into 2D instance masks. Then, our problem is transferred to estimate the 3D shapes and poses from given instance masks. Despite the high quality of SAM masks, the problem is still extremely difficult to solve. This is because multiple 3D shapes can be mapped to the same 2D mask (Fig. 2(a)), making it a highly ill-posed problem and posing a significant challenge for auto-labeler methods.", + "bbox": [ + 212, + 409, + 787, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Optimization Objective", + "text_level": 1, + "bbox": [ + 214, + 520, + 455, + 535 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We employ a pose vector $\\mathbf{p} = (x,y,z,\\theta) \\in \\mathbb{R}^4$ and a latent shape vector $\\mathbf{s} \\in \\mathbb{R}^d$ to represent a 3D object, where $(x,y,z)$ denote the position of the object center in the global coordinate system, $\\theta$ denotes the heading angle around the $z$ axis, and the latent dimension $d$ is empirically set to 5. We assume all instances are on the ground (the $x-y$ plane) and omit rotations around the $x$ and $y$ axes. Our goal is to minimize an energy function that corresponds to the posterior probability of the 3D reconstruction given $\\mathbf{p}$ and $\\mathbf{s}$ as follows:", + "bbox": [ + 212, + 542, + 787, + 648 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {p} ^ {*}, \\mathbf {s} ^ {*} = \\underset {\\mathbf {p}, \\mathbf {s}} {\\arg \\min } E (Y, \\mathcal {X} _ {f}, \\mathbf {p}, \\mathbf {s}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 657, + 785, + 680 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $Y \\in [0,1]^{h \\times w}$ is the 2D instance mask predicted by SAM, and $\\mathcal{X}_f \\in \\mathbb{R}^{N \\times 3}$ is a subset of the whole scene point cloud $\\mathcal{X}$ . $\\mathcal{X}_f$ is comprised of $N$ three-dimensional coordinates of points inside the camera frustum corresponding to the 2D mask. To solve the optimization problem in Eq. (1), we build a differentiable renderer to optimize the energy function $E$ via gradient descent.", + "bbox": [ + 214, + 690, + 787, + 767 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3 Lift: Differentiable Rendering", + "text_level": 1, + "bbox": [ + 214, + 786, + 506, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Representation of 3D objects. We use the Signed Distance Function (SDF) [22, 37] to represent 3D objects. Formally, an SDF is a three-dimensional scalar field", + "bbox": [ + 214, + 809, + 795, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Segment, Lift and Fit: Automatic 3D Shape Labeling from 2D Prompts", + "bbox": [ + 253, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6bd121f5a2fda56b47886bf8930e25e0255479b32030c4ded30316ae91319d6a.jpg", + "image_caption": [ + "Fig. 3: We perform PCA on a diverse collection of 3D shapes to obtain the shape latent code. Left: Example of the basis 3D shapes in the collection. Right: Interpolation along the latent space of the shape code (each row) shows smooth shape variations." + ], + "image_footnote": [], + "bbox": [ + 225, + 146, + 782, + 258 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\phi (\\mathbf{x})$ in which the value of a particular spatial coordinate $\\mathbf{x}$ is given by its distance to the closest object surface, with positive and negative values meaning exterior and interior area. In practice, we use the discrete approximation of continuous SDF. Consider a cubic divided into a 3D mesh grid with the size of $l\\times w\\times h$ . We can represent the inscribed object with the SDF value in each mesh grid, i.e., $\\mathbf{m} = [f(\\mathbf{x}_1),\\dots,f(\\mathbf{x}_i)\\dots,f(\\mathbf{x}_n)]^T$ , where $n = l\\times w\\times h$ .", + "bbox": [ + 212, + 347, + 787, + 439 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Shape prior. We aim to condense the nuanced differences present in class-specific 3D shapes into a low-dimensional latent space, where traversing along the manifold of the latent vector $\\mathbf{s}$ corresponds to the exploration of feasible shapes of the object. We focus on the \"car\" class in this paper, where every instance in the class possesses an overall shape prior and exhibits minor variations in characteristics, such as the shape of the front face or the shape of the cabin.", + "bbox": [ + 212, + 440, + 787, + 531 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To obtain the shape prior, we first assemble a collection of 3D vehicle models from the Apollo-Car3D dataset [33], which consists of 79 models with exhibit significant variations in shapes and sizes (see the left of Fig. 3). However, using the Signed Distance Function (SDF) representation necessitates that objects be water-tight to avoid ambiguity between their interior and exterior. Some models contain intricate internal structures and are not initially water-tight. To obtain the water-tight surface of 3D models, we employ the SoftRas [21] algorithm to remove redundant vertices and faces inside the model. Finally, we convert the original triangular mesh representation of each 3D model into an SDF representation to obtain a set of models $\\mathcal{M} = \\{\\mathbf{m}_i\\}_{i=1}^{79}$ .", + "bbox": [ + 212, + 532, + 787, + 685 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given the collection of representative 3D models, we perform the PCA method on $\\mathcal{M}$ , and keep the top- $d$ ( $d = 5$ ) principle components to obtain a projection matrix $V \\in \\mathbb{R}^{d \\times n}$ . Accordingly, we can use the projection matrix $V$ to render a 3D shape $\\mathbf{m}$ from a given shape vector $\\mathbf{s}$ . Besides, we can encode a 3D model $\\mathbf{m}$ into a shape vector $\\mathbf{s}$ by $\\mathbf{s} = V^{\\top}(\\mathbf{m} - \\bar{\\mathbf{m}})$ and $\\mathbf{m} = V\\mathbf{s} + \\bar{\\mathbf{m}}$ , where $\\bar{\\mathbf{m}}$ indicates the mean vector of the collection $\\mathcal{M}$ . In this work, we select a compact representation $d = 5$ . The right of Fig. 3 illustrates that interpolation along the latent space of $\\mathbf{s}$ leads to smooth variation in shapes.", + "bbox": [ + 212, + 686, + 787, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Differentiable rendering. After recovering the 3D shape in a local coordinate system, we can put the model in the global coordinate system by adding its", + "bbox": [ + 214, + 809, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "J. Li et al.", + "bbox": [ + 271, + 114, + 343, + 126 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1bf734251b80c0387ca1a65e835a9dfe5f06b88d1287ef4481da445f07fbb1a0.jpg", + "image_caption": [ + "(a) RGB image" + ], + "image_footnote": [], + "bbox": [ + 232, + 143, + 364, + 223 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9f55c4409305f9243c5efe18982a98d28fe326148c91b9593c0edba93d3c94eb.jpg", + "image_caption": [ + "(b) SAM mask" + ], + "image_footnote": [], + "bbox": [ + 367, + 143, + 500, + 223 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/312ae667a1f489734491527cf4776831a541092c1cfb7067d5d1306c6e610356.jpg", + "image_caption": [ + "(c) Target mask", + "Fig. 4: Generation of occlusion map $O$ for the mask alignment objective." + ], + "image_footnote": [], + "bbox": [ + 501, + 143, + 635, + 223 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e8306911fa36c3706b1d6ade31fa809ab948e3120d8e2b6caf55b1a4f89b052d.jpg", + "image_caption": [ + "(d) Occ. mask" + ], + "image_footnote": [], + "bbox": [ + 705, + 162, + 771, + 223 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "pose $\\mathbf{p}$ . This gives us an unbounded SDF $\\phi(\\mathbf{x})$ of the 3D model, in which the values outside the pre-defined cubic are filled with negative infinity. Notably, the rendering process of shape is differentiable w.r.t. both $\\mathbf{s}$ and $\\mathbf{p}$ .", + "bbox": [ + 212, + 290, + 784, + 335 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Fit: Optimization over Pose and Shape", + "text_level": 1, + "bbox": [ + 214, + 356, + 583, + 372 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Given the 2D instance mask and the point cloud scene, we randomly initialize pose $\\mathbf{p}$ and shape $\\mathbf{s}$ and optimize them using Eq. (1). Specifically, the optimization objective consists of three terms: the mask alignment term, the point cloud alignment term, and the ground alignment term.", + "bbox": [ + 212, + 378, + 784, + 439 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Mask alignment term. The mask alignment term measures the dissimilarity between the target instance segmentation mask $Y$ and the mask $Y_{proj}$ obtained by projecting the 3D model onto the image plane. To make the projection process differentiable, we follow [22, 37] to define the projection as $\\pi(p_j) = 1 - \\prod_{\\mathbf{x} \\in \\mathcal{R}_{p_j}} \\frac{1}{e^{\\phi(\\mathbf{x})\\zeta + 1}}$ , where $p_j$ is the $j$ -th pixel in the image plane and $\\pi(p_j) \\in (0,1)$ is the mask value. A ray is cast through $p_j$ from the ego camera, and a set of points $\\mathcal{R}p_j$ is collected by sampling along the ray. If the ray does not hit the surface of the 3D model, the SDF values of all sampled points $\\mathbf{x} \\in \\mathcal{R}p_j$ will be negative, resulting in a mask value close to zero. Conversely, if the ray intersects with the 3D model, there must exist sampled points with $\\phi(\\mathbf{x}) > 0$ , leading to a mask value close to 1. A hyperparameter $\\zeta$ controls the \"hardness\" of the mask values, where a larger $\\zeta$ leads to harder values (close to either 0 or 1). Finally, we define the mask alignment term using the dice loss [34] as follows:", + "bbox": [ + 212, + 440, + 785, + 640 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nE _ {\\text {m a s k}} = 1 - \\frac {2 | (Y _ {\\text {p r o j}} * O) \\bigcap Y |}{| Y _ {\\text {p r o j}} * O | + | Y |}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 647, + 784, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $O$ is a binary map indicating if the pixel is occluded by a nearer object. As shown in Fig. 4, to obtain the occlusion map, 3D LiDAR points are associated with each 2D mask based on projections, and the depth of each mask is computed as the median depth value of the associated LiDAR points. Masks are sorted by their depth values and then used to compute the occlusion maps.", + "bbox": [ + 212, + 688, + 784, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Point cloud alignment term. Solely optimizing the mask alignment term leads to inaccurate estimation due to the missing depth information. To avoid undesired local optimum, we introduce an additional point cloud alignment term. This term measures how well the 3D object's surface aligns with the point cloud inside the corresponding frustum.", + "bbox": [ + 212, + 765, + 785, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Segment, Lift and Fit: Automatic 3D Shape Labeling from 2D Prompts", + "bbox": [ + 253, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Intuitively, we can use the $\\ell-1$ norm of the SDF $\\phi(\\mathcal{X}_f)$ to determine how well a 3D model fits a given frustum point cloud, denoted as $\\mathcal{X}_f$ . However, this term alone might lead to sub-optimal solutions, as the points could be associated with the back face of the object. To overcome this issue, a second term is introduced that involves casting a ray for each point $\\mathbf{x} \\in \\mathcal{X}_f$ starting from the ego camera and recording the coordinate $\\mathbf{y}$ when the first time the ray intersects with the 3D model. This term penalizes the predicted 3D model if the distance between $\\mathbf{x}$ and $\\mathbf{y}$ is considerable. The point cloud alignment term can be written as follows:", + "bbox": [ + 212, + 146, + 787, + 268 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nE _ {p c} = \\| \\phi (\\mathcal {X} _ {f}) \\| _ {1} + \\frac {1}{| \\hat {\\mathcal {X}} _ {f} |} \\sum_ {\\mathbf {x} \\in \\hat {\\mathcal {X}} _ {f}} \\| \\mathbf {x} - \\mathbf {y} \\| _ {2} ^ {2}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 364, + 279, + 785, + 319 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$\\hat{\\mathcal{X}}_f$ is a subset of $\\subseteq \\mathcal{X}_f$ and represents points whose corresponding ray intersects with the model surface.", + "bbox": [ + 214, + 327, + 784, + 356 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ground alignment term. Another strong cue that can be leveraged to refine the 3D pose is the height of the ground since we can assume that vehicles run on the road surface. Specifically, we regress a ground plane function $z = g(x,y)$ using RANSAC [6] and then define a ground alignment objective as follows:", + "bbox": [ + 214, + 359, + 785, + 419 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nE _ {\\text {g r o u n d}} = \\| \\hat {z} - \\frac {\\hat {h}}{2} - g (\\hat {x}, \\hat {y}) \\| _ {2} ^ {2}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 431, + 785, + 464 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $(\\hat{x},\\hat{y},\\hat{z})$ is the predicted object center, and $\\hat{h}$ is the height of the object inferred from the predicted 3D shape.", + "bbox": [ + 212, + 476, + 784, + 507 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Optimization. The optimization objective is formulated as a weighted sum of $E_{mask}$ , $E_{pc}$ , and $E_{ground}$ . We initialize pose $\\mathbf{p}$ with the median value of frustum points and shape $\\mathbf{s}$ with mean shape $\\bar{\\mathbf{m}}$ . Based on the Adam optimizer [11] with a fixed learning rate of 0.1 and 150 iterations, a mini-batch comprising all instances (typically $< 100$ ) in a scene is employed for optimization, as opposed to individual objects. This approach yields a $5 \\times$ acceleration with negligible accuracy drop.", + "bbox": [ + 214, + 507, + 787, + 612 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 637, + 375, + 655 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Previous auto-labelers can be divided into supervised ones and unsupervised ones. The proposed SLF is an unsupervised method. Therefore, we primarily compare it with existing unsupervised methods. Nonetheless, we also compare it with supervised methods and demonstrate that SLF exhibits superior transfer ability across datasets. In addition to bounding box accuracy, we also evaluate the quality of the predicted 3D shape.", + "bbox": [ + 212, + 669, + 787, + 760 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Comparison with Unsupervised Auto-labelers", + "text_level": 1, + "bbox": [ + 214, + 782, + 638, + 799 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Dataset and metric. We conduct experiments on the commonly used KITTI [7] dataset. We follow the official training and validation split to get 3712 training", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "J. Li et al.", + "bbox": [ + 271, + 114, + 343, + 127 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a270bfb8047f925141e774688462a67c5a608cb74fb8551a96eaf51d6bfe021f.jpg", + "table_caption": [ + "Table 1: Comparison with unsupervised auto-labelers." + ], + "table_footnote": [], + "table_body": "
MethodAPBEV(IoU=0.5)|R40AP3D(IoU=0.5)|R40
Easy Mod.HardEasy Mod.Hard
VS3D [31]74.566.757.640.337.431.1
SDFLabel [44]77.859.8-62.342.2-
McCraith et al. [23]75.576.668.6---
FGR [39]93.777.870.591.377.870.6
SLFGTmask89.288.987.187.687.883.8
SLFpoint88.588.584.284.884.980.8
SLFbox89.289.384.787.988.283.9
", + "bbox": [ + 289, + 170, + 715, + 325 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "frames and 3768 validation frames. To align with prior works [15,39], the \"Car\" category is the focus of auto-labeling efforts. The Bird's Eye View (BEV) AP and the 3D AP are selected as evaluation metrics for a comprehensive comparison. Results are reported for all three levels of difficulty (Easy, Moderate, and Hard). The commonly used KITTI metrics are evaluated at a strict threshold of 0.7 IoU. As observed in [23,44], the 3D boxes in KITTI have varying spatial padding, making it challenging for human annotators to determine a precise size for a significant number of instances. Therefore, we follow the previous work [23,44] and set the IoU threshold as 0.5 and calculate AP by 40 recall positions.", + "bbox": [ + 212, + 354, + 787, + 491 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Direct evaluation on generated pseudo labels. A straightforward evaluation is to compare the quality of auto-labeled 3D bounding boxes in terms of the canonical metric, Average Precision (AP). However, the AP metric needs a confidence score for each bounding box for thresholding and integration, but FGR [39] and SLF do not output confidence scores. To remedy this issue, we simply define IoU between the predicted mask and the target mask in SLF as a confidence score. As for FGR, we first project a 3D bounding box to image plane to get a 2D box, and the confidence score is defined by IoU between the predicted 2D box and the ground truth 2D box.", + "bbox": [ + 212, + 491, + 787, + 627 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Tab. 1 compares the proposed SLF with existing unsupervised methods on KITTI validation set. $\\mathrm{SLF}_{\\mathrm{point}}$ uses 2D points as input, corresponding to the semi-automatic labeling scenario with humans in the loop. The point prompts are simulated by random sampling 3 points inside the ground truth masks [9]. $\\mathrm{SLF}_{\\mathrm{box}}$ indicates 2D bounding boxes as input, corresponding to the full-automatic labeling scenario in which 2D detectors generate the prompts. For all methods that accept 2D boxes as inputs, we use ground truth boxes for a fair comparison. We also provide a variant of our method in which the SAM masks are replaced with ground truth masks from [9], denoted as $\\mathrm{SLF}_{\\mathrm{GTmask}}$ .", + "bbox": [ + 212, + 628, + 787, + 763 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The last three rows in Tab. 1 show the results of SLF with different 2D prompts. $\\mathrm{SLF}_{\\mathrm{point}}$ and $\\mathrm{SLF}_{\\mathrm{box}}$ perform roughly the same in terms of $\\mathrm{AP}_{BEV}$ , and $\\mathrm{SLF}_{\\mathrm{box}}$ shows superior $\\mathrm{AP}_{3D}$ performance. To our surprise, $\\mathrm{SLF}_{\\mathrm{box}}$ is slightly better than $\\mathrm{SLF}_{\\mathrm{GTmask}}$ in terms of both $\\mathrm{AP}_{3D}$ and $\\mathrm{AP}_{BEV}$ except on hard samples, which means the 2D mask generated by SAM from box prompts are even", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Segment, Lift and Fit: Automatic 3D Shape Labeling from 2D Prompts", + "bbox": [ + 253, + 114, + 732, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/9f4e65eae4c5415ae6867108ae6611b04cc8fb57665011b933509da89f71a3eb.jpg", + "table_caption": [ + "Table 2: Detector performance trained with pseudo labels." + ], + "table_footnote": [], + "table_body": "
DetectorsLabels\\(AP_{BEV}(IoU=0.5)\\) \\(|_{R40}\\)\\(AP_{3D}(IoU=0.5)\\) \\(|_{R40}\\)
EasyMod.HardEasyMod.Hard
PointPillar [14]GT95.794.393.295.793.691.4
FGR96.590.287.194.989.484.8
SLF95.394.191.195.292.188.8
SECOND [42]GT95.994.893.995.894.691.9
FGR95.488.986.295.388.685.7
SLF95.694.291.795.592.489.3
Part-\\(A^2\\) [32]GT96.094.393.996.094.291.9
FGR96.192.187.596.090.085.1
SLF96.094.591.895.993.989.6
Voxel R-CNN [3]GT98.494.994.598.394.894.4
FGR98.790.988.298.690.788.0
SLF96.295.192.496.194.791.9
", + "bbox": [ + 274, + 170, + 728, + 390 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "better than those annotated by humans. This suggests that SLF benefits from more accurate input prompts, and more accurate instance masks.", + "bbox": [ + 212, + 422, + 784, + 454 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We compare $\\mathrm{SLF}_{\\mathrm{box}}$ with other unsupervised methods as they accept the same input prompts. It can be observed that SLF outperforms FGR and SDFLabel by a significant margin on the Moderate and Hard samples. However, SLF performs worse than FGR on the Easy samples, and we assume that easy cases in KITTI have much denser sub-points, which are beneficial to the segmentation and box estimation stage of FGR. Moreover, FGR consumes ground-truth amodal bounding box, while in KITTI these boxes are obtained by projecting 3D ground-truth boxes to 2D images, which may leak the oracle occlusion and size information of partially visible objects. SDFLabel uses a similar 3D shape representation to SLF but relies on a synthetic dataset for training, which may limit its performance on real-world data. In contrast, SLF does not require synthetic training data, and generalizes well to real-world scenarios.", + "bbox": [ + 212, + 455, + 785, + 637 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Detector performance trained with pseudo labels. Following previous work [15,24,39], to further evaluate the quality of generated pseudo labels, we use them as training samples to train off-the-shelf detectors and compare the results of different detectors. We mainly compare with the previous best-performing FGR auto-labeler and train 4 different 3D detectors. Specifically, for all detectors, we use the OpenPCDet [36] with default hyperparameters. In Tab. 2, we compare the results of detectors trained on KITTI training set with ground truth (GT) labels, FGR pseudo labels, and SLF pseudo labels, respectively. If not specified, SLF refers to $\\mathrm{SLF}_{\\mathrm{box}}$ in the remaining sections.", + "bbox": [ + 212, + 640, + 787, + 777 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Tab. 2 shows SLF pseudo labels significantly improve upon FGR pseudo labels on all detectors on Moderate and Hard samples. Regarding Easy samples, SLF pseudo labels are on par with FGR pseudo labels on all detectors except Voxel R-CNN [3]. When comparing specific detectors, the one trained with SLF", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "J. Li et al.", + "bbox": [ + 271, + 114, + 343, + 126 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "labels performs almost as well as the one trained with ground truth labels on both Easy and Moderate samples, and slightly worse on Hard samples, which shows the effectiveness of our method.", + "bbox": [ + 212, + 146, + 787, + 191 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Cross Dataset Generalization", + "text_level": 1, + "bbox": [ + 214, + 217, + 503, + 231 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Apart from KITTI, we conduct experiments on more challenging nuScenes [1]. We transform the validation set of nuScenes to KITTI format, resulting in 6019 frames with corresponding front view images, point cloud and 3D bounding box annotations. On KITTI, we compare the pseudo label with ground truths and report $\\mathrm{AP}_{3D}$ , and on nuScenes, we report the official metrics including mAP, NDS, mATE, mASE and mAOE. We compare with the best-performing unsupervised auto-labeler FGR, and the supervised auto-labeler MTrans [15]. Both FGR and Mtrans require 2D amodal boxes as prompts, so we project 3d ground truth boxes into 2D as their input. Our SLF supports a more practical setting that only requires regular 2D modal boxes as prompts. Thus we can generate them with an off-the-shelf detector HTC [2] pre-trained on nuImages [1].", + "bbox": [ + 212, + 244, + 787, + 412 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/03b9a91528b17c10fe27dcc404f3b77b0055cfaa7b4da56cff4729c92fb405f6.jpg", + "table_caption": [ + "Table 3: Cross dataset evaluation compared with state-of-the-art unsupervised auto-labeler FGR and supervised auto-labeler MTrans. Results in gray are less informative because the NDS and error metrics are computed only on recalled samples, while FGR fails on recalling a majority of input samples (low mAP). $\\mathrm{mAOE}^{\\dagger}$ means to change the periodicity of orientation to $[0,\\pi]$ ." + ], + "table_footnote": [], + "table_body": "
MethodTraining DataAP3D@KITTInuScenes Metrics
EasyMod.HardmAPNDSmATEmASEmAOEmAOE†
FGR [39]-91.377.870.611.640.4.152.188.010.010
MTrans [15]KITTI99.699.697.145.949.4.358.349.702.225
SLF-87.988.283.957.255.2.372.192.836.113
", + "bbox": [ + 220, + 521, + 782, + 614 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In Tab. 3, on KITTI, MTrans shows a nearly perfect performance, this is not surprising since MTrans is a learnable neural network trained on KITTI, it easily learns the specific annotation patterns in training labels. Despite the high accuracy, we argue that learnable auto-labelers like MTrans are prone to overfit the biased pattern in the training data and generalize poorly across datasets. When we apply MTrans to the challenging nuScenes dataset, its performance dramatically decreases. SLF performs better in terms of mAP and NDS. Regarding the detailed error terms, we find SLF outperforms MTrans on size estimation (mASE), and is on par with MTrans on location estimation (mATE). Regarding orientation estimation, SLF underperforms MTrans by 0.134. However, we find a typical orientation error is the model mistakes the front end of the car for the rear end, which is usually caused when the LiDAR points are rather sparse. To explore how good the orientation predictions are when the $\\pi$", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Segment, Lift and Fit: Automatic 3D Shape Labeling from 2D Prompts", + "bbox": [ + 253, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "phase lag is disentangled, we change the periodicity of orientation to $[0,\\pi ]$ and re-measure the orientation error, denoted as $\\mathrm{mAOE}^{\\dagger}$ . SLF outperforms MTrans on $\\mathrm{mAOE}^{\\dagger}$ , suggesting that SLF predicts more accurate orientation, the heading direction aligning better with ground truths. When applying FGR on the nuScences dataset, it fails on a majority of samples, resulting in a low mAP of $11.6\\%$ . Consequently, the NDS score and error terms are less informative since they are computed only on successfully labeled samples, typically easy samples.", + "bbox": [ + 212, + 146, + 787, + 253 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/d19ca68f87c4ad4daa7b11dc6d06115a0a5e400f7ec0cb62ee4e5aee6daf1a74.jpg", + "image_caption": [ + "Fig.5: Qualitative results of estimated shapes. From left to right: projected mask of the initial 3D model; projected mask of the optimized 3D model; mean shape and surrounding point cloud; optimized 3D shape and surrounding point cloud." + ], + "image_footnote": [], + "bbox": [ + 217, + 282, + 354, + 425 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/20522f0553e92b0d5d8ff3489ac7cd44a50b017711bcb56b61c898726f4f7988.jpg", + "image_caption": [ + "Fig.6: Comparison of occupancy annotation. From left to right: reference instance image; occupancy labels from OFN [25]; occupancy labels from OpenOccupancy [38]; occupancy labels generated by our SLF." + ], + "image_footnote": [], + "bbox": [ + 354, + 284, + 490, + 425 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/2cc134221cba5ecd9ac29cb81169f87978f12b8e02370433672c154fabaee569.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 514, + 282, + 784, + 422 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Voxel Occupancy Annotation", + "text_level": 1, + "bbox": [ + 215, + 580, + 503, + 597 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Since benchmarks [1,7] (i.e., KITTI and nuScences) do not support quantitative evaluation on 3D shapes, we show qualitative results in Fig. 5 to better show the effect of SLF on shape labeling. Specifically, we replace the estimated shape with a mean shape and show comparisons. A clear improvement can be observed by comparing the initial shape and the final optimized shape, both in terms of mask alignment and LiDAR point alignment. Besides, we observe that SLF can serve as a good auto-labeler for occupancy annotation of dynamic objects such as vehicles. We compare two versions of occupancy labels for nuScenes, OpenOccupancy [38] and OFN [25]. Considering that existing occupancy labels are still of low quality, quantitative evaluation may be misleading. Therefore we show qualitative visualizations in Fig. 6, and we observe that OpenOccupancy and OFN labels fail to depict the inner occupancy of the cars. This is mainly attributed to that they label the occupancy by voxelizing the LiDAR points, and therefore self-occlusion can not be avoided. In contrast, SLF can predict dense occupancy more accurately, with more detailed shapes.", + "bbox": [ + 212, + 613, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "J. Li et al.", + "bbox": [ + 271, + 114, + 343, + 127 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 146, + 390, + 162 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In this section, we conduct comprehensive ablation studies on the KITTI validation set and report the $\\mathrm{AP}_{3D}(\\mathrm{IoU} = 0.5)|_{R40}$ results. Except in Tab. 6, we input 2D points to SLF, other results are obtained by using 2D bounding boxes as prompt. Effect of energy terms. In Tab. 4, we observe that introducing the mask alignment term significantly improves pseudo-label quality rather than solely applying the point cloud alignment term, and adding the ground alignment term further improves performance. However, as we mentioned above, using mask alignment alone can easily be trapped in sub-optimal solutions, thus it can not obtain valid results (Easy: 6.9, Mod.: 5.9, Hard: 5.6).", + "bbox": [ + 212, + 174, + 787, + 310 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/7bf9328807e4c352ac3a33e7e1817156a69c9949f3c8f7e2fcbd2d653ff4c0b6.jpg", + "table_caption": [ + "Table 4: Ablation on energy terms." + ], + "table_footnote": [], + "table_body": "
ObjectiveEasyMod.Hard
Epc70.564.862.9
Emask+Epc87.388.083.7
Emask+Epc+Eground87.988.283.9
", + "bbox": [ + 225, + 364, + 521, + 438 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/71d5b5540cfd1634fb124f7d721e52f98ba874ff07ad7b12665da3c0bbc98090.jpg", + "table_caption": [ + "Table 5: Ablation on 3D models." + ], + "table_footnote": [], + "table_body": "
# modelsEasyMod.Hard
574.770.164.0
3084.283.677.3
7987.988.283.9
", + "bbox": [ + 527, + 364, + 750, + 436 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effect of the number of CAD models. In Tab. 5, we randomly sample 5 and 30 CAD models out of the total 79 samples to investigate the influence of the number of CAD models used for PCA. From the results, we can conclude that better result comes with more models.", + "bbox": [ + 212, + 465, + 785, + 526 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effect of the PCA dimensions. In Tab. 7, we show how different PCA dimensions affect results. Performance improves when we increase the dimension from 3 to 5. More dimensions like 10 do not further improve. This is possibly attributed to that a higher dimension of shape representation may complicate the optimization process, which means it may need more steps to converge.", + "bbox": [ + 212, + 527, + 787, + 602 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effect of the number of points. In Tab. 6, we analyze the effect of the number of points input to SAM [12]. When we increase the points from 3 to 8, the improvement of pseudo-label quality is minor. However, we observe a significant decrease in the easy cases of KITTI when using one point as the prompt, which is reasonable as easy cases often have large masks and a single point can not provide sufficient context for SAM to generate high-quality results.", + "bbox": [ + 212, + 604, + 787, + 696 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.5 Further Analysis", + "text_level": 1, + "bbox": [ + 215, + 720, + 401, + 734 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Robustness of LiDAR Beam. Point clouds in KITTI [7] are collected by 64 beams LiDAR, we downsample point cloud to 32, 16 and 8 beams following [40]. $\\mathrm{AP}_{3D}$ results on KITTI validation set in Tab. 8 show SLF can still get decent results on 32 and 16 beams LiDAR, and the performances downgrade rapidly on Mod. and Hard cases, since these cases usually far away from ego vehicle which exacerbating the sparsity of point clouds.", + "bbox": [ + 212, + 748, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Segment, Lift and Fit: Automatic 3D Shape Labeling from 2D Prompts", + "bbox": [ + 251, + 113, + 730, + 130 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 127 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/521d705aa31c6f2b46a0bf16ece126c389ee520ccec3130d9b6bf27474faf0ba.jpg", + "table_caption": [ + "Table 6: No. points." + ], + "table_footnote": [], + "table_body": "
# PointsEasyMod.Hard
384.884.980.8
178.181.178.8
584.885.081.0
885.085.280.9
", + "bbox": [ + 215, + 167, + 406, + 263 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/7f63de3023bd2e66e875b77c341cae3e2d9db046bf6ad3ac45e8872d89da1ebb.jpg", + "table_caption": [ + "Table 7: No. dim." + ], + "table_footnote": [], + "table_body": "
Dim.EasyMod.Hard
587.988.283.9
387.187.883.6
886.987.184.1
1084.885.481.1
", + "bbox": [ + 419, + 170, + 583, + 263 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/babc2a323788f96873ef2aa47b376d2a25ebba9db6ad44349e1ac25aa264445d.jpg", + "table_caption": [ + "Table 8: No. beam." + ], + "table_footnote": [], + "table_body": "
BeamEasyMod.Hard
6487.988.283.9
3281.377.577.3
1680.776.370.0
875.160.056.7
", + "bbox": [ + 612, + 170, + 781, + 263 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/e9e3a0a35a8b5949b4666db5fb71b900e3867a261d207f7c82902028055f4377.jpg", + "table_caption": [ + "Table 9: Dilation on instance mask." + ], + "table_footnote": [], + "table_body": "
KernelEasyMod.Hard
FGR(no dil.)91.377.870.6
SLF(no dil.)87.988.283.9
5x586.385.681.2
9x983.980.576.6
", + "bbox": [ + 241, + 304, + 486, + 398 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/5b5f281986e619c47caf3245321b722464ed195d54203c41fbee0df1402400ff.jpg", + "table_caption": [ + "Table 10: Erosion on instance mask." + ], + "table_footnote": [], + "table_body": "
KernelEasyMod.Hard
FGR(noERO.)91.377.870.6
SLF(noERO.)87.988.283.9
5x585.283.979.5
9x982.178.272.1
", + "bbox": [ + 516, + 304, + 763, + 398 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Robustness of Mask Quality. We simulate low-quality masks by applying erosion and dilation operations on SAM [12] masks and report $\\mathrm{AP}_{3D}$ results of KITTI validation set in Tab. 9 and Tab. 10. We observe even with severely corrupted masks, SLF still outperforms the existing unsupervised auto-annotation methods in Mod. and Hard cases, which show the robustness of SLF.", + "bbox": [ + 212, + 428, + 787, + 506 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Limitation", + "text_level": 1, + "bbox": [ + 215, + 534, + 354, + 550 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The proposed SLF uses a strong shape prior learned by PCA from given 3D CAD models. The shape prior is category-specific, which means we need to prepare a collection of 3D models when labeling a specific semantic class.", + "bbox": [ + 212, + 573, + 785, + 619 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 650, + 359, + 666 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This paper proposes a Segment, Lift, and Fit paradigm for automatic 3D shape annotation of objects from the 2D point or box prompts. We first segment high-quality instance masks from the prompts using the SAM, and then lift 2D masks to 3D forms, where gradient descent is used to adjust their poses and shapes until the projections fits the masks and the surfaces conform to surrounding LiDAR points. Extensive experimental results on multiple datasets for 3D detection demonstrate the effectiveness of our SLF method. Moreover, the SLF auto-labeler achieves promising results in detailed shape predictions, which can be used for occupancy annotation. In the future, we will investigate the generalization abilities by applying our SLF to more categories.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "J. Li et al.", + "bbox": [ + 271, + 114, + 343, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. arXiv preprint arXiv:1903.11027 (2019)", + "2. Chen, K., Pang, J., Wang, J., Xiong, Y., Li, X., Sun, S., Feng, W., Liu, Z., Shi, J., Ouyang, W., et al.: Hybrid task cascade for instance segmentation. In: CVPR. pp. 4974-4983 (2019)", + "3. Deng, J., Shi, S., Li, P., Zhou, W., Zhang, Y., Li, H.: Voxel r-cnn: Towards high performance voxel-based 3d object detection. arXiv:2012.15712 (2020)", + "4. Engelmann, F., Stickler, J., Leibe, B.: Joint object pose estimation and shape reconstruction in urban street scenes using 3d shape priors. In: Pattern Recognition: 38th German Conference, GCPR 2016, Hannover, Germany, September 12-15, 2016, Proceedings 38. pp. 219-230. Springer (2016)", + "5. Engelmann, F., Stuckler, J., Leibe, B.: Samp: shape and motion priors for 4d vehicle reconstruction. In: 2017 IEEE Winter Conference on Applications of Computer Vision (WACV). pp. 400-408. IEEE (2017)", + "6. Fischler, M.A., Bolles, R.C.: Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM 24(6), 381-395 (1981)", + "7. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: CVPR (2012)", + "8. Guo, J., Liu, J., Xu, D.: 3d-pruning: A model compression framework for efficient 3d action recognition. IEEE Transactions on Circuits and Systems for Video Technology 32(12), 8717-8729 (2022)", + "9. Heylen, J., De Wolf, M., Dawagne, B., Proesmans, M., Van Gool, L., Abbeloos, W., Abdelkawy, H., Reino, D.O.: Monocinis: Camera independent monocular 3d object detection using instance segmentation. In: ICCV. pp. 923-934 (2021)", + "0. Huang, Y., Zheng, W., Zhang, Y., Zhou, J., Lu, J.: Tri-perspective view for vision-based 3d semantic occupancy prediction. arXiv preprint arXiv:2302.07817 (2023)", + "1. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. In: ICLR (2015)", + "2. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. arXiv:2304.02643 (2023)", + "3. Kundu, A., Li, Y., Rehg, J.M.: 3d-rcnn: Instance-level 3d object reconstruction via render-and-compare. In: CVPR (2018). https://doi.org/10.1109/CVPR.2018.00375, https://doi.org/10.1109/CVPR.2018.00375", + "4. Lang, A.H., Vora, S., Caesar, H., Zhou, L., Yang, J., Beijbom, O.: Pointpillars: Fast encoders for object detection from point clouds. In: CVPR. pp. 12697-12705 (2019)", + "5. Liu, C., Qian, X., Huang, B., Qi, X., Lam, E., Tan, S.C., Wong, N.: Multimodal transformer for automatic 3d annotation and object detection. In: ECCV (2022)", + "6. Liu, J., Guo, J., Xu, D.: Aptsnet: Toward adaptive point sampling for efficient 3d action recognition. IEEE Transactions on Image Processing 31, 5287-5302 (2022)", + "7. Liu, J., Guo, J., Xu, D.: Geometrhythmotion-transformer: An end-to-end framework for 3d action recognition. IEEE Transactions on Multimedia (2022)", + "8. Liu, J., He, T., Yang, H., Su, R., Tian, J., Wu, J., Guo, H., Xu, K., Ouyang, W.: 3d-queryis: A query-based framework for 3d instance segmentation. arXiv preprint arXiv:2211.09375 (2022)" + ], + "bbox": [ + 225, + 176, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Segment, Lift and Fit: Automatic 3D Shape Labeling from 2D Prompts", + "bbox": [ + 253, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "19. Liu, J., Li, J., Wang, K., Guo, H., Yang, J., Peng, J., Xu, K., Liu, X., Guo, J.: Lta-pcs: Learnable task-agnostic point cloud sampling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 28035-28045 (June 2024)", + "20. Liu, J., Xu, D.: Geometrhythmotion-net: A strong two-stream baseline for 3d action recognition. IEEE Transactions on Circuits and Systems for Video Technology 31(12), 4711-4721 (2021)", + "21. Liu, S., Li, T., Chen, W., Li, H.: Soft rasterizer: A differentiable renderer for image-based 3d reasoning. ICCV (Oct 2019)", + "22. Liu, Z., Zhou, D., Lu, F., Fang, J., Zhang, L.: Autoshape: Real-time shape-aware monocular 3d object detection. In: ICCV. pp. 15641-15650 (2021)", + "23. McCraith, R., Insafutdinov, E., Neumann, L., Vedaldi, A.: Lifting 2d object locations to 3d by discounting lidar outliers across objects and views (2022)", + "24. Meng, Q., Wang, W., Zhou, T., Shen, J., Van Gool, L., Dai, D.: Weakly supervised 3d object detection from lidar point cloud. In: ECCV (2020)", + "25. Ming Fang, Z.L.: Occupancy dataset for nuscenes. https://github.com/FANG-MING/occupancy-for-nuscenes (2023)", + "26. Osher, S., Fedkiw, R.: Signed Distance Functions. Springer New York (2003). https://doi.org/10.1007/0-387-22746-6_2, https://doi.org/10.1007/0-387-22746-6_2", + "27. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: CVPR (June 2019)", + "28. Peng, L., Liu, F., Yu, Z., Yan, S., Deng, D., Yang, Z., Liu, H., Cai, D.: Lidar point cloud guided monocular 3d object detection. In: European Conference on Computer Vision. pp. 123-139. Springer (2022)", + "29. Prisacariu, V.A., Segal, A.V., Reid, I.: Simultaneous monocular 2d segmentation, 3d pose recovery and 3d reconstruction. In: Lee, K.M., Matsushita, Y., Rehg, J.M., Hu, Z. (eds.) ACCV. pp. 593-606. Springer Berlin Heidelberg, Berlin, Heidelberg (2013)", + "30. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. arXiv preprint arXiv:1612.00593 (2016)", + "31. Qin, Z., Wang, J., Lu, Y.: Weakly supervised 3d object detection from point clouds. ACMMM (2020)", + "32. Shi, S., Wang, Z., Shi, J., Wang, X., Li, H.: From points to parts: 3d object detection from point cloud with part-aware and part-aggregation network. IEEE transactions on pattern analysis and machine intelligence 43(8), 2647-2664 (2020)", + "33. Song, X., Wang, P., Zhou, D., Zhu, R., Guan, C., Dai, Y., Su, H., Li, H., Yang, R.: Apollocar3d: A large 3d car instance understanding benchmark for autonomous driving. In: CVPR. pp. 5452-5462 (2019)", + "34. Sudre, C.H., Li, W., Vercauteren, T., Ourselin, S., Jorge Cardoso, M.: Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations. In: Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: Third International Workshop, DLMIA 2017, and 7th International Workshop, ML-CDS 2017, Held in Conjunction with MICCAI 2017, Quebec City, QC, Canada, September 14, Proceedings 3. pp. 240-248. Springer (2017)", + "35. Tang, Y.S., Lee, G.H.: Transferable semi-supervised 3d object detection from rgb-d data. In: ICCV. pp. 1931-1940 (2019)", + "36. Team, O.D.: Openpcdet: An open-source toolbox for 3d object detection from point clouds. https://github.com/open-mmlab/OpenPCDet (2020)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "J. Li et al.", + "bbox": [ + 271, + 114, + 341, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "37. Wang, R., Yang, N., Stueckler, J., Cremers, D.: Directshape: Photometric alignment of shape priors for visual vehicle pose and shape estimation (2020)", + "38. Wang, X., Zhu, Z., Xu, W., Zhang, Y., Wei, Y., Chi, X., Ye, Y., Du, D., Lu, J., Wang, X.: Openoccupancy: A large scale benchmark for surrounding semantic occupancy perception. arXiv preprint arXiv:2303.03991 (2023)", + "39. Wei, Y., Su, S., Lu, J., Zhou, J.: Fgr: Frustum-aware geometric reasoning for weakly supervised 3d vehicle detection (2021)", + "40. Wei, Y., Wei, Z., Rao, Y., Li, J., Zhou, J., Lu, J.: Lidar distillation: Bridging the beam-induced domain gap for 3d object detection. In: ECCV. pp. 179-195. Springer (2022)", + "41. Wei, Y., Zhao, L., Zheng, W., Zhu, Z., Zhou, J., Lu, J.: Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving. arXiv preprint arXiv:2303.09551 (2023)", + "42. Yan, Y., Mao, Y., Li, B.: Second: Sparsely embedded convolutional detection. Sensors 18(10), 3337 (2018)", + "43. Yang, H., He, T., Liu, J., Chen, H., Wu, B., Lin, B., He, X., Ouyang, W.: Gdmae: generative decoder for mae pre-training on lidar point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9403-9414 (2023)", + "44. Zakharov, S., Kehl, W., Bhargava, A., Gaidon, A.: Autolabeling 3d objects with differentiable rendering of sdf shape priors. In: CVPR (2020)" + ], + "bbox": [ + 215, + 146, + 784, + 438 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Segment, Lift and Fit: Automatic 3D Shape Labeling from 2D Prompts", + "bbox": [ + 253, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Segment, Lift and Fit_ Automatic 3D Shape Labeling from 2D Prompts/478f2a06-4d04-42d7-815f-2c137ffb864d_model.json b/2024/Segment, Lift and Fit_ Automatic 3D Shape Labeling from 2D Prompts/478f2a06-4d04-42d7-815f-2c137ffb864d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..de0f885e537b4cfd6896f185da3e5d211267cbf1 --- /dev/null +++ b/2024/Segment, Lift and Fit_ Automatic 3D Shape Labeling from 2D Prompts/478f2a06-4d04-42d7-815f-2c137ffb864d_model.json @@ -0,0 +1,2412 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.25, + 0.141, + 0.756, + 0.187 + ], + "angle": 0, + "content": "Segment, Lift and Fit: Automatic 3D Shape Labeling from 2D Prompts" + }, + { + "type": "text", + "bbox": [ + 0.243, + 0.212, + 0.761, + 0.245 + ], + "angle": 0, + "content": "Jianhao Li\\(^{1}\\), Tianyu Sun\\(^{2}\\), Zhongdao Wang\\(^{3\\dagger}\\), Enze Xie\\(^{3}\\), Bailan Feng\\(^{3}\\), Hongbo Zhang\\(^{3}\\), Ze Yuan\\(^{1}\\), Ke Xu\\(^{1}\\), Jiaheng Liu\\(^{4\\dagger}\\), and Ping Luo\\(^{5}\\)" + }, + { + "type": "text", + "bbox": [ + 0.375, + 0.255, + 0.626, + 0.269 + ], + "angle": 0, + "content": "1 Beihang University, Beijing, China" + }, + { + "type": "text", + "bbox": [ + 0.375, + 0.27, + 0.627, + 0.283 + ], + "angle": 0, + "content": "2 Tsinghua University, Beijing, China" + }, + { + "type": "text", + "bbox": [ + 0.411, + 0.283, + 0.591, + 0.297 + ], + "angle": 0, + "content": "3 Noah's Ark Lab, Huawei" + }, + { + "type": "text", + "bbox": [ + 0.431, + 0.297, + 0.573, + 0.311 + ], + "angle": 0, + "content": "4 Nanjing University" + }, + { + "type": "text", + "bbox": [ + 0.396, + 0.311, + 0.609, + 0.325 + ], + "angle": 0, + "content": "5 The University of Hong Kong" + }, + { + "type": "text", + "bbox": [ + 0.343, + 0.326, + 0.661, + 0.339 + ], + "angle": 0, + "content": "{lijianhao, yuanze1024, kexu}@buaa.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.253, + 0.34, + 0.75, + 0.353 + ], + "angle": 0, + "content": "{wangzhongdao, Johnny_ez, fengbailan, zhanghongbo888}@huawei.com" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.354, + 0.77, + 0.367 + ], + "angle": 0, + "content": "sty21@mails.tsinghua.edu.cn, buaaljiaheng@gmail.com, pluo.lhi@gmail.com" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.393, + 0.741, + 0.712 + ], + "angle": 0, + "content": "Abstract. This paper proposes an algorithm for automatically labeling 3D objects from 2D point or box prompts, especially focusing on applications in autonomous driving. Unlike previous arts, our auto-labeler predicts 3D shapes instead of bounding boxes and does not require training on a specific dataset. We propose a Segment, Lift, and Fit (SLF) paradigm to achieve this goal. Firstly, we segment high-quality instance masks from the prompts using the Segment Anything Model (SAM) and transform the remaining problem into predicting 3D shapes from given 2D masks. Due to the ill-posed nature of this problem, it presents a significant challenge as multiple 3D shapes can project into an identical mask. To tackle this issue, we then lift 2D masks to 3D forms and employ gradient descent to adjust their poses and shapes until the projections fit the masks and the surfaces conform to surrounding LiDAR points. Notably, since we do not train on a specific dataset, the SLF auto-labeler does not overfit to biased annotation patterns in the training set as other methods do. Thus, the generalization ability across different datasets improves. Experimental results on the KITTI dataset demonstrate that the SLF auto-labeler produces high-quality bounding box annotations, achieving an AP@0.5 IoU of nearly \\(90\\%\\). Detectors trained with the generated pseudo-labels perform nearly as well as those trained with actual ground-truth annotations. Furthermore, the SLF auto-labeler shows promising results in detailed shape predictions, providing a potential alternative for the occupancy annotation of dynamic objects." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.725, + 0.717, + 0.739 + ], + "angle": 0, + "content": "Keywords: 3D Auto-labelers \\(\\cdot\\) Shape Optimization \\(\\cdot\\) Training Free" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.76, + 0.377, + 0.777 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.788, + 0.788, + 0.819 + ], + "angle": 0, + "content": "In modern robotics and autonomous driving, a significant amount of labeled data is required for supervised learning to understand the 3D scene [8,16,17,19," + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.825, + 0.614, + 0.841 + ], + "angle": 0, + "content": "† Corresponding authors: Zhongdao Wang; Jiaheng Liu." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.344, + 0.127 + ], + "angle": 0, + "content": "J. Li et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.147, + 0.458, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.268, + 0.264, + 0.424, + 0.278 + ], + "angle": 0, + "content": "(a) Human Annotation" + }, + { + "type": "image", + "bbox": [ + 0.479, + 0.145, + 0.785, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.265, + 0.78, + 0.279 + ], + "angle": 0, + "content": "(b) Auto-labeler Annotation using SLF" + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.29, + 0.785, + 0.305 + ], + "angle": 0, + "content": "Fig. 1: Comparison between human annotation and auto-labeler annotation using SLF." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.337, + 0.784, + 0.429 + ], + "angle": 0, + "content": "[20], emphasizing dynamic objects such as vehicles and pedestrians. Therefore, there is a pressing need to improve the degree of automation for 3D labeling. The preferred labeling format is typically the 3D bounding box, which can be represented compactly with as few as 7-9 degrees of freedom. However, manually labeling a large number of 3D boxes remains a laborious and costly task, posing challenges for scaling up 3D object detectors as shown in Fig. 1(a)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.785, + 0.551 + ], + "angle": 0, + "content": "Meanwhile, as 3D perception models evolve, there is an increasing demand for more precise labeling granularity. While bounding boxes offer a compact representation, they are not optimal for 3D objects and are being replaced by finer representations. For instance, voxel occupancy [10,38,41] has emerged as a dominant format for 3D scenes in autonomous driving, as it unifies the representation of static road elements and dynamic objects. However, finer-grained labels further complicate the annotation pipeline, exacerbating the annotation efficiency problem and presenting a significant obstacle to scaling up 3D models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.567, + 0.785, + 0.763 + ], + "angle": 0, + "content": "In this paper, we propose a novel approach for 3D auto-labeling. The proposed auto-labeler follows a Segment, Lift, and Fit (SLF) paradigm as shown in Fig. 1(b). Specifically, SLF utilizes the 2D points and bounding boxes as prompts and outputs the estimated 3D shapes and poses of target objects, thus supporting semi- or fully automatic labeling depending on whether the prompts come from humans or machines. To accomplish this, we leverage the Segment Anything Model (SAM) [12], a foundational vision model, to generate high-quality instance masks from the input prompts. Subsequently, we lift 2D instance masks into 3D forms with Signed Distance Functions (SDFs) [26], and then iteratively optimize the shapes and poses until their projections fit the mask and its surface conforms to the surrounding LiDAR points. Notably, the proposed SLF auto-labeler does not require supervised training on a specific dataset, and is therefore not susceptible to overfitting biased annotation patterns in the training dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Experimental results show that SLF can generate high-quality 3D labels. As our method outputs detailed shapes of objects, while existing benchmarks [1,7] only provide bounding box annotations, for a fair comparison with existing auto-labelers, we convert shapes to boxes and assess the label quality regarding box accuracy. When compared among unsupervised auto-labelers [23,39,44] that do" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.254, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Segment, Lift and Fit: Automatic 3D Shape Labeling from 2D Prompts" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.283 + ], + "angle": 0, + "content": "not train with 3D labels as ours, our method outperforms the best-performing FGR [39] auto-labeler in both settings of direct evaluation on pseudo labels and training object detectors with generated pseudo labels. Moreover, because SLF does not learn biased annotation patterns from the training data, it significantly outperforms state of art supervised model MTrans [15] on \"unseen\" datasets, demonstrating its superior generalization capacity. We also show qualitatively that the predicted shapes align well with the actual geometry of the target objects, underpinning a new possibility for voxel occupancy annotation [38,41]. Our contribution can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.291, + 0.788, + 0.367 + ], + "angle": 0, + "content": "- We propose SLF, a 3D auto-labeler that improves existing methods by predicting detailed shapes and poses rather than just bounding boxes. The shape and pose are directly optimized using gradient descent without requiring a neural network trained on a small portion of labels, which enables a \"cold-start\" scenario for auto-labeling." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.366, + 0.788, + 0.412 + ], + "angle": 0, + "content": "- We demonstrate that SLF generates superior labels to other unsupervised auto-labelers. Additionally, it shows good generalization ability across different datasets, surpassing supervised auto-labelers." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.411, + 0.788, + 0.457 + ], + "angle": 0, + "content": "- The predicted shapes align well with the actual geometries of the objects, suggesting that SLF can serve as a viable alternative solution for annotating fine-grained occupancy of dynamic objects." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.291, + 0.788, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.478, + 0.398, + 0.494 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.79, + 0.844 + ], + "angle": 0, + "content": "3D Auto-labelers. Compared to supervised 3D object detectors [14,18,42,43], auto-labelers require fewer or no training labels. Supervised auto-labelers learn to annotate from limited samples with ground-truth 3D labels, often in the hundreds. For instance, WS3D [24] generates cylindrical object proposals on bird's eye view (BEV) maps using center clicks in BEV and fine-tunes a 3D detector pre-trained on a few human-annotated data. MTrans [15] employs a multi-task design of segmentation, point generation, and box regression to train a 3D detector that takes ground truth 2D boxes and LiDAR points as input. Unsupervised auto-labelers do not require 3D labels for training. CC [35] uses PointNet [30] to segment and regress 3D boxes from frustum points. SDFLabel [44] trains a DeepSDF [27] network supervised with image patch and CAD models of cars on a synthetic dataset, then uses RANSAC [6] to fit the predicted shape to frustum sub-points. VS3D [31] employs an unsupervised UPM module to generate 3D proposals and filter them using a pre-trained image model. Our SLF, is also an unsupervised approach that does not require training with 3D labels. SLF achieves finer annotations with weaker prompts compared with a closely related unsupervised and non-learning approach, FGR [39]. FGR requires amodal boxes as inputs, in which oracle occlusion information has already been considered. LPCG [28] finds that concerning label accuracy, the 3D location part in the label is preferred compared to other parts of labels in monocular 3D object detection and proposed training-free and high-cost label pipeline, which is similar to FGR and MTrans respectively. In contrast, SLF takes normal 2D" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.344, + 0.127 + ], + "angle": 0, + "content": "J. Li et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.191 + ], + "angle": 0, + "content": "modal bounding boxes as input, which an off-the-shelf 2D detector can easily obtain. Furthermore, SLF can produce 3D object shapes rather than just bounding boxes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.192, + 0.788, + 0.541 + ], + "angle": 0, + "content": "Shape Prior. To reconstruct the 3D geometry for instances from a specific semantic class, e.g., cars or pedestrians, the shape prior of the given class is often utilized to ease the estimation. For instance, McCraith et al. [23] uses a fixed car shape as the most straightforward shape prior. Autoshape [22] develops a shape-labeler for the car class, but it requires ground truth 3D bounding box and amodal instance mask which contains oracle occlusion information as input and it uses part-based deformable representation as shape prior which needs to label the key points and different part of CAD model. Another usage of shape prior is to encode it into a compact low-dimensional space with PCA [4,5,29,37] or neural network [13,44]. SDFLabel [44] pretrains a DeepSDF [27] model to predict the hidden representation of shape conditioned on cropped image patches. 3D-RCNN [13] regresses poses and shape parameters in a single forward pass. However, in 3D-RCNN shape representation is not differentiable, and the finite differences that are used for gradient approximation cause ambiguity during training. Upon the compact PCA representation, Engelmann et al. [4] and DirectShape [37] optimize the shape and pose concurrently with the stereo reconstruction of the object to refine the 3D detection result. SAMP [5] takes sequence of stereo depth maps and 3D trajectory to generate target's shape and refine its trajectory. Prisacariu et al. [29] requires sequence of instance masks and results of 3D tracker to optimize 3D model. Similar to previous PCA based method, our method adopts the PCA shape prior in a differentiable optimization process, however, SLF is training-free and only requires a single raw point cloud and RGB image with 2D prompts, which is more practical." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.564, + 0.347, + 0.58 + ], + "angle": 0, + "content": "3 Approach" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.595, + 0.788, + 0.732 + ], + "angle": 0, + "content": "As shown in Fig. 2(a), we aim to recover the 3D shape and pose of an interested object by asking a labeler to click one or more points or draw a box on the image as the prompt for the target object. A labeler could be a human in semiautomatic labeling and a machine algorithm such as a 2D object detector in case of full-automatic labeling. Fig. 2(a) shows a brief illustration. We propose a Segment, Lift, and Fit (SLF) paradigm to address this problem. We first segment a 2D mask of the target object using the input prompts, then lift it into a 3D model and iteratively optimize its shape and pose until it fits the 2D mask and surrounding LiDAR points. An overall illustration is shown in Fig. 2(b)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.755, + 0.65, + 0.77 + ], + "angle": 0, + "content": "3.1 Segment: From 2D Prompts to Instance Masks" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The problem of auto-labeling 3D shapes from 2D prompts is a significant issue in autonomous driving because it would be highly beneficial to effectively utilize the already established robust 2D instance segmentation models for this purpose. This work employs the strong Segment Anything Model (SAM) [12] to transfer" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.254, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Segment, Lift and Fit: Automatic 3D Shape Labeling from 2D Prompts" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.146, + 0.401, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.234, + 0.288, + 0.386, + 0.301 + ], + "angle": 0, + "content": "(a) Definition & challenge" + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.147, + 0.568, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.458, + 0.288, + 0.733, + 0.301 + ], + "angle": 0, + "content": "(b) Overall illustration of the proposed method" + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.148, + 0.784, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.312, + 0.789, + 0.382 + ], + "angle": 0, + "content": "Fig. 2: (a) We aim to recover the 3D shape and pose of an interested object, given 2D box or point as prompts. This problem is highly ill-posed. (b) In our SLF, we propose to firstly segment the 2D mask of the target object, and then lift the mask to a 3D form and optimize over its shape and pose by gradient descent until the 3D object fits the mask and LiDAR points." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.41, + 0.788, + 0.501 + ], + "angle": 0, + "content": "the input point or box prompts into 2D instance masks. Then, our problem is transferred to estimate the 3D shapes and poses from given instance masks. Despite the high quality of SAM masks, the problem is still extremely difficult to solve. This is because multiple 3D shapes can be mapped to the same 2D mask (Fig. 2(a)), making it a highly ill-posed problem and posing a significant challenge for auto-labeler methods." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.521, + 0.456, + 0.536 + ], + "angle": 0, + "content": "3.2 Optimization Objective" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.543, + 0.788, + 0.65 + ], + "angle": 0, + "content": "We employ a pose vector \\(\\mathbf{p} = (x,y,z,\\theta) \\in \\mathbb{R}^4\\) and a latent shape vector \\(\\mathbf{s} \\in \\mathbb{R}^d\\) to represent a 3D object, where \\((x,y,z)\\) denote the position of the object center in the global coordinate system, \\(\\theta\\) denotes the heading angle around the \\(z\\) axis, and the latent dimension \\(d\\) is empirically set to 5. We assume all instances are on the ground (the \\(x-y\\) plane) and omit rotations around the \\(x\\) and \\(y\\) axes. Our goal is to minimize an energy function that corresponds to the posterior probability of the 3D reconstruction given \\(\\mathbf{p}\\) and \\(\\mathbf{s}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.389, + 0.659, + 0.786, + 0.681 + ], + "angle": 0, + "content": "\\[\n\\mathbf {p} ^ {*}, \\mathbf {s} ^ {*} = \\underset {\\mathbf {p}, \\mathbf {s}} {\\arg \\min } E (Y, \\mathcal {X} _ {f}, \\mathbf {p}, \\mathbf {s}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.691, + 0.788, + 0.768 + ], + "angle": 0, + "content": "where \\( Y \\in [0,1]^{h \\times w} \\) is the 2D instance mask predicted by SAM, and \\( \\mathcal{X}_f \\in \\mathbb{R}^{N \\times 3} \\) is a subset of the whole scene point cloud \\( \\mathcal{X} \\). \\( \\mathcal{X}_f \\) is comprised of \\( N \\) three-dimensional coordinates of points inside the camera frustum corresponding to the 2D mask. To solve the optimization problem in Eq. (1), we build a differentiable renderer to optimize the energy function \\( E \\) via gradient descent." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.787, + 0.507, + 0.803 + ], + "angle": 0, + "content": "3.3 Lift: Differentiable Rendering" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.796, + 0.842 + ], + "angle": 0, + "content": "Representation of 3D objects. We use the Signed Distance Function (SDF) [22, 37] to represent 3D objects. Formally, an SDF is a three-dimensional scalar field" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.344, + 0.127 + ], + "angle": 0, + "content": "J. Li et al." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.147, + 0.784, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.27, + 0.788, + 0.314 + ], + "angle": 0, + "content": "Fig. 3: We perform PCA on a diverse collection of 3D shapes to obtain the shape latent code. Left: Example of the basis 3D shapes in the collection. Right: Interpolation along the latent space of the shape code (each row) shows smooth shape variations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.348, + 0.788, + 0.44 + ], + "angle": 0, + "content": "\\(\\phi (\\mathbf{x})\\) in which the value of a particular spatial coordinate \\(\\mathbf{x}\\) is given by its distance to the closest object surface, with positive and negative values meaning exterior and interior area. In practice, we use the discrete approximation of continuous SDF. Consider a cubic divided into a 3D mesh grid with the size of \\(l\\times w\\times h\\). We can represent the inscribed object with the SDF value in each mesh grid, i.e., \\(\\mathbf{m} = [f(\\mathbf{x}_1),\\dots,f(\\mathbf{x}_i)\\dots,f(\\mathbf{x}_n)]^T\\), where \\(n = l\\times w\\times h\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.441, + 0.788, + 0.532 + ], + "angle": 0, + "content": "Shape prior. We aim to condense the nuanced differences present in class-specific 3D shapes into a low-dimensional latent space, where traversing along the manifold of the latent vector \\(\\mathbf{s}\\) corresponds to the exploration of feasible shapes of the object. We focus on the \"car\" class in this paper, where every instance in the class possesses an overall shape prior and exhibits minor variations in characteristics, such as the shape of the front face or the shape of the cabin." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.534, + 0.788, + 0.686 + ], + "angle": 0, + "content": "To obtain the shape prior, we first assemble a collection of 3D vehicle models from the Apollo-Car3D dataset [33], which consists of 79 models with exhibit significant variations in shapes and sizes (see the left of Fig. 3). However, using the Signed Distance Function (SDF) representation necessitates that objects be water-tight to avoid ambiguity between their interior and exterior. Some models contain intricate internal structures and are not initially water-tight. To obtain the water-tight surface of 3D models, we employ the SoftRas [21] algorithm to remove redundant vertices and faces inside the model. Finally, we convert the original triangular mesh representation of each 3D model into an SDF representation to obtain a set of models \\(\\mathcal{M} = \\{\\mathbf{m}_i\\}_{i=1}^{79}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.687, + 0.788, + 0.809 + ], + "angle": 0, + "content": "Given the collection of representative 3D models, we perform the PCA method on \\(\\mathcal{M}\\), and keep the top-\\(d\\) (\\(d = 5\\)) principle components to obtain a projection matrix \\(V \\in \\mathbb{R}^{d \\times n}\\). Accordingly, we can use the projection matrix \\(V\\) to render a 3D shape \\(\\mathbf{m}\\) from a given shape vector \\(\\mathbf{s}\\). Besides, we can encode a 3D model \\(\\mathbf{m}\\) into a shape vector \\(\\mathbf{s}\\) by \\(\\mathbf{s} = V^{\\top}(\\mathbf{m} - \\bar{\\mathbf{m}})\\) and \\(\\mathbf{m} = V\\mathbf{s} + \\bar{\\mathbf{m}}\\), where \\(\\bar{\\mathbf{m}}\\) indicates the mean vector of the collection \\(\\mathcal{M}\\). In this work, we select a compact representation \\(d = 5\\). The right of Fig. 3 illustrates that interpolation along the latent space of \\(\\mathbf{s}\\) leads to smooth variation in shapes." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Differentiable rendering. After recovering the 3D shape in a local coordinate system, we can put the model in the global coordinate system by adding its" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.254, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Segment, Lift and Fit: Automatic 3D Shape Labeling from 2D Prompts" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.144, + 0.366, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.255, + 0.226, + 0.345, + 0.237 + ], + "angle": 0, + "content": "(a) RGB image" + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.144, + 0.501, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.392, + 0.226, + 0.479, + 0.237 + ], + "angle": 0, + "content": "(b) SAM mask" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.144, + 0.636, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.523, + 0.226, + 0.618, + 0.237 + ], + "angle": 0, + "content": "(c) Target mask" + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.164, + 0.772, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.663, + 0.226, + 0.747, + 0.237 + ], + "angle": 0, + "content": "(d) Occ. mask" + }, + { + "type": "image_caption", + "bbox": [ + 0.257, + 0.248, + 0.744, + 0.263 + ], + "angle": 0, + "content": "Fig. 4: Generation of occlusion map \\( O \\) for the mask alignment objective." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.291, + 0.785, + 0.337 + ], + "angle": 0, + "content": "pose \\(\\mathbf{p}\\). This gives us an unbounded SDF \\(\\phi(\\mathbf{x})\\) of the 3D model, in which the values outside the pre-defined cubic are filled with negative infinity. Notably, the rendering process of shape is differentiable w.r.t. both \\(\\mathbf{s}\\) and \\(\\mathbf{p}\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.357, + 0.584, + 0.373 + ], + "angle": 0, + "content": "3.4 Fit: Optimization over Pose and Shape" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.38, + 0.785, + 0.44 + ], + "angle": 0, + "content": "Given the 2D instance mask and the point cloud scene, we randomly initialize pose \\(\\mathbf{p}\\) and shape \\(\\mathbf{s}\\) and optimize them using Eq. (1). Specifically, the optimization objective consists of three terms: the mask alignment term, the point cloud alignment term, and the ground alignment term." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.441, + 0.787, + 0.641 + ], + "angle": 0, + "content": "Mask alignment term. The mask alignment term measures the dissimilarity between the target instance segmentation mask \\( Y \\) and the mask \\( Y_{proj} \\) obtained by projecting the 3D model onto the image plane. To make the projection process differentiable, we follow [22, 37] to define the projection as \\( \\pi(p_j) = 1 - \\prod_{\\mathbf{x} \\in \\mathcal{R}_{p_j}} \\frac{1}{e^{\\phi(\\mathbf{x})\\zeta + 1}} \\), where \\( p_j \\) is the \\( j \\)-th pixel in the image plane and \\( \\pi(p_j) \\in (0,1) \\) is the mask value. A ray is cast through \\( p_j \\) from the ego camera, and a set of points \\( \\mathcal{R}p_j \\) is collected by sampling along the ray. If the ray does not hit the surface of the 3D model, the SDF values of all sampled points \\( \\mathbf{x} \\in \\mathcal{R}p_j \\) will be negative, resulting in a mask value close to zero. Conversely, if the ray intersects with the 3D model, there must exist sampled points with \\( \\phi(\\mathbf{x}) > 0 \\), leading to a mask value close to 1. A hyperparameter \\( \\zeta \\) controls the \"hardness\" of the mask values, where a larger \\( \\zeta \\) leads to harder values (close to either 0 or 1). Finally, we define the mask alignment term using the dice loss [34] as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.382, + 0.648, + 0.785, + 0.683 + ], + "angle": 0, + "content": "\\[\nE _ {\\text {m a s k}} = 1 - \\frac {2 | (Y _ {\\text {p r o j}} * O) \\bigcap Y |}{| Y _ {\\text {p r o j}} * O | + | Y |}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.785, + 0.764 + ], + "angle": 0, + "content": "where \\( O \\) is a binary map indicating if the pixel is occluded by a nearer object. As shown in Fig. 4, to obtain the occlusion map, 3D LiDAR points are associated with each 2D mask based on projections, and the depth of each mask is computed as the median depth value of the associated LiDAR points. Masks are sorted by their depth values and then used to compute the occlusion maps." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Point cloud alignment term. Solely optimizing the mask alignment term leads to inaccurate estimation due to the missing depth information. To avoid undesired local optimum, we introduce an additional point cloud alignment term. This term measures how well the 3D object's surface aligns with the point cloud inside the corresponding frustum." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.344, + 0.128 + ], + "angle": 0, + "content": "J. Li et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.269 + ], + "angle": 0, + "content": "Intuitively, we can use the \\(\\ell-1\\) norm of the SDF \\(\\phi(\\mathcal{X}_f)\\) to determine how well a 3D model fits a given frustum point cloud, denoted as \\(\\mathcal{X}_f\\). However, this term alone might lead to sub-optimal solutions, as the points could be associated with the back face of the object. To overcome this issue, a second term is introduced that involves casting a ray for each point \\(\\mathbf{x} \\in \\mathcal{X}_f\\) starting from the ego camera and recording the coordinate \\(\\mathbf{y}\\) when the first time the ray intersects with the 3D model. This term penalizes the predicted 3D model if the distance between \\(\\mathbf{x}\\) and \\(\\mathbf{y}\\) is considerable. The point cloud alignment term can be written as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.366, + 0.28, + 0.786, + 0.32 + ], + "angle": 0, + "content": "\\[\nE _ {p c} = \\| \\phi (\\mathcal {X} _ {f}) \\| _ {1} + \\frac {1}{| \\hat {\\mathcal {X}} _ {f} |} \\sum_ {\\mathbf {x} \\in \\hat {\\mathcal {X}} _ {f}} \\| \\mathbf {x} - \\mathbf {y} \\| _ {2} ^ {2}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.328, + 0.785, + 0.357 + ], + "angle": 0, + "content": "\\(\\hat{\\mathcal{X}}_f\\) is a subset of \\(\\subseteq \\mathcal{X}_f\\) and represents points whose corresponding ray intersects with the model surface." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.36, + 0.787, + 0.42 + ], + "angle": 0, + "content": "Ground alignment term. Another strong cue that can be leveraged to refine the 3D pose is the height of the ground since we can assume that vehicles run on the road surface. Specifically, we regress a ground plane function \\( z = g(x,y) \\) using RANSAC [6] and then define a ground alignment objective as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.397, + 0.433, + 0.786, + 0.465 + ], + "angle": 0, + "content": "\\[\nE _ {\\text {g r o u n d}} = \\| \\hat {z} - \\frac {\\hat {h}}{2} - g (\\hat {x}, \\hat {y}) \\| _ {2} ^ {2}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.477, + 0.785, + 0.508 + ], + "angle": 0, + "content": "where \\((\\hat{x},\\hat{y},\\hat{z})\\) is the predicted object center, and \\(\\hat{h}\\) is the height of the object inferred from the predicted 3D shape." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.508, + 0.788, + 0.613 + ], + "angle": 0, + "content": "Optimization. The optimization objective is formulated as a weighted sum of \\( E_{mask} \\), \\( E_{pc} \\), and \\( E_{ground} \\). We initialize pose \\( \\mathbf{p} \\) with the median value of frustum points and shape \\( \\mathbf{s} \\) with mean shape \\( \\bar{\\mathbf{m}} \\). Based on the Adam optimizer [11] with a fixed learning rate of 0.1 and 150 iterations, a mini-batch comprising all instances (typically \\( < 100 \\)) in a scene is employed for optimization, as opposed to individual objects. This approach yields a \\( 5 \\times \\) acceleration with negligible accuracy drop." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.638, + 0.376, + 0.656 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.67, + 0.788, + 0.761 + ], + "angle": 0, + "content": "Previous auto-labelers can be divided into supervised ones and unsupervised ones. The proposed SLF is an unsupervised method. Therefore, we primarily compare it with existing unsupervised methods. Nonetheless, we also compare it with supervised methods and demonstrate that SLF exhibits superior transfer ability across datasets. In addition to bounding box accuracy, we also evaluate the quality of the predicted 3D shape." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.784, + 0.64, + 0.8 + ], + "angle": 0, + "content": "4.1 Comparison with Unsupervised Auto-labelers" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Dataset and metric. We conduct experiments on the commonly used KITTI [7] dataset. We follow the official training and validation split to get 3712 training" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.254, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Segment, Lift and Fit: Automatic 3D Shape Labeling from 2D Prompts" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.318, + 0.145, + 0.684, + 0.16 + ], + "angle": 0, + "content": "Table 1: Comparison with unsupervised auto-labelers." + }, + { + "type": "table", + "bbox": [ + 0.29, + 0.171, + 0.716, + 0.327 + ], + "angle": 0, + "content": "
MethodAPBEV(IoU=0.5)|R40AP3D(IoU=0.5)|R40
Easy Mod.HardEasy Mod.Hard
VS3D [31]74.566.757.640.337.431.1
SDFLabel [44]77.859.8-62.342.2-
McCraith et al. [23]75.576.668.6---
FGR [39]93.777.870.591.377.870.6
SLFGTmask89.288.987.187.687.883.8
SLFpoint88.588.584.284.884.980.8
SLFbox89.289.384.787.988.283.9
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.355, + 0.788, + 0.492 + ], + "angle": 0, + "content": "frames and 3768 validation frames. To align with prior works [15,39], the \"Car\" category is the focus of auto-labeling efforts. The Bird's Eye View (BEV) AP and the 3D AP are selected as evaluation metrics for a comprehensive comparison. Results are reported for all three levels of difficulty (Easy, Moderate, and Hard). The commonly used KITTI metrics are evaluated at a strict threshold of 0.7 IoU. As observed in [23,44], the 3D boxes in KITTI have varying spatial padding, making it challenging for human annotators to determine a precise size for a significant number of instances. Therefore, we follow the previous work [23,44] and set the IoU threshold as 0.5 and calculate AP by 40 recall positions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.492, + 0.788, + 0.628 + ], + "angle": 0, + "content": "Direct evaluation on generated pseudo labels. A straightforward evaluation is to compare the quality of auto-labeled 3D bounding boxes in terms of the canonical metric, Average Precision (AP). However, the AP metric needs a confidence score for each bounding box for thresholding and integration, but FGR [39] and SLF do not output confidence scores. To remedy this issue, we simply define IoU between the predicted mask and the target mask in SLF as a confidence score. As for FGR, we first project a 3D bounding box to image plane to get a 2D box, and the confidence score is defined by IoU between the predicted 2D box and the ground truth 2D box." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.765 + ], + "angle": 0, + "content": "Tab. 1 compares the proposed SLF with existing unsupervised methods on KITTI validation set. \\(\\mathrm{SLF}_{\\mathrm{point}}\\) uses 2D points as input, corresponding to the semi-automatic labeling scenario with humans in the loop. The point prompts are simulated by random sampling 3 points inside the ground truth masks [9]. \\(\\mathrm{SLF}_{\\mathrm{box}}\\) indicates 2D bounding boxes as input, corresponding to the full-automatic labeling scenario in which 2D detectors generate the prompts. For all methods that accept 2D boxes as inputs, we use ground truth boxes for a fair comparison. We also provide a variant of our method in which the SAM masks are replaced with ground truth masks from [9], denoted as \\(\\mathrm{SLF}_{\\mathrm{GTmask}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.842 + ], + "angle": 0, + "content": "The last three rows in Tab. 1 show the results of SLF with different 2D prompts. \\(\\mathrm{SLF}_{\\mathrm{point}}\\) and \\(\\mathrm{SLF}_{\\mathrm{box}}\\) perform roughly the same in terms of \\(\\mathrm{AP}_{BEV}\\), and \\(\\mathrm{SLF}_{\\mathrm{box}}\\) shows superior \\(\\mathrm{AP}_{3D}\\) performance. To our surprise, \\(\\mathrm{SLF}_{\\mathrm{box}}\\) is slightly better than \\(\\mathrm{SLF}_{\\mathrm{GTmask}}\\) in terms of both \\(\\mathrm{AP}_{3D}\\) and \\(\\mathrm{AP}_{BEV}\\) except on hard samples, which means the 2D mask generated by SAM from box prompts are even" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.344, + 0.127 + ], + "angle": 0, + "content": "J. Li et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.302, + 0.145, + 0.699, + 0.16 + ], + "angle": 0, + "content": "Table 2: Detector performance trained with pseudo labels." + }, + { + "type": "table", + "bbox": [ + 0.276, + 0.171, + 0.73, + 0.391 + ], + "angle": 0, + "content": "
DetectorsLabels\\(AP_{BEV}(IoU=0.5)\\) \\(|_{R40}\\)\\(AP_{3D}(IoU=0.5)\\) \\(|_{R40}\\)
EasyMod.HardEasyMod.Hard
PointPillar [14]GT95.794.393.295.793.691.4
FGR96.590.287.194.989.484.8
SLF95.394.191.195.292.188.8
SECOND [42]GT95.994.893.995.894.691.9
FGR95.488.986.295.388.685.7
SLF95.694.291.795.592.489.3
Part-\\(A^2\\) [32]GT96.094.393.996.094.291.9
FGR96.192.187.596.090.085.1
SLF96.094.591.895.993.989.6
Voxel R-CNN [3]GT98.494.994.598.394.894.4
FGR98.790.988.298.690.788.0
SLF96.295.192.496.194.791.9
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.424, + 0.785, + 0.455 + ], + "angle": 0, + "content": "better than those annotated by humans. This suggests that SLF benefits from more accurate input prompts, and more accurate instance masks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.457, + 0.787, + 0.638 + ], + "angle": 0, + "content": "We compare \\(\\mathrm{SLF}_{\\mathrm{box}}\\) with other unsupervised methods as they accept the same input prompts. It can be observed that SLF outperforms FGR and SDFLabel by a significant margin on the Moderate and Hard samples. However, SLF performs worse than FGR on the Easy samples, and we assume that easy cases in KITTI have much denser sub-points, which are beneficial to the segmentation and box estimation stage of FGR. Moreover, FGR consumes ground-truth amodal bounding box, while in KITTI these boxes are obtained by projecting 3D ground-truth boxes to 2D images, which may leak the oracle occlusion and size information of partially visible objects. SDFLabel uses a similar 3D shape representation to SLF but relies on a synthetic dataset for training, which may limit its performance on real-world data. In contrast, SLF does not require synthetic training data, and generalizes well to real-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.641, + 0.788, + 0.779 + ], + "angle": 0, + "content": "Detector performance trained with pseudo labels. Following previous work [15,24,39], to further evaluate the quality of generated pseudo labels, we use them as training samples to train off-the-shelf detectors and compare the results of different detectors. We mainly compare with the previous best-performing FGR auto-labeler and train 4 different 3D detectors. Specifically, for all detectors, we use the OpenPCDet [36] with default hyperparameters. In Tab. 2, we compare the results of detectors trained on KITTI training set with ground truth (GT) labels, FGR pseudo labels, and SLF pseudo labels, respectively. If not specified, SLF refers to \\(\\mathrm{SLF}_{\\mathrm{box}}\\) in the remaining sections." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Tab. 2 shows SLF pseudo labels significantly improve upon FGR pseudo labels on all detectors on Moderate and Hard samples. Regarding Easy samples, SLF pseudo labels are on par with FGR pseudo labels on all detectors except Voxel R-CNN [3]. When comparing specific detectors, the one trained with SLF" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.254, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Segment, Lift and Fit: Automatic 3D Shape Labeling from 2D Prompts" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.193 + ], + "angle": 0, + "content": "labels performs almost as well as the one trained with ground truth labels on both Easy and Moderate samples, and slightly worse on Hard samples, which shows the effectiveness of our method." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.218, + 0.504, + 0.232 + ], + "angle": 0, + "content": "4.2 Cross Dataset Generalization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.245, + 0.789, + 0.414 + ], + "angle": 0, + "content": "Apart from KITTI, we conduct experiments on more challenging nuScenes [1]. We transform the validation set of nuScenes to KITTI format, resulting in 6019 frames with corresponding front view images, point cloud and 3D bounding box annotations. On KITTI, we compare the pseudo label with ground truths and report \\(\\mathrm{AP}_{3D}\\), and on nuScenes, we report the official metrics including mAP, NDS, mATE, mASE and mAOE. We compare with the best-performing unsupervised auto-labeler FGR, and the supervised auto-labeler MTrans [15]. Both FGR and Mtrans require 2D amodal boxes as prompts, so we project 3d ground truth boxes into 2D as their input. Our SLF supports a more practical setting that only requires regular 2D modal boxes as prompts. Thus we can generate them with an off-the-shelf detector HTC [2] pre-trained on nuImages [1]." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.44, + 0.79, + 0.511 + ], + "angle": 0, + "content": "Table 3: Cross dataset evaluation compared with state-of-the-art unsupervised auto-labeler FGR and supervised auto-labeler MTrans. Results in gray are less informative because the NDS and error metrics are computed only on recalled samples, while FGR fails on recalling a majority of input samples (low mAP). \\(\\mathrm{mAOE}^{\\dagger}\\) means to change the periodicity of orientation to \\([0,\\pi]\\)." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.522, + 0.784, + 0.616 + ], + "angle": 0, + "content": "
MethodTraining DataAP3D@KITTInuScenes Metrics
EasyMod.HardmAPNDSmATEmASEmAOEmAOE†
FGR [39]-91.377.870.611.640.4.152.188.010.010
MTrans [15]KITTI99.699.697.145.949.4.358.349.702.225
SLF-87.988.283.957.255.2.372.192.836.113
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.789, + 0.842 + ], + "angle": 0, + "content": "In Tab. 3, on KITTI, MTrans shows a nearly perfect performance, this is not surprising since MTrans is a learnable neural network trained on KITTI, it easily learns the specific annotation patterns in training labels. Despite the high accuracy, we argue that learnable auto-labelers like MTrans are prone to overfit the biased pattern in the training data and generalize poorly across datasets. When we apply MTrans to the challenging nuScenes dataset, its performance dramatically decreases. SLF performs better in terms of mAP and NDS. Regarding the detailed error terms, we find SLF outperforms MTrans on size estimation (mASE), and is on par with MTrans on location estimation (mATE). Regarding orientation estimation, SLF underperforms MTrans by 0.134. However, we find a typical orientation error is the model mistakes the front end of the car for the rear end, which is usually caused when the LiDAR points are rather sparse. To explore how good the orientation predictions are when the \\(\\pi\\)" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.344, + 0.128 + ], + "angle": 0, + "content": "J. Li et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.254 + ], + "angle": 0, + "content": "phase lag is disentangled, we change the periodicity of orientation to \\([0,\\pi ]\\) and re-measure the orientation error, denoted as \\(\\mathrm{mAOE}^{\\dagger}\\). SLF outperforms MTrans on \\(\\mathrm{mAOE}^{\\dagger}\\), suggesting that SLF predicts more accurate orientation, the heading direction aligning better with ground truths. When applying FGR on the nuScences dataset, it fails on a majority of samples, resulting in a low mAP of \\(11.6\\%\\). Consequently, the NDS score and error terms are less informative since they are computed only on successfully labeled samples, typically easy samples." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.283, + 0.355, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.355, + 0.285, + 0.491, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.438, + 0.496, + 0.522 + ], + "angle": 0, + "content": "Fig.5: Qualitative results of estimated shapes. From left to right: projected mask of the initial 3D model; projected mask of the optimized 3D model; mean shape and surrounding point cloud; optimized 3D shape and surrounding point cloud." + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.284, + 0.785, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.438, + 0.788, + 0.522 + ], + "angle": 0, + "content": "Fig.6: Comparison of occupancy annotation. From left to right: reference instance image; occupancy labels from OFN [25]; occupancy labels from OpenOccupancy [38]; occupancy labels generated by our SLF." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.582, + 0.504, + 0.598 + ], + "angle": 0, + "content": "4.3 Voxel Occupancy Annotation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Since benchmarks [1,7] (i.e., KITTI and nuScences) do not support quantitative evaluation on 3D shapes, we show qualitative results in Fig. 5 to better show the effect of SLF on shape labeling. Specifically, we replace the estimated shape with a mean shape and show comparisons. A clear improvement can be observed by comparing the initial shape and the final optimized shape, both in terms of mask alignment and LiDAR point alignment. Besides, we observe that SLF can serve as a good auto-labeler for occupancy annotation of dynamic objects such as vehicles. We compare two versions of occupancy labels for nuScenes, OpenOccupancy [38] and OFN [25]. Considering that existing occupancy labels are still of low quality, quantitative evaluation may be misleading. Therefore we show qualitative visualizations in Fig. 6, and we observe that OpenOccupancy and OFN labels fail to depict the inner occupancy of the cars. This is mainly attributed to that they label the occupancy by voxelizing the LiDAR points, and therefore self-occlusion can not be avoided. In contrast, SLF can predict dense occupancy more accurately, with more detailed shapes." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.253, + 0.114, + 0.732, + 0.131 + ], + "angle": 0, + "content": "Segment, Lift and Fit: Automatic 3D Shape Labeling from 2D Prompts" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.128 + ], + "angle": 0, + "content": "13" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.391, + 0.163 + ], + "angle": 0, + "content": "4.4 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.311 + ], + "angle": 0, + "content": "In this section, we conduct comprehensive ablation studies on the KITTI validation set and report the \\(\\mathrm{AP}_{3D}(\\mathrm{IoU} = 0.5)|_{R40}\\) results. Except in Tab. 6, we input 2D points to SLF, other results are obtained by using 2D bounding boxes as prompt. Effect of energy terms. In Tab. 4, we observe that introducing the mask alignment term significantly improves pseudo-label quality rather than solely applying the point cloud alignment term, and adding the ground alignment term further improves performance. However, as we mentioned above, using mask alignment alone can easily be trapped in sub-optimal solutions, thus it can not obtain valid results (Easy: 6.9, Mod.: 5.9, Hard: 5.6)." + }, + { + "type": "table_caption", + "bbox": [ + 0.241, + 0.34, + 0.484, + 0.354 + ], + "angle": 0, + "content": "Table 4: Ablation on energy terms." + }, + { + "type": "table", + "bbox": [ + 0.226, + 0.365, + 0.522, + 0.439 + ], + "angle": 0, + "content": "
ObjectiveEasyMod.Hard
Epc70.564.862.9
Emask+Epc87.388.083.7
Emask+Epc+Eground87.988.283.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.526, + 0.34, + 0.752, + 0.354 + ], + "angle": 0, + "content": "Table 5: Ablation on 3D models." + }, + { + "type": "table", + "bbox": [ + 0.528, + 0.366, + 0.75, + 0.438 + ], + "angle": 0, + "content": "
# modelsEasyMod.Hard
574.770.164.0
3084.283.677.3
7987.988.283.9
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.467, + 0.787, + 0.527 + ], + "angle": 0, + "content": "Effect of the number of CAD models. In Tab. 5, we randomly sample 5 and 30 CAD models out of the total 79 samples to investigate the influence of the number of CAD models used for PCA. From the results, we can conclude that better result comes with more models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.528, + 0.788, + 0.603 + ], + "angle": 0, + "content": "Effect of the PCA dimensions. In Tab. 7, we show how different PCA dimensions affect results. Performance improves when we increase the dimension from 3 to 5. More dimensions like 10 do not further improve. This is possibly attributed to that a higher dimension of shape representation may complicate the optimization process, which means it may need more steps to converge." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.605, + 0.788, + 0.697 + ], + "angle": 0, + "content": "Effect of the number of points. In Tab. 6, we analyze the effect of the number of points input to SAM [12]. When we increase the points from 3 to 8, the improvement of pseudo-label quality is minor. However, we observe a significant decrease in the easy cases of KITTI when using one point as the prompt, which is reasonable as easy cases often have large masks and a single point can not provide sufficient context for SAM to generate high-quality results." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.721, + 0.403, + 0.736 + ], + "angle": 0, + "content": "4.5 Further Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.749, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Robustness of LiDAR Beam. Point clouds in KITTI [7] are collected by 64 beams LiDAR, we downsample point cloud to 32, 16 and 8 beams following [40]. \\(\\mathrm{AP}_{3D}\\) results on KITTI validation set in Tab. 8 show SLF can still get decent results on 32 and 16 beams LiDAR, and the performances downgrade rapidly on Mod. and Hard cases, since these cases usually far away from ego vehicle which exacerbating the sparsity of point clouds." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.344, + 0.127 + ], + "angle": 0, + "content": "J. Li et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.232, + 0.145, + 0.375, + 0.159 + ], + "angle": 0, + "content": "Table 6: No. points." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.169, + 0.407, + 0.264 + ], + "angle": 0, + "content": "
# PointsEasyMod.Hard
384.884.980.8
178.181.178.8
584.885.081.0
885.085.280.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.438, + 0.145, + 0.566, + 0.159 + ], + "angle": 0, + "content": "Table 7: No. dim." + }, + { + "type": "table", + "bbox": [ + 0.42, + 0.171, + 0.584, + 0.264 + ], + "angle": 0, + "content": "
Dim.EasyMod.Hard
587.988.283.9
387.187.883.6
886.987.184.1
1084.885.481.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.63, + 0.145, + 0.768, + 0.159 + ], + "angle": 0, + "content": "Table 8: No. beam." + }, + { + "type": "table", + "bbox": [ + 0.613, + 0.171, + 0.782, + 0.264 + ], + "angle": 0, + "content": "
BeamEasyMod.Hard
6487.988.283.9
3281.377.577.3
1680.776.370.0
875.160.056.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.24, + 0.28, + 0.486, + 0.294 + ], + "angle": 0, + "content": "Table 9: Dilation on instance mask." + }, + { + "type": "table", + "bbox": [ + 0.243, + 0.305, + 0.487, + 0.399 + ], + "angle": 0, + "content": "
KernelEasyMod.Hard
FGR(no dil.)91.377.870.6
SLF(no dil.)87.988.283.9
5x586.385.681.2
9x983.980.576.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.28, + 0.764, + 0.294 + ], + "angle": 0, + "content": "Table 10: Erosion on instance mask." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.305, + 0.764, + 0.399 + ], + "angle": 0, + "content": "
KernelEasyMod.Hard
FGR(noERO.)91.377.870.6
SLF(noERO.)87.988.283.9
5x585.283.979.5
9x982.178.272.1
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.429, + 0.788, + 0.507 + ], + "angle": 0, + "content": "Robustness of Mask Quality. We simulate low-quality masks by applying erosion and dilation operations on SAM [12] masks and report \\(\\mathrm{AP}_{3D}\\) results of KITTI validation set in Tab. 9 and Tab. 10. We observe even with severely corrupted masks, SLF still outperforms the existing unsupervised auto-annotation methods in Mod. and Hard cases, which show the robustness of SLF." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.535, + 0.356, + 0.551 + ], + "angle": 0, + "content": "5 Limitation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.574, + 0.787, + 0.621 + ], + "angle": 0, + "content": "The proposed SLF uses a strong shape prior learned by PCA from given 3D CAD models. The shape prior is category-specific, which means we need to prepare a collection of 3D models when labeling a specific semantic class." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.651, + 0.36, + 0.667 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "This paper proposes a Segment, Lift, and Fit paradigm for automatic 3D shape annotation of objects from the 2D point or box prompts. We first segment high-quality instance masks from the prompts using the SAM, and then lift 2D masks to 3D forms, where gradient descent is used to adjust their poses and shapes until the projections fits the masks and the surfaces conform to surrounding LiDAR points. Extensive experimental results on multiple datasets for 3D detection demonstrate the effectiveness of our SLF method. Moreover, the SLF auto-labeler achieves promising results in detailed shape predictions, which can be used for occupancy annotation. In the future, we will investigate the generalization abilities by applying our SLF to more categories." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.254, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Segment, Lift and Fit: Automatic 3D Shape Labeling from 2D Prompts" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.218 + ], + "angle": 0, + "content": "1. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. arXiv preprint arXiv:1903.11027 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.219, + 0.785, + 0.26 + ], + "angle": 0, + "content": "2. Chen, K., Pang, J., Wang, J., Xiong, Y., Li, X., Sun, S., Feng, W., Liu, Z., Shi, J., Ouyang, W., et al.: Hybrid task cascade for instance segmentation. In: CVPR. pp. 4974-4983 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.261, + 0.785, + 0.287 + ], + "angle": 0, + "content": "3. Deng, J., Shi, S., Li, P., Zhou, W., Zhang, Y., Li, H.: Voxel r-cnn: Towards high performance voxel-based 3d object detection. arXiv:2012.15712 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.288, + 0.785, + 0.343 + ], + "angle": 0, + "content": "4. Engelmann, F., Stickler, J., Leibe, B.: Joint object pose estimation and shape reconstruction in urban street scenes using 3d shape priors. In: Pattern Recognition: 38th German Conference, GCPR 2016, Hannover, Germany, September 12-15, 2016, Proceedings 38. pp. 219-230. Springer (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.344, + 0.785, + 0.384 + ], + "angle": 0, + "content": "5. Engelmann, F., Stuckler, J., Leibe, B.: Samp: shape and motion priors for 4d vehicle reconstruction. In: 2017 IEEE Winter Conference on Applications of Computer Vision (WACV). pp. 400-408. IEEE (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.385, + 0.785, + 0.425 + ], + "angle": 0, + "content": "6. Fischler, M.A., Bolles, R.C.: Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM 24(6), 381-395 (1981)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.426, + 0.785, + 0.453 + ], + "angle": 0, + "content": "7. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: CVPR (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.454, + 0.785, + 0.495 + ], + "angle": 0, + "content": "8. Guo, J., Liu, J., Xu, D.: 3d-pruning: A model compression framework for efficient 3d action recognition. IEEE Transactions on Circuits and Systems for Video Technology 32(12), 8717-8729 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.495, + 0.785, + 0.537 + ], + "angle": 0, + "content": "9. Heylen, J., De Wolf, M., Dawagne, B., Proesmans, M., Van Gool, L., Abbeloos, W., Abdelkawy, H., Reino, D.O.: Monocinis: Camera independent monocular 3d object detection using instance segmentation. In: ICCV. pp. 923-934 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.537, + 0.785, + 0.564 + ], + "angle": 0, + "content": "0. Huang, Y., Zheng, W., Zhang, Y., Zhou, J., Lu, J.: Tri-perspective view for vision-based 3d semantic occupancy prediction. arXiv preprint arXiv:2302.07817 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.565, + 0.785, + 0.592 + ], + "angle": 0, + "content": "1. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. In: ICLR (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.592, + 0.785, + 0.633 + ], + "angle": 0, + "content": "2. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. arXiv:2304.02643 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.633, + 0.785, + 0.674 + ], + "angle": 0, + "content": "3. Kundu, A., Li, Y., Rehg, J.M.: 3d-rcnn: Instance-level 3d object reconstruction via render-and-compare. In: CVPR (2018). https://doi.org/10.1109/CVPR.2018.00375, https://doi.org/10.1109/CVPR.2018.00375" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.675, + 0.785, + 0.716 + ], + "angle": 0, + "content": "4. Lang, A.H., Vora, S., Caesar, H., Zhou, L., Yang, J., Beijbom, O.: Pointpillars: Fast encoders for object detection from point clouds. In: CVPR. pp. 12697-12705 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.716, + 0.785, + 0.744 + ], + "angle": 0, + "content": "5. Liu, C., Qian, X., Huang, B., Qi, X., Lam, E., Tan, S.C., Wong, N.: Multimodal transformer for automatic 3d annotation and object detection. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.744, + 0.785, + 0.771 + ], + "angle": 0, + "content": "6. Liu, J., Guo, J., Xu, D.: Aptsnet: Toward adaptive point sampling for efficient 3d action recognition. IEEE Transactions on Image Processing 31, 5287-5302 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.771, + 0.785, + 0.799 + ], + "angle": 0, + "content": "7. Liu, J., Guo, J., Xu, D.: Geometrhythmotion-transformer: An end-to-end framework for 3d action recognition. IEEE Transactions on Multimedia (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "8. Liu, J., He, T., Yang, H., Su, R., Tian, J., Wu, J., Guo, H., Xu, K., Ouyang, W.: 3d-queryis: A query-based framework for 3d instance segmentation. arXiv preprint arXiv:2211.09375 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.343, + 0.127 + ], + "angle": 0, + "content": "J. Li et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.204 + ], + "angle": 0, + "content": "19. Liu, J., Li, J., Wang, K., Guo, H., Yang, J., Peng, J., Xu, K., Liu, X., Guo, J.: Lta-pcs: Learnable task-agnostic point cloud sampling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 28035-28045 (June 2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.205, + 0.786, + 0.245 + ], + "angle": 0, + "content": "20. Liu, J., Xu, D.: Geometrhythmotion-net: A strong two-stream baseline for 3d action recognition. IEEE Transactions on Circuits and Systems for Video Technology 31(12), 4711-4721 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.786, + 0.273 + ], + "angle": 0, + "content": "21. Liu, S., Li, T., Chen, W., Li, H.: Soft rasterizer: A differentiable renderer for image-based 3d reasoning. ICCV (Oct 2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.786, + 0.3 + ], + "angle": 0, + "content": "22. Liu, Z., Zhou, D., Lu, F., Fang, J., Zhang, L.: Autoshape: Real-time shape-aware monocular 3d object detection. In: ICCV. pp. 15641-15650 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.786, + 0.328 + ], + "angle": 0, + "content": "23. McCraith, R., Insafutdinov, E., Neumann, L., Vedaldi, A.: Lifting 2d object locations to 3d by discounting lidar outliers across objects and views (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.786, + 0.356 + ], + "angle": 0, + "content": "24. Meng, Q., Wang, W., Zhou, T., Shen, J., Van Gool, L., Dai, D.: Weakly supervised 3d object detection from lidar point cloud. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.786, + 0.384 + ], + "angle": 0, + "content": "25. Ming Fang, Z.L.: Occupancy dataset for nuscenes. https://github.com/FANG-MING/occupancy-for-nuscenes (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.786, + 0.424 + ], + "angle": 0, + "content": "26. Osher, S., Fedkiw, R.: Signed Distance Functions. Springer New York (2003). https://doi.org/10.1007/0-387-22746-6_2, https://doi.org/10.1007/0-387-22746-6_2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.424, + 0.786, + 0.466 + ], + "angle": 0, + "content": "27. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: CVPR (June 2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.466, + 0.786, + 0.509 + ], + "angle": 0, + "content": "28. Peng, L., Liu, F., Yu, Z., Yan, S., Deng, D., Yang, Z., Liu, H., Cai, D.: Lidar point cloud guided monocular 3d object detection. In: European Conference on Computer Vision. pp. 123-139. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.786, + 0.563 + ], + "angle": 0, + "content": "29. Prisacariu, V.A., Segal, A.V., Reid, I.: Simultaneous monocular 2d segmentation, 3d pose recovery and 3d reconstruction. In: Lee, K.M., Matsushita, Y., Rehg, J.M., Hu, Z. (eds.) ACCV. pp. 593-606. Springer Berlin Heidelberg, Berlin, Heidelberg (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.786, + 0.592 + ], + "angle": 0, + "content": "30. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. arXiv preprint arXiv:1612.00593 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.786, + 0.619 + ], + "angle": 0, + "content": "31. Qin, Z., Wang, J., Lu, Y.: Weakly supervised 3d object detection from point clouds. ACMMM (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.786, + 0.661 + ], + "angle": 0, + "content": "32. Shi, S., Wang, Z., Shi, J., Wang, X., Li, H.: From points to parts: 3d object detection from point cloud with part-aware and part-aggregation network. IEEE transactions on pattern analysis and machine intelligence 43(8), 2647-2664 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.786, + 0.702 + ], + "angle": 0, + "content": "33. Song, X., Wang, P., Zhou, D., Zhu, R., Guan, C., Dai, Y., Su, H., Li, H., Yang, R.: Apollocar3d: A large 3d car instance understanding benchmark for autonomous driving. In: CVPR. pp. 5452-5462 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.786, + 0.785 + ], + "angle": 0, + "content": "34. Sudre, C.H., Li, W., Vercauteren, T., Ourselin, S., Jorge Cardoso, M.: Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations. In: Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: Third International Workshop, DLMIA 2017, and 7th International Workshop, ML-CDS 2017, Held in Conjunction with MICCAI 2017, Quebec City, QC, Canada, September 14, Proceedings 3. pp. 240-248. Springer (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.786, + 0.813 + ], + "angle": 0, + "content": "35. Tang, Y.S., Lee, G.H.: Transferable semi-supervised 3d object detection from rgb-d data. In: ICCV. pp. 1931-1940 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.786, + 0.841 + ], + "angle": 0, + "content": "36. Team, O.D.: Openpcdet: An open-source toolbox for 3d object detection from point clouds. https://github.com/open-mmlab/OpenPCDet (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.254, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Segment, Lift and Fit: Automatic 3D Shape Labeling from 2D Prompts" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "37. Wang, R., Yang, N., Stueckler, J., Cremers, D.: Directshape: Photometric alignment of shape priors for visual vehicle pose and shape estimation (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.175, + 0.785, + 0.217 + ], + "angle": 0, + "content": "38. Wang, X., Zhu, Z., Xu, W., Zhang, Y., Wei, Y., Chi, X., Ye, Y., Du, D., Lu, J., Wang, X.: Openoccupancy: A large scale benchmark for surrounding semantic occupancy perception. arXiv preprint arXiv:2303.03991 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.244 + ], + "angle": 0, + "content": "39. Wei, Y., Su, S., Lu, J., Zhou, J.: Fgr: Frustum-aware geometric reasoning for weakly supervised 3d vehicle detection (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.785, + 0.286 + ], + "angle": 0, + "content": "40. Wei, Y., Wei, Z., Rao, Y., Li, J., Zhou, J., Lu, J.: Lidar distillation: Bridging the beam-induced domain gap for 3d object detection. In: ECCV. pp. 179-195. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.327 + ], + "angle": 0, + "content": "41. Wei, Y., Zhao, L., Zheng, W., Zhu, Z., Zhou, J., Lu, J.: Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving. arXiv preprint arXiv:2303.09551 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.355 + ], + "angle": 0, + "content": "42. Yan, Y., Mao, Y., Li, B.: Second: Sparsely embedded convolutional detection. Sensors 18(10), 3337 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.785, + 0.41 + ], + "angle": 0, + "content": "43. Yang, H., He, T., Liu, J., Chen, H., Wu, B., Lin, B., He, X., Ouyang, W.: Gdmae: generative decoder for mae pre-training on lidar point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9403-9414 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.785, + 0.439 + ], + "angle": 0, + "content": "44. Zakharov, S., Kehl, W., Bhargava, A., Gaidon, A.: Autolabeling 3d objects with differentiable rendering of sdf shape priors. In: CVPR (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.439 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Segment, Lift and Fit_ Automatic 3D Shape Labeling from 2D Prompts/478f2a06-4d04-42d7-815f-2c137ffb864d_origin.pdf b/2024/Segment, Lift and Fit_ Automatic 3D Shape Labeling from 2D Prompts/478f2a06-4d04-42d7-815f-2c137ffb864d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..727e4c7beb460741b6c907603b2dcd287825078f --- /dev/null +++ b/2024/Segment, Lift and Fit_ Automatic 3D Shape Labeling from 2D Prompts/478f2a06-4d04-42d7-815f-2c137ffb864d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:96771b1de8faa1ba6b2653f2c6044ed56d31998a63bb1bcd27027c8e3cb70d04 +size 1833576 diff --git a/2024/Segment, Lift and Fit_ Automatic 3D Shape Labeling from 2D Prompts/full.md b/2024/Segment, Lift and Fit_ Automatic 3D Shape Labeling from 2D Prompts/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5e189b6f7bee4cee8c6cc4adcf8c361e39c43e17 --- /dev/null +++ b/2024/Segment, Lift and Fit_ Automatic 3D Shape Labeling from 2D Prompts/full.md @@ -0,0 +1,308 @@ +# Segment, Lift and Fit: Automatic 3D Shape Labeling from 2D Prompts + +Jianhao Li $^{1}$ , Tianyu Sun $^{2}$ , Zhongdao Wang $^{3\dagger}$ , Enze Xie $^{3}$ , Bailan Feng $^{3}$ , Hongbo Zhang $^{3}$ , Ze Yuan $^{1}$ , Ke Xu $^{1}$ , Jiaheng Liu $^{4\dagger}$ , and Ping Luo $^{5}$ + +1 Beihang University, Beijing, China + +2 Tsinghua University, Beijing, China + +3 Noah's Ark Lab, Huawei + +4 Nanjing University + +5 The University of Hong Kong + +{lijianhao, yuanze1024, kexu}@buaa.edu.cn + +{wangzhongdao, Johnny_ez, fengbailan, zhanghongbo888}@huawei.com + +sty21@mails.tsinghua.edu.cn, buaaljiaheng@gmail.com, pluo.lhi@gmail.com + +Abstract. This paper proposes an algorithm for automatically labeling 3D objects from 2D point or box prompts, especially focusing on applications in autonomous driving. Unlike previous arts, our auto-labeler predicts 3D shapes instead of bounding boxes and does not require training on a specific dataset. We propose a Segment, Lift, and Fit (SLF) paradigm to achieve this goal. Firstly, we segment high-quality instance masks from the prompts using the Segment Anything Model (SAM) and transform the remaining problem into predicting 3D shapes from given 2D masks. Due to the ill-posed nature of this problem, it presents a significant challenge as multiple 3D shapes can project into an identical mask. To tackle this issue, we then lift 2D masks to 3D forms and employ gradient descent to adjust their poses and shapes until the projections fit the masks and the surfaces conform to surrounding LiDAR points. Notably, since we do not train on a specific dataset, the SLF auto-labeler does not overfit to biased annotation patterns in the training set as other methods do. Thus, the generalization ability across different datasets improves. Experimental results on the KITTI dataset demonstrate that the SLF auto-labeler produces high-quality bounding box annotations, achieving an AP@0.5 IoU of nearly $90\%$ . Detectors trained with the generated pseudo-labels perform nearly as well as those trained with actual ground-truth annotations. Furthermore, the SLF auto-labeler shows promising results in detailed shape predictions, providing a potential alternative for the occupancy annotation of dynamic objects. + +Keywords: 3D Auto-labelers $\cdot$ Shape Optimization $\cdot$ Training Free + +# 1 Introduction + +In modern robotics and autonomous driving, a significant amount of labeled data is required for supervised learning to understand the 3D scene [8,16,17,19, + +![](images/5e3f82b4b5a52d18b0696738ee5adf02a82ff4230af217d24bf50743cb51ac8d.jpg) +(a) Human Annotation + +![](images/30ab53df3ef06079ff2931a42e8d06cb898243baa547769543e4a9482ab4c689.jpg) +(b) Auto-labeler Annotation using SLF +Fig. 1: Comparison between human annotation and auto-labeler annotation using SLF. + +[20], emphasizing dynamic objects such as vehicles and pedestrians. Therefore, there is a pressing need to improve the degree of automation for 3D labeling. The preferred labeling format is typically the 3D bounding box, which can be represented compactly with as few as 7-9 degrees of freedom. However, manually labeling a large number of 3D boxes remains a laborious and costly task, posing challenges for scaling up 3D object detectors as shown in Fig. 1(a). + +Meanwhile, as 3D perception models evolve, there is an increasing demand for more precise labeling granularity. While bounding boxes offer a compact representation, they are not optimal for 3D objects and are being replaced by finer representations. For instance, voxel occupancy [10,38,41] has emerged as a dominant format for 3D scenes in autonomous driving, as it unifies the representation of static road elements and dynamic objects. However, finer-grained labels further complicate the annotation pipeline, exacerbating the annotation efficiency problem and presenting a significant obstacle to scaling up 3D models. + +In this paper, we propose a novel approach for 3D auto-labeling. The proposed auto-labeler follows a Segment, Lift, and Fit (SLF) paradigm as shown in Fig. 1(b). Specifically, SLF utilizes the 2D points and bounding boxes as prompts and outputs the estimated 3D shapes and poses of target objects, thus supporting semi- or fully automatic labeling depending on whether the prompts come from humans or machines. To accomplish this, we leverage the Segment Anything Model (SAM) [12], a foundational vision model, to generate high-quality instance masks from the input prompts. Subsequently, we lift 2D instance masks into 3D forms with Signed Distance Functions (SDFs) [26], and then iteratively optimize the shapes and poses until their projections fit the mask and its surface conforms to the surrounding LiDAR points. Notably, the proposed SLF auto-labeler does not require supervised training on a specific dataset, and is therefore not susceptible to overfitting biased annotation patterns in the training dataset. + +Experimental results show that SLF can generate high-quality 3D labels. As our method outputs detailed shapes of objects, while existing benchmarks [1,7] only provide bounding box annotations, for a fair comparison with existing auto-labelers, we convert shapes to boxes and assess the label quality regarding box accuracy. When compared among unsupervised auto-labelers [23,39,44] that do + +not train with 3D labels as ours, our method outperforms the best-performing FGR [39] auto-labeler in both settings of direct evaluation on pseudo labels and training object detectors with generated pseudo labels. Moreover, because SLF does not learn biased annotation patterns from the training data, it significantly outperforms state of art supervised model MTrans [15] on "unseen" datasets, demonstrating its superior generalization capacity. We also show qualitatively that the predicted shapes align well with the actual geometry of the target objects, underpinning a new possibility for voxel occupancy annotation [38,41]. Our contribution can be summarized as follows: + +- We propose SLF, a 3D auto-labeler that improves existing methods by predicting detailed shapes and poses rather than just bounding boxes. The shape and pose are directly optimized using gradient descent without requiring a neural network trained on a small portion of labels, which enables a "cold-start" scenario for auto-labeling. +- We demonstrate that SLF generates superior labels to other unsupervised auto-labelers. Additionally, it shows good generalization ability across different datasets, surpassing supervised auto-labelers. +- The predicted shapes align well with the actual geometries of the objects, suggesting that SLF can serve as a viable alternative solution for annotating fine-grained occupancy of dynamic objects. + +# 2 Related Works + +3D Auto-labelers. Compared to supervised 3D object detectors [14,18,42,43], auto-labelers require fewer or no training labels. Supervised auto-labelers learn to annotate from limited samples with ground-truth 3D labels, often in the hundreds. For instance, WS3D [24] generates cylindrical object proposals on bird's eye view (BEV) maps using center clicks in BEV and fine-tunes a 3D detector pre-trained on a few human-annotated data. MTrans [15] employs a multi-task design of segmentation, point generation, and box regression to train a 3D detector that takes ground truth 2D boxes and LiDAR points as input. Unsupervised auto-labelers do not require 3D labels for training. CC [35] uses PointNet [30] to segment and regress 3D boxes from frustum points. SDFLabel [44] trains a DeepSDF [27] network supervised with image patch and CAD models of cars on a synthetic dataset, then uses RANSAC [6] to fit the predicted shape to frustum sub-points. VS3D [31] employs an unsupervised UPM module to generate 3D proposals and filter them using a pre-trained image model. Our SLF, is also an unsupervised approach that does not require training with 3D labels. SLF achieves finer annotations with weaker prompts compared with a closely related unsupervised and non-learning approach, FGR [39]. FGR requires amodal boxes as inputs, in which oracle occlusion information has already been considered. LPCG [28] finds that concerning label accuracy, the 3D location part in the label is preferred compared to other parts of labels in monocular 3D object detection and proposed training-free and high-cost label pipeline, which is similar to FGR and MTrans respectively. In contrast, SLF takes normal 2D + +modal bounding boxes as input, which an off-the-shelf 2D detector can easily obtain. Furthermore, SLF can produce 3D object shapes rather than just bounding boxes. + +Shape Prior. To reconstruct the 3D geometry for instances from a specific semantic class, e.g., cars or pedestrians, the shape prior of the given class is often utilized to ease the estimation. For instance, McCraith et al. [23] uses a fixed car shape as the most straightforward shape prior. Autoshape [22] develops a shape-labeler for the car class, but it requires ground truth 3D bounding box and amodal instance mask which contains oracle occlusion information as input and it uses part-based deformable representation as shape prior which needs to label the key points and different part of CAD model. Another usage of shape prior is to encode it into a compact low-dimensional space with PCA [4,5,29,37] or neural network [13,44]. SDFLabel [44] pretrains a DeepSDF [27] model to predict the hidden representation of shape conditioned on cropped image patches. 3D-RCNN [13] regresses poses and shape parameters in a single forward pass. However, in 3D-RCNN shape representation is not differentiable, and the finite differences that are used for gradient approximation cause ambiguity during training. Upon the compact PCA representation, Engelmann et al. [4] and DirectShape [37] optimize the shape and pose concurrently with the stereo reconstruction of the object to refine the 3D detection result. SAMP [5] takes sequence of stereo depth maps and 3D trajectory to generate target's shape and refine its trajectory. Prisacariu et al. [29] requires sequence of instance masks and results of 3D tracker to optimize 3D model. Similar to previous PCA based method, our method adopts the PCA shape prior in a differentiable optimization process, however, SLF is training-free and only requires a single raw point cloud and RGB image with 2D prompts, which is more practical. + +# 3 Approach + +As shown in Fig. 2(a), we aim to recover the 3D shape and pose of an interested object by asking a labeler to click one or more points or draw a box on the image as the prompt for the target object. A labeler could be a human in semiautomatic labeling and a machine algorithm such as a 2D object detector in case of full-automatic labeling. Fig. 2(a) shows a brief illustration. We propose a Segment, Lift, and Fit (SLF) paradigm to address this problem. We first segment a 2D mask of the target object using the input prompts, then lift it into a 3D model and iteratively optimize its shape and pose until it fits the 2D mask and surrounding LiDAR points. An overall illustration is shown in Fig. 2(b). + +# 3.1 Segment: From 2D Prompts to Instance Masks + +The problem of auto-labeling 3D shapes from 2D prompts is a significant issue in autonomous driving because it would be highly beneficial to effectively utilize the already established robust 2D instance segmentation models for this purpose. This work employs the strong Segment Anything Model (SAM) [12] to transfer + +![](images/e4e36a39259e7796dca42e05a0704ea8c03dda066acfc7caa8da949a34313bdf.jpg) +(a) Definition & challenge + +![](images/b4978b9de0169b630f9a0e8efb1bf8581ca0b749110a2477a8b876175954700b.jpg) +(b) Overall illustration of the proposed method + +![](images/13046f29a8c3bb633932bc7285a2df9e30b8e2375e6ca34107f801167761e2a0.jpg) +Fig. 2: (a) We aim to recover the 3D shape and pose of an interested object, given 2D box or point as prompts. This problem is highly ill-posed. (b) In our SLF, we propose to firstly segment the 2D mask of the target object, and then lift the mask to a 3D form and optimize over its shape and pose by gradient descent until the 3D object fits the mask and LiDAR points. + +the input point or box prompts into 2D instance masks. Then, our problem is transferred to estimate the 3D shapes and poses from given instance masks. Despite the high quality of SAM masks, the problem is still extremely difficult to solve. This is because multiple 3D shapes can be mapped to the same 2D mask (Fig. 2(a)), making it a highly ill-posed problem and posing a significant challenge for auto-labeler methods. + +# 3.2 Optimization Objective + +We employ a pose vector $\mathbf{p} = (x,y,z,\theta) \in \mathbb{R}^4$ and a latent shape vector $\mathbf{s} \in \mathbb{R}^d$ to represent a 3D object, where $(x,y,z)$ denote the position of the object center in the global coordinate system, $\theta$ denotes the heading angle around the $z$ axis, and the latent dimension $d$ is empirically set to 5. We assume all instances are on the ground (the $x-y$ plane) and omit rotations around the $x$ and $y$ axes. Our goal is to minimize an energy function that corresponds to the posterior probability of the 3D reconstruction given $\mathbf{p}$ and $\mathbf{s}$ as follows: + +$$ +\mathbf {p} ^ {*}, \mathbf {s} ^ {*} = \underset {\mathbf {p}, \mathbf {s}} {\arg \min } E (Y, \mathcal {X} _ {f}, \mathbf {p}, \mathbf {s}), \tag {1} +$$ + +where $Y \in [0,1]^{h \times w}$ is the 2D instance mask predicted by SAM, and $\mathcal{X}_f \in \mathbb{R}^{N \times 3}$ is a subset of the whole scene point cloud $\mathcal{X}$ . $\mathcal{X}_f$ is comprised of $N$ three-dimensional coordinates of points inside the camera frustum corresponding to the 2D mask. To solve the optimization problem in Eq. (1), we build a differentiable renderer to optimize the energy function $E$ via gradient descent. + +# 3.3 Lift: Differentiable Rendering + +Representation of 3D objects. We use the Signed Distance Function (SDF) [22, 37] to represent 3D objects. Formally, an SDF is a three-dimensional scalar field + +![](images/6bd121f5a2fda56b47886bf8930e25e0255479b32030c4ded30316ae91319d6a.jpg) +Fig. 3: We perform PCA on a diverse collection of 3D shapes to obtain the shape latent code. Left: Example of the basis 3D shapes in the collection. Right: Interpolation along the latent space of the shape code (each row) shows smooth shape variations. + +$\phi (\mathbf{x})$ in which the value of a particular spatial coordinate $\mathbf{x}$ is given by its distance to the closest object surface, with positive and negative values meaning exterior and interior area. In practice, we use the discrete approximation of continuous SDF. Consider a cubic divided into a 3D mesh grid with the size of $l\times w\times h$ . We can represent the inscribed object with the SDF value in each mesh grid, i.e., $\mathbf{m} = [f(\mathbf{x}_1),\dots,f(\mathbf{x}_i)\dots,f(\mathbf{x}_n)]^T$ , where $n = l\times w\times h$ . + +Shape prior. We aim to condense the nuanced differences present in class-specific 3D shapes into a low-dimensional latent space, where traversing along the manifold of the latent vector $\mathbf{s}$ corresponds to the exploration of feasible shapes of the object. We focus on the "car" class in this paper, where every instance in the class possesses an overall shape prior and exhibits minor variations in characteristics, such as the shape of the front face or the shape of the cabin. + +To obtain the shape prior, we first assemble a collection of 3D vehicle models from the Apollo-Car3D dataset [33], which consists of 79 models with exhibit significant variations in shapes and sizes (see the left of Fig. 3). However, using the Signed Distance Function (SDF) representation necessitates that objects be water-tight to avoid ambiguity between their interior and exterior. Some models contain intricate internal structures and are not initially water-tight. To obtain the water-tight surface of 3D models, we employ the SoftRas [21] algorithm to remove redundant vertices and faces inside the model. Finally, we convert the original triangular mesh representation of each 3D model into an SDF representation to obtain a set of models $\mathcal{M} = \{\mathbf{m}_i\}_{i=1}^{79}$ . + +Given the collection of representative 3D models, we perform the PCA method on $\mathcal{M}$ , and keep the top- $d$ ( $d = 5$ ) principle components to obtain a projection matrix $V \in \mathbb{R}^{d \times n}$ . Accordingly, we can use the projection matrix $V$ to render a 3D shape $\mathbf{m}$ from a given shape vector $\mathbf{s}$ . Besides, we can encode a 3D model $\mathbf{m}$ into a shape vector $\mathbf{s}$ by $\mathbf{s} = V^{\top}(\mathbf{m} - \bar{\mathbf{m}})$ and $\mathbf{m} = V\mathbf{s} + \bar{\mathbf{m}}$ , where $\bar{\mathbf{m}}$ indicates the mean vector of the collection $\mathcal{M}$ . In this work, we select a compact representation $d = 5$ . The right of Fig. 3 illustrates that interpolation along the latent space of $\mathbf{s}$ leads to smooth variation in shapes. + +Differentiable rendering. After recovering the 3D shape in a local coordinate system, we can put the model in the global coordinate system by adding its + +![](images/1bf734251b80c0387ca1a65e835a9dfe5f06b88d1287ef4481da445f07fbb1a0.jpg) +(a) RGB image + +![](images/9f55c4409305f9243c5efe18982a98d28fe326148c91b9593c0edba93d3c94eb.jpg) +(b) SAM mask + +![](images/312ae667a1f489734491527cf4776831a541092c1cfb7067d5d1306c6e610356.jpg) +(c) Target mask +Fig. 4: Generation of occlusion map $O$ for the mask alignment objective. + +![](images/e8306911fa36c3706b1d6ade31fa809ab948e3120d8e2b6caf55b1a4f89b052d.jpg) +(d) Occ. mask + +pose $\mathbf{p}$ . This gives us an unbounded SDF $\phi(\mathbf{x})$ of the 3D model, in which the values outside the pre-defined cubic are filled with negative infinity. Notably, the rendering process of shape is differentiable w.r.t. both $\mathbf{s}$ and $\mathbf{p}$ . + +# 3.4 Fit: Optimization over Pose and Shape + +Given the 2D instance mask and the point cloud scene, we randomly initialize pose $\mathbf{p}$ and shape $\mathbf{s}$ and optimize them using Eq. (1). Specifically, the optimization objective consists of three terms: the mask alignment term, the point cloud alignment term, and the ground alignment term. + +Mask alignment term. The mask alignment term measures the dissimilarity between the target instance segmentation mask $Y$ and the mask $Y_{proj}$ obtained by projecting the 3D model onto the image plane. To make the projection process differentiable, we follow [22, 37] to define the projection as $\pi(p_j) = 1 - \prod_{\mathbf{x} \in \mathcal{R}_{p_j}} \frac{1}{e^{\phi(\mathbf{x})\zeta + 1}}$ , where $p_j$ is the $j$ -th pixel in the image plane and $\pi(p_j) \in (0,1)$ is the mask value. A ray is cast through $p_j$ from the ego camera, and a set of points $\mathcal{R}p_j$ is collected by sampling along the ray. If the ray does not hit the surface of the 3D model, the SDF values of all sampled points $\mathbf{x} \in \mathcal{R}p_j$ will be negative, resulting in a mask value close to zero. Conversely, if the ray intersects with the 3D model, there must exist sampled points with $\phi(\mathbf{x}) > 0$ , leading to a mask value close to 1. A hyperparameter $\zeta$ controls the "hardness" of the mask values, where a larger $\zeta$ leads to harder values (close to either 0 or 1). Finally, we define the mask alignment term using the dice loss [34] as follows: + +$$ +E _ {\text {m a s k}} = 1 - \frac {2 | (Y _ {\text {p r o j}} * O) \bigcap Y |}{| Y _ {\text {p r o j}} * O | + | Y |}, \tag {2} +$$ + +where $O$ is a binary map indicating if the pixel is occluded by a nearer object. As shown in Fig. 4, to obtain the occlusion map, 3D LiDAR points are associated with each 2D mask based on projections, and the depth of each mask is computed as the median depth value of the associated LiDAR points. Masks are sorted by their depth values and then used to compute the occlusion maps. + +Point cloud alignment term. Solely optimizing the mask alignment term leads to inaccurate estimation due to the missing depth information. To avoid undesired local optimum, we introduce an additional point cloud alignment term. This term measures how well the 3D object's surface aligns with the point cloud inside the corresponding frustum. + +Intuitively, we can use the $\ell-1$ norm of the SDF $\phi(\mathcal{X}_f)$ to determine how well a 3D model fits a given frustum point cloud, denoted as $\mathcal{X}_f$ . However, this term alone might lead to sub-optimal solutions, as the points could be associated with the back face of the object. To overcome this issue, a second term is introduced that involves casting a ray for each point $\mathbf{x} \in \mathcal{X}_f$ starting from the ego camera and recording the coordinate $\mathbf{y}$ when the first time the ray intersects with the 3D model. This term penalizes the predicted 3D model if the distance between $\mathbf{x}$ and $\mathbf{y}$ is considerable. The point cloud alignment term can be written as follows: + +$$ +E _ {p c} = \| \phi (\mathcal {X} _ {f}) \| _ {1} + \frac {1}{| \hat {\mathcal {X}} _ {f} |} \sum_ {\mathbf {x} \in \hat {\mathcal {X}} _ {f}} \| \mathbf {x} - \mathbf {y} \| _ {2} ^ {2}. \tag {3} +$$ + +$\hat{\mathcal{X}}_f$ is a subset of $\subseteq \mathcal{X}_f$ and represents points whose corresponding ray intersects with the model surface. + +Ground alignment term. Another strong cue that can be leveraged to refine the 3D pose is the height of the ground since we can assume that vehicles run on the road surface. Specifically, we regress a ground plane function $z = g(x,y)$ using RANSAC [6] and then define a ground alignment objective as follows: + +$$ +E _ {\text {g r o u n d}} = \| \hat {z} - \frac {\hat {h}}{2} - g (\hat {x}, \hat {y}) \| _ {2} ^ {2}, \tag {4} +$$ + +where $(\hat{x},\hat{y},\hat{z})$ is the predicted object center, and $\hat{h}$ is the height of the object inferred from the predicted 3D shape. + +Optimization. The optimization objective is formulated as a weighted sum of $E_{mask}$ , $E_{pc}$ , and $E_{ground}$ . We initialize pose $\mathbf{p}$ with the median value of frustum points and shape $\mathbf{s}$ with mean shape $\bar{\mathbf{m}}$ . Based on the Adam optimizer [11] with a fixed learning rate of 0.1 and 150 iterations, a mini-batch comprising all instances (typically $< 100$ ) in a scene is employed for optimization, as opposed to individual objects. This approach yields a $5 \times$ acceleration with negligible accuracy drop. + +# 4 Experiments + +Previous auto-labelers can be divided into supervised ones and unsupervised ones. The proposed SLF is an unsupervised method. Therefore, we primarily compare it with existing unsupervised methods. Nonetheless, we also compare it with supervised methods and demonstrate that SLF exhibits superior transfer ability across datasets. In addition to bounding box accuracy, we also evaluate the quality of the predicted 3D shape. + +# 4.1 Comparison with Unsupervised Auto-labelers + +Dataset and metric. We conduct experiments on the commonly used KITTI [7] dataset. We follow the official training and validation split to get 3712 training + +Table 1: Comparison with unsupervised auto-labelers. + +
MethodAPBEV(IoU=0.5)|R40AP3D(IoU=0.5)|R40
Easy Mod.HardEasy Mod.Hard
VS3D [31]74.566.757.640.337.431.1
SDFLabel [44]77.859.8-62.342.2-
McCraith et al. [23]75.576.668.6---
FGR [39]93.777.870.591.377.870.6
SLFGTmask89.288.987.187.687.883.8
SLFpoint88.588.584.284.884.980.8
SLFbox89.289.384.787.988.283.9
+ +frames and 3768 validation frames. To align with prior works [15,39], the "Car" category is the focus of auto-labeling efforts. The Bird's Eye View (BEV) AP and the 3D AP are selected as evaluation metrics for a comprehensive comparison. Results are reported for all three levels of difficulty (Easy, Moderate, and Hard). The commonly used KITTI metrics are evaluated at a strict threshold of 0.7 IoU. As observed in [23,44], the 3D boxes in KITTI have varying spatial padding, making it challenging for human annotators to determine a precise size for a significant number of instances. Therefore, we follow the previous work [23,44] and set the IoU threshold as 0.5 and calculate AP by 40 recall positions. + +Direct evaluation on generated pseudo labels. A straightforward evaluation is to compare the quality of auto-labeled 3D bounding boxes in terms of the canonical metric, Average Precision (AP). However, the AP metric needs a confidence score for each bounding box for thresholding and integration, but FGR [39] and SLF do not output confidence scores. To remedy this issue, we simply define IoU between the predicted mask and the target mask in SLF as a confidence score. As for FGR, we first project a 3D bounding box to image plane to get a 2D box, and the confidence score is defined by IoU between the predicted 2D box and the ground truth 2D box. + +Tab. 1 compares the proposed SLF with existing unsupervised methods on KITTI validation set. $\mathrm{SLF}_{\mathrm{point}}$ uses 2D points as input, corresponding to the semi-automatic labeling scenario with humans in the loop. The point prompts are simulated by random sampling 3 points inside the ground truth masks [9]. $\mathrm{SLF}_{\mathrm{box}}$ indicates 2D bounding boxes as input, corresponding to the full-automatic labeling scenario in which 2D detectors generate the prompts. For all methods that accept 2D boxes as inputs, we use ground truth boxes for a fair comparison. We also provide a variant of our method in which the SAM masks are replaced with ground truth masks from [9], denoted as $\mathrm{SLF}_{\mathrm{GTmask}}$ . + +The last three rows in Tab. 1 show the results of SLF with different 2D prompts. $\mathrm{SLF}_{\mathrm{point}}$ and $\mathrm{SLF}_{\mathrm{box}}$ perform roughly the same in terms of $\mathrm{AP}_{BEV}$ , and $\mathrm{SLF}_{\mathrm{box}}$ shows superior $\mathrm{AP}_{3D}$ performance. To our surprise, $\mathrm{SLF}_{\mathrm{box}}$ is slightly better than $\mathrm{SLF}_{\mathrm{GTmask}}$ in terms of both $\mathrm{AP}_{3D}$ and $\mathrm{AP}_{BEV}$ except on hard samples, which means the 2D mask generated by SAM from box prompts are even + +Table 2: Detector performance trained with pseudo labels. + +
DetectorsLabels\(AP_{BEV}(IoU=0.5)\) \(|_{R40}\)\(AP_{3D}(IoU=0.5)\) \(|_{R40}\)
EasyMod.HardEasyMod.Hard
PointPillar [14]GT95.794.393.295.793.691.4
FGR96.590.287.194.989.484.8
SLF95.394.191.195.292.188.8
SECOND [42]GT95.994.893.995.894.691.9
FGR95.488.986.295.388.685.7
SLF95.694.291.795.592.489.3
Part-\(A^2\) [32]GT96.094.393.996.094.291.9
FGR96.192.187.596.090.085.1
SLF96.094.591.895.993.989.6
Voxel R-CNN [3]GT98.494.994.598.394.894.4
FGR98.790.988.298.690.788.0
SLF96.295.192.496.194.791.9
+ +better than those annotated by humans. This suggests that SLF benefits from more accurate input prompts, and more accurate instance masks. + +We compare $\mathrm{SLF}_{\mathrm{box}}$ with other unsupervised methods as they accept the same input prompts. It can be observed that SLF outperforms FGR and SDFLabel by a significant margin on the Moderate and Hard samples. However, SLF performs worse than FGR on the Easy samples, and we assume that easy cases in KITTI have much denser sub-points, which are beneficial to the segmentation and box estimation stage of FGR. Moreover, FGR consumes ground-truth amodal bounding box, while in KITTI these boxes are obtained by projecting 3D ground-truth boxes to 2D images, which may leak the oracle occlusion and size information of partially visible objects. SDFLabel uses a similar 3D shape representation to SLF but relies on a synthetic dataset for training, which may limit its performance on real-world data. In contrast, SLF does not require synthetic training data, and generalizes well to real-world scenarios. + +Detector performance trained with pseudo labels. Following previous work [15,24,39], to further evaluate the quality of generated pseudo labels, we use them as training samples to train off-the-shelf detectors and compare the results of different detectors. We mainly compare with the previous best-performing FGR auto-labeler and train 4 different 3D detectors. Specifically, for all detectors, we use the OpenPCDet [36] with default hyperparameters. In Tab. 2, we compare the results of detectors trained on KITTI training set with ground truth (GT) labels, FGR pseudo labels, and SLF pseudo labels, respectively. If not specified, SLF refers to $\mathrm{SLF}_{\mathrm{box}}$ in the remaining sections. + +Tab. 2 shows SLF pseudo labels significantly improve upon FGR pseudo labels on all detectors on Moderate and Hard samples. Regarding Easy samples, SLF pseudo labels are on par with FGR pseudo labels on all detectors except Voxel R-CNN [3]. When comparing specific detectors, the one trained with SLF + +labels performs almost as well as the one trained with ground truth labels on both Easy and Moderate samples, and slightly worse on Hard samples, which shows the effectiveness of our method. + +# 4.2 Cross Dataset Generalization + +Apart from KITTI, we conduct experiments on more challenging nuScenes [1]. We transform the validation set of nuScenes to KITTI format, resulting in 6019 frames with corresponding front view images, point cloud and 3D bounding box annotations. On KITTI, we compare the pseudo label with ground truths and report $\mathrm{AP}_{3D}$ , and on nuScenes, we report the official metrics including mAP, NDS, mATE, mASE and mAOE. We compare with the best-performing unsupervised auto-labeler FGR, and the supervised auto-labeler MTrans [15]. Both FGR and Mtrans require 2D amodal boxes as prompts, so we project 3d ground truth boxes into 2D as their input. Our SLF supports a more practical setting that only requires regular 2D modal boxes as prompts. Thus we can generate them with an off-the-shelf detector HTC [2] pre-trained on nuImages [1]. + +Table 3: Cross dataset evaluation compared with state-of-the-art unsupervised auto-labeler FGR and supervised auto-labeler MTrans. Results in gray are less informative because the NDS and error metrics are computed only on recalled samples, while FGR fails on recalling a majority of input samples (low mAP). $\mathrm{mAOE}^{\dagger}$ means to change the periodicity of orientation to $[0,\pi]$ . + +
MethodTraining DataAP3D@KITTInuScenes Metrics
EasyMod.HardmAPNDSmATEmASEmAOEmAOE†
FGR [39]-91.377.870.611.640.4.152.188.010.010
MTrans [15]KITTI99.699.697.145.949.4.358.349.702.225
SLF-87.988.283.957.255.2.372.192.836.113
+ +In Tab. 3, on KITTI, MTrans shows a nearly perfect performance, this is not surprising since MTrans is a learnable neural network trained on KITTI, it easily learns the specific annotation patterns in training labels. Despite the high accuracy, we argue that learnable auto-labelers like MTrans are prone to overfit the biased pattern in the training data and generalize poorly across datasets. When we apply MTrans to the challenging nuScenes dataset, its performance dramatically decreases. SLF performs better in terms of mAP and NDS. Regarding the detailed error terms, we find SLF outperforms MTrans on size estimation (mASE), and is on par with MTrans on location estimation (mATE). Regarding orientation estimation, SLF underperforms MTrans by 0.134. However, we find a typical orientation error is the model mistakes the front end of the car for the rear end, which is usually caused when the LiDAR points are rather sparse. To explore how good the orientation predictions are when the $\pi$ + +phase lag is disentangled, we change the periodicity of orientation to $[0,\pi ]$ and re-measure the orientation error, denoted as $\mathrm{mAOE}^{\dagger}$ . SLF outperforms MTrans on $\mathrm{mAOE}^{\dagger}$ , suggesting that SLF predicts more accurate orientation, the heading direction aligning better with ground truths. When applying FGR on the nuScences dataset, it fails on a majority of samples, resulting in a low mAP of $11.6\%$ . Consequently, the NDS score and error terms are less informative since they are computed only on successfully labeled samples, typically easy samples. + +![](images/d19ca68f87c4ad4daa7b11dc6d06115a0a5e400f7ec0cb62ee4e5aee6daf1a74.jpg) +Fig.5: Qualitative results of estimated shapes. From left to right: projected mask of the initial 3D model; projected mask of the optimized 3D model; mean shape and surrounding point cloud; optimized 3D shape and surrounding point cloud. + +![](images/20522f0553e92b0d5d8ff3489ac7cd44a50b017711bcb56b61c898726f4f7988.jpg) +Fig.6: Comparison of occupancy annotation. From left to right: reference instance image; occupancy labels from OFN [25]; occupancy labels from OpenOccupancy [38]; occupancy labels generated by our SLF. + +![](images/2cc134221cba5ecd9ac29cb81169f87978f12b8e02370433672c154fabaee569.jpg) + +# 4.3 Voxel Occupancy Annotation + +Since benchmarks [1,7] (i.e., KITTI and nuScences) do not support quantitative evaluation on 3D shapes, we show qualitative results in Fig. 5 to better show the effect of SLF on shape labeling. Specifically, we replace the estimated shape with a mean shape and show comparisons. A clear improvement can be observed by comparing the initial shape and the final optimized shape, both in terms of mask alignment and LiDAR point alignment. Besides, we observe that SLF can serve as a good auto-labeler for occupancy annotation of dynamic objects such as vehicles. We compare two versions of occupancy labels for nuScenes, OpenOccupancy [38] and OFN [25]. Considering that existing occupancy labels are still of low quality, quantitative evaluation may be misleading. Therefore we show qualitative visualizations in Fig. 6, and we observe that OpenOccupancy and OFN labels fail to depict the inner occupancy of the cars. This is mainly attributed to that they label the occupancy by voxelizing the LiDAR points, and therefore self-occlusion can not be avoided. In contrast, SLF can predict dense occupancy more accurately, with more detailed shapes. + +# 4.4 Ablation Study + +In this section, we conduct comprehensive ablation studies on the KITTI validation set and report the $\mathrm{AP}_{3D}(\mathrm{IoU} = 0.5)|_{R40}$ results. Except in Tab. 6, we input 2D points to SLF, other results are obtained by using 2D bounding boxes as prompt. Effect of energy terms. In Tab. 4, we observe that introducing the mask alignment term significantly improves pseudo-label quality rather than solely applying the point cloud alignment term, and adding the ground alignment term further improves performance. However, as we mentioned above, using mask alignment alone can easily be trapped in sub-optimal solutions, thus it can not obtain valid results (Easy: 6.9, Mod.: 5.9, Hard: 5.6). + +Table 4: Ablation on energy terms. + +
ObjectiveEasyMod.Hard
Epc70.564.862.9
Emask+Epc87.388.083.7
Emask+Epc+Eground87.988.283.9
+ +Table 5: Ablation on 3D models. + +
# modelsEasyMod.Hard
574.770.164.0
3084.283.677.3
7987.988.283.9
+ +Effect of the number of CAD models. In Tab. 5, we randomly sample 5 and 30 CAD models out of the total 79 samples to investigate the influence of the number of CAD models used for PCA. From the results, we can conclude that better result comes with more models. + +Effect of the PCA dimensions. In Tab. 7, we show how different PCA dimensions affect results. Performance improves when we increase the dimension from 3 to 5. More dimensions like 10 do not further improve. This is possibly attributed to that a higher dimension of shape representation may complicate the optimization process, which means it may need more steps to converge. + +Effect of the number of points. In Tab. 6, we analyze the effect of the number of points input to SAM [12]. When we increase the points from 3 to 8, the improvement of pseudo-label quality is minor. However, we observe a significant decrease in the easy cases of KITTI when using one point as the prompt, which is reasonable as easy cases often have large masks and a single point can not provide sufficient context for SAM to generate high-quality results. + +# 4.5 Further Analysis + +Robustness of LiDAR Beam. Point clouds in KITTI [7] are collected by 64 beams LiDAR, we downsample point cloud to 32, 16 and 8 beams following [40]. $\mathrm{AP}_{3D}$ results on KITTI validation set in Tab. 8 show SLF can still get decent results on 32 and 16 beams LiDAR, and the performances downgrade rapidly on Mod. and Hard cases, since these cases usually far away from ego vehicle which exacerbating the sparsity of point clouds. + +Table 6: No. points. + +
# PointsEasyMod.Hard
384.884.980.8
178.181.178.8
584.885.081.0
885.085.280.9
+ +Table 7: No. dim. + +
Dim.EasyMod.Hard
587.988.283.9
387.187.883.6
886.987.184.1
1084.885.481.1
+ +Table 8: No. beam. + +
BeamEasyMod.Hard
6487.988.283.9
3281.377.577.3
1680.776.370.0
875.160.056.7
+ +Table 9: Dilation on instance mask. + +
KernelEasyMod.Hard
FGR(no dil.)91.377.870.6
SLF(no dil.)87.988.283.9
5x586.385.681.2
9x983.980.576.6
+ +Table 10: Erosion on instance mask. + +
KernelEasyMod.Hard
FGR(noERO.)91.377.870.6
SLF(noERO.)87.988.283.9
5x585.283.979.5
9x982.178.272.1
+ +Robustness of Mask Quality. We simulate low-quality masks by applying erosion and dilation operations on SAM [12] masks and report $\mathrm{AP}_{3D}$ results of KITTI validation set in Tab. 9 and Tab. 10. We observe even with severely corrupted masks, SLF still outperforms the existing unsupervised auto-annotation methods in Mod. and Hard cases, which show the robustness of SLF. + +# 5 Limitation + +The proposed SLF uses a strong shape prior learned by PCA from given 3D CAD models. The shape prior is category-specific, which means we need to prepare a collection of 3D models when labeling a specific semantic class. + +# 6 Conclusion + +This paper proposes a Segment, Lift, and Fit paradigm for automatic 3D shape annotation of objects from the 2D point or box prompts. We first segment high-quality instance masks from the prompts using the SAM, and then lift 2D masks to 3D forms, where gradient descent is used to adjust their poses and shapes until the projections fits the masks and the surfaces conform to surrounding LiDAR points. Extensive experimental results on multiple datasets for 3D detection demonstrate the effectiveness of our SLF method. Moreover, the SLF auto-labeler achieves promising results in detailed shape predictions, which can be used for occupancy annotation. In the future, we will investigate the generalization abilities by applying our SLF to more categories. + +# References + +1. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. arXiv preprint arXiv:1903.11027 (2019) +2. Chen, K., Pang, J., Wang, J., Xiong, Y., Li, X., Sun, S., Feng, W., Liu, Z., Shi, J., Ouyang, W., et al.: Hybrid task cascade for instance segmentation. In: CVPR. pp. 4974-4983 (2019) +3. Deng, J., Shi, S., Li, P., Zhou, W., Zhang, Y., Li, H.: Voxel r-cnn: Towards high performance voxel-based 3d object detection. arXiv:2012.15712 (2020) +4. Engelmann, F., Stickler, J., Leibe, B.: Joint object pose estimation and shape reconstruction in urban street scenes using 3d shape priors. In: Pattern Recognition: 38th German Conference, GCPR 2016, Hannover, Germany, September 12-15, 2016, Proceedings 38. pp. 219-230. Springer (2016) +5. Engelmann, F., Stuckler, J., Leibe, B.: Samp: shape and motion priors for 4d vehicle reconstruction. In: 2017 IEEE Winter Conference on Applications of Computer Vision (WACV). pp. 400-408. IEEE (2017) +6. Fischler, M.A., Bolles, R.C.: Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM 24(6), 381-395 (1981) +7. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: CVPR (2012) +8. Guo, J., Liu, J., Xu, D.: 3d-pruning: A model compression framework for efficient 3d action recognition. IEEE Transactions on Circuits and Systems for Video Technology 32(12), 8717-8729 (2022) +9. Heylen, J., De Wolf, M., Dawagne, B., Proesmans, M., Van Gool, L., Abbeloos, W., Abdelkawy, H., Reino, D.O.: Monocinis: Camera independent monocular 3d object detection using instance segmentation. In: ICCV. pp. 923-934 (2021) +0. Huang, Y., Zheng, W., Zhang, Y., Zhou, J., Lu, J.: Tri-perspective view for vision-based 3d semantic occupancy prediction. arXiv preprint arXiv:2302.07817 (2023) +1. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. In: ICLR (2015) +2. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. arXiv:2304.02643 (2023) +3. Kundu, A., Li, Y., Rehg, J.M.: 3d-rcnn: Instance-level 3d object reconstruction via render-and-compare. In: CVPR (2018). https://doi.org/10.1109/CVPR.2018.00375, https://doi.org/10.1109/CVPR.2018.00375 +4. Lang, A.H., Vora, S., Caesar, H., Zhou, L., Yang, J., Beijbom, O.: Pointpillars: Fast encoders for object detection from point clouds. In: CVPR. pp. 12697-12705 (2019) +5. Liu, C., Qian, X., Huang, B., Qi, X., Lam, E., Tan, S.C., Wong, N.: Multimodal transformer for automatic 3d annotation and object detection. In: ECCV (2022) +6. Liu, J., Guo, J., Xu, D.: Aptsnet: Toward adaptive point sampling for efficient 3d action recognition. IEEE Transactions on Image Processing 31, 5287-5302 (2022) +7. Liu, J., Guo, J., Xu, D.: Geometrhythmotion-transformer: An end-to-end framework for 3d action recognition. IEEE Transactions on Multimedia (2022) +8. Liu, J., He, T., Yang, H., Su, R., Tian, J., Wu, J., Guo, H., Xu, K., Ouyang, W.: 3d-queryis: A query-based framework for 3d instance segmentation. arXiv preprint arXiv:2211.09375 (2022) + +19. Liu, J., Li, J., Wang, K., Guo, H., Yang, J., Peng, J., Xu, K., Liu, X., Guo, J.: Lta-pcs: Learnable task-agnostic point cloud sampling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 28035-28045 (June 2024) +20. Liu, J., Xu, D.: Geometrhythmotion-net: A strong two-stream baseline for 3d action recognition. IEEE Transactions on Circuits and Systems for Video Technology 31(12), 4711-4721 (2021) +21. Liu, S., Li, T., Chen, W., Li, H.: Soft rasterizer: A differentiable renderer for image-based 3d reasoning. ICCV (Oct 2019) +22. Liu, Z., Zhou, D., Lu, F., Fang, J., Zhang, L.: Autoshape: Real-time shape-aware monocular 3d object detection. In: ICCV. pp. 15641-15650 (2021) +23. McCraith, R., Insafutdinov, E., Neumann, L., Vedaldi, A.: Lifting 2d object locations to 3d by discounting lidar outliers across objects and views (2022) +24. Meng, Q., Wang, W., Zhou, T., Shen, J., Van Gool, L., Dai, D.: Weakly supervised 3d object detection from lidar point cloud. In: ECCV (2020) +25. Ming Fang, Z.L.: Occupancy dataset for nuscenes. https://github.com/FANG-MING/occupancy-for-nuscenes (2023) +26. Osher, S., Fedkiw, R.: Signed Distance Functions. Springer New York (2003). https://doi.org/10.1007/0-387-22746-6_2, https://doi.org/10.1007/0-387-22746-6_2 +27. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: CVPR (June 2019) +28. Peng, L., Liu, F., Yu, Z., Yan, S., Deng, D., Yang, Z., Liu, H., Cai, D.: Lidar point cloud guided monocular 3d object detection. In: European Conference on Computer Vision. pp. 123-139. Springer (2022) +29. Prisacariu, V.A., Segal, A.V., Reid, I.: Simultaneous monocular 2d segmentation, 3d pose recovery and 3d reconstruction. In: Lee, K.M., Matsushita, Y., Rehg, J.M., Hu, Z. (eds.) ACCV. pp. 593-606. Springer Berlin Heidelberg, Berlin, Heidelberg (2013) +30. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. arXiv preprint arXiv:1612.00593 (2016) +31. Qin, Z., Wang, J., Lu, Y.: Weakly supervised 3d object detection from point clouds. ACMMM (2020) +32. Shi, S., Wang, Z., Shi, J., Wang, X., Li, H.: From points to parts: 3d object detection from point cloud with part-aware and part-aggregation network. IEEE transactions on pattern analysis and machine intelligence 43(8), 2647-2664 (2020) +33. Song, X., Wang, P., Zhou, D., Zhu, R., Guan, C., Dai, Y., Su, H., Li, H., Yang, R.: Apollocar3d: A large 3d car instance understanding benchmark for autonomous driving. In: CVPR. pp. 5452-5462 (2019) +34. Sudre, C.H., Li, W., Vercauteren, T., Ourselin, S., Jorge Cardoso, M.: Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations. In: Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: Third International Workshop, DLMIA 2017, and 7th International Workshop, ML-CDS 2017, Held in Conjunction with MICCAI 2017, Quebec City, QC, Canada, September 14, Proceedings 3. pp. 240-248. Springer (2017) +35. Tang, Y.S., Lee, G.H.: Transferable semi-supervised 3d object detection from rgb-d data. In: ICCV. pp. 1931-1940 (2019) +36. Team, O.D.: Openpcdet: An open-source toolbox for 3d object detection from point clouds. https://github.com/open-mmlab/OpenPCDet (2020) + +37. Wang, R., Yang, N., Stueckler, J., Cremers, D.: Directshape: Photometric alignment of shape priors for visual vehicle pose and shape estimation (2020) +38. Wang, X., Zhu, Z., Xu, W., Zhang, Y., Wei, Y., Chi, X., Ye, Y., Du, D., Lu, J., Wang, X.: Openoccupancy: A large scale benchmark for surrounding semantic occupancy perception. arXiv preprint arXiv:2303.03991 (2023) +39. Wei, Y., Su, S., Lu, J., Zhou, J.: Fgr: Frustum-aware geometric reasoning for weakly supervised 3d vehicle detection (2021) +40. Wei, Y., Wei, Z., Rao, Y., Li, J., Zhou, J., Lu, J.: Lidar distillation: Bridging the beam-induced domain gap for 3d object detection. In: ECCV. pp. 179-195. Springer (2022) +41. Wei, Y., Zhao, L., Zheng, W., Zhu, Z., Zhou, J., Lu, J.: Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving. arXiv preprint arXiv:2303.09551 (2023) +42. Yan, Y., Mao, Y., Li, B.: Second: Sparsely embedded convolutional detection. Sensors 18(10), 3337 (2018) +43. Yang, H., He, T., Liu, J., Chen, H., Wu, B., Lin, B., He, X., Ouyang, W.: Gdmae: generative decoder for mae pre-training on lidar point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9403-9414 (2023) +44. Zakharov, S., Kehl, W., Bhargava, A., Gaidon, A.: Autolabeling 3d objects with differentiable rendering of sdf shape priors. In: CVPR (2020) \ No newline at end of file diff --git a/2024/Segment, Lift and Fit_ Automatic 3D Shape Labeling from 2D Prompts/images.zip b/2024/Segment, Lift and Fit_ Automatic 3D Shape Labeling from 2D Prompts/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..19b293ca29e338a6fd2cd69b37385349d7c6171d --- /dev/null +++ b/2024/Segment, Lift and Fit_ Automatic 3D Shape Labeling from 2D Prompts/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d9287ff400f2355db5d088d87f8692e3ab585da2a62cf0cf425c985ff2b3eab8 +size 478201 diff --git a/2024/Segment, Lift and Fit_ Automatic 3D Shape Labeling from 2D Prompts/layout.json b/2024/Segment, Lift and Fit_ Automatic 3D Shape Labeling from 2D Prompts/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..22c8f60fef4bce94d4a508c8c477ddf0b8d7239c --- /dev/null +++ b/2024/Segment, Lift and Fit_ Automatic 3D Shape Labeling from 2D Prompts/layout.json @@ -0,0 +1,9835 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 153, + 111, + 462, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 111, + 462, + 148 + ], + "spans": [ + { + "bbox": [ + 153, + 111, + 462, + 148 + ], + "type": "text", + "content": "Segment, Lift and Fit: Automatic 3D Shape Labeling from 2D Prompts" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "spans": [ + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "text", + "content": "Jianhao Li" + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "text", + "content": ", Tianyu Sun" + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "text", + "content": ", Zhongdao Wang" + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "inline_equation", + "content": "^{3\\dagger}" + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "text", + "content": ", Enze Xie" + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "text", + "content": ", Bailan Feng" + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "text", + "content": ", Hongbo Zhang" + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "text", + "content": ", Ze Yuan" + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "text", + "content": ", Ke Xu" + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "text", + "content": ", Jiaheng Liu" + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "inline_equation", + "content": "^{4\\dagger}" + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "text", + "content": ", and Ping Luo" + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "inline_equation", + "content": "^{5}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 229, + 201, + 383, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 229, + 201, + 383, + 213 + ], + "spans": [ + { + "bbox": [ + 229, + 201, + 383, + 213 + ], + "type": "text", + "content": "1 Beihang University, Beijing, China" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 229, + 213, + 383, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 229, + 213, + 383, + 224 + ], + "spans": [ + { + "bbox": [ + 229, + 213, + 383, + 224 + ], + "type": "text", + "content": "2 Tsinghua University, Beijing, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 251, + 224, + 361, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 224, + 361, + 235 + ], + "spans": [ + { + "bbox": [ + 251, + 224, + 361, + 235 + ], + "type": "text", + "content": "3 Noah's Ark Lab, Huawei" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 263, + 235, + 350, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 263, + 235, + 350, + 246 + ], + "spans": [ + { + "bbox": [ + 263, + 235, + 350, + 246 + ], + "type": "text", + "content": "4 Nanjing University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 242, + 246, + 372, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 246, + 372, + 257 + ], + "spans": [ + { + "bbox": [ + 242, + 246, + 372, + 257 + ], + "type": "text", + "content": "5 The University of Hong Kong" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 209, + 258, + 404, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 258, + 404, + 268 + ], + "spans": [ + { + "bbox": [ + 209, + 258, + 404, + 268 + ], + "type": "text", + "content": "{lijianhao, yuanze1024, kexu}@buaa.edu.cn" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 154, + 269, + 459, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 269, + 459, + 279 + ], + "spans": [ + { + "bbox": [ + 154, + 269, + 459, + 279 + ], + "type": "text", + "content": "{wangzhongdao, Johnny_ez, fengbailan, zhanghongbo888}@huawei.com" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 135, + 280, + 471, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 280, + 471, + 290 + ], + "spans": [ + { + "bbox": [ + 135, + 280, + 471, + 290 + ], + "type": "text", + "content": "sty21@mails.tsinghua.edu.cn, buaaljiaheng@gmail.com, pluo.lhi@gmail.com" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 159, + 311, + 453, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 311, + 453, + 563 + ], + "spans": [ + { + "bbox": [ + 159, + 311, + 453, + 563 + ], + "type": "text", + "content": "Abstract. This paper proposes an algorithm for automatically labeling 3D objects from 2D point or box prompts, especially focusing on applications in autonomous driving. Unlike previous arts, our auto-labeler predicts 3D shapes instead of bounding boxes and does not require training on a specific dataset. We propose a Segment, Lift, and Fit (SLF) paradigm to achieve this goal. Firstly, we segment high-quality instance masks from the prompts using the Segment Anything Model (SAM) and transform the remaining problem into predicting 3D shapes from given 2D masks. Due to the ill-posed nature of this problem, it presents a significant challenge as multiple 3D shapes can project into an identical mask. To tackle this issue, we then lift 2D masks to 3D forms and employ gradient descent to adjust their poses and shapes until the projections fit the masks and the surfaces conform to surrounding LiDAR points. Notably, since we do not train on a specific dataset, the SLF auto-labeler does not overfit to biased annotation patterns in the training set as other methods do. Thus, the generalization ability across different datasets improves. Experimental results on the KITTI dataset demonstrate that the SLF auto-labeler produces high-quality bounding box annotations, achieving an AP@0.5 IoU of nearly " + }, + { + "bbox": [ + 159, + 311, + 453, + 563 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 159, + 311, + 453, + 563 + ], + "type": "text", + "content": ". Detectors trained with the generated pseudo-labels perform nearly as well as those trained with actual ground-truth annotations. Furthermore, the SLF auto-labeler shows promising results in detailed shape predictions, providing a potential alternative for the occupancy annotation of dynamic objects." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 160, + 574, + 438, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 574, + 438, + 585 + ], + "spans": [ + { + "bbox": [ + 160, + 574, + 438, + 585 + ], + "type": "text", + "content": "Keywords: 3D Auto-labelers " + }, + { + "bbox": [ + 160, + 574, + 438, + 585 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 574, + 438, + 585 + ], + "type": "text", + "content": " Shape Optimization " + }, + { + "bbox": [ + 160, + 574, + 438, + 585 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 574, + 438, + 585 + ], + "type": "text", + "content": " Training Free" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 601, + 230, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 230, + 615 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 230, + 615 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 624, + 482, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 624, + 482, + 648 + ], + "spans": [ + { + "bbox": [ + 130, + 624, + 482, + 648 + ], + "type": "text", + "content": "In modern robotics and autonomous driving, a significant amount of labeled data is required for supervised learning to understand the 3D scene [8,16,17,19," + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 653, + 375, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 653, + 375, + 666 + ], + "spans": [ + { + "bbox": [ + 141, + 653, + 375, + 666 + ], + "type": "text", + "content": "† Corresponding authors: Zhongdao Wang; Jiaheng Liu." + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 116, + 280, + 205 + ], + "blocks": [ + { + "bbox": [ + 133, + 116, + 280, + 205 + ], + "lines": [ + { + "bbox": [ + 133, + 116, + 280, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 280, + 205 + ], + "type": "image", + "image_path": "5e3f82b4b5a52d18b0696738ee5adf02a82ff4230af217d24bf50743cb51ac8d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 164, + 209, + 259, + 220 + ], + "lines": [ + { + "bbox": [ + 164, + 209, + 259, + 220 + ], + "spans": [ + { + "bbox": [ + 164, + 209, + 259, + 220 + ], + "type": "text", + "content": "(a) Human Annotation" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 293, + 114, + 480, + 209 + ], + "blocks": [ + { + "bbox": [ + 293, + 114, + 480, + 209 + ], + "lines": [ + { + "bbox": [ + 293, + 114, + 480, + 209 + ], + "spans": [ + { + "bbox": [ + 293, + 114, + 480, + 209 + ], + "type": "image", + "image_path": "30ab53df3ef06079ff2931a42e8d06cb898243baa547769543e4a9482ab4c689.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 316, + 209, + 477, + 220 + ], + "lines": [ + { + "bbox": [ + 316, + 209, + 477, + 220 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 477, + 220 + ], + "type": "text", + "content": "(b) Auto-labeler Annotation using SLF" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 229, + 480, + 241 + ], + "lines": [ + { + "bbox": [ + 132, + 229, + 480, + 241 + ], + "spans": [ + { + "bbox": [ + 132, + 229, + 480, + 241 + ], + "type": "text", + "content": "Fig. 1: Comparison between human annotation and auto-labeler annotation using SLF." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 266, + 479, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 266, + 479, + 339 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 479, + 339 + ], + "type": "text", + "content": "[20], emphasizing dynamic objects such as vehicles and pedestrians. Therefore, there is a pressing need to improve the degree of automation for 3D labeling. The preferred labeling format is typically the 3D bounding box, which can be represented compactly with as few as 7-9 degrees of freedom. However, manually labeling a large number of 3D boxes remains a laborious and costly task, posing challenges for scaling up 3D object detectors as shown in Fig. 1(a)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 340, + 480, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 480, + 436 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 480, + 436 + ], + "type": "text", + "content": "Meanwhile, as 3D perception models evolve, there is an increasing demand for more precise labeling granularity. While bounding boxes offer a compact representation, they are not optimal for 3D objects and are being replaced by finer representations. For instance, voxel occupancy [10,38,41] has emerged as a dominant format for 3D scenes in autonomous driving, as it unifies the representation of static road elements and dynamic objects. However, finer-grained labels further complicate the annotation pipeline, exacerbating the annotation efficiency problem and presenting a significant obstacle to scaling up 3D models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 449, + 480, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 480, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 480, + 604 + ], + "type": "text", + "content": "In this paper, we propose a novel approach for 3D auto-labeling. The proposed auto-labeler follows a Segment, Lift, and Fit (SLF) paradigm as shown in Fig. 1(b). Specifically, SLF utilizes the 2D points and bounding boxes as prompts and outputs the estimated 3D shapes and poses of target objects, thus supporting semi- or fully automatic labeling depending on whether the prompts come from humans or machines. To accomplish this, we leverage the Segment Anything Model (SAM) [12], a foundational vision model, to generate high-quality instance masks from the input prompts. Subsequently, we lift 2D instance masks into 3D forms with Signed Distance Functions (SDFs) [26], and then iteratively optimize the shapes and poses until their projections fit the mask and its surface conforms to the surrounding LiDAR points. Notably, the proposed SLF auto-labeler does not require supervised training on a specific dataset, and is therefore not susceptible to overfitting biased annotation patterns in the training dataset." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": "Experimental results show that SLF can generate high-quality 3D labels. As our method outputs detailed shapes of objects, while existing benchmarks [1,7] only provide bounding box annotations, for a fair comparison with existing auto-labelers, we convert shapes to boxes and assess the label quality regarding box accuracy. When compared among unsupervised auto-labelers [23,39,44] that do" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 210, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 210, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 210, + 100 + ], + "type": "text", + "content": "J. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "not train with 3D labels as ours, our method outperforms the best-performing FGR [39] auto-labeler in both settings of direct evaluation on pseudo labels and training object detectors with generated pseudo labels. Moreover, because SLF does not learn biased annotation patterns from the training data, it significantly outperforms state of art supervised model MTrans [15] on \"unseen\" datasets, demonstrating its superior generalization capacity. We also show qualitatively that the predicted shapes align well with the actual geometry of the target objects, underpinning a new possibility for voxel occupancy annotation [38,41]. Our contribution can be summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 230, + 482, + 361 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 138, + 230, + 482, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 230, + 482, + 290 + ], + "spans": [ + { + "bbox": [ + 138, + 230, + 482, + 290 + ], + "type": "text", + "content": "- We propose SLF, a 3D auto-labeler that improves existing methods by predicting detailed shapes and poses rather than just bounding boxes. The shape and pose are directly optimized using gradient descent without requiring a neural network trained on a small portion of labels, which enables a \"cold-start\" scenario for auto-labeling." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 289, + 482, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 289, + 482, + 326 + ], + "spans": [ + { + "bbox": [ + 138, + 289, + 482, + 326 + ], + "type": "text", + "content": "- We demonstrate that SLF generates superior labels to other unsupervised auto-labelers. Additionally, it shows good generalization ability across different datasets, surpassing supervised auto-labelers." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 325, + 482, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 325, + 482, + 361 + ], + "spans": [ + { + "bbox": [ + 138, + 325, + 482, + 361 + ], + "type": "text", + "content": "- The predicted shapes align well with the actual geometries of the objects, suggesting that SLF can serve as a viable alternative solution for annotating fine-grained occupancy of dynamic objects." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 378, + 243, + 391 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 378, + 243, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 378, + 243, + 391 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 402, + 483, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 483, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 483, + 668 + ], + "type": "text", + "content": "3D Auto-labelers. Compared to supervised 3D object detectors [14,18,42,43], auto-labelers require fewer or no training labels. Supervised auto-labelers learn to annotate from limited samples with ground-truth 3D labels, often in the hundreds. For instance, WS3D [24] generates cylindrical object proposals on bird's eye view (BEV) maps using center clicks in BEV and fine-tunes a 3D detector pre-trained on a few human-annotated data. MTrans [15] employs a multi-task design of segmentation, point generation, and box regression to train a 3D detector that takes ground truth 2D boxes and LiDAR points as input. Unsupervised auto-labelers do not require 3D labels for training. CC [35] uses PointNet [30] to segment and regress 3D boxes from frustum points. SDFLabel [44] trains a DeepSDF [27] network supervised with image patch and CAD models of cars on a synthetic dataset, then uses RANSAC [6] to fit the predicted shape to frustum sub-points. VS3D [31] employs an unsupervised UPM module to generate 3D proposals and filter them using a pre-trained image model. Our SLF, is also an unsupervised approach that does not require training with 3D labels. SLF achieves finer annotations with weaker prompts compared with a closely related unsupervised and non-learning approach, FGR [39]. FGR requires amodal boxes as inputs, in which oracle occlusion information has already been considered. LPCG [28] finds that concerning label accuracy, the 3D location part in the label is preferred compared to other parts of labels in monocular 3D object detection and proposed training-free and high-cost label pipeline, which is similar to FGR and MTrans respectively. In contrast, SLF takes normal 2D" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 155, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 155, + 91, + 448, + 103 + ], + "type": "text", + "content": "Segment, Lift and Fit: Automatic 3D Shape Labeling from 2D Prompts" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "content": "modal bounding boxes as input, which an off-the-shelf 2D detector can easily obtain. Furthermore, SLF can produce 3D object shapes rather than just bounding boxes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 482, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 482, + 428 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 482, + 428 + ], + "type": "text", + "content": "Shape Prior. To reconstruct the 3D geometry for instances from a specific semantic class, e.g., cars or pedestrians, the shape prior of the given class is often utilized to ease the estimation. For instance, McCraith et al. [23] uses a fixed car shape as the most straightforward shape prior. Autoshape [22] develops a shape-labeler for the car class, but it requires ground truth 3D bounding box and amodal instance mask which contains oracle occlusion information as input and it uses part-based deformable representation as shape prior which needs to label the key points and different part of CAD model. Another usage of shape prior is to encode it into a compact low-dimensional space with PCA [4,5,29,37] or neural network [13,44]. SDFLabel [44] pretrains a DeepSDF [27] model to predict the hidden representation of shape conditioned on cropped image patches. 3D-RCNN [13] regresses poses and shape parameters in a single forward pass. However, in 3D-RCNN shape representation is not differentiable, and the finite differences that are used for gradient approximation cause ambiguity during training. Upon the compact PCA representation, Engelmann et al. [4] and DirectShape [37] optimize the shape and pose concurrently with the stereo reconstruction of the object to refine the 3D detection result. SAMP [5] takes sequence of stereo depth maps and 3D trajectory to generate target's shape and refine its trajectory. Prisacariu et al. [29] requires sequence of instance masks and results of 3D tracker to optimize 3D model. Similar to previous PCA based method, our method adopts the PCA shape prior in a differentiable optimization process, however, SLF is training-free and only requires a single raw point cloud and RGB image with 2D prompts, which is more practical." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 446, + 212, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 212, + 459 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 212, + 459 + ], + "type": "text", + "content": "3 Approach" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 471, + 482, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 471, + 482, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 471, + 482, + 579 + ], + "type": "text", + "content": "As shown in Fig. 2(a), we aim to recover the 3D shape and pose of an interested object by asking a labeler to click one or more points or draw a box on the image as the prompt for the target object. A labeler could be a human in semiautomatic labeling and a machine algorithm such as a 2D object detector in case of full-automatic labeling. Fig. 2(a) shows a brief illustration. We propose a Segment, Lift, and Fit (SLF) paradigm to address this problem. We first segment a 2D mask of the target object using the input prompts, then lift it into a 3D model and iteratively optimize its shape and pose until it fits the 2D mask and surrounding LiDAR points. An overall illustration is shown in Fig. 2(b)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 597, + 397, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 597, + 397, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 597, + 397, + 609 + ], + "type": "text", + "content": "3.1 Segment: From 2D Prompts to Instance Masks" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "The problem of auto-labeling 3D shapes from 2D prompts is a significant issue in autonomous driving because it would be highly beneficial to effectively utilize the already established robust 2D instance segmentation models for this purpose. This work employs the strong Segment Anything Model (SAM) [12] to transfer" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 210, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 210, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 210, + 100 + ], + "type": "text", + "content": "J. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 115, + 245, + 225 + ], + "blocks": [ + { + "bbox": [ + 134, + 115, + 245, + 225 + ], + "lines": [ + { + "bbox": [ + 134, + 115, + 245, + 225 + ], + "spans": [ + { + "bbox": [ + 134, + 115, + 245, + 225 + ], + "type": "image", + "image_path": "e4e36a39259e7796dca42e05a0704ea8c03dda066acfc7caa8da949a34313bdf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 143, + 228, + 236, + 238 + ], + "lines": [ + { + "bbox": [ + 143, + 228, + 236, + 238 + ], + "spans": [ + { + "bbox": [ + 143, + 228, + 236, + 238 + ], + "type": "text", + "content": "(a) Definition & challenge" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 249, + 116, + 347, + 224 + ], + "blocks": [ + { + "bbox": [ + 249, + 116, + 347, + 224 + ], + "lines": [ + { + "bbox": [ + 249, + 116, + 347, + 224 + ], + "spans": [ + { + "bbox": [ + 249, + 116, + 347, + 224 + ], + "type": "image", + "image_path": "b4978b9de0169b630f9a0e8efb1bf8581ca0b749110a2477a8b876175954700b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 280, + 228, + 448, + 238 + ], + "lines": [ + { + "bbox": [ + 280, + 228, + 448, + 238 + ], + "spans": [ + { + "bbox": [ + 280, + 228, + 448, + 238 + ], + "type": "text", + "content": "(b) Overall illustration of the proposed method" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 349, + 117, + 479, + 224 + ], + "blocks": [ + { + "bbox": [ + 349, + 117, + 479, + 224 + ], + "lines": [ + { + "bbox": [ + 349, + 117, + 479, + 224 + ], + "spans": [ + { + "bbox": [ + 349, + 117, + 479, + 224 + ], + "type": "image", + "image_path": "13046f29a8c3bb633932bc7285a2df9e30b8e2375e6ca34107f801167761e2a0.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 247, + 482, + 302 + ], + "lines": [ + { + "bbox": [ + 131, + 247, + 482, + 302 + ], + "spans": [ + { + "bbox": [ + 131, + 247, + 482, + 302 + ], + "type": "text", + "content": "Fig. 2: (a) We aim to recover the 3D shape and pose of an interested object, given 2D box or point as prompts. This problem is highly ill-posed. (b) In our SLF, we propose to firstly segment the 2D mask of the target object, and then lift the mask to a 3D form and optimize over its shape and pose by gradient descent until the 3D object fits the mask and LiDAR points." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 324, + 482, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 324, + 482, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 324, + 482, + 396 + ], + "type": "text", + "content": "the input point or box prompts into 2D instance masks. Then, our problem is transferred to estimate the 3D shapes and poses from given instance masks. Despite the high quality of SAM masks, the problem is still extremely difficult to solve. This is because multiple 3D shapes can be mapped to the same 2D mask (Fig. 2(a)), making it a highly ill-posed problem and posing a significant challenge for auto-labeler methods." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 412, + 279, + 424 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 412, + 279, + 424 + ], + "spans": [ + { + "bbox": [ + 131, + 412, + 279, + 424 + ], + "type": "text", + "content": "3.2 Optimization Objective" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 430, + 482, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 482, + 514 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 482, + 514 + ], + "type": "text", + "content": "We employ a pose vector " + }, + { + "bbox": [ + 130, + 430, + 482, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{p} = (x,y,z,\\theta) \\in \\mathbb{R}^4" + }, + { + "bbox": [ + 130, + 430, + 482, + 514 + ], + "type": "text", + "content": " and a latent shape vector " + }, + { + "bbox": [ + 130, + 430, + 482, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{s} \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 130, + 430, + 482, + 514 + ], + "type": "text", + "content": " to represent a 3D object, where " + }, + { + "bbox": [ + 130, + 430, + 482, + 514 + ], + "type": "inline_equation", + "content": "(x,y,z)" + }, + { + "bbox": [ + 130, + 430, + 482, + 514 + ], + "type": "text", + "content": " denote the position of the object center in the global coordinate system, " + }, + { + "bbox": [ + 130, + 430, + 482, + 514 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 430, + 482, + 514 + ], + "type": "text", + "content": " denotes the heading angle around the " + }, + { + "bbox": [ + 130, + 430, + 482, + 514 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 430, + 482, + 514 + ], + "type": "text", + "content": " axis, and the latent dimension " + }, + { + "bbox": [ + 130, + 430, + 482, + 514 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 430, + 482, + 514 + ], + "type": "text", + "content": " is empirically set to 5. We assume all instances are on the ground (the " + }, + { + "bbox": [ + 130, + 430, + 482, + 514 + ], + "type": "inline_equation", + "content": "x-y" + }, + { + "bbox": [ + 130, + 430, + 482, + 514 + ], + "type": "text", + "content": " plane) and omit rotations around the " + }, + { + "bbox": [ + 130, + 430, + 482, + 514 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 430, + 482, + 514 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 430, + 482, + 514 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 430, + 482, + 514 + ], + "type": "text", + "content": " axes. Our goal is to minimize an energy function that corresponds to the posterior probability of the 3D reconstruction given " + }, + { + "bbox": [ + 130, + 430, + 482, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 130, + 430, + 482, + 514 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 430, + 482, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 130, + 430, + 482, + 514 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 238, + 521, + 481, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 521, + 481, + 539 + ], + "spans": [ + { + "bbox": [ + 238, + 521, + 481, + 539 + ], + "type": "interline_equation", + "content": "\\mathbf {p} ^ {*}, \\mathbf {s} ^ {*} = \\underset {\\mathbf {p}, \\mathbf {s}} {\\arg \\min } E (Y, \\mathcal {X} _ {f}, \\mathbf {p}, \\mathbf {s}), \\tag {1}", + "image_path": "52a43f47137ad790013d79ff7b92871e3d1e1319a3023256adaefb10ff444aec.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 547, + 482, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 547, + 482, + 608 + ], + "spans": [ + { + "bbox": [ + 131, + 547, + 482, + 608 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 547, + 482, + 608 + ], + "type": "inline_equation", + "content": "Y \\in [0,1]^{h \\times w}" + }, + { + "bbox": [ + 131, + 547, + 482, + 608 + ], + "type": "text", + "content": " is the 2D instance mask predicted by SAM, and " + }, + { + "bbox": [ + 131, + 547, + 482, + 608 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_f \\in \\mathbb{R}^{N \\times 3}" + }, + { + "bbox": [ + 131, + 547, + 482, + 608 + ], + "type": "text", + "content": " is a subset of the whole scene point cloud " + }, + { + "bbox": [ + 131, + 547, + 482, + 608 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 131, + 547, + 482, + 608 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 131, + 547, + 482, + 608 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_f" + }, + { + "bbox": [ + 131, + 547, + 482, + 608 + ], + "type": "text", + "content": " is comprised of " + }, + { + "bbox": [ + 131, + 547, + 482, + 608 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 131, + 547, + 482, + 608 + ], + "type": "text", + "content": " three-dimensional coordinates of points inside the camera frustum corresponding to the 2D mask. To solve the optimization problem in Eq. (1), we build a differentiable renderer to optimize the energy function " + }, + { + "bbox": [ + 131, + 547, + 482, + 608 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 131, + 547, + 482, + 608 + ], + "type": "text", + "content": " via gradient descent." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 623, + 310, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 623, + 310, + 635 + ], + "spans": [ + { + "bbox": [ + 131, + 623, + 310, + 635 + ], + "type": "text", + "content": "3.3 Lift: Differentiable Rendering" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 641, + 487, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 487, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 487, + 666 + ], + "type": "text", + "content": "Representation of 3D objects. We use the Signed Distance Function (SDF) [22, 37] to represent 3D objects. Formally, an SDF is a three-dimensional scalar field" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 155, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 155, + 91, + 448, + 102 + ], + "type": "text", + "content": "Segment, Lift and Fit: Automatic 3D Shape Labeling from 2D Prompts" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 116, + 479, + 205 + ], + "blocks": [ + { + "bbox": [ + 138, + 116, + 479, + 205 + ], + "lines": [ + { + "bbox": [ + 138, + 116, + 479, + 205 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 479, + 205 + ], + "type": "image", + "image_path": "6bd121f5a2fda56b47886bf8930e25e0255479b32030c4ded30316ae91319d6a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 213, + 482, + 248 + ], + "lines": [ + { + "bbox": [ + 131, + 213, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 131, + 213, + 482, + 248 + ], + "type": "text", + "content": "Fig. 3: We perform PCA on a diverse collection of 3D shapes to obtain the shape latent code. Left: Example of the basis 3D shapes in the collection. Right: Interpolation along the latent space of the shape code (each row) shows smooth shape variations." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 275, + 482, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 275, + 482, + 348 + ], + "spans": [ + { + "bbox": [ + 130, + 275, + 482, + 348 + ], + "type": "inline_equation", + "content": "\\phi (\\mathbf{x})" + }, + { + "bbox": [ + 130, + 275, + 482, + 348 + ], + "type": "text", + "content": " in which the value of a particular spatial coordinate " + }, + { + "bbox": [ + 130, + 275, + 482, + 348 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 275, + 482, + 348 + ], + "type": "text", + "content": " is given by its distance to the closest object surface, with positive and negative values meaning exterior and interior area. In practice, we use the discrete approximation of continuous SDF. Consider a cubic divided into a 3D mesh grid with the size of " + }, + { + "bbox": [ + 130, + 275, + 482, + 348 + ], + "type": "inline_equation", + "content": "l\\times w\\times h" + }, + { + "bbox": [ + 130, + 275, + 482, + 348 + ], + "type": "text", + "content": ". We can represent the inscribed object with the SDF value in each mesh grid, i.e., " + }, + { + "bbox": [ + 130, + 275, + 482, + 348 + ], + "type": "inline_equation", + "content": "\\mathbf{m} = [f(\\mathbf{x}_1),\\dots,f(\\mathbf{x}_i)\\dots,f(\\mathbf{x}_n)]^T" + }, + { + "bbox": [ + 130, + 275, + 482, + 348 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 275, + 482, + 348 + ], + "type": "inline_equation", + "content": "n = l\\times w\\times h" + }, + { + "bbox": [ + 130, + 275, + 482, + 348 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 349, + 482, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 349, + 482, + 421 + ], + "spans": [ + { + "bbox": [ + 130, + 349, + 482, + 421 + ], + "type": "text", + "content": "Shape prior. We aim to condense the nuanced differences present in class-specific 3D shapes into a low-dimensional latent space, where traversing along the manifold of the latent vector " + }, + { + "bbox": [ + 130, + 349, + 482, + 421 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 130, + 349, + 482, + 421 + ], + "type": "text", + "content": " corresponds to the exploration of feasible shapes of the object. We focus on the \"car\" class in this paper, where every instance in the class possesses an overall shape prior and exhibits minor variations in characteristics, such as the shape of the front face or the shape of the cabin." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 422, + 482, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 482, + 543 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 482, + 543 + ], + "type": "text", + "content": "To obtain the shape prior, we first assemble a collection of 3D vehicle models from the Apollo-Car3D dataset [33], which consists of 79 models with exhibit significant variations in shapes and sizes (see the left of Fig. 3). However, using the Signed Distance Function (SDF) representation necessitates that objects be water-tight to avoid ambiguity between their interior and exterior. Some models contain intricate internal structures and are not initially water-tight. To obtain the water-tight surface of 3D models, we employ the SoftRas [21] algorithm to remove redundant vertices and faces inside the model. Finally, we convert the original triangular mesh representation of each 3D model into an SDF representation to obtain a set of models " + }, + { + "bbox": [ + 130, + 422, + 482, + 543 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = \\{\\mathbf{m}_i\\}_{i=1}^{79}" + }, + { + "bbox": [ + 130, + 422, + 482, + 543 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 544, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 482, + 640 + ], + "type": "text", + "content": "Given the collection of representative 3D models, we perform the PCA method on " + }, + { + "bbox": [ + 130, + 544, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 130, + 544, + 482, + 640 + ], + "type": "text", + "content": ", and keep the top-" + }, + { + "bbox": [ + 130, + 544, + 482, + 640 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 544, + 482, + 640 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 544, + 482, + 640 + ], + "type": "inline_equation", + "content": "d = 5" + }, + { + "bbox": [ + 130, + 544, + 482, + 640 + ], + "type": "text", + "content": ") principle components to obtain a projection matrix " + }, + { + "bbox": [ + 130, + 544, + 482, + 640 + ], + "type": "inline_equation", + "content": "V \\in \\mathbb{R}^{d \\times n}" + }, + { + "bbox": [ + 130, + 544, + 482, + 640 + ], + "type": "text", + "content": ". Accordingly, we can use the projection matrix " + }, + { + "bbox": [ + 130, + 544, + 482, + 640 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 544, + 482, + 640 + ], + "type": "text", + "content": " to render a 3D shape " + }, + { + "bbox": [ + 130, + 544, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{m}" + }, + { + "bbox": [ + 130, + 544, + 482, + 640 + ], + "type": "text", + "content": " from a given shape vector " + }, + { + "bbox": [ + 130, + 544, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 130, + 544, + 482, + 640 + ], + "type": "text", + "content": ". Besides, we can encode a 3D model " + }, + { + "bbox": [ + 130, + 544, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{m}" + }, + { + "bbox": [ + 130, + 544, + 482, + 640 + ], + "type": "text", + "content": " into a shape vector " + }, + { + "bbox": [ + 130, + 544, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 130, + 544, + 482, + 640 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 130, + 544, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{s} = V^{\\top}(\\mathbf{m} - \\bar{\\mathbf{m}})" + }, + { + "bbox": [ + 130, + 544, + 482, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 544, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{m} = V\\mathbf{s} + \\bar{\\mathbf{m}}" + }, + { + "bbox": [ + 130, + 544, + 482, + 640 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 544, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{m}}" + }, + { + "bbox": [ + 130, + 544, + 482, + 640 + ], + "type": "text", + "content": " indicates the mean vector of the collection " + }, + { + "bbox": [ + 130, + 544, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 130, + 544, + 482, + 640 + ], + "type": "text", + "content": ". In this work, we select a compact representation " + }, + { + "bbox": [ + 130, + 544, + 482, + 640 + ], + "type": "inline_equation", + "content": "d = 5" + }, + { + "bbox": [ + 130, + 544, + 482, + 640 + ], + "type": "text", + "content": ". The right of Fig. 3 illustrates that interpolation along the latent space of " + }, + { + "bbox": [ + 130, + 544, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 130, + 544, + 482, + 640 + ], + "type": "text", + "content": " leads to smooth variation in shapes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": "Differentiable rendering. After recovering the 3D shape in a local coordinate system, we can put the model in the global coordinate system by adding its" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 210, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 210, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 210, + 100 + ], + "type": "text", + "content": "J. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 142, + 114, + 223, + 177 + ], + "blocks": [ + { + "bbox": [ + 142, + 114, + 223, + 177 + ], + "lines": [ + { + "bbox": [ + 142, + 114, + 223, + 177 + ], + "spans": [ + { + "bbox": [ + 142, + 114, + 223, + 177 + ], + "type": "image", + "image_path": "1bf734251b80c0387ca1a65e835a9dfe5f06b88d1287ef4481da445f07fbb1a0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 156, + 178, + 211, + 187 + ], + "lines": [ + { + "bbox": [ + 156, + 178, + 211, + 187 + ], + "spans": [ + { + "bbox": [ + 156, + 178, + 211, + 187 + ], + "type": "text", + "content": "(a) RGB image" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 225, + 114, + 306, + 177 + ], + "blocks": [ + { + "bbox": [ + 225, + 114, + 306, + 177 + ], + "lines": [ + { + "bbox": [ + 225, + 114, + 306, + 177 + ], + "spans": [ + { + "bbox": [ + 225, + 114, + 306, + 177 + ], + "type": "image", + "image_path": "9f55c4409305f9243c5efe18982a98d28fe326148c91b9593c0edba93d3c94eb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 239, + 178, + 293, + 187 + ], + "lines": [ + { + "bbox": [ + 239, + 178, + 293, + 187 + ], + "spans": [ + { + "bbox": [ + 239, + 178, + 293, + 187 + ], + "type": "text", + "content": "(b) SAM mask" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 114, + 389, + 177 + ], + "blocks": [ + { + "bbox": [ + 307, + 114, + 389, + 177 + ], + "lines": [ + { + "bbox": [ + 307, + 114, + 389, + 177 + ], + "spans": [ + { + "bbox": [ + 307, + 114, + 389, + 177 + ], + "type": "image", + "image_path": "312ae667a1f489734491527cf4776831a541092c1cfb7067d5d1306c6e610356.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 178, + 378, + 187 + ], + "lines": [ + { + "bbox": [ + 320, + 178, + 378, + 187 + ], + "spans": [ + { + "bbox": [ + 320, + 178, + 378, + 187 + ], + "type": "text", + "content": "(c) Target mask" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 157, + 196, + 455, + 208 + ], + "lines": [ + { + "bbox": [ + 157, + 196, + 455, + 208 + ], + "spans": [ + { + "bbox": [ + 157, + 196, + 455, + 208 + ], + "type": "text", + "content": "Fig. 4: Generation of occlusion map " + }, + { + "bbox": [ + 157, + 196, + 455, + 208 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 157, + 196, + 455, + 208 + ], + "type": "text", + "content": " for the mask alignment objective." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 432, + 129, + 472, + 177 + ], + "blocks": [ + { + "bbox": [ + 432, + 129, + 472, + 177 + ], + "lines": [ + { + "bbox": [ + 432, + 129, + 472, + 177 + ], + "spans": [ + { + "bbox": [ + 432, + 129, + 472, + 177 + ], + "type": "image", + "image_path": "e8306911fa36c3706b1d6ade31fa809ab948e3120d8e2b6caf55b1a4f89b052d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 405, + 178, + 457, + 187 + ], + "lines": [ + { + "bbox": [ + 405, + 178, + 457, + 187 + ], + "spans": [ + { + "bbox": [ + 405, + 178, + 457, + 187 + ], + "type": "text", + "content": "(d) Occ. mask" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 230, + 480, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 230, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 130, + 230, + 480, + 266 + ], + "type": "text", + "content": "pose " + }, + { + "bbox": [ + 130, + 230, + 480, + 266 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 130, + 230, + 480, + 266 + ], + "type": "text", + "content": ". This gives us an unbounded SDF " + }, + { + "bbox": [ + 130, + 230, + 480, + 266 + ], + "type": "inline_equation", + "content": "\\phi(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 230, + 480, + 266 + ], + "type": "text", + "content": " of the 3D model, in which the values outside the pre-defined cubic are filled with negative infinity. Notably, the rendering process of shape is differentiable w.r.t. both " + }, + { + "bbox": [ + 130, + 230, + 480, + 266 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 130, + 230, + 480, + 266 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 230, + 480, + 266 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 130, + 230, + 480, + 266 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 282, + 357, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 282, + 357, + 295 + ], + "spans": [ + { + "bbox": [ + 131, + 282, + 357, + 295 + ], + "type": "text", + "content": "3.4 Fit: Optimization over Pose and Shape" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 300, + 480, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 480, + 348 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 480, + 348 + ], + "type": "text", + "content": "Given the 2D instance mask and the point cloud scene, we randomly initialize pose " + }, + { + "bbox": [ + 130, + 300, + 480, + 348 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 130, + 300, + 480, + 348 + ], + "type": "text", + "content": " and shape " + }, + { + "bbox": [ + 130, + 300, + 480, + 348 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 130, + 300, + 480, + 348 + ], + "type": "text", + "content": " and optimize them using Eq. (1). Specifically, the optimization objective consists of three terms: the mask alignment term, the point cloud alignment term, and the ground alignment term." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 349, + 481, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 349, + 481, + 507 + ], + "spans": [ + { + "bbox": [ + 130, + 349, + 481, + 507 + ], + "type": "text", + "content": "Mask alignment term. The mask alignment term measures the dissimilarity between the target instance segmentation mask " + }, + { + "bbox": [ + 130, + 349, + 481, + 507 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 130, + 349, + 481, + 507 + ], + "type": "text", + "content": " and the mask " + }, + { + "bbox": [ + 130, + 349, + 481, + 507 + ], + "type": "inline_equation", + "content": "Y_{proj}" + }, + { + "bbox": [ + 130, + 349, + 481, + 507 + ], + "type": "text", + "content": " obtained by projecting the 3D model onto the image plane. To make the projection process differentiable, we follow [22, 37] to define the projection as " + }, + { + "bbox": [ + 130, + 349, + 481, + 507 + ], + "type": "inline_equation", + "content": "\\pi(p_j) = 1 - \\prod_{\\mathbf{x} \\in \\mathcal{R}_{p_j}} \\frac{1}{e^{\\phi(\\mathbf{x})\\zeta + 1}}" + }, + { + "bbox": [ + 130, + 349, + 481, + 507 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 349, + 481, + 507 + ], + "type": "inline_equation", + "content": "p_j" + }, + { + "bbox": [ + 130, + 349, + 481, + 507 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 130, + 349, + 481, + 507 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 349, + 481, + 507 + ], + "type": "text", + "content": "-th pixel in the image plane and " + }, + { + "bbox": [ + 130, + 349, + 481, + 507 + ], + "type": "inline_equation", + "content": "\\pi(p_j) \\in (0,1)" + }, + { + "bbox": [ + 130, + 349, + 481, + 507 + ], + "type": "text", + "content": " is the mask value. A ray is cast through " + }, + { + "bbox": [ + 130, + 349, + 481, + 507 + ], + "type": "inline_equation", + "content": "p_j" + }, + { + "bbox": [ + 130, + 349, + 481, + 507 + ], + "type": "text", + "content": " from the ego camera, and a set of points " + }, + { + "bbox": [ + 130, + 349, + 481, + 507 + ], + "type": "inline_equation", + "content": "\\mathcal{R}p_j" + }, + { + "bbox": [ + 130, + 349, + 481, + 507 + ], + "type": "text", + "content": " is collected by sampling along the ray. If the ray does not hit the surface of the 3D model, the SDF values of all sampled points " + }, + { + "bbox": [ + 130, + 349, + 481, + 507 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathcal{R}p_j" + }, + { + "bbox": [ + 130, + 349, + 481, + 507 + ], + "type": "text", + "content": " will be negative, resulting in a mask value close to zero. Conversely, if the ray intersects with the 3D model, there must exist sampled points with " + }, + { + "bbox": [ + 130, + 349, + 481, + 507 + ], + "type": "inline_equation", + "content": "\\phi(\\mathbf{x}) > 0" + }, + { + "bbox": [ + 130, + 349, + 481, + 507 + ], + "type": "text", + "content": ", leading to a mask value close to 1. A hyperparameter " + }, + { + "bbox": [ + 130, + 349, + 481, + 507 + ], + "type": "inline_equation", + "content": "\\zeta" + }, + { + "bbox": [ + 130, + 349, + 481, + 507 + ], + "type": "text", + "content": " controls the \"hardness\" of the mask values, where a larger " + }, + { + "bbox": [ + 130, + 349, + 481, + 507 + ], + "type": "inline_equation", + "content": "\\zeta" + }, + { + "bbox": [ + 130, + 349, + 481, + 507 + ], + "type": "text", + "content": " leads to harder values (close to either 0 or 1). Finally, we define the mask alignment term using the dice loss [34] as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 233, + 513, + 480, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 513, + 480, + 540 + ], + "spans": [ + { + "bbox": [ + 233, + 513, + 480, + 540 + ], + "type": "interline_equation", + "content": "E _ {\\text {m a s k}} = 1 - \\frac {2 | (Y _ {\\text {p r o j}} * O) \\bigcap Y |}{| Y _ {\\text {p r o j}} * O | + | Y |}, \\tag {2}", + "image_path": "4f6de3cb164e6af544815a9f2b3de1e42cf026aa2f471ccd474f4038d9495825.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 545, + 480, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 480, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 480, + 605 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 545, + 480, + 605 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 130, + 545, + 480, + 605 + ], + "type": "text", + "content": " is a binary map indicating if the pixel is occluded by a nearer object. As shown in Fig. 4, to obtain the occlusion map, 3D LiDAR points are associated with each 2D mask based on projections, and the depth of each mask is computed as the median depth value of the associated LiDAR points. Masks are sorted by their depth values and then used to compute the occlusion maps." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 606, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 481, + 665 + ], + "type": "text", + "content": "Point cloud alignment term. Solely optimizing the mask alignment term leads to inaccurate estimation due to the missing depth information. To avoid undesired local optimum, we introduce an additional point cloud alignment term. This term measures how well the 3D object's surface aligns with the point cloud inside the corresponding frustum." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 155, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 155, + 91, + 448, + 102 + ], + "type": "text", + "content": "Segment, Lift and Fit: Automatic 3D Shape Labeling from 2D Prompts" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "Intuitively, we can use the " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\ell-1" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " norm of the SDF " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\phi(\\mathcal{X}_f)" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " to determine how well a 3D model fits a given frustum point cloud, denoted as " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_f" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": ". However, this term alone might lead to sub-optimal solutions, as the points could be associated with the back face of the object. To overcome this issue, a second term is introduced that involves casting a ray for each point " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathcal{X}_f" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " starting from the ego camera and recording the coordinate " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " when the first time the ray intersects with the 3D model. This term penalizes the predicted 3D model if the distance between " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " is considerable. The point cloud alignment term can be written as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 223, + 221, + 481, + 253 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 221, + 481, + 253 + ], + "spans": [ + { + "bbox": [ + 223, + 221, + 481, + 253 + ], + "type": "interline_equation", + "content": "E _ {p c} = \\| \\phi (\\mathcal {X} _ {f}) \\| _ {1} + \\frac {1}{| \\hat {\\mathcal {X}} _ {f} |} \\sum_ {\\mathbf {x} \\in \\hat {\\mathcal {X}} _ {f}} \\| \\mathbf {x} - \\mathbf {y} \\| _ {2} ^ {2}. \\tag {3}", + "image_path": "a1d6c93256f585dc8b77b2002ee284f5bf3bacf7f4bbeec97f657787f86ffeb8.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 259, + 480, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 259, + 480, + 282 + ], + "spans": [ + { + "bbox": [ + 131, + 259, + 480, + 282 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{X}}_f" + }, + { + "bbox": [ + 131, + 259, + 480, + 282 + ], + "type": "text", + "content": " is a subset of " + }, + { + "bbox": [ + 131, + 259, + 480, + 282 + ], + "type": "inline_equation", + "content": "\\subseteq \\mathcal{X}_f" + }, + { + "bbox": [ + 131, + 259, + 480, + 282 + ], + "type": "text", + "content": " and represents points whose corresponding ray intersects with the model surface." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 285, + 481, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 285, + 481, + 332 + ], + "spans": [ + { + "bbox": [ + 131, + 285, + 481, + 332 + ], + "type": "text", + "content": "Ground alignment term. Another strong cue that can be leveraged to refine the 3D pose is the height of the ground since we can assume that vehicles run on the road surface. Specifically, we regress a ground plane function " + }, + { + "bbox": [ + 131, + 285, + 481, + 332 + ], + "type": "inline_equation", + "content": "z = g(x,y)" + }, + { + "bbox": [ + 131, + 285, + 481, + 332 + ], + "type": "text", + "content": " using RANSAC [6] and then define a ground alignment objective as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 242, + 342, + 481, + 368 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 342, + 481, + 368 + ], + "spans": [ + { + "bbox": [ + 242, + 342, + 481, + 368 + ], + "type": "interline_equation", + "content": "E _ {\\text {g r o u n d}} = \\| \\hat {z} - \\frac {\\hat {h}}{2} - g (\\hat {x}, \\hat {y}) \\| _ {2} ^ {2}, \\tag {4}", + "image_path": "5c0c79c7d0cb6ef00da6d9e9914c3bd1cccbbd254d062a09a61b91dd4a88a488.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 377, + 480, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 377, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 377, + 480, + 402 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 377, + 480, + 402 + ], + "type": "inline_equation", + "content": "(\\hat{x},\\hat{y},\\hat{z})" + }, + { + "bbox": [ + 130, + 377, + 480, + 402 + ], + "type": "text", + "content": " is the predicted object center, and " + }, + { + "bbox": [ + 130, + 377, + 480, + 402 + ], + "type": "inline_equation", + "content": "\\hat{h}" + }, + { + "bbox": [ + 130, + 377, + 480, + 402 + ], + "type": "text", + "content": " is the height of the object inferred from the predicted 3D shape." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 402, + 482, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 402, + 482, + 485 + ], + "spans": [ + { + "bbox": [ + 131, + 402, + 482, + 485 + ], + "type": "text", + "content": "Optimization. The optimization objective is formulated as a weighted sum of " + }, + { + "bbox": [ + 131, + 402, + 482, + 485 + ], + "type": "inline_equation", + "content": "E_{mask}" + }, + { + "bbox": [ + 131, + 402, + 482, + 485 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 402, + 482, + 485 + ], + "type": "inline_equation", + "content": "E_{pc}" + }, + { + "bbox": [ + 131, + 402, + 482, + 485 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 402, + 482, + 485 + ], + "type": "inline_equation", + "content": "E_{ground}" + }, + { + "bbox": [ + 131, + 402, + 482, + 485 + ], + "type": "text", + "content": ". We initialize pose " + }, + { + "bbox": [ + 131, + 402, + 482, + 485 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 131, + 402, + 482, + 485 + ], + "type": "text", + "content": " with the median value of frustum points and shape " + }, + { + "bbox": [ + 131, + 402, + 482, + 485 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 131, + 402, + 482, + 485 + ], + "type": "text", + "content": " with mean shape " + }, + { + "bbox": [ + 131, + 402, + 482, + 485 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{m}}" + }, + { + "bbox": [ + 131, + 402, + 482, + 485 + ], + "type": "text", + "content": ". Based on the Adam optimizer [11] with a fixed learning rate of 0.1 and 150 iterations, a mini-batch comprising all instances (typically " + }, + { + "bbox": [ + 131, + 402, + 482, + 485 + ], + "type": "inline_equation", + "content": "< 100" + }, + { + "bbox": [ + 131, + 402, + 482, + 485 + ], + "type": "text", + "content": ") in a scene is employed for optimization, as opposed to individual objects. This approach yields a " + }, + { + "bbox": [ + 131, + 402, + 482, + 485 + ], + "type": "inline_equation", + "content": "5 \\times" + }, + { + "bbox": [ + 131, + 402, + 482, + 485 + ], + "type": "text", + "content": " acceleration with negligible accuracy drop." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 505, + 230, + 519 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 505, + 230, + 519 + ], + "spans": [ + { + "bbox": [ + 132, + 505, + 230, + 519 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 530, + 482, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 530, + 482, + 602 + ], + "spans": [ + { + "bbox": [ + 130, + 530, + 482, + 602 + ], + "type": "text", + "content": "Previous auto-labelers can be divided into supervised ones and unsupervised ones. The proposed SLF is an unsupervised method. Therefore, we primarily compare it with existing unsupervised methods. Nonetheless, we also compare it with supervised methods and demonstrate that SLF exhibits superior transfer ability across datasets. In addition to bounding box accuracy, we also evaluate the quality of the predicted 3D shape." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 620, + 391, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 620, + 391, + 633 + ], + "spans": [ + { + "bbox": [ + 131, + 620, + 391, + 633 + ], + "type": "text", + "content": "4.1 Comparison with Unsupervised Auto-labelers" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Dataset and metric. We conduct experiments on the commonly used KITTI [7] dataset. We follow the official training and validation split to get 3712 training" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 210, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 210, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 210, + 101 + ], + "type": "text", + "content": "J. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 177, + 135, + 438, + 258 + ], + "blocks": [ + { + "bbox": [ + 194, + 114, + 418, + 126 + ], + "lines": [ + { + "bbox": [ + 194, + 114, + 418, + 126 + ], + "spans": [ + { + "bbox": [ + 194, + 114, + 418, + 126 + ], + "type": "text", + "content": "Table 1: Comparison with unsupervised auto-labelers." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 177, + 135, + 438, + 258 + ], + "lines": [ + { + "bbox": [ + 177, + 135, + 438, + 258 + ], + "spans": [ + { + "bbox": [ + 177, + 135, + 438, + 258 + ], + "type": "table", + "html": "
MethodAPBEV(IoU=0.5)|R40AP3D(IoU=0.5)|R40
Easy Mod.HardEasy Mod.Hard
VS3D [31]74.566.757.640.337.431.1
SDFLabel [44]77.859.8-62.342.2-
McCraith et al. [23]75.576.668.6---
FGR [39]93.777.870.591.377.870.6
SLFGTmask89.288.987.187.687.883.8
SLFpoint88.588.584.284.884.980.8
SLFbox89.289.384.787.988.283.9
", + "image_path": "a270bfb8047f925141e774688462a67c5a608cb74fb8551a96eaf51d6bfe021f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 281, + 482, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 389 + ], + "type": "text", + "content": "frames and 3768 validation frames. To align with prior works [15,39], the \"Car\" category is the focus of auto-labeling efforts. The Bird's Eye View (BEV) AP and the 3D AP are selected as evaluation metrics for a comprehensive comparison. Results are reported for all three levels of difficulty (Easy, Moderate, and Hard). The commonly used KITTI metrics are evaluated at a strict threshold of 0.7 IoU. As observed in [23,44], the 3D boxes in KITTI have varying spatial padding, making it challenging for human annotators to determine a precise size for a significant number of instances. Therefore, we follow the previous work [23,44] and set the IoU threshold as 0.5 and calculate AP by 40 recall positions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 389, + 482, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 389, + 482, + 497 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 482, + 497 + ], + "type": "text", + "content": "Direct evaluation on generated pseudo labels. A straightforward evaluation is to compare the quality of auto-labeled 3D bounding boxes in terms of the canonical metric, Average Precision (AP). However, the AP metric needs a confidence score for each bounding box for thresholding and integration, but FGR [39] and SLF do not output confidence scores. To remedy this issue, we simply define IoU between the predicted mask and the target mask in SLF as a confidence score. As for FGR, we first project a 3D bounding box to image plane to get a 2D box, and the confidence score is defined by IoU between the predicted 2D box and the ground truth 2D box." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "type": "text", + "content": "Tab. 1 compares the proposed SLF with existing unsupervised methods on KITTI validation set. " + }, + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\mathrm{SLF}_{\\mathrm{point}}" + }, + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "type": "text", + "content": " uses 2D points as input, corresponding to the semi-automatic labeling scenario with humans in the loop. The point prompts are simulated by random sampling 3 points inside the ground truth masks [9]. " + }, + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\mathrm{SLF}_{\\mathrm{box}}" + }, + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "type": "text", + "content": " indicates 2D bounding boxes as input, corresponding to the full-automatic labeling scenario in which 2D detectors generate the prompts. For all methods that accept 2D boxes as inputs, we use ground truth boxes for a fair comparison. We also provide a variant of our method in which the SAM masks are replaced with ground truth masks from [9], denoted as " + }, + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\mathrm{SLF}_{\\mathrm{GTmask}}" + }, + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "The last three rows in Tab. 1 show the results of SLF with different 2D prompts. " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{SLF}_{\\mathrm{point}}" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{SLF}_{\\mathrm{box}}" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " perform roughly the same in terms of " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{BEV}" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{SLF}_{\\mathrm{box}}" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " shows superior " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{3D}" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " performance. To our surprise, " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{SLF}_{\\mathrm{box}}" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " is slightly better than " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{SLF}_{\\mathrm{GTmask}}" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " in terms of both " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{3D}" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{BEV}" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " except on hard samples, which means the 2D mask generated by SAM from box prompts are even" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 155, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 155, + 91, + 448, + 103 + ], + "type": "text", + "content": "Segment, Lift and Fit: Automatic 3D Shape Labeling from 2D Prompts" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 168, + 135, + 446, + 309 + ], + "blocks": [ + { + "bbox": [ + 184, + 114, + 427, + 126 + ], + "lines": [ + { + "bbox": [ + 184, + 114, + 427, + 126 + ], + "spans": [ + { + "bbox": [ + 184, + 114, + 427, + 126 + ], + "type": "text", + "content": "Table 2: Detector performance trained with pseudo labels." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 168, + 135, + 446, + 309 + ], + "lines": [ + { + "bbox": [ + 168, + 135, + 446, + 309 + ], + "spans": [ + { + "bbox": [ + 168, + 135, + 446, + 309 + ], + "type": "table", + "html": "
DetectorsLabels\\(AP_{BEV}(IoU=0.5)\\) \\(|_{R40}\\)\\(AP_{3D}(IoU=0.5)\\) \\(|_{R40}\\)
EasyMod.HardEasyMod.Hard
PointPillar [14]GT95.794.393.295.793.691.4
FGR96.590.287.194.989.484.8
SLF95.394.191.195.292.188.8
SECOND [42]GT95.994.893.995.894.691.9
FGR95.488.986.295.388.685.7
SLF95.694.291.795.592.489.3
Part-\\(A^2\\) [32]GT96.094.393.996.094.291.9
FGR96.192.187.596.090.085.1
SLF96.094.591.895.993.989.6
Voxel R-CNN [3]GT98.494.994.598.394.894.4
FGR98.790.988.298.690.788.0
SLF96.295.192.496.194.791.9
", + "image_path": "9f4e65eae4c5415ae6867108ae6611b04cc8fb57665011b933509da89f71a3eb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 335, + 480, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 480, + 360 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 480, + 360 + ], + "type": "text", + "content": "better than those annotated by humans. This suggests that SLF benefits from more accurate input prompts, and more accurate instance masks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 361, + 481, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 481, + 505 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 481, + 505 + ], + "type": "text", + "content": "We compare " + }, + { + "bbox": [ + 130, + 361, + 481, + 505 + ], + "type": "inline_equation", + "content": "\\mathrm{SLF}_{\\mathrm{box}}" + }, + { + "bbox": [ + 130, + 361, + 481, + 505 + ], + "type": "text", + "content": " with other unsupervised methods as they accept the same input prompts. It can be observed that SLF outperforms FGR and SDFLabel by a significant margin on the Moderate and Hard samples. However, SLF performs worse than FGR on the Easy samples, and we assume that easy cases in KITTI have much denser sub-points, which are beneficial to the segmentation and box estimation stage of FGR. Moreover, FGR consumes ground-truth amodal bounding box, while in KITTI these boxes are obtained by projecting 3D ground-truth boxes to 2D images, which may leak the oracle occlusion and size information of partially visible objects. SDFLabel uses a similar 3D shape representation to SLF but relies on a synthetic dataset for training, which may limit its performance on real-world data. In contrast, SLF does not require synthetic training data, and generalizes well to real-world scenarios." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 507, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 507, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 507, + 482, + 616 + ], + "type": "text", + "content": "Detector performance trained with pseudo labels. Following previous work [15,24,39], to further evaluate the quality of generated pseudo labels, we use them as training samples to train off-the-shelf detectors and compare the results of different detectors. We mainly compare with the previous best-performing FGR auto-labeler and train 4 different 3D detectors. Specifically, for all detectors, we use the OpenPCDet [36] with default hyperparameters. In Tab. 2, we compare the results of detectors trained on KITTI training set with ground truth (GT) labels, FGR pseudo labels, and SLF pseudo labels, respectively. If not specified, SLF refers to " + }, + { + "bbox": [ + 130, + 507, + 482, + 616 + ], + "type": "inline_equation", + "content": "\\mathrm{SLF}_{\\mathrm{box}}" + }, + { + "bbox": [ + 130, + 507, + 482, + 616 + ], + "type": "text", + "content": " in the remaining sections." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Tab. 2 shows SLF pseudo labels significantly improve upon FGR pseudo labels on all detectors on Moderate and Hard samples. Regarding Easy samples, SLF pseudo labels are on par with FGR pseudo labels on all detectors except Voxel R-CNN [3]. When comparing specific detectors, the one trained with SLF" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 210, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 210, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 210, + 100 + ], + "type": "text", + "content": "J. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": "labels performs almost as well as the one trained with ground truth labels on both Easy and Moderate samples, and slightly worse on Hard samples, which shows the effectiveness of our method." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 172, + 308, + 183 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 172, + 308, + 183 + ], + "spans": [ + { + "bbox": [ + 131, + 172, + 308, + 183 + ], + "type": "text", + "content": "4.2 Cross Dataset Generalization" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 194, + 482, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 194, + 482, + 327 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 482, + 327 + ], + "type": "text", + "content": "Apart from KITTI, we conduct experiments on more challenging nuScenes [1]. We transform the validation set of nuScenes to KITTI format, resulting in 6019 frames with corresponding front view images, point cloud and 3D bounding box annotations. On KITTI, we compare the pseudo label with ground truths and report " + }, + { + "bbox": [ + 130, + 194, + 482, + 327 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{3D}" + }, + { + "bbox": [ + 130, + 194, + 482, + 327 + ], + "type": "text", + "content": ", and on nuScenes, we report the official metrics including mAP, NDS, mATE, mASE and mAOE. We compare with the best-performing unsupervised auto-labeler FGR, and the supervised auto-labeler MTrans [15]. Both FGR and Mtrans require 2D amodal boxes as prompts, so we project 3d ground truth boxes into 2D as their input. Our SLF supports a more practical setting that only requires regular 2D modal boxes as prompts. Thus we can generate them with an off-the-shelf detector HTC [2] pre-trained on nuImages [1]." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 135, + 413, + 479, + 487 + ], + "blocks": [ + { + "bbox": [ + 130, + 348, + 483, + 404 + ], + "lines": [ + { + "bbox": [ + 130, + 348, + 483, + 404 + ], + "spans": [ + { + "bbox": [ + 130, + 348, + 483, + 404 + ], + "type": "text", + "content": "Table 3: Cross dataset evaluation compared with state-of-the-art unsupervised auto-labeler FGR and supervised auto-labeler MTrans. Results in gray are less informative because the NDS and error metrics are computed only on recalled samples, while FGR fails on recalling a majority of input samples (low mAP). " + }, + { + "bbox": [ + 130, + 348, + 483, + 404 + ], + "type": "inline_equation", + "content": "\\mathrm{mAOE}^{\\dagger}" + }, + { + "bbox": [ + 130, + 348, + 483, + 404 + ], + "type": "text", + "content": " means to change the periodicity of orientation to " + }, + { + "bbox": [ + 130, + 348, + 483, + 404 + ], + "type": "inline_equation", + "content": "[0,\\pi]" + }, + { + "bbox": [ + 130, + 348, + 483, + 404 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 413, + 479, + 487 + ], + "lines": [ + { + "bbox": [ + 135, + 413, + 479, + 487 + ], + "spans": [ + { + "bbox": [ + 135, + 413, + 479, + 487 + ], + "type": "table", + "html": "
MethodTraining DataAP3D@KITTInuScenes Metrics
EasyMod.HardmAPNDSmATEmASEmAOEmAOE†
FGR [39]-91.377.870.611.640.4.152.188.010.010
MTrans [15]KITTI99.699.697.145.949.4.358.349.702.225
SLF-87.988.283.957.255.2.372.192.836.113
", + "image_path": "03b9a91528b17c10fe27dcc404f3b77b0055cfaa7b4da56cff4729c92fb405f6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "In Tab. 3, on KITTI, MTrans shows a nearly perfect performance, this is not surprising since MTrans is a learnable neural network trained on KITTI, it easily learns the specific annotation patterns in training labels. Despite the high accuracy, we argue that learnable auto-labelers like MTrans are prone to overfit the biased pattern in the training data and generalize poorly across datasets. When we apply MTrans to the challenging nuScenes dataset, its performance dramatically decreases. SLF performs better in terms of mAP and NDS. Regarding the detailed error terms, we find SLF outperforms MTrans on size estimation (mASE), and is on par with MTrans on location estimation (mATE). Regarding orientation estimation, SLF underperforms MTrans by 0.134. However, we find a typical orientation error is the model mistakes the front end of the car for the rear end, which is usually caused when the LiDAR points are rather sparse. To explore how good the orientation predictions are when the " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pi" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 155, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 155, + 91, + 448, + 102 + ], + "type": "text", + "content": "Segment, Lift and Fit: Automatic 3D Shape Labeling from 2D Prompts" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": "phase lag is disentangled, we change the periodicity of orientation to " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "[0,\\pi ]" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": " and re-measure the orientation error, denoted as " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "\\mathrm{mAOE}^{\\dagger}" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": ". SLF outperforms MTrans on " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "\\mathrm{mAOE}^{\\dagger}" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": ", suggesting that SLF predicts more accurate orientation, the heading direction aligning better with ground truths. When applying FGR on the nuScences dataset, it fails on a majority of samples, resulting in a low mAP of " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "11.6\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": ". Consequently, the NDS score and error terms are less informative since they are computed only on successfully labeled samples, typically easy samples." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 224, + 217, + 337 + ], + "blocks": [ + { + "bbox": [ + 133, + 224, + 217, + 337 + ], + "lines": [ + { + "bbox": [ + 133, + 224, + 217, + 337 + ], + "spans": [ + { + "bbox": [ + 133, + 224, + 217, + 337 + ], + "type": "image", + "image_path": "d19ca68f87c4ad4daa7b11dc6d06115a0a5e400f7ec0cb62ee4e5aee6daf1a74.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 346, + 303, + 413 + ], + "lines": [ + { + "bbox": [ + 130, + 346, + 303, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 303, + 413 + ], + "type": "text", + "content": "Fig.5: Qualitative results of estimated shapes. From left to right: projected mask of the initial 3D model; projected mask of the optimized 3D model; mean shape and surrounding point cloud; optimized 3D shape and surrounding point cloud." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 217, + 225, + 300, + 337 + ], + "blocks": [ + { + "bbox": [ + 217, + 225, + 300, + 337 + ], + "lines": [ + { + "bbox": [ + 217, + 225, + 300, + 337 + ], + "spans": [ + { + "bbox": [ + 217, + 225, + 300, + 337 + ], + "type": "image", + "image_path": "20522f0553e92b0d5d8ff3489ac7cd44a50b017711bcb56b61c898726f4f7988.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 311, + 346, + 482, + 413 + ], + "lines": [ + { + "bbox": [ + 311, + 346, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 311, + 346, + 482, + 413 + ], + "type": "text", + "content": "Fig.6: Comparison of occupancy annotation. From left to right: reference instance image; occupancy labels from OFN [25]; occupancy labels from OpenOccupancy [38]; occupancy labels generated by our SLF." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 315, + 224, + 480, + 335 + ], + "blocks": [ + { + "bbox": [ + 315, + 224, + 480, + 335 + ], + "lines": [ + { + "bbox": [ + 315, + 224, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 315, + 224, + 480, + 335 + ], + "type": "image", + "image_path": "2cc134221cba5ecd9ac29cb81169f87978f12b8e02370433672c154fabaee569.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 460, + 308, + 473 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 460, + 308, + 473 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 308, + 473 + ], + "type": "text", + "content": "4.3 Voxel Occupancy Annotation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": "Since benchmarks [1,7] (i.e., KITTI and nuScences) do not support quantitative evaluation on 3D shapes, we show qualitative results in Fig. 5 to better show the effect of SLF on shape labeling. Specifically, we replace the estimated shape with a mean shape and show comparisons. A clear improvement can be observed by comparing the initial shape and the final optimized shape, both in terms of mask alignment and LiDAR point alignment. Besides, we observe that SLF can serve as a good auto-labeler for occupancy annotation of dynamic objects such as vehicles. We compare two versions of occupancy labels for nuScenes, OpenOccupancy [38] and OFN [25]. Considering that existing occupancy labels are still of low quality, quantitative evaluation may be misleading. Therefore we show qualitative visualizations in Fig. 6, and we observe that OpenOccupancy and OFN labels fail to depict the inner occupancy of the cars. This is mainly attributed to that they label the occupancy by voxelizing the LiDAR points, and therefore self-occlusion can not be avoided. In contrast, SLF can predict dense occupancy more accurately, with more detailed shapes." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 210, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 210, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 210, + 101 + ], + "type": "text", + "content": "J. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 239, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 239, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 239, + 129 + ], + "type": "text", + "content": "4.4 Ablation Study" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 246 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 246 + ], + "type": "text", + "content": "In this section, we conduct comprehensive ablation studies on the KITTI validation set and report the " + }, + { + "bbox": [ + 130, + 138, + 482, + 246 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{3D}(\\mathrm{IoU} = 0.5)|_{R40}" + }, + { + "bbox": [ + 130, + 138, + 482, + 246 + ], + "type": "text", + "content": " results. Except in Tab. 6, we input 2D points to SLF, other results are obtained by using 2D bounding boxes as prompt. Effect of energy terms. In Tab. 4, we observe that introducing the mask alignment term significantly improves pseudo-label quality rather than solely applying the point cloud alignment term, and adding the ground alignment term further improves performance. However, as we mentioned above, using mask alignment alone can easily be trapped in sub-optimal solutions, thus it can not obtain valid results (Easy: 6.9, Mod.: 5.9, Hard: 5.6)." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 138, + 289, + 319, + 347 + ], + "blocks": [ + { + "bbox": [ + 147, + 269, + 296, + 280 + ], + "lines": [ + { + "bbox": [ + 147, + 269, + 296, + 280 + ], + "spans": [ + { + "bbox": [ + 147, + 269, + 296, + 280 + ], + "type": "text", + "content": "Table 4: Ablation on energy terms." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 289, + 319, + 347 + ], + "lines": [ + { + "bbox": [ + 138, + 289, + 319, + 347 + ], + "spans": [ + { + "bbox": [ + 138, + 289, + 319, + 347 + ], + "type": "table", + "html": "
ObjectiveEasyMod.Hard
Epc70.564.862.9
Emask+Epc87.388.083.7
Emask+Epc+Eground87.988.283.9
", + "image_path": "7bf9328807e4c352ac3a33e7e1817156a69c9949f3c8f7e2fcbd2d653ff4c0b6.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 323, + 289, + 459, + 346 + ], + "blocks": [ + { + "bbox": [ + 321, + 269, + 460, + 280 + ], + "lines": [ + { + "bbox": [ + 321, + 269, + 460, + 280 + ], + "spans": [ + { + "bbox": [ + 321, + 269, + 460, + 280 + ], + "type": "text", + "content": "Table 5: Ablation on 3D models." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 323, + 289, + 459, + 346 + ], + "lines": [ + { + "bbox": [ + 323, + 289, + 459, + 346 + ], + "spans": [ + { + "bbox": [ + 323, + 289, + 459, + 346 + ], + "type": "table", + "html": "
# modelsEasyMod.Hard
574.770.164.0
3084.283.677.3
7987.988.283.9
", + "image_path": "71d5b5540cfd1634fb124f7d721e52f98ba874ff07ad7b12665da3c0bbc98090.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 369, + 481, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 481, + 417 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 481, + 417 + ], + "type": "text", + "content": "Effect of the number of CAD models. In Tab. 5, we randomly sample 5 and 30 CAD models out of the total 79 samples to investigate the influence of the number of CAD models used for PCA. From the results, we can conclude that better result comes with more models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 418, + 482, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 482, + 477 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 482, + 477 + ], + "type": "text", + "content": "Effect of the PCA dimensions. In Tab. 7, we show how different PCA dimensions affect results. Performance improves when we increase the dimension from 3 to 5. More dimensions like 10 do not further improve. This is possibly attributed to that a higher dimension of shape representation may complicate the optimization process, which means it may need more steps to converge." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 479, + 482, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 482, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 482, + 552 + ], + "type": "text", + "content": "Effect of the number of points. In Tab. 6, we analyze the effect of the number of points input to SAM [12]. When we increase the points from 3 to 8, the improvement of pseudo-label quality is minor. However, we observe a significant decrease in the easy cases of KITTI when using one point as the prompt, which is reasonable as easy cases often have large masks and a single point can not provide sufficient context for SAM to generate high-quality results." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 571, + 246, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 571, + 246, + 582 + ], + "spans": [ + { + "bbox": [ + 132, + 571, + 246, + 582 + ], + "type": "text", + "content": "4.5 Further Analysis" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "content": "Robustness of LiDAR Beam. Point clouds in KITTI [7] are collected by 64 beams LiDAR, we downsample point cloud to 32, 16 and 8 beams following [40]. " + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{3D}" + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "content": " results on KITTI validation set in Tab. 8 show SLF can still get decent results on 32 and 16 beams LiDAR, and the performances downgrade rapidly on Mod. and Hard cases, since these cases usually far away from ego vehicle which exacerbating the sparsity of point clouds." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 154, + 90, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 90, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 154, + 90, + 447, + 103 + ], + "type": "text", + "content": "Segment, Lift and Fit: Automatic 3D Shape Labeling from 2D Prompts" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 101 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 101 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 133, + 249, + 209 + ], + "blocks": [ + { + "bbox": [ + 141, + 114, + 229, + 125 + ], + "lines": [ + { + "bbox": [ + 141, + 114, + 229, + 125 + ], + "spans": [ + { + "bbox": [ + 141, + 114, + 229, + 125 + ], + "type": "text", + "content": "Table 6: No. points." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 133, + 249, + 209 + ], + "lines": [ + { + "bbox": [ + 132, + 133, + 249, + 209 + ], + "spans": [ + { + "bbox": [ + 132, + 133, + 249, + 209 + ], + "type": "table", + "html": "
# PointsEasyMod.Hard
384.884.980.8
178.181.178.8
584.885.081.0
885.085.280.9
", + "image_path": "521d705aa31c6f2b46a0bf16ece126c389ee520ccec3130d9b6bf27474faf0ba.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 257, + 135, + 357, + 209 + ], + "blocks": [ + { + "bbox": [ + 268, + 114, + 346, + 125 + ], + "lines": [ + { + "bbox": [ + 268, + 114, + 346, + 125 + ], + "spans": [ + { + "bbox": [ + 268, + 114, + 346, + 125 + ], + "type": "text", + "content": "Table 7: No. dim." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 257, + 135, + 357, + 209 + ], + "lines": [ + { + "bbox": [ + 257, + 135, + 357, + 209 + ], + "spans": [ + { + "bbox": [ + 257, + 135, + 357, + 209 + ], + "type": "table", + "html": "
Dim.EasyMod.Hard
587.988.283.9
387.187.883.6
886.987.184.1
1084.885.481.1
", + "image_path": "7f63de3023bd2e66e875b77c341cae3e2d9db046bf6ad3ac45e8872d89da1ebb.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 375, + 135, + 478, + 209 + ], + "blocks": [ + { + "bbox": [ + 385, + 114, + 470, + 125 + ], + "lines": [ + { + "bbox": [ + 385, + 114, + 470, + 125 + ], + "spans": [ + { + "bbox": [ + 385, + 114, + 470, + 125 + ], + "type": "text", + "content": "Table 8: No. beam." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 375, + 135, + 478, + 209 + ], + "lines": [ + { + "bbox": [ + 375, + 135, + 478, + 209 + ], + "spans": [ + { + "bbox": [ + 375, + 135, + 478, + 209 + ], + "type": "table", + "html": "
BeamEasyMod.Hard
6487.988.283.9
3281.377.577.3
1680.776.370.0
875.160.056.7
", + "image_path": "babc2a323788f96873ef2aa47b376d2a25ebba9db6ad44349e1ac25aa264445d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 148, + 241, + 298, + 316 + ], + "blocks": [ + { + "bbox": [ + 146, + 221, + 297, + 232 + ], + "lines": [ + { + "bbox": [ + 146, + 221, + 297, + 232 + ], + "spans": [ + { + "bbox": [ + 146, + 221, + 297, + 232 + ], + "type": "text", + "content": "Table 9: Dilation on instance mask." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 148, + 241, + 298, + 316 + ], + "lines": [ + { + "bbox": [ + 148, + 241, + 298, + 316 + ], + "spans": [ + { + "bbox": [ + 148, + 241, + 298, + 316 + ], + "type": "table", + "html": "
KernelEasyMod.Hard
FGR(no dil.)91.377.870.6
SLF(no dil.)87.988.283.9
5x586.385.681.2
9x983.980.576.6
", + "image_path": "e9e3a0a35a8b5949b4666db5fb71b900e3867a261d207f7c82902028055f4377.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 316, + 241, + 467, + 316 + ], + "blocks": [ + { + "bbox": [ + 314, + 221, + 467, + 232 + ], + "lines": [ + { + "bbox": [ + 314, + 221, + 467, + 232 + ], + "spans": [ + { + "bbox": [ + 314, + 221, + 467, + 232 + ], + "type": "text", + "content": "Table 10: Erosion on instance mask." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 241, + 467, + 316 + ], + "lines": [ + { + "bbox": [ + 316, + 241, + 467, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 467, + 316 + ], + "type": "table", + "html": "
KernelEasyMod.Hard
FGR(noERO.)91.377.870.6
SLF(noERO.)87.988.283.9
5x585.283.979.5
9x982.178.272.1
", + "image_path": "5b5f281986e619c47caf3245321b722464ed195d54203c41fbee0df1402400ff.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 339, + 482, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 339, + 482, + 401 + ], + "spans": [ + { + "bbox": [ + 130, + 339, + 482, + 401 + ], + "type": "text", + "content": "Robustness of Mask Quality. We simulate low-quality masks by applying erosion and dilation operations on SAM [12] masks and report " + }, + { + "bbox": [ + 130, + 339, + 482, + 401 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{3D}" + }, + { + "bbox": [ + 130, + 339, + 482, + 401 + ], + "type": "text", + "content": " results of KITTI validation set in Tab. 9 and Tab. 10. We observe even with severely corrupted masks, SLF still outperforms the existing unsupervised auto-annotation methods in Mod. and Hard cases, which show the robustness of SLF." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 423, + 217, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 423, + 217, + 436 + ], + "spans": [ + { + "bbox": [ + 132, + 423, + 217, + 436 + ], + "type": "text", + "content": "5 Limitation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 454, + 481, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 454, + 481, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 454, + 481, + 491 + ], + "type": "text", + "content": "The proposed SLF uses a strong shape prior learned by PCA from given 3D CAD models. The shape prior is category-specific, which means we need to prepare a collection of 3D models when labeling a specific semantic class." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 515, + 220, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 515, + 220, + 528 + ], + "spans": [ + { + "bbox": [ + 132, + 515, + 220, + 528 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "This paper proposes a Segment, Lift, and Fit paradigm for automatic 3D shape annotation of objects from the 2D point or box prompts. We first segment high-quality instance masks from the prompts using the SAM, and then lift 2D masks to 3D forms, where gradient descent is used to adjust their poses and shapes until the projections fits the masks and the surfaces conform to surrounding LiDAR points. Extensive experimental results on multiple datasets for 3D detection demonstrate the effectiveness of our SLF method. Moreover, the SLF auto-labeler achieves promising results in detailed shape predictions, which can be used for occupancy annotation. In the future, we will investigate the generalization abilities by applying our SLF to more categories." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 210, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 210, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 210, + 100 + ], + "type": "text", + "content": "J. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "type": "text", + "content": "1. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. arXiv preprint arXiv:1903.11027 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 173, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 173, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 138, + 173, + 480, + 205 + ], + "type": "text", + "content": "2. Chen, K., Pang, J., Wang, J., Xiong, Y., Li, X., Sun, S., Feng, W., Liu, Z., Shi, J., Ouyang, W., et al.: Hybrid task cascade for instance segmentation. In: CVPR. pp. 4974-4983 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 206, + 480, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 206, + 480, + 227 + ], + "spans": [ + { + "bbox": [ + 138, + 206, + 480, + 227 + ], + "type": "text", + "content": "3. Deng, J., Shi, S., Li, P., Zhou, W., Zhang, Y., Li, H.: Voxel r-cnn: Towards high performance voxel-based 3d object detection. arXiv:2012.15712 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 228, + 480, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 480, + 271 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 480, + 271 + ], + "type": "text", + "content": "4. Engelmann, F., Stickler, J., Leibe, B.: Joint object pose estimation and shape reconstruction in urban street scenes using 3d shape priors. In: Pattern Recognition: 38th German Conference, GCPR 2016, Hannover, Germany, September 12-15, 2016, Proceedings 38. pp. 219-230. Springer (2016)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 272, + 480, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 272, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 138, + 272, + 480, + 304 + ], + "type": "text", + "content": "5. Engelmann, F., Stuckler, J., Leibe, B.: Samp: shape and motion priors for 4d vehicle reconstruction. In: 2017 IEEE Winter Conference on Applications of Computer Vision (WACV). pp. 400-408. IEEE (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 304, + 480, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 304, + 480, + 336 + ], + "spans": [ + { + "bbox": [ + 138, + 304, + 480, + 336 + ], + "type": "text", + "content": "6. Fischler, M.A., Bolles, R.C.: Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM 24(6), 381-395 (1981)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 337, + 480, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 337, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 138, + 337, + 480, + 358 + ], + "type": "text", + "content": "7. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: CVPR (2012)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 359, + 480, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 359, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 138, + 359, + 480, + 392 + ], + "type": "text", + "content": "8. Guo, J., Liu, J., Xu, D.: 3d-pruning: A model compression framework for efficient 3d action recognition. IEEE Transactions on Circuits and Systems for Video Technology 32(12), 8717-8729 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 392, + 480, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 392, + 480, + 425 + ], + "spans": [ + { + "bbox": [ + 138, + 392, + 480, + 425 + ], + "type": "text", + "content": "9. Heylen, J., De Wolf, M., Dawagne, B., Proesmans, M., Van Gool, L., Abbeloos, W., Abdelkawy, H., Reino, D.O.: Monocinis: Camera independent monocular 3d object detection using instance segmentation. In: ICCV. pp. 923-934 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 425, + 480, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 425, + 480, + 446 + ], + "spans": [ + { + "bbox": [ + 138, + 425, + 480, + 446 + ], + "type": "text", + "content": "0. Huang, Y., Zheng, W., Zhang, Y., Zhou, J., Lu, J.: Tri-perspective view for vision-based 3d semantic occupancy prediction. arXiv preprint arXiv:2302.07817 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 447, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 447, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 447, + 480, + 468 + ], + "type": "text", + "content": "1. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. In: ICLR (2015)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 468, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 468, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 138, + 468, + 480, + 501 + ], + "type": "text", + "content": "2. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. arXiv:2304.02643 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 501, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 501, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 501, + 480, + 533 + ], + "type": "text", + "content": "3. Kundu, A., Li, Y., Rehg, J.M.: 3d-rcnn: Instance-level 3d object reconstruction via render-and-compare. In: CVPR (2018). https://doi.org/10.1109/CVPR.2018.00375, https://doi.org/10.1109/CVPR.2018.00375" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 534, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 534, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 534, + 480, + 567 + ], + "type": "text", + "content": "4. Lang, A.H., Vora, S., Caesar, H., Zhou, L., Yang, J., Beijbom, O.: Pointpillars: Fast encoders for object detection from point clouds. In: CVPR. pp. 12697-12705 (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 567, + 480, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 480, + 589 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 480, + 589 + ], + "type": "text", + "content": "5. Liu, C., Qian, X., Huang, B., Qi, X., Lam, E., Tan, S.C., Wong, N.: Multimodal transformer for automatic 3d annotation and object detection. In: ECCV (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 589, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 589, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 138, + 589, + 480, + 610 + ], + "type": "text", + "content": "6. Liu, J., Guo, J., Xu, D.: Aptsnet: Toward adaptive point sampling for efficient 3d action recognition. IEEE Transactions on Image Processing 31, 5287-5302 (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 610, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 610, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 610, + 480, + 632 + ], + "type": "text", + "content": "7. Liu, J., Guo, J., Xu, D.: Geometrhythmotion-transformer: An end-to-end framework for 3d action recognition. IEEE Transactions on Multimedia (2022)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "text", + "content": "8. Liu, J., He, T., Yang, H., Su, R., Tian, J., Wu, J., Guo, H., Xu, K., Ouyang, W.: 3d-queryis: A query-based framework for 3d instance segmentation. arXiv preprint arXiv:2211.09375 (2022)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 155, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 155, + 91, + 447, + 102 + ], + "type": "text", + "content": "Segment, Lift and Fit: Automatic 3D Shape Labeling from 2D Prompts" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "text", + "content": "19. Liu, J., Li, J., Wang, K., Guo, H., Yang, J., Peng, J., Xu, K., Liu, X., Guo, J.: Lta-pcs: Learnable task-agnostic point cloud sampling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 28035-28045 (June 2024)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 162, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 162, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 162, + 481, + 194 + ], + "type": "text", + "content": "20. Liu, J., Xu, D.: Geometrhythmotion-net: A strong two-stream baseline for 3d action recognition. IEEE Transactions on Circuits and Systems for Video Technology 31(12), 4711-4721 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "text", + "content": "21. Liu, S., Li, T., Chen, W., Li, H.: Soft rasterizer: A differentiable renderer for image-based 3d reasoning. ICCV (Oct 2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "type": "text", + "content": "22. Liu, Z., Zhou, D., Lu, F., Fang, J., Zhang, L.: Autoshape: Real-time shape-aware monocular 3d object detection. In: ICCV. pp. 15641-15650 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 481, + 259 + ], + "type": "text", + "content": "23. McCraith, R., Insafutdinov, E., Neumann, L., Vedaldi, A.: Lifting 2d object locations to 3d by discounting lidar outliers across objects and views (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "type": "text", + "content": "24. Meng, Q., Wang, W., Zhou, T., Shen, J., Van Gool, L., Dai, D.: Weakly supervised 3d object detection from lidar point cloud. In: ECCV (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 281, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 481, + 304 + ], + "type": "text", + "content": "25. Ming Fang, Z.L.: Occupancy dataset for nuscenes. https://github.com/FANG-MING/occupancy-for-nuscenes (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "type": "text", + "content": "26. Osher, S., Fedkiw, R.: Signed Distance Functions. Springer New York (2003). https://doi.org/10.1007/0-387-22746-6_2, https://doi.org/10.1007/0-387-22746-6_2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 335, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 335, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 335, + 481, + 369 + ], + "type": "text", + "content": "27. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: CVPR (June 2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 369, + 481, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 481, + 403 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 481, + 403 + ], + "type": "text", + "content": "28. Peng, L., Liu, F., Yu, Z., Yan, S., Deng, D., Yang, Z., Liu, H., Cai, D.: Lidar point cloud guided monocular 3d object detection. In: European Conference on Computer Vision. pp. 123-139. Springer (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 403, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 481, + 445 + ], + "type": "text", + "content": "29. Prisacariu, V.A., Segal, A.V., Reid, I.: Simultaneous monocular 2d segmentation, 3d pose recovery and 3d reconstruction. In: Lee, K.M., Matsushita, Y., Rehg, J.M., Hu, Z. (eds.) ACCV. pp. 593-606. Springer Berlin Heidelberg, Berlin, Heidelberg (2013)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 446, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 481, + 468 + ], + "type": "text", + "content": "30. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. arXiv preprint arXiv:1612.00593 (2016)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 468, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 490 + ], + "type": "text", + "content": "31. Qin, Z., Wang, J., Lu, Y.: Weakly supervised 3d object detection from point clouds. ACMMM (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 490, + 481, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 481, + 523 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 481, + 523 + ], + "type": "text", + "content": "32. Shi, S., Wang, Z., Shi, J., Wang, X., Li, H.: From points to parts: 3d object detection from point cloud with part-aware and part-aggregation network. IEEE transactions on pattern analysis and machine intelligence 43(8), 2647-2664 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 523, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 481, + 555 + ], + "type": "text", + "content": "33. Song, X., Wang, P., Zhou, D., Zhu, R., Guan, C., Dai, Y., Su, H., Li, H., Yang, R.: Apollocar3d: A large 3d car instance understanding benchmark for autonomous driving. In: CVPR. pp. 5452-5462 (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 555, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 481, + 621 + ], + "type": "text", + "content": "34. Sudre, C.H., Li, W., Vercauteren, T., Ourselin, S., Jorge Cardoso, M.: Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations. In: Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: Third International Workshop, DLMIA 2017, and 7th International Workshop, ML-CDS 2017, Held in Conjunction with MICCAI 2017, Quebec City, QC, Canada, September 14, Proceedings 3. pp. 240-248. Springer (2017)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 621, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 481, + 643 + ], + "type": "text", + "content": "35. Tang, Y.S., Lee, G.H.: Transferable semi-supervised 3d object detection from rgb-d data. In: ICCV. pp. 1931-1940 (2019)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "text", + "content": "36. Team, O.D.: Openpcdet: An open-source toolbox for 3d object detection from point clouds. https://github.com/open-mmlab/OpenPCDet (2020)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 209, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 209, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 209, + 100 + ], + "type": "text", + "content": "J. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 347 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "37. Wang, R., Yang, N., Stueckler, J., Cremers, D.: Directshape: Photometric alignment of shape priors for visual vehicle pose and shape estimation (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 138, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 138, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 138, + 480, + 171 + ], + "type": "text", + "content": "38. Wang, X., Zhu, Z., Xu, W., Zhang, Y., Wei, Y., Chi, X., Ye, Y., Du, D., Lu, J., Wang, X.: Openoccupancy: A large scale benchmark for surrounding semantic occupancy perception. arXiv preprint arXiv:2303.03991 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "type": "text", + "content": "39. Wei, Y., Su, S., Lu, J., Zhou, J.: Fgr: Frustum-aware geometric reasoning for weakly supervised 3d vehicle detection (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "text", + "content": "40. Wei, Y., Wei, Z., Rao, Y., Li, J., Zhou, J., Lu, J.: Lidar distillation: Bridging the beam-induced domain gap for 3d object detection. In: ECCV. pp. 179-195. Springer (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "text", + "content": "41. Wei, Y., Zhao, L., Zheng, W., Zhu, Z., Zhou, J., Lu, J.: Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving. arXiv preprint arXiv:2303.09551 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "text", + "content": "42. Yan, Y., Mao, Y., Li, B.: Second: Sparsely embedded convolutional detection. Sensors 18(10), 3337 (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 281, + 480, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 480, + 324 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 480, + 324 + ], + "type": "text", + "content": "43. Yang, H., He, T., Liu, J., Chen, H., Wu, B., Lin, B., He, X., Ouyang, W.: Gdmae: generative decoder for mae pre-training on lidar point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9403-9414 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 325, + 480, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 480, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 480, + 347 + ], + "type": "text", + "content": "44. Zakharov, S., Kehl, W., Bhargava, A., Gaidon, A.: Autolabeling 3d objects with differentiable rendering of sdf shape priors. In: CVPR (2020)" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 155, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 155, + 91, + 447, + 102 + ], + "type": "text", + "content": "Segment, Lift and Fit: Automatic 3D Shape Labeling from 2D Prompts" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Segment3D_ Learning Fine-Grained Class-Agnostic 3D Segmentation without Manual Labels/a98903ad-e70d-49e3-988e-6063b3741036_content_list.json b/2024/Segment3D_ Learning Fine-Grained Class-Agnostic 3D Segmentation without Manual Labels/a98903ad-e70d-49e3-988e-6063b3741036_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d71e07c043b121f6a05519b3e66bfe641aaa6230 --- /dev/null +++ b/2024/Segment3D_ Learning Fine-Grained Class-Agnostic 3D Segmentation without Manual Labels/a98903ad-e70d-49e3-988e-6063b3741036_content_list.json @@ -0,0 +1,1677 @@ +[ + { + "type": "text", + "text": "Segment3D: Learning Fine-Grained Class-Agnostic 3D Segmentation without Manual Labels", + "text_level": 1, + "bbox": [ + 297, + 140, + 704, + 205 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rui Huang $^{1}$ , Songyou Peng $^{2}$ , Ayça Takmaz $^{2}$ , Federico Tombari $^{3}$ , Marc Pollefeys $^{2,4}$ , Shiji Song $^{1}$ , Gao Huang $^{1,*}$ , and Francis Engelmann $^{2,3}$", + "bbox": [ + 217, + 233, + 782, + 265 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/60d9244e1dc0824a5af4c5e3ed84654cf14df2985f7b03550f194dbe0c75c040.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 270, + 362, + 362 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Tsinghua University $^{2}$ ETH Zurich $^{3}$ Google $^{4}$ Microsoft *Corr. Author https://segment3d.github.io", + "bbox": [ + 372, + 275, + 620, + 321 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4a65f75423c7b6cc674929c485768bfeadcf11d3c765b2ccad32071a2e91278f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 643, + 271, + 785, + 362 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/02a083dd30a7cdc2269f2c2a4655f07b03e7b64862df8f3540abe08103c970e7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 220, + 354, + 362, + 446 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/23deca315088794a2003b1809fb7c064cce75576b6ffdc223bb8ba918eafa4ec.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 362, + 325, + 495, + 414 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/977f807793e24d4383695bb2f8c5801da5a67eae6ec4df571485bde96d8d6b35.jpg", + "image_caption": [ + "Mask3D [46] trained on manual labels." + ], + "image_footnote": [], + "bbox": [ + 362, + 412, + 495, + 500 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ca2a3ddc14d4f8421e7737521a083b1df213c7e8ac3a1d7135b381ab047c1ff6.jpg", + "image_caption": [ + "Segment3D (Ours) trained on automatic labels." + ], + "image_footnote": [], + "bbox": [ + 500, + 325, + 647, + 500 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/53e1e0c7a9520f86a8a02c5e3eaabc759f92bbcb36f5119192bff79bb450aea8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 357, + 784, + 446 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/aabda92570de3afea7e8a391da2c6c58414d59976b9338cb44aa8c9538da057f.jpg", + "image_caption": [ + "Fig. 1: Fine-Grained Class-Agnostic 3D Point Cloud Segmentation. Segment3D predicts highly accurate segmentation masks (right), improves over state-of-the-art 3D segmentation methods (e.g., Mask3D [46], left), and does not require manually labeled 3D training data. This is achieved through the automatic generation of high-quality training masks using foundation models for image segmentation [28]." + ], + "image_footnote": [], + "bbox": [ + 645, + 441, + 785, + 546 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Current 3D scene segmentation methods are heavily dependent on manually annotated 3D training datasets. Such manual annotations are labor-intensive, and often lack fine-grained details. Furthermore, models trained on this data typically struggle to recognize object classes beyond the annotated training classes, i.e., they do not generalize well to unseen domains and require additional domain-specific annotations. In contrast, recent 2D foundation models have demonstrated strong generalization and impressive zero-shot abilities, inspiring us to incorporate these characteristics from 2D models into 3D models. Therefore, we explore the use of image segmentation foundation models to automatically generate high-quality training labels for 3D segmentation models. The resulting model, Segment3D, generalizes significantly better than the models trained on costly manual 3D labels and enables easily adding new training data to further boost the segmentation performance.", + "bbox": [ + 259, + 645, + 740, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 186, + 375, + 202 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose Segment3D, a method for fine-grained class-agnostic 3D segmentation. In particular, dividing the space into coherent segments aligned with both the scene geometry and its semantics is a key challenge. This ability to accurately segment and interpret 3D scenes is fundamental for numerous downstream tasks [52,61], intelligent assistants and autonomous robots [30,64].", + "bbox": [ + 212, + 217, + 785, + 292 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Current methods for 3D indoor-scene understanding mostly focus on semantic [35, 39, 40, 47, 53] and instance segmentation [9, 17, 31, 32, 46, 54]. These approaches, while effective on popular benchmarks [13, 15, 44, 50], have limitations stemming from their training. Primarily, they depend on extensive manually labeled 3D training sets that are both time-consuming and challenging to annotate. Importantly, their performance often deteriorates when applied to scenarios beyond their training data, limiting their effectiveness in diverse, real-world scenarios. This becomes particularly apparent under the recently emerging task of open-vocabulary 3D scene understanding [18, 27, 33, 38, 51] that aims to segment arbitrary user queries, which naturally go beyond the pre-defined set of training-set classes. Concurrently, the recent surge in foundation models, especially 2D vision-language models [25, 28, 41, 60], demonstrates remarkable potential. Trained on internet-scale data, these models exhibit an extraordinary ability to generalize, even in a zero-shot setting, to new and different input distributions. However, their application has been predominantly confined to 2D data. For instance, SAM [28] has shown impressive results in 2D image segmentation, but its applicability to 3D scene understanding remains largely under-explored. All these factors give rise to the research question:", + "bbox": [ + 212, + 294, + 785, + 565 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "How to leverage 2D foundation models for class-agnostic 3D scene segmentation without requiring manually labeled 3D data?", + "text_level": 1, + "bbox": [ + 261, + 574, + 740, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, SAM3D [57] has proposed a straightforward method that uses posed RGB-D images corresponding to a 3D scene. They predict segmentation masks for all input RGB-D images with SAM [28], and projected them into the 3D space. Next, through an iterative bottom-up process, the 3D masks of partial scenes are merged to derive the final 3D segmentation (see Fig. 2). However, variations in perspectives across frames can lead to conflicting segmentations, which introduces inconsistency during the merging process. Besides, the inference speed of SAM3D is slow, due to running SAM on the fly for a large number of RGB images and the cumbersome non-learned merging process. In this paper, we seek to circumvent both the slow speed and the issue of merging inconsistent segmentation across frames.", + "bbox": [ + 212, + 613, + 785, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Towards this goal, we introduce Segment3D, a clean and efficient approach that utilizes a 3D model to achieve class-agnostic, fine-grained 3D segmentation at speed. Segment3D employs a two-stage training approach that requires no hand-annotated labels. We first train our class-agnostic 3D segmentation model", + "bbox": [ + 212, + 779, + 785, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Huang et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Keywords: Class-agnostic 3D segmentation $\\cdot$ 3D Scene Understanding", + "bbox": [ + 261, + 147, + 740, + 161 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "with RGB-D data, which are transformed into partial RGB-D point clouds. The supervision is provided by the segmentation masks produced by SAM and projected into 3D. While SAM can automatically generate masks for entire images, sometimes the results may suffer from over-segmentation (see Fig. 4), which is not desired in our context. Therefore, we propose a Mask Generation Module to generate high-quality, complete masks of objects for a given RGB image. Since there is readily available large-scale RGB-D data, additional data can be effortlessly integrated to further enhance segmentation performance. This pretraining stage lays the groundwork for understanding the 3D structure from 2D annotations. However, as our ultimate objective is the segmentation of full 3D scenes, we must bridge the domain gap between partial point clouds and the more comprehensive 3D point clouds obtained from 3D scanners or reconstruction techniques [14,24]. To this end, in the second stage, we fine-tune the model on full 3D point clouds in a self-supervised manner, utilizing high-confidence mask predictions from the pre-trained model as training signal.", + "bbox": [ + 212, + 145, + 787, + 375 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We demonstrate strong performance in class-agnostic segmentation on ScanNet [13, 44] and the newly released ScanNet++ [58]. Though Segment3D is trained on the indoor dataset, transferring it directly to the outdoor nuScenes dataset [4] also yields surprisingly good results. Finally, we show the use of Segment3D for improving open-vocabulary 3D instance segmentation [51]. Overall, the contributions of this paper are as follows:", + "bbox": [ + 212, + 375, + 787, + 465 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce Segment3D, a novel approach and training strategy for fine-grained class-agnostic 3D point cloud segmentation without manually annotated labels.", + "- We propose a Mask Generation Module that utilizes the 2D foundation model to automatically generate high-quality, complete training masks.", + "- We show that Segment3D demonstrates strong generalization compared to a wide range of baselines, including fully supervised methods trained on carefully annotated datasets." + ], + "bbox": [ + 214, + 470, + 782, + 588 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 214, + 625, + 387, + 641 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Instance Segmentation. Current models have seen a significant development over the last years, from proposal-based [56, 59], over grouping-based [9, 26, 31, 54], to recent Transformer-based [32, 46] methods. In this work, we follow the currently best-performing Transformer-based paradigm. Despite impressive advancements, a shared limitation is the dependence on costly manual ground-truth annotations. Recently, there have been efforts to automate the annotation process by bundling state-of-the-art segmentation models [55], but they are still limited to pre-defined object classes. However, in the context of open-world 3D scene understanding, the importance of semantic classification for closed-set categories has diminished. Instead, in this work, we propose a general 3D segmentation method trained on automatically generated labels.", + "bbox": [ + 212, + 672, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Segment3D", + "bbox": [ + 651, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Foundation Models. Foundation models, particularly those that are multimodal [25,41], have revolutionized the field of AI by leveraging extensive image-text pre-training. These models can derive rich image representations guided by natural language descriptions, enabling a variety of downstream tasks [21,37,42]. Another line of work [6,36], based on self-supervised learning, employs image training and yields high-performance features directly applicable as inputs for linear classifiers. Segment Anything Model (SAM) [28] has recently advanced the performance of foundation models for image segmentation. SAM has undergone training on a diverse, high-quality dataset comprising more than 1 billion masks. This training equips SAM with the ability to generalize to novel object types and images, surpassing the scope of its observations during the training process. Additionally, SAM can generate high-quality, fine-grained masks. Our work leverages the power of SAM for pre-training a 3D segmentation model, which is later also used for generating supervision signals for fine-tuning on scene-level point clouds.", + "bbox": [ + 212, + 145, + 787, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Open-Vocabulary 3D Scene Understanding. Recently, there has been an increased interest in 3D open-vocabulary scene understanding. This new field utilizes the zero-shot recognition abilities of 2D vision-language models [25,41], enabling a more comprehensive understanding of diverse and previously unseen 3D environments [8,16,22,23,27,29,33,38,51,62,63]. PLA [16] aligns point cloud features with captions extracted from multi-view images of a scene to enable open-vocabulary recognition. OpenScene [38] distills per-pixel image features to 3D point clouds, generating point-wise scene representations co-embedded with text and image pixels in CLIP feature space. However, it mainly focuses on semantic segmentation and exhibits a limited understanding of object instances. To this end, OpenMask3D [51] predicts class-agnostic 3D instance masks and aggregates per-mask features via multi-view fusion of CLIP-based image embeddings. Nevertheless, the segmentation model it uses is trained on closed-set labels, lacking generalization to the open world. Our method contributes to this area by providing open-set class-agnostic masks, which can serve as foundational inputs for models such as OpenMask3D.", + "bbox": [ + 212, + 378, + 787, + 623 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 642, + 330, + 659 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We aim to develop a method capable of segmenting any object within a given 3D scene. Relying on existing 3D training datasets cannot accomplish this goal, as their mask annotations are limited to a predefined set of object classes [1,2,13,48]. Consequently, models trained on such datasets fail to generalize effectively to new, unseen classes. Recently, with SAM [28] showing extraordinary generalization ability in 2D segmentation, SAM3D [57] has proposed to simply project and merge SAM predictions in 3D from posed RGB-D images, as illustrated in Fig. 2. However, it suffers from inconsistent results and slow inference speed, limiting its practical applicability.", + "bbox": [ + 212, + 672, + 787, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To fully exploit SAM for consistent 3D segmentation at speed, we introduce a clean and efficient method which utilizes a 3D model to directly segment the", + "bbox": [ + 214, + 809, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Huang et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f86f7851e28f780ecb81a3e158d663ff50c97018776cce971b8d5f1e6a05a28e.jpg", + "image_caption": [ + "Fig. 2: SAM3D vs. Segment3D. SAM3D [57] (left) merges 2D segmentation masks of RGB-D images generated by SAM to obtain 3D segmentation of entire scenes. However, conflicting segmentations across frames introduce inconsistency during the merging process. Moreover, it is slow because of extensive image inference and the cumbersome merging procedure. Instead, our Segment3D (right) utilizes a 3D model to directly segment entire 3D scenes which is clean and efficient." + ], + "image_footnote": [], + "bbox": [ + 287, + 147, + 715, + 315 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "entire scene. First, we propose a Mask Generation Module to generate high-quality, complete object masks from SAM predictions (Sec. 3.1). Next, we train our 3D segmentation model with RGB-D data and the generated object masks, which are transformed into partial point clouds (Sec. 3.2). Since the final goal is to segment 3D scenes, we need to bridge the inevitable domain gap between partial RGB-D point clouds and full 3D point clouds from 3D scanners or reconstruction methods. We therefore fine-tune our model on full 3D point clouds using high-confident mask predictions from our pre-trained model (Sec. 3.3). The overall framework is illustrated in Fig. 3.", + "bbox": [ + 212, + 426, + 787, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Generating Pseudo Ground-Truth Masks", + "text_level": 1, + "bbox": [ + 214, + 583, + 602, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Review on SAM. We first review the automatic mask generation pipeline of SAM [28]. SAM is prompted with a regular grid of $32 \\times 32$ points on the RGB image. Each point predicts masks at three different granularities: whole, part, and subpart. Next, all three masks for all points are filtered by the model's predicted IoU score as well as the stability of the masks with respect to the binarization threshold. Then, the remaining masks undergo Non-Maximum Suppression (NMS). Finally, if a pixel is encompassed by multiple masks, the mask ID with the highest predicted IoU is assigned to the pixel.", + "bbox": [ + 212, + 613, + 787, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, as shown in Fig. 4, the results obtained in this way could suffer from over-segmentation. A naive idea is to only use the masks generated by the \"whole\" branch. Nevertheless, we find that the \"whole\" branch typically generates high-quality masks for large objects while overlooking small objects in the image. In contrast, the \"part\" and \"subpart\" branches not only generate masks for parts of large objects but also predict masks for small and tiny objects. Thus, to enable our 3D model to achieve fine-grained segmentation capabilities for details and", + "bbox": [ + 212, + 734, + 787, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Segment3D", + "bbox": [ + 651, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a5671d4a17d0178f55d40f687cb4dd4ff3a744096b4a4ba325d0b3d4c7e53496.jpg", + "image_caption": [ + "Stage1: Pre-Training on Partial Point Clouds", + "Stage2: Fine-Tuning on Full Point Clouds", + "Fig. 3: Method Overview. Training Segment3D involves two stages: The first stage (left) relies on largely available RGB-D image sequences and SAM $\\square$ , a pre-trained foundation model for 2D image segmentation [28]. Segment3D $\\square$ is pre-trained on partial RGB-D point clouds and supervised with pseudo ground-truth masks generated by Mask Generation Module (MGM) $\\square$ . Due to the domain gap between partial and full point clouds, in the second stage (right), Segment3D $\\square$ is fine-tuned with confident masks predicted by the pre-trained Segment3D $\\square$ ." + ], + "image_footnote": [], + "bbox": [ + 218, + 160, + 781, + 286 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "small objects, we must retain the outputs from the \"part\" and \"subpart\" branches. The problem is how to preserve their segmentation for small and tiny objects while removing their segmentation for parts of large objects.", + "bbox": [ + 214, + 414, + 784, + 460 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Mask Generation Module (MGM). Here we introduce our Mask Generation Module (MGM). Since SAM's automatic mask generation pipeline can provide reasonably high-quality masks for a given image, we start based on it. Next, we compare the result with the output of the \"whole\" branch to determine which masks are parts of large objects and remove them. Specifically, for each mask in the automatic output $m_i^{\\mathrm{auto}}$ , we iterate through the masks of the entire branch, calculating the intersection $m_{ij}^{\\mathrm{inter}}$ between $m_i^{\\mathrm{auto}}$ and $m_j^{\\mathrm{whole}}$ . If $m_{ij}^{\\mathrm{inter}} / m_i^{\\mathrm{auto}} \\geq t$ , then we consider $m_i^{\\mathrm{auto}}$ to be a part of $m_j^{\\mathrm{whole}}$ . Our preliminary results show that the process of extracting valid masks is highly robust to the hyper-parameter $t$ and we set it as 0.9 by default. We remove the part masks and combine the remaining masks with the output of the \"whole\" branch, which will serve as our training supervision. With MGM, we observe notable improvements at all object mask sizes, especially on medium-sized and large objects (see Table 6).", + "bbox": [ + 214, + 465, + 787, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Stage 1: Pre-Training on RGB-D Point Clouds", + "text_level": 1, + "bbox": [ + 214, + 690, + 645, + 705 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In contrast to the relatively scarce availability of 3D data, there is an abundance of 2D data, particularly RGB-D images, which are readily accessible. For example, ScanNet [13] comprises merely 1513 3D scans, compared to the substantially larger collection of 2.5 million RGB-D images. Therefore, we first pre-train our 3D segmentation model on partial RGB-D point clouds.", + "bbox": [ + 212, + 712, + 787, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training-Set Preparation. Starting from a collection of RGB-D frames, we create partial 3D point clouds and their corresponding pseudo ground-truth 3D masks. Note that the labels can be automatically obtained without manual effort.", + "bbox": [ + 214, + 794, + 787, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Huang et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e931db91a5ab2cba7dfd94ce873d48d635125c3f64094c8e679b9324cdb26036.jpg", + "image_caption": [ + "Automatic-SAM Generated Masks" + ], + "image_footnote": [], + "bbox": [ + 223, + 171, + 346, + 239 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/69a033e5891fced37f67f592bb79ceaeeb2344d035c1e81b3881aa49f045ca13.jpg", + "image_caption": [ + "Fig. 4: Automatic-SAM Generated (left) vs. MGM Generated (right) Masks from ScanNet [13]. The former exhibits over-segmentation, while the latter does not." + ], + "image_footnote": [], + "bbox": [ + 223, + 242, + 344, + 310 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d4e3faaf194ed016d08c2395d4390adc664698488d391d639d8560861534abcd.jpg", + "image_caption": [ + "MGM rated Masks" + ], + "image_footnote": [], + "bbox": [ + 361, + 171, + 483, + 239 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f143a5be6e9ef1ffc065014faf122eb7e7f09aa654e81953cea21d48573234ec.jpg", + "image_caption": [ + "Fig. 5: Partial (center) vs. Full (right) Point Clouds from ScanNet [13]. Full point clouds are more complete and exhibit fewer occlusions due to reconstruction over multiple viewpoints." + ], + "image_footnote": [], + "bbox": [ + 361, + 242, + 483, + 310 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4da37cbae5a0269038b2b95a3af6fe4c3a9130f6185a19847f5a6f98d73d0ec6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 511, + 148, + 784, + 327 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For each frame in a large RGB-D dataset, we first transform the 2D depth map to a partial 3D point cloud. To do so, we need to know the intrinsic matrix $\\mathbf{K} \\in \\mathbb{R}^{3 \\times 3}$ and extrinsic matrix $\\mathbf{T} = [\\mathbf{R} \\mathbf{t}] \\in \\mathbb{R}^{3 \\times 4}$ . For each pixel $\\mathbf{p} = (u, v)$ , we can transform it with its depth value $D_{\\mathbf{p}}$ into a 3D point $\\mathbf{P}$ in world coordinates as follows:", + "bbox": [ + 212, + 430, + 784, + 503 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} = \\mathbf {R} ^ {\\top} \\cdot \\left(D _ {\\mathbf {p}} \\cdot \\mathbf {K} ^ {- 1} \\cdot \\tilde {\\mathbf {p}}\\right) - \\mathbf {R} ^ {\\top} \\mathbf {t}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 510, + 784, + 530 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\tilde{\\mathbf{p}}$ is the homogeneous coordinate of $\\mathbf{p}$ . By applying Eq. (1) to all pixels in the depth map $D$ and associating their per-pixel RGB value, we obtain the input partial 3D point cloud. Next, we obtain the pseudo ground-truth 3D segmentation masks for this point cloud with the 2D masks generated by MGM in the same way. Since we know the one-to-one mapping between 2D pixel and 3D points from Eq. (1), we directly obtain the per-point 3D mask labels.", + "bbox": [ + 212, + 541, + 787, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Model Architecture. We use a model inspired by Mask3D [46] to train a class-agnostic 3D segmentation model. The model comprises a sparse convolutional backbone derived from MinkowskiUNet [12] and a transformer decoder, as in MaskFormer [10, 11]. We adopt a set of queries to represent the masks, each of which is initialized with a positional embedding. Specifically, we select query positions with furthest point sampling (FPS) and use their Fourier positional encodings as the query embeddings. Leveraging the transformer decoder, all the mask queries are refined by progressively attending to point cloud features across multiple scales in parallel. Each mask query is subsequently decoded into both a mask feature and a binary label to predict whether the given query corresponds to a valid object or not. By computing cosine similarity scores between a mask feature and all point features within the point cloud, a heatmap is generated over", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Segment3D", + "bbox": [ + 651, + 114, + 730, + 127 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "at the point cloud. This heatmap is input to a sigmoid function, and thresholded at 0.5, resulting in the final binary mask.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Training with SAM Generated Masks. We supervise the model with two losses: the per-point supervision loss $\\mathcal{L}_{\\mathrm{mask}}$ and a per-query supervision loss $\\mathcal{L}_{\\mathrm{obj}}$ . The loss $\\mathcal{L}_{\\mathrm{mask}}$ enables learning a foreground-background segmentation for each mask, and is composed of a dice loss $\\mathcal{L}_{\\mathrm{dice}}$ [34] and a binary cross-entropy loss $\\mathcal{L}_{\\mathrm{ce}}$ for each point. The $\\mathcal{L}_{\\mathrm{obj}}$ is a binary classification loss that indicates whether a query represents a valid \"object\" or \"no object\". This mechanism allows for the prediction of a variable number of masks, depending on the underlying scene content and geometry. Following prior work [5, 10, 46], we first adopt bipartite graph matching to establish correspondences between the set of predicted masks and the set of target masks provided by MGM as described before. If the predicted instance finds a matching target mask, then we assign it an \"object\" label; conversely, if there is no match, we assign \"no object\". In summary, we optimize the following losses:", + "bbox": [ + 212, + 183, + 787, + 378 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {m a s k}} + \\lambda_ {\\text {o b j}} \\mathcal {L} _ {\\text {o b j}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 421, + 380, + 784, + 395 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {m a s k}} = \\lambda_ {\\text {d i c e}} \\mathcal {L} _ {\\text {d i c e}} + \\lambda_ {\\text {c e}} \\mathcal {L} _ {\\text {c e}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 401, + 398, + 784, + 414 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\lambda_{*}$ are hyperparameters that balance the contribution of each component in the loss. The binary classification loss $\\mathcal{L}_{\\mathrm{obj}}$ is applied to all queries, while the mask loss $\\mathcal{L}_{\\mathrm{mask}}$ is specifically applied to masks labeled as \"object\".", + "bbox": [ + 212, + 420, + 782, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 Stage 2: Self-Supervised Scene Fine-Tuning", + "text_level": 1, + "bbox": [ + 214, + 489, + 619, + 505 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "After Stage 1, we obtain a class-agnostic 3D segmentation model by pre-training solely on RGB-D images and automatically generated labels from MGM. However, a fundamental domain gap persists between partial point clouds derived from RGB-D data and full point clouds acquired through 3D scanners or reconstruction methods [14,24] (See Fig. 5). This gap exists mostly because of object occlusions, but also due to challenges of depth cameras to capture dark or reflective surfaces from a single viewpoint. Hence, depending solely on RGB-D frames for training a 3D segmentation model intended for full 3D scans proves inadequate. Therefore, we propose to further fine-tune our model on scene-level full 3D point clouds. The key idea to obtain 3D mask annotations for training on full point clouds is to use selected, high-confidence masks generated by the pre-trained model itself. Note that this stage requires no manual labels on full 3D scenes and proves essential for the performance of Segment3D (see Tab. 5).", + "bbox": [ + 212, + 516, + 787, + 712 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Confidence-Based Mask Generation. Next, we outline the process of generating the supervision signal for the fine-tuning stage. The pre-trained model processes point clouds independently of their nature, be it partial or full. Therefore, when presented with a full 3D point cloud, the pre-trained 3D model produces a set of masks, each with a binary classification (valid or not) and a heatmap over all points, just as in Sec. 3.2. To assess the quality of the predicted masks, we compute a confidence score based on the confidence map $\\sigma(\\mathbf{h})$ , where $\\mathbf{h}$ is the predicted heatmap and $\\sigma$ is the sigmoid function. We then compute the average", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Huang et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "confidence of those points for which $\\sigma(\\mathbf{h}) > 0.5$ as the confidence score of the predicted mask, denoted as $c_{\\mathrm{mask}}$ . We also consider the classification as a valid object and use the probability from the binary classification assigned to the \"object\" category as the confidence score, denoted as $c_{\\mathrm{obj}}$ . The final confidence score for each predicted mask is then the product of the two scores, $c = c_{\\mathrm{mask}} \\cdot c_{\\mathrm{obj}}$ . For fine-tuning our 3D segmentation model in Stage 2, we select the most confidently predicted masks above a threshold $\\tau_c$ .", + "bbox": [ + 218, + 145, + 787, + 252 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Training with the High-Confidence Generated Masks. For fine-tuning, we follow the same procedure as before and use $\\mathcal{L}_{\\mathrm{mask}}$ as defined in Eq. 3. In contrast to the pre-training stage, the binary classification loss $\\mathcal{L}_{\\mathrm{obj}}$ , responsible for categorizing queries into valid or invalid, is omitted. Since we only select masks with high confidence for supervision, it can happen that some objects in the scene have no assigned ground truth mask. In such instances, deeming a correctly predicted mask for those objects as invalid would be detrimental. Table 5 illustrates the efficacy of the self-supervised fine-tuning process in comparison to pre-training alone.", + "bbox": [ + 218, + 258, + 787, + 393 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 218, + 417, + 374, + 434 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We firstly evaluate Segment3D on indoor ScanNet++ [58], ScanNet200 [13,44] and outdoor nuScenes [4] in a class-agnostic segmentation setting (Sec. 4.1). We then show the advantage of our method in segmenting small objects, and provide analysis to understand the importance of fine-tuning, the effectiveness of Mask Generation Module and the potential of training on more data (Sec. 4.2). Finally, we demonstrate its application for the task of open-set 3D instance segmentation as proposed in OpenMask3D [51] (Sec. 4.3).", + "bbox": [ + 218, + 446, + 787, + 553 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Comparing with State-of-the-Art Methods", + "text_level": 1, + "bbox": [ + 218, + 575, + 612, + 590 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Datasets. We first evaluate on the recently released ScanNet++ [58] which includes high-resolution 3D scans captured at sub-millimeter precision, and fine-grained annotations covering objects of varying sizes. Due to its comprehensive data annotation, all of our analytical experiments in Sec. 4.2 are also based on it. Next, we adopt the ScanNet200 [13, 44], which builds upon the classic ScanNet [13] dataset and extends the semantic annotations to 200 classes. Then, we test our model on the nuScenes [4] outdoor dataset. For training our model, we employ ScanNet [13, 44], which is collected through a lightweight RGB-D scanning process. See the appendix for the implementation details of our method.", + "bbox": [ + 218, + 606, + 787, + 742 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Methods in Comparison. We compare with a wide range of prior art methods from different categories. Mask3D [46] is a state-of-the-art fully-supervised method trained on manually annotated 3D segmentation masks. Segment3D has the same backbone as Mask3D but instead of training on manually annotated 3D masks, it learns from automatically generated masks. Felzenszwalb et al. [20] proposed a graph-based method for segmentation which operates directly on", + "bbox": [ + 218, + 748, + 787, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Segment3D", + "bbox": [ + 651, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/ef572c3db0b06ab09f19eeac9ea547efee88bdce58a0b1824ac46389914c5a78.jpg", + "table_caption": [ + "Table 1: Segmentation Scores on ScanNet++ [58]. The metric is average precision (AP) on the validation split. We report scores with and without post-processing (more details in Sec. 4.1). For reference, we show results on Mask3D trained on manual ScanNet200 labels." + ], + "table_footnote": [], + "table_body": "
ModelAvg. Inference Time/swithout post-processingwith post-processing
AP\\( AP_{50} \\)\\( AP_{25} \\)AP\\( AP_{50} \\)\\( AP_{25} \\)
Mask3D [46]0.78.715.527.214.321.329.9
SAM3D [57]386.73.99.322.18.416.130.0
Felzenszwalb et al. [20]12.65.811.627.2---
Segment3D (Ours)0.713.023.838.320.230.942.7
", + "bbox": [ + 232, + 205, + 761, + 309 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "the 3D geometry. UnScene3D [45] leverages self-supervised color and geometry features to obtain coarse segmentation and refines it through self-training. SAM3D [57] merges the segmentation masks of RGB-D images generated by SAM [28] to obtain the 3D segmentation result.", + "bbox": [ + 212, + 339, + 782, + 400 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Metrics. We report average precision (AP) scores at IoU thresholds of $25\\%$ , $50\\%$ , and averaged over the range [0.5:0.95:05] between predicted and ground truth masks. Consistent with common practices in the field [47, 49, 54], we also report scores after post-processing. This involves smoothing the predicted masks through graph-based oversegmentation [20], and splitting distant parts of the same mask via connected component clustering DBSCAN [19]. We also report the average inference time for these methods on ScanNet++ tested on A100.", + "bbox": [ + 212, + 407, + 784, + 513 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Results on ScanNet++. Scores are reported in Table 1. Segment3D outperforms all previous methods by at least +4.3 AP and up to +12.7 AP $_{25}$ . Notably, we achieve such improvements without any ground-truth mask annotation contrary to Mask3D trained on ScanNet200. In general, the performance of Mask3D (and other fully-supervised methods) depends on the quality of the annotated training dataset; often the manual annotation of small objects (pens, cellphones) or other fine-details is challenging. Instead, Segment3D relies on automatically generated high-quality masks from MGM using the 2D foundation model, which can capture fine-grained details without human annotation effort. Table 4 highlights that Segment3D excels particularly in predicting the more challenging small object masks. Furthermore, our method significantly outperforms SAM3D, indicating that Segment3D can circumvent the noise that arises during the merging process of SAM3D. Additionally, As shown in Table 1, Segment3D is three-order-of-magnitude faster than SAM3D in inference speed.", + "bbox": [ + 212, + 521, + 787, + 733 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Fig. 6 shows segmentation results of several representative examples on ScanNet++. As can be seen, the scenes are quite diverse, presenting multiple challenges such as clutter and a wide range of mask sizes. Despite these challenges, our model predicts quite accurate and well-localized segmentation masks. For example, compared to Mask3D, our method is able to segment the finer-grained objects on top of the bed and the shelf. The masks of the computer screen and the chair in front of it are also less fragmented than predictions of the baselines.", + "bbox": [ + 212, + 734, + 787, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Huang et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/70a7a84c1b43dda22c98994091f8e91e916973c390047cff985be18d92af8d34.jpg", + "table_caption": [ + "Table 2: Segmentation Scores on ScanNet200 [13, 44]. The evaluation metric is average precision on the val split." + ], + "table_footnote": [], + "table_body": "
ModelAPAP50
Mask3D [46]34.143.1
Felzenszwalb et al. [20]6.112.1
UnScene3D [45]15.932.2
SAM3D [57]19.032.5
Segment3D (Ours)27.739.8
", + "bbox": [ + 238, + 200, + 460, + 295 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/34ab496a9b18383d234278effd7a94b97b0d6901e6c09d2dccb7be15d3e10149.jpg", + "table_caption": [ + "Table 3: Segmentation Scores on nuScenes [4]. The evaluation metric is average precision on the val split." + ], + "table_footnote": [], + "table_body": "
ModelGT LabelsAP50AP25
zero-shot test
Mask3D×15.325.5
Segment3D×37.848.0
fine-tuned on nuScenes
Mask3D✓46.557.8
Segment3D×42.555.5
", + "bbox": [ + 524, + 198, + 763, + 297 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Results on ScanNet200. As illustrated in Table 2, Segment3D outperforms all baselines that require no manual 3D labels, and even shows competitive results with Mask3D, which is trained on ScanNet200. Compared to Unscene3D, which leverages self-supervised color and geometry features to obtain coarse segmentation, we achieve superior results by using the Mask Generation Module to generate high-quality complete masks.", + "bbox": [ + 212, + 324, + 782, + 416 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Results on Outdoor Data. The performance on nuScenes is demonstrated in Table 3. We observe that even though Segment3D is trained using purely indoor data, applying it directly to outdoor data produces surprisingly good results. This indicates the stronger generalizability of Segment3D over the fully-supervised method Mask3D when transferring from indoor to outdoor scenes. Furthermore, we report scores with fine-tuning on nuScenes which leads to overall improved results. Similar to Table 2, Mask3D performs best, however, Segment3D achieves very competitive performance without any hand-labeled training data. Refer to the appendix for additional results on STPLS3D [7] and Paris-Lille-3D [43] datasets.", + "bbox": [ + 212, + 421, + 787, + 571 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Analysis Experiments", + "text_level": 1, + "bbox": [ + 215, + 594, + 441, + 609 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Performance on Different Mask Sizes. We proceed with an analysis of the performance of our Segment3D and Mask3D (the best-performing baseline) on object masks of various sizes. We categorize the size of masks based on the number of points they contain (small: [0, 2k], medium: [2k, 15k] large: [15k, $\\infty$ ]). The results are reported in Table 4. Our method yields significantly improved", + "bbox": [ + 212, + 625, + 787, + 702 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/67b15ec73e95f4e75d9603feeee909df272bf19bb8dfb6142c5b9942b3ff3d7a.jpg", + "table_caption": [ + "Table 4: Segmentation Scores on Different Mask Sizes. Segmented3D improves over Mask3D, especially on small object masks. Details are in Sec. 4.2." + ], + "table_footnote": [], + "table_body": "
Mask SizeLargeMediumSmall
ScanNet++APAP50APAP50APAP50
Mask3D13.825.113.623.31.84.3
Segment3D18.331.916.629.88.417.3
(+4.5)(+6.8)(+3.0)(+6.5)(+6.6)(+13.0)
", + "bbox": [ + 279, + 758, + 727, + 837 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Segment3D", + "bbox": [ + 651, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/4a6319bc0d817b372e927c3ebd7431955843bfdfb283c6eb2c5ebf05c64311c1.jpg", + "image_caption": [ + "Fig. 6: Qualitative Results on ScanNet++. From top to bottom, we show the colored input 3D scenes, the segmentation masks predicted by SAM3D [57], Mask3D [46], our Segment3D and the ground truth 3D mask annotations." + ], + "image_footnote": [], + "bbox": [ + 232, + 142, + 777, + 542 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "segmentation results on objects of various sizes, especially on small objects. This confirms our intuition that Mask3D performs poorly on small-sized object masks as those are typically harder to manually annotate. In contrast, Segment3D utilizes masks generated by MGM with the foundation model as supervision, capturing fine-grained scene details. This showcases the usefulness of foundation models and raises the question if manually labeled large-scale 3D datasets are necessary for training 3D point cloud segmentation models.", + "bbox": [ + 212, + 632, + 787, + 739 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The Importance of Two-Stage Training. Next, we compare the performance of Segment3D pre-trained solely on partial RGB-D point clouds (Stage 1) and with additional fine-tuning on full point clouds (Stage 2). Scores are reported in Table 5. The additional fine-tuning stage almost doubles the segmentation performance of our model on the most challenging AP metric. By training with the predicted high-confidence masks, Segment3D effectively reduces the inherent", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Huang et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/c5d921e880bc0eb545d040fc90cf581b581d268c7b7eca14dbf1d22db4ac190e.jpg", + "table_caption": [ + "Table 5: The Importance of Two-Stage Training. Fine-tuning on full point clouds supervised by confident predicted 3D masks significantly surpasses pre-training on projected 2D SAM masks alone." + ], + "table_footnote": [], + "table_body": "
Training StagesAPAP50AP25
Pre-Training (Stage 1)6.814.230.6
+ Fine-Tuning (Stage 2)13.0 (+6.2)23.8 (+9.6)38.3 (+7.7)
", + "bbox": [ + 303, + 191, + 694, + 246 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/fa69ff3d223ec6460dff8da5433237344844bef36b54fff7b5cc1bfe320accb2.jpg", + "table_caption": [ + "Table 6: Effectiveness of Mask Generation Module. Compared to the model trained with masks generated by Automatic-SAM, our model shows improvement across all object mask sizes, particularly for medium-sized and large objects." + ], + "table_footnote": [], + "table_body": "
Mask SizeLargeMediumSmall
Masks Generated byAPAP50APAP50APAP50
Automatic-SAM16.830.614.227.27.516.2
MGM18.331.916.629.88.417.3
(+1.5)(+1.3)(+2.4)(+2.6)(+0.9)(+1.1)
", + "bbox": [ + 274, + 306, + 718, + 393 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "domain gap between the partial point clouds derived from RGB-D images and full 3D point clouds.", + "bbox": [ + 212, + 420, + 784, + 450 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The Effectiveness of Mask Generation Module. To verify whether the MGM enhances the model's performance by reducing the over-segmentation issue and providing more complete masks, we conducted evaluations on objects of different sizes. As shown in Table 6, we can see that compared to the model trained with masks generated by Automatic-SAM, our model notably improves the performance of large and medium-sized objects which demonstrates the effectiveness of MGM.", + "bbox": [ + 212, + 455, + 787, + 561 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Pre-Training with Additional Data. Since RGB-D data is available in abundance and masks can be automatically generated, it is natural to ask if pretraining on additional masks will further improve the overall performance. To that end, we perform a first experiment where we select additional frames from the training set of the ScanNet++ dataset (to increase the variety of training data). Table 7 shows an impressive performance boost of +4.1 AP. For future work, considering that this improvement is obtained simply by adding more automatically generated masks to the pre-training, our approach seems promising to train on even more data. It could even be plausible to train on internet images combined with monocular depth estimation, such as ZoeDepth [3], to compute the partial point clouds.", + "bbox": [ + 212, + 569, + 787, + 734 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.3 Application: Open-Set Scene Understanding", + "text_level": 1, + "bbox": [ + 214, + 755, + 627, + 771 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "A real-world application of our class-agnostic 3D segmentation method is open-vocabulary 3D scene understanding, as implemented recently by OpenMask3D [51]. Given a 3D scene, a user can search for arbitrary objects via text prompts (see Fig. 7). A core component of OpenMask3D is Mask3D, which segments the", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Segment3D", + "bbox": [ + 651, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/b9fedc4e335787867d2b18f26983f43048f69596ac8e5bcccf80411ce605e54f.jpg", + "table_caption": [ + "Table 7: Results with Additional Training Data. The performance is further boosted with more training data." + ], + "table_footnote": [], + "table_body": "
Pre-Training DataAPAP50
ScanNet13.023.8
ScanNet, ScanNet++17.1 (+4.1)26.7 (+2.9)
", + "bbox": [ + 222, + 199, + 483, + 247 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/b5f76b5009d9e688053dbd0296afccc2e3702e5ccbb98c5f7622e1c7ba587879.jpg", + "table_caption": [ + "Table 8: Open-Set 3D Scene Understanding. The evaluation metric is average precision with an IoU threshold of $50\\%$ ." + ], + "table_footnote": [], + "table_body": "
SegmentorScanNet++Replica
Mask3D [46]15.018.0
Segment3D (Ours)18.5 (+3.5)18.4 (+0.4)
", + "bbox": [ + 508, + 199, + 772, + 247 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/4ae3e9804388c5cc2a75354f70b7afaaae5bf91bfc44cffa88e504391e2b8334.jpg", + "image_caption": [ + "\"aback eraser\"" + ], + "image_footnote": [], + "bbox": [ + 245, + 255, + 413, + 333 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/9e756c4fe06ba932d5c39e49f1b57229a930481421b91c7320fc831f7cbddd62.jpg", + "image_caption": [ + "\"paper on the laptop\"", + "Fig. 7: Open-Set 3D Object Retrieval Results. Given a text prompt (bottom), OpenMask3D [51] finds the corresponding object masks in a given 3D scene (top). We adapt OpenMask3D and use fine-grained masks from our Segment3D method. We show 3D reconstructions and RGB images for better visualization (top left corner)." + ], + "image_footnote": [], + "bbox": [ + 416, + 256, + 586, + 333 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/f35f83da57a01168c4ea50e9a8fbbf7999838c0387258d7e92ac5a0bb3f83d8c.jpg", + "image_caption": [ + "\"scissors\"" + ], + "image_footnote": [], + "bbox": [ + 589, + 256, + 756, + 333 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "scene into a set of object masks. Since Mask3D is trained on the closed-set of labeled annotations from ScanNet, its masks are not truly class-agnostic or open-vocabulary. We therefore replace Mask3D with our class-agnostic Segment3D. We evaluate on the closed-set labels of $\\mathrm{ScanNet}++$ and Replica in Table 8 and report an improvement of up to $+3.5\\mathrm{AP}_{50}$ . Since Segment3D can accurately generate fine-grained masks in a class-agnostic manner, we are able to retrieve objects such as \"eraser\", \"paper\" and \"scissors\", which are small in scale.", + "bbox": [ + 212, + 436, + 787, + 544 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 566, + 359, + 583 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We have presented Segment3D, a simple, yet powerful class-agnostic 3D segmentation model. Segment3D employs a two-stage training framework that requires no manually annotated labels. We propose a Mask Generation Module to automatically generate high-quality, complete training masks with the 2D foundation model. The model is first pre-trained on RGB-D data with predicted 2D masks, then fine-tuned on full point clouds. Segment3D shows strong generalization and even outperforms existing 3D segmentation models that rely on hand-labeled 3D training scenes. Indeed, this raises the question of whether hand-labeled 3D training datasets are as essential as they were once thought to be.", + "bbox": [ + 212, + 585, + 787, + 722 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement. Gao Huang is supported in part by the National Natural Science Foundation of China under grants (62321005 and 42327901). Francis Engelmann is partially supported by an ETH AI Center postdoctoral research fellowship and an ETH Zurich Career Seed Award. Songyou Peng is supported by an Innosuisse funding (100.567 IP-ICT), and Ayça Takmaz is supported by an Innosuisse grant (48727.1 IP-ICT).", + "bbox": [ + 212, + 727, + 787, + 819 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Huang et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 323, + 160 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Armeni, I., Sener, O., Zamir, A.R., Jiang, H., Brilakis, I., Fischer, M., Savarese, S.: 3D Semantic Parsing of Large-Scale Indoor Spaces. In: CVPR (2016) 4", + "2. Baruch, G., Chen, Z., Dehghan, A., Dimry, T., Feigin, Y., Fu, P., Gebauer, T., Joffe, B., Kurz, D., Schwartz, A., et al.: ARKitScenes: A Diverse Real-World Dataset for 3D Indoor Scene Understanding using Mobile RGB-D Data. arXiv preprint arXiv:2111.08897 (2021) 4", + "3. Bhat, S.F., Birkl, R., Wofk, D., Wonka, P., Müller, M.: ZoeDepth: Zero-Shot Transfer by Combining Relative and Metric Depth. arXiv preprint arXiv:2302.12288 (2023) 13", + "4. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuScenes: A Multimodal Dataset for Autonomous Driving. In: CVPR (2020) 3, 9, 11", + "5. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-End Object Detection with Transformers. In: ECCV (2020) 8", + "6. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging Properties in Self-Supervised Vision Transformers. In: ICCV (2021) 4", + "7. Chen, M., Hu, Q., Yu, Z., Thomas, H., Feng, A., Hou, Y., McCullough, K., Ren, F., Soibelman, L.: STPLS3D: A Large-Scale Synthetic and Real Aerial Photogrammetry 3D Point Cloud Dataset. arXiv preprint arXiv:2203.09065 (2022) 11", + "8. Chen, R., Liu, Y., Kong, L., Zhu, X., Ma, Y., Li, Y., Hou, Y., Qiao, Y., Wang, W.: CLIP2Scene: Towards Label-Efficient 3D Scene Understanding by CLIP. In: CVPR (2023) 4", + "9. Chen, S., Fang, J., Zhang, Q., Liu, W., Wang, X.: Hierarchical Aggregation for 3D Instance Segmentation. In: ICCV (2021) 2, 3", + "0. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-Attention Mask Transformer for Universal Image Segmentation. In: CVPR (2022) 7, 8", + "1. Cheng, B., Schwing, A., Kirillov, A.: Per-Pixel Classification is Not All You Need for Semantic Segmentation. In: NeurIPS (2021) 7", + "2. Choy, C., Gwak, J., Savarese, S.: 4D Spatio-Temporal ConvNets: Minkowski Convolutional Neural Networks. In: CVPR (2019) 7", + "3. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: ScanNet: Richly-Annotated 3D Reconstructions of Indoor Scenes. In: CVPR (2017) 2, 3, 4, 6, 7, 9, 11", + "4. Dai, A., Nießner, M., Zollhöfer, M., Izadi, S., Theobalt, C.: BundleFusion: Real-time Globally Consistent 3D Reconstruction using On-the-fly Surface Re-integration. TOG (2017) 3, 8", + "5. Delitzas, A., Takmaz, A., Tombari, F., Sumner, R., Pollefeys, M., Engelmann, F.: SceneFun3D: Fine-Grained Functionality and Affordance Understanding in 3D Scenes. In: CVPR (2024) 2", + "6. Ding, R., Yang, J., Xue, C., Zhang, W., Bai, S., Qi, X.: PLA: Language-Driven Open-Vocabulary 3D Scene Understanding. In: CVPR (2023) 4", + "7. Engelmann, F., Bokeloh, M., Fathi, A., Leibe, B., Nießner, M.: 3D-MPA: Multi-Proposal Aggregation for 3D Semantic Instance Segmentation. In: CVPR (2020) 2", + "8. Engelmann, F., Manhardt, F., Niemeyer, M., Tateno, K., Tombari, F.: OpenNeRF: Open Set 3D Neural Scene Segmentation with Pixel-Wise Features and Rendered Novel Views. In: ICLR (2024) 2" + ], + "bbox": [ + 225, + 176, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Segment3D", + "bbox": [ + 651, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "19. Ester, M., Kriegel, H.P., Sander, J., Xu, X., et al.: A Density-Based Algorithm for Discovering Clusters in Large Spatial Databases with Noise. In: KDD (1996) 10", + "20. Felzenszwalb, P.F., Huttenlocher, D.P.: Efficient Graph-Based Image Segmentation. IJCV (2004) 9, 10, 11", + "21. Gu, X., Lin, T.Y., Kuo, W., Cui, Y.: Open-Vocabulary Object Detection via Vision and Language Knowledge Distillation. In: ICLR (2022) 4", + "22. Ha, H., Song, S.: Semantic Abstraction: Open-World 3D Scene Understanding from 2D Vision-Language Models. In: CoRL (2022) 4", + "23. Huang, T., Dong, B., Yang, Y., Huang, X., Lau, R.W., Ouyang, W., Zuo, W.: CLIP2Point: Transfer CLIP to Point Cloud Classification with Image-Depth Pre-Training. In: ICCV (2023) 4", + "24. Izadi, S., Kim, D., Hilliges, O., Molyneaux, D., Newcombe, R., Kohli, P., Shotton, J., Hodges, S., Freeman, D., Davison, A., et al.: KinectFusion: Real-time 3D Reconstruction and Interaction using a Moving Depth Camera. In: UIST (2011) 3, 8", + "25. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up Visual and Vision-Language Representation Learning with Noisy Text Supervision. In: ICML (2021) 2, 4", + "26. Jiang, L., Zhao, H., Shi, S., Liu, S., Fu, C.W., Jia, J.: PointGroup: Dual-Set Point Grouping for 3D Instance Segmentation. In: CVPR (2020) 3", + "27. Kerr, J., Kim, C.M., Goldberg, K., Kanazawa, A., Tancik, M.: LERF: Language Embedded Radiance Fields. In: ICCV (2023) 2, 4", + "28. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment Anything. In: ICCV (2023) 1, 2, 4, 5, 6, 10", + "29. Kobayashi, S., Matsumoto, E., Sitzmann, V.: Decomposing Nerf for Editing via Feature Field Distillation. In: NeurIPS (2022) 4", + "30. Lemke, O., Bauer, Z., Zurbrügg, R., Pollefeys, M., Engelmann, F., Blum, H.: Spotcompose: A framework for open-vocabulary object retrieval and drawer manipulation in point clouds. In: 2nd Workshop on Mobile Manipulation and Embodied Intelligence at ICRA 2024 (2024) 2", + "31. Liang, Z., Li, Z., Xu, S., Tan, M., Jia, K.: Instance Segmentation in 3D Scenes using Semantic Superpoint Tree Networks. In: CVPR (2021) 2, 3", + "32. Lu, J., Deng, J., Wang, C., He, J., Zhang, T.: Query Refinement Transformer for 3D Instance Segmentation. In: ICCV (2023) 2, 3", + "33. Lu, Y., Xu, C., Wei, X., Xie, X., Tomizuka, M., Keutzer, K., Zhang, S.: Open-Vocabulary Point-Cloud Object Detection without 3D Annotation. In: CVPR (2023) 2, 4", + "34. Milletari, F., Navab, N., Ahmadi, S.A.: V-Net: Fully Convolutional Neural Networks for Volumetric Medical Image Segmentation. In: 3DV (2016) 8", + "35. Nekrasov, A., Schult, J., Litany, O., Leibe, B., Engelmann, F.: Mix3D: Out-of-context data augmentation for 3D scenes. In: 3DV (2021) 2", + "36. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: DINOv2: Learning Robust Visual Features without Supervision. arXiv preprint arXiv:2304.07193 (2023) 4", + "37. Patashnik, O., Wu, Z., Shechtman, E., Cohen-Or, D., Lischinski, D.: StyleCLIP: Text-Driven Manipulation of StyleGAN Imagery. In: ICCV (2021) 4", + "38. Peng, S., Genova, K., Jiang, C., Tagliasacchi, A., Pollefeys, M., Funkhouser, T., et al.: OpenScene: 3D Scene Understanding with Open Vocabularies. In: CVPR (2023) 2, 4" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Huang et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "39. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation. In: CVPR (2017) 2", + "40. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In: NeurIPS (2017) 2", + "41. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning Transferable Visual Models from Natural Language Supervision. In: ICML (2021) 2, 4", + "42. Rao, Y., Zhao, W., Chen, G., Tang, Y., Zhu, Z., Huang, G., Zhou, J., Lu, J.: DenseCLIP: Language-Guided Dense Prediction with Context-Aware Prompting. In: CVPR (2022) 4", + "43. Roynard, X., Deschaud, J.E., Goulette, F.: Paris-Lille-3D: A Large and High-Quality Ground-Truth Urban Point Cloud Dataset for Automatic Segmentation and Classification. *IJRR* (2018) 11", + "44. Rozenberszki, D., Litany, O., Dai, A.: Language-Grounded Indoor 3D Semantic Segmentation in the Wild. In: ECCV (2022) 2, 3, 9, 11", + "45. Rozenberszki, D., Litany, O., Dai, A.: UnScene3D: Unsupervised 3D Instance Segmentation for Indoor Scenes. In: CVPR (2024) 10, 11", + "46. Schult, J., Engelmann, F., Hermans, A., Litany, O., Tang, S., Leibe, B.: Mask3D: Mask Transformer for 3D Semantic Instance Segmentation. In: ICRA (2023) 1, 2, 3, 7, 8, 9, 10, 11, 12, 14", + "47. Schult, J., Engelmann, F., Kontogianni, T., Leibe, B.: DualConvMesh-Net: Joint Geodesic and Euclidean Convolutions on 3D Meshes. In: CVPR (2020) 2, 10", + "48. Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur-Artal, R., Ren, C., Verma, S., et al.: The Replica Dataset: A Digital Replica of Indoor Spaces. arXiv preprint arXiv:1906.05797 (2019) 4", + "49. Sun, J., Qing, C., Tan, J., Xu, X.: Superpoint Transformer for 3D Scene Instance Segmentation. In: AAAI (2023) 10", + "50. Sun, T., Hao, Y., Huang, S., Savarese, S., Schindler, K., Pollefeys, M., Armeni, I.: Nothing stands still: A spatiotemporal benchmark on 3d point cloud registration under large geometric and temporal change. arXiv preprint arXiv:2311.09346 (2023) 2", + "51. Takmaz, A., Fedele, E., Sumner, R.W., Pollefeys, M., Tombari, F., Engelmann, F.: OpenMask3D: Open-Vocabulary 3D Instance Segmentation. In: NeurIPS (2023) 2, 3, 4, 9, 13, 14", + "52. Takmaz, A., Schult, J., Kaftan, I., Akçay, M., Leibe, B., Sumner, R., Engelmann, F., Tang, S.: 3D Segmentation of Humans in Point Clouds with Synthetic Data. In: ICCV (2023) 2", + "53. Thomas, H., Qi, C.R., Deschaud, J.E., Marcotegui, B., Goulette, F., Guibas, L.J.: KPCov: Flexible and Deformable Convolution for Point Clouds. In: ICCV (2019) 2", + "54. Vu, T., Kim, K., Luu, T.M., Nguyen, T., Yoo, C.D.: SoftGroup for 3D Instance Segmentation on Point Clouds. In: CVPR (2022) 2, 3, 10", + "55. Weder, S., Blum, H., Engelmann, F., Pollefeys, M.: LabelMaker: Automatic Semantic Label Generation from RGB-D Trajectories. In: 3DV (2024) 3", + "56. Yang, B., Wang, J., Clark, R., Hu, Q., Wang, S., Markham, A., Trigoni, N.: Learning Object Bounding Boxes for 3D Instance Segmentation on Point Clouds. In: NeurIPS (2019) 3", + "57. Yang, Y., Wu, X., He, T., Zhao, H., Liu, X.: SAM3D: Segment Anything in 3D Scenes. In: ICCVW (2023) 2, 4, 5, 10, 11, 12", + "58. Yeshwanth, C., Liu, Y.C., Nießner, M., Dai, A.: ScanNet++: A High-Fidelity Dataset of 3D Indoor Scenes. In: ICCV (2023) 3, 9, 10" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Segment3D", + "bbox": [ + 651, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "59. Yi, L., Zhao, W., Wang, H., Sung, M., Guibas, L.J.: GSPN: Generative Shape Proposal Network for 3D Instance Segmentation in Point Cloud. In: CVPR (2019) 3", + "60. Yue, Y., Das, A., Engelmann, F., Tang, S., Lenssen, J.: Improving 2D Feature Representations by 3D-Aware Fine-Tuning. In: ECCV (2024) 2", + "61. Yue, Y., Kontogianni, T., Schindler, K., Engelmann, F.: Connecting the Dots: Floorplan Reconstruction Using Two-Level Queries. In: CVPR (2023) 2", + "62. Zeng, Y., Jiang, C., Mao, J., Han, J., Ye, C., Huang, Q., Yeung, D.Y., Yang, Z., Liang, X., Xu, H.: CLIP2: Contrastive Language-Image-Point Pretraining from Real-World Point Cloud Data. In: CVPR (2023) 4", + "63. Zhang, R., Guo, Z., Zhang, W., Li, K., Miao, X., Cui, B., Qiao, Y., Gao, P., Li, H.: PointCLIP: Point Cloud Understanding by CLIP. In: CVPR (2022) 4", + "64. Zurbrugg, R., Liu, Y., Engelmann, F., Kumar, S., Hutter, M., Patil, V., Yu, F.: ICGNet: A Unified Approach for Instance-Centric Grasping. In: ICRA (2024) 2" + ], + "bbox": [ + 215, + 146, + 784, + 340 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Huang et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Segment3D_ Learning Fine-Grained Class-Agnostic 3D Segmentation without Manual Labels/a98903ad-e70d-49e3-988e-6063b3741036_model.json b/2024/Segment3D_ Learning Fine-Grained Class-Agnostic 3D Segmentation without Manual Labels/a98903ad-e70d-49e3-988e-6063b3741036_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3d9393d7f0e2d234c5af7e6afdf7d4425180b15d --- /dev/null +++ b/2024/Segment3D_ Learning Fine-Grained Class-Agnostic 3D Segmentation without Manual Labels/a98903ad-e70d-49e3-988e-6063b3741036_model.json @@ -0,0 +1,2513 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.299, + 0.141, + 0.705, + 0.207 + ], + "angle": 0, + "content": "Segment3D: Learning Fine-Grained Class-Agnostic 3D Segmentation without Manual Labels" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.234, + 0.784, + 0.266 + ], + "angle": 0, + "content": "Rui Huang\\(^{1}\\), Songyou Peng\\(^{2}\\), Ayça Takmaz\\(^{2}\\), Federico Tombari\\(^{3}\\), Marc Pollefeys\\(^{2,4}\\), Shiji Song\\(^{1}\\), Gao Huang\\(^{1,*}\\), and Francis Engelmann\\(^{2,3}\\)" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.271, + 0.363, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.374, + 0.276, + 0.621, + 0.322 + ], + "angle": 0, + "content": "\\(^{1}\\)Tsinghua University \\(^{2}\\)ETH Zurich \\(^{3}\\)Google \\(^{4}\\)Microsoft *Corr. Author https://segment3d.github.io" + }, + { + "type": "image", + "bbox": [ + 0.645, + 0.272, + 0.787, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.356, + 0.363, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.363, + 0.326, + 0.496, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.363, + 0.414, + 0.496, + 0.501 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.385, + 0.503, + 0.474, + 0.534 + ], + "angle": 0, + "content": "Mask3D [46] trained on manual labels." + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.326, + 0.648, + 0.501 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.516, + 0.502, + 0.64, + 0.533 + ], + "angle": 0, + "content": "Segment3D (Ours) trained on automatic labels." + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.358, + 0.785, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.646, + 0.442, + 0.787, + 0.547 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.555, + 0.788, + 0.626 + ], + "angle": 0, + "content": "Fig. 1: Fine-Grained Class-Agnostic 3D Point Cloud Segmentation. Segment3D predicts highly accurate segmentation masks (right), improves over state-of-the-art 3D segmentation methods (e.g., Mask3D [46], left), and does not require manually labeled 3D training data. This is achieved through the automatic generation of high-quality training masks using foundation models for image segmentation [28]." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.646, + 0.741, + 0.84 + ], + "angle": 0, + "content": "Abstract. Current 3D scene segmentation methods are heavily dependent on manually annotated 3D training datasets. Such manual annotations are labor-intensive, and often lack fine-grained details. Furthermore, models trained on this data typically struggle to recognize object classes beyond the annotated training classes, i.e., they do not generalize well to unseen domains and require additional domain-specific annotations. In contrast, recent 2D foundation models have demonstrated strong generalization and impressive zero-shot abilities, inspiring us to incorporate these characteristics from 2D models into 3D models. Therefore, we explore the use of image segmentation foundation models to automatically generate high-quality training labels for 3D segmentation models. The resulting model, Segment3D, generalizes significantly better than the models trained on costly manual 3D labels and enables easily adding new training data to further boost the segmentation performance." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.129 + ], + "angle": 0, + "content": "Huang et al." + }, + { + "type": "header", + "bbox": [ + 0.262, + 0.148, + 0.741, + 0.162 + ], + "angle": 0, + "content": "Keywords: Class-agnostic 3D segmentation \\(\\cdot\\) 3D Scene Understanding" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.187, + 0.377, + 0.203 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.218, + 0.786, + 0.294 + ], + "angle": 0, + "content": "In this work, we propose Segment3D, a method for fine-grained class-agnostic 3D segmentation. In particular, dividing the space into coherent segments aligned with both the scene geometry and its semantics is a key challenge. This ability to accurately segment and interpret 3D scenes is fundamental for numerous downstream tasks [52,61], intelligent assistants and autonomous robots [30,64]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.295, + 0.787, + 0.566 + ], + "angle": 0, + "content": "Current methods for 3D indoor-scene understanding mostly focus on semantic [35, 39, 40, 47, 53] and instance segmentation [9, 17, 31, 32, 46, 54]. These approaches, while effective on popular benchmarks [13, 15, 44, 50], have limitations stemming from their training. Primarily, they depend on extensive manually labeled 3D training sets that are both time-consuming and challenging to annotate. Importantly, their performance often deteriorates when applied to scenarios beyond their training data, limiting their effectiveness in diverse, real-world scenarios. This becomes particularly apparent under the recently emerging task of open-vocabulary 3D scene understanding [18, 27, 33, 38, 51] that aims to segment arbitrary user queries, which naturally go beyond the pre-defined set of training-set classes. Concurrently, the recent surge in foundation models, especially 2D vision-language models [25, 28, 41, 60], demonstrates remarkable potential. Trained on internet-scale data, these models exhibit an extraordinary ability to generalize, even in a zero-shot setting, to new and different input distributions. However, their application has been predominantly confined to 2D data. For instance, SAM [28] has shown impressive results in 2D image segmentation, but its applicability to 3D scene understanding remains largely under-explored. All these factors give rise to the research question:" + }, + { + "type": "title", + "bbox": [ + 0.263, + 0.575, + 0.741, + 0.605 + ], + "angle": 0, + "content": "How to leverage 2D foundation models for class-agnostic 3D scene segmentation without requiring manually labeled 3D data?" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.787, + 0.779 + ], + "angle": 0, + "content": "Recently, SAM3D [57] has proposed a straightforward method that uses posed RGB-D images corresponding to a 3D scene. They predict segmentation masks for all input RGB-D images with SAM [28], and projected them into the 3D space. Next, through an iterative bottom-up process, the 3D masks of partial scenes are merged to derive the final 3D segmentation (see Fig. 2). However, variations in perspectives across frames can lead to conflicting segmentations, which introduces inconsistency during the merging process. Besides, the inference speed of SAM3D is slow, due to running SAM on the fly for a large number of RGB images and the cumbersome non-learned merging process. In this paper, we seek to circumvent both the slow speed and the issue of merging inconsistent segmentation across frames." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Towards this goal, we introduce Segment3D, a clean and efficient approach that utilizes a 3D model to achieve class-agnostic, fine-grained 3D segmentation at speed. Segment3D employs a two-stage training approach that requires no hand-annotated labels. We first train our class-agnostic 3D segmentation model" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.653, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Segment3D" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.375 + ], + "angle": 0, + "content": "with RGB-D data, which are transformed into partial RGB-D point clouds. The supervision is provided by the segmentation masks produced by SAM and projected into 3D. While SAM can automatically generate masks for entire images, sometimes the results may suffer from over-segmentation (see Fig. 4), which is not desired in our context. Therefore, we propose a Mask Generation Module to generate high-quality, complete masks of objects for a given RGB image. Since there is readily available large-scale RGB-D data, additional data can be effortlessly integrated to further enhance segmentation performance. This pretraining stage lays the groundwork for understanding the 3D structure from 2D annotations. However, as our ultimate objective is the segmentation of full 3D scenes, we must bridge the domain gap between partial point clouds and the more comprehensive 3D point clouds obtained from 3D scanners or reconstruction techniques [14,24]. To this end, in the second stage, we fine-tune the model on full 3D point clouds in a self-supervised manner, utilizing high-confidence mask predictions from the pre-trained model as training signal." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.376, + 0.788, + 0.466 + ], + "angle": 0, + "content": "We demonstrate strong performance in class-agnostic segmentation on ScanNet [13, 44] and the newly released ScanNet++ [58]. Though Segment3D is trained on the indoor dataset, transferring it directly to the outdoor nuScenes dataset [4] also yields surprisingly good results. Finally, we show the use of Segment3D for improving open-vocabulary 3D instance segmentation [51]. Overall, the contributions of this paper are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.471, + 0.784, + 0.514 + ], + "angle": 0, + "content": "- We introduce Segment3D, a novel approach and training strategy for fine-grained class-agnostic 3D point cloud segmentation without manually annotated labels." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.516, + 0.784, + 0.544 + ], + "angle": 0, + "content": "- We propose a Mask Generation Module that utilizes the 2D foundation model to automatically generate high-quality, complete training masks." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.547, + 0.784, + 0.589 + ], + "angle": 0, + "content": "- We show that Segment3D demonstrates strong generalization compared to a wide range of baselines, including fully supervised methods trained on carefully annotated datasets." + }, + { + "type": "list", + "bbox": [ + 0.215, + 0.471, + 0.784, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.625, + 0.388, + 0.642 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.84 + ], + "angle": 0, + "content": "3D Instance Segmentation. Current models have seen a significant development over the last years, from proposal-based [56, 59], over grouping-based [9, 26, 31, 54], to recent Transformer-based [32, 46] methods. In this work, we follow the currently best-performing Transformer-based paradigm. Despite impressive advancements, a shared limitation is the dependence on costly manual ground-truth annotations. Recently, there have been efforts to automate the annotation process by bundling state-of-the-art segmentation models [55], but they are still limited to pre-defined object classes. However, in the context of open-world 3D scene understanding, the importance of semantic classification for closed-set categories has diminished. Instead, in this work, we propose a general 3D segmentation method trained on automatically generated labels." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.129 + ], + "angle": 0, + "content": "Huang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.375 + ], + "angle": 0, + "content": "Foundation Models. Foundation models, particularly those that are multimodal [25,41], have revolutionized the field of AI by leveraging extensive image-text pre-training. These models can derive rich image representations guided by natural language descriptions, enabling a variety of downstream tasks [21,37,42]. Another line of work [6,36], based on self-supervised learning, employs image training and yields high-performance features directly applicable as inputs for linear classifiers. Segment Anything Model (SAM) [28] has recently advanced the performance of foundation models for image segmentation. SAM has undergone training on a diverse, high-quality dataset comprising more than 1 billion masks. This training equips SAM with the ability to generalize to novel object types and images, surpassing the scope of its observations during the training process. Additionally, SAM can generate high-quality, fine-grained masks. Our work leverages the power of SAM for pre-training a 3D segmentation model, which is later also used for generating supervision signals for fine-tuning on scene-level point clouds." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.379, + 0.788, + 0.624 + ], + "angle": 0, + "content": "Open-Vocabulary 3D Scene Understanding. Recently, there has been an increased interest in 3D open-vocabulary scene understanding. This new field utilizes the zero-shot recognition abilities of 2D vision-language models [25,41], enabling a more comprehensive understanding of diverse and previously unseen 3D environments [8,16,22,23,27,29,33,38,51,62,63]. PLA [16] aligns point cloud features with captions extracted from multi-view images of a scene to enable open-vocabulary recognition. OpenScene [38] distills per-pixel image features to 3D point clouds, generating point-wise scene representations co-embedded with text and image pixels in CLIP feature space. However, it mainly focuses on semantic segmentation and exhibits a limited understanding of object instances. To this end, OpenMask3D [51] predicts class-agnostic 3D instance masks and aggregates per-mask features via multi-view fusion of CLIP-based image embeddings. Nevertheless, the segmentation model it uses is trained on closed-set labels, lacking generalization to the open world. Our method contributes to this area by providing open-set class-agnostic masks, which can serve as foundational inputs for models such as OpenMask3D." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.643, + 0.331, + 0.66 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.673, + 0.788, + 0.81 + ], + "angle": 0, + "content": "We aim to develop a method capable of segmenting any object within a given 3D scene. Relying on existing 3D training datasets cannot accomplish this goal, as their mask annotations are limited to a predefined set of object classes [1,2,13,48]. Consequently, models trained on such datasets fail to generalize effectively to new, unseen classes. Recently, with SAM [28] showing extraordinary generalization ability in 2D segmentation, SAM3D [57] has proposed to simply project and merge SAM predictions in 3D from posed RGB-D images, as illustrated in Fig. 2. However, it suffers from inconsistent results and slow inference speed, limiting its practical applicability." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "To fully exploit SAM for consistent 3D segmentation at speed, we introduce a clean and efficient method which utilizes a 3D model to directly segment the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.653, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Segment3D" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.148, + 0.716, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.326, + 0.788, + 0.41 + ], + "angle": 0, + "content": "Fig. 2: SAM3D vs. Segment3D. SAM3D [57] (left) merges 2D segmentation masks of RGB-D images generated by SAM to obtain 3D segmentation of entire scenes. However, conflicting segmentations across frames introduce inconsistency during the merging process. Moreover, it is slow because of extensive image inference and the cumbersome merging procedure. Instead, our Segment3D (right) utilizes a 3D model to directly segment entire 3D scenes which is clean and efficient." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.428, + 0.788, + 0.563 + ], + "angle": 0, + "content": "entire scene. First, we propose a Mask Generation Module to generate high-quality, complete object masks from SAM predictions (Sec. 3.1). Next, we train our 3D segmentation model with RGB-D data and the generated object masks, which are transformed into partial point clouds (Sec. 3.2). Since the final goal is to segment 3D scenes, we need to bridge the inevitable domain gap between partial RGB-D point clouds and full 3D point clouds from 3D scanners or reconstruction methods. We therefore fine-tune our model on full 3D point clouds using high-confident mask predictions from our pre-trained model (Sec. 3.3). The overall framework is illustrated in Fig. 3." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.584, + 0.603, + 0.6 + ], + "angle": 0, + "content": "3.1 Generating Pseudo Ground-Truth Masks" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.735 + ], + "angle": 0, + "content": "Review on SAM. We first review the automatic mask generation pipeline of SAM [28]. SAM is prompted with a regular grid of \\(32 \\times 32\\) points on the RGB image. Each point predicts masks at three different granularities: whole, part, and subpart. Next, all three masks for all points are filtered by the model's predicted IoU score as well as the stability of the masks with respect to the binarization threshold. Then, the remaining masks undergo Non-Maximum Suppression (NMS). Finally, if a pixel is encompassed by multiple masks, the mask ID with the highest predicted IoU is assigned to the pixel." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.84 + ], + "angle": 0, + "content": "However, as shown in Fig. 4, the results obtained in this way could suffer from over-segmentation. A naive idea is to only use the masks generated by the \"whole\" branch. Nevertheless, we find that the \"whole\" branch typically generates high-quality masks for large objects while overlooking small objects in the image. In contrast, the \"part\" and \"subpart\" branches not only generate masks for parts of large objects but also predict masks for small and tiny objects. Thus, to enable our 3D model to achieve fine-grained segmentation capabilities for details and" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.129 + ], + "angle": 0, + "content": "Huang et al." + }, + { + "type": "image_caption", + "bbox": [ + 0.277, + 0.148, + 0.502, + 0.159 + ], + "angle": 0, + "content": "Stage1: Pre-Training on Partial Point Clouds" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.161, + 0.782, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.565, + 0.148, + 0.774, + 0.159 + ], + "angle": 0, + "content": "Stage2: Fine-Tuning on Full Point Clouds" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.299, + 0.788, + 0.397 + ], + "angle": 0, + "content": "Fig. 3: Method Overview. Training Segment3D involves two stages: The first stage (left) relies on largely available RGB-D image sequences and SAM \\(\\square\\), a pre-trained foundation model for 2D image segmentation [28]. Segment3D \\(\\square\\) is pre-trained on partial RGB-D point clouds and supervised with pseudo ground-truth masks generated by Mask Generation Module (MGM) \\(\\square\\). Due to the domain gap between partial and full point clouds, in the second stage (right), Segment3D \\(\\square\\) is fine-tuned with confident masks predicted by the pre-trained Segment3D \\(\\square\\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.415, + 0.785, + 0.461 + ], + "angle": 0, + "content": "small objects, we must retain the outputs from the \"part\" and \"subpart\" branches. The problem is how to preserve their segmentation for small and tiny objects while removing their segmentation for parts of large objects." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.467, + 0.789, + 0.668 + ], + "angle": 0, + "content": "Mask Generation Module (MGM). Here we introduce our Mask Generation Module (MGM). Since SAM's automatic mask generation pipeline can provide reasonably high-quality masks for a given image, we start based on it. Next, we compare the result with the output of the \"whole\" branch to determine which masks are parts of large objects and remove them. Specifically, for each mask in the automatic output \\( m_i^{\\mathrm{auto}} \\), we iterate through the masks of the entire branch, calculating the intersection \\( m_{ij}^{\\mathrm{inter}} \\) between \\( m_i^{\\mathrm{auto}} \\) and \\( m_j^{\\mathrm{whole}} \\). If \\( m_{ij}^{\\mathrm{inter}} / m_i^{\\mathrm{auto}} \\geq t \\), then we consider \\( m_i^{\\mathrm{auto}} \\) to be a part of \\( m_j^{\\mathrm{whole}} \\). Our preliminary results show that the process of extracting valid masks is highly robust to the hyper-parameter \\( t \\) and we set it as 0.9 by default. We remove the part masks and combine the remaining masks with the output of the \"whole\" branch, which will serve as our training supervision. With MGM, we observe notable improvements at all object mask sizes, especially on medium-sized and large objects (see Table 6)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.691, + 0.646, + 0.707 + ], + "angle": 0, + "content": "3.2 Stage 1: Pre-Training on RGB-D Point Clouds" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.713, + 0.788, + 0.789 + ], + "angle": 0, + "content": "In contrast to the relatively scarce availability of 3D data, there is an abundance of 2D data, particularly RGB-D images, which are readily accessible. For example, ScanNet [13] comprises merely 1513 3D scans, compared to the substantially larger collection of 2.5 million RGB-D images. Therefore, we first pre-train our 3D segmentation model on partial RGB-D point clouds." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.795, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Training-Set Preparation. Starting from a collection of RGB-D frames, we create partial 3D point clouds and their corresponding pseudo ground-truth 3D masks. Note that the labels can be automatically obtained without manual effort." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.653, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Segment3D" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image_caption", + "bbox": [ + 0.245, + 0.149, + 0.328, + 0.17 + ], + "angle": 0, + "content": "Automatic-SAM Generated Masks" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.172, + 0.347, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.243, + 0.346, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.408, + 0.149, + 0.468, + 0.17 + ], + "angle": 0, + "content": "MGM rated Masks" + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.172, + 0.485, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.243, + 0.485, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.322, + 0.493, + 0.39 + ], + "angle": 0, + "content": "Fig. 4: Automatic-SAM Generated (left) vs. MGM Generated (right) Masks from ScanNet [13]. The former exhibits over-segmentation, while the latter does not." + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.149, + 0.785, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.339, + 0.788, + 0.408 + ], + "angle": 0, + "content": "Fig. 5: Partial (center) vs. Full (right) Point Clouds from ScanNet [13]. Full point clouds are more complete and exhibit fewer occlusions due to reconstruction over multiple viewpoints." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.431, + 0.785, + 0.505 + ], + "angle": 0, + "content": "For each frame in a large RGB-D dataset, we first transform the 2D depth map to a partial 3D point cloud. To do so, we need to know the intrinsic matrix \\(\\mathbf{K} \\in \\mathbb{R}^{3 \\times 3}\\) and extrinsic matrix \\(\\mathbf{T} = [\\mathbf{R} \\mathbf{t}] \\in \\mathbb{R}^{3 \\times 4}\\). For each pixel \\(\\mathbf{p} = (u, v)\\), we can transform it with its depth value \\(D_{\\mathbf{p}}\\) into a 3D point \\(\\mathbf{P}\\) in world coordinates as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.382, + 0.511, + 0.785, + 0.531 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} = \\mathbf {R} ^ {\\top} \\cdot \\left(D _ {\\mathbf {p}} \\cdot \\mathbf {K} ^ {- 1} \\cdot \\tilde {\\mathbf {p}}\\right) - \\mathbf {R} ^ {\\top} \\mathbf {t}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.542, + 0.788, + 0.634 + ], + "angle": 0, + "content": "where \\(\\tilde{\\mathbf{p}}\\) is the homogeneous coordinate of \\(\\mathbf{p}\\). By applying Eq. (1) to all pixels in the depth map \\(D\\) and associating their per-pixel RGB value, we obtain the input partial 3D point cloud. Next, we obtain the pseudo ground-truth 3D segmentation masks for this point cloud with the 2D masks generated by MGM in the same way. Since we know the one-to-one mapping between 2D pixel and 3D points from Eq. (1), we directly obtain the per-point 3D mask labels." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Model Architecture. We use a model inspired by Mask3D [46] to train a class-agnostic 3D segmentation model. The model comprises a sparse convolutional backbone derived from MinkowskiUNet [12] and a transformer decoder, as in MaskFormer [10, 11]. We adopt a set of queries to represent the masks, each of which is initialized with a positional embedding. Specifically, we select query positions with furthest point sampling (FPS) and use their Fourier positional encodings as the query embeddings. Leveraging the transformer decoder, all the mask queries are refined by progressively attending to point cloud features across multiple scales in parallel. Each mask query is subsequently decoded into both a mask feature and a binary label to predict whether the given query corresponds to a valid object or not. By computing cosine similarity scores between a mask feature and all point features within the point cloud, a heatmap is generated over" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.129 + ], + "angle": 0, + "content": "Huang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "at the point cloud. This heatmap is input to a sigmoid function, and thresholded at 0.5, resulting in the final binary mask." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.184, + 0.788, + 0.379 + ], + "angle": 0, + "content": "Training with SAM Generated Masks. We supervise the model with two losses: the per-point supervision loss \\(\\mathcal{L}_{\\mathrm{mask}}\\) and a per-query supervision loss \\(\\mathcal{L}_{\\mathrm{obj}}\\). The loss \\(\\mathcal{L}_{\\mathrm{mask}}\\) enables learning a foreground-background segmentation for each mask, and is composed of a dice loss \\(\\mathcal{L}_{\\mathrm{dice}}\\) [34] and a binary cross-entropy loss \\(\\mathcal{L}_{\\mathrm{ce}}\\) for each point. The \\(\\mathcal{L}_{\\mathrm{obj}}\\) is a binary classification loss that indicates whether a query represents a valid \"object\" or \"no object\". This mechanism allows for the prediction of a variable number of masks, depending on the underlying scene content and geometry. Following prior work [5, 10, 46], we first adopt bipartite graph matching to establish correspondences between the set of predicted masks and the set of target masks provided by MGM as described before. If the predicted instance finds a matching target mask, then we assign it an \"object\" label; conversely, if there is no match, we assign \"no object\". In summary, we optimize the following losses:" + }, + { + "type": "equation", + "bbox": [ + 0.423, + 0.381, + 0.785, + 0.396 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {m a s k}} + \\lambda_ {\\text {o b j}} \\mathcal {L} _ {\\text {o b j}}, \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.402, + 0.4, + 0.785, + 0.415 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {m a s k}} = \\lambda_ {\\text {d i c e}} \\mathcal {L} _ {\\text {d i c e}} + \\lambda_ {\\text {c e}} \\mathcal {L} _ {\\text {c e}}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.421, + 0.784, + 0.466 + ], + "angle": 0, + "content": "where \\(\\lambda_{*}\\) are hyperparameters that balance the contribution of each component in the loss. The binary classification loss \\(\\mathcal{L}_{\\mathrm{obj}}\\) is applied to all queries, while the mask loss \\(\\mathcal{L}_{\\mathrm{mask}}\\) is specifically applied to masks labeled as \"object\"." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.49, + 0.62, + 0.506 + ], + "angle": 0, + "content": "3.3 Stage 2: Self-Supervised Scene Fine-Tuning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.517, + 0.788, + 0.713 + ], + "angle": 0, + "content": "After Stage 1, we obtain a class-agnostic 3D segmentation model by pre-training solely on RGB-D images and automatically generated labels from MGM. However, a fundamental domain gap persists between partial point clouds derived from RGB-D data and full point clouds acquired through 3D scanners or reconstruction methods [14,24] (See Fig. 5). This gap exists mostly because of object occlusions, but also due to challenges of depth cameras to capture dark or reflective surfaces from a single viewpoint. Hence, depending solely on RGB-D frames for training a 3D segmentation model intended for full 3D scans proves inadequate. Therefore, we propose to further fine-tune our model on scene-level full 3D point clouds. The key idea to obtain 3D mask annotations for training on full point clouds is to use selected, high-confidence masks generated by the pre-trained model itself. Note that this stage requires no manual labels on full 3D scenes and proves essential for the performance of Segment3D (see Tab. 5)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Confidence-Based Mask Generation. Next, we outline the process of generating the supervision signal for the fine-tuning stage. The pre-trained model processes point clouds independently of their nature, be it partial or full. Therefore, when presented with a full 3D point cloud, the pre-trained 3D model produces a set of masks, each with a binary classification (valid or not) and a heatmap over all points, just as in Sec. 3.2. To assess the quality of the predicted masks, we compute a confidence score based on the confidence map \\(\\sigma(\\mathbf{h})\\), where \\(\\mathbf{h}\\) is the predicted heatmap and \\(\\sigma\\) is the sigmoid function. We then compute the average" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.653, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Segment3D" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.146, + 0.788, + 0.253 + ], + "angle": 0, + "content": "confidence of those points for which \\(\\sigma(\\mathbf{h}) > 0.5\\) as the confidence score of the predicted mask, denoted as \\(c_{\\mathrm{mask}}\\). We also consider the classification as a valid object and use the probability from the binary classification assigned to the \"object\" category as the confidence score, denoted as \\(c_{\\mathrm{obj}}\\). The final confidence score for each predicted mask is then the product of the two scores, \\(c = c_{\\mathrm{mask}} \\cdot c_{\\mathrm{obj}}\\). For fine-tuning our 3D segmentation model in Stage 2, we select the most confidently predicted masks above a threshold \\(\\tau_c\\)." + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.259, + 0.788, + 0.394 + ], + "angle": 0, + "content": "Training with the High-Confidence Generated Masks. For fine-tuning, we follow the same procedure as before and use \\(\\mathcal{L}_{\\mathrm{mask}}\\) as defined in Eq. 3. In contrast to the pre-training stage, the binary classification loss \\(\\mathcal{L}_{\\mathrm{obj}}\\), responsible for categorizing queries into valid or invalid, is omitted. Since we only select masks with high confidence for supervision, it can happen that some objects in the scene have no assigned ground truth mask. In such instances, deeming a correctly predicted mask for those objects as invalid would be detrimental. Table 5 illustrates the efficacy of the self-supervised fine-tuning process in comparison to pre-training alone." + }, + { + "type": "title", + "bbox": [ + 0.219, + 0.418, + 0.375, + 0.435 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.448, + 0.788, + 0.554 + ], + "angle": 0, + "content": "We firstly evaluate Segment3D on indoor ScanNet++ [58], ScanNet200 [13,44] and outdoor nuScenes [4] in a class-agnostic segmentation setting (Sec. 4.1). We then show the advantage of our method in segmenting small objects, and provide analysis to understand the importance of fine-tuning, the effectiveness of Mask Generation Module and the potential of training on more data (Sec. 4.2). Finally, we demonstrate its application for the task of open-set 3D instance segmentation as proposed in OpenMask3D [51] (Sec. 4.3)." + }, + { + "type": "title", + "bbox": [ + 0.219, + 0.576, + 0.614, + 0.592 + ], + "angle": 0, + "content": "4.1 Comparing with State-of-the-Art Methods" + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.607, + 0.788, + 0.743 + ], + "angle": 0, + "content": "Datasets. We first evaluate on the recently released ScanNet++ [58] which includes high-resolution 3D scans captured at sub-millimeter precision, and fine-grained annotations covering objects of varying sizes. Due to its comprehensive data annotation, all of our analytical experiments in Sec. 4.2 are also based on it. Next, we adopt the ScanNet200 [13, 44], which builds upon the classic ScanNet [13] dataset and extends the semantic annotations to 200 classes. Then, we test our model on the nuScenes [4] outdoor dataset. For training our model, we employ ScanNet [13, 44], which is collected through a lightweight RGB-D scanning process. See the appendix for the implementation details of our method." + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.749, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Methods in Comparison. We compare with a wide range of prior art methods from different categories. Mask3D [46] is a state-of-the-art fully-supervised method trained on manually annotated 3D segmentation masks. Segment3D has the same backbone as Mask3D but instead of training on manually annotated 3D masks, it learns from automatically generated masks. Felzenszwalb et al. [20] proposed a graph-based method for segmentation which operates directly on" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.129 + ], + "angle": 0, + "content": "Huang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.2 + ], + "angle": 0, + "content": "Table 1: Segmentation Scores on ScanNet++ [58]. The metric is average precision (AP) on the validation split. We report scores with and without post-processing (more details in Sec. 4.1). For reference, we show results on Mask3D trained on manual ScanNet200 labels." + }, + { + "type": "table", + "bbox": [ + 0.233, + 0.207, + 0.763, + 0.31 + ], + "angle": 0, + "content": "
ModelAvg. Inference Time/swithout post-processingwith post-processing
AP\\( AP_{50} \\)\\( AP_{25} \\)AP\\( AP_{50} \\)\\( AP_{25} \\)
Mask3D [46]0.78.715.527.214.321.329.9
SAM3D [57]386.73.99.322.18.416.130.0
Felzenszwalb et al. [20]12.65.811.627.2---
Segment3D (Ours)0.713.023.838.320.230.942.7
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.34, + 0.784, + 0.401 + ], + "angle": 0, + "content": "the 3D geometry. UnScene3D [45] leverages self-supervised color and geometry features to obtain coarse segmentation and refines it through self-training. SAM3D [57] merges the segmentation masks of RGB-D images generated by SAM [28] to obtain the 3D segmentation result." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.408, + 0.785, + 0.514 + ], + "angle": 0, + "content": "Metrics. We report average precision (AP) scores at IoU thresholds of \\(25\\%\\), \\(50\\%\\), and averaged over the range [0.5:0.95:05] between predicted and ground truth masks. Consistent with common practices in the field [47, 49, 54], we also report scores after post-processing. This involves smoothing the predicted masks through graph-based oversegmentation [20], and splitting distant parts of the same mask via connected component clustering DBSCAN [19]. We also report the average inference time for these methods on ScanNet++ tested on A100." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.522, + 0.788, + 0.734 + ], + "angle": 0, + "content": "Results on ScanNet++. Scores are reported in Table 1. Segment3D outperforms all previous methods by at least +4.3 AP and up to +12.7 AP\\(_{25}\\). Notably, we achieve such improvements without any ground-truth mask annotation contrary to Mask3D trained on ScanNet200. In general, the performance of Mask3D (and other fully-supervised methods) depends on the quality of the annotated training dataset; often the manual annotation of small objects (pens, cellphones) or other fine-details is challenging. Instead, Segment3D relies on automatically generated high-quality masks from MGM using the 2D foundation model, which can capture fine-grained details without human annotation effort. Table 4 highlights that Segment3D excels particularly in predicting the more challenging small object masks. Furthermore, our method significantly outperforms SAM3D, indicating that Segment3D can circumvent the noise that arises during the merging process of SAM3D. Additionally, As shown in Table 1, Segment3D is three-order-of-magnitude faster than SAM3D in inference speed." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Fig. 6 shows segmentation results of several representative examples on ScanNet++. As can be seen, the scenes are quite diverse, presenting multiple challenges such as clutter and a wide range of mask sizes. Despite these challenges, our model predicts quite accurate and well-localized segmentation masks. For example, compared to Mask3D, our method is able to segment the finer-grained objects on top of the bed and the shelf. The masks of the computer screen and the chair in front of it are also less fragmented than predictions of the baselines." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.653, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Segment3D" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.147, + 0.493, + 0.19 + ], + "angle": 0, + "content": "Table 2: Segmentation Scores on ScanNet200 [13, 44]. The evaluation metric is average precision on the val split." + }, + { + "type": "table", + "bbox": [ + 0.239, + 0.202, + 0.462, + 0.296 + ], + "angle": 0, + "content": "
ModelAPAP50
Mask3D [46]34.143.1
Felzenszwalb et al. [20]6.112.1
UnScene3D [45]15.932.2
SAM3D [57]19.032.5
Segment3D (Ours)27.739.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.51, + 0.145, + 0.788, + 0.188 + ], + "angle": 0, + "content": "Table 3: Segmentation Scores on nuScenes [4]. The evaluation metric is average precision on the val split." + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.199, + 0.764, + 0.298 + ], + "angle": 0, + "content": "
ModelGT LabelsAP50AP25
zero-shot test
Mask3D×15.325.5
Segment3D×37.848.0
fine-tuned on nuScenes
Mask3D✓46.557.8
Segment3D×42.555.5
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.325, + 0.784, + 0.417 + ], + "angle": 0, + "content": "Results on ScanNet200. As illustrated in Table 2, Segment3D outperforms all baselines that require no manual 3D labels, and even shows competitive results with Mask3D, which is trained on ScanNet200. Compared to Unscene3D, which leverages self-supervised color and geometry features to obtain coarse segmentation, we achieve superior results by using the Mask Generation Module to generate high-quality complete masks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.422, + 0.788, + 0.573 + ], + "angle": 0, + "content": "Results on Outdoor Data. The performance on nuScenes is demonstrated in Table 3. We observe that even though Segment3D is trained using purely indoor data, applying it directly to outdoor data produces surprisingly good results. This indicates the stronger generalizability of Segment3D over the fully-supervised method Mask3D when transferring from indoor to outdoor scenes. Furthermore, we report scores with fine-tuning on nuScenes which leads to overall improved results. Similar to Table 2, Mask3D performs best, however, Segment3D achieves very competitive performance without any hand-labeled training data. Refer to the appendix for additional results on STPLS3D [7] and Paris-Lille-3D [43] datasets." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.595, + 0.442, + 0.61 + ], + "angle": 0, + "content": "4.2 Analysis Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.625, + 0.788, + 0.703 + ], + "angle": 0, + "content": "Performance on Different Mask Sizes. We proceed with an analysis of the performance of our Segment3D and Mask3D (the best-performing baseline) on object masks of various sizes. We categorize the size of masks based on the number of points they contain (small: [0, 2k], medium: [2k, 15k] large: [15k, \\(\\infty\\)]). The results are reported in Table 4. Our method yields significantly improved" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.726, + 0.787, + 0.755 + ], + "angle": 0, + "content": "Table 4: Segmentation Scores on Different Mask Sizes. Segmented3D improves over Mask3D, especially on small object masks. Details are in Sec. 4.2." + }, + { + "type": "table", + "bbox": [ + 0.28, + 0.76, + 0.728, + 0.838 + ], + "angle": 0, + "content": "
Mask SizeLargeMediumSmall
ScanNet++APAP50APAP50APAP50
Mask3D13.825.113.623.31.84.3
Segment3D18.331.916.629.88.417.3
(+4.5)(+6.8)(+3.0)(+6.5)(+6.6)(+13.0)
" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.129 + ], + "angle": 0, + "content": "Huang et al." + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.143, + 0.778, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.551, + 0.788, + 0.595 + ], + "angle": 0, + "content": "Fig. 6: Qualitative Results on ScanNet++. From top to bottom, we show the colored input 3D scenes, the segmentation masks predicted by SAM3D [57], Mask3D [46], our Segment3D and the ground truth 3D mask annotations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.633, + 0.788, + 0.74 + ], + "angle": 0, + "content": "segmentation results on objects of various sizes, especially on small objects. This confirms our intuition that Mask3D performs poorly on small-sized object masks as those are typically harder to manually annotate. In contrast, Segment3D utilizes masks generated by MGM with the foundation model as supervision, capturing fine-grained scene details. This showcases the usefulness of foundation models and raises the question if manually labeled large-scale 3D datasets are necessary for training 3D point cloud segmentation models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.842 + ], + "angle": 0, + "content": "The Importance of Two-Stage Training. Next, we compare the performance of Segment3D pre-trained solely on partial RGB-D point clouds (Stage 1) and with additional fine-tuning on full point clouds (Stage 2). Scores are reported in Table 5. The additional fine-tuning stage almost doubles the segmentation performance of our model on the most challenging AP metric. By training with the predicted high-confidence masks, Segment3D effectively reduces the inherent" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.653, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Segment3D" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 5: The Importance of Two-Stage Training. Fine-tuning on full point clouds supervised by confident predicted 3D masks significantly surpasses pre-training on projected 2D SAM masks alone." + }, + { + "type": "table", + "bbox": [ + 0.304, + 0.193, + 0.695, + 0.247 + ], + "angle": 0, + "content": "
Training StagesAPAP50AP25
Pre-Training (Stage 1)6.814.230.6
+ Fine-Tuning (Stage 2)13.0 (+6.2)23.8 (+9.6)38.3 (+7.7)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.259, + 0.785, + 0.302 + ], + "angle": 0, + "content": "Table 6: Effectiveness of Mask Generation Module. Compared to the model trained with masks generated by Automatic-SAM, our model shows improvement across all object mask sizes, particularly for medium-sized and large objects." + }, + { + "type": "table", + "bbox": [ + 0.276, + 0.307, + 0.72, + 0.395 + ], + "angle": 0, + "content": "
Mask SizeLargeMediumSmall
Masks Generated byAPAP50APAP50APAP50
Automatic-SAM16.830.614.227.27.516.2
MGM18.331.916.629.88.417.3
(+1.5)(+1.3)(+2.4)(+2.6)(+0.9)(+1.1)
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.421, + 0.785, + 0.451 + ], + "angle": 0, + "content": "domain gap between the partial point clouds derived from RGB-D images and full 3D point clouds." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.457, + 0.788, + 0.563 + ], + "angle": 0, + "content": "The Effectiveness of Mask Generation Module. To verify whether the MGM enhances the model's performance by reducing the over-segmentation issue and providing more complete masks, we conducted evaluations on objects of different sizes. As shown in Table 6, we can see that compared to the model trained with masks generated by Automatic-SAM, our model notably improves the performance of large and medium-sized objects which demonstrates the effectiveness of MGM." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.57, + 0.788, + 0.736 + ], + "angle": 0, + "content": "Pre-Training with Additional Data. Since RGB-D data is available in abundance and masks can be automatically generated, it is natural to ask if pretraining on additional masks will further improve the overall performance. To that end, we perform a first experiment where we select additional frames from the training set of the ScanNet++ dataset (to increase the variety of training data). Table 7 shows an impressive performance boost of +4.1 AP. For future work, considering that this improvement is obtained simply by adding more automatically generated masks to the pre-training, our approach seems promising to train on even more data. It could even be plausible to train on internet images combined with monocular depth estimation, such as ZoeDepth [3], to compute the partial point clouds." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.756, + 0.628, + 0.772 + ], + "angle": 0, + "content": "4.3 Application: Open-Set Scene Understanding" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "A real-world application of our class-agnostic 3D segmentation method is open-vocabulary 3D scene understanding, as implemented recently by OpenMask3D [51]. Given a 3D scene, a user can search for arbitrary objects via text prompts (see Fig. 7). A core component of OpenMask3D is Mask3D, which segments the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.129 + ], + "angle": 0, + "content": "Huang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.492, + 0.188 + ], + "angle": 0, + "content": "Table 7: Results with Additional Training Data. The performance is further boosted with more training data." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.2, + 0.484, + 0.248 + ], + "angle": 0, + "content": "
Pre-Training DataAPAP50
ScanNet13.023.8
ScanNet, ScanNet++17.1 (+4.1)26.7 (+2.9)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.145, + 0.787, + 0.188 + ], + "angle": 0, + "content": "Table 8: Open-Set 3D Scene Understanding. The evaluation metric is average precision with an IoU threshold of \\(50\\%\\)." + }, + { + "type": "table", + "bbox": [ + 0.509, + 0.2, + 0.773, + 0.248 + ], + "angle": 0, + "content": "
SegmentorScanNet++Replica
Mask3D [46]15.018.0
Segment3D (Ours)18.5 (+3.5)18.4 (+0.4)
" + }, + { + "type": "image", + "bbox": [ + 0.246, + 0.256, + 0.414, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.292, + 0.336, + 0.372, + 0.344 + ], + "angle": 0, + "content": "\"aback eraser\"" + }, + { + "type": "image", + "bbox": [ + 0.418, + 0.257, + 0.587, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.449, + 0.336, + 0.557, + 0.345 + ], + "angle": 0, + "content": "\"paper on the laptop\"" + }, + { + "type": "image", + "bbox": [ + 0.59, + 0.257, + 0.757, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.648, + 0.336, + 0.698, + 0.344 + ], + "angle": 0, + "content": "\"scissors\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.352, + 0.788, + 0.409 + ], + "angle": 0, + "content": "Fig. 7: Open-Set 3D Object Retrieval Results. Given a text prompt (bottom), OpenMask3D [51] finds the corresponding object masks in a given 3D scene (top). We adapt OpenMask3D and use fine-grained masks from our Segment3D method. We show 3D reconstructions and RGB images for better visualization (top left corner)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.438, + 0.788, + 0.545 + ], + "angle": 0, + "content": "scene into a set of object masks. Since Mask3D is trained on the closed-set of labeled annotations from ScanNet, its masks are not truly class-agnostic or open-vocabulary. We therefore replace Mask3D with our class-agnostic Segment3D. We evaluate on the closed-set labels of \\(\\mathrm{ScanNet}++\\) and Replica in Table 8 and report an improvement of up to \\(+3.5\\mathrm{AP}_{50}\\). Since Segment3D can accurately generate fine-grained masks in a class-agnostic manner, we are able to retrieve objects such as \"eraser\", \"paper\" and \"scissors\", which are small in scale." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.567, + 0.36, + 0.584 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.587, + 0.789, + 0.723 + ], + "angle": 0, + "content": "We have presented Segment3D, a simple, yet powerful class-agnostic 3D segmentation model. Segment3D employs a two-stage training framework that requires no manually annotated labels. We propose a Mask Generation Module to automatically generate high-quality, complete training masks with the 2D foundation model. The model is first pre-trained on RGB-D data with predicted 2D masks, then fine-tuned on full point clouds. Segment3D shows strong generalization and even outperforms existing 3D segmentation models that rely on hand-labeled 3D training scenes. Indeed, this raises the question of whether hand-labeled 3D training datasets are as essential as they were once thought to be." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.728, + 0.789, + 0.82 + ], + "angle": 0, + "content": "Acknowledgement. Gao Huang is supported in part by the National Natural Science Foundation of China under grants (62321005 and 42327901). Francis Engelmann is partially supported by an ETH AI Center postdoctoral research fellowship and an ETH Zurich Career Seed Award. Songyou Peng is supported by an Innosuisse funding (100.567 IP-ICT), and Ayça Takmaz is supported by an Innosuisse grant (48727.1 IP-ICT)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.653, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Segment3D" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.324, + 0.161 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.204 + ], + "angle": 0, + "content": "1. Armeni, I., Sener, O., Zamir, A.R., Jiang, H., Brilakis, I., Fischer, M., Savarese, S.: 3D Semantic Parsing of Large-Scale Indoor Spaces. In: CVPR (2016) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.205, + 0.785, + 0.259 + ], + "angle": 0, + "content": "2. Baruch, G., Chen, Z., Dehghan, A., Dimry, T., Feigin, Y., Fu, P., Gebauer, T., Joffe, B., Kurz, D., Schwartz, A., et al.: ARKitScenes: A Diverse Real-World Dataset for 3D Indoor Scene Understanding using Mobile RGB-D Data. arXiv preprint arXiv:2111.08897 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.261, + 0.785, + 0.3 + ], + "angle": 0, + "content": "3. Bhat, S.F., Birkl, R., Wofk, D., Wonka, P., Müller, M.: ZoeDepth: Zero-Shot Transfer by Combining Relative and Metric Depth. arXiv preprint arXiv:2302.12288 (2023) 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.302, + 0.785, + 0.342 + ], + "angle": 0, + "content": "4. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuScenes: A Multimodal Dataset for Autonomous Driving. In: CVPR (2020) 3, 9, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.344, + 0.785, + 0.37 + ], + "angle": 0, + "content": "5. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-End Object Detection with Transformers. In: ECCV (2020) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.372, + 0.785, + 0.41 + ], + "angle": 0, + "content": "6. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging Properties in Self-Supervised Vision Transformers. In: ICCV (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.412, + 0.785, + 0.452 + ], + "angle": 0, + "content": "7. Chen, M., Hu, Q., Yu, Z., Thomas, H., Feng, A., Hou, Y., McCullough, K., Ren, F., Soibelman, L.: STPLS3D: A Large-Scale Synthetic and Real Aerial Photogrammetry 3D Point Cloud Dataset. arXiv preprint arXiv:2203.09065 (2022) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.454, + 0.785, + 0.494 + ], + "angle": 0, + "content": "8. Chen, R., Liu, Y., Kong, L., Zhu, X., Ma, Y., Li, Y., Hou, Y., Qiao, Y., Wang, W.: CLIP2Scene: Towards Label-Efficient 3D Scene Understanding by CLIP. In: CVPR (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.496, + 0.785, + 0.522 + ], + "angle": 0, + "content": "9. Chen, S., Fang, J., Zhang, Q., Liu, W., Wang, X.: Hierarchical Aggregation for 3D Instance Segmentation. In: ICCV (2021) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.524, + 0.785, + 0.55 + ], + "angle": 0, + "content": "0. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-Attention Mask Transformer for Universal Image Segmentation. In: CVPR (2022) 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.551, + 0.785, + 0.577 + ], + "angle": 0, + "content": "1. Cheng, B., Schwing, A., Kirillov, A.: Per-Pixel Classification is Not All You Need for Semantic Segmentation. In: NeurIPS (2021) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.579, + 0.785, + 0.605 + ], + "angle": 0, + "content": "2. Choy, C., Gwak, J., Savarese, S.: 4D Spatio-Temporal ConvNets: Minkowski Convolutional Neural Networks. In: CVPR (2019) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.607, + 0.785, + 0.646 + ], + "angle": 0, + "content": "3. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: ScanNet: Richly-Annotated 3D Reconstructions of Indoor Scenes. In: CVPR (2017) 2, 3, 4, 6, 7, 9, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.648, + 0.785, + 0.688 + ], + "angle": 0, + "content": "4. Dai, A., Nießner, M., Zollhöfer, M., Izadi, S., Theobalt, C.: BundleFusion: Real-time Globally Consistent 3D Reconstruction using On-the-fly Surface Re-integration. TOG (2017) 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.689, + 0.785, + 0.729 + ], + "angle": 0, + "content": "5. Delitzas, A., Takmaz, A., Tombari, F., Sumner, R., Pollefeys, M., Engelmann, F.: SceneFun3D: Fine-Grained Functionality and Affordance Understanding in 3D Scenes. In: CVPR (2024) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.731, + 0.785, + 0.757 + ], + "angle": 0, + "content": "6. Ding, R., Yang, J., Xue, C., Zhang, W., Bai, S., Qi, X.: PLA: Language-Driven Open-Vocabulary 3D Scene Understanding. In: CVPR (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.758, + 0.785, + 0.796 + ], + "angle": 0, + "content": "7. Engelmann, F., Bokeloh, M., Fathi, A., Leibe, B., Nießner, M.: 3D-MPA: Multi-Proposal Aggregation for 3D Semantic Instance Segmentation. In: CVPR (2020) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "8. Engelmann, F., Manhardt, F., Niemeyer, M., Tateno, K., Tombari, F.: OpenNeRF: Open Set 3D Neural Scene Segmentation with Pixel-Wise Features and Rendered Novel Views. In: ICLR (2024) 2" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.129 + ], + "angle": 0, + "content": "Huang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "19. Ester, M., Kriegel, H.P., Sander, J., Xu, X., et al.: A Density-Based Algorithm for Discovering Clusters in Large Spatial Databases with Noise. In: KDD (1996) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.203 + ], + "angle": 0, + "content": "20. Felzenszwalb, P.F., Huttenlocher, D.P.: Efficient Graph-Based Image Segmentation. IJCV (2004) 9, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.785, + 0.231 + ], + "angle": 0, + "content": "21. Gu, X., Lin, T.Y., Kuo, W., Cui, Y.: Open-Vocabulary Object Detection via Vision and Language Knowledge Distillation. In: ICLR (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.259 + ], + "angle": 0, + "content": "22. Ha, H., Song, S.: Semantic Abstraction: Open-World 3D Scene Understanding from 2D Vision-Language Models. In: CoRL (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.785, + 0.3 + ], + "angle": 0, + "content": "23. Huang, T., Dong, B., Yang, Y., Huang, X., Lau, R.W., Ouyang, W., Zuo, W.: CLIP2Point: Transfer CLIP to Point Cloud Classification with Image-Depth Pre-Training. In: ICCV (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.301, + 0.785, + 0.355 + ], + "angle": 0, + "content": "24. Izadi, S., Kim, D., Hilliges, O., Molyneaux, D., Newcombe, R., Kohli, P., Shotton, J., Hodges, S., Freeman, D., Davison, A., et al.: KinectFusion: Real-time 3D Reconstruction and Interaction using a Moving Depth Camera. In: UIST (2011) 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.785, + 0.397 + ], + "angle": 0, + "content": "25. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up Visual and Vision-Language Representation Learning with Noisy Text Supervision. In: ICML (2021) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.424 + ], + "angle": 0, + "content": "26. Jiang, L., Zhao, H., Shi, S., Liu, S., Fu, C.W., Jia, J.: PointGroup: Dual-Set Point Grouping for 3D Instance Segmentation. In: CVPR (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.452 + ], + "angle": 0, + "content": "27. Kerr, J., Kim, C.M., Goldberg, K., Kanazawa, A., Tancik, M.: LERF: Language Embedded Radiance Fields. In: ICCV (2023) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.493 + ], + "angle": 0, + "content": "28. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment Anything. In: ICCV (2023) 1, 2, 4, 5, 6, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.494, + 0.785, + 0.522 + ], + "angle": 0, + "content": "29. Kobayashi, S., Matsumoto, E., Sitzmann, V.: Decomposing Nerf for Editing via Feature Field Distillation. In: NeurIPS (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.785, + 0.577 + ], + "angle": 0, + "content": "30. Lemke, O., Bauer, Z., Zurbrügg, R., Pollefeys, M., Engelmann, F., Blum, H.: Spotcompose: A framework for open-vocabulary object retrieval and drawer manipulation in point clouds. In: 2nd Workshop on Mobile Manipulation and Embodied Intelligence at ICRA 2024 (2024) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.605 + ], + "angle": 0, + "content": "31. Liang, Z., Li, Z., Xu, S., Tan, M., Jia, K.: Instance Segmentation in 3D Scenes using Semantic Superpoint Tree Networks. In: CVPR (2021) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.785, + 0.633 + ], + "angle": 0, + "content": "32. Lu, J., Deng, J., Wang, C., He, J., Zhang, T.: Query Refinement Transformer for 3D Instance Segmentation. In: ICCV (2023) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.634, + 0.785, + 0.673 + ], + "angle": 0, + "content": "33. Lu, Y., Xu, C., Wei, X., Xie, X., Tomizuka, M., Keutzer, K., Zhang, S.: Open-Vocabulary Point-Cloud Object Detection without 3D Annotation. In: CVPR (2023) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.785, + 0.702 + ], + "angle": 0, + "content": "34. Milletari, F., Navab, N., Ahmadi, S.A.: V-Net: Fully Convolutional Neural Networks for Volumetric Medical Image Segmentation. In: 3DV (2016) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.703, + 0.785, + 0.73 + ], + "angle": 0, + "content": "35. Nekrasov, A., Schult, J., Litany, O., Leibe, B., Engelmann, F.: Mix3D: Out-of-context data augmentation for 3D scenes. In: 3DV (2021) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.731, + 0.785, + 0.77 + ], + "angle": 0, + "content": "36. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: DINOv2: Learning Robust Visual Features without Supervision. arXiv preprint arXiv:2304.07193 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.799 + ], + "angle": 0, + "content": "37. Patashnik, O., Wu, Z., Shechtman, E., Cohen-Or, D., Lischinski, D.: StyleCLIP: Text-Driven Manipulation of StyleGAN Imagery. In: ICCV (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "38. Peng, S., Genova, K., Jiang, C., Tagliasacchi, A., Pollefeys, M., Funkhouser, T., et al.: OpenScene: 3D Scene Understanding with Open Vocabularies. In: CVPR (2023) 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.653, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Segment3D" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.175 + ], + "angle": 0, + "content": "39. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation. In: CVPR (2017) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.202 + ], + "angle": 0, + "content": "40. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In: NeurIPS (2017) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.203, + 0.788, + 0.243 + ], + "angle": 0, + "content": "41. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning Transferable Visual Models from Natural Language Supervision. In: ICML (2021) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.243, + 0.788, + 0.284 + ], + "angle": 0, + "content": "42. Rao, Y., Zhao, W., Chen, G., Tang, Y., Zhu, Z., Huang, G., Zhou, J., Lu, J.: DenseCLIP: Language-Guided Dense Prediction with Context-Aware Prompting. In: CVPR (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.285, + 0.788, + 0.324 + ], + "angle": 0, + "content": "43. Roynard, X., Deschaud, J.E., Goulette, F.: Paris-Lille-3D: A Large and High-Quality Ground-Truth Urban Point Cloud Dataset for Automatic Segmentation and Classification. *IJRR* (2018) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.325, + 0.788, + 0.352 + ], + "angle": 0, + "content": "44. Rozenberszki, D., Litany, O., Dai, A.: Language-Grounded Indoor 3D Semantic Segmentation in the Wild. In: ECCV (2022) 2, 3, 9, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.353, + 0.788, + 0.379 + ], + "angle": 0, + "content": "45. Rozenberszki, D., Litany, O., Dai, A.: UnScene3D: Unsupervised 3D Instance Segmentation for Indoor Scenes. In: CVPR (2024) 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.379, + 0.788, + 0.419 + ], + "angle": 0, + "content": "46. Schult, J., Engelmann, F., Hermans, A., Litany, O., Tang, S., Leibe, B.: Mask3D: Mask Transformer for 3D Semantic Instance Segmentation. In: ICRA (2023) 1, 2, 3, 7, 8, 9, 10, 11, 12, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.42, + 0.788, + 0.446 + ], + "angle": 0, + "content": "47. Schult, J., Engelmann, F., Kontogianni, T., Leibe, B.: DualConvMesh-Net: Joint Geodesic and Euclidean Convolutions on 3D Meshes. In: CVPR (2020) 2, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.447, + 0.788, + 0.488 + ], + "angle": 0, + "content": "48. Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur-Artal, R., Ren, C., Verma, S., et al.: The Replica Dataset: A Digital Replica of Indoor Spaces. arXiv preprint arXiv:1906.05797 (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.488, + 0.788, + 0.514 + ], + "angle": 0, + "content": "49. Sun, J., Qing, C., Tan, J., Xu, X.: Superpoint Transformer for 3D Scene Instance Segmentation. In: AAAI (2023) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.514, + 0.788, + 0.569 + ], + "angle": 0, + "content": "50. Sun, T., Hao, Y., Huang, S., Savarese, S., Schindler, K., Pollefeys, M., Armeni, I.: Nothing stands still: A spatiotemporal benchmark on 3d point cloud registration under large geometric and temporal change. arXiv preprint arXiv:2311.09346 (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.569, + 0.788, + 0.609 + ], + "angle": 0, + "content": "51. Takmaz, A., Fedele, E., Sumner, R.W., Pollefeys, M., Tombari, F., Engelmann, F.: OpenMask3D: Open-Vocabulary 3D Instance Segmentation. In: NeurIPS (2023) 2, 3, 4, 9, 13, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.61, + 0.788, + 0.651 + ], + "angle": 0, + "content": "52. Takmaz, A., Schult, J., Kaftan, I., Akçay, M., Leibe, B., Sumner, R., Engelmann, F., Tang, S.: 3D Segmentation of Humans in Point Clouds with Synthetic Data. In: ICCV (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.651, + 0.788, + 0.69 + ], + "angle": 0, + "content": "53. Thomas, H., Qi, C.R., Deschaud, J.E., Marcotegui, B., Goulette, F., Guibas, L.J.: KPCov: Flexible and Deformable Convolution for Point Clouds. In: ICCV (2019) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.691, + 0.788, + 0.718 + ], + "angle": 0, + "content": "54. Vu, T., Kim, K., Luu, T.M., Nguyen, T., Yoo, C.D.: SoftGroup for 3D Instance Segmentation on Point Clouds. In: CVPR (2022) 2, 3, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.719, + 0.788, + 0.746 + ], + "angle": 0, + "content": "55. Weder, S., Blum, H., Engelmann, F., Pollefeys, M.: LabelMaker: Automatic Semantic Label Generation from RGB-D Trajectories. In: 3DV (2024) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.746, + 0.788, + 0.786 + ], + "angle": 0, + "content": "56. Yang, B., Wang, J., Clark, R., Hu, Q., Wang, S., Markham, A., Trigoni, N.: Learning Object Bounding Boxes for 3D Instance Segmentation on Point Clouds. In: NeurIPS (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.786, + 0.788, + 0.813 + ], + "angle": 0, + "content": "57. Yang, Y., Wu, X., He, T., Zhao, H., Liu, X.: SAM3D: Segment Anything in 3D Scenes. In: ICCVW (2023) 2, 4, 5, 10, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.813, + 0.788, + 0.84 + ], + "angle": 0, + "content": "58. Yeshwanth, C., Liu, Y.C., Nießner, M., Dai, A.: ScanNet++: A High-Fidelity Dataset of 3D Indoor Scenes. In: ICCV (2023) 3, 9, 10" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.129 + ], + "angle": 0, + "content": "Huang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.188 + ], + "angle": 0, + "content": "59. Yi, L., Zhao, W., Wang, H., Sung, M., Guibas, L.J.: GSPN: Generative Shape Proposal Network for 3D Instance Segmentation in Point Cloud. In: CVPR (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.189, + 0.785, + 0.217 + ], + "angle": 0, + "content": "60. Yue, Y., Das, A., Engelmann, F., Tang, S., Lenssen, J.: Improving 2D Feature Representations by 3D-Aware Fine-Tuning. In: ECCV (2024) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.244 + ], + "angle": 0, + "content": "61. Yue, Y., Kontogianni, T., Schindler, K., Engelmann, F.: Connecting the Dots: Floorplan Reconstruction Using Two-Level Queries. In: CVPR (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.785, + 0.285 + ], + "angle": 0, + "content": "62. Zeng, Y., Jiang, C., Mao, J., Han, J., Ye, C., Huang, Q., Yeung, D.Y., Yang, Z., Liang, X., Xu, H.: CLIP2: Contrastive Language-Image-Point Pretraining from Real-World Point Cloud Data. In: CVPR (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.314 + ], + "angle": 0, + "content": "63. Zhang, R., Guo, Z., Zhang, W., Li, K., Miao, X., Cui, B., Qiao, Y., Gao, P., Li, H.: PointCLIP: Point Cloud Understanding by CLIP. In: CVPR (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.342 + ], + "angle": 0, + "content": "64. Zurbrugg, R., Liu, Y., Engelmann, F., Kumar, S., Hutter, M., Patil, V., Yu, F.: ICGNet: A Unified Approach for Instance-Centric Grasping. In: ICRA (2024) 2" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.342 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Segment3D_ Learning Fine-Grained Class-Agnostic 3D Segmentation without Manual Labels/a98903ad-e70d-49e3-988e-6063b3741036_origin.pdf b/2024/Segment3D_ Learning Fine-Grained Class-Agnostic 3D Segmentation without Manual Labels/a98903ad-e70d-49e3-988e-6063b3741036_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..791c947a226018793259dd86e6fd3eb02b2539ba --- /dev/null +++ b/2024/Segment3D_ Learning Fine-Grained Class-Agnostic 3D Segmentation without Manual Labels/a98903ad-e70d-49e3-988e-6063b3741036_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4fe139345d805ccc5ecbc6d64951ad42e2df095db77edac553e8ddd7f727b24b +size 8725387 diff --git a/2024/Segment3D_ Learning Fine-Grained Class-Agnostic 3D Segmentation without Manual Labels/full.md b/2024/Segment3D_ Learning Fine-Grained Class-Agnostic 3D Segmentation without Manual Labels/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5bafe72f06daaafef429ce531624614829fa8c10 --- /dev/null +++ b/2024/Segment3D_ Learning Fine-Grained Class-Agnostic 3D Segmentation without Manual Labels/full.md @@ -0,0 +1,297 @@ +# Segment3D: Learning Fine-Grained Class-Agnostic 3D Segmentation without Manual Labels + +Rui Huang $^{1}$ , Songyou Peng $^{2}$ , Ayça Takmaz $^{2}$ , Federico Tombari $^{3}$ , Marc Pollefeys $^{2,4}$ , Shiji Song $^{1}$ , Gao Huang $^{1,*}$ , and Francis Engelmann $^{2,3}$ + +![](images/60d9244e1dc0824a5af4c5e3ed84654cf14df2985f7b03550f194dbe0c75c040.jpg) + +$^{1}$ Tsinghua University $^{2}$ ETH Zurich $^{3}$ Google $^{4}$ Microsoft *Corr. Author https://segment3d.github.io + +![](images/4a65f75423c7b6cc674929c485768bfeadcf11d3c765b2ccad32071a2e91278f.jpg) + +![](images/02a083dd30a7cdc2269f2c2a4655f07b03e7b64862df8f3540abe08103c970e7.jpg) + +![](images/23deca315088794a2003b1809fb7c064cce75576b6ffdc223bb8ba918eafa4ec.jpg) + +![](images/977f807793e24d4383695bb2f8c5801da5a67eae6ec4df571485bde96d8d6b35.jpg) +Mask3D [46] trained on manual labels. + +![](images/ca2a3ddc14d4f8421e7737521a083b1df213c7e8ac3a1d7135b381ab047c1ff6.jpg) +Segment3D (Ours) trained on automatic labels. + +![](images/53e1e0c7a9520f86a8a02c5e3eaabc759f92bbcb36f5119192bff79bb450aea8.jpg) + +![](images/aabda92570de3afea7e8a391da2c6c58414d59976b9338cb44aa8c9538da057f.jpg) +Fig. 1: Fine-Grained Class-Agnostic 3D Point Cloud Segmentation. Segment3D predicts highly accurate segmentation masks (right), improves over state-of-the-art 3D segmentation methods (e.g., Mask3D [46], left), and does not require manually labeled 3D training data. This is achieved through the automatic generation of high-quality training masks using foundation models for image segmentation [28]. + +Abstract. Current 3D scene segmentation methods are heavily dependent on manually annotated 3D training datasets. Such manual annotations are labor-intensive, and often lack fine-grained details. Furthermore, models trained on this data typically struggle to recognize object classes beyond the annotated training classes, i.e., they do not generalize well to unseen domains and require additional domain-specific annotations. In contrast, recent 2D foundation models have demonstrated strong generalization and impressive zero-shot abilities, inspiring us to incorporate these characteristics from 2D models into 3D models. Therefore, we explore the use of image segmentation foundation models to automatically generate high-quality training labels for 3D segmentation models. The resulting model, Segment3D, generalizes significantly better than the models trained on costly manual 3D labels and enables easily adding new training data to further boost the segmentation performance. + +# 1 Introduction + +In this work, we propose Segment3D, a method for fine-grained class-agnostic 3D segmentation. In particular, dividing the space into coherent segments aligned with both the scene geometry and its semantics is a key challenge. This ability to accurately segment and interpret 3D scenes is fundamental for numerous downstream tasks [52,61], intelligent assistants and autonomous robots [30,64]. + +Current methods for 3D indoor-scene understanding mostly focus on semantic [35, 39, 40, 47, 53] and instance segmentation [9, 17, 31, 32, 46, 54]. These approaches, while effective on popular benchmarks [13, 15, 44, 50], have limitations stemming from their training. Primarily, they depend on extensive manually labeled 3D training sets that are both time-consuming and challenging to annotate. Importantly, their performance often deteriorates when applied to scenarios beyond their training data, limiting their effectiveness in diverse, real-world scenarios. This becomes particularly apparent under the recently emerging task of open-vocabulary 3D scene understanding [18, 27, 33, 38, 51] that aims to segment arbitrary user queries, which naturally go beyond the pre-defined set of training-set classes. Concurrently, the recent surge in foundation models, especially 2D vision-language models [25, 28, 41, 60], demonstrates remarkable potential. Trained on internet-scale data, these models exhibit an extraordinary ability to generalize, even in a zero-shot setting, to new and different input distributions. However, their application has been predominantly confined to 2D data. For instance, SAM [28] has shown impressive results in 2D image segmentation, but its applicability to 3D scene understanding remains largely under-explored. All these factors give rise to the research question: + +# How to leverage 2D foundation models for class-agnostic 3D scene segmentation without requiring manually labeled 3D data? + +Recently, SAM3D [57] has proposed a straightforward method that uses posed RGB-D images corresponding to a 3D scene. They predict segmentation masks for all input RGB-D images with SAM [28], and projected them into the 3D space. Next, through an iterative bottom-up process, the 3D masks of partial scenes are merged to derive the final 3D segmentation (see Fig. 2). However, variations in perspectives across frames can lead to conflicting segmentations, which introduces inconsistency during the merging process. Besides, the inference speed of SAM3D is slow, due to running SAM on the fly for a large number of RGB images and the cumbersome non-learned merging process. In this paper, we seek to circumvent both the slow speed and the issue of merging inconsistent segmentation across frames. + +Towards this goal, we introduce Segment3D, a clean and efficient approach that utilizes a 3D model to achieve class-agnostic, fine-grained 3D segmentation at speed. Segment3D employs a two-stage training approach that requires no hand-annotated labels. We first train our class-agnostic 3D segmentation model + +with RGB-D data, which are transformed into partial RGB-D point clouds. The supervision is provided by the segmentation masks produced by SAM and projected into 3D. While SAM can automatically generate masks for entire images, sometimes the results may suffer from over-segmentation (see Fig. 4), which is not desired in our context. Therefore, we propose a Mask Generation Module to generate high-quality, complete masks of objects for a given RGB image. Since there is readily available large-scale RGB-D data, additional data can be effortlessly integrated to further enhance segmentation performance. This pretraining stage lays the groundwork for understanding the 3D structure from 2D annotations. However, as our ultimate objective is the segmentation of full 3D scenes, we must bridge the domain gap between partial point clouds and the more comprehensive 3D point clouds obtained from 3D scanners or reconstruction techniques [14,24]. To this end, in the second stage, we fine-tune the model on full 3D point clouds in a self-supervised manner, utilizing high-confidence mask predictions from the pre-trained model as training signal. + +We demonstrate strong performance in class-agnostic segmentation on ScanNet [13, 44] and the newly released ScanNet++ [58]. Though Segment3D is trained on the indoor dataset, transferring it directly to the outdoor nuScenes dataset [4] also yields surprisingly good results. Finally, we show the use of Segment3D for improving open-vocabulary 3D instance segmentation [51]. Overall, the contributions of this paper are as follows: + +- We introduce Segment3D, a novel approach and training strategy for fine-grained class-agnostic 3D point cloud segmentation without manually annotated labels. +- We propose a Mask Generation Module that utilizes the 2D foundation model to automatically generate high-quality, complete training masks. +- We show that Segment3D demonstrates strong generalization compared to a wide range of baselines, including fully supervised methods trained on carefully annotated datasets. + +# 2 Related Work + +3D Instance Segmentation. Current models have seen a significant development over the last years, from proposal-based [56, 59], over grouping-based [9, 26, 31, 54], to recent Transformer-based [32, 46] methods. In this work, we follow the currently best-performing Transformer-based paradigm. Despite impressive advancements, a shared limitation is the dependence on costly manual ground-truth annotations. Recently, there have been efforts to automate the annotation process by bundling state-of-the-art segmentation models [55], but they are still limited to pre-defined object classes. However, in the context of open-world 3D scene understanding, the importance of semantic classification for closed-set categories has diminished. Instead, in this work, we propose a general 3D segmentation method trained on automatically generated labels. + +Foundation Models. Foundation models, particularly those that are multimodal [25,41], have revolutionized the field of AI by leveraging extensive image-text pre-training. These models can derive rich image representations guided by natural language descriptions, enabling a variety of downstream tasks [21,37,42]. Another line of work [6,36], based on self-supervised learning, employs image training and yields high-performance features directly applicable as inputs for linear classifiers. Segment Anything Model (SAM) [28] has recently advanced the performance of foundation models for image segmentation. SAM has undergone training on a diverse, high-quality dataset comprising more than 1 billion masks. This training equips SAM with the ability to generalize to novel object types and images, surpassing the scope of its observations during the training process. Additionally, SAM can generate high-quality, fine-grained masks. Our work leverages the power of SAM for pre-training a 3D segmentation model, which is later also used for generating supervision signals for fine-tuning on scene-level point clouds. + +Open-Vocabulary 3D Scene Understanding. Recently, there has been an increased interest in 3D open-vocabulary scene understanding. This new field utilizes the zero-shot recognition abilities of 2D vision-language models [25,41], enabling a more comprehensive understanding of diverse and previously unseen 3D environments [8,16,22,23,27,29,33,38,51,62,63]. PLA [16] aligns point cloud features with captions extracted from multi-view images of a scene to enable open-vocabulary recognition. OpenScene [38] distills per-pixel image features to 3D point clouds, generating point-wise scene representations co-embedded with text and image pixels in CLIP feature space. However, it mainly focuses on semantic segmentation and exhibits a limited understanding of object instances. To this end, OpenMask3D [51] predicts class-agnostic 3D instance masks and aggregates per-mask features via multi-view fusion of CLIP-based image embeddings. Nevertheless, the segmentation model it uses is trained on closed-set labels, lacking generalization to the open world. Our method contributes to this area by providing open-set class-agnostic masks, which can serve as foundational inputs for models such as OpenMask3D. + +# 3 Method + +We aim to develop a method capable of segmenting any object within a given 3D scene. Relying on existing 3D training datasets cannot accomplish this goal, as their mask annotations are limited to a predefined set of object classes [1,2,13,48]. Consequently, models trained on such datasets fail to generalize effectively to new, unseen classes. Recently, with SAM [28] showing extraordinary generalization ability in 2D segmentation, SAM3D [57] has proposed to simply project and merge SAM predictions in 3D from posed RGB-D images, as illustrated in Fig. 2. However, it suffers from inconsistent results and slow inference speed, limiting its practical applicability. + +To fully exploit SAM for consistent 3D segmentation at speed, we introduce a clean and efficient method which utilizes a 3D model to directly segment the + +![](images/f86f7851e28f780ecb81a3e158d663ff50c97018776cce971b8d5f1e6a05a28e.jpg) +Fig. 2: SAM3D vs. Segment3D. SAM3D [57] (left) merges 2D segmentation masks of RGB-D images generated by SAM to obtain 3D segmentation of entire scenes. However, conflicting segmentations across frames introduce inconsistency during the merging process. Moreover, it is slow because of extensive image inference and the cumbersome merging procedure. Instead, our Segment3D (right) utilizes a 3D model to directly segment entire 3D scenes which is clean and efficient. + +entire scene. First, we propose a Mask Generation Module to generate high-quality, complete object masks from SAM predictions (Sec. 3.1). Next, we train our 3D segmentation model with RGB-D data and the generated object masks, which are transformed into partial point clouds (Sec. 3.2). Since the final goal is to segment 3D scenes, we need to bridge the inevitable domain gap between partial RGB-D point clouds and full 3D point clouds from 3D scanners or reconstruction methods. We therefore fine-tune our model on full 3D point clouds using high-confident mask predictions from our pre-trained model (Sec. 3.3). The overall framework is illustrated in Fig. 3. + +# 3.1 Generating Pseudo Ground-Truth Masks + +Review on SAM. We first review the automatic mask generation pipeline of SAM [28]. SAM is prompted with a regular grid of $32 \times 32$ points on the RGB image. Each point predicts masks at three different granularities: whole, part, and subpart. Next, all three masks for all points are filtered by the model's predicted IoU score as well as the stability of the masks with respect to the binarization threshold. Then, the remaining masks undergo Non-Maximum Suppression (NMS). Finally, if a pixel is encompassed by multiple masks, the mask ID with the highest predicted IoU is assigned to the pixel. + +However, as shown in Fig. 4, the results obtained in this way could suffer from over-segmentation. A naive idea is to only use the masks generated by the "whole" branch. Nevertheless, we find that the "whole" branch typically generates high-quality masks for large objects while overlooking small objects in the image. In contrast, the "part" and "subpart" branches not only generate masks for parts of large objects but also predict masks for small and tiny objects. Thus, to enable our 3D model to achieve fine-grained segmentation capabilities for details and + +![](images/a5671d4a17d0178f55d40f687cb4dd4ff3a744096b4a4ba325d0b3d4c7e53496.jpg) +Stage1: Pre-Training on Partial Point Clouds +Stage2: Fine-Tuning on Full Point Clouds +Fig. 3: Method Overview. Training Segment3D involves two stages: The first stage (left) relies on largely available RGB-D image sequences and SAM $\square$ , a pre-trained foundation model for 2D image segmentation [28]. Segment3D $\square$ is pre-trained on partial RGB-D point clouds and supervised with pseudo ground-truth masks generated by Mask Generation Module (MGM) $\square$ . Due to the domain gap between partial and full point clouds, in the second stage (right), Segment3D $\square$ is fine-tuned with confident masks predicted by the pre-trained Segment3D $\square$ . + +small objects, we must retain the outputs from the "part" and "subpart" branches. The problem is how to preserve their segmentation for small and tiny objects while removing their segmentation for parts of large objects. + +Mask Generation Module (MGM). Here we introduce our Mask Generation Module (MGM). Since SAM's automatic mask generation pipeline can provide reasonably high-quality masks for a given image, we start based on it. Next, we compare the result with the output of the "whole" branch to determine which masks are parts of large objects and remove them. Specifically, for each mask in the automatic output $m_i^{\mathrm{auto}}$ , we iterate through the masks of the entire branch, calculating the intersection $m_{ij}^{\mathrm{inter}}$ between $m_i^{\mathrm{auto}}$ and $m_j^{\mathrm{whole}}$ . If $m_{ij}^{\mathrm{inter}} / m_i^{\mathrm{auto}} \geq t$ , then we consider $m_i^{\mathrm{auto}}$ to be a part of $m_j^{\mathrm{whole}}$ . Our preliminary results show that the process of extracting valid masks is highly robust to the hyper-parameter $t$ and we set it as 0.9 by default. We remove the part masks and combine the remaining masks with the output of the "whole" branch, which will serve as our training supervision. With MGM, we observe notable improvements at all object mask sizes, especially on medium-sized and large objects (see Table 6). + +# 3.2 Stage 1: Pre-Training on RGB-D Point Clouds + +In contrast to the relatively scarce availability of 3D data, there is an abundance of 2D data, particularly RGB-D images, which are readily accessible. For example, ScanNet [13] comprises merely 1513 3D scans, compared to the substantially larger collection of 2.5 million RGB-D images. Therefore, we first pre-train our 3D segmentation model on partial RGB-D point clouds. + +Training-Set Preparation. Starting from a collection of RGB-D frames, we create partial 3D point clouds and their corresponding pseudo ground-truth 3D masks. Note that the labels can be automatically obtained without manual effort. + +![](images/e931db91a5ab2cba7dfd94ce873d48d635125c3f64094c8e679b9324cdb26036.jpg) +Automatic-SAM Generated Masks + +![](images/69a033e5891fced37f67f592bb79ceaeeb2344d035c1e81b3881aa49f045ca13.jpg) +Fig. 4: Automatic-SAM Generated (left) vs. MGM Generated (right) Masks from ScanNet [13]. The former exhibits over-segmentation, while the latter does not. + +![](images/d4e3faaf194ed016d08c2395d4390adc664698488d391d639d8560861534abcd.jpg) +MGM rated Masks + +![](images/f143a5be6e9ef1ffc065014faf122eb7e7f09aa654e81953cea21d48573234ec.jpg) +Fig. 5: Partial (center) vs. Full (right) Point Clouds from ScanNet [13]. Full point clouds are more complete and exhibit fewer occlusions due to reconstruction over multiple viewpoints. + +![](images/4da37cbae5a0269038b2b95a3af6fe4c3a9130f6185a19847f5a6f98d73d0ec6.jpg) + +For each frame in a large RGB-D dataset, we first transform the 2D depth map to a partial 3D point cloud. To do so, we need to know the intrinsic matrix $\mathbf{K} \in \mathbb{R}^{3 \times 3}$ and extrinsic matrix $\mathbf{T} = [\mathbf{R} \mathbf{t}] \in \mathbb{R}^{3 \times 4}$ . For each pixel $\mathbf{p} = (u, v)$ , we can transform it with its depth value $D_{\mathbf{p}}$ into a 3D point $\mathbf{P}$ in world coordinates as follows: + +$$ +\mathbf {P} = \mathbf {R} ^ {\top} \cdot \left(D _ {\mathbf {p}} \cdot \mathbf {K} ^ {- 1} \cdot \tilde {\mathbf {p}}\right) - \mathbf {R} ^ {\top} \mathbf {t}, \tag {1} +$$ + +where $\tilde{\mathbf{p}}$ is the homogeneous coordinate of $\mathbf{p}$ . By applying Eq. (1) to all pixels in the depth map $D$ and associating their per-pixel RGB value, we obtain the input partial 3D point cloud. Next, we obtain the pseudo ground-truth 3D segmentation masks for this point cloud with the 2D masks generated by MGM in the same way. Since we know the one-to-one mapping between 2D pixel and 3D points from Eq. (1), we directly obtain the per-point 3D mask labels. + +Model Architecture. We use a model inspired by Mask3D [46] to train a class-agnostic 3D segmentation model. The model comprises a sparse convolutional backbone derived from MinkowskiUNet [12] and a transformer decoder, as in MaskFormer [10, 11]. We adopt a set of queries to represent the masks, each of which is initialized with a positional embedding. Specifically, we select query positions with furthest point sampling (FPS) and use their Fourier positional encodings as the query embeddings. Leveraging the transformer decoder, all the mask queries are refined by progressively attending to point cloud features across multiple scales in parallel. Each mask query is subsequently decoded into both a mask feature and a binary label to predict whether the given query corresponds to a valid object or not. By computing cosine similarity scores between a mask feature and all point features within the point cloud, a heatmap is generated over + +at the point cloud. This heatmap is input to a sigmoid function, and thresholded at 0.5, resulting in the final binary mask. + +Training with SAM Generated Masks. We supervise the model with two losses: the per-point supervision loss $\mathcal{L}_{\mathrm{mask}}$ and a per-query supervision loss $\mathcal{L}_{\mathrm{obj}}$ . The loss $\mathcal{L}_{\mathrm{mask}}$ enables learning a foreground-background segmentation for each mask, and is composed of a dice loss $\mathcal{L}_{\mathrm{dice}}$ [34] and a binary cross-entropy loss $\mathcal{L}_{\mathrm{ce}}$ for each point. The $\mathcal{L}_{\mathrm{obj}}$ is a binary classification loss that indicates whether a query represents a valid "object" or "no object". This mechanism allows for the prediction of a variable number of masks, depending on the underlying scene content and geometry. Following prior work [5, 10, 46], we first adopt bipartite graph matching to establish correspondences between the set of predicted masks and the set of target masks provided by MGM as described before. If the predicted instance finds a matching target mask, then we assign it an "object" label; conversely, if there is no match, we assign "no object". In summary, we optimize the following losses: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {m a s k}} + \lambda_ {\text {o b j}} \mathcal {L} _ {\text {o b j}}, \tag {2} +$$ + +$$ +\mathcal {L} _ {\text {m a s k}} = \lambda_ {\text {d i c e}} \mathcal {L} _ {\text {d i c e}} + \lambda_ {\text {c e}} \mathcal {L} _ {\text {c e}}, \tag {3} +$$ + +where $\lambda_{*}$ are hyperparameters that balance the contribution of each component in the loss. The binary classification loss $\mathcal{L}_{\mathrm{obj}}$ is applied to all queries, while the mask loss $\mathcal{L}_{\mathrm{mask}}$ is specifically applied to masks labeled as "object". + +# 3.3 Stage 2: Self-Supervised Scene Fine-Tuning + +After Stage 1, we obtain a class-agnostic 3D segmentation model by pre-training solely on RGB-D images and automatically generated labels from MGM. However, a fundamental domain gap persists between partial point clouds derived from RGB-D data and full point clouds acquired through 3D scanners or reconstruction methods [14,24] (See Fig. 5). This gap exists mostly because of object occlusions, but also due to challenges of depth cameras to capture dark or reflective surfaces from a single viewpoint. Hence, depending solely on RGB-D frames for training a 3D segmentation model intended for full 3D scans proves inadequate. Therefore, we propose to further fine-tune our model on scene-level full 3D point clouds. The key idea to obtain 3D mask annotations for training on full point clouds is to use selected, high-confidence masks generated by the pre-trained model itself. Note that this stage requires no manual labels on full 3D scenes and proves essential for the performance of Segment3D (see Tab. 5). + +Confidence-Based Mask Generation. Next, we outline the process of generating the supervision signal for the fine-tuning stage. The pre-trained model processes point clouds independently of their nature, be it partial or full. Therefore, when presented with a full 3D point cloud, the pre-trained 3D model produces a set of masks, each with a binary classification (valid or not) and a heatmap over all points, just as in Sec. 3.2. To assess the quality of the predicted masks, we compute a confidence score based on the confidence map $\sigma(\mathbf{h})$ , where $\mathbf{h}$ is the predicted heatmap and $\sigma$ is the sigmoid function. We then compute the average + +confidence of those points for which $\sigma(\mathbf{h}) > 0.5$ as the confidence score of the predicted mask, denoted as $c_{\mathrm{mask}}$ . We also consider the classification as a valid object and use the probability from the binary classification assigned to the "object" category as the confidence score, denoted as $c_{\mathrm{obj}}$ . The final confidence score for each predicted mask is then the product of the two scores, $c = c_{\mathrm{mask}} \cdot c_{\mathrm{obj}}$ . For fine-tuning our 3D segmentation model in Stage 2, we select the most confidently predicted masks above a threshold $\tau_c$ . + +Training with the High-Confidence Generated Masks. For fine-tuning, we follow the same procedure as before and use $\mathcal{L}_{\mathrm{mask}}$ as defined in Eq. 3. In contrast to the pre-training stage, the binary classification loss $\mathcal{L}_{\mathrm{obj}}$ , responsible for categorizing queries into valid or invalid, is omitted. Since we only select masks with high confidence for supervision, it can happen that some objects in the scene have no assigned ground truth mask. In such instances, deeming a correctly predicted mask for those objects as invalid would be detrimental. Table 5 illustrates the efficacy of the self-supervised fine-tuning process in comparison to pre-training alone. + +# 4 Experiments + +We firstly evaluate Segment3D on indoor ScanNet++ [58], ScanNet200 [13,44] and outdoor nuScenes [4] in a class-agnostic segmentation setting (Sec. 4.1). We then show the advantage of our method in segmenting small objects, and provide analysis to understand the importance of fine-tuning, the effectiveness of Mask Generation Module and the potential of training on more data (Sec. 4.2). Finally, we demonstrate its application for the task of open-set 3D instance segmentation as proposed in OpenMask3D [51] (Sec. 4.3). + +# 4.1 Comparing with State-of-the-Art Methods + +Datasets. We first evaluate on the recently released ScanNet++ [58] which includes high-resolution 3D scans captured at sub-millimeter precision, and fine-grained annotations covering objects of varying sizes. Due to its comprehensive data annotation, all of our analytical experiments in Sec. 4.2 are also based on it. Next, we adopt the ScanNet200 [13, 44], which builds upon the classic ScanNet [13] dataset and extends the semantic annotations to 200 classes. Then, we test our model on the nuScenes [4] outdoor dataset. For training our model, we employ ScanNet [13, 44], which is collected through a lightweight RGB-D scanning process. See the appendix for the implementation details of our method. + +Methods in Comparison. We compare with a wide range of prior art methods from different categories. Mask3D [46] is a state-of-the-art fully-supervised method trained on manually annotated 3D segmentation masks. Segment3D has the same backbone as Mask3D but instead of training on manually annotated 3D masks, it learns from automatically generated masks. Felzenszwalb et al. [20] proposed a graph-based method for segmentation which operates directly on + +Table 1: Segmentation Scores on ScanNet++ [58]. The metric is average precision (AP) on the validation split. We report scores with and without post-processing (more details in Sec. 4.1). For reference, we show results on Mask3D trained on manual ScanNet200 labels. + +
ModelAvg. Inference Time/swithout post-processingwith post-processing
AP\( AP_{50} \)\( AP_{25} \)AP\( AP_{50} \)\( AP_{25} \)
Mask3D [46]0.78.715.527.214.321.329.9
SAM3D [57]386.73.99.322.18.416.130.0
Felzenszwalb et al. [20]12.65.811.627.2---
Segment3D (Ours)0.713.023.838.320.230.942.7
+ +the 3D geometry. UnScene3D [45] leverages self-supervised color and geometry features to obtain coarse segmentation and refines it through self-training. SAM3D [57] merges the segmentation masks of RGB-D images generated by SAM [28] to obtain the 3D segmentation result. + +Metrics. We report average precision (AP) scores at IoU thresholds of $25\%$ , $50\%$ , and averaged over the range [0.5:0.95:05] between predicted and ground truth masks. Consistent with common practices in the field [47, 49, 54], we also report scores after post-processing. This involves smoothing the predicted masks through graph-based oversegmentation [20], and splitting distant parts of the same mask via connected component clustering DBSCAN [19]. We also report the average inference time for these methods on ScanNet++ tested on A100. + +Results on ScanNet++. Scores are reported in Table 1. Segment3D outperforms all previous methods by at least +4.3 AP and up to +12.7 AP $_{25}$ . Notably, we achieve such improvements without any ground-truth mask annotation contrary to Mask3D trained on ScanNet200. In general, the performance of Mask3D (and other fully-supervised methods) depends on the quality of the annotated training dataset; often the manual annotation of small objects (pens, cellphones) or other fine-details is challenging. Instead, Segment3D relies on automatically generated high-quality masks from MGM using the 2D foundation model, which can capture fine-grained details without human annotation effort. Table 4 highlights that Segment3D excels particularly in predicting the more challenging small object masks. Furthermore, our method significantly outperforms SAM3D, indicating that Segment3D can circumvent the noise that arises during the merging process of SAM3D. Additionally, As shown in Table 1, Segment3D is three-order-of-magnitude faster than SAM3D in inference speed. + +Fig. 6 shows segmentation results of several representative examples on ScanNet++. As can be seen, the scenes are quite diverse, presenting multiple challenges such as clutter and a wide range of mask sizes. Despite these challenges, our model predicts quite accurate and well-localized segmentation masks. For example, compared to Mask3D, our method is able to segment the finer-grained objects on top of the bed and the shelf. The masks of the computer screen and the chair in front of it are also less fragmented than predictions of the baselines. + +Table 2: Segmentation Scores on ScanNet200 [13, 44]. The evaluation metric is average precision on the val split. + +
ModelAPAP50
Mask3D [46]34.143.1
Felzenszwalb et al. [20]6.112.1
UnScene3D [45]15.932.2
SAM3D [57]19.032.5
Segment3D (Ours)27.739.8
+ +Table 3: Segmentation Scores on nuScenes [4]. The evaluation metric is average precision on the val split. + +
ModelGT LabelsAP50AP25
zero-shot test
Mask3D×15.325.5
Segment3D×37.848.0
fine-tuned on nuScenes
Mask3D✓46.557.8
Segment3D×42.555.5
+ +Results on ScanNet200. As illustrated in Table 2, Segment3D outperforms all baselines that require no manual 3D labels, and even shows competitive results with Mask3D, which is trained on ScanNet200. Compared to Unscene3D, which leverages self-supervised color and geometry features to obtain coarse segmentation, we achieve superior results by using the Mask Generation Module to generate high-quality complete masks. + +Results on Outdoor Data. The performance on nuScenes is demonstrated in Table 3. We observe that even though Segment3D is trained using purely indoor data, applying it directly to outdoor data produces surprisingly good results. This indicates the stronger generalizability of Segment3D over the fully-supervised method Mask3D when transferring from indoor to outdoor scenes. Furthermore, we report scores with fine-tuning on nuScenes which leads to overall improved results. Similar to Table 2, Mask3D performs best, however, Segment3D achieves very competitive performance without any hand-labeled training data. Refer to the appendix for additional results on STPLS3D [7] and Paris-Lille-3D [43] datasets. + +# 4.2 Analysis Experiments + +Performance on Different Mask Sizes. We proceed with an analysis of the performance of our Segment3D and Mask3D (the best-performing baseline) on object masks of various sizes. We categorize the size of masks based on the number of points they contain (small: [0, 2k], medium: [2k, 15k] large: [15k, $\infty$ ]). The results are reported in Table 4. Our method yields significantly improved + +Table 4: Segmentation Scores on Different Mask Sizes. Segmented3D improves over Mask3D, especially on small object masks. Details are in Sec. 4.2. + +
Mask SizeLargeMediumSmall
ScanNet++APAP50APAP50APAP50
Mask3D13.825.113.623.31.84.3
Segment3D18.331.916.629.88.417.3
(+4.5)(+6.8)(+3.0)(+6.5)(+6.6)(+13.0)
+ +![](images/4a6319bc0d817b372e927c3ebd7431955843bfdfb283c6eb2c5ebf05c64311c1.jpg) +Fig. 6: Qualitative Results on ScanNet++. From top to bottom, we show the colored input 3D scenes, the segmentation masks predicted by SAM3D [57], Mask3D [46], our Segment3D and the ground truth 3D mask annotations. + +segmentation results on objects of various sizes, especially on small objects. This confirms our intuition that Mask3D performs poorly on small-sized object masks as those are typically harder to manually annotate. In contrast, Segment3D utilizes masks generated by MGM with the foundation model as supervision, capturing fine-grained scene details. This showcases the usefulness of foundation models and raises the question if manually labeled large-scale 3D datasets are necessary for training 3D point cloud segmentation models. + +The Importance of Two-Stage Training. Next, we compare the performance of Segment3D pre-trained solely on partial RGB-D point clouds (Stage 1) and with additional fine-tuning on full point clouds (Stage 2). Scores are reported in Table 5. The additional fine-tuning stage almost doubles the segmentation performance of our model on the most challenging AP metric. By training with the predicted high-confidence masks, Segment3D effectively reduces the inherent + +Table 5: The Importance of Two-Stage Training. Fine-tuning on full point clouds supervised by confident predicted 3D masks significantly surpasses pre-training on projected 2D SAM masks alone. + +
Training StagesAPAP50AP25
Pre-Training (Stage 1)6.814.230.6
+ Fine-Tuning (Stage 2)13.0 (+6.2)23.8 (+9.6)38.3 (+7.7)
+ +Table 6: Effectiveness of Mask Generation Module. Compared to the model trained with masks generated by Automatic-SAM, our model shows improvement across all object mask sizes, particularly for medium-sized and large objects. + +
Mask SizeLargeMediumSmall
Masks Generated byAPAP50APAP50APAP50
Automatic-SAM16.830.614.227.27.516.2
MGM18.331.916.629.88.417.3
(+1.5)(+1.3)(+2.4)(+2.6)(+0.9)(+1.1)
+ +domain gap between the partial point clouds derived from RGB-D images and full 3D point clouds. + +The Effectiveness of Mask Generation Module. To verify whether the MGM enhances the model's performance by reducing the over-segmentation issue and providing more complete masks, we conducted evaluations on objects of different sizes. As shown in Table 6, we can see that compared to the model trained with masks generated by Automatic-SAM, our model notably improves the performance of large and medium-sized objects which demonstrates the effectiveness of MGM. + +Pre-Training with Additional Data. Since RGB-D data is available in abundance and masks can be automatically generated, it is natural to ask if pretraining on additional masks will further improve the overall performance. To that end, we perform a first experiment where we select additional frames from the training set of the ScanNet++ dataset (to increase the variety of training data). Table 7 shows an impressive performance boost of +4.1 AP. For future work, considering that this improvement is obtained simply by adding more automatically generated masks to the pre-training, our approach seems promising to train on even more data. It could even be plausible to train on internet images combined with monocular depth estimation, such as ZoeDepth [3], to compute the partial point clouds. + +# 4.3 Application: Open-Set Scene Understanding + +A real-world application of our class-agnostic 3D segmentation method is open-vocabulary 3D scene understanding, as implemented recently by OpenMask3D [51]. Given a 3D scene, a user can search for arbitrary objects via text prompts (see Fig. 7). A core component of OpenMask3D is Mask3D, which segments the + +Table 7: Results with Additional Training Data. The performance is further boosted with more training data. + +
Pre-Training DataAPAP50
ScanNet13.023.8
ScanNet, ScanNet++17.1 (+4.1)26.7 (+2.9)
+ +Table 8: Open-Set 3D Scene Understanding. The evaluation metric is average precision with an IoU threshold of $50\%$ . + +
SegmentorScanNet++Replica
Mask3D [46]15.018.0
Segment3D (Ours)18.5 (+3.5)18.4 (+0.4)
+ +![](images/4ae3e9804388c5cc2a75354f70b7afaaae5bf91bfc44cffa88e504391e2b8334.jpg) +"aback eraser" + +![](images/9e756c4fe06ba932d5c39e49f1b57229a930481421b91c7320fc831f7cbddd62.jpg) +"paper on the laptop" +Fig. 7: Open-Set 3D Object Retrieval Results. Given a text prompt (bottom), OpenMask3D [51] finds the corresponding object masks in a given 3D scene (top). We adapt OpenMask3D and use fine-grained masks from our Segment3D method. We show 3D reconstructions and RGB images for better visualization (top left corner). + +![](images/f35f83da57a01168c4ea50e9a8fbbf7999838c0387258d7e92ac5a0bb3f83d8c.jpg) +"scissors" + +scene into a set of object masks. Since Mask3D is trained on the closed-set of labeled annotations from ScanNet, its masks are not truly class-agnostic or open-vocabulary. We therefore replace Mask3D with our class-agnostic Segment3D. We evaluate on the closed-set labels of $\mathrm{ScanNet}++$ and Replica in Table 8 and report an improvement of up to $+3.5\mathrm{AP}_{50}$ . Since Segment3D can accurately generate fine-grained masks in a class-agnostic manner, we are able to retrieve objects such as "eraser", "paper" and "scissors", which are small in scale. + +# 5 Conclusion + +We have presented Segment3D, a simple, yet powerful class-agnostic 3D segmentation model. Segment3D employs a two-stage training framework that requires no manually annotated labels. We propose a Mask Generation Module to automatically generate high-quality, complete training masks with the 2D foundation model. The model is first pre-trained on RGB-D data with predicted 2D masks, then fine-tuned on full point clouds. Segment3D shows strong generalization and even outperforms existing 3D segmentation models that rely on hand-labeled 3D training scenes. Indeed, this raises the question of whether hand-labeled 3D training datasets are as essential as they were once thought to be. + +Acknowledgement. Gao Huang is supported in part by the National Natural Science Foundation of China under grants (62321005 and 42327901). Francis Engelmann is partially supported by an ETH AI Center postdoctoral research fellowship and an ETH Zurich Career Seed Award. Songyou Peng is supported by an Innosuisse funding (100.567 IP-ICT), and Ayça Takmaz is supported by an Innosuisse grant (48727.1 IP-ICT). + +# References + +1. Armeni, I., Sener, O., Zamir, A.R., Jiang, H., Brilakis, I., Fischer, M., Savarese, S.: 3D Semantic Parsing of Large-Scale Indoor Spaces. In: CVPR (2016) 4 +2. Baruch, G., Chen, Z., Dehghan, A., Dimry, T., Feigin, Y., Fu, P., Gebauer, T., Joffe, B., Kurz, D., Schwartz, A., et al.: ARKitScenes: A Diverse Real-World Dataset for 3D Indoor Scene Understanding using Mobile RGB-D Data. arXiv preprint arXiv:2111.08897 (2021) 4 +3. Bhat, S.F., Birkl, R., Wofk, D., Wonka, P., Müller, M.: ZoeDepth: Zero-Shot Transfer by Combining Relative and Metric Depth. arXiv preprint arXiv:2302.12288 (2023) 13 +4. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuScenes: A Multimodal Dataset for Autonomous Driving. In: CVPR (2020) 3, 9, 11 +5. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-End Object Detection with Transformers. In: ECCV (2020) 8 +6. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging Properties in Self-Supervised Vision Transformers. In: ICCV (2021) 4 +7. Chen, M., Hu, Q., Yu, Z., Thomas, H., Feng, A., Hou, Y., McCullough, K., Ren, F., Soibelman, L.: STPLS3D: A Large-Scale Synthetic and Real Aerial Photogrammetry 3D Point Cloud Dataset. arXiv preprint arXiv:2203.09065 (2022) 11 +8. Chen, R., Liu, Y., Kong, L., Zhu, X., Ma, Y., Li, Y., Hou, Y., Qiao, Y., Wang, W.: CLIP2Scene: Towards Label-Efficient 3D Scene Understanding by CLIP. In: CVPR (2023) 4 +9. Chen, S., Fang, J., Zhang, Q., Liu, W., Wang, X.: Hierarchical Aggregation for 3D Instance Segmentation. In: ICCV (2021) 2, 3 +0. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-Attention Mask Transformer for Universal Image Segmentation. In: CVPR (2022) 7, 8 +1. Cheng, B., Schwing, A., Kirillov, A.: Per-Pixel Classification is Not All You Need for Semantic Segmentation. In: NeurIPS (2021) 7 +2. Choy, C., Gwak, J., Savarese, S.: 4D Spatio-Temporal ConvNets: Minkowski Convolutional Neural Networks. In: CVPR (2019) 7 +3. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: ScanNet: Richly-Annotated 3D Reconstructions of Indoor Scenes. In: CVPR (2017) 2, 3, 4, 6, 7, 9, 11 +4. Dai, A., Nießner, M., Zollhöfer, M., Izadi, S., Theobalt, C.: BundleFusion: Real-time Globally Consistent 3D Reconstruction using On-the-fly Surface Re-integration. TOG (2017) 3, 8 +5. Delitzas, A., Takmaz, A., Tombari, F., Sumner, R., Pollefeys, M., Engelmann, F.: SceneFun3D: Fine-Grained Functionality and Affordance Understanding in 3D Scenes. In: CVPR (2024) 2 +6. Ding, R., Yang, J., Xue, C., Zhang, W., Bai, S., Qi, X.: PLA: Language-Driven Open-Vocabulary 3D Scene Understanding. In: CVPR (2023) 4 +7. Engelmann, F., Bokeloh, M., Fathi, A., Leibe, B., Nießner, M.: 3D-MPA: Multi-Proposal Aggregation for 3D Semantic Instance Segmentation. In: CVPR (2020) 2 +8. Engelmann, F., Manhardt, F., Niemeyer, M., Tateno, K., Tombari, F.: OpenNeRF: Open Set 3D Neural Scene Segmentation with Pixel-Wise Features and Rendered Novel Views. In: ICLR (2024) 2 + +19. Ester, M., Kriegel, H.P., Sander, J., Xu, X., et al.: A Density-Based Algorithm for Discovering Clusters in Large Spatial Databases with Noise. In: KDD (1996) 10 +20. Felzenszwalb, P.F., Huttenlocher, D.P.: Efficient Graph-Based Image Segmentation. IJCV (2004) 9, 10, 11 +21. Gu, X., Lin, T.Y., Kuo, W., Cui, Y.: Open-Vocabulary Object Detection via Vision and Language Knowledge Distillation. In: ICLR (2022) 4 +22. Ha, H., Song, S.: Semantic Abstraction: Open-World 3D Scene Understanding from 2D Vision-Language Models. In: CoRL (2022) 4 +23. Huang, T., Dong, B., Yang, Y., Huang, X., Lau, R.W., Ouyang, W., Zuo, W.: CLIP2Point: Transfer CLIP to Point Cloud Classification with Image-Depth Pre-Training. In: ICCV (2023) 4 +24. Izadi, S., Kim, D., Hilliges, O., Molyneaux, D., Newcombe, R., Kohli, P., Shotton, J., Hodges, S., Freeman, D., Davison, A., et al.: KinectFusion: Real-time 3D Reconstruction and Interaction using a Moving Depth Camera. In: UIST (2011) 3, 8 +25. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up Visual and Vision-Language Representation Learning with Noisy Text Supervision. In: ICML (2021) 2, 4 +26. Jiang, L., Zhao, H., Shi, S., Liu, S., Fu, C.W., Jia, J.: PointGroup: Dual-Set Point Grouping for 3D Instance Segmentation. In: CVPR (2020) 3 +27. Kerr, J., Kim, C.M., Goldberg, K., Kanazawa, A., Tancik, M.: LERF: Language Embedded Radiance Fields. In: ICCV (2023) 2, 4 +28. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment Anything. In: ICCV (2023) 1, 2, 4, 5, 6, 10 +29. Kobayashi, S., Matsumoto, E., Sitzmann, V.: Decomposing Nerf for Editing via Feature Field Distillation. In: NeurIPS (2022) 4 +30. Lemke, O., Bauer, Z., Zurbrügg, R., Pollefeys, M., Engelmann, F., Blum, H.: Spotcompose: A framework for open-vocabulary object retrieval and drawer manipulation in point clouds. In: 2nd Workshop on Mobile Manipulation and Embodied Intelligence at ICRA 2024 (2024) 2 +31. Liang, Z., Li, Z., Xu, S., Tan, M., Jia, K.: Instance Segmentation in 3D Scenes using Semantic Superpoint Tree Networks. In: CVPR (2021) 2, 3 +32. Lu, J., Deng, J., Wang, C., He, J., Zhang, T.: Query Refinement Transformer for 3D Instance Segmentation. In: ICCV (2023) 2, 3 +33. Lu, Y., Xu, C., Wei, X., Xie, X., Tomizuka, M., Keutzer, K., Zhang, S.: Open-Vocabulary Point-Cloud Object Detection without 3D Annotation. In: CVPR (2023) 2, 4 +34. Milletari, F., Navab, N., Ahmadi, S.A.: V-Net: Fully Convolutional Neural Networks for Volumetric Medical Image Segmentation. In: 3DV (2016) 8 +35. Nekrasov, A., Schult, J., Litany, O., Leibe, B., Engelmann, F.: Mix3D: Out-of-context data augmentation for 3D scenes. In: 3DV (2021) 2 +36. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: DINOv2: Learning Robust Visual Features without Supervision. arXiv preprint arXiv:2304.07193 (2023) 4 +37. Patashnik, O., Wu, Z., Shechtman, E., Cohen-Or, D., Lischinski, D.: StyleCLIP: Text-Driven Manipulation of StyleGAN Imagery. In: ICCV (2021) 4 +38. Peng, S., Genova, K., Jiang, C., Tagliasacchi, A., Pollefeys, M., Funkhouser, T., et al.: OpenScene: 3D Scene Understanding with Open Vocabularies. In: CVPR (2023) 2, 4 + +39. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation. In: CVPR (2017) 2 +40. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In: NeurIPS (2017) 2 +41. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning Transferable Visual Models from Natural Language Supervision. In: ICML (2021) 2, 4 +42. Rao, Y., Zhao, W., Chen, G., Tang, Y., Zhu, Z., Huang, G., Zhou, J., Lu, J.: DenseCLIP: Language-Guided Dense Prediction with Context-Aware Prompting. In: CVPR (2022) 4 +43. Roynard, X., Deschaud, J.E., Goulette, F.: Paris-Lille-3D: A Large and High-Quality Ground-Truth Urban Point Cloud Dataset for Automatic Segmentation and Classification. *IJRR* (2018) 11 +44. Rozenberszki, D., Litany, O., Dai, A.: Language-Grounded Indoor 3D Semantic Segmentation in the Wild. In: ECCV (2022) 2, 3, 9, 11 +45. Rozenberszki, D., Litany, O., Dai, A.: UnScene3D: Unsupervised 3D Instance Segmentation for Indoor Scenes. In: CVPR (2024) 10, 11 +46. Schult, J., Engelmann, F., Hermans, A., Litany, O., Tang, S., Leibe, B.: Mask3D: Mask Transformer for 3D Semantic Instance Segmentation. In: ICRA (2023) 1, 2, 3, 7, 8, 9, 10, 11, 12, 14 +47. Schult, J., Engelmann, F., Kontogianni, T., Leibe, B.: DualConvMesh-Net: Joint Geodesic and Euclidean Convolutions on 3D Meshes. In: CVPR (2020) 2, 10 +48. Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur-Artal, R., Ren, C., Verma, S., et al.: The Replica Dataset: A Digital Replica of Indoor Spaces. arXiv preprint arXiv:1906.05797 (2019) 4 +49. Sun, J., Qing, C., Tan, J., Xu, X.: Superpoint Transformer for 3D Scene Instance Segmentation. In: AAAI (2023) 10 +50. Sun, T., Hao, Y., Huang, S., Savarese, S., Schindler, K., Pollefeys, M., Armeni, I.: Nothing stands still: A spatiotemporal benchmark on 3d point cloud registration under large geometric and temporal change. arXiv preprint arXiv:2311.09346 (2023) 2 +51. Takmaz, A., Fedele, E., Sumner, R.W., Pollefeys, M., Tombari, F., Engelmann, F.: OpenMask3D: Open-Vocabulary 3D Instance Segmentation. In: NeurIPS (2023) 2, 3, 4, 9, 13, 14 +52. Takmaz, A., Schult, J., Kaftan, I., Akçay, M., Leibe, B., Sumner, R., Engelmann, F., Tang, S.: 3D Segmentation of Humans in Point Clouds with Synthetic Data. In: ICCV (2023) 2 +53. Thomas, H., Qi, C.R., Deschaud, J.E., Marcotegui, B., Goulette, F., Guibas, L.J.: KPCov: Flexible and Deformable Convolution for Point Clouds. In: ICCV (2019) 2 +54. Vu, T., Kim, K., Luu, T.M., Nguyen, T., Yoo, C.D.: SoftGroup for 3D Instance Segmentation on Point Clouds. In: CVPR (2022) 2, 3, 10 +55. Weder, S., Blum, H., Engelmann, F., Pollefeys, M.: LabelMaker: Automatic Semantic Label Generation from RGB-D Trajectories. In: 3DV (2024) 3 +56. Yang, B., Wang, J., Clark, R., Hu, Q., Wang, S., Markham, A., Trigoni, N.: Learning Object Bounding Boxes for 3D Instance Segmentation on Point Clouds. In: NeurIPS (2019) 3 +57. Yang, Y., Wu, X., He, T., Zhao, H., Liu, X.: SAM3D: Segment Anything in 3D Scenes. In: ICCVW (2023) 2, 4, 5, 10, 11, 12 +58. Yeshwanth, C., Liu, Y.C., Nießner, M., Dai, A.: ScanNet++: A High-Fidelity Dataset of 3D Indoor Scenes. In: ICCV (2023) 3, 9, 10 + +59. Yi, L., Zhao, W., Wang, H., Sung, M., Guibas, L.J.: GSPN: Generative Shape Proposal Network for 3D Instance Segmentation in Point Cloud. In: CVPR (2019) 3 +60. Yue, Y., Das, A., Engelmann, F., Tang, S., Lenssen, J.: Improving 2D Feature Representations by 3D-Aware Fine-Tuning. In: ECCV (2024) 2 +61. Yue, Y., Kontogianni, T., Schindler, K., Engelmann, F.: Connecting the Dots: Floorplan Reconstruction Using Two-Level Queries. In: CVPR (2023) 2 +62. Zeng, Y., Jiang, C., Mao, J., Han, J., Ye, C., Huang, Q., Yeung, D.Y., Yang, Z., Liang, X., Xu, H.: CLIP2: Contrastive Language-Image-Point Pretraining from Real-World Point Cloud Data. In: CVPR (2023) 4 +63. Zhang, R., Guo, Z., Zhang, W., Li, K., Miao, X., Cui, B., Qiao, Y., Gao, P., Li, H.: PointCLIP: Point Cloud Understanding by CLIP. In: CVPR (2022) 4 +64. Zurbrugg, R., Liu, Y., Engelmann, F., Kumar, S., Hutter, M., Patil, V., Yu, F.: ICGNet: A Unified Approach for Instance-Centric Grasping. In: ICRA (2024) 2 \ No newline at end of file diff --git a/2024/Segment3D_ Learning Fine-Grained Class-Agnostic 3D Segmentation without Manual Labels/images.zip b/2024/Segment3D_ Learning Fine-Grained Class-Agnostic 3D Segmentation without Manual Labels/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2fb692e0ce7e72f9dc9f87b6d5e41ffafc3cf428 --- /dev/null +++ b/2024/Segment3D_ Learning Fine-Grained Class-Agnostic 3D Segmentation without Manual Labels/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:87fb96d8e35088d15ae6703761465c58ce3761f483c35ff25deb9aa8cb53b8bb +size 574005 diff --git a/2024/Segment3D_ Learning Fine-Grained Class-Agnostic 3D Segmentation without Manual Labels/layout.json b/2024/Segment3D_ Learning Fine-Grained Class-Agnostic 3D Segmentation without Manual Labels/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..11aec261fe4940d6b72b1f6a3210403352d67788 --- /dev/null +++ b/2024/Segment3D_ Learning Fine-Grained Class-Agnostic 3D Segmentation without Manual Labels/layout.json @@ -0,0 +1,9068 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 182, + 111, + 431, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 111, + 431, + 163 + ], + "spans": [ + { + "bbox": [ + 182, + 111, + 431, + 163 + ], + "type": "text", + "content": "Segment3D: Learning Fine-Grained Class-Agnostic 3D Segmentation without Manual Labels" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 133, + 185, + 479, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 185, + 479, + 210 + ], + "spans": [ + { + "bbox": [ + 133, + 185, + 479, + 210 + ], + "type": "text", + "content": "Rui Huang" + }, + { + "bbox": [ + 133, + 185, + 479, + 210 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 133, + 185, + 479, + 210 + ], + "type": "text", + "content": ", Songyou Peng" + }, + { + "bbox": [ + 133, + 185, + 479, + 210 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 133, + 185, + 479, + 210 + ], + "type": "text", + "content": ", Ayça Takmaz" + }, + { + "bbox": [ + 133, + 185, + 479, + 210 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 133, + 185, + 479, + 210 + ], + "type": "text", + "content": ", Federico Tombari" + }, + { + "bbox": [ + 133, + 185, + 479, + 210 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 133, + 185, + 479, + 210 + ], + "type": "text", + "content": ", Marc Pollefeys" + }, + { + "bbox": [ + 133, + 185, + 479, + 210 + ], + "type": "inline_equation", + "content": "^{2,4}" + }, + { + "bbox": [ + 133, + 185, + 479, + 210 + ], + "type": "text", + "content": ", Shiji Song" + }, + { + "bbox": [ + 133, + 185, + 479, + 210 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 133, + 185, + 479, + 210 + ], + "type": "text", + "content": ", Gao Huang" + }, + { + "bbox": [ + 133, + 185, + 479, + 210 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 133, + 185, + 479, + 210 + ], + "type": "text", + "content": ", and Francis Engelmann" + }, + { + "bbox": [ + 133, + 185, + 479, + 210 + ], + "type": "inline_equation", + "content": "^{2,3}" + } + ] + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 134, + 214, + 222, + 287 + ], + "blocks": [ + { + "bbox": [ + 134, + 214, + 222, + 287 + ], + "lines": [ + { + "bbox": [ + 134, + 214, + 222, + 287 + ], + "spans": [ + { + "bbox": [ + 134, + 214, + 222, + 287 + ], + "type": "image", + "image_path": "60d9244e1dc0824a5af4c5e3ed84654cf14df2985f7b03550f194dbe0c75c040.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 228, + 218, + 380, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 218, + 380, + 255 + ], + "spans": [ + { + "bbox": [ + 228, + 218, + 380, + 255 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 228, + 218, + 380, + 255 + ], + "type": "text", + "content": "Tsinghua University " + }, + { + "bbox": [ + 228, + 218, + 380, + 255 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 228, + 218, + 380, + 255 + ], + "type": "text", + "content": "ETH Zurich " + }, + { + "bbox": [ + 228, + 218, + 380, + 255 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 228, + 218, + 380, + 255 + ], + "type": "text", + "content": "Google " + }, + { + "bbox": [ + 228, + 218, + 380, + 255 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 228, + 218, + 380, + 255 + ], + "type": "text", + "content": "Microsoft *Corr. Author https://segment3d.github.io" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 394, + 215, + 481, + 287 + ], + "blocks": [ + { + "bbox": [ + 394, + 215, + 481, + 287 + ], + "lines": [ + { + "bbox": [ + 394, + 215, + 481, + 287 + ], + "spans": [ + { + "bbox": [ + 394, + 215, + 481, + 287 + ], + "type": "image", + "image_path": "4a65f75423c7b6cc674929c485768bfeadcf11d3c765b2ccad32071a2e91278f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 135, + 281, + 222, + 354 + ], + "blocks": [ + { + "bbox": [ + 135, + 281, + 222, + 354 + ], + "lines": [ + { + "bbox": [ + 135, + 281, + 222, + 354 + ], + "spans": [ + { + "bbox": [ + 135, + 281, + 222, + 354 + ], + "type": "image", + "image_path": "02a083dd30a7cdc2269f2c2a4655f07b03e7b64862df8f3540abe08103c970e7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 222, + 258, + 303, + 328 + ], + "blocks": [ + { + "bbox": [ + 222, + 258, + 303, + 328 + ], + "lines": [ + { + "bbox": [ + 222, + 258, + 303, + 328 + ], + "spans": [ + { + "bbox": [ + 222, + 258, + 303, + 328 + ], + "type": "image", + "image_path": "23deca315088794a2003b1809fb7c064cce75576b6ffdc223bb8ba918eafa4ec.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 222, + 327, + 303, + 396 + ], + "blocks": [ + { + "bbox": [ + 222, + 327, + 303, + 396 + ], + "lines": [ + { + "bbox": [ + 222, + 327, + 303, + 396 + ], + "spans": [ + { + "bbox": [ + 222, + 327, + 303, + 396 + ], + "type": "image", + "image_path": "977f807793e24d4383695bb2f8c5801da5a67eae6ec4df571485bde96d8d6b35.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 235, + 398, + 290, + 422 + ], + "lines": [ + { + "bbox": [ + 235, + 398, + 290, + 422 + ], + "spans": [ + { + "bbox": [ + 235, + 398, + 290, + 422 + ], + "type": "text", + "content": "Mask3D [46] trained on manual labels." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 306, + 258, + 396, + 396 + ], + "blocks": [ + { + "bbox": [ + 306, + 258, + 396, + 396 + ], + "lines": [ + { + "bbox": [ + 306, + 258, + 396, + 396 + ], + "spans": [ + { + "bbox": [ + 306, + 258, + 396, + 396 + ], + "type": "image", + "image_path": "ca2a3ddc14d4f8421e7737521a083b1df213c7e8ac3a1d7135b381ab047c1ff6.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 397, + 391, + 422 + ], + "lines": [ + { + "bbox": [ + 315, + 397, + 391, + 422 + ], + "spans": [ + { + "bbox": [ + 315, + 397, + 391, + 422 + ], + "type": "text", + "content": "Segment3D (Ours) trained on automatic labels." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 395, + 283, + 480, + 354 + ], + "blocks": [ + { + "bbox": [ + 395, + 283, + 480, + 354 + ], + "lines": [ + { + "bbox": [ + 395, + 283, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 395, + 283, + 480, + 354 + ], + "type": "image", + "image_path": "53e1e0c7a9520f86a8a02c5e3eaabc759f92bbcb36f5119192bff79bb450aea8.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 395, + 350, + 481, + 433 + ], + "blocks": [ + { + "bbox": [ + 395, + 350, + 481, + 433 + ], + "lines": [ + { + "bbox": [ + 395, + 350, + 481, + 433 + ], + "spans": [ + { + "bbox": [ + 395, + 350, + 481, + 433 + ], + "type": "image", + "image_path": "aabda92570de3afea7e8a391da2c6c58414d59976b9338cb44aa8c9538da057f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 439, + 482, + 495 + ], + "lines": [ + { + "bbox": [ + 130, + 439, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 439, + 482, + 495 + ], + "type": "text", + "content": "Fig. 1: Fine-Grained Class-Agnostic 3D Point Cloud Segmentation. Segment3D predicts highly accurate segmentation masks (right), improves over state-of-the-art 3D segmentation methods (e.g., Mask3D [46], left), and does not require manually labeled 3D training data. This is achieved through the automatic generation of high-quality training masks using foundation models for image segmentation [28]." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 159, + 511, + 453, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 511, + 453, + 665 + ], + "spans": [ + { + "bbox": [ + 159, + 511, + 453, + 665 + ], + "type": "text", + "content": "Abstract. Current 3D scene segmentation methods are heavily dependent on manually annotated 3D training datasets. Such manual annotations are labor-intensive, and often lack fine-grained details. Furthermore, models trained on this data typically struggle to recognize object classes beyond the annotated training classes, i.e., they do not generalize well to unseen domains and require additional domain-specific annotations. In contrast, recent 2D foundation models have demonstrated strong generalization and impressive zero-shot abilities, inspiring us to incorporate these characteristics from 2D models into 3D models. Therefore, we explore the use of image segmentation foundation models to automatically generate high-quality training labels for 3D segmentation models. The resulting model, Segment3D, generalizes significantly better than the models trained on costly manual 3D labels and enables easily adding new training data to further boost the segmentation performance." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 148, + 230, + 160 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 148, + 230, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 148, + 230, + 160 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 172, + 481, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 172, + 481, + 232 + ], + "spans": [ + { + "bbox": [ + 130, + 172, + 481, + 232 + ], + "type": "text", + "content": "In this work, we propose Segment3D, a method for fine-grained class-agnostic 3D segmentation. In particular, dividing the space into coherent segments aligned with both the scene geometry and its semantics is a key challenge. This ability to accurately segment and interpret 3D scenes is fundamental for numerous downstream tasks [52,61], intelligent assistants and autonomous robots [30,64]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 233, + 481, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 233, + 481, + 448 + ], + "spans": [ + { + "bbox": [ + 130, + 233, + 481, + 448 + ], + "type": "text", + "content": "Current methods for 3D indoor-scene understanding mostly focus on semantic [35, 39, 40, 47, 53] and instance segmentation [9, 17, 31, 32, 46, 54]. These approaches, while effective on popular benchmarks [13, 15, 44, 50], have limitations stemming from their training. Primarily, they depend on extensive manually labeled 3D training sets that are both time-consuming and challenging to annotate. Importantly, their performance often deteriorates when applied to scenarios beyond their training data, limiting their effectiveness in diverse, real-world scenarios. This becomes particularly apparent under the recently emerging task of open-vocabulary 3D scene understanding [18, 27, 33, 38, 51] that aims to segment arbitrary user queries, which naturally go beyond the pre-defined set of training-set classes. Concurrently, the recent surge in foundation models, especially 2D vision-language models [25, 28, 41, 60], demonstrates remarkable potential. Trained on internet-scale data, these models exhibit an extraordinary ability to generalize, even in a zero-shot setting, to new and different input distributions. However, their application has been predominantly confined to 2D data. For instance, SAM [28] has shown impressive results in 2D image segmentation, but its applicability to 3D scene understanding remains largely under-explored. All these factors give rise to the research question:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 160, + 455, + 453, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 455, + 453, + 479 + ], + "spans": [ + { + "bbox": [ + 160, + 455, + 453, + 479 + ], + "type": "text", + "content": "How to leverage 2D foundation models for class-agnostic 3D scene segmentation without requiring manually labeled 3D data?" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 486, + 481, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 481, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 481, + 616 + ], + "type": "text", + "content": "Recently, SAM3D [57] has proposed a straightforward method that uses posed RGB-D images corresponding to a 3D scene. They predict segmentation masks for all input RGB-D images with SAM [28], and projected them into the 3D space. Next, through an iterative bottom-up process, the 3D masks of partial scenes are merged to derive the final 3D segmentation (see Fig. 2). However, variations in perspectives across frames can lead to conflicting segmentations, which introduces inconsistency during the merging process. Besides, the inference speed of SAM3D is slow, due to running SAM on the fly for a large number of RGB images and the cumbersome non-learned merging process. In this paper, we seek to circumvent both the slow speed and the issue of merging inconsistent segmentation across frames." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "content": "Towards this goal, we introduce Segment3D, a clean and efficient approach that utilizes a 3D model to achieve class-agnostic, fine-grained 3D segmentation at speed. Segment3D employs a two-stage training approach that requires no hand-annotated labels. We first train our class-agnostic 3D segmentation model" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Huang et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 160, + 117, + 453, + 128 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 117, + 453, + 128 + ], + "spans": [ + { + "bbox": [ + 160, + 117, + 453, + 128 + ], + "type": "text", + "content": "Keywords: Class-agnostic 3D segmentation " + }, + { + "bbox": [ + 160, + 117, + 453, + 128 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 117, + 453, + 128 + ], + "type": "text", + "content": " 3D Scene Understanding" + } + ] + } + ], + "index": 2 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 297 + ], + "type": "text", + "content": "with RGB-D data, which are transformed into partial RGB-D point clouds. The supervision is provided by the segmentation masks produced by SAM and projected into 3D. While SAM can automatically generate masks for entire images, sometimes the results may suffer from over-segmentation (see Fig. 4), which is not desired in our context. Therefore, we propose a Mask Generation Module to generate high-quality, complete masks of objects for a given RGB image. Since there is readily available large-scale RGB-D data, additional data can be effortlessly integrated to further enhance segmentation performance. This pretraining stage lays the groundwork for understanding the 3D structure from 2D annotations. However, as our ultimate objective is the segmentation of full 3D scenes, we must bridge the domain gap between partial point clouds and the more comprehensive 3D point clouds obtained from 3D scanners or reconstruction techniques [14,24]. To this end, in the second stage, we fine-tune the model on full 3D point clouds in a self-supervised manner, utilizing high-confidence mask predictions from the pre-trained model as training signal." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 297, + 482, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 482, + 369 + ], + "type": "text", + "content": "We demonstrate strong performance in class-agnostic segmentation on ScanNet [13, 44] and the newly released ScanNet++ [58]. Though Segment3D is trained on the indoor dataset, transferring it directly to the outdoor nuScenes dataset [4] also yields surprisingly good results. Finally, we show the use of Segment3D for improving open-vocabulary 3D instance segmentation [51]. Overall, the contributions of this paper are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 373, + 479, + 466 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 131, + 373, + 479, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 373, + 479, + 407 + ], + "spans": [ + { + "bbox": [ + 131, + 373, + 479, + 407 + ], + "type": "text", + "content": "- We introduce Segment3D, a novel approach and training strategy for fine-grained class-agnostic 3D point cloud segmentation without manually annotated labels." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 408, + 479, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 408, + 479, + 430 + ], + "spans": [ + { + "bbox": [ + 131, + 408, + 479, + 430 + ], + "type": "text", + "content": "- We propose a Mask Generation Module that utilizes the 2D foundation model to automatically generate high-quality, complete training masks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 433, + 479, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 433, + 479, + 466 + ], + "spans": [ + { + "bbox": [ + 131, + 433, + 479, + 466 + ], + "type": "text", + "content": "- We show that Segment3D demonstrates strong generalization compared to a wide range of baselines, including fully supervised methods trained on carefully annotated datasets." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 131, + 495, + 237, + 508 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 495, + 237, + 508 + ], + "spans": [ + { + "bbox": [ + 131, + 495, + 237, + 508 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "type": "text", + "content": "3D Instance Segmentation. Current models have seen a significant development over the last years, from proposal-based [56, 59], over grouping-based [9, 26, 31, 54], to recent Transformer-based [32, 46] methods. In this work, we follow the currently best-performing Transformer-based paradigm. Despite impressive advancements, a shared limitation is the dependence on costly manual ground-truth annotations. Recently, there have been efforts to automate the annotation process by bundling state-of-the-art segmentation models [55], but they are still limited to pre-defined object classes. However, in the context of open-world 3D scene understanding, the importance of semantic classification for closed-set categories has diminished. Instead, in this work, we propose a general 3D segmentation method trained on automatically generated labels." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 399, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 399, + 91, + 447, + 102 + ], + "type": "text", + "content": "Segment3D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 297 + ], + "type": "text", + "content": "Foundation Models. Foundation models, particularly those that are multimodal [25,41], have revolutionized the field of AI by leveraging extensive image-text pre-training. These models can derive rich image representations guided by natural language descriptions, enabling a variety of downstream tasks [21,37,42]. Another line of work [6,36], based on self-supervised learning, employs image training and yields high-performance features directly applicable as inputs for linear classifiers. Segment Anything Model (SAM) [28] has recently advanced the performance of foundation models for image segmentation. SAM has undergone training on a diverse, high-quality dataset comprising more than 1 billion masks. This training equips SAM with the ability to generalize to novel object types and images, surpassing the scope of its observations during the training process. Additionally, SAM can generate high-quality, fine-grained masks. Our work leverages the power of SAM for pre-training a 3D segmentation model, which is later also used for generating supervision signals for fine-tuning on scene-level point clouds." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 300, + 482, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 482, + 494 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 482, + 494 + ], + "type": "text", + "content": "Open-Vocabulary 3D Scene Understanding. Recently, there has been an increased interest in 3D open-vocabulary scene understanding. This new field utilizes the zero-shot recognition abilities of 2D vision-language models [25,41], enabling a more comprehensive understanding of diverse and previously unseen 3D environments [8,16,22,23,27,29,33,38,51,62,63]. PLA [16] aligns point cloud features with captions extracted from multi-view images of a scene to enable open-vocabulary recognition. OpenScene [38] distills per-pixel image features to 3D point clouds, generating point-wise scene representations co-embedded with text and image pixels in CLIP feature space. However, it mainly focuses on semantic segmentation and exhibits a limited understanding of object instances. To this end, OpenMask3D [51] predicts class-agnostic 3D instance masks and aggregates per-mask features via multi-view fusion of CLIP-based image embeddings. Nevertheless, the segmentation model it uses is trained on closed-set labels, lacking generalization to the open world. Our method contributes to this area by providing open-set class-agnostic masks, which can serve as foundational inputs for models such as OpenMask3D." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 509, + 202, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 509, + 202, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 509, + 202, + 522 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "text", + "content": "We aim to develop a method capable of segmenting any object within a given 3D scene. Relying on existing 3D training datasets cannot accomplish this goal, as their mask annotations are limited to a predefined set of object classes [1,2,13,48]. Consequently, models trained on such datasets fail to generalize effectively to new, unseen classes. Recently, with SAM [28] showing extraordinary generalization ability in 2D segmentation, SAM3D [57] has proposed to simply project and merge SAM predictions in 3D from posed RGB-D images, as illustrated in Fig. 2. However, it suffers from inconsistent results and slow inference speed, limiting its practical applicability." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": "To fully exploit SAM for consistent 3D segmentation at speed, we introduce a clean and efficient method which utilizes a 3D model to directly segment the" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 176, + 117, + 438, + 250 + ], + "blocks": [ + { + "bbox": [ + 176, + 117, + 438, + 250 + ], + "lines": [ + { + "bbox": [ + 176, + 117, + 438, + 250 + ], + "spans": [ + { + "bbox": [ + 176, + 117, + 438, + 250 + ], + "type": "image", + "image_path": "f86f7851e28f780ecb81a3e158d663ff50c97018776cce971b8d5f1e6a05a28e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 258, + 482, + 324 + ], + "lines": [ + { + "bbox": [ + 130, + 258, + 482, + 324 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 482, + 324 + ], + "type": "text", + "content": "Fig. 2: SAM3D vs. Segment3D. SAM3D [57] (left) merges 2D segmentation masks of RGB-D images generated by SAM to obtain 3D segmentation of entire scenes. However, conflicting segmentations across frames introduce inconsistency during the merging process. Moreover, it is slow because of extensive image inference and the cumbersome merging procedure. Instead, our Segment3D (right) utilizes a 3D model to directly segment entire 3D scenes which is clean and efficient." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 338, + 482, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 338, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 482, + 445 + ], + "type": "text", + "content": "entire scene. First, we propose a Mask Generation Module to generate high-quality, complete object masks from SAM predictions (Sec. 3.1). Next, we train our 3D segmentation model with RGB-D data and the generated object masks, which are transformed into partial point clouds (Sec. 3.2). Since the final goal is to segment 3D scenes, we need to bridge the inevitable domain gap between partial RGB-D point clouds and full 3D point clouds from 3D scanners or reconstruction methods. We therefore fine-tune our model on full 3D point clouds using high-confident mask predictions from our pre-trained model (Sec. 3.3). The overall framework is illustrated in Fig. 3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 462, + 369, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 462, + 369, + 475 + ], + "spans": [ + { + "bbox": [ + 131, + 462, + 369, + 475 + ], + "type": "text", + "content": "3.1 Generating Pseudo Ground-Truth Masks" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 486, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 582 + ], + "type": "text", + "content": "Review on SAM. We first review the automatic mask generation pipeline of SAM [28]. SAM is prompted with a regular grid of " + }, + { + "bbox": [ + 130, + 486, + 482, + 582 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 130, + 486, + 482, + 582 + ], + "type": "text", + "content": " points on the RGB image. Each point predicts masks at three different granularities: whole, part, and subpart. Next, all three masks for all points are filtered by the model's predicted IoU score as well as the stability of the masks with respect to the binarization threshold. Then, the remaining masks undergo Non-Maximum Suppression (NMS). Finally, if a pixel is encompassed by multiple masks, the mask ID with the highest predicted IoU is assigned to the pixel." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": "However, as shown in Fig. 4, the results obtained in this way could suffer from over-segmentation. A naive idea is to only use the masks generated by the \"whole\" branch. Nevertheless, we find that the \"whole\" branch typically generates high-quality masks for large objects while overlooking small objects in the image. In contrast, the \"part\" and \"subpart\" branches not only generate masks for parts of large objects but also predict masks for small and tiny objects. Thus, to enable our 3D model to achieve fine-grained segmentation capabilities for details and" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 399, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 399, + 91, + 447, + 102 + ], + "type": "text", + "content": "Segment3D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 127, + 478, + 227 + ], + "blocks": [ + { + "bbox": [ + 169, + 117, + 307, + 125 + ], + "lines": [ + { + "bbox": [ + 169, + 117, + 307, + 125 + ], + "spans": [ + { + "bbox": [ + 169, + 117, + 307, + 125 + ], + "type": "text", + "content": "Stage1: Pre-Training on Partial Point Clouds" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 134, + 127, + 478, + 227 + ], + "lines": [ + { + "bbox": [ + 134, + 127, + 478, + 227 + ], + "spans": [ + { + "bbox": [ + 134, + 127, + 478, + 227 + ], + "type": "image", + "image_path": "a5671d4a17d0178f55d40f687cb4dd4ff3a744096b4a4ba325d0b3d4c7e53496.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 117, + 473, + 125 + ], + "lines": [ + { + "bbox": [ + 345, + 117, + 473, + 125 + ], + "spans": [ + { + "bbox": [ + 345, + 117, + 473, + 125 + ], + "type": "text", + "content": "Stage2: Fine-Tuning on Full Point Clouds" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 236, + 482, + 314 + ], + "lines": [ + { + "bbox": [ + 131, + 236, + 482, + 314 + ], + "spans": [ + { + "bbox": [ + 131, + 236, + 482, + 314 + ], + "type": "text", + "content": "Fig. 3: Method Overview. Training Segment3D involves two stages: The first stage (left) relies on largely available RGB-D image sequences and SAM " + }, + { + "bbox": [ + 131, + 236, + 482, + 314 + ], + "type": "inline_equation", + "content": "\\square" + }, + { + "bbox": [ + 131, + 236, + 482, + 314 + ], + "type": "text", + "content": ", a pre-trained foundation model for 2D image segmentation [28]. Segment3D " + }, + { + "bbox": [ + 131, + 236, + 482, + 314 + ], + "type": "inline_equation", + "content": "\\square" + }, + { + "bbox": [ + 131, + 236, + 482, + 314 + ], + "type": "text", + "content": " is pre-trained on partial RGB-D point clouds and supervised with pseudo ground-truth masks generated by Mask Generation Module (MGM) " + }, + { + "bbox": [ + 131, + 236, + 482, + 314 + ], + "type": "inline_equation", + "content": "\\square" + }, + { + "bbox": [ + 131, + 236, + 482, + 314 + ], + "type": "text", + "content": ". Due to the domain gap between partial and full point clouds, in the second stage (right), Segment3D " + }, + { + "bbox": [ + 131, + 236, + 482, + 314 + ], + "type": "inline_equation", + "content": "\\square" + }, + { + "bbox": [ + 131, + 236, + 482, + 314 + ], + "type": "text", + "content": " is fine-tuned with confident masks predicted by the pre-trained Segment3D " + }, + { + "bbox": [ + 131, + 236, + 482, + 314 + ], + "type": "inline_equation", + "content": "\\square" + }, + { + "bbox": [ + 131, + 236, + 482, + 314 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 328, + 480, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 328, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 131, + 328, + 480, + 365 + ], + "type": "text", + "content": "small objects, we must retain the outputs from the \"part\" and \"subpart\" branches. The problem is how to preserve their segmentation for small and tiny objects while removing their segmentation for parts of large objects." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 369, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 369, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 131, + 369, + 482, + 529 + ], + "type": "text", + "content": "Mask Generation Module (MGM). Here we introduce our Mask Generation Module (MGM). Since SAM's automatic mask generation pipeline can provide reasonably high-quality masks for a given image, we start based on it. Next, we compare the result with the output of the \"whole\" branch to determine which masks are parts of large objects and remove them. Specifically, for each mask in the automatic output " + }, + { + "bbox": [ + 131, + 369, + 482, + 529 + ], + "type": "inline_equation", + "content": "m_i^{\\mathrm{auto}}" + }, + { + "bbox": [ + 131, + 369, + 482, + 529 + ], + "type": "text", + "content": ", we iterate through the masks of the entire branch, calculating the intersection " + }, + { + "bbox": [ + 131, + 369, + 482, + 529 + ], + "type": "inline_equation", + "content": "m_{ij}^{\\mathrm{inter}}" + }, + { + "bbox": [ + 131, + 369, + 482, + 529 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 131, + 369, + 482, + 529 + ], + "type": "inline_equation", + "content": "m_i^{\\mathrm{auto}}" + }, + { + "bbox": [ + 131, + 369, + 482, + 529 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 369, + 482, + 529 + ], + "type": "inline_equation", + "content": "m_j^{\\mathrm{whole}}" + }, + { + "bbox": [ + 131, + 369, + 482, + 529 + ], + "type": "text", + "content": ". If " + }, + { + "bbox": [ + 131, + 369, + 482, + 529 + ], + "type": "inline_equation", + "content": "m_{ij}^{\\mathrm{inter}} / m_i^{\\mathrm{auto}} \\geq t" + }, + { + "bbox": [ + 131, + 369, + 482, + 529 + ], + "type": "text", + "content": ", then we consider " + }, + { + "bbox": [ + 131, + 369, + 482, + 529 + ], + "type": "inline_equation", + "content": "m_i^{\\mathrm{auto}}" + }, + { + "bbox": [ + 131, + 369, + 482, + 529 + ], + "type": "text", + "content": " to be a part of " + }, + { + "bbox": [ + 131, + 369, + 482, + 529 + ], + "type": "inline_equation", + "content": "m_j^{\\mathrm{whole}}" + }, + { + "bbox": [ + 131, + 369, + 482, + 529 + ], + "type": "text", + "content": ". Our preliminary results show that the process of extracting valid masks is highly robust to the hyper-parameter " + }, + { + "bbox": [ + 131, + 369, + 482, + 529 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 369, + 482, + 529 + ], + "type": "text", + "content": " and we set it as 0.9 by default. We remove the part masks and combine the remaining masks with the output of the \"whole\" branch, which will serve as our training supervision. With MGM, we observe notable improvements at all object mask sizes, especially on medium-sized and large objects (see Table 6)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 547, + 395, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 547, + 395, + 559 + ], + "spans": [ + { + "bbox": [ + 131, + 547, + 395, + 559 + ], + "type": "text", + "content": "3.2 Stage 1: Pre-Training on RGB-D Point Clouds" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 564, + 482, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 564, + 482, + 624 + ], + "spans": [ + { + "bbox": [ + 130, + 564, + 482, + 624 + ], + "type": "text", + "content": "In contrast to the relatively scarce availability of 3D data, there is an abundance of 2D data, particularly RGB-D images, which are readily accessible. For example, ScanNet [13] comprises merely 1513 3D scans, compared to the substantially larger collection of 2.5 million RGB-D images. Therefore, we first pre-train our 3D segmentation model on partial RGB-D point clouds." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 629, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 629, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 629, + 482, + 665 + ], + "type": "text", + "content": "Training-Set Preparation. Starting from a collection of RGB-D frames, we create partial 3D point clouds and their corresponding pseudo ground-truth 3D masks. Note that the labels can be automatically obtained without manual effort." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 136, + 212, + 190 + ], + "blocks": [ + { + "bbox": [ + 149, + 118, + 200, + 134 + ], + "lines": [ + { + "bbox": [ + 149, + 118, + 200, + 134 + ], + "spans": [ + { + "bbox": [ + 149, + 118, + 200, + 134 + ], + "type": "text", + "content": "Automatic-SAM Generated Masks" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 137, + 136, + 212, + 190 + ], + "lines": [ + { + "bbox": [ + 137, + 136, + 212, + 190 + ], + "spans": [ + { + "bbox": [ + 137, + 136, + 212, + 190 + ], + "type": "image", + "image_path": "e931db91a5ab2cba7dfd94ce873d48d635125c3f64094c8e679b9324cdb26036.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 137, + 192, + 211, + 246 + ], + "blocks": [ + { + "bbox": [ + 137, + 192, + 211, + 246 + ], + "lines": [ + { + "bbox": [ + 137, + 192, + 211, + 246 + ], + "spans": [ + { + "bbox": [ + 137, + 192, + 211, + 246 + ], + "type": "image", + "image_path": "69a033e5891fced37f67f592bb79ceaeeb2344d035c1e81b3881aa49f045ca13.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 255, + 301, + 308 + ], + "lines": [ + { + "bbox": [ + 132, + 255, + 301, + 308 + ], + "spans": [ + { + "bbox": [ + 132, + 255, + 301, + 308 + ], + "type": "text", + "content": "Fig. 4: Automatic-SAM Generated (left) vs. MGM Generated (right) Masks from ScanNet [13]. The former exhibits over-segmentation, while the latter does not." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 221, + 136, + 296, + 190 + ], + "blocks": [ + { + "bbox": [ + 249, + 118, + 286, + 134 + ], + "lines": [ + { + "bbox": [ + 249, + 118, + 286, + 134 + ], + "spans": [ + { + "bbox": [ + 249, + 118, + 286, + 134 + ], + "type": "text", + "content": "MGM rated Masks" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 221, + 136, + 296, + 190 + ], + "lines": [ + { + "bbox": [ + 221, + 136, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 221, + 136, + 296, + 190 + ], + "type": "image", + "image_path": "d4e3faaf194ed016d08c2395d4390adc664698488d391d639d8560861534abcd.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 221, + 192, + 296, + 246 + ], + "blocks": [ + { + "bbox": [ + 221, + 192, + 296, + 246 + ], + "lines": [ + { + "bbox": [ + 221, + 192, + 296, + 246 + ], + "spans": [ + { + "bbox": [ + 221, + 192, + 296, + 246 + ], + "type": "image", + "image_path": "f143a5be6e9ef1ffc065014faf122eb7e7f09aa654e81953cea21d48573234ec.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 311, + 268, + 482, + 323 + ], + "lines": [ + { + "bbox": [ + 311, + 268, + 482, + 323 + ], + "spans": [ + { + "bbox": [ + 311, + 268, + 482, + 323 + ], + "type": "text", + "content": "Fig. 5: Partial (center) vs. Full (right) Point Clouds from ScanNet [13]. Full point clouds are more complete and exhibit fewer occlusions due to reconstruction over multiple viewpoints." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 313, + 118, + 480, + 259 + ], + "blocks": [ + { + "bbox": [ + 313, + 118, + 480, + 259 + ], + "lines": [ + { + "bbox": [ + 313, + 118, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 313, + 118, + 480, + 259 + ], + "type": "image", + "image_path": "4da37cbae5a0269038b2b95a3af6fe4c3a9130f6185a19847f5a6f98d73d0ec6.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 341, + 480, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 341, + 480, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 341, + 480, + 399 + ], + "type": "text", + "content": "For each frame in a large RGB-D dataset, we first transform the 2D depth map to a partial 3D point cloud. To do so, we need to know the intrinsic matrix " + }, + { + "bbox": [ + 130, + 341, + 480, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{K} \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 130, + 341, + 480, + 399 + ], + "type": "text", + "content": " and extrinsic matrix " + }, + { + "bbox": [ + 130, + 341, + 480, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{T} = [\\mathbf{R} \\mathbf{t}] \\in \\mathbb{R}^{3 \\times 4}" + }, + { + "bbox": [ + 130, + 341, + 480, + 399 + ], + "type": "text", + "content": ". For each pixel " + }, + { + "bbox": [ + 130, + 341, + 480, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{p} = (u, v)" + }, + { + "bbox": [ + 130, + 341, + 480, + 399 + ], + "type": "text", + "content": ", we can transform it with its depth value " + }, + { + "bbox": [ + 130, + 341, + 480, + 399 + ], + "type": "inline_equation", + "content": "D_{\\mathbf{p}}" + }, + { + "bbox": [ + 130, + 341, + 480, + 399 + ], + "type": "text", + "content": " into a 3D point " + }, + { + "bbox": [ + 130, + 341, + 480, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 130, + 341, + 480, + 399 + ], + "type": "text", + "content": " in world coordinates as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 233, + 404, + 480, + 420 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 404, + 480, + 420 + ], + "spans": [ + { + "bbox": [ + 233, + 404, + 480, + 420 + ], + "type": "interline_equation", + "content": "\\mathbf {P} = \\mathbf {R} ^ {\\top} \\cdot \\left(D _ {\\mathbf {p}} \\cdot \\mathbf {K} ^ {- 1} \\cdot \\tilde {\\mathbf {p}}\\right) - \\mathbf {R} ^ {\\top} \\mathbf {t}, \\tag {1}", + "image_path": "a8ddbd52cf0c425f1eff0e2e0e6538d96d25ded05f97f35b560a097bd099a587.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 429, + 482, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 429, + 482, + 502 + ], + "spans": [ + { + "bbox": [ + 130, + 429, + 482, + 502 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 429, + 482, + 502 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{p}}" + }, + { + "bbox": [ + 130, + 429, + 482, + 502 + ], + "type": "text", + "content": " is the homogeneous coordinate of " + }, + { + "bbox": [ + 130, + 429, + 482, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 130, + 429, + 482, + 502 + ], + "type": "text", + "content": ". By applying Eq. (1) to all pixels in the depth map " + }, + { + "bbox": [ + 130, + 429, + 482, + 502 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 429, + 482, + 502 + ], + "type": "text", + "content": " and associating their per-pixel RGB value, we obtain the input partial 3D point cloud. Next, we obtain the pseudo ground-truth 3D segmentation masks for this point cloud with the 2D masks generated by MGM in the same way. Since we know the one-to-one mapping between 2D pixel and 3D points from Eq. (1), we directly obtain the per-point 3D mask labels." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "Model Architecture. We use a model inspired by Mask3D [46] to train a class-agnostic 3D segmentation model. The model comprises a sparse convolutional backbone derived from MinkowskiUNet [12] and a transformer decoder, as in MaskFormer [10, 11]. We adopt a set of queries to represent the masks, each of which is initialized with a positional embedding. Specifically, we select query positions with furthest point sampling (FPS) and use their Fourier positional encodings as the query embeddings. Leveraging the transformer decoder, all the mask queries are refined by progressively attending to point cloud features across multiple scales in parallel. Each mask query is subsequently decoded into both a mask feature and a binary label to predict whether the given query corresponds to a valid object or not. By computing cosine similarity scores between a mask feature and all point features within the point cloud, a heatmap is generated over" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 399, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 399, + 91, + 447, + 101 + ], + "type": "text", + "content": "Segment3D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "at the point cloud. This heatmap is input to a sigmoid function, and thresholded at 0.5, resulting in the final binary mask." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 145, + 482, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 145, + 482, + 300 + ], + "spans": [ + { + "bbox": [ + 130, + 145, + 482, + 300 + ], + "type": "text", + "content": "Training with SAM Generated Masks. We supervise the model with two losses: the per-point supervision loss " + }, + { + "bbox": [ + 130, + 145, + 482, + 300 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{mask}}" + }, + { + "bbox": [ + 130, + 145, + 482, + 300 + ], + "type": "text", + "content": " and a per-query supervision loss " + }, + { + "bbox": [ + 130, + 145, + 482, + 300 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{obj}}" + }, + { + "bbox": [ + 130, + 145, + 482, + 300 + ], + "type": "text", + "content": ". The loss " + }, + { + "bbox": [ + 130, + 145, + 482, + 300 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{mask}}" + }, + { + "bbox": [ + 130, + 145, + 482, + 300 + ], + "type": "text", + "content": " enables learning a foreground-background segmentation for each mask, and is composed of a dice loss " + }, + { + "bbox": [ + 130, + 145, + 482, + 300 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{dice}}" + }, + { + "bbox": [ + 130, + 145, + 482, + 300 + ], + "type": "text", + "content": " [34] and a binary cross-entropy loss " + }, + { + "bbox": [ + 130, + 145, + 482, + 300 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ce}}" + }, + { + "bbox": [ + 130, + 145, + 482, + 300 + ], + "type": "text", + "content": " for each point. The " + }, + { + "bbox": [ + 130, + 145, + 482, + 300 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{obj}}" + }, + { + "bbox": [ + 130, + 145, + 482, + 300 + ], + "type": "text", + "content": " is a binary classification loss that indicates whether a query represents a valid \"object\" or \"no object\". This mechanism allows for the prediction of a variable number of masks, depending on the underlying scene content and geometry. Following prior work [5, 10, 46], we first adopt bipartite graph matching to establish correspondences between the set of predicted masks and the set of target masks provided by MGM as described before. If the predicted instance finds a matching target mask, then we assign it an \"object\" label; conversely, if there is no match, we assign \"no object\". In summary, we optimize the following losses:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 258, + 301, + 480, + 313 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 301, + 480, + 313 + ], + "spans": [ + { + "bbox": [ + 258, + 301, + 480, + 313 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {m a s k}} + \\lambda_ {\\text {o b j}} \\mathcal {L} _ {\\text {o b j}}, \\tag {2}", + "image_path": "f82501df3b8ae702ac89a302a75897a21da9ba486e48a592c0259680d5432352.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 246, + 316, + 480, + 328 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 316, + 480, + 328 + ], + "spans": [ + { + "bbox": [ + 246, + 316, + 480, + 328 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {m a s k}} = \\lambda_ {\\text {d i c e}} \\mathcal {L} _ {\\text {d i c e}} + \\lambda_ {\\text {c e}} \\mathcal {L} _ {\\text {c e}}, \\tag {3}", + "image_path": "634c37ac1bab557f637ceb0eaaade333974a5dff4da8ab8513b916461a831908.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 333, + 479, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 333, + 479, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 333, + 479, + 369 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 333, + 479, + 369 + ], + "type": "inline_equation", + "content": "\\lambda_{*}" + }, + { + "bbox": [ + 130, + 333, + 479, + 369 + ], + "type": "text", + "content": " are hyperparameters that balance the contribution of each component in the loss. The binary classification loss " + }, + { + "bbox": [ + 130, + 333, + 479, + 369 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{obj}}" + }, + { + "bbox": [ + 130, + 333, + 479, + 369 + ], + "type": "text", + "content": " is applied to all queries, while the mask loss " + }, + { + "bbox": [ + 130, + 333, + 479, + 369 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{mask}}" + }, + { + "bbox": [ + 130, + 333, + 479, + 369 + ], + "type": "text", + "content": " is specifically applied to masks labeled as \"object\"." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 388, + 379, + 400 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 388, + 379, + 400 + ], + "spans": [ + { + "bbox": [ + 131, + 388, + 379, + 400 + ], + "type": "text", + "content": "3.3 Stage 2: Self-Supervised Scene Fine-Tuning" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 409, + 482, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 409, + 482, + 564 + ], + "spans": [ + { + "bbox": [ + 130, + 409, + 482, + 564 + ], + "type": "text", + "content": "After Stage 1, we obtain a class-agnostic 3D segmentation model by pre-training solely on RGB-D images and automatically generated labels from MGM. However, a fundamental domain gap persists between partial point clouds derived from RGB-D data and full point clouds acquired through 3D scanners or reconstruction methods [14,24] (See Fig. 5). This gap exists mostly because of object occlusions, but also due to challenges of depth cameras to capture dark or reflective surfaces from a single viewpoint. Hence, depending solely on RGB-D frames for training a 3D segmentation model intended for full 3D scans proves inadequate. Therefore, we propose to further fine-tune our model on scene-level full 3D point clouds. The key idea to obtain 3D mask annotations for training on full point clouds is to use selected, high-confidence masks generated by the pre-trained model itself. Note that this stage requires no manual labels on full 3D scenes and proves essential for the performance of Segment3D (see Tab. 5)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Confidence-Based Mask Generation. Next, we outline the process of generating the supervision signal for the fine-tuning stage. The pre-trained model processes point clouds independently of their nature, be it partial or full. Therefore, when presented with a full 3D point cloud, the pre-trained 3D model produces a set of masks, each with a binary classification (valid or not) and a heatmap over all points, just as in Sec. 3.2. To assess the quality of the predicted masks, we compute a confidence score based on the confidence map " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\sigma(\\mathbf{h})" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{h}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " is the predicted heatmap and " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " is the sigmoid function. We then compute the average" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 134, + 115, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 115, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 134, + 115, + 482, + 200 + ], + "type": "text", + "content": "confidence of those points for which " + }, + { + "bbox": [ + 134, + 115, + 482, + 200 + ], + "type": "inline_equation", + "content": "\\sigma(\\mathbf{h}) > 0.5" + }, + { + "bbox": [ + 134, + 115, + 482, + 200 + ], + "type": "text", + "content": " as the confidence score of the predicted mask, denoted as " + }, + { + "bbox": [ + 134, + 115, + 482, + 200 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{mask}}" + }, + { + "bbox": [ + 134, + 115, + 482, + 200 + ], + "type": "text", + "content": ". We also consider the classification as a valid object and use the probability from the binary classification assigned to the \"object\" category as the confidence score, denoted as " + }, + { + "bbox": [ + 134, + 115, + 482, + 200 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{obj}}" + }, + { + "bbox": [ + 134, + 115, + 482, + 200 + ], + "type": "text", + "content": ". The final confidence score for each predicted mask is then the product of the two scores, " + }, + { + "bbox": [ + 134, + 115, + 482, + 200 + ], + "type": "inline_equation", + "content": "c = c_{\\mathrm{mask}} \\cdot c_{\\mathrm{obj}}" + }, + { + "bbox": [ + 134, + 115, + 482, + 200 + ], + "type": "text", + "content": ". For fine-tuning our 3D segmentation model in Stage 2, we select the most confidently predicted masks above a threshold " + }, + { + "bbox": [ + 134, + 115, + 482, + 200 + ], + "type": "inline_equation", + "content": "\\tau_c" + }, + { + "bbox": [ + 134, + 115, + 482, + 200 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 205, + 482, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 205, + 482, + 312 + ], + "spans": [ + { + "bbox": [ + 134, + 205, + 482, + 312 + ], + "type": "text", + "content": "Training with the High-Confidence Generated Masks. For fine-tuning, we follow the same procedure as before and use " + }, + { + "bbox": [ + 134, + 205, + 482, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{mask}}" + }, + { + "bbox": [ + 134, + 205, + 482, + 312 + ], + "type": "text", + "content": " as defined in Eq. 3. In contrast to the pre-training stage, the binary classification loss " + }, + { + "bbox": [ + 134, + 205, + 482, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{obj}}" + }, + { + "bbox": [ + 134, + 205, + 482, + 312 + ], + "type": "text", + "content": ", responsible for categorizing queries into valid or invalid, is omitted. Since we only select masks with high confidence for supervision, it can happen that some objects in the scene have no assigned ground truth mask. In such instances, deeming a correctly predicted mask for those objects as invalid would be detrimental. Table 5 illustrates the efficacy of the self-supervised fine-tuning process in comparison to pre-training alone." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 134, + 331, + 229, + 344 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 331, + 229, + 344 + ], + "spans": [ + { + "bbox": [ + 134, + 331, + 229, + 344 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 354, + 482, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 354, + 482, + 438 + ], + "spans": [ + { + "bbox": [ + 134, + 354, + 482, + 438 + ], + "type": "text", + "content": "We firstly evaluate Segment3D on indoor ScanNet++ [58], ScanNet200 [13,44] and outdoor nuScenes [4] in a class-agnostic segmentation setting (Sec. 4.1). We then show the advantage of our method in segmenting small objects, and provide analysis to understand the importance of fine-tuning, the effectiveness of Mask Generation Module and the potential of training on more data (Sec. 4.2). Finally, we demonstrate its application for the task of open-set 3D instance segmentation as proposed in OpenMask3D [51] (Sec. 4.3)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 134, + 456, + 375, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 456, + 375, + 468 + ], + "spans": [ + { + "bbox": [ + 134, + 456, + 375, + 468 + ], + "type": "text", + "content": "4.1 Comparing with State-of-the-Art Methods" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 134, + 480, + 482, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 480, + 482, + 588 + ], + "spans": [ + { + "bbox": [ + 134, + 480, + 482, + 588 + ], + "type": "text", + "content": "Datasets. We first evaluate on the recently released ScanNet++ [58] which includes high-resolution 3D scans captured at sub-millimeter precision, and fine-grained annotations covering objects of varying sizes. Due to its comprehensive data annotation, all of our analytical experiments in Sec. 4.2 are also based on it. Next, we adopt the ScanNet200 [13, 44], which builds upon the classic ScanNet [13] dataset and extends the semantic annotations to 200 classes. Then, we test our model on the nuScenes [4] outdoor dataset. For training our model, we employ ScanNet [13, 44], which is collected through a lightweight RGB-D scanning process. See the appendix for the implementation details of our method." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 134, + 593, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 593, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 593, + 482, + 665 + ], + "type": "text", + "content": "Methods in Comparison. We compare with a wide range of prior art methods from different categories. Mask3D [46] is a state-of-the-art fully-supervised method trained on manually annotated 3D segmentation masks. Segment3D has the same backbone as Mask3D but instead of training on manually annotated 3D masks, it learns from automatically generated masks. Felzenszwalb et al. [20] proposed a graph-based method for segmentation which operates directly on" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 399, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 399, + 91, + 447, + 102 + ], + "type": "text", + "content": "Segment3D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 142, + 163, + 466, + 245 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 158 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 158 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 158 + ], + "type": "text", + "content": "Table 1: Segmentation Scores on ScanNet++ [58]. The metric is average precision (AP) on the validation split. We report scores with and without post-processing (more details in Sec. 4.1). For reference, we show results on Mask3D trained on manual ScanNet200 labels." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 142, + 163, + 466, + 245 + ], + "lines": [ + { + "bbox": [ + 142, + 163, + 466, + 245 + ], + "spans": [ + { + "bbox": [ + 142, + 163, + 466, + 245 + ], + "type": "table", + "html": "
ModelAvg. Inference Time/swithout post-processingwith post-processing
AP\\( AP_{50} \\)\\( AP_{25} \\)AP\\( AP_{50} \\)\\( AP_{25} \\)
Mask3D [46]0.78.715.527.214.321.329.9
SAM3D [57]386.73.99.322.18.416.130.0
Felzenszwalb et al. [20]12.65.811.627.2---
Segment3D (Ours)0.713.023.838.320.230.942.7
", + "image_path": "ef572c3db0b06ab09f19eeac9ea547efee88bdce58a0b1824ac46389914c5a78.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 269, + 479, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 269, + 479, + 317 + ], + "spans": [ + { + "bbox": [ + 130, + 269, + 479, + 317 + ], + "type": "text", + "content": "the 3D geometry. UnScene3D [45] leverages self-supervised color and geometry features to obtain coarse segmentation and refines it through self-training. SAM3D [57] merges the segmentation masks of RGB-D images generated by SAM [28] to obtain the 3D segmentation result." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 323, + 480, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 323, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 480, + 407 + ], + "type": "text", + "content": "Metrics. We report average precision (AP) scores at IoU thresholds of " + }, + { + "bbox": [ + 130, + 323, + 480, + 407 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 130, + 323, + 480, + 407 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 323, + 480, + 407 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 323, + 480, + 407 + ], + "type": "text", + "content": ", and averaged over the range [0.5:0.95:05] between predicted and ground truth masks. Consistent with common practices in the field [47, 49, 54], we also report scores after post-processing. This involves smoothing the predicted masks through graph-based oversegmentation [20], and splitting distant parts of the same mask via connected component clustering DBSCAN [19]. We also report the average inference time for these methods on ScanNet++ tested on A100." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 413, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 482, + 581 + ], + "type": "text", + "content": "Results on ScanNet++. Scores are reported in Table 1. Segment3D outperforms all previous methods by at least +4.3 AP and up to +12.7 AP" + }, + { + "bbox": [ + 130, + 413, + 482, + 581 + ], + "type": "inline_equation", + "content": "_{25}" + }, + { + "bbox": [ + 130, + 413, + 482, + 581 + ], + "type": "text", + "content": ". Notably, we achieve such improvements without any ground-truth mask annotation contrary to Mask3D trained on ScanNet200. In general, the performance of Mask3D (and other fully-supervised methods) depends on the quality of the annotated training dataset; often the manual annotation of small objects (pens, cellphones) or other fine-details is challenging. Instead, Segment3D relies on automatically generated high-quality masks from MGM using the 2D foundation model, which can capture fine-grained details without human annotation effort. Table 4 highlights that Segment3D excels particularly in predicting the more challenging small object masks. Furthermore, our method significantly outperforms SAM3D, indicating that Segment3D can circumvent the noise that arises during the merging process of SAM3D. Additionally, As shown in Table 1, Segment3D is three-order-of-magnitude faster than SAM3D in inference speed." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": "Fig. 6 shows segmentation results of several representative examples on ScanNet++. As can be seen, the scenes are quite diverse, presenting multiple challenges such as clutter and a wide range of mask sizes. Despite these challenges, our model predicts quite accurate and well-localized segmentation masks. For example, compared to Mask3D, our method is able to segment the finer-grained objects on top of the bed and the shelf. The masks of the computer screen and the chair in front of it are also less fragmented than predictions of the baselines." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 146, + 159, + 282, + 234 + ], + "blocks": [ + { + "bbox": [ + 132, + 116, + 301, + 150 + ], + "lines": [ + { + "bbox": [ + 132, + 116, + 301, + 150 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 301, + 150 + ], + "type": "text", + "content": "Table 2: Segmentation Scores on ScanNet200 [13, 44]. The evaluation metric is average precision on the val split." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 146, + 159, + 282, + 234 + ], + "lines": [ + { + "bbox": [ + 146, + 159, + 282, + 234 + ], + "spans": [ + { + "bbox": [ + 146, + 159, + 282, + 234 + ], + "type": "table", + "html": "
ModelAPAP50
Mask3D [46]34.143.1
Felzenszwalb et al. [20]6.112.1
UnScene3D [45]15.932.2
SAM3D [57]19.032.5
Segment3D (Ours)27.739.8
", + "image_path": "70a7a84c1b43dda22c98994091f8e91e916973c390047cff985be18d92af8d34.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 321, + 157, + 467, + 236 + ], + "blocks": [ + { + "bbox": [ + 312, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 312, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 312, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 3: Segmentation Scores on nuScenes [4]. The evaluation metric is average precision on the val split." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 157, + 467, + 236 + ], + "lines": [ + { + "bbox": [ + 321, + 157, + 467, + 236 + ], + "spans": [ + { + "bbox": [ + 321, + 157, + 467, + 236 + ], + "type": "table", + "html": "
ModelGT LabelsAP50AP25
zero-shot test
Mask3D×15.325.5
Segment3D×37.848.0
fine-tuned on nuScenes
Mask3D✓46.557.8
Segment3D×42.555.5
", + "image_path": "34ab496a9b18383d234278effd7a94b97b0d6901e6c09d2dccb7be15d3e10149.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 257, + 479, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 257, + 479, + 330 + ], + "spans": [ + { + "bbox": [ + 130, + 257, + 479, + 330 + ], + "type": "text", + "content": "Results on ScanNet200. As illustrated in Table 2, Segment3D outperforms all baselines that require no manual 3D labels, and even shows competitive results with Mask3D, which is trained on ScanNet200. Compared to Unscene3D, which leverages self-supervised color and geometry features to obtain coarse segmentation, we achieve superior results by using the Mask Generation Module to generate high-quality complete masks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 334, + 482, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 334, + 482, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 334, + 482, + 453 + ], + "type": "text", + "content": "Results on Outdoor Data. The performance on nuScenes is demonstrated in Table 3. We observe that even though Segment3D is trained using purely indoor data, applying it directly to outdoor data produces surprisingly good results. This indicates the stronger generalizability of Segment3D over the fully-supervised method Mask3D when transferring from indoor to outdoor scenes. Furthermore, we report scores with fine-tuning on nuScenes which leads to overall improved results. Similar to Table 2, Mask3D performs best, however, Segment3D achieves very competitive performance without any hand-labeled training data. Refer to the appendix for additional results on STPLS3D [7] and Paris-Lille-3D [43] datasets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 471, + 270, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 471, + 270, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 471, + 270, + 483 + ], + "type": "text", + "content": "4.2 Analysis Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 495, + 482, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 482, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 482, + 556 + ], + "type": "text", + "content": "Performance on Different Mask Sizes. We proceed with an analysis of the performance of our Segment3D and Mask3D (the best-performing baseline) on object masks of various sizes. We categorize the size of masks based on the number of points they contain (small: [0, 2k], medium: [2k, 15k] large: [15k, " + }, + { + "bbox": [ + 130, + 495, + 482, + 556 + ], + "type": "inline_equation", + "content": "\\infty" + }, + { + "bbox": [ + 130, + 495, + 482, + 556 + ], + "type": "text", + "content": "]). The results are reported in Table 4. Our method yields significantly improved" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 171, + 601, + 445, + 663 + ], + "blocks": [ + { + "bbox": [ + 132, + 574, + 481, + 597 + ], + "lines": [ + { + "bbox": [ + 132, + 574, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 574, + 481, + 597 + ], + "type": "text", + "content": "Table 4: Segmentation Scores on Different Mask Sizes. Segmented3D improves over Mask3D, especially on small object masks. Details are in Sec. 4.2." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 171, + 601, + 445, + 663 + ], + "lines": [ + { + "bbox": [ + 171, + 601, + 445, + 663 + ], + "spans": [ + { + "bbox": [ + 171, + 601, + 445, + 663 + ], + "type": "table", + "html": "
Mask SizeLargeMediumSmall
ScanNet++APAP50APAP50APAP50
Mask3D13.825.113.623.31.84.3
Segment3D18.331.916.629.88.417.3
(+4.5)(+6.8)(+3.0)(+6.5)(+6.6)(+13.0)
", + "image_path": "67b15ec73e95f4e75d9603feeee909df272bf19bb8dfb6142c5b9942b3ff3d7a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 399, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 399, + 91, + 447, + 102 + ], + "type": "text", + "content": "Segment3D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 142, + 113, + 476, + 430 + ], + "blocks": [ + { + "bbox": [ + 142, + 113, + 476, + 430 + ], + "lines": [ + { + "bbox": [ + 142, + 113, + 476, + 430 + ], + "spans": [ + { + "bbox": [ + 142, + 113, + 476, + 430 + ], + "type": "image", + "image_path": "4a6319bc0d817b372e927c3ebd7431955843bfdfb283c6eb2c5ebf05c64311c1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 436, + 482, + 471 + ], + "lines": [ + { + "bbox": [ + 130, + 436, + 482, + 471 + ], + "spans": [ + { + "bbox": [ + 130, + 436, + 482, + 471 + ], + "type": "text", + "content": "Fig. 6: Qualitative Results on ScanNet++. From top to bottom, we show the colored input 3D scenes, the segmentation masks predicted by SAM3D [57], Mask3D [46], our Segment3D and the ground truth 3D mask annotations." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 501, + 482, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 501, + 482, + 586 + ], + "spans": [ + { + "bbox": [ + 130, + 501, + 482, + 586 + ], + "type": "text", + "content": "segmentation results on objects of various sizes, especially on small objects. This confirms our intuition that Mask3D performs poorly on small-sized object masks as those are typically harder to manually annotate. In contrast, Segment3D utilizes masks generated by MGM with the foundation model as supervision, capturing fine-grained scene details. This showcases the usefulness of foundation models and raises the question if manually labeled large-scale 3D datasets are necessary for training 3D point cloud segmentation models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "The Importance of Two-Stage Training. Next, we compare the performance of Segment3D pre-trained solely on partial RGB-D point clouds (Stage 1) and with additional fine-tuning on full point clouds (Stage 2). Scores are reported in Table 5. The additional fine-tuning stage almost doubles the segmentation performance of our model on the most challenging AP metric. By training with the predicted high-confidence masks, Segment3D effectively reduces the inherent" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 186, + 152, + 425, + 195 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 5: The Importance of Two-Stage Training. Fine-tuning on full point clouds supervised by confident predicted 3D masks significantly surpasses pre-training on projected 2D SAM masks alone." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 186, + 152, + 425, + 195 + ], + "lines": [ + { + "bbox": [ + 186, + 152, + 425, + 195 + ], + "spans": [ + { + "bbox": [ + 186, + 152, + 425, + 195 + ], + "type": "table", + "html": "
Training StagesAPAP50AP25
Pre-Training (Stage 1)6.814.230.6
+ Fine-Tuning (Stage 2)13.0 (+6.2)23.8 (+9.6)38.3 (+7.7)
", + "image_path": "c5d921e880bc0eb545d040fc90cf581b581d268c7b7eca14dbf1d22db4ac190e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 168, + 243, + 440, + 312 + ], + "blocks": [ + { + "bbox": [ + 130, + 205, + 480, + 239 + ], + "lines": [ + { + "bbox": [ + 130, + 205, + 480, + 239 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 480, + 239 + ], + "type": "text", + "content": "Table 6: Effectiveness of Mask Generation Module. Compared to the model trained with masks generated by Automatic-SAM, our model shows improvement across all object mask sizes, particularly for medium-sized and large objects." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 168, + 243, + 440, + 312 + ], + "lines": [ + { + "bbox": [ + 168, + 243, + 440, + 312 + ], + "spans": [ + { + "bbox": [ + 168, + 243, + 440, + 312 + ], + "type": "table", + "html": "
Mask SizeLargeMediumSmall
Masks Generated byAPAP50APAP50APAP50
Automatic-SAM16.830.614.227.27.516.2
MGM18.331.916.629.88.417.3
(+1.5)(+1.3)(+2.4)(+2.6)(+0.9)(+1.1)
", + "image_path": "fa69ff3d223ec6460dff8da5433237344844bef36b54fff7b5cc1bfe320accb2.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 333, + 480, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 333, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 333, + 480, + 357 + ], + "type": "text", + "content": "domain gap between the partial point clouds derived from RGB-D images and full 3D point clouds." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 361, + 482, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 482, + 445 + ], + "type": "text", + "content": "The Effectiveness of Mask Generation Module. To verify whether the MGM enhances the model's performance by reducing the over-segmentation issue and providing more complete masks, we conducted evaluations on objects of different sizes. As shown in Table 6, we can see that compared to the model trained with masks generated by Automatic-SAM, our model notably improves the performance of large and medium-sized objects which demonstrates the effectiveness of MGM." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 451, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 451, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 451, + 482, + 582 + ], + "type": "text", + "content": "Pre-Training with Additional Data. Since RGB-D data is available in abundance and masks can be automatically generated, it is natural to ask if pretraining on additional masks will further improve the overall performance. To that end, we perform a first experiment where we select additional frames from the training set of the ScanNet++ dataset (to increase the variety of training data). Table 7 shows an impressive performance boost of +4.1 AP. For future work, considering that this improvement is obtained simply by adding more automatically generated masks to the pre-training, our approach seems promising to train on even more data. It could even be plausible to train on internet images combined with monocular depth estimation, such as ZoeDepth [3], to compute the partial point clouds." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 598, + 384, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 598, + 384, + 611 + ], + "spans": [ + { + "bbox": [ + 131, + 598, + 384, + 611 + ], + "type": "text", + "content": "4.3 Application: Open-Set Scene Understanding" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "A real-world application of our class-agnostic 3D segmentation method is open-vocabulary 3D scene understanding, as implemented recently by OpenMask3D [51]. Given a 3D scene, a user can search for arbitrary objects via text prompts (see Fig. 7). A core component of OpenMask3D is Mask3D, which segments the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 399, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 399, + 91, + 447, + 102 + ], + "type": "text", + "content": "Segment3D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 158, + 296, + 196 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 301, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 301, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 301, + 148 + ], + "type": "text", + "content": "Table 7: Results with Additional Training Data. The performance is further boosted with more training data." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 158, + 296, + 196 + ], + "lines": [ + { + "bbox": [ + 136, + 158, + 296, + 196 + ], + "spans": [ + { + "bbox": [ + 136, + 158, + 296, + 196 + ], + "type": "table", + "html": "
Pre-Training DataAPAP50
ScanNet13.023.8
ScanNet, ScanNet++17.1 (+4.1)26.7 (+2.9)
", + "image_path": "b9fedc4e335787867d2b18f26983f43048f69596ac8e5bcccf80411ce605e54f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 311, + 158, + 473, + 196 + ], + "blocks": [ + { + "bbox": [ + 306, + 114, + 481, + 148 + ], + "lines": [ + { + "bbox": [ + 306, + 114, + 481, + 148 + ], + "spans": [ + { + "bbox": [ + 306, + 114, + 481, + 148 + ], + "type": "text", + "content": "Table 8: Open-Set 3D Scene Understanding. The evaluation metric is average precision with an IoU threshold of " + }, + { + "bbox": [ + 306, + 114, + 481, + 148 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 306, + 114, + 481, + 148 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 311, + 158, + 473, + 196 + ], + "lines": [ + { + "bbox": [ + 311, + 158, + 473, + 196 + ], + "spans": [ + { + "bbox": [ + 311, + 158, + 473, + 196 + ], + "type": "table", + "html": "
SegmentorScanNet++Replica
Mask3D [46]15.018.0
Segment3D (Ours)18.5 (+3.5)18.4 (+0.4)
", + "image_path": "b5f76b5009d9e688053dbd0296afccc2e3702e5ccbb98c5f7622e1c7ba587879.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 150, + 202, + 253, + 264 + ], + "blocks": [ + { + "bbox": [ + 150, + 202, + 253, + 264 + ], + "lines": [ + { + "bbox": [ + 150, + 202, + 253, + 264 + ], + "spans": [ + { + "bbox": [ + 150, + 202, + 253, + 264 + ], + "type": "image", + "image_path": "4ae3e9804388c5cc2a75354f70b7afaaae5bf91bfc44cffa88e504391e2b8334.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 178, + 266, + 227, + 272 + ], + "lines": [ + { + "bbox": [ + 178, + 266, + 227, + 272 + ], + "spans": [ + { + "bbox": [ + 178, + 266, + 227, + 272 + ], + "type": "text", + "content": "\"aback eraser\"" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 255, + 203, + 359, + 264 + ], + "blocks": [ + { + "bbox": [ + 255, + 203, + 359, + 264 + ], + "lines": [ + { + "bbox": [ + 255, + 203, + 359, + 264 + ], + "spans": [ + { + "bbox": [ + 255, + 203, + 359, + 264 + ], + "type": "image", + "image_path": "9e756c4fe06ba932d5c39e49f1b57229a930481421b91c7320fc831f7cbddd62.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 274, + 266, + 340, + 273 + ], + "lines": [ + { + "bbox": [ + 274, + 266, + 340, + 273 + ], + "spans": [ + { + "bbox": [ + 274, + 266, + 340, + 273 + ], + "type": "text", + "content": "\"paper on the laptop\"" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 278, + 482, + 323 + ], + "lines": [ + { + "bbox": [ + 131, + 278, + 482, + 323 + ], + "spans": [ + { + "bbox": [ + 131, + 278, + 482, + 323 + ], + "type": "text", + "content": "Fig. 7: Open-Set 3D Object Retrieval Results. Given a text prompt (bottom), OpenMask3D [51] finds the corresponding object masks in a given 3D scene (top). We adapt OpenMask3D and use fine-grained masks from our Segment3D method. We show 3D reconstructions and RGB images for better visualization (top left corner)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 361, + 203, + 463, + 264 + ], + "blocks": [ + { + "bbox": [ + 361, + 203, + 463, + 264 + ], + "lines": [ + { + "bbox": [ + 361, + 203, + 463, + 264 + ], + "spans": [ + { + "bbox": [ + 361, + 203, + 463, + 264 + ], + "type": "image", + "image_path": "f35f83da57a01168c4ea50e9a8fbbf7999838c0387258d7e92ac5a0bb3f83d8c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 396, + 266, + 427, + 272 + ], + "lines": [ + { + "bbox": [ + 396, + 266, + 427, + 272 + ], + "spans": [ + { + "bbox": [ + 396, + 266, + 427, + 272 + ], + "type": "text", + "content": "\"scissors\"" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 346, + 482, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 482, + 431 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 482, + 431 + ], + "type": "text", + "content": "scene into a set of object masks. Since Mask3D is trained on the closed-set of labeled annotations from ScanNet, its masks are not truly class-agnostic or open-vocabulary. We therefore replace Mask3D with our class-agnostic Segment3D. We evaluate on the closed-set labels of " + }, + { + "bbox": [ + 130, + 346, + 482, + 431 + ], + "type": "inline_equation", + "content": "\\mathrm{ScanNet}++" + }, + { + "bbox": [ + 130, + 346, + 482, + 431 + ], + "type": "text", + "content": " and Replica in Table 8 and report an improvement of up to " + }, + { + "bbox": [ + 130, + 346, + 482, + 431 + ], + "type": "inline_equation", + "content": "+3.5\\mathrm{AP}_{50}" + }, + { + "bbox": [ + 130, + 346, + 482, + 431 + ], + "type": "text", + "content": ". Since Segment3D can accurately generate fine-grained masks in a class-agnostic manner, we are able to retrieve objects such as \"eraser\", \"paper\" and \"scissors\", which are small in scale." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 449, + 220, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 449, + 220, + 462 + ], + "spans": [ + { + "bbox": [ + 132, + 449, + 220, + 462 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 464, + 482, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 464, + 482, + 572 + ], + "spans": [ + { + "bbox": [ + 130, + 464, + 482, + 572 + ], + "type": "text", + "content": "We have presented Segment3D, a simple, yet powerful class-agnostic 3D segmentation model. Segment3D employs a two-stage training framework that requires no manually annotated labels. We propose a Mask Generation Module to automatically generate high-quality, complete training masks with the 2D foundation model. The model is first pre-trained on RGB-D data with predicted 2D masks, then fine-tuned on full point clouds. Segment3D shows strong generalization and even outperforms existing 3D segmentation models that rely on hand-labeled 3D training scenes. Indeed, this raises the question of whether hand-labeled 3D training datasets are as essential as they were once thought to be." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 576, + 482, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 576, + 482, + 649 + ], + "spans": [ + { + "bbox": [ + 130, + 576, + 482, + 649 + ], + "type": "text", + "content": "Acknowledgement. Gao Huang is supported in part by the National Natural Science Foundation of China under grants (62321005 and 42327901). Francis Engelmann is partially supported by an ETH AI Center postdoctoral research fellowship and an ETH Zurich Career Seed Award. Songyou Peng is supported by an Innosuisse funding (100.567 IP-ICT), and Ayça Takmaz is supported by an Innosuisse grant (48727.1 IP-ICT)." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "type": "text", + "content": "1. Armeni, I., Sener, O., Zamir, A.R., Jiang, H., Brilakis, I., Fischer, M., Savarese, S.: 3D Semantic Parsing of Large-Scale Indoor Spaces. In: CVPR (2016) 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 162, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 162, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 138, + 162, + 480, + 205 + ], + "type": "text", + "content": "2. Baruch, G., Chen, Z., Dehghan, A., Dimry, T., Feigin, Y., Fu, P., Gebauer, T., Joffe, B., Kurz, D., Schwartz, A., et al.: ARKitScenes: A Diverse Real-World Dataset for 3D Indoor Scene Understanding using Mobile RGB-D Data. arXiv preprint arXiv:2111.08897 (2021) 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 206, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 206, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 138, + 206, + 480, + 237 + ], + "type": "text", + "content": "3. Bhat, S.F., Birkl, R., Wofk, D., Wonka, P., Müller, M.: ZoeDepth: Zero-Shot Transfer by Combining Relative and Metric Depth. arXiv preprint arXiv:2302.12288 (2023) 13" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 239, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 239, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 138, + 239, + 480, + 270 + ], + "type": "text", + "content": "4. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuScenes: A Multimodal Dataset for Autonomous Driving. In: CVPR (2020) 3, 9, 11" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 272, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 272, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 138, + 272, + 480, + 293 + ], + "type": "text", + "content": "5. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-End Object Detection with Transformers. In: ECCV (2020) 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 294, + 480, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 294, + 480, + 324 + ], + "spans": [ + { + "bbox": [ + 138, + 294, + 480, + 324 + ], + "type": "text", + "content": "6. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging Properties in Self-Supervised Vision Transformers. In: ICCV (2021) 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 326, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 326, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 138, + 326, + 480, + 357 + ], + "type": "text", + "content": "7. Chen, M., Hu, Q., Yu, Z., Thomas, H., Feng, A., Hou, Y., McCullough, K., Ren, F., Soibelman, L.: STPLS3D: A Large-Scale Synthetic and Real Aerial Photogrammetry 3D Point Cloud Dataset. arXiv preprint arXiv:2203.09065 (2022) 11" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 359, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 359, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 138, + 359, + 480, + 391 + ], + "type": "text", + "content": "8. Chen, R., Liu, Y., Kong, L., Zhu, X., Ma, Y., Li, Y., Hou, Y., Qiao, Y., Wang, W.: CLIP2Scene: Towards Label-Efficient 3D Scene Understanding by CLIP. In: CVPR (2023) 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 392, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 392, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 138, + 392, + 480, + 413 + ], + "type": "text", + "content": "9. Chen, S., Fang, J., Zhang, Q., Liu, W., Wang, X.: Hierarchical Aggregation for 3D Instance Segmentation. In: ICCV (2021) 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 415, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 415, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 138, + 415, + 480, + 435 + ], + "type": "text", + "content": "0. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-Attention Mask Transformer for Universal Image Segmentation. In: CVPR (2022) 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 436, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 436, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 138, + 436, + 480, + 456 + ], + "type": "text", + "content": "1. Cheng, B., Schwing, A., Kirillov, A.: Per-Pixel Classification is Not All You Need for Semantic Segmentation. In: NeurIPS (2021) 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 458, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 458, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 138, + 458, + 480, + 479 + ], + "type": "text", + "content": "2. Choy, C., Gwak, J., Savarese, S.: 4D Spatio-Temporal ConvNets: Minkowski Convolutional Neural Networks. In: CVPR (2019) 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 480, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 480, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 138, + 480, + 480, + 511 + ], + "type": "text", + "content": "3. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: ScanNet: Richly-Annotated 3D Reconstructions of Indoor Scenes. In: CVPR (2017) 2, 3, 4, 6, 7, 9, 11" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 513, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 513, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 513, + 480, + 544 + ], + "type": "text", + "content": "4. Dai, A., Nießner, M., Zollhöfer, M., Izadi, S., Theobalt, C.: BundleFusion: Real-time Globally Consistent 3D Reconstruction using On-the-fly Surface Re-integration. TOG (2017) 3, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 545, + 480, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 545, + 480, + 577 + ], + "spans": [ + { + "bbox": [ + 138, + 545, + 480, + 577 + ], + "type": "text", + "content": "5. Delitzas, A., Takmaz, A., Tombari, F., Sumner, R., Pollefeys, M., Engelmann, F.: SceneFun3D: Fine-Grained Functionality and Affordance Understanding in 3D Scenes. In: CVPR (2024) 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 578, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 480, + 599 + ], + "type": "text", + "content": "6. Ding, R., Yang, J., Xue, C., Zhang, W., Bai, S., Qi, X.: PLA: Language-Driven Open-Vocabulary 3D Scene Understanding. In: CVPR (2023) 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 600, + 480, + 630 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 600, + 480, + 630 + ], + "spans": [ + { + "bbox": [ + 138, + 600, + 480, + 630 + ], + "type": "text", + "content": "7. Engelmann, F., Bokeloh, M., Fathi, A., Leibe, B., Nießner, M.: 3D-MPA: Multi-Proposal Aggregation for 3D Semantic Instance Segmentation. In: CVPR (2020) 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "text", + "content": "8. Engelmann, F., Manhardt, F., Niemeyer, M., Tateno, K., Tombari, F.: OpenNeRF: Open Set 3D Neural Scene Segmentation with Pixel-Wise Features and Rendered Novel Views. In: ICLR (2024) 2" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 399, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 399, + 91, + 447, + 102 + ], + "type": "text", + "content": "Segment3D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "19. Ester, M., Kriegel, H.P., Sander, J., Xu, X., et al.: A Density-Based Algorithm for Discovering Clusters in Large Spatial Databases with Noise. In: KDD (1996) 10" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "type": "text", + "content": "20. Felzenszwalb, P.F., Huttenlocher, D.P.: Efficient Graph-Based Image Segmentation. IJCV (2004) 9, 10, 11" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 161, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 480, + 182 + ], + "type": "text", + "content": "21. Gu, X., Lin, T.Y., Kuo, W., Cui, Y.: Open-Vocabulary Object Detection via Vision and Language Knowledge Distillation. In: ICLR (2022) 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "type": "text", + "content": "22. Ha, H., Song, S.: Semantic Abstraction: Open-World 3D Scene Understanding from 2D Vision-Language Models. In: CoRL (2022) 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "text", + "content": "23. Huang, T., Dong, B., Yang, Y., Huang, X., Lau, R.W., Ouyang, W., Zuo, W.: CLIP2Point: Transfer CLIP to Point Cloud Classification with Image-Depth Pre-Training. In: ICCV (2023) 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 238, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 480, + 281 + ], + "type": "text", + "content": "24. Izadi, S., Kim, D., Hilliges, O., Molyneaux, D., Newcombe, R., Kohli, P., Shotton, J., Hodges, S., Freeman, D., Davison, A., et al.: KinectFusion: Real-time 3D Reconstruction and Interaction using a Moving Depth Camera. In: UIST (2011) 3, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 281, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 480, + 314 + ], + "type": "text", + "content": "25. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up Visual and Vision-Language Representation Learning with Noisy Text Supervision. In: ICML (2021) 2, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "type": "text", + "content": "26. Jiang, L., Zhao, H., Shi, S., Liu, S., Fu, C.W., Jia, J.: PointGroup: Dual-Set Point Grouping for 3D Instance Segmentation. In: CVPR (2020) 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "type": "text", + "content": "27. Kerr, J., Kim, C.M., Goldberg, K., Kanazawa, A., Tancik, M.: LERF: Language Embedded Radiance Fields. In: ICCV (2023) 2, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 358, + 480, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 390 + ], + "type": "text", + "content": "28. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment Anything. In: ICCV (2023) 1, 2, 4, 5, 6, 10" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 391, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 391, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 391, + 480, + 413 + ], + "type": "text", + "content": "29. Kobayashi, S., Matsumoto, E., Sitzmann, V.: Decomposing Nerf for Editing via Feature Field Distillation. In: NeurIPS (2022) 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 414, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 480, + 456 + ], + "type": "text", + "content": "30. Lemke, O., Bauer, Z., Zurbrügg, R., Pollefeys, M., Engelmann, F., Blum, H.: Spotcompose: A framework for open-vocabulary object retrieval and drawer manipulation in point clouds. In: 2nd Workshop on Mobile Manipulation and Embodied Intelligence at ICRA 2024 (2024) 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "type": "text", + "content": "31. Liang, Z., Li, Z., Xu, S., Tan, M., Jia, K.: Instance Segmentation in 3D Scenes using Semantic Superpoint Tree Networks. In: CVPR (2021) 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "type": "text", + "content": "32. Lu, J., Deng, J., Wang, C., He, J., Zhang, T.: Query Refinement Transformer for 3D Instance Segmentation. In: ICCV (2023) 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 502, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 480, + 533 + ], + "type": "text", + "content": "33. Lu, Y., Xu, C., Wei, X., Xie, X., Tomizuka, M., Keutzer, K., Zhang, S.: Open-Vocabulary Point-Cloud Object Detection without 3D Annotation. In: CVPR (2023) 2, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "type": "text", + "content": "34. Milletari, F., Navab, N., Ahmadi, S.A.: V-Net: Fully Convolutional Neural Networks for Volumetric Medical Image Segmentation. In: 3DV (2016) 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 556, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 556, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 556, + 480, + 578 + ], + "type": "text", + "content": "35. Nekrasov, A., Schult, J., Litany, O., Leibe, B., Engelmann, F.: Mix3D: Out-of-context data augmentation for 3D scenes. In: 3DV (2021) 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 578, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 480, + 609 + ], + "type": "text", + "content": "36. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: DINOv2: Learning Robust Visual Features without Supervision. arXiv preprint arXiv:2304.07193 (2023) 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "text", + "content": "37. Patashnik, O., Wu, Z., Shechtman, E., Cohen-Or, D., Lischinski, D.: StyleCLIP: Text-Driven Manipulation of StyleGAN Imagery. In: ICCV (2021) 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "text", + "content": "38. Peng, S., Genova, K., Jiang, C., Tagliasacchi, A., Pollefeys, M., Funkhouser, T., et al.: OpenScene: 3D Scene Understanding with Open Vocabularies. In: CVPR (2023) 2, 4" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "text", + "content": "39. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation. In: CVPR (2017) 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 159 + ], + "type": "text", + "content": "40. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In: NeurIPS (2017) 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 160, + 482, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 160, + 482, + 192 + ], + "spans": [ + { + "bbox": [ + 130, + 160, + 482, + 192 + ], + "type": "text", + "content": "41. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning Transferable Visual Models from Natural Language Supervision. In: ICML (2021) 2, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 192, + 482, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 192, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 192, + 482, + 224 + ], + "type": "text", + "content": "42. Rao, Y., Zhao, W., Chen, G., Tang, Y., Zhu, Z., Huang, G., Zhou, J., Lu, J.: DenseCLIP: Language-Guided Dense Prediction with Context-Aware Prompting. In: CVPR (2022) 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 225, + 482, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 225, + 482, + 256 + ], + "spans": [ + { + "bbox": [ + 130, + 225, + 482, + 256 + ], + "type": "text", + "content": "43. Roynard, X., Deschaud, J.E., Goulette, F.: Paris-Lille-3D: A Large and High-Quality Ground-Truth Urban Point Cloud Dataset for Automatic Segmentation and Classification. *IJRR* (2018) 11" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 257, + 482, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 257, + 482, + 278 + ], + "spans": [ + { + "bbox": [ + 130, + 257, + 482, + 278 + ], + "type": "text", + "content": "44. Rozenberszki, D., Litany, O., Dai, A.: Language-Grounded Indoor 3D Semantic Segmentation in the Wild. In: ECCV (2022) 2, 3, 9, 11" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 279, + 482, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 279, + 482, + 300 + ], + "spans": [ + { + "bbox": [ + 130, + 279, + 482, + 300 + ], + "type": "text", + "content": "45. Rozenberszki, D., Litany, O., Dai, A.: UnScene3D: Unsupervised 3D Instance Segmentation for Indoor Scenes. In: CVPR (2024) 10, 11" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 300, + 482, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 482, + 331 + ], + "type": "text", + "content": "46. Schult, J., Engelmann, F., Hermans, A., Litany, O., Tang, S., Leibe, B.: Mask3D: Mask Transformer for 3D Semantic Instance Segmentation. In: ICRA (2023) 1, 2, 3, 7, 8, 9, 10, 11, 12, 14" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 332, + 482, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 332, + 482, + 353 + ], + "spans": [ + { + "bbox": [ + 130, + 332, + 482, + 353 + ], + "type": "text", + "content": "47. Schult, J., Engelmann, F., Kontogianni, T., Leibe, B.: DualConvMesh-Net: Joint Geodesic and Euclidean Convolutions on 3D Meshes. In: CVPR (2020) 2, 10" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 354, + 482, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 386 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 386 + ], + "type": "text", + "content": "48. Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur-Artal, R., Ren, C., Verma, S., et al.: The Replica Dataset: A Digital Replica of Indoor Spaces. arXiv preprint arXiv:1906.05797 (2019) 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 386, + 482, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 386, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 386, + 482, + 407 + ], + "type": "text", + "content": "49. Sun, J., Qing, C., Tan, J., Xu, X.: Superpoint Transformer for 3D Scene Instance Segmentation. In: AAAI (2023) 10" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 407, + 482, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 482, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 482, + 450 + ], + "type": "text", + "content": "50. Sun, T., Hao, Y., Huang, S., Savarese, S., Schindler, K., Pollefeys, M., Armeni, I.: Nothing stands still: A spatiotemporal benchmark on 3d point cloud registration under large geometric and temporal change. arXiv preprint arXiv:2311.09346 (2023) 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 450, + 482, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 482 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 482 + ], + "type": "text", + "content": "51. Takmaz, A., Fedele, E., Sumner, R.W., Pollefeys, M., Tombari, F., Engelmann, F.: OpenMask3D: Open-Vocabulary 3D Instance Segmentation. In: NeurIPS (2023) 2, 3, 4, 9, 13, 14" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 483, + 482, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 483, + 482, + 515 + ], + "spans": [ + { + "bbox": [ + 130, + 483, + 482, + 515 + ], + "type": "text", + "content": "52. Takmaz, A., Schult, J., Kaftan, I., Akçay, M., Leibe, B., Sumner, R., Engelmann, F., Tang, S.: 3D Segmentation of Humans in Point Clouds with Synthetic Data. In: ICCV (2023) 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 515, + 482, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 515, + 482, + 546 + ], + "spans": [ + { + "bbox": [ + 130, + 515, + 482, + 546 + ], + "type": "text", + "content": "53. Thomas, H., Qi, C.R., Deschaud, J.E., Marcotegui, B., Goulette, F., Guibas, L.J.: KPCov: Flexible and Deformable Convolution for Point Clouds. In: ICCV (2019) 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 547, + 482, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 547, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 547, + 482, + 568 + ], + "type": "text", + "content": "54. Vu, T., Kim, K., Luu, T.M., Nguyen, T., Yoo, C.D.: SoftGroup for 3D Instance Segmentation on Point Clouds. In: CVPR (2022) 2, 3, 10" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 569, + 482, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 590 + ], + "type": "text", + "content": "55. Weder, S., Blum, H., Engelmann, F., Pollefeys, M.: LabelMaker: Automatic Semantic Label Generation from RGB-D Trajectories. In: 3DV (2024) 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 590, + 482, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 590, + 482, + 622 + ], + "spans": [ + { + "bbox": [ + 130, + 590, + 482, + 622 + ], + "type": "text", + "content": "56. Yang, B., Wang, J., Clark, R., Hu, Q., Wang, S., Markham, A., Trigoni, N.: Learning Object Bounding Boxes for 3D Instance Segmentation on Point Clouds. In: NeurIPS (2019) 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 622, + 482, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 622, + 482, + 643 + ], + "spans": [ + { + "bbox": [ + 130, + 622, + 482, + 643 + ], + "type": "text", + "content": "57. Yang, Y., Wu, X., He, T., Zhao, H., Liu, X.: SAM3D: Segment Anything in 3D Scenes. In: ICCVW (2023) 2, 4, 5, 10, 11, 12" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 643, + 482, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 643, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 643, + 482, + 665 + ], + "type": "text", + "content": "58. Yeshwanth, C., Liu, Y.C., Nießner, M., Dai, A.: ScanNet++: A High-Fidelity Dataset of 3D Indoor Scenes. In: ICCV (2023) 3, 9, 10" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 399, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 399, + 91, + 447, + 102 + ], + "type": "text", + "content": "Segment3D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 270 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 148 + ], + "type": "text", + "content": "59. Yi, L., Zhao, W., Wang, H., Sung, M., Guibas, L.J.: GSPN: Generative Shape Proposal Network for 3D Instance Segmentation in Point Cloud. In: CVPR (2019) 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 149, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 480, + 171 + ], + "type": "text", + "content": "60. Yue, Y., Das, A., Engelmann, F., Tang, S., Lenssen, J.: Improving 2D Feature Representations by 3D-Aware Fine-Tuning. In: ECCV (2024) 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "type": "text", + "content": "61. Yue, Y., Kontogianni, T., Schindler, K., Engelmann, F.: Connecting the Dots: Floorplan Reconstruction Using Two-Level Queries. In: CVPR (2023) 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 225 + ], + "type": "text", + "content": "62. Zeng, Y., Jiang, C., Mao, J., Han, J., Ye, C., Huang, Q., Yeung, D.Y., Yang, Z., Liang, X., Xu, H.: CLIP2: Contrastive Language-Image-Point Pretraining from Real-World Point Cloud Data. In: CVPR (2023) 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "text", + "content": "63. Zhang, R., Guo, Z., Zhang, W., Li, K., Miao, X., Cui, B., Qiao, Y., Gao, P., Li, H.: PointCLIP: Point Cloud Understanding by CLIP. In: CVPR (2022) 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "text", + "content": "64. Zurbrugg, R., Liu, Y., Engelmann, F., Kumar, S., Hutter, M., Patil, V., Yu, F.: ICGNet: A Unified Approach for Instance-Centric Grasping. In: ICRA (2024) 2" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Segmentation-guided Layer-wise Image Vectorization with Gradient Fills/01f8c0e2-d236-4d33-947c-2747a133c0bb_content_list.json b/2024/Segmentation-guided Layer-wise Image Vectorization with Gradient Fills/01f8c0e2-d236-4d33-947c-2747a133c0bb_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d3b7dd65013ccb44b945f364a2ef44bce1793afe --- /dev/null +++ b/2024/Segmentation-guided Layer-wise Image Vectorization with Gradient Fills/01f8c0e2-d236-4d33-947c-2747a133c0bb_content_list.json @@ -0,0 +1,2308 @@ +[ + { + "type": "text", + "text": "Segmentation-guided Layer-wise Image Vectorization with Gradient Fills", + "text_level": 1, + "bbox": [ + 277, + 140, + 725, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hengyu Zhou, Hui Zhang, and Bin Wang", + "bbox": [ + 320, + 210, + 681, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "School of Software, Tsinghua University, P. R. China \nzhouhy22@mails.tsinghua.edu.cn, {huizhang, wangbins}@tsinghua.edu.cn", + "bbox": [ + 241, + 239, + 759, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. The widespread use of vector graphics creates a significant demand for vectorization methods. While recent learning-based techniques have shown their capability to create vector images of clear topology, filling these primitives with gradients remains a challenge. In this paper, we propose a segmentation-guided vectorization framework to convert raster images into concise vector graphics with radial gradient fills. With the guidance of an embedded gradient-aware segmentation subroutine, our approach progressively appends gradient-filled Bézier paths to the output, where primitive parameters are initiated with our newly designed initialization technique and are optimized to minimize our novel loss function. We build our method on a differentiable renderer with traditional segmentation algorithms to develop it as a model-free tool for raster-to-vector conversion. It is tested on various inputs to demonstrate its feasibility, independent of datasets, to synthesize vector graphics with improved visual quality and layer-wise topology compared to prior work.", + "bbox": [ + 261, + 306, + 743, + 517 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Vectorization $\\cdot$ Segmentation $\\cdot$ Differentiable rendering", + "bbox": [ + 261, + 529, + 718, + 544 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 570, + 375, + 585 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vector graphics offer great flexibility in digital design as they can be easily edited and arbitrarily scaled. Vectorization, the procedure of converting raster images to vector ones, serves as a second way to creating vector graphics other than designing from scratch where extensive artistic skills are required, and also as a bridge between the rapidly developing raster image generation [27] and the relatively less studied vector generation [3,8,10,24,32].", + "bbox": [ + 212, + 602, + 787, + 694 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vectorization has been explored for decades with various representations proposed to divide images into non-overlapping regions [17, 29, 33, 35, 36]. Despite their capability of generating vivid vectorization of realistic images, the complex primitives and lack of hierarchy make the vector output less intuitive for manipulation.", + "bbox": [ + 212, + 694, + 787, + 768 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent learning-based methods show their potential in preserving image hierarchy, where a vector image is often considered as a list of primitives and their order indicates the structure of the input. Many deep learning approaches", + "bbox": [ + 212, + 770, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\boxtimes$ Corresponding authors.", + "bbox": [ + 215, + 823, + 392, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "have been proposed to vectorize simple inputs [24] or images of a specific field of interest [6, 18, 25, 28], but their dependency on models limits them to a particular domain and are not trivially generalizable. LIVE [20], in contrast, utilizes DiffVG [16], a differentiable renderer, to present a model-free vectorization framework. It progressively translates a raster image into an SVG in a layer-wise hierarchy, through which the topology of the input is preserved within the order of geometric primitives. However, its lack of support for gradients results in excessive primitives being added in case of images with rich gradient effects, and adding such support is not as simple as replacing RGBA colors with gradient parameters, as elaborated in Sec. 3.4.", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a novel segmentation-guided vectorization framework that extends the capability of LIVE to support radial gradients. The additional parameters of a radial gradient pose an increased challenge to optimization, where an effective method to determine whether a pixel contributes to a path's gradient fill is necessary. The key insight behind our idea is the similarity between finding contributing pixels and segmentation tasks, based on which we designed a segmentation-guided initialization procedure to progressively append new shapes to the vector output, and a novel loss function to optimize their geometric and gradient parameters. We evaluated our framework on several datasets with quantitative metrics and a user study, to show its effectiveness and superior performance compared to previous work.", + "bbox": [ + 212, + 297, + 787, + 464 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize our contributions:", + "bbox": [ + 240, + 464, + 477, + 477 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a segmentation-guided vectorization framework to create vector graphics automatically with layer-wise hierarchy and radial gradients.", + "- We propose a gradient-aware segmentation method to evaluate the pixel-wise contribution to the geometric and gradient parameters of a path.", + "- We take the segmentation as guidance for our new initialization technique and as a part of our novel segmentation-guided loss." + ], + "bbox": [ + 223, + 488, + 782, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 601, + 387, + 617 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1 Image Vectorization", + "text_level": 1, + "bbox": [ + 215, + 633, + 428, + 648 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Most traditional vectorization methods aim to create vector images of high fidelity with different representations, which could be roughly categorized into mesh-based ones and curve-based ones. The former representations divide the input image into non-overlapping 2D patches across which colors are interpolated [30]. Shapes of patches include triangular [17,33,36], rectangular [1,29] or irregular ones such as bezigons [35]. The different selection of mesh shapes determines how patches are organized and how colors are interpolated within patches. The curved-based representations decompose the image into curves. Diffusion curves, for instance, use curves as geometric primitives with colors defined on sides of the curves [22,34]. While these methods can yield near-photo-realistic results, they may fall short when it comes to ease of editing due to the complex primitives and loss of topological information, compared to a plain list of", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "H. Zhou et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "simple primitives like how an SVG organizes a vector image. Recent learning-based vectorization work mainly takes the list-of-primitive approach, where the ordered primitives are seen as a sequence of drawing operations, and sequential prediction methods are applied to synthesize vector graphics. Models are used including recurrent neural networks [10,24], transformers [3], and are often combined with variational autoencoders [10,15,19,24]. DiffVG [16] as a differentiable renderer fills the gap between raster and vector graphics, with which loss functions on raster images could be used directly to optimize the vector images. A vectorization method optimizing all paths at once is also proposed in the work.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 Image Topology", + "text_level": 1, + "bbox": [ + 214, + 303, + 395, + 318 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "When synthesizing vector graphics, both similarity to the original image and correct topology are important for the ease of subsequent human manipulation. A related problem is layer decomposition, where images are decomposed into semi-transparent layers. With these layers being vector paths, such a method could serve as a layered vectorization approach. An interactive method [26] is among the first proposed to convert bitmaps into layered vector graphics. Photo2ClipArt [7] and other similar work [5] replace the heavy manual interaction with a user-provided segmentation input, from which the decomposition could be automatically determined. However, these methods require a concise segmentation for efficient and effective vectorization into a relatively small number of paths, where automatic segmentation algorithms struggle to serve the purpose [7].", + "bbox": [ + 212, + 325, + 787, + 507 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Many learning-based methods formulate vector synthesis as sequential prediction problems, through which the order of primitives is naturally preserved. Some methods predict the primitive parameters directly. For example, Egiazarian et al. [6] use a transformer-based network to directly generate the vectorization of technical line drawings; MARVEL [28] applies deep reinforcement learning to predict strokes in black-and-white comics. Meanwhile, some approaches use variational autoencoders (VAE) [15] to create vector graphics. SketchRNN [10] is the first to introduce a Long Short-Term Memory (LSTM) [12] based VAE to the representation of vector sketches. The later DeepSVG [3] proposed a transformer-based VAE network that also encodes a vector image into a latent code, which could be decoded to reconstruct the vector paths. SVG-VAE [19] is the first to use raster images as the input to its encoder, thus it could be used to vectorize images. Im2Vec [24] also takes raster images as its input but drops the need for vector supervision with the help of DiffVG [16]. While these methods show great potential in creating topology-aware vector graphics, they rely on specific training datasets and may overlook the fine details in a complex input. The most similar work to ours is LIVE [20], in which a progressive framework is proposed to convert raster images into layered vector paths utilizing DiffVG for optimization, but its missing support for color gradients is not trivially achievable by simply optimizing gradient parameters. Our novel segmentation-guided method takes a step further to handle gradient effects properly, at the same level of simplicity as LIVE, requiring neither additional user input nor deep models.", + "bbox": [ + 212, + 508, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Segmentation-guided Layer-wise Image Vectorization with Gradient Fills", + "bbox": [ + 246, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 143, + 330, + 159 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Method Overview", + "text_level": 1, + "bbox": [ + 215, + 181, + 410, + 195 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3dec88129d8fb75647595f2a55b5411babe110ce346b8e2277518d2d7d74b425.jpg", + "image_caption": [ + "Fig. 1: Overview of our framework. $U + 1F61A$ from Noto Emoji [9] is used for demonstration." + ], + "image_footnote": [], + "bbox": [ + 220, + 236, + 785, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We present a new framework to generate SVG figures from rasterized images, with support for radial gradient fills via a segmentation-guided initialization and optimization process.", + "bbox": [ + 212, + 388, + 782, + 434 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our framework works progressively, where at each epoch, single or multiple Bézier paths are added and optimized. Fig. 1 provides an overview of our method. At the beginning of each epoch $i$ , we calculate the difference between the input raster image and the output $I_{i-1}$ from the previous epoch. We segment it with our gradient-aware segmentation (Sec. 3.2), from which $n_i$ segmented regions are selected for initialization of new paths (Sec. 3.3). All added paths, including those added in previous epochs, are optimized to minimize the vectorization loss (Sec. 3.4). The parameters to be optimized consist of geometric parameters including positions of curve control points and gradient parameters including their center, radius, and color stops. $n_i$ , or the number of paths to be appended at $i$ -th epoch is specified by the user along with the number of iteration epochs.", + "bbox": [ + 212, + 436, + 787, + 603 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2 Gradient-aware Segmentation", + "text_level": 1, + "bbox": [ + 215, + 628, + 504, + 645 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a9b54d52e89be1addffb28f0d37b1c8c91702903a19b122f40156d2b543e3d94.jpg", + "image_caption": [ + "(a) Input" + ], + "image_footnote": [], + "bbox": [ + 282, + 686, + 362, + 753 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0c44a11ab887e4de40cfa39657c07e24d819e25682a4e5895ace2abc6176219f.jpg", + "image_caption": [ + "(b) LIVE" + ], + "image_footnote": [], + "bbox": [ + 372, + 686, + 450, + 752 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ef37de576e9696f8f12ab2de5d11fbfa263c22519b1fbe56a557d2703addd118.jpg", + "image_caption": [ + "(c) LIVE-30" + ], + "image_footnote": [], + "bbox": [ + 462, + 686, + 540, + 752 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/272a1c1e33a91fac29924bdad3320a8bea65678f33dc786df825f3a1b9af4993.jpg", + "image_caption": [ + "Fig. 2: Comparisons on segmentation methods. (a) is the raster input. (b) is by LIVE under default settings clustering colors into 200 bins. (c) is also by LIVE but with the number of bins set to 30. (d) is segmented using the Mean-shift [4] algorithm. (e) is by our method." + ], + "image_footnote": [], + "bbox": [ + 550, + 686, + 630, + 753 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b81f23cd872c02c150cefc4cebc05aef534eba67305943d94207dfa791c7f6ef.jpg", + "image_caption": [ + "(d) Mean shift", + "(e) Ours" + ], + "image_footnote": [], + "bbox": [ + 642, + 686, + 720, + 753 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "H. Zhou et al.", + "bbox": [ + 271, + 114, + 366, + 126 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9bee42c81b824927e2bd089eb81b535cf4f6e949900277643d8dcc8ef4f4217d.jpg", + "image_caption": [ + "Fig. 3: Step-by-step outputs of our gradient-aware segmentation" + ], + "image_footnote": [], + "bbox": [ + 230, + 143, + 769, + 327 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Quality of path initialization is the key to effective vectorization, hence we present a gradient-aware segmentation method to better estimate the area that a path is likely to span over. Its result is used in the subsequent operations including initialization and optimization (Sec. 3.3, Sec. 3.4).", + "bbox": [ + 212, + 381, + 784, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "When paths are filled with a solid color, a connected component of similar or identical colors has a good chance of being covered by one path. LIVE [20] designed a component-wise initialization method, where pixels are clustered into buckets based on their l2-length over RGB channels, and connected pixels in the same bucket are considered one component. As gradients are considered, the clustering algorithm used by LIVE may result in an excessive segmentation, as shown by Fig. 2. Other clustering methods including Mean-shift [4] also suffer from over-segmentation for the same reason that colors within a path may vary more than colors between paths.", + "bbox": [ + 212, + 441, + 784, + 577 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With the limitations of previous methods, our approach is designed to detect edges of gradients. Colors should derive smoothly inside a region filled with the same gradient fill and are likely to change abruptly at its boundary. Thus, we calculate the secondary spatial gradient with a discrete Laplacian filter $L = \\left[ \\begin{array}{lll}1 & 1 & 1\\\\ 1 & -8 & 1\\\\ 1 & 1 & 1 \\end{array} \\right]$ to identify such rapid changes at gradient boundaries. Our segmentation method features the following steps:", + "bbox": [ + 212, + 578, + 784, + 675 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- The cross-correlation $S_0 \\in \\mathbb{R}_{w \\times h \\times 3}$ between the difference to the input and the 2D Laplacian filter is calculated as", + "bbox": [ + 223, + 681, + 785, + 712 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS _ {0} = \\operatorname {c o r r e l a t e} ((I - \\hat {I}) \\mathbf {1} _ {| | \\hat {I} - I | | _ {2} > \\epsilon}, L),\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 720, + 645, + 741 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{I}$ is the target and $I$ is the synthesized image. Pixels with an error below $\\epsilon = 0.1$ are excluded.", + "bbox": [ + 240, + 750, + 782, + 777 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- $S_0$ is then summed over its RGB channels for a grayscale image $S_1 \\in \\mathbb{R}_{w \\times h}$ , where each pixel of $S_1$ has a value of $(S_1)_{ij} = \\sum_{c=1}^{3} |(S_0)_{jc}|$ .", + "- The grayscale image $S_{1}$ is converted to a binary image $S_{2} = \\mathbf{1}_{S_{1} > \\mathrm{Otsu}(S_{1})}$ with the threshold determined via Otsu's method [23]." + ], + "bbox": [ + 223, + 780, + 782, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Segmentation-guided Layer-wise Image Vectorization with Gradient Fills", + "bbox": [ + 246, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/83c3d80cb12b90f85cfc81b0155549a3862c5323e21685942860f29fcb242084.jpg", + "image_caption": [ + "Fig. 4: With paths being progressively added, the overall difference between the current output and the raster input decreases, thus we apply Otsu's method [23] to automatically determine the threshold for binarization. In this example, the red cheeks get segmented after more significant differences including eyes and brows are fitted, thanks to the dynamic threshold. The differences are normalized and mapped to grayscale." + ], + "image_footnote": [], + "bbox": [ + 220, + 143, + 759, + 291 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- Morphological closing [11] and the watershed algorithm [2] are then applied to $S_{2}$ for the final segmentation $S \\in \\mathbb{N}_{h \\times w}$ .", + "bbox": [ + 223, + 402, + 787, + 435 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Fig. 3 we showcase some intermediate results after each step. Since we are segmenting the difference between the output and the target, the already fitted pixels are ignored and the automatically determined threshold via Otsu's method decreases in response to the descending overall difference, as shown in Fig. 4. Compared to a fixed threshold, our method makes fewer assumptions about the input and avoids hyperparameters.", + "bbox": [ + 212, + 450, + 789, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Segmentation-guided Initialization", + "text_level": 1, + "bbox": [ + 215, + 569, + 545, + 585 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our vectorization method works in a progressive manner, throughout which we add one or more paths at each epoch. These newly added paths are initialized using our segmentation $S$ as the guidance. For each path to be added, we select the segmentation region with the largest accumulated square error calculated as", + "bbox": [ + 212, + 599, + 787, + 660 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nw _ {i} = \\sum_ {p \\in \\tilde {S} [ i ]} | | I _ {p} - \\hat {I} _ {p} | | ^ {2} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 421, + 674, + 785, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $p$ is iterated over all pixels in $i$ -th region. This approach prioritizes larger regions to encourage a hierarchical initialization order and prevents regions that are almost properly filled from being chosen. A circle path of four cubic Bezier curves [20] is added at the selected region's center of mass $p_m$ . We fill the path with a radial gradient, which centers at $p_m$ , with a diameter equal to the geometric mean of the width and the height of the region's bounding box clipped to [0.2, 1.0]. The two stop colors at offsets $0\\%$ and $100\\%$ are both initialized to the color of the input image $I$ at $p_m$ .", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "H. Zhou et al.", + "bbox": [ + 271, + 114, + 367, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/271aab27b9072f3beb784dde32be0362dbebad789c2aabff804240410a5615f5.jpg", + "image_caption": [ + "Fig. 5: Optimizing course for our method in comparison with UDF loss from LIVE. With our proposed segmentation-guided weight (SG Weight), the gradient fill is optimized to minimize the color error not only on the contour as UDF weight focuses on but also on colors inside the path, while excluding pixels occluded by eyes, brows, and the mouth." + ], + "image_footnote": [], + "bbox": [ + 251, + 146, + 754, + 345 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Loss Function with Segmentation-guided weight", + "text_level": 1, + "bbox": [ + 215, + 462, + 656, + 478 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A common approach to measuring the difference between the synthesized image $\\hat{I}$ and the target image $I$ is through mean squared error (MSE). However, simply minimizing MSE results in colors biasing towards the average color inside a path, as suggested by LIVE [20]. LIVE tackles the problem with its designed UDF loss to focus on pixels on the contour, but for radial gradients, correct colors on the contour do not mean correctness inside, as indicated by Fig. 5. When gradients are optimized with UDF loss, colors at the center are not fitted well, since those pixels are ignored by the UDF weight. We draw from the insight by LIVE emphasizing the significance of pixels on edges and extend this concept to include all pixels within a path, except for those occluded by other paths, which is well estimated through our gradient-aware segmentation.", + "bbox": [ + 212, + 491, + 787, + 656 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We formulate our segmentation-guided loss with this estimation of unoccluded pixels within a path. For each added path $p_i$ , given the set of pixels covered by the path and the set of pixels within the segment from which the path is initialized, we take their intersection and mark these pixels as being focused. The union of all focused pixels forms a set $F$ . We define the weight $w_{\\mathrm{SG}}$ as:", + "bbox": [ + 212, + 657, + 787, + 746 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nw _ {\\mathrm {S G}} (i) = \\left\\{ \\begin{array}{l l} \\max \\left(d _ {i} ^ {\\prime}, \\alpha_ {s}\\right), & i \\in F \\\\ d _ {i} ^ {\\prime} \\left(1 - \\alpha_ {s}\\right), & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 366, + 746, + 784, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $i$ is an index of pixel and $d_{i}^{\\prime}$ is the UDF weight from [20]. The UDF weight is introduced to give the pixels near the path contour a higher weight, as depicted in Fig. 5. $\\alpha_{s}$ balances between the UDF weight $d_{i}^{\\prime}$ and our segmentation weight,", + "bbox": [ + 212, + 794, + 787, + 842 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Segmentation-guided Layer-wise Image Vectorization with Gradient Fills", + "bbox": [ + 246, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1d72db136b7302ef3da7da709d00c1691c76b0f2b3bab6cc3232ee9ed74d1299.jpg", + "image_caption": [ + "Fig. 6: Bézier paths with four control points" + ], + "image_footnote": [], + "bbox": [ + 356, + 143, + 506, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d2801467b160c455860087ca738f3f599776b5f96a363dc31336c188c6357722.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 521, + 143, + 643, + 219 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and is set to 0.6 empirically. $d_i^\\prime$ is defined with:", + "bbox": [ + 215, + 272, + 557, + 287 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nd _ {i} ^ {\\prime} = \\frac {\\operatorname {R e L U} (\\tau - | d _ {i} |)}{\\sum_ {j = 1} ^ {w \\times h} \\operatorname {R e L U} (\\tau - | d _ {j} |)}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 397, + 295, + 785, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $d_{i}$ is the distance from the pixel $i$ to the nearest path contour, which is then normalized and saturated with threshold $\\tau = 10$ .", + "bbox": [ + 214, + 342, + 782, + 371 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We re-weight the pixel-wise color error with our proposed weight $w_{\\mathrm{SG}}$ for the final segmentation-guided loss:", + "bbox": [ + 214, + 372, + 784, + 402 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {S G}} = \\frac {1}{3} \\sum_ {i = 1} ^ {w \\times h} w _ {\\mathrm {S G}} (i) \\sum_ {c = 1} ^ {3} \\left(I _ {c, i} - \\hat {I} _ {c, i}\\right) ^ {2}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 412, + 785, + 454 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $i$ is iterated over all pixels and squared color differences are averaged over the RGB channels.", + "bbox": [ + 214, + 460, + 782, + 489 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also introduce Xing loss [20] to help relieve self-intersection. Given a cubic Bézier curve with four control points $A, B, C$ , and $D$ , self-intersection is more likely to occur when the angle between $AB$ and $CD$ is greater than $180^{\\circ}$ , as shown in Fig. 6. The Xing loss $\\mathcal{L}_{\\mathrm{Xing}}$ is defined to penalize angles over $180^{\\circ}$ :", + "bbox": [ + 214, + 491, + 784, + 551 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nD _ {1} = \\left\\{ \\begin{array}{l} 1, \\boldsymbol {A B} \\times \\boldsymbol {B C} > 0 \\\\ 0, \\text {o t h e r w i s e} \\end{array} , D _ {2} = \\frac {\\boldsymbol {A B} \\times \\boldsymbol {C D}}{\\| \\boldsymbol {A B} \\| \\| \\boldsymbol {C D} \\|} \\right. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 318, + 561, + 785, + 604 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {X i n g}} = D _ {1} \\left(\\operatorname {R e L U} (- D _ {2})\\right) + (1 - D _ {1}) \\left(\\operatorname {R e L U} \\left(D _ {2}\\right)\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 320, + 616, + 785, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our optimization objective is defined as follows:", + "bbox": [ + 238, + 638, + 588, + 652 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {S G}} + \\lambda \\mathcal {L} _ {\\text {X i n g}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 429, + 662, + 785, + 680 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\lambda$ is set to 0.05 empirically.", + "bbox": [ + 215, + 689, + 460, + 704 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 726, + 375, + 743 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 755, + 455, + 771 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We implement our framework with DiffVG [16] renderer based on PyTorch. Adam optimizer [14] is used with a learning rate of $10^{-2}$ and 1 for gradient parameters and path points respectively. We use scikit-image [31] for its implementation of morphological operations and the watershed algorithm [2].", + "bbox": [ + 214, + 779, + 782, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "H. Zhou et al.", + "bbox": [ + 271, + 114, + 367, + 127 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5a939afbd6df4c936c9d83d10a475c62c34953ff75ff87491ceb8177d71491c8.jpg", + "image_caption": [ + "Noto" + ], + "image_footnote": [], + "bbox": [ + 300, + 147, + 728, + 195 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/6fe04ddb176c3692179b4337c8e29b9dfbc4b736e9553b6cc93df308dda46f7f.jpg", + "image_caption": [ + "Fluent" + ], + "image_footnote": [], + "bbox": [ + 300, + 200, + 725, + 247 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/1c6ae29aafc1fbee37bd2291f5399d7e2510bf77a3b8d2f95c9f38b43bdab7cf.jpg", + "image_caption": [ + "onont", + "Fig. 7: Exemplars from our chosen datasets for evaluation" + ], + "image_footnote": [], + "bbox": [ + 299, + 251, + 727, + 297 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Datasets", + "text_level": 1, + "bbox": [ + 215, + 356, + 333, + 369 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Noto Emoji We randomly select 256 emojis from Noto Emoji [9](Unicode 15.0) varying in colors and genres. In contrast to previous work [20,24], we take a more recent version of the dataset, consisting of resigned emojis filled with gradients instead of single colors. Emojis from this dataset consist of a relatively small number of paths with a clear hierarchy, thus we evaluate on this dataset to show our method's capability to decompose images into layers with gradients.", + "bbox": [ + 212, + 381, + 784, + 472 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Fluent Emojis We select an additional 256 images at random from Microsoft's Fluent Emojis [21] collection. This dataset features colorful emojis with rich gradient details. We test on this dataset to elaborate on our method's ability to match up with target images using fewer paths compared to previous work [16, 20].", + "bbox": [ + 212, + 488, + 784, + 550 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Iconfont A set of 128 vector arts are fetched from the online Iconfont Illustration Library [13]. Images in the dataset contain an average of 1020 paths with great details but no gradients. We use these images to exhibit that our segmentation-guided approach is also suitable for vectorizing complex inputs, and performs on par with or better than previous work even without gradients.", + "bbox": [ + 212, + 564, + 784, + 638 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "It is worth noticing that our framework is model-free, thus datasets mentioned above are for sole evaluation purposes.", + "bbox": [ + 212, + 640, + 784, + 670 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3 Qualitative Comparison", + "text_level": 1, + "bbox": [ + 215, + 693, + 457, + 708 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "With our initial goal of supporting gradient fills in a topology-preserving vectorization approach, we conducted a comparative analysis of the visual quality of our reconstructed vector graphics against LIVE [20]. Our method achieves superior visual quality in vectorization while utilizing the same number of paths as LIVE, as evidenced in Fig. 8.", + "bbox": [ + 212, + 719, + 784, + 794 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "As depicted in Fig. 9, when presented with input containing gradients, LIVE struggles to accurately vectorize gradient-filled facial features, despite correctly capturing other facial parts. In contrast, our approach accurately reconstructs", + "bbox": [ + 212, + 795, + 784, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Segmentation-guided Layer-wise Image Vectorization with Gradient Fills", + "bbox": [ + 246, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/3a2a806be6bbfa4f15d50be9e7c6b8f40eb06747cc1efb06baa1b77796f4938c.jpg", + "image_caption": [ + "Fig. 8: Qualitative comparisons with LIVE" + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 787, + 329 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "the input using concise paths. Additionally, attempts to optimize gradient parameters without guidance, as seen in the 'w/o guidance' rows, result in degraded outcomes.", + "bbox": [ + 212, + 387, + 787, + 433 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.4 Quantitative Comparison", + "text_level": 1, + "bbox": [ + 215, + 465, + 470, + 479 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We further adopt quantitative metrics in a quantitative comparison with LIVE. We calculated PSNR to measure the difference between the rendered vectorized outputs and their corresponding raster inputs. For Noto and Fluent Emoji datasets, we add $\\mathrm{clamp}(2^{i - 2},1,32)$ paths at $i$ -th initialization to reach a total of 256 paths. For the Iconfont dataset, 512 paths are used for each image with $\\mathrm{clamp}(2^{i - 2},1,64)$ paths being added at $i$ -th epoch, for its higher complexity.", + "bbox": [ + 212, + 500, + 787, + 590 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "As reported in Fig. 10, our method achieves generally faster convergence than LIVE, especially when a small number of paths are added. Our segmentation-guided framework can capture large-scale gradient features where LIVE tends to emit superfluous paths. With excessive paths being added, both methods yield the same level of quality in terms of PSNR.", + "bbox": [ + 212, + 592, + 787, + 667 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.5 Layer Decomposition", + "text_level": 1, + "bbox": [ + 215, + 699, + 434, + 715 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The framework we propose captures layer-wise structure during our progressive vectorization. The segmentation-guided initialization prioritizes segments that are more significant in terms of accumulated error, while details with relatively small errors are captured later by our adaptive gradient-aware segmentation, as shown in Fig. 4. Given an input with a clear hierarchy, these progressively added ordered paths resemble the handcrafted vector graphics, creating an easy-to-edit output, as shown in Fig. 11.", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "H. Zhou et al.", + "bbox": [ + 271, + 114, + 366, + 126 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/314e5d14161c9505484e34ac184b4e1e451d87b390e33cab8a1651ac2de3ddac.jpg", + "image_caption": [ + "Fig.9: A comparison on number of paths needed to reconstruct major components. 'w/o guidance' refers to gradient fills being added without our proposed segmentation guidance. For the first input ( $U + 1F61A$ from Noto Emoji), our method vectorizes all elements with 8 paths, while adding more paths does not deteriorate the output. For the second input ( $U + 1F605$ from Fluent Emoji), ours reconstructs the facial parts with 6 paths and achieves a close gradient effect to the input using 16 paths." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 787, + 407 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/0e2b8445c99e9d3f2c42d5c891e7b42b7fc8769a9748d9491eb379dea2eb6b9b.jpg", + "image_caption": [ + "Fig. 10: Quantitative comparison between our method and LIVE. Our method achieves significantly lower error when the numbers of paths are small and converge to equal performance with excessive paths being added." + ], + "image_footnote": [], + "bbox": [ + 227, + 521, + 776, + 657 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/29055c6a678f70d459984fcc6d2ca4d905d4e55a1170323c7b79c466b734bc67.jpg", + "image_caption": [ + "(a) Source" + ], + "image_footnote": [], + "bbox": [ + 282, + 727, + 364, + 792 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/7f5f729c936f9d9b4bfd42085c78e71cdc104b9874ef4fae90b49ef079d797c3.jpg", + "image_caption": [ + "(b) Vectorized" + ], + "image_footnote": [], + "bbox": [ + 372, + 728, + 455, + 792 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/e47813adba443833613a02bcac7fabb566f617b7658d3c28498cf3f4dc675496.jpg", + "image_caption": [ + "(c) Decomposed layers" + ], + "image_footnote": [], + "bbox": [ + 467, + 728, + 620, + 792 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/799b3b2d269561edaed1789aa892fa575fcf94f1e92a9072fddca9e3cfdec989.jpg", + "image_caption": [ + "(d) Edited", + "Fig. 11: Decomposed layers and recoloring" + ], + "image_footnote": [], + "bbox": [ + 635, + 728, + 720, + 792 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Segmentation-guided Layer-wise Image Vectorization with Gradient Fills", + "bbox": [ + 246, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 1: Results collected from our user study. The row 'Total' stands for the total number of answers collected for all the images in the dataset vectorized given the number of paths. As shown by the results, users in general prefer the vectorization by our method.", + "bbox": [ + 212, + 143, + 491, + 241 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/3aba846de821633cdeda0e4bdf6ba1c90c11b1e3c907ec4e43095d2f82858e96.jpg", + "table_caption": [ + "(a) Noto Emoji" + ], + "table_footnote": [], + "table_body": "
#Paths8163264Overall
LIVE39.0%39.0%40.8%43.0%40.4%
Ours61.0%61.0%59.2%57.0%59.6%
Total4494414414541785
", + "bbox": [ + 215, + 290, + 493, + 357 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/38c2cc24012b8921aecf766084d7b29cac7bb3c0361c28b77159a8d77407bc3c.jpg", + "table_caption": [ + "(b) Fluent Emoji" + ], + "table_footnote": [], + "table_body": "
#Paths8163264Overall
LIVE40.2%38.4%30.8%29.3%34.7%
Ours59.8%61.6%69.2%70.7%65.3%
Total4334454524261756
", + "bbox": [ + 215, + 391, + 493, + 458 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/490990bdd7bfcb69848a54ce86644b70cae4aff48f226dc428a49f733430790c.jpg", + "table_caption": [ + "(c) Iconfont" + ], + "table_footnote": [], + "table_body": "
#Paths3264128256Overall
LIVE32.2%42.7%48.5%44.3%42.1%
Ours67.8%57.3%51.5%55.7%57.9%
Total214241227237919
", + "bbox": [ + 217, + 492, + 493, + 559 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 2: Results of our ablation study in PSNR. 'G' in the header is for 'with gradients' and 'S' is for 'with segmentation guidance'. The first row, with neither gradients nor segmentation, stands for LIVE.", + "bbox": [ + 508, + 143, + 787, + 214 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/6dd969033e304d5d203acc118fbacfe22194c67481a296cab1e67bce02bb7cc6.jpg", + "table_caption": [ + "(a) Noto Emoji" + ], + "table_footnote": [], + "table_body": "
G SN=8163264128256
24.9228.2131.4233.8335.8937.16
✓25.3328.6731.6334.1336.1237.34
✓25.1928.5131.8034.1736.0537.29
✓25.6928.9531.9234.2236.3037.46
", + "bbox": [ + 511, + 263, + 785, + 340 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/0a17e529b8d09f1e799f9cb383b5397646b0c46ee6226c15f399989408aa1043.jpg", + "table_caption": [ + "(b) Fluent Emoji" + ], + "table_footnote": [], + "table_body": "
G SN=8163264128256
25.1928.7431.7434.4636.8538.38
✓26.2629.3632.2234.6736.9638.52
26.1529.5932.6435.2837.5438.91
✓26.7330.0732.9535.4337.6138.84
", + "bbox": [ + 511, + 376, + 785, + 455 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/f2a417a30ad9b4ebe1abcfa6a20f3feef9c2cc7d98f1d060345ddbe7520b4226.jpg", + "table_caption": [ + "(c) Iconfont" + ], + "table_footnote": [], + "table_body": "
G SN=163264128256512
19.2221.6924.3827.2630.2532.36
✓20.0222.0224.2926.8629.4331.44
✓19.4221.9624.7227.6030.6732.72
✓20.3122.3524.7327.2829.7631.74
", + "bbox": [ + 511, + 489, + 785, + 569 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.6 User Study", + "text_level": 1, + "bbox": [ + 215, + 602, + 356, + 617 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We conducted a questionnaire survey to obtain subjective feedback from users. Each participant is presented with 20 questions randomly selected from a total of 2,560 questions, that is, 4 numbers of paths multiplied by 640 images from the three datasets. In each question, participants are shown a raster input and two vectorized versions of it, one using LIVE and the other using our method, with the same number of paths. They are asked to choose the one that looks closer to the original image or, at an equivalent level of similarity, looks more appealing. Figure 12 shows examples.", + "bbox": [ + 212, + 640, + 787, + 762 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In our study, 223 people engaged, contributing a total of 4,460 votes. The results, presented in Tab. 1, reveal a clear preference for our approach. In the Fluent Emoji dataset, our method is notably preferred for these images with rich gradients. Despite the absence of gradients in many images from Noto Emoji and Iconfont, our method remains the preferred choice.", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "H. Zhou et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/3f03120f097d551ba6d703a1612e5360658d4314e6a4262652f60d0439d342af.jpg", + "image_caption": [ + "$U + 1F31A$ from Noto $(\\mathrm{N} = 8)$ B is ours" + ], + "image_footnote": [], + "bbox": [ + 217, + 146, + 277, + 191 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/ac831756a9ef06ce24a81dbc99e9f7ed824d9e3bfbc07784e87fed7449a995ba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 279, + 146, + 339, + 191 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/ab8bebc98ecda06a5b0bba83f7f139751adc9ad04a056ec164111a3235d056fd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 344, + 146, + 403, + 191 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/dec3af428d2f61b369eef8f6c6ce3af1afd5fdc83187a77a1c31bc0307f24a82.jpg", + "image_caption": [ + "$U + 1F613$ from Fluent $(\\mathrm{N} = 16)$ B is ours" + ], + "image_footnote": [], + "bbox": [ + 406, + 146, + 465, + 191 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/124a12716deafdc0cb609d07de31b5c6fd55d35e9ca6080d67271a436a4a76c6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 467, + 146, + 527, + 191 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/94ab970c50e82372c67ee63562aaddfe82377609c16ca5e013ac2812d658ba84.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 534, + 146, + 591, + 191 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/0c6574c16804ba2649d8e8095898efde1e2942f971bcb1f6e9e93fb722a949da.jpg", + "image_caption": [ + "46087_07 from Iconfont (N=128); A is ours" + ], + "image_footnote": [], + "bbox": [ + 596, + 146, + 658, + 191 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/caea1c5ace2af1cc8865fb1579beb42d05d035c3e52a68bc34571a898ccf0aa8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 660, + 146, + 722, + 191 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/bcefd55b40ec584b502c25a521b7e59b3ad09c1cae25cd6d9b6bbc1f27a3f4e5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 725, + 146, + 784, + 191 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/cebfb94fb8bf20e7cb172033d3dd08cfd276237e3f938cddb50a59fe9f9b8fcb.jpg", + "image_caption": [ + "$U + 1F31A$ from Noto $(\\mathrm{N} = 16)$ ; A is ours" + ], + "image_footnote": [], + "bbox": [ + 217, + 207, + 277, + 253 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/fa11a3a6768dc361cea0231ae016dc5f5f60cc69a37d4a77bd7526a16e6ce6a4.jpg", + "image_caption": [ + "Fig. 12: Examples from our questionnaire. Participants are asked to choose the better one for each pair of outputs." + ], + "image_footnote": [], + "bbox": [ + 279, + 207, + 339, + 253 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/78a901925005728c8111e070baee41bcc8463775406c219abecc7b9b269f0c42.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 344, + 207, + 403, + 253 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/3da92b7fc92ed0453ec77b4b09121fe4756f0b56392c0abb8cfb19be793a2a88.jpg", + "image_caption": [ + "$U + 1F613$ from Fluent $(\\mathrm{N} = 32)$ ; A is ours" + ], + "image_footnote": [], + "bbox": [ + 410, + 208, + 465, + 253 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/d81d572b93b621faf1286427918b3beb56fd865e2e51fe083cb22fe83b03f103.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 468, + 208, + 527, + 252 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/68a743928cffc3cb37b0fee6e4258e88ddbf60045d93dfd861ef480df9d2a1a0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 534, + 208, + 591, + 253 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/0be836b6ebf0f0d9e8e7b9c9a97641604c09306c2c283a077d4b000013008ef7.jpg", + "image_caption": [ + "46087 07 from Iconfont (N=256); B is ours" + ], + "image_footnote": [], + "bbox": [ + 596, + 207, + 658, + 253 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/855a590285f2b9350f8761c5bfd3b2ef94d10f4938f3cba4f7b8b28cbaa3893b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 660, + 207, + 722, + 253 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/45dc4111100e898d4d44fde5653e43c2d9e836343ca6cd2327df5d240ba7d13c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 725, + 207, + 784, + 253 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.7 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 337, + 387, + 352 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "To verify the effectiveness of our contributions, we conducted an ablation study, by vectorizing without gradients, without segmentation guidance, or both, as shown in Fig. 13. The similarity between vectorized results and their corresponding inputs is measured by PSNR. Results are tabulated in Tab. 2.", + "bbox": [ + 212, + 361, + 784, + 422 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Gradients For images with rich gradient effects, such as those in Fluent Emoji, the introduction of gradients significantly improves the outcome. For Noto Emoji, where most images do not have gradients, the improvement is less noticeable. Iconfont is an interesting case; although its paths are filled with solid colors, a form of approximate color transition is achieved through multiple paths with progressively changing colors. Introducing gradients also contributes to a noticeable enhancement.", + "bbox": [ + 212, + 436, + 784, + 541 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Segmentation guidance Compared to solely introducing gradients, incorporating segmentation guidance has proven more effective. As illustrated in Tab. 2, a combination of gradients and segmentation yields superior results on both Noto and Fluent datasets. For the Iconfont dataset, the imitated color transition is", + "bbox": [ + 212, + 556, + 785, + 617 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/41e3cc32362c02b3e1f5084e1f7b541cf628154234caa72e27fb7fc3fc3b4db8.jpg", + "image_caption": [ + "Fig. 13: Comparison between four combinations tested in the ablation study. When neither gradients nor segmentation guidance is applied, our method works the same as LIVE (1 $^{\\text{st}}$ row)." + ], + "image_footnote": [], + "bbox": [ + 222, + 640, + 782, + 785 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Segmentation-guided Layer-wise Image Vectorization with Gradient Fills", + "bbox": [ + 246, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "considered a gradient by the segmentation; thus, with fewer paths, the guided variants achieve a more approximate visual effect through large gradient paths. When more paths are added, using smaller, precise solid color paths yields better results, while the guided vectorization remains competitive in terms of accuracy, and is preferred in the user study.", + "bbox": [ + 212, + 146, + 787, + 223 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "4.8 Limitations and Future Work", + "text_level": 1, + "bbox": [ + 215, + 239, + 504, + 253 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/32f07586b986c6a14aa440df4a6d4131642643c453afff1022d2b238873584ce.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 292, + 282, + 356, + 335 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/9c5843823faf84e41ffc9cbb5ead80817a5c02d86dd5fbbc2b4a82e6ba5970a2.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 292, + 338, + 354, + 388 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/82501da6cabfd05a24469809b458386f1772e9f29829091ede3ec37dae16f4c0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 282, + 424, + 330 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/527f92b36da004b16a4fbc7d0110e039ec7d47d56cfc8a4ad6e7735d77b980d5.jpg", + "image_caption": [ + "N=1" + ], + "image_footnote": [], + "bbox": [ + 364, + 347, + 424, + 388 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/533546488afd40e530fe05a2a042d9d89f45c6161ea9c40f84993ded5fdb9fd1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 436, + 282, + 496, + 330 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/baa1e215c3da4677aa28b621c516dd267ebb521f1113cdb3a1327c0cf0604152.jpg", + "image_caption": [ + "N=2", + "Fig. 14: Intricate contours that our simple initial paths fail to converge to. Extra paths are added to mask the overflowing colors." + ], + "image_footnote": [], + "bbox": [ + 436, + 338, + 495, + 388 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/80d75e55e0ff609700329c8cb544346771599352603365ddb2028f024300e070.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 282, + 568, + 330 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/e2b24f2811cf04a777189474e454efeac8350463483da55e28c9dcde2646ac81.jpg", + "image_caption": [ + "N=4" + ], + "image_footnote": [], + "bbox": [ + 506, + 338, + 566, + 388 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/568d41cb898bdabcb4518222b82af754ecf767b26675c2862f82a6ca3f3ea360.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 576, + 282, + 638, + 335 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/79b1b7c39580d21f88f52a73667f71cc78f502affc0c82a152b477b3ec12eca1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 578, + 338, + 637, + 388 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/700c9d2434b6811f066371519a78e3558d85ba1b96562e86b8eeffa7c76ae85d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 648, + 282, + 710, + 337 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/88edb6154ccfe6408c8b1e577c6ed5bc99b95994aa3ebd7dab254252af6c35cb.jpg", + "image_caption": [ + "N=8" + ], + "image_footnote": [], + "bbox": [ + 650, + 338, + 707, + 388 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "While our segmentation guidance effectively captures large-scale features, the initial shape, consisting of a loop formed by four connected Bezier curves, encounters challenges in accurately fitting intricate shapes, as illustrated in Figure 14. Achieving a balance between the number of paths and the number of control points per path needs further deliberation.", + "bbox": [ + 212, + 476, + 787, + 551 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Additionally, our implementation employs radial gradients with two color stops for simplicity, since a linear gradient can be interpreted as a radial gradient with its center positioned outside the path. Our method is not confined to a specific type of gradient. A more comprehensive implementation could involve dynamically determining gradient types and adjusting the number of color stops accordingly.", + "bbox": [ + 212, + 551, + 787, + 643 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 661, + 359, + 678 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we propose a segmentation-guided layer-wise vectorization framework, which synthesizes vector images by progressively adding paths and filling the paths with radial gradients. We design a gradient-aware segmentation method out of traditional algorithms to guide our novel initialization approach and newly designed loss function to address the obstacles in optimizing gradient parameters. We test our method on Noto [9] and Fluent Emoji [21] to show its capability in decomposing raster images into concise layers of gradient-filled paths when the topology is clear, and on Iconfont [13] to demonstrate its effectiveness in case of complex inputs. Our method shows superior performance compared to previous work.", + "bbox": [ + 212, + 689, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "H. Zhou et al.", + "bbox": [ + 271, + 114, + 366, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported by the NSFC under Grant 62072271.", + "bbox": [ + 215, + 176, + 666, + 191 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 214, + 321, + 229 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Baksteen, S.D., Hettinga, G.J., Echevarria, J., Kosinka, J.: Mesh colours for gradient meshes. STAG: Smart Tools and Applications in Graphics (2021)", + "2. Beucher, S., Meyer, F.: The morphological approach to segmentation: the watershed transformation. Mathematical morphology in image processing 34(1993), 49 (1993)", + "3. Carlier, A., Danelljan, M., Alahi, A., Timofte, R.: Deepsvg: A hierarchical generative network for vector graphics animation. Advances in Neural Information Processing Systems 33, 16351-16361 (2020)", + "4. Comaniciu, D., Meer, P.: Mean shift: A robust approach toward feature space analysis. IEEE Transactions on pattern analysis and machine intelligence 24(5), 603-619 (2002)", + "5. Du, Z.J., Kang, L.F., Tan, J., Gingold, Y., Xu, K.: Image vectorization and editing via linear gradient layer decomposition. ACM Transactions on Graphics (TOG) 42(4), 1-13 (2023)", + "6. Egiazarian, V., Voynov, O., Artemov, A., Volkhonskiy, D., Safin, A., Taktasheva, M., Zorin, D., Burnaev, E.: Deep vectorization of technical drawings. In: European conference on computer vision. pp. 582-598. Springer (2020)", + "7. Favreau, J.D., Lafarge, F., Bousseau, A.: Photo2clipart: Image abstraction and vectorization using layered linear gradients. ACM Transactions on Graphics (TOG) 36(6), 1-11 (2017)", + "8. Frans, K., Soros, L., Witkowski, O.: Clipdraw: Exploring text-to-drawing synthesis through language-image encoders. Advances in Neural Information Processing Systems 35, 5207-5218 (2022)", + "9. Noto emoji, https://github.com/googlefonts/sono-emoji Accessed: 2023-09-19", + "0. Ha, D., Eck, D.: A neural representation of sketch drawings. In: International Conference on Learning Representations (2018)", + "1. Haralick, R.M., Sternberg, S.R., Zhuang, X.: Image analysis using mathematical morphology. IEEE transactions on pattern analysis and machine intelligence (4), 532-550 (1987)", + "2. Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural Computation 9(8), 1735-1780 (1997)", + "3. Vector illustrations library, https://www iconfont.cn/illustrations/index Accessed: 2023-11-02", + "4. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. In: International Conference on Learning Representations (2015)", + "5. Kingma, D.P., Welling, M.: Auto-Encoding Variational Bayes. In: International Conference on Learning Representations (2014)", + "6. Li, T.M., Lukáč, M., Gharbi, M., Ragan-Kelley, J.: Differentiable vector graphics rasterization for editing and learning. ACM Transactions on Graphics (TOG) 39(6), 1-15 (2020)", + "7. Liao, Z., Hoppe, H., Forsyth, D., Yu, Y.: A subdivision-based representation for vector image editing. IEEE Transactions on Visualization and Computer Graphics 18(11), 1858-1867 (2012)" + ], + "bbox": [ + 225, + 244, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Segmentation-guided Layer-wise Image Vectorization with Gradient Fills", + "bbox": [ + 246, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Liu, Y.T., Zhang, Z., Guo, Y.C., Fisher, M., Wang, Z., Zhang, S.H.: Dualvector: Unsupervised vector font synthesis with dual-part representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14193-14202 (2023)", + "19. Lopes, R.G., Ha, D., Eck, D., Shlens, J.: A Learned Representation for Scalable Vector Graphics. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7929-7938. IEEE, Seoul, Korea (South) (Oct 2019)", + "20. Ma, X., Zhou, Y., Xu, X., Sun, B., Filev, V., Orlov, N., Fu, Y., Shi, H.: Towards layer-wise image vectorization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16314-16323 (2022)", + "21. Fluent emoji, https://github.com/microsoft/fluentui-emoji Accessed: 2023-10-27", + "22. Orzan, A., Bousseau, A., Winnemöller, H., Barla, P., Thollot, J., Salesin, D.: Diffusion curves: a vector representation for smooth-shaded images. ACM Transactions on Graphics (TOG) 27(3), 1-8 (2008)", + "23. Otsu, N.: A threshold selection method from gray-level histograms. IEEE transactions on systems, man, and cybernetics 9(1), 62-66 (1979)", + "24. Reddy, P., Gharbi, M., Lukac, M., Mitra, N.J.: Im2vec: Synthesizing vector graphics without vector supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7342-7351 (2021)", + "25. Reddy, P., Zhang, Z., Wang, Z., Fisher, M., Jin, H., Mitra, N.: A multi-implicit neural representation for fonts. Advances in Neural Information Processing Systems 34, 12637-12647 (2021)", + "26. Richardt, C., Lopez-Moreno, J., Bousseau, A., Agrawala, M., Drettakis, G.: Vectorising bitmaps into semi-transparent gradient layers. Computer Graphics Forum (Proceedings of EGSR) 33(4), 11-19 (July 2014)", + "27. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)", + "28. Su, H., Liu, X., Niu, J., Cui, J., Wan, J., Wu, X., Wang, N.: Marvel: Raster gray-level manga vectorization via primitive-wise deep reinforcement learning. IEEE Transactions on Circuits and Systems for Video Technology (2023)", + "29. Sun, J., Liang, L., Wen, F., Shum, H.Y.: Image vectorization using optimized gradient meshes. ACM Transactions on Graphics (TOG) 26(3), 11-es (2007)", + "30. Tian, X., Günther, T.: A survey of smooth vector graphics: Recent advances in representation, creation, rasterization and image vectorization. IEEE Transactions on Visualization and Computer Graphics (2022)", + "31. Van der Walt, S., Schonberger, J.L., Nunez-Iglesias, J., Boulogne, F., Warner, J.D., Yager, N., Gouillart, E., Yu, T.: scikit-image: image processing in python. PeerJ 2, e453 (2014)", + "32. Wang, Y., Lian, Z.: Deepvecfont: Synthesizing high-quality vector fonts via dual-modality learning. ACM Transactions on Graphics (TOG) 40(6), 1-15 (2021)", + "33. Xia, T., Liao, B., Yu, Y.: Patch-based image vectorization with automatic curvilinear feature alignment. ACM Transactions on Graphics (TOG) 28(5), 1-10 (2009)", + "34. Xie, G., Sun, X., Tong, X., Nowrouzezahrai, D.: Hierarchical diffusion curves for accurate automatic image vectorization. ACM Transactions on Graphics (TOG) 33(6), 1-11 (2014)", + "35. Yang, M., Chao, H., Zhang, C., Guo, J., Yuan, L., Sun, J.: Effective clipart image vectorization through direct optimization of bezigons. IEEE transactions on visualization and computer graphics 22(2), 1063-1075 (2015)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "H. Zhou et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 15 + }, + { + "type": "ref_text", + "text": "36. Zhu, H., Cao, J., Xiao, Y., Chen, Z., Zhong, Z., Zhang, Y.J.: Tcb-spline-based image vectorization. ACM Transactions on Graphics (TOG) 41(3), 1-17 (2022)", + "bbox": [ + 217, + 146, + 784, + 176 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Segmentation-guided Layer-wise Image Vectorization with Gradient Fills", + "bbox": [ + 246, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Segmentation-guided Layer-wise Image Vectorization with Gradient Fills/01f8c0e2-d236-4d33-947c-2747a133c0bb_model.json b/2024/Segmentation-guided Layer-wise Image Vectorization with Gradient Fills/01f8c0e2-d236-4d33-947c-2747a133c0bb_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7e39d326523682c23750fc667b83196595375f5a --- /dev/null +++ b/2024/Segmentation-guided Layer-wise Image Vectorization with Gradient Fills/01f8c0e2-d236-4d33-947c-2747a133c0bb_model.json @@ -0,0 +1,2973 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.279, + 0.141, + 0.726, + 0.187 + ], + "angle": 0, + "content": "Segmentation-guided Layer-wise Image Vectorization with Gradient Fills" + }, + { + "type": "text", + "bbox": [ + 0.321, + 0.212, + 0.682, + 0.229 + ], + "angle": 0, + "content": "Hengyu Zhou, Hui Zhang, and Bin Wang" + }, + { + "type": "text", + "bbox": [ + 0.242, + 0.24, + 0.761, + 0.269 + ], + "angle": 0, + "content": "School of Software, Tsinghua University, P. R. China \nzhouhy22@mails.tsinghua.edu.cn, {huizhang, wangbins}@tsinghua.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.308, + 0.744, + 0.518 + ], + "angle": 0, + "content": "Abstract. The widespread use of vector graphics creates a significant demand for vectorization methods. While recent learning-based techniques have shown their capability to create vector images of clear topology, filling these primitives with gradients remains a challenge. In this paper, we propose a segmentation-guided vectorization framework to convert raster images into concise vector graphics with radial gradient fills. With the guidance of an embedded gradient-aware segmentation subroutine, our approach progressively appends gradient-filled Bézier paths to the output, where primitive parameters are initiated with our newly designed initialization technique and are optimized to minimize our novel loss function. We build our method on a differentiable renderer with traditional segmentation algorithms to develop it as a model-free tool for raster-to-vector conversion. It is tested on various inputs to demonstrate its feasibility, independent of datasets, to synthesize vector graphics with improved visual quality and layer-wise topology compared to prior work." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.53, + 0.719, + 0.545 + ], + "angle": 0, + "content": "Keywords: Vectorization \\(\\cdot\\) Segmentation \\(\\cdot\\) Differentiable rendering" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.571, + 0.377, + 0.587 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.603, + 0.788, + 0.695 + ], + "angle": 0, + "content": "Vector graphics offer great flexibility in digital design as they can be easily edited and arbitrarily scaled. Vectorization, the procedure of converting raster images to vector ones, serves as a second way to creating vector graphics other than designing from scratch where extensive artistic skills are required, and also as a bridge between the rapidly developing raster image generation [27] and the relatively less studied vector generation [3,8,10,24,32]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.695, + 0.788, + 0.77 + ], + "angle": 0, + "content": "Vectorization has been explored for decades with various representations proposed to divide images into non-overlapping regions [17, 29, 33, 35, 36]. Despite their capability of generating vivid vectorization of realistic images, the complex primitives and lack of hierarchy make the vector output less intuitive for manipulation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.771, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Recent learning-based methods show their potential in preserving image hierarchy, where a vector image is often considered as a list of primitives and their order indicates the structure of the input. Many deep learning approaches" + }, + { + "type": "page_footnote", + "bbox": [ + 0.216, + 0.824, + 0.393, + 0.841 + ], + "angle": 0, + "content": "\\(\\boxtimes\\) Corresponding authors." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.127 + ], + "angle": 0, + "content": "H. Zhou et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.298 + ], + "angle": 0, + "content": "have been proposed to vectorize simple inputs [24] or images of a specific field of interest [6, 18, 25, 28], but their dependency on models limits them to a particular domain and are not trivially generalizable. LIVE [20], in contrast, utilizes DiffVG [16], a differentiable renderer, to present a model-free vectorization framework. It progressively translates a raster image into an SVG in a layer-wise hierarchy, through which the topology of the input is preserved within the order of geometric primitives. However, its lack of support for gradients results in excessive primitives being added in case of images with rich gradient effects, and adding such support is not as simple as replacing RGBA colors with gradient parameters, as elaborated in Sec. 3.4." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.299, + 0.788, + 0.465 + ], + "angle": 0, + "content": "In this paper, we propose a novel segmentation-guided vectorization framework that extends the capability of LIVE to support radial gradients. The additional parameters of a radial gradient pose an increased challenge to optimization, where an effective method to determine whether a pixel contributes to a path's gradient fill is necessary. The key insight behind our idea is the similarity between finding contributing pixels and segmentation tasks, based on which we designed a segmentation-guided initialization procedure to progressively append new shapes to the vector output, and a novel loss function to optimize their geometric and gradient parameters. We evaluated our framework on several datasets with quantitative metrics and a user study, to show its effectiveness and superior performance compared to previous work." + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.465, + 0.478, + 0.478 + ], + "angle": 0, + "content": "To summarize our contributions:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.489, + 0.784, + 0.517 + ], + "angle": 0, + "content": "- We introduce a segmentation-guided vectorization framework to create vector graphics automatically with layer-wise hierarchy and radial gradients." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.519, + 0.784, + 0.547 + ], + "angle": 0, + "content": "- We propose a gradient-aware segmentation method to evaluate the pixel-wise contribution to the geometric and gradient parameters of a path." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.549, + 0.784, + 0.577 + ], + "angle": 0, + "content": "- We take the segmentation as guidance for our new initialization technique and as a part of our novel segmentation-guided loss." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.489, + 0.784, + 0.577 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.602, + 0.388, + 0.618 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.634, + 0.429, + 0.649 + ], + "angle": 0, + "content": "2.1 Image Vectorization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Most traditional vectorization methods aim to create vector images of high fidelity with different representations, which could be roughly categorized into mesh-based ones and curve-based ones. The former representations divide the input image into non-overlapping 2D patches across which colors are interpolated [30]. Shapes of patches include triangular [17,33,36], rectangular [1,29] or irregular ones such as bezigons [35]. The different selection of mesh shapes determines how patches are organized and how colors are interpolated within patches. The curved-based representations decompose the image into curves. Diffusion curves, for instance, use curves as geometric primitives with colors defined on sides of the curves [22,34]. While these methods can yield near-photo-realistic results, they may fall short when it comes to ease of editing due to the complex primitives and loss of topological information, compared to a plain list of" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.248, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Segmentation-guided Layer-wise Image Vectorization with Gradient Fills" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.284 + ], + "angle": 0, + "content": "simple primitives like how an SVG organizes a vector image. Recent learning-based vectorization work mainly takes the list-of-primitive approach, where the ordered primitives are seen as a sequence of drawing operations, and sequential prediction methods are applied to synthesize vector graphics. Models are used including recurrent neural networks [10,24], transformers [3], and are often combined with variational autoencoders [10,15,19,24]. DiffVG [16] as a differentiable renderer fills the gap between raster and vector graphics, with which loss functions on raster images could be used directly to optimize the vector images. A vectorization method optimizing all paths at once is also proposed in the work." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.304, + 0.396, + 0.319 + ], + "angle": 0, + "content": "2.2 Image Topology" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.327, + 0.788, + 0.508 + ], + "angle": 0, + "content": "When synthesizing vector graphics, both similarity to the original image and correct topology are important for the ease of subsequent human manipulation. A related problem is layer decomposition, where images are decomposed into semi-transparent layers. With these layers being vector paths, such a method could serve as a layered vectorization approach. An interactive method [26] is among the first proposed to convert bitmaps into layered vector graphics. Photo2ClipArt [7] and other similar work [5] replace the heavy manual interaction with a user-provided segmentation input, from which the decomposition could be automatically determined. However, these methods require a concise segmentation for efficient and effective vectorization into a relatively small number of paths, where automatic segmentation algorithms struggle to serve the purpose [7]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.509, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Many learning-based methods formulate vector synthesis as sequential prediction problems, through which the order of primitives is naturally preserved. Some methods predict the primitive parameters directly. For example, Egiazarian et al. [6] use a transformer-based network to directly generate the vectorization of technical line drawings; MARVEL [28] applies deep reinforcement learning to predict strokes in black-and-white comics. Meanwhile, some approaches use variational autoencoders (VAE) [15] to create vector graphics. SketchRNN [10] is the first to introduce a Long Short-Term Memory (LSTM) [12] based VAE to the representation of vector sketches. The later DeepSVG [3] proposed a transformer-based VAE network that also encodes a vector image into a latent code, which could be decoded to reconstruct the vector paths. SVG-VAE [19] is the first to use raster images as the input to its encoder, thus it could be used to vectorize images. Im2Vec [24] also takes raster images as its input but drops the need for vector supervision with the help of DiffVG [16]. While these methods show great potential in creating topology-aware vector graphics, they rely on specific training datasets and may overlook the fine details in a complex input. The most similar work to ours is LIVE [20], in which a progressive framework is proposed to convert raster images into layered vector paths utilizing DiffVG for optimization, but its missing support for color gradients is not trivially achievable by simply optimizing gradient parameters. Our novel segmentation-guided method takes a step further to handle gradient effects properly, at the same level of simplicity as LIVE, requiring neither additional user input nor deep models." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.367, + 0.127 + ], + "angle": 0, + "content": "H. Zhou et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.331, + 0.16 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.182, + 0.411, + 0.196 + ], + "angle": 0, + "content": "3.1 Method Overview" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.237, + 0.787, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.327, + 0.785, + 0.355 + ], + "angle": 0, + "content": "Fig. 1: Overview of our framework. \\( U + 1F61A \\) from Noto Emoji [9] is used for demonstration." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.39, + 0.784, + 0.435 + ], + "angle": 0, + "content": "We present a new framework to generate SVG figures from rasterized images, with support for radial gradient fills via a segmentation-guided initialization and optimization process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.437, + 0.788, + 0.604 + ], + "angle": 0, + "content": "Our framework works progressively, where at each epoch, single or multiple Bézier paths are added and optimized. Fig. 1 provides an overview of our method. At the beginning of each epoch \\( i \\), we calculate the difference between the input raster image and the output \\( I_{i-1} \\) from the previous epoch. We segment it with our gradient-aware segmentation (Sec. 3.2), from which \\( n_i \\) segmented regions are selected for initialization of new paths (Sec. 3.3). All added paths, including those added in previous epochs, are optimized to minimize the vectorization loss (Sec. 3.4). The parameters to be optimized consist of geometric parameters including positions of curve control points and gradient parameters including their center, radius, and color stops. \\( n_i \\), or the number of paths to be appended at \\( i \\)-th epoch is specified by the user along with the number of iteration epochs." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.63, + 0.506, + 0.646 + ], + "angle": 0, + "content": "3.2 Gradient-aware Segmentation" + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.687, + 0.363, + 0.754 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.294, + 0.757, + 0.352, + 0.769 + ], + "angle": 0, + "content": "(a) Input" + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.688, + 0.452, + 0.753 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.383, + 0.757, + 0.442, + 0.769 + ], + "angle": 0, + "content": "(b) LIVE" + }, + { + "type": "image", + "bbox": [ + 0.464, + 0.688, + 0.541, + 0.753 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.466, + 0.757, + 0.539, + 0.768 + ], + "angle": 0, + "content": "(c) LIVE-30" + }, + { + "type": "image", + "bbox": [ + 0.552, + 0.688, + 0.631, + 0.754 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.55, + 0.757, + 0.635, + 0.768 + ], + "angle": 0, + "content": "(d) Mean shift" + }, + { + "type": "image", + "bbox": [ + 0.643, + 0.688, + 0.721, + 0.754 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.656, + 0.757, + 0.708, + 0.768 + ], + "angle": 0, + "content": "(e) Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.836 + ], + "angle": 0, + "content": "Fig. 2: Comparisons on segmentation methods. (a) is the raster input. (b) is by LIVE under default settings clustering colors into 200 bins. (c) is also by LIVE but with the number of bins set to 30. (d) is segmented using the Mean-shift [4] algorithm. (e) is by our method." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.248, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Segmentation-guided Layer-wise Image Vectorization with Gradient Fills" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.145, + 0.77, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.285, + 0.341, + 0.719, + 0.355 + ], + "angle": 0, + "content": "Fig. 3: Step-by-step outputs of our gradient-aware segmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.382, + 0.785, + 0.442 + ], + "angle": 0, + "content": "Quality of path initialization is the key to effective vectorization, hence we present a gradient-aware segmentation method to better estimate the area that a path is likely to span over. Its result is used in the subsequent operations including initialization and optimization (Sec. 3.3, Sec. 3.4)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.442, + 0.785, + 0.578 + ], + "angle": 0, + "content": "When paths are filled with a solid color, a connected component of similar or identical colors has a good chance of being covered by one path. LIVE [20] designed a component-wise initialization method, where pixels are clustered into buckets based on their l2-length over RGB channels, and connected pixels in the same bucket are considered one component. As gradients are considered, the clustering algorithm used by LIVE may result in an excessive segmentation, as shown by Fig. 2. Other clustering methods including Mean-shift [4] also suffer from over-segmentation for the same reason that colors within a path may vary more than colors between paths." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.579, + 0.785, + 0.676 + ], + "angle": 0, + "content": "With the limitations of previous methods, our approach is designed to detect edges of gradients. Colors should derive smoothly inside a region filled with the same gradient fill and are likely to change abruptly at its boundary. Thus, we calculate the secondary spatial gradient with a discrete Laplacian filter \\( L = \\left[ \\begin{array}{lll}1 & 1 & 1\\\\ 1 & -8 & 1\\\\ 1 & 1 & 1 \\end{array} \\right] \\) to identify such rapid changes at gradient boundaries. Our segmentation method features the following steps:" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.683, + 0.787, + 0.713 + ], + "angle": 0, + "content": "- The cross-correlation \\( S_0 \\in \\mathbb{R}_{w \\times h \\times 3} \\) between the difference to the input and the 2D Laplacian filter is calculated as" + }, + { + "type": "equation", + "bbox": [ + 0.382, + 0.721, + 0.646, + 0.742 + ], + "angle": 0, + "content": "\\[\nS _ {0} = \\operatorname {c o r r e l a t e} ((I - \\hat {I}) \\mathbf {1} _ {| | \\hat {I} - I | | _ {2} > \\epsilon}, L),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.75, + 0.784, + 0.779 + ], + "angle": 0, + "content": "where \\(\\hat{I}\\) is the target and \\(I\\) is the synthesized image. Pixels with an error below \\(\\epsilon = 0.1\\) are excluded." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.781, + 0.784, + 0.811 + ], + "angle": 0, + "content": "- \\( S_0 \\) is then summed over its RGB channels for a grayscale image \\( S_1 \\in \\mathbb{R}_{w \\times h} \\), where each pixel of \\( S_1 \\) has a value of \\( (S_1)_{ij} = \\sum_{c=1}^{3} |(S_0)_{jc}| \\)." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.811, + 0.784, + 0.84 + ], + "angle": 0, + "content": "- The grayscale image \\( S_{1} \\) is converted to a binary image \\( S_{2} = \\mathbf{1}_{S_{1} > \\mathrm{Otsu}(S_{1})} \\) with the threshold determined via Otsu's method [23]." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.781, + 0.784, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.128 + ], + "angle": 0, + "content": "H. Zhou et al." + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.145, + 0.761, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.302, + 0.789, + 0.373 + ], + "angle": 0, + "content": "Fig. 4: With paths being progressively added, the overall difference between the current output and the raster input decreases, thus we apply Otsu's method [23] to automatically determine the threshold for binarization. In this example, the red cheeks get segmented after more significant differences including eyes and brows are fitted, thanks to the dynamic threshold. The differences are normalized and mapped to grayscale." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.404, + 0.789, + 0.436 + ], + "angle": 0, + "content": "- Morphological closing [11] and the watershed algorithm [2] are then applied to \\( S_{2} \\) for the final segmentation \\( S \\in \\mathbb{N}_{h \\times w} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.451, + 0.79, + 0.543 + ], + "angle": 0, + "content": "In Fig. 3 we showcase some intermediate results after each step. Since we are segmenting the difference between the output and the target, the already fitted pixels are ignored and the automatically determined threshold via Otsu's method decreases in response to the descending overall difference, as shown in Fig. 4. Compared to a fixed threshold, our method makes fewer assumptions about the input and avoids hyperparameters." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.57, + 0.547, + 0.586 + ], + "angle": 0, + "content": "3.3 Segmentation-guided Initialization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.6, + 0.789, + 0.661 + ], + "angle": 0, + "content": "Our vectorization method works in a progressive manner, throughout which we add one or more paths at each epoch. These newly added paths are initialized using our segmentation \\( S \\) as the guidance. For each path to be added, we select the segmentation region with the largest accumulated square error calculated as" + }, + { + "type": "equation", + "bbox": [ + 0.422, + 0.675, + 0.787, + 0.711 + ], + "angle": 0, + "content": "\\[\nw _ {i} = \\sum_ {p \\in \\tilde {S} [ i ]} | | I _ {p} - \\hat {I} _ {p} | | ^ {2} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.842 + ], + "angle": 0, + "content": "where \\( p \\) is iterated over all pixels in \\( i \\)-th region. This approach prioritizes larger regions to encourage a hierarchical initialization order and prevents regions that are almost properly filled from being chosen. A circle path of four cubic Bezier curves [20] is added at the selected region's center of mass \\( p_m \\). We fill the path with a radial gradient, which centers at \\( p_m \\), with a diameter equal to the geometric mean of the width and the height of the region's bounding box clipped to [0.2, 1.0]. The two stop colors at offsets \\( 0\\% \\) and \\( 100\\% \\) are both initialized to the color of the input image \\( I \\) at \\( p_m \\)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.248, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Segmentation-guided Layer-wise Image Vectorization with Gradient Fills" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.252, + 0.147, + 0.756, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.362, + 0.788, + 0.432 + ], + "angle": 0, + "content": "Fig. 5: Optimizing course for our method in comparison with UDF loss from LIVE. With our proposed segmentation-guided weight (SG Weight), the gradient fill is optimized to minimize the color error not only on the contour as UDF weight focuses on but also on colors inside the path, while excluding pixels occluded by eyes, brows, and the mouth." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.463, + 0.657, + 0.479 + ], + "angle": 0, + "content": "3.4 Loss Function with Segmentation-guided weight" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.492, + 0.788, + 0.657 + ], + "angle": 0, + "content": "A common approach to measuring the difference between the synthesized image \\(\\hat{I}\\) and the target image \\(I\\) is through mean squared error (MSE). However, simply minimizing MSE results in colors biasing towards the average color inside a path, as suggested by LIVE [20]. LIVE tackles the problem with its designed UDF loss to focus on pixels on the contour, but for radial gradients, correct colors on the contour do not mean correctness inside, as indicated by Fig. 5. When gradients are optimized with UDF loss, colors at the center are not fitted well, since those pixels are ignored by the UDF weight. We draw from the insight by LIVE emphasizing the significance of pixels on edges and extend this concept to include all pixels within a path, except for those occluded by other paths, which is well estimated through our gradient-aware segmentation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.658, + 0.788, + 0.747 + ], + "angle": 0, + "content": "We formulate our segmentation-guided loss with this estimation of unoccluded pixels within a path. For each added path \\( p_i \\), given the set of pixels covered by the path and the set of pixels within the segment from which the path is initialized, we take their intersection and mark these pixels as being focused. The union of all focused pixels forms a set \\( F \\). We define the weight \\( w_{\\mathrm{SG}} \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.367, + 0.747, + 0.785, + 0.788 + ], + "angle": 0, + "content": "\\[\nw _ {\\mathrm {S G}} (i) = \\left\\{ \\begin{array}{l l} \\max \\left(d _ {i} ^ {\\prime}, \\alpha_ {s}\\right), & i \\in F \\\\ d _ {i} ^ {\\prime} \\left(1 - \\alpha_ {s}\\right), & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.843 + ], + "angle": 0, + "content": "where \\( i \\) is an index of pixel and \\( d_{i}^{\\prime} \\) is the UDF weight from [20]. The UDF weight is introduced to give the pixels near the path contour a higher weight, as depicted in Fig. 5. \\( \\alpha_{s} \\) balances between the UDF weight \\( d_{i}^{\\prime} \\) and our segmentation weight," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.128 + ], + "angle": 0, + "content": "H. Zhou et al." + }, + { + "type": "image", + "bbox": [ + 0.357, + 0.145, + 0.508, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.144, + 0.644, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.351, + 0.231, + 0.653, + 0.246 + ], + "angle": 0, + "content": "Fig. 6: Bézier paths with four control points" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.273, + 0.558, + 0.289 + ], + "angle": 0, + "content": "and is set to 0.6 empirically. \\(d_i^\\prime\\) is defined with:" + }, + { + "type": "equation", + "bbox": [ + 0.398, + 0.296, + 0.786, + 0.336 + ], + "angle": 0, + "content": "\\[\nd _ {i} ^ {\\prime} = \\frac {\\operatorname {R e L U} (\\tau - | d _ {i} |)}{\\sum_ {j = 1} ^ {w \\times h} \\operatorname {R e L U} (\\tau - | d _ {j} |)}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.343, + 0.784, + 0.372 + ], + "angle": 0, + "content": "where \\( d_{i} \\) is the distance from the pixel \\( i \\) to the nearest path contour, which is then normalized and saturated with threshold \\( \\tau = 10 \\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.373, + 0.785, + 0.403 + ], + "angle": 0, + "content": "We re-weight the pixel-wise color error with our proposed weight \\( w_{\\mathrm{SG}} \\) for the final segmentation-guided loss:" + }, + { + "type": "equation", + "bbox": [ + 0.368, + 0.413, + 0.786, + 0.455 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {S G}} = \\frac {1}{3} \\sum_ {i = 1} ^ {w \\times h} w _ {\\mathrm {S G}} (i) \\sum_ {c = 1} ^ {3} \\left(I _ {c, i} - \\hat {I} _ {c, i}\\right) ^ {2}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.462, + 0.784, + 0.491 + ], + "angle": 0, + "content": "where \\( i \\) is iterated over all pixels and squared color differences are averaged over the RGB channels." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.492, + 0.785, + 0.553 + ], + "angle": 0, + "content": "We also introduce Xing loss [20] to help relieve self-intersection. Given a cubic Bézier curve with four control points \\(A, B, C\\), and \\(D\\), self-intersection is more likely to occur when the angle between \\(AB\\) and \\(CD\\) is greater than \\(180^{\\circ}\\), as shown in Fig. 6. The Xing loss \\(\\mathcal{L}_{\\mathrm{Xing}}\\) is defined to penalize angles over \\(180^{\\circ}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.32, + 0.563, + 0.786, + 0.605 + ], + "angle": 0, + "content": "\\[\nD _ {1} = \\left\\{ \\begin{array}{l} 1, \\boldsymbol {A B} \\times \\boldsymbol {B C} > 0 \\\\ 0, \\text {o t h e r w i s e} \\end{array} , D _ {2} = \\frac {\\boldsymbol {A B} \\times \\boldsymbol {C D}}{\\| \\boldsymbol {A B} \\| \\| \\boldsymbol {C D} \\|} \\right. \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.321, + 0.617, + 0.786, + 0.634 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {X i n g}} = D _ {1} \\left(\\operatorname {R e L U} (- D _ {2})\\right) + (1 - D _ {1}) \\left(\\operatorname {R e L U} \\left(D _ {2}\\right)\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.639, + 0.589, + 0.654 + ], + "angle": 0, + "content": "Our optimization objective is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.431, + 0.664, + 0.786, + 0.681 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {S G}} + \\lambda \\mathcal {L} _ {\\text {X i n g}}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.69, + 0.462, + 0.705 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is set to 0.05 empirically." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.727, + 0.376, + 0.744 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.756, + 0.457, + 0.772 + ], + "angle": 0, + "content": "4.1 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.78, + 0.784, + 0.843 + ], + "angle": 0, + "content": "We implement our framework with DiffVG [16] renderer based on PyTorch. Adam optimizer [14] is used with a learning rate of \\( 10^{-2} \\) and 1 for gradient parameters and path points respectively. We use scikit-image [31] for its implementation of morphological operations and the watershed algorithm [2]." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.248, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Segmentation-guided Layer-wise Image Vectorization with Gradient Fills" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image_caption", + "bbox": [ + 0.278, + 0.155, + 0.294, + 0.184 + ], + "angle": 0, + "content": "Noto" + }, + { + "type": "image", + "bbox": [ + 0.301, + 0.148, + 0.73, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.281, + 0.206, + 0.297, + 0.24 + ], + "angle": 0, + "content": "Fluent" + }, + { + "type": "image", + "bbox": [ + 0.302, + 0.202, + 0.726, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.282, + 0.254, + 0.297, + 0.295 + ], + "angle": 0, + "content": "onont" + }, + { + "type": "image", + "bbox": [ + 0.3, + 0.252, + 0.728, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.306, + 0.312, + 0.696, + 0.325 + ], + "angle": 0, + "content": "Fig. 7: Exemplars from our chosen datasets for evaluation" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.357, + 0.334, + 0.37 + ], + "angle": 0, + "content": "4.2 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.382, + 0.785, + 0.473 + ], + "angle": 0, + "content": "Noto Emoji We randomly select 256 emojis from Noto Emoji [9](Unicode 15.0) varying in colors and genres. In contrast to previous work [20,24], we take a more recent version of the dataset, consisting of resigned emojis filled with gradients instead of single colors. Emojis from this dataset consist of a relatively small number of paths with a clear hierarchy, thus we evaluate on this dataset to show our method's capability to decompose images into layers with gradients." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.489, + 0.785, + 0.551 + ], + "angle": 0, + "content": "Fluent Emojis We select an additional 256 images at random from Microsoft's Fluent Emojis [21] collection. This dataset features colorful emojis with rich gradient details. We test on this dataset to elaborate on our method's ability to match up with target images using fewer paths compared to previous work [16, 20]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.565, + 0.785, + 0.64 + ], + "angle": 0, + "content": "Iconfont A set of 128 vector arts are fetched from the online Iconfont Illustration Library [13]. Images in the dataset contain an average of 1020 paths with great details but no gradients. We use these images to exhibit that our segmentation-guided approach is also suitable for vectorizing complex inputs, and performs on par with or better than previous work even without gradients." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.641, + 0.785, + 0.671 + ], + "angle": 0, + "content": "It is worth noticing that our framework is model-free, thus datasets mentioned above are for sole evaluation purposes." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.694, + 0.458, + 0.709 + ], + "angle": 0, + "content": "4.3 Qualitative Comparison" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.785, + 0.795 + ], + "angle": 0, + "content": "With our initial goal of supporting gradient fills in a topology-preserving vectorization approach, we conducted a comparative analysis of the visual quality of our reconstructed vector graphics against LIVE [20]. Our method achieves superior visual quality in vectorization while utilizing the same number of paths as LIVE, as evidenced in Fig. 8." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.785, + 0.84 + ], + "angle": 0, + "content": "As depicted in Fig. 9, when presented with input containing gradients, LIVE struggles to accurately vectorize gradient-filled facial features, despite correctly capturing other facial parts. In contrast, our approach accurately reconstructs" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.367, + 0.127 + ], + "angle": 0, + "content": "H. Zhou et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.789, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.356, + 0.34, + 0.647, + 0.355 + ], + "angle": 0, + "content": "Fig. 8: Qualitative comparisons with LIVE" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.388, + 0.788, + 0.434 + ], + "angle": 0, + "content": "the input using concise paths. Additionally, attempts to optimize gradient parameters without guidance, as seen in the 'w/o guidance' rows, result in degraded outcomes." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.466, + 0.471, + 0.481 + ], + "angle": 0, + "content": "4.4 Quantitative Comparison" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.5, + 0.788, + 0.592 + ], + "angle": 0, + "content": "We further adopt quantitative metrics in a quantitative comparison with LIVE. We calculated PSNR to measure the difference between the rendered vectorized outputs and their corresponding raster inputs. For Noto and Fluent Emoji datasets, we add \\(\\mathrm{clamp}(2^{i - 2},1,32)\\) paths at \\(i\\)-th initialization to reach a total of 256 paths. For the Iconfont dataset, 512 paths are used for each image with \\(\\mathrm{clamp}(2^{i - 2},1,64)\\) paths being added at \\(i\\)-th epoch, for its higher complexity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.593, + 0.788, + 0.669 + ], + "angle": 0, + "content": "As reported in Fig. 10, our method achieves generally faster convergence than LIVE, especially when a small number of paths are added. Our segmentation-guided framework can capture large-scale gradient features where LIVE tends to emit superfluous paths. With excessive paths being added, both methods yield the same level of quality in terms of PSNR." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.7, + 0.436, + 0.716 + ], + "angle": 0, + "content": "4.5 Layer Decomposition" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The framework we propose captures layer-wise structure during our progressive vectorization. The segmentation-guided initialization prioritizes segments that are more significant in terms of accumulated error, while details with relatively small errors are captured later by our adaptive gradient-aware segmentation, as shown in Fig. 4. Given an input with a clear hierarchy, these progressively added ordered paths resemble the handcrafted vector graphics, creating an easy-to-edit output, as shown in Fig. 11." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.248, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Segmentation-guided Layer-wise Image Vectorization with Gradient Fills" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.145, + 0.789, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.418, + 0.788, + 0.504 + ], + "angle": 0, + "content": "Fig.9: A comparison on number of paths needed to reconstruct major components. 'w/o guidance' refers to gradient fills being added without our proposed segmentation guidance. For the first input (\\(U + 1F61A\\) from Noto Emoji), our method vectorizes all elements with 8 paths, while adding more paths does not deteriorate the output. For the second input (\\(U + 1F605\\) from Fluent Emoji), ours reconstructs the facial parts with 6 paths and achieves a close gradient effect to the input using 16 paths." + }, + { + "type": "image", + "bbox": [ + 0.228, + 0.522, + 0.777, + 0.659 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.673, + 0.788, + 0.716 + ], + "angle": 0, + "content": "Fig. 10: Quantitative comparison between our method and LIVE. Our method achieves significantly lower error when the numbers of paths are small and converge to equal performance with excessive paths being added." + }, + { + "type": "image", + "bbox": [ + 0.283, + 0.728, + 0.366, + 0.794 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.293, + 0.797, + 0.357, + 0.808 + ], + "angle": 0, + "content": "(a) Source" + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.729, + 0.457, + 0.794 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.373, + 0.797, + 0.458, + 0.808 + ], + "angle": 0, + "content": "(b) Vectorized" + }, + { + "type": "image", + "bbox": [ + 0.468, + 0.729, + 0.622, + 0.794 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.482, + 0.797, + 0.613, + 0.809 + ], + "angle": 0, + "content": "(c) Decomposed layers" + }, + { + "type": "image", + "bbox": [ + 0.637, + 0.729, + 0.721, + 0.794 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.648, + 0.797, + 0.712, + 0.809 + ], + "angle": 0, + "content": "(d) Edited" + }, + { + "type": "image_caption", + "bbox": [ + 0.356, + 0.82, + 0.646, + 0.835 + ], + "angle": 0, + "content": "Fig. 11: Decomposed layers and recoloring" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.127 + ], + "angle": 0, + "content": "H. Zhou et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.145, + 0.493, + 0.242 + ], + "angle": 0, + "content": "Table 1: Results collected from our user study. The row 'Total' stands for the total number of answers collected for all the images in the dataset vectorized given the number of paths. As shown by the results, users in general prefer the vectorization by our method." + }, + { + "type": "table_caption", + "bbox": [ + 0.309, + 0.267, + 0.402, + 0.28 + ], + "angle": 0, + "content": "(a) Noto Emoji" + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.291, + 0.495, + 0.358 + ], + "angle": 0, + "content": "
#Paths8163264Overall
LIVE39.0%39.0%40.8%43.0%40.4%
Ours61.0%61.0%59.2%57.0%59.6%
Total4494414414541785
" + }, + { + "type": "table_caption", + "bbox": [ + 0.305, + 0.368, + 0.407, + 0.38 + ], + "angle": 0, + "content": "(b) Fluent Emoji" + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.392, + 0.495, + 0.459 + ], + "angle": 0, + "content": "
#Paths8163264Overall
LIVE40.2%38.4%30.8%29.3%34.7%
Ours59.8%61.6%69.2%70.7%65.3%
Total4334454524261756
" + }, + { + "type": "table_caption", + "bbox": [ + 0.32, + 0.469, + 0.391, + 0.481 + ], + "angle": 0, + "content": "(c) Iconfont" + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.493, + 0.495, + 0.56 + ], + "angle": 0, + "content": "
#Paths3264128256Overall
LIVE32.2%42.7%48.5%44.3%42.1%
Ours67.8%57.3%51.5%55.7%57.9%
Total214241227237919
" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.145, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 2: Results of our ablation study in PSNR. 'G' in the header is for 'with gradients' and 'S' is for 'with segmentation guidance'. The first row, with neither gradients nor segmentation, stands for LIVE." + }, + { + "type": "table_caption", + "bbox": [ + 0.603, + 0.239, + 0.696, + 0.252 + ], + "angle": 0, + "content": "(a) Noto Emoji" + }, + { + "type": "table", + "bbox": [ + 0.512, + 0.264, + 0.787, + 0.342 + ], + "angle": 0, + "content": "
G SN=8163264128256
24.9228.2131.4233.8335.8937.16
✓25.3328.6731.6334.1336.1237.34
✓25.1928.5131.8034.1736.0537.29
✓25.6928.9531.9234.2236.3037.46
" + }, + { + "type": "table_caption", + "bbox": [ + 0.598, + 0.353, + 0.701, + 0.365 + ], + "angle": 0, + "content": "(b) Fluent Emoji" + }, + { + "type": "table", + "bbox": [ + 0.512, + 0.377, + 0.787, + 0.457 + ], + "angle": 0, + "content": "
G SN=8163264128256
25.1928.7431.7434.4636.8538.38
✓26.2629.3632.2234.6736.9638.52
26.1529.5932.6435.2837.5438.91
✓26.7330.0732.9535.4337.6138.84
" + }, + { + "type": "table_caption", + "bbox": [ + 0.613, + 0.467, + 0.685, + 0.479 + ], + "angle": 0, + "content": "(c) Iconfont" + }, + { + "type": "table", + "bbox": [ + 0.512, + 0.491, + 0.787, + 0.57 + ], + "angle": 0, + "content": "
G SN=163264128256512
19.2221.6924.3827.2630.2532.36
✓20.0222.0224.2926.8629.4331.44
✓19.4221.9624.7227.6030.6732.72
✓20.3122.3524.7327.2829.7631.74
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.603, + 0.357, + 0.618 + ], + "angle": 0, + "content": "4.6 User Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.641, + 0.788, + 0.763 + ], + "angle": 0, + "content": "We conducted a questionnaire survey to obtain subjective feedback from users. Each participant is presented with 20 questions randomly selected from a total of 2,560 questions, that is, 4 numbers of paths multiplied by 640 images from the three datasets. In each question, participants are shown a raster input and two vectorized versions of it, one using LIVE and the other using our method, with the same number of paths. They are asked to choose the one that looks closer to the original image or, at an equivalent level of similarity, looks more appealing. Figure 12 shows examples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "In our study, 223 people engaged, contributing a total of 4,460 votes. The results, presented in Tab. 1, reveal a clear preference for our approach. In the Fluent Emoji dataset, our method is notably preferred for these images with rich gradients. Despite the absence of gradients in many images from Noto Emoji and Iconfont, our method remains the preferred choice." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.248, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Segmentation-guided Layer-wise Image Vectorization with Gradient Fills" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.147, + 0.278, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.232, + 0.195, + 0.393, + 0.205 + ], + "angle": 0, + "content": "\\(U + 1F31A\\) from Noto \\((\\mathrm{N} = 8)\\) B is ours" + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.147, + 0.341, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.345, + 0.147, + 0.404, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.147, + 0.466, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.418, + 0.195, + 0.587, + 0.205 + ], + "angle": 0, + "content": "\\(U + 1F613\\) from Fluent \\((\\mathrm{N} = 16)\\) B is ours" + }, + { + "type": "image", + "bbox": [ + 0.468, + 0.147, + 0.528, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.535, + 0.147, + 0.593, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.147, + 0.659, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.661, + 0.147, + 0.723, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.726, + 0.147, + 0.785, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.601, + 0.195, + 0.784, + 0.206 + ], + "angle": 0, + "content": "46087_07 from Iconfont (N=128); A is ours" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.208, + 0.278, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.23, + 0.256, + 0.394, + 0.267 + ], + "angle": 0, + "content": "\\(U + 1F31A\\) from Noto \\((\\mathrm{N} = 16)\\) ; A is ours" + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.208, + 0.34, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.345, + 0.208, + 0.404, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.209, + 0.466, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.47, + 0.209, + 0.528, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.535, + 0.209, + 0.592, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.417, + 0.256, + 0.587, + 0.267 + ], + "angle": 0, + "content": "\\(U + 1F613\\) from Fluent \\((\\mathrm{N} = 32)\\) ; A is ours" + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.208, + 0.66, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.661, + 0.208, + 0.723, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.726, + 0.208, + 0.785, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.601, + 0.256, + 0.784, + 0.267 + ], + "angle": 0, + "content": "46087 07 from Iconfont (N=256); B is ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.28, + 0.785, + 0.309 + ], + "angle": 0, + "content": "Fig. 12: Examples from our questionnaire. Participants are asked to choose the better one for each pair of outputs." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.338, + 0.388, + 0.353 + ], + "angle": 0, + "content": "4.7 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.362, + 0.785, + 0.424 + ], + "angle": 0, + "content": "To verify the effectiveness of our contributions, we conducted an ablation study, by vectorizing without gradients, without segmentation guidance, or both, as shown in Fig. 13. The similarity between vectorized results and their corresponding inputs is measured by PSNR. Results are tabulated in Tab. 2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.437, + 0.785, + 0.542 + ], + "angle": 0, + "content": "Gradients For images with rich gradient effects, such as those in Fluent Emoji, the introduction of gradients significantly improves the outcome. For Noto Emoji, where most images do not have gradients, the improvement is less noticeable. Iconfont is an interesting case; although its paths are filled with solid colors, a form of approximate color transition is achieved through multiple paths with progressively changing colors. Introducing gradients also contributes to a noticeable enhancement." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.557, + 0.787, + 0.618 + ], + "angle": 0, + "content": "Segmentation guidance Compared to solely introducing gradients, incorporating segmentation guidance has proven more effective. As illustrated in Tab. 2, a combination of gradients and segmentation yields superior results on both Noto and Fluent datasets. For the Iconfont dataset, the imitated color transition is" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.641, + 0.784, + 0.786 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.794, + 0.785, + 0.836 + ], + "angle": 0, + "content": "Fig. 13: Comparison between four combinations tested in the ablation study. When neither gradients nor segmentation guidance is applied, our method works the same as LIVE (1\\(^{\\text{st}}\\) row)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.367, + 0.128 + ], + "angle": 0, + "content": "H. Zhou et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "considered a gradient by the segmentation; thus, with fewer paths, the guided variants achieve a more approximate visual effect through large gradient paths. When more paths are added, using smaller, precise solid color paths yields better results, while the guided vectorization remains competitive in terms of accuracy, and is preferred in the user study." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.24, + 0.506, + 0.255 + ], + "angle": 0, + "content": "4.8 Limitations and Future Work" + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.283, + 0.357, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.339, + 0.356, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.31, + 0.393, + 0.341, + 0.404 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.284, + 0.425, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.348, + 0.425, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.382, + 0.393, + 0.407, + 0.404 + ], + "angle": 0, + "content": "N=1" + }, + { + "type": "image", + "bbox": [ + 0.437, + 0.284, + 0.497, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.437, + 0.339, + 0.496, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.453, + 0.393, + 0.48, + 0.403 + ], + "angle": 0, + "content": "N=2" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.284, + 0.569, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.339, + 0.567, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.524, + 0.393, + 0.55, + 0.403 + ], + "angle": 0, + "content": "N=4" + }, + { + "type": "image", + "bbox": [ + 0.578, + 0.284, + 0.639, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.579, + 0.339, + 0.638, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.639, + 0.393, + 0.661, + 0.403 + ], + "angle": 0, + "content": "N=8" + }, + { + "type": "image", + "bbox": [ + 0.649, + 0.284, + 0.711, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.651, + 0.339, + 0.709, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.419, + 0.789, + 0.447 + ], + "angle": 0, + "content": "Fig. 14: Intricate contours that our simple initial paths fail to converge to. Extra paths are added to mask the overflowing colors." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.477, + 0.789, + 0.553 + ], + "angle": 0, + "content": "While our segmentation guidance effectively captures large-scale features, the initial shape, consisting of a loop formed by four connected Bezier curves, encounters challenges in accurately fitting intricate shapes, as illustrated in Figure 14. Achieving a balance between the number of paths and the number of control points per path needs further deliberation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.789, + 0.645 + ], + "angle": 0, + "content": "Additionally, our implementation employs radial gradients with two color stops for simplicity, since a linear gradient can be interpreted as a radial gradient with its center positioned outside the path. Our method is not confined to a specific type of gradient. A more comprehensive implementation could involve dynamically determining gradient types and adjusting the number of color stops accordingly." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.662, + 0.36, + 0.679 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.789, + 0.841 + ], + "angle": 0, + "content": "In this paper, we propose a segmentation-guided layer-wise vectorization framework, which synthesizes vector images by progressively adding paths and filling the paths with radial gradients. We design a gradient-aware segmentation method out of traditional algorithms to guide our novel initialization approach and newly designed loss function to address the obstacles in optimizing gradient parameters. We test our method on Noto [9] and Fluent Emoji [21] to show its capability in decomposing raster images into concise layers of gradient-filled paths when the topology is clear, and on Iconfont [13] to demonstrate its effectiveness in case of complex inputs. Our method shows superior performance compared to previous work." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.248, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Segmentation-guided Layer-wise Image Vectorization with Gradient Fills" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.177, + 0.667, + 0.192 + ], + "angle": 0, + "content": "This work was supported by the NSFC under Grant 62072271." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.215, + 0.323, + 0.23 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.246, + 0.785, + 0.274 + ], + "angle": 0, + "content": "1. Baksteen, S.D., Hettinga, G.J., Echevarria, J., Kosinka, J.: Mesh colours for gradient meshes. STAG: Smart Tools and Applications in Graphics (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.275, + 0.785, + 0.315 + ], + "angle": 0, + "content": "2. Beucher, S., Meyer, F.: The morphological approach to segmentation: the watershed transformation. Mathematical morphology in image processing 34(1993), 49 (1993)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.316, + 0.785, + 0.357 + ], + "angle": 0, + "content": "3. Carlier, A., Danelljan, M., Alahi, A., Timofte, R.: Deepsvg: A hierarchical generative network for vector graphics animation. Advances in Neural Information Processing Systems 33, 16351-16361 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.358, + 0.785, + 0.398 + ], + "angle": 0, + "content": "4. Comaniciu, D., Meer, P.: Mean shift: A robust approach toward feature space analysis. IEEE Transactions on pattern analysis and machine intelligence 24(5), 603-619 (2002)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.399, + 0.785, + 0.44 + ], + "angle": 0, + "content": "5. Du, Z.J., Kang, L.F., Tan, J., Gingold, Y., Xu, K.: Image vectorization and editing via linear gradient layer decomposition. ACM Transactions on Graphics (TOG) 42(4), 1-13 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.441, + 0.785, + 0.482 + ], + "angle": 0, + "content": "6. Egiazarian, V., Voynov, O., Artemov, A., Volkhonskiy, D., Safin, A., Taktasheva, M., Zorin, D., Burnaev, E.: Deep vectorization of technical drawings. In: European conference on computer vision. pp. 582-598. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.483, + 0.785, + 0.523 + ], + "angle": 0, + "content": "7. Favreau, J.D., Lafarge, F., Bousseau, A.: Photo2clipart: Image abstraction and vectorization using layered linear gradients. ACM Transactions on Graphics (TOG) 36(6), 1-11 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.524, + 0.785, + 0.564 + ], + "angle": 0, + "content": "8. Frans, K., Soros, L., Witkowski, O.: Clipdraw: Exploring text-to-drawing synthesis through language-image encoders. Advances in Neural Information Processing Systems 35, 5207-5218 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.565, + 0.785, + 0.578 + ], + "angle": 0, + "content": "9. Noto emoji, https://github.com/googlefonts/sono-emoji Accessed: 2023-09-19" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.579, + 0.785, + 0.605 + ], + "angle": 0, + "content": "0. Ha, D., Eck, D.: A neural representation of sketch drawings. In: International Conference on Learning Representations (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.606, + 0.785, + 0.646 + ], + "angle": 0, + "content": "1. Haralick, R.M., Sternberg, S.R., Zhuang, X.: Image analysis using mathematical morphology. IEEE transactions on pattern analysis and machine intelligence (4), 532-550 (1987)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.647, + 0.785, + 0.674 + ], + "angle": 0, + "content": "2. Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural Computation 9(8), 1735-1780 (1997)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.675, + 0.785, + 0.701 + ], + "angle": 0, + "content": "3. Vector illustrations library, https://www iconfont.cn/illustrations/index Accessed: 2023-11-02" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.702, + 0.785, + 0.73 + ], + "angle": 0, + "content": "4. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. In: International Conference on Learning Representations (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.731, + 0.785, + 0.757 + ], + "angle": 0, + "content": "5. Kingma, D.P., Welling, M.: Auto-Encoding Variational Bayes. In: International Conference on Learning Representations (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "6. Li, T.M., Lukáč, M., Gharbi, M., Ragan-Kelley, J.: Differentiable vector graphics rasterization for editing and learning. ACM Transactions on Graphics (TOG) 39(6), 1-15 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "7. Liao, Z., Hoppe, H., Forsyth, D., Yu, Y.: A subdivision-based representation for vector image editing. IEEE Transactions on Visualization and Computer Graphics 18(11), 1858-1867 (2012)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.246, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.127 + ], + "angle": 0, + "content": "H. Zhou et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.203 + ], + "angle": 0, + "content": "18. Liu, Y.T., Zhang, Z., Guo, Y.C., Fisher, M., Wang, Z., Zhang, S.H.: Dualvector: Unsupervised vector font synthesis with dual-part representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14193-14202 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.203, + 0.787, + 0.257 + ], + "angle": 0, + "content": "19. Lopes, R.G., Ha, D., Eck, D., Shlens, J.: A Learned Representation for Scalable Vector Graphics. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7929-7938. IEEE, Seoul, Korea (South) (Oct 2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.257, + 0.787, + 0.299 + ], + "angle": 0, + "content": "20. Ma, X., Zhou, Y., Xu, X., Sun, B., Filev, V., Orlov, N., Fu, Y., Shi, H.: Towards layer-wise image vectorization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16314-16323 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.299, + 0.787, + 0.324 + ], + "angle": 0, + "content": "21. Fluent emoji, https://github.com/microsoft/fluentui-emoji Accessed: 2023-10-27" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.325, + 0.787, + 0.367 + ], + "angle": 0, + "content": "22. Orzan, A., Bousseau, A., Winnemöller, H., Barla, P., Thollot, J., Salesin, D.: Diffusion curves: a vector representation for smooth-shaded images. ACM Transactions on Graphics (TOG) 27(3), 1-8 (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.367, + 0.787, + 0.393 + ], + "angle": 0, + "content": "23. Otsu, N.: A threshold selection method from gray-level histograms. IEEE transactions on systems, man, and cybernetics 9(1), 62-66 (1979)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.393, + 0.787, + 0.434 + ], + "angle": 0, + "content": "24. Reddy, P., Gharbi, M., Lukac, M., Mitra, N.J.: Im2vec: Synthesizing vector graphics without vector supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7342-7351 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.434, + 0.787, + 0.473 + ], + "angle": 0, + "content": "25. Reddy, P., Zhang, Z., Wang, Z., Fisher, M., Jin, H., Mitra, N.: A multi-implicit neural representation for fonts. Advances in Neural Information Processing Systems 34, 12637-12647 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.787, + 0.515 + ], + "angle": 0, + "content": "26. Richardt, C., Lopez-Moreno, J., Bousseau, A., Agrawala, M., Drettakis, G.: Vectorising bitmaps into semi-transparent gradient layers. Computer Graphics Forum (Proceedings of EGSR) 33(4), 11-19 (July 2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.515, + 0.787, + 0.556 + ], + "angle": 0, + "content": "27. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.556, + 0.787, + 0.597 + ], + "angle": 0, + "content": "28. Su, H., Liu, X., Niu, J., Cui, J., Wan, J., Wu, X., Wang, N.: Marvel: Raster gray-level manga vectorization via primitive-wise deep reinforcement learning. IEEE Transactions on Circuits and Systems for Video Technology (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.597, + 0.787, + 0.623 + ], + "angle": 0, + "content": "29. Sun, J., Liang, L., Wen, F., Shum, H.Y.: Image vectorization using optimized gradient meshes. ACM Transactions on Graphics (TOG) 26(3), 11-es (2007)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.623, + 0.787, + 0.664 + ], + "angle": 0, + "content": "30. Tian, X., Günther, T.: A survey of smooth vector graphics: Recent advances in representation, creation, rasterization and image vectorization. IEEE Transactions on Visualization and Computer Graphics (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.664, + 0.787, + 0.704 + ], + "angle": 0, + "content": "31. Van der Walt, S., Schonberger, J.L., Nunez-Iglesias, J., Boulogne, F., Warner, J.D., Yager, N., Gouillart, E., Yu, T.: scikit-image: image processing in python. PeerJ 2, e453 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.705, + 0.787, + 0.732 + ], + "angle": 0, + "content": "32. Wang, Y., Lian, Z.: Deepvecfont: Synthesizing high-quality vector fonts via dual-modality learning. ACM Transactions on Graphics (TOG) 40(6), 1-15 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.732, + 0.787, + 0.759 + ], + "angle": 0, + "content": "33. Xia, T., Liao, B., Yu, Y.: Patch-based image vectorization with automatic curvilinear feature alignment. ACM Transactions on Graphics (TOG) 28(5), 1-10 (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.787, + 0.799 + ], + "angle": 0, + "content": "34. Xie, G., Sun, X., Tong, X., Nowrouzezahrai, D.: Hierarchical diffusion curves for accurate automatic image vectorization. ACM Transactions on Graphics (TOG) 33(6), 1-11 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "35. Yang, M., Chao, H., Zhang, C., Guo, J., Yuan, L., Sun, J.: Effective clipart image vectorization through direct optimization of bezigons. IEEE transactions on visualization and computer graphics 22(2), 1063-1075 (2015)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.248, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Segmentation-guided Layer-wise Image Vectorization with Gradient Fills" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.177 + ], + "angle": 0, + "content": "36. Zhu, H., Cao, J., Xiao, Y., Chen, Z., Zhong, Z., Zhang, Y.J.: Tcb-spline-based image vectorization. ACM Transactions on Graphics (TOG) 41(3), 1-17 (2022)" + } + ] +] \ No newline at end of file diff --git a/2024/Segmentation-guided Layer-wise Image Vectorization with Gradient Fills/01f8c0e2-d236-4d33-947c-2747a133c0bb_origin.pdf b/2024/Segmentation-guided Layer-wise Image Vectorization with Gradient Fills/01f8c0e2-d236-4d33-947c-2747a133c0bb_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f6b39db210fe50179625d62f1682655b476518ed --- /dev/null +++ b/2024/Segmentation-guided Layer-wise Image Vectorization with Gradient Fills/01f8c0e2-d236-4d33-947c-2747a133c0bb_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fd6f3c46fb4491b08cf0f1ef1fc229e4509d5e0d56020e7f808c9181ec7ec463 +size 7298525 diff --git a/2024/Segmentation-guided Layer-wise Image Vectorization with Gradient Fills/full.md b/2024/Segmentation-guided Layer-wise Image Vectorization with Gradient Fills/full.md new file mode 100644 index 0000000000000000000000000000000000000000..987807042bf5ed9b7e43c617cf08214688479e53 --- /dev/null +++ b/2024/Segmentation-guided Layer-wise Image Vectorization with Gradient Fills/full.md @@ -0,0 +1,403 @@ +# Segmentation-guided Layer-wise Image Vectorization with Gradient Fills + +Hengyu Zhou, Hui Zhang, and Bin Wang + +School of Software, Tsinghua University, P. R. China +zhouhy22@mails.tsinghua.edu.cn, {huizhang, wangbins}@tsinghua.edu.cn + +Abstract. The widespread use of vector graphics creates a significant demand for vectorization methods. While recent learning-based techniques have shown their capability to create vector images of clear topology, filling these primitives with gradients remains a challenge. In this paper, we propose a segmentation-guided vectorization framework to convert raster images into concise vector graphics with radial gradient fills. With the guidance of an embedded gradient-aware segmentation subroutine, our approach progressively appends gradient-filled Bézier paths to the output, where primitive parameters are initiated with our newly designed initialization technique and are optimized to minimize our novel loss function. We build our method on a differentiable renderer with traditional segmentation algorithms to develop it as a model-free tool for raster-to-vector conversion. It is tested on various inputs to demonstrate its feasibility, independent of datasets, to synthesize vector graphics with improved visual quality and layer-wise topology compared to prior work. + +Keywords: Vectorization $\cdot$ Segmentation $\cdot$ Differentiable rendering + +# 1 Introduction + +Vector graphics offer great flexibility in digital design as they can be easily edited and arbitrarily scaled. Vectorization, the procedure of converting raster images to vector ones, serves as a second way to creating vector graphics other than designing from scratch where extensive artistic skills are required, and also as a bridge between the rapidly developing raster image generation [27] and the relatively less studied vector generation [3,8,10,24,32]. + +Vectorization has been explored for decades with various representations proposed to divide images into non-overlapping regions [17, 29, 33, 35, 36]. Despite their capability of generating vivid vectorization of realistic images, the complex primitives and lack of hierarchy make the vector output less intuitive for manipulation. + +Recent learning-based methods show their potential in preserving image hierarchy, where a vector image is often considered as a list of primitives and their order indicates the structure of the input. Many deep learning approaches + +have been proposed to vectorize simple inputs [24] or images of a specific field of interest [6, 18, 25, 28], but their dependency on models limits them to a particular domain and are not trivially generalizable. LIVE [20], in contrast, utilizes DiffVG [16], a differentiable renderer, to present a model-free vectorization framework. It progressively translates a raster image into an SVG in a layer-wise hierarchy, through which the topology of the input is preserved within the order of geometric primitives. However, its lack of support for gradients results in excessive primitives being added in case of images with rich gradient effects, and adding such support is not as simple as replacing RGBA colors with gradient parameters, as elaborated in Sec. 3.4. + +In this paper, we propose a novel segmentation-guided vectorization framework that extends the capability of LIVE to support radial gradients. The additional parameters of a radial gradient pose an increased challenge to optimization, where an effective method to determine whether a pixel contributes to a path's gradient fill is necessary. The key insight behind our idea is the similarity between finding contributing pixels and segmentation tasks, based on which we designed a segmentation-guided initialization procedure to progressively append new shapes to the vector output, and a novel loss function to optimize their geometric and gradient parameters. We evaluated our framework on several datasets with quantitative metrics and a user study, to show its effectiveness and superior performance compared to previous work. + +To summarize our contributions: + +- We introduce a segmentation-guided vectorization framework to create vector graphics automatically with layer-wise hierarchy and radial gradients. +- We propose a gradient-aware segmentation method to evaluate the pixel-wise contribution to the geometric and gradient parameters of a path. +- We take the segmentation as guidance for our new initialization technique and as a part of our novel segmentation-guided loss. + +# 2 Related Work + +# 2.1 Image Vectorization + +Most traditional vectorization methods aim to create vector images of high fidelity with different representations, which could be roughly categorized into mesh-based ones and curve-based ones. The former representations divide the input image into non-overlapping 2D patches across which colors are interpolated [30]. Shapes of patches include triangular [17,33,36], rectangular [1,29] or irregular ones such as bezigons [35]. The different selection of mesh shapes determines how patches are organized and how colors are interpolated within patches. The curved-based representations decompose the image into curves. Diffusion curves, for instance, use curves as geometric primitives with colors defined on sides of the curves [22,34]. While these methods can yield near-photo-realistic results, they may fall short when it comes to ease of editing due to the complex primitives and loss of topological information, compared to a plain list of + +simple primitives like how an SVG organizes a vector image. Recent learning-based vectorization work mainly takes the list-of-primitive approach, where the ordered primitives are seen as a sequence of drawing operations, and sequential prediction methods are applied to synthesize vector graphics. Models are used including recurrent neural networks [10,24], transformers [3], and are often combined with variational autoencoders [10,15,19,24]. DiffVG [16] as a differentiable renderer fills the gap between raster and vector graphics, with which loss functions on raster images could be used directly to optimize the vector images. A vectorization method optimizing all paths at once is also proposed in the work. + +# 2.2 Image Topology + +When synthesizing vector graphics, both similarity to the original image and correct topology are important for the ease of subsequent human manipulation. A related problem is layer decomposition, where images are decomposed into semi-transparent layers. With these layers being vector paths, such a method could serve as a layered vectorization approach. An interactive method [26] is among the first proposed to convert bitmaps into layered vector graphics. Photo2ClipArt [7] and other similar work [5] replace the heavy manual interaction with a user-provided segmentation input, from which the decomposition could be automatically determined. However, these methods require a concise segmentation for efficient and effective vectorization into a relatively small number of paths, where automatic segmentation algorithms struggle to serve the purpose [7]. + +Many learning-based methods formulate vector synthesis as sequential prediction problems, through which the order of primitives is naturally preserved. Some methods predict the primitive parameters directly. For example, Egiazarian et al. [6] use a transformer-based network to directly generate the vectorization of technical line drawings; MARVEL [28] applies deep reinforcement learning to predict strokes in black-and-white comics. Meanwhile, some approaches use variational autoencoders (VAE) [15] to create vector graphics. SketchRNN [10] is the first to introduce a Long Short-Term Memory (LSTM) [12] based VAE to the representation of vector sketches. The later DeepSVG [3] proposed a transformer-based VAE network that also encodes a vector image into a latent code, which could be decoded to reconstruct the vector paths. SVG-VAE [19] is the first to use raster images as the input to its encoder, thus it could be used to vectorize images. Im2Vec [24] also takes raster images as its input but drops the need for vector supervision with the help of DiffVG [16]. While these methods show great potential in creating topology-aware vector graphics, they rely on specific training datasets and may overlook the fine details in a complex input. The most similar work to ours is LIVE [20], in which a progressive framework is proposed to convert raster images into layered vector paths utilizing DiffVG for optimization, but its missing support for color gradients is not trivially achievable by simply optimizing gradient parameters. Our novel segmentation-guided method takes a step further to handle gradient effects properly, at the same level of simplicity as LIVE, requiring neither additional user input nor deep models. + +# 3 Method + +# 3.1 Method Overview + +![](images/3dec88129d8fb75647595f2a55b5411babe110ce346b8e2277518d2d7d74b425.jpg) +Fig. 1: Overview of our framework. $U + 1F61A$ from Noto Emoji [9] is used for demonstration. + +We present a new framework to generate SVG figures from rasterized images, with support for radial gradient fills via a segmentation-guided initialization and optimization process. + +Our framework works progressively, where at each epoch, single or multiple Bézier paths are added and optimized. Fig. 1 provides an overview of our method. At the beginning of each epoch $i$ , we calculate the difference between the input raster image and the output $I_{i-1}$ from the previous epoch. We segment it with our gradient-aware segmentation (Sec. 3.2), from which $n_i$ segmented regions are selected for initialization of new paths (Sec. 3.3). All added paths, including those added in previous epochs, are optimized to minimize the vectorization loss (Sec. 3.4). The parameters to be optimized consist of geometric parameters including positions of curve control points and gradient parameters including their center, radius, and color stops. $n_i$ , or the number of paths to be appended at $i$ -th epoch is specified by the user along with the number of iteration epochs. + +# 3.2 Gradient-aware Segmentation + +![](images/a9b54d52e89be1addffb28f0d37b1c8c91702903a19b122f40156d2b543e3d94.jpg) +(a) Input + +![](images/0c44a11ab887e4de40cfa39657c07e24d819e25682a4e5895ace2abc6176219f.jpg) +(b) LIVE + +![](images/ef37de576e9696f8f12ab2de5d11fbfa263c22519b1fbe56a557d2703addd118.jpg) +(c) LIVE-30 + +![](images/272a1c1e33a91fac29924bdad3320a8bea65678f33dc786df825f3a1b9af4993.jpg) +Fig. 2: Comparisons on segmentation methods. (a) is the raster input. (b) is by LIVE under default settings clustering colors into 200 bins. (c) is also by LIVE but with the number of bins set to 30. (d) is segmented using the Mean-shift [4] algorithm. (e) is by our method. + +![](images/b81f23cd872c02c150cefc4cebc05aef534eba67305943d94207dfa791c7f6ef.jpg) +(d) Mean shift +(e) Ours + +![](images/9bee42c81b824927e2bd089eb81b535cf4f6e949900277643d8dcc8ef4f4217d.jpg) +Fig. 3: Step-by-step outputs of our gradient-aware segmentation + +Quality of path initialization is the key to effective vectorization, hence we present a gradient-aware segmentation method to better estimate the area that a path is likely to span over. Its result is used in the subsequent operations including initialization and optimization (Sec. 3.3, Sec. 3.4). + +When paths are filled with a solid color, a connected component of similar or identical colors has a good chance of being covered by one path. LIVE [20] designed a component-wise initialization method, where pixels are clustered into buckets based on their l2-length over RGB channels, and connected pixels in the same bucket are considered one component. As gradients are considered, the clustering algorithm used by LIVE may result in an excessive segmentation, as shown by Fig. 2. Other clustering methods including Mean-shift [4] also suffer from over-segmentation for the same reason that colors within a path may vary more than colors between paths. + +With the limitations of previous methods, our approach is designed to detect edges of gradients. Colors should derive smoothly inside a region filled with the same gradient fill and are likely to change abruptly at its boundary. Thus, we calculate the secondary spatial gradient with a discrete Laplacian filter $L = \left[ \begin{array}{lll}1 & 1 & 1\\ 1 & -8 & 1\\ 1 & 1 & 1 \end{array} \right]$ to identify such rapid changes at gradient boundaries. Our segmentation method features the following steps: + +- The cross-correlation $S_0 \in \mathbb{R}_{w \times h \times 3}$ between the difference to the input and the 2D Laplacian filter is calculated as + +$$ +S _ {0} = \operatorname {c o r r e l a t e} ((I - \hat {I}) \mathbf {1} _ {| | \hat {I} - I | | _ {2} > \epsilon}, L), +$$ + +where $\hat{I}$ is the target and $I$ is the synthesized image. Pixels with an error below $\epsilon = 0.1$ are excluded. + +- $S_0$ is then summed over its RGB channels for a grayscale image $S_1 \in \mathbb{R}_{w \times h}$ , where each pixel of $S_1$ has a value of $(S_1)_{ij} = \sum_{c=1}^{3} |(S_0)_{jc}|$ . +- The grayscale image $S_{1}$ is converted to a binary image $S_{2} = \mathbf{1}_{S_{1} > \mathrm{Otsu}(S_{1})}$ with the threshold determined via Otsu's method [23]. + +![](images/83c3d80cb12b90f85cfc81b0155549a3862c5323e21685942860f29fcb242084.jpg) +Fig. 4: With paths being progressively added, the overall difference between the current output and the raster input decreases, thus we apply Otsu's method [23] to automatically determine the threshold for binarization. In this example, the red cheeks get segmented after more significant differences including eyes and brows are fitted, thanks to the dynamic threshold. The differences are normalized and mapped to grayscale. + +- Morphological closing [11] and the watershed algorithm [2] are then applied to $S_{2}$ for the final segmentation $S \in \mathbb{N}_{h \times w}$ . + +In Fig. 3 we showcase some intermediate results after each step. Since we are segmenting the difference between the output and the target, the already fitted pixels are ignored and the automatically determined threshold via Otsu's method decreases in response to the descending overall difference, as shown in Fig. 4. Compared to a fixed threshold, our method makes fewer assumptions about the input and avoids hyperparameters. + +# 3.3 Segmentation-guided Initialization + +Our vectorization method works in a progressive manner, throughout which we add one or more paths at each epoch. These newly added paths are initialized using our segmentation $S$ as the guidance. For each path to be added, we select the segmentation region with the largest accumulated square error calculated as + +$$ +w _ {i} = \sum_ {p \in \tilde {S} [ i ]} | | I _ {p} - \hat {I} _ {p} | | ^ {2} \tag {1} +$$ + +where $p$ is iterated over all pixels in $i$ -th region. This approach prioritizes larger regions to encourage a hierarchical initialization order and prevents regions that are almost properly filled from being chosen. A circle path of four cubic Bezier curves [20] is added at the selected region's center of mass $p_m$ . We fill the path with a radial gradient, which centers at $p_m$ , with a diameter equal to the geometric mean of the width and the height of the region's bounding box clipped to [0.2, 1.0]. The two stop colors at offsets $0\%$ and $100\%$ are both initialized to the color of the input image $I$ at $p_m$ . + +![](images/271aab27b9072f3beb784dde32be0362dbebad789c2aabff804240410a5615f5.jpg) +Fig. 5: Optimizing course for our method in comparison with UDF loss from LIVE. With our proposed segmentation-guided weight (SG Weight), the gradient fill is optimized to minimize the color error not only on the contour as UDF weight focuses on but also on colors inside the path, while excluding pixels occluded by eyes, brows, and the mouth. + +# 3.4 Loss Function with Segmentation-guided weight + +A common approach to measuring the difference between the synthesized image $\hat{I}$ and the target image $I$ is through mean squared error (MSE). However, simply minimizing MSE results in colors biasing towards the average color inside a path, as suggested by LIVE [20]. LIVE tackles the problem with its designed UDF loss to focus on pixels on the contour, but for radial gradients, correct colors on the contour do not mean correctness inside, as indicated by Fig. 5. When gradients are optimized with UDF loss, colors at the center are not fitted well, since those pixels are ignored by the UDF weight. We draw from the insight by LIVE emphasizing the significance of pixels on edges and extend this concept to include all pixels within a path, except for those occluded by other paths, which is well estimated through our gradient-aware segmentation. + +We formulate our segmentation-guided loss with this estimation of unoccluded pixels within a path. For each added path $p_i$ , given the set of pixels covered by the path and the set of pixels within the segment from which the path is initialized, we take their intersection and mark these pixels as being focused. The union of all focused pixels forms a set $F$ . We define the weight $w_{\mathrm{SG}}$ as: + +$$ +w _ {\mathrm {S G}} (i) = \left\{ \begin{array}{l l} \max \left(d _ {i} ^ {\prime}, \alpha_ {s}\right), & i \in F \\ d _ {i} ^ {\prime} \left(1 - \alpha_ {s}\right), & \text {o t h e r w i s e} \end{array} , \right. \tag {2} +$$ + +where $i$ is an index of pixel and $d_{i}^{\prime}$ is the UDF weight from [20]. The UDF weight is introduced to give the pixels near the path contour a higher weight, as depicted in Fig. 5. $\alpha_{s}$ balances between the UDF weight $d_{i}^{\prime}$ and our segmentation weight, + +![](images/1d72db136b7302ef3da7da709d00c1691c76b0f2b3bab6cc3232ee9ed74d1299.jpg) +Fig. 6: Bézier paths with four control points + +![](images/d2801467b160c455860087ca738f3f599776b5f96a363dc31336c188c6357722.jpg) + +and is set to 0.6 empirically. $d_i^\prime$ is defined with: + +$$ +d _ {i} ^ {\prime} = \frac {\operatorname {R e L U} (\tau - | d _ {i} |)}{\sum_ {j = 1} ^ {w \times h} \operatorname {R e L U} (\tau - | d _ {j} |)}, \tag {3} +$$ + +where $d_{i}$ is the distance from the pixel $i$ to the nearest path contour, which is then normalized and saturated with threshold $\tau = 10$ . + +We re-weight the pixel-wise color error with our proposed weight $w_{\mathrm{SG}}$ for the final segmentation-guided loss: + +$$ +\mathcal {L} _ {\mathrm {S G}} = \frac {1}{3} \sum_ {i = 1} ^ {w \times h} w _ {\mathrm {S G}} (i) \sum_ {c = 1} ^ {3} \left(I _ {c, i} - \hat {I} _ {c, i}\right) ^ {2}, \tag {4} +$$ + +where $i$ is iterated over all pixels and squared color differences are averaged over the RGB channels. + +We also introduce Xing loss [20] to help relieve self-intersection. Given a cubic Bézier curve with four control points $A, B, C$ , and $D$ , self-intersection is more likely to occur when the angle between $AB$ and $CD$ is greater than $180^{\circ}$ , as shown in Fig. 6. The Xing loss $\mathcal{L}_{\mathrm{Xing}}$ is defined to penalize angles over $180^{\circ}$ : + +$$ +D _ {1} = \left\{ \begin{array}{l} 1, \boldsymbol {A B} \times \boldsymbol {B C} > 0 \\ 0, \text {o t h e r w i s e} \end{array} , D _ {2} = \frac {\boldsymbol {A B} \times \boldsymbol {C D}}{\| \boldsymbol {A B} \| \| \boldsymbol {C D} \|} \right. \tag {5} +$$ + +$$ +\mathcal {L} _ {\mathrm {X i n g}} = D _ {1} \left(\operatorname {R e L U} (- D _ {2})\right) + (1 - D _ {1}) \left(\operatorname {R e L U} \left(D _ {2}\right)\right). \tag {6} +$$ + +Our optimization objective is defined as follows: + +$$ +\mathcal {L} = \mathcal {L} _ {\mathrm {S G}} + \lambda \mathcal {L} _ {\text {X i n g}}, \tag {7} +$$ + +where $\lambda$ is set to 0.05 empirically. + +# 4 Experiments + +# 4.1 Implementation Details + +We implement our framework with DiffVG [16] renderer based on PyTorch. Adam optimizer [14] is used with a learning rate of $10^{-2}$ and 1 for gradient parameters and path points respectively. We use scikit-image [31] for its implementation of morphological operations and the watershed algorithm [2]. + +![](images/5a939afbd6df4c936c9d83d10a475c62c34953ff75ff87491ceb8177d71491c8.jpg) +Noto + +![](images/6fe04ddb176c3692179b4337c8e29b9dfbc4b736e9553b6cc93df308dda46f7f.jpg) +Fluent + +![](images/1c6ae29aafc1fbee37bd2291f5399d7e2510bf77a3b8d2f95c9f38b43bdab7cf.jpg) +onont +Fig. 7: Exemplars from our chosen datasets for evaluation + +# 4.2 Datasets + +Noto Emoji We randomly select 256 emojis from Noto Emoji [9](Unicode 15.0) varying in colors and genres. In contrast to previous work [20,24], we take a more recent version of the dataset, consisting of resigned emojis filled with gradients instead of single colors. Emojis from this dataset consist of a relatively small number of paths with a clear hierarchy, thus we evaluate on this dataset to show our method's capability to decompose images into layers with gradients. + +Fluent Emojis We select an additional 256 images at random from Microsoft's Fluent Emojis [21] collection. This dataset features colorful emojis with rich gradient details. We test on this dataset to elaborate on our method's ability to match up with target images using fewer paths compared to previous work [16, 20]. + +Iconfont A set of 128 vector arts are fetched from the online Iconfont Illustration Library [13]. Images in the dataset contain an average of 1020 paths with great details but no gradients. We use these images to exhibit that our segmentation-guided approach is also suitable for vectorizing complex inputs, and performs on par with or better than previous work even without gradients. + +It is worth noticing that our framework is model-free, thus datasets mentioned above are for sole evaluation purposes. + +# 4.3 Qualitative Comparison + +With our initial goal of supporting gradient fills in a topology-preserving vectorization approach, we conducted a comparative analysis of the visual quality of our reconstructed vector graphics against LIVE [20]. Our method achieves superior visual quality in vectorization while utilizing the same number of paths as LIVE, as evidenced in Fig. 8. + +As depicted in Fig. 9, when presented with input containing gradients, LIVE struggles to accurately vectorize gradient-filled facial features, despite correctly capturing other facial parts. In contrast, our approach accurately reconstructs + +![](images/3a2a806be6bbfa4f15d50be9e7c6b8f40eb06747cc1efb06baa1b77796f4938c.jpg) +Fig. 8: Qualitative comparisons with LIVE + +the input using concise paths. Additionally, attempts to optimize gradient parameters without guidance, as seen in the 'w/o guidance' rows, result in degraded outcomes. + +# 4.4 Quantitative Comparison + +We further adopt quantitative metrics in a quantitative comparison with LIVE. We calculated PSNR to measure the difference between the rendered vectorized outputs and their corresponding raster inputs. For Noto and Fluent Emoji datasets, we add $\mathrm{clamp}(2^{i - 2},1,32)$ paths at $i$ -th initialization to reach a total of 256 paths. For the Iconfont dataset, 512 paths are used for each image with $\mathrm{clamp}(2^{i - 2},1,64)$ paths being added at $i$ -th epoch, for its higher complexity. + +As reported in Fig. 10, our method achieves generally faster convergence than LIVE, especially when a small number of paths are added. Our segmentation-guided framework can capture large-scale gradient features where LIVE tends to emit superfluous paths. With excessive paths being added, both methods yield the same level of quality in terms of PSNR. + +# 4.5 Layer Decomposition + +The framework we propose captures layer-wise structure during our progressive vectorization. The segmentation-guided initialization prioritizes segments that are more significant in terms of accumulated error, while details with relatively small errors are captured later by our adaptive gradient-aware segmentation, as shown in Fig. 4. Given an input with a clear hierarchy, these progressively added ordered paths resemble the handcrafted vector graphics, creating an easy-to-edit output, as shown in Fig. 11. + +![](images/314e5d14161c9505484e34ac184b4e1e451d87b390e33cab8a1651ac2de3ddac.jpg) +Fig.9: A comparison on number of paths needed to reconstruct major components. 'w/o guidance' refers to gradient fills being added without our proposed segmentation guidance. For the first input ( $U + 1F61A$ from Noto Emoji), our method vectorizes all elements with 8 paths, while adding more paths does not deteriorate the output. For the second input ( $U + 1F605$ from Fluent Emoji), ours reconstructs the facial parts with 6 paths and achieves a close gradient effect to the input using 16 paths. + +![](images/0e2b8445c99e9d3f2c42d5c891e7b42b7fc8769a9748d9491eb379dea2eb6b9b.jpg) +Fig. 10: Quantitative comparison between our method and LIVE. Our method achieves significantly lower error when the numbers of paths are small and converge to equal performance with excessive paths being added. + +![](images/29055c6a678f70d459984fcc6d2ca4d905d4e55a1170323c7b79c466b734bc67.jpg) +(a) Source + +![](images/7f5f729c936f9d9b4bfd42085c78e71cdc104b9874ef4fae90b49ef079d797c3.jpg) +(b) Vectorized + +![](images/e47813adba443833613a02bcac7fabb566f617b7658d3c28498cf3f4dc675496.jpg) +(c) Decomposed layers + +![](images/799b3b2d269561edaed1789aa892fa575fcf94f1e92a9072fddca9e3cfdec989.jpg) +(d) Edited +Fig. 11: Decomposed layers and recoloring + +Table 1: Results collected from our user study. The row 'Total' stands for the total number of answers collected for all the images in the dataset vectorized given the number of paths. As shown by the results, users in general prefer the vectorization by our method. + +(a) Noto Emoji + +
#Paths8163264Overall
LIVE39.0%39.0%40.8%43.0%40.4%
Ours61.0%61.0%59.2%57.0%59.6%
Total4494414414541785
+ +(b) Fluent Emoji + +
#Paths8163264Overall
LIVE40.2%38.4%30.8%29.3%34.7%
Ours59.8%61.6%69.2%70.7%65.3%
Total4334454524261756
+ +(c) Iconfont + +
#Paths3264128256Overall
LIVE32.2%42.7%48.5%44.3%42.1%
Ours67.8%57.3%51.5%55.7%57.9%
Total214241227237919
+ +Table 2: Results of our ablation study in PSNR. 'G' in the header is for 'with gradients' and 'S' is for 'with segmentation guidance'. The first row, with neither gradients nor segmentation, stands for LIVE. + +(a) Noto Emoji + +
G SN=8163264128256
24.9228.2131.4233.8335.8937.16
✓25.3328.6731.6334.1336.1237.34
✓25.1928.5131.8034.1736.0537.29
✓25.6928.9531.9234.2236.3037.46
+ +(b) Fluent Emoji + +
G SN=8163264128256
25.1928.7431.7434.4636.8538.38
✓26.2629.3632.2234.6736.9638.52
26.1529.5932.6435.2837.5438.91
✓26.7330.0732.9535.4337.6138.84
+ +(c) Iconfont + +
G SN=163264128256512
19.2221.6924.3827.2630.2532.36
✓20.0222.0224.2926.8629.4331.44
✓19.4221.9624.7227.6030.6732.72
✓20.3122.3524.7327.2829.7631.74
+ +# 4.6 User Study + +We conducted a questionnaire survey to obtain subjective feedback from users. Each participant is presented with 20 questions randomly selected from a total of 2,560 questions, that is, 4 numbers of paths multiplied by 640 images from the three datasets. In each question, participants are shown a raster input and two vectorized versions of it, one using LIVE and the other using our method, with the same number of paths. They are asked to choose the one that looks closer to the original image or, at an equivalent level of similarity, looks more appealing. Figure 12 shows examples. + +In our study, 223 people engaged, contributing a total of 4,460 votes. The results, presented in Tab. 1, reveal a clear preference for our approach. In the Fluent Emoji dataset, our method is notably preferred for these images with rich gradients. Despite the absence of gradients in many images from Noto Emoji and Iconfont, our method remains the preferred choice. + +![](images/3f03120f097d551ba6d703a1612e5360658d4314e6a4262652f60d0439d342af.jpg) +$U + 1F31A$ from Noto $(\mathrm{N} = 8)$ B is ours + +![](images/ac831756a9ef06ce24a81dbc99e9f7ed824d9e3bfbc07784e87fed7449a995ba.jpg) + +![](images/ab8bebc98ecda06a5b0bba83f7f139751adc9ad04a056ec164111a3235d056fd.jpg) + +![](images/dec3af428d2f61b369eef8f6c6ce3af1afd5fdc83187a77a1c31bc0307f24a82.jpg) +$U + 1F613$ from Fluent $(\mathrm{N} = 16)$ B is ours + +![](images/124a12716deafdc0cb609d07de31b5c6fd55d35e9ca6080d67271a436a4a76c6.jpg) + +![](images/94ab970c50e82372c67ee63562aaddfe82377609c16ca5e013ac2812d658ba84.jpg) + +![](images/0c6574c16804ba2649d8e8095898efde1e2942f971bcb1f6e9e93fb722a949da.jpg) +46087_07 from Iconfont (N=128); A is ours + +![](images/caea1c5ace2af1cc8865fb1579beb42d05d035c3e52a68bc34571a898ccf0aa8.jpg) + +![](images/bcefd55b40ec584b502c25a521b7e59b3ad09c1cae25cd6d9b6bbc1f27a3f4e5.jpg) + +![](images/cebfb94fb8bf20e7cb172033d3dd08cfd276237e3f938cddb50a59fe9f9b8fcb.jpg) +$U + 1F31A$ from Noto $(\mathrm{N} = 16)$ ; A is ours + +![](images/fa11a3a6768dc361cea0231ae016dc5f5f60cc69a37d4a77bd7526a16e6ce6a4.jpg) +Fig. 12: Examples from our questionnaire. Participants are asked to choose the better one for each pair of outputs. + +![](images/78a901925005728c8111e070baee41bcc8463775406c219abecc7b9b269f0c42.jpg) + +![](images/3da92b7fc92ed0453ec77b4b09121fe4756f0b56392c0abb8cfb19be793a2a88.jpg) +$U + 1F613$ from Fluent $(\mathrm{N} = 32)$ ; A is ours + +![](images/d81d572b93b621faf1286427918b3beb56fd865e2e51fe083cb22fe83b03f103.jpg) + +![](images/68a743928cffc3cb37b0fee6e4258e88ddbf60045d93dfd861ef480df9d2a1a0.jpg) + +![](images/0be836b6ebf0f0d9e8e7b9c9a97641604c09306c2c283a077d4b000013008ef7.jpg) +46087 07 from Iconfont (N=256); B is ours + +![](images/855a590285f2b9350f8761c5bfd3b2ef94d10f4938f3cba4f7b8b28cbaa3893b.jpg) + +![](images/45dc4111100e898d4d44fde5653e43c2d9e836343ca6cd2327df5d240ba7d13c.jpg) + +# 4.7 Ablation Study + +To verify the effectiveness of our contributions, we conducted an ablation study, by vectorizing without gradients, without segmentation guidance, or both, as shown in Fig. 13. The similarity between vectorized results and their corresponding inputs is measured by PSNR. Results are tabulated in Tab. 2. + +Gradients For images with rich gradient effects, such as those in Fluent Emoji, the introduction of gradients significantly improves the outcome. For Noto Emoji, where most images do not have gradients, the improvement is less noticeable. Iconfont is an interesting case; although its paths are filled with solid colors, a form of approximate color transition is achieved through multiple paths with progressively changing colors. Introducing gradients also contributes to a noticeable enhancement. + +Segmentation guidance Compared to solely introducing gradients, incorporating segmentation guidance has proven more effective. As illustrated in Tab. 2, a combination of gradients and segmentation yields superior results on both Noto and Fluent datasets. For the Iconfont dataset, the imitated color transition is + +![](images/41e3cc32362c02b3e1f5084e1f7b541cf628154234caa72e27fb7fc3fc3b4db8.jpg) +Fig. 13: Comparison between four combinations tested in the ablation study. When neither gradients nor segmentation guidance is applied, our method works the same as LIVE (1 $^{\text{st}}$ row). + +considered a gradient by the segmentation; thus, with fewer paths, the guided variants achieve a more approximate visual effect through large gradient paths. When more paths are added, using smaller, precise solid color paths yields better results, while the guided vectorization remains competitive in terms of accuracy, and is preferred in the user study. + +# 4.8 Limitations and Future Work + +![](images/32f07586b986c6a14aa440df4a6d4131642643c453afff1022d2b238873584ce.jpg) + +![](images/9c5843823faf84e41ffc9cbb5ead80817a5c02d86dd5fbbc2b4a82e6ba5970a2.jpg) +Input + +![](images/82501da6cabfd05a24469809b458386f1772e9f29829091ede3ec37dae16f4c0.jpg) + +![](images/527f92b36da004b16a4fbc7d0110e039ec7d47d56cfc8a4ad6e7735d77b980d5.jpg) +N=1 + +![](images/533546488afd40e530fe05a2a042d9d89f45c6161ea9c40f84993ded5fdb9fd1.jpg) + +![](images/baa1e215c3da4677aa28b621c516dd267ebb521f1113cdb3a1327c0cf0604152.jpg) +N=2 +Fig. 14: Intricate contours that our simple initial paths fail to converge to. Extra paths are added to mask the overflowing colors. + +![](images/80d75e55e0ff609700329c8cb544346771599352603365ddb2028f024300e070.jpg) + +![](images/e2b24f2811cf04a777189474e454efeac8350463483da55e28c9dcde2646ac81.jpg) +N=4 + +![](images/568d41cb898bdabcb4518222b82af754ecf767b26675c2862f82a6ca3f3ea360.jpg) + +![](images/79b1b7c39580d21f88f52a73667f71cc78f502affc0c82a152b477b3ec12eca1.jpg) + +![](images/700c9d2434b6811f066371519a78e3558d85ba1b96562e86b8eeffa7c76ae85d.jpg) + +![](images/88edb6154ccfe6408c8b1e577c6ed5bc99b95994aa3ebd7dab254252af6c35cb.jpg) +N=8 + +While our segmentation guidance effectively captures large-scale features, the initial shape, consisting of a loop formed by four connected Bezier curves, encounters challenges in accurately fitting intricate shapes, as illustrated in Figure 14. Achieving a balance between the number of paths and the number of control points per path needs further deliberation. + +Additionally, our implementation employs radial gradients with two color stops for simplicity, since a linear gradient can be interpreted as a radial gradient with its center positioned outside the path. Our method is not confined to a specific type of gradient. A more comprehensive implementation could involve dynamically determining gradient types and adjusting the number of color stops accordingly. + +# 5 Conclusion + +In this paper, we propose a segmentation-guided layer-wise vectorization framework, which synthesizes vector images by progressively adding paths and filling the paths with radial gradients. We design a gradient-aware segmentation method out of traditional algorithms to guide our novel initialization approach and newly designed loss function to address the obstacles in optimizing gradient parameters. We test our method on Noto [9] and Fluent Emoji [21] to show its capability in decomposing raster images into concise layers of gradient-filled paths when the topology is clear, and on Iconfont [13] to demonstrate its effectiveness in case of complex inputs. Our method shows superior performance compared to previous work. + +# Acknowledgements + +This work was supported by the NSFC under Grant 62072271. + +# References + +1. Baksteen, S.D., Hettinga, G.J., Echevarria, J., Kosinka, J.: Mesh colours for gradient meshes. STAG: Smart Tools and Applications in Graphics (2021) +2. Beucher, S., Meyer, F.: The morphological approach to segmentation: the watershed transformation. Mathematical morphology in image processing 34(1993), 49 (1993) +3. Carlier, A., Danelljan, M., Alahi, A., Timofte, R.: Deepsvg: A hierarchical generative network for vector graphics animation. Advances in Neural Information Processing Systems 33, 16351-16361 (2020) +4. Comaniciu, D., Meer, P.: Mean shift: A robust approach toward feature space analysis. IEEE Transactions on pattern analysis and machine intelligence 24(5), 603-619 (2002) +5. Du, Z.J., Kang, L.F., Tan, J., Gingold, Y., Xu, K.: Image vectorization and editing via linear gradient layer decomposition. ACM Transactions on Graphics (TOG) 42(4), 1-13 (2023) +6. Egiazarian, V., Voynov, O., Artemov, A., Volkhonskiy, D., Safin, A., Taktasheva, M., Zorin, D., Burnaev, E.: Deep vectorization of technical drawings. In: European conference on computer vision. pp. 582-598. Springer (2020) +7. Favreau, J.D., Lafarge, F., Bousseau, A.: Photo2clipart: Image abstraction and vectorization using layered linear gradients. ACM Transactions on Graphics (TOG) 36(6), 1-11 (2017) +8. Frans, K., Soros, L., Witkowski, O.: Clipdraw: Exploring text-to-drawing synthesis through language-image encoders. Advances in Neural Information Processing Systems 35, 5207-5218 (2022) +9. Noto emoji, https://github.com/googlefonts/sono-emoji Accessed: 2023-09-19 +0. Ha, D., Eck, D.: A neural representation of sketch drawings. In: International Conference on Learning Representations (2018) +1. Haralick, R.M., Sternberg, S.R., Zhuang, X.: Image analysis using mathematical morphology. IEEE transactions on pattern analysis and machine intelligence (4), 532-550 (1987) +2. Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural Computation 9(8), 1735-1780 (1997) +3. Vector illustrations library, https://www iconfont.cn/illustrations/index Accessed: 2023-11-02 +4. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. In: International Conference on Learning Representations (2015) +5. Kingma, D.P., Welling, M.: Auto-Encoding Variational Bayes. In: International Conference on Learning Representations (2014) +6. Li, T.M., Lukáč, M., Gharbi, M., Ragan-Kelley, J.: Differentiable vector graphics rasterization for editing and learning. ACM Transactions on Graphics (TOG) 39(6), 1-15 (2020) +7. Liao, Z., Hoppe, H., Forsyth, D., Yu, Y.: A subdivision-based representation for vector image editing. IEEE Transactions on Visualization and Computer Graphics 18(11), 1858-1867 (2012) + +18. Liu, Y.T., Zhang, Z., Guo, Y.C., Fisher, M., Wang, Z., Zhang, S.H.: Dualvector: Unsupervised vector font synthesis with dual-part representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14193-14202 (2023) +19. Lopes, R.G., Ha, D., Eck, D., Shlens, J.: A Learned Representation for Scalable Vector Graphics. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7929-7938. IEEE, Seoul, Korea (South) (Oct 2019) +20. Ma, X., Zhou, Y., Xu, X., Sun, B., Filev, V., Orlov, N., Fu, Y., Shi, H.: Towards layer-wise image vectorization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16314-16323 (2022) +21. Fluent emoji, https://github.com/microsoft/fluentui-emoji Accessed: 2023-10-27 +22. Orzan, A., Bousseau, A., Winnemöller, H., Barla, P., Thollot, J., Salesin, D.: Diffusion curves: a vector representation for smooth-shaded images. ACM Transactions on Graphics (TOG) 27(3), 1-8 (2008) +23. Otsu, N.: A threshold selection method from gray-level histograms. IEEE transactions on systems, man, and cybernetics 9(1), 62-66 (1979) +24. Reddy, P., Gharbi, M., Lukac, M., Mitra, N.J.: Im2vec: Synthesizing vector graphics without vector supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7342-7351 (2021) +25. Reddy, P., Zhang, Z., Wang, Z., Fisher, M., Jin, H., Mitra, N.: A multi-implicit neural representation for fonts. Advances in Neural Information Processing Systems 34, 12637-12647 (2021) +26. Richardt, C., Lopez-Moreno, J., Bousseau, A., Agrawala, M., Drettakis, G.: Vectorising bitmaps into semi-transparent gradient layers. Computer Graphics Forum (Proceedings of EGSR) 33(4), 11-19 (July 2014) +27. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022) +28. Su, H., Liu, X., Niu, J., Cui, J., Wan, J., Wu, X., Wang, N.: Marvel: Raster gray-level manga vectorization via primitive-wise deep reinforcement learning. IEEE Transactions on Circuits and Systems for Video Technology (2023) +29. Sun, J., Liang, L., Wen, F., Shum, H.Y.: Image vectorization using optimized gradient meshes. ACM Transactions on Graphics (TOG) 26(3), 11-es (2007) +30. Tian, X., Günther, T.: A survey of smooth vector graphics: Recent advances in representation, creation, rasterization and image vectorization. IEEE Transactions on Visualization and Computer Graphics (2022) +31. Van der Walt, S., Schonberger, J.L., Nunez-Iglesias, J., Boulogne, F., Warner, J.D., Yager, N., Gouillart, E., Yu, T.: scikit-image: image processing in python. PeerJ 2, e453 (2014) +32. Wang, Y., Lian, Z.: Deepvecfont: Synthesizing high-quality vector fonts via dual-modality learning. ACM Transactions on Graphics (TOG) 40(6), 1-15 (2021) +33. Xia, T., Liao, B., Yu, Y.: Patch-based image vectorization with automatic curvilinear feature alignment. ACM Transactions on Graphics (TOG) 28(5), 1-10 (2009) +34. Xie, G., Sun, X., Tong, X., Nowrouzezahrai, D.: Hierarchical diffusion curves for accurate automatic image vectorization. ACM Transactions on Graphics (TOG) 33(6), 1-11 (2014) +35. Yang, M., Chao, H., Zhang, C., Guo, J., Yuan, L., Sun, J.: Effective clipart image vectorization through direct optimization of bezigons. IEEE transactions on visualization and computer graphics 22(2), 1063-1075 (2015) + +36. Zhu, H., Cao, J., Xiao, Y., Chen, Z., Zhong, Z., Zhang, Y.J.: Tcb-spline-based image vectorization. ACM Transactions on Graphics (TOG) 41(3), 1-17 (2022) \ No newline at end of file diff --git a/2024/Segmentation-guided Layer-wise Image Vectorization with Gradient Fills/images.zip b/2024/Segmentation-guided Layer-wise Image Vectorization with Gradient Fills/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..17df4589007971e9048e6f1265dd34ac3238aca0 --- /dev/null +++ b/2024/Segmentation-guided Layer-wise Image Vectorization with Gradient Fills/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:268780250576e26b79092754bba2bafb3f531ba05c133ff3944ab34429d9fb55 +size 710569 diff --git a/2024/Segmentation-guided Layer-wise Image Vectorization with Gradient Fills/layout.json b/2024/Segmentation-guided Layer-wise Image Vectorization with Gradient Fills/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c52afe37f4d7c098eac567848e1b8705a9bd5cd6 --- /dev/null +++ b/2024/Segmentation-guided Layer-wise Image Vectorization with Gradient Fills/layout.json @@ -0,0 +1,10988 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 170, + 111, + 444, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 111, + 444, + 148 + ], + "spans": [ + { + "bbox": [ + 170, + 111, + 444, + 148 + ], + "type": "text", + "content": "Segmentation-guided Layer-wise Image Vectorization with Gradient Fills" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 196, + 167, + 417, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 167, + 417, + 181 + ], + "spans": [ + { + "bbox": [ + 196, + 167, + 417, + 181 + ], + "type": "text", + "content": "Hengyu Zhou, Hui Zhang, and Bin Wang" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 148, + 190, + 465, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 190, + 465, + 213 + ], + "spans": [ + { + "bbox": [ + 148, + 190, + 465, + 213 + ], + "type": "text", + "content": "School of Software, Tsinghua University, P. R. China \nzhouhy22@mails.tsinghua.edu.cn, {huizhang, wangbins}@tsinghua.edu.cn" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 243, + 455, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 243, + 455, + 410 + ], + "spans": [ + { + "bbox": [ + 160, + 243, + 455, + 410 + ], + "type": "text", + "content": "Abstract. The widespread use of vector graphics creates a significant demand for vectorization methods. While recent learning-based techniques have shown their capability to create vector images of clear topology, filling these primitives with gradients remains a challenge. In this paper, we propose a segmentation-guided vectorization framework to convert raster images into concise vector graphics with radial gradient fills. With the guidance of an embedded gradient-aware segmentation subroutine, our approach progressively appends gradient-filled Bézier paths to the output, where primitive parameters are initiated with our newly designed initialization technique and are optimized to minimize our novel loss function. We build our method on a differentiable renderer with traditional segmentation algorithms to develop it as a model-free tool for raster-to-vector conversion. It is tested on various inputs to demonstrate its feasibility, independent of datasets, to synthesize vector graphics with improved visual quality and layer-wise topology compared to prior work." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 419, + 440, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 419, + 440, + 431 + ], + "spans": [ + { + "bbox": [ + 160, + 419, + 440, + 431 + ], + "type": "text", + "content": "Keywords: Vectorization " + }, + { + "bbox": [ + 160, + 419, + 440, + 431 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 419, + 440, + 431 + ], + "type": "text", + "content": " Segmentation " + }, + { + "bbox": [ + 160, + 419, + 440, + 431 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 419, + 440, + 431 + ], + "type": "text", + "content": " Differentiable rendering" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 452, + 230, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 452, + 230, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 452, + 230, + 464 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 477, + 482, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 477, + 482, + 550 + ], + "spans": [ + { + "bbox": [ + 130, + 477, + 482, + 550 + ], + "type": "text", + "content": "Vector graphics offer great flexibility in digital design as they can be easily edited and arbitrarily scaled. Vectorization, the procedure of converting raster images to vector ones, serves as a second way to creating vector graphics other than designing from scratch where extensive artistic skills are required, and also as a bridge between the rapidly developing raster image generation [27] and the relatively less studied vector generation [3,8,10,24,32]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 550, + 482, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 550, + 482, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 550, + 482, + 609 + ], + "type": "text", + "content": "Vectorization has been explored for decades with various representations proposed to divide images into non-overlapping regions [17, 29, 33, 35, 36]. Despite their capability of generating vivid vectorization of realistic images, the complex primitives and lack of hierarchy make the vector output less intuitive for manipulation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 610, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 610, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 610, + 482, + 647 + ], + "type": "text", + "content": "Recent learning-based methods show their potential in preserving image hierarchy, where a vector image is often considered as a list of primitives and their order indicates the structure of the input. Many deep learning approaches" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 132, + 652, + 240, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 652, + 240, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 652, + 240, + 666 + ], + "type": "inline_equation", + "content": "\\boxtimes" + }, + { + "bbox": [ + 132, + 652, + 240, + 666 + ], + "type": "text", + "content": " Corresponding authors." + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "have been proposed to vectorize simple inputs [24] or images of a specific field of interest [6, 18, 25, 28], but their dependency on models limits them to a particular domain and are not trivially generalizable. LIVE [20], in contrast, utilizes DiffVG [16], a differentiable renderer, to present a model-free vectorization framework. It progressively translates a raster image into an SVG in a layer-wise hierarchy, through which the topology of the input is preserved within the order of geometric primitives. However, its lack of support for gradients results in excessive primitives being added in case of images with rich gradient effects, and adding such support is not as simple as replacing RGBA colors with gradient parameters, as elaborated in Sec. 3.4." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 236, + 482, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 482, + 368 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 482, + 368 + ], + "type": "text", + "content": "In this paper, we propose a novel segmentation-guided vectorization framework that extends the capability of LIVE to support radial gradients. The additional parameters of a radial gradient pose an increased challenge to optimization, where an effective method to determine whether a pixel contributes to a path's gradient fill is necessary. The key insight behind our idea is the similarity between finding contributing pixels and segmentation tasks, based on which we designed a segmentation-guided initialization procedure to progressively append new shapes to the vector output, and a novel loss function to optimize their geometric and gradient parameters. We evaluated our framework on several datasets with quantitative metrics and a user study, to show its effectiveness and superior performance compared to previous work." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 147, + 368, + 292, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 368, + 292, + 378 + ], + "spans": [ + { + "bbox": [ + 147, + 368, + 292, + 378 + ], + "type": "text", + "content": "To summarize our contributions:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 387, + 479, + 456 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 387, + 479, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 387, + 479, + 409 + ], + "spans": [ + { + "bbox": [ + 138, + 387, + 479, + 409 + ], + "type": "text", + "content": "- We introduce a segmentation-guided vectorization framework to create vector graphics automatically with layer-wise hierarchy and radial gradients." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 411, + 479, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 411, + 479, + 433 + ], + "spans": [ + { + "bbox": [ + 138, + 411, + 479, + 433 + ], + "type": "text", + "content": "- We propose a gradient-aware segmentation method to evaluate the pixel-wise contribution to the geometric and gradient parameters of a path." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 434, + 479, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 434, + 479, + 456 + ], + "spans": [ + { + "bbox": [ + 137, + 434, + 479, + 456 + ], + "type": "text", + "content": "- We take the segmentation as guidance for our new initialization technique and as a part of our novel segmentation-guided loss." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 476, + 237, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 476, + 237, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 476, + 237, + 489 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 502, + 262, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 262, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 262, + 514 + ], + "type": "text", + "content": "2.1 Image Vectorization" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "Most traditional vectorization methods aim to create vector images of high fidelity with different representations, which could be roughly categorized into mesh-based ones and curve-based ones. The former representations divide the input image into non-overlapping 2D patches across which colors are interpolated [30]. Shapes of patches include triangular [17,33,36], rectangular [1,29] or irregular ones such as bezigons [35]. The different selection of mesh shapes determines how patches are organized and how colors are interpolated within patches. The curved-based representations decompose the image into curves. Diffusion curves, for instance, use curves as geometric primitives with colors defined on sides of the curves [22,34]. While these methods can yield near-photo-realistic results, they may fall short when it comes to ease of editing due to the complex primitives and loss of topological information, compared to a plain list of" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "H. Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "simple primitives like how an SVG organizes a vector image. Recent learning-based vectorization work mainly takes the list-of-primitive approach, where the ordered primitives are seen as a sequence of drawing operations, and sequential prediction methods are applied to synthesize vector graphics. Models are used including recurrent neural networks [10,24], transformers [3], and are often combined with variational autoencoders [10,15,19,24]. DiffVG [16] as a differentiable renderer fills the gap between raster and vector graphics, with which loss functions on raster images could be used directly to optimize the vector images. A vectorization method optimizing all paths at once is also proposed in the work." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 240, + 242, + 252 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 240, + 242, + 252 + ], + "spans": [ + { + "bbox": [ + 131, + 240, + 242, + 252 + ], + "type": "text", + "content": "2.2 Image Topology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 258, + 482, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 482, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 482, + 402 + ], + "type": "text", + "content": "When synthesizing vector graphics, both similarity to the original image and correct topology are important for the ease of subsequent human manipulation. A related problem is layer decomposition, where images are decomposed into semi-transparent layers. With these layers being vector paths, such a method could serve as a layered vectorization approach. An interactive method [26] is among the first proposed to convert bitmaps into layered vector graphics. Photo2ClipArt [7] and other similar work [5] replace the heavy manual interaction with a user-provided segmentation input, from which the decomposition could be automatically determined. However, these methods require a concise segmentation for efficient and effective vectorization into a relatively small number of paths, where automatic segmentation algorithms struggle to serve the purpose [7]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 403, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 666 + ], + "type": "text", + "content": "Many learning-based methods formulate vector synthesis as sequential prediction problems, through which the order of primitives is naturally preserved. Some methods predict the primitive parameters directly. For example, Egiazarian et al. [6] use a transformer-based network to directly generate the vectorization of technical line drawings; MARVEL [28] applies deep reinforcement learning to predict strokes in black-and-white comics. Meanwhile, some approaches use variational autoencoders (VAE) [15] to create vector graphics. SketchRNN [10] is the first to introduce a Long Short-Term Memory (LSTM) [12] based VAE to the representation of vector sketches. The later DeepSVG [3] proposed a transformer-based VAE network that also encodes a vector image into a latent code, which could be decoded to reconstruct the vector paths. SVG-VAE [19] is the first to use raster images as the input to its encoder, thus it could be used to vectorize images. Im2Vec [24] also takes raster images as its input but drops the need for vector supervision with the help of DiffVG [16]. While these methods show great potential in creating topology-aware vector graphics, they rely on specific training datasets and may overlook the fine details in a complex input. The most similar work to ours is LIVE [20], in which a progressive framework is proposed to convert raster images into layered vector paths utilizing DiffVG for optimization, but its missing support for color gradients is not trivially achievable by simply optimizing gradient parameters. Our novel segmentation-guided method takes a step further to handle gradient effects properly, at the same level of simplicity as LIVE, requiring neither additional user input nor deep models." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 151, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 151, + 91, + 447, + 102 + ], + "type": "text", + "content": "Segmentation-guided Layer-wise Image Vectorization with Gradient Fills" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 202, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 202, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 202, + 126 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 144, + 251, + 155 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 144, + 251, + 155 + ], + "spans": [ + { + "bbox": [ + 132, + 144, + 251, + 155 + ], + "type": "text", + "content": "3.1 Method Overview" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 135, + 187, + 481, + 251 + ], + "blocks": [ + { + "bbox": [ + 135, + 187, + 481, + 251 + ], + "lines": [ + { + "bbox": [ + 135, + 187, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 135, + 187, + 481, + 251 + ], + "type": "image", + "image_path": "3dec88129d8fb75647595f2a55b5411babe110ce346b8e2277518d2d7d74b425.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 258, + 480, + 281 + ], + "lines": [ + { + "bbox": [ + 130, + 258, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 480, + 281 + ], + "type": "text", + "content": "Fig. 1: Overview of our framework. " + }, + { + "bbox": [ + 130, + 258, + 480, + 281 + ], + "type": "inline_equation", + "content": "U + 1F61A" + }, + { + "bbox": [ + 130, + 258, + 480, + 281 + ], + "type": "text", + "content": " from Noto Emoji [9] is used for demonstration." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 308, + 479, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 479, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 479, + 344 + ], + "type": "text", + "content": "We present a new framework to generate SVG figures from rasterized images, with support for radial gradient fills via a segmentation-guided initialization and optimization process." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 346, + 482, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 482, + 478 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 482, + 478 + ], + "type": "text", + "content": "Our framework works progressively, where at each epoch, single or multiple Bézier paths are added and optimized. Fig. 1 provides an overview of our method. At the beginning of each epoch " + }, + { + "bbox": [ + 130, + 346, + 482, + 478 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 346, + 482, + 478 + ], + "type": "text", + "content": ", we calculate the difference between the input raster image and the output " + }, + { + "bbox": [ + 130, + 346, + 482, + 478 + ], + "type": "inline_equation", + "content": "I_{i-1}" + }, + { + "bbox": [ + 130, + 346, + 482, + 478 + ], + "type": "text", + "content": " from the previous epoch. We segment it with our gradient-aware segmentation (Sec. 3.2), from which " + }, + { + "bbox": [ + 130, + 346, + 482, + 478 + ], + "type": "inline_equation", + "content": "n_i" + }, + { + "bbox": [ + 130, + 346, + 482, + 478 + ], + "type": "text", + "content": " segmented regions are selected for initialization of new paths (Sec. 3.3). All added paths, including those added in previous epochs, are optimized to minimize the vectorization loss (Sec. 3.4). The parameters to be optimized consist of geometric parameters including positions of curve control points and gradient parameters including their center, radius, and color stops. " + }, + { + "bbox": [ + 130, + 346, + 482, + 478 + ], + "type": "inline_equation", + "content": "n_i" + }, + { + "bbox": [ + 130, + 346, + 482, + 478 + ], + "type": "text", + "content": ", or the number of paths to be appended at " + }, + { + "bbox": [ + 130, + 346, + 482, + 478 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 346, + 482, + 478 + ], + "type": "text", + "content": "-th epoch is specified by the user along with the number of iteration epochs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 498, + 309, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 309, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 309, + 511 + ], + "type": "text", + "content": "3.2 Gradient-aware Segmentation" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 173, + 544, + 222, + 597 + ], + "blocks": [ + { + "bbox": [ + 173, + 544, + 222, + 597 + ], + "lines": [ + { + "bbox": [ + 173, + 544, + 222, + 597 + ], + "spans": [ + { + "bbox": [ + 173, + 544, + 222, + 597 + ], + "type": "image", + "image_path": "a9b54d52e89be1addffb28f0d37b1c8c91702903a19b122f40156d2b543e3d94.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 179, + 599, + 215, + 609 + ], + "lines": [ + { + "bbox": [ + 179, + 599, + 215, + 609 + ], + "spans": [ + { + "bbox": [ + 179, + 599, + 215, + 609 + ], + "type": "text", + "content": "(a) Input" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 228, + 544, + 276, + 596 + ], + "blocks": [ + { + "bbox": [ + 228, + 544, + 276, + 596 + ], + "lines": [ + { + "bbox": [ + 228, + 544, + 276, + 596 + ], + "spans": [ + { + "bbox": [ + 228, + 544, + 276, + 596 + ], + "type": "image", + "image_path": "0c44a11ab887e4de40cfa39657c07e24d819e25682a4e5895ace2abc6176219f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 234, + 599, + 270, + 609 + ], + "lines": [ + { + "bbox": [ + 234, + 599, + 270, + 609 + ], + "spans": [ + { + "bbox": [ + 234, + 599, + 270, + 609 + ], + "type": "text", + "content": "(b) LIVE" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 283, + 544, + 331, + 596 + ], + "blocks": [ + { + "bbox": [ + 283, + 544, + 331, + 596 + ], + "lines": [ + { + "bbox": [ + 283, + 544, + 331, + 596 + ], + "spans": [ + { + "bbox": [ + 283, + 544, + 331, + 596 + ], + "type": "image", + "image_path": "ef37de576e9696f8f12ab2de5d11fbfa263c22519b1fbe56a557d2703addd118.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 285, + 599, + 329, + 608 + ], + "lines": [ + { + "bbox": [ + 285, + 599, + 329, + 608 + ], + "spans": [ + { + "bbox": [ + 285, + 599, + 329, + 608 + ], + "type": "text", + "content": "(c) LIVE-30" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 337, + 544, + 386, + 597 + ], + "blocks": [ + { + "bbox": [ + 337, + 544, + 386, + 597 + ], + "lines": [ + { + "bbox": [ + 337, + 544, + 386, + 597 + ], + "spans": [ + { + "bbox": [ + 337, + 544, + 386, + 597 + ], + "type": "image", + "image_path": "272a1c1e33a91fac29924bdad3320a8bea65678f33dc786df825f3a1b9af4993.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 617, + 482, + 662 + ], + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 662 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 662 + ], + "type": "text", + "content": "Fig. 2: Comparisons on segmentation methods. (a) is the raster input. (b) is by LIVE under default settings clustering colors into 200 bins. (c) is also by LIVE but with the number of bins set to 30. (d) is segmented using the Mean-shift [4] algorithm. (e) is by our method." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 393, + 544, + 441, + 597 + ], + "blocks": [ + { + "bbox": [ + 336, + 599, + 388, + 608 + ], + "lines": [ + { + "bbox": [ + 336, + 599, + 388, + 608 + ], + "spans": [ + { + "bbox": [ + 336, + 599, + 388, + 608 + ], + "type": "text", + "content": "(d) Mean shift" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 393, + 544, + 441, + 597 + ], + "lines": [ + { + "bbox": [ + 393, + 544, + 441, + 597 + ], + "spans": [ + { + "bbox": [ + 393, + 544, + 441, + 597 + ], + "type": "image", + "image_path": "b81f23cd872c02c150cefc4cebc05aef534eba67305943d94207dfa791c7f6ef.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 401, + 599, + 433, + 608 + ], + "lines": [ + { + "bbox": [ + 401, + 599, + 433, + 608 + ], + "spans": [ + { + "bbox": [ + 401, + 599, + 433, + 608 + ], + "type": "text", + "content": "(e) Ours" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 224, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 224, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 224, + 100 + ], + "type": "text", + "content": "H. Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 141, + 114, + 471, + 259 + ], + "blocks": [ + { + "bbox": [ + 141, + 114, + 471, + 259 + ], + "lines": [ + { + "bbox": [ + 141, + 114, + 471, + 259 + ], + "spans": [ + { + "bbox": [ + 141, + 114, + 471, + 259 + ], + "type": "image", + "image_path": "9bee42c81b824927e2bd089eb81b535cf4f6e949900277643d8dcc8ef4f4217d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 174, + 270, + 440, + 281 + ], + "lines": [ + { + "bbox": [ + 174, + 270, + 440, + 281 + ], + "spans": [ + { + "bbox": [ + 174, + 270, + 440, + 281 + ], + "type": "text", + "content": "Fig. 3: Step-by-step outputs of our gradient-aware segmentation" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 302, + 480, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 302, + 480, + 350 + ], + "spans": [ + { + "bbox": [ + 130, + 302, + 480, + 350 + ], + "type": "text", + "content": "Quality of path initialization is the key to effective vectorization, hence we present a gradient-aware segmentation method to better estimate the area that a path is likely to span over. Its result is used in the subsequent operations including initialization and optimization (Sec. 3.3, Sec. 3.4)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 350, + 480, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 480, + 457 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 480, + 457 + ], + "type": "text", + "content": "When paths are filled with a solid color, a connected component of similar or identical colors has a good chance of being covered by one path. LIVE [20] designed a component-wise initialization method, where pixels are clustered into buckets based on their l2-length over RGB channels, and connected pixels in the same bucket are considered one component. As gradients are considered, the clustering algorithm used by LIVE may result in an excessive segmentation, as shown by Fig. 2. Other clustering methods including Mean-shift [4] also suffer from over-segmentation for the same reason that colors within a path may vary more than colors between paths." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 458, + 480, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 458, + 480, + 535 + ], + "spans": [ + { + "bbox": [ + 130, + 458, + 480, + 535 + ], + "type": "text", + "content": "With the limitations of previous methods, our approach is designed to detect edges of gradients. Colors should derive smoothly inside a region filled with the same gradient fill and are likely to change abruptly at its boundary. Thus, we calculate the secondary spatial gradient with a discrete Laplacian filter " + }, + { + "bbox": [ + 130, + 458, + 480, + 535 + ], + "type": "inline_equation", + "content": "L = \\left[ \\begin{array}{lll}1 & 1 & 1\\\\ 1 & -8 & 1\\\\ 1 & 1 & 1 \\end{array} \\right]" + }, + { + "bbox": [ + 130, + 458, + 480, + 535 + ], + "type": "text", + "content": " to identify such rapid changes at gradient boundaries. Our segmentation method features the following steps:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 540, + 481, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 540, + 481, + 564 + ], + "spans": [ + { + "bbox": [ + 137, + 540, + 481, + 564 + ], + "type": "text", + "content": "- The cross-correlation " + }, + { + "bbox": [ + 137, + 540, + 481, + 564 + ], + "type": "inline_equation", + "content": "S_0 \\in \\mathbb{R}_{w \\times h \\times 3}" + }, + { + "bbox": [ + 137, + 540, + 481, + 564 + ], + "type": "text", + "content": " between the difference to the input and the 2D Laplacian filter is calculated as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 233, + 571, + 395, + 587 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 571, + 395, + 587 + ], + "spans": [ + { + "bbox": [ + 233, + 571, + 395, + 587 + ], + "type": "interline_equation", + "content": "S _ {0} = \\operatorname {c o r r e l a t e} ((I - \\hat {I}) \\mathbf {1} _ {| | \\hat {I} - I | | _ {2} > \\epsilon}, L),", + "image_path": "608e48968dcef13f9bfd318c54bbf6e0ec1fdd7f3704ab43e5a454ce6e0ead62.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 147, + 594, + 479, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 594, + 479, + 616 + ], + "spans": [ + { + "bbox": [ + 147, + 594, + 479, + 616 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 147, + 594, + 479, + 616 + ], + "type": "inline_equation", + "content": "\\hat{I}" + }, + { + "bbox": [ + 147, + 594, + 479, + 616 + ], + "type": "text", + "content": " is the target and " + }, + { + "bbox": [ + 147, + 594, + 479, + 616 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 147, + 594, + 479, + 616 + ], + "type": "text", + "content": " is the synthesized image. Pixels with an error below " + }, + { + "bbox": [ + 147, + 594, + 479, + 616 + ], + "type": "inline_equation", + "content": "\\epsilon = 0.1" + }, + { + "bbox": [ + 147, + 594, + 479, + 616 + ], + "type": "text", + "content": " are excluded." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 137, + 618, + 479, + 665 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 137, + 618, + 479, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 618, + 479, + 642 + ], + "spans": [ + { + "bbox": [ + 137, + 618, + 479, + 642 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 137, + 618, + 479, + 642 + ], + "type": "inline_equation", + "content": "S_0" + }, + { + "bbox": [ + 137, + 618, + 479, + 642 + ], + "type": "text", + "content": " is then summed over its RGB channels for a grayscale image " + }, + { + "bbox": [ + 137, + 618, + 479, + 642 + ], + "type": "inline_equation", + "content": "S_1 \\in \\mathbb{R}_{w \\times h}" + }, + { + "bbox": [ + 137, + 618, + 479, + 642 + ], + "type": "text", + "content": ", where each pixel of " + }, + { + "bbox": [ + 137, + 618, + 479, + 642 + ], + "type": "inline_equation", + "content": "S_1" + }, + { + "bbox": [ + 137, + 618, + 479, + 642 + ], + "type": "text", + "content": " has a value of " + }, + { + "bbox": [ + 137, + 618, + 479, + 642 + ], + "type": "inline_equation", + "content": "(S_1)_{ij} = \\sum_{c=1}^{3} |(S_0)_{jc}|" + }, + { + "bbox": [ + 137, + 618, + 479, + 642 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 137, + 642, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 642, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 137, + 642, + 479, + 665 + ], + "type": "text", + "content": "- The grayscale image " + }, + { + "bbox": [ + 137, + 642, + 479, + 665 + ], + "type": "inline_equation", + "content": "S_{1}" + }, + { + "bbox": [ + 137, + 642, + 479, + 665 + ], + "type": "text", + "content": " is converted to a binary image " + }, + { + "bbox": [ + 137, + 642, + 479, + 665 + ], + "type": "inline_equation", + "content": "S_{2} = \\mathbf{1}_{S_{1} > \\mathrm{Otsu}(S_{1})}" + }, + { + "bbox": [ + 137, + 642, + 479, + 665 + ], + "type": "text", + "content": " with the threshold determined via Otsu's method [23]." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 151, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 151, + 91, + 448, + 102 + ], + "type": "text", + "content": "Segmentation-guided Layer-wise Image Vectorization with Gradient Fills" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 114, + 465, + 231 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 465, + 231 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 465, + 231 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 465, + 231 + ], + "type": "image", + "image_path": "83c3d80cb12b90f85cfc81b0155549a3862c5323e21685942860f29fcb242084.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 239, + 482, + 295 + ], + "lines": [ + { + "bbox": [ + 130, + 239, + 482, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 482, + 295 + ], + "type": "text", + "content": "Fig. 4: With paths being progressively added, the overall difference between the current output and the raster input decreases, thus we apply Otsu's method [23] to automatically determine the threshold for binarization. In this example, the red cheeks get segmented after more significant differences including eyes and brows are fitted, thanks to the dynamic threshold. The differences are normalized and mapped to grayscale." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 319, + 482, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 319, + 482, + 345 + ], + "spans": [ + { + "bbox": [ + 137, + 319, + 482, + 345 + ], + "type": "text", + "content": "- Morphological closing [11] and the watershed algorithm [2] are then applied to " + }, + { + "bbox": [ + 137, + 319, + 482, + 345 + ], + "type": "inline_equation", + "content": "S_{2}" + }, + { + "bbox": [ + 137, + 319, + 482, + 345 + ], + "type": "text", + "content": " for the final segmentation " + }, + { + "bbox": [ + 137, + 319, + 482, + 345 + ], + "type": "inline_equation", + "content": "S \\in \\mathbb{N}_{h \\times w}" + }, + { + "bbox": [ + 137, + 319, + 482, + 345 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 357, + 483, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 357, + 483, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 357, + 483, + 430 + ], + "type": "text", + "content": "In Fig. 3 we showcase some intermediate results after each step. Since we are segmenting the difference between the output and the target, the already fitted pixels are ignored and the automatically determined threshold via Otsu's method decreases in response to the descending overall difference, as shown in Fig. 4. Compared to a fixed threshold, our method makes fewer assumptions about the input and avoids hyperparameters." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 451, + 334, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 451, + 334, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 451, + 334, + 464 + ], + "type": "text", + "content": "3.3 Segmentation-guided Initialization" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 475, + 482, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 475, + 482, + 523 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 482, + 523 + ], + "type": "text", + "content": "Our vectorization method works in a progressive manner, throughout which we add one or more paths at each epoch. These newly added paths are initialized using our segmentation " + }, + { + "bbox": [ + 130, + 475, + 482, + 523 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 475, + 482, + 523 + ], + "type": "text", + "content": " as the guidance. For each path to be added, we select the segmentation region with the largest accumulated square error calculated as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 258, + 534, + 481, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 534, + 481, + 563 + ], + "spans": [ + { + "bbox": [ + 258, + 534, + 481, + 563 + ], + "type": "interline_equation", + "content": "w _ {i} = \\sum_ {p \\in \\tilde {S} [ i ]} | | I _ {p} - \\hat {I} _ {p} | | ^ {2} \\tag {1}", + "image_path": "71e59c84e1e7614695df560ce115bc5eb9e472adcbb447a810fb04b61bffca04.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " is iterated over all pixels in " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "-th region. This approach prioritizes larger regions to encourage a hierarchical initialization order and prevents regions that are almost properly filled from being chosen. A circle path of four cubic Bezier curves [20] is added at the selected region's center of mass " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "p_m" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": ". We fill the path with a radial gradient, which centers at " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "p_m" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": ", with a diameter equal to the geometric mean of the width and the height of the region's bounding box clipped to [0.2, 1.0]. The two stop colors at offsets " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "0\\%" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " are both initialized to the color of the input image " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "p_m" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "text", + "content": "H. Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 154, + 116, + 462, + 274 + ], + "blocks": [ + { + "bbox": [ + 154, + 116, + 462, + 274 + ], + "lines": [ + { + "bbox": [ + 154, + 116, + 462, + 274 + ], + "spans": [ + { + "bbox": [ + 154, + 116, + 462, + 274 + ], + "type": "image", + "image_path": "271aab27b9072f3beb784dde32be0362dbebad789c2aabff804240410a5615f5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 286, + 482, + 342 + ], + "lines": [ + { + "bbox": [ + 130, + 286, + 482, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 286, + 482, + 342 + ], + "type": "text", + "content": "Fig. 5: Optimizing course for our method in comparison with UDF loss from LIVE. With our proposed segmentation-guided weight (SG Weight), the gradient fill is optimized to minimize the color error not only on the contour as UDF weight focuses on but also on colors inside the path, while excluding pixels occluded by eyes, brows, and the mouth." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 366, + 402, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 366, + 402, + 379 + ], + "spans": [ + { + "bbox": [ + 132, + 366, + 402, + 379 + ], + "type": "text", + "content": "3.4 Loss Function with Segmentation-guided weight" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 389, + 482, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 389, + 482, + 520 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 482, + 520 + ], + "type": "text", + "content": "A common approach to measuring the difference between the synthesized image " + }, + { + "bbox": [ + 130, + 389, + 482, + 520 + ], + "type": "inline_equation", + "content": "\\hat{I}" + }, + { + "bbox": [ + 130, + 389, + 482, + 520 + ], + "type": "text", + "content": " and the target image " + }, + { + "bbox": [ + 130, + 389, + 482, + 520 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 389, + 482, + 520 + ], + "type": "text", + "content": " is through mean squared error (MSE). However, simply minimizing MSE results in colors biasing towards the average color inside a path, as suggested by LIVE [20]. LIVE tackles the problem with its designed UDF loss to focus on pixels on the contour, but for radial gradients, correct colors on the contour do not mean correctness inside, as indicated by Fig. 5. When gradients are optimized with UDF loss, colors at the center are not fitted well, since those pixels are ignored by the UDF weight. We draw from the insight by LIVE emphasizing the significance of pixels on edges and extend this concept to include all pixels within a path, except for those occluded by other paths, which is well estimated through our gradient-aware segmentation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 521, + 482, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 591 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 591 + ], + "type": "text", + "content": "We formulate our segmentation-guided loss with this estimation of unoccluded pixels within a path. For each added path " + }, + { + "bbox": [ + 130, + 521, + 482, + 591 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 130, + 521, + 482, + 591 + ], + "type": "text", + "content": ", given the set of pixels covered by the path and the set of pixels within the segment from which the path is initialized, we take their intersection and mark these pixels as being focused. The union of all focused pixels forms a set " + }, + { + "bbox": [ + 130, + 521, + 482, + 591 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 521, + 482, + 591 + ], + "type": "text", + "content": ". We define the weight " + }, + { + "bbox": [ + 130, + 521, + 482, + 591 + ], + "type": "inline_equation", + "content": "w_{\\mathrm{SG}}" + }, + { + "bbox": [ + 130, + 521, + 482, + 591 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 224, + 591, + 480, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 591, + 480, + 624 + ], + "spans": [ + { + "bbox": [ + 224, + 591, + 480, + 624 + ], + "type": "interline_equation", + "content": "w _ {\\mathrm {S G}} (i) = \\left\\{ \\begin{array}{l l} \\max \\left(d _ {i} ^ {\\prime}, \\alpha_ {s}\\right), & i \\in F \\\\ d _ {i} ^ {\\prime} \\left(1 - \\alpha_ {s}\\right), & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {2}", + "image_path": "821940b89ddf18b7bacfa3435daf276b3f9dae49423d92f43b2b3aa83fa196f9.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " is an index of pixel and " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "d_{i}^{\\prime}" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " is the UDF weight from [20]. The UDF weight is introduced to give the pixels near the path contour a higher weight, as depicted in Fig. 5. " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\alpha_{s}" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " balances between the UDF weight " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "d_{i}^{\\prime}" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " and our segmentation weight," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 151, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 151, + 91, + 448, + 102 + ], + "type": "text", + "content": "Segmentation-guided Layer-wise Image Vectorization with Gradient Fills" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 218, + 114, + 310, + 175 + ], + "blocks": [ + { + "bbox": [ + 218, + 114, + 310, + 175 + ], + "lines": [ + { + "bbox": [ + 218, + 114, + 310, + 175 + ], + "spans": [ + { + "bbox": [ + 218, + 114, + 310, + 175 + ], + "type": "image", + "image_path": "1d72db136b7302ef3da7da709d00c1691c76b0f2b3bab6cc3232ee9ed74d1299.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 214, + 182, + 399, + 194 + ], + "lines": [ + { + "bbox": [ + 214, + 182, + 399, + 194 + ], + "spans": [ + { + "bbox": [ + 214, + 182, + 399, + 194 + ], + "type": "text", + "content": "Fig. 6: Bézier paths with four control points" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 319, + 114, + 394, + 174 + ], + "blocks": [ + { + "bbox": [ + 319, + 114, + 394, + 174 + ], + "lines": [ + { + "bbox": [ + 319, + 114, + 394, + 174 + ], + "spans": [ + { + "bbox": [ + 319, + 114, + 394, + 174 + ], + "type": "image", + "image_path": "d2801467b160c455860087ca738f3f599776b5f96a363dc31336c188c6357722.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 216, + 341, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 341, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 341, + 228 + ], + "type": "text", + "content": "and is set to 0.6 empirically. " + }, + { + "bbox": [ + 132, + 216, + 341, + 228 + ], + "type": "inline_equation", + "content": "d_i^\\prime" + }, + { + "bbox": [ + 132, + 216, + 341, + 228 + ], + "type": "text", + "content": " is defined with:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 243, + 234, + 481, + 266 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 243, + 234, + 481, + 266 + ], + "spans": [ + { + "bbox": [ + 243, + 234, + 481, + 266 + ], + "type": "interline_equation", + "content": "d _ {i} ^ {\\prime} = \\frac {\\operatorname {R e L U} (\\tau - | d _ {i} |)}{\\sum_ {j = 1} ^ {w \\times h} \\operatorname {R e L U} (\\tau - | d _ {j} |)}, \\tag {3}", + "image_path": "9bd2faf5b90105415385f292d85061d0053b0bc97e2f448fe623f9c77d51c86e.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 271, + 479, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 271, + 479, + 294 + ], + "spans": [ + { + "bbox": [ + 131, + 271, + 479, + 294 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 271, + 479, + 294 + ], + "type": "inline_equation", + "content": "d_{i}" + }, + { + "bbox": [ + 131, + 271, + 479, + 294 + ], + "type": "text", + "content": " is the distance from the pixel " + }, + { + "bbox": [ + 131, + 271, + 479, + 294 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 271, + 479, + 294 + ], + "type": "text", + "content": " to the nearest path contour, which is then normalized and saturated with threshold " + }, + { + "bbox": [ + 131, + 271, + 479, + 294 + ], + "type": "inline_equation", + "content": "\\tau = 10" + }, + { + "bbox": [ + 131, + 271, + 479, + 294 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 295, + 480, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 295, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 131, + 295, + 480, + 319 + ], + "type": "text", + "content": "We re-weight the pixel-wise color error with our proposed weight " + }, + { + "bbox": [ + 131, + 295, + 480, + 319 + ], + "type": "inline_equation", + "content": "w_{\\mathrm{SG}}" + }, + { + "bbox": [ + 131, + 295, + 480, + 319 + ], + "type": "text", + "content": " for the final segmentation-guided loss:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 225, + 327, + 481, + 360 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 327, + 481, + 360 + ], + "spans": [ + { + "bbox": [ + 225, + 327, + 481, + 360 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {S G}} = \\frac {1}{3} \\sum_ {i = 1} ^ {w \\times h} w _ {\\mathrm {S G}} (i) \\sum_ {c = 1} ^ {3} \\left(I _ {c, i} - \\hat {I} _ {c, i}\\right) ^ {2}, \\tag {4}", + "image_path": "9119b2e34140df6e8075fc411b51c89e830f166233f2940f95ba2b4ed2379ce4.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 365, + 479, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 365, + 479, + 388 + ], + "spans": [ + { + "bbox": [ + 131, + 365, + 479, + 388 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 365, + 479, + 388 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 365, + 479, + 388 + ], + "type": "text", + "content": " is iterated over all pixels and squared color differences are averaged over the RGB channels." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 389, + 480, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 389, + 480, + 437 + ], + "spans": [ + { + "bbox": [ + 131, + 389, + 480, + 437 + ], + "type": "text", + "content": "We also introduce Xing loss [20] to help relieve self-intersection. Given a cubic Bézier curve with four control points " + }, + { + "bbox": [ + 131, + 389, + 480, + 437 + ], + "type": "inline_equation", + "content": "A, B, C" + }, + { + "bbox": [ + 131, + 389, + 480, + 437 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 389, + 480, + 437 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 131, + 389, + 480, + 437 + ], + "type": "text", + "content": ", self-intersection is more likely to occur when the angle between " + }, + { + "bbox": [ + 131, + 389, + 480, + 437 + ], + "type": "inline_equation", + "content": "AB" + }, + { + "bbox": [ + 131, + 389, + 480, + 437 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 389, + 480, + 437 + ], + "type": "inline_equation", + "content": "CD" + }, + { + "bbox": [ + 131, + 389, + 480, + 437 + ], + "type": "text", + "content": " is greater than " + }, + { + "bbox": [ + 131, + 389, + 480, + 437 + ], + "type": "inline_equation", + "content": "180^{\\circ}" + }, + { + "bbox": [ + 131, + 389, + 480, + 437 + ], + "type": "text", + "content": ", as shown in Fig. 6. The Xing loss " + }, + { + "bbox": [ + 131, + 389, + 480, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Xing}}" + }, + { + "bbox": [ + 131, + 389, + 480, + 437 + ], + "type": "text", + "content": " is defined to penalize angles over " + }, + { + "bbox": [ + 131, + 389, + 480, + 437 + ], + "type": "inline_equation", + "content": "180^{\\circ}" + }, + { + "bbox": [ + 131, + 389, + 480, + 437 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 195, + 445, + 481, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 445, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 195, + 445, + 481, + 479 + ], + "type": "interline_equation", + "content": "D _ {1} = \\left\\{ \\begin{array}{l} 1, \\boldsymbol {A B} \\times \\boldsymbol {B C} > 0 \\\\ 0, \\text {o t h e r w i s e} \\end{array} , D _ {2} = \\frac {\\boldsymbol {A B} \\times \\boldsymbol {C D}}{\\| \\boldsymbol {A B} \\| \\| \\boldsymbol {C D} \\|} \\right. \\tag {5}", + "image_path": "f1ee7aeb442fa0a9bfa902f401f42b141b0b9075f2a805cf5a0e055011faaf94.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 196, + 488, + 481, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 488, + 481, + 502 + ], + "spans": [ + { + "bbox": [ + 196, + 488, + 481, + 502 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {X i n g}} = D _ {1} \\left(\\operatorname {R e L U} (- D _ {2})\\right) + (1 - D _ {1}) \\left(\\operatorname {R e L U} \\left(D _ {2}\\right)\\right). \\tag {6}", + "image_path": "aaa34b73433abbb531bb699deeaa209b50cf3735a276a7d6691e60f223b64d65.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 146, + 506, + 360, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 506, + 360, + 517 + ], + "spans": [ + { + "bbox": [ + 146, + 506, + 360, + 517 + ], + "type": "text", + "content": "Our optimization objective is defined as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 263, + 525, + 481, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 263, + 525, + 481, + 539 + ], + "spans": [ + { + "bbox": [ + 263, + 525, + 481, + 539 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\mathrm {S G}} + \\lambda \\mathcal {L} _ {\\text {X i n g}}, \\tag {7}", + "image_path": "fbb74d9614ba7835c7f78fe0806ec25908fc0f8e410970344571517fde481fb4.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 546, + 282, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 546, + 282, + 558 + ], + "spans": [ + { + "bbox": [ + 132, + 546, + 282, + 558 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 546, + 282, + 558 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 132, + 546, + 282, + 558 + ], + "type": "text", + "content": " is set to 0.05 empirically." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 575, + 230, + 589 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 575, + 230, + 589 + ], + "spans": [ + { + "bbox": [ + 132, + 575, + 230, + 589 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 598, + 279, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 598, + 279, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 598, + 279, + 611 + ], + "type": "text", + "content": "4.1 Implementation Details" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 131, + 617, + 479, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 617, + 479, + 667 + ], + "spans": [ + { + "bbox": [ + 131, + 617, + 479, + 667 + ], + "type": "text", + "content": "We implement our framework with DiffVG [16] renderer based on PyTorch. Adam optimizer [14] is used with a learning rate of " + }, + { + "bbox": [ + 131, + 617, + 479, + 667 + ], + "type": "inline_equation", + "content": "10^{-2}" + }, + { + "bbox": [ + 131, + 617, + 479, + 667 + ], + "type": "text", + "content": " and 1 for gradient parameters and path points respectively. We use scikit-image [31] for its implementation of morphological operations and the watershed algorithm [2]." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "text", + "content": "H. Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 184, + 117, + 446, + 155 + ], + "blocks": [ + { + "bbox": [ + 170, + 122, + 179, + 145 + ], + "lines": [ + { + "bbox": [ + 170, + 122, + 179, + 145 + ], + "spans": [ + { + "bbox": [ + 170, + 122, + 179, + 145 + ], + "type": "text", + "content": "Noto" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 184, + 117, + 446, + 155 + ], + "lines": [ + { + "bbox": [ + 184, + 117, + 446, + 155 + ], + "spans": [ + { + "bbox": [ + 184, + 117, + 446, + 155 + ], + "type": "image", + "image_path": "5a939afbd6df4c936c9d83d10a475c62c34953ff75ff87491ceb8177d71491c8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 184, + 159, + 444, + 196 + ], + "blocks": [ + { + "bbox": [ + 171, + 163, + 181, + 190 + ], + "lines": [ + { + "bbox": [ + 171, + 163, + 181, + 190 + ], + "spans": [ + { + "bbox": [ + 171, + 163, + 181, + 190 + ], + "type": "text", + "content": "Fluent" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 184, + 159, + 444, + 196 + ], + "lines": [ + { + "bbox": [ + 184, + 159, + 444, + 196 + ], + "spans": [ + { + "bbox": [ + 184, + 159, + 444, + 196 + ], + "type": "image", + "image_path": "6fe04ddb176c3692179b4337c8e29b9dfbc4b736e9553b6cc93df308dda46f7f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 183, + 199, + 445, + 236 + ], + "blocks": [ + { + "bbox": [ + 172, + 201, + 181, + 233 + ], + "lines": [ + { + "bbox": [ + 172, + 201, + 181, + 233 + ], + "spans": [ + { + "bbox": [ + 172, + 201, + 181, + 233 + ], + "type": "text", + "content": "onont" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 183, + 199, + 445, + 236 + ], + "lines": [ + { + "bbox": [ + 183, + 199, + 445, + 236 + ], + "spans": [ + { + "bbox": [ + 183, + 199, + 445, + 236 + ], + "type": "image", + "image_path": "1c6ae29aafc1fbee37bd2291f5399d7e2510bf77a3b8d2f95c9f38b43bdab7cf.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 187, + 247, + 425, + 257 + ], + "lines": [ + { + "bbox": [ + 187, + 247, + 425, + 257 + ], + "spans": [ + { + "bbox": [ + 187, + 247, + 425, + 257 + ], + "type": "text", + "content": "Fig. 7: Exemplars from our chosen datasets for evaluation" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 282, + 204, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 204, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 204, + 293 + ], + "type": "text", + "content": "4.2 Datasets" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 302, + 480, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 302, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 130, + 302, + 480, + 374 + ], + "type": "text", + "content": "Noto Emoji We randomly select 256 emojis from Noto Emoji [9](Unicode 15.0) varying in colors and genres. In contrast to previous work [20,24], we take a more recent version of the dataset, consisting of resigned emojis filled with gradients instead of single colors. Emojis from this dataset consist of a relatively small number of paths with a clear hierarchy, thus we evaluate on this dataset to show our method's capability to decompose images into layers with gradients." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 387, + 480, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 387, + 480, + 436 + ], + "spans": [ + { + "bbox": [ + 130, + 387, + 480, + 436 + ], + "type": "text", + "content": "Fluent Emojis We select an additional 256 images at random from Microsoft's Fluent Emojis [21] collection. This dataset features colorful emojis with rich gradient details. We test on this dataset to elaborate on our method's ability to match up with target images using fewer paths compared to previous work [16, 20]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 447, + 480, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 447, + 480, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 447, + 480, + 506 + ], + "type": "text", + "content": "Iconfont A set of 128 vector arts are fetched from the online Iconfont Illustration Library [13]. Images in the dataset contain an average of 1020 paths with great details but no gradients. We use these images to exhibit that our segmentation-guided approach is also suitable for vectorizing complex inputs, and performs on par with or better than previous work even without gradients." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 507, + 480, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 507, + 480, + 531 + ], + "spans": [ + { + "bbox": [ + 130, + 507, + 480, + 531 + ], + "type": "text", + "content": "It is worth noticing that our framework is model-free, thus datasets mentioned above are for sole evaluation purposes." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 549, + 280, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 549, + 280, + 561 + ], + "spans": [ + { + "bbox": [ + 132, + 549, + 280, + 561 + ], + "type": "text", + "content": "4.3 Qualitative Comparison" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 570, + 480, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 480, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 480, + 629 + ], + "type": "text", + "content": "With our initial goal of supporting gradient fills in a topology-preserving vectorization approach, we conducted a comparative analysis of the visual quality of our reconstructed vector graphics against LIVE [20]. Our method achieves superior visual quality in vectorization while utilizing the same number of paths as LIVE, as evidenced in Fig. 8." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 630, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 480, + 665 + ], + "type": "text", + "content": "As depicted in Fig. 9, when presented with input containing gradients, LIVE struggles to accurately vectorize gradient-filled facial features, despite correctly capturing other facial parts. In contrast, our approach accurately reconstructs" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 151, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 151, + 91, + 447, + 102 + ], + "type": "text", + "content": "Segmentation-guided Layer-wise Image Vectorization with Gradient Fills" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 482, + 261 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 482, + 261 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 482, + 261 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 482, + 261 + ], + "type": "image", + "image_path": "3a2a806be6bbfa4f15d50be9e7c6b8f40eb06747cc1efb06baa1b77796f4938c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 217, + 269, + 395, + 281 + ], + "lines": [ + { + "bbox": [ + 217, + 269, + 395, + 281 + ], + "spans": [ + { + "bbox": [ + 217, + 269, + 395, + 281 + ], + "type": "text", + "content": "Fig. 8: Qualitative comparisons with LIVE" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 307, + 482, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 307, + 482, + 343 + ], + "spans": [ + { + "bbox": [ + 130, + 307, + 482, + 343 + ], + "type": "text", + "content": "the input using concise paths. Additionally, attempts to optimize gradient parameters without guidance, as seen in the 'w/o guidance' rows, result in degraded outcomes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 369, + 288, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 288, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 288, + 380 + ], + "type": "text", + "content": "4.4 Quantitative Comparison" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 396, + 482, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 396, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 482, + 468 + ], + "type": "text", + "content": "We further adopt quantitative metrics in a quantitative comparison with LIVE. We calculated PSNR to measure the difference between the rendered vectorized outputs and their corresponding raster inputs. For Noto and Fluent Emoji datasets, we add " + }, + { + "bbox": [ + 130, + 396, + 482, + 468 + ], + "type": "inline_equation", + "content": "\\mathrm{clamp}(2^{i - 2},1,32)" + }, + { + "bbox": [ + 130, + 396, + 482, + 468 + ], + "type": "text", + "content": " paths at " + }, + { + "bbox": [ + 130, + 396, + 482, + 468 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 396, + 482, + 468 + ], + "type": "text", + "content": "-th initialization to reach a total of 256 paths. For the Iconfont dataset, 512 paths are used for each image with " + }, + { + "bbox": [ + 130, + 396, + 482, + 468 + ], + "type": "inline_equation", + "content": "\\mathrm{clamp}(2^{i - 2},1,64)" + }, + { + "bbox": [ + 130, + 396, + 482, + 468 + ], + "type": "text", + "content": " paths being added at " + }, + { + "bbox": [ + 130, + 396, + 482, + 468 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 396, + 482, + 468 + ], + "type": "text", + "content": "-th epoch, for its higher complexity." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 469, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 469, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 469, + 482, + 529 + ], + "type": "text", + "content": "As reported in Fig. 10, our method achieves generally faster convergence than LIVE, especially when a small number of paths are added. Our segmentation-guided framework can capture large-scale gradient features where LIVE tends to emit superfluous paths. With excessive paths being added, both methods yield the same level of quality in terms of PSNR." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 554, + 266, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 554, + 266, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 554, + 266, + 567 + ], + "type": "text", + "content": "4.5 Layer Decomposition" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "The framework we propose captures layer-wise structure during our progressive vectorization. The segmentation-guided initialization prioritizes segments that are more significant in terms of accumulated error, while details with relatively small errors are captured later by our adaptive gradient-aware segmentation, as shown in Fig. 4. Given an input with a clear hierarchy, these progressively added ordered paths resemble the handcrafted vector graphics, creating an easy-to-edit output, as shown in Fig. 11." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 224, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 224, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 224, + 100 + ], + "type": "text", + "content": "H. Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 482, + 323 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 482, + 323 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 482, + 323 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 482, + 323 + ], + "type": "image", + "image_path": "314e5d14161c9505484e34ac184b4e1e451d87b390e33cab8a1651ac2de3ddac.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 331, + 482, + 399 + ], + "lines": [ + { + "bbox": [ + 130, + 331, + 482, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 482, + 399 + ], + "type": "text", + "content": "Fig.9: A comparison on number of paths needed to reconstruct major components. 'w/o guidance' refers to gradient fills being added without our proposed segmentation guidance. For the first input (" + }, + { + "bbox": [ + 130, + 331, + 482, + 399 + ], + "type": "inline_equation", + "content": "U + 1F61A" + }, + { + "bbox": [ + 130, + 331, + 482, + 399 + ], + "type": "text", + "content": " from Noto Emoji), our method vectorizes all elements with 8 paths, while adding more paths does not deteriorate the output. For the second input (" + }, + { + "bbox": [ + 130, + 331, + 482, + 399 + ], + "type": "inline_equation", + "content": "U + 1F605" + }, + { + "bbox": [ + 130, + 331, + 482, + 399 + ], + "type": "text", + "content": " from Fluent Emoji), ours reconstructs the facial parts with 6 paths and achieves a close gradient effect to the input using 16 paths." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 139, + 413, + 475, + 521 + ], + "blocks": [ + { + "bbox": [ + 139, + 413, + 475, + 521 + ], + "lines": [ + { + "bbox": [ + 139, + 413, + 475, + 521 + ], + "spans": [ + { + "bbox": [ + 139, + 413, + 475, + 521 + ], + "type": "image", + "image_path": "0e2b8445c99e9d3f2c42d5c891e7b42b7fc8769a9748d9491eb379dea2eb6b9b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 533, + 482, + 567 + ], + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 567 + ], + "type": "text", + "content": "Fig. 10: Quantitative comparison between our method and LIVE. Our method achieves significantly lower error when the numbers of paths are small and converge to equal performance with excessive paths being added." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 173, + 576, + 223, + 628 + ], + "blocks": [ + { + "bbox": [ + 173, + 576, + 223, + 628 + ], + "lines": [ + { + "bbox": [ + 173, + 576, + 223, + 628 + ], + "spans": [ + { + "bbox": [ + 173, + 576, + 223, + 628 + ], + "type": "image", + "image_path": "29055c6a678f70d459984fcc6d2ca4d905d4e55a1170323c7b79c466b734bc67.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 179, + 631, + 218, + 639 + ], + "lines": [ + { + "bbox": [ + 179, + 631, + 218, + 639 + ], + "spans": [ + { + "bbox": [ + 179, + 631, + 218, + 639 + ], + "type": "text", + "content": "(a) Source" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 228, + 577, + 279, + 628 + ], + "blocks": [ + { + "bbox": [ + 228, + 577, + 279, + 628 + ], + "lines": [ + { + "bbox": [ + 228, + 577, + 279, + 628 + ], + "spans": [ + { + "bbox": [ + 228, + 577, + 279, + 628 + ], + "type": "image", + "image_path": "7f5f729c936f9d9b4bfd42085c78e71cdc104b9874ef4fae90b49ef079d797c3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 228, + 631, + 280, + 639 + ], + "lines": [ + { + "bbox": [ + 228, + 631, + 280, + 639 + ], + "spans": [ + { + "bbox": [ + 228, + 631, + 280, + 639 + ], + "type": "text", + "content": "(b) Vectorized" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 286, + 577, + 380, + 628 + ], + "blocks": [ + { + "bbox": [ + 286, + 577, + 380, + 628 + ], + "lines": [ + { + "bbox": [ + 286, + 577, + 380, + 628 + ], + "spans": [ + { + "bbox": [ + 286, + 577, + 380, + 628 + ], + "type": "image", + "image_path": "e47813adba443833613a02bcac7fabb566f617b7658d3c28498cf3f4dc675496.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 294, + 631, + 375, + 640 + ], + "lines": [ + { + "bbox": [ + 294, + 631, + 375, + 640 + ], + "spans": [ + { + "bbox": [ + 294, + 631, + 375, + 640 + ], + "type": "text", + "content": "(c) Decomposed layers" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 389, + 577, + 441, + 628 + ], + "blocks": [ + { + "bbox": [ + 389, + 577, + 441, + 628 + ], + "lines": [ + { + "bbox": [ + 389, + 577, + 441, + 628 + ], + "spans": [ + { + "bbox": [ + 389, + 577, + 441, + 628 + ], + "type": "image", + "image_path": "799b3b2d269561edaed1789aa892fa575fcf94f1e92a9072fddca9e3cfdec989.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 396, + 631, + 435, + 640 + ], + "lines": [ + { + "bbox": [ + 396, + 631, + 435, + 640 + ], + "spans": [ + { + "bbox": [ + 396, + 631, + 435, + 640 + ], + "type": "text", + "content": "(d) Edited" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 217, + 649, + 395, + 661 + ], + "lines": [ + { + "bbox": [ + 217, + 649, + 395, + 661 + ], + "spans": [ + { + "bbox": [ + 217, + 649, + 395, + 661 + ], + "type": "text", + "content": "Fig. 11: Decomposed layers and recoloring" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 151, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 151, + 91, + 448, + 102 + ], + "type": "text", + "content": "Segmentation-guided Layer-wise Image Vectorization with Gradient Fills" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 114, + 301, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 114, + 301, + 191 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 301, + 191 + ], + "type": "text", + "content": "Table 1: Results collected from our user study. The row 'Total' stands for the total number of answers collected for all the images in the dataset vectorized given the number of paths. As shown by the results, users in general prefer the vectorization by our method." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 132, + 230, + 302, + 283 + ], + "blocks": [ + { + "bbox": [ + 189, + 211, + 246, + 221 + ], + "lines": [ + { + "bbox": [ + 189, + 211, + 246, + 221 + ], + "spans": [ + { + "bbox": [ + 189, + 211, + 246, + 221 + ], + "type": "text", + "content": "(a) Noto Emoji" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 230, + 302, + 283 + ], + "lines": [ + { + "bbox": [ + 132, + 230, + 302, + 283 + ], + "spans": [ + { + "bbox": [ + 132, + 230, + 302, + 283 + ], + "type": "table", + "html": "
#Paths8163264Overall
LIVE39.0%39.0%40.8%43.0%40.4%
Ours61.0%61.0%59.2%57.0%59.6%
Total4494414414541785
", + "image_path": "3aba846de821633cdeda0e4bdf6ba1c90c11b1e3c907ec4e43095d2f82858e96.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 132, + 310, + 302, + 363 + ], + "blocks": [ + { + "bbox": [ + 186, + 291, + 249, + 300 + ], + "lines": [ + { + "bbox": [ + 186, + 291, + 249, + 300 + ], + "spans": [ + { + "bbox": [ + 186, + 291, + 249, + 300 + ], + "type": "text", + "content": "(b) Fluent Emoji" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 310, + 302, + 363 + ], + "lines": [ + { + "bbox": [ + 132, + 310, + 302, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 310, + 302, + 363 + ], + "type": "table", + "html": "
#Paths8163264Overall
LIVE40.2%38.4%30.8%29.3%34.7%
Ours59.8%61.6%69.2%70.7%65.3%
Total4334454524261756
", + "image_path": "38c2cc24012b8921aecf766084d7b29cac7bb3c0361c28b77159a8d77407bc3c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 133, + 390, + 302, + 443 + ], + "blocks": [ + { + "bbox": [ + 195, + 371, + 239, + 380 + ], + "lines": [ + { + "bbox": [ + 195, + 371, + 239, + 380 + ], + "spans": [ + { + "bbox": [ + 195, + 371, + 239, + 380 + ], + "type": "text", + "content": "(c) Iconfont" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 390, + 302, + 443 + ], + "lines": [ + { + "bbox": [ + 133, + 390, + 302, + 443 + ], + "spans": [ + { + "bbox": [ + 133, + 390, + 302, + 443 + ], + "type": "table", + "html": "
#Paths3264128256Overall
LIVE32.2%42.7%48.5%44.3%42.1%
Ours67.8%57.3%51.5%55.7%57.9%
Total214241227237919
", + "image_path": "490990bdd7bfcb69848a54ce86644b70cae4aff48f226dc428a49f733430790c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 311, + 114, + 482, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 311, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 2: Results of our ablation study in PSNR. 'G' in the header is for 'with gradients' and 'S' is for 'with segmentation guidance'. The first row, with neither gradients nor segmentation, stands for LIVE." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 313, + 209, + 481, + 270 + ], + "blocks": [ + { + "bbox": [ + 369, + 189, + 425, + 199 + ], + "lines": [ + { + "bbox": [ + 369, + 189, + 425, + 199 + ], + "spans": [ + { + "bbox": [ + 369, + 189, + 425, + 199 + ], + "type": "text", + "content": "(a) Noto Emoji" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 209, + 481, + 270 + ], + "lines": [ + { + "bbox": [ + 313, + 209, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 313, + 209, + 481, + 270 + ], + "type": "table", + "html": "
G SN=8163264128256
24.9228.2131.4233.8335.8937.16
✓25.3328.6731.6334.1336.1237.34
✓25.1928.5131.8034.1736.0537.29
✓25.6928.9531.9234.2236.3037.46
", + "image_path": "6dd969033e304d5d203acc118fbacfe22194c67481a296cab1e67bce02bb7cc6.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 313, + 298, + 481, + 361 + ], + "blocks": [ + { + "bbox": [ + 365, + 279, + 429, + 289 + ], + "lines": [ + { + "bbox": [ + 365, + 279, + 429, + 289 + ], + "spans": [ + { + "bbox": [ + 365, + 279, + 429, + 289 + ], + "type": "text", + "content": "(b) Fluent Emoji" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 298, + 481, + 361 + ], + "lines": [ + { + "bbox": [ + 313, + 298, + 481, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 298, + 481, + 361 + ], + "type": "table", + "html": "
G SN=8163264128256
25.1928.7431.7434.4636.8538.38
✓26.2629.3632.2234.6736.9638.52
26.1529.5932.6435.2837.5438.91
✓26.7330.0732.9535.4337.6138.84
", + "image_path": "0a17e529b8d09f1e799f9cb383b5397646b0c46ee6226c15f399989408aa1043.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 313, + 388, + 481, + 451 + ], + "blocks": [ + { + "bbox": [ + 375, + 369, + 419, + 379 + ], + "lines": [ + { + "bbox": [ + 375, + 369, + 419, + 379 + ], + "spans": [ + { + "bbox": [ + 375, + 369, + 419, + 379 + ], + "type": "text", + "content": "(c) Iconfont" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 388, + 481, + 451 + ], + "lines": [ + { + "bbox": [ + 313, + 388, + 481, + 451 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 481, + 451 + ], + "type": "table", + "html": "
G SN=163264128256512
19.2221.6924.3827.2630.2532.36
✓20.0222.0224.2926.8629.4331.44
✓19.4221.9624.7227.6030.6732.72
✓20.3122.3524.7327.2829.7631.74
", + "image_path": "f2a417a30ad9b4ebe1abcfa6a20f3feef9c2cc7d98f1d060345ddbe7520b4226.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 477, + 218, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 477, + 218, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 477, + 218, + 489 + ], + "type": "text", + "content": "4.6 User Study" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 507, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 507, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 507, + 482, + 604 + ], + "type": "text", + "content": "We conducted a questionnaire survey to obtain subjective feedback from users. Each participant is presented with 20 questions randomly selected from a total of 2,560 questions, that is, 4 numbers of paths multiplied by 640 images from the three datasets. In each question, participants are shown a raster input and two vectorized versions of it, one using LIVE and the other using our method, with the same number of paths. They are asked to choose the one that looks closer to the original image or, at an equivalent level of similarity, looks more appealing. Figure 12 shows examples." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "In our study, 223 people engaged, contributing a total of 4,460 votes. The results, presented in Tab. 1, reveal a clear preference for our approach. In the Fluent Emoji dataset, our method is notably preferred for these images with rich gradients. Despite the absence of gradients in many images from Noto Emoji and Iconfont, our method remains the preferred choice." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "H. Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 116, + 170, + 152 + ], + "blocks": [ + { + "bbox": [ + 133, + 116, + 170, + 152 + ], + "lines": [ + { + "bbox": [ + 133, + 116, + 170, + 152 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 170, + 152 + ], + "type": "image", + "image_path": "3f03120f097d551ba6d703a1612e5360658d4314e6a4262652f60d0439d342af.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 141, + 154, + 240, + 162 + ], + "lines": [ + { + "bbox": [ + 141, + 154, + 240, + 162 + ], + "spans": [ + { + "bbox": [ + 141, + 154, + 240, + 162 + ], + "type": "inline_equation", + "content": "U + 1F31A" + }, + { + "bbox": [ + 141, + 154, + 240, + 162 + ], + "type": "text", + "content": " from Noto " + }, + { + "bbox": [ + 141, + 154, + 240, + 162 + ], + "type": "inline_equation", + "content": "(\\mathrm{N} = 8)" + }, + { + "bbox": [ + 141, + 154, + 240, + 162 + ], + "type": "text", + "content": " B is ours" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 171, + 116, + 208, + 152 + ], + "blocks": [ + { + "bbox": [ + 171, + 116, + 208, + 152 + ], + "lines": [ + { + "bbox": [ + 171, + 116, + 208, + 152 + ], + "spans": [ + { + "bbox": [ + 171, + 116, + 208, + 152 + ], + "type": "image", + "image_path": "ac831756a9ef06ce24a81dbc99e9f7ed824d9e3bfbc07784e87fed7449a995ba.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 211, + 116, + 247, + 152 + ], + "blocks": [ + { + "bbox": [ + 211, + 116, + 247, + 152 + ], + "lines": [ + { + "bbox": [ + 211, + 116, + 247, + 152 + ], + "spans": [ + { + "bbox": [ + 211, + 116, + 247, + 152 + ], + "type": "image", + "image_path": "ab8bebc98ecda06a5b0bba83f7f139751adc9ad04a056ec164111a3235d056fd.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 249, + 116, + 285, + 152 + ], + "blocks": [ + { + "bbox": [ + 249, + 116, + 285, + 152 + ], + "lines": [ + { + "bbox": [ + 249, + 116, + 285, + 152 + ], + "spans": [ + { + "bbox": [ + 249, + 116, + 285, + 152 + ], + "type": "image", + "image_path": "dec3af428d2f61b369eef8f6c6ce3af1afd5fdc83187a77a1c31bc0307f24a82.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 255, + 154, + 359, + 162 + ], + "lines": [ + { + "bbox": [ + 255, + 154, + 359, + 162 + ], + "spans": [ + { + "bbox": [ + 255, + 154, + 359, + 162 + ], + "type": "inline_equation", + "content": "U + 1F613" + }, + { + "bbox": [ + 255, + 154, + 359, + 162 + ], + "type": "text", + "content": " from Fluent " + }, + { + "bbox": [ + 255, + 154, + 359, + 162 + ], + "type": "inline_equation", + "content": "(\\mathrm{N} = 16)" + }, + { + "bbox": [ + 255, + 154, + 359, + 162 + ], + "type": "text", + "content": " B is ours" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 286, + 116, + 323, + 152 + ], + "blocks": [ + { + "bbox": [ + 286, + 116, + 323, + 152 + ], + "lines": [ + { + "bbox": [ + 286, + 116, + 323, + 152 + ], + "spans": [ + { + "bbox": [ + 286, + 116, + 323, + 152 + ], + "type": "image", + "image_path": "124a12716deafdc0cb609d07de31b5c6fd55d35e9ca6080d67271a436a4a76c6.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 327, + 116, + 362, + 152 + ], + "blocks": [ + { + "bbox": [ + 327, + 116, + 362, + 152 + ], + "lines": [ + { + "bbox": [ + 327, + 116, + 362, + 152 + ], + "spans": [ + { + "bbox": [ + 327, + 116, + 362, + 152 + ], + "type": "image", + "image_path": "94ab970c50e82372c67ee63562aaddfe82377609c16ca5e013ac2812d658ba84.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 365, + 116, + 403, + 152 + ], + "blocks": [ + { + "bbox": [ + 365, + 116, + 403, + 152 + ], + "lines": [ + { + "bbox": [ + 365, + 116, + 403, + 152 + ], + "spans": [ + { + "bbox": [ + 365, + 116, + 403, + 152 + ], + "type": "image", + "image_path": "0c6574c16804ba2649d8e8095898efde1e2942f971bcb1f6e9e93fb722a949da.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 367, + 154, + 479, + 163 + ], + "lines": [ + { + "bbox": [ + 367, + 154, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 367, + 154, + 479, + 163 + ], + "type": "text", + "content": "46087_07 from Iconfont (N=128); A is ours" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 404, + 116, + 442, + 152 + ], + "blocks": [ + { + "bbox": [ + 404, + 116, + 442, + 152 + ], + "lines": [ + { + "bbox": [ + 404, + 116, + 442, + 152 + ], + "spans": [ + { + "bbox": [ + 404, + 116, + 442, + 152 + ], + "type": "image", + "image_path": "caea1c5ace2af1cc8865fb1579beb42d05d035c3e52a68bc34571a898ccf0aa8.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 444, + 116, + 480, + 152 + ], + "blocks": [ + { + "bbox": [ + 444, + 116, + 480, + 152 + ], + "lines": [ + { + "bbox": [ + 444, + 116, + 480, + 152 + ], + "spans": [ + { + "bbox": [ + 444, + 116, + 480, + 152 + ], + "type": "image", + "image_path": "bcefd55b40ec584b502c25a521b7e59b3ad09c1cae25cd6d9b6bbc1f27a3f4e5.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 133, + 164, + 170, + 201 + ], + "blocks": [ + { + "bbox": [ + 133, + 164, + 170, + 201 + ], + "lines": [ + { + "bbox": [ + 133, + 164, + 170, + 201 + ], + "spans": [ + { + "bbox": [ + 133, + 164, + 170, + 201 + ], + "type": "image", + "image_path": "cebfb94fb8bf20e7cb172033d3dd08cfd276237e3f938cddb50a59fe9f9b8fcb.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 140, + 202, + 241, + 211 + ], + "lines": [ + { + "bbox": [ + 140, + 202, + 241, + 211 + ], + "spans": [ + { + "bbox": [ + 140, + 202, + 241, + 211 + ], + "type": "inline_equation", + "content": "U + 1F31A" + }, + { + "bbox": [ + 140, + 202, + 241, + 211 + ], + "type": "text", + "content": " from Noto " + }, + { + "bbox": [ + 140, + 202, + 241, + 211 + ], + "type": "inline_equation", + "content": "(\\mathrm{N} = 16)" + }, + { + "bbox": [ + 140, + 202, + 241, + 211 + ], + "type": "text", + "content": " ; A is ours" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 171, + 164, + 208, + 201 + ], + "blocks": [ + { + "bbox": [ + 171, + 164, + 208, + 201 + ], + "lines": [ + { + "bbox": [ + 171, + 164, + 208, + 201 + ], + "spans": [ + { + "bbox": [ + 171, + 164, + 208, + 201 + ], + "type": "image", + "image_path": "fa11a3a6768dc361cea0231ae016dc5f5f60cc69a37d4a77bd7526a16e6ce6a4.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 221, + 480, + 244 + ], + "lines": [ + { + "bbox": [ + 131, + 221, + 480, + 244 + ], + "spans": [ + { + "bbox": [ + 131, + 221, + 480, + 244 + ], + "type": "text", + "content": "Fig. 12: Examples from our questionnaire. Participants are asked to choose the better one for each pair of outputs." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 211, + 164, + 247, + 201 + ], + "blocks": [ + { + "bbox": [ + 211, + 164, + 247, + 201 + ], + "lines": [ + { + "bbox": [ + 211, + 164, + 247, + 201 + ], + "spans": [ + { + "bbox": [ + 211, + 164, + 247, + 201 + ], + "type": "image", + "image_path": "78a901925005728c8111e070baee41bcc8463775406c219abecc7b9b269f0c42.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 251, + 165, + 285, + 201 + ], + "blocks": [ + { + "bbox": [ + 251, + 165, + 285, + 201 + ], + "lines": [ + { + "bbox": [ + 251, + 165, + 285, + 201 + ], + "spans": [ + { + "bbox": [ + 251, + 165, + 285, + 201 + ], + "type": "image", + "image_path": "3da92b7fc92ed0453ec77b4b09121fe4756f0b56392c0abb8cfb19be793a2a88.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 255, + 202, + 359, + 211 + ], + "lines": [ + { + "bbox": [ + 255, + 202, + 359, + 211 + ], + "spans": [ + { + "bbox": [ + 255, + 202, + 359, + 211 + ], + "type": "inline_equation", + "content": "U + 1F613" + }, + { + "bbox": [ + 255, + 202, + 359, + 211 + ], + "type": "text", + "content": " from Fluent " + }, + { + "bbox": [ + 255, + 202, + 359, + 211 + ], + "type": "inline_equation", + "content": "(\\mathrm{N} = 32)" + }, + { + "bbox": [ + 255, + 202, + 359, + 211 + ], + "type": "text", + "content": " ; A is ours" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 287, + 165, + 323, + 200 + ], + "blocks": [ + { + "bbox": [ + 287, + 165, + 323, + 200 + ], + "lines": [ + { + "bbox": [ + 287, + 165, + 323, + 200 + ], + "spans": [ + { + "bbox": [ + 287, + 165, + 323, + 200 + ], + "type": "image", + "image_path": "d81d572b93b621faf1286427918b3beb56fd865e2e51fe083cb22fe83b03f103.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 327, + 165, + 362, + 201 + ], + "blocks": [ + { + "bbox": [ + 327, + 165, + 362, + 201 + ], + "lines": [ + { + "bbox": [ + 327, + 165, + 362, + 201 + ], + "spans": [ + { + "bbox": [ + 327, + 165, + 362, + 201 + ], + "type": "image", + "image_path": "68a743928cffc3cb37b0fee6e4258e88ddbf60045d93dfd861ef480df9d2a1a0.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 365, + 164, + 403, + 201 + ], + "blocks": [ + { + "bbox": [ + 365, + 164, + 403, + 201 + ], + "lines": [ + { + "bbox": [ + 365, + 164, + 403, + 201 + ], + "spans": [ + { + "bbox": [ + 365, + 164, + 403, + 201 + ], + "type": "image", + "image_path": "0be836b6ebf0f0d9e8e7b9c9a97641604c09306c2c283a077d4b000013008ef7.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 367, + 202, + 479, + 211 + ], + "lines": [ + { + "bbox": [ + 367, + 202, + 479, + 211 + ], + "spans": [ + { + "bbox": [ + 367, + 202, + 479, + 211 + ], + "type": "text", + "content": "46087 07 from Iconfont (N=256); B is ours" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 404, + 164, + 442, + 201 + ], + "blocks": [ + { + "bbox": [ + 404, + 164, + 442, + 201 + ], + "lines": [ + { + "bbox": [ + 404, + 164, + 442, + 201 + ], + "spans": [ + { + "bbox": [ + 404, + 164, + 442, + 201 + ], + "type": "image", + "image_path": "855a590285f2b9350f8761c5bfd3b2ef94d10f4938f3cba4f7b8b28cbaa3893b.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 444, + 164, + 480, + 201 + ], + "blocks": [ + { + "bbox": [ + 444, + 164, + 480, + 201 + ], + "lines": [ + { + "bbox": [ + 444, + 164, + 480, + 201 + ], + "spans": [ + { + "bbox": [ + 444, + 164, + 480, + 201 + ], + "type": "image", + "image_path": "45dc4111100e898d4d44fde5653e43c2d9e836343ca6cd2327df5d240ba7d13c.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 132, + 267, + 237, + 279 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 267, + 237, + 279 + ], + "spans": [ + { + "bbox": [ + 132, + 267, + 237, + 279 + ], + "type": "text", + "content": "4.7 Ablation Study" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 130, + 286, + 480, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 286, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 286, + 480, + 335 + ], + "type": "text", + "content": "To verify the effectiveness of our contributions, we conducted an ablation study, by vectorizing without gradients, without segmentation guidance, or both, as shown in Fig. 13. The similarity between vectorized results and their corresponding inputs is measured by PSNR. Results are tabulated in Tab. 2." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 130, + 346, + 480, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 480, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 480, + 429 + ], + "type": "text", + "content": "Gradients For images with rich gradient effects, such as those in Fluent Emoji, the introduction of gradients significantly improves the outcome. For Noto Emoji, where most images do not have gradients, the improvement is less noticeable. Iconfont is an interesting case; although its paths are filled with solid colors, a form of approximate color transition is achieved through multiple paths with progressively changing colors. Introducing gradients also contributes to a noticeable enhancement." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 130, + 441, + 481, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 441, + 481, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 441, + 481, + 489 + ], + "type": "text", + "content": "Segmentation guidance Compared to solely introducing gradients, incorporating segmentation guidance has proven more effective. As illustrated in Tab. 2, a combination of gradients and segmentation yields superior results on both Noto and Fluent datasets. For the Iconfont dataset, the imitated color transition is" + } + ] + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 136, + 507, + 479, + 622 + ], + "blocks": [ + { + "bbox": [ + 136, + 507, + 479, + 622 + ], + "lines": [ + { + "bbox": [ + 136, + 507, + 479, + 622 + ], + "spans": [ + { + "bbox": [ + 136, + 507, + 479, + 622 + ], + "type": "image", + "image_path": "41e3cc32362c02b3e1f5084e1f7b541cf628154234caa72e27fb7fc3fc3b4db8.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 628, + 480, + 662 + ], + "lines": [ + { + "bbox": [ + 130, + 628, + 480, + 662 + ], + "spans": [ + { + "bbox": [ + 130, + 628, + 480, + 662 + ], + "type": "text", + "content": "Fig. 13: Comparison between four combinations tested in the ablation study. When neither gradients nor segmentation guidance is applied, our method works the same as LIVE (1" + }, + { + "bbox": [ + 130, + 628, + 480, + 662 + ], + "type": "inline_equation", + "content": "^{\\text{st}}" + }, + { + "bbox": [ + 130, + 628, + 480, + 662 + ], + "type": "text", + "content": " row)." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 151, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 151, + 91, + 448, + 102 + ], + "type": "text", + "content": "Segmentation-guided Layer-wise Image Vectorization with Gradient Fills" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": "considered a gradient by the segmentation; thus, with fewer paths, the guided variants achieve a more approximate visual effect through large gradient paths. When more paths are added, using smaller, precise solid color paths yields better results, while the guided vectorization remains competitive in terms of accuracy, and is preferred in the user study." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 190, + 309, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 190, + 309, + 201 + ], + "spans": [ + { + "bbox": [ + 132, + 190, + 309, + 201 + ], + "type": "text", + "content": "4.8 Limitations and Future Work" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 179, + 224, + 218, + 266 + ], + "blocks": [ + { + "bbox": [ + 179, + 224, + 218, + 266 + ], + "lines": [ + { + "bbox": [ + 179, + 224, + 218, + 266 + ], + "spans": [ + { + "bbox": [ + 179, + 224, + 218, + 266 + ], + "type": "image", + "image_path": "32f07586b986c6a14aa440df4a6d4131642643c453afff1022d2b238873584ce.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 179, + 268, + 217, + 308 + ], + "blocks": [ + { + "bbox": [ + 179, + 268, + 217, + 308 + ], + "lines": [ + { + "bbox": [ + 179, + 268, + 217, + 308 + ], + "spans": [ + { + "bbox": [ + 179, + 268, + 217, + 308 + ], + "type": "image", + "image_path": "9c5843823faf84e41ffc9cbb5ead80817a5c02d86dd5fbbc2b4a82e6ba5970a2.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 189, + 311, + 208, + 319 + ], + "lines": [ + { + "bbox": [ + 189, + 311, + 208, + 319 + ], + "spans": [ + { + "bbox": [ + 189, + 311, + 208, + 319 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 223, + 224, + 260, + 262 + ], + "blocks": [ + { + "bbox": [ + 223, + 224, + 260, + 262 + ], + "lines": [ + { + "bbox": [ + 223, + 224, + 260, + 262 + ], + "spans": [ + { + "bbox": [ + 223, + 224, + 260, + 262 + ], + "type": "image", + "image_path": "82501da6cabfd05a24469809b458386f1772e9f29829091ede3ec37dae16f4c0.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 223, + 275, + 260, + 308 + ], + "blocks": [ + { + "bbox": [ + 223, + 275, + 260, + 308 + ], + "lines": [ + { + "bbox": [ + 223, + 275, + 260, + 308 + ], + "spans": [ + { + "bbox": [ + 223, + 275, + 260, + 308 + ], + "type": "image", + "image_path": "527f92b36da004b16a4fbc7d0110e039ec7d47d56cfc8a4ad6e7735d77b980d5.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 233, + 311, + 249, + 319 + ], + "lines": [ + { + "bbox": [ + 233, + 311, + 249, + 319 + ], + "spans": [ + { + "bbox": [ + 233, + 311, + 249, + 319 + ], + "type": "text", + "content": "N=1" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 267, + 224, + 304, + 262 + ], + "blocks": [ + { + "bbox": [ + 267, + 224, + 304, + 262 + ], + "lines": [ + { + "bbox": [ + 267, + 224, + 304, + 262 + ], + "spans": [ + { + "bbox": [ + 267, + 224, + 304, + 262 + ], + "type": "image", + "image_path": "533546488afd40e530fe05a2a042d9d89f45c6161ea9c40f84993ded5fdb9fd1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 267, + 268, + 303, + 308 + ], + "blocks": [ + { + "bbox": [ + 267, + 268, + 303, + 308 + ], + "lines": [ + { + "bbox": [ + 267, + 268, + 303, + 308 + ], + "spans": [ + { + "bbox": [ + 267, + 268, + 303, + 308 + ], + "type": "image", + "image_path": "baa1e215c3da4677aa28b621c516dd267ebb521f1113cdb3a1327c0cf0604152.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 277, + 311, + 293, + 319 + ], + "lines": [ + { + "bbox": [ + 277, + 311, + 293, + 319 + ], + "spans": [ + { + "bbox": [ + 277, + 311, + 293, + 319 + ], + "type": "text", + "content": "N=2" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 331, + 482, + 354 + ], + "lines": [ + { + "bbox": [ + 131, + 331, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 131, + 331, + 482, + 354 + ], + "type": "text", + "content": "Fig. 14: Intricate contours that our simple initial paths fail to converge to. Extra paths are added to mask the overflowing colors." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 310, + 224, + 348, + 262 + ], + "blocks": [ + { + "bbox": [ + 310, + 224, + 348, + 262 + ], + "lines": [ + { + "bbox": [ + 310, + 224, + 348, + 262 + ], + "spans": [ + { + "bbox": [ + 310, + 224, + 348, + 262 + ], + "type": "image", + "image_path": "80d75e55e0ff609700329c8cb544346771599352603365ddb2028f024300e070.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 310, + 268, + 347, + 308 + ], + "blocks": [ + { + "bbox": [ + 310, + 268, + 347, + 308 + ], + "lines": [ + { + "bbox": [ + 310, + 268, + 347, + 308 + ], + "spans": [ + { + "bbox": [ + 310, + 268, + 347, + 308 + ], + "type": "image", + "image_path": "e2b24f2811cf04a777189474e454efeac8350463483da55e28c9dcde2646ac81.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 311, + 336, + 319 + ], + "lines": [ + { + "bbox": [ + 320, + 311, + 336, + 319 + ], + "spans": [ + { + "bbox": [ + 320, + 311, + 336, + 319 + ], + "type": "text", + "content": "N=4" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 353, + 224, + 391, + 266 + ], + "blocks": [ + { + "bbox": [ + 353, + 224, + 391, + 266 + ], + "lines": [ + { + "bbox": [ + 353, + 224, + 391, + 266 + ], + "spans": [ + { + "bbox": [ + 353, + 224, + 391, + 266 + ], + "type": "image", + "image_path": "568d41cb898bdabcb4518222b82af754ecf767b26675c2862f82a6ca3f3ea360.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 354, + 268, + 390, + 308 + ], + "blocks": [ + { + "bbox": [ + 354, + 268, + 390, + 308 + ], + "lines": [ + { + "bbox": [ + 354, + 268, + 390, + 308 + ], + "spans": [ + { + "bbox": [ + 354, + 268, + 390, + 308 + ], + "type": "image", + "image_path": "79b1b7c39580d21f88f52a73667f71cc78f502affc0c82a152b477b3ec12eca1.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 397, + 224, + 435, + 267 + ], + "blocks": [ + { + "bbox": [ + 397, + 224, + 435, + 267 + ], + "lines": [ + { + "bbox": [ + 397, + 224, + 435, + 267 + ], + "spans": [ + { + "bbox": [ + 397, + 224, + 435, + 267 + ], + "type": "image", + "image_path": "700c9d2434b6811f066371519a78e3558d85ba1b96562e86b8eeffa7c76ae85d.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 398, + 268, + 433, + 308 + ], + "blocks": [ + { + "bbox": [ + 391, + 311, + 404, + 319 + ], + "lines": [ + { + "bbox": [ + 391, + 311, + 404, + 319 + ], + "spans": [ + { + "bbox": [ + 391, + 311, + 404, + 319 + ], + "type": "text", + "content": "N=8" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 398, + 268, + 433, + 308 + ], + "lines": [ + { + "bbox": [ + 398, + 268, + 433, + 308 + ], + "spans": [ + { + "bbox": [ + 398, + 268, + 433, + 308 + ], + "type": "image", + "image_path": "88edb6154ccfe6408c8b1e577c6ed5bc99b95994aa3ebd7dab254252af6c35cb.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 377, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 377, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 377, + 482, + 437 + ], + "type": "text", + "content": "While our segmentation guidance effectively captures large-scale features, the initial shape, consisting of a loop formed by four connected Bezier curves, encounters challenges in accurately fitting intricate shapes, as illustrated in Figure 14. Achieving a balance between the number of paths and the number of control points per path needs further deliberation." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 130, + 437, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 510 + ], + "type": "text", + "content": "Additionally, our implementation employs radial gradients with two color stops for simplicity, since a linear gradient can be interpreted as a radial gradient with its center positioned outside the path. Our method is not confined to a specific type of gradient. A more comprehensive implementation could involve dynamically determining gradient types and adjusting the number of color stops accordingly." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 132, + 524, + 220, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 524, + 220, + 537 + ], + "spans": [ + { + "bbox": [ + 132, + 524, + 220, + 537 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we propose a segmentation-guided layer-wise vectorization framework, which synthesizes vector images by progressively adding paths and filling the paths with radial gradients. We design a gradient-aware segmentation method out of traditional algorithms to guide our novel initialization approach and newly designed loss function to address the obstacles in optimizing gradient parameters. We test our method on Noto [9] and Fluent Emoji [21] to show its capability in decomposing raster images into concise layers of gradient-filled paths when the topology is clear, and on Iconfont [13] to demonstrate its effectiveness in case of complex inputs. Our method shows superior performance compared to previous work." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 224, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 224, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 224, + 101 + ], + "type": "text", + "content": "H. Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 408, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 408, + 152 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 408, + 152 + ], + "type": "text", + "content": "This work was supported by the NSFC under Grant 62072271." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 170, + 197, + 182 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 170, + 197, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 170, + 197, + 182 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 194, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 138, + 194, + 480, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 194, + 480, + 217 + ], + "spans": [ + { + "bbox": [ + 138, + 194, + 480, + 217 + ], + "type": "text", + "content": "1. Baksteen, S.D., Hettinga, G.J., Echevarria, J., Kosinka, J.: Mesh colours for gradient meshes. STAG: Smart Tools and Applications in Graphics (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 217, + 480, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 217, + 480, + 249 + ], + "spans": [ + { + "bbox": [ + 138, + 217, + 480, + 249 + ], + "type": "text", + "content": "2. Beucher, S., Meyer, F.: The morphological approach to segmentation: the watershed transformation. Mathematical morphology in image processing 34(1993), 49 (1993)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 250, + 480, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 250, + 480, + 282 + ], + "spans": [ + { + "bbox": [ + 138, + 250, + 480, + 282 + ], + "type": "text", + "content": "3. Carlier, A., Danelljan, M., Alahi, A., Timofte, R.: Deepsvg: A hierarchical generative network for vector graphics animation. Advances in Neural Information Processing Systems 33, 16351-16361 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 283, + 480, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 283, + 480, + 315 + ], + "spans": [ + { + "bbox": [ + 138, + 283, + 480, + 315 + ], + "type": "text", + "content": "4. Comaniciu, D., Meer, P.: Mean shift: A robust approach toward feature space analysis. IEEE Transactions on pattern analysis and machine intelligence 24(5), 603-619 (2002)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 316, + 480, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 316, + 480, + 348 + ], + "spans": [ + { + "bbox": [ + 138, + 316, + 480, + 348 + ], + "type": "text", + "content": "5. Du, Z.J., Kang, L.F., Tan, J., Gingold, Y., Xu, K.: Image vectorization and editing via linear gradient layer decomposition. ACM Transactions on Graphics (TOG) 42(4), 1-13 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 349, + 480, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 349, + 480, + 381 + ], + "spans": [ + { + "bbox": [ + 138, + 349, + 480, + 381 + ], + "type": "text", + "content": "6. Egiazarian, V., Voynov, O., Artemov, A., Volkhonskiy, D., Safin, A., Taktasheva, M., Zorin, D., Burnaev, E.: Deep vectorization of technical drawings. In: European conference on computer vision. pp. 582-598. Springer (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 382, + 480, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 382, + 480, + 414 + ], + "spans": [ + { + "bbox": [ + 138, + 382, + 480, + 414 + ], + "type": "text", + "content": "7. Favreau, J.D., Lafarge, F., Bousseau, A.: Photo2clipart: Image abstraction and vectorization using layered linear gradients. ACM Transactions on Graphics (TOG) 36(6), 1-11 (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 415, + 480, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 415, + 480, + 446 + ], + "spans": [ + { + "bbox": [ + 138, + 415, + 480, + 446 + ], + "type": "text", + "content": "8. Frans, K., Soros, L., Witkowski, O.: Clipdraw: Exploring text-to-drawing synthesis through language-image encoders. Advances in Neural Information Processing Systems 35, 5207-5218 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 447, + 480, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 447, + 480, + 457 + ], + "spans": [ + { + "bbox": [ + 138, + 447, + 480, + 457 + ], + "type": "text", + "content": "9. Noto emoji, https://github.com/googlefonts/sono-emoji Accessed: 2023-09-19" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 458, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 458, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 138, + 458, + 480, + 479 + ], + "type": "text", + "content": "0. Ha, D., Eck, D.: A neural representation of sketch drawings. In: International Conference on Learning Representations (2018)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 479, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 479, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 138, + 479, + 480, + 511 + ], + "type": "text", + "content": "1. Haralick, R.M., Sternberg, S.R., Zhuang, X.: Image analysis using mathematical morphology. IEEE transactions on pattern analysis and machine intelligence (4), 532-550 (1987)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 512, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 512, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 512, + 480, + 533 + ], + "type": "text", + "content": "2. Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural Computation 9(8), 1735-1780 (1997)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 534, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 534, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 534, + 480, + 555 + ], + "type": "text", + "content": "3. Vector illustrations library, https://www iconfont.cn/illustrations/index Accessed: 2023-11-02" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 555, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 480, + 578 + ], + "type": "text", + "content": "4. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. In: International Conference on Learning Representations (2015)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 578, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 480, + 599 + ], + "type": "text", + "content": "5. Kingma, D.P., Welling, M.: Auto-Encoding Variational Bayes. In: International Conference on Learning Representations (2014)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "type": "text", + "content": "6. Li, T.M., Lukáč, M., Gharbi, M., Ragan-Kelley, J.: Differentiable vector graphics rasterization for editing and learning. ACM Transactions on Graphics (TOG) 39(6), 1-15 (2020)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "text", + "content": "7. Liao, Z., Hoppe, H., Forsyth, D., Yu, Y.: A subdivision-based representation for vector image editing. IEEE Transactions on Visualization and Computer Graphics 18(11), 1858-1867 (2012)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 151, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 151, + 91, + 447, + 102 + ], + "type": "text", + "content": "Segmentation-guided Layer-wise Image Vectorization with Gradient Fills" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "text", + "content": "18. Liu, Y.T., Zhang, Z., Guo, Y.C., Fisher, M., Wang, Z., Zhang, S.H.: Dualvector: Unsupervised vector font synthesis with dual-part representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14193-14202 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 160, + 481, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 160, + 481, + 203 + ], + "spans": [ + { + "bbox": [ + 133, + 160, + 481, + 203 + ], + "type": "text", + "content": "19. Lopes, R.G., Ha, D., Eck, D., Shlens, J.: A Learned Representation for Scalable Vector Graphics. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7929-7938. IEEE, Seoul, Korea (South) (Oct 2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 203, + 481, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 203, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 481, + 236 + ], + "type": "text", + "content": "20. Ma, X., Zhou, Y., Xu, X., Sun, B., Filev, V., Orlov, N., Fu, Y., Shi, H.: Towards layer-wise image vectorization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16314-16323 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 236, + 481, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 481, + 256 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 481, + 256 + ], + "type": "text", + "content": "21. Fluent emoji, https://github.com/microsoft/fluentui-emoji Accessed: 2023-10-27" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 257, + 481, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 257, + 481, + 290 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 481, + 290 + ], + "type": "text", + "content": "22. Orzan, A., Bousseau, A., Winnemöller, H., Barla, P., Thollot, J., Salesin, D.: Diffusion curves: a vector representation for smooth-shaded images. ACM Transactions on Graphics (TOG) 27(3), 1-8 (2008)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 290, + 481, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 290, + 481, + 311 + ], + "spans": [ + { + "bbox": [ + 132, + 290, + 481, + 311 + ], + "type": "text", + "content": "23. Otsu, N.: A threshold selection method from gray-level histograms. IEEE transactions on systems, man, and cybernetics 9(1), 62-66 (1979)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 311, + 481, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 311, + 481, + 343 + ], + "spans": [ + { + "bbox": [ + 132, + 311, + 481, + 343 + ], + "type": "text", + "content": "24. Reddy, P., Gharbi, M., Lukac, M., Mitra, N.J.: Im2vec: Synthesizing vector graphics without vector supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7342-7351 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 343, + 481, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 481, + 374 + ], + "type": "text", + "content": "25. Reddy, P., Zhang, Z., Wang, Z., Fisher, M., Jin, H., Mitra, N.: A multi-implicit neural representation for fonts. Advances in Neural Information Processing Systems 34, 12637-12647 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "type": "text", + "content": "26. Richardt, C., Lopez-Moreno, J., Bousseau, A., Agrawala, M., Drettakis, G.: Vectorising bitmaps into semi-transparent gradient layers. Computer Graphics Forum (Proceedings of EGSR) 33(4), 11-19 (July 2014)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 407, + 481, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 481, + 440 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 481, + 440 + ], + "type": "text", + "content": "27. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 440, + 481, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 440, + 481, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 440, + 481, + 472 + ], + "type": "text", + "content": "28. Su, H., Liu, X., Niu, J., Cui, J., Wan, J., Wu, X., Wang, N.: Marvel: Raster gray-level manga vectorization via primitive-wise deep reinforcement learning. IEEE Transactions on Circuits and Systems for Video Technology (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 472, + 481, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 481, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 481, + 493 + ], + "type": "text", + "content": "29. Sun, J., Liang, L., Wen, F., Shum, H.Y.: Image vectorization using optimized gradient meshes. ACM Transactions on Graphics (TOG) 26(3), 11-es (2007)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 493, + 481, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 493, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 493, + 481, + 525 + ], + "type": "text", + "content": "30. Tian, X., Günther, T.: A survey of smooth vector graphics: Recent advances in representation, creation, rasterization and image vectorization. IEEE Transactions on Visualization and Computer Graphics (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 525, + 481, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 481, + 557 + ], + "type": "text", + "content": "31. Van der Walt, S., Schonberger, J.L., Nunez-Iglesias, J., Boulogne, F., Warner, J.D., Yager, N., Gouillart, E., Yu, T.: scikit-image: image processing in python. PeerJ 2, e453 (2014)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 558, + 481, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 558, + 481, + 579 + ], + "spans": [ + { + "bbox": [ + 132, + 558, + 481, + 579 + ], + "type": "text", + "content": "32. Wang, Y., Lian, Z.: Deepvecfont: Synthesizing high-quality vector fonts via dual-modality learning. ACM Transactions on Graphics (TOG) 40(6), 1-15 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 579, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 579, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 579, + 481, + 601 + ], + "type": "text", + "content": "33. Xia, T., Liao, B., Yu, Y.: Patch-based image vectorization with automatic curvilinear feature alignment. ACM Transactions on Graphics (TOG) 28(5), 1-10 (2009)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 601, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 481, + 632 + ], + "type": "text", + "content": "34. Xie, G., Sun, X., Tong, X., Nowrouzezahrai, D.: Hierarchical diffusion curves for accurate automatic image vectorization. ACM Transactions on Graphics (TOG) 33(6), 1-11 (2014)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "35. Yang, M., Chao, H., Zhang, C., Guo, J., Yuan, L., Sun, J.: Effective clipart image vectorization through direct optimization of bezigons. IEEE transactions on visualization and computer graphics 22(2), 1063-1075 (2015)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "H. Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 140 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 140 + ], + "type": "text", + "content": "36. Zhu, H., Cao, J., Xiao, Y., Chen, Z., Zhong, Z., Zhang, Y.J.: Tcb-spline-based image vectorization. ACM Transactions on Graphics (TOG) 41(3), 1-17 (2022)" + } + ] + } + ], + "index": 2 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 151, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 151, + 91, + 447, + 102 + ], + "type": "text", + "content": "Segmentation-guided Layer-wise Image Vectorization with Gradient Fills" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/SeiT++_ Masked Token Modeling Improves Storage-efficient Training/a6a99e76-474a-47eb-80d4-d9020aeb146a_content_list.json b/2024/SeiT++_ Masked Token Modeling Improves Storage-efficient Training/a6a99e76-474a-47eb-80d4-d9020aeb146a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5817215ac1e0a6c716b3eb9bcc8071131aee9a86 --- /dev/null +++ b/2024/SeiT++_ Masked Token Modeling Improves Storage-efficient Training/a6a99e76-474a-47eb-80d4-d9020aeb146a_content_list.json @@ -0,0 +1,1707 @@ +[ + { + "type": "text", + "text": "SeiT++: Masked Token Modeling Improves Storage-efficient Training", + "text_level": 1, + "bbox": [ + 259, + 140, + 743, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Minhyun Lee $^{1,\\dagger}$ Song Park $^{2,\\dagger}$ Byeongho Heo $^{2}$ Dongyoon Han $^{2}$ Hyunjung Shim $^{3,\\star}$", + "bbox": [ + 230, + 212, + 772, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Yonsei University $^{2}$ Naver AI Lab $^{3}$ KAIST AI $^{4}$ Equal contribution. Work done at Naver AI Lab.", + "bbox": [ + 300, + 250, + 700, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Recent advancements in Deep Neural Network (DNN) models have significantly improved performance across computer vision tasks. However, achieving highly generalizable and high-performing vision models requires expansive datasets, resulting in significant storage requirements. This storage challenge is a critical bottleneck for scaling up models. A recent breakthrough by SeiT proposed the use of Vector-Quantized (VQ) feature vectors (i.e., tokens) as network inputs for vision classification. This approach achieved $90\\%$ of the performance of a model trained on full-pixel images with only $1\\%$ of the storage. While SeiT needs labeled data, its potential in scenarios beyond fully supervised learning remains largely untapped. In this paper, we extend SeiT by integrating Masked Token Modeling (MTM) for self-supervised pre-training. Recognizing that self-supervised approaches often demand more data due to the lack of labels, we introduce TokenAdapt and ColorAdapt. These methods facilitate comprehensive token-friendly data augmentation, effectively addressing the increased data requirements of self-supervised learning. We evaluate our approach across various scenarios, including storage-efficient ImageNet-1k classification, fine-grained classification, ADE-20k semantic segmentation, and robustness benchmarks. Experimental results demonstrate consistent performance improvement in diverse experiments, validating the effectiveness of our method. Code is available at https://github.com/naver-ai/seit.", + "bbox": [ + 261, + 323, + 743, + 628 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 657, + 375, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advancements in Deep Neural Network (DNN) [17,56] have significantly improved their performance across various computer vision tasks. However, achieving highly generalizable and high-performing vision models demands an extensive dataset, often containing billions of data points [28,41,53] with large storage requirements. For example, the LAION-5B [50] dataset requires a storage size of 240TB. This storage challenge emerges as a critical bottleneck for scaling up vision models. Consequently, optimizing storage efficiency becomes essential for the practical advancement of scaled-up vision models.", + "bbox": [ + 212, + 691, + 794, + 815 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Hyunjung Shim is a corresponding author.", + "bbox": [ + 217, + 824, + 517, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/048273a0b26487a7811143950997cf5c723f7911823c28f1746567b200c0ae27.jpg", + "image_caption": [ + "Fig.1: Over $70\\%$ top-1 accuracy is achievable with just 1GB data. On ImageNet-1k, we visualize the trade-off of training data storage vs. top-1 accuracy using the fixed ViT-B/16 for a controlled comparison. Each accuracy metric is individually trained with different data types. While the entire ImageNet-1k dataset requires approximately 140GB for training with images, our approach demonstrates significant storage efficiency over competitors." + ], + "image_footnote": [], + "bbox": [ + 274, + 143, + 727, + 289 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Images require more storage than texts due to their intrinsic complexity in converting continuous natural light signals into discrete pixels for computer representation. However, a considerable portion of image data corresponds to redundant or unnecessary details, such as repeated patterns or imperceptible details [47]. These properties lead to large storage requirements and also induce negative impacts on visual recognition tasks as mentioned in [1,11,21,22,40,51]. Therefore, several attempts [9, 29, 45, 69, 70] have been made to reduce storage requirements by optimizing the efficiency of image data. A standard approach to reduce storage requirements is to decrease the total number of training data by removing less important samples [9, 29, 45] or condensing the data into smaller synthetic sets [69, 70].", + "bbox": [ + 212, + 415, + 787, + 582 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, as SeiT [44] illustrates, these approaches may incur a significant performance drop or become infeasible for large-scale datasets due to their complexity. Alternatively, some methods aim to reduce the size of each image through compression techniques or resolution adjustments [50]. However, they still face a considerable performance gap with the original dataset. Recently, SeiT [44] has made a significant breakthrough in this area by enabling the direct learning of ViTs [17] from discrete representations (i.e., tokens) pre-extracted by an offline tokenizer [66]. By using vector-quantized features instead of traditional pixel images, SeiT achieves $90\\%$ of the performance of a full storage model while requiring only $1\\%$ of the storage.", + "bbox": [ + 212, + 582, + 787, + 734 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While SeiT [44] has demonstrated the high potential of token-based training, a performance gap with the image-based models still persists, and a prospective solution - exploration beyond fully supervised learning scenarios - remains untapped. To leverage the advantages of large-scale data beyond human-annotated datasets, exploring label-free learning methods is imperative. In this paper, we investigate the compatibility of SeiT with self-supervised learning approaches in token-based frameworks. Among many different self-supervised methods [6,7,23],", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Lee and Park et al.", + "bbox": [ + 271, + 114, + 400, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Masked Image Modeling (MIM) [24,64] derived from Masked Language Modeling (MLM) [15] has brought significant success and has quickly become a popular approach for learning visual representations.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inspired by this, we introduce Masked Token Modeling (MTM), a pre-training method that performs token masking and reconstruction in token-based frameworks, and validate its effectiveness for storage efficiency on various benchmarks. To the best of our knowledge, this is the first approach for learning an MLM-based model directly from offline tokens, offering the potential to optimize storage efficiency for large-scale training tasks. As previous studies [12, 13, 56] have highlighted, data augmentation plays a crucial role in large-scale model training and is a fundamental element in various training setups [6, 7, 23, 24, 37, 56, 62, 64]. However, applying existing data augmentation techniques [13, 67, 71] designed for pixel images to tokens poses a significant challenge due to the input domain shift. Therefore, addressing this challenge is key to overcoming major performance bottlenecks and improving the storage efficiency of token-based frameworks.", + "bbox": [ + 212, + 200, + 787, + 383 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To this end, we introduce TokenAdapt and ColorAdapt, two novel augmentation strategies designed for token-based training. TokenAdapt facilitates the integration of traditional pixel-based augmentation policies by transforming token embeddings into a feature space compatible with pixel-based augmentations. After applying augmentation to the transformed token embeddings, the augmented token embeddings are then reverted to their original token space. ColorAdapt, inspired by Adaptive Instance Normalization (AdaIN [27]), alters the color attributes of token embeddings while preserving object structure by adjusting their statistics. By integrating TokenAdapt and ColorAdapt into our token augmentation process, we enhance the effectiveness of token-based training across both fully supervised and unsupervised learning scenarios, demonstrating significant improvements in model performance.", + "bbox": [ + 212, + 392, + 787, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our SeiT++ framework combines Masked Token Modeling (MTM) with the proposed token augmentation strategies, TokenAdapt and ColorAdapt. We validate the effectiveness of SeiT++ on various scenarios: (1) storage-efficient ImageNet-1k classification, (2) transfer learning on fine-grained classification, (3) ADE-20k semantic segmentation, and (4) evaluation on robustness benchmarks. As shown in Figure 1, SeiT++ consistently outperforms comparison methods in storage-efficient ImageNet-1k classification with limited storage sizes, achieving a top-1 accuracy of $77.8\\%$ with only 1.4GB of data. These results confirm that token-based training can take advantage of self-supervised learning to improve storage efficiency. We also demonstrate that our proposed token augmentations significantly contribute to storage efficiency by augmenting data effectively. We reveal that our token-based augmentation methods basically yield significant boosts in supervised learning and further amplify improvements in MTM learning. Furthermore, we present how our strategies enhance the generalizability of trained models, as evidenced by our experimental results on fine-grained classification and robustness evaluation. Lastly, the applicability of our method to an alternative input format (e.g., DCT coefficients) demonstrates its extensibility.", + "bbox": [ + 212, + 583, + 795, + 843 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "SeiT++: Masked Token Modeling Improves Storage-efficient Training", + "bbox": [ + 276, + 114, + 730, + 130 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 217, + 143, + 387, + 161 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Learning with Tokenization. The use of discrete tokens for training vision models has recently gained attention, particularly in self-supervised representation learning. BeiT [2] employs the Masked Image Modeling (MIM) framework to recover discrete tokens from masked images and CIM [19] utilizes tokens in Corrupted Image Modeling. MAGE [34] demonstrates the incorporation of tokens from VQGAN [18] for both generative training and representation learning within a unified training framework. However, existing studies mainly focus on self-supervised representation learning and rely on online tokenization, converting images to tokens during each iteration of model training. This approach incurs notable memory consumption and computational costs, limiting the full exploitation of the storage efficiency of discrete tokens.", + "bbox": [ + 217, + 175, + 785, + 340 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Storage-efficient Vision Training. Efficient model learning is a crucial research area, with studies focusing on optimizing target models [35, 36] or minimizing training data and labels [39, 42]. Storage-efficient training aims to reduce the dataset storage size. To this end, dataset distillation (DD) [60] creates a condensed dataset by leveraging the knowledge of the original dataset. Recent research [33, 48, 49, 69, 70] highlights the effectiveness of dataset distillation in improving training efficiency. However, its high computational complexity poses challenges, particularly when applied to large-scale datasets like ImageNet-1k [14]. Sampling-based methods [9, 29, 45] address this by selecting a subset of the most representative samples from the full dataset. Paul et al. [45] identify important samples very early in training. Jiang et al. [29] utilize a consistency score based on statistics collected during training. However, the diverse sample selection is not guaranteed, particularly in low data regimes, leading to suboptimal performance compared to dataset distillation (DD) methods [70]. To tackle these issues, SeiT [44] proposes a storage-efficient vision training framework by leveraging the storage efficiency of tokens. They utilize the ViT-VQGAN [66] tokenizer to extract tokens from images and store them for vision model training. As a result, they outperform other storage-efficient training methods, significantly reducing the dataset storage size.", + "bbox": [ + 217, + 363, + 785, + 651 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Data Augmentation. Data augmentation serves as a fundamental building block to improve the model performance across diverse computer vision tasks. Various data augmentation techniques have been employed for improved performance [16, 71] or model robustness [20, 38]. Mixup [68] introduces convex combinations of pairs of images and their labels to create a smoother decision boundary. CutMix [67] replaces certain regions of an image with a patch from another image, enhancing the model's localization ability. AutoAugment [12] leverages reinforcement learning to automate optimal augmentation policy discovery. RandAugment [13] introduces a potent policy based on random combinations of different data augmentation methods. DeiT [56] emphasizes the crucial role of data augmentation in vision transformers for improved model performance.", + "bbox": [ + 217, + 672, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 230, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Lee and Park et al.", + "bbox": [ + 271, + 114, + 400, + 126 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/93f70da2a7709e419b589aaf11d5588d6735a9a08eb5ccbdab7920fa8989e29c.jpg", + "image_caption": [ + "Fig. 2: Masked Token Modeling (MTM) pipeline. MTM is a self-supervised learning approach in token-based frameworks. (a) The tokenized dataset is saved on storage before model training. Then, (b) using only the pre-stored tokens, a storage-efficient vision model (MTM) is trained without relying on labeled datasets." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 782, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Previous studies extensively explore the effective data augmentation policies for pixel-based vision model training. However, we observe that pixel-based data augmentations are not well-suited for tokens, posing a critical performance bottleneck. Motivated by this, we tackle the data augmentation challenge, which has not been explored in previous studies for vision model training with tokens.", + "bbox": [ + 212, + 369, + 787, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 469, + 330, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Preliminary: Storage-efficient Vision Training (SeiT)", + "text_level": 1, + "bbox": [ + 215, + 502, + 694, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SeiT [44] aims to develop a scalable and high-performance vision classifier with minimal storage requirements. It introduces a storage-efficient vision training framework that enables the direct learning of ViT [17] from discrete representations (i.e., tokens) pre-extracted by an offline tokenizer. In addition, SeiT further improves training by incorporating two token-specific augmentations: Token-EDA, drawing inspiration from easy data augmentation for language [61], and Emb-Noise, which introduces Gaussian noise into discrete representations. By using vector-quantized features instead of pixel images, SeiT retains $90\\%$ of the precision of a model trained on full-pixel images while using only $1\\%$ of the storage space. SeiT consists of two parts: (1) preparing the token dataset and (2) training a model using the token dataset.", + "bbox": [ + 212, + 527, + 787, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Preparing the Token Dataset. SeiT extracts tokens using the ImageNet-1k-trained ViT-VQGAN tokenizer [66] to prepare the token dataset for training. As shown in Figure 2(a), each pixel image $I$ is transformed into a sequence of tokens $T = [t_1, t_2, \\dots, t_n]$ . Since the tokens are stored in quantized form, each token means an embedding vector in a codebook $\\mathcal{Z} = \\{z_k \\in \\mathbb{R}^d | k = 1, \\dots, K\\}$ . i.e., $t_i \\in \\mathbb{Q}$ where $\\mathbb{Q} = \\{1, 2, \\dots, K\\}$ for a codebook with size $K$ . Here, given the tokens $T$ and the codebook $\\mathcal{Z}$ , the corresponding token embeddings $Z_T = [z_{t_1}, z_{t_2}, \\dots, z_{t_n}]$ can be computed by lookup process. This conversion compresses each image", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "SeiT++: Masked Token Modeling Improves Storage-efficient Training", + "bbox": [ + 276, + 114, + 732, + 130 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$I \\in \\mathbb{R}^{H \\times W \\times 3}$ into a more compact token sequence $T \\in \\mathbb{R}^n$ , where $n$ is much smaller than $H \\times W \\times 3$ , thus significantly reducing storage requirements. For example, storing the full-pixel ImageNet-1k requires 140GB, but its tokenized form needs only 1.4GB. Notably, the tokens are stored on storage before model training with the codebook $\\mathcal{Z}$ . Thus, during the training phase, the pre-stored tokens are used directly as input instead of the original pixel images.", + "bbox": [ + 212, + 145, + 787, + 238 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training Pipeline. To train a vision classifier using tokens, SeiT initially loads the token dataset and applies Token-EDA to the tokens. The tokens are then converted to a one-hot format and randomly resized and cropped. Subsequently, the processed one-hot tokens are transformed into token embeddings using a pretrained codebook. Techniques such as CutMix and Embedding-noise are applied to these token embeddings, and the resulting augmented tensor is fed into the model for training. Through the proposed framework, SeiT demonstrates the high potential of token-based training, but it lacks exploration beyond fully supervised learning scenarios. In addition, while token-specific augmentation is proposed, the types of token augmentation are limited, and the analysis of data augmentation compared to pixel-based training is insufficient.", + "bbox": [ + 212, + 266, + 789, + 434 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Masked Token Modeling", + "text_level": 1, + "bbox": [ + 215, + 462, + 465, + 478 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we introduce Masked Token Modeling (MTM), a pre-training method in token-based frameworks without relying on labeled datasets. MTM shares a similar concept to Masked Language Modeling (MLM) [15], which works by reconstructing original tokens from their masked versions, employing an autoencoding strategy. Figure 2(b) presents an overview of how our MTM operates.", + "bbox": [ + 212, + 493, + 787, + 571 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Masking. We employ a variable masking ratio that follows a truncated normal distribution, as investigated in previous studies [4,5,34]. The tokens are randomly dropped according to the sampled masking ratio, and then the visible, unmasked token embeddings are fed into the model. By processing only a subset of token embeddings in the encoder, we significantly reduce the overall pre-training time and memory consumption, which aligns with the findings of MAE [24].", + "bbox": [ + 212, + 598, + 787, + 691 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MTM Encoder & Decoder. We use ViT [17] as our encoder because it is able to process both the visible token embeddings and the original token embeddings, seamlessly accommodating their different lengths. During training, the encoder takes the visible token embeddings and outputs latent features that become the decoder input. Before decoding, the latent features are padded to the length of the original token embeddings using mask tokens M. Each mask token is a shared, learned vector indicating the presence of a missing token to be predicted. The decoder then estimates the original tokens $T$ from the padded features.", + "bbox": [ + 212, + 718, + 787, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Lee and Park et al.", + "bbox": [ + 271, + 114, + 400, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bfff7f2568ee3f97bd988c21ad6ac5254aa5d362922dfaad3a67390dedcc233a.jpg", + "image_caption": [ + "(a) Failure cases visualization" + ], + "image_footnote": [], + "bbox": [ + 222, + 148, + 303, + 281 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/37f0cc45980f6f5ece0cb7f7ce5a3e901c84c72294edd0d97c36cc4fba1c2563.jpg", + "image_caption": [ + "Fig. 3: Data augmentations with tokens. Each ViT-VQGAN decoded image is reconstructed from a given RGB image after undergoing specific data augmentations. We observe that naively adopting these methods results in incorrect tokenization: 1) Token w/ hFlip demonstrates spatial information collapse during tokenization; 2) Token w/ RRC shows interdependence between neighboring token embeddings. We note the reconstructed images fail to preserve the images' details, suggesting that this incurs ineffectiveness of tokenization. Furthermore, we report top-1 accuracies (ViT-B) on ImageNet-1k according to the augmentations applied during training. All the augmentations are de-facto default training setups for vision transformers [37,54,56]." + ], + "image_footnote": [], + "bbox": [ + 310, + 148, + 390, + 281 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f34aa1f77be4ef74f7d0b9bad29499ebe5a9428bb29c4b47f7478758d29a519f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 393, + 148, + 473, + 281 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/371a44b8363fc6f0990aa4993e9f607dd47d1a289ae80827c36e3443b9448f1a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 477, + 148, + 555, + 281 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/de59c1cd87ba0ee0c3bff7f9980f770a286ed455da551754ed64ea2337fb05a3.jpg", + "image_caption": [ + "(b) Performance degradation" + ], + "image_footnote": [], + "bbox": [ + 570, + 145, + 767, + 282 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Training Objective. MTM learns to accurately predict the original tokens of each masked token from the visible tokens. In particular, we use a cross-entropy loss between the decoder output and the ground truth tokens as the training objective, which is denoted by $\\mathcal{L}_{\\text{recon}}$ :", + "bbox": [ + 212, + 465, + 787, + 527 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e c o n}} = \\mathrm {C E} \\left(T _ {M} ^ {\\prime}, T _ {M}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 416, + 540, + 784, + 558 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $T_M'$ denotes the predictions of the model for the masked tokens, and $T_M$ corresponds to the ground truth tokens for $T_M'$ . Following MAE [24], this loss is optimized only for the masked tokens.", + "bbox": [ + 212, + 561, + 787, + 608 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Data Augmentation for Tokens", + "text_level": 1, + "bbox": [ + 214, + 628, + 517, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "While Masked Token Modeling (MTM) improves the storage efficiency of the token-based framework, there is still a performance bottleneck in improving token-based training. Since self-supervised approaches often require more data due to the lack of labels, exploring data augmentation is important to take full advantage of MTM. In this section, we analyze why existing data augmentations designed for pixel images are ineffective in the token domain. We then present TokenAdapt and ColorAdapt as solutions for token-based data augmentation.", + "bbox": [ + 212, + 652, + 787, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Limited Data Augmentations for Tokens. We explore the challenges of pixel-based augmentation in the token domain and identify two main reasons: (1) spatial information collapse during tokenization and (2) inter-token dependencies. The discrete tokenization process condenses a $n \\times n$ 2D image patch", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "SeiT++: Masked Token Modeling Improves Storage-efficient Training", + "bbox": [ + 276, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d9b09dd8744d67ef668a6f82106c8a5a3c91181751e91e7fb32529c4fbd1f02b.jpg", + "image_caption": [ + "Fig. 4: TokenAdapt processing pipeline. TokenAdapt aims to enhance the compatibility of token embeddings with pixel-based data augmentations by converting them into augmentation-compatible space, applying augmentations, and reverting them back to the original token embedding space." + ], + "image_footnote": [], + "bbox": [ + 259, + 143, + 743, + 227 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "into a single $d$ -dimensional 1D vector (i.e., token embedding), causing a collapse of spatial information. This collapse hinders augmentations that rely on spatial details (e.g., horizontal flip). Furthermore, unlike independent image pixels, token embeddings are inter-correlated. This complicates augmentations involving interpolation (e.g., Resize or Mixup [68]), by introducing undesired artifacts.", + "bbox": [ + 212, + 327, + 782, + 402 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 3 shows the limitations of pixel-based augmentations in the token domain. A horizontal Flip (hFlip) distorts both the diagonal line (top row) and the cat's silhouette (bottom row) due to spatial information collapse. This occurs because only the inter-token spatial relationship is flipped, while the inner-token spatial arrangement remains unchanged. Augmentations involving interpolation can lead to undesired artifacts, such as color leakage in the case of Random Resized Crop (RRC), due to interdependence between neighboring embeddings. Therefore, we suspect that the direct application of pixel-based augmentations to token embedding may negatively affect token-based training.", + "bbox": [ + 212, + 404, + 787, + 541 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "TokenAdapt. The concept of TokenAdapt is to establish a feature space that allows for conventional pixel-based augmentations while facilitating the conversion and inversion of pre-extracted tokens with minimal computational overhead. This approach offers several advantages for our TokenAdapt. Firstly, it enables the direct reuse of well-designed pixel-based data augmentation policies. Additionally, the augmented token can be easily computed through the conversion-augmentation-reverse process. Finally, it ensures that the feature spaces for both the original and augmented tokens are preserved identically. This allows for the seamless use of existing model architectures or pre-trained weights.", + "bbox": [ + 212, + 566, + 787, + 702 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The goal of TokenAdapt is to design an augmentation function that converts the input tokens $T$ to augmented tokens $T^{\\mathbf{A}} = [t_1^{\\mathbf{A}}, t_2^{\\mathbf{A}}, \\dots, t_n^{\\mathbf{A}}]$ . Note that the augmented tokens should also be quantized $t_i^{\\mathbf{A}} \\in \\mathbb{Q}$ . Figure 4 depicts the overall pipeline of TokenAdapt. The initial step of TokenAdapt is transforming token embeddings $Z_T$ , computed from $T$ and codebook $\\mathcal{Z}$ , into a representation $S_T \\in \\mathbb{R}^{h \\times w \\times d}$ in augmentation-compatible space $S$ . This transformation is achieved using a conversion function $f$ , i.e., $S_T = f(Z_T)$ . We apply augmentation $\\mathbf{A}$ at augmentation-compatible space $S$ and build an augmented representation $S_T^{\\mathbf{A}} = \\mathbf{A}(S_T)$ . Then, the augmented representation is reverted to", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Lee and Park et al.", + "bbox": [ + 271, + 114, + 400, + 126 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "token embedding space by a reverse transformation $g$ , $Z_T^{\\mathbf{A}} = g(S_T^{\\mathbf{A}})$ . At last, the augmented tokens in embedding space $Z_T^{\\mathbf{A}}$ are quantized to codebook index as $T^{\\mathbf{A}} = \\mathbf{q}_{\\mathcal{Z}}(Z_T^{\\mathbf{A}})$ , where $\\mathbf{q}_{\\mathcal{Z}}$ indicates the vector quantization process. In summary, the augmentation process of TokenAdapt is as follows:", + "bbox": [ + 214, + 145, + 782, + 205 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nZ _ {T} ^ {\\mathbf {A}} = g (\\mathbf {A} (f (Z _ {T}))), \\quad T ^ {\\mathbf {A}} = \\mathbf {q} _ {\\mathcal {Z}} \\left(Z _ {T} ^ {\\mathbf {A}}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 220, + 784, + 239 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The conversion function $f$ and the reverse transformation $g$ are learned from paired data of tokens $(T_x, T_{\\mathbf{A}(x)})$ , where $T_x$ and $T_{\\mathbf{A}(x)}$ denote tokens from an image $x$ and its augmented image $\\mathbf{A}(x)$ , respectively. We use the cross-entropy loss between $Z_{T_x}^{\\mathbf{A}}$ and $Z_{T_{\\mathbf{A}(x)}}$ as our objective function $\\mathcal{L}_{\\mathrm{TA}}$ for learning our TokenAdapt module:", + "bbox": [ + 214, + 244, + 782, + 319 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {T A}} = \\operatorname {C E} \\left(Z _ {T _ {x}} ^ {\\mathbf {A}}, Z _ {T _ {\\mathbf {A} (x)}}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 416, + 335, + 784, + 353 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "TokenAdapt deals with geometric data augmentations (e.g., flip, resize, crop, affine) and Mixup [68]. These augmentations are commonly used in various training methods, but they can cause a performance drop in token-based training, as shown in Figure 3(b).", + "bbox": [ + 212, + 359, + 782, + 417 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Once trained, the TokenAdapt module is used to augment a sequence of tokens $T$ into an augmented sequence $T^{\\mathbf{A}}$ across different datasets and downstream tasks. Importantly, the augmentation-compatible feature space, which does not rely on specific semantic knowledge or alignment with downstream tasks, makes it dataset and task-agnostic. The inherent flexibility of TokenAdapt makes it highly generalizable, allowing direct application to other datasets without the need for additional training procedures.", + "bbox": [ + 212, + 419, + 784, + 525 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "ColorAdapt. As witnessed in Cubuk et al. [12], color-based augmentations are often more effective than geometric ones in certain datasets and are an important component of data augmentation strategies. For example, color-based augmentations are useful for dealing with various environmental changes caused by different lighting conditions (e.g., day and night) or weather conditions (e.g., snow). However, integrating them into TokenAdapt is challenging because most color-based augmentations are designed to work only in the pixel domain. To address this, we propose ColorAdapt, a color-based token augmentation inspired by the method [27]. Formally, the ColorAdapt function $\\mathcal{C}$ is defined as:", + "bbox": [ + 214, + 547, + 782, + 683 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {C} \\left(Z _ {T _ {1}}, Z _ {T _ {2}}\\right) = \\sigma \\left(Z _ {T _ {2}}\\right) \\left(\\frac {Z _ {T _ {1}} - \\mu \\left(Z _ {T _ {1}}\\right)}{\\sigma \\left(Z _ {T _ {1}}\\right)}\\right) + \\mu \\left(Z _ {T _ {2}}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 696, + 784, + 728 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\mu(Z_T)$ and $\\sigma(Z_T)$ are the mean and standard deviation of each token embedding $Z_T$ across its spatial dimension, computed for each channel.", + "bbox": [ + 212, + 733, + 782, + 763 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "ColorAdapt mimics color-changing augmentations in the pixel domain while preserving the encoded object structure by transforming the global statistics within each channel of the token embedding. Figure 5 shows the images decoded from various color-augmented tokens. This illustrates how ColorAdapt affects the visual properties of the augmented tokens. By incorporating ColorAdapt", + "bbox": [ + 212, + 763, + 784, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "SeiT++: Masked Token Modeling Improves Storage-efficient Training", + "bbox": [ + 274, + 114, + 732, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/4508aa871791343080e71200dba7cddbff76462a5c0cfec5d0e51c2e3adcc971.jpg", + "image_caption": [ + "Input data" + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 290, + 200 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/c85dc23a13cec7b64d1d81d12f85aabd7a074eebf84fc4511f8d8e097d465407.jpg", + "image_caption": [ + "Brightness" + ], + "image_footnote": [], + "bbox": [ + 292, + 145, + 359, + 200 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/0fa165838fc446140e9278ad22cdc9f518d5854db63f031d700b19e6e2f663b6.jpg", + "image_caption": [ + "Emb-Noise", + "Fig. 5: ColorAdapt provides more reasonable color changes. We present ViT-VQGAN decoded images to verify the quality of tokenizations after color changes. We use the brightness function with a factor of 0.2 following the implementation [46]. Emb-Noise [44] is the color-based token augmentation. Notably, our ColorAdapt effectively preserves object structure in contrast to the failure of the counterparts." + ], + "image_footnote": [], + "bbox": [ + 359, + 145, + 426, + 200 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/27d01be32f7de5336c7f944d86429cd3033221d1c3f83fbe4d3bc5a07dc46a6c.jpg", + "image_caption": [ + "ColorAdapt" + ], + "image_footnote": [], + "bbox": [ + 428, + 145, + 501, + 200 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/713fa19fa1f0869bdc8ae3bea4ea6c0f65b67ac67890e165cef9cfa69459ea4c.jpg", + "image_caption": [ + "Input data" + ], + "image_footnote": [], + "bbox": [ + 501, + 145, + 575, + 200 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/89c295adbe5f70ab45ac7762e93e29835192b012fc0b4ef63da1f2f6039e30dc.jpg", + "image_caption": [ + "Brightness" + ], + "image_footnote": [], + "bbox": [ + 576, + 145, + 643, + 200 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/afac642f0f9ea9ed6a1849c22bda5a2a73eef4b6b99b8eeec05ee8df522fd560.jpg", + "image_caption": [ + "Emb-Noise" + ], + "image_footnote": [], + "bbox": [ + 645, + 145, + 707, + 200 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/26bde3657dd73f534634fef3fe4a0463330841d217d7c65981b2f4c66f8b696e.jpg", + "image_caption": [ + "ColorAdapt" + ], + "image_footnote": [], + "bbox": [ + 712, + 145, + 785, + 200 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/b529451ec4ec4543b14dc4c12a79edfa626422d0c2437fc791edc2804a150a16.jpg", + "table_caption": [ + "Table 1: Storage-efficient ImageNet-1k classification. We report top-1 accuracies (ViT-B) and compression ratio on ImageNet-1k using various data storage reduction methods. Our SeiT++ shows significant storage efficiency compared to competitors." + ], + "table_footnote": [], + "table_body": "
MethodInput # of images Top 1 Acc.Dataset storage size
Full pixelsImage1,281 k81.8140.0 GB
Uniform random sampling512 k74.054.6 GB (39%)
C-score [29] based samplingImage512 k73.353.3 GB (38%)
Adjusting JPEG quality factor to 51,281 k74.611.0 GB (8%)
Adjusting image resolution (× 0.2)1,281 k75.29.6 GB (7%)
SRe2L [65]Image200 k65.92.6 GB (2%)
G-VBSM [52]100 k63.71.6 GB (1%)
SeiT [44]Token1,281 k74.01.4 GB (1%)
SeiT++1,281 k77.81.4 GB (1%)
", + "bbox": [ + 232, + 369, + 767, + 547 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "into our token augmentation strategy, we improve the data diversity, exploring new possibilities for training more robust and generalizable vision models.", + "bbox": [ + 212, + 579, + 787, + 609 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiment", + "text_level": 1, + "bbox": [ + 215, + 645, + 366, + 662 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In this section, we validate the effectiveness of SeiT++ on various scenarios. First, we evaluate the performance of our method in the context of storage-efficient ImageNet-1k classification and transfer learning on fine-grained datasets. Then, we demonstrate the adaptability of token-based vision training to ADE-20k semantic segmentation, showing additional improvements facilitated by SeiT++. In addition, we present experimental results illustrating how our strategies contribute to enhanced model generalizability in robustness evaluation. We also explore the applicability of our augmentation strategies to alternative input formats, demonstrating its extensibility. Unless otherwise stated, SeiT++ refers to the combination of MTM and our token augmentation strategies.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Lee and Park et al.", + "bbox": [ + 271, + 114, + 400, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/6c90a73afb3ff613605618605b88a1f768cb01d5a368b96ddbf3bac710ff0da5.jpg", + "table_caption": [ + "Table 2: Token-based ImageNet-1k Classification. We report top-1 accuracies (ViT-B) on ImageNet-1k. Note that SeiT++ w/o MTM means the original SeiT [44] training with our token augmentation strategies." + ], + "table_footnote": [], + "table_body": "
StorageSeiTSeiT++
w/o MTMw/ MTMw/o MTMw/ MTM
1.4 GB74.075.175.5 (+1.5)77.8 (+2.7)
1.1 GB70.274.173.1 (+2.9)76.5 (+2.4)
0.8 GB66.370.669.1 (+2.8)74.1 (+3.5)
0.5 GB59.764.563.8 (+4.1)70.0 (+5.5)
0.3 GB47.253.951.2 (+4.0)60.6 (+6.7)
", + "bbox": [ + 233, + 196, + 769, + 325 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/ea67dffa43c22b4796eed92aa6a34ba0d4154243d99d1bab00dffe4986640f28.jpg", + "table_caption": [ + "Table 3: Fine-grained classification. We report top-1 accuracies (ViT-B) on various fine-grained datasets. Our method consistently outperforms its counterpart through effective augmentations for tokens." + ], + "table_footnote": [], + "table_body": "
MethodFlowersCarsiNat18iNat19
SeiT [44]93.579.743.150.1
SeiT++95.286.952.458.9
(+1.7)(+7.2)(+9.3)(+8.8)
", + "bbox": [ + 272, + 390, + 732, + 465 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.1 Storage-efficient ImageNet-1k", + "text_level": 1, + "bbox": [ + 215, + 492, + 506, + 508 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 1 summarizes the performance and compression ratio of storage-efficient ImageNet-1k classification. Our method is compared with image-based approaches, including uniform random sampling, C-score [29]-based sampling, adjusting image resolution, and adjusting JPEG quality factor. Despite a $1\\%$ compression ratio, when a ViT-B/16 model is pre-trained on tokenized ImageNet-1k using MTM and the ViT model is fine-tuned, our method achieves a $77.8\\%$ top-1 accuracy, outperforming various image-based approaches. In Table 2, our method consistently outperforms the token-based learning baseline SeiT by $1.5 - 6.7\\% p$ in top-1 accuracies. Notably, the effectiveness of our method becomes more pronounced as the available storage size decreases, highlighting its efficiency in low-storage scenarios. Furthermore, SeiT++ shows more significant performance gains with MTM. For a storage size of 1.4GB, SeiT++ achieves an improvement of $1.5\\% p$ over SeiT without MTM, while with MTM the improvement increases to $2.7\\% p$ . These results confirm that SeiT++ effectively takes advantage of MTM.", + "bbox": [ + 212, + 518, + 792, + 731 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Fine-grained Classification", + "text_level": 1, + "bbox": [ + 215, + 753, + 480, + 768 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "To further investigate the generalizability of SeiT++, we conduct experiments on four different fine-grained datasets: Flowers [43], StandfordCars [31], iNat-18 [57], and iNat-19 [58]. Specifically, we tokenize each dataset and then fine-tune the token-trained model (ViT-B) on ImageNet-1k with the tokenized dataset. Since", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "SeiT++: Masked Token Modeling Improves Storage-efficient Training", + "bbox": [ + 274, + 114, + 730, + 130 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/da413636c5fd51feb7c9d7e334180eddf849c0c0504fab907e68fa4a720a1209.jpg", + "table_caption": [ + "Table 4: ADE-20k Semantic Segmentation. We report mIoU on the ADE-20k validation set, highlighting the effectiveness of our approach in pixel-level classification." + ], + "table_footnote": [], + "table_body": "
MethodBackboneSeiTSeiT++
UPerNet [63]ViT-B [17]39.043.2 (+4.2)
", + "bbox": [ + 287, + 184, + 715, + 229 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/5b3a877d70faebe76834efa817f8e898385604871845f72914414420ae0d1d50.jpg", + "table_caption": [ + "Table 5: Robustness evaluation. We report top-1 robust accuracies (ViT-B) against corruptions and domain shifts for each model trained on ImageNet-1k. Our method consistently outperforms its counterpart on various robustness benchmarks." + ], + "table_footnote": [], + "table_body": "
MethodCleanGauss. NoiseGauss. BlurImageNet-RSketchObjectNet
SeiT [44]74.050.762.625.522.615.7
SeiT++75.558.666.830.227.718.3
(+1.5)(+7.9)(+4.2)(+4.7)(+5.1)(+2.6)
", + "bbox": [ + 233, + 297, + 767, + 364 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "fine-grained datasets require the classification of objects with intricate characteristics, they are particularly sensitive to undesirable noise (i.e., RRC in SeiT) or excessive perturbation (i.e., Emb-Noise) resulting from data augmentation in the token embedding space. Table 3 shows the quantitative performance (top-1 accuracies) on fine-grained datasets. SeiT++ consistently outperforms the performance of SeiT across various datasets. These results confirm the effectiveness of our proposed method in fine-grained classification.", + "bbox": [ + 212, + 393, + 787, + 500 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 ADE-20k Semantic Segmentation", + "text_level": 1, + "bbox": [ + 215, + 523, + 537, + 539 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "A limitation of token-based learning is often substantiated by learning dense prediction tasks. We conjecture that our method could facilitate training for those tasks, such as semantic segmentation. We extract tokens from ADE-20k [72] and use the pre-trained ViT-B/16 model on tokenized ImageNet-1k to evaluate the effectiveness of our method in semantic segmentation. We then fine-tune the pre-trained model using the tokenized ADE-20k dataset. Table 4 reports the mean intersection-over-union (mIoU) as an evaluation metric. The results show a significant performance improvement $(+4.2\\%)$ p mIoU) over SeiT. These results indicate that SeiT++ significantly improves model performance in the token-based dense prediction task.", + "bbox": [ + 212, + 547, + 787, + 700 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Evaluation on Robustness Datasets", + "text_level": 1, + "bbox": [ + 215, + 723, + 552, + 738 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To validate the effectiveness of our augmentation strategies in terms of robustness, we further evaluate the robustness of our method on a set of benchmarks such as adding Gaussian noise [26] or Gaussian blur [26], ImageNet-R [25], Sketch [59], and ObjectNet [3]. Notably, we adopt SeiT++ w/o MTM to validate the impact of our augmentation strategies on model robustness. The results in Table 5 show that SeiT++ consistently outperforms SeiT in terms of robustness", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Lee and Park et al.", + "bbox": [ + 271, + 114, + 400, + 127 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/ec32eea75f984532940cb2aee2f606decfc428649032a65501c2765cb1a2eac7.jpg", + "table_caption": [ + "Table 6: Impact of the proposed augmentation strategies. We report the top-1 accuracies (ViT-S) on the ImageNet-100 validation set for various combinations of our token augmentation strategies. Notably, our TokenAdapt and ColorAdapt consistently improve model performance, exhibiting synergy when used together." + ], + "table_footnote": [], + "table_body": "
FrameworkColorAdaptTokenAdaptTop 1 Acc.
SeiT [44]XX77.3
✓X78.3
X✓80.4
✓✓81.4
", + "bbox": [ + 338, + 212, + 663, + 299 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/6c88c27a1e09ec262a943a2f6b2dc461a978e06a43f0c3ed04af1d089ff37b6f.jpg", + "table_caption": [ + "Table 7: Adaptibility of our method to different input. Input denotes tokenizers for our model. We report the top-1 accuracies (ViT-S) on ImageNet-100. For VQGAN, we used a publicly available VQGAN [18] trained on OpenImages [32] as a tokenizer." + ], + "table_footnote": [], + "table_body": "
InputSeiTSeiT++
Token (ViT-VQGAN [66])77.381.4 (+4.1)
Token (VQGAN [18])81.883.9 (+2.1)
DCT coefficients69.571.0 (+1.5)
", + "bbox": [ + 346, + 366, + 656, + 439 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "accuracy. Previous studies [8, 55] observe that the key to the input pixel robustness depends on the pixel-level augmentations. Since our method enables pixel-based data augmentation to tokens, it achieves higher robustness than the baseline, which does not effectively exploit the pixel-based data augmentation.", + "bbox": [ + 212, + 467, + 787, + 529 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.5 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 550, + 390, + 566 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Impact of Each Augmentation Strategy. In Table 6, we report the impact of proposed token augmentations. In particular, incorporating ColorAdapt leads to a $1.0\\%$ p performance improvement, highlighting the positive effect of preserving object structure during model training. Additionally, incorporating TokenAdapt yields an extra $3.1\\%$ p improvement. These results confirm that exploring token augmentations is a key factor in fully exploiting pre-extracted tokens.", + "bbox": [ + 212, + 575, + 787, + 667 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Extensibility of Our Augmentation Strategy. In this section, we conjecture the broad applicability of our augmentation strategies across various scenarios. To explore its extensibility, we delve into two distinct situations. Firstly, we examine the integration of our method with an alternative tokenizer—the OpenImages-trained VQGAN tokenizer [18], alongside the ImageNet-1k-trained ViT-VQGAN tokenizer [66]. The results in Table 7 consistently demonstrate improved performance on ImageNet-100 across different tokenizers. Additionally, we investigate the adaptability of our approach to a different input format by focusing on Discrete Cosine Transform (DCT) coefficients. Inspired by JPEG compression, we apply DCT to each $8 \\times 8$ image patch, yielding quantized DCT", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "SeiT++: Masked Token Modeling Improves Storage-efficient Training", + "bbox": [ + 274, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "coefficients utilized as input data during model training. As shown in Table 7, our approach enhances top-1 accuracy, similar to the tokenizer scenario. This consistent improvement across diverse input formats underscores the universal applicability of our method, highlighting its potential as a widely applicable solution for diverse input representations.", + "bbox": [ + 212, + 146, + 787, + 223 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "4.6 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 241, + 457, + 257 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We use the pre-trained ViT-VQGAN tokenizer [66] for token extraction. For our TokenAdapt module, we employ a single transformer block for conversion and reverse function. The TokenAdapt module is trained for one epoch on ImageNet-1k with a batch size of 128, a learning rate of 0.001 with cosine scheduling, and a weight decay of 0.05. We follow the training recipes from MAGE [34], SeiT [44], and mmsegmentation [10] for masked token modeling (MTM), token-based classification, and semantic segmentation, respectively. Detailed hyperparameters and settings are available in the Appendix.", + "bbox": [ + 212, + 262, + 787, + 383 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 402, + 359, + 420 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we investigate the compatibility of SeiT with self-supervised learning approaches in token-based frameworks. Among various self-supervised algorithms, Masked Image Modeling (MIM) [24,64] has recently gained popularity for learning visual representations. Inspired by this, we extend SeiT by integrating Masked Token Modeling (MTM) for self-supervised pre-training. Additionally, recognizing that self-supervised approaches often require more data due to the lack of labels, we propose simple yet effective token-based augmentation strategies: TokenAdapt and ColorAdapt. TokenAdapt aligns the token embedding space with another embedding space compatible with pixel-based data augmentations. In addition, we introduce ColorAdapt, a color-changing augmentation for tokens. Our experimental results across various scenarios, including storage-efficient ImageNet-1k classification, fine-grained classification, ADE-20k semantic segmentation, and robustness benchmarks, consistently validate the effectiveness of our proposed method. Furthermore, the consistent performance improvement across diverse input formats highlights the potential of our approach as a widely applicable solution for various input representations.", + "bbox": [ + 212, + 431, + 787, + 675 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations. While SeiT++ is capable of handling large-scale data without labels with optimized storage requirements, experiments on such large datasets (e.g., ImageNet-21k [14]) and huge models are not contained in this paper due to resource constraints. We believe the advantages of MTM and our data augmentation strategies can be seamlessly integrated with scalability.", + "bbox": [ + 212, + 688, + 787, + 765 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements. Experiments are based on the Naver Smart Machine Learning NSML [30] platform. This research was supported by IITP grant funded by the Korea government(MSIT) and KEIT grant funded by the Korea government(MOTIE) (No. 2022-0-00680, 2022-0-01045).", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Lee and Park et al.", + "bbox": [ + 271, + 114, + 400, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bahng, H., Chun, S., Yun, S., Choo, J., Oh, S.J.: Learning de-biased representations with biased representations. In: International Conference on Machine Learning. pp. 528-539. PMLR (2020)", + "2. Bao, H., Dong, L., Piao, S., Wei, F.: Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254 (2021)", + "3. Barbu, A., Mayo, D., Alverio, J., Luo, W., Wang, C., Gutfreund, D., Tenenbaum, J., Katz, B.: Objectnet: A large-scale bias-controlled dataset for pushing the limits of object recognition models. Advances in neural information processing systems 32 (2019)", + "4. Chang, H., Zhang, H., Barber, J., Maschinot, A., Lezama, J., Jiang, L., Yang, M.H., Murphy, K., Freeman, W.T., Rubinstein, M., et al.: Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704 (2023)", + "5. Chang, H., Zhang, H., Jiang, L., Liu, C., Freeman, W.T.: Maskgit: Masked generative image transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11315-11325 (2022)", + "6. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International conference on machine learning. pp. 1597-1607. PMLR (2020)", + "7. Chen, X., Xie, S., He, K.: An empirical study of training self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9640-9649 (2021)", + "8. Chun, S., Oh, S.J., Yun, S., Han, D., Choe, J., Yoo, Y.: An empirical evaluation on robustness and uncertainty of regularization methods. arXiv preprint arXiv:2003.03879 (2020)", + "9. Coleman, C., Yeh, C., Mussmann, S., Mirzasoleiman, B., Bailis, P., Liang, P., Leskovec, J., Zaharia, M.: Selection via proxy: Efficient data selection for deep learning. arXiv preprint arXiv:1906.11829 (2019)", + "0. Contributors, M.: MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark. https://github.com/open-mmlab/mmsegmentation (2020)", + "1. Croce, F., Hein, M.: Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In: International conference on machine learning. pp. 2206-2216. PMLR (2020)", + "2. Cubuk, E.D., Zoph, B., Mane, D., Vasudevan, V., Le, Q.V.: Autoaugment: Learning augmentation strategies from data. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 113-123 (2019)", + "3. Cubuk, E.D., Zoph, B., Shlens, J., Le, Q.V.: Randaugment: Practical automated data augmentation with a reduced search space. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 702-703 (2020)", + "4. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)", + "5. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)", + "6. DeVries, T., Taylor, G.W.: Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552 (2017)" + ], + "bbox": [ + 225, + 175, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "SeiT++: Masked Token Modeling Improves Storage-efficient Training", + "bbox": [ + 276, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "17. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)", + "18. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12873-12883 (2021)", + "19. Fang, Y., Dong, L., Bao, H., Wang, X., Wei, F.: Corrupted image modeling for self-supervised visual pre-training. arXiv preprint arXiv:2202.03382 (2022)", + "20. Ford, N., Gilmer, J., Carlini, N., Cubuk, D.: Adversarial examples are a natural consequence of test error in noise. arXiv preprint arXiv:1901.10513 (2019)", + "21. Geirhos, R., Rubisch, P., Michaelis, C., Bethge, M., Wichmann, F.A., Brendel, W.: Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. arXiv preprint arXiv:1811.12231 (2018)", + "22. Goodfellow, I.J., Shlens, J., Szegedy, C.: Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572 (2014)", + "23. Grill, J.B., Strub, F., Altché, F., Tallec, C., Richemond, P., Buchatskaya, E., Doersch, C., Avila Pires, B., Guo, Z., Gheshlaghi Azar, M., et al.: Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems 33, 21271-21284 (2020)", + "24. He, K., Chen, X., Xie, S., Li, Y., Dollár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022)", + "25. Hendrycks, D., Basart, S., Mu, N., Kadavath, S., Wang, F., Dorundo, E., Desai, R., Zhu, T., Parajuli, S., Guo, M., Song, D., Steinhardt, J., Gilmer, J.: The many faces of robustness: A critical analysis of out-of-distribution generalization. ICCV (2021)", + "26. Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261 (2019)", + "27. Huang, X., Belongie, S.: Arbitrary style transfer in real-time with adaptive instance normalization. In: Proceedings of the IEEE international conference on computer vision. pp. 1501-1510 (2017)", + "28. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International conference on machine learning. pp. 4904-4916. PMLR (2021)", + "29. Jiang, Z., Zhang, C., Talwar, K., Mozer, M.C.: Characterizing structural regularities of labeled data in overparameterized models. arXiv preprint arXiv:2002.03206 (2020)", + "30. Kim, H., Kim, M., Seo, D., Kim, J., Park, H., Park, S., Jo, H., Kim, K., Yang, Y., Kim, Y., et al.: Nsml: Meet the mlaas platform with a real-world case study. arXiv preprint arXiv:1810.09957 (2018)", + "31. Krause, J., Stark, M., Deng, J., Fei-Fei, L.: 3d object representations for fine-grained categorization. In: Proceedings of the IEEE international conference on computer vision workshops. pp. 554-561 (2013)", + "32. Kuznetsova, A., Rom, H., Alldrin, N., Uijlings, J., Krasin, I., Pont-Tuset, J., Kamali, S., Popov, S., Malloci, M., Kolesnikov, A., et al.: The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision 128(7), 1956-1981 (2020)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Lee and Park et al.", + "bbox": [ + 271, + 114, + 400, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "33. Lee, S., Chun, S., Jung, S., Yun, S., Yoon, S.: Dataset condensation with contrastive signals. In: International Conference on Machine Learning. pp. 12352-12364. PMLR (2022)", + "34. Li, T., Chang, H., Mishra, S., Zhang, H., Katabi, D., Krishnan, D.: Mage: Masked generative encoder to unify representation learning and image synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2142-2152 (2023)", + "35. Lin, S., Ji, R., Yan, C., Zhang, B., Cao, L., Ye, Q., Huang, F., Doermann, D.: Towards optimal structured cnn pruning via generative adversarial learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2790-2799 (2019)", + "36. Liu, J., Guo, J., Xu, D.: Apsnet: Toward adaptive point sampling for efficient 3d action recognition. IEEE Transactions on Image Processing 31, 5287-5302 (2022)", + "37. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021)", + "38. Lopes, R.G., Yin, D., Poole, B., Gilmer, J., Cubuk, E.D.: Improving robustness without sacrificing accuracy with patch gaussian augmentation. arXiv preprint arXiv:1906.02611 (2019)", + "39. Ma, Y., Li, H., Zhang, Z., Guo, J., Zhang, S., Gong, R., Liu, X.: Annealing-based label-transfer learning for open world object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11454-11463 (2023)", + "40. Madry, A., Makelov, A., Schmidt, L., Tsipras, D., Vladu, A.: Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083 (2017)", + "41. Mahajan, D., Girshick, R., Ramanathan, V., He, K., Paluri, M., Li, Y., Bharambe, A., Van Der Maaten, L.: Exploring the limits of weakly supervised pretraining. In: Proceedings of the European conference on computer vision (ECCV). pp. 181-196 (2018)", + "42. Mei, Z., Ye, P., Ye, H., Li, B., Guo, J., Chen, T., Ouyang, W.: Automatic loss function search for adversarial unsupervised domain adaptation. IEEE Transactions on Circuits and Systems for Video Technology 33(10), 5868-5881 (2023)", + "43. Nilsback, M.E., Zisserman, A.: Automated flower classification over a large number of classes. In: 2008 Sixth Indian conference on computer vision, graphics & image processing. pp. 722-729. IEEE (2008)", + "44. Park, S., Chun, S., Heo, B., Kim, W., Yun, S.: Seit: Storage-efficient vision training with tokens using $1\\%$ of pixel storage. arXiv preprint arXiv:2303.11114 (2023)", + "45. Paul, M., Ganguli, S., Dziugaite, G.K.: Deep learning on a data diet: Finding important examples early in training. Advances in Neural Information Processing Systems 34, 20596-20607 (2021)", + "46. Riba, E., Mishkin, D., Ponsa, D., Rublee, E., Bradski, G.: Kornia: an open source differentiable computer vision library for pytorch. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3674-3683 (2020)", + "47. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10684-10695 (2022)", + "48. Rosasco, A., Carta, A., Cossu, A., Lomonaco, V., Bacciu, D.: Distilled replay: Overcoming forgetting through synthetic samples. In: International Workshop on Continual Semi-Supervised Learning. pp. 104-117. Springer (2021)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "SeiT++: Masked Token Modeling Improves Storage-efficient Training", + "bbox": [ + 276, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "49. Sangermano, M., Carta, A., Cossu, A., Bacciu, D.: Sample condensation in online continual learning. In: 2022 International Joint Conference on Neural Networks (IJCNN). pp. 01-08. IEEE (2022)", + "50. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022)", + "51. Scimeca, L., Oh, S.J., Chun, S., Poli, M., Yun, S.: Which shortcut cues will dnns choose? a study from the parameter-space perspective. arXiv preprint arXiv:2110.03095 (2021)", + "52. Shao, S., Yin, Z., Zhou, M., Zhang, X., Shen, Z.: Generalized large-scale data condensation via various backbone and statistical matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16709-16718 (2024)", + "53. Singh, M., Gustafson, L., Adcock, A., de Freitas Reis, V., Gedik, B., Kosaraju, R.P., Mahajan, D., Girshick, R., Dollar, P., Van Der Maaten, L.: Revisiting weakly supervised pre-training of visual perception models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 804-814 (2022)", + "54. Steiner, A., Kolesnikov, A., Zhai, X., Wightman, R., Uszkoreit, J., Beyer, L.: How to train your vit? data, augmentation, and regularization in vision transformers. arXiv preprint arXiv:2106.10270 (2021)", + "55. Taori, R., Dave, A., Shankar, V., Carlini, N., Recht, B., Schmidt, L.: Measuring robustness to natural distribution shifts in image classification. Advances in Neural Information Processing Systems 33, 18583-18599 (2020)", + "56. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jégou, H.: Training data-efficient image transformers & distillation through attention. In: International conference on machine learning. pp. 10347-10357. PMLR (2021)", + "57. Van Horn, G., Mac Aodha, O., Song, Y., Cui, Y., Sun, C., Shepard, A., Adam, H., Perona, P., Belongie, S.: The inaturalist challenge 2018 dataset. arXiv preprint arXiv:1707.06642 (2018)", + "58. Van Horn, G., Mac Aodha, O., Song, Y., Cui, Y., Sun, C., Shepard, A., Adam, H., Perona, P., Belongie, S.: The inaturalist challenge 2019 dataset. arXiv preprint arXiv:1707.06642 (2019)", + "59. Wang, H., Ge, S., Lipton, Z., Xing, E.P.: Learning robust global representations by penalizing local predictive power. In: Advances in Neural Information Processing Systems. pp. 10506-10518 (2019)", + "60. Wang, T., Zhu, J.Y., Torralba, A., Efros, A.A.: Dataset distillation. arXiv preprint arXiv:1811.10959 (2018)", + "61. Wei, J., Zou, K.: Eda: Easy data augmentation techniques for boosting performance on text classification tasks. arXiv preprint arXiv:1901.11196 (2019)", + "62. Wightman, R., Touvron, H., Jégou, H.: Resnet strikes back: An improved training procedure in timm. arXiv preprint arXiv:2110.00476 (2021)", + "63. Xiao, T., Liu, Y., Zhou, B., Jiang, Y., Sun, J.: Unified perceptual parsing for scene understanding. In: Proceedings of the European conference on computer vision (ECCV). pp. 418-434 (2018)", + "64. Xie, Z., Zhang, Z., Cao, Y., Lin, Y., Bao, J., Yao, Z., Dai, Q., Hu, H.: Simmim: A simple framework for masked image modeling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9653-9663 (2022)", + "65. Yin, Z., Xing, E., Shen, Z.: Squeeze, recover and relabel: Dataset condensation at imagenet scale from a new perspective. In: Thirty-seventh Conference on Neural Information Processing Systems (2023)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Lee and Park et al.", + "bbox": [ + 271, + 114, + 400, + 126 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "66. Yu, J., Li, X., Koh, J.Y., Zhang, H., Pang, R., Qin, J., Ku, A., Xu, Y., Baldridge, J., Wu, Y.: Vector-quantized image modeling with improved vqgan. arXiv preprint arXiv:2110.04627 (2021)", + "67. Yun, S., Han, D., Oh, S.J., Chun, S., Choe, J., Yoo, Y.: Cutmix: Regularization strategy to train strong classifiers with localizable features. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6023-6032 (2019)", + "68. Zhang, H., Cisse, M., Dauphin, Y.N., Lopez-Paz, D.: mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412 (2017)", + "69. Zhao, B., Bilen, H.: Dataset condensation with differentiable siamese augmentation. In: International Conference on Machine Learning. pp. 12674-12685. PMLR (2021)", + "70. Zhao, B., Mopuri, K.R., Bilen, H.: Dataset condensation with gradient matching. arXiv preprint arXiv:2006.05929 (2020)", + "71. Zhong, Z., Zheng, L., Kang, G., Li, S., Yang, Y.: Random erasing data augmentation. In: Proceedings of the AAAI Conference on Artificial Intelligence (AAAI) (2020)", + "72. Zhou, B., Zhao, H., Puig, X., Xiao, T., Fidler, S., Barriuso, A., Torralba, A.: Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision 127, 302-321 (2019)" + ], + "bbox": [ + 212, + 146, + 787, + 410 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "SeiT++: Masked Token Modeling Improves Storage-efficient Training", + "bbox": [ + 276, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/SeiT++_ Masked Token Modeling Improves Storage-efficient Training/a6a99e76-474a-47eb-80d4-d9020aeb146a_model.json b/2024/SeiT++_ Masked Token Modeling Improves Storage-efficient Training/a6a99e76-474a-47eb-80d4-d9020aeb146a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8327f6cb22bcfcd7eedb4dc2c98a71e4059cd66a --- /dev/null +++ b/2024/SeiT++_ Masked Token Modeling Improves Storage-efficient Training/a6a99e76-474a-47eb-80d4-d9020aeb146a_model.json @@ -0,0 +1,2581 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.26, + 0.141, + 0.745, + 0.187 + ], + "angle": 0, + "content": "SeiT++: Masked Token Modeling Improves Storage-efficient Training" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.213, + 0.773, + 0.245 + ], + "angle": 0, + "content": "Minhyun Lee\\(^{1,\\dagger}\\) Song Park\\(^{2,\\dagger}\\) Byeongho Heo\\(^{2}\\) Dongyoon Han\\(^{2}\\) Hyunjung Shim\\(^{3,\\star}\\)" + }, + { + "type": "text", + "bbox": [ + 0.301, + 0.25, + 0.702, + 0.282 + ], + "angle": 0, + "content": "\\(^{1}\\) Yonsei University \\(^{2}\\) Naver AI Lab \\(^{3}\\) KAIST AI \\(^{4}\\) Equal contribution. Work done at Naver AI Lab." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.324, + 0.744, + 0.63 + ], + "angle": 0, + "content": "Abstract. Recent advancements in Deep Neural Network (DNN) models have significantly improved performance across computer vision tasks. However, achieving highly generalizable and high-performing vision models requires expansive datasets, resulting in significant storage requirements. This storage challenge is a critical bottleneck for scaling up models. A recent breakthrough by SeiT proposed the use of Vector-Quantized (VQ) feature vectors (i.e., tokens) as network inputs for vision classification. This approach achieved \\(90\\%\\) of the performance of a model trained on full-pixel images with only \\(1\\%\\) of the storage. While SeiT needs labeled data, its potential in scenarios beyond fully supervised learning remains largely untapped. In this paper, we extend SeiT by integrating Masked Token Modeling (MTM) for self-supervised pre-training. Recognizing that self-supervised approaches often demand more data due to the lack of labels, we introduce TokenAdapt and ColorAdapt. These methods facilitate comprehensive token-friendly data augmentation, effectively addressing the increased data requirements of self-supervised learning. We evaluate our approach across various scenarios, including storage-efficient ImageNet-1k classification, fine-grained classification, ADE-20k semantic segmentation, and robustness benchmarks. Experimental results demonstrate consistent performance improvement in diverse experiments, validating the effectiveness of our method. Code is available at https://github.com/naver-ai/seit." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.658, + 0.376, + 0.674 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.693, + 0.795, + 0.816 + ], + "angle": 0, + "content": "Recent advancements in Deep Neural Network (DNN) [17,56] have significantly improved their performance across various computer vision tasks. However, achieving highly generalizable and high-performing vision models demands an extensive dataset, often containing billions of data points [28,41,53] with large storage requirements. For example, the LAION-5B [50] dataset requires a storage size of 240TB. This storage challenge emerges as a critical bottleneck for scaling up vision models. Consequently, optimizing storage efficiency becomes essential for the practical advancement of scaled-up vision models." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.518, + 0.84 + ], + "angle": 0, + "content": "* Hyunjung Shim is a corresponding author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.401, + 0.127 + ], + "angle": 0, + "content": "Lee and Park et al." + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.145, + 0.728, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.3, + 0.788, + 0.385 + ], + "angle": 0, + "content": "Fig.1: Over \\(70\\%\\) top-1 accuracy is achievable with just 1GB data. On ImageNet-1k, we visualize the trade-off of training data storage vs. top-1 accuracy using the fixed ViT-B/16 for a controlled comparison. Each accuracy metric is individually trained with different data types. While the entire ImageNet-1k dataset requires approximately 140GB for training with images, our approach demonstrates significant storage efficiency over competitors." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.416, + 0.788, + 0.583 + ], + "angle": 0, + "content": "Images require more storage than texts due to their intrinsic complexity in converting continuous natural light signals into discrete pixels for computer representation. However, a considerable portion of image data corresponds to redundant or unnecessary details, such as repeated patterns or imperceptible details [47]. These properties lead to large storage requirements and also induce negative impacts on visual recognition tasks as mentioned in [1,11,21,22,40,51]. Therefore, several attempts [9, 29, 45, 69, 70] have been made to reduce storage requirements by optimizing the efficiency of image data. A standard approach to reduce storage requirements is to decrease the total number of training data by removing less important samples [9, 29, 45] or condensing the data into smaller synthetic sets [69, 70]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.788, + 0.735 + ], + "angle": 0, + "content": "However, as SeiT [44] illustrates, these approaches may incur a significant performance drop or become infeasible for large-scale datasets due to their complexity. Alternatively, some methods aim to reduce the size of each image through compression techniques or resolution adjustments [50]. However, they still face a considerable performance gap with the original dataset. Recently, SeiT [44] has made a significant breakthrough in this area by enabling the direct learning of ViTs [17] from discrete representations (i.e., tokens) pre-extracted by an offline tokenizer [66]. By using vector-quantized features instead of traditional pixel images, SeiT achieves \\(90\\%\\) of the performance of a full storage model while requiring only \\(1\\%\\) of the storage." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.842 + ], + "angle": 0, + "content": "While SeiT [44] has demonstrated the high potential of token-based training, a performance gap with the image-based models still persists, and a prospective solution - exploration beyond fully supervised learning scenarios - remains untapped. To leverage the advantages of large-scale data beyond human-annotated datasets, exploring label-free learning methods is imperative. In this paper, we investigate the compatibility of SeiT with self-supervised learning approaches in token-based frameworks. Among many different self-supervised methods [6,7,23]," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.277, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "SeiT++: Masked Token Modeling Improves Storage-efficient Training" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "Masked Image Modeling (MIM) [24,64] derived from Masked Language Modeling (MLM) [15] has brought significant success and has quickly become a popular approach for learning visual representations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.202, + 0.788, + 0.385 + ], + "angle": 0, + "content": "Inspired by this, we introduce Masked Token Modeling (MTM), a pre-training method that performs token masking and reconstruction in token-based frameworks, and validate its effectiveness for storage efficiency on various benchmarks. To the best of our knowledge, this is the first approach for learning an MLM-based model directly from offline tokens, offering the potential to optimize storage efficiency for large-scale training tasks. As previous studies [12, 13, 56] have highlighted, data augmentation plays a crucial role in large-scale model training and is a fundamental element in various training setups [6, 7, 23, 24, 37, 56, 62, 64]. However, applying existing data augmentation techniques [13, 67, 71] designed for pixel images to tokens poses a significant challenge due to the input domain shift. Therefore, addressing this challenge is key to overcoming major performance bottlenecks and improving the storage efficiency of token-based frameworks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.393, + 0.788, + 0.575 + ], + "angle": 0, + "content": "To this end, we introduce TokenAdapt and ColorAdapt, two novel augmentation strategies designed for token-based training. TokenAdapt facilitates the integration of traditional pixel-based augmentation policies by transforming token embeddings into a feature space compatible with pixel-based augmentations. After applying augmentation to the transformed token embeddings, the augmented token embeddings are then reverted to their original token space. ColorAdapt, inspired by Adaptive Instance Normalization (AdaIN [27]), alters the color attributes of token embeddings while preserving object structure by adjusting their statistics. By integrating TokenAdapt and ColorAdapt into our token augmentation process, we enhance the effectiveness of token-based training across both fully supervised and unsupervised learning scenarios, demonstrating significant improvements in model performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.797, + 0.844 + ], + "angle": 0, + "content": "Our SeiT++ framework combines Masked Token Modeling (MTM) with the proposed token augmentation strategies, TokenAdapt and ColorAdapt. We validate the effectiveness of SeiT++ on various scenarios: (1) storage-efficient ImageNet-1k classification, (2) transfer learning on fine-grained classification, (3) ADE-20k semantic segmentation, and (4) evaluation on robustness benchmarks. As shown in Figure 1, SeiT++ consistently outperforms comparison methods in storage-efficient ImageNet-1k classification with limited storage sizes, achieving a top-1 accuracy of \\(77.8\\%\\) with only 1.4GB of data. These results confirm that token-based training can take advantage of self-supervised learning to improve storage efficiency. We also demonstrate that our proposed token augmentations significantly contribute to storage efficiency by augmenting data effectively. We reveal that our token-based augmentation methods basically yield significant boosts in supervised learning and further amplify improvements in MTM learning. Furthermore, we present how our strategies enhance the generalizability of trained models, as evidenced by our experimental results on fine-grained classification and robustness evaluation. Lastly, the applicability of our method to an alternative input format (e.g., DCT coefficients) demonstrates its extensibility." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.232, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.401, + 0.127 + ], + "angle": 0, + "content": "Lee and Park et al." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.144, + 0.388, + 0.162 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.176, + 0.786, + 0.342 + ], + "angle": 0, + "content": "Learning with Tokenization. The use of discrete tokens for training vision models has recently gained attention, particularly in self-supervised representation learning. BeiT [2] employs the Masked Image Modeling (MIM) framework to recover discrete tokens from masked images and CIM [19] utilizes tokens in Corrupted Image Modeling. MAGE [34] demonstrates the incorporation of tokens from VQGAN [18] for both generative training and representation learning within a unified training framework. However, existing studies mainly focus on self-supervised representation learning and rely on online tokenization, converting images to tokens during each iteration of model training. This approach incurs notable memory consumption and computational costs, limiting the full exploitation of the storage efficiency of discrete tokens." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.364, + 0.786, + 0.652 + ], + "angle": 0, + "content": "Storage-efficient Vision Training. Efficient model learning is a crucial research area, with studies focusing on optimizing target models [35, 36] or minimizing training data and labels [39, 42]. Storage-efficient training aims to reduce the dataset storage size. To this end, dataset distillation (DD) [60] creates a condensed dataset by leveraging the knowledge of the original dataset. Recent research [33, 48, 49, 69, 70] highlights the effectiveness of dataset distillation in improving training efficiency. However, its high computational complexity poses challenges, particularly when applied to large-scale datasets like ImageNet-1k [14]. Sampling-based methods [9, 29, 45] address this by selecting a subset of the most representative samples from the full dataset. Paul et al. [45] identify important samples very early in training. Jiang et al. [29] utilize a consistency score based on statistics collected during training. However, the diverse sample selection is not guaranteed, particularly in low data regimes, leading to suboptimal performance compared to dataset distillation (DD) methods [70]. To tackle these issues, SeiT [44] proposes a storage-efficient vision training framework by leveraging the storage efficiency of tokens. They utilize the ViT-VQGAN [66] tokenizer to extract tokens from images and store them for vision model training. As a result, they outperform other storage-efficient training methods, significantly reducing the dataset storage size." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.673, + 0.786, + 0.84 + ], + "angle": 0, + "content": "Data Augmentation. Data augmentation serves as a fundamental building block to improve the model performance across diverse computer vision tasks. Various data augmentation techniques have been employed for improved performance [16, 71] or model robustness [20, 38]. Mixup [68] introduces convex combinations of pairs of images and their labels to create a smoother decision boundary. CutMix [67] replaces certain regions of an image with a patch from another image, enhancing the model's localization ability. AutoAugment [12] leverages reinforcement learning to automate optimal augmentation policy discovery. RandAugment [13] introduces a potent policy based on random combinations of different data augmentation methods. DeiT [56] emphasizes the crucial role of data augmentation in vision transformers for improved model performance." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.277, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "SeiT++: Masked Token Modeling Improves Storage-efficient Training" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.145, + 0.784, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.283, + 0.788, + 0.34 + ], + "angle": 0, + "content": "Fig. 2: Masked Token Modeling (MTM) pipeline. MTM is a self-supervised learning approach in token-based frameworks. (a) The tokenized dataset is saved on storage before model training. Then, (b) using only the pre-stored tokens, a storage-efficient vision model (MTM) is trained without relying on labeled datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.37, + 0.788, + 0.447 + ], + "angle": 0, + "content": "Previous studies extensively explore the effective data augmentation policies for pixel-based vision model training. However, we observe that pixel-based data augmentations are not well-suited for tokens, posing a critical performance bottleneck. Motivated by this, we tackle the data augmentation challenge, which has not been explored in previous studies for vision model training with tokens." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.47, + 0.331, + 0.485 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.503, + 0.695, + 0.52 + ], + "angle": 0, + "content": "3.1 Preliminary: Storage-efficient Vision Training (SeiT)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.529, + 0.789, + 0.697 + ], + "angle": 0, + "content": "SeiT [44] aims to develop a scalable and high-performance vision classifier with minimal storage requirements. It introduces a storage-efficient vision training framework that enables the direct learning of ViT [17] from discrete representations (i.e., tokens) pre-extracted by an offline tokenizer. In addition, SeiT further improves training by incorporating two token-specific augmentations: Token-EDA, drawing inspiration from easy data augmentation for language [61], and Emb-Noise, which introduces Gaussian noise into discrete representations. By using vector-quantized features instead of pixel images, SeiT retains \\(90\\%\\) of the precision of a model trained on full-pixel images while using only \\(1\\%\\) of the storage space. SeiT consists of two parts: (1) preparing the token dataset and (2) training a model using the token dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Preparing the Token Dataset. SeiT extracts tokens using the ImageNet-1k-trained ViT-VQGAN tokenizer [66] to prepare the token dataset for training. As shown in Figure 2(a), each pixel image \\(I\\) is transformed into a sequence of tokens \\(T = [t_1, t_2, \\dots, t_n]\\). Since the tokens are stored in quantized form, each token means an embedding vector in a codebook \\(\\mathcal{Z} = \\{z_k \\in \\mathbb{R}^d | k = 1, \\dots, K\\}\\). i.e., \\(t_i \\in \\mathbb{Q}\\) where \\(\\mathbb{Q} = \\{1, 2, \\dots, K\\}\\) for a codebook with size \\(K\\). Here, given the tokens \\(T\\) and the codebook \\(\\mathcal{Z}\\), the corresponding token embeddings \\(Z_T = [z_{t_1}, z_{t_2}, \\dots, z_{t_n}]\\) can be computed by lookup process. This conversion compresses each image" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.401, + 0.128 + ], + "angle": 0, + "content": "Lee and Park et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.239 + ], + "angle": 0, + "content": "\\(I \\in \\mathbb{R}^{H \\times W \\times 3}\\) into a more compact token sequence \\(T \\in \\mathbb{R}^n\\), where \\(n\\) is much smaller than \\(H \\times W \\times 3\\), thus significantly reducing storage requirements. For example, storing the full-pixel ImageNet-1k requires 140GB, but its tokenized form needs only 1.4GB. Notably, the tokens are stored on storage before model training with the codebook \\(\\mathcal{Z}\\). Thus, during the training phase, the pre-stored tokens are used directly as input instead of the original pixel images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.267, + 0.79, + 0.435 + ], + "angle": 0, + "content": "Training Pipeline. To train a vision classifier using tokens, SeiT initially loads the token dataset and applies Token-EDA to the tokens. The tokens are then converted to a one-hot format and randomly resized and cropped. Subsequently, the processed one-hot tokens are transformed into token embeddings using a pretrained codebook. Techniques such as CutMix and Embedding-noise are applied to these token embeddings, and the resulting augmented tensor is fed into the model for training. Through the proposed framework, SeiT demonstrates the high potential of token-based training, but it lacks exploration beyond fully supervised learning scenarios. In addition, while token-specific augmentation is proposed, the types of token augmentation are limited, and the analysis of data augmentation compared to pixel-based training is insufficient." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.463, + 0.466, + 0.479 + ], + "angle": 0, + "content": "3.2 Masked Token Modeling" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.494, + 0.788, + 0.572 + ], + "angle": 0, + "content": "In this section, we introduce Masked Token Modeling (MTM), a pre-training method in token-based frameworks without relying on labeled datasets. MTM shares a similar concept to Masked Language Modeling (MLM) [15], which works by reconstructing original tokens from their masked versions, employing an autoencoding strategy. Figure 2(b) presents an overview of how our MTM operates." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.788, + 0.693 + ], + "angle": 0, + "content": "Masking. We employ a variable masking ratio that follows a truncated normal distribution, as investigated in previous studies [4,5,34]. The tokens are randomly dropped according to the sampled masking ratio, and then the visible, unmasked token embeddings are fed into the model. By processing only a subset of token embeddings in the encoder, we significantly reduce the overall pre-training time and memory consumption, which aligns with the findings of MAE [24]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.788, + 0.843 + ], + "angle": 0, + "content": "MTM Encoder & Decoder. We use ViT [17] as our encoder because it is able to process both the visible token embeddings and the original token embeddings, seamlessly accommodating their different lengths. During training, the encoder takes the visible token embeddings and outputs latent features that become the decoder input. Before decoding, the latent features are padded to the length of the original token embeddings using mask tokens M. Each mask token is a shared, learned vector indicating the presence of a missing token to be predicted. The decoder then estimates the original tokens \\( T \\) from the padded features." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.277, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "SeiT++: Masked Token Modeling Improves Storage-efficient Training" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.149, + 0.305, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.312, + 0.149, + 0.391, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.395, + 0.15, + 0.474, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.478, + 0.15, + 0.557, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.291, + 0.286, + 0.504, + 0.3 + ], + "angle": 0, + "content": "(a) Failure cases visualization" + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.146, + 0.768, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.573, + 0.286, + 0.783, + 0.301 + ], + "angle": 0, + "content": "(b) Performance degradation" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.312, + 0.788, + 0.438 + ], + "angle": 0, + "content": "Fig. 3: Data augmentations with tokens. Each ViT-VQGAN decoded image is reconstructed from a given RGB image after undergoing specific data augmentations. We observe that naively adopting these methods results in incorrect tokenization: 1) Token w/ hFlip demonstrates spatial information collapse during tokenization; 2) Token w/ RRC shows interdependence between neighboring token embeddings. We note the reconstructed images fail to preserve the images' details, suggesting that this incurs ineffectiveness of tokenization. Furthermore, we report top-1 accuracies (ViT-B) on ImageNet-1k according to the augmentations applied during training. All the augmentations are de-facto default training setups for vision transformers [37,54,56]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.466, + 0.788, + 0.528 + ], + "angle": 0, + "content": "Training Objective. MTM learns to accurately predict the original tokens of each masked token from the visible tokens. In particular, we use a cross-entropy loss between the decoder output and the ground truth tokens as the training objective, which is denoted by \\(\\mathcal{L}_{\\text{recon}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.417, + 0.541, + 0.785, + 0.559 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e c o n}} = \\mathrm {C E} \\left(T _ {M} ^ {\\prime}, T _ {M}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.563, + 0.788, + 0.609 + ], + "angle": 0, + "content": "where \\( T_M' \\) denotes the predictions of the model for the masked tokens, and \\( T_M \\) corresponds to the ground truth tokens for \\( T_M' \\). Following MAE [24], this loss is optimized only for the masked tokens." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.63, + 0.519, + 0.645 + ], + "angle": 0, + "content": "3.3 Data Augmentation for Tokens" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.653, + 0.788, + 0.76 + ], + "angle": 0, + "content": "While Masked Token Modeling (MTM) improves the storage efficiency of the token-based framework, there is still a performance bottleneck in improving token-based training. Since self-supervised approaches often require more data due to the lack of labels, exploring data augmentation is important to take full advantage of MTM. In this section, we analyze why existing data augmentations designed for pixel images are ineffective in the token domain. We then present TokenAdapt and ColorAdapt as solutions for token-based data augmentation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Limited Data Augmentations for Tokens. We explore the challenges of pixel-based augmentation in the token domain and identify two main reasons: (1) spatial information collapse during tokenization and (2) inter-token dependencies. The discrete tokenization process condenses a \\( n \\times n \\) 2D image patch" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.401, + 0.127 + ], + "angle": 0, + "content": "Lee and Park et al." + }, + { + "type": "image", + "bbox": [ + 0.261, + 0.144, + 0.744, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.24, + 0.788, + 0.298 + ], + "angle": 0, + "content": "Fig. 4: TokenAdapt processing pipeline. TokenAdapt aims to enhance the compatibility of token embeddings with pixel-based data augmentations by converting them into augmentation-compatible space, applying augmentations, and reverting them back to the original token embedding space." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.328, + 0.784, + 0.404 + ], + "angle": 0, + "content": "into a single \\(d\\)-dimensional 1D vector (i.e., token embedding), causing a collapse of spatial information. This collapse hinders augmentations that rely on spatial details (e.g., horizontal flip). Furthermore, unlike independent image pixels, token embeddings are inter-correlated. This complicates augmentations involving interpolation (e.g., Resize or Mixup [68]), by introducing undesired artifacts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.405, + 0.788, + 0.542 + ], + "angle": 0, + "content": "Figure 3 shows the limitations of pixel-based augmentations in the token domain. A horizontal Flip (hFlip) distorts both the diagonal line (top row) and the cat's silhouette (bottom row) due to spatial information collapse. This occurs because only the inter-token spatial relationship is flipped, while the inner-token spatial arrangement remains unchanged. Augmentations involving interpolation can lead to undesired artifacts, such as color leakage in the case of Random Resized Crop (RRC), due to interdependence between neighboring embeddings. Therefore, we suspect that the direct application of pixel-based augmentations to token embedding may negatively affect token-based training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.567, + 0.788, + 0.703 + ], + "angle": 0, + "content": "TokenAdapt. The concept of TokenAdapt is to establish a feature space that allows for conventional pixel-based augmentations while facilitating the conversion and inversion of pre-extracted tokens with minimal computational overhead. This approach offers several advantages for our TokenAdapt. Firstly, it enables the direct reuse of well-designed pixel-based data augmentation policies. Additionally, the augmented token can be easily computed through the conversion-augmentation-reverse process. Finally, it ensures that the feature spaces for both the original and augmented tokens are preserved identically. This allows for the seamless use of existing model architectures or pre-trained weights." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.789, + 0.842 + ], + "angle": 0, + "content": "The goal of TokenAdapt is to design an augmentation function that converts the input tokens \\( T \\) to augmented tokens \\( T^{\\mathbf{A}} = [t_1^{\\mathbf{A}}, t_2^{\\mathbf{A}}, \\dots, t_n^{\\mathbf{A}}] \\). Note that the augmented tokens should also be quantized \\( t_i^{\\mathbf{A}} \\in \\mathbb{Q} \\). Figure 4 depicts the overall pipeline of TokenAdapt. The initial step of TokenAdapt is transforming token embeddings \\( Z_T \\), computed from \\( T \\) and codebook \\( \\mathcal{Z} \\), into a representation \\( S_T \\in \\mathbb{R}^{h \\times w \\times d} \\) in augmentation-compatible space \\( S \\). This transformation is achieved using a conversion function \\( f \\), i.e., \\( S_T = f(Z_T) \\). We apply augmentation \\( \\mathbf{A} \\) at augmentation-compatible space \\( S \\) and build an augmented representation \\( S_T^{\\mathbf{A}} = \\mathbf{A}(S_T) \\). Then, the augmented representation is reverted to" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.276, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "SeiT++: Masked Token Modeling Improves Storage-efficient Training" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.146, + 0.784, + 0.207 + ], + "angle": 0, + "content": "token embedding space by a reverse transformation \\( g \\), \\( Z_T^{\\mathbf{A}} = g(S_T^{\\mathbf{A}}) \\). At last, the augmented tokens in embedding space \\( Z_T^{\\mathbf{A}} \\) are quantized to codebook index as \\( T^{\\mathbf{A}} = \\mathbf{q}_{\\mathcal{Z}}(Z_T^{\\mathbf{A}}) \\), where \\( \\mathbf{q}_{\\mathcal{Z}} \\) indicates the vector quantization process. In summary, the augmentation process of TokenAdapt is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.368, + 0.222, + 0.785, + 0.24 + ], + "angle": 0, + "content": "\\[\nZ _ {T} ^ {\\mathbf {A}} = g (\\mathbf {A} (f (Z _ {T}))), \\quad T ^ {\\mathbf {A}} = \\mathbf {q} _ {\\mathcal {Z}} \\left(Z _ {T} ^ {\\mathbf {A}}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.245, + 0.784, + 0.32 + ], + "angle": 0, + "content": "The conversion function \\( f \\) and the reverse transformation \\( g \\) are learned from paired data of tokens \\( (T_x, T_{\\mathbf{A}(x)}) \\), where \\( T_x \\) and \\( T_{\\mathbf{A}(x)} \\) denote tokens from an image \\( x \\) and its augmented image \\( \\mathbf{A}(x) \\), respectively. We use the cross-entropy loss between \\( Z_{T_x}^{\\mathbf{A}} \\) and \\( Z_{T_{\\mathbf{A}(x)}} \\) as our objective function \\( \\mathcal{L}_{\\mathrm{TA}} \\) for learning our TokenAdapt module:" + }, + { + "type": "equation", + "bbox": [ + 0.418, + 0.336, + 0.785, + 0.354 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {T A}} = \\operatorname {C E} \\left(Z _ {T _ {x}} ^ {\\mathbf {A}}, Z _ {T _ {\\mathbf {A} (x)}}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.36, + 0.784, + 0.419 + ], + "angle": 0, + "content": "TokenAdapt deals with geometric data augmentations (e.g., flip, resize, crop, affine) and Mixup [68]. These augmentations are commonly used in various training methods, but they can cause a performance drop in token-based training, as shown in Figure 3(b)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.42, + 0.785, + 0.526 + ], + "angle": 0, + "content": "Once trained, the TokenAdapt module is used to augment a sequence of tokens \\( T \\) into an augmented sequence \\( T^{\\mathbf{A}} \\) across different datasets and downstream tasks. Importantly, the augmentation-compatible feature space, which does not rely on specific semantic knowledge or alignment with downstream tasks, makes it dataset and task-agnostic. The inherent flexibility of TokenAdapt makes it highly generalizable, allowing direct application to other datasets without the need for additional training procedures." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.548, + 0.784, + 0.684 + ], + "angle": 0, + "content": "ColorAdapt. As witnessed in Cubuk et al. [12], color-based augmentations are often more effective than geometric ones in certain datasets and are an important component of data augmentation strategies. For example, color-based augmentations are useful for dealing with various environmental changes caused by different lighting conditions (e.g., day and night) or weather conditions (e.g., snow). However, integrating them into TokenAdapt is challenging because most color-based augmentations are designed to work only in the pixel domain. To address this, we propose ColorAdapt, a color-based token augmentation inspired by the method [27]. Formally, the ColorAdapt function \\(\\mathcal{C}\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.327, + 0.697, + 0.785, + 0.729 + ], + "angle": 0, + "content": "\\[\n\\mathcal {C} \\left(Z _ {T _ {1}}, Z _ {T _ {2}}\\right) = \\sigma \\left(Z _ {T _ {2}}\\right) \\left(\\frac {Z _ {T _ {1}} - \\mu \\left(Z _ {T _ {1}}\\right)}{\\sigma \\left(Z _ {T _ {1}}\\right)}\\right) + \\mu \\left(Z _ {T _ {2}}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.783, + 0.764 + ], + "angle": 0, + "content": "where \\(\\mu(Z_T)\\) and \\(\\sigma(Z_T)\\) are the mean and standard deviation of each token embedding \\(Z_T\\) across its spatial dimension, computed for each channel." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.785, + 0.84 + ], + "angle": 0, + "content": "ColorAdapt mimics color-changing augmentations in the pixel domain while preserving the encoded object structure by transforming the global statistics within each channel of the token embedding. Figure 5 shows the images decoded from various color-augmented tokens. This illustrates how ColorAdapt affects the visual properties of the augmented tokens. By incorporating ColorAdapt" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.401, + 0.127 + ], + "angle": 0, + "content": "Lee and Park et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.292, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.233, + 0.203, + 0.276, + 0.213 + ], + "angle": 0, + "content": "Input data" + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.146, + 0.36, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.303, + 0.203, + 0.348, + 0.212 + ], + "angle": 0, + "content": "Brightness" + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.146, + 0.428, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.371, + 0.203, + 0.418, + 0.212 + ], + "angle": 0, + "content": "Emb-Noise" + }, + { + "type": "image", + "bbox": [ + 0.429, + 0.146, + 0.503, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.442, + 0.203, + 0.491, + 0.213 + ], + "angle": 0, + "content": "ColorAdapt" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.146, + 0.576, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.203, + 0.561, + 0.213 + ], + "angle": 0, + "content": "Input data" + }, + { + "type": "image", + "bbox": [ + 0.577, + 0.146, + 0.645, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.587, + 0.203, + 0.632, + 0.213 + ], + "angle": 0, + "content": "Brightness" + }, + { + "type": "image", + "bbox": [ + 0.646, + 0.146, + 0.709, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.655, + 0.203, + 0.703, + 0.213 + ], + "angle": 0, + "content": "Emb-Noise" + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.146, + 0.787, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.726, + 0.203, + 0.775, + 0.213 + ], + "angle": 0, + "content": "ColorAdapt" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.226, + 0.788, + 0.298 + ], + "angle": 0, + "content": "Fig. 5: ColorAdapt provides more reasonable color changes. We present ViT-VQGAN decoded images to verify the quality of tokenizations after color changes. We use the brightness function with a factor of 0.2 following the implementation [46]. Emb-Noise [44] is the color-based token augmentation. Notably, our ColorAdapt effectively preserves object structure in contrast to the failure of the counterparts." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.315, + 0.789, + 0.36 + ], + "angle": 0, + "content": "Table 1: Storage-efficient ImageNet-1k classification. We report top-1 accuracies (ViT-B) and compression ratio on ImageNet-1k using various data storage reduction methods. Our SeiT++ shows significant storage efficiency compared to competitors." + }, + { + "type": "table", + "bbox": [ + 0.233, + 0.371, + 0.769, + 0.548 + ], + "angle": 0, + "content": "
MethodInput # of images Top 1 Acc.Dataset storage size
Full pixelsImage1,281 k81.8140.0 GB
Uniform random sampling512 k74.054.6 GB (39%)
C-score [29] based samplingImage512 k73.353.3 GB (38%)
Adjusting JPEG quality factor to 51,281 k74.611.0 GB (8%)
Adjusting image resolution (× 0.2)1,281 k75.29.6 GB (7%)
SRe2L [65]Image200 k65.92.6 GB (2%)
G-VBSM [52]100 k63.71.6 GB (1%)
SeiT [44]Token1,281 k74.01.4 GB (1%)
SeiT++1,281 k77.81.4 GB (1%)
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.58, + 0.788, + 0.611 + ], + "angle": 0, + "content": "into our token augmentation strategy, we improve the data diversity, exploring new possibilities for training more robust and generalizable vision models." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.646, + 0.367, + 0.663 + ], + "angle": 0, + "content": "4 Experiment" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.841 + ], + "angle": 0, + "content": "In this section, we validate the effectiveness of SeiT++ on various scenarios. First, we evaluate the performance of our method in the context of storage-efficient ImageNet-1k classification and transfer learning on fine-grained datasets. Then, we demonstrate the adaptability of token-based vision training to ADE-20k semantic segmentation, showing additional improvements facilitated by SeiT++. In addition, we present experimental results illustrating how our strategies contribute to enhanced model generalizability in robustness evaluation. We also explore the applicability of our augmentation strategies to alternative input formats, demonstrating its extensibility. Unless otherwise stated, SeiT++ refers to the combination of MTM and our token augmentation strategies." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.276, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "SeiT++: Masked Token Modeling Improves Storage-efficient Training" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.189 + ], + "angle": 0, + "content": "Table 2: Token-based ImageNet-1k Classification. We report top-1 accuracies (ViT-B) on ImageNet-1k. Note that SeiT++ w/o MTM means the original SeiT [44] training with our token augmentation strategies." + }, + { + "type": "table", + "bbox": [ + 0.235, + 0.198, + 0.77, + 0.326 + ], + "angle": 0, + "content": "
StorageSeiTSeiT++
w/o MTMw/ MTMw/o MTMw/ MTM
1.4 GB74.075.175.5 (+1.5)77.8 (+2.7)
1.1 GB70.274.173.1 (+2.9)76.5 (+2.4)
0.8 GB66.370.669.1 (+2.8)74.1 (+3.5)
0.5 GB59.764.563.8 (+4.1)70.0 (+5.5)
0.3 GB47.253.951.2 (+4.0)60.6 (+6.7)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.337, + 0.788, + 0.38 + ], + "angle": 0, + "content": "Table 3: Fine-grained classification. We report top-1 accuracies (ViT-B) on various fine-grained datasets. Our method consistently outperforms its counterpart through effective augmentations for tokens." + }, + { + "type": "table", + "bbox": [ + 0.274, + 0.391, + 0.733, + 0.466 + ], + "angle": 0, + "content": "
MethodFlowersCarsiNat18iNat19
SeiT [44]93.579.743.150.1
SeiT++95.286.952.458.9
(+1.7)(+7.2)(+9.3)(+8.8)
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.493, + 0.507, + 0.509 + ], + "angle": 0, + "content": "4.1 Storage-efficient ImageNet-1k" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.519, + 0.794, + 0.732 + ], + "angle": 0, + "content": "Table 1 summarizes the performance and compression ratio of storage-efficient ImageNet-1k classification. Our method is compared with image-based approaches, including uniform random sampling, C-score [29]-based sampling, adjusting image resolution, and adjusting JPEG quality factor. Despite a \\(1\\%\\) compression ratio, when a ViT-B/16 model is pre-trained on tokenized ImageNet-1k using MTM and the ViT model is fine-tuned, our method achieves a \\(77.8\\%\\) top-1 accuracy, outperforming various image-based approaches. In Table 2, our method consistently outperforms the token-based learning baseline SeiT by \\(1.5 - 6.7\\% p\\) in top-1 accuracies. Notably, the effectiveness of our method becomes more pronounced as the available storage size decreases, highlighting its efficiency in low-storage scenarios. Furthermore, SeiT++ shows more significant performance gains with MTM. For a storage size of 1.4GB, SeiT++ achieves an improvement of \\(1.5\\% p\\) over SeiT without MTM, while with MTM the improvement increases to \\(2.7\\% p\\). These results confirm that SeiT++ effectively takes advantage of MTM." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.754, + 0.482, + 0.77 + ], + "angle": 0, + "content": "4.2 Fine-grained Classification" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "To further investigate the generalizability of SeiT++, we conduct experiments on four different fine-grained datasets: Flowers [43], StandfordCars [31], iNat-18 [57], and iNat-19 [58]. Specifically, we tokenize each dataset and then fine-tune the token-trained model (ViT-B) on ImageNet-1k with the tokenized dataset. Since" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.401, + 0.128 + ], + "angle": 0, + "content": "Lee and Park et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 4: ADE-20k Semantic Segmentation. We report mIoU on the ADE-20k validation set, highlighting the effectiveness of our approach in pixel-level classification." + }, + { + "type": "table", + "bbox": [ + 0.288, + 0.185, + 0.716, + 0.23 + ], + "angle": 0, + "content": "
MethodBackboneSeiTSeiT++
UPerNet [63]ViT-B [17]39.043.2 (+4.2)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.244, + 0.787, + 0.286 + ], + "angle": 0, + "content": "Table 5: Robustness evaluation. We report top-1 robust accuracies (ViT-B) against corruptions and domain shifts for each model trained on ImageNet-1k. Our method consistently outperforms its counterpart on various robustness benchmarks." + }, + { + "type": "table", + "bbox": [ + 0.235, + 0.298, + 0.769, + 0.366 + ], + "angle": 0, + "content": "
MethodCleanGauss. NoiseGauss. BlurImageNet-RSketchObjectNet
SeiT [44]74.050.762.625.522.615.7
SeiT++75.558.666.830.227.718.3
(+1.5)(+7.9)(+4.2)(+4.7)(+5.1)(+2.6)
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.395, + 0.788, + 0.501 + ], + "angle": 0, + "content": "fine-grained datasets require the classification of objects with intricate characteristics, they are particularly sensitive to undesirable noise (i.e., RRC in SeiT) or excessive perturbation (i.e., Emb-Noise) resulting from data augmentation in the token embedding space. Table 3 shows the quantitative performance (top-1 accuracies) on fine-grained datasets. SeiT++ consistently outperforms the performance of SeiT across various datasets. These results confirm the effectiveness of our proposed method in fine-grained classification." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.524, + 0.538, + 0.54 + ], + "angle": 0, + "content": "4.3 ADE-20k Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.549, + 0.789, + 0.702 + ], + "angle": 0, + "content": "A limitation of token-based learning is often substantiated by learning dense prediction tasks. We conjecture that our method could facilitate training for those tasks, such as semantic segmentation. We extract tokens from ADE-20k [72] and use the pre-trained ViT-B/16 model on tokenized ImageNet-1k to evaluate the effectiveness of our method in semantic segmentation. We then fine-tune the pre-trained model using the tokenized ADE-20k dataset. Table 4 reports the mean intersection-over-union (mIoU) as an evaluation metric. The results show a significant performance improvement \\((+4.2\\%)\\) p mIoU) over SeiT. These results indicate that SeiT++ significantly improves model performance in the token-based dense prediction task." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.724, + 0.553, + 0.739 + ], + "angle": 0, + "content": "4.4 Evaluation on Robustness Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To validate the effectiveness of our augmentation strategies in terms of robustness, we further evaluate the robustness of our method on a set of benchmarks such as adding Gaussian noise [26] or Gaussian blur [26], ImageNet-R [25], Sketch [59], and ObjectNet [3]. Notably, we adopt SeiT++ w/o MTM to validate the impact of our augmentation strategies on model robustness. The results in Table 5 show that SeiT++ consistently outperforms SeiT in terms of robustness" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.276, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "SeiT++: Masked Token Modeling Improves Storage-efficient Training" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.203 + ], + "angle": 0, + "content": "Table 6: Impact of the proposed augmentation strategies. We report the top-1 accuracies (ViT-S) on the ImageNet-100 validation set for various combinations of our token augmentation strategies. Notably, our TokenAdapt and ColorAdapt consistently improve model performance, exhibiting synergy when used together." + }, + { + "type": "table", + "bbox": [ + 0.339, + 0.213, + 0.665, + 0.3 + ], + "angle": 0, + "content": "
FrameworkColorAdaptTokenAdaptTop 1 Acc.
SeiT [44]XX77.3
✓X78.3
X✓80.4
✓✓81.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.313, + 0.789, + 0.357 + ], + "angle": 0, + "content": "Table 7: Adaptibility of our method to different input. Input denotes tokenizers for our model. We report the top-1 accuracies (ViT-S) on ImageNet-100. For VQGAN, we used a publicly available VQGAN [18] trained on OpenImages [32] as a tokenizer." + }, + { + "type": "table", + "bbox": [ + 0.348, + 0.367, + 0.657, + 0.44 + ], + "angle": 0, + "content": "
InputSeiTSeiT++
Token (ViT-VQGAN [66])77.381.4 (+4.1)
Token (VQGAN [18])81.883.9 (+2.1)
DCT coefficients69.571.0 (+1.5)
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.468, + 0.788, + 0.53 + ], + "angle": 0, + "content": "accuracy. Previous studies [8, 55] observe that the key to the input pixel robustness depends on the pixel-level augmentations. Since our method enables pixel-based data augmentation to tokens, it achieves higher robustness than the baseline, which does not effectively exploit the pixel-based data augmentation." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.551, + 0.391, + 0.567 + ], + "angle": 0, + "content": "4.5 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.576, + 0.788, + 0.668 + ], + "angle": 0, + "content": "Impact of Each Augmentation Strategy. In Table 6, we report the impact of proposed token augmentations. In particular, incorporating ColorAdapt leads to a \\(1.0\\%\\) p performance improvement, highlighting the positive effect of preserving object structure during model training. Additionally, incorporating TokenAdapt yields an extra \\(3.1\\%\\) p improvement. These results confirm that exploring token augmentations is a key factor in fully exploiting pre-extracted tokens." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Extensibility of Our Augmentation Strategy. In this section, we conjecture the broad applicability of our augmentation strategies across various scenarios. To explore its extensibility, we delve into two distinct situations. Firstly, we examine the integration of our method with an alternative tokenizer—the OpenImages-trained VQGAN tokenizer [18], alongside the ImageNet-1k-trained ViT-VQGAN tokenizer [66]. The results in Table 7 consistently demonstrate improved performance on ImageNet-100 across different tokenizers. Additionally, we investigate the adaptability of our approach to a different input format by focusing on Discrete Cosine Transform (DCT) coefficients. Inspired by JPEG compression, we apply DCT to each \\(8 \\times 8\\) image patch, yielding quantized DCT" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.401, + 0.128 + ], + "angle": 0, + "content": "Lee and Park et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "coefficients utilized as input data during model training. As shown in Table 7, our approach enhances top-1 accuracy, similar to the tokenizer scenario. This consistent improvement across diverse input formats underscores the universal applicability of our method, highlighting its potential as a widely applicable solution for diverse input representations." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.242, + 0.458, + 0.258 + ], + "angle": 0, + "content": "4.6 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.263, + 0.788, + 0.385 + ], + "angle": 0, + "content": "We use the pre-trained ViT-VQGAN tokenizer [66] for token extraction. For our TokenAdapt module, we employ a single transformer block for conversion and reverse function. The TokenAdapt module is trained for one epoch on ImageNet-1k with a batch size of 128, a learning rate of 0.001 with cosine scheduling, and a weight decay of 0.05. We follow the training recipes from MAGE [34], SeiT [44], and mmsegmentation [10] for masked token modeling (MTM), token-based classification, and semantic segmentation, respectively. Detailed hyperparameters and settings are available in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.404, + 0.36, + 0.421 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.433, + 0.788, + 0.676 + ], + "angle": 0, + "content": "In this paper, we investigate the compatibility of SeiT with self-supervised learning approaches in token-based frameworks. Among various self-supervised algorithms, Masked Image Modeling (MIM) [24,64] has recently gained popularity for learning visual representations. Inspired by this, we extend SeiT by integrating Masked Token Modeling (MTM) for self-supervised pre-training. Additionally, recognizing that self-supervised approaches often require more data due to the lack of labels, we propose simple yet effective token-based augmentation strategies: TokenAdapt and ColorAdapt. TokenAdapt aligns the token embedding space with another embedding space compatible with pixel-based data augmentations. In addition, we introduce ColorAdapt, a color-changing augmentation for tokens. Our experimental results across various scenarios, including storage-efficient ImageNet-1k classification, fine-grained classification, ADE-20k semantic segmentation, and robustness benchmarks, consistently validate the effectiveness of our proposed method. Furthermore, the consistent performance improvement across diverse input formats highlights the potential of our approach as a widely applicable solution for various input representations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.766 + ], + "angle": 0, + "content": "Limitations. While SeiT++ is capable of handling large-scale data without labels with optimized storage requirements, experiments on such large datasets (e.g., ImageNet-21k [14]) and huge models are not contained in this paper due to resource constraints. We believe the advantages of MTM and our data augmentation strategies can be seamlessly integrated with scalability." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Acknowledgements. Experiments are based on the Naver Smart Machine Learning NSML [30] platform. This research was supported by IITP grant funded by the Korea government(MSIT) and KEIT grant funded by the Korea government(MOTIE) (No. 2022-0-00680, 2022-0-01045)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.277, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SeiT++: Masked Token Modeling Improves Storage-efficient Training" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.176, + 0.785, + 0.218 + ], + "angle": 0, + "content": "1. Bahng, H., Chun, S., Yun, S., Choo, J., Oh, S.J.: Learning de-biased representations with biased representations. In: International Conference on Machine Learning. pp. 528-539. PMLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.219, + 0.785, + 0.246 + ], + "angle": 0, + "content": "2. Bao, H., Dong, L., Piao, S., Wei, F.: Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.246, + 0.785, + 0.3 + ], + "angle": 0, + "content": "3. Barbu, A., Mayo, D., Alverio, J., Luo, W., Wang, C., Gutfreund, D., Tenenbaum, J., Katz, B.: Objectnet: A large-scale bias-controlled dataset for pushing the limits of object recognition models. Advances in neural information processing systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.301, + 0.785, + 0.355 + ], + "angle": 0, + "content": "4. Chang, H., Zhang, H., Barber, J., Maschinot, A., Lezama, J., Jiang, L., Yang, M.H., Murphy, K., Freeman, W.T., Rubinstein, M., et al.: Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.357, + 0.785, + 0.398 + ], + "angle": 0, + "content": "5. Chang, H., Zhang, H., Jiang, L., Liu, C., Freeman, W.T.: Maskgit: Masked generative image transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11315-11325 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.399, + 0.785, + 0.44 + ], + "angle": 0, + "content": "6. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International conference on machine learning. pp. 1597-1607. PMLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.44, + 0.785, + 0.481 + ], + "angle": 0, + "content": "7. Chen, X., Xie, S., He, K.: An empirical study of training self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9640-9649 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.481, + 0.785, + 0.522 + ], + "angle": 0, + "content": "8. Chun, S., Oh, S.J., Yun, S., Han, D., Choe, J., Yoo, Y.: An empirical evaluation on robustness and uncertainty of regularization methods. arXiv preprint arXiv:2003.03879 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.522, + 0.785, + 0.564 + ], + "angle": 0, + "content": "9. Coleman, C., Yeh, C., Mussmann, S., Mirzasoleiman, B., Bailis, P., Liang, P., Leskovec, J., Zaharia, M.: Selection via proxy: Efficient data selection for deep learning. arXiv preprint arXiv:1906.11829 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.565, + 0.785, + 0.592 + ], + "angle": 0, + "content": "0. Contributors, M.: MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark. https://github.com/open-mmlab/mmsegmentation (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.592, + 0.785, + 0.633 + ], + "angle": 0, + "content": "1. Croce, F., Hein, M.: Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In: International conference on machine learning. pp. 2206-2216. PMLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.633, + 0.785, + 0.675 + ], + "angle": 0, + "content": "2. Cubuk, E.D., Zoph, B., Mane, D., Vasudevan, V., Le, Q.V.: Autoaugment: Learning augmentation strategies from data. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 113-123 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.675, + 0.785, + 0.729 + ], + "angle": 0, + "content": "3. Cubuk, E.D., Zoph, B., Shlens, J., Le, Q.V.: Randaugment: Practical automated data augmentation with a reduced search space. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 702-703 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.729, + 0.785, + 0.771 + ], + "angle": 0, + "content": "4. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.771, + 0.785, + 0.813 + ], + "angle": 0, + "content": "5. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "6. DeVries, T., Taylor, G.W.: Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552 (2017)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.176, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.401, + 0.127 + ], + "angle": 0, + "content": "Lee and Park et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.204 + ], + "angle": 0, + "content": "17. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.205, + 0.787, + 0.247 + ], + "angle": 0, + "content": "18. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12873-12883 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.248, + 0.787, + 0.275 + ], + "angle": 0, + "content": "19. Fang, Y., Dong, L., Bao, H., Wang, X., Wei, F.: Corrupted image modeling for self-supervised visual pre-training. arXiv preprint arXiv:2202.03382 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.276, + 0.787, + 0.303 + ], + "angle": 0, + "content": "20. Ford, N., Gilmer, J., Carlini, N., Cubuk, D.: Adversarial examples are a natural consequence of test error in noise. arXiv preprint arXiv:1901.10513 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.304, + 0.787, + 0.346 + ], + "angle": 0, + "content": "21. Geirhos, R., Rubisch, P., Michaelis, C., Bethge, M., Wichmann, F.A., Brendel, W.: Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. arXiv preprint arXiv:1811.12231 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.346, + 0.787, + 0.374 + ], + "angle": 0, + "content": "22. Goodfellow, I.J., Shlens, J., Szegedy, C.: Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.375, + 0.787, + 0.431 + ], + "angle": 0, + "content": "23. Grill, J.B., Strub, F., Altché, F., Tallec, C., Richemond, P., Buchatskaya, E., Doersch, C., Avila Pires, B., Guo, Z., Gheshlaghi Azar, M., et al.: Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems 33, 21271-21284 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.432, + 0.787, + 0.473 + ], + "angle": 0, + "content": "24. He, K., Chen, X., Xie, S., Li, Y., Dollár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.474, + 0.787, + 0.529 + ], + "angle": 0, + "content": "25. Hendrycks, D., Basart, S., Mu, N., Kadavath, S., Wang, F., Dorundo, E., Desai, R., Zhu, T., Parajuli, S., Guo, M., Song, D., Steinhardt, J., Gilmer, J.: The many faces of robustness: A critical analysis of out-of-distribution generalization. ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.53, + 0.787, + 0.558 + ], + "angle": 0, + "content": "26. Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.559, + 0.787, + 0.601 + ], + "angle": 0, + "content": "27. Huang, X., Belongie, S.: Arbitrary style transfer in real-time with adaptive instance normalization. In: Proceedings of the IEEE international conference on computer vision. pp. 1501-1510 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.602, + 0.787, + 0.657 + ], + "angle": 0, + "content": "28. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International conference on machine learning. pp. 4904-4916. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.658, + 0.787, + 0.699 + ], + "angle": 0, + "content": "29. Jiang, Z., Zhang, C., Talwar, K., Mozer, M.C.: Characterizing structural regularities of labeled data in overparameterized models. arXiv preprint arXiv:2002.03206 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.7, + 0.787, + 0.742 + ], + "angle": 0, + "content": "30. Kim, H., Kim, M., Seo, D., Kim, J., Park, H., Park, S., Jo, H., Kim, K., Yang, Y., Kim, Y., et al.: Nsml: Meet the mlaas platform with a real-world case study. arXiv preprint arXiv:1810.09957 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.743, + 0.787, + 0.785 + ], + "angle": 0, + "content": "31. Krause, J., Stark, M., Deng, J., Fei-Fei, L.: 3d object representations for fine-grained categorization. In: Proceedings of the IEEE international conference on computer vision workshops. pp. 554-561 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.787, + 0.841 + ], + "angle": 0, + "content": "32. Kuznetsova, A., Rom, H., Alldrin, N., Uijlings, J., Krasin, I., Pont-Tuset, J., Kamali, S., Popov, S., Malloci, M., Kolesnikov, A., et al.: The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision 128(7), 1956-1981 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.277, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SeiT++: Masked Token Modeling Improves Storage-efficient Training" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.786, + 0.189 + ], + "angle": 0, + "content": "33. Lee, S., Chun, S., Jung, S., Yun, S., Yoon, S.: Dataset condensation with contrastive signals. In: International Conference on Machine Learning. pp. 12352-12364. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.189, + 0.787, + 0.244 + ], + "angle": 0, + "content": "34. Li, T., Chang, H., Mishra, S., Zhang, H., Katabi, D., Krishnan, D.: Mage: Masked generative encoder to unify representation learning and image synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2142-2152 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.244, + 0.786, + 0.298 + ], + "angle": 0, + "content": "35. Lin, S., Ji, R., Yan, C., Zhang, B., Cao, L., Ye, Q., Huang, F., Doermann, D.: Towards optimal structured cnn pruning via generative adversarial learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2790-2799 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.298, + 0.786, + 0.325 + ], + "angle": 0, + "content": "36. Liu, J., Guo, J., Xu, D.: Apsnet: Toward adaptive point sampling for efficient 3d action recognition. IEEE Transactions on Image Processing 31, 5287-5302 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.325, + 0.786, + 0.379 + ], + "angle": 0, + "content": "37. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.786, + 0.42 + ], + "angle": 0, + "content": "38. Lopes, R.G., Yin, D., Poole, B., Gilmer, J., Cubuk, E.D.: Improving robustness without sacrificing accuracy with patch gaussian augmentation. arXiv preprint arXiv:1906.02611 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.42, + 0.786, + 0.473 + ], + "angle": 0, + "content": "39. Ma, Y., Li, H., Zhang, Z., Guo, J., Zhang, S., Gong, R., Liu, X.: Annealing-based label-transfer learning for open world object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11454-11463 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.786, + 0.502 + ], + "angle": 0, + "content": "40. Madry, A., Makelov, A., Schmidt, L., Tsipras, D., Vladu, A.: Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.502, + 0.786, + 0.555 + ], + "angle": 0, + "content": "41. Mahajan, D., Girshick, R., Ramanathan, V., He, K., Paluri, M., Li, Y., Bharambe, A., Van Der Maaten, L.: Exploring the limits of weakly supervised pretraining. In: Proceedings of the European conference on computer vision (ECCV). pp. 181-196 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.555, + 0.786, + 0.597 + ], + "angle": 0, + "content": "42. Mei, Z., Ye, P., Ye, H., Li, B., Guo, J., Chen, T., Ouyang, W.: Automatic loss function search for adversarial unsupervised domain adaptation. IEEE Transactions on Circuits and Systems for Video Technology 33(10), 5868-5881 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.597, + 0.786, + 0.638 + ], + "angle": 0, + "content": "43. Nilsback, M.E., Zisserman, A.: Automated flower classification over a large number of classes. In: 2008 Sixth Indian conference on computer vision, graphics & image processing. pp. 722-729. IEEE (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.638, + 0.786, + 0.665 + ], + "angle": 0, + "content": "44. Park, S., Chun, S., Heo, B., Kim, W., Yun, S.: Seit: Storage-efficient vision training with tokens using \\(1\\%\\) of pixel storage. arXiv preprint arXiv:2303.11114 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.664, + 0.786, + 0.704 + ], + "angle": 0, + "content": "45. Paul, M., Ganguli, S., Dziugaite, G.K.: Deep learning on a data diet: Finding important examples early in training. Advances in Neural Information Processing Systems 34, 20596-20607 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.704, + 0.786, + 0.759 + ], + "angle": 0, + "content": "46. Riba, E., Mishkin, D., Ponsa, D., Rublee, E., Bradski, G.: Kornia: an open source differentiable computer vision library for pytorch. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3674-3683 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.786, + 0.8 + ], + "angle": 0, + "content": "47. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10684-10695 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.786, + 0.841 + ], + "angle": 0, + "content": "48. Rosasco, A., Carta, A., Cossu, A., Lomonaco, V., Bacciu, D.: Distilled replay: Overcoming forgetting through synthetic samples. In: International Workshop on Continual Semi-Supervised Learning. pp. 104-117. Springer (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.401, + 0.127 + ], + "angle": 0, + "content": "Lee and Park et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.188 + ], + "angle": 0, + "content": "49. Sangermano, M., Carta, A., Cossu, A., Bacciu, D.: Sample condensation in online continual learning. In: 2022 International Joint Conference on Neural Networks (IJCNN). pp. 01-08. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.189, + 0.785, + 0.243 + ], + "angle": 0, + "content": "50. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.244, + 0.785, + 0.284 + ], + "angle": 0, + "content": "51. Scimeca, L., Oh, S.J., Chun, S., Poli, M., Yun, S.: Which shortcut cues will dnns choose? a study from the parameter-space perspective. arXiv preprint arXiv:2110.03095 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.285, + 0.785, + 0.338 + ], + "angle": 0, + "content": "52. Shao, S., Yin, Z., Zhou, M., Zhang, X., Shen, Z.: Generalized large-scale data condensation via various backbone and statistical matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16709-16718 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.339, + 0.785, + 0.393 + ], + "angle": 0, + "content": "53. Singh, M., Gustafson, L., Adcock, A., de Freitas Reis, V., Gedik, B., Kosaraju, R.P., Mahajan, D., Girshick, R., Dollar, P., Van Der Maaten, L.: Revisiting weakly supervised pre-training of visual perception models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 804-814 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.394, + 0.785, + 0.434 + ], + "angle": 0, + "content": "54. Steiner, A., Kolesnikov, A., Zhai, X., Wightman, R., Uszkoreit, J., Beyer, L.: How to train your vit? data, augmentation, and regularization in vision transformers. arXiv preprint arXiv:2106.10270 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.435, + 0.785, + 0.474 + ], + "angle": 0, + "content": "55. Taori, R., Dave, A., Shankar, V., Carlini, N., Recht, B., Schmidt, L.: Measuring robustness to natural distribution shifts in image classification. Advances in Neural Information Processing Systems 33, 18583-18599 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.476, + 0.785, + 0.515 + ], + "angle": 0, + "content": "56. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jégou, H.: Training data-efficient image transformers & distillation through attention. In: International conference on machine learning. pp. 10347-10357. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.516, + 0.785, + 0.555 + ], + "angle": 0, + "content": "57. Van Horn, G., Mac Aodha, O., Song, Y., Cui, Y., Sun, C., Shepard, A., Adam, H., Perona, P., Belongie, S.: The inaturalist challenge 2018 dataset. arXiv preprint arXiv:1707.06642 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.557, + 0.785, + 0.596 + ], + "angle": 0, + "content": "58. Van Horn, G., Mac Aodha, O., Song, Y., Cui, Y., Sun, C., Shepard, A., Adam, H., Perona, P., Belongie, S.: The inaturalist challenge 2019 dataset. arXiv preprint arXiv:1707.06642 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.597, + 0.785, + 0.637 + ], + "angle": 0, + "content": "59. Wang, H., Ge, S., Lipton, Z., Xing, E.P.: Learning robust global representations by penalizing local predictive power. In: Advances in Neural Information Processing Systems. pp. 10506-10518 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.638, + 0.785, + 0.664 + ], + "angle": 0, + "content": "60. Wang, T., Zhu, J.Y., Torralba, A., Efros, A.A.: Dataset distillation. arXiv preprint arXiv:1811.10959 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.665, + 0.785, + 0.691 + ], + "angle": 0, + "content": "61. Wei, J., Zou, K.: Eda: Easy data augmentation techniques for boosting performance on text classification tasks. arXiv preprint arXiv:1901.11196 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.692, + 0.785, + 0.718 + ], + "angle": 0, + "content": "62. Wightman, R., Touvron, H., Jégou, H.: Resnet strikes back: An improved training procedure in timm. arXiv preprint arXiv:2110.00476 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.719, + 0.785, + 0.759 + ], + "angle": 0, + "content": "63. Xiao, T., Liu, Y., Zhou, B., Jiang, Y., Sun, J.: Unified perceptual parsing for scene understanding. In: Proceedings of the European conference on computer vision (ECCV). pp. 418-434 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.785, + 0.799 + ], + "angle": 0, + "content": "64. Xie, Z., Zhang, Z., Cao, Y., Lin, Y., Bao, J., Yao, Z., Dai, Q., Hu, H.: Simmim: A simple framework for masked image modeling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9653-9663 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "65. Yin, Z., Xing, E., Shen, Z.: Squeeze, recover and relabel: Dataset condensation at imagenet scale from a new perspective. In: Thirty-seventh Conference on Neural Information Processing Systems (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.277, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SeiT++: Masked Token Modeling Improves Storage-efficient Training" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "66. Yu, J., Li, X., Koh, J.Y., Zhang, H., Pang, R., Qin, J., Ku, A., Xu, Y., Baldridge, J., Wu, Y.: Vector-quantized image modeling with improved vqgan. arXiv preprint arXiv:2110.04627 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.788, + 0.231 + ], + "angle": 0, + "content": "67. Yun, S., Han, D., Oh, S.J., Chun, S., Choe, J., Yoo, Y.: Cutmix: Regularization strategy to train strong classifiers with localizable features. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6023-6032 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.231, + 0.788, + 0.259 + ], + "angle": 0, + "content": "68. Zhang, H., Cisse, M., Dauphin, Y.N., Lopez-Paz, D.: mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.259, + 0.788, + 0.3 + ], + "angle": 0, + "content": "69. Zhao, B., Bilen, H.: Dataset condensation with differentiable siamese augmentation. In: International Conference on Machine Learning. pp. 12674-12685. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.3, + 0.788, + 0.328 + ], + "angle": 0, + "content": "70. Zhao, B., Mopuri, K.R., Bilen, H.: Dataset condensation with gradient matching. arXiv preprint arXiv:2006.05929 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.328, + 0.788, + 0.37 + ], + "angle": 0, + "content": "71. Zhong, Z., Zheng, L., Kang, G., Li, S., Yang, Y.: Random erasing data augmentation. In: Proceedings of the AAAI Conference on Artificial Intelligence (AAAI) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.37, + 0.788, + 0.411 + ], + "angle": 0, + "content": "72. Zhou, B., Zhao, H., Puig, X., Xiao, T., Fidler, S., Barriuso, A., Torralba, A.: Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision 127, 302-321 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.411 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/SeiT++_ Masked Token Modeling Improves Storage-efficient Training/a6a99e76-474a-47eb-80d4-d9020aeb146a_origin.pdf b/2024/SeiT++_ Masked Token Modeling Improves Storage-efficient Training/a6a99e76-474a-47eb-80d4-d9020aeb146a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c872994600616da5e3353f225ef74ba2be00eb30 --- /dev/null +++ b/2024/SeiT++_ Masked Token Modeling Improves Storage-efficient Training/a6a99e76-474a-47eb-80d4-d9020aeb146a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:87d9cc9ccc463c35ba215a19b838098b73f7820311ca643b5488340cf5202d31 +size 1347715 diff --git a/2024/SeiT++_ Masked Token Modeling Improves Storage-efficient Training/full.md b/2024/SeiT++_ Masked Token Modeling Improves Storage-efficient Training/full.md new file mode 100644 index 0000000000000000000000000000000000000000..92a70b989db139b6de3d259dd1fee3c053b1bf4c --- /dev/null +++ b/2024/SeiT++_ Masked Token Modeling Improves Storage-efficient Training/full.md @@ -0,0 +1,303 @@ +# SeiT++: Masked Token Modeling Improves Storage-efficient Training + +Minhyun Lee $^{1,\dagger}$ Song Park $^{2,\dagger}$ Byeongho Heo $^{2}$ Dongyoon Han $^{2}$ Hyunjung Shim $^{3,\star}$ + +$^{1}$ Yonsei University $^{2}$ Naver AI Lab $^{3}$ KAIST AI $^{4}$ Equal contribution. Work done at Naver AI Lab. + +Abstract. Recent advancements in Deep Neural Network (DNN) models have significantly improved performance across computer vision tasks. However, achieving highly generalizable and high-performing vision models requires expansive datasets, resulting in significant storage requirements. This storage challenge is a critical bottleneck for scaling up models. A recent breakthrough by SeiT proposed the use of Vector-Quantized (VQ) feature vectors (i.e., tokens) as network inputs for vision classification. This approach achieved $90\%$ of the performance of a model trained on full-pixel images with only $1\%$ of the storage. While SeiT needs labeled data, its potential in scenarios beyond fully supervised learning remains largely untapped. In this paper, we extend SeiT by integrating Masked Token Modeling (MTM) for self-supervised pre-training. Recognizing that self-supervised approaches often demand more data due to the lack of labels, we introduce TokenAdapt and ColorAdapt. These methods facilitate comprehensive token-friendly data augmentation, effectively addressing the increased data requirements of self-supervised learning. We evaluate our approach across various scenarios, including storage-efficient ImageNet-1k classification, fine-grained classification, ADE-20k semantic segmentation, and robustness benchmarks. Experimental results demonstrate consistent performance improvement in diverse experiments, validating the effectiveness of our method. Code is available at https://github.com/naver-ai/seit. + +# 1 Introduction + +Recent advancements in Deep Neural Network (DNN) [17,56] have significantly improved their performance across various computer vision tasks. However, achieving highly generalizable and high-performing vision models demands an extensive dataset, often containing billions of data points [28,41,53] with large storage requirements. For example, the LAION-5B [50] dataset requires a storage size of 240TB. This storage challenge emerges as a critical bottleneck for scaling up vision models. Consequently, optimizing storage efficiency becomes essential for the practical advancement of scaled-up vision models. + +![](images/048273a0b26487a7811143950997cf5c723f7911823c28f1746567b200c0ae27.jpg) +Fig.1: Over $70\%$ top-1 accuracy is achievable with just 1GB data. On ImageNet-1k, we visualize the trade-off of training data storage vs. top-1 accuracy using the fixed ViT-B/16 for a controlled comparison. Each accuracy metric is individually trained with different data types. While the entire ImageNet-1k dataset requires approximately 140GB for training with images, our approach demonstrates significant storage efficiency over competitors. + +Images require more storage than texts due to their intrinsic complexity in converting continuous natural light signals into discrete pixels for computer representation. However, a considerable portion of image data corresponds to redundant or unnecessary details, such as repeated patterns or imperceptible details [47]. These properties lead to large storage requirements and also induce negative impacts on visual recognition tasks as mentioned in [1,11,21,22,40,51]. Therefore, several attempts [9, 29, 45, 69, 70] have been made to reduce storage requirements by optimizing the efficiency of image data. A standard approach to reduce storage requirements is to decrease the total number of training data by removing less important samples [9, 29, 45] or condensing the data into smaller synthetic sets [69, 70]. + +However, as SeiT [44] illustrates, these approaches may incur a significant performance drop or become infeasible for large-scale datasets due to their complexity. Alternatively, some methods aim to reduce the size of each image through compression techniques or resolution adjustments [50]. However, they still face a considerable performance gap with the original dataset. Recently, SeiT [44] has made a significant breakthrough in this area by enabling the direct learning of ViTs [17] from discrete representations (i.e., tokens) pre-extracted by an offline tokenizer [66]. By using vector-quantized features instead of traditional pixel images, SeiT achieves $90\%$ of the performance of a full storage model while requiring only $1\%$ of the storage. + +While SeiT [44] has demonstrated the high potential of token-based training, a performance gap with the image-based models still persists, and a prospective solution - exploration beyond fully supervised learning scenarios - remains untapped. To leverage the advantages of large-scale data beyond human-annotated datasets, exploring label-free learning methods is imperative. In this paper, we investigate the compatibility of SeiT with self-supervised learning approaches in token-based frameworks. Among many different self-supervised methods [6,7,23], + +Masked Image Modeling (MIM) [24,64] derived from Masked Language Modeling (MLM) [15] has brought significant success and has quickly become a popular approach for learning visual representations. + +Inspired by this, we introduce Masked Token Modeling (MTM), a pre-training method that performs token masking and reconstruction in token-based frameworks, and validate its effectiveness for storage efficiency on various benchmarks. To the best of our knowledge, this is the first approach for learning an MLM-based model directly from offline tokens, offering the potential to optimize storage efficiency for large-scale training tasks. As previous studies [12, 13, 56] have highlighted, data augmentation plays a crucial role in large-scale model training and is a fundamental element in various training setups [6, 7, 23, 24, 37, 56, 62, 64]. However, applying existing data augmentation techniques [13, 67, 71] designed for pixel images to tokens poses a significant challenge due to the input domain shift. Therefore, addressing this challenge is key to overcoming major performance bottlenecks and improving the storage efficiency of token-based frameworks. + +To this end, we introduce TokenAdapt and ColorAdapt, two novel augmentation strategies designed for token-based training. TokenAdapt facilitates the integration of traditional pixel-based augmentation policies by transforming token embeddings into a feature space compatible with pixel-based augmentations. After applying augmentation to the transformed token embeddings, the augmented token embeddings are then reverted to their original token space. ColorAdapt, inspired by Adaptive Instance Normalization (AdaIN [27]), alters the color attributes of token embeddings while preserving object structure by adjusting their statistics. By integrating TokenAdapt and ColorAdapt into our token augmentation process, we enhance the effectiveness of token-based training across both fully supervised and unsupervised learning scenarios, demonstrating significant improvements in model performance. + +Our SeiT++ framework combines Masked Token Modeling (MTM) with the proposed token augmentation strategies, TokenAdapt and ColorAdapt. We validate the effectiveness of SeiT++ on various scenarios: (1) storage-efficient ImageNet-1k classification, (2) transfer learning on fine-grained classification, (3) ADE-20k semantic segmentation, and (4) evaluation on robustness benchmarks. As shown in Figure 1, SeiT++ consistently outperforms comparison methods in storage-efficient ImageNet-1k classification with limited storage sizes, achieving a top-1 accuracy of $77.8\%$ with only 1.4GB of data. These results confirm that token-based training can take advantage of self-supervised learning to improve storage efficiency. We also demonstrate that our proposed token augmentations significantly contribute to storage efficiency by augmenting data effectively. We reveal that our token-based augmentation methods basically yield significant boosts in supervised learning and further amplify improvements in MTM learning. Furthermore, we present how our strategies enhance the generalizability of trained models, as evidenced by our experimental results on fine-grained classification and robustness evaluation. Lastly, the applicability of our method to an alternative input format (e.g., DCT coefficients) demonstrates its extensibility. + +# 2 Related Work + +Learning with Tokenization. The use of discrete tokens for training vision models has recently gained attention, particularly in self-supervised representation learning. BeiT [2] employs the Masked Image Modeling (MIM) framework to recover discrete tokens from masked images and CIM [19] utilizes tokens in Corrupted Image Modeling. MAGE [34] demonstrates the incorporation of tokens from VQGAN [18] for both generative training and representation learning within a unified training framework. However, existing studies mainly focus on self-supervised representation learning and rely on online tokenization, converting images to tokens during each iteration of model training. This approach incurs notable memory consumption and computational costs, limiting the full exploitation of the storage efficiency of discrete tokens. + +Storage-efficient Vision Training. Efficient model learning is a crucial research area, with studies focusing on optimizing target models [35, 36] or minimizing training data and labels [39, 42]. Storage-efficient training aims to reduce the dataset storage size. To this end, dataset distillation (DD) [60] creates a condensed dataset by leveraging the knowledge of the original dataset. Recent research [33, 48, 49, 69, 70] highlights the effectiveness of dataset distillation in improving training efficiency. However, its high computational complexity poses challenges, particularly when applied to large-scale datasets like ImageNet-1k [14]. Sampling-based methods [9, 29, 45] address this by selecting a subset of the most representative samples from the full dataset. Paul et al. [45] identify important samples very early in training. Jiang et al. [29] utilize a consistency score based on statistics collected during training. However, the diverse sample selection is not guaranteed, particularly in low data regimes, leading to suboptimal performance compared to dataset distillation (DD) methods [70]. To tackle these issues, SeiT [44] proposes a storage-efficient vision training framework by leveraging the storage efficiency of tokens. They utilize the ViT-VQGAN [66] tokenizer to extract tokens from images and store them for vision model training. As a result, they outperform other storage-efficient training methods, significantly reducing the dataset storage size. + +Data Augmentation. Data augmentation serves as a fundamental building block to improve the model performance across diverse computer vision tasks. Various data augmentation techniques have been employed for improved performance [16, 71] or model robustness [20, 38]. Mixup [68] introduces convex combinations of pairs of images and their labels to create a smoother decision boundary. CutMix [67] replaces certain regions of an image with a patch from another image, enhancing the model's localization ability. AutoAugment [12] leverages reinforcement learning to automate optimal augmentation policy discovery. RandAugment [13] introduces a potent policy based on random combinations of different data augmentation methods. DeiT [56] emphasizes the crucial role of data augmentation in vision transformers for improved model performance. + +![](images/93f70da2a7709e419b589aaf11d5588d6735a9a08eb5ccbdab7920fa8989e29c.jpg) +Fig. 2: Masked Token Modeling (MTM) pipeline. MTM is a self-supervised learning approach in token-based frameworks. (a) The tokenized dataset is saved on storage before model training. Then, (b) using only the pre-stored tokens, a storage-efficient vision model (MTM) is trained without relying on labeled datasets. + +Previous studies extensively explore the effective data augmentation policies for pixel-based vision model training. However, we observe that pixel-based data augmentations are not well-suited for tokens, posing a critical performance bottleneck. Motivated by this, we tackle the data augmentation challenge, which has not been explored in previous studies for vision model training with tokens. + +# 3 Method + +# 3.1 Preliminary: Storage-efficient Vision Training (SeiT) + +SeiT [44] aims to develop a scalable and high-performance vision classifier with minimal storage requirements. It introduces a storage-efficient vision training framework that enables the direct learning of ViT [17] from discrete representations (i.e., tokens) pre-extracted by an offline tokenizer. In addition, SeiT further improves training by incorporating two token-specific augmentations: Token-EDA, drawing inspiration from easy data augmentation for language [61], and Emb-Noise, which introduces Gaussian noise into discrete representations. By using vector-quantized features instead of pixel images, SeiT retains $90\%$ of the precision of a model trained on full-pixel images while using only $1\%$ of the storage space. SeiT consists of two parts: (1) preparing the token dataset and (2) training a model using the token dataset. + +Preparing the Token Dataset. SeiT extracts tokens using the ImageNet-1k-trained ViT-VQGAN tokenizer [66] to prepare the token dataset for training. As shown in Figure 2(a), each pixel image $I$ is transformed into a sequence of tokens $T = [t_1, t_2, \dots, t_n]$ . Since the tokens are stored in quantized form, each token means an embedding vector in a codebook $\mathcal{Z} = \{z_k \in \mathbb{R}^d | k = 1, \dots, K\}$ . i.e., $t_i \in \mathbb{Q}$ where $\mathbb{Q} = \{1, 2, \dots, K\}$ for a codebook with size $K$ . Here, given the tokens $T$ and the codebook $\mathcal{Z}$ , the corresponding token embeddings $Z_T = [z_{t_1}, z_{t_2}, \dots, z_{t_n}]$ can be computed by lookup process. This conversion compresses each image + +$I \in \mathbb{R}^{H \times W \times 3}$ into a more compact token sequence $T \in \mathbb{R}^n$ , where $n$ is much smaller than $H \times W \times 3$ , thus significantly reducing storage requirements. For example, storing the full-pixel ImageNet-1k requires 140GB, but its tokenized form needs only 1.4GB. Notably, the tokens are stored on storage before model training with the codebook $\mathcal{Z}$ . Thus, during the training phase, the pre-stored tokens are used directly as input instead of the original pixel images. + +Training Pipeline. To train a vision classifier using tokens, SeiT initially loads the token dataset and applies Token-EDA to the tokens. The tokens are then converted to a one-hot format and randomly resized and cropped. Subsequently, the processed one-hot tokens are transformed into token embeddings using a pretrained codebook. Techniques such as CutMix and Embedding-noise are applied to these token embeddings, and the resulting augmented tensor is fed into the model for training. Through the proposed framework, SeiT demonstrates the high potential of token-based training, but it lacks exploration beyond fully supervised learning scenarios. In addition, while token-specific augmentation is proposed, the types of token augmentation are limited, and the analysis of data augmentation compared to pixel-based training is insufficient. + +# 3.2 Masked Token Modeling + +In this section, we introduce Masked Token Modeling (MTM), a pre-training method in token-based frameworks without relying on labeled datasets. MTM shares a similar concept to Masked Language Modeling (MLM) [15], which works by reconstructing original tokens from their masked versions, employing an autoencoding strategy. Figure 2(b) presents an overview of how our MTM operates. + +Masking. We employ a variable masking ratio that follows a truncated normal distribution, as investigated in previous studies [4,5,34]. The tokens are randomly dropped according to the sampled masking ratio, and then the visible, unmasked token embeddings are fed into the model. By processing only a subset of token embeddings in the encoder, we significantly reduce the overall pre-training time and memory consumption, which aligns with the findings of MAE [24]. + +MTM Encoder & Decoder. We use ViT [17] as our encoder because it is able to process both the visible token embeddings and the original token embeddings, seamlessly accommodating their different lengths. During training, the encoder takes the visible token embeddings and outputs latent features that become the decoder input. Before decoding, the latent features are padded to the length of the original token embeddings using mask tokens M. Each mask token is a shared, learned vector indicating the presence of a missing token to be predicted. The decoder then estimates the original tokens $T$ from the padded features. + +![](images/bfff7f2568ee3f97bd988c21ad6ac5254aa5d362922dfaad3a67390dedcc233a.jpg) +(a) Failure cases visualization + +![](images/37f0cc45980f6f5ece0cb7f7ce5a3e901c84c72294edd0d97c36cc4fba1c2563.jpg) +Fig. 3: Data augmentations with tokens. Each ViT-VQGAN decoded image is reconstructed from a given RGB image after undergoing specific data augmentations. We observe that naively adopting these methods results in incorrect tokenization: 1) Token w/ hFlip demonstrates spatial information collapse during tokenization; 2) Token w/ RRC shows interdependence between neighboring token embeddings. We note the reconstructed images fail to preserve the images' details, suggesting that this incurs ineffectiveness of tokenization. Furthermore, we report top-1 accuracies (ViT-B) on ImageNet-1k according to the augmentations applied during training. All the augmentations are de-facto default training setups for vision transformers [37,54,56]. + +![](images/f34aa1f77be4ef74f7d0b9bad29499ebe5a9428bb29c4b47f7478758d29a519f.jpg) + +![](images/371a44b8363fc6f0990aa4993e9f607dd47d1a289ae80827c36e3443b9448f1a.jpg) + +![](images/de59c1cd87ba0ee0c3bff7f9980f770a286ed455da551754ed64ea2337fb05a3.jpg) +(b) Performance degradation + +Training Objective. MTM learns to accurately predict the original tokens of each masked token from the visible tokens. In particular, we use a cross-entropy loss between the decoder output and the ground truth tokens as the training objective, which is denoted by $\mathcal{L}_{\text{recon}}$ : + +$$ +\mathcal {L} _ {\text {r e c o n}} = \mathrm {C E} \left(T _ {M} ^ {\prime}, T _ {M}\right), \tag {1} +$$ + +where $T_M'$ denotes the predictions of the model for the masked tokens, and $T_M$ corresponds to the ground truth tokens for $T_M'$ . Following MAE [24], this loss is optimized only for the masked tokens. + +# 3.3 Data Augmentation for Tokens + +While Masked Token Modeling (MTM) improves the storage efficiency of the token-based framework, there is still a performance bottleneck in improving token-based training. Since self-supervised approaches often require more data due to the lack of labels, exploring data augmentation is important to take full advantage of MTM. In this section, we analyze why existing data augmentations designed for pixel images are ineffective in the token domain. We then present TokenAdapt and ColorAdapt as solutions for token-based data augmentation. + +Limited Data Augmentations for Tokens. We explore the challenges of pixel-based augmentation in the token domain and identify two main reasons: (1) spatial information collapse during tokenization and (2) inter-token dependencies. The discrete tokenization process condenses a $n \times n$ 2D image patch + +![](images/d9b09dd8744d67ef668a6f82106c8a5a3c91181751e91e7fb32529c4fbd1f02b.jpg) +Fig. 4: TokenAdapt processing pipeline. TokenAdapt aims to enhance the compatibility of token embeddings with pixel-based data augmentations by converting them into augmentation-compatible space, applying augmentations, and reverting them back to the original token embedding space. + +into a single $d$ -dimensional 1D vector (i.e., token embedding), causing a collapse of spatial information. This collapse hinders augmentations that rely on spatial details (e.g., horizontal flip). Furthermore, unlike independent image pixels, token embeddings are inter-correlated. This complicates augmentations involving interpolation (e.g., Resize or Mixup [68]), by introducing undesired artifacts. + +Figure 3 shows the limitations of pixel-based augmentations in the token domain. A horizontal Flip (hFlip) distorts both the diagonal line (top row) and the cat's silhouette (bottom row) due to spatial information collapse. This occurs because only the inter-token spatial relationship is flipped, while the inner-token spatial arrangement remains unchanged. Augmentations involving interpolation can lead to undesired artifacts, such as color leakage in the case of Random Resized Crop (RRC), due to interdependence between neighboring embeddings. Therefore, we suspect that the direct application of pixel-based augmentations to token embedding may negatively affect token-based training. + +TokenAdapt. The concept of TokenAdapt is to establish a feature space that allows for conventional pixel-based augmentations while facilitating the conversion and inversion of pre-extracted tokens with minimal computational overhead. This approach offers several advantages for our TokenAdapt. Firstly, it enables the direct reuse of well-designed pixel-based data augmentation policies. Additionally, the augmented token can be easily computed through the conversion-augmentation-reverse process. Finally, it ensures that the feature spaces for both the original and augmented tokens are preserved identically. This allows for the seamless use of existing model architectures or pre-trained weights. + +The goal of TokenAdapt is to design an augmentation function that converts the input tokens $T$ to augmented tokens $T^{\mathbf{A}} = [t_1^{\mathbf{A}}, t_2^{\mathbf{A}}, \dots, t_n^{\mathbf{A}}]$ . Note that the augmented tokens should also be quantized $t_i^{\mathbf{A}} \in \mathbb{Q}$ . Figure 4 depicts the overall pipeline of TokenAdapt. The initial step of TokenAdapt is transforming token embeddings $Z_T$ , computed from $T$ and codebook $\mathcal{Z}$ , into a representation $S_T \in \mathbb{R}^{h \times w \times d}$ in augmentation-compatible space $S$ . This transformation is achieved using a conversion function $f$ , i.e., $S_T = f(Z_T)$ . We apply augmentation $\mathbf{A}$ at augmentation-compatible space $S$ and build an augmented representation $S_T^{\mathbf{A}} = \mathbf{A}(S_T)$ . Then, the augmented representation is reverted to + +token embedding space by a reverse transformation $g$ , $Z_T^{\mathbf{A}} = g(S_T^{\mathbf{A}})$ . At last, the augmented tokens in embedding space $Z_T^{\mathbf{A}}$ are quantized to codebook index as $T^{\mathbf{A}} = \mathbf{q}_{\mathcal{Z}}(Z_T^{\mathbf{A}})$ , where $\mathbf{q}_{\mathcal{Z}}$ indicates the vector quantization process. In summary, the augmentation process of TokenAdapt is as follows: + +$$ +Z _ {T} ^ {\mathbf {A}} = g (\mathbf {A} (f (Z _ {T}))), \quad T ^ {\mathbf {A}} = \mathbf {q} _ {\mathcal {Z}} \left(Z _ {T} ^ {\mathbf {A}}\right). \tag {2} +$$ + +The conversion function $f$ and the reverse transformation $g$ are learned from paired data of tokens $(T_x, T_{\mathbf{A}(x)})$ , where $T_x$ and $T_{\mathbf{A}(x)}$ denote tokens from an image $x$ and its augmented image $\mathbf{A}(x)$ , respectively. We use the cross-entropy loss between $Z_{T_x}^{\mathbf{A}}$ and $Z_{T_{\mathbf{A}(x)}}$ as our objective function $\mathcal{L}_{\mathrm{TA}}$ for learning our TokenAdapt module: + +$$ +\mathcal {L} _ {\mathrm {T A}} = \operatorname {C E} \left(Z _ {T _ {x}} ^ {\mathbf {A}}, Z _ {T _ {\mathbf {A} (x)}}\right). \tag {3} +$$ + +TokenAdapt deals with geometric data augmentations (e.g., flip, resize, crop, affine) and Mixup [68]. These augmentations are commonly used in various training methods, but they can cause a performance drop in token-based training, as shown in Figure 3(b). + +Once trained, the TokenAdapt module is used to augment a sequence of tokens $T$ into an augmented sequence $T^{\mathbf{A}}$ across different datasets and downstream tasks. Importantly, the augmentation-compatible feature space, which does not rely on specific semantic knowledge or alignment with downstream tasks, makes it dataset and task-agnostic. The inherent flexibility of TokenAdapt makes it highly generalizable, allowing direct application to other datasets without the need for additional training procedures. + +ColorAdapt. As witnessed in Cubuk et al. [12], color-based augmentations are often more effective than geometric ones in certain datasets and are an important component of data augmentation strategies. For example, color-based augmentations are useful for dealing with various environmental changes caused by different lighting conditions (e.g., day and night) or weather conditions (e.g., snow). However, integrating them into TokenAdapt is challenging because most color-based augmentations are designed to work only in the pixel domain. To address this, we propose ColorAdapt, a color-based token augmentation inspired by the method [27]. Formally, the ColorAdapt function $\mathcal{C}$ is defined as: + +$$ +\mathcal {C} \left(Z _ {T _ {1}}, Z _ {T _ {2}}\right) = \sigma \left(Z _ {T _ {2}}\right) \left(\frac {Z _ {T _ {1}} - \mu \left(Z _ {T _ {1}}\right)}{\sigma \left(Z _ {T _ {1}}\right)}\right) + \mu \left(Z _ {T _ {2}}\right), \tag {4} +$$ + +where $\mu(Z_T)$ and $\sigma(Z_T)$ are the mean and standard deviation of each token embedding $Z_T$ across its spatial dimension, computed for each channel. + +ColorAdapt mimics color-changing augmentations in the pixel domain while preserving the encoded object structure by transforming the global statistics within each channel of the token embedding. Figure 5 shows the images decoded from various color-augmented tokens. This illustrates how ColorAdapt affects the visual properties of the augmented tokens. By incorporating ColorAdapt + +![](images/4508aa871791343080e71200dba7cddbff76462a5c0cfec5d0e51c2e3adcc971.jpg) +Input data + +![](images/c85dc23a13cec7b64d1d81d12f85aabd7a074eebf84fc4511f8d8e097d465407.jpg) +Brightness + +![](images/0fa165838fc446140e9278ad22cdc9f518d5854db63f031d700b19e6e2f663b6.jpg) +Emb-Noise +Fig. 5: ColorAdapt provides more reasonable color changes. We present ViT-VQGAN decoded images to verify the quality of tokenizations after color changes. We use the brightness function with a factor of 0.2 following the implementation [46]. Emb-Noise [44] is the color-based token augmentation. Notably, our ColorAdapt effectively preserves object structure in contrast to the failure of the counterparts. + +![](images/27d01be32f7de5336c7f944d86429cd3033221d1c3f83fbe4d3bc5a07dc46a6c.jpg) +ColorAdapt + +![](images/713fa19fa1f0869bdc8ae3bea4ea6c0f65b67ac67890e165cef9cfa69459ea4c.jpg) +Input data + +![](images/89c295adbe5f70ab45ac7762e93e29835192b012fc0b4ef63da1f2f6039e30dc.jpg) +Brightness + +![](images/afac642f0f9ea9ed6a1849c22bda5a2a73eef4b6b99b8eeec05ee8df522fd560.jpg) +Emb-Noise + +![](images/26bde3657dd73f534634fef3fe4a0463330841d217d7c65981b2f4c66f8b696e.jpg) +ColorAdapt + +Table 1: Storage-efficient ImageNet-1k classification. We report top-1 accuracies (ViT-B) and compression ratio on ImageNet-1k using various data storage reduction methods. Our SeiT++ shows significant storage efficiency compared to competitors. + +
MethodInput # of images Top 1 Acc.Dataset storage size
Full pixelsImage1,281 k81.8140.0 GB
Uniform random sampling512 k74.054.6 GB (39%)
C-score [29] based samplingImage512 k73.353.3 GB (38%)
Adjusting JPEG quality factor to 51,281 k74.611.0 GB (8%)
Adjusting image resolution (× 0.2)1,281 k75.29.6 GB (7%)
SRe2L [65]Image200 k65.92.6 GB (2%)
G-VBSM [52]100 k63.71.6 GB (1%)
SeiT [44]Token1,281 k74.01.4 GB (1%)
SeiT++1,281 k77.81.4 GB (1%)
+ +into our token augmentation strategy, we improve the data diversity, exploring new possibilities for training more robust and generalizable vision models. + +# 4 Experiment + +In this section, we validate the effectiveness of SeiT++ on various scenarios. First, we evaluate the performance of our method in the context of storage-efficient ImageNet-1k classification and transfer learning on fine-grained datasets. Then, we demonstrate the adaptability of token-based vision training to ADE-20k semantic segmentation, showing additional improvements facilitated by SeiT++. In addition, we present experimental results illustrating how our strategies contribute to enhanced model generalizability in robustness evaluation. We also explore the applicability of our augmentation strategies to alternative input formats, demonstrating its extensibility. Unless otherwise stated, SeiT++ refers to the combination of MTM and our token augmentation strategies. + +Table 2: Token-based ImageNet-1k Classification. We report top-1 accuracies (ViT-B) on ImageNet-1k. Note that SeiT++ w/o MTM means the original SeiT [44] training with our token augmentation strategies. + +
StorageSeiTSeiT++
w/o MTMw/ MTMw/o MTMw/ MTM
1.4 GB74.075.175.5 (+1.5)77.8 (+2.7)
1.1 GB70.274.173.1 (+2.9)76.5 (+2.4)
0.8 GB66.370.669.1 (+2.8)74.1 (+3.5)
0.5 GB59.764.563.8 (+4.1)70.0 (+5.5)
0.3 GB47.253.951.2 (+4.0)60.6 (+6.7)
+ +Table 3: Fine-grained classification. We report top-1 accuracies (ViT-B) on various fine-grained datasets. Our method consistently outperforms its counterpart through effective augmentations for tokens. + +
MethodFlowersCarsiNat18iNat19
SeiT [44]93.579.743.150.1
SeiT++95.286.952.458.9
(+1.7)(+7.2)(+9.3)(+8.8)
+ +# 4.1 Storage-efficient ImageNet-1k + +Table 1 summarizes the performance and compression ratio of storage-efficient ImageNet-1k classification. Our method is compared with image-based approaches, including uniform random sampling, C-score [29]-based sampling, adjusting image resolution, and adjusting JPEG quality factor. Despite a $1\%$ compression ratio, when a ViT-B/16 model is pre-trained on tokenized ImageNet-1k using MTM and the ViT model is fine-tuned, our method achieves a $77.8\%$ top-1 accuracy, outperforming various image-based approaches. In Table 2, our method consistently outperforms the token-based learning baseline SeiT by $1.5 - 6.7\% p$ in top-1 accuracies. Notably, the effectiveness of our method becomes more pronounced as the available storage size decreases, highlighting its efficiency in low-storage scenarios. Furthermore, SeiT++ shows more significant performance gains with MTM. For a storage size of 1.4GB, SeiT++ achieves an improvement of $1.5\% p$ over SeiT without MTM, while with MTM the improvement increases to $2.7\% p$ . These results confirm that SeiT++ effectively takes advantage of MTM. + +# 4.2 Fine-grained Classification + +To further investigate the generalizability of SeiT++, we conduct experiments on four different fine-grained datasets: Flowers [43], StandfordCars [31], iNat-18 [57], and iNat-19 [58]. Specifically, we tokenize each dataset and then fine-tune the token-trained model (ViT-B) on ImageNet-1k with the tokenized dataset. Since + +Table 4: ADE-20k Semantic Segmentation. We report mIoU on the ADE-20k validation set, highlighting the effectiveness of our approach in pixel-level classification. + +
MethodBackboneSeiTSeiT++
UPerNet [63]ViT-B [17]39.043.2 (+4.2)
+ +Table 5: Robustness evaluation. We report top-1 robust accuracies (ViT-B) against corruptions and domain shifts for each model trained on ImageNet-1k. Our method consistently outperforms its counterpart on various robustness benchmarks. + +
MethodCleanGauss. NoiseGauss. BlurImageNet-RSketchObjectNet
SeiT [44]74.050.762.625.522.615.7
SeiT++75.558.666.830.227.718.3
(+1.5)(+7.9)(+4.2)(+4.7)(+5.1)(+2.6)
+ +fine-grained datasets require the classification of objects with intricate characteristics, they are particularly sensitive to undesirable noise (i.e., RRC in SeiT) or excessive perturbation (i.e., Emb-Noise) resulting from data augmentation in the token embedding space. Table 3 shows the quantitative performance (top-1 accuracies) on fine-grained datasets. SeiT++ consistently outperforms the performance of SeiT across various datasets. These results confirm the effectiveness of our proposed method in fine-grained classification. + +# 4.3 ADE-20k Semantic Segmentation + +A limitation of token-based learning is often substantiated by learning dense prediction tasks. We conjecture that our method could facilitate training for those tasks, such as semantic segmentation. We extract tokens from ADE-20k [72] and use the pre-trained ViT-B/16 model on tokenized ImageNet-1k to evaluate the effectiveness of our method in semantic segmentation. We then fine-tune the pre-trained model using the tokenized ADE-20k dataset. Table 4 reports the mean intersection-over-union (mIoU) as an evaluation metric. The results show a significant performance improvement $(+4.2\%)$ p mIoU) over SeiT. These results indicate that SeiT++ significantly improves model performance in the token-based dense prediction task. + +# 4.4 Evaluation on Robustness Datasets + +To validate the effectiveness of our augmentation strategies in terms of robustness, we further evaluate the robustness of our method on a set of benchmarks such as adding Gaussian noise [26] or Gaussian blur [26], ImageNet-R [25], Sketch [59], and ObjectNet [3]. Notably, we adopt SeiT++ w/o MTM to validate the impact of our augmentation strategies on model robustness. The results in Table 5 show that SeiT++ consistently outperforms SeiT in terms of robustness + +Table 6: Impact of the proposed augmentation strategies. We report the top-1 accuracies (ViT-S) on the ImageNet-100 validation set for various combinations of our token augmentation strategies. Notably, our TokenAdapt and ColorAdapt consistently improve model performance, exhibiting synergy when used together. + +
FrameworkColorAdaptTokenAdaptTop 1 Acc.
SeiT [44]XX77.3
✓X78.3
X✓80.4
✓✓81.4
+ +Table 7: Adaptibility of our method to different input. Input denotes tokenizers for our model. We report the top-1 accuracies (ViT-S) on ImageNet-100. For VQGAN, we used a publicly available VQGAN [18] trained on OpenImages [32] as a tokenizer. + +
InputSeiTSeiT++
Token (ViT-VQGAN [66])77.381.4 (+4.1)
Token (VQGAN [18])81.883.9 (+2.1)
DCT coefficients69.571.0 (+1.5)
+ +accuracy. Previous studies [8, 55] observe that the key to the input pixel robustness depends on the pixel-level augmentations. Since our method enables pixel-based data augmentation to tokens, it achieves higher robustness than the baseline, which does not effectively exploit the pixel-based data augmentation. + +# 4.5 Ablation Study + +Impact of Each Augmentation Strategy. In Table 6, we report the impact of proposed token augmentations. In particular, incorporating ColorAdapt leads to a $1.0\%$ p performance improvement, highlighting the positive effect of preserving object structure during model training. Additionally, incorporating TokenAdapt yields an extra $3.1\%$ p improvement. These results confirm that exploring token augmentations is a key factor in fully exploiting pre-extracted tokens. + +Extensibility of Our Augmentation Strategy. In this section, we conjecture the broad applicability of our augmentation strategies across various scenarios. To explore its extensibility, we delve into two distinct situations. Firstly, we examine the integration of our method with an alternative tokenizer—the OpenImages-trained VQGAN tokenizer [18], alongside the ImageNet-1k-trained ViT-VQGAN tokenizer [66]. The results in Table 7 consistently demonstrate improved performance on ImageNet-100 across different tokenizers. Additionally, we investigate the adaptability of our approach to a different input format by focusing on Discrete Cosine Transform (DCT) coefficients. Inspired by JPEG compression, we apply DCT to each $8 \times 8$ image patch, yielding quantized DCT + +coefficients utilized as input data during model training. As shown in Table 7, our approach enhances top-1 accuracy, similar to the tokenizer scenario. This consistent improvement across diverse input formats underscores the universal applicability of our method, highlighting its potential as a widely applicable solution for diverse input representations. + +# 4.6 Implementation Details + +We use the pre-trained ViT-VQGAN tokenizer [66] for token extraction. For our TokenAdapt module, we employ a single transformer block for conversion and reverse function. The TokenAdapt module is trained for one epoch on ImageNet-1k with a batch size of 128, a learning rate of 0.001 with cosine scheduling, and a weight decay of 0.05. We follow the training recipes from MAGE [34], SeiT [44], and mmsegmentation [10] for masked token modeling (MTM), token-based classification, and semantic segmentation, respectively. Detailed hyperparameters and settings are available in the Appendix. + +# 5 Conclusion + +In this paper, we investigate the compatibility of SeiT with self-supervised learning approaches in token-based frameworks. Among various self-supervised algorithms, Masked Image Modeling (MIM) [24,64] has recently gained popularity for learning visual representations. Inspired by this, we extend SeiT by integrating Masked Token Modeling (MTM) for self-supervised pre-training. Additionally, recognizing that self-supervised approaches often require more data due to the lack of labels, we propose simple yet effective token-based augmentation strategies: TokenAdapt and ColorAdapt. TokenAdapt aligns the token embedding space with another embedding space compatible with pixel-based data augmentations. In addition, we introduce ColorAdapt, a color-changing augmentation for tokens. Our experimental results across various scenarios, including storage-efficient ImageNet-1k classification, fine-grained classification, ADE-20k semantic segmentation, and robustness benchmarks, consistently validate the effectiveness of our proposed method. Furthermore, the consistent performance improvement across diverse input formats highlights the potential of our approach as a widely applicable solution for various input representations. + +Limitations. While SeiT++ is capable of handling large-scale data without labels with optimized storage requirements, experiments on such large datasets (e.g., ImageNet-21k [14]) and huge models are not contained in this paper due to resource constraints. We believe the advantages of MTM and our data augmentation strategies can be seamlessly integrated with scalability. + +Acknowledgements. Experiments are based on the Naver Smart Machine Learning NSML [30] platform. This research was supported by IITP grant funded by the Korea government(MSIT) and KEIT grant funded by the Korea government(MOTIE) (No. 2022-0-00680, 2022-0-01045). + +# References + +1. Bahng, H., Chun, S., Yun, S., Choo, J., Oh, S.J.: Learning de-biased representations with biased representations. In: International Conference on Machine Learning. pp. 528-539. PMLR (2020) +2. Bao, H., Dong, L., Piao, S., Wei, F.: Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254 (2021) +3. Barbu, A., Mayo, D., Alverio, J., Luo, W., Wang, C., Gutfreund, D., Tenenbaum, J., Katz, B.: Objectnet: A large-scale bias-controlled dataset for pushing the limits of object recognition models. Advances in neural information processing systems 32 (2019) +4. Chang, H., Zhang, H., Barber, J., Maschinot, A., Lezama, J., Jiang, L., Yang, M.H., Murphy, K., Freeman, W.T., Rubinstein, M., et al.: Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704 (2023) +5. Chang, H., Zhang, H., Jiang, L., Liu, C., Freeman, W.T.: Maskgit: Masked generative image transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11315-11325 (2022) +6. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International conference on machine learning. pp. 1597-1607. PMLR (2020) +7. Chen, X., Xie, S., He, K.: An empirical study of training self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9640-9649 (2021) +8. Chun, S., Oh, S.J., Yun, S., Han, D., Choe, J., Yoo, Y.: An empirical evaluation on robustness and uncertainty of regularization methods. arXiv preprint arXiv:2003.03879 (2020) +9. Coleman, C., Yeh, C., Mussmann, S., Mirzasoleiman, B., Bailis, P., Liang, P., Leskovec, J., Zaharia, M.: Selection via proxy: Efficient data selection for deep learning. arXiv preprint arXiv:1906.11829 (2019) +0. Contributors, M.: MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark. https://github.com/open-mmlab/mmsegmentation (2020) +1. Croce, F., Hein, M.: Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In: International conference on machine learning. pp. 2206-2216. PMLR (2020) +2. Cubuk, E.D., Zoph, B., Mane, D., Vasudevan, V., Le, Q.V.: Autoaugment: Learning augmentation strategies from data. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 113-123 (2019) +3. Cubuk, E.D., Zoph, B., Shlens, J., Le, Q.V.: Randaugment: Practical automated data augmentation with a reduced search space. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 702-703 (2020) +4. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009) +5. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018) +6. DeVries, T., Taylor, G.W.: Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552 (2017) + +17. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) +18. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12873-12883 (2021) +19. Fang, Y., Dong, L., Bao, H., Wang, X., Wei, F.: Corrupted image modeling for self-supervised visual pre-training. arXiv preprint arXiv:2202.03382 (2022) +20. Ford, N., Gilmer, J., Carlini, N., Cubuk, D.: Adversarial examples are a natural consequence of test error in noise. arXiv preprint arXiv:1901.10513 (2019) +21. Geirhos, R., Rubisch, P., Michaelis, C., Bethge, M., Wichmann, F.A., Brendel, W.: Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. arXiv preprint arXiv:1811.12231 (2018) +22. Goodfellow, I.J., Shlens, J., Szegedy, C.: Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572 (2014) +23. Grill, J.B., Strub, F., Altché, F., Tallec, C., Richemond, P., Buchatskaya, E., Doersch, C., Avila Pires, B., Guo, Z., Gheshlaghi Azar, M., et al.: Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems 33, 21271-21284 (2020) +24. He, K., Chen, X., Xie, S., Li, Y., Dollár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022) +25. Hendrycks, D., Basart, S., Mu, N., Kadavath, S., Wang, F., Dorundo, E., Desai, R., Zhu, T., Parajuli, S., Guo, M., Song, D., Steinhardt, J., Gilmer, J.: The many faces of robustness: A critical analysis of out-of-distribution generalization. ICCV (2021) +26. Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261 (2019) +27. Huang, X., Belongie, S.: Arbitrary style transfer in real-time with adaptive instance normalization. In: Proceedings of the IEEE international conference on computer vision. pp. 1501-1510 (2017) +28. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International conference on machine learning. pp. 4904-4916. PMLR (2021) +29. Jiang, Z., Zhang, C., Talwar, K., Mozer, M.C.: Characterizing structural regularities of labeled data in overparameterized models. arXiv preprint arXiv:2002.03206 (2020) +30. Kim, H., Kim, M., Seo, D., Kim, J., Park, H., Park, S., Jo, H., Kim, K., Yang, Y., Kim, Y., et al.: Nsml: Meet the mlaas platform with a real-world case study. arXiv preprint arXiv:1810.09957 (2018) +31. Krause, J., Stark, M., Deng, J., Fei-Fei, L.: 3d object representations for fine-grained categorization. In: Proceedings of the IEEE international conference on computer vision workshops. pp. 554-561 (2013) +32. Kuznetsova, A., Rom, H., Alldrin, N., Uijlings, J., Krasin, I., Pont-Tuset, J., Kamali, S., Popov, S., Malloci, M., Kolesnikov, A., et al.: The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision 128(7), 1956-1981 (2020) + +33. Lee, S., Chun, S., Jung, S., Yun, S., Yoon, S.: Dataset condensation with contrastive signals. In: International Conference on Machine Learning. pp. 12352-12364. PMLR (2022) +34. Li, T., Chang, H., Mishra, S., Zhang, H., Katabi, D., Krishnan, D.: Mage: Masked generative encoder to unify representation learning and image synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2142-2152 (2023) +35. Lin, S., Ji, R., Yan, C., Zhang, B., Cao, L., Ye, Q., Huang, F., Doermann, D.: Towards optimal structured cnn pruning via generative adversarial learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2790-2799 (2019) +36. Liu, J., Guo, J., Xu, D.: Apsnet: Toward adaptive point sampling for efficient 3d action recognition. IEEE Transactions on Image Processing 31, 5287-5302 (2022) +37. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021) +38. Lopes, R.G., Yin, D., Poole, B., Gilmer, J., Cubuk, E.D.: Improving robustness without sacrificing accuracy with patch gaussian augmentation. arXiv preprint arXiv:1906.02611 (2019) +39. Ma, Y., Li, H., Zhang, Z., Guo, J., Zhang, S., Gong, R., Liu, X.: Annealing-based label-transfer learning for open world object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11454-11463 (2023) +40. Madry, A., Makelov, A., Schmidt, L., Tsipras, D., Vladu, A.: Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083 (2017) +41. Mahajan, D., Girshick, R., Ramanathan, V., He, K., Paluri, M., Li, Y., Bharambe, A., Van Der Maaten, L.: Exploring the limits of weakly supervised pretraining. In: Proceedings of the European conference on computer vision (ECCV). pp. 181-196 (2018) +42. Mei, Z., Ye, P., Ye, H., Li, B., Guo, J., Chen, T., Ouyang, W.: Automatic loss function search for adversarial unsupervised domain adaptation. IEEE Transactions on Circuits and Systems for Video Technology 33(10), 5868-5881 (2023) +43. Nilsback, M.E., Zisserman, A.: Automated flower classification over a large number of classes. In: 2008 Sixth Indian conference on computer vision, graphics & image processing. pp. 722-729. IEEE (2008) +44. Park, S., Chun, S., Heo, B., Kim, W., Yun, S.: Seit: Storage-efficient vision training with tokens using $1\%$ of pixel storage. arXiv preprint arXiv:2303.11114 (2023) +45. Paul, M., Ganguli, S., Dziugaite, G.K.: Deep learning on a data diet: Finding important examples early in training. Advances in Neural Information Processing Systems 34, 20596-20607 (2021) +46. Riba, E., Mishkin, D., Ponsa, D., Rublee, E., Bradski, G.: Kornia: an open source differentiable computer vision library for pytorch. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3674-3683 (2020) +47. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10684-10695 (2022) +48. Rosasco, A., Carta, A., Cossu, A., Lomonaco, V., Bacciu, D.: Distilled replay: Overcoming forgetting through synthetic samples. In: International Workshop on Continual Semi-Supervised Learning. pp. 104-117. Springer (2021) + +49. Sangermano, M., Carta, A., Cossu, A., Bacciu, D.: Sample condensation in online continual learning. In: 2022 International Joint Conference on Neural Networks (IJCNN). pp. 01-08. IEEE (2022) +50. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022) +51. Scimeca, L., Oh, S.J., Chun, S., Poli, M., Yun, S.: Which shortcut cues will dnns choose? a study from the parameter-space perspective. arXiv preprint arXiv:2110.03095 (2021) +52. Shao, S., Yin, Z., Zhou, M., Zhang, X., Shen, Z.: Generalized large-scale data condensation via various backbone and statistical matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16709-16718 (2024) +53. Singh, M., Gustafson, L., Adcock, A., de Freitas Reis, V., Gedik, B., Kosaraju, R.P., Mahajan, D., Girshick, R., Dollar, P., Van Der Maaten, L.: Revisiting weakly supervised pre-training of visual perception models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 804-814 (2022) +54. Steiner, A., Kolesnikov, A., Zhai, X., Wightman, R., Uszkoreit, J., Beyer, L.: How to train your vit? data, augmentation, and regularization in vision transformers. arXiv preprint arXiv:2106.10270 (2021) +55. Taori, R., Dave, A., Shankar, V., Carlini, N., Recht, B., Schmidt, L.: Measuring robustness to natural distribution shifts in image classification. Advances in Neural Information Processing Systems 33, 18583-18599 (2020) +56. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jégou, H.: Training data-efficient image transformers & distillation through attention. In: International conference on machine learning. pp. 10347-10357. PMLR (2021) +57. Van Horn, G., Mac Aodha, O., Song, Y., Cui, Y., Sun, C., Shepard, A., Adam, H., Perona, P., Belongie, S.: The inaturalist challenge 2018 dataset. arXiv preprint arXiv:1707.06642 (2018) +58. Van Horn, G., Mac Aodha, O., Song, Y., Cui, Y., Sun, C., Shepard, A., Adam, H., Perona, P., Belongie, S.: The inaturalist challenge 2019 dataset. arXiv preprint arXiv:1707.06642 (2019) +59. Wang, H., Ge, S., Lipton, Z., Xing, E.P.: Learning robust global representations by penalizing local predictive power. In: Advances in Neural Information Processing Systems. pp. 10506-10518 (2019) +60. Wang, T., Zhu, J.Y., Torralba, A., Efros, A.A.: Dataset distillation. arXiv preprint arXiv:1811.10959 (2018) +61. Wei, J., Zou, K.: Eda: Easy data augmentation techniques for boosting performance on text classification tasks. arXiv preprint arXiv:1901.11196 (2019) +62. Wightman, R., Touvron, H., Jégou, H.: Resnet strikes back: An improved training procedure in timm. arXiv preprint arXiv:2110.00476 (2021) +63. Xiao, T., Liu, Y., Zhou, B., Jiang, Y., Sun, J.: Unified perceptual parsing for scene understanding. In: Proceedings of the European conference on computer vision (ECCV). pp. 418-434 (2018) +64. Xie, Z., Zhang, Z., Cao, Y., Lin, Y., Bao, J., Yao, Z., Dai, Q., Hu, H.: Simmim: A simple framework for masked image modeling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9653-9663 (2022) +65. Yin, Z., Xing, E., Shen, Z.: Squeeze, recover and relabel: Dataset condensation at imagenet scale from a new perspective. In: Thirty-seventh Conference on Neural Information Processing Systems (2023) + +66. Yu, J., Li, X., Koh, J.Y., Zhang, H., Pang, R., Qin, J., Ku, A., Xu, Y., Baldridge, J., Wu, Y.: Vector-quantized image modeling with improved vqgan. arXiv preprint arXiv:2110.04627 (2021) +67. Yun, S., Han, D., Oh, S.J., Chun, S., Choe, J., Yoo, Y.: Cutmix: Regularization strategy to train strong classifiers with localizable features. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6023-6032 (2019) +68. Zhang, H., Cisse, M., Dauphin, Y.N., Lopez-Paz, D.: mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412 (2017) +69. Zhao, B., Bilen, H.: Dataset condensation with differentiable siamese augmentation. In: International Conference on Machine Learning. pp. 12674-12685. PMLR (2021) +70. Zhao, B., Mopuri, K.R., Bilen, H.: Dataset condensation with gradient matching. arXiv preprint arXiv:2006.05929 (2020) +71. Zhong, Z., Zheng, L., Kang, G., Li, S., Yang, Y.: Random erasing data augmentation. In: Proceedings of the AAAI Conference on Artificial Intelligence (AAAI) (2020) +72. Zhou, B., Zhao, H., Puig, X., Xiao, T., Fidler, S., Barriuso, A., Torralba, A.: Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision 127, 302-321 (2019) \ No newline at end of file diff --git a/2024/SeiT++_ Masked Token Modeling Improves Storage-efficient Training/images.zip b/2024/SeiT++_ Masked Token Modeling Improves Storage-efficient Training/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8273a41a2a047d479c4de422984759da4d14aba6 --- /dev/null +++ b/2024/SeiT++_ Masked Token Modeling Improves Storage-efficient Training/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a53dd346c1077d475e0082fc6dcf4da4778dfe044d628982de62cab8868e53ff +size 348317 diff --git a/2024/SeiT++_ Masked Token Modeling Improves Storage-efficient Training/layout.json b/2024/SeiT++_ Masked Token Modeling Improves Storage-efficient Training/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..215798a61552ed955b3bc72b7f202ee71547187b --- /dev/null +++ b/2024/SeiT++_ Masked Token Modeling Improves Storage-efficient Training/layout.json @@ -0,0 +1,9640 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 159, + 111, + 455, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 111, + 455, + 148 + ], + "spans": [ + { + "bbox": [ + 159, + 111, + 455, + 148 + ], + "type": "text", + "content": "SeiT++: Masked Token Modeling Improves Storage-efficient Training" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 141, + 168, + 473, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 168, + 473, + 194 + ], + "spans": [ + { + "bbox": [ + 141, + 168, + 473, + 194 + ], + "type": "text", + "content": "Minhyun Lee" + }, + { + "bbox": [ + 141, + 168, + 473, + 194 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + }, + { + "bbox": [ + 141, + 168, + 473, + 194 + ], + "type": "text", + "content": " Song Park" + }, + { + "bbox": [ + 141, + 168, + 473, + 194 + ], + "type": "inline_equation", + "content": "^{2,\\dagger}" + }, + { + "bbox": [ + 141, + 168, + 473, + 194 + ], + "type": "text", + "content": " Byeongho Heo" + }, + { + "bbox": [ + 141, + 168, + 473, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 141, + 168, + 473, + 194 + ], + "type": "text", + "content": " Dongyoon Han" + }, + { + "bbox": [ + 141, + 168, + 473, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 141, + 168, + 473, + 194 + ], + "type": "text", + "content": " Hyunjung Shim" + }, + { + "bbox": [ + 141, + 168, + 473, + 194 + ], + "type": "inline_equation", + "content": "^{3,\\star}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 184, + 198, + 429, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 198, + 429, + 223 + ], + "spans": [ + { + "bbox": [ + 184, + 198, + 429, + 223 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 184, + 198, + 429, + 223 + ], + "type": "text", + "content": " Yonsei University " + }, + { + "bbox": [ + 184, + 198, + 429, + 223 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 184, + 198, + 429, + 223 + ], + "type": "text", + "content": " Naver AI Lab " + }, + { + "bbox": [ + 184, + 198, + 429, + 223 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 184, + 198, + 429, + 223 + ], + "type": "text", + "content": " KAIST AI " + }, + { + "bbox": [ + 184, + 198, + 429, + 223 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 184, + 198, + 429, + 223 + ], + "type": "text", + "content": " Equal contribution. Work done at Naver AI Lab." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 256, + 455, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 256, + 455, + 498 + ], + "spans": [ + { + "bbox": [ + 160, + 256, + 455, + 498 + ], + "type": "text", + "content": "Abstract. Recent advancements in Deep Neural Network (DNN) models have significantly improved performance across computer vision tasks. However, achieving highly generalizable and high-performing vision models requires expansive datasets, resulting in significant storage requirements. This storage challenge is a critical bottleneck for scaling up models. A recent breakthrough by SeiT proposed the use of Vector-Quantized (VQ) feature vectors (i.e., tokens) as network inputs for vision classification. This approach achieved " + }, + { + "bbox": [ + 160, + 256, + 455, + 498 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 160, + 256, + 455, + 498 + ], + "type": "text", + "content": " of the performance of a model trained on full-pixel images with only " + }, + { + "bbox": [ + 160, + 256, + 455, + 498 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 160, + 256, + 455, + 498 + ], + "type": "text", + "content": " of the storage. While SeiT needs labeled data, its potential in scenarios beyond fully supervised learning remains largely untapped. In this paper, we extend SeiT by integrating Masked Token Modeling (MTM) for self-supervised pre-training. Recognizing that self-supervised approaches often demand more data due to the lack of labels, we introduce TokenAdapt and ColorAdapt. These methods facilitate comprehensive token-friendly data augmentation, effectively addressing the increased data requirements of self-supervised learning. We evaluate our approach across various scenarios, including storage-efficient ImageNet-1k classification, fine-grained classification, ADE-20k semantic segmentation, and robustness benchmarks. Experimental results demonstrate consistent performance improvement in diverse experiments, validating the effectiveness of our method. Code is available at https://github.com/naver-ai/seit." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 521, + 230, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 521, + 230, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 521, + 230, + 533 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 548, + 486, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 548, + 486, + 646 + ], + "spans": [ + { + "bbox": [ + 130, + 548, + 486, + 646 + ], + "type": "text", + "content": "Recent advancements in Deep Neural Network (DNN) [17,56] have significantly improved their performance across various computer vision tasks. However, achieving highly generalizable and high-performing vision models demands an extensive dataset, often containing billions of data points [28,41,53] with large storage requirements. For example, the LAION-5B [50] dataset requires a storage size of 240TB. This storage challenge emerges as a critical bottleneck for scaling up vision models. Consequently, optimizing storage efficiency becomes essential for the practical advancement of scaled-up vision models." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 317, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 317, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 317, + 665 + ], + "type": "text", + "content": "* Hyunjung Shim is a corresponding author." + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 168, + 114, + 445, + 229 + ], + "blocks": [ + { + "bbox": [ + 168, + 114, + 445, + 229 + ], + "lines": [ + { + "bbox": [ + 168, + 114, + 445, + 229 + ], + "spans": [ + { + "bbox": [ + 168, + 114, + 445, + 229 + ], + "type": "image", + "image_path": "048273a0b26487a7811143950997cf5c723f7911823c28f1746567b200c0ae27.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 237, + 482, + 304 + ], + "lines": [ + { + "bbox": [ + 130, + 237, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 482, + 304 + ], + "type": "text", + "content": "Fig.1: Over " + }, + { + "bbox": [ + 130, + 237, + 482, + 304 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 130, + 237, + 482, + 304 + ], + "type": "text", + "content": " top-1 accuracy is achievable with just 1GB data. On ImageNet-1k, we visualize the trade-off of training data storage vs. top-1 accuracy using the fixed ViT-B/16 for a controlled comparison. Each accuracy metric is individually trained with different data types. While the entire ImageNet-1k dataset requires approximately 140GB for training with images, our approach demonstrates significant storage efficiency over competitors." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 329, + 482, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 329, + 482, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 329, + 482, + 461 + ], + "type": "text", + "content": "Images require more storage than texts due to their intrinsic complexity in converting continuous natural light signals into discrete pixels for computer representation. However, a considerable portion of image data corresponds to redundant or unnecessary details, such as repeated patterns or imperceptible details [47]. These properties lead to large storage requirements and also induce negative impacts on visual recognition tasks as mentioned in [1,11,21,22,40,51]. Therefore, several attempts [9, 29, 45, 69, 70] have been made to reduce storage requirements by optimizing the efficiency of image data. A standard approach to reduce storage requirements is to decrease the total number of training data by removing less important samples [9, 29, 45] or condensing the data into smaller synthetic sets [69, 70]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 461, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 582 + ], + "type": "text", + "content": "However, as SeiT [44] illustrates, these approaches may incur a significant performance drop or become infeasible for large-scale datasets due to their complexity. Alternatively, some methods aim to reduce the size of each image through compression techniques or resolution adjustments [50]. However, they still face a considerable performance gap with the original dataset. Recently, SeiT [44] has made a significant breakthrough in this area by enabling the direct learning of ViTs [17] from discrete representations (i.e., tokens) pre-extracted by an offline tokenizer [66]. By using vector-quantized features instead of traditional pixel images, SeiT achieves " + }, + { + "bbox": [ + 130, + 461, + 482, + 582 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 130, + 461, + 482, + 582 + ], + "type": "text", + "content": " of the performance of a full storage model while requiring only " + }, + { + "bbox": [ + 130, + 461, + 482, + 582 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 461, + 482, + 582 + ], + "type": "text", + "content": " of the storage." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "While SeiT [44] has demonstrated the high potential of token-based training, a performance gap with the image-based models still persists, and a prospective solution - exploration beyond fully supervised learning scenarios - remains untapped. To leverage the advantages of large-scale data beyond human-annotated datasets, exploring label-free learning methods is imperative. In this paper, we investigate the compatibility of SeiT with self-supervised learning approaches in token-based frameworks. Among many different self-supervised methods [6,7,23]," + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 245, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 245, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 245, + 100 + ], + "type": "text", + "content": "Lee and Park et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "Masked Image Modeling (MIM) [24,64] derived from Masked Language Modeling (MLM) [15] has brought significant success and has quickly become a popular approach for learning visual representations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 159, + 482, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 159, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 159, + 482, + 304 + ], + "type": "text", + "content": "Inspired by this, we introduce Masked Token Modeling (MTM), a pre-training method that performs token masking and reconstruction in token-based frameworks, and validate its effectiveness for storage efficiency on various benchmarks. To the best of our knowledge, this is the first approach for learning an MLM-based model directly from offline tokens, offering the potential to optimize storage efficiency for large-scale training tasks. As previous studies [12, 13, 56] have highlighted, data augmentation plays a crucial role in large-scale model training and is a fundamental element in various training setups [6, 7, 23, 24, 37, 56, 62, 64]. However, applying existing data augmentation techniques [13, 67, 71] designed for pixel images to tokens poses a significant challenge due to the input domain shift. Therefore, addressing this challenge is key to overcoming major performance bottlenecks and improving the storage efficiency of token-based frameworks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 311, + 482, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 311, + 482, + 455 + ], + "spans": [ + { + "bbox": [ + 130, + 311, + 482, + 455 + ], + "type": "text", + "content": "To this end, we introduce TokenAdapt and ColorAdapt, two novel augmentation strategies designed for token-based training. TokenAdapt facilitates the integration of traditional pixel-based augmentation policies by transforming token embeddings into a feature space compatible with pixel-based augmentations. After applying augmentation to the transformed token embeddings, the augmented token embeddings are then reverted to their original token space. ColorAdapt, inspired by Adaptive Instance Normalization (AdaIN [27]), alters the color attributes of token embeddings while preserving object structure by adjusting their statistics. By integrating TokenAdapt and ColorAdapt into our token augmentation process, we enhance the effectiveness of token-based training across both fully supervised and unsupervised learning scenarios, demonstrating significant improvements in model performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 462, + 487, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 487, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 487, + 668 + ], + "type": "text", + "content": "Our SeiT++ framework combines Masked Token Modeling (MTM) with the proposed token augmentation strategies, TokenAdapt and ColorAdapt. We validate the effectiveness of SeiT++ on various scenarios: (1) storage-efficient ImageNet-1k classification, (2) transfer learning on fine-grained classification, (3) ADE-20k semantic segmentation, and (4) evaluation on robustness benchmarks. As shown in Figure 1, SeiT++ consistently outperforms comparison methods in storage-efficient ImageNet-1k classification with limited storage sizes, achieving a top-1 accuracy of " + }, + { + "bbox": [ + 130, + 462, + 487, + 668 + ], + "type": "inline_equation", + "content": "77.8\\%" + }, + { + "bbox": [ + 130, + 462, + 487, + 668 + ], + "type": "text", + "content": " with only 1.4GB of data. These results confirm that token-based training can take advantage of self-supervised learning to improve storage efficiency. We also demonstrate that our proposed token augmentations significantly contribute to storage efficiency by augmenting data effectively. We reveal that our token-based augmentation methods basically yield significant boosts in supervised learning and further amplify improvements in MTM learning. Furthermore, we present how our strategies enhance the generalizability of trained models, as evidenced by our experimental results on fine-grained classification and robustness evaluation. Lastly, the applicability of our method to an alternative input format (e.g., DCT coefficients) demonstrates its extensibility." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 169, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 169, + 91, + 447, + 103 + ], + "type": "text", + "content": "SeiT++: Masked Token Modeling Improves Storage-efficient Training" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 237, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 237, + 128 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 237, + 128 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 481, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 481, + 270 + ], + "type": "text", + "content": "Learning with Tokenization. The use of discrete tokens for training vision models has recently gained attention, particularly in self-supervised representation learning. BeiT [2] employs the Masked Image Modeling (MIM) framework to recover discrete tokens from masked images and CIM [19] utilizes tokens in Corrupted Image Modeling. MAGE [34] demonstrates the incorporation of tokens from VQGAN [18] for both generative training and representation learning within a unified training framework. However, existing studies mainly focus on self-supervised representation learning and rely on online tokenization, converting images to tokens during each iteration of model training. This approach incurs notable memory consumption and computational costs, limiting the full exploitation of the storage efficiency of discrete tokens." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 288, + 481, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 288, + 481, + 516 + ], + "spans": [ + { + "bbox": [ + 133, + 288, + 481, + 516 + ], + "type": "text", + "content": "Storage-efficient Vision Training. Efficient model learning is a crucial research area, with studies focusing on optimizing target models [35, 36] or minimizing training data and labels [39, 42]. Storage-efficient training aims to reduce the dataset storage size. To this end, dataset distillation (DD) [60] creates a condensed dataset by leveraging the knowledge of the original dataset. Recent research [33, 48, 49, 69, 70] highlights the effectiveness of dataset distillation in improving training efficiency. However, its high computational complexity poses challenges, particularly when applied to large-scale datasets like ImageNet-1k [14]. Sampling-based methods [9, 29, 45] address this by selecting a subset of the most representative samples from the full dataset. Paul et al. [45] identify important samples very early in training. Jiang et al. [29] utilize a consistency score based on statistics collected during training. However, the diverse sample selection is not guaranteed, particularly in low data regimes, leading to suboptimal performance compared to dataset distillation (DD) methods [70]. To tackle these issues, SeiT [44] proposes a storage-efficient vision training framework by leveraging the storage efficiency of tokens. They utilize the ViT-VQGAN [66] tokenizer to extract tokens from images and store them for vision model training. As a result, they outperform other storage-efficient training methods, significantly reducing the dataset storage size." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 533, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 533, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 533, + 481, + 665 + ], + "type": "text", + "content": "Data Augmentation. Data augmentation serves as a fundamental building block to improve the model performance across diverse computer vision tasks. Various data augmentation techniques have been employed for improved performance [16, 71] or model robustness [20, 38]. Mixup [68] introduces convex combinations of pairs of images and their labels to create a smoother decision boundary. CutMix [67] replaces certain regions of an image with a patch from another image, enhancing the model's localization ability. AutoAugment [12] leverages reinforcement learning to automate optimal augmentation policy discovery. RandAugment [13] introduces a potent policy based on random combinations of different data augmentation methods. DeiT [56] emphasizes the crucial role of data augmentation in vision transformers for improved model performance." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 245, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 245, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 245, + 100 + ], + "type": "text", + "content": "Lee and Park et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 479, + 217 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 217 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 217 + ], + "type": "image", + "image_path": "93f70da2a7709e419b589aaf11d5588d6735a9a08eb5ccbdab7920fa8989e29c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 224, + 482, + 269 + ], + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 269 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 269 + ], + "type": "text", + "content": "Fig. 2: Masked Token Modeling (MTM) pipeline. MTM is a self-supervised learning approach in token-based frameworks. (a) The tokenized dataset is saved on storage before model training. Then, (b) using only the pre-stored tokens, a storage-efficient vision model (MTM) is trained without relying on labeled datasets." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 293, + 482, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 354 + ], + "type": "text", + "content": "Previous studies extensively explore the effective data augmentation policies for pixel-based vision model training. However, we observe that pixel-based data augmentations are not well-suited for tokens, posing a critical performance bottleneck. Motivated by this, we tackle the data augmentation challenge, which has not been explored in previous studies for vision model training with tokens." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 372, + 202, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 372, + 202, + 384 + ], + "spans": [ + { + "bbox": [ + 132, + 372, + 202, + 384 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 398, + 425, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 398, + 425, + 411 + ], + "spans": [ + { + "bbox": [ + 132, + 398, + 425, + 411 + ], + "type": "text", + "content": "3.1 Preliminary: Storage-efficient Vision Training (SeiT)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 418, + 482, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 482, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 482, + 552 + ], + "type": "text", + "content": "SeiT [44] aims to develop a scalable and high-performance vision classifier with minimal storage requirements. It introduces a storage-efficient vision training framework that enables the direct learning of ViT [17] from discrete representations (i.e., tokens) pre-extracted by an offline tokenizer. In addition, SeiT further improves training by incorporating two token-specific augmentations: Token-EDA, drawing inspiration from easy data augmentation for language [61], and Emb-Noise, which introduces Gaussian noise into discrete representations. By using vector-quantized features instead of pixel images, SeiT retains " + }, + { + "bbox": [ + 130, + 418, + 482, + 552 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 130, + 418, + 482, + 552 + ], + "type": "text", + "content": " of the precision of a model trained on full-pixel images while using only " + }, + { + "bbox": [ + 130, + 418, + 482, + 552 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 418, + 482, + 552 + ], + "type": "text", + "content": " of the storage space. SeiT consists of two parts: (1) preparing the token dataset and (2) training a model using the token dataset." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Preparing the Token Dataset. SeiT extracts tokens using the ImageNet-1k-trained ViT-VQGAN tokenizer [66] to prepare the token dataset for training. As shown in Figure 2(a), each pixel image " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " is transformed into a sequence of tokens " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "T = [t_1, t_2, \\dots, t_n]" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": ". Since the tokens are stored in quantized form, each token means an embedding vector in a codebook " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{Z} = \\{z_k \\in \\mathbb{R}^d | k = 1, \\dots, K\\}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": ". i.e., " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "t_i \\in \\mathbb{Q}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{Q} = \\{1, 2, \\dots, K\\}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " for a codebook with size " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": ". Here, given the tokens " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " and the codebook " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": ", the corresponding token embeddings " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "Z_T = [z_{t_1}, z_{t_2}, \\dots, z_{t_n}]" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " can be computed by lookup process. This conversion compresses each image" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 169, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 169, + 91, + 448, + 103 + ], + "type": "text", + "content": "SeiT++: Masked Token Modeling Improves Storage-efficient Training" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "inline_equation", + "content": "I \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "text", + "content": " into a more compact token sequence " + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "inline_equation", + "content": "T \\in \\mathbb{R}^n" + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "text", + "content": " is much smaller than " + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "inline_equation", + "content": "H \\times W \\times 3" + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "text", + "content": ", thus significantly reducing storage requirements. For example, storing the full-pixel ImageNet-1k requires 140GB, but its tokenized form needs only 1.4GB. Notably, the tokens are stored on storage before model training with the codebook " + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "text", + "content": ". Thus, during the training phase, the pre-stored tokens are used directly as input instead of the original pixel images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 211, + 483, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 211, + 483, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 211, + 483, + 344 + ], + "type": "text", + "content": "Training Pipeline. To train a vision classifier using tokens, SeiT initially loads the token dataset and applies Token-EDA to the tokens. The tokens are then converted to a one-hot format and randomly resized and cropped. Subsequently, the processed one-hot tokens are transformed into token embeddings using a pretrained codebook. Techniques such as CutMix and Embedding-noise are applied to these token embeddings, and the resulting augmented tensor is fed into the model for training. Through the proposed framework, SeiT demonstrates the high potential of token-based training, but it lacks exploration beyond fully supervised learning scenarios. In addition, while token-specific augmentation is proposed, the types of token augmentation are limited, and the analysis of data augmentation compared to pixel-based training is insufficient." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 366, + 285, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 366, + 285, + 379 + ], + "spans": [ + { + "bbox": [ + 132, + 366, + 285, + 379 + ], + "type": "text", + "content": "3.2 Masked Token Modeling" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 391, + 482, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 391, + 482, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 482, + 453 + ], + "type": "text", + "content": "In this section, we introduce Masked Token Modeling (MTM), a pre-training method in token-based frameworks without relying on labeled datasets. MTM shares a similar concept to Masked Language Modeling (MLM) [15], which works by reconstructing original tokens from their masked versions, employing an autoencoding strategy. Figure 2(b) presents an overview of how our MTM operates." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 474, + 482, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 548 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 548 + ], + "type": "text", + "content": "Masking. We employ a variable masking ratio that follows a truncated normal distribution, as investigated in previous studies [4,5,34]. The tokens are randomly dropped according to the sampled masking ratio, and then the visible, unmasked token embeddings are fed into the model. By processing only a subset of token embeddings in the encoder, we significantly reduce the overall pre-training time and memory consumption, which aligns with the findings of MAE [24]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "text", + "content": "MTM Encoder & Decoder. We use ViT [17] as our encoder because it is able to process both the visible token embeddings and the original token embeddings, seamlessly accommodating their different lengths. During training, the encoder takes the visible token embeddings and outputs latent features that become the decoder input. Before decoding, the latent features are padded to the length of the original token embeddings using mask tokens M. Each mask token is a shared, learned vector indicating the presence of a missing token to be predicted. The decoder then estimates the original tokens " + }, + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "text", + "content": " from the padded features." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 245, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 245, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 245, + 101 + ], + "type": "text", + "content": "Lee and Park et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 118, + 186, + 223 + ], + "blocks": [ + { + "bbox": [ + 136, + 118, + 186, + 223 + ], + "lines": [ + { + "bbox": [ + 136, + 118, + 186, + 223 + ], + "spans": [ + { + "bbox": [ + 136, + 118, + 186, + 223 + ], + "type": "image", + "image_path": "bfff7f2568ee3f97bd988c21ad6ac5254aa5d362922dfaad3a67390dedcc233a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 178, + 226, + 308, + 237 + ], + "lines": [ + { + "bbox": [ + 178, + 226, + 308, + 237 + ], + "spans": [ + { + "bbox": [ + 178, + 226, + 308, + 237 + ], + "type": "text", + "content": "(a) Failure cases visualization" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 190, + 118, + 239, + 223 + ], + "blocks": [ + { + "bbox": [ + 190, + 118, + 239, + 223 + ], + "lines": [ + { + "bbox": [ + 190, + 118, + 239, + 223 + ], + "spans": [ + { + "bbox": [ + 190, + 118, + 239, + 223 + ], + "type": "image", + "image_path": "37f0cc45980f6f5ece0cb7f7ce5a3e901c84c72294edd0d97c36cc4fba1c2563.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 247, + 482, + 346 + ], + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 346 + ], + "type": "text", + "content": "Fig. 3: Data augmentations with tokens. Each ViT-VQGAN decoded image is reconstructed from a given RGB image after undergoing specific data augmentations. We observe that naively adopting these methods results in incorrect tokenization: 1) Token w/ hFlip demonstrates spatial information collapse during tokenization; 2) Token w/ RRC shows interdependence between neighboring token embeddings. We note the reconstructed images fail to preserve the images' details, suggesting that this incurs ineffectiveness of tokenization. Furthermore, we report top-1 accuracies (ViT-B) on ImageNet-1k according to the augmentations applied during training. All the augmentations are de-facto default training setups for vision transformers [37,54,56]." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 241, + 118, + 290, + 223 + ], + "blocks": [ + { + "bbox": [ + 241, + 118, + 290, + 223 + ], + "lines": [ + { + "bbox": [ + 241, + 118, + 290, + 223 + ], + "spans": [ + { + "bbox": [ + 241, + 118, + 290, + 223 + ], + "type": "image", + "image_path": "f34aa1f77be4ef74f7d0b9bad29499ebe5a9428bb29c4b47f7478758d29a519f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 292, + 118, + 340, + 223 + ], + "blocks": [ + { + "bbox": [ + 292, + 118, + 340, + 223 + ], + "lines": [ + { + "bbox": [ + 292, + 118, + 340, + 223 + ], + "spans": [ + { + "bbox": [ + 292, + 118, + 340, + 223 + ], + "type": "image", + "image_path": "371a44b8363fc6f0990aa4993e9f607dd47d1a289ae80827c36e3443b9448f1a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 349, + 115, + 470, + 224 + ], + "blocks": [ + { + "bbox": [ + 349, + 115, + 470, + 224 + ], + "lines": [ + { + "bbox": [ + 349, + 115, + 470, + 224 + ], + "spans": [ + { + "bbox": [ + 349, + 115, + 470, + 224 + ], + "type": "image", + "image_path": "de59c1cd87ba0ee0c3bff7f9980f770a286ed455da551754ed64ea2337fb05a3.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 350, + 226, + 479, + 238 + ], + "lines": [ + { + "bbox": [ + 350, + 226, + 479, + 238 + ], + "spans": [ + { + "bbox": [ + 350, + 226, + 479, + 238 + ], + "type": "text", + "content": "(b) Performance degradation" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 369, + 482, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 482, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 482, + 418 + ], + "type": "text", + "content": "Training Objective. MTM learns to accurately predict the original tokens of each masked token from the visible tokens. In particular, we use a cross-entropy loss between the decoder output and the ground truth tokens as the training objective, which is denoted by " + }, + { + "bbox": [ + 130, + 369, + 482, + 418 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{recon}}" + }, + { + "bbox": [ + 130, + 369, + 482, + 418 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 255, + 428, + 480, + 442 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 428, + 480, + 442 + ], + "spans": [ + { + "bbox": [ + 255, + 428, + 480, + 442 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e c o n}} = \\mathrm {C E} \\left(T _ {M} ^ {\\prime}, T _ {M}\\right), \\tag {1}", + "image_path": "a0662a919fc4b76d6a816ccdce582396ea1c50fb83bb73a6fed17799ba8922b7.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 445, + 482, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 445, + 482, + 482 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 482, + 482 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 445, + 482, + 482 + ], + "type": "inline_equation", + "content": "T_M'" + }, + { + "bbox": [ + 130, + 445, + 482, + 482 + ], + "type": "text", + "content": " denotes the predictions of the model for the masked tokens, and " + }, + { + "bbox": [ + 130, + 445, + 482, + 482 + ], + "type": "inline_equation", + "content": "T_M" + }, + { + "bbox": [ + 130, + 445, + 482, + 482 + ], + "type": "text", + "content": " corresponds to the ground truth tokens for " + }, + { + "bbox": [ + 130, + 445, + 482, + 482 + ], + "type": "inline_equation", + "content": "T_M'" + }, + { + "bbox": [ + 130, + 445, + 482, + 482 + ], + "type": "text", + "content": ". Following MAE [24], this loss is optimized only for the masked tokens." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 498, + 317, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 498, + 317, + 510 + ], + "spans": [ + { + "bbox": [ + 131, + 498, + 317, + 510 + ], + "type": "text", + "content": "3.3 Data Augmentation for Tokens" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 517, + 482, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 517, + 482, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 517, + 482, + 601 + ], + "type": "text", + "content": "While Masked Token Modeling (MTM) improves the storage efficiency of the token-based framework, there is still a performance bottleneck in improving token-based training. Since self-supervised approaches often require more data due to the lack of labels, exploring data augmentation is important to take full advantage of MTM. In this section, we analyze why existing data augmentations designed for pixel images are ineffective in the token domain. We then present TokenAdapt and ColorAdapt as solutions for token-based data augmentation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Limited Data Augmentations for Tokens. We explore the challenges of pixel-based augmentation in the token domain and identify two main reasons: (1) spatial information collapse during tokenization and (2) inter-token dependencies. The discrete tokenization process condenses a " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "n \\times n" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " 2D image patch" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 169, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 169, + 91, + 448, + 102 + ], + "type": "text", + "content": "SeiT++: Masked Token Modeling Improves Storage-efficient Training" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 159, + 114, + 455, + 180 + ], + "blocks": [ + { + "bbox": [ + 159, + 114, + 455, + 180 + ], + "lines": [ + { + "bbox": [ + 159, + 114, + 455, + 180 + ], + "spans": [ + { + "bbox": [ + 159, + 114, + 455, + 180 + ], + "type": "image", + "image_path": "d9b09dd8744d67ef668a6f82106c8a5a3c91181751e91e7fb32529c4fbd1f02b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 190, + 482, + 236 + ], + "lines": [ + { + "bbox": [ + 130, + 190, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 190, + 482, + 236 + ], + "type": "text", + "content": "Fig. 4: TokenAdapt processing pipeline. TokenAdapt aims to enhance the compatibility of token embeddings with pixel-based data augmentations by converting them into augmentation-compatible space, applying augmentations, and reverting them back to the original token embedding space." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 259, + 479, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 479, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 479, + 319 + ], + "type": "text", + "content": "into a single " + }, + { + "bbox": [ + 130, + 259, + 479, + 319 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 259, + 479, + 319 + ], + "type": "text", + "content": "-dimensional 1D vector (i.e., token embedding), causing a collapse of spatial information. This collapse hinders augmentations that rely on spatial details (e.g., horizontal flip). Furthermore, unlike independent image pixels, token embeddings are inter-correlated. This complicates augmentations involving interpolation (e.g., Resize or Mixup [68]), by introducing undesired artifacts." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 320, + 482, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 320, + 482, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 320, + 482, + 429 + ], + "type": "text", + "content": "Figure 3 shows the limitations of pixel-based augmentations in the token domain. A horizontal Flip (hFlip) distorts both the diagonal line (top row) and the cat's silhouette (bottom row) due to spatial information collapse. This occurs because only the inter-token spatial relationship is flipped, while the inner-token spatial arrangement remains unchanged. Augmentations involving interpolation can lead to undesired artifacts, such as color leakage in the case of Random Resized Crop (RRC), due to interdependence between neighboring embeddings. Therefore, we suspect that the direct application of pixel-based augmentations to token embedding may negatively affect token-based training." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 449, + 482, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 556 + ], + "type": "text", + "content": "TokenAdapt. The concept of TokenAdapt is to establish a feature space that allows for conventional pixel-based augmentations while facilitating the conversion and inversion of pre-extracted tokens with minimal computational overhead. This approach offers several advantages for our TokenAdapt. Firstly, it enables the direct reuse of well-designed pixel-based data augmentation policies. Additionally, the augmented token can be easily computed through the conversion-augmentation-reverse process. Finally, it ensures that the feature spaces for both the original and augmented tokens are preserved identically. This allows for the seamless use of existing model architectures or pre-trained weights." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "The goal of TokenAdapt is to design an augmentation function that converts the input tokens " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " to augmented tokens " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "T^{\\mathbf{A}} = [t_1^{\\mathbf{A}}, t_2^{\\mathbf{A}}, \\dots, t_n^{\\mathbf{A}}]" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": ". Note that the augmented tokens should also be quantized " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "t_i^{\\mathbf{A}} \\in \\mathbb{Q}" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": ". Figure 4 depicts the overall pipeline of TokenAdapt. The initial step of TokenAdapt is transforming token embeddings " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "Z_T" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": ", computed from " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " and codebook " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": ", into a representation " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "S_T \\in \\mathbb{R}^{h \\times w \\times d}" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " in augmentation-compatible space " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": ". This transformation is achieved using a conversion function " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "S_T = f(Z_T)" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": ". We apply augmentation " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " at augmentation-compatible space " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " and build an augmented representation " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "S_T^{\\mathbf{A}} = \\mathbf{A}(S_T)" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": ". Then, the augmented representation is reverted to" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 245, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 245, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 245, + 100 + ], + "type": "text", + "content": "Lee and Park et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 115, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 115, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 131, + 115, + 479, + 163 + ], + "type": "text", + "content": "token embedding space by a reverse transformation " + }, + { + "bbox": [ + 131, + 115, + 479, + 163 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 131, + 115, + 479, + 163 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 115, + 479, + 163 + ], + "type": "inline_equation", + "content": "Z_T^{\\mathbf{A}} = g(S_T^{\\mathbf{A}})" + }, + { + "bbox": [ + 131, + 115, + 479, + 163 + ], + "type": "text", + "content": ". At last, the augmented tokens in embedding space " + }, + { + "bbox": [ + 131, + 115, + 479, + 163 + ], + "type": "inline_equation", + "content": "Z_T^{\\mathbf{A}}" + }, + { + "bbox": [ + 131, + 115, + 479, + 163 + ], + "type": "text", + "content": " are quantized to codebook index as " + }, + { + "bbox": [ + 131, + 115, + 479, + 163 + ], + "type": "inline_equation", + "content": "T^{\\mathbf{A}} = \\mathbf{q}_{\\mathcal{Z}}(Z_T^{\\mathbf{A}})" + }, + { + "bbox": [ + 131, + 115, + 479, + 163 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 131, + 115, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_{\\mathcal{Z}}" + }, + { + "bbox": [ + 131, + 115, + 479, + 163 + ], + "type": "text", + "content": " indicates the vector quantization process. In summary, the augmentation process of TokenAdapt is as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 225, + 175, + 480, + 190 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 175, + 480, + 190 + ], + "spans": [ + { + "bbox": [ + 225, + 175, + 480, + 190 + ], + "type": "interline_equation", + "content": "Z _ {T} ^ {\\mathbf {A}} = g (\\mathbf {A} (f (Z _ {T}))), \\quad T ^ {\\mathbf {A}} = \\mathbf {q} _ {\\mathcal {Z}} \\left(Z _ {T} ^ {\\mathbf {A}}\\right). \\tag {2}", + "image_path": "9e26186be20059541a6ee8a766835d93d9905c9e21d73680c6ac8d87205735d1.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 194, + 479, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 194, + 479, + 253 + ], + "spans": [ + { + "bbox": [ + 131, + 194, + 479, + 253 + ], + "type": "text", + "content": "The conversion function " + }, + { + "bbox": [ + 131, + 194, + 479, + 253 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 131, + 194, + 479, + 253 + ], + "type": "text", + "content": " and the reverse transformation " + }, + { + "bbox": [ + 131, + 194, + 479, + 253 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 131, + 194, + 479, + 253 + ], + "type": "text", + "content": " are learned from paired data of tokens " + }, + { + "bbox": [ + 131, + 194, + 479, + 253 + ], + "type": "inline_equation", + "content": "(T_x, T_{\\mathbf{A}(x)})" + }, + { + "bbox": [ + 131, + 194, + 479, + 253 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 131, + 194, + 479, + 253 + ], + "type": "inline_equation", + "content": "T_x" + }, + { + "bbox": [ + 131, + 194, + 479, + 253 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 194, + 479, + 253 + ], + "type": "inline_equation", + "content": "T_{\\mathbf{A}(x)}" + }, + { + "bbox": [ + 131, + 194, + 479, + 253 + ], + "type": "text", + "content": " denote tokens from an image " + }, + { + "bbox": [ + 131, + 194, + 479, + 253 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 131, + 194, + 479, + 253 + ], + "type": "text", + "content": " and its augmented image " + }, + { + "bbox": [ + 131, + 194, + 479, + 253 + ], + "type": "inline_equation", + "content": "\\mathbf{A}(x)" + }, + { + "bbox": [ + 131, + 194, + 479, + 253 + ], + "type": "text", + "content": ", respectively. We use the cross-entropy loss between " + }, + { + "bbox": [ + 131, + 194, + 479, + 253 + ], + "type": "inline_equation", + "content": "Z_{T_x}^{\\mathbf{A}}" + }, + { + "bbox": [ + 131, + 194, + 479, + 253 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 194, + 479, + 253 + ], + "type": "inline_equation", + "content": "Z_{T_{\\mathbf{A}(x)}}" + }, + { + "bbox": [ + 131, + 194, + 479, + 253 + ], + "type": "text", + "content": " as our objective function " + }, + { + "bbox": [ + 131, + 194, + 479, + 253 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{TA}}" + }, + { + "bbox": [ + 131, + 194, + 479, + 253 + ], + "type": "text", + "content": " for learning our TokenAdapt module:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 255, + 266, + 480, + 280 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 266, + 480, + 280 + ], + "spans": [ + { + "bbox": [ + 255, + 266, + 480, + 280 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {T A}} = \\operatorname {C E} \\left(Z _ {T _ {x}} ^ {\\mathbf {A}}, Z _ {T _ {\\mathbf {A} (x)}}\\right). \\tag {3}", + "image_path": "6b862db4da7a7d500523c9ccbf57cbb8ff003a3c92b61ec710697e960c4a441e.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 285, + 479, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 479, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 479, + 331 + ], + "type": "text", + "content": "TokenAdapt deals with geometric data augmentations (e.g., flip, resize, crop, affine) and Mixup [68]. These augmentations are commonly used in various training methods, but they can cause a performance drop in token-based training, as shown in Figure 3(b)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 332, + 480, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 332, + 480, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 332, + 480, + 416 + ], + "type": "text", + "content": "Once trained, the TokenAdapt module is used to augment a sequence of tokens " + }, + { + "bbox": [ + 130, + 332, + 480, + 416 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 332, + 480, + 416 + ], + "type": "text", + "content": " into an augmented sequence " + }, + { + "bbox": [ + 130, + 332, + 480, + 416 + ], + "type": "inline_equation", + "content": "T^{\\mathbf{A}}" + }, + { + "bbox": [ + 130, + 332, + 480, + 416 + ], + "type": "text", + "content": " across different datasets and downstream tasks. Importantly, the augmentation-compatible feature space, which does not rely on specific semantic knowledge or alignment with downstream tasks, makes it dataset and task-agnostic. The inherent flexibility of TokenAdapt makes it highly generalizable, allowing direct application to other datasets without the need for additional training procedures." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 434, + 479, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 434, + 479, + 541 + ], + "spans": [ + { + "bbox": [ + 131, + 434, + 479, + 541 + ], + "type": "text", + "content": "ColorAdapt. As witnessed in Cubuk et al. [12], color-based augmentations are often more effective than geometric ones in certain datasets and are an important component of data augmentation strategies. For example, color-based augmentations are useful for dealing with various environmental changes caused by different lighting conditions (e.g., day and night) or weather conditions (e.g., snow). However, integrating them into TokenAdapt is challenging because most color-based augmentations are designed to work only in the pixel domain. To address this, we propose ColorAdapt, a color-based token augmentation inspired by the method [27]. Formally, the ColorAdapt function " + }, + { + "bbox": [ + 131, + 434, + 479, + 541 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 131, + 434, + 479, + 541 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 200, + 552, + 480, + 577 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 552, + 480, + 577 + ], + "spans": [ + { + "bbox": [ + 200, + 552, + 480, + 577 + ], + "type": "interline_equation", + "content": "\\mathcal {C} \\left(Z _ {T _ {1}}, Z _ {T _ {2}}\\right) = \\sigma \\left(Z _ {T _ {2}}\\right) \\left(\\frac {Z _ {T _ {1}} - \\mu \\left(Z _ {T _ {1}}\\right)}{\\sigma \\left(Z _ {T _ {1}}\\right)}\\right) + \\mu \\left(Z _ {T _ {2}}\\right), \\tag {4}", + "image_path": "6ba307447d56dd48f657ad2e49cf91d10f9e0dfdc15809e5c16958df3760315a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 581, + 479, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 479, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 479, + 605 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 581, + 479, + 605 + ], + "type": "inline_equation", + "content": "\\mu(Z_T)" + }, + { + "bbox": [ + 130, + 581, + 479, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 581, + 479, + 605 + ], + "type": "inline_equation", + "content": "\\sigma(Z_T)" + }, + { + "bbox": [ + 130, + 581, + 479, + 605 + ], + "type": "text", + "content": " are the mean and standard deviation of each token embedding " + }, + { + "bbox": [ + 130, + 581, + 479, + 605 + ], + "type": "inline_equation", + "content": "Z_T" + }, + { + "bbox": [ + 130, + 581, + 479, + 605 + ], + "type": "text", + "content": " across its spatial dimension, computed for each channel." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "type": "text", + "content": "ColorAdapt mimics color-changing augmentations in the pixel domain while preserving the encoded object structure by transforming the global statistics within each channel of the token embedding. Figure 5 shows the images decoded from various color-augmented tokens. This illustrates how ColorAdapt affects the visual properties of the augmented tokens. By incorporating ColorAdapt" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 168, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 168, + 91, + 448, + 103 + ], + "type": "text", + "content": "SeiT++: Masked Token Modeling Improves Storage-efficient Training" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 178, + 159 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 178, + 159 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 178, + 159 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 178, + 159 + ], + "type": "image", + "image_path": "4508aa871791343080e71200dba7cddbff76462a5c0cfec5d0e51c2e3adcc971.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 142, + 160, + 168, + 168 + ], + "lines": [ + { + "bbox": [ + 142, + 160, + 168, + 168 + ], + "spans": [ + { + "bbox": [ + 142, + 160, + 168, + 168 + ], + "type": "text", + "content": "Input data" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 179, + 115, + 220, + 159 + ], + "blocks": [ + { + "bbox": [ + 179, + 115, + 220, + 159 + ], + "lines": [ + { + "bbox": [ + 179, + 115, + 220, + 159 + ], + "spans": [ + { + "bbox": [ + 179, + 115, + 220, + 159 + ], + "type": "image", + "image_path": "c85dc23a13cec7b64d1d81d12f85aabd7a074eebf84fc4511f8d8e097d465407.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 185, + 160, + 212, + 167 + ], + "lines": [ + { + "bbox": [ + 185, + 160, + 212, + 167 + ], + "spans": [ + { + "bbox": [ + 185, + 160, + 212, + 167 + ], + "type": "text", + "content": "Brightness" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 220, + 115, + 261, + 159 + ], + "blocks": [ + { + "bbox": [ + 220, + 115, + 261, + 159 + ], + "lines": [ + { + "bbox": [ + 220, + 115, + 261, + 159 + ], + "spans": [ + { + "bbox": [ + 220, + 115, + 261, + 159 + ], + "type": "image", + "image_path": "0fa165838fc446140e9278ad22cdc9f518d5854db63f031d700b19e6e2f663b6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 227, + 160, + 255, + 167 + ], + "lines": [ + { + "bbox": [ + 227, + 160, + 255, + 167 + ], + "spans": [ + { + "bbox": [ + 227, + 160, + 255, + 167 + ], + "type": "text", + "content": "Emb-Noise" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 178, + 482, + 236 + ], + "lines": [ + { + "bbox": [ + 130, + 178, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 178, + 482, + 236 + ], + "type": "text", + "content": "Fig. 5: ColorAdapt provides more reasonable color changes. We present ViT-VQGAN decoded images to verify the quality of tokenizations after color changes. We use the brightness function with a factor of 0.2 following the implementation [46]. Emb-Noise [44] is the color-based token augmentation. Notably, our ColorAdapt effectively preserves object structure in contrast to the failure of the counterparts." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 262, + 115, + 307, + 159 + ], + "blocks": [ + { + "bbox": [ + 262, + 115, + 307, + 159 + ], + "lines": [ + { + "bbox": [ + 262, + 115, + 307, + 159 + ], + "spans": [ + { + "bbox": [ + 262, + 115, + 307, + 159 + ], + "type": "image", + "image_path": "27d01be32f7de5336c7f944d86429cd3033221d1c3f83fbe4d3bc5a07dc46a6c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 270, + 160, + 300, + 168 + ], + "lines": [ + { + "bbox": [ + 270, + 160, + 300, + 168 + ], + "spans": [ + { + "bbox": [ + 270, + 160, + 300, + 168 + ], + "type": "text", + "content": "ColorAdapt" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 115, + 352, + 159 + ], + "blocks": [ + { + "bbox": [ + 307, + 115, + 352, + 159 + ], + "lines": [ + { + "bbox": [ + 307, + 115, + 352, + 159 + ], + "spans": [ + { + "bbox": [ + 307, + 115, + 352, + 159 + ], + "type": "image", + "image_path": "713fa19fa1f0869bdc8ae3bea4ea6c0f65b67ac67890e165cef9cfa69459ea4c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 316, + 160, + 343, + 168 + ], + "lines": [ + { + "bbox": [ + 316, + 160, + 343, + 168 + ], + "spans": [ + { + "bbox": [ + 316, + 160, + 343, + 168 + ], + "type": "text", + "content": "Input data" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 353, + 115, + 394, + 159 + ], + "blocks": [ + { + "bbox": [ + 353, + 115, + 394, + 159 + ], + "lines": [ + { + "bbox": [ + 353, + 115, + 394, + 159 + ], + "spans": [ + { + "bbox": [ + 353, + 115, + 394, + 159 + ], + "type": "image", + "image_path": "89c295adbe5f70ab45ac7762e93e29835192b012fc0b4ef63da1f2f6039e30dc.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 359, + 160, + 386, + 168 + ], + "lines": [ + { + "bbox": [ + 359, + 160, + 386, + 168 + ], + "spans": [ + { + "bbox": [ + 359, + 160, + 386, + 168 + ], + "type": "text", + "content": "Brightness" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 395, + 115, + 433, + 159 + ], + "blocks": [ + { + "bbox": [ + 395, + 115, + 433, + 159 + ], + "lines": [ + { + "bbox": [ + 395, + 115, + 433, + 159 + ], + "spans": [ + { + "bbox": [ + 395, + 115, + 433, + 159 + ], + "type": "image", + "image_path": "afac642f0f9ea9ed6a1849c22bda5a2a73eef4b6b99b8eeec05ee8df522fd560.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 400, + 160, + 430, + 168 + ], + "lines": [ + { + "bbox": [ + 400, + 160, + 430, + 168 + ], + "spans": [ + { + "bbox": [ + 400, + 160, + 430, + 168 + ], + "type": "text", + "content": "Emb-Noise" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 436, + 115, + 481, + 159 + ], + "blocks": [ + { + "bbox": [ + 436, + 115, + 481, + 159 + ], + "lines": [ + { + "bbox": [ + 436, + 115, + 481, + 159 + ], + "spans": [ + { + "bbox": [ + 436, + 115, + 481, + 159 + ], + "type": "image", + "image_path": "26bde3657dd73f534634fef3fe4a0463330841d217d7c65981b2f4c66f8b696e.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 444, + 160, + 474, + 168 + ], + "lines": [ + { + "bbox": [ + 444, + 160, + 474, + 168 + ], + "spans": [ + { + "bbox": [ + 444, + 160, + 474, + 168 + ], + "type": "text", + "content": "ColorAdapt" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "table", + "bbox": [ + 142, + 293, + 470, + 434 + ], + "blocks": [ + { + "bbox": [ + 130, + 249, + 482, + 285 + ], + "lines": [ + { + "bbox": [ + 130, + 249, + 482, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 249, + 482, + 285 + ], + "type": "text", + "content": "Table 1: Storage-efficient ImageNet-1k classification. We report top-1 accuracies (ViT-B) and compression ratio on ImageNet-1k using various data storage reduction methods. Our SeiT++ shows significant storage efficiency compared to competitors." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 142, + 293, + 470, + 434 + ], + "lines": [ + { + "bbox": [ + 142, + 293, + 470, + 434 + ], + "spans": [ + { + "bbox": [ + 142, + 293, + 470, + 434 + ], + "type": "table", + "html": "
MethodInput # of images Top 1 Acc.Dataset storage size
Full pixelsImage1,281 k81.8140.0 GB
Uniform random sampling512 k74.054.6 GB (39%)
C-score [29] based samplingImage512 k73.353.3 GB (38%)
Adjusting JPEG quality factor to 51,281 k74.611.0 GB (8%)
Adjusting image resolution (× 0.2)1,281 k75.29.6 GB (7%)
SRe2L [65]Image200 k65.92.6 GB (2%)
G-VBSM [52]100 k63.71.6 GB (1%)
SeiT [44]Token1,281 k74.01.4 GB (1%)
SeiT++1,281 k77.81.4 GB (1%)
", + "image_path": "b529451ec4ec4543b14dc4c12a79edfa626422d0c2437fc791edc2804a150a16.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "table_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 459, + 482, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 459, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 459, + 482, + 483 + ], + "type": "text", + "content": "into our token augmentation strategy, we improve the data diversity, exploring new possibilities for training more robust and generalizable vision models." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 511, + 224, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 511, + 224, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 511, + 224, + 525 + ], + "type": "text", + "content": "4 Experiment" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "In this section, we validate the effectiveness of SeiT++ on various scenarios. First, we evaluate the performance of our method in the context of storage-efficient ImageNet-1k classification and transfer learning on fine-grained datasets. Then, we demonstrate the adaptability of token-based vision training to ADE-20k semantic segmentation, showing additional improvements facilitated by SeiT++. In addition, we present experimental results illustrating how our strategies contribute to enhanced model generalizability in robustness evaluation. We also explore the applicability of our augmentation strategies to alternative input formats, demonstrating its extensibility. Unless otherwise stated, SeiT++ refers to the combination of MTM and our token augmentation strategies." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 245, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 245, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 245, + 100 + ], + "type": "text", + "content": "Lee and Park et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 143, + 156, + 471, + 258 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 149 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 149 + ], + "type": "text", + "content": "Table 2: Token-based ImageNet-1k Classification. We report top-1 accuracies (ViT-B) on ImageNet-1k. Note that SeiT++ w/o MTM means the original SeiT [44] training with our token augmentation strategies." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 143, + 156, + 471, + 258 + ], + "lines": [ + { + "bbox": [ + 143, + 156, + 471, + 258 + ], + "spans": [ + { + "bbox": [ + 143, + 156, + 471, + 258 + ], + "type": "table", + "html": "
StorageSeiTSeiT++
w/o MTMw/ MTMw/o MTMw/ MTM
1.4 GB74.075.175.5 (+1.5)77.8 (+2.7)
1.1 GB70.274.173.1 (+2.9)76.5 (+2.4)
0.8 GB66.370.669.1 (+2.8)74.1 (+3.5)
0.5 GB59.764.563.8 (+4.1)70.0 (+5.5)
0.3 GB47.253.951.2 (+4.0)60.6 (+6.7)
", + "image_path": "6c90a73afb3ff613605618605b88a1f768cb01d5a368b96ddbf3bac710ff0da5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 167, + 309, + 448, + 369 + ], + "blocks": [ + { + "bbox": [ + 130, + 266, + 482, + 300 + ], + "lines": [ + { + "bbox": [ + 130, + 266, + 482, + 300 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 482, + 300 + ], + "type": "text", + "content": "Table 3: Fine-grained classification. We report top-1 accuracies (ViT-B) on various fine-grained datasets. Our method consistently outperforms its counterpart through effective augmentations for tokens." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 167, + 309, + 448, + 369 + ], + "lines": [ + { + "bbox": [ + 167, + 309, + 448, + 369 + ], + "spans": [ + { + "bbox": [ + 167, + 309, + 448, + 369 + ], + "type": "table", + "html": "
MethodFlowersCarsiNat18iNat19
SeiT [44]93.579.743.150.1
SeiT++95.286.952.458.9
(+1.7)(+7.2)(+9.3)(+8.8)
", + "image_path": "ea67dffa43c22b4796eed92aa6a34ba0d4154243d99d1bab00dffe4986640f28.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 390, + 310, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 390, + 310, + 403 + ], + "spans": [ + { + "bbox": [ + 132, + 390, + 310, + 403 + ], + "type": "text", + "content": "4.1 Storage-efficient ImageNet-1k" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 411, + 485, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 485, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 485, + 579 + ], + "type": "text", + "content": "Table 1 summarizes the performance and compression ratio of storage-efficient ImageNet-1k classification. Our method is compared with image-based approaches, including uniform random sampling, C-score [29]-based sampling, adjusting image resolution, and adjusting JPEG quality factor. Despite a " + }, + { + "bbox": [ + 130, + 411, + 485, + 579 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 411, + 485, + 579 + ], + "type": "text", + "content": " compression ratio, when a ViT-B/16 model is pre-trained on tokenized ImageNet-1k using MTM and the ViT model is fine-tuned, our method achieves a " + }, + { + "bbox": [ + 130, + 411, + 485, + 579 + ], + "type": "inline_equation", + "content": "77.8\\%" + }, + { + "bbox": [ + 130, + 411, + 485, + 579 + ], + "type": "text", + "content": " top-1 accuracy, outperforming various image-based approaches. In Table 2, our method consistently outperforms the token-based learning baseline SeiT by " + }, + { + "bbox": [ + 130, + 411, + 485, + 579 + ], + "type": "inline_equation", + "content": "1.5 - 6.7\\% p" + }, + { + "bbox": [ + 130, + 411, + 485, + 579 + ], + "type": "text", + "content": " in top-1 accuracies. Notably, the effectiveness of our method becomes more pronounced as the available storage size decreases, highlighting its efficiency in low-storage scenarios. Furthermore, SeiT++ shows more significant performance gains with MTM. For a storage size of 1.4GB, SeiT++ achieves an improvement of " + }, + { + "bbox": [ + 130, + 411, + 485, + 579 + ], + "type": "inline_equation", + "content": "1.5\\% p" + }, + { + "bbox": [ + 130, + 411, + 485, + 579 + ], + "type": "text", + "content": " over SeiT without MTM, while with MTM the improvement increases to " + }, + { + "bbox": [ + 130, + 411, + 485, + 579 + ], + "type": "inline_equation", + "content": "2.7\\% p" + }, + { + "bbox": [ + 130, + 411, + 485, + 579 + ], + "type": "text", + "content": ". These results confirm that SeiT++ effectively takes advantage of MTM." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 597, + 294, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 597, + 294, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 597, + 294, + 609 + ], + "type": "text", + "content": "4.2 Fine-grained Classification" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "To further investigate the generalizability of SeiT++, we conduct experiments on four different fine-grained datasets: Flowers [43], StandfordCars [31], iNat-18 [57], and iNat-19 [58]. Specifically, we tokenize each dataset and then fine-tune the token-trained model (ViT-B) on ImageNet-1k with the tokenized dataset. Since" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 168, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 168, + 91, + 447, + 103 + ], + "type": "text", + "content": "SeiT++: Masked Token Modeling Improves Storage-efficient Training" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 176, + 146, + 438, + 182 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 4: ADE-20k Semantic Segmentation. We report mIoU on the ADE-20k validation set, highlighting the effectiveness of our approach in pixel-level classification." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 176, + 146, + 438, + 182 + ], + "lines": [ + { + "bbox": [ + 176, + 146, + 438, + 182 + ], + "spans": [ + { + "bbox": [ + 176, + 146, + 438, + 182 + ], + "type": "table", + "html": "
MethodBackboneSeiTSeiT++
UPerNet [63]ViT-B [17]39.043.2 (+4.2)
", + "image_path": "da413636c5fd51feb7c9d7e334180eddf849c0c0504fab907e68fa4a720a1209.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 143, + 236, + 470, + 289 + ], + "blocks": [ + { + "bbox": [ + 132, + 193, + 481, + 226 + ], + "lines": [ + { + "bbox": [ + 132, + 193, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 193, + 481, + 226 + ], + "type": "text", + "content": "Table 5: Robustness evaluation. We report top-1 robust accuracies (ViT-B) against corruptions and domain shifts for each model trained on ImageNet-1k. Our method consistently outperforms its counterpart on various robustness benchmarks." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 143, + 236, + 470, + 289 + ], + "lines": [ + { + "bbox": [ + 143, + 236, + 470, + 289 + ], + "spans": [ + { + "bbox": [ + 143, + 236, + 470, + 289 + ], + "type": "table", + "html": "
MethodCleanGauss. NoiseGauss. BlurImageNet-RSketchObjectNet
SeiT [44]74.050.762.625.522.615.7
SeiT++75.558.666.830.227.718.3
(+1.5)(+7.9)(+4.2)(+4.7)(+5.1)(+2.6)
", + "image_path": "5b3a877d70faebe76834efa817f8e898385604871845f72914414420ae0d1d50.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 312, + 482, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 482, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 482, + 396 + ], + "type": "text", + "content": "fine-grained datasets require the classification of objects with intricate characteristics, they are particularly sensitive to undesirable noise (i.e., RRC in SeiT) or excessive perturbation (i.e., Emb-Noise) resulting from data augmentation in the token embedding space. Table 3 shows the quantitative performance (top-1 accuracies) on fine-grained datasets. SeiT++ consistently outperforms the performance of SeiT across various datasets. These results confirm the effectiveness of our proposed method in fine-grained classification." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 415, + 329, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 415, + 329, + 427 + ], + "spans": [ + { + "bbox": [ + 132, + 415, + 329, + 427 + ], + "type": "text", + "content": "4.3 ADE-20k Semantic Segmentation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 434, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 482, + 555 + ], + "type": "text", + "content": "A limitation of token-based learning is often substantiated by learning dense prediction tasks. We conjecture that our method could facilitate training for those tasks, such as semantic segmentation. We extract tokens from ADE-20k [72] and use the pre-trained ViT-B/16 model on tokenized ImageNet-1k to evaluate the effectiveness of our method in semantic segmentation. We then fine-tune the pre-trained model using the tokenized ADE-20k dataset. Table 4 reports the mean intersection-over-union (mIoU) as an evaluation metric. The results show a significant performance improvement " + }, + { + "bbox": [ + 130, + 434, + 482, + 555 + ], + "type": "inline_equation", + "content": "(+4.2\\%)" + }, + { + "bbox": [ + 130, + 434, + 482, + 555 + ], + "type": "text", + "content": " p mIoU) over SeiT. These results indicate that SeiT++ significantly improves model performance in the token-based dense prediction task." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 573, + 338, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 573, + 338, + 585 + ], + "spans": [ + { + "bbox": [ + 132, + 573, + 338, + 585 + ], + "type": "text", + "content": "4.4 Evaluation on Robustness Datasets" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "To validate the effectiveness of our augmentation strategies in terms of robustness, we further evaluate the robustness of our method on a set of benchmarks such as adding Gaussian noise [26] or Gaussian blur [26], ImageNet-R [25], Sketch [59], and ObjectNet [3]. Notably, we adopt SeiT++ w/o MTM to validate the impact of our augmentation strategies on model robustness. The results in Table 5 show that SeiT++ consistently outperforms SeiT in terms of robustness" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 245, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 245, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 245, + 101 + ], + "type": "text", + "content": "Lee and Park et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 207, + 168, + 406, + 237 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "text", + "content": "Table 6: Impact of the proposed augmentation strategies. We report the top-1 accuracies (ViT-S) on the ImageNet-100 validation set for various combinations of our token augmentation strategies. Notably, our TokenAdapt and ColorAdapt consistently improve model performance, exhibiting synergy when used together." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 207, + 168, + 406, + 237 + ], + "lines": [ + { + "bbox": [ + 207, + 168, + 406, + 237 + ], + "spans": [ + { + "bbox": [ + 207, + 168, + 406, + 237 + ], + "type": "table", + "html": "
FrameworkColorAdaptTokenAdaptTop 1 Acc.
SeiT [44]XX77.3
✓X78.3
X✓80.4
✓✓81.4
", + "image_path": "ec32eea75f984532940cb2aee2f606decfc428649032a65501c2765cb1a2eac7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 212, + 290, + 402, + 348 + ], + "blocks": [ + { + "bbox": [ + 130, + 247, + 482, + 282 + ], + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 282 + ], + "type": "text", + "content": "Table 7: Adaptibility of our method to different input. Input denotes tokenizers for our model. We report the top-1 accuracies (ViT-S) on ImageNet-100. For VQGAN, we used a publicly available VQGAN [18] trained on OpenImages [32] as a tokenizer." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 212, + 290, + 402, + 348 + ], + "lines": [ + { + "bbox": [ + 212, + 290, + 402, + 348 + ], + "spans": [ + { + "bbox": [ + 212, + 290, + 402, + 348 + ], + "type": "table", + "html": "
InputSeiTSeiT++
Token (ViT-VQGAN [66])77.381.4 (+4.1)
Token (VQGAN [18])81.883.9 (+2.1)
DCT coefficients69.571.0 (+1.5)
", + "image_path": "6c88c27a1e09ec262a943a2f6b2dc461a978e06a43f0c3ed04af1d089ff37b6f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 370, + 482, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 370, + 482, + 419 + ], + "spans": [ + { + "bbox": [ + 130, + 370, + 482, + 419 + ], + "type": "text", + "content": "accuracy. Previous studies [8, 55] observe that the key to the input pixel robustness depends on the pixel-level augmentations. Since our method enables pixel-based data augmentation to tokens, it achieves higher robustness than the baseline, which does not effectively exploit the pixel-based data augmentation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 436, + 239, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 436, + 239, + 449 + ], + "spans": [ + { + "bbox": [ + 131, + 436, + 239, + 449 + ], + "type": "text", + "content": "4.5 Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 456, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 456, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 456, + 482, + 529 + ], + "type": "text", + "content": "Impact of Each Augmentation Strategy. In Table 6, we report the impact of proposed token augmentations. In particular, incorporating ColorAdapt leads to a " + }, + { + "bbox": [ + 130, + 456, + 482, + 529 + ], + "type": "inline_equation", + "content": "1.0\\%" + }, + { + "bbox": [ + 130, + 456, + 482, + 529 + ], + "type": "text", + "content": " p performance improvement, highlighting the positive effect of preserving object structure during model training. Additionally, incorporating TokenAdapt yields an extra " + }, + { + "bbox": [ + 130, + 456, + 482, + 529 + ], + "type": "inline_equation", + "content": "3.1\\%" + }, + { + "bbox": [ + 130, + 456, + 482, + 529 + ], + "type": "text", + "content": " p improvement. These results confirm that exploring token augmentations is a key factor in fully exploiting pre-extracted tokens." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Extensibility of Our Augmentation Strategy. In this section, we conjecture the broad applicability of our augmentation strategies across various scenarios. To explore its extensibility, we delve into two distinct situations. Firstly, we examine the integration of our method with an alternative tokenizer—the OpenImages-trained VQGAN tokenizer [18], alongside the ImageNet-1k-trained ViT-VQGAN tokenizer [66]. The results in Table 7 consistently demonstrate improved performance on ImageNet-100 across different tokenizers. Additionally, we investigate the adaptability of our approach to a different input format by focusing on Discrete Cosine Transform (DCT) coefficients. Inspired by JPEG compression, we apply DCT to each " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " image patch, yielding quantized DCT" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 168, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 168, + 91, + 447, + 102 + ], + "type": "text", + "content": "SeiT++: Masked Token Modeling Improves Storage-efficient Training" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": "coefficients utilized as input data during model training. As shown in Table 7, our approach enhances top-1 accuracy, similar to the tokenizer scenario. This consistent improvement across diverse input formats underscores the universal applicability of our method, highlighting its potential as a widely applicable solution for diverse input representations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 191, + 280, + 204 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 191, + 280, + 204 + ], + "spans": [ + { + "bbox": [ + 132, + 191, + 280, + 204 + ], + "type": "text", + "content": "4.6 Implementation Details" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 208, + 482, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 208, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 208, + 482, + 304 + ], + "type": "text", + "content": "We use the pre-trained ViT-VQGAN tokenizer [66] for token extraction. For our TokenAdapt module, we employ a single transformer block for conversion and reverse function. The TokenAdapt module is trained for one epoch on ImageNet-1k with a batch size of 128, a learning rate of 0.001 with cosine scheduling, and a weight decay of 0.05. We follow the training recipes from MAGE [34], SeiT [44], and mmsegmentation [10] for masked token modeling (MTM), token-based classification, and semantic segmentation, respectively. Detailed hyperparameters and settings are available in the Appendix." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 319, + 220, + 333 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 319, + 220, + 333 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 220, + 333 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 342, + 482, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 535 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 535 + ], + "type": "text", + "content": "In this paper, we investigate the compatibility of SeiT with self-supervised learning approaches in token-based frameworks. Among various self-supervised algorithms, Masked Image Modeling (MIM) [24,64] has recently gained popularity for learning visual representations. Inspired by this, we extend SeiT by integrating Masked Token Modeling (MTM) for self-supervised pre-training. Additionally, recognizing that self-supervised approaches often require more data due to the lack of labels, we propose simple yet effective token-based augmentation strategies: TokenAdapt and ColorAdapt. TokenAdapt aligns the token embedding space with another embedding space compatible with pixel-based data augmentations. In addition, we introduce ColorAdapt, a color-changing augmentation for tokens. Our experimental results across various scenarios, including storage-efficient ImageNet-1k classification, fine-grained classification, ADE-20k semantic segmentation, and robustness benchmarks, consistently validate the effectiveness of our proposed method. Furthermore, the consistent performance improvement across diverse input formats highlights the potential of our approach as a widely applicable solution for various input representations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 545, + 482, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 606 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 606 + ], + "type": "text", + "content": "Limitations. While SeiT++ is capable of handling large-scale data without labels with optimized storage requirements, experiments on such large datasets (e.g., ImageNet-21k [14]) and huge models are not contained in this paper due to resource constraints. We believe the advantages of MTM and our data augmentation strategies can be seamlessly integrated with scalability." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Acknowledgements. Experiments are based on the Naver Smart Machine Learning NSML [30] platform. This research was supported by IITP grant funded by the Korea government(MSIT) and KEIT grant funded by the Korea government(MOTIE) (No. 2022-0-00680, 2022-0-01045)." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 245, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 245, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 245, + 101 + ], + "type": "text", + "content": "Lee and Park et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 139, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 139, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 139, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 138, + 139, + 480, + 172 + ], + "type": "text", + "content": "1. Bahng, H., Chun, S., Yun, S., Choo, J., Oh, S.J.: Learning de-biased representations with biased representations. In: International Conference on Machine Learning. pp. 528-539. PMLR (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 173, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 173, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 138, + 173, + 480, + 194 + ], + "type": "text", + "content": "2. Bao, H., Dong, L., Piao, S., Wei, F.: Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 194, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 194, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 138, + 194, + 480, + 237 + ], + "type": "text", + "content": "3. Barbu, A., Mayo, D., Alverio, J., Luo, W., Wang, C., Gutfreund, D., Tenenbaum, J., Katz, B.: Objectnet: A large-scale bias-controlled dataset for pushing the limits of object recognition models. Advances in neural information processing systems 32 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 238, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 238, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 138, + 238, + 480, + 281 + ], + "type": "text", + "content": "4. Chang, H., Zhang, H., Barber, J., Maschinot, A., Lezama, J., Jiang, L., Yang, M.H., Murphy, K., Freeman, W.T., Rubinstein, M., et al.: Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 282, + 480, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 282, + 480, + 315 + ], + "spans": [ + { + "bbox": [ + 138, + 282, + 480, + 315 + ], + "type": "text", + "content": "5. Chang, H., Zhang, H., Jiang, L., Liu, C., Freeman, W.T.: Maskgit: Masked generative image transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11315-11325 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 316, + 480, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 316, + 480, + 348 + ], + "spans": [ + { + "bbox": [ + 138, + 316, + 480, + 348 + ], + "type": "text", + "content": "6. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International conference on machine learning. pp. 1597-1607. PMLR (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 348, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 348, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 138, + 348, + 480, + 380 + ], + "type": "text", + "content": "7. Chen, X., Xie, S., He, K.: An empirical study of training self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9640-9649 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 380, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 380, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 138, + 380, + 480, + 413 + ], + "type": "text", + "content": "8. Chun, S., Oh, S.J., Yun, S., Han, D., Choe, J., Yoo, Y.: An empirical evaluation on robustness and uncertainty of regularization methods. arXiv preprint arXiv:2003.03879 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 413, + 480, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 413, + 480, + 446 + ], + "spans": [ + { + "bbox": [ + 138, + 413, + 480, + 446 + ], + "type": "text", + "content": "9. Coleman, C., Yeh, C., Mussmann, S., Mirzasoleiman, B., Bailis, P., Liang, P., Leskovec, J., Zaharia, M.: Selection via proxy: Efficient data selection for deep learning. arXiv preprint arXiv:1906.11829 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 447, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 447, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 447, + 480, + 468 + ], + "type": "text", + "content": "0. Contributors, M.: MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark. https://github.com/open-mmlab/mmsegmentation (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 468, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 468, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 138, + 468, + 480, + 501 + ], + "type": "text", + "content": "1. Croce, F., Hein, M.: Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In: International conference on machine learning. pp. 2206-2216. PMLR (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 501, + 480, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 501, + 480, + 534 + ], + "spans": [ + { + "bbox": [ + 138, + 501, + 480, + 534 + ], + "type": "text", + "content": "2. Cubuk, E.D., Zoph, B., Mane, D., Vasudevan, V., Le, Q.V.: Autoaugment: Learning augmentation strategies from data. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 113-123 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 534, + 480, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 534, + 480, + 577 + ], + "spans": [ + { + "bbox": [ + 138, + 534, + 480, + 577 + ], + "type": "text", + "content": "3. Cubuk, E.D., Zoph, B., Shlens, J., Le, Q.V.: Randaugment: Practical automated data augmentation with a reduced search space. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 702-703 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 577, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 577, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 138, + 577, + 480, + 610 + ], + "type": "text", + "content": "4. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 610, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 610, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 610, + 480, + 643 + ], + "type": "text", + "content": "5. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "6. DeVries, T., Taylor, G.W.: Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552 (2017)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 169, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 169, + 91, + 447, + 102 + ], + "type": "text", + "content": "SeiT++: Masked Token Modeling Improves Storage-efficient Training" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "text", + "content": "17. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 162, + 481, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 162, + 481, + 195 + ], + "spans": [ + { + "bbox": [ + 132, + 162, + 481, + 195 + ], + "type": "text", + "content": "18. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12873-12883 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 196, + 481, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 196, + 481, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 196, + 481, + 217 + ], + "type": "text", + "content": "19. Fang, Y., Dong, L., Bao, H., Wang, X., Wei, F.: Corrupted image modeling for self-supervised visual pre-training. arXiv preprint arXiv:2202.03382 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 218, + 481, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 218, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 218, + 481, + 239 + ], + "type": "text", + "content": "20. Ford, N., Gilmer, J., Carlini, N., Cubuk, D.: Adversarial examples are a natural consequence of test error in noise. arXiv preprint arXiv:1901.10513 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 240, + 481, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 240, + 481, + 274 + ], + "spans": [ + { + "bbox": [ + 132, + 240, + 481, + 274 + ], + "type": "text", + "content": "21. Geirhos, R., Rubisch, P., Michaelis, C., Bethge, M., Wichmann, F.A., Brendel, W.: Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. arXiv preprint arXiv:1811.12231 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 274, + 481, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 481, + 296 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 481, + 296 + ], + "type": "text", + "content": "22. Goodfellow, I.J., Shlens, J., Szegedy, C.: Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572 (2014)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 297, + 481, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 297, + 481, + 341 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 481, + 341 + ], + "type": "text", + "content": "23. Grill, J.B., Strub, F., Altché, F., Tallec, C., Richemond, P., Buchatskaya, E., Doersch, C., Avila Pires, B., Guo, Z., Gheshlaghi Azar, M., et al.: Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems 33, 21271-21284 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 342, + 481, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 481, + 374 + ], + "type": "text", + "content": "24. He, K., Chen, X., Xie, S., Li, Y., Dollár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 375, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 375, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 375, + 481, + 418 + ], + "type": "text", + "content": "25. Hendrycks, D., Basart, S., Mu, N., Kadavath, S., Wang, F., Dorundo, E., Desai, R., Zhu, T., Parajuli, S., Guo, M., Song, D., Steinhardt, J., Gilmer, J.: The many faces of robustness: A critical analysis of out-of-distribution generalization. ICCV (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 419, + 481, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 419, + 481, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 419, + 481, + 441 + ], + "type": "text", + "content": "26. Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 442, + 481, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 442, + 481, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 442, + 481, + 475 + ], + "type": "text", + "content": "27. Huang, X., Belongie, S.: Arbitrary style transfer in real-time with adaptive instance normalization. In: Proceedings of the IEEE international conference on computer vision. pp. 1501-1510 (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 476, + 481, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 476, + 481, + 520 + ], + "spans": [ + { + "bbox": [ + 132, + 476, + 481, + 520 + ], + "type": "text", + "content": "28. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International conference on machine learning. pp. 4904-4916. PMLR (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 521, + 481, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 521, + 481, + 553 + ], + "spans": [ + { + "bbox": [ + 132, + 521, + 481, + 553 + ], + "type": "text", + "content": "29. Jiang, Z., Zhang, C., Talwar, K., Mozer, M.C.: Characterizing structural regularities of labeled data in overparameterized models. arXiv preprint arXiv:2002.03206 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 554, + 481, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 554, + 481, + 587 + ], + "spans": [ + { + "bbox": [ + 132, + 554, + 481, + 587 + ], + "type": "text", + "content": "30. Kim, H., Kim, M., Seo, D., Kim, J., Park, H., Park, S., Jo, H., Kim, K., Yang, Y., Kim, Y., et al.: Nsml: Meet the mlaas platform with a real-world case study. arXiv preprint arXiv:1810.09957 (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 588, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 588, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 481, + 621 + ], + "type": "text", + "content": "31. Krause, J., Stark, M., Deng, J., Fei-Fei, L.: 3d object representations for fine-grained categorization. In: Proceedings of the IEEE international conference on computer vision workshops. pp. 554-561 (2013)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 621, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 481, + 666 + ], + "type": "text", + "content": "32. Kuznetsova, A., Rom, H., Alldrin, N., Uijlings, J., Krasin, I., Pont-Tuset, J., Kamali, S., Popov, S., Malloci, M., Kolesnikov, A., et al.: The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision 128(7), 1956-1981 (2020)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 245, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 245, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 245, + 100 + ], + "type": "text", + "content": "Lee and Park et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "type": "text", + "content": "33. Lee, S., Chun, S., Jung, S., Yun, S., Yoon, S.: Dataset condensation with contrastive signals. In: International Conference on Machine Learning. pp. 12352-12364. PMLR (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 149, + 481, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 481, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 481, + 193 + ], + "type": "text", + "content": "34. Li, T., Chang, H., Mishra, S., Zhang, H., Katabi, D., Krishnan, D.: Mage: Masked generative encoder to unify representation learning and image synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2142-2152 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 193, + 481, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 193, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 193, + 481, + 236 + ], + "type": "text", + "content": "35. Lin, S., Ji, R., Yan, C., Zhang, B., Cao, L., Ye, Q., Huang, F., Doermann, D.: Towards optimal structured cnn pruning via generative adversarial learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2790-2799 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 236, + 481, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 481, + 257 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 481, + 257 + ], + "type": "text", + "content": "36. Liu, J., Guo, J., Xu, D.: Apsnet: Toward adaptive point sampling for efficient 3d action recognition. IEEE Transactions on Image Processing 31, 5287-5302 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 257, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 257, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 481, + 300 + ], + "type": "text", + "content": "37. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 300, + 481, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 481, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 481, + 332 + ], + "type": "text", + "content": "38. Lopes, R.G., Yin, D., Poole, B., Gilmer, J., Cubuk, E.D.: Improving robustness without sacrificing accuracy with patch gaussian augmentation. arXiv preprint arXiv:1906.02611 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 332, + 481, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 481, + 374 + ], + "type": "text", + "content": "39. Ma, Y., Li, H., Zhang, Z., Guo, J., Zhang, S., Gong, R., Liu, X.: Annealing-based label-transfer learning for open world object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11454-11463 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 374, + 481, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 481, + 397 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 481, + 397 + ], + "type": "text", + "content": "40. Madry, A., Makelov, A., Schmidt, L., Tsipras, D., Vladu, A.: Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083 (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 397, + 481, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 481, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 481, + 439 + ], + "type": "text", + "content": "41. Mahajan, D., Girshick, R., Ramanathan, V., He, K., Paluri, M., Li, Y., Bharambe, A., Van Der Maaten, L.: Exploring the limits of weakly supervised pretraining. In: Proceedings of the European conference on computer vision (ECCV). pp. 181-196 (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 439, + 481, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 439, + 481, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 439, + 481, + 472 + ], + "type": "text", + "content": "42. Mei, Z., Ye, P., Ye, H., Li, B., Guo, J., Chen, T., Ouyang, W.: Automatic loss function search for adversarial unsupervised domain adaptation. IEEE Transactions on Circuits and Systems for Video Technology 33(10), 5868-5881 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 472, + 481, + 505 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 481, + 505 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 481, + 505 + ], + "type": "text", + "content": "43. Nilsback, M.E., Zisserman, A.: Automated flower classification over a large number of classes. In: 2008 Sixth Indian conference on computer vision, graphics & image processing. pp. 722-729. IEEE (2008)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 505, + 481, + 526 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 505, + 481, + 526 + ], + "spans": [ + { + "bbox": [ + 132, + 505, + 481, + 526 + ], + "type": "text", + "content": "44. Park, S., Chun, S., Heo, B., Kim, W., Yun, S.: Seit: Storage-efficient vision training with tokens using " + }, + { + "bbox": [ + 132, + 505, + 481, + 526 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 132, + 505, + 481, + 526 + ], + "type": "text", + "content": " of pixel storage. arXiv preprint arXiv:2303.11114 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 525, + 481, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 481, + 557 + ], + "type": "text", + "content": "45. Paul, M., Ganguli, S., Dziugaite, G.K.: Deep learning on a data diet: Finding important examples early in training. Advances in Neural Information Processing Systems 34, 20596-20607 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 557, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 481, + 601 + ], + "type": "text", + "content": "46. Riba, E., Mishkin, D., Ponsa, D., Rublee, E., Bradski, G.: Kornia: an open source differentiable computer vision library for pytorch. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3674-3683 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "type": "text", + "content": "47. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10684-10695 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "type": "text", + "content": "48. Rosasco, A., Carta, A., Cossu, A., Lomonaco, V., Bacciu, D.: Distilled replay: Overcoming forgetting through synthetic samples. In: International Workshop on Continual Semi-Supervised Learning. pp. 104-117. Springer (2021)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 169, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 169, + 91, + 447, + 102 + ], + "type": "text", + "content": "SeiT++: Masked Token Modeling Improves Storage-efficient Training" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 148 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 148 + ], + "type": "text", + "content": "49. Sangermano, M., Carta, A., Cossu, A., Bacciu, D.: Sample condensation in online continual learning. In: 2022 International Joint Conference on Neural Networks (IJCNN). pp. 01-08. IEEE (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 149, + 480, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 480, + 192 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 480, + 192 + ], + "type": "text", + "content": "50. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 193, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 193, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 193, + 480, + 224 + ], + "type": "text", + "content": "51. Scimeca, L., Oh, S.J., Chun, S., Poli, M., Yun, S.: Which shortcut cues will dnns choose? a study from the parameter-space perspective. arXiv preprint arXiv:2110.03095 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 225, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 225, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 225, + 480, + 267 + ], + "type": "text", + "content": "52. Shao, S., Yin, Z., Zhou, M., Zhang, X., Shen, Z.: Generalized large-scale data condensation via various backbone and statistical matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16709-16718 (2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 268, + 480, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 268, + 480, + 311 + ], + "spans": [ + { + "bbox": [ + 132, + 268, + 480, + 311 + ], + "type": "text", + "content": "53. Singh, M., Gustafson, L., Adcock, A., de Freitas Reis, V., Gedik, B., Kosaraju, R.P., Mahajan, D., Girshick, R., Dollar, P., Van Der Maaten, L.: Revisiting weakly supervised pre-training of visual perception models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 804-814 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 312, + 480, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 312, + 480, + 343 + ], + "spans": [ + { + "bbox": [ + 132, + 312, + 480, + 343 + ], + "type": "text", + "content": "54. Steiner, A., Kolesnikov, A., Zhai, X., Wightman, R., Uszkoreit, J., Beyer, L.: How to train your vit? data, augmentation, and regularization in vision transformers. arXiv preprint arXiv:2106.10270 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 344, + 480, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 344, + 480, + 375 + ], + "spans": [ + { + "bbox": [ + 132, + 344, + 480, + 375 + ], + "type": "text", + "content": "55. Taori, R., Dave, A., Shankar, V., Carlini, N., Recht, B., Schmidt, L.: Measuring robustness to natural distribution shifts in image classification. Advances in Neural Information Processing Systems 33, 18583-18599 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 376, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 376, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 376, + 480, + 407 + ], + "type": "text", + "content": "56. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jégou, H.: Training data-efficient image transformers & distillation through attention. In: International conference on machine learning. pp. 10347-10357. PMLR (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 408, + 480, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 408, + 480, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 408, + 480, + 439 + ], + "type": "text", + "content": "57. Van Horn, G., Mac Aodha, O., Song, Y., Cui, Y., Sun, C., Shepard, A., Adam, H., Perona, P., Belongie, S.: The inaturalist challenge 2018 dataset. arXiv preprint arXiv:1707.06642 (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 441, + 480, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 441, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 441, + 480, + 472 + ], + "type": "text", + "content": "58. Van Horn, G., Mac Aodha, O., Song, Y., Cui, Y., Sun, C., Shepard, A., Adam, H., Perona, P., Belongie, S.: The inaturalist challenge 2019 dataset. arXiv preprint arXiv:1707.06642 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 472, + 480, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 480, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 480, + 504 + ], + "type": "text", + "content": "59. Wang, H., Ge, S., Lipton, Z., Xing, E.P.: Learning robust global representations by penalizing local predictive power. In: Advances in Neural Information Processing Systems. pp. 10506-10518 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 505, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 505, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 505, + 480, + 525 + ], + "type": "text", + "content": "60. Wang, T., Zhu, J.Y., Torralba, A., Efros, A.A.: Dataset distillation. arXiv preprint arXiv:1811.10959 (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 526, + 480, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 526, + 480, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 526, + 480, + 547 + ], + "type": "text", + "content": "61. Wei, J., Zou, K.: Eda: Easy data augmentation techniques for boosting performance on text classification tasks. arXiv preprint arXiv:1901.11196 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 548, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 548, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 548, + 480, + 568 + ], + "type": "text", + "content": "62. Wightman, R., Touvron, H., Jégou, H.: Resnet strikes back: An improved training procedure in timm. arXiv preprint arXiv:2110.00476 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 569, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 569, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 569, + 480, + 601 + ], + "type": "text", + "content": "63. Xiao, T., Liu, Y., Zhou, B., Jiang, Y., Sun, J.: Unified perceptual parsing for scene understanding. In: Proceedings of the European conference on computer vision (ECCV). pp. 418-434 (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 601, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 480, + 632 + ], + "type": "text", + "content": "64. Xie, Z., Zhang, Z., Cao, Y., Lin, Y., Bao, J., Yao, Z., Dai, Q., Hu, H.: Simmim: A simple framework for masked image modeling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9653-9663 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "text", + "content": "65. Yin, Z., Xing, E., Shen, Z.: Squeeze, recover and relabel: Dataset condensation at imagenet scale from a new perspective. In: Thirty-seventh Conference on Neural Information Processing Systems (2023)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 245, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 245, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 245, + 100 + ], + "type": "text", + "content": "Lee and Park et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 325 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "66. Yu, J., Li, X., Koh, J.Y., Zhang, H., Pang, R., Qin, J., Ku, A., Xu, Y., Baldridge, J., Wu, Y.: Vector-quantized image modeling with improved vqgan. arXiv preprint arXiv:2110.04627 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "type": "text", + "content": "67. Yun, S., Han, D., Oh, S.J., Chun, S., Choe, J., Yoo, Y.: Cutmix: Regularization strategy to train strong classifiers with localizable features. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6023-6032 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 182, + 482, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 482, + 205 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 482, + 205 + ], + "type": "text", + "content": "68. Zhang, H., Cisse, M., Dauphin, Y.N., Lopez-Paz, D.: mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412 (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 205, + 482, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 482, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 482, + 237 + ], + "type": "text", + "content": "69. Zhao, B., Bilen, H.: Dataset condensation with differentiable siamese augmentation. In: International Conference on Machine Learning. pp. 12674-12685. PMLR (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 237, + 482, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 482, + 259 + ], + "type": "text", + "content": "70. Zhao, B., Mopuri, K.R., Bilen, H.: Dataset condensation with gradient matching. arXiv preprint arXiv:2006.05929 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "type": "text", + "content": "71. Zhong, Z., Zheng, L., Kang, G., Li, S., Yang, Y.: Random erasing data augmentation. In: Proceedings of the AAAI Conference on Artificial Intelligence (AAAI) (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 293, + 482, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 325 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 325 + ], + "type": "text", + "content": "72. Zhou, B., Zhao, H., Puig, X., Xiao, T., Fidler, S., Barriuso, A., Torralba, A.: Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision 127, 302-321 (2019)" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 169, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 169, + 91, + 447, + 102 + ], + "type": "text", + "content": "SeiT++: Masked Token Modeling Improves Storage-efficient Training" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file